diff --git a/.gitattributes b/.gitattributes index 1fcffcb2691c3a2e91f4932d0f79d8e1f643f418..0d89f0087ecea566ffa808f5c1d12c02e39070ec 100644 --- a/.gitattributes +++ b/.gitattributes @@ -1781,3 +1781,67 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text 2024/Realistic[[:space:]]Human[[:space:]]Motion[[:space:]]Generation[[:space:]]with[[:space:]]Cross-Diffusion[[:space:]]Models/5b1805dd-89b3-4df9-9e0a-1ed375e3f44c_origin.pdf filter=lfs diff=lfs merge=lfs -text 2024/Reason2Drive_[[:space:]]Towards[[:space:]]Interpretable[[:space:]]and[[:space:]]Chain-based[[:space:]]Reasoning[[:space:]]for[[:space:]]Autonomous[[:space:]]Driving/66f968e7-ebed-40e1-a3ff-b107a9813c0e_origin.pdf filter=lfs diff=lfs merge=lfs -text 2024/Rebalancing[[:space:]]Using[[:space:]]Estimated[[:space:]]Class[[:space:]]Distribution[[:space:]]for[[:space:]]Imbalanced[[:space:]]Semi-Supervised[[:space:]]Learning[[:space:]]under[[:space:]]Class[[:space:]]Distribution[[:space:]]Mismatch/c1f56d15-2542-45a0-80c7-7ebab6520f43_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Receler_[[:space:]]Reliable[[:space:]]Concept[[:space:]]Erasing[[:space:]]of[[:space:]]Text-to-Image[[:space:]]Diffusion[[:space:]]Models[[:space:]]via[[:space:]]Lightweight[[:space:]]Erasers/e96e787a-46f5-4d12-8c30-eca8de901ff9_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Reconstruction[[:space:]]and[[:space:]]Simulation[[:space:]]of[[:space:]]Elastic[[:space:]]Objects[[:space:]]with[[:space:]]Spring-Mass[[:space:]]3D[[:space:]]Gaussians/2eb8194f-eadd-43df-b4b2-bbf94906f9cc_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Rectify[[:space:]]the[[:space:]]Regression[[:space:]]Bias[[:space:]]in[[:space:]]Long-Tailed[[:space:]]Object[[:space:]]Detection/07bdb2f4-ff12-4767-80ed-dec2d3b31ce7_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/RecurrentBEV_[[:space:]]A[[:space:]]Long-term[[:space:]]Temporal[[:space:]]Fusion[[:space:]]Framework[[:space:]]for[[:space:]]Multi-view[[:space:]]3D[[:space:]]Detection/c66016d3-668d-4d64-aa39-f278cd973b7f_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Recursive[[:space:]]Visual[[:space:]]Programming/e57bdabb-77ec-4030-98f0-ae871f741b2d_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Ref-AVS_[[:space:]]Refer[[:space:]]and[[:space:]]Segment[[:space:]]Objects[[:space:]]in[[:space:]]Audio-Visual[[:space:]]Scenes/a65e28a6-a79d-4af3-846c-76b29927236e_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Referring[[:space:]]Atomic[[:space:]]Video[[:space:]]Action[[:space:]]Recognition/3c97fc3b-c407-40be-ae83-9d7a4bd31254_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Refine,[[:space:]]Discriminate[[:space:]]and[[:space:]]Align_[[:space:]]Stealing[[:space:]]Encoders[[:space:]]via[[:space:]]Sample-Wise[[:space:]]Prototypes[[:space:]]and[[:space:]]Multi-Relational[[:space:]]Extraction/2d1e904f-54f5-49e9-a836-f9f59c16b169_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Reflective[[:space:]]Instruction[[:space:]]Tuning_[[:space:]]Mitigating[[:space:]]Hallucinations[[:space:]]in[[:space:]]Large[[:space:]]Vision-Language[[:space:]]Models/8b065bde-2d53-44c1-a90a-d68b84da76bf_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Region-Adaptive[[:space:]]Transform[[:space:]]with[[:space:]]Segmentation[[:space:]]Prior[[:space:]]for[[:space:]]Image[[:space:]]Compression/7ddd6f0f-062f-4fcd-9635-2169878bc97b_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Region-Aware[[:space:]]Sequence-to-Sequence[[:space:]]Learning[[:space:]]for[[:space:]]Hyperspectral[[:space:]]Denoising/fb5f87c2-230f-4062-9b75-422a250a1410_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Region-Native[[:space:]]Visual[[:space:]]Tokenization/a10ef64d-036f-4197-bc89-c920baeb7773_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Region-aware[[:space:]]Distribution[[:space:]]Contrast_[[:space:]]A[[:space:]]Novel[[:space:]]Approach[[:space:]]to[[:space:]]Multi-Task[[:space:]]Partially[[:space:]]Supervised[[:space:]]Learning/b09b7220-34e4-48d5-bc47-6cde62cb650d_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Region-centric[[:space:]]Image-Language[[:space:]]Pretraining[[:space:]]for[[:space:]]Open-Vocabulary[[:space:]]Detection/be279718-23a8-4b0a-ac0c-2bca133ca6e7_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/RegionDrag_[[:space:]]Fast[[:space:]]Region-Based[[:space:]]Image[[:space:]]Editing[[:space:]]with[[:space:]]Diffusion[[:space:]]Models/7e845ddf-5fd4-480c-a68c-61d4b586b518_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Regularizing[[:space:]]Dynamic[[:space:]]Radiance[[:space:]]Fields[[:space:]]with[[:space:]]Kinematic[[:space:]]Fields/79573c22-a91b-4733-9726-bd12faf03469_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Regulating[[:space:]]Model[[:space:]]Reliance[[:space:]]on[[:space:]]Non-Robust[[:space:]]Features[[:space:]]by[[:space:]]Smoothing[[:space:]]Input[[:space:]]Marginal[[:space:]]Density/1125a5c6-c219-425e-8e55-ef1e3d508033_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Reinforcement[[:space:]]Learning[[:space:]]Friendly[[:space:]]Vision-Language[[:space:]]Model[[:space:]]for[[:space:]]Minecraft/c4770d1a-703c-494e-ba79-18744626ca70_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Reinforcement[[:space:]]Learning[[:space:]]Meets[[:space:]]Visual[[:space:]]Odometry/4903ec8e-7c0d-4a2f-9524-76e74784ea76_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Reinforcement[[:space:]]Learning[[:space:]]via[[:space:]]Auxillary[[:space:]]Task[[:space:]]Distillation/0d1f16d8-0d98-408f-b999-123ebb200f38_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Rejection[[:space:]]Sampling[[:space:]]IMLE_[[:space:]]Designing[[:space:]]Priors[[:space:]]for[[:space:]]Better[[:space:]]Few-Shot[[:space:]]Image[[:space:]]Synthesis/0d5c28ac-bfc0-4636-af16-85b85c6b5148_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Relation[[:space:]]DETR_[[:space:]]Exploring[[:space:]]Explicit[[:space:]]Position[[:space:]]Relation[[:space:]]Prior[[:space:]]for[[:space:]]Object[[:space:]]Detection/650a9d12-9a9f-48d6-b77e-bd08f66392b3_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Reliability[[:space:]]in[[:space:]]Semantic[[:space:]]Segmentation_[[:space:]]Can[[:space:]]We[[:space:]]Use[[:space:]]Synthetic[[:space:]]Data_/3d29ee45-5380-4719-adfd-255f20df0905_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Reliable[[:space:]]Spatial-Temporal[[:space:]]Voxels[[:space:]]For[[:space:]]Multi-Modal[[:space:]]Test-Time[[:space:]]Adaptation/c906fc92-ebc1-45f1-bc7d-fa0568ba9181_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Reliable[[:space:]]and[[:space:]]Efficient[[:space:]]Concept[[:space:]]Erasure[[:space:]]of[[:space:]]Text-to-Image[[:space:]]Diffusion[[:space:]]Models/d5988da3-2c65-4cde-838f-fe6bf86e8cd3_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Relightable[[:space:]]3D[[:space:]]Gaussians_[[:space:]]Realistic[[:space:]]Point[[:space:]]Cloud[[:space:]]Relighting[[:space:]]with[[:space:]]BRDF[[:space:]]Decomposition[[:space:]]and[[:space:]]Ray[[:space:]]Tracing/8f8cf818-147d-408c-857b-28f5e4e05e99_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Relightable[[:space:]]Neural[[:space:]]Actor[[:space:]]with[[:space:]]Intrinsic[[:space:]]Decomposition[[:space:]]and[[:space:]]Pose[[:space:]]Control/b67bc924-ef1e-4a87-8446-f01e4ef31ad6_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Remove[[:space:]]Projective[[:space:]]LiDAR[[:space:]]Depthmap[[:space:]]Artifacts[[:space:]]via[[:space:]]Exploiting[[:space:]]Epipolar[[:space:]]Geometry/0a1ce34c-ca2a-42ac-a914-2a125b92f644_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Removing[[:space:]]Distributional[[:space:]]Discrepancies[[:space:]]in[[:space:]]Captions[[:space:]]Improves[[:space:]]Image-Text[[:space:]]Alignment/0409d8d6-db45-4397-97c7-eeaa2a4590eb_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Removing[[:space:]]Rows[[:space:]]and[[:space:]]Columns[[:space:]]of[[:space:]]Tokens[[:space:]]in[[:space:]]Vision[[:space:]]Transformer[[:space:]]enables[[:space:]]Faster[[:space:]]Dense[[:space:]]Prediction[[:space:]]without[[:space:]]Retraining/8e177374-55c7-4790-9a57-d4f7dcccc52f_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/RepVF_[[:space:]]A[[:space:]]Unified[[:space:]]Vector[[:space:]]Fields[[:space:]]Representation[[:space:]]for[[:space:]]Multi-task[[:space:]]3D[[:space:]]Perception/9c289191-ecfe-49ca-a348-d2cad5492a03_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Repaint123_[[:space:]]Fast[[:space:]]and[[:space:]]High-quality[[:space:]]One[[:space:]]Image[[:space:]]to[[:space:]]3D[[:space:]]Generation[[:space:]]with[[:space:]]Progressive[[:space:]]Controllable[[:space:]]Repainting/a8257f3f-735a-49a4-ab19-4aa6628f29e5_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Representation[[:space:]]Enhancement-Stabilization_[[:space:]]Reducing[[:space:]]Bias-Variance[[:space:]]of[[:space:]]Domain[[:space:]]Generalization/f28a3a3b-4152-48f4-99b3-107ae4c33ff2_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Representing[[:space:]]Topological[[:space:]]Self-Similarity[[:space:]]Using[[:space:]]Fractal[[:space:]]Feature[[:space:]]Maps[[:space:]]for[[:space:]]Accurate[[:space:]]Segmentation[[:space:]]of[[:space:]]Tubular[[:space:]]Structures/494e6709-0fe6-4415-a992-88fda17780d8_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Reprojection[[:space:]]Errors[[:space:]]as[[:space:]]Prompts[[:space:]]for[[:space:]]Efficient[[:space:]]Scene[[:space:]]Coordinate[[:space:]]Regression/91ff6348-fcc0-42f1-b6d0-08cddca0de4a_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Reshaping[[:space:]]the[[:space:]]Online[[:space:]]Data[[:space:]]Buffering[[:space:]]and[[:space:]]Organizing[[:space:]]Mechanism[[:space:]]for[[:space:]]Continual[[:space:]]Test-Time[[:space:]]Adaptation/69885bc8-fe2a-47b1-a255-a982ec605708_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Resilience[[:space:]]of[[:space:]]Entropy[[:space:]]Model[[:space:]]in[[:space:]]Distributed[[:space:]]Neural[[:space:]]Networks/ad486483-132b-4d7f-84db-26c1a65f3dae_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Resolving[[:space:]]Scale[[:space:]]Ambiguity[[:space:]]in[[:space:]]Multi-view[[:space:]]3D[[:space:]]Reconstruction[[:space:]]using[[:space:]]Dual-Pixel[[:space:]]Sensors/518914c5-d998-41c3-963e-0840d0e1f2c8_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Responsible[[:space:]]Visual[[:space:]]Editing/a719b1ec-b6ce-4b23-ac8b-598c2b587aef_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Restore[[:space:]]Anything[[:space:]]with[[:space:]]Masks_[[:space:]]Leveraging[[:space:]]Mask[[:space:]]Image[[:space:]]Modeling[[:space:]]for[[:space:]]Blind[[:space:]]All-in-One[[:space:]]Image[[:space:]]Restoration/0de4c62c-e2d6-4f18-8a4e-c05352d6d956_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Restoring[[:space:]]Images[[:space:]]in[[:space:]]Adverse[[:space:]]Weather[[:space:]]Conditions[[:space:]]via[[:space:]]Histogram[[:space:]]Transformer/95e1934a-8ade-4d7d-8947-f969acbf3910_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Retargeting[[:space:]]Visual[[:space:]]Data[[:space:]]with[[:space:]]Deformation[[:space:]]Fields/547a2a3d-69aa-4867-932e-8316bae28b65_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Rethinking[[:space:]]Data[[:space:]]Augmentation[[:space:]]for[[:space:]]Robust[[:space:]]LiDAR[[:space:]]Semantic[[:space:]]Segmentation[[:space:]]in[[:space:]]Adverse[[:space:]]Weather/60de04d8-9845-4ba5-9238-a2dd5b5c4b07_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Rethinking[[:space:]]Data[[:space:]]Bias_[[:space:]]Dataset[[:space:]]Copyright[[:space:]]Protection[[:space:]]via[[:space:]]Embedding[[:space:]]Class-wise[[:space:]]Hidden[[:space:]]Bias/0a3fef5c-d927-4d80-a493-071bf703c4f6_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Rethinking[[:space:]]Deep[[:space:]]Unrolled[[:space:]]Model[[:space:]]for[[:space:]]Accelerated[[:space:]]MRI[[:space:]]Reconstruction/27de6684-978b-4680-80b1-f68f3bc7d6a8_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Rethinking[[:space:]]Directional[[:space:]]Parameterization[[:space:]]in[[:space:]]Neural[[:space:]]Implicit[[:space:]]Surface[[:space:]]Reconstruction/8b296952-4d44-42db-affd-9e9843feb2a6_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Rethinking[[:space:]]Fast[[:space:]]Adversarial[[:space:]]Training_[[:space:]]A[[:space:]]Splitting[[:space:]]Technique[[:space:]]To[[:space:]]Overcome[[:space:]]Catastrophic[[:space:]]Overfitting/54a9d439-23cc-4380-af56-e12fd67a1f40_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Rethinking[[:space:]]Features-Fused-Pyramid-Neck[[:space:]]for[[:space:]]Object[[:space:]]Detection/8cefd2b8-bab0-4ef4-a711-578971c2749a_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Rethinking[[:space:]]Few-shot[[:space:]]Class-incremental[[:space:]]Learning_[[:space:]]Learning[[:space:]]from[[:space:]]Yourself/a34c73cf-0382-420e-bbe9-4f775aa32eab_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Rethinking[[:space:]]Image[[:space:]]Super[[:space:]]Resolution[[:space:]]from[[:space:]]Training[[:space:]]Data[[:space:]]Perspectives/8d3fa075-356f-4f4c-b1a5-4318bed20e1d_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Rethinking[[:space:]]Image-to-Video[[:space:]]Adaptation_[[:space:]]An[[:space:]]Object-centric[[:space:]]Perspective/17973f40-dece-4cfa-8ceb-59145fcea296_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Rethinking[[:space:]]LiDAR[[:space:]]Domain[[:space:]]Generalization_[[:space:]]Single[[:space:]]Source[[:space:]]as[[:space:]]Multiple[[:space:]]Density[[:space:]]Domains/e618c429-3f75-446b-a895-903ce242d321_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Rethinking[[:space:]]Normalization[[:space:]]Layers[[:space:]]for[[:space:]]Domain[[:space:]]Generalizable[[:space:]]Person[[:space:]]Re-identification/02192c3d-6e2f-475e-b7a0-eaf65cc6bd4c_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Rethinking[[:space:]]Tree-Ring[[:space:]]Watermarking[[:space:]]for[[:space:]]Enhanced[[:space:]]Multi-Key[[:space:]]Identification/dd0cc8f0-17b3-4928-9aeb-d79e0a1054db_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Rethinking[[:space:]]Unsupervised[[:space:]]Outlier[[:space:]]Detection[[:space:]]via[[:space:]]Multiple[[:space:]]Thresholding/7bb2035e-e079-470b-b56b-69d7d8631ccd_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Rethinking[[:space:]]Video[[:space:]]Deblurring[[:space:]]with[[:space:]]Wavelet-Aware[[:space:]]Dynamic[[:space:]]Transformer[[:space:]]and[[:space:]]Diffusion[[:space:]]Model/7afdf756-63a6-48fd-b654-3bb54bd5d976_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Rethinking[[:space:]]Video-Text[[:space:]]Understanding_[[:space:]]Retrieval[[:space:]]from[[:space:]]Counterfactually[[:space:]]Augmented[[:space:]]Data/21cc04b7-ea52-4351-82c2-826a88e983b2_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Rethinking[[:space:]]Weakly-supervised[[:space:]]Video[[:space:]]Temporal[[:space:]]Grounding[[:space:]]From[[:space:]]a[[:space:]]Game[[:space:]]Perspective/4f2540b2-441f-49e9-a4f9-06c96b70a305_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Rethinking[[:space:]]and[[:space:]]Improving[[:space:]]Visual[[:space:]]Prompt[[:space:]]Selection[[:space:]]for[[:space:]]In-Context[[:space:]]Learning[[:space:]]Segmentation[[:space:]]Framework/c5a744a9-bd8f-4567-93ec-5cbfe907a7d5_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Retrieval[[:space:]]Robust[[:space:]]to[[:space:]]Object[[:space:]]Motion[[:space:]]Blur/191ce9d4-edfe-414c-a611-8012902a3458_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Revising[[:space:]]Densification[[:space:]]in[[:space:]]Gaussian[[:space:]]Splatting/918310b7-b91b-411d-ade7-18df636ce501_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Revisit[[:space:]]Anything_[[:space:]]Visual[[:space:]]Place[[:space:]]Recognition[[:space:]]via[[:space:]]Image[[:space:]]Segment[[:space:]]Retrieval/32b8503b-ae81-47c5-bdc0-0efb4201113d_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Revisit[[:space:]]Event[[:space:]]Generation[[:space:]]Model_[[:space:]]Self-Supervised[[:space:]]Learning[[:space:]]of[[:space:]]Event-to-Video[[:space:]]Reconstruction[[:space:]]with[[:space:]]Implicit[[:space:]]Neural[[:space:]]Representations/a41e408e-cbb3-4c38-85c3-298d71f3cf9d_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Revisit[[:space:]]Human-Scene[[:space:]]Interaction[[:space:]]via[[:space:]]Space[[:space:]]Occupancy/f5977317-ef8e-4afa-8ef9-88b9ecbe0b04_origin.pdf filter=lfs diff=lfs merge=lfs -text diff --git a/2024/Receler_ Reliable Concept Erasing of Text-to-Image Diffusion Models via Lightweight Erasers/e96e787a-46f5-4d12-8c30-eca8de901ff9_content_list.json b/2024/Receler_ Reliable Concept Erasing of Text-to-Image Diffusion Models via Lightweight Erasers/e96e787a-46f5-4d12-8c30-eca8de901ff9_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..3fbb7ab0ddd0b5ee0491e89aea68407e12b11494 --- /dev/null +++ b/2024/Receler_ Reliable Concept Erasing of Text-to-Image Diffusion Models via Lightweight Erasers/e96e787a-46f5-4d12-8c30-eca8de901ff9_content_list.json @@ -0,0 +1,1602 @@ +[ + { + "type": "text", + "text": "Receler: Reliable Concept Erasing of Text-to-Image Diffusion Models via Lightweight Erasers", + "text_level": 1, + "bbox": [ + 292, + 140, + 715, + 209 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Chi-Pin Huang\\*1,\\*, Kai-Po Chang\\*1, Chung-Ting Tsai2, Yung-Hsuan Lai2, Fu-En Yang3, and Yu-Chiang Frank Wang\\*1,3,\\*", + "bbox": [ + 250, + 234, + 753, + 267 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Graduate Institute of Communication Engineering, National Taiwan University", + "bbox": [ + 228, + 276, + 772, + 290 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ National Taiwan University, $^{3}$ NVIDIA", + "bbox": [ + 362, + 290, + 640, + 305 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "† f11942097@ntu.edu.tw, ‡ frankwang@nvidia.com", + "bbox": [ + 328, + 305, + 673, + 319 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Concept erasure in text-to-image diffusion models aims to disable pre-trained diffusion models from generating images related to a target concept. To perform reliable concept erasure, the properties of robustness and locality are desirable. The former refrains the model from producing images associated with the target concept for any paraphrased or learned prompts, while the latter preserves its ability in generating images with non-target concepts. In this paper, we propose Reliable Concept Erasing via Lightweight Erasers (Receler). It learns a lightweight Eraser to perform concept erasing while satisfying the above desirable properties through the proposed concept-localized regularization and adversarial prompt learning scheme. Experiments with various concepts verify the superiority of Receler over previous methods. Code is available at https://github.com/jasper0314-huang/Receler.", + "bbox": [ + 259, + 358, + 743, + 539 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Concept Erasing $\\cdot$ Diffusion Models $\\cdot$ Adversarial Learning", + "bbox": [ + 261, + 551, + 741, + 566 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 616, + 377, + 633 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In recent years, text-to-image generation models [4, 32, 34, 36, 38] have achieved unprecedented success in generating photo-realistic images which benefit various industrial applications [34, 38]. Despite their apparent success and convenience, these models may produce images that are deemed NSFW (Not Safe for Work) [21] or infringe upon intellectual property and portrait rights [1], e.g., generating nudity, violent content, or imitating the style of well-known artists. This issue is mainly due to the memorization of the large-scale training data sourced from the web [8, 43]. To address the above problem, an intuitive solution is to manually filter out the inappropriate images and re-train the model. However, as pointed out in [16], this may lead to unpredictable results such as exposing more inappropriate content to be memorized [7] or incomplete visual", + "bbox": [ + 212, + 648, + 787, + 816 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal Contribution", + "bbox": [ + 236, + 825, + 380, + 840 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/d57102dd2aa42d591ce0cbe6f0ce18cf90f5f02e07c5a13dea826be849246218.jpg", + "image_caption": [ + "Fig. 1: Illustration of (a) robustness and (b) locality-preserving in concept erasing. The former requires models to be robust against paraphrased attacks from the target concept (i.e., \"nudity\"), while the latter aims to preserve the visual content of non-target concepts (e.g., \"tennis player\" or \"airplane\"). Note that SD denotes Stable Diffusion [36]; UCE [16] and ESD [15] are recent works on concept erasing." + ], + "image_footnote": [], + "bbox": [ + 225, + 143, + 777, + 324 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "concept removal [35]. Moreover, even though filtered image data can be collected, re-training generative models is still computationally expensive [15,23,42].", + "bbox": [ + 212, + 440, + 782, + 470 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Instead of processing training data and re-training models, an alternative is to erase or unlearn specific concepts from the pre-trained model [15,16,23,52]. That is, given a concept described in text, the pre-trained model is fine-tuned to forget that concept so that the associated image content cannot be generated from the fine-tuned model. In practice, it would be desirable to perform concept erasing with sufficient reliability, which suggests two desirable properties: locality and robustness. Locality indicates the ability to preserve the model generalization in synthesizing content not associated with the target concept [31,42]. Robustness requires erased models to effectively remove the target concept [31,42], while not be circumvented by paraphrased prompts that aim to recover the target concept (e.g., \"car\" vs. \"jeep\"). While both locality and robustness have been recently discussed in the field of NLP [10,12,20,29,50], the developed techniques cannot be directly applied to text-to-image generative models for performing concept erasing.", + "bbox": [ + 212, + 474, + 787, + 685 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recently, a number of methods for concept erasing or unlearning for diffusion models have been proposed [15, 16, 23, 52]. For example, Ablating [23] predefines an anchor concept for each target concept that needs to be unlearned and then achieves model unlearning by mapping the image distribution of the target concept to that of the anchor concept, e.g. mapping \"grumpy cat\" to \"cat.\" Inspired by classifier-free guidance [19], ESD [15] fine-tunes the model to predict negatively guided noise. In other words, it decreases the probability of generating images of the target concept, thus unlearning that concept. FMN [52] designs a computationally efficient unlearning method by directly minimizing the cross-attention weights corresponding to the target concept in the input text prompt,", + "bbox": [ + 212, + 689, + 787, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "Huang et al.", + "bbox": [ + 271, + 114, + 354, + 128 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "encouraging the model to ignore the concept. UCE [16] employs a closed-form editing approach to optimize the projection matrices of keys and values in cross-attention layers, ensuring the model maintains locality when unlearning the target concept.", + "bbox": [ + 212, + 146, + 784, + 205 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Although promising progress has been made in erasing specific target concepts, most existing works are not specifically designed to preserve model locality and robustness. For example, Ablating [23] and ESD [15] fine-tune a considerable amount of parameters within pre-trained diffusion models to achieve concept erasure, which inevitably compromises the original capabilities of the model. On the other hand, methods such as UCE [16] and FMN [52] only modify specific parameters (i.e., the projection matrices of keys and values in cross-attention layers) responsible for encoding input textual features instead of visual ones. These methods would be vulnerable to rephrased target concepts since they only learn to dissociate textual prompts in the cross-attention layers (i.e., lack of ability to recognize that paraphrased queries are semantically similar to the erased target concept). For example, a diffusion model that has been erased of the concept of \"car\" may still produce images of jeeps due to its inability to recognize that jeeps fall under the category of cars; hence, it cannot remove the attention to jeep in cross-attention layers. As a result, proposing a concept-erasing method that addresses locality and robustness continues to pose a crucial challenge.", + "bbox": [ + 212, + 207, + 787, + 448 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this paper, we propose Reliable Concept Erasing via Lightweight Erasers (Receler) for erasing concepts from pre-trained diffusion models, exhibiting sufficient locality and robustness properties. Receler involves a lightweight eraser (only $0.37\\%$ of the U-Net parameters), which is designed to remove a target concept from the outputs of cross-attention layers. During this unlearning process, we train the eraser while preserving the image generation capability of the pre-trained diffusion models. Furthermore, a concept-localized regularization is introduced to ensure that the eraser focuses on erasing the target concept. This regularization prevents the generation of non-target concepts from being affected, thereby preserving locality. As for robustness, we advance adversarial prompt learning, which optimizes the adversarial prompts that induce the model to generate images of the target concept and then fine-tunes the eraser to erase images generated with these prompts. By training our eraser with concept-localized regularization and adversarial prompt learning, we are able to preserve the image generation capability of non-target concepts and robustly refrain the model from generating images with target-concept content.", + "bbox": [ + 212, + 449, + 787, + 690 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We now summarize the contributions of this work below:", + "bbox": [ + 238, + 690, + 650, + 704 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We present Reliable Concept Erasing via Lightweight Erasers (Receler), a novel approach using a lightweight eraser (only $0.37\\%$ of the U-Net parameters) for reliable and efficient concept erasing.", + "- Locality is introduced through concept-localized regularization, which constrains the eraser for precise erasing of the target concept without affecting the generation of non-target ones.", + "- Receler is trained against adversarial prompts, imitating paraphrased prompts of target concepts, resulting in improved robustness in concept erasure." + ], + "bbox": [ + 222, + 712, + 790, + 839 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "Receler", + "bbox": [ + 679, + 114, + 730, + 126 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 114, + 785, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Works", + "text_level": 1, + "bbox": [ + 215, + 143, + 397, + 160 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.1 Erasing Concepts from Diffusion Models", + "text_level": 1, + "bbox": [ + 215, + 176, + 596, + 191 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Text-to-image diffusion models [4, 32, 34, 36, 38] have been shown to generate high-quality images with impressive generalization. However, such models are typically trained on extensive web-crawled data (e.g., LAION-5B [41]), which could memorize NSFW or copyrighted content, leading to the generation or replication of undesired images [9, 44]. To address this issue, some works explore the solution without the need to update the model weights. For instance, Stable Diffusion [36] employs an unsafe content classifier to filter out risky outputs, while SLD [39] uses negative guidance to prevent inappropriate content generation. However, the former relies on the reliability of pre-trained classifiers, while the latter only suppresses undesired concepts without complete removal.", + "bbox": [ + 212, + 202, + 787, + 353 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In response, several works focus on fine-tuning the diffusion model to erase the target concepts [15, 16, 23, 52]. Ablating [23] associates each concept to be erased, e.g., \"grumpy cat,\" with a broader, predefined anchor concept, e.g., \"cat\" and fine-tunes the diffusion model to map the generated image of the target concept to that of the anchor concept by minimizing the L2 distance of predicted noises. Inspired by classifier-free guidance [19], ESD [15] proposes to decrease the likelihood of generating images belonging to the target concept. This is achieved by fine-tuning the diffusion model to predict negatively guided noises, effectively steering the model's conditional prediction away from the erased concept. FMN [52] adopts attention resteing, a computationally efficient unlearning method, to identify attention maps associated with the target concept in the diffusion U-Net's cross-attention layers. By minimizing the attention weights corresponding to the target concept, the diffusion model gradually disregards the target concept during image generation, facilitating the erasure of the concept. UCE [16] employs a closed-form editing method to optimize the projection matrices of keys and values in cross-attention. The objective is to align the embedding of a source prompt (e.g., \"a photo of an airplane\") more closely with that of a destination (e.g., an empty string), while leaving other unrelated concepts unchanged. Despite their effectiveness in erasing concepts, most current methods are not particularly designed to preserve robustness against paraphrased prompts (e.g., \"nudity\" vs. \"without clothes\"). Moreover, both [23] and [15] require fine-tuning a considerable number of model parameters, which might affect the model capability and consequently compromise the locality property.", + "bbox": [ + 212, + 354, + 787, + 703 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.2 Controlling Text-to-Image Diffusion Models", + "text_level": 1, + "bbox": [ + 215, + 723, + 624, + 739 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Parameter-Efficient Fine-Tuning (PEFT) is a training scheme that addresses the challenges of extensive parameter updates, especially for large language models. These approaches update only a small subset of parameters, thereby reducing the risk of compromising the pre-trained capabilities of the model. Recent researches [14,24,30,37,51,53,55] have applied PEFT to control text-to-image diffusion models. For instance, some studies [14,24,37] empower the model to learn", + "bbox": [ + 212, + 750, + 787, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 230, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "Huang et al.", + "bbox": [ + 271, + 114, + 357, + 128 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "personalized or unseen concepts by fine-tuning a new text token and a small number of parameters using few user-provided exemplar images. Meanwhile, other works [30, 51, 53, 55] aim to enable diffusion models to generate images based on additional conditions, e.g., edge maps, depth maps, or segmentation masks. They fine-tune lightweight task-specific modules with condition-image pairs to achieve control over the image generation process. Despite the effectiveness of these methods, they achieve learning new concepts by accessing the associated data of interest during training. As for concept erasing, since one only observes the description of the concept to be unlearned, existing PEFT-based methods cannot be directly applied.", + "bbox": [ + 212, + 146, + 787, + 297 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "2.3 Adversarial Attack & Training", + "text_level": 1, + "bbox": [ + 215, + 324, + 514, + 339 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In adversarial attacks [2, 13, 17, 25], adversarial examples are deliberately constructed inputs, which would deceive models into making incorrect predictions. Popular methods such as Fast Gradient Sign Method (FGSM [13]) and its variants (I-FGSM [13] and MI-FGSM [13]) targeted at attacking classification models, utilizing the resulting loss gradients to produce imperceptible perturbations and induce misclassification. In contrast, recent approaches [11, 47, 54] introduce prompt-based adversarial attacks tailored to provoke seemingly unlearned diffusion models into generating images of the unlearned concepts. For instance, P4D [11] learns prompts to reconstruct noise associated with the target concept in diffusion models, and Ring-A-Bell [47] extracts holistic concept representations from CLIP model [33] to generate model-agnostic attack prompts. These learned attack prompts have been shown to provoke unlearned models to regenerate images of the erased concept, posing potential issues in text-to-image generative models.", + "bbox": [ + 212, + 353, + 787, + 564 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To defend against such adversarial attacks, various adversarial training strategies have been proposed [3, 17, 28]. Such adversarial training schemes expose models to adversarial examples during training, enabling the models to learn and recognize these examples. Consequently, these adversially trained models can respond accurately when encountering adversarial examples during inference. Inspired by adversarial training approaches, we employ an adversarial erasing learning scheme to introduce additional robustness to the unlearned model. We will detail our proposed framework in the following section.", + "bbox": [ + 212, + 565, + 787, + 686 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3 Method", + "text_level": 1, + "bbox": [ + 215, + 713, + 330, + 729 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Problem formulation. We first define the setting and notations of our Receler. Given a pre-trained text-to-image diffusion model, parameterized by $\\theta$ , we aim to erase a textual concept $c$ from the model without requiring access to the corresponding image data. The erasure is considered successful when the model no longer generates images that contain or represent the concept $c$ (e.g., \"nudity\"-erased model should not generate any images with exposed body parts).", + "bbox": [ + 212, + 750, + 787, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "Receler", + "bbox": [ + 679, + 114, + 730, + 126 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 114, + 785, + 126 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/75e522671b89ce6f7b06b60d99906beccdb442bb70123b6778b11acf04fe5179.jpg", + "image_caption": [ + "Fig. 2: Overview of Receler. (a) Receler involves iterative learning of a lightweight Eraser $E$ and adversarial prompt embedding $e_{Adv}$ . The former is trained to erase the target concept $c$ while preserving non-target concepts, and the latter learns to imitate the prompts to recover visual content associated with the concept previously erased. (b) The Eraser $E$ is inserted after each cross attention layer of Diffusion U-Net to remove the target concept from its outputs, with prediction $o^l$ directly added to the cross attention output." + ], + "image_footnote": [], + "bbox": [ + 223, + 141, + 338, + 335 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/75e57cece7844947bb7b0a5404923736184dd37d6ea7795eb4a111171f4f5dd7.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 341, + 141, + 620, + 335 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/243f92659970b1ae4d644fca56012aca6b72b6146e9bc3d4b349f35a0eb7f45e.jpg", + "image_caption": [ + "(b)" + ], + "image_footnote": [], + "bbox": [ + 620, + 141, + 781, + 335 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "As depicted in Fig. 2, our method employs a lightweight Eraser $E$ , parameterized by $\\theta_E$ , to learn to erase the target concept $c$ , as discussed in Sec. 3.1. To introduce the desirable locality and robustness to our model, we incorporate concept-localized regularization and adversarial prompt learning schemes into our framework, as detailed in Sec. 3.2 and Sec. 3.3.", + "bbox": [ + 212, + 487, + 787, + 563 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.1 Concept Erasing with Lightweight Eraser", + "text_level": 1, + "bbox": [ + 214, + 585, + 602, + 602 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In order to erase particular visual concepts from a pre-trained diffusion model, we introduce a lightweight adapter-based eraser. As depicted in Fig. 2, by solely finetuning the newly introduced $\\theta_{E}$ , our goal is to unlearn the target visual concept while preserving model generalization on non-target concepts. To be specific, our eraser is designed to remove the target concept from the output visual features of each cross-attention layer within the diffusion U-Net. This is based on the fact that these layers are responsible for incorporating the input text concept into the visual features. Thus, our eraser is positioned subsequent to each of the cross-attention layers, which is in line with the empirical analysis presented in [49]. To erase the concept, the eraser is trained to predict the negatively guided noises [15, 19] that move the model's prediction away from the erased concept. The objective is defined as:", + "bbox": [ + 212, + 613, + 787, + 795 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {E r a s e} = \\mathbb {E} _ {x _ {t}, t} \\left[ \\lVert \\epsilon_ {\\theta^ {\\prime}} (x _ {t}, e _ {c}, t) - \\epsilon_ {E} \\rVert^ {2} \\right],\n$$\n", + "text_format": "latex", + "bbox": [ + 346, + 805, + 614, + 825 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\epsilon_{E} = \\epsilon_{\\theta}(x_{t},t) - \\eta [\\epsilon_{\\theta}(x_{t},e_{c},t) - \\epsilon_{\\theta}(x_{t},t)]$", + "bbox": [ + 326, + 825, + 674, + 842 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "(1)", + "bbox": [ + 638, + 816, + 784, + 832 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "Huang et al.", + "bbox": [ + 271, + 114, + 354, + 128 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Note that $\\theta' = \\{\\theta, \\theta_E\\}$ represents the parameters of the diffusion model plugged with eraser, $x_t \\in \\mathbb{R}^{W \\times H \\times d}$ is the denoised image at timestep $t$ sampled from $\\theta'$ conditioned on $c$ , $e_c$ is the text embedding of concept $c$ , and $\\epsilon_E$ is the negatively guided noises [15, 19] predicted by $\\theta$ . By minimizing the L2 distance between $\\epsilon_{\\theta'}(x_t, e_c, t)$ and $\\epsilon_E$ , the eraser learns to reduce the probability of the generated image $x$ belongs to the target concept $c$ , thus effectively erasing the concept.", + "bbox": [ + 214, + 146, + 787, + 238 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.2 Concept-Localized Regularization for Erasing Locality", + "text_level": 1, + "bbox": [ + 214, + 258, + 707, + 273 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "As the first desirable property in concept erasing, locality refers to preserving the model's ability to synthesize content unrelated to the target, which is realized by enforcing the eraser to affect the image synthesis process only if the target concept is present. To achieve this, we introduce concept-localized regularization into our Receler by leveraging the spatial information associated with the target concept's text tokens to regularize the eraser outputs. Specifically, inspired by [6, 46], we obtain the binary target concept mask $M \\in \\mathbb{R}^{W/4 \\times H/4}$ by thresholding the attention maps when predicting $\\epsilon_{\\theta}(x_t, e_c, t)$ as follows:", + "bbox": [ + 214, + 281, + 787, + 402 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\nM _ {i, j} = \\left\\{ \\begin{array}{l l} 1, & \\text {i f} \\frac {1}{| \\mathcal {S} |} \\sum_ {s \\in \\mathcal {S}} A _ {i, j} ^ {s} \\geq \\tau , \\\\ 0, & \\text {o t h e r w i s e .} \\end{array} \\right. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 383, + 412, + 785, + 467 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $\\mathcal{S}$ is the set of indices of all U-Net's mid-layers with resolution $(\\frac{W}{4},\\frac{H}{4})$ , $A^s\\in \\mathbb{R}^{W / 4\\times H / 4}$ is the cross-attention map in the $s$ -th layer of the text tokens corresponding to the concept $c$ , and $\\tau$ is a pre-defined threshold. With $M$ obtained, we calculate the following regularization loss to regularize the outputs of the eraser as follows:", + "bbox": [ + 214, + 476, + 787, + 554 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {R e g} = \\frac {1}{L} \\sum_ {l = 1} ^ {L} \\| o ^ {l} \\odot (1 - \\tilde {M}) \\| ^ {2}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 393, + 561, + 785, + 601 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $L$ is the number of U-Net's layers, $\\odot$ is the element-wise product, $o^l \\in \\mathbb{R}^{w^l \\times h^l \\times d}$ is the output of the eraser in the $l$ -th layer with the resolution $(w^l, h^l)$ , and $d$ is the feature dimension. Note that $\\tilde{M}$ is the shorthand of $M$ being bicubically upscaled to the same resolution as $o^l$ . With the above regularization introduced, our Receler is enforced to preserve the generation of non-target concepts. Thus, the diversity and fidelity of the original model can be maintained.", + "bbox": [ + 214, + 612, + 787, + 705 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.3 Adversarial Prompt Learning for Erasing Robustness", + "text_level": 1, + "bbox": [ + 214, + 724, + 699, + 741 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To further ensure our Receler being robust against prompting attacks (e.g., by paraphrased or learned prompts), we introduce an adversarial learning strategy into our framework to enrich such model robustness. In our framework, we present a unique adversarial prompt learning scheme, which learns prompting attacks that would induce the diffusion model to synthesize the images containing previously erased concepts. To achieve this, we design the adversarial loss,", + "bbox": [ + 212, + 750, + 787, + 840 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "Receler", + "bbox": [ + 679, + 114, + 730, + 126 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 114, + 785, + 126 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "which optimizes the continuous soft prompts $e_{Adv}$ by encouraging such learned prompts to imitate the malicious prompts, as illustrated in Fig. 2. More precisely, the objective $\\mathcal{L}_{Adv}$ is defined as follows:", + "bbox": [ + 212, + 146, + 787, + 191 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {A d v} = \\mathbb {E} _ {x _ {t}, t} \\left[ \\| \\epsilon_ {\\theta^ {\\prime}} (x _ {t}, e ^ {\\prime}, t) - \\epsilon_ {M} \\| ^ {2} \\right], \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 372, + 205, + 785, + 224 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $e' = [e_c; e_{Adv}]$ is the concatenated prompts of the erased target concept embedding and the learned soft prompts, and $\\epsilon_M = \\epsilon_\\theta(x_t, e_c, t)$ represents the malicious noise predicted by the pre-trained diffusion model (without safety mechanism) conditioned on the target concept. By minimizing this adversarial loss, $e_{Adv}$ learns to regenerate the unlearned concept from the erased model. The optimization of the soft prompt $e_{Adv}$ and the eraser $\\theta_E$ is performed iteratively, with each being fixed while the other is trained. Thus, $\\theta_E$ and $e_{Adv}$ are trained against each other to improve model robustness. For more details, please refer to the pseudo-algorithm in the supplementary material.", + "bbox": [ + 212, + 236, + 789, + 373 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 215, + 397, + 375, + 416 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this section, we first conduct quantitative experiments to assess the robustness and locality of Receler compared to state-of-the-art baselines, followed by ablation studies of our method. Lastly, we present qualitative comparisons and visualizations to demonstrate its effectiveness.", + "bbox": [ + 212, + 431, + 787, + 491 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Datasets. We conduct experiments on erasing objects defined in the CIFAR-10 dataset [22] and on erasing inappropriate contents from the Inappropriate Image Prompts (I2P) dataset [39]:", + "bbox": [ + 212, + 515, + 787, + 561 + ], + "page_idx": 7 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Object Erasure. To evaluate the effectiveness of erasure methods in erasing common visual concepts, we choose to erase ten class labels from CIFAR-10 [22]. Note that during our experiment, we only utilize the label set, not the images. For comprehensive assessment, we devise two types of evaluation prompts for each class: Firstly, we use simple prompts formatted as \"A photo of {class}\" to evaluate the efficacy of erasure in removing the target concept. Secondly, to further assess the robustness, we generate 50 paraphrased prompts for each class using ChatGPT3 to simulate real-world scenarios where prompts are typically more complex and target concepts may not be explicitly mentioned. For example, a paraphrased prompt for \"airplane\" is \"A sleek, black stealth bomber flying low over a desert landscape at dusk.\" More details can be found in the supplementary material.", + "- Inappropriate Content Erasure. The I2P dataset [39], collected from a text-to-image prompt dataset DiffusionDB [48], comprises 4,703 real-world, user-generated prompts that produce inappropriate images, including hate, harassment, violence, self-harm, shocking, sexual, and illegal content." + ], + "bbox": [ + 222, + 571, + 787, + 818 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "Huang et al.", + "bbox": [ + 271, + 114, + 354, + 128 + ], + "page_idx": 7 + }, + { + "type": "page_footnote", + "text": "3https://chat.openai.com/", + "bbox": [ + 236, + 824, + 418, + 840 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Evaluation Setup. We assess the robustness and locality of the erasure methods for object erasure and inappropriate content erasure as follows:", + "bbox": [ + 215, + 146, + 782, + 176 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "- Object Erasure. For each method, we fine-tune ten models, each erasing one CIFAR-10 class. Each model is then evaluated by: 1) Efficacy $(\\mathrm{Acc}_E)$ : the percentage of the target class being erased when inputting simple prompts; lower values are better. 2) Robustness $(\\mathrm{Acc}_R)$ : the percentage of the target class being erased when inputting paraphrased prompts; lower values are better. 3) Locality $(\\mathrm{Acc}_L)$ : the percentage of non-target classes being preserved; higher values are better. To assess efficacy and robustness, we generate 150 images using simple and paraphrased prompts for the target class, respectively. For Locality, we generate 50 images for each of the nine non-target classes using paraphrased prompts. We then use GroundingDINO [27] to detect if the corresponding class is presenting in the image, thereby assessing $\\mathrm{Acc}_E$ , $\\mathrm{Acc}_R$ , and $\\mathrm{Acc}_L$ . To further evaluate the overall performance, we calculate the harmonic mean (H) of $100 - \\mathrm{Acc}_E$ , $100 - \\mathrm{Acc}_R$ , and $\\mathrm{Acc}_L$ .", + "bbox": [ + 222, + 185, + 787, + 381 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "- Inappropriate Content Erasure. Following ESD [15], we fine-tune two models for each erasure method: one for \"nudity\" and the other for the predefined inappropriate concepts e.g. hate, harassment, and violence. We use the NudeNet detector [5] to detect nudity and both the NudeNet and the Q16 detector [40] to identify the inappropriate concepts. Model robustness is evaluated by using real-world prompts in I2P [39]. Locality is evaluated using COCO-30K [26], a nudity-free dataset, by employing the nudity-erased model to generate safe contents from COCO-30K prompts and evaluating the quality of the generated images in terms of FID [18] and CLIP [33].", + "bbox": [ + 222, + 381, + 787, + 517 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Comparisons. We compare Receler to state-of-the-art erasing methods, including FMN [52], SLD [39], Ablating [23], ESD [15], and UCE [16]. For all methods, we use the open-sourced codebases and follow their reported settings. We fine-tune all models from SD v1.4 [36], and for all image generation, we employ DDIM sampler [45] over 50 steps and a guidance scale of 7.5. Following SD v1.4, the image resolution in all our experiments is $512 \\times 512$ . Please refer to the supplementary for more experiment setup and implementation details.", + "bbox": [ + 215, + 537, + 787, + 643 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.1 Quantitative Evaluation", + "text_level": 1, + "bbox": [ + 215, + 665, + 459, + 680 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Object Erasure. In Tab. 1, we show that Receler surpasses previous state-of-the-art methods in erasing common visual concepts from CIFAR-10 [22] class labels. Notably, Receler achieves the highest harmonic mean $(H)$ and exceeds the second-best method by 11.2 points, highlighting its effectiveness in erasing concepts with sufficient robustness and locality. Specifically, when assessing method efficacy using the simple prompts $(\\mathrm{Acc}_E)$ , Receler achieves an average accuracy of $14.9\\%$ across the erased classes, $2.2\\%$ better than the second-best method, ESD [15]. In addition, when evaluating method robustness with paraphrased prompts $(\\mathrm{Acc}_R)$ , Receler reaches $17.6\\%$ on average, outperforming ESD by $22.3\\%$ in the erased classes. We evaluate the locality of the erased model by", + "bbox": [ + 215, + 688, + 787, + 840 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "Receler", + "bbox": [ + 679, + 114, + 730, + 126 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/6a56c92a563c8407a2dbab671512b9086b026234652264dfc393f11e89a900da.jpg", + "table_caption": [ + "Table 1: Evaluation of erasing common objects in CIFAR-10 classes. $\\mathrm{Acc}_E$ and $\\mathrm{Acc}_R$ represent efficacy and robustness, indicating accuracy of target class (which should be minimized) on simple and paraphrased prompts, respectively. Locality, $\\mathrm{Acc}_L$ , is accuracy of non-target classes (which should be maximized) using paraphrased prompts. The harmonic mean $H$ reflects overall assessment of $\\mathrm{Acc}_E$ , $\\mathrm{Acc}_R$ , and $\\mathrm{Acc}_L$ ." + ], + "table_footnote": [], + "table_body": "
MethodsMetricsErased conceptsavg.
air-planeauto-mobilebirdcatdeerdogfroghorseshiptruck
SD v1.4AccE89.399.393.396.799.398.796.097.395.396.096.1
AccR79.394.096.088.098.792.088.792.765.384.087.9
AccL88.887.287.087.986.787.487.887.390.488.387.9
FMN [52]AccE ↓93.397.390.092.098.095.384.795.388.794.792.9
AccR ↓80.796.793.370.795.386.767.395.360.784.083.1
AccL ↑88.088.286.087.684.488.086.485.890.488.987.4
H↑14.14.411.517.64.110.027.96.924.011.414.2
Ablating [23]AccE ↓78.074.776.793.392.797.394.7100.090.786.788.5
AccR ↓67.390.093.372.795.387.371.390.758.076.080.2
AccL ↑87.883.884.987.384.085.886.085.388.286.486.0
H↑34.319.814.715.28.36.512.80.021.023.420.1
ESD [15]AccE ↓20.044.011.314.019.320.013.38.716.04.717.1
AccR ↓33.381.354.018.040.727.338.741.332.032.739.9
AccL ↑83.679.872.971.868.070.079.368.286.779.175.9
H↑76.035.864.279.568.274.074.270.378.679.071.6
UCE [16]AccE ↓34.746.08.716.74.011.311.36.022.010.017.1
AccR ↓58.079.363.316.015.349.328.034.041.339.342.4
AccL ↑84.979.181.882.078.082.283.375.887.381.681.6
H↑58.937.859.583.185.669.580.777.072.675.372.0
Receler (Ours)AccE ↓10.046.73.311.32.76.723.37.324.014.014.9
AccR ↓16.762.026.70.72.04.717.36.020.719.317.6
AccL ↑88.481.382.280.476.774.783.880.488.284.782.1
H↑87.152.383.088.889.586.780.988.680.883.783.2
", + "bbox": [ + 225, + 227, + 776, + 570 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "examining its accuracy in the remaining classes $(\\mathrm{Acc}_L)$ aside from the erased class, focusing on whether erasing one CIFAR-10 class affects the image synthesis of the other unrelated classes. Although FMN [52] and Ablating [23] exhibit high average $\\mathrm{Acc}_L$ , appearing effective in preserving model locality, they struggle to erase the target objects, with only a $3.2\\%$ and $7.6\\%$ drop in $\\mathrm{Acc}_E$ from SD, compared to our $81.2\\%$ drop.", + "bbox": [ + 212, + 603, + 787, + 694 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Erasure of Inappropriate Content. In Tab. 2 and Tab. 3, we evaluate the robustness of erasing inappropriate content with real-world prompts from I2P dataset [39]. Compared to the second-best result, Receler stands out by achieving $4.3\\%$ overall improvement in erasing sensitive concepts on I2P and a $3.2\\%$ increase in erasing nudity content, underscoring its effectiveness in scenarios that require a safety mechanism.", + "bbox": [ + 212, + 718, + 787, + 808 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "In Tab. 4, in addition to robustness, we assess locality using COCO-30K [26], a nudity-free dataset. We employ the \"nudity\"-erased model to produce safe", + "bbox": [ + 214, + 809, + 785, + 839 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "Huang et al.", + "bbox": [ + 271, + 114, + 354, + 128 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/eab7af37e90c2144c0d485e6dfe197ea5a65a508b101cb31b27e3fd5039a395c.jpg", + "table_caption": [ + "Table 2: Quantitative results on Inappropriate Image Prompts (I2P) dataset. We follow SLD [39] and apply the ratio of inappropriate images as the metric. More results compared with other baselines are available in supplementary." + ], + "table_footnote": [], + "table_body": "
Class nameInappropriate proportion (%) (↓)
SDFMNSLDESDUCEReceler
Hate44.237.722.526.836.428.6
Harassment37.525.022.124.029.521.7
Violence46.347.831.835.134.127.1
Self-harm47.946.830.033.730.824.8
Sexual60.259.152.435.025.529.4
Shocking59.558.140.540.141.134.8
Illegal activity40.037.022.126.729.021.3
Overall48.947.833.732.831.327.0
", + "bbox": [ + 225, + 239, + 498, + 367 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/f17d5180a696d5ee883f82a5f269dd3515959f01f47f99ff8c3ca53d816e057e.jpg", + "table_caption": [ + "Table 4: Assessment of reliability of nudity-erased models. Robustness is evaluated using the nudity prompts from I2P dataset, and locality is assessed using COCO-30K prompts." + ], + "table_footnote": [], + "table_body": "
MethodRobustness Nudity-erased ratio(↑)Locality
CLIP-30K(↑)FID-30K(↓)
SD-31.3214.27
FMN44.2%30.3913.52
SLD71.6%30.9016.34
ESD81.3%30.2415.31
UCE75.9%30.8514.07
Receler84.5%31.0214.10
", + "bbox": [ + 225, + 463, + 493, + 571 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/90b3a8614e84a89000e609cb46c36c968cb2c1a6f3c32736206820e06a8bb24e.jpg", + "table_caption": [ + "Table 3: Quantitative results on nudity prompts from I2P dataset. We report the number of nudity images detected by the NudeNet [5]. F- and M- refer to female and male, respectively." + ], + "table_footnote": [], + "table_body": "
Class nameNumber of nudity detected (↓)
SDFMNSLDESDUCEReceler
Armpits1484246312939
Belly17011670206026
F-Breast26615539323513
M-Breast421730151212
Buttocks29123975
Feet635619242910
F-Genitalia18151151
M-Genitalia723749
Total743415211139179115
Erasing ratio%--44.2-71.6-81.3-75.9-84.5
", + "bbox": [ + 521, + 228, + 777, + 364 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/20a8e8b65861368a967262b66abea1fca2d28948b42be53c8559ca46754515c8.jpg", + "table_caption": [ + "Table 5: Evaluation of robustness against learned attack prompts. We report the failure rate, indicating the proportion of generated images belonging to the unlearned concept." + ], + "table_footnote": [], + "table_body": "
MethodP4D [11]Ring-A-Bell [47]
cifar10 avg.nudityviolencenudity
FMN88.3%89.4%98.8%94.7%
Ablating85.4%82.8%100.0%96.8%
SLD60.5%56.3%80.4%86.3%
ESD48.1%54.3%86.0%55.8%
UCE53.8%51.9%76.8%49.5%
Receler13.7%31.2%59.2%1.1%
", + "bbox": [ + 521, + 468, + 777, + 566 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "content using COCO-30K prompts and evaluate the quality of the generated images using FID [18] and CLIP [33] metrics. As shown in the last two columns, Receler secures the highest CLIP-30K and nearly matches the top result on FID-30K. It is noteworthy that while Receler performs comparably to FMN in FID-30K, it surpasses FMN in robustness by erasing $40.3\\%$ more nudity content.", + "bbox": [ + 212, + 599, + 787, + 678 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Learned Attack Prompts. To demonstrate the robustness of Receler in safeguarding against the potential and unprecedented malicious attacks, we employ P4D [11] and Ring-A-Bell [47]. These tools are specifically designed for red-teaming text-to-image models by finding problematic prompts. As illustrated in Tab. 5, Receler is significantly more reliable than other concept erasing methods. Specifically, it achieves lower failure rates $-34.4\\%$ for CIFAR-10 and $20.7\\%$ for nudity against P4D prompts, and $17.6\\%$ for violence and $48.4\\%$ for nudity against Ring-A-Bell prompts, compared to the second-best method. These results highlight the robustness of Receler in defending against malicious attacks.", + "bbox": [ + 212, + 703, + 787, + 840 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "Receler", + "bbox": [ + 679, + 114, + 730, + 126 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 784, + 126 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/567d5cece8a2374a36a1fb5f19c59e553473057643b65bd0c0e97e879b49402d.jpg", + "image_caption": [ + "Fig. 3: Qualitative comparison of concept erasure methods. Note that erased concepts are listed at the top, and images generated from each method are shown in each row. Input prompts used for image generation are provided in supplementary." + ], + "image_footnote": [], + "bbox": [ + 274, + 152, + 730, + 339 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Ablation Study. In Tab. 6, we ablate the three proposed components in Receler: the lightweight eraser, concept-localized regularization, and adversarial prompt learning. We establish a simple baseline by fine-tuning the whole model with Eq. (1) (i.e., the first row). With the eraser introduced, both $\\mathrm{Acc}_R$ and $\\mathrm{Acc}_L$ improve. Adding concept-localized regularization further increases the $\\mathrm{Acc}_L$ from $77.4\\%$ to $79.8\\%$ , demonstrating its effectiveness in enhancing locality. On the other hand, coupling adversarial prompting learning with the eraser boosts $\\mathrm{Acc}_R$ by $14\\%$ , albeit with a slight decrease in $\\mathrm{Acc}_L$ . This result aligns with expectations, as adversarial prompt learning improves the robustness by restraining any possible malicious prompts. Therefore, Receler, which integrates both adversarial prompt learning and concept-localized regularization, yields the best experimental results. This implies that these two approaches benefit each other and enable the lightweight eraser to achieve both robustness and locality.", + "bbox": [ + 212, + 419, + 787, + 617 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "4.2 Qualitative Evaluation", + "text_level": 1, + "bbox": [ + 215, + 636, + 447, + 651 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Visualization of Erased Concepts with Paraphrased Prompts. In Fig. 3, we show examples of erasing different artist styles (e.g., van Gogh), objects (e.g., automobile) and high-level concepts (e.g., nudity). As observed in the figure, ESD [15] and UCE [16] fail to erase the target concept, with the outputs from these methods closely resembling the original images from SD [1]. On the contrary, Receler successfully erases all target concepts and is able to generate images that are visually close to the original ones, e.g., same background with the car removed, a human in the same posture with nudity removed.", + "bbox": [ + 212, + 657, + 787, + 777 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "In Fig. 4, we qualitatively validate the robustness and locality of Receler. The diagonal orange boxes show its robustness against paraphrased prompts where the erased concept is not explicitly mentioned. For instance, airplane-erased Receler successfully prevents the generation of an \"airplane\" image when", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "Huang et al.", + "bbox": [ + 271, + 114, + 354, + 128 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/9e7ccc830afb3aca565f969c79f1383d247b1eb19d85133a02f242e754a93124.jpg", + "image_caption": [ + "Fig. 4: Visualization of robustness and locality from Receler on CIFAR-10." + ], + "image_footnote": [], + "bbox": [ + 258, + 155, + 748, + 445 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "The red strikethrough at the top indicates the erased concepts. On the left, the input paraphrased prompts are provided. Images enclosed within the diagonal orange borders shows robustness while others show the locality.", + "bbox": [ + 212, + 469, + 785, + 510 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/406d958736678147da0a6e2d2fdbe48445ada524f875ea931ce8133084ccb08b.jpg", + "image_caption": [ + "(a) nudity" + ], + "image_footnote": [], + "bbox": [ + 243, + 532, + 495, + 643 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/dc9a7319f4a7dadc2a2f807de0af117bdb1ed0591aa98df7796317c05fcb9437.jpg", + "image_caption": [ + "(b) violence", + "Fig. 5: Visualization of erasure methods against learned attack prompts. We use Ring-A-Bell [47] to generate adversarial prompts for nudity and violence concepts." + ], + "image_footnote": [], + "bbox": [ + 503, + 532, + 756, + 645 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "the prompt is paraphrased to \"jet\". Concerning locality, as shown in the non-diagonal boxes, Receler generates images that not only faithfully adhere to the descriptions but also closely resemble the images from the original diffusion model. Notably, it can be seen that Receler is able to replace the unlearned objects with reasonable alternatives rather than simply removing them, e.g., two white puppies are substituted with two people wearing white furry clothes and gloves (see the fourth row in Fig. 4).", + "bbox": [ + 212, + 734, + 787, + 840 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "Receler", + "bbox": [ + 679, + 114, + 730, + 126 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/99ad01acf7e02f44ca8dc78587435f0e06a5aa3907ec144960b79271a4c48e9e.jpg", + "table_caption": [ + "Table 6: Ablation study on CIFAR-10. We ablate the components of Receler and report the robustness and locality metrics. The first row refers to fine-tuning all parameters with only $\\mathcal{L}_{\\text{Erase}}$ in Eq. (1)." + ], + "table_footnote": [], + "table_body": "
ComponentsMetrics
Eraser\\( \\mathcal{L}_{\\text{Reg}} \\)\\( \\mathcal{L}_{\\text{Adv}} \\)\\( \\text{Acc}_R(\\downarrow) \\)\\( \\text{Acc}_L(\\uparrow) \\)
XXX39.475.2
✓XX34.977.4
✓✓X26.379.8
✓X✓20.976.2
✓✓✓17.682.1
", + "bbox": [ + 233, + 244, + 464, + 335 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/4a8a0b671d733ca8446d6c0e3ccc63968bb8ba7e63e6715a3ee7b625abeb5702.jpg", + "image_caption": [ + "Fig.6: Examples of erasing multiple concepts. Instead of training an eraser for multiple concepts from scratch, we combine existing erasers for multi-concept erasure." + ], + "image_footnote": [], + "bbox": [ + 500, + 143, + 593, + 268 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/66a508933f86f395f8fbaa82c3981e47b4928dd161b9ac48c7dfb776c92b8cc6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 596, + 143, + 687, + 268 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/132671481899a7f4e16153ac210d8d2be4ae2e0079e1fab1dae41192398d50d5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 691, + 145, + 782, + 268 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Visualization of Robustness against Learned Attack Prompts. In addition to the quantitative results shown in Tab. 5, we further qualitatively demonstrate the robustness of Receler against learned attack prompts in Fig. 5. The attack prompts are learned to induce the concepts of \"nudity\" and \"violence\" from models that should have erased these concepts using Ring-A-Bell [47]. For both nudity and violence attack prompts, Receler successfully prevents the generation of erased concepts, whereas other methods like ESD and UCE fail and generate images with the supposedly forbidden concepts (e.g., nudity or blood).", + "bbox": [ + 212, + 367, + 787, + 491 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Compositional Concept Erasure. In Fig. 6, we showcase examples of Receler in performing compositional concept erasure. This is achieved by combining outputs from separately trained erasers, each targeting a specific unlearned concept, during inference. By averaging these outputs, compositional concept erasure is accomplished without necessitating retraining. This approach notably offers the flexibility to selectively determine which concepts are to be erased, as required.", + "bbox": [ + 212, + 511, + 787, + 604 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 626, + 359, + 642 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In this paper, we proposed Reliable Concept Erasing via Lightweight Erasers (Receler) to erase target concepts entirely from the pre-trained diffusion model against prompting attacks (i.e., robustness), while preserving its image generation ability of other concepts (i.e., locality). In Receler, we employ concept-localized regularization to enforce the eraser to only affect the visual features related to the target concept. To enhance model robustness to paraphrased or attack prompts, we present an adversarial prompt learning scheme to induce the model to produce images of previously erased concepts, followed by optimizing the model against such image generation. We conducted extensive quantitative and qualitative evaluations on Receler, validating its superior robustness and locality-preserving ability over previous concept-erasing methods.", + "bbox": [ + 212, + 657, + 787, + 825 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "Huang et al.", + "bbox": [ + 271, + 114, + 354, + 128 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgements. This research was supported in part by the National Science and Technology Council via grant NSTC 112-2634-F-002-007, NSTC 113-2640-E-002-003 and the Featured Area Research Center Program within the framework of the Higher Education Sprout Project by the Ministry of Education 113L900902. We also thank the National Center for High-performance Computing (NCHC) for providing computational and storage resources.", + "bbox": [ + 212, + 146, + 787, + 238 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 215, + 261, + 323, + 277 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Sarah andersen. et al v. stability ai ltd. et al. case no.3:2023cv00201. us district court for the northern district of california. (2023)", + "2. Akhtar, N., Mian, A., Kardan, N., Shah, M.: Advances in adversarial attacks and defenses in computer vision: A survey. IEEE Access 9, 155161-155196 (2021)", + "3. Bai, T., Luo, J., Zhao, J., Wen, B., Wang, Q.: Recent advances in adversarial training for adversarial robustness. arXiv preprint arXiv:2102.01356 (2021)", + "4. Balaji, Y., Nah, S., Huang, X., Vahdat, A., Song, J., Kreis, K., Aittala, M., Aila, T., Laine, S., Catanzaro, B., et al.: edifi: Text-to-image diffusion models with an ensemble of expert denoisers. arXiv preprint arXiv:2211.01324 (2022)", + "5. Bedapudi Praneeth, b.k., lireza Ayinmehr: Nudenet: Neural nets for nudity classification, detection and selective censoring (2019)", + "6. Cao, M., Wang, X., Qi, Z., Shan, Y., Qie, X., Zheng, Y.: Masactrl: Tuning-free mutual self-attention control for consistent image synthesis and editing. arXiv preprint arXiv:2304.08465 (2023)", + "7. Carlini, N., Jagielski, M., Zhang, C., Papernot, N., Terzis, A., Tramer, F.: The privacy onion effect: Memorization is relative. In: NeurIPS (2022)", + "8. Carlini, N., Hayes, J., Nasr, M., Jagielski, M., Sehwag, V., Tramer, F., Balle, B., Ippolito, D., Wallace, E.: Extracting training data from diffusion models. In: 32nd USENIX Security Symposium (USENIX Security 23). pp. 5253-5270 (2023)", + "9. Carlini, N., Hayes, J., Nasr, M., Jagielski, M., Sehwag, V., Tramer, F., Balle, B., Ippolito, D., Wallace, E.: Extracting training data from diffusion models. In: 32nd USENIX Security Symposium (USENIX Security 23). pp. 5253-5270 (2023)", + "0. Chen, J., Yang, D.: Unlearn what you want to forget: Efficient unlearning for LLMs. In: Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing. Association for Computational Linguistics (Dec 2023)", + "1. Chin, Z.Y., Jiang, C.M., Huang, C.C., Chen, P.Y., Chiu, W.C.: Prompting4debugging: Red-teaming text-to-image diffusion models by finding problematic prompts. arXiv preprint arXiv:2309.06135 (2023)", + "2. De Cao, N., Aziz, W., Titov, I.: Editing factual knowledge in language models. arXiv preprint arXiv:2104.08164 (2021)", + "3. Dong, Y., Liao, F., Pang, T., Su, H., Zhu, J., Hu, X., Li, J.: Boosting adversarial attacks with momentum. In: CVPR (2018)", + "4. Gal, R., Alaluf, Y., Atzmon, Y., Patashnik, O., Bermano, A.H., Chechik, G., Cohen-Or, D.: An image is worth one word: Personalizing text-to-image generation using textual inversion. arXiv preprint arXiv:2208.01618 (2022)", + "5. Gandikota, R., Materzynska, J., Fiotto-Kaufman, J., Bau, D.: Erasing concepts from diffusion models. arXiv preprint arXiv:2303.07345 (2023)", + "6. Gandikota, R., Orgad, H., Belinkov, Y., Materzyńska, J., Bau, D.: Unified concept editing in diffusion models. arXiv preprint arXiv:2308.14761 (2023)" + ], + "bbox": [ + 225, + 294, + 784, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "Receler", + "bbox": [ + 679, + 114, + 730, + 126 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "17. Goodfellow, I.J., Shlens, J., Szegedy, C.: Explaining and harnessing adversarial examples. arXiv preprint arXiv:1412.6572 (2014)", + "18. Heusel, M., Ramsauer, H., Unterthiner, T., Nessler, B., Hochreiter, S.: Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems 30 (2017)", + "19. Ho, J., Salimans, T.: Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598 (2022)", + "20. Huang, Z., Shen, Y., Zhang, X., Zhou, J., Rong, W., Xiong, Z.: Transformer-patcher: One mistake worth one neuron. arXiv preprint arXiv:2301.09785 (2023)", + "21. Hunter, T.: Ai porn is easy to make now. for women, that's a nightmare. The Washington Post pp. NA-NA (2023)", + "22. Krizhevsky, A., Hinton, G.: Learning multiple layers of features from tiny images. Master's thesis, Department of Computer Science, University of Toronto (2009)", + "23. Kumari, N., Zhang, B., Wang, S.Y., Shechtman, E., Zhang, R., Zhu, J.Y.: Ablating concepts in text-to-image diffusion models. In: ICCV (2023)", + "24. Kumari, N., Zhang, B., Zhang, R., Shechtman, E., Zhu, J.Y.: Multi-concept customization of text-to-image diffusion. In: CVPR (2023)", + "25. Kurakin, A., Goodfellow, I.J., Bengio, S.: Adversarial examples in the physical world. In: Artificial intelligence safety and security, pp. 99-112. Chapman and Hall/CRC (2018)", + "26. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollar, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13. pp. 740-755. Springer (2014)", + "27. Liu, S., Zeng, Z., Ren, T., Li, F., Zhang, H., Yang, J., Li, C., Yang, J., Su, H., Zhu, J., et al.: Grounding dino: Marrying dino with grounded pre-training for open-set object detection. arXiv preprint arXiv:2303.05499 (2023)", + "28. Madry, A., Makelov, A., Schmidt, L., Tsipras, D., Vladu, A.: Towards deep learning models resistant to adversarial attacks. arXiv preprint arXiv:1706.06083 (2017)", + "29. Meng, K., Bau, D., Andonian, A., Belinkov, Y.: Locating and editing factual associations in gpt. In: NeurIPS (2022)", + "30. Mou, C., Wang, X., Xie, L., Zhang, J., Qi, Z., Shan, Y., Qie, X.: T2i-adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models. arXiv preprint arXiv:2302.08453 (2023)", + "31. Ni, M., Wu, C., Wang, X., Yin, S., Wang, L., Liu, Z., Duan, N.: Orens: Open-vocabulary responsible visual synthesis. arXiv preprint arXiv:2308.13785 (2023)", + "32. Nichol, A., Dhariwal, P., Ramesh, A., Shyam, P., Mishkin, P., McGrew, B., Sutskever, I., Chen, M.: Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741 (2021)", + "33. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021)", + "34. Ramesh, A., Dhariwal, P., Nichol, A., Chu, C., Chen, M.: Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125 1(2), 3 (2022)", + "35. Rombach, R.: Stable diffusion 2.0 release (Nov 2022)", + "36. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models (2021)" + ], + "bbox": [ + 215, + 147, + 784, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "Huang et al.", + "bbox": [ + 271, + 114, + 354, + 128 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "37. Ruiz, N., Li, Y., Jampani, V., Pritch, Y., Rubinstein, M., Aberman, K.: Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In: CVPR (2023)", + "38. Sahara, C., Chan, W., Saxena, S., Li, L., Whang, J., Denton, E.L., Ghasemipour, K., Gontijo Lopes, R., Karagol Ayan, B., Salimans, T., et al.: Photorealistic text-to-image diffusion models with deep language understanding. In: NeurIPS (2022)", + "39. Schramowski, P., Brack, M., Deiseroth, B., Kersting, K.: Safe latent diffusion: Mitigating inappropriate degeneration in diffusion models. In: CVPR (2023)", + "40. Schramowski, P., Tauchmann, C., Kersting, K.: Can machines help us answering question 16 in datasheets, and in turn reflecting on inappropriate content? In: Proceedings of the 2022 ACM Conference on Fairness, Accountability, and Transparency. pp. 1350-1361 (2022)", + "41. Schuhmann, C., Beaumont, R., Vencu, R., Gordon, C., Wightman, R., Cherti, M., Coombes, T., Katta, A., Mullis, C., Wortsman, M., et al.: Laion-5b: An open large-scale dataset for training next generation image-text models. In: NeurIPS (2022)", + "42. Sinitsin, A., Plokhotnyuk, V., Pyrkin, D., Popov, S., Babenko, A.: Editable neural networks. arXiv preprint arXiv:2004.00345 (2020)", + "43. Somepalli, G., Singla, V., Goldblum, M., Geiping, J., Goldstein, T.: Diffusion art or digital forgery? investigating data replication in diffusion models. In: CVPR (2023)", + "44. Somepalli, G., Singla, V., Goldblum, M., Geiping, J., Goldstein, T.: Diffusion art or digital forgery? investigating data replication in diffusion models. In: CVPR (2023)", + "45. Song, J., Meng, C., Ermon, S.: Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502 (2020)", + "46. Tang, R., Liu, L., Pandey, A., Jiang, Z., Yang, G., Kumar, K., Stenetorp, P., Lin, J., Ture, F.: What the daam: Interpreting stable diffusion using cross attention. arXiv preprint arXiv:2210.04885 (2022)", + "47. Tsai, Y.L., Hsu, C.Y., Xie, C., Lin, C.H., Chen, J.Y., Li, B., Chen, P.Y., Yu, C.M., Huang, C.Y.: Ring-a-bell! how reliable are concept removal methods for diffusion models? arXiv preprint arXiv:2310.10012 (2023)", + "48. Wang, Z.J., Montoya, E., Munechika, D., Yang, H., Hoover, B., Chau, D.H.: Diffusiondb: A large-scale prompt gallery dataset for text-to-image generative models. arXiv preprint arXiv:2210.14896 (2022)", + "49. Xiang, C., Bao, F., Li, C., Su, H., Zhu, J.: A closer look at parameter-efficient tuning in diffusion models. arXiv preprint arXiv:2303.18181 (2023)", + "50. Yao, Y., Wang, P., Tian, B., Cheng, S., Li, Z., Deng, S., Chen, H., Zhang, N.: Editing large language models: Problems, methods, and opportunities. arXiv preprint arXiv:2305.13172 (2023)", + "51. Ye, H., Zhang, J., Liu, S., Han, X., Yang, W.: Ip-adapter: Text compatible image prompt adapter for text-to-image diffusion models. arXiv preprint arXiv:2308.06721 (2023)", + "52. Zhang, E., Wang, K., Xu, X., Wang, Z., Shi, H.: Forget-me-not: Learning to forget in text-to-image diffusion models. arXiv preprint arXiv:2303.17591 (2023)", + "53. Zhang, L., Rao, A., Agrawala, M.: Adding conditional control to text-to-image diffusion models. In: ICCV (2023)", + "54. Zhang, Y., Jia, J., Chen, X., Chen, A., Zhang, Y., Liu, J., Ding, K., Liu, S.: To generate or not? safety-driven unlearned diffusion models are still easy to generate unsafe images... for now. arXiv preprint arXiv:2310.11868 (2023)" + ], + "bbox": [ + 212, + 146, + 787, + 840 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "Receler", + "bbox": [ + 679, + 114, + 730, + 126 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + }, + { + "type": "ref_text", + "text": "55. Zhao, S., Chen, D., Chen, Y.C., Bao, J., Hao, S., Yuan, L., Wong, K.Y.K.: Unictrlnet: All-in-one control to text-to-image diffusion models. arXiv preprint arXiv:2305.16322 (2023)", + "bbox": [ + 215, + 146, + 785, + 189 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "Huang et al.", + "bbox": [ + 271, + 114, + 354, + 128 + ], + "page_idx": 17 + } +] \ No newline at end of file diff --git a/2024/Receler_ Reliable Concept Erasing of Text-to-Image Diffusion Models via Lightweight Erasers/e96e787a-46f5-4d12-8c30-eca8de901ff9_model.json b/2024/Receler_ Reliable Concept Erasing of Text-to-Image Diffusion Models via Lightweight Erasers/e96e787a-46f5-4d12-8c30-eca8de901ff9_model.json new file mode 100644 index 0000000000000000000000000000000000000000..134a88a3a762b18faa5568ecab24e00d4559d949 --- /dev/null +++ b/2024/Receler_ Reliable Concept Erasing of Text-to-Image Diffusion Models via Lightweight Erasers/e96e787a-46f5-4d12-8c30-eca8de901ff9_model.json @@ -0,0 +1,2304 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.293, + 0.141, + 0.717, + 0.21 + ], + "angle": 0, + "content": "Receler: Reliable Concept Erasing of Text-to-Image Diffusion Models via Lightweight Erasers" + }, + { + "type": "text", + "bbox": [ + 0.251, + 0.235, + 0.754, + 0.268 + ], + "angle": 0, + "content": "Chi-Pin Huang\\*1,\\*, Kai-Po Chang\\*1, Chung-Ting Tsai2, Yung-Hsuan Lai2, Fu-En Yang3, and Yu-Chiang Frank Wang\\*1,3,\\*" + }, + { + "type": "text", + "bbox": [ + 0.23, + 0.277, + 0.773, + 0.291 + ], + "angle": 0, + "content": "1 Graduate Institute of Communication Engineering, National Taiwan University" + }, + { + "type": "text", + "bbox": [ + 0.363, + 0.291, + 0.641, + 0.306 + ], + "angle": 0, + "content": "\\(^{2}\\) National Taiwan University, \\(^{3}\\) NVIDIA" + }, + { + "type": "text", + "bbox": [ + 0.33, + 0.306, + 0.674, + 0.32 + ], + "angle": 0, + "content": "† f11942097@ntu.edu.tw, ‡ frankwang@nvidia.com" + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.359, + 0.744, + 0.54 + ], + "angle": 0, + "content": "Abstract. Concept erasure in text-to-image diffusion models aims to disable pre-trained diffusion models from generating images related to a target concept. To perform reliable concept erasure, the properties of robustness and locality are desirable. The former refrains the model from producing images associated with the target concept for any paraphrased or learned prompts, while the latter preserves its ability in generating images with non-target concepts. In this paper, we propose Reliable Concept Erasing via Lightweight Erasers (Receler). It learns a lightweight Eraser to perform concept erasing while satisfying the above desirable properties through the proposed concept-localized regularization and adversarial prompt learning scheme. Experiments with various concepts verify the superiority of Receler over previous methods. Code is available at https://github.com/jasper0314-huang/Receler." + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.552, + 0.743, + 0.568 + ], + "angle": 0, + "content": "Keywords: Concept Erasing \\(\\cdot\\) Diffusion Models \\(\\cdot\\) Adversarial Learning" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.617, + 0.378, + 0.634 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.649, + 0.788, + 0.817 + ], + "angle": 0, + "content": "In recent years, text-to-image generation models [4, 32, 34, 36, 38] have achieved unprecedented success in generating photo-realistic images which benefit various industrial applications [34, 38]. Despite their apparent success and convenience, these models may produce images that are deemed NSFW (Not Safe for Work) [21] or infringe upon intellectual property and portrait rights [1], e.g., generating nudity, violent content, or imitating the style of well-known artists. This issue is mainly due to the memorization of the large-scale training data sourced from the web [8, 43]. To address the above problem, an intuitive solution is to manually filter out the inappropriate images and re-train the model. However, as pointed out in [16], this may lead to unpredictable results such as exposing more inappropriate content to be memorized [7] or incomplete visual" + }, + { + "type": "page_footnote", + "bbox": [ + 0.238, + 0.826, + 0.381, + 0.841 + ], + "angle": 0, + "content": "*Equal Contribution" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.356, + 0.129 + ], + "angle": 0, + "content": "Huang et al." + }, + { + "type": "image", + "bbox": [ + 0.226, + 0.144, + 0.778, + 0.325 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.335, + 0.788, + 0.406 + ], + "angle": 0, + "content": "Fig. 1: Illustration of (a) robustness and (b) locality-preserving in concept erasing. The former requires models to be robust against paraphrased attacks from the target concept (i.e., \"nudity\"), while the latter aims to preserve the visual content of non-target concepts (e.g., \"tennis player\" or \"airplane\"). Note that SD denotes Stable Diffusion [36]; UCE [16] and ESD [15] are recent works on concept erasing." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.441, + 0.784, + 0.472 + ], + "angle": 0, + "content": "concept removal [35]. Moreover, even though filtered image data can be collected, re-training generative models is still computationally expensive [15,23,42]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.476, + 0.788, + 0.686 + ], + "angle": 0, + "content": "Instead of processing training data and re-training models, an alternative is to erase or unlearn specific concepts from the pre-trained model [15,16,23,52]. That is, given a concept described in text, the pre-trained model is fine-tuned to forget that concept so that the associated image content cannot be generated from the fine-tuned model. In practice, it would be desirable to perform concept erasing with sufficient reliability, which suggests two desirable properties: locality and robustness. Locality indicates the ability to preserve the model generalization in synthesizing content not associated with the target concept [31,42]. Robustness requires erased models to effectively remove the target concept [31,42], while not be circumvented by paraphrased prompts that aim to recover the target concept (e.g., \"car\" vs. \"jeep\"). While both locality and robustness have been recently discussed in the field of NLP [10,12,20,29,50], the developed techniques cannot be directly applied to text-to-image generative models for performing concept erasing." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.69, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Recently, a number of methods for concept erasing or unlearning for diffusion models have been proposed [15, 16, 23, 52]. For example, Ablating [23] predefines an anchor concept for each target concept that needs to be unlearned and then achieves model unlearning by mapping the image distribution of the target concept to that of the anchor concept, e.g. mapping \"grumpy cat\" to \"cat.\" Inspired by classifier-free guidance [19], ESD [15] fine-tunes the model to predict negatively guided noise. In other words, it decreases the probability of generating images of the target concept, thus unlearning that concept. FMN [52] designs a computationally efficient unlearning method by directly minimizing the cross-attention weights corresponding to the target concept in the input text prompt," + } + ], + [ + { + "type": "header", + "bbox": [ + 0.68, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "Receler" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.787, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.785, + 0.207 + ], + "angle": 0, + "content": "encouraging the model to ignore the concept. UCE [16] employs a closed-form editing approach to optimize the projection matrices of keys and values in cross-attention layers, ensuring the model maintains locality when unlearning the target concept." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.208, + 0.788, + 0.449 + ], + "angle": 0, + "content": "Although promising progress has been made in erasing specific target concepts, most existing works are not specifically designed to preserve model locality and robustness. For example, Ablating [23] and ESD [15] fine-tune a considerable amount of parameters within pre-trained diffusion models to achieve concept erasure, which inevitably compromises the original capabilities of the model. On the other hand, methods such as UCE [16] and FMN [52] only modify specific parameters (i.e., the projection matrices of keys and values in cross-attention layers) responsible for encoding input textual features instead of visual ones. These methods would be vulnerable to rephrased target concepts since they only learn to dissociate textual prompts in the cross-attention layers (i.e., lack of ability to recognize that paraphrased queries are semantically similar to the erased target concept). For example, a diffusion model that has been erased of the concept of \"car\" may still produce images of jeeps due to its inability to recognize that jeeps fall under the category of cars; hence, it cannot remove the attention to jeep in cross-attention layers. As a result, proposing a concept-erasing method that addresses locality and robustness continues to pose a crucial challenge." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.45, + 0.788, + 0.691 + ], + "angle": 0, + "content": "In this paper, we propose Reliable Concept Erasing via Lightweight Erasers (Receler) for erasing concepts from pre-trained diffusion models, exhibiting sufficient locality and robustness properties. Receler involves a lightweight eraser (only \\(0.37\\%\\) of the U-Net parameters), which is designed to remove a target concept from the outputs of cross-attention layers. During this unlearning process, we train the eraser while preserving the image generation capability of the pre-trained diffusion models. Furthermore, a concept-localized regularization is introduced to ensure that the eraser focuses on erasing the target concept. This regularization prevents the generation of non-target concepts from being affected, thereby preserving locality. As for robustness, we advance adversarial prompt learning, which optimizes the adversarial prompts that induce the model to generate images of the target concept and then fine-tunes the eraser to erase images generated with these prompts. By training our eraser with concept-localized regularization and adversarial prompt learning, we are able to preserve the image generation capability of non-target concepts and robustly refrain the model from generating images with target-concept content." + }, + { + "type": "text", + "bbox": [ + 0.24, + 0.691, + 0.651, + 0.705 + ], + "angle": 0, + "content": "We now summarize the contributions of this work below:" + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.713, + 0.787, + 0.759 + ], + "angle": 0, + "content": "- We present Reliable Concept Erasing via Lightweight Erasers (Receler), a novel approach using a lightweight eraser (only \\(0.37\\%\\) of the U-Net parameters) for reliable and efficient concept erasing." + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.761, + 0.787, + 0.806 + ], + "angle": 0, + "content": "- Locality is introduced through concept-localized regularization, which constrains the eraser for precise erasing of the target concept without affecting the generation of non-target ones." + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.81, + 0.792, + 0.84 + ], + "angle": 0, + "content": "- Receler is trained against adversarial prompts, imitating paraphrased prompts of target concepts, resulting in improved robustness in concept erasure." + }, + { + "type": "list", + "bbox": [ + 0.223, + 0.713, + 0.792, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.232, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.358, + 0.129 + ], + "angle": 0, + "content": "Huang et al." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.145, + 0.398, + 0.161 + ], + "angle": 0, + "content": "2 Related Works" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.177, + 0.597, + 0.193 + ], + "angle": 0, + "content": "2.1 Erasing Concepts from Diffusion Models" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.203, + 0.788, + 0.354 + ], + "angle": 0, + "content": "Text-to-image diffusion models [4, 32, 34, 36, 38] have been shown to generate high-quality images with impressive generalization. However, such models are typically trained on extensive web-crawled data (e.g., LAION-5B [41]), which could memorize NSFW or copyrighted content, leading to the generation or replication of undesired images [9, 44]. To address this issue, some works explore the solution without the need to update the model weights. For instance, Stable Diffusion [36] employs an unsafe content classifier to filter out risky outputs, while SLD [39] uses negative guidance to prevent inappropriate content generation. However, the former relies on the reliability of pre-trained classifiers, while the latter only suppresses undesired concepts without complete removal." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.355, + 0.789, + 0.704 + ], + "angle": 0, + "content": "In response, several works focus on fine-tuning the diffusion model to erase the target concepts [15, 16, 23, 52]. Ablating [23] associates each concept to be erased, e.g., \"grumpy cat,\" with a broader, predefined anchor concept, e.g., \"cat\" and fine-tunes the diffusion model to map the generated image of the target concept to that of the anchor concept by minimizing the L2 distance of predicted noises. Inspired by classifier-free guidance [19], ESD [15] proposes to decrease the likelihood of generating images belonging to the target concept. This is achieved by fine-tuning the diffusion model to predict negatively guided noises, effectively steering the model's conditional prediction away from the erased concept. FMN [52] adopts attention resteing, a computationally efficient unlearning method, to identify attention maps associated with the target concept in the diffusion U-Net's cross-attention layers. By minimizing the attention weights corresponding to the target concept, the diffusion model gradually disregards the target concept during image generation, facilitating the erasure of the concept. UCE [16] employs a closed-form editing method to optimize the projection matrices of keys and values in cross-attention. The objective is to align the embedding of a source prompt (e.g., \"a photo of an airplane\") more closely with that of a destination (e.g., an empty string), while leaving other unrelated concepts unchanged. Despite their effectiveness in erasing concepts, most current methods are not particularly designed to preserve robustness against paraphrased prompts (e.g., \"nudity\" vs. \"without clothes\"). Moreover, both [23] and [15] require fine-tuning a considerable number of model parameters, which might affect the model capability and consequently compromise the locality property." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.724, + 0.625, + 0.74 + ], + "angle": 0, + "content": "2.2 Controlling Text-to-Image Diffusion Models" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Parameter-Efficient Fine-Tuning (PEFT) is a training scheme that addresses the challenges of extensive parameter updates, especially for large language models. These approaches update only a small subset of parameters, thereby reducing the risk of compromising the pre-trained capabilities of the model. Recent researches [14,24,30,37,51,53,55] have applied PEFT to control text-to-image diffusion models. For instance, some studies [14,24,37] empower the model to learn" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.681, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "Receler" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.787, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.298 + ], + "angle": 0, + "content": "personalized or unseen concepts by fine-tuning a new text token and a small number of parameters using few user-provided exemplar images. Meanwhile, other works [30, 51, 53, 55] aim to enable diffusion models to generate images based on additional conditions, e.g., edge maps, depth maps, or segmentation masks. They fine-tune lightweight task-specific modules with condition-image pairs to achieve control over the image generation process. Despite the effectiveness of these methods, they achieve learning new concepts by accessing the associated data of interest during training. As for concept erasing, since one only observes the description of the concept to be unlearned, existing PEFT-based methods cannot be directly applied." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.325, + 0.516, + 0.34 + ], + "angle": 0, + "content": "2.3 Adversarial Attack & Training" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.354, + 0.788, + 0.565 + ], + "angle": 0, + "content": "In adversarial attacks [2, 13, 17, 25], adversarial examples are deliberately constructed inputs, which would deceive models into making incorrect predictions. Popular methods such as Fast Gradient Sign Method (FGSM [13]) and its variants (I-FGSM [13] and MI-FGSM [13]) targeted at attacking classification models, utilizing the resulting loss gradients to produce imperceptible perturbations and induce misclassification. In contrast, recent approaches [11, 47, 54] introduce prompt-based adversarial attacks tailored to provoke seemingly unlearned diffusion models into generating images of the unlearned concepts. For instance, P4D [11] learns prompts to reconstruct noise associated with the target concept in diffusion models, and Ring-A-Bell [47] extracts holistic concept representations from CLIP model [33] to generate model-agnostic attack prompts. These learned attack prompts have been shown to provoke unlearned models to regenerate images of the erased concept, posing potential issues in text-to-image generative models." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.566, + 0.788, + 0.687 + ], + "angle": 0, + "content": "To defend against such adversarial attacks, various adversarial training strategies have been proposed [3, 17, 28]. Such adversarial training schemes expose models to adversarial examples during training, enabling the models to learn and recognize these examples. Consequently, these adversially trained models can respond accurately when encountering adversarial examples during inference. Inspired by adversarial training approaches, we employ an adversarial erasing learning scheme to introduce additional robustness to the unlearned model. We will detail our proposed framework in the following section." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.714, + 0.331, + 0.73 + ], + "angle": 0, + "content": "3 Method" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Problem formulation. We first define the setting and notations of our Receler. Given a pre-trained text-to-image diffusion model, parameterized by \\(\\theta\\), we aim to erase a textual concept \\(c\\) from the model without requiring access to the corresponding image data. The erasure is considered successful when the model no longer generates images that contain or represent the concept \\(c\\) (e.g., \"nudity\"-erased model should not generate any images with exposed body parts)." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.356, + 0.129 + ], + "angle": 0, + "content": "Huang et al." + }, + { + "type": "image", + "bbox": [ + 0.224, + 0.142, + 0.339, + 0.337 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.342, + 0.142, + 0.621, + 0.337 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.417, + 0.337, + 0.432, + 0.348 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.621, + 0.142, + 0.782, + 0.336 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.696, + 0.337, + 0.712, + 0.347 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.359, + 0.788, + 0.457 + ], + "angle": 0, + "content": "Fig. 2: Overview of Receler. (a) Receler involves iterative learning of a lightweight Eraser \\( E \\) and adversarial prompt embedding \\( e_{Adv} \\). The former is trained to erase the target concept \\( c \\) while preserving non-target concepts, and the latter learns to imitate the prompts to recover visual content associated with the concept previously erased. (b) The Eraser \\( E \\) is inserted after each cross attention layer of Diffusion U-Net to remove the target concept from its outputs, with prediction \\( o^l \\) directly added to the cross attention output." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.488, + 0.788, + 0.564 + ], + "angle": 0, + "content": "As depicted in Fig. 2, our method employs a lightweight Eraser \\( E \\), parameterized by \\( \\theta_E \\), to learn to erase the target concept \\( c \\), as discussed in Sec. 3.1. To introduce the desirable locality and robustness to our model, we incorporate concept-localized regularization and adversarial prompt learning schemes into our framework, as detailed in Sec. 3.2 and Sec. 3.3." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.587, + 0.604, + 0.603 + ], + "angle": 0, + "content": "3.1 Concept Erasing with Lightweight Eraser" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.614, + 0.789, + 0.796 + ], + "angle": 0, + "content": "In order to erase particular visual concepts from a pre-trained diffusion model, we introduce a lightweight adapter-based eraser. As depicted in Fig. 2, by solely finetuning the newly introduced \\(\\theta_{E}\\), our goal is to unlearn the target visual concept while preserving model generalization on non-target concepts. To be specific, our eraser is designed to remove the target concept from the output visual features of each cross-attention layer within the diffusion U-Net. This is based on the fact that these layers are responsible for incorporating the input text concept into the visual features. Thus, our eraser is positioned subsequent to each of the cross-attention layers, which is in line with the empirical analysis presented in [49]. To erase the concept, the eraser is trained to predict the negatively guided noises [15, 19] that move the model's prediction away from the erased concept. The objective is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.348, + 0.806, + 0.616, + 0.826 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {E r a s e} = \\mathbb {E} _ {x _ {t}, t} \\left[ \\lVert \\epsilon_ {\\theta^ {\\prime}} (x _ {t}, e _ {c}, t) - \\epsilon_ {E} \\rVert^ {2} \\right],\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.328, + 0.826, + 0.676, + 0.843 + ], + "angle": 0, + "content": "where \\(\\epsilon_{E} = \\epsilon_{\\theta}(x_{t},t) - \\eta [\\epsilon_{\\theta}(x_{t},e_{c},t) - \\epsilon_{\\theta}(x_{t},t)]\\)" + }, + { + "type": "text", + "bbox": [ + 0.639, + 0.818, + 0.785, + 0.833 + ], + "angle": 0, + "content": "(1)" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.68, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "Receler" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.787, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.147, + 0.788, + 0.239 + ], + "angle": 0, + "content": "Note that \\(\\theta' = \\{\\theta, \\theta_E\\}\\) represents the parameters of the diffusion model plugged with eraser, \\(x_t \\in \\mathbb{R}^{W \\times H \\times d}\\) is the denoised image at timestep \\(t\\) sampled from \\(\\theta'\\) conditioned on \\(c\\), \\(e_c\\) is the text embedding of concept \\(c\\), and \\(\\epsilon_E\\) is the negatively guided noises [15, 19] predicted by \\(\\theta\\). By minimizing the L2 distance between \\(\\epsilon_{\\theta'}(x_t, e_c, t)\\) and \\(\\epsilon_E\\), the eraser learns to reduce the probability of the generated image \\(x\\) belongs to the target concept \\(c\\), thus effectively erasing the concept." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.259, + 0.709, + 0.274 + ], + "angle": 0, + "content": "3.2 Concept-Localized Regularization for Erasing Locality" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.282, + 0.789, + 0.404 + ], + "angle": 0, + "content": "As the first desirable property in concept erasing, locality refers to preserving the model's ability to synthesize content unrelated to the target, which is realized by enforcing the eraser to affect the image synthesis process only if the target concept is present. To achieve this, we introduce concept-localized regularization into our Receler by leveraging the spatial information associated with the target concept's text tokens to regularize the eraser outputs. Specifically, inspired by [6, 46], we obtain the binary target concept mask \\( M \\in \\mathbb{R}^{W/4 \\times H/4} \\) by thresholding the attention maps when predicting \\( \\epsilon_{\\theta}(x_t, e_c, t) \\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.384, + 0.413, + 0.787, + 0.468 + ], + "angle": 0, + "content": "\\[\nM _ {i, j} = \\left\\{ \\begin{array}{l l} 1, & \\text {i f} \\frac {1}{| \\mathcal {S} |} \\sum_ {s \\in \\mathcal {S}} A _ {i, j} ^ {s} \\geq \\tau , \\\\ 0, & \\text {o t h e r w i s e .} \\end{array} \\right. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.477, + 0.789, + 0.555 + ], + "angle": 0, + "content": "where \\(\\mathcal{S}\\) is the set of indices of all U-Net's mid-layers with resolution \\((\\frac{W}{4},\\frac{H}{4})\\), \\(A^s\\in \\mathbb{R}^{W / 4\\times H / 4}\\) is the cross-attention map in the \\(s\\)-th layer of the text tokens corresponding to the concept \\(c\\), and \\(\\tau\\) is a pre-defined threshold. With \\(M\\) obtained, we calculate the following regularization loss to regularize the outputs of the eraser as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.394, + 0.563, + 0.787, + 0.602 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {R e g} = \\frac {1}{L} \\sum_ {l = 1} ^ {L} \\| o ^ {l} \\odot (1 - \\tilde {M}) \\| ^ {2}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.613, + 0.788, + 0.707 + ], + "angle": 0, + "content": "where \\( L \\) is the number of U-Net's layers, \\( \\odot \\) is the element-wise product, \\( o^l \\in \\mathbb{R}^{w^l \\times h^l \\times d} \\) is the output of the eraser in the \\( l \\)-th layer with the resolution \\( (w^l, h^l) \\), and \\( d \\) is the feature dimension. Note that \\( \\tilde{M} \\) is the shorthand of \\( M \\) being bicubically upscaled to the same resolution as \\( o^l \\). With the above regularization introduced, our Receler is enforced to preserve the generation of non-target concepts. Thus, the diversity and fidelity of the original model can be maintained." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.726, + 0.7, + 0.742 + ], + "angle": 0, + "content": "3.3 Adversarial Prompt Learning for Erasing Robustness" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.841 + ], + "angle": 0, + "content": "To further ensure our Receler being robust against prompting attacks (e.g., by paraphrased or learned prompts), we introduce an adversarial learning strategy into our framework to enrich such model robustness. In our framework, we present a unique adversarial prompt learning scheme, which learns prompting attacks that would induce the diffusion model to synthesize the images containing previously erased concepts. To achieve this, we design the adversarial loss," + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.356, + 0.129 + ], + "angle": 0, + "content": "Huang et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.193 + ], + "angle": 0, + "content": "which optimizes the continuous soft prompts \\( e_{Adv} \\) by encouraging such learned prompts to imitate the malicious prompts, as illustrated in Fig. 2. More precisely, the objective \\( \\mathcal{L}_{Adv} \\) is defined as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.374, + 0.207, + 0.786, + 0.226 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {A d v} = \\mathbb {E} _ {x _ {t}, t} \\left[ \\| \\epsilon_ {\\theta^ {\\prime}} (x _ {t}, e ^ {\\prime}, t) - \\epsilon_ {M} \\| ^ {2} \\right], \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.237, + 0.79, + 0.374 + ], + "angle": 0, + "content": "where \\( e' = [e_c; e_{Adv}] \\) is the concatenated prompts of the erased target concept embedding and the learned soft prompts, and \\( \\epsilon_M = \\epsilon_\\theta(x_t, e_c, t) \\) represents the malicious noise predicted by the pre-trained diffusion model (without safety mechanism) conditioned on the target concept. By minimizing this adversarial loss, \\( e_{Adv} \\) learns to regenerate the unlearned concept from the erased model. The optimization of the soft prompt \\( e_{Adv} \\) and the eraser \\( \\theta_E \\) is performed iteratively, with each being fixed while the other is trained. Thus, \\( \\theta_E \\) and \\( e_{Adv} \\) are trained against each other to improve model robustness. For more details, please refer to the pseudo-algorithm in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.398, + 0.377, + 0.417 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.432, + 0.788, + 0.492 + ], + "angle": 0, + "content": "In this section, we first conduct quantitative experiments to assess the robustness and locality of Receler compared to state-of-the-art baselines, followed by ablation studies of our method. Lastly, we present qualitative comparisons and visualizations to demonstrate its effectiveness." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.516, + 0.788, + 0.562 + ], + "angle": 0, + "content": "Datasets. We conduct experiments on erasing objects defined in the CIFAR-10 dataset [22] and on erasing inappropriate contents from the Inappropriate Image Prompts (I2P) dataset [39]:" + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.573, + 0.788, + 0.753 + ], + "angle": 0, + "content": "- Object Erasure. To evaluate the effectiveness of erasure methods in erasing common visual concepts, we choose to erase ten class labels from CIFAR-10 [22]. Note that during our experiment, we only utilize the label set, not the images. For comprehensive assessment, we devise two types of evaluation prompts for each class: Firstly, we use simple prompts formatted as \"A photo of {class}\" to evaluate the efficacy of erasure in removing the target concept. Secondly, to further assess the robustness, we generate 50 paraphrased prompts for each class using ChatGPT3 to simulate real-world scenarios where prompts are typically more complex and target concepts may not be explicitly mentioned. For example, a paraphrased prompt for \"airplane\" is \"A sleek, black stealth bomber flying low over a desert landscape at dusk.\" More details can be found in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.758, + 0.788, + 0.819 + ], + "angle": 0, + "content": "- Inappropriate Content Erasure. The I2P dataset [39], collected from a text-to-image prompt dataset DiffusionDB [48], comprises 4,703 real-world, user-generated prompts that produce inappropriate images, including hate, harassment, violence, self-harm, shocking, sexual, and illegal content." + }, + { + "type": "list", + "bbox": [ + 0.223, + 0.573, + 0.788, + 0.819 + ], + "angle": 0, + "content": null + }, + { + "type": "page_footnote", + "bbox": [ + 0.238, + 0.825, + 0.419, + 0.842 + ], + "angle": 0, + "content": "3https://chat.openai.com/" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.681, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "Receler" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.147, + 0.784, + 0.177 + ], + "angle": 0, + "content": "Evaluation Setup. We assess the robustness and locality of the erasure methods for object erasure and inappropriate content erasure as follows:" + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.186, + 0.788, + 0.382 + ], + "angle": 0, + "content": "- Object Erasure. For each method, we fine-tune ten models, each erasing one CIFAR-10 class. Each model is then evaluated by: 1) Efficacy \\((\\mathrm{Acc}_E)\\): the percentage of the target class being erased when inputting simple prompts; lower values are better. 2) Robustness \\((\\mathrm{Acc}_R)\\): the percentage of the target class being erased when inputting paraphrased prompts; lower values are better. 3) Locality \\((\\mathrm{Acc}_L)\\): the percentage of non-target classes being preserved; higher values are better. To assess efficacy and robustness, we generate 150 images using simple and paraphrased prompts for the target class, respectively. For Locality, we generate 50 images for each of the nine non-target classes using paraphrased prompts. We then use GroundingDINO [27] to detect if the corresponding class is presenting in the image, thereby assessing \\(\\mathrm{Acc}_E\\), \\(\\mathrm{Acc}_R\\), and \\(\\mathrm{Acc}_L\\). To further evaluate the overall performance, we calculate the harmonic mean (H) of \\(100 - \\mathrm{Acc}_E\\), \\(100 - \\mathrm{Acc}_R\\), and \\(\\mathrm{Acc}_L\\)." + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.382, + 0.788, + 0.518 + ], + "angle": 0, + "content": "- Inappropriate Content Erasure. Following ESD [15], we fine-tune two models for each erasure method: one for \"nudity\" and the other for the predefined inappropriate concepts e.g. hate, harassment, and violence. We use the NudeNet detector [5] to detect nudity and both the NudeNet and the Q16 detector [40] to identify the inappropriate concepts. Model robustness is evaluated by using real-world prompts in I2P [39]. Locality is evaluated using COCO-30K [26], a nudity-free dataset, by employing the nudity-erased model to generate safe contents from COCO-30K prompts and evaluating the quality of the generated images in terms of FID [18] and CLIP [33]." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.539, + 0.788, + 0.644 + ], + "angle": 0, + "content": "Comparisons. We compare Receler to state-of-the-art erasing methods, including FMN [52], SLD [39], Ablating [23], ESD [15], and UCE [16]. For all methods, we use the open-sourced codebases and follow their reported settings. We fine-tune all models from SD v1.4 [36], and for all image generation, we employ DDIM sampler [45] over 50 steps and a guidance scale of 7.5. Following SD v1.4, the image resolution in all our experiments is \\(512 \\times 512\\). Please refer to the supplementary for more experiment setup and implementation details." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.666, + 0.46, + 0.681 + ], + "angle": 0, + "content": "4.1 Quantitative Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.689, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Object Erasure. In Tab. 1, we show that Receler surpasses previous state-of-the-art methods in erasing common visual concepts from CIFAR-10 [22] class labels. Notably, Receler achieves the highest harmonic mean \\((H)\\) and exceeds the second-best method by 11.2 points, highlighting its effectiveness in erasing concepts with sufficient robustness and locality. Specifically, when assessing method efficacy using the simple prompts \\((\\mathrm{Acc}_E)\\), Receler achieves an average accuracy of \\(14.9\\%\\) across the erased classes, \\(2.2\\%\\) better than the second-best method, ESD [15]. In addition, when evaluating method robustness with paraphrased prompts \\((\\mathrm{Acc}_R)\\), Receler reaches \\(17.6\\%\\) on average, outperforming ESD by \\(22.3\\%\\) in the erased classes. We evaluate the locality of the erased model by" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.356, + 0.129 + ], + "angle": 0, + "content": "Huang et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.788, + 0.215 + ], + "angle": 0, + "content": "Table 1: Evaluation of erasing common objects in CIFAR-10 classes. \\(\\mathrm{Acc}_E\\) and \\(\\mathrm{Acc}_R\\) represent efficacy and robustness, indicating accuracy of target class (which should be minimized) on simple and paraphrased prompts, respectively. Locality, \\(\\mathrm{Acc}_L\\), is accuracy of non-target classes (which should be maximized) using paraphrased prompts. The harmonic mean \\(H\\) reflects overall assessment of \\(\\mathrm{Acc}_E\\), \\(\\mathrm{Acc}_R\\), and \\(\\mathrm{Acc}_L\\)." + }, + { + "type": "table", + "bbox": [ + 0.227, + 0.228, + 0.777, + 0.571 + ], + "angle": 0, + "content": "
MethodsMetricsErased conceptsavg.
air-planeauto-mobilebirdcatdeerdogfroghorseshiptruck
SD v1.4AccE89.399.393.396.799.398.796.097.395.396.096.1
AccR79.394.096.088.098.792.088.792.765.384.087.9
AccL88.887.287.087.986.787.487.887.390.488.387.9
FMN [52]AccE ↓93.397.390.092.098.095.384.795.388.794.792.9
AccR ↓80.796.793.370.795.386.767.395.360.784.083.1
AccL ↑88.088.286.087.684.488.086.485.890.488.987.4
H↑14.14.411.517.64.110.027.96.924.011.414.2
Ablating [23]AccE ↓78.074.776.793.392.797.394.7100.090.786.788.5
AccR ↓67.390.093.372.795.387.371.390.758.076.080.2
AccL ↑87.883.884.987.384.085.886.085.388.286.486.0
H↑34.319.814.715.28.36.512.80.021.023.420.1
ESD [15]AccE ↓20.044.011.314.019.320.013.38.716.04.717.1
AccR ↓33.381.354.018.040.727.338.741.332.032.739.9
AccL ↑83.679.872.971.868.070.079.368.286.779.175.9
H↑76.035.864.279.568.274.074.270.378.679.071.6
UCE [16]AccE ↓34.746.08.716.74.011.311.36.022.010.017.1
AccR ↓58.079.363.316.015.349.328.034.041.339.342.4
AccL ↑84.979.181.882.078.082.283.375.887.381.681.6
H↑58.937.859.583.185.669.580.777.072.675.372.0
Receler (Ours)AccE ↓10.046.73.311.32.76.723.37.324.014.014.9
AccR ↓16.762.026.70.72.04.717.36.020.719.317.6
AccL ↑88.481.382.280.476.774.783.880.488.284.782.1
H↑87.152.383.088.889.586.780.988.680.883.783.2
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.604, + 0.788, + 0.695 + ], + "angle": 0, + "content": "examining its accuracy in the remaining classes \\((\\mathrm{Acc}_L)\\) aside from the erased class, focusing on whether erasing one CIFAR-10 class affects the image synthesis of the other unrelated classes. Although FMN [52] and Ablating [23] exhibit high average \\(\\mathrm{Acc}_L\\), appearing effective in preserving model locality, they struggle to erase the target objects, with only a \\(3.2\\%\\) and \\(7.6\\%\\) drop in \\(\\mathrm{Acc}_E\\) from SD, compared to our \\(81.2\\%\\) drop." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.719, + 0.788, + 0.809 + ], + "angle": 0, + "content": "Erasure of Inappropriate Content. In Tab. 2 and Tab. 3, we evaluate the robustness of erasing inappropriate content with real-world prompts from I2P dataset [39]. Compared to the second-best result, Receler stands out by achieving \\(4.3\\%\\) overall improvement in erasing sensitive concepts on I2P and a \\(3.2\\%\\) increase in erasing nudity content, underscoring its effectiveness in scenarios that require a safety mechanism." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.81, + 0.787, + 0.84 + ], + "angle": 0, + "content": "In Tab. 4, in addition to robustness, we assess locality using COCO-30K [26], a nudity-free dataset. We employ the \"nudity\"-erased model to produce safe" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.68, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "Receler" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.505, + 0.228 + ], + "angle": 0, + "content": "Table 2: Quantitative results on Inappropriate Image Prompts (I2P) dataset. We follow SLD [39] and apply the ratio of inappropriate images as the metric. More results compared with other baselines are available in supplementary." + }, + { + "type": "table", + "bbox": [ + 0.227, + 0.24, + 0.499, + 0.368 + ], + "angle": 0, + "content": "
Class nameInappropriate proportion (%) (↓)
SDFMNSLDESDUCEReceler
Hate44.237.722.526.836.428.6
Harassment37.525.022.124.029.521.7
Violence46.347.831.835.134.127.1
Self-harm47.946.830.033.730.824.8
Sexual60.259.152.435.025.529.4
Shocking59.558.140.540.141.134.8
Illegal activity40.037.022.126.729.021.3
Overall48.947.833.732.831.327.0
" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.382, + 0.505, + 0.452 + ], + "angle": 0, + "content": "Table 4: Assessment of reliability of nudity-erased models. Robustness is evaluated using the nudity prompts from I2P dataset, and locality is assessed using COCO-30K prompts." + }, + { + "type": "table", + "bbox": [ + 0.227, + 0.464, + 0.495, + 0.572 + ], + "angle": 0, + "content": "
MethodRobustness Nudity-erased ratio(↑)Locality
CLIP-30K(↑)FID-30K(↓)
SD-31.3214.27
FMN44.2%30.3913.52
SLD71.6%30.9016.34
ESD81.3%30.2415.31
UCE75.9%30.8514.07
Receler84.5%31.0214.10
" + }, + { + "type": "table_caption", + "bbox": [ + 0.51, + 0.147, + 0.788, + 0.218 + ], + "angle": 0, + "content": "Table 3: Quantitative results on nudity prompts from I2P dataset. We report the number of nudity images detected by the NudeNet [5]. F- and M- refer to female and male, respectively." + }, + { + "type": "table", + "bbox": [ + 0.522, + 0.229, + 0.778, + 0.365 + ], + "angle": 0, + "content": "
Class nameNumber of nudity detected (↓)
SDFMNSLDESDUCEReceler
Armpits1484246312939
Belly17011670206026
F-Breast26615539323513
M-Breast421730151212
Buttocks29123975
Feet635619242910
F-Genitalia18151151
M-Genitalia723749
Total743415211139179115
Erasing ratio%--44.2-71.6-81.3-75.9-84.5
" + }, + { + "type": "table_caption", + "bbox": [ + 0.51, + 0.387, + 0.789, + 0.457 + ], + "angle": 0, + "content": "Table 5: Evaluation of robustness against learned attack prompts. We report the failure rate, indicating the proportion of generated images belonging to the unlearned concept." + }, + { + "type": "table", + "bbox": [ + 0.522, + 0.469, + 0.778, + 0.567 + ], + "angle": 0, + "content": "
MethodP4D [11]Ring-A-Bell [47]
cifar10 avg.nudityviolencenudity
FMN88.3%89.4%98.8%94.7%
Ablating85.4%82.8%100.0%96.8%
SLD60.5%56.3%80.4%86.3%
ESD48.1%54.3%86.0%55.8%
UCE53.8%51.9%76.8%49.5%
Receler13.7%31.2%59.2%1.1%
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.601, + 0.788, + 0.679 + ], + "angle": 0, + "content": "content using COCO-30K prompts and evaluate the quality of the generated images using FID [18] and CLIP [33] metrics. As shown in the last two columns, Receler secures the highest CLIP-30K and nearly matches the top result on FID-30K. It is noteworthy that while Receler performs comparably to FMN in FID-30K, it surpasses FMN in robustness by erasing \\(40.3\\%\\) more nudity content." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.704, + 0.789, + 0.841 + ], + "angle": 0, + "content": "Learned Attack Prompts. To demonstrate the robustness of Receler in safeguarding against the potential and unprecedented malicious attacks, we employ P4D [11] and Ring-A-Bell [47]. These tools are specifically designed for red-teaming text-to-image models by finding problematic prompts. As illustrated in Tab. 5, Receler is significantly more reliable than other concept erasing methods. Specifically, it achieves lower failure rates \\(-34.4\\%\\) for CIFAR-10 and \\(20.7\\%\\) for nudity against P4D prompts, and \\(17.6\\%\\) for violence and \\(48.4\\%\\) for nudity against Ring-A-Bell prompts, compared to the second-best method. These results highlight the robustness of Receler in defending against malicious attacks." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.356, + 0.129 + ], + "angle": 0, + "content": "Huang et al." + }, + { + "type": "image", + "bbox": [ + 0.276, + 0.154, + 0.731, + 0.34 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.351, + 0.788, + 0.394 + ], + "angle": 0, + "content": "Fig. 3: Qualitative comparison of concept erasure methods. Note that erased concepts are listed at the top, and images generated from each method are shown in each row. Input prompts used for image generation are provided in supplementary." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.42, + 0.789, + 0.618 + ], + "angle": 0, + "content": "Ablation Study. In Tab. 6, we ablate the three proposed components in Receler: the lightweight eraser, concept-localized regularization, and adversarial prompt learning. We establish a simple baseline by fine-tuning the whole model with Eq. (1) (i.e., the first row). With the eraser introduced, both \\(\\mathrm{Acc}_R\\) and \\(\\mathrm{Acc}_L\\) improve. Adding concept-localized regularization further increases the \\(\\mathrm{Acc}_L\\) from \\(77.4\\%\\) to \\(79.8\\%\\), demonstrating its effectiveness in enhancing locality. On the other hand, coupling adversarial prompting learning with the eraser boosts \\(\\mathrm{Acc}_R\\) by \\(14\\%\\), albeit with a slight decrease in \\(\\mathrm{Acc}_L\\). This result aligns with expectations, as adversarial prompt learning improves the robustness by restraining any possible malicious prompts. Therefore, Receler, which integrates both adversarial prompt learning and concept-localized regularization, yields the best experimental results. This implies that these two approaches benefit each other and enable the lightweight eraser to achieve both robustness and locality." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.637, + 0.449, + 0.652 + ], + "angle": 0, + "content": "4.2 Qualitative Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.659, + 0.788, + 0.779 + ], + "angle": 0, + "content": "Visualization of Erased Concepts with Paraphrased Prompts. In Fig. 3, we show examples of erasing different artist styles (e.g., van Gogh), objects (e.g., automobile) and high-level concepts (e.g., nudity). As observed in the figure, ESD [15] and UCE [16] fail to erase the target concept, with the outputs from these methods closely resembling the original images from SD [1]. On the contrary, Receler successfully erases all target concepts and is able to generate images that are visually close to the original ones, e.g., same background with the car removed, a human in the same posture with nudity removed." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.841 + ], + "angle": 0, + "content": "In Fig. 4, we qualitatively validate the robustness and locality of Receler. The diagonal orange boxes show its robustness against paraphrased prompts where the erased concept is not explicitly mentioned. For instance, airplane-erased Receler successfully prevents the generation of an \"airplane\" image when" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.681, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "Receler" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "image", + "bbox": [ + 0.259, + 0.156, + 0.749, + 0.446 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.456, + 0.788, + 0.469 + ], + "angle": 0, + "content": "Fig. 4: Visualization of robustness and locality from Receler on CIFAR-10." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.47, + 0.786, + 0.511 + ], + "angle": 0, + "content": "The red strikethrough at the top indicates the erased concepts. On the left, the input paraphrased prompts are provided. Images enclosed within the diagonal orange borders shows robustness while others show the locality." + }, + { + "type": "image", + "bbox": [ + 0.245, + 0.533, + 0.496, + 0.645 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.341, + 0.647, + 0.395, + 0.658 + ], + "angle": 0, + "content": "(a) nudity" + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.534, + 0.758, + 0.646 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.595, + 0.647, + 0.658, + 0.658 + ], + "angle": 0, + "content": "(b) violence" + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.671, + 0.788, + 0.701 + ], + "angle": 0, + "content": "Fig. 5: Visualization of erasure methods against learned attack prompts. We use Ring-A-Bell [47] to generate adversarial prompts for nudity and violence concepts." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.788, + 0.841 + ], + "angle": 0, + "content": "the prompt is paraphrased to \"jet\". Concerning locality, as shown in the non-diagonal boxes, Receler generates images that not only faithfully adhere to the descriptions but also closely resemble the images from the original diffusion model. Notably, it can be seen that Receler is able to replace the unlearned objects with reasonable alternatives rather than simply removing them, e.g., two white puppies are substituted with two people wearing white furry clothes and gloves (see the fourth row in Fig. 4)." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.356, + 0.129 + ], + "angle": 0, + "content": "Huang et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.149, + 0.482, + 0.233 + ], + "angle": 0, + "content": "Table 6: Ablation study on CIFAR-10. We ablate the components of Receler and report the robustness and locality metrics. The first row refers to fine-tuning all parameters with only \\(\\mathcal{L}_{\\text{Erase}}\\) in Eq. (1)." + }, + { + "type": "table", + "bbox": [ + 0.235, + 0.245, + 0.465, + 0.336 + ], + "angle": 0, + "content": "
ComponentsMetrics
Eraser\\( \\mathcal{L}_{\\text{Reg}} \\)\\( \\mathcal{L}_{\\text{Adv}} \\)\\( \\text{Acc}_R(\\downarrow) \\)\\( \\text{Acc}_L(\\uparrow) \\)
XXX39.475.2
✓XX34.977.4
✓✓X26.379.8
✓X✓20.976.2
✓✓✓17.682.1
" + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.145, + 0.594, + 0.269 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.597, + 0.145, + 0.688, + 0.27 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.692, + 0.146, + 0.784, + 0.269 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.271, + 0.787, + 0.327 + ], + "angle": 0, + "content": "Fig.6: Examples of erasing multiple concepts. Instead of training an eraser for multiple concepts from scratch, we combine existing erasers for multi-concept erasure." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.368, + 0.788, + 0.492 + ], + "angle": 0, + "content": "Visualization of Robustness against Learned Attack Prompts. In addition to the quantitative results shown in Tab. 5, we further qualitatively demonstrate the robustness of Receler against learned attack prompts in Fig. 5. The attack prompts are learned to induce the concepts of \"nudity\" and \"violence\" from models that should have erased these concepts using Ring-A-Bell [47]. For both nudity and violence attack prompts, Receler successfully prevents the generation of erased concepts, whereas other methods like ESD and UCE fail and generate images with the supposedly forbidden concepts (e.g., nudity or blood)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.512, + 0.788, + 0.605 + ], + "angle": 0, + "content": "Compositional Concept Erasure. In Fig. 6, we showcase examples of Receler in performing compositional concept erasure. This is achieved by combining outputs from separately trained erasers, each targeting a specific unlearned concept, during inference. By averaging these outputs, compositional concept erasure is accomplished without necessitating retraining. This approach notably offers the flexibility to selectively determine which concepts are to be erased, as required." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.627, + 0.36, + 0.643 + ], + "angle": 0, + "content": "5 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.659, + 0.788, + 0.827 + ], + "angle": 0, + "content": "In this paper, we proposed Reliable Concept Erasing via Lightweight Erasers (Receler) to erase target concepts entirely from the pre-trained diffusion model against prompting attacks (i.e., robustness), while preserving its image generation ability of other concepts (i.e., locality). In Receler, we employ concept-localized regularization to enforce the eraser to only affect the visual features related to the target concept. To enhance model robustness to paraphrased or attack prompts, we present an adversarial prompt learning scheme to induce the model to produce images of previously erased concepts, followed by optimizing the model against such image generation. We conducted extensive quantitative and qualitative evaluations on Receler, validating its superior robustness and locality-preserving ability over previous concept-erasing methods." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.68, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "Receler" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.239 + ], + "angle": 0, + "content": "Acknowledgements. This research was supported in part by the National Science and Technology Council via grant NSTC 112-2634-F-002-007, NSTC 113-2640-E-002-003 and the Featured Area Research Center Program within the framework of the Higher Education Sprout Project by the Ministry of Education 113L900902. We also thank the National Center for High-performance Computing (NCHC) for providing computational and storage resources." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.262, + 0.324, + 0.279 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.295, + 0.785, + 0.323 + ], + "angle": 0, + "content": "1. Sarah andersen. et al v. stability ai ltd. et al. case no.3:2023cv00201. us district court for the northern district of california. (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.324, + 0.785, + 0.351 + ], + "angle": 0, + "content": "2. Akhtar, N., Mian, A., Kardan, N., Shah, M.: Advances in adversarial attacks and defenses in computer vision: A survey. IEEE Access 9, 155161-155196 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.352, + 0.785, + 0.379 + ], + "angle": 0, + "content": "3. Bai, T., Luo, J., Zhao, J., Wen, B., Wang, Q.: Recent advances in adversarial training for adversarial robustness. arXiv preprint arXiv:2102.01356 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.38, + 0.785, + 0.421 + ], + "angle": 0, + "content": "4. Balaji, Y., Nah, S., Huang, X., Vahdat, A., Song, J., Kreis, K., Aittala, M., Aila, T., Laine, S., Catanzaro, B., et al.: edifi: Text-to-image diffusion models with an ensemble of expert denoisers. arXiv preprint arXiv:2211.01324 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.422, + 0.785, + 0.449 + ], + "angle": 0, + "content": "5. Bedapudi Praneeth, b.k., lireza Ayinmehr: Nudenet: Neural nets for nudity classification, detection and selective censoring (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.45, + 0.785, + 0.491 + ], + "angle": 0, + "content": "6. Cao, M., Wang, X., Qi, Z., Shan, Y., Qie, X., Zheng, Y.: Masactrl: Tuning-free mutual self-attention control for consistent image synthesis and editing. arXiv preprint arXiv:2304.08465 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.492, + 0.785, + 0.518 + ], + "angle": 0, + "content": "7. Carlini, N., Jagielski, M., Zhang, C., Papernot, N., Terzis, A., Tramer, F.: The privacy onion effect: Memorization is relative. In: NeurIPS (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.519, + 0.785, + 0.561 + ], + "angle": 0, + "content": "8. Carlini, N., Hayes, J., Nasr, M., Jagielski, M., Sehwag, V., Tramer, F., Balle, B., Ippolito, D., Wallace, E.: Extracting training data from diffusion models. In: 32nd USENIX Security Symposium (USENIX Security 23). pp. 5253-5270 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.562, + 0.785, + 0.603 + ], + "angle": 0, + "content": "9. Carlini, N., Hayes, J., Nasr, M., Jagielski, M., Sehwag, V., Tramer, F., Balle, B., Ippolito, D., Wallace, E.: Extracting training data from diffusion models. In: 32nd USENIX Security Symposium (USENIX Security 23). pp. 5253-5270 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.604, + 0.785, + 0.644 + ], + "angle": 0, + "content": "0. Chen, J., Yang, D.: Unlearn what you want to forget: Efficient unlearning for LLMs. In: Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing. Association for Computational Linguistics (Dec 2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.645, + 0.785, + 0.686 + ], + "angle": 0, + "content": "1. Chin, Z.Y., Jiang, C.M., Huang, C.C., Chen, P.Y., Chiu, W.C.: Prompting4debugging: Red-teaming text-to-image diffusion models by finding problematic prompts. arXiv preprint arXiv:2309.06135 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.687, + 0.785, + 0.714 + ], + "angle": 0, + "content": "2. De Cao, N., Aziz, W., Titov, I.: Editing factual knowledge in language models. arXiv preprint arXiv:2104.08164 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.715, + 0.785, + 0.742 + ], + "angle": 0, + "content": "3. Dong, Y., Liao, F., Pang, T., Su, H., Zhu, J., Hu, X., Li, J.: Boosting adversarial attacks with momentum. In: CVPR (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.743, + 0.785, + 0.784 + ], + "angle": 0, + "content": "4. Gal, R., Alaluf, Y., Atzmon, Y., Patashnik, O., Bermano, A.H., Chechik, G., Cohen-Or, D.: An image is worth one word: Personalizing text-to-image generation using textual inversion. arXiv preprint arXiv:2208.01618 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.785, + 0.785, + 0.812 + ], + "angle": 0, + "content": "5. Gandikota, R., Materzynska, J., Fiotto-Kaufman, J., Bau, D.: Erasing concepts from diffusion models. arXiv preprint arXiv:2303.07345 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.813, + 0.785, + 0.84 + ], + "angle": 0, + "content": "6. Gandikota, R., Orgad, H., Belinkov, Y., Materzyńska, J., Bau, D.: Unified concept editing in diffusion models. arXiv preprint arXiv:2308.14761 (2023)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.295, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.356, + 0.129 + ], + "angle": 0, + "content": "Huang et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.148, + 0.785, + 0.175 + ], + "angle": 0, + "content": "17. Goodfellow, I.J., Shlens, J., Szegedy, C.: Explaining and harnessing adversarial examples. arXiv preprint arXiv:1412.6572 (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.177, + 0.785, + 0.218 + ], + "angle": 0, + "content": "18. Heusel, M., Ramsauer, H., Unterthiner, T., Nessler, B., Hochreiter, S.: Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems 30 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.219, + 0.785, + 0.246 + ], + "angle": 0, + "content": "19. Ho, J., Salimans, T.: Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.247, + 0.785, + 0.275 + ], + "angle": 0, + "content": "20. Huang, Z., Shen, Y., Zhang, X., Zhou, J., Rong, W., Xiong, Z.: Transformer-patcher: One mistake worth one neuron. arXiv preprint arXiv:2301.09785 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.276, + 0.785, + 0.303 + ], + "angle": 0, + "content": "21. Hunter, T.: Ai porn is easy to make now. for women, that's a nightmare. The Washington Post pp. NA-NA (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.304, + 0.785, + 0.332 + ], + "angle": 0, + "content": "22. Krizhevsky, A., Hinton, G.: Learning multiple layers of features from tiny images. Master's thesis, Department of Computer Science, University of Toronto (2009)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.333, + 0.785, + 0.36 + ], + "angle": 0, + "content": "23. Kumari, N., Zhang, B., Wang, S.Y., Shechtman, E., Zhang, R., Zhu, J.Y.: Ablating concepts in text-to-image diffusion models. In: ICCV (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.361, + 0.785, + 0.388 + ], + "angle": 0, + "content": "24. Kumari, N., Zhang, B., Zhang, R., Shechtman, E., Zhu, J.Y.: Multi-concept customization of text-to-image diffusion. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.39, + 0.785, + 0.431 + ], + "angle": 0, + "content": "25. Kurakin, A., Goodfellow, I.J., Bengio, S.: Adversarial examples in the physical world. In: Artificial intelligence safety and security, pp. 99-112. Chapman and Hall/CRC (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.432, + 0.785, + 0.487 + ], + "angle": 0, + "content": "26. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollar, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13. pp. 740-755. Springer (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.488, + 0.785, + 0.529 + ], + "angle": 0, + "content": "27. Liu, S., Zeng, Z., Ren, T., Li, F., Zhang, H., Yang, J., Li, C., Yang, J., Su, H., Zhu, J., et al.: Grounding dino: Marrying dino with grounded pre-training for open-set object detection. arXiv preprint arXiv:2303.05499 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.53, + 0.785, + 0.558 + ], + "angle": 0, + "content": "28. Madry, A., Makelov, A., Schmidt, L., Tsipras, D., Vladu, A.: Towards deep learning models resistant to adversarial attacks. arXiv preprint arXiv:1706.06083 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.559, + 0.785, + 0.586 + ], + "angle": 0, + "content": "29. Meng, K., Bau, D., Andonian, A., Belinkov, Y.: Locating and editing factual associations in gpt. In: NeurIPS (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.587, + 0.785, + 0.628 + ], + "angle": 0, + "content": "30. Mou, C., Wang, X., Xie, L., Zhang, J., Qi, Z., Shan, Y., Qie, X.: T2i-adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models. arXiv preprint arXiv:2302.08453 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.629, + 0.785, + 0.657 + ], + "angle": 0, + "content": "31. Ni, M., Wu, C., Wang, X., Yin, S., Wang, L., Liu, Z., Duan, N.: Orens: Open-vocabulary responsible visual synthesis. arXiv preprint arXiv:2308.13785 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.658, + 0.785, + 0.699 + ], + "angle": 0, + "content": "32. Nichol, A., Dhariwal, P., Ramesh, A., Shyam, P., Mishkin, P., McGrew, B., Sutskever, I., Chen, M.: Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.7, + 0.785, + 0.755 + ], + "angle": 0, + "content": "33. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.756, + 0.785, + 0.797 + ], + "angle": 0, + "content": "34. Ramesh, A., Dhariwal, P., Nichol, A., Chu, C., Chen, M.: Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125 1(2), 3 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.798, + 0.595, + 0.812 + ], + "angle": 0, + "content": "35. Rombach, R.: Stable diffusion 2.0 release (Nov 2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.813, + 0.785, + 0.84 + ], + "angle": 0, + "content": "36. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models (2021)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.148, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.681, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "Receler" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.189 + ], + "angle": 0, + "content": "37. Ruiz, N., Li, Y., Jampani, V., Pritch, Y., Rubinstein, M., Aberman, K.: Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.189, + 0.788, + 0.231 + ], + "angle": 0, + "content": "38. Sahara, C., Chan, W., Saxena, S., Li, L., Whang, J., Denton, E.L., Ghasemipour, K., Gontijo Lopes, R., Karagol Ayan, B., Salimans, T., et al.: Photorealistic text-to-image diffusion models with deep language understanding. In: NeurIPS (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.231, + 0.788, + 0.259 + ], + "angle": 0, + "content": "39. Schramowski, P., Brack, M., Deiseroth, B., Kersting, K.: Safe latent diffusion: Mitigating inappropriate degeneration in diffusion models. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.259, + 0.788, + 0.313 + ], + "angle": 0, + "content": "40. Schramowski, P., Tauchmann, C., Kersting, K.: Can machines help us answering question 16 in datasheets, and in turn reflecting on inappropriate content? In: Proceedings of the 2022 ACM Conference on Fairness, Accountability, and Transparency. pp. 1350-1361 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.313, + 0.788, + 0.369 + ], + "angle": 0, + "content": "41. Schuhmann, C., Beaumont, R., Vencu, R., Gordon, C., Wightman, R., Cherti, M., Coombes, T., Katta, A., Mullis, C., Wortsman, M., et al.: Laion-5b: An open large-scale dataset for training next generation image-text models. In: NeurIPS (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.37, + 0.788, + 0.397 + ], + "angle": 0, + "content": "42. Sinitsin, A., Plokhotnyuk, V., Pyrkin, D., Popov, S., Babenko, A.: Editable neural networks. arXiv preprint arXiv:2004.00345 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.397, + 0.788, + 0.438 + ], + "angle": 0, + "content": "43. Somepalli, G., Singla, V., Goldblum, M., Geiping, J., Goldstein, T.: Diffusion art or digital forgery? investigating data replication in diffusion models. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.438, + 0.788, + 0.481 + ], + "angle": 0, + "content": "44. Somepalli, G., Singla, V., Goldblum, M., Geiping, J., Goldstein, T.: Diffusion art or digital forgery? investigating data replication in diffusion models. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.481, + 0.788, + 0.508 + ], + "angle": 0, + "content": "45. Song, J., Meng, C., Ermon, S.: Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.508, + 0.788, + 0.55 + ], + "angle": 0, + "content": "46. Tang, R., Liu, L., Pandey, A., Jiang, Z., Yang, G., Kumar, K., Stenetorp, P., Lin, J., Ture, F.: What the daam: Interpreting stable diffusion using cross attention. arXiv preprint arXiv:2210.04885 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.55, + 0.788, + 0.591 + ], + "angle": 0, + "content": "47. Tsai, Y.L., Hsu, C.Y., Xie, C., Lin, C.H., Chen, J.Y., Li, B., Chen, P.Y., Yu, C.M., Huang, C.Y.: Ring-a-bell! how reliable are concept removal methods for diffusion models? arXiv preprint arXiv:2310.10012 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.591, + 0.788, + 0.633 + ], + "angle": 0, + "content": "48. Wang, Z.J., Montoya, E., Munechika, D., Yang, H., Hoover, B., Chau, D.H.: Diffusiondb: A large-scale prompt gallery dataset for text-to-image generative models. arXiv preprint arXiv:2210.14896 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.633, + 0.788, + 0.66 + ], + "angle": 0, + "content": "49. Xiang, C., Bao, F., Li, C., Su, H., Zhu, J.: A closer look at parameter-efficient tuning in diffusion models. arXiv preprint arXiv:2303.18181 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.66, + 0.788, + 0.702 + ], + "angle": 0, + "content": "50. Yao, Y., Wang, P., Tian, B., Cheng, S., Li, Z., Deng, S., Chen, H., Zhang, N.: Editing large language models: Problems, methods, and opportunities. arXiv preprint arXiv:2305.13172 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.702, + 0.788, + 0.743 + ], + "angle": 0, + "content": "51. Ye, H., Zhang, J., Liu, S., Han, X., Yang, W.: Ip-adapter: Text compatible image prompt adapter for text-to-image diffusion models. arXiv preprint arXiv:2308.06721 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.743, + 0.788, + 0.771 + ], + "angle": 0, + "content": "52. Zhang, E., Wang, K., Xu, X., Wang, Z., Shi, H.: Forget-me-not: Learning to forget in text-to-image diffusion models. arXiv preprint arXiv:2303.17591 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.771, + 0.788, + 0.799 + ], + "angle": 0, + "content": "53. Zhang, L., Rao, A., Agrawala, M.: Adding conditional control to text-to-image diffusion models. In: ICCV (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.799, + 0.788, + 0.841 + ], + "angle": 0, + "content": "54. Zhang, Y., Jia, J., Chen, X., Chen, A., Zhang, Y., Liu, J., Ding, K., Liu, S.: To generate or not? safety-driven unlearned diffusion models are still easy to generate unsafe images... for now. arXiv preprint arXiv:2310.11868 (2023)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.356, + 0.129 + ], + "angle": 0, + "content": "Huang et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.19 + ], + "angle": 0, + "content": "55. Zhao, S., Chen, D., Chen, Y.C., Bao, J., Hao, S., Yuan, L., Wong, K.Y.K.: Unictrlnet: All-in-one control to text-to-image diffusion models. arXiv preprint arXiv:2305.16322 (2023)" + } + ] +] \ No newline at end of file diff --git a/2024/Receler_ Reliable Concept Erasing of Text-to-Image Diffusion Models via Lightweight Erasers/e96e787a-46f5-4d12-8c30-eca8de901ff9_origin.pdf b/2024/Receler_ Reliable Concept Erasing of Text-to-Image Diffusion Models via Lightweight Erasers/e96e787a-46f5-4d12-8c30-eca8de901ff9_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..5b2d4b186cc8d0e59d1e42b97e9a360ffd93d9c9 --- /dev/null +++ b/2024/Receler_ Reliable Concept Erasing of Text-to-Image Diffusion Models via Lightweight Erasers/e96e787a-46f5-4d12-8c30-eca8de901ff9_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d3ad84633c20204b94140f8433a0828315099be6418292caffe6cc2ee8a31b97 +size 2558818 diff --git a/2024/Receler_ Reliable Concept Erasing of Text-to-Image Diffusion Models via Lightweight Erasers/full.md b/2024/Receler_ Reliable Concept Erasing of Text-to-Image Diffusion Models via Lightweight Erasers/full.md new file mode 100644 index 0000000000000000000000000000000000000000..617ebf6f707a5d681f9b7563aee54a9ff454d5c5 --- /dev/null +++ b/2024/Receler_ Reliable Concept Erasing of Text-to-Image Diffusion Models via Lightweight Erasers/full.md @@ -0,0 +1,273 @@ +# Receler: Reliable Concept Erasing of Text-to-Image Diffusion Models via Lightweight Erasers + +Chi-Pin Huang\*1,\*, Kai-Po Chang\*1, Chung-Ting Tsai2, Yung-Hsuan Lai2, Fu-En Yang3, and Yu-Chiang Frank Wang\*1,3,\* + +1 Graduate Institute of Communication Engineering, National Taiwan University + +$^{2}$ National Taiwan University, $^{3}$ NVIDIA + +† f11942097@ntu.edu.tw, ‡ frankwang@nvidia.com + +Abstract. Concept erasure in text-to-image diffusion models aims to disable pre-trained diffusion models from generating images related to a target concept. To perform reliable concept erasure, the properties of robustness and locality are desirable. The former refrains the model from producing images associated with the target concept for any paraphrased or learned prompts, while the latter preserves its ability in generating images with non-target concepts. In this paper, we propose Reliable Concept Erasing via Lightweight Erasers (Receler). It learns a lightweight Eraser to perform concept erasing while satisfying the above desirable properties through the proposed concept-localized regularization and adversarial prompt learning scheme. Experiments with various concepts verify the superiority of Receler over previous methods. Code is available at https://github.com/jasper0314-huang/Receler. + +Keywords: Concept Erasing $\cdot$ Diffusion Models $\cdot$ Adversarial Learning + +# 1 Introduction + +In recent years, text-to-image generation models [4, 32, 34, 36, 38] have achieved unprecedented success in generating photo-realistic images which benefit various industrial applications [34, 38]. Despite their apparent success and convenience, these models may produce images that are deemed NSFW (Not Safe for Work) [21] or infringe upon intellectual property and portrait rights [1], e.g., generating nudity, violent content, or imitating the style of well-known artists. This issue is mainly due to the memorization of the large-scale training data sourced from the web [8, 43]. To address the above problem, an intuitive solution is to manually filter out the inappropriate images and re-train the model. However, as pointed out in [16], this may lead to unpredictable results such as exposing more inappropriate content to be memorized [7] or incomplete visual + +![](images/d57102dd2aa42d591ce0cbe6f0ce18cf90f5f02e07c5a13dea826be849246218.jpg) +Fig. 1: Illustration of (a) robustness and (b) locality-preserving in concept erasing. The former requires models to be robust against paraphrased attacks from the target concept (i.e., "nudity"), while the latter aims to preserve the visual content of non-target concepts (e.g., "tennis player" or "airplane"). Note that SD denotes Stable Diffusion [36]; UCE [16] and ESD [15] are recent works on concept erasing. + +concept removal [35]. Moreover, even though filtered image data can be collected, re-training generative models is still computationally expensive [15,23,42]. + +Instead of processing training data and re-training models, an alternative is to erase or unlearn specific concepts from the pre-trained model [15,16,23,52]. That is, given a concept described in text, the pre-trained model is fine-tuned to forget that concept so that the associated image content cannot be generated from the fine-tuned model. In practice, it would be desirable to perform concept erasing with sufficient reliability, which suggests two desirable properties: locality and robustness. Locality indicates the ability to preserve the model generalization in synthesizing content not associated with the target concept [31,42]. Robustness requires erased models to effectively remove the target concept [31,42], while not be circumvented by paraphrased prompts that aim to recover the target concept (e.g., "car" vs. "jeep"). While both locality and robustness have been recently discussed in the field of NLP [10,12,20,29,50], the developed techniques cannot be directly applied to text-to-image generative models for performing concept erasing. + +Recently, a number of methods for concept erasing or unlearning for diffusion models have been proposed [15, 16, 23, 52]. For example, Ablating [23] predefines an anchor concept for each target concept that needs to be unlearned and then achieves model unlearning by mapping the image distribution of the target concept to that of the anchor concept, e.g. mapping "grumpy cat" to "cat." Inspired by classifier-free guidance [19], ESD [15] fine-tunes the model to predict negatively guided noise. In other words, it decreases the probability of generating images of the target concept, thus unlearning that concept. FMN [52] designs a computationally efficient unlearning method by directly minimizing the cross-attention weights corresponding to the target concept in the input text prompt, + +encouraging the model to ignore the concept. UCE [16] employs a closed-form editing approach to optimize the projection matrices of keys and values in cross-attention layers, ensuring the model maintains locality when unlearning the target concept. + +Although promising progress has been made in erasing specific target concepts, most existing works are not specifically designed to preserve model locality and robustness. For example, Ablating [23] and ESD [15] fine-tune a considerable amount of parameters within pre-trained diffusion models to achieve concept erasure, which inevitably compromises the original capabilities of the model. On the other hand, methods such as UCE [16] and FMN [52] only modify specific parameters (i.e., the projection matrices of keys and values in cross-attention layers) responsible for encoding input textual features instead of visual ones. These methods would be vulnerable to rephrased target concepts since they only learn to dissociate textual prompts in the cross-attention layers (i.e., lack of ability to recognize that paraphrased queries are semantically similar to the erased target concept). For example, a diffusion model that has been erased of the concept of "car" may still produce images of jeeps due to its inability to recognize that jeeps fall under the category of cars; hence, it cannot remove the attention to jeep in cross-attention layers. As a result, proposing a concept-erasing method that addresses locality and robustness continues to pose a crucial challenge. + +In this paper, we propose Reliable Concept Erasing via Lightweight Erasers (Receler) for erasing concepts from pre-trained diffusion models, exhibiting sufficient locality and robustness properties. Receler involves a lightweight eraser (only $0.37\%$ of the U-Net parameters), which is designed to remove a target concept from the outputs of cross-attention layers. During this unlearning process, we train the eraser while preserving the image generation capability of the pre-trained diffusion models. Furthermore, a concept-localized regularization is introduced to ensure that the eraser focuses on erasing the target concept. This regularization prevents the generation of non-target concepts from being affected, thereby preserving locality. As for robustness, we advance adversarial prompt learning, which optimizes the adversarial prompts that induce the model to generate images of the target concept and then fine-tunes the eraser to erase images generated with these prompts. By training our eraser with concept-localized regularization and adversarial prompt learning, we are able to preserve the image generation capability of non-target concepts and robustly refrain the model from generating images with target-concept content. + +We now summarize the contributions of this work below: + +- We present Reliable Concept Erasing via Lightweight Erasers (Receler), a novel approach using a lightweight eraser (only $0.37\%$ of the U-Net parameters) for reliable and efficient concept erasing. +- Locality is introduced through concept-localized regularization, which constrains the eraser for precise erasing of the target concept without affecting the generation of non-target ones. +- Receler is trained against adversarial prompts, imitating paraphrased prompts of target concepts, resulting in improved robustness in concept erasure. + +# 2 Related Works + +# 2.1 Erasing Concepts from Diffusion Models + +Text-to-image diffusion models [4, 32, 34, 36, 38] have been shown to generate high-quality images with impressive generalization. However, such models are typically trained on extensive web-crawled data (e.g., LAION-5B [41]), which could memorize NSFW or copyrighted content, leading to the generation or replication of undesired images [9, 44]. To address this issue, some works explore the solution without the need to update the model weights. For instance, Stable Diffusion [36] employs an unsafe content classifier to filter out risky outputs, while SLD [39] uses negative guidance to prevent inappropriate content generation. However, the former relies on the reliability of pre-trained classifiers, while the latter only suppresses undesired concepts without complete removal. + +In response, several works focus on fine-tuning the diffusion model to erase the target concepts [15, 16, 23, 52]. Ablating [23] associates each concept to be erased, e.g., "grumpy cat," with a broader, predefined anchor concept, e.g., "cat" and fine-tunes the diffusion model to map the generated image of the target concept to that of the anchor concept by minimizing the L2 distance of predicted noises. Inspired by classifier-free guidance [19], ESD [15] proposes to decrease the likelihood of generating images belonging to the target concept. This is achieved by fine-tuning the diffusion model to predict negatively guided noises, effectively steering the model's conditional prediction away from the erased concept. FMN [52] adopts attention resteing, a computationally efficient unlearning method, to identify attention maps associated with the target concept in the diffusion U-Net's cross-attention layers. By minimizing the attention weights corresponding to the target concept, the diffusion model gradually disregards the target concept during image generation, facilitating the erasure of the concept. UCE [16] employs a closed-form editing method to optimize the projection matrices of keys and values in cross-attention. The objective is to align the embedding of a source prompt (e.g., "a photo of an airplane") more closely with that of a destination (e.g., an empty string), while leaving other unrelated concepts unchanged. Despite their effectiveness in erasing concepts, most current methods are not particularly designed to preserve robustness against paraphrased prompts (e.g., "nudity" vs. "without clothes"). Moreover, both [23] and [15] require fine-tuning a considerable number of model parameters, which might affect the model capability and consequently compromise the locality property. + +# 2.2 Controlling Text-to-Image Diffusion Models + +Parameter-Efficient Fine-Tuning (PEFT) is a training scheme that addresses the challenges of extensive parameter updates, especially for large language models. These approaches update only a small subset of parameters, thereby reducing the risk of compromising the pre-trained capabilities of the model. Recent researches [14,24,30,37,51,53,55] have applied PEFT to control text-to-image diffusion models. For instance, some studies [14,24,37] empower the model to learn + +personalized or unseen concepts by fine-tuning a new text token and a small number of parameters using few user-provided exemplar images. Meanwhile, other works [30, 51, 53, 55] aim to enable diffusion models to generate images based on additional conditions, e.g., edge maps, depth maps, or segmentation masks. They fine-tune lightweight task-specific modules with condition-image pairs to achieve control over the image generation process. Despite the effectiveness of these methods, they achieve learning new concepts by accessing the associated data of interest during training. As for concept erasing, since one only observes the description of the concept to be unlearned, existing PEFT-based methods cannot be directly applied. + +# 2.3 Adversarial Attack & Training + +In adversarial attacks [2, 13, 17, 25], adversarial examples are deliberately constructed inputs, which would deceive models into making incorrect predictions. Popular methods such as Fast Gradient Sign Method (FGSM [13]) and its variants (I-FGSM [13] and MI-FGSM [13]) targeted at attacking classification models, utilizing the resulting loss gradients to produce imperceptible perturbations and induce misclassification. In contrast, recent approaches [11, 47, 54] introduce prompt-based adversarial attacks tailored to provoke seemingly unlearned diffusion models into generating images of the unlearned concepts. For instance, P4D [11] learns prompts to reconstruct noise associated with the target concept in diffusion models, and Ring-A-Bell [47] extracts holistic concept representations from CLIP model [33] to generate model-agnostic attack prompts. These learned attack prompts have been shown to provoke unlearned models to regenerate images of the erased concept, posing potential issues in text-to-image generative models. + +To defend against such adversarial attacks, various adversarial training strategies have been proposed [3, 17, 28]. Such adversarial training schemes expose models to adversarial examples during training, enabling the models to learn and recognize these examples. Consequently, these adversially trained models can respond accurately when encountering adversarial examples during inference. Inspired by adversarial training approaches, we employ an adversarial erasing learning scheme to introduce additional robustness to the unlearned model. We will detail our proposed framework in the following section. + +# 3 Method + +Problem formulation. We first define the setting and notations of our Receler. Given a pre-trained text-to-image diffusion model, parameterized by $\theta$ , we aim to erase a textual concept $c$ from the model without requiring access to the corresponding image data. The erasure is considered successful when the model no longer generates images that contain or represent the concept $c$ (e.g., "nudity"-erased model should not generate any images with exposed body parts). + +![](images/75e522671b89ce6f7b06b60d99906beccdb442bb70123b6778b11acf04fe5179.jpg) +Fig. 2: Overview of Receler. (a) Receler involves iterative learning of a lightweight Eraser $E$ and adversarial prompt embedding $e_{Adv}$ . The former is trained to erase the target concept $c$ while preserving non-target concepts, and the latter learns to imitate the prompts to recover visual content associated with the concept previously erased. (b) The Eraser $E$ is inserted after each cross attention layer of Diffusion U-Net to remove the target concept from its outputs, with prediction $o^l$ directly added to the cross attention output. + +![](images/75e57cece7844947bb7b0a5404923736184dd37d6ea7795eb4a111171f4f5dd7.jpg) +(a) + +![](images/243f92659970b1ae4d644fca56012aca6b72b6146e9bc3d4b349f35a0eb7f45e.jpg) +(b) + +As depicted in Fig. 2, our method employs a lightweight Eraser $E$ , parameterized by $\theta_E$ , to learn to erase the target concept $c$ , as discussed in Sec. 3.1. To introduce the desirable locality and robustness to our model, we incorporate concept-localized regularization and adversarial prompt learning schemes into our framework, as detailed in Sec. 3.2 and Sec. 3.3. + +# 3.1 Concept Erasing with Lightweight Eraser + +In order to erase particular visual concepts from a pre-trained diffusion model, we introduce a lightweight adapter-based eraser. As depicted in Fig. 2, by solely finetuning the newly introduced $\theta_{E}$ , our goal is to unlearn the target visual concept while preserving model generalization on non-target concepts. To be specific, our eraser is designed to remove the target concept from the output visual features of each cross-attention layer within the diffusion U-Net. This is based on the fact that these layers are responsible for incorporating the input text concept into the visual features. Thus, our eraser is positioned subsequent to each of the cross-attention layers, which is in line with the empirical analysis presented in [49]. To erase the concept, the eraser is trained to predict the negatively guided noises [15, 19] that move the model's prediction away from the erased concept. The objective is defined as: + +$$ +\mathcal {L} _ {E r a s e} = \mathbb {E} _ {x _ {t}, t} \left[ \lVert \epsilon_ {\theta^ {\prime}} (x _ {t}, e _ {c}, t) - \epsilon_ {E} \rVert^ {2} \right], +$$ + +where $\epsilon_{E} = \epsilon_{\theta}(x_{t},t) - \eta [\epsilon_{\theta}(x_{t},e_{c},t) - \epsilon_{\theta}(x_{t},t)]$ + +(1) + +Note that $\theta' = \{\theta, \theta_E\}$ represents the parameters of the diffusion model plugged with eraser, $x_t \in \mathbb{R}^{W \times H \times d}$ is the denoised image at timestep $t$ sampled from $\theta'$ conditioned on $c$ , $e_c$ is the text embedding of concept $c$ , and $\epsilon_E$ is the negatively guided noises [15, 19] predicted by $\theta$ . By minimizing the L2 distance between $\epsilon_{\theta'}(x_t, e_c, t)$ and $\epsilon_E$ , the eraser learns to reduce the probability of the generated image $x$ belongs to the target concept $c$ , thus effectively erasing the concept. + +# 3.2 Concept-Localized Regularization for Erasing Locality + +As the first desirable property in concept erasing, locality refers to preserving the model's ability to synthesize content unrelated to the target, which is realized by enforcing the eraser to affect the image synthesis process only if the target concept is present. To achieve this, we introduce concept-localized regularization into our Receler by leveraging the spatial information associated with the target concept's text tokens to regularize the eraser outputs. Specifically, inspired by [6, 46], we obtain the binary target concept mask $M \in \mathbb{R}^{W/4 \times H/4}$ by thresholding the attention maps when predicting $\epsilon_{\theta}(x_t, e_c, t)$ as follows: + +$$ +M _ {i, j} = \left\{ \begin{array}{l l} 1, & \text {i f} \frac {1}{| \mathcal {S} |} \sum_ {s \in \mathcal {S}} A _ {i, j} ^ {s} \geq \tau , \\ 0, & \text {o t h e r w i s e .} \end{array} \right. \tag {2} +$$ + +where $\mathcal{S}$ is the set of indices of all U-Net's mid-layers with resolution $(\frac{W}{4},\frac{H}{4})$ , $A^s\in \mathbb{R}^{W / 4\times H / 4}$ is the cross-attention map in the $s$ -th layer of the text tokens corresponding to the concept $c$ , and $\tau$ is a pre-defined threshold. With $M$ obtained, we calculate the following regularization loss to regularize the outputs of the eraser as follows: + +$$ +\mathcal {L} _ {R e g} = \frac {1}{L} \sum_ {l = 1} ^ {L} \| o ^ {l} \odot (1 - \tilde {M}) \| ^ {2}, \tag {3} +$$ + +where $L$ is the number of U-Net's layers, $\odot$ is the element-wise product, $o^l \in \mathbb{R}^{w^l \times h^l \times d}$ is the output of the eraser in the $l$ -th layer with the resolution $(w^l, h^l)$ , and $d$ is the feature dimension. Note that $\tilde{M}$ is the shorthand of $M$ being bicubically upscaled to the same resolution as $o^l$ . With the above regularization introduced, our Receler is enforced to preserve the generation of non-target concepts. Thus, the diversity and fidelity of the original model can be maintained. + +# 3.3 Adversarial Prompt Learning for Erasing Robustness + +To further ensure our Receler being robust against prompting attacks (e.g., by paraphrased or learned prompts), we introduce an adversarial learning strategy into our framework to enrich such model robustness. In our framework, we present a unique adversarial prompt learning scheme, which learns prompting attacks that would induce the diffusion model to synthesize the images containing previously erased concepts. To achieve this, we design the adversarial loss, + +which optimizes the continuous soft prompts $e_{Adv}$ by encouraging such learned prompts to imitate the malicious prompts, as illustrated in Fig. 2. More precisely, the objective $\mathcal{L}_{Adv}$ is defined as follows: + +$$ +\mathcal {L} _ {A d v} = \mathbb {E} _ {x _ {t}, t} \left[ \| \epsilon_ {\theta^ {\prime}} (x _ {t}, e ^ {\prime}, t) - \epsilon_ {M} \| ^ {2} \right], \tag {4} +$$ + +where $e' = [e_c; e_{Adv}]$ is the concatenated prompts of the erased target concept embedding and the learned soft prompts, and $\epsilon_M = \epsilon_\theta(x_t, e_c, t)$ represents the malicious noise predicted by the pre-trained diffusion model (without safety mechanism) conditioned on the target concept. By minimizing this adversarial loss, $e_{Adv}$ learns to regenerate the unlearned concept from the erased model. The optimization of the soft prompt $e_{Adv}$ and the eraser $\theta_E$ is performed iteratively, with each being fixed while the other is trained. Thus, $\theta_E$ and $e_{Adv}$ are trained against each other to improve model robustness. For more details, please refer to the pseudo-algorithm in the supplementary material. + +# 4 Experiments + +In this section, we first conduct quantitative experiments to assess the robustness and locality of Receler compared to state-of-the-art baselines, followed by ablation studies of our method. Lastly, we present qualitative comparisons and visualizations to demonstrate its effectiveness. + +Datasets. We conduct experiments on erasing objects defined in the CIFAR-10 dataset [22] and on erasing inappropriate contents from the Inappropriate Image Prompts (I2P) dataset [39]: + +- Object Erasure. To evaluate the effectiveness of erasure methods in erasing common visual concepts, we choose to erase ten class labels from CIFAR-10 [22]. Note that during our experiment, we only utilize the label set, not the images. For comprehensive assessment, we devise two types of evaluation prompts for each class: Firstly, we use simple prompts formatted as "A photo of {class}" to evaluate the efficacy of erasure in removing the target concept. Secondly, to further assess the robustness, we generate 50 paraphrased prompts for each class using ChatGPT3 to simulate real-world scenarios where prompts are typically more complex and target concepts may not be explicitly mentioned. For example, a paraphrased prompt for "airplane" is "A sleek, black stealth bomber flying low over a desert landscape at dusk." More details can be found in the supplementary material. +- Inappropriate Content Erasure. The I2P dataset [39], collected from a text-to-image prompt dataset DiffusionDB [48], comprises 4,703 real-world, user-generated prompts that produce inappropriate images, including hate, harassment, violence, self-harm, shocking, sexual, and illegal content. + +Evaluation Setup. We assess the robustness and locality of the erasure methods for object erasure and inappropriate content erasure as follows: + +- Object Erasure. For each method, we fine-tune ten models, each erasing one CIFAR-10 class. Each model is then evaluated by: 1) Efficacy $(\mathrm{Acc}_E)$ : the percentage of the target class being erased when inputting simple prompts; lower values are better. 2) Robustness $(\mathrm{Acc}_R)$ : the percentage of the target class being erased when inputting paraphrased prompts; lower values are better. 3) Locality $(\mathrm{Acc}_L)$ : the percentage of non-target classes being preserved; higher values are better. To assess efficacy and robustness, we generate 150 images using simple and paraphrased prompts for the target class, respectively. For Locality, we generate 50 images for each of the nine non-target classes using paraphrased prompts. We then use GroundingDINO [27] to detect if the corresponding class is presenting in the image, thereby assessing $\mathrm{Acc}_E$ , $\mathrm{Acc}_R$ , and $\mathrm{Acc}_L$ . To further evaluate the overall performance, we calculate the harmonic mean (H) of $100 - \mathrm{Acc}_E$ , $100 - \mathrm{Acc}_R$ , and $\mathrm{Acc}_L$ . + +- Inappropriate Content Erasure. Following ESD [15], we fine-tune two models for each erasure method: one for "nudity" and the other for the predefined inappropriate concepts e.g. hate, harassment, and violence. We use the NudeNet detector [5] to detect nudity and both the NudeNet and the Q16 detector [40] to identify the inappropriate concepts. Model robustness is evaluated by using real-world prompts in I2P [39]. Locality is evaluated using COCO-30K [26], a nudity-free dataset, by employing the nudity-erased model to generate safe contents from COCO-30K prompts and evaluating the quality of the generated images in terms of FID [18] and CLIP [33]. + +Comparisons. We compare Receler to state-of-the-art erasing methods, including FMN [52], SLD [39], Ablating [23], ESD [15], and UCE [16]. For all methods, we use the open-sourced codebases and follow their reported settings. We fine-tune all models from SD v1.4 [36], and for all image generation, we employ DDIM sampler [45] over 50 steps and a guidance scale of 7.5. Following SD v1.4, the image resolution in all our experiments is $512 \times 512$ . Please refer to the supplementary for more experiment setup and implementation details. + +# 4.1 Quantitative Evaluation + +Object Erasure. In Tab. 1, we show that Receler surpasses previous state-of-the-art methods in erasing common visual concepts from CIFAR-10 [22] class labels. Notably, Receler achieves the highest harmonic mean $(H)$ and exceeds the second-best method by 11.2 points, highlighting its effectiveness in erasing concepts with sufficient robustness and locality. Specifically, when assessing method efficacy using the simple prompts $(\mathrm{Acc}_E)$ , Receler achieves an average accuracy of $14.9\%$ across the erased classes, $2.2\%$ better than the second-best method, ESD [15]. In addition, when evaluating method robustness with paraphrased prompts $(\mathrm{Acc}_R)$ , Receler reaches $17.6\%$ on average, outperforming ESD by $22.3\%$ in the erased classes. We evaluate the locality of the erased model by + +Table 1: Evaluation of erasing common objects in CIFAR-10 classes. $\mathrm{Acc}_E$ and $\mathrm{Acc}_R$ represent efficacy and robustness, indicating accuracy of target class (which should be minimized) on simple and paraphrased prompts, respectively. Locality, $\mathrm{Acc}_L$ , is accuracy of non-target classes (which should be maximized) using paraphrased prompts. The harmonic mean $H$ reflects overall assessment of $\mathrm{Acc}_E$ , $\mathrm{Acc}_R$ , and $\mathrm{Acc}_L$ . + +
MethodsMetricsErased conceptsavg.
air-planeauto-mobilebirdcatdeerdogfroghorseshiptruck
SD v1.4AccE89.399.393.396.799.398.796.097.395.396.096.1
AccR79.394.096.088.098.792.088.792.765.384.087.9
AccL88.887.287.087.986.787.487.887.390.488.387.9
FMN [52]AccE ↓93.397.390.092.098.095.384.795.388.794.792.9
AccR ↓80.796.793.370.795.386.767.395.360.784.083.1
AccL ↑88.088.286.087.684.488.086.485.890.488.987.4
H↑14.14.411.517.64.110.027.96.924.011.414.2
Ablating [23]AccE ↓78.074.776.793.392.797.394.7100.090.786.788.5
AccR ↓67.390.093.372.795.387.371.390.758.076.080.2
AccL ↑87.883.884.987.384.085.886.085.388.286.486.0
H↑34.319.814.715.28.36.512.80.021.023.420.1
ESD [15]AccE ↓20.044.011.314.019.320.013.38.716.04.717.1
AccR ↓33.381.354.018.040.727.338.741.332.032.739.9
AccL ↑83.679.872.971.868.070.079.368.286.779.175.9
H↑76.035.864.279.568.274.074.270.378.679.071.6
UCE [16]AccE ↓34.746.08.716.74.011.311.36.022.010.017.1
AccR ↓58.079.363.316.015.349.328.034.041.339.342.4
AccL ↑84.979.181.882.078.082.283.375.887.381.681.6
H↑58.937.859.583.185.669.580.777.072.675.372.0
Receler (Ours)AccE ↓10.046.73.311.32.76.723.37.324.014.014.9
AccR ↓16.762.026.70.72.04.717.36.020.719.317.6
AccL ↑88.481.382.280.476.774.783.880.488.284.782.1
H↑87.152.383.088.889.586.780.988.680.883.783.2
+ +examining its accuracy in the remaining classes $(\mathrm{Acc}_L)$ aside from the erased class, focusing on whether erasing one CIFAR-10 class affects the image synthesis of the other unrelated classes. Although FMN [52] and Ablating [23] exhibit high average $\mathrm{Acc}_L$ , appearing effective in preserving model locality, they struggle to erase the target objects, with only a $3.2\%$ and $7.6\%$ drop in $\mathrm{Acc}_E$ from SD, compared to our $81.2\%$ drop. + +Erasure of Inappropriate Content. In Tab. 2 and Tab. 3, we evaluate the robustness of erasing inappropriate content with real-world prompts from I2P dataset [39]. Compared to the second-best result, Receler stands out by achieving $4.3\%$ overall improvement in erasing sensitive concepts on I2P and a $3.2\%$ increase in erasing nudity content, underscoring its effectiveness in scenarios that require a safety mechanism. + +In Tab. 4, in addition to robustness, we assess locality using COCO-30K [26], a nudity-free dataset. We employ the "nudity"-erased model to produce safe + +Table 2: Quantitative results on Inappropriate Image Prompts (I2P) dataset. We follow SLD [39] and apply the ratio of inappropriate images as the metric. More results compared with other baselines are available in supplementary. + +
Class nameInappropriate proportion (%) (↓)
SDFMNSLDESDUCEReceler
Hate44.237.722.526.836.428.6
Harassment37.525.022.124.029.521.7
Violence46.347.831.835.134.127.1
Self-harm47.946.830.033.730.824.8
Sexual60.259.152.435.025.529.4
Shocking59.558.140.540.141.134.8
Illegal activity40.037.022.126.729.021.3
Overall48.947.833.732.831.327.0
+ +Table 4: Assessment of reliability of nudity-erased models. Robustness is evaluated using the nudity prompts from I2P dataset, and locality is assessed using COCO-30K prompts. + +
MethodRobustness Nudity-erased ratio(↑)Locality
CLIP-30K(↑)FID-30K(↓)
SD-31.3214.27
FMN44.2%30.3913.52
SLD71.6%30.9016.34
ESD81.3%30.2415.31
UCE75.9%30.8514.07
Receler84.5%31.0214.10
+ +Table 3: Quantitative results on nudity prompts from I2P dataset. We report the number of nudity images detected by the NudeNet [5]. F- and M- refer to female and male, respectively. + +
Class nameNumber of nudity detected (↓)
SDFMNSLDESDUCEReceler
Armpits1484246312939
Belly17011670206026
F-Breast26615539323513
M-Breast421730151212
Buttocks29123975
Feet635619242910
F-Genitalia18151151
M-Genitalia723749
Total743415211139179115
Erasing ratio%--44.2-71.6-81.3-75.9-84.5
+ +Table 5: Evaluation of robustness against learned attack prompts. We report the failure rate, indicating the proportion of generated images belonging to the unlearned concept. + +
MethodP4D [11]Ring-A-Bell [47]
cifar10 avg.nudityviolencenudity
FMN88.3%89.4%98.8%94.7%
Ablating85.4%82.8%100.0%96.8%
SLD60.5%56.3%80.4%86.3%
ESD48.1%54.3%86.0%55.8%
UCE53.8%51.9%76.8%49.5%
Receler13.7%31.2%59.2%1.1%
+ +content using COCO-30K prompts and evaluate the quality of the generated images using FID [18] and CLIP [33] metrics. As shown in the last two columns, Receler secures the highest CLIP-30K and nearly matches the top result on FID-30K. It is noteworthy that while Receler performs comparably to FMN in FID-30K, it surpasses FMN in robustness by erasing $40.3\%$ more nudity content. + +Learned Attack Prompts. To demonstrate the robustness of Receler in safeguarding against the potential and unprecedented malicious attacks, we employ P4D [11] and Ring-A-Bell [47]. These tools are specifically designed for red-teaming text-to-image models by finding problematic prompts. As illustrated in Tab. 5, Receler is significantly more reliable than other concept erasing methods. Specifically, it achieves lower failure rates $-34.4\%$ for CIFAR-10 and $20.7\%$ for nudity against P4D prompts, and $17.6\%$ for violence and $48.4\%$ for nudity against Ring-A-Bell prompts, compared to the second-best method. These results highlight the robustness of Receler in defending against malicious attacks. + +![](images/567d5cece8a2374a36a1fb5f19c59e553473057643b65bd0c0e97e879b49402d.jpg) +Fig. 3: Qualitative comparison of concept erasure methods. Note that erased concepts are listed at the top, and images generated from each method are shown in each row. Input prompts used for image generation are provided in supplementary. + +Ablation Study. In Tab. 6, we ablate the three proposed components in Receler: the lightweight eraser, concept-localized regularization, and adversarial prompt learning. We establish a simple baseline by fine-tuning the whole model with Eq. (1) (i.e., the first row). With the eraser introduced, both $\mathrm{Acc}_R$ and $\mathrm{Acc}_L$ improve. Adding concept-localized regularization further increases the $\mathrm{Acc}_L$ from $77.4\%$ to $79.8\%$ , demonstrating its effectiveness in enhancing locality. On the other hand, coupling adversarial prompting learning with the eraser boosts $\mathrm{Acc}_R$ by $14\%$ , albeit with a slight decrease in $\mathrm{Acc}_L$ . This result aligns with expectations, as adversarial prompt learning improves the robustness by restraining any possible malicious prompts. Therefore, Receler, which integrates both adversarial prompt learning and concept-localized regularization, yields the best experimental results. This implies that these two approaches benefit each other and enable the lightweight eraser to achieve both robustness and locality. + +# 4.2 Qualitative Evaluation + +Visualization of Erased Concepts with Paraphrased Prompts. In Fig. 3, we show examples of erasing different artist styles (e.g., van Gogh), objects (e.g., automobile) and high-level concepts (e.g., nudity). As observed in the figure, ESD [15] and UCE [16] fail to erase the target concept, with the outputs from these methods closely resembling the original images from SD [1]. On the contrary, Receler successfully erases all target concepts and is able to generate images that are visually close to the original ones, e.g., same background with the car removed, a human in the same posture with nudity removed. + +In Fig. 4, we qualitatively validate the robustness and locality of Receler. The diagonal orange boxes show its robustness against paraphrased prompts where the erased concept is not explicitly mentioned. For instance, airplane-erased Receler successfully prevents the generation of an "airplane" image when + +![](images/9e7ccc830afb3aca565f969c79f1383d247b1eb19d85133a02f242e754a93124.jpg) +Fig. 4: Visualization of robustness and locality from Receler on CIFAR-10. + +The red strikethrough at the top indicates the erased concepts. On the left, the input paraphrased prompts are provided. Images enclosed within the diagonal orange borders shows robustness while others show the locality. + +![](images/406d958736678147da0a6e2d2fdbe48445ada524f875ea931ce8133084ccb08b.jpg) +(a) nudity + +![](images/dc9a7319f4a7dadc2a2f807de0af117bdb1ed0591aa98df7796317c05fcb9437.jpg) +(b) violence +Fig. 5: Visualization of erasure methods against learned attack prompts. We use Ring-A-Bell [47] to generate adversarial prompts for nudity and violence concepts. + +the prompt is paraphrased to "jet". Concerning locality, as shown in the non-diagonal boxes, Receler generates images that not only faithfully adhere to the descriptions but also closely resemble the images from the original diffusion model. Notably, it can be seen that Receler is able to replace the unlearned objects with reasonable alternatives rather than simply removing them, e.g., two white puppies are substituted with two people wearing white furry clothes and gloves (see the fourth row in Fig. 4). + +Table 6: Ablation study on CIFAR-10. We ablate the components of Receler and report the robustness and locality metrics. The first row refers to fine-tuning all parameters with only $\mathcal{L}_{\text{Erase}}$ in Eq. (1). + +
ComponentsMetrics
Eraser\( \mathcal{L}_{\text{Reg}} \)\( \mathcal{L}_{\text{Adv}} \)\( \text{Acc}_R(\downarrow) \)\( \text{Acc}_L(\uparrow) \)
XXX39.475.2
✓XX34.977.4
✓✓X26.379.8
✓X✓20.976.2
✓✓✓17.682.1
+ +![](images/4a8a0b671d733ca8446d6c0e3ccc63968bb8ba7e63e6715a3ee7b625abeb5702.jpg) +Fig.6: Examples of erasing multiple concepts. Instead of training an eraser for multiple concepts from scratch, we combine existing erasers for multi-concept erasure. + +![](images/66a508933f86f395f8fbaa82c3981e47b4928dd161b9ac48c7dfb776c92b8cc6.jpg) + +![](images/132671481899a7f4e16153ac210d8d2be4ae2e0079e1fab1dae41192398d50d5.jpg) + +Visualization of Robustness against Learned Attack Prompts. In addition to the quantitative results shown in Tab. 5, we further qualitatively demonstrate the robustness of Receler against learned attack prompts in Fig. 5. The attack prompts are learned to induce the concepts of "nudity" and "violence" from models that should have erased these concepts using Ring-A-Bell [47]. For both nudity and violence attack prompts, Receler successfully prevents the generation of erased concepts, whereas other methods like ESD and UCE fail and generate images with the supposedly forbidden concepts (e.g., nudity or blood). + +Compositional Concept Erasure. In Fig. 6, we showcase examples of Receler in performing compositional concept erasure. This is achieved by combining outputs from separately trained erasers, each targeting a specific unlearned concept, during inference. By averaging these outputs, compositional concept erasure is accomplished without necessitating retraining. This approach notably offers the flexibility to selectively determine which concepts are to be erased, as required. + +# 5 Conclusion + +In this paper, we proposed Reliable Concept Erasing via Lightweight Erasers (Receler) to erase target concepts entirely from the pre-trained diffusion model against prompting attacks (i.e., robustness), while preserving its image generation ability of other concepts (i.e., locality). In Receler, we employ concept-localized regularization to enforce the eraser to only affect the visual features related to the target concept. To enhance model robustness to paraphrased or attack prompts, we present an adversarial prompt learning scheme to induce the model to produce images of previously erased concepts, followed by optimizing the model against such image generation. We conducted extensive quantitative and qualitative evaluations on Receler, validating its superior robustness and locality-preserving ability over previous concept-erasing methods. + +Acknowledgements. This research was supported in part by the National Science and Technology Council via grant NSTC 112-2634-F-002-007, NSTC 113-2640-E-002-003 and the Featured Area Research Center Program within the framework of the Higher Education Sprout Project by the Ministry of Education 113L900902. We also thank the National Center for High-performance Computing (NCHC) for providing computational and storage resources. + +# References + +1. Sarah andersen. et al v. stability ai ltd. et al. case no.3:2023cv00201. us district court for the northern district of california. (2023) +2. Akhtar, N., Mian, A., Kardan, N., Shah, M.: Advances in adversarial attacks and defenses in computer vision: A survey. IEEE Access 9, 155161-155196 (2021) +3. Bai, T., Luo, J., Zhao, J., Wen, B., Wang, Q.: Recent advances in adversarial training for adversarial robustness. arXiv preprint arXiv:2102.01356 (2021) +4. Balaji, Y., Nah, S., Huang, X., Vahdat, A., Song, J., Kreis, K., Aittala, M., Aila, T., Laine, S., Catanzaro, B., et al.: edifi: Text-to-image diffusion models with an ensemble of expert denoisers. arXiv preprint arXiv:2211.01324 (2022) +5. Bedapudi Praneeth, b.k., lireza Ayinmehr: Nudenet: Neural nets for nudity classification, detection and selective censoring (2019) +6. Cao, M., Wang, X., Qi, Z., Shan, Y., Qie, X., Zheng, Y.: Masactrl: Tuning-free mutual self-attention control for consistent image synthesis and editing. arXiv preprint arXiv:2304.08465 (2023) +7. Carlini, N., Jagielski, M., Zhang, C., Papernot, N., Terzis, A., Tramer, F.: The privacy onion effect: Memorization is relative. In: NeurIPS (2022) +8. Carlini, N., Hayes, J., Nasr, M., Jagielski, M., Sehwag, V., Tramer, F., Balle, B., Ippolito, D., Wallace, E.: Extracting training data from diffusion models. In: 32nd USENIX Security Symposium (USENIX Security 23). pp. 5253-5270 (2023) +9. Carlini, N., Hayes, J., Nasr, M., Jagielski, M., Sehwag, V., Tramer, F., Balle, B., Ippolito, D., Wallace, E.: Extracting training data from diffusion models. In: 32nd USENIX Security Symposium (USENIX Security 23). pp. 5253-5270 (2023) +0. Chen, J., Yang, D.: Unlearn what you want to forget: Efficient unlearning for LLMs. In: Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing. Association for Computational Linguistics (Dec 2023) +1. Chin, Z.Y., Jiang, C.M., Huang, C.C., Chen, P.Y., Chiu, W.C.: Prompting4debugging: Red-teaming text-to-image diffusion models by finding problematic prompts. arXiv preprint arXiv:2309.06135 (2023) +2. De Cao, N., Aziz, W., Titov, I.: Editing factual knowledge in language models. arXiv preprint arXiv:2104.08164 (2021) +3. Dong, Y., Liao, F., Pang, T., Su, H., Zhu, J., Hu, X., Li, J.: Boosting adversarial attacks with momentum. In: CVPR (2018) +4. Gal, R., Alaluf, Y., Atzmon, Y., Patashnik, O., Bermano, A.H., Chechik, G., Cohen-Or, D.: An image is worth one word: Personalizing text-to-image generation using textual inversion. arXiv preprint arXiv:2208.01618 (2022) +5. Gandikota, R., Materzynska, J., Fiotto-Kaufman, J., Bau, D.: Erasing concepts from diffusion models. arXiv preprint arXiv:2303.07345 (2023) +6. Gandikota, R., Orgad, H., Belinkov, Y., Materzyńska, J., Bau, D.: Unified concept editing in diffusion models. arXiv preprint arXiv:2308.14761 (2023) + +17. Goodfellow, I.J., Shlens, J., Szegedy, C.: Explaining and harnessing adversarial examples. arXiv preprint arXiv:1412.6572 (2014) +18. Heusel, M., Ramsauer, H., Unterthiner, T., Nessler, B., Hochreiter, S.: Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems 30 (2017) +19. Ho, J., Salimans, T.: Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598 (2022) +20. Huang, Z., Shen, Y., Zhang, X., Zhou, J., Rong, W., Xiong, Z.: Transformer-patcher: One mistake worth one neuron. arXiv preprint arXiv:2301.09785 (2023) +21. Hunter, T.: Ai porn is easy to make now. for women, that's a nightmare. The Washington Post pp. NA-NA (2023) +22. Krizhevsky, A., Hinton, G.: Learning multiple layers of features from tiny images. Master's thesis, Department of Computer Science, University of Toronto (2009) +23. Kumari, N., Zhang, B., Wang, S.Y., Shechtman, E., Zhang, R., Zhu, J.Y.: Ablating concepts in text-to-image diffusion models. In: ICCV (2023) +24. Kumari, N., Zhang, B., Zhang, R., Shechtman, E., Zhu, J.Y.: Multi-concept customization of text-to-image diffusion. In: CVPR (2023) +25. Kurakin, A., Goodfellow, I.J., Bengio, S.: Adversarial examples in the physical world. In: Artificial intelligence safety and security, pp. 99-112. Chapman and Hall/CRC (2018) +26. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollar, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13. pp. 740-755. Springer (2014) +27. Liu, S., Zeng, Z., Ren, T., Li, F., Zhang, H., Yang, J., Li, C., Yang, J., Su, H., Zhu, J., et al.: Grounding dino: Marrying dino with grounded pre-training for open-set object detection. arXiv preprint arXiv:2303.05499 (2023) +28. Madry, A., Makelov, A., Schmidt, L., Tsipras, D., Vladu, A.: Towards deep learning models resistant to adversarial attacks. arXiv preprint arXiv:1706.06083 (2017) +29. Meng, K., Bau, D., Andonian, A., Belinkov, Y.: Locating and editing factual associations in gpt. In: NeurIPS (2022) +30. Mou, C., Wang, X., Xie, L., Zhang, J., Qi, Z., Shan, Y., Qie, X.: T2i-adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models. arXiv preprint arXiv:2302.08453 (2023) +31. Ni, M., Wu, C., Wang, X., Yin, S., Wang, L., Liu, Z., Duan, N.: Orens: Open-vocabulary responsible visual synthesis. arXiv preprint arXiv:2308.13785 (2023) +32. Nichol, A., Dhariwal, P., Ramesh, A., Shyam, P., Mishkin, P., McGrew, B., Sutskever, I., Chen, M.: Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741 (2021) +33. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021) +34. Ramesh, A., Dhariwal, P., Nichol, A., Chu, C., Chen, M.: Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125 1(2), 3 (2022) +35. Rombach, R.: Stable diffusion 2.0 release (Nov 2022) +36. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models (2021) + +37. Ruiz, N., Li, Y., Jampani, V., Pritch, Y., Rubinstein, M., Aberman, K.: Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In: CVPR (2023) +38. Sahara, C., Chan, W., Saxena, S., Li, L., Whang, J., Denton, E.L., Ghasemipour, K., Gontijo Lopes, R., Karagol Ayan, B., Salimans, T., et al.: Photorealistic text-to-image diffusion models with deep language understanding. In: NeurIPS (2022) +39. Schramowski, P., Brack, M., Deiseroth, B., Kersting, K.: Safe latent diffusion: Mitigating inappropriate degeneration in diffusion models. In: CVPR (2023) +40. Schramowski, P., Tauchmann, C., Kersting, K.: Can machines help us answering question 16 in datasheets, and in turn reflecting on inappropriate content? In: Proceedings of the 2022 ACM Conference on Fairness, Accountability, and Transparency. pp. 1350-1361 (2022) +41. Schuhmann, C., Beaumont, R., Vencu, R., Gordon, C., Wightman, R., Cherti, M., Coombes, T., Katta, A., Mullis, C., Wortsman, M., et al.: Laion-5b: An open large-scale dataset for training next generation image-text models. In: NeurIPS (2022) +42. Sinitsin, A., Plokhotnyuk, V., Pyrkin, D., Popov, S., Babenko, A.: Editable neural networks. arXiv preprint arXiv:2004.00345 (2020) +43. Somepalli, G., Singla, V., Goldblum, M., Geiping, J., Goldstein, T.: Diffusion art or digital forgery? investigating data replication in diffusion models. In: CVPR (2023) +44. Somepalli, G., Singla, V., Goldblum, M., Geiping, J., Goldstein, T.: Diffusion art or digital forgery? investigating data replication in diffusion models. In: CVPR (2023) +45. Song, J., Meng, C., Ermon, S.: Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502 (2020) +46. Tang, R., Liu, L., Pandey, A., Jiang, Z., Yang, G., Kumar, K., Stenetorp, P., Lin, J., Ture, F.: What the daam: Interpreting stable diffusion using cross attention. arXiv preprint arXiv:2210.04885 (2022) +47. Tsai, Y.L., Hsu, C.Y., Xie, C., Lin, C.H., Chen, J.Y., Li, B., Chen, P.Y., Yu, C.M., Huang, C.Y.: Ring-a-bell! how reliable are concept removal methods for diffusion models? arXiv preprint arXiv:2310.10012 (2023) +48. Wang, Z.J., Montoya, E., Munechika, D., Yang, H., Hoover, B., Chau, D.H.: Diffusiondb: A large-scale prompt gallery dataset for text-to-image generative models. arXiv preprint arXiv:2210.14896 (2022) +49. Xiang, C., Bao, F., Li, C., Su, H., Zhu, J.: A closer look at parameter-efficient tuning in diffusion models. arXiv preprint arXiv:2303.18181 (2023) +50. Yao, Y., Wang, P., Tian, B., Cheng, S., Li, Z., Deng, S., Chen, H., Zhang, N.: Editing large language models: Problems, methods, and opportunities. arXiv preprint arXiv:2305.13172 (2023) +51. Ye, H., Zhang, J., Liu, S., Han, X., Yang, W.: Ip-adapter: Text compatible image prompt adapter for text-to-image diffusion models. arXiv preprint arXiv:2308.06721 (2023) +52. Zhang, E., Wang, K., Xu, X., Wang, Z., Shi, H.: Forget-me-not: Learning to forget in text-to-image diffusion models. arXiv preprint arXiv:2303.17591 (2023) +53. Zhang, L., Rao, A., Agrawala, M.: Adding conditional control to text-to-image diffusion models. In: ICCV (2023) +54. Zhang, Y., Jia, J., Chen, X., Chen, A., Zhang, Y., Liu, J., Ding, K., Liu, S.: To generate or not? safety-driven unlearned diffusion models are still easy to generate unsafe images... for now. arXiv preprint arXiv:2310.11868 (2023) + +55. Zhao, S., Chen, D., Chen, Y.C., Bao, J., Hao, S., Yuan, L., Wong, K.Y.K.: Unictrlnet: All-in-one control to text-to-image diffusion models. arXiv preprint arXiv:2305.16322 (2023) \ No newline at end of file diff --git a/2024/Receler_ Reliable Concept Erasing of Text-to-Image Diffusion Models via Lightweight Erasers/images.zip b/2024/Receler_ Reliable Concept Erasing of Text-to-Image Diffusion Models via Lightweight Erasers/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..100d509528c9d4169ed05e066e68841d866404ce --- /dev/null +++ b/2024/Receler_ Reliable Concept Erasing of Text-to-Image Diffusion Models via Lightweight Erasers/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e6fc737faf4ab4d45b5788f8a86d0fe1ccda441aabc5817617de195a37ac378f +size 696141 diff --git a/2024/Receler_ Reliable Concept Erasing of Text-to-Image Diffusion Models via Lightweight Erasers/layout.json b/2024/Receler_ Reliable Concept Erasing of Text-to-Image Diffusion Models via Lightweight Erasers/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..c7d8c0342b020066c55e3a9b7a2e5bf0ce31e03a --- /dev/null +++ b/2024/Receler_ Reliable Concept Erasing of Text-to-Image Diffusion Models via Lightweight Erasers/layout.json @@ -0,0 +1,9201 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 179, + 111, + 438, + 166 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 179, + 111, + 438, + 166 + ], + "spans": [ + { + "bbox": [ + 179, + 111, + 438, + 166 + ], + "type": "text", + "content": "Receler: Reliable Concept Erasing of Text-to-Image Diffusion Models via Lightweight Erasers" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 153, + 186, + 461, + 212 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 153, + 186, + 461, + 212 + ], + "spans": [ + { + "bbox": [ + 153, + 186, + 461, + 212 + ], + "type": "text", + "content": "Chi-Pin Huang\\*1,\\*, Kai-Po Chang\\*1, Chung-Ting Tsai2, Yung-Hsuan Lai2, Fu-En Yang3, and Yu-Chiang Frank Wang\\*1,3,\\*" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 140, + 219, + 473, + 230 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 140, + 219, + 473, + 230 + ], + "spans": [ + { + "bbox": [ + 140, + 219, + 473, + 230 + ], + "type": "text", + "content": "1 Graduate Institute of Communication Engineering, National Taiwan University" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 222, + 230, + 392, + 242 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 222, + 230, + 392, + 242 + ], + "spans": [ + { + "bbox": [ + 222, + 230, + 392, + 242 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 222, + 230, + 392, + 242 + ], + "type": "text", + "content": " National Taiwan University, " + }, + { + "bbox": [ + 222, + 230, + 392, + 242 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 222, + 230, + 392, + 242 + ], + "type": "text", + "content": " NVIDIA" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 201, + 242, + 412, + 253 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 201, + 242, + 412, + 253 + ], + "spans": [ + { + "bbox": [ + 201, + 242, + 412, + 253 + ], + "type": "text", + "content": "† f11942097@ntu.edu.tw, ‡ frankwang@nvidia.com" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 159, + 284, + 455, + 427 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 284, + 455, + 427 + ], + "spans": [ + { + "bbox": [ + 159, + 284, + 455, + 427 + ], + "type": "text", + "content": "Abstract. Concept erasure in text-to-image diffusion models aims to disable pre-trained diffusion models from generating images related to a target concept. To perform reliable concept erasure, the properties of robustness and locality are desirable. The former refrains the model from producing images associated with the target concept for any paraphrased or learned prompts, while the latter preserves its ability in generating images with non-target concepts. In this paper, we propose Reliable Concept Erasing via Lightweight Erasers (Receler). It learns a lightweight Eraser to perform concept erasing while satisfying the above desirable properties through the proposed concept-localized regularization and adversarial prompt learning scheme. Experiments with various concepts verify the superiority of Receler over previous methods. Code is available at https://github.com/jasper0314-huang/Receler." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 160, + 437, + 454, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 437, + 454, + 449 + ], + "spans": [ + { + "bbox": [ + 160, + 437, + 454, + 449 + ], + "type": "text", + "content": "Keywords: Concept Erasing " + }, + { + "bbox": [ + 160, + 437, + 454, + 449 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 437, + 454, + 449 + ], + "type": "text", + "content": " Diffusion Models " + }, + { + "bbox": [ + 160, + 437, + 454, + 449 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 437, + 454, + 449 + ], + "type": "text", + "content": " Adversarial Learning" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 488, + 231, + 502 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 488, + 231, + 502 + ], + "spans": [ + { + "bbox": [ + 132, + 488, + 231, + 502 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 514, + 482, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 514, + 482, + 647 + ], + "spans": [ + { + "bbox": [ + 130, + 514, + 482, + 647 + ], + "type": "text", + "content": "In recent years, text-to-image generation models [4, 32, 34, 36, 38] have achieved unprecedented success in generating photo-realistic images which benefit various industrial applications [34, 38]. Despite their apparent success and convenience, these models may produce images that are deemed NSFW (Not Safe for Work) [21] or infringe upon intellectual property and portrait rights [1], e.g., generating nudity, violent content, or imitating the style of well-known artists. This issue is mainly due to the memorization of the large-scale training data sourced from the web [8, 43]. To address the above problem, an intuitive solution is to manually filter out the inappropriate images and re-train the model. However, as pointed out in [16], this may lead to unpredictable results such as exposing more inappropriate content to be memorized [7] or incomplete visual" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 145, + 654, + 233, + 666 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 654, + 233, + 666 + ], + "spans": [ + { + "bbox": [ + 145, + 654, + 233, + 666 + ], + "type": "text", + "content": "*Equal Contribution" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 138, + 114, + 476, + 257 + ], + "blocks": [ + { + "bbox": [ + 138, + 114, + 476, + 257 + ], + "lines": [ + { + "bbox": [ + 138, + 114, + 476, + 257 + ], + "spans": [ + { + "bbox": [ + 138, + 114, + 476, + 257 + ], + "type": "image", + "image_path": "d57102dd2aa42d591ce0cbe6f0ce18cf90f5f02e07c5a13dea826be849246218.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 265, + 482, + 321 + ], + "lines": [ + { + "bbox": [ + 130, + 265, + 482, + 321 + ], + "spans": [ + { + "bbox": [ + 130, + 265, + 482, + 321 + ], + "type": "text", + "content": "Fig. 1: Illustration of (a) robustness and (b) locality-preserving in concept erasing. The former requires models to be robust against paraphrased attacks from the target concept (i.e., \"nudity\"), while the latter aims to preserve the visual content of non-target concepts (e.g., \"tennis player\" or \"airplane\"). Note that SD denotes Stable Diffusion [36]; UCE [16] and ESD [15] are recent works on concept erasing." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 349, + 479, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 349, + 479, + 373 + ], + "spans": [ + { + "bbox": [ + 130, + 349, + 479, + 373 + ], + "type": "text", + "content": "concept removal [35]. Moreover, even though filtered image data can be collected, re-training generative models is still computationally expensive [15,23,42]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 376, + 482, + 543 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 376, + 482, + 543 + ], + "spans": [ + { + "bbox": [ + 130, + 376, + 482, + 543 + ], + "type": "text", + "content": "Instead of processing training data and re-training models, an alternative is to erase or unlearn specific concepts from the pre-trained model [15,16,23,52]. That is, given a concept described in text, the pre-trained model is fine-tuned to forget that concept so that the associated image content cannot be generated from the fine-tuned model. In practice, it would be desirable to perform concept erasing with sufficient reliability, which suggests two desirable properties: locality and robustness. Locality indicates the ability to preserve the model generalization in synthesizing content not associated with the target concept [31,42]. Robustness requires erased models to effectively remove the target concept [31,42], while not be circumvented by paraphrased prompts that aim to recover the target concept (e.g., \"car\" vs. \"jeep\"). While both locality and robustness have been recently discussed in the field of NLP [10,12,20,29,50], the developed techniques cannot be directly applied to text-to-image generative models for performing concept erasing." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "type": "text", + "content": "Recently, a number of methods for concept erasing or unlearning for diffusion models have been proposed [15, 16, 23, 52]. For example, Ablating [23] predefines an anchor concept for each target concept that needs to be unlearned and then achieves model unlearning by mapping the image distribution of the target concept to that of the anchor concept, e.g. mapping \"grumpy cat\" to \"cat.\" Inspired by classifier-free guidance [19], ESD [15] fine-tunes the model to predict negatively guided noise. In other words, it decreases the probability of generating images of the target concept, thus unlearning that concept. FMN [52] designs a computationally efficient unlearning method by directly minimizing the cross-attention weights corresponding to the target concept in the input text prompt," + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "type": "text", + "content": "Huang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 480, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 480, + 163 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 480, + 163 + ], + "type": "text", + "content": "encouraging the model to ignore the concept. UCE [16] employs a closed-form editing approach to optimize the projection matrices of keys and values in cross-attention layers, ensuring the model maintains locality when unlearning the target concept." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 164, + 482, + 355 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 164, + 482, + 355 + ], + "spans": [ + { + "bbox": [ + 130, + 164, + 482, + 355 + ], + "type": "text", + "content": "Although promising progress has been made in erasing specific target concepts, most existing works are not specifically designed to preserve model locality and robustness. For example, Ablating [23] and ESD [15] fine-tune a considerable amount of parameters within pre-trained diffusion models to achieve concept erasure, which inevitably compromises the original capabilities of the model. On the other hand, methods such as UCE [16] and FMN [52] only modify specific parameters (i.e., the projection matrices of keys and values in cross-attention layers) responsible for encoding input textual features instead of visual ones. These methods would be vulnerable to rephrased target concepts since they only learn to dissociate textual prompts in the cross-attention layers (i.e., lack of ability to recognize that paraphrased queries are semantically similar to the erased target concept). For example, a diffusion model that has been erased of the concept of \"car\" may still produce images of jeeps due to its inability to recognize that jeeps fall under the category of cars; hence, it cannot remove the attention to jeep in cross-attention layers. As a result, proposing a concept-erasing method that addresses locality and robustness continues to pose a crucial challenge." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 356, + 482, + 547 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 356, + 482, + 547 + ], + "spans": [ + { + "bbox": [ + 130, + 356, + 482, + 547 + ], + "type": "text", + "content": "In this paper, we propose Reliable Concept Erasing via Lightweight Erasers (Receler) for erasing concepts from pre-trained diffusion models, exhibiting sufficient locality and robustness properties. Receler involves a lightweight eraser (only " + }, + { + "bbox": [ + 130, + 356, + 482, + 547 + ], + "type": "inline_equation", + "content": "0.37\\%" + }, + { + "bbox": [ + 130, + 356, + 482, + 547 + ], + "type": "text", + "content": " of the U-Net parameters), which is designed to remove a target concept from the outputs of cross-attention layers. During this unlearning process, we train the eraser while preserving the image generation capability of the pre-trained diffusion models. Furthermore, a concept-localized regularization is introduced to ensure that the eraser focuses on erasing the target concept. This regularization prevents the generation of non-target concepts from being affected, thereby preserving locality. As for robustness, we advance adversarial prompt learning, which optimizes the adversarial prompts that induce the model to generate images of the target concept and then fine-tunes the eraser to erase images generated with these prompts. By training our eraser with concept-localized regularization and adversarial prompt learning, we are able to preserve the image generation capability of non-target concepts and robustly refrain the model from generating images with target-concept content." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 146, + 547, + 398, + 558 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 547, + 398, + 558 + ], + "spans": [ + { + "bbox": [ + 146, + 547, + 398, + 558 + ], + "type": "text", + "content": "We now summarize the contributions of this work below:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 136, + 564, + 484, + 665 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 136, + 564, + 481, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 564, + 481, + 601 + ], + "spans": [ + { + "bbox": [ + 136, + 564, + 481, + 601 + ], + "type": "text", + "content": "- We present Reliable Concept Erasing via Lightweight Erasers (Receler), a novel approach using a lightweight eraser (only " + }, + { + "bbox": [ + 136, + 564, + 481, + 601 + ], + "type": "inline_equation", + "content": "0.37\\%" + }, + { + "bbox": [ + 136, + 564, + 481, + 601 + ], + "type": "text", + "content": " of the U-Net parameters) for reliable and efficient concept erasing." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 136, + 602, + 481, + 638 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 602, + 481, + 638 + ], + "spans": [ + { + "bbox": [ + 136, + 602, + 481, + 638 + ], + "type": "text", + "content": "- Locality is introduced through concept-localized regularization, which constrains the eraser for precise erasing of the target concept without affecting the generation of non-target ones." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 136, + 641, + 484, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 641, + 484, + 665 + ], + "spans": [ + { + "bbox": [ + 136, + 641, + 484, + 665 + ], + "type": "text", + "content": "- Receler is trained against adversarial prompts, imitating paraphrased prompts of target concepts, resulting in improved robustness in concept erasure." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 416, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 416, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 416, + 91, + 447, + 100 + ], + "type": "text", + "content": "Receler" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 243, + 127 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 243, + 127 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 243, + 127 + ], + "type": "text", + "content": "2 Related Works" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 140, + 365, + 152 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 140, + 365, + 152 + ], + "spans": [ + { + "bbox": [ + 132, + 140, + 365, + 152 + ], + "type": "text", + "content": "2.1 Erasing Concepts from Diffusion Models" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 160, + 482, + 280 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 160, + 482, + 280 + ], + "spans": [ + { + "bbox": [ + 130, + 160, + 482, + 280 + ], + "type": "text", + "content": "Text-to-image diffusion models [4, 32, 34, 36, 38] have been shown to generate high-quality images with impressive generalization. However, such models are typically trained on extensive web-crawled data (e.g., LAION-5B [41]), which could memorize NSFW or copyrighted content, leading to the generation or replication of undesired images [9, 44]. To address this issue, some works explore the solution without the need to update the model weights. For instance, Stable Diffusion [36] employs an unsafe content classifier to filter out risky outputs, while SLD [39] uses negative guidance to prevent inappropriate content generation. However, the former relies on the reliability of pre-trained classifiers, while the latter only suppresses undesired concepts without complete removal." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 281, + 482, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 281, + 482, + 557 + ], + "spans": [ + { + "bbox": [ + 130, + 281, + 482, + 557 + ], + "type": "text", + "content": "In response, several works focus on fine-tuning the diffusion model to erase the target concepts [15, 16, 23, 52]. Ablating [23] associates each concept to be erased, e.g., \"grumpy cat,\" with a broader, predefined anchor concept, e.g., \"cat\" and fine-tunes the diffusion model to map the generated image of the target concept to that of the anchor concept by minimizing the L2 distance of predicted noises. Inspired by classifier-free guidance [19], ESD [15] proposes to decrease the likelihood of generating images belonging to the target concept. This is achieved by fine-tuning the diffusion model to predict negatively guided noises, effectively steering the model's conditional prediction away from the erased concept. FMN [52] adopts attention resteing, a computationally efficient unlearning method, to identify attention maps associated with the target concept in the diffusion U-Net's cross-attention layers. By minimizing the attention weights corresponding to the target concept, the diffusion model gradually disregards the target concept during image generation, facilitating the erasure of the concept. UCE [16] employs a closed-form editing method to optimize the projection matrices of keys and values in cross-attention. The objective is to align the embedding of a source prompt (e.g., \"a photo of an airplane\") more closely with that of a destination (e.g., an empty string), while leaving other unrelated concepts unchanged. Despite their effectiveness in erasing concepts, most current methods are not particularly designed to preserve robustness against paraphrased prompts (e.g., \"nudity\" vs. \"without clothes\"). Moreover, both [23] and [15] require fine-tuning a considerable number of model parameters, which might affect the model capability and consequently compromise the locality property." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 573, + 382, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 573, + 382, + 586 + ], + "spans": [ + { + "bbox": [ + 132, + 573, + 382, + 586 + ], + "type": "text", + "content": "2.2 Controlling Text-to-Image Diffusion Models" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": "Parameter-Efficient Fine-Tuning (PEFT) is a training scheme that addresses the challenges of extensive parameter updates, especially for large language models. These approaches update only a small subset of parameters, thereby reducing the risk of compromising the pre-trained capabilities of the model. Recent researches [14,24,30,37,51,53,55] have applied PEFT to control text-to-image diffusion models. For instance, some studies [14,24,37] empower the model to learn" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 141, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 141, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 141, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 219, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 219, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 219, + 102 + ], + "type": "text", + "content": "Huang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "text", + "content": "personalized or unseen concepts by fine-tuning a new text token and a small number of parameters using few user-provided exemplar images. Meanwhile, other works [30, 51, 53, 55] aim to enable diffusion models to generate images based on additional conditions, e.g., edge maps, depth maps, or segmentation masks. They fine-tune lightweight task-specific modules with condition-image pairs to achieve control over the image generation process. Despite the effectiveness of these methods, they achieve learning new concepts by accessing the associated data of interest during training. As for concept erasing, since one only observes the description of the concept to be unlearned, existing PEFT-based methods cannot be directly applied." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 257, + 315, + 269 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 257, + 315, + 269 + ], + "spans": [ + { + "bbox": [ + 132, + 257, + 315, + 269 + ], + "type": "text", + "content": "2.3 Adversarial Attack & Training" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 280, + 482, + 447 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 280, + 482, + 447 + ], + "spans": [ + { + "bbox": [ + 130, + 280, + 482, + 447 + ], + "type": "text", + "content": "In adversarial attacks [2, 13, 17, 25], adversarial examples are deliberately constructed inputs, which would deceive models into making incorrect predictions. Popular methods such as Fast Gradient Sign Method (FGSM [13]) and its variants (I-FGSM [13] and MI-FGSM [13]) targeted at attacking classification models, utilizing the resulting loss gradients to produce imperceptible perturbations and induce misclassification. In contrast, recent approaches [11, 47, 54] introduce prompt-based adversarial attacks tailored to provoke seemingly unlearned diffusion models into generating images of the unlearned concepts. For instance, P4D [11] learns prompts to reconstruct noise associated with the target concept in diffusion models, and Ring-A-Bell [47] extracts holistic concept representations from CLIP model [33] to generate model-agnostic attack prompts. These learned attack prompts have been shown to provoke unlearned models to regenerate images of the erased concept, posing potential issues in text-to-image generative models." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 448, + 482, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 448, + 482, + 544 + ], + "spans": [ + { + "bbox": [ + 130, + 448, + 482, + 544 + ], + "type": "text", + "content": "To defend against such adversarial attacks, various adversarial training strategies have been proposed [3, 17, 28]. Such adversarial training schemes expose models to adversarial examples during training, enabling the models to learn and recognize these examples. Consequently, these adversially trained models can respond accurately when encountering adversarial examples during inference. Inspired by adversarial training approaches, we employ an adversarial erasing learning scheme to introduce additional robustness to the unlearned model. We will detail our proposed framework in the following section." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 565, + 202, + 578 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 565, + 202, + 578 + ], + "spans": [ + { + "bbox": [ + 132, + 565, + 202, + 578 + ], + "type": "text", + "content": "3 Method" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": "Problem formulation. We first define the setting and notations of our Receler. Given a pre-trained text-to-image diffusion model, parameterized by " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": ", we aim to erase a textual concept " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": " from the model without requiring access to the corresponding image data. The erasure is considered successful when the model no longer generates images that contain or represent the concept " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": " (e.g., \"nudity\"-erased model should not generate any images with exposed body parts)." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 416, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 416, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 416, + 91, + 447, + 100 + ], + "type": "text", + "content": "Receler" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 137, + 112, + 207, + 266 + ], + "blocks": [ + { + "bbox": [ + 137, + 112, + 207, + 266 + ], + "lines": [ + { + "bbox": [ + 137, + 112, + 207, + 266 + ], + "spans": [ + { + "bbox": [ + 137, + 112, + 207, + 266 + ], + "type": "image", + "image_path": "75e522671b89ce6f7b06b60d99906beccdb442bb70123b6778b11acf04fe5179.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 284, + 482, + 361 + ], + "lines": [ + { + "bbox": [ + 130, + 284, + 482, + 361 + ], + "spans": [ + { + "bbox": [ + 130, + 284, + 482, + 361 + ], + "type": "text", + "content": "Fig. 2: Overview of Receler. (a) Receler involves iterative learning of a lightweight Eraser " + }, + { + "bbox": [ + 130, + 284, + 482, + 361 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 130, + 284, + 482, + 361 + ], + "type": "text", + "content": " and adversarial prompt embedding " + }, + { + "bbox": [ + 130, + 284, + 482, + 361 + ], + "type": "inline_equation", + "content": "e_{Adv}" + }, + { + "bbox": [ + 130, + 284, + 482, + 361 + ], + "type": "text", + "content": ". The former is trained to erase the target concept " + }, + { + "bbox": [ + 130, + 284, + 482, + 361 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 130, + 284, + 482, + 361 + ], + "type": "text", + "content": " while preserving non-target concepts, and the latter learns to imitate the prompts to recover visual content associated with the concept previously erased. (b) The Eraser " + }, + { + "bbox": [ + 130, + 284, + 482, + 361 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 130, + 284, + 482, + 361 + ], + "type": "text", + "content": " is inserted after each cross attention layer of Diffusion U-Net to remove the target concept from its outputs, with prediction " + }, + { + "bbox": [ + 130, + 284, + 482, + 361 + ], + "type": "inline_equation", + "content": "o^l" + }, + { + "bbox": [ + 130, + 284, + 482, + 361 + ], + "type": "text", + "content": " directly added to the cross attention output." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 209, + 112, + 380, + 266 + ], + "blocks": [ + { + "bbox": [ + 209, + 112, + 380, + 266 + ], + "lines": [ + { + "bbox": [ + 209, + 112, + 380, + 266 + ], + "spans": [ + { + "bbox": [ + 209, + 112, + 380, + 266 + ], + "type": "image", + "image_path": "75e57cece7844947bb7b0a5404923736184dd37d6ea7795eb4a111171f4f5dd7.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 255, + 266, + 264, + 275 + ], + "lines": [ + { + "bbox": [ + 255, + 266, + 264, + 275 + ], + "spans": [ + { + "bbox": [ + 255, + 266, + 264, + 275 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 380, + 112, + 478, + 266 + ], + "blocks": [ + { + "bbox": [ + 380, + 112, + 478, + 266 + ], + "lines": [ + { + "bbox": [ + 380, + 112, + 478, + 266 + ], + "spans": [ + { + "bbox": [ + 380, + 112, + 478, + 266 + ], + "type": "image", + "image_path": "243f92659970b1ae4d644fca56012aca6b72b6146e9bc3d4b349f35a0eb7f45e.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 425, + 266, + 435, + 274 + ], + "lines": [ + { + "bbox": [ + 425, + 266, + 435, + 274 + ], + "spans": [ + { + "bbox": [ + 425, + 266, + 435, + 274 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 386, + 482, + 446 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 386, + 482, + 446 + ], + "spans": [ + { + "bbox": [ + 130, + 386, + 482, + 446 + ], + "type": "text", + "content": "As depicted in Fig. 2, our method employs a lightweight Eraser " + }, + { + "bbox": [ + 130, + 386, + 482, + 446 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 130, + 386, + 482, + 446 + ], + "type": "text", + "content": ", parameterized by " + }, + { + "bbox": [ + 130, + 386, + 482, + 446 + ], + "type": "inline_equation", + "content": "\\theta_E" + }, + { + "bbox": [ + 130, + 386, + 482, + 446 + ], + "type": "text", + "content": ", to learn to erase the target concept " + }, + { + "bbox": [ + 130, + 386, + 482, + 446 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 130, + 386, + 482, + 446 + ], + "type": "text", + "content": ", as discussed in Sec. 3.1. To introduce the desirable locality and robustness to our model, we incorporate concept-localized regularization and adversarial prompt learning schemes into our framework, as detailed in Sec. 3.2 and Sec. 3.3." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 464, + 369, + 477 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 464, + 369, + 477 + ], + "spans": [ + { + "bbox": [ + 131, + 464, + 369, + 477 + ], + "type": "text", + "content": "3.1 Concept Erasing with Lightweight Eraser" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 486, + 482, + 630 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 486, + 482, + 630 + ], + "spans": [ + { + "bbox": [ + 130, + 486, + 482, + 630 + ], + "type": "text", + "content": "In order to erase particular visual concepts from a pre-trained diffusion model, we introduce a lightweight adapter-based eraser. As depicted in Fig. 2, by solely finetuning the newly introduced " + }, + { + "bbox": [ + 130, + 486, + 482, + 630 + ], + "type": "inline_equation", + "content": "\\theta_{E}" + }, + { + "bbox": [ + 130, + 486, + 482, + 630 + ], + "type": "text", + "content": ", our goal is to unlearn the target visual concept while preserving model generalization on non-target concepts. To be specific, our eraser is designed to remove the target concept from the output visual features of each cross-attention layer within the diffusion U-Net. This is based on the fact that these layers are responsible for incorporating the input text concept into the visual features. Thus, our eraser is positioned subsequent to each of the cross-attention layers, which is in line with the empirical analysis presented in [49]. To erase the concept, the eraser is trained to predict the negatively guided noises [15, 19] that move the model's prediction away from the erased concept. The objective is defined as:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 212, + 638, + 376, + 654 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 212, + 638, + 376, + 654 + ], + "spans": [ + { + "bbox": [ + 212, + 638, + 376, + 654 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {E r a s e} = \\mathbb {E} _ {x _ {t}, t} \\left[ \\lVert \\epsilon_ {\\theta^ {\\prime}} (x _ {t}, e _ {c}, t) - \\epsilon_ {E} \\rVert^ {2} \\right],", + "image_path": "a624acd1b32ffc0b65eb535b771014e480a24ce887e582539455cac9090ea1e8.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 200, + 654, + 413, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 200, + 654, + 413, + 667 + ], + "spans": [ + { + "bbox": [ + 200, + 654, + 413, + 667 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 200, + 654, + 413, + 667 + ], + "type": "inline_equation", + "content": "\\epsilon_{E} = \\epsilon_{\\theta}(x_{t},t) - \\eta [\\epsilon_{\\theta}(x_{t},e_{c},t) - \\epsilon_{\\theta}(x_{t},t)]" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 391, + 647, + 480, + 659 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 391, + 647, + 480, + 659 + ], + "spans": [ + { + "bbox": [ + 391, + 647, + 480, + 659 + ], + "type": "text", + "content": "(1)" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "type": "text", + "content": "Huang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 131, + 116, + 482, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 116, + 482, + 189 + ], + "spans": [ + { + "bbox": [ + 131, + 116, + 482, + 189 + ], + "type": "text", + "content": "Note that " + }, + { + "bbox": [ + 131, + 116, + 482, + 189 + ], + "type": "inline_equation", + "content": "\\theta' = \\{\\theta, \\theta_E\\}" + }, + { + "bbox": [ + 131, + 116, + 482, + 189 + ], + "type": "text", + "content": " represents the parameters of the diffusion model plugged with eraser, " + }, + { + "bbox": [ + 131, + 116, + 482, + 189 + ], + "type": "inline_equation", + "content": "x_t \\in \\mathbb{R}^{W \\times H \\times d}" + }, + { + "bbox": [ + 131, + 116, + 482, + 189 + ], + "type": "text", + "content": " is the denoised image at timestep " + }, + { + "bbox": [ + 131, + 116, + 482, + 189 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 131, + 116, + 482, + 189 + ], + "type": "text", + "content": " sampled from " + }, + { + "bbox": [ + 131, + 116, + 482, + 189 + ], + "type": "inline_equation", + "content": "\\theta'" + }, + { + "bbox": [ + 131, + 116, + 482, + 189 + ], + "type": "text", + "content": " conditioned on " + }, + { + "bbox": [ + 131, + 116, + 482, + 189 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 131, + 116, + 482, + 189 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 131, + 116, + 482, + 189 + ], + "type": "inline_equation", + "content": "e_c" + }, + { + "bbox": [ + 131, + 116, + 482, + 189 + ], + "type": "text", + "content": " is the text embedding of concept " + }, + { + "bbox": [ + 131, + 116, + 482, + 189 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 131, + 116, + 482, + 189 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 131, + 116, + 482, + 189 + ], + "type": "inline_equation", + "content": "\\epsilon_E" + }, + { + "bbox": [ + 131, + 116, + 482, + 189 + ], + "type": "text", + "content": " is the negatively guided noises [15, 19] predicted by " + }, + { + "bbox": [ + 131, + 116, + 482, + 189 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 131, + 116, + 482, + 189 + ], + "type": "text", + "content": ". By minimizing the L2 distance between " + }, + { + "bbox": [ + 131, + 116, + 482, + 189 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta'}(x_t, e_c, t)" + }, + { + "bbox": [ + 131, + 116, + 482, + 189 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 116, + 482, + 189 + ], + "type": "inline_equation", + "content": "\\epsilon_E" + }, + { + "bbox": [ + 131, + 116, + 482, + 189 + ], + "type": "text", + "content": ", the eraser learns to reduce the probability of the generated image " + }, + { + "bbox": [ + 131, + 116, + 482, + 189 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 131, + 116, + 482, + 189 + ], + "type": "text", + "content": " belongs to the target concept " + }, + { + "bbox": [ + 131, + 116, + 482, + 189 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 131, + 116, + 482, + 189 + ], + "type": "text", + "content": ", thus effectively erasing the concept." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 205, + 433, + 217 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 205, + 433, + 217 + ], + "spans": [ + { + "bbox": [ + 131, + 205, + 433, + 217 + ], + "type": "text", + "content": "3.2 Concept-Localized Regularization for Erasing Locality" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 223, + 482, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 223, + 482, + 319 + ], + "spans": [ + { + "bbox": [ + 131, + 223, + 482, + 319 + ], + "type": "text", + "content": "As the first desirable property in concept erasing, locality refers to preserving the model's ability to synthesize content unrelated to the target, which is realized by enforcing the eraser to affect the image synthesis process only if the target concept is present. To achieve this, we introduce concept-localized regularization into our Receler by leveraging the spatial information associated with the target concept's text tokens to regularize the eraser outputs. Specifically, inspired by [6, 46], we obtain the binary target concept mask " + }, + { + "bbox": [ + 131, + 223, + 482, + 319 + ], + "type": "inline_equation", + "content": "M \\in \\mathbb{R}^{W/4 \\times H/4}" + }, + { + "bbox": [ + 131, + 223, + 482, + 319 + ], + "type": "text", + "content": " by thresholding the attention maps when predicting " + }, + { + "bbox": [ + 131, + 223, + 482, + 319 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}(x_t, e_c, t)" + }, + { + "bbox": [ + 131, + 223, + 482, + 319 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 235, + 327, + 481, + 370 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 235, + 327, + 481, + 370 + ], + "spans": [ + { + "bbox": [ + 235, + 327, + 481, + 370 + ], + "type": "interline_equation", + "content": "M _ {i, j} = \\left\\{ \\begin{array}{l l} 1, & \\text {i f} \\frac {1}{| \\mathcal {S} |} \\sum_ {s \\in \\mathcal {S}} A _ {i, j} ^ {s} \\geq \\tau , \\\\ 0, & \\text {o t h e r w i s e .} \\end{array} \\right. \\tag {2}", + "image_path": "767e9160cbd6417e3c7a7fb6e39dc94fea86744fab4f05c26bf30002e24d7588.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 377, + 482, + 439 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 377, + 482, + 439 + ], + "spans": [ + { + "bbox": [ + 131, + 377, + 482, + 439 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 377, + 482, + 439 + ], + "type": "inline_equation", + "content": "\\mathcal{S}" + }, + { + "bbox": [ + 131, + 377, + 482, + 439 + ], + "type": "text", + "content": " is the set of indices of all U-Net's mid-layers with resolution " + }, + { + "bbox": [ + 131, + 377, + 482, + 439 + ], + "type": "inline_equation", + "content": "(\\frac{W}{4},\\frac{H}{4})" + }, + { + "bbox": [ + 131, + 377, + 482, + 439 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 131, + 377, + 482, + 439 + ], + "type": "inline_equation", + "content": "A^s\\in \\mathbb{R}^{W / 4\\times H / 4}" + }, + { + "bbox": [ + 131, + 377, + 482, + 439 + ], + "type": "text", + "content": " is the cross-attention map in the " + }, + { + "bbox": [ + 131, + 377, + 482, + 439 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 131, + 377, + 482, + 439 + ], + "type": "text", + "content": "-th layer of the text tokens corresponding to the concept " + }, + { + "bbox": [ + 131, + 377, + 482, + 439 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 131, + 377, + 482, + 439 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 131, + 377, + 482, + 439 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 131, + 377, + 482, + 439 + ], + "type": "text", + "content": " is a pre-defined threshold. With " + }, + { + "bbox": [ + 131, + 377, + 482, + 439 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 131, + 377, + 482, + 439 + ], + "type": "text", + "content": " obtained, we calculate the following regularization loss to regularize the outputs of the eraser as follows:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 241, + 445, + 481, + 476 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 241, + 445, + 481, + 476 + ], + "spans": [ + { + "bbox": [ + 241, + 445, + 481, + 476 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {R e g} = \\frac {1}{L} \\sum_ {l = 1} ^ {L} \\| o ^ {l} \\odot (1 - \\tilde {M}) \\| ^ {2}, \\tag {3}", + "image_path": "2da81f5b01303b9e9c6663515848cda28a5e8a7d921f37646302f418bb2fa5b3.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 485, + 482, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 485, + 482, + 559 + ], + "spans": [ + { + "bbox": [ + 131, + 485, + 482, + 559 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 485, + 482, + 559 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 131, + 485, + 482, + 559 + ], + "type": "text", + "content": " is the number of U-Net's layers, " + }, + { + "bbox": [ + 131, + 485, + 482, + 559 + ], + "type": "inline_equation", + "content": "\\odot" + }, + { + "bbox": [ + 131, + 485, + 482, + 559 + ], + "type": "text", + "content": " is the element-wise product, " + }, + { + "bbox": [ + 131, + 485, + 482, + 559 + ], + "type": "inline_equation", + "content": "o^l \\in \\mathbb{R}^{w^l \\times h^l \\times d}" + }, + { + "bbox": [ + 131, + 485, + 482, + 559 + ], + "type": "text", + "content": " is the output of the eraser in the " + }, + { + "bbox": [ + 131, + 485, + 482, + 559 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 131, + 485, + 482, + 559 + ], + "type": "text", + "content": "-th layer with the resolution " + }, + { + "bbox": [ + 131, + 485, + 482, + 559 + ], + "type": "inline_equation", + "content": "(w^l, h^l)" + }, + { + "bbox": [ + 131, + 485, + 482, + 559 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 131, + 485, + 482, + 559 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 131, + 485, + 482, + 559 + ], + "type": "text", + "content": " is the feature dimension. Note that " + }, + { + "bbox": [ + 131, + 485, + 482, + 559 + ], + "type": "inline_equation", + "content": "\\tilde{M}" + }, + { + "bbox": [ + 131, + 485, + 482, + 559 + ], + "type": "text", + "content": " is the shorthand of " + }, + { + "bbox": [ + 131, + 485, + 482, + 559 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 131, + 485, + 482, + 559 + ], + "type": "text", + "content": " being bicubically upscaled to the same resolution as " + }, + { + "bbox": [ + 131, + 485, + 482, + 559 + ], + "type": "inline_equation", + "content": "o^l" + }, + { + "bbox": [ + 131, + 485, + 482, + 559 + ], + "type": "text", + "content": ". With the above regularization introduced, our Receler is enforced to preserve the generation of non-target concepts. Thus, the diversity and fidelity of the original model can be maintained." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 574, + 428, + 587 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 574, + 428, + 587 + ], + "spans": [ + { + "bbox": [ + 131, + 574, + 428, + 587 + ], + "type": "text", + "content": "3.3 Adversarial Prompt Learning for Erasing Robustness" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": "To further ensure our Receler being robust against prompting attacks (e.g., by paraphrased or learned prompts), we introduce an adversarial learning strategy into our framework to enrich such model robustness. In our framework, we present a unique adversarial prompt learning scheme, which learns prompting attacks that would induce the diffusion model to synthesize the images containing previously erased concepts. To achieve this, we design the adversarial loss," + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 416, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 416, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 416, + 91, + 447, + 100 + ], + "type": "text", + "content": "Receler" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 152 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 152 + ], + "type": "text", + "content": "which optimizes the continuous soft prompts " + }, + { + "bbox": [ + 130, + 116, + 482, + 152 + ], + "type": "inline_equation", + "content": "e_{Adv}" + }, + { + "bbox": [ + 130, + 116, + 482, + 152 + ], + "type": "text", + "content": " by encouraging such learned prompts to imitate the malicious prompts, as illustrated in Fig. 2. More precisely, the objective " + }, + { + "bbox": [ + 130, + 116, + 482, + 152 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{Adv}" + }, + { + "bbox": [ + 130, + 116, + 482, + 152 + ], + "type": "text", + "content": " is defined as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 228, + 163, + 481, + 178 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 228, + 163, + 481, + 178 + ], + "spans": [ + { + "bbox": [ + 228, + 163, + 481, + 178 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {A d v} = \\mathbb {E} _ {x _ {t}, t} \\left[ \\| \\epsilon_ {\\theta^ {\\prime}} (x _ {t}, e ^ {\\prime}, t) - \\epsilon_ {M} \\| ^ {2} \\right], \\tag {4}", + "image_path": "97e5b76b49f02cb2e7d6f84f5634b002ea84f6686d807823da1f36cfcacc545d.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 187, + 483, + 296 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 187, + 483, + 296 + ], + "spans": [ + { + "bbox": [ + 130, + 187, + 483, + 296 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 187, + 483, + 296 + ], + "type": "inline_equation", + "content": "e' = [e_c; e_{Adv}]" + }, + { + "bbox": [ + 130, + 187, + 483, + 296 + ], + "type": "text", + "content": " is the concatenated prompts of the erased target concept embedding and the learned soft prompts, and " + }, + { + "bbox": [ + 130, + 187, + 483, + 296 + ], + "type": "inline_equation", + "content": "\\epsilon_M = \\epsilon_\\theta(x_t, e_c, t)" + }, + { + "bbox": [ + 130, + 187, + 483, + 296 + ], + "type": "text", + "content": " represents the malicious noise predicted by the pre-trained diffusion model (without safety mechanism) conditioned on the target concept. By minimizing this adversarial loss, " + }, + { + "bbox": [ + 130, + 187, + 483, + 296 + ], + "type": "inline_equation", + "content": "e_{Adv}" + }, + { + "bbox": [ + 130, + 187, + 483, + 296 + ], + "type": "text", + "content": " learns to regenerate the unlearned concept from the erased model. The optimization of the soft prompt " + }, + { + "bbox": [ + 130, + 187, + 483, + 296 + ], + "type": "inline_equation", + "content": "e_{Adv}" + }, + { + "bbox": [ + 130, + 187, + 483, + 296 + ], + "type": "text", + "content": " and the eraser " + }, + { + "bbox": [ + 130, + 187, + 483, + 296 + ], + "type": "inline_equation", + "content": "\\theta_E" + }, + { + "bbox": [ + 130, + 187, + 483, + 296 + ], + "type": "text", + "content": " is performed iteratively, with each being fixed while the other is trained. Thus, " + }, + { + "bbox": [ + 130, + 187, + 483, + 296 + ], + "type": "inline_equation", + "content": "\\theta_E" + }, + { + "bbox": [ + 130, + 187, + 483, + 296 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 187, + 483, + 296 + ], + "type": "inline_equation", + "content": "e_{Adv}" + }, + { + "bbox": [ + 130, + 187, + 483, + 296 + ], + "type": "text", + "content": " are trained against each other to improve model robustness. For more details, please refer to the pseudo-algorithm in the supplementary material." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 315, + 230, + 330 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 315, + 230, + 330 + ], + "spans": [ + { + "bbox": [ + 132, + 315, + 230, + 330 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 342, + 482, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 342, + 482, + 389 + ], + "spans": [ + { + "bbox": [ + 130, + 342, + 482, + 389 + ], + "type": "text", + "content": "In this section, we first conduct quantitative experiments to assess the robustness and locality of Receler compared to state-of-the-art baselines, followed by ablation studies of our method. Lastly, we present qualitative comparisons and visualizations to demonstrate its effectiveness." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 408, + 482, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 408, + 482, + 445 + ], + "spans": [ + { + "bbox": [ + 130, + 408, + 482, + 445 + ], + "type": "text", + "content": "Datasets. We conduct experiments on erasing objects defined in the CIFAR-10 dataset [22] and on erasing inappropriate contents from the Inappropriate Image Prompts (I2P) dataset [39]:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 136, + 453, + 482, + 648 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 136, + 453, + 482, + 596 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 453, + 482, + 596 + ], + "spans": [ + { + "bbox": [ + 136, + 453, + 482, + 596 + ], + "type": "text", + "content": "- Object Erasure. To evaluate the effectiveness of erasure methods in erasing common visual concepts, we choose to erase ten class labels from CIFAR-10 [22]. Note that during our experiment, we only utilize the label set, not the images. For comprehensive assessment, we devise two types of evaluation prompts for each class: Firstly, we use simple prompts formatted as \"A photo of {class}\" to evaluate the efficacy of erasure in removing the target concept. Secondly, to further assess the robustness, we generate 50 paraphrased prompts for each class using ChatGPT3 to simulate real-world scenarios where prompts are typically more complex and target concepts may not be explicitly mentioned. For example, a paraphrased prompt for \"airplane\" is \"A sleek, black stealth bomber flying low over a desert landscape at dusk.\" More details can be found in the supplementary material." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 136, + 600, + 482, + 648 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 600, + 482, + 648 + ], + "spans": [ + { + "bbox": [ + 136, + 600, + 482, + 648 + ], + "type": "text", + "content": "- Inappropriate Content Erasure. The I2P dataset [39], collected from a text-to-image prompt dataset DiffusionDB [48], comprises 4,703 real-world, user-generated prompts that produce inappropriate images, including hate, harassment, violence, self-harm, shocking, sexual, and illegal content." + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "type": "text", + "content": "Huang et al." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 145, + 653, + 256, + 666 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 653, + 256, + 666 + ], + "spans": [ + { + "bbox": [ + 145, + 653, + 256, + 666 + ], + "type": "text", + "content": "3https://chat.openai.com/" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 479, + 140 + ], + "type": "text", + "content": "Evaluation Setup. We assess the robustness and locality of the erasure methods for object erasure and inappropriate content erasure as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 136, + 147, + 482, + 302 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 147, + 482, + 302 + ], + "spans": [ + { + "bbox": [ + 136, + 147, + 482, + 302 + ], + "type": "text", + "content": "- Object Erasure. For each method, we fine-tune ten models, each erasing one CIFAR-10 class. Each model is then evaluated by: 1) Efficacy " + }, + { + "bbox": [ + 136, + 147, + 482, + 302 + ], + "type": "inline_equation", + "content": "(\\mathrm{Acc}_E)" + }, + { + "bbox": [ + 136, + 147, + 482, + 302 + ], + "type": "text", + "content": ": the percentage of the target class being erased when inputting simple prompts; lower values are better. 2) Robustness " + }, + { + "bbox": [ + 136, + 147, + 482, + 302 + ], + "type": "inline_equation", + "content": "(\\mathrm{Acc}_R)" + }, + { + "bbox": [ + 136, + 147, + 482, + 302 + ], + "type": "text", + "content": ": the percentage of the target class being erased when inputting paraphrased prompts; lower values are better. 3) Locality " + }, + { + "bbox": [ + 136, + 147, + 482, + 302 + ], + "type": "inline_equation", + "content": "(\\mathrm{Acc}_L)" + }, + { + "bbox": [ + 136, + 147, + 482, + 302 + ], + "type": "text", + "content": ": the percentage of non-target classes being preserved; higher values are better. To assess efficacy and robustness, we generate 150 images using simple and paraphrased prompts for the target class, respectively. For Locality, we generate 50 images for each of the nine non-target classes using paraphrased prompts. We then use GroundingDINO [27] to detect if the corresponding class is presenting in the image, thereby assessing " + }, + { + "bbox": [ + 136, + 147, + 482, + 302 + ], + "type": "inline_equation", + "content": "\\mathrm{Acc}_E" + }, + { + "bbox": [ + 136, + 147, + 482, + 302 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 136, + 147, + 482, + 302 + ], + "type": "inline_equation", + "content": "\\mathrm{Acc}_R" + }, + { + "bbox": [ + 136, + 147, + 482, + 302 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 136, + 147, + 482, + 302 + ], + "type": "inline_equation", + "content": "\\mathrm{Acc}_L" + }, + { + "bbox": [ + 136, + 147, + 482, + 302 + ], + "type": "text", + "content": ". To further evaluate the overall performance, we calculate the harmonic mean (H) of " + }, + { + "bbox": [ + 136, + 147, + 482, + 302 + ], + "type": "inline_equation", + "content": "100 - \\mathrm{Acc}_E" + }, + { + "bbox": [ + 136, + 147, + 482, + 302 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 136, + 147, + 482, + 302 + ], + "type": "inline_equation", + "content": "100 - \\mathrm{Acc}_R" + }, + { + "bbox": [ + 136, + 147, + 482, + 302 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 136, + 147, + 482, + 302 + ], + "type": "inline_equation", + "content": "\\mathrm{Acc}_L" + }, + { + "bbox": [ + 136, + 147, + 482, + 302 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 136, + 302, + 482, + 410 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 302, + 482, + 410 + ], + "spans": [ + { + "bbox": [ + 136, + 302, + 482, + 410 + ], + "type": "text", + "content": "- Inappropriate Content Erasure. Following ESD [15], we fine-tune two models for each erasure method: one for \"nudity\" and the other for the predefined inappropriate concepts e.g. hate, harassment, and violence. We use the NudeNet detector [5] to detect nudity and both the NudeNet and the Q16 detector [40] to identify the inappropriate concepts. Model robustness is evaluated by using real-world prompts in I2P [39]. Locality is evaluated using COCO-30K [26], a nudity-free dataset, by employing the nudity-erased model to generate safe contents from COCO-30K prompts and evaluating the quality of the generated images in terms of FID [18] and CLIP [33]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 426, + 482, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 426, + 482, + 510 + ], + "spans": [ + { + "bbox": [ + 132, + 426, + 482, + 510 + ], + "type": "text", + "content": "Comparisons. We compare Receler to state-of-the-art erasing methods, including FMN [52], SLD [39], Ablating [23], ESD [15], and UCE [16]. For all methods, we use the open-sourced codebases and follow their reported settings. We fine-tune all models from SD v1.4 [36], and for all image generation, we employ DDIM sampler [45] over 50 steps and a guidance scale of 7.5. Following SD v1.4, the image resolution in all our experiments is " + }, + { + "bbox": [ + 132, + 426, + 482, + 510 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 132, + 426, + 482, + 510 + ], + "type": "text", + "content": ". Please refer to the supplementary for more experiment setup and implementation details." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 527, + 281, + 539 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 527, + 281, + 539 + ], + "spans": [ + { + "bbox": [ + 132, + 527, + 281, + 539 + ], + "type": "text", + "content": "4.1 Quantitative Evaluation" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 545, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 545, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 545, + 482, + 666 + ], + "type": "text", + "content": "Object Erasure. In Tab. 1, we show that Receler surpasses previous state-of-the-art methods in erasing common visual concepts from CIFAR-10 [22] class labels. Notably, Receler achieves the highest harmonic mean " + }, + { + "bbox": [ + 132, + 545, + 482, + 666 + ], + "type": "inline_equation", + "content": "(H)" + }, + { + "bbox": [ + 132, + 545, + 482, + 666 + ], + "type": "text", + "content": " and exceeds the second-best method by 11.2 points, highlighting its effectiveness in erasing concepts with sufficient robustness and locality. Specifically, when assessing method efficacy using the simple prompts " + }, + { + "bbox": [ + 132, + 545, + 482, + 666 + ], + "type": "inline_equation", + "content": "(\\mathrm{Acc}_E)" + }, + { + "bbox": [ + 132, + 545, + 482, + 666 + ], + "type": "text", + "content": ", Receler achieves an average accuracy of " + }, + { + "bbox": [ + 132, + 545, + 482, + 666 + ], + "type": "inline_equation", + "content": "14.9\\%" + }, + { + "bbox": [ + 132, + 545, + 482, + 666 + ], + "type": "text", + "content": " across the erased classes, " + }, + { + "bbox": [ + 132, + 545, + 482, + 666 + ], + "type": "inline_equation", + "content": "2.2\\%" + }, + { + "bbox": [ + 132, + 545, + 482, + 666 + ], + "type": "text", + "content": " better than the second-best method, ESD [15]. In addition, when evaluating method robustness with paraphrased prompts " + }, + { + "bbox": [ + 132, + 545, + 482, + 666 + ], + "type": "inline_equation", + "content": "(\\mathrm{Acc}_R)" + }, + { + "bbox": [ + 132, + 545, + 482, + 666 + ], + "type": "text", + "content": ", Receler reaches " + }, + { + "bbox": [ + 132, + 545, + 482, + 666 + ], + "type": "inline_equation", + "content": "17.6\\%" + }, + { + "bbox": [ + 132, + 545, + 482, + 666 + ], + "type": "text", + "content": " on average, outperforming ESD by " + }, + { + "bbox": [ + 132, + 545, + 482, + 666 + ], + "type": "inline_equation", + "content": "22.3\\%" + }, + { + "bbox": [ + 132, + 545, + 482, + 666 + ], + "type": "text", + "content": " in the erased classes. We evaluate the locality of the erased model by" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 416, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 416, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 416, + 91, + 447, + 100 + ], + "type": "text", + "content": "Receler" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 138, + 180, + 475, + 452 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 482, + 170 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 482, + 170 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 482, + 170 + ], + "type": "text", + "content": "Table 1: Evaluation of erasing common objects in CIFAR-10 classes. " + }, + { + "bbox": [ + 132, + 114, + 482, + 170 + ], + "type": "inline_equation", + "content": "\\mathrm{Acc}_E" + }, + { + "bbox": [ + 132, + 114, + 482, + 170 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 132, + 114, + 482, + 170 + ], + "type": "inline_equation", + "content": "\\mathrm{Acc}_R" + }, + { + "bbox": [ + 132, + 114, + 482, + 170 + ], + "type": "text", + "content": " represent efficacy and robustness, indicating accuracy of target class (which should be minimized) on simple and paraphrased prompts, respectively. Locality, " + }, + { + "bbox": [ + 132, + 114, + 482, + 170 + ], + "type": "inline_equation", + "content": "\\mathrm{Acc}_L" + }, + { + "bbox": [ + 132, + 114, + 482, + 170 + ], + "type": "text", + "content": ", is accuracy of non-target classes (which should be maximized) using paraphrased prompts. The harmonic mean " + }, + { + "bbox": [ + 132, + 114, + 482, + 170 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 132, + 114, + 482, + 170 + ], + "type": "text", + "content": " reflects overall assessment of " + }, + { + "bbox": [ + 132, + 114, + 482, + 170 + ], + "type": "inline_equation", + "content": "\\mathrm{Acc}_E" + }, + { + "bbox": [ + 132, + 114, + 482, + 170 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 132, + 114, + 482, + 170 + ], + "type": "inline_equation", + "content": "\\mathrm{Acc}_R" + }, + { + "bbox": [ + 132, + 114, + 482, + 170 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 132, + 114, + 482, + 170 + ], + "type": "inline_equation", + "content": "\\mathrm{Acc}_L" + }, + { + "bbox": [ + 132, + 114, + 482, + 170 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 138, + 180, + 475, + 452 + ], + "lines": [ + { + "bbox": [ + 138, + 180, + 475, + 452 + ], + "spans": [ + { + "bbox": [ + 138, + 180, + 475, + 452 + ], + "type": "table", + "html": "
MethodsMetricsErased conceptsavg.
air-planeauto-mobilebirdcatdeerdogfroghorseshiptruck
SD v1.4AccE89.399.393.396.799.398.796.097.395.396.096.1
AccR79.394.096.088.098.792.088.792.765.384.087.9
AccL88.887.287.087.986.787.487.887.390.488.387.9
FMN [52]AccE ↓93.397.390.092.098.095.384.795.388.794.792.9
AccR ↓80.796.793.370.795.386.767.395.360.784.083.1
AccL ↑88.088.286.087.684.488.086.485.890.488.987.4
H↑14.14.411.517.64.110.027.96.924.011.414.2
Ablating [23]AccE ↓78.074.776.793.392.797.394.7100.090.786.788.5
AccR ↓67.390.093.372.795.387.371.390.758.076.080.2
AccL ↑87.883.884.987.384.085.886.085.388.286.486.0
H↑34.319.814.715.28.36.512.80.021.023.420.1
ESD [15]AccE ↓20.044.011.314.019.320.013.38.716.04.717.1
AccR ↓33.381.354.018.040.727.338.741.332.032.739.9
AccL ↑83.679.872.971.868.070.079.368.286.779.175.9
H↑76.035.864.279.568.274.074.270.378.679.071.6
UCE [16]AccE ↓34.746.08.716.74.011.311.36.022.010.017.1
AccR ↓58.079.363.316.015.349.328.034.041.339.342.4
AccL ↑84.979.181.882.078.082.283.375.887.381.681.6
H↑58.937.859.583.185.669.580.777.072.675.372.0
Receler (Ours)AccE ↓10.046.73.311.32.76.723.37.324.014.014.9
AccR ↓16.762.026.70.72.04.717.36.020.719.317.6
AccL ↑88.481.382.280.476.774.783.880.488.284.782.1
H↑87.152.383.088.889.586.780.988.680.883.783.2
", + "image_path": "6a56c92a563c8407a2dbab671512b9086b026234652264dfc393f11e89a900da.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 478, + 482, + 550 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 478, + 482, + 550 + ], + "spans": [ + { + "bbox": [ + 130, + 478, + 482, + 550 + ], + "type": "text", + "content": "examining its accuracy in the remaining classes " + }, + { + "bbox": [ + 130, + 478, + 482, + 550 + ], + "type": "inline_equation", + "content": "(\\mathrm{Acc}_L)" + }, + { + "bbox": [ + 130, + 478, + 482, + 550 + ], + "type": "text", + "content": " aside from the erased class, focusing on whether erasing one CIFAR-10 class affects the image synthesis of the other unrelated classes. Although FMN [52] and Ablating [23] exhibit high average " + }, + { + "bbox": [ + 130, + 478, + 482, + 550 + ], + "type": "inline_equation", + "content": "\\mathrm{Acc}_L" + }, + { + "bbox": [ + 130, + 478, + 482, + 550 + ], + "type": "text", + "content": ", appearing effective in preserving model locality, they struggle to erase the target objects, with only a " + }, + { + "bbox": [ + 130, + 478, + 482, + 550 + ], + "type": "inline_equation", + "content": "3.2\\%" + }, + { + "bbox": [ + 130, + 478, + 482, + 550 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 478, + 482, + 550 + ], + "type": "inline_equation", + "content": "7.6\\%" + }, + { + "bbox": [ + 130, + 478, + 482, + 550 + ], + "type": "text", + "content": " drop in " + }, + { + "bbox": [ + 130, + 478, + 482, + 550 + ], + "type": "inline_equation", + "content": "\\mathrm{Acc}_E" + }, + { + "bbox": [ + 130, + 478, + 482, + 550 + ], + "type": "text", + "content": " from SD, compared to our " + }, + { + "bbox": [ + 130, + 478, + 482, + 550 + ], + "type": "inline_equation", + "content": "81.2\\%" + }, + { + "bbox": [ + 130, + 478, + 482, + 550 + ], + "type": "text", + "content": " drop." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 569, + 482, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 569, + 482, + 640 + ], + "spans": [ + { + "bbox": [ + 130, + 569, + 482, + 640 + ], + "type": "text", + "content": "Erasure of Inappropriate Content. In Tab. 2 and Tab. 3, we evaluate the robustness of erasing inappropriate content with real-world prompts from I2P dataset [39]. Compared to the second-best result, Receler stands out by achieving " + }, + { + "bbox": [ + 130, + 569, + 482, + 640 + ], + "type": "inline_equation", + "content": "4.3\\%" + }, + { + "bbox": [ + 130, + 569, + 482, + 640 + ], + "type": "text", + "content": " overall improvement in erasing sensitive concepts on I2P and a " + }, + { + "bbox": [ + 130, + 569, + 482, + 640 + ], + "type": "inline_equation", + "content": "3.2\\%" + }, + { + "bbox": [ + 130, + 569, + 482, + 640 + ], + "type": "text", + "content": " increase in erasing nudity content, underscoring its effectiveness in scenarios that require a safety mechanism." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 641, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 641, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 131, + 641, + 481, + 665 + ], + "type": "text", + "content": "In Tab. 4, in addition to robustness, we assess locality using COCO-30K [26], a nudity-free dataset. We employ the \"nudity\"-erased model to produce safe" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "type": "text", + "content": "Huang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 138, + 190, + 305, + 291 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 309, + 180 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 309, + 180 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 309, + 180 + ], + "type": "text", + "content": "Table 2: Quantitative results on Inappropriate Image Prompts (I2P) dataset. We follow SLD [39] and apply the ratio of inappropriate images as the metric. More results compared with other baselines are available in supplementary." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 138, + 190, + 305, + 291 + ], + "lines": [ + { + "bbox": [ + 138, + 190, + 305, + 291 + ], + "spans": [ + { + "bbox": [ + 138, + 190, + 305, + 291 + ], + "type": "table", + "html": "
Class nameInappropriate proportion (%) (↓)
SDFMNSLDESDUCEReceler
Hate44.237.722.526.836.428.6
Harassment37.525.022.124.029.521.7
Violence46.347.831.835.134.127.1
Self-harm47.946.830.033.730.824.8
Sexual60.259.152.435.025.529.4
Shocking59.558.140.540.141.134.8
Illegal activity40.037.022.126.729.021.3
Overall48.947.833.732.831.327.0
", + "image_path": "eab7af37e90c2144c0d485e6dfe197ea5a65a508b101cb31b27e3fd5039a395c.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 138, + 367, + 302, + 453 + ], + "blocks": [ + { + "bbox": [ + 132, + 302, + 309, + 357 + ], + "lines": [ + { + "bbox": [ + 132, + 302, + 309, + 357 + ], + "spans": [ + { + "bbox": [ + 132, + 302, + 309, + 357 + ], + "type": "text", + "content": "Table 4: Assessment of reliability of nudity-erased models. Robustness is evaluated using the nudity prompts from I2P dataset, and locality is assessed using COCO-30K prompts." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 138, + 367, + 302, + 453 + ], + "lines": [ + { + "bbox": [ + 138, + 367, + 302, + 453 + ], + "spans": [ + { + "bbox": [ + 138, + 367, + 302, + 453 + ], + "type": "table", + "html": "
MethodRobustness Nudity-erased ratio(↑)Locality
CLIP-30K(↑)FID-30K(↓)
SD-31.3214.27
FMN44.2%30.3913.52
SLD71.6%30.9016.34
ESD81.3%30.2415.31
UCE75.9%30.8514.07
Receler84.5%31.0214.10
", + "image_path": "f17d5180a696d5ee883f82a5f269dd3515959f01f47f99ff8c3ca53d816e057e.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 319, + 181, + 476, + 289 + ], + "blocks": [ + { + "bbox": [ + 312, + 116, + 482, + 172 + ], + "lines": [ + { + "bbox": [ + 312, + 116, + 482, + 172 + ], + "spans": [ + { + "bbox": [ + 312, + 116, + 482, + 172 + ], + "type": "text", + "content": "Table 3: Quantitative results on nudity prompts from I2P dataset. We report the number of nudity images detected by the NudeNet [5]. F- and M- refer to female and male, respectively." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 319, + 181, + 476, + 289 + ], + "lines": [ + { + "bbox": [ + 319, + 181, + 476, + 289 + ], + "spans": [ + { + "bbox": [ + 319, + 181, + 476, + 289 + ], + "type": "table", + "html": "
Class nameNumber of nudity detected (↓)
SDFMNSLDESDUCEReceler
Armpits1484246312939
Belly17011670206026
F-Breast26615539323513
M-Breast421730151212
Buttocks29123975
Feet635619242910
F-Genitalia18151151
M-Genitalia723749
Total743415211139179115
Erasing ratio%--44.2-71.6-81.3-75.9-84.5
", + "image_path": "90b3a8614e84a89000e609cb46c36c968cb2c1a6f3c32736206820e06a8bb24e.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 319, + 371, + 476, + 449 + ], + "blocks": [ + { + "bbox": [ + 312, + 306, + 482, + 361 + ], + "lines": [ + { + "bbox": [ + 312, + 306, + 482, + 361 + ], + "spans": [ + { + "bbox": [ + 312, + 306, + 482, + 361 + ], + "type": "text", + "content": "Table 5: Evaluation of robustness against learned attack prompts. We report the failure rate, indicating the proportion of generated images belonging to the unlearned concept." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 319, + 371, + 476, + 449 + ], + "lines": [ + { + "bbox": [ + 319, + 371, + 476, + 449 + ], + "spans": [ + { + "bbox": [ + 319, + 371, + 476, + 449 + ], + "type": "table", + "html": "
MethodP4D [11]Ring-A-Bell [47]
cifar10 avg.nudityviolencenudity
FMN88.3%89.4%98.8%94.7%
Ablating85.4%82.8%100.0%96.8%
SLD60.5%56.3%80.4%86.3%
ESD48.1%54.3%86.0%55.8%
UCE53.8%51.9%76.8%49.5%
Receler13.7%31.2%59.2%1.1%
", + "image_path": "20a8e8b65861368a967262b66abea1fca2d28948b42be53c8559ca46754515c8.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 475, + 482, + 537 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 475, + 482, + 537 + ], + "spans": [ + { + "bbox": [ + 130, + 475, + 482, + 537 + ], + "type": "text", + "content": "content using COCO-30K prompts and evaluate the quality of the generated images using FID [18] and CLIP [33] metrics. As shown in the last two columns, Receler secures the highest CLIP-30K and nearly matches the top result on FID-30K. It is noteworthy that while Receler performs comparably to FMN in FID-30K, it surpasses FMN in robustness by erasing " + }, + { + "bbox": [ + 130, + 475, + 482, + 537 + ], + "type": "inline_equation", + "content": "40.3\\%" + }, + { + "bbox": [ + 130, + 475, + 482, + 537 + ], + "type": "text", + "content": " more nudity content." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "text", + "content": "Learned Attack Prompts. To demonstrate the robustness of Receler in safeguarding against the potential and unprecedented malicious attacks, we employ P4D [11] and Ring-A-Bell [47]. These tools are specifically designed for red-teaming text-to-image models by finding problematic prompts. As illustrated in Tab. 5, Receler is significantly more reliable than other concept erasing methods. Specifically, it achieves lower failure rates " + }, + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "inline_equation", + "content": "-34.4\\%" + }, + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "text", + "content": " for CIFAR-10 and " + }, + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "inline_equation", + "content": "20.7\\%" + }, + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "text", + "content": " for nudity against P4D prompts, and " + }, + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "inline_equation", + "content": "17.6\\%" + }, + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "text", + "content": " for violence and " + }, + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "inline_equation", + "content": "48.4\\%" + }, + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "text", + "content": " for nudity against Ring-A-Bell prompts, compared to the second-best method. These results highlight the robustness of Receler in defending against malicious attacks." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 416, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 416, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 416, + 91, + 447, + 100 + ], + "type": "text", + "content": "Receler" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 168, + 121, + 447, + 269 + ], + "blocks": [ + { + "bbox": [ + 168, + 121, + 447, + 269 + ], + "lines": [ + { + "bbox": [ + 168, + 121, + 447, + 269 + ], + "spans": [ + { + "bbox": [ + 168, + 121, + 447, + 269 + ], + "type": "image", + "image_path": "567d5cece8a2374a36a1fb5f19c59e553473057643b65bd0c0e97e879b49402d.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 277, + 482, + 312 + ], + "lines": [ + { + "bbox": [ + 130, + 277, + 482, + 312 + ], + "spans": [ + { + "bbox": [ + 130, + 277, + 482, + 312 + ], + "type": "text", + "content": "Fig. 3: Qualitative comparison of concept erasure methods. Note that erased concepts are listed at the top, and images generated from each method are shown in each row. Input prompts used for image generation are provided in supplementary." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 332, + 482, + 489 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 332, + 482, + 489 + ], + "spans": [ + { + "bbox": [ + 130, + 332, + 482, + 489 + ], + "type": "text", + "content": "Ablation Study. In Tab. 6, we ablate the three proposed components in Receler: the lightweight eraser, concept-localized regularization, and adversarial prompt learning. We establish a simple baseline by fine-tuning the whole model with Eq. (1) (i.e., the first row). With the eraser introduced, both " + }, + { + "bbox": [ + 130, + 332, + 482, + 489 + ], + "type": "inline_equation", + "content": "\\mathrm{Acc}_R" + }, + { + "bbox": [ + 130, + 332, + 482, + 489 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 332, + 482, + 489 + ], + "type": "inline_equation", + "content": "\\mathrm{Acc}_L" + }, + { + "bbox": [ + 130, + 332, + 482, + 489 + ], + "type": "text", + "content": " improve. Adding concept-localized regularization further increases the " + }, + { + "bbox": [ + 130, + 332, + 482, + 489 + ], + "type": "inline_equation", + "content": "\\mathrm{Acc}_L" + }, + { + "bbox": [ + 130, + 332, + 482, + 489 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 130, + 332, + 482, + 489 + ], + "type": "inline_equation", + "content": "77.4\\%" + }, + { + "bbox": [ + 130, + 332, + 482, + 489 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 130, + 332, + 482, + 489 + ], + "type": "inline_equation", + "content": "79.8\\%" + }, + { + "bbox": [ + 130, + 332, + 482, + 489 + ], + "type": "text", + "content": ", demonstrating its effectiveness in enhancing locality. On the other hand, coupling adversarial prompting learning with the eraser boosts " + }, + { + "bbox": [ + 130, + 332, + 482, + 489 + ], + "type": "inline_equation", + "content": "\\mathrm{Acc}_R" + }, + { + "bbox": [ + 130, + 332, + 482, + 489 + ], + "type": "text", + "content": " by " + }, + { + "bbox": [ + 130, + 332, + 482, + 489 + ], + "type": "inline_equation", + "content": "14\\%" + }, + { + "bbox": [ + 130, + 332, + 482, + 489 + ], + "type": "text", + "content": ", albeit with a slight decrease in " + }, + { + "bbox": [ + 130, + 332, + 482, + 489 + ], + "type": "inline_equation", + "content": "\\mathrm{Acc}_L" + }, + { + "bbox": [ + 130, + 332, + 482, + 489 + ], + "type": "text", + "content": ". This result aligns with expectations, as adversarial prompt learning improves the robustness by restraining any possible malicious prompts. Therefore, Receler, which integrates both adversarial prompt learning and concept-localized regularization, yields the best experimental results. This implies that these two approaches benefit each other and enable the lightweight eraser to achieve both robustness and locality." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 504, + 274, + 516 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 504, + 274, + 516 + ], + "spans": [ + { + "bbox": [ + 132, + 504, + 274, + 516 + ], + "type": "text", + "content": "4.2 Qualitative Evaluation" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 521, + 482, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 482, + 616 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 482, + 616 + ], + "type": "text", + "content": "Visualization of Erased Concepts with Paraphrased Prompts. In Fig. 3, we show examples of erasing different artist styles (e.g., van Gogh), objects (e.g., automobile) and high-level concepts (e.g., nudity). As observed in the figure, ESD [15] and UCE [16] fail to erase the target concept, with the outputs from these methods closely resembling the original images from SD [1]. On the contrary, Receler successfully erases all target concepts and is able to generate images that are visually close to the original ones, e.g., same background with the car removed, a human in the same posture with nudity removed." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "In Fig. 4, we qualitatively validate the robustness and locality of Receler. The diagonal orange boxes show its robustness against paraphrased prompts where the erased concept is not explicitly mentioned. For instance, airplane-erased Receler successfully prevents the generation of an \"airplane\" image when" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "type": "text", + "content": "Huang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 158, + 123, + 458, + 353 + ], + "blocks": [ + { + "bbox": [ + 158, + 123, + 458, + 353 + ], + "lines": [ + { + "bbox": [ + 158, + 123, + 458, + 353 + ], + "spans": [ + { + "bbox": [ + 158, + 123, + 458, + 353 + ], + "type": "image", + "image_path": "9e7ccc830afb3aca565f969c79f1383d247b1eb19d85133a02f242e754a93124.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 361, + 482, + 371 + ], + "lines": [ + { + "bbox": [ + 130, + 361, + 482, + 371 + ], + "spans": [ + { + "bbox": [ + 130, + 361, + 482, + 371 + ], + "type": "text", + "content": "Fig. 4: Visualization of robustness and locality from Receler on CIFAR-10." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 372, + 481, + 404 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 372, + 481, + 404 + ], + "spans": [ + { + "bbox": [ + 130, + 372, + 481, + 404 + ], + "type": "text", + "content": "The red strikethrough at the top indicates the erased concepts. On the left, the input paraphrased prompts are provided. Images enclosed within the diagonal orange borders shows robustness while others show the locality." + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 149, + 422, + 303, + 510 + ], + "blocks": [ + { + "bbox": [ + 149, + 422, + 303, + 510 + ], + "lines": [ + { + "bbox": [ + 149, + 422, + 303, + 510 + ], + "spans": [ + { + "bbox": [ + 149, + 422, + 303, + 510 + ], + "type": "image", + "image_path": "406d958736678147da0a6e2d2fdbe48445ada524f875ea931ce8133084ccb08b.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 208, + 512, + 241, + 521 + ], + "lines": [ + { + "bbox": [ + 208, + 512, + 241, + 521 + ], + "spans": [ + { + "bbox": [ + 208, + 512, + 241, + 521 + ], + "type": "text", + "content": "(a) nudity" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 308, + 422, + 463, + 511 + ], + "blocks": [ + { + "bbox": [ + 308, + 422, + 463, + 511 + ], + "lines": [ + { + "bbox": [ + 308, + 422, + 463, + 511 + ], + "spans": [ + { + "bbox": [ + 308, + 422, + 463, + 511 + ], + "type": "image", + "image_path": "dc9a7319f4a7dadc2a2f807de0af117bdb1ed0591aa98df7796317c05fcb9437.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 364, + 512, + 402, + 521 + ], + "lines": [ + { + "bbox": [ + 364, + 512, + 402, + 521 + ], + "spans": [ + { + "bbox": [ + 364, + 512, + 402, + 521 + ], + "type": "text", + "content": "(b) violence" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 131, + 531, + 482, + 555 + ], + "lines": [ + { + "bbox": [ + 131, + 531, + 482, + 555 + ], + "spans": [ + { + "bbox": [ + 131, + 531, + 482, + 555 + ], + "type": "text", + "content": "Fig. 5: Visualization of erasure methods against learned attack prompts. We use Ring-A-Bell [47] to generate adversarial prompts for nudity and violence concepts." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "content": "the prompt is paraphrased to \"jet\". Concerning locality, as shown in the non-diagonal boxes, Receler generates images that not only faithfully adhere to the descriptions but also closely resemble the images from the original diffusion model. Notably, it can be seen that Receler is able to replace the unlearned objects with reasonable alternatives rather than simply removing them, e.g., two white puppies are substituted with two people wearing white furry clothes and gloves (see the fourth row in Fig. 4)." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 416, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 416, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 416, + 91, + 447, + 100 + ], + "type": "text", + "content": "Receler" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 143, + 194, + 284, + 266 + ], + "blocks": [ + { + "bbox": [ + 132, + 118, + 294, + 184 + ], + "lines": [ + { + "bbox": [ + 132, + 118, + 294, + 184 + ], + "spans": [ + { + "bbox": [ + 132, + 118, + 294, + 184 + ], + "type": "text", + "content": "Table 6: Ablation study on CIFAR-10. We ablate the components of Receler and report the robustness and locality metrics. The first row refers to fine-tuning all parameters with only " + }, + { + "bbox": [ + 132, + 118, + 294, + 184 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\text{Erase}}" + }, + { + "bbox": [ + 132, + 118, + 294, + 184 + ], + "type": "text", + "content": " in Eq. (1)." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 143, + 194, + 284, + 266 + ], + "lines": [ + { + "bbox": [ + 143, + 194, + 284, + 266 + ], + "spans": [ + { + "bbox": [ + 143, + 194, + 284, + 266 + ], + "type": "table", + "html": "
ComponentsMetrics
Eraser\\( \\mathcal{L}_{\\text{Reg}} \\)\\( \\mathcal{L}_{\\text{Adv}} \\)\\( \\text{Acc}_R(\\downarrow) \\)\\( \\text{Acc}_L(\\uparrow) \\)
XXX39.475.2
✓XX34.977.4
✓✓X26.379.8
✓X✓20.976.2
✓✓✓17.682.1
", + "image_path": "99ad01acf7e02f44ca8dc78587435f0e06a5aa3907ec144960b79271a4c48e9e.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 306, + 114, + 363, + 213 + ], + "blocks": [ + { + "bbox": [ + 306, + 114, + 363, + 213 + ], + "lines": [ + { + "bbox": [ + 306, + 114, + 363, + 213 + ], + "spans": [ + { + "bbox": [ + 306, + 114, + 363, + 213 + ], + "type": "image", + "image_path": "4a8a0b671d733ca8446d6c0e3ccc63968bb8ba7e63e6715a3ee7b625abeb5702.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 214, + 481, + 258 + ], + "lines": [ + { + "bbox": [ + 304, + 214, + 481, + 258 + ], + "spans": [ + { + "bbox": [ + 304, + 214, + 481, + 258 + ], + "type": "text", + "content": "Fig.6: Examples of erasing multiple concepts. Instead of training an eraser for multiple concepts from scratch, we combine existing erasers for multi-concept erasure." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 365, + 114, + 421, + 213 + ], + "blocks": [ + { + "bbox": [ + 365, + 114, + 421, + 213 + ], + "lines": [ + { + "bbox": [ + 365, + 114, + 421, + 213 + ], + "spans": [ + { + "bbox": [ + 365, + 114, + 421, + 213 + ], + "type": "image", + "image_path": "66a508933f86f395f8fbaa82c3981e47b4928dd161b9ac48c7dfb776c92b8cc6.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 423, + 115, + 479, + 213 + ], + "blocks": [ + { + "bbox": [ + 423, + 115, + 479, + 213 + ], + "lines": [ + { + "bbox": [ + 423, + 115, + 479, + 213 + ], + "spans": [ + { + "bbox": [ + 423, + 115, + 479, + 213 + ], + "type": "image", + "image_path": "132671481899a7f4e16153ac210d8d2be4ae2e0079e1fab1dae41192398d50d5.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 291, + 482, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 291, + 482, + 389 + ], + "spans": [ + { + "bbox": [ + 130, + 291, + 482, + 389 + ], + "type": "text", + "content": "Visualization of Robustness against Learned Attack Prompts. In addition to the quantitative results shown in Tab. 5, we further qualitatively demonstrate the robustness of Receler against learned attack prompts in Fig. 5. The attack prompts are learned to induce the concepts of \"nudity\" and \"violence\" from models that should have erased these concepts using Ring-A-Bell [47]. For both nudity and violence attack prompts, Receler successfully prevents the generation of erased concepts, whereas other methods like ESD and UCE fail and generate images with the supposedly forbidden concepts (e.g., nudity or blood)." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 405, + 482, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 405, + 482, + 479 + ], + "spans": [ + { + "bbox": [ + 130, + 405, + 482, + 479 + ], + "type": "text", + "content": "Compositional Concept Erasure. In Fig. 6, we showcase examples of Receler in performing compositional concept erasure. This is achieved by combining outputs from separately trained erasers, each targeting a specific unlearned concept, during inference. By averaging these outputs, compositional concept erasure is accomplished without necessitating retraining. This approach notably offers the flexibility to selectively determine which concepts are to be erased, as required." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 496, + 220, + 509 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 496, + 220, + 509 + ], + "spans": [ + { + "bbox": [ + 132, + 496, + 220, + 509 + ], + "type": "text", + "content": "5 Conclusion" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 521, + 482, + 654 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 482, + 654 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 482, + 654 + ], + "type": "text", + "content": "In this paper, we proposed Reliable Concept Erasing via Lightweight Erasers (Receler) to erase target concepts entirely from the pre-trained diffusion model against prompting attacks (i.e., robustness), while preserving its image generation ability of other concepts (i.e., locality). In Receler, we employ concept-localized regularization to enforce the eraser to only affect the visual features related to the target concept. To enhance model robustness to paraphrased or attack prompts, we present an adversarial prompt learning scheme to induce the model to produce images of previously erased concepts, followed by optimizing the model against such image generation. We conducted extensive quantitative and qualitative evaluations on Receler, validating its superior robustness and locality-preserving ability over previous concept-erasing methods." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "type": "text", + "content": "Huang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "text", + "content": "Acknowledgements. This research was supported in part by the National Science and Technology Council via grant NSTC 112-2634-F-002-007, NSTC 113-2640-E-002-003 and the Featured Area Research Center Program within the framework of the Higher Education Sprout Project by the Ministry of Education 113L900902. We also thank the National Center for High-performance Computing (NCHC) for providing computational and storage resources." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 207, + 198, + 220 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 207, + 198, + 220 + ], + "spans": [ + { + "bbox": [ + 132, + 207, + 198, + 220 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 233, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 138, + 233, + 480, + 255 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 233, + 480, + 255 + ], + "spans": [ + { + "bbox": [ + 138, + 233, + 480, + 255 + ], + "type": "text", + "content": "1. Sarah andersen. et al v. stability ai ltd. et al. case no.3:2023cv00201. us district court for the northern district of california. (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 256, + 480, + 277 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 256, + 480, + 277 + ], + "spans": [ + { + "bbox": [ + 138, + 256, + 480, + 277 + ], + "type": "text", + "content": "2. Akhtar, N., Mian, A., Kardan, N., Shah, M.: Advances in adversarial attacks and defenses in computer vision: A survey. IEEE Access 9, 155161-155196 (2021)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 278, + 480, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 278, + 480, + 300 + ], + "spans": [ + { + "bbox": [ + 138, + 278, + 480, + 300 + ], + "type": "text", + "content": "3. Bai, T., Luo, J., Zhao, J., Wen, B., Wang, Q.: Recent advances in adversarial training for adversarial robustness. arXiv preprint arXiv:2102.01356 (2021)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 300, + 480, + 333 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 300, + 480, + 333 + ], + "spans": [ + { + "bbox": [ + 138, + 300, + 480, + 333 + ], + "type": "text", + "content": "4. Balaji, Y., Nah, S., Huang, X., Vahdat, A., Song, J., Kreis, K., Aittala, M., Aila, T., Laine, S., Catanzaro, B., et al.: edifi: Text-to-image diffusion models with an ensemble of expert denoisers. arXiv preprint arXiv:2211.01324 (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 334, + 480, + 355 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 334, + 480, + 355 + ], + "spans": [ + { + "bbox": [ + 138, + 334, + 480, + 355 + ], + "type": "text", + "content": "5. Bedapudi Praneeth, b.k., lireza Ayinmehr: Nudenet: Neural nets for nudity classification, detection and selective censoring (2019)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 356, + 480, + 388 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 356, + 480, + 388 + ], + "spans": [ + { + "bbox": [ + 138, + 356, + 480, + 388 + ], + "type": "text", + "content": "6. Cao, M., Wang, X., Qi, Z., Shan, Y., Qie, X., Zheng, Y.: Masactrl: Tuning-free mutual self-attention control for consistent image synthesis and editing. arXiv preprint arXiv:2304.08465 (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 389, + 480, + 410 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 389, + 480, + 410 + ], + "spans": [ + { + "bbox": [ + 138, + 389, + 480, + 410 + ], + "type": "text", + "content": "7. Carlini, N., Jagielski, M., Zhang, C., Papernot, N., Terzis, A., Tramer, F.: The privacy onion effect: Memorization is relative. In: NeurIPS (2022)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 411, + 480, + 444 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 411, + 480, + 444 + ], + "spans": [ + { + "bbox": [ + 138, + 411, + 480, + 444 + ], + "type": "text", + "content": "8. Carlini, N., Hayes, J., Nasr, M., Jagielski, M., Sehwag, V., Tramer, F., Balle, B., Ippolito, D., Wallace, E.: Extracting training data from diffusion models. In: 32nd USENIX Security Symposium (USENIX Security 23). pp. 5253-5270 (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 445, + 480, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 445, + 480, + 477 + ], + "spans": [ + { + "bbox": [ + 138, + 445, + 480, + 477 + ], + "type": "text", + "content": "9. Carlini, N., Hayes, J., Nasr, M., Jagielski, M., Sehwag, V., Tramer, F., Balle, B., Ippolito, D., Wallace, E.: Extracting training data from diffusion models. In: 32nd USENIX Security Symposium (USENIX Security 23). pp. 5253-5270 (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 478, + 480, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 478, + 480, + 510 + ], + "spans": [ + { + "bbox": [ + 138, + 478, + 480, + 510 + ], + "type": "text", + "content": "0. Chen, J., Yang, D.: Unlearn what you want to forget: Efficient unlearning for LLMs. In: Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing. Association for Computational Linguistics (Dec 2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 510, + 480, + 543 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 510, + 480, + 543 + ], + "spans": [ + { + "bbox": [ + 138, + 510, + 480, + 543 + ], + "type": "text", + "content": "1. Chin, Z.Y., Jiang, C.M., Huang, C.C., Chen, P.Y., Chiu, W.C.: Prompting4debugging: Red-teaming text-to-image diffusion models by finding problematic prompts. arXiv preprint arXiv:2309.06135 (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 544, + 480, + 565 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 544, + 480, + 565 + ], + "spans": [ + { + "bbox": [ + 138, + 544, + 480, + 565 + ], + "type": "text", + "content": "2. De Cao, N., Aziz, W., Titov, I.: Editing factual knowledge in language models. arXiv preprint arXiv:2104.08164 (2021)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 566, + 480, + 587 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 566, + 480, + 587 + ], + "spans": [ + { + "bbox": [ + 138, + 566, + 480, + 587 + ], + "type": "text", + "content": "3. Dong, Y., Liao, F., Pang, T., Su, H., Zhu, J., Hu, X., Li, J.: Boosting adversarial attacks with momentum. In: CVPR (2018)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 588, + 480, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 588, + 480, + 620 + ], + "spans": [ + { + "bbox": [ + 138, + 588, + 480, + 620 + ], + "type": "text", + "content": "4. Gal, R., Alaluf, Y., Atzmon, Y., Patashnik, O., Bermano, A.H., Chechik, G., Cohen-Or, D.: An image is worth one word: Personalizing text-to-image generation using textual inversion. arXiv preprint arXiv:2208.01618 (2022)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 621, + 480, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 621, + 480, + 643 + ], + "spans": [ + { + "bbox": [ + 138, + 621, + 480, + 643 + ], + "type": "text", + "content": "5. Gandikota, R., Materzynska, J., Fiotto-Kaufman, J., Bau, D.: Erasing concepts from diffusion models. arXiv preprint arXiv:2303.07345 (2023)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 138, + 643, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 643, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 643, + 480, + 665 + ], + "type": "text", + "content": "6. Gandikota, R., Orgad, H., Belinkov, Y., Materzyńska, J., Bau, D.: Unified concept editing in diffusion models. arXiv preprint arXiv:2308.14761 (2023)" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 416, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 416, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 416, + 91, + 447, + 100 + ], + "type": "text", + "content": "Receler" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 117, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 22, + "blocks": [ + { + "bbox": [ + 133, + 117, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 117, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 133, + 117, + 480, + 138 + ], + "type": "text", + "content": "17. Goodfellow, I.J., Shlens, J., Szegedy, C.: Explaining and harnessing adversarial examples. arXiv preprint arXiv:1412.6572 (2014)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 140, + 480, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 140, + 480, + 172 + ], + "spans": [ + { + "bbox": [ + 133, + 140, + 480, + 172 + ], + "type": "text", + "content": "18. Heusel, M., Ramsauer, H., Unterthiner, T., Nessler, B., Hochreiter, S.: Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems 30 (2017)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 173, + 480, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 173, + 480, + 194 + ], + "spans": [ + { + "bbox": [ + 133, + 173, + 480, + 194 + ], + "type": "text", + "content": "19. Ho, J., Salimans, T.: Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598 (2022)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 195, + 480, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 195, + 480, + 217 + ], + "spans": [ + { + "bbox": [ + 132, + 195, + 480, + 217 + ], + "type": "text", + "content": "20. Huang, Z., Shen, Y., Zhang, X., Zhou, J., Rong, W., Xiong, Z.: Transformer-patcher: One mistake worth one neuron. arXiv preprint arXiv:2301.09785 (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 218, + 480, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 218, + 480, + 239 + ], + "spans": [ + { + "bbox": [ + 132, + 218, + 480, + 239 + ], + "type": "text", + "content": "21. Hunter, T.: Ai porn is easy to make now. for women, that's a nightmare. The Washington Post pp. NA-NA (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 240, + 480, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 240, + 480, + 262 + ], + "spans": [ + { + "bbox": [ + 132, + 240, + 480, + 262 + ], + "type": "text", + "content": "22. Krizhevsky, A., Hinton, G.: Learning multiple layers of features from tiny images. Master's thesis, Department of Computer Science, University of Toronto (2009)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 263, + 480, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 263, + 480, + 285 + ], + "spans": [ + { + "bbox": [ + 132, + 263, + 480, + 285 + ], + "type": "text", + "content": "23. Kumari, N., Zhang, B., Wang, S.Y., Shechtman, E., Zhang, R., Zhu, J.Y.: Ablating concepts in text-to-image diffusion models. In: ICCV (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 285, + 480, + 307 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 285, + 480, + 307 + ], + "spans": [ + { + "bbox": [ + 132, + 285, + 480, + 307 + ], + "type": "text", + "content": "24. Kumari, N., Zhang, B., Zhang, R., Shechtman, E., Zhu, J.Y.: Multi-concept customization of text-to-image diffusion. In: CVPR (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 308, + 480, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 308, + 480, + 341 + ], + "spans": [ + { + "bbox": [ + 132, + 308, + 480, + 341 + ], + "type": "text", + "content": "25. Kurakin, A., Goodfellow, I.J., Bengio, S.: Adversarial examples in the physical world. In: Artificial intelligence safety and security, pp. 99-112. Chapman and Hall/CRC (2018)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 342, + 480, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 342, + 480, + 385 + ], + "spans": [ + { + "bbox": [ + 132, + 342, + 480, + 385 + ], + "type": "text", + "content": "26. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollar, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13. pp. 740-755. Springer (2014)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 386, + 480, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 386, + 480, + 418 + ], + "spans": [ + { + "bbox": [ + 132, + 386, + 480, + 418 + ], + "type": "text", + "content": "27. Liu, S., Zeng, Z., Ren, T., Li, F., Zhang, H., Yang, J., Li, C., Yang, J., Su, H., Zhu, J., et al.: Grounding dino: Marrying dino with grounded pre-training for open-set object detection. arXiv preprint arXiv:2303.05499 (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 419, + 480, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 419, + 480, + 441 + ], + "spans": [ + { + "bbox": [ + 132, + 419, + 480, + 441 + ], + "type": "text", + "content": "28. Madry, A., Makelov, A., Schmidt, L., Tsipras, D., Vladu, A.: Towards deep learning models resistant to adversarial attacks. arXiv preprint arXiv:1706.06083 (2017)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 442, + 480, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 442, + 480, + 464 + ], + "spans": [ + { + "bbox": [ + 132, + 442, + 480, + 464 + ], + "type": "text", + "content": "29. Meng, K., Bau, D., Andonian, A., Belinkov, Y.: Locating and editing factual associations in gpt. In: NeurIPS (2022)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 464, + 480, + 497 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 464, + 480, + 497 + ], + "spans": [ + { + "bbox": [ + 132, + 464, + 480, + 497 + ], + "type": "text", + "content": "30. Mou, C., Wang, X., Xie, L., Zhang, J., Qi, Z., Shan, Y., Qie, X.: T2i-adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models. arXiv preprint arXiv:2302.08453 (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 498, + 480, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 498, + 480, + 520 + ], + "spans": [ + { + "bbox": [ + 132, + 498, + 480, + 520 + ], + "type": "text", + "content": "31. Ni, M., Wu, C., Wang, X., Yin, S., Wang, L., Liu, Z., Duan, N.: Orens: Open-vocabulary responsible visual synthesis. arXiv preprint arXiv:2308.13785 (2023)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 521, + 480, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 521, + 480, + 553 + ], + "spans": [ + { + "bbox": [ + 132, + 521, + 480, + 553 + ], + "type": "text", + "content": "32. Nichol, A., Dhariwal, P., Ramesh, A., Shyam, P., Mishkin, P., McGrew, B., Sutskever, I., Chen, M.: Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741 (2021)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 554, + 480, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 554, + 480, + 597 + ], + "spans": [ + { + "bbox": [ + 132, + 554, + 480, + 597 + ], + "type": "text", + "content": "33. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 598, + 480, + 631 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 598, + 480, + 631 + ], + "spans": [ + { + "bbox": [ + 132, + 598, + 480, + 631 + ], + "type": "text", + "content": "34. Ramesh, A., Dhariwal, P., Nichol, A., Chu, C., Chen, M.: Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125 1(2), 3 (2022)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 132, + 632, + 364, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 364, + 643 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 364, + 643 + ], + "type": "text", + "content": "35. Rombach, R.: Stable diffusion 2.0 release (Nov 2022)" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 132, + 643, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 643, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 643, + 480, + 665 + ], + "type": "text", + "content": "36. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models (2021)" + } + ] + } + ], + "index": 21 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "type": "text", + "content": "Huang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 666 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "text", + "content": "37. Ruiz, N., Li, Y., Jampani, V., Pritch, Y., Rubinstein, M., Aberman, K.: Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In: CVPR (2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 149, + 482, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 149, + 482, + 182 + ], + "spans": [ + { + "bbox": [ + 130, + 149, + 482, + 182 + ], + "type": "text", + "content": "38. Sahara, C., Chan, W., Saxena, S., Li, L., Whang, J., Denton, E.L., Ghasemipour, K., Gontijo Lopes, R., Karagol Ayan, B., Salimans, T., et al.: Photorealistic text-to-image diffusion models with deep language understanding. In: NeurIPS (2022)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 182, + 482, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 182, + 482, + 205 + ], + "spans": [ + { + "bbox": [ + 130, + 182, + 482, + 205 + ], + "type": "text", + "content": "39. Schramowski, P., Brack, M., Deiseroth, B., Kersting, K.: Safe latent diffusion: Mitigating inappropriate degeneration in diffusion models. In: CVPR (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 205, + 482, + 247 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 205, + 482, + 247 + ], + "spans": [ + { + "bbox": [ + 130, + 205, + 482, + 247 + ], + "type": "text", + "content": "40. Schramowski, P., Tauchmann, C., Kersting, K.: Can machines help us answering question 16 in datasheets, and in turn reflecting on inappropriate content? In: Proceedings of the 2022 ACM Conference on Fairness, Accountability, and Transparency. pp. 1350-1361 (2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 247, + 482, + 292 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 247, + 482, + 292 + ], + "spans": [ + { + "bbox": [ + 130, + 247, + 482, + 292 + ], + "type": "text", + "content": "41. Schuhmann, C., Beaumont, R., Vencu, R., Gordon, C., Wightman, R., Cherti, M., Coombes, T., Katta, A., Mullis, C., Wortsman, M., et al.: Laion-5b: An open large-scale dataset for training next generation image-text models. In: NeurIPS (2022)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 293, + 482, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 293, + 482, + 314 + ], + "spans": [ + { + "bbox": [ + 130, + 293, + 482, + 314 + ], + "type": "text", + "content": "42. Sinitsin, A., Plokhotnyuk, V., Pyrkin, D., Popov, S., Babenko, A.: Editable neural networks. arXiv preprint arXiv:2004.00345 (2020)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 314, + 482, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 314, + 482, + 346 + ], + "spans": [ + { + "bbox": [ + 130, + 314, + 482, + 346 + ], + "type": "text", + "content": "43. Somepalli, G., Singla, V., Goldblum, M., Geiping, J., Goldstein, T.: Diffusion art or digital forgery? investigating data replication in diffusion models. In: CVPR (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 346, + 482, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 346, + 482, + 380 + ], + "spans": [ + { + "bbox": [ + 130, + 346, + 482, + 380 + ], + "type": "text", + "content": "44. Somepalli, G., Singla, V., Goldblum, M., Geiping, J., Goldstein, T.: Diffusion art or digital forgery? investigating data replication in diffusion models. In: CVPR (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 380, + 482, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 380, + 482, + 402 + ], + "spans": [ + { + "bbox": [ + 130, + 380, + 482, + 402 + ], + "type": "text", + "content": "45. Song, J., Meng, C., Ermon, S.: Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502 (2020)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 402, + 482, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 402, + 482, + 435 + ], + "spans": [ + { + "bbox": [ + 130, + 402, + 482, + 435 + ], + "type": "text", + "content": "46. Tang, R., Liu, L., Pandey, A., Jiang, Z., Yang, G., Kumar, K., Stenetorp, P., Lin, J., Ture, F.: What the daam: Interpreting stable diffusion using cross attention. arXiv preprint arXiv:2210.04885 (2022)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 435, + 482, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 435, + 482, + 468 + ], + "spans": [ + { + "bbox": [ + 130, + 435, + 482, + 468 + ], + "type": "text", + "content": "47. Tsai, Y.L., Hsu, C.Y., Xie, C., Lin, C.H., Chen, J.Y., Li, B., Chen, P.Y., Yu, C.M., Huang, C.Y.: Ring-a-bell! how reliable are concept removal methods for diffusion models? arXiv preprint arXiv:2310.10012 (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 468, + 482, + 501 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 468, + 482, + 501 + ], + "spans": [ + { + "bbox": [ + 130, + 468, + 482, + 501 + ], + "type": "text", + "content": "48. Wang, Z.J., Montoya, E., Munechika, D., Yang, H., Hoover, B., Chau, D.H.: Diffusiondb: A large-scale prompt gallery dataset for text-to-image generative models. arXiv preprint arXiv:2210.14896 (2022)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 501, + 482, + 522 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 501, + 482, + 522 + ], + "spans": [ + { + "bbox": [ + 130, + 501, + 482, + 522 + ], + "type": "text", + "content": "49. Xiang, C., Bao, F., Li, C., Su, H., Zhu, J.: A closer look at parameter-efficient tuning in diffusion models. arXiv preprint arXiv:2303.18181 (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 522, + 482, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 522, + 482, + 555 + ], + "spans": [ + { + "bbox": [ + 130, + 522, + 482, + 555 + ], + "type": "text", + "content": "50. Yao, Y., Wang, P., Tian, B., Cheng, S., Li, Z., Deng, S., Chen, H., Zhang, N.: Editing large language models: Problems, methods, and opportunities. arXiv preprint arXiv:2305.13172 (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 555, + 482, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 555, + 482, + 588 + ], + "spans": [ + { + "bbox": [ + 130, + 555, + 482, + 588 + ], + "type": "text", + "content": "51. Ye, H., Zhang, J., Liu, S., Han, X., Yang, W.: Ip-adapter: Text compatible image prompt adapter for text-to-image diffusion models. arXiv preprint arXiv:2308.06721 (2023)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 130, + 588, + 482, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 588, + 482, + 610 + ], + "spans": [ + { + "bbox": [ + 130, + 588, + 482, + 610 + ], + "type": "text", + "content": "52. Zhang, E., Wang, K., Xu, X., Wang, Z., Shi, H.: Forget-me-not: Learning to forget in text-to-image diffusion models. arXiv preprint arXiv:2303.17591 (2023)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 130, + 610, + 482, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 610, + 482, + 632 + ], + "spans": [ + { + "bbox": [ + 130, + 610, + 482, + 632 + ], + "type": "text", + "content": "53. Zhang, L., Rao, A., Agrawala, M.: Adding conditional control to text-to-image diffusion models. In: ICCV (2023)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 130, + 632, + 482, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 632, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 632, + 482, + 666 + ], + "type": "text", + "content": "54. Zhang, Y., Jia, J., Chen, X., Chen, A., Zhang, Y., Liu, J., Ding, K., Liu, S.: To generate or not? safety-driven unlearned diffusion models are still easy to generate unsafe images... for now. arXiv preprint arXiv:2310.11868 (2023)" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 416, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 416, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 416, + 91, + 447, + 100 + ], + "type": "text", + "content": "Receler" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 481, + 150 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 481, + 150 + ], + "type": "text", + "content": "55. Zhao, S., Chen, D., Chen, Y.C., Bao, J., Hao, S., Yuan, L., Wong, K.Y.K.: Unictrlnet: All-in-one control to text-to-image diffusion models. arXiv preprint arXiv:2305.16322 (2023)" + } + ] + } + ], + "index": 2 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "type": "text", + "content": "Huang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Reconstruction and Simulation of Elastic Objects with Spring-Mass 3D Gaussians/2eb8194f-eadd-43df-b4b2-bbf94906f9cc_content_list.json b/2024/Reconstruction and Simulation of Elastic Objects with Spring-Mass 3D Gaussians/2eb8194f-eadd-43df-b4b2-bbf94906f9cc_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..410bfd4391a60dbb867cbd50f901aad5e42efd43 --- /dev/null +++ b/2024/Reconstruction and Simulation of Elastic Objects with Spring-Mass 3D Gaussians/2eb8194f-eadd-43df-b4b2-bbf94906f9cc_content_list.json @@ -0,0 +1,1878 @@ +[ + { + "type": "text", + "text": "Reconstruction and Simulation of Elastic Objects with Spring-Mass 3D Gaussians", + "text_level": 1, + "bbox": [ + 220, + 140, + 782, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Licheng Zhong $^{1*}$ , Hong-Xing Yu $^{1}$ , Jiajun Wu $^{1}$ , and Yunzhu Li $^{1,2,3}$", + "bbox": [ + 259, + 210, + 743, + 228 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Stanford University, Stanford, USA", + "bbox": [ + 375, + 238, + 627, + 253 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "2 Columbia University, New York, USA", + "bbox": [ + 367, + 253, + 633, + 267 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3 University of Illinois Urbana-Champaign, Champaign, USA", + "bbox": [ + 295, + 267, + 705, + 281 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Reconstructing and simulating elastic objects from visual observations is crucial for applications in computer vision and robotics. Existing methods, such as 3D Gaussians, model 3D appearance and geometry, but lack the ability to estimate physical properties for objects and simulate them. The core challenge lies in integrating an expressive yet efficient physical dynamics model. We propose Spring-Gaus, a 3D physical object representation for reconstructing and simulating elastic objects from videos of the object from multiple viewpoints. In particular, we develop and integrate a 3D Spring-Mass model into 3D Gaussian kernels, enabling the reconstruction of the visual appearance, shape, and physical dynamics of the object. Our approach enables future prediction and simulation under various initial states and environmental properties. We evaluate Spring-Gaus on both synthetic and real-world datasets, demonstrating accurate reconstruction and simulation of elastic objects. Project page: https://zlicheng.com/spring_gaus.", + "bbox": [ + 259, + 319, + 743, + 527 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: System Identification $\\cdot$ Spring-Mass Model $\\cdot$ 3D Gaussians", + "bbox": [ + 261, + 540, + 740, + 555 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 603, + 375, + 619 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Reconstructing and simulating elastic objects from visual observations poses a fundamental challenge in computer vision and robotics, with applications spanning virtual reality, augmented reality, and robotic manipulation. Accurately modeling the elasticity of objects is crucial for creating immersive experiences and enabling embodied agents to understand and interact with the elastic objects commonly encountered in our daily lives. However, accurately identifying the dynamics from vision still presents considerable challenges.", + "bbox": [ + 212, + 635, + 787, + 739 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Existing methods for dynamic scene reconstruction, such as 3D Gaussians and their dynamic extensions [14, 26, 44, 49], have made significant progress in capturing the temporal changes of the appearance and geometry of objects. However, these methods do not capture the physical properties of the reconstructed", + "bbox": [ + 212, + 741, + 787, + 801 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* The work was done while L. Zhong was a visiting student at Stanford University. L. Zhong is now at Shanghai Jiao Tong University.", + "bbox": [ + 217, + 810, + 785, + 840 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/d477639a8b9254b185e6cd3fc9c49c626e3b8e907ec7ac3f3524e68e64bded53.jpg", + "image_caption": [ + "Fig. 1: Spring-Gaus reconstructs the appearance, geometry, and physical dynamic properties of elastic objects from video observations. Spring-Gaus enables future predictions and simulations under different initial states and environmental conditions." + ], + "image_footnote": [], + "bbox": [ + 217, + 143, + 785, + 340 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "objects; thus, they typically cannot predict the future dynamics of these objects. While a few recent approaches using MPM [11,39] have attempted to integrate physics-based priors into 3D object representations, such as PAC-NeRF [18]. Their ability to handle real, especially heterogeneous, objects is limited, as they assume a known material model and only assigns a global physical parameter to the entire object which restricts its adaptability to real objects. Assigning learnable physical parameters to each particle in an MPM is theoretically possible. However, in practice, it incurs extreme computational costs since MPM requires tens of thousands of dense points. In addition, PAC-NeRF [18] use a implicit grid representation, due to the computational demands, the resolution of the grid is limited, which can lead to a loss of detail in the appearance modeling when using real or noisy data. Thus, the core challenge in reconstructing and predicting object dynamics lies in developing and integrating an expressive and efficient physical model for the dynamics. The physical dynamics model should be expressive enough to capture the motions of elastic objects, including collisions, deformations, and bouncing. It must also be efficient and conducive to inverse parameter estimation through gradient-based optimization.", + "bbox": [ + 212, + 428, + 787, + 684 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this work, we propose Spring-Gaus, a 3D object representation that integrates a 3D Spring-Mass model. Our model represents the elastic object dynamics properties through a learnable system of mass points and springs. Our design is expressive in that it assumes a general and widely applicable physical model class. With a learnable topology and physical parameters, Spring-Gaus can model complex deformation and motion for heterogeneous elastic objects. In addition to expressiveness, Spring-Gaus is also highly efficient for the inverse optimization of physical parameters due to its differentiable nature.", + "bbox": [ + 212, + 686, + 787, + 808 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Spring-Gaus enables reconstructing and simulating elastic objects from sparse multi-view videos (Fig. 1). To overcome the intrinsic difficulty in optimization,", + "bbox": [ + 212, + 809, + 787, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "L. Zhong et al.", + "bbox": [ + 271, + 114, + 372, + 128 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "we propose a reconstruction pipeline that decouples physical parameter reconstruction from appearance and geometry reconstruction. Spring-Gaus requires only a few multi-view videos for physical parameter identification and is robust to the quality of geometry reconstruction.", + "bbox": [ + 212, + 146, + 782, + 205 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We evaluate the effectiveness of Spring-Gaus on both synthetic and real-world datasets, demonstrating its ability to accurately reconstruct and simulate elastic objects. Our Spring-Gaus allows accurate future prediction and simulations under varying initial states and environmental parameters, showcasing its potential for applications in predictive visual perception and immersive experiences. In summary, the main contributions of this work are threefold:", + "bbox": [ + 212, + 207, + 784, + 296 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose Spring-Gaus, which incorporates an expressive yet efficient 3D Spring-Mass model for reconstructing and simulating elastic objects.", + "- We introduce a pipeline to reconstruct Spring-Gaus from multi-view videos of the object. We decouple the appearance and geometry reconstruction from the physical dynamics reconstruction for more effective optimization.", + "- We demonstrate the effectiveness of Spring-Gaus on both synthetic and real-world datasets, showcasing accurate reconstruction and simulation of elastic objects. This includes capabilities for future prediction and simulation under varying initial configurations." + ], + "bbox": [ + 225, + 305, + 782, + 438 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 462, + 385, + 478 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3D Object Representations: Traditionally, 3D objects are often represented by point clouds, meshes, and voxels. Recently, neural 3D object representation has become popular due to the efficiency and flexibility. For example, Scene Representation Networks (SRNs) [38] and DeepSDF [29] represent significant advancements in 3D scene and object representation, treating 3D objects as continuous functions that map world coordinates to a feature representation of local object properties. Over the past few years, NeRF [28] and its successors [1, 2, 25, 47, 50] have demonstrated the efficacy of neural networks in capturing continuous 3D scenes and objects through implicit representation. DirectVoxGO [40] accelerates NeRF's approach by substituting the MLP with a voxel grid. Furthermore, 3D Gaussian Splatting [14] has emerged as a method for real-time differentiable rendering, representing scenes and objects with 3D Gaussians. Extending this approach, DreamGaussian [41] applies 3D Gaussians for 3D object generation. Unlike these approaches, our method focuses on reconstructing simulatable 3D objects. Recently, PhysGaussian [46] integrated physical simulation into 3D Gaussians using a customized Material Point Method, allowing forward simulation of reconstructed objects. In contrast to PhysGaussian, our work focuses on system identification from raw videos and supports both the forward simulation and the inverse reconstruction of physical objects.", + "bbox": [ + 212, + 492, + 784, + 777 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Dynamic Scene Reconstruction: The modeling of dynamic scenes has seen significant progress with the adoption of NeRF [8, 10, 18, 20, 22, 23, 30-32, 42, 43] and 3D Gaussian [12, 16, 24, 26, 44, 48, 49, 51] representations. D-NeRF [32] introduces an extension to NeRF, capable of modeling dynamic scenes from monocu", + "bbox": [ + 212, + 779, + 784, + 839 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "Spring-Gaus", + "bbox": [ + 645, + 114, + 730, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "lar views by optimizing an underlying deformable volumetric function. Furthermore, Dynamic 3D Gaussians [26] focus on optimizing the motion of Gaussian kernels for each frame, presenting an efficient method to capture scene dynamics. Deformable 3D Gaussians [49] propose a novel approach by learning Gaussian distributions in a canonical space, complemented with a deformation field for modeling monocular dynamic scenes. Meanwhile, 4D Gaussian Splatting [44] introduces a hybrid model combining 3D Gaussians with 4D neural voxels. PAC-NeRF [18] delves into the integration of Lagrangian particle simulation with Eulerian scene representation, exploring a new aspect of scene dynamics. Among them, PAC-NeRF [18] is the most relevant to our work. PAC-NeRF learns physical parameters used in the Material Point Method (MPM) for better dynamic reconstruction and prediction. However, PAC-NeRF assumes known material models that restrict its adaptability and applicability to complex real objects. In contrast, our work integrates a 3D Spring-Mass model that is both expressive and efficient, allowing reconstructing and simulating real elastic objects.", + "bbox": [ + 212, + 146, + 787, + 372 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Physics-Informed Learning: Physics-Informed Learning has emerged as a prominent research direction since the introduction of Physics-Informed Neural Networks (PINNs) [33]. Li et al. [21] and Sanchez-Gonzalez et al. [34] introduced Graph Network-based simulators within a machine learning framework. INSR-PDE [3] tackles time-dependent partial differential equations (PDEs) using implicit neural spatial representations. NCLaw [27] focuses on learning neural constitutive laws for PDE dynamics. DiffPD [9] presents a differentiable soft-body simulator. Li et al. [19] have made strides in learning preconditioners for conjugate gradient PDE solvers. Chu et al. [4] and Yu et al. [52] modeled smoke in neural density and velocity fields. Neural Flow Maps [7] integrate fluid simulation with neural implicit representations. Deng et al. [6] introduced a novel differentiable vortex particle method for fluid dynamics inference. DANO [17] is the most relevant work to ours. In particular, DANO develops a differentiable simulation for rigid objects represented by NeRFs, allowing reconstruction and simulation from a few videos. In contrast, our work focuses on elastic objects, which involve a fundamentally new set of challenges and require a redesign of the physical models.", + "bbox": [ + 212, + 373, + 787, + 630 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3 Approach", + "text_level": 1, + "bbox": [ + 215, + 652, + 346, + 672 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We propose Spring-Gaus that integrates a 3D Spring-Mass model with 3D Gaussians for elastic object reconstruction and simulation. Specifically, we are given a set of $O$ calibrated videos from multiple viewpoints of an elastic object in motion, denoted as $\\{I^{o,f}\\}_{o = 1,f = 1}^{O,F}$ , where $F$ is the number of video frames and $I^{o,f}$ denotes a 2D image at video $o$ and frame $f$ . Our goal is to reconstruct the appearance, geometry, and physical dynamics parameters of Spring-Gaus.", + "bbox": [ + 212, + 686, + 785, + 779 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In the following, we first introduce the representation for the static properties, i.e., appearance and geometry, and then the dynamics properties, i.e., the 3D Spring-Mass model. Then we describe our reconstruction pipeline. We show an overview of our reconstruction pipeline in Fig. 2.", + "bbox": [ + 212, + 780, + 785, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "L. Zhong et al.", + "bbox": [ + 271, + 114, + 370, + 128 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1 Appearance and Geometry Representation", + "text_level": 1, + "bbox": [ + 215, + 146, + 614, + 162 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Following 3D Gaussians Splatting [14], we use a set of 3D Gaussian kernels to represent the appearance and shape of an object. These kernels are described by a full 3D covariance matrix $\\Sigma$ , defined in world space and centered at point $\\pmb{\\mu}$ :", + "bbox": [ + 212, + 169, + 782, + 215 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nG (\\boldsymbol {\\tau}) = e ^ {- \\frac {1}{2} (\\boldsymbol {\\tau}) ^ {T} \\Sigma^ {- 1} (\\boldsymbol {\\tau})}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 419, + 224, + 785, + 244 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\pmb{\\mu}$ is the mean of the Gaussian distribution, and $\\pmb{\\tau}$ is the independent variable of the Gaussian distribution. In our formulation, we assume all Gaussian kernels are isotropic, the covariance matrix $\\Sigma$ can be controlled by a scalar $s\\in \\mathbb{R}$ .", + "bbox": [ + 212, + 253, + 782, + 297 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The splatting process, designed to render Gaussian kernels into a 2D image, involves two main steps: (1) projecting the Gaussian kernels into camera space, and (2) rendering the projected kernels into image space. The projection process is defined as:", + "bbox": [ + 212, + 299, + 782, + 356 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\Sigma^ {\\prime} = J W \\Sigma W ^ {T} J ^ {T}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 434, + 358, + 784, + 376 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\Sigma'$ represents the covariance matrix in camera coordinates, $W$ is the transformation matrix, and $J$ denotes the Jacobian of the affine approximation of the projective transformation. The color $C$ of each pixel is rendered as:", + "bbox": [ + 212, + 380, + 782, + 426 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nC = \\sum_ {i \\in N} T _ {i} \\alpha_ {i} \\boldsymbol {c} _ {i}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 444, + 436, + 784, + 467 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $N$ is the total number of kernels and $T_{i}$ represents the transmittance, defined as $\\Pi_{j = 1}^{i - 1}\\alpha_{j}$ . The term $\\alpha_{i}$ denotes the alpha value for each Gaussian, which is calculated using the expression $1 - e^{-\\sigma_i\\delta_i}$ , where $\\sigma_{i}$ is the opacity factor. Additionally, $\\pmb{c}_i$ refers to the color of the Gaussian along the ray within the interval $\\delta_{i}$ .", + "bbox": [ + 212, + 474, + 782, + 551 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For each kernel, the learnable parameters include a point center $\\pmb{\\mu}$ , a scaling scalar $s$ , a color vector $\\pmb{c}$ , and an opacity value $\\sigma$ .", + "bbox": [ + 212, + 553, + 782, + 584 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2 Physics-Based Dynamics with 3D Spring-Mass Model", + "text_level": 1, + "bbox": [ + 214, + 604, + 702, + 619 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We aim to develop and integrate an expressive yet efficient physics-based dynamics model. With these two design goals in mind, we introduce the 3D Spring-Mass model which represents the elastic object dynamics by a learnable spring-mass system. Our Spring-Gaus does not assume any material type, and it is expressive to model real heterogeneous elastic objects. Besides being expressive, our dynamics model should also be efficient and amenable to gradient-based optimization for parameter estimation.", + "bbox": [ + 212, + 628, + 782, + 733 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Recall that the appearance and geometry representations include a set of Gaussian kernels $\\{G_i\\}_{i=1}^N$ , parameterized by point centers $\\mathbf{X} = \\{\\pmb{\\mu}_i\\}_{i=1}^N$ , scaling scalars $\\{s_i\\}_{i=1}^N$ , color vectors $\\{\\pmb{c}_i\\}_{i=1}^N$ , and opacity values $\\{\\sigma_i\\}_{i=1}^N$ . To manage complexity for efficient simulation, we introduce volume sampling to generate a set of anchor points $\\mathbf{A} = \\{\\pmb{x}_i\\}_{i=1}^{N_A}$ (each $\\pmb{x}$ represents a spatial point), defined by:", + "bbox": [ + 212, + 734, + 782, + 810 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {A} = \\left\\{\\boldsymbol {x} _ {i} \\right\\} _ {i = 1} ^ {N _ {A}} = \\mathcal {V} (\\boldsymbol {X}), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 423, + 823, + 784, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "Spring-Gaus", + "bbox": [ + 645, + 114, + 730, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/dc0f7fb28cd35083fb30769d417e6b17eb077004dc422f078951c8e33f244093.jpg", + "image_caption": [ + "Fig.2: Overview of Spring-Gaus reconstruction pipeline: (a) Static Scene Reconstruction: We start by reconstructing static 3D Gaussians from the first frames of the multiview videos. (b) Refining 3D Gaussians: We extract a set of anchor points to allow efficient simulation, which leads to appearance drift. We refine the 3D Gaussians to better model the appearance during simulation. (c) Dynamic Reconstruction: Our 3D Spring-Mass model simulates anchor points and updates the positions of Gaussian kernels. Upon completion of optimization, we obtain a simulatable 3D object that accurately models its dynamics." + ], + "image_footnote": [], + "bbox": [ + 223, + 146, + 781, + 407 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $N_{A}$ denotes the number of anchors, and $\\nu$ denotes the sampling function.", + "bbox": [ + 212, + 577, + 784, + 592 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Our 3D Spring-Mass physical model can simulate the motion of anchor points $\\mathbf{A}$ , assuming each anchor has a mass $m_{i}$ and an initial velocity $\\mathbf{v}_{i}$ . Each anchor $\\mathbf{x}_{i}$ connects to its $n_{k}$ nearest neighbors $\\mathbf{N}_{i} = \\{\\mathbf{x}_{i,j}\\}_{j=1}^{n_{k}}$ through springs $\\mathbf{L}$ :", + "bbox": [ + 212, + 593, + 785, + 640 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {L} = \\left\\{l _ {i, j} \\right\\} _ {i = 1, j = 1} ^ {N _ {A}, n _ {k}} = \\operatorname {k n n} (\\boldsymbol {A}, \\boldsymbol {A}, n _ {k}), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 374, + 650, + 784, + 671 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $l_{i,j}$ denotes the spring's length between $\\pmb{x}_i$ and $\\pmb{x}_{i,j}$ , and knn denotes the k-nearest neighbors function. Each spring is characterized by a stiffness $k_{i,j}$ and a damping factor $\\zeta_{i,j}$ .", + "bbox": [ + 212, + 680, + 785, + 724 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To update the positions of the kernels, we first measure the distance between each kernel center and its $n_b$ nearest anchors at the dynamic simulation's onset:", + "bbox": [ + 212, + 726, + 785, + 756 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\left\\{d _ {i, j} \\right\\} _ {i = 1, j = 1} ^ {N, n _ {b}} = \\operatorname {k n n} (\\boldsymbol {X}, \\boldsymbol {A}, n _ {b}). \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 390, + 768, + 784, + 789 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For each timestep $t$ , the forces $\\pmb{F}_i^t$ acting on each anchor point $\\pmb{x}_i^t$ are calculated as follows:", + "bbox": [ + 212, + 794, + 785, + 821 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {F} _ {i} ^ {t} = \\boldsymbol {F} _ {\\boldsymbol {k} i} ^ {t} + \\boldsymbol {F} _ {\\boldsymbol {\\zeta} i} ^ {t} + m _ {i} \\boldsymbol {g}, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 415, + 823, + 784, + 842 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "L. Zhong et al.", + "bbox": [ + 271, + 114, + 372, + 128 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\pmb{F}_{\\pmb{k}_i}^t, \\pmb{F}_{\\pmb{\\zeta}_i}^t$ , and $\\pmb{g}$ represent the spring force, the damping force, and gravitational acceleration, respectively.", + "bbox": [ + 212, + 145, + 782, + 176 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Then, for each spring $\\pmb{L}_{i,j}$ , the spring force $\\pmb{F}_{\\pmb{k}_{i,j}}^{t}$ and damping force $\\pmb{F}_{\\pmb{\\zeta}_{i,j}}^{t}$ are determined by:", + "bbox": [ + 212, + 176, + 784, + 205 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {F} _ {\\boldsymbol {k} _ {i, j}} ^ {t} = - \\eta_ {j} \\cdot k _ {i, j} \\left(\\left\\| \\boldsymbol {x} _ {i} ^ {t} - \\boldsymbol {x} _ {i, j} ^ {t} \\right\\| - l _ {i, j}\\right) \\frac {\\boldsymbol {x} _ {i} ^ {t} - \\boldsymbol {x} _ {i , j} ^ {t}}{\\left\\| \\boldsymbol {x} _ {i} ^ {t} - \\boldsymbol {x} _ {i , j} ^ {t} \\right\\|} \\cdot \\left| \\left\\| \\boldsymbol {x} _ {i} ^ {t} - \\boldsymbol {x} _ {i, j} ^ {t} \\right\\| - l _ {i, j} \\right| ^ {p _ {k}}, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 235, + 218, + 782, + 253 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {F} _ {\\zeta_ {i, j}} ^ {t} = \\left(- \\zeta_ {i, j} \\left(\\boldsymbol {v} _ {i} ^ {t} - \\boldsymbol {v} _ {i, j} ^ {t}\\right) \\frac {\\boldsymbol {x} _ {i} ^ {t} - \\boldsymbol {x} _ {i , j} ^ {t}}{\\left\\| \\boldsymbol {x} _ {i} ^ {t} - \\boldsymbol {x} _ {i , j} ^ {t} \\right\\|}\\right) \\cdot \\frac {\\boldsymbol {x} _ {i} ^ {t} - \\boldsymbol {x} _ {i , j} ^ {t}}{\\left\\| \\boldsymbol {x} _ {i} ^ {t} - \\boldsymbol {x} _ {i , j} ^ {t} \\right\\|}, \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 308, + 257, + 782, + 296 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $\\eta$ is a soft vector which will be discussed later and $p_k$ is a hyperparameter that determines the nonlinearity of the spring force. When $p_k$ is set to 0, Eq. (8) becomes Hooke's law, and for positive values of $p_k$ , the spring force becomes a nonlinear function of the spring's length. The cumulative forces acting on each anchor point $\\boldsymbol{x}_i^t$ are expressed as:", + "bbox": [ + 212, + 306, + 782, + 381 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {F} _ {i} ^ {t} = \\sum_ {j = 1} ^ {n _ {k}} \\boldsymbol {F} _ {\\boldsymbol {k} _ {i, j}} ^ {t} + \\sum_ {j = 1} ^ {n _ {k}} \\boldsymbol {F} _ {\\zeta_ {i, j}} ^ {t} + m _ {i} \\boldsymbol {g}. \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 383, + 393, + 782, + 433 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Anchor points $\\mathbf{A}$ 's positions and velocities are updated using semi-implicit Euler integration:", + "bbox": [ + 214, + 443, + 782, + 472 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\boldsymbol {v}} _ {i} ^ {t + 1} = \\boldsymbol {v} _ {i} ^ {t} + \\frac {\\boldsymbol {F} _ {i} ^ {t}}{m _ {i}} \\Delta t, \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 431, + 484, + 782, + 515 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\boldsymbol {x}} _ {i} ^ {t + 1} = \\boldsymbol {x} _ {i} ^ {t} + \\boldsymbol {v} _ {i} ^ {t + 1} \\Delta t, \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 426, + 517, + 782, + 535 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "and a boundary condition $\\mathcal{B}$ is applied to the anchor points $\\pmb{A}$ to model the interactions with the environment:", + "bbox": [ + 214, + 546, + 782, + 575 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {x} _ {i} ^ {t + 1}, \\boldsymbol {v} _ {i} ^ {t + 1} = \\mathcal {B} \\left(\\hat {\\boldsymbol {x}} _ {i} ^ {t + 1}, \\hat {\\boldsymbol {v}} _ {i} ^ {t + 1}\\right). \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 403, + 587, + 782, + 604 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The position of each Gaussian kernel $\\pmb{\\mu}_i$ is updated through Inverse Distance Weighting (IDW) interpolation to reflect the dynamic changes accurately:", + "bbox": [ + 214, + 614, + 782, + 646 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {\\mu} _ {i} ^ {t + 1} = \\frac {\\sum_ {j = 1} ^ {n _ {b}} \\boldsymbol {x} _ {i , j} ^ {t + 1} \\cdot \\left(1 / \\left(d _ {i , j}\\right) ^ {p _ {b}}\\right)}{\\sum_ {j = 1} ^ {n _ {b}} \\left(1 / \\left(d _ {i , j}\\right) ^ {p _ {b}}\\right)}, \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 385, + 657, + 782, + 694 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $p_b$ is a positive real number that determines the diminishing influence of anchor points with distance. To render the image $\\hat{I}$ at a specific camera and frame, we use the updated positions of the Gaussian kernels following the rendering equation Eq. (3).", + "bbox": [ + 212, + 704, + 782, + 763 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Soft Vector for Springs Connection: In our formulation above, the $n_k$ is a hyperparameter which is the number of connected springs for each anchor. However, the choice of $n_k$ will directly affect the simulation results markedly. The bigger value of $n_k$ , the object will behave more rigidly, and a smaller value of $n_k$ leads to a noticeably softer behavior of the point cloud. To address the", + "bbox": [ + 212, + 765, + 784, + 839 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "Spring-Gaus", + "bbox": [ + 645, + 114, + 730, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "significant impact that the value of $n_k$ has on the simulation, we introduce a mitigation strategy by applying a soft vector $\\eta = [\\eta_0, \\eta_1, \\dots, \\eta_{n_k}]$ . This vector controlled by a learnable parameter $\\kappa$ (shared by all anchors) is used to modulate the number of connected springs, thereby adjusting the system's response to a different object. Given an empirical value $n_c$ , the soft vector $\\eta$ is calculated as:", + "bbox": [ + 212, + 146, + 787, + 223 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\eta_ {j} = \\left\\{ \\begin{array}{l l} 1 & j \\leq n _ {c}, \\\\ \\operatorname {c l a m p} (2 - \\exp (\\operatorname {s o f t p l u s} (\\kappa)) ^ {j - n _ {c}}, 0, 1) & n _ {c} < j \\leq n _ {k}. \\end{array} \\right. \\tag {15}\n$$\n", + "text_format": "latex", + "bbox": [ + 290, + 232, + 785, + 273 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "3.3 Optimization", + "text_level": 1, + "bbox": [ + 215, + 290, + 372, + 306 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To allow efficient optimization, we simplify our model by reducing the number of learnable parameters without changing the essential expressiveness. In our simplified approach, we standardize the mass of every anchor to a constant value $m_0$ , and control all damping factors using a singular parameter $\\zeta_0$ . These two parameters are fixed, eliminating their variability from the optimization process. Furthermore, we introduce a unified parameter $k_{i}$ for each anchor $\\pmb{x}_i$ to control the spring stiffness of the springs connected to it, simplifying the model without compromising its functional integrity. The spring stiffness $k_{i,j}$ and damping factor $\\zeta_{i,j}$ are thus given as follows:", + "bbox": [ + 212, + 314, + 787, + 450 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\nk _ {i, j} = k _ {i} / l _ {i, j}, \\tag {16}\n$$\n", + "text_format": "latex", + "bbox": [ + 455, + 462, + 784, + 477 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\zeta_ {i, j} = \\zeta_ {0} / l _ {i, j}. \\tag {17}\n$$\n", + "text_format": "latex", + "bbox": [ + 455, + 482, + 784, + 497 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Note that, in Eq. (17) stiffness is defined at anchor points rather than on springs themselves to simplify the optimization—we only need to optimize $N_{A}$ stiffness coefficients instead of all the springs, which is on the order of $n_k \\cdot N_A$ . By this simplification, our model maintains computational efficiency and ease of optimization, while still capturing the essential dynamics of the object.", + "bbox": [ + 212, + 506, + 787, + 580 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Summarized, the learnable parameters in our model now include:", + "bbox": [ + 238, + 582, + 712, + 597 + ], + "page_idx": 7 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- $\\pmb{v}_0$ : the initial velocity vector, providing a baseline movement pattern for the simulation;", + "- $\\{k_i\\}_{i=1}^{N_A}$ : the individual stiffness parameters for each anchor, allowing for localized adjustments to spring stiffness;", + "- $\\kappa$ : a parameter governing the modulation of the soft vector, facilitating finetuned control over the spring dynamics;", + "- $\\Theta(\\mathcal{B}(\\cdot))$ : parameters defining the boundary conditions ( $\\Theta$ is learnable parameters), such as the friction coefficient, which influence the simulation's physical realism." + ], + "bbox": [ + 225, + 604, + 784, + 739 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "There exist $n_t$ timesteps between each of the two keyframes. At each timestep, $\\boldsymbol{x}^t$ is computed from $\\boldsymbol{x}^{t-1}$ using the update rule Eqs. (11) and (12). We only optimize physical parameters at each keyframe (when $t = 0, n_t, 2n_t, 3n_t, \\ldots$ ) based on the visual observations until that time. We increase the value of $n_t$ as the parameters converge. This approach balances identification accuracy and computational demand.", + "bbox": [ + 212, + 750, + 787, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "L. Zhong et al.", + "bbox": [ + 271, + 114, + 372, + 128 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Same with 3D Gaussian Splatting [14], we define our loss function as a weighted combination of the $\\mathcal{L}_1$ norm and the Structural Similarity Index Measure (D-SSIM) $\\mathcal{L}_{\\mathrm{d - ssim}}$ , applied between the input images $I$ and the rendered images $\\hat{I}$ . This is formally expressed as:", + "bbox": [ + 212, + 146, + 787, + 207 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\left(1 - \\lambda_ {\\mathrm {d - s s i m}}\\right) \\mathcal {L} _ {1} + \\lambda_ {\\mathrm {d - s s i m}} \\mathcal {L} _ {\\mathrm {d - s s i m}}, \\tag {18}\n$$\n", + "text_format": "latex", + "bbox": [ + 372, + 223, + 785, + 238 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where $\\lambda_{\\mathrm{d - ssim}}$ is a weighting coefficient that balances the contribution of the $\\mathcal{L}_1$ norm and the D-SSIM term to the overall loss. We decouple our optimization into a few stages, as illustrated in Fig. 2.", + "bbox": [ + 212, + 244, + 785, + 290 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Static Reconstruction: Our optimization starts by taking the first frames of the multi-view videos and using them to reconstruct the appearance and geometry of the object.", + "bbox": [ + 212, + 290, + 785, + 335 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "3D Gaussians Refinement: During dynamic simulation of the anchor points, the 3D Gaussian centers $\\mathbf{X}_0 = \\{\\pmb{\\mu}_i^0\\}_{i=1}^N$ are computed by IDW interpolation outlined in Eq. (14). This leads to a slight appearance drift from the static reconstruction. Therefore, we refine the parameters of the Gaussian kernels—scaling scalar $s$ , color vector $c$ , and opacity value $\\sigma$ , excluding points center $\\pmb{\\mu}$ —also at the first frame, before the dynamic simulation. Since the Gaussian kernels' position from IDW interpolation is slightly different from the static reconstruction results, this refinement process enables the Gaussian kernels with the interpolated spatial position to render the correct object appearance.", + "bbox": [ + 212, + 335, + 787, + 470 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Dynamic Reconstruction: Since our efficient dynamics simulation is fully differentiable with respect to the learnable parameters, we can optimize the physical parameters through differentiable simulation and differentiable rendering with our loss function $\\mathcal{L}$ .", + "bbox": [ + 212, + 472, + 787, + 531 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4 Experiment", + "text_level": 1, + "bbox": [ + 215, + 556, + 366, + 571 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.1 Datasets", + "text_level": 1, + "bbox": [ + 215, + 587, + 334, + 601 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Synthetic Data. We evaluate Spring-Gaus using a synthetic dataset. We first collected fourteen 3D models, including point clouds and meshes, to use as initial point clouds for our simulations. Some of them are sourced from PAC-NeRF [18] and OmniObject3D [45]. Following PAC-NeRF, we employ the Material Point Method [11,39] to simulate the dynamics of elastic objects to generate synthetic data. Our dataset features elastic objects with various stiffness levels and diverse geometric forms. Multi-view RGB videos are rendered using Blender [5], with each sequence comprising 10 viewpoints and 30 frames at a resolution of $512 \\times 512$ . The initial 20 frames are utilized for dynamic reconstruction, while the subsequent 10 frames are dedicated to evaluating future prediction performance.", + "bbox": [ + 212, + 612, + 787, + 763 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Real-World Data. In addition to synthetic datasets, we further assess Spring-Gaus using captured real-world examples. Our collection process distinguishes between static scenes and dynamic multi-view videos. For static scenes, we position each object on a table and capture 50-70 images from various viewpoints within the upper hemisphere surrounding the object. To obtain camera", + "bbox": [ + 212, + 765, + 787, + 840 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "Spring-Gaus", + "bbox": [ + 645, + 114, + 730, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/f94c41c50fd8443bb7e3cbda1bfb71df2f2b25a9084a735fa1dc61606ee2d19d.jpg", + "image_caption": [ + "Fig. 3: Registration from static scene to dynamic scene for real-world sample." + ], + "image_footnote": [], + "bbox": [ + 285, + 143, + 725, + 247 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "poses for static scenes, we utilize the off-the-shelf Structure-from-Motion toolkit COLMAP [35, 36]. The dynamic aspect of our dataset is represented through multi-view RGB videos, recorded from three distinct viewpoints, at a resolution of $1920 \\times 1080$ . The camera parameters for dynamic scenes are obtained through calibration using a checkerboard. SAM [15] is used to obtain object masks.", + "bbox": [ + 212, + 310, + 787, + 387 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.2 Implementation Details", + "text_level": 1, + "bbox": [ + 215, + 406, + 455, + 422 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "When learning the 3D Gaussians, the distribution of Gaussian kernels is highly dependent on the initial points. Gaussian kernels tend to concentrate on the surface if the initial points are derived from SfM results. However, we initialize a large number of points inside a cube, resulting in a more uniform distribution of the final kernels.", + "bbox": [ + 212, + 428, + 787, + 502 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "For real-world samples, we collect static and dynamic scenes separately because three viewpoints are insufficient for effective 3D Gaussian reconstruction [14]. This results in the Gaussian kernels for static and dynamic scenes being in different coordinate systems. Therefore, for real samples, we employ a registration network before dynamic reconstruction to align the Gaussian kernels from static scene coordinates to dynamic scene coordinates, as shown in Fig. 3. Specifically, we optimize a scale factor $s_r$ , a translation vector $\\pmb{t}_r$ , and a rotation vector $\\pmb{r}_r$ for the registration. We represent 3D rotations using a continuous 6D vector $\\pmb{r}_r \\in \\mathbb{R}^6$ , which has been shown to be more amenable for gradient-based optimization [53]. However, slight deformations of objects and variations in lighting conditions and exposure times during data capture are inevitable. These variations can result in color discrepancies between frames across time and viewpoints. The color discrepancy is the major noise source that hinders registration and reconstruction. Thus, we refine our approach by computing the loss function between mask images, incorporating both mask center loss and perceptual loss into our model. Consequently, the revised loss function for real samples is expressed as:", + "bbox": [ + 212, + 503, + 787, + 762 + ], + "page_idx": 9 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\left(1 - \\lambda_ {\\mathrm {d - s s i m}}\\right) \\mathcal {L} _ {1} + \\lambda_ {\\mathrm {d - s s i m}} \\mathcal {L} _ {\\mathrm {d - s s i m}} + \\lambda_ {\\text {c e n t e r}} \\mathcal {L} _ {\\text {c e n t e r}} + \\lambda_ {\\text {p e r c e p}} \\mathcal {L} _ {\\text {p e r c e p}}, \\tag {19}\n$$\n", + "text_format": "latex", + "bbox": [ + 240, + 785, + 785, + 801 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "where $\\lambda_{\\mathrm{d - ssim}} = 0.8$ , $\\lambda_{\\mathrm{center}} = 1.0$ , and $\\lambda_{\\mathrm{percep}} = 0.1$ are the weighting coefficients. Here, $\\mathcal{L}_{\\mathrm{center}}$ quantifies the discrepancy between the center coordinates of", + "bbox": [ + 212, + 809, + 787, + 840 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "L. Zhong et al.", + "bbox": [ + 271, + 114, + 372, + 128 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/66cb43f6edbce48e9fde54d4f86a42cd10259a064ecfa80dc2bfaa63701528a1.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
TorusCrossCreamApplePasteChessBananaMean
CD↓Spring-Gaus (ours)2.381.572.221.877.032.5918.485.16
PAC-NeRF [18]2.473.872.214.6937.708.2066.4317.94
EMD↓Spring-Gaus (ours)0.0870.0510.0940.0760.1260.0950.1350.095
PAC-NeRF [18]0.0550.1110.0830.1080.1920.1550.2340.134
PSNR↑Spring-Gaus (ours)16.8316.9315.4221.5514.7116.0817.8917.06
PAC-NeRF [18]17.4614.1515.3719.9412.3215.0816.0415.77
SSIM↑Spring-Gaus (ours)0.9190.9400.8620.9020.8720.8810.9040.897
PAC-NeRF [18]0.9130.9060.8580.8780.8190.8480.8660.870
", + "bbox": [ + 217, + 143, + 782, + 275 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/4e8a200bb47f559c2de5432b2ae6b223c85ce1e1429e28f1916e77b942097618.jpg", + "table_caption": [ + "Table 1: Quantitative results of future prediction on synthetic data. Spring-Gaus excels in short-term future prediction. Meanwhile, since we separate appearance and dynamics modeling, Spring-Gaus also maintains good rendering quality." + ], + "table_footnote": [], + "table_body": "
TorusCrossCreamApplePasteChessBananaMean
CD↓Spring-Gaus (ours)0.170.480.360.380.191.802.600.85
PAC-NeRF [18]4.921.100.771.113.140.962.772.11
Dy-Gaus [26]579773479727284976429631305
4D-Gaus [44]11.121.772.872.231.953.977.134.43
EMD↓Spring-Gaus (ours)0.0400.0370.0310.0330.0220.0630.0520.040
PAC-NeRF [18]0.0560.0520.0410.0450.0540.0520.0620.052
Dy-Gaus [26]0.8570.9550.7830.9031.7390.9851.5911.116
4D-Gaus [44]0.1300.0780.0890.0880.0700.0970.1120.095
", + "bbox": [ + 225, + 344, + 771, + 474 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Table 2: Quantitative results of dynamic reconstruction on synthetic data. Spring-Gaus has excellent geometric accuracy in dynamic reconstruction.", + "bbox": [ + 215, + 476, + 782, + 502 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "the rendered and ground truth images, while $\\mathcal{L}_{\\mathrm{percep}}$ represents the perceptual loss [13], which is based on the VGG16 architecture [37].", + "bbox": [ + 215, + 546, + 784, + 575 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4.3 Qualitative and Quantitative Results", + "text_level": 1, + "bbox": [ + 217, + 601, + 563, + 616 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "We assess the performance of our approach using both synthetic and real-world datasets. Firstly, we validate the effectiveness of our method on synthetic datasets, utilizing the first 20 frames as our observation set for training dynamic modeling capabilities. For future prediction, we employ the subsequent 10 frames, comparing the ground truth with our model's predictions of future frames. Additionally, we benchmark our approach against the most relevant methods in dynamic scene modeling and physics-informed learning, including PAC-NeRF [18], Dynamic 3D Gaussians [26], and 4D Gaussian Splatting [44].", + "bbox": [ + 215, + 628, + 784, + 748 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "The qualitative results are presented in Fig. 4. We also report quantitative results by computing the Chamfer Distance (CD) and Earth Mover's Distance (EMD). In all tables, the Chamfer Distance is measured based on squared distance, with units expressed as $1 \\times 10^{3}mm^{2}$ . The quantitative analysis of dynamic reconstruction, shown in Tab. 2, reveals that Spring-Gaus can accurately simulate object dynamics for the synthetic data.", + "bbox": [ + 215, + 750, + 784, + 839 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "Spring-Gaus", + "bbox": [ + 645, + 114, + 730, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/34cf5ad2c67f59e147f820c12520acb53b250549f8d061319f616d4a63d3a380.jpg", + "image_caption": [ + "Fig.4: Qualitative results on synthetic data. Compared with PAC-NeRF [18], Dynamic 3D Gaussians [26] and 4D Gaussian Splatting [44], Spring-Gaus can maintain a good geometry and appearance while reconstructing reasonable dynamics." + ], + "image_footnote": [], + "bbox": [ + 217, + 143, + 787, + 720 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "L. Zhong et al.", + "bbox": [ + 271, + 114, + 372, + 128 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/27a05d1c66777d2fd18a08deff01dff3718388705359443008465c2f3b43667c.jpg", + "image_caption": [ + "Fig. 5: Qualitative results of future prediction on real-world samples. Predicted dynamics closely follow real observations." + ], + "image_footnote": [], + "bbox": [ + 218, + 142, + 501, + 273 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/8d99da0f958e146ff55a5265e9359c90e49105e0361bde02951c869b35470d31.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 501, + 142, + 651, + 273 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/5bfb97ef473387211a109ed2e243fb1d4a2ad13dfbcccd07566af81eb1cc0427.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 651, + 142, + 787, + 273 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "In Tab. 1, we show our method's capability in predicting future frames. Our method outperforms PAC-NeRF across CD, EMD, Peak Signal-to-Noise Ratio (PSNR), and Structural Similarity Index (SSIM) metrics. We show qualitative results in Fig. 4. We observe that Spring-Gaus gives faithful reconstruction and future prediction that closely aligns with the ground truth, even though the simulation engines are different. This shows that Spring-Gaus is not only expressive but also allows efficient identification through gradient-based optimization.", + "bbox": [ + 212, + 364, + 784, + 470 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "We also evaluate Spring-Gaus on real-world samples. It is difficult for both NeRF [28] and 3D Gaussian Splatting [14] to reconstruct the correct geometric information under extremely sparse camera views. However, due to 3D Gaussians' explicit representation that we can directly operate Gaussian kernels, we could reconstruct the static scenes and dynamic scenes in a different coordinate system and then align them using a registration network mentioned in Sec. 4.2, which is hard to do under an implicit representation. We show the registration process in Fig. 3. Results on real-world samples can be found in Fig. 5 and Fig. 1.", + "bbox": [ + 212, + 474, + 785, + 595 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "4.4 Generalization to New Conditions", + "text_level": 1, + "bbox": [ + 215, + 636, + 542, + 651 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "In addition to future prediction, we also show that our method essentially creates a digital asset of the object from the multi-view videos, allowing dynamic simulation under different unseen environmental conditions. In Fig. 6, we edit the boundary conditions, such as adjusting the positions and the stickiness of the ground plane. In Fig. 7, we edit physical conditions, initialization conditions, and environmental conditions, such as object properties (making them softer or harder), initial velocities, and environmental gravity. Please check our project website* for videos of the results for a more expressive illustration.", + "bbox": [ + 212, + 680, + 787, + 801 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "Spring-Gaus", + "bbox": [ + 645, + 114, + 730, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "page_footnote", + "text": "* https://zlicheng.com/spring_gaus", + "bbox": [ + 217, + 824, + 482, + 839 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/9f337b1a4dfdecbd32704379c1f3ec4bd5c7167f08bf89dff96b245df825ec26.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 217, + 143, + 495, + 215 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/d001745f6a0c502ff1a4cddedd883c09c35e7bec2847e6c7057958dafb16b536.jpg", + "image_caption": [ + "High ground", + "Mid ground", + "Low ground", + "Normal", + "Smooth", + "Sticky", + "Fig. 6: We edit boundary conditions in these demos. Changing grounds' position and using smooth or sticky ground." + ], + "image_footnote": [], + "bbox": [ + 217, + 227, + 493, + 297 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/9cde7ca8efb75f7977e40c20db31ec234cd393d79e98fc96a6b7a4ba94e707ad.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 508, + 143, + 787, + 215 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/1de20ed7fdca102612c668c48950a28f36722e6c834a0cd2963d87fc952e0ca3.jpg", + "image_caption": [ + "High ground", + "Mid ground", + "Low ground", + "Normal", + "Smooth", + "Sticky" + ], + "image_footnote": [], + "bbox": [ + 508, + 227, + 785, + 297 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/0a88e24a00358f1c56ae036d6b43b306c630fb103b27f88cc53fec03e1f6f5b4.jpg", + "image_caption": [ + "Original", + "Random init velocity Smaller & Softer" + ], + "image_footnote": [], + "bbox": [ + 217, + 371, + 493, + 444 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/06428440e0a26eb043d0b775992f8e03c68576ec37876ffec49a59072a3474be.jpg", + "image_caption": [ + "Original", + "Zero gravity", + "Random init velocity", + "Fig. 7: We edit physical conditions, initial velocities and gravities in these demos." + ], + "image_footnote": [], + "bbox": [ + 217, + 454, + 493, + 527 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/5a55546295d3f00f0a1fbcac37791e9955fd20b9112c0c70f4c6f4b29f6b2ff4.jpg", + "image_caption": [ + "Bigger & Softer" + ], + "image_footnote": [], + "bbox": [ + 508, + 371, + 785, + 444 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/a93dea573f604609d7f3fb16ccc63210efc1a2f50bb65ac310d9ebca036de7be.jpg", + "image_caption": [ + "Original", + "Windy", + "Original", + "Hypergravity", + "Bigger & Softer" + ], + "image_footnote": [], + "bbox": [ + 508, + 454, + 785, + 527 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 602, + 357, + 618 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In this paper, we introduce Spring-Gaus, a novel framework designed to acquire simulatable digital assets of elastic objects from video observations. By integrating a 3D Spring-Mass model, Spring-Gaus allows the reconstruction of object appearance and dynamics. A key feature of our approach is the distinct separation between the learning processes for appearance and physics, thereby circumventing potential issues with optimization interference. We evaluate Spring-Gaus on both synthetic and real-world datasets, demonstrating its capability to reconstruct geometry, appearance, and physical dynamic properties. Moreover, our method demonstrates improved capabilities by predicting short-term future dynamics under different environmental conditions. This showcases its strength in identifying physical properties from observational data and predicting the dynamics of reconstructed digital assets.", + "bbox": [ + 212, + 657, + 787, + 840 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "L. Zhong et al.", + "bbox": [ + 271, + 114, + 372, + 128 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgments. The work is in part supported by the Amazon AICE Award, NSF RI #2211258, #2338203, ONR YIP N00014-24-1-2117, and ONR MURI N00014-22-1-2740.", + "bbox": [ + 215, + 146, + 787, + 191 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 212, + 321, + 228 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Barron, J.T., Mildenhall, B., Tancik, M., Hedman, P., Martin-Brualla, R., Srinivasan, P.P.: Mip-NeRF: A multiscale representation for anti-aliasing neural radiance fields. In: CVPR (2021)", + "2. Chen, A., Xu, Z., Zhao, F., Zhang, X., Xiang, F., Yu, J., Su, H.: Mvsnerf: Fast generalizable radiance field reconstruction from multi-view stereo. In: ICCV (2021)", + "3. Chen, H., Wu, R., Grinspun, E., Zheng, C., Chen, P.Y.: Implicit neural spatial representations for time-dependent pdes. In: ICML (2023)", + "4. Chu, M., Liu, L., Zheng, Q., Franz, E., Seidel, H.P., Theobalt, C., Zayer, R.: Physics informed neural fields for smoke reconstruction with sparse data. ACM TOG 41(4) (jul 2022). https://doi.org/10.1145/3528223.3530169", + "5. Community, B.O.: Blender - a 3D modelling and rendering package. Blender Foundation, Stichting Blender Foundation, Amsterdam (2018), http://www.blender.org", + "6. Deng, Y., Yu, H.X., Wu, J., Zhu, B.: Learning vortex dynamics for fluid inference and prediction. In: ICLR (2023)", + "7. Deng, Y., Yu, H.X., Zhang, D., Wu, J., Zhu, B.: Fluid simulation on neural flow maps. ACM TOG 42(6) (2023)", + "8. Driess, D., Huang, Z., Li, Y., Tedrake, R., Toussaint, M.: Learning multi-object dynamics with compositional neural radiance fields. In: Conference on Robot Learning. pp. 1755-1768. PMLR (2023)", + "9. Du, T., Wu, K., Ma, P., Wah, S., Spielberg, A., Rus, D., Matusik, W.: Diffpd: Differentiable projective dynamics. ACM TOG 41(2) (nov 2021). https://doi.org/10.1145/3490168", + "10. Guan, S., Deng, H., Wang, Y., Yang, X.: Neurofluid: Fluid dynamics grounding with particle-driven neural radiance fields. In: ICML (2022)", + "1. Hu, Y., Fang, Y., Ge, Z., Qu, Z., Zhu, Y., Pradhana, A., Jiang, C.: A moving least squares material point method with displacement discontinuity and two-way rigid body coupling. ACM TOG 37(4) (jul 2018). https://doi.org/10.1145/3197517.3201293", + "2. Huang, Y.H., Sun, Y.T., Yang, Z., Lyu, X., Cao, Y.P., Qi, X.: Sc-gs: Sparse-controlled gaussian splatting for editable dynamic scenes. CVPR (2024)", + "3. Johnson, J., Alahi, A., Fei-Fei, L.: Perceptual losses for real-time style transfer and super-resolution. In: ECCV (2016)", + "4. Kerbl, B., Kopanas, G., Leimkuhler, T., Drettakis, G.: 3d gaussian splatting for real-time radiance field rendering. ACM TOG 42(4) (July 2023)", + "5. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., Dollár, P., Girshick, R.: Segment Anything. arXiv preprint arXiv:2304.02643 (2023)", + "6. Kratimenos, A., Lei, J., Daniilidis, K.: Dynmf: Neural motion factorization for real-time dynamic view synthesis with 3d gaussian splattering. arXiv preprint arXiv:2312.00112 (2023)", + "7. Le Cleac'h, S., Yu, H.X., Guo, M., Howell, T., Gao, R., Wu, J., Manchester, Z., Schwager, M.: Differentiable physics simulation of dynamics-augmented neural objects. IEEE Robotics and Automation Letters 8(5), 2780-2787 (2023)" + ], + "bbox": [ + 225, + 239, + 784, + 840 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "Spring-Gaus", + "bbox": [ + 645, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "18. Li, X., Qiao, Y.L., Chen, P.Y., Jatavallabhula, K.M., Lin, M., Jiang, C., Gan, C.: PAC-NeRF: Physics augmented continuum neural radiance fields for geometry-agnostic system identification. In: ICLR (2023)", + "19. Li, Y., Chen, P.Y., Du, T., Matusik, W.: Learning preconditioners for conjugate gradient pde solvers. In: ICML (2023)", + "20. Li, Y., Li, S., Sitzmann, V., Agrawal, P., Torralba, A.: 3d neural scene representations for visuomotor control. In: Conference on Robot Learning. pp. 112-123. PMLR (2022)", + "21. Li, Y., Wu, J., Tedrake, R., Tenenbaum, J.B., Torralba, A.: Learning particle dynamics for manipulating rigid bodies, deformable objects, and fluids. arXiv preprint arXiv:1810.01566 (2018)", + "22. Li, Z., Niklaus, S., Snavely, N., Wang, O.: Neural scene flow fields for space-time view synthesis of dynamic scenes. In: CVPR (2021)", + "23. Li, Z., Wang, Q., Cole, F., Tucker, R., Snavely, N.: Dynibar: Neural dynamic image-based rendering. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4273-4284 (2023)", + "24. Lin, Y., Dai, Z., Zhu, S., Yao, Y.: Gaussian-flow: 4d reconstruction with dynamic 3d gaussian particle. arXiv preprint arXiv:2312.03431 (2023)", + "25. Liu, Y., Peng, S., Liu, L., Wang, Q., Wang, P., Christian, T., Zhou, X., Wang, W.: Neural rays for occlusion-aware image-based rendering. In: CVPR (2022)", + "26. Luiten, J., Kopanas, G., Leibe, B., Ramanan, D.: Dynamic 3d gaussians: Tracking by persistent dynamic view synthesis. In: 3DV (2024)", + "27. Ma, P., Chen, P.Y., Deng, B., Tenenbaum, J.B., Du, T., Gan, C., Matusik, W.: Learning neural constitutive laws from motion observations for generalizable pde dynamics. In: ICML (2023)", + "28. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: NeRF: Representing scenes as neural radiance fields for view synthesis. In: ECCV (2020)", + "29. Park, J.J., Florence, P., Straub, J., Newcombe, R., Lovegrove, S.: Deepsdf: Learning continuous signed distance functions for shape representation. In: CVPR (2019)", + "30. Park, K., Sinha, U., Barron, J.T., Bouaziz, S., Goldman, D.B., Seitz, S.M., Martin-Brualla, R.: Nerfies: Deformable neural radiance fields. In: ICCV (2021)", + "31. Park, K., Sinha, U., Hedman, P., Barron, J.T., Bouaziz, S., Goldman, D.B., Martin-Brualla, R., Seitz, S.M.: Hypernerf: A higher-dimensional representation for topologically varying neural radiance fields. arXiv preprint arXiv:2106.13228 (2021)", + "32. Pumarola, A., Corona, E., Pons-Moll, G., Moreno-Noguer, F.: D-NeRF: Neural Radiance Fields for Dynamic Scenes. In: CVPR (2021)", + "33. Raissi, M., Perdikaris, P., Karniadakis, G.: Physics-informed neural networks: A deep learning framework for solving forward and inverse problems involving nonlinear partial differential equations. Journal of Computational Physics 378, 686-707 (2019). https://doi.org/https://doi.org/10.1016/j.jcp.2018.10.045", + "34. Sanchez-Gonzalez, A., Godwin, J., Pfaff, T., Ying, R., Leskovec, J., Battaglia, P.W.: Learning to simulate complex physics with graph networks. In: ICML (2020)", + "35. Schonberger, J.L., Frahm, J.M.: Structure-from-motion revisited. In: CVPR (2016)", + "36. Schonberger, J.L., Zheng, E., Pollefeys, M., Frahm, J.M.: Pixelwise view selection for unstructured multi-view stereo. In: ECCV (2016)", + "37. Simonyan, K., Zisserman, A.: Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556 (2014)", + "38. Sitzmann, V., Zollhöfer, M., Wetzstein, G.: Scene representation networks: Continuous 3d-structure-aware neural scene representations. In: NeurIPS (2019)" + ], + "bbox": [ + 215, + 146, + 784, + 840 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "L. Zhong et al.", + "bbox": [ + 271, + 114, + 370, + 127 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "39. Stomakhin, A., Schroeder, C., Chai, L., Teran, J., Selle, A.: A material point method for snow simulation. ACM TOG 32(4) (jul 2013). https://doi.org/10.1145/2461912.2461948", + "40. Sun, C., Sun, M., Chen, H.: Direct voxel grid optimization: Super-fast convergence for radiance fields reconstruction. In: CVPR (2022)", + "41. Tang, J., Ren, J., Zhou, H., Liu, Z., Zeng, G.: Dreamgaussian: Generative gaussian splatting for efficient 3d content creation. ICLR (2024)", + "42. Tretschk, E., Tewari, A., Golyanik, V., Zollhöfer, M., Lassner, C., Theobalt, C.: Non-rigid neural radiance fields: Reconstruction and novel view synthesis of a dynamic scene from monocular video. In: ICCV (2021)", + "43. Wang, C., MacDonald, L.E., Jeni, L.A., Lucey, S.: Flow supervision for deformable nerf. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 21128-21137 (2023)", + "44. Wu, G., Yi, T., Fang, J., Xie, L., Zhang, X., Wei, W., Liu, W., Tian, Q., Xinggang, W.: 4d gaussian splatting for real-time dynamic scene rendering. In: CVPR (2024)", + "45. Wu, T., Zhang, J., Fu, X., Wang, Y., Ren, J., Pan, L., Wu, W., Yang, L., Wang, J., Qian, C., Lin, D., Liu, Z.: OmniObject3D: Large-vocabulary 3d object dataset for realistic perception, reconstruction and generation. In: CVPR (2023)", + "46. Xie, T., Zong, Z., Qiu, Y., Li, X., Feng, Y., Yang, Y., Jiang, C.: Physsgaussian: Physics-integrated 3d gaussians for generative dynamics. In: CVPR (2024)", + "47. Xu, Q., Xu, Z., Philip, J., Bi, S., Shu, Z., Sunkavalli, K., Neumann, U.: Point-NeRF: Point-based neural radiance fields. In: CVPR (2022)", + "48. Yang, Z., Yang, H., Pan, Z., Zhang, L.: Real-time photorealistic dynamic scene representation and rendering with 4d gaussian splatting. In: ICLR (2024)", + "49. Yang, Z., Gao, X., Zhou, W., Jiao, S., Zhang, Y., Jin, X.: Deformable 3d gaussians for high-fidelity monocular dynamic scene reconstruction. In: CVPR (2024)", + "50. Yu, A., Ye, V., Tancik, M., Kanazawa, A.: pixelNeRF: Neural radiance fields from one or few images. In: CVPR (2021)", + "51. Yu, H., Julin, J., Milacski, Z.A., Niinuma, K., Jeni, L.A.: Cogs: Controllable gaussian splatting. arXiv preprint arXiv:2312.05664 (2023)", + "52. Yu, H.X., Zheng, Y., Gao, Y., Deng, Y., Zhu, B., Wu, J.: Inferring hybrid neural fluid fields from videos. Advances in Neural Information Processing Systems (2023)", + "53. Zhou, Y., Barnes, C., Lu, J., Yang, J., Li, H.: On the continuity of rotation representations in neural networks. In: CVPR (2019)" + ], + "bbox": [ + 212, + 146, + 787, + 618 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "Spring-Gaus", + "bbox": [ + 645, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + } +] \ No newline at end of file diff --git a/2024/Reconstruction and Simulation of Elastic Objects with Spring-Mass 3D Gaussians/2eb8194f-eadd-43df-b4b2-bbf94906f9cc_model.json b/2024/Reconstruction and Simulation of Elastic Objects with Spring-Mass 3D Gaussians/2eb8194f-eadd-43df-b4b2-bbf94906f9cc_model.json new file mode 100644 index 0000000000000000000000000000000000000000..46c2a89d3270f5ed457783df8cd84497a3497ec1 --- /dev/null +++ b/2024/Reconstruction and Simulation of Elastic Objects with Spring-Mass 3D Gaussians/2eb8194f-eadd-43df-b4b2-bbf94906f9cc_model.json @@ -0,0 +1,2731 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.221, + 0.141, + 0.783, + 0.187 + ], + "angle": 0, + "content": "Reconstruction and Simulation of Elastic Objects with Spring-Mass 3D Gaussians" + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.212, + 0.744, + 0.229 + ], + "angle": 0, + "content": "Licheng Zhong\\(^{1*}\\), Hong-Xing Yu\\(^{1}\\), Jiajun Wu\\(^{1}\\), and Yunzhu Li\\(^{1,2,3}\\)" + }, + { + "type": "text", + "bbox": [ + 0.376, + 0.239, + 0.628, + 0.254 + ], + "angle": 0, + "content": "1 Stanford University, Stanford, USA" + }, + { + "type": "text", + "bbox": [ + 0.369, + 0.254, + 0.634, + 0.268 + ], + "angle": 0, + "content": "2 Columbia University, New York, USA" + }, + { + "type": "text", + "bbox": [ + 0.296, + 0.268, + 0.706, + 0.282 + ], + "angle": 0, + "content": "3 University of Illinois Urbana-Champaign, Champaign, USA" + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.32, + 0.744, + 0.528 + ], + "angle": 0, + "content": "Abstract. Reconstructing and simulating elastic objects from visual observations is crucial for applications in computer vision and robotics. Existing methods, such as 3D Gaussians, model 3D appearance and geometry, but lack the ability to estimate physical properties for objects and simulate them. The core challenge lies in integrating an expressive yet efficient physical dynamics model. We propose Spring-Gaus, a 3D physical object representation for reconstructing and simulating elastic objects from videos of the object from multiple viewpoints. In particular, we develop and integrate a 3D Spring-Mass model into 3D Gaussian kernels, enabling the reconstruction of the visual appearance, shape, and physical dynamics of the object. Our approach enables future prediction and simulation under various initial states and environmental properties. We evaluate Spring-Gaus on both synthetic and real-world datasets, demonstrating accurate reconstruction and simulation of elastic objects. Project page: https://zlicheng.com/spring_gaus." + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.541, + 0.741, + 0.556 + ], + "angle": 0, + "content": "Keywords: System Identification \\(\\cdot\\) Spring-Mass Model \\(\\cdot\\) 3D Gaussians" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.604, + 0.377, + 0.62 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.636, + 0.788, + 0.741 + ], + "angle": 0, + "content": "Reconstructing and simulating elastic objects from visual observations poses a fundamental challenge in computer vision and robotics, with applications spanning virtual reality, augmented reality, and robotic manipulation. Accurately modeling the elasticity of objects is crucial for creating immersive experiences and enabling embodied agents to understand and interact with the elastic objects commonly encountered in our daily lives. However, accurately identifying the dynamics from vision still presents considerable challenges." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.742, + 0.788, + 0.803 + ], + "angle": 0, + "content": "Existing methods for dynamic scene reconstruction, such as 3D Gaussians and their dynamic extensions [14, 26, 44, 49], have made significant progress in capturing the temporal changes of the appearance and geometry of objects. However, these methods do not capture the physical properties of the reconstructed" + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.811, + 0.787, + 0.841 + ], + "angle": 0, + "content": "* The work was done while L. Zhong was a visiting student at Stanford University. L. Zhong is now at Shanghai Jiao Tong University." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.373, + 0.129 + ], + "angle": 0, + "content": "L. Zhong et al." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.145, + 0.787, + 0.341 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.352, + 0.788, + 0.394 + ], + "angle": 0, + "content": "Fig. 1: Spring-Gaus reconstructs the appearance, geometry, and physical dynamic properties of elastic objects from video observations. Spring-Gaus enables future predictions and simulations under different initial states and environmental conditions." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.429, + 0.789, + 0.685 + ], + "angle": 0, + "content": "objects; thus, they typically cannot predict the future dynamics of these objects. While a few recent approaches using MPM [11,39] have attempted to integrate physics-based priors into 3D object representations, such as PAC-NeRF [18]. Their ability to handle real, especially heterogeneous, objects is limited, as they assume a known material model and only assigns a global physical parameter to the entire object which restricts its adaptability to real objects. Assigning learnable physical parameters to each particle in an MPM is theoretically possible. However, in practice, it incurs extreme computational costs since MPM requires tens of thousands of dense points. In addition, PAC-NeRF [18] use a implicit grid representation, due to the computational demands, the resolution of the grid is limited, which can lead to a loss of detail in the appearance modeling when using real or noisy data. Thus, the core challenge in reconstructing and predicting object dynamics lies in developing and integrating an expressive and efficient physical model for the dynamics. The physical dynamics model should be expressive enough to capture the motions of elastic objects, including collisions, deformations, and bouncing. It must also be efficient and conducive to inverse parameter estimation through gradient-based optimization." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.687, + 0.789, + 0.809 + ], + "angle": 0, + "content": "In this work, we propose Spring-Gaus, a 3D object representation that integrates a 3D Spring-Mass model. Our model represents the elastic object dynamics properties through a learnable system of mass points and springs. Our design is expressive in that it assumes a general and widely applicable physical model class. With a learnable topology and physical parameters, Spring-Gaus can model complex deformation and motion for heterogeneous elastic objects. In addition to expressiveness, Spring-Gaus is also highly efficient for the inverse optimization of physical parameters due to its differentiable nature." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.789, + 0.842 + ], + "angle": 0, + "content": "Spring-Gaus enables reconstructing and simulating elastic objects from sparse multi-view videos (Fig. 1). To overcome the intrinsic difficulty in optimization," + } + ], + [ + { + "type": "header", + "bbox": [ + 0.646, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Spring-Gaus" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.207 + ], + "angle": 0, + "content": "we propose a reconstruction pipeline that decouples physical parameter reconstruction from appearance and geometry reconstruction. Spring-Gaus requires only a few multi-view videos for physical parameter identification and is robust to the quality of geometry reconstruction." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.208, + 0.785, + 0.297 + ], + "angle": 0, + "content": "We evaluate the effectiveness of Spring-Gaus on both synthetic and real-world datasets, demonstrating its ability to accurately reconstruct and simulate elastic objects. Our Spring-Gaus allows accurate future prediction and simulations under varying initial states and environmental parameters, showcasing its potential for applications in predictive visual perception and immersive experiences. In summary, the main contributions of this work are threefold:" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.306, + 0.784, + 0.335 + ], + "angle": 0, + "content": "- We propose Spring-Gaus, which incorporates an expressive yet efficient 3D Spring-Mass model for reconstructing and simulating elastic objects." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.337, + 0.784, + 0.379 + ], + "angle": 0, + "content": "- We introduce a pipeline to reconstruct Spring-Gaus from multi-view videos of the object. We decouple the appearance and geometry reconstruction from the physical dynamics reconstruction for more effective optimization." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.381, + 0.784, + 0.439 + ], + "angle": 0, + "content": "- We demonstrate the effectiveness of Spring-Gaus on both synthetic and real-world datasets, showcasing accurate reconstruction and simulation of elastic objects. This includes capabilities for future prediction and simulation under varying initial configurations." + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.306, + 0.784, + 0.439 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.463, + 0.387, + 0.479 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.493, + 0.785, + 0.779 + ], + "angle": 0, + "content": "3D Object Representations: Traditionally, 3D objects are often represented by point clouds, meshes, and voxels. Recently, neural 3D object representation has become popular due to the efficiency and flexibility. For example, Scene Representation Networks (SRNs) [38] and DeepSDF [29] represent significant advancements in 3D scene and object representation, treating 3D objects as continuous functions that map world coordinates to a feature representation of local object properties. Over the past few years, NeRF [28] and its successors [1, 2, 25, 47, 50] have demonstrated the efficacy of neural networks in capturing continuous 3D scenes and objects through implicit representation. DirectVoxGO [40] accelerates NeRF's approach by substituting the MLP with a voxel grid. Furthermore, 3D Gaussian Splatting [14] has emerged as a method for real-time differentiable rendering, representing scenes and objects with 3D Gaussians. Extending this approach, DreamGaussian [41] applies 3D Gaussians for 3D object generation. Unlike these approaches, our method focuses on reconstructing simulatable 3D objects. Recently, PhysGaussian [46] integrated physical simulation into 3D Gaussians using a customized Material Point Method, allowing forward simulation of reconstructed objects. In contrast to PhysGaussian, our work focuses on system identification from raw videos and supports both the forward simulation and the inverse reconstruction of physical objects." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.785, + 0.84 + ], + "angle": 0, + "content": "Dynamic Scene Reconstruction: The modeling of dynamic scenes has seen significant progress with the adoption of NeRF [8, 10, 18, 20, 22, 23, 30-32, 42, 43] and 3D Gaussian [12, 16, 24, 26, 44, 48, 49, 51] representations. D-NeRF [32] introduces an extension to NeRF, capable of modeling dynamic scenes from monocu" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.372, + 0.129 + ], + "angle": 0, + "content": "L. Zhong et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.373 + ], + "angle": 0, + "content": "lar views by optimizing an underlying deformable volumetric function. Furthermore, Dynamic 3D Gaussians [26] focus on optimizing the motion of Gaussian kernels for each frame, presenting an efficient method to capture scene dynamics. Deformable 3D Gaussians [49] propose a novel approach by learning Gaussian distributions in a canonical space, complemented with a deformation field for modeling monocular dynamic scenes. Meanwhile, 4D Gaussian Splatting [44] introduces a hybrid model combining 3D Gaussians with 4D neural voxels. PAC-NeRF [18] delves into the integration of Lagrangian particle simulation with Eulerian scene representation, exploring a new aspect of scene dynamics. Among them, PAC-NeRF [18] is the most relevant to our work. PAC-NeRF learns physical parameters used in the Material Point Method (MPM) for better dynamic reconstruction and prediction. However, PAC-NeRF assumes known material models that restrict its adaptability and applicability to complex real objects. In contrast, our work integrates a 3D Spring-Mass model that is both expressive and efficient, allowing reconstructing and simulating real elastic objects." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.374, + 0.788, + 0.631 + ], + "angle": 0, + "content": "Physics-Informed Learning: Physics-Informed Learning has emerged as a prominent research direction since the introduction of Physics-Informed Neural Networks (PINNs) [33]. Li et al. [21] and Sanchez-Gonzalez et al. [34] introduced Graph Network-based simulators within a machine learning framework. INSR-PDE [3] tackles time-dependent partial differential equations (PDEs) using implicit neural spatial representations. NCLaw [27] focuses on learning neural constitutive laws for PDE dynamics. DiffPD [9] presents a differentiable soft-body simulator. Li et al. [19] have made strides in learning preconditioners for conjugate gradient PDE solvers. Chu et al. [4] and Yu et al. [52] modeled smoke in neural density and velocity fields. Neural Flow Maps [7] integrate fluid simulation with neural implicit representations. Deng et al. [6] introduced a novel differentiable vortex particle method for fluid dynamics inference. DANO [17] is the most relevant work to ours. In particular, DANO develops a differentiable simulation for rigid objects represented by NeRFs, allowing reconstruction and simulation from a few videos. In contrast, our work focuses on elastic objects, which involve a fundamentally new set of challenges and require a redesign of the physical models." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.654, + 0.348, + 0.673 + ], + "angle": 0, + "content": "3 Approach" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.687, + 0.787, + 0.78 + ], + "angle": 0, + "content": "We propose Spring-Gaus that integrates a 3D Spring-Mass model with 3D Gaussians for elastic object reconstruction and simulation. Specifically, we are given a set of \\(O\\) calibrated videos from multiple viewpoints of an elastic object in motion, denoted as \\(\\{I^{o,f}\\}_{o = 1,f = 1}^{O,F}\\), where \\(F\\) is the number of video frames and \\(I^{o,f}\\) denotes a 2D image at video \\(o\\) and frame \\(f\\). Our goal is to reconstruct the appearance, geometry, and physical dynamics parameters of Spring-Gaus." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.781, + 0.787, + 0.842 + ], + "angle": 0, + "content": "In the following, we first introduce the representation for the static properties, i.e., appearance and geometry, and then the dynamics properties, i.e., the 3D Spring-Mass model. Then we describe our reconstruction pipeline. We show an overview of our reconstruction pipeline in Fig. 2." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.646, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Spring-Gaus" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.147, + 0.616, + 0.163 + ], + "angle": 0, + "content": "3.1 Appearance and Geometry Representation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.17, + 0.784, + 0.217 + ], + "angle": 0, + "content": "Following 3D Gaussians Splatting [14], we use a set of 3D Gaussian kernels to represent the appearance and shape of an object. These kernels are described by a full 3D covariance matrix \\(\\Sigma\\), defined in world space and centered at point \\(\\pmb{\\mu}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.421, + 0.226, + 0.786, + 0.246 + ], + "angle": 0, + "content": "\\[\nG (\\boldsymbol {\\tau}) = e ^ {- \\frac {1}{2} (\\boldsymbol {\\tau}) ^ {T} \\Sigma^ {- 1} (\\boldsymbol {\\tau})}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.254, + 0.784, + 0.299 + ], + "angle": 0, + "content": "where \\(\\pmb{\\mu}\\) is the mean of the Gaussian distribution, and \\(\\pmb{\\tau}\\) is the independent variable of the Gaussian distribution. In our formulation, we assume all Gaussian kernels are isotropic, the covariance matrix \\(\\Sigma\\) can be controlled by a scalar \\(s\\in \\mathbb{R}\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.3, + 0.784, + 0.357 + ], + "angle": 0, + "content": "The splatting process, designed to render Gaussian kernels into a 2D image, involves two main steps: (1) projecting the Gaussian kernels into camera space, and (2) rendering the projected kernels into image space. The projection process is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.435, + 0.359, + 0.785, + 0.377 + ], + "angle": 0, + "content": "\\[\n\\Sigma^ {\\prime} = J W \\Sigma W ^ {T} J ^ {T}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.381, + 0.784, + 0.427 + ], + "angle": 0, + "content": "where \\(\\Sigma'\\) represents the covariance matrix in camera coordinates, \\(W\\) is the transformation matrix, and \\(J\\) denotes the Jacobian of the affine approximation of the projective transformation. The color \\(C\\) of each pixel is rendered as:" + }, + { + "type": "equation", + "bbox": [ + 0.446, + 0.437, + 0.785, + 0.468 + ], + "angle": 0, + "content": "\\[\nC = \\sum_ {i \\in N} T _ {i} \\alpha_ {i} \\boldsymbol {c} _ {i}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.476, + 0.784, + 0.553 + ], + "angle": 0, + "content": "where \\(N\\) is the total number of kernels and \\(T_{i}\\) represents the transmittance, defined as \\(\\Pi_{j = 1}^{i - 1}\\alpha_{j}\\). The term \\(\\alpha_{i}\\) denotes the alpha value for each Gaussian, which is calculated using the expression \\(1 - e^{-\\sigma_i\\delta_i}\\), where \\(\\sigma_{i}\\) is the opacity factor. Additionally, \\(\\pmb{c}_i\\) refers to the color of the Gaussian along the ray within the interval \\(\\delta_{i}\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.554, + 0.784, + 0.585 + ], + "angle": 0, + "content": "For each kernel, the learnable parameters include a point center \\(\\pmb{\\mu}\\), a scaling scalar \\(s\\), a color vector \\(\\pmb{c}\\), and an opacity value \\(\\sigma\\)." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.606, + 0.704, + 0.621 + ], + "angle": 0, + "content": "3.2 Physics-Based Dynamics with 3D Spring-Mass Model" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.629, + 0.784, + 0.734 + ], + "angle": 0, + "content": "We aim to develop and integrate an expressive yet efficient physics-based dynamics model. With these two design goals in mind, we introduce the 3D Spring-Mass model which represents the elastic object dynamics by a learnable spring-mass system. Our Spring-Gaus does not assume any material type, and it is expressive to model real heterogeneous elastic objects. Besides being expressive, our dynamics model should also be efficient and amenable to gradient-based optimization for parameter estimation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.784, + 0.811 + ], + "angle": 0, + "content": "Recall that the appearance and geometry representations include a set of Gaussian kernels \\(\\{G_i\\}_{i=1}^N\\), parameterized by point centers \\(\\mathbf{X} = \\{\\pmb{\\mu}_i\\}_{i=1}^N\\), scaling scalars \\(\\{s_i\\}_{i=1}^N\\), color vectors \\(\\{\\pmb{c}_i\\}_{i=1}^N\\), and opacity values \\(\\{\\sigma_i\\}_{i=1}^N\\). To manage complexity for efficient simulation, we introduce volume sampling to generate a set of anchor points \\(\\mathbf{A} = \\{\\pmb{x}_i\\}_{i=1}^{N_A}\\) (each \\(\\pmb{x}\\) represents a spatial point), defined by:" + }, + { + "type": "equation", + "bbox": [ + 0.424, + 0.824, + 0.785, + 0.842 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {A} = \\left\\{\\boldsymbol {x} _ {i} \\right\\} _ {i = 1} ^ {N _ {A}} = \\mathcal {V} (\\boldsymbol {X}), \\tag {4}\n\\]" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.373, + 0.129 + ], + "angle": 0, + "content": "L. Zhong et al." + }, + { + "type": "image", + "bbox": [ + 0.224, + 0.147, + 0.782, + 0.409 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.437, + 0.788, + 0.549 + ], + "angle": 0, + "content": "Fig.2: Overview of Spring-Gaus reconstruction pipeline: (a) Static Scene Reconstruction: We start by reconstructing static 3D Gaussians from the first frames of the multiview videos. (b) Refining 3D Gaussians: We extract a set of anchor points to allow efficient simulation, which leads to appearance drift. We refine the 3D Gaussians to better model the appearance during simulation. (c) Dynamic Reconstruction: Our 3D Spring-Mass model simulates anchor points and updates the positions of Gaussian kernels. Upon completion of optimization, we obtain a simulatable 3D object that accurately models its dynamics." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.578, + 0.785, + 0.593 + ], + "angle": 0, + "content": "where \\(N_{A}\\) denotes the number of anchors, and \\(\\nu\\) denotes the sampling function." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.594, + 0.787, + 0.641 + ], + "angle": 0, + "content": "Our 3D Spring-Mass physical model can simulate the motion of anchor points \\(\\mathbf{A}\\), assuming each anchor has a mass \\(m_{i}\\) and an initial velocity \\(\\mathbf{v}_{i}\\). Each anchor \\(\\mathbf{x}_{i}\\) connects to its \\(n_{k}\\) nearest neighbors \\(\\mathbf{N}_{i} = \\{\\mathbf{x}_{i,j}\\}_{j=1}^{n_{k}}\\) through springs \\(\\mathbf{L}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.375, + 0.651, + 0.785, + 0.672 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {L} = \\left\\{l _ {i, j} \\right\\} _ {i = 1, j = 1} ^ {N _ {A}, n _ {k}} = \\operatorname {k n n} (\\boldsymbol {A}, \\boldsymbol {A}, n _ {k}), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.681, + 0.787, + 0.726 + ], + "angle": 0, + "content": "where \\( l_{i,j} \\) denotes the spring's length between \\( \\pmb{x}_i \\) and \\( \\pmb{x}_{i,j} \\), and knn denotes the k-nearest neighbors function. Each spring is characterized by a stiffness \\( k_{i,j} \\) and a damping factor \\( \\zeta_{i,j} \\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.727, + 0.787, + 0.757 + ], + "angle": 0, + "content": "To update the positions of the kernels, we first measure the distance between each kernel center and its \\( n_b \\) nearest anchors at the dynamic simulation's onset:" + }, + { + "type": "equation", + "bbox": [ + 0.391, + 0.769, + 0.785, + 0.79 + ], + "angle": 0, + "content": "\\[\n\\left\\{d _ {i, j} \\right\\} _ {i = 1, j = 1} ^ {N, n _ {b}} = \\operatorname {k n n} (\\boldsymbol {X}, \\boldsymbol {A}, n _ {b}). \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.787, + 0.823 + ], + "angle": 0, + "content": "For each timestep \\(t\\), the forces \\(\\pmb{F}_i^t\\) acting on each anchor point \\(\\pmb{x}_i^t\\) are calculated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.416, + 0.824, + 0.785, + 0.843 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {F} _ {i} ^ {t} = \\boldsymbol {F} _ {\\boldsymbol {k} i} ^ {t} + \\boldsymbol {F} _ {\\boldsymbol {\\zeta} i} ^ {t} + m _ {i} \\boldsymbol {g}, \\tag {7}\n\\]" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.646, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Spring-Gaus" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.146, + 0.784, + 0.177 + ], + "angle": 0, + "content": "where \\( \\pmb{F}_{\\pmb{k}_i}^t, \\pmb{F}_{\\pmb{\\zeta}_i}^t \\), and \\( \\pmb{g} \\) represent the spring force, the damping force, and gravitational acceleration, respectively." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.177, + 0.785, + 0.207 + ], + "angle": 0, + "content": "Then, for each spring \\( \\pmb{L}_{i,j} \\), the spring force \\( \\pmb{F}_{\\pmb{k}_{i,j}}^{t} \\) and damping force \\( \\pmb{F}_{\\pmb{\\zeta}_{i,j}}^{t} \\) are determined by:" + }, + { + "type": "equation", + "bbox": [ + 0.236, + 0.219, + 0.784, + 0.255 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {F} _ {\\boldsymbol {k} _ {i, j}} ^ {t} = - \\eta_ {j} \\cdot k _ {i, j} \\left(\\left\\| \\boldsymbol {x} _ {i} ^ {t} - \\boldsymbol {x} _ {i, j} ^ {t} \\right\\| - l _ {i, j}\\right) \\frac {\\boldsymbol {x} _ {i} ^ {t} - \\boldsymbol {x} _ {i , j} ^ {t}}{\\left\\| \\boldsymbol {x} _ {i} ^ {t} - \\boldsymbol {x} _ {i , j} ^ {t} \\right\\|} \\cdot \\left| \\left\\| \\boldsymbol {x} _ {i} ^ {t} - \\boldsymbol {x} _ {i, j} ^ {t} \\right\\| - l _ {i, j} \\right| ^ {p _ {k}}, \\tag {8}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.31, + 0.258, + 0.784, + 0.297 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {F} _ {\\zeta_ {i, j}} ^ {t} = \\left(- \\zeta_ {i, j} \\left(\\boldsymbol {v} _ {i} ^ {t} - \\boldsymbol {v} _ {i, j} ^ {t}\\right) \\frac {\\boldsymbol {x} _ {i} ^ {t} - \\boldsymbol {x} _ {i , j} ^ {t}}{\\left\\| \\boldsymbol {x} _ {i} ^ {t} - \\boldsymbol {x} _ {i , j} ^ {t} \\right\\|}\\right) \\cdot \\frac {\\boldsymbol {x} _ {i} ^ {t} - \\boldsymbol {x} _ {i , j} ^ {t}}{\\left\\| \\boldsymbol {x} _ {i} ^ {t} - \\boldsymbol {x} _ {i , j} ^ {t} \\right\\|}, \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.307, + 0.784, + 0.382 + ], + "angle": 0, + "content": "where \\(\\eta\\) is a soft vector which will be discussed later and \\(p_k\\) is a hyperparameter that determines the nonlinearity of the spring force. When \\(p_k\\) is set to 0, Eq. (8) becomes Hooke's law, and for positive values of \\(p_k\\), the spring force becomes a nonlinear function of the spring's length. The cumulative forces acting on each anchor point \\(\\boldsymbol{x}_i^t\\) are expressed as:" + }, + { + "type": "equation", + "bbox": [ + 0.384, + 0.395, + 0.784, + 0.434 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {F} _ {i} ^ {t} = \\sum_ {j = 1} ^ {n _ {k}} \\boldsymbol {F} _ {\\boldsymbol {k} _ {i, j}} ^ {t} + \\sum_ {j = 1} ^ {n _ {k}} \\boldsymbol {F} _ {\\zeta_ {i, j}} ^ {t} + m _ {i} \\boldsymbol {g}. \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.444, + 0.784, + 0.473 + ], + "angle": 0, + "content": "Anchor points \\(\\mathbf{A}\\)'s positions and velocities are updated using semi-implicit Euler integration:" + }, + { + "type": "equation", + "bbox": [ + 0.433, + 0.485, + 0.784, + 0.516 + ], + "angle": 0, + "content": "\\[\n\\hat {\\boldsymbol {v}} _ {i} ^ {t + 1} = \\boldsymbol {v} _ {i} ^ {t} + \\frac {\\boldsymbol {F} _ {i} ^ {t}}{m _ {i}} \\Delta t, \\tag {11}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.428, + 0.518, + 0.784, + 0.536 + ], + "angle": 0, + "content": "\\[\n\\hat {\\boldsymbol {x}} _ {i} ^ {t + 1} = \\boldsymbol {x} _ {i} ^ {t} + \\boldsymbol {v} _ {i} ^ {t + 1} \\Delta t, \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.547, + 0.784, + 0.576 + ], + "angle": 0, + "content": "and a boundary condition \\(\\mathcal{B}\\) is applied to the anchor points \\(\\pmb{A}\\) to model the interactions with the environment:" + }, + { + "type": "equation", + "bbox": [ + 0.404, + 0.588, + 0.784, + 0.605 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {x} _ {i} ^ {t + 1}, \\boldsymbol {v} _ {i} ^ {t + 1} = \\mathcal {B} \\left(\\hat {\\boldsymbol {x}} _ {i} ^ {t + 1}, \\hat {\\boldsymbol {v}} _ {i} ^ {t + 1}\\right). \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.616, + 0.784, + 0.647 + ], + "angle": 0, + "content": "The position of each Gaussian kernel \\(\\pmb{\\mu}_i\\) is updated through Inverse Distance Weighting (IDW) interpolation to reflect the dynamic changes accurately:" + }, + { + "type": "equation", + "bbox": [ + 0.386, + 0.658, + 0.784, + 0.695 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {\\mu} _ {i} ^ {t + 1} = \\frac {\\sum_ {j = 1} ^ {n _ {b}} \\boldsymbol {x} _ {i , j} ^ {t + 1} \\cdot \\left(1 / \\left(d _ {i , j}\\right) ^ {p _ {b}}\\right)}{\\sum_ {j = 1} ^ {n _ {b}} \\left(1 / \\left(d _ {i , j}\\right) ^ {p _ {b}}\\right)}, \\tag {14}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.705, + 0.784, + 0.764 + ], + "angle": 0, + "content": "where \\( p_b \\) is a positive real number that determines the diminishing influence of anchor points with distance. To render the image \\( \\hat{I} \\) at a specific camera and frame, we use the updated positions of the Gaussian kernels following the rendering equation Eq. (3)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.766, + 0.785, + 0.84 + ], + "angle": 0, + "content": "Soft Vector for Springs Connection: In our formulation above, the \\( n_k \\) is a hyperparameter which is the number of connected springs for each anchor. However, the choice of \\( n_k \\) will directly affect the simulation results markedly. The bigger value of \\( n_k \\), the object will behave more rigidly, and a smaller value of \\( n_k \\) leads to a noticeably softer behavior of the point cloud. To address the" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.373, + 0.129 + ], + "angle": 0, + "content": "L. Zhong et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.224 + ], + "angle": 0, + "content": "significant impact that the value of \\( n_k \\) has on the simulation, we introduce a mitigation strategy by applying a soft vector \\( \\eta = [\\eta_0, \\eta_1, \\dots, \\eta_{n_k}] \\). This vector controlled by a learnable parameter \\( \\kappa \\) (shared by all anchors) is used to modulate the number of connected springs, thereby adjusting the system's response to a different object. Given an empirical value \\( n_c \\), the soft vector \\( \\eta \\) is calculated as:" + }, + { + "type": "equation", + "bbox": [ + 0.292, + 0.233, + 0.786, + 0.274 + ], + "angle": 0, + "content": "\\[\n\\eta_ {j} = \\left\\{ \\begin{array}{l l} 1 & j \\leq n _ {c}, \\\\ \\operatorname {c l a m p} (2 - \\exp (\\operatorname {s o f t p l u s} (\\kappa)) ^ {j - n _ {c}}, 0, 1) & n _ {c} < j \\leq n _ {k}. \\end{array} \\right. \\tag {15}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.291, + 0.373, + 0.307 + ], + "angle": 0, + "content": "3.3 Optimization" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.315, + 0.789, + 0.452 + ], + "angle": 0, + "content": "To allow efficient optimization, we simplify our model by reducing the number of learnable parameters without changing the essential expressiveness. In our simplified approach, we standardize the mass of every anchor to a constant value \\( m_0 \\), and control all damping factors using a singular parameter \\( \\zeta_0 \\). These two parameters are fixed, eliminating their variability from the optimization process. Furthermore, we introduce a unified parameter \\( k_{i} \\) for each anchor \\( \\pmb{x}_i \\) to control the spring stiffness of the springs connected to it, simplifying the model without compromising its functional integrity. The spring stiffness \\( k_{i,j} \\) and damping factor \\( \\zeta_{i,j} \\) are thus given as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.456, + 0.463, + 0.785, + 0.478 + ], + "angle": 0, + "content": "\\[\nk _ {i, j} = k _ {i} / l _ {i, j}, \\tag {16}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.456, + 0.483, + 0.785, + 0.498 + ], + "angle": 0, + "content": "\\[\n\\zeta_ {i, j} = \\zeta_ {0} / l _ {i, j}. \\tag {17}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.507, + 0.789, + 0.582 + ], + "angle": 0, + "content": "Note that, in Eq. (17) stiffness is defined at anchor points rather than on springs themselves to simplify the optimization—we only need to optimize \\( N_{A} \\) stiffness coefficients instead of all the springs, which is on the order of \\( n_k \\cdot N_A \\). By this simplification, our model maintains computational efficiency and ease of optimization, while still capturing the essential dynamics of the object." + }, + { + "type": "text", + "bbox": [ + 0.239, + 0.583, + 0.714, + 0.598 + ], + "angle": 0, + "content": "Summarized, the learnable parameters in our model now include:" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.606, + 0.785, + 0.634 + ], + "angle": 0, + "content": "- \\( \\pmb{v}_0 \\): the initial velocity vector, providing a baseline movement pattern for the simulation;" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.635, + 0.785, + 0.666 + ], + "angle": 0, + "content": "- \\(\\{k_i\\}_{i=1}^{N_A}\\): the individual stiffness parameters for each anchor, allowing for localized adjustments to spring stiffness;" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.667, + 0.785, + 0.695 + ], + "angle": 0, + "content": "- \\(\\kappa\\): a parameter governing the modulation of the soft vector, facilitating finetuned control over the spring dynamics;" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.696, + 0.785, + 0.741 + ], + "angle": 0, + "content": "- \\(\\Theta(\\mathcal{B}(\\cdot))\\): parameters defining the boundary conditions (\\(\\Theta\\) is learnable parameters), such as the friction coefficient, which influence the simulation's physical realism." + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.606, + 0.785, + 0.741 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.841 + ], + "angle": 0, + "content": "There exist \\( n_t \\) timesteps between each of the two keyframes. At each timestep, \\( \\boldsymbol{x}^t \\) is computed from \\( \\boldsymbol{x}^{t-1} \\) using the update rule Eqs. (11) and (12). We only optimize physical parameters at each keyframe (when \\( t = 0, n_t, 2n_t, 3n_t, \\ldots \\)) based on the visual observations until that time. We increase the value of \\( n_t \\) as the parameters converge. This approach balances identification accuracy and computational demand." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.646, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Spring-Gaus" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.208 + ], + "angle": 0, + "content": "Same with 3D Gaussian Splatting [14], we define our loss function as a weighted combination of the \\(\\mathcal{L}_1\\) norm and the Structural Similarity Index Measure (D-SSIM) \\(\\mathcal{L}_{\\mathrm{d - ssim}}\\), applied between the input images \\(I\\) and the rendered images \\(\\hat{I}\\). This is formally expressed as:" + }, + { + "type": "equation", + "bbox": [ + 0.373, + 0.224, + 0.786, + 0.239 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\left(1 - \\lambda_ {\\mathrm {d - s s i m}}\\right) \\mathcal {L} _ {1} + \\lambda_ {\\mathrm {d - s s i m}} \\mathcal {L} _ {\\mathrm {d - s s i m}}, \\tag {18}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.245, + 0.786, + 0.291 + ], + "angle": 0, + "content": "where \\(\\lambda_{\\mathrm{d - ssim}}\\) is a weighting coefficient that balances the contribution of the \\(\\mathcal{L}_1\\) norm and the D-SSIM term to the overall loss. We decouple our optimization into a few stages, as illustrated in Fig. 2." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.291, + 0.787, + 0.336 + ], + "angle": 0, + "content": "Static Reconstruction: Our optimization starts by taking the first frames of the multi-view videos and using them to reconstruct the appearance and geometry of the object." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.337, + 0.788, + 0.472 + ], + "angle": 0, + "content": "3D Gaussians Refinement: During dynamic simulation of the anchor points, the 3D Gaussian centers \\( \\mathbf{X}_0 = \\{\\pmb{\\mu}_i^0\\}_{i=1}^N \\) are computed by IDW interpolation outlined in Eq. (14). This leads to a slight appearance drift from the static reconstruction. Therefore, we refine the parameters of the Gaussian kernels—scaling scalar \\( s \\), color vector \\( c \\), and opacity value \\( \\sigma \\), excluding points center \\( \\pmb{\\mu} \\)—also at the first frame, before the dynamic simulation. Since the Gaussian kernels' position from IDW interpolation is slightly different from the static reconstruction results, this refinement process enables the Gaussian kernels with the interpolated spatial position to render the correct object appearance." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.473, + 0.788, + 0.532 + ], + "angle": 0, + "content": "Dynamic Reconstruction: Since our efficient dynamics simulation is fully differentiable with respect to the learnable parameters, we can optimize the physical parameters through differentiable simulation and differentiable rendering with our loss function \\(\\mathcal{L}\\)." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.557, + 0.367, + 0.573 + ], + "angle": 0, + "content": "4 Experiment" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.588, + 0.336, + 0.602 + ], + "angle": 0, + "content": "4.1 Datasets" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.613, + 0.788, + 0.765 + ], + "angle": 0, + "content": "Synthetic Data. We evaluate Spring-Gaus using a synthetic dataset. We first collected fourteen 3D models, including point clouds and meshes, to use as initial point clouds for our simulations. Some of them are sourced from PAC-NeRF [18] and OmniObject3D [45]. Following PAC-NeRF, we employ the Material Point Method [11,39] to simulate the dynamics of elastic objects to generate synthetic data. Our dataset features elastic objects with various stiffness levels and diverse geometric forms. Multi-view RGB videos are rendered using Blender [5], with each sequence comprising 10 viewpoints and 30 frames at a resolution of \\(512 \\times 512\\). The initial 20 frames are utilized for dynamic reconstruction, while the subsequent 10 frames are dedicated to evaluating future prediction performance." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.766, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Real-World Data. In addition to synthetic datasets, we further assess Spring-Gaus using captured real-world examples. Our collection process distinguishes between static scenes and dynamic multi-view videos. For static scenes, we position each object on a table and capture 50-70 images from various viewpoints within the upper hemisphere surrounding the object. To obtain camera" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.373, + 0.129 + ], + "angle": 0, + "content": "L. Zhong et al." + }, + { + "type": "image", + "bbox": [ + 0.286, + 0.145, + 0.726, + 0.248 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.241, + 0.269, + 0.761, + 0.285 + ], + "angle": 0, + "content": "Fig. 3: Registration from static scene to dynamic scene for real-world sample." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.311, + 0.788, + 0.388 + ], + "angle": 0, + "content": "poses for static scenes, we utilize the off-the-shelf Structure-from-Motion toolkit COLMAP [35, 36]. The dynamic aspect of our dataset is represented through multi-view RGB videos, recorded from three distinct viewpoints, at a resolution of \\(1920 \\times 1080\\). The camera parameters for dynamic scenes are obtained through calibration using a checkerboard. SAM [15] is used to obtain object masks." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.407, + 0.457, + 0.423 + ], + "angle": 0, + "content": "4.2 Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.429, + 0.788, + 0.503 + ], + "angle": 0, + "content": "When learning the 3D Gaussians, the distribution of Gaussian kernels is highly dependent on the initial points. Gaussian kernels tend to concentrate on the surface if the initial points are derived from SfM results. However, we initialize a large number of points inside a cube, resulting in a more uniform distribution of the final kernels." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.505, + 0.789, + 0.763 + ], + "angle": 0, + "content": "For real-world samples, we collect static and dynamic scenes separately because three viewpoints are insufficient for effective 3D Gaussian reconstruction [14]. This results in the Gaussian kernels for static and dynamic scenes being in different coordinate systems. Therefore, for real samples, we employ a registration network before dynamic reconstruction to align the Gaussian kernels from static scene coordinates to dynamic scene coordinates, as shown in Fig. 3. Specifically, we optimize a scale factor \\( s_r \\), a translation vector \\( \\pmb{t}_r \\), and a rotation vector \\( \\pmb{r}_r \\) for the registration. We represent 3D rotations using a continuous 6D vector \\( \\pmb{r}_r \\in \\mathbb{R}^6 \\), which has been shown to be more amenable for gradient-based optimization [53]. However, slight deformations of objects and variations in lighting conditions and exposure times during data capture are inevitable. These variations can result in color discrepancies between frames across time and viewpoints. The color discrepancy is the major noise source that hinders registration and reconstruction. Thus, we refine our approach by computing the loss function between mask images, incorporating both mask center loss and perceptual loss into our model. Consequently, the revised loss function for real samples is expressed as:" + }, + { + "type": "equation", + "bbox": [ + 0.241, + 0.786, + 0.787, + 0.803 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\left(1 - \\lambda_ {\\mathrm {d - s s i m}}\\right) \\mathcal {L} _ {1} + \\lambda_ {\\mathrm {d - s s i m}} \\mathcal {L} _ {\\mathrm {d - s s i m}} + \\lambda_ {\\text {c e n t e r}} \\mathcal {L} _ {\\text {c e n t e r}} + \\lambda_ {\\text {p e r c e p}} \\mathcal {L} _ {\\text {p e r c e p}}, \\tag {19}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.789, + 0.842 + ], + "angle": 0, + "content": "where \\(\\lambda_{\\mathrm{d - ssim}} = 0.8\\), \\(\\lambda_{\\mathrm{center}} = 1.0\\), and \\(\\lambda_{\\mathrm{percep}} = 0.1\\) are the weighting coefficients. Here, \\(\\mathcal{L}_{\\mathrm{center}}\\) quantifies the discrepancy between the center coordinates of" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.647, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Spring-Gaus" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "table", + "bbox": [ + 0.218, + 0.144, + 0.784, + 0.276 + ], + "angle": 0, + "content": "
TorusCrossCreamApplePasteChessBananaMean
CD↓Spring-Gaus (ours)2.381.572.221.877.032.5918.485.16
PAC-NeRF [18]2.473.872.214.6937.708.2066.4317.94
EMD↓Spring-Gaus (ours)0.0870.0510.0940.0760.1260.0950.1350.095
PAC-NeRF [18]0.0550.1110.0830.1080.1920.1550.2340.134
PSNR↑Spring-Gaus (ours)16.8316.9315.4221.5514.7116.0817.8917.06
PAC-NeRF [18]17.4614.1515.3719.9412.3215.0816.0415.77
SSIM↑Spring-Gaus (ours)0.9190.9400.8620.9020.8720.8810.9040.897
PAC-NeRF [18]0.9130.9060.8580.8780.8190.8480.8660.870
" + }, + { + "type": "table_caption", + "bbox": [ + 0.217, + 0.277, + 0.785, + 0.318 + ], + "angle": 0, + "content": "Table 1: Quantitative results of future prediction on synthetic data. Spring-Gaus excels in short-term future prediction. Meanwhile, since we separate appearance and dynamics modeling, Spring-Gaus also maintains good rendering quality." + }, + { + "type": "table", + "bbox": [ + 0.227, + 0.345, + 0.772, + 0.476 + ], + "angle": 0, + "content": "
TorusCrossCreamApplePasteChessBananaMean
CD↓Spring-Gaus (ours)0.170.480.360.380.191.802.600.85
PAC-NeRF [18]4.921.100.771.113.140.962.772.11
Dy-Gaus [26]579773479727284976429631305
4D-Gaus [44]11.121.772.872.231.953.977.134.43
EMD↓Spring-Gaus (ours)0.0400.0370.0310.0330.0220.0630.0520.040
PAC-NeRF [18]0.0560.0520.0410.0450.0540.0520.0620.052
Dy-Gaus [26]0.8570.9550.7830.9031.7390.9851.5911.116
4D-Gaus [44]0.1300.0780.0890.0880.0700.0970.1120.095
" + }, + { + "type": "table_caption", + "bbox": [ + 0.217, + 0.477, + 0.784, + 0.503 + ], + "angle": 0, + "content": "Table 2: Quantitative results of dynamic reconstruction on synthetic data. Spring-Gaus has excellent geometric accuracy in dynamic reconstruction." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.547, + 0.785, + 0.577 + ], + "angle": 0, + "content": "the rendered and ground truth images, while \\(\\mathcal{L}_{\\mathrm{percep}}\\) represents the perceptual loss [13], which is based on the VGG16 architecture [37]." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.602, + 0.565, + 0.617 + ], + "angle": 0, + "content": "4.3 Qualitative and Quantitative Results" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.629, + 0.785, + 0.749 + ], + "angle": 0, + "content": "We assess the performance of our approach using both synthetic and real-world datasets. Firstly, we validate the effectiveness of our method on synthetic datasets, utilizing the first 20 frames as our observation set for training dynamic modeling capabilities. For future prediction, we employ the subsequent 10 frames, comparing the ground truth with our model's predictions of future frames. Additionally, we benchmark our approach against the most relevant methods in dynamic scene modeling and physics-informed learning, including PAC-NeRF [18], Dynamic 3D Gaussians [26], and 4D Gaussian Splatting [44]." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.75, + 0.785, + 0.84 + ], + "angle": 0, + "content": "The qualitative results are presented in Fig. 4. We also report quantitative results by computing the Chamfer Distance (CD) and Earth Mover's Distance (EMD). In all tables, the Chamfer Distance is measured based on squared distance, with units expressed as \\(1 \\times 10^{3}mm^{2}\\). The quantitative analysis of dynamic reconstruction, shown in Tab. 2, reveals that Spring-Gaus can accurately simulate object dynamics for the synthetic data." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.373, + 0.129 + ], + "angle": 0, + "content": "L. Zhong et al." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.145, + 0.788, + 0.721 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.743, + 0.788, + 0.788 + ], + "angle": 0, + "content": "Fig.4: Qualitative results on synthetic data. Compared with PAC-NeRF [18], Dynamic 3D Gaussians [26] and 4D Gaussian Splatting [44], Spring-Gaus can maintain a good geometry and appearance while reconstructing reasonable dynamics." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.646, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Spring-Gaus" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "image", + "bbox": [ + 0.219, + 0.143, + 0.502, + 0.275 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.143, + 0.653, + 0.275 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.653, + 0.143, + 0.788, + 0.275 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.3, + 0.785, + 0.329 + ], + "angle": 0, + "content": "Fig. 5: Qualitative results of future prediction on real-world samples. Predicted dynamics closely follow real observations." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.366, + 0.785, + 0.472 + ], + "angle": 0, + "content": "In Tab. 1, we show our method's capability in predicting future frames. Our method outperforms PAC-NeRF across CD, EMD, Peak Signal-to-Noise Ratio (PSNR), and Structural Similarity Index (SSIM) metrics. We show qualitative results in Fig. 4. We observe that Spring-Gaus gives faithful reconstruction and future prediction that closely aligns with the ground truth, even though the simulation engines are different. This shows that Spring-Gaus is not only expressive but also allows efficient identification through gradient-based optimization." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.476, + 0.787, + 0.597 + ], + "angle": 0, + "content": "We also evaluate Spring-Gaus on real-world samples. It is difficult for both NeRF [28] and 3D Gaussian Splatting [14] to reconstruct the correct geometric information under extremely sparse camera views. However, due to 3D Gaussians' explicit representation that we can directly operate Gaussian kernels, we could reconstruct the static scenes and dynamic scenes in a different coordinate system and then align them using a registration network mentioned in Sec. 4.2, which is hard to do under an implicit representation. We show the registration process in Fig. 3. Results on real-world samples can be found in Fig. 5 and Fig. 1." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.637, + 0.544, + 0.652 + ], + "angle": 0, + "content": "4.4 Generalization to New Conditions" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.681, + 0.788, + 0.803 + ], + "angle": 0, + "content": "In addition to future prediction, we also show that our method essentially creates a digital asset of the object from the multi-view videos, allowing dynamic simulation under different unseen environmental conditions. In Fig. 6, we edit the boundary conditions, such as adjusting the positions and the stickiness of the ground plane. In Fig. 7, we edit physical conditions, initialization conditions, and environmental conditions, such as object properties (making them softer or harder), initial velocities, and environmental gravity. Please check our project website* for videos of the results for a more expressive illustration." + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.825, + 0.483, + 0.84 + ], + "angle": 0, + "content": "* https://zlicheng.com/spring_gaus" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.373, + 0.129 + ], + "angle": 0, + "content": "L. Zhong et al." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.144, + 0.496, + 0.217 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.23, + 0.218, + 0.293, + 0.228 + ], + "angle": 0, + "content": "High ground" + }, + { + "type": "image_caption", + "bbox": [ + 0.329, + 0.218, + 0.387, + 0.228 + ], + "angle": 0, + "content": "Mid ground" + }, + { + "type": "image_caption", + "bbox": [ + 0.419, + 0.218, + 0.479, + 0.228 + ], + "angle": 0, + "content": "Low ground" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.228, + 0.495, + 0.299 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.245, + 0.3, + 0.284, + 0.308 + ], + "angle": 0, + "content": "Normal" + }, + { + "type": "image_caption", + "bbox": [ + 0.337, + 0.3, + 0.376, + 0.308 + ], + "angle": 0, + "content": "Smooth" + }, + { + "type": "image_caption", + "bbox": [ + 0.435, + 0.3, + 0.466, + 0.309 + ], + "angle": 0, + "content": "Sticky" + }, + { + "type": "image", + "bbox": [ + 0.509, + 0.144, + 0.788, + 0.217 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.524, + 0.218, + 0.585, + 0.228 + ], + "angle": 0, + "content": "High ground" + }, + { + "type": "image_caption", + "bbox": [ + 0.621, + 0.218, + 0.678, + 0.228 + ], + "angle": 0, + "content": "Mid ground" + }, + { + "type": "image_caption", + "bbox": [ + 0.712, + 0.218, + 0.771, + 0.228 + ], + "angle": 0, + "content": "Low ground" + }, + { + "type": "image", + "bbox": [ + 0.509, + 0.228, + 0.787, + 0.299 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.538, + 0.3, + 0.577, + 0.308 + ], + "angle": 0, + "content": "Normal" + }, + { + "type": "image_caption", + "bbox": [ + 0.628, + 0.3, + 0.668, + 0.309 + ], + "angle": 0, + "content": "Smooth" + }, + { + "type": "image_caption", + "bbox": [ + 0.725, + 0.3, + 0.757, + 0.309 + ], + "angle": 0, + "content": "Sticky" + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.32, + 0.785, + 0.35 + ], + "angle": 0, + "content": "Fig. 6: We edit boundary conditions in these demos. Changing grounds' position and using smooth or sticky ground." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.372, + 0.495, + 0.445 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.246, + 0.446, + 0.284, + 0.454 + ], + "angle": 0, + "content": "Original" + }, + { + "type": "image_caption", + "bbox": [ + 0.31, + 0.446, + 0.487, + 0.454 + ], + "angle": 0, + "content": "Random init velocity Smaller & Softer" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.455, + 0.495, + 0.528 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.246, + 0.529, + 0.284, + 0.539 + ], + "angle": 0, + "content": "Original" + }, + { + "type": "image_caption", + "bbox": [ + 0.327, + 0.529, + 0.382, + 0.539 + ], + "angle": 0, + "content": "Zero gravity" + }, + { + "type": "image_caption", + "bbox": [ + 0.401, + 0.529, + 0.495, + 0.539 + ], + "angle": 0, + "content": "Random init velocity" + }, + { + "type": "image", + "bbox": [ + 0.509, + 0.372, + 0.787, + 0.445 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.536, + 0.446, + 0.574, + 0.455 + ], + "angle": 0, + "content": "Original" + }, + { + "type": "image_caption", + "bbox": [ + 0.631, + 0.446, + 0.661, + 0.455 + ], + "angle": 0, + "content": "Windy" + }, + { + "type": "image_caption", + "bbox": [ + 0.707, + 0.446, + 0.775, + 0.454 + ], + "angle": 0, + "content": "Bigger & Softer" + }, + { + "type": "image", + "bbox": [ + 0.509, + 0.455, + 0.787, + 0.528 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.536, + 0.529, + 0.574, + 0.539 + ], + "angle": 0, + "content": "Original" + }, + { + "type": "image_caption", + "bbox": [ + 0.62, + 0.529, + 0.681, + 0.54 + ], + "angle": 0, + "content": "Hypergravity" + }, + { + "type": "image_caption", + "bbox": [ + 0.706, + 0.529, + 0.776, + 0.539 + ], + "angle": 0, + "content": "Bigger & Softer" + }, + { + "type": "image_caption", + "bbox": [ + 0.228, + 0.55, + 0.773, + 0.565 + ], + "angle": 0, + "content": "Fig. 7: We edit physical conditions, initial velocities and gravities in these demos." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.603, + 0.359, + 0.619 + ], + "angle": 0, + "content": "5 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.659, + 0.788, + 0.841 + ], + "angle": 0, + "content": "In this paper, we introduce Spring-Gaus, a novel framework designed to acquire simulatable digital assets of elastic objects from video observations. By integrating a 3D Spring-Mass model, Spring-Gaus allows the reconstruction of object appearance and dynamics. A key feature of our approach is the distinct separation between the learning processes for appearance and physics, thereby circumventing potential issues with optimization interference. We evaluate Spring-Gaus on both synthetic and real-world datasets, demonstrating its capability to reconstruct geometry, appearance, and physical dynamic properties. Moreover, our method demonstrates improved capabilities by predicting short-term future dynamics under different environmental conditions. This showcases its strength in identifying physical properties from observational data and predicting the dynamics of reconstructed digital assets." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.646, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Spring-Gaus" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.147, + 0.788, + 0.193 + ], + "angle": 0, + "content": "Acknowledgments. The work is in part supported by the Amazon AICE Award, NSF RI #2211258, #2338203, ONR YIP N00014-24-1-2117, and ONR MURI N00014-22-1-2740." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.213, + 0.323, + 0.229 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.241, + 0.785, + 0.283 + ], + "angle": 0, + "content": "1. Barron, J.T., Mildenhall, B., Tancik, M., Hedman, P., Martin-Brualla, R., Srinivasan, P.P.: Mip-NeRF: A multiscale representation for anti-aliasing neural radiance fields. In: CVPR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.283, + 0.785, + 0.311 + ], + "angle": 0, + "content": "2. Chen, A., Xu, Z., Zhao, F., Zhang, X., Xiang, F., Yu, J., Su, H.: Mvsnerf: Fast generalizable radiance field reconstruction from multi-view stereo. In: ICCV (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.311, + 0.785, + 0.338 + ], + "angle": 0, + "content": "3. Chen, H., Wu, R., Grinspun, E., Zheng, C., Chen, P.Y.: Implicit neural spatial representations for time-dependent pdes. In: ICML (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.338, + 0.785, + 0.378 + ], + "angle": 0, + "content": "4. Chu, M., Liu, L., Zheng, Q., Franz, E., Seidel, H.P., Theobalt, C., Zayer, R.: Physics informed neural fields for smoke reconstruction with sparse data. ACM TOG 41(4) (jul 2022). https://doi.org/10.1145/3528223.3530169" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.379, + 0.785, + 0.419 + ], + "angle": 0, + "content": "5. Community, B.O.: Blender - a 3D modelling and rendering package. Blender Foundation, Stichting Blender Foundation, Amsterdam (2018), http://www.blender.org" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.42, + 0.785, + 0.446 + ], + "angle": 0, + "content": "6. Deng, Y., Yu, H.X., Wu, J., Zhu, B.: Learning vortex dynamics for fluid inference and prediction. In: ICLR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.446, + 0.785, + 0.473 + ], + "angle": 0, + "content": "7. Deng, Y., Yu, H.X., Zhang, D., Wu, J., Zhu, B.: Fluid simulation on neural flow maps. ACM TOG 42(6) (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.473, + 0.785, + 0.514 + ], + "angle": 0, + "content": "8. Driess, D., Huang, Z., Li, Y., Tedrake, R., Toussaint, M.: Learning multi-object dynamics with compositional neural radiance fields. In: Conference on Robot Learning. pp. 1755-1768. PMLR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.514, + 0.785, + 0.554 + ], + "angle": 0, + "content": "9. Du, T., Wu, K., Ma, P., Wah, S., Spielberg, A., Rus, D., Matusik, W.: Diffpd: Differentiable projective dynamics. ACM TOG 41(2) (nov 2021). https://doi.org/10.1145/3490168" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.555, + 0.785, + 0.582 + ], + "angle": 0, + "content": "10. Guan, S., Deng, H., Wang, Y., Yang, X.: Neurofluid: Fluid dynamics grounding with particle-driven neural radiance fields. In: ICML (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.582, + 0.785, + 0.635 + ], + "angle": 0, + "content": "1. Hu, Y., Fang, Y., Ge, Z., Qu, Z., Zhu, Y., Pradhana, A., Jiang, C.: A moving least squares material point method with displacement discontinuity and two-way rigid body coupling. ACM TOG 37(4) (jul 2018). https://doi.org/10.1145/3197517.3201293" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.635, + 0.785, + 0.664 + ], + "angle": 0, + "content": "2. Huang, Y.H., Sun, Y.T., Yang, Z., Lyu, X., Cao, Y.P., Qi, X.: Sc-gs: Sparse-controlled gaussian splatting for editable dynamic scenes. CVPR (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.664, + 0.785, + 0.691 + ], + "angle": 0, + "content": "3. Johnson, J., Alahi, A., Fei-Fei, L.: Perceptual losses for real-time style transfer and super-resolution. In: ECCV (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.691, + 0.785, + 0.718 + ], + "angle": 0, + "content": "4. Kerbl, B., Kopanas, G., Leimkuhler, T., Drettakis, G.: 3d gaussian splatting for real-time radiance field rendering. ACM TOG 42(4) (July 2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.718, + 0.785, + 0.759 + ], + "angle": 0, + "content": "5. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., Dollár, P., Girshick, R.: Segment Anything. arXiv preprint arXiv:2304.02643 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.759, + 0.785, + 0.8 + ], + "angle": 0, + "content": "6. Kratimenos, A., Lei, J., Daniilidis, K.: Dynmf: Neural motion factorization for real-time dynamic view synthesis with 3d gaussian splattering. arXiv preprint arXiv:2312.00112 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.8, + 0.785, + 0.841 + ], + "angle": 0, + "content": "7. Le Cleac'h, S., Yu, H.X., Guo, M., Howell, T., Gao, R., Wu, J., Manchester, Z., Schwager, M.: Differentiable physics simulation of dynamics-augmented neural objects. IEEE Robotics and Automation Letters 8(5), 2780-2787 (2023)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.241, + 0.785, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.372, + 0.128 + ], + "angle": 0, + "content": "L. Zhong et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.785, + 0.189 + ], + "angle": 0, + "content": "18. Li, X., Qiao, Y.L., Chen, P.Y., Jatavallabhula, K.M., Lin, M., Jiang, C., Gan, C.: PAC-NeRF: Physics augmented continuum neural radiance fields for geometry-agnostic system identification. In: ICLR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.19, + 0.785, + 0.218 + ], + "angle": 0, + "content": "19. Li, Y., Chen, P.Y., Du, T., Matusik, W.: Learning preconditioners for conjugate gradient pde solvers. In: ICML (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.218, + 0.785, + 0.258 + ], + "angle": 0, + "content": "20. Li, Y., Li, S., Sitzmann, V., Agrawal, P., Torralba, A.: 3d neural scene representations for visuomotor control. In: Conference on Robot Learning. pp. 112-123. PMLR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.259, + 0.785, + 0.3 + ], + "angle": 0, + "content": "21. Li, Y., Wu, J., Tedrake, R., Tenenbaum, J.B., Torralba, A.: Learning particle dynamics for manipulating rigid bodies, deformable objects, and fluids. arXiv preprint arXiv:1810.01566 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.3, + 0.785, + 0.328 + ], + "angle": 0, + "content": "22. Li, Z., Niklaus, S., Snavely, N., Wang, O.: Neural scene flow fields for space-time view synthesis of dynamic scenes. In: CVPR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.328, + 0.785, + 0.37 + ], + "angle": 0, + "content": "23. Li, Z., Wang, Q., Cole, F., Tucker, R., Snavely, N.: Dynibar: Neural dynamic image-based rendering. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4273-4284 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.371, + 0.785, + 0.398 + ], + "angle": 0, + "content": "24. Lin, Y., Dai, Z., Zhu, S., Yao, Y.: Gaussian-flow: 4d reconstruction with dynamic 3d gaussian particle. arXiv preprint arXiv:2312.03431 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.398, + 0.785, + 0.426 + ], + "angle": 0, + "content": "25. Liu, Y., Peng, S., Liu, L., Wang, Q., Wang, P., Christian, T., Zhou, X., Wang, W.: Neural rays for occlusion-aware image-based rendering. In: CVPR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.426, + 0.785, + 0.452 + ], + "angle": 0, + "content": "26. Luiten, J., Kopanas, G., Leibe, B., Ramanan, D.: Dynamic 3d gaussians: Tracking by persistent dynamic view synthesis. In: 3DV (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.453, + 0.785, + 0.494 + ], + "angle": 0, + "content": "27. Ma, P., Chen, P.Y., Deng, B., Tenenbaum, J.B., Du, T., Gan, C., Matusik, W.: Learning neural constitutive laws from motion observations for generalizable pde dynamics. In: ICML (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.495, + 0.785, + 0.536 + ], + "angle": 0, + "content": "28. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: NeRF: Representing scenes as neural radiance fields for view synthesis. In: ECCV (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.536, + 0.785, + 0.564 + ], + "angle": 0, + "content": "29. Park, J.J., Florence, P., Straub, J., Newcombe, R., Lovegrove, S.: Deepsdf: Learning continuous signed distance functions for shape representation. In: CVPR (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.564, + 0.785, + 0.592 + ], + "angle": 0, + "content": "30. Park, K., Sinha, U., Barron, J.T., Bouaziz, S., Goldman, D.B., Seitz, S.M., Martin-Brualla, R.: Nerfies: Deformable neural radiance fields. In: ICCV (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.592, + 0.785, + 0.633 + ], + "angle": 0, + "content": "31. Park, K., Sinha, U., Hedman, P., Barron, J.T., Bouaziz, S., Goldman, D.B., Martin-Brualla, R., Seitz, S.M.: Hypernerf: A higher-dimensional representation for topologically varying neural radiance fields. arXiv preprint arXiv:2106.13228 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.633, + 0.785, + 0.661 + ], + "angle": 0, + "content": "32. Pumarola, A., Corona, E., Pons-Moll, G., Moreno-Noguer, F.: D-NeRF: Neural Radiance Fields for Dynamic Scenes. In: CVPR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.661, + 0.785, + 0.716 + ], + "angle": 0, + "content": "33. Raissi, M., Perdikaris, P., Karniadakis, G.: Physics-informed neural networks: A deep learning framework for solving forward and inverse problems involving nonlinear partial differential equations. Journal of Computational Physics 378, 686-707 (2019). https://doi.org/https://doi.org/10.1016/j.jcp.2018.10.045" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.716, + 0.785, + 0.744 + ], + "angle": 0, + "content": "34. Sanchez-Gonzalez, A., Godwin, J., Pfaff, T., Ying, R., Leskovec, J., Battaglia, P.W.: Learning to simulate complex physics with graph networks. In: ICML (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.744, + 0.785, + 0.758 + ], + "angle": 0, + "content": "35. Schonberger, J.L., Frahm, J.M.: Structure-from-motion revisited. In: CVPR (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.758, + 0.785, + 0.785 + ], + "angle": 0, + "content": "36. Schonberger, J.L., Zheng, E., Pollefeys, M., Frahm, J.M.: Pixelwise view selection for unstructured multi-view stereo. In: ECCV (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.785, + 0.785, + 0.813 + ], + "angle": 0, + "content": "37. Simonyan, K., Zisserman, A.: Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556 (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.813, + 0.785, + 0.841 + ], + "angle": 0, + "content": "38. Sitzmann, V., Zollhöfer, M., Wetzstein, G.: Scene representation networks: Continuous 3d-structure-aware neural scene representations. In: NeurIPS (2019)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.646, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Spring-Gaus" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.188 + ], + "angle": 0, + "content": "39. Stomakhin, A., Schroeder, C., Chai, L., Teran, J., Selle, A.: A material point method for snow simulation. ACM TOG 32(4) (jul 2013). https://doi.org/10.1145/2461912.2461948" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.189, + 0.788, + 0.217 + ], + "angle": 0, + "content": "40. Sun, C., Sun, M., Chen, H.: Direct voxel grid optimization: Super-fast convergence for radiance fields reconstruction. In: CVPR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.218, + 0.786, + 0.245 + ], + "angle": 0, + "content": "41. Tang, J., Ren, J., Zhou, H., Liu, Z., Zeng, G.: Dreamgaussian: Generative gaussian splatting for efficient 3d content creation. ICLR (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.246, + 0.787, + 0.286 + ], + "angle": 0, + "content": "42. Tretschk, E., Tewari, A., Golyanik, V., Zollhöfer, M., Lassner, C., Theobalt, C.: Non-rigid neural radiance fields: Reconstruction and novel view synthesis of a dynamic scene from monocular video. In: ICCV (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.287, + 0.787, + 0.327 + ], + "angle": 0, + "content": "43. Wang, C., MacDonald, L.E., Jeni, L.A., Lucey, S.: Flow supervision for deformable nerf. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 21128-21137 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.328, + 0.787, + 0.356 + ], + "angle": 0, + "content": "44. Wu, G., Yi, T., Fang, J., Xie, L., Zhang, X., Wei, W., Liu, W., Tian, Q., Xinggang, W.: 4d gaussian splatting for real-time dynamic scene rendering. In: CVPR (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.357, + 0.787, + 0.397 + ], + "angle": 0, + "content": "45. Wu, T., Zhang, J., Fu, X., Wang, Y., Ren, J., Pan, L., Wu, W., Yang, L., Wang, J., Qian, C., Lin, D., Liu, Z.: OmniObject3D: Large-vocabulary 3d object dataset for realistic perception, reconstruction and generation. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.398, + 0.787, + 0.425 + ], + "angle": 0, + "content": "46. Xie, T., Zong, Z., Qiu, Y., Li, X., Feng, Y., Yang, Y., Jiang, C.: Physsgaussian: Physics-integrated 3d gaussians for generative dynamics. In: CVPR (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.426, + 0.787, + 0.452 + ], + "angle": 0, + "content": "47. Xu, Q., Xu, Z., Philip, J., Bi, S., Shu, Z., Sunkavalli, K., Neumann, U.: Point-NeRF: Point-based neural radiance fields. In: CVPR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.453, + 0.787, + 0.48 + ], + "angle": 0, + "content": "48. Yang, Z., Yang, H., Pan, Z., Zhang, L.: Real-time photorealistic dynamic scene representation and rendering with 4d gaussian splatting. In: ICLR (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.481, + 0.787, + 0.508 + ], + "angle": 0, + "content": "49. Yang, Z., Gao, X., Zhou, W., Jiao, S., Zhang, Y., Jin, X.: Deformable 3d gaussians for high-fidelity monocular dynamic scene reconstruction. In: CVPR (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.509, + 0.787, + 0.536 + ], + "angle": 0, + "content": "50. Yu, A., Ye, V., Tancik, M., Kanazawa, A.: pixelNeRF: Neural radiance fields from one or few images. In: CVPR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.537, + 0.787, + 0.563 + ], + "angle": 0, + "content": "51. Yu, H., Julin, J., Milacski, Z.A., Niinuma, K., Jeni, L.A.: Cogs: Controllable gaussian splatting. arXiv preprint arXiv:2312.05664 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.565, + 0.787, + 0.591 + ], + "angle": 0, + "content": "52. Yu, H.X., Zheng, Y., Gao, Y., Deng, Y., Zhu, B., Wu, J.: Inferring hybrid neural fluid fields from videos. Advances in Neural Information Processing Systems (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.592, + 0.787, + 0.619 + ], + "angle": 0, + "content": "53. Zhou, Y., Barnes, C., Lu, J., Yang, J., Li, H.: On the continuity of rotation representations in neural networks. In: CVPR (2019)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.619 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/Reconstruction and Simulation of Elastic Objects with Spring-Mass 3D Gaussians/2eb8194f-eadd-43df-b4b2-bbf94906f9cc_origin.pdf b/2024/Reconstruction and Simulation of Elastic Objects with Spring-Mass 3D Gaussians/2eb8194f-eadd-43df-b4b2-bbf94906f9cc_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..2d12670f3242daa82b075590aa946a091df9bac0 --- /dev/null +++ b/2024/Reconstruction and Simulation of Elastic Objects with Spring-Mass 3D Gaussians/2eb8194f-eadd-43df-b4b2-bbf94906f9cc_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c22c05ee59174764e47c09194e9c26f74bd58a445891654c0107008fef8f6d43 +size 7612896 diff --git a/2024/Reconstruction and Simulation of Elastic Objects with Spring-Mass 3D Gaussians/full.md b/2024/Reconstruction and Simulation of Elastic Objects with Spring-Mass 3D Gaussians/full.md new file mode 100644 index 0000000000000000000000000000000000000000..7084a240c257f374439ba5d7465ffae4691146a6 --- /dev/null +++ b/2024/Reconstruction and Simulation of Elastic Objects with Spring-Mass 3D Gaussians/full.md @@ -0,0 +1,364 @@ +# Reconstruction and Simulation of Elastic Objects with Spring-Mass 3D Gaussians + +Licheng Zhong $^{1*}$ , Hong-Xing Yu $^{1}$ , Jiajun Wu $^{1}$ , and Yunzhu Li $^{1,2,3}$ + +1 Stanford University, Stanford, USA + +2 Columbia University, New York, USA + +3 University of Illinois Urbana-Champaign, Champaign, USA + +Abstract. Reconstructing and simulating elastic objects from visual observations is crucial for applications in computer vision and robotics. Existing methods, such as 3D Gaussians, model 3D appearance and geometry, but lack the ability to estimate physical properties for objects and simulate them. The core challenge lies in integrating an expressive yet efficient physical dynamics model. We propose Spring-Gaus, a 3D physical object representation for reconstructing and simulating elastic objects from videos of the object from multiple viewpoints. In particular, we develop and integrate a 3D Spring-Mass model into 3D Gaussian kernels, enabling the reconstruction of the visual appearance, shape, and physical dynamics of the object. Our approach enables future prediction and simulation under various initial states and environmental properties. We evaluate Spring-Gaus on both synthetic and real-world datasets, demonstrating accurate reconstruction and simulation of elastic objects. Project page: https://zlicheng.com/spring_gaus. + +Keywords: System Identification $\cdot$ Spring-Mass Model $\cdot$ 3D Gaussians + +# 1 Introduction + +Reconstructing and simulating elastic objects from visual observations poses a fundamental challenge in computer vision and robotics, with applications spanning virtual reality, augmented reality, and robotic manipulation. Accurately modeling the elasticity of objects is crucial for creating immersive experiences and enabling embodied agents to understand and interact with the elastic objects commonly encountered in our daily lives. However, accurately identifying the dynamics from vision still presents considerable challenges. + +Existing methods for dynamic scene reconstruction, such as 3D Gaussians and their dynamic extensions [14, 26, 44, 49], have made significant progress in capturing the temporal changes of the appearance and geometry of objects. However, these methods do not capture the physical properties of the reconstructed + +![](images/d477639a8b9254b185e6cd3fc9c49c626e3b8e907ec7ac3f3524e68e64bded53.jpg) +Fig. 1: Spring-Gaus reconstructs the appearance, geometry, and physical dynamic properties of elastic objects from video observations. Spring-Gaus enables future predictions and simulations under different initial states and environmental conditions. + +objects; thus, they typically cannot predict the future dynamics of these objects. While a few recent approaches using MPM [11,39] have attempted to integrate physics-based priors into 3D object representations, such as PAC-NeRF [18]. Their ability to handle real, especially heterogeneous, objects is limited, as they assume a known material model and only assigns a global physical parameter to the entire object which restricts its adaptability to real objects. Assigning learnable physical parameters to each particle in an MPM is theoretically possible. However, in practice, it incurs extreme computational costs since MPM requires tens of thousands of dense points. In addition, PAC-NeRF [18] use a implicit grid representation, due to the computational demands, the resolution of the grid is limited, which can lead to a loss of detail in the appearance modeling when using real or noisy data. Thus, the core challenge in reconstructing and predicting object dynamics lies in developing and integrating an expressive and efficient physical model for the dynamics. The physical dynamics model should be expressive enough to capture the motions of elastic objects, including collisions, deformations, and bouncing. It must also be efficient and conducive to inverse parameter estimation through gradient-based optimization. + +In this work, we propose Spring-Gaus, a 3D object representation that integrates a 3D Spring-Mass model. Our model represents the elastic object dynamics properties through a learnable system of mass points and springs. Our design is expressive in that it assumes a general and widely applicable physical model class. With a learnable topology and physical parameters, Spring-Gaus can model complex deformation and motion for heterogeneous elastic objects. In addition to expressiveness, Spring-Gaus is also highly efficient for the inverse optimization of physical parameters due to its differentiable nature. + +Spring-Gaus enables reconstructing and simulating elastic objects from sparse multi-view videos (Fig. 1). To overcome the intrinsic difficulty in optimization, + +we propose a reconstruction pipeline that decouples physical parameter reconstruction from appearance and geometry reconstruction. Spring-Gaus requires only a few multi-view videos for physical parameter identification and is robust to the quality of geometry reconstruction. + +We evaluate the effectiveness of Spring-Gaus on both synthetic and real-world datasets, demonstrating its ability to accurately reconstruct and simulate elastic objects. Our Spring-Gaus allows accurate future prediction and simulations under varying initial states and environmental parameters, showcasing its potential for applications in predictive visual perception and immersive experiences. In summary, the main contributions of this work are threefold: + +- We propose Spring-Gaus, which incorporates an expressive yet efficient 3D Spring-Mass model for reconstructing and simulating elastic objects. +- We introduce a pipeline to reconstruct Spring-Gaus from multi-view videos of the object. We decouple the appearance and geometry reconstruction from the physical dynamics reconstruction for more effective optimization. +- We demonstrate the effectiveness of Spring-Gaus on both synthetic and real-world datasets, showcasing accurate reconstruction and simulation of elastic objects. This includes capabilities for future prediction and simulation under varying initial configurations. + +# 2 Related Work + +3D Object Representations: Traditionally, 3D objects are often represented by point clouds, meshes, and voxels. Recently, neural 3D object representation has become popular due to the efficiency and flexibility. For example, Scene Representation Networks (SRNs) [38] and DeepSDF [29] represent significant advancements in 3D scene and object representation, treating 3D objects as continuous functions that map world coordinates to a feature representation of local object properties. Over the past few years, NeRF [28] and its successors [1, 2, 25, 47, 50] have demonstrated the efficacy of neural networks in capturing continuous 3D scenes and objects through implicit representation. DirectVoxGO [40] accelerates NeRF's approach by substituting the MLP with a voxel grid. Furthermore, 3D Gaussian Splatting [14] has emerged as a method for real-time differentiable rendering, representing scenes and objects with 3D Gaussians. Extending this approach, DreamGaussian [41] applies 3D Gaussians for 3D object generation. Unlike these approaches, our method focuses on reconstructing simulatable 3D objects. Recently, PhysGaussian [46] integrated physical simulation into 3D Gaussians using a customized Material Point Method, allowing forward simulation of reconstructed objects. In contrast to PhysGaussian, our work focuses on system identification from raw videos and supports both the forward simulation and the inverse reconstruction of physical objects. + +Dynamic Scene Reconstruction: The modeling of dynamic scenes has seen significant progress with the adoption of NeRF [8, 10, 18, 20, 22, 23, 30-32, 42, 43] and 3D Gaussian [12, 16, 24, 26, 44, 48, 49, 51] representations. D-NeRF [32] introduces an extension to NeRF, capable of modeling dynamic scenes from monocu + +lar views by optimizing an underlying deformable volumetric function. Furthermore, Dynamic 3D Gaussians [26] focus on optimizing the motion of Gaussian kernels for each frame, presenting an efficient method to capture scene dynamics. Deformable 3D Gaussians [49] propose a novel approach by learning Gaussian distributions in a canonical space, complemented with a deformation field for modeling monocular dynamic scenes. Meanwhile, 4D Gaussian Splatting [44] introduces a hybrid model combining 3D Gaussians with 4D neural voxels. PAC-NeRF [18] delves into the integration of Lagrangian particle simulation with Eulerian scene representation, exploring a new aspect of scene dynamics. Among them, PAC-NeRF [18] is the most relevant to our work. PAC-NeRF learns physical parameters used in the Material Point Method (MPM) for better dynamic reconstruction and prediction. However, PAC-NeRF assumes known material models that restrict its adaptability and applicability to complex real objects. In contrast, our work integrates a 3D Spring-Mass model that is both expressive and efficient, allowing reconstructing and simulating real elastic objects. + +Physics-Informed Learning: Physics-Informed Learning has emerged as a prominent research direction since the introduction of Physics-Informed Neural Networks (PINNs) [33]. Li et al. [21] and Sanchez-Gonzalez et al. [34] introduced Graph Network-based simulators within a machine learning framework. INSR-PDE [3] tackles time-dependent partial differential equations (PDEs) using implicit neural spatial representations. NCLaw [27] focuses on learning neural constitutive laws for PDE dynamics. DiffPD [9] presents a differentiable soft-body simulator. Li et al. [19] have made strides in learning preconditioners for conjugate gradient PDE solvers. Chu et al. [4] and Yu et al. [52] modeled smoke in neural density and velocity fields. Neural Flow Maps [7] integrate fluid simulation with neural implicit representations. Deng et al. [6] introduced a novel differentiable vortex particle method for fluid dynamics inference. DANO [17] is the most relevant work to ours. In particular, DANO develops a differentiable simulation for rigid objects represented by NeRFs, allowing reconstruction and simulation from a few videos. In contrast, our work focuses on elastic objects, which involve a fundamentally new set of challenges and require a redesign of the physical models. + +# 3 Approach + +We propose Spring-Gaus that integrates a 3D Spring-Mass model with 3D Gaussians for elastic object reconstruction and simulation. Specifically, we are given a set of $O$ calibrated videos from multiple viewpoints of an elastic object in motion, denoted as $\{I^{o,f}\}_{o = 1,f = 1}^{O,F}$ , where $F$ is the number of video frames and $I^{o,f}$ denotes a 2D image at video $o$ and frame $f$ . Our goal is to reconstruct the appearance, geometry, and physical dynamics parameters of Spring-Gaus. + +In the following, we first introduce the representation for the static properties, i.e., appearance and geometry, and then the dynamics properties, i.e., the 3D Spring-Mass model. Then we describe our reconstruction pipeline. We show an overview of our reconstruction pipeline in Fig. 2. + +# 3.1 Appearance and Geometry Representation + +Following 3D Gaussians Splatting [14], we use a set of 3D Gaussian kernels to represent the appearance and shape of an object. These kernels are described by a full 3D covariance matrix $\Sigma$ , defined in world space and centered at point $\pmb{\mu}$ : + +$$ +G (\boldsymbol {\tau}) = e ^ {- \frac {1}{2} (\boldsymbol {\tau}) ^ {T} \Sigma^ {- 1} (\boldsymbol {\tau})}, \tag {1} +$$ + +where $\pmb{\mu}$ is the mean of the Gaussian distribution, and $\pmb{\tau}$ is the independent variable of the Gaussian distribution. In our formulation, we assume all Gaussian kernels are isotropic, the covariance matrix $\Sigma$ can be controlled by a scalar $s\in \mathbb{R}$ . + +The splatting process, designed to render Gaussian kernels into a 2D image, involves two main steps: (1) projecting the Gaussian kernels into camera space, and (2) rendering the projected kernels into image space. The projection process is defined as: + +$$ +\Sigma^ {\prime} = J W \Sigma W ^ {T} J ^ {T}, \tag {2} +$$ + +where $\Sigma'$ represents the covariance matrix in camera coordinates, $W$ is the transformation matrix, and $J$ denotes the Jacobian of the affine approximation of the projective transformation. The color $C$ of each pixel is rendered as: + +$$ +C = \sum_ {i \in N} T _ {i} \alpha_ {i} \boldsymbol {c} _ {i}, \tag {3} +$$ + +where $N$ is the total number of kernels and $T_{i}$ represents the transmittance, defined as $\Pi_{j = 1}^{i - 1}\alpha_{j}$ . The term $\alpha_{i}$ denotes the alpha value for each Gaussian, which is calculated using the expression $1 - e^{-\sigma_i\delta_i}$ , where $\sigma_{i}$ is the opacity factor. Additionally, $\pmb{c}_i$ refers to the color of the Gaussian along the ray within the interval $\delta_{i}$ . + +For each kernel, the learnable parameters include a point center $\pmb{\mu}$ , a scaling scalar $s$ , a color vector $\pmb{c}$ , and an opacity value $\sigma$ . + +# 3.2 Physics-Based Dynamics with 3D Spring-Mass Model + +We aim to develop and integrate an expressive yet efficient physics-based dynamics model. With these two design goals in mind, we introduce the 3D Spring-Mass model which represents the elastic object dynamics by a learnable spring-mass system. Our Spring-Gaus does not assume any material type, and it is expressive to model real heterogeneous elastic objects. Besides being expressive, our dynamics model should also be efficient and amenable to gradient-based optimization for parameter estimation. + +Recall that the appearance and geometry representations include a set of Gaussian kernels $\{G_i\}_{i=1}^N$ , parameterized by point centers $\mathbf{X} = \{\pmb{\mu}_i\}_{i=1}^N$ , scaling scalars $\{s_i\}_{i=1}^N$ , color vectors $\{\pmb{c}_i\}_{i=1}^N$ , and opacity values $\{\sigma_i\}_{i=1}^N$ . To manage complexity for efficient simulation, we introduce volume sampling to generate a set of anchor points $\mathbf{A} = \{\pmb{x}_i\}_{i=1}^{N_A}$ (each $\pmb{x}$ represents a spatial point), defined by: + +$$ +\boldsymbol {A} = \left\{\boldsymbol {x} _ {i} \right\} _ {i = 1} ^ {N _ {A}} = \mathcal {V} (\boldsymbol {X}), \tag {4} +$$ + +![](images/dc0f7fb28cd35083fb30769d417e6b17eb077004dc422f078951c8e33f244093.jpg) +Fig.2: Overview of Spring-Gaus reconstruction pipeline: (a) Static Scene Reconstruction: We start by reconstructing static 3D Gaussians from the first frames of the multiview videos. (b) Refining 3D Gaussians: We extract a set of anchor points to allow efficient simulation, which leads to appearance drift. We refine the 3D Gaussians to better model the appearance during simulation. (c) Dynamic Reconstruction: Our 3D Spring-Mass model simulates anchor points and updates the positions of Gaussian kernels. Upon completion of optimization, we obtain a simulatable 3D object that accurately models its dynamics. + +where $N_{A}$ denotes the number of anchors, and $\nu$ denotes the sampling function. + +Our 3D Spring-Mass physical model can simulate the motion of anchor points $\mathbf{A}$ , assuming each anchor has a mass $m_{i}$ and an initial velocity $\mathbf{v}_{i}$ . Each anchor $\mathbf{x}_{i}$ connects to its $n_{k}$ nearest neighbors $\mathbf{N}_{i} = \{\mathbf{x}_{i,j}\}_{j=1}^{n_{k}}$ through springs $\mathbf{L}$ : + +$$ +\boldsymbol {L} = \left\{l _ {i, j} \right\} _ {i = 1, j = 1} ^ {N _ {A}, n _ {k}} = \operatorname {k n n} (\boldsymbol {A}, \boldsymbol {A}, n _ {k}), \tag {5} +$$ + +where $l_{i,j}$ denotes the spring's length between $\pmb{x}_i$ and $\pmb{x}_{i,j}$ , and knn denotes the k-nearest neighbors function. Each spring is characterized by a stiffness $k_{i,j}$ and a damping factor $\zeta_{i,j}$ . + +To update the positions of the kernels, we first measure the distance between each kernel center and its $n_b$ nearest anchors at the dynamic simulation's onset: + +$$ +\left\{d _ {i, j} \right\} _ {i = 1, j = 1} ^ {N, n _ {b}} = \operatorname {k n n} (\boldsymbol {X}, \boldsymbol {A}, n _ {b}). \tag {6} +$$ + +For each timestep $t$ , the forces $\pmb{F}_i^t$ acting on each anchor point $\pmb{x}_i^t$ are calculated as follows: + +$$ +\boldsymbol {F} _ {i} ^ {t} = \boldsymbol {F} _ {\boldsymbol {k} i} ^ {t} + \boldsymbol {F} _ {\boldsymbol {\zeta} i} ^ {t} + m _ {i} \boldsymbol {g}, \tag {7} +$$ + +where $\pmb{F}_{\pmb{k}_i}^t, \pmb{F}_{\pmb{\zeta}_i}^t$ , and $\pmb{g}$ represent the spring force, the damping force, and gravitational acceleration, respectively. + +Then, for each spring $\pmb{L}_{i,j}$ , the spring force $\pmb{F}_{\pmb{k}_{i,j}}^{t}$ and damping force $\pmb{F}_{\pmb{\zeta}_{i,j}}^{t}$ are determined by: + +$$ +\boldsymbol {F} _ {\boldsymbol {k} _ {i, j}} ^ {t} = - \eta_ {j} \cdot k _ {i, j} \left(\left\| \boldsymbol {x} _ {i} ^ {t} - \boldsymbol {x} _ {i, j} ^ {t} \right\| - l _ {i, j}\right) \frac {\boldsymbol {x} _ {i} ^ {t} - \boldsymbol {x} _ {i , j} ^ {t}}{\left\| \boldsymbol {x} _ {i} ^ {t} - \boldsymbol {x} _ {i , j} ^ {t} \right\|} \cdot \left| \left\| \boldsymbol {x} _ {i} ^ {t} - \boldsymbol {x} _ {i, j} ^ {t} \right\| - l _ {i, j} \right| ^ {p _ {k}}, \tag {8} +$$ + +$$ +\boldsymbol {F} _ {\zeta_ {i, j}} ^ {t} = \left(- \zeta_ {i, j} \left(\boldsymbol {v} _ {i} ^ {t} - \boldsymbol {v} _ {i, j} ^ {t}\right) \frac {\boldsymbol {x} _ {i} ^ {t} - \boldsymbol {x} _ {i , j} ^ {t}}{\left\| \boldsymbol {x} _ {i} ^ {t} - \boldsymbol {x} _ {i , j} ^ {t} \right\|}\right) \cdot \frac {\boldsymbol {x} _ {i} ^ {t} - \boldsymbol {x} _ {i , j} ^ {t}}{\left\| \boldsymbol {x} _ {i} ^ {t} - \boldsymbol {x} _ {i , j} ^ {t} \right\|}, \tag {9} +$$ + +where $\eta$ is a soft vector which will be discussed later and $p_k$ is a hyperparameter that determines the nonlinearity of the spring force. When $p_k$ is set to 0, Eq. (8) becomes Hooke's law, and for positive values of $p_k$ , the spring force becomes a nonlinear function of the spring's length. The cumulative forces acting on each anchor point $\boldsymbol{x}_i^t$ are expressed as: + +$$ +\boldsymbol {F} _ {i} ^ {t} = \sum_ {j = 1} ^ {n _ {k}} \boldsymbol {F} _ {\boldsymbol {k} _ {i, j}} ^ {t} + \sum_ {j = 1} ^ {n _ {k}} \boldsymbol {F} _ {\zeta_ {i, j}} ^ {t} + m _ {i} \boldsymbol {g}. \tag {10} +$$ + +Anchor points $\mathbf{A}$ 's positions and velocities are updated using semi-implicit Euler integration: + +$$ +\hat {\boldsymbol {v}} _ {i} ^ {t + 1} = \boldsymbol {v} _ {i} ^ {t} + \frac {\boldsymbol {F} _ {i} ^ {t}}{m _ {i}} \Delta t, \tag {11} +$$ + +$$ +\hat {\boldsymbol {x}} _ {i} ^ {t + 1} = \boldsymbol {x} _ {i} ^ {t} + \boldsymbol {v} _ {i} ^ {t + 1} \Delta t, \tag {12} +$$ + +and a boundary condition $\mathcal{B}$ is applied to the anchor points $\pmb{A}$ to model the interactions with the environment: + +$$ +\boldsymbol {x} _ {i} ^ {t + 1}, \boldsymbol {v} _ {i} ^ {t + 1} = \mathcal {B} \left(\hat {\boldsymbol {x}} _ {i} ^ {t + 1}, \hat {\boldsymbol {v}} _ {i} ^ {t + 1}\right). \tag {13} +$$ + +The position of each Gaussian kernel $\pmb{\mu}_i$ is updated through Inverse Distance Weighting (IDW) interpolation to reflect the dynamic changes accurately: + +$$ +\boldsymbol {\mu} _ {i} ^ {t + 1} = \frac {\sum_ {j = 1} ^ {n _ {b}} \boldsymbol {x} _ {i , j} ^ {t + 1} \cdot \left(1 / \left(d _ {i , j}\right) ^ {p _ {b}}\right)}{\sum_ {j = 1} ^ {n _ {b}} \left(1 / \left(d _ {i , j}\right) ^ {p _ {b}}\right)}, \tag {14} +$$ + +where $p_b$ is a positive real number that determines the diminishing influence of anchor points with distance. To render the image $\hat{I}$ at a specific camera and frame, we use the updated positions of the Gaussian kernels following the rendering equation Eq. (3). + +Soft Vector for Springs Connection: In our formulation above, the $n_k$ is a hyperparameter which is the number of connected springs for each anchor. However, the choice of $n_k$ will directly affect the simulation results markedly. The bigger value of $n_k$ , the object will behave more rigidly, and a smaller value of $n_k$ leads to a noticeably softer behavior of the point cloud. To address the + +significant impact that the value of $n_k$ has on the simulation, we introduce a mitigation strategy by applying a soft vector $\eta = [\eta_0, \eta_1, \dots, \eta_{n_k}]$ . This vector controlled by a learnable parameter $\kappa$ (shared by all anchors) is used to modulate the number of connected springs, thereby adjusting the system's response to a different object. Given an empirical value $n_c$ , the soft vector $\eta$ is calculated as: + +$$ +\eta_ {j} = \left\{ \begin{array}{l l} 1 & j \leq n _ {c}, \\ \operatorname {c l a m p} (2 - \exp (\operatorname {s o f t p l u s} (\kappa)) ^ {j - n _ {c}}, 0, 1) & n _ {c} < j \leq n _ {k}. \end{array} \right. \tag {15} +$$ + +# 3.3 Optimization + +To allow efficient optimization, we simplify our model by reducing the number of learnable parameters without changing the essential expressiveness. In our simplified approach, we standardize the mass of every anchor to a constant value $m_0$ , and control all damping factors using a singular parameter $\zeta_0$ . These two parameters are fixed, eliminating their variability from the optimization process. Furthermore, we introduce a unified parameter $k_{i}$ for each anchor $\pmb{x}_i$ to control the spring stiffness of the springs connected to it, simplifying the model without compromising its functional integrity. The spring stiffness $k_{i,j}$ and damping factor $\zeta_{i,j}$ are thus given as follows: + +$$ +k _ {i, j} = k _ {i} / l _ {i, j}, \tag {16} +$$ + +$$ +\zeta_ {i, j} = \zeta_ {0} / l _ {i, j}. \tag {17} +$$ + +Note that, in Eq. (17) stiffness is defined at anchor points rather than on springs themselves to simplify the optimization—we only need to optimize $N_{A}$ stiffness coefficients instead of all the springs, which is on the order of $n_k \cdot N_A$ . By this simplification, our model maintains computational efficiency and ease of optimization, while still capturing the essential dynamics of the object. + +Summarized, the learnable parameters in our model now include: + +- $\pmb{v}_0$ : the initial velocity vector, providing a baseline movement pattern for the simulation; +- $\{k_i\}_{i=1}^{N_A}$ : the individual stiffness parameters for each anchor, allowing for localized adjustments to spring stiffness; +- $\kappa$ : a parameter governing the modulation of the soft vector, facilitating finetuned control over the spring dynamics; +- $\Theta(\mathcal{B}(\cdot))$ : parameters defining the boundary conditions ( $\Theta$ is learnable parameters), such as the friction coefficient, which influence the simulation's physical realism. + +There exist $n_t$ timesteps between each of the two keyframes. At each timestep, $\boldsymbol{x}^t$ is computed from $\boldsymbol{x}^{t-1}$ using the update rule Eqs. (11) and (12). We only optimize physical parameters at each keyframe (when $t = 0, n_t, 2n_t, 3n_t, \ldots$ ) based on the visual observations until that time. We increase the value of $n_t$ as the parameters converge. This approach balances identification accuracy and computational demand. + +Same with 3D Gaussian Splatting [14], we define our loss function as a weighted combination of the $\mathcal{L}_1$ norm and the Structural Similarity Index Measure (D-SSIM) $\mathcal{L}_{\mathrm{d - ssim}}$ , applied between the input images $I$ and the rendered images $\hat{I}$ . This is formally expressed as: + +$$ +\mathcal {L} = \left(1 - \lambda_ {\mathrm {d - s s i m}}\right) \mathcal {L} _ {1} + \lambda_ {\mathrm {d - s s i m}} \mathcal {L} _ {\mathrm {d - s s i m}}, \tag {18} +$$ + +where $\lambda_{\mathrm{d - ssim}}$ is a weighting coefficient that balances the contribution of the $\mathcal{L}_1$ norm and the D-SSIM term to the overall loss. We decouple our optimization into a few stages, as illustrated in Fig. 2. + +Static Reconstruction: Our optimization starts by taking the first frames of the multi-view videos and using them to reconstruct the appearance and geometry of the object. + +3D Gaussians Refinement: During dynamic simulation of the anchor points, the 3D Gaussian centers $\mathbf{X}_0 = \{\pmb{\mu}_i^0\}_{i=1}^N$ are computed by IDW interpolation outlined in Eq. (14). This leads to a slight appearance drift from the static reconstruction. Therefore, we refine the parameters of the Gaussian kernels—scaling scalar $s$ , color vector $c$ , and opacity value $\sigma$ , excluding points center $\pmb{\mu}$ —also at the first frame, before the dynamic simulation. Since the Gaussian kernels' position from IDW interpolation is slightly different from the static reconstruction results, this refinement process enables the Gaussian kernels with the interpolated spatial position to render the correct object appearance. + +Dynamic Reconstruction: Since our efficient dynamics simulation is fully differentiable with respect to the learnable parameters, we can optimize the physical parameters through differentiable simulation and differentiable rendering with our loss function $\mathcal{L}$ . + +# 4 Experiment + +# 4.1 Datasets + +Synthetic Data. We evaluate Spring-Gaus using a synthetic dataset. We first collected fourteen 3D models, including point clouds and meshes, to use as initial point clouds for our simulations. Some of them are sourced from PAC-NeRF [18] and OmniObject3D [45]. Following PAC-NeRF, we employ the Material Point Method [11,39] to simulate the dynamics of elastic objects to generate synthetic data. Our dataset features elastic objects with various stiffness levels and diverse geometric forms. Multi-view RGB videos are rendered using Blender [5], with each sequence comprising 10 viewpoints and 30 frames at a resolution of $512 \times 512$ . The initial 20 frames are utilized for dynamic reconstruction, while the subsequent 10 frames are dedicated to evaluating future prediction performance. + +Real-World Data. In addition to synthetic datasets, we further assess Spring-Gaus using captured real-world examples. Our collection process distinguishes between static scenes and dynamic multi-view videos. For static scenes, we position each object on a table and capture 50-70 images from various viewpoints within the upper hemisphere surrounding the object. To obtain camera + +![](images/f94c41c50fd8443bb7e3cbda1bfb71df2f2b25a9084a735fa1dc61606ee2d19d.jpg) +Fig. 3: Registration from static scene to dynamic scene for real-world sample. + +poses for static scenes, we utilize the off-the-shelf Structure-from-Motion toolkit COLMAP [35, 36]. The dynamic aspect of our dataset is represented through multi-view RGB videos, recorded from three distinct viewpoints, at a resolution of $1920 \times 1080$ . The camera parameters for dynamic scenes are obtained through calibration using a checkerboard. SAM [15] is used to obtain object masks. + +# 4.2 Implementation Details + +When learning the 3D Gaussians, the distribution of Gaussian kernels is highly dependent on the initial points. Gaussian kernels tend to concentrate on the surface if the initial points are derived from SfM results. However, we initialize a large number of points inside a cube, resulting in a more uniform distribution of the final kernels. + +For real-world samples, we collect static and dynamic scenes separately because three viewpoints are insufficient for effective 3D Gaussian reconstruction [14]. This results in the Gaussian kernels for static and dynamic scenes being in different coordinate systems. Therefore, for real samples, we employ a registration network before dynamic reconstruction to align the Gaussian kernels from static scene coordinates to dynamic scene coordinates, as shown in Fig. 3. Specifically, we optimize a scale factor $s_r$ , a translation vector $\pmb{t}_r$ , and a rotation vector $\pmb{r}_r$ for the registration. We represent 3D rotations using a continuous 6D vector $\pmb{r}_r \in \mathbb{R}^6$ , which has been shown to be more amenable for gradient-based optimization [53]. However, slight deformations of objects and variations in lighting conditions and exposure times during data capture are inevitable. These variations can result in color discrepancies between frames across time and viewpoints. The color discrepancy is the major noise source that hinders registration and reconstruction. Thus, we refine our approach by computing the loss function between mask images, incorporating both mask center loss and perceptual loss into our model. Consequently, the revised loss function for real samples is expressed as: + +$$ +\mathcal {L} = \left(1 - \lambda_ {\mathrm {d - s s i m}}\right) \mathcal {L} _ {1} + \lambda_ {\mathrm {d - s s i m}} \mathcal {L} _ {\mathrm {d - s s i m}} + \lambda_ {\text {c e n t e r}} \mathcal {L} _ {\text {c e n t e r}} + \lambda_ {\text {p e r c e p}} \mathcal {L} _ {\text {p e r c e p}}, \tag {19} +$$ + +where $\lambda_{\mathrm{d - ssim}} = 0.8$ , $\lambda_{\mathrm{center}} = 1.0$ , and $\lambda_{\mathrm{percep}} = 0.1$ are the weighting coefficients. Here, $\mathcal{L}_{\mathrm{center}}$ quantifies the discrepancy between the center coordinates of + +
TorusCrossCreamApplePasteChessBananaMean
CD↓Spring-Gaus (ours)2.381.572.221.877.032.5918.485.16
PAC-NeRF [18]2.473.872.214.6937.708.2066.4317.94
EMD↓Spring-Gaus (ours)0.0870.0510.0940.0760.1260.0950.1350.095
PAC-NeRF [18]0.0550.1110.0830.1080.1920.1550.2340.134
PSNR↑Spring-Gaus (ours)16.8316.9315.4221.5514.7116.0817.8917.06
PAC-NeRF [18]17.4614.1515.3719.9412.3215.0816.0415.77
SSIM↑Spring-Gaus (ours)0.9190.9400.8620.9020.8720.8810.9040.897
PAC-NeRF [18]0.9130.9060.8580.8780.8190.8480.8660.870
+ +Table 1: Quantitative results of future prediction on synthetic data. Spring-Gaus excels in short-term future prediction. Meanwhile, since we separate appearance and dynamics modeling, Spring-Gaus also maintains good rendering quality. + +
TorusCrossCreamApplePasteChessBananaMean
CD↓Spring-Gaus (ours)0.170.480.360.380.191.802.600.85
PAC-NeRF [18]4.921.100.771.113.140.962.772.11
Dy-Gaus [26]579773479727284976429631305
4D-Gaus [44]11.121.772.872.231.953.977.134.43
EMD↓Spring-Gaus (ours)0.0400.0370.0310.0330.0220.0630.0520.040
PAC-NeRF [18]0.0560.0520.0410.0450.0540.0520.0620.052
Dy-Gaus [26]0.8570.9550.7830.9031.7390.9851.5911.116
4D-Gaus [44]0.1300.0780.0890.0880.0700.0970.1120.095
+ +Table 2: Quantitative results of dynamic reconstruction on synthetic data. Spring-Gaus has excellent geometric accuracy in dynamic reconstruction. + +the rendered and ground truth images, while $\mathcal{L}_{\mathrm{percep}}$ represents the perceptual loss [13], which is based on the VGG16 architecture [37]. + +# 4.3 Qualitative and Quantitative Results + +We assess the performance of our approach using both synthetic and real-world datasets. Firstly, we validate the effectiveness of our method on synthetic datasets, utilizing the first 20 frames as our observation set for training dynamic modeling capabilities. For future prediction, we employ the subsequent 10 frames, comparing the ground truth with our model's predictions of future frames. Additionally, we benchmark our approach against the most relevant methods in dynamic scene modeling and physics-informed learning, including PAC-NeRF [18], Dynamic 3D Gaussians [26], and 4D Gaussian Splatting [44]. + +The qualitative results are presented in Fig. 4. We also report quantitative results by computing the Chamfer Distance (CD) and Earth Mover's Distance (EMD). In all tables, the Chamfer Distance is measured based on squared distance, with units expressed as $1 \times 10^{3}mm^{2}$ . The quantitative analysis of dynamic reconstruction, shown in Tab. 2, reveals that Spring-Gaus can accurately simulate object dynamics for the synthetic data. + +![](images/34cf5ad2c67f59e147f820c12520acb53b250549f8d061319f616d4a63d3a380.jpg) +Fig.4: Qualitative results on synthetic data. Compared with PAC-NeRF [18], Dynamic 3D Gaussians [26] and 4D Gaussian Splatting [44], Spring-Gaus can maintain a good geometry and appearance while reconstructing reasonable dynamics. + +![](images/27a05d1c66777d2fd18a08deff01dff3718388705359443008465c2f3b43667c.jpg) +Fig. 5: Qualitative results of future prediction on real-world samples. Predicted dynamics closely follow real observations. + +![](images/8d99da0f958e146ff55a5265e9359c90e49105e0361bde02951c869b35470d31.jpg) + +![](images/5bfb97ef473387211a109ed2e243fb1d4a2ad13dfbcccd07566af81eb1cc0427.jpg) + +In Tab. 1, we show our method's capability in predicting future frames. Our method outperforms PAC-NeRF across CD, EMD, Peak Signal-to-Noise Ratio (PSNR), and Structural Similarity Index (SSIM) metrics. We show qualitative results in Fig. 4. We observe that Spring-Gaus gives faithful reconstruction and future prediction that closely aligns with the ground truth, even though the simulation engines are different. This shows that Spring-Gaus is not only expressive but also allows efficient identification through gradient-based optimization. + +We also evaluate Spring-Gaus on real-world samples. It is difficult for both NeRF [28] and 3D Gaussian Splatting [14] to reconstruct the correct geometric information under extremely sparse camera views. However, due to 3D Gaussians' explicit representation that we can directly operate Gaussian kernels, we could reconstruct the static scenes and dynamic scenes in a different coordinate system and then align them using a registration network mentioned in Sec. 4.2, which is hard to do under an implicit representation. We show the registration process in Fig. 3. Results on real-world samples can be found in Fig. 5 and Fig. 1. + +# 4.4 Generalization to New Conditions + +In addition to future prediction, we also show that our method essentially creates a digital asset of the object from the multi-view videos, allowing dynamic simulation under different unseen environmental conditions. In Fig. 6, we edit the boundary conditions, such as adjusting the positions and the stickiness of the ground plane. In Fig. 7, we edit physical conditions, initialization conditions, and environmental conditions, such as object properties (making them softer or harder), initial velocities, and environmental gravity. Please check our project website* for videos of the results for a more expressive illustration. + +![](images/9f337b1a4dfdecbd32704379c1f3ec4bd5c7167f08bf89dff96b245df825ec26.jpg) + +![](images/d001745f6a0c502ff1a4cddedd883c09c35e7bec2847e6c7057958dafb16b536.jpg) +High ground +Mid ground +Low ground +Normal +Smooth +Sticky +Fig. 6: We edit boundary conditions in these demos. Changing grounds' position and using smooth or sticky ground. + +![](images/9cde7ca8efb75f7977e40c20db31ec234cd393d79e98fc96a6b7a4ba94e707ad.jpg) + +![](images/1de20ed7fdca102612c668c48950a28f36722e6c834a0cd2963d87fc952e0ca3.jpg) +High ground +Mid ground +Low ground +Normal +Smooth +Sticky + +![](images/0a88e24a00358f1c56ae036d6b43b306c630fb103b27f88cc53fec03e1f6f5b4.jpg) +Original +Random init velocity Smaller & Softer + +![](images/06428440e0a26eb043d0b775992f8e03c68576ec37876ffec49a59072a3474be.jpg) +Original +Zero gravity +Random init velocity +Fig. 7: We edit physical conditions, initial velocities and gravities in these demos. + +![](images/5a55546295d3f00f0a1fbcac37791e9955fd20b9112c0c70f4c6f4b29f6b2ff4.jpg) +Bigger & Softer + +![](images/a93dea573f604609d7f3fb16ccc63210efc1a2f50bb65ac310d9ebca036de7be.jpg) +Original +Windy +Original +Hypergravity +Bigger & Softer + +# 5 Conclusion + +In this paper, we introduce Spring-Gaus, a novel framework designed to acquire simulatable digital assets of elastic objects from video observations. By integrating a 3D Spring-Mass model, Spring-Gaus allows the reconstruction of object appearance and dynamics. A key feature of our approach is the distinct separation between the learning processes for appearance and physics, thereby circumventing potential issues with optimization interference. We evaluate Spring-Gaus on both synthetic and real-world datasets, demonstrating its capability to reconstruct geometry, appearance, and physical dynamic properties. Moreover, our method demonstrates improved capabilities by predicting short-term future dynamics under different environmental conditions. This showcases its strength in identifying physical properties from observational data and predicting the dynamics of reconstructed digital assets. + +Acknowledgments. The work is in part supported by the Amazon AICE Award, NSF RI #2211258, #2338203, ONR YIP N00014-24-1-2117, and ONR MURI N00014-22-1-2740. + +# References + +1. Barron, J.T., Mildenhall, B., Tancik, M., Hedman, P., Martin-Brualla, R., Srinivasan, P.P.: Mip-NeRF: A multiscale representation for anti-aliasing neural radiance fields. In: CVPR (2021) +2. Chen, A., Xu, Z., Zhao, F., Zhang, X., Xiang, F., Yu, J., Su, H.: Mvsnerf: Fast generalizable radiance field reconstruction from multi-view stereo. In: ICCV (2021) +3. Chen, H., Wu, R., Grinspun, E., Zheng, C., Chen, P.Y.: Implicit neural spatial representations for time-dependent pdes. In: ICML (2023) +4. Chu, M., Liu, L., Zheng, Q., Franz, E., Seidel, H.P., Theobalt, C., Zayer, R.: Physics informed neural fields for smoke reconstruction with sparse data. ACM TOG 41(4) (jul 2022). https://doi.org/10.1145/3528223.3530169 +5. Community, B.O.: Blender - a 3D modelling and rendering package. Blender Foundation, Stichting Blender Foundation, Amsterdam (2018), http://www.blender.org +6. Deng, Y., Yu, H.X., Wu, J., Zhu, B.: Learning vortex dynamics for fluid inference and prediction. In: ICLR (2023) +7. Deng, Y., Yu, H.X., Zhang, D., Wu, J., Zhu, B.: Fluid simulation on neural flow maps. ACM TOG 42(6) (2023) +8. Driess, D., Huang, Z., Li, Y., Tedrake, R., Toussaint, M.: Learning multi-object dynamics with compositional neural radiance fields. In: Conference on Robot Learning. pp. 1755-1768. PMLR (2023) +9. Du, T., Wu, K., Ma, P., Wah, S., Spielberg, A., Rus, D., Matusik, W.: Diffpd: Differentiable projective dynamics. ACM TOG 41(2) (nov 2021). https://doi.org/10.1145/3490168 +10. Guan, S., Deng, H., Wang, Y., Yang, X.: Neurofluid: Fluid dynamics grounding with particle-driven neural radiance fields. In: ICML (2022) +1. Hu, Y., Fang, Y., Ge, Z., Qu, Z., Zhu, Y., Pradhana, A., Jiang, C.: A moving least squares material point method with displacement discontinuity and two-way rigid body coupling. ACM TOG 37(4) (jul 2018). https://doi.org/10.1145/3197517.3201293 +2. Huang, Y.H., Sun, Y.T., Yang, Z., Lyu, X., Cao, Y.P., Qi, X.: Sc-gs: Sparse-controlled gaussian splatting for editable dynamic scenes. CVPR (2024) +3. Johnson, J., Alahi, A., Fei-Fei, L.: Perceptual losses for real-time style transfer and super-resolution. In: ECCV (2016) +4. Kerbl, B., Kopanas, G., Leimkuhler, T., Drettakis, G.: 3d gaussian splatting for real-time radiance field rendering. ACM TOG 42(4) (July 2023) +5. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., Dollár, P., Girshick, R.: Segment Anything. arXiv preprint arXiv:2304.02643 (2023) +6. Kratimenos, A., Lei, J., Daniilidis, K.: Dynmf: Neural motion factorization for real-time dynamic view synthesis with 3d gaussian splattering. arXiv preprint arXiv:2312.00112 (2023) +7. Le Cleac'h, S., Yu, H.X., Guo, M., Howell, T., Gao, R., Wu, J., Manchester, Z., Schwager, M.: Differentiable physics simulation of dynamics-augmented neural objects. IEEE Robotics and Automation Letters 8(5), 2780-2787 (2023) + +18. Li, X., Qiao, Y.L., Chen, P.Y., Jatavallabhula, K.M., Lin, M., Jiang, C., Gan, C.: PAC-NeRF: Physics augmented continuum neural radiance fields for geometry-agnostic system identification. In: ICLR (2023) +19. Li, Y., Chen, P.Y., Du, T., Matusik, W.: Learning preconditioners for conjugate gradient pde solvers. In: ICML (2023) +20. Li, Y., Li, S., Sitzmann, V., Agrawal, P., Torralba, A.: 3d neural scene representations for visuomotor control. In: Conference on Robot Learning. pp. 112-123. PMLR (2022) +21. Li, Y., Wu, J., Tedrake, R., Tenenbaum, J.B., Torralba, A.: Learning particle dynamics for manipulating rigid bodies, deformable objects, and fluids. arXiv preprint arXiv:1810.01566 (2018) +22. Li, Z., Niklaus, S., Snavely, N., Wang, O.: Neural scene flow fields for space-time view synthesis of dynamic scenes. In: CVPR (2021) +23. Li, Z., Wang, Q., Cole, F., Tucker, R., Snavely, N.: Dynibar: Neural dynamic image-based rendering. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4273-4284 (2023) +24. Lin, Y., Dai, Z., Zhu, S., Yao, Y.: Gaussian-flow: 4d reconstruction with dynamic 3d gaussian particle. arXiv preprint arXiv:2312.03431 (2023) +25. Liu, Y., Peng, S., Liu, L., Wang, Q., Wang, P., Christian, T., Zhou, X., Wang, W.: Neural rays for occlusion-aware image-based rendering. In: CVPR (2022) +26. Luiten, J., Kopanas, G., Leibe, B., Ramanan, D.: Dynamic 3d gaussians: Tracking by persistent dynamic view synthesis. In: 3DV (2024) +27. Ma, P., Chen, P.Y., Deng, B., Tenenbaum, J.B., Du, T., Gan, C., Matusik, W.: Learning neural constitutive laws from motion observations for generalizable pde dynamics. In: ICML (2023) +28. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: NeRF: Representing scenes as neural radiance fields for view synthesis. In: ECCV (2020) +29. Park, J.J., Florence, P., Straub, J., Newcombe, R., Lovegrove, S.: Deepsdf: Learning continuous signed distance functions for shape representation. In: CVPR (2019) +30. Park, K., Sinha, U., Barron, J.T., Bouaziz, S., Goldman, D.B., Seitz, S.M., Martin-Brualla, R.: Nerfies: Deformable neural radiance fields. In: ICCV (2021) +31. Park, K., Sinha, U., Hedman, P., Barron, J.T., Bouaziz, S., Goldman, D.B., Martin-Brualla, R., Seitz, S.M.: Hypernerf: A higher-dimensional representation for topologically varying neural radiance fields. arXiv preprint arXiv:2106.13228 (2021) +32. Pumarola, A., Corona, E., Pons-Moll, G., Moreno-Noguer, F.: D-NeRF: Neural Radiance Fields for Dynamic Scenes. In: CVPR (2021) +33. Raissi, M., Perdikaris, P., Karniadakis, G.: Physics-informed neural networks: A deep learning framework for solving forward and inverse problems involving nonlinear partial differential equations. Journal of Computational Physics 378, 686-707 (2019). https://doi.org/https://doi.org/10.1016/j.jcp.2018.10.045 +34. Sanchez-Gonzalez, A., Godwin, J., Pfaff, T., Ying, R., Leskovec, J., Battaglia, P.W.: Learning to simulate complex physics with graph networks. In: ICML (2020) +35. Schonberger, J.L., Frahm, J.M.: Structure-from-motion revisited. In: CVPR (2016) +36. Schonberger, J.L., Zheng, E., Pollefeys, M., Frahm, J.M.: Pixelwise view selection for unstructured multi-view stereo. In: ECCV (2016) +37. Simonyan, K., Zisserman, A.: Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556 (2014) +38. Sitzmann, V., Zollhöfer, M., Wetzstein, G.: Scene representation networks: Continuous 3d-structure-aware neural scene representations. In: NeurIPS (2019) + +39. Stomakhin, A., Schroeder, C., Chai, L., Teran, J., Selle, A.: A material point method for snow simulation. ACM TOG 32(4) (jul 2013). https://doi.org/10.1145/2461912.2461948 +40. Sun, C., Sun, M., Chen, H.: Direct voxel grid optimization: Super-fast convergence for radiance fields reconstruction. In: CVPR (2022) +41. Tang, J., Ren, J., Zhou, H., Liu, Z., Zeng, G.: Dreamgaussian: Generative gaussian splatting for efficient 3d content creation. ICLR (2024) +42. Tretschk, E., Tewari, A., Golyanik, V., Zollhöfer, M., Lassner, C., Theobalt, C.: Non-rigid neural radiance fields: Reconstruction and novel view synthesis of a dynamic scene from monocular video. In: ICCV (2021) +43. Wang, C., MacDonald, L.E., Jeni, L.A., Lucey, S.: Flow supervision for deformable nerf. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 21128-21137 (2023) +44. Wu, G., Yi, T., Fang, J., Xie, L., Zhang, X., Wei, W., Liu, W., Tian, Q., Xinggang, W.: 4d gaussian splatting for real-time dynamic scene rendering. In: CVPR (2024) +45. Wu, T., Zhang, J., Fu, X., Wang, Y., Ren, J., Pan, L., Wu, W., Yang, L., Wang, J., Qian, C., Lin, D., Liu, Z.: OmniObject3D: Large-vocabulary 3d object dataset for realistic perception, reconstruction and generation. In: CVPR (2023) +46. Xie, T., Zong, Z., Qiu, Y., Li, X., Feng, Y., Yang, Y., Jiang, C.: Physsgaussian: Physics-integrated 3d gaussians for generative dynamics. In: CVPR (2024) +47. Xu, Q., Xu, Z., Philip, J., Bi, S., Shu, Z., Sunkavalli, K., Neumann, U.: Point-NeRF: Point-based neural radiance fields. In: CVPR (2022) +48. Yang, Z., Yang, H., Pan, Z., Zhang, L.: Real-time photorealistic dynamic scene representation and rendering with 4d gaussian splatting. In: ICLR (2024) +49. Yang, Z., Gao, X., Zhou, W., Jiao, S., Zhang, Y., Jin, X.: Deformable 3d gaussians for high-fidelity monocular dynamic scene reconstruction. In: CVPR (2024) +50. Yu, A., Ye, V., Tancik, M., Kanazawa, A.: pixelNeRF: Neural radiance fields from one or few images. In: CVPR (2021) +51. Yu, H., Julin, J., Milacski, Z.A., Niinuma, K., Jeni, L.A.: Cogs: Controllable gaussian splatting. arXiv preprint arXiv:2312.05664 (2023) +52. Yu, H.X., Zheng, Y., Gao, Y., Deng, Y., Zhu, B., Wu, J.: Inferring hybrid neural fluid fields from videos. Advances in Neural Information Processing Systems (2023) +53. Zhou, Y., Barnes, C., Lu, J., Yang, J., Li, H.: On the continuity of rotation representations in neural networks. In: CVPR (2019) \ No newline at end of file diff --git a/2024/Reconstruction and Simulation of Elastic Objects with Spring-Mass 3D Gaussians/images.zip b/2024/Reconstruction and Simulation of Elastic Objects with Spring-Mass 3D Gaussians/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..d2a16d24645bea0a747c7ea594923157b5219695 --- /dev/null +++ b/2024/Reconstruction and Simulation of Elastic Objects with Spring-Mass 3D Gaussians/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4d27c8f23d726fa9bb2e0e291a2907879cf6290c0944e2b0c4fccd8d4930ed75 +size 659830 diff --git a/2024/Reconstruction and Simulation of Elastic Objects with Spring-Mass 3D Gaussians/layout.json b/2024/Reconstruction and Simulation of Elastic Objects with Spring-Mass 3D Gaussians/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..00a6cca23d2c67ea58ff0701e1f59947f5065fc3 --- /dev/null +++ b/2024/Reconstruction and Simulation of Elastic Objects with Spring-Mass 3D Gaussians/layout.json @@ -0,0 +1,10918 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 135, + 111, + 479, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 111, + 479, + 148 + ], + "spans": [ + { + "bbox": [ + 135, + 111, + 479, + 148 + ], + "type": "text", + "content": "Reconstruction and Simulation of Elastic Objects with Spring-Mass 3D Gaussians" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 159, + 167, + 455, + 181 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 167, + 455, + 181 + ], + "spans": [ + { + "bbox": [ + 159, + 167, + 455, + 181 + ], + "type": "text", + "content": "Licheng Zhong" + }, + { + "bbox": [ + 159, + 167, + 455, + 181 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 159, + 167, + 455, + 181 + ], + "type": "text", + "content": ", Hong-Xing Yu" + }, + { + "bbox": [ + 159, + 167, + 455, + 181 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 159, + 167, + 455, + 181 + ], + "type": "text", + "content": ", Jiajun Wu" + }, + { + "bbox": [ + 159, + 167, + 455, + 181 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 159, + 167, + 455, + 181 + ], + "type": "text", + "content": ", and Yunzhu Li" + }, + { + "bbox": [ + 159, + 167, + 455, + 181 + ], + "type": "inline_equation", + "content": "^{1,2,3}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 230, + 189, + 384, + 201 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 230, + 189, + 384, + 201 + ], + "spans": [ + { + "bbox": [ + 230, + 189, + 384, + 201 + ], + "type": "text", + "content": "1 Stanford University, Stanford, USA" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 225, + 201, + 388, + 212 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 225, + 201, + 388, + 212 + ], + "spans": [ + { + "bbox": [ + 225, + 201, + 388, + 212 + ], + "type": "text", + "content": "2 Columbia University, New York, USA" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 181, + 212, + 432, + 223 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 181, + 212, + 432, + 223 + ], + "spans": [ + { + "bbox": [ + 181, + 212, + 432, + 223 + ], + "type": "text", + "content": "3 University of Illinois Urbana-Champaign, Champaign, USA" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 159, + 253, + 455, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 253, + 455, + 418 + ], + "spans": [ + { + "bbox": [ + 159, + 253, + 455, + 418 + ], + "type": "text", + "content": "Abstract. Reconstructing and simulating elastic objects from visual observations is crucial for applications in computer vision and robotics. Existing methods, such as 3D Gaussians, model 3D appearance and geometry, but lack the ability to estimate physical properties for objects and simulate them. The core challenge lies in integrating an expressive yet efficient physical dynamics model. We propose Spring-Gaus, a 3D physical object representation for reconstructing and simulating elastic objects from videos of the object from multiple viewpoints. In particular, we develop and integrate a 3D Spring-Mass model into 3D Gaussian kernels, enabling the reconstruction of the visual appearance, shape, and physical dynamics of the object. Our approach enables future prediction and simulation under various initial states and environmental properties. We evaluate Spring-Gaus on both synthetic and real-world datasets, demonstrating accurate reconstruction and simulation of elastic objects. Project page: https://zlicheng.com/spring_gaus." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 160, + 428, + 453, + 440 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 428, + 453, + 440 + ], + "spans": [ + { + "bbox": [ + 160, + 428, + 453, + 440 + ], + "type": "text", + "content": "Keywords: System Identification " + }, + { + "bbox": [ + 160, + 428, + 453, + 440 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 428, + 453, + 440 + ], + "type": "text", + "content": " Spring-Mass Model " + }, + { + "bbox": [ + 160, + 428, + 453, + 440 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 428, + 453, + 440 + ], + "type": "text", + "content": " 3D Gaussians" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 478, + 230, + 491 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 478, + 230, + 491 + ], + "spans": [ + { + "bbox": [ + 132, + 478, + 230, + 491 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 503, + 482, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 503, + 482, + 586 + ], + "spans": [ + { + "bbox": [ + 130, + 503, + 482, + 586 + ], + "type": "text", + "content": "Reconstructing and simulating elastic objects from visual observations poses a fundamental challenge in computer vision and robotics, with applications spanning virtual reality, augmented reality, and robotic manipulation. Accurately modeling the elasticity of objects is crucial for creating immersive experiences and enabling embodied agents to understand and interact with the elastic objects commonly encountered in our daily lives. However, accurately identifying the dynamics from vision still presents considerable challenges." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 587, + 482, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 587, + 482, + 635 + ], + "spans": [ + { + "bbox": [ + 130, + 587, + 482, + 635 + ], + "type": "text", + "content": "Existing methods for dynamic scene reconstruction, such as 3D Gaussians and their dynamic extensions [14, 26, 44, 49], have made significant progress in capturing the temporal changes of the appearance and geometry of objects. However, these methods do not capture the physical properties of the reconstructed" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 642, + 481, + 666 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 642, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 133, + 642, + 481, + 666 + ], + "type": "text", + "content": "* The work was done while L. Zhong was a visiting student at Stanford University. L. Zhong is now at Shanghai Jiao Tong University." + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 114, + 481, + 270 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 481, + 270 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 481, + 270 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 481, + 270 + ], + "type": "image", + "image_path": "d477639a8b9254b185e6cd3fc9c49c626e3b8e907ec7ac3f3524e68e64bded53.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 278, + 482, + 312 + ], + "lines": [ + { + "bbox": [ + 130, + 278, + 482, + 312 + ], + "spans": [ + { + "bbox": [ + 130, + 278, + 482, + 312 + ], + "type": "text", + "content": "Fig. 1: Spring-Gaus reconstructs the appearance, geometry, and physical dynamic properties of elastic objects from video observations. Spring-Gaus enables future predictions and simulations under different initial states and environmental conditions." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 339, + 482, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 339, + 482, + 542 + ], + "spans": [ + { + "bbox": [ + 130, + 339, + 482, + 542 + ], + "type": "text", + "content": "objects; thus, they typically cannot predict the future dynamics of these objects. While a few recent approaches using MPM [11,39] have attempted to integrate physics-based priors into 3D object representations, such as PAC-NeRF [18]. Their ability to handle real, especially heterogeneous, objects is limited, as they assume a known material model and only assigns a global physical parameter to the entire object which restricts its adaptability to real objects. Assigning learnable physical parameters to each particle in an MPM is theoretically possible. However, in practice, it incurs extreme computational costs since MPM requires tens of thousands of dense points. In addition, PAC-NeRF [18] use a implicit grid representation, due to the computational demands, the resolution of the grid is limited, which can lead to a loss of detail in the appearance modeling when using real or noisy data. Thus, the core challenge in reconstructing and predicting object dynamics lies in developing and integrating an expressive and efficient physical model for the dynamics. The physical dynamics model should be expressive enough to capture the motions of elastic objects, including collisions, deformations, and bouncing. It must also be efficient and conducive to inverse parameter estimation through gradient-based optimization." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 544, + 482, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 544, + 482, + 640 + ], + "spans": [ + { + "bbox": [ + 130, + 544, + 482, + 640 + ], + "type": "text", + "content": "In this work, we propose Spring-Gaus, a 3D object representation that integrates a 3D Spring-Mass model. Our model represents the elastic object dynamics properties through a learnable system of mass points and springs. Our design is expressive in that it assumes a general and widely applicable physical model class. With a learnable topology and physical parameters, Spring-Gaus can model complex deformation and motion for heterogeneous elastic objects. In addition to expressiveness, Spring-Gaus is also highly efficient for the inverse optimization of physical parameters due to its differentiable nature." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "content": "Spring-Gaus enables reconstructing and simulating elastic objects from sparse multi-view videos (Fig. 1). To overcome the intrinsic difficulty in optimization," + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 228, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 228, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 228, + 102 + ], + "type": "text", + "content": "L. Zhong et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "text", + "content": "we propose a reconstruction pipeline that decouples physical parameter reconstruction from appearance and geometry reconstruction. Spring-Gaus requires only a few multi-view videos for physical parameter identification and is robust to the quality of geometry reconstruction." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 164, + 480, + 235 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 164, + 480, + 235 + ], + "spans": [ + { + "bbox": [ + 130, + 164, + 480, + 235 + ], + "type": "text", + "content": "We evaluate the effectiveness of Spring-Gaus on both synthetic and real-world datasets, demonstrating its ability to accurately reconstruct and simulate elastic objects. Our Spring-Gaus allows accurate future prediction and simulations under varying initial states and environmental parameters, showcasing its potential for applications in predictive visual perception and immersive experiences. In summary, the main contributions of this work are threefold:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 242, + 479, + 347 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 138, + 242, + 479, + 265 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 242, + 479, + 265 + ], + "spans": [ + { + "bbox": [ + 138, + 242, + 479, + 265 + ], + "type": "text", + "content": "- We propose Spring-Gaus, which incorporates an expressive yet efficient 3D Spring-Mass model for reconstructing and simulating elastic objects." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 266, + 479, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 266, + 479, + 300 + ], + "spans": [ + { + "bbox": [ + 138, + 266, + 479, + 300 + ], + "type": "text", + "content": "- We introduce a pipeline to reconstruct Spring-Gaus from multi-view videos of the object. We decouple the appearance and geometry reconstruction from the physical dynamics reconstruction for more effective optimization." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 301, + 479, + 347 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 301, + 479, + 347 + ], + "spans": [ + { + "bbox": [ + 138, + 301, + 479, + 347 + ], + "type": "text", + "content": "- We demonstrate the effectiveness of Spring-Gaus on both synthetic and real-world datasets, showcasing accurate reconstruction and simulation of elastic objects. This includes capabilities for future prediction and simulation under varying initial configurations." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 132, + 366, + 236, + 379 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 366, + 236, + 379 + ], + "spans": [ + { + "bbox": [ + 132, + 366, + 236, + 379 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 390, + 480, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 390, + 480, + 616 + ], + "spans": [ + { + "bbox": [ + 130, + 390, + 480, + 616 + ], + "type": "text", + "content": "3D Object Representations: Traditionally, 3D objects are often represented by point clouds, meshes, and voxels. Recently, neural 3D object representation has become popular due to the efficiency and flexibility. For example, Scene Representation Networks (SRNs) [38] and DeepSDF [29] represent significant advancements in 3D scene and object representation, treating 3D objects as continuous functions that map world coordinates to a feature representation of local object properties. Over the past few years, NeRF [28] and its successors [1, 2, 25, 47, 50] have demonstrated the efficacy of neural networks in capturing continuous 3D scenes and objects through implicit representation. DirectVoxGO [40] accelerates NeRF's approach by substituting the MLP with a voxel grid. Furthermore, 3D Gaussian Splatting [14] has emerged as a method for real-time differentiable rendering, representing scenes and objects with 3D Gaussians. Extending this approach, DreamGaussian [41] applies 3D Gaussians for 3D object generation. Unlike these approaches, our method focuses on reconstructing simulatable 3D objects. Recently, PhysGaussian [46] integrated physical simulation into 3D Gaussians using a customized Material Point Method, allowing forward simulation of reconstructed objects. In contrast to PhysGaussian, our work focuses on system identification from raw videos and supports both the forward simulation and the inverse reconstruction of physical objects." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 617, + 480, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 480, + 665 + ], + "type": "text", + "content": "Dynamic Scene Reconstruction: The modeling of dynamic scenes has seen significant progress with the adoption of NeRF [8, 10, 18, 20, 22, 23, 30-32, 42, 43] and 3D Gaussian [12, 16, 24, 26, 44, 48, 49, 51] representations. D-NeRF [32] introduces an extension to NeRF, capable of modeling dynamic scenes from monocu" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 395, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 395, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 395, + 91, + 447, + 102 + ], + "type": "text", + "content": "Spring-Gaus" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 295 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 295 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 295 + ], + "type": "text", + "content": "lar views by optimizing an underlying deformable volumetric function. Furthermore, Dynamic 3D Gaussians [26] focus on optimizing the motion of Gaussian kernels for each frame, presenting an efficient method to capture scene dynamics. Deformable 3D Gaussians [49] propose a novel approach by learning Gaussian distributions in a canonical space, complemented with a deformation field for modeling monocular dynamic scenes. Meanwhile, 4D Gaussian Splatting [44] introduces a hybrid model combining 3D Gaussians with 4D neural voxels. PAC-NeRF [18] delves into the integration of Lagrangian particle simulation with Eulerian scene representation, exploring a new aspect of scene dynamics. Among them, PAC-NeRF [18] is the most relevant to our work. PAC-NeRF learns physical parameters used in the Material Point Method (MPM) for better dynamic reconstruction and prediction. However, PAC-NeRF assumes known material models that restrict its adaptability and applicability to complex real objects. In contrast, our work integrates a 3D Spring-Mass model that is both expressive and efficient, allowing reconstructing and simulating real elastic objects." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 296, + 482, + 499 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 296, + 482, + 499 + ], + "spans": [ + { + "bbox": [ + 130, + 296, + 482, + 499 + ], + "type": "text", + "content": "Physics-Informed Learning: Physics-Informed Learning has emerged as a prominent research direction since the introduction of Physics-Informed Neural Networks (PINNs) [33]. Li et al. [21] and Sanchez-Gonzalez et al. [34] introduced Graph Network-based simulators within a machine learning framework. INSR-PDE [3] tackles time-dependent partial differential equations (PDEs) using implicit neural spatial representations. NCLaw [27] focuses on learning neural constitutive laws for PDE dynamics. DiffPD [9] presents a differentiable soft-body simulator. Li et al. [19] have made strides in learning preconditioners for conjugate gradient PDE solvers. Chu et al. [4] and Yu et al. [52] modeled smoke in neural density and velocity fields. Neural Flow Maps [7] integrate fluid simulation with neural implicit representations. Deng et al. [6] introduced a novel differentiable vortex particle method for fluid dynamics inference. DANO [17] is the most relevant work to ours. In particular, DANO develops a differentiable simulation for rigid objects represented by NeRFs, allowing reconstruction and simulation from a few videos. In contrast, our work focuses on elastic objects, which involve a fundamentally new set of challenges and require a redesign of the physical models." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 517, + 212, + 533 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 517, + 212, + 533 + ], + "spans": [ + { + "bbox": [ + 132, + 517, + 212, + 533 + ], + "type": "text", + "content": "3 Approach" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 544, + 481, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 544, + 481, + 617 + ], + "spans": [ + { + "bbox": [ + 130, + 544, + 481, + 617 + ], + "type": "text", + "content": "We propose Spring-Gaus that integrates a 3D Spring-Mass model with 3D Gaussians for elastic object reconstruction and simulation. Specifically, we are given a set of " + }, + { + "bbox": [ + 130, + 544, + 481, + 617 + ], + "type": "inline_equation", + "content": "O" + }, + { + "bbox": [ + 130, + 544, + 481, + 617 + ], + "type": "text", + "content": " calibrated videos from multiple viewpoints of an elastic object in motion, denoted as " + }, + { + "bbox": [ + 130, + 544, + 481, + 617 + ], + "type": "inline_equation", + "content": "\\{I^{o,f}\\}_{o = 1,f = 1}^{O,F}" + }, + { + "bbox": [ + 130, + 544, + 481, + 617 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 544, + 481, + 617 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 130, + 544, + 481, + 617 + ], + "type": "text", + "content": " is the number of video frames and " + }, + { + "bbox": [ + 130, + 544, + 481, + 617 + ], + "type": "inline_equation", + "content": "I^{o,f}" + }, + { + "bbox": [ + 130, + 544, + 481, + 617 + ], + "type": "text", + "content": " denotes a 2D image at video " + }, + { + "bbox": [ + 130, + 544, + 481, + 617 + ], + "type": "inline_equation", + "content": "o" + }, + { + "bbox": [ + 130, + 544, + 481, + 617 + ], + "type": "text", + "content": " and frame " + }, + { + "bbox": [ + 130, + 544, + 481, + 617 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 130, + 544, + 481, + 617 + ], + "type": "text", + "content": ". Our goal is to reconstruct the appearance, geometry, and physical dynamics parameters of Spring-Gaus." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 618, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 618, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 618, + 481, + 666 + ], + "type": "text", + "content": "In the following, we first introduce the representation for the static properties, i.e., appearance and geometry, and then the dynamics properties, i.e., the 3D Spring-Mass model. Then we describe our reconstruction pipeline. We show an overview of our reconstruction pipeline in Fig. 2." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "type": "text", + "content": "L. Zhong et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 376, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 376, + 129 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 376, + 129 + ], + "type": "text", + "content": "3.1 Appearance and Geometry Representation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 134, + 479, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 134, + 479, + 171 + ], + "spans": [ + { + "bbox": [ + 130, + 134, + 479, + 171 + ], + "type": "text", + "content": "Following 3D Gaussians Splatting [14], we use a set of 3D Gaussian kernels to represent the appearance and shape of an object. These kernels are described by a full 3D covariance matrix " + }, + { + "bbox": [ + 130, + 134, + 479, + 171 + ], + "type": "inline_equation", + "content": "\\Sigma" + }, + { + "bbox": [ + 130, + 134, + 479, + 171 + ], + "type": "text", + "content": ", defined in world space and centered at point " + }, + { + "bbox": [ + 130, + 134, + 479, + 171 + ], + "type": "inline_equation", + "content": "\\pmb{\\mu}" + }, + { + "bbox": [ + 130, + 134, + 479, + 171 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 257, + 178, + 481, + 194 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 257, + 178, + 481, + 194 + ], + "spans": [ + { + "bbox": [ + 257, + 178, + 481, + 194 + ], + "type": "interline_equation", + "content": "G (\\boldsymbol {\\tau}) = e ^ {- \\frac {1}{2} (\\boldsymbol {\\tau}) ^ {T} \\Sigma^ {- 1} (\\boldsymbol {\\tau})}, \\tag {1}", + "image_path": "bc2823da920e5769b87256c0e2bdd5b46efbcf498da419090e0aa44799f0bd97.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 201, + 479, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 201, + 479, + 236 + ], + "spans": [ + { + "bbox": [ + 130, + 201, + 479, + 236 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 201, + 479, + 236 + ], + "type": "inline_equation", + "content": "\\pmb{\\mu}" + }, + { + "bbox": [ + 130, + 201, + 479, + 236 + ], + "type": "text", + "content": " is the mean of the Gaussian distribution, and " + }, + { + "bbox": [ + 130, + 201, + 479, + 236 + ], + "type": "inline_equation", + "content": "\\pmb{\\tau}" + }, + { + "bbox": [ + 130, + 201, + 479, + 236 + ], + "type": "text", + "content": " is the independent variable of the Gaussian distribution. In our formulation, we assume all Gaussian kernels are isotropic, the covariance matrix " + }, + { + "bbox": [ + 130, + 201, + 479, + 236 + ], + "type": "inline_equation", + "content": "\\Sigma" + }, + { + "bbox": [ + 130, + 201, + 479, + 236 + ], + "type": "text", + "content": " can be controlled by a scalar " + }, + { + "bbox": [ + 130, + 201, + 479, + 236 + ], + "type": "inline_equation", + "content": "s\\in \\mathbb{R}" + }, + { + "bbox": [ + 130, + 201, + 479, + 236 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 237, + 479, + 282 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 237, + 479, + 282 + ], + "spans": [ + { + "bbox": [ + 130, + 237, + 479, + 282 + ], + "type": "text", + "content": "The splatting process, designed to render Gaussian kernels into a 2D image, involves two main steps: (1) projecting the Gaussian kernels into camera space, and (2) rendering the projected kernels into image space. The projection process is defined as:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 266, + 284, + 480, + 298 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 266, + 284, + 480, + 298 + ], + "spans": [ + { + "bbox": [ + 266, + 284, + 480, + 298 + ], + "type": "interline_equation", + "content": "\\Sigma^ {\\prime} = J W \\Sigma W ^ {T} J ^ {T}, \\tag {2}", + "image_path": "7d5069fa790b8782562ff6358f793ca76bb5dd1db7df66515f55d06b4d19fb76.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 301, + 479, + 338 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 301, + 479, + 338 + ], + "spans": [ + { + "bbox": [ + 130, + 301, + 479, + 338 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 301, + 479, + 338 + ], + "type": "inline_equation", + "content": "\\Sigma'" + }, + { + "bbox": [ + 130, + 301, + 479, + 338 + ], + "type": "text", + "content": " represents the covariance matrix in camera coordinates, " + }, + { + "bbox": [ + 130, + 301, + 479, + 338 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 130, + 301, + 479, + 338 + ], + "type": "text", + "content": " is the transformation matrix, and " + }, + { + "bbox": [ + 130, + 301, + 479, + 338 + ], + "type": "inline_equation", + "content": "J" + }, + { + "bbox": [ + 130, + 301, + 479, + 338 + ], + "type": "text", + "content": " denotes the Jacobian of the affine approximation of the projective transformation. The color " + }, + { + "bbox": [ + 130, + 301, + 479, + 338 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 130, + 301, + 479, + 338 + ], + "type": "text", + "content": " of each pixel is rendered as:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 272, + 346, + 480, + 370 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 272, + 346, + 480, + 370 + ], + "spans": [ + { + "bbox": [ + 272, + 346, + 480, + 370 + ], + "type": "interline_equation", + "content": "C = \\sum_ {i \\in N} T _ {i} \\alpha_ {i} \\boldsymbol {c} _ {i}, \\tag {3}", + "image_path": "f7c129777ffe7773bc3e7608fcfae50f3df04fbcd95052080e368f74a9355bcc.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 376, + 479, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 376, + 479, + 437 + ], + "spans": [ + { + "bbox": [ + 130, + 376, + 479, + 437 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 376, + 479, + 437 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 376, + 479, + 437 + ], + "type": "text", + "content": " is the total number of kernels and " + }, + { + "bbox": [ + 130, + 376, + 479, + 437 + ], + "type": "inline_equation", + "content": "T_{i}" + }, + { + "bbox": [ + 130, + 376, + 479, + 437 + ], + "type": "text", + "content": " represents the transmittance, defined as " + }, + { + "bbox": [ + 130, + 376, + 479, + 437 + ], + "type": "inline_equation", + "content": "\\Pi_{j = 1}^{i - 1}\\alpha_{j}" + }, + { + "bbox": [ + 130, + 376, + 479, + 437 + ], + "type": "text", + "content": ". The term " + }, + { + "bbox": [ + 130, + 376, + 479, + 437 + ], + "type": "inline_equation", + "content": "\\alpha_{i}" + }, + { + "bbox": [ + 130, + 376, + 479, + 437 + ], + "type": "text", + "content": " denotes the alpha value for each Gaussian, which is calculated using the expression " + }, + { + "bbox": [ + 130, + 376, + 479, + 437 + ], + "type": "inline_equation", + "content": "1 - e^{-\\sigma_i\\delta_i}" + }, + { + "bbox": [ + 130, + 376, + 479, + 437 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 376, + 479, + 437 + ], + "type": "inline_equation", + "content": "\\sigma_{i}" + }, + { + "bbox": [ + 130, + 376, + 479, + 437 + ], + "type": "text", + "content": " is the opacity factor. Additionally, " + }, + { + "bbox": [ + 130, + 376, + 479, + 437 + ], + "type": "inline_equation", + "content": "\\pmb{c}_i" + }, + { + "bbox": [ + 130, + 376, + 479, + 437 + ], + "type": "text", + "content": " refers to the color of the Gaussian along the ray within the interval " + }, + { + "bbox": [ + 130, + 376, + 479, + 437 + ], + "type": "inline_equation", + "content": "\\delta_{i}" + }, + { + "bbox": [ + 130, + 376, + 479, + 437 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 438, + 479, + 463 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 438, + 479, + 463 + ], + "spans": [ + { + "bbox": [ + 130, + 438, + 479, + 463 + ], + "type": "text", + "content": "For each kernel, the learnable parameters include a point center " + }, + { + "bbox": [ + 130, + 438, + 479, + 463 + ], + "type": "inline_equation", + "content": "\\pmb{\\mu}" + }, + { + "bbox": [ + 130, + 438, + 479, + 463 + ], + "type": "text", + "content": ", a scaling scalar " + }, + { + "bbox": [ + 130, + 438, + 479, + 463 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 130, + 438, + 479, + 463 + ], + "type": "text", + "content": ", a color vector " + }, + { + "bbox": [ + 130, + 438, + 479, + 463 + ], + "type": "inline_equation", + "content": "\\pmb{c}" + }, + { + "bbox": [ + 130, + 438, + 479, + 463 + ], + "type": "text", + "content": ", and an opacity value " + }, + { + "bbox": [ + 130, + 438, + 479, + 463 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 130, + 438, + 479, + 463 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 131, + 479, + 430, + 491 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 479, + 430, + 491 + ], + "spans": [ + { + "bbox": [ + 131, + 479, + 430, + 491 + ], + "type": "text", + "content": "3.2 Physics-Based Dynamics with 3D Spring-Mass Model" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 498, + 479, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 498, + 479, + 581 + ], + "spans": [ + { + "bbox": [ + 130, + 498, + 479, + 581 + ], + "type": "text", + "content": "We aim to develop and integrate an expressive yet efficient physics-based dynamics model. With these two design goals in mind, we introduce the 3D Spring-Mass model which represents the elastic object dynamics by a learnable spring-mass system. Our Spring-Gaus does not assume any material type, and it is expressive to model real heterogeneous elastic objects. Besides being expressive, our dynamics model should also be efficient and amenable to gradient-based optimization for parameter estimation." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 582, + 479, + 642 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 479, + 642 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 479, + 642 + ], + "type": "text", + "content": "Recall that the appearance and geometry representations include a set of Gaussian kernels " + }, + { + "bbox": [ + 130, + 582, + 479, + 642 + ], + "type": "inline_equation", + "content": "\\{G_i\\}_{i=1}^N" + }, + { + "bbox": [ + 130, + 582, + 479, + 642 + ], + "type": "text", + "content": ", parameterized by point centers " + }, + { + "bbox": [ + 130, + 582, + 479, + 642 + ], + "type": "inline_equation", + "content": "\\mathbf{X} = \\{\\pmb{\\mu}_i\\}_{i=1}^N" + }, + { + "bbox": [ + 130, + 582, + 479, + 642 + ], + "type": "text", + "content": ", scaling scalars " + }, + { + "bbox": [ + 130, + 582, + 479, + 642 + ], + "type": "inline_equation", + "content": "\\{s_i\\}_{i=1}^N" + }, + { + "bbox": [ + 130, + 582, + 479, + 642 + ], + "type": "text", + "content": ", color vectors " + }, + { + "bbox": [ + 130, + 582, + 479, + 642 + ], + "type": "inline_equation", + "content": "\\{\\pmb{c}_i\\}_{i=1}^N" + }, + { + "bbox": [ + 130, + 582, + 479, + 642 + ], + "type": "text", + "content": ", and opacity values " + }, + { + "bbox": [ + 130, + 582, + 479, + 642 + ], + "type": "inline_equation", + "content": "\\{\\sigma_i\\}_{i=1}^N" + }, + { + "bbox": [ + 130, + 582, + 479, + 642 + ], + "type": "text", + "content": ". To manage complexity for efficient simulation, we introduce volume sampling to generate a set of anchor points " + }, + { + "bbox": [ + 130, + 582, + 479, + 642 + ], + "type": "inline_equation", + "content": "\\mathbf{A} = \\{\\pmb{x}_i\\}_{i=1}^{N_A}" + }, + { + "bbox": [ + 130, + 582, + 479, + 642 + ], + "type": "text", + "content": " (each " + }, + { + "bbox": [ + 130, + 582, + 479, + 642 + ], + "type": "inline_equation", + "content": "\\pmb{x}" + }, + { + "bbox": [ + 130, + 582, + 479, + 642 + ], + "type": "text", + "content": " represents a spatial point), defined by:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 259, + 652, + 480, + 666 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 259, + 652, + 480, + 666 + ], + "spans": [ + { + "bbox": [ + 259, + 652, + 480, + 666 + ], + "type": "interline_equation", + "content": "\\boldsymbol {A} = \\left\\{\\boldsymbol {x} _ {i} \\right\\} _ {i = 1} ^ {N _ {A}} = \\mathcal {V} (\\boldsymbol {X}), \\tag {4}", + "image_path": "0e5a2bc5adb467136e45438506945425f2e42fcdfe66359bef4feb806882b76b.jpg" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 395, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 395, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 395, + 91, + 447, + 102 + ], + "type": "text", + "content": "Spring-Gaus" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 137, + 116, + 478, + 323 + ], + "blocks": [ + { + "bbox": [ + 137, + 116, + 478, + 323 + ], + "lines": [ + { + "bbox": [ + 137, + 116, + 478, + 323 + ], + "spans": [ + { + "bbox": [ + 137, + 116, + 478, + 323 + ], + "type": "image", + "image_path": "dc0f7fb28cd35083fb30769d417e6b17eb077004dc422f078951c8e33f244093.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 346, + 482, + 434 + ], + "lines": [ + { + "bbox": [ + 130, + 346, + 482, + 434 + ], + "spans": [ + { + "bbox": [ + 130, + 346, + 482, + 434 + ], + "type": "text", + "content": "Fig.2: Overview of Spring-Gaus reconstruction pipeline: (a) Static Scene Reconstruction: We start by reconstructing static 3D Gaussians from the first frames of the multiview videos. (b) Refining 3D Gaussians: We extract a set of anchor points to allow efficient simulation, which leads to appearance drift. We refine the 3D Gaussians to better model the appearance during simulation. (c) Dynamic Reconstruction: Our 3D Spring-Mass model simulates anchor points and updates the positions of Gaussian kernels. Upon completion of optimization, we obtain a simulatable 3D object that accurately models its dynamics." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 457, + 480, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 457, + 480, + 469 + ], + "spans": [ + { + "bbox": [ + 130, + 457, + 480, + 469 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 457, + 480, + 469 + ], + "type": "inline_equation", + "content": "N_{A}" + }, + { + "bbox": [ + 130, + 457, + 480, + 469 + ], + "type": "text", + "content": " denotes the number of anchors, and " + }, + { + "bbox": [ + 130, + 457, + 480, + 469 + ], + "type": "inline_equation", + "content": "\\nu" + }, + { + "bbox": [ + 130, + 457, + 480, + 469 + ], + "type": "text", + "content": " denotes the sampling function." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 470, + 481, + 507 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 470, + 481, + 507 + ], + "spans": [ + { + "bbox": [ + 130, + 470, + 481, + 507 + ], + "type": "text", + "content": "Our 3D Spring-Mass physical model can simulate the motion of anchor points " + }, + { + "bbox": [ + 130, + 470, + 481, + 507 + ], + "type": "inline_equation", + "content": "\\mathbf{A}" + }, + { + "bbox": [ + 130, + 470, + 481, + 507 + ], + "type": "text", + "content": ", assuming each anchor has a mass " + }, + { + "bbox": [ + 130, + 470, + 481, + 507 + ], + "type": "inline_equation", + "content": "m_{i}" + }, + { + "bbox": [ + 130, + 470, + 481, + 507 + ], + "type": "text", + "content": " and an initial velocity " + }, + { + "bbox": [ + 130, + 470, + 481, + 507 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_{i}" + }, + { + "bbox": [ + 130, + 470, + 481, + 507 + ], + "type": "text", + "content": ". Each anchor " + }, + { + "bbox": [ + 130, + 470, + 481, + 507 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{i}" + }, + { + "bbox": [ + 130, + 470, + 481, + 507 + ], + "type": "text", + "content": " connects to its " + }, + { + "bbox": [ + 130, + 470, + 481, + 507 + ], + "type": "inline_equation", + "content": "n_{k}" + }, + { + "bbox": [ + 130, + 470, + 481, + 507 + ], + "type": "text", + "content": " nearest neighbors " + }, + { + "bbox": [ + 130, + 470, + 481, + 507 + ], + "type": "inline_equation", + "content": "\\mathbf{N}_{i} = \\{\\mathbf{x}_{i,j}\\}_{j=1}^{n_{k}}" + }, + { + "bbox": [ + 130, + 470, + 481, + 507 + ], + "type": "text", + "content": " through springs " + }, + { + "bbox": [ + 130, + 470, + 481, + 507 + ], + "type": "inline_equation", + "content": "\\mathbf{L}" + }, + { + "bbox": [ + 130, + 470, + 481, + 507 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 229, + 515, + 480, + 532 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 229, + 515, + 480, + 532 + ], + "spans": [ + { + "bbox": [ + 229, + 515, + 480, + 532 + ], + "type": "interline_equation", + "content": "\\boldsymbol {L} = \\left\\{l _ {i, j} \\right\\} _ {i = 1, j = 1} ^ {N _ {A}, n _ {k}} = \\operatorname {k n n} (\\boldsymbol {A}, \\boldsymbol {A}, n _ {k}), \\tag {5}", + "image_path": "bc697241cb0686d1ae015fd057fef33b153d6c7f53726cfbda1759d6a1066014.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 539, + 481, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 539, + 481, + 574 + ], + "spans": [ + { + "bbox": [ + 130, + 539, + 481, + 574 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 539, + 481, + 574 + ], + "type": "inline_equation", + "content": "l_{i,j}" + }, + { + "bbox": [ + 130, + 539, + 481, + 574 + ], + "type": "text", + "content": " denotes the spring's length between " + }, + { + "bbox": [ + 130, + 539, + 481, + 574 + ], + "type": "inline_equation", + "content": "\\pmb{x}_i" + }, + { + "bbox": [ + 130, + 539, + 481, + 574 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 539, + 481, + 574 + ], + "type": "inline_equation", + "content": "\\pmb{x}_{i,j}" + }, + { + "bbox": [ + 130, + 539, + 481, + 574 + ], + "type": "text", + "content": ", and knn denotes the k-nearest neighbors function. Each spring is characterized by a stiffness " + }, + { + "bbox": [ + 130, + 539, + 481, + 574 + ], + "type": "inline_equation", + "content": "k_{i,j}" + }, + { + "bbox": [ + 130, + 539, + 481, + 574 + ], + "type": "text", + "content": " and a damping factor " + }, + { + "bbox": [ + 130, + 539, + 481, + 574 + ], + "type": "inline_equation", + "content": "\\zeta_{i,j}" + }, + { + "bbox": [ + 130, + 539, + 481, + 574 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 575, + 481, + 599 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 575, + 481, + 599 + ], + "spans": [ + { + "bbox": [ + 130, + 575, + 481, + 599 + ], + "type": "text", + "content": "To update the positions of the kernels, we first measure the distance between each kernel center and its " + }, + { + "bbox": [ + 130, + 575, + 481, + 599 + ], + "type": "inline_equation", + "content": "n_b" + }, + { + "bbox": [ + 130, + 575, + 481, + 599 + ], + "type": "text", + "content": " nearest anchors at the dynamic simulation's onset:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 239, + 609, + 480, + 625 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 239, + 609, + 480, + 625 + ], + "spans": [ + { + "bbox": [ + 239, + 609, + 480, + 625 + ], + "type": "interline_equation", + "content": "\\left\\{d _ {i, j} \\right\\} _ {i = 1, j = 1} ^ {N, n _ {b}} = \\operatorname {k n n} (\\boldsymbol {X}, \\boldsymbol {A}, n _ {b}). \\tag {6}", + "image_path": "340b8936ed0b3572010a5c1462a2a7684570f02ea530f9670e8ee0dcfbb0d058.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 629, + 481, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 481, + 651 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 481, + 651 + ], + "type": "text", + "content": "For each timestep " + }, + { + "bbox": [ + 130, + 629, + 481, + 651 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 629, + 481, + 651 + ], + "type": "text", + "content": ", the forces " + }, + { + "bbox": [ + 130, + 629, + 481, + 651 + ], + "type": "inline_equation", + "content": "\\pmb{F}_i^t" + }, + { + "bbox": [ + 130, + 629, + 481, + 651 + ], + "type": "text", + "content": " acting on each anchor point " + }, + { + "bbox": [ + 130, + 629, + 481, + 651 + ], + "type": "inline_equation", + "content": "\\pmb{x}_i^t" + }, + { + "bbox": [ + 130, + 629, + 481, + 651 + ], + "type": "text", + "content": " are calculated as follows:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 254, + 652, + 480, + 667 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 254, + 652, + 480, + 667 + ], + "spans": [ + { + "bbox": [ + 254, + 652, + 480, + 667 + ], + "type": "interline_equation", + "content": "\\boldsymbol {F} _ {i} ^ {t} = \\boldsymbol {F} _ {\\boldsymbol {k} i} ^ {t} + \\boldsymbol {F} _ {\\boldsymbol {\\zeta} i} ^ {t} + m _ {i} \\boldsymbol {g}, \\tag {7}", + "image_path": "d1f057e022bf11752076302669113503bbfa9d0585d72fe5ebc76366aec73db1.jpg" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 228, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 228, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 228, + 102 + ], + "type": "text", + "content": "L. Zhong et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 115, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 115, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 130, + 115, + 479, + 140 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 115, + 479, + 140 + ], + "type": "inline_equation", + "content": "\\pmb{F}_{\\pmb{k}_i}^t, \\pmb{F}_{\\pmb{\\zeta}_i}^t" + }, + { + "bbox": [ + 130, + 115, + 479, + 140 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 115, + 479, + 140 + ], + "type": "inline_equation", + "content": "\\pmb{g}" + }, + { + "bbox": [ + 130, + 115, + 479, + 140 + ], + "type": "text", + "content": " represent the spring force, the damping force, and gravitational acceleration, respectively." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 140, + 480, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 140, + 480, + 163 + ], + "spans": [ + { + "bbox": [ + 130, + 140, + 480, + 163 + ], + "type": "text", + "content": "Then, for each spring " + }, + { + "bbox": [ + 130, + 140, + 480, + 163 + ], + "type": "inline_equation", + "content": "\\pmb{L}_{i,j}" + }, + { + "bbox": [ + 130, + 140, + 480, + 163 + ], + "type": "text", + "content": ", the spring force " + }, + { + "bbox": [ + 130, + 140, + 480, + 163 + ], + "type": "inline_equation", + "content": "\\pmb{F}_{\\pmb{k}_{i,j}}^{t}" + }, + { + "bbox": [ + 130, + 140, + 480, + 163 + ], + "type": "text", + "content": " and damping force " + }, + { + "bbox": [ + 130, + 140, + 480, + 163 + ], + "type": "inline_equation", + "content": "\\pmb{F}_{\\pmb{\\zeta}_{i,j}}^{t}" + }, + { + "bbox": [ + 130, + 140, + 480, + 163 + ], + "type": "text", + "content": " are determined by:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 144, + 173, + 479, + 201 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 144, + 173, + 479, + 201 + ], + "spans": [ + { + "bbox": [ + 144, + 173, + 479, + 201 + ], + "type": "interline_equation", + "content": "\\boldsymbol {F} _ {\\boldsymbol {k} _ {i, j}} ^ {t} = - \\eta_ {j} \\cdot k _ {i, j} \\left(\\left\\| \\boldsymbol {x} _ {i} ^ {t} - \\boldsymbol {x} _ {i, j} ^ {t} \\right\\| - l _ {i, j}\\right) \\frac {\\boldsymbol {x} _ {i} ^ {t} - \\boldsymbol {x} _ {i , j} ^ {t}}{\\left\\| \\boldsymbol {x} _ {i} ^ {t} - \\boldsymbol {x} _ {i , j} ^ {t} \\right\\|} \\cdot \\left| \\left\\| \\boldsymbol {x} _ {i} ^ {t} - \\boldsymbol {x} _ {i, j} ^ {t} \\right\\| - l _ {i, j} \\right| ^ {p _ {k}}, \\tag {8}", + "image_path": "b28fc85d37b44b38631a45d652af87861164da64d89146d50d9009056bd6fd2e.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 189, + 204, + 479, + 235 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 189, + 204, + 479, + 235 + ], + "spans": [ + { + "bbox": [ + 189, + 204, + 479, + 235 + ], + "type": "interline_equation", + "content": "\\boldsymbol {F} _ {\\zeta_ {i, j}} ^ {t} = \\left(- \\zeta_ {i, j} \\left(\\boldsymbol {v} _ {i} ^ {t} - \\boldsymbol {v} _ {i, j} ^ {t}\\right) \\frac {\\boldsymbol {x} _ {i} ^ {t} - \\boldsymbol {x} _ {i , j} ^ {t}}{\\left\\| \\boldsymbol {x} _ {i} ^ {t} - \\boldsymbol {x} _ {i , j} ^ {t} \\right\\|}\\right) \\cdot \\frac {\\boldsymbol {x} _ {i} ^ {t} - \\boldsymbol {x} _ {i , j} ^ {t}}{\\left\\| \\boldsymbol {x} _ {i} ^ {t} - \\boldsymbol {x} _ {i , j} ^ {t} \\right\\|}, \\tag {9}", + "image_path": "b9f7aa9b2ee55614de0eaf1e8a7a1d0b404421068011643aa30de6bee70ed51a.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 243, + 479, + 302 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 243, + 479, + 302 + ], + "spans": [ + { + "bbox": [ + 130, + 243, + 479, + 302 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 243, + 479, + 302 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 130, + 243, + 479, + 302 + ], + "type": "text", + "content": " is a soft vector which will be discussed later and " + }, + { + "bbox": [ + 130, + 243, + 479, + 302 + ], + "type": "inline_equation", + "content": "p_k" + }, + { + "bbox": [ + 130, + 243, + 479, + 302 + ], + "type": "text", + "content": " is a hyperparameter that determines the nonlinearity of the spring force. When " + }, + { + "bbox": [ + 130, + 243, + 479, + 302 + ], + "type": "inline_equation", + "content": "p_k" + }, + { + "bbox": [ + 130, + 243, + 479, + 302 + ], + "type": "text", + "content": " is set to 0, Eq. (8) becomes Hooke's law, and for positive values of " + }, + { + "bbox": [ + 130, + 243, + 479, + 302 + ], + "type": "inline_equation", + "content": "p_k" + }, + { + "bbox": [ + 130, + 243, + 479, + 302 + ], + "type": "text", + "content": ", the spring force becomes a nonlinear function of the spring's length. The cumulative forces acting on each anchor point " + }, + { + "bbox": [ + 130, + 243, + 479, + 302 + ], + "type": "inline_equation", + "content": "\\boldsymbol{x}_i^t" + }, + { + "bbox": [ + 130, + 243, + 479, + 302 + ], + "type": "text", + "content": " are expressed as:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 235, + 312, + 479, + 343 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 235, + 312, + 479, + 343 + ], + "spans": [ + { + "bbox": [ + 235, + 312, + 479, + 343 + ], + "type": "interline_equation", + "content": "\\boldsymbol {F} _ {i} ^ {t} = \\sum_ {j = 1} ^ {n _ {k}} \\boldsymbol {F} _ {\\boldsymbol {k} _ {i, j}} ^ {t} + \\sum_ {j = 1} ^ {n _ {k}} \\boldsymbol {F} _ {\\zeta_ {i, j}} ^ {t} + m _ {i} \\boldsymbol {g}. \\tag {10}", + "image_path": "71bc3a6737282ec2eebe765843174ee59f59fa771a4ebb8c17b5dbbd23b7fd5a.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 351, + 479, + 374 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 351, + 479, + 374 + ], + "spans": [ + { + "bbox": [ + 131, + 351, + 479, + 374 + ], + "type": "text", + "content": "Anchor points " + }, + { + "bbox": [ + 131, + 351, + 479, + 374 + ], + "type": "inline_equation", + "content": "\\mathbf{A}" + }, + { + "bbox": [ + 131, + 351, + 479, + 374 + ], + "type": "text", + "content": "'s positions and velocities are updated using semi-implicit Euler integration:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 264, + 384, + 479, + 408 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 264, + 384, + 479, + 408 + ], + "spans": [ + { + "bbox": [ + 264, + 384, + 479, + 408 + ], + "type": "interline_equation", + "content": "\\hat {\\boldsymbol {v}} _ {i} ^ {t + 1} = \\boldsymbol {v} _ {i} ^ {t} + \\frac {\\boldsymbol {F} _ {i} ^ {t}}{m _ {i}} \\Delta t, \\tag {11}", + "image_path": "074dada16048b9d2345b1f6bb31fb9dc62d6ca67a635ffd7cf55610578ee5afe.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 261, + 410, + 479, + 424 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 261, + 410, + 479, + 424 + ], + "spans": [ + { + "bbox": [ + 261, + 410, + 479, + 424 + ], + "type": "interline_equation", + "content": "\\hat {\\boldsymbol {x}} _ {i} ^ {t + 1} = \\boldsymbol {x} _ {i} ^ {t} + \\boldsymbol {v} _ {i} ^ {t + 1} \\Delta t, \\tag {12}", + "image_path": "5bf368b08f4966f9fe999f70892d1793d50cc85b2edc15ed4f9a925ac59ee626.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 131, + 433, + 479, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 433, + 479, + 456 + ], + "spans": [ + { + "bbox": [ + 131, + 433, + 479, + 456 + ], + "type": "text", + "content": "and a boundary condition " + }, + { + "bbox": [ + 131, + 433, + 479, + 456 + ], + "type": "inline_equation", + "content": "\\mathcal{B}" + }, + { + "bbox": [ + 131, + 433, + 479, + 456 + ], + "type": "text", + "content": " is applied to the anchor points " + }, + { + "bbox": [ + 131, + 433, + 479, + 456 + ], + "type": "inline_equation", + "content": "\\pmb{A}" + }, + { + "bbox": [ + 131, + 433, + 479, + 456 + ], + "type": "text", + "content": " to model the interactions with the environment:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 247, + 465, + 479, + 479 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 247, + 465, + 479, + 479 + ], + "spans": [ + { + "bbox": [ + 247, + 465, + 479, + 479 + ], + "type": "interline_equation", + "content": "\\boldsymbol {x} _ {i} ^ {t + 1}, \\boldsymbol {v} _ {i} ^ {t + 1} = \\mathcal {B} \\left(\\hat {\\boldsymbol {x}} _ {i} ^ {t + 1}, \\hat {\\boldsymbol {v}} _ {i} ^ {t + 1}\\right). \\tag {13}", + "image_path": "45044e3c04b08ef073e997de1f2d15ac5c4f36ab615357361962d8fc02452a10.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 131, + 487, + 479, + 512 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 487, + 479, + 512 + ], + "spans": [ + { + "bbox": [ + 131, + 487, + 479, + 512 + ], + "type": "text", + "content": "The position of each Gaussian kernel " + }, + { + "bbox": [ + 131, + 487, + 479, + 512 + ], + "type": "inline_equation", + "content": "\\pmb{\\mu}_i" + }, + { + "bbox": [ + 131, + 487, + 479, + 512 + ], + "type": "text", + "content": " is updated through Inverse Distance Weighting (IDW) interpolation to reflect the dynamic changes accurately:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 236, + 521, + 479, + 550 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 236, + 521, + 479, + 550 + ], + "spans": [ + { + "bbox": [ + 236, + 521, + 479, + 550 + ], + "type": "interline_equation", + "content": "\\boldsymbol {\\mu} _ {i} ^ {t + 1} = \\frac {\\sum_ {j = 1} ^ {n _ {b}} \\boldsymbol {x} _ {i , j} ^ {t + 1} \\cdot \\left(1 / \\left(d _ {i , j}\\right) ^ {p _ {b}}\\right)}{\\sum_ {j = 1} ^ {n _ {b}} \\left(1 / \\left(d _ {i , j}\\right) ^ {p _ {b}}\\right)}, \\tag {14}", + "image_path": "552fed2162b4f2d56de5867bd673b62a3d03fc45588bd80ef566581c1a40c5a9.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 558, + 479, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 558, + 479, + 605 + ], + "spans": [ + { + "bbox": [ + 130, + 558, + 479, + 605 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 558, + 479, + 605 + ], + "type": "inline_equation", + "content": "p_b" + }, + { + "bbox": [ + 130, + 558, + 479, + 605 + ], + "type": "text", + "content": " is a positive real number that determines the diminishing influence of anchor points with distance. To render the image " + }, + { + "bbox": [ + 130, + 558, + 479, + 605 + ], + "type": "inline_equation", + "content": "\\hat{I}" + }, + { + "bbox": [ + 130, + 558, + 479, + 605 + ], + "type": "text", + "content": " at a specific camera and frame, we use the updated positions of the Gaussian kernels following the rendering equation Eq. (3)." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 606, + 480, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 606, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 606, + 480, + 665 + ], + "type": "text", + "content": "Soft Vector for Springs Connection: In our formulation above, the " + }, + { + "bbox": [ + 130, + 606, + 480, + 665 + ], + "type": "inline_equation", + "content": "n_k" + }, + { + "bbox": [ + 130, + 606, + 480, + 665 + ], + "type": "text", + "content": " is a hyperparameter which is the number of connected springs for each anchor. However, the choice of " + }, + { + "bbox": [ + 130, + 606, + 480, + 665 + ], + "type": "inline_equation", + "content": "n_k" + }, + { + "bbox": [ + 130, + 606, + 480, + 665 + ], + "type": "text", + "content": " will directly affect the simulation results markedly. The bigger value of " + }, + { + "bbox": [ + 130, + 606, + 480, + 665 + ], + "type": "inline_equation", + "content": "n_k" + }, + { + "bbox": [ + 130, + 606, + 480, + 665 + ], + "type": "text", + "content": ", the object will behave more rigidly, and a smaller value of " + }, + { + "bbox": [ + 130, + 606, + 480, + 665 + ], + "type": "inline_equation", + "content": "n_k" + }, + { + "bbox": [ + 130, + 606, + 480, + 665 + ], + "type": "text", + "content": " leads to a noticeably softer behavior of the point cloud. To address the" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 395, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 395, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 395, + 91, + 447, + 102 + ], + "type": "text", + "content": "Spring-Gaus" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "type": "text", + "content": "significant impact that the value of " + }, + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "type": "inline_equation", + "content": "n_k" + }, + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "type": "text", + "content": " has on the simulation, we introduce a mitigation strategy by applying a soft vector " + }, + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "type": "inline_equation", + "content": "\\eta = [\\eta_0, \\eta_1, \\dots, \\eta_{n_k}]" + }, + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "type": "text", + "content": ". This vector controlled by a learnable parameter " + }, + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "type": "inline_equation", + "content": "\\kappa" + }, + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "type": "text", + "content": " (shared by all anchors) is used to modulate the number of connected springs, thereby adjusting the system's response to a different object. Given an empirical value " + }, + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "type": "inline_equation", + "content": "n_c" + }, + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "type": "text", + "content": ", the soft vector " + }, + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "type": "text", + "content": " is calculated as:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 178, + 184, + 481, + 217 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 178, + 184, + 481, + 217 + ], + "spans": [ + { + "bbox": [ + 178, + 184, + 481, + 217 + ], + "type": "interline_equation", + "content": "\\eta_ {j} = \\left\\{ \\begin{array}{l l} 1 & j \\leq n _ {c}, \\\\ \\operatorname {c l a m p} (2 - \\exp (\\operatorname {s o f t p l u s} (\\kappa)) ^ {j - n _ {c}}, 0, 1) & n _ {c} < j \\leq n _ {k}. \\end{array} \\right. \\tag {15}", + "image_path": "a02697575bb0b03957baa607373ae6372d460aed19946df0efa638f06d3b6f4c.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 230, + 228, + 243 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 230, + 228, + 243 + ], + "spans": [ + { + "bbox": [ + 132, + 230, + 228, + 243 + ], + "type": "text", + "content": "3.3 Optimization" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 249, + 482, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 249, + 482, + 357 + ], + "spans": [ + { + "bbox": [ + 130, + 249, + 482, + 357 + ], + "type": "text", + "content": "To allow efficient optimization, we simplify our model by reducing the number of learnable parameters without changing the essential expressiveness. In our simplified approach, we standardize the mass of every anchor to a constant value " + }, + { + "bbox": [ + 130, + 249, + 482, + 357 + ], + "type": "inline_equation", + "content": "m_0" + }, + { + "bbox": [ + 130, + 249, + 482, + 357 + ], + "type": "text", + "content": ", and control all damping factors using a singular parameter " + }, + { + "bbox": [ + 130, + 249, + 482, + 357 + ], + "type": "inline_equation", + "content": "\\zeta_0" + }, + { + "bbox": [ + 130, + 249, + 482, + 357 + ], + "type": "text", + "content": ". These two parameters are fixed, eliminating their variability from the optimization process. Furthermore, we introduce a unified parameter " + }, + { + "bbox": [ + 130, + 249, + 482, + 357 + ], + "type": "inline_equation", + "content": "k_{i}" + }, + { + "bbox": [ + 130, + 249, + 482, + 357 + ], + "type": "text", + "content": " for each anchor " + }, + { + "bbox": [ + 130, + 249, + 482, + 357 + ], + "type": "inline_equation", + "content": "\\pmb{x}_i" + }, + { + "bbox": [ + 130, + 249, + 482, + 357 + ], + "type": "text", + "content": " to control the spring stiffness of the springs connected to it, simplifying the model without compromising its functional integrity. The spring stiffness " + }, + { + "bbox": [ + 130, + 249, + 482, + 357 + ], + "type": "inline_equation", + "content": "k_{i,j}" + }, + { + "bbox": [ + 130, + 249, + 482, + 357 + ], + "type": "text", + "content": " and damping factor " + }, + { + "bbox": [ + 130, + 249, + 482, + 357 + ], + "type": "inline_equation", + "content": "\\zeta_{i,j}" + }, + { + "bbox": [ + 130, + 249, + 482, + 357 + ], + "type": "text", + "content": " are thus given as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 279, + 366, + 480, + 378 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 279, + 366, + 480, + 378 + ], + "spans": [ + { + "bbox": [ + 279, + 366, + 480, + 378 + ], + "type": "interline_equation", + "content": "k _ {i, j} = k _ {i} / l _ {i, j}, \\tag {16}", + "image_path": "dfe2521a40c8a86fd5d66d1560e805d012629ddd5f8ec0e3f17ee9e9db80c8e2.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 279, + 382, + 480, + 394 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 279, + 382, + 480, + 394 + ], + "spans": [ + { + "bbox": [ + 279, + 382, + 480, + 394 + ], + "type": "interline_equation", + "content": "\\zeta_ {i, j} = \\zeta_ {0} / l _ {i, j}. \\tag {17}", + "image_path": "26b6d6fdb40d8c74fed1a62a7fe4f1c88dbb7b2ce94f0b4f07cc3397b54e6a90.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 401, + 482, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 401, + 482, + 460 + ], + "spans": [ + { + "bbox": [ + 130, + 401, + 482, + 460 + ], + "type": "text", + "content": "Note that, in Eq. (17) stiffness is defined at anchor points rather than on springs themselves to simplify the optimization—we only need to optimize " + }, + { + "bbox": [ + 130, + 401, + 482, + 460 + ], + "type": "inline_equation", + "content": "N_{A}" + }, + { + "bbox": [ + 130, + 401, + 482, + 460 + ], + "type": "text", + "content": " stiffness coefficients instead of all the springs, which is on the order of " + }, + { + "bbox": [ + 130, + 401, + 482, + 460 + ], + "type": "inline_equation", + "content": "n_k \\cdot N_A" + }, + { + "bbox": [ + 130, + 401, + 482, + 460 + ], + "type": "text", + "content": ". By this simplification, our model maintains computational efficiency and ease of optimization, while still capturing the essential dynamics of the object." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 146, + 461, + 436, + 473 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 461, + 436, + 473 + ], + "spans": [ + { + "bbox": [ + 146, + 461, + 436, + 473 + ], + "type": "text", + "content": "Summarized, the learnable parameters in our model now include:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 479, + 480, + 586 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 138, + 479, + 480, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 479, + 480, + 502 + ], + "spans": [ + { + "bbox": [ + 138, + 479, + 480, + 502 + ], + "type": "text", + "content": "- " + }, + { + "bbox": [ + 138, + 479, + 480, + 502 + ], + "type": "inline_equation", + "content": "\\pmb{v}_0" + }, + { + "bbox": [ + 138, + 479, + 480, + 502 + ], + "type": "text", + "content": ": the initial velocity vector, providing a baseline movement pattern for the simulation;" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 502, + 480, + 527 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 502, + 480, + 527 + ], + "spans": [ + { + "bbox": [ + 138, + 502, + 480, + 527 + ], + "type": "text", + "content": "- " + }, + { + "bbox": [ + 138, + 502, + 480, + 527 + ], + "type": "inline_equation", + "content": "\\{k_i\\}_{i=1}^{N_A}" + }, + { + "bbox": [ + 138, + 502, + 480, + 527 + ], + "type": "text", + "content": ": the individual stiffness parameters for each anchor, allowing for localized adjustments to spring stiffness;" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 528, + 480, + 550 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 528, + 480, + 550 + ], + "spans": [ + { + "bbox": [ + 138, + 528, + 480, + 550 + ], + "type": "text", + "content": "- " + }, + { + "bbox": [ + 138, + 528, + 480, + 550 + ], + "type": "inline_equation", + "content": "\\kappa" + }, + { + "bbox": [ + 138, + 528, + 480, + 550 + ], + "type": "text", + "content": ": a parameter governing the modulation of the soft vector, facilitating finetuned control over the spring dynamics;" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 551, + 480, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 551, + 480, + 586 + ], + "spans": [ + { + "bbox": [ + 138, + 551, + 480, + 586 + ], + "type": "text", + "content": "- " + }, + { + "bbox": [ + 138, + 551, + 480, + 586 + ], + "type": "inline_equation", + "content": "\\Theta(\\mathcal{B}(\\cdot))" + }, + { + "bbox": [ + 138, + 551, + 480, + 586 + ], + "type": "text", + "content": ": parameters defining the boundary conditions (" + }, + { + "bbox": [ + 138, + 551, + 480, + 586 + ], + "type": "inline_equation", + "content": "\\Theta" + }, + { + "bbox": [ + 138, + 551, + 480, + 586 + ], + "type": "text", + "content": " is learnable parameters), such as the friction coefficient, which influence the simulation's physical realism." + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": "There exist " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "n_t" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": " timesteps between each of the two keyframes. At each timestep, " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\boldsymbol{x}^t" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": " is computed from " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\boldsymbol{x}^{t-1}" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": " using the update rule Eqs. (11) and (12). We only optimize physical parameters at each keyframe (when " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "t = 0, n_t, 2n_t, 3n_t, \\ldots" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": ") based on the visual observations until that time. We increase the value of " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "n_t" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": " as the parameters converge. This approach balances identification accuracy and computational demand." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 228, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 228, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 228, + 102 + ], + "type": "text", + "content": "L. Zhong et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 164 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 164 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 164 + ], + "type": "text", + "content": "Same with 3D Gaussian Splatting [14], we define our loss function as a weighted combination of the " + }, + { + "bbox": [ + 130, + 116, + 482, + 164 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_1" + }, + { + "bbox": [ + 130, + 116, + 482, + 164 + ], + "type": "text", + "content": " norm and the Structural Similarity Index Measure (D-SSIM) " + }, + { + "bbox": [ + 130, + 116, + 482, + 164 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{d - ssim}}" + }, + { + "bbox": [ + 130, + 116, + 482, + 164 + ], + "type": "text", + "content": ", applied between the input images " + }, + { + "bbox": [ + 130, + 116, + 482, + 164 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 130, + 116, + 482, + 164 + ], + "type": "text", + "content": " and the rendered images " + }, + { + "bbox": [ + 130, + 116, + 482, + 164 + ], + "type": "inline_equation", + "content": "\\hat{I}" + }, + { + "bbox": [ + 130, + 116, + 482, + 164 + ], + "type": "text", + "content": ". This is formally expressed as:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 228, + 177, + 481, + 189 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 228, + 177, + 481, + 189 + ], + "spans": [ + { + "bbox": [ + 228, + 177, + 481, + 189 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\left(1 - \\lambda_ {\\mathrm {d - s s i m}}\\right) \\mathcal {L} _ {1} + \\lambda_ {\\mathrm {d - s s i m}} \\mathcal {L} _ {\\mathrm {d - s s i m}}, \\tag {18}", + "image_path": "d6778aa37890fed010d3f7cef92f032ed40324c0572b5143563db346edf06389.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 194, + 481, + 230 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 194, + 481, + 230 + ], + "spans": [ + { + "bbox": [ + 130, + 194, + 481, + 230 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 194, + 481, + 230 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{d - ssim}}" + }, + { + "bbox": [ + 130, + 194, + 481, + 230 + ], + "type": "text", + "content": " is a weighting coefficient that balances the contribution of the " + }, + { + "bbox": [ + 130, + 194, + 481, + 230 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_1" + }, + { + "bbox": [ + 130, + 194, + 481, + 230 + ], + "type": "text", + "content": " norm and the D-SSIM term to the overall loss. We decouple our optimization into a few stages, as illustrated in Fig. 2." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 230, + 481, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 230, + 481, + 266 + ], + "spans": [ + { + "bbox": [ + 130, + 230, + 481, + 266 + ], + "type": "text", + "content": "Static Reconstruction: Our optimization starts by taking the first frames of the multi-view videos and using them to reconstruct the appearance and geometry of the object." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 266, + 482, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 266, + 482, + 373 + ], + "spans": [ + { + "bbox": [ + 130, + 266, + 482, + 373 + ], + "type": "text", + "content": "3D Gaussians Refinement: During dynamic simulation of the anchor points, the 3D Gaussian centers " + }, + { + "bbox": [ + 130, + 266, + 482, + 373 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_0 = \\{\\pmb{\\mu}_i^0\\}_{i=1}^N" + }, + { + "bbox": [ + 130, + 266, + 482, + 373 + ], + "type": "text", + "content": " are computed by IDW interpolation outlined in Eq. (14). This leads to a slight appearance drift from the static reconstruction. Therefore, we refine the parameters of the Gaussian kernels—scaling scalar " + }, + { + "bbox": [ + 130, + 266, + 482, + 373 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 130, + 266, + 482, + 373 + ], + "type": "text", + "content": ", color vector " + }, + { + "bbox": [ + 130, + 266, + 482, + 373 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 130, + 266, + 482, + 373 + ], + "type": "text", + "content": ", and opacity value " + }, + { + "bbox": [ + 130, + 266, + 482, + 373 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 130, + 266, + 482, + 373 + ], + "type": "text", + "content": ", excluding points center " + }, + { + "bbox": [ + 130, + 266, + 482, + 373 + ], + "type": "inline_equation", + "content": "\\pmb{\\mu}" + }, + { + "bbox": [ + 130, + 266, + 482, + 373 + ], + "type": "text", + "content": "—also at the first frame, before the dynamic simulation. Since the Gaussian kernels' position from IDW interpolation is slightly different from the static reconstruction results, this refinement process enables the Gaussian kernels with the interpolated spatial position to render the correct object appearance." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 374, + 482, + 421 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 374, + 482, + 421 + ], + "spans": [ + { + "bbox": [ + 130, + 374, + 482, + 421 + ], + "type": "text", + "content": "Dynamic Reconstruction: Since our efficient dynamics simulation is fully differentiable with respect to the learnable parameters, we can optimize the physical parameters through differentiable simulation and differentiable rendering with our loss function " + }, + { + "bbox": [ + 130, + 374, + 482, + 421 + ], + "type": "inline_equation", + "content": "\\mathcal{L}" + }, + { + "bbox": [ + 130, + 374, + 482, + 421 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 441, + 224, + 453 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 441, + 224, + 453 + ], + "spans": [ + { + "bbox": [ + 132, + 441, + 224, + 453 + ], + "type": "text", + "content": "4 Experiment" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 465, + 205, + 476 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 465, + 205, + 476 + ], + "spans": [ + { + "bbox": [ + 132, + 465, + 205, + 476 + ], + "type": "text", + "content": "4.1 Datasets" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 485, + 482, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 485, + 482, + 605 + ], + "spans": [ + { + "bbox": [ + 130, + 485, + 482, + 605 + ], + "type": "text", + "content": "Synthetic Data. We evaluate Spring-Gaus using a synthetic dataset. We first collected fourteen 3D models, including point clouds and meshes, to use as initial point clouds for our simulations. Some of them are sourced from PAC-NeRF [18] and OmniObject3D [45]. Following PAC-NeRF, we employ the Material Point Method [11,39] to simulate the dynamics of elastic objects to generate synthetic data. Our dataset features elastic objects with various stiffness levels and diverse geometric forms. Multi-view RGB videos are rendered using Blender [5], with each sequence comprising 10 viewpoints and 30 frames at a resolution of " + }, + { + "bbox": [ + 130, + 485, + 482, + 605 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 130, + 485, + 482, + 605 + ], + "type": "text", + "content": ". The initial 20 frames are utilized for dynamic reconstruction, while the subsequent 10 frames are dedicated to evaluating future prediction performance." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 606, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 606, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 606, + 482, + 666 + ], + "type": "text", + "content": "Real-World Data. In addition to synthetic datasets, we further assess Spring-Gaus using captured real-world examples. Our collection process distinguishes between static scenes and dynamic multi-view videos. For static scenes, we position each object on a table and capture 50-70 images from various viewpoints within the upper hemisphere surrounding the object. To obtain camera" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 395, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 395, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 395, + 91, + 447, + 102 + ], + "type": "text", + "content": "Spring-Gaus" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 175, + 114, + 444, + 196 + ], + "blocks": [ + { + "bbox": [ + 175, + 114, + 444, + 196 + ], + "lines": [ + { + "bbox": [ + 175, + 114, + 444, + 196 + ], + "spans": [ + { + "bbox": [ + 175, + 114, + 444, + 196 + ], + "type": "image", + "image_path": "f94c41c50fd8443bb7e3cbda1bfb71df2f2b25a9084a735fa1dc61606ee2d19d.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 147, + 213, + 465, + 225 + ], + "lines": [ + { + "bbox": [ + 147, + 213, + 465, + 225 + ], + "spans": [ + { + "bbox": [ + 147, + 213, + 465, + 225 + ], + "type": "text", + "content": "Fig. 3: Registration from static scene to dynamic scene for real-world sample." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 246, + 482, + 307 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 246, + 482, + 307 + ], + "spans": [ + { + "bbox": [ + 130, + 246, + 482, + 307 + ], + "type": "text", + "content": "poses for static scenes, we utilize the off-the-shelf Structure-from-Motion toolkit COLMAP [35, 36]. The dynamic aspect of our dataset is represented through multi-view RGB videos, recorded from three distinct viewpoints, at a resolution of " + }, + { + "bbox": [ + 130, + 246, + 482, + 307 + ], + "type": "inline_equation", + "content": "1920 \\times 1080" + }, + { + "bbox": [ + 130, + 246, + 482, + 307 + ], + "type": "text", + "content": ". The camera parameters for dynamic scenes are obtained through calibration using a checkerboard. SAM [15] is used to obtain object masks." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 322, + 279, + 335 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 322, + 279, + 335 + ], + "spans": [ + { + "bbox": [ + 132, + 322, + 279, + 335 + ], + "type": "text", + "content": "4.2 Implementation Details" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 339, + 482, + 398 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 339, + 482, + 398 + ], + "spans": [ + { + "bbox": [ + 130, + 339, + 482, + 398 + ], + "type": "text", + "content": "When learning the 3D Gaussians, the distribution of Gaussian kernels is highly dependent on the initial points. Gaussian kernels tend to concentrate on the surface if the initial points are derived from SfM results. However, we initialize a large number of points inside a cube, resulting in a more uniform distribution of the final kernels." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 399, + 482, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 399, + 482, + 604 + ], + "spans": [ + { + "bbox": [ + 130, + 399, + 482, + 604 + ], + "type": "text", + "content": "For real-world samples, we collect static and dynamic scenes separately because three viewpoints are insufficient for effective 3D Gaussian reconstruction [14]. This results in the Gaussian kernels for static and dynamic scenes being in different coordinate systems. Therefore, for real samples, we employ a registration network before dynamic reconstruction to align the Gaussian kernels from static scene coordinates to dynamic scene coordinates, as shown in Fig. 3. Specifically, we optimize a scale factor " + }, + { + "bbox": [ + 130, + 399, + 482, + 604 + ], + "type": "inline_equation", + "content": "s_r" + }, + { + "bbox": [ + 130, + 399, + 482, + 604 + ], + "type": "text", + "content": ", a translation vector " + }, + { + "bbox": [ + 130, + 399, + 482, + 604 + ], + "type": "inline_equation", + "content": "\\pmb{t}_r" + }, + { + "bbox": [ + 130, + 399, + 482, + 604 + ], + "type": "text", + "content": ", and a rotation vector " + }, + { + "bbox": [ + 130, + 399, + 482, + 604 + ], + "type": "inline_equation", + "content": "\\pmb{r}_r" + }, + { + "bbox": [ + 130, + 399, + 482, + 604 + ], + "type": "text", + "content": " for the registration. We represent 3D rotations using a continuous 6D vector " + }, + { + "bbox": [ + 130, + 399, + 482, + 604 + ], + "type": "inline_equation", + "content": "\\pmb{r}_r \\in \\mathbb{R}^6" + }, + { + "bbox": [ + 130, + 399, + 482, + 604 + ], + "type": "text", + "content": ", which has been shown to be more amenable for gradient-based optimization [53]. However, slight deformations of objects and variations in lighting conditions and exposure times during data capture are inevitable. These variations can result in color discrepancies between frames across time and viewpoints. The color discrepancy is the major noise source that hinders registration and reconstruction. Thus, we refine our approach by computing the loss function between mask images, incorporating both mask center loss and perceptual loss into our model. Consequently, the revised loss function for real samples is expressed as:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 147, + 622, + 481, + 635 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 147, + 622, + 481, + 635 + ], + "spans": [ + { + "bbox": [ + 147, + 622, + 481, + 635 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\left(1 - \\lambda_ {\\mathrm {d - s s i m}}\\right) \\mathcal {L} _ {1} + \\lambda_ {\\mathrm {d - s s i m}} \\mathcal {L} _ {\\mathrm {d - s s i m}} + \\lambda_ {\\text {c e n t e r}} \\mathcal {L} _ {\\text {c e n t e r}} + \\lambda_ {\\text {p e r c e p}} \\mathcal {L} _ {\\text {p e r c e p}}, \\tag {19}", + "image_path": "3eaf5cf7220ad6c73c52f76fa5430c9363d179e4af52e124b15d8dc25de248fc.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{d - ssim}} = 0.8" + }, + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{center}} = 1.0" + }, + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{percep}} = 0.1" + }, + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "content": " are the weighting coefficients. Here, " + }, + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{center}}" + }, + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "content": " quantifies the discrepancy between the center coordinates of" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 228, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 228, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 228, + 102 + ], + "type": "text", + "content": "L. Zhong et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 133, + 114, + 479, + 218 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 479, + 218 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 479, + 218 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 479, + 218 + ], + "type": "table", + "html": "
TorusCrossCreamApplePasteChessBananaMean
CD↓Spring-Gaus (ours)2.381.572.221.877.032.5918.485.16
PAC-NeRF [18]2.473.872.214.6937.708.2066.4317.94
EMD↓Spring-Gaus (ours)0.0870.0510.0940.0760.1260.0950.1350.095
PAC-NeRF [18]0.0550.1110.0830.1080.1920.1550.2340.134
PSNR↑Spring-Gaus (ours)16.8316.9315.4221.5514.7116.0817.8917.06
PAC-NeRF [18]17.4614.1515.3719.9412.3215.0816.0415.77
SSIM↑Spring-Gaus (ours)0.9190.9400.8620.9020.8720.8810.9040.897
PAC-NeRF [18]0.9130.9060.8580.8780.8190.8480.8660.870
", + "image_path": "66cb43f6edbce48e9fde54d4f86a42cd10259a064ecfa80dc2bfaa63701528a1.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 138, + 273, + 472, + 376 + ], + "blocks": [ + { + "bbox": [ + 132, + 219, + 480, + 251 + ], + "lines": [ + { + "bbox": [ + 132, + 219, + 480, + 251 + ], + "spans": [ + { + "bbox": [ + 132, + 219, + 480, + 251 + ], + "type": "text", + "content": "Table 1: Quantitative results of future prediction on synthetic data. Spring-Gaus excels in short-term future prediction. Meanwhile, since we separate appearance and dynamics modeling, Spring-Gaus also maintains good rendering quality." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 138, + 273, + 472, + 376 + ], + "lines": [ + { + "bbox": [ + 138, + 273, + 472, + 376 + ], + "spans": [ + { + "bbox": [ + 138, + 273, + 472, + 376 + ], + "type": "table", + "html": "
TorusCrossCreamApplePasteChessBananaMean
CD↓Spring-Gaus (ours)0.170.480.360.380.191.802.600.85
PAC-NeRF [18]4.921.100.771.113.140.962.772.11
Dy-Gaus [26]579773479727284976429631305
4D-Gaus [44]11.121.772.872.231.953.977.134.43
EMD↓Spring-Gaus (ours)0.0400.0370.0310.0330.0220.0630.0520.040
PAC-NeRF [18]0.0560.0520.0410.0450.0540.0520.0620.052
Dy-Gaus [26]0.8570.9550.7830.9031.7390.9851.5911.116
4D-Gaus [44]0.1300.0780.0890.0880.0700.0970.1120.095
", + "image_path": "4e8a200bb47f559c2de5432b2ae6b223c85ce1e1429e28f1916e77b942097618.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 377, + 479, + 398 + ], + "lines": [ + { + "bbox": [ + 132, + 377, + 479, + 398 + ], + "spans": [ + { + "bbox": [ + 132, + 377, + 479, + 398 + ], + "type": "text", + "content": "Table 2: Quantitative results of dynamic reconstruction on synthetic data. Spring-Gaus has excellent geometric accuracy in dynamic reconstruction." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 132, + 433, + 480, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 433, + 480, + 456 + ], + "spans": [ + { + "bbox": [ + 132, + 433, + 480, + 456 + ], + "type": "text", + "content": "the rendered and ground truth images, while " + }, + { + "bbox": [ + 132, + 433, + 480, + 456 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{percep}}" + }, + { + "bbox": [ + 132, + 433, + 480, + 456 + ], + "type": "text", + "content": " represents the perceptual loss [13], which is based on the VGG16 architecture [37]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 133, + 476, + 345, + 488 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 476, + 345, + 488 + ], + "spans": [ + { + "bbox": [ + 133, + 476, + 345, + 488 + ], + "type": "text", + "content": "4.3 Qualitative and Quantitative Results" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 498, + 480, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 498, + 480, + 593 + ], + "spans": [ + { + "bbox": [ + 132, + 498, + 480, + 593 + ], + "type": "text", + "content": "We assess the performance of our approach using both synthetic and real-world datasets. Firstly, we validate the effectiveness of our method on synthetic datasets, utilizing the first 20 frames as our observation set for training dynamic modeling capabilities. For future prediction, we employ the subsequent 10 frames, comparing the ground truth with our model's predictions of future frames. Additionally, we benchmark our approach against the most relevant methods in dynamic scene modeling and physics-informed learning, including PAC-NeRF [18], Dynamic 3D Gaussians [26], and 4D Gaussian Splatting [44]." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 594, + 480, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 594, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 594, + 480, + 665 + ], + "type": "text", + "content": "The qualitative results are presented in Fig. 4. We also report quantitative results by computing the Chamfer Distance (CD) and Earth Mover's Distance (EMD). In all tables, the Chamfer Distance is measured based on squared distance, with units expressed as " + }, + { + "bbox": [ + 132, + 594, + 480, + 665 + ], + "type": "inline_equation", + "content": "1 \\times 10^{3}mm^{2}" + }, + { + "bbox": [ + 132, + 594, + 480, + 665 + ], + "type": "text", + "content": ". The quantitative analysis of dynamic reconstruction, shown in Tab. 2, reveals that Spring-Gaus can accurately simulate object dynamics for the synthetic data." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 395, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 395, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 395, + 91, + 447, + 102 + ], + "type": "text", + "content": "Spring-Gaus" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 114, + 482, + 571 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 482, + 571 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 482, + 571 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 482, + 571 + ], + "type": "image", + "image_path": "34cf5ad2c67f59e147f820c12520acb53b250549f8d061319f616d4a63d3a380.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 588, + 482, + 624 + ], + "lines": [ + { + "bbox": [ + 130, + 588, + 482, + 624 + ], + "spans": [ + { + "bbox": [ + 130, + 588, + 482, + 624 + ], + "type": "text", + "content": "Fig.4: Qualitative results on synthetic data. Compared with PAC-NeRF [18], Dynamic 3D Gaussians [26] and 4D Gaussian Splatting [44], Spring-Gaus can maintain a good geometry and appearance while reconstructing reasonable dynamics." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 228, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 228, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 228, + 102 + ], + "type": "text", + "content": "L. Zhong et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 134, + 113, + 307, + 217 + ], + "blocks": [ + { + "bbox": [ + 134, + 113, + 307, + 217 + ], + "lines": [ + { + "bbox": [ + 134, + 113, + 307, + 217 + ], + "spans": [ + { + "bbox": [ + 134, + 113, + 307, + 217 + ], + "type": "image", + "image_path": "27a05d1c66777d2fd18a08deff01dff3718388705359443008465c2f3b43667c.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 237, + 480, + 260 + ], + "lines": [ + { + "bbox": [ + 131, + 237, + 480, + 260 + ], + "spans": [ + { + "bbox": [ + 131, + 237, + 480, + 260 + ], + "type": "text", + "content": "Fig. 5: Qualitative results of future prediction on real-world samples. Predicted dynamics closely follow real observations." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 307, + 113, + 399, + 217 + ], + "blocks": [ + { + "bbox": [ + 307, + 113, + 399, + 217 + ], + "lines": [ + { + "bbox": [ + 307, + 113, + 399, + 217 + ], + "spans": [ + { + "bbox": [ + 307, + 113, + 399, + 217 + ], + "type": "image", + "image_path": "8d99da0f958e146ff55a5265e9359c90e49105e0361bde02951c869b35470d31.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 399, + 113, + 482, + 217 + ], + "blocks": [ + { + "bbox": [ + 399, + 113, + 482, + 217 + ], + "lines": [ + { + "bbox": [ + 399, + 113, + 482, + 217 + ], + "spans": [ + { + "bbox": [ + 399, + 113, + 482, + 217 + ], + "type": "image", + "image_path": "5bfb97ef473387211a109ed2e243fb1d4a2ad13dfbcccd07566af81eb1cc0427.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 289, + 480, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 289, + 480, + 373 + ], + "spans": [ + { + "bbox": [ + 130, + 289, + 480, + 373 + ], + "type": "text", + "content": "In Tab. 1, we show our method's capability in predicting future frames. Our method outperforms PAC-NeRF across CD, EMD, Peak Signal-to-Noise Ratio (PSNR), and Structural Similarity Index (SSIM) metrics. We show qualitative results in Fig. 4. We observe that Spring-Gaus gives faithful reconstruction and future prediction that closely aligns with the ground truth, even though the simulation engines are different. This shows that Spring-Gaus is not only expressive but also allows efficient identification through gradient-based optimization." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 376, + 481, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 376, + 481, + 472 + ], + "spans": [ + { + "bbox": [ + 130, + 376, + 481, + 472 + ], + "type": "text", + "content": "We also evaluate Spring-Gaus on real-world samples. It is difficult for both NeRF [28] and 3D Gaussian Splatting [14] to reconstruct the correct geometric information under extremely sparse camera views. However, due to 3D Gaussians' explicit representation that we can directly operate Gaussian kernels, we could reconstruct the static scenes and dynamic scenes in a different coordinate system and then align them using a registration network mentioned in Sec. 4.2, which is hard to do under an implicit representation. We show the registration process in Fig. 3. Results on real-world samples can be found in Fig. 5 and Fig. 1." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 504, + 332, + 516 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 504, + 332, + 516 + ], + "spans": [ + { + "bbox": [ + 132, + 504, + 332, + 516 + ], + "type": "text", + "content": "4.4 Generalization to New Conditions" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 539, + 482, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 539, + 482, + 635 + ], + "spans": [ + { + "bbox": [ + 130, + 539, + 482, + 635 + ], + "type": "text", + "content": "In addition to future prediction, we also show that our method essentially creates a digital asset of the object from the multi-view videos, allowing dynamic simulation under different unseen environmental conditions. In Fig. 6, we edit the boundary conditions, such as adjusting the positions and the stickiness of the ground plane. In Fig. 7, we edit physical conditions, initialization conditions, and environmental conditions, such as object properties (making them softer or harder), initial velocities, and environmental gravity. Please check our project website* for videos of the results for a more expressive illustration." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 395, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 395, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 395, + 91, + 447, + 102 + ], + "type": "text", + "content": "Spring-Gaus" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 133, + 653, + 295, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 653, + 295, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 653, + 295, + 665 + ], + "type": "text", + "content": "* https://zlicheng.com/spring_gaus" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 114, + 303, + 171 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 303, + 171 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 303, + 171 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 303, + 171 + ], + "type": "image", + "image_path": "9f337b1a4dfdecbd32704379c1f3ec4bd5c7167f08bf89dff96b245df825ec26.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 133, + 180, + 302, + 236 + ], + "blocks": [ + { + "bbox": [ + 140, + 172, + 179, + 180 + ], + "lines": [ + { + "bbox": [ + 140, + 172, + 179, + 180 + ], + "spans": [ + { + "bbox": [ + 140, + 172, + 179, + 180 + ], + "type": "text", + "content": "High ground" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 201, + 172, + 236, + 180 + ], + "lines": [ + { + "bbox": [ + 201, + 172, + 236, + 180 + ], + "spans": [ + { + "bbox": [ + 201, + 172, + 236, + 180 + ], + "type": "text", + "content": "Mid ground" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 256, + 172, + 293, + 180 + ], + "lines": [ + { + "bbox": [ + 256, + 172, + 293, + 180 + ], + "spans": [ + { + "bbox": [ + 256, + 172, + 293, + 180 + ], + "type": "text", + "content": "Low ground" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 133, + 180, + 302, + 236 + ], + "lines": [ + { + "bbox": [ + 133, + 180, + 302, + 236 + ], + "spans": [ + { + "bbox": [ + 133, + 180, + 302, + 236 + ], + "type": "image", + "image_path": "d001745f6a0c502ff1a4cddedd883c09c35e7bec2847e6c7057958dafb16b536.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 149, + 237, + 173, + 243 + ], + "lines": [ + { + "bbox": [ + 149, + 237, + 173, + 243 + ], + "spans": [ + { + "bbox": [ + 149, + 237, + 173, + 243 + ], + "type": "text", + "content": "Normal" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 206, + 237, + 230, + 243 + ], + "lines": [ + { + "bbox": [ + 206, + 237, + 230, + 243 + ], + "spans": [ + { + "bbox": [ + 206, + 237, + 230, + 243 + ], + "type": "text", + "content": "Smooth" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 266, + 237, + 285, + 244 + ], + "lines": [ + { + "bbox": [ + 266, + 237, + 285, + 244 + ], + "spans": [ + { + "bbox": [ + 266, + 237, + 285, + 244 + ], + "type": "text", + "content": "Sticky" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 132, + 253, + 480, + 277 + ], + "lines": [ + { + "bbox": [ + 132, + 253, + 480, + 277 + ], + "spans": [ + { + "bbox": [ + 132, + 253, + 480, + 277 + ], + "type": "text", + "content": "Fig. 6: We edit boundary conditions in these demos. Changing grounds' position and using smooth or sticky ground." + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 311, + 114, + 482, + 171 + ], + "blocks": [ + { + "bbox": [ + 311, + 114, + 482, + 171 + ], + "lines": [ + { + "bbox": [ + 311, + 114, + 482, + 171 + ], + "spans": [ + { + "bbox": [ + 311, + 114, + 482, + 171 + ], + "type": "image", + "image_path": "9cde7ca8efb75f7977e40c20db31ec234cd393d79e98fc96a6b7a4ba94e707ad.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 311, + 180, + 481, + 236 + ], + "blocks": [ + { + "bbox": [ + 320, + 172, + 358, + 180 + ], + "lines": [ + { + "bbox": [ + 320, + 172, + 358, + 180 + ], + "spans": [ + { + "bbox": [ + 320, + 172, + 358, + 180 + ], + "type": "text", + "content": "High ground" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 380, + 172, + 414, + 180 + ], + "lines": [ + { + "bbox": [ + 380, + 172, + 414, + 180 + ], + "spans": [ + { + "bbox": [ + 380, + 172, + 414, + 180 + ], + "type": "text", + "content": "Mid ground" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 435, + 172, + 471, + 180 + ], + "lines": [ + { + "bbox": [ + 435, + 172, + 471, + 180 + ], + "spans": [ + { + "bbox": [ + 435, + 172, + 471, + 180 + ], + "type": "text", + "content": "Low ground" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 311, + 180, + 481, + 236 + ], + "lines": [ + { + "bbox": [ + 311, + 180, + 481, + 236 + ], + "spans": [ + { + "bbox": [ + 311, + 180, + 481, + 236 + ], + "type": "image", + "image_path": "1de20ed7fdca102612c668c48950a28f36722e6c834a0cd2963d87fc952e0ca3.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 329, + 237, + 353, + 243 + ], + "lines": [ + { + "bbox": [ + 329, + 237, + 353, + 243 + ], + "spans": [ + { + "bbox": [ + 329, + 237, + 353, + 243 + ], + "type": "text", + "content": "Normal" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 384, + 237, + 408, + 244 + ], + "lines": [ + { + "bbox": [ + 384, + 237, + 408, + 244 + ], + "spans": [ + { + "bbox": [ + 384, + 237, + 408, + 244 + ], + "type": "text", + "content": "Smooth" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 443, + 237, + 463, + 244 + ], + "lines": [ + { + "bbox": [ + 443, + 237, + 463, + 244 + ], + "spans": [ + { + "bbox": [ + 443, + 237, + 463, + 244 + ], + "type": "text", + "content": "Sticky" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 133, + 294, + 302, + 352 + ], + "blocks": [ + { + "bbox": [ + 133, + 294, + 302, + 352 + ], + "lines": [ + { + "bbox": [ + 133, + 294, + 302, + 352 + ], + "spans": [ + { + "bbox": [ + 133, + 294, + 302, + 352 + ], + "type": "image", + "image_path": "0a88e24a00358f1c56ae036d6b43b306c630fb103b27f88cc53fec03e1f6f5b4.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 150, + 353, + 173, + 359 + ], + "lines": [ + { + "bbox": [ + 150, + 353, + 173, + 359 + ], + "spans": [ + { + "bbox": [ + 150, + 353, + 173, + 359 + ], + "type": "text", + "content": "Original" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 189, + 353, + 298, + 359 + ], + "lines": [ + { + "bbox": [ + 189, + 353, + 298, + 359 + ], + "spans": [ + { + "bbox": [ + 189, + 353, + 298, + 359 + ], + "type": "text", + "content": "Random init velocity Smaller & Softer" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 133, + 360, + 302, + 418 + ], + "blocks": [ + { + "bbox": [ + 133, + 360, + 302, + 418 + ], + "lines": [ + { + "bbox": [ + 133, + 360, + 302, + 418 + ], + "spans": [ + { + "bbox": [ + 133, + 360, + 302, + 418 + ], + "type": "image", + "image_path": "06428440e0a26eb043d0b775992f8e03c68576ec37876ffec49a59072a3474be.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 150, + 418, + 173, + 426 + ], + "lines": [ + { + "bbox": [ + 150, + 418, + 173, + 426 + ], + "spans": [ + { + "bbox": [ + 150, + 418, + 173, + 426 + ], + "type": "text", + "content": "Original" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 200, + 418, + 233, + 426 + ], + "lines": [ + { + "bbox": [ + 200, + 418, + 233, + 426 + ], + "spans": [ + { + "bbox": [ + 200, + 418, + 233, + 426 + ], + "type": "text", + "content": "Zero gravity" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 245, + 418, + 302, + 426 + ], + "lines": [ + { + "bbox": [ + 245, + 418, + 302, + 426 + ], + "spans": [ + { + "bbox": [ + 245, + 418, + 302, + 426 + ], + "type": "text", + "content": "Random init velocity" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 139, + 435, + 473, + 447 + ], + "lines": [ + { + "bbox": [ + 139, + 435, + 473, + 447 + ], + "spans": [ + { + "bbox": [ + 139, + 435, + 473, + 447 + ], + "type": "text", + "content": "Fig. 7: We edit physical conditions, initial velocities and gravities in these demos." + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 311, + 294, + 481, + 352 + ], + "blocks": [ + { + "bbox": [ + 311, + 294, + 481, + 352 + ], + "lines": [ + { + "bbox": [ + 311, + 294, + 481, + 352 + ], + "spans": [ + { + "bbox": [ + 311, + 294, + 481, + 352 + ], + "type": "image", + "image_path": "5a55546295d3f00f0a1fbcac37791e9955fd20b9112c0c70f4c6f4b29f6b2ff4.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 432, + 353, + 474, + 359 + ], + "lines": [ + { + "bbox": [ + 432, + 353, + 474, + 359 + ], + "spans": [ + { + "bbox": [ + 432, + 353, + 474, + 359 + ], + "type": "text", + "content": "Bigger & Softer" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_caption" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 311, + 360, + 481, + 418 + ], + "blocks": [ + { + "bbox": [ + 328, + 353, + 351, + 360 + ], + "lines": [ + { + "bbox": [ + 328, + 353, + 351, + 360 + ], + "spans": [ + { + "bbox": [ + 328, + 353, + 351, + 360 + ], + "type": "text", + "content": "Original" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 386, + 353, + 404, + 360 + ], + "lines": [ + { + "bbox": [ + 386, + 353, + 404, + 360 + ], + "spans": [ + { + "bbox": [ + 386, + 353, + 404, + 360 + ], + "type": "text", + "content": "Windy" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 311, + 360, + 481, + 418 + ], + "lines": [ + { + "bbox": [ + 311, + 360, + 481, + 418 + ], + "spans": [ + { + "bbox": [ + 311, + 360, + 481, + 418 + ], + "type": "image", + "image_path": "a93dea573f604609d7f3fb16ccc63210efc1a2f50bb65ac310d9ebca036de7be.jpg" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 328, + 418, + 351, + 426 + ], + "lines": [ + { + "bbox": [ + 328, + 418, + 351, + 426 + ], + "spans": [ + { + "bbox": [ + 328, + 418, + 351, + 426 + ], + "type": "text", + "content": "Original" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 379, + 418, + 416, + 427 + ], + "lines": [ + { + "bbox": [ + 379, + 418, + 416, + 427 + ], + "spans": [ + { + "bbox": [ + 379, + 418, + 416, + 427 + ], + "type": "text", + "content": "Hypergravity" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 432, + 418, + 474, + 426 + ], + "lines": [ + { + "bbox": [ + 432, + 418, + 474, + 426 + ], + "spans": [ + { + "bbox": [ + 432, + 418, + 474, + 426 + ], + "type": "text", + "content": "Bigger & Softer" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_caption" + } + ], + "index": 30 + }, + { + "bbox": [ + 132, + 477, + 219, + 490 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 477, + 219, + 490 + ], + "spans": [ + { + "bbox": [ + 132, + 477, + 219, + 490 + ], + "type": "text", + "content": "5 Conclusion" + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "content": "In this paper, we introduce Spring-Gaus, a novel framework designed to acquire simulatable digital assets of elastic objects from video observations. By integrating a 3D Spring-Mass model, Spring-Gaus allows the reconstruction of object appearance and dynamics. A key feature of our approach is the distinct separation between the learning processes for appearance and physics, thereby circumventing potential issues with optimization interference. We evaluate Spring-Gaus on both synthetic and real-world datasets, demonstrating its capability to reconstruct geometry, appearance, and physical dynamic properties. Moreover, our method demonstrates improved capabilities by predicting short-term future dynamics under different environmental conditions. This showcases its strength in identifying physical properties from observational data and predicting the dynamics of reconstructed digital assets." + } + ] + } + ], + "index": 36 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 228, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 228, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 228, + 102 + ], + "type": "text", + "content": "L. Zhong et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 482, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 482, + 152 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 482, + 152 + ], + "type": "text", + "content": "Acknowledgments. The work is in part supported by the Amazon AICE Award, NSF RI #2211258, #2338203, ONR YIP N00014-24-1-2117, and ONR MURI N00014-22-1-2740." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 168, + 197, + 181 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 168, + 197, + 181 + ], + "spans": [ + { + "bbox": [ + 133, + 168, + 197, + 181 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 190, + 480, + 666 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 138, + 190, + 480, + 224 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 190, + 480, + 224 + ], + "spans": [ + { + "bbox": [ + 138, + 190, + 480, + 224 + ], + "type": "text", + "content": "1. Barron, J.T., Mildenhall, B., Tancik, M., Hedman, P., Martin-Brualla, R., Srinivasan, P.P.: Mip-NeRF: A multiscale representation for anti-aliasing neural radiance fields. In: CVPR (2021)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 224, + 480, + 246 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 224, + 480, + 246 + ], + "spans": [ + { + "bbox": [ + 138, + 224, + 480, + 246 + ], + "type": "text", + "content": "2. Chen, A., Xu, Z., Zhao, F., Zhang, X., Xiang, F., Yu, J., Su, H.: Mvsnerf: Fast generalizable radiance field reconstruction from multi-view stereo. In: ICCV (2021)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 246, + 480, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 246, + 480, + 267 + ], + "spans": [ + { + "bbox": [ + 138, + 246, + 480, + 267 + ], + "type": "text", + "content": "3. Chen, H., Wu, R., Grinspun, E., Zheng, C., Chen, P.Y.: Implicit neural spatial representations for time-dependent pdes. In: ICML (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 267, + 480, + 299 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 267, + 480, + 299 + ], + "spans": [ + { + "bbox": [ + 138, + 267, + 480, + 299 + ], + "type": "text", + "content": "4. Chu, M., Liu, L., Zheng, Q., Franz, E., Seidel, H.P., Theobalt, C., Zayer, R.: Physics informed neural fields for smoke reconstruction with sparse data. ACM TOG 41(4) (jul 2022). https://doi.org/10.1145/3528223.3530169" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 300, + 480, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 300, + 480, + 331 + ], + "spans": [ + { + "bbox": [ + 138, + 300, + 480, + 331 + ], + "type": "text", + "content": "5. Community, B.O.: Blender - a 3D modelling and rendering package. Blender Foundation, Stichting Blender Foundation, Amsterdam (2018), http://www.blender.org" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 332, + 480, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 332, + 480, + 353 + ], + "spans": [ + { + "bbox": [ + 138, + 332, + 480, + 353 + ], + "type": "text", + "content": "6. Deng, Y., Yu, H.X., Wu, J., Zhu, B.: Learning vortex dynamics for fluid inference and prediction. In: ICLR (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 353, + 480, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 353, + 480, + 374 + ], + "spans": [ + { + "bbox": [ + 138, + 353, + 480, + 374 + ], + "type": "text", + "content": "7. Deng, Y., Yu, H.X., Zhang, D., Wu, J., Zhu, B.: Fluid simulation on neural flow maps. ACM TOG 42(6) (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 374, + 480, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 374, + 480, + 407 + ], + "spans": [ + { + "bbox": [ + 138, + 374, + 480, + 407 + ], + "type": "text", + "content": "8. Driess, D., Huang, Z., Li, Y., Tedrake, R., Toussaint, M.: Learning multi-object dynamics with compositional neural radiance fields. In: Conference on Robot Learning. pp. 1755-1768. PMLR (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 407, + 480, + 438 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 407, + 480, + 438 + ], + "spans": [ + { + "bbox": [ + 138, + 407, + 480, + 438 + ], + "type": "text", + "content": "9. Du, T., Wu, K., Ma, P., Wah, S., Spielberg, A., Rus, D., Matusik, W.: Diffpd: Differentiable projective dynamics. ACM TOG 41(2) (nov 2021). https://doi.org/10.1145/3490168" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 439, + 480, + 460 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 439, + 480, + 460 + ], + "spans": [ + { + "bbox": [ + 138, + 439, + 480, + 460 + ], + "type": "text", + "content": "10. Guan, S., Deng, H., Wang, Y., Yang, X.: Neurofluid: Fluid dynamics grounding with particle-driven neural radiance fields. In: ICML (2022)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 460, + 480, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 460, + 480, + 502 + ], + "spans": [ + { + "bbox": [ + 138, + 460, + 480, + 502 + ], + "type": "text", + "content": "1. Hu, Y., Fang, Y., Ge, Z., Qu, Z., Zhu, Y., Pradhana, A., Jiang, C.: A moving least squares material point method with displacement discontinuity and two-way rigid body coupling. ACM TOG 37(4) (jul 2018). https://doi.org/10.1145/3197517.3201293" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 502, + 480, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 502, + 480, + 525 + ], + "spans": [ + { + "bbox": [ + 138, + 502, + 480, + 525 + ], + "type": "text", + "content": "2. Huang, Y.H., Sun, Y.T., Yang, Z., Lyu, X., Cao, Y.P., Qi, X.: Sc-gs: Sparse-controlled gaussian splatting for editable dynamic scenes. CVPR (2024)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 525, + 480, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 525, + 480, + 547 + ], + "spans": [ + { + "bbox": [ + 138, + 525, + 480, + 547 + ], + "type": "text", + "content": "3. Johnson, J., Alahi, A., Fei-Fei, L.: Perceptual losses for real-time style transfer and super-resolution. In: ECCV (2016)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 547, + 480, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 547, + 480, + 568 + ], + "spans": [ + { + "bbox": [ + 138, + 547, + 480, + 568 + ], + "type": "text", + "content": "4. Kerbl, B., Kopanas, G., Leimkuhler, T., Drettakis, G.: 3d gaussian splatting for real-time radiance field rendering. ACM TOG 42(4) (July 2023)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 568, + 480, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 568, + 480, + 601 + ], + "spans": [ + { + "bbox": [ + 138, + 568, + 480, + 601 + ], + "type": "text", + "content": "5. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., Dollár, P., Girshick, R.: Segment Anything. arXiv preprint arXiv:2304.02643 (2023)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 138, + 601, + 480, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 601, + 480, + 633 + ], + "spans": [ + { + "bbox": [ + 138, + 601, + 480, + 633 + ], + "type": "text", + "content": "6. Kratimenos, A., Lei, J., Daniilidis, K.: Dynmf: Neural motion factorization for real-time dynamic view synthesis with 3d gaussian splattering. arXiv preprint arXiv:2312.00112 (2023)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 138, + 633, + 480, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 633, + 480, + 666 + ], + "spans": [ + { + "bbox": [ + 138, + 633, + 480, + 666 + ], + "type": "text", + "content": "7. Le Cleac'h, S., Yu, H.X., Guo, M., Howell, T., Gao, R., Wu, J., Manchester, Z., Schwager, M.: Differentiable physics simulation of dynamics-augmented neural objects. IEEE Robotics and Automation Letters 8(5), 2780-2787 (2023)" + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 395, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 395, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 395, + 91, + 447, + 102 + ], + "type": "text", + "content": "Spring-Gaus" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 666 + ], + "type": "list", + "angle": 0, + "index": 23, + "blocks": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "type": "text", + "content": "18. Li, X., Qiao, Y.L., Chen, P.Y., Jatavallabhula, K.M., Lin, M., Jiang, C., Gan, C.: PAC-NeRF: Physics augmented continuum neural radiance fields for geometry-agnostic system identification. In: ICLR (2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 150, + 480, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 150, + 480, + 172 + ], + "spans": [ + { + "bbox": [ + 133, + 150, + 480, + 172 + ], + "type": "text", + "content": "19. Li, Y., Chen, P.Y., Du, T., Matusik, W.: Learning preconditioners for conjugate gradient pde solvers. In: ICML (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 172, + 480, + 204 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 172, + 480, + 204 + ], + "spans": [ + { + "bbox": [ + 133, + 172, + 480, + 204 + ], + "type": "text", + "content": "20. Li, Y., Li, S., Sitzmann, V., Agrawal, P., Torralba, A.: 3d neural scene representations for visuomotor control. In: Conference on Robot Learning. pp. 112-123. PMLR (2022)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 205, + 480, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 205, + 480, + 237 + ], + "spans": [ + { + "bbox": [ + 132, + 205, + 480, + 237 + ], + "type": "text", + "content": "21. Li, Y., Wu, J., Tedrake, R., Tenenbaum, J.B., Torralba, A.: Learning particle dynamics for manipulating rigid bodies, deformable objects, and fluids. arXiv preprint arXiv:1810.01566 (2018)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 237, + 480, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 237, + 480, + 259 + ], + "spans": [ + { + "bbox": [ + 132, + 237, + 480, + 259 + ], + "type": "text", + "content": "22. Li, Z., Niklaus, S., Snavely, N., Wang, O.: Neural scene flow fields for space-time view synthesis of dynamic scenes. In: CVPR (2021)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 259, + 480, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 259, + 480, + 293 + ], + "spans": [ + { + "bbox": [ + 132, + 259, + 480, + 293 + ], + "type": "text", + "content": "23. Li, Z., Wang, Q., Cole, F., Tucker, R., Snavely, N.: Dynibar: Neural dynamic image-based rendering. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4273-4284 (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 293, + 480, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 293, + 480, + 315 + ], + "spans": [ + { + "bbox": [ + 132, + 293, + 480, + 315 + ], + "type": "text", + "content": "24. Lin, Y., Dai, Z., Zhu, S., Yao, Y.: Gaussian-flow: 4d reconstruction with dynamic 3d gaussian particle. arXiv preprint arXiv:2312.03431 (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 315, + 480, + 337 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 315, + 480, + 337 + ], + "spans": [ + { + "bbox": [ + 132, + 315, + 480, + 337 + ], + "type": "text", + "content": "25. Liu, Y., Peng, S., Liu, L., Wang, Q., Wang, P., Christian, T., Zhou, X., Wang, W.: Neural rays for occlusion-aware image-based rendering. In: CVPR (2022)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 337, + 480, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 337, + 480, + 357 + ], + "spans": [ + { + "bbox": [ + 132, + 337, + 480, + 357 + ], + "type": "text", + "content": "26. Luiten, J., Kopanas, G., Leibe, B., Ramanan, D.: Dynamic 3d gaussians: Tracking by persistent dynamic view synthesis. In: 3DV (2024)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 358, + 480, + 391 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 358, + 480, + 391 + ], + "spans": [ + { + "bbox": [ + 132, + 358, + 480, + 391 + ], + "type": "text", + "content": "27. Ma, P., Chen, P.Y., Deng, B., Tenenbaum, J.B., Du, T., Gan, C., Matusik, W.: Learning neural constitutive laws from motion observations for generalizable pde dynamics. In: ICML (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 392, + 480, + 424 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 392, + 480, + 424 + ], + "spans": [ + { + "bbox": [ + 132, + 392, + 480, + 424 + ], + "type": "text", + "content": "28. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: NeRF: Representing scenes as neural radiance fields for view synthesis. In: ECCV (2020)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 424, + 480, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 424, + 480, + 446 + ], + "spans": [ + { + "bbox": [ + 132, + 424, + 480, + 446 + ], + "type": "text", + "content": "29. Park, J.J., Florence, P., Straub, J., Newcombe, R., Lovegrove, S.: Deepsdf: Learning continuous signed distance functions for shape representation. In: CVPR (2019)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 446, + 480, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 446, + 480, + 468 + ], + "spans": [ + { + "bbox": [ + 132, + 446, + 480, + 468 + ], + "type": "text", + "content": "30. Park, K., Sinha, U., Barron, J.T., Bouaziz, S., Goldman, D.B., Seitz, S.M., Martin-Brualla, R.: Nerfies: Deformable neural radiance fields. In: ICCV (2021)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 468, + 480, + 501 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 468, + 480, + 501 + ], + "spans": [ + { + "bbox": [ + 132, + 468, + 480, + 501 + ], + "type": "text", + "content": "31. Park, K., Sinha, U., Hedman, P., Barron, J.T., Bouaziz, S., Goldman, D.B., Martin-Brualla, R., Seitz, S.M.: Hypernerf: A higher-dimensional representation for topologically varying neural radiance fields. arXiv preprint arXiv:2106.13228 (2021)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 501, + 480, + 523 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 501, + 480, + 523 + ], + "spans": [ + { + "bbox": [ + 132, + 501, + 480, + 523 + ], + "type": "text", + "content": "32. Pumarola, A., Corona, E., Pons-Moll, G., Moreno-Noguer, F.: D-NeRF: Neural Radiance Fields for Dynamic Scenes. In: CVPR (2021)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 523, + 480, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 523, + 480, + 567 + ], + "spans": [ + { + "bbox": [ + 132, + 523, + 480, + 567 + ], + "type": "text", + "content": "33. Raissi, M., Perdikaris, P., Karniadakis, G.: Physics-informed neural networks: A deep learning framework for solving forward and inverse problems involving nonlinear partial differential equations. Journal of Computational Physics 378, 686-707 (2019). https://doi.org/https://doi.org/10.1016/j.jcp.2018.10.045" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 567, + 480, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 567, + 480, + 589 + ], + "spans": [ + { + "bbox": [ + 132, + 567, + 480, + 589 + ], + "type": "text", + "content": "34. Sanchez-Gonzalez, A., Godwin, J., Pfaff, T., Ying, R., Leskovec, J., Battaglia, P.W.: Learning to simulate complex physics with graph networks. In: ICML (2020)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 589, + 480, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 589, + 480, + 600 + ], + "spans": [ + { + "bbox": [ + 132, + 589, + 480, + 600 + ], + "type": "text", + "content": "35. Schonberger, J.L., Frahm, J.M.: Structure-from-motion revisited. In: CVPR (2016)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 132, + 600, + 480, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 600, + 480, + 621 + ], + "spans": [ + { + "bbox": [ + 132, + 600, + 480, + 621 + ], + "type": "text", + "content": "36. Schonberger, J.L., Zheng, E., Pollefeys, M., Frahm, J.M.: Pixelwise view selection for unstructured multi-view stereo. In: ECCV (2016)" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 132, + 621, + 480, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 621, + 480, + 643 + ], + "spans": [ + { + "bbox": [ + 132, + 621, + 480, + 643 + ], + "type": "text", + "content": "37. Simonyan, K., Zisserman, A.: Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556 (2014)" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 132, + 643, + 480, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 643, + 480, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 643, + 480, + 666 + ], + "type": "text", + "content": "38. Sitzmann, V., Zollhöfer, M., Wetzstein, G.: Scene representation networks: Continuous 3d-structure-aware neural scene representations. In: NeurIPS (2019)" + } + ] + } + ], + "index": 22 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 227, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 227, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 227, + 101 + ], + "type": "text", + "content": "L. Zhong et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 490 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 148 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 148 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 148 + ], + "type": "text", + "content": "39. Stomakhin, A., Schroeder, C., Chai, L., Teran, J., Selle, A.: A material point method for snow simulation. ACM TOG 32(4) (jul 2013). https://doi.org/10.1145/2461912.2461948" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 149, + 482, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 149, + 482, + 171 + ], + "spans": [ + { + "bbox": [ + 130, + 149, + 482, + 171 + ], + "type": "text", + "content": "40. Sun, C., Sun, M., Chen, H.: Direct voxel grid optimization: Super-fast convergence for radiance fields reconstruction. In: CVPR (2022)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 172, + 481, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 172, + 481, + 194 + ], + "spans": [ + { + "bbox": [ + 132, + 172, + 481, + 194 + ], + "type": "text", + "content": "41. Tang, J., Ren, J., Zhou, H., Liu, Z., Zeng, G.: Dreamgaussian: Generative gaussian splatting for efficient 3d content creation. ICLR (2024)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 194, + 481, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 194, + 481, + 226 + ], + "spans": [ + { + "bbox": [ + 132, + 194, + 481, + 226 + ], + "type": "text", + "content": "42. Tretschk, E., Tewari, A., Golyanik, V., Zollhöfer, M., Lassner, C., Theobalt, C.: Non-rigid neural radiance fields: Reconstruction and novel view synthesis of a dynamic scene from monocular video. In: ICCV (2021)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 227, + 481, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 227, + 481, + 258 + ], + "spans": [ + { + "bbox": [ + 132, + 227, + 481, + 258 + ], + "type": "text", + "content": "43. Wang, C., MacDonald, L.E., Jeni, L.A., Lucey, S.: Flow supervision for deformable nerf. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 21128-21137 (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 259, + 481, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 259, + 481, + 281 + ], + "spans": [ + { + "bbox": [ + 132, + 259, + 481, + 281 + ], + "type": "text", + "content": "44. Wu, G., Yi, T., Fang, J., Xie, L., Zhang, X., Wei, W., Liu, W., Tian, Q., Xinggang, W.: 4d gaussian splatting for real-time dynamic scene rendering. In: CVPR (2024)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 282, + 481, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 282, + 481, + 314 + ], + "spans": [ + { + "bbox": [ + 132, + 282, + 481, + 314 + ], + "type": "text", + "content": "45. Wu, T., Zhang, J., Fu, X., Wang, Y., Ren, J., Pan, L., Wu, W., Yang, L., Wang, J., Qian, C., Lin, D., Liu, Z.: OmniObject3D: Large-vocabulary 3d object dataset for realistic perception, reconstruction and generation. In: CVPR (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 315, + 481, + 336 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 315, + 481, + 336 + ], + "spans": [ + { + "bbox": [ + 132, + 315, + 481, + 336 + ], + "type": "text", + "content": "46. Xie, T., Zong, Z., Qiu, Y., Li, X., Feng, Y., Yang, Y., Jiang, C.: Physsgaussian: Physics-integrated 3d gaussians for generative dynamics. In: CVPR (2024)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 337, + 481, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 337, + 481, + 357 + ], + "spans": [ + { + "bbox": [ + 132, + 337, + 481, + 357 + ], + "type": "text", + "content": "47. Xu, Q., Xu, Z., Philip, J., Bi, S., Shu, Z., Sunkavalli, K., Neumann, U.: Point-NeRF: Point-based neural radiance fields. In: CVPR (2022)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 358, + 481, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 358, + 481, + 380 + ], + "spans": [ + { + "bbox": [ + 132, + 358, + 481, + 380 + ], + "type": "text", + "content": "48. Yang, Z., Yang, H., Pan, Z., Zhang, L.: Real-time photorealistic dynamic scene representation and rendering with 4d gaussian splatting. In: ICLR (2024)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 380, + 481, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 380, + 481, + 402 + ], + "spans": [ + { + "bbox": [ + 132, + 380, + 481, + 402 + ], + "type": "text", + "content": "49. Yang, Z., Gao, X., Zhou, W., Jiao, S., Zhang, Y., Jin, X.: Deformable 3d gaussians for high-fidelity monocular dynamic scene reconstruction. In: CVPR (2024)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 403, + 481, + 424 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 403, + 481, + 424 + ], + "spans": [ + { + "bbox": [ + 132, + 403, + 481, + 424 + ], + "type": "text", + "content": "50. Yu, A., Ye, V., Tancik, M., Kanazawa, A.: pixelNeRF: Neural radiance fields from one or few images. In: CVPR (2021)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 425, + 481, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 425, + 481, + 445 + ], + "spans": [ + { + "bbox": [ + 132, + 425, + 481, + 445 + ], + "type": "text", + "content": "51. Yu, H., Julin, J., Milacski, Z.A., Niinuma, K., Jeni, L.A.: Cogs: Controllable gaussian splatting. arXiv preprint arXiv:2312.05664 (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 447, + 481, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 447, + 481, + 468 + ], + "spans": [ + { + "bbox": [ + 132, + 447, + 481, + 468 + ], + "type": "text", + "content": "52. Yu, H.X., Zheng, Y., Gao, Y., Deng, Y., Zhu, B., Wu, J.: Inferring hybrid neural fluid fields from videos. Advances in Neural Information Processing Systems (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 468, + 481, + 490 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 468, + 481, + 490 + ], + "spans": [ + { + "bbox": [ + 132, + 468, + 481, + 490 + ], + "type": "text", + "content": "53. Zhou, Y., Barnes, C., Lu, J., Yang, J., Li, H.: On the continuity of rotation representations in neural networks. In: CVPR (2019)" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 395, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 395, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 395, + 91, + 447, + 102 + ], + "type": "text", + "content": "Spring-Gaus" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Rectify the Regression Bias in Long-Tailed Object Detection/07bdb2f4-ff12-4767-80ed-dec2d3b31ce7_content_list.json b/2024/Rectify the Regression Bias in Long-Tailed Object Detection/07bdb2f4-ff12-4767-80ed-dec2d3b31ce7_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..399c65e9d088f99ab5437d7bbce50a885587cba2 --- /dev/null +++ b/2024/Rectify the Regression Bias in Long-Tailed Object Detection/07bdb2f4-ff12-4767-80ed-dec2d3b31ce7_content_list.json @@ -0,0 +1,1686 @@ +[ + { + "type": "text", + "text": "Rectify the Regression Bias in Long-Tailed Object Detection", + "text_level": 1, + "bbox": [ + 256, + 140, + 746, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Ke Zhu, Minghao Fu, Jie Shao, Tianyu Liu, and Jianxin Wu", + "bbox": [ + 250, + 212, + 750, + 228 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "State Key Laboratory for Novel Software Technology", + "bbox": [ + 323, + 239, + 676, + 253 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Nanjing University, Nanjing, China", + "bbox": [ + 382, + 255, + 617, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "School of Artificial Intelligence, Nanjing University, China", + "bbox": [ + 307, + 268, + 692, + 281 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{zhuk, fumh, shaoj, liuty}@lamda.nju.edu.cn, wujx2001@nju.edu.cn", + "bbox": [ + 254, + 282, + 746, + 296 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Long-tailed object detection faces great challenges because of its extremely imbalanced class distribution. Recent methods mainly focus on the classification bias and its loss function design, while ignoring the subtle influence of the regression branch. This paper shows that the regression bias exists and does adversely and seriously impact the detection accuracy. While existing methods fail to handle the regression bias, the class-specific regression head for rare classes is hypothesized to be the main cause of it in this paper. As a result, three kinds of viable solutions to cater for the rare categories are proposed, including adding a class-agnostic branch, clustering heads and merging heads. The proposed methods brings in consistent and significant improvements over existing long-tailed detection methods, especially in rare and common classes. The proposed method achieves state-of-the-art performance in the large vocabulary LVIS dataset with different backbones and architectures. It generalizes well to more difficult evaluation metrics, relatively balanced datasets, and the mask branch. This is the first attempt to reveal and explore rectifying of the regression bias in long-tailed object detection.", + "bbox": [ + 259, + 337, + 740, + 571 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 599, + 375, + 616 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Long-tailed object detection draws great attention [12] for its practical utility recently. Numerous efforts [1,7,34,35,37] have been made to tackle this challenging task, such as re-weighting [33,35,37], over-sampling [11,12,39], and balanced grouping [19,24]. These methods are proposed to prevent the tail classes from being overwhelmed due to discouraging gradients [33, 37], inferior predicting scores [11,24] or insufficient samples [11,28].", + "bbox": [ + 212, + 632, + 787, + 723 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Long-tailed detection often involves both classification and regression heads. While almost all existing methods focus on mitigating the classification bias (e.g., adjusting the classification structure in detection branches), little or no attention has been paid to the regression branch. We will show in this paper that the regression bias has significant adverse effects in long-tailed object detection, but previous methods failed to identify this important issue.", + "bbox": [ + 212, + 724, + 787, + 816 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* J. Wu is the corresponding author", + "bbox": [ + 217, + 824, + 464, + 839 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/cb55291d3410932d064ef0546897379e5704c6b2049bd8f4de0a558c7c6a3032.jpg", + "image_caption": [ + "(a) RCNN loss" + ], + "image_footnote": [], + "bbox": [ + 233, + 143, + 377, + 232 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/86ed05ecb93ccadeb52977fa1c71d2cdc8ab58bb374d63f4c205230d844b63ab.jpg", + "image_caption": [ + "(b) RPN loss", + "Fig. 1: 1a shows the RCNN regression train loss of frequent, common and rare categories. 1b shows the RPN regression loss. 1c shows distribution of per class mean object scales in LVIS1.0. 'delta' in 1c is the negative of difference between train and validation set sizes for different classes." + ], + "image_footnote": [], + "bbox": [ + 385, + 143, + 531, + 233 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/ce00f1ff39dd760bd8410b0995ffbbf80c95a55a864f2c7e91201e52872539a9.jpg", + "image_caption": [ + "(c) Mean scale in LVIS train and val set" + ], + "image_footnote": [], + "bbox": [ + 540, + 145, + 769, + 232 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Fig. 1 clearly showcases this issue. For different methods (EQLv2 [33], See-Saw [35], ECM [7] and CrossEntropy CE [12]) trained on the LVIS1.0 [12] dataset, we plot the regression branch's losses for the final detection RCNN head in Fig. 1a. It is clear that the regression losses of rare categories are significantly higher than those of frequent and common categories, which inevitably will lead to poor regression results (and hence detection results) for rare classes. We name this finding as the regression bias, but existing long-tailed detection methods all fail to deal with or even identify this issue.", + "bbox": [ + 212, + 348, + 787, + 469 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To further demonstrate the importance of regression, we calculate the class-wise mean scale of GT (groundtruth) boxes in LVIS train and validation sets, as well as their differences (cf. Fig. 1c). The scale shift of rare classes is much larger than that of frequent classes. Since regression is highly correlated with box scale [30], it is thus inherently difficult for a rare class to learn a good bounding box regressor with both few samples and large scale shift. In short, it is crucial to properly handle the regression bias in long-tailed object detection.", + "bbox": [ + 212, + 470, + 787, + 577 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our solution to rectify the regression bias is motivated by Fig. 1b. We find that the regression loss in RPN is balanced where rare, common and frequent categories have almost the same regression loss, which is almost immune to the regression bias when comparing Fig. 1b with Fig. 1a.", + "bbox": [ + 212, + 579, + 787, + 640 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The key difference between RPN and RCNN regression is that the former is class-agnostic (i.e., all classes share the same regression parameters), while the latter is class-specific. Then, one natural question arises: Can class-agnostic head improve a tail class's generalization ability and handle the regression bias? Our hypothesis is that rare classes do favor class-agnostic regression heads.", + "bbox": [ + 212, + 641, + 787, + 715 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our hypothesis is supported by experiments in Table 1, where we compared class-specific and -agnostic regression heads in RCNN. By replacing the classification head with groundtruth class labels, experiments in Table 1 disentangled the impact of the classification head and focus on the regression heads. It is clear that the class-agnostic head possesses substantial advantages: $\\mathrm{AP}_r^b$ (for rare) surges from 0.7 to 54.6, even surpassing $\\mathrm{AP}_f^b$ (for frequent classes)! However, the agnostic head will bring a small drop in frequent classes (e.g., from 40.7 to 40.0 in $\\mathrm{AP}_f^b$ ). The final version of our conjecture is: the rare (and possibly", + "bbox": [ + 212, + 719, + 787, + 842 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "K. Zhu et al.", + "bbox": [ + 271, + 114, + 359, + 127 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "common) classes indeed favor class-agnostic regression, while the frequent classes prefers a class-specific regression, and there should be a trade-off between the two to optimize for all three types of categories.", + "bbox": [ + 212, + 146, + 787, + 191 + ], + "page_idx": 2 + }, + { + "type": "table", + "img_path": "images/27ecd21c59bee80b0918849d16462b26f8b7b5827d5775457ef45cab397c25ee.jpg", + "table_caption": [ + "Table 1: Results on LVIS1.0 based on the CE baseline. We trained a Mask-RCNN R50-FPN detector [14] with class-specific (the default setting) and class-agnostic regression heads in the RCNN head. Please note that during test, the predicted proposals were provided groundtruth (GT) classification results." + ], + "table_footnote": [], + "table_body": "
Reg. HeadGTAPAPbAPrAPpcAPfb
Specificno18.719.70.316.531.7
yes37.239.834.141.340.7
Agnosticno18.019.30.715.531.5
yes38.445.654.647.040.0
", + "bbox": [ + 357, + 287, + 640, + 381 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Accordingly, we design three different methods to fully rectify the regression bias, including adding a class-agnostic head, clustering similar heads, or merging heads. All three methods bring in positive effects (cf. Table 2), which verifies that rectifying regression bias is indeed crucial. We choose to adopt 'adding a class-agnostic head' in our main experiment for its simplicity, which leads to consistent and significant improvements over previous long-tailed detection pipeline, and has achieved state-of-the-art performance with various backbones and architectures. Moreover, our method shows robust generalization ability (cf. Tables 5-8) under varying settings, including different datasets (COCO/COCOLT [36]), different evaluation metrics and even adapts to the mask branch design. Furthermore, visualizations show that the proposed method leads to more accurate bounding box predictions (cf. Fig. 4) and indeed alleviates the regression bias (cf. Fig. 5). In summary, our contributions are:", + "bbox": [ + 212, + 410, + 787, + 606 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. For the first time, we reveal and successfully handle the regression bias in long-tailed object detection.", + "2. We propose three remedies to alleviate this bias, all of which produce consistent improvements over existing methods.", + "3. Our method achieves state-of-the-art results on LVIS as well as generalizing across datasets, metrics and even the mask branch. Visualizations qualitatively verify our hypothesis, too." + ], + "bbox": [ + 220, + 616, + 784, + 722 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 214, + 746, + 387, + 762 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Vast number of pipelines and techniques have been invented in the field of general purpose object detection. Although most of these detectors can possibly be adapted to balanced detection datasets [25], they can nevertheless hardly handle extremely long-tailed distribution. Dedicated methods are hence required.", + "bbox": [ + 212, + 779, + 787, + 839 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "Rectify the Regression Bias", + "bbox": [ + 545, + 114, + 730, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Modern long-tailed learning methods aim to solve two important tasks: image classification [31] and object detection/segmentation [12, 30].", + "bbox": [ + 212, + 146, + 782, + 176 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The first task, long-tailed image classification [26], is well explored and displays great diversity in terms of specific methods [3,29], training pipelines [20,44], post-hoc analysis [22,27] and network architecture [16,23,38]. Early attempts in this family focused on the details of re-weighting [3] or re-sampling [8] techniques to provide a relatively balanced (data) distribution for tail classes. Later on, the decoupling pipeline [20] for long-tailed learning becomes popular, paving the way for various successors [43,44]. This decoupled paradigm are based on the insight that instance sampling is beneficial for representation, while the classifier needs to be calibrated to alleviate its bias. Some works try to unify long-tailed learning with theoretical analysis [22,27], which involves label distribution [29] and generalization bound [27]. Most recent methods are much more diverse in ideas: they involve knowledge distillation [16,38,46], self-supervised learning [23,45] or statistical approach [10] to better handle long-tailed recognition.", + "bbox": [ + 212, + 176, + 785, + 372 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The combination of long-tailed learning and object detection or instance segmentation proves to be more challenging [12], because the class distribution in LVIS [12] is extremely imbalanced and naively applying common object detection techniques leads to un-satisfactory results [18, 24, 34, 35, 37]. Dominating solutions in this area are re-sampling [12, 39] and re-weighting [17, 21, 33, 35, 37]. There are variants of them that adopt balanced grouping [24], class incremental learning [19], augmented feature sampling [11, 36] or extra data source [41]. A common characteristic of these long-tailed detection methods is that they only focus on the classification task, and ignore the subtle influence of the regression branch. In this paper, we thoroughly explore the previously unvisited regression bias in long-tailed object detection, and propose a simple yet novel method to tackle long-tailed object detection.", + "bbox": [ + 212, + 373, + 787, + 551 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In [36], there is an experiment with similar design as ours in Table 1. But, we emphasize that they are fundamentally different. On one hand, conclusions in [36] is that 'performance drop in LVIS is mainly caused by the proposal classification [36]' and hence they focused on classification. On the other hand, their class agnostic-head is a supplementary to support their classification hypothesis, while our design and experiments lead to a novel finding: the regression bias.", + "bbox": [ + 212, + 554, + 787, + 645 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3 Method", + "text_level": 1, + "bbox": [ + 215, + 667, + 331, + 684 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Now we elaborate three different remedies to alleviate the regression bias, and start from the background on long-tailed object detection.", + "bbox": [ + 212, + 700, + 785, + 731 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1 Preliminaries", + "text_level": 1, + "bbox": [ + 215, + 753, + 372, + 768 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We take Faster-RCNN [14] as an example. For a scene image $I$ , it is first fed into a backbone network $\\phi(\\cdot)$ (e.g., ResNet [15]) to get the image feature $f \\in \\mathbb{R}^{d \\times w \\times h}$ : $f = \\phi(I)$ , with dimensionality, width and height denoted as $d$ , $w$ , $h$ , respectively. A region proposal network (RPN) [30] that contains both agnostic", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "K. Zhu et al.", + "bbox": [ + 271, + 114, + 359, + 127 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/48d9ce9b44f1f0519ef1709b476b10bded8ad48d8f78d5d48de7414b2e12ff0c.jpg", + "image_caption": [ + "Fig. 2: Illustration of the regular two-stage detection pipeline and the proposed regression methods. Previous methods (the left figure) generally focus on the final classification branch (the yellow arrow), while we focus on rectifying the regression bias (the red arrow). The right part shows our three regression methods, including adding an extra branch $W_{0}$ , clustering regression heads (e.g., from $W_{1},\\ldots ,W_{C}$ to $W_1',W_2'$ ) and merging (e.g., merging rare categories into one regression head $W_{\\text{rare}}$ , cf. Table 2). In our main experiments, we choose 'adding an extra branch' for its simplicity. This figure needs to be viewed in color." + ], + "image_footnote": [], + "bbox": [ + 233, + 142, + 769, + 268 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "classification and regression branches is then applied on the feature tensor $f$ to produce proposals $p$ from pre-defined anchors. The ROAlign [14] then extracts proposal features:", + "bbox": [ + 212, + 421, + 784, + 465 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\np = \\operatorname {R P N} (f), \\quad f _ {p} = \\operatorname {R O I A l i g n} (f, p), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 364, + 479, + 785, + 497 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $p$ represents a large set of proposals, and $f_{p}$ is the set of aligned proposal features. $f_{p}$ goes into post-process modules (e.g., NMS [30]) before being sent to the RCNN head to get the final features set $f_{n}$ : $f_{n} = \\mathrm{RCNN}(f_{p})$ , which is fed to classification and regression branches (e.g., linear layers) to produce the prediction results.", + "bbox": [ + 212, + 507, + 787, + 583 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2 Our three remedies for the regression bias", + "text_level": 1, + "bbox": [ + 214, + 607, + 609, + 622 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In Faster RCNN, there is a dedicated regression head for each class:", + "bbox": [ + 214, + 633, + 705, + 648 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nr _ {i} = W _ {i} ^ {T} f _ {n}, \\quad i = 1, 2, \\dots , C, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 390, + 660, + 785, + 679 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $r_i = (\\delta x_i, \\delta y_i, \\delta w_i, \\delta h_i)$ represents the regression offset for class $i$ , and $W_i$ is the class-specific regression head (a linear layer). We will now present our approaches to rectify the regression bias. The existing pipeline and our proposed methods are shown in Fig. 2.", + "bbox": [ + 212, + 690, + 784, + 750 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Extra class-agnostic branch. This is a simple remedy to cope with the regression bias. Since rare classes favor a class-agnostic head while class-specific ones are slightly preferable to frequent classes in our hypothesis, we propose to use both heads. For class $i$ , its regression head changes to:", + "bbox": [ + 212, + 751, + 787, + 811 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nW _ {i} ^ {\\prime} = \\alpha W _ {0} + (1 - \\alpha) W _ {i}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 408, + 823, + 785, + 842 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "Rectify the Regression Bias", + "bbox": [ + 545, + 114, + 732, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/4743e126ba433b257a10cf7bb27612f59cdb5080e2838897aa757d1b9debca26.jpg", + "table_caption": [ + "Table 2: Our three different methods to alleviate the regression bias. In 2a, we add an extra class-agnostic regression head during training and testing, where the shared class-agonistic head and the class-specific head are combined using a weight $\\alpha$ . In 2b, we perform regression head clustering according to the number of instances (' $n'$ ) or mean box scale (' $s'$ ). In 2c, we directly merge the regression heads of rare, common or frequent classes. The original class-specific heads in 2b and 2c are replaced by the new heads. Note that 'base' means the baseline method, i.e., with only class-specific regression heads. In 2a, $\\alpha = 0$ is equivalent to the baseline method, and $\\alpha = 1.0$ is equivalent to merging all $r, c, f$ heads into a single head (i.e., the last row in 2c)." + ], + "table_footnote": [ + "(a) class-agnostic head", + "(b) Clustering heads", + "(c) Merging heads" + ], + "table_body": "
(α) APAPrAPbAPrknsAPAPrAPbAPrmerge APAPrAPbAPr
0.023.714.224.713.4base23.714.224.713.4base23.714.224.713.4
0.224.115.825.415.110✓24.514.526.314.4r25.116.326.716.7
0.525.117.527.018.010✓24.212.826.012.6c25.517.727.217.2
0.824.417.025.916.4100✓24.413.026.212.9r, c25.317.127.318.2
1.024.716.726.718.3100✓25.216.726.916.7r, c, f24.716.726.718.3
", + "bbox": [ + 228, + 280, + 785, + 383 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $W_0$ is a shared class-agnostic regression head for all classes, and $\\alpha$ is a hyper-parameter to balance class-agnostic and class-specific heads. We empirically find that this simple change leads to consistent improvements over the default class-specific regression head (cf. Table 2a), while simply setting $\\alpha = 0.5$ gives the optimal trade-off.", + "bbox": [ + 212, + 436, + 784, + 511 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Clustering heads. This method is motivated by the analysis in Fig. 1c. Since some categories have similar statistics, we can assign them a shared regression head to improve the generalization ability. We implement it by following three steps: sorting, grouping and assigning. First, we use the number of instance or mean box scale to sort the original categories in descending order ( $C = 1203$ in LVIS):", + "bbox": [ + 212, + 512, + 787, + 602 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nW _ {1}, \\dots , W _ {C} \\stackrel {\\text {s o r t}} {\\Longrightarrow} W _ {1} ^ {s}, \\dots , W _ {C} ^ {s}. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 393, + 603, + 784, + 622 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Then, we cluster them into $K$ groups. During clustering, we do not rely on time-cost algorithms such as K-means [4] or GMM [40]), but simply put adjacent classes into one group, and each group has the same number of classes $N = \\frac{C}{K}$ :", + "bbox": [ + 212, + 627, + 787, + 675 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\{(W _ {N \\times i + 1} ^ {s}, W _ {N \\times i + 2} ^ {s}, \\dots , W _ {N \\times i + N} ^ {s}) \\} _ {i = 0} ^ {K - 1}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 356, + 683, + 784, + 702 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Finally, each group $i$ share one regression head:", + "bbox": [ + 214, + 709, + 560, + 724 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nW _ {g i} \\stackrel {\\text {r e p l a c e}} {\\Longleftarrow} \\left(W _ {N \\times i + 1} ^ {s}, W _ {N \\times i + 2} ^ {s}, \\dots , W _ {N \\times i + N} ^ {s}\\right). \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 338, + 736, + 784, + 756 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "These shared regression matrix are then used for both training and testing. As shown in Table 2b, clustering heads with similar scale statistics brings robust improvements over the baseline methods (see more discussion in Sec. 4.4).", + "bbox": [ + 212, + 763, + 787, + 809 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Merge heads. This approach has similar motivation to the previous one, which clusters regression heads, but is even more straightforward. We simply", + "bbox": [ + 212, + 810, + 787, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "K. Zhu et al.", + "bbox": [ + 271, + 114, + 359, + 127 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "group the regression heads in pre-defined clusters. For example, we let all rare categories share a common regression head $W_{\\text{rare}}$ , and the same for common and frequent classes. Experimental results can be found in Table 2c, where we try four different combinations (note that $r, c$ means merging rare and common classes all together into one regression head). The results indicate that merging always leads to performance gains, especially for the rare categories. The most significant improvements come from merging only the common class (i.e., the row denoted as 'c' in Table 2c, also cf. Sec. 4.4 for discussion on partition shift).", + "bbox": [ + 212, + 146, + 787, + 267 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "All our proposed remedies improve the accuracy of long-tailed object detection and instance segmentation, which verifies the importance of rectifying the regression bias. We will show that all these methods lead to consistent and nontrivial margin over various existing approaches (cf. Table 3). We then choose adding a class-agnostic branch (the first approach, cf. Table 2a) to conduct the rest of our experiments for its simplicity.", + "bbox": [ + 212, + 268, + 787, + 358 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Attentive readers might advocate that all proposed three methods are to be verified on all datasets. However, we argue that this is not necessary and our choice has its advantages. On one hand, it does not need any dataset statistics, which is especially appreciated when the number of categories and the data distribution is unknown. On the other hand, by combining both class-agnostic and class-specific heads, it fully exploits object priors and per class knowledge: each type of head has its own merit, as Table 1 shows. Moreover, the validity of all three methods are already verified in Table 2-3.", + "bbox": [ + 212, + 359, + 787, + 479 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4 Experiment", + "text_level": 1, + "bbox": [ + 215, + 507, + 366, + 523 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We first combine 'Class-Agnostic Branch' (CAB), 'Clustering Head' (CH) and 'Merging Head' (MH) with existing long-tailed methods (cf. Table 3), then choose 'SeeSaw [35] + CAB' as 'Our' to compete with state-of-the-art methods (cf. Table 4). Although SeeSaw is a lower baseline than ECM [7], it is more stable in reproducing. Finally, we generalize our methods to various evaluation metrics, different datasets and the mask branch.", + "bbox": [ + 212, + 542, + 787, + 633 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.1 Experimental settings", + "text_level": 1, + "bbox": [ + 215, + 660, + 442, + 674 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Datasets. We use the large vocabulary dataset LVIS1.0 [12] as our main dataset, which contains 100k training and 20k validation images. Rare $(r)$ , common $(c)$ and frequent $(f)$ classes are defined by how many images they occur [12]: [0, 10] for rare, [11, 100] for common, and $(100, +\\infty)$ for frequent, respectively. We also adopt COCO-LT [25] and COCO2017 [36] to verify the generalization ability of our approach. COCO2017 is a large object detection dataset, containing 118k training and 5k validation images. It is relatively balanced in comparison with LVIS1.0. The COCO-LT dataset is an artificially sampled subset of COCO, with the same validation set but a long-tailed training set. It has about 99k training and 5k validation images. Following previous works [36], we partition COCO-LT", + "bbox": [ + 212, + 688, + 787, + 840 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "Rectify the Regression Bias", + "bbox": [ + 545, + 114, + 730, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/fdcf6c15fdb078283ae2bc1302ae9dac22b0e9d54db0530639b16ab475e76f1f.jpg", + "table_caption": [ + "Table 3: Experiments on LVIS1.0. We combine four existing methods with our approach 'adding a class-agnostic branch' (CAB), 'Clustering Head' (CH) or 'Merging Head' (MH). We reproduced RFS [12], EQLv2 [33], SeeSaw [35] and ECM [7] using their official code. For clarity, here we list the object detection metrics $\\mathrm{AP}^b$ in the first four columns while putting the instance segmentation metrics AP in the latter ones." + ], + "table_footnote": [], + "table_body": "
Method+Our APbAPRbAPCbAPf bAPAPrAPcAPf
RFS [12]base24.713.423.131.423.714.222.9
CAB27.018.025.332.925.117.523.9
CH26.917.225.533.025.216.724.3
MH27.318.225.832.825.317.124.5
EQLv2 [33]base26.016.124.032.525.217.424.1
CAB28.120.426.333.526.019.524.9
CH27.318.225.833.125.717.724.8
MH27.117.325.733.225.617.524.7
SeeSaw [35]base27.318.226.532.326.919.626.8
CAB28.919.928.333.627.720.227.3
CH29.119.828.933.427.820.627.7
MH28.518.428.433.727.419.827.4
ECM [7]base27.717.726.633.127.219.626.6
CAB29.118.428.933.927.819.128.0
CH28.718.028.333.727.819.627.7
MH28.618.628.233.527.719.827.7
", + "bbox": [ + 292, + 226, + 705, + 482 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "into 4 evaluation subsets according to the number of training instances per class, with bins of [1, 20), [20, 400), [400, 8000) and [8000, -), respectively.", + "bbox": [ + 212, + 508, + 782, + 537 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Training details. We reproduce four different methods as our baselines, including RFS [12], EQLv2 [33], SeeSaw [35] and ECM [7], following their default experiment settings. We employ MMDetection [5] as our detection framework to conduct our experiment, and train detection models of Faster-RCNN, Mask-RCNN and Cascade R-CNN for 1x or 2x scheduler (except Swin-Transformer based detectors), following previous works [7, 17]. The batch size and learning rate are set as 16 and 0.02, and the data augmentation strictly follows previous long-tailed detection methods [7, 33, 35]. During training, we use FP16 mixed precision training and the warmup strategy to stabilize the learning process. For the evaluation metrics, we adopt AP and $\\mathrm{AP}^b$ for instance segmentation and object detection, respectively, and adopt $\\mathrm{AP}_1^b$ , $\\mathrm{AP}_2^b$ , $\\mathrm{AP}_3^b$ and $\\mathrm{AP}_4^b$ on COCOLT, corresponding to its 4 different subsets. With the suggested practice in LVIS official website, we run all our experiment 3 times on 8 RTX3090 GPUs to reduce the variance. Please refer to our supplementary material for more detailed information.", + "bbox": [ + 212, + 539, + 787, + 765 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.2 LVIS detection and segmentation", + "text_level": 1, + "bbox": [ + 214, + 786, + 535, + 801 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Consistent improvements. we first evaluate the effectiveness of our method on the LVIS1.0 dataset by combing the proposed approach 'adding a class-agnostic", + "bbox": [ + 212, + 809, + 785, + 839 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "K. Zhu et al.", + "bbox": [ + 271, + 114, + 359, + 127 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/b629e40fb6d9c6f3aba4acc2ef3955b5b45d437cc6dcf63931ac4d0aa1806c64.jpg", + "table_caption": [ + "Table 4: Comparison with state-of-the-art methods on the LVIS1.0 dataset. The results of 'Our' means our proposed CAB regression method combined with the SeeSaw loss ('SeeSaw + CAB', cf. Sec. 4). We show the metrics of instance segmentation AP and object detection $\\mathrm{AP}^b$ . The results with a $\\dagger$ are copied from [7], while others are reproduced by us using their official released code." + ], + "table_footnote": [], + "table_body": "
ArchitectBackboneMethodAPAPrAPfAPcAPb
Mask-RCNN [14]R50-FPNCE18.70.416.529.319.7
RFS [13]23.714.222.929.324.7
BSCE [29]24.415.723.629.125.5
EQLv2 [33]25.217.424.129.926.0
ECM [7]†27.419.727.031.127.9
ECM27.219.626.631.327.7
SeeSaw [35]26.919.626.830.527.3
ROG [42]26.920.126.830.027.2
Our27.720.227.331.328.9
Mask-RCNN [14]R101-FPNCE [7]†25.516.624.530.626.6
EQLv2 [33]†27.220.625.931.427.9
ECM [7]†28.721.927.932.329.4
ECM28.620.928.332.229.3
SeeSaw [35]28.220.328.131.829.0
Our29.021.028.932.430.7
Cascade R-CNN [2]Swin-TECM [7]34.124.034.938.037.6
SeeSaw [35]34.224.634.737.837.8
Our34.624.735.338.138.2
Cascade R-CNN [2]Swin-BECM [7]†39.733.540.641.443.6
Our39.934.640.741.244.2
", + "bbox": [ + 259, + 226, + 736, + 554 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "branch' (CAB) with existing long-tailed object detection methods. Since our main focus is on bounding box regression, we list the object detection results in early columns and segmentation results in later ones. As shown in Table 3, using CAB leads to consistent $\\mathrm{AP}^b$ and AP improvement over existing classification-based methods, surpassing all of them with large margins. For object detection, Our CAB benefits the rare class a lot, with an increase of $4.6\\mathrm{AP}^b$ and $4.3\\mathrm{AP}^b$ on RFS and EQLv2. The same is true for instance segmentation, where a growing trend can be observed on all metrics, showing that CAB is also beneficial for later mask pixel predictions. Interestingly, the method 'RFS+CAB' (which uses a CE loss) can almost achieve the same object detection accuracy as the SeeSaw method, and surpasses EQLv2 for about $1\\mathrm{AP}^b$ . We thus conjecture that: besides merely focusing on classification, our regression methods can serve as strong alternatives that also strongly boost long-tailed detection accuracy.", + "bbox": [ + 212, + 593, + 787, + 790 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Comparison with SOTA. We then compare our method with state-of-the-art methods using different object detection framework (Mask-RCNN, Cascade R-CNN) and backbones (ResNet-50, ResNet-101, Swin-T and Swin-B). Note", + "bbox": [ + 212, + 794, + 787, + 840 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "Rectify the Regression Bias", + "bbox": [ + 545, + 114, + 730, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 114, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/434d8f4356dc88563624947052da4159241b439bcf73a490bd85615d33f4d328.jpg", + "table_caption": [ + "Table 5: More evaluation metrics in LVIS1.0 dataset with a 1x scheduler using Mask RCNN. AP ${}_{\\text{boundary }}^{fixed}$ and ${\\mathrm{{AP}}}_{\\text{box }}^{fixed}$ are two newly proposed challenging metrics [6,9]." + ], + "table_footnote": [], + "table_body": "
MethodAPAPbAPfixedboundaryAPfixedbbox
BAGS [24]23.123.7-26.2
EQLv2 [33]23.924.020.325.9
SeeSaw [35]25.225.419.826.5
ECM [7]26.326.721.427.4
Our26.928.022.128.2
", + "bbox": [ + 354, + 185, + 647, + 295 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "that 'Our' means 'SeeSaw + CAB'. For fair comparison, we reproduce majority of existing methods using their official released code unless specialized symbols $(e.g.,^{\\dagger})$ appears after a method's name. As can be seen in Table 4, our method achieves the overall highest accuracy in AP and $\\mathrm{AP}^b$ . For ResNet series models, our regression technique easily surpass the best competitor ECM [7], especially in $\\mathrm{AP}^b$ (an increase of $1.2\\mathrm{AP}^b$ for ResNet-50 and $1.1\\mathrm{AP}^b$ for ResNet-101). The advantage also holds for ViT-based object detectors, where we surpassed the best competitor ECM [7] with both Swin-Tiny and Swin-Base backbone architectures. Following LVIS's common practice, we didn't list $\\mathrm{AP}_r^b$ and $\\mathrm{AP}_c^b$ here, but we want to emphasize that the advantage of our regression methods can be further enlarged when more metrics are listed (cf. Table 3). This is also true if we replace 'adding a Class-Agnostic Branch' (CAB) with the 'merging heads' regression alternative (cf. the best accuracy in Table 2).", + "bbox": [ + 212, + 325, + 787, + 521 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.3 Generalization ability", + "text_level": 1, + "bbox": [ + 215, + 549, + 441, + 566 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "In this section, we will show the generalization ability of our regression methods in various aspects, including different evaluation metrics, datasets and its benefits on the mask branch.", + "bbox": [ + 212, + 579, + 785, + 622 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Different metrics. We first explore how different metrics affect our model's accuracy. Two additional metrics are $\\mathrm{AP}_{\\text {boundary }}$ (a more strict metric in instance segmentation [6]) and $\\mathrm{AP}_{\\text {bbox }}^{\\text {fixed }}$ (constraining 10,000 predicted bounding boxes per class across the dataset [9]). Object detector that obtain decent results in traditional metrics may not perform as well in these criterions. As shown in Table 5, our regression methods adapts well and surpasses all existing methods on both traditional and these challenging metrics.", + "bbox": [ + 212, + 625, + 787, + 733 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "The COCO-LT dataset. We also transfer our regression to another long-tailed dataset: COCO-LT. This is an artificially sampled [36] subset of original COCO [25]. We calculate the over bounding box metrics $\\mathrm{AP}^b$ and more fine-grained results: $\\mathrm{AP}_1^b$ , $\\mathrm{AP}_2^b$ , $\\mathrm{AP}_3^b$ and $\\mathrm{AP}_4^b$ (ranging from the rarest to the most frequent class). As shown in Table 6, our CAB leads to consistent improvements on all metrics (especially for the rarest categories $\\mathrm{AP}_1^b$ ) under different repeat-factor-sampling rates, showing the great power of CAB to help the rare", + "bbox": [ + 212, + 734, + 787, + 840 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "K. Zhu et al.", + "bbox": [ + 271, + 114, + 359, + 127 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/dd6040b02b5971ed0249a00111e944b1ca4c8343d051f7661fc501a38a64487d.jpg", + "table_caption": [ + "Table 6: Experiments on COCO-LT [36]. Following LVIS [12], we use a similar repeat-factor-sampling (RFS) strategy with different sampling rates (3e-3, 5e-3 and 1e-2)." + ], + "table_footnote": [], + "table_body": "
Rate +CABAP1bAP2bAP3bAP4bAPb
3e-3✓2.316.926.730.423.5
6.219.127.030.524.4
5e-3✓3.619.326.730.624.3
7.520.027.330.725.0
1e-2✓8.720.227.830.325.2
10.821.628.030.525.8
", + "bbox": [ + 364, + 184, + 632, + 313 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/8880ea6f3a05acb6c18f64dbe90cfc08e1a5d90ff6edcb7f171a5f41d22cae19.jpg", + "table_caption": [ + "Table 7: Experiments on the (relatively) balanced COCO dataset. We adopted the Faster RCNN [30] detector and tried three different backbones." + ], + "table_footnote": [], + "table_body": "
Backbone+CABAP\\( AP_{50} \\)\\( AP_{75} \\)\\( AP_s \\)\\( AP_m \\)\\( AP_l \\)
Res-50✓37.358.340.321.741.048.2
38.358.942.122.441.450.4
Res-101✓39.460.343.022.943.451.0
39.960.343.622.743.752.8
Res-32x4d✓41.062.244.623.945.352.9
41.762.545.423.645.954.8
", + "bbox": [ + 325, + 386, + 671, + 513 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "classes. In fact, this conclusion generally holds true for all sampling rates in our experiments. We only listed three here for simplicity and clarity.", + "bbox": [ + 212, + 563, + 782, + 593 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "On balanced training set. Furthermore, we validate how our method performs on the relatively balanced dataset MS-COCO2017. We adopted Faster RCNN with three different backbones (ResNet-50, ResNet-101 and ResNext101-32x4d). Results in Table 7 clearly show that CAB generalizes well to datasets with more balanced distributions. Interestingly, CAB boosts metrics of large and medium objects while shows similar accuracy on small objects. This is possibly because that small objects occupy over $60\\%$ of the total instances while large objects only consumes roughly $15\\%$ [32]. Since CAB is more beneficial to less-frequent classes, it brings higher gains in large objects than in small ones.", + "bbox": [ + 212, + 603, + 784, + 739 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "The mask branch. Finally, we apply our CAB to the segmentation branch to test whether adding an class-agnostic prior is suitable for mask prediction. Experimental results are in Table 8, where we add a class-agnostic mask prediction head and combines it with each class-specific mask head. As shown in the table, our CAB generalizes well to segmentation tasks. Our main experimental results may be further improved if CAB was applied in both box and mask predictions.", + "bbox": [ + 212, + 750, + 785, + 839 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "Rectify the Regression Bias", + "bbox": [ + 545, + 114, + 730, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/b7ef28d692341c326de78d12442b70a31acde8784292232536f677107f1ca702.jpg", + "table_caption": [ + "Table 8: Generalization to mask prediction. We added a class-agnostic mask prediction branch and combines it with each class-specific mask prediction results." + ], + "table_footnote": [], + "table_body": "
Branch +CABAPAPrAPbAPr
Box✓23.714.224.713.4
25.117.527.018.0
Mask✓23.714.224.713.4
24.114.625.013.9
", + "bbox": [ + 375, + 184, + 622, + 280 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "4.4 Discussion of our three remedies", + "text_level": 1, + "bbox": [ + 215, + 309, + 532, + 324 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Now we further discuss the proposed three methods (CAB, CH and MH) in detail, hoping to provide a more in-depth understanding of our hypothesis.", + "bbox": [ + 212, + 338, + 782, + 369 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Relationship of CAB to the objectness branch. At start, we want to clarify the differences between our CAB and the objectness branch approach [7, 35] adopted in the classification head. We have argued that although they seemingly share similar structures with ours, they are essentially different. First, the purpose of objectness branch is to deal with the imbalanced distribution between foreground and background classification samples [35], while our CAB aim to tackle the regression bias. Second, unlike classification where each class must preserve its own classifier, the regression heads can be clustered or merged, which provide more diverse solutions to reduce the regression bias besides CAB (cf. Table 2). Lastly, the effect of the objectness branch is yet to be proved because it will decrease the performance of rare classes even when combined with CE loss [35], while CAB leads to consistent improvements, especially in rare class.", + "bbox": [ + 212, + 369, + 784, + 549 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "How many clusters are the best? Attentive readers might have questions why larger clusters K (such as 100) in CH achieve better results in Table 2b for rare class, since enlarging K means the (useful) class-agnostic information is reduced. We thus empirically verify that: besides objects numbers per group, the scale variance could potentially affect the final AP. We test CH with 'object scales' in more K. For each K, we calculate standard deviation of scales in each group, average, and get representative scale std. The same is done for average object. As Fig. 3a shows: as K decreases, training objects per group increases, which benefit rare classes. But, the variation of scales increases a lot, which makes regression head learning more difficult—imagine one head to fit offsets that vary a lot, and we also observed higher training loss for smaller K. Hence, $\\max AP_r^b$ is not at $K = 1$ or 10, but at $K = 50$ in our experiments (cf. Fig. 3b).", + "bbox": [ + 212, + 551, + 785, + 733 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Why merging 'Common' class is the best?. We observe in Table 2c that merging common classes leads to the best improvement in $\\mathrm{AP}_r$ , which is somehow counter-intuitive. We conjecture that this is due to the partition shift of rare, common and frequent in LVIS train and validation sets. In the LVIS1.0 train dataset, the size ranges for frequent, common and rare are [0, 404], [405, 865] and [866, 1202], respectively. While for the validation set, they become [0, 212], [213, 536] and [537, 1202], respectively (cf. Fig. 1c). Hence, when we", + "bbox": [ + 212, + 734, + 785, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "K. Zhu et al.", + "bbox": [ + 271, + 114, + 359, + 127 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/84c82224482063638ab8a7c3ea99fe566a81d31ec2e85c041f63c656378608d4.jpg", + "image_caption": [ + "Fig. 3: The statistics (averaged standard deviation, averaged objects per group) of different K and its corresponding AP box." + ], + "image_footnote": [], + "bbox": [ + 330, + 145, + 491, + 256 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/3e8c2b949e449956d09858474ec0f3d9bcde517125df2e5d907997c6849105f2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 504, + 145, + 666, + 255 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/8a80c217608a3d3fdc3e34bd446db360da2a9eedb0590cf9bb416e8b5ee0a547.jpg", + "image_caption": [ + "Fig. 4: Visualizations of detection results before (in the left of each group) and after (in the right) using our CAB. We adopted RFS [12] as the baseline in LVIS1.0 and combine it with our CAB regression method. In comparison, the proposed method is good at detecting missing objects, filtering duplicated objects away, as well as rectifying bounding box predictions. This figure needs to be viewed in color." + ], + "image_footnote": [], + "bbox": [ + 225, + 310, + 779, + 411 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "use training set statistics to merge the common class, a large portion of rare categories in the validation set are also potentially merged, thus contributing to the validation improvement on both $\\mathrm{AP}_r$ and $\\mathrm{AP}_c$ .", + "bbox": [ + 214, + 523, + 787, + 568 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "4.5 Visualization and ablation", + "text_level": 1, + "bbox": [ + 215, + 595, + 478, + 609 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Flatten distribution. We first plot the RCNN regression loss of each category before and after combining with our regression CAB. The baseline methods we choose are EQLv2 and CE. As shown in Fig. 5, the regression loss for rare has seen a noticeable drop after adding our CAB, and the overall loss distribution has become much more balanced, too. We thus believe that the pipeline of our regression remedies can relieve the regression bias, further verifying that our hypothesis in Sec. 1 in indeed valid.", + "bbox": [ + 212, + 627, + 787, + 732 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "More accurate box/mask. We then calculate AP50-AP95 of boxes and masks to find whether our CAB behaves well on stricter IoU thresholds. As show in Fig. 5c-5d, adding CAB achieves consistent accuracy gains over all IoU thresholds for both box and mask prediction, and is especially helpful for those hard IoU threshold (e.g., AP75-90) in box evaluation (cf. Fig. 5c). Since a higher threshold requires more precise box prediction, these results have shown that our regression methods is capable of producing precise boxes with better quality.", + "bbox": [ + 212, + 734, + 787, + 840 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "Rectify the Regression Bias", + "bbox": [ + 545, + 114, + 730, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 784, + 126 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/9c1c9872b6b3fb04803c144e4903357762f5c4e2c4511321a81c709ffb11410c.jpg", + "image_caption": [ + "(a) w/o CAB" + ], + "image_footnote": [], + "bbox": [ + 227, + 143, + 356, + 218 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/af6b51a96e98acf6c4fc699c7eb46616b82bae858e2b38b01d02dc6d4fc08d62.jpg", + "image_caption": [ + "(b) w/CAB", + "Fig. 5: The loss distribution shift before and after combining with our CAB (Fig. 5a, 5b), and AP improvement of combining our CAB with baseline in LVIS (Fig. 5c, 5d)." + ], + "image_footnote": [], + "bbox": [ + 367, + 145, + 496, + 218 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/66697af2349b380f646825e9f6a47b6feb50d1bb2a16c978d59c70f5e58c5096.jpg", + "image_caption": [ + "(c) box delta" + ], + "image_footnote": [], + "bbox": [ + 508, + 145, + 635, + 218 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/3b99aeb85f883781aa25c1e924f5c2ca04c72c54ec401407a61c7a055c18345f.jpg", + "image_caption": [ + "(d) mask delta" + ], + "image_footnote": [], + "bbox": [ + 648, + 145, + 776, + 218 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Qualitative results. Last but not least, we provide example detection images sampled from the LVIS1.0 validation dataset. For simplicity, we choose RFS [12] with CE loss as baseline and combine it with our CAB method. As shown in Fig. 4, both baseline and our CAB detect most of the objects in an image, but CAB generally caters for more details. For example, CAB can help discover missed boxes, like the cabinet in the lower left images. It is also clearly illustrated that CAB helps filter duplicate boxes away in the final predictions (e.g., the pictures with elephant and cow with grass background). Since CAB brings in better predictions, duplicate boxes will have larger overlap that may be suppressed by NMS (non-maximum suppression). If we zoom this figure (cf. the last column in Fig. 4), we will find that CAB is capable of rectifying the predicted boxes, and this can empirically explain why our CAB achieves much better AP under higher and more difficult IoU thresholds (cf. Fig. 5c).", + "bbox": [ + 212, + 306, + 787, + 503 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Conclusions and Limitations", + "text_level": 1, + "bbox": [ + 215, + 530, + 527, + 546 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In this paper, we discovered that the regression bias (imbalanced regression loss distribution on the RCNN head) exists in long-tailed object detection, and adversely affects detection results. We thus proposed three remedies for rectifying the regression bias. The proposed method significantly boosts the performance of rare class $\\mathrm{AP}^b$ , and achieves state-of-the-art results. We also generalize our regression methods to balanced dataset, different evaluation metrics and the mask branch. Finally, visualizations show that our method indeed produces better predicted bounding boxes.", + "bbox": [ + 212, + 566, + 787, + 686 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "As for the limitations, it remains unclear why the boosted accuracy becomes lower when the baseline are higher (e.g., the improvement of CAB on ECM is lower than that on RFS, cf. Table 3). This may relate to the upper bound a backbone model can achieve. Since large network generally better fit the dataset (cf. Table 4), the performance a ResNet-50 can achieve is limited. We thus call for involving larger vision models on the LVIS dataset to better handle the long-tailed object detection problem. Besides involving stronger vision backbones, we will try to testify the adaptability of the proposed methods when different pretraining information is integrated into the backbone network (e.g. self-supervised learning), which is left as our future work.", + "bbox": [ + 212, + 689, + 787, + 840 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "K. Zhu et al.", + "bbox": [ + 271, + 114, + 359, + 127 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 217, + 143, + 401, + 162 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "We thank Yin-Yin He (M.Sc. degree in Nanjing University, now in ByteDance) for providing helpful discussions in long-tailed object detection.", + "bbox": [ + 215, + 176, + 782, + 205 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 229, + 321, + 244 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Alexandridis, K.P., Deng, J., Nguyen, A., Luo, S.: Long-tailed instance segmentation using gumbel optimized loss. In: ECCV. Lecture Notes in Computer Science, vol. 13670, pp. 353-369. Springer (2022)", + "2. Cai, Z., Vasconcelos, N.: Cascade r-cnn: Delving into high quality object detection. In: CVPR. pp. 6154-6162 (2018)", + "3. Cao, K., Wei, C., Gaidon, A., Arechiga, N., Ma, T.: Learning imbalanced datasets with label-distribution-aware margin loss. In: NeurIPS. pp. 1565-1576 (2019)", + "4. Caron, M., Bojanowski, P., Joulin, A., Douze, M.: Deep clustering for unsupervised learning of visual features. In: Ferrari, V., Hebert, M., Sminchisescu, C., Weiss, Y. (eds.) ECCV. Lecture Notes in Computer Science, vol. 11218, pp. 139-156. Springer (2018)", + "5. Chen, K., Wang, J., Pang, J., Cao, Y., Xiong, Y., Li, X., Sun, S., Feng, W., Liu, Z., Xu, J., Zhang, Z., Cheng, D., Zhu, C., Cheng, T., Zhao, Q., Li, B., Lu, X., Zhu, R., Wu, Y., Dai, J., Wang, J., Shi, J., Ouyang, W., Loy, C.C., Lin, D.: MMDetection: Open mmlab detection toolbox and benchmark. arXiv preprint arXiv:1906.07155 (2019)", + "6. Cheng, B., Girshick, R., Dollar, P., Berg, A.C., Kirillov, A.: Boundary iou: Improving object-centric image segmentation evaluation. In: CVPR. pp. 15334-15342 (2021)", + "7. Cho, J.H., Krahenbuhl, P.: Long-tail detection with effective class-margins. In: ECCV. Lecture Notes in Computer Science, vol. 13668, pp. 698-714. Springer (2022)", + "8. Cui, Y., Jia, M., Lin, T.Y., Song, Y., Belongie, S.: Class-balanced loss based on effective number of samples. In: CVPR. pp. 9268-9277 (2019)", + "9. Dave, A., Dóllár, P., Ramanan, D., Kirillov, A., Girshick, R.: Evaluating large-vocabulary object detectors: The devil is in the details. arXiv preprint arXiv:2102.01066 (2021)", + "10. Du, Y., Wu, J.: No one left behind: Improving the worst categories in long-tailed learning. In: CVPR. pp. 15804-15813 (2023)", + "1. Feng, C., Zhong, Y., Huang, W.: Exploring classification equilibrium in long-tailed object detection. In: ICCV. pp. 3417-3426 (2021)", + "2. Gupta, A., Dollar, P., Girshick, R.: Lvis: A dataset for large vocabulary instance segmentation. In: CVPR. pp. 5356-5364 (2019)", + "3. Gupta, A., Dollar, P., Girshick, R.: Lvis: A dataset for large vocabulary instance segmentation. In: CVPR. pp. 5356-5364 (2019)", + "4. He, K., Gkioxari, G., Dollár, P., Girshick, R.: Mask R-CNN. In: ICCV. pp. 2961-2969 (2017)", + "5. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 770-778 (2016)", + "6. He, Y.Y., Wu, J., Wei, X.S.: Distilling virtual examples for long-tailed recognition. In: ICCV. pp. 235-244 (2021)" + ], + "bbox": [ + 225, + 260, + 784, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "Rectify the Regression Bias", + "bbox": [ + 545, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "17. He, Y.Y., Zhang, P., Wei, X.S., Zhang, X., Sun, J.: Relieving long-tailed instance segmentation via pairwise class balance. In: CVPR. pp. 7000-7009 (2022)", + "18. Hsieh, T., Robb, E., Chen, H., Huang, J.: Droploss for long-tail instance segmentation. In: AAAI. pp. 1549-1557 (2021)", + "19. Hu, X., Jiang, Y., Tang, K., Chen, J., Miao, C., Zhang, H.: Learning to segment the tail. In: CVPR (2020)", + "20. Kang, B., Xie, S., Rohrbach, M., Yan, Z., Gordo, A., Feng, J., Kalantidis, Y.: Decoupling representation and classifier for long-tailed recognition. In: ICLR (2020)", + "21. Li, B., Yao, Y., Tan, J., Zhang, G., Yu, F., Lu, J., Luo, Y.: Equalized focal loss for dense long-tailed object detection. In: CVPR. pp. 6990-6999 (2022)", + "22. Li, M., Cheung, Y.m., Lu, Y.: Long-tailed visual recognition via gaussian clouded logit adjustment. In: CVPR. pp. 6929-6938 (2022)", + "23. Li, T., Wang, L., Wu, G.: Self supervision to distillation for long-tailed visual recognition. In: ICCV. pp. 630-639 (2021)", + "24. Li, Y., Wang, T., Kang, B., Tang, S., Wang, C., Li, J., Feng, J.: Overcoming classifier imbalance for long-tail object detection with balanced group softmax. In: CVPR (2020)", + "25. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dólár, P., Zitnick, C.L.: Microsoft COCO: Common objects in context. In: ECCV. Lecture Notes in Computer Science, vol. 8693, pp. 740-755. Springer (2014)", + "26. Liu, Z., Miao, Z., Zhan, X., Wang, J., Gong, B., Yu, S.X.: Large-scale long-tailed recognition in an open world. In: CVPR. pp. 2537-2546 (2019)", + "27. Menon, A.K., Jayasumana, S., Rawat, A.S., Jain, H., Veit, A., Kumar, S.: Long-tail learning via logit adjustment. In: ICLR (2021)", + "28. Pang, J., Chen, K., Shi, J., Feng, H., Ouyang, W., Lin, D.: Libra r-cnn: Towards balanced learning for object detection. In: CVPR. pp. 821-830 (2019)", + "29. Ren, J., Yu, C., sheng, s., Ma, X., Zhao, H., Yi, S., Li, h.: Balanced meta-softmax for long-tailed visual recognition. In: NeurIPS. pp. 4175-4186 (2020)", + "30. Ren, S., He, K., Girshick, R., Sun, J.: Faster R-CNN: Towards real-time object detection with region proposal networks. In: NeurIPS. pp. 91-99 (2015)", + "31. Russakovsky, O., Deng, J., Su, H., Krause, J., Satheesh, S., Ma, S., Huang, Z., Karpathy, A., Khosla, A., Bernstein, M., Berg, A., Fei-Fei, L.: ImageNet large scale visual recognition challenge. International Journal of Computer Vision 115(3), 211-252 (2015)", + "32. Singh, B., Davis, L.S.: An analysis of scale invariance in object detection SNIP. In: CVPR. pp. 3578-3587 (2018)", + "33. Tan, J., Lu, X., Zhang, G., Yin, C., Li, Q.: Equalization loss v2: A new gradient balance approach for long-tailed object detection. In: CVPR. pp. 1685-1694 (2021)", + "34. Tan, J., Wang, C., Li, B., Li, Q., Ouyang, W., Yin, C., Yan, J.: Equalization loss for long-tailed object recognition. In: CVPR (2020)", + "35. Wang, J., Zhang, W., Zang, Y., Cao, Y., Pang, J., Gong, T., Chen, K., Liu, Z., Loy, C.C., Lin, D.: Seesaw loss for long-tailed instance segmentation. In: CVPR. pp. 9695-9704 (2021)", + "36. Wang, T., Li, Y., Kang, B., Li, J., Liew, J.H., Tang, S., Hoi, S.C.H., Feng, J.: The devil is in classification: A simple framework for long-tail instance segmentation. In: ECCV. Lecture Notes in Computer Science, vol. 12359, pp. 728-744. Springer (2020)", + "37. Wang, T., Zhu, Y., Zhao, C., Zeng, W., Wang, J., Tang, M.: Adaptive class suppression loss for long-tail object detection. In: CVPR. pp. 3103-3112 (2021)", + "38. Wang, X., Lian, L., Miao, Z., Liu, Z., Yu, S.: Long-tailed recognition by routing diverse distribution-aware experts. In: ICLR (2021)" + ], + "bbox": [ + 215, + 146, + 784, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "K. Zhu et al.", + "bbox": [ + 271, + 114, + 357, + 126 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "39. Wu, J., Song, L., Wang, T., Zhang, Q., Yuan, J.: Forest R-CNN: large-vocabulary long-tailed object detection and instance segmentation. In: ACM MM. pp. 1570-1578 (2020)", + "40. Xuan, G., Zhang, W., Chai, P.: EM algorithms of gaussian mixture model and hidden markov model. In: ICIP. pp. 145-148 (2001)", + "41. Zhang, C., Pan, T.Y., Li, Y., Hu, H., Xuan, D., Changpinyo, S., Gong, B., Chao, W.L.: Mosaicos: A simple and effective use of object-centric images for long-tailed object detection. In: ICCV. pp. 417-427 (2021)", + "42. Zhang, S., Chen, C., Peng, S.: Reconciling object-level and global-level objectives for long-tail detection. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 18982-18992 (2023)", + "43. Zhong, Z., Cui, J., Liu, S., Jia, J.: Improving calibration for long-tailed recognition. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 16489-16498 (2021)", + "44. Zhou, B., Cui, Q., Wei, X.S., Chen, Z.M.: Bbn: Bilateral-branch network with cumulative learning for long-tailed visual recognition. In: CVPR. pp. 9716-9725 (2020)", + "45. Zhu, K., Fu, M., Wu, J.: Multi-label self-supervised learning with scene images. In: ICCV. pp. 6694-6703 (2023)", + "46. Zhu, K., He, Y., Wu, J.: Quantized feature distillation for network quantization. In: AAAI. pp. 11452-11460 (2023)" + ], + "bbox": [ + 212, + 146, + 787, + 438 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "Rectify the Regression Bias", + "bbox": [ + 545, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + } +] \ No newline at end of file diff --git a/2024/Rectify the Regression Bias in Long-Tailed Object Detection/07bdb2f4-ff12-4767-80ed-dec2d3b31ce7_model.json b/2024/Rectify the Regression Bias in Long-Tailed Object Detection/07bdb2f4-ff12-4767-80ed-dec2d3b31ce7_model.json new file mode 100644 index 0000000000000000000000000000000000000000..107760e846f4afa479f2aedd218269c1166f92f4 --- /dev/null +++ b/2024/Rectify the Regression Bias in Long-Tailed Object Detection/07bdb2f4-ff12-4767-80ed-dec2d3b31ce7_model.json @@ -0,0 +1,2346 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.258, + 0.141, + 0.747, + 0.187 + ], + "angle": 0, + "content": "Rectify the Regression Bias in Long-Tailed Object Detection" + }, + { + "type": "text", + "bbox": [ + 0.251, + 0.213, + 0.751, + 0.229 + ], + "angle": 0, + "content": "Ke Zhu, Minghao Fu, Jie Shao, Tianyu Liu, and Jianxin Wu" + }, + { + "type": "text", + "bbox": [ + 0.324, + 0.241, + 0.678, + 0.255 + ], + "angle": 0, + "content": "State Key Laboratory for Novel Software Technology" + }, + { + "type": "text", + "bbox": [ + 0.383, + 0.256, + 0.619, + 0.269 + ], + "angle": 0, + "content": "Nanjing University, Nanjing, China" + }, + { + "type": "text", + "bbox": [ + 0.308, + 0.27, + 0.694, + 0.282 + ], + "angle": 0, + "content": "School of Artificial Intelligence, Nanjing University, China" + }, + { + "type": "text", + "bbox": [ + 0.255, + 0.284, + 0.748, + 0.297 + ], + "angle": 0, + "content": "{zhuk, fumh, shaoj, liuty}@lamda.nju.edu.cn, wujx2001@nju.edu.cn" + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.338, + 0.741, + 0.573 + ], + "angle": 0, + "content": "Abstract. Long-tailed object detection faces great challenges because of its extremely imbalanced class distribution. Recent methods mainly focus on the classification bias and its loss function design, while ignoring the subtle influence of the regression branch. This paper shows that the regression bias exists and does adversely and seriously impact the detection accuracy. While existing methods fail to handle the regression bias, the class-specific regression head for rare classes is hypothesized to be the main cause of it in this paper. As a result, three kinds of viable solutions to cater for the rare categories are proposed, including adding a class-agnostic branch, clustering heads and merging heads. The proposed methods brings in consistent and significant improvements over existing long-tailed detection methods, especially in rare and common classes. The proposed method achieves state-of-the-art performance in the large vocabulary LVIS dataset with different backbones and architectures. It generalizes well to more difficult evaluation metrics, relatively balanced datasets, and the mask branch. This is the first attempt to reveal and explore rectifying of the regression bias in long-tailed object detection." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.601, + 0.377, + 0.617 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.633, + 0.788, + 0.724 + ], + "angle": 0, + "content": "Long-tailed object detection draws great attention [12] for its practical utility recently. Numerous efforts [1,7,34,35,37] have been made to tackle this challenging task, such as re-weighting [33,35,37], over-sampling [11,12,39], and balanced grouping [19,24]. These methods are proposed to prevent the tail classes from being overwhelmed due to discouraging gradients [33, 37], inferior predicting scores [11,24] or insufficient samples [11,28]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.725, + 0.788, + 0.817 + ], + "angle": 0, + "content": "Long-tailed detection often involves both classification and regression heads. While almost all existing methods focus on mitigating the classification bias (e.g., adjusting the classification structure in detection branches), little or no attention has been paid to the regression branch. We will show in this paper that the regression bias has significant adverse effects in long-tailed object detection, but previous methods failed to identify this important issue." + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.825, + 0.465, + 0.84 + ], + "angle": 0, + "content": "* J. Wu is the corresponding author" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.128 + ], + "angle": 0, + "content": "K. Zhu et al." + }, + { + "type": "image", + "bbox": [ + 0.235, + 0.145, + 0.379, + 0.233 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.265, + 0.236, + 0.354, + 0.248 + ], + "angle": 0, + "content": "(a) RCNN loss" + }, + { + "type": "image", + "bbox": [ + 0.387, + 0.144, + 0.532, + 0.234 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.423, + 0.236, + 0.503, + 0.248 + ], + "angle": 0, + "content": "(b) RPN loss" + }, + { + "type": "image", + "bbox": [ + 0.541, + 0.146, + 0.77, + 0.233 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.541, + 0.235, + 0.769, + 0.248 + ], + "angle": 0, + "content": "(c) Mean scale in LVIS train and val set" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.259, + 0.788, + 0.314 + ], + "angle": 0, + "content": "Fig. 1: 1a shows the RCNN regression train loss of frequent, common and rare categories. 1b shows the RPN regression loss. 1c shows distribution of per class mean object scales in LVIS1.0. 'delta' in 1c is the negative of difference between train and validation set sizes for different classes." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.349, + 0.788, + 0.47 + ], + "angle": 0, + "content": "Fig. 1 clearly showcases this issue. For different methods (EQLv2 [33], See-Saw [35], ECM [7] and CrossEntropy CE [12]) trained on the LVIS1.0 [12] dataset, we plot the regression branch's losses for the final detection RCNN head in Fig. 1a. It is clear that the regression losses of rare categories are significantly higher than those of frequent and common categories, which inevitably will lead to poor regression results (and hence detection results) for rare classes. We name this finding as the regression bias, but existing long-tailed detection methods all fail to deal with or even identify this issue." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.472, + 0.788, + 0.578 + ], + "angle": 0, + "content": "To further demonstrate the importance of regression, we calculate the class-wise mean scale of GT (groundtruth) boxes in LVIS train and validation sets, as well as their differences (cf. Fig. 1c). The scale shift of rare classes is much larger than that of frequent classes. Since regression is highly correlated with box scale [30], it is thus inherently difficult for a rare class to learn a good bounding box regressor with both few samples and large scale shift. In short, it is crucial to properly handle the regression bias in long-tailed object detection." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.58, + 0.788, + 0.641 + ], + "angle": 0, + "content": "Our solution to rectify the regression bias is motivated by Fig. 1b. We find that the regression loss in RPN is balanced where rare, common and frequent categories have almost the same regression loss, which is almost immune to the regression bias when comparing Fig. 1b with Fig. 1a." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.642, + 0.788, + 0.717 + ], + "angle": 0, + "content": "The key difference between RPN and RCNN regression is that the former is class-agnostic (i.e., all classes share the same regression parameters), while the latter is class-specific. Then, one natural question arises: Can class-agnostic head improve a tail class's generalization ability and handle the regression bias? Our hypothesis is that rare classes do favor class-agnostic regression heads." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.788, + 0.843 + ], + "angle": 0, + "content": "Our hypothesis is supported by experiments in Table 1, where we compared class-specific and -agnostic regression heads in RCNN. By replacing the classification head with groundtruth class labels, experiments in Table 1 disentangled the impact of the classification head and focus on the regression heads. It is clear that the class-agnostic head possesses substantial advantages: \\(\\mathrm{AP}_r^b\\) (for rare) surges from 0.7 to 54.6, even surpassing \\(\\mathrm{AP}_f^b\\) (for frequent classes)! However, the agnostic head will bring a small drop in frequent classes (e.g., from 40.7 to 40.0 in \\(\\mathrm{AP}_f^b\\)). The final version of our conjecture is: the rare (and possibly" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.546, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Rectify the Regression Bias" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.193 + ], + "angle": 0, + "content": "common) classes indeed favor class-agnostic regression, while the frequent classes prefers a class-specific regression, and there should be a trade-off between the two to optimize for all three types of categories." + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.219, + 0.788, + 0.275 + ], + "angle": 0, + "content": "Table 1: Results on LVIS1.0 based on the CE baseline. We trained a Mask-RCNN R50-FPN detector [14] with class-specific (the default setting) and class-agnostic regression heads in the RCNN head. Please note that during test, the predicted proposals were provided groundtruth (GT) classification results." + }, + { + "type": "table", + "bbox": [ + 0.359, + 0.288, + 0.642, + 0.382 + ], + "angle": 0, + "content": "
Reg. HeadGTAPAPbAPrAPpcAPfb
Specificno18.719.70.316.531.7
yes37.239.834.141.340.7
Agnosticno18.019.30.715.531.5
yes38.445.654.647.040.0
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.411, + 0.788, + 0.607 + ], + "angle": 0, + "content": "Accordingly, we design three different methods to fully rectify the regression bias, including adding a class-agnostic head, clustering similar heads, or merging heads. All three methods bring in positive effects (cf. Table 2), which verifies that rectifying regression bias is indeed crucial. We choose to adopt 'adding a class-agnostic head' in our main experiment for its simplicity, which leads to consistent and significant improvements over previous long-tailed detection pipeline, and has achieved state-of-the-art performance with various backbones and architectures. Moreover, our method shows robust generalization ability (cf. Tables 5-8) under varying settings, including different datasets (COCO/COCOLT [36]), different evaluation metrics and even adapts to the mask branch design. Furthermore, visualizations show that the proposed method leads to more accurate bounding box predictions (cf. Fig. 4) and indeed alleviates the regression bias (cf. Fig. 5). In summary, our contributions are:" + }, + { + "type": "text", + "bbox": [ + 0.222, + 0.617, + 0.785, + 0.645 + ], + "angle": 0, + "content": "1. For the first time, we reveal and successfully handle the regression bias in long-tailed object detection." + }, + { + "type": "text", + "bbox": [ + 0.222, + 0.648, + 0.785, + 0.677 + ], + "angle": 0, + "content": "2. We propose three remedies to alleviate this bias, all of which produce consistent improvements over existing methods." + }, + { + "type": "text", + "bbox": [ + 0.222, + 0.678, + 0.785, + 0.723 + ], + "angle": 0, + "content": "3. Our method achieves state-of-the-art results on LVIS as well as generalizing across datasets, metrics and even the mask branch. Visualizations qualitatively verify our hypothesis, too." + }, + { + "type": "list", + "bbox": [ + 0.222, + 0.617, + 0.785, + 0.723 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.747, + 0.388, + 0.763 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Vast number of pipelines and techniques have been invented in the field of general purpose object detection. Although most of these detectors can possibly be adapted to balanced detection datasets [25], they can nevertheless hardly handle extremely long-tailed distribution. Dedicated methods are hence required." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.128 + ], + "angle": 0, + "content": "K. Zhu et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.177 + ], + "angle": 0, + "content": "Modern long-tailed learning methods aim to solve two important tasks: image classification [31] and object detection/segmentation [12, 30]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.178, + 0.787, + 0.373 + ], + "angle": 0, + "content": "The first task, long-tailed image classification [26], is well explored and displays great diversity in terms of specific methods [3,29], training pipelines [20,44], post-hoc analysis [22,27] and network architecture [16,23,38]. Early attempts in this family focused on the details of re-weighting [3] or re-sampling [8] techniques to provide a relatively balanced (data) distribution for tail classes. Later on, the decoupling pipeline [20] for long-tailed learning becomes popular, paving the way for various successors [43,44]. This decoupled paradigm are based on the insight that instance sampling is beneficial for representation, while the classifier needs to be calibrated to alleviate its bias. Some works try to unify long-tailed learning with theoretical analysis [22,27], which involves label distribution [29] and generalization bound [27]. Most recent methods are much more diverse in ideas: they involve knowledge distillation [16,38,46], self-supervised learning [23,45] or statistical approach [10] to better handle long-tailed recognition." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.374, + 0.788, + 0.553 + ], + "angle": 0, + "content": "The combination of long-tailed learning and object detection or instance segmentation proves to be more challenging [12], because the class distribution in LVIS [12] is extremely imbalanced and naively applying common object detection techniques leads to un-satisfactory results [18, 24, 34, 35, 37]. Dominating solutions in this area are re-sampling [12, 39] and re-weighting [17, 21, 33, 35, 37]. There are variants of them that adopt balanced grouping [24], class incremental learning [19], augmented feature sampling [11, 36] or extra data source [41]. A common characteristic of these long-tailed detection methods is that they only focus on the classification task, and ignore the subtle influence of the regression branch. In this paper, we thoroughly explore the previously unvisited regression bias in long-tailed object detection, and propose a simple yet novel method to tackle long-tailed object detection." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.555, + 0.788, + 0.646 + ], + "angle": 0, + "content": "In [36], there is an experiment with similar design as ours in Table 1. But, we emphasize that they are fundamentally different. On one hand, conclusions in [36] is that 'performance drop in LVIS is mainly caused by the proposal classification [36]' and hence they focused on classification. On the other hand, their class agnostic-head is a supplementary to support their classification hypothesis, while our design and experiments lead to a novel finding: the regression bias." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.669, + 0.332, + 0.685 + ], + "angle": 0, + "content": "3 Method" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.702, + 0.787, + 0.732 + ], + "angle": 0, + "content": "Now we elaborate three different remedies to alleviate the regression bias, and start from the background on long-tailed object detection." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.755, + 0.374, + 0.769 + ], + "angle": 0, + "content": "3.1 Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.842 + ], + "angle": 0, + "content": "We take Faster-RCNN [14] as an example. For a scene image \\(I\\), it is first fed into a backbone network \\(\\phi(\\cdot)\\) (e.g., ResNet [15]) to get the image feature \\(f \\in \\mathbb{R}^{d \\times w \\times h}\\): \\(f = \\phi(I)\\), with dimensionality, width and height denoted as \\(d\\), \\(w\\), \\(h\\), respectively. A region proposal network (RPN) [30] that contains both agnostic" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.547, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "Rectify the Regression Bias" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "image", + "bbox": [ + 0.235, + 0.143, + 0.77, + 0.269 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.28, + 0.788, + 0.39 + ], + "angle": 0, + "content": "Fig. 2: Illustration of the regular two-stage detection pipeline and the proposed regression methods. Previous methods (the left figure) generally focus on the final classification branch (the yellow arrow), while we focus on rectifying the regression bias (the red arrow). The right part shows our three regression methods, including adding an extra branch \\( W_{0} \\), clustering regression heads (e.g., from \\( W_{1},\\ldots ,W_{C} \\) to \\( W_1',W_2' \\)) and merging (e.g., merging rare categories into one regression head \\( W_{\\text{rare}} \\), cf. Table 2). In our main experiments, we choose 'adding an extra branch' for its simplicity. This figure needs to be viewed in color." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.422, + 0.785, + 0.467 + ], + "angle": 0, + "content": "classification and regression branches is then applied on the feature tensor \\( f \\) to produce proposals \\( p \\) from pre-defined anchors. The ROAlign [14] then extracts proposal features:" + }, + { + "type": "equation", + "bbox": [ + 0.365, + 0.48, + 0.786, + 0.498 + ], + "angle": 0, + "content": "\\[\np = \\operatorname {R P N} (f), \\quad f _ {p} = \\operatorname {R O I A l i g n} (f, p), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.508, + 0.788, + 0.584 + ], + "angle": 0, + "content": "where \\( p \\) represents a large set of proposals, and \\( f_{p} \\) is the set of aligned proposal features. \\( f_{p} \\) goes into post-process modules (e.g., NMS [30]) before being sent to the RCNN head to get the final features set \\( f_{n} \\): \\( f_{n} = \\mathrm{RCNN}(f_{p}) \\), which is fed to classification and regression branches (e.g., linear layers) to produce the prediction results." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.608, + 0.611, + 0.623 + ], + "angle": 0, + "content": "3.2 Our three remedies for the regression bias" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.635, + 0.706, + 0.65 + ], + "angle": 0, + "content": "In Faster RCNN, there is a dedicated regression head for each class:" + }, + { + "type": "equation", + "bbox": [ + 0.392, + 0.661, + 0.787, + 0.68 + ], + "angle": 0, + "content": "\\[\nr _ {i} = W _ {i} ^ {T} f _ {n}, \\quad i = 1, 2, \\dots , C, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.691, + 0.785, + 0.751 + ], + "angle": 0, + "content": "where \\( r_i = (\\delta x_i, \\delta y_i, \\delta w_i, \\delta h_i) \\) represents the regression offset for class \\( i \\), and \\( W_i \\) is the class-specific regression head (a linear layer). We will now present our approaches to rectify the regression bias. The existing pipeline and our proposed methods are shown in Fig. 2." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.752, + 0.788, + 0.813 + ], + "angle": 0, + "content": "Extra class-agnostic branch. This is a simple remedy to cope with the regression bias. Since rare classes favor a class-agnostic head while class-specific ones are slightly preferable to frequent classes in our hypothesis, we propose to use both heads. For class \\( i \\), its regression head changes to:" + }, + { + "type": "equation", + "bbox": [ + 0.41, + 0.824, + 0.787, + 0.843 + ], + "angle": 0, + "content": "\\[\nW _ {i} ^ {\\prime} = \\alpha W _ {0} + (1 - \\alpha) W _ {i}, \\tag {3}\n\\]" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.128 + ], + "angle": 0, + "content": "K. Zhu et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.145, + 0.788, + 0.271 + ], + "angle": 0, + "content": "Table 2: Our three different methods to alleviate the regression bias. In 2a, we add an extra class-agnostic regression head during training and testing, where the shared class-agonistic head and the class-specific head are combined using a weight \\(\\alpha\\). In 2b, we perform regression head clustering according to the number of instances ('\\(n'\\)) or mean box scale ('\\(s'\\)). In 2c, we directly merge the regression heads of rare, common or frequent classes. The original class-specific heads in 2b and 2c are replaced by the new heads. Note that 'base' means the baseline method, i.e., with only class-specific regression heads. In 2a, \\(\\alpha = 0\\) is equivalent to the baseline method, and \\(\\alpha = 1.0\\) is equivalent to merging all \\(r, c, f\\) heads into a single head (i.e., the last row in 2c)." + }, + { + "type": "table", + "bbox": [ + 0.229, + 0.281, + 0.787, + 0.384 + ], + "angle": 0, + "content": "
(α) APAPrAPbAPrknsAPAPrAPbAPrmerge APAPrAPbAPr
0.023.714.224.713.4base23.714.224.713.4base23.714.224.713.4
0.224.115.825.415.110✓24.514.526.314.4r25.116.326.716.7
0.525.117.527.018.010✓24.212.826.012.6c25.517.727.217.2
0.824.417.025.916.4100✓24.413.026.212.9r, c25.317.127.318.2
1.024.716.726.718.3100✓25.216.726.916.7r, c, f24.716.726.718.3
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.231, + 0.385, + 0.38, + 0.397 + ], + "angle": 0, + "content": "(a) class-agnostic head" + }, + { + "type": "table_footnote", + "bbox": [ + 0.427, + 0.385, + 0.562, + 0.397 + ], + "angle": 0, + "content": "(b) Clustering heads" + }, + { + "type": "table_footnote", + "bbox": [ + 0.63, + 0.385, + 0.751, + 0.397 + ], + "angle": 0, + "content": "(c) Merging heads" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.437, + 0.785, + 0.512 + ], + "angle": 0, + "content": "where \\( W_0 \\) is a shared class-agnostic regression head for all classes, and \\( \\alpha \\) is a hyper-parameter to balance class-agnostic and class-specific heads. We empirically find that this simple change leads to consistent improvements over the default class-specific regression head (cf. Table 2a), while simply setting \\( \\alpha = 0.5 \\) gives the optimal trade-off." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.513, + 0.788, + 0.603 + ], + "angle": 0, + "content": "Clustering heads. This method is motivated by the analysis in Fig. 1c. Since some categories have similar statistics, we can assign them a shared regression head to improve the generalization ability. We implement it by following three steps: sorting, grouping and assigning. First, we use the number of instance or mean box scale to sort the original categories in descending order (\\( C = 1203 \\) in LVIS):" + }, + { + "type": "equation", + "bbox": [ + 0.395, + 0.604, + 0.785, + 0.623 + ], + "angle": 0, + "content": "\\[\nW _ {1}, \\dots , W _ {C} \\stackrel {\\text {s o r t}} {\\Longrightarrow} W _ {1} ^ {s}, \\dots , W _ {C} ^ {s}. \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.628, + 0.788, + 0.676 + ], + "angle": 0, + "content": "Then, we cluster them into \\( K \\) groups. During clustering, we do not rely on time-cost algorithms such as K-means [4] or GMM [40]), but simply put adjacent classes into one group, and each group has the same number of classes \\( N = \\frac{C}{K} \\):" + }, + { + "type": "equation", + "bbox": [ + 0.357, + 0.684, + 0.785, + 0.703 + ], + "angle": 0, + "content": "\\[\n\\{(W _ {N \\times i + 1} ^ {s}, W _ {N \\times i + 2} ^ {s}, \\dots , W _ {N \\times i + N} ^ {s}) \\} _ {i = 0} ^ {K - 1}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.71, + 0.561, + 0.726 + ], + "angle": 0, + "content": "Finally, each group \\(i\\) share one regression head:" + }, + { + "type": "equation", + "bbox": [ + 0.339, + 0.737, + 0.785, + 0.757 + ], + "angle": 0, + "content": "\\[\nW _ {g i} \\stackrel {\\text {r e p l a c e}} {\\Longleftarrow} \\left(W _ {N \\times i + 1} ^ {s}, W _ {N \\times i + 2} ^ {s}, \\dots , W _ {N \\times i + N} ^ {s}\\right). \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.788, + 0.81 + ], + "angle": 0, + "content": "These shared regression matrix are then used for both training and testing. As shown in Table 2b, clustering heads with similar scale statistics brings robust improvements over the baseline methods (see more discussion in Sec. 4.4)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.811, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Merge heads. This approach has similar motivation to the previous one, which clusters regression heads, but is even more straightforward. We simply" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.547, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Rectify the Regression Bias" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.268 + ], + "angle": 0, + "content": "group the regression heads in pre-defined clusters. For example, we let all rare categories share a common regression head \\( W_{\\text{rare}} \\), and the same for common and frequent classes. Experimental results can be found in Table 2c, where we try four different combinations (note that \\( r, c \\) means merging rare and common classes all together into one regression head). The results indicate that merging always leads to performance gains, especially for the rare categories. The most significant improvements come from merging only the common class (i.e., the row denoted as 'c' in Table 2c, also cf. Sec. 4.4 for discussion on partition shift)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.269, + 0.788, + 0.359 + ], + "angle": 0, + "content": "All our proposed remedies improve the accuracy of long-tailed object detection and instance segmentation, which verifies the importance of rectifying the regression bias. We will show that all these methods lead to consistent and nontrivial margin over various existing approaches (cf. Table 3). We then choose adding a class-agnostic branch (the first approach, cf. Table 2a) to conduct the rest of our experiments for its simplicity." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.36, + 0.788, + 0.48 + ], + "angle": 0, + "content": "Attentive readers might advocate that all proposed three methods are to be verified on all datasets. However, we argue that this is not necessary and our choice has its advantages. On one hand, it does not need any dataset statistics, which is especially appreciated when the number of categories and the data distribution is unknown. On the other hand, by combining both class-agnostic and class-specific heads, it fully exploits object priors and per class knowledge: each type of head has its own merit, as Table 1 shows. Moreover, the validity of all three methods are already verified in Table 2-3." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.508, + 0.367, + 0.525 + ], + "angle": 0, + "content": "4 Experiment" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.543, + 0.788, + 0.634 + ], + "angle": 0, + "content": "We first combine 'Class-Agnostic Branch' (CAB), 'Clustering Head' (CH) and 'Merging Head' (MH) with existing long-tailed methods (cf. Table 3), then choose 'SeeSaw [35] + CAB' as 'Our' to compete with state-of-the-art methods (cf. Table 4). Although SeeSaw is a lower baseline than ECM [7], it is more stable in reproducing. Finally, we generalize our methods to various evaluation metrics, different datasets and the mask branch." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.661, + 0.444, + 0.675 + ], + "angle": 0, + "content": "4.1 Experimental settings" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Datasets. We use the large vocabulary dataset LVIS1.0 [12] as our main dataset, which contains 100k training and 20k validation images. Rare \\((r)\\), common \\((c)\\) and frequent \\((f)\\) classes are defined by how many images they occur [12]: [0, 10] for rare, [11, 100] for common, and \\((100, +\\infty)\\) for frequent, respectively. We also adopt COCO-LT [25] and COCO2017 [36] to verify the generalization ability of our approach. COCO2017 is a large object detection dataset, containing 118k training and 5k validation images. It is relatively balanced in comparison with LVIS1.0. The COCO-LT dataset is an artificially sampled subset of COCO, with the same validation set but a long-tailed training set. It has about 99k training and 5k validation images. Following previous works [36], we partition COCO-LT" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.128 + ], + "angle": 0, + "content": "K. Zhu et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.145, + 0.788, + 0.215 + ], + "angle": 0, + "content": "Table 3: Experiments on LVIS1.0. We combine four existing methods with our approach 'adding a class-agnostic branch' (CAB), 'Clustering Head' (CH) or 'Merging Head' (MH). We reproduced RFS [12], EQLv2 [33], SeeSaw [35] and ECM [7] using their official code. For clarity, here we list the object detection metrics \\(\\mathrm{AP}^b\\) in the first four columns while putting the instance segmentation metrics AP in the latter ones." + }, + { + "type": "table", + "bbox": [ + 0.293, + 0.227, + 0.706, + 0.483 + ], + "angle": 0, + "content": "
Method+Our APbAPRbAPCbAPf bAPAPrAPcAPf
RFS [12]base24.713.423.131.423.714.222.9
CAB27.018.025.332.925.117.523.9
CH26.917.225.533.025.216.724.3
MH27.318.225.832.825.317.124.5
EQLv2 [33]base26.016.124.032.525.217.424.1
CAB28.120.426.333.526.019.524.9
CH27.318.225.833.125.717.724.8
MH27.117.325.733.225.617.524.7
SeeSaw [35]base27.318.226.532.326.919.626.8
CAB28.919.928.333.627.720.227.3
CH29.119.828.933.427.820.627.7
MH28.518.428.433.727.419.827.4
ECM [7]base27.717.726.633.127.219.626.6
CAB29.118.428.933.927.819.128.0
CH28.718.028.333.727.819.627.7
MH28.618.628.233.527.719.827.7
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.509, + 0.784, + 0.539 + ], + "angle": 0, + "content": "into 4 evaluation subsets according to the number of training instances per class, with bins of [1, 20), [20, 400), [400, 8000) and [8000, -), respectively." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.54, + 0.788, + 0.766 + ], + "angle": 0, + "content": "Training details. We reproduce four different methods as our baselines, including RFS [12], EQLv2 [33], SeeSaw [35] and ECM [7], following their default experiment settings. We employ MMDetection [5] as our detection framework to conduct our experiment, and train detection models of Faster-RCNN, Mask-RCNN and Cascade R-CNN for 1x or 2x scheduler (except Swin-Transformer based detectors), following previous works [7, 17]. The batch size and learning rate are set as 16 and 0.02, and the data augmentation strictly follows previous long-tailed detection methods [7, 33, 35]. During training, we use FP16 mixed precision training and the warmup strategy to stabilize the learning process. For the evaluation metrics, we adopt AP and \\(\\mathrm{AP}^b\\) for instance segmentation and object detection, respectively, and adopt \\(\\mathrm{AP}_1^b\\), \\(\\mathrm{AP}_2^b\\), \\(\\mathrm{AP}_3^b\\) and \\(\\mathrm{AP}_4^b\\) on COCOLT, corresponding to its 4 different subsets. With the suggested practice in LVIS official website, we run all our experiment 3 times on 8 RTX3090 GPUs to reduce the variance. Please refer to our supplementary material for more detailed information." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.787, + 0.537, + 0.803 + ], + "angle": 0, + "content": "4.2 LVIS detection and segmentation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Consistent improvements. we first evaluate the effectiveness of our method on the LVIS1.0 dataset by combing the proposed approach 'adding a class-agnostic" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.546, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Rectify the Regression Bias" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.145, + 0.788, + 0.215 + ], + "angle": 0, + "content": "Table 4: Comparison with state-of-the-art methods on the LVIS1.0 dataset. The results of 'Our' means our proposed CAB regression method combined with the SeeSaw loss ('SeeSaw + CAB', cf. Sec. 4). We show the metrics of instance segmentation AP and object detection \\(\\mathrm{AP}^b\\). The results with a \\(\\dagger\\) are copied from [7], while others are reproduced by us using their official released code." + }, + { + "type": "table", + "bbox": [ + 0.261, + 0.227, + 0.737, + 0.555 + ], + "angle": 0, + "content": "
ArchitectBackboneMethodAPAPrAPfAPcAPb
Mask-RCNN [14]R50-FPNCE18.70.416.529.319.7
RFS [13]23.714.222.929.324.7
BSCE [29]24.415.723.629.125.5
EQLv2 [33]25.217.424.129.926.0
ECM [7]†27.419.727.031.127.9
ECM27.219.626.631.327.7
SeeSaw [35]26.919.626.830.527.3
ROG [42]26.920.126.830.027.2
Our27.720.227.331.328.9
Mask-RCNN [14]R101-FPNCE [7]†25.516.624.530.626.6
EQLv2 [33]†27.220.625.931.427.9
ECM [7]†28.721.927.932.329.4
ECM28.620.928.332.229.3
SeeSaw [35]28.220.328.131.829.0
Our29.021.028.932.430.7
Cascade R-CNN [2]Swin-TECM [7]34.124.034.938.037.6
SeeSaw [35]34.224.634.737.837.8
Our34.624.735.338.138.2
Cascade R-CNN [2]Swin-BECM [7]†39.733.540.641.443.6
Our39.934.640.741.244.2
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.594, + 0.788, + 0.791 + ], + "angle": 0, + "content": "branch' (CAB) with existing long-tailed object detection methods. Since our main focus is on bounding box regression, we list the object detection results in early columns and segmentation results in later ones. As shown in Table 3, using CAB leads to consistent \\(\\mathrm{AP}^b\\) and AP improvement over existing classification-based methods, surpassing all of them with large margins. For object detection, Our CAB benefits the rare class a lot, with an increase of \\(4.6\\mathrm{AP}^b\\) and \\(4.3\\mathrm{AP}^b\\) on RFS and EQLv2. The same is true for instance segmentation, where a growing trend can be observed on all metrics, showing that CAB is also beneficial for later mask pixel predictions. Interestingly, the method 'RFS+CAB' (which uses a CE loss) can almost achieve the same object detection accuracy as the SeeSaw method, and surpasses EQLv2 for about \\(1\\mathrm{AP}^b\\). We thus conjecture that: besides merely focusing on classification, our regression methods can serve as strong alternatives that also strongly boost long-tailed detection accuracy." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Comparison with SOTA. We then compare our method with state-of-the-art methods using different object detection framework (Mask-RCNN, Cascade R-CNN) and backbones (ResNet-50, ResNet-101, Swin-T and Swin-B). Note" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.128 + ], + "angle": 0, + "content": "K. Zhu et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.788, + 0.176 + ], + "angle": 0, + "content": "Table 5: More evaluation metrics in LVIS1.0 dataset with a 1x scheduler using Mask RCNN. AP \\( {}_{\\text{boundary }}^{fixed} \\) and \\( {\\mathrm{{AP}}}_{\\text{box }}^{fixed} \\) are two newly proposed challenging metrics [6,9]." + }, + { + "type": "table", + "bbox": [ + 0.356, + 0.186, + 0.648, + 0.296 + ], + "angle": 0, + "content": "
MethodAPAPbAPfixedboundaryAPfixedbbox
BAGS [24]23.123.7-26.2
EQLv2 [33]23.924.020.325.9
SeeSaw [35]25.225.419.826.5
ECM [7]26.326.721.427.4
Our26.928.022.128.2
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.326, + 0.789, + 0.522 + ], + "angle": 0, + "content": "that 'Our' means 'SeeSaw + CAB'. For fair comparison, we reproduce majority of existing methods using their official released code unless specialized symbols \\((e.g.,^{\\dagger})\\) appears after a method's name. As can be seen in Table 4, our method achieves the overall highest accuracy in AP and \\(\\mathrm{AP}^b\\). For ResNet series models, our regression technique easily surpass the best competitor ECM [7], especially in \\(\\mathrm{AP}^b\\) (an increase of \\(1.2\\mathrm{AP}^b\\) for ResNet-50 and \\(1.1\\mathrm{AP}^b\\) for ResNet-101). The advantage also holds for ViT-based object detectors, where we surpassed the best competitor ECM [7] with both Swin-Tiny and Swin-Base backbone architectures. Following LVIS's common practice, we didn't list \\(\\mathrm{AP}_r^b\\) and \\(\\mathrm{AP}_c^b\\) here, but we want to emphasize that the advantage of our regression methods can be further enlarged when more metrics are listed (cf. Table 3). This is also true if we replace 'adding a Class-Agnostic Branch' (CAB) with the 'merging heads' regression alternative (cf. the best accuracy in Table 2)." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.55, + 0.442, + 0.567 + ], + "angle": 0, + "content": "4.3 Generalization ability" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.58, + 0.787, + 0.623 + ], + "angle": 0, + "content": "In this section, we will show the generalization ability of our regression methods in various aspects, including different evaluation metrics, datasets and its benefits on the mask branch." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.626, + 0.788, + 0.734 + ], + "angle": 0, + "content": "Different metrics. We first explore how different metrics affect our model's accuracy. Two additional metrics are \\(\\mathrm{AP}_{\\text {boundary }}\\) (a more strict metric in instance segmentation [6]) and \\(\\mathrm{AP}_{\\text {bbox }}^{\\text {fixed }}\\) (constraining 10,000 predicted bounding boxes per class across the dataset [9]). Object detector that obtain decent results in traditional metrics may not perform as well in these criterions. As shown in Table 5, our regression methods adapts well and surpasses all existing methods on both traditional and these challenging metrics." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.788, + 0.842 + ], + "angle": 0, + "content": "The COCO-LT dataset. We also transfer our regression to another long-tailed dataset: COCO-LT. This is an artificially sampled [36] subset of original COCO [25]. We calculate the over bounding box metrics \\(\\mathrm{AP}^b\\) and more fine-grained results: \\(\\mathrm{AP}_1^b\\), \\(\\mathrm{AP}_2^b\\), \\(\\mathrm{AP}_3^b\\) and \\(\\mathrm{AP}_4^b\\) (ranging from the rarest to the most frequent class). As shown in Table 6, our CAB leads to consistent improvements on all metrics (especially for the rarest categories \\(\\mathrm{AP}_1^b\\)) under different repeat-factor-sampling rates, showing the great power of CAB to help the rare" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.546, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Rectify the Regression Bias" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.144, + 0.784, + 0.174 + ], + "angle": 0, + "content": "Table 6: Experiments on COCO-LT [36]. Following LVIS [12], we use a similar repeat-factor-sampling (RFS) strategy with different sampling rates (3e-3, 5e-3 and 1e-2)." + }, + { + "type": "table", + "bbox": [ + 0.366, + 0.185, + 0.633, + 0.314 + ], + "angle": 0, + "content": "
Rate +CABAP1bAP2bAP3bAP4bAPb
3e-3✓2.316.926.730.423.5
6.219.127.030.524.4
5e-3✓3.619.326.730.624.3
7.520.027.330.725.0
1e-2✓8.720.227.830.325.2
10.821.628.030.525.8
" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.346, + 0.784, + 0.374 + ], + "angle": 0, + "content": "Table 7: Experiments on the (relatively) balanced COCO dataset. We adopted the Faster RCNN [30] detector and tried three different backbones." + }, + { + "type": "table", + "bbox": [ + 0.326, + 0.387, + 0.673, + 0.515 + ], + "angle": 0, + "content": "
Backbone+CABAP\\( AP_{50} \\)\\( AP_{75} \\)\\( AP_s \\)\\( AP_m \\)\\( AP_l \\)
Res-50✓37.358.340.321.741.048.2
38.358.942.122.441.450.4
Res-101✓39.460.343.022.943.451.0
39.960.343.622.743.752.8
Res-32x4d✓41.062.244.623.945.352.9
41.762.545.423.645.954.8
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.564, + 0.784, + 0.594 + ], + "angle": 0, + "content": "classes. In fact, this conclusion generally holds true for all sampling rates in our experiments. We only listed three here for simplicity and clarity." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.604, + 0.785, + 0.74 + ], + "angle": 0, + "content": "On balanced training set. Furthermore, we validate how our method performs on the relatively balanced dataset MS-COCO2017. We adopted Faster RCNN with three different backbones (ResNet-50, ResNet-101 and ResNext101-32x4d). Results in Table 7 clearly show that CAB generalizes well to datasets with more balanced distributions. Interestingly, CAB boosts metrics of large and medium objects while shows similar accuracy on small objects. This is possibly because that small objects occupy over \\(60\\%\\) of the total instances while large objects only consumes roughly \\(15\\%\\) [32]. Since CAB is more beneficial to less-frequent classes, it brings higher gains in large objects than in small ones." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.787, + 0.84 + ], + "angle": 0, + "content": "The mask branch. Finally, we apply our CAB to the segmentation branch to test whether adding an class-agnostic prior is suitable for mask prediction. Experimental results are in Table 8, where we add a class-agnostic mask prediction head and combines it with each class-specific mask head. As shown in the table, our CAB generalizes well to segmentation tasks. Our main experimental results may be further improved if CAB was applied in both box and mask predictions." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.128 + ], + "angle": 0, + "content": "K. Zhu et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.788, + 0.174 + ], + "angle": 0, + "content": "Table 8: Generalization to mask prediction. We added a class-agnostic mask prediction branch and combines it with each class-specific mask prediction results." + }, + { + "type": "table", + "bbox": [ + 0.377, + 0.185, + 0.623, + 0.281 + ], + "angle": 0, + "content": "
Branch +CABAPAPrAPbAPr
Box✓23.714.224.713.4
25.117.527.018.0
Mask✓23.714.224.713.4
24.114.625.013.9
" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.31, + 0.533, + 0.325 + ], + "angle": 0, + "content": "4.4 Discussion of our three remedies" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.339, + 0.784, + 0.37 + ], + "angle": 0, + "content": "Now we further discuss the proposed three methods (CAB, CH and MH) in detail, hoping to provide a more in-depth understanding of our hypothesis." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.371, + 0.785, + 0.55 + ], + "angle": 0, + "content": "Relationship of CAB to the objectness branch. At start, we want to clarify the differences between our CAB and the objectness branch approach [7, 35] adopted in the classification head. We have argued that although they seemingly share similar structures with ours, they are essentially different. First, the purpose of objectness branch is to deal with the imbalanced distribution between foreground and background classification samples [35], while our CAB aim to tackle the regression bias. Second, unlike classification where each class must preserve its own classifier, the regression heads can be clustered or merged, which provide more diverse solutions to reduce the regression bias besides CAB (cf. Table 2). Lastly, the effect of the objectness branch is yet to be proved because it will decrease the performance of rare classes even when combined with CE loss [35], while CAB leads to consistent improvements, especially in rare class." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.553, + 0.786, + 0.734 + ], + "angle": 0, + "content": "How many clusters are the best? Attentive readers might have questions why larger clusters K (such as 100) in CH achieve better results in Table 2b for rare class, since enlarging K means the (useful) class-agnostic information is reduced. We thus empirically verify that: besides objects numbers per group, the scale variance could potentially affect the final AP. We test CH with 'object scales' in more K. For each K, we calculate standard deviation of scales in each group, average, and get representative scale std. The same is done for average object. As Fig. 3a shows: as K decreases, training objects per group increases, which benefit rare classes. But, the variation of scales increases a lot, which makes regression head learning more difficult—imagine one head to fit offsets that vary a lot, and we also observed higher training loss for smaller K. Hence, \\(\\max AP_r^b\\) is not at \\(K = 1\\) or 10, but at \\(K = 50\\) in our experiments (cf. Fig. 3b)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.787, + 0.842 + ], + "angle": 0, + "content": "Why merging 'Common' class is the best?. We observe in Table 2c that merging common classes leads to the best improvement in \\(\\mathrm{AP}_r\\), which is somehow counter-intuitive. We conjecture that this is due to the partition shift of rare, common and frequent in LVIS train and validation sets. In the LVIS1.0 train dataset, the size ranges for frequent, common and rare are [0, 404], [405, 865] and [866, 1202], respectively. While for the validation set, they become [0, 212], [213, 536] and [537, 1202], respectively (cf. Fig. 1c). Hence, when we" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.547, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Rectify the Regression Bias" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "image", + "bbox": [ + 0.331, + 0.146, + 0.493, + 0.257 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.506, + 0.146, + 0.668, + 0.256 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.268, + 0.788, + 0.296 + ], + "angle": 0, + "content": "Fig. 3: The statistics (averaged standard deviation, averaged objects per group) of different K and its corresponding AP box." + }, + { + "type": "image", + "bbox": [ + 0.226, + 0.311, + 0.78, + 0.412 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.421, + 0.788, + 0.492 + ], + "angle": 0, + "content": "Fig. 4: Visualizations of detection results before (in the left of each group) and after (in the right) using our CAB. We adopted RFS [12] as the baseline in LVIS1.0 and combine it with our CAB regression method. In comparison, the proposed method is good at detecting missing objects, filtering duplicated objects away, as well as rectifying bounding box predictions. This figure needs to be viewed in color." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.524, + 0.788, + 0.569 + ], + "angle": 0, + "content": "use training set statistics to merge the common class, a large portion of rare categories in the validation set are also potentially merged, thus contributing to the validation improvement on both \\(\\mathrm{AP}_r\\) and \\(\\mathrm{AP}_c\\)." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.597, + 0.479, + 0.61 + ], + "angle": 0, + "content": "4.5 Visualization and ablation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.628, + 0.788, + 0.733 + ], + "angle": 0, + "content": "Flatten distribution. We first plot the RCNN regression loss of each category before and after combining with our regression CAB. The baseline methods we choose are EQLv2 and CE. As shown in Fig. 5, the regression loss for rare has seen a noticeable drop after adding our CAB, and the overall loss distribution has become much more balanced, too. We thus believe that the pipeline of our regression remedies can relieve the regression bias, further verifying that our hypothesis in Sec. 1 in indeed valid." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.788, + 0.841 + ], + "angle": 0, + "content": "More accurate box/mask. We then calculate AP50-AP95 of boxes and masks to find whether our CAB behaves well on stricter IoU thresholds. As show in Fig. 5c-5d, adding CAB achieves consistent accuracy gains over all IoU thresholds for both box and mask prediction, and is especially helpful for those hard IoU threshold (e.g., AP75-90) in box evaluation (cf. Fig. 5c). Since a higher threshold requires more precise box prediction, these results have shown that our regression methods is capable of producing precise boxes with better quality." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.128 + ], + "angle": 0, + "content": "K. Zhu et al." + }, + { + "type": "image", + "bbox": [ + 0.228, + 0.145, + 0.357, + 0.219 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.251, + 0.223, + 0.332, + 0.235 + ], + "angle": 0, + "content": "(a) w/o CAB" + }, + { + "type": "image", + "bbox": [ + 0.368, + 0.146, + 0.497, + 0.219 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.395, + 0.223, + 0.469, + 0.235 + ], + "angle": 0, + "content": "(b) w/CAB" + }, + { + "type": "image", + "bbox": [ + 0.509, + 0.146, + 0.637, + 0.219 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.534, + 0.223, + 0.612, + 0.235 + ], + "angle": 0, + "content": "(c) box delta" + }, + { + "type": "image", + "bbox": [ + 0.649, + 0.146, + 0.777, + 0.219 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.669, + 0.223, + 0.758, + 0.235 + ], + "angle": 0, + "content": "(d) mask delta" + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.245, + 0.785, + 0.275 + ], + "angle": 0, + "content": "Fig. 5: The loss distribution shift before and after combining with our CAB (Fig. 5a, 5b), and AP improvement of combining our CAB with baseline in LVIS (Fig. 5c, 5d)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.307, + 0.789, + 0.504 + ], + "angle": 0, + "content": "Qualitative results. Last but not least, we provide example detection images sampled from the LVIS1.0 validation dataset. For simplicity, we choose RFS [12] with CE loss as baseline and combine it with our CAB method. As shown in Fig. 4, both baseline and our CAB detect most of the objects in an image, but CAB generally caters for more details. For example, CAB can help discover missed boxes, like the cabinet in the lower left images. It is also clearly illustrated that CAB helps filter duplicate boxes away in the final predictions (e.g., the pictures with elephant and cow with grass background). Since CAB brings in better predictions, duplicate boxes will have larger overlap that may be suppressed by NMS (non-maximum suppression). If we zoom this figure (cf. the last column in Fig. 4), we will find that CAB is capable of rectifying the predicted boxes, and this can empirically explain why our CAB achieves much better AP under higher and more difficult IoU thresholds (cf. Fig. 5c)." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.531, + 0.528, + 0.547 + ], + "angle": 0, + "content": "5 Conclusions and Limitations" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.568, + 0.788, + 0.688 + ], + "angle": 0, + "content": "In this paper, we discovered that the regression bias (imbalanced regression loss distribution on the RCNN head) exists in long-tailed object detection, and adversely affects detection results. We thus proposed three remedies for rectifying the regression bias. The proposed method significantly boosts the performance of rare class \\(\\mathrm{AP}^b\\), and achieves state-of-the-art results. We also generalize our regression methods to balanced dataset, different evaluation metrics and the mask branch. Finally, visualizations show that our method indeed produces better predicted bounding boxes." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.69, + 0.789, + 0.841 + ], + "angle": 0, + "content": "As for the limitations, it remains unclear why the boosted accuracy becomes lower when the baseline are higher (e.g., the improvement of CAB on ECM is lower than that on RFS, cf. Table 3). This may relate to the upper bound a backbone model can achieve. Since large network generally better fit the dataset (cf. Table 4), the performance a ResNet-50 can achieve is limited. We thus call for involving larger vision models on the LVIS dataset to better handle the long-tailed object detection problem. Besides involving stronger vision backbones, we will try to testify the adaptability of the proposed methods when different pretraining information is integrated into the backbone network (e.g. self-supervised learning), which is left as our future work." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.547, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Rectify the Regression Bias" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.403, + 0.163 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.177, + 0.784, + 0.207 + ], + "angle": 0, + "content": "We thank Yin-Yin He (M.Sc. degree in Nanjing University, now in ByteDance) for providing helpful discussions in long-tailed object detection." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.23, + 0.323, + 0.245 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.261, + 0.785, + 0.302 + ], + "angle": 0, + "content": "1. Alexandridis, K.P., Deng, J., Nguyen, A., Luo, S.: Long-tailed instance segmentation using gumbel optimized loss. In: ECCV. Lecture Notes in Computer Science, vol. 13670, pp. 353-369. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.303, + 0.785, + 0.329 + ], + "angle": 0, + "content": "2. Cai, Z., Vasconcelos, N.: Cascade r-cnn: Delving into high quality object detection. In: CVPR. pp. 6154-6162 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.33, + 0.785, + 0.357 + ], + "angle": 0, + "content": "3. Cao, K., Wei, C., Gaidon, A., Arechiga, N., Ma, T.: Learning imbalanced datasets with label-distribution-aware margin loss. In: NeurIPS. pp. 1565-1576 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.359, + 0.785, + 0.412 + ], + "angle": 0, + "content": "4. Caron, M., Bojanowski, P., Joulin, A., Douze, M.: Deep clustering for unsupervised learning of visual features. In: Ferrari, V., Hebert, M., Sminchisescu, C., Weiss, Y. (eds.) ECCV. Lecture Notes in Computer Science, vol. 11218, pp. 139-156. Springer (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.413, + 0.785, + 0.481 + ], + "angle": 0, + "content": "5. Chen, K., Wang, J., Pang, J., Cao, Y., Xiong, Y., Li, X., Sun, S., Feng, W., Liu, Z., Xu, J., Zhang, Z., Cheng, D., Zhu, C., Cheng, T., Zhao, Q., Li, B., Lu, X., Zhu, R., Wu, Y., Dai, J., Wang, J., Shi, J., Ouyang, W., Loy, C.C., Lin, D.: MMDetection: Open mmlab detection toolbox and benchmark. arXiv preprint arXiv:1906.07155 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.482, + 0.785, + 0.523 + ], + "angle": 0, + "content": "6. Cheng, B., Girshick, R., Dollar, P., Berg, A.C., Kirillov, A.: Boundary iou: Improving object-centric image segmentation evaluation. In: CVPR. pp. 15334-15342 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.524, + 0.785, + 0.564 + ], + "angle": 0, + "content": "7. Cho, J.H., Krahenbuhl, P.: Long-tail detection with effective class-margins. In: ECCV. Lecture Notes in Computer Science, vol. 13668, pp. 698-714. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.565, + 0.785, + 0.592 + ], + "angle": 0, + "content": "8. Cui, Y., Jia, M., Lin, T.Y., Song, Y., Belongie, S.: Class-balanced loss based on effective number of samples. In: CVPR. pp. 9268-9277 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.593, + 0.785, + 0.633 + ], + "angle": 0, + "content": "9. Dave, A., Dóllár, P., Ramanan, D., Kirillov, A., Girshick, R.: Evaluating large-vocabulary object detectors: The devil is in the details. arXiv preprint arXiv:2102.01066 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.634, + 0.785, + 0.661 + ], + "angle": 0, + "content": "10. Du, Y., Wu, J.: No one left behind: Improving the worst categories in long-tailed learning. In: CVPR. pp. 15804-15813 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.662, + 0.785, + 0.689 + ], + "angle": 0, + "content": "1. Feng, C., Zhong, Y., Huang, W.: Exploring classification equilibrium in long-tailed object detection. In: ICCV. pp. 3417-3426 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.689, + 0.785, + 0.716 + ], + "angle": 0, + "content": "2. Gupta, A., Dollar, P., Girshick, R.: Lvis: A dataset for large vocabulary instance segmentation. In: CVPR. pp. 5356-5364 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.717, + 0.785, + 0.744 + ], + "angle": 0, + "content": "3. Gupta, A., Dollar, P., Girshick, R.: Lvis: A dataset for large vocabulary instance segmentation. In: CVPR. pp. 5356-5364 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.745, + 0.785, + 0.771 + ], + "angle": 0, + "content": "4. He, K., Gkioxari, G., Dollár, P., Girshick, R.: Mask R-CNN. In: ICCV. pp. 2961-2969 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.772, + 0.785, + 0.813 + ], + "angle": 0, + "content": "5. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 770-778 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.814, + 0.785, + 0.84 + ], + "angle": 0, + "content": "6. He, Y.Y., Wu, J., Wei, X.S.: Distilling virtual examples for long-tailed recognition. In: ICCV. pp. 235-244 (2021)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.261, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.359, + 0.127 + ], + "angle": 0, + "content": "K. Zhu et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.785, + 0.175 + ], + "angle": 0, + "content": "17. He, Y.Y., Zhang, P., Wei, X.S., Zhang, X., Sun, J.: Relieving long-tailed instance segmentation via pairwise class balance. In: CVPR. pp. 7000-7009 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.176, + 0.785, + 0.202 + ], + "angle": 0, + "content": "18. Hsieh, T., Robb, E., Chen, H., Huang, J.: Droploss for long-tail instance segmentation. In: AAAI. pp. 1549-1557 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.203, + 0.785, + 0.229 + ], + "angle": 0, + "content": "19. Hu, X., Jiang, Y., Tang, K., Chen, J., Miao, C., Zhang, H.: Learning to segment the tail. In: CVPR (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.23, + 0.785, + 0.256 + ], + "angle": 0, + "content": "20. Kang, B., Xie, S., Rohrbach, M., Yan, Z., Gordo, A., Feng, J., Kalantidis, Y.: Decoupling representation and classifier for long-tailed recognition. In: ICLR (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.257, + 0.785, + 0.284 + ], + "angle": 0, + "content": "21. Li, B., Yao, Y., Tan, J., Zhang, G., Yu, F., Lu, J., Luo, Y.: Equalized focal loss for dense long-tailed object detection. In: CVPR. pp. 6990-6999 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.285, + 0.785, + 0.31 + ], + "angle": 0, + "content": "22. Li, M., Cheung, Y.m., Lu, Y.: Long-tailed visual recognition via gaussian clouded logit adjustment. In: CVPR. pp. 6929-6938 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.311, + 0.785, + 0.338 + ], + "angle": 0, + "content": "23. Li, T., Wang, L., Wu, G.: Self supervision to distillation for long-tailed visual recognition. In: ICCV. pp. 630-639 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.339, + 0.785, + 0.378 + ], + "angle": 0, + "content": "24. Li, Y., Wang, T., Kang, B., Tang, S., Wang, C., Li, J., Feng, J.: Overcoming classifier imbalance for long-tail object detection with balanced group softmax. In: CVPR (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.379, + 0.785, + 0.419 + ], + "angle": 0, + "content": "25. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dólár, P., Zitnick, C.L.: Microsoft COCO: Common objects in context. In: ECCV. Lecture Notes in Computer Science, vol. 8693, pp. 740-755. Springer (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.42, + 0.785, + 0.446 + ], + "angle": 0, + "content": "26. Liu, Z., Miao, Z., Zhan, X., Wang, J., Gong, B., Yu, S.X.: Large-scale long-tailed recognition in an open world. In: CVPR. pp. 2537-2546 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.447, + 0.785, + 0.473 + ], + "angle": 0, + "content": "27. Menon, A.K., Jayasumana, S., Rawat, A.S., Jain, H., Veit, A., Kumar, S.: Long-tail learning via logit adjustment. In: ICLR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.474, + 0.785, + 0.5 + ], + "angle": 0, + "content": "28. Pang, J., Chen, K., Shi, J., Feng, H., Ouyang, W., Lin, D.: Libra r-cnn: Towards balanced learning for object detection. In: CVPR. pp. 821-830 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.501, + 0.785, + 0.528 + ], + "angle": 0, + "content": "29. Ren, J., Yu, C., sheng, s., Ma, X., Zhao, H., Yi, S., Li, h.: Balanced meta-softmax for long-tailed visual recognition. In: NeurIPS. pp. 4175-4186 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.529, + 0.785, + 0.554 + ], + "angle": 0, + "content": "30. Ren, S., He, K., Girshick, R., Sun, J.: Faster R-CNN: Towards real-time object detection with region proposal networks. In: NeurIPS. pp. 91-99 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.555, + 0.785, + 0.609 + ], + "angle": 0, + "content": "31. Russakovsky, O., Deng, J., Su, H., Krause, J., Satheesh, S., Ma, S., Huang, Z., Karpathy, A., Khosla, A., Bernstein, M., Berg, A., Fei-Fei, L.: ImageNet large scale visual recognition challenge. International Journal of Computer Vision 115(3), 211-252 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.61, + 0.785, + 0.636 + ], + "angle": 0, + "content": "32. Singh, B., Davis, L.S.: An analysis of scale invariance in object detection SNIP. In: CVPR. pp. 3578-3587 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.637, + 0.785, + 0.663 + ], + "angle": 0, + "content": "33. Tan, J., Lu, X., Zhang, G., Yin, C., Li, Q.: Equalization loss v2: A new gradient balance approach for long-tailed object detection. In: CVPR. pp. 1685-1694 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.664, + 0.785, + 0.69 + ], + "angle": 0, + "content": "34. Tan, J., Wang, C., Li, B., Li, Q., Ouyang, W., Yin, C., Yan, J.: Equalization loss for long-tailed object recognition. In: CVPR (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.691, + 0.785, + 0.731 + ], + "angle": 0, + "content": "35. Wang, J., Zhang, W., Zang, Y., Cao, Y., Pang, J., Gong, T., Chen, K., Liu, Z., Loy, C.C., Lin, D.: Seesaw loss for long-tailed instance segmentation. In: CVPR. pp. 9695-9704 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.732, + 0.785, + 0.786 + ], + "angle": 0, + "content": "36. Wang, T., Li, Y., Kang, B., Li, J., Liew, J.H., Tang, S., Hoi, S.C.H., Feng, J.: The devil is in classification: A simple framework for long-tail instance segmentation. In: ECCV. Lecture Notes in Computer Science, vol. 12359, pp. 728-744. Springer (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.787, + 0.785, + 0.813 + ], + "angle": 0, + "content": "37. Wang, T., Zhu, Y., Zhao, C., Zeng, W., Wang, J., Tang, M.: Adaptive class suppression loss for long-tail object detection. In: CVPR. pp. 3103-3112 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.814, + 0.785, + 0.84 + ], + "angle": 0, + "content": "38. Wang, X., Lian, L., Miao, Z., Liu, Z., Yu, S.: Long-tailed recognition by routing diverse distribution-aware experts. In: ICLR (2021)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.547, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Rectify the Regression Bias" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.189 + ], + "angle": 0, + "content": "39. Wu, J., Song, L., Wang, T., Zhang, Q., Yuan, J.: Forest R-CNN: large-vocabulary long-tailed object detection and instance segmentation. In: ACM MM. pp. 1570-1578 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.19, + 0.787, + 0.218 + ], + "angle": 0, + "content": "40. Xuan, G., Zhang, W., Chai, P.: EM algorithms of gaussian mixture model and hidden markov model. In: ICIP. pp. 145-148 (2001)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.218, + 0.786, + 0.259 + ], + "angle": 0, + "content": "41. Zhang, C., Pan, T.Y., Li, Y., Hu, H., Xuan, D., Changpinyo, S., Gong, B., Chao, W.L.: Mosaicos: A simple and effective use of object-centric images for long-tailed object detection. In: ICCV. pp. 417-427 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.26, + 0.787, + 0.3 + ], + "angle": 0, + "content": "42. Zhang, S., Chen, C., Peng, S.: Reconciling object-level and global-level objectives for long-tail detection. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 18982-18992 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.301, + 0.787, + 0.342 + ], + "angle": 0, + "content": "43. Zhong, Z., Cui, J., Liu, S., Jia, J.: Improving calibration for long-tailed recognition. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 16489-16498 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.343, + 0.787, + 0.383 + ], + "angle": 0, + "content": "44. Zhou, B., Cui, Q., Wei, X.S., Chen, Z.M.: Bbn: Bilateral-branch network with cumulative learning for long-tailed visual recognition. In: CVPR. pp. 9716-9725 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.384, + 0.787, + 0.411 + ], + "angle": 0, + "content": "45. Zhu, K., Fu, M., Wu, J.: Multi-label self-supervised learning with scene images. In: ICCV. pp. 6694-6703 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.412, + 0.787, + 0.439 + ], + "angle": 0, + "content": "46. Zhu, K., He, Y., Wu, J.: Quantized feature distillation for network quantization. In: AAAI. pp. 11452-11460 (2023)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.439 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/Rectify the Regression Bias in Long-Tailed Object Detection/07bdb2f4-ff12-4767-80ed-dec2d3b31ce7_origin.pdf b/2024/Rectify the Regression Bias in Long-Tailed Object Detection/07bdb2f4-ff12-4767-80ed-dec2d3b31ce7_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..1a8149301196fecbb51426905544448f44396408 --- /dev/null +++ b/2024/Rectify the Regression Bias in Long-Tailed Object Detection/07bdb2f4-ff12-4767-80ed-dec2d3b31ce7_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b924a2724c4f310a8c6386ea0a72e666ee9b52ff1ee5e55b3d9f3eb019d5f376 +size 5003363 diff --git a/2024/Rectify the Regression Bias in Long-Tailed Object Detection/full.md b/2024/Rectify the Regression Bias in Long-Tailed Object Detection/full.md new file mode 100644 index 0000000000000000000000000000000000000000..c0f1afe7cb7770a8f3355feb2428c7092920b6f8 --- /dev/null +++ b/2024/Rectify the Regression Bias in Long-Tailed Object Detection/full.md @@ -0,0 +1,298 @@ +# Rectify the Regression Bias in Long-Tailed Object Detection + +Ke Zhu, Minghao Fu, Jie Shao, Tianyu Liu, and Jianxin Wu + +State Key Laboratory for Novel Software Technology + +Nanjing University, Nanjing, China + +School of Artificial Intelligence, Nanjing University, China + +{zhuk, fumh, shaoj, liuty}@lamda.nju.edu.cn, wujx2001@nju.edu.cn + +Abstract. Long-tailed object detection faces great challenges because of its extremely imbalanced class distribution. Recent methods mainly focus on the classification bias and its loss function design, while ignoring the subtle influence of the regression branch. This paper shows that the regression bias exists and does adversely and seriously impact the detection accuracy. While existing methods fail to handle the regression bias, the class-specific regression head for rare classes is hypothesized to be the main cause of it in this paper. As a result, three kinds of viable solutions to cater for the rare categories are proposed, including adding a class-agnostic branch, clustering heads and merging heads. The proposed methods brings in consistent and significant improvements over existing long-tailed detection methods, especially in rare and common classes. The proposed method achieves state-of-the-art performance in the large vocabulary LVIS dataset with different backbones and architectures. It generalizes well to more difficult evaluation metrics, relatively balanced datasets, and the mask branch. This is the first attempt to reveal and explore rectifying of the regression bias in long-tailed object detection. + +# 1 Introduction + +Long-tailed object detection draws great attention [12] for its practical utility recently. Numerous efforts [1,7,34,35,37] have been made to tackle this challenging task, such as re-weighting [33,35,37], over-sampling [11,12,39], and balanced grouping [19,24]. These methods are proposed to prevent the tail classes from being overwhelmed due to discouraging gradients [33, 37], inferior predicting scores [11,24] or insufficient samples [11,28]. + +Long-tailed detection often involves both classification and regression heads. While almost all existing methods focus on mitigating the classification bias (e.g., adjusting the classification structure in detection branches), little or no attention has been paid to the regression branch. We will show in this paper that the regression bias has significant adverse effects in long-tailed object detection, but previous methods failed to identify this important issue. + +![](images/cb55291d3410932d064ef0546897379e5704c6b2049bd8f4de0a558c7c6a3032.jpg) +(a) RCNN loss + +![](images/86ed05ecb93ccadeb52977fa1c71d2cdc8ab58bb374d63f4c205230d844b63ab.jpg) +(b) RPN loss +Fig. 1: 1a shows the RCNN regression train loss of frequent, common and rare categories. 1b shows the RPN regression loss. 1c shows distribution of per class mean object scales in LVIS1.0. 'delta' in 1c is the negative of difference between train and validation set sizes for different classes. + +![](images/ce00f1ff39dd760bd8410b0995ffbbf80c95a55a864f2c7e91201e52872539a9.jpg) +(c) Mean scale in LVIS train and val set + +Fig. 1 clearly showcases this issue. For different methods (EQLv2 [33], See-Saw [35], ECM [7] and CrossEntropy CE [12]) trained on the LVIS1.0 [12] dataset, we plot the regression branch's losses for the final detection RCNN head in Fig. 1a. It is clear that the regression losses of rare categories are significantly higher than those of frequent and common categories, which inevitably will lead to poor regression results (and hence detection results) for rare classes. We name this finding as the regression bias, but existing long-tailed detection methods all fail to deal with or even identify this issue. + +To further demonstrate the importance of regression, we calculate the class-wise mean scale of GT (groundtruth) boxes in LVIS train and validation sets, as well as their differences (cf. Fig. 1c). The scale shift of rare classes is much larger than that of frequent classes. Since regression is highly correlated with box scale [30], it is thus inherently difficult for a rare class to learn a good bounding box regressor with both few samples and large scale shift. In short, it is crucial to properly handle the regression bias in long-tailed object detection. + +Our solution to rectify the regression bias is motivated by Fig. 1b. We find that the regression loss in RPN is balanced where rare, common and frequent categories have almost the same regression loss, which is almost immune to the regression bias when comparing Fig. 1b with Fig. 1a. + +The key difference between RPN and RCNN regression is that the former is class-agnostic (i.e., all classes share the same regression parameters), while the latter is class-specific. Then, one natural question arises: Can class-agnostic head improve a tail class's generalization ability and handle the regression bias? Our hypothesis is that rare classes do favor class-agnostic regression heads. + +Our hypothesis is supported by experiments in Table 1, where we compared class-specific and -agnostic regression heads in RCNN. By replacing the classification head with groundtruth class labels, experiments in Table 1 disentangled the impact of the classification head and focus on the regression heads. It is clear that the class-agnostic head possesses substantial advantages: $\mathrm{AP}_r^b$ (for rare) surges from 0.7 to 54.6, even surpassing $\mathrm{AP}_f^b$ (for frequent classes)! However, the agnostic head will bring a small drop in frequent classes (e.g., from 40.7 to 40.0 in $\mathrm{AP}_f^b$ ). The final version of our conjecture is: the rare (and possibly + +common) classes indeed favor class-agnostic regression, while the frequent classes prefers a class-specific regression, and there should be a trade-off between the two to optimize for all three types of categories. + +Table 1: Results on LVIS1.0 based on the CE baseline. We trained a Mask-RCNN R50-FPN detector [14] with class-specific (the default setting) and class-agnostic regression heads in the RCNN head. Please note that during test, the predicted proposals were provided groundtruth (GT) classification results. + +
Reg. HeadGTAPAPbAPrAPpcAPfb
Specificno18.719.70.316.531.7
yes37.239.834.141.340.7
Agnosticno18.019.30.715.531.5
yes38.445.654.647.040.0
+ +Accordingly, we design three different methods to fully rectify the regression bias, including adding a class-agnostic head, clustering similar heads, or merging heads. All three methods bring in positive effects (cf. Table 2), which verifies that rectifying regression bias is indeed crucial. We choose to adopt 'adding a class-agnostic head' in our main experiment for its simplicity, which leads to consistent and significant improvements over previous long-tailed detection pipeline, and has achieved state-of-the-art performance with various backbones and architectures. Moreover, our method shows robust generalization ability (cf. Tables 5-8) under varying settings, including different datasets (COCO/COCOLT [36]), different evaluation metrics and even adapts to the mask branch design. Furthermore, visualizations show that the proposed method leads to more accurate bounding box predictions (cf. Fig. 4) and indeed alleviates the regression bias (cf. Fig. 5). In summary, our contributions are: + +1. For the first time, we reveal and successfully handle the regression bias in long-tailed object detection. +2. We propose three remedies to alleviate this bias, all of which produce consistent improvements over existing methods. +3. Our method achieves state-of-the-art results on LVIS as well as generalizing across datasets, metrics and even the mask branch. Visualizations qualitatively verify our hypothesis, too. + +# 2 Related Work + +Vast number of pipelines and techniques have been invented in the field of general purpose object detection. Although most of these detectors can possibly be adapted to balanced detection datasets [25], they can nevertheless hardly handle extremely long-tailed distribution. Dedicated methods are hence required. + +Modern long-tailed learning methods aim to solve two important tasks: image classification [31] and object detection/segmentation [12, 30]. + +The first task, long-tailed image classification [26], is well explored and displays great diversity in terms of specific methods [3,29], training pipelines [20,44], post-hoc analysis [22,27] and network architecture [16,23,38]. Early attempts in this family focused on the details of re-weighting [3] or re-sampling [8] techniques to provide a relatively balanced (data) distribution for tail classes. Later on, the decoupling pipeline [20] for long-tailed learning becomes popular, paving the way for various successors [43,44]. This decoupled paradigm are based on the insight that instance sampling is beneficial for representation, while the classifier needs to be calibrated to alleviate its bias. Some works try to unify long-tailed learning with theoretical analysis [22,27], which involves label distribution [29] and generalization bound [27]. Most recent methods are much more diverse in ideas: they involve knowledge distillation [16,38,46], self-supervised learning [23,45] or statistical approach [10] to better handle long-tailed recognition. + +The combination of long-tailed learning and object detection or instance segmentation proves to be more challenging [12], because the class distribution in LVIS [12] is extremely imbalanced and naively applying common object detection techniques leads to un-satisfactory results [18, 24, 34, 35, 37]. Dominating solutions in this area are re-sampling [12, 39] and re-weighting [17, 21, 33, 35, 37]. There are variants of them that adopt balanced grouping [24], class incremental learning [19], augmented feature sampling [11, 36] or extra data source [41]. A common characteristic of these long-tailed detection methods is that they only focus on the classification task, and ignore the subtle influence of the regression branch. In this paper, we thoroughly explore the previously unvisited regression bias in long-tailed object detection, and propose a simple yet novel method to tackle long-tailed object detection. + +In [36], there is an experiment with similar design as ours in Table 1. But, we emphasize that they are fundamentally different. On one hand, conclusions in [36] is that 'performance drop in LVIS is mainly caused by the proposal classification [36]' and hence they focused on classification. On the other hand, their class agnostic-head is a supplementary to support their classification hypothesis, while our design and experiments lead to a novel finding: the regression bias. + +# 3 Method + +Now we elaborate three different remedies to alleviate the regression bias, and start from the background on long-tailed object detection. + +# 3.1 Preliminaries + +We take Faster-RCNN [14] as an example. For a scene image $I$ , it is first fed into a backbone network $\phi(\cdot)$ (e.g., ResNet [15]) to get the image feature $f \in \mathbb{R}^{d \times w \times h}$ : $f = \phi(I)$ , with dimensionality, width and height denoted as $d$ , $w$ , $h$ , respectively. A region proposal network (RPN) [30] that contains both agnostic + +![](images/48d9ce9b44f1f0519ef1709b476b10bded8ad48d8f78d5d48de7414b2e12ff0c.jpg) +Fig. 2: Illustration of the regular two-stage detection pipeline and the proposed regression methods. Previous methods (the left figure) generally focus on the final classification branch (the yellow arrow), while we focus on rectifying the regression bias (the red arrow). The right part shows our three regression methods, including adding an extra branch $W_{0}$ , clustering regression heads (e.g., from $W_{1},\ldots ,W_{C}$ to $W_1',W_2'$ ) and merging (e.g., merging rare categories into one regression head $W_{\text{rare}}$ , cf. Table 2). In our main experiments, we choose 'adding an extra branch' for its simplicity. This figure needs to be viewed in color. + +classification and regression branches is then applied on the feature tensor $f$ to produce proposals $p$ from pre-defined anchors. The ROAlign [14] then extracts proposal features: + +$$ +p = \operatorname {R P N} (f), \quad f _ {p} = \operatorname {R O I A l i g n} (f, p), \tag {1} +$$ + +where $p$ represents a large set of proposals, and $f_{p}$ is the set of aligned proposal features. $f_{p}$ goes into post-process modules (e.g., NMS [30]) before being sent to the RCNN head to get the final features set $f_{n}$ : $f_{n} = \mathrm{RCNN}(f_{p})$ , which is fed to classification and regression branches (e.g., linear layers) to produce the prediction results. + +# 3.2 Our three remedies for the regression bias + +In Faster RCNN, there is a dedicated regression head for each class: + +$$ +r _ {i} = W _ {i} ^ {T} f _ {n}, \quad i = 1, 2, \dots , C, \tag {2} +$$ + +where $r_i = (\delta x_i, \delta y_i, \delta w_i, \delta h_i)$ represents the regression offset for class $i$ , and $W_i$ is the class-specific regression head (a linear layer). We will now present our approaches to rectify the regression bias. The existing pipeline and our proposed methods are shown in Fig. 2. + +Extra class-agnostic branch. This is a simple remedy to cope with the regression bias. Since rare classes favor a class-agnostic head while class-specific ones are slightly preferable to frequent classes in our hypothesis, we propose to use both heads. For class $i$ , its regression head changes to: + +$$ +W _ {i} ^ {\prime} = \alpha W _ {0} + (1 - \alpha) W _ {i}, \tag {3} +$$ + +Table 2: Our three different methods to alleviate the regression bias. In 2a, we add an extra class-agnostic regression head during training and testing, where the shared class-agonistic head and the class-specific head are combined using a weight $\alpha$ . In 2b, we perform regression head clustering according to the number of instances (' $n'$ ) or mean box scale (' $s'$ ). In 2c, we directly merge the regression heads of rare, common or frequent classes. The original class-specific heads in 2b and 2c are replaced by the new heads. Note that 'base' means the baseline method, i.e., with only class-specific regression heads. In 2a, $\alpha = 0$ is equivalent to the baseline method, and $\alpha = 1.0$ is equivalent to merging all $r, c, f$ heads into a single head (i.e., the last row in 2c). + +
(α) APAPrAPbAPrknsAPAPrAPbAPrmerge APAPrAPbAPr
0.023.714.224.713.4base23.714.224.713.4base23.714.224.713.4
0.224.115.825.415.110✓24.514.526.314.4r25.116.326.716.7
0.525.117.527.018.010✓24.212.826.012.6c25.517.727.217.2
0.824.417.025.916.4100✓24.413.026.212.9r, c25.317.127.318.2
1.024.716.726.718.3100✓25.216.726.916.7r, c, f24.716.726.718.3
+ +(a) class-agnostic head +(b) Clustering heads +(c) Merging heads + +where $W_0$ is a shared class-agnostic regression head for all classes, and $\alpha$ is a hyper-parameter to balance class-agnostic and class-specific heads. We empirically find that this simple change leads to consistent improvements over the default class-specific regression head (cf. Table 2a), while simply setting $\alpha = 0.5$ gives the optimal trade-off. + +Clustering heads. This method is motivated by the analysis in Fig. 1c. Since some categories have similar statistics, we can assign them a shared regression head to improve the generalization ability. We implement it by following three steps: sorting, grouping and assigning. First, we use the number of instance or mean box scale to sort the original categories in descending order ( $C = 1203$ in LVIS): + +$$ +W _ {1}, \dots , W _ {C} \stackrel {\text {s o r t}} {\Longrightarrow} W _ {1} ^ {s}, \dots , W _ {C} ^ {s}. \tag {4} +$$ + +Then, we cluster them into $K$ groups. During clustering, we do not rely on time-cost algorithms such as K-means [4] or GMM [40]), but simply put adjacent classes into one group, and each group has the same number of classes $N = \frac{C}{K}$ : + +$$ +\{(W _ {N \times i + 1} ^ {s}, W _ {N \times i + 2} ^ {s}, \dots , W _ {N \times i + N} ^ {s}) \} _ {i = 0} ^ {K - 1}, \tag {5} +$$ + +Finally, each group $i$ share one regression head: + +$$ +W _ {g i} \stackrel {\text {r e p l a c e}} {\Longleftarrow} \left(W _ {N \times i + 1} ^ {s}, W _ {N \times i + 2} ^ {s}, \dots , W _ {N \times i + N} ^ {s}\right). \tag {6} +$$ + +These shared regression matrix are then used for both training and testing. As shown in Table 2b, clustering heads with similar scale statistics brings robust improvements over the baseline methods (see more discussion in Sec. 4.4). + +Merge heads. This approach has similar motivation to the previous one, which clusters regression heads, but is even more straightforward. We simply + +group the regression heads in pre-defined clusters. For example, we let all rare categories share a common regression head $W_{\text{rare}}$ , and the same for common and frequent classes. Experimental results can be found in Table 2c, where we try four different combinations (note that $r, c$ means merging rare and common classes all together into one regression head). The results indicate that merging always leads to performance gains, especially for the rare categories. The most significant improvements come from merging only the common class (i.e., the row denoted as 'c' in Table 2c, also cf. Sec. 4.4 for discussion on partition shift). + +All our proposed remedies improve the accuracy of long-tailed object detection and instance segmentation, which verifies the importance of rectifying the regression bias. We will show that all these methods lead to consistent and nontrivial margin over various existing approaches (cf. Table 3). We then choose adding a class-agnostic branch (the first approach, cf. Table 2a) to conduct the rest of our experiments for its simplicity. + +Attentive readers might advocate that all proposed three methods are to be verified on all datasets. However, we argue that this is not necessary and our choice has its advantages. On one hand, it does not need any dataset statistics, which is especially appreciated when the number of categories and the data distribution is unknown. On the other hand, by combining both class-agnostic and class-specific heads, it fully exploits object priors and per class knowledge: each type of head has its own merit, as Table 1 shows. Moreover, the validity of all three methods are already verified in Table 2-3. + +# 4 Experiment + +We first combine 'Class-Agnostic Branch' (CAB), 'Clustering Head' (CH) and 'Merging Head' (MH) with existing long-tailed methods (cf. Table 3), then choose 'SeeSaw [35] + CAB' as 'Our' to compete with state-of-the-art methods (cf. Table 4). Although SeeSaw is a lower baseline than ECM [7], it is more stable in reproducing. Finally, we generalize our methods to various evaluation metrics, different datasets and the mask branch. + +# 4.1 Experimental settings + +Datasets. We use the large vocabulary dataset LVIS1.0 [12] as our main dataset, which contains 100k training and 20k validation images. Rare $(r)$ , common $(c)$ and frequent $(f)$ classes are defined by how many images they occur [12]: [0, 10] for rare, [11, 100] for common, and $(100, +\infty)$ for frequent, respectively. We also adopt COCO-LT [25] and COCO2017 [36] to verify the generalization ability of our approach. COCO2017 is a large object detection dataset, containing 118k training and 5k validation images. It is relatively balanced in comparison with LVIS1.0. The COCO-LT dataset is an artificially sampled subset of COCO, with the same validation set but a long-tailed training set. It has about 99k training and 5k validation images. Following previous works [36], we partition COCO-LT + +Table 3: Experiments on LVIS1.0. We combine four existing methods with our approach 'adding a class-agnostic branch' (CAB), 'Clustering Head' (CH) or 'Merging Head' (MH). We reproduced RFS [12], EQLv2 [33], SeeSaw [35] and ECM [7] using their official code. For clarity, here we list the object detection metrics $\mathrm{AP}^b$ in the first four columns while putting the instance segmentation metrics AP in the latter ones. + +
Method+Our APbAPRbAPCbAPf bAPAPrAPcAPf
RFS [12]base24.713.423.131.423.714.222.9
CAB27.018.025.332.925.117.523.9
CH26.917.225.533.025.216.724.3
MH27.318.225.832.825.317.124.5
EQLv2 [33]base26.016.124.032.525.217.424.1
CAB28.120.426.333.526.019.524.9
CH27.318.225.833.125.717.724.8
MH27.117.325.733.225.617.524.7
SeeSaw [35]base27.318.226.532.326.919.626.8
CAB28.919.928.333.627.720.227.3
CH29.119.828.933.427.820.627.7
MH28.518.428.433.727.419.827.4
ECM [7]base27.717.726.633.127.219.626.6
CAB29.118.428.933.927.819.128.0
CH28.718.028.333.727.819.627.7
MH28.618.628.233.527.719.827.7
+ +into 4 evaluation subsets according to the number of training instances per class, with bins of [1, 20), [20, 400), [400, 8000) and [8000, -), respectively. + +Training details. We reproduce four different methods as our baselines, including RFS [12], EQLv2 [33], SeeSaw [35] and ECM [7], following their default experiment settings. We employ MMDetection [5] as our detection framework to conduct our experiment, and train detection models of Faster-RCNN, Mask-RCNN and Cascade R-CNN for 1x or 2x scheduler (except Swin-Transformer based detectors), following previous works [7, 17]. The batch size and learning rate are set as 16 and 0.02, and the data augmentation strictly follows previous long-tailed detection methods [7, 33, 35]. During training, we use FP16 mixed precision training and the warmup strategy to stabilize the learning process. For the evaluation metrics, we adopt AP and $\mathrm{AP}^b$ for instance segmentation and object detection, respectively, and adopt $\mathrm{AP}_1^b$ , $\mathrm{AP}_2^b$ , $\mathrm{AP}_3^b$ and $\mathrm{AP}_4^b$ on COCOLT, corresponding to its 4 different subsets. With the suggested practice in LVIS official website, we run all our experiment 3 times on 8 RTX3090 GPUs to reduce the variance. Please refer to our supplementary material for more detailed information. + +# 4.2 LVIS detection and segmentation + +Consistent improvements. we first evaluate the effectiveness of our method on the LVIS1.0 dataset by combing the proposed approach 'adding a class-agnostic + +Table 4: Comparison with state-of-the-art methods on the LVIS1.0 dataset. The results of 'Our' means our proposed CAB regression method combined with the SeeSaw loss ('SeeSaw + CAB', cf. Sec. 4). We show the metrics of instance segmentation AP and object detection $\mathrm{AP}^b$ . The results with a $\dagger$ are copied from [7], while others are reproduced by us using their official released code. + +
ArchitectBackboneMethodAPAPrAPfAPcAPb
Mask-RCNN [14]R50-FPNCE18.70.416.529.319.7
RFS [13]23.714.222.929.324.7
BSCE [29]24.415.723.629.125.5
EQLv2 [33]25.217.424.129.926.0
ECM [7]†27.419.727.031.127.9
ECM27.219.626.631.327.7
SeeSaw [35]26.919.626.830.527.3
ROG [42]26.920.126.830.027.2
Our27.720.227.331.328.9
Mask-RCNN [14]R101-FPNCE [7]†25.516.624.530.626.6
EQLv2 [33]†27.220.625.931.427.9
ECM [7]†28.721.927.932.329.4
ECM28.620.928.332.229.3
SeeSaw [35]28.220.328.131.829.0
Our29.021.028.932.430.7
Cascade R-CNN [2]Swin-TECM [7]34.124.034.938.037.6
SeeSaw [35]34.224.634.737.837.8
Our34.624.735.338.138.2
Cascade R-CNN [2]Swin-BECM [7]†39.733.540.641.443.6
Our39.934.640.741.244.2
+ +branch' (CAB) with existing long-tailed object detection methods. Since our main focus is on bounding box regression, we list the object detection results in early columns and segmentation results in later ones. As shown in Table 3, using CAB leads to consistent $\mathrm{AP}^b$ and AP improvement over existing classification-based methods, surpassing all of them with large margins. For object detection, Our CAB benefits the rare class a lot, with an increase of $4.6\mathrm{AP}^b$ and $4.3\mathrm{AP}^b$ on RFS and EQLv2. The same is true for instance segmentation, where a growing trend can be observed on all metrics, showing that CAB is also beneficial for later mask pixel predictions. Interestingly, the method 'RFS+CAB' (which uses a CE loss) can almost achieve the same object detection accuracy as the SeeSaw method, and surpasses EQLv2 for about $1\mathrm{AP}^b$ . We thus conjecture that: besides merely focusing on classification, our regression methods can serve as strong alternatives that also strongly boost long-tailed detection accuracy. + +Comparison with SOTA. We then compare our method with state-of-the-art methods using different object detection framework (Mask-RCNN, Cascade R-CNN) and backbones (ResNet-50, ResNet-101, Swin-T and Swin-B). Note + +Table 5: More evaluation metrics in LVIS1.0 dataset with a 1x scheduler using Mask RCNN. AP ${}_{\text{boundary }}^{fixed}$ and ${\mathrm{{AP}}}_{\text{box }}^{fixed}$ are two newly proposed challenging metrics [6,9]. + +
MethodAPAPbAPfixedboundaryAPfixedbbox
BAGS [24]23.123.7-26.2
EQLv2 [33]23.924.020.325.9
SeeSaw [35]25.225.419.826.5
ECM [7]26.326.721.427.4
Our26.928.022.128.2
+ +that 'Our' means 'SeeSaw + CAB'. For fair comparison, we reproduce majority of existing methods using their official released code unless specialized symbols $(e.g.,^{\dagger})$ appears after a method's name. As can be seen in Table 4, our method achieves the overall highest accuracy in AP and $\mathrm{AP}^b$ . For ResNet series models, our regression technique easily surpass the best competitor ECM [7], especially in $\mathrm{AP}^b$ (an increase of $1.2\mathrm{AP}^b$ for ResNet-50 and $1.1\mathrm{AP}^b$ for ResNet-101). The advantage also holds for ViT-based object detectors, where we surpassed the best competitor ECM [7] with both Swin-Tiny and Swin-Base backbone architectures. Following LVIS's common practice, we didn't list $\mathrm{AP}_r^b$ and $\mathrm{AP}_c^b$ here, but we want to emphasize that the advantage of our regression methods can be further enlarged when more metrics are listed (cf. Table 3). This is also true if we replace 'adding a Class-Agnostic Branch' (CAB) with the 'merging heads' regression alternative (cf. the best accuracy in Table 2). + +# 4.3 Generalization ability + +In this section, we will show the generalization ability of our regression methods in various aspects, including different evaluation metrics, datasets and its benefits on the mask branch. + +Different metrics. We first explore how different metrics affect our model's accuracy. Two additional metrics are $\mathrm{AP}_{\text {boundary }}$ (a more strict metric in instance segmentation [6]) and $\mathrm{AP}_{\text {bbox }}^{\text {fixed }}$ (constraining 10,000 predicted bounding boxes per class across the dataset [9]). Object detector that obtain decent results in traditional metrics may not perform as well in these criterions. As shown in Table 5, our regression methods adapts well and surpasses all existing methods on both traditional and these challenging metrics. + +The COCO-LT dataset. We also transfer our regression to another long-tailed dataset: COCO-LT. This is an artificially sampled [36] subset of original COCO [25]. We calculate the over bounding box metrics $\mathrm{AP}^b$ and more fine-grained results: $\mathrm{AP}_1^b$ , $\mathrm{AP}_2^b$ , $\mathrm{AP}_3^b$ and $\mathrm{AP}_4^b$ (ranging from the rarest to the most frequent class). As shown in Table 6, our CAB leads to consistent improvements on all metrics (especially for the rarest categories $\mathrm{AP}_1^b$ ) under different repeat-factor-sampling rates, showing the great power of CAB to help the rare + +Table 6: Experiments on COCO-LT [36]. Following LVIS [12], we use a similar repeat-factor-sampling (RFS) strategy with different sampling rates (3e-3, 5e-3 and 1e-2). + +
Rate +CABAP1bAP2bAP3bAP4bAPb
3e-3✓2.316.926.730.423.5
6.219.127.030.524.4
5e-3✓3.619.326.730.624.3
7.520.027.330.725.0
1e-2✓8.720.227.830.325.2
10.821.628.030.525.8
+ +Table 7: Experiments on the (relatively) balanced COCO dataset. We adopted the Faster RCNN [30] detector and tried three different backbones. + +
Backbone+CABAP\( AP_{50} \)\( AP_{75} \)\( AP_s \)\( AP_m \)\( AP_l \)
Res-50✓37.358.340.321.741.048.2
38.358.942.122.441.450.4
Res-101✓39.460.343.022.943.451.0
39.960.343.622.743.752.8
Res-32x4d✓41.062.244.623.945.352.9
41.762.545.423.645.954.8
+ +classes. In fact, this conclusion generally holds true for all sampling rates in our experiments. We only listed three here for simplicity and clarity. + +On balanced training set. Furthermore, we validate how our method performs on the relatively balanced dataset MS-COCO2017. We adopted Faster RCNN with three different backbones (ResNet-50, ResNet-101 and ResNext101-32x4d). Results in Table 7 clearly show that CAB generalizes well to datasets with more balanced distributions. Interestingly, CAB boosts metrics of large and medium objects while shows similar accuracy on small objects. This is possibly because that small objects occupy over $60\%$ of the total instances while large objects only consumes roughly $15\%$ [32]. Since CAB is more beneficial to less-frequent classes, it brings higher gains in large objects than in small ones. + +The mask branch. Finally, we apply our CAB to the segmentation branch to test whether adding an class-agnostic prior is suitable for mask prediction. Experimental results are in Table 8, where we add a class-agnostic mask prediction head and combines it with each class-specific mask head. As shown in the table, our CAB generalizes well to segmentation tasks. Our main experimental results may be further improved if CAB was applied in both box and mask predictions. + +Table 8: Generalization to mask prediction. We added a class-agnostic mask prediction branch and combines it with each class-specific mask prediction results. + +
Branch +CABAPAPrAPbAPr
Box✓23.714.224.713.4
25.117.527.018.0
Mask✓23.714.224.713.4
24.114.625.013.9
+ +# 4.4 Discussion of our three remedies + +Now we further discuss the proposed three methods (CAB, CH and MH) in detail, hoping to provide a more in-depth understanding of our hypothesis. + +Relationship of CAB to the objectness branch. At start, we want to clarify the differences between our CAB and the objectness branch approach [7, 35] adopted in the classification head. We have argued that although they seemingly share similar structures with ours, they are essentially different. First, the purpose of objectness branch is to deal with the imbalanced distribution between foreground and background classification samples [35], while our CAB aim to tackle the regression bias. Second, unlike classification where each class must preserve its own classifier, the regression heads can be clustered or merged, which provide more diverse solutions to reduce the regression bias besides CAB (cf. Table 2). Lastly, the effect of the objectness branch is yet to be proved because it will decrease the performance of rare classes even when combined with CE loss [35], while CAB leads to consistent improvements, especially in rare class. + +How many clusters are the best? Attentive readers might have questions why larger clusters K (such as 100) in CH achieve better results in Table 2b for rare class, since enlarging K means the (useful) class-agnostic information is reduced. We thus empirically verify that: besides objects numbers per group, the scale variance could potentially affect the final AP. We test CH with 'object scales' in more K. For each K, we calculate standard deviation of scales in each group, average, and get representative scale std. The same is done for average object. As Fig. 3a shows: as K decreases, training objects per group increases, which benefit rare classes. But, the variation of scales increases a lot, which makes regression head learning more difficult—imagine one head to fit offsets that vary a lot, and we also observed higher training loss for smaller K. Hence, $\max AP_r^b$ is not at $K = 1$ or 10, but at $K = 50$ in our experiments (cf. Fig. 3b). + +Why merging 'Common' class is the best?. We observe in Table 2c that merging common classes leads to the best improvement in $\mathrm{AP}_r$ , which is somehow counter-intuitive. We conjecture that this is due to the partition shift of rare, common and frequent in LVIS train and validation sets. In the LVIS1.0 train dataset, the size ranges for frequent, common and rare are [0, 404], [405, 865] and [866, 1202], respectively. While for the validation set, they become [0, 212], [213, 536] and [537, 1202], respectively (cf. Fig. 1c). Hence, when we + +![](images/84c82224482063638ab8a7c3ea99fe566a81d31ec2e85c041f63c656378608d4.jpg) +Fig. 3: The statistics (averaged standard deviation, averaged objects per group) of different K and its corresponding AP box. + +![](images/3e8c2b949e449956d09858474ec0f3d9bcde517125df2e5d907997c6849105f2.jpg) + +![](images/8a80c217608a3d3fdc3e34bd446db360da2a9eedb0590cf9bb416e8b5ee0a547.jpg) +Fig. 4: Visualizations of detection results before (in the left of each group) and after (in the right) using our CAB. We adopted RFS [12] as the baseline in LVIS1.0 and combine it with our CAB regression method. In comparison, the proposed method is good at detecting missing objects, filtering duplicated objects away, as well as rectifying bounding box predictions. This figure needs to be viewed in color. + +use training set statistics to merge the common class, a large portion of rare categories in the validation set are also potentially merged, thus contributing to the validation improvement on both $\mathrm{AP}_r$ and $\mathrm{AP}_c$ . + +# 4.5 Visualization and ablation + +Flatten distribution. We first plot the RCNN regression loss of each category before and after combining with our regression CAB. The baseline methods we choose are EQLv2 and CE. As shown in Fig. 5, the regression loss for rare has seen a noticeable drop after adding our CAB, and the overall loss distribution has become much more balanced, too. We thus believe that the pipeline of our regression remedies can relieve the regression bias, further verifying that our hypothesis in Sec. 1 in indeed valid. + +More accurate box/mask. We then calculate AP50-AP95 of boxes and masks to find whether our CAB behaves well on stricter IoU thresholds. As show in Fig. 5c-5d, adding CAB achieves consistent accuracy gains over all IoU thresholds for both box and mask prediction, and is especially helpful for those hard IoU threshold (e.g., AP75-90) in box evaluation (cf. Fig. 5c). Since a higher threshold requires more precise box prediction, these results have shown that our regression methods is capable of producing precise boxes with better quality. + +![](images/9c1c9872b6b3fb04803c144e4903357762f5c4e2c4511321a81c709ffb11410c.jpg) +(a) w/o CAB + +![](images/af6b51a96e98acf6c4fc699c7eb46616b82bae858e2b38b01d02dc6d4fc08d62.jpg) +(b) w/CAB +Fig. 5: The loss distribution shift before and after combining with our CAB (Fig. 5a, 5b), and AP improvement of combining our CAB with baseline in LVIS (Fig. 5c, 5d). + +![](images/66697af2349b380f646825e9f6a47b6feb50d1bb2a16c978d59c70f5e58c5096.jpg) +(c) box delta + +![](images/3b99aeb85f883781aa25c1e924f5c2ca04c72c54ec401407a61c7a055c18345f.jpg) +(d) mask delta + +Qualitative results. Last but not least, we provide example detection images sampled from the LVIS1.0 validation dataset. For simplicity, we choose RFS [12] with CE loss as baseline and combine it with our CAB method. As shown in Fig. 4, both baseline and our CAB detect most of the objects in an image, but CAB generally caters for more details. For example, CAB can help discover missed boxes, like the cabinet in the lower left images. It is also clearly illustrated that CAB helps filter duplicate boxes away in the final predictions (e.g., the pictures with elephant and cow with grass background). Since CAB brings in better predictions, duplicate boxes will have larger overlap that may be suppressed by NMS (non-maximum suppression). If we zoom this figure (cf. the last column in Fig. 4), we will find that CAB is capable of rectifying the predicted boxes, and this can empirically explain why our CAB achieves much better AP under higher and more difficult IoU thresholds (cf. Fig. 5c). + +# 5 Conclusions and Limitations + +In this paper, we discovered that the regression bias (imbalanced regression loss distribution on the RCNN head) exists in long-tailed object detection, and adversely affects detection results. We thus proposed three remedies for rectifying the regression bias. The proposed method significantly boosts the performance of rare class $\mathrm{AP}^b$ , and achieves state-of-the-art results. We also generalize our regression methods to balanced dataset, different evaluation metrics and the mask branch. Finally, visualizations show that our method indeed produces better predicted bounding boxes. + +As for the limitations, it remains unclear why the boosted accuracy becomes lower when the baseline are higher (e.g., the improvement of CAB on ECM is lower than that on RFS, cf. Table 3). This may relate to the upper bound a backbone model can achieve. Since large network generally better fit the dataset (cf. Table 4), the performance a ResNet-50 can achieve is limited. We thus call for involving larger vision models on the LVIS dataset to better handle the long-tailed object detection problem. Besides involving stronger vision backbones, we will try to testify the adaptability of the proposed methods when different pretraining information is integrated into the backbone network (e.g. self-supervised learning), which is left as our future work. + +# Acknowledgements + +We thank Yin-Yin He (M.Sc. degree in Nanjing University, now in ByteDance) for providing helpful discussions in long-tailed object detection. + +# References + +1. Alexandridis, K.P., Deng, J., Nguyen, A., Luo, S.: Long-tailed instance segmentation using gumbel optimized loss. In: ECCV. Lecture Notes in Computer Science, vol. 13670, pp. 353-369. Springer (2022) +2. Cai, Z., Vasconcelos, N.: Cascade r-cnn: Delving into high quality object detection. In: CVPR. pp. 6154-6162 (2018) +3. Cao, K., Wei, C., Gaidon, A., Arechiga, N., Ma, T.: Learning imbalanced datasets with label-distribution-aware margin loss. In: NeurIPS. pp. 1565-1576 (2019) +4. Caron, M., Bojanowski, P., Joulin, A., Douze, M.: Deep clustering for unsupervised learning of visual features. In: Ferrari, V., Hebert, M., Sminchisescu, C., Weiss, Y. (eds.) ECCV. Lecture Notes in Computer Science, vol. 11218, pp. 139-156. Springer (2018) +5. Chen, K., Wang, J., Pang, J., Cao, Y., Xiong, Y., Li, X., Sun, S., Feng, W., Liu, Z., Xu, J., Zhang, Z., Cheng, D., Zhu, C., Cheng, T., Zhao, Q., Li, B., Lu, X., Zhu, R., Wu, Y., Dai, J., Wang, J., Shi, J., Ouyang, W., Loy, C.C., Lin, D.: MMDetection: Open mmlab detection toolbox and benchmark. arXiv preprint arXiv:1906.07155 (2019) +6. Cheng, B., Girshick, R., Dollar, P., Berg, A.C., Kirillov, A.: Boundary iou: Improving object-centric image segmentation evaluation. In: CVPR. pp. 15334-15342 (2021) +7. Cho, J.H., Krahenbuhl, P.: Long-tail detection with effective class-margins. In: ECCV. Lecture Notes in Computer Science, vol. 13668, pp. 698-714. Springer (2022) +8. Cui, Y., Jia, M., Lin, T.Y., Song, Y., Belongie, S.: Class-balanced loss based on effective number of samples. In: CVPR. pp. 9268-9277 (2019) +9. Dave, A., Dóllár, P., Ramanan, D., Kirillov, A., Girshick, R.: Evaluating large-vocabulary object detectors: The devil is in the details. arXiv preprint arXiv:2102.01066 (2021) +10. Du, Y., Wu, J.: No one left behind: Improving the worst categories in long-tailed learning. In: CVPR. pp. 15804-15813 (2023) +1. Feng, C., Zhong, Y., Huang, W.: Exploring classification equilibrium in long-tailed object detection. In: ICCV. pp. 3417-3426 (2021) +2. Gupta, A., Dollar, P., Girshick, R.: Lvis: A dataset for large vocabulary instance segmentation. In: CVPR. pp. 5356-5364 (2019) +3. Gupta, A., Dollar, P., Girshick, R.: Lvis: A dataset for large vocabulary instance segmentation. In: CVPR. pp. 5356-5364 (2019) +4. He, K., Gkioxari, G., Dollár, P., Girshick, R.: Mask R-CNN. In: ICCV. pp. 2961-2969 (2017) +5. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 770-778 (2016) +6. He, Y.Y., Wu, J., Wei, X.S.: Distilling virtual examples for long-tailed recognition. In: ICCV. pp. 235-244 (2021) + +17. He, Y.Y., Zhang, P., Wei, X.S., Zhang, X., Sun, J.: Relieving long-tailed instance segmentation via pairwise class balance. In: CVPR. pp. 7000-7009 (2022) +18. Hsieh, T., Robb, E., Chen, H., Huang, J.: Droploss for long-tail instance segmentation. In: AAAI. pp. 1549-1557 (2021) +19. Hu, X., Jiang, Y., Tang, K., Chen, J., Miao, C., Zhang, H.: Learning to segment the tail. In: CVPR (2020) +20. Kang, B., Xie, S., Rohrbach, M., Yan, Z., Gordo, A., Feng, J., Kalantidis, Y.: Decoupling representation and classifier for long-tailed recognition. In: ICLR (2020) +21. Li, B., Yao, Y., Tan, J., Zhang, G., Yu, F., Lu, J., Luo, Y.: Equalized focal loss for dense long-tailed object detection. In: CVPR. pp. 6990-6999 (2022) +22. Li, M., Cheung, Y.m., Lu, Y.: Long-tailed visual recognition via gaussian clouded logit adjustment. In: CVPR. pp. 6929-6938 (2022) +23. Li, T., Wang, L., Wu, G.: Self supervision to distillation for long-tailed visual recognition. In: ICCV. pp. 630-639 (2021) +24. Li, Y., Wang, T., Kang, B., Tang, S., Wang, C., Li, J., Feng, J.: Overcoming classifier imbalance for long-tail object detection with balanced group softmax. In: CVPR (2020) +25. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dólár, P., Zitnick, C.L.: Microsoft COCO: Common objects in context. In: ECCV. Lecture Notes in Computer Science, vol. 8693, pp. 740-755. Springer (2014) +26. Liu, Z., Miao, Z., Zhan, X., Wang, J., Gong, B., Yu, S.X.: Large-scale long-tailed recognition in an open world. In: CVPR. pp. 2537-2546 (2019) +27. Menon, A.K., Jayasumana, S., Rawat, A.S., Jain, H., Veit, A., Kumar, S.: Long-tail learning via logit adjustment. In: ICLR (2021) +28. Pang, J., Chen, K., Shi, J., Feng, H., Ouyang, W., Lin, D.: Libra r-cnn: Towards balanced learning for object detection. In: CVPR. pp. 821-830 (2019) +29. Ren, J., Yu, C., sheng, s., Ma, X., Zhao, H., Yi, S., Li, h.: Balanced meta-softmax for long-tailed visual recognition. In: NeurIPS. pp. 4175-4186 (2020) +30. Ren, S., He, K., Girshick, R., Sun, J.: Faster R-CNN: Towards real-time object detection with region proposal networks. In: NeurIPS. pp. 91-99 (2015) +31. Russakovsky, O., Deng, J., Su, H., Krause, J., Satheesh, S., Ma, S., Huang, Z., Karpathy, A., Khosla, A., Bernstein, M., Berg, A., Fei-Fei, L.: ImageNet large scale visual recognition challenge. International Journal of Computer Vision 115(3), 211-252 (2015) +32. Singh, B., Davis, L.S.: An analysis of scale invariance in object detection SNIP. In: CVPR. pp. 3578-3587 (2018) +33. Tan, J., Lu, X., Zhang, G., Yin, C., Li, Q.: Equalization loss v2: A new gradient balance approach for long-tailed object detection. In: CVPR. pp. 1685-1694 (2021) +34. Tan, J., Wang, C., Li, B., Li, Q., Ouyang, W., Yin, C., Yan, J.: Equalization loss for long-tailed object recognition. In: CVPR (2020) +35. Wang, J., Zhang, W., Zang, Y., Cao, Y., Pang, J., Gong, T., Chen, K., Liu, Z., Loy, C.C., Lin, D.: Seesaw loss for long-tailed instance segmentation. In: CVPR. pp. 9695-9704 (2021) +36. Wang, T., Li, Y., Kang, B., Li, J., Liew, J.H., Tang, S., Hoi, S.C.H., Feng, J.: The devil is in classification: A simple framework for long-tail instance segmentation. In: ECCV. Lecture Notes in Computer Science, vol. 12359, pp. 728-744. Springer (2020) +37. Wang, T., Zhu, Y., Zhao, C., Zeng, W., Wang, J., Tang, M.: Adaptive class suppression loss for long-tail object detection. In: CVPR. pp. 3103-3112 (2021) +38. Wang, X., Lian, L., Miao, Z., Liu, Z., Yu, S.: Long-tailed recognition by routing diverse distribution-aware experts. In: ICLR (2021) + +39. Wu, J., Song, L., Wang, T., Zhang, Q., Yuan, J.: Forest R-CNN: large-vocabulary long-tailed object detection and instance segmentation. In: ACM MM. pp. 1570-1578 (2020) +40. Xuan, G., Zhang, W., Chai, P.: EM algorithms of gaussian mixture model and hidden markov model. In: ICIP. pp. 145-148 (2001) +41. Zhang, C., Pan, T.Y., Li, Y., Hu, H., Xuan, D., Changpinyo, S., Gong, B., Chao, W.L.: Mosaicos: A simple and effective use of object-centric images for long-tailed object detection. In: ICCV. pp. 417-427 (2021) +42. Zhang, S., Chen, C., Peng, S.: Reconciling object-level and global-level objectives for long-tail detection. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 18982-18992 (2023) +43. Zhong, Z., Cui, J., Liu, S., Jia, J.: Improving calibration for long-tailed recognition. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 16489-16498 (2021) +44. Zhou, B., Cui, Q., Wei, X.S., Chen, Z.M.: Bbn: Bilateral-branch network with cumulative learning for long-tailed visual recognition. In: CVPR. pp. 9716-9725 (2020) +45. Zhu, K., Fu, M., Wu, J.: Multi-label self-supervised learning with scene images. In: ICCV. pp. 6694-6703 (2023) +46. Zhu, K., He, Y., Wu, J.: Quantized feature distillation for network quantization. In: AAAI. pp. 11452-11460 (2023) \ No newline at end of file diff --git a/2024/Rectify the Regression Bias in Long-Tailed Object Detection/images.zip b/2024/Rectify the Regression Bias in Long-Tailed Object Detection/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..7fe811325aaf190b226a8a355b493835200c2cae --- /dev/null +++ b/2024/Rectify the Regression Bias in Long-Tailed Object Detection/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:44c12e4fc375ba6e5dd5f63fde10b8858623f5eb9d8fe1e0536df9d0cc32fb7e +size 536854 diff --git a/2024/Rectify the Regression Bias in Long-Tailed Object Detection/layout.json b/2024/Rectify the Regression Bias in Long-Tailed Object Detection/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..bc335c6a85a7ce20c2acdb2bd9b2ccf006426383 --- /dev/null +++ b/2024/Rectify the Regression Bias in Long-Tailed Object Detection/layout.json @@ -0,0 +1,8929 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 157, + 111, + 457, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 157, + 111, + 457, + 148 + ], + "spans": [ + { + "bbox": [ + 157, + 111, + 457, + 148 + ], + "type": "text", + "content": "Rectify the Regression Bias in Long-Tailed Object Detection" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 153, + 168, + 459, + 181 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 153, + 168, + 459, + 181 + ], + "spans": [ + { + "bbox": [ + 153, + 168, + 459, + 181 + ], + "type": "text", + "content": "Ke Zhu, Minghao Fu, Jie Shao, Tianyu Liu, and Jianxin Wu" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 198, + 190, + 414, + 201 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 198, + 190, + 414, + 201 + ], + "spans": [ + { + "bbox": [ + 198, + 190, + 414, + 201 + ], + "type": "text", + "content": "State Key Laboratory for Novel Software Technology" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 234, + 202, + 378, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 234, + 202, + 378, + 213 + ], + "spans": [ + { + "bbox": [ + 234, + 202, + 378, + 213 + ], + "type": "text", + "content": "Nanjing University, Nanjing, China" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 188, + 213, + 424, + 223 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 188, + 213, + 424, + 223 + ], + "spans": [ + { + "bbox": [ + 188, + 213, + 424, + 223 + ], + "type": "text", + "content": "School of Artificial Intelligence, Nanjing University, China" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 156, + 224, + 457, + 235 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 156, + 224, + 457, + 235 + ], + "spans": [ + { + "bbox": [ + 156, + 224, + 457, + 235 + ], + "type": "text", + "content": "{zhuk, fumh, shaoj, liuty}@lamda.nju.edu.cn, wujx2001@nju.edu.cn" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 159, + 267, + 453, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 267, + 453, + 453 + ], + "spans": [ + { + "bbox": [ + 159, + 267, + 453, + 453 + ], + "type": "text", + "content": "Abstract. Long-tailed object detection faces great challenges because of its extremely imbalanced class distribution. Recent methods mainly focus on the classification bias and its loss function design, while ignoring the subtle influence of the regression branch. This paper shows that the regression bias exists and does adversely and seriously impact the detection accuracy. While existing methods fail to handle the regression bias, the class-specific regression head for rare classes is hypothesized to be the main cause of it in this paper. As a result, three kinds of viable solutions to cater for the rare categories are proposed, including adding a class-agnostic branch, clustering heads and merging heads. The proposed methods brings in consistent and significant improvements over existing long-tailed detection methods, especially in rare and common classes. The proposed method achieves state-of-the-art performance in the large vocabulary LVIS dataset with different backbones and architectures. It generalizes well to more difficult evaluation metrics, relatively balanced datasets, and the mask branch. This is the first attempt to reveal and explore rectifying of the regression bias in long-tailed object detection." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 475, + 230, + 488 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 475, + 230, + 488 + ], + "spans": [ + { + "bbox": [ + 132, + 475, + 230, + 488 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 501, + 482, + 573 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 501, + 482, + 573 + ], + "spans": [ + { + "bbox": [ + 130, + 501, + 482, + 573 + ], + "type": "text", + "content": "Long-tailed object detection draws great attention [12] for its practical utility recently. Numerous efforts [1,7,34,35,37] have been made to tackle this challenging task, such as re-weighting [33,35,37], over-sampling [11,12,39], and balanced grouping [19,24]. These methods are proposed to prevent the tail classes from being overwhelmed due to discouraging gradients [33, 37], inferior predicting scores [11,24] or insufficient samples [11,28]." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 574, + 482, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 574, + 482, + 647 + ], + "spans": [ + { + "bbox": [ + 130, + 574, + 482, + 647 + ], + "type": "text", + "content": "Long-tailed detection often involves both classification and regression heads. While almost all existing methods focus on mitigating the classification bias (e.g., adjusting the classification structure in detection branches), little or no attention has been paid to the regression branch. We will show in this paper that the regression bias has significant adverse effects in long-tailed object detection, but previous methods failed to identify this important issue." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 653, + 284, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 653, + 284, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 653, + 284, + 665 + ], + "type": "text", + "content": "* J. Wu is the corresponding author" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 143, + 114, + 231, + 184 + ], + "blocks": [ + { + "bbox": [ + 143, + 114, + 231, + 184 + ], + "lines": [ + { + "bbox": [ + 143, + 114, + 231, + 184 + ], + "spans": [ + { + "bbox": [ + 143, + 114, + 231, + 184 + ], + "type": "image", + "image_path": "cb55291d3410932d064ef0546897379e5704c6b2049bd8f4de0a558c7c6a3032.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 162, + 186, + 216, + 196 + ], + "lines": [ + { + "bbox": [ + 162, + 186, + 216, + 196 + ], + "spans": [ + { + "bbox": [ + 162, + 186, + 216, + 196 + ], + "type": "text", + "content": "(a) RCNN loss" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 236, + 114, + 325, + 185 + ], + "blocks": [ + { + "bbox": [ + 236, + 114, + 325, + 185 + ], + "lines": [ + { + "bbox": [ + 236, + 114, + 325, + 185 + ], + "spans": [ + { + "bbox": [ + 236, + 114, + 325, + 185 + ], + "type": "image", + "image_path": "86ed05ecb93ccadeb52977fa1c71d2cdc8ab58bb374d63f4c205230d844b63ab.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 258, + 186, + 307, + 196 + ], + "lines": [ + { + "bbox": [ + 258, + 186, + 307, + 196 + ], + "spans": [ + { + "bbox": [ + 258, + 186, + 307, + 196 + ], + "type": "text", + "content": "(b) RPN loss" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 205, + 482, + 248 + ], + "lines": [ + { + "bbox": [ + 130, + 205, + 482, + 248 + ], + "spans": [ + { + "bbox": [ + 130, + 205, + 482, + 248 + ], + "type": "text", + "content": "Fig. 1: 1a shows the RCNN regression train loss of frequent, common and rare categories. 1b shows the RPN regression loss. 1c shows distribution of per class mean object scales in LVIS1.0. 'delta' in 1c is the negative of difference between train and validation set sizes for different classes." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 331, + 115, + 471, + 184 + ], + "blocks": [ + { + "bbox": [ + 331, + 115, + 471, + 184 + ], + "lines": [ + { + "bbox": [ + 331, + 115, + 471, + 184 + ], + "spans": [ + { + "bbox": [ + 331, + 115, + 471, + 184 + ], + "type": "image", + "image_path": "ce00f1ff39dd760bd8410b0995ffbbf80c95a55a864f2c7e91201e52872539a9.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 331, + 186, + 470, + 196 + ], + "lines": [ + { + "bbox": [ + 331, + 186, + 470, + 196 + ], + "spans": [ + { + "bbox": [ + 331, + 186, + 470, + 196 + ], + "type": "text", + "content": "(c) Mean scale in LVIS train and val set" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 276, + 482, + 372 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 276, + 482, + 372 + ], + "spans": [ + { + "bbox": [ + 130, + 276, + 482, + 372 + ], + "type": "text", + "content": "Fig. 1 clearly showcases this issue. For different methods (EQLv2 [33], See-Saw [35], ECM [7] and CrossEntropy CE [12]) trained on the LVIS1.0 [12] dataset, we plot the regression branch's losses for the final detection RCNN head in Fig. 1a. It is clear that the regression losses of rare categories are significantly higher than those of frequent and common categories, which inevitably will lead to poor regression results (and hence detection results) for rare classes. We name this finding as the regression bias, but existing long-tailed detection methods all fail to deal with or even identify this issue." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 373, + 482, + 457 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 373, + 482, + 457 + ], + "spans": [ + { + "bbox": [ + 130, + 373, + 482, + 457 + ], + "type": "text", + "content": "To further demonstrate the importance of regression, we calculate the class-wise mean scale of GT (groundtruth) boxes in LVIS train and validation sets, as well as their differences (cf. Fig. 1c). The scale shift of rare classes is much larger than that of frequent classes. Since regression is highly correlated with box scale [30], it is thus inherently difficult for a rare class to learn a good bounding box regressor with both few samples and large scale shift. In short, it is crucial to properly handle the regression bias in long-tailed object detection." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 459, + 482, + 507 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 459, + 482, + 507 + ], + "spans": [ + { + "bbox": [ + 130, + 459, + 482, + 507 + ], + "type": "text", + "content": "Our solution to rectify the regression bias is motivated by Fig. 1b. We find that the regression loss in RPN is balanced where rare, common and frequent categories have almost the same regression loss, which is almost immune to the regression bias when comparing Fig. 1b with Fig. 1a." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 508, + 482, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 508, + 482, + 567 + ], + "spans": [ + { + "bbox": [ + 130, + 508, + 482, + 567 + ], + "type": "text", + "content": "The key difference between RPN and RCNN regression is that the former is class-agnostic (i.e., all classes share the same regression parameters), while the latter is class-specific. Then, one natural question arises: Can class-agnostic head improve a tail class's generalization ability and handle the regression bias? Our hypothesis is that rare classes do favor class-agnostic regression heads." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 570, + 482, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 667 + ], + "type": "text", + "content": "Our hypothesis is supported by experiments in Table 1, where we compared class-specific and -agnostic regression heads in RCNN. By replacing the classification head with groundtruth class labels, experiments in Table 1 disentangled the impact of the classification head and focus on the regression heads. It is clear that the class-agnostic head possesses substantial advantages: " + }, + { + "bbox": [ + 130, + 570, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\mathrm{AP}_r^b" + }, + { + "bbox": [ + 130, + 570, + 482, + 667 + ], + "type": "text", + "content": " (for rare) surges from 0.7 to 54.6, even surpassing " + }, + { + "bbox": [ + 130, + 570, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\mathrm{AP}_f^b" + }, + { + "bbox": [ + 130, + 570, + 482, + 667 + ], + "type": "text", + "content": " (for frequent classes)! However, the agnostic head will bring a small drop in frequent classes (e.g., from 40.7 to 40.0 in " + }, + { + "bbox": [ + 130, + 570, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\mathrm{AP}_f^b" + }, + { + "bbox": [ + 130, + 570, + 482, + 667 + ], + "type": "text", + "content": "). The final version of our conjecture is: the rare (and possibly" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "type": "text", + "content": "K. Zhu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 152 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 152 + ], + "type": "text", + "content": "common) classes indeed favor class-agnostic regression, while the frequent classes prefers a class-specific regression, and there should be a trade-off between the two to optimize for all three types of categories." + } + ] + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 219, + 228, + 392, + 302 + ], + "blocks": [ + { + "bbox": [ + 130, + 173, + 482, + 217 + ], + "lines": [ + { + "bbox": [ + 130, + 173, + 482, + 217 + ], + "spans": [ + { + "bbox": [ + 130, + 173, + 482, + 217 + ], + "type": "text", + "content": "Table 1: Results on LVIS1.0 based on the CE baseline. We trained a Mask-RCNN R50-FPN detector [14] with class-specific (the default setting) and class-agnostic regression heads in the RCNN head. Please note that during test, the predicted proposals were provided groundtruth (GT) classification results." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 219, + 228, + 392, + 302 + ], + "lines": [ + { + "bbox": [ + 219, + 228, + 392, + 302 + ], + "spans": [ + { + "bbox": [ + 219, + 228, + 392, + 302 + ], + "type": "table", + "html": "
Reg. HeadGTAPAPbAPrAPpcAPfb
Specificno18.719.70.316.531.7
yes37.239.834.141.340.7
Agnosticno18.019.30.715.531.5
yes38.445.654.647.040.0
", + "image_path": "27ecd21c59bee80b0918849d16462b26f8b7b5827d5775457ef45cab397c25ee.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 325, + 482, + 480 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 325, + 482, + 480 + ], + "spans": [ + { + "bbox": [ + 130, + 325, + 482, + 480 + ], + "type": "text", + "content": "Accordingly, we design three different methods to fully rectify the regression bias, including adding a class-agnostic head, clustering similar heads, or merging heads. All three methods bring in positive effects (cf. Table 2), which verifies that rectifying regression bias is indeed crucial. We choose to adopt 'adding a class-agnostic head' in our main experiment for its simplicity, which leads to consistent and significant improvements over previous long-tailed detection pipeline, and has achieved state-of-the-art performance with various backbones and architectures. Moreover, our method shows robust generalization ability (cf. Tables 5-8) under varying settings, including different datasets (COCO/COCOLT [36]), different evaluation metrics and even adapts to the mask branch design. Furthermore, visualizations show that the proposed method leads to more accurate bounding box predictions (cf. Fig. 4) and indeed alleviates the regression bias (cf. Fig. 5). In summary, our contributions are:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 135, + 488, + 480, + 572 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 135, + 488, + 480, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 488, + 480, + 510 + ], + "spans": [ + { + "bbox": [ + 135, + 488, + 480, + 510 + ], + "type": "text", + "content": "1. For the first time, we reveal and successfully handle the regression bias in long-tailed object detection." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 135, + 513, + 480, + 536 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 513, + 480, + 536 + ], + "spans": [ + { + "bbox": [ + 135, + 513, + 480, + 536 + ], + "type": "text", + "content": "2. We propose three remedies to alleviate this bias, all of which produce consistent improvements over existing methods." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 135, + 536, + 480, + 572 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 536, + 480, + 572 + ], + "spans": [ + { + "bbox": [ + 135, + 536, + 480, + 572 + ], + "type": "text", + "content": "3. Our method achieves state-of-the-art results on LVIS as well as generalizing across datasets, metrics and even the mask branch. Visualizations qualitatively verify our hypothesis, too." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 131, + 591, + 237, + 604 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 591, + 237, + 604 + ], + "spans": [ + { + "bbox": [ + 131, + 591, + 237, + 604 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 617, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 665 + ], + "type": "text", + "content": "Vast number of pipelines and techniques have been invented in the field of general purpose object detection. Although most of these detectors can possibly be adapted to balanced detection datasets [25], they can nevertheless hardly handle extremely long-tailed distribution. Dedicated methods are hence required." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 334, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 334, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 334, + 91, + 447, + 102 + ], + "type": "text", + "content": "Rectify the Regression Bias" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": "Modern long-tailed learning methods aim to solve two important tasks: image classification [31] and object detection/segmentation [12, 30]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 140, + 481, + 295 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 140, + 481, + 295 + ], + "spans": [ + { + "bbox": [ + 130, + 140, + 481, + 295 + ], + "type": "text", + "content": "The first task, long-tailed image classification [26], is well explored and displays great diversity in terms of specific methods [3,29], training pipelines [20,44], post-hoc analysis [22,27] and network architecture [16,23,38]. Early attempts in this family focused on the details of re-weighting [3] or re-sampling [8] techniques to provide a relatively balanced (data) distribution for tail classes. Later on, the decoupling pipeline [20] for long-tailed learning becomes popular, paving the way for various successors [43,44]. This decoupled paradigm are based on the insight that instance sampling is beneficial for representation, while the classifier needs to be calibrated to alleviate its bias. Some works try to unify long-tailed learning with theoretical analysis [22,27], which involves label distribution [29] and generalization bound [27]. Most recent methods are much more diverse in ideas: they involve knowledge distillation [16,38,46], self-supervised learning [23,45] or statistical approach [10] to better handle long-tailed recognition." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 296, + 482, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 296, + 482, + 437 + ], + "spans": [ + { + "bbox": [ + 130, + 296, + 482, + 437 + ], + "type": "text", + "content": "The combination of long-tailed learning and object detection or instance segmentation proves to be more challenging [12], because the class distribution in LVIS [12] is extremely imbalanced and naively applying common object detection techniques leads to un-satisfactory results [18, 24, 34, 35, 37]. Dominating solutions in this area are re-sampling [12, 39] and re-weighting [17, 21, 33, 35, 37]. There are variants of them that adopt balanced grouping [24], class incremental learning [19], augmented feature sampling [11, 36] or extra data source [41]. A common characteristic of these long-tailed detection methods is that they only focus on the classification task, and ignore the subtle influence of the regression branch. In this paper, we thoroughly explore the previously unvisited regression bias in long-tailed object detection, and propose a simple yet novel method to tackle long-tailed object detection." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 439, + 482, + 511 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 439, + 482, + 511 + ], + "spans": [ + { + "bbox": [ + 130, + 439, + 482, + 511 + ], + "type": "text", + "content": "In [36], there is an experiment with similar design as ours in Table 1. But, we emphasize that they are fundamentally different. On one hand, conclusions in [36] is that 'performance drop in LVIS is mainly caused by the proposal classification [36]' and hence they focused on classification. On the other hand, their class agnostic-head is a supplementary to support their classification hypothesis, while our design and experiments lead to a novel finding: the regression bias." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 529, + 203, + 542 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 529, + 203, + 542 + ], + "spans": [ + { + "bbox": [ + 132, + 529, + 203, + 542 + ], + "type": "text", + "content": "3 Method" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 555, + 481, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 555, + 481, + 579 + ], + "spans": [ + { + "bbox": [ + 130, + 555, + 481, + 579 + ], + "type": "text", + "content": "Now we elaborate three different remedies to alleviate the regression bias, and start from the background on long-tailed object detection." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 597, + 228, + 609 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 597, + 228, + 609 + ], + "spans": [ + { + "bbox": [ + 132, + 597, + 228, + 609 + ], + "type": "text", + "content": "3.1 Preliminaries" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "We take Faster-RCNN [14] as an example. For a scene image " + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": ", it is first fed into a backbone network " + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\phi(\\cdot)" + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": " (e.g., ResNet [15]) to get the image feature " + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "inline_equation", + "content": "f \\in \\mathbb{R}^{d \\times w \\times h}" + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": ": " + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "inline_equation", + "content": "f = \\phi(I)" + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": ", with dimensionality, width and height denoted as " + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": ", respectively. A region proposal network (RPN) [30] that contains both agnostic" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "type": "text", + "content": "K. Zhu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 143, + 113, + 471, + 213 + ], + "blocks": [ + { + "bbox": [ + 143, + 113, + 471, + 213 + ], + "lines": [ + { + "bbox": [ + 143, + 113, + 471, + 213 + ], + "spans": [ + { + "bbox": [ + 143, + 113, + 471, + 213 + ], + "type": "image", + "image_path": "48d9ce9b44f1f0519ef1709b476b10bded8ad48d8f78d5d48de7414b2e12ff0c.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 221, + 482, + 308 + ], + "lines": [ + { + "bbox": [ + 130, + 221, + 482, + 308 + ], + "spans": [ + { + "bbox": [ + 130, + 221, + 482, + 308 + ], + "type": "text", + "content": "Fig. 2: Illustration of the regular two-stage detection pipeline and the proposed regression methods. Previous methods (the left figure) generally focus on the final classification branch (the yellow arrow), while we focus on rectifying the regression bias (the red arrow). The right part shows our three regression methods, including adding an extra branch " + }, + { + "bbox": [ + 130, + 221, + 482, + 308 + ], + "type": "inline_equation", + "content": "W_{0}" + }, + { + "bbox": [ + 130, + 221, + 482, + 308 + ], + "type": "text", + "content": ", clustering regression heads (e.g., from " + }, + { + "bbox": [ + 130, + 221, + 482, + 308 + ], + "type": "inline_equation", + "content": "W_{1},\\ldots ,W_{C}" + }, + { + "bbox": [ + 130, + 221, + 482, + 308 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 130, + 221, + 482, + 308 + ], + "type": "inline_equation", + "content": "W_1',W_2'" + }, + { + "bbox": [ + 130, + 221, + 482, + 308 + ], + "type": "text", + "content": ") and merging (e.g., merging rare categories into one regression head " + }, + { + "bbox": [ + 130, + 221, + 482, + 308 + ], + "type": "inline_equation", + "content": "W_{\\text{rare}}" + }, + { + "bbox": [ + 130, + 221, + 482, + 308 + ], + "type": "text", + "content": ", cf. Table 2). In our main experiments, we choose 'adding an extra branch' for its simplicity. This figure needs to be viewed in color." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 334, + 480, + 369 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 334, + 480, + 369 + ], + "spans": [ + { + "bbox": [ + 130, + 334, + 480, + 369 + ], + "type": "text", + "content": "classification and regression branches is then applied on the feature tensor " + }, + { + "bbox": [ + 130, + 334, + 480, + 369 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 130, + 334, + 480, + 369 + ], + "type": "text", + "content": " to produce proposals " + }, + { + "bbox": [ + 130, + 334, + 480, + 369 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 130, + 334, + 480, + 369 + ], + "type": "text", + "content": " from pre-defined anchors. The ROAlign [14] then extracts proposal features:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 223, + 380, + 481, + 394 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 223, + 380, + 481, + 394 + ], + "spans": [ + { + "bbox": [ + 223, + 380, + 481, + 394 + ], + "type": "interline_equation", + "content": "p = \\operatorname {R P N} (f), \\quad f _ {p} = \\operatorname {R O I A l i g n} (f, p), \\tag {1}", + "image_path": "0f732739488f23ec64885e0f34efc49916c24a2a647b868d7da8ccc5fdd60a7e.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 402, + 482, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 402, + 482, + 462 + ], + "spans": [ + { + "bbox": [ + 130, + 402, + 482, + 462 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 402, + 482, + 462 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 130, + 402, + 482, + 462 + ], + "type": "text", + "content": " represents a large set of proposals, and " + }, + { + "bbox": [ + 130, + 402, + 482, + 462 + ], + "type": "inline_equation", + "content": "f_{p}" + }, + { + "bbox": [ + 130, + 402, + 482, + 462 + ], + "type": "text", + "content": " is the set of aligned proposal features. " + }, + { + "bbox": [ + 130, + 402, + 482, + 462 + ], + "type": "inline_equation", + "content": "f_{p}" + }, + { + "bbox": [ + 130, + 402, + 482, + 462 + ], + "type": "text", + "content": " goes into post-process modules (e.g., NMS [30]) before being sent to the RCNN head to get the final features set " + }, + { + "bbox": [ + 130, + 402, + 482, + 462 + ], + "type": "inline_equation", + "content": "f_{n}" + }, + { + "bbox": [ + 130, + 402, + 482, + 462 + ], + "type": "text", + "content": ": " + }, + { + "bbox": [ + 130, + 402, + 482, + 462 + ], + "type": "inline_equation", + "content": "f_{n} = \\mathrm{RCNN}(f_{p})" + }, + { + "bbox": [ + 130, + 402, + 482, + 462 + ], + "type": "text", + "content": ", which is fed to classification and regression branches (e.g., linear layers) to produce the prediction results." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 481, + 373, + 493 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 481, + 373, + 493 + ], + "spans": [ + { + "bbox": [ + 131, + 481, + 373, + 493 + ], + "type": "text", + "content": "3.2 Our three remedies for the regression bias" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 502, + 432, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 502, + 432, + 514 + ], + "spans": [ + { + "bbox": [ + 131, + 502, + 432, + 514 + ], + "type": "text", + "content": "In Faster RCNN, there is a dedicated regression head for each class:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 239, + 523, + 481, + 538 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 239, + 523, + 481, + 538 + ], + "spans": [ + { + "bbox": [ + 239, + 523, + 481, + 538 + ], + "type": "interline_equation", + "content": "r _ {i} = W _ {i} ^ {T} f _ {n}, \\quad i = 1, 2, \\dots , C, \\tag {2}", + "image_path": "43c454a18f164b8ab6b88217ef9827b266f80cb440271940c10ff93ab8986714.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 547, + 480, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 547, + 480, + 594 + ], + "spans": [ + { + "bbox": [ + 130, + 547, + 480, + 594 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 547, + 480, + 594 + ], + "type": "inline_equation", + "content": "r_i = (\\delta x_i, \\delta y_i, \\delta w_i, \\delta h_i)" + }, + { + "bbox": [ + 130, + 547, + 480, + 594 + ], + "type": "text", + "content": " represents the regression offset for class " + }, + { + "bbox": [ + 130, + 547, + 480, + 594 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 547, + 480, + 594 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 547, + 480, + 594 + ], + "type": "inline_equation", + "content": "W_i" + }, + { + "bbox": [ + 130, + 547, + 480, + 594 + ], + "type": "text", + "content": " is the class-specific regression head (a linear layer). We will now present our approaches to rectify the regression bias. The existing pipeline and our proposed methods are shown in Fig. 2." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 595, + 482, + 643 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 595, + 482, + 643 + ], + "spans": [ + { + "bbox": [ + 130, + 595, + 482, + 643 + ], + "type": "text", + "content": "Extra class-agnostic branch. This is a simple remedy to cope with the regression bias. Since rare classes favor a class-agnostic head while class-specific ones are slightly preferable to frequent classes in our hypothesis, we propose to use both heads. For class " + }, + { + "bbox": [ + 130, + 595, + 482, + 643 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 595, + 482, + 643 + ], + "type": "text", + "content": ", its regression head changes to:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 250, + 652, + 481, + 667 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 250, + 652, + 481, + 667 + ], + "spans": [ + { + "bbox": [ + 250, + 652, + 481, + 667 + ], + "type": "interline_equation", + "content": "W _ {i} ^ {\\prime} = \\alpha W _ {0} + (1 - \\alpha) W _ {i}, \\tag {3}", + "image_path": "5ce3cd7ce72ca00d83580af99604443f7c154b8ef7e26df76109849e2ba9f189.jpg" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 334, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 334, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 334, + 91, + 448, + 102 + ], + "type": "text", + "content": "Rectify the Regression Bias" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 140, + 222, + 481, + 304 + ], + "blocks": [ + { + "bbox": [ + 130, + 114, + 482, + 214 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 482, + 214 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 482, + 214 + ], + "type": "text", + "content": "Table 2: Our three different methods to alleviate the regression bias. In 2a, we add an extra class-agnostic regression head during training and testing, where the shared class-agonistic head and the class-specific head are combined using a weight " + }, + { + "bbox": [ + 130, + 114, + 482, + 214 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 130, + 114, + 482, + 214 + ], + "type": "text", + "content": ". In 2b, we perform regression head clustering according to the number of instances ('" + }, + { + "bbox": [ + 130, + 114, + 482, + 214 + ], + "type": "inline_equation", + "content": "n'" + }, + { + "bbox": [ + 130, + 114, + 482, + 214 + ], + "type": "text", + "content": ") or mean box scale ('" + }, + { + "bbox": [ + 130, + 114, + 482, + 214 + ], + "type": "inline_equation", + "content": "s'" + }, + { + "bbox": [ + 130, + 114, + 482, + 214 + ], + "type": "text", + "content": "). In 2c, we directly merge the regression heads of rare, common or frequent classes. The original class-specific heads in 2b and 2c are replaced by the new heads. Note that 'base' means the baseline method, i.e., with only class-specific regression heads. In 2a, " + }, + { + "bbox": [ + 130, + 114, + 482, + 214 + ], + "type": "inline_equation", + "content": "\\alpha = 0" + }, + { + "bbox": [ + 130, + 114, + 482, + 214 + ], + "type": "text", + "content": " is equivalent to the baseline method, and " + }, + { + "bbox": [ + 130, + 114, + 482, + 214 + ], + "type": "inline_equation", + "content": "\\alpha = 1.0" + }, + { + "bbox": [ + 130, + 114, + 482, + 214 + ], + "type": "text", + "content": " is equivalent to merging all " + }, + { + "bbox": [ + 130, + 114, + 482, + 214 + ], + "type": "inline_equation", + "content": "r, c, f" + }, + { + "bbox": [ + 130, + 114, + 482, + 214 + ], + "type": "text", + "content": " heads into a single head (i.e., the last row in 2c)." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 140, + 222, + 481, + 304 + ], + "lines": [ + { + "bbox": [ + 140, + 222, + 481, + 304 + ], + "spans": [ + { + "bbox": [ + 140, + 222, + 481, + 304 + ], + "type": "table", + "html": "
(α) APAPrAPbAPrknsAPAPrAPbAPrmerge APAPrAPbAPr
0.023.714.224.713.4base23.714.224.713.4base23.714.224.713.4
0.224.115.825.415.110✓24.514.526.314.4r25.116.326.716.7
0.525.117.527.018.010✓24.212.826.012.6c25.517.727.217.2
0.824.417.025.916.4100✓24.413.026.212.9r, c25.317.127.318.2
1.024.716.726.718.3100✓25.216.726.916.7r, c, f24.716.726.718.3
", + "image_path": "4743e126ba433b257a10cf7bb27612f59cdb5080e2838897aa757d1b9debca26.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 141, + 304, + 232, + 314 + ], + "lines": [ + { + "bbox": [ + 141, + 304, + 232, + 314 + ], + "spans": [ + { + "bbox": [ + 141, + 304, + 232, + 314 + ], + "type": "text", + "content": "(a) class-agnostic head" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_footnote" + }, + { + "bbox": [ + 261, + 304, + 343, + 314 + ], + "lines": [ + { + "bbox": [ + 261, + 304, + 343, + 314 + ], + "spans": [ + { + "bbox": [ + 261, + 304, + 343, + 314 + ], + "type": "text", + "content": "(b) Clustering heads" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_footnote" + }, + { + "bbox": [ + 385, + 304, + 459, + 314 + ], + "lines": [ + { + "bbox": [ + 385, + 304, + 459, + 314 + ], + "spans": [ + { + "bbox": [ + 385, + 304, + 459, + 314 + ], + "type": "text", + "content": "(c) Merging heads" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 346, + 480, + 405 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 346, + 480, + 405 + ], + "spans": [ + { + "bbox": [ + 130, + 346, + 480, + 405 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 346, + 480, + 405 + ], + "type": "inline_equation", + "content": "W_0" + }, + { + "bbox": [ + 130, + 346, + 480, + 405 + ], + "type": "text", + "content": " is a shared class-agnostic regression head for all classes, and " + }, + { + "bbox": [ + 130, + 346, + 480, + 405 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 130, + 346, + 480, + 405 + ], + "type": "text", + "content": " is a hyper-parameter to balance class-agnostic and class-specific heads. We empirically find that this simple change leads to consistent improvements over the default class-specific regression head (cf. Table 2a), while simply setting " + }, + { + "bbox": [ + 130, + 346, + 480, + 405 + ], + "type": "inline_equation", + "content": "\\alpha = 0.5" + }, + { + "bbox": [ + 130, + 346, + 480, + 405 + ], + "type": "text", + "content": " gives the optimal trade-off." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 406, + 482, + 477 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 406, + 482, + 477 + ], + "spans": [ + { + "bbox": [ + 130, + 406, + 482, + 477 + ], + "type": "text", + "content": "Clustering heads. This method is motivated by the analysis in Fig. 1c. Since some categories have similar statistics, we can assign them a shared regression head to improve the generalization ability. We implement it by following three steps: sorting, grouping and assigning. First, we use the number of instance or mean box scale to sort the original categories in descending order (" + }, + { + "bbox": [ + 130, + 406, + 482, + 477 + ], + "type": "inline_equation", + "content": "C = 1203" + }, + { + "bbox": [ + 130, + 406, + 482, + 477 + ], + "type": "text", + "content": " in LVIS):" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 241, + 478, + 480, + 493 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 241, + 478, + 480, + 493 + ], + "spans": [ + { + "bbox": [ + 241, + 478, + 480, + 493 + ], + "type": "interline_equation", + "content": "W _ {1}, \\dots , W _ {C} \\stackrel {\\text {s o r t}} {\\Longrightarrow} W _ {1} ^ {s}, \\dots , W _ {C} ^ {s}. \\tag {4}", + "image_path": "17632642b619d84ea2a551802af99dae302202ece8398464a0de34635ac91d8e.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 497, + 482, + 535 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 497, + 482, + 535 + ], + "spans": [ + { + "bbox": [ + 130, + 497, + 482, + 535 + ], + "type": "text", + "content": "Then, we cluster them into " + }, + { + "bbox": [ + 130, + 497, + 482, + 535 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 497, + 482, + 535 + ], + "type": "text", + "content": " groups. During clustering, we do not rely on time-cost algorithms such as K-means [4] or GMM [40]), but simply put adjacent classes into one group, and each group has the same number of classes " + }, + { + "bbox": [ + 130, + 497, + 482, + 535 + ], + "type": "inline_equation", + "content": "N = \\frac{C}{K}" + }, + { + "bbox": [ + 130, + 497, + 482, + 535 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 218, + 541, + 480, + 556 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 218, + 541, + 480, + 556 + ], + "spans": [ + { + "bbox": [ + 218, + 541, + 480, + 556 + ], + "type": "interline_equation", + "content": "\\{(W _ {N \\times i + 1} ^ {s}, W _ {N \\times i + 2} ^ {s}, \\dots , W _ {N \\times i + N} ^ {s}) \\} _ {i = 0} ^ {K - 1}, \\tag {5}", + "image_path": "cb6a9c56690b7d0c9cb2c46193b59ff7ea04a23ac0c2c30cd5fdeb112189255b.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 131, + 562, + 343, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 562, + 343, + 574 + ], + "spans": [ + { + "bbox": [ + 131, + 562, + 343, + 574 + ], + "type": "text", + "content": "Finally, each group " + }, + { + "bbox": [ + 131, + 562, + 343, + 574 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 131, + 562, + 343, + 574 + ], + "type": "text", + "content": " share one regression head:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 207, + 583, + 480, + 599 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 207, + 583, + 480, + 599 + ], + "spans": [ + { + "bbox": [ + 207, + 583, + 480, + 599 + ], + "type": "interline_equation", + "content": "W _ {g i} \\stackrel {\\text {r e p l a c e}} {\\Longleftarrow} \\left(W _ {N \\times i + 1} ^ {s}, W _ {N \\times i + 2} ^ {s}, \\dots , W _ {N \\times i + N} ^ {s}\\right). \\tag {6}", + "image_path": "d0e02f5904255d0e19f8d58eaefd12d0cd111b5dd3603542b22d7be45906ba83.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 605, + 482, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 641 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 641 + ], + "type": "text", + "content": "These shared regression matrix are then used for both training and testing. As shown in Table 2b, clustering heads with similar scale statistics brings robust improvements over the baseline methods (see more discussion in Sec. 4.4)." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 642, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 642, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 642, + 482, + 666 + ], + "type": "text", + "content": "Merge heads. This approach has similar motivation to the previous one, which clusters regression heads, but is even more straightforward. We simply" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "type": "text", + "content": "K. Zhu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 212 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 212 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 212 + ], + "type": "text", + "content": "group the regression heads in pre-defined clusters. For example, we let all rare categories share a common regression head " + }, + { + "bbox": [ + 130, + 116, + 482, + 212 + ], + "type": "inline_equation", + "content": "W_{\\text{rare}}" + }, + { + "bbox": [ + 130, + 116, + 482, + 212 + ], + "type": "text", + "content": ", and the same for common and frequent classes. Experimental results can be found in Table 2c, where we try four different combinations (note that " + }, + { + "bbox": [ + 130, + 116, + 482, + 212 + ], + "type": "inline_equation", + "content": "r, c" + }, + { + "bbox": [ + 130, + 116, + 482, + 212 + ], + "type": "text", + "content": " means merging rare and common classes all together into one regression head). The results indicate that merging always leads to performance gains, especially for the rare categories. The most significant improvements come from merging only the common class (i.e., the row denoted as 'c' in Table 2c, also cf. Sec. 4.4 for discussion on partition shift)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 213, + 482, + 284 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 213, + 482, + 284 + ], + "spans": [ + { + "bbox": [ + 130, + 213, + 482, + 284 + ], + "type": "text", + "content": "All our proposed remedies improve the accuracy of long-tailed object detection and instance segmentation, which verifies the importance of rectifying the regression bias. We will show that all these methods lead to consistent and nontrivial margin over various existing approaches (cf. Table 3). We then choose adding a class-agnostic branch (the first approach, cf. Table 2a) to conduct the rest of our experiments for its simplicity." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 285, + 482, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 285, + 482, + 380 + ], + "spans": [ + { + "bbox": [ + 130, + 285, + 482, + 380 + ], + "type": "text", + "content": "Attentive readers might advocate that all proposed three methods are to be verified on all datasets. However, we argue that this is not necessary and our choice has its advantages. On one hand, it does not need any dataset statistics, which is especially appreciated when the number of categories and the data distribution is unknown. On the other hand, by combining both class-agnostic and class-specific heads, it fully exploits object priors and per class knowledge: each type of head has its own merit, as Table 1 shows. Moreover, the validity of all three methods are already verified in Table 2-3." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 402, + 224, + 415 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 402, + 224, + 415 + ], + "spans": [ + { + "bbox": [ + 132, + 402, + 224, + 415 + ], + "type": "text", + "content": "4 Experiment" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 430, + 482, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 430, + 482, + 502 + ], + "spans": [ + { + "bbox": [ + 130, + 430, + 482, + 502 + ], + "type": "text", + "content": "We first combine 'Class-Agnostic Branch' (CAB), 'Clustering Head' (CH) and 'Merging Head' (MH) with existing long-tailed methods (cf. Table 3), then choose 'SeeSaw [35] + CAB' as 'Our' to compete with state-of-the-art methods (cf. Table 4). Although SeeSaw is a lower baseline than ECM [7], it is more stable in reproducing. Finally, we generalize our methods to various evaluation metrics, different datasets and the mask branch." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 523, + 271, + 534 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 523, + 271, + 534 + ], + "spans": [ + { + "bbox": [ + 132, + 523, + 271, + 534 + ], + "type": "text", + "content": "4.1 Experimental settings" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": "Datasets. We use the large vocabulary dataset LVIS1.0 [12] as our main dataset, which contains 100k training and 20k validation images. Rare " + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "inline_equation", + "content": "(r)" + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": ", common " + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "inline_equation", + "content": "(c)" + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": " and frequent " + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "inline_equation", + "content": "(f)" + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": " classes are defined by how many images they occur [12]: [0, 10] for rare, [11, 100] for common, and " + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "inline_equation", + "content": "(100, +\\infty)" + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": " for frequent, respectively. We also adopt COCO-LT [25] and COCO2017 [36] to verify the generalization ability of our approach. COCO2017 is a large object detection dataset, containing 118k training and 5k validation images. It is relatively balanced in comparison with LVIS1.0. The COCO-LT dataset is an artificially sampled subset of COCO, with the same validation set but a long-tailed training set. It has about 99k training and 5k validation images. Following previous works [36], we partition COCO-LT" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 334, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 334, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 334, + 91, + 447, + 102 + ], + "type": "text", + "content": "Rectify the Regression Bias" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 179, + 179, + 432, + 382 + ], + "blocks": [ + { + "bbox": [ + 130, + 114, + 482, + 170 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 482, + 170 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 482, + 170 + ], + "type": "text", + "content": "Table 3: Experiments on LVIS1.0. We combine four existing methods with our approach 'adding a class-agnostic branch' (CAB), 'Clustering Head' (CH) or 'Merging Head' (MH). We reproduced RFS [12], EQLv2 [33], SeeSaw [35] and ECM [7] using their official code. For clarity, here we list the object detection metrics " + }, + { + "bbox": [ + 130, + 114, + 482, + 170 + ], + "type": "inline_equation", + "content": "\\mathrm{AP}^b" + }, + { + "bbox": [ + 130, + 114, + 482, + 170 + ], + "type": "text", + "content": " in the first four columns while putting the instance segmentation metrics AP in the latter ones." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 179, + 179, + 432, + 382 + ], + "lines": [ + { + "bbox": [ + 179, + 179, + 432, + 382 + ], + "spans": [ + { + "bbox": [ + 179, + 179, + 432, + 382 + ], + "type": "table", + "html": "
Method+Our APbAPRbAPCbAPf bAPAPrAPcAPf
RFS [12]base24.713.423.131.423.714.222.9
CAB27.018.025.332.925.117.523.9
CH26.917.225.533.025.216.724.3
MH27.318.225.832.825.317.124.5
EQLv2 [33]base26.016.124.032.525.217.424.1
CAB28.120.426.333.526.019.524.9
CH27.318.225.833.125.717.724.8
MH27.117.325.733.225.617.524.7
SeeSaw [35]base27.318.226.532.326.919.626.8
CAB28.919.928.333.627.720.227.3
CH29.119.828.933.427.820.627.7
MH28.518.428.433.727.419.827.4
ECM [7]base27.717.726.633.127.219.626.6
CAB29.118.428.933.927.819.128.0
CH28.718.028.333.727.819.627.7
MH28.618.628.233.527.719.827.7
", + "image_path": "fdcf6c15fdb078283ae2bc1302ae9dac22b0e9d54db0530639b16ab475e76f1f.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 403, + 479, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 403, + 479, + 426 + ], + "spans": [ + { + "bbox": [ + 130, + 403, + 479, + 426 + ], + "type": "text", + "content": "into 4 evaluation subsets according to the number of training instances per class, with bins of [1, 20), [20, 400), [400, 8000) and [8000, -), respectively." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 427, + 482, + 606 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 427, + 482, + 606 + ], + "spans": [ + { + "bbox": [ + 130, + 427, + 482, + 606 + ], + "type": "text", + "content": "Training details. We reproduce four different methods as our baselines, including RFS [12], EQLv2 [33], SeeSaw [35] and ECM [7], following their default experiment settings. We employ MMDetection [5] as our detection framework to conduct our experiment, and train detection models of Faster-RCNN, Mask-RCNN and Cascade R-CNN for 1x or 2x scheduler (except Swin-Transformer based detectors), following previous works [7, 17]. The batch size and learning rate are set as 16 and 0.02, and the data augmentation strictly follows previous long-tailed detection methods [7, 33, 35]. During training, we use FP16 mixed precision training and the warmup strategy to stabilize the learning process. For the evaluation metrics, we adopt AP and " + }, + { + "bbox": [ + 130, + 427, + 482, + 606 + ], + "type": "inline_equation", + "content": "\\mathrm{AP}^b" + }, + { + "bbox": [ + 130, + 427, + 482, + 606 + ], + "type": "text", + "content": " for instance segmentation and object detection, respectively, and adopt " + }, + { + "bbox": [ + 130, + 427, + 482, + 606 + ], + "type": "inline_equation", + "content": "\\mathrm{AP}_1^b" + }, + { + "bbox": [ + 130, + 427, + 482, + 606 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 427, + 482, + 606 + ], + "type": "inline_equation", + "content": "\\mathrm{AP}_2^b" + }, + { + "bbox": [ + 130, + 427, + 482, + 606 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 427, + 482, + 606 + ], + "type": "inline_equation", + "content": "\\mathrm{AP}_3^b" + }, + { + "bbox": [ + 130, + 427, + 482, + 606 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 427, + 482, + 606 + ], + "type": "inline_equation", + "content": "\\mathrm{AP}_4^b" + }, + { + "bbox": [ + 130, + 427, + 482, + 606 + ], + "type": "text", + "content": " on COCOLT, corresponding to its 4 different subsets. With the suggested practice in LVIS official website, we run all our experiment 3 times on 8 RTX3090 GPUs to reduce the variance. Please refer to our supplementary material for more detailed information." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 623, + 328, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 623, + 328, + 635 + ], + "spans": [ + { + "bbox": [ + 131, + 623, + 328, + 635 + ], + "type": "text", + "content": "4.2 LVIS detection and segmentation" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "content": "Consistent improvements. we first evaluate the effectiveness of our method on the LVIS1.0 dataset by combing the proposed approach 'adding a class-agnostic" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "type": "text", + "content": "K. Zhu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 159, + 179, + 451, + 439 + ], + "blocks": [ + { + "bbox": [ + 130, + 114, + 482, + 170 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 482, + 170 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 482, + 170 + ], + "type": "text", + "content": "Table 4: Comparison with state-of-the-art methods on the LVIS1.0 dataset. The results of 'Our' means our proposed CAB regression method combined with the SeeSaw loss ('SeeSaw + CAB', cf. Sec. 4). We show the metrics of instance segmentation AP and object detection " + }, + { + "bbox": [ + 130, + 114, + 482, + 170 + ], + "type": "inline_equation", + "content": "\\mathrm{AP}^b" + }, + { + "bbox": [ + 130, + 114, + 482, + 170 + ], + "type": "text", + "content": ". The results with a " + }, + { + "bbox": [ + 130, + 114, + 482, + 170 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 130, + 114, + 482, + 170 + ], + "type": "text", + "content": " are copied from [7], while others are reproduced by us using their official released code." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 159, + 179, + 451, + 439 + ], + "lines": [ + { + "bbox": [ + 159, + 179, + 451, + 439 + ], + "spans": [ + { + "bbox": [ + 159, + 179, + 451, + 439 + ], + "type": "table", + "html": "
ArchitectBackboneMethodAPAPrAPfAPcAPb
Mask-RCNN [14]R50-FPNCE18.70.416.529.319.7
RFS [13]23.714.222.929.324.7
BSCE [29]24.415.723.629.125.5
EQLv2 [33]25.217.424.129.926.0
ECM [7]†27.419.727.031.127.9
ECM27.219.626.631.327.7
SeeSaw [35]26.919.626.830.527.3
ROG [42]26.920.126.830.027.2
Our27.720.227.331.328.9
Mask-RCNN [14]R101-FPNCE [7]†25.516.624.530.626.6
EQLv2 [33]†27.220.625.931.427.9
ECM [7]†28.721.927.932.329.4
ECM28.620.928.332.229.3
SeeSaw [35]28.220.328.131.829.0
Our29.021.028.932.430.7
Cascade R-CNN [2]Swin-TECM [7]34.124.034.938.037.6
SeeSaw [35]34.224.634.737.837.8
Our34.624.735.338.138.2
Cascade R-CNN [2]Swin-BECM [7]†39.733.540.641.443.6
Our39.934.640.741.244.2
", + "image_path": "b629e40fb6d9c6f3aba4acc2ef3955b5b45d437cc6dcf63931ac4d0aa1806c64.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 470, + 482, + 626 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 470, + 482, + 626 + ], + "spans": [ + { + "bbox": [ + 130, + 470, + 482, + 626 + ], + "type": "text", + "content": "branch' (CAB) with existing long-tailed object detection methods. Since our main focus is on bounding box regression, we list the object detection results in early columns and segmentation results in later ones. As shown in Table 3, using CAB leads to consistent " + }, + { + "bbox": [ + 130, + 470, + 482, + 626 + ], + "type": "inline_equation", + "content": "\\mathrm{AP}^b" + }, + { + "bbox": [ + 130, + 470, + 482, + 626 + ], + "type": "text", + "content": " and AP improvement over existing classification-based methods, surpassing all of them with large margins. For object detection, Our CAB benefits the rare class a lot, with an increase of " + }, + { + "bbox": [ + 130, + 470, + 482, + 626 + ], + "type": "inline_equation", + "content": "4.6\\mathrm{AP}^b" + }, + { + "bbox": [ + 130, + 470, + 482, + 626 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 470, + 482, + 626 + ], + "type": "inline_equation", + "content": "4.3\\mathrm{AP}^b" + }, + { + "bbox": [ + 130, + 470, + 482, + 626 + ], + "type": "text", + "content": " on RFS and EQLv2. The same is true for instance segmentation, where a growing trend can be observed on all metrics, showing that CAB is also beneficial for later mask pixel predictions. Interestingly, the method 'RFS+CAB' (which uses a CE loss) can almost achieve the same object detection accuracy as the SeeSaw method, and surpasses EQLv2 for about " + }, + { + "bbox": [ + 130, + 470, + 482, + 626 + ], + "type": "inline_equation", + "content": "1\\mathrm{AP}^b" + }, + { + "bbox": [ + 130, + 470, + 482, + 626 + ], + "type": "text", + "content": ". We thus conjecture that: besides merely focusing on classification, our regression methods can serve as strong alternatives that also strongly boost long-tailed detection accuracy." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "text", + "content": "Comparison with SOTA. We then compare our method with state-of-the-art methods using different object detection framework (Mask-RCNN, Cascade R-CNN) and backbones (ResNet-50, ResNet-101, Swin-T and Swin-B). Note" + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 334, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 334, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 334, + 91, + 447, + 102 + ], + "type": "text", + "content": "Rectify the Regression Bias" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 217, + 147, + 396, + 234 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 482, + 139 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 482, + 139 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 482, + 139 + ], + "type": "text", + "content": "Table 5: More evaluation metrics in LVIS1.0 dataset with a 1x scheduler using Mask RCNN. AP " + }, + { + "bbox": [ + 132, + 114, + 482, + 139 + ], + "type": "inline_equation", + "content": "{}_{\\text{boundary }}^{fixed}" + }, + { + "bbox": [ + 132, + 114, + 482, + 139 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 132, + 114, + 482, + 139 + ], + "type": "inline_equation", + "content": "{\\mathrm{{AP}}}_{\\text{box }}^{fixed}" + }, + { + "bbox": [ + 132, + 114, + 482, + 139 + ], + "type": "text", + "content": " are two newly proposed challenging metrics [6,9]." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 217, + 147, + 396, + 234 + ], + "lines": [ + { + "bbox": [ + 217, + 147, + 396, + 234 + ], + "spans": [ + { + "bbox": [ + 217, + 147, + 396, + 234 + ], + "type": "table", + "html": "
MethodAPAPbAPfixedboundaryAPfixedbbox
BAGS [24]23.123.7-26.2
EQLv2 [33]23.924.020.325.9
SeeSaw [35]25.225.419.826.5
ECM [7]26.326.721.427.4
Our26.928.022.128.2
", + "image_path": "434d8f4356dc88563624947052da4159241b439bcf73a490bd85615d33f4d328.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 258, + 482, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 258, + 482, + 413 + ], + "spans": [ + { + "bbox": [ + 130, + 258, + 482, + 413 + ], + "type": "text", + "content": "that 'Our' means 'SeeSaw + CAB'. For fair comparison, we reproduce majority of existing methods using their official released code unless specialized symbols " + }, + { + "bbox": [ + 130, + 258, + 482, + 413 + ], + "type": "inline_equation", + "content": "(e.g.,^{\\dagger})" + }, + { + "bbox": [ + 130, + 258, + 482, + 413 + ], + "type": "text", + "content": " appears after a method's name. As can be seen in Table 4, our method achieves the overall highest accuracy in AP and " + }, + { + "bbox": [ + 130, + 258, + 482, + 413 + ], + "type": "inline_equation", + "content": "\\mathrm{AP}^b" + }, + { + "bbox": [ + 130, + 258, + 482, + 413 + ], + "type": "text", + "content": ". For ResNet series models, our regression technique easily surpass the best competitor ECM [7], especially in " + }, + { + "bbox": [ + 130, + 258, + 482, + 413 + ], + "type": "inline_equation", + "content": "\\mathrm{AP}^b" + }, + { + "bbox": [ + 130, + 258, + 482, + 413 + ], + "type": "text", + "content": " (an increase of " + }, + { + "bbox": [ + 130, + 258, + 482, + 413 + ], + "type": "inline_equation", + "content": "1.2\\mathrm{AP}^b" + }, + { + "bbox": [ + 130, + 258, + 482, + 413 + ], + "type": "text", + "content": " for ResNet-50 and " + }, + { + "bbox": [ + 130, + 258, + 482, + 413 + ], + "type": "inline_equation", + "content": "1.1\\mathrm{AP}^b" + }, + { + "bbox": [ + 130, + 258, + 482, + 413 + ], + "type": "text", + "content": " for ResNet-101). The advantage also holds for ViT-based object detectors, where we surpassed the best competitor ECM [7] with both Swin-Tiny and Swin-Base backbone architectures. Following LVIS's common practice, we didn't list " + }, + { + "bbox": [ + 130, + 258, + 482, + 413 + ], + "type": "inline_equation", + "content": "\\mathrm{AP}_r^b" + }, + { + "bbox": [ + 130, + 258, + 482, + 413 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 258, + 482, + 413 + ], + "type": "inline_equation", + "content": "\\mathrm{AP}_c^b" + }, + { + "bbox": [ + 130, + 258, + 482, + 413 + ], + "type": "text", + "content": " here, but we want to emphasize that the advantage of our regression methods can be further enlarged when more metrics are listed (cf. Table 3). This is also true if we replace 'adding a Class-Agnostic Branch' (CAB) with the 'merging heads' regression alternative (cf. the best accuracy in Table 2)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 435, + 270, + 449 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 435, + 270, + 449 + ], + "spans": [ + { + "bbox": [ + 132, + 435, + 270, + 449 + ], + "type": "text", + "content": "4.3 Generalization ability" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 459, + 481, + 493 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 459, + 481, + 493 + ], + "spans": [ + { + "bbox": [ + 130, + 459, + 481, + 493 + ], + "type": "text", + "content": "In this section, we will show the generalization ability of our regression methods in various aspects, including different evaluation metrics, datasets and its benefits on the mask branch." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 495, + 482, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 495, + 482, + 581 + ], + "spans": [ + { + "bbox": [ + 130, + 495, + 482, + 581 + ], + "type": "text", + "content": "Different metrics. We first explore how different metrics affect our model's accuracy. Two additional metrics are " + }, + { + "bbox": [ + 130, + 495, + 482, + 581 + ], + "type": "inline_equation", + "content": "\\mathrm{AP}_{\\text {boundary }}" + }, + { + "bbox": [ + 130, + 495, + 482, + 581 + ], + "type": "text", + "content": " (a more strict metric in instance segmentation [6]) and " + }, + { + "bbox": [ + 130, + 495, + 482, + 581 + ], + "type": "inline_equation", + "content": "\\mathrm{AP}_{\\text {bbox }}^{\\text {fixed }}" + }, + { + "bbox": [ + 130, + 495, + 482, + 581 + ], + "type": "text", + "content": " (constraining 10,000 predicted bounding boxes per class across the dataset [9]). Object detector that obtain decent results in traditional metrics may not perform as well in these criterions. As shown in Table 5, our regression methods adapts well and surpasses all existing methods on both traditional and these challenging metrics." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "content": "The COCO-LT dataset. We also transfer our regression to another long-tailed dataset: COCO-LT. This is an artificially sampled [36] subset of original COCO [25]. We calculate the over bounding box metrics " + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathrm{AP}^b" + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "content": " and more fine-grained results: " + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathrm{AP}_1^b" + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathrm{AP}_2^b" + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathrm{AP}_3^b" + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathrm{AP}_4^b" + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "content": " (ranging from the rarest to the most frequent class). As shown in Table 6, our CAB leads to consistent improvements on all metrics (especially for the rarest categories " + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathrm{AP}_1^b" + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "content": ") under different repeat-factor-sampling rates, showing the great power of CAB to help the rare" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "type": "text", + "content": "K. Zhu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 223, + 146, + 387, + 248 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 479, + 137 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 479, + 137 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 479, + 137 + ], + "type": "text", + "content": "Table 6: Experiments on COCO-LT [36]. Following LVIS [12], we use a similar repeat-factor-sampling (RFS) strategy with different sampling rates (3e-3, 5e-3 and 1e-2)." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 223, + 146, + 387, + 248 + ], + "lines": [ + { + "bbox": [ + 223, + 146, + 387, + 248 + ], + "spans": [ + { + "bbox": [ + 223, + 146, + 387, + 248 + ], + "type": "table", + "html": "
Rate +CABAP1bAP2bAP3bAP4bAPb
3e-3✓2.316.926.730.423.5
6.219.127.030.524.4
5e-3✓3.619.326.730.624.3
7.520.027.330.725.0
1e-2✓8.720.227.830.325.2
10.821.628.030.525.8
", + "image_path": "dd6040b02b5971ed0249a00111e944b1ca4c8343d051f7661fc501a38a64487d.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 199, + 306, + 411, + 407 + ], + "blocks": [ + { + "bbox": [ + 132, + 274, + 479, + 296 + ], + "lines": [ + { + "bbox": [ + 132, + 274, + 479, + 296 + ], + "spans": [ + { + "bbox": [ + 132, + 274, + 479, + 296 + ], + "type": "text", + "content": "Table 7: Experiments on the (relatively) balanced COCO dataset. We adopted the Faster RCNN [30] detector and tried three different backbones." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 199, + 306, + 411, + 407 + ], + "lines": [ + { + "bbox": [ + 199, + 306, + 411, + 407 + ], + "spans": [ + { + "bbox": [ + 199, + 306, + 411, + 407 + ], + "type": "table", + "html": "
Backbone+CABAP\\( AP_{50} \\)\\( AP_{75} \\)\\( AP_s \\)\\( AP_m \\)\\( AP_l \\)
Res-50✓37.358.340.321.741.048.2
38.358.942.122.441.450.4
Res-101✓39.460.343.022.943.451.0
39.960.343.622.743.752.8
Res-32x4d✓41.062.244.623.945.352.9
41.762.545.423.645.954.8
", + "image_path": "8880ea6f3a05acb6c18f64dbe90cfc08e1a5d90ff6edcb7f171a5f41d22cae19.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 446, + 479, + 470 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 446, + 479, + 470 + ], + "spans": [ + { + "bbox": [ + 130, + 446, + 479, + 470 + ], + "type": "text", + "content": "classes. In fact, this conclusion generally holds true for all sampling rates in our experiments. We only listed three here for simplicity and clarity." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 478, + 480, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 478, + 480, + 586 + ], + "spans": [ + { + "bbox": [ + 130, + 478, + 480, + 586 + ], + "type": "text", + "content": "On balanced training set. Furthermore, we validate how our method performs on the relatively balanced dataset MS-COCO2017. We adopted Faster RCNN with three different backbones (ResNet-50, ResNet-101 and ResNext101-32x4d). Results in Table 7 clearly show that CAB generalizes well to datasets with more balanced distributions. Interestingly, CAB boosts metrics of large and medium objects while shows similar accuracy on small objects. This is possibly because that small objects occupy over " + }, + { + "bbox": [ + 130, + 478, + 480, + 586 + ], + "type": "inline_equation", + "content": "60\\%" + }, + { + "bbox": [ + 130, + 478, + 480, + 586 + ], + "type": "text", + "content": " of the total instances while large objects only consumes roughly " + }, + { + "bbox": [ + 130, + 478, + 480, + 586 + ], + "type": "inline_equation", + "content": "15\\%" + }, + { + "bbox": [ + 130, + 478, + 480, + 586 + ], + "type": "text", + "content": " [32]. Since CAB is more beneficial to less-frequent classes, it brings higher gains in large objects than in small ones." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 594, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 481, + 665 + ], + "type": "text", + "content": "The mask branch. Finally, we apply our CAB to the segmentation branch to test whether adding an class-agnostic prior is suitable for mask prediction. Experimental results are in Table 8, where we add a class-agnostic mask prediction head and combines it with each class-specific mask head. As shown in the table, our CAB generalizes well to segmentation tasks. Our main experimental results may be further improved if CAB was applied in both box and mask predictions." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 334, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 334, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 334, + 91, + 447, + 102 + ], + "type": "text", + "content": "Rectify the Regression Bias" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 230, + 146, + 381, + 222 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 482, + 137 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 482, + 137 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 482, + 137 + ], + "type": "text", + "content": "Table 8: Generalization to mask prediction. We added a class-agnostic mask prediction branch and combines it with each class-specific mask prediction results." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 230, + 146, + 381, + 222 + ], + "lines": [ + { + "bbox": [ + 230, + 146, + 381, + 222 + ], + "spans": [ + { + "bbox": [ + 230, + 146, + 381, + 222 + ], + "type": "table", + "html": "
Branch +CABAPAPrAPbAPr
Box✓23.714.224.713.4
25.117.527.018.0
Mask✓23.714.224.713.4
24.114.625.013.9
", + "image_path": "b7ef28d692341c326de78d12442b70a31acde8784292232536f677107f1ca702.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 245, + 326, + 257 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 245, + 326, + 257 + ], + "spans": [ + { + "bbox": [ + 132, + 245, + 326, + 257 + ], + "type": "text", + "content": "4.4 Discussion of our three remedies" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 268, + 479, + 293 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 268, + 479, + 293 + ], + "spans": [ + { + "bbox": [ + 130, + 268, + 479, + 293 + ], + "type": "text", + "content": "Now we further discuss the proposed three methods (CAB, CH and MH) in detail, hoping to provide a more in-depth understanding of our hypothesis." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 293, + 480, + 435 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 293, + 480, + 435 + ], + "spans": [ + { + "bbox": [ + 130, + 293, + 480, + 435 + ], + "type": "text", + "content": "Relationship of CAB to the objectness branch. At start, we want to clarify the differences between our CAB and the objectness branch approach [7, 35] adopted in the classification head. We have argued that although they seemingly share similar structures with ours, they are essentially different. First, the purpose of objectness branch is to deal with the imbalanced distribution between foreground and background classification samples [35], while our CAB aim to tackle the regression bias. Second, unlike classification where each class must preserve its own classifier, the regression heads can be clustered or merged, which provide more diverse solutions to reduce the regression bias besides CAB (cf. Table 2). Lastly, the effect of the objectness branch is yet to be proved because it will decrease the performance of rare classes even when combined with CE loss [35], while CAB leads to consistent improvements, especially in rare class." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 437, + 481, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 437, + 481, + 581 + ], + "spans": [ + { + "bbox": [ + 130, + 437, + 481, + 581 + ], + "type": "text", + "content": "How many clusters are the best? Attentive readers might have questions why larger clusters K (such as 100) in CH achieve better results in Table 2b for rare class, since enlarging K means the (useful) class-agnostic information is reduced. We thus empirically verify that: besides objects numbers per group, the scale variance could potentially affect the final AP. We test CH with 'object scales' in more K. For each K, we calculate standard deviation of scales in each group, average, and get representative scale std. The same is done for average object. As Fig. 3a shows: as K decreases, training objects per group increases, which benefit rare classes. But, the variation of scales increases a lot, which makes regression head learning more difficult—imagine one head to fit offsets that vary a lot, and we also observed higher training loss for smaller K. Hence, " + }, + { + "bbox": [ + 130, + 437, + 481, + 581 + ], + "type": "inline_equation", + "content": "\\max AP_r^b" + }, + { + "bbox": [ + 130, + 437, + 481, + 581 + ], + "type": "text", + "content": " is not at " + }, + { + "bbox": [ + 130, + 437, + 481, + 581 + ], + "type": "inline_equation", + "content": "K = 1" + }, + { + "bbox": [ + 130, + 437, + 481, + 581 + ], + "type": "text", + "content": " or 10, but at " + }, + { + "bbox": [ + 130, + 437, + 481, + 581 + ], + "type": "inline_equation", + "content": "K = 50" + }, + { + "bbox": [ + 130, + 437, + 481, + 581 + ], + "type": "text", + "content": " in our experiments (cf. Fig. 3b)." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 582, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 481, + 666 + ], + "type": "text", + "content": "Why merging 'Common' class is the best?. We observe in Table 2c that merging common classes leads to the best improvement in " + }, + { + "bbox": [ + 130, + 582, + 481, + 666 + ], + "type": "inline_equation", + "content": "\\mathrm{AP}_r" + }, + { + "bbox": [ + 130, + 582, + 481, + 666 + ], + "type": "text", + "content": ", which is somehow counter-intuitive. We conjecture that this is due to the partition shift of rare, common and frequent in LVIS train and validation sets. In the LVIS1.0 train dataset, the size ranges for frequent, common and rare are [0, 404], [405, 865] and [866, 1202], respectively. While for the validation set, they become [0, 212], [213, 536] and [537, 1202], respectively (cf. Fig. 1c). Hence, when we" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "type": "text", + "content": "K. Zhu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 202, + 115, + 301, + 203 + ], + "blocks": [ + { + "bbox": [ + 202, + 115, + 301, + 203 + ], + "lines": [ + { + "bbox": [ + 202, + 115, + 301, + 203 + ], + "spans": [ + { + "bbox": [ + 202, + 115, + 301, + 203 + ], + "type": "image", + "image_path": "84c82224482063638ab8a7c3ea99fe566a81d31ec2e85c041f63c656378608d4.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 212, + 482, + 234 + ], + "lines": [ + { + "bbox": [ + 132, + 212, + 482, + 234 + ], + "spans": [ + { + "bbox": [ + 132, + 212, + 482, + 234 + ], + "type": "text", + "content": "Fig. 3: The statistics (averaged standard deviation, averaged objects per group) of different K and its corresponding AP box." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 309, + 115, + 408, + 202 + ], + "blocks": [ + { + "bbox": [ + 309, + 115, + 408, + 202 + ], + "lines": [ + { + "bbox": [ + 309, + 115, + 408, + 202 + ], + "spans": [ + { + "bbox": [ + 309, + 115, + 408, + 202 + ], + "type": "image", + "image_path": "3e8c2b949e449956d09858474ec0f3d9bcde517125df2e5d907997c6849105f2.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 138, + 246, + 477, + 326 + ], + "blocks": [ + { + "bbox": [ + 138, + 246, + 477, + 326 + ], + "lines": [ + { + "bbox": [ + 138, + 246, + 477, + 326 + ], + "spans": [ + { + "bbox": [ + 138, + 246, + 477, + 326 + ], + "type": "image", + "image_path": "8a80c217608a3d3fdc3e34bd446db360da2a9eedb0590cf9bb416e8b5ee0a547.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 333, + 482, + 389 + ], + "lines": [ + { + "bbox": [ + 131, + 333, + 482, + 389 + ], + "spans": [ + { + "bbox": [ + 131, + 333, + 482, + 389 + ], + "type": "text", + "content": "Fig. 4: Visualizations of detection results before (in the left of each group) and after (in the right) using our CAB. We adopted RFS [12] as the baseline in LVIS1.0 and combine it with our CAB regression method. In comparison, the proposed method is good at detecting missing objects, filtering duplicated objects away, as well as rectifying bounding box predictions. This figure needs to be viewed in color." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 415, + 482, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 415, + 482, + 450 + ], + "spans": [ + { + "bbox": [ + 131, + 415, + 482, + 450 + ], + "type": "text", + "content": "use training set statistics to merge the common class, a large portion of rare categories in the validation set are also potentially merged, thus contributing to the validation improvement on both " + }, + { + "bbox": [ + 131, + 415, + 482, + 450 + ], + "type": "inline_equation", + "content": "\\mathrm{AP}_r" + }, + { + "bbox": [ + 131, + 415, + 482, + 450 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 415, + 482, + 450 + ], + "type": "inline_equation", + "content": "\\mathrm{AP}_c" + }, + { + "bbox": [ + 131, + 415, + 482, + 450 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 472, + 293, + 483 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 472, + 293, + 483 + ], + "spans": [ + { + "bbox": [ + 132, + 472, + 293, + 483 + ], + "type": "text", + "content": "4.5 Visualization and ablation" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 497, + 482, + 580 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 497, + 482, + 580 + ], + "spans": [ + { + "bbox": [ + 130, + 497, + 482, + 580 + ], + "type": "text", + "content": "Flatten distribution. We first plot the RCNN regression loss of each category before and after combining with our regression CAB. The baseline methods we choose are EQLv2 and CE. As shown in Fig. 5, the regression loss for rare has seen a noticeable drop after adding our CAB, and the overall loss distribution has become much more balanced, too. We thus believe that the pipeline of our regression remedies can relieve the regression bias, further verifying that our hypothesis in Sec. 1 in indeed valid." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "content": "More accurate box/mask. We then calculate AP50-AP95 of boxes and masks to find whether our CAB behaves well on stricter IoU thresholds. As show in Fig. 5c-5d, adding CAB achieves consistent accuracy gains over all IoU thresholds for both box and mask prediction, and is especially helpful for those hard IoU threshold (e.g., AP75-90) in box evaluation (cf. Fig. 5c). Since a higher threshold requires more precise box prediction, these results have shown that our regression methods is capable of producing precise boxes with better quality." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 334, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 334, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 334, + 91, + 447, + 102 + ], + "type": "text", + "content": "Rectify the Regression Bias" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 139, + 114, + 218, + 173 + ], + "blocks": [ + { + "bbox": [ + 139, + 114, + 218, + 173 + ], + "lines": [ + { + "bbox": [ + 139, + 114, + 218, + 173 + ], + "spans": [ + { + "bbox": [ + 139, + 114, + 218, + 173 + ], + "type": "image", + "image_path": "9c1c9872b6b3fb04803c144e4903357762f5c4e2c4511321a81c709ffb11410c.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 153, + 176, + 203, + 186 + ], + "lines": [ + { + "bbox": [ + 153, + 176, + 203, + 186 + ], + "spans": [ + { + "bbox": [ + 153, + 176, + 203, + 186 + ], + "type": "text", + "content": "(a) w/o CAB" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 225, + 115, + 304, + 173 + ], + "blocks": [ + { + "bbox": [ + 225, + 115, + 304, + 173 + ], + "lines": [ + { + "bbox": [ + 225, + 115, + 304, + 173 + ], + "spans": [ + { + "bbox": [ + 225, + 115, + 304, + 173 + ], + "type": "image", + "image_path": "af6b51a96e98acf6c4fc699c7eb46616b82bae858e2b38b01d02dc6d4fc08d62.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 241, + 176, + 287, + 186 + ], + "lines": [ + { + "bbox": [ + 241, + 176, + 287, + 186 + ], + "spans": [ + { + "bbox": [ + 241, + 176, + 287, + 186 + ], + "type": "text", + "content": "(b) w/CAB" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 131, + 194, + 480, + 217 + ], + "lines": [ + { + "bbox": [ + 131, + 194, + 480, + 217 + ], + "spans": [ + { + "bbox": [ + 131, + 194, + 480, + 217 + ], + "type": "text", + "content": "Fig. 5: The loss distribution shift before and after combining with our CAB (Fig. 5a, 5b), and AP improvement of combining our CAB with baseline in LVIS (Fig. 5c, 5d)." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 311, + 115, + 389, + 173 + ], + "blocks": [ + { + "bbox": [ + 311, + 115, + 389, + 173 + ], + "lines": [ + { + "bbox": [ + 311, + 115, + 389, + 173 + ], + "spans": [ + { + "bbox": [ + 311, + 115, + 389, + 173 + ], + "type": "image", + "image_path": "66697af2349b380f646825e9f6a47b6feb50d1bb2a16c978d59c70f5e58c5096.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 326, + 176, + 374, + 186 + ], + "lines": [ + { + "bbox": [ + 326, + 176, + 374, + 186 + ], + "spans": [ + { + "bbox": [ + 326, + 176, + 374, + 186 + ], + "type": "text", + "content": "(c) box delta" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 397, + 115, + 475, + 173 + ], + "blocks": [ + { + "bbox": [ + 397, + 115, + 475, + 173 + ], + "lines": [ + { + "bbox": [ + 397, + 115, + 475, + 173 + ], + "spans": [ + { + "bbox": [ + 397, + 115, + 475, + 173 + ], + "type": "image", + "image_path": "3b99aeb85f883781aa25c1e924f5c2ca04c72c54ec401407a61c7a055c18345f.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 409, + 176, + 463, + 186 + ], + "lines": [ + { + "bbox": [ + 409, + 176, + 463, + 186 + ], + "spans": [ + { + "bbox": [ + 409, + 176, + 463, + 186 + ], + "type": "text", + "content": "(d) mask delta" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 243, + 482, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 243, + 482, + 399 + ], + "spans": [ + { + "bbox": [ + 130, + 243, + 482, + 399 + ], + "type": "text", + "content": "Qualitative results. Last but not least, we provide example detection images sampled from the LVIS1.0 validation dataset. For simplicity, we choose RFS [12] with CE loss as baseline and combine it with our CAB method. As shown in Fig. 4, both baseline and our CAB detect most of the objects in an image, but CAB generally caters for more details. For example, CAB can help discover missed boxes, like the cabinet in the lower left images. It is also clearly illustrated that CAB helps filter duplicate boxes away in the final predictions (e.g., the pictures with elephant and cow with grass background). Since CAB brings in better predictions, duplicate boxes will have larger overlap that may be suppressed by NMS (non-maximum suppression). If we zoom this figure (cf. the last column in Fig. 4), we will find that CAB is capable of rectifying the predicted boxes, and this can empirically explain why our CAB achieves much better AP under higher and more difficult IoU thresholds (cf. Fig. 5c)." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 420, + 323, + 433 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 420, + 323, + 433 + ], + "spans": [ + { + "bbox": [ + 132, + 420, + 323, + 433 + ], + "type": "text", + "content": "5 Conclusions and Limitations" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 449, + 482, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 449, + 482, + 544 + ], + "spans": [ + { + "bbox": [ + 130, + 449, + 482, + 544 + ], + "type": "text", + "content": "In this paper, we discovered that the regression bias (imbalanced regression loss distribution on the RCNN head) exists in long-tailed object detection, and adversely affects detection results. We thus proposed three remedies for rectifying the regression bias. The proposed method significantly boosts the performance of rare class " + }, + { + "bbox": [ + 130, + 449, + 482, + 544 + ], + "type": "inline_equation", + "content": "\\mathrm{AP}^b" + }, + { + "bbox": [ + 130, + 449, + 482, + 544 + ], + "type": "text", + "content": ", and achieves state-of-the-art results. We also generalize our regression methods to balanced dataset, different evaluation metrics and the mask branch. Finally, visualizations show that our method indeed produces better predicted bounding boxes." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "type": "text", + "content": "As for the limitations, it remains unclear why the boosted accuracy becomes lower when the baseline are higher (e.g., the improvement of CAB on ECM is lower than that on RFS, cf. Table 3). This may relate to the upper bound a backbone model can achieve. Since large network generally better fit the dataset (cf. Table 4), the performance a ResNet-50 can achieve is limited. We thus call for involving larger vision models on the LVIS dataset to better handle the long-tailed object detection problem. Besides involving stronger vision backbones, we will try to testify the adaptability of the proposed methods when different pretraining information is integrated into the backbone network (e.g. self-supervised learning), which is left as our future work." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "type": "text", + "content": "K. Zhu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 140, + 479, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 140, + 479, + 163 + ], + "spans": [ + { + "bbox": [ + 132, + 140, + 479, + 163 + ], + "type": "text", + "content": "We thank Yin-Yin He (M.Sc. degree in Nanjing University, now in ByteDance) for providing helpful discussions in long-tailed object detection." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 182, + 197, + 194 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 182, + 197, + 194 + ], + "spans": [ + { + "bbox": [ + 133, + 182, + 197, + 194 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 206, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 138, + 206, + 480, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 206, + 480, + 239 + ], + "spans": [ + { + "bbox": [ + 138, + 206, + 480, + 239 + ], + "type": "text", + "content": "1. Alexandridis, K.P., Deng, J., Nguyen, A., Luo, S.: Long-tailed instance segmentation using gumbel optimized loss. In: ECCV. Lecture Notes in Computer Science, vol. 13670, pp. 353-369. Springer (2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 239, + 480, + 260 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 239, + 480, + 260 + ], + "spans": [ + { + "bbox": [ + 138, + 239, + 480, + 260 + ], + "type": "text", + "content": "2. Cai, Z., Vasconcelos, N.: Cascade r-cnn: Delving into high quality object detection. In: CVPR. pp. 6154-6162 (2018)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 261, + 480, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 261, + 480, + 282 + ], + "spans": [ + { + "bbox": [ + 138, + 261, + 480, + 282 + ], + "type": "text", + "content": "3. Cao, K., Wei, C., Gaidon, A., Arechiga, N., Ma, T.: Learning imbalanced datasets with label-distribution-aware margin loss. In: NeurIPS. pp. 1565-1576 (2019)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 284, + 480, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 284, + 480, + 326 + ], + "spans": [ + { + "bbox": [ + 138, + 284, + 480, + 326 + ], + "type": "text", + "content": "4. Caron, M., Bojanowski, P., Joulin, A., Douze, M.: Deep clustering for unsupervised learning of visual features. In: Ferrari, V., Hebert, M., Sminchisescu, C., Weiss, Y. (eds.) ECCV. Lecture Notes in Computer Science, vol. 11218, pp. 139-156. Springer (2018)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 327, + 480, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 327, + 480, + 380 + ], + "spans": [ + { + "bbox": [ + 138, + 327, + 480, + 380 + ], + "type": "text", + "content": "5. Chen, K., Wang, J., Pang, J., Cao, Y., Xiong, Y., Li, X., Sun, S., Feng, W., Liu, Z., Xu, J., Zhang, Z., Cheng, D., Zhu, C., Cheng, T., Zhao, Q., Li, B., Lu, X., Zhu, R., Wu, Y., Dai, J., Wang, J., Shi, J., Ouyang, W., Loy, C.C., Lin, D.: MMDetection: Open mmlab detection toolbox and benchmark. arXiv preprint arXiv:1906.07155 (2019)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 381, + 480, + 414 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 381, + 480, + 414 + ], + "spans": [ + { + "bbox": [ + 138, + 381, + 480, + 414 + ], + "type": "text", + "content": "6. Cheng, B., Girshick, R., Dollar, P., Berg, A.C., Kirillov, A.: Boundary iou: Improving object-centric image segmentation evaluation. In: CVPR. pp. 15334-15342 (2021)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 415, + 480, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 415, + 480, + 446 + ], + "spans": [ + { + "bbox": [ + 138, + 415, + 480, + 446 + ], + "type": "text", + "content": "7. Cho, J.H., Krahenbuhl, P.: Long-tail detection with effective class-margins. In: ECCV. Lecture Notes in Computer Science, vol. 13668, pp. 698-714. Springer (2022)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 447, + 480, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 447, + 480, + 468 + ], + "spans": [ + { + "bbox": [ + 138, + 447, + 480, + 468 + ], + "type": "text", + "content": "8. Cui, Y., Jia, M., Lin, T.Y., Song, Y., Belongie, S.: Class-balanced loss based on effective number of samples. In: CVPR. pp. 9268-9277 (2019)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 469, + 480, + 501 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 469, + 480, + 501 + ], + "spans": [ + { + "bbox": [ + 138, + 469, + 480, + 501 + ], + "type": "text", + "content": "9. Dave, A., Dóllár, P., Ramanan, D., Kirillov, A., Girshick, R.: Evaluating large-vocabulary object detectors: The devil is in the details. arXiv preprint arXiv:2102.01066 (2021)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 502, + 480, + 523 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 502, + 480, + 523 + ], + "spans": [ + { + "bbox": [ + 138, + 502, + 480, + 523 + ], + "type": "text", + "content": "10. Du, Y., Wu, J.: No one left behind: Improving the worst categories in long-tailed learning. In: CVPR. pp. 15804-15813 (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 524, + 480, + 545 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 524, + 480, + 545 + ], + "spans": [ + { + "bbox": [ + 138, + 524, + 480, + 545 + ], + "type": "text", + "content": "1. Feng, C., Zhong, Y., Huang, W.: Exploring classification equilibrium in long-tailed object detection. In: ICCV. pp. 3417-3426 (2021)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 545, + 480, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 545, + 480, + 567 + ], + "spans": [ + { + "bbox": [ + 138, + 545, + 480, + 567 + ], + "type": "text", + "content": "2. Gupta, A., Dollar, P., Girshick, R.: Lvis: A dataset for large vocabulary instance segmentation. In: CVPR. pp. 5356-5364 (2019)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 567, + 480, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 567, + 480, + 589 + ], + "spans": [ + { + "bbox": [ + 138, + 567, + 480, + 589 + ], + "type": "text", + "content": "3. Gupta, A., Dollar, P., Girshick, R.: Lvis: A dataset for large vocabulary instance segmentation. In: CVPR. pp. 5356-5364 (2019)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 590, + 480, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 590, + 480, + 610 + ], + "spans": [ + { + "bbox": [ + 138, + 590, + 480, + 610 + ], + "type": "text", + "content": "4. He, K., Gkioxari, G., Dollár, P., Girshick, R.: Mask R-CNN. In: ICCV. pp. 2961-2969 (2017)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 138, + 611, + 480, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 611, + 480, + 643 + ], + "spans": [ + { + "bbox": [ + 138, + 611, + 480, + 643 + ], + "type": "text", + "content": "5. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 770-778 (2016)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 138, + 644, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 644, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 644, + 480, + 665 + ], + "type": "text", + "content": "6. He, Y.Y., Wu, J., Wei, X.S.: Distilling virtual examples for long-tailed recognition. In: ICCV. pp. 235-244 (2021)" + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 334, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 334, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 334, + 91, + 447, + 102 + ], + "type": "text", + "content": "Rectify the Regression Bias" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 24, + "blocks": [ + { + "bbox": [ + 133, + 116, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 480, + 138 + ], + "type": "text", + "content": "17. He, Y.Y., Zhang, P., Wei, X.S., Zhang, X., Sun, J.: Relieving long-tailed instance segmentation via pairwise class balance. In: CVPR. pp. 7000-7009 (2022)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 139, + 480, + 159 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 139, + 480, + 159 + ], + "spans": [ + { + "bbox": [ + 133, + 139, + 480, + 159 + ], + "type": "text", + "content": "18. Hsieh, T., Robb, E., Chen, H., Huang, J.: Droploss for long-tail instance segmentation. In: AAAI. pp. 1549-1557 (2021)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 160, + 480, + 181 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 160, + 480, + 181 + ], + "spans": [ + { + "bbox": [ + 133, + 160, + 480, + 181 + ], + "type": "text", + "content": "19. Hu, X., Jiang, Y., Tang, K., Chen, J., Miao, C., Zhang, H.: Learning to segment the tail. In: CVPR (2020)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 182, + 480, + 202 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 182, + 480, + 202 + ], + "spans": [ + { + "bbox": [ + 132, + 182, + 480, + 202 + ], + "type": "text", + "content": "20. Kang, B., Xie, S., Rohrbach, M., Yan, Z., Gordo, A., Feng, J., Kalantidis, Y.: Decoupling representation and classifier for long-tailed recognition. In: ICLR (2020)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 203, + 480, + 224 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 203, + 480, + 224 + ], + "spans": [ + { + "bbox": [ + 132, + 203, + 480, + 224 + ], + "type": "text", + "content": "21. Li, B., Yao, Y., Tan, J., Zhang, G., Yu, F., Lu, J., Luo, Y.: Equalized focal loss for dense long-tailed object detection. In: CVPR. pp. 6990-6999 (2022)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 225, + 480, + 245 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 225, + 480, + 245 + ], + "spans": [ + { + "bbox": [ + 132, + 225, + 480, + 245 + ], + "type": "text", + "content": "22. Li, M., Cheung, Y.m., Lu, Y.: Long-tailed visual recognition via gaussian clouded logit adjustment. In: CVPR. pp. 6929-6938 (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 246, + 480, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 246, + 480, + 267 + ], + "spans": [ + { + "bbox": [ + 132, + 246, + 480, + 267 + ], + "type": "text", + "content": "23. Li, T., Wang, L., Wu, G.: Self supervision to distillation for long-tailed visual recognition. In: ICCV. pp. 630-639 (2021)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 268, + 480, + 299 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 268, + 480, + 299 + ], + "spans": [ + { + "bbox": [ + 132, + 268, + 480, + 299 + ], + "type": "text", + "content": "24. Li, Y., Wang, T., Kang, B., Tang, S., Wang, C., Li, J., Feng, J.: Overcoming classifier imbalance for long-tail object detection with balanced group softmax. In: CVPR (2020)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 300, + 480, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 300, + 480, + 331 + ], + "spans": [ + { + "bbox": [ + 132, + 300, + 480, + 331 + ], + "type": "text", + "content": "25. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dólár, P., Zitnick, C.L.: Microsoft COCO: Common objects in context. In: ECCV. Lecture Notes in Computer Science, vol. 8693, pp. 740-755. Springer (2014)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 332, + 480, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 332, + 480, + 353 + ], + "spans": [ + { + "bbox": [ + 132, + 332, + 480, + 353 + ], + "type": "text", + "content": "26. Liu, Z., Miao, Z., Zhan, X., Wang, J., Gong, B., Yu, S.X.: Large-scale long-tailed recognition in an open world. In: CVPR. pp. 2537-2546 (2019)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 354, + 480, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 354, + 480, + 374 + ], + "spans": [ + { + "bbox": [ + 132, + 354, + 480, + 374 + ], + "type": "text", + "content": "27. Menon, A.K., Jayasumana, S., Rawat, A.S., Jain, H., Veit, A., Kumar, S.: Long-tail learning via logit adjustment. In: ICLR (2021)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 375, + 480, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 375, + 480, + 396 + ], + "spans": [ + { + "bbox": [ + 132, + 375, + 480, + 396 + ], + "type": "text", + "content": "28. Pang, J., Chen, K., Shi, J., Feng, H., Ouyang, W., Lin, D.: Libra r-cnn: Towards balanced learning for object detection. In: CVPR. pp. 821-830 (2019)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 396, + 480, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 396, + 480, + 418 + ], + "spans": [ + { + "bbox": [ + 132, + 396, + 480, + 418 + ], + "type": "text", + "content": "29. Ren, J., Yu, C., sheng, s., Ma, X., Zhao, H., Yi, S., Li, h.: Balanced meta-softmax for long-tailed visual recognition. In: NeurIPS. pp. 4175-4186 (2020)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 418, + 480, + 438 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 418, + 480, + 438 + ], + "spans": [ + { + "bbox": [ + 132, + 418, + 480, + 438 + ], + "type": "text", + "content": "30. Ren, S., He, K., Girshick, R., Sun, J.: Faster R-CNN: Towards real-time object detection with region proposal networks. In: NeurIPS. pp. 91-99 (2015)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 439, + 480, + 482 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 439, + 480, + 482 + ], + "spans": [ + { + "bbox": [ + 132, + 439, + 480, + 482 + ], + "type": "text", + "content": "31. Russakovsky, O., Deng, J., Su, H., Krause, J., Satheesh, S., Ma, S., Huang, Z., Karpathy, A., Khosla, A., Bernstein, M., Berg, A., Fei-Fei, L.: ImageNet large scale visual recognition challenge. International Journal of Computer Vision 115(3), 211-252 (2015)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 483, + 480, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 483, + 480, + 503 + ], + "spans": [ + { + "bbox": [ + 132, + 483, + 480, + 503 + ], + "type": "text", + "content": "32. Singh, B., Davis, L.S.: An analysis of scale invariance in object detection SNIP. In: CVPR. pp. 3578-3587 (2018)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 504, + 480, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 504, + 480, + 525 + ], + "spans": [ + { + "bbox": [ + 132, + 504, + 480, + 525 + ], + "type": "text", + "content": "33. Tan, J., Lu, X., Zhang, G., Yin, C., Li, Q.: Equalization loss v2: A new gradient balance approach for long-tailed object detection. In: CVPR. pp. 1685-1694 (2021)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 525, + 480, + 546 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 525, + 480, + 546 + ], + "spans": [ + { + "bbox": [ + 132, + 525, + 480, + 546 + ], + "type": "text", + "content": "34. Tan, J., Wang, C., Li, B., Li, Q., Ouyang, W., Yin, C., Yan, J.: Equalization loss for long-tailed object recognition. In: CVPR (2020)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 132, + 547, + 480, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 547, + 480, + 578 + ], + "spans": [ + { + "bbox": [ + 132, + 547, + 480, + 578 + ], + "type": "text", + "content": "35. Wang, J., Zhang, W., Zang, Y., Cao, Y., Pang, J., Gong, T., Chen, K., Liu, Z., Loy, C.C., Lin, D.: Seesaw loss for long-tailed instance segmentation. In: CVPR. pp. 9695-9704 (2021)" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 132, + 579, + 480, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 579, + 480, + 622 + ], + "spans": [ + { + "bbox": [ + 132, + 579, + 480, + 622 + ], + "type": "text", + "content": "36. Wang, T., Li, Y., Kang, B., Li, J., Liew, J.H., Tang, S., Hoi, S.C.H., Feng, J.: The devil is in classification: A simple framework for long-tail instance segmentation. In: ECCV. Lecture Notes in Computer Science, vol. 12359, pp. 728-744. Springer (2020)" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 132, + 623, + 480, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 623, + 480, + 643 + ], + "spans": [ + { + "bbox": [ + 132, + 623, + 480, + 643 + ], + "type": "text", + "content": "37. Wang, T., Zhu, Y., Zhao, C., Zeng, W., Wang, J., Tang, M.: Adaptive class suppression loss for long-tail object detection. In: CVPR. pp. 3103-3112 (2021)" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 132, + 644, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 644, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 644, + 480, + 665 + ], + "type": "text", + "content": "38. Wang, X., Lian, L., Miao, Z., Liu, Z., Yu, S.: Long-tailed recognition by routing diverse distribution-aware experts. In: ICLR (2021)" + } + ] + } + ], + "index": 23 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "text", + "content": "K. Zhu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 347 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "text", + "content": "39. Wu, J., Song, L., Wang, T., Zhang, Q., Yuan, J.: Forest R-CNN: large-vocabulary long-tailed object detection and instance segmentation. In: ACM MM. pp. 1570-1578 (2020)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 150, + 481, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 150, + 481, + 172 + ], + "spans": [ + { + "bbox": [ + 130, + 150, + 481, + 172 + ], + "type": "text", + "content": "40. Xuan, G., Zhang, W., Chai, P.: EM algorithms of gaussian mixture model and hidden markov model. In: ICIP. pp. 145-148 (2001)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 172, + 481, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 172, + 481, + 205 + ], + "spans": [ + { + "bbox": [ + 132, + 172, + 481, + 205 + ], + "type": "text", + "content": "41. Zhang, C., Pan, T.Y., Li, Y., Hu, H., Xuan, D., Changpinyo, S., Gong, B., Chao, W.L.: Mosaicos: A simple and effective use of object-centric images for long-tailed object detection. In: ICCV. pp. 417-427 (2021)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 205, + 481, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 205, + 481, + 237 + ], + "spans": [ + { + "bbox": [ + 132, + 205, + 481, + 237 + ], + "type": "text", + "content": "42. Zhang, S., Chen, C., Peng, S.: Reconciling object-level and global-level objectives for long-tail detection. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 18982-18992 (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 238, + 481, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 238, + 481, + 270 + ], + "spans": [ + { + "bbox": [ + 132, + 238, + 481, + 270 + ], + "type": "text", + "content": "43. Zhong, Z., Cui, J., Liu, S., Jia, J.: Improving calibration for long-tailed recognition. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 16489-16498 (2021)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 133, + 271, + 481, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 271, + 481, + 303 + ], + "spans": [ + { + "bbox": [ + 133, + 271, + 481, + 303 + ], + "type": "text", + "content": "44. Zhou, B., Cui, Q., Wei, X.S., Chen, Z.M.: Bbn: Bilateral-branch network with cumulative learning for long-tailed visual recognition. In: CVPR. pp. 9716-9725 (2020)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 304, + 481, + 325 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 304, + 481, + 325 + ], + "spans": [ + { + "bbox": [ + 132, + 304, + 481, + 325 + ], + "type": "text", + "content": "45. Zhu, K., Fu, M., Wu, J.: Multi-label self-supervised learning with scene images. In: ICCV. pp. 6694-6703 (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 326, + 481, + 347 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 326, + 481, + 347 + ], + "spans": [ + { + "bbox": [ + 132, + 326, + 481, + 347 + ], + "type": "text", + "content": "46. Zhu, K., He, Y., Wu, J.: Quantized feature distillation for network quantization. In: AAAI. pp. 11452-11460 (2023)" + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 334, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 334, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 334, + 91, + 447, + 102 + ], + "type": "text", + "content": "Rectify the Regression Bias" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/RecurrentBEV_ A Long-term Temporal Fusion Framework for Multi-view 3D Detection/c66016d3-668d-4d64-aa39-f278cd973b7f_content_list.json b/2024/RecurrentBEV_ A Long-term Temporal Fusion Framework for Multi-view 3D Detection/c66016d3-668d-4d64-aa39-f278cd973b7f_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..184e6b594d09c3add01d3e1980afa12d0ddb398f --- /dev/null +++ b/2024/RecurrentBEV_ A Long-term Temporal Fusion Framework for Multi-view 3D Detection/c66016d3-668d-4d64-aa39-f278cd973b7f_content_list.json @@ -0,0 +1,1858 @@ +[ + { + "type": "text", + "text": "RecurrentBEV: A Long-term Temporal Fusion Framework for Multi-view 3D Detection", + "text_level": 1, + "bbox": [ + 236, + 140, + 767, + 185 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Ming Chang $^{1}$ , Xishan Zhang $^{1,2\\star}$ , Rui Zhang $^{2}$ , Zhipeng Zhao $^{1}$ , Guanhua He $^{1,2,3}$ , and Shaoli Liu $^{1}$", + "bbox": [ + 241, + 210, + 759, + 243 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Cambricon Technologies {changming,zhaozhipeng,liushaoli}@cambricon.com", + "bbox": [ + 228, + 253, + 772, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "2 SKL of Processors, Institute of Computing Technology, CAS, Beijing, China {zhangxishan,zhangrui}@ict.ac.cn, heguanhua22@mails.ucas.ac.cn", + "bbox": [ + 238, + 268, + 761, + 296 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3 University of Chinese Academy of Sciences, UCAS, Beijing, China", + "bbox": [ + 272, + 296, + 728, + 310 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Long-term temporal fusion is frequently employed in camera-based Bird's-Eye-View (BEV) 3D object detection to improve detection of occluded objects. Existing methods can be divided into two categories, parallel fusion and recurrent fusion. Recurrent fusion reduces inference latency and memory consumption but fails to exploit the long-term information as well as parallel fusion. In this paper, we first find two shortcomings of recurrent fusion paradigm: (1) Gradients of previous BEV features cannot directly contribute to the fusion module. (2) Semantic ambiguity are caused by coarse granularity of the BEV grids during aligning BEV features. Then based on the above analysis, we propose RecurrentBEV, a novel recurrent temporal fusion method for BEV based 3D object detector. By adopting RNN-style back-propagation and new-designed inner grid transformation, RecurrentBEV improves the long-term fusion ability while still enjoying efficient inference latency and memory consumption during inference. Extensive experiments on the nuScenes benchmark demonstrate its effectiveness, achieving a new state-of-the-art performance of $57.4\\%$ mAP and $65.1\\%$ NDS on the test set. The real-time version (25.6 FPS) achieves $44.5\\%$ mAP and $54.9\\%$ NDS without external dataset, outperforming the previous best method StreamPETR by $1.3\\%$ mAP and $0.9\\%$ NDS. The code is available at https://github.com/lucifer443/RecurrentBEV.", + "bbox": [ + 261, + 348, + 738, + 640 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Autonomous Driving $\\cdot$ Bird's-Eye-View $\\cdot$ 3D Object Detection $\\cdot$ Temporal Fusion", + "bbox": [ + 259, + 652, + 740, + 681 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 708, + 375, + 724 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Camera-only 3D detection presents a formidable challenge for autonomous driving and has garnered significant attention in recent years owing to its low cost of deployment. The Bird's-Eye-View (BEV) representation-based approach has recently demonstrated notable improvement by leveraging temporal information [8,11,21,22,33].", + "bbox": [ + 212, + 739, + 787, + 816 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Corresponding Author.", + "bbox": [ + 217, + 824, + 388, + 839 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/a0c2c30d14ddc03080547d6318c6c605f57895cc2abc4416a369c72668279d1c.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 218, + 146, + 488, + 306 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/1e566910a3ade95c9c7492688b428a3a9751acd0ed968f9126a98c27112afe4a.jpg", + "image_caption": [ + "(b)", + "Fig. 1: (a) Comparison of improvement $(\\Delta \\mathrm{mAP}\\uparrow)$ from long-term fusion between parallel paradigm (SOLOFusion [33]) and recurrent paradigm (VideoBEV [8], Ours). (b) Performance comparison on the validation set of nuScenes [2]. All methods use ResNet50 [9] pre-trained on ImageNet [6] as backbone with $256\\times 704$ as input size. Frames per second (FPS) is measured on a single RTX 3090 with the PyTorch fp32." + ], + "image_footnote": [], + "bbox": [ + 514, + 148, + 782, + 306 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The temporal fusion mechanisms utilised in the BEV-based 3D detector can be broadly categorised into parallel fusion [11,33,39], and recurrent fusion [8,22]. Parallel fusion [11,33,39] caches and updates multiple historical features in learning the visual features for current frames. Therefore, the overhead of computing and storage grows as the length of the historical frame feature increases. Since the superior performance requires longer historical information, this fusion paradigm faces a dilemma between effectiveness and efficiency in deployment. In contrast, recurrent fusion [22] encodes all historical features into a single recurrent feature to cache and update, so it largely reduces inference latency and memory consumption. For instance, when employing 16 historical frames, the recurrent fusion method VideoBEV [8] exhibits $6 \\times$ times faster in latency and uses $12 \\times$ less memory compared to the state-of-the-art parallel fusion method SOLOFusion [33]. However, we observe that recurrent fusion fails to exploit the long-term information as well as parallel fusion. As depicted in Fig.1a, VideoBEV benefits less ( $\\Delta$ mAP↑) from utilizing historical frames than that of SOLOFusion ( $\\Delta$ mAP↑). This phenomenon becomes even more pronounced, when utilizing more than four historical frames.", + "bbox": [ + 212, + 441, + 787, + 699 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Why is the performance of recurrent fusion not as good as parallel fusion? We conducted a detailed analysis of the differences between the two paradigms and found two reasons: (1) Gradient propagation. During the training process, only the recent frames contribute to gradients of fusion module in the recurrent paradigm, while all historical frames contribute equally in the parallel paradigm. This prevents the temporal fusion module of recurrent paradigm from learning fusion pattern from long-term temporal information of all historical frames. (2) Semantic ambiguity. Due to the coarse granularity of the BEV grids, the recurrent paradigm generates errors when aligning historical features multiple", + "bbox": [ + 212, + 704, + 787, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "M. Chang et al.", + "bbox": [ + 271, + 114, + 379, + 128 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "times, resulting in semantic ambiguity for the same object. As a result, recurrent paradigm may fuse different objects on the singular BEV grid, confusing the recognition of the same object. For the above two reasons, the previous recurrent method cannot fully and correctly utilize the long-term temporal information.", + "bbox": [ + 212, + 146, + 784, + 205 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Motivated by the above analysis, in this paper, we propose RecurrentBEV, a novel temporal fusion framework for BEV-based 3D object detector to improve the long-term fusion ability of recurrent paradigm while still has efficient latency and memory consumption in inference. First, RecurrentBEV adopts an RNN-style back-propagation during training, enabling historical frames to contribute gradients to the fusion module directly. This improvement solves the first problem while still maintaining inference efficiency. Second, we design an inner grid transformation before fusion, which can learn fine-grained semantic changes from ego-motion matrix. This effectively reduces semantic differences in the inter-frame alignment process.", + "bbox": [ + 212, + 207, + 785, + 356 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Based on the RNN-style back-propagation and the inner grid transformation, the proposed RecurrentBEV effectively improves long-term fusion in the recurrent paradigm, as shown in Fig.1a. Extensive experiments on the nuScenes 3D detection dataset evaluate the effectiveness of the proposed RecurrentBEV. As shown in Fig.1b, it outperforms both existing BEV-based algorithms such as SOLOFusion, VideoBEV and query-based algorithms such as StreamPETR [35], SparseBEV [27], Sparse4Dv2 [26] in both accuracy and efficiency. To summarize, our contributions are:", + "bbox": [ + 212, + 358, + 787, + 478 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "- We provide an in-depth analysis for two temporal fusion paradigms and point out two factors that render the recurrent fusion less effective: issues with gradient propagation and semantic ambiguity.", + "bbox": [ + 225, + 488, + 784, + 532 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "- We propose RecurrentBEV, a novel temporal fusion framework for BEV-based 3D object detector. It adapots RNN-style back-propagation and inner grid transformation, attaining state-of-the-art results (57.4% mAP and 65.1% NDS) on the nuScenes benchmark. It represents the first recurrent fusion method with explicit BEV features that surpasses parallel fusion performance.", + "bbox": [ + 225, + 534, + 785, + 622 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Works", + "text_level": 1, + "bbox": [ + 215, + 647, + 397, + 662 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.1 Multi-view 3D Object Detection", + "text_level": 1, + "bbox": [ + 214, + 678, + 529, + 694 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Multi-view 3D object detection is a fundamental task for autonomous driving. Previous works mainly focus on how to transform the 2D features into a unified 3D space. This transformation can be divided to two categories, BEV-based methods [10, 13, 18, 22, 23, 41, 42] and query-based methods [25, 28, 36, 37], depending on whether an explicit 3D representation is employed.", + "bbox": [ + 212, + 703, + 784, + 779 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "BEVDet [13] is one typical method of BEV-based multi-view 3D object detection. This method involves converting 2D features to BEV features based on the Lift-Splat-Shoot [34] paradigm. BEVDepth [21] is a further example, which notably refines the transformation process via explicit depth supervision.", + "bbox": [ + 212, + 779, + 785, + 840 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "RecurrentBEV", + "bbox": [ + 630, + 114, + 730, + 127 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 114, + 785, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "BEVFormer [22] further introduces a novel attention mechanism to gather 2D features for BEV grids.", + "bbox": [ + 212, + 146, + 782, + 175 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Query-based methods typically define several object queries in 3D space to gather features from multi-view image features. DETR3D [36] projects object queries to image coordinates to sample features then predicts 3D bounding boxes directly. DETR4D [31] and Sparse4D [25] enhance it by leveraging historical information. PETR [28] encodes the position information of 3D coordinates into image feature, facilitating the detection of 3D objects similar to DETR [3].", + "bbox": [ + 212, + 176, + 787, + 268 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.2 Temporal Fusion for BEV-based Detectors", + "text_level": 1, + "bbox": [ + 214, + 290, + 612, + 304 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Recent works find that incorporating temporal information can enhance the stability in perception results and improve overall performance. BEVFormer [22] first proposes the temporal self-attention mechanism to fuse the historical BEV features in a recurrent paradigm. BEVDet4D [11] extends BEVDet [13] to a temporal version by employing a concatenation operation to integrate previous BEV features, usually called parallel paradigm. SOLOFusion [33] extends these methods by increasing the number of previous frames to 16, resulting in significant benefits. BEVFormerv2 [39] switches to parallel paradigm same as SOLOFusion. However, the inference latency and memory consumption of the parallel paradigm increase linearly as the historical frame lengths. VideoBEV [8] applies recurrent fusion to the LSS-based 3D detector, resulting in significant reduction of inference latency and memory consumption. However, the improvement of long-term fusion in VideoBEV is limited in comparison to SOLOFusion. In order to combine the advantages of the two paradigms, we construct a novel temporal fusion framework based on efficient recurrent paradigm.", + "bbox": [ + 212, + 314, + 787, + 542 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3 Method", + "text_level": 1, + "bbox": [ + 214, + 564, + 330, + 579 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Converting multi-camera image features to bird's-eye-view (BEV) features can provide a unified surrounding environment representation for autonomous driving perception tasks. It involves transforming image data into a top-down representation of the environment, which often denoted as BEV features. The BEV detectors contain an image encoder, a 2D-3D view transformation, and a temporal fusion module. In this work, we present a new temporal fusion module, which can effectively aggregate spatio-temporal BEV features.", + "bbox": [ + 212, + 595, + 787, + 702 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1 Analysis of BEV-based Temporal Fusion Paradigms", + "text_level": 1, + "bbox": [ + 214, + 724, + 687, + 739 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "There are two kinds of temporal fusion paradigms in BEV detection, which are the effective parallel fusion and efficient recurrent fusion. The parallel fusion caches and updates multiple historical BEV features in learning the visual features for current frames. The recurrent fusion encodes all historical features into a single recurrent feature to cache and update. We conduct a detailed analysis to identify the disparities among them and pinpoint two crucial factors that cause", + "bbox": [ + 212, + 750, + 787, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "M. Chang et al.", + "bbox": [ + 271, + 114, + 379, + 128 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/327b47b23a4df406692a42d40688da02e529a5cc08c09b32bb2e34625d40be16.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 222, + 141, + 488, + 353 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/4396a3b608b529bd6704b4f031f15f2427c34399b933dd6f36f98335dffae64b.jpg", + "image_caption": [ + "(b)", + "Fig. 2: (a) Semantic ambiguity caused by grid sample operation. When ego vehicle turning right, BEV feature change from $B_{t-1}$ to $B_t$ . In fusion, the grid sample operation used for aligning $B_{t-1}$ and $B_t$ only aligns the position but ignores the orientation as shown in the blue and red dash lines. (b) Visualization of the semantic ambiguity. The green boxes represent the BEV feature change from $B_{t-1}$ , which have been transformed to align with the $B_t$ and decoded into bounding boxes. The blue boxes depict the actual bounding boxes in the $B_t$ . Notably, for static objects, there is a mismatch in the orientation of the green and blue boxes." + ], + "image_footnote": [], + "bbox": [ + 493, + 142, + 777, + 357 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "the sub-optimal performance of recurrent fusion: issues with gradient propagation and semantic ambiguity.", + "bbox": [ + 212, + 521, + 784, + 551 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Gradient propagation Aligning historical frames to current frame is essential for BEV-based temporal fusion. Given the previous BEV feature $B_{j}$ , this process can be written as:", + "bbox": [ + 212, + 574, + 787, + 619 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\widetilde {B} _ {j, i} = \\zeta \\left(B _ {j}, P _ {j, i}\\right) \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 436, + 628, + 785, + 648 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\widetilde{B}_{j,i}$ is the $j$ -th frame BEV feature in the ego coordinate of $i$ -th frame, $P_{j,i}$ is the ego-motion matrix between the $j$ -th frame and $i$ -th frame and $\\zeta$ refers to the ego-motion transformation function, specifically grid sample operation [15]. Based on Eq.(1), We can express the parallel fusion of the $i$ -th frame as:", + "bbox": [ + 212, + 660, + 787, + 723 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\bar {B} _ {i} = \\varphi \\left(\\widetilde {B} _ {i - k, i}, \\dots , \\widetilde {B} _ {i - 1, i}, B _ {i}, W\\right) \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 379, + 734, + 784, + 753 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\overline{B}_i$ is the fused BEV feature of $i$ -th frame, $\\varphi$ is the temporal fusion function (convolution [11,21,33,39] or deformable attention [22]) and $W$ refers the parameters of it. Similarly, the recurrent paradigm for the $i$ -th frame can be written as:", + "bbox": [ + 212, + 763, + 787, + 823 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\bar {B} _ {i} = \\varphi (\\zeta (\\bar {B} _ {i - 1}, P _ {i - 1, i}), B _ {i}, W) \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 388, + 823, + 784, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "RecurrentBEV", + "bbox": [ + 630, + 114, + 730, + 126 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 114, + 784, + 125 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/2164b740a825f3d9c2171889420426ab1448c78963eccbb54182acc39e379e06.jpg", + "image_caption": [ + "Fig. 3: Overall architecture of the proposed RecurrentBEV. It shows a training process in the $t$ -th frame, while arrows indicate forward propagation and backward propagation. The red arrows indicate the additional back-propagation introduced by our method. The details of the Transformation-aware Gate Recurrent Unit (TA-GRU) is illustrated in dotted box." + ], + "image_footnote": [], + "bbox": [ + 240, + 143, + 759, + 313 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Based on Eq.(3), we can calculate the gradient for the parameters in the temporal fusion function for parallel fusion.", + "bbox": [ + 212, + 424, + 787, + 455 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\frac {\\partial L}{\\partial W} = \\frac {\\partial L}{\\partial \\bar {B} _ {i}} \\cdot \\varphi^ {\\prime} \\left(\\widetilde {B} _ {i - k, i}, \\dots , \\widetilde {B} _ {i - 1, i}, B _ {i}\\right) \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 361, + 460, + 785, + 494 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In recurrent fusion, $B_{i-1}$ is detached so that the gradient of $W$ in recurrent fusion can be written as:", + "bbox": [ + 212, + 501, + 787, + 531 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\frac {\\partial L}{\\partial W} = \\frac {\\partial L}{\\partial \\bar {B} _ {i}} \\cdot \\varphi^ {\\prime} (\\zeta (\\bar {B} _ {i - 1}, P _ {i - 1, i}), B _ {i}) \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 370, + 536, + 785, + 570 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Comparing Eq.(4) and Eq.(5), it is evident that only the current BEV feature $B_{i}$ and one previous BEV feature $\\overline{B}_{i-1}$ affect the gradient of $W$ directly in recurrent paradigm while all previous BEV features within a window size $k$ contribute to gradient of $W$ in parallel paradigm. This is one reason that recurrent paradigm makes limited progress in long-term fusion.", + "bbox": [ + 212, + 577, + 787, + 654 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Semantic Ambiguity To optimize memory usage, the recurrent fusion approach only caches a single historical 3D BEV feature instead of storing a sequence of 2D features. Consequently, this approach encounters challenges with temporal alignment in the 3D space.", + "bbox": [ + 212, + 672, + 787, + 734 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In the current recurrent fusion method, we have observed semantic ambiguity issues arising during the temporal alignment process in the 3D space. As illustrated in Fig.2a, the coarse granularity of BEV grids results in inaccurate transformation of orientation information within each grid during the temporal alignment process represented by Eq.(1). Consequently, this process introduces incorrect object orientations in real-world detection scenarios, as evident in Fig.2b.", + "bbox": [ + 212, + 734, + 787, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "M. Chang et al.", + "bbox": [ + 271, + 114, + 379, + 128 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In the parallel fusion counterpart, temporal alignment in the 3D space is avoided by directly projecting a sequence of 2D view feature onto the current 3D BEV feature space [33, 39]. Assuming that the past perspective 2D view feature at time step $j$ is denoted as $C_j$ , the formulation for projecting it onto the BEV feature at time step $i$ is as follows:", + "bbox": [ + 214, + 146, + 787, + 220 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\widetilde {B} _ {j, i} = \\operatorname {V i e w T r a n s} \\left(C _ {j}, I _ {j}, P _ {j, i} \\cdot E _ {j}\\right) \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 377, + 232, + 785, + 250 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $I_{j}$ and $E_{j}$ denote camera intrinsic and extrinsic respectively. This additional view transformation in 2D space effectively aligns objects between frames, subsequently enhancing the performance of parallel fusion. Nevertheless, this approach is not feasible for recurrent fusion due to the unavailability of 2D features during the recurrent fusion process. This discrepancy serves as another contributing factor to the performance gap between the two fusion methods.", + "bbox": [ + 214, + 258, + 787, + 349 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.2 RecurrentBEV", + "text_level": 1, + "bbox": [ + 215, + 369, + 385, + 386 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Based on the two above-mentioned possible reasons, we design a novel temporal fusion framework called RecurrentBEV. Similar to existing BEV-based 3D object detectors [8,11,21,33], RecurrentBEV contains an image encoder, a 2D-3D view transformation, a temporal fusion module, a BEV encoder and a task head, as illustrated in Fig.3. The main difference between RecurrentBEV and previous recurrent fusion is the temporal fusion module which can learn from all historical frames directly and learn fine-grained ego-motion transform.", + "bbox": [ + 212, + 395, + 787, + 501 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "RNN-style Back-propagation In prior recurrent fusion methods [22], the $\\overline{B}_{i-1}$ in Eq.(3) is solely computed for forward propagation, but not for backward propagation. Thus, historical BEV features do not contribute to the gradient of temporal fusion module in Eq.(5), preventing optimizing the temporal fusion module with all the relevant historical BEV features. To address this issue, RecurrentBEV adopts RNN-style Back-propagation that all the relevant historical features back-propagate gradients to adjacent layers. In this context, the gradients of parameters in temporal fusion function is written as:", + "bbox": [ + 214, + 521, + 787, + 643 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\frac {\\partial L}{\\partial W} = \\sum_ {j = i - k} ^ {i} \\frac {\\partial L}{\\partial \\overline {{B}} _ {j}} \\cdot \\varphi^ {'} (\\zeta (\\overline {{B}} _ {j - 1}, P _ {j - 1, j}), B _ {j}) \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 351, + 652, + 785, + 696 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $k$ is the number of total historical frames using in a iteration of training process. As a result, each historical BEV feature makes an equal contribution to the gradient of the temporal fusion module, akin to the parallel paradigm.", + "bbox": [ + 212, + 704, + 785, + 750 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Computing the back-propagation along the temporal dimension for the entire network places significant demands on GPU memory. As shown in Table 1 (Setting 1), if all modules, including image encoders, undergo back-propagation along the temporal dimension, memory overflow issues arise even when the frame number reaches only 2. However, it is imperative for the training process to incorporate long-term frames to effectively learn temporal information (refer to", + "bbox": [ + 214, + 750, + 787, + 840 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "RecurrentBEV", + "bbox": [ + 630, + 114, + 730, + 127 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 114, + 785, + 126 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/a64676f109229dfd434eb1766195e494153d04208e13c200eee795cc5f70e197.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Training frames0124816
Setting 115,06730,105OOMOOMOOMOOM
Setting 215,06717,66118,73519,97523,57930,941
", + "bbox": [ + 305, + 143, + 697, + 200 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 1: The memory usage (MB) during training. OOM means out of memory on A100-40G with FP32 precision.", + "bbox": [ + 215, + 200, + 785, + 227 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Sec.4.6 for details). Therefore, we need to strike a balance between GPU memory usage (determined by the modules involved in back-propagation) and the length of the time sequence. In our design, as depicted by the red arrows in Figure 3, RecurrentBEV further computes all the gradients of the temporal fusion module (TA-GRU in the following) via back-propagation along the temporal dimension. Consequently, RecurrentBEV (Setting 2 in Table 1) can be trained with a batch size of 8 with FP32 precision on a single A100-40G GPU, effectively utilizing up to 16 consecutive frames.", + "bbox": [ + 212, + 268, + 787, + 388 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Transformation-aware Gated Recurrent Unit Because of the coarse granularity of the BEV grids, the semantic feature in a grid can not project correctly when model transforms the previous BEV features to the coordinate system of the current ego. We design an inner grid transformation to solve this problem. For simplicity, we use a linear layer to model this transformation:", + "bbox": [ + 212, + 409, + 787, + 486 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\widetilde {B} _ {j, i} = \\operatorname {L i n e a r} \\left(\\zeta \\left(B _ {j}, P _ {j, i}\\right), W _ {p}\\right) \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 393, + 493, + 785, + 512 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $W_{p}$ is the parameter of Linear. For simplicity, we assume different channels in a projected feature share same motion embedding, so that the $W_{p}$ can be decoupled as:", + "bbox": [ + 212, + 518, + 787, + 563 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\nW _ {p} = E _ {j, i} \\odot \\widetilde {W} _ {p} \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 444, + 563, + 785, + 580 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $E_{j,i} \\in \\mathbb{R}^{C_{in} \\times 1}$ is ego motion embedding from timestamp $j$ to timestamp $i$ , $\\widetilde{W}_p \\in \\mathbb{R}^{C_{in} \\times C_{out}}$ is a learnable weight and $\\odot$ denotes a Hadamard Product. Motion embedding $E_{j,i}$ is a high dimensional vector converted from the given ego motion matrix $P_{j,i}$ :", + "bbox": [ + 212, + 585, + 787, + 648 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\nE _ {j, i} = \\sigma (\\psi (P _ {j, i})) \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 441, + 648, + 785, + 666 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $\\sigma$ is a sigmoid activation and $\\psi$ is a multi-layer perception (MLP). Combining Eq.(8), Eq.(9) and Eq.(10), the complete inner grid transformation can be written as:", + "bbox": [ + 212, + 669, + 787, + 714 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\widetilde {B} _ {j, i} = \\operatorname {L i n e a r} \\left(\\zeta \\left(\\bar {B} _ {j}, P _ {j, i}\\right), \\sigma \\left(\\psi \\left(P _ {j, i}\\right)\\right) \\odot \\widetilde {W} _ {p}\\right) \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 346, + 737, + 785, + 757 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "For the RNN modules, we adopted the typical Convolutional Gated Recurrent Unit (ConvGRU) [1]. By combining the aforementioned inner grid transformation, we introduce a novel temporal fusion module named the Transformation-Aware Gated Recurrent Unit (TA-GRU). The entire process of temporal fusion is formulated as follows:", + "bbox": [ + 212, + 763, + 787, + 839 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "M. Chang et al.", + "bbox": [ + 271, + 114, + 379, + 128 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\nW _ {p} = \\sigma \\left(\\psi \\left(P _ {i - 1, i}\\right)\\right) \\odot \\widetilde {W} _ {p} \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 367, + 172, + 785, + 191 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\widetilde {B} _ {i - 1, i} = \\operatorname {L i n e a r} \\left(\\zeta \\left(\\bar {B} _ {i - 1}, P _ {i - 1, i}\\right), W _ {p}\\right) \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 346, + 194, + 785, + 213 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\nZ _ {i} = \\sigma \\left(\\operatorname {C o n v} \\left(\\left[ \\widetilde {B} _ {i - 1, i}, B _ {i} \\right], W _ {z}\\right)\\right) \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 372, + 217, + 785, + 234 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\nR _ {i} = \\sigma \\left(\\operatorname {C o n v} \\left(\\left[ \\widetilde {B} _ {i - 1, i}, B _ {i} \\right], W _ {r}\\right)\\right) \\tag {15}\n$$\n", + "text_format": "latex", + "bbox": [ + 370, + 238, + 785, + 256 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\nO _ {i} = \\tanh \\left(\\operatorname {C o n v} \\left(\\left[ R _ {i} \\odot \\widetilde {B} _ {i - 1, i}, B _ {i} \\right], W _ {o}\\right)\\right) \\tag {16}\n$$\n", + "text_format": "latex", + "bbox": [ + 370, + 260, + 785, + 279 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\bar {B} _ {i} = \\left(1 - Z _ {i}\\right) \\odot \\tilde {B} _ {i - 1, i} + Z _ {i} \\odot O _ {i} \\tag {17}\n$$\n", + "text_format": "latex", + "bbox": [ + 370, + 282, + 785, + 300 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where $[\\cdot, \\cdot]$ represents the concatenation of feature maps and $W_{z}, W_{z}, W_{o}$ are 2D-convolution kernels. Because of inner grid transformation, TA-GRU can learn fine-grained semantic changes from ego-motion matrix, thus removing wrong orientation.", + "bbox": [ + 212, + 311, + 787, + 371 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Inference Efficiency Comparing Eq.(2) and Eq.(3), parallel fusion requires $k$ historical frames to be cached while recurrent fusion requires only one. Additionally, the transformation process described in Eq.(1) or Eq.(6) also needs to be computed $k$ times in parallel fusion, whereas in recurrent fusion it only needs to be computed once. As a result, recurrent fusion, which has lower inference latency and smaller memory consumption, is more suitable for deployment in the industry.", + "bbox": [ + 212, + 396, + 787, + 501 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Similar to previous recurrent fusion methods [8, 22], RecurrentBEV adopts sequential inference which only needs to cache a single historical BEV feature $\\overline{B}$ , initialized with zero and updated at the end of the fusion.", + "bbox": [ + 212, + 502, + 787, + 547 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 215, + 573, + 375, + 590 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.1 Dataset and Metrics", + "text_level": 1, + "bbox": [ + 215, + 606, + 431, + 619 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We evaluate our method on the challenging nuScenes dataset [2]. The dataset comprises 1000 driving scenes, each lasting about 20 seconds, collected in Boston and Singapore, which are divided into 700, 150 and 150 scenes respectively for training, validation, and testing. The vehicle is equipped with 6 cameras for a complete 360-degree field of view. The nuScenes dataset annotates 23 object classes with accurate 3D bounding boxes for each scene at $2\\mathrm{Hz}$ . For evaluation purposes, 10 classes are selected, namely: car, truck, bus, trailer, construction vehicle, pedestrian, motorcycle, bicycle, barrier, and traffic cone. As for evaluation, we use the following metrics, mean average precision (mAP), the nuScenes detection score (NDS), and five true positive metrics including average translation error (ATE), average scale error (ASE), average orientation error (AOE), average velocity error (AVE), average attribute error (AAE).", + "bbox": [ + 212, + 632, + 787, + 815 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "RecurrentBEV", + "bbox": [ + 630, + 114, + 730, + 126 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 114, + 785, + 126 + ], + "page_idx": 8 + }, + { + "type": "page_footnote", + "text": "4 details of evaluation metrics: https://www.nuscenes.org/object-detection", + "bbox": [ + 217, + 824, + 718, + 840 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/682d094a8c6516605fd042a08d954b5db7042d23f8e8bda63b629d8db9a0b3a5.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodsBackboneImage SizeEDmAP↑NDS↑mATE↓mASE↓mAOE↓mAVE↓mAAE↓
BEVDet [13]ResNet50256 × 704X0.2980.3790.7250.2790.5890.8600.245
BEVDet4D [11]ResNet50256 × 704X0.3220.4570.7030.2780.4950.3540.206
BEVDepth [21]ResNet50256 × 704X0.3510.4750.6390.2670.4790.4280.198
BEVStereo [20]ResNet50256 × 704X0.3720.5000.5980.2700.4380.3670.190
BEVPoolv2 [12]ResNet50256 × 704X0.4060.5260.5720.2750.4630.2750.188
SOLOFusion [33]ResNet50256 × 704X0.4270.5340.5670.2740.5110.2520.181
VideoBEV [8]ResNet50256 × 704X0.4220.5350.5640.2760.4400.2860.198
PETRv2 [29]ResNet50256 × 704X0.3490.4560.7000.2750.5800.4370.187
Sparse4Dv2 [26]ResNet50256 × 704X0.4390.5390.5980.2700.4750.2820.179
SparseBEV [27]ResNet50256 × 704X0.4320.5450.6060.2740.3870.2510.186
StreamPETR [35]ResNet50256 × 704X0.4320.5400.5810.2720.4130.2950.195
RecurrentBEVResNet50256 × 704X0.4450.5490.5550.2720.4510.2560.204
BEVDepth [13]ResNet101512 × 1408X0.4120.5350.5650.2660.3580.3310.190
SOLOFusion [33]ResNet101512 × 1408X0.4830.5820.5030.2640.3810.2460.207
Sparse4Dv2 [26]ResNet101512 × 1408X0.4850.5800.5550.2720.3670.2560.182
RecurrentBEVResNet101512 × 1408X0.5090.5990.4870.2670.3580.2470.199
SparseBEV [27]ResNet101512 × 1408✓0.5010.5920.5620.2650.3210.2430.195
StreamPETR [35]ResNet101512 × 1408✓0.5040.5920.5690.2620.3150.2570.199
Sparse4Dv2 [26]ResNet101512 × 1408✓0.5050.5940.5480.2680.3480.2390.184
RecurrentBEVResNet101512 × 1408✓0.5280.6120.4810.2560.3550.2260.201
", + "bbox": [ + 222, + 143, + 779, + 404 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Table 2: Comparison results on 3D detection on the nuScenes validation set. ED is external data which indicates to use pre-trained weights from the external dataset i.e. nuImages.", + "bbox": [ + 215, + 405, + 784, + 445 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.2 Implementation Details", + "text_level": 1, + "bbox": [ + 215, + 502, + 455, + 517 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "We conduct experiments based on the BEVDet4D-Depth [11, 21] with the same learning rate, optimizer, and data augmentation. Drawing upon previous methods [21, 33, 35], we utilize ResNet50 [9], ResNet101, ConNeXt-B [30] and V2-99 [17] as backbones. On the nuScenes validation set, we report results of ResNet50 and ResNet101 models with pre-trained weights ImageNet [6] and nuImages [35]. On the nuScenes test set, we provide the performance of V2-99 model initialized from DD3D [32] and ConvNeXt-B model benefited from COCO [24] pre-training. The models in the ablation study are trained for 24 epochs without CBGS [43]. When compared with others, we follow the training setting of StreamPETR [35], extending the training epochs to 60 and using 2D auxiliary supervision for a fair comparison.", + "bbox": [ + 212, + 560, + 787, + 728 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "To achieve real-time inference speed, we optimize the BEVDet4D-Depth baseline by making several key modifications. Firstly, we reduce the number of depth bins. Secondly, we eliminate the ASPP [5] module from the DepthNet to minimize the latency associated with view transformation. Additionally, we merge 6 detection tasks performed by CenterHead [40] into a single task, significantly reducing post-processing time. These optimizations combined enable us to achieve a more efficient inference speed.", + "bbox": [ + 212, + 733, + 787, + 840 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "M. Chang et al.", + "bbox": [ + 271, + 114, + 377, + 128 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/65ec0d5ad1b4872404010cc113b0b4ffd7c5ad8c94d6b1b7f220ac1e410e5e58.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodsBackbonemAP↑NDS↑mATE↓mASE↓mAOE↓mAVE↓mAAE↓Conference
DETR3D* [36]V2-990.4120.4790.6410.2550.3940.8450.133CoRL2021
PETR [28]V2-990.4410.5040.5930.2490.3830.8080.132ECCV2022
UVTR [19]V2-990.4720.5510.5770.2530.3910.5080.123NIPS2022
BEVFormer [22]V2-990.4810.5690.5820.2560.3750.3780.126ECCV2022
PETRv2 [29]V2-990.4900.5820.5610.2430.3610.3430.120ICCV2023
PolarFormer [16]V2-990.4930.5720.5560.2560.3640.4390.127AAAI2023
BEVDepth [21]V2-990.5030.6000.4450.2450.3780.3200.126AAAI2023
VEDet [4]V2-990.5050.5850.5450.2440.3460.4210.123CVPR2023
MV2D [37]V2-990.5110.5960.5250.2430.3570.3570.120ICCV2023
CAPE [38]V2-990.5250.6100.5030.2420.3610.3060.114CVPR2023
BEVStereo [20]V2-990.5250.6100.4310.2460.3580.3570.138AAAI2023
SA-BEV [41]V2-990.5330.6240.4300.2410.3380.2820.139ICCV2023
FB-BEV [23]V2-990.5370.6240.4390.2500.3580.2700.128ICCV2023
StreamPETR [35]V2-990.5500.6360.4790.2390.3170.2410.119ICCV2023
SparseBEV [27]V2-990.5560.6360.4850.2440.3320.2460.117ICCV2023
Sparse4Dv2 [26]V2-990.5570.6380.4620.2380.3280.2640.115-
RecurrentBEVV2-990.5730.6510.4100.2470.3230.2490.123-
BEVDepth [21]ConvNeXt-B0.5200.6090.4450.2430.3520.3470.127AAAI2023
AeDet* [7]ConvNeXt-B0.5310.6200.4390.2470.3440.2920.130CVPR2023
SOLOFusion [33]ConvNeXt-B0.5400.6190.4530.2570.3760.2760.148ICLR2023
VCD-A [14]ConvNeXt-B0.5480.6310.4360.2440.3430.2900.120NIPS2023
StreamPETR [35]ConvNeXt-B0.5500.6310.4930.2410.3430.2430.123ICCV2023
VideoBEV [8]ConvNeXt-B0.5540.6290.4570.2490.3810.2660.132-
RecurrentBEVConvNeXt-B0.5740.6510.4110.2480.3330.2570.116-
", + "bbox": [ + 225, + 143, + 776, + 444 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Table 3: Comparison results on 3D detection on the nuScenes test set. * indicates using test-time augmentation in test process. No future frames are used in any of the methods.", + "bbox": [ + 215, + 445, + 784, + 484 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4.3 Main Results", + "text_level": 1, + "bbox": [ + 215, + 532, + 372, + 546 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "We compare the proposed RecurrentBEV with previous state-of-the-art multiview 3D detectors on the nuScenes validation and test set.", + "bbox": [ + 212, + 564, + 784, + 594 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "On the nuScenes validation set, as shown in Fig.1b and Tab.2, the real-time version of RecurrentBEV (25.6 FPS), using ResNet50 as backbone and $256 \\times 704$ as the input size, outperforms the previous state-of-the-art method StreamPETR by $1.3\\%$ mAP and $0.9\\%$ NDS on the nuScenes validation set. To scale up our method, we adopt ResNet101 backbone with nuImages pre-training and extend input size to $512 \\times 1408$ , outperforming the same setting StreamPETR by $2.4\\%$ mAP and $2.0\\%$ NDS. It's worth noting that RecurrentBEV makes remarkable progress on the mAP and mATE metrics, indicating the improvement of detection of occlusion objects. We'll delve into the reasons behind this in Sec.4.6.", + "bbox": [ + 212, + 595, + 785, + 733 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "On the nuScenes test set in Tab.3, we compare with all the recent methods in multi-view 3D object detection. Following the settings in StreamPETR, we adopt V2-99 and ConvNeXt-B as backbone and $640 \\times 1600$ as the input size. Using V2-99 as backbone, RecurrentBEV can surpass the previous state-of-the-art method Sparse4Dv2 by $1.6\\%$ mAP and $1.3\\%$ NDS. When the backbone is scaled up to ConvNeXt-B, RecurrentBEV demonstrates comparable performance, surpassing other methods.", + "bbox": [ + 212, + 734, + 785, + 839 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "RecurrentBEV", + "bbox": [ + 630, + 114, + 730, + 126 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/e1fb0d106c14c50d0db68c81aaa2e50ada8fe706391face9337d818eb8e45b1e.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
methodmAP↑NDS↑memory usage↓latency↑
parallel fusion [33]0.4270.534945MB96ms
recurrent fusion baseline0.3960.515516MB34ms
RecurrentBEV0.4450.549549MB39ms
", + "bbox": [ + 289, + 143, + 712, + 214 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Table 4: Comparison of time and memory in inference. All data are measured on a single RTX 3090 with PyTorch fp32. Implementation details are the same as Tab.2 ResNet50 Backbone. For more latency analysis, please refer to supplementary.", + "bbox": [ + 212, + 214, + 785, + 257 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "4.4 Time and Memory Comparison", + "text_level": 1, + "bbox": [ + 215, + 299, + 521, + 315 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "To demonstrate the time and memory efficiency in inference, we compare the proposed RecurrentBEV with parallel fusion methods and previous recurrent fusion baselines. As shown in Tab.4 RecurrentBEV is 2.5 times faster than the normal parallel fusion method SOLOFusion [33], and the memory usage is reduced by $42\\%$ . When compared to the normal recurrent baseline, RecurrentBEV requires only an additional 33MB of memory while significantly improving the mean Average Precision (mAP) from 0.396 to 0.445.", + "bbox": [ + 212, + 325, + 787, + 431 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "4.5 Ablation Study", + "text_level": 1, + "bbox": [ + 215, + 455, + 388, + 470 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "In the ablation study, we delve into three aspects: the components of RecurrentBEV, the receptive field of TA-GRU, and the length of frame sequence in training. We conduct the ablation experiments with ResNet-50 backbone.", + "bbox": [ + 212, + 483, + 785, + 527 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Ablation of Each Component. We conduct a component-wise ablation in Tab.5. Exp.1 starts with a ConvGRU as temporal fusion module, achieving similar performance to VideoBEV [8]. In Exp.2 and Exp.3, we enable the backward propagation of historical BEV features in training process. In Exp.3 and Exp.4, we replace the temporal fusion module with TA-GRU. The overall components (BP and TA-GRU) brings significant performance gains $(+1.7\\% \\mathrm{mAP}, +1.6\\%$ NDS,-4.7% mAOE) over baseline.", + "bbox": [ + 212, + 527, + 787, + 633 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Specifically, the TA-GRU demonstrates its ability to enhance the mAP and mAOE metrics, regardless of whether backward propagation of historical BEV features is enabled. This proves the TA-GRU's consistent ability to alleviate semantic ambiguity for identical objects. Notably, the mean average orientation error (mAOE) has a notable increase, indicating that the inner grid transformation, facilitated by the alignment of historical BEV features, effectively mitigates semantic ambiguity by refining the orientation of objects.", + "bbox": [ + 212, + 633, + 787, + 739 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "As demonstrated in Exp.4, when the historical BEV feature is detached, the performance suffers, highlighting the importance of RNN-style back-propagation. However, relying solely on back-propagation of historical BEV features during training results in limited improvement. This shows the necessity of employing", + "bbox": [ + 212, + 741, + 787, + 801 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "M. Chang et al.", + "bbox": [ + 271, + 114, + 380, + 128 + ], + "page_idx": 11 + }, + { + "type": "page_footnote", + "text": "similar performance to VideoBEV [8] without stereo matching (VideoBEV-D) with 60 training epochs.", + "bbox": [ + 217, + 809, + 785, + 840 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/226a2d686539e788e37dd02f792775c258931569d2d5252d7ab37d6b2008bc02.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Exp|BP TG|mAP↑|NDS↑|mAOE↓|mAVE↓
10.3710.4760.6460.317
2✓0.3750.4790.6340.319
3✓✓0.3880.4920.5960.281
4✓0.3780.4840.6020.316
", + "bbox": [ + 354, + 143, + 647, + 227 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/204c57375c992e61c6076c20f9467a8446a4f7c659c16c2b441cb5c4ec2ef4c5.jpg", + "table_caption": [ + "Table 5: Ablation of each component. BP means the historical BEV feature backward propagate. TG means TA-GRU." + ], + "table_footnote": [], + "table_body": "
Kernel sizemAP↑NDS↑mATE↓mAVE↓
1 × 10.3880.4920.6470.281
3 × 30.3970.5000.6320.284
5 × 50.4030.5060.6150.276
7 × 70.3980.5010.6300.273
9 × 90.3940.4960.6330.280
", + "bbox": [ + 235, + 273, + 478, + 358 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/f79157ca373f0cf5aff8882bc5d3d7a702e94e704ac0940d010d56f832deb659.jpg", + "table_caption": [ + "Table 6: Ablation of receptive field." + ], + "table_footnote": [], + "table_body": "
Training framesmAP↑NDS↑mATE↓mAVE↓
00.3180.3800.7420.835
10.3460.4570.6960.389
20.3720.4790.6650.310
40.3900.4930.6460.283
80.4030.5060.6150.276
160.4020.490.6280.280
", + "bbox": [ + 506, + 263, + 774, + 358 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Table 7: Ablation of sequence length.", + "bbox": [ + 509, + 358, + 767, + 371 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "the correct temporal fusion module for effective back-propagation, as the combination of these two components can achieve a synergistic effect that surpasses the sum of their individual contributions.", + "bbox": [ + 212, + 387, + 784, + 431 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Ablation of Receptive Field. Receptive field is critical for temporal fusion of moving objects. We explore the kernel size of convolutions in TA-GRU in Tab.6. As we can see, an excessively large or small receptive field can have a detrimental effect on performance. Our suspect is that a fusion module with an overlarge receptive field gathers noisy information, while a small receptive field leads to mismatches for moving objects. Consequently, we use a $5 \\times 5$ kernel size in TA-GRU, unless otherwise specified.", + "bbox": [ + 212, + 431, + 784, + 537 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Ablation of Training Sequence Length. We use the BEVDet-Depth with no temporal fusion as baseline and apply our method with different training sequence length in Tab.7. The fusion of a single frame greatly enhances velocity prediction (mAVE). The fusion of four to eight frames mainly enhances localisation metrics mAP and mATE, with an increase of $8.5\\%$ and $12.7\\%$ respectively. Further increasing the number of frames to 16 offers gains limited performance improvement. Therefore, we train our models on 8 frames.", + "bbox": [ + 212, + 537, + 785, + 643 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "4.6 Effect of Recurrent Long-term Fusion", + "text_level": 1, + "bbox": [ + 215, + 665, + 571, + 680 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Detection of Occlusion Objects. Intuitively, long-term fusion has the potential to aid in the detection of occluded objects. Therefore, we conducted an analysis of object occlusions on the nuScenes validation set. Firstly, we define a 'distinct object' as one where the visibility of the entire object exceeds $60\\%$ . Subsequently, we determined that if an object is classified as a 'distinct object' in any frame of the video, it is considered visible throughout the video. Finally, we divided the number of visible objects by the total number of objects present to get the overall visibility of the video. As shown in Fig.4a, the visibility of the video has the same trend with mAP, proving the improvement of mAP in long-term fusion originates from the detection of occlusion objects.", + "bbox": [ + 212, + 688, + 787, + 840 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "RecurrentBEV", + "bbox": [ + 630, + 114, + 730, + 127 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/812dfcd2a3b4fa1490e5d506dfd24ae98b085b72a1abee35dca80c25252e598c.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 218, + 146, + 500, + 294 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/398adb6d4e1c1b0ed0a2dc7fd1f5ab3add939cdb6f29f70d65c2dd144c0e846d.jpg", + "image_caption": [ + "(b)", + "Fig. 4: (a) Relation between visibility and mAP. Visibility refers to the proportion of distinct objects in a video, expressed as a percentage of the total number of objects present. The Pearson correlation coefficient between visibility and mAP is 0.977.(b) The mATE metrics with objects at different location on the nuScenes validation set. The blue one is a result from single frame baseline and the orange one is RecurrentBEV. The objects between 25 meters and 30 meters improve the mATE metric by $24\\%$ mostly." + ], + "image_footnote": [], + "bbox": [ + 526, + 146, + 774, + 292 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Localization of Distant Objects. The localization of objects is a major challenge for camera-based 3D detector. Due to the uncertainty of depth estimation. We speculate long-term fusion can more accurately localize objects that move from near to far as the ego vehicle moves. Therefore, we calculate the mATE metrics of objects at different distances before and after long-term fusion, as in Fig.4b. Notably, objects between 25 and 30 meters exhibit the most significant improvement in mATE through long-term fusion, whereas nearby objects show marginal gains. This finding demonstrates the efficacy of long-term fusion in accurately localizing distant objects.", + "bbox": [ + 212, + 422, + 787, + 559 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Conclusion and Future Work", + "text_level": 1, + "bbox": [ + 215, + 588, + 532, + 604 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In this paper, we investigate the differences between recurrent fusion and parallel fusion of BEV-based 3D object detectors, and propose RecurrentBEV, a simple and effective recurrent temporal fusion method. By incorporating RNN-style back-propagation and a newly designed inner grid transformation, RecurrentBEV enhances long-term fusion capabilities while maintaining latency and memory advantages. Rigorous experiments conducted on the nuScenes dataset demonstrate that RecurrentBEV outperforms all prior methods, while maintaining competitive inference speeds, thereby facilitating deployment. To our knowledge, it represents the first recurrent fusion method with explicit BEV features that surpasses parallel fusion performance.", + "bbox": [ + 212, + 627, + 787, + 777 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In the future, we will generalize recurrent fusion idea to more BEV architectures such as different BEV view transform backbones and query-based methods. We will also try to extend RecurrentBEV to other 3D perceptual tasks and datasets such as object tracking and trajectory prediction.", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "M. Chang et al.", + "bbox": [ + 271, + 114, + 380, + 128 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 217, + 143, + 401, + 162 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "This work is partially supported by the National Key R&D Program of China (under Grant 2023YFB4502200), the NSF of China (under Grants U22A2028, 61925208, 62102399, 62222214, 62341411, 62102398, U20A20227, 62372436, 623-02478, 62302482, 62302483, 62302480), Strategic Priority Research Program of the Chinese Academy of Sciences, (Grant No.XDB0660200, XDB0660201, XDB0660202), CAS Project for Young Scientists in Basic Research(YSBR-029), Youth Innovation Promotion Association CAS and Xplore Prize.", + "bbox": [ + 212, + 175, + 787, + 282 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 301, + 321, + 318 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Ballas, N., Yao, L., Pal, C., Courville, A.: Delving deeper into convolutional networks for learning video representations. In: ICLR (2016)", + "2. Caesar, H., Bankiti, V., Lang, A.H., Vora, S., Liong, V.E., Xu, Q., Krishnan, A., Pan, Y., Baldan, G., Beijbom, O.: nuscenes: A multimodal dataset for autonomous driving. In: CVPR (2020)", + "3. Carion, N., Massa, F., Synnaeve, G., Usunier, N., Kirillov, A., Zagoruyko, S.: End-to-end object detection with transformers. In: ECCV (2020)", + "4. Chen, D., Li, J., Guizilini, V., Ambrus, R.A., Gaidon, A.: Viewpoint equivariance for multi-view 3d object detection. In: CVPR (2023)", + "5. Chen, L.C., Papandreou, G., Kokkinos, I., Murphy, K., Yuille, A.L.: Deeplab: Semantic image segmentation with deep convolutional nets, atrous convolution, and fully connected crfs. IEEE TPAMI (2017)", + "6. Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: Imagenet: A large-scale hierarchical image database. In: CVPR (2009)", + "7. Feng, C., Jie, Z., Zhong, Y., Chu, X., Ma, L.: Aedet: Azimuth-invariant multi-view 3d object detection. In: CVPR (2023)", + "8. Han, C., Sun, J., Ge, Z., Yang, J., Dong, R., Zhou, H., Mao, W., Peng, Y., Zhang, X.: Exploring recurrent long-term temporal fusion for multi-view 3d perception. arXiv preprint arXiv:2303.05970 (2023)", + "9. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: CVPR (2016)", + "0. Huang, B., Li, Y., Xie, E., Liang, F., Wang, L., Shen, M., Liu, F., Wang, T., Luo, P., Shao, J.: Fast-bev: Towards real-time on-vehicle bird's-eye view perception. arXiv preprint arXiv:2301.07870 (2023)", + "1. Huang, J., Huang, G.: Bevdet4d: Exploit temporal cues in multi-camera 3d object detection. arXiv preprint arXiv:2203.17054 (2022)", + "2. Huang, J., Huang, G.: Bevpoolv2: A cutting-edge implementation of bevdet toward deployment. arXiv preprint arXiv:2211.17111 (2022)", + "3. Huang, J., Huang, G., Zhu, Z., Ye, Y., Du, D.: Bevdet: High-performance multicamera 3d object detection in bird-eye-view. arXiv preprint arXiv:2112.11790 (2021)", + "4. Huang, L., Li, Z., Sima, C., Wang, W., Wang, J., Qiao, Y., Li, H.: Leveraging vision-centric multi-modal expertise for 3d object detection. In: NeurIPS (2023)", + "5. Jaderberg, M., Simonyan, K., Zisserman, A., et al.: Spatial transformer networks. In: NeurIPS (2015)", + "6. Jiang, Y., Zhang, L., Miao, Z., Zhu, X., Gao, J., Hu, W., Jiang, Y.G.: Polarformer: Multi-camera 3d object detection with polar transformer. In: AAAI (2023)" + ], + "bbox": [ + 225, + 332, + 784, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "RecurrentBEV", + "bbox": [ + 630, + 114, + 730, + 127 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "17. Lee, Y., Hwang, J.w., Lee, S., Bae, Y., Park, J.: An energy andgpu-computation efficient backbone network for real-time object detection. In: CVPRW (2019)", + "18. Li, H., Zhang, H., Zeng, Z., Liu, S., Li, F., Ren, T., Zhang, L.: Dfa3d: 3d deformable attention for 2d-to-3d feature lifting. In: ICCV (2023)", + "19. Li, Y., Chen, Y., Qi, X., Li, Z., Sun, J., Jia, J.: Unifying voxel-based representation with transformer for 3d object detection. In: NeurIPS (2022)", + "20. Li, Y., Bao, H., Ge, Z., Yang, J., Sun, J., Li, Z.: Bevstereo: enhancing depth estimation in multi-view 3d object detection with temporal stereo. In: AAAI (2023)", + "21. Li, Y., Ge, Z., Yu, G., Yang, J., Wang, Z., Shi, Y., Sun, J., Li, Z.: Bevdepth: Acquisition of reliable depth for multi-view 3d object detection. In: AAAI (2023)", + "22. Li, Z., Wang, W., Li, H., Xie, E., Sima, C., Lu, T., Qiao, Y., Dai, J.: Bevformer: Learning bird's-eye-view representation from multi-camera images via spatiotemporal transformers. In: ECCV (2022)", + "23. Li, Z., Yu, Z., Wang, W., Anandkumar, A., Lu, T., Alvarez, J.M.: Fb-bev: Bev representation from forward-backward view transformations. In: ICCV (2023)", + "24. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dóllar, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: ECCV (2014)", + "25. Lin, X., Lin, T., Pei, Z., Huang, L., Su, Z.: Sparse4d: Multi-view 3d object detection with sparse spatial-temporal fusion. arXiv preprint arXiv:2211.10581 (2022)", + "26. Lin, X., Lin, T., Pei, Z., Huang, L., Su, Z.: Sparse4d v2: Recurrent temporal fusion with sparse model. arXiv preprint arXiv:2305.14018 (2023)", + "27. Liu, H., Teng, Y., Lu, T., Wang, H., Wang, L.: Sparsebev: High-performance sparse 3d object detection from multi-camera videos. In: ICCV (2023)", + "28. Liu, Y., Wang, T., Zhang, X., Sun, J.: Petr: Position embedding transformation for multi-view 3d object detection. In: ECCV (2022)", + "29. Liu, Y., Yan, J., Jia, F., Li, S., Gao, A., Wang, T., Zhang, X.: Petrov2: A unified framework for 3d perception from multi-camera images. In: ICCV (2023)", + "30. Liu, Z., Mao, H., Wu, C.Y., Feichtenhofer, C., Darrell, T., Xie, S.: A convnet for the 2020s. In: CVPR (2022)", + "31. Luo, Z., Zhou, C., Zhang, G., Lu, S.: Detr4d: Direct multi-view 3d object detection with sparse attention. arXiv preprint arXiv:2212.07849 (2022)", + "32. Park, D., Ambrus, R., Guizilini, V., Li, J., Gaidon, A.: Is pseudo-lidar needed for monocular 3d object detection? In: ICCV (2021)", + "33. Park, J., Xu, C., Yang, S., Keutzer, K., Kitani, K., Tomizuka, M., Zhan, W.: Time will tell: New outlooks and a baseline for temporal multi-view 3d object detection. In: ICLR (2023)", + "34. Philion, J., Fidler, S.: Lift, splat, shoot: Encoding images from arbitrary camera rigs by implicitly unprojecting to 3d. In: ECCV (2020)", + "35. Wang, S., Liu, Y., Wang, T., Li, Y., Zhang, X.: Exploring object-centric temporal modeling for efficient multi-view 3d object detection. In: ICCV (2023)", + "36. Wang, Y., Guizilini, V.C., Zhang, T., Wang, Y., Zhao, H., Solomon, J.: Detr3d: 3d object detection from multi-view images via 3d-to-2d queries. In: Conference on Robot Learning (2022)", + "37. Wang, Z., Huang, Z., Fu, J., Wang, N., Liu, S.: Object as query: Lifting any 2d object detector to 3d detection. In: ICCV (2023)", + "38. Xiong, K., Gong, S., Ye, X., Tan, X., Wan, J., Ding, E., Wang, J., Bai, X.: Cape: Camera view position embedding for multi-view 3d object detection. In: CVPR (2023)", + "39. Yang, C., Chen, Y., Tian, H., Tao, C., Zhu, X., Zhang, Z., Huang, G., Li, H., Qiao, Y., Lu, L., et al.: Bevformer v2: Adapting modern image backbones to bird's-eye-view recognition via perspective supervision. In: CVPR (2023)" + ], + "bbox": [ + 215, + 147, + 784, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "M. Chang et al.", + "bbox": [ + 271, + 114, + 377, + 128 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "40. Yin, T., Zhou, X., Krahenbuhl, P.: Center-based 3d object detection and tracking. In: CVPR (2021)", + "41. Zhang, J., Zhang, Y., Liu, Q., Wang, Y.: Sa-bev: Generating semantic-aware bird's-eye-view feature for multi-view 3d object detection. In: ICCV (2023)", + "42. Zhou, H., Ge, Z., Li, Z., Zhang, X.: Matrixvt: Efficient multi-camera to bev transformation for 3d perception. In: ICCV (2023)", + "43. Zhu, B., Jiang, Z., Zhou, X., Li, Z., Yu, G.: Class-balanced grouping and sampling for point cloud 3d object detection. arXiv preprint arXiv:1908.09492 (2019)" + ], + "bbox": [ + 215, + 146, + 784, + 258 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "RecurrentBEV", + "bbox": [ + 630, + 114, + 730, + 126 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + } +] \ No newline at end of file diff --git a/2024/RecurrentBEV_ A Long-term Temporal Fusion Framework for Multi-view 3D Detection/c66016d3-668d-4d64-aa39-f278cd973b7f_model.json b/2024/RecurrentBEV_ A Long-term Temporal Fusion Framework for Multi-view 3D Detection/c66016d3-668d-4d64-aa39-f278cd973b7f_model.json new file mode 100644 index 0000000000000000000000000000000000000000..ddacca41494cef1d4da0d709a6725b42c168e82e --- /dev/null +++ b/2024/RecurrentBEV_ A Long-term Temporal Fusion Framework for Multi-view 3D Detection/c66016d3-668d-4d64-aa39-f278cd973b7f_model.json @@ -0,0 +1,2357 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.237, + 0.141, + 0.768, + 0.186 + ], + "angle": 0, + "content": "RecurrentBEV: A Long-term Temporal Fusion Framework for Multi-view 3D Detection" + }, + { + "type": "text", + "bbox": [ + 0.243, + 0.212, + 0.761, + 0.244 + ], + "angle": 0, + "content": "Ming Chang\\(^{1}\\), Xishan Zhang\\(^{1,2\\star}\\), Rui Zhang\\(^{2}\\), Zhipeng Zhao\\(^{1}\\), Guanhua He\\(^{1,2,3}\\), and Shaoli Liu\\(^{1}\\)" + }, + { + "type": "text", + "bbox": [ + 0.23, + 0.255, + 0.774, + 0.269 + ], + "angle": 0, + "content": "1 Cambricon Technologies {changming,zhaozhipeng,liushaoli}@cambricon.com" + }, + { + "type": "text", + "bbox": [ + 0.239, + 0.269, + 0.762, + 0.297 + ], + "angle": 0, + "content": "2 SKL of Processors, Institute of Computing Technology, CAS, Beijing, China {zhangxishan,zhangrui}@ict.ac.cn, heguanhua22@mails.ucas.ac.cn" + }, + { + "type": "text", + "bbox": [ + 0.273, + 0.297, + 0.729, + 0.311 + ], + "angle": 0, + "content": "3 University of Chinese Academy of Sciences, UCAS, Beijing, China" + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.349, + 0.74, + 0.641 + ], + "angle": 0, + "content": "Abstract. Long-term temporal fusion is frequently employed in camera-based Bird's-Eye-View (BEV) 3D object detection to improve detection of occluded objects. Existing methods can be divided into two categories, parallel fusion and recurrent fusion. Recurrent fusion reduces inference latency and memory consumption but fails to exploit the long-term information as well as parallel fusion. In this paper, we first find two shortcomings of recurrent fusion paradigm: (1) Gradients of previous BEV features cannot directly contribute to the fusion module. (2) Semantic ambiguity are caused by coarse granularity of the BEV grids during aligning BEV features. Then based on the above analysis, we propose RecurrentBEV, a novel recurrent temporal fusion method for BEV based 3D object detector. By adopting RNN-style back-propagation and new-designed inner grid transformation, RecurrentBEV improves the long-term fusion ability while still enjoying efficient inference latency and memory consumption during inference. Extensive experiments on the nuScenes benchmark demonstrate its effectiveness, achieving a new state-of-the-art performance of \\(57.4\\%\\) mAP and \\(65.1\\%\\) NDS on the test set. The real-time version (25.6 FPS) achieves \\(44.5\\%\\) mAP and \\(54.9\\%\\) NDS without external dataset, outperforming the previous best method StreamPETR by \\(1.3\\%\\) mAP and \\(0.9\\%\\) NDS. The code is available at https://github.com/lucifer443/RecurrentBEV." + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.654, + 0.741, + 0.682 + ], + "angle": 0, + "content": "Keywords: Autonomous Driving \\(\\cdot\\) Bird's-Eye-View \\(\\cdot\\) 3D Object Detection \\(\\cdot\\) Temporal Fusion" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.709, + 0.377, + 0.725 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.74, + 0.788, + 0.817 + ], + "angle": 0, + "content": "Camera-only 3D detection presents a formidable challenge for autonomous driving and has garnered significant attention in recent years owing to its low cost of deployment. The Bird's-Eye-View (BEV) representation-based approach has recently demonstrated notable improvement by leveraging temporal information [8,11,21,22,33]." + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.825, + 0.39, + 0.84 + ], + "angle": 0, + "content": "* Corresponding Author." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.38, + 0.129 + ], + "angle": 0, + "content": "M. Chang et al." + }, + { + "type": "image", + "bbox": [ + 0.219, + 0.147, + 0.49, + 0.308 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.345, + 0.313, + 0.365, + 0.325 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.515, + 0.149, + 0.783, + 0.308 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.638, + 0.313, + 0.66, + 0.325 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.336, + 0.788, + 0.407 + ], + "angle": 0, + "content": "Fig. 1: (a) Comparison of improvement \\((\\Delta \\mathrm{mAP}\\uparrow)\\) from long-term fusion between parallel paradigm (SOLOFusion [33]) and recurrent paradigm (VideoBEV [8], Ours). (b) Performance comparison on the validation set of nuScenes [2]. All methods use ResNet50 [9] pre-trained on ImageNet [6] as backbone with \\(256\\times 704\\) as input size. Frames per second (FPS) is measured on a single RTX 3090 with the PyTorch fp32." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.443, + 0.789, + 0.7 + ], + "angle": 0, + "content": "The temporal fusion mechanisms utilised in the BEV-based 3D detector can be broadly categorised into parallel fusion [11,33,39], and recurrent fusion [8,22]. Parallel fusion [11,33,39] caches and updates multiple historical features in learning the visual features for current frames. Therefore, the overhead of computing and storage grows as the length of the historical frame feature increases. Since the superior performance requires longer historical information, this fusion paradigm faces a dilemma between effectiveness and efficiency in deployment. In contrast, recurrent fusion [22] encodes all historical features into a single recurrent feature to cache and update, so it largely reduces inference latency and memory consumption. For instance, when employing 16 historical frames, the recurrent fusion method VideoBEV [8] exhibits \\(6 \\times\\) times faster in latency and uses \\(12 \\times\\) less memory compared to the state-of-the-art parallel fusion method SOLOFusion [33]. However, we observe that recurrent fusion fails to exploit the long-term information as well as parallel fusion. As depicted in Fig.1a, VideoBEV benefits less (\\(\\Delta\\) mAP↑) from utilizing historical frames than that of SOLOFusion (\\(\\Delta\\) mAP↑). This phenomenon becomes even more pronounced, when utilizing more than four historical frames." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.705, + 0.789, + 0.842 + ], + "angle": 0, + "content": "Why is the performance of recurrent fusion not as good as parallel fusion? We conducted a detailed analysis of the differences between the two paradigms and found two reasons: (1) Gradient propagation. During the training process, only the recent frames contribute to gradients of fusion module in the recurrent paradigm, while all historical frames contribute equally in the parallel paradigm. This prevents the temporal fusion module of recurrent paradigm from learning fusion pattern from long-term temporal information of all historical frames. (2) Semantic ambiguity. Due to the coarse granularity of the BEV grids, the recurrent paradigm generates errors when aligning historical features multiple" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.631, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "RecurrentBEV" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.787, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.785, + 0.207 + ], + "angle": 0, + "content": "times, resulting in semantic ambiguity for the same object. As a result, recurrent paradigm may fuse different objects on the singular BEV grid, confusing the recognition of the same object. For the above two reasons, the previous recurrent method cannot fully and correctly utilize the long-term temporal information." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.208, + 0.787, + 0.357 + ], + "angle": 0, + "content": "Motivated by the above analysis, in this paper, we propose RecurrentBEV, a novel temporal fusion framework for BEV-based 3D object detector to improve the long-term fusion ability of recurrent paradigm while still has efficient latency and memory consumption in inference. First, RecurrentBEV adopts an RNN-style back-propagation during training, enabling historical frames to contribute gradients to the fusion module directly. This improvement solves the first problem while still maintaining inference efficiency. Second, we design an inner grid transformation before fusion, which can learn fine-grained semantic changes from ego-motion matrix. This effectively reduces semantic differences in the inter-frame alignment process." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.359, + 0.788, + 0.479 + ], + "angle": 0, + "content": "Based on the RNN-style back-propagation and the inner grid transformation, the proposed RecurrentBEV effectively improves long-term fusion in the recurrent paradigm, as shown in Fig.1a. Extensive experiments on the nuScenes 3D detection dataset evaluate the effectiveness of the proposed RecurrentBEV. As shown in Fig.1b, it outperforms both existing BEV-based algorithms such as SOLOFusion, VideoBEV and query-based algorithms such as StreamPETR [35], SparseBEV [27], Sparse4Dv2 [26] in both accuracy and efficiency. To summarize, our contributions are:" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.489, + 0.785, + 0.534 + ], + "angle": 0, + "content": "- We provide an in-depth analysis for two temporal fusion paradigms and point out two factors that render the recurrent fusion less effective: issues with gradient propagation and semantic ambiguity." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.535, + 0.787, + 0.623 + ], + "angle": 0, + "content": "- We propose RecurrentBEV, a novel temporal fusion framework for BEV-based 3D object detector. It adapots RNN-style back-propagation and inner grid transformation, attaining state-of-the-art results (57.4% mAP and 65.1% NDS) on the nuScenes benchmark. It represents the first recurrent fusion method with explicit BEV features that surpasses parallel fusion performance." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.648, + 0.398, + 0.663 + ], + "angle": 0, + "content": "2 Related Works" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.679, + 0.531, + 0.695 + ], + "angle": 0, + "content": "2.1 Multi-view 3D Object Detection" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.704, + 0.785, + 0.78 + ], + "angle": 0, + "content": "Multi-view 3D object detection is a fundamental task for autonomous driving. Previous works mainly focus on how to transform the 2D features into a unified 3D space. This transformation can be divided to two categories, BEV-based methods [10, 13, 18, 22, 23, 41, 42] and query-based methods [25, 28, 36, 37], depending on whether an explicit 3D representation is employed." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.787, + 0.841 + ], + "angle": 0, + "content": "BEVDet [13] is one typical method of BEV-based multi-view 3D object detection. This method involves converting 2D features to BEV features based on the Lift-Splat-Shoot [34] paradigm. BEVDepth [21] is a further example, which notably refines the transformation process via explicit depth supervision." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.38, + 0.129 + ], + "angle": 0, + "content": "M. Chang et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.176 + ], + "angle": 0, + "content": "BEVFormer [22] further introduces a novel attention mechanism to gather 2D features for BEV grids." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.178, + 0.788, + 0.269 + ], + "angle": 0, + "content": "Query-based methods typically define several object queries in 3D space to gather features from multi-view image features. DETR3D [36] projects object queries to image coordinates to sample features then predicts 3D bounding boxes directly. DETR4D [31] and Sparse4D [25] enhance it by leveraging historical information. PETR [28] encodes the position information of 3D coordinates into image feature, facilitating the detection of 3D objects similar to DETR [3]." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.291, + 0.613, + 0.305 + ], + "angle": 0, + "content": "2.2 Temporal Fusion for BEV-based Detectors" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.315, + 0.789, + 0.543 + ], + "angle": 0, + "content": "Recent works find that incorporating temporal information can enhance the stability in perception results and improve overall performance. BEVFormer [22] first proposes the temporal self-attention mechanism to fuse the historical BEV features in a recurrent paradigm. BEVDet4D [11] extends BEVDet [13] to a temporal version by employing a concatenation operation to integrate previous BEV features, usually called parallel paradigm. SOLOFusion [33] extends these methods by increasing the number of previous frames to 16, resulting in significant benefits. BEVFormerv2 [39] switches to parallel paradigm same as SOLOFusion. However, the inference latency and memory consumption of the parallel paradigm increase linearly as the historical frame lengths. VideoBEV [8] applies recurrent fusion to the LSS-based 3D detector, resulting in significant reduction of inference latency and memory consumption. However, the improvement of long-term fusion in VideoBEV is limited in comparison to SOLOFusion. In order to combine the advantages of the two paradigms, we construct a novel temporal fusion framework based on efficient recurrent paradigm." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.565, + 0.331, + 0.58 + ], + "angle": 0, + "content": "3 Method" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.597, + 0.788, + 0.703 + ], + "angle": 0, + "content": "Converting multi-camera image features to bird's-eye-view (BEV) features can provide a unified surrounding environment representation for autonomous driving perception tasks. It involves transforming image data into a top-down representation of the environment, which often denoted as BEV features. The BEV detectors contain an image encoder, a 2D-3D view transformation, and a temporal fusion module. In this work, we present a new temporal fusion module, which can effectively aggregate spatio-temporal BEV features." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.725, + 0.688, + 0.74 + ], + "angle": 0, + "content": "3.1 Analysis of BEV-based Temporal Fusion Paradigms" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.841 + ], + "angle": 0, + "content": "There are two kinds of temporal fusion paradigms in BEV detection, which are the effective parallel fusion and efficient recurrent fusion. The parallel fusion caches and updates multiple historical BEV features in learning the visual features for current frames. The recurrent fusion encodes all historical features into a single recurrent feature to cache and update. We conduct a detailed analysis to identify the disparities among them and pinpoint two crucial factors that cause" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.631, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "RecurrentBEV" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.126 + ], + "angle": 0, + "content": "5" + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.142, + 0.49, + 0.354 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.345, + 0.357, + 0.365, + 0.37 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.494, + 0.143, + 0.778, + 0.358 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.621, + 0.358, + 0.642, + 0.37 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.38, + 0.788, + 0.492 + ], + "angle": 0, + "content": "Fig. 2: (a) Semantic ambiguity caused by grid sample operation. When ego vehicle turning right, BEV feature change from \\( B_{t-1} \\) to \\( B_t \\). In fusion, the grid sample operation used for aligning \\( B_{t-1} \\) and \\( B_t \\) only aligns the position but ignores the orientation as shown in the blue and red dash lines. (b) Visualization of the semantic ambiguity. The green boxes represent the BEV feature change from \\( B_{t-1} \\), which have been transformed to align with the \\( B_t \\) and decoded into bounding boxes. The blue boxes depict the actual bounding boxes in the \\( B_t \\). Notably, for static objects, there is a mismatch in the orientation of the green and blue boxes." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.522, + 0.785, + 0.553 + ], + "angle": 0, + "content": "the sub-optimal performance of recurrent fusion: issues with gradient propagation and semantic ambiguity." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.575, + 0.788, + 0.621 + ], + "angle": 0, + "content": "Gradient propagation Aligning historical frames to current frame is essential for BEV-based temporal fusion. Given the previous BEV feature \\( B_{j} \\), this process can be written as:" + }, + { + "type": "equation", + "bbox": [ + 0.437, + 0.63, + 0.786, + 0.65 + ], + "angle": 0, + "content": "\\[\n\\widetilde {B} _ {j, i} = \\zeta \\left(B _ {j}, P _ {j, i}\\right) \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.661, + 0.788, + 0.724 + ], + "angle": 0, + "content": "where \\(\\widetilde{B}_{j,i}\\) is the \\(j\\)-th frame BEV feature in the ego coordinate of \\(i\\)-th frame, \\(P_{j,i}\\) is the ego-motion matrix between the \\(j\\)-th frame and \\(i\\)-th frame and \\(\\zeta\\) refers to the ego-motion transformation function, specifically grid sample operation [15]. Based on Eq.(1), We can express the parallel fusion of the \\(i\\)-th frame as:" + }, + { + "type": "equation", + "bbox": [ + 0.38, + 0.735, + 0.785, + 0.754 + ], + "angle": 0, + "content": "\\[\n\\bar {B} _ {i} = \\varphi \\left(\\widetilde {B} _ {i - k, i}, \\dots , \\widetilde {B} _ {i - 1, i}, B _ {i}, W\\right) \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.764, + 0.788, + 0.824 + ], + "angle": 0, + "content": "where \\(\\overline{B}_i\\) is the fused BEV feature of \\(i\\)-th frame, \\(\\varphi\\) is the temporal fusion function (convolution [11,21,33,39] or deformable attention [22]) and \\(W\\) refers the parameters of it. Similarly, the recurrent paradigm for the \\(i\\)-th frame can be written as:" + }, + { + "type": "equation", + "bbox": [ + 0.389, + 0.824, + 0.785, + 0.842 + ], + "angle": 0, + "content": "\\[\n\\bar {B} _ {i} = \\varphi (\\zeta (\\bar {B} _ {i - 1}, P _ {i - 1, i}), B _ {i}, W) \\tag {3}\n\\]" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.38, + 0.129 + ], + "angle": 0, + "content": "M. Chang et al." + }, + { + "type": "image", + "bbox": [ + 0.241, + 0.144, + 0.761, + 0.314 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.327, + 0.788, + 0.397 + ], + "angle": 0, + "content": "Fig. 3: Overall architecture of the proposed RecurrentBEV. It shows a training process in the \\(t\\)-th frame, while arrows indicate forward propagation and backward propagation. The red arrows indicate the additional back-propagation introduced by our method. The details of the Transformation-aware Gate Recurrent Unit (TA-GRU) is illustrated in dotted box." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.425, + 0.789, + 0.456 + ], + "angle": 0, + "content": "Based on Eq.(3), we can calculate the gradient for the parameters in the temporal fusion function for parallel fusion." + }, + { + "type": "equation", + "bbox": [ + 0.362, + 0.462, + 0.787, + 0.495 + ], + "angle": 0, + "content": "\\[\n\\frac {\\partial L}{\\partial W} = \\frac {\\partial L}{\\partial \\bar {B} _ {i}} \\cdot \\varphi^ {\\prime} \\left(\\widetilde {B} _ {i - k, i}, \\dots , \\widetilde {B} _ {i - 1, i}, B _ {i}\\right) \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.502, + 0.788, + 0.532 + ], + "angle": 0, + "content": "In recurrent fusion, \\( B_{i-1} \\) is detached so that the gradient of \\( W \\) in recurrent fusion can be written as:" + }, + { + "type": "equation", + "bbox": [ + 0.372, + 0.537, + 0.787, + 0.571 + ], + "angle": 0, + "content": "\\[\n\\frac {\\partial L}{\\partial W} = \\frac {\\partial L}{\\partial \\bar {B} _ {i}} \\cdot \\varphi^ {\\prime} (\\zeta (\\bar {B} _ {i - 1}, P _ {i - 1, i}), B _ {i}) \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.578, + 0.788, + 0.655 + ], + "angle": 0, + "content": "Comparing Eq.(4) and Eq.(5), it is evident that only the current BEV feature \\( B_{i} \\) and one previous BEV feature \\( \\overline{B}_{i-1} \\) affect the gradient of \\( W \\) directly in recurrent paradigm while all previous BEV features within a window size \\( k \\) contribute to gradient of \\( W \\) in parallel paradigm. This is one reason that recurrent paradigm makes limited progress in long-term fusion." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.788, + 0.735 + ], + "angle": 0, + "content": "Semantic Ambiguity To optimize memory usage, the recurrent fusion approach only caches a single historical 3D BEV feature instead of storing a sequence of 2D features. Consequently, this approach encounters challenges with temporal alignment in the 3D space." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.788, + 0.841 + ], + "angle": 0, + "content": "In the current recurrent fusion method, we have observed semantic ambiguity issues arising during the temporal alignment process in the 3D space. As illustrated in Fig.2a, the coarse granularity of BEV grids results in inaccurate transformation of orientation information within each grid during the temporal alignment process represented by Eq.(1). Consequently, this process introduces incorrect object orientations in real-world detection scenarios, as evident in Fig.2b." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.631, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "RecurrentBEV" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.116, + 0.787, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.147, + 0.788, + 0.222 + ], + "angle": 0, + "content": "In the parallel fusion counterpart, temporal alignment in the 3D space is avoided by directly projecting a sequence of 2D view feature onto the current 3D BEV feature space [33, 39]. Assuming that the past perspective 2D view feature at time step \\(j\\) is denoted as \\(C_j\\), the formulation for projecting it onto the BEV feature at time step \\(i\\) is as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.379, + 0.233, + 0.786, + 0.251 + ], + "angle": 0, + "content": "\\[\n\\widetilde {B} _ {j, i} = \\operatorname {V i e w T r a n s} \\left(C _ {j}, I _ {j}, P _ {j, i} \\cdot E _ {j}\\right) \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.259, + 0.788, + 0.351 + ], + "angle": 0, + "content": "where \\( I_{j} \\) and \\( E_{j} \\) denote camera intrinsic and extrinsic respectively. This additional view transformation in 2D space effectively aligns objects between frames, subsequently enhancing the performance of parallel fusion. Nevertheless, this approach is not feasible for recurrent fusion due to the unavailability of 2D features during the recurrent fusion process. This discrepancy serves as another contributing factor to the performance gap between the two fusion methods." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.371, + 0.387, + 0.387 + ], + "angle": 0, + "content": "3.2 RecurrentBEV" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.396, + 0.789, + 0.502 + ], + "angle": 0, + "content": "Based on the two above-mentioned possible reasons, we design a novel temporal fusion framework called RecurrentBEV. Similar to existing BEV-based 3D object detectors [8,11,21,33], RecurrentBEV contains an image encoder, a 2D-3D view transformation, a temporal fusion module, a BEV encoder and a task head, as illustrated in Fig.3. The main difference between RecurrentBEV and previous recurrent fusion is the temporal fusion module which can learn from all historical frames directly and learn fine-grained ego-motion transform." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.522, + 0.789, + 0.645 + ], + "angle": 0, + "content": "RNN-style Back-propagation In prior recurrent fusion methods [22], the \\(\\overline{B}_{i-1}\\) in Eq.(3) is solely computed for forward propagation, but not for backward propagation. Thus, historical BEV features do not contribute to the gradient of temporal fusion module in Eq.(5), preventing optimizing the temporal fusion module with all the relevant historical BEV features. To address this issue, RecurrentBEV adopts RNN-style Back-propagation that all the relevant historical features back-propagate gradients to adjacent layers. In this context, the gradients of parameters in temporal fusion function is written as:" + }, + { + "type": "equation", + "bbox": [ + 0.352, + 0.654, + 0.787, + 0.697 + ], + "angle": 0, + "content": "\\[\n\\frac {\\partial L}{\\partial W} = \\sum_ {j = i - k} ^ {i} \\frac {\\partial L}{\\partial \\overline {{B}} _ {j}} \\cdot \\varphi^ {'} (\\zeta (\\overline {{B}} _ {j - 1}, P _ {j - 1, j}), B _ {j}) \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.705, + 0.787, + 0.75 + ], + "angle": 0, + "content": "where \\( k \\) is the number of total historical frames using in a iteration of training process. As a result, each historical BEV feature makes an equal contribution to the gradient of the temporal fusion module, akin to the parallel paradigm." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.751, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Computing the back-propagation along the temporal dimension for the entire network places significant demands on GPU memory. As shown in Table 1 (Setting 1), if all modules, including image encoders, undergo back-propagation along the temporal dimension, memory overflow issues arise even when the frame number reaches only 2. However, it is imperative for the training process to incorporate long-term frames to effectively learn temporal information (refer to" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.38, + 0.129 + ], + "angle": 0, + "content": "M. Chang et al." + }, + { + "type": "table", + "bbox": [ + 0.307, + 0.144, + 0.699, + 0.201 + ], + "angle": 0, + "content": "
Training frames0124816
Setting 115,06730,105OOMOOMOOMOOM
Setting 215,06717,66118,73519,97523,57930,941
" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.201, + 0.787, + 0.228 + ], + "angle": 0, + "content": "Table 1: The memory usage (MB) during training. OOM means out of memory on A100-40G with FP32 precision." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.269, + 0.788, + 0.39 + ], + "angle": 0, + "content": "Sec.4.6 for details). Therefore, we need to strike a balance between GPU memory usage (determined by the modules involved in back-propagation) and the length of the time sequence. In our design, as depicted by the red arrows in Figure 3, RecurrentBEV further computes all the gradients of the temporal fusion module (TA-GRU in the following) via back-propagation along the temporal dimension. Consequently, RecurrentBEV (Setting 2 in Table 1) can be trained with a batch size of 8 with FP32 precision on a single A100-40G GPU, effectively utilizing up to 16 consecutive frames." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.41, + 0.788, + 0.487 + ], + "angle": 0, + "content": "Transformation-aware Gated Recurrent Unit Because of the coarse granularity of the BEV grids, the semantic feature in a grid can not project correctly when model transforms the previous BEV features to the coordinate system of the current ego. We design an inner grid transformation to solve this problem. For simplicity, we use a linear layer to model this transformation:" + }, + { + "type": "equation", + "bbox": [ + 0.395, + 0.494, + 0.787, + 0.513 + ], + "angle": 0, + "content": "\\[\n\\widetilde {B} _ {j, i} = \\operatorname {L i n e a r} \\left(\\zeta \\left(B _ {j}, P _ {j, i}\\right), W _ {p}\\right) \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.52, + 0.788, + 0.564 + ], + "angle": 0, + "content": "where \\( W_{p} \\) is the parameter of Linear. For simplicity, we assume different channels in a projected feature share same motion embedding, so that the \\( W_{p} \\) can be decoupled as:" + }, + { + "type": "equation", + "bbox": [ + 0.445, + 0.564, + 0.786, + 0.582 + ], + "angle": 0, + "content": "\\[\nW _ {p} = E _ {j, i} \\odot \\widetilde {W} _ {p} \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.586, + 0.788, + 0.65 + ], + "angle": 0, + "content": "where \\( E_{j,i} \\in \\mathbb{R}^{C_{in} \\times 1} \\) is ego motion embedding from timestamp \\( j \\) to timestamp \\( i \\), \\( \\widetilde{W}_p \\in \\mathbb{R}^{C_{in} \\times C_{out}} \\) is a learnable weight and \\( \\odot \\) denotes a Hadamard Product. Motion embedding \\( E_{j,i} \\) is a high dimensional vector converted from the given ego motion matrix \\( P_{j,i} \\):" + }, + { + "type": "equation", + "bbox": [ + 0.442, + 0.65, + 0.786, + 0.667 + ], + "angle": 0, + "content": "\\[\nE _ {j, i} = \\sigma (\\psi (P _ {j, i})) \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.67, + 0.788, + 0.715 + ], + "angle": 0, + "content": "where \\(\\sigma\\) is a sigmoid activation and \\(\\psi\\) is a multi-layer perception (MLP). Combining Eq.(8), Eq.(9) and Eq.(10), the complete inner grid transformation can be written as:" + }, + { + "type": "equation", + "bbox": [ + 0.348, + 0.738, + 0.787, + 0.758 + ], + "angle": 0, + "content": "\\[\n\\widetilde {B} _ {j, i} = \\operatorname {L i n e a r} \\left(\\zeta \\left(\\bar {B} _ {j}, P _ {j, i}\\right), \\sigma \\left(\\psi \\left(P _ {j, i}\\right)\\right) \\odot \\widetilde {W} _ {p}\\right) \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.788, + 0.84 + ], + "angle": 0, + "content": "For the RNN modules, we adopted the typical Convolutional Gated Recurrent Unit (ConvGRU) [1]. By combining the aforementioned inner grid transformation, we introduce a novel temporal fusion module named the Transformation-Aware Gated Recurrent Unit (TA-GRU). The entire process of temporal fusion is formulated as follows:" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.631, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "RecurrentBEV" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "equation", + "bbox": [ + 0.368, + 0.173, + 0.786, + 0.192 + ], + "angle": 0, + "content": "\\[\nW _ {p} = \\sigma \\left(\\psi \\left(P _ {i - 1, i}\\right)\\right) \\odot \\widetilde {W} _ {p} \\tag {12}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.347, + 0.195, + 0.786, + 0.214 + ], + "angle": 0, + "content": "\\[\n\\widetilde {B} _ {i - 1, i} = \\operatorname {L i n e a r} \\left(\\zeta \\left(\\bar {B} _ {i - 1}, P _ {i - 1, i}\\right), W _ {p}\\right) \\tag {13}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.374, + 0.218, + 0.786, + 0.236 + ], + "angle": 0, + "content": "\\[\nZ _ {i} = \\sigma \\left(\\operatorname {C o n v} \\left(\\left[ \\widetilde {B} _ {i - 1, i}, B _ {i} \\right], W _ {z}\\right)\\right) \\tag {14}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.372, + 0.239, + 0.786, + 0.257 + ], + "angle": 0, + "content": "\\[\nR _ {i} = \\sigma \\left(\\operatorname {C o n v} \\left(\\left[ \\widetilde {B} _ {i - 1, i}, B _ {i} \\right], W _ {r}\\right)\\right) \\tag {15}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.372, + 0.261, + 0.786, + 0.28 + ], + "angle": 0, + "content": "\\[\nO _ {i} = \\tanh \\left(\\operatorname {C o n v} \\left(\\left[ R _ {i} \\odot \\widetilde {B} _ {i - 1, i}, B _ {i} \\right], W _ {o}\\right)\\right) \\tag {16}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.371, + 0.283, + 0.786, + 0.301 + ], + "angle": 0, + "content": "\\[\n\\bar {B} _ {i} = \\left(1 - Z _ {i}\\right) \\odot \\tilde {B} _ {i - 1, i} + Z _ {i} \\odot O _ {i} \\tag {17}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.312, + 0.788, + 0.372 + ], + "angle": 0, + "content": "where \\([\\cdot, \\cdot]\\) represents the concatenation of feature maps and \\(W_{z}, W_{z}, W_{o}\\) are 2D-convolution kernels. Because of inner grid transformation, TA-GRU can learn fine-grained semantic changes from ego-motion matrix, thus removing wrong orientation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.397, + 0.788, + 0.502 + ], + "angle": 0, + "content": "Inference Efficiency Comparing Eq.(2) and Eq.(3), parallel fusion requires \\( k \\) historical frames to be cached while recurrent fusion requires only one. Additionally, the transformation process described in Eq.(1) or Eq.(6) also needs to be computed \\( k \\) times in parallel fusion, whereas in recurrent fusion it only needs to be computed once. As a result, recurrent fusion, which has lower inference latency and smaller memory consumption, is more suitable for deployment in the industry." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.503, + 0.788, + 0.549 + ], + "angle": 0, + "content": "Similar to previous recurrent fusion methods [8, 22], RecurrentBEV adopts sequential inference which only needs to cache a single historical BEV feature \\(\\overline{B}\\), initialized with zero and updated at the end of the fusion." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.574, + 0.377, + 0.591 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.607, + 0.433, + 0.621 + ], + "angle": 0, + "content": "4.1 Dataset and Metrics" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.633, + 0.788, + 0.816 + ], + "angle": 0, + "content": "We evaluate our method on the challenging nuScenes dataset [2]. The dataset comprises 1000 driving scenes, each lasting about 20 seconds, collected in Boston and Singapore, which are divided into 700, 150 and 150 scenes respectively for training, validation, and testing. The vehicle is equipped with 6 cameras for a complete 360-degree field of view. The nuScenes dataset annotates 23 object classes with accurate 3D bounding boxes for each scene at \\(2\\mathrm{Hz}\\). For evaluation purposes, 10 classes are selected, namely: car, truck, bus, trailer, construction vehicle, pedestrian, motorcycle, bicycle, barrier, and traffic cone. As for evaluation, we use the following metrics, mean average precision (mAP), the nuScenes detection score (NDS), and five true positive metrics including average translation error (ATE), average scale error (ASE), average orientation error (AOE), average velocity error (AVE), average attribute error (AAE)." + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.825, + 0.719, + 0.841 + ], + "angle": 0, + "content": "4 details of evaluation metrics: https://www.nuscenes.org/object-detection" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.378, + 0.129 + ], + "angle": 0, + "content": "M. Chang et al." + }, + { + "type": "table", + "bbox": [ + 0.223, + 0.144, + 0.781, + 0.405 + ], + "angle": 0, + "content": "
MethodsBackboneImage SizeEDmAP↑NDS↑mATE↓mASE↓mAOE↓mAVE↓mAAE↓
BEVDet [13]ResNet50256 × 704X0.2980.3790.7250.2790.5890.8600.245
BEVDet4D [11]ResNet50256 × 704X0.3220.4570.7030.2780.4950.3540.206
BEVDepth [21]ResNet50256 × 704X0.3510.4750.6390.2670.4790.4280.198
BEVStereo [20]ResNet50256 × 704X0.3720.5000.5980.2700.4380.3670.190
BEVPoolv2 [12]ResNet50256 × 704X0.4060.5260.5720.2750.4630.2750.188
SOLOFusion [33]ResNet50256 × 704X0.4270.5340.5670.2740.5110.2520.181
VideoBEV [8]ResNet50256 × 704X0.4220.5350.5640.2760.4400.2860.198
PETRv2 [29]ResNet50256 × 704X0.3490.4560.7000.2750.5800.4370.187
Sparse4Dv2 [26]ResNet50256 × 704X0.4390.5390.5980.2700.4750.2820.179
SparseBEV [27]ResNet50256 × 704X0.4320.5450.6060.2740.3870.2510.186
StreamPETR [35]ResNet50256 × 704X0.4320.5400.5810.2720.4130.2950.195
RecurrentBEVResNet50256 × 704X0.4450.5490.5550.2720.4510.2560.204
BEVDepth [13]ResNet101512 × 1408X0.4120.5350.5650.2660.3580.3310.190
SOLOFusion [33]ResNet101512 × 1408X0.4830.5820.5030.2640.3810.2460.207
Sparse4Dv2 [26]ResNet101512 × 1408X0.4850.5800.5550.2720.3670.2560.182
RecurrentBEVResNet101512 × 1408X0.5090.5990.4870.2670.3580.2470.199
SparseBEV [27]ResNet101512 × 1408✓0.5010.5920.5620.2650.3210.2430.195
StreamPETR [35]ResNet101512 × 1408✓0.5040.5920.5690.2620.3150.2570.199
Sparse4Dv2 [26]ResNet101512 × 1408✓0.5050.5940.5480.2680.3480.2390.184
RecurrentBEVResNet101512 × 1408✓0.5280.6120.4810.2560.3550.2260.201
" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.406, + 0.785, + 0.446 + ], + "angle": 0, + "content": "Table 2: Comparison results on 3D detection on the nuScenes validation set. ED is external data which indicates to use pre-trained weights from the external dataset i.e. nuImages." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.503, + 0.456, + 0.518 + ], + "angle": 0, + "content": "4.2 Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.561, + 0.788, + 0.729 + ], + "angle": 0, + "content": "We conduct experiments based on the BEVDet4D-Depth [11, 21] with the same learning rate, optimizer, and data augmentation. Drawing upon previous methods [21, 33, 35], we utilize ResNet50 [9], ResNet101, ConNeXt-B [30] and V2-99 [17] as backbones. On the nuScenes validation set, we report results of ResNet50 and ResNet101 models with pre-trained weights ImageNet [6] and nuImages [35]. On the nuScenes test set, we provide the performance of V2-99 model initialized from DD3D [32] and ConvNeXt-B model benefited from COCO [24] pre-training. The models in the ablation study are trained for 24 epochs without CBGS [43]. When compared with others, we follow the training setting of StreamPETR [35], extending the training epochs to 60 and using 2D auxiliary supervision for a fair comparison." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.734, + 0.788, + 0.841 + ], + "angle": 0, + "content": "To achieve real-time inference speed, we optimize the BEVDet4D-Depth baseline by making several key modifications. Firstly, we reduce the number of depth bins. Secondly, we eliminate the ASPP [5] module from the DepthNet to minimize the latency associated with view transformation. Additionally, we merge 6 detection tasks performed by CenterHead [40] into a single task, significantly reducing post-processing time. These optimizations combined enable us to achieve a more efficient inference speed." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.631, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "RecurrentBEV" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "table", + "bbox": [ + 0.227, + 0.144, + 0.777, + 0.445 + ], + "angle": 0, + "content": "
MethodsBackbonemAP↑NDS↑mATE↓mASE↓mAOE↓mAVE↓mAAE↓Conference
DETR3D* [36]V2-990.4120.4790.6410.2550.3940.8450.133CoRL2021
PETR [28]V2-990.4410.5040.5930.2490.3830.8080.132ECCV2022
UVTR [19]V2-990.4720.5510.5770.2530.3910.5080.123NIPS2022
BEVFormer [22]V2-990.4810.5690.5820.2560.3750.3780.126ECCV2022
PETRv2 [29]V2-990.4900.5820.5610.2430.3610.3430.120ICCV2023
PolarFormer [16]V2-990.4930.5720.5560.2560.3640.4390.127AAAI2023
BEVDepth [21]V2-990.5030.6000.4450.2450.3780.3200.126AAAI2023
VEDet [4]V2-990.5050.5850.5450.2440.3460.4210.123CVPR2023
MV2D [37]V2-990.5110.5960.5250.2430.3570.3570.120ICCV2023
CAPE [38]V2-990.5250.6100.5030.2420.3610.3060.114CVPR2023
BEVStereo [20]V2-990.5250.6100.4310.2460.3580.3570.138AAAI2023
SA-BEV [41]V2-990.5330.6240.4300.2410.3380.2820.139ICCV2023
FB-BEV [23]V2-990.5370.6240.4390.2500.3580.2700.128ICCV2023
StreamPETR [35]V2-990.5500.6360.4790.2390.3170.2410.119ICCV2023
SparseBEV [27]V2-990.5560.6360.4850.2440.3320.2460.117ICCV2023
Sparse4Dv2 [26]V2-990.5570.6380.4620.2380.3280.2640.115-
RecurrentBEVV2-990.5730.6510.4100.2470.3230.2490.123-
BEVDepth [21]ConvNeXt-B0.5200.6090.4450.2430.3520.3470.127AAAI2023
AeDet* [7]ConvNeXt-B0.5310.6200.4390.2470.3440.2920.130CVPR2023
SOLOFusion [33]ConvNeXt-B0.5400.6190.4530.2570.3760.2760.148ICLR2023
VCD-A [14]ConvNeXt-B0.5480.6310.4360.2440.3430.2900.120NIPS2023
StreamPETR [35]ConvNeXt-B0.5500.6310.4930.2410.3430.2430.123ICCV2023
VideoBEV [8]ConvNeXt-B0.5540.6290.4570.2490.3810.2660.132-
RecurrentBEVConvNeXt-B0.5740.6510.4110.2480.3330.2570.116-
" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.446, + 0.785, + 0.486 + ], + "angle": 0, + "content": "Table 3: Comparison results on 3D detection on the nuScenes test set. * indicates using test-time augmentation in test process. No future frames are used in any of the methods." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.533, + 0.373, + 0.547 + ], + "angle": 0, + "content": "4.3 Main Results" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.565, + 0.785, + 0.595 + ], + "angle": 0, + "content": "We compare the proposed RecurrentBEV with previous state-of-the-art multiview 3D detectors on the nuScenes validation and test set." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.597, + 0.787, + 0.734 + ], + "angle": 0, + "content": "On the nuScenes validation set, as shown in Fig.1b and Tab.2, the real-time version of RecurrentBEV (25.6 FPS), using ResNet50 as backbone and \\(256 \\times 704\\) as the input size, outperforms the previous state-of-the-art method StreamPETR by \\(1.3\\%\\) mAP and \\(0.9\\%\\) NDS on the nuScenes validation set. To scale up our method, we adopt ResNet101 backbone with nuImages pre-training and extend input size to \\(512 \\times 1408\\), outperforming the same setting StreamPETR by \\(2.4\\%\\) mAP and \\(2.0\\%\\) NDS. It's worth noting that RecurrentBEV makes remarkable progress on the mAP and mATE metrics, indicating the improvement of detection of occlusion objects. We'll delve into the reasons behind this in Sec.4.6." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.787, + 0.84 + ], + "angle": 0, + "content": "On the nuScenes test set in Tab.3, we compare with all the recent methods in multi-view 3D object detection. Following the settings in StreamPETR, we adopt V2-99 and ConvNeXt-B as backbone and \\(640 \\times 1600\\) as the input size. Using V2-99 as backbone, RecurrentBEV can surpass the previous state-of-the-art method Sparse4Dv2 by \\(1.6\\%\\) mAP and \\(1.3\\%\\) NDS. When the backbone is scaled up to ConvNeXt-B, RecurrentBEV demonstrates comparable performance, surpassing other methods." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.381, + 0.129 + ], + "angle": 0, + "content": "M. Chang et al." + }, + { + "type": "table", + "bbox": [ + 0.29, + 0.144, + 0.713, + 0.215 + ], + "angle": 0, + "content": "
methodmAP↑NDS↑memory usage↓latency↑
parallel fusion [33]0.4270.534945MB96ms
recurrent fusion baseline0.3960.515516MB34ms
RecurrentBEV0.4450.549549MB39ms
" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.215, + 0.787, + 0.258 + ], + "angle": 0, + "content": "Table 4: Comparison of time and memory in inference. All data are measured on a single RTX 3090 with PyTorch fp32. Implementation details are the same as Tab.2 ResNet50 Backbone. For more latency analysis, please refer to supplementary." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.3, + 0.522, + 0.316 + ], + "angle": 0, + "content": "4.4 Time and Memory Comparison" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.327, + 0.788, + 0.433 + ], + "angle": 0, + "content": "To demonstrate the time and memory efficiency in inference, we compare the proposed RecurrentBEV with parallel fusion methods and previous recurrent fusion baselines. As shown in Tab.4 RecurrentBEV is 2.5 times faster than the normal parallel fusion method SOLOFusion [33], and the memory usage is reduced by \\(42\\%\\). When compared to the normal recurrent baseline, RecurrentBEV requires only an additional 33MB of memory while significantly improving the mean Average Precision (mAP) from 0.396 to 0.445." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.457, + 0.39, + 0.472 + ], + "angle": 0, + "content": "4.5 Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.484, + 0.787, + 0.528 + ], + "angle": 0, + "content": "In the ablation study, we delve into three aspects: the components of RecurrentBEV, the receptive field of TA-GRU, and the length of frame sequence in training. We conduct the ablation experiments with ResNet-50 backbone." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.529, + 0.788, + 0.634 + ], + "angle": 0, + "content": "Ablation of Each Component. We conduct a component-wise ablation in Tab.5. Exp.1 starts with a ConvGRU as temporal fusion module, achieving similar performance to VideoBEV [8]. In Exp.2 and Exp.3, we enable the backward propagation of historical BEV features in training process. In Exp.3 and Exp.4, we replace the temporal fusion module with TA-GRU. The overall components (BP and TA-GRU) brings significant performance gains \\((+1.7\\% \\mathrm{mAP}, +1.6\\%\\) NDS,-4.7% mAOE) over baseline." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.635, + 0.788, + 0.741 + ], + "angle": 0, + "content": "Specifically, the TA-GRU demonstrates its ability to enhance the mAP and mAOE metrics, regardless of whether backward propagation of historical BEV features is enabled. This proves the TA-GRU's consistent ability to alleviate semantic ambiguity for identical objects. Notably, the mean average orientation error (mAOE) has a notable increase, indicating that the inner grid transformation, facilitated by the alignment of historical BEV features, effectively mitigates semantic ambiguity by refining the orientation of objects." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.742, + 0.788, + 0.803 + ], + "angle": 0, + "content": "As demonstrated in Exp.4, when the historical BEV feature is detached, the performance suffers, highlighting the importance of RNN-style back-propagation. However, relying solely on back-propagation of historical BEV features during training results in limited improvement. This shows the necessity of employing" + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.81, + 0.787, + 0.842 + ], + "angle": 0, + "content": "similar performance to VideoBEV [8] without stereo matching (VideoBEV-D) with 60 training epochs." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.631, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "RecurrentBEV" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "table", + "bbox": [ + 0.356, + 0.144, + 0.648, + 0.228 + ], + "angle": 0, + "content": "
Exp|BP TG|mAP↑|NDS↑|mAOE↓|mAVE↓
10.3710.4760.6460.317
2✓0.3750.4790.6340.319
3✓✓0.3880.4920.5960.281
4✓0.3780.4840.6020.316
" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.228, + 0.784, + 0.256 + ], + "angle": 0, + "content": "Table 5: Ablation of each component. BP means the historical BEV feature backward propagate. TG means TA-GRU." + }, + { + "type": "table", + "bbox": [ + 0.236, + 0.274, + 0.479, + 0.359 + ], + "angle": 0, + "content": "
Kernel sizemAP↑NDS↑mATE↓mAVE↓
1 × 10.3880.4920.6470.281
3 × 30.3970.5000.6320.284
5 × 50.4030.5060.6150.276
7 × 70.3980.5010.6300.273
9 × 90.3940.4960.6330.280
" + }, + { + "type": "table_caption", + "bbox": [ + 0.236, + 0.359, + 0.48, + 0.372 + ], + "angle": 0, + "content": "Table 6: Ablation of receptive field." + }, + { + "type": "table", + "bbox": [ + 0.507, + 0.264, + 0.776, + 0.359 + ], + "angle": 0, + "content": "
Training framesmAP↑NDS↑mATE↓mAVE↓
00.3180.3800.7420.835
10.3460.4570.6960.389
20.3720.4790.6650.310
40.3900.4930.6460.283
80.4030.5060.6150.276
160.4020.490.6280.280
" + }, + { + "type": "table_caption", + "bbox": [ + 0.511, + 0.359, + 0.768, + 0.372 + ], + "angle": 0, + "content": "Table 7: Ablation of sequence length." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.388, + 0.785, + 0.432 + ], + "angle": 0, + "content": "the correct temporal fusion module for effective back-propagation, as the combination of these two components can achieve a synergistic effect that surpasses the sum of their individual contributions." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.433, + 0.785, + 0.538 + ], + "angle": 0, + "content": "Ablation of Receptive Field. Receptive field is critical for temporal fusion of moving objects. We explore the kernel size of convolutions in TA-GRU in Tab.6. As we can see, an excessively large or small receptive field can have a detrimental effect on performance. Our suspect is that a fusion module with an overlarge receptive field gathers noisy information, while a small receptive field leads to mismatches for moving objects. Consequently, we use a \\(5 \\times 5\\) kernel size in TA-GRU, unless otherwise specified." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.539, + 0.787, + 0.644 + ], + "angle": 0, + "content": "Ablation of Training Sequence Length. We use the BEVDet-Depth with no temporal fusion as baseline and apply our method with different training sequence length in Tab.7. The fusion of a single frame greatly enhances velocity prediction (mAVE). The fusion of four to eight frames mainly enhances localisation metrics mAP and mATE, with an increase of \\(8.5\\%\\) and \\(12.7\\%\\) respectively. Further increasing the number of frames to 16 offers gains limited performance improvement. Therefore, we train our models on 8 frames." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.666, + 0.573, + 0.681 + ], + "angle": 0, + "content": "4.6 Effect of Recurrent Long-term Fusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Detection of Occlusion Objects. Intuitively, long-term fusion has the potential to aid in the detection of occluded objects. Therefore, we conducted an analysis of object occlusions on the nuScenes validation set. Firstly, we define a 'distinct object' as one where the visibility of the entire object exceeds \\(60\\%\\). Subsequently, we determined that if an object is classified as a 'distinct object' in any frame of the video, it is considered visible throughout the video. Finally, we divided the number of visible objects by the total number of objects present to get the overall visibility of the video. As shown in Fig.4a, the visibility of the video has the same trend with mAP, proving the improvement of mAP in long-term fusion originates from the detection of occlusion objects." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.381, + 0.13 + ], + "angle": 0, + "content": "M. Chang et al." + }, + { + "type": "image", + "bbox": [ + 0.219, + 0.147, + 0.501, + 0.295 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.346, + 0.299, + 0.365, + 0.31 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.527, + 0.147, + 0.776, + 0.293 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.642, + 0.295, + 0.663, + 0.307 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.308, + 0.788, + 0.392 + ], + "angle": 0, + "content": "Fig. 4: (a) Relation between visibility and mAP. Visibility refers to the proportion of distinct objects in a video, expressed as a percentage of the total number of objects present. The Pearson correlation coefficient between visibility and mAP is 0.977.(b) The mATE metrics with objects at different location on the nuScenes validation set. The blue one is a result from single frame baseline and the orange one is RecurrentBEV. The objects between 25 meters and 30 meters improve the mATE metric by \\(24\\%\\) mostly." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.424, + 0.788, + 0.56 + ], + "angle": 0, + "content": "Localization of Distant Objects. The localization of objects is a major challenge for camera-based 3D detector. Due to the uncertainty of depth estimation. We speculate long-term fusion can more accurately localize objects that move from near to far as the ego vehicle moves. Therefore, we calculate the mATE metrics of objects at different distances before and after long-term fusion, as in Fig.4b. Notably, objects between 25 and 30 meters exhibit the most significant improvement in mATE through long-term fusion, whereas nearby objects show marginal gains. This finding demonstrates the efficacy of long-term fusion in accurately localizing distant objects." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.589, + 0.533, + 0.606 + ], + "angle": 0, + "content": "5 Conclusion and Future Work" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.628, + 0.788, + 0.779 + ], + "angle": 0, + "content": "In this paper, we investigate the differences between recurrent fusion and parallel fusion of BEV-based 3D object detectors, and propose RecurrentBEV, a simple and effective recurrent temporal fusion method. By incorporating RNN-style back-propagation and a newly designed inner grid transformation, RecurrentBEV enhances long-term fusion capabilities while maintaining latency and memory advantages. Rigorous experiments conducted on the nuScenes dataset demonstrate that RecurrentBEV outperforms all prior methods, while maintaining competitive inference speeds, thereby facilitating deployment. To our knowledge, it represents the first recurrent fusion method with explicit BEV features that surpasses parallel fusion performance." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.841 + ], + "angle": 0, + "content": "In the future, we will generalize recurrent fusion idea to more BEV architectures such as different BEV view transform backbones and query-based methods. We will also try to extend RecurrentBEV to other 3D perceptual tasks and datasets such as object tracking and trajectory prediction." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.631, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "RecurrentBEV" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.403, + 0.163 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.176, + 0.788, + 0.283 + ], + "angle": 0, + "content": "This work is partially supported by the National Key R&D Program of China (under Grant 2023YFB4502200), the NSF of China (under Grants U22A2028, 61925208, 62102399, 62222214, 62341411, 62102398, U20A20227, 62372436, 623-02478, 62302482, 62302483, 62302480), Strategic Priority Research Program of the Chinese Academy of Sciences, (Grant No.XDB0660200, XDB0660201, XDB0660202), CAS Project for Young Scientists in Basic Research(YSBR-029), Youth Innovation Promotion Association CAS and Xplore Prize." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.303, + 0.323, + 0.319 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.333, + 0.785, + 0.36 + ], + "angle": 0, + "content": "1. Ballas, N., Yao, L., Pal, C., Courville, A.: Delving deeper into convolutional networks for learning video representations. In: ICLR (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.361, + 0.785, + 0.401 + ], + "angle": 0, + "content": "2. Caesar, H., Bankiti, V., Lang, A.H., Vora, S., Liong, V.E., Xu, Q., Krishnan, A., Pan, Y., Baldan, G., Beijbom, O.: nuscenes: A multimodal dataset for autonomous driving. In: CVPR (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.403, + 0.785, + 0.429 + ], + "angle": 0, + "content": "3. Carion, N., Massa, F., Synnaeve, G., Usunier, N., Kirillov, A., Zagoruyko, S.: End-to-end object detection with transformers. In: ECCV (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.43, + 0.785, + 0.456 + ], + "angle": 0, + "content": "4. Chen, D., Li, J., Guizilini, V., Ambrus, R.A., Gaidon, A.: Viewpoint equivariance for multi-view 3d object detection. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.457, + 0.785, + 0.498 + ], + "angle": 0, + "content": "5. Chen, L.C., Papandreou, G., Kokkinos, I., Murphy, K., Yuille, A.L.: Deeplab: Semantic image segmentation with deep convolutional nets, atrous convolution, and fully connected crfs. IEEE TPAMI (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.499, + 0.785, + 0.525 + ], + "angle": 0, + "content": "6. Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: Imagenet: A large-scale hierarchical image database. In: CVPR (2009)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.526, + 0.785, + 0.552 + ], + "angle": 0, + "content": "7. Feng, C., Jie, Z., Zhong, Y., Chu, X., Ma, L.: Aedet: Azimuth-invariant multi-view 3d object detection. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.553, + 0.785, + 0.594 + ], + "angle": 0, + "content": "8. Han, C., Sun, J., Ge, Z., Yang, J., Dong, R., Zhou, H., Mao, W., Peng, Y., Zhang, X.: Exploring recurrent long-term temporal fusion for multi-view 3d perception. arXiv preprint arXiv:2303.05970 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.595, + 0.785, + 0.621 + ], + "angle": 0, + "content": "9. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: CVPR (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.622, + 0.785, + 0.662 + ], + "angle": 0, + "content": "0. Huang, B., Li, Y., Xie, E., Liang, F., Wang, L., Shen, M., Liu, F., Wang, T., Luo, P., Shao, J.: Fast-bev: Towards real-time on-vehicle bird's-eye view perception. arXiv preprint arXiv:2301.07870 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.663, + 0.785, + 0.689 + ], + "angle": 0, + "content": "1. Huang, J., Huang, G.: Bevdet4d: Exploit temporal cues in multi-camera 3d object detection. arXiv preprint arXiv:2203.17054 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.69, + 0.785, + 0.717 + ], + "angle": 0, + "content": "2. Huang, J., Huang, G.: Bevpoolv2: A cutting-edge implementation of bevdet toward deployment. arXiv preprint arXiv:2211.17111 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.718, + 0.785, + 0.758 + ], + "angle": 0, + "content": "3. Huang, J., Huang, G., Zhu, Z., Ye, Y., Du, D.: Bevdet: High-performance multicamera 3d object detection in bird-eye-view. arXiv preprint arXiv:2112.11790 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.759, + 0.785, + 0.785 + ], + "angle": 0, + "content": "4. Huang, L., Li, Z., Sima, C., Wang, W., Wang, J., Qiao, Y., Li, H.: Leveraging vision-centric multi-modal expertise for 3d object detection. In: NeurIPS (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.786, + 0.785, + 0.812 + ], + "angle": 0, + "content": "5. Jaderberg, M., Simonyan, K., Zisserman, A., et al.: Spatial transformer networks. In: NeurIPS (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.813, + 0.785, + 0.84 + ], + "angle": 0, + "content": "6. Jiang, Y., Zhang, L., Miao, Z., Zhu, X., Gao, J., Hu, W., Jiang, Y.G.: Polarformer: Multi-camera 3d object detection with polar transformer. In: AAAI (2023)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.333, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.378, + 0.129 + ], + "angle": 0, + "content": "M. Chang et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.148, + 0.785, + 0.175 + ], + "angle": 0, + "content": "17. Lee, Y., Hwang, J.w., Lee, S., Bae, Y., Park, J.: An energy andgpu-computation efficient backbone network for real-time object detection. In: CVPRW (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.176, + 0.785, + 0.202 + ], + "angle": 0, + "content": "18. Li, H., Zhang, H., Zeng, Z., Liu, S., Li, F., Ren, T., Zhang, L.: Dfa3d: 3d deformable attention for 2d-to-3d feature lifting. In: ICCV (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.203, + 0.785, + 0.229 + ], + "angle": 0, + "content": "19. Li, Y., Chen, Y., Qi, X., Li, Z., Sun, J., Jia, J.: Unifying voxel-based representation with transformer for 3d object detection. In: NeurIPS (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.23, + 0.785, + 0.256 + ], + "angle": 0, + "content": "20. Li, Y., Bao, H., Ge, Z., Yang, J., Sun, J., Li, Z.: Bevstereo: enhancing depth estimation in multi-view 3d object detection with temporal stereo. In: AAAI (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.257, + 0.785, + 0.284 + ], + "angle": 0, + "content": "21. Li, Y., Ge, Z., Yu, G., Yang, J., Wang, Z., Shi, Y., Sun, J., Li, Z.: Bevdepth: Acquisition of reliable depth for multi-view 3d object detection. In: AAAI (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.285, + 0.785, + 0.324 + ], + "angle": 0, + "content": "22. Li, Z., Wang, W., Li, H., Xie, E., Sima, C., Lu, T., Qiao, Y., Dai, J.: Bevformer: Learning bird's-eye-view representation from multi-camera images via spatiotemporal transformers. In: ECCV (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.325, + 0.785, + 0.352 + ], + "angle": 0, + "content": "23. Li, Z., Yu, Z., Wang, W., Anandkumar, A., Lu, T., Alvarez, J.M.: Fb-bev: Bev representation from forward-backward view transformations. In: ICCV (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.353, + 0.785, + 0.379 + ], + "angle": 0, + "content": "24. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dóllar, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: ECCV (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.379, + 0.785, + 0.405 + ], + "angle": 0, + "content": "25. Lin, X., Lin, T., Pei, Z., Huang, L., Su, Z.: Sparse4d: Multi-view 3d object detection with sparse spatial-temporal fusion. arXiv preprint arXiv:2211.10581 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.407, + 0.785, + 0.433 + ], + "angle": 0, + "content": "26. Lin, X., Lin, T., Pei, Z., Huang, L., Su, Z.: Sparse4d v2: Recurrent temporal fusion with sparse model. arXiv preprint arXiv:2305.14018 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.434, + 0.785, + 0.46 + ], + "angle": 0, + "content": "27. Liu, H., Teng, Y., Lu, T., Wang, H., Wang, L.: Sparsebev: High-performance sparse 3d object detection from multi-camera videos. In: ICCV (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.461, + 0.785, + 0.486 + ], + "angle": 0, + "content": "28. Liu, Y., Wang, T., Zhang, X., Sun, J.: Petr: Position embedding transformation for multi-view 3d object detection. In: ECCV (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.487, + 0.785, + 0.514 + ], + "angle": 0, + "content": "29. Liu, Y., Yan, J., Jia, F., Li, S., Gao, A., Wang, T., Zhang, X.: Petrov2: A unified framework for 3d perception from multi-camera images. In: ICCV (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.515, + 0.785, + 0.541 + ], + "angle": 0, + "content": "30. Liu, Z., Mao, H., Wu, C.Y., Feichtenhofer, C., Darrell, T., Xie, S.: A convnet for the 2020s. In: CVPR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.542, + 0.785, + 0.568 + ], + "angle": 0, + "content": "31. Luo, Z., Zhou, C., Zhang, G., Lu, S.: Detr4d: Direct multi-view 3d object detection with sparse attention. arXiv preprint arXiv:2212.07849 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.569, + 0.785, + 0.595 + ], + "angle": 0, + "content": "32. Park, D., Ambrus, R., Guizilini, V., Li, J., Gaidon, A.: Is pseudo-lidar needed for monocular 3d object detection? In: ICCV (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.596, + 0.785, + 0.636 + ], + "angle": 0, + "content": "33. Park, J., Xu, C., Yang, S., Keutzer, K., Kitani, K., Tomizuka, M., Zhan, W.: Time will tell: New outlooks and a baseline for temporal multi-view 3d object detection. In: ICLR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.637, + 0.785, + 0.663 + ], + "angle": 0, + "content": "34. Philion, J., Fidler, S.: Lift, splat, shoot: Encoding images from arbitrary camera rigs by implicitly unprojecting to 3d. In: ECCV (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.664, + 0.785, + 0.69 + ], + "angle": 0, + "content": "35. Wang, S., Liu, Y., Wang, T., Li, Y., Zhang, X.: Exploring object-centric temporal modeling for efficient multi-view 3d object detection. In: ICCV (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.691, + 0.785, + 0.731 + ], + "angle": 0, + "content": "36. Wang, Y., Guizilini, V.C., Zhang, T., Wang, Y., Zhao, H., Solomon, J.: Detr3d: 3d object detection from multi-view images via 3d-to-2d queries. In: Conference on Robot Learning (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.732, + 0.785, + 0.758 + ], + "angle": 0, + "content": "37. Wang, Z., Huang, Z., Fu, J., Wang, N., Liu, S.: Object as query: Lifting any 2d object detector to 3d detection. In: ICCV (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.759, + 0.785, + 0.799 + ], + "angle": 0, + "content": "38. Xiong, K., Gong, S., Ye, X., Tan, X., Wan, J., Ding, E., Wang, J., Bai, X.: Cape: Camera view position embedding for multi-view 3d object detection. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.8, + 0.785, + 0.84 + ], + "angle": 0, + "content": "39. Yang, C., Chen, Y., Tian, H., Tao, C., Zhu, X., Zhang, Z., Huang, G., Li, H., Qiao, Y., Lu, L., et al.: Bevformer v2: Adapting modern image backbones to bird's-eye-view recognition via perspective supervision. In: CVPR (2023)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.148, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.631, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "RecurrentBEV" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.175 + ], + "angle": 0, + "content": "40. Yin, T., Zhou, X., Krahenbuhl, P.: Center-based 3d object detection and tracking. In: CVPR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.175, + 0.785, + 0.203 + ], + "angle": 0, + "content": "41. Zhang, J., Zhang, Y., Liu, Q., Wang, Y.: Sa-bev: Generating semantic-aware bird's-eye-view feature for multi-view 3d object detection. In: ICCV (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.203, + 0.785, + 0.231 + ], + "angle": 0, + "content": "42. Zhou, H., Ge, Z., Li, Z., Zhang, X.: Matrixvt: Efficient multi-camera to bev transformation for 3d perception. In: ICCV (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.231, + 0.785, + 0.259 + ], + "angle": 0, + "content": "43. Zhu, B., Jiang, Z., Zhou, X., Li, Z., Yu, G.: Class-balanced grouping and sampling for point cloud 3d object detection. arXiv preprint arXiv:1908.09492 (2019)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.259 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/RecurrentBEV_ A Long-term Temporal Fusion Framework for Multi-view 3D Detection/c66016d3-668d-4d64-aa39-f278cd973b7f_origin.pdf b/2024/RecurrentBEV_ A Long-term Temporal Fusion Framework for Multi-view 3D Detection/c66016d3-668d-4d64-aa39-f278cd973b7f_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..b520d68139fc06c1bbc389adb32ddbf75e2c1e75 --- /dev/null +++ b/2024/RecurrentBEV_ A Long-term Temporal Fusion Framework for Multi-view 3D Detection/c66016d3-668d-4d64-aa39-f278cd973b7f_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:09f839db6347ee80240f078ac7c18af319484551b2707a65961aba318738e803 +size 856995 diff --git a/2024/RecurrentBEV_ A Long-term Temporal Fusion Framework for Multi-view 3D Detection/full.md b/2024/RecurrentBEV_ A Long-term Temporal Fusion Framework for Multi-view 3D Detection/full.md new file mode 100644 index 0000000000000000000000000000000000000000..d1f5b67090e8d0495fcea9d5eefe5d5ee094a8f3 --- /dev/null +++ b/2024/RecurrentBEV_ A Long-term Temporal Fusion Framework for Multi-view 3D Detection/full.md @@ -0,0 +1,329 @@ +# RecurrentBEV: A Long-term Temporal Fusion Framework for Multi-view 3D Detection + +Ming Chang $^{1}$ , Xishan Zhang $^{1,2\star}$ , Rui Zhang $^{2}$ , Zhipeng Zhao $^{1}$ , Guanhua He $^{1,2,3}$ , and Shaoli Liu $^{1}$ + +1 Cambricon Technologies {changming,zhaozhipeng,liushaoli}@cambricon.com + +2 SKL of Processors, Institute of Computing Technology, CAS, Beijing, China {zhangxishan,zhangrui}@ict.ac.cn, heguanhua22@mails.ucas.ac.cn + +3 University of Chinese Academy of Sciences, UCAS, Beijing, China + +Abstract. Long-term temporal fusion is frequently employed in camera-based Bird's-Eye-View (BEV) 3D object detection to improve detection of occluded objects. Existing methods can be divided into two categories, parallel fusion and recurrent fusion. Recurrent fusion reduces inference latency and memory consumption but fails to exploit the long-term information as well as parallel fusion. In this paper, we first find two shortcomings of recurrent fusion paradigm: (1) Gradients of previous BEV features cannot directly contribute to the fusion module. (2) Semantic ambiguity are caused by coarse granularity of the BEV grids during aligning BEV features. Then based on the above analysis, we propose RecurrentBEV, a novel recurrent temporal fusion method for BEV based 3D object detector. By adopting RNN-style back-propagation and new-designed inner grid transformation, RecurrentBEV improves the long-term fusion ability while still enjoying efficient inference latency and memory consumption during inference. Extensive experiments on the nuScenes benchmark demonstrate its effectiveness, achieving a new state-of-the-art performance of $57.4\%$ mAP and $65.1\%$ NDS on the test set. The real-time version (25.6 FPS) achieves $44.5\%$ mAP and $54.9\%$ NDS without external dataset, outperforming the previous best method StreamPETR by $1.3\%$ mAP and $0.9\%$ NDS. The code is available at https://github.com/lucifer443/RecurrentBEV. + +Keywords: Autonomous Driving $\cdot$ Bird's-Eye-View $\cdot$ 3D Object Detection $\cdot$ Temporal Fusion + +# 1 Introduction + +Camera-only 3D detection presents a formidable challenge for autonomous driving and has garnered significant attention in recent years owing to its low cost of deployment. The Bird's-Eye-View (BEV) representation-based approach has recently demonstrated notable improvement by leveraging temporal information [8,11,21,22,33]. + +![](images/a0c2c30d14ddc03080547d6318c6c605f57895cc2abc4416a369c72668279d1c.jpg) +(a) + +![](images/1e566910a3ade95c9c7492688b428a3a9751acd0ed968f9126a98c27112afe4a.jpg) +(b) +Fig. 1: (a) Comparison of improvement $(\Delta \mathrm{mAP}\uparrow)$ from long-term fusion between parallel paradigm (SOLOFusion [33]) and recurrent paradigm (VideoBEV [8], Ours). (b) Performance comparison on the validation set of nuScenes [2]. All methods use ResNet50 [9] pre-trained on ImageNet [6] as backbone with $256\times 704$ as input size. Frames per second (FPS) is measured on a single RTX 3090 with the PyTorch fp32. + +The temporal fusion mechanisms utilised in the BEV-based 3D detector can be broadly categorised into parallel fusion [11,33,39], and recurrent fusion [8,22]. Parallel fusion [11,33,39] caches and updates multiple historical features in learning the visual features for current frames. Therefore, the overhead of computing and storage grows as the length of the historical frame feature increases. Since the superior performance requires longer historical information, this fusion paradigm faces a dilemma between effectiveness and efficiency in deployment. In contrast, recurrent fusion [22] encodes all historical features into a single recurrent feature to cache and update, so it largely reduces inference latency and memory consumption. For instance, when employing 16 historical frames, the recurrent fusion method VideoBEV [8] exhibits $6 \times$ times faster in latency and uses $12 \times$ less memory compared to the state-of-the-art parallel fusion method SOLOFusion [33]. However, we observe that recurrent fusion fails to exploit the long-term information as well as parallel fusion. As depicted in Fig.1a, VideoBEV benefits less ( $\Delta$ mAP↑) from utilizing historical frames than that of SOLOFusion ( $\Delta$ mAP↑). This phenomenon becomes even more pronounced, when utilizing more than four historical frames. + +Why is the performance of recurrent fusion not as good as parallel fusion? We conducted a detailed analysis of the differences between the two paradigms and found two reasons: (1) Gradient propagation. During the training process, only the recent frames contribute to gradients of fusion module in the recurrent paradigm, while all historical frames contribute equally in the parallel paradigm. This prevents the temporal fusion module of recurrent paradigm from learning fusion pattern from long-term temporal information of all historical frames. (2) Semantic ambiguity. Due to the coarse granularity of the BEV grids, the recurrent paradigm generates errors when aligning historical features multiple + +times, resulting in semantic ambiguity for the same object. As a result, recurrent paradigm may fuse different objects on the singular BEV grid, confusing the recognition of the same object. For the above two reasons, the previous recurrent method cannot fully and correctly utilize the long-term temporal information. + +Motivated by the above analysis, in this paper, we propose RecurrentBEV, a novel temporal fusion framework for BEV-based 3D object detector to improve the long-term fusion ability of recurrent paradigm while still has efficient latency and memory consumption in inference. First, RecurrentBEV adopts an RNN-style back-propagation during training, enabling historical frames to contribute gradients to the fusion module directly. This improvement solves the first problem while still maintaining inference efficiency. Second, we design an inner grid transformation before fusion, which can learn fine-grained semantic changes from ego-motion matrix. This effectively reduces semantic differences in the inter-frame alignment process. + +Based on the RNN-style back-propagation and the inner grid transformation, the proposed RecurrentBEV effectively improves long-term fusion in the recurrent paradigm, as shown in Fig.1a. Extensive experiments on the nuScenes 3D detection dataset evaluate the effectiveness of the proposed RecurrentBEV. As shown in Fig.1b, it outperforms both existing BEV-based algorithms such as SOLOFusion, VideoBEV and query-based algorithms such as StreamPETR [35], SparseBEV [27], Sparse4Dv2 [26] in both accuracy and efficiency. To summarize, our contributions are: + +- We provide an in-depth analysis for two temporal fusion paradigms and point out two factors that render the recurrent fusion less effective: issues with gradient propagation and semantic ambiguity. + +- We propose RecurrentBEV, a novel temporal fusion framework for BEV-based 3D object detector. It adapots RNN-style back-propagation and inner grid transformation, attaining state-of-the-art results (57.4% mAP and 65.1% NDS) on the nuScenes benchmark. It represents the first recurrent fusion method with explicit BEV features that surpasses parallel fusion performance. + +# 2 Related Works + +# 2.1 Multi-view 3D Object Detection + +Multi-view 3D object detection is a fundamental task for autonomous driving. Previous works mainly focus on how to transform the 2D features into a unified 3D space. This transformation can be divided to two categories, BEV-based methods [10, 13, 18, 22, 23, 41, 42] and query-based methods [25, 28, 36, 37], depending on whether an explicit 3D representation is employed. + +BEVDet [13] is one typical method of BEV-based multi-view 3D object detection. This method involves converting 2D features to BEV features based on the Lift-Splat-Shoot [34] paradigm. BEVDepth [21] is a further example, which notably refines the transformation process via explicit depth supervision. + +BEVFormer [22] further introduces a novel attention mechanism to gather 2D features for BEV grids. + +Query-based methods typically define several object queries in 3D space to gather features from multi-view image features. DETR3D [36] projects object queries to image coordinates to sample features then predicts 3D bounding boxes directly. DETR4D [31] and Sparse4D [25] enhance it by leveraging historical information. PETR [28] encodes the position information of 3D coordinates into image feature, facilitating the detection of 3D objects similar to DETR [3]. + +# 2.2 Temporal Fusion for BEV-based Detectors + +Recent works find that incorporating temporal information can enhance the stability in perception results and improve overall performance. BEVFormer [22] first proposes the temporal self-attention mechanism to fuse the historical BEV features in a recurrent paradigm. BEVDet4D [11] extends BEVDet [13] to a temporal version by employing a concatenation operation to integrate previous BEV features, usually called parallel paradigm. SOLOFusion [33] extends these methods by increasing the number of previous frames to 16, resulting in significant benefits. BEVFormerv2 [39] switches to parallel paradigm same as SOLOFusion. However, the inference latency and memory consumption of the parallel paradigm increase linearly as the historical frame lengths. VideoBEV [8] applies recurrent fusion to the LSS-based 3D detector, resulting in significant reduction of inference latency and memory consumption. However, the improvement of long-term fusion in VideoBEV is limited in comparison to SOLOFusion. In order to combine the advantages of the two paradigms, we construct a novel temporal fusion framework based on efficient recurrent paradigm. + +# 3 Method + +Converting multi-camera image features to bird's-eye-view (BEV) features can provide a unified surrounding environment representation for autonomous driving perception tasks. It involves transforming image data into a top-down representation of the environment, which often denoted as BEV features. The BEV detectors contain an image encoder, a 2D-3D view transformation, and a temporal fusion module. In this work, we present a new temporal fusion module, which can effectively aggregate spatio-temporal BEV features. + +# 3.1 Analysis of BEV-based Temporal Fusion Paradigms + +There are two kinds of temporal fusion paradigms in BEV detection, which are the effective parallel fusion and efficient recurrent fusion. The parallel fusion caches and updates multiple historical BEV features in learning the visual features for current frames. The recurrent fusion encodes all historical features into a single recurrent feature to cache and update. We conduct a detailed analysis to identify the disparities among them and pinpoint two crucial factors that cause + +![](images/327b47b23a4df406692a42d40688da02e529a5cc08c09b32bb2e34625d40be16.jpg) +(a) + +![](images/4396a3b608b529bd6704b4f031f15f2427c34399b933dd6f36f98335dffae64b.jpg) +(b) +Fig. 2: (a) Semantic ambiguity caused by grid sample operation. When ego vehicle turning right, BEV feature change from $B_{t-1}$ to $B_t$ . In fusion, the grid sample operation used for aligning $B_{t-1}$ and $B_t$ only aligns the position but ignores the orientation as shown in the blue and red dash lines. (b) Visualization of the semantic ambiguity. The green boxes represent the BEV feature change from $B_{t-1}$ , which have been transformed to align with the $B_t$ and decoded into bounding boxes. The blue boxes depict the actual bounding boxes in the $B_t$ . Notably, for static objects, there is a mismatch in the orientation of the green and blue boxes. + +the sub-optimal performance of recurrent fusion: issues with gradient propagation and semantic ambiguity. + +Gradient propagation Aligning historical frames to current frame is essential for BEV-based temporal fusion. Given the previous BEV feature $B_{j}$ , this process can be written as: + +$$ +\widetilde {B} _ {j, i} = \zeta \left(B _ {j}, P _ {j, i}\right) \tag {1} +$$ + +where $\widetilde{B}_{j,i}$ is the $j$ -th frame BEV feature in the ego coordinate of $i$ -th frame, $P_{j,i}$ is the ego-motion matrix between the $j$ -th frame and $i$ -th frame and $\zeta$ refers to the ego-motion transformation function, specifically grid sample operation [15]. Based on Eq.(1), We can express the parallel fusion of the $i$ -th frame as: + +$$ +\bar {B} _ {i} = \varphi \left(\widetilde {B} _ {i - k, i}, \dots , \widetilde {B} _ {i - 1, i}, B _ {i}, W\right) \tag {2} +$$ + +where $\overline{B}_i$ is the fused BEV feature of $i$ -th frame, $\varphi$ is the temporal fusion function (convolution [11,21,33,39] or deformable attention [22]) and $W$ refers the parameters of it. Similarly, the recurrent paradigm for the $i$ -th frame can be written as: + +$$ +\bar {B} _ {i} = \varphi (\zeta (\bar {B} _ {i - 1}, P _ {i - 1, i}), B _ {i}, W) \tag {3} +$$ + +![](images/2164b740a825f3d9c2171889420426ab1448c78963eccbb54182acc39e379e06.jpg) +Fig. 3: Overall architecture of the proposed RecurrentBEV. It shows a training process in the $t$ -th frame, while arrows indicate forward propagation and backward propagation. The red arrows indicate the additional back-propagation introduced by our method. The details of the Transformation-aware Gate Recurrent Unit (TA-GRU) is illustrated in dotted box. + +Based on Eq.(3), we can calculate the gradient for the parameters in the temporal fusion function for parallel fusion. + +$$ +\frac {\partial L}{\partial W} = \frac {\partial L}{\partial \bar {B} _ {i}} \cdot \varphi^ {\prime} \left(\widetilde {B} _ {i - k, i}, \dots , \widetilde {B} _ {i - 1, i}, B _ {i}\right) \tag {4} +$$ + +In recurrent fusion, $B_{i-1}$ is detached so that the gradient of $W$ in recurrent fusion can be written as: + +$$ +\frac {\partial L}{\partial W} = \frac {\partial L}{\partial \bar {B} _ {i}} \cdot \varphi^ {\prime} (\zeta (\bar {B} _ {i - 1}, P _ {i - 1, i}), B _ {i}) \tag {5} +$$ + +Comparing Eq.(4) and Eq.(5), it is evident that only the current BEV feature $B_{i}$ and one previous BEV feature $\overline{B}_{i-1}$ affect the gradient of $W$ directly in recurrent paradigm while all previous BEV features within a window size $k$ contribute to gradient of $W$ in parallel paradigm. This is one reason that recurrent paradigm makes limited progress in long-term fusion. + +Semantic Ambiguity To optimize memory usage, the recurrent fusion approach only caches a single historical 3D BEV feature instead of storing a sequence of 2D features. Consequently, this approach encounters challenges with temporal alignment in the 3D space. + +In the current recurrent fusion method, we have observed semantic ambiguity issues arising during the temporal alignment process in the 3D space. As illustrated in Fig.2a, the coarse granularity of BEV grids results in inaccurate transformation of orientation information within each grid during the temporal alignment process represented by Eq.(1). Consequently, this process introduces incorrect object orientations in real-world detection scenarios, as evident in Fig.2b. + +In the parallel fusion counterpart, temporal alignment in the 3D space is avoided by directly projecting a sequence of 2D view feature onto the current 3D BEV feature space [33, 39]. Assuming that the past perspective 2D view feature at time step $j$ is denoted as $C_j$ , the formulation for projecting it onto the BEV feature at time step $i$ is as follows: + +$$ +\widetilde {B} _ {j, i} = \operatorname {V i e w T r a n s} \left(C _ {j}, I _ {j}, P _ {j, i} \cdot E _ {j}\right) \tag {6} +$$ + +where $I_{j}$ and $E_{j}$ denote camera intrinsic and extrinsic respectively. This additional view transformation in 2D space effectively aligns objects between frames, subsequently enhancing the performance of parallel fusion. Nevertheless, this approach is not feasible for recurrent fusion due to the unavailability of 2D features during the recurrent fusion process. This discrepancy serves as another contributing factor to the performance gap between the two fusion methods. + +# 3.2 RecurrentBEV + +Based on the two above-mentioned possible reasons, we design a novel temporal fusion framework called RecurrentBEV. Similar to existing BEV-based 3D object detectors [8,11,21,33], RecurrentBEV contains an image encoder, a 2D-3D view transformation, a temporal fusion module, a BEV encoder and a task head, as illustrated in Fig.3. The main difference between RecurrentBEV and previous recurrent fusion is the temporal fusion module which can learn from all historical frames directly and learn fine-grained ego-motion transform. + +RNN-style Back-propagation In prior recurrent fusion methods [22], the $\overline{B}_{i-1}$ in Eq.(3) is solely computed for forward propagation, but not for backward propagation. Thus, historical BEV features do not contribute to the gradient of temporal fusion module in Eq.(5), preventing optimizing the temporal fusion module with all the relevant historical BEV features. To address this issue, RecurrentBEV adopts RNN-style Back-propagation that all the relevant historical features back-propagate gradients to adjacent layers. In this context, the gradients of parameters in temporal fusion function is written as: + +$$ +\frac {\partial L}{\partial W} = \sum_ {j = i - k} ^ {i} \frac {\partial L}{\partial \overline {{B}} _ {j}} \cdot \varphi^ {'} (\zeta (\overline {{B}} _ {j - 1}, P _ {j - 1, j}), B _ {j}) \tag {7} +$$ + +where $k$ is the number of total historical frames using in a iteration of training process. As a result, each historical BEV feature makes an equal contribution to the gradient of the temporal fusion module, akin to the parallel paradigm. + +Computing the back-propagation along the temporal dimension for the entire network places significant demands on GPU memory. As shown in Table 1 (Setting 1), if all modules, including image encoders, undergo back-propagation along the temporal dimension, memory overflow issues arise even when the frame number reaches only 2. However, it is imperative for the training process to incorporate long-term frames to effectively learn temporal information (refer to + +
Training frames0124816
Setting 115,06730,105OOMOOMOOMOOM
Setting 215,06717,66118,73519,97523,57930,941
+ +Table 1: The memory usage (MB) during training. OOM means out of memory on A100-40G with FP32 precision. + +Sec.4.6 for details). Therefore, we need to strike a balance between GPU memory usage (determined by the modules involved in back-propagation) and the length of the time sequence. In our design, as depicted by the red arrows in Figure 3, RecurrentBEV further computes all the gradients of the temporal fusion module (TA-GRU in the following) via back-propagation along the temporal dimension. Consequently, RecurrentBEV (Setting 2 in Table 1) can be trained with a batch size of 8 with FP32 precision on a single A100-40G GPU, effectively utilizing up to 16 consecutive frames. + +Transformation-aware Gated Recurrent Unit Because of the coarse granularity of the BEV grids, the semantic feature in a grid can not project correctly when model transforms the previous BEV features to the coordinate system of the current ego. We design an inner grid transformation to solve this problem. For simplicity, we use a linear layer to model this transformation: + +$$ +\widetilde {B} _ {j, i} = \operatorname {L i n e a r} \left(\zeta \left(B _ {j}, P _ {j, i}\right), W _ {p}\right) \tag {8} +$$ + +where $W_{p}$ is the parameter of Linear. For simplicity, we assume different channels in a projected feature share same motion embedding, so that the $W_{p}$ can be decoupled as: + +$$ +W _ {p} = E _ {j, i} \odot \widetilde {W} _ {p} \tag {9} +$$ + +where $E_{j,i} \in \mathbb{R}^{C_{in} \times 1}$ is ego motion embedding from timestamp $j$ to timestamp $i$ , $\widetilde{W}_p \in \mathbb{R}^{C_{in} \times C_{out}}$ is a learnable weight and $\odot$ denotes a Hadamard Product. Motion embedding $E_{j,i}$ is a high dimensional vector converted from the given ego motion matrix $P_{j,i}$ : + +$$ +E _ {j, i} = \sigma (\psi (P _ {j, i})) \tag {10} +$$ + +where $\sigma$ is a sigmoid activation and $\psi$ is a multi-layer perception (MLP). Combining Eq.(8), Eq.(9) and Eq.(10), the complete inner grid transformation can be written as: + +$$ +\widetilde {B} _ {j, i} = \operatorname {L i n e a r} \left(\zeta \left(\bar {B} _ {j}, P _ {j, i}\right), \sigma \left(\psi \left(P _ {j, i}\right)\right) \odot \widetilde {W} _ {p}\right) \tag {11} +$$ + +For the RNN modules, we adopted the typical Convolutional Gated Recurrent Unit (ConvGRU) [1]. By combining the aforementioned inner grid transformation, we introduce a novel temporal fusion module named the Transformation-Aware Gated Recurrent Unit (TA-GRU). The entire process of temporal fusion is formulated as follows: + +$$ +W _ {p} = \sigma \left(\psi \left(P _ {i - 1, i}\right)\right) \odot \widetilde {W} _ {p} \tag {12} +$$ + +$$ +\widetilde {B} _ {i - 1, i} = \operatorname {L i n e a r} \left(\zeta \left(\bar {B} _ {i - 1}, P _ {i - 1, i}\right), W _ {p}\right) \tag {13} +$$ + +$$ +Z _ {i} = \sigma \left(\operatorname {C o n v} \left(\left[ \widetilde {B} _ {i - 1, i}, B _ {i} \right], W _ {z}\right)\right) \tag {14} +$$ + +$$ +R _ {i} = \sigma \left(\operatorname {C o n v} \left(\left[ \widetilde {B} _ {i - 1, i}, B _ {i} \right], W _ {r}\right)\right) \tag {15} +$$ + +$$ +O _ {i} = \tanh \left(\operatorname {C o n v} \left(\left[ R _ {i} \odot \widetilde {B} _ {i - 1, i}, B _ {i} \right], W _ {o}\right)\right) \tag {16} +$$ + +$$ +\bar {B} _ {i} = \left(1 - Z _ {i}\right) \odot \tilde {B} _ {i - 1, i} + Z _ {i} \odot O _ {i} \tag {17} +$$ + +where $[\cdot, \cdot]$ represents the concatenation of feature maps and $W_{z}, W_{z}, W_{o}$ are 2D-convolution kernels. Because of inner grid transformation, TA-GRU can learn fine-grained semantic changes from ego-motion matrix, thus removing wrong orientation. + +Inference Efficiency Comparing Eq.(2) and Eq.(3), parallel fusion requires $k$ historical frames to be cached while recurrent fusion requires only one. Additionally, the transformation process described in Eq.(1) or Eq.(6) also needs to be computed $k$ times in parallel fusion, whereas in recurrent fusion it only needs to be computed once. As a result, recurrent fusion, which has lower inference latency and smaller memory consumption, is more suitable for deployment in the industry. + +Similar to previous recurrent fusion methods [8, 22], RecurrentBEV adopts sequential inference which only needs to cache a single historical BEV feature $\overline{B}$ , initialized with zero and updated at the end of the fusion. + +# 4 Experiments + +# 4.1 Dataset and Metrics + +We evaluate our method on the challenging nuScenes dataset [2]. The dataset comprises 1000 driving scenes, each lasting about 20 seconds, collected in Boston and Singapore, which are divided into 700, 150 and 150 scenes respectively for training, validation, and testing. The vehicle is equipped with 6 cameras for a complete 360-degree field of view. The nuScenes dataset annotates 23 object classes with accurate 3D bounding boxes for each scene at $2\mathrm{Hz}$ . For evaluation purposes, 10 classes are selected, namely: car, truck, bus, trailer, construction vehicle, pedestrian, motorcycle, bicycle, barrier, and traffic cone. As for evaluation, we use the following metrics, mean average precision (mAP), the nuScenes detection score (NDS), and five true positive metrics including average translation error (ATE), average scale error (ASE), average orientation error (AOE), average velocity error (AVE), average attribute error (AAE). + +
MethodsBackboneImage SizeEDmAP↑NDS↑mATE↓mASE↓mAOE↓mAVE↓mAAE↓
BEVDet [13]ResNet50256 × 704X0.2980.3790.7250.2790.5890.8600.245
BEVDet4D [11]ResNet50256 × 704X0.3220.4570.7030.2780.4950.3540.206
BEVDepth [21]ResNet50256 × 704X0.3510.4750.6390.2670.4790.4280.198
BEVStereo [20]ResNet50256 × 704X0.3720.5000.5980.2700.4380.3670.190
BEVPoolv2 [12]ResNet50256 × 704X0.4060.5260.5720.2750.4630.2750.188
SOLOFusion [33]ResNet50256 × 704X0.4270.5340.5670.2740.5110.2520.181
VideoBEV [8]ResNet50256 × 704X0.4220.5350.5640.2760.4400.2860.198
PETRv2 [29]ResNet50256 × 704X0.3490.4560.7000.2750.5800.4370.187
Sparse4Dv2 [26]ResNet50256 × 704X0.4390.5390.5980.2700.4750.2820.179
SparseBEV [27]ResNet50256 × 704X0.4320.5450.6060.2740.3870.2510.186
StreamPETR [35]ResNet50256 × 704X0.4320.5400.5810.2720.4130.2950.195
RecurrentBEVResNet50256 × 704X0.4450.5490.5550.2720.4510.2560.204
BEVDepth [13]ResNet101512 × 1408X0.4120.5350.5650.2660.3580.3310.190
SOLOFusion [33]ResNet101512 × 1408X0.4830.5820.5030.2640.3810.2460.207
Sparse4Dv2 [26]ResNet101512 × 1408X0.4850.5800.5550.2720.3670.2560.182
RecurrentBEVResNet101512 × 1408X0.5090.5990.4870.2670.3580.2470.199
SparseBEV [27]ResNet101512 × 1408✓0.5010.5920.5620.2650.3210.2430.195
StreamPETR [35]ResNet101512 × 1408✓0.5040.5920.5690.2620.3150.2570.199
Sparse4Dv2 [26]ResNet101512 × 1408✓0.5050.5940.5480.2680.3480.2390.184
RecurrentBEVResNet101512 × 1408✓0.5280.6120.4810.2560.3550.2260.201
+ +Table 2: Comparison results on 3D detection on the nuScenes validation set. ED is external data which indicates to use pre-trained weights from the external dataset i.e. nuImages. + +# 4.2 Implementation Details + +We conduct experiments based on the BEVDet4D-Depth [11, 21] with the same learning rate, optimizer, and data augmentation. Drawing upon previous methods [21, 33, 35], we utilize ResNet50 [9], ResNet101, ConNeXt-B [30] and V2-99 [17] as backbones. On the nuScenes validation set, we report results of ResNet50 and ResNet101 models with pre-trained weights ImageNet [6] and nuImages [35]. On the nuScenes test set, we provide the performance of V2-99 model initialized from DD3D [32] and ConvNeXt-B model benefited from COCO [24] pre-training. The models in the ablation study are trained for 24 epochs without CBGS [43]. When compared with others, we follow the training setting of StreamPETR [35], extending the training epochs to 60 and using 2D auxiliary supervision for a fair comparison. + +To achieve real-time inference speed, we optimize the BEVDet4D-Depth baseline by making several key modifications. Firstly, we reduce the number of depth bins. Secondly, we eliminate the ASPP [5] module from the DepthNet to minimize the latency associated with view transformation. Additionally, we merge 6 detection tasks performed by CenterHead [40] into a single task, significantly reducing post-processing time. These optimizations combined enable us to achieve a more efficient inference speed. + +
MethodsBackbonemAP↑NDS↑mATE↓mASE↓mAOE↓mAVE↓mAAE↓Conference
DETR3D* [36]V2-990.4120.4790.6410.2550.3940.8450.133CoRL2021
PETR [28]V2-990.4410.5040.5930.2490.3830.8080.132ECCV2022
UVTR [19]V2-990.4720.5510.5770.2530.3910.5080.123NIPS2022
BEVFormer [22]V2-990.4810.5690.5820.2560.3750.3780.126ECCV2022
PETRv2 [29]V2-990.4900.5820.5610.2430.3610.3430.120ICCV2023
PolarFormer [16]V2-990.4930.5720.5560.2560.3640.4390.127AAAI2023
BEVDepth [21]V2-990.5030.6000.4450.2450.3780.3200.126AAAI2023
VEDet [4]V2-990.5050.5850.5450.2440.3460.4210.123CVPR2023
MV2D [37]V2-990.5110.5960.5250.2430.3570.3570.120ICCV2023
CAPE [38]V2-990.5250.6100.5030.2420.3610.3060.114CVPR2023
BEVStereo [20]V2-990.5250.6100.4310.2460.3580.3570.138AAAI2023
SA-BEV [41]V2-990.5330.6240.4300.2410.3380.2820.139ICCV2023
FB-BEV [23]V2-990.5370.6240.4390.2500.3580.2700.128ICCV2023
StreamPETR [35]V2-990.5500.6360.4790.2390.3170.2410.119ICCV2023
SparseBEV [27]V2-990.5560.6360.4850.2440.3320.2460.117ICCV2023
Sparse4Dv2 [26]V2-990.5570.6380.4620.2380.3280.2640.115-
RecurrentBEVV2-990.5730.6510.4100.2470.3230.2490.123-
BEVDepth [21]ConvNeXt-B0.5200.6090.4450.2430.3520.3470.127AAAI2023
AeDet* [7]ConvNeXt-B0.5310.6200.4390.2470.3440.2920.130CVPR2023
SOLOFusion [33]ConvNeXt-B0.5400.6190.4530.2570.3760.2760.148ICLR2023
VCD-A [14]ConvNeXt-B0.5480.6310.4360.2440.3430.2900.120NIPS2023
StreamPETR [35]ConvNeXt-B0.5500.6310.4930.2410.3430.2430.123ICCV2023
VideoBEV [8]ConvNeXt-B0.5540.6290.4570.2490.3810.2660.132-
RecurrentBEVConvNeXt-B0.5740.6510.4110.2480.3330.2570.116-
+ +Table 3: Comparison results on 3D detection on the nuScenes test set. * indicates using test-time augmentation in test process. No future frames are used in any of the methods. + +# 4.3 Main Results + +We compare the proposed RecurrentBEV with previous state-of-the-art multiview 3D detectors on the nuScenes validation and test set. + +On the nuScenes validation set, as shown in Fig.1b and Tab.2, the real-time version of RecurrentBEV (25.6 FPS), using ResNet50 as backbone and $256 \times 704$ as the input size, outperforms the previous state-of-the-art method StreamPETR by $1.3\%$ mAP and $0.9\%$ NDS on the nuScenes validation set. To scale up our method, we adopt ResNet101 backbone with nuImages pre-training and extend input size to $512 \times 1408$ , outperforming the same setting StreamPETR by $2.4\%$ mAP and $2.0\%$ NDS. It's worth noting that RecurrentBEV makes remarkable progress on the mAP and mATE metrics, indicating the improvement of detection of occlusion objects. We'll delve into the reasons behind this in Sec.4.6. + +On the nuScenes test set in Tab.3, we compare with all the recent methods in multi-view 3D object detection. Following the settings in StreamPETR, we adopt V2-99 and ConvNeXt-B as backbone and $640 \times 1600$ as the input size. Using V2-99 as backbone, RecurrentBEV can surpass the previous state-of-the-art method Sparse4Dv2 by $1.6\%$ mAP and $1.3\%$ NDS. When the backbone is scaled up to ConvNeXt-B, RecurrentBEV demonstrates comparable performance, surpassing other methods. + +
methodmAP↑NDS↑memory usage↓latency↑
parallel fusion [33]0.4270.534945MB96ms
recurrent fusion baseline0.3960.515516MB34ms
RecurrentBEV0.4450.549549MB39ms
+ +Table 4: Comparison of time and memory in inference. All data are measured on a single RTX 3090 with PyTorch fp32. Implementation details are the same as Tab.2 ResNet50 Backbone. For more latency analysis, please refer to supplementary. + +# 4.4 Time and Memory Comparison + +To demonstrate the time and memory efficiency in inference, we compare the proposed RecurrentBEV with parallel fusion methods and previous recurrent fusion baselines. As shown in Tab.4 RecurrentBEV is 2.5 times faster than the normal parallel fusion method SOLOFusion [33], and the memory usage is reduced by $42\%$ . When compared to the normal recurrent baseline, RecurrentBEV requires only an additional 33MB of memory while significantly improving the mean Average Precision (mAP) from 0.396 to 0.445. + +# 4.5 Ablation Study + +In the ablation study, we delve into three aspects: the components of RecurrentBEV, the receptive field of TA-GRU, and the length of frame sequence in training. We conduct the ablation experiments with ResNet-50 backbone. + +Ablation of Each Component. We conduct a component-wise ablation in Tab.5. Exp.1 starts with a ConvGRU as temporal fusion module, achieving similar performance to VideoBEV [8]. In Exp.2 and Exp.3, we enable the backward propagation of historical BEV features in training process. In Exp.3 and Exp.4, we replace the temporal fusion module with TA-GRU. The overall components (BP and TA-GRU) brings significant performance gains $(+1.7\% \mathrm{mAP}, +1.6\%$ NDS,-4.7% mAOE) over baseline. + +Specifically, the TA-GRU demonstrates its ability to enhance the mAP and mAOE metrics, regardless of whether backward propagation of historical BEV features is enabled. This proves the TA-GRU's consistent ability to alleviate semantic ambiguity for identical objects. Notably, the mean average orientation error (mAOE) has a notable increase, indicating that the inner grid transformation, facilitated by the alignment of historical BEV features, effectively mitigates semantic ambiguity by refining the orientation of objects. + +As demonstrated in Exp.4, when the historical BEV feature is detached, the performance suffers, highlighting the importance of RNN-style back-propagation. However, relying solely on back-propagation of historical BEV features during training results in limited improvement. This shows the necessity of employing + +
Exp|BP TG|mAP↑|NDS↑|mAOE↓|mAVE↓
10.3710.4760.6460.317
2✓0.3750.4790.6340.319
3✓✓0.3880.4920.5960.281
4✓0.3780.4840.6020.316
+ +Table 5: Ablation of each component. BP means the historical BEV feature backward propagate. TG means TA-GRU. + +
Kernel sizemAP↑NDS↑mATE↓mAVE↓
1 × 10.3880.4920.6470.281
3 × 30.3970.5000.6320.284
5 × 50.4030.5060.6150.276
7 × 70.3980.5010.6300.273
9 × 90.3940.4960.6330.280
+ +Table 6: Ablation of receptive field. + +
Training framesmAP↑NDS↑mATE↓mAVE↓
00.3180.3800.7420.835
10.3460.4570.6960.389
20.3720.4790.6650.310
40.3900.4930.6460.283
80.4030.5060.6150.276
160.4020.490.6280.280
+ +Table 7: Ablation of sequence length. + +the correct temporal fusion module for effective back-propagation, as the combination of these two components can achieve a synergistic effect that surpasses the sum of their individual contributions. + +Ablation of Receptive Field. Receptive field is critical for temporal fusion of moving objects. We explore the kernel size of convolutions in TA-GRU in Tab.6. As we can see, an excessively large or small receptive field can have a detrimental effect on performance. Our suspect is that a fusion module with an overlarge receptive field gathers noisy information, while a small receptive field leads to mismatches for moving objects. Consequently, we use a $5 \times 5$ kernel size in TA-GRU, unless otherwise specified. + +Ablation of Training Sequence Length. We use the BEVDet-Depth with no temporal fusion as baseline and apply our method with different training sequence length in Tab.7. The fusion of a single frame greatly enhances velocity prediction (mAVE). The fusion of four to eight frames mainly enhances localisation metrics mAP and mATE, with an increase of $8.5\%$ and $12.7\%$ respectively. Further increasing the number of frames to 16 offers gains limited performance improvement. Therefore, we train our models on 8 frames. + +# 4.6 Effect of Recurrent Long-term Fusion + +Detection of Occlusion Objects. Intuitively, long-term fusion has the potential to aid in the detection of occluded objects. Therefore, we conducted an analysis of object occlusions on the nuScenes validation set. Firstly, we define a 'distinct object' as one where the visibility of the entire object exceeds $60\%$ . Subsequently, we determined that if an object is classified as a 'distinct object' in any frame of the video, it is considered visible throughout the video. Finally, we divided the number of visible objects by the total number of objects present to get the overall visibility of the video. As shown in Fig.4a, the visibility of the video has the same trend with mAP, proving the improvement of mAP in long-term fusion originates from the detection of occlusion objects. + +![](images/812dfcd2a3b4fa1490e5d506dfd24ae98b085b72a1abee35dca80c25252e598c.jpg) +(a) + +![](images/398adb6d4e1c1b0ed0a2dc7fd1f5ab3add939cdb6f29f70d65c2dd144c0e846d.jpg) +(b) +Fig. 4: (a) Relation between visibility and mAP. Visibility refers to the proportion of distinct objects in a video, expressed as a percentage of the total number of objects present. The Pearson correlation coefficient between visibility and mAP is 0.977.(b) The mATE metrics with objects at different location on the nuScenes validation set. The blue one is a result from single frame baseline and the orange one is RecurrentBEV. The objects between 25 meters and 30 meters improve the mATE metric by $24\%$ mostly. + +Localization of Distant Objects. The localization of objects is a major challenge for camera-based 3D detector. Due to the uncertainty of depth estimation. We speculate long-term fusion can more accurately localize objects that move from near to far as the ego vehicle moves. Therefore, we calculate the mATE metrics of objects at different distances before and after long-term fusion, as in Fig.4b. Notably, objects between 25 and 30 meters exhibit the most significant improvement in mATE through long-term fusion, whereas nearby objects show marginal gains. This finding demonstrates the efficacy of long-term fusion in accurately localizing distant objects. + +# 5 Conclusion and Future Work + +In this paper, we investigate the differences between recurrent fusion and parallel fusion of BEV-based 3D object detectors, and propose RecurrentBEV, a simple and effective recurrent temporal fusion method. By incorporating RNN-style back-propagation and a newly designed inner grid transformation, RecurrentBEV enhances long-term fusion capabilities while maintaining latency and memory advantages. Rigorous experiments conducted on the nuScenes dataset demonstrate that RecurrentBEV outperforms all prior methods, while maintaining competitive inference speeds, thereby facilitating deployment. To our knowledge, it represents the first recurrent fusion method with explicit BEV features that surpasses parallel fusion performance. + +In the future, we will generalize recurrent fusion idea to more BEV architectures such as different BEV view transform backbones and query-based methods. We will also try to extend RecurrentBEV to other 3D perceptual tasks and datasets such as object tracking and trajectory prediction. + +# Acknowledgements + +This work is partially supported by the National Key R&D Program of China (under Grant 2023YFB4502200), the NSF of China (under Grants U22A2028, 61925208, 62102399, 62222214, 62341411, 62102398, U20A20227, 62372436, 623-02478, 62302482, 62302483, 62302480), Strategic Priority Research Program of the Chinese Academy of Sciences, (Grant No.XDB0660200, XDB0660201, XDB0660202), CAS Project for Young Scientists in Basic Research(YSBR-029), Youth Innovation Promotion Association CAS and Xplore Prize. + +# References + +1. Ballas, N., Yao, L., Pal, C., Courville, A.: Delving deeper into convolutional networks for learning video representations. In: ICLR (2016) +2. Caesar, H., Bankiti, V., Lang, A.H., Vora, S., Liong, V.E., Xu, Q., Krishnan, A., Pan, Y., Baldan, G., Beijbom, O.: nuscenes: A multimodal dataset for autonomous driving. In: CVPR (2020) +3. Carion, N., Massa, F., Synnaeve, G., Usunier, N., Kirillov, A., Zagoruyko, S.: End-to-end object detection with transformers. In: ECCV (2020) +4. Chen, D., Li, J., Guizilini, V., Ambrus, R.A., Gaidon, A.: Viewpoint equivariance for multi-view 3d object detection. In: CVPR (2023) +5. Chen, L.C., Papandreou, G., Kokkinos, I., Murphy, K., Yuille, A.L.: Deeplab: Semantic image segmentation with deep convolutional nets, atrous convolution, and fully connected crfs. IEEE TPAMI (2017) +6. Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: Imagenet: A large-scale hierarchical image database. In: CVPR (2009) +7. Feng, C., Jie, Z., Zhong, Y., Chu, X., Ma, L.: Aedet: Azimuth-invariant multi-view 3d object detection. In: CVPR (2023) +8. Han, C., Sun, J., Ge, Z., Yang, J., Dong, R., Zhou, H., Mao, W., Peng, Y., Zhang, X.: Exploring recurrent long-term temporal fusion for multi-view 3d perception. arXiv preprint arXiv:2303.05970 (2023) +9. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: CVPR (2016) +0. Huang, B., Li, Y., Xie, E., Liang, F., Wang, L., Shen, M., Liu, F., Wang, T., Luo, P., Shao, J.: Fast-bev: Towards real-time on-vehicle bird's-eye view perception. arXiv preprint arXiv:2301.07870 (2023) +1. Huang, J., Huang, G.: Bevdet4d: Exploit temporal cues in multi-camera 3d object detection. arXiv preprint arXiv:2203.17054 (2022) +2. Huang, J., Huang, G.: Bevpoolv2: A cutting-edge implementation of bevdet toward deployment. arXiv preprint arXiv:2211.17111 (2022) +3. Huang, J., Huang, G., Zhu, Z., Ye, Y., Du, D.: Bevdet: High-performance multicamera 3d object detection in bird-eye-view. arXiv preprint arXiv:2112.11790 (2021) +4. Huang, L., Li, Z., Sima, C., Wang, W., Wang, J., Qiao, Y., Li, H.: Leveraging vision-centric multi-modal expertise for 3d object detection. In: NeurIPS (2023) +5. Jaderberg, M., Simonyan, K., Zisserman, A., et al.: Spatial transformer networks. In: NeurIPS (2015) +6. Jiang, Y., Zhang, L., Miao, Z., Zhu, X., Gao, J., Hu, W., Jiang, Y.G.: Polarformer: Multi-camera 3d object detection with polar transformer. In: AAAI (2023) + +17. Lee, Y., Hwang, J.w., Lee, S., Bae, Y., Park, J.: An energy andgpu-computation efficient backbone network for real-time object detection. In: CVPRW (2019) +18. Li, H., Zhang, H., Zeng, Z., Liu, S., Li, F., Ren, T., Zhang, L.: Dfa3d: 3d deformable attention for 2d-to-3d feature lifting. In: ICCV (2023) +19. Li, Y., Chen, Y., Qi, X., Li, Z., Sun, J., Jia, J.: Unifying voxel-based representation with transformer for 3d object detection. In: NeurIPS (2022) +20. Li, Y., Bao, H., Ge, Z., Yang, J., Sun, J., Li, Z.: Bevstereo: enhancing depth estimation in multi-view 3d object detection with temporal stereo. In: AAAI (2023) +21. Li, Y., Ge, Z., Yu, G., Yang, J., Wang, Z., Shi, Y., Sun, J., Li, Z.: Bevdepth: Acquisition of reliable depth for multi-view 3d object detection. In: AAAI (2023) +22. Li, Z., Wang, W., Li, H., Xie, E., Sima, C., Lu, T., Qiao, Y., Dai, J.: Bevformer: Learning bird's-eye-view representation from multi-camera images via spatiotemporal transformers. In: ECCV (2022) +23. Li, Z., Yu, Z., Wang, W., Anandkumar, A., Lu, T., Alvarez, J.M.: Fb-bev: Bev representation from forward-backward view transformations. In: ICCV (2023) +24. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dóllar, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: ECCV (2014) +25. Lin, X., Lin, T., Pei, Z., Huang, L., Su, Z.: Sparse4d: Multi-view 3d object detection with sparse spatial-temporal fusion. arXiv preprint arXiv:2211.10581 (2022) +26. Lin, X., Lin, T., Pei, Z., Huang, L., Su, Z.: Sparse4d v2: Recurrent temporal fusion with sparse model. arXiv preprint arXiv:2305.14018 (2023) +27. Liu, H., Teng, Y., Lu, T., Wang, H., Wang, L.: Sparsebev: High-performance sparse 3d object detection from multi-camera videos. In: ICCV (2023) +28. Liu, Y., Wang, T., Zhang, X., Sun, J.: Petr: Position embedding transformation for multi-view 3d object detection. In: ECCV (2022) +29. Liu, Y., Yan, J., Jia, F., Li, S., Gao, A., Wang, T., Zhang, X.: Petrov2: A unified framework for 3d perception from multi-camera images. In: ICCV (2023) +30. Liu, Z., Mao, H., Wu, C.Y., Feichtenhofer, C., Darrell, T., Xie, S.: A convnet for the 2020s. In: CVPR (2022) +31. Luo, Z., Zhou, C., Zhang, G., Lu, S.: Detr4d: Direct multi-view 3d object detection with sparse attention. arXiv preprint arXiv:2212.07849 (2022) +32. Park, D., Ambrus, R., Guizilini, V., Li, J., Gaidon, A.: Is pseudo-lidar needed for monocular 3d object detection? In: ICCV (2021) +33. Park, J., Xu, C., Yang, S., Keutzer, K., Kitani, K., Tomizuka, M., Zhan, W.: Time will tell: New outlooks and a baseline for temporal multi-view 3d object detection. In: ICLR (2023) +34. Philion, J., Fidler, S.: Lift, splat, shoot: Encoding images from arbitrary camera rigs by implicitly unprojecting to 3d. In: ECCV (2020) +35. Wang, S., Liu, Y., Wang, T., Li, Y., Zhang, X.: Exploring object-centric temporal modeling for efficient multi-view 3d object detection. In: ICCV (2023) +36. Wang, Y., Guizilini, V.C., Zhang, T., Wang, Y., Zhao, H., Solomon, J.: Detr3d: 3d object detection from multi-view images via 3d-to-2d queries. In: Conference on Robot Learning (2022) +37. Wang, Z., Huang, Z., Fu, J., Wang, N., Liu, S.: Object as query: Lifting any 2d object detector to 3d detection. In: ICCV (2023) +38. Xiong, K., Gong, S., Ye, X., Tan, X., Wan, J., Ding, E., Wang, J., Bai, X.: Cape: Camera view position embedding for multi-view 3d object detection. In: CVPR (2023) +39. Yang, C., Chen, Y., Tian, H., Tao, C., Zhu, X., Zhang, Z., Huang, G., Li, H., Qiao, Y., Lu, L., et al.: Bevformer v2: Adapting modern image backbones to bird's-eye-view recognition via perspective supervision. In: CVPR (2023) + +40. Yin, T., Zhou, X., Krahenbuhl, P.: Center-based 3d object detection and tracking. In: CVPR (2021) +41. Zhang, J., Zhang, Y., Liu, Q., Wang, Y.: Sa-bev: Generating semantic-aware bird's-eye-view feature for multi-view 3d object detection. In: ICCV (2023) +42. Zhou, H., Ge, Z., Li, Z., Zhang, X.: Matrixvt: Efficient multi-camera to bev transformation for 3d perception. In: ICCV (2023) +43. Zhu, B., Jiang, Z., Zhou, X., Li, Z., Yu, G.: Class-balanced grouping and sampling for point cloud 3d object detection. arXiv preprint arXiv:1908.09492 (2019) \ No newline at end of file diff --git a/2024/RecurrentBEV_ A Long-term Temporal Fusion Framework for Multi-view 3D Detection/images.zip b/2024/RecurrentBEV_ A Long-term Temporal Fusion Framework for Multi-view 3D Detection/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..803eb51d36178f5736d0ad6c282e882e01bc77ba --- /dev/null +++ b/2024/RecurrentBEV_ A Long-term Temporal Fusion Framework for Multi-view 3D Detection/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a89fd73a7697e041e6e256fba003123268b0b726ee9e3e19a5d1d6233a26a6a0 +size 663257 diff --git a/2024/RecurrentBEV_ A Long-term Temporal Fusion Framework for Multi-view 3D Detection/layout.json b/2024/RecurrentBEV_ A Long-term Temporal Fusion Framework for Multi-view 3D Detection/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..8ea0c1afe5a9a19d99e2179e4eb80201e244e836 --- /dev/null +++ b/2024/RecurrentBEV_ A Long-term Temporal Fusion Framework for Multi-view 3D Detection/layout.json @@ -0,0 +1,9081 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 145, + 111, + 470, + 147 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 111, + 470, + 147 + ], + "spans": [ + { + "bbox": [ + 145, + 111, + 470, + 147 + ], + "type": "text", + "content": "RecurrentBEV: A Long-term Temporal Fusion Framework for Multi-view 3D Detection" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 148, + 167, + 465, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 148, + 167, + 465, + 193 + ], + "spans": [ + { + "bbox": [ + 148, + 167, + 465, + 193 + ], + "type": "text", + "content": "Ming Chang" + }, + { + "bbox": [ + 148, + 167, + 465, + 193 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 148, + 167, + 465, + 193 + ], + "type": "text", + "content": ", Xishan Zhang" + }, + { + "bbox": [ + 148, + 167, + 465, + 193 + ], + "type": "inline_equation", + "content": "^{1,2\\star}" + }, + { + "bbox": [ + 148, + 167, + 465, + 193 + ], + "type": "text", + "content": ", Rui Zhang" + }, + { + "bbox": [ + 148, + 167, + 465, + 193 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 148, + 167, + 465, + 193 + ], + "type": "text", + "content": ", Zhipeng Zhao" + }, + { + "bbox": [ + 148, + 167, + 465, + 193 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 148, + 167, + 465, + 193 + ], + "type": "text", + "content": ", Guanhua He" + }, + { + "bbox": [ + 148, + 167, + 465, + 193 + ], + "type": "inline_equation", + "content": "^{1,2,3}" + }, + { + "bbox": [ + 148, + 167, + 465, + 193 + ], + "type": "text", + "content": ", and Shaoli Liu" + }, + { + "bbox": [ + 148, + 167, + 465, + 193 + ], + "type": "inline_equation", + "content": "^{1}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 140, + 201, + 473, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 140, + 201, + 473, + 213 + ], + "spans": [ + { + "bbox": [ + 140, + 201, + 473, + 213 + ], + "type": "text", + "content": "1 Cambricon Technologies {changming,zhaozhipeng,liushaoli}@cambricon.com" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 146, + 213, + 466, + 235 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 213, + 466, + 235 + ], + "spans": [ + { + "bbox": [ + 146, + 213, + 466, + 235 + ], + "type": "text", + "content": "2 SKL of Processors, Institute of Computing Technology, CAS, Beijing, China {zhangxishan,zhangrui}@ict.ac.cn, heguanhua22@mails.ucas.ac.cn" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 167, + 235, + 446, + 246 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 167, + 235, + 446, + 246 + ], + "spans": [ + { + "bbox": [ + 167, + 235, + 446, + 246 + ], + "type": "text", + "content": "3 University of Chinese Academy of Sciences, UCAS, Beijing, China" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 160, + 276, + 452, + 507 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 276, + 452, + 507 + ], + "spans": [ + { + "bbox": [ + 160, + 276, + 452, + 507 + ], + "type": "text", + "content": "Abstract. Long-term temporal fusion is frequently employed in camera-based Bird's-Eye-View (BEV) 3D object detection to improve detection of occluded objects. Existing methods can be divided into two categories, parallel fusion and recurrent fusion. Recurrent fusion reduces inference latency and memory consumption but fails to exploit the long-term information as well as parallel fusion. In this paper, we first find two shortcomings of recurrent fusion paradigm: (1) Gradients of previous BEV features cannot directly contribute to the fusion module. (2) Semantic ambiguity are caused by coarse granularity of the BEV grids during aligning BEV features. Then based on the above analysis, we propose RecurrentBEV, a novel recurrent temporal fusion method for BEV based 3D object detector. By adopting RNN-style back-propagation and new-designed inner grid transformation, RecurrentBEV improves the long-term fusion ability while still enjoying efficient inference latency and memory consumption during inference. Extensive experiments on the nuScenes benchmark demonstrate its effectiveness, achieving a new state-of-the-art performance of " + }, + { + "bbox": [ + 160, + 276, + 452, + 507 + ], + "type": "inline_equation", + "content": "57.4\\%" + }, + { + "bbox": [ + 160, + 276, + 452, + 507 + ], + "type": "text", + "content": " mAP and " + }, + { + "bbox": [ + 160, + 276, + 452, + 507 + ], + "type": "inline_equation", + "content": "65.1\\%" + }, + { + "bbox": [ + 160, + 276, + 452, + 507 + ], + "type": "text", + "content": " NDS on the test set. The real-time version (25.6 FPS) achieves " + }, + { + "bbox": [ + 160, + 276, + 452, + 507 + ], + "type": "inline_equation", + "content": "44.5\\%" + }, + { + "bbox": [ + 160, + 276, + 452, + 507 + ], + "type": "text", + "content": " mAP and " + }, + { + "bbox": [ + 160, + 276, + 452, + 507 + ], + "type": "inline_equation", + "content": "54.9\\%" + }, + { + "bbox": [ + 160, + 276, + 452, + 507 + ], + "type": "text", + "content": " NDS without external dataset, outperforming the previous best method StreamPETR by " + }, + { + "bbox": [ + 160, + 276, + 452, + 507 + ], + "type": "inline_equation", + "content": "1.3\\%" + }, + { + "bbox": [ + 160, + 276, + 452, + 507 + ], + "type": "text", + "content": " mAP and " + }, + { + "bbox": [ + 160, + 276, + 452, + 507 + ], + "type": "inline_equation", + "content": "0.9\\%" + }, + { + "bbox": [ + 160, + 276, + 452, + 507 + ], + "type": "text", + "content": " NDS. The code is available at https://github.com/lucifer443/RecurrentBEV." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 159, + 517, + 453, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 517, + 453, + 540 + ], + "spans": [ + { + "bbox": [ + 159, + 517, + 453, + 540 + ], + "type": "text", + "content": "Keywords: Autonomous Driving " + }, + { + "bbox": [ + 159, + 517, + 453, + 540 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 517, + 453, + 540 + ], + "type": "text", + "content": " Bird's-Eye-View " + }, + { + "bbox": [ + 159, + 517, + 453, + 540 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 517, + 453, + 540 + ], + "type": "text", + "content": " 3D Object Detection " + }, + { + "bbox": [ + 159, + 517, + 453, + 540 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 517, + 453, + 540 + ], + "type": "text", + "content": " Temporal Fusion" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 561, + 230, + 574 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 561, + 230, + 574 + ], + "spans": [ + { + "bbox": [ + 132, + 561, + 230, + 574 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 586, + 482, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 586, + 482, + 647 + ], + "spans": [ + { + "bbox": [ + 130, + 586, + 482, + 647 + ], + "type": "text", + "content": "Camera-only 3D detection presents a formidable challenge for autonomous driving and has garnered significant attention in recent years owing to its low cost of deployment. The Bird's-Eye-View (BEV) representation-based approach has recently demonstrated notable improvement by leveraging temporal information [8,11,21,22,33]." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 653, + 238, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 653, + 238, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 653, + 238, + 665 + ], + "type": "text", + "content": "* Corresponding Author." + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 134, + 116, + 299, + 243 + ], + "blocks": [ + { + "bbox": [ + 134, + 116, + 299, + 243 + ], + "lines": [ + { + "bbox": [ + 134, + 116, + 299, + 243 + ], + "spans": [ + { + "bbox": [ + 134, + 116, + 299, + 243 + ], + "type": "image", + "image_path": "a0c2c30d14ddc03080547d6318c6c605f57895cc2abc4416a369c72668279d1c.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 211, + 247, + 223, + 257 + ], + "lines": [ + { + "bbox": [ + 211, + 247, + 223, + 257 + ], + "spans": [ + { + "bbox": [ + 211, + 247, + 223, + 257 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 315, + 118, + 479, + 243 + ], + "blocks": [ + { + "bbox": [ + 315, + 118, + 479, + 243 + ], + "lines": [ + { + "bbox": [ + 315, + 118, + 479, + 243 + ], + "spans": [ + { + "bbox": [ + 315, + 118, + 479, + 243 + ], + "type": "image", + "image_path": "1e566910a3ade95c9c7492688b428a3a9751acd0ed968f9126a98c27112afe4a.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 390, + 247, + 403, + 257 + ], + "lines": [ + { + "bbox": [ + 390, + 247, + 403, + 257 + ], + "spans": [ + { + "bbox": [ + 390, + 247, + 403, + 257 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 266, + 482, + 322 + ], + "lines": [ + { + "bbox": [ + 130, + 266, + 482, + 322 + ], + "spans": [ + { + "bbox": [ + 130, + 266, + 482, + 322 + ], + "type": "text", + "content": "Fig. 1: (a) Comparison of improvement " + }, + { + "bbox": [ + 130, + 266, + 482, + 322 + ], + "type": "inline_equation", + "content": "(\\Delta \\mathrm{mAP}\\uparrow)" + }, + { + "bbox": [ + 130, + 266, + 482, + 322 + ], + "type": "text", + "content": " from long-term fusion between parallel paradigm (SOLOFusion [33]) and recurrent paradigm (VideoBEV [8], Ours). (b) Performance comparison on the validation set of nuScenes [2]. All methods use ResNet50 [9] pre-trained on ImageNet [6] as backbone with " + }, + { + "bbox": [ + 130, + 266, + 482, + 322 + ], + "type": "inline_equation", + "content": "256\\times 704" + }, + { + "bbox": [ + 130, + 266, + 482, + 322 + ], + "type": "text", + "content": " as input size. Frames per second (FPS) is measured on a single RTX 3090 with the PyTorch fp32." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 350, + 482, + 554 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 350, + 482, + 554 + ], + "spans": [ + { + "bbox": [ + 130, + 350, + 482, + 554 + ], + "type": "text", + "content": "The temporal fusion mechanisms utilised in the BEV-based 3D detector can be broadly categorised into parallel fusion [11,33,39], and recurrent fusion [8,22]. Parallel fusion [11,33,39] caches and updates multiple historical features in learning the visual features for current frames. Therefore, the overhead of computing and storage grows as the length of the historical frame feature increases. Since the superior performance requires longer historical information, this fusion paradigm faces a dilemma between effectiveness and efficiency in deployment. In contrast, recurrent fusion [22] encodes all historical features into a single recurrent feature to cache and update, so it largely reduces inference latency and memory consumption. For instance, when employing 16 historical frames, the recurrent fusion method VideoBEV [8] exhibits " + }, + { + "bbox": [ + 130, + 350, + 482, + 554 + ], + "type": "inline_equation", + "content": "6 \\times" + }, + { + "bbox": [ + 130, + 350, + 482, + 554 + ], + "type": "text", + "content": " times faster in latency and uses " + }, + { + "bbox": [ + 130, + 350, + 482, + 554 + ], + "type": "inline_equation", + "content": "12 \\times" + }, + { + "bbox": [ + 130, + 350, + 482, + 554 + ], + "type": "text", + "content": " less memory compared to the state-of-the-art parallel fusion method SOLOFusion [33]. However, we observe that recurrent fusion fails to exploit the long-term information as well as parallel fusion. As depicted in Fig.1a, VideoBEV benefits less (" + }, + { + "bbox": [ + 130, + 350, + 482, + 554 + ], + "type": "inline_equation", + "content": "\\Delta" + }, + { + "bbox": [ + 130, + 350, + 482, + 554 + ], + "type": "text", + "content": " mAP↑) from utilizing historical frames than that of SOLOFusion (" + }, + { + "bbox": [ + 130, + 350, + 482, + 554 + ], + "type": "inline_equation", + "content": "\\Delta" + }, + { + "bbox": [ + 130, + 350, + 482, + 554 + ], + "type": "text", + "content": " mAP↑). This phenomenon becomes even more pronounced, when utilizing more than four historical frames." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "text", + "content": "Why is the performance of recurrent fusion not as good as parallel fusion? We conducted a detailed analysis of the differences between the two paradigms and found two reasons: (1) Gradient propagation. During the training process, only the recent frames contribute to gradients of fusion module in the recurrent paradigm, while all historical frames contribute equally in the parallel paradigm. This prevents the temporal fusion module of recurrent paradigm from learning fusion pattern from long-term temporal information of all historical frames. (2) Semantic ambiguity. Due to the coarse granularity of the BEV grids, the recurrent paradigm generates errors when aligning historical features multiple" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 232, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 232, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 232, + 102 + ], + "type": "text", + "content": "M. Chang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 480, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 480, + 163 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 480, + 163 + ], + "type": "text", + "content": "times, resulting in semantic ambiguity for the same object. As a result, recurrent paradigm may fuse different objects on the singular BEV grid, confusing the recognition of the same object. For the above two reasons, the previous recurrent method cannot fully and correctly utilize the long-term temporal information." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 164, + 481, + 282 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 164, + 481, + 282 + ], + "spans": [ + { + "bbox": [ + 130, + 164, + 481, + 282 + ], + "type": "text", + "content": "Motivated by the above analysis, in this paper, we propose RecurrentBEV, a novel temporal fusion framework for BEV-based 3D object detector to improve the long-term fusion ability of recurrent paradigm while still has efficient latency and memory consumption in inference. First, RecurrentBEV adopts an RNN-style back-propagation during training, enabling historical frames to contribute gradients to the fusion module directly. This improvement solves the first problem while still maintaining inference efficiency. Second, we design an inner grid transformation before fusion, which can learn fine-grained semantic changes from ego-motion matrix. This effectively reduces semantic differences in the inter-frame alignment process." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 284, + 482, + 379 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 284, + 482, + 379 + ], + "spans": [ + { + "bbox": [ + 130, + 284, + 482, + 379 + ], + "type": "text", + "content": "Based on the RNN-style back-propagation and the inner grid transformation, the proposed RecurrentBEV effectively improves long-term fusion in the recurrent paradigm, as shown in Fig.1a. Extensive experiments on the nuScenes 3D detection dataset evaluate the effectiveness of the proposed RecurrentBEV. As shown in Fig.1b, it outperforms both existing BEV-based algorithms such as SOLOFusion, VideoBEV and query-based algorithms such as StreamPETR [35], SparseBEV [27], Sparse4Dv2 [26] in both accuracy and efficiency. To summarize, our contributions are:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 387, + 480, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 387, + 480, + 422 + ], + "spans": [ + { + "bbox": [ + 138, + 387, + 480, + 422 + ], + "type": "text", + "content": "- We provide an in-depth analysis for two temporal fusion paradigms and point out two factors that render the recurrent fusion less effective: issues with gradient propagation and semantic ambiguity." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 423, + 481, + 493 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 423, + 481, + 493 + ], + "spans": [ + { + "bbox": [ + 138, + 423, + 481, + 493 + ], + "type": "text", + "content": "- We propose RecurrentBEV, a novel temporal fusion framework for BEV-based 3D object detector. It adapots RNN-style back-propagation and inner grid transformation, attaining state-of-the-art results (57.4% mAP and 65.1% NDS) on the nuScenes benchmark. It represents the first recurrent fusion method with explicit BEV features that surpasses parallel fusion performance." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 513, + 243, + 525 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 513, + 243, + 525 + ], + "spans": [ + { + "bbox": [ + 132, + 513, + 243, + 525 + ], + "type": "text", + "content": "2 Related Works" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 537, + 324, + 550 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 537, + 324, + 550 + ], + "spans": [ + { + "bbox": [ + 131, + 537, + 324, + 550 + ], + "type": "text", + "content": "2.1 Multi-view 3D Object Detection" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 557, + 480, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 557, + 480, + 617 + ], + "spans": [ + { + "bbox": [ + 130, + 557, + 480, + 617 + ], + "type": "text", + "content": "Multi-view 3D object detection is a fundamental task for autonomous driving. Previous works mainly focus on how to transform the 2D features into a unified 3D space. This transformation can be divided to two categories, BEV-based methods [10, 13, 18, 22, 23, 41, 42] and query-based methods [25, 28, 36, 37], depending on whether an explicit 3D representation is employed." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "type": "text", + "content": "BEVDet [13] is one typical method of BEV-based multi-view 3D object detection. This method involves converting 2D features to BEV features based on the Lift-Splat-Shoot [34] paradigm. BEVDepth [21] is a further example, which notably refines the transformation process via explicit depth supervision." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 386, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 386, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 386, + 91, + 447, + 101 + ], + "type": "text", + "content": "RecurrentBEV" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 139 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 139 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 139 + ], + "type": "text", + "content": "BEVFormer [22] further introduces a novel attention mechanism to gather 2D features for BEV grids." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 140, + 482, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 140, + 482, + 213 + ], + "spans": [ + { + "bbox": [ + 130, + 140, + 482, + 213 + ], + "type": "text", + "content": "Query-based methods typically define several object queries in 3D space to gather features from multi-view image features. DETR3D [36] projects object queries to image coordinates to sample features then predicts 3D bounding boxes directly. DETR4D [31] and Sparse4D [25] enhance it by leveraging historical information. PETR [28] encodes the position information of 3D coordinates into image feature, facilitating the detection of 3D objects similar to DETR [3]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 230, + 375, + 241 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 230, + 375, + 241 + ], + "spans": [ + { + "bbox": [ + 131, + 230, + 375, + 241 + ], + "type": "text", + "content": "2.2 Temporal Fusion for BEV-based Detectors" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 249, + 482, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 249, + 482, + 430 + ], + "spans": [ + { + "bbox": [ + 130, + 249, + 482, + 430 + ], + "type": "text", + "content": "Recent works find that incorporating temporal information can enhance the stability in perception results and improve overall performance. BEVFormer [22] first proposes the temporal self-attention mechanism to fuse the historical BEV features in a recurrent paradigm. BEVDet4D [11] extends BEVDet [13] to a temporal version by employing a concatenation operation to integrate previous BEV features, usually called parallel paradigm. SOLOFusion [33] extends these methods by increasing the number of previous frames to 16, resulting in significant benefits. BEVFormerv2 [39] switches to parallel paradigm same as SOLOFusion. However, the inference latency and memory consumption of the parallel paradigm increase linearly as the historical frame lengths. VideoBEV [8] applies recurrent fusion to the LSS-based 3D detector, resulting in significant reduction of inference latency and memory consumption. However, the improvement of long-term fusion in VideoBEV is limited in comparison to SOLOFusion. In order to combine the advantages of the two paradigms, we construct a novel temporal fusion framework based on efficient recurrent paradigm." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 447, + 202, + 459 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 447, + 202, + 459 + ], + "spans": [ + { + "bbox": [ + 131, + 447, + 202, + 459 + ], + "type": "text", + "content": "3 Method" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 472, + 482, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 472, + 482, + 556 + ], + "spans": [ + { + "bbox": [ + 130, + 472, + 482, + 556 + ], + "type": "text", + "content": "Converting multi-camera image features to bird's-eye-view (BEV) features can provide a unified surrounding environment representation for autonomous driving perception tasks. It involves transforming image data into a top-down representation of the environment, which often denoted as BEV features. The BEV detectors contain an image encoder, a 2D-3D view transformation, and a temporal fusion module. In this work, we present a new temporal fusion module, which can effectively aggregate spatio-temporal BEV features." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 574, + 421, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 574, + 421, + 586 + ], + "spans": [ + { + "bbox": [ + 131, + 574, + 421, + 586 + ], + "type": "text", + "content": "3.1 Analysis of BEV-based Temporal Fusion Paradigms" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": "There are two kinds of temporal fusion paradigms in BEV detection, which are the effective parallel fusion and efficient recurrent fusion. The parallel fusion caches and updates multiple historical BEV features in learning the visual features for current frames. The recurrent fusion encodes all historical features into a single recurrent feature to cache and update. We conduct a detailed analysis to identify the disparities among them and pinpoint two crucial factors that cause" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 232, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 232, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 232, + 102 + ], + "type": "text", + "content": "M. Chang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 136, + 112, + 299, + 280 + ], + "blocks": [ + { + "bbox": [ + 136, + 112, + 299, + 280 + ], + "lines": [ + { + "bbox": [ + 136, + 112, + 299, + 280 + ], + "spans": [ + { + "bbox": [ + 136, + 112, + 299, + 280 + ], + "type": "image", + "image_path": "327b47b23a4df406692a42d40688da02e529a5cc08c09b32bb2e34625d40be16.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 211, + 282, + 223, + 293 + ], + "lines": [ + { + "bbox": [ + 211, + 282, + 223, + 293 + ], + "spans": [ + { + "bbox": [ + 211, + 282, + 223, + 293 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 302, + 113, + 476, + 283 + ], + "blocks": [ + { + "bbox": [ + 302, + 113, + 476, + 283 + ], + "lines": [ + { + "bbox": [ + 302, + 113, + 476, + 283 + ], + "spans": [ + { + "bbox": [ + 302, + 113, + 476, + 283 + ], + "type": "image", + "image_path": "4396a3b608b529bd6704b4f031f15f2427c34399b933dd6f36f98335dffae64b.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 380, + 283, + 392, + 293 + ], + "lines": [ + { + "bbox": [ + 380, + 283, + 392, + 293 + ], + "spans": [ + { + "bbox": [ + 380, + 283, + 392, + 293 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 300, + 482, + 389 + ], + "lines": [ + { + "bbox": [ + 130, + 300, + 482, + 389 + ], + "spans": [ + { + "bbox": [ + 130, + 300, + 482, + 389 + ], + "type": "text", + "content": "Fig. 2: (a) Semantic ambiguity caused by grid sample operation. When ego vehicle turning right, BEV feature change from " + }, + { + "bbox": [ + 130, + 300, + 482, + 389 + ], + "type": "inline_equation", + "content": "B_{t-1}" + }, + { + "bbox": [ + 130, + 300, + 482, + 389 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 130, + 300, + 482, + 389 + ], + "type": "inline_equation", + "content": "B_t" + }, + { + "bbox": [ + 130, + 300, + 482, + 389 + ], + "type": "text", + "content": ". In fusion, the grid sample operation used for aligning " + }, + { + "bbox": [ + 130, + 300, + 482, + 389 + ], + "type": "inline_equation", + "content": "B_{t-1}" + }, + { + "bbox": [ + 130, + 300, + 482, + 389 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 300, + 482, + 389 + ], + "type": "inline_equation", + "content": "B_t" + }, + { + "bbox": [ + 130, + 300, + 482, + 389 + ], + "type": "text", + "content": " only aligns the position but ignores the orientation as shown in the blue and red dash lines. (b) Visualization of the semantic ambiguity. The green boxes represent the BEV feature change from " + }, + { + "bbox": [ + 130, + 300, + 482, + 389 + ], + "type": "inline_equation", + "content": "B_{t-1}" + }, + { + "bbox": [ + 130, + 300, + 482, + 389 + ], + "type": "text", + "content": ", which have been transformed to align with the " + }, + { + "bbox": [ + 130, + 300, + 482, + 389 + ], + "type": "inline_equation", + "content": "B_t" + }, + { + "bbox": [ + 130, + 300, + 482, + 389 + ], + "type": "text", + "content": " and decoded into bounding boxes. The blue boxes depict the actual bounding boxes in the " + }, + { + "bbox": [ + 130, + 300, + 482, + 389 + ], + "type": "inline_equation", + "content": "B_t" + }, + { + "bbox": [ + 130, + 300, + 482, + 389 + ], + "type": "text", + "content": ". Notably, for static objects, there is a mismatch in the orientation of the green and blue boxes." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 413, + 480, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 413, + 480, + 437 + ], + "spans": [ + { + "bbox": [ + 130, + 413, + 480, + 437 + ], + "type": "text", + "content": "the sub-optimal performance of recurrent fusion: issues with gradient propagation and semantic ambiguity." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 455, + 482, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 455, + 482, + 491 + ], + "spans": [ + { + "bbox": [ + 130, + 455, + 482, + 491 + ], + "type": "text", + "content": "Gradient propagation Aligning historical frames to current frame is essential for BEV-based temporal fusion. Given the previous BEV feature " + }, + { + "bbox": [ + 130, + 455, + 482, + 491 + ], + "type": "inline_equation", + "content": "B_{j}" + }, + { + "bbox": [ + 130, + 455, + 482, + 491 + ], + "type": "text", + "content": ", this process can be written as:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 267, + 498, + 481, + 514 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 267, + 498, + 481, + 514 + ], + "spans": [ + { + "bbox": [ + 267, + 498, + 481, + 514 + ], + "type": "interline_equation", + "content": "\\widetilde {B} _ {j, i} = \\zeta \\left(B _ {j}, P _ {j, i}\\right) \\tag {1}", + "image_path": "ea04c5d8715a9532d92408a43cd827b7e26f28adace14e84fce5301be9f57fd0.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 523, + 482, + 573 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 523, + 482, + 573 + ], + "spans": [ + { + "bbox": [ + 130, + 523, + 482, + 573 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 523, + 482, + 573 + ], + "type": "inline_equation", + "content": "\\widetilde{B}_{j,i}" + }, + { + "bbox": [ + 130, + 523, + 482, + 573 + ], + "type": "text", + "content": " is the " + }, + { + "bbox": [ + 130, + 523, + 482, + 573 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 130, + 523, + 482, + 573 + ], + "type": "text", + "content": "-th frame BEV feature in the ego coordinate of " + }, + { + "bbox": [ + 130, + 523, + 482, + 573 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 523, + 482, + 573 + ], + "type": "text", + "content": "-th frame, " + }, + { + "bbox": [ + 130, + 523, + 482, + 573 + ], + "type": "inline_equation", + "content": "P_{j,i}" + }, + { + "bbox": [ + 130, + 523, + 482, + 573 + ], + "type": "text", + "content": " is the ego-motion matrix between the " + }, + { + "bbox": [ + 130, + 523, + 482, + 573 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 130, + 523, + 482, + 573 + ], + "type": "text", + "content": "-th frame and " + }, + { + "bbox": [ + 130, + 523, + 482, + 573 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 523, + 482, + 573 + ], + "type": "text", + "content": "-th frame and " + }, + { + "bbox": [ + 130, + 523, + 482, + 573 + ], + "type": "inline_equation", + "content": "\\zeta" + }, + { + "bbox": [ + 130, + 523, + 482, + 573 + ], + "type": "text", + "content": " refers to the ego-motion transformation function, specifically grid sample operation [15]. Based on Eq.(1), We can express the parallel fusion of the " + }, + { + "bbox": [ + 130, + 523, + 482, + 573 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 523, + 482, + 573 + ], + "type": "text", + "content": "-th frame as:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 232, + 582, + 480, + 597 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 232, + 582, + 480, + 597 + ], + "spans": [ + { + "bbox": [ + 232, + 582, + 480, + 597 + ], + "type": "interline_equation", + "content": "\\bar {B} _ {i} = \\varphi \\left(\\widetilde {B} _ {i - k, i}, \\dots , \\widetilde {B} _ {i - 1, i}, B _ {i}, W\\right) \\tag {2}", + "image_path": "7750d9270342a14520e4fbff7e3292c8d98215f4cb84f1730595b11e5b1c3f47.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 605, + 482, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 652 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 652 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 605, + 482, + 652 + ], + "type": "inline_equation", + "content": "\\overline{B}_i" + }, + { + "bbox": [ + 130, + 605, + 482, + 652 + ], + "type": "text", + "content": " is the fused BEV feature of " + }, + { + "bbox": [ + 130, + 605, + 482, + 652 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 605, + 482, + 652 + ], + "type": "text", + "content": "-th frame, " + }, + { + "bbox": [ + 130, + 605, + 482, + 652 + ], + "type": "inline_equation", + "content": "\\varphi" + }, + { + "bbox": [ + 130, + 605, + 482, + 652 + ], + "type": "text", + "content": " is the temporal fusion function (convolution [11,21,33,39] or deformable attention [22]) and " + }, + { + "bbox": [ + 130, + 605, + 482, + 652 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 130, + 605, + 482, + 652 + ], + "type": "text", + "content": " refers the parameters of it. Similarly, the recurrent paradigm for the " + }, + { + "bbox": [ + 130, + 605, + 482, + 652 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 605, + 482, + 652 + ], + "type": "text", + "content": "-th frame can be written as:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 238, + 652, + 480, + 666 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 238, + 652, + 480, + 666 + ], + "spans": [ + { + "bbox": [ + 238, + 652, + 480, + 666 + ], + "type": "interline_equation", + "content": "\\bar {B} _ {i} = \\varphi (\\zeta (\\bar {B} _ {i - 1}, P _ {i - 1, i}), B _ {i}, W) \\tag {3}", + "image_path": "88cdbc00d6bad435e48ed4fde53531a1b3cae87580386ff60ecda385935a46e1.jpg" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 386, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 386, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 386, + 91, + 447, + 100 + ], + "type": "text", + "content": "RecurrentBEV" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 99 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 99 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 99 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 147, + 114, + 465, + 248 + ], + "blocks": [ + { + "bbox": [ + 147, + 114, + 465, + 248 + ], + "lines": [ + { + "bbox": [ + 147, + 114, + 465, + 248 + ], + "spans": [ + { + "bbox": [ + 147, + 114, + 465, + 248 + ], + "type": "image", + "image_path": "2164b740a825f3d9c2171889420426ab1448c78963eccbb54182acc39e379e06.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 258, + 482, + 314 + ], + "lines": [ + { + "bbox": [ + 130, + 258, + 482, + 314 + ], + "spans": [ + { + "bbox": [ + 130, + 258, + 482, + 314 + ], + "type": "text", + "content": "Fig. 3: Overall architecture of the proposed RecurrentBEV. It shows a training process in the " + }, + { + "bbox": [ + 130, + 258, + 482, + 314 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 258, + 482, + 314 + ], + "type": "text", + "content": "-th frame, while arrows indicate forward propagation and backward propagation. The red arrows indicate the additional back-propagation introduced by our method. The details of the Transformation-aware Gate Recurrent Unit (TA-GRU) is illustrated in dotted box." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 336, + 482, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 336, + 482, + 361 + ], + "spans": [ + { + "bbox": [ + 130, + 336, + 482, + 361 + ], + "type": "text", + "content": "Based on Eq.(3), we can calculate the gradient for the parameters in the temporal fusion function for parallel fusion." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 221, + 365, + 481, + 392 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 221, + 365, + 481, + 392 + ], + "spans": [ + { + "bbox": [ + 221, + 365, + 481, + 392 + ], + "type": "interline_equation", + "content": "\\frac {\\partial L}{\\partial W} = \\frac {\\partial L}{\\partial \\bar {B} _ {i}} \\cdot \\varphi^ {\\prime} \\left(\\widetilde {B} _ {i - k, i}, \\dots , \\widetilde {B} _ {i - 1, i}, B _ {i}\\right) \\tag {4}", + "image_path": "8166e54cd40fab3b6d1a86069425547eef5ef1a9f1267f86a43356cffa440ae4.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 397, + 482, + 421 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 397, + 482, + 421 + ], + "spans": [ + { + "bbox": [ + 130, + 397, + 482, + 421 + ], + "type": "text", + "content": "In recurrent fusion, " + }, + { + "bbox": [ + 130, + 397, + 482, + 421 + ], + "type": "inline_equation", + "content": "B_{i-1}" + }, + { + "bbox": [ + 130, + 397, + 482, + 421 + ], + "type": "text", + "content": " is detached so that the gradient of " + }, + { + "bbox": [ + 130, + 397, + 482, + 421 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 130, + 397, + 482, + 421 + ], + "type": "text", + "content": " in recurrent fusion can be written as:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 227, + 425, + 481, + 452 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 227, + 425, + 481, + 452 + ], + "spans": [ + { + "bbox": [ + 227, + 425, + 481, + 452 + ], + "type": "interline_equation", + "content": "\\frac {\\partial L}{\\partial W} = \\frac {\\partial L}{\\partial \\bar {B} _ {i}} \\cdot \\varphi^ {\\prime} (\\zeta (\\bar {B} _ {i - 1}, P _ {i - 1, i}), B _ {i}) \\tag {5}", + "image_path": "9b92f6121a5ea5a3fd43f380349ed24dc016238f8a4d00fdfa7d655449d0b5f5.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 457, + 482, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 457, + 482, + 518 + ], + "spans": [ + { + "bbox": [ + 130, + 457, + 482, + 518 + ], + "type": "text", + "content": "Comparing Eq.(4) and Eq.(5), it is evident that only the current BEV feature " + }, + { + "bbox": [ + 130, + 457, + 482, + 518 + ], + "type": "inline_equation", + "content": "B_{i}" + }, + { + "bbox": [ + 130, + 457, + 482, + 518 + ], + "type": "text", + "content": " and one previous BEV feature " + }, + { + "bbox": [ + 130, + 457, + 482, + 518 + ], + "type": "inline_equation", + "content": "\\overline{B}_{i-1}" + }, + { + "bbox": [ + 130, + 457, + 482, + 518 + ], + "type": "text", + "content": " affect the gradient of " + }, + { + "bbox": [ + 130, + 457, + 482, + 518 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 130, + 457, + 482, + 518 + ], + "type": "text", + "content": " directly in recurrent paradigm while all previous BEV features within a window size " + }, + { + "bbox": [ + 130, + 457, + 482, + 518 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 457, + 482, + 518 + ], + "type": "text", + "content": " contribute to gradient of " + }, + { + "bbox": [ + 130, + 457, + 482, + 518 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 130, + 457, + 482, + 518 + ], + "type": "text", + "content": " in parallel paradigm. This is one reason that recurrent paradigm makes limited progress in long-term fusion." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 533, + 482, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 482, + 582 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 482, + 582 + ], + "type": "text", + "content": "Semantic Ambiguity To optimize memory usage, the recurrent fusion approach only caches a single historical 3D BEV feature instead of storing a sequence of 2D features. Consequently, this approach encounters challenges with temporal alignment in the 3D space." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "content": "In the current recurrent fusion method, we have observed semantic ambiguity issues arising during the temporal alignment process in the 3D space. As illustrated in Fig.2a, the coarse granularity of BEV grids results in inaccurate transformation of orientation information within each grid during the temporal alignment process represented by Eq.(1). Consequently, this process introduces incorrect object orientations in real-world detection scenarios, as evident in Fig.2b." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 232, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 232, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 232, + 102 + ], + "type": "text", + "content": "M. Chang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 131, + 116, + 482, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 116, + 482, + 175 + ], + "spans": [ + { + "bbox": [ + 131, + 116, + 482, + 175 + ], + "type": "text", + "content": "In the parallel fusion counterpart, temporal alignment in the 3D space is avoided by directly projecting a sequence of 2D view feature onto the current 3D BEV feature space [33, 39]. Assuming that the past perspective 2D view feature at time step " + }, + { + "bbox": [ + 131, + 116, + 482, + 175 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 131, + 116, + 482, + 175 + ], + "type": "text", + "content": " is denoted as " + }, + { + "bbox": [ + 131, + 116, + 482, + 175 + ], + "type": "inline_equation", + "content": "C_j" + }, + { + "bbox": [ + 131, + 116, + 482, + 175 + ], + "type": "text", + "content": ", the formulation for projecting it onto the BEV feature at time step " + }, + { + "bbox": [ + 131, + 116, + 482, + 175 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 131, + 116, + 482, + 175 + ], + "type": "text", + "content": " is as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 231, + 184, + 481, + 198 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 231, + 184, + 481, + 198 + ], + "spans": [ + { + "bbox": [ + 231, + 184, + 481, + 198 + ], + "type": "interline_equation", + "content": "\\widetilde {B} _ {j, i} = \\operatorname {V i e w T r a n s} \\left(C _ {j}, I _ {j}, P _ {j, i} \\cdot E _ {j}\\right) \\tag {6}", + "image_path": "e937c214294746585a12e3f8638c56e156cfa963f791b74c4354ca1ef686346c.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 205, + 482, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 205, + 482, + 277 + ], + "spans": [ + { + "bbox": [ + 131, + 205, + 482, + 277 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 205, + 482, + 277 + ], + "type": "inline_equation", + "content": "I_{j}" + }, + { + "bbox": [ + 131, + 205, + 482, + 277 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 205, + 482, + 277 + ], + "type": "inline_equation", + "content": "E_{j}" + }, + { + "bbox": [ + 131, + 205, + 482, + 277 + ], + "type": "text", + "content": " denote camera intrinsic and extrinsic respectively. This additional view transformation in 2D space effectively aligns objects between frames, subsequently enhancing the performance of parallel fusion. Nevertheless, this approach is not feasible for recurrent fusion due to the unavailability of 2D features during the recurrent fusion process. This discrepancy serves as another contributing factor to the performance gap between the two fusion methods." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 293, + 236, + 306 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 293, + 236, + 306 + ], + "spans": [ + { + "bbox": [ + 132, + 293, + 236, + 306 + ], + "type": "text", + "content": "3.2 RecurrentBEV" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 313, + 482, + 397 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 313, + 482, + 397 + ], + "spans": [ + { + "bbox": [ + 130, + 313, + 482, + 397 + ], + "type": "text", + "content": "Based on the two above-mentioned possible reasons, we design a novel temporal fusion framework called RecurrentBEV. Similar to existing BEV-based 3D object detectors [8,11,21,33], RecurrentBEV contains an image encoder, a 2D-3D view transformation, a temporal fusion module, a BEV encoder and a task head, as illustrated in Fig.3. The main difference between RecurrentBEV and previous recurrent fusion is the temporal fusion module which can learn from all historical frames directly and learn fine-grained ego-motion transform." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 413, + 482, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 413, + 482, + 510 + ], + "spans": [ + { + "bbox": [ + 131, + 413, + 482, + 510 + ], + "type": "text", + "content": "RNN-style Back-propagation In prior recurrent fusion methods [22], the " + }, + { + "bbox": [ + 131, + 413, + 482, + 510 + ], + "type": "inline_equation", + "content": "\\overline{B}_{i-1}" + }, + { + "bbox": [ + 131, + 413, + 482, + 510 + ], + "type": "text", + "content": " in Eq.(3) is solely computed for forward propagation, but not for backward propagation. Thus, historical BEV features do not contribute to the gradient of temporal fusion module in Eq.(5), preventing optimizing the temporal fusion module with all the relevant historical BEV features. To address this issue, RecurrentBEV adopts RNN-style Back-propagation that all the relevant historical features back-propagate gradients to adjacent layers. In this context, the gradients of parameters in temporal fusion function is written as:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 215, + 517, + 481, + 552 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 215, + 517, + 481, + 552 + ], + "spans": [ + { + "bbox": [ + 215, + 517, + 481, + 552 + ], + "type": "interline_equation", + "content": "\\frac {\\partial L}{\\partial W} = \\sum_ {j = i - k} ^ {i} \\frac {\\partial L}{\\partial \\overline {{B}} _ {j}} \\cdot \\varphi^ {'} (\\zeta (\\overline {{B}} _ {j - 1}, P _ {j - 1, j}), B _ {j}) \\tag {7}", + "image_path": "87ae80c6aa3467e49f8835b3d923402fe0733ea26c91cc1738f7f281d54ac4a6.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 558, + 481, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 558, + 481, + 594 + ], + "spans": [ + { + "bbox": [ + 130, + 558, + 481, + 594 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 558, + 481, + 594 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 558, + 481, + 594 + ], + "type": "text", + "content": " is the number of total historical frames using in a iteration of training process. As a result, each historical BEV feature makes an equal contribution to the gradient of the temporal fusion module, akin to the parallel paradigm." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 594, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 594, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 131, + 594, + 482, + 666 + ], + "type": "text", + "content": "Computing the back-propagation along the temporal dimension for the entire network places significant demands on GPU memory. As shown in Table 1 (Setting 1), if all modules, including image encoders, undergo back-propagation along the temporal dimension, memory overflow issues arise even when the frame number reaches only 2. However, it is imperative for the training process to incorporate long-term frames to effectively learn temporal information (refer to" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 386, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 386, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 386, + 91, + 447, + 101 + ], + "type": "text", + "content": "RecurrentBEV" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 187, + 114, + 427, + 159 + ], + "blocks": [ + { + "bbox": [ + 187, + 114, + 427, + 159 + ], + "lines": [ + { + "bbox": [ + 187, + 114, + 427, + 159 + ], + "spans": [ + { + "bbox": [ + 187, + 114, + 427, + 159 + ], + "type": "table", + "html": "
Training frames0124816
Setting 115,06730,105OOMOOMOOMOOM
Setting 215,06717,66118,73519,97523,57930,941
", + "image_path": "a64676f109229dfd434eb1766195e494153d04208e13c200eee795cc5f70e197.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 159, + 481, + 180 + ], + "lines": [ + { + "bbox": [ + 132, + 159, + 481, + 180 + ], + "spans": [ + { + "bbox": [ + 132, + 159, + 481, + 180 + ], + "type": "text", + "content": "Table 1: The memory usage (MB) during training. OOM means out of memory on A100-40G with FP32 precision." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 130, + 213, + 482, + 308 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 213, + 482, + 308 + ], + "spans": [ + { + "bbox": [ + 130, + 213, + 482, + 308 + ], + "type": "text", + "content": "Sec.4.6 for details). Therefore, we need to strike a balance between GPU memory usage (determined by the modules involved in back-propagation) and the length of the time sequence. In our design, as depicted by the red arrows in Figure 3, RecurrentBEV further computes all the gradients of the temporal fusion module (TA-GRU in the following) via back-propagation along the temporal dimension. Consequently, RecurrentBEV (Setting 2 in Table 1) can be trained with a batch size of 8 with FP32 precision on a single A100-40G GPU, effectively utilizing up to 16 consecutive frames." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 324, + 482, + 385 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 324, + 482, + 385 + ], + "spans": [ + { + "bbox": [ + 130, + 324, + 482, + 385 + ], + "type": "text", + "content": "Transformation-aware Gated Recurrent Unit Because of the coarse granularity of the BEV grids, the semantic feature in a grid can not project correctly when model transforms the previous BEV features to the coordinate system of the current ego. We design an inner grid transformation to solve this problem. For simplicity, we use a linear layer to model this transformation:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 241, + 391, + 481, + 406 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 241, + 391, + 481, + 406 + ], + "spans": [ + { + "bbox": [ + 241, + 391, + 481, + 406 + ], + "type": "interline_equation", + "content": "\\widetilde {B} _ {j, i} = \\operatorname {L i n e a r} \\left(\\zeta \\left(B _ {j}, P _ {j, i}\\right), W _ {p}\\right) \\tag {8}", + "image_path": "b4d2dcd9a1fc225b951a2111f0e8d030ca5de38ea8e7c662da74d4515fe3dcf4.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 411, + 482, + 446 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 411, + 482, + 446 + ], + "spans": [ + { + "bbox": [ + 130, + 411, + 482, + 446 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 411, + 482, + 446 + ], + "type": "inline_equation", + "content": "W_{p}" + }, + { + "bbox": [ + 130, + 411, + 482, + 446 + ], + "type": "text", + "content": " is the parameter of Linear. For simplicity, we assume different channels in a projected feature share same motion embedding, so that the " + }, + { + "bbox": [ + 130, + 411, + 482, + 446 + ], + "type": "inline_equation", + "content": "W_{p}" + }, + { + "bbox": [ + 130, + 411, + 482, + 446 + ], + "type": "text", + "content": " can be decoupled as:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 272, + 446, + 481, + 460 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 272, + 446, + 481, + 460 + ], + "spans": [ + { + "bbox": [ + 272, + 446, + 481, + 460 + ], + "type": "interline_equation", + "content": "W _ {p} = E _ {j, i} \\odot \\widetilde {W} _ {p} \\tag {9}", + "image_path": "931fcbdb47c82abdd0271ee99d46ea0965f62311a0d1f00f09894ec715fb120b.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 464, + 482, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 464, + 482, + 514 + ], + "spans": [ + { + "bbox": [ + 130, + 464, + 482, + 514 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 464, + 482, + 514 + ], + "type": "inline_equation", + "content": "E_{j,i} \\in \\mathbb{R}^{C_{in} \\times 1}" + }, + { + "bbox": [ + 130, + 464, + 482, + 514 + ], + "type": "text", + "content": " is ego motion embedding from timestamp " + }, + { + "bbox": [ + 130, + 464, + 482, + 514 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 130, + 464, + 482, + 514 + ], + "type": "text", + "content": " to timestamp " + }, + { + "bbox": [ + 130, + 464, + 482, + 514 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 464, + 482, + 514 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 464, + 482, + 514 + ], + "type": "inline_equation", + "content": "\\widetilde{W}_p \\in \\mathbb{R}^{C_{in} \\times C_{out}}" + }, + { + "bbox": [ + 130, + 464, + 482, + 514 + ], + "type": "text", + "content": " is a learnable weight and " + }, + { + "bbox": [ + 130, + 464, + 482, + 514 + ], + "type": "inline_equation", + "content": "\\odot" + }, + { + "bbox": [ + 130, + 464, + 482, + 514 + ], + "type": "text", + "content": " denotes a Hadamard Product. Motion embedding " + }, + { + "bbox": [ + 130, + 464, + 482, + 514 + ], + "type": "inline_equation", + "content": "E_{j,i}" + }, + { + "bbox": [ + 130, + 464, + 482, + 514 + ], + "type": "text", + "content": " is a high dimensional vector converted from the given ego motion matrix " + }, + { + "bbox": [ + 130, + 464, + 482, + 514 + ], + "type": "inline_equation", + "content": "P_{j,i}" + }, + { + "bbox": [ + 130, + 464, + 482, + 514 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 270, + 514, + 481, + 528 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 270, + 514, + 481, + 528 + ], + "spans": [ + { + "bbox": [ + 270, + 514, + 481, + 528 + ], + "type": "interline_equation", + "content": "E _ {j, i} = \\sigma (\\psi (P _ {j, i})) \\tag {10}", + "image_path": "c919fbb81dfa88a5b8d7be82c2846642ad63a9925f04cd63eaca023323213135.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 530, + 482, + 566 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 530, + 482, + 566 + ], + "spans": [ + { + "bbox": [ + 130, + 530, + 482, + 566 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 530, + 482, + 566 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 130, + 530, + 482, + 566 + ], + "type": "text", + "content": " is a sigmoid activation and " + }, + { + "bbox": [ + 130, + 530, + 482, + 566 + ], + "type": "inline_equation", + "content": "\\psi" + }, + { + "bbox": [ + 130, + 530, + 482, + 566 + ], + "type": "text", + "content": " is a multi-layer perception (MLP). Combining Eq.(8), Eq.(9) and Eq.(10), the complete inner grid transformation can be written as:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 212, + 584, + 481, + 600 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 212, + 584, + 481, + 600 + ], + "spans": [ + { + "bbox": [ + 212, + 584, + 481, + 600 + ], + "type": "interline_equation", + "content": "\\widetilde {B} _ {j, i} = \\operatorname {L i n e a r} \\left(\\zeta \\left(\\bar {B} _ {j}, P _ {j, i}\\right), \\sigma \\left(\\psi \\left(P _ {j, i}\\right)\\right) \\odot \\widetilde {W} _ {p}\\right) \\tag {11}", + "image_path": "3c28889e8dd7cc596952d9073825ead620b491e7e3d56f5668d4ed6c6588652e.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "type": "text", + "content": "For the RNN modules, we adopted the typical Convolutional Gated Recurrent Unit (ConvGRU) [1]. By combining the aforementioned inner grid transformation, we introduce a novel temporal fusion module named the Transformation-Aware Gated Recurrent Unit (TA-GRU). The entire process of temporal fusion is formulated as follows:" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 232, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 232, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 232, + 102 + ], + "type": "text", + "content": "M. Chang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 225, + 137, + 481, + 152 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 225, + 137, + 481, + 152 + ], + "spans": [ + { + "bbox": [ + 225, + 137, + 481, + 152 + ], + "type": "interline_equation", + "content": "W _ {p} = \\sigma \\left(\\psi \\left(P _ {i - 1, i}\\right)\\right) \\odot \\widetilde {W} _ {p} \\tag {12}", + "image_path": "b8197276022f85cd981ba2f09e127d41f6743a1d7a7d418e46ccd98084ce512e.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 212, + 154, + 481, + 169 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 212, + 154, + 481, + 169 + ], + "spans": [ + { + "bbox": [ + 212, + 154, + 481, + 169 + ], + "type": "interline_equation", + "content": "\\widetilde {B} _ {i - 1, i} = \\operatorname {L i n e a r} \\left(\\zeta \\left(\\bar {B} _ {i - 1}, P _ {i - 1, i}\\right), W _ {p}\\right) \\tag {13}", + "image_path": "27ca0f5b051f9c28b30bad5d32edee2a9421103353002636a55aa75994a2abd6.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 228, + 172, + 481, + 186 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 228, + 172, + 481, + 186 + ], + "spans": [ + { + "bbox": [ + 228, + 172, + 481, + 186 + ], + "type": "interline_equation", + "content": "Z _ {i} = \\sigma \\left(\\operatorname {C o n v} \\left(\\left[ \\widetilde {B} _ {i - 1, i}, B _ {i} \\right], W _ {z}\\right)\\right) \\tag {14}", + "image_path": "c13fb3b4ed686ee317b68d9ebf174813893bb1855f857842e8b4a2b9281f7337.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 227, + 189, + 481, + 203 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 227, + 189, + 481, + 203 + ], + "spans": [ + { + "bbox": [ + 227, + 189, + 481, + 203 + ], + "type": "interline_equation", + "content": "R _ {i} = \\sigma \\left(\\operatorname {C o n v} \\left(\\left[ \\widetilde {B} _ {i - 1, i}, B _ {i} \\right], W _ {r}\\right)\\right) \\tag {15}", + "image_path": "e6d23d2f2836de86b6b11656a96a8a2897bcd641a13a58e7083258bbde4ce273.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 227, + 206, + 481, + 221 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 227, + 206, + 481, + 221 + ], + "spans": [ + { + "bbox": [ + 227, + 206, + 481, + 221 + ], + "type": "interline_equation", + "content": "O _ {i} = \\tanh \\left(\\operatorname {C o n v} \\left(\\left[ R _ {i} \\odot \\widetilde {B} _ {i - 1, i}, B _ {i} \\right], W _ {o}\\right)\\right) \\tag {16}", + "image_path": "c0cc0d8766c51eedef7ae3752dd014a9f8660349255f8a97786031582bf6cbc1.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 227, + 224, + 481, + 238 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 227, + 224, + 481, + 238 + ], + "spans": [ + { + "bbox": [ + 227, + 224, + 481, + 238 + ], + "type": "interline_equation", + "content": "\\bar {B} _ {i} = \\left(1 - Z _ {i}\\right) \\odot \\tilde {B} _ {i - 1, i} + Z _ {i} \\odot O _ {i} \\tag {17}", + "image_path": "0673f9419a76ad3ca58ec7f9e7283cfe198c6dc21644437ae9530afb7f0e8df8.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 247, + 482, + 294 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 247, + 482, + 294 + ], + "spans": [ + { + "bbox": [ + 130, + 247, + 482, + 294 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 247, + 482, + 294 + ], + "type": "inline_equation", + "content": "[\\cdot, \\cdot]" + }, + { + "bbox": [ + 130, + 247, + 482, + 294 + ], + "type": "text", + "content": " represents the concatenation of feature maps and " + }, + { + "bbox": [ + 130, + 247, + 482, + 294 + ], + "type": "inline_equation", + "content": "W_{z}, W_{z}, W_{o}" + }, + { + "bbox": [ + 130, + 247, + 482, + 294 + ], + "type": "text", + "content": " are 2D-convolution kernels. Because of inner grid transformation, TA-GRU can learn fine-grained semantic changes from ego-motion matrix, thus removing wrong orientation." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 314, + 482, + 397 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 314, + 482, + 397 + ], + "spans": [ + { + "bbox": [ + 130, + 314, + 482, + 397 + ], + "type": "text", + "content": "Inference Efficiency Comparing Eq.(2) and Eq.(3), parallel fusion requires " + }, + { + "bbox": [ + 130, + 314, + 482, + 397 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 314, + 482, + 397 + ], + "type": "text", + "content": " historical frames to be cached while recurrent fusion requires only one. Additionally, the transformation process described in Eq.(1) or Eq.(6) also needs to be computed " + }, + { + "bbox": [ + 130, + 314, + 482, + 397 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 314, + 482, + 397 + ], + "type": "text", + "content": " times in parallel fusion, whereas in recurrent fusion it only needs to be computed once. As a result, recurrent fusion, which has lower inference latency and smaller memory consumption, is more suitable for deployment in the industry." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 398, + 482, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 398, + 482, + 434 + ], + "spans": [ + { + "bbox": [ + 130, + 398, + 482, + 434 + ], + "type": "text", + "content": "Similar to previous recurrent fusion methods [8, 22], RecurrentBEV adopts sequential inference which only needs to cache a single historical BEV feature " + }, + { + "bbox": [ + 130, + 398, + 482, + 434 + ], + "type": "inline_equation", + "content": "\\overline{B}" + }, + { + "bbox": [ + 130, + 398, + 482, + 434 + ], + "type": "text", + "content": ", initialized with zero and updated at the end of the fusion." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 454, + 230, + 468 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 454, + 230, + 468 + ], + "spans": [ + { + "bbox": [ + 132, + 454, + 230, + 468 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 480, + 264, + 491 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 480, + 264, + 491 + ], + "spans": [ + { + "bbox": [ + 132, + 480, + 264, + 491 + ], + "type": "text", + "content": "4.1 Dataset and Metrics" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 501, + 482, + 646 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 501, + 482, + 646 + ], + "spans": [ + { + "bbox": [ + 130, + 501, + 482, + 646 + ], + "type": "text", + "content": "We evaluate our method on the challenging nuScenes dataset [2]. The dataset comprises 1000 driving scenes, each lasting about 20 seconds, collected in Boston and Singapore, which are divided into 700, 150 and 150 scenes respectively for training, validation, and testing. The vehicle is equipped with 6 cameras for a complete 360-degree field of view. The nuScenes dataset annotates 23 object classes with accurate 3D bounding boxes for each scene at " + }, + { + "bbox": [ + 130, + 501, + 482, + 646 + ], + "type": "inline_equation", + "content": "2\\mathrm{Hz}" + }, + { + "bbox": [ + 130, + 501, + 482, + 646 + ], + "type": "text", + "content": ". For evaluation purposes, 10 classes are selected, namely: car, truck, bus, trailer, construction vehicle, pedestrian, motorcycle, bicycle, barrier, and traffic cone. As for evaluation, we use the following metrics, mean average precision (mAP), the nuScenes detection score (NDS), and five true positive metrics including average translation error (ATE), average scale error (ASE), average orientation error (AOE), average velocity error (AVE), average attribute error (AAE)." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 386, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 386, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 386, + 91, + 447, + 100 + ], + "type": "text", + "content": "RecurrentBEV" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 133, + 653, + 440, + 666 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 653, + 440, + 666 + ], + "spans": [ + { + "bbox": [ + 133, + 653, + 440, + 666 + ], + "type": "text", + "content": "4 details of evaluation metrics: https://www.nuscenes.org/object-detection" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 136, + 114, + 477, + 320 + ], + "blocks": [ + { + "bbox": [ + 136, + 114, + 477, + 320 + ], + "lines": [ + { + "bbox": [ + 136, + 114, + 477, + 320 + ], + "spans": [ + { + "bbox": [ + 136, + 114, + 477, + 320 + ], + "type": "table", + "html": "
MethodsBackboneImage SizeEDmAP↑NDS↑mATE↓mASE↓mAOE↓mAVE↓mAAE↓
BEVDet [13]ResNet50256 × 704X0.2980.3790.7250.2790.5890.8600.245
BEVDet4D [11]ResNet50256 × 704X0.3220.4570.7030.2780.4950.3540.206
BEVDepth [21]ResNet50256 × 704X0.3510.4750.6390.2670.4790.4280.198
BEVStereo [20]ResNet50256 × 704X0.3720.5000.5980.2700.4380.3670.190
BEVPoolv2 [12]ResNet50256 × 704X0.4060.5260.5720.2750.4630.2750.188
SOLOFusion [33]ResNet50256 × 704X0.4270.5340.5670.2740.5110.2520.181
VideoBEV [8]ResNet50256 × 704X0.4220.5350.5640.2760.4400.2860.198
PETRv2 [29]ResNet50256 × 704X0.3490.4560.7000.2750.5800.4370.187
Sparse4Dv2 [26]ResNet50256 × 704X0.4390.5390.5980.2700.4750.2820.179
SparseBEV [27]ResNet50256 × 704X0.4320.5450.6060.2740.3870.2510.186
StreamPETR [35]ResNet50256 × 704X0.4320.5400.5810.2720.4130.2950.195
RecurrentBEVResNet50256 × 704X0.4450.5490.5550.2720.4510.2560.204
BEVDepth [13]ResNet101512 × 1408X0.4120.5350.5650.2660.3580.3310.190
SOLOFusion [33]ResNet101512 × 1408X0.4830.5820.5030.2640.3810.2460.207
Sparse4Dv2 [26]ResNet101512 × 1408X0.4850.5800.5550.2720.3670.2560.182
RecurrentBEVResNet101512 × 1408X0.5090.5990.4870.2670.3580.2470.199
SparseBEV [27]ResNet101512 × 1408✓0.5010.5920.5620.2650.3210.2430.195
StreamPETR [35]ResNet101512 × 1408✓0.5040.5920.5690.2620.3150.2570.199
Sparse4Dv2 [26]ResNet101512 × 1408✓0.5050.5940.5480.2680.3480.2390.184
RecurrentBEVResNet101512 × 1408✓0.5280.6120.4810.2560.3550.2260.201
", + "image_path": "682d094a8c6516605fd042a08d954b5db7042d23f8e8bda63b629d8db9a0b3a5.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 321, + 480, + 353 + ], + "lines": [ + { + "bbox": [ + 132, + 321, + 480, + 353 + ], + "spans": [ + { + "bbox": [ + 132, + 321, + 480, + 353 + ], + "type": "text", + "content": "Table 2: Comparison results on 3D detection on the nuScenes validation set. ED is external data which indicates to use pre-trained weights from the external dataset i.e. nuImages." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 132, + 398, + 279, + 410 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 398, + 279, + 410 + ], + "spans": [ + { + "bbox": [ + 132, + 398, + 279, + 410 + ], + "type": "text", + "content": "4.2 Implementation Details" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 444, + 482, + 577 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 444, + 482, + 577 + ], + "spans": [ + { + "bbox": [ + 130, + 444, + 482, + 577 + ], + "type": "text", + "content": "We conduct experiments based on the BEVDet4D-Depth [11, 21] with the same learning rate, optimizer, and data augmentation. Drawing upon previous methods [21, 33, 35], we utilize ResNet50 [9], ResNet101, ConNeXt-B [30] and V2-99 [17] as backbones. On the nuScenes validation set, we report results of ResNet50 and ResNet101 models with pre-trained weights ImageNet [6] and nuImages [35]. On the nuScenes test set, we provide the performance of V2-99 model initialized from DD3D [32] and ConvNeXt-B model benefited from COCO [24] pre-training. The models in the ablation study are trained for 24 epochs without CBGS [43]. When compared with others, we follow the training setting of StreamPETR [35], extending the training epochs to 60 and using 2D auxiliary supervision for a fair comparison." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "content": "To achieve real-time inference speed, we optimize the BEVDet4D-Depth baseline by making several key modifications. Firstly, we reduce the number of depth bins. Secondly, we eliminate the ASPP [5] module from the DepthNet to minimize the latency associated with view transformation. Additionally, we merge 6 detection tasks performed by CenterHead [40] into a single task, significantly reducing post-processing time. These optimizations combined enable us to achieve a more efficient inference speed." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 231, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 231, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 231, + 102 + ], + "type": "text", + "content": "M. Chang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 138, + 114, + 475, + 352 + ], + "blocks": [ + { + "bbox": [ + 138, + 114, + 475, + 352 + ], + "lines": [ + { + "bbox": [ + 138, + 114, + 475, + 352 + ], + "spans": [ + { + "bbox": [ + 138, + 114, + 475, + 352 + ], + "type": "table", + "html": "
MethodsBackbonemAP↑NDS↑mATE↓mASE↓mAOE↓mAVE↓mAAE↓Conference
DETR3D* [36]V2-990.4120.4790.6410.2550.3940.8450.133CoRL2021
PETR [28]V2-990.4410.5040.5930.2490.3830.8080.132ECCV2022
UVTR [19]V2-990.4720.5510.5770.2530.3910.5080.123NIPS2022
BEVFormer [22]V2-990.4810.5690.5820.2560.3750.3780.126ECCV2022
PETRv2 [29]V2-990.4900.5820.5610.2430.3610.3430.120ICCV2023
PolarFormer [16]V2-990.4930.5720.5560.2560.3640.4390.127AAAI2023
BEVDepth [21]V2-990.5030.6000.4450.2450.3780.3200.126AAAI2023
VEDet [4]V2-990.5050.5850.5450.2440.3460.4210.123CVPR2023
MV2D [37]V2-990.5110.5960.5250.2430.3570.3570.120ICCV2023
CAPE [38]V2-990.5250.6100.5030.2420.3610.3060.114CVPR2023
BEVStereo [20]V2-990.5250.6100.4310.2460.3580.3570.138AAAI2023
SA-BEV [41]V2-990.5330.6240.4300.2410.3380.2820.139ICCV2023
FB-BEV [23]V2-990.5370.6240.4390.2500.3580.2700.128ICCV2023
StreamPETR [35]V2-990.5500.6360.4790.2390.3170.2410.119ICCV2023
SparseBEV [27]V2-990.5560.6360.4850.2440.3320.2460.117ICCV2023
Sparse4Dv2 [26]V2-990.5570.6380.4620.2380.3280.2640.115-
RecurrentBEVV2-990.5730.6510.4100.2470.3230.2490.123-
BEVDepth [21]ConvNeXt-B0.5200.6090.4450.2430.3520.3470.127AAAI2023
AeDet* [7]ConvNeXt-B0.5310.6200.4390.2470.3440.2920.130CVPR2023
SOLOFusion [33]ConvNeXt-B0.5400.6190.4530.2570.3760.2760.148ICLR2023
VCD-A [14]ConvNeXt-B0.5480.6310.4360.2440.3430.2900.120NIPS2023
StreamPETR [35]ConvNeXt-B0.5500.6310.4930.2410.3430.2430.123ICCV2023
VideoBEV [8]ConvNeXt-B0.5540.6290.4570.2490.3810.2660.132-
RecurrentBEVConvNeXt-B0.5740.6510.4110.2480.3330.2570.116-
", + "image_path": "65ec0d5ad1b4872404010cc113b0b4ffd7c5ad8c94d6b1b7f220ac1e410e5e58.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 353, + 480, + 384 + ], + "lines": [ + { + "bbox": [ + 132, + 353, + 480, + 384 + ], + "spans": [ + { + "bbox": [ + 132, + 353, + 480, + 384 + ], + "type": "text", + "content": "Table 3: Comparison results on 3D detection on the nuScenes test set. * indicates using test-time augmentation in test process. No future frames are used in any of the methods." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 132, + 422, + 228, + 433 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 422, + 228, + 433 + ], + "spans": [ + { + "bbox": [ + 132, + 422, + 228, + 433 + ], + "type": "text", + "content": "4.3 Main Results" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 447, + 480, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 447, + 480, + 471 + ], + "spans": [ + { + "bbox": [ + 130, + 447, + 480, + 471 + ], + "type": "text", + "content": "We compare the proposed RecurrentBEV with previous state-of-the-art multiview 3D detectors on the nuScenes validation and test set." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 472, + 481, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 472, + 481, + 581 + ], + "spans": [ + { + "bbox": [ + 130, + 472, + 481, + 581 + ], + "type": "text", + "content": "On the nuScenes validation set, as shown in Fig.1b and Tab.2, the real-time version of RecurrentBEV (25.6 FPS), using ResNet50 as backbone and " + }, + { + "bbox": [ + 130, + 472, + 481, + 581 + ], + "type": "inline_equation", + "content": "256 \\times 704" + }, + { + "bbox": [ + 130, + 472, + 481, + 581 + ], + "type": "text", + "content": " as the input size, outperforms the previous state-of-the-art method StreamPETR by " + }, + { + "bbox": [ + 130, + 472, + 481, + 581 + ], + "type": "inline_equation", + "content": "1.3\\%" + }, + { + "bbox": [ + 130, + 472, + 481, + 581 + ], + "type": "text", + "content": " mAP and " + }, + { + "bbox": [ + 130, + 472, + 481, + 581 + ], + "type": "inline_equation", + "content": "0.9\\%" + }, + { + "bbox": [ + 130, + 472, + 481, + 581 + ], + "type": "text", + "content": " NDS on the nuScenes validation set. To scale up our method, we adopt ResNet101 backbone with nuImages pre-training and extend input size to " + }, + { + "bbox": [ + 130, + 472, + 481, + 581 + ], + "type": "inline_equation", + "content": "512 \\times 1408" + }, + { + "bbox": [ + 130, + 472, + 481, + 581 + ], + "type": "text", + "content": ", outperforming the same setting StreamPETR by " + }, + { + "bbox": [ + 130, + 472, + 481, + 581 + ], + "type": "inline_equation", + "content": "2.4\\%" + }, + { + "bbox": [ + 130, + 472, + 481, + 581 + ], + "type": "text", + "content": " mAP and " + }, + { + "bbox": [ + 130, + 472, + 481, + 581 + ], + "type": "inline_equation", + "content": "2.0\\%" + }, + { + "bbox": [ + 130, + 472, + 481, + 581 + ], + "type": "text", + "content": " NDS. It's worth noting that RecurrentBEV makes remarkable progress on the mAP and mATE metrics, indicating the improvement of detection of occlusion objects. We'll delve into the reasons behind this in Sec.4.6." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 582, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 481, + 665 + ], + "type": "text", + "content": "On the nuScenes test set in Tab.3, we compare with all the recent methods in multi-view 3D object detection. Following the settings in StreamPETR, we adopt V2-99 and ConvNeXt-B as backbone and " + }, + { + "bbox": [ + 130, + 582, + 481, + 665 + ], + "type": "inline_equation", + "content": "640 \\times 1600" + }, + { + "bbox": [ + 130, + 582, + 481, + 665 + ], + "type": "text", + "content": " as the input size. Using V2-99 as backbone, RecurrentBEV can surpass the previous state-of-the-art method Sparse4Dv2 by " + }, + { + "bbox": [ + 130, + 582, + 481, + 665 + ], + "type": "inline_equation", + "content": "1.6\\%" + }, + { + "bbox": [ + 130, + 582, + 481, + 665 + ], + "type": "text", + "content": " mAP and " + }, + { + "bbox": [ + 130, + 582, + 481, + 665 + ], + "type": "inline_equation", + "content": "1.3\\%" + }, + { + "bbox": [ + 130, + 582, + 481, + 665 + ], + "type": "text", + "content": " NDS. When the backbone is scaled up to ConvNeXt-B, RecurrentBEV demonstrates comparable performance, surpassing other methods." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 386, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 386, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 386, + 91, + 447, + 100 + ], + "type": "text", + "content": "RecurrentBEV" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 177, + 114, + 436, + 170 + ], + "blocks": [ + { + "bbox": [ + 177, + 114, + 436, + 170 + ], + "lines": [ + { + "bbox": [ + 177, + 114, + 436, + 170 + ], + "spans": [ + { + "bbox": [ + 177, + 114, + 436, + 170 + ], + "type": "table", + "html": "
methodmAP↑NDS↑memory usage↓latency↑
parallel fusion [33]0.4270.534945MB96ms
recurrent fusion baseline0.3960.515516MB34ms
RecurrentBEV0.4450.549549MB39ms
", + "image_path": "e1fb0d106c14c50d0db68c81aaa2e50ada8fe706391face9337d818eb8e45b1e.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 170, + 481, + 204 + ], + "lines": [ + { + "bbox": [ + 130, + 170, + 481, + 204 + ], + "spans": [ + { + "bbox": [ + 130, + 170, + 481, + 204 + ], + "type": "text", + "content": "Table 4: Comparison of time and memory in inference. All data are measured on a single RTX 3090 with PyTorch fp32. Implementation details are the same as Tab.2 ResNet50 Backbone. For more latency analysis, please refer to supplementary." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 132, + 237, + 319, + 250 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 237, + 319, + 250 + ], + "spans": [ + { + "bbox": [ + 132, + 237, + 319, + 250 + ], + "type": "text", + "content": "4.4 Time and Memory Comparison" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 258, + 482, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 258, + 482, + 342 + ], + "spans": [ + { + "bbox": [ + 130, + 258, + 482, + 342 + ], + "type": "text", + "content": "To demonstrate the time and memory efficiency in inference, we compare the proposed RecurrentBEV with parallel fusion methods and previous recurrent fusion baselines. As shown in Tab.4 RecurrentBEV is 2.5 times faster than the normal parallel fusion method SOLOFusion [33], and the memory usage is reduced by " + }, + { + "bbox": [ + 130, + 258, + 482, + 342 + ], + "type": "inline_equation", + "content": "42\\%" + }, + { + "bbox": [ + 130, + 258, + 482, + 342 + ], + "type": "text", + "content": ". When compared to the normal recurrent baseline, RecurrentBEV requires only an additional 33MB of memory while significantly improving the mean Average Precision (mAP) from 0.396 to 0.445." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 361, + 238, + 373 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 361, + 238, + 373 + ], + "spans": [ + { + "bbox": [ + 132, + 361, + 238, + 373 + ], + "type": "text", + "content": "4.5 Ablation Study" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 383, + 481, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 383, + 481, + 418 + ], + "spans": [ + { + "bbox": [ + 130, + 383, + 481, + 418 + ], + "type": "text", + "content": "In the ablation study, we delve into three aspects: the components of RecurrentBEV, the receptive field of TA-GRU, and the length of frame sequence in training. We conduct the ablation experiments with ResNet-50 backbone." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 418, + 482, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 418, + 482, + 502 + ], + "spans": [ + { + "bbox": [ + 130, + 418, + 482, + 502 + ], + "type": "text", + "content": "Ablation of Each Component. We conduct a component-wise ablation in Tab.5. Exp.1 starts with a ConvGRU as temporal fusion module, achieving similar performance to VideoBEV [8]. In Exp.2 and Exp.3, we enable the backward propagation of historical BEV features in training process. In Exp.3 and Exp.4, we replace the temporal fusion module with TA-GRU. The overall components (BP and TA-GRU) brings significant performance gains " + }, + { + "bbox": [ + 130, + 418, + 482, + 502 + ], + "type": "inline_equation", + "content": "(+1.7\\% \\mathrm{mAP}, +1.6\\%" + }, + { + "bbox": [ + 130, + 418, + 482, + 502 + ], + "type": "text", + "content": " NDS,-4.7% mAOE) over baseline." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 502, + 482, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 502, + 482, + 586 + ], + "spans": [ + { + "bbox": [ + 130, + 502, + 482, + 586 + ], + "type": "text", + "content": "Specifically, the TA-GRU demonstrates its ability to enhance the mAP and mAOE metrics, regardless of whether backward propagation of historical BEV features is enabled. This proves the TA-GRU's consistent ability to alleviate semantic ambiguity for identical objects. Notably, the mean average orientation error (mAOE) has a notable increase, indicating that the inner grid transformation, facilitated by the alignment of historical BEV features, effectively mitigates semantic ambiguity by refining the orientation of objects." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 587, + 482, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 587, + 482, + 635 + ], + "spans": [ + { + "bbox": [ + 130, + 587, + 482, + 635 + ], + "type": "text", + "content": "As demonstrated in Exp.4, when the historical BEV feature is detached, the performance suffers, highlighting the importance of RNN-style back-propagation. However, relying solely on back-propagation of historical BEV features during training results in limited improvement. This shows the necessity of employing" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 233, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 233, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 233, + 102 + ], + "type": "text", + "content": "M. Chang et al." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 133, + 641, + 481, + 666 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 641, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 133, + 641, + 481, + 666 + ], + "type": "text", + "content": "similar performance to VideoBEV [8] without stereo matching (VideoBEV-D) with 60 training epochs." + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 217, + 114, + 396, + 180 + ], + "blocks": [ + { + "bbox": [ + 217, + 114, + 396, + 180 + ], + "lines": [ + { + "bbox": [ + 217, + 114, + 396, + 180 + ], + "spans": [ + { + "bbox": [ + 217, + 114, + 396, + 180 + ], + "type": "table", + "html": "
Exp|BP TG|mAP↑|NDS↑|mAOE↓|mAVE↓
10.3710.4760.6460.317
2✓0.3750.4790.6340.319
3✓✓0.3880.4920.5960.281
4✓0.3780.4840.6020.316
", + "image_path": "226a2d686539e788e37dd02f792775c258931569d2d5252d7ab37d6b2008bc02.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 144, + 217, + 293, + 284 + ], + "blocks": [ + { + "bbox": [ + 132, + 180, + 479, + 202 + ], + "lines": [ + { + "bbox": [ + 132, + 180, + 479, + 202 + ], + "spans": [ + { + "bbox": [ + 132, + 180, + 479, + 202 + ], + "type": "text", + "content": "Table 5: Ablation of each component. BP means the historical BEV feature backward propagate. TG means TA-GRU." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 144, + 217, + 293, + 284 + ], + "lines": [ + { + "bbox": [ + 144, + 217, + 293, + 284 + ], + "spans": [ + { + "bbox": [ + 144, + 217, + 293, + 284 + ], + "type": "table", + "html": "
Kernel sizemAP↑NDS↑mATE↓mAVE↓
1 × 10.3880.4920.6470.281
3 × 30.3970.5000.6320.284
5 × 50.4030.5060.6150.276
7 × 70.3980.5010.6300.273
9 × 90.3940.4960.6330.280
", + "image_path": "204c57375c992e61c6076c20f9467a8446a4f7c659c16c2b441cb5c4ec2ef4c5.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 310, + 209, + 474, + 284 + ], + "blocks": [ + { + "bbox": [ + 144, + 284, + 293, + 294 + ], + "lines": [ + { + "bbox": [ + 144, + 284, + 293, + 294 + ], + "spans": [ + { + "bbox": [ + 144, + 284, + 293, + 294 + ], + "type": "text", + "content": "Table 6: Ablation of receptive field." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 310, + 209, + 474, + 284 + ], + "lines": [ + { + "bbox": [ + 310, + 209, + 474, + 284 + ], + "spans": [ + { + "bbox": [ + 310, + 209, + 474, + 284 + ], + "type": "table", + "html": "
Training framesmAP↑NDS↑mATE↓mAVE↓
00.3180.3800.7420.835
10.3460.4570.6960.389
20.3720.4790.6650.310
40.3900.4930.6460.283
80.4030.5060.6150.276
160.4020.490.6280.280
", + "image_path": "f79157ca373f0cf5aff8882bc5d3d7a702e94e704ac0940d010d56f832deb659.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 312, + 284, + 470, + 294 + ], + "lines": [ + { + "bbox": [ + 312, + 284, + 470, + 294 + ], + "spans": [ + { + "bbox": [ + 312, + 284, + 470, + 294 + ], + "type": "text", + "content": "Table 7: Ablation of sequence length." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 130, + 307, + 480, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 307, + 480, + 342 + ], + "spans": [ + { + "bbox": [ + 130, + 307, + 480, + 342 + ], + "type": "text", + "content": "the correct temporal fusion module for effective back-propagation, as the combination of these two components can achieve a synergistic effect that surpasses the sum of their individual contributions." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 342, + 480, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 342, + 480, + 426 + ], + "spans": [ + { + "bbox": [ + 130, + 342, + 480, + 426 + ], + "type": "text", + "content": "Ablation of Receptive Field. Receptive field is critical for temporal fusion of moving objects. We explore the kernel size of convolutions in TA-GRU in Tab.6. As we can see, an excessively large or small receptive field can have a detrimental effect on performance. Our suspect is that a fusion module with an overlarge receptive field gathers noisy information, while a small receptive field leads to mismatches for moving objects. Consequently, we use a " + }, + { + "bbox": [ + 130, + 342, + 480, + 426 + ], + "type": "inline_equation", + "content": "5 \\times 5" + }, + { + "bbox": [ + 130, + 342, + 480, + 426 + ], + "type": "text", + "content": " kernel size in TA-GRU, unless otherwise specified." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 426, + 481, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 426, + 481, + 510 + ], + "spans": [ + { + "bbox": [ + 130, + 426, + 481, + 510 + ], + "type": "text", + "content": "Ablation of Training Sequence Length. We use the BEVDet-Depth with no temporal fusion as baseline and apply our method with different training sequence length in Tab.7. The fusion of a single frame greatly enhances velocity prediction (mAVE). The fusion of four to eight frames mainly enhances localisation metrics mAP and mATE, with an increase of " + }, + { + "bbox": [ + 130, + 426, + 481, + 510 + ], + "type": "inline_equation", + "content": "8.5\\%" + }, + { + "bbox": [ + 130, + 426, + 481, + 510 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 426, + 481, + 510 + ], + "type": "inline_equation", + "content": "12.7\\%" + }, + { + "bbox": [ + 130, + 426, + 481, + 510 + ], + "type": "text", + "content": " respectively. Further increasing the number of frames to 16 offers gains limited performance improvement. Therefore, we train our models on 8 frames." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 527, + 350, + 539 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 527, + 350, + 539 + ], + "spans": [ + { + "bbox": [ + 132, + 527, + 350, + 539 + ], + "type": "text", + "content": "4.6 Effect of Recurrent Long-term Fusion" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": "Detection of Occlusion Objects. Intuitively, long-term fusion has the potential to aid in the detection of occluded objects. Therefore, we conducted an analysis of object occlusions on the nuScenes validation set. Firstly, we define a 'distinct object' as one where the visibility of the entire object exceeds " + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "inline_equation", + "content": "60\\%" + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": ". Subsequently, we determined that if an object is classified as a 'distinct object' in any frame of the video, it is considered visible throughout the video. Finally, we divided the number of visible objects by the total number of objects present to get the overall visibility of the video. As shown in Fig.4a, the visibility of the video has the same trend with mAP, proving the improvement of mAP in long-term fusion originates from the detection of occlusion objects." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 386, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 386, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 386, + 91, + 447, + 101 + ], + "type": "text", + "content": "RecurrentBEV" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 134, + 116, + 306, + 233 + ], + "blocks": [ + { + "bbox": [ + 134, + 116, + 306, + 233 + ], + "lines": [ + { + "bbox": [ + 134, + 116, + 306, + 233 + ], + "spans": [ + { + "bbox": [ + 134, + 116, + 306, + 233 + ], + "type": "image", + "image_path": "812dfcd2a3b4fa1490e5d506dfd24ae98b085b72a1abee35dca80c25252e598c.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 211, + 236, + 223, + 245 + ], + "lines": [ + { + "bbox": [ + 211, + 236, + 223, + 245 + ], + "spans": [ + { + "bbox": [ + 211, + 236, + 223, + 245 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 322, + 116, + 474, + 232 + ], + "blocks": [ + { + "bbox": [ + 322, + 116, + 474, + 232 + ], + "lines": [ + { + "bbox": [ + 322, + 116, + 474, + 232 + ], + "spans": [ + { + "bbox": [ + 322, + 116, + 474, + 232 + ], + "type": "image", + "image_path": "398adb6d4e1c1b0ed0a2dc7fd1f5ab3add939cdb6f29f70d65c2dd144c0e846d.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 392, + 233, + 405, + 243 + ], + "lines": [ + { + "bbox": [ + 392, + 233, + 405, + 243 + ], + "spans": [ + { + "bbox": [ + 392, + 233, + 405, + 243 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 243, + 482, + 310 + ], + "lines": [ + { + "bbox": [ + 130, + 243, + 482, + 310 + ], + "spans": [ + { + "bbox": [ + 130, + 243, + 482, + 310 + ], + "type": "text", + "content": "Fig. 4: (a) Relation between visibility and mAP. Visibility refers to the proportion of distinct objects in a video, expressed as a percentage of the total number of objects present. The Pearson correlation coefficient between visibility and mAP is 0.977.(b) The mATE metrics with objects at different location on the nuScenes validation set. The blue one is a result from single frame baseline and the orange one is RecurrentBEV. The objects between 25 meters and 30 meters improve the mATE metric by " + }, + { + "bbox": [ + 130, + 243, + 482, + 310 + ], + "type": "inline_equation", + "content": "24\\%" + }, + { + "bbox": [ + 130, + 243, + 482, + 310 + ], + "type": "text", + "content": " mostly." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 335, + 482, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 335, + 482, + 443 + ], + "spans": [ + { + "bbox": [ + 130, + 335, + 482, + 443 + ], + "type": "text", + "content": "Localization of Distant Objects. The localization of objects is a major challenge for camera-based 3D detector. Due to the uncertainty of depth estimation. We speculate long-term fusion can more accurately localize objects that move from near to far as the ego vehicle moves. Therefore, we calculate the mATE metrics of objects at different distances before and after long-term fusion, as in Fig.4b. Notably, objects between 25 and 30 meters exhibit the most significant improvement in mATE through long-term fusion, whereas nearby objects show marginal gains. This finding demonstrates the efficacy of long-term fusion in accurately localizing distant objects." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 466, + 326, + 479 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 466, + 326, + 479 + ], + "spans": [ + { + "bbox": [ + 132, + 466, + 326, + 479 + ], + "type": "text", + "content": "5 Conclusion and Future Work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 497, + 482, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 497, + 482, + 616 + ], + "spans": [ + { + "bbox": [ + 130, + 497, + 482, + 616 + ], + "type": "text", + "content": "In this paper, we investigate the differences between recurrent fusion and parallel fusion of BEV-based 3D object detectors, and propose RecurrentBEV, a simple and effective recurrent temporal fusion method. By incorporating RNN-style back-propagation and a newly designed inner grid transformation, RecurrentBEV enhances long-term fusion capabilities while maintaining latency and memory advantages. Rigorous experiments conducted on the nuScenes dataset demonstrate that RecurrentBEV outperforms all prior methods, while maintaining competitive inference speeds, thereby facilitating deployment. To our knowledge, it represents the first recurrent fusion method with explicit BEV features that surpasses parallel fusion performance." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "In the future, we will generalize recurrent fusion idea to more BEV architectures such as different BEV view transform backbones and query-based methods. We will also try to extend RecurrentBEV to other 3D perceptual tasks and datasets such as object tracking and trajectory prediction." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 233, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 233, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 233, + 102 + ], + "type": "text", + "content": "M. Chang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 139, + 482, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 139, + 482, + 224 + ], + "spans": [ + { + "bbox": [ + 130, + 139, + 482, + 224 + ], + "type": "text", + "content": "This work is partially supported by the National Key R&D Program of China (under Grant 2023YFB4502200), the NSF of China (under Grants U22A2028, 61925208, 62102399, 62222214, 62341411, 62102398, U20A20227, 62372436, 623-02478, 62302482, 62302483, 62302480), Strategic Priority Research Program of the Chinese Academy of Sciences, (Grant No.XDB0660200, XDB0660201, XDB0660202), CAS Project for Young Scientists in Basic Research(YSBR-029), Youth Innovation Promotion Association CAS and Xplore Prize." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 239, + 197, + 252 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 239, + 197, + 252 + ], + "spans": [ + { + "bbox": [ + 133, + 239, + 197, + 252 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 263, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 138, + 263, + 480, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 263, + 480, + 285 + ], + "spans": [ + { + "bbox": [ + 138, + 263, + 480, + 285 + ], + "type": "text", + "content": "1. Ballas, N., Yao, L., Pal, C., Courville, A.: Delving deeper into convolutional networks for learning video representations. In: ICLR (2016)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 285, + 480, + 317 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 285, + 480, + 317 + ], + "spans": [ + { + "bbox": [ + 138, + 285, + 480, + 317 + ], + "type": "text", + "content": "2. Caesar, H., Bankiti, V., Lang, A.H., Vora, S., Liong, V.E., Xu, Q., Krishnan, A., Pan, Y., Baldan, G., Beijbom, O.: nuscenes: A multimodal dataset for autonomous driving. In: CVPR (2020)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 319, + 480, + 339 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 319, + 480, + 339 + ], + "spans": [ + { + "bbox": [ + 138, + 319, + 480, + 339 + ], + "type": "text", + "content": "3. Carion, N., Massa, F., Synnaeve, G., Usunier, N., Kirillov, A., Zagoruyko, S.: End-to-end object detection with transformers. In: ECCV (2020)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 340, + 480, + 361 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 340, + 480, + 361 + ], + "spans": [ + { + "bbox": [ + 138, + 340, + 480, + 361 + ], + "type": "text", + "content": "4. Chen, D., Li, J., Guizilini, V., Ambrus, R.A., Gaidon, A.: Viewpoint equivariance for multi-view 3d object detection. In: CVPR (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 361, + 480, + 394 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 361, + 480, + 394 + ], + "spans": [ + { + "bbox": [ + 138, + 361, + 480, + 394 + ], + "type": "text", + "content": "5. Chen, L.C., Papandreou, G., Kokkinos, I., Murphy, K., Yuille, A.L.: Deeplab: Semantic image segmentation with deep convolutional nets, atrous convolution, and fully connected crfs. IEEE TPAMI (2017)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 395, + 480, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 395, + 480, + 415 + ], + "spans": [ + { + "bbox": [ + 138, + 395, + 480, + 415 + ], + "type": "text", + "content": "6. Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: Imagenet: A large-scale hierarchical image database. In: CVPR (2009)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 416, + 480, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 416, + 480, + 437 + ], + "spans": [ + { + "bbox": [ + 138, + 416, + 480, + 437 + ], + "type": "text", + "content": "7. Feng, C., Jie, Z., Zhong, Y., Chu, X., Ma, L.: Aedet: Azimuth-invariant multi-view 3d object detection. In: CVPR (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 437, + 480, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 437, + 480, + 470 + ], + "spans": [ + { + "bbox": [ + 138, + 437, + 480, + 470 + ], + "type": "text", + "content": "8. Han, C., Sun, J., Ge, Z., Yang, J., Dong, R., Zhou, H., Mao, W., Peng, Y., Zhang, X.: Exploring recurrent long-term temporal fusion for multi-view 3d perception. arXiv preprint arXiv:2303.05970 (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 471, + 480, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 471, + 480, + 491 + ], + "spans": [ + { + "bbox": [ + 138, + 471, + 480, + 491 + ], + "type": "text", + "content": "9. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: CVPR (2016)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 492, + 480, + 524 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 492, + 480, + 524 + ], + "spans": [ + { + "bbox": [ + 138, + 492, + 480, + 524 + ], + "type": "text", + "content": "0. Huang, B., Li, Y., Xie, E., Liang, F., Wang, L., Shen, M., Liu, F., Wang, T., Luo, P., Shao, J.: Fast-bev: Towards real-time on-vehicle bird's-eye view perception. arXiv preprint arXiv:2301.07870 (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 525, + 480, + 545 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 525, + 480, + 545 + ], + "spans": [ + { + "bbox": [ + 138, + 525, + 480, + 545 + ], + "type": "text", + "content": "1. Huang, J., Huang, G.: Bevdet4d: Exploit temporal cues in multi-camera 3d object detection. arXiv preprint arXiv:2203.17054 (2022)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 546, + 480, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 546, + 480, + 567 + ], + "spans": [ + { + "bbox": [ + 138, + 546, + 480, + 567 + ], + "type": "text", + "content": "2. Huang, J., Huang, G.: Bevpoolv2: A cutting-edge implementation of bevdet toward deployment. arXiv preprint arXiv:2211.17111 (2022)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 568, + 480, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 568, + 480, + 600 + ], + "spans": [ + { + "bbox": [ + 138, + 568, + 480, + 600 + ], + "type": "text", + "content": "3. Huang, J., Huang, G., Zhu, Z., Ye, Y., Du, D.: Bevdet: High-performance multicamera 3d object detection in bird-eye-view. arXiv preprint arXiv:2112.11790 (2021)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 601, + 480, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 601, + 480, + 621 + ], + "spans": [ + { + "bbox": [ + 138, + 601, + 480, + 621 + ], + "type": "text", + "content": "4. Huang, L., Li, Z., Sima, C., Wang, W., Wang, J., Qiao, Y., Li, H.: Leveraging vision-centric multi-modal expertise for 3d object detection. In: NeurIPS (2023)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 138, + 622, + 480, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 622, + 480, + 643 + ], + "spans": [ + { + "bbox": [ + 138, + 622, + 480, + 643 + ], + "type": "text", + "content": "5. Jaderberg, M., Simonyan, K., Zisserman, A., et al.: Spatial transformer networks. In: NeurIPS (2015)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 138, + 643, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 643, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 643, + 480, + 665 + ], + "type": "text", + "content": "6. Jiang, Y., Zhang, L., Miao, Z., Zhu, X., Gao, J., Hu, W., Jiang, Y.G.: Polarformer: Multi-camera 3d object detection with polar transformer. In: AAAI (2023)" + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 386, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 386, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 386, + 91, + 447, + 101 + ], + "type": "text", + "content": "RecurrentBEV" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 117, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 133, + 117, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 117, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 133, + 117, + 480, + 138 + ], + "type": "text", + "content": "17. Lee, Y., Hwang, J.w., Lee, S., Bae, Y., Park, J.: An energy andgpu-computation efficient backbone network for real-time object detection. In: CVPRW (2019)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 139, + 480, + 159 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 139, + 480, + 159 + ], + "spans": [ + { + "bbox": [ + 133, + 139, + 480, + 159 + ], + "type": "text", + "content": "18. Li, H., Zhang, H., Zeng, Z., Liu, S., Li, F., Ren, T., Zhang, L.: Dfa3d: 3d deformable attention for 2d-to-3d feature lifting. In: ICCV (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 160, + 480, + 181 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 160, + 480, + 181 + ], + "spans": [ + { + "bbox": [ + 133, + 160, + 480, + 181 + ], + "type": "text", + "content": "19. Li, Y., Chen, Y., Qi, X., Li, Z., Sun, J., Jia, J.: Unifying voxel-based representation with transformer for 3d object detection. In: NeurIPS (2022)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 182, + 480, + 202 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 182, + 480, + 202 + ], + "spans": [ + { + "bbox": [ + 132, + 182, + 480, + 202 + ], + "type": "text", + "content": "20. Li, Y., Bao, H., Ge, Z., Yang, J., Sun, J., Li, Z.: Bevstereo: enhancing depth estimation in multi-view 3d object detection with temporal stereo. In: AAAI (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 203, + 480, + 224 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 203, + 480, + 224 + ], + "spans": [ + { + "bbox": [ + 132, + 203, + 480, + 224 + ], + "type": "text", + "content": "21. Li, Y., Ge, Z., Yu, G., Yang, J., Wang, Z., Shi, Y., Sun, J., Li, Z.: Bevdepth: Acquisition of reliable depth for multi-view 3d object detection. In: AAAI (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 225, + 480, + 256 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 225, + 480, + 256 + ], + "spans": [ + { + "bbox": [ + 132, + 225, + 480, + 256 + ], + "type": "text", + "content": "22. Li, Z., Wang, W., Li, H., Xie, E., Sima, C., Lu, T., Qiao, Y., Dai, J.: Bevformer: Learning bird's-eye-view representation from multi-camera images via spatiotemporal transformers. In: ECCV (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 257, + 480, + 278 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 257, + 480, + 278 + ], + "spans": [ + { + "bbox": [ + 132, + 257, + 480, + 278 + ], + "type": "text", + "content": "23. Li, Z., Yu, Z., Wang, W., Anandkumar, A., Lu, T., Alvarez, J.M.: Fb-bev: Bev representation from forward-backward view transformations. In: ICCV (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 279, + 480, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 279, + 480, + 300 + ], + "spans": [ + { + "bbox": [ + 132, + 279, + 480, + 300 + ], + "type": "text", + "content": "24. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dóllar, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: ECCV (2014)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 300, + 480, + 320 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 300, + 480, + 320 + ], + "spans": [ + { + "bbox": [ + 132, + 300, + 480, + 320 + ], + "type": "text", + "content": "25. Lin, X., Lin, T., Pei, Z., Huang, L., Su, Z.: Sparse4d: Multi-view 3d object detection with sparse spatial-temporal fusion. arXiv preprint arXiv:2211.10581 (2022)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 322, + 480, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 322, + 480, + 342 + ], + "spans": [ + { + "bbox": [ + 132, + 322, + 480, + 342 + ], + "type": "text", + "content": "26. Lin, X., Lin, T., Pei, Z., Huang, L., Su, Z.: Sparse4d v2: Recurrent temporal fusion with sparse model. arXiv preprint arXiv:2305.14018 (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 343, + 480, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 343, + 480, + 364 + ], + "spans": [ + { + "bbox": [ + 132, + 343, + 480, + 364 + ], + "type": "text", + "content": "27. Liu, H., Teng, Y., Lu, T., Wang, H., Wang, L.: Sparsebev: High-performance sparse 3d object detection from multi-camera videos. In: ICCV (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 365, + 480, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 365, + 480, + 384 + ], + "spans": [ + { + "bbox": [ + 132, + 365, + 480, + 384 + ], + "type": "text", + "content": "28. Liu, Y., Wang, T., Zhang, X., Sun, J.: Petr: Position embedding transformation for multi-view 3d object detection. In: ECCV (2022)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 385, + 480, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 385, + 480, + 407 + ], + "spans": [ + { + "bbox": [ + 132, + 385, + 480, + 407 + ], + "type": "text", + "content": "29. Liu, Y., Yan, J., Jia, F., Li, S., Gao, A., Wang, T., Zhang, X.: Petrov2: A unified framework for 3d perception from multi-camera images. In: ICCV (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 407, + 480, + 428 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 407, + 480, + 428 + ], + "spans": [ + { + "bbox": [ + 132, + 407, + 480, + 428 + ], + "type": "text", + "content": "30. Liu, Z., Mao, H., Wu, C.Y., Feichtenhofer, C., Darrell, T., Xie, S.: A convnet for the 2020s. In: CVPR (2022)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 429, + 480, + 449 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 429, + 480, + 449 + ], + "spans": [ + { + "bbox": [ + 132, + 429, + 480, + 449 + ], + "type": "text", + "content": "31. Luo, Z., Zhou, C., Zhang, G., Lu, S.: Detr4d: Direct multi-view 3d object detection with sparse attention. arXiv preprint arXiv:2212.07849 (2022)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 450, + 480, + 471 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 450, + 480, + 471 + ], + "spans": [ + { + "bbox": [ + 132, + 450, + 480, + 471 + ], + "type": "text", + "content": "32. Park, D., Ambrus, R., Guizilini, V., Li, J., Gaidon, A.: Is pseudo-lidar needed for monocular 3d object detection? In: ICCV (2021)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 472, + 480, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 472, + 480, + 503 + ], + "spans": [ + { + "bbox": [ + 132, + 472, + 480, + 503 + ], + "type": "text", + "content": "33. Park, J., Xu, C., Yang, S., Keutzer, K., Kitani, K., Tomizuka, M., Zhan, W.: Time will tell: New outlooks and a baseline for temporal multi-view 3d object detection. In: ICLR (2023)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 504, + 480, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 504, + 480, + 525 + ], + "spans": [ + { + "bbox": [ + 132, + 504, + 480, + 525 + ], + "type": "text", + "content": "34. Philion, J., Fidler, S.: Lift, splat, shoot: Encoding images from arbitrary camera rigs by implicitly unprojecting to 3d. In: ECCV (2020)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 132, + 525, + 480, + 546 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 525, + 480, + 546 + ], + "spans": [ + { + "bbox": [ + 132, + 525, + 480, + 546 + ], + "type": "text", + "content": "35. Wang, S., Liu, Y., Wang, T., Li, Y., Zhang, X.: Exploring object-centric temporal modeling for efficient multi-view 3d object detection. In: ICCV (2023)" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 132, + 547, + 480, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 547, + 480, + 578 + ], + "spans": [ + { + "bbox": [ + 132, + 547, + 480, + 578 + ], + "type": "text", + "content": "36. Wang, Y., Guizilini, V.C., Zhang, T., Wang, Y., Zhao, H., Solomon, J.: Detr3d: 3d object detection from multi-view images via 3d-to-2d queries. In: Conference on Robot Learning (2022)" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 132, + 579, + 480, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 579, + 480, + 600 + ], + "spans": [ + { + "bbox": [ + 132, + 579, + 480, + 600 + ], + "type": "text", + "content": "37. Wang, Z., Huang, Z., Fu, J., Wang, N., Liu, S.: Object as query: Lifting any 2d object detector to 3d detection. In: ICCV (2023)" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 132, + 601, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 601, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 601, + 480, + 632 + ], + "type": "text", + "content": "38. Xiong, K., Gong, S., Ye, X., Tan, X., Wan, J., Ding, E., Wang, J., Bai, X.: Cape: Camera view position embedding for multi-view 3d object detection. In: CVPR (2023)" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 132, + 633, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 633, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 633, + 480, + 665 + ], + "type": "text", + "content": "39. Yang, C., Chen, Y., Tian, H., Tao, C., Zhu, X., Zhang, Z., Huang, G., Li, H., Qiao, Y., Lu, L., et al.: Bevformer v2: Adapting modern image backbones to bird's-eye-view recognition via perspective supervision. In: CVPR (2023)" + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 231, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 231, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 231, + 102 + ], + "type": "text", + "content": "M. Chang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 205 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 480, + 138 + ], + "type": "text", + "content": "40. Yin, T., Zhou, X., Krahenbuhl, P.: Center-based 3d object detection and tracking. In: CVPR (2021)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 138, + 480, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 138, + 480, + 160 + ], + "spans": [ + { + "bbox": [ + 132, + 138, + 480, + 160 + ], + "type": "text", + "content": "41. Zhang, J., Zhang, Y., Liu, Q., Wang, Y.: Sa-bev: Generating semantic-aware bird's-eye-view feature for multi-view 3d object detection. In: ICCV (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 160, + 480, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 160, + 480, + 182 + ], + "spans": [ + { + "bbox": [ + 132, + 160, + 480, + 182 + ], + "type": "text", + "content": "42. Zhou, H., Ge, Z., Li, Z., Zhang, X.: Matrixvt: Efficient multi-camera to bev transformation for 3d perception. In: ICCV (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 182, + 480, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 182, + 480, + 205 + ], + "spans": [ + { + "bbox": [ + 132, + 182, + 480, + 205 + ], + "type": "text", + "content": "43. Zhu, B., Jiang, Z., Zhou, X., Li, Z., Yu, G.: Class-balanced grouping and sampling for point cloud 3d object detection. arXiv preprint arXiv:1908.09492 (2019)" + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 386, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 386, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 386, + 91, + 447, + 100 + ], + "type": "text", + "content": "RecurrentBEV" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Recursive Visual Programming/e57bdabb-77ec-4030-98f0-ae871f741b2d_content_list.json b/2024/Recursive Visual Programming/e57bdabb-77ec-4030-98f0-ae871f741b2d_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..1f0a324af53e03f55bebfe73a88a4272b6c123dd --- /dev/null +++ b/2024/Recursive Visual Programming/e57bdabb-77ec-4030-98f0-ae871f741b2d_content_list.json @@ -0,0 +1,1668 @@ +[ + { + "type": "text", + "text": "Recursive Visual Programming", + "text_level": 1, + "bbox": [ + 321, + 141, + 679, + 162 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jiaxin Ge1 Sanjay Subramanian1 Baifeng Shi1 Roei Herzig1 Trevor Darrell1", + "bbox": [ + 316, + 188, + 683, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "UC Berkeley, CA, USA", + "bbox": [ + 421, + 232, + 580, + 244 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Visual Programming (VP) has emerged as a powerful framework for Visual Question Answering (VQA). By generating and executing bespoke code for each question, these methods show advancement in leveraging Large Language Models (LLMs) for complex problem-solving. Despite their potential, existing VP methods generate all code in a single function, which does not fully utilize LLM's reasoning capacity and the modular adaptability of code. This results in code that is suboptimal in terms of both accuracy and interpretability. Inspired by human coding practices, we propose Recursive Visual Programming (RVP), which better harnesses the reasoning capacity of LLMs, provides modular code structure between code pieces, and assigns different return types for the sub-problems elegantly. RVP approaches VQA tasks with an top-down recursive code generation approach, allowing decomposition of complicated problems into smaller parts. We show RVP's efficacy through extensive experiments on benchmarks including VSR, COVR, GQA, and NextQA, underscoring the value of adopting human-like recursive and modular programming techniques for solving VQA tasks. Our code is available at https://github.com/para-lost/RVP.", + "bbox": [ + 261, + 281, + 738, + 530 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Visual Programming $\\cdot$ Vision-Language $\\cdot$ VQA", + "bbox": [ + 261, + 542, + 661, + 556 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 582, + 375, + 598 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Visual Question Answering (VQA) lies at the intersection of computer vision and natural language processing, posing the challenge of interpreting visual data to answer questions [3,16,26]. Following the recent progress on code generation using Large Language Models (LLMs) [6,27,33], Visual Programming (VP) methods have notably advanced in this area, successfully using LLMs to generate and execute code in few-shot and zero-shot scenarios [14,30,31]. In these techniques, various vision models are provided to LLM as APIs, and the LLM is used as a planner to utilize these vision models to perform reasoning over images.", + "bbox": [ + 212, + 613, + 785, + 734 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Despite these advancements, such planning remains very challenging for Visual Programming. These methods require the model to generate a single code block using the predefined APIs at once. As a result, the model must handle logic, be aware of all details, and utilize the vision APIs appropriately. This has two main drawbacks. First, such approaches do not fully utilize LLM's reasoning capability, which would benefit from breaking down the problem into step-by-step sub-problems, as demonstrated in recent NLP research [4]. Secondly, they", + "bbox": [ + 212, + 734, + 785, + 839 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/cc5ac8908a42096357813ed66e120177bf0638873d475f7f0f643930020c20da.jpg", + "image_caption": [ + "Fig. 1: A breakdown of Recursive Visual Programming for a visual question is illustrated with an image where scenes within it contain smaller versions of themselves. To locate a man wearing a hat at the third level of an image, the model first generates code that has two recursive_query API calls. The initial call identifies the largest mirror (representing the next image level), and the subsequent call seeks \"the man with a hat\" at the second level. The model iteratively generates and executes code through these calls until the final answer is produced without additional recursive calls." + ], + "image_footnote": [], + "bbox": [ + 222, + 143, + 782, + 409 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "fail to utilize the full potential of code because they lack modularity between the code pieces, particularly when the code is one long piece.", + "bbox": [ + 212, + 549, + 785, + 580 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We consider how humans tackle these issues. Humans tend to code in a top-down manner: first they break down complex tasks into manageable subproblems, and then they tackle each through focused code segments. This approach allows the programmer to concentrate on the main logic flow rather than figuring out all the details at once, resulting in a more clear and elegant structure.", + "bbox": [ + 212, + 580, + 785, + 656 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Inspired by this, we hypothesize that incorporating a similar recursive coding strategy would also help solve complex visual question-answering problems. To investigate whether LLMs could indeed benefit from recursive coding, we conduct preliminary experiment of the recursive coding approach on standard symbolic reasoning benchmarks, such as Dyck Language [32] and Games of 24 [38]. This is done by simply prompting the LLM to code recursively. We find that utilizing LLMs to code recursively leads to a notable improvement in accuracy, as shown in Figure 2. Motivated by this result, we explore in this paper how a recursive coding methodology can be applied to VQA tasks.", + "bbox": [ + 212, + 657, + 787, + 792 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Specifically, we propose a novel VP method for VQA tasks, named Recursive Visual Programming (RVP). Instead of querying LLM only once to generate a single piece of code, we ask LLM to decompose the question into sub-questions", + "bbox": [ + 212, + 794, + 787, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "Ge et al.", + "bbox": [ + 271, + 114, + 331, + 126 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/a11a8f66fb1cbde1cb07d94fb4fa00b38ebd21d81121ba7bbb9dd4d01c6c283c.jpg", + "image_caption": [ + "Fig. 2: Motivating Examples. Adding the prompt \"solve the problem recursively\" to the input significantly improves the model's performance on traditional coding tasks, Dyck language [32] and Games of 24 [38]. \"NR\" refers to non-recursive approach and \"R\" refers to recursive approach. This shows that LLMs can potentially benefit from human's recursive programming approach." + ], + "image_footnote": [], + "bbox": [ + 218, + 145, + 781, + 297 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "and recursively query itself for each sub-question to generate modular code pieces. For each sub-question, the same process repeats, until reaching an atom level where the current question can be solved by only calling external APIs. See example in Figure 1. In contrast to querying LLM only once, we find that using recursive queries to generate code for each sub-question can improve the readability and accuracy of the code, as shown in Figure 3. Moreover, MVP allows assigning dynamic types to each sub-question, enabling new code pieces to return a variety of types based on the current need. Remarkably, MVP assigns return types unseen in in-context examples, such as List[str], List[ImagePatch].", + "bbox": [ + 212, + 393, + 787, + 529 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our contribution can be summarized as follows: (i) We introduce a novel recursive approach based on human programming. This approach is a top-down programming method that better harnesses LLM's reasoning capacity and enhances modularity between multiple code pieces. It is easily applicable to any existing VP model for visual question answering task. (ii) We allow dynamic return types for the sub-problems, which enhances flexibility and adaptability. (iii) We demonstrate improved performance on several standard benchmarks, such as GQA, VSR, NextQA and COVR, demonstrating the effectiveness of our approach. We also show better interpretability than non-recursive VP methods.", + "bbox": [ + 212, + 530, + 787, + 667 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 688, + 388, + 704 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Modular Visual Reasoning. There is a long line of work that seeks to combine modularity with the expressivity of deep neural networks. Early work on this topic converts the output of a parser [1] to a program or trains a program generator either separately from or jointly with the execution modules [2, 15]. Later work finds that such models are often challenging to train on real-world datasets, particularly when interpretability of individual module outputs is expected [29]. Other work improves performance on multi-hop visual reasoning without explicitly defining a set of modules, e.g., by predicting and operating", + "bbox": [ + 212, + 719, + 787, + 840 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "Recursive Visual Programming", + "bbox": [ + 524, + 114, + 730, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/a8b7ab2795a081c333bac9386e43bd962bfd8fe27c57853ee5b1b9e61417c441.jpg", + "image_caption": [ + "Fig. 3: Example from COVR (Upper). Current VP methods fail to provide binary 'yes' or 'no' answers, while Recursive Visual Programming method outputs the correct answer. Example from GQA (Middle and Bottom). Recursive outperforms current non-recursive methods by correctly addressing all details and their associated logic." + ], + "image_footnote": [], + "bbox": [ + 220, + 143, + 785, + 412 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "over a scene graph [18].", + "bbox": [ + 212, + 537, + 387, + 553 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Visual Programming. Recent language models can accurately generate code from text descriptions when prompted with a small number of examples [8]. This advance has enabled a few-shot modular visual reasoning approach, commonly called visual programming [14], which consists of prompting a language model for a program and executing the program with pre-trained vision models. This paradigm has led to impressive results in visual question answering, video question answering, visual and tabular question answering, referring expression comprehension, and text-based image editing [11, 12, 23, 30, 31]. Programming with LLMs has also been applied to produce plans and reward functions for robots [21, 24, 40]. Our work builds on these techniques by enabling the LLM to generate code in multiple steps-first planning the overall process and then filling in the details-rather than all at once.", + "bbox": [ + 212, + 553, + 787, + 733 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Language Models for Reasoning Tasks. In the text domain, there is a large body of work on using language models for reasoning tasks. In particular, several papers have shown the efficacy of code generation as a scaffold for such tasks [9, 13, 25, 34]. Aside from code generation, prior work has also shown that decomposing complex queries into simpler ones improves the performance of language models [28, 42]. Our RVP approach applies this insight to improve the code generation component of visual programming.", + "bbox": [ + 212, + 734, + 789, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "Ge et al.", + "bbox": [ + 271, + 114, + 331, + 127 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/289e4bc1dc3f4c4c4b0e6b14f23b1457f7727744791a3995762557724aba68b7.jpg", + "image_caption": [ + "Fig. 4: An overview of our approach. When a recursive_query API call is occurred, the model generates and executes new code based on the query, returning the output back to the original call." + ], + "image_footnote": [], + "bbox": [ + 222, + 143, + 782, + 397 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3 Recursive Visual Programming", + "text_level": 1, + "bbox": [ + 214, + 477, + 553, + 494 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we introduce our method, Recursive Visual Programming (RVP). We first review the problem setup and ingredients of visual programming as established by prior work (see Section 3.1). We then introduce our recursive coding framework (see Section 3.2), which involves a multi-step prompting scheme, as well as a system for dynamic type assignment (see Section 3.3).", + "bbox": [ + 212, + 507, + 787, + 584 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1 Preliminaries", + "text_level": 1, + "bbox": [ + 214, + 604, + 372, + 619 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In Visual Question Answering (VQA), the system $f$ is given a visual input $I$ (i.e. an image, a set of images, or a video), as well as a question $Q$ relating to the visual content. The expected output is a textual answer $A$ to the question:", + "bbox": [ + 214, + 628, + 787, + 675 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nA \\leftarrow f (I, Q) \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 452, + 685, + 785, + 703 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Visual Programming methods [14,30,31] represent $f$ as the composition of two functions: a code generator $g$ and a visual executor $h$ . In the first part, the code generator generates a program $P$ conditioned on the question, and in the second part, the executor executes the program on the visual input:", + "bbox": [ + 212, + 710, + 787, + 772 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nP \\leftarrow g (Q), A \\leftarrow h (P, I) \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 415, + 782, + 785, + 800 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We focus on the few-shot setting in which we have only a few (at most 50) training examples—each consisting of a question, program, and answer—for a given", + "bbox": [ + 214, + 809, + 785, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "Recursive Visual Programming", + "bbox": [ + 524, + 114, + 730, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "task. Therefore, the code generator $g$ is composed of a language model and a prompt that includes a description of the API available to the executor and in-context examples of questions along with the corresponding correct programs. We use Python as the programming language, similar to [30, 31].", + "bbox": [ + 212, + 146, + 782, + 205 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We evaluate our approach using the ViperGPT [31] API. For questions over images or sets of images, these functions can be called on an ImagePatch object, which can be any crop of the image:", + "bbox": [ + 212, + 208, + 782, + 253 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. find(object: str): Returns a list of ImagePatch objects matching the description.", + "2. exists(object: str): Returns a bool representing whether the described object exists.", + "3. verify_property(object: str, property: str): Returns a bool representing whether the described object has the specified property.", + "4. simple_query(question: str): Returns an answer to the provided question based on the given ImagePatch.", + "5. compute_depth(): Returns the median depth of the region in the given ImagePatch.", + "6. crop(left: int, lower: int, right: int, upper: int): Crops a part of the given ImagePatch and returns a new ImagePatch representing that crop." + ], + "bbox": [ + 222, + 268, + 785, + 472 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For questions about a video, the API includes three other functions: trim(), frame_from_index(), and frame_iterator(), which return an abbreviated version of the video, a specific frame, and an iterator over the frames respectively.", + "bbox": [ + 212, + 488, + 782, + 534 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.2 Recursive Visual Programming", + "text_level": 1, + "bbox": [ + 214, + 564, + 519, + 579 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We introduce a new method for code generation in visual programming: Recursive Visual Programming (RVP). RVP enables the language model to break the code generation process into multiple sub-questions and recursively call itself to generate a program for each sub-question. As shown in Figure 4, we operationalize RVP by adding a new function to the API: recursive_query(image, sub-question). During execution, when recursive_query is called, it prompts the language model to generate a new code given the sub-question. This code is then executed, and its return value is returned by recursive_query. An example break down of the usage of recursive_query call is shown in Figure 1. We set the recursive termination condition by (1) setting max depth of 10 and (2) switching to direct simple query when generated code contains a recursive call to the same input question.", + "bbox": [ + 212, + 595, + 785, + 777 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To enable this process, we add examples of calling recursive_query in the prompt to show the model how to decompose the problem and utilize the answers returned by recursive_query. We provide examples of the in-context prompt of recursive_query we used in the supplementary material.", + "bbox": [ + 212, + 779, + 785, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "Ge et al.", + "bbox": [ + 271, + 114, + 331, + 127 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/e847bf6c00ebf60a2f11847f5d049b303e64d5069766beb8bd8f16965ff7d17b.jpg", + "image_caption": [ + "Fig.5: Examples of dynamic type assignment in recursive_query calls. The model autonomously determines the appropriate return type in its code generation and generalize to new types unseen in in-context examples such as List[ImagePatch]." + ], + "image_footnote": [], + "bbox": [ + 223, + 146, + 779, + 363 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.3 Dynamic Type Assignment", + "text_level": 1, + "bbox": [ + 215, + 455, + 485, + 472 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In previous visual programming frameworks, code generation mechanisms are constrained by static type assignments, i.e., the return type of a given function is fixed. For instance, simple_query always returns a string. However, this rigid structure limits the model's flexibility to adapt the code to different contexts within the VQA task. RVP introduces dynamic type assignment to address this limitation. When the code generator writes a recursive_query function, it can include the expected return type as a part of the question that is passed to the recursive_query. To facilitate this, in the in-context examples that include recursive_query, we let the input 'question' specify a type. As shown in Figure 5, we give examples of type ImagePatch, bool, float, and str being specified within the queries. When receiving this question, the model discerns the required type and generates code that declares and returns the correct type. For instance, a query like \"Return a bool, is there a black cat in the picture?\" will prompt the generation of code with the signature.", + "bbox": [ + 212, + 507, + 787, + 719 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {e x e c u t e} (\\operatorname {i m a g e}) \\rightarrow \\operatorname {b o o l} \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 415, + 741, + 785, + 758 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "This signature indicates that the generated code $P'$ , when executed, will return a boolean value $A_{\\mathrm{bool}}$ which answers the sub-question $Q'$ , illustrating the dynamic and context-sensitive nature of RVP's code generation capabilities that help to produce syntactically and semantically appropriate code for complex reasoning.", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "Recursive Visual Programming", + "bbox": [ + 524, + 114, + 730, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 215, + 143, + 375, + 162 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this section, we introduce the basic settings including the task formulation and parameter details. Then, we provide experiment results and in-depth analysis.", + "bbox": [ + 215, + 175, + 784, + 205 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.1 Experiment Setup", + "text_level": 1, + "bbox": [ + 215, + 215, + 413, + 231 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Datasets. We evaluate on four datasets spanning a wide range of visual reasoning skills and types of visual input: Visual Spatial Reasoning (VSR) [22], GQA [17], COVR [5], and NextQA [36]. For NextQA, we present results on the \"Hard Temporal\" subset which was curated to consist of questions that truly require temporal reasoning [7]. VSR and GQA include queries about single images, while COVR includes queries about multiple images, and NextQA includes queries about video. On each dataset, we evaluate methods by their accuracy in producing an answer that exactly matches the ground-truth answer. For NextQA, we adopt a multiple-choice setup in which we provide methods with the set of possible answer choices for each question. For VSR, a spatial statement of an image is made, and the question is framed as : \"Is it true or false,\" followed by the statement. In this way, a VSR statement is framed as a VQA question and the model will answer either true or false.", + "bbox": [ + 212, + 239, + 787, + 435 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Implementation Details. We implement RVP using the ViperGPT API and the accompanying prompt template. For each dataset, we curate a set of in-context examples which are added to the prompt. The set of question-answer pairs is shared between the recursive and non-recursive methods. We use seven in-context examples for GQA, six for NextQA, six for VSR, and eight for COVR. For code generation, use GPT-3.5-turbo via the OpenAI API1.", + "bbox": [ + 212, + 436, + 787, + 527 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.2 Quantitative Results", + "text_level": 1, + "bbox": [ + 215, + 549, + 433, + 564 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 1 presents our results across four datasets. In all settings, we outperform ViperGPT, which is identical to RVP except that RVP includes recursive_query function in the API and in-context examples. This indicates that recursive programming indeed improves accuracy in visual reasoning tasks. Moreover, our method outperforms previous zero/few-shot methods in all but one of the settings. While our primary aim was not to push the state-of-the-art results, we show that integrating the clarity and elegance of recursive coding into standard VP methods not only doesn't impair performance, but can enhance it.", + "bbox": [ + 212, + 573, + 790, + 694 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.3 Qualitative Results", + "text_level": 1, + "bbox": [ + 215, + 715, + 421, + 729 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We provide qualitative examples in Figure 3,6,7. In the example from COVR in Figure 3, the non-recursive approach answers with neither \"yes\" nor \"no,\" while the recursive approach does confirm to these choices. One possible explanation for this issue is that LLMs struggle to maintain coherence across long contexts [37]. RVP circumvents this issue by pushing the LLM to commit to the answer", + "bbox": [ + 215, + 739, + 787, + 816 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "Ge et al.", + "bbox": [ + 271, + 114, + 331, + 127 + ], + "page_idx": 7 + }, + { + "type": "page_footnote", + "text": "1 wwwPLATFORM.openai.com", + "bbox": [ + 217, + 824, + 408, + 839 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/496ab97313fb4e20d67846ce56a2339f9cdcdbe1199030901a99f073198dd51d.jpg", + "table_caption": [ + "Table 1: Results. We report exact-match accuracy on each dataset. Best zero/few-shot result for each dataset is bolded. †Fully supervised results are from different models: ViLT [19] for VSR, HiTeA [39] for NextQA, VinVL-Base [41] for GQA, and VisualBERT [20] for COVR. ‡We report the results of our reproduction of ViperGPT using the officially released code and the same in-context examples used for our method (since the original in-context examples from ViperGPT were not released). We use GPT3.5-turbo while CodeVQA uses Codex, which is not available anymore." + ], + "table_footnote": [], + "table_body": "
MethodVSRNextQAGQA Test-devCOVR Test
Random SplitZero-shot SplitHard Split-TTest-dev
Fully supervised†69.363.048.665.157.9
Zero/few-shot methods
CLIP56.054.5---
BLIPv2---42.31-
CodeVQA---49.050.7
ViperGPT‡61.2561.5947.2144.6351.69
RVP (ours)63.5366.0948.8245.6252.67
", + "bbox": [ + 253, + 253, + 743, + 368 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "types earlier in its generation (the length from the beginning of the code to the point \"return xxx\" is shorter). In Figure 6, the non-recursive approach fails to handle the logic correctly while the recursive approach correctly does. This is likely due to that RVP breaks down the code logic and each code piece only needs to focus on a single logic, while the non-recursive method attempts to handle all the logic at once. In the GQA example, a common pattern involves handling complex logical problems and multiple properties. Non-recursive methods often fell short in these cases, either overlooking critical details or processing them in a logically flawed manner. RVP manages these intricate questions by addressing each detail accurately. This difference may stem from the inherent difficulty LLMs face in processing complicated logic over extended contexts. Recursive methods simplify this by focusing on the overall logic and gradually breaking the problems down into smaller, more manageable segments. This facilitates easier and more accurate resolution.", + "bbox": [ + 212, + 397, + 787, + 607 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We provide a dynamic type assignment example in Figure 5. In this example, RVP assigns the return type List[ImagePatch] in the recursive_query call. It's worth noting that the return type List[ImagePatch] is unseen in the in-context examples. This suggests that RVP has the potential to utilize complicated and diverse data structures, highlight the effectiveness of our approach.", + "bbox": [ + 212, + 608, + 787, + 686 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.4 Ablation Study", + "text_level": 1, + "bbox": [ + 214, + 707, + 388, + 722 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Dynamic Type This section studies integrating dynamic type assignment within the recursive_query API. For the GQA dataset, we examined all instances with recursive patterns in the test_dev split. In COVR, where the test set is unavailable, we used the val_1000 samples from the validation set, following [30].", + "bbox": [ + 212, + 733, + 787, + 794 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We compare four approaches: Non-recursive Baseline: Standard approach without recursive features. Fixed-Type Recursive: The recursive_query API is constrained to return a string. Dynamic-Implicit Recursive: The API", + "bbox": [ + 212, + 795, + 787, + 839 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "Recursive Visual Programming", + "bbox": [ + 522, + 114, + 730, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/96032daf77e87f35d0d7e22e44648709dd0d14974fd7eb030569176dd9312844.jpg", + "image_caption": [ + "Fig. 6: Examples from COVR. RVP allows correct logic flow while traditional nonrecursive VP fails to handle the logic correctly." + ], + "image_footnote": [], + "bbox": [ + 218, + 145, + 784, + 527 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "can return various types, but the specific type is not predefined in the query. For example, \"Is there a black cat?\" could yield either a boolean or a string. Dynamic-Explicit Recursive: Our primary approach, where the API's input query specifies the return type, like \"Return a bool, is there a black cat?\"", + "bbox": [ + 212, + 603, + 787, + 666 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Table 2 shows our findings. The dynamic-explicit type assignment outperformed other methods, underscoring the need for clear type specification. The dynamic-implicit method showed lower performance than non-recursive due to type ambiguities, as exemplified in Figure 8, where the first code piece assumes the return type is a bool, while the second code piece returns a str. Fixed-type assignment proves to be beneficial over non-recursive methods, indicating that recursion itself adds value, but its rigidity in type handling limits performance. This study highlights the benefits of flexible, dynamic type assignment and the crucial role of explicit type declaration for accurate and consistent responses. These insights mark a step towards more adaptable visual reasoning frameworks, emphasizing the need for precision in dynamic environments.", + "bbox": [ + 212, + 672, + 789, + 840 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 218, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "Ge et al.", + "bbox": [ + 271, + 114, + 331, + 126 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/68ee3982eab4da0234e24377af0769c55cfb81cf4fbef81d73669132562e818b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 223, + 143, + 779, + 348 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/7289d0e956b6012d08344c50375c505214c6fb3d6a09b6bebfa0f820f4e5b31e.jpg", + "image_caption": [ + "Fig. 7: Examples from GQA. RVP allows handling the details more elegantly and produces logically correct code." + ], + "image_footnote": [], + "bbox": [ + 223, + 349, + 779, + 583 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "In-Context Examples We conduct ablation on using retrieval-based in-context example selection and using different number of in-context examples.", + "bbox": [ + 212, + 657, + 784, + 688 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Retrieval-based in-context example selection. We manage to write 9 programs using the recursive method from the 50 provided by CodeVQA [30], and add our original 3 recursive programs to them. We keep the non-recursive in-context example unchanged and choose 3 recursive in-context examples from these 12 for each question. We follow [35] to use the embedding-based retrieval technique that calculates a score for each example, and then select the top 3 examples for each question. The results of GQA_val_2000 are in Table 2. We find that the recursive pattern occurrence rate is significantly higher than using fixed in-context examples. This is likely because by retrieving similar recursively-decomposed questions, the model can more easily simulate their pat", + "bbox": [ + 212, + 689, + 787, + 840 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "Recursive Visual Programming", + "bbox": [ + 524, + 114, + 730, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 116, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "terns, thereby attempting to solve more problems recursively. However, the overall accuracy does not improve. This suggests that although the retrieval-based method encourages more diverse and suitable patterns of recursive decomposition, there is a need for improved techniques in selecting the example pool—such as focusing on those most suitable for recursive decomposition instead of randomly selecting—along with careful tuning of the example programs.", + "bbox": [ + 212, + 146, + 787, + 236 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Number of in-context examples. We randomly add one example/delete one example from the covr_1000_val, the results are shown in Table 2. The results indicate that adding the number of examples can enhance the performance, but does not have a significant impact.", + "bbox": [ + 212, + 237, + 787, + 297 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Error feedback loop Recent studies have studied the debugging ability of LLMs on code [10]. We study whether modular visual code structure enables better bug identification for visual programming. We sample 20 incorrect programs for 10 questions that have type/logic errors, and use GPT3.5 to (1) identify the bug, and (2) write the correct code given the bug. For identification, the model correctly finds 7 bugs correctly for RVP and 2 for non-recursive visual programming. For correction, the model fixes 7 bugs correctly for RVP and 7 for visual programming. This suggests that clearer and more modular code structure, as presented RVP, can potentially aids bug identification.", + "bbox": [ + 212, + 311, + 787, + 448 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Table 2: Ablation study. (a) We provide ablation study results for dynamic type assignment and found that specified dynamic type assignment in RVP achieves the highest performance. (b) We provide results on using retrieval-based in-context examples and fixed examples. Retrieval-based method enables more diverse recursive patterns but requires careful choosing the prompt pool. (c) We provide results of using different number of examples and found this doesn't influence performance significantly.", + "bbox": [ + 212, + 459, + 787, + 544 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/c7572b10aeb5498e15731da1d5b06aaab0a3a3cd9bc31fcf04faf953bd090d35.jpg", + "table_caption": [ + "(a) Dynamic Type Ablation." + ], + "table_footnote": [], + "table_body": "
MethodGQACOVR
Non-Recursive55.9960.71
Fixed-Type65.0560.12
Dynamic-Implicit55.3453.57
Dynamic-Explicit70.2367.86
", + "bbox": [ + 225, + 584, + 439, + 670 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/893966c64d7c77f5fe3a0a53d68e6a634e877000ec23d7becc5181806f12162d.jpg", + "table_caption": [ + "(b) Retrieval-Based and Fixed Examples" + ], + "table_footnote": [], + "table_body": "
FixedRetrieval
Accuracy54.7553.79
Rec Rate3.55%14.85%
Rec Acc71.8357.91
", + "bbox": [ + 455, + 595, + 630, + 670 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/bebc2669a832fa6f24c63efb460af383106230a1eed24bcad68d5b818ea4993b.jpg", + "table_caption": [ + "(c) In-Context Example Num" + ], + "table_footnote": [], + "table_body": "
MethodAccuracy
Del 150.65
Original51.05
Add 151.35
", + "bbox": [ + 650, + 597, + 777, + 670 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "4.5 Readability Study", + "text_level": 1, + "bbox": [ + 214, + 695, + 413, + 710 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "This study investigates how recursive coding impacts code readability and comprehension for human programmers. Our goal is to move from complex code structures to more modular, clear, and understandable code. To evaluate this, we focus on code quality in terms of its understandability.", + "bbox": [ + 212, + 719, + 785, + 779 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "From the COVR dataset, we randomly select one example from each five different question types where recursive patterns occur. We present the main function of both their recursive and non-recursive code versions to 15 participants and ask them to grasp the high-level intent of the code as quickly as", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "Ge et al.", + "bbox": [ + 271, + 114, + 331, + 127 + ], + "page_idx": 11 + }, + { + "type": "code", + "sub_type": "code", + "code_caption": [ + "Return Type Confusion: Expect Bool but Actually Str." + ], + "code_body": "def execute(image) -> str: image_batch = ImagePatch(image) is.black_skateboard= image_batchrecursive_query(\"Are there any black skateboards?\") is_motorcycle $=$ image_batchrecursive_query(\"Arethereanymotorcycles?\") return bool to yesno(is.black Skateboard or is_motorcycle) \ndef execute(image) -> str: image_batch $=$ ImagePatch(image) skateboard_patches $=$ image_batch.find(\"skateboard\") for skateboard_batch in skateboard_patches: If skateboard_batchverified_property(\"skateboard\", \"black\"): return \"yes\" return \"no\"", + "guess_lang": "python", + "bbox": [ + 222, + 143, + 486, + 300 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/5d1427f2a49da9b18a5f78c59270b89ceb435daaf18ee769d6a2d32c42e9e657.jpg", + "image_caption": [ + "Fig.8: Failure example of dynamic type assignment due to unspecified type, The model confuses the type and return a str instead of a bool.", + "Fig.9: Comparison of average understanding time for recursive and non-recursive codes. RVP demonstrates better readability and is easy to understand." + ], + "image_footnote": [], + "bbox": [ + 504, + 156, + 781, + 313 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "possible. For the non-recursive case, the main function is just the single code snippet generated by the LLM. For the recursive case, the main function consists of the code generated by the first LLM call (the high level code) but not the functions generated by the recursive queries. This helps us measure how fast users can understand the code's intention by looking at the main function, and reflects the self-explanatory nature of the code. Figure 9 shows the results. Participants consistently understand the recursive code faster than the non-recursive versions. This quicker comprehension can be attributed to the recursive code's structured and clear approach. Non-recursive code, while correct, sometimes lacks immediate clarity in conveying its primary goal.", + "bbox": [ + 212, + 415, + 787, + 566 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "We also calculate the average line length for all main and sub functions. For RVP, the average length of all functions is 8.47. For non-recursive VP, the average length of the functions is 15.4. This suggests that RVP decreases the average length the code pieces and enhances modularity.", + "bbox": [ + 212, + 566, + 787, + 626 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Our findings suggest that recursive coding not only simplifies complex code but also enhances its accessibility and ease of understanding. Its orderly structure and clear logic accelerate comprehension, leading to more efficient and user-friendly programming. We provide more details of this study in the supp.", + "bbox": [ + 212, + 627, + 787, + 688 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "4.6 Recursive Analysis", + "text_level": 1, + "bbox": [ + 215, + 709, + 419, + 724 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Question Type. We study which types of questions tend to use recursive and found that recursive patterns appeared in five question types within both the COVR test set and the GQA test-dev set, as shown in Table 3. The 'Verify Property' and 'Query' types in GQA, and 'Compare' and 'Multi-Reference' in COVR, were most common. This suggests recursive methods tend to occur in tasks that require comparative and referential analysis.", + "bbox": [ + 212, + 733, + 787, + 824 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Dynamic Type. We study which types of dynamic return types occurs and", + "bbox": [ + 215, + 824, + 785, + 840 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "Recursive Visual Programming", + "bbox": [ + 524, + 114, + 730, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 784, + 126 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Table 3: Recursive analysis. Table (a) provides the distribution of recursive patterns by question type in GQA and COVR. We have found that recursive patterns occur commonly in question types that require reasoning. Table (b) studies different recursive query return types and their frequency in GQA and COVR. The model generalizes to unseen types. Table (c) compares accuracy across question types. Recursive methods outperform non-recursive particularly in 'Choose' and 'Query' questions.", + "bbox": [ + 212, + 143, + 787, + 229 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "(a) Dist. of Recursive Patterns", + "bbox": [ + 228, + 238, + 434, + 253 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "(b) Diverse Return Types", + "bbox": [ + 442, + 238, + 617, + 253 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "(c) Type Accuracy", + "bbox": [ + 635, + 238, + 764, + 253 + ], + "page_idx": 13 + }, + { + "type": "table", + "img_path": "images/c3ec4fc02156296064b0cc19376ec9964bd0c0a61d3507bb99e8ff47e76d118c.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
COVR/GQAQuestion Types
COVRGQA
Compare45715
Mult. Ref282-
Quant. Attr12-
Spec Attr357-
Choose125
Verify-68
Query-167
Logical-34
", + "bbox": [ + 232, + 263, + 431, + 388 + ], + "page_idx": 13 + }, + { + "type": "table", + "img_path": "images/c64a9c635a2dca6e0741c5521f4b7e3ea2d55902cf6856c7c7b06de6f69fa8b1.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Data TypeGQACOVR
Str3191828
Bool521592
List[str]16-
ImagePatch30-
List[ImgPatch]15
Float6-
Int-914
", + "bbox": [ + 441, + 265, + 616, + 362 + ], + "page_idx": 13 + }, + { + "type": "table", + "img_path": "images/ddc4faaf47e79e153aacf695e7aeb7a74d5690893d1cf14b51e72d0894afff2a.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
QuestionAccuracy (%)
Non-RecRec
Verify52.9464.71
Query56.8971.86
Logical55.8870.59
Compare66.6753.33
Choose52.0084.00
", + "bbox": [ + 620, + 265, + 779, + 359 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "their frequently in GQA and COVR. In Table 3, we found that RVP generalizes to new types like List[str] and List[ImagePatch], which are not included in the in-context examples. This flexibility and ability to generalize beyond in-context examples highlight the inherent adaptability of recursive methods.", + "bbox": [ + 212, + 398, + 785, + 458 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Type Accuracy. We evaluate the accuracy of recursive methods across different question types. An analysis of the GQA test-dev set is presented in Table 3. It reveals that recursive methods significantly improve accuracy in 'Query', 'Logical', and 'Choose' questions. This underscores their suitability for complex reasoning and decision-making tasks.", + "bbox": [ + 212, + 459, + 785, + 534 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Error Source. We sample 50 incorrect recursive examples from GQA2000val and manually check the error source: $22\\%$ are caused by logic error inside subfunctions; $18\\%$ are caused by logic error between functions; $52\\%$ are caused by API model inaccuracy (i.e. detection model error); and $4\\%$ are caused by other problems like the groundtruth answer unclear.", + "bbox": [ + 212, + 535, + 785, + 609 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Cost And Runtime. For the questions which require recursive query, RVP takes about 2.1 times (runtime) and about 3 times (cost) compared to ViperGPT.", + "bbox": [ + 212, + 609, + 785, + 641 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Conclusion and Limitation", + "text_level": 1, + "bbox": [ + 214, + 657, + 511, + 672 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In this work, we present RVP, a visual programming method that adopts recursive coding to generate concise, elegant and easy-to-understand code, handling details and intricate logic more accurately compared with traditional visual programming methods. We conduct extensive experiments across various benchmarks and provide a comprehensive analysis of RVP. Our work is the first to explore recursive programming in the visual domain and dynamic-type code generation for visual reasoning; we believe the recursive programming concept would have a potential greater impact on other domains as well. While RVP is effective for complicated visual questions, most questions in current VQA benchmarks tend to be short and non-recursive. We believe that the benefit of our method could be demonstrated better with a new benchmark or a more complex split.", + "bbox": [ + 212, + 674, + 787, + 840 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "Ge et al.", + "bbox": [ + 271, + 114, + 331, + 127 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 217, + 143, + 401, + 162 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "We would like to thank Ben Bogin for helping us test on the private COVR test set, and Guangyuan Jiang for helping us proof read the paper and providing valuable feedback.", + "bbox": [ + 212, + 175, + 785, + 219 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 243, + 321, + 258 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Andreas, J., Rohrbach, M., Darrell, T., Klein, D.: Neural module networks. 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR) pp. 39-48 (2015), https://api-semanticscholar.org/CorpusID:5276660", + "2. Andreas, J., Rohrbach, M., Darrell, T., Klein, D.: Learning to compose neural networks for question answering. ArXiv abs/1601.01705 (2016), https://apisemantic scholar.org/CorpusID:3130692", + "3. Antol, S., Agrawal, A., Lu, J., Mitchell, M., Batra, D., Zitnick, C.L., Parikh, D.: Vqa: Visual question answering. In: Proceedings of the IEEE international conference on computer vision. pp. 2425-2433 (2015)", + "4. Besta, M., Blach, N., Kubicek, A., Gerstenberger, R., Podstawski, M., Gianinazzi, L., Gajda, J., Lehmann, T., Niewiadomski, H., Nczyk, P., Hoefler, T.: Graph of thoughts: Solving elaborate problems with large language models (2024)", + "5. Bogin, B., Gupta, S., Gardner, M., Berant, J.: Covr: A test-bed for visually grounded compositional generalization with real images. ArXiv abs/2109.10613 (2021), https://api.sementicscholar.org/CorpusID:237592834", + "6. Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J.D., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., Agarwal, S., Herbert-Voss, A., Krueger, G., Henighan, T., Child, R., Ramesh, A., Ziegler, D., Wu, J., Winter, C., Hesse, C., Chen, M., Sigler, E., Litwin, M., Gray, S., Chess, B., Clark, J., Berner, C., McCandlish, S., Radford, A., Sutskever, I., Amodei, D.: Language models are few-shot learners. In: Larochelle, H., Ranzato, M., Hadsell, R., Balcan, M., Lin, H. (eds.) Advances in Neural Information Processing Systems. vol. 33, pp. 1877-1901. Curran Associates, Inc. (2020), https://proceedings.neurips.cc/paper/2020/file/1457c0d6bfb4967418bf8ac142f64a-Paper.pdf", + "7. Buch, S., Eyzaguirre, C., Gaidon, A., Wu, J., Fei-Fei, L., Niebles, J.C.: Revisiting the \"Video\" in Video-Language Understanding. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2022)", + "8. Chen, M., Tworek, J., Jun, H., Yuan, Q., Ponde, H., Kaplan, J., Edwards, H., Burda, Y., Joseph, N., Brockman, G., Ray, A., Puri, R., Krueger, G., Petrov, M., Khlaaf, H., Sastry, G., Mishkin, P., Chan, B., Gray, S., Ryder, N., Pavlov, M., Power, A., Kaiser, L., Bavarian, M., Winter, C., Tillet, P., Such, F.P., Cummings, D.W., Plappert, M., Chantzis, F., Barnes, E., Herbert-Voss, A., Guss, W.H., Nichol, A., Babuschkin, I., Balaji, S.A., Jain, S., Carr, A., Leike, J., Achiam, J., Misra, V., Morikawa, E., Radford, A., Knight, M.M., Brundage, M., Murati, M., Mayer, K., Welinder, P., McGrew, B., Amodei, D., McCandlish, S., Sutskever, I., Zaremba, W.: Evaluating large language models trained on code. ArXiv abs/2107.03374 (2021), https://api-semanticscholar.org/CorpusID: 235755472", + "9. Chen, W., Ma, X., Wang, X., Cohen, W.W.: Program of thoughts prompting: Disentangling computation from reasoning for numerical reasoning tasks. Transactions on Machine Learning Research (2023)" + ], + "bbox": [ + 225, + 273, + 785, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "Recursive Visual Programming", + "bbox": [ + 522, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "10. Chen, X., Lin, M., Scharli, N., Zhou, D.: Teaching large language models to self-debug. arXiv preprint arXiv:2304.05128 (2023)", + "11. Cheng, Z., Xie, T., Shi, P., Li, C., Nadkarni, R., Hu, Y., Xiong, C., Radev, D.R., Ostendorf, M., Zettlemoyer, L., Smith, N.A., Yu, T.: Binding language models in symbolic languages. ArXiv abs/2210.02875 (2022), https://api-semanticscholar.org/CorpusID:252734772", + "12. Cho, J., Zala, A., Bansal, M.: Visual programming for text-to-image generation and evaluation. NeurIPS (2023)", + "13. Gao, L., Madaan, A., Zhou, S., Alon, U., Liu, P., Yang, Y., Callan, J., Neubig, G.: Pal: Program-aided language models. arXiv preprint arXiv:2211.10435 (2022)", + "14. Gupta, T., Kembhavi, A.: Visual programming: Compositional visual reasoning without training. 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) pp. 14953-14962 (2022), https://api-semanticscholar.org/CorpusID:253734854", + "15. Hu, R., Andreas, J., Rohrbach, M., Darrell, T., Saenko, K.: Learning to reason: End-to-end module networks for visual question answering. 2017 IEEE International Conference on Computer Vision (ICCV) pp. 804-813 (2017), https://api_semanticscholar.org/CorpusID:18682", + "16. Hudson, D.A., Manning, C.D.: Gqa: A new dataset for real-world visual reasoning and compositional question answering. 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) pp. 6693-6702 (2019), https://apisemantic scholar.org/CorpusID:152282269", + "17. Hudson, D.A., Manning, C.D.: Gqa: A new dataset for real-world visual reasoning and compositional question answering. 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) pp. 6693-6702 (2019), https://apisemantic scholar.org/CorpusID:152282269", + "18. Hudson, D.A., Manning, C.D.: Learning by abstraction: The neural state machine. In: Neural Information Processing Systems (2019), https://api-semanticscholar.org/CorpusID:195847902", + "19. Kim, W., Son, B., Kim, I.: Vilt: Vision-and-language transformer without convolution or region supervision. In: International Conference on Machine Learning (2021), https://api-semanticscholar.org/CorpusID:231839613", + "20. Li, L.H., Yatskar, M., Yin, D., Hsieh, C.J., Chang, K.W.: Visualbert: A simple and performant baseline for vision and language. ArXiv abs/1908.03557 (2019), https://api_semanticscholar.org/CorpusID:199528533", + "21. Liang, J., Huang, W., Xia, F., Xu, P., Hausman, K., Ichter, B., Florence, P., Zeng, A.: Code as policies: Language model programs for embodied control. In: arXiv preprint arXiv:2209.07753 (2022)", + "22. Liu, F., Emerson, G.E.T., Collier, N.: Visual spatial reasoning. Transactions of the Association for Computational Linguistics 11, 635-651 (2022), https://apisemantic scholar.org/CorpusID:248496506", + "23. Lu, P., Peng, B., Cheng, H., Galley, M., Chang, K.W., Wu, Y.N., Zhu, S.C., Gao, J.: Chameleon: Plug-and-play compositional reasoning with large language models. arXiv preprint arXiv:2304.09842 (2023)", + "24. Ma, Y.J., Liang, W., Wang, G., Huang, D.A., Bastani, O., Jayaraman, D., Zhu, Y., Fan, L., Anandkumar, A.: Eureka: Human-level reward design via coding large language models. arXiv preprint arXiv: Arxiv-2310.12931 (2023)", + "25. Madaan, A., Zhou, S., Alon, U., Yang, Y., Neubig, G.: Language models of code are few-shot commonsense learners. In: Goldberg, Y., Kozareva, Z., Zhang, Y. (eds.) Proceedings of the 2022 Conference on Empirical Methods in Natural Language" + ], + "bbox": [ + 215, + 146, + 785, + 840 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "Ge et al.", + "bbox": [ + 271, + 114, + 330, + 126 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Processing. pp. 1384-1403. Association for Computational Linguistics, Abu Dhabi, United Arab Emirates (Dec 2022). https://doi.org/10.18653/v1/2022.emnlp-main.90, https://aclanthology.org/2022.emnlp-main.90", + "26. Marino, K., Rastegari, M., Farhadi, A., Mottaghi, R.: Ok-vqa: A visual question answering benchmark requiring external knowledge. 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) pp. 3190-3199 (2019), https://api-semanticscholar.org/CorpusID:173991173", + "27. OpenAI: Gpt-4 technical report. ArXiv abs/2303.08774 (2023), https://apisemantic scholar.org/CorpusID:257532815", + "28. Press, O., Zhang, M., Min, S., Schmidt, L., Smith, N.A., Lewis, M.: Measuring and narrowing the compositionality gap in language models. ArXiv abs/2210.03350 (2022), https://api-semanticscholar.org/CorpusID:252762102", + "29. Subramanian, S., Bogin, B., Gupta, N., Wolfson, T., Singh, S., Berant, J., Gardner, M.: Obtaining faithful interpretations from compositional neural networks. In: Annual Meeting of the Association for Computational Linguistics (2020), https://apisemantic scholar.org/CorpusID:218487535", + "30. Subramanian, S., Narasimhan, M., Khangaonkar, K., Yang, K., Nagrani, A., Schmid, C., Zeng, A., Darrell, T., Klein, D.: Modular visual question answering via code generation. In: Rogers, A., Boyd-Graber, J., Okazaki, N. (eds.) Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers). pp. 747-761. Association for Computational Linguistics, Toronto, Canada (Jul 2023). https://doi.org/10.18653/v1/2023.acl-short.65, https://aclanthology.org/2023.acl-short.65", + "31. Sur'is, D., Menon, S., Vondrick, C.: Vipergpt: Visual inference via python execution for reasoning. ArXiv abs/2303.08128 (2023), https://api_semanticscholar.org/CorpusID:257505358", + "32. Suzgun, M., Scales, N., Scharli, N., Gehrmann, S., Tay, Y., Chung, H.W., Chowdhery, A., Le, Q.V., Chi, E.H., Zhou, D., et al.: Challenging big-bench tasks and whether chain-of-thought can solve them. arXiv preprint arXiv:2210.09261 (2022)", + "33. Touvron, H., Lavril, T., Izacard, G., Martinet, X., Lachaux, M.A., Lacroix, T., Rozière, B., Goyal, N., Hambro, E., Azhar, F., Rodriguez, A., Joulin, A., Grave, E., Lample, G.: Llama: Open and efficient foundation language models. ArXiv abs/2302.13971 (2023), https://apisemanticscholar.org/CorpusID: 257219404", + "34. Wang, X., Li, S., Ji, H.: Code4struct: Code generation for few-shot structured prediction from natural language. arXiv preprint arXiv:2210.12810 (2022)", + "35. Wang, Z., Li, M., Xu, R., Zhou, L., Lei, J., Lin, X., Wang, S., Yang, Z., Zhu, C., Hoiem, D., et al.: Language models with image descriptors are strong few-shot video-language learners. Advances in Neural Information Processing Systems 35, 8483-8497 (2022)", + "36. Xiao, J., Shang, X., Yao, A., Chua, T.S.: Next-qa: Next phase of question-answering to explaining temporal actions. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 9777–9786 (June 2021)", + "37. Yang, K., Klein, D., Peng, N., Tian, Y.: Doc: Improving long story coherence with detailed outline control. In: Annual Meeting of the Association for Computational Linguistics (2023), https://api-semanticscholar.org/CorpusID:254877751", + "38. Yao, S., Yu, D., Zhao, J., Shafran, I., Griffiths, T.L., Cao, Y., Narasimhan, K.: Tree of thoughts: Deliberate problem solving with large language models. arXiv preprint arXiv:2305.10601 (2023)" + ], + "bbox": [ + 215, + 146, + 785, + 839 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "Recursive Visual Programming", + "bbox": [ + 522, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 784, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "39. Ye, Q., Xu, G., Yan, M., Xu, H., Qian, Q., Zhang, J., Huang, F.: Hitea: Hierarchical temporal-aware video-language pre-training. ArXiv abs/2212.14546 (2022), https://api-semanticscholar.org/CorpusID:255340506", + "40. Yu, W., Gileadi, N., Fu, C., Kirmani, S., Lee, K.H., Gonzalez Arenas, M., Lewis Chiang, H.T., Erez, T., Hasenclever, L., Humplik, J., Ichter, B., Xiao, T., Xu, P., Zeng, A., Zhang, T., Heess, N., Sadigh, D., Tan, J., Tassa, Y., Xia, F.: Language to rewards for robotic skill synthesis. Arxiv preprint arXiv:2306.08647 (2023)", + "41. Zhang, P., Li, X., Hu, X., Yang, J., Zhang, L., Wang, L., Choi, Y., Gao, J.: Vinvl: Revisiting visual representations in vision-language models. 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) pp. 5575-5584 (2021), https://api_semanticscholar.org/CorpusID:235692795", + "42. Zhou, D., Scharli, N., Hou, L., Wei, J., Scales, N., Wang, X., Schuurmans, D., Bousquet, O., Le, Q., hsin Chi, E.H.: Least-to-most prompting enables complex reasoning in large language models. ArXiv abs/2205.10625 (2022), https://api_semanticscholar.org/CorpusID:248986239" + ], + "bbox": [ + 215, + 146, + 785, + 369 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "Ge et al.", + "bbox": [ + 271, + 114, + 331, + 127 + ], + "page_idx": 17 + } +] \ No newline at end of file diff --git a/2024/Recursive Visual Programming/e57bdabb-77ec-4030-98f0-ae871f741b2d_model.json b/2024/Recursive Visual Programming/e57bdabb-77ec-4030-98f0-ae871f741b2d_model.json new file mode 100644 index 0000000000000000000000000000000000000000..d2c73467444a5a4eb8cf43b9d06afaedb039b993 --- /dev/null +++ b/2024/Recursive Visual Programming/e57bdabb-77ec-4030-98f0-ae871f741b2d_model.json @@ -0,0 +1,2249 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.323, + 0.142, + 0.681, + 0.163 + ], + "angle": 0, + "content": "Recursive Visual Programming" + }, + { + "type": "text", + "bbox": [ + 0.318, + 0.189, + 0.684, + 0.221 + ], + "angle": 0, + "content": "Jiaxin Ge1 Sanjay Subramanian1 Baifeng Shi1 Roei Herzig1 Trevor Darrell1" + }, + { + "type": "text", + "bbox": [ + 0.422, + 0.233, + 0.581, + 0.246 + ], + "angle": 0, + "content": "UC Berkeley, CA, USA" + }, + { + "type": "text", + "bbox": [ + 0.263, + 0.282, + 0.74, + 0.531 + ], + "angle": 0, + "content": "Abstract. Visual Programming (VP) has emerged as a powerful framework for Visual Question Answering (VQA). By generating and executing bespoke code for each question, these methods show advancement in leveraging Large Language Models (LLMs) for complex problem-solving. Despite their potential, existing VP methods generate all code in a single function, which does not fully utilize LLM's reasoning capacity and the modular adaptability of code. This results in code that is suboptimal in terms of both accuracy and interpretability. Inspired by human coding practices, we propose Recursive Visual Programming (RVP), which better harnesses the reasoning capacity of LLMs, provides modular code structure between code pieces, and assigns different return types for the sub-problems elegantly. RVP approaches VQA tasks with an top-down recursive code generation approach, allowing decomposition of complicated problems into smaller parts. We show RVP's efficacy through extensive experiments on benchmarks including VSR, COVR, GQA, and NextQA, underscoring the value of adopting human-like recursive and modular programming techniques for solving VQA tasks. Our code is available at https://github.com/para-lost/RVP." + }, + { + "type": "text", + "bbox": [ + 0.263, + 0.544, + 0.663, + 0.558 + ], + "angle": 0, + "content": "Keywords: Visual Programming \\(\\cdot\\) Vision-Language \\(\\cdot\\) VQA" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.583, + 0.377, + 0.599 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.614, + 0.787, + 0.735 + ], + "angle": 0, + "content": "Visual Question Answering (VQA) lies at the intersection of computer vision and natural language processing, posing the challenge of interpreting visual data to answer questions [3,16,26]. Following the recent progress on code generation using Large Language Models (LLMs) [6,27,33], Visual Programming (VP) methods have notably advanced in this area, successfully using LLMs to generate and execute code in few-shot and zero-shot scenarios [14,30,31]. In these techniques, various vision models are provided to LLM as APIs, and the LLM is used as a planner to utilize these vision models to perform reasoning over images." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Despite these advancements, such planning remains very challenging for Visual Programming. These methods require the model to generate a single code block using the predefined APIs at once. As a result, the model must handle logic, be aware of all details, and utilize the vision APIs appropriately. This has two main drawbacks. First, such approaches do not fully utilize LLM's reasoning capability, which would benefit from breaking down the problem into step-by-step sub-problems, as demonstrated in recent NLP research [4]. Secondly, they" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.333, + 0.127 + ], + "angle": 0, + "content": "Ge et al." + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.144, + 0.784, + 0.41 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.419, + 0.788, + 0.518 + ], + "angle": 0, + "content": "Fig. 1: A breakdown of Recursive Visual Programming for a visual question is illustrated with an image where scenes within it contain smaller versions of themselves. To locate a man wearing a hat at the third level of an image, the model first generates code that has two recursive_query API calls. The initial call identifies the largest mirror (representing the next image level), and the subsequent call seeks \"the man with a hat\" at the second level. The model iteratively generates and executes code through these calls until the final answer is produced without additional recursive calls." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.55, + 0.787, + 0.581 + ], + "angle": 0, + "content": "fail to utilize the full potential of code because they lack modularity between the code pieces, particularly when the code is one long piece." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.582, + 0.787, + 0.657 + ], + "angle": 0, + "content": "We consider how humans tackle these issues. Humans tend to code in a top-down manner: first they break down complex tasks into manageable subproblems, and then they tackle each through focused code segments. This approach allows the programmer to concentrate on the main logic flow rather than figuring out all the details at once, resulting in a more clear and elegant structure." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.659, + 0.788, + 0.794 + ], + "angle": 0, + "content": "Inspired by this, we hypothesize that incorporating a similar recursive coding strategy would also help solve complex visual question-answering problems. To investigate whether LLMs could indeed benefit from recursive coding, we conduct preliminary experiment of the recursive coding approach on standard symbolic reasoning benchmarks, such as Dyck Language [32] and Games of 24 [38]. This is done by simply prompting the LLM to code recursively. We find that utilizing LLMs to code recursively leads to a notable improvement in accuracy, as shown in Figure 2. Motivated by this result, we explore in this paper how a recursive coding methodology can be applied to VQA tasks." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Specifically, we propose a novel VP method for VQA tasks, named Recursive Visual Programming (RVP). Instead of querying LLM only once to generate a single piece of code, we ask LLM to decompose the question into sub-questions" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.525, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Recursive Visual Programming" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "image", + "bbox": [ + 0.22, + 0.146, + 0.782, + 0.298 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.308, + 0.788, + 0.379 + ], + "angle": 0, + "content": "Fig. 2: Motivating Examples. Adding the prompt \"solve the problem recursively\" to the input significantly improves the model's performance on traditional coding tasks, Dyck language [32] and Games of 24 [38]. \"NR\" refers to non-recursive approach and \"R\" refers to recursive approach. This shows that LLMs can potentially benefit from human's recursive programming approach." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.394, + 0.788, + 0.53 + ], + "angle": 0, + "content": "and recursively query itself for each sub-question to generate modular code pieces. For each sub-question, the same process repeats, until reaching an atom level where the current question can be solved by only calling external APIs. See example in Figure 1. In contrast to querying LLM only once, we find that using recursive queries to generate code for each sub-question can improve the readability and accuracy of the code, as shown in Figure 3. Moreover, MVP allows assigning dynamic types to each sub-question, enabling new code pieces to return a variety of types based on the current need. Remarkably, MVP assigns return types unseen in in-context examples, such as List[str], List[ImagePatch]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.531, + 0.788, + 0.668 + ], + "angle": 0, + "content": "Our contribution can be summarized as follows: (i) We introduce a novel recursive approach based on human programming. This approach is a top-down programming method that better harnesses LLM's reasoning capacity and enhances modularity between multiple code pieces. It is easily applicable to any existing VP model for visual question answering task. (ii) We allow dynamic return types for the sub-problems, which enhances flexibility and adaptability. (iii) We demonstrate improved performance on several standard benchmarks, such as GQA, VSR, NextQA and COVR, demonstrating the effectiveness of our approach. We also show better interpretability than non-recursive VP methods." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.689, + 0.389, + 0.705 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Modular Visual Reasoning. There is a long line of work that seeks to combine modularity with the expressivity of deep neural networks. Early work on this topic converts the output of a parser [1] to a program or trains a program generator either separately from or jointly with the execution modules [2, 15]. Later work finds that such models are often challenging to train on real-world datasets, particularly when interpretability of individual module outputs is expected [29]. Other work improves performance on multi-hop visual reasoning without explicitly defining a set of modules, e.g., by predicting and operating" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.333, + 0.128 + ], + "angle": 0, + "content": "Ge et al." + }, + { + "type": "image", + "bbox": [ + 0.221, + 0.145, + 0.787, + 0.414 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.424, + 0.789, + 0.482 + ], + "angle": 0, + "content": "Fig. 3: Example from COVR (Upper). Current VP methods fail to provide binary 'yes' or 'no' answers, while Recursive Visual Programming method outputs the correct answer. Example from GQA (Middle and Bottom). Recursive outperforms current non-recursive methods by correctly addressing all details and their associated logic." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.538, + 0.388, + 0.554 + ], + "angle": 0, + "content": "over a scene graph [18]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.554, + 0.789, + 0.734 + ], + "angle": 0, + "content": "Visual Programming. Recent language models can accurately generate code from text descriptions when prompted with a small number of examples [8]. This advance has enabled a few-shot modular visual reasoning approach, commonly called visual programming [14], which consists of prompting a language model for a program and executing the program with pre-trained vision models. This paradigm has led to impressive results in visual question answering, video question answering, visual and tabular question answering, referring expression comprehension, and text-based image editing [11, 12, 23, 30, 31]. Programming with LLMs has also been applied to produce plans and reward functions for robots [21, 24, 40]. Our work builds on these techniques by enabling the LLM to generate code in multiple steps-first planning the overall process and then filling in the details-rather than all at once." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.79, + 0.842 + ], + "angle": 0, + "content": "Language Models for Reasoning Tasks. In the text domain, there is a large body of work on using language models for reasoning tasks. In particular, several papers have shown the efficacy of code generation as a scaffold for such tasks [9, 13, 25, 34]. Aside from code generation, prior work has also shown that decomposing complex queries into simpler ones improves the performance of language models [28, 42]. Our RVP approach applies this insight to improve the code generation component of visual programming." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.525, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Recursive Visual Programming" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.144, + 0.784, + 0.398 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.409, + 0.788, + 0.451 + ], + "angle": 0, + "content": "Fig. 4: An overview of our approach. When a recursive_query API call is occurred, the model generates and executes new code based on the query, returning the output back to the original call." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.478, + 0.555, + 0.496 + ], + "angle": 0, + "content": "3 Recursive Visual Programming" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.508, + 0.788, + 0.585 + ], + "angle": 0, + "content": "In this section, we introduce our method, Recursive Visual Programming (RVP). We first review the problem setup and ingredients of visual programming as established by prior work (see Section 3.1). We then introduce our recursive coding framework (see Section 3.2), which involves a multi-step prompting scheme, as well as a system for dynamic type assignment (see Section 3.3)." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.606, + 0.373, + 0.62 + ], + "angle": 0, + "content": "3.1 Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.63, + 0.788, + 0.676 + ], + "angle": 0, + "content": "In Visual Question Answering (VQA), the system \\( f \\) is given a visual input \\( I \\) (i.e. an image, a set of images, or a video), as well as a question \\( Q \\) relating to the visual content. The expected output is a textual answer \\( A \\) to the question:" + }, + { + "type": "equation", + "bbox": [ + 0.454, + 0.686, + 0.787, + 0.704 + ], + "angle": 0, + "content": "\\[\nA \\leftarrow f (I, Q) \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.712, + 0.788, + 0.773 + ], + "angle": 0, + "content": "Visual Programming methods [14,30,31] represent \\( f \\) as the composition of two functions: a code generator \\( g \\) and a visual executor \\( h \\). In the first part, the code generator generates a program \\( P \\) conditioned on the question, and in the second part, the executor executes the program on the visual input:" + }, + { + "type": "equation", + "bbox": [ + 0.416, + 0.784, + 0.787, + 0.801 + ], + "angle": 0, + "content": "\\[\nP \\leftarrow g (Q), A \\leftarrow h (P, I) \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.81, + 0.787, + 0.841 + ], + "angle": 0, + "content": "We focus on the few-shot setting in which we have only a few (at most 50) training examples—each consisting of a question, program, and answer—for a given" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.333, + 0.128 + ], + "angle": 0, + "content": "Ge et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.783, + 0.207 + ], + "angle": 0, + "content": "task. Therefore, the code generator \\( g \\) is composed of a language model and a prompt that includes a description of the API available to the executor and in-context examples of questions along with the corresponding correct programs. We use Python as the programming language, similar to [30, 31]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.209, + 0.784, + 0.255 + ], + "angle": 0, + "content": "We evaluate our approach using the ViperGPT [31] API. For questions over images or sets of images, these functions can be called on an ImagePatch object, which can be any crop of the image:" + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.27, + 0.786, + 0.3 + ], + "angle": 0, + "content": "1. find(object: str): Returns a list of ImagePatch objects matching the description." + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.302, + 0.786, + 0.331 + ], + "angle": 0, + "content": "2. exists(object: str): Returns a bool representing whether the described object exists." + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.334, + 0.786, + 0.363 + ], + "angle": 0, + "content": "3. verify_property(object: str, property: str): Returns a bool representing whether the described object has the specified property." + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.366, + 0.786, + 0.395 + ], + "angle": 0, + "content": "4. simple_query(question: str): Returns an answer to the provided question based on the given ImagePatch." + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.398, + 0.786, + 0.427 + ], + "angle": 0, + "content": "5. compute_depth(): Returns the median depth of the region in the given ImagePatch." + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.43, + 0.786, + 0.473 + ], + "angle": 0, + "content": "6. crop(left: int, lower: int, right: int, upper: int): Crops a part of the given ImagePatch and returns a new ImagePatch representing that crop." + }, + { + "type": "list", + "bbox": [ + 0.223, + 0.27, + 0.786, + 0.473 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.489, + 0.784, + 0.535 + ], + "angle": 0, + "content": "For questions about a video, the API includes three other functions: trim(), frame_from_index(), and frame_iterator(), which return an abbreviated version of the video, a specific frame, and an iterator over the frames respectively." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.565, + 0.52, + 0.58 + ], + "angle": 0, + "content": "3.2 Recursive Visual Programming" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.597, + 0.787, + 0.779 + ], + "angle": 0, + "content": "We introduce a new method for code generation in visual programming: Recursive Visual Programming (RVP). RVP enables the language model to break the code generation process into multiple sub-questions and recursively call itself to generate a program for each sub-question. As shown in Figure 4, we operationalize RVP by adding a new function to the API: recursive_query(image, sub-question). During execution, when recursive_query is called, it prompts the language model to generate a new code given the sub-question. This code is then executed, and its return value is returned by recursive_query. An example break down of the usage of recursive_query call is shown in Figure 1. We set the recursive termination condition by (1) setting max depth of 10 and (2) switching to direct simple query when generated code contains a recursive call to the same input question." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.787, + 0.841 + ], + "angle": 0, + "content": "To enable this process, we add examples of calling recursive_query in the prompt to show the model how to decompose the problem and utilize the answers returned by recursive_query. We provide examples of the in-context prompt of recursive_query we used in the supplementary material." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.525, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Recursive Visual Programming" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "image", + "bbox": [ + 0.225, + 0.147, + 0.78, + 0.364 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.375, + 0.788, + 0.418 + ], + "angle": 0, + "content": "Fig.5: Examples of dynamic type assignment in recursive_query calls. The model autonomously determines the appropriate return type in its code generation and generalize to new types unseen in in-context examples such as List[ImagePatch]." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.457, + 0.486, + 0.473 + ], + "angle": 0, + "content": "3.3 Dynamic Type Assignment" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.508, + 0.789, + 0.72 + ], + "angle": 0, + "content": "In previous visual programming frameworks, code generation mechanisms are constrained by static type assignments, i.e., the return type of a given function is fixed. For instance, simple_query always returns a string. However, this rigid structure limits the model's flexibility to adapt the code to different contexts within the VQA task. RVP introduces dynamic type assignment to address this limitation. When the code generator writes a recursive_query function, it can include the expected return type as a part of the question that is passed to the recursive_query. To facilitate this, in the in-context examples that include recursive_query, we let the input 'question' specify a type. As shown in Figure 5, we give examples of type ImagePatch, bool, float, and str being specified within the queries. When receiving this question, the model discerns the required type and generates code that declares and returns the correct type. For instance, a query like \"Return a bool, is there a black cat in the picture?\" will prompt the generation of code with the signature." + }, + { + "type": "equation", + "bbox": [ + 0.416, + 0.742, + 0.786, + 0.759 + ], + "angle": 0, + "content": "\\[\n\\operatorname {e x e c u t e} (\\operatorname {i m a g e}) \\rightarrow \\operatorname {b o o l} \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.841 + ], + "angle": 0, + "content": "This signature indicates that the generated code \\( P' \\), when executed, will return a boolean value \\( A_{\\mathrm{bool}} \\) which answers the sub-question \\( Q' \\), illustrating the dynamic and context-sensitive nature of RVP's code generation capabilities that help to produce syntactically and semantically appropriate code for complex reasoning." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.333, + 0.128 + ], + "angle": 0, + "content": "Ge et al." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.145, + 0.377, + 0.163 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.176, + 0.785, + 0.207 + ], + "angle": 0, + "content": "In this section, we introduce the basic settings including the task formulation and parameter details. Then, we provide experiment results and in-depth analysis." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.216, + 0.414, + 0.232 + ], + "angle": 0, + "content": "4.1 Experiment Setup" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.24, + 0.788, + 0.436 + ], + "angle": 0, + "content": "Datasets. We evaluate on four datasets spanning a wide range of visual reasoning skills and types of visual input: Visual Spatial Reasoning (VSR) [22], GQA [17], COVR [5], and NextQA [36]. For NextQA, we present results on the \"Hard Temporal\" subset which was curated to consist of questions that truly require temporal reasoning [7]. VSR and GQA include queries about single images, while COVR includes queries about multiple images, and NextQA includes queries about video. On each dataset, we evaluate methods by their accuracy in producing an answer that exactly matches the ground-truth answer. For NextQA, we adopt a multiple-choice setup in which we provide methods with the set of possible answer choices for each question. For VSR, a spatial statement of an image is made, and the question is framed as : \"Is it true or false,\" followed by the statement. In this way, a VSR statement is framed as a VQA question and the model will answer either true or false." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.437, + 0.788, + 0.529 + ], + "angle": 0, + "content": "Implementation Details. We implement RVP using the ViperGPT API and the accompanying prompt template. For each dataset, we curate a set of in-context examples which are added to the prompt. The set of question-answer pairs is shared between the recursive and non-recursive methods. We use seven in-context examples for GQA, six for NextQA, six for VSR, and eight for COVR. For code generation, use GPT-3.5-turbo via the OpenAI API1." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.55, + 0.434, + 0.565 + ], + "angle": 0, + "content": "4.2 Quantitative Results" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.574, + 0.791, + 0.695 + ], + "angle": 0, + "content": "Table 1 presents our results across four datasets. In all settings, we outperform ViperGPT, which is identical to RVP except that RVP includes recursive_query function in the API and in-context examples. This indicates that recursive programming indeed improves accuracy in visual reasoning tasks. Moreover, our method outperforms previous zero/few-shot methods in all but one of the settings. While our primary aim was not to push the state-of-the-art results, we show that integrating the clarity and elegance of recursive coding into standard VP methods not only doesn't impair performance, but can enhance it." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.716, + 0.422, + 0.731 + ], + "angle": 0, + "content": "4.3 Qualitative Results" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.741, + 0.788, + 0.817 + ], + "angle": 0, + "content": "We provide qualitative examples in Figure 3,6,7. In the example from COVR in Figure 3, the non-recursive approach answers with neither \"yes\" nor \"no,\" while the recursive approach does confirm to these choices. One possible explanation for this issue is that LLMs struggle to maintain coherence across long contexts [37]. RVP circumvents this issue by pushing the LLM to commit to the answer" + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.825, + 0.409, + 0.84 + ], + "angle": 0, + "content": "1 wwwPLATFORM.openai.com" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.524, + 0.115, + 0.732, + 0.13 + ], + "angle": 0, + "content": "Recursive Visual Programming" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.145, + 0.788, + 0.244 + ], + "angle": 0, + "content": "Table 1: Results. We report exact-match accuracy on each dataset. Best zero/few-shot result for each dataset is bolded. †Fully supervised results are from different models: ViLT [19] for VSR, HiTeA [39] for NextQA, VinVL-Base [41] for GQA, and VisualBERT [20] for COVR. ‡We report the results of our reproduction of ViperGPT using the officially released code and the same in-context examples used for our method (since the original in-context examples from ViperGPT were not released). We use GPT3.5-turbo while CodeVQA uses Codex, which is not available anymore." + }, + { + "type": "table", + "bbox": [ + 0.254, + 0.255, + 0.744, + 0.369 + ], + "angle": 0, + "content": "
MethodVSRNextQAGQA Test-devCOVR Test
Random SplitZero-shot SplitHard Split-TTest-dev
Fully supervised†69.363.048.665.157.9
Zero/few-shot methods
CLIP56.054.5---
BLIPv2---42.31-
CodeVQA---49.050.7
ViperGPT‡61.2561.5947.2144.6351.69
RVP (ours)63.5366.0948.8245.6252.67
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.398, + 0.789, + 0.608 + ], + "angle": 0, + "content": "types earlier in its generation (the length from the beginning of the code to the point \"return xxx\" is shorter). In Figure 6, the non-recursive approach fails to handle the logic correctly while the recursive approach correctly does. This is likely due to that RVP breaks down the code logic and each code piece only needs to focus on a single logic, while the non-recursive method attempts to handle all the logic at once. In the GQA example, a common pattern involves handling complex logical problems and multiple properties. Non-recursive methods often fell short in these cases, either overlooking critical details or processing them in a logically flawed manner. RVP manages these intricate questions by addressing each detail accurately. This difference may stem from the inherent difficulty LLMs face in processing complicated logic over extended contexts. Recursive methods simplify this by focusing on the overall logic and gradually breaking the problems down into smaller, more manageable segments. This facilitates easier and more accurate resolution." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.609, + 0.789, + 0.687 + ], + "angle": 0, + "content": "We provide a dynamic type assignment example in Figure 5. In this example, RVP assigns the return type List[ImagePatch] in the recursive_query call. It's worth noting that the return type List[ImagePatch] is unseen in the in-context examples. This suggests that RVP has the potential to utilize complicated and diverse data structures, highlight the effectiveness of our approach." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.708, + 0.39, + 0.723 + ], + "angle": 0, + "content": "4.4 Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.734, + 0.788, + 0.795 + ], + "angle": 0, + "content": "Dynamic Type This section studies integrating dynamic type assignment within the recursive_query API. For the GQA dataset, we examined all instances with recursive patterns in the test_dev split. In COVR, where the test set is unavailable, we used the val_1000 samples from the validation set, following [30]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.796, + 0.788, + 0.84 + ], + "angle": 0, + "content": "We compare four approaches: Non-recursive Baseline: Standard approach without recursive features. Fixed-Type Recursive: The recursive_query API is constrained to return a string. Dynamic-Implicit Recursive: The API" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.219, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.333, + 0.127 + ], + "angle": 0, + "content": "Ge et al." + }, + { + "type": "image", + "bbox": [ + 0.219, + 0.146, + 0.785, + 0.529 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.539, + 0.788, + 0.569 + ], + "angle": 0, + "content": "Fig. 6: Examples from COVR. RVP allows correct logic flow while traditional nonrecursive VP fails to handle the logic correctly." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.604, + 0.789, + 0.667 + ], + "angle": 0, + "content": "can return various types, but the specific type is not predefined in the query. For example, \"Is there a black cat?\" could yield either a boolean or a string. Dynamic-Explicit Recursive: Our primary approach, where the API's input query specifies the return type, like \"Return a bool, is there a black cat?\"" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.79, + 0.842 + ], + "angle": 0, + "content": "Table 2 shows our findings. The dynamic-explicit type assignment outperformed other methods, underscoring the need for clear type specification. The dynamic-implicit method showed lower performance than non-recursive due to type ambiguities, as exemplified in Figure 8, where the first code piece assumes the return type is a bool, while the second code piece returns a str. Fixed-type assignment proves to be beneficial over non-recursive methods, indicating that recursion itself adds value, but its rigidity in type handling limits performance. This study highlights the benefits of flexible, dynamic type assignment and the crucial role of explicit type declaration for accurate and consistent responses. These insights mark a step towards more adaptable visual reasoning frameworks, emphasizing the need for precision in dynamic environments." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.525, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Recursive Visual Programming" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "image", + "bbox": [ + 0.224, + 0.145, + 0.781, + 0.349 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.224, + 0.35, + 0.78, + 0.584 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.596, + 0.787, + 0.625 + ], + "angle": 0, + "content": "Fig. 7: Examples from GQA. RVP allows handling the details more elegantly and produces logically correct code." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.659, + 0.785, + 0.689 + ], + "angle": 0, + "content": "In-Context Examples We conduct ablation on using retrieval-based in-context example selection and using different number of in-context examples." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.69, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Retrieval-based in-context example selection. We manage to write 9 programs using the recursive method from the 50 provided by CodeVQA [30], and add our original 3 recursive programs to them. We keep the non-recursive in-context example unchanged and choose 3 recursive in-context examples from these 12 for each question. We follow [35] to use the embedding-based retrieval technique that calculates a score for each example, and then select the top 3 examples for each question. The results of GQA_val_2000 are in Table 2. We find that the recursive pattern occurrence rate is significantly higher than using fixed in-context examples. This is likely because by retrieving similar recursively-decomposed questions, the model can more easily simulate their pat" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.333, + 0.128 + ], + "angle": 0, + "content": "Ge et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.237 + ], + "angle": 0, + "content": "terns, thereby attempting to solve more problems recursively. However, the overall accuracy does not improve. This suggests that although the retrieval-based method encourages more diverse and suitable patterns of recursive decomposition, there is a need for improved techniques in selecting the example pool—such as focusing on those most suitable for recursive decomposition instead of randomly selecting—along with careful tuning of the example programs." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.238, + 0.788, + 0.299 + ], + "angle": 0, + "content": "Number of in-context examples. We randomly add one example/delete one example from the covr_1000_val, the results are shown in Table 2. The results indicate that adding the number of examples can enhance the performance, but does not have a significant impact." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.312, + 0.788, + 0.449 + ], + "angle": 0, + "content": "Error feedback loop Recent studies have studied the debugging ability of LLMs on code [10]. We study whether modular visual code structure enables better bug identification for visual programming. We sample 20 incorrect programs for 10 questions that have type/logic errors, and use GPT3.5 to (1) identify the bug, and (2) write the correct code given the bug. For identification, the model correctly finds 7 bugs correctly for RVP and 2 for non-recursive visual programming. For correction, the model fixes 7 bugs correctly for RVP and 7 for visual programming. This suggests that clearer and more modular code structure, as presented RVP, can potentially aids bug identification." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.46, + 0.788, + 0.545 + ], + "angle": 0, + "content": "Table 2: Ablation study. (a) We provide ablation study results for dynamic type assignment and found that specified dynamic type assignment in RVP achieves the highest performance. (b) We provide results on using retrieval-based in-context examples and fixed examples. Retrieval-based method enables more diverse recursive patterns but requires careful choosing the prompt pool. (c) We provide results of using different number of examples and found this doesn't influence performance significantly." + }, + { + "type": "table_caption", + "bbox": [ + 0.236, + 0.557, + 0.428, + 0.573 + ], + "angle": 0, + "content": "(a) Dynamic Type Ablation." + }, + { + "type": "table", + "bbox": [ + 0.226, + 0.585, + 0.44, + 0.671 + ], + "angle": 0, + "content": "
MethodGQACOVR
Non-Recursive55.9960.71
Fixed-Type65.0560.12
Dynamic-Implicit55.3453.57
Dynamic-Explicit70.2367.86
" + }, + { + "type": "table_caption", + "bbox": [ + 0.458, + 0.557, + 0.631, + 0.586 + ], + "angle": 0, + "content": "(b) Retrieval-Based and Fixed Examples" + }, + { + "type": "table", + "bbox": [ + 0.457, + 0.597, + 0.632, + 0.671 + ], + "angle": 0, + "content": "
FixedRetrieval
Accuracy54.7553.79
Rec Rate3.55%14.85%
Rec Acc71.8357.91
" + }, + { + "type": "table_caption", + "bbox": [ + 0.64, + 0.557, + 0.787, + 0.586 + ], + "angle": 0, + "content": "(c) In-Context Example Num" + }, + { + "type": "table", + "bbox": [ + 0.651, + 0.598, + 0.778, + 0.671 + ], + "angle": 0, + "content": "
MethodAccuracy
Del 150.65
Original51.05
Add 151.35
" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.696, + 0.414, + 0.712 + ], + "angle": 0, + "content": "4.5 Readability Study" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.787, + 0.78 + ], + "angle": 0, + "content": "This study investigates how recursive coding impacts code readability and comprehension for human programmers. Our goal is to move from complex code structures to more modular, clear, and understandable code. To evaluate this, we focus on code quality in terms of its understandability." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.841 + ], + "angle": 0, + "content": "From the COVR dataset, we randomly select one example from each five different question types where recursive patterns occur. We present the main function of both their recursive and non-recursive code versions to 15 participants and ask them to grasp the high-level intent of the code as quickly as" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.525, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Recursive Visual Programming" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "code", + "bbox": [ + 0.223, + 0.145, + 0.488, + 0.301 + ], + "angle": 0, + "content": "def execute(image) -> str: image_batch = ImagePatch(image) is.black_skateboard= image_batchrecursive_query(\"Are there any black skateboards?\") is_motorcycle \\(=\\) image_batchrecursive_query(\"Arethereanymotorcycles?\") return bool to yesno(is.black Skateboard or is_motorcycle) \ndef execute(image) -> str: image_batch \\(=\\) ImagePatch(image) skateboard_patches \\(=\\) image_batch.find(\"skateboard\") for skateboard_batch in skateboard_patches: If skateboard_batchverified_property(\"skateboard\", \"black\"): return \"yes\" return \"no\"" + }, + { + "type": "code_caption", + "bbox": [ + 0.218, + 0.307, + 0.49, + 0.319 + ], + "angle": 0, + "content": "Return Type Confusion: Expect Bool but Actually Str." + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.157, + 0.782, + 0.314 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.33, + 0.493, + 0.386 + ], + "angle": 0, + "content": "Fig.8: Failure example of dynamic type assignment due to unspecified type, The model confuses the type and return a str instead of a bool." + }, + { + "type": "image_caption", + "bbox": [ + 0.497, + 0.33, + 0.788, + 0.386 + ], + "angle": 0, + "content": "Fig.9: Comparison of average understanding time for recursive and non-recursive codes. RVP demonstrates better readability and is easy to understand." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.416, + 0.789, + 0.568 + ], + "angle": 0, + "content": "possible. For the non-recursive case, the main function is just the single code snippet generated by the LLM. For the recursive case, the main function consists of the code generated by the first LLM call (the high level code) but not the functions generated by the recursive queries. This helps us measure how fast users can understand the code's intention by looking at the main function, and reflects the self-explanatory nature of the code. Figure 9 shows the results. Participants consistently understand the recursive code faster than the non-recursive versions. This quicker comprehension can be attributed to the recursive code's structured and clear approach. Non-recursive code, while correct, sometimes lacks immediate clarity in conveying its primary goal." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.567, + 0.789, + 0.627 + ], + "angle": 0, + "content": "We also calculate the average line length for all main and sub functions. For RVP, the average length of all functions is 8.47. For non-recursive VP, the average length of the functions is 15.4. This suggests that RVP decreases the average length the code pieces and enhances modularity." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.628, + 0.789, + 0.689 + ], + "angle": 0, + "content": "Our findings suggest that recursive coding not only simplifies complex code but also enhances its accessibility and ease of understanding. Its orderly structure and clear logic accelerate comprehension, leading to more efficient and user-friendly programming. We provide more details of this study in the supp." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.71, + 0.42, + 0.726 + ], + "angle": 0, + "content": "4.6 Recursive Analysis" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.734, + 0.789, + 0.825 + ], + "angle": 0, + "content": "Question Type. We study which types of questions tend to use recursive and found that recursive patterns appeared in five question types within both the COVR test set and the GQA test-dev set, as shown in Table 3. The 'Verify Property' and 'Query' types in GQA, and 'Compare' and 'Multi-Reference' in COVR, were most common. This suggests recursive methods tend to occur in tasks that require comparative and referential analysis." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.825, + 0.787, + 0.841 + ], + "angle": 0, + "content": "Dynamic Type. We study which types of dynamic return types occurs and" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.333, + 0.128 + ], + "angle": 0, + "content": "Ge et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.145, + 0.788, + 0.23 + ], + "angle": 0, + "content": "Table 3: Recursive analysis. Table (a) provides the distribution of recursive patterns by question type in GQA and COVR. We have found that recursive patterns occur commonly in question types that require reasoning. Table (b) studies different recursive query return types and their frequency in GQA and COVR. The model generalizes to unseen types. Table (c) compares accuracy across question types. Recursive methods outperform non-recursive particularly in 'Choose' and 'Query' questions." + }, + { + "type": "text", + "bbox": [ + 0.229, + 0.239, + 0.436, + 0.255 + ], + "angle": 0, + "content": "(a) Dist. of Recursive Patterns" + }, + { + "type": "text", + "bbox": [ + 0.444, + 0.239, + 0.619, + 0.255 + ], + "angle": 0, + "content": "(b) Diverse Return Types" + }, + { + "type": "text", + "bbox": [ + 0.637, + 0.239, + 0.765, + 0.255 + ], + "angle": 0, + "content": "(c) Type Accuracy" + }, + { + "type": "table", + "bbox": [ + 0.233, + 0.265, + 0.432, + 0.39 + ], + "angle": 0, + "content": "
COVR/GQAQuestion Types
COVRGQA
Compare45715
Mult. Ref282-
Quant. Attr12-
Spec Attr357-
Choose125
Verify-68
Query-167
Logical-34
" + }, + { + "type": "table", + "bbox": [ + 0.442, + 0.266, + 0.617, + 0.363 + ], + "angle": 0, + "content": "
Data TypeGQACOVR
Str3191828
Bool521592
List[str]16-
ImagePatch30-
List[ImgPatch]15
Float6-
Int-914
" + }, + { + "type": "table", + "bbox": [ + 0.622, + 0.266, + 0.78, + 0.36 + ], + "angle": 0, + "content": "
QuestionAccuracy (%)
Non-RecRec
Verify52.9464.71
Query56.8971.86
Logical55.8870.59
Compare66.6753.33
Choose52.0084.00
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.399, + 0.787, + 0.459 + ], + "angle": 0, + "content": "their frequently in GQA and COVR. In Table 3, we found that RVP generalizes to new types like List[str] and List[ImagePatch], which are not included in the in-context examples. This flexibility and ability to generalize beyond in-context examples highlight the inherent adaptability of recursive methods." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.46, + 0.787, + 0.535 + ], + "angle": 0, + "content": "Type Accuracy. We evaluate the accuracy of recursive methods across different question types. An analysis of the GQA test-dev set is presented in Table 3. It reveals that recursive methods significantly improve accuracy in 'Query', 'Logical', and 'Choose' questions. This underscores their suitability for complex reasoning and decision-making tasks." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.536, + 0.787, + 0.61 + ], + "angle": 0, + "content": "Error Source. We sample 50 incorrect recursive examples from GQA2000val and manually check the error source: \\(22\\%\\) are caused by logic error inside subfunctions; \\(18\\%\\) are caused by logic error between functions; \\(52\\%\\) are caused by API model inaccuracy (i.e. detection model error); and \\(4\\%\\) are caused by other problems like the groundtruth answer unclear." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.61, + 0.787, + 0.642 + ], + "angle": 0, + "content": "Cost And Runtime. For the questions which require recursive query, RVP takes about 2.1 times (runtime) and about 3 times (cost) compared to ViperGPT." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.658, + 0.512, + 0.673 + ], + "angle": 0, + "content": "5 Conclusion and Limitation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.675, + 0.788, + 0.841 + ], + "angle": 0, + "content": "In this work, we present RVP, a visual programming method that adopts recursive coding to generate concise, elegant and easy-to-understand code, handling details and intricate logic more accurately compared with traditional visual programming methods. We conduct extensive experiments across various benchmarks and provide a comprehensive analysis of RVP. Our work is the first to explore recursive programming in the visual domain and dynamic-type code generation for visual reasoning; we believe the recursive programming concept would have a potential greater impact on other domains as well. While RVP is effective for complicated visual questions, most questions in current VQA benchmarks tend to be short and non-recursive. We believe that the benefit of our method could be demonstrated better with a new benchmark or a more complex split." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.524, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Recursive Visual Programming" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.403, + 0.163 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.176, + 0.786, + 0.22 + ], + "angle": 0, + "content": "We would like to thank Ben Bogin for helping us test on the private COVR test set, and Guangyuan Jiang for helping us proof read the paper and providing valuable feedback." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.244, + 0.323, + 0.26 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.274, + 0.786, + 0.317 + ], + "angle": 0, + "content": "1. Andreas, J., Rohrbach, M., Darrell, T., Klein, D.: Neural module networks. 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR) pp. 39-48 (2015), https://api-semanticscholar.org/CorpusID:5276660" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.318, + 0.787, + 0.357 + ], + "angle": 0, + "content": "2. Andreas, J., Rohrbach, M., Darrell, T., Klein, D.: Learning to compose neural networks for question answering. ArXiv abs/1601.01705 (2016), https://apisemantic scholar.org/CorpusID:3130692" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.358, + 0.786, + 0.398 + ], + "angle": 0, + "content": "3. Antol, S., Agrawal, A., Lu, J., Mitchell, M., Batra, D., Zitnick, C.L., Parikh, D.: Vqa: Visual question answering. In: Proceedings of the IEEE international conference on computer vision. pp. 2425-2433 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.399, + 0.787, + 0.44 + ], + "angle": 0, + "content": "4. Besta, M., Blach, N., Kubicek, A., Gerstenberger, R., Podstawski, M., Gianinazzi, L., Gajda, J., Lehmann, T., Niewiadomski, H., Nczyk, P., Hoefler, T.: Graph of thoughts: Solving elaborate problems with large language models (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.441, + 0.787, + 0.482 + ], + "angle": 0, + "content": "5. Bogin, B., Gupta, S., Gardner, M., Berant, J.: Covr: A test-bed for visually grounded compositional generalization with real images. ArXiv abs/2109.10613 (2021), https://api.sementicscholar.org/CorpusID:237592834" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.482, + 0.787, + 0.605 + ], + "angle": 0, + "content": "6. Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J.D., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., Agarwal, S., Herbert-Voss, A., Krueger, G., Henighan, T., Child, R., Ramesh, A., Ziegler, D., Wu, J., Winter, C., Hesse, C., Chen, M., Sigler, E., Litwin, M., Gray, S., Chess, B., Clark, J., Berner, C., McCandlish, S., Radford, A., Sutskever, I., Amodei, D.: Language models are few-shot learners. In: Larochelle, H., Ranzato, M., Hadsell, R., Balcan, M., Lin, H. (eds.) Advances in Neural Information Processing Systems. vol. 33, pp. 1877-1901. Curran Associates, Inc. (2020), https://proceedings.neurips.cc/paper/2020/file/1457c0d6bfb4967418bf8ac142f64a-Paper.pdf" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.606, + 0.787, + 0.647 + ], + "angle": 0, + "content": "7. Buch, S., Eyzaguirre, C., Gaidon, A., Wu, J., Fei-Fei, L., Niebles, J.C.: Revisiting the \"Video\" in Video-Language Understanding. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.647, + 0.787, + 0.797 + ], + "angle": 0, + "content": "8. Chen, M., Tworek, J., Jun, H., Yuan, Q., Ponde, H., Kaplan, J., Edwards, H., Burda, Y., Joseph, N., Brockman, G., Ray, A., Puri, R., Krueger, G., Petrov, M., Khlaaf, H., Sastry, G., Mishkin, P., Chan, B., Gray, S., Ryder, N., Pavlov, M., Power, A., Kaiser, L., Bavarian, M., Winter, C., Tillet, P., Such, F.P., Cummings, D.W., Plappert, M., Chantzis, F., Barnes, E., Herbert-Voss, A., Guss, W.H., Nichol, A., Babuschkin, I., Balaji, S.A., Jain, S., Carr, A., Leike, J., Achiam, J., Misra, V., Morikawa, E., Radford, A., Knight, M.M., Brundage, M., Murati, M., Mayer, K., Welinder, P., McGrew, B., Amodei, D., McCandlish, S., Sutskever, I., Zaremba, W.: Evaluating large language models trained on code. ArXiv abs/2107.03374 (2021), https://api-semanticscholar.org/CorpusID: 235755472" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.799, + 0.787, + 0.84 + ], + "angle": 0, + "content": "9. Chen, W., Ma, X., Wang, X., Cohen, W.W.: Program of thoughts prompting: Disentangling computation from reasoning for numerical reasoning tasks. Transactions on Machine Learning Research (2023)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.274, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.331, + 0.127 + ], + "angle": 0, + "content": "Ge et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.785, + 0.175 + ], + "angle": 0, + "content": "10. Chen, X., Lin, M., Scharli, N., Zhou, D.: Teaching large language models to self-debug. arXiv preprint arXiv:2304.05128 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.176, + 0.787, + 0.23 + ], + "angle": 0, + "content": "11. Cheng, Z., Xie, T., Shi, P., Li, C., Nadkarni, R., Hu, Y., Xiong, C., Radev, D.R., Ostendorf, M., Zettlemoyer, L., Smith, N.A., Yu, T.: Binding language models in symbolic languages. ArXiv abs/2210.02875 (2022), https://api-semanticscholar.org/CorpusID:252734772" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.231, + 0.785, + 0.259 + ], + "angle": 0, + "content": "12. Cho, J., Zala, A., Bansal, M.: Visual programming for text-to-image generation and evaluation. NeurIPS (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.26, + 0.785, + 0.286 + ], + "angle": 0, + "content": "13. Gao, L., Madaan, A., Zhou, S., Alon, U., Liu, P., Yang, Y., Callan, J., Neubig, G.: Pal: Program-aided language models. arXiv preprint arXiv:2211.10435 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.287, + 0.785, + 0.342 + ], + "angle": 0, + "content": "14. Gupta, T., Kembhavi, A.: Visual programming: Compositional visual reasoning without training. 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) pp. 14953-14962 (2022), https://api-semanticscholar.org/CorpusID:253734854" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.343, + 0.785, + 0.398 + ], + "angle": 0, + "content": "15. Hu, R., Andreas, J., Rohrbach, M., Darrell, T., Saenko, K.: Learning to reason: End-to-end module networks for visual question answering. 2017 IEEE International Conference on Computer Vision (ICCV) pp. 804-813 (2017), https://api_semanticscholar.org/CorpusID:18682" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.399, + 0.785, + 0.453 + ], + "angle": 0, + "content": "16. Hudson, D.A., Manning, C.D.: Gqa: A new dataset for real-world visual reasoning and compositional question answering. 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) pp. 6693-6702 (2019), https://apisemantic scholar.org/CorpusID:152282269" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.453, + 0.785, + 0.508 + ], + "angle": 0, + "content": "17. Hudson, D.A., Manning, C.D.: Gqa: A new dataset for real-world visual reasoning and compositional question answering. 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) pp. 6693-6702 (2019), https://apisemantic scholar.org/CorpusID:152282269" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.509, + 0.785, + 0.55 + ], + "angle": 0, + "content": "18. Hudson, D.A., Manning, C.D.: Learning by abstraction: The neural state machine. In: Neural Information Processing Systems (2019), https://api-semanticscholar.org/CorpusID:195847902" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.55, + 0.785, + 0.592 + ], + "angle": 0, + "content": "19. Kim, W., Son, B., Kim, I.: Vilt: Vision-and-language transformer without convolution or region supervision. In: International Conference on Machine Learning (2021), https://api-semanticscholar.org/CorpusID:231839613" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.592, + 0.785, + 0.633 + ], + "angle": 0, + "content": "20. Li, L.H., Yatskar, M., Yin, D., Hsieh, C.J., Chang, K.W.: Visualbert: A simple and performant baseline for vision and language. ArXiv abs/1908.03557 (2019), https://api_semanticscholar.org/CorpusID:199528533" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.633, + 0.785, + 0.675 + ], + "angle": 0, + "content": "21. Liang, J., Huang, W., Xia, F., Xu, P., Hausman, K., Ichter, B., Florence, P., Zeng, A.: Code as policies: Language model programs for embodied control. In: arXiv preprint arXiv:2209.07753 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.675, + 0.785, + 0.715 + ], + "angle": 0, + "content": "22. Liu, F., Emerson, G.E.T., Collier, N.: Visual spatial reasoning. Transactions of the Association for Computational Linguistics 11, 635-651 (2022), https://apisemantic scholar.org/CorpusID:248496506" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.716, + 0.785, + 0.758 + ], + "angle": 0, + "content": "23. Lu, P., Peng, B., Cheng, H., Galley, M., Chang, K.W., Wu, Y.N., Zhu, S.C., Gao, J.: Chameleon: Plug-and-play compositional reasoning with large language models. arXiv preprint arXiv:2304.09842 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.759, + 0.785, + 0.799 + ], + "angle": 0, + "content": "24. Ma, Y.J., Liang, W., Wang, G., Huang, D.A., Bastani, O., Jayaraman, D., Zhu, Y., Fan, L., Anandkumar, A.: Eureka: Human-level reward design via coding large language models. arXiv preprint arXiv: Arxiv-2310.12931 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.799, + 0.785, + 0.841 + ], + "angle": 0, + "content": "25. Madaan, A., Zhou, S., Alon, U., Yang, Y., Neubig, G.: Language models of code are few-shot commonsense learners. In: Goldberg, Y., Kozareva, Z., Zhang, Y. (eds.) Proceedings of the 2022 Conference on Empirical Methods in Natural Language" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.524, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Recursive Visual Programming" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.245, + 0.147, + 0.785, + 0.189 + ], + "angle": 0, + "content": "Processing. pp. 1384-1403. Association for Computational Linguistics, Abu Dhabi, United Arab Emirates (Dec 2022). https://doi.org/10.18653/v1/2022.emnlp-main.90, https://aclanthology.org/2022.emnlp-main.90" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.191, + 0.787, + 0.247 + ], + "angle": 0, + "content": "26. Marino, K., Rastegari, M., Farhadi, A., Mottaghi, R.: Ok-vqa: A visual question answering benchmark requiring external knowledge. 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) pp. 3190-3199 (2019), https://api-semanticscholar.org/CorpusID:173991173" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.248, + 0.785, + 0.275 + ], + "angle": 0, + "content": "27. OpenAI: Gpt-4 technical report. ArXiv abs/2303.08774 (2023), https://apisemantic scholar.org/CorpusID:257532815" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.277, + 0.785, + 0.318 + ], + "angle": 0, + "content": "28. Press, O., Zhang, M., Min, S., Schmidt, L., Smith, N.A., Lewis, M.: Measuring and narrowing the compositionality gap in language models. ArXiv abs/2210.03350 (2022), https://api-semanticscholar.org/CorpusID:252762102" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.319, + 0.785, + 0.374 + ], + "angle": 0, + "content": "29. Subramanian, S., Bogin, B., Gupta, N., Wolfson, T., Singh, S., Berant, J., Gardner, M.: Obtaining faithful interpretations from compositional neural networks. In: Annual Meeting of the Association for Computational Linguistics (2020), https://apisemantic scholar.org/CorpusID:218487535" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.375, + 0.787, + 0.472 + ], + "angle": 0, + "content": "30. Subramanian, S., Narasimhan, M., Khangaonkar, K., Yang, K., Nagrani, A., Schmid, C., Zeng, A., Darrell, T., Klein, D.: Modular visual question answering via code generation. In: Rogers, A., Boyd-Graber, J., Okazaki, N. (eds.) Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers). pp. 747-761. Association for Computational Linguistics, Toronto, Canada (Jul 2023). https://doi.org/10.18653/v1/2023.acl-short.65, https://aclanthology.org/2023.acl-short.65" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.473, + 0.785, + 0.514 + ], + "angle": 0, + "content": "31. Sur'is, D., Menon, S., Vondrick, C.: Vipergpt: Visual inference via python execution for reasoning. ArXiv abs/2303.08128 (2023), https://api_semanticscholar.org/CorpusID:257505358" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.516, + 0.785, + 0.557 + ], + "angle": 0, + "content": "32. Suzgun, M., Scales, N., Scharli, N., Gehrmann, S., Tay, Y., Chung, H.W., Chowdhery, A., Le, Q.V., Chi, E.H., Zhou, D., et al.: Challenging big-bench tasks and whether chain-of-thought can solve them. arXiv preprint arXiv:2210.09261 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.558, + 0.785, + 0.626 + ], + "angle": 0, + "content": "33. Touvron, H., Lavril, T., Izacard, G., Martinet, X., Lachaux, M.A., Lacroix, T., Rozière, B., Goyal, N., Hambro, E., Azhar, F., Rodriguez, A., Joulin, A., Grave, E., Lample, G.: Llama: Open and efficient foundation language models. ArXiv abs/2302.13971 (2023), https://apisemanticscholar.org/CorpusID: 257219404" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.628, + 0.785, + 0.656 + ], + "angle": 0, + "content": "34. Wang, X., Li, S., Ji, H.: Code4struct: Code generation for few-shot structured prediction from natural language. arXiv preprint arXiv:2210.12810 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.657, + 0.785, + 0.712 + ], + "angle": 0, + "content": "35. Wang, Z., Li, M., Xu, R., Zhou, L., Lei, J., Lin, X., Wang, S., Yang, Z., Zhu, C., Hoiem, D., et al.: Language models with image descriptors are strong few-shot video-language learners. Advances in Neural Information Processing Systems 35, 8483-8497 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.714, + 0.785, + 0.755 + ], + "angle": 0, + "content": "36. Xiao, J., Shang, X., Yao, A., Chua, T.S.: Next-qa: Next phase of question-answering to explaining temporal actions. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 9777–9786 (June 2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.757, + 0.785, + 0.798 + ], + "angle": 0, + "content": "37. Yang, K., Klein, D., Peng, N., Tian, Y.: Doc: Improving long story coherence with detailed outline control. In: Annual Meeting of the Association for Computational Linguistics (2023), https://api-semanticscholar.org/CorpusID:254877751" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.799, + 0.785, + 0.84 + ], + "angle": 0, + "content": "38. Yao, S., Yu, D., Zhao, J., Shafran, I., Griffiths, T.L., Cao, Y., Narasimhan, K.: Tree of thoughts: Deliberate problem solving with large language models. arXiv preprint arXiv:2305.10601 (2023)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.333, + 0.128 + ], + "angle": 0, + "content": "Ge et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.189 + ], + "angle": 0, + "content": "39. Ye, Q., Xu, G., Yan, M., Xu, H., Qian, Q., Zhang, J., Huang, F.: Hitea: Hierarchical temporal-aware video-language pre-training. ArXiv abs/2212.14546 (2022), https://api-semanticscholar.org/CorpusID:255340506" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.19, + 0.787, + 0.259 + ], + "angle": 0, + "content": "40. Yu, W., Gileadi, N., Fu, C., Kirmani, S., Lee, K.H., Gonzalez Arenas, M., Lewis Chiang, H.T., Erez, T., Hasenclever, L., Humplik, J., Ichter, B., Xiao, T., Xu, P., Zeng, A., Zhang, T., Heess, N., Sadigh, D., Tan, J., Tassa, Y., Xia, F.: Language to rewards for robotic skill synthesis. Arxiv preprint arXiv:2306.08647 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.259, + 0.787, + 0.314 + ], + "angle": 0, + "content": "41. Zhang, P., Li, X., Hu, X., Yang, J., Zhang, L., Wang, L., Choi, Y., Gao, J.: Vinvl: Revisiting visual representations in vision-language models. 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) pp. 5575-5584 (2021), https://api_semanticscholar.org/CorpusID:235692795" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.315, + 0.787, + 0.37 + ], + "angle": 0, + "content": "42. Zhou, D., Scharli, N., Hou, L., Wei, J., Scales, N., Wang, X., Schuurmans, D., Bousquet, O., Le, Q., hsin Chi, E.H.: Least-to-most prompting enables complex reasoning in large language models. ArXiv abs/2205.10625 (2022), https://api_semanticscholar.org/CorpusID:248986239" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.37 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/Recursive Visual Programming/e57bdabb-77ec-4030-98f0-ae871f741b2d_origin.pdf b/2024/Recursive Visual Programming/e57bdabb-77ec-4030-98f0-ae871f741b2d_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..f9d546ff055e11995ff5b70ca56d7d0c693c75db --- /dev/null +++ b/2024/Recursive Visual Programming/e57bdabb-77ec-4030-98f0-ae871f741b2d_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ffbac4766d3ed52b15a8d230595965d2cba554f04ea16c85d408942557b5e95e +size 2601804 diff --git a/2024/Recursive Visual Programming/full.md b/2024/Recursive Visual Programming/full.md new file mode 100644 index 0000000000000000000000000000000000000000..b472ede1bcbad2d14bbc30f21e8f264c5bf2d582 --- /dev/null +++ b/2024/Recursive Visual Programming/full.md @@ -0,0 +1,276 @@ +# Recursive Visual Programming + +Jiaxin Ge1 Sanjay Subramanian1 Baifeng Shi1 Roei Herzig1 Trevor Darrell1 + +UC Berkeley, CA, USA + +Abstract. Visual Programming (VP) has emerged as a powerful framework for Visual Question Answering (VQA). By generating and executing bespoke code for each question, these methods show advancement in leveraging Large Language Models (LLMs) for complex problem-solving. Despite their potential, existing VP methods generate all code in a single function, which does not fully utilize LLM's reasoning capacity and the modular adaptability of code. This results in code that is suboptimal in terms of both accuracy and interpretability. Inspired by human coding practices, we propose Recursive Visual Programming (RVP), which better harnesses the reasoning capacity of LLMs, provides modular code structure between code pieces, and assigns different return types for the sub-problems elegantly. RVP approaches VQA tasks with an top-down recursive code generation approach, allowing decomposition of complicated problems into smaller parts. We show RVP's efficacy through extensive experiments on benchmarks including VSR, COVR, GQA, and NextQA, underscoring the value of adopting human-like recursive and modular programming techniques for solving VQA tasks. Our code is available at https://github.com/para-lost/RVP. + +Keywords: Visual Programming $\cdot$ Vision-Language $\cdot$ VQA + +# 1 Introduction + +Visual Question Answering (VQA) lies at the intersection of computer vision and natural language processing, posing the challenge of interpreting visual data to answer questions [3,16,26]. Following the recent progress on code generation using Large Language Models (LLMs) [6,27,33], Visual Programming (VP) methods have notably advanced in this area, successfully using LLMs to generate and execute code in few-shot and zero-shot scenarios [14,30,31]. In these techniques, various vision models are provided to LLM as APIs, and the LLM is used as a planner to utilize these vision models to perform reasoning over images. + +Despite these advancements, such planning remains very challenging for Visual Programming. These methods require the model to generate a single code block using the predefined APIs at once. As a result, the model must handle logic, be aware of all details, and utilize the vision APIs appropriately. This has two main drawbacks. First, such approaches do not fully utilize LLM's reasoning capability, which would benefit from breaking down the problem into step-by-step sub-problems, as demonstrated in recent NLP research [4]. Secondly, they + +![](images/cc5ac8908a42096357813ed66e120177bf0638873d475f7f0f643930020c20da.jpg) +Fig. 1: A breakdown of Recursive Visual Programming for a visual question is illustrated with an image where scenes within it contain smaller versions of themselves. To locate a man wearing a hat at the third level of an image, the model first generates code that has two recursive_query API calls. The initial call identifies the largest mirror (representing the next image level), and the subsequent call seeks "the man with a hat" at the second level. The model iteratively generates and executes code through these calls until the final answer is produced without additional recursive calls. + +fail to utilize the full potential of code because they lack modularity between the code pieces, particularly when the code is one long piece. + +We consider how humans tackle these issues. Humans tend to code in a top-down manner: first they break down complex tasks into manageable subproblems, and then they tackle each through focused code segments. This approach allows the programmer to concentrate on the main logic flow rather than figuring out all the details at once, resulting in a more clear and elegant structure. + +Inspired by this, we hypothesize that incorporating a similar recursive coding strategy would also help solve complex visual question-answering problems. To investigate whether LLMs could indeed benefit from recursive coding, we conduct preliminary experiment of the recursive coding approach on standard symbolic reasoning benchmarks, such as Dyck Language [32] and Games of 24 [38]. This is done by simply prompting the LLM to code recursively. We find that utilizing LLMs to code recursively leads to a notable improvement in accuracy, as shown in Figure 2. Motivated by this result, we explore in this paper how a recursive coding methodology can be applied to VQA tasks. + +Specifically, we propose a novel VP method for VQA tasks, named Recursive Visual Programming (RVP). Instead of querying LLM only once to generate a single piece of code, we ask LLM to decompose the question into sub-questions + +![](images/a11a8f66fb1cbde1cb07d94fb4fa00b38ebd21d81121ba7bbb9dd4d01c6c283c.jpg) +Fig. 2: Motivating Examples. Adding the prompt "solve the problem recursively" to the input significantly improves the model's performance on traditional coding tasks, Dyck language [32] and Games of 24 [38]. "NR" refers to non-recursive approach and "R" refers to recursive approach. This shows that LLMs can potentially benefit from human's recursive programming approach. + +and recursively query itself for each sub-question to generate modular code pieces. For each sub-question, the same process repeats, until reaching an atom level where the current question can be solved by only calling external APIs. See example in Figure 1. In contrast to querying LLM only once, we find that using recursive queries to generate code for each sub-question can improve the readability and accuracy of the code, as shown in Figure 3. Moreover, MVP allows assigning dynamic types to each sub-question, enabling new code pieces to return a variety of types based on the current need. Remarkably, MVP assigns return types unseen in in-context examples, such as List[str], List[ImagePatch]. + +Our contribution can be summarized as follows: (i) We introduce a novel recursive approach based on human programming. This approach is a top-down programming method that better harnesses LLM's reasoning capacity and enhances modularity between multiple code pieces. It is easily applicable to any existing VP model for visual question answering task. (ii) We allow dynamic return types for the sub-problems, which enhances flexibility and adaptability. (iii) We demonstrate improved performance on several standard benchmarks, such as GQA, VSR, NextQA and COVR, demonstrating the effectiveness of our approach. We also show better interpretability than non-recursive VP methods. + +# 2 Related Work + +Modular Visual Reasoning. There is a long line of work that seeks to combine modularity with the expressivity of deep neural networks. Early work on this topic converts the output of a parser [1] to a program or trains a program generator either separately from or jointly with the execution modules [2, 15]. Later work finds that such models are often challenging to train on real-world datasets, particularly when interpretability of individual module outputs is expected [29]. Other work improves performance on multi-hop visual reasoning without explicitly defining a set of modules, e.g., by predicting and operating + +![](images/a8b7ab2795a081c333bac9386e43bd962bfd8fe27c57853ee5b1b9e61417c441.jpg) +Fig. 3: Example from COVR (Upper). Current VP methods fail to provide binary 'yes' or 'no' answers, while Recursive Visual Programming method outputs the correct answer. Example from GQA (Middle and Bottom). Recursive outperforms current non-recursive methods by correctly addressing all details and their associated logic. + +over a scene graph [18]. + +Visual Programming. Recent language models can accurately generate code from text descriptions when prompted with a small number of examples [8]. This advance has enabled a few-shot modular visual reasoning approach, commonly called visual programming [14], which consists of prompting a language model for a program and executing the program with pre-trained vision models. This paradigm has led to impressive results in visual question answering, video question answering, visual and tabular question answering, referring expression comprehension, and text-based image editing [11, 12, 23, 30, 31]. Programming with LLMs has also been applied to produce plans and reward functions for robots [21, 24, 40]. Our work builds on these techniques by enabling the LLM to generate code in multiple steps-first planning the overall process and then filling in the details-rather than all at once. + +Language Models for Reasoning Tasks. In the text domain, there is a large body of work on using language models for reasoning tasks. In particular, several papers have shown the efficacy of code generation as a scaffold for such tasks [9, 13, 25, 34]. Aside from code generation, prior work has also shown that decomposing complex queries into simpler ones improves the performance of language models [28, 42]. Our RVP approach applies this insight to improve the code generation component of visual programming. + +![](images/289e4bc1dc3f4c4c4b0e6b14f23b1457f7727744791a3995762557724aba68b7.jpg) +Fig. 4: An overview of our approach. When a recursive_query API call is occurred, the model generates and executes new code based on the query, returning the output back to the original call. + +# 3 Recursive Visual Programming + +In this section, we introduce our method, Recursive Visual Programming (RVP). We first review the problem setup and ingredients of visual programming as established by prior work (see Section 3.1). We then introduce our recursive coding framework (see Section 3.2), which involves a multi-step prompting scheme, as well as a system for dynamic type assignment (see Section 3.3). + +# 3.1 Preliminaries + +In Visual Question Answering (VQA), the system $f$ is given a visual input $I$ (i.e. an image, a set of images, or a video), as well as a question $Q$ relating to the visual content. The expected output is a textual answer $A$ to the question: + +$$ +A \leftarrow f (I, Q) \tag {1} +$$ + +Visual Programming methods [14,30,31] represent $f$ as the composition of two functions: a code generator $g$ and a visual executor $h$ . In the first part, the code generator generates a program $P$ conditioned on the question, and in the second part, the executor executes the program on the visual input: + +$$ +P \leftarrow g (Q), A \leftarrow h (P, I) \tag {2} +$$ + +We focus on the few-shot setting in which we have only a few (at most 50) training examples—each consisting of a question, program, and answer—for a given + +task. Therefore, the code generator $g$ is composed of a language model and a prompt that includes a description of the API available to the executor and in-context examples of questions along with the corresponding correct programs. We use Python as the programming language, similar to [30, 31]. + +We evaluate our approach using the ViperGPT [31] API. For questions over images or sets of images, these functions can be called on an ImagePatch object, which can be any crop of the image: + +1. find(object: str): Returns a list of ImagePatch objects matching the description. +2. exists(object: str): Returns a bool representing whether the described object exists. +3. verify_property(object: str, property: str): Returns a bool representing whether the described object has the specified property. +4. simple_query(question: str): Returns an answer to the provided question based on the given ImagePatch. +5. compute_depth(): Returns the median depth of the region in the given ImagePatch. +6. crop(left: int, lower: int, right: int, upper: int): Crops a part of the given ImagePatch and returns a new ImagePatch representing that crop. + +For questions about a video, the API includes three other functions: trim(), frame_from_index(), and frame_iterator(), which return an abbreviated version of the video, a specific frame, and an iterator over the frames respectively. + +# 3.2 Recursive Visual Programming + +We introduce a new method for code generation in visual programming: Recursive Visual Programming (RVP). RVP enables the language model to break the code generation process into multiple sub-questions and recursively call itself to generate a program for each sub-question. As shown in Figure 4, we operationalize RVP by adding a new function to the API: recursive_query(image, sub-question). During execution, when recursive_query is called, it prompts the language model to generate a new code given the sub-question. This code is then executed, and its return value is returned by recursive_query. An example break down of the usage of recursive_query call is shown in Figure 1. We set the recursive termination condition by (1) setting max depth of 10 and (2) switching to direct simple query when generated code contains a recursive call to the same input question. + +To enable this process, we add examples of calling recursive_query in the prompt to show the model how to decompose the problem and utilize the answers returned by recursive_query. We provide examples of the in-context prompt of recursive_query we used in the supplementary material. + +![](images/e847bf6c00ebf60a2f11847f5d049b303e64d5069766beb8bd8f16965ff7d17b.jpg) +Fig.5: Examples of dynamic type assignment in recursive_query calls. The model autonomously determines the appropriate return type in its code generation and generalize to new types unseen in in-context examples such as List[ImagePatch]. + +# 3.3 Dynamic Type Assignment + +In previous visual programming frameworks, code generation mechanisms are constrained by static type assignments, i.e., the return type of a given function is fixed. For instance, simple_query always returns a string. However, this rigid structure limits the model's flexibility to adapt the code to different contexts within the VQA task. RVP introduces dynamic type assignment to address this limitation. When the code generator writes a recursive_query function, it can include the expected return type as a part of the question that is passed to the recursive_query. To facilitate this, in the in-context examples that include recursive_query, we let the input 'question' specify a type. As shown in Figure 5, we give examples of type ImagePatch, bool, float, and str being specified within the queries. When receiving this question, the model discerns the required type and generates code that declares and returns the correct type. For instance, a query like "Return a bool, is there a black cat in the picture?" will prompt the generation of code with the signature. + +$$ +\operatorname {e x e c u t e} (\operatorname {i m a g e}) \rightarrow \operatorname {b o o l} \tag {3} +$$ + +This signature indicates that the generated code $P'$ , when executed, will return a boolean value $A_{\mathrm{bool}}$ which answers the sub-question $Q'$ , illustrating the dynamic and context-sensitive nature of RVP's code generation capabilities that help to produce syntactically and semantically appropriate code for complex reasoning. + +# 4 Experiments + +In this section, we introduce the basic settings including the task formulation and parameter details. Then, we provide experiment results and in-depth analysis. + +# 4.1 Experiment Setup + +Datasets. We evaluate on four datasets spanning a wide range of visual reasoning skills and types of visual input: Visual Spatial Reasoning (VSR) [22], GQA [17], COVR [5], and NextQA [36]. For NextQA, we present results on the "Hard Temporal" subset which was curated to consist of questions that truly require temporal reasoning [7]. VSR and GQA include queries about single images, while COVR includes queries about multiple images, and NextQA includes queries about video. On each dataset, we evaluate methods by their accuracy in producing an answer that exactly matches the ground-truth answer. For NextQA, we adopt a multiple-choice setup in which we provide methods with the set of possible answer choices for each question. For VSR, a spatial statement of an image is made, and the question is framed as : "Is it true or false," followed by the statement. In this way, a VSR statement is framed as a VQA question and the model will answer either true or false. + +Implementation Details. We implement RVP using the ViperGPT API and the accompanying prompt template. For each dataset, we curate a set of in-context examples which are added to the prompt. The set of question-answer pairs is shared between the recursive and non-recursive methods. We use seven in-context examples for GQA, six for NextQA, six for VSR, and eight for COVR. For code generation, use GPT-3.5-turbo via the OpenAI API1. + +# 4.2 Quantitative Results + +Table 1 presents our results across four datasets. In all settings, we outperform ViperGPT, which is identical to RVP except that RVP includes recursive_query function in the API and in-context examples. This indicates that recursive programming indeed improves accuracy in visual reasoning tasks. Moreover, our method outperforms previous zero/few-shot methods in all but one of the settings. While our primary aim was not to push the state-of-the-art results, we show that integrating the clarity and elegance of recursive coding into standard VP methods not only doesn't impair performance, but can enhance it. + +# 4.3 Qualitative Results + +We provide qualitative examples in Figure 3,6,7. In the example from COVR in Figure 3, the non-recursive approach answers with neither "yes" nor "no," while the recursive approach does confirm to these choices. One possible explanation for this issue is that LLMs struggle to maintain coherence across long contexts [37]. RVP circumvents this issue by pushing the LLM to commit to the answer + +Table 1: Results. We report exact-match accuracy on each dataset. Best zero/few-shot result for each dataset is bolded. †Fully supervised results are from different models: ViLT [19] for VSR, HiTeA [39] for NextQA, VinVL-Base [41] for GQA, and VisualBERT [20] for COVR. ‡We report the results of our reproduction of ViperGPT using the officially released code and the same in-context examples used for our method (since the original in-context examples from ViperGPT were not released). We use GPT3.5-turbo while CodeVQA uses Codex, which is not available anymore. + +
MethodVSRNextQAGQA Test-devCOVR Test
Random SplitZero-shot SplitHard Split-TTest-dev
Fully supervised†69.363.048.665.157.9
Zero/few-shot methods
CLIP56.054.5---
BLIPv2---42.31-
CodeVQA---49.050.7
ViperGPT‡61.2561.5947.2144.6351.69
RVP (ours)63.5366.0948.8245.6252.67
+ +types earlier in its generation (the length from the beginning of the code to the point "return xxx" is shorter). In Figure 6, the non-recursive approach fails to handle the logic correctly while the recursive approach correctly does. This is likely due to that RVP breaks down the code logic and each code piece only needs to focus on a single logic, while the non-recursive method attempts to handle all the logic at once. In the GQA example, a common pattern involves handling complex logical problems and multiple properties. Non-recursive methods often fell short in these cases, either overlooking critical details or processing them in a logically flawed manner. RVP manages these intricate questions by addressing each detail accurately. This difference may stem from the inherent difficulty LLMs face in processing complicated logic over extended contexts. Recursive methods simplify this by focusing on the overall logic and gradually breaking the problems down into smaller, more manageable segments. This facilitates easier and more accurate resolution. + +We provide a dynamic type assignment example in Figure 5. In this example, RVP assigns the return type List[ImagePatch] in the recursive_query call. It's worth noting that the return type List[ImagePatch] is unseen in the in-context examples. This suggests that RVP has the potential to utilize complicated and diverse data structures, highlight the effectiveness of our approach. + +# 4.4 Ablation Study + +Dynamic Type This section studies integrating dynamic type assignment within the recursive_query API. For the GQA dataset, we examined all instances with recursive patterns in the test_dev split. In COVR, where the test set is unavailable, we used the val_1000 samples from the validation set, following [30]. + +We compare four approaches: Non-recursive Baseline: Standard approach without recursive features. Fixed-Type Recursive: The recursive_query API is constrained to return a string. Dynamic-Implicit Recursive: The API + +![](images/96032daf77e87f35d0d7e22e44648709dd0d14974fd7eb030569176dd9312844.jpg) +Fig. 6: Examples from COVR. RVP allows correct logic flow while traditional nonrecursive VP fails to handle the logic correctly. + +can return various types, but the specific type is not predefined in the query. For example, "Is there a black cat?" could yield either a boolean or a string. Dynamic-Explicit Recursive: Our primary approach, where the API's input query specifies the return type, like "Return a bool, is there a black cat?" + +Table 2 shows our findings. The dynamic-explicit type assignment outperformed other methods, underscoring the need for clear type specification. The dynamic-implicit method showed lower performance than non-recursive due to type ambiguities, as exemplified in Figure 8, where the first code piece assumes the return type is a bool, while the second code piece returns a str. Fixed-type assignment proves to be beneficial over non-recursive methods, indicating that recursion itself adds value, but its rigidity in type handling limits performance. This study highlights the benefits of flexible, dynamic type assignment and the crucial role of explicit type declaration for accurate and consistent responses. These insights mark a step towards more adaptable visual reasoning frameworks, emphasizing the need for precision in dynamic environments. + +![](images/68ee3982eab4da0234e24377af0769c55cfb81cf4fbef81d73669132562e818b.jpg) + +![](images/7289d0e956b6012d08344c50375c505214c6fb3d6a09b6bebfa0f820f4e5b31e.jpg) +Fig. 7: Examples from GQA. RVP allows handling the details more elegantly and produces logically correct code. + +In-Context Examples We conduct ablation on using retrieval-based in-context example selection and using different number of in-context examples. + +Retrieval-based in-context example selection. We manage to write 9 programs using the recursive method from the 50 provided by CodeVQA [30], and add our original 3 recursive programs to them. We keep the non-recursive in-context example unchanged and choose 3 recursive in-context examples from these 12 for each question. We follow [35] to use the embedding-based retrieval technique that calculates a score for each example, and then select the top 3 examples for each question. The results of GQA_val_2000 are in Table 2. We find that the recursive pattern occurrence rate is significantly higher than using fixed in-context examples. This is likely because by retrieving similar recursively-decomposed questions, the model can more easily simulate their pat + +terns, thereby attempting to solve more problems recursively. However, the overall accuracy does not improve. This suggests that although the retrieval-based method encourages more diverse and suitable patterns of recursive decomposition, there is a need for improved techniques in selecting the example pool—such as focusing on those most suitable for recursive decomposition instead of randomly selecting—along with careful tuning of the example programs. + +Number of in-context examples. We randomly add one example/delete one example from the covr_1000_val, the results are shown in Table 2. The results indicate that adding the number of examples can enhance the performance, but does not have a significant impact. + +Error feedback loop Recent studies have studied the debugging ability of LLMs on code [10]. We study whether modular visual code structure enables better bug identification for visual programming. We sample 20 incorrect programs for 10 questions that have type/logic errors, and use GPT3.5 to (1) identify the bug, and (2) write the correct code given the bug. For identification, the model correctly finds 7 bugs correctly for RVP and 2 for non-recursive visual programming. For correction, the model fixes 7 bugs correctly for RVP and 7 for visual programming. This suggests that clearer and more modular code structure, as presented RVP, can potentially aids bug identification. + +Table 2: Ablation study. (a) We provide ablation study results for dynamic type assignment and found that specified dynamic type assignment in RVP achieves the highest performance. (b) We provide results on using retrieval-based in-context examples and fixed examples. Retrieval-based method enables more diverse recursive patterns but requires careful choosing the prompt pool. (c) We provide results of using different number of examples and found this doesn't influence performance significantly. + +(a) Dynamic Type Ablation. + +
MethodGQACOVR
Non-Recursive55.9960.71
Fixed-Type65.0560.12
Dynamic-Implicit55.3453.57
Dynamic-Explicit70.2367.86
+ +(b) Retrieval-Based and Fixed Examples + +
FixedRetrieval
Accuracy54.7553.79
Rec Rate3.55%14.85%
Rec Acc71.8357.91
+ +(c) In-Context Example Num + +
MethodAccuracy
Del 150.65
Original51.05
Add 151.35
+ +# 4.5 Readability Study + +This study investigates how recursive coding impacts code readability and comprehension for human programmers. Our goal is to move from complex code structures to more modular, clear, and understandable code. To evaluate this, we focus on code quality in terms of its understandability. + +From the COVR dataset, we randomly select one example from each five different question types where recursive patterns occur. We present the main function of both their recursive and non-recursive code versions to 15 participants and ask them to grasp the high-level intent of the code as quickly as + +Return Type Confusion: Expect Bool but Actually Str. +```python +def execute(image) -> str: image_batch = ImagePatch(image) is.black_skateboard= image_batchrecursive_query("Are there any black skateboards?") is_motorcycle $=$ image_batchrecursive_query("Arethereanymotorcycles?") return bool to yesno(is.black Skateboard or is_motorcycle) +def execute(image) -> str: image_batch $=$ ImagePatch(image) skateboard_patches $=$ image_batch.find("skateboard") for skateboard_batch in skateboard_patches: If skateboard_batchverified_property("skateboard", "black"): return "yes" return "no" +``` + +![](images/5d1427f2a49da9b18a5f78c59270b89ceb435daaf18ee769d6a2d32c42e9e657.jpg) +Fig.8: Failure example of dynamic type assignment due to unspecified type, The model confuses the type and return a str instead of a bool. +Fig.9: Comparison of average understanding time for recursive and non-recursive codes. RVP demonstrates better readability and is easy to understand. + +possible. For the non-recursive case, the main function is just the single code snippet generated by the LLM. For the recursive case, the main function consists of the code generated by the first LLM call (the high level code) but not the functions generated by the recursive queries. This helps us measure how fast users can understand the code's intention by looking at the main function, and reflects the self-explanatory nature of the code. Figure 9 shows the results. Participants consistently understand the recursive code faster than the non-recursive versions. This quicker comprehension can be attributed to the recursive code's structured and clear approach. Non-recursive code, while correct, sometimes lacks immediate clarity in conveying its primary goal. + +We also calculate the average line length for all main and sub functions. For RVP, the average length of all functions is 8.47. For non-recursive VP, the average length of the functions is 15.4. This suggests that RVP decreases the average length the code pieces and enhances modularity. + +Our findings suggest that recursive coding not only simplifies complex code but also enhances its accessibility and ease of understanding. Its orderly structure and clear logic accelerate comprehension, leading to more efficient and user-friendly programming. We provide more details of this study in the supp. + +# 4.6 Recursive Analysis + +Question Type. We study which types of questions tend to use recursive and found that recursive patterns appeared in five question types within both the COVR test set and the GQA test-dev set, as shown in Table 3. The 'Verify Property' and 'Query' types in GQA, and 'Compare' and 'Multi-Reference' in COVR, were most common. This suggests recursive methods tend to occur in tasks that require comparative and referential analysis. + +Dynamic Type. We study which types of dynamic return types occurs and + +Table 3: Recursive analysis. Table (a) provides the distribution of recursive patterns by question type in GQA and COVR. We have found that recursive patterns occur commonly in question types that require reasoning. Table (b) studies different recursive query return types and their frequency in GQA and COVR. The model generalizes to unseen types. Table (c) compares accuracy across question types. Recursive methods outperform non-recursive particularly in 'Choose' and 'Query' questions. + +(a) Dist. of Recursive Patterns + +(b) Diverse Return Types + +(c) Type Accuracy + +
COVR/GQAQuestion Types
COVRGQA
Compare45715
Mult. Ref282-
Quant. Attr12-
Spec Attr357-
Choose125
Verify-68
Query-167
Logical-34
+ +
Data TypeGQACOVR
Str3191828
Bool521592
List[str]16-
ImagePatch30-
List[ImgPatch]15
Float6-
Int-914
+ +
QuestionAccuracy (%)
Non-RecRec
Verify52.9464.71
Query56.8971.86
Logical55.8870.59
Compare66.6753.33
Choose52.0084.00
+ +their frequently in GQA and COVR. In Table 3, we found that RVP generalizes to new types like List[str] and List[ImagePatch], which are not included in the in-context examples. This flexibility and ability to generalize beyond in-context examples highlight the inherent adaptability of recursive methods. + +Type Accuracy. We evaluate the accuracy of recursive methods across different question types. An analysis of the GQA test-dev set is presented in Table 3. It reveals that recursive methods significantly improve accuracy in 'Query', 'Logical', and 'Choose' questions. This underscores their suitability for complex reasoning and decision-making tasks. + +Error Source. We sample 50 incorrect recursive examples from GQA2000val and manually check the error source: $22\%$ are caused by logic error inside subfunctions; $18\%$ are caused by logic error between functions; $52\%$ are caused by API model inaccuracy (i.e. detection model error); and $4\%$ are caused by other problems like the groundtruth answer unclear. + +Cost And Runtime. For the questions which require recursive query, RVP takes about 2.1 times (runtime) and about 3 times (cost) compared to ViperGPT. + +# 5 Conclusion and Limitation + +In this work, we present RVP, a visual programming method that adopts recursive coding to generate concise, elegant and easy-to-understand code, handling details and intricate logic more accurately compared with traditional visual programming methods. We conduct extensive experiments across various benchmarks and provide a comprehensive analysis of RVP. Our work is the first to explore recursive programming in the visual domain and dynamic-type code generation for visual reasoning; we believe the recursive programming concept would have a potential greater impact on other domains as well. While RVP is effective for complicated visual questions, most questions in current VQA benchmarks tend to be short and non-recursive. We believe that the benefit of our method could be demonstrated better with a new benchmark or a more complex split. + +# Acknowledgements + +We would like to thank Ben Bogin for helping us test on the private COVR test set, and Guangyuan Jiang for helping us proof read the paper and providing valuable feedback. + +# References + +1. Andreas, J., Rohrbach, M., Darrell, T., Klein, D.: Neural module networks. 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR) pp. 39-48 (2015), https://api-semanticscholar.org/CorpusID:5276660 +2. Andreas, J., Rohrbach, M., Darrell, T., Klein, D.: Learning to compose neural networks for question answering. ArXiv abs/1601.01705 (2016), https://apisemantic scholar.org/CorpusID:3130692 +3. Antol, S., Agrawal, A., Lu, J., Mitchell, M., Batra, D., Zitnick, C.L., Parikh, D.: Vqa: Visual question answering. In: Proceedings of the IEEE international conference on computer vision. pp. 2425-2433 (2015) +4. Besta, M., Blach, N., Kubicek, A., Gerstenberger, R., Podstawski, M., Gianinazzi, L., Gajda, J., Lehmann, T., Niewiadomski, H., Nczyk, P., Hoefler, T.: Graph of thoughts: Solving elaborate problems with large language models (2024) +5. Bogin, B., Gupta, S., Gardner, M., Berant, J.: Covr: A test-bed for visually grounded compositional generalization with real images. ArXiv abs/2109.10613 (2021), https://api.sementicscholar.org/CorpusID:237592834 +6. Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J.D., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., Agarwal, S., Herbert-Voss, A., Krueger, G., Henighan, T., Child, R., Ramesh, A., Ziegler, D., Wu, J., Winter, C., Hesse, C., Chen, M., Sigler, E., Litwin, M., Gray, S., Chess, B., Clark, J., Berner, C., McCandlish, S., Radford, A., Sutskever, I., Amodei, D.: Language models are few-shot learners. In: Larochelle, H., Ranzato, M., Hadsell, R., Balcan, M., Lin, H. (eds.) Advances in Neural Information Processing Systems. vol. 33, pp. 1877-1901. Curran Associates, Inc. (2020), https://proceedings.neurips.cc/paper/2020/file/1457c0d6bfb4967418bf8ac142f64a-Paper.pdf +7. Buch, S., Eyzaguirre, C., Gaidon, A., Wu, J., Fei-Fei, L., Niebles, J.C.: Revisiting the "Video" in Video-Language Understanding. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2022) +8. Chen, M., Tworek, J., Jun, H., Yuan, Q., Ponde, H., Kaplan, J., Edwards, H., Burda, Y., Joseph, N., Brockman, G., Ray, A., Puri, R., Krueger, G., Petrov, M., Khlaaf, H., Sastry, G., Mishkin, P., Chan, B., Gray, S., Ryder, N., Pavlov, M., Power, A., Kaiser, L., Bavarian, M., Winter, C., Tillet, P., Such, F.P., Cummings, D.W., Plappert, M., Chantzis, F., Barnes, E., Herbert-Voss, A., Guss, W.H., Nichol, A., Babuschkin, I., Balaji, S.A., Jain, S., Carr, A., Leike, J., Achiam, J., Misra, V., Morikawa, E., Radford, A., Knight, M.M., Brundage, M., Murati, M., Mayer, K., Welinder, P., McGrew, B., Amodei, D., McCandlish, S., Sutskever, I., Zaremba, W.: Evaluating large language models trained on code. ArXiv abs/2107.03374 (2021), https://api-semanticscholar.org/CorpusID: 235755472 +9. Chen, W., Ma, X., Wang, X., Cohen, W.W.: Program of thoughts prompting: Disentangling computation from reasoning for numerical reasoning tasks. Transactions on Machine Learning Research (2023) + +10. Chen, X., Lin, M., Scharli, N., Zhou, D.: Teaching large language models to self-debug. arXiv preprint arXiv:2304.05128 (2023) +11. Cheng, Z., Xie, T., Shi, P., Li, C., Nadkarni, R., Hu, Y., Xiong, C., Radev, D.R., Ostendorf, M., Zettlemoyer, L., Smith, N.A., Yu, T.: Binding language models in symbolic languages. ArXiv abs/2210.02875 (2022), https://api-semanticscholar.org/CorpusID:252734772 +12. Cho, J., Zala, A., Bansal, M.: Visual programming for text-to-image generation and evaluation. NeurIPS (2023) +13. Gao, L., Madaan, A., Zhou, S., Alon, U., Liu, P., Yang, Y., Callan, J., Neubig, G.: Pal: Program-aided language models. arXiv preprint arXiv:2211.10435 (2022) +14. Gupta, T., Kembhavi, A.: Visual programming: Compositional visual reasoning without training. 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) pp. 14953-14962 (2022), https://api-semanticscholar.org/CorpusID:253734854 +15. Hu, R., Andreas, J., Rohrbach, M., Darrell, T., Saenko, K.: Learning to reason: End-to-end module networks for visual question answering. 2017 IEEE International Conference on Computer Vision (ICCV) pp. 804-813 (2017), https://api_semanticscholar.org/CorpusID:18682 +16. Hudson, D.A., Manning, C.D.: Gqa: A new dataset for real-world visual reasoning and compositional question answering. 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) pp. 6693-6702 (2019), https://apisemantic scholar.org/CorpusID:152282269 +17. Hudson, D.A., Manning, C.D.: Gqa: A new dataset for real-world visual reasoning and compositional question answering. 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) pp. 6693-6702 (2019), https://apisemantic scholar.org/CorpusID:152282269 +18. Hudson, D.A., Manning, C.D.: Learning by abstraction: The neural state machine. In: Neural Information Processing Systems (2019), https://api-semanticscholar.org/CorpusID:195847902 +19. Kim, W., Son, B., Kim, I.: Vilt: Vision-and-language transformer without convolution or region supervision. In: International Conference on Machine Learning (2021), https://api-semanticscholar.org/CorpusID:231839613 +20. Li, L.H., Yatskar, M., Yin, D., Hsieh, C.J., Chang, K.W.: Visualbert: A simple and performant baseline for vision and language. ArXiv abs/1908.03557 (2019), https://api_semanticscholar.org/CorpusID:199528533 +21. Liang, J., Huang, W., Xia, F., Xu, P., Hausman, K., Ichter, B., Florence, P., Zeng, A.: Code as policies: Language model programs for embodied control. In: arXiv preprint arXiv:2209.07753 (2022) +22. Liu, F., Emerson, G.E.T., Collier, N.: Visual spatial reasoning. Transactions of the Association for Computational Linguistics 11, 635-651 (2022), https://apisemantic scholar.org/CorpusID:248496506 +23. Lu, P., Peng, B., Cheng, H., Galley, M., Chang, K.W., Wu, Y.N., Zhu, S.C., Gao, J.: Chameleon: Plug-and-play compositional reasoning with large language models. arXiv preprint arXiv:2304.09842 (2023) +24. Ma, Y.J., Liang, W., Wang, G., Huang, D.A., Bastani, O., Jayaraman, D., Zhu, Y., Fan, L., Anandkumar, A.: Eureka: Human-level reward design via coding large language models. arXiv preprint arXiv: Arxiv-2310.12931 (2023) +25. Madaan, A., Zhou, S., Alon, U., Yang, Y., Neubig, G.: Language models of code are few-shot commonsense learners. In: Goldberg, Y., Kozareva, Z., Zhang, Y. (eds.) Proceedings of the 2022 Conference on Empirical Methods in Natural Language + +Processing. pp. 1384-1403. Association for Computational Linguistics, Abu Dhabi, United Arab Emirates (Dec 2022). https://doi.org/10.18653/v1/2022.emnlp-main.90, https://aclanthology.org/2022.emnlp-main.90 +26. Marino, K., Rastegari, M., Farhadi, A., Mottaghi, R.: Ok-vqa: A visual question answering benchmark requiring external knowledge. 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) pp. 3190-3199 (2019), https://api-semanticscholar.org/CorpusID:173991173 +27. OpenAI: Gpt-4 technical report. ArXiv abs/2303.08774 (2023), https://apisemantic scholar.org/CorpusID:257532815 +28. Press, O., Zhang, M., Min, S., Schmidt, L., Smith, N.A., Lewis, M.: Measuring and narrowing the compositionality gap in language models. ArXiv abs/2210.03350 (2022), https://api-semanticscholar.org/CorpusID:252762102 +29. Subramanian, S., Bogin, B., Gupta, N., Wolfson, T., Singh, S., Berant, J., Gardner, M.: Obtaining faithful interpretations from compositional neural networks. In: Annual Meeting of the Association for Computational Linguistics (2020), https://apisemantic scholar.org/CorpusID:218487535 +30. Subramanian, S., Narasimhan, M., Khangaonkar, K., Yang, K., Nagrani, A., Schmid, C., Zeng, A., Darrell, T., Klein, D.: Modular visual question answering via code generation. In: Rogers, A., Boyd-Graber, J., Okazaki, N. (eds.) Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers). pp. 747-761. Association for Computational Linguistics, Toronto, Canada (Jul 2023). https://doi.org/10.18653/v1/2023.acl-short.65, https://aclanthology.org/2023.acl-short.65 +31. Sur'is, D., Menon, S., Vondrick, C.: Vipergpt: Visual inference via python execution for reasoning. ArXiv abs/2303.08128 (2023), https://api_semanticscholar.org/CorpusID:257505358 +32. Suzgun, M., Scales, N., Scharli, N., Gehrmann, S., Tay, Y., Chung, H.W., Chowdhery, A., Le, Q.V., Chi, E.H., Zhou, D., et al.: Challenging big-bench tasks and whether chain-of-thought can solve them. arXiv preprint arXiv:2210.09261 (2022) +33. Touvron, H., Lavril, T., Izacard, G., Martinet, X., Lachaux, M.A., Lacroix, T., Rozière, B., Goyal, N., Hambro, E., Azhar, F., Rodriguez, A., Joulin, A., Grave, E., Lample, G.: Llama: Open and efficient foundation language models. ArXiv abs/2302.13971 (2023), https://apisemanticscholar.org/CorpusID: 257219404 +34. Wang, X., Li, S., Ji, H.: Code4struct: Code generation for few-shot structured prediction from natural language. arXiv preprint arXiv:2210.12810 (2022) +35. Wang, Z., Li, M., Xu, R., Zhou, L., Lei, J., Lin, X., Wang, S., Yang, Z., Zhu, C., Hoiem, D., et al.: Language models with image descriptors are strong few-shot video-language learners. Advances in Neural Information Processing Systems 35, 8483-8497 (2022) +36. Xiao, J., Shang, X., Yao, A., Chua, T.S.: Next-qa: Next phase of question-answering to explaining temporal actions. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 9777–9786 (June 2021) +37. Yang, K., Klein, D., Peng, N., Tian, Y.: Doc: Improving long story coherence with detailed outline control. In: Annual Meeting of the Association for Computational Linguistics (2023), https://api-semanticscholar.org/CorpusID:254877751 +38. Yao, S., Yu, D., Zhao, J., Shafran, I., Griffiths, T.L., Cao, Y., Narasimhan, K.: Tree of thoughts: Deliberate problem solving with large language models. arXiv preprint arXiv:2305.10601 (2023) + +39. Ye, Q., Xu, G., Yan, M., Xu, H., Qian, Q., Zhang, J., Huang, F.: Hitea: Hierarchical temporal-aware video-language pre-training. ArXiv abs/2212.14546 (2022), https://api-semanticscholar.org/CorpusID:255340506 +40. Yu, W., Gileadi, N., Fu, C., Kirmani, S., Lee, K.H., Gonzalez Arenas, M., Lewis Chiang, H.T., Erez, T., Hasenclever, L., Humplik, J., Ichter, B., Xiao, T., Xu, P., Zeng, A., Zhang, T., Heess, N., Sadigh, D., Tan, J., Tassa, Y., Xia, F.: Language to rewards for robotic skill synthesis. Arxiv preprint arXiv:2306.08647 (2023) +41. Zhang, P., Li, X., Hu, X., Yang, J., Zhang, L., Wang, L., Choi, Y., Gao, J.: Vinvl: Revisiting visual representations in vision-language models. 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) pp. 5575-5584 (2021), https://api_semanticscholar.org/CorpusID:235692795 +42. Zhou, D., Scharli, N., Hou, L., Wei, J., Scales, N., Wang, X., Schuurmans, D., Bousquet, O., Le, Q., hsin Chi, E.H.: Least-to-most prompting enables complex reasoning in large language models. ArXiv abs/2205.10625 (2022), https://api_semanticscholar.org/CorpusID:248986239 \ No newline at end of file diff --git a/2024/Recursive Visual Programming/images.zip b/2024/Recursive Visual Programming/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..1b57f27f4771b22fb74184a38eed43c803b5c252 --- /dev/null +++ b/2024/Recursive Visual Programming/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:69706bc6fc5669ddc9222815c1a14321d6dce21b590462f9a057ba2ae887f881 +size 735660 diff --git a/2024/Recursive Visual Programming/layout.json b/2024/Recursive Visual Programming/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..0719975bad3baa55637322fc15618c25784f2e66 --- /dev/null +++ b/2024/Recursive Visual Programming/layout.json @@ -0,0 +1,7377 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 197, + 112, + 416, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 197, + 112, + 416, + 129 + ], + "spans": [ + { + "bbox": [ + 197, + 112, + 416, + 129 + ], + "type": "text", + "content": "Recursive Visual Programming" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 194, + 149, + 418, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 194, + 149, + 418, + 175 + ], + "spans": [ + { + "bbox": [ + 194, + 149, + 418, + 175 + ], + "type": "text", + "content": "Jiaxin Ge1 Sanjay Subramanian1 Baifeng Shi1 Roei Herzig1 Trevor Darrell1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 258, + 184, + 355, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 258, + 184, + 355, + 194 + ], + "spans": [ + { + "bbox": [ + 258, + 184, + 355, + 194 + ], + "type": "text", + "content": "UC Berkeley, CA, USA" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 160, + 223, + 452, + 420 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 223, + 452, + 420 + ], + "spans": [ + { + "bbox": [ + 160, + 223, + 452, + 420 + ], + "type": "text", + "content": "Abstract. Visual Programming (VP) has emerged as a powerful framework for Visual Question Answering (VQA). By generating and executing bespoke code for each question, these methods show advancement in leveraging Large Language Models (LLMs) for complex problem-solving. Despite their potential, existing VP methods generate all code in a single function, which does not fully utilize LLM's reasoning capacity and the modular adaptability of code. This results in code that is suboptimal in terms of both accuracy and interpretability. Inspired by human coding practices, we propose Recursive Visual Programming (RVP), which better harnesses the reasoning capacity of LLMs, provides modular code structure between code pieces, and assigns different return types for the sub-problems elegantly. RVP approaches VQA tasks with an top-down recursive code generation approach, allowing decomposition of complicated problems into smaller parts. We show RVP's efficacy through extensive experiments on benchmarks including VSR, COVR, GQA, and NextQA, underscoring the value of adopting human-like recursive and modular programming techniques for solving VQA tasks. Our code is available at https://github.com/para-lost/RVP." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 160, + 430, + 405, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 430, + 405, + 441 + ], + "spans": [ + { + "bbox": [ + 160, + 430, + 405, + 441 + ], + "type": "text", + "content": "Keywords: Visual Programming " + }, + { + "bbox": [ + 160, + 430, + 405, + 441 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 430, + 405, + 441 + ], + "type": "text", + "content": " Vision-Language " + }, + { + "bbox": [ + 160, + 430, + 405, + 441 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 430, + 405, + 441 + ], + "type": "text", + "content": " VQA" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 461, + 230, + 474 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 461, + 230, + 474 + ], + "spans": [ + { + "bbox": [ + 132, + 461, + 230, + 474 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 486, + 481, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 486, + 481, + 582 + ], + "spans": [ + { + "bbox": [ + 130, + 486, + 481, + 582 + ], + "type": "text", + "content": "Visual Question Answering (VQA) lies at the intersection of computer vision and natural language processing, posing the challenge of interpreting visual data to answer questions [3,16,26]. Following the recent progress on code generation using Large Language Models (LLMs) [6,27,33], Visual Programming (VP) methods have notably advanced in this area, successfully using LLMs to generate and execute code in few-shot and zero-shot scenarios [14,30,31]. In these techniques, various vision models are provided to LLM as APIs, and the LLM is used as a planner to utilize these vision models to perform reasoning over images." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 582, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 481, + 665 + ], + "type": "text", + "content": "Despite these advancements, such planning remains very challenging for Visual Programming. These methods require the model to generate a single code block using the predefined APIs at once. As a result, the model must handle logic, be aware of all details, and utilize the vision APIs appropriately. This has two main drawbacks. First, such approaches do not fully utilize LLM's reasoning capability, which would benefit from breaking down the problem into step-by-step sub-problems, as demonstrated in recent NLP research [4]. Secondly, they" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 136, + 114, + 479, + 324 + ], + "blocks": [ + { + "bbox": [ + 136, + 114, + 479, + 324 + ], + "lines": [ + { + "bbox": [ + 136, + 114, + 479, + 324 + ], + "spans": [ + { + "bbox": [ + 136, + 114, + 479, + 324 + ], + "type": "image", + "image_path": "cc5ac8908a42096357813ed66e120177bf0638873d475f7f0f643930020c20da.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 331, + 482, + 410 + ], + "lines": [ + { + "bbox": [ + 130, + 331, + 482, + 410 + ], + "spans": [ + { + "bbox": [ + 130, + 331, + 482, + 410 + ], + "type": "text", + "content": "Fig. 1: A breakdown of Recursive Visual Programming for a visual question is illustrated with an image where scenes within it contain smaller versions of themselves. To locate a man wearing a hat at the third level of an image, the model first generates code that has two recursive_query API calls. The initial call identifies the largest mirror (representing the next image level), and the subsequent call seeks \"the man with a hat\" at the second level. The model iteratively generates and executes code through these calls until the final answer is produced without additional recursive calls." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 435, + 481, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 435, + 481, + 460 + ], + "spans": [ + { + "bbox": [ + 130, + 435, + 481, + 460 + ], + "type": "text", + "content": "fail to utilize the full potential of code because they lack modularity between the code pieces, particularly when the code is one long piece." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 460, + 481, + 520 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 460, + 481, + 520 + ], + "spans": [ + { + "bbox": [ + 130, + 460, + 481, + 520 + ], + "type": "text", + "content": "We consider how humans tackle these issues. Humans tend to code in a top-down manner: first they break down complex tasks into manageable subproblems, and then they tackle each through focused code segments. This approach allows the programmer to concentrate on the main logic flow rather than figuring out all the details at once, resulting in a more clear and elegant structure." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 521, + 482, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 482, + 628 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 482, + 628 + ], + "type": "text", + "content": "Inspired by this, we hypothesize that incorporating a similar recursive coding strategy would also help solve complex visual question-answering problems. To investigate whether LLMs could indeed benefit from recursive coding, we conduct preliminary experiment of the recursive coding approach on standard symbolic reasoning benchmarks, such as Dyck Language [32] and Games of 24 [38]. This is done by simply prompting the LLM to code recursively. We find that utilizing LLMs to code recursively leads to a notable improvement in accuracy, as shown in Figure 2. Motivated by this result, we explore in this paper how a recursive coding methodology can be applied to VQA tasks." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "text", + "content": "Specifically, we propose a novel VP method for VQA tasks, named Recursive Visual Programming (RVP). Instead of querying LLM only once to generate a single piece of code, we ask LLM to decompose the question into sub-questions" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 203, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 203, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 203, + 100 + ], + "type": "text", + "content": "Ge et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 134, + 115, + 478, + 236 + ], + "blocks": [ + { + "bbox": [ + 134, + 115, + 478, + 236 + ], + "lines": [ + { + "bbox": [ + 134, + 115, + 478, + 236 + ], + "spans": [ + { + "bbox": [ + 134, + 115, + 478, + 236 + ], + "type": "image", + "image_path": "a11a8f66fb1cbde1cb07d94fb4fa00b38ebd21d81121ba7bbb9dd4d01c6c283c.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 243, + 482, + 300 + ], + "lines": [ + { + "bbox": [ + 130, + 243, + 482, + 300 + ], + "spans": [ + { + "bbox": [ + 130, + 243, + 482, + 300 + ], + "type": "text", + "content": "Fig. 2: Motivating Examples. Adding the prompt \"solve the problem recursively\" to the input significantly improves the model's performance on traditional coding tasks, Dyck language [32] and Games of 24 [38]. \"NR\" refers to non-recursive approach and \"R\" refers to recursive approach. This shows that LLMs can potentially benefit from human's recursive programming approach." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 312, + 482, + 419 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 312, + 482, + 419 + ], + "spans": [ + { + "bbox": [ + 130, + 312, + 482, + 419 + ], + "type": "text", + "content": "and recursively query itself for each sub-question to generate modular code pieces. For each sub-question, the same process repeats, until reaching an atom level where the current question can be solved by only calling external APIs. See example in Figure 1. In contrast to querying LLM only once, we find that using recursive queries to generate code for each sub-question can improve the readability and accuracy of the code, as shown in Figure 3. Moreover, MVP allows assigning dynamic types to each sub-question, enabling new code pieces to return a variety of types based on the current need. Remarkably, MVP assigns return types unseen in in-context examples, such as List[str], List[ImagePatch]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 420, + 482, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 420, + 482, + 529 + ], + "spans": [ + { + "bbox": [ + 130, + 420, + 482, + 529 + ], + "type": "text", + "content": "Our contribution can be summarized as follows: (i) We introduce a novel recursive approach based on human programming. This approach is a top-down programming method that better harnesses LLM's reasoning capacity and enhances modularity between multiple code pieces. It is easily applicable to any existing VP model for visual question answering task. (ii) We allow dynamic return types for the sub-problems, which enhances flexibility and adaptability. (iii) We demonstrate improved performance on several standard benchmarks, such as GQA, VSR, NextQA and COVR, demonstrating the effectiveness of our approach. We also show better interpretability than non-recursive VP methods." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 545, + 238, + 558 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 545, + 238, + 558 + ], + "spans": [ + { + "bbox": [ + 132, + 545, + 238, + 558 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": "Modular Visual Reasoning. There is a long line of work that seeks to combine modularity with the expressivity of deep neural networks. Early work on this topic converts the output of a parser [1] to a program or trains a program generator either separately from or jointly with the execution modules [2, 15]. Later work finds that such models are often challenging to train on real-world datasets, particularly when interpretability of individual module outputs is expected [29]. Other work improves performance on multi-hop visual reasoning without explicitly defining a set of modules, e.g., by predicting and operating" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 321, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 321, + 91, + 447, + 102 + ], + "type": "text", + "content": "Recursive Visual Programming" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 135, + 114, + 481, + 327 + ], + "blocks": [ + { + "bbox": [ + 135, + 114, + 481, + 327 + ], + "lines": [ + { + "bbox": [ + 135, + 114, + 481, + 327 + ], + "spans": [ + { + "bbox": [ + 135, + 114, + 481, + 327 + ], + "type": "image", + "image_path": "a8b7ab2795a081c333bac9386e43bd962bfd8fe27c57853ee5b1b9e61417c441.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 335, + 482, + 381 + ], + "lines": [ + { + "bbox": [ + 130, + 335, + 482, + 381 + ], + "spans": [ + { + "bbox": [ + 130, + 335, + 482, + 381 + ], + "type": "text", + "content": "Fig. 3: Example from COVR (Upper). Current VP methods fail to provide binary 'yes' or 'no' answers, while Recursive Visual Programming method outputs the correct answer. Example from GQA (Middle and Bottom). Recursive outperforms current non-recursive methods by correctly addressing all details and their associated logic." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 426, + 237, + 438 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 426, + 237, + 438 + ], + "spans": [ + { + "bbox": [ + 130, + 426, + 237, + 438 + ], + "type": "text", + "content": "over a scene graph [18]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 438, + 482, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 438, + 482, + 581 + ], + "spans": [ + { + "bbox": [ + 130, + 438, + 482, + 581 + ], + "type": "text", + "content": "Visual Programming. Recent language models can accurately generate code from text descriptions when prompted with a small number of examples [8]. This advance has enabled a few-shot modular visual reasoning approach, commonly called visual programming [14], which consists of prompting a language model for a program and executing the program with pre-trained vision models. This paradigm has led to impressive results in visual question answering, video question answering, visual and tabular question answering, referring expression comprehension, and text-based image editing [11, 12, 23, 30, 31]. Programming with LLMs has also been applied to produce plans and reward functions for robots [21, 24, 40]. Our work builds on these techniques by enabling the LLM to generate code in multiple steps-first planning the overall process and then filling in the details-rather than all at once." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 582, + 483, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 483, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 483, + 666 + ], + "type": "text", + "content": "Language Models for Reasoning Tasks. In the text domain, there is a large body of work on using language models for reasoning tasks. In particular, several papers have shown the efficacy of code generation as a scaffold for such tasks [9, 13, 25, 34]. Aside from code generation, prior work has also shown that decomposing complex queries into simpler ones improves the performance of language models [28, 42]. Our RVP approach applies this insight to improve the code generation component of visual programming." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 203, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 203, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 203, + 101 + ], + "type": "text", + "content": "Ge et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 136, + 114, + 479, + 315 + ], + "blocks": [ + { + "bbox": [ + 136, + 114, + 479, + 315 + ], + "lines": [ + { + "bbox": [ + 136, + 114, + 479, + 315 + ], + "spans": [ + { + "bbox": [ + 136, + 114, + 479, + 315 + ], + "type": "image", + "image_path": "289e4bc1dc3f4c4c4b0e6b14f23b1457f7727744791a3995762557724aba68b7.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 323, + 482, + 357 + ], + "lines": [ + { + "bbox": [ + 130, + 323, + 482, + 357 + ], + "spans": [ + { + "bbox": [ + 130, + 323, + 482, + 357 + ], + "type": "text", + "content": "Fig. 4: An overview of our approach. When a recursive_query API call is occurred, the model generates and executes new code based on the query, returning the output back to the original call." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 378, + 339, + 392 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 378, + 339, + 392 + ], + "spans": [ + { + "bbox": [ + 131, + 378, + 339, + 392 + ], + "type": "text", + "content": "3 Recursive Visual Programming" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 402, + 482, + 463 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 402, + 482, + 463 + ], + "spans": [ + { + "bbox": [ + 130, + 402, + 482, + 463 + ], + "type": "text", + "content": "In this section, we introduce our method, Recursive Visual Programming (RVP). We first review the problem setup and ingredients of visual programming as established by prior work (see Section 3.1). We then introduce our recursive coding framework (see Section 3.2), which involves a multi-step prompting scheme, as well as a system for dynamic type assignment (see Section 3.3)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 479, + 228, + 491 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 479, + 228, + 491 + ], + "spans": [ + { + "bbox": [ + 131, + 479, + 228, + 491 + ], + "type": "text", + "content": "3.1 Preliminaries" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 498, + 482, + 535 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 498, + 482, + 535 + ], + "spans": [ + { + "bbox": [ + 131, + 498, + 482, + 535 + ], + "type": "text", + "content": "In Visual Question Answering (VQA), the system " + }, + { + "bbox": [ + 131, + 498, + 482, + 535 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 131, + 498, + 482, + 535 + ], + "type": "text", + "content": " is given a visual input " + }, + { + "bbox": [ + 131, + 498, + 482, + 535 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 131, + 498, + 482, + 535 + ], + "type": "text", + "content": " (i.e. an image, a set of images, or a video), as well as a question " + }, + { + "bbox": [ + 131, + 498, + 482, + 535 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 131, + 498, + 482, + 535 + ], + "type": "text", + "content": " relating to the visual content. The expected output is a textual answer " + }, + { + "bbox": [ + 131, + 498, + 482, + 535 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 131, + 498, + 482, + 535 + ], + "type": "text", + "content": " to the question:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 277, + 543, + 481, + 557 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 277, + 543, + 481, + 557 + ], + "spans": [ + { + "bbox": [ + 277, + 543, + 481, + 557 + ], + "type": "interline_equation", + "content": "A \\leftarrow f (I, Q) \\tag {1}", + "image_path": "dba73d81bc1633f66faa4e0f74d3d1b92900824b150f96d02745820cd8f1ea3e.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 563, + 482, + 612 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 563, + 482, + 612 + ], + "spans": [ + { + "bbox": [ + 130, + 563, + 482, + 612 + ], + "type": "text", + "content": "Visual Programming methods [14,30,31] represent " + }, + { + "bbox": [ + 130, + 563, + 482, + 612 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 130, + 563, + 482, + 612 + ], + "type": "text", + "content": " as the composition of two functions: a code generator " + }, + { + "bbox": [ + 130, + 563, + 482, + 612 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 130, + 563, + 482, + 612 + ], + "type": "text", + "content": " and a visual executor " + }, + { + "bbox": [ + 130, + 563, + 482, + 612 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 130, + 563, + 482, + 612 + ], + "type": "text", + "content": ". In the first part, the code generator generates a program " + }, + { + "bbox": [ + 130, + 563, + 482, + 612 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 130, + 563, + 482, + 612 + ], + "type": "text", + "content": " conditioned on the question, and in the second part, the executor executes the program on the visual input:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 254, + 620, + 481, + 634 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 254, + 620, + 481, + 634 + ], + "spans": [ + { + "bbox": [ + 254, + 620, + 481, + 634 + ], + "type": "interline_equation", + "content": "P \\leftarrow g (Q), A \\leftarrow h (P, I) \\tag {2}", + "image_path": "e63d2122334e0aa08529b7cd53aac01dd8645d71ad73f15e611c2a111aa5e254.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 131, + 641, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 641, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 131, + 641, + 481, + 666 + ], + "type": "text", + "content": "We focus on the few-shot setting in which we have only a few (at most 50) training examples—each consisting of a question, program, and answer—for a given" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 321, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 321, + 91, + 447, + 102 + ], + "type": "text", + "content": "Recursive Visual Programming" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "text", + "content": "task. Therefore, the code generator " + }, + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "text", + "content": " is composed of a language model and a prompt that includes a description of the API available to the executor and in-context examples of questions along with the corresponding correct programs. We use Python as the programming language, similar to [30, 31]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 165, + 479, + 201 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 165, + 479, + 201 + ], + "spans": [ + { + "bbox": [ + 130, + 165, + 479, + 201 + ], + "type": "text", + "content": "We evaluate our approach using the ViperGPT [31] API. For questions over images or sets of images, these functions can be called on an ImagePatch object, which can be any crop of the image:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 136, + 213, + 481, + 374 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 136, + 213, + 481, + 237 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 213, + 481, + 237 + ], + "spans": [ + { + "bbox": [ + 136, + 213, + 481, + 237 + ], + "type": "text", + "content": "1. find(object: str): Returns a list of ImagePatch objects matching the description." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 136, + 239, + 481, + 262 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 239, + 481, + 262 + ], + "spans": [ + { + "bbox": [ + 136, + 239, + 481, + 262 + ], + "type": "text", + "content": "2. exists(object: str): Returns a bool representing whether the described object exists." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 136, + 264, + 481, + 287 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 264, + 481, + 287 + ], + "spans": [ + { + "bbox": [ + 136, + 264, + 481, + 287 + ], + "type": "text", + "content": "3. verify_property(object: str, property: str): Returns a bool representing whether the described object has the specified property." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 136, + 289, + 481, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 289, + 481, + 312 + ], + "spans": [ + { + "bbox": [ + 136, + 289, + 481, + 312 + ], + "type": "text", + "content": "4. simple_query(question: str): Returns an answer to the provided question based on the given ImagePatch." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 136, + 315, + 481, + 338 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 315, + 481, + 338 + ], + "spans": [ + { + "bbox": [ + 136, + 315, + 481, + 338 + ], + "type": "text", + "content": "5. compute_depth(): Returns the median depth of the region in the given ImagePatch." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 136, + 340, + 481, + 374 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 340, + 481, + 374 + ], + "spans": [ + { + "bbox": [ + 136, + 340, + 481, + 374 + ], + "type": "text", + "content": "6. crop(left: int, lower: int, right: int, upper: int): Crops a part of the given ImagePatch and returns a new ImagePatch representing that crop." + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 130, + 387, + 479, + 423 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 387, + 479, + 423 + ], + "spans": [ + { + "bbox": [ + 130, + 387, + 479, + 423 + ], + "type": "text", + "content": "For questions about a video, the API includes three other functions: trim(), frame_from_index(), and frame_iterator(), which return an abbreviated version of the video, a specific frame, and an iterator over the frames respectively." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 131, + 447, + 318, + 459 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 447, + 318, + 459 + ], + "spans": [ + { + "bbox": [ + 131, + 447, + 318, + 459 + ], + "type": "text", + "content": "3.2 Recursive Visual Programming" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 472, + 481, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 472, + 481, + 616 + ], + "spans": [ + { + "bbox": [ + 130, + 472, + 481, + 616 + ], + "type": "text", + "content": "We introduce a new method for code generation in visual programming: Recursive Visual Programming (RVP). RVP enables the language model to break the code generation process into multiple sub-questions and recursively call itself to generate a program for each sub-question. As shown in Figure 4, we operationalize RVP by adding a new function to the API: recursive_query(image, sub-question). During execution, when recursive_query is called, it prompts the language model to generate a new code given the sub-question. This code is then executed, and its return value is returned by recursive_query. An example break down of the usage of recursive_query call is shown in Figure 1. We set the recursive termination condition by (1) setting max depth of 10 and (2) switching to direct simple query when generated code contains a recursive call to the same input question." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "type": "text", + "content": "To enable this process, we add examples of calling recursive_query in the prompt to show the model how to decompose the problem and utilize the answers returned by recursive_query. We provide examples of the in-context prompt of recursive_query we used in the supplementary material." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 203, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 203, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 203, + 101 + ], + "type": "text", + "content": "Ge et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 137, + 116, + 477, + 288 + ], + "blocks": [ + { + "bbox": [ + 137, + 116, + 477, + 288 + ], + "lines": [ + { + "bbox": [ + 137, + 116, + 477, + 288 + ], + "spans": [ + { + "bbox": [ + 137, + 116, + 477, + 288 + ], + "type": "image", + "image_path": "e847bf6c00ebf60a2f11847f5d049b303e64d5069766beb8bd8f16965ff7d17b.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 297, + 482, + 331 + ], + "lines": [ + { + "bbox": [ + 130, + 297, + 482, + 331 + ], + "spans": [ + { + "bbox": [ + 130, + 297, + 482, + 331 + ], + "type": "text", + "content": "Fig.5: Examples of dynamic type assignment in recursive_query calls. The model autonomously determines the appropriate return type in its code generation and generalize to new types unseen in in-context examples such as List[ImagePatch]." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 361, + 297, + 374 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 361, + 297, + 374 + ], + "spans": [ + { + "bbox": [ + 132, + 361, + 297, + 374 + ], + "type": "text", + "content": "3.3 Dynamic Type Assignment" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 402, + 482, + 570 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 402, + 482, + 570 + ], + "spans": [ + { + "bbox": [ + 130, + 402, + 482, + 570 + ], + "type": "text", + "content": "In previous visual programming frameworks, code generation mechanisms are constrained by static type assignments, i.e., the return type of a given function is fixed. For instance, simple_query always returns a string. However, this rigid structure limits the model's flexibility to adapt the code to different contexts within the VQA task. RVP introduces dynamic type assignment to address this limitation. When the code generator writes a recursive_query function, it can include the expected return type as a part of the question that is passed to the recursive_query. To facilitate this, in the in-context examples that include recursive_query, we let the input 'question' specify a type. As shown in Figure 5, we give examples of type ImagePatch, bool, float, and str being specified within the queries. When receiving this question, the model discerns the required type and generates code that declares and returns the correct type. For instance, a query like \"Return a bool, is there a black cat in the picture?\" will prompt the generation of code with the signature." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 254, + 587, + 481, + 601 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 254, + 587, + 481, + 601 + ], + "spans": [ + { + "bbox": [ + 254, + 587, + 481, + 601 + ], + "type": "interline_equation", + "content": "\\operatorname {e x e c u t e} (\\operatorname {i m a g e}) \\rightarrow \\operatorname {b o o l} \\tag {3}", + "image_path": "cb4dde4afbd4993cffaf3879144c3b569d93740abb7ab4fa3ef178746782995f.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "This signature indicates that the generated code " + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "inline_equation", + "content": "P'" + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": ", when executed, will return a boolean value " + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "inline_equation", + "content": "A_{\\mathrm{bool}}" + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": " which answers the sub-question " + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "inline_equation", + "content": "Q'" + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": ", illustrating the dynamic and context-sensitive nature of RVP's code generation capabilities that help to produce syntactically and semantically appropriate code for complex reasoning." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 321, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 321, + 91, + 447, + 102 + ], + "type": "text", + "content": "Recursive Visual Programming" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 230, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 230, + 129 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 230, + 129 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 139, + 480, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 139, + 480, + 163 + ], + "spans": [ + { + "bbox": [ + 132, + 139, + 480, + 163 + ], + "type": "text", + "content": "In this section, we introduce the basic settings including the task formulation and parameter details. Then, we provide experiment results and in-depth analysis." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 171, + 253, + 183 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 171, + 253, + 183 + ], + "spans": [ + { + "bbox": [ + 132, + 171, + 253, + 183 + ], + "type": "text", + "content": "4.1 Experiment Setup" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 190, + 482, + 345 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 190, + 482, + 345 + ], + "spans": [ + { + "bbox": [ + 130, + 190, + 482, + 345 + ], + "type": "text", + "content": "Datasets. We evaluate on four datasets spanning a wide range of visual reasoning skills and types of visual input: Visual Spatial Reasoning (VSR) [22], GQA [17], COVR [5], and NextQA [36]. For NextQA, we present results on the \"Hard Temporal\" subset which was curated to consist of questions that truly require temporal reasoning [7]. VSR and GQA include queries about single images, while COVR includes queries about multiple images, and NextQA includes queries about video. On each dataset, we evaluate methods by their accuracy in producing an answer that exactly matches the ground-truth answer. For NextQA, we adopt a multiple-choice setup in which we provide methods with the set of possible answer choices for each question. For VSR, a spatial statement of an image is made, and the question is framed as : \"Is it true or false,\" followed by the statement. In this way, a VSR statement is framed as a VQA question and the model will answer either true or false." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 346, + 482, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 346, + 482, + 418 + ], + "spans": [ + { + "bbox": [ + 130, + 346, + 482, + 418 + ], + "type": "text", + "content": "Implementation Details. We implement RVP using the ViperGPT API and the accompanying prompt template. For each dataset, we curate a set of in-context examples which are added to the prompt. The set of question-answer pairs is shared between the recursive and non-recursive methods. We use seven in-context examples for GQA, six for NextQA, six for VSR, and eight for COVR. For code generation, use GPT-3.5-turbo via the OpenAI API1." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 435, + 265, + 447 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 435, + 265, + 447 + ], + "spans": [ + { + "bbox": [ + 132, + 435, + 265, + 447 + ], + "type": "text", + "content": "4.2 Quantitative Results" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 454, + 484, + 550 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 454, + 484, + 550 + ], + "spans": [ + { + "bbox": [ + 130, + 454, + 484, + 550 + ], + "type": "text", + "content": "Table 1 presents our results across four datasets. In all settings, we outperform ViperGPT, which is identical to RVP except that RVP includes recursive_query function in the API and in-context examples. This indicates that recursive programming indeed improves accuracy in visual reasoning tasks. Moreover, our method outperforms previous zero/few-shot methods in all but one of the settings. While our primary aim was not to push the state-of-the-art results, we show that integrating the clarity and elegance of recursive coding into standard VP methods not only doesn't impair performance, but can enhance it." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 567, + 258, + 578 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 567, + 258, + 578 + ], + "spans": [ + { + "bbox": [ + 132, + 567, + 258, + 578 + ], + "type": "text", + "content": "4.3 Qualitative Results" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 586, + 482, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 586, + 482, + 647 + ], + "spans": [ + { + "bbox": [ + 132, + 586, + 482, + 647 + ], + "type": "text", + "content": "We provide qualitative examples in Figure 3,6,7. In the example from COVR in Figure 3, the non-recursive approach answers with neither \"yes\" nor \"no,\" while the recursive approach does confirm to these choices. One possible explanation for this issue is that LLMs struggle to maintain coherence across long contexts [37]. RVP circumvents this issue by pushing the LLM to commit to the answer" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 203, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 203, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 203, + 101 + ], + "type": "text", + "content": "Ge et al." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 133, + 653, + 250, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 653, + 250, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 653, + 250, + 665 + ], + "type": "text", + "content": "1 wwwPLATFORM.openai.com" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 155, + 201, + 455, + 292 + ], + "blocks": [ + { + "bbox": [ + 130, + 114, + 482, + 193 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 482, + 193 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 482, + 193 + ], + "type": "text", + "content": "Table 1: Results. We report exact-match accuracy on each dataset. Best zero/few-shot result for each dataset is bolded. †Fully supervised results are from different models: ViLT [19] for VSR, HiTeA [39] for NextQA, VinVL-Base [41] for GQA, and VisualBERT [20] for COVR. ‡We report the results of our reproduction of ViperGPT using the officially released code and the same in-context examples used for our method (since the original in-context examples from ViperGPT were not released). We use GPT3.5-turbo while CodeVQA uses Codex, which is not available anymore." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 155, + 201, + 455, + 292 + ], + "lines": [ + { + "bbox": [ + 155, + 201, + 455, + 292 + ], + "spans": [ + { + "bbox": [ + 155, + 201, + 455, + 292 + ], + "type": "table", + "html": "
MethodVSRNextQAGQA Test-devCOVR Test
Random SplitZero-shot SplitHard Split-TTest-dev
Fully supervised†69.363.048.665.157.9
Zero/few-shot methods
CLIP56.054.5---
BLIPv2---42.31-
CodeVQA---49.050.7
ViperGPT‡61.2561.5947.2144.6351.69
RVP (ours)63.5366.0948.8245.6252.67
", + "image_path": "496ab97313fb4e20d67846ce56a2339f9cdcdbe1199030901a99f073198dd51d.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 315, + 482, + 481 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 315, + 482, + 481 + ], + "spans": [ + { + "bbox": [ + 130, + 315, + 482, + 481 + ], + "type": "text", + "content": "types earlier in its generation (the length from the beginning of the code to the point \"return xxx\" is shorter). In Figure 6, the non-recursive approach fails to handle the logic correctly while the recursive approach correctly does. This is likely due to that RVP breaks down the code logic and each code piece only needs to focus on a single logic, while the non-recursive method attempts to handle all the logic at once. In the GQA example, a common pattern involves handling complex logical problems and multiple properties. Non-recursive methods often fell short in these cases, either overlooking critical details or processing them in a logically flawed manner. RVP manages these intricate questions by addressing each detail accurately. This difference may stem from the inherent difficulty LLMs face in processing complicated logic over extended contexts. Recursive methods simplify this by focusing on the overall logic and gradually breaking the problems down into smaller, more manageable segments. This facilitates easier and more accurate resolution." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 482, + 482, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 482, + 482, + 544 + ], + "spans": [ + { + "bbox": [ + 130, + 482, + 482, + 544 + ], + "type": "text", + "content": "We provide a dynamic type assignment example in Figure 5. In this example, RVP assigns the return type List[ImagePatch] in the recursive_query call. It's worth noting that the return type List[ImagePatch] is unseen in the in-context examples. This suggests that RVP has the potential to utilize complicated and diverse data structures, highlight the effectiveness of our approach." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 560, + 238, + 572 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 560, + 238, + 572 + ], + "spans": [ + { + "bbox": [ + 131, + 560, + 238, + 572 + ], + "type": "text", + "content": "4.4 Ablation Study" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 581, + 482, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 581, + 482, + 629 + ], + "spans": [ + { + "bbox": [ + 130, + 581, + 482, + 629 + ], + "type": "text", + "content": "Dynamic Type This section studies integrating dynamic type assignment within the recursive_query API. For the GQA dataset, we examined all instances with recursive patterns in the test_dev split. In COVR, where the test set is unavailable, we used the val_1000 samples from the validation set, following [30]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 630, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 630, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 630, + 482, + 665 + ], + "type": "text", + "content": "We compare four approaches: Non-recursive Baseline: Standard approach without recursive features. Fixed-Type Recursive: The recursive_query API is constrained to return a string. Dynamic-Implicit Recursive: The API" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 320, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 320, + 91, + 447, + 102 + ], + "type": "text", + "content": "Recursive Visual Programming" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 134, + 115, + 480, + 418 + ], + "blocks": [ + { + "bbox": [ + 134, + 115, + 480, + 418 + ], + "lines": [ + { + "bbox": [ + 134, + 115, + 480, + 418 + ], + "spans": [ + { + "bbox": [ + 134, + 115, + 480, + 418 + ], + "type": "image", + "image_path": "96032daf77e87f35d0d7e22e44648709dd0d14974fd7eb030569176dd9312844.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 426, + 482, + 450 + ], + "lines": [ + { + "bbox": [ + 132, + 426, + 482, + 450 + ], + "spans": [ + { + "bbox": [ + 132, + 426, + 482, + 450 + ], + "type": "text", + "content": "Fig. 6: Examples from COVR. RVP allows correct logic flow while traditional nonrecursive VP fails to handle the logic correctly." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 478, + 482, + 528 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 478, + 482, + 528 + ], + "spans": [ + { + "bbox": [ + 130, + 478, + 482, + 528 + ], + "type": "text", + "content": "can return various types, but the specific type is not predefined in the query. For example, \"Is there a black cat?\" could yield either a boolean or a string. Dynamic-Explicit Recursive: Our primary approach, where the API's input query specifies the return type, like \"Return a bool, is there a black cat?\"" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 533, + 483, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 483, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 483, + 666 + ], + "type": "text", + "content": "Table 2 shows our findings. The dynamic-explicit type assignment outperformed other methods, underscoring the need for clear type specification. The dynamic-implicit method showed lower performance than non-recursive due to type ambiguities, as exemplified in Figure 8, where the first code piece assumes the return type is a bool, while the second code piece returns a str. Fixed-type assignment proves to be beneficial over non-recursive methods, indicating that recursion itself adds value, but its rigidity in type handling limits performance. This study highlights the benefits of flexible, dynamic type assignment and the crucial role of explicit type declaration for accurate and consistent responses. These insights mark a step towards more adaptable visual reasoning frameworks, emphasizing the need for precision in dynamic environments." + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 134, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 134, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 203, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 203, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 203, + 100 + ], + "type": "text", + "content": "Ge et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 137, + 114, + 477, + 276 + ], + "blocks": [ + { + "bbox": [ + 137, + 114, + 477, + 276 + ], + "lines": [ + { + "bbox": [ + 137, + 114, + 477, + 276 + ], + "spans": [ + { + "bbox": [ + 137, + 114, + 477, + 276 + ], + "type": "image", + "image_path": "68ee3982eab4da0234e24377af0769c55cfb81cf4fbef81d73669132562e818b.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 137, + 277, + 477, + 462 + ], + "blocks": [ + { + "bbox": [ + 137, + 277, + 477, + 462 + ], + "lines": [ + { + "bbox": [ + 137, + 277, + 477, + 462 + ], + "spans": [ + { + "bbox": [ + 137, + 277, + 477, + 462 + ], + "type": "image", + "image_path": "7289d0e956b6012d08344c50375c505214c6fb3d6a09b6bebfa0f820f4e5b31e.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 472, + 481, + 495 + ], + "lines": [ + { + "bbox": [ + 132, + 472, + 481, + 495 + ], + "spans": [ + { + "bbox": [ + 132, + 472, + 481, + 495 + ], + "type": "text", + "content": "Fig. 7: Examples from GQA. RVP allows handling the details more elegantly and produces logically correct code." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 521, + 480, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 480, + 545 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 480, + 545 + ], + "type": "text", + "content": "In-Context Examples We conduct ablation on using retrieval-based in-context example selection and using different number of in-context examples." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "type": "text", + "content": "Retrieval-based in-context example selection. We manage to write 9 programs using the recursive method from the 50 provided by CodeVQA [30], and add our original 3 recursive programs to them. We keep the non-recursive in-context example unchanged and choose 3 recursive in-context examples from these 12 for each question. We follow [35] to use the embedding-based retrieval technique that calculates a score for each example, and then select the top 3 examples for each question. The results of GQA_val_2000 are in Table 2. We find that the recursive pattern occurrence rate is significantly higher than using fixed in-context examples. This is likely because by retrieving similar recursively-decomposed questions, the model can more easily simulate their pat" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 321, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 321, + 91, + 447, + 102 + ], + "type": "text", + "content": "Recursive Visual Programming" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 187 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 187 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 187 + ], + "type": "text", + "content": "terns, thereby attempting to solve more problems recursively. However, the overall accuracy does not improve. This suggests that although the retrieval-based method encourages more diverse and suitable patterns of recursive decomposition, there is a need for improved techniques in selecting the example pool—such as focusing on those most suitable for recursive decomposition instead of randomly selecting—along with careful tuning of the example programs." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 188, + 482, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 188, + 482, + 236 + ], + "spans": [ + { + "bbox": [ + 130, + 188, + 482, + 236 + ], + "type": "text", + "content": "Number of in-context examples. We randomly add one example/delete one example from the covr_1000_val, the results are shown in Table 2. The results indicate that adding the number of examples can enhance the performance, but does not have a significant impact." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 247, + 482, + 355 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 247, + 482, + 355 + ], + "spans": [ + { + "bbox": [ + 130, + 247, + 482, + 355 + ], + "type": "text", + "content": "Error feedback loop Recent studies have studied the debugging ability of LLMs on code [10]. We study whether modular visual code structure enables better bug identification for visual programming. We sample 20 incorrect programs for 10 questions that have type/logic errors, and use GPT3.5 to (1) identify the bug, and (2) write the correct code given the bug. For identification, the model correctly finds 7 bugs correctly for RVP and 2 for non-recursive visual programming. For correction, the model fixes 7 bugs correctly for RVP and 7 for visual programming. This suggests that clearer and more modular code structure, as presented RVP, can potentially aids bug identification." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 364, + 482, + 431 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 364, + 482, + 431 + ], + "spans": [ + { + "bbox": [ + 130, + 364, + 482, + 431 + ], + "type": "text", + "content": "Table 2: Ablation study. (a) We provide ablation study results for dynamic type assignment and found that specified dynamic type assignment in RVP achieves the highest performance. (b) We provide results on using retrieval-based in-context examples and fixed examples. Retrieval-based method enables more diverse recursive patterns but requires careful choosing the prompt pool. (c) We provide results of using different number of examples and found this doesn't influence performance significantly." + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 138, + 463, + 269, + 531 + ], + "blocks": [ + { + "bbox": [ + 144, + 441, + 261, + 453 + ], + "lines": [ + { + "bbox": [ + 144, + 441, + 261, + 453 + ], + "spans": [ + { + "bbox": [ + 144, + 441, + 261, + 453 + ], + "type": "text", + "content": "(a) Dynamic Type Ablation." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 138, + 463, + 269, + 531 + ], + "lines": [ + { + "bbox": [ + 138, + 463, + 269, + 531 + ], + "spans": [ + { + "bbox": [ + 138, + 463, + 269, + 531 + ], + "type": "table", + "html": "
MethodGQACOVR
Non-Recursive55.9960.71
Fixed-Type65.0560.12
Dynamic-Implicit55.3453.57
Dynamic-Explicit70.2367.86
", + "image_path": "c7572b10aeb5498e15731da1d5b06aaab0a3a3cd9bc31fcf04faf953bd090d35.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 279, + 472, + 386, + 531 + ], + "blocks": [ + { + "bbox": [ + 280, + 441, + 386, + 464 + ], + "lines": [ + { + "bbox": [ + 280, + 441, + 386, + 464 + ], + "spans": [ + { + "bbox": [ + 280, + 441, + 386, + 464 + ], + "type": "text", + "content": "(b) Retrieval-Based and Fixed Examples" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 279, + 472, + 386, + 531 + ], + "lines": [ + { + "bbox": [ + 279, + 472, + 386, + 531 + ], + "spans": [ + { + "bbox": [ + 279, + 472, + 386, + 531 + ], + "type": "table", + "html": "
FixedRetrieval
Accuracy54.7553.79
Rec Rate3.55%14.85%
Rec Acc71.8357.91
", + "image_path": "893966c64d7c77f5fe3a0a53d68e6a634e877000ec23d7becc5181806f12162d.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 398, + 473, + 476, + 531 + ], + "blocks": [ + { + "bbox": [ + 391, + 441, + 481, + 464 + ], + "lines": [ + { + "bbox": [ + 391, + 441, + 481, + 464 + ], + "spans": [ + { + "bbox": [ + 391, + 441, + 481, + 464 + ], + "type": "text", + "content": "(c) In-Context Example Num" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 398, + 473, + 476, + 531 + ], + "lines": [ + { + "bbox": [ + 398, + 473, + 476, + 531 + ], + "spans": [ + { + "bbox": [ + 398, + 473, + 476, + 531 + ], + "type": "table", + "html": "
MethodAccuracy
Del 150.65
Original51.05
Add 151.35
", + "image_path": "bebc2669a832fa6f24c63efb460af383106230a1eed24bcad68d5b818ea4993b.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 131, + 551, + 253, + 563 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 551, + 253, + 563 + ], + "spans": [ + { + "bbox": [ + 131, + 551, + 253, + 563 + ], + "type": "text", + "content": "4.5 Readability Study" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 570, + 481, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 481, + 617 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 481, + 617 + ], + "type": "text", + "content": "This study investigates how recursive coding impacts code readability and comprehension for human programmers. Our goal is to move from complex code structures to more modular, clear, and understandable code. To evaluate this, we focus on code quality in terms of its understandability." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "From the COVR dataset, we randomly select one example from each five different question types where recursive patterns occur. We present the main function of both their recursive and non-recursive code versions to 15 participants and ask them to grasp the high-level intent of the code as quickly as" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 203, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 203, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 203, + 101 + ], + "type": "text", + "content": "Ge et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "code", + "bbox": [ + 136, + 114, + 298, + 238 + ], + "blocks": [ + { + "bbox": [ + 136, + 114, + 298, + 238 + ], + "lines": [ + { + "bbox": [ + 136, + 114, + 298, + 238 + ], + "spans": [ + { + "bbox": [ + 136, + 114, + 298, + 238 + ], + "type": "text", + "content": "def execute(image) -> str: image_batch = ImagePatch(image) is.black_skateboard= image_batchrecursive_query(\"Are there any black skateboards?\") is_motorcycle " + }, + { + "bbox": [ + 136, + 114, + 298, + 238 + ], + "type": "inline_equation", + "content": "=" + }, + { + "bbox": [ + 136, + 114, + 298, + 238 + ], + "type": "text", + "content": " image_batchrecursive_query(\"Arethereanymotorcycles?\") return bool to yesno(is.black Skateboard or is_motorcycle) \ndef execute(image) -> str: image_batch " + }, + { + "bbox": [ + 136, + 114, + 298, + 238 + ], + "type": "inline_equation", + "content": "=" + }, + { + "bbox": [ + 136, + 114, + 298, + 238 + ], + "type": "text", + "content": " ImagePatch(image) skateboard_patches " + }, + { + "bbox": [ + 136, + 114, + 298, + 238 + ], + "type": "inline_equation", + "content": "=" + }, + { + "bbox": [ + 136, + 114, + 298, + 238 + ], + "type": "text", + "content": " image_batch.find(\"skateboard\") for skateboard_batch in skateboard_patches: If skateboard_batchverified_property(\"skateboard\", \"black\"): return \"yes\" return \"no\"" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "code_body" + }, + { + "bbox": [ + 133, + 243, + 299, + 252 + ], + "lines": [ + { + "bbox": [ + 133, + 243, + 299, + 252 + ], + "spans": [ + { + "bbox": [ + 133, + 243, + 299, + 252 + ], + "type": "text", + "content": "Return Type Confusion: Expect Bool but Actually Str." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "code_caption" + } + ], + "index": 2, + "sub_type": "code", + "guess_lang": "python" + }, + { + "type": "image", + "bbox": [ + 309, + 124, + 478, + 248 + ], + "blocks": [ + { + "bbox": [ + 309, + 124, + 478, + 248 + ], + "lines": [ + { + "bbox": [ + 309, + 124, + 478, + 248 + ], + "spans": [ + { + "bbox": [ + 309, + 124, + 478, + 248 + ], + "type": "image", + "image_path": "5d1427f2a49da9b18a5f78c59270b89ceb435daaf18ee769d6a2d32c42e9e657.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 261, + 301, + 305 + ], + "lines": [ + { + "bbox": [ + 130, + 261, + 301, + 305 + ], + "spans": [ + { + "bbox": [ + 130, + 261, + 301, + 305 + ], + "type": "text", + "content": "Fig.8: Failure example of dynamic type assignment due to unspecified type, The model confuses the type and return a str instead of a bool." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 304, + 261, + 482, + 305 + ], + "lines": [ + { + "bbox": [ + 304, + 261, + 482, + 305 + ], + "spans": [ + { + "bbox": [ + 304, + 261, + 482, + 305 + ], + "type": "text", + "content": "Fig.9: Comparison of average understanding time for recursive and non-recursive codes. RVP demonstrates better readability and is easy to understand." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 329, + 482, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 329, + 482, + 449 + ], + "spans": [ + { + "bbox": [ + 130, + 329, + 482, + 449 + ], + "type": "text", + "content": "possible. For the non-recursive case, the main function is just the single code snippet generated by the LLM. For the recursive case, the main function consists of the code generated by the first LLM call (the high level code) but not the functions generated by the recursive queries. This helps us measure how fast users can understand the code's intention by looking at the main function, and reflects the self-explanatory nature of the code. Figure 9 shows the results. Participants consistently understand the recursive code faster than the non-recursive versions. This quicker comprehension can be attributed to the recursive code's structured and clear approach. Non-recursive code, while correct, sometimes lacks immediate clarity in conveying its primary goal." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 449, + 482, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 449, + 482, + 496 + ], + "spans": [ + { + "bbox": [ + 130, + 449, + 482, + 496 + ], + "type": "text", + "content": "We also calculate the average line length for all main and sub functions. For RVP, the average length of all functions is 8.47. For non-recursive VP, the average length of the functions is 15.4. This suggests that RVP decreases the average length the code pieces and enhances modularity." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 497, + 482, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 497, + 482, + 545 + ], + "spans": [ + { + "bbox": [ + 130, + 497, + 482, + 545 + ], + "type": "text", + "content": "Our findings suggest that recursive coding not only simplifies complex code but also enhances its accessibility and ease of understanding. Its orderly structure and clear logic accelerate comprehension, leading to more efficient and user-friendly programming. We provide more details of this study in the supp." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 562, + 257, + 574 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 562, + 257, + 574 + ], + "spans": [ + { + "bbox": [ + 132, + 562, + 257, + 574 + ], + "type": "text", + "content": "4.6 Recursive Analysis" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 581, + 482, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 581, + 482, + 653 + ], + "spans": [ + { + "bbox": [ + 130, + 581, + 482, + 653 + ], + "type": "text", + "content": "Question Type. We study which types of questions tend to use recursive and found that recursive patterns appeared in five question types within both the COVR test set and the GQA test-dev set, as shown in Table 3. The 'Verify Property' and 'Query' types in GQA, and 'Compare' and 'Multi-Reference' in COVR, were most common. This suggests recursive methods tend to occur in tasks that require comparative and referential analysis." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 653, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 653, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 653, + 481, + 666 + ], + "type": "text", + "content": "Dynamic Type. We study which types of dynamic return types occurs and" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 321, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 321, + 91, + 447, + 102 + ], + "type": "text", + "content": "Recursive Visual Programming" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 114, + 482, + 182 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 114, + 482, + 182 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 482, + 182 + ], + "type": "text", + "content": "Table 3: Recursive analysis. Table (a) provides the distribution of recursive patterns by question type in GQA and COVR. We have found that recursive patterns occur commonly in question types that require reasoning. Table (b) studies different recursive query return types and their frequency in GQA and COVR. The model generalizes to unseen types. Table (c) compares accuracy across question types. Recursive methods outperform non-recursive particularly in 'Choose' and 'Query' questions." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 140, + 189, + 266, + 201 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 140, + 189, + 266, + 201 + ], + "spans": [ + { + "bbox": [ + 140, + 189, + 266, + 201 + ], + "type": "text", + "content": "(a) Dist. of Recursive Patterns" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 271, + 189, + 378, + 201 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 271, + 189, + 378, + 201 + ], + "spans": [ + { + "bbox": [ + 271, + 189, + 378, + 201 + ], + "type": "text", + "content": "(b) Diverse Return Types" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 389, + 189, + 468, + 201 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 389, + 189, + 468, + 201 + ], + "spans": [ + { + "bbox": [ + 389, + 189, + 468, + 201 + ], + "type": "text", + "content": "(c) Type Accuracy" + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 142, + 209, + 264, + 308 + ], + "blocks": [ + { + "bbox": [ + 142, + 209, + 264, + 308 + ], + "lines": [ + { + "bbox": [ + 142, + 209, + 264, + 308 + ], + "spans": [ + { + "bbox": [ + 142, + 209, + 264, + 308 + ], + "type": "table", + "html": "
COVR/GQAQuestion Types
COVRGQA
Compare45715
Mult. Ref282-
Quant. Attr12-
Spec Attr357-
Choose125
Verify-68
Query-167
Logical-34
", + "image_path": "c3ec4fc02156296064b0cc19376ec9964bd0c0a61d3507bb99e8ff47e76d118c.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 270, + 210, + 377, + 287 + ], + "blocks": [ + { + "bbox": [ + 270, + 210, + 377, + 287 + ], + "lines": [ + { + "bbox": [ + 270, + 210, + 377, + 287 + ], + "spans": [ + { + "bbox": [ + 270, + 210, + 377, + 287 + ], + "type": "table", + "html": "
Data TypeGQACOVR
Str3191828
Bool521592
List[str]16-
ImagePatch30-
List[ImgPatch]15
Float6-
Int-914
", + "image_path": "c64a9c635a2dca6e0741c5521f4b7e3ea2d55902cf6856c7c7b06de6f69fa8b1.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 380, + 210, + 477, + 285 + ], + "blocks": [ + { + "bbox": [ + 380, + 210, + 477, + 285 + ], + "lines": [ + { + "bbox": [ + 380, + 210, + 477, + 285 + ], + "spans": [ + { + "bbox": [ + 380, + 210, + 477, + 285 + ], + "type": "table", + "html": "
QuestionAccuracy (%)
Non-RecRec
Verify52.9464.71
Query56.8971.86
Logical55.8870.59
Compare66.6753.33
Choose52.0084.00
", + "image_path": "ddc4faaf47e79e153aacf695e7aeb7a74d5690893d1cf14b51e72d0894afff2a.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 316, + 481, + 363 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 316, + 481, + 363 + ], + "spans": [ + { + "bbox": [ + 130, + 316, + 481, + 363 + ], + "type": "text", + "content": "their frequently in GQA and COVR. In Table 3, we found that RVP generalizes to new types like List[str] and List[ImagePatch], which are not included in the in-context examples. This flexibility and ability to generalize beyond in-context examples highlight the inherent adaptability of recursive methods." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 364, + 481, + 423 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 364, + 481, + 423 + ], + "spans": [ + { + "bbox": [ + 130, + 364, + 481, + 423 + ], + "type": "text", + "content": "Type Accuracy. We evaluate the accuracy of recursive methods across different question types. An analysis of the GQA test-dev set is presented in Table 3. It reveals that recursive methods significantly improve accuracy in 'Query', 'Logical', and 'Choose' questions. This underscores their suitability for complex reasoning and decision-making tasks." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 424, + 481, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 424, + 481, + 483 + ], + "spans": [ + { + "bbox": [ + 130, + 424, + 481, + 483 + ], + "type": "text", + "content": "Error Source. We sample 50 incorrect recursive examples from GQA2000val and manually check the error source: " + }, + { + "bbox": [ + 130, + 424, + 481, + 483 + ], + "type": "inline_equation", + "content": "22\\%" + }, + { + "bbox": [ + 130, + 424, + 481, + 483 + ], + "type": "text", + "content": " are caused by logic error inside subfunctions; " + }, + { + "bbox": [ + 130, + 424, + 481, + 483 + ], + "type": "inline_equation", + "content": "18\\%" + }, + { + "bbox": [ + 130, + 424, + 481, + 483 + ], + "type": "text", + "content": " are caused by logic error between functions; " + }, + { + "bbox": [ + 130, + 424, + 481, + 483 + ], + "type": "inline_equation", + "content": "52\\%" + }, + { + "bbox": [ + 130, + 424, + 481, + 483 + ], + "type": "text", + "content": " are caused by API model inaccuracy (i.e. detection model error); and " + }, + { + "bbox": [ + 130, + 424, + 481, + 483 + ], + "type": "inline_equation", + "content": "4\\%" + }, + { + "bbox": [ + 130, + 424, + 481, + 483 + ], + "type": "text", + "content": " are caused by other problems like the groundtruth answer unclear." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 483, + 481, + 508 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 483, + 481, + 508 + ], + "spans": [ + { + "bbox": [ + 130, + 483, + 481, + 508 + ], + "type": "text", + "content": "Cost And Runtime. For the questions which require recursive query, RVP takes about 2.1 times (runtime) and about 3 times (cost) compared to ViperGPT." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 131, + 521, + 313, + 533 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 521, + 313, + 533 + ], + "spans": [ + { + "bbox": [ + 131, + 521, + 313, + 533 + ], + "type": "text", + "content": "5 Conclusion and Limitation" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 534, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 534, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 534, + 482, + 666 + ], + "type": "text", + "content": "In this work, we present RVP, a visual programming method that adopts recursive coding to generate concise, elegant and easy-to-understand code, handling details and intricate logic more accurately compared with traditional visual programming methods. We conduct extensive experiments across various benchmarks and provide a comprehensive analysis of RVP. Our work is the first to explore recursive programming in the visual domain and dynamic-type code generation for visual reasoning; we believe the recursive programming concept would have a potential greater impact on other domains as well. While RVP is effective for complicated visual questions, most questions in current VQA benchmarks tend to be short and non-recursive. We believe that the benefit of our method could be demonstrated better with a new benchmark or a more complex split." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 203, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 203, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 203, + 101 + ], + "type": "text", + "content": "Ge et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 139, + 481, + 174 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 139, + 481, + 174 + ], + "spans": [ + { + "bbox": [ + 130, + 139, + 481, + 174 + ], + "type": "text", + "content": "We would like to thank Ben Bogin for helping us test on the private COVR test set, and Guangyuan Jiang for helping us proof read the paper and providing valuable feedback." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 193, + 197, + 205 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 193, + 197, + 205 + ], + "spans": [ + { + "bbox": [ + 133, + 193, + 197, + 205 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 217, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 138, + 217, + 481, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 217, + 481, + 251 + ], + "spans": [ + { + "bbox": [ + 138, + 217, + 481, + 251 + ], + "type": "text", + "content": "1. Andreas, J., Rohrbach, M., Darrell, T., Klein, D.: Neural module networks. 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR) pp. 39-48 (2015), https://api-semanticscholar.org/CorpusID:5276660" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 251, + 481, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 251, + 481, + 282 + ], + "spans": [ + { + "bbox": [ + 138, + 251, + 481, + 282 + ], + "type": "text", + "content": "2. Andreas, J., Rohrbach, M., Darrell, T., Klein, D.: Learning to compose neural networks for question answering. ArXiv abs/1601.01705 (2016), https://apisemantic scholar.org/CorpusID:3130692" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 283, + 481, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 283, + 481, + 315 + ], + "spans": [ + { + "bbox": [ + 138, + 283, + 481, + 315 + ], + "type": "text", + "content": "3. Antol, S., Agrawal, A., Lu, J., Mitchell, M., Batra, D., Zitnick, C.L., Parikh, D.: Vqa: Visual question answering. In: Proceedings of the IEEE international conference on computer vision. pp. 2425-2433 (2015)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 316, + 481, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 316, + 481, + 348 + ], + "spans": [ + { + "bbox": [ + 138, + 316, + 481, + 348 + ], + "type": "text", + "content": "4. Besta, M., Blach, N., Kubicek, A., Gerstenberger, R., Podstawski, M., Gianinazzi, L., Gajda, J., Lehmann, T., Niewiadomski, H., Nczyk, P., Hoefler, T.: Graph of thoughts: Solving elaborate problems with large language models (2024)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 349, + 481, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 349, + 481, + 381 + ], + "spans": [ + { + "bbox": [ + 138, + 349, + 481, + 381 + ], + "type": "text", + "content": "5. Bogin, B., Gupta, S., Gardner, M., Berant, J.: Covr: A test-bed for visually grounded compositional generalization with real images. ArXiv abs/2109.10613 (2021), https://api.sementicscholar.org/CorpusID:237592834" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 381, + 481, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 381, + 481, + 479 + ], + "spans": [ + { + "bbox": [ + 138, + 381, + 481, + 479 + ], + "type": "text", + "content": "6. Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J.D., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., Agarwal, S., Herbert-Voss, A., Krueger, G., Henighan, T., Child, R., Ramesh, A., Ziegler, D., Wu, J., Winter, C., Hesse, C., Chen, M., Sigler, E., Litwin, M., Gray, S., Chess, B., Clark, J., Berner, C., McCandlish, S., Radford, A., Sutskever, I., Amodei, D.: Language models are few-shot learners. In: Larochelle, H., Ranzato, M., Hadsell, R., Balcan, M., Lin, H. (eds.) Advances in Neural Information Processing Systems. vol. 33, pp. 1877-1901. Curran Associates, Inc. (2020), https://proceedings.neurips.cc/paper/2020/file/1457c0d6bfb4967418bf8ac142f64a-Paper.pdf" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 479, + 481, + 512 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 479, + 481, + 512 + ], + "spans": [ + { + "bbox": [ + 138, + 479, + 481, + 512 + ], + "type": "text", + "content": "7. Buch, S., Eyzaguirre, C., Gaidon, A., Wu, J., Fei-Fei, L., Niebles, J.C.: Revisiting the \"Video\" in Video-Language Understanding. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2022)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 512, + 481, + 631 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 512, + 481, + 631 + ], + "spans": [ + { + "bbox": [ + 138, + 512, + 481, + 631 + ], + "type": "text", + "content": "8. Chen, M., Tworek, J., Jun, H., Yuan, Q., Ponde, H., Kaplan, J., Edwards, H., Burda, Y., Joseph, N., Brockman, G., Ray, A., Puri, R., Krueger, G., Petrov, M., Khlaaf, H., Sastry, G., Mishkin, P., Chan, B., Gray, S., Ryder, N., Pavlov, M., Power, A., Kaiser, L., Bavarian, M., Winter, C., Tillet, P., Such, F.P., Cummings, D.W., Plappert, M., Chantzis, F., Barnes, E., Herbert-Voss, A., Guss, W.H., Nichol, A., Babuschkin, I., Balaji, S.A., Jain, S., Carr, A., Leike, J., Achiam, J., Misra, V., Morikawa, E., Radford, A., Knight, M.M., Brundage, M., Murati, M., Mayer, K., Welinder, P., McGrew, B., Amodei, D., McCandlish, S., Sutskever, I., Zaremba, W.: Evaluating large language models trained on code. ArXiv abs/2107.03374 (2021), https://api-semanticscholar.org/CorpusID: 235755472" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 632, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 632, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 632, + 481, + 665 + ], + "type": "text", + "content": "9. Chen, W., Ma, X., Wang, X., Cohen, W.W.: Program of thoughts prompting: Disentangling computation from reasoning for numerical reasoning tasks. Transactions on Machine Learning Research (2023)" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 320, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 320, + 91, + 447, + 102 + ], + "type": "text", + "content": "Recursive Visual Programming" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 666 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 133, + 116, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 480, + 138 + ], + "type": "text", + "content": "10. Chen, X., Lin, M., Scharli, N., Zhou, D.: Teaching large language models to self-debug. arXiv preprint arXiv:2304.05128 (2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 139, + 481, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 139, + 481, + 182 + ], + "spans": [ + { + "bbox": [ + 133, + 139, + 481, + 182 + ], + "type": "text", + "content": "11. Cheng, Z., Xie, T., Shi, P., Li, C., Nadkarni, R., Hu, Y., Xiong, C., Radev, D.R., Ostendorf, M., Zettlemoyer, L., Smith, N.A., Yu, T.: Binding language models in symbolic languages. ArXiv abs/2210.02875 (2022), https://api-semanticscholar.org/CorpusID:252734772" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 182, + 480, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 182, + 480, + 205 + ], + "spans": [ + { + "bbox": [ + 132, + 182, + 480, + 205 + ], + "type": "text", + "content": "12. Cho, J., Zala, A., Bansal, M.: Visual programming for text-to-image generation and evaluation. NeurIPS (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 205, + 480, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 205, + 480, + 226 + ], + "spans": [ + { + "bbox": [ + 132, + 205, + 480, + 226 + ], + "type": "text", + "content": "13. Gao, L., Madaan, A., Zhou, S., Alon, U., Liu, P., Yang, Y., Callan, J., Neubig, G.: Pal: Program-aided language models. arXiv preprint arXiv:2211.10435 (2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 227, + 480, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 227, + 480, + 270 + ], + "spans": [ + { + "bbox": [ + 132, + 227, + 480, + 270 + ], + "type": "text", + "content": "14. Gupta, T., Kembhavi, A.: Visual programming: Compositional visual reasoning without training. 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) pp. 14953-14962 (2022), https://api-semanticscholar.org/CorpusID:253734854" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 271, + 480, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 271, + 480, + 315 + ], + "spans": [ + { + "bbox": [ + 132, + 271, + 480, + 315 + ], + "type": "text", + "content": "15. Hu, R., Andreas, J., Rohrbach, M., Darrell, T., Saenko, K.: Learning to reason: End-to-end module networks for visual question answering. 2017 IEEE International Conference on Computer Vision (ICCV) pp. 804-813 (2017), https://api_semanticscholar.org/CorpusID:18682" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 316, + 480, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 316, + 480, + 358 + ], + "spans": [ + { + "bbox": [ + 132, + 316, + 480, + 358 + ], + "type": "text", + "content": "16. Hudson, D.A., Manning, C.D.: Gqa: A new dataset for real-world visual reasoning and compositional question answering. 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) pp. 6693-6702 (2019), https://apisemantic scholar.org/CorpusID:152282269" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 358, + 480, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 358, + 480, + 402 + ], + "spans": [ + { + "bbox": [ + 132, + 358, + 480, + 402 + ], + "type": "text", + "content": "17. Hudson, D.A., Manning, C.D.: Gqa: A new dataset for real-world visual reasoning and compositional question answering. 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) pp. 6693-6702 (2019), https://apisemantic scholar.org/CorpusID:152282269" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 403, + 480, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 403, + 480, + 435 + ], + "spans": [ + { + "bbox": [ + 132, + 403, + 480, + 435 + ], + "type": "text", + "content": "18. Hudson, D.A., Manning, C.D.: Learning by abstraction: The neural state machine. In: Neural Information Processing Systems (2019), https://api-semanticscholar.org/CorpusID:195847902" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 435, + 480, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 435, + 480, + 468 + ], + "spans": [ + { + "bbox": [ + 132, + 435, + 480, + 468 + ], + "type": "text", + "content": "19. Kim, W., Son, B., Kim, I.: Vilt: Vision-and-language transformer without convolution or region supervision. In: International Conference on Machine Learning (2021), https://api-semanticscholar.org/CorpusID:231839613" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 468, + 480, + 501 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 468, + 480, + 501 + ], + "spans": [ + { + "bbox": [ + 132, + 468, + 480, + 501 + ], + "type": "text", + "content": "20. Li, L.H., Yatskar, M., Yin, D., Hsieh, C.J., Chang, K.W.: Visualbert: A simple and performant baseline for vision and language. ArXiv abs/1908.03557 (2019), https://api_semanticscholar.org/CorpusID:199528533" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 501, + 480, + 534 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 501, + 480, + 534 + ], + "spans": [ + { + "bbox": [ + 132, + 501, + 480, + 534 + ], + "type": "text", + "content": "21. Liang, J., Huang, W., Xia, F., Xu, P., Hausman, K., Ichter, B., Florence, P., Zeng, A.: Code as policies: Language model programs for embodied control. In: arXiv preprint arXiv:2209.07753 (2022)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 534, + 480, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 534, + 480, + 566 + ], + "spans": [ + { + "bbox": [ + 132, + 534, + 480, + 566 + ], + "type": "text", + "content": "22. Liu, F., Emerson, G.E.T., Collier, N.: Visual spatial reasoning. Transactions of the Association for Computational Linguistics 11, 635-651 (2022), https://apisemantic scholar.org/CorpusID:248496506" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 567, + 480, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 567, + 480, + 600 + ], + "spans": [ + { + "bbox": [ + 132, + 567, + 480, + 600 + ], + "type": "text", + "content": "23. Lu, P., Peng, B., Cheng, H., Galley, M., Chang, K.W., Wu, Y.N., Zhu, S.C., Gao, J.: Chameleon: Plug-and-play compositional reasoning with large language models. arXiv preprint arXiv:2304.09842 (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 601, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 601, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 601, + 480, + 632 + ], + "type": "text", + "content": "24. Ma, Y.J., Liang, W., Wang, G., Huang, D.A., Bastani, O., Jayaraman, D., Zhu, Y., Fan, L., Anandkumar, A.: Eureka: Human-level reward design via coding large language models. arXiv preprint arXiv: Arxiv-2310.12931 (2023)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 632, + 480, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 480, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 480, + 666 + ], + "type": "text", + "content": "25. Madaan, A., Zhou, S., Alon, U., Yang, Y., Neubig, G.: Language models of code are few-shot commonsense learners. In: Goldberg, Y., Kozareva, Z., Zhang, Y. (eds.) Proceedings of the 2022 Conference on Empirical Methods in Natural Language" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 202, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 202, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 202, + 100 + ], + "type": "text", + "content": "Ge et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 149, + 116, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 116, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 149, + 116, + 480, + 149 + ], + "type": "text", + "content": "Processing. pp. 1384-1403. Association for Computational Linguistics, Abu Dhabi, United Arab Emirates (Dec 2022). https://doi.org/10.18653/v1/2022.emnlp-main.90, https://aclanthology.org/2022.emnlp-main.90" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 151, + 481, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 151, + 481, + 195 + ], + "spans": [ + { + "bbox": [ + 133, + 151, + 481, + 195 + ], + "type": "text", + "content": "26. Marino, K., Rastegari, M., Farhadi, A., Mottaghi, R.: Ok-vqa: A visual question answering benchmark requiring external knowledge. 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) pp. 3190-3199 (2019), https://api-semanticscholar.org/CorpusID:173991173" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 196, + 480, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 196, + 480, + 217 + ], + "spans": [ + { + "bbox": [ + 132, + 196, + 480, + 217 + ], + "type": "text", + "content": "27. OpenAI: Gpt-4 technical report. ArXiv abs/2303.08774 (2023), https://apisemantic scholar.org/CorpusID:257532815" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 219, + 480, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 219, + 480, + 251 + ], + "spans": [ + { + "bbox": [ + 132, + 219, + 480, + 251 + ], + "type": "text", + "content": "28. Press, O., Zhang, M., Min, S., Schmidt, L., Smith, N.A., Lewis, M.: Measuring and narrowing the compositionality gap in language models. ArXiv abs/2210.03350 (2022), https://api-semanticscholar.org/CorpusID:252762102" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 252, + 480, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 252, + 480, + 296 + ], + "spans": [ + { + "bbox": [ + 132, + 252, + 480, + 296 + ], + "type": "text", + "content": "29. Subramanian, S., Bogin, B., Gupta, N., Wolfson, T., Singh, S., Berant, J., Gardner, M.: Obtaining faithful interpretations from compositional neural networks. In: Annual Meeting of the Association for Computational Linguistics (2020), https://apisemantic scholar.org/CorpusID:218487535" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 297, + 481, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 297, + 481, + 373 + ], + "spans": [ + { + "bbox": [ + 132, + 297, + 481, + 373 + ], + "type": "text", + "content": "30. Subramanian, S., Narasimhan, M., Khangaonkar, K., Yang, K., Nagrani, A., Schmid, C., Zeng, A., Darrell, T., Klein, D.: Modular visual question answering via code generation. In: Rogers, A., Boyd-Graber, J., Okazaki, N. (eds.) Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers). pp. 747-761. Association for Computational Linguistics, Toronto, Canada (Jul 2023). https://doi.org/10.18653/v1/2023.acl-short.65, https://aclanthology.org/2023.acl-short.65" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 374, + 480, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 374, + 480, + 407 + ], + "spans": [ + { + "bbox": [ + 132, + 374, + 480, + 407 + ], + "type": "text", + "content": "31. Sur'is, D., Menon, S., Vondrick, C.: Vipergpt: Visual inference via python execution for reasoning. ArXiv abs/2303.08128 (2023), https://api_semanticscholar.org/CorpusID:257505358" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 408, + 480, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 408, + 480, + 441 + ], + "spans": [ + { + "bbox": [ + 132, + 408, + 480, + 441 + ], + "type": "text", + "content": "32. Suzgun, M., Scales, N., Scharli, N., Gehrmann, S., Tay, Y., Chung, H.W., Chowdhery, A., Le, Q.V., Chi, E.H., Zhou, D., et al.: Challenging big-bench tasks and whether chain-of-thought can solve them. arXiv preprint arXiv:2210.09261 (2022)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 441, + 480, + 495 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 441, + 480, + 495 + ], + "spans": [ + { + "bbox": [ + 132, + 441, + 480, + 495 + ], + "type": "text", + "content": "33. Touvron, H., Lavril, T., Izacard, G., Martinet, X., Lachaux, M.A., Lacroix, T., Rozière, B., Goyal, N., Hambro, E., Azhar, F., Rodriguez, A., Joulin, A., Grave, E., Lample, G.: Llama: Open and efficient foundation language models. ArXiv abs/2302.13971 (2023), https://apisemanticscholar.org/CorpusID: 257219404" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 497, + 480, + 519 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 497, + 480, + 519 + ], + "spans": [ + { + "bbox": [ + 132, + 497, + 480, + 519 + ], + "type": "text", + "content": "34. Wang, X., Li, S., Ji, H.: Code4struct: Code generation for few-shot structured prediction from natural language. arXiv preprint arXiv:2210.12810 (2022)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 520, + 480, + 563 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 520, + 480, + 563 + ], + "spans": [ + { + "bbox": [ + 132, + 520, + 480, + 563 + ], + "type": "text", + "content": "35. Wang, Z., Li, M., Xu, R., Zhou, L., Lei, J., Lin, X., Wang, S., Yang, Z., Zhu, C., Hoiem, D., et al.: Language models with image descriptors are strong few-shot video-language learners. Advances in Neural Information Processing Systems 35, 8483-8497 (2022)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 565, + 480, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 565, + 480, + 597 + ], + "spans": [ + { + "bbox": [ + 132, + 565, + 480, + 597 + ], + "type": "text", + "content": "36. Xiao, J., Shang, X., Yao, A., Chua, T.S.: Next-qa: Next phase of question-answering to explaining temporal actions. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 9777–9786 (June 2021)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 599, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 599, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 599, + 480, + 632 + ], + "type": "text", + "content": "37. Yang, K., Klein, D., Peng, N., Tian, Y.: Doc: Improving long story coherence with detailed outline control. In: Annual Meeting of the Association for Computational Linguistics (2023), https://api-semanticscholar.org/CorpusID:254877751" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "type": "text", + "content": "38. Yao, S., Yu, D., Zhao, J., Shafran, I., Griffiths, T.L., Cao, Y., Narasimhan, K.: Tree of thoughts: Deliberate problem solving with large language models. arXiv preprint arXiv:2305.10601 (2023)" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 320, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 320, + 91, + 447, + 102 + ], + "type": "text", + "content": "Recursive Visual Programming" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 293 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 481, + 149 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 481, + 149 + ], + "type": "text", + "content": "39. Ye, Q., Xu, G., Yan, M., Xu, H., Qian, Q., Zhang, J., Huang, F.: Hitea: Hierarchical temporal-aware video-language pre-training. ArXiv abs/2212.14546 (2022), https://api-semanticscholar.org/CorpusID:255340506" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 150, + 481, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 150, + 481, + 205 + ], + "spans": [ + { + "bbox": [ + 132, + 150, + 481, + 205 + ], + "type": "text", + "content": "40. Yu, W., Gileadi, N., Fu, C., Kirmani, S., Lee, K.H., Gonzalez Arenas, M., Lewis Chiang, H.T., Erez, T., Hasenclever, L., Humplik, J., Ichter, B., Xiao, T., Xu, P., Zeng, A., Zhang, T., Heess, N., Sadigh, D., Tan, J., Tassa, Y., Xia, F.: Language to rewards for robotic skill synthesis. Arxiv preprint arXiv:2306.08647 (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 205, + 481, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 205, + 481, + 248 + ], + "spans": [ + { + "bbox": [ + 132, + 205, + 481, + 248 + ], + "type": "text", + "content": "41. Zhang, P., Li, X., Hu, X., Yang, J., Zhang, L., Wang, L., Choi, Y., Gao, J.: Vinvl: Revisiting visual representations in vision-language models. 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) pp. 5575-5584 (2021), https://api_semanticscholar.org/CorpusID:235692795" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 249, + 481, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 249, + 481, + 293 + ], + "spans": [ + { + "bbox": [ + 132, + 249, + 481, + 293 + ], + "type": "text", + "content": "42. Zhou, D., Scharli, N., Hou, L., Wei, J., Scales, N., Wang, X., Schuurmans, D., Bousquet, O., Le, Q., hsin Chi, E.H.: Least-to-most prompting enables complex reasoning in large language models. ArXiv abs/2205.10625 (2022), https://api_semanticscholar.org/CorpusID:248986239" + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 203, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 203, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 203, + 101 + ], + "type": "text", + "content": "Ge et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Ref-AVS_ Refer and Segment Objects in Audio-Visual Scenes/a65e28a6-a79d-4af3-846c-76b29927236e_content_list.json b/2024/Ref-AVS_ Refer and Segment Objects in Audio-Visual Scenes/a65e28a6-a79d-4af3-846c-76b29927236e_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..ce4faeac7f0dfa5048a3605f8a3ef58e9c2bdca3 --- /dev/null +++ b/2024/Ref-AVS_ Refer and Segment Objects in Audio-Visual Scenes/a65e28a6-a79d-4af3-846c-76b29927236e_content_list.json @@ -0,0 +1,1773 @@ +[ + { + "type": "text", + "text": "Ref-AVS: Refer and Segment Objects in Audio-Visual Scenes", + "text_level": 1, + "bbox": [ + 271, + 140, + 732, + 184 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yaoting Wang $^{1\\dagger}$ , Peiwen Sun $^{2\\dagger}$ , Dongzhan Zhou $^{3\\dagger}$ , Guangyao Li $^{1}$ , Honggang Zhang $^{2}$ , and Di Hu $^{\\boxtimes 1,4}$", + "bbox": [ + 238, + 210, + 759, + 243 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Gaoling School of Artificial Intelligence, Renmin University of China, China \nyaoting.wang@outlook.com \n{guangyaoli, dihu}@ruc.edu.cn", + "bbox": [ + 241, + 253, + 759, + 296 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ Beijing University of Posts and Telecommunications, Beijing, China {sunpeiwen,zhhg}@bupt.edu.cn", + "bbox": [ + 267, + 296, + 732, + 324 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3 Shanghai Artificial Intelligence Laboratory, Shanghai, China zhoudongzhan@pjlab.org.cn", + "bbox": [ + 294, + 324, + 707, + 352 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "4 Engineering Research Center of Next-Generation Search and Recommendation", + "bbox": [ + 233, + 352, + 766, + 364 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Traditional reference segmentation tasks have predominantly focused on silent visual scenes, neglecting the integral role of multimodal perception and interaction in human experiences. In this work, we introduce a novel task called Reference Audio-Visual Segmentation (Ref-AVS), which seeks to segment objects within the visual domain based on expressions containing multimodal cues. Such expressions are articulated in natural language forms but are enriched with multimodal cues, including audio and visual descriptions. To facilitate this research, we construct the first Ref-AVS benchmark, which provides pixel-level annotations for objects described in corresponding multimodal-cue expressions. To tackle the Ref-AVS task, we propose a new method that adequately utilizes multimodal cues to offer precise segmentation guidance. Finally, we conduct quantitative and qualitative experiments on three test subsets to compare our approach with existing methods from related tasks. The results demonstrate the effectiveness of our method, highlighting its capability to precisely segment objects using multimodal-cue expressions. Dataset is available at https://gewu-lab.github.io/Ref-AVS.", + "bbox": [ + 259, + 391, + 740, + 627 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Referring Audio-Visual Segmentation $\\cdot$ Audio-Visual Segmentation $\\cdot$ Multimodal Learning", + "bbox": [ + 259, + 638, + 740, + 667 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 705, + 374, + 722 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In the real world, visual scenes are often accompanied by diverse multimodal information1, including audio and text modalities. This additional multimodal", + "bbox": [ + 212, + 732, + 787, + 762 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "$\\dagger$ Equal contribution.", + "bbox": [ + 232, + 768, + 377, + 782 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "$\\mathbb{O}$ Corresponding author.", + "bbox": [ + 232, + 782, + 398, + 797 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "To clarify, we refer to the sound and vision in the scene as multimodal information, while the attribute description of audio and vision (e.g., louder and on the right) in the reference expression is named multimodal cues.", + "bbox": [ + 222, + 797, + 785, + 839 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/a67defb349bfe0b763649657c54f8dd938206205b45a2524bf2bbaaf0657a43a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 218, + 145, + 413, + 239 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/cd9cec3a02c62acd3090fd1bad3797227fee6a27567cfa90a67e3f50386d30e2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 225, + 241, + 413, + 275 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/ec91475b65546273e18449d9e230e94cb936a1cba736fbc33d20772d544e4613.jpg", + "image_caption": [ + "Fig. 1: Comparison of the Ref-AVS task with other related tasks. Ref-AVS challenges machines to locate objects of interest in the visual space using multimodal cues, just like humans do in the real world." + ], + "image_footnote": [], + "bbox": [ + 225, + 276, + 413, + 311 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/fb5bb9a31be2cd4ad32118cdca9b00a7deb3ce89104918e35cc46f4af2e37afc.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 418, + 146, + 586, + 239 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/7523bf8a058148df62ae39bf991715d14ea28b372cd5d773bb365adfa0109c45.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 424, + 241, + 586, + 275 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/475650cdc4a48b1e34e9ee625e8afc84651bf7755e4c0c2aef3f11bb54be7c54.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 406, + 276, + 571, + 311 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Which one should the camera focus?", + "bbox": [ + 591, + 148, + 771, + 160 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "An ideal machine needs to know always going for the player who is actually playing.", + "bbox": [ + 591, + 162, + 774, + 183 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In the view of camera.", + "bbox": [ + 591, + 193, + 702, + 203 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "How to focus the player who is actually playing.", + "bbox": [ + 591, + 210, + 753, + 229 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/bcd12c67cc482a48575faaca83c62933c2394a9d1f8649d825a00f089b3bde81.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 591, + 231, + 767, + 299 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "information offers valuable spatiotemporal and semantic cues, aiding individuals in locating objects of interest. Similarly, the ability to locate objects based on multimodal cues plays a crucial role in the functioning of machines. For example, stage cameras should be able to locate the musician playing instruments, while movie cameras should locate the superhero in front of the yelling villain. In the above scenarios, using a single modality for segmentation alone cannot achieve the desired behavior.", + "bbox": [ + 212, + 398, + 787, + 503 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To date, researchers have studied the reference segmentation problem from somewhat limited scenarios. As shown in the coordinate system in Fig. 1, there are currently three mainstream developments in segmentation methods for different modalities. Firstly, from the perspective of visual cues, Video Object Segmentation (VOS) [9,30-32,40] emerges with annotated first frame mask as the reference, guiding the segmentation of specific objects in subsequent video frames. While VOS can achieve arbitrary object segmentation in videos, it heavily relies on precise annotation of the first frame, which can be challenging and time-consuming in practice. Secondly, from the perspective of text cues, Referring Video Object Segmentation (R-VOS) [8,11,13,20] explores the segmentation ability by using attribute descriptions as guidance. As R-VOS has successfully replaced the mask annotation used in VOS with natural language, it provides a more accessible and user-friendly reference form but the capability to locate objects in more natural dynamic audio-visual scenes is still limited. In contrast, from the perspective of audio cues, Audio-Visual Segmentation (AVS) [26,37,44-47] leverages audio as a guide to segment sounding objects. This approach effectively addresses the challenge of locating objects in dynamic audio-visual scenes. However, AVS still faces limitations in its ability to focus on general objects that do not generate sound and cannot effectively locate objects of specific interest.", + "bbox": [ + 212, + 506, + 787, + 808 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In simple terms, current work falls short of enabling machines to locate objects of interest in natural dynamic audio-visual scenes. For instance, as illus", + "bbox": [ + 212, + 809, + 787, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "Y. Wang et al.", + "bbox": [ + 271, + 114, + 370, + 128 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "trated in Fig. 1, how can machines accurately locate a person who is genuinely playing a musical instrument over time? This requires machines to infer which instrument is making a sound and who is playing that instrument. Though Yan et al. [41] utilize either audio or language modality for multimodal reference segmentation, their method does not integrate multimodal information simultaneously nor support the aforementioned scenario. Therefore, introducing a task to explore the possibility of locating interested objects in natural dynamic audio-visual scenes holds significant practical potential.", + "bbox": [ + 212, + 146, + 782, + 263 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this paper, we propose a pixel-level segmentation task called Referring Audio-Visual Segmentation (Ref-AVS), which requires the network to densely predict whether each pixel corresponds to the given multimodal-cue expression, including dynamic audio-visual information. Top-left of Fig. 1 highlights the distinctions between Ref-AVS and previous tasks. The Ref-AVS task poses a greater challenge as it requires the network to accurately locate and segment objects in a more complex and stereoscopic modality space. To this end, a computation model with comprehensive multimodal understanding capabilities is necessary. To foster research advancements in this field, we introduce the Ref-AVS Bench, the pioneering benchmark that focuses on locating and segmenting objects of interest using referring multimodal-cue expressions. Considering the complexity of real-world audio-visual scenes, we collect about 4,000 audible video clips from YouTube of which more than $60\\%$ are multi-source sound scenarios. More than 20,000 reference expressions are collected and verified by experts to ensure accuracy and reliability, which adopt multimodal cues to describe objects in diverse and dynamic audio-visual scenes. Moreover, a special unseen test set is considered to evaluate the model's generalization ability in the growing demand for zero-shot scenarios.", + "bbox": [ + 212, + 267, + 784, + 536 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our contributions can be summarized as follows:", + "bbox": [ + 238, + 540, + 589, + 554 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose Ref-AVS as a challenging scene understanding task that segments objects of interest with multimodal-cue expressions and provide the corresponding Ref-AVS benchmark for performance training and validation.", + "- We design an end-to-end framework for Ref-AVS that efficiently processes the multimodal cues with a crossmodal transformer, serving as a feasible research framework for future development.", + "- We conduct extensive experiments to demonstrate the advantages of considering multimodal cues for visual segmentation, which also indicate the superiority of our approach in all subsets." + ], + "bbox": [ + 223, + 566, + 782, + 700 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 731, + 387, + 747 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.1 Referring Video Object Segmentation (R-VOS)", + "text_level": 1, + "bbox": [ + 214, + 766, + 651, + 782 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The main objective of R-VOS is to perform object segmentation in streaming videos based on given natural language expressions. Initially proposed in 2018, A2D-Sentence [13] aims to segment actors in video content based on", + "bbox": [ + 212, + 794, + 782, + 839 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "Ref-AVS", + "bbox": [ + 671, + 114, + 730, + 126 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "descriptions of their actions. Subsequently, Refer-DAVIS17 [20] and JHMDB-Sentences [11] use extra language as references to replace previous VOS settings. To facilitate large-scale R-VOS training, Seo et al. [35] develop the ReferYouTube-VOS dataset upon the YouTube-VOS-2019 dataset [40] and Henghui et al. [8] build the MeViS for complex motion expression. Overall, the primary R-VOS methods have evolved from semi-supervised VOS and incorporate attention mechanisms to enhance modalities interaction. For example, MAttNet [43] decomposes the expression information into language- and word-level attention, while URVOS [35] introduces crossmodal attention and memory attention. ReferFormer [39] and MTTR [1] utilize the crossmodal transformer to connect language expression with image regions. These datasets typically offer expressions that encompass the action or appearance attributes of the target object. However, these expressions primarily focus on visual information, which fails to meet the increasing multimodal demands in today's diverse audio-visual scenes.", + "bbox": [ + 212, + 146, + 787, + 359 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.2 Audio-Visual Segmentation (AVS)", + "text_level": 1, + "bbox": [ + 215, + 378, + 545, + 395 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The multimodal segmentation in audio-visual scenes, which is of great concern to us, has received some attention in AVS [12, 37, 46]. AVS aims to obtain finer pixel-level masks corresponding to sound-emitting objects in the visual space. In contrast, previous studies on audio-visual localization [17, 36] use heat maps for coarse localization in an unsupervised learning manner. The existing works on AVS can be broadly categorized into fusion-based methods [12, 18, 22, 24, 25, 46] and prompt-based methods [26, 28, 29, 37]. The former primarily focuses on localizing sounding objects by fusing audio and visual features, while the latter emphasizes constructing effective audio prompts for the visual foundation model. However, AVS simply segments all sound-emitting objects in the visual space, without the flexibility to combine the multimodal cues to segment specific objects of interest.", + "bbox": [ + 212, + 402, + 787, + 583 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.3 Language-aided Audio-Visual Scene Understanding", + "text_level": 1, + "bbox": [ + 215, + 604, + 684, + 619 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Currently, there is a limited number of public tasks offering datasets for audio-visual scene comprehension accompanied by language assistance. Two notable examples of such datasets are Music-AVQA [15] and AVQA [42], both encompassing audio, visual, and language information. These datasets contain rich audio-visual components and focus on annotations for textual questions and answers, emphasizing temporal (Before/After), spatial cues (Left/Right), etc.", + "bbox": [ + 212, + 628, + 785, + 718 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Researchers [4, 19, 21] acknowledge the significance of the provided question, as it guides the feature extraction process for both audio and visual signals. Consequently, they have made efforts to identify pertinent segments by evaluating the semantic similarity between the question and temporal segments. Furthermore, [21] has been dedicated to locating crucial areas by leveraging semantic similarity between questions and visual tokens. Clearly, existing explorations have significantly advanced the research on language-aided audio-visual scene understanding. Nevertheless, the works that rely on the aforementioned datasets", + "bbox": [ + 212, + 719, + 787, + 839 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "Y. Wang et al.", + "bbox": [ + 271, + 114, + 370, + 128 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "are still unable to offer pixel-level scene understanding. This encourages us to begin exploring fine-grained segmentation of dynamic audio-visual scenes in the real world by constructing appropriate datasets.", + "bbox": [ + 212, + 146, + 787, + 191 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3 Refer-AVS Dataset", + "text_level": 1, + "bbox": [ + 214, + 215, + 437, + 233 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1 Object Category", + "text_level": 1, + "bbox": [ + 214, + 247, + 401, + 263 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To ensure a diverse range of referred objects, we have carefully selected a wide variety of audible objects in 48 categories as well as a smaller selection of static, unsoundable objects in 3 categories. For the objects that can produce sound, we have chosen 20 categories of musical instruments, 8 of animals, 15 of machines, and 5 of humans. In the special case of humans, considering their diverse appearances, voices, and actions, we have employed the concept of morphology and classified them into 5 categories, based on age and gender.", + "bbox": [ + 212, + 273, + 787, + 381 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2 Video Selection", + "text_level": 1, + "bbox": [ + 214, + 402, + 390, + 417 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "During the process of video collection, we employed the techniques presented in [3,46] to ensure the alignment of audio and visual snippets with the intended semantics. All videos were sourced from YouTube under the Creative Commons license, and each video was trimmed to a duration of 10 seconds. Throughout the manual collection process, we deliberately avoided videos falling into several categories (detailed in the appendix): 1) Videos with a large number of instances of the same semantics; 2) Videos characterized by extensive editing and camera switching; 3) Non-realistic videos containing synthetic artifacts.", + "bbox": [ + 212, + 428, + 787, + 547 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To raise the alignment with real-world distributions, we carefully select videos that contribute to the diversification of scenes within our dataset. Specifically, we focus on selecting videos that involve interactions between multiple objects, such as musical instruments, people, vehicles, etc. The rich combination of categories indicates that our dataset is not limited to a narrow set of scenarios but rather encompasses a broad spectrum of real-life scenes where such objects are likely to naturally appear together. Refer to the supplementary materials for details.", + "bbox": [ + 212, + 549, + 787, + 654 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In addition to diversity, we also carefully filter the videos to ensure that the dataset includes scenes with greater complexity and a larger number of objects. Specifically, $56\\%$ of the total videos contain two or more objects, while $13\\%$ of the total videos contain three or more objects.", + "bbox": [ + 212, + 655, + 787, + 715 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3 Expressions", + "text_level": 1, + "bbox": [ + 214, + 738, + 359, + 753 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The diversity of expressions is one of the core factors in the assembly of the Ref-AVS dataset. The expressions consist of three dimensions of information, namely audio, vision, and time. The auditory dimension incorporates characteristics such as volume and rhythm while the visual dimension encompasses attributes like the appearance and spatial configuration of objects. We also leverage temporal", + "bbox": [ + 212, + 763, + 787, + 842 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "Ref-AVS", + "bbox": [ + 671, + 114, + 730, + 126 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/2b6172fe72406fe8f653881dc334d9ef979652a16e93ef30ca099e791e1f29ce.jpg", + "image_caption": [ + "Fig. 2: The illustration of our Ref-AVS benchmark. Note, for $\\bullet \\bullet \\bullet$ , the sound volume increases successively from silent to loud. Our benchmark is meticulously designed to encompass multimodal expressions from multiple dimensions. By combining various types of modality expressions, we achieve a dataset that exhibits great diversity." + ], + "image_footnote": [], + "bbox": [ + 218, + 145, + 782, + 311 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "cues to generate references with sequential hints, e.g., \"the one making sound first\" or \"the one appearing later\". By integrating auditory, visual, and temporal information, we craft a rich array of expressions that not only accurately reflect multimodal scenarios but also meet users' specific needs for precise references. Fig. 2 illustrates examples of the combination of different modalities.", + "bbox": [ + 212, + 412, + 787, + 487 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The accuracy of expressions is also a core concern. We follow three rules to generate high-quality expressions: 1) Uniqueness: The object referred to by an expression must be unique and an expression cannot refer to multiple objects simultaneously. 2) Necessity: Complex expressions can be used for reference, but each adjective in the sentence should narrow down the scope of target objects, avoiding unnecessary and redundant descriptions of the object. 3) Clarity: Certain expression templates involve subjective factors, such as \"the _ with a louder sound\". The use of such expressions should only occur when the situation is clear enough to avoid ambiguous references.", + "bbox": [ + 212, + 488, + 787, + 623 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Besides diversity and accuracy, we have graded the difficulty of expressions based on the number of cues involved, with the proportions of easy, medium, and hard samples in the overall dataset being $20\\%$ , $60\\%$ , and $20\\%$ respectively. This gradation could be a great benefit to future research like curriculum learning. Please refer to the supplementary material for more details.", + "bbox": [ + 212, + 625, + 787, + 700 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.4 Segmentation Masks", + "text_level": 1, + "bbox": [ + 214, + 722, + 433, + 738 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We divide each 10-second video into ten equal 1-second snippets and the objective of annotation is to acquire the first frame's mask for each snippet. For these sampled frames, the ground truth labels are binary masks indicating the target object, according to expressions and multimodal information.", + "bbox": [ + 212, + 750, + 787, + 809 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To obtain such a mask, initially, we manually select the pivotal frames for each 10-second video in which the target object is present. These pivotal frames", + "bbox": [ + 212, + 809, + 787, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "Y. Wang et al.", + "bbox": [ + 271, + 114, + 370, + 128 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/ccd758c3dd4e9ba974151aa7efb008c1fdc9f38dfc9405bc490b566531fec39f.jpg", + "table_caption": [ + "Table 1: Datasets comparison with related tasks." + ], + "table_footnote": [], + "table_body": "
TaskDatasetModalityVideoFrameObjectExpressionObject/VideoPub.
AVLFlickr-SoundNet [34]A+V5,0005,000---CVPR‘2018
VGG-SS [2]A+V5,1585,158---CVPR‘2021
AVSAVS [46]A+V12,35682,97213,500~-1.09~ECCV‘2022
VPO [44]A+V-34,87440,000~-1.14 (Object/Frame)CVPR‘2024
R-VOSJ-HMDB Sentences [11]T+V9289289289281.28CVPR‘2018
A2D Sentences [13]T+V3,78211,9364,8256,6561CVPR‘2018
Refer-DAVIS16 [20]T+V503,455501001ACCV‘2018
Refer-DAVIS17 [20]T+V9013,5432051,5442.27ACCV‘2018
Refer-Youtu-VOS [35]T+V3,97511,9367,45127,8991.86ECCV‘2020
MeViS [8]T+V2,006443,0008,17128,5704.28ICCV‘2023
Ref-AVSRef-AVS BenchA+T+V4,00240,0206,88820,2611.72ECCV‘2024
", + "bbox": [ + 220, + 170, + 792, + 306 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "may occur at the beginning, middle, or end of the video, depending on when the target object is most clearly visible. Subsequently, we utilize Grounding SAM [33] to segment and label the pivotal frames, which will be manually checked and corrected subsequently. This process allows us to generate masks and labels for multiple target objects within the pivotal frames. Once the masks for the pivotal frames are established, we apply a tracking algorithm [6] to track the target object across the preceding and subsequent frames and obtain the ultimate mask and label for the target object in a span of 10 frames.", + "bbox": [ + 212, + 338, + 787, + 460 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.5 Dataset Statistics", + "text_level": 1, + "bbox": [ + 215, + 488, + 408, + 502 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In addition, we compare Refer-AVS with other popular audio-visual benchmarks in the Tab. 1. The Flickr-SoundNet [34] and VGG-SS [2] benchmarks are labeled at a patch level through bounding boxes and possess around 5,000 frame-level annotations. Compared with our dataset with pixel-level annotations, these benchmarks have a significantly lower quantity of annotations. Regarding the AVS dataset [44-46], the videos in our dataset contain a higher average number of objects about 1.72 objects per video, implying that our scenarios are more challenging with multiple sound sources and multiple semantics. Within such scenarios, our Ref-AVS benchmark is particularly valuable as it demonstrates the ability to effectively focus on objects of real interest. Our dataset also has more uniform video durations and a more refined video selection process than AVS. When it comes to datasets for R-VOS tasks, we maintain a consistent advantage in the number of videos compared to other datasets [8,11,13,20,35]. Moreover, we possess a considerable volume of data encompassing a large number of objects, expressions, and complex scenes.", + "bbox": [ + 212, + 521, + 787, + 747 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Overall, our Ref-AVS dataset encompasses a substantial collection of 20,000 expressions and pixel-level annotations spread across 4,000 videos, totaling more than 11 hours. As a result, we believe that this benchmark adequately fulfills the requirements for facilitating research on the Ref-AVS task, while also presenting a significant level of challenge in this domain. We aim to continuously expand the dataset for broader community needs, like training larger foundation models.", + "bbox": [ + 212, + 750, + 787, + 840 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "Ref-AVS", + "bbox": [ + 671, + 114, + 730, + 126 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 114, + 784, + 125 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/5bea5c731abc0fb5bd3f3b4a405932e917303750bcbcbd9ae2f36e9ebb19df74.jpg", + "table_caption": [ + "Table 2: Dataset split of Ref-AVS Bench." + ], + "table_footnote": [ + "† The categories here contains \"background\"." + ], + "table_body": "
Subsetalltrainvaltesttest (seen)test (unseen)test (null)
Category†5239395239131
Video4,0022,908276818292269257
Object6,8885,3665181,004478269257
Expression20,26114,1171,3494,7702,2881,4541,028
", + "bbox": [ + 241, + 171, + 756, + 266 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "3.6 Dataset Split", + "text_level": 1, + "bbox": [ + 215, + 306, + 372, + 321 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "As shown in Tab. 2, the complete dataset is divided into three sets, i.e., a training set of 2908 videos, a validation set of 276 videos, and a test set of 818 videos. The videos in the test set and their corresponding annotations undergo a meticulous review and re-annotation process conducted by experienced workers. In order to thoroughly evaluate the models' performance in the Ref-AVS task, the test set is further divided into three distinct subsets, each serving a specific purpose.", + "bbox": [ + 212, + 329, + 785, + 419 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Seen: The first test subset, referred to as the \"seen test set\", comprises object categories that have already appeared in the training set. The purpose of establishing this subset is to evaluate the model's fundamental performance and assess how well it generalizes to familiar object categories.", + "bbox": [ + 212, + 421, + 785, + 479 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Unseen: To address the growing demand for the generalization capabilities of various models in an open-world scenario, an additional test set was created specifically to assess the model's ability to generalize to unseen audio-visual scenes. This \"unseen test set\" consists of object categories that did not appear in the training set, although their super-categories (e.g., animal, vehicle) may have been present in the training set. The purpose of this test set is to evaluate the model's capacity to generalize to novel object categories while leveraging its understanding of broader super-categories.", + "bbox": [ + 212, + 481, + 785, + 601 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Null: A \"null reference problem\" arises when an expression refers to an object that either does not exist or is not visible in the given context. A model that accurately understands expression guidance should not segment any object in scenarios of a null reference [38]. Based on this consideration, we have developed a \"null subset\" to test the model's robustness against such challenges. This subset comprises object categories that were present during training but are paired with expressions that do not correlate, ensuring all objects in the video frames are irrelevant to the given reference expression. Therefore, the true masks for this subset are empty, and the model should refrain from segmenting any object.", + "bbox": [ + 212, + 601, + 785, + 738 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4 Expression Enhancing with Multimodal Cues", + "text_level": 1, + "bbox": [ + 215, + 758, + 689, + 776 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.1 Overall Architecture", + "text_level": 1, + "bbox": [ + 215, + 787, + 433, + 801 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Ref-AVS is designed to locate objects of interest in dynamic audio-visual scenes by utilizing multimodal cues. To solve this challenging task, we propose the", + "bbox": [ + 212, + 809, + 785, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "Y. Wang et al.", + "bbox": [ + 271, + 114, + 370, + 128 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/cffc8aac645895b33be4e35ec7dc2eb120bcc870c8b35fe8892544ca4d6deb8e.jpg", + "image_caption": [ + "Fig. 3: The overall architecture of our method EEMC. We utilize cached memory to preserve temporal information, enabling the model to capture temporal mutations. Modality encoding provides a more clear context for the multimodal cue features $Q_{m}$ , using the modality identify tokens [aud] and [vis]. Lastly, we achieve efficient prompting of the visual foundation model by employing cross-attention, where mask queries $q$ serve as the target and multimodal cue features $Q_{m}$ act as the source." + ], + "image_footnote": [], + "bbox": [ + 220, + 145, + 782, + 397 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Expression Enhancing with Multimodal Cues (EEMC) method, which focuses on integrating multimodal information in the dynamic audio-visual scenes into reference expressions with the corresponding multimodal cues. This approach enables the formation of a comprehensive multimodal reference feature. Additionally, we employ an attention mechanism to utilize the multimodal reference cues as a prompt for the visual foundation model, thereby facilitating the final segmentation process.", + "bbox": [ + 212, + 523, + 787, + 630 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.2 Multimodal Representations", + "text_level": 1, + "bbox": [ + 215, + 650, + 496, + 666 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Audio: Similar to the video processing, we divide the audio input into clips at 1-second intervals. Audio representations $F_{A} \\in \\mathbb{R}^{t \\times d_{A}}$ is encoded from VGGish [14,16], where $t$ represents the duration of the audio in seconds and matches the number of video frames. The audio representations are extracted offline and the audio encoder is not fine-tuned.", + "bbox": [ + 212, + 672, + 787, + 747 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Visual: We sample $t$ frames from the video input at 1-second intervals and extract visual representations $F_{V} \\in \\mathbb{R}^{t \\times d_{V} \\times H \\times W}$ using a pre-trained Swin-base model [10]. The visual encoder is not fine-tuned.", + "bbox": [ + 212, + 750, + 787, + 794 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Expression: We use RoBERTa [7, 27] as our text encoder to extract textual expression features $F_{T} \\in \\mathbb{R}^{L \\times d_{T}}$ , where $L$ denotes the length of the expression. The expression representations are obtained off-the-shelf without fine-tuning.", + "bbox": [ + 212, + 794, + 787, + 840 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "Ref-AVS", + "bbox": [ + 671, + 114, + 730, + 126 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 114, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.3 Temporal Bi-modal Transformer", + "text_level": 1, + "bbox": [ + 215, + 146, + 529, + 161 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Temporal A-T & V-T Fusion: This module is introduced to retrieve the expression-related information of each modality. Firstly, for the convenience of the following multimodal fusion, we prepare the projected audio feature $F_{A'} \\in \\mathbb{R}^{t \\times 1 \\times d_V}$ , downsampled and flattened visual feature $F_{V'} \\in \\mathbb{R}^{t \\times (\\frac{W}{8} \\times \\frac{H}{8}) \\times d_V}$ , and projected text feature $F_{T'} \\in \\mathbb{R}^{t \\times L \\times d_V}$ which is expanded along the temporal dimension. Subsequently, the features are combined to yield the expression-related multimodal features $\\hat{F}_A$ , $\\hat{F}_V$ and modality-aware expression features $\\hat{F}_{T_A}$ , $\\hat{F}_{T_V}$ , with the same dimension as $F_{A'}$ , $F_{V'}$ , $F_{T'}$ , respectively:", + "bbox": [ + 212, + 172, + 787, + 294 + ], + "page_idx": 9 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {F} _ {M}, \\hat {F} _ {T _ {M}} = M F (C o n c a t \\left(F _ {M ^ {\\prime}}; F _ {T ^ {\\prime}}\\right)), M \\in \\{A, V \\} \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 328, + 308, + 785, + 325 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "where $\\text{Concat}(\\cdot)$ denotes the concatenate operation, and $MF(\\cdot)$ denotes the modality fusion function, which is employed as self-attention.", + "bbox": [ + 212, + 332, + 785, + 362 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Cached Memory: Although the aforementioned fusion method can combine cross-modal information, it does not give much emphasis to strong temporal dynamics. To address this issue, we introduce a straightforward cached memory $\\mathcal{C}_A\\in \\mathbb{R}^{t\\times 2\\times d_T},\\mathcal{C}_V\\in \\mathbb{R}^{t\\times (W\\times H + 1)\\times d_T}$ to explicitly capture the temporal mutations as they occur. The cache memory is required to store the mean modality features in the temporal domain from the beginning to the current moment. Then, the multimodal cues feature $Q_{A}$ and $Q_{V}$ output from temporal-aware transformers can be calculated as", + "bbox": [ + 212, + 363, + 787, + 483 + ], + "page_idx": 9 + }, + { + "type": "equation", + "text": "\n$$\nQ _ {M} ^ {\\tau} = (\\beta + 1) \\hat {F} _ {M} ^ {\\tau} - \\beta \\mathcal {C} _ {M} ^ {\\tau}, M \\in \\{A, V \\}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 362, + 510, + 785, + 529 + ], + "page_idx": 9 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {C} _ {M} ^ {\\tau} = \\left\\{ \\begin{array}{l l} 0, & \\text {i f} \\tau = 1, \\\\ \\frac {1}{\\tau - 1} \\sum_ {i = 0} ^ {\\tau - 1} \\hat {F} _ {M} ^ {i}, & \\text {i f} \\tau > 1, \\end{array} \\right. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 366, + 531, + 785, + 571 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "while $\\tau$ is the time stamp of the specific feature at the temporal dimension and $\\beta$ is a hyper-parameter to control the sensitivity to temporal dynamics. When the current audio or visual features do not change significantly compared to the mean of past features (i.e., $\\hat{F}_M^\\tau = \\mathcal{C}_M^\\tau$ ), $Q_M^\\tau$ remains nearly unchanged. However, in cases where the current change is drastic (i.e., $\\hat{F}_M^\\tau \\neq \\mathcal{C}_M^\\tau$ ), the cached memory can amplify the difference in the current feature, leading to an output with salient features. It is important to note that text cue feature $Q_T$ inherently possesses a highly abstracted semantic nature, which means manipulation may not be necessary at this stage. Therefore, at this stage, we combine and average the original textual features with the text features enhanced by multimodal information, resulting in a comprehensive enhanced text feature $Q_T = mean(F_{T'} + \\hat{F}_{T_A} + \\hat{F}_{T_V})$ . Finally, we obtain the processed cue feature of each modality $Q_A$ , $Q_V$ and $Q_T$ . Modality Encoding: The modality encoding approach involves incorporating modality identity tokens $[aud] \\in \\mathbb{R}^{1 \\times d_V}$ for audio modality and $[vis] \\in \\mathbb{R}^{1 \\times d_V}$ for visual modality into the following multimodal cues integration process. Since only two tokens are required to split the sequence into three segments, we only need to manipulate the audio and visual cue features.", + "bbox": [ + 212, + 580, + 787, + 840 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "Y. Wang et al.", + "bbox": [ + 271, + 114, + 370, + 127 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.4 Prompting with Multimodal Cues", + "text_level": 1, + "bbox": [ + 215, + 146, + 542, + 161 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "In this Prompting with Multimodal Cues (PMC) phrase, when we obtain the final multimodal cues features, we can concatenate the multimodal cues together and then apply self-attention to obtain comprehensive multimodal cues $Q_{m} \\in \\mathbb{R}^{t \\times (H \\times W + L + 3) \\times d_V}$ through the Multimodal Integration Transformer:", + "bbox": [ + 212, + 176, + 787, + 237 + ], + "page_idx": 10 + }, + { + "type": "equation", + "text": "\n$$\nQ _ {m} = M F (C o n c a t ([ Q _ {A}; [ a u d ]; Q _ {V}; [ v i s ]; Q _ {T} ])), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 334, + 252, + 785, + 268 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "where $MF$ is the multimodal fusion function serving as the self-attention for multimodal cues interaction. Then we utilize these comprehensive multinational cues to prompt the learnable mask queries $q \\in \\mathbb{R}^{N \\times d_V}$ in the transformer-based segmentation decoder with cross-attention, where $N$ is the number of mask queries:", + "bbox": [ + 214, + 280, + 787, + 342 + ], + "page_idx": 10 + }, + { + "type": "equation", + "text": "\n$$\nq _ {Q} = C A T F (q u e r y = q, k e y = Q _ {m}, v a l u e = Q _ {m}), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 326, + 362, + 785, + 376 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "where $q_{Q} \\in \\mathbb{R}^{N \\times d_{V}}$ is the updated mask queries. $CATF(\\cdot)$ is the cross-attention transformer with $q$ and $Q_{m}$ as the information target and source, respectively.", + "bbox": [ + 214, + 385, + 787, + 417 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "5 Experiments", + "text_level": 1, + "bbox": [ + 215, + 444, + 375, + 462 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "5.1 Implementation Details", + "text_level": 1, + "bbox": [ + 215, + 479, + 455, + 494 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Mask2Former [5] serves as our visual foundation model to provide the commonly used transformer-based segmentation decoder. By default, in this work, all input video frames are scaled to $384 \\times 384$ . The shape of the visual features is $[H = 64, W = 64, d_V = 256]$ , and we employ 8-fold downsampling to reduce the computation costs. The audio features with $d_A = 128$ are extracted from the mono-channel waveform. The text features have a shape of $[L = 25, d_T = 768]$ . For convenience, we map all modality feature dimensions to $d_V$ . Hyper-parameter $\\beta$ is set to 1 by default. Transformer layers $N_L$ of the Temporal Bi-modal Transformer, Multimodal Integration Transformer and the cross-attention transformer CATF are set to 4 by default. The number of mask queries $N_q$ is fixed to 100.", + "bbox": [ + 212, + 510, + 787, + 679 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "5.2 Evaluation Metrics", + "text_level": 1, + "bbox": [ + 215, + 704, + 419, + 718 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "To conduct a comprehensive evaluation of our Ref-AVS method, we employ the Jaccard index $(\\mathcal{I})$ and F-score $(\\mathcal{F})$ as performance metrics. Additionally, we introduce $S$ in the null reference test set to assess the efficacy of expression guidance in the model. $S$ denotes the square root of the ratio between the predicted mask area and the background area, a higher value indicates a larger proportion of the mask relative to the background area, suggesting a lack of precise expression guidance provided to the model.", + "bbox": [ + 212, + 733, + 787, + 840 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "Ref-AVS", + "bbox": [ + 669, + 114, + 730, + 126 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/2a37dd1590d1f5f46cd9dd701362eaf7a20b2dcb3ad6aeebc59a82e6eeeb90c8.jpg", + "table_caption": [ + "Table 3: Performance on Ref-AVS Bench. We use three test subsets to evaluate the comprehensive ability of Ref-AVS methods. Mix is the average performance of Seen and Unseen test set. We also use Null test set to evaluate the robustness of multimodal-cue expression guidance." + ], + "table_footnote": [], + "table_body": "
MethodTaskSeenUnseenMix(S+U)Null S (↓)Pub.
J(%)FJ(%)FJ(%)F
AVSbench [46] +textAVS21.220.45727.670.50924.450.483-ECCV'2022
23.200.51132.360.54727.780.5290.208
AVSegFormer [12] +textAVS29.160.42332.710.45130.940.437-AAAI'2024
33.470.47036.050.50134.760.4860.171
GAVS [37] +textAVS24.850.45527.860.46026.360.457-AAAI'2024
28.930.49829.820.49729.380.4980.190
ReferFormer [39] +audioRef-VOS27.460.47724.940.50426.20.491-CVPR'2022
31.310.50130.400.48830.860.4950.176
R2VOS [23] +audioRef-VOS21.340.30923.910.43122.630.370-ICCV'2023
25.010.41027.930.49826.470.4540.183
EEMCRef-AVS34.200.51349.540.64841.870.5810.007ECCV'2024
", + "bbox": [ + 220, + 212, + 784, + 397 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "5.3 Quantitative Results", + "text_level": 1, + "bbox": [ + 215, + 431, + 431, + 446 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "We compare our Ref-AVS method with existing approaches in the relevant field using our Ref-AVS benchmark. The results from the seen test set demonstrate the superior performance of our Ref-AVS method on this dataset, outperforming other methods by a significant margin. To ensure fairness, we equip the other methods with the same inputs from both the audio and visual modalities as ours. However, even with this additional information, these methods still fail to achieve satisfactory results. The result indicates that simple modal fusion is inadequate for addressing the challenges of understanding multimodal cues within the Ref-AVS task. Instead of directly incorporating audiovisual information, our approach selects textual representations as carriers of multimodal information because they contain rich semantics and cues that are contextually relevant to the current audiovisual environment.", + "bbox": [ + 212, + 473, + 787, + 654 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "We conduct tests on the unseen test set and the null reference test set to explore the generalization and multimodal cues-following ability. Our method still maintains a leading position in the unseen test set because we leverage text with highly abstract semantic capabilities as a multimodal information carrier, instead of directly fusing audio and visual information. Therefore, the obtained multimodal cues can provide more robust semantic guidance. On the null test set, we obtain leading results among all other methods, indicating that our method can perceive multimodal cues to a reasonably accurate degree. Compared to other transformer-based and query-based frameworks, AVSBench, as a classic AVS method, obtains inferior results. This may be attributed to the model simply fuses multimodal cues and visual features in a progressive fashion, resulting in relatively weak guidance over the visual space.", + "bbox": [ + 212, + 657, + 787, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "Y. Wang et al.", + "bbox": [ + 271, + 114, + 370, + 128 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/8a53bb49ce7fbca751068fdbeb6a515b6dc1fdab9bff663dc95ac3199a192dd2.jpg", + "image_caption": [ + "Fig. 4: Segmentation masks of the referred objects in the Ref-AVS Bench. Note, for $\\bullet \\bullet \\bullet$ , the sound volume increases successively from silent to loud. We compare our methodology with AVSegformer and ReferFormer. More cases can be found in the supplementary material." + ], + "image_footnote": [], + "bbox": [ + 220, + 142, + 519, + 410 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/b101478385e0d87e0a8d8c96351d50fade3155d9a35b30135bd8b637994afd51.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 534, + 142, + 787, + 410 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "5.4 Qualitative Results", + "text_level": 1, + "bbox": [ + 215, + 506, + 419, + 522 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "We visualize the segmentation masks on the test set of the Ref-AVS Bench and compare them with AVSegFormer and ReferFormer, as shown in Fig. 4. From these qualitative results, we observe that both ReferFormer in the Ref-VOS task and AVSegFormer in the AVS task fail to accurately segment the object described in the expression. Specifically, AVSegformer struggles to fully comprehend the expression and tends to directly generate the sound source. This issue is exemplified in the bottom-left sample, where AVSegformer erroneously segments the vacuum cleaner instead of the boy. On the other hand, Ref-VOS may not adequately understand the audio-visual scene and thus mistakenly identify the toddler as the piano player, as shown in the top-right sample. In contrast, our Ref-AVS method demonstrates a superior capability to simultaneously process multi-modal expressions and scenes, enabling it to accurately interpret the user instruction and segment the intended object of interest.", + "bbox": [ + 212, + 532, + 787, + 729 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "5.5 Ablation Study", + "text_level": 1, + "bbox": [ + 215, + 753, + 390, + 768 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "We conduct ablation studies to investigate the impact of two modalities (audio and text) information on the Ref-AVS task, as well as the effectiveness of the proposed method. As shown in the Tab. 4, we observe that in setting $②$ , removing the text information results in a significant performance degradation of $11.44\\%$ in", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "Ref-AVS", + "bbox": [ + 671, + 114, + 730, + 126 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/a2fa8cf2d398642af24bcbb26672de4e0d060dc7c42c7156445e2a064129a74d.jpg", + "table_caption": [ + "Table 4: Ablation study on modalities and method modules." + ], + "table_footnote": [], + "table_body": "
SettingSeenUnseenMixNull
J(%)FJ(%)FJ(%)FS(↓)
① EEMC34.200.51349.540.64841.870.5810.007
② (-) Text23.670.46837.190.64130.430.5550.011
③ (-) Audio27.270.49243.640.65635.460.5740.010
④ (-) Cached Memory25.300.49436.730.66031.050.5770.009
⑤ (-) Modality Encoding27.200.47044.970.63736.090.5540.009
", + "bbox": [ + 254, + 170, + 750, + 268 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "$\\mathcal{I}$ and $2.60\\%$ in $\\mathcal{F}$ . This degradation is much higher compared to setting ③, where the removal of the audio information results in a combined mixed drop of $6.41\\%$ in $\\mathcal{I}$ and $0.70\\%$ in $\\mathcal{F}$ . This phenomenon primarily results from the clarity and directness of textual information as a reference source. In contrast, when relying solely on audio information, the model often overlooks the referential content and instead focuses on identifying objects visually associated with audible behaviors. We also conduct experiments to examine the function of each module in this framework. The cached memory is designed to track significant changes in the temporal domain. Meanwhile, the purpose of modality encoding is to extract more distinct and thorough features from multimodal cues, enhancing modal perception. Results from settings ④ and ⑤ in Tab. 4 demonstrate the effectiveness of these modules.", + "bbox": [ + 212, + 306, + 787, + 487 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "6 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 507, + 359, + 523 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In this work, we introduce a novel task, Ref-AVS, which requires the machine to segment objects of interest in dynamic audio-visual scenes based on expressions that incorporate multimodal cues and temporal information. To support this field, we develop the first benchmark, Ref-AVS Bench, for performance training and evaluation. Our method, EEMC, establishes a robust baseline with an advanced multimodal cues fusion module. We compare our method with several existing approaches on the Ref-AVS Bench dataset and demonstrate its promising performance in accurately locating objects using multimodal reference expressions. Essentially, Ref-AVS leverages audio, visual, and textual data to enrich real-world scene comprehension. Our task has potential in applications such as attention focusing and object editing in immersive audio-visual scenarios like extended reality, thereby enhancing user interaction experiences. In the future, we plan to further expand the scale of the current dataset to meet the growing data demands in the era of large language models.", + "bbox": [ + 212, + 535, + 787, + 748 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 215, + 766, + 403, + 784 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "This research was supported by National Natural Science Foundation of China (NO.62106272), and Public Computing Cloud, Renmin University of China. This research was partially supported by Shanghai Artificial Intelligence Laboratory.", + "bbox": [ + 212, + 794, + 785, + 840 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "Y. Wang et al.", + "bbox": [ + 271, + 114, + 370, + 128 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 143, + 321, + 159 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Botach, A., Zheltonozhskii, E., Baskin, C.: End-to-end referring video object segmentation with multimodal transformers. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4985-4995 (2022)", + "2. Chen, H., Xie, W., Afouras, T., Nagrani, A., Vedaldi, A., Zisserman, A.: Localizing visual sounds the hard way. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16867-16876 (2021)", + "3. Chen, H., Xie, W., Vedaldi, A., Zisserman, A.: Vggsound: A large-scale audiovisual dataset. In: ICASSP 2020-2020 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). pp. 721-725. IEEE (2020)", + "4. Chen, Z., Wang, L., Wang, P., Gao, P.: Question-aware global-local video understanding network for audio-visual question answering. IEEE Transactions on Circuits and Systems for Video Technology (2023)", + "5. Cheng, B., Choudhuri, A., Misra, I., Kirillov, A., Girdhar, R., Schwing, A.G.: Mask2former for video instance segmentation (2021)", + "6. Cheng, H.K., Oh, S.W., Price, B., Schwing, A., Lee, J.Y.: Tracking anything with decoupled video segmentation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 1316-1326 (2023)", + "7. Devlin, J., Chang, M.W., Lee, K., Toutanova, K.: Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805 (2018)", + "8. Ding, H., Liu, C., He, S., Jiang, X., Loy, C.C.: Mevis: A large-scale benchmark for video segmentation with motion expressions. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 2694-2703 (October 2023)", + "9. Ding, H., Liu, C., He, S., Jiang, X., Torr, P.H., Bai, S.: Mose: A new dataset for video object segmentation in complex scenes. arXiv preprint arXiv:2302.01872 (2023)", + "10. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929 (2020)", + "1. Dutta, U.K., Harandi, M., Sekhar, C.C.: Unsupervised deep metric learning via orthogonality based probabilistic loss. IEEE Transactions on Artificial Intelligence 1(1), 74-84 (2020)", + "2. Gao, S., Chen, Z., Chen, G., Wang, W., Lu, T.: Avsegformer: Audio-visual segmentation with transformer (2023)", + "3. Gavrilyuk, K., Ghodrati, A., Li, Z., Snoek, C.G.: Actor and action video segmentation from a sentence. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 5958-5966 (2018)", + "4. Gemmeke, J.F., Ellis, D.P., Freedman, D., Jansen, A., Lawrence, W., Moore, R.C., Plakal, M., Ritter, M.: Audio set: An ontology and human-labeled dataset for audio events. In: 2017 IEEE international conference on acoustics, speech and signal processing (ICASSP). pp. 776-780. IEEE (2017)", + "5. Guangyao li, Jake Wei, Y.T.C.X.J.R.W.D.H.: Learning to answer questions in dynamic audio-visual scenarios. IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2022)", + "6. Hershey, S., Chaudhuri, S., Ellis, D.P., Gemmeke, J.F., Jansen, A., Moore, R.C., Plakal, M., Platt, D., Saurous, R.A., Seybold, B., et al.: Cnn architectures for" + ], + "bbox": [ + 225, + 175, + 784, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "Ref-AVS", + "bbox": [ + 671, + 114, + 730, + 126 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 785, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "large-scale audio classification. In: 2017 IEEE international conference on acoustics, speech and signal processing (ICASSP). pp. 131-135. IEEE (2017)", + "17. Hu, D., Qian, R., Jiang, M., Tan, X., Wen, S., Ding, E., Lin, W., Dou, D.: Discriminative sounding objects localization via self-supervised audiovisual matching. Advances in Neural Information Processing Systems 33, 10077-10087 (2020)", + "18. Huang, S., Li, H., Wang, Y., Zhu, H., Dai, J., Han, J., Rong, W., Liu, S.: Discovering sounding objects by audio queries for audio visual segmentation (2023)", + "19. Jiang, Y., Yin, J.: Target-aware spatio-temporal reasoning via answering questions in dynamics audio-visual scenarios. arXiv preprint arXiv:2305.12397 (2023)", + "20. Khoreva, A., Rohrbach, A., Schiele, B.: Video object segmentation with language referring expressions. In: Computer Vision-ACCV 2018: 14th Asian Conference on Computer Vision, Perth, Australia, December 2-6, 2018, Revised Selected Papers, Part IV 14. pp. 123-141. Springer (2019)", + "21. Li, G., Hou, W., Hu, D.: Progressive spatio-temporal perception for audio-visual question answering. In: Proceedings of the 31st ACM International Conference on Multimedia. p. 7808-7816. MM '23, Association for Computing Machinery, New York, NY, USA (2023). https://doi.org/10.1145/3581783.3612293, https://doi.org/10.1145/3581783.3612293", + "22. Li, K., Yang, Z., Chen, L., Yang, Y., Xun, J.: Catr: Combinatorial-dependence audio-queried transformer for audio-visual video segmentation. arXiv preprint arXiv:2309.09709 (2023)", + "23. Li, X., Wang, J., Xu, X., Li, X., Raj, B., Lu, Y.: Robust referring video object segmentation with cyclic structural consensus. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 22236-22245 (2023)", + "24. Ling, Y., Li, Y., Gan, Z., Zhang, J., Chi, M., Wang, Y.: Hear to segment: Unmixing the audio to guide the semantic segmentation (2023)", + "25. Liu, J., Ju, C., Ma, C., Wang, Y., Wang, Y., Zhang, Y.: Audio-aware query-enhanced transformer for audio-visual segmentation (2023)", + "26. Liu, J., Wang, Y., Ju, C., Zhang, Y., Xie, W.: Annotation-free audio-visual segmentation. arXiv preprint arXiv:2305.11019 (2023)", + "27. Liu, Y., Ott, M., Goyal, N., Du, J., Joshi, M., Chen, D., Levy, O., Lewis, M., Zettlemoyer, L., Stoyanov, V.: Roberta: A robustly optimized bert pretraining approach. arXiv preprint arXiv:1907.11692 (2019)", + "28. Ma, J., Sun, P., Wang, Y., Hu, D.: Stepping stones: A progressive training strategy for audio-visual semantic segmentation. IEEE European Conference on Computer Vision (ECCV) (2024)", + "29. Mo, S., Tian, Y.: Av-sam: Segment anything model meets audio-visual localization and segmentation. arXiv preprint arXiv:2305.01836 (2023)", + "30. Perazzi, F., Khoreva, A., Benenson, R., Schiele, B., Sorkine-Hornung, A.: Learning video object segmentation from static images. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 2663-2672 (2017)", + "31. Perazzi, F., Pont-Tuset, J., McWilliams, B., Van Gool, L., Gross, M., Sorkine-Hornung, A.: A benchmark dataset and evaluation methodology for video object segmentation. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 724-732 (2016)", + "32. Pont-Tuset, J., Perazzi, F., Caelles, S., Arbeláez, P., Sorkine-Hornung, A., Van Gool, L.: The 2017 davis challenge on video object segmentation. arXiv preprint arXiv:1704.00675 (2017)", + "33. Ren, T., Liu, S., Zeng, A., Lin, J., Li, K., Cao, H., Chen, J., Huang, X., Chen, Y., Yan, F., Zeng, Z., Zhang, H., Li, F., Yang, J., Li, H., Jiang, Q., Zhang, L.: Grounded sam: Assembling open-world models for diverse visual tasks (2024)" + ], + "bbox": [ + 215, + 147, + 785, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "Y. Wang et al.", + "bbox": [ + 271, + 114, + 370, + 127 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "34. Senocak, A., Oh, T.H., Kim, J., Yang, M.H., So Kweon, I.: Learning to localize sound source in visual scenes. In: The IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (June 2018)", + "35. Seo, S., Lee, J.Y., Han, B.: Urvos: Unified referring video object segmentation network with a large-scale benchmark. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XV 16. pp. 208-223. Springer (2020)", + "36. Tian, Y., Shi, J., Li, B., Duan, Z., Xu, C.: Audio-visual event localization in unconstrained videos. In: Proceedings of the European Conference on Computer Vision (ECCV) (September 2018)", + "37. Wang, Y., Liu, W., Li, G., Ding, J., Hu, D., Li, X.: Prompting segmentation with sound is generalizable audio-visual source localizer. arXiv preprint arXiv:2309.07929 (2023)", + "38. Wang, Y., Sun, P., Li, Y., Zhang, H., Hu, D.: Can textual semantics mitigate sounding object segmentation preference? IEEE European Conference on Computer Vision (ECCV) (2024)", + "39. Wu, J., Jiang, Y., Sun, P., Yuan, Z., Luo, P.: Language as queries for referring video object segmentation (2022)", + "40. Xu, N., Yang, L., Fan, Y., Yue, D., Liang, Y., Yang, J., Huang, T.: Youtube-vos: A large-scale video object segmentation benchmark. arXiv preprint arXiv:1809.03327 (2018)", + "41. Yan, S., Zhang, R., Guo, Z., Chen, W., Zhang, W., Li, H., Qiao, Y., He, Z., Gao, P.: Referred by multi-modality: A unified temporal transformer for video object segmentation (2023)", + "42. Yang, P., Wang, X., Duan, X., Chen, H., Hou, R., Jin, C., Zhu, W.: Avqa: A dataset for audio-visual question answering on videos. In: Proceedings of the 30th ACM International Conference on Multimedia. pp. 3480-3491 (2022)", + "43. Yu, L., Lin, Z., Shen, X., Yang, J., Lu, X., Bansal, M., Berg, T.L.: Mattnet: Modular attention network for referring expression comprehension. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1307-1315 (2018)", + "44. Yuanhong, C., Yuyuan, L., Hu, W., Fengbei, L., Chong, W., Gustavo, C.: A closer look at audio-visual semantic segmentation. arXiv preprint arXiv:2304.02970 (2023)", + "45. Zhou, J., Shen, X., Wang, J., Zhang, J., Sun, W., Zhang, J., Birchfield, S., Guo, D., Kong, L., Wang, M., Zhong, Y.: Audio-visual segmentation with semantics. arXiv preprint arXiv:2301.13190 (2023)", + "46. Zhou, J., Wang, J., Zhang, J., Sun, W., Zhang, J., Birchfield, S., Guo, D., Kong, L., Wang, M., Zhong, Y.: Audio-visual segmentation. In: European Conference on Computer Vision (2022)", + "47. Zhou, X., Zhou, D., Hu, D., Zhou, H., Ouyang, W.: Exploiting visual context semantics for sound source localization. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 5199-5208 (2023)" + ], + "bbox": [ + 212, + 146, + 787, + 742 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "Ref-AVS", + "bbox": [ + 671, + 114, + 730, + 126 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + } +] \ No newline at end of file diff --git a/2024/Ref-AVS_ Refer and Segment Objects in Audio-Visual Scenes/a65e28a6-a79d-4af3-846c-76b29927236e_model.json b/2024/Ref-AVS_ Refer and Segment Objects in Audio-Visual Scenes/a65e28a6-a79d-4af3-846c-76b29927236e_model.json new file mode 100644 index 0000000000000000000000000000000000000000..90f24239eefb73a83170b9d11ac9165ad11cc179 --- /dev/null +++ b/2024/Ref-AVS_ Refer and Segment Objects in Audio-Visual Scenes/a65e28a6-a79d-4af3-846c-76b29927236e_model.json @@ -0,0 +1,2335 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.272, + 0.141, + 0.733, + 0.185 + ], + "angle": 0, + "content": "Ref-AVS: Refer and Segment Objects in Audio-Visual Scenes" + }, + { + "type": "text", + "bbox": [ + 0.24, + 0.212, + 0.761, + 0.244 + ], + "angle": 0, + "content": "Yaoting Wang\\(^{1\\dagger}\\), Peiwen Sun\\(^{2\\dagger}\\), Dongzhan Zhou\\(^{3\\dagger}\\), Guangyao Li\\(^{1}\\), Honggang Zhang\\(^{2}\\), and Di Hu\\(^{\\boxtimes 1,4}\\)" + }, + { + "type": "text", + "bbox": [ + 0.242, + 0.255, + 0.761, + 0.297 + ], + "angle": 0, + "content": "1 Gaoling School of Artificial Intelligence, Renmin University of China, China \nyaoting.wang@outlook.com \n{guangyaoli, dihu}@ruc.edu.cn" + }, + { + "type": "text", + "bbox": [ + 0.269, + 0.297, + 0.733, + 0.325 + ], + "angle": 0, + "content": "\\(^{2}\\) Beijing University of Posts and Telecommunications, Beijing, China {sunpeiwen,zhhg}@bupt.edu.cn" + }, + { + "type": "text", + "bbox": [ + 0.295, + 0.325, + 0.709, + 0.353 + ], + "angle": 0, + "content": "3 Shanghai Artificial Intelligence Laboratory, Shanghai, China zhoudongzhan@pjlab.org.cn" + }, + { + "type": "text", + "bbox": [ + 0.235, + 0.353, + 0.767, + 0.366 + ], + "angle": 0, + "content": "4 Engineering Research Center of Next-Generation Search and Recommendation" + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.392, + 0.741, + 0.628 + ], + "angle": 0, + "content": "Abstract. Traditional reference segmentation tasks have predominantly focused on silent visual scenes, neglecting the integral role of multimodal perception and interaction in human experiences. In this work, we introduce a novel task called Reference Audio-Visual Segmentation (Ref-AVS), which seeks to segment objects within the visual domain based on expressions containing multimodal cues. Such expressions are articulated in natural language forms but are enriched with multimodal cues, including audio and visual descriptions. To facilitate this research, we construct the first Ref-AVS benchmark, which provides pixel-level annotations for objects described in corresponding multimodal-cue expressions. To tackle the Ref-AVS task, we propose a new method that adequately utilizes multimodal cues to offer precise segmentation guidance. Finally, we conduct quantitative and qualitative experiments on three test subsets to compare our approach with existing methods from related tasks. The results demonstrate the effectiveness of our method, highlighting its capability to precisely segment objects using multimodal-cue expressions. Dataset is available at https://gewu-lab.github.io/Ref-AVS." + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.64, + 0.741, + 0.668 + ], + "angle": 0, + "content": "Keywords: Referring Audio-Visual Segmentation \\(\\cdot\\) Audio-Visual Segmentation \\(\\cdot\\) Multimodal Learning" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.707, + 0.375, + 0.723 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.733, + 0.788, + 0.763 + ], + "angle": 0, + "content": "In the real world, visual scenes are often accompanied by diverse multimodal information1, including audio and text modalities. This additional multimodal" + }, + { + "type": "page_footnote", + "bbox": [ + 0.233, + 0.77, + 0.378, + 0.784 + ], + "angle": 0, + "content": "\\(\\dagger\\) Equal contribution." + }, + { + "type": "page_footnote", + "bbox": [ + 0.233, + 0.784, + 0.399, + 0.798 + ], + "angle": 0, + "content": "\\(\\mathbb{O}\\) Corresponding author." + }, + { + "type": "page_footnote", + "bbox": [ + 0.223, + 0.799, + 0.786, + 0.84 + ], + "angle": 0, + "content": "To clarify, we refer to the sound and vision in the scene as multimodal information, while the attribute description of audio and vision (e.g., louder and on the right) in the reference expression is named multimodal cues." + }, + { + "type": "list", + "bbox": [ + 0.223, + 0.77, + 0.786, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.371, + 0.129 + ], + "angle": 0, + "content": "Y. Wang et al." + }, + { + "type": "image", + "bbox": [ + 0.22, + 0.146, + 0.415, + 0.241 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.227, + 0.242, + 0.415, + 0.276 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.227, + 0.277, + 0.415, + 0.313 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.419, + 0.147, + 0.588, + 0.241 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.425, + 0.242, + 0.587, + 0.276 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.407, + 0.277, + 0.572, + 0.313 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.593, + 0.15, + 0.772, + 0.161 + ], + "angle": 0, + "content": "Which one should the camera focus?" + }, + { + "type": "text", + "bbox": [ + 0.593, + 0.164, + 0.776, + 0.184 + ], + "angle": 0, + "content": "An ideal machine needs to know always going for the player who is actually playing." + }, + { + "type": "text", + "bbox": [ + 0.593, + 0.194, + 0.703, + 0.204 + ], + "angle": 0, + "content": "In the view of camera." + }, + { + "type": "text", + "bbox": [ + 0.593, + 0.211, + 0.754, + 0.231 + ], + "angle": 0, + "content": "How to focus the player who is actually playing." + }, + { + "type": "image", + "bbox": [ + 0.593, + 0.232, + 0.768, + 0.3 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.325, + 0.788, + 0.368 + ], + "angle": 0, + "content": "Fig. 1: Comparison of the Ref-AVS task with other related tasks. Ref-AVS challenges machines to locate objects of interest in the visual space using multimodal cues, just like humans do in the real world." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.4, + 0.789, + 0.505 + ], + "angle": 0, + "content": "information offers valuable spatiotemporal and semantic cues, aiding individuals in locating objects of interest. Similarly, the ability to locate objects based on multimodal cues plays a crucial role in the functioning of machines. For example, stage cameras should be able to locate the musician playing instruments, while movie cameras should locate the superhero in front of the yelling villain. In the above scenarios, using a single modality for segmentation alone cannot achieve the desired behavior." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.507, + 0.789, + 0.809 + ], + "angle": 0, + "content": "To date, researchers have studied the reference segmentation problem from somewhat limited scenarios. As shown in the coordinate system in Fig. 1, there are currently three mainstream developments in segmentation methods for different modalities. Firstly, from the perspective of visual cues, Video Object Segmentation (VOS) [9,30-32,40] emerges with annotated first frame mask as the reference, guiding the segmentation of specific objects in subsequent video frames. While VOS can achieve arbitrary object segmentation in videos, it heavily relies on precise annotation of the first frame, which can be challenging and time-consuming in practice. Secondly, from the perspective of text cues, Referring Video Object Segmentation (R-VOS) [8,11,13,20] explores the segmentation ability by using attribute descriptions as guidance. As R-VOS has successfully replaced the mask annotation used in VOS with natural language, it provides a more accessible and user-friendly reference form but the capability to locate objects in more natural dynamic audio-visual scenes is still limited. In contrast, from the perspective of audio cues, Audio-Visual Segmentation (AVS) [26,37,44-47] leverages audio as a guide to segment sounding objects. This approach effectively addresses the challenge of locating objects in dynamic audio-visual scenes. However, AVS still faces limitations in its ability to focus on general objects that do not generate sound and cannot effectively locate objects of specific interest." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.789, + 0.841 + ], + "angle": 0, + "content": "In simple terms, current work falls short of enabling machines to locate objects of interest in natural dynamic audio-visual scenes. For instance, as illus" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.672, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "Ref-AVS" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.265 + ], + "angle": 0, + "content": "trated in Fig. 1, how can machines accurately locate a person who is genuinely playing a musical instrument over time? This requires machines to infer which instrument is making a sound and who is playing that instrument. Though Yan et al. [41] utilize either audio or language modality for multimodal reference segmentation, their method does not integrate multimodal information simultaneously nor support the aforementioned scenario. Therefore, introducing a task to explore the possibility of locating interested objects in natural dynamic audio-visual scenes holds significant practical potential." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.268, + 0.785, + 0.537 + ], + "angle": 0, + "content": "In this paper, we propose a pixel-level segmentation task called Referring Audio-Visual Segmentation (Ref-AVS), which requires the network to densely predict whether each pixel corresponds to the given multimodal-cue expression, including dynamic audio-visual information. Top-left of Fig. 1 highlights the distinctions between Ref-AVS and previous tasks. The Ref-AVS task poses a greater challenge as it requires the network to accurately locate and segment objects in a more complex and stereoscopic modality space. To this end, a computation model with comprehensive multimodal understanding capabilities is necessary. To foster research advancements in this field, we introduce the Ref-AVS Bench, the pioneering benchmark that focuses on locating and segmenting objects of interest using referring multimodal-cue expressions. Considering the complexity of real-world audio-visual scenes, we collect about 4,000 audible video clips from YouTube of which more than \\(60\\%\\) are multi-source sound scenarios. More than 20,000 reference expressions are collected and verified by experts to ensure accuracy and reliability, which adopt multimodal cues to describe objects in diverse and dynamic audio-visual scenes. Moreover, a special unseen test set is considered to evaluate the model's generalization ability in the growing demand for zero-shot scenarios." + }, + { + "type": "text", + "bbox": [ + 0.24, + 0.541, + 0.591, + 0.555 + ], + "angle": 0, + "content": "Our contributions can be summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.568, + 0.784, + 0.611 + ], + "angle": 0, + "content": "- We propose Ref-AVS as a challenging scene understanding task that segments objects of interest with multimodal-cue expressions and provide the corresponding Ref-AVS benchmark for performance training and validation." + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.614, + 0.784, + 0.656 + ], + "angle": 0, + "content": "- We design an end-to-end framework for Ref-AVS that efficiently processes the multimodal cues with a crossmodal transformer, serving as a feasible research framework for future development." + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.66, + 0.784, + 0.702 + ], + "angle": 0, + "content": "- We conduct extensive experiments to demonstrate the advantages of considering multimodal cues for visual segmentation, which also indicate the superiority of our approach in all subsets." + }, + { + "type": "list", + "bbox": [ + 0.225, + 0.568, + 0.784, + 0.702 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.732, + 0.388, + 0.748 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.767, + 0.652, + 0.783 + ], + "angle": 0, + "content": "2.1 Referring Video Object Segmentation (R-VOS)" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.784, + 0.84 + ], + "angle": 0, + "content": "The main objective of R-VOS is to perform object segmentation in streaming videos based on given natural language expressions. Initially proposed in 2018, A2D-Sentence [13] aims to segment actors in video content based on" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.371, + 0.129 + ], + "angle": 0, + "content": "Y. Wang et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.36 + ], + "angle": 0, + "content": "descriptions of their actions. Subsequently, Refer-DAVIS17 [20] and JHMDB-Sentences [11] use extra language as references to replace previous VOS settings. To facilitate large-scale R-VOS training, Seo et al. [35] develop the ReferYouTube-VOS dataset upon the YouTube-VOS-2019 dataset [40] and Henghui et al. [8] build the MeViS for complex motion expression. Overall, the primary R-VOS methods have evolved from semi-supervised VOS and incorporate attention mechanisms to enhance modalities interaction. For example, MAttNet [43] decomposes the expression information into language- and word-level attention, while URVOS [35] introduces crossmodal attention and memory attention. ReferFormer [39] and MTTR [1] utilize the crossmodal transformer to connect language expression with image regions. These datasets typically offer expressions that encompass the action or appearance attributes of the target object. However, these expressions primarily focus on visual information, which fails to meet the increasing multimodal demands in today's diverse audio-visual scenes." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.379, + 0.546, + 0.396 + ], + "angle": 0, + "content": "2.2 Audio-Visual Segmentation (AVS)" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.403, + 0.789, + 0.584 + ], + "angle": 0, + "content": "The multimodal segmentation in audio-visual scenes, which is of great concern to us, has received some attention in AVS [12, 37, 46]. AVS aims to obtain finer pixel-level masks corresponding to sound-emitting objects in the visual space. In contrast, previous studies on audio-visual localization [17, 36] use heat maps for coarse localization in an unsupervised learning manner. The existing works on AVS can be broadly categorized into fusion-based methods [12, 18, 22, 24, 25, 46] and prompt-based methods [26, 28, 29, 37]. The former primarily focuses on localizing sounding objects by fusing audio and visual features, while the latter emphasizes constructing effective audio prompts for the visual foundation model. However, AVS simply segments all sound-emitting objects in the visual space, without the flexibility to combine the multimodal cues to segment specific objects of interest." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.605, + 0.686, + 0.621 + ], + "angle": 0, + "content": "2.3 Language-aided Audio-Visual Scene Understanding" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.629, + 0.787, + 0.719 + ], + "angle": 0, + "content": "Currently, there is a limited number of public tasks offering datasets for audio-visual scene comprehension accompanied by language assistance. Two notable examples of such datasets are Music-AVQA [15] and AVQA [42], both encompassing audio, visual, and language information. These datasets contain rich audio-visual components and focus on annotations for textual questions and answers, emphasizing temporal (Before/After), spatial cues (Left/Right), etc." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Researchers [4, 19, 21] acknowledge the significance of the provided question, as it guides the feature extraction process for both audio and visual signals. Consequently, they have made efforts to identify pertinent segments by evaluating the semantic similarity between the question and temporal segments. Furthermore, [21] has been dedicated to locating crucial areas by leveraging semantic similarity between questions and visual tokens. Clearly, existing explorations have significantly advanced the research on language-aided audio-visual scene understanding. Nevertheless, the works that rely on the aforementioned datasets" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.672, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "Ref-AVS" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.193 + ], + "angle": 0, + "content": "are still unable to offer pixel-level scene understanding. This encourages us to begin exploring fine-grained segmentation of dynamic audio-visual scenes in the real world by constructing appropriate datasets." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.216, + 0.439, + 0.234 + ], + "angle": 0, + "content": "3 Refer-AVS Dataset" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.248, + 0.403, + 0.265 + ], + "angle": 0, + "content": "3.1 Object Category" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.275, + 0.788, + 0.382 + ], + "angle": 0, + "content": "To ensure a diverse range of referred objects, we have carefully selected a wide variety of audible objects in 48 categories as well as a smaller selection of static, unsoundable objects in 3 categories. For the objects that can produce sound, we have chosen 20 categories of musical instruments, 8 of animals, 15 of machines, and 5 of humans. In the special case of humans, considering their diverse appearances, voices, and actions, we have employed the concept of morphology and classified them into 5 categories, based on age and gender." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.403, + 0.392, + 0.418 + ], + "angle": 0, + "content": "3.2 Video Selection" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.429, + 0.788, + 0.549 + ], + "angle": 0, + "content": "During the process of video collection, we employed the techniques presented in [3,46] to ensure the alignment of audio and visual snippets with the intended semantics. All videos were sourced from YouTube under the Creative Commons license, and each video was trimmed to a duration of 10 seconds. Throughout the manual collection process, we deliberately avoided videos falling into several categories (detailed in the appendix): 1) Videos with a large number of instances of the same semantics; 2) Videos characterized by extensive editing and camera switching; 3) Non-realistic videos containing synthetic artifacts." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.55, + 0.788, + 0.655 + ], + "angle": 0, + "content": "To raise the alignment with real-world distributions, we carefully select videos that contribute to the diversification of scenes within our dataset. Specifically, we focus on selecting videos that involve interactions between multiple objects, such as musical instruments, people, vehicles, etc. The rich combination of categories indicates that our dataset is not limited to a narrow set of scenarios but rather encompasses a broad spectrum of real-life scenes where such objects are likely to naturally appear together. Refer to the supplementary materials for details." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.656, + 0.788, + 0.716 + ], + "angle": 0, + "content": "In addition to diversity, we also carefully filter the videos to ensure that the dataset includes scenes with greater complexity and a larger number of objects. Specifically, \\(56\\%\\) of the total videos contain two or more objects, while \\(13\\%\\) of the total videos contain three or more objects." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.739, + 0.36, + 0.755 + ], + "angle": 0, + "content": "3.3 Expressions" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.788, + 0.843 + ], + "angle": 0, + "content": "The diversity of expressions is one of the core factors in the assembly of the Ref-AVS dataset. The expressions consist of three dimensions of information, namely audio, vision, and time. The auditory dimension incorporates characteristics such as volume and rhythm while the visual dimension encompasses attributes like the appearance and spatial configuration of objects. We also leverage temporal" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.371, + 0.129 + ], + "angle": 0, + "content": "Y. Wang et al." + }, + { + "type": "image", + "bbox": [ + 0.219, + 0.146, + 0.784, + 0.313 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.325, + 0.788, + 0.385 + ], + "angle": 0, + "content": "Fig. 2: The illustration of our Ref-AVS benchmark. Note, for \\(\\bullet \\bullet \\bullet\\), the sound volume increases successively from silent to loud. Our benchmark is meticulously designed to encompass multimodal expressions from multiple dimensions. By combining various types of modality expressions, we achieve a dataset that exhibits great diversity." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.413, + 0.789, + 0.488 + ], + "angle": 0, + "content": "cues to generate references with sequential hints, e.g., \"the one making sound first\" or \"the one appearing later\". By integrating auditory, visual, and temporal information, we craft a rich array of expressions that not only accurately reflect multimodal scenarios but also meet users' specific needs for precise references. Fig. 2 illustrates examples of the combination of different modalities." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.489, + 0.789, + 0.624 + ], + "angle": 0, + "content": "The accuracy of expressions is also a core concern. We follow three rules to generate high-quality expressions: 1) Uniqueness: The object referred to by an expression must be unique and an expression cannot refer to multiple objects simultaneously. 2) Necessity: Complex expressions can be used for reference, but each adjective in the sentence should narrow down the scope of target objects, avoiding unnecessary and redundant descriptions of the object. 3) Clarity: Certain expression templates involve subjective factors, such as \"the _ with a louder sound\". The use of such expressions should only occur when the situation is clear enough to avoid ambiguous references." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.625, + 0.789, + 0.701 + ], + "angle": 0, + "content": "Besides diversity and accuracy, we have graded the difficulty of expressions based on the number of cues involved, with the proportions of easy, medium, and hard samples in the overall dataset being \\(20\\%\\), \\(60\\%\\), and \\(20\\%\\) respectively. This gradation could be a great benefit to future research like curriculum learning. Please refer to the supplementary material for more details." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.723, + 0.434, + 0.739 + ], + "angle": 0, + "content": "3.4 Segmentation Masks" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.81 + ], + "angle": 0, + "content": "We divide each 10-second video into ten equal 1-second snippets and the objective of annotation is to acquire the first frame's mask for each snippet. For these sampled frames, the ground truth labels are binary masks indicating the target object, according to expressions and multimodal information." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.789, + 0.842 + ], + "angle": 0, + "content": "To obtain such a mask, initially, we manually select the pivotal frames for each 10-second video in which the target object is present. These pivotal frames" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.672, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "Ref-AVS" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.126 + ], + "angle": 0, + "content": "7" + }, + { + "type": "table_caption", + "bbox": [ + 0.335, + 0.145, + 0.668, + 0.159 + ], + "angle": 0, + "content": "Table 1: Datasets comparison with related tasks." + }, + { + "type": "table", + "bbox": [ + 0.221, + 0.171, + 0.794, + 0.307 + ], + "angle": 0, + "content": "
TaskDatasetModalityVideoFrameObjectExpressionObject/VideoPub.
AVLFlickr-SoundNet [34]A+V5,0005,000---CVPR‘2018
VGG-SS [2]A+V5,1585,158---CVPR‘2021
AVSAVS [46]A+V12,35682,97213,500~-1.09~ECCV‘2022
VPO [44]A+V-34,87440,000~-1.14 (Object/Frame)CVPR‘2024
R-VOSJ-HMDB Sentences [11]T+V9289289289281.28CVPR‘2018
A2D Sentences [13]T+V3,78211,9364,8256,6561CVPR‘2018
Refer-DAVIS16 [20]T+V503,455501001ACCV‘2018
Refer-DAVIS17 [20]T+V9013,5432051,5442.27ACCV‘2018
Refer-Youtu-VOS [35]T+V3,97511,9367,45127,8991.86ECCV‘2020
MeViS [8]T+V2,006443,0008,17128,5704.28ICCV‘2023
Ref-AVSRef-AVS BenchA+T+V4,00240,0206,88820,2611.72ECCV‘2024
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.339, + 0.788, + 0.462 + ], + "angle": 0, + "content": "may occur at the beginning, middle, or end of the video, depending on when the target object is most clearly visible. Subsequently, we utilize Grounding SAM [33] to segment and label the pivotal frames, which will be manually checked and corrected subsequently. This process allows us to generate masks and labels for multiple target objects within the pivotal frames. Once the masks for the pivotal frames are established, we apply a tracking algorithm [6] to track the target object across the preceding and subsequent frames and obtain the ultimate mask and label for the target object in a span of 10 frames." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.489, + 0.409, + 0.503 + ], + "angle": 0, + "content": "3.5 Dataset Statistics" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.522, + 0.788, + 0.748 + ], + "angle": 0, + "content": "In addition, we compare Refer-AVS with other popular audio-visual benchmarks in the Tab. 1. The Flickr-SoundNet [34] and VGG-SS [2] benchmarks are labeled at a patch level through bounding boxes and possess around 5,000 frame-level annotations. Compared with our dataset with pixel-level annotations, these benchmarks have a significantly lower quantity of annotations. Regarding the AVS dataset [44-46], the videos in our dataset contain a higher average number of objects about 1.72 objects per video, implying that our scenarios are more challenging with multiple sound sources and multiple semantics. Within such scenarios, our Ref-AVS benchmark is particularly valuable as it demonstrates the ability to effectively focus on objects of real interest. Our dataset also has more uniform video durations and a more refined video selection process than AVS. When it comes to datasets for R-VOS tasks, we maintain a consistent advantage in the number of videos compared to other datasets [8,11,13,20,35]. Moreover, we possess a considerable volume of data encompassing a large number of objects, expressions, and complex scenes." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Overall, our Ref-AVS dataset encompasses a substantial collection of 20,000 expressions and pixel-level annotations spread across 4,000 videos, totaling more than 11 hours. As a result, we believe that this benchmark adequately fulfills the requirements for facilitating research on the Ref-AVS task, while also presenting a significant level of challenge in this domain. We aim to continuously expand the dataset for broader community needs, like training larger foundation models." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.371, + 0.129 + ], + "angle": 0, + "content": "Y. Wang et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.36, + 0.145, + 0.645, + 0.16 + ], + "angle": 0, + "content": "Table 2: Dataset split of Ref-AVS Bench." + }, + { + "type": "table", + "bbox": [ + 0.242, + 0.172, + 0.758, + 0.267 + ], + "angle": 0, + "content": "
Subsetalltrainvaltesttest (seen)test (unseen)test (null)
Category†5239395239131
Video4,0022,908276818292269257
Object6,8885,3665181,004478269257
Expression20,26114,1171,3494,7702,2881,4541,028
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.252, + 0.268, + 0.551, + 0.282 + ], + "angle": 0, + "content": "† The categories here contains \"background\"." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.308, + 0.373, + 0.323 + ], + "angle": 0, + "content": "3.6 Dataset Split" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.33, + 0.787, + 0.42 + ], + "angle": 0, + "content": "As shown in Tab. 2, the complete dataset is divided into three sets, i.e., a training set of 2908 videos, a validation set of 276 videos, and a test set of 818 videos. The videos in the test set and their corresponding annotations undergo a meticulous review and re-annotation process conducted by experienced workers. In order to thoroughly evaluate the models' performance in the Ref-AVS task, the test set is further divided into three distinct subsets, each serving a specific purpose." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.422, + 0.787, + 0.481 + ], + "angle": 0, + "content": "Seen: The first test subset, referred to as the \"seen test set\", comprises object categories that have already appeared in the training set. The purpose of establishing this subset is to evaluate the model's fundamental performance and assess how well it generalizes to familiar object categories." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.482, + 0.787, + 0.602 + ], + "angle": 0, + "content": "Unseen: To address the growing demand for the generalization capabilities of various models in an open-world scenario, an additional test set was created specifically to assess the model's ability to generalize to unseen audio-visual scenes. This \"unseen test set\" consists of object categories that did not appear in the training set, although their super-categories (e.g., animal, vehicle) may have been present in the training set. The purpose of this test set is to evaluate the model's capacity to generalize to novel object categories while leveraging its understanding of broader super-categories." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.602, + 0.787, + 0.739 + ], + "angle": 0, + "content": "Null: A \"null reference problem\" arises when an expression refers to an object that either does not exist or is not visible in the given context. A model that accurately understands expression guidance should not segment any object in scenarios of a null reference [38]. Based on this consideration, we have developed a \"null subset\" to test the model's robustness against such challenges. This subset comprises object categories that were present during training but are paired with expressions that do not correlate, ensuring all objects in the video frames are irrelevant to the given reference expression. Therefore, the true masks for this subset are empty, and the model should refrain from segmenting any object." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.759, + 0.691, + 0.777 + ], + "angle": 0, + "content": "4 Expression Enhancing with Multimodal Cues" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.788, + 0.434, + 0.803 + ], + "angle": 0, + "content": "4.1 Overall Architecture" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.842 + ], + "angle": 0, + "content": "Ref-AVS is designed to locate objects of interest in dynamic audio-visual scenes by utilizing multimodal cues. To solve this challenging task, we propose the" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.672, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "Ref-AVS" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "image", + "bbox": [ + 0.221, + 0.146, + 0.784, + 0.398 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.412, + 0.788, + 0.496 + ], + "angle": 0, + "content": "Fig. 3: The overall architecture of our method EEMC. We utilize cached memory to preserve temporal information, enabling the model to capture temporal mutations. Modality encoding provides a more clear context for the multimodal cue features \\( Q_{m} \\), using the modality identify tokens [aud] and [vis]. Lastly, we achieve efficient prompting of the visual foundation model by employing cross-attention, where mask queries \\( q \\) serve as the target and multimodal cue features \\( Q_{m} \\) act as the source." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.524, + 0.788, + 0.631 + ], + "angle": 0, + "content": "Expression Enhancing with Multimodal Cues (EEMC) method, which focuses on integrating multimodal information in the dynamic audio-visual scenes into reference expressions with the corresponding multimodal cues. This approach enables the formation of a comprehensive multimodal reference feature. Additionally, we employ an attention mechanism to utilize the multimodal reference cues as a prompt for the visual foundation model, thereby facilitating the final segmentation process." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.651, + 0.498, + 0.667 + ], + "angle": 0, + "content": "4.2 Multimodal Representations" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.788, + 0.748 + ], + "angle": 0, + "content": "Audio: Similar to the video processing, we divide the audio input into clips at 1-second intervals. Audio representations \\( F_{A} \\in \\mathbb{R}^{t \\times d_{A}} \\) is encoded from VGGish [14,16], where \\( t \\) represents the duration of the audio in seconds and matches the number of video frames. The audio representations are extracted offline and the audio encoder is not fine-tuned." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.795 + ], + "angle": 0, + "content": "Visual: We sample \\(t\\) frames from the video input at 1-second intervals and extract visual representations \\(F_{V} \\in \\mathbb{R}^{t \\times d_{V} \\times H \\times W}\\) using a pre-trained Swin-base model [10]. The visual encoder is not fine-tuned." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Expression: We use RoBERTa [7, 27] as our text encoder to extract textual expression features \\( F_{T} \\in \\mathbb{R}^{L \\times d_{T}} \\), where \\( L \\) denotes the length of the expression. The expression representations are obtained off-the-shelf without fine-tuning." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.371, + 0.128 + ], + "angle": 0, + "content": "Y. Wang et al." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.147, + 0.53, + 0.162 + ], + "angle": 0, + "content": "4.3 Temporal Bi-modal Transformer" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.173, + 0.788, + 0.295 + ], + "angle": 0, + "content": "Temporal A-T & V-T Fusion: This module is introduced to retrieve the expression-related information of each modality. Firstly, for the convenience of the following multimodal fusion, we prepare the projected audio feature \\( F_{A'} \\in \\mathbb{R}^{t \\times 1 \\times d_V} \\), downsampled and flattened visual feature \\( F_{V'} \\in \\mathbb{R}^{t \\times (\\frac{W}{8} \\times \\frac{H}{8}) \\times d_V} \\), and projected text feature \\( F_{T'} \\in \\mathbb{R}^{t \\times L \\times d_V} \\) which is expanded along the temporal dimension. Subsequently, the features are combined to yield the expression-related multimodal features \\( \\hat{F}_A \\), \\( \\hat{F}_V \\) and modality-aware expression features \\( \\hat{F}_{T_A} \\), \\( \\hat{F}_{T_V} \\), with the same dimension as \\( F_{A'} \\), \\( F_{V'} \\), \\( F_{T'} \\), respectively:" + }, + { + "type": "equation", + "bbox": [ + 0.329, + 0.309, + 0.786, + 0.327 + ], + "angle": 0, + "content": "\\[\n\\hat {F} _ {M}, \\hat {F} _ {T _ {M}} = M F (C o n c a t \\left(F _ {M ^ {\\prime}}; F _ {T ^ {\\prime}}\\right)), M \\in \\{A, V \\} \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.333, + 0.786, + 0.363 + ], + "angle": 0, + "content": "where \\( \\text{Concat}(\\cdot) \\) denotes the concatenate operation, and \\( MF(\\cdot) \\) denotes the modality fusion function, which is employed as self-attention." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.364, + 0.788, + 0.484 + ], + "angle": 0, + "content": "Cached Memory: Although the aforementioned fusion method can combine cross-modal information, it does not give much emphasis to strong temporal dynamics. To address this issue, we introduce a straightforward cached memory \\(\\mathcal{C}_A\\in \\mathbb{R}^{t\\times 2\\times d_T},\\mathcal{C}_V\\in \\mathbb{R}^{t\\times (W\\times H + 1)\\times d_T}\\) to explicitly capture the temporal mutations as they occur. The cache memory is required to store the mean modality features in the temporal domain from the beginning to the current moment. Then, the multimodal cues feature \\(Q_{A}\\) and \\(Q_{V}\\) output from temporal-aware transformers can be calculated as" + }, + { + "type": "equation", + "bbox": [ + 0.364, + 0.511, + 0.786, + 0.53 + ], + "angle": 0, + "content": "\\[\nQ _ {M} ^ {\\tau} = (\\beta + 1) \\hat {F} _ {M} ^ {\\tau} - \\beta \\mathcal {C} _ {M} ^ {\\tau}, M \\in \\{A, V \\}, \\tag {2}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.367, + 0.532, + 0.786, + 0.572 + ], + "angle": 0, + "content": "\\[\n\\mathcal {C} _ {M} ^ {\\tau} = \\left\\{ \\begin{array}{l l} 0, & \\text {i f} \\tau = 1, \\\\ \\frac {1}{\\tau - 1} \\sum_ {i = 0} ^ {\\tau - 1} \\hat {F} _ {M} ^ {i}, & \\text {i f} \\tau > 1, \\end{array} \\right. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.582, + 0.789, + 0.841 + ], + "angle": 0, + "content": "while \\(\\tau\\) is the time stamp of the specific feature at the temporal dimension and \\(\\beta\\) is a hyper-parameter to control the sensitivity to temporal dynamics. When the current audio or visual features do not change significantly compared to the mean of past features (i.e., \\(\\hat{F}_M^\\tau = \\mathcal{C}_M^\\tau\\)), \\(Q_M^\\tau\\) remains nearly unchanged. However, in cases where the current change is drastic (i.e., \\(\\hat{F}_M^\\tau \\neq \\mathcal{C}_M^\\tau\\)), the cached memory can amplify the difference in the current feature, leading to an output with salient features. It is important to note that text cue feature \\(Q_T\\) inherently possesses a highly abstracted semantic nature, which means manipulation may not be necessary at this stage. Therefore, at this stage, we combine and average the original textual features with the text features enhanced by multimodal information, resulting in a comprehensive enhanced text feature \\(Q_T = mean(F_{T'} + \\hat{F}_{T_A} + \\hat{F}_{T_V})\\). Finally, we obtain the processed cue feature of each modality \\(Q_A\\), \\(Q_V\\) and \\(Q_T\\). Modality Encoding: The modality encoding approach involves incorporating modality identity tokens \\([aud] \\in \\mathbb{R}^{1 \\times d_V}\\) for audio modality and \\([vis] \\in \\mathbb{R}^{1 \\times d_V}\\) for visual modality into the following multimodal cues integration process. Since only two tokens are required to split the sequence into three segments, we only need to manipulate the audio and visual cue features." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.671, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "Ref-AVS" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.147, + 0.544, + 0.162 + ], + "angle": 0, + "content": "4.4 Prompting with Multimodal Cues" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.177, + 0.788, + 0.238 + ], + "angle": 0, + "content": "In this Prompting with Multimodal Cues (PMC) phrase, when we obtain the final multimodal cues features, we can concatenate the multimodal cues together and then apply self-attention to obtain comprehensive multimodal cues \\( Q_{m} \\in \\mathbb{R}^{t \\times (H \\times W + L + 3) \\times d_V} \\) through the Multimodal Integration Transformer:" + }, + { + "type": "equation", + "bbox": [ + 0.335, + 0.253, + 0.786, + 0.269 + ], + "angle": 0, + "content": "\\[\nQ _ {m} = M F (C o n c a t ([ Q _ {A}; [ a u d ]; Q _ {V}; [ v i s ]; Q _ {T} ])), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.281, + 0.789, + 0.343 + ], + "angle": 0, + "content": "where \\(MF\\) is the multimodal fusion function serving as the self-attention for multimodal cues interaction. Then we utilize these comprehensive multinational cues to prompt the learnable mask queries \\(q \\in \\mathbb{R}^{N \\times d_V}\\) in the transformer-based segmentation decoder with cross-attention, where \\(N\\) is the number of mask queries:" + }, + { + "type": "equation", + "bbox": [ + 0.327, + 0.363, + 0.786, + 0.377 + ], + "angle": 0, + "content": "\\[\nq _ {Q} = C A T F (q u e r y = q, k e y = Q _ {m}, v a l u e = Q _ {m}), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.386, + 0.788, + 0.418 + ], + "angle": 0, + "content": "where \\( q_{Q} \\in \\mathbb{R}^{N \\times d_{V}} \\) is the updated mask queries. \\( CATF(\\cdot) \\) is the cross-attention transformer with \\( q \\) and \\( Q_{m} \\) as the information target and source, respectively." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.445, + 0.377, + 0.463 + ], + "angle": 0, + "content": "5 Experiments" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.481, + 0.457, + 0.496 + ], + "angle": 0, + "content": "5.1 Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.511, + 0.789, + 0.68 + ], + "angle": 0, + "content": "Mask2Former [5] serves as our visual foundation model to provide the commonly used transformer-based segmentation decoder. By default, in this work, all input video frames are scaled to \\(384 \\times 384\\). The shape of the visual features is \\([H = 64, W = 64, d_V = 256]\\), and we employ 8-fold downsampling to reduce the computation costs. The audio features with \\(d_A = 128\\) are extracted from the mono-channel waveform. The text features have a shape of \\([L = 25, d_T = 768]\\). For convenience, we map all modality feature dimensions to \\(d_V\\). Hyper-parameter \\(\\beta\\) is set to 1 by default. Transformer layers \\(N_L\\) of the Temporal Bi-modal Transformer, Multimodal Integration Transformer and the cross-attention transformer CATF are set to 4 by default. The number of mask queries \\(N_q\\) is fixed to 100." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.705, + 0.421, + 0.719 + ], + "angle": 0, + "content": "5.2 Evaluation Metrics" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.734, + 0.789, + 0.841 + ], + "angle": 0, + "content": "To conduct a comprehensive evaluation of our Ref-AVS method, we employ the Jaccard index \\((\\mathcal{I})\\) and F-score \\((\\mathcal{F})\\) as performance metrics. Additionally, we introduce \\(S\\) in the null reference test set to assess the efficacy of expression guidance in the model. \\(S\\) denotes the square root of the ratio between the predicted mask area and the background area, a higher value indicates a larger proportion of the mask relative to the background area, suggesting a lack of precise expression guidance provided to the model." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.371, + 0.129 + ], + "angle": 0, + "content": "Y. Wang et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.145, + 0.788, + 0.201 + ], + "angle": 0, + "content": "Table 3: Performance on Ref-AVS Bench. We use three test subsets to evaluate the comprehensive ability of Ref-AVS methods. Mix is the average performance of Seen and Unseen test set. We also use Null test set to evaluate the robustness of multimodal-cue expression guidance." + }, + { + "type": "table", + "bbox": [ + 0.221, + 0.213, + 0.785, + 0.398 + ], + "angle": 0, + "content": "
MethodTaskSeenUnseenMix(S+U)Null S (↓)Pub.
J(%)FJ(%)FJ(%)F
AVSbench [46] +textAVS21.220.45727.670.50924.450.483-ECCV'2022
23.200.51132.360.54727.780.5290.208
AVSegFormer [12] +textAVS29.160.42332.710.45130.940.437-AAAI'2024
33.470.47036.050.50134.760.4860.171
GAVS [37] +textAVS24.850.45527.860.46026.360.457-AAAI'2024
28.930.49829.820.49729.380.4980.190
ReferFormer [39] +audioRef-VOS27.460.47724.940.50426.20.491-CVPR'2022
31.310.50130.400.48830.860.4950.176
R2VOS [23] +audioRef-VOS21.340.30923.910.43122.630.370-ICCV'2023
25.010.41027.930.49826.470.4540.183
EEMCRef-AVS34.200.51349.540.64841.870.5810.007ECCV'2024
" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.433, + 0.433, + 0.448 + ], + "angle": 0, + "content": "5.3 Quantitative Results" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.474, + 0.788, + 0.655 + ], + "angle": 0, + "content": "We compare our Ref-AVS method with existing approaches in the relevant field using our Ref-AVS benchmark. The results from the seen test set demonstrate the superior performance of our Ref-AVS method on this dataset, outperforming other methods by a significant margin. To ensure fairness, we equip the other methods with the same inputs from both the audio and visual modalities as ours. However, even with this additional information, these methods still fail to achieve satisfactory results. The result indicates that simple modal fusion is inadequate for addressing the challenges of understanding multimodal cues within the Ref-AVS task. Instead of directly incorporating audiovisual information, our approach selects textual representations as carriers of multimodal information because they contain rich semantics and cues that are contextually relevant to the current audiovisual environment." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.659, + 0.788, + 0.841 + ], + "angle": 0, + "content": "We conduct tests on the unseen test set and the null reference test set to explore the generalization and multimodal cues-following ability. Our method still maintains a leading position in the unseen test set because we leverage text with highly abstract semantic capabilities as a multimodal information carrier, instead of directly fusing audio and visual information. Therefore, the obtained multimodal cues can provide more robust semantic guidance. On the null test set, we obtain leading results among all other methods, indicating that our method can perceive multimodal cues to a reasonably accurate degree. Compared to other transformer-based and query-based frameworks, AVSBench, as a classic AVS method, obtains inferior results. This may be attributed to the model simply fuses multimodal cues and visual features in a progressive fashion, resulting in relatively weak guidance over the visual space." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.672, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "Ref-AVS" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "image", + "bbox": [ + 0.221, + 0.143, + 0.52, + 0.411 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.535, + 0.143, + 0.788, + 0.411 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.421, + 0.788, + 0.477 + ], + "angle": 0, + "content": "Fig. 4: Segmentation masks of the referred objects in the Ref-AVS Bench. Note, for \\(\\bullet \\bullet \\bullet\\), the sound volume increases successively from silent to loud. We compare our methodology with AVSegformer and ReferFormer. More cases can be found in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.507, + 0.421, + 0.523 + ], + "angle": 0, + "content": "5.4 Qualitative Results" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.533, + 0.788, + 0.731 + ], + "angle": 0, + "content": "We visualize the segmentation masks on the test set of the Ref-AVS Bench and compare them with AVSegFormer and ReferFormer, as shown in Fig. 4. From these qualitative results, we observe that both ReferFormer in the Ref-VOS task and AVSegFormer in the AVS task fail to accurately segment the object described in the expression. Specifically, AVSegformer struggles to fully comprehend the expression and tends to directly generate the sound source. This issue is exemplified in the bottom-left sample, where AVSegformer erroneously segments the vacuum cleaner instead of the boy. On the other hand, Ref-VOS may not adequately understand the audio-visual scene and thus mistakenly identify the toddler as the piano player, as shown in the top-right sample. In contrast, our Ref-AVS method demonstrates a superior capability to simultaneously process multi-modal expressions and scenes, enabling it to accurately interpret the user instruction and segment the intended object of interest." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.754, + 0.391, + 0.77 + ], + "angle": 0, + "content": "5.5 Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.841 + ], + "angle": 0, + "content": "We conduct ablation studies to investigate the impact of two modalities (audio and text) information on the Ref-AVS task, as well as the effectiveness of the proposed method. As shown in the Tab. 4, we observe that in setting \\(②\\), removing the text information results in a significant performance degradation of \\(11.44\\%\\) in" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.371, + 0.129 + ], + "angle": 0, + "content": "Y. Wang et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.296, + 0.145, + 0.704, + 0.159 + ], + "angle": 0, + "content": "Table 4: Ablation study on modalities and method modules." + }, + { + "type": "table", + "bbox": [ + 0.255, + 0.171, + 0.751, + 0.27 + ], + "angle": 0, + "content": "
SettingSeenUnseenMixNull
J(%)FJ(%)FJ(%)FS(↓)
① EEMC34.200.51349.540.64841.870.5810.007
② (-) Text23.670.46837.190.64130.430.5550.011
③ (-) Audio27.270.49243.640.65635.460.5740.010
④ (-) Cached Memory25.300.49436.730.66031.050.5770.009
⑤ (-) Modality Encoding27.200.47044.970.63736.090.5540.009
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.307, + 0.788, + 0.488 + ], + "angle": 0, + "content": "\\(\\mathcal{I}\\) and \\(2.60\\%\\) in \\(\\mathcal{F}\\). This degradation is much higher compared to setting ③, where the removal of the audio information results in a combined mixed drop of \\(6.41\\%\\) in \\(\\mathcal{I}\\) and \\(0.70\\%\\) in \\(\\mathcal{F}\\). This phenomenon primarily results from the clarity and directness of textual information as a reference source. In contrast, when relying solely on audio information, the model often overlooks the referential content and instead focuses on identifying objects visually associated with audible behaviors. We also conduct experiments to examine the function of each module in this framework. The cached memory is designed to track significant changes in the temporal domain. Meanwhile, the purpose of modality encoding is to extract more distinct and thorough features from multimodal cues, enhancing modal perception. Results from settings ④ and ⑤ in Tab. 4 demonstrate the effectiveness of these modules." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.508, + 0.36, + 0.525 + ], + "angle": 0, + "content": "6 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.536, + 0.788, + 0.749 + ], + "angle": 0, + "content": "In this work, we introduce a novel task, Ref-AVS, which requires the machine to segment objects of interest in dynamic audio-visual scenes based on expressions that incorporate multimodal cues and temporal information. To support this field, we develop the first benchmark, Ref-AVS Bench, for performance training and evaluation. Our method, EEMC, establishes a robust baseline with an advanced multimodal cues fusion module. We compare our method with several existing approaches on the Ref-AVS Bench dataset and demonstrate its promising performance in accurately locating objects using multimodal reference expressions. Essentially, Ref-AVS leverages audio, visual, and textual data to enrich real-world scene comprehension. Our task has potential in applications such as attention focusing and object editing in immersive audio-visual scenarios like extended reality, thereby enhancing user interaction experiences. In the future, we plan to further expand the scale of the current dataset to meet the growing data demands in the era of large language models." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.767, + 0.405, + 0.785 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.787, + 0.842 + ], + "angle": 0, + "content": "This research was supported by National Natural Science Foundation of China (NO.62106272), and Public Computing Cloud, Renmin University of China. This research was partially supported by Shanghai Artificial Intelligence Laboratory." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.672, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "Ref-AVS" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.323, + 0.16 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.176, + 0.785, + 0.218 + ], + "angle": 0, + "content": "1. Botach, A., Zheltonozhskii, E., Baskin, C.: End-to-end referring video object segmentation with multimodal transformers. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4985-4995 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.219, + 0.785, + 0.259 + ], + "angle": 0, + "content": "2. Chen, H., Xie, W., Afouras, T., Nagrani, A., Vedaldi, A., Zisserman, A.: Localizing visual sounds the hard way. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16867-16876 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.26, + 0.785, + 0.301 + ], + "angle": 0, + "content": "3. Chen, H., Xie, W., Vedaldi, A., Zisserman, A.: Vggsound: A large-scale audiovisual dataset. In: ICASSP 2020-2020 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). pp. 721-725. IEEE (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.302, + 0.785, + 0.342 + ], + "angle": 0, + "content": "4. Chen, Z., Wang, L., Wang, P., Gao, P.: Question-aware global-local video understanding network for audio-visual question answering. IEEE Transactions on Circuits and Systems for Video Technology (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.343, + 0.785, + 0.37 + ], + "angle": 0, + "content": "5. Cheng, B., Choudhuri, A., Misra, I., Kirillov, A., Girdhar, R., Schwing, A.G.: Mask2former for video instance segmentation (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.371, + 0.785, + 0.411 + ], + "angle": 0, + "content": "6. Cheng, H.K., Oh, S.W., Price, B., Schwing, A., Lee, J.Y.: Tracking anything with decoupled video segmentation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 1316-1326 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.412, + 0.785, + 0.452 + ], + "angle": 0, + "content": "7. Devlin, J., Chang, M.W., Lee, K., Toutanova, K.: Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.453, + 0.785, + 0.508 + ], + "angle": 0, + "content": "8. Ding, H., Liu, C., He, S., Jiang, X., Loy, C.C.: Mevis: A large-scale benchmark for video segmentation with motion expressions. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 2694-2703 (October 2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.509, + 0.785, + 0.55 + ], + "angle": 0, + "content": "9. Ding, H., Liu, C., He, S., Jiang, X., Torr, P.H., Bai, S.: Mose: A new dataset for video object segmentation in complex scenes. arXiv preprint arXiv:2302.01872 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.551, + 0.785, + 0.605 + ], + "angle": 0, + "content": "10. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.606, + 0.785, + 0.646 + ], + "angle": 0, + "content": "1. Dutta, U.K., Harandi, M., Sekhar, C.C.: Unsupervised deep metric learning via orthogonality based probabilistic loss. IEEE Transactions on Artificial Intelligence 1(1), 74-84 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.647, + 0.785, + 0.674 + ], + "angle": 0, + "content": "2. Gao, S., Chen, Z., Chen, G., Wang, W., Lu, T.: Avsegformer: Audio-visual segmentation with transformer (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.675, + 0.785, + 0.715 + ], + "angle": 0, + "content": "3. Gavrilyuk, K., Ghodrati, A., Li, Z., Snoek, C.G.: Actor and action video segmentation from a sentence. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 5958-5966 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.716, + 0.785, + 0.77 + ], + "angle": 0, + "content": "4. Gemmeke, J.F., Ellis, D.P., Freedman, D., Jansen, A., Lawrence, W., Moore, R.C., Plakal, M., Ritter, M.: Audio set: An ontology and human-labeled dataset for audio events. In: 2017 IEEE international conference on acoustics, speech and signal processing (ICASSP). pp. 776-780. IEEE (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.772, + 0.785, + 0.812 + ], + "angle": 0, + "content": "5. Guangyao li, Jake Wei, Y.T.C.X.J.R.W.D.H.: Learning to answer questions in dynamic audio-visual scenarios. IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.813, + 0.785, + 0.84 + ], + "angle": 0, + "content": "6. Hershey, S., Chaudhuri, S., Ellis, D.P., Gemmeke, J.F., Jansen, A., Moore, R.C., Plakal, M., Platt, D., Saurous, R.A., Seybold, B., et al.: Cnn architectures for" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.176, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.371, + 0.128 + ], + "angle": 0, + "content": "Y. Wang et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.241, + 0.148, + 0.785, + 0.175 + ], + "angle": 0, + "content": "large-scale audio classification. In: 2017 IEEE international conference on acoustics, speech and signal processing (ICASSP). pp. 131-135. IEEE (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.176, + 0.786, + 0.217 + ], + "angle": 0, + "content": "17. Hu, D., Qian, R., Jiang, M., Tan, X., Wen, S., Ding, E., Lin, W., Dou, D.: Discriminative sounding objects localization via self-supervised audiovisual matching. Advances in Neural Information Processing Systems 33, 10077-10087 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.217, + 0.786, + 0.243 + ], + "angle": 0, + "content": "18. Huang, S., Li, H., Wang, Y., Zhu, H., Dai, J., Han, J., Rong, W., Liu, S.: Discovering sounding objects by audio queries for audio visual segmentation (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.243, + 0.786, + 0.27 + ], + "angle": 0, + "content": "19. Jiang, Y., Yin, J.: Target-aware spatio-temporal reasoning via answering questions in dynamics audio-visual scenarios. arXiv preprint arXiv:2305.12397 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.27, + 0.786, + 0.325 + ], + "angle": 0, + "content": "20. Khoreva, A., Rohrbach, A., Schiele, B.: Video object segmentation with language referring expressions. In: Computer Vision-ACCV 2018: 14th Asian Conference on Computer Vision, Perth, Australia, December 2-6, 2018, Revised Selected Papers, Part IV 14. pp. 123-141. Springer (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.325, + 0.786, + 0.392 + ], + "angle": 0, + "content": "21. Li, G., Hou, W., Hu, D.: Progressive spatio-temporal perception for audio-visual question answering. In: Proceedings of the 31st ACM International Conference on Multimedia. p. 7808-7816. MM '23, Association for Computing Machinery, New York, NY, USA (2023). https://doi.org/10.1145/3581783.3612293, https://doi.org/10.1145/3581783.3612293" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.392, + 0.786, + 0.434 + ], + "angle": 0, + "content": "22. Li, K., Yang, Z., Chen, L., Yang, Y., Xun, J.: Catr: Combinatorial-dependence audio-queried transformer for audio-visual video segmentation. arXiv preprint arXiv:2309.09709 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.434, + 0.786, + 0.474 + ], + "angle": 0, + "content": "23. Li, X., Wang, J., Xu, X., Li, X., Raj, B., Lu, Y.: Robust referring video object segmentation with cyclic structural consensus. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 22236-22245 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.474, + 0.786, + 0.501 + ], + "angle": 0, + "content": "24. Ling, Y., Li, Y., Gan, Z., Zhang, J., Chi, M., Wang, Y.: Hear to segment: Unmixing the audio to guide the semantic segmentation (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.501, + 0.786, + 0.528 + ], + "angle": 0, + "content": "25. Liu, J., Ju, C., Ma, C., Wang, Y., Wang, Y., Zhang, Y.: Audio-aware query-enhanced transformer for audio-visual segmentation (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.528, + 0.786, + 0.555 + ], + "angle": 0, + "content": "26. Liu, J., Wang, Y., Ju, C., Zhang, Y., Xie, W.: Annotation-free audio-visual segmentation. arXiv preprint arXiv:2305.11019 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.555, + 0.786, + 0.596 + ], + "angle": 0, + "content": "27. Liu, Y., Ott, M., Goyal, N., Du, J., Joshi, M., Chen, D., Levy, O., Lewis, M., Zettlemoyer, L., Stoyanov, V.: Roberta: A robustly optimized bert pretraining approach. arXiv preprint arXiv:1907.11692 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.596, + 0.786, + 0.637 + ], + "angle": 0, + "content": "28. Ma, J., Sun, P., Wang, Y., Hu, D.: Stepping stones: A progressive training strategy for audio-visual semantic segmentation. IEEE European Conference on Computer Vision (ECCV) (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.637, + 0.786, + 0.663 + ], + "angle": 0, + "content": "29. Mo, S., Tian, Y.: Av-sam: Segment anything model meets audio-visual localization and segmentation. arXiv preprint arXiv:2305.01836 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.663, + 0.786, + 0.703 + ], + "angle": 0, + "content": "30. Perazzi, F., Khoreva, A., Benenson, R., Schiele, B., Sorkine-Hornung, A.: Learning video object segmentation from static images. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 2663-2672 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.703, + 0.786, + 0.759 + ], + "angle": 0, + "content": "31. Perazzi, F., Pont-Tuset, J., McWilliams, B., Van Gool, L., Gross, M., Sorkine-Hornung, A.: A benchmark dataset and evaluation methodology for video object segmentation. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 724-732 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.759, + 0.786, + 0.8 + ], + "angle": 0, + "content": "32. Pont-Tuset, J., Perazzi, F., Caelles, S., Arbeláez, P., Sorkine-Hornung, A., Van Gool, L.: The 2017 davis challenge on video object segmentation. arXiv preprint arXiv:1704.00675 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.8, + 0.786, + 0.84 + ], + "angle": 0, + "content": "33. Ren, T., Liu, S., Zeng, A., Lin, J., Li, K., Cao, H., Chen, J., Huang, X., Chen, Y., Yan, F., Zeng, Z., Zhang, H., Li, F., Yang, J., Li, H., Jiang, Q., Zhang, L.: Grounded sam: Assembling open-world models for diverse visual tasks (2024)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.148, + 0.786, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.672, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "Ref-AVS" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.189 + ], + "angle": 0, + "content": "34. Senocak, A., Oh, T.H., Kim, J., Yang, M.H., So Kweon, I.: Learning to localize sound source in visual scenes. In: The IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (June 2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.19, + 0.788, + 0.245 + ], + "angle": 0, + "content": "35. Seo, S., Lee, J.Y., Han, B.: Urvos: Unified referring video object segmentation network with a large-scale benchmark. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XV 16. pp. 208-223. Springer (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.246, + 0.786, + 0.286 + ], + "angle": 0, + "content": "36. Tian, Y., Shi, J., Li, B., Duan, Z., Xu, C.: Audio-visual event localization in unconstrained videos. In: Proceedings of the European Conference on Computer Vision (ECCV) (September 2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.288, + 0.787, + 0.327 + ], + "angle": 0, + "content": "37. Wang, Y., Liu, W., Li, G., Ding, J., Hu, D., Li, X.: Prompting segmentation with sound is generalizable audio-visual source localizer. arXiv preprint arXiv:2309.07929 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.328, + 0.787, + 0.37 + ], + "angle": 0, + "content": "38. Wang, Y., Sun, P., Li, Y., Zhang, H., Hu, D.: Can textual semantics mitigate sounding object segmentation preference? IEEE European Conference on Computer Vision (ECCV) (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.371, + 0.787, + 0.397 + ], + "angle": 0, + "content": "39. Wu, J., Jiang, Y., Sun, P., Yuan, Z., Luo, P.: Language as queries for referring video object segmentation (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.398, + 0.787, + 0.438 + ], + "angle": 0, + "content": "40. Xu, N., Yang, L., Fan, Y., Yue, D., Liang, Y., Yang, J., Huang, T.: Youtube-vos: A large-scale video object segmentation benchmark. arXiv preprint arXiv:1809.03327 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.439, + 0.787, + 0.481 + ], + "angle": 0, + "content": "41. Yan, S., Zhang, R., Guo, Z., Chen, W., Zhang, W., Li, H., Qiao, Y., He, Z., Gao, P.: Referred by multi-modality: A unified temporal transformer for video object segmentation (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.481, + 0.787, + 0.522 + ], + "angle": 0, + "content": "42. Yang, P., Wang, X., Duan, X., Chen, H., Hou, R., Jin, C., Zhu, W.: Avqa: A dataset for audio-visual question answering on videos. In: Proceedings of the 30th ACM International Conference on Multimedia. pp. 3480-3491 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.522, + 0.787, + 0.577 + ], + "angle": 0, + "content": "43. Yu, L., Lin, Z., Shen, X., Yang, J., Lu, X., Bansal, M., Berg, T.L.: Mattnet: Modular attention network for referring expression comprehension. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1307-1315 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.578, + 0.787, + 0.618 + ], + "angle": 0, + "content": "44. Yuanhong, C., Yuyuan, L., Hu, W., Fengbei, L., Chong, W., Gustavo, C.: A closer look at audio-visual semantic segmentation. arXiv preprint arXiv:2304.02970 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.619, + 0.787, + 0.66 + ], + "angle": 0, + "content": "45. Zhou, J., Shen, X., Wang, J., Zhang, J., Sun, W., Zhang, J., Birchfield, S., Guo, D., Kong, L., Wang, M., Zhong, Y.: Audio-visual segmentation with semantics. arXiv preprint arXiv:2301.13190 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.661, + 0.787, + 0.702 + ], + "angle": 0, + "content": "46. Zhou, J., Wang, J., Zhang, J., Sun, W., Zhang, J., Birchfield, S., Guo, D., Kong, L., Wang, M., Zhong, Y.: Audio-visual segmentation. In: European Conference on Computer Vision (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.702, + 0.787, + 0.743 + ], + "angle": 0, + "content": "47. Zhou, X., Zhou, D., Hu, D., Zhou, H., Ouyang, W.: Exploiting visual context semantics for sound source localization. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 5199-5208 (2023)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.743 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/Ref-AVS_ Refer and Segment Objects in Audio-Visual Scenes/a65e28a6-a79d-4af3-846c-76b29927236e_origin.pdf b/2024/Ref-AVS_ Refer and Segment Objects in Audio-Visual Scenes/a65e28a6-a79d-4af3-846c-76b29927236e_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..ec68a7b9d8a565b8873b0d0b5a0e920f4b969ecd --- /dev/null +++ b/2024/Ref-AVS_ Refer and Segment Objects in Audio-Visual Scenes/a65e28a6-a79d-4af3-846c-76b29927236e_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:07033ac5515032bcf98d793b5d5b3c59bb7ad29ed2754d154f0c74e30c9cf9c3 +size 10180188 diff --git a/2024/Ref-AVS_ Refer and Segment Objects in Audio-Visual Scenes/full.md b/2024/Ref-AVS_ Refer and Segment Objects in Audio-Visual Scenes/full.md new file mode 100644 index 0000000000000000000000000000000000000000..2ea9cf58a021d29b348e84a6875dfd68dc3b7dba --- /dev/null +++ b/2024/Ref-AVS_ Refer and Segment Objects in Audio-Visual Scenes/full.md @@ -0,0 +1,298 @@ +# Ref-AVS: Refer and Segment Objects in Audio-Visual Scenes + +Yaoting Wang $^{1\dagger}$ , Peiwen Sun $^{2\dagger}$ , Dongzhan Zhou $^{3\dagger}$ , Guangyao Li $^{1}$ , Honggang Zhang $^{2}$ , and Di Hu $^{\boxtimes 1,4}$ + +1 Gaoling School of Artificial Intelligence, Renmin University of China, China +yaoting.wang@outlook.com +{guangyaoli, dihu}@ruc.edu.cn + +$^{2}$ Beijing University of Posts and Telecommunications, Beijing, China {sunpeiwen,zhhg}@bupt.edu.cn + +3 Shanghai Artificial Intelligence Laboratory, Shanghai, China zhoudongzhan@pjlab.org.cn + +4 Engineering Research Center of Next-Generation Search and Recommendation + +Abstract. Traditional reference segmentation tasks have predominantly focused on silent visual scenes, neglecting the integral role of multimodal perception and interaction in human experiences. In this work, we introduce a novel task called Reference Audio-Visual Segmentation (Ref-AVS), which seeks to segment objects within the visual domain based on expressions containing multimodal cues. Such expressions are articulated in natural language forms but are enriched with multimodal cues, including audio and visual descriptions. To facilitate this research, we construct the first Ref-AVS benchmark, which provides pixel-level annotations for objects described in corresponding multimodal-cue expressions. To tackle the Ref-AVS task, we propose a new method that adequately utilizes multimodal cues to offer precise segmentation guidance. Finally, we conduct quantitative and qualitative experiments on three test subsets to compare our approach with existing methods from related tasks. The results demonstrate the effectiveness of our method, highlighting its capability to precisely segment objects using multimodal-cue expressions. Dataset is available at https://gewu-lab.github.io/Ref-AVS. + +Keywords: Referring Audio-Visual Segmentation $\cdot$ Audio-Visual Segmentation $\cdot$ Multimodal Learning + +# 1 Introduction + +In the real world, visual scenes are often accompanied by diverse multimodal information1, including audio and text modalities. This additional multimodal + +![](images/a67defb349bfe0b763649657c54f8dd938206205b45a2524bf2bbaaf0657a43a.jpg) + +![](images/cd9cec3a02c62acd3090fd1bad3797227fee6a27567cfa90a67e3f50386d30e2.jpg) + +![](images/ec91475b65546273e18449d9e230e94cb936a1cba736fbc33d20772d544e4613.jpg) +Fig. 1: Comparison of the Ref-AVS task with other related tasks. Ref-AVS challenges machines to locate objects of interest in the visual space using multimodal cues, just like humans do in the real world. + +![](images/fb5bb9a31be2cd4ad32118cdca9b00a7deb3ce89104918e35cc46f4af2e37afc.jpg) + +![](images/7523bf8a058148df62ae39bf991715d14ea28b372cd5d773bb365adfa0109c45.jpg) + +![](images/475650cdc4a48b1e34e9ee625e8afc84651bf7755e4c0c2aef3f11bb54be7c54.jpg) + +Which one should the camera focus? + +An ideal machine needs to know always going for the player who is actually playing. + +In the view of camera. + +How to focus the player who is actually playing. + +![](images/bcd12c67cc482a48575faaca83c62933c2394a9d1f8649d825a00f089b3bde81.jpg) + +information offers valuable spatiotemporal and semantic cues, aiding individuals in locating objects of interest. Similarly, the ability to locate objects based on multimodal cues plays a crucial role in the functioning of machines. For example, stage cameras should be able to locate the musician playing instruments, while movie cameras should locate the superhero in front of the yelling villain. In the above scenarios, using a single modality for segmentation alone cannot achieve the desired behavior. + +To date, researchers have studied the reference segmentation problem from somewhat limited scenarios. As shown in the coordinate system in Fig. 1, there are currently three mainstream developments in segmentation methods for different modalities. Firstly, from the perspective of visual cues, Video Object Segmentation (VOS) [9,30-32,40] emerges with annotated first frame mask as the reference, guiding the segmentation of specific objects in subsequent video frames. While VOS can achieve arbitrary object segmentation in videos, it heavily relies on precise annotation of the first frame, which can be challenging and time-consuming in practice. Secondly, from the perspective of text cues, Referring Video Object Segmentation (R-VOS) [8,11,13,20] explores the segmentation ability by using attribute descriptions as guidance. As R-VOS has successfully replaced the mask annotation used in VOS with natural language, it provides a more accessible and user-friendly reference form but the capability to locate objects in more natural dynamic audio-visual scenes is still limited. In contrast, from the perspective of audio cues, Audio-Visual Segmentation (AVS) [26,37,44-47] leverages audio as a guide to segment sounding objects. This approach effectively addresses the challenge of locating objects in dynamic audio-visual scenes. However, AVS still faces limitations in its ability to focus on general objects that do not generate sound and cannot effectively locate objects of specific interest. + +In simple terms, current work falls short of enabling machines to locate objects of interest in natural dynamic audio-visual scenes. For instance, as illus + +trated in Fig. 1, how can machines accurately locate a person who is genuinely playing a musical instrument over time? This requires machines to infer which instrument is making a sound and who is playing that instrument. Though Yan et al. [41] utilize either audio or language modality for multimodal reference segmentation, their method does not integrate multimodal information simultaneously nor support the aforementioned scenario. Therefore, introducing a task to explore the possibility of locating interested objects in natural dynamic audio-visual scenes holds significant practical potential. + +In this paper, we propose a pixel-level segmentation task called Referring Audio-Visual Segmentation (Ref-AVS), which requires the network to densely predict whether each pixel corresponds to the given multimodal-cue expression, including dynamic audio-visual information. Top-left of Fig. 1 highlights the distinctions between Ref-AVS and previous tasks. The Ref-AVS task poses a greater challenge as it requires the network to accurately locate and segment objects in a more complex and stereoscopic modality space. To this end, a computation model with comprehensive multimodal understanding capabilities is necessary. To foster research advancements in this field, we introduce the Ref-AVS Bench, the pioneering benchmark that focuses on locating and segmenting objects of interest using referring multimodal-cue expressions. Considering the complexity of real-world audio-visual scenes, we collect about 4,000 audible video clips from YouTube of which more than $60\%$ are multi-source sound scenarios. More than 20,000 reference expressions are collected and verified by experts to ensure accuracy and reliability, which adopt multimodal cues to describe objects in diverse and dynamic audio-visual scenes. Moreover, a special unseen test set is considered to evaluate the model's generalization ability in the growing demand for zero-shot scenarios. + +Our contributions can be summarized as follows: + +- We propose Ref-AVS as a challenging scene understanding task that segments objects of interest with multimodal-cue expressions and provide the corresponding Ref-AVS benchmark for performance training and validation. +- We design an end-to-end framework for Ref-AVS that efficiently processes the multimodal cues with a crossmodal transformer, serving as a feasible research framework for future development. +- We conduct extensive experiments to demonstrate the advantages of considering multimodal cues for visual segmentation, which also indicate the superiority of our approach in all subsets. + +# 2 Related Work + +# 2.1 Referring Video Object Segmentation (R-VOS) + +The main objective of R-VOS is to perform object segmentation in streaming videos based on given natural language expressions. Initially proposed in 2018, A2D-Sentence [13] aims to segment actors in video content based on + +descriptions of their actions. Subsequently, Refer-DAVIS17 [20] and JHMDB-Sentences [11] use extra language as references to replace previous VOS settings. To facilitate large-scale R-VOS training, Seo et al. [35] develop the ReferYouTube-VOS dataset upon the YouTube-VOS-2019 dataset [40] and Henghui et al. [8] build the MeViS for complex motion expression. Overall, the primary R-VOS methods have evolved from semi-supervised VOS and incorporate attention mechanisms to enhance modalities interaction. For example, MAttNet [43] decomposes the expression information into language- and word-level attention, while URVOS [35] introduces crossmodal attention and memory attention. ReferFormer [39] and MTTR [1] utilize the crossmodal transformer to connect language expression with image regions. These datasets typically offer expressions that encompass the action or appearance attributes of the target object. However, these expressions primarily focus on visual information, which fails to meet the increasing multimodal demands in today's diverse audio-visual scenes. + +# 2.2 Audio-Visual Segmentation (AVS) + +The multimodal segmentation in audio-visual scenes, which is of great concern to us, has received some attention in AVS [12, 37, 46]. AVS aims to obtain finer pixel-level masks corresponding to sound-emitting objects in the visual space. In contrast, previous studies on audio-visual localization [17, 36] use heat maps for coarse localization in an unsupervised learning manner. The existing works on AVS can be broadly categorized into fusion-based methods [12, 18, 22, 24, 25, 46] and prompt-based methods [26, 28, 29, 37]. The former primarily focuses on localizing sounding objects by fusing audio and visual features, while the latter emphasizes constructing effective audio prompts for the visual foundation model. However, AVS simply segments all sound-emitting objects in the visual space, without the flexibility to combine the multimodal cues to segment specific objects of interest. + +# 2.3 Language-aided Audio-Visual Scene Understanding + +Currently, there is a limited number of public tasks offering datasets for audio-visual scene comprehension accompanied by language assistance. Two notable examples of such datasets are Music-AVQA [15] and AVQA [42], both encompassing audio, visual, and language information. These datasets contain rich audio-visual components and focus on annotations for textual questions and answers, emphasizing temporal (Before/After), spatial cues (Left/Right), etc. + +Researchers [4, 19, 21] acknowledge the significance of the provided question, as it guides the feature extraction process for both audio and visual signals. Consequently, they have made efforts to identify pertinent segments by evaluating the semantic similarity between the question and temporal segments. Furthermore, [21] has been dedicated to locating crucial areas by leveraging semantic similarity between questions and visual tokens. Clearly, existing explorations have significantly advanced the research on language-aided audio-visual scene understanding. Nevertheless, the works that rely on the aforementioned datasets + +are still unable to offer pixel-level scene understanding. This encourages us to begin exploring fine-grained segmentation of dynamic audio-visual scenes in the real world by constructing appropriate datasets. + +# 3 Refer-AVS Dataset + +# 3.1 Object Category + +To ensure a diverse range of referred objects, we have carefully selected a wide variety of audible objects in 48 categories as well as a smaller selection of static, unsoundable objects in 3 categories. For the objects that can produce sound, we have chosen 20 categories of musical instruments, 8 of animals, 15 of machines, and 5 of humans. In the special case of humans, considering their diverse appearances, voices, and actions, we have employed the concept of morphology and classified them into 5 categories, based on age and gender. + +# 3.2 Video Selection + +During the process of video collection, we employed the techniques presented in [3,46] to ensure the alignment of audio and visual snippets with the intended semantics. All videos were sourced from YouTube under the Creative Commons license, and each video was trimmed to a duration of 10 seconds. Throughout the manual collection process, we deliberately avoided videos falling into several categories (detailed in the appendix): 1) Videos with a large number of instances of the same semantics; 2) Videos characterized by extensive editing and camera switching; 3) Non-realistic videos containing synthetic artifacts. + +To raise the alignment with real-world distributions, we carefully select videos that contribute to the diversification of scenes within our dataset. Specifically, we focus on selecting videos that involve interactions between multiple objects, such as musical instruments, people, vehicles, etc. The rich combination of categories indicates that our dataset is not limited to a narrow set of scenarios but rather encompasses a broad spectrum of real-life scenes where such objects are likely to naturally appear together. Refer to the supplementary materials for details. + +In addition to diversity, we also carefully filter the videos to ensure that the dataset includes scenes with greater complexity and a larger number of objects. Specifically, $56\%$ of the total videos contain two or more objects, while $13\%$ of the total videos contain three or more objects. + +# 3.3 Expressions + +The diversity of expressions is one of the core factors in the assembly of the Ref-AVS dataset. The expressions consist of three dimensions of information, namely audio, vision, and time. The auditory dimension incorporates characteristics such as volume and rhythm while the visual dimension encompasses attributes like the appearance and spatial configuration of objects. We also leverage temporal + +![](images/2b6172fe72406fe8f653881dc334d9ef979652a16e93ef30ca099e791e1f29ce.jpg) +Fig. 2: The illustration of our Ref-AVS benchmark. Note, for $\bullet \bullet \bullet$ , the sound volume increases successively from silent to loud. Our benchmark is meticulously designed to encompass multimodal expressions from multiple dimensions. By combining various types of modality expressions, we achieve a dataset that exhibits great diversity. + +cues to generate references with sequential hints, e.g., "the one making sound first" or "the one appearing later". By integrating auditory, visual, and temporal information, we craft a rich array of expressions that not only accurately reflect multimodal scenarios but also meet users' specific needs for precise references. Fig. 2 illustrates examples of the combination of different modalities. + +The accuracy of expressions is also a core concern. We follow three rules to generate high-quality expressions: 1) Uniqueness: The object referred to by an expression must be unique and an expression cannot refer to multiple objects simultaneously. 2) Necessity: Complex expressions can be used for reference, but each adjective in the sentence should narrow down the scope of target objects, avoiding unnecessary and redundant descriptions of the object. 3) Clarity: Certain expression templates involve subjective factors, such as "the _ with a louder sound". The use of such expressions should only occur when the situation is clear enough to avoid ambiguous references. + +Besides diversity and accuracy, we have graded the difficulty of expressions based on the number of cues involved, with the proportions of easy, medium, and hard samples in the overall dataset being $20\%$ , $60\%$ , and $20\%$ respectively. This gradation could be a great benefit to future research like curriculum learning. Please refer to the supplementary material for more details. + +# 3.4 Segmentation Masks + +We divide each 10-second video into ten equal 1-second snippets and the objective of annotation is to acquire the first frame's mask for each snippet. For these sampled frames, the ground truth labels are binary masks indicating the target object, according to expressions and multimodal information. + +To obtain such a mask, initially, we manually select the pivotal frames for each 10-second video in which the target object is present. These pivotal frames + +Table 1: Datasets comparison with related tasks. + +
TaskDatasetModalityVideoFrameObjectExpressionObject/VideoPub.
AVLFlickr-SoundNet [34]A+V5,0005,000---CVPR‘2018
VGG-SS [2]A+V5,1585,158---CVPR‘2021
AVSAVS [46]A+V12,35682,97213,500~-1.09~ECCV‘2022
VPO [44]A+V-34,87440,000~-1.14 (Object/Frame)CVPR‘2024
R-VOSJ-HMDB Sentences [11]T+V9289289289281.28CVPR‘2018
A2D Sentences [13]T+V3,78211,9364,8256,6561CVPR‘2018
Refer-DAVIS16 [20]T+V503,455501001ACCV‘2018
Refer-DAVIS17 [20]T+V9013,5432051,5442.27ACCV‘2018
Refer-Youtu-VOS [35]T+V3,97511,9367,45127,8991.86ECCV‘2020
MeViS [8]T+V2,006443,0008,17128,5704.28ICCV‘2023
Ref-AVSRef-AVS BenchA+T+V4,00240,0206,88820,2611.72ECCV‘2024
+ +may occur at the beginning, middle, or end of the video, depending on when the target object is most clearly visible. Subsequently, we utilize Grounding SAM [33] to segment and label the pivotal frames, which will be manually checked and corrected subsequently. This process allows us to generate masks and labels for multiple target objects within the pivotal frames. Once the masks for the pivotal frames are established, we apply a tracking algorithm [6] to track the target object across the preceding and subsequent frames and obtain the ultimate mask and label for the target object in a span of 10 frames. + +# 3.5 Dataset Statistics + +In addition, we compare Refer-AVS with other popular audio-visual benchmarks in the Tab. 1. The Flickr-SoundNet [34] and VGG-SS [2] benchmarks are labeled at a patch level through bounding boxes and possess around 5,000 frame-level annotations. Compared with our dataset with pixel-level annotations, these benchmarks have a significantly lower quantity of annotations. Regarding the AVS dataset [44-46], the videos in our dataset contain a higher average number of objects about 1.72 objects per video, implying that our scenarios are more challenging with multiple sound sources and multiple semantics. Within such scenarios, our Ref-AVS benchmark is particularly valuable as it demonstrates the ability to effectively focus on objects of real interest. Our dataset also has more uniform video durations and a more refined video selection process than AVS. When it comes to datasets for R-VOS tasks, we maintain a consistent advantage in the number of videos compared to other datasets [8,11,13,20,35]. Moreover, we possess a considerable volume of data encompassing a large number of objects, expressions, and complex scenes. + +Overall, our Ref-AVS dataset encompasses a substantial collection of 20,000 expressions and pixel-level annotations spread across 4,000 videos, totaling more than 11 hours. As a result, we believe that this benchmark adequately fulfills the requirements for facilitating research on the Ref-AVS task, while also presenting a significant level of challenge in this domain. We aim to continuously expand the dataset for broader community needs, like training larger foundation models. + +Table 2: Dataset split of Ref-AVS Bench. + +
Subsetalltrainvaltesttest (seen)test (unseen)test (null)
Category†5239395239131
Video4,0022,908276818292269257
Object6,8885,3665181,004478269257
Expression20,26114,1171,3494,7702,2881,4541,028
+ +† The categories here contains "background". + +# 3.6 Dataset Split + +As shown in Tab. 2, the complete dataset is divided into three sets, i.e., a training set of 2908 videos, a validation set of 276 videos, and a test set of 818 videos. The videos in the test set and their corresponding annotations undergo a meticulous review and re-annotation process conducted by experienced workers. In order to thoroughly evaluate the models' performance in the Ref-AVS task, the test set is further divided into three distinct subsets, each serving a specific purpose. + +Seen: The first test subset, referred to as the "seen test set", comprises object categories that have already appeared in the training set. The purpose of establishing this subset is to evaluate the model's fundamental performance and assess how well it generalizes to familiar object categories. + +Unseen: To address the growing demand for the generalization capabilities of various models in an open-world scenario, an additional test set was created specifically to assess the model's ability to generalize to unseen audio-visual scenes. This "unseen test set" consists of object categories that did not appear in the training set, although their super-categories (e.g., animal, vehicle) may have been present in the training set. The purpose of this test set is to evaluate the model's capacity to generalize to novel object categories while leveraging its understanding of broader super-categories. + +Null: A "null reference problem" arises when an expression refers to an object that either does not exist or is not visible in the given context. A model that accurately understands expression guidance should not segment any object in scenarios of a null reference [38]. Based on this consideration, we have developed a "null subset" to test the model's robustness against such challenges. This subset comprises object categories that were present during training but are paired with expressions that do not correlate, ensuring all objects in the video frames are irrelevant to the given reference expression. Therefore, the true masks for this subset are empty, and the model should refrain from segmenting any object. + +# 4 Expression Enhancing with Multimodal Cues + +# 4.1 Overall Architecture + +Ref-AVS is designed to locate objects of interest in dynamic audio-visual scenes by utilizing multimodal cues. To solve this challenging task, we propose the + +![](images/cffc8aac645895b33be4e35ec7dc2eb120bcc870c8b35fe8892544ca4d6deb8e.jpg) +Fig. 3: The overall architecture of our method EEMC. We utilize cached memory to preserve temporal information, enabling the model to capture temporal mutations. Modality encoding provides a more clear context for the multimodal cue features $Q_{m}$ , using the modality identify tokens [aud] and [vis]. Lastly, we achieve efficient prompting of the visual foundation model by employing cross-attention, where mask queries $q$ serve as the target and multimodal cue features $Q_{m}$ act as the source. + +Expression Enhancing with Multimodal Cues (EEMC) method, which focuses on integrating multimodal information in the dynamic audio-visual scenes into reference expressions with the corresponding multimodal cues. This approach enables the formation of a comprehensive multimodal reference feature. Additionally, we employ an attention mechanism to utilize the multimodal reference cues as a prompt for the visual foundation model, thereby facilitating the final segmentation process. + +# 4.2 Multimodal Representations + +Audio: Similar to the video processing, we divide the audio input into clips at 1-second intervals. Audio representations $F_{A} \in \mathbb{R}^{t \times d_{A}}$ is encoded from VGGish [14,16], where $t$ represents the duration of the audio in seconds and matches the number of video frames. The audio representations are extracted offline and the audio encoder is not fine-tuned. + +Visual: We sample $t$ frames from the video input at 1-second intervals and extract visual representations $F_{V} \in \mathbb{R}^{t \times d_{V} \times H \times W}$ using a pre-trained Swin-base model [10]. The visual encoder is not fine-tuned. + +Expression: We use RoBERTa [7, 27] as our text encoder to extract textual expression features $F_{T} \in \mathbb{R}^{L \times d_{T}}$ , where $L$ denotes the length of the expression. The expression representations are obtained off-the-shelf without fine-tuning. + +# 4.3 Temporal Bi-modal Transformer + +Temporal A-T & V-T Fusion: This module is introduced to retrieve the expression-related information of each modality. Firstly, for the convenience of the following multimodal fusion, we prepare the projected audio feature $F_{A'} \in \mathbb{R}^{t \times 1 \times d_V}$ , downsampled and flattened visual feature $F_{V'} \in \mathbb{R}^{t \times (\frac{W}{8} \times \frac{H}{8}) \times d_V}$ , and projected text feature $F_{T'} \in \mathbb{R}^{t \times L \times d_V}$ which is expanded along the temporal dimension. Subsequently, the features are combined to yield the expression-related multimodal features $\hat{F}_A$ , $\hat{F}_V$ and modality-aware expression features $\hat{F}_{T_A}$ , $\hat{F}_{T_V}$ , with the same dimension as $F_{A'}$ , $F_{V'}$ , $F_{T'}$ , respectively: + +$$ +\hat {F} _ {M}, \hat {F} _ {T _ {M}} = M F (C o n c a t \left(F _ {M ^ {\prime}}; F _ {T ^ {\prime}}\right)), M \in \{A, V \} \tag {1} +$$ + +where $\text{Concat}(\cdot)$ denotes the concatenate operation, and $MF(\cdot)$ denotes the modality fusion function, which is employed as self-attention. + +Cached Memory: Although the aforementioned fusion method can combine cross-modal information, it does not give much emphasis to strong temporal dynamics. To address this issue, we introduce a straightforward cached memory $\mathcal{C}_A\in \mathbb{R}^{t\times 2\times d_T},\mathcal{C}_V\in \mathbb{R}^{t\times (W\times H + 1)\times d_T}$ to explicitly capture the temporal mutations as they occur. The cache memory is required to store the mean modality features in the temporal domain from the beginning to the current moment. Then, the multimodal cues feature $Q_{A}$ and $Q_{V}$ output from temporal-aware transformers can be calculated as + +$$ +Q _ {M} ^ {\tau} = (\beta + 1) \hat {F} _ {M} ^ {\tau} - \beta \mathcal {C} _ {M} ^ {\tau}, M \in \{A, V \}, \tag {2} +$$ + +$$ +\mathcal {C} _ {M} ^ {\tau} = \left\{ \begin{array}{l l} 0, & \text {i f} \tau = 1, \\ \frac {1}{\tau - 1} \sum_ {i = 0} ^ {\tau - 1} \hat {F} _ {M} ^ {i}, & \text {i f} \tau > 1, \end{array} \right. \tag {3} +$$ + +while $\tau$ is the time stamp of the specific feature at the temporal dimension and $\beta$ is a hyper-parameter to control the sensitivity to temporal dynamics. When the current audio or visual features do not change significantly compared to the mean of past features (i.e., $\hat{F}_M^\tau = \mathcal{C}_M^\tau$ ), $Q_M^\tau$ remains nearly unchanged. However, in cases where the current change is drastic (i.e., $\hat{F}_M^\tau \neq \mathcal{C}_M^\tau$ ), the cached memory can amplify the difference in the current feature, leading to an output with salient features. It is important to note that text cue feature $Q_T$ inherently possesses a highly abstracted semantic nature, which means manipulation may not be necessary at this stage. Therefore, at this stage, we combine and average the original textual features with the text features enhanced by multimodal information, resulting in a comprehensive enhanced text feature $Q_T = mean(F_{T'} + \hat{F}_{T_A} + \hat{F}_{T_V})$ . Finally, we obtain the processed cue feature of each modality $Q_A$ , $Q_V$ and $Q_T$ . Modality Encoding: The modality encoding approach involves incorporating modality identity tokens $[aud] \in \mathbb{R}^{1 \times d_V}$ for audio modality and $[vis] \in \mathbb{R}^{1 \times d_V}$ for visual modality into the following multimodal cues integration process. Since only two tokens are required to split the sequence into three segments, we only need to manipulate the audio and visual cue features. + +# 4.4 Prompting with Multimodal Cues + +In this Prompting with Multimodal Cues (PMC) phrase, when we obtain the final multimodal cues features, we can concatenate the multimodal cues together and then apply self-attention to obtain comprehensive multimodal cues $Q_{m} \in \mathbb{R}^{t \times (H \times W + L + 3) \times d_V}$ through the Multimodal Integration Transformer: + +$$ +Q _ {m} = M F (C o n c a t ([ Q _ {A}; [ a u d ]; Q _ {V}; [ v i s ]; Q _ {T} ])), \tag {4} +$$ + +where $MF$ is the multimodal fusion function serving as the self-attention for multimodal cues interaction. Then we utilize these comprehensive multinational cues to prompt the learnable mask queries $q \in \mathbb{R}^{N \times d_V}$ in the transformer-based segmentation decoder with cross-attention, where $N$ is the number of mask queries: + +$$ +q _ {Q} = C A T F (q u e r y = q, k e y = Q _ {m}, v a l u e = Q _ {m}), \tag {5} +$$ + +where $q_{Q} \in \mathbb{R}^{N \times d_{V}}$ is the updated mask queries. $CATF(\cdot)$ is the cross-attention transformer with $q$ and $Q_{m}$ as the information target and source, respectively. + +# 5 Experiments + +# 5.1 Implementation Details + +Mask2Former [5] serves as our visual foundation model to provide the commonly used transformer-based segmentation decoder. By default, in this work, all input video frames are scaled to $384 \times 384$ . The shape of the visual features is $[H = 64, W = 64, d_V = 256]$ , and we employ 8-fold downsampling to reduce the computation costs. The audio features with $d_A = 128$ are extracted from the mono-channel waveform. The text features have a shape of $[L = 25, d_T = 768]$ . For convenience, we map all modality feature dimensions to $d_V$ . Hyper-parameter $\beta$ is set to 1 by default. Transformer layers $N_L$ of the Temporal Bi-modal Transformer, Multimodal Integration Transformer and the cross-attention transformer CATF are set to 4 by default. The number of mask queries $N_q$ is fixed to 100. + +# 5.2 Evaluation Metrics + +To conduct a comprehensive evaluation of our Ref-AVS method, we employ the Jaccard index $(\mathcal{I})$ and F-score $(\mathcal{F})$ as performance metrics. Additionally, we introduce $S$ in the null reference test set to assess the efficacy of expression guidance in the model. $S$ denotes the square root of the ratio between the predicted mask area and the background area, a higher value indicates a larger proportion of the mask relative to the background area, suggesting a lack of precise expression guidance provided to the model. + +Table 3: Performance on Ref-AVS Bench. We use three test subsets to evaluate the comprehensive ability of Ref-AVS methods. Mix is the average performance of Seen and Unseen test set. We also use Null test set to evaluate the robustness of multimodal-cue expression guidance. + +
MethodTaskSeenUnseenMix(S+U)Null S (↓)Pub.
J(%)FJ(%)FJ(%)F
AVSbench [46] +textAVS21.220.45727.670.50924.450.483-ECCV'2022
23.200.51132.360.54727.780.5290.208
AVSegFormer [12] +textAVS29.160.42332.710.45130.940.437-AAAI'2024
33.470.47036.050.50134.760.4860.171
GAVS [37] +textAVS24.850.45527.860.46026.360.457-AAAI'2024
28.930.49829.820.49729.380.4980.190
ReferFormer [39] +audioRef-VOS27.460.47724.940.50426.20.491-CVPR'2022
31.310.50130.400.48830.860.4950.176
R2VOS [23] +audioRef-VOS21.340.30923.910.43122.630.370-ICCV'2023
25.010.41027.930.49826.470.4540.183
EEMCRef-AVS34.200.51349.540.64841.870.5810.007ECCV'2024
+ +# 5.3 Quantitative Results + +We compare our Ref-AVS method with existing approaches in the relevant field using our Ref-AVS benchmark. The results from the seen test set demonstrate the superior performance of our Ref-AVS method on this dataset, outperforming other methods by a significant margin. To ensure fairness, we equip the other methods with the same inputs from both the audio and visual modalities as ours. However, even with this additional information, these methods still fail to achieve satisfactory results. The result indicates that simple modal fusion is inadequate for addressing the challenges of understanding multimodal cues within the Ref-AVS task. Instead of directly incorporating audiovisual information, our approach selects textual representations as carriers of multimodal information because they contain rich semantics and cues that are contextually relevant to the current audiovisual environment. + +We conduct tests on the unseen test set and the null reference test set to explore the generalization and multimodal cues-following ability. Our method still maintains a leading position in the unseen test set because we leverage text with highly abstract semantic capabilities as a multimodal information carrier, instead of directly fusing audio and visual information. Therefore, the obtained multimodal cues can provide more robust semantic guidance. On the null test set, we obtain leading results among all other methods, indicating that our method can perceive multimodal cues to a reasonably accurate degree. Compared to other transformer-based and query-based frameworks, AVSBench, as a classic AVS method, obtains inferior results. This may be attributed to the model simply fuses multimodal cues and visual features in a progressive fashion, resulting in relatively weak guidance over the visual space. + +![](images/8a53bb49ce7fbca751068fdbeb6a515b6dc1fdab9bff663dc95ac3199a192dd2.jpg) +Fig. 4: Segmentation masks of the referred objects in the Ref-AVS Bench. Note, for $\bullet \bullet \bullet$ , the sound volume increases successively from silent to loud. We compare our methodology with AVSegformer and ReferFormer. More cases can be found in the supplementary material. + +![](images/b101478385e0d87e0a8d8c96351d50fade3155d9a35b30135bd8b637994afd51.jpg) + +# 5.4 Qualitative Results + +We visualize the segmentation masks on the test set of the Ref-AVS Bench and compare them with AVSegFormer and ReferFormer, as shown in Fig. 4. From these qualitative results, we observe that both ReferFormer in the Ref-VOS task and AVSegFormer in the AVS task fail to accurately segment the object described in the expression. Specifically, AVSegformer struggles to fully comprehend the expression and tends to directly generate the sound source. This issue is exemplified in the bottom-left sample, where AVSegformer erroneously segments the vacuum cleaner instead of the boy. On the other hand, Ref-VOS may not adequately understand the audio-visual scene and thus mistakenly identify the toddler as the piano player, as shown in the top-right sample. In contrast, our Ref-AVS method demonstrates a superior capability to simultaneously process multi-modal expressions and scenes, enabling it to accurately interpret the user instruction and segment the intended object of interest. + +# 5.5 Ablation Study + +We conduct ablation studies to investigate the impact of two modalities (audio and text) information on the Ref-AVS task, as well as the effectiveness of the proposed method. As shown in the Tab. 4, we observe that in setting $②$ , removing the text information results in a significant performance degradation of $11.44\%$ in + +Table 4: Ablation study on modalities and method modules. + +
SettingSeenUnseenMixNull
J(%)FJ(%)FJ(%)FS(↓)
① EEMC34.200.51349.540.64841.870.5810.007
② (-) Text23.670.46837.190.64130.430.5550.011
③ (-) Audio27.270.49243.640.65635.460.5740.010
④ (-) Cached Memory25.300.49436.730.66031.050.5770.009
⑤ (-) Modality Encoding27.200.47044.970.63736.090.5540.009
+ +$\mathcal{I}$ and $2.60\%$ in $\mathcal{F}$ . This degradation is much higher compared to setting ③, where the removal of the audio information results in a combined mixed drop of $6.41\%$ in $\mathcal{I}$ and $0.70\%$ in $\mathcal{F}$ . This phenomenon primarily results from the clarity and directness of textual information as a reference source. In contrast, when relying solely on audio information, the model often overlooks the referential content and instead focuses on identifying objects visually associated with audible behaviors. We also conduct experiments to examine the function of each module in this framework. The cached memory is designed to track significant changes in the temporal domain. Meanwhile, the purpose of modality encoding is to extract more distinct and thorough features from multimodal cues, enhancing modal perception. Results from settings ④ and ⑤ in Tab. 4 demonstrate the effectiveness of these modules. + +# 6 Conclusion + +In this work, we introduce a novel task, Ref-AVS, which requires the machine to segment objects of interest in dynamic audio-visual scenes based on expressions that incorporate multimodal cues and temporal information. To support this field, we develop the first benchmark, Ref-AVS Bench, for performance training and evaluation. Our method, EEMC, establishes a robust baseline with an advanced multimodal cues fusion module. We compare our method with several existing approaches on the Ref-AVS Bench dataset and demonstrate its promising performance in accurately locating objects using multimodal reference expressions. Essentially, Ref-AVS leverages audio, visual, and textual data to enrich real-world scene comprehension. Our task has potential in applications such as attention focusing and object editing in immersive audio-visual scenarios like extended reality, thereby enhancing user interaction experiences. In the future, we plan to further expand the scale of the current dataset to meet the growing data demands in the era of large language models. + +# Acknowledgements + +This research was supported by National Natural Science Foundation of China (NO.62106272), and Public Computing Cloud, Renmin University of China. This research was partially supported by Shanghai Artificial Intelligence Laboratory. + +# References + +1. Botach, A., Zheltonozhskii, E., Baskin, C.: End-to-end referring video object segmentation with multimodal transformers. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4985-4995 (2022) +2. Chen, H., Xie, W., Afouras, T., Nagrani, A., Vedaldi, A., Zisserman, A.: Localizing visual sounds the hard way. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16867-16876 (2021) +3. Chen, H., Xie, W., Vedaldi, A., Zisserman, A.: Vggsound: A large-scale audiovisual dataset. In: ICASSP 2020-2020 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). pp. 721-725. IEEE (2020) +4. Chen, Z., Wang, L., Wang, P., Gao, P.: Question-aware global-local video understanding network for audio-visual question answering. IEEE Transactions on Circuits and Systems for Video Technology (2023) +5. Cheng, B., Choudhuri, A., Misra, I., Kirillov, A., Girdhar, R., Schwing, A.G.: Mask2former for video instance segmentation (2021) +6. Cheng, H.K., Oh, S.W., Price, B., Schwing, A., Lee, J.Y.: Tracking anything with decoupled video segmentation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 1316-1326 (2023) +7. Devlin, J., Chang, M.W., Lee, K., Toutanova, K.: Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805 (2018) +8. Ding, H., Liu, C., He, S., Jiang, X., Loy, C.C.: Mevis: A large-scale benchmark for video segmentation with motion expressions. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 2694-2703 (October 2023) +9. Ding, H., Liu, C., He, S., Jiang, X., Torr, P.H., Bai, S.: Mose: A new dataset for video object segmentation in complex scenes. arXiv preprint arXiv:2302.01872 (2023) +10. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929 (2020) +1. Dutta, U.K., Harandi, M., Sekhar, C.C.: Unsupervised deep metric learning via orthogonality based probabilistic loss. IEEE Transactions on Artificial Intelligence 1(1), 74-84 (2020) +2. Gao, S., Chen, Z., Chen, G., Wang, W., Lu, T.: Avsegformer: Audio-visual segmentation with transformer (2023) +3. Gavrilyuk, K., Ghodrati, A., Li, Z., Snoek, C.G.: Actor and action video segmentation from a sentence. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 5958-5966 (2018) +4. Gemmeke, J.F., Ellis, D.P., Freedman, D., Jansen, A., Lawrence, W., Moore, R.C., Plakal, M., Ritter, M.: Audio set: An ontology and human-labeled dataset for audio events. In: 2017 IEEE international conference on acoustics, speech and signal processing (ICASSP). pp. 776-780. IEEE (2017) +5. Guangyao li, Jake Wei, Y.T.C.X.J.R.W.D.H.: Learning to answer questions in dynamic audio-visual scenarios. IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2022) +6. Hershey, S., Chaudhuri, S., Ellis, D.P., Gemmeke, J.F., Jansen, A., Moore, R.C., Plakal, M., Platt, D., Saurous, R.A., Seybold, B., et al.: Cnn architectures for + +large-scale audio classification. In: 2017 IEEE international conference on acoustics, speech and signal processing (ICASSP). pp. 131-135. IEEE (2017) +17. Hu, D., Qian, R., Jiang, M., Tan, X., Wen, S., Ding, E., Lin, W., Dou, D.: Discriminative sounding objects localization via self-supervised audiovisual matching. Advances in Neural Information Processing Systems 33, 10077-10087 (2020) +18. Huang, S., Li, H., Wang, Y., Zhu, H., Dai, J., Han, J., Rong, W., Liu, S.: Discovering sounding objects by audio queries for audio visual segmentation (2023) +19. Jiang, Y., Yin, J.: Target-aware spatio-temporal reasoning via answering questions in dynamics audio-visual scenarios. arXiv preprint arXiv:2305.12397 (2023) +20. Khoreva, A., Rohrbach, A., Schiele, B.: Video object segmentation with language referring expressions. In: Computer Vision-ACCV 2018: 14th Asian Conference on Computer Vision, Perth, Australia, December 2-6, 2018, Revised Selected Papers, Part IV 14. pp. 123-141. Springer (2019) +21. Li, G., Hou, W., Hu, D.: Progressive spatio-temporal perception for audio-visual question answering. In: Proceedings of the 31st ACM International Conference on Multimedia. p. 7808-7816. MM '23, Association for Computing Machinery, New York, NY, USA (2023). https://doi.org/10.1145/3581783.3612293, https://doi.org/10.1145/3581783.3612293 +22. Li, K., Yang, Z., Chen, L., Yang, Y., Xun, J.: Catr: Combinatorial-dependence audio-queried transformer for audio-visual video segmentation. arXiv preprint arXiv:2309.09709 (2023) +23. Li, X., Wang, J., Xu, X., Li, X., Raj, B., Lu, Y.: Robust referring video object segmentation with cyclic structural consensus. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 22236-22245 (2023) +24. Ling, Y., Li, Y., Gan, Z., Zhang, J., Chi, M., Wang, Y.: Hear to segment: Unmixing the audio to guide the semantic segmentation (2023) +25. Liu, J., Ju, C., Ma, C., Wang, Y., Wang, Y., Zhang, Y.: Audio-aware query-enhanced transformer for audio-visual segmentation (2023) +26. Liu, J., Wang, Y., Ju, C., Zhang, Y., Xie, W.: Annotation-free audio-visual segmentation. arXiv preprint arXiv:2305.11019 (2023) +27. Liu, Y., Ott, M., Goyal, N., Du, J., Joshi, M., Chen, D., Levy, O., Lewis, M., Zettlemoyer, L., Stoyanov, V.: Roberta: A robustly optimized bert pretraining approach. arXiv preprint arXiv:1907.11692 (2019) +28. Ma, J., Sun, P., Wang, Y., Hu, D.: Stepping stones: A progressive training strategy for audio-visual semantic segmentation. IEEE European Conference on Computer Vision (ECCV) (2024) +29. Mo, S., Tian, Y.: Av-sam: Segment anything model meets audio-visual localization and segmentation. arXiv preprint arXiv:2305.01836 (2023) +30. Perazzi, F., Khoreva, A., Benenson, R., Schiele, B., Sorkine-Hornung, A.: Learning video object segmentation from static images. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 2663-2672 (2017) +31. Perazzi, F., Pont-Tuset, J., McWilliams, B., Van Gool, L., Gross, M., Sorkine-Hornung, A.: A benchmark dataset and evaluation methodology for video object segmentation. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 724-732 (2016) +32. Pont-Tuset, J., Perazzi, F., Caelles, S., Arbeláez, P., Sorkine-Hornung, A., Van Gool, L.: The 2017 davis challenge on video object segmentation. arXiv preprint arXiv:1704.00675 (2017) +33. Ren, T., Liu, S., Zeng, A., Lin, J., Li, K., Cao, H., Chen, J., Huang, X., Chen, Y., Yan, F., Zeng, Z., Zhang, H., Li, F., Yang, J., Li, H., Jiang, Q., Zhang, L.: Grounded sam: Assembling open-world models for diverse visual tasks (2024) + +34. Senocak, A., Oh, T.H., Kim, J., Yang, M.H., So Kweon, I.: Learning to localize sound source in visual scenes. In: The IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (June 2018) +35. Seo, S., Lee, J.Y., Han, B.: Urvos: Unified referring video object segmentation network with a large-scale benchmark. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XV 16. pp. 208-223. Springer (2020) +36. Tian, Y., Shi, J., Li, B., Duan, Z., Xu, C.: Audio-visual event localization in unconstrained videos. In: Proceedings of the European Conference on Computer Vision (ECCV) (September 2018) +37. Wang, Y., Liu, W., Li, G., Ding, J., Hu, D., Li, X.: Prompting segmentation with sound is generalizable audio-visual source localizer. arXiv preprint arXiv:2309.07929 (2023) +38. Wang, Y., Sun, P., Li, Y., Zhang, H., Hu, D.: Can textual semantics mitigate sounding object segmentation preference? IEEE European Conference on Computer Vision (ECCV) (2024) +39. Wu, J., Jiang, Y., Sun, P., Yuan, Z., Luo, P.: Language as queries for referring video object segmentation (2022) +40. Xu, N., Yang, L., Fan, Y., Yue, D., Liang, Y., Yang, J., Huang, T.: Youtube-vos: A large-scale video object segmentation benchmark. arXiv preprint arXiv:1809.03327 (2018) +41. Yan, S., Zhang, R., Guo, Z., Chen, W., Zhang, W., Li, H., Qiao, Y., He, Z., Gao, P.: Referred by multi-modality: A unified temporal transformer for video object segmentation (2023) +42. Yang, P., Wang, X., Duan, X., Chen, H., Hou, R., Jin, C., Zhu, W.: Avqa: A dataset for audio-visual question answering on videos. In: Proceedings of the 30th ACM International Conference on Multimedia. pp. 3480-3491 (2022) +43. Yu, L., Lin, Z., Shen, X., Yang, J., Lu, X., Bansal, M., Berg, T.L.: Mattnet: Modular attention network for referring expression comprehension. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1307-1315 (2018) +44. Yuanhong, C., Yuyuan, L., Hu, W., Fengbei, L., Chong, W., Gustavo, C.: A closer look at audio-visual semantic segmentation. arXiv preprint arXiv:2304.02970 (2023) +45. Zhou, J., Shen, X., Wang, J., Zhang, J., Sun, W., Zhang, J., Birchfield, S., Guo, D., Kong, L., Wang, M., Zhong, Y.: Audio-visual segmentation with semantics. arXiv preprint arXiv:2301.13190 (2023) +46. Zhou, J., Wang, J., Zhang, J., Sun, W., Zhang, J., Birchfield, S., Guo, D., Kong, L., Wang, M., Zhong, Y.: Audio-visual segmentation. In: European Conference on Computer Vision (2022) +47. Zhou, X., Zhou, D., Hu, D., Zhou, H., Ouyang, W.: Exploiting visual context semantics for sound source localization. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 5199-5208 (2023) \ No newline at end of file diff --git a/2024/Ref-AVS_ Refer and Segment Objects in Audio-Visual Scenes/images.zip b/2024/Ref-AVS_ Refer and Segment Objects in Audio-Visual Scenes/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..09876d7002e160c0da102256ad530101d0959256 --- /dev/null +++ b/2024/Ref-AVS_ Refer and Segment Objects in Audio-Visual Scenes/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5120ecd5343e2499dea6631b1dd9799cf6bea9ba8d6402de0e1b217690a430e7 +size 567812 diff --git a/2024/Ref-AVS_ Refer and Segment Objects in Audio-Visual Scenes/layout.json b/2024/Ref-AVS_ Refer and Segment Objects in Audio-Visual Scenes/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..cdf74902a8342472119e13eb5c645227b19d7d67 --- /dev/null +++ b/2024/Ref-AVS_ Refer and Segment Objects in Audio-Visual Scenes/layout.json @@ -0,0 +1,8800 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 166, + 111, + 448, + 146 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 111, + 448, + 146 + ], + "spans": [ + { + "bbox": [ + 166, + 111, + 448, + 146 + ], + "type": "text", + "content": "Ref-AVS: Refer and Segment Objects in Audio-Visual Scenes" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 167, + 465, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 167, + 465, + 193 + ], + "spans": [ + { + "bbox": [ + 146, + 167, + 465, + 193 + ], + "type": "text", + "content": "Yaoting Wang" + }, + { + "bbox": [ + 146, + 167, + 465, + 193 + ], + "type": "inline_equation", + "content": "^{1\\dagger}" + }, + { + "bbox": [ + 146, + 167, + 465, + 193 + ], + "type": "text", + "content": ", Peiwen Sun" + }, + { + "bbox": [ + 146, + 167, + 465, + 193 + ], + "type": "inline_equation", + "content": "^{2\\dagger}" + }, + { + "bbox": [ + 146, + 167, + 465, + 193 + ], + "type": "text", + "content": ", Dongzhan Zhou" + }, + { + "bbox": [ + 146, + 167, + 465, + 193 + ], + "type": "inline_equation", + "content": "^{3\\dagger}" + }, + { + "bbox": [ + 146, + 167, + 465, + 193 + ], + "type": "text", + "content": ", Guangyao Li" + }, + { + "bbox": [ + 146, + 167, + 465, + 193 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 146, + 167, + 465, + 193 + ], + "type": "text", + "content": ", Honggang Zhang" + }, + { + "bbox": [ + 146, + 167, + 465, + 193 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 146, + 167, + 465, + 193 + ], + "type": "text", + "content": ", and Di Hu" + }, + { + "bbox": [ + 146, + 167, + 465, + 193 + ], + "type": "inline_equation", + "content": "^{\\boxtimes 1,4}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 148, + 201, + 465, + 235 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 148, + 201, + 465, + 235 + ], + "spans": [ + { + "bbox": [ + 148, + 201, + 465, + 235 + ], + "type": "text", + "content": "1 Gaoling School of Artificial Intelligence, Renmin University of China, China \nyaoting.wang@outlook.com \n{guangyaoli, dihu}@ruc.edu.cn" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 164, + 235, + 448, + 257 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 164, + 235, + 448, + 257 + ], + "spans": [ + { + "bbox": [ + 164, + 235, + 448, + 257 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 164, + 235, + 448, + 257 + ], + "type": "text", + "content": " Beijing University of Posts and Telecommunications, Beijing, China {sunpeiwen,zhhg}@bupt.edu.cn" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 180, + 257, + 433, + 279 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 180, + 257, + 433, + 279 + ], + "spans": [ + { + "bbox": [ + 180, + 257, + 433, + 279 + ], + "type": "text", + "content": "3 Shanghai Artificial Intelligence Laboratory, Shanghai, China zhoudongzhan@pjlab.org.cn" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 143, + 279, + 469, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 279, + 469, + 289 + ], + "spans": [ + { + "bbox": [ + 143, + 279, + 469, + 289 + ], + "type": "text", + "content": "4 Engineering Research Center of Next-Generation Search and Recommendation" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 159, + 310, + 453, + 497 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 310, + 453, + 497 + ], + "spans": [ + { + "bbox": [ + 159, + 310, + 453, + 497 + ], + "type": "text", + "content": "Abstract. Traditional reference segmentation tasks have predominantly focused on silent visual scenes, neglecting the integral role of multimodal perception and interaction in human experiences. In this work, we introduce a novel task called Reference Audio-Visual Segmentation (Ref-AVS), which seeks to segment objects within the visual domain based on expressions containing multimodal cues. Such expressions are articulated in natural language forms but are enriched with multimodal cues, including audio and visual descriptions. To facilitate this research, we construct the first Ref-AVS benchmark, which provides pixel-level annotations for objects described in corresponding multimodal-cue expressions. To tackle the Ref-AVS task, we propose a new method that adequately utilizes multimodal cues to offer precise segmentation guidance. Finally, we conduct quantitative and qualitative experiments on three test subsets to compare our approach with existing methods from related tasks. The results demonstrate the effectiveness of our method, highlighting its capability to precisely segment objects using multimodal-cue expressions. Dataset is available at https://gewu-lab.github.io/Ref-AVS." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 159, + 506, + 453, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 506, + 453, + 529 + ], + "spans": [ + { + "bbox": [ + 159, + 506, + 453, + 529 + ], + "type": "text", + "content": "Keywords: Referring Audio-Visual Segmentation " + }, + { + "bbox": [ + 159, + 506, + 453, + 529 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 506, + 453, + 529 + ], + "type": "text", + "content": " Audio-Visual Segmentation " + }, + { + "bbox": [ + 159, + 506, + 453, + 529 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 506, + 453, + 529 + ], + "type": "text", + "content": " Multimodal Learning" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 559, + 229, + 572 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 559, + 229, + 572 + ], + "spans": [ + { + "bbox": [ + 132, + 559, + 229, + 572 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 580, + 482, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 580, + 482, + 604 + ], + "spans": [ + { + "bbox": [ + 130, + 580, + 482, + 604 + ], + "type": "text", + "content": "In the real world, visual scenes are often accompanied by diverse multimodal information1, including audio and text modalities. This additional multimodal" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 142, + 609, + 231, + 620 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 142, + 609, + 231, + 620 + ], + "spans": [ + { + "bbox": [ + 142, + 609, + 231, + 620 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 142, + 609, + 231, + 620 + ], + "type": "text", + "content": " Equal contribution." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 142, + 620, + 244, + 632 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 142, + 620, + 244, + 632 + ], + "spans": [ + { + "bbox": [ + 142, + 620, + 244, + 632 + ], + "type": "inline_equation", + "content": "\\mathbb{O}" + }, + { + "bbox": [ + 142, + 620, + 244, + 632 + ], + "type": "text", + "content": " Corresponding author." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 136, + 632, + 481, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 632, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 136, + 632, + 481, + 665 + ], + "type": "text", + "content": "To clarify, we refer to the sound and vision in the scene as multimodal information, while the attribute description of audio and vision (e.g., louder and on the right) in the reference expression is named multimodal cues." + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 134, + 115, + 253, + 190 + ], + "blocks": [ + { + "bbox": [ + 134, + 115, + 253, + 190 + ], + "lines": [ + { + "bbox": [ + 134, + 115, + 253, + 190 + ], + "spans": [ + { + "bbox": [ + 134, + 115, + 253, + 190 + ], + "type": "image", + "image_path": "a67defb349bfe0b763649657c54f8dd938206205b45a2524bf2bbaaf0657a43a.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 138, + 191, + 253, + 218 + ], + "blocks": [ + { + "bbox": [ + 138, + 191, + 253, + 218 + ], + "lines": [ + { + "bbox": [ + 138, + 191, + 253, + 218 + ], + "spans": [ + { + "bbox": [ + 138, + 191, + 253, + 218 + ], + "type": "image", + "image_path": "cd9cec3a02c62acd3090fd1bad3797227fee6a27567cfa90a67e3f50386d30e2.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 138, + 219, + 253, + 247 + ], + "blocks": [ + { + "bbox": [ + 138, + 219, + 253, + 247 + ], + "lines": [ + { + "bbox": [ + 138, + 219, + 253, + 247 + ], + "spans": [ + { + "bbox": [ + 138, + 219, + 253, + 247 + ], + "type": "image", + "image_path": "ec91475b65546273e18449d9e230e94cb936a1cba736fbc33d20772d544e4613.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 257, + 482, + 291 + ], + "lines": [ + { + "bbox": [ + 131, + 257, + 482, + 291 + ], + "spans": [ + { + "bbox": [ + 131, + 257, + 482, + 291 + ], + "type": "text", + "content": "Fig. 1: Comparison of the Ref-AVS task with other related tasks. Ref-AVS challenges machines to locate objects of interest in the visual space using multimodal cues, just like humans do in the real world." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 256, + 116, + 359, + 190 + ], + "blocks": [ + { + "bbox": [ + 256, + 116, + 359, + 190 + ], + "lines": [ + { + "bbox": [ + 256, + 116, + 359, + 190 + ], + "spans": [ + { + "bbox": [ + 256, + 116, + 359, + 190 + ], + "type": "image", + "image_path": "fb5bb9a31be2cd4ad32118cdca9b00a7deb3ce89104918e35cc46f4af2e37afc.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 260, + 191, + 359, + 218 + ], + "blocks": [ + { + "bbox": [ + 260, + 191, + 359, + 218 + ], + "lines": [ + { + "bbox": [ + 260, + 191, + 359, + 218 + ], + "spans": [ + { + "bbox": [ + 260, + 191, + 359, + 218 + ], + "type": "image", + "image_path": "7523bf8a058148df62ae39bf991715d14ea28b372cd5d773bb365adfa0109c45.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 249, + 219, + 350, + 247 + ], + "blocks": [ + { + "bbox": [ + 249, + 219, + 350, + 247 + ], + "lines": [ + { + "bbox": [ + 249, + 219, + 350, + 247 + ], + "spans": [ + { + "bbox": [ + 249, + 219, + 350, + 247 + ], + "type": "image", + "image_path": "475650cdc4a48b1e34e9ee625e8afc84651bf7755e4c0c2aef3f11bb54be7c54.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 362, + 118, + 472, + 127 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 362, + 118, + 472, + 127 + ], + "spans": [ + { + "bbox": [ + 362, + 118, + 472, + 127 + ], + "type": "text", + "content": "Which one should the camera focus?" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 362, + 129, + 474, + 145 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 362, + 129, + 474, + 145 + ], + "spans": [ + { + "bbox": [ + 362, + 129, + 474, + 145 + ], + "type": "text", + "content": "An ideal machine needs to know always going for the player who is actually playing." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 362, + 153, + 430, + 161 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 362, + 153, + 430, + 161 + ], + "spans": [ + { + "bbox": [ + 362, + 153, + 430, + 161 + ], + "type": "text", + "content": "In the view of camera." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 362, + 167, + 461, + 182 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 362, + 167, + 461, + 182 + ], + "spans": [ + { + "bbox": [ + 362, + 167, + 461, + 182 + ], + "type": "text", + "content": "How to focus the player who is actually playing." + } + ] + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 362, + 183, + 470, + 237 + ], + "blocks": [ + { + "bbox": [ + 362, + 183, + 470, + 237 + ], + "lines": [ + { + "bbox": [ + 362, + 183, + 470, + 237 + ], + "spans": [ + { + "bbox": [ + 362, + 183, + 470, + 237 + ], + "type": "image", + "image_path": "bcd12c67cc482a48575faaca83c62933c2394a9d1f8649d825a00f089b3bde81.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 316, + 482, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 316, + 482, + 399 + ], + "spans": [ + { + "bbox": [ + 130, + 316, + 482, + 399 + ], + "type": "text", + "content": "information offers valuable spatiotemporal and semantic cues, aiding individuals in locating objects of interest. Similarly, the ability to locate objects based on multimodal cues plays a crucial role in the functioning of machines. For example, stage cameras should be able to locate the musician playing instruments, while movie cameras should locate the superhero in front of the yelling villain. In the above scenarios, using a single modality for segmentation alone cannot achieve the desired behavior." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 401, + 482, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 401, + 482, + 640 + ], + "spans": [ + { + "bbox": [ + 130, + 401, + 482, + 640 + ], + "type": "text", + "content": "To date, researchers have studied the reference segmentation problem from somewhat limited scenarios. As shown in the coordinate system in Fig. 1, there are currently three mainstream developments in segmentation methods for different modalities. Firstly, from the perspective of visual cues, Video Object Segmentation (VOS) [9,30-32,40] emerges with annotated first frame mask as the reference, guiding the segmentation of specific objects in subsequent video frames. While VOS can achieve arbitrary object segmentation in videos, it heavily relies on precise annotation of the first frame, which can be challenging and time-consuming in practice. Secondly, from the perspective of text cues, Referring Video Object Segmentation (R-VOS) [8,11,13,20] explores the segmentation ability by using attribute descriptions as guidance. As R-VOS has successfully replaced the mask annotation used in VOS with natural language, it provides a more accessible and user-friendly reference form but the capability to locate objects in more natural dynamic audio-visual scenes is still limited. In contrast, from the perspective of audio cues, Audio-Visual Segmentation (AVS) [26,37,44-47] leverages audio as a guide to segment sounding objects. This approach effectively addresses the challenge of locating objects in dynamic audio-visual scenes. However, AVS still faces limitations in its ability to focus on general objects that do not generate sound and cannot effectively locate objects of specific interest." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "content": "In simple terms, current work falls short of enabling machines to locate objects of interest in natural dynamic audio-visual scenes. For instance, as illus" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "type": "text", + "content": "Y. Wang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 209 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 209 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 209 + ], + "type": "text", + "content": "trated in Fig. 1, how can machines accurately locate a person who is genuinely playing a musical instrument over time? This requires machines to infer which instrument is making a sound and who is playing that instrument. Though Yan et al. [41] utilize either audio or language modality for multimodal reference segmentation, their method does not integrate multimodal information simultaneously nor support the aforementioned scenario. Therefore, introducing a task to explore the possibility of locating interested objects in natural dynamic audio-visual scenes holds significant practical potential." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 212, + 480, + 425 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 212, + 480, + 425 + ], + "spans": [ + { + "bbox": [ + 130, + 212, + 480, + 425 + ], + "type": "text", + "content": "In this paper, we propose a pixel-level segmentation task called Referring Audio-Visual Segmentation (Ref-AVS), which requires the network to densely predict whether each pixel corresponds to the given multimodal-cue expression, including dynamic audio-visual information. Top-left of Fig. 1 highlights the distinctions between Ref-AVS and previous tasks. The Ref-AVS task poses a greater challenge as it requires the network to accurately locate and segment objects in a more complex and stereoscopic modality space. To this end, a computation model with comprehensive multimodal understanding capabilities is necessary. To foster research advancements in this field, we introduce the Ref-AVS Bench, the pioneering benchmark that focuses on locating and segmenting objects of interest using referring multimodal-cue expressions. Considering the complexity of real-world audio-visual scenes, we collect about 4,000 audible video clips from YouTube of which more than " + }, + { + "bbox": [ + 130, + 212, + 480, + 425 + ], + "type": "inline_equation", + "content": "60\\%" + }, + { + "bbox": [ + 130, + 212, + 480, + 425 + ], + "type": "text", + "content": " are multi-source sound scenarios. More than 20,000 reference expressions are collected and verified by experts to ensure accuracy and reliability, which adopt multimodal cues to describe objects in diverse and dynamic audio-visual scenes. Moreover, a special unseen test set is considered to evaluate the model's generalization ability in the growing demand for zero-shot scenarios." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 146, + 428, + 361, + 439 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 428, + 361, + 439 + ], + "spans": [ + { + "bbox": [ + 146, + 428, + 361, + 439 + ], + "type": "text", + "content": "Our contributions can be summarized as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 137, + 449, + 479, + 555 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 138, + 449, + 479, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 449, + 479, + 483 + ], + "spans": [ + { + "bbox": [ + 138, + 449, + 479, + 483 + ], + "type": "text", + "content": "- We propose Ref-AVS as a challenging scene understanding task that segments objects of interest with multimodal-cue expressions and provide the corresponding Ref-AVS benchmark for performance training and validation." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 137, + 486, + 479, + 519 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 486, + 479, + 519 + ], + "spans": [ + { + "bbox": [ + 137, + 486, + 479, + 519 + ], + "type": "text", + "content": "- We design an end-to-end framework for Ref-AVS that efficiently processes the multimodal cues with a crossmodal transformer, serving as a feasible research framework for future development." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 137, + 522, + 479, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 522, + 479, + 555 + ], + "spans": [ + { + "bbox": [ + 137, + 522, + 479, + 555 + ], + "type": "text", + "content": "- We conduct extensive experiments to demonstrate the advantages of considering multimodal cues for visual segmentation, which also indicate the superiority of our approach in all subsets." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 132, + 579, + 237, + 592 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 579, + 237, + 592 + ], + "spans": [ + { + "bbox": [ + 132, + 579, + 237, + 592 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 607, + 399, + 620 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 607, + 399, + 620 + ], + "spans": [ + { + "bbox": [ + 131, + 607, + 399, + 620 + ], + "type": "text", + "content": "2.1 Referring Video Object Segmentation (R-VOS)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 629, + 479, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 479, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 479, + 665 + ], + "type": "text", + "content": "The main objective of R-VOS is to perform object segmentation in streaming videos based on given natural language expressions. Initially proposed in 2018, A2D-Sentence [13] aims to segment actors in video content based on" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 411, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 411, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 411, + 91, + 447, + 100 + ], + "type": "text", + "content": "Ref-AVS" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 285 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 285 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 285 + ], + "type": "text", + "content": "descriptions of their actions. Subsequently, Refer-DAVIS17 [20] and JHMDB-Sentences [11] use extra language as references to replace previous VOS settings. To facilitate large-scale R-VOS training, Seo et al. [35] develop the ReferYouTube-VOS dataset upon the YouTube-VOS-2019 dataset [40] and Henghui et al. [8] build the MeViS for complex motion expression. Overall, the primary R-VOS methods have evolved from semi-supervised VOS and incorporate attention mechanisms to enhance modalities interaction. For example, MAttNet [43] decomposes the expression information into language- and word-level attention, while URVOS [35] introduces crossmodal attention and memory attention. ReferFormer [39] and MTTR [1] utilize the crossmodal transformer to connect language expression with image regions. These datasets typically offer expressions that encompass the action or appearance attributes of the target object. However, these expressions primarily focus on visual information, which fails to meet the increasing multimodal demands in today's diverse audio-visual scenes." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 300, + 334, + 313 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 300, + 334, + 313 + ], + "spans": [ + { + "bbox": [ + 132, + 300, + 334, + 313 + ], + "type": "text", + "content": "2.2 Audio-Visual Segmentation (AVS)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 319, + 482, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 319, + 482, + 462 + ], + "spans": [ + { + "bbox": [ + 130, + 319, + 482, + 462 + ], + "type": "text", + "content": "The multimodal segmentation in audio-visual scenes, which is of great concern to us, has received some attention in AVS [12, 37, 46]. AVS aims to obtain finer pixel-level masks corresponding to sound-emitting objects in the visual space. In contrast, previous studies on audio-visual localization [17, 36] use heat maps for coarse localization in an unsupervised learning manner. The existing works on AVS can be broadly categorized into fusion-based methods [12, 18, 22, 24, 25, 46] and prompt-based methods [26, 28, 29, 37]. The former primarily focuses on localizing sounding objects by fusing audio and visual features, while the latter emphasizes constructing effective audio prompts for the visual foundation model. However, AVS simply segments all sound-emitting objects in the visual space, without the flexibility to combine the multimodal cues to segment specific objects of interest." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 479, + 419, + 491 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 479, + 419, + 491 + ], + "spans": [ + { + "bbox": [ + 132, + 479, + 419, + 491 + ], + "type": "text", + "content": "2.3 Language-aided Audio-Visual Scene Understanding" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 498, + 481, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 498, + 481, + 569 + ], + "spans": [ + { + "bbox": [ + 130, + 498, + 481, + 569 + ], + "type": "text", + "content": "Currently, there is a limited number of public tasks offering datasets for audio-visual scene comprehension accompanied by language assistance. Two notable examples of such datasets are Music-AVQA [15] and AVQA [42], both encompassing audio, visual, and language information. These datasets contain rich audio-visual components and focus on annotations for textual questions and answers, emphasizing temporal (Before/After), spatial cues (Left/Right), etc." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "text", + "content": "Researchers [4, 19, 21] acknowledge the significance of the provided question, as it guides the feature extraction process for both audio and visual signals. Consequently, they have made efforts to identify pertinent segments by evaluating the semantic similarity between the question and temporal segments. Furthermore, [21] has been dedicated to locating crucial areas by leveraging semantic similarity between questions and visual tokens. Clearly, existing explorations have significantly advanced the research on language-aided audio-visual scene understanding. Nevertheless, the works that rely on the aforementioned datasets" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "type": "text", + "content": "Y. Wang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 152 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 152 + ], + "type": "text", + "content": "are still unable to offer pixel-level scene understanding. This encourages us to begin exploring fine-grained segmentation of dynamic audio-visual scenes in the real world by constructing appropriate datasets." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 171, + 268, + 185 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 171, + 268, + 185 + ], + "spans": [ + { + "bbox": [ + 131, + 171, + 268, + 185 + ], + "type": "text", + "content": "3 Refer-AVS Dataset" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 196, + 246, + 209 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 196, + 246, + 209 + ], + "spans": [ + { + "bbox": [ + 131, + 196, + 246, + 209 + ], + "type": "text", + "content": "3.1 Object Category" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 217, + 482, + 302 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 217, + 482, + 302 + ], + "spans": [ + { + "bbox": [ + 130, + 217, + 482, + 302 + ], + "type": "text", + "content": "To ensure a diverse range of referred objects, we have carefully selected a wide variety of audible objects in 48 categories as well as a smaller selection of static, unsoundable objects in 3 categories. For the objects that can produce sound, we have chosen 20 categories of musical instruments, 8 of animals, 15 of machines, and 5 of humans. In the special case of humans, considering their diverse appearances, voices, and actions, we have employed the concept of morphology and classified them into 5 categories, based on age and gender." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 319, + 239, + 331 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 319, + 239, + 331 + ], + "spans": [ + { + "bbox": [ + 131, + 319, + 239, + 331 + ], + "type": "text", + "content": "3.2 Video Selection" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 339, + 482, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 339, + 482, + 434 + ], + "spans": [ + { + "bbox": [ + 130, + 339, + 482, + 434 + ], + "type": "text", + "content": "During the process of video collection, we employed the techniques presented in [3,46] to ensure the alignment of audio and visual snippets with the intended semantics. All videos were sourced from YouTube under the Creative Commons license, and each video was trimmed to a duration of 10 seconds. Throughout the manual collection process, we deliberately avoided videos falling into several categories (detailed in the appendix): 1) Videos with a large number of instances of the same semantics; 2) Videos characterized by extensive editing and camera switching; 3) Non-realistic videos containing synthetic artifacts." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 435, + 482, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 435, + 482, + 518 + ], + "spans": [ + { + "bbox": [ + 130, + 435, + 482, + 518 + ], + "type": "text", + "content": "To raise the alignment with real-world distributions, we carefully select videos that contribute to the diversification of scenes within our dataset. Specifically, we focus on selecting videos that involve interactions between multiple objects, such as musical instruments, people, vehicles, etc. The rich combination of categories indicates that our dataset is not limited to a narrow set of scenarios but rather encompasses a broad spectrum of real-life scenes where such objects are likely to naturally appear together. Refer to the supplementary materials for details." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 519, + 482, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 519, + 482, + 567 + ], + "spans": [ + { + "bbox": [ + 130, + 519, + 482, + 567 + ], + "type": "text", + "content": "In addition to diversity, we also carefully filter the videos to ensure that the dataset includes scenes with greater complexity and a larger number of objects. Specifically, " + }, + { + "bbox": [ + 130, + 519, + 482, + 567 + ], + "type": "inline_equation", + "content": "56\\%" + }, + { + "bbox": [ + 130, + 519, + 482, + 567 + ], + "type": "text", + "content": " of the total videos contain two or more objects, while " + }, + { + "bbox": [ + 130, + 519, + 482, + 567 + ], + "type": "inline_equation", + "content": "13\\%" + }, + { + "bbox": [ + 130, + 519, + 482, + 567 + ], + "type": "text", + "content": " of the total videos contain three or more objects." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 585, + 220, + 597 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 585, + 220, + 597 + ], + "spans": [ + { + "bbox": [ + 131, + 585, + 220, + 597 + ], + "type": "text", + "content": "3.3 Expressions" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "text", + "content": "The diversity of expressions is one of the core factors in the assembly of the Ref-AVS dataset. The expressions consist of three dimensions of information, namely audio, vision, and time. The auditory dimension incorporates characteristics such as volume and rhythm while the visual dimension encompasses attributes like the appearance and spatial configuration of objects. We also leverage temporal" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 411, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 411, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 411, + 91, + 447, + 100 + ], + "type": "text", + "content": "Ref-AVS" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 134, + 115, + 479, + 247 + ], + "blocks": [ + { + "bbox": [ + 134, + 115, + 479, + 247 + ], + "lines": [ + { + "bbox": [ + 134, + 115, + 479, + 247 + ], + "spans": [ + { + "bbox": [ + 134, + 115, + 479, + 247 + ], + "type": "image", + "image_path": "2b6172fe72406fe8f653881dc334d9ef979652a16e93ef30ca099e791e1f29ce.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 257, + 482, + 304 + ], + "lines": [ + { + "bbox": [ + 130, + 257, + 482, + 304 + ], + "spans": [ + { + "bbox": [ + 130, + 257, + 482, + 304 + ], + "type": "text", + "content": "Fig. 2: The illustration of our Ref-AVS benchmark. Note, for " + }, + { + "bbox": [ + 130, + 257, + 482, + 304 + ], + "type": "inline_equation", + "content": "\\bullet \\bullet \\bullet" + }, + { + "bbox": [ + 130, + 257, + 482, + 304 + ], + "type": "text", + "content": ", the sound volume increases successively from silent to loud. Our benchmark is meticulously designed to encompass multimodal expressions from multiple dimensions. By combining various types of modality expressions, we achieve a dataset that exhibits great diversity." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 327, + 482, + 386 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 327, + 482, + 386 + ], + "spans": [ + { + "bbox": [ + 130, + 327, + 482, + 386 + ], + "type": "text", + "content": "cues to generate references with sequential hints, e.g., \"the one making sound first\" or \"the one appearing later\". By integrating auditory, visual, and temporal information, we craft a rich array of expressions that not only accurately reflect multimodal scenarios but also meet users' specific needs for precise references. Fig. 2 illustrates examples of the combination of different modalities." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 387, + 482, + 494 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 387, + 482, + 494 + ], + "spans": [ + { + "bbox": [ + 130, + 387, + 482, + 494 + ], + "type": "text", + "content": "The accuracy of expressions is also a core concern. We follow three rules to generate high-quality expressions: 1) Uniqueness: The object referred to by an expression must be unique and an expression cannot refer to multiple objects simultaneously. 2) Necessity: Complex expressions can be used for reference, but each adjective in the sentence should narrow down the scope of target objects, avoiding unnecessary and redundant descriptions of the object. 3) Clarity: Certain expression templates involve subjective factors, such as \"the _ with a louder sound\". The use of such expressions should only occur when the situation is clear enough to avoid ambiguous references." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 495, + 482, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 495, + 482, + 555 + ], + "spans": [ + { + "bbox": [ + 130, + 495, + 482, + 555 + ], + "type": "text", + "content": "Besides diversity and accuracy, we have graded the difficulty of expressions based on the number of cues involved, with the proportions of easy, medium, and hard samples in the overall dataset being " + }, + { + "bbox": [ + 130, + 495, + 482, + 555 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 130, + 495, + 482, + 555 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 495, + 482, + 555 + ], + "type": "inline_equation", + "content": "60\\%" + }, + { + "bbox": [ + 130, + 495, + 482, + 555 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 495, + 482, + 555 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 130, + 495, + 482, + 555 + ], + "type": "text", + "content": " respectively. This gradation could be a great benefit to future research like curriculum learning. Please refer to the supplementary material for more details." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 572, + 265, + 585 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 572, + 265, + 585 + ], + "spans": [ + { + "bbox": [ + 131, + 572, + 265, + 585 + ], + "type": "text", + "content": "3.4 Segmentation Masks" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 594, + 482, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 641 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 641 + ], + "type": "text", + "content": "We divide each 10-second video into ten equal 1-second snippets and the objective of annotation is to acquire the first frame's mask for each snippet. For these sampled frames, the ground truth labels are binary masks indicating the target object, according to expressions and multimodal information." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "content": "To obtain such a mask, initially, we manually select the pivotal frames for each 10-second video in which the target object is present. These pivotal frames" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "type": "text", + "content": "Y. Wang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 135, + 135, + 485, + 243 + ], + "blocks": [ + { + "bbox": [ + 205, + 114, + 408, + 125 + ], + "lines": [ + { + "bbox": [ + 205, + 114, + 408, + 125 + ], + "spans": [ + { + "bbox": [ + 205, + 114, + 408, + 125 + ], + "type": "text", + "content": "Table 1: Datasets comparison with related tasks." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 135, + 135, + 485, + 243 + ], + "lines": [ + { + "bbox": [ + 135, + 135, + 485, + 243 + ], + "spans": [ + { + "bbox": [ + 135, + 135, + 485, + 243 + ], + "type": "table", + "html": "
TaskDatasetModalityVideoFrameObjectExpressionObject/VideoPub.
AVLFlickr-SoundNet [34]A+V5,0005,000---CVPR‘2018
VGG-SS [2]A+V5,1585,158---CVPR‘2021
AVSAVS [46]A+V12,35682,97213,500~-1.09~ECCV‘2022
VPO [44]A+V-34,87440,000~-1.14 (Object/Frame)CVPR‘2024
R-VOSJ-HMDB Sentences [11]T+V9289289289281.28CVPR‘2018
A2D Sentences [13]T+V3,78211,9364,8256,6561CVPR‘2018
Refer-DAVIS16 [20]T+V503,455501001ACCV‘2018
Refer-DAVIS17 [20]T+V9013,5432051,5442.27ACCV‘2018
Refer-Youtu-VOS [35]T+V3,97511,9367,45127,8991.86ECCV‘2020
MeViS [8]T+V2,006443,0008,17128,5704.28ICCV‘2023
Ref-AVSRef-AVS BenchA+T+V4,00240,0206,88820,2611.72ECCV‘2024
", + "image_path": "ccd758c3dd4e9ba974151aa7efb008c1fdc9f38dfc9405bc490b566531fec39f.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 268, + 482, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 268, + 482, + 365 + ], + "spans": [ + { + "bbox": [ + 130, + 268, + 482, + 365 + ], + "type": "text", + "content": "may occur at the beginning, middle, or end of the video, depending on when the target object is most clearly visible. Subsequently, we utilize Grounding SAM [33] to segment and label the pivotal frames, which will be manually checked and corrected subsequently. This process allows us to generate masks and labels for multiple target objects within the pivotal frames. Once the masks for the pivotal frames are established, we apply a tracking algorithm [6] to track the target object across the preceding and subsequent frames and obtain the ultimate mask and label for the target object in a span of 10 frames." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 387, + 250, + 398 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 387, + 250, + 398 + ], + "spans": [ + { + "bbox": [ + 132, + 387, + 250, + 398 + ], + "type": "text", + "content": "3.5 Dataset Statistics" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 413, + 482, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 413, + 482, + 592 + ], + "spans": [ + { + "bbox": [ + 130, + 413, + 482, + 592 + ], + "type": "text", + "content": "In addition, we compare Refer-AVS with other popular audio-visual benchmarks in the Tab. 1. The Flickr-SoundNet [34] and VGG-SS [2] benchmarks are labeled at a patch level through bounding boxes and possess around 5,000 frame-level annotations. Compared with our dataset with pixel-level annotations, these benchmarks have a significantly lower quantity of annotations. Regarding the AVS dataset [44-46], the videos in our dataset contain a higher average number of objects about 1.72 objects per video, implying that our scenarios are more challenging with multiple sound sources and multiple semantics. Within such scenarios, our Ref-AVS benchmark is particularly valuable as it demonstrates the ability to effectively focus on objects of real interest. Our dataset also has more uniform video durations and a more refined video selection process than AVS. When it comes to datasets for R-VOS tasks, we maintain a consistent advantage in the number of videos compared to other datasets [8,11,13,20,35]. Moreover, we possess a considerable volume of data encompassing a large number of objects, expressions, and complex scenes." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": "Overall, our Ref-AVS dataset encompasses a substantial collection of 20,000 expressions and pixel-level annotations spread across 4,000 videos, totaling more than 11 hours. As a result, we believe that this benchmark adequately fulfills the requirements for facilitating research on the Ref-AVS task, while also presenting a significant level of challenge in this domain. We aim to continuously expand the dataset for broader community needs, like training larger foundation models." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 411, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 411, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 411, + 91, + 447, + 100 + ], + "type": "text", + "content": "Ref-AVS" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 99 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 99 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 99 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 148, + 136, + 463, + 211 + ], + "blocks": [ + { + "bbox": [ + 220, + 114, + 394, + 126 + ], + "lines": [ + { + "bbox": [ + 220, + 114, + 394, + 126 + ], + "spans": [ + { + "bbox": [ + 220, + 114, + 394, + 126 + ], + "type": "text", + "content": "Table 2: Dataset split of Ref-AVS Bench." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 148, + 136, + 463, + 211 + ], + "lines": [ + { + "bbox": [ + 148, + 136, + 463, + 211 + ], + "spans": [ + { + "bbox": [ + 148, + 136, + 463, + 211 + ], + "type": "table", + "html": "
Subsetalltrainvaltesttest (seen)test (unseen)test (null)
Category†5239395239131
Video4,0022,908276818292269257
Object6,8885,3665181,004478269257
Expression20,26114,1171,3494,7702,2881,4541,028
", + "image_path": "5bea5c731abc0fb5bd3f3b4a405932e917303750bcbcbd9ae2f36e9ebb19df74.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 154, + 212, + 337, + 223 + ], + "lines": [ + { + "bbox": [ + 154, + 212, + 337, + 223 + ], + "spans": [ + { + "bbox": [ + 154, + 212, + 337, + 223 + ], + "type": "text", + "content": "† The categories here contains \"background\"." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 243, + 228, + 255 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 243, + 228, + 255 + ], + "spans": [ + { + "bbox": [ + 132, + 243, + 228, + 255 + ], + "type": "text", + "content": "3.6 Dataset Split" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 261, + 481, + 332 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 261, + 481, + 332 + ], + "spans": [ + { + "bbox": [ + 130, + 261, + 481, + 332 + ], + "type": "text", + "content": "As shown in Tab. 2, the complete dataset is divided into three sets, i.e., a training set of 2908 videos, a validation set of 276 videos, and a test set of 818 videos. The videos in the test set and their corresponding annotations undergo a meticulous review and re-annotation process conducted by experienced workers. In order to thoroughly evaluate the models' performance in the Ref-AVS task, the test set is further divided into three distinct subsets, each serving a specific purpose." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 334, + 481, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 334, + 481, + 380 + ], + "spans": [ + { + "bbox": [ + 130, + 334, + 481, + 380 + ], + "type": "text", + "content": "Seen: The first test subset, referred to as the \"seen test set\", comprises object categories that have already appeared in the training set. The purpose of establishing this subset is to evaluate the model's fundamental performance and assess how well it generalizes to familiar object categories." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 381, + 481, + 476 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 381, + 481, + 476 + ], + "spans": [ + { + "bbox": [ + 130, + 381, + 481, + 476 + ], + "type": "text", + "content": "Unseen: To address the growing demand for the generalization capabilities of various models in an open-world scenario, an additional test set was created specifically to assess the model's ability to generalize to unseen audio-visual scenes. This \"unseen test set\" consists of object categories that did not appear in the training set, although their super-categories (e.g., animal, vehicle) may have been present in the training set. The purpose of this test set is to evaluate the model's capacity to generalize to novel object categories while leveraging its understanding of broader super-categories." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 476, + 481, + 585 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 476, + 481, + 585 + ], + "spans": [ + { + "bbox": [ + 130, + 476, + 481, + 585 + ], + "type": "text", + "content": "Null: A \"null reference problem\" arises when an expression refers to an object that either does not exist or is not visible in the given context. A model that accurately understands expression guidance should not segment any object in scenarios of a null reference [38]. Based on this consideration, we have developed a \"null subset\" to test the model's robustness against such challenges. This subset comprises object categories that were present during training but are paired with expressions that do not correlate, ensuring all objects in the video frames are irrelevant to the given reference expression. Therefore, the true masks for this subset are empty, and the model should refrain from segmenting any object." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 601, + 422, + 615 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 601, + 422, + 615 + ], + "spans": [ + { + "bbox": [ + 132, + 601, + 422, + 615 + ], + "type": "text", + "content": "4 Expression Enhancing with Multimodal Cues" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 624, + 265, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 624, + 265, + 635 + ], + "spans": [ + { + "bbox": [ + 132, + 624, + 265, + 635 + ], + "type": "text", + "content": "4.1 Overall Architecture" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "content": "Ref-AVS is designed to locate objects of interest in dynamic audio-visual scenes by utilizing multimodal cues. To solve this challenging task, we propose the" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "type": "text", + "content": "Y. Wang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 135, + 115, + 479, + 315 + ], + "blocks": [ + { + "bbox": [ + 135, + 115, + 479, + 315 + ], + "lines": [ + { + "bbox": [ + 135, + 115, + 479, + 315 + ], + "spans": [ + { + "bbox": [ + 135, + 115, + 479, + 315 + ], + "type": "image", + "image_path": "cffc8aac645895b33be4e35ec7dc2eb120bcc870c8b35fe8892544ca4d6deb8e.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 326, + 482, + 392 + ], + "lines": [ + { + "bbox": [ + 130, + 326, + 482, + 392 + ], + "spans": [ + { + "bbox": [ + 130, + 326, + 482, + 392 + ], + "type": "text", + "content": "Fig. 3: The overall architecture of our method EEMC. We utilize cached memory to preserve temporal information, enabling the model to capture temporal mutations. Modality encoding provides a more clear context for the multimodal cue features " + }, + { + "bbox": [ + 130, + 326, + 482, + 392 + ], + "type": "inline_equation", + "content": "Q_{m}" + }, + { + "bbox": [ + 130, + 326, + 482, + 392 + ], + "type": "text", + "content": ", using the modality identify tokens [aud] and [vis]. Lastly, we achieve efficient prompting of the visual foundation model by employing cross-attention, where mask queries " + }, + { + "bbox": [ + 130, + 326, + 482, + 392 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 130, + 326, + 482, + 392 + ], + "type": "text", + "content": " serve as the target and multimodal cue features " + }, + { + "bbox": [ + 130, + 326, + 482, + 392 + ], + "type": "inline_equation", + "content": "Q_{m}" + }, + { + "bbox": [ + 130, + 326, + 482, + 392 + ], + "type": "text", + "content": " act as the source." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 415, + 482, + 499 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 415, + 482, + 499 + ], + "spans": [ + { + "bbox": [ + 130, + 415, + 482, + 499 + ], + "type": "text", + "content": "Expression Enhancing with Multimodal Cues (EEMC) method, which focuses on integrating multimodal information in the dynamic audio-visual scenes into reference expressions with the corresponding multimodal cues. This approach enables the formation of a comprehensive multimodal reference feature. Additionally, we employ an attention mechanism to utilize the multimodal reference cues as a prompt for the visual foundation model, thereby facilitating the final segmentation process." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 515, + 304, + 528 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 515, + 304, + 528 + ], + "spans": [ + { + "bbox": [ + 132, + 515, + 304, + 528 + ], + "type": "text", + "content": "4.2 Multimodal Representations" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 533, + 482, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 482, + 592 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 482, + 592 + ], + "type": "text", + "content": "Audio: Similar to the video processing, we divide the audio input into clips at 1-second intervals. Audio representations " + }, + { + "bbox": [ + 130, + 533, + 482, + 592 + ], + "type": "inline_equation", + "content": "F_{A} \\in \\mathbb{R}^{t \\times d_{A}}" + }, + { + "bbox": [ + 130, + 533, + 482, + 592 + ], + "type": "text", + "content": " is encoded from VGGish [14,16], where " + }, + { + "bbox": [ + 130, + 533, + 482, + 592 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 533, + 482, + 592 + ], + "type": "text", + "content": " represents the duration of the audio in seconds and matches the number of video frames. The audio representations are extracted offline and the audio encoder is not fine-tuned." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 594, + 482, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 629 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 629 + ], + "type": "text", + "content": "Visual: We sample " + }, + { + "bbox": [ + 130, + 594, + 482, + 629 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 594, + 482, + 629 + ], + "type": "text", + "content": " frames from the video input at 1-second intervals and extract visual representations " + }, + { + "bbox": [ + 130, + 594, + 482, + 629 + ], + "type": "inline_equation", + "content": "F_{V} \\in \\mathbb{R}^{t \\times d_{V} \\times H \\times W}" + }, + { + "bbox": [ + 130, + 594, + 482, + 629 + ], + "type": "text", + "content": " using a pre-trained Swin-base model [10]. The visual encoder is not fine-tuned." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "text", + "content": "Expression: We use RoBERTa [7, 27] as our text encoder to extract textual expression features " + }, + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "inline_equation", + "content": "F_{T} \\in \\mathbb{R}^{L \\times d_{T}}" + }, + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "text", + "content": " denotes the length of the expression. The expression representations are obtained off-the-shelf without fine-tuning." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 411, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 411, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 411, + 91, + 447, + 100 + ], + "type": "text", + "content": "Ref-AVS" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 324, + 128 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 324, + 128 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 324, + 128 + ], + "type": "text", + "content": "4.3 Temporal Bi-modal Transformer" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 137, + 482, + 233 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 137, + 482, + 233 + ], + "spans": [ + { + "bbox": [ + 130, + 137, + 482, + 233 + ], + "type": "text", + "content": "Temporal A-T & V-T Fusion: This module is introduced to retrieve the expression-related information of each modality. Firstly, for the convenience of the following multimodal fusion, we prepare the projected audio feature " + }, + { + "bbox": [ + 130, + 137, + 482, + 233 + ], + "type": "inline_equation", + "content": "F_{A'} \\in \\mathbb{R}^{t \\times 1 \\times d_V}" + }, + { + "bbox": [ + 130, + 137, + 482, + 233 + ], + "type": "text", + "content": ", downsampled and flattened visual feature " + }, + { + "bbox": [ + 130, + 137, + 482, + 233 + ], + "type": "inline_equation", + "content": "F_{V'} \\in \\mathbb{R}^{t \\times (\\frac{W}{8} \\times \\frac{H}{8}) \\times d_V}" + }, + { + "bbox": [ + 130, + 137, + 482, + 233 + ], + "type": "text", + "content": ", and projected text feature " + }, + { + "bbox": [ + 130, + 137, + 482, + 233 + ], + "type": "inline_equation", + "content": "F_{T'} \\in \\mathbb{R}^{t \\times L \\times d_V}" + }, + { + "bbox": [ + 130, + 137, + 482, + 233 + ], + "type": "text", + "content": " which is expanded along the temporal dimension. Subsequently, the features are combined to yield the expression-related multimodal features " + }, + { + "bbox": [ + 130, + 137, + 482, + 233 + ], + "type": "inline_equation", + "content": "\\hat{F}_A" + }, + { + "bbox": [ + 130, + 137, + 482, + 233 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 137, + 482, + 233 + ], + "type": "inline_equation", + "content": "\\hat{F}_V" + }, + { + "bbox": [ + 130, + 137, + 482, + 233 + ], + "type": "text", + "content": " and modality-aware expression features " + }, + { + "bbox": [ + 130, + 137, + 482, + 233 + ], + "type": "inline_equation", + "content": "\\hat{F}_{T_A}" + }, + { + "bbox": [ + 130, + 137, + 482, + 233 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 137, + 482, + 233 + ], + "type": "inline_equation", + "content": "\\hat{F}_{T_V}" + }, + { + "bbox": [ + 130, + 137, + 482, + 233 + ], + "type": "text", + "content": ", with the same dimension as " + }, + { + "bbox": [ + 130, + 137, + 482, + 233 + ], + "type": "inline_equation", + "content": "F_{A'}" + }, + { + "bbox": [ + 130, + 137, + 482, + 233 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 137, + 482, + 233 + ], + "type": "inline_equation", + "content": "F_{V'}" + }, + { + "bbox": [ + 130, + 137, + 482, + 233 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 137, + 482, + 233 + ], + "type": "inline_equation", + "content": "F_{T'}" + }, + { + "bbox": [ + 130, + 137, + 482, + 233 + ], + "type": "text", + "content": ", respectively:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 201, + 244, + 481, + 258 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 201, + 244, + 481, + 258 + ], + "spans": [ + { + "bbox": [ + 201, + 244, + 481, + 258 + ], + "type": "interline_equation", + "content": "\\hat {F} _ {M}, \\hat {F} _ {T _ {M}} = M F (C o n c a t \\left(F _ {M ^ {\\prime}}; F _ {T ^ {\\prime}}\\right)), M \\in \\{A, V \\} \\tag {1}", + "image_path": "313b6ab5ea34809e27256c9da700caf883b16489c72e55bef14bc3bd938584b5.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 263, + 481, + 287 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 263, + 481, + 287 + ], + "spans": [ + { + "bbox": [ + 130, + 263, + 481, + 287 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 263, + 481, + 287 + ], + "type": "inline_equation", + "content": "\\text{Concat}(\\cdot)" + }, + { + "bbox": [ + 130, + 263, + 481, + 287 + ], + "type": "text", + "content": " denotes the concatenate operation, and " + }, + { + "bbox": [ + 130, + 263, + 481, + 287 + ], + "type": "inline_equation", + "content": "MF(\\cdot)" + }, + { + "bbox": [ + 130, + 263, + 481, + 287 + ], + "type": "text", + "content": " denotes the modality fusion function, which is employed as self-attention." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 288, + 482, + 383 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 288, + 482, + 383 + ], + "spans": [ + { + "bbox": [ + 130, + 288, + 482, + 383 + ], + "type": "text", + "content": "Cached Memory: Although the aforementioned fusion method can combine cross-modal information, it does not give much emphasis to strong temporal dynamics. To address this issue, we introduce a straightforward cached memory " + }, + { + "bbox": [ + 130, + 288, + 482, + 383 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_A\\in \\mathbb{R}^{t\\times 2\\times d_T},\\mathcal{C}_V\\in \\mathbb{R}^{t\\times (W\\times H + 1)\\times d_T}" + }, + { + "bbox": [ + 130, + 288, + 482, + 383 + ], + "type": "text", + "content": " to explicitly capture the temporal mutations as they occur. The cache memory is required to store the mean modality features in the temporal domain from the beginning to the current moment. Then, the multimodal cues feature " + }, + { + "bbox": [ + 130, + 288, + 482, + 383 + ], + "type": "inline_equation", + "content": "Q_{A}" + }, + { + "bbox": [ + 130, + 288, + 482, + 383 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 288, + 482, + 383 + ], + "type": "inline_equation", + "content": "Q_{V}" + }, + { + "bbox": [ + 130, + 288, + 482, + 383 + ], + "type": "text", + "content": " output from temporal-aware transformers can be calculated as" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 222, + 404, + 481, + 419 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 222, + 404, + 481, + 419 + ], + "spans": [ + { + "bbox": [ + 222, + 404, + 481, + 419 + ], + "type": "interline_equation", + "content": "Q _ {M} ^ {\\tau} = (\\beta + 1) \\hat {F} _ {M} ^ {\\tau} - \\beta \\mathcal {C} _ {M} ^ {\\tau}, M \\in \\{A, V \\}, \\tag {2}", + "image_path": "d55e60633215c88fdcec742a320638e2a3f971debc4e883f37806864fc9882ef.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 224, + 421, + 481, + 453 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 224, + 421, + 481, + 453 + ], + "spans": [ + { + "bbox": [ + 224, + 421, + 481, + 453 + ], + "type": "interline_equation", + "content": "\\mathcal {C} _ {M} ^ {\\tau} = \\left\\{ \\begin{array}{l l} 0, & \\text {i f} \\tau = 1, \\\\ \\frac {1}{\\tau - 1} \\sum_ {i = 0} ^ {\\tau - 1} \\hat {F} _ {M} ^ {i}, & \\text {i f} \\tau > 1, \\end{array} \\right. \\tag {3}", + "image_path": "c3dca2ed1d51d54b61435331e12defc914ba6be5f6dd0cf376aff2609f9d79d0.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 460, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 460, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 460, + 482, + 666 + ], + "type": "text", + "content": "while " + }, + { + "bbox": [ + 130, + 460, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 130, + 460, + 482, + 666 + ], + "type": "text", + "content": " is the time stamp of the specific feature at the temporal dimension and " + }, + { + "bbox": [ + 130, + 460, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 130, + 460, + 482, + 666 + ], + "type": "text", + "content": " is a hyper-parameter to control the sensitivity to temporal dynamics. When the current audio or visual features do not change significantly compared to the mean of past features (i.e., " + }, + { + "bbox": [ + 130, + 460, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\hat{F}_M^\\tau = \\mathcal{C}_M^\\tau" + }, + { + "bbox": [ + 130, + 460, + 482, + 666 + ], + "type": "text", + "content": "), " + }, + { + "bbox": [ + 130, + 460, + 482, + 666 + ], + "type": "inline_equation", + "content": "Q_M^\\tau" + }, + { + "bbox": [ + 130, + 460, + 482, + 666 + ], + "type": "text", + "content": " remains nearly unchanged. However, in cases where the current change is drastic (i.e., " + }, + { + "bbox": [ + 130, + 460, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\hat{F}_M^\\tau \\neq \\mathcal{C}_M^\\tau" + }, + { + "bbox": [ + 130, + 460, + 482, + 666 + ], + "type": "text", + "content": "), the cached memory can amplify the difference in the current feature, leading to an output with salient features. It is important to note that text cue feature " + }, + { + "bbox": [ + 130, + 460, + 482, + 666 + ], + "type": "inline_equation", + "content": "Q_T" + }, + { + "bbox": [ + 130, + 460, + 482, + 666 + ], + "type": "text", + "content": " inherently possesses a highly abstracted semantic nature, which means manipulation may not be necessary at this stage. Therefore, at this stage, we combine and average the original textual features with the text features enhanced by multimodal information, resulting in a comprehensive enhanced text feature " + }, + { + "bbox": [ + 130, + 460, + 482, + 666 + ], + "type": "inline_equation", + "content": "Q_T = mean(F_{T'} + \\hat{F}_{T_A} + \\hat{F}_{T_V})" + }, + { + "bbox": [ + 130, + 460, + 482, + 666 + ], + "type": "text", + "content": ". Finally, we obtain the processed cue feature of each modality " + }, + { + "bbox": [ + 130, + 460, + 482, + 666 + ], + "type": "inline_equation", + "content": "Q_A" + }, + { + "bbox": [ + 130, + 460, + 482, + 666 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 460, + 482, + 666 + ], + "type": "inline_equation", + "content": "Q_V" + }, + { + "bbox": [ + 130, + 460, + 482, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 460, + 482, + 666 + ], + "type": "inline_equation", + "content": "Q_T" + }, + { + "bbox": [ + 130, + 460, + 482, + 666 + ], + "type": "text", + "content": ". Modality Encoding: The modality encoding approach involves incorporating modality identity tokens " + }, + { + "bbox": [ + 130, + 460, + 482, + 666 + ], + "type": "inline_equation", + "content": "[aud] \\in \\mathbb{R}^{1 \\times d_V}" + }, + { + "bbox": [ + 130, + 460, + 482, + 666 + ], + "type": "text", + "content": " for audio modality and " + }, + { + "bbox": [ + 130, + 460, + 482, + 666 + ], + "type": "inline_equation", + "content": "[vis] \\in \\mathbb{R}^{1 \\times d_V}" + }, + { + "bbox": [ + 130, + 460, + 482, + 666 + ], + "type": "text", + "content": " for visual modality into the following multimodal cues integration process. Since only two tokens are required to split the sequence into three segments, we only need to manipulate the audio and visual cue features." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 227, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 227, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 227, + 101 + ], + "type": "text", + "content": "Y. Wang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 332, + 128 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 332, + 128 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 332, + 128 + ], + "type": "text", + "content": "4.4 Prompting with Multimodal Cues" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 140, + 482, + 188 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 140, + 482, + 188 + ], + "spans": [ + { + "bbox": [ + 130, + 140, + 482, + 188 + ], + "type": "text", + "content": "In this Prompting with Multimodal Cues (PMC) phrase, when we obtain the final multimodal cues features, we can concatenate the multimodal cues together and then apply self-attention to obtain comprehensive multimodal cues " + }, + { + "bbox": [ + 130, + 140, + 482, + 188 + ], + "type": "inline_equation", + "content": "Q_{m} \\in \\mathbb{R}^{t \\times (H \\times W + L + 3) \\times d_V}" + }, + { + "bbox": [ + 130, + 140, + 482, + 188 + ], + "type": "text", + "content": " through the Multimodal Integration Transformer:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 205, + 200, + 481, + 213 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 205, + 200, + 481, + 213 + ], + "spans": [ + { + "bbox": [ + 205, + 200, + 481, + 213 + ], + "type": "interline_equation", + "content": "Q _ {m} = M F (C o n c a t ([ Q _ {A}; [ a u d ]; Q _ {V}; [ v i s ]; Q _ {T} ])), \\tag {4}", + "image_path": "fa585be1d22b9fbc93b0c9b46cc3d12b91926be3aca98cdc2086a01b06644630.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 222, + 482, + 271 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 222, + 482, + 271 + ], + "spans": [ + { + "bbox": [ + 131, + 222, + 482, + 271 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 222, + 482, + 271 + ], + "type": "inline_equation", + "content": "MF" + }, + { + "bbox": [ + 131, + 222, + 482, + 271 + ], + "type": "text", + "content": " is the multimodal fusion function serving as the self-attention for multimodal cues interaction. Then we utilize these comprehensive multinational cues to prompt the learnable mask queries " + }, + { + "bbox": [ + 131, + 222, + 482, + 271 + ], + "type": "inline_equation", + "content": "q \\in \\mathbb{R}^{N \\times d_V}" + }, + { + "bbox": [ + 131, + 222, + 482, + 271 + ], + "type": "text", + "content": " in the transformer-based segmentation decoder with cross-attention, where " + }, + { + "bbox": [ + 131, + 222, + 482, + 271 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 131, + 222, + 482, + 271 + ], + "type": "text", + "content": " is the number of mask queries:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 200, + 287, + 481, + 298 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 200, + 287, + 481, + 298 + ], + "spans": [ + { + "bbox": [ + 200, + 287, + 481, + 298 + ], + "type": "interline_equation", + "content": "q _ {Q} = C A T F (q u e r y = q, k e y = Q _ {m}, v a l u e = Q _ {m}), \\tag {5}", + "image_path": "8e440ae6b4f1ee69094cacc7756bc03b7f74cb9611f0fb64b9e7a9a5139b2b45.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 305, + 482, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 305, + 482, + 331 + ], + "spans": [ + { + "bbox": [ + 131, + 305, + 482, + 331 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 305, + 482, + 331 + ], + "type": "inline_equation", + "content": "q_{Q} \\in \\mathbb{R}^{N \\times d_{V}}" + }, + { + "bbox": [ + 131, + 305, + 482, + 331 + ], + "type": "text", + "content": " is the updated mask queries. " + }, + { + "bbox": [ + 131, + 305, + 482, + 331 + ], + "type": "inline_equation", + "content": "CATF(\\cdot)" + }, + { + "bbox": [ + 131, + 305, + 482, + 331 + ], + "type": "text", + "content": " is the cross-attention transformer with " + }, + { + "bbox": [ + 131, + 305, + 482, + 331 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 131, + 305, + 482, + 331 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 305, + 482, + 331 + ], + "type": "inline_equation", + "content": "Q_{m}" + }, + { + "bbox": [ + 131, + 305, + 482, + 331 + ], + "type": "text", + "content": " as the information target and source, respectively." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 352, + 230, + 366 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 352, + 230, + 366 + ], + "spans": [ + { + "bbox": [ + 132, + 352, + 230, + 366 + ], + "type": "text", + "content": "5 Experiments" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 380, + 279, + 392 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 380, + 279, + 392 + ], + "spans": [ + { + "bbox": [ + 132, + 380, + 279, + 392 + ], + "type": "text", + "content": "5.1 Implementation Details" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 404, + 482, + 538 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 404, + 482, + 538 + ], + "spans": [ + { + "bbox": [ + 130, + 404, + 482, + 538 + ], + "type": "text", + "content": "Mask2Former [5] serves as our visual foundation model to provide the commonly used transformer-based segmentation decoder. By default, in this work, all input video frames are scaled to " + }, + { + "bbox": [ + 130, + 404, + 482, + 538 + ], + "type": "inline_equation", + "content": "384 \\times 384" + }, + { + "bbox": [ + 130, + 404, + 482, + 538 + ], + "type": "text", + "content": ". The shape of the visual features is " + }, + { + "bbox": [ + 130, + 404, + 482, + 538 + ], + "type": "inline_equation", + "content": "[H = 64, W = 64, d_V = 256]" + }, + { + "bbox": [ + 130, + 404, + 482, + 538 + ], + "type": "text", + "content": ", and we employ 8-fold downsampling to reduce the computation costs. The audio features with " + }, + { + "bbox": [ + 130, + 404, + 482, + 538 + ], + "type": "inline_equation", + "content": "d_A = 128" + }, + { + "bbox": [ + 130, + 404, + 482, + 538 + ], + "type": "text", + "content": " are extracted from the mono-channel waveform. The text features have a shape of " + }, + { + "bbox": [ + 130, + 404, + 482, + 538 + ], + "type": "inline_equation", + "content": "[L = 25, d_T = 768]" + }, + { + "bbox": [ + 130, + 404, + 482, + 538 + ], + "type": "text", + "content": ". For convenience, we map all modality feature dimensions to " + }, + { + "bbox": [ + 130, + 404, + 482, + 538 + ], + "type": "inline_equation", + "content": "d_V" + }, + { + "bbox": [ + 130, + 404, + 482, + 538 + ], + "type": "text", + "content": ". Hyper-parameter " + }, + { + "bbox": [ + 130, + 404, + 482, + 538 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 130, + 404, + 482, + 538 + ], + "type": "text", + "content": " is set to 1 by default. Transformer layers " + }, + { + "bbox": [ + 130, + 404, + 482, + 538 + ], + "type": "inline_equation", + "content": "N_L" + }, + { + "bbox": [ + 130, + 404, + 482, + 538 + ], + "type": "text", + "content": " of the Temporal Bi-modal Transformer, Multimodal Integration Transformer and the cross-attention transformer CATF are set to 4 by default. The number of mask queries " + }, + { + "bbox": [ + 130, + 404, + 482, + 538 + ], + "type": "inline_equation", + "content": "N_q" + }, + { + "bbox": [ + 130, + 404, + 482, + 538 + ], + "type": "text", + "content": " is fixed to 100." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 558, + 257, + 569 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 558, + 257, + 569 + ], + "spans": [ + { + "bbox": [ + 132, + 558, + 257, + 569 + ], + "type": "text", + "content": "5.2 Evaluation Metrics" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "content": "To conduct a comprehensive evaluation of our Ref-AVS method, we employ the Jaccard index " + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "inline_equation", + "content": "(\\mathcal{I})" + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "content": " and F-score " + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "inline_equation", + "content": "(\\mathcal{F})" + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "content": " as performance metrics. Additionally, we introduce " + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "content": " in the null reference test set to assess the efficacy of expression guidance in the model. " + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "content": " denotes the square root of the ratio between the predicted mask area and the background area, a higher value indicates a larger proportion of the mask relative to the background area, suggesting a lack of precise expression guidance provided to the model." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 410, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 410, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 410, + 91, + 447, + 100 + ], + "type": "text", + "content": "Ref-AVS" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 135, + 168, + 480, + 315 + ], + "blocks": [ + { + "bbox": [ + 130, + 114, + 482, + 159 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 482, + 159 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 482, + 159 + ], + "type": "text", + "content": "Table 3: Performance on Ref-AVS Bench. We use three test subsets to evaluate the comprehensive ability of Ref-AVS methods. Mix is the average performance of Seen and Unseen test set. We also use Null test set to evaluate the robustness of multimodal-cue expression guidance." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 135, + 168, + 480, + 315 + ], + "lines": [ + { + "bbox": [ + 135, + 168, + 480, + 315 + ], + "spans": [ + { + "bbox": [ + 135, + 168, + 480, + 315 + ], + "type": "table", + "html": "
MethodTaskSeenUnseenMix(S+U)Null S (↓)Pub.
J(%)FJ(%)FJ(%)F
AVSbench [46] +textAVS21.220.45727.670.50924.450.483-ECCV'2022
23.200.51132.360.54727.780.5290.208
AVSegFormer [12] +textAVS29.160.42332.710.45130.940.437-AAAI'2024
33.470.47036.050.50134.760.4860.171
GAVS [37] +textAVS24.850.45527.860.46026.360.457-AAAI'2024
28.930.49829.820.49729.380.4980.190
ReferFormer [39] +audioRef-VOS27.460.47724.940.50426.20.491-CVPR'2022
31.310.50130.400.48830.860.4950.176
R2VOS [23] +audioRef-VOS21.340.30923.910.43122.630.370-ICCV'2023
25.010.41027.930.49826.470.4540.183
EEMCRef-AVS34.200.51349.540.64841.870.5810.007ECCV'2024
", + "image_path": "2a37dd1590d1f5f46cd9dd701362eaf7a20b2dcb3ad6aeebc59a82e6eeeb90c8.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 342, + 264, + 354 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 342, + 264, + 354 + ], + "spans": [ + { + "bbox": [ + 132, + 342, + 264, + 354 + ], + "type": "text", + "content": "5.3 Quantitative Results" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 375, + 482, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 375, + 482, + 518 + ], + "spans": [ + { + "bbox": [ + 130, + 375, + 482, + 518 + ], + "type": "text", + "content": "We compare our Ref-AVS method with existing approaches in the relevant field using our Ref-AVS benchmark. The results from the seen test set demonstrate the superior performance of our Ref-AVS method on this dataset, outperforming other methods by a significant margin. To ensure fairness, we equip the other methods with the same inputs from both the audio and visual modalities as ours. However, even with this additional information, these methods still fail to achieve satisfactory results. The result indicates that simple modal fusion is inadequate for addressing the challenges of understanding multimodal cues within the Ref-AVS task. Instead of directly incorporating audiovisual information, our approach selects textual representations as carriers of multimodal information because they contain rich semantics and cues that are contextually relevant to the current audiovisual environment." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "content": "We conduct tests on the unseen test set and the null reference test set to explore the generalization and multimodal cues-following ability. Our method still maintains a leading position in the unseen test set because we leverage text with highly abstract semantic capabilities as a multimodal information carrier, instead of directly fusing audio and visual information. Therefore, the obtained multimodal cues can provide more robust semantic guidance. On the null test set, we obtain leading results among all other methods, indicating that our method can perceive multimodal cues to a reasonably accurate degree. Compared to other transformer-based and query-based frameworks, AVSBench, as a classic AVS method, obtains inferior results. This may be attributed to the model simply fuses multimodal cues and visual features in a progressive fashion, resulting in relatively weak guidance over the visual space." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "type": "text", + "content": "Y. Wang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 135, + 113, + 318, + 325 + ], + "blocks": [ + { + "bbox": [ + 135, + 113, + 318, + 325 + ], + "lines": [ + { + "bbox": [ + 135, + 113, + 318, + 325 + ], + "spans": [ + { + "bbox": [ + 135, + 113, + 318, + 325 + ], + "type": "image", + "image_path": "8a53bb49ce7fbca751068fdbeb6a515b6dc1fdab9bff663dc95ac3199a192dd2.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 333, + 482, + 377 + ], + "lines": [ + { + "bbox": [ + 130, + 333, + 482, + 377 + ], + "spans": [ + { + "bbox": [ + 130, + 333, + 482, + 377 + ], + "type": "text", + "content": "Fig. 4: Segmentation masks of the referred objects in the Ref-AVS Bench. Note, for " + }, + { + "bbox": [ + 130, + 333, + 482, + 377 + ], + "type": "inline_equation", + "content": "\\bullet \\bullet \\bullet" + }, + { + "bbox": [ + 130, + 333, + 482, + 377 + ], + "type": "text", + "content": ", the sound volume increases successively from silent to loud. We compare our methodology with AVSegformer and ReferFormer. More cases can be found in the supplementary material." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 327, + 113, + 482, + 325 + ], + "blocks": [ + { + "bbox": [ + 327, + 113, + 482, + 325 + ], + "lines": [ + { + "bbox": [ + 327, + 113, + 482, + 325 + ], + "spans": [ + { + "bbox": [ + 327, + 113, + 482, + 325 + ], + "type": "image", + "image_path": "b101478385e0d87e0a8d8c96351d50fade3155d9a35b30135bd8b637994afd51.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 401, + 257, + 414 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 401, + 257, + 414 + ], + "spans": [ + { + "bbox": [ + 132, + 401, + 257, + 414 + ], + "type": "text", + "content": "5.4 Qualitative Results" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 422, + 482, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 422, + 482, + 578 + ], + "spans": [ + { + "bbox": [ + 130, + 422, + 482, + 578 + ], + "type": "text", + "content": "We visualize the segmentation masks on the test set of the Ref-AVS Bench and compare them with AVSegFormer and ReferFormer, as shown in Fig. 4. From these qualitative results, we observe that both ReferFormer in the Ref-VOS task and AVSegFormer in the AVS task fail to accurately segment the object described in the expression. Specifically, AVSegformer struggles to fully comprehend the expression and tends to directly generate the sound source. This issue is exemplified in the bottom-left sample, where AVSegformer erroneously segments the vacuum cleaner instead of the boy. On the other hand, Ref-VOS may not adequately understand the audio-visual scene and thus mistakenly identify the toddler as the piano player, as shown in the top-right sample. In contrast, our Ref-AVS method demonstrates a superior capability to simultaneously process multi-modal expressions and scenes, enabling it to accurately interpret the user instruction and segment the intended object of interest." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 597, + 239, + 609 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 597, + 239, + 609 + ], + "spans": [ + { + "bbox": [ + 132, + 597, + 239, + 609 + ], + "type": "text", + "content": "5.5 Ablation Study" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "We conduct ablation studies to investigate the impact of two modalities (audio and text) information on the Ref-AVS task, as well as the effectiveness of the proposed method. As shown in the Tab. 4, we observe that in setting " + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "inline_equation", + "content": "②" + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": ", removing the text information results in a significant performance degradation of " + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "inline_equation", + "content": "11.44\\%" + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": " in" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 411, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 411, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 411, + 91, + 447, + 100 + ], + "type": "text", + "content": "Ref-AVS" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 156, + 135, + 459, + 213 + ], + "blocks": [ + { + "bbox": [ + 181, + 114, + 430, + 125 + ], + "lines": [ + { + "bbox": [ + 181, + 114, + 430, + 125 + ], + "spans": [ + { + "bbox": [ + 181, + 114, + 430, + 125 + ], + "type": "text", + "content": "Table 4: Ablation study on modalities and method modules." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 156, + 135, + 459, + 213 + ], + "lines": [ + { + "bbox": [ + 156, + 135, + 459, + 213 + ], + "spans": [ + { + "bbox": [ + 156, + 135, + 459, + 213 + ], + "type": "table", + "html": "
SettingSeenUnseenMixNull
J(%)FJ(%)FJ(%)FS(↓)
① EEMC34.200.51349.540.64841.870.5810.007
② (-) Text23.670.46837.190.64130.430.5550.011
③ (-) Audio27.270.49243.640.65635.460.5740.010
④ (-) Cached Memory25.300.49436.730.66031.050.5770.009
⑤ (-) Modality Encoding27.200.47044.970.63736.090.5540.009
", + "image_path": "a2fa8cf2d398642af24bcbb26672de4e0d060dc7c42c7156445e2a064129a74d.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 243, + 482, + 386 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 243, + 482, + 386 + ], + "spans": [ + { + "bbox": [ + 130, + 243, + 482, + 386 + ], + "type": "inline_equation", + "content": "\\mathcal{I}" + }, + { + "bbox": [ + 130, + 243, + 482, + 386 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 243, + 482, + 386 + ], + "type": "inline_equation", + "content": "2.60\\%" + }, + { + "bbox": [ + 130, + 243, + 482, + 386 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 130, + 243, + 482, + 386 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 130, + 243, + 482, + 386 + ], + "type": "text", + "content": ". This degradation is much higher compared to setting ③, where the removal of the audio information results in a combined mixed drop of " + }, + { + "bbox": [ + 130, + 243, + 482, + 386 + ], + "type": "inline_equation", + "content": "6.41\\%" + }, + { + "bbox": [ + 130, + 243, + 482, + 386 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 130, + 243, + 482, + 386 + ], + "type": "inline_equation", + "content": "\\mathcal{I}" + }, + { + "bbox": [ + 130, + 243, + 482, + 386 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 243, + 482, + 386 + ], + "type": "inline_equation", + "content": "0.70\\%" + }, + { + "bbox": [ + 130, + 243, + 482, + 386 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 130, + 243, + 482, + 386 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 130, + 243, + 482, + 386 + ], + "type": "text", + "content": ". This phenomenon primarily results from the clarity and directness of textual information as a reference source. In contrast, when relying solely on audio information, the model often overlooks the referential content and instead focuses on identifying objects visually associated with audible behaviors. We also conduct experiments to examine the function of each module in this framework. The cached memory is designed to track significant changes in the temporal domain. Meanwhile, the purpose of modality encoding is to extract more distinct and thorough features from multimodal cues, enhancing modal perception. Results from settings ④ and ⑤ in Tab. 4 demonstrate the effectiveness of these modules." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 402, + 220, + 415 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 402, + 220, + 415 + ], + "spans": [ + { + "bbox": [ + 132, + 402, + 220, + 415 + ], + "type": "text", + "content": "6 Conclusion" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 424, + 482, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 424, + 482, + 593 + ], + "spans": [ + { + "bbox": [ + 130, + 424, + 482, + 593 + ], + "type": "text", + "content": "In this work, we introduce a novel task, Ref-AVS, which requires the machine to segment objects of interest in dynamic audio-visual scenes based on expressions that incorporate multimodal cues and temporal information. To support this field, we develop the first benchmark, Ref-AVS Bench, for performance training and evaluation. Our method, EEMC, establishes a robust baseline with an advanced multimodal cues fusion module. We compare our method with several existing approaches on the Ref-AVS Bench dataset and demonstrate its promising performance in accurately locating objects using multimodal reference expressions. Essentially, Ref-AVS leverages audio, visual, and textual data to enrich real-world scene comprehension. Our task has potential in applications such as attention focusing and object editing in immersive audio-visual scenarios like extended reality, thereby enhancing user interaction experiences. In the future, we plan to further expand the scale of the current dataset to meet the growing data demands in the era of large language models." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 607, + 247, + 621 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 607, + 247, + 621 + ], + "spans": [ + { + "bbox": [ + 132, + 607, + 247, + 621 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "type": "text", + "content": "This research was supported by National Natural Science Foundation of China (NO.62106272), and Public Computing Cloud, Renmin University of China. This research was partially supported by Shanghai Artificial Intelligence Laboratory." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "type": "text", + "content": "Y. Wang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 197, + 126 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 197, + 126 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 197, + 126 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 139, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 138, + 139, + 480, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 139, + 480, + 172 + ], + "spans": [ + { + "bbox": [ + 138, + 139, + 480, + 172 + ], + "type": "text", + "content": "1. Botach, A., Zheltonozhskii, E., Baskin, C.: End-to-end referring video object segmentation with multimodal transformers. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4985-4995 (2022)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 173, + 480, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 173, + 480, + 205 + ], + "spans": [ + { + "bbox": [ + 138, + 173, + 480, + 205 + ], + "type": "text", + "content": "2. Chen, H., Xie, W., Afouras, T., Nagrani, A., Vedaldi, A., Zisserman, A.: Localizing visual sounds the hard way. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16867-16876 (2021)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 205, + 480, + 238 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 205, + 480, + 238 + ], + "spans": [ + { + "bbox": [ + 138, + 205, + 480, + 238 + ], + "type": "text", + "content": "3. Chen, H., Xie, W., Vedaldi, A., Zisserman, A.: Vggsound: A large-scale audiovisual dataset. In: ICASSP 2020-2020 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). pp. 721-725. IEEE (2020)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 239, + 480, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 239, + 480, + 270 + ], + "spans": [ + { + "bbox": [ + 138, + 239, + 480, + 270 + ], + "type": "text", + "content": "4. Chen, Z., Wang, L., Wang, P., Gao, P.: Question-aware global-local video understanding network for audio-visual question answering. IEEE Transactions on Circuits and Systems for Video Technology (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 271, + 480, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 271, + 480, + 293 + ], + "spans": [ + { + "bbox": [ + 138, + 271, + 480, + 293 + ], + "type": "text", + "content": "5. Cheng, B., Choudhuri, A., Misra, I., Kirillov, A., Girdhar, R., Schwing, A.G.: Mask2former for video instance segmentation (2021)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 293, + 480, + 325 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 293, + 480, + 325 + ], + "spans": [ + { + "bbox": [ + 138, + 293, + 480, + 325 + ], + "type": "text", + "content": "6. Cheng, H.K., Oh, S.W., Price, B., Schwing, A., Lee, J.Y.: Tracking anything with decoupled video segmentation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 1316-1326 (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 326, + 480, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 326, + 480, + 357 + ], + "spans": [ + { + "bbox": [ + 138, + 326, + 480, + 357 + ], + "type": "text", + "content": "7. Devlin, J., Chang, M.W., Lee, K., Toutanova, K.: Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805 (2018)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 358, + 480, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 358, + 480, + 402 + ], + "spans": [ + { + "bbox": [ + 138, + 358, + 480, + 402 + ], + "type": "text", + "content": "8. Ding, H., Liu, C., He, S., Jiang, X., Loy, C.C.: Mevis: A large-scale benchmark for video segmentation with motion expressions. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 2694-2703 (October 2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 403, + 480, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 403, + 480, + 435 + ], + "spans": [ + { + "bbox": [ + 138, + 403, + 480, + 435 + ], + "type": "text", + "content": "9. Ding, H., Liu, C., He, S., Jiang, X., Torr, P.H., Bai, S.: Mose: A new dataset for video object segmentation in complex scenes. arXiv preprint arXiv:2302.01872 (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 436, + 480, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 436, + 480, + 479 + ], + "spans": [ + { + "bbox": [ + 138, + 436, + 480, + 479 + ], + "type": "text", + "content": "10. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929 (2020)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 479, + 480, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 479, + 480, + 511 + ], + "spans": [ + { + "bbox": [ + 138, + 479, + 480, + 511 + ], + "type": "text", + "content": "1. Dutta, U.K., Harandi, M., Sekhar, C.C.: Unsupervised deep metric learning via orthogonality based probabilistic loss. IEEE Transactions on Artificial Intelligence 1(1), 74-84 (2020)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 512, + 480, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 512, + 480, + 533 + ], + "spans": [ + { + "bbox": [ + 138, + 512, + 480, + 533 + ], + "type": "text", + "content": "2. Gao, S., Chen, Z., Chen, G., Wang, W., Lu, T.: Avsegformer: Audio-visual segmentation with transformer (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 534, + 480, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 534, + 480, + 566 + ], + "spans": [ + { + "bbox": [ + 138, + 534, + 480, + 566 + ], + "type": "text", + "content": "3. Gavrilyuk, K., Ghodrati, A., Li, Z., Snoek, C.G.: Actor and action video segmentation from a sentence. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 5958-5966 (2018)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 567, + 480, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 567, + 480, + 609 + ], + "spans": [ + { + "bbox": [ + 138, + 567, + 480, + 609 + ], + "type": "text", + "content": "4. Gemmeke, J.F., Ellis, D.P., Freedman, D., Jansen, A., Lawrence, W., Moore, R.C., Plakal, M., Ritter, M.: Audio set: An ontology and human-labeled dataset for audio events. In: 2017 IEEE international conference on acoustics, speech and signal processing (ICASSP). pp. 776-780. IEEE (2017)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 611, + 480, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 611, + 480, + 643 + ], + "spans": [ + { + "bbox": [ + 138, + 611, + 480, + 643 + ], + "type": "text", + "content": "5. Guangyao li, Jake Wei, Y.T.C.X.J.R.W.D.H.: Learning to answer questions in dynamic audio-visual scenarios. IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2022)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 643, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 643, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 643, + 480, + 665 + ], + "type": "text", + "content": "6. Hershey, S., Chaudhuri, S., Ellis, D.P., Gemmeke, J.F., Jansen, A., Moore, R.C., Plakal, M., Platt, D., Saurous, R.A., Seybold, B., et al.: Cnn architectures for" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 411, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 411, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 411, + 91, + 447, + 100 + ], + "type": "text", + "content": "Ref-AVS" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 117, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 147, + 117, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 147, + 117, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 147, + 117, + 480, + 138 + ], + "type": "text", + "content": "large-scale audio classification. In: 2017 IEEE international conference on acoustics, speech and signal processing (ICASSP). pp. 131-135. IEEE (2017)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 139, + 481, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 139, + 481, + 171 + ], + "spans": [ + { + "bbox": [ + 133, + 139, + 481, + 171 + ], + "type": "text", + "content": "17. Hu, D., Qian, R., Jiang, M., Tan, X., Wen, S., Ding, E., Lin, W., Dou, D.: Discriminative sounding objects localization via self-supervised audiovisual matching. Advances in Neural Information Processing Systems 33, 10077-10087 (2020)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 171, + 481, + 192 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 171, + 481, + 192 + ], + "spans": [ + { + "bbox": [ + 133, + 171, + 481, + 192 + ], + "type": "text", + "content": "18. Huang, S., Li, H., Wang, Y., Zhu, H., Dai, J., Han, J., Rong, W., Liu, S.: Discovering sounding objects by audio queries for audio visual segmentation (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 192, + 481, + 213 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 192, + 481, + 213 + ], + "spans": [ + { + "bbox": [ + 133, + 192, + 481, + 213 + ], + "type": "text", + "content": "19. Jiang, Y., Yin, J.: Target-aware spatio-temporal reasoning via answering questions in dynamics audio-visual scenarios. arXiv preprint arXiv:2305.12397 (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 213, + 481, + 257 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 213, + 481, + 257 + ], + "spans": [ + { + "bbox": [ + 132, + 213, + 481, + 257 + ], + "type": "text", + "content": "20. Khoreva, A., Rohrbach, A., Schiele, B.: Video object segmentation with language referring expressions. In: Computer Vision-ACCV 2018: 14th Asian Conference on Computer Vision, Perth, Australia, December 2-6, 2018, Revised Selected Papers, Part IV 14. pp. 123-141. Springer (2019)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 257, + 481, + 310 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 257, + 481, + 310 + ], + "spans": [ + { + "bbox": [ + 132, + 257, + 481, + 310 + ], + "type": "text", + "content": "21. Li, G., Hou, W., Hu, D.: Progressive spatio-temporal perception for audio-visual question answering. In: Proceedings of the 31st ACM International Conference on Multimedia. p. 7808-7816. MM '23, Association for Computing Machinery, New York, NY, USA (2023). https://doi.org/10.1145/3581783.3612293, https://doi.org/10.1145/3581783.3612293" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 310, + 481, + 343 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 310, + 481, + 343 + ], + "spans": [ + { + "bbox": [ + 132, + 310, + 481, + 343 + ], + "type": "text", + "content": "22. Li, K., Yang, Z., Chen, L., Yang, Y., Xun, J.: Catr: Combinatorial-dependence audio-queried transformer for audio-visual video segmentation. arXiv preprint arXiv:2309.09709 (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 343, + 481, + 375 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 343, + 481, + 375 + ], + "spans": [ + { + "bbox": [ + 132, + 343, + 481, + 375 + ], + "type": "text", + "content": "23. Li, X., Wang, J., Xu, X., Li, X., Raj, B., Lu, Y.: Robust referring video object segmentation with cyclic structural consensus. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 22236-22245 (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 375, + 481, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 375, + 481, + 396 + ], + "spans": [ + { + "bbox": [ + 132, + 375, + 481, + 396 + ], + "type": "text", + "content": "24. Ling, Y., Li, Y., Gan, Z., Zhang, J., Chi, M., Wang, Y.: Hear to segment: Unmixing the audio to guide the semantic segmentation (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 396, + 481, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 396, + 481, + 418 + ], + "spans": [ + { + "bbox": [ + 132, + 396, + 481, + 418 + ], + "type": "text", + "content": "25. Liu, J., Ju, C., Ma, C., Wang, Y., Wang, Y., Zhang, Y.: Audio-aware query-enhanced transformer for audio-visual segmentation (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 418, + 481, + 439 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 418, + 481, + 439 + ], + "spans": [ + { + "bbox": [ + 132, + 418, + 481, + 439 + ], + "type": "text", + "content": "26. Liu, J., Wang, Y., Ju, C., Zhang, Y., Xie, W.: Annotation-free audio-visual segmentation. arXiv preprint arXiv:2305.11019 (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 439, + 481, + 472 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 439, + 481, + 472 + ], + "spans": [ + { + "bbox": [ + 132, + 439, + 481, + 472 + ], + "type": "text", + "content": "27. Liu, Y., Ott, M., Goyal, N., Du, J., Joshi, M., Chen, D., Levy, O., Lewis, M., Zettlemoyer, L., Stoyanov, V.: Roberta: A robustly optimized bert pretraining approach. arXiv preprint arXiv:1907.11692 (2019)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 472, + 481, + 504 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 472, + 481, + 504 + ], + "spans": [ + { + "bbox": [ + 132, + 472, + 481, + 504 + ], + "type": "text", + "content": "28. Ma, J., Sun, P., Wang, Y., Hu, D.: Stepping stones: A progressive training strategy for audio-visual semantic segmentation. IEEE European Conference on Computer Vision (ECCV) (2024)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 504, + 481, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 504, + 481, + 525 + ], + "spans": [ + { + "bbox": [ + 132, + 504, + 481, + 525 + ], + "type": "text", + "content": "29. Mo, S., Tian, Y.: Av-sam: Segment anything model meets audio-visual localization and segmentation. arXiv preprint arXiv:2305.01836 (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 525, + 481, + 556 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 525, + 481, + 556 + ], + "spans": [ + { + "bbox": [ + 132, + 525, + 481, + 556 + ], + "type": "text", + "content": "30. Perazzi, F., Khoreva, A., Benenson, R., Schiele, B., Sorkine-Hornung, A.: Learning video object segmentation from static images. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 2663-2672 (2017)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 556, + 481, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 556, + 481, + 601 + ], + "spans": [ + { + "bbox": [ + 132, + 556, + 481, + 601 + ], + "type": "text", + "content": "31. Perazzi, F., Pont-Tuset, J., McWilliams, B., Van Gool, L., Gross, M., Sorkine-Hornung, A.: A benchmark dataset and evaluation methodology for video object segmentation. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 724-732 (2016)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 601, + 481, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 601, + 481, + 633 + ], + "spans": [ + { + "bbox": [ + 132, + 601, + 481, + 633 + ], + "type": "text", + "content": "32. Pont-Tuset, J., Perazzi, F., Caelles, S., Arbeláez, P., Sorkine-Hornung, A., Van Gool, L.: The 2017 davis challenge on video object segmentation. arXiv preprint arXiv:1704.00675 (2017)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 633, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 633, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 633, + 481, + 665 + ], + "type": "text", + "content": "33. Ren, T., Liu, S., Zeng, A., Lin, J., Li, K., Cao, H., Chen, J., Huang, X., Chen, Y., Yan, F., Zeng, Z., Zhang, H., Li, F., Yang, J., Li, H., Jiang, Q., Zhang, L.: Grounded sam: Assembling open-world models for diverse visual tasks (2024)" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 227, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 227, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 227, + 101 + ], + "type": "text", + "content": "Y. Wang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 588 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "text", + "content": "34. Senocak, A., Oh, T.H., Kim, J., Yang, M.H., So Kweon, I.: Learning to localize sound source in visual scenes. In: The IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (June 2018)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 150, + 482, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 150, + 482, + 194 + ], + "spans": [ + { + "bbox": [ + 130, + 150, + 482, + 194 + ], + "type": "text", + "content": "35. Seo, S., Lee, J.Y., Han, B.: Urvos: Unified referring video object segmentation network with a large-scale benchmark. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XV 16. pp. 208-223. Springer (2020)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 194, + 481, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 194, + 481, + 226 + ], + "spans": [ + { + "bbox": [ + 132, + 194, + 481, + 226 + ], + "type": "text", + "content": "36. Tian, Y., Shi, J., Li, B., Duan, Z., Xu, C.: Audio-visual event localization in unconstrained videos. In: Proceedings of the European Conference on Computer Vision (ECCV) (September 2018)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 228, + 481, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 228, + 481, + 258 + ], + "spans": [ + { + "bbox": [ + 132, + 228, + 481, + 258 + ], + "type": "text", + "content": "37. Wang, Y., Liu, W., Li, G., Ding, J., Hu, D., Li, X.: Prompting segmentation with sound is generalizable audio-visual source localizer. arXiv preprint arXiv:2309.07929 (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 259, + 481, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 259, + 481, + 293 + ], + "spans": [ + { + "bbox": [ + 132, + 259, + 481, + 293 + ], + "type": "text", + "content": "38. Wang, Y., Sun, P., Li, Y., Zhang, H., Hu, D.: Can textual semantics mitigate sounding object segmentation preference? IEEE European Conference on Computer Vision (ECCV) (2024)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 293, + 481, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 293, + 481, + 314 + ], + "spans": [ + { + "bbox": [ + 132, + 293, + 481, + 314 + ], + "type": "text", + "content": "39. Wu, J., Jiang, Y., Sun, P., Yuan, Z., Luo, P.: Language as queries for referring video object segmentation (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 315, + 481, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 315, + 481, + 346 + ], + "spans": [ + { + "bbox": [ + 132, + 315, + 481, + 346 + ], + "type": "text", + "content": "40. Xu, N., Yang, L., Fan, Y., Yue, D., Liang, Y., Yang, J., Huang, T.: Youtube-vos: A large-scale video object segmentation benchmark. arXiv preprint arXiv:1809.03327 (2018)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 347, + 481, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 347, + 481, + 380 + ], + "spans": [ + { + "bbox": [ + 132, + 347, + 481, + 380 + ], + "type": "text", + "content": "41. Yan, S., Zhang, R., Guo, Z., Chen, W., Zhang, W., Li, H., Qiao, Y., He, Z., Gao, P.: Referred by multi-modality: A unified temporal transformer for video object segmentation (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 380, + 481, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 380, + 481, + 413 + ], + "spans": [ + { + "bbox": [ + 132, + 380, + 481, + 413 + ], + "type": "text", + "content": "42. Yang, P., Wang, X., Duan, X., Chen, H., Hou, R., Jin, C., Zhu, W.: Avqa: A dataset for audio-visual question answering on videos. In: Proceedings of the 30th ACM International Conference on Multimedia. pp. 3480-3491 (2022)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 413, + 481, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 413, + 481, + 456 + ], + "spans": [ + { + "bbox": [ + 132, + 413, + 481, + 456 + ], + "type": "text", + "content": "43. Yu, L., Lin, Z., Shen, X., Yang, J., Lu, X., Bansal, M., Berg, T.L.: Mattnet: Modular attention network for referring expression comprehension. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1307-1315 (2018)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 457, + 481, + 489 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 457, + 481, + 489 + ], + "spans": [ + { + "bbox": [ + 132, + 457, + 481, + 489 + ], + "type": "text", + "content": "44. Yuanhong, C., Yuyuan, L., Hu, W., Fengbei, L., Chong, W., Gustavo, C.: A closer look at audio-visual semantic segmentation. arXiv preprint arXiv:2304.02970 (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 490, + 481, + 522 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 490, + 481, + 522 + ], + "spans": [ + { + "bbox": [ + 132, + 490, + 481, + 522 + ], + "type": "text", + "content": "45. Zhou, J., Shen, X., Wang, J., Zhang, J., Sun, W., Zhang, J., Birchfield, S., Guo, D., Kong, L., Wang, M., Zhong, Y.: Audio-visual segmentation with semantics. arXiv preprint arXiv:2301.13190 (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 523, + 481, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 523, + 481, + 555 + ], + "spans": [ + { + "bbox": [ + 132, + 523, + 481, + 555 + ], + "type": "text", + "content": "46. Zhou, J., Wang, J., Zhang, J., Sun, W., Zhang, J., Birchfield, S., Guo, D., Kong, L., Wang, M., Zhong, Y.: Audio-visual segmentation. In: European Conference on Computer Vision (2022)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 555, + 481, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 555, + 481, + 588 + ], + "spans": [ + { + "bbox": [ + 132, + 555, + 481, + 588 + ], + "type": "text", + "content": "47. Zhou, X., Zhou, D., Hu, D., Zhou, H., Ouyang, W.: Exploiting visual context semantics for sound source localization. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 5199-5208 (2023)" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 411, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 411, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 411, + 91, + 447, + 100 + ], + "type": "text", + "content": "Ref-AVS" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Referring Atomic Video Action Recognition/3c97fc3b-c407-40be-ae83-9d7a4bd31254_content_list.json b/2024/Referring Atomic Video Action Recognition/3c97fc3b-c407-40be-ae83-9d7a4bd31254_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..1a1e50d246172df8d15cf04e6466a9b12612ddf4 --- /dev/null +++ b/2024/Referring Atomic Video Action Recognition/3c97fc3b-c407-40be-ae83-9d7a4bd31254_content_list.json @@ -0,0 +1,2015 @@ +[ + { + "type": "text", + "text": "Referring Atomic Video Action Recognition", + "text_level": 1, + "bbox": [ + 250, + 141, + 751, + 162 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Kunyu Peng $^{1,*}$ , Jia Fu $^{3,4,*}$ , Kailun Yang $^{2,\\dagger}$ , Di Wen $^{1}$ , Yufan Chen $^{1}$ , Ruiping Liu $^{1}$ , Junwei Zheng $^{1}$ , Jiaming Zhang $^{1}$ , M. Saqib Sarfraz $^{1,6}$ , Rainer Stiefelhagen $^{1}$ , and Alina Roitberg $^{5}$", + "bbox": [ + 225, + 188, + 774, + 236 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Karlsruhe Institute of Technology, $^{2}$ Hunan University, $^{3}$ RISE Research Institutes of Sweden, $^{4}$ KTH Royal Institute of Technology, $^{5}$ University of Stuttgart, $^{6}$ Mercedes-Benz Tech Innovation", + "bbox": [ + 250, + 246, + 751, + 290 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. We introduce a new task called Referring Atomic Video Action Recognition (RAVAR), aimed at identifying atomic actions of a particular person based on a textual description and the video data of this person. This task differs from traditional action recognition and localization, where predictions are delivered for all present individuals. In contrast, we focus on recognizing the correct atomic action of a specific individual, guided by text. To explore this task, we present the RefAVA dataset, containing 36,630 instances with manually annotated textual descriptions of the individuals. To establish a strong initial benchmark, we implement and validate baselines from various domains, e.g., atomic action localization, video question answering, and text-video retrieval. Since these existing methods underperform on RAVAR, we introduce RefAtomNet - a novel cross-stream attention-driven method specialized for the unique challenges of RAVAR: the need to interpret a textual referring expression for the targeted individual, utilize this reference to guide the spatial localization and harvest the prediction of the atomic actions for the referring person. The key ingredients are: (1) a multi-stream architecture that connects video, text, and a new location-semantic stream, and (2) cross-stream agent attention fusion and agent token fusion which amplify the most relevant information across these streams and consistently surpasses standard attention-based fusion on RAVAR. Extensive experiments demonstrate the effectiveness of RefAtomNet and its building blocks for recognizing the action of the described individual. The dataset and code will be made publicly available at RAVAR.", + "bbox": [ + 261, + 316, + 743, + 650 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 670, + 375, + 686 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Referring scene understanding [50, 52, 66, 93] aims to solve the underlying computer vision task for the particular scene element specified via a natural language referring expression. Context-driven queries are vital in information retrieval, assistive systems, and multimedia analysis, but incorporating them into computer vision models is challenging due to a complex entanglement of linguistic descriptions, localization, and visual recognition itself. Several benchmarks for referring scene understanding have been successfully established for multi-object", + "bbox": [ + 212, + 698, + 787, + 806 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Equal contribution", + "bbox": [ + 217, + 811, + 362, + 824 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Correspondence: kailun.yang@hnu.edu.cn", + "bbox": [ + 222, + 824, + 508, + 839 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/f616d00f7bfd11a1e4da9bb0839d5a77f557c291b420ae8731c50413dc1fc6fb.jpg", + "image_caption": [ + "Fig. 1: A comparison of AAL Task (left) and RAVAR task (right)." + ], + "image_footnote": [], + "bbox": [ + 222, + 145, + 782, + 238 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "tracking [86], semantic segmentation [30, 46, 72, 75, 76], medical imaging [71], and object detection [12, 64]. While these efforts focus on object-centric referring expressions, many applications require a human-centric understanding of human actions, e.g., in rehabilitation assistance [35, 70] and human-robot interaction [27, 37]. Our work addresses a new problem - reference-driven recognition of atomic actions performed by the human described via a textual reference, which has been overlooked in the past.", + "bbox": [ + 212, + 295, + 785, + 400 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The field of human action recognition has made exciting advances, from new approaches stemming from the rise of visual transformers [23,47,63,69,81,83,85] to new large-scale datasets [4,22,34,73,78]. Datasets utilized for the prediction of atomic actions often feature recordings with multiple individuals [24]. Nevertheless, the vast majority of previously published works [23,31,68,69,81] either rely on manually window crop of a particular person of interest or automatically generated region of interests to predict the atomic actions for all the individuals in one video, which results in large pre- or post-processing effort to focus on one specific person. This workflow is also very inconvenient in certain applications, such as assistive systems for users with visual impairments [51,62,97], who must understand the state of each present in the scene for effective interactions. Leveraging succinct textual descriptions, which may include broad positional indicators (e.g., left, center, and right), appearance attributes (e.g., hair color and clothing), or gender, to steer the model towards delivering atomic action recognition outcomes for the targeted individual, presents a promising solution to the outlined challenges. Employing these textual cues as a reference, the model can facilitate an end-to-end retrieval of the specified instance across the entire video, subsequently providing precise atomic action recognition results.", + "bbox": [ + 212, + 402, + 785, + 672 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To close this gap, we formalize the new task of Referring Atomic Video Action Recognition (RAVAR). The differences between RAVAR and atomic action localization are shown in Fig. 1. The conventional atomic action recognition in multi-person scenarios is often formalized in an action localization manner, and requires the region of interest to localize the human available a priori, after which atomic action recognition is carried out for each individual. Post-processing is required for individuals of interest. In contrast, RAVAR assumes a textual reference and a video as inputs, delivering the referring individual's atomic actions along with the location. This comparison shows the efficiency of RAVAR, leveraging textual information for less labor-intensive subject identification to assist atomic action analysis for the individual of interest. We further", + "bbox": [ + 212, + 674, + 787, + 839 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "K. Peng, J. Fu et al.", + "bbox": [ + 271, + 114, + 408, + 128 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "introduce the new RefAVA dataset for RAVAR. RefAVA is established using the 17,946 video clips of the public AVA dataset [24] initially designed for atomic action localization, which we extend with 36,630 textual descriptions of the corresponding individuals used as the referring expression. The dataset covers diverse settings, e.g., in/out door and day/night time scenarios, and numerous multi-person scenes, constituting an ideal testbed for our task.", + "bbox": [ + 212, + 146, + 787, + 236 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To establish an initial benchmark, we assess 15 well-established approaches from multiple related research domains. Our experiments reveal that none of the benchmarked methods delivers sufficient performance. For example, previous methods within the domains of VQA and VTR typically yield predictions of a coarse analysis regarding actions, whereas strategies in AAL face challenges in spatially localizing the correct referring individual. We attribute this phenomenon to the massive irrelevant information provided by the visual data, distracting the model from the referring location. How to suppress the referring irrelevant information becomes a critical challenge to pursue better RAVAR performances.", + "bbox": [ + 212, + 239, + 787, + 388 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To address this, we introduce the new RefAtomNet approach specialized in simultaneously solving the tasks of understanding textual referring expressions and using them as guides for spatial localization and classification of atomic actions. We begin by computing textual, visual, and newly proposed location-semantic tokens using a large foundation model and well-established object detector. The latter stream is designed to harvest semantic cues of different scene entities: it leverages a pre-trained object detector to localize and identify objects, after which the location-semantic tokens are computed as a fusion of visual bounding box coordinates and text embeddings of the object category. Another novel aspect of our work is the introduction of cross-stream agent attention and agent token fusions, which builds on the concept of agent attention [26] and re-defines it for cross-stream compatibility: agent tokens are reformulated into a 1D sequence format through the use of fully connected layers, bypassing the initial need for 2D pooling, removing the depthwise convolution branch, and forgiving 2D positional encodings. This mechanism enhances the model's ability to filter and emphasize relevant information from multiple data sources effectively. RefAtomNet achieves the highest performance for RAVAR, while a detailed ablation study demonstrates the effectiveness of each component.", + "bbox": [ + 212, + 392, + 787, + 664 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our contributions can be summarized as follows:", + "bbox": [ + 238, + 667, + 591, + 680 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We introduce the new Referring Atomic Video Action Recognition (RAVAR) task and the RefAVA benchmark with 36,630 atomic action instances manually annotated with suitable language expressions.", + "- To establish a competitive benchmark, we examine 11 well-established models from different related fields, namely atomic action localization, video question answering, and video-text retrieval.", + "- We propose RefAtomNet, a new RAVAR approach that uses a novel agent-based semantic-location aware attentional fusion to integrate multi-modal tokens, suppressing the irrelevant visual cues. RefAtomNet delivers improve" + ], + "bbox": [ + 223, + 699, + 790, + 839 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "RAVAR", + "bbox": [ + 674, + 114, + 730, + 126 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 114, + 785, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "ments of $3.85\\%$ and $3.17\\%$ of mAP, and $4.33\\%$ and $4.03\\%$ of AUROC, on val and test sets, respectively, compared with the leading baseline, BLIPv2 [42].", + "bbox": [ + 241, + 145, + 787, + 178 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 200, + 388, + 217 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Referring Scene Understanding. Referring scene understanding aims to locate parts of interest within images or videos guided by natural language, exhibiting the utility in many computer vision tasks including autonomous driving [86] and video editing [6]. The development of this field cannot be advanced without the contribution of high-quality open-source datasets and benchmarks [2, 12, 30, 48, 72, 80, 86, 91]. For example, the CLEVR-Ref+ benchmark is proposed by Liu et al. [52] to achieve visual reasoning with referring expressions. Li et al. [46] proposed referring image segmentation by using a recurrent refinement network. Wu et al. [86] proposed referring multi-object tracking benchmark. However, there are no referring understanding works focusing on atomic video action analysis, whereupon we first conduct RAVAR benchmarking with our RefAVA dataset. We further clarify that Referring Video Object Segmentation (RVOS) task [21, 53, 58, 79] differs us from multiple perspectives, e.g., RVOS includes action name in input textual reference.", + "bbox": [ + 212, + 234, + 787, + 445 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Video Text Retrieval. Video Text Retrieval (VTR) aims at matching relevant video content with text. Amidst the rise of Vision-Language foundation models like CLIP [67] and BLIP [43], efforts [9,42,49,55-57,77,82,87,95] are being made to apply powerful pre-trained models for promoting their competencies in VTR. XCLIP [56] introduced multi-grained contrastive learning for end-to-end VTR, by counting all the video-sentence, video-word, sentence-frame, and frame-word contrasts. BLIPv2 [42] pre-trained a Querying Transformer (QFormer) to bootstrap vision-and-language representation learning and vision-to-language generative learning. The key difference between VTR and RAVAR lies in their inference tasks: VTR queries text or video to localize elements, while RAVAR integrates textual and visual data for fine-grained subject retrieval and atomic action prediction in videos.", + "bbox": [ + 212, + 446, + 789, + 626 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Video Question Answering. Video Question Answering (VQA) focuses on generating answers to questions posed in natural language for a given video. Depending on the emphasis of the question, factoid VQA [1,5,7,19,20,25,28,36,39, 45,54,88,89,96] straightforward queries visual facts. Lei et al. [38] proposed that a single-frame trained transformer-based model, with large-scale pre-training and a frame ensemble at the inference stage, can perform better than existing multi-frame trained models in factoid VQA tasks. On the other hand, inference VQA [18,41] delves into logical reasoning. Li et al. [44] released a video-centric instruction dataset and leveraged a neural interface to integrate video foundation models and Large Language Models (LLM), showcasing capability in temporal reasoning, causal inference, and event localization. In addition, some multimodal VQA frameworks [20,40,90] explore information-invoking scenarios that incorporate visual, audio, subtitle, and external knowledge. However, compared with the RAVAR task, most of the existing VQA approaches do not particularly focus", + "bbox": [ + 212, + 628, + 789, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "K. Peng, J. Fu et al.", + "bbox": [ + 271, + 114, + 408, + 128 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/9a26a67e3a8f3b1b1ded71c2923706c13761f81a525867ae06384322ec1bc114.jpg", + "table_caption": [ + "Table 1: An overview of the referring scene understanding datasets and our RefAVA dataset; Our task is Referring Atomic Video Action Recognition (RAVAR), whereas existing benchmarks focus on Referring Object Detection (ROD), Referring Video-based Object Segmentation (RVOS), and Referring Multi-Object Tracking (RMOT)." + ], + "table_footnote": [], + "table_body": "
DatasetRefCOCO [91]RefCOCO+ [91]RefCOCOg [91]Talk2Car [13]VID-Sentence [10]Refer-DAVUS17 [30]Refer-YV [94]Refer-KITTI [86]RefAVA
TaskRODRODRODRODRODRVOSRMOTRMOTRAVAR
NFrames26,71119,99226,7119,21759,2384,21993,8696,6501,615,140
NInstance26,71119,99226,71110,5197,6543,9787,451-36,630
", + "bbox": [ + 215, + 212, + 782, + 257 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "on atomic action recognition and will only tend to give a coarse textual description, which limits the applications requiring precise prediction, e.g., human-robot assistance [27, 37].", + "bbox": [ + 212, + 286, + 782, + 330 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Atomic Video Action Recognition and Localization. Atomic video-based action recognition [11, 23] and localization [24] involves the identification and analysis of the most fundamental, indivisible actions or movements performed by humans for single and multiple-person scenarios. Compared with the general video action recognition task, atomic video-based action localization is much more fine-grained and is always formulated in a multi-label manner with bounding box predictions. Most of the existing convolutional neural networks (CNN) [4, 16, 17, 81] and the transformer [23, 47, 63, 69, 81, 83, 85] networks for human action recognition are commonly used in the atomic video-based action localization by changing the classification head into multi-label manner, adding additional bounding box prediction head, and integrating region of interest features from human detector. Ryali et al. [69] proposed a hierarchical vision transformer with high efficiency and precision within the realm of the existing methods using video as input. Wang et al. [81] proposed scaling video-masked autoencoders with dual masking. Current methods for predicting individual actions in multi-person scenarios often require manual video cropping for atomic action recognition or generate predictions for all detected individuals, necessitating further human selection and reducing practicality [65, 68, 84]. Most existing AAL methods are not specifically designed for the RAVAR task. We thereby introduce RefAtomNet, a novel method that utilizes location semantics atop predicted location and scene semantic information derived from the scenario together with the textual reference and visual cues, then further uses cross-stream agent attention and agent token fusions to suppress redundant information.", + "bbox": [ + 212, + 330, + 787, + 680 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3 RAVAR: Established Benchmark", + "text_level": 1, + "bbox": [ + 214, + 703, + 570, + 719 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1 Introduction of the RefAVA Dataset", + "text_level": 1, + "bbox": [ + 214, + 737, + 560, + 750 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Textual Annotations. To acquire precise textual annotations for the individuals of interest, 7 annotators manually provided the textual annotations according to the key frame bounding boxes presented in the AVA dataset [24]. Cross-checking among all the annotators for the annotated individuals is conducted to deliver high textual annotation quality.", + "bbox": [ + 212, + 763, + 782, + 839 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "RAVAR", + "bbox": [ + 674, + 114, + 730, + 126 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/3e5fe960a155d8fceeda66643690b3c23e8f632cccd672b5859fa1e8c986b73c.jpg", + "image_caption": [ + "(a) Chord visualization of several keywords." + ], + "image_footnote": [], + "bbox": [ + 225, + 148, + 509, + 291 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/230b39ea41268350f85d973bcbaa8ac917c594a95c1db74f928394e6e59eff56.jpg", + "image_caption": [ + "(b) Annotation statistics: #The number of annotated persons contained in the scenario (x-axis); #The number of instances (y-axis).", + "Fig. 2: An chord visualization of several keywords from the textual references, shown on the left, and the instance amount in video clips containing different numbers of annotated persons in our RefAVA dataset, shown on the right." + ], + "image_footnote": [], + "bbox": [ + 516, + 143, + 784, + 276 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Dataset. We selected 17,946 video clips from 127 movies of the AVA dataset, preserving the most complex scenarios, and annotated each person on each center frame of every 90s video clip. In total, RefAVA has 36,630 labeled instances, and we split them into 22,658 train instances, 10,916 validation instances, and 3,056 test instances, where the samples from different sets are from different movie scenarios. The textual annotations cover information on approximate age, gender, appearance, relative position in the center frame, etc., and without action description. The chord visualization in Fig. 2a reveals the quantitative causal relations between the most frequent notional words in different categories from the textual references. Further, Fig. 2b shows the statistics of instance amount towards the annotated person number inside the scenario where each instance belongs. We deliver the comparison of the statistics between representative existing referring scene understanding datasets and our RefAVA dataset in Tab. 1. The atomic actions involve 80 categories covering Object Manipulation (OM), Person Interactions (PI), and Person Movement (PM). The videos cover diverse scenarios. The test set is sourced from 26 movies different from those in the train set (67 different movies) and val set (34 different movies) to achieve the evaluation of generalizability.", + "bbox": [ + 212, + 395, + 787, + 667 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.2 The Baselines for the RAVAR Benchmark", + "text_level": 1, + "bbox": [ + 214, + 691, + 609, + 705 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We have adapted methodologies from similar fields as our baselines for RAVAR. AAL Baselines. We first reformulate the existing methods from the general atomic video action localization field by integrating the textual reference embeddings into the visual branch. Some action recognition approaches, e.g., I3D [4] and X3D [16], are adapted to AAL following [17]. BERT [14] is used for textual embedding extraction. The selected baselines from this domain can be grouped into CNN-based approaches [4,16] and transformer-based approaches [47,69,81]. All these approaches leverage the pre-trained weight on the Kinetics400 [4].", + "bbox": [ + 212, + 718, + 787, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "K. Peng, J. Fu et al.", + "bbox": [ + 271, + 114, + 408, + 128 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "VQA Baselines. The second group of approaches comes from the video question answering domain, which contains the GPT-based model, i.e., AskAnything [44], and conventional transformer-based model [38]. We input the reference sentence in a questioning manner and add a classification head and a bounding box regression head after the acquisition of the logits. We finetuned the two models for the RAVAR task based on their pre-trained weight on ActivityNetQA [92].", + "bbox": [ + 212, + 142, + 785, + 236 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "VTR Baselines. The last group of the foundation model baselines comes from the video-text retrieval task, where XCLIP [56], CLIP4CLIP [55], BLIPv2 [42], and MeVTR [95] are selected. These foundation models are pre-trained on a combination of numerous datasets, incorporating Conceptual Captions [74], SBU Captions [61], and COCO Captions [8], etc. We reformulate these approaches in the same way as the approaches in VQA.", + "bbox": [ + 212, + 237, + 785, + 327 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "SF Baselines. Some Single Frame (SF) baselines are adopted, where SAM [33], DETR [3], and REFCLIP [29] are utilized together with the CLIP [67] feature extraction backbone.", + "bbox": [ + 212, + 327, + 785, + 372 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "VOS Baseline. We adopt one Video Object Segmentation (VOS) baseline using the encoder of the approach from Su et al. [79] and our prediction head.", + "bbox": [ + 214, + 373, + 785, + 402 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4 RefAtomNet: Proposed Method", + "text_level": 1, + "bbox": [ + 214, + 428, + 562, + 445 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.1 Overview", + "text_level": 1, + "bbox": [ + 214, + 460, + 339, + 474 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We propose a new model, RefAtomNet, with an overview provided in Fig. 3. It leverages three token streams: visual, textual reference, and location-semantic streams. Visual tokens are extracted from the video using ViT [15], while textual reference tokens are obtained from the text using BERT [14]. Both streams are enhanced by QFormer [42]. Additionally, in the location-semantic stream, location-semantic tokens are extracted from the center frame by fusing predicted object coordinates and semantic embeddings of the object categories estimated by a frozen object detector and BERT [14]. To suppress irrelevant information for each stream, we propose the agent-based location-semantic aware attentional fusion to achieve better amplification of relevant information during the cross-stream exchange. The detailed structure is discussed in the following subsections.", + "bbox": [ + 212, + 487, + 785, + 654 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2 RefAtomNet", + "text_level": 1, + "bbox": [ + 214, + 676, + 367, + 691 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Background of QFormer. We rely on BLIPv2 to extract the visual features, where QFormer is the most essential component. Multiple learnable queries are initialized in QFormer as trainable parameters and interact with input data via Transformer attention mechanisms. Each query selectively attends to input parts, capturing task-specific features and updating based on attention outputs. During training, these queries are optimized for specialized information extraction. Queries then directly contribute to generating outputs.", + "bbox": [ + 212, + 703, + 785, + 809 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Extraction of Visual and Textual Reference Tokens. Similar to BLIPv2 [42], we use a pre-trained multimodal model for token extraction from both the video", + "bbox": [ + 212, + 809, + 795, + 839 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "RAVAR", + "bbox": [ + 674, + 114, + 730, + 126 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 114, + 784, + 125 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "and textual reference streams. This model leverages the QFormer [42] architecture to effectively extract and integrate multimodal embeddings from the visual and textual reference data. Specifically, the visual stream employs a ViT [15] backbone encoder to process visual inputs, while the textual branch utilizes BERT [14] to encode textual reference cues. This combined approach facilitates a holistic understanding of both visual content and textual descriptions. The model extracts visual tokens $(\\mathbf{t}^{VT})$ and textual reference tokens $(\\mathbf{t}^{RT})$ through Eq. 1:", + "bbox": [ + 212, + 146, + 787, + 265 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {t} ^ {V T}, \\mathbf {t} ^ {R T} = \\mathcal {V} _ {V L} \\left(\\mathcal {V} _ {V T} \\left(\\mathbf {x} ^ {V T}\\right), \\mathcal {V} _ {R T} \\left(\\mathbf {x} ^ {R T}\\right)\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 354, + 266, + 785, + 284 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $\\mathcal{V}_{VL}$ represents the visual-textual integration model (i.e., QFormer [42]). $\\mathcal{V}_{VT}$ , and $\\mathcal{V}_{RT}$ denote the backbones for extracting visual tokens (i.e., ViT [15]) and textual reference tokens (i.e., BERT [14]), respectively. $\\mathbf{x}^{VT}$ and $\\mathbf{x}^{RT}$ represent the input video and textual reference caption, which are then fed into liner projection layers, i.e., $\\mathbf{P}_{VT}$ and $\\mathbf{P}_{RT}$ , respectively.", + "bbox": [ + 212, + 290, + 787, + 366 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Extraction of Location-Semantic Aware Tokens. To integrate more location and semantic information into the token representations, we leverage a well-established object detector DETR [3] to deliver $N_{o}$ detection results based on the input of the center frame of a video clip, noted as the keyframe. Note that, the detections provided by the object detector contain either the human or other objects with a high confidence score according to Eq. 2.", + "bbox": [ + 212, + 366, + 787, + 455 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {r} _ {\\text {b o x e s}}, \\mathbf {r} _ {\\text {c a t s}} = \\mathcal {V} _ {\\text {d e t s}} (\\mathbf {x} _ {k}), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 415, + 470, + 785, + 486 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $\\mathbf{x}_k$ , $\\mathcal{V}_{\\text{dets}}$ indicates the keyframe and the detection network. $\\mathbf{r}_{\\text{boxes}} \\in \\mathbb{R}^{N_o \\times 4}$ are detected 2D corner coordinates (top left and bottom right corners) of the bounding boxes. $\\mathbf{r}_{\\text{cats}}$ is the predicted category for each bounding box represented in textual format. These category labels of the detected objects are passed to a text encoder (BERT [14]) to extract semantic embeddings. We concatenate these two types of tokens together along the channel dimension and use a single linear projection layer to obtain the aggregated location-semantic aware tokens $(\\mathbf{t}^{LS})$ as Eq. 3.", + "bbox": [ + 212, + 497, + 787, + 618 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {t} ^ {L S} = \\mathbf {P} _ {L S} \\left(\\text {C o n c a t} \\left[ \\mathcal {V} _ {R T} \\left(\\mathbf {r} _ {\\text {c a t s}}\\right), \\mathbf {r} _ {\\text {b o x e s}} \\right]\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 359, + 628, + 785, + 647 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $\\mathbf{P}_{LS}$ indicates a fully connected layer, $\\mathcal{V}_{RT}$ indicates the language feature extraction backbone, i.e., BERT [14], and Concat indicates the concatenation.", + "bbox": [ + 212, + 657, + 787, + 688 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Agent-Based Location Semantic Aware Attentional Fusion.", + "text_level": 1, + "bbox": [ + 214, + 688, + 702, + 702 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Agent Tokens and Sequential Processing. We use agent tokens, inspired by agent attention [26]. Derived through fully connected layers from input tokens, similar to Query, Key, and Value extraction in transformers, agent tokens aggregate essential information via agent-key and agent-query pairs. To enable agent attention for 1D sequential tokens from our streams, we redefine the 2D agent spatial tokens proposed in [26] into a 1D sequential format for multi-stream fusion. Specifically, we replace 2D pooling with linear projection and eliminate the depthwise convolutional branch and biased position encoding to better suit agent acquisition in our model. These 1D sequential agent tokens serve dual functions:", + "bbox": [ + 212, + 704, + 787, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "K. Peng, J. Fu et al.", + "bbox": [ + 271, + 114, + 408, + 128 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/5b0adef20eec56d83631eb637cb8d669f8ebd6cf8f797bf7b2af8c7e511379d0.jpg", + "image_caption": [ + "Fig. 3: An overview of the RefAtomNet architecture." + ], + "image_footnote": [], + "bbox": [ + 218, + 145, + 782, + 340 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "aggregating crucial information from agent-key and agent-query pairs and efficiently redistributing this information to the original value tokens, therefore, enhancing attention focus and reducing irrelevant cues.", + "bbox": [ + 212, + 392, + 782, + 436 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Agent-Based Attention Mechanism. The agent-based attention mechanism for visual, textual reference, and location-semantic streams is computed as:", + "bbox": [ + 212, + 436, + 782, + 468 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {Q} ^ {\\phi}, \\mathbf {K} ^ {\\phi}, \\mathbf {V} ^ {\\phi}, \\mathbf {A} ^ {\\phi} = \\mathbf {W} _ {Q} ^ {\\phi} (\\mathbf {t} ^ {\\phi}), \\mathbf {W} _ {K} ^ {\\phi} (\\mathbf {t} ^ {\\phi}), \\mathbf {W} _ {V} ^ {\\phi} (\\mathbf {t} ^ {\\phi}), \\mathbf {W} _ {A} ^ {\\phi} (\\mathbf {t} ^ {\\phi}), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 297, + 474, + 785, + 494 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where for better readability, we use $\\phi$ to represent $[RT, VT, LS]$ , where RT indicates the reference tokens, VT indicates the visual tokens, and LS indicates the location-semantic tokens. $\\mathbf{Q}^{\\phi}$ , $\\mathbf{K}^{\\phi}$ , $\\mathbf{V}^{\\phi}$ , and $\\mathbf{A}^{\\phi}$ are the query, key, value, and agent tokens. $t^{\\phi}$ depicts the input. $\\mathbf{W}_Q^{\\phi}$ , $\\mathbf{W}_K^{\\phi}$ , $\\mathbf{W}_V^{\\phi}$ , and $\\mathbf{W}_A^{\\phi}$ are constructed by linear projection layers. We then project the agent tokens using fully connected layers $\\mathbf{P}_A^\\phi$ through $\\mathbf{A}_*^{\\phi} = \\mathbf{P}_A^\\phi (\\mathbf{A}^\\phi)$ as aforementioned. To mitigate redundancy cues of the textual reference and location-semantic streams, we leverage agent attention for both of these two streams. The agent query attention mask $\\mathbf{M}_{QA}^{\\pi}$ and the agent key attention mask $\\mathbf{M}_{KA}^{\\pi}$ are obtained by matrix multiplication (MatMul) and SoftMax operations along the channel dimension (indicated by $\\sigma_c$ ) as shown in Eq. 5, where $\\pi \\in [RT, LS]$ and $\\alpha$ indicates a fixed scale factor. The agent attention masks and tokens are computed and refined to ensure that only pertinent information influences the attention mechanism, as shown in Eq. 5 and Eq. 6:", + "bbox": [ + 212, + 500, + 787, + 712 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {M} _ {Q A} ^ {\\pi}, \\mathbf {M} _ {K A} ^ {\\pi} = \\sigma_ {c} (M a t M u l [ \\alpha * \\mathbf {A} _ {*} ^ {\\pi}, \\mathbf {Q} ^ {\\pi} ]), \\sigma_ {c} (M a t M u l [ \\alpha * \\mathbf {A} _ {*} ^ {\\pi}, \\mathbf {K} ^ {\\pi} ]), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 256, + 719, + 784, + 738 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {t} _ {*} ^ {\\pi} = F F N \\left(M a t M u l \\left[ \\mathbf {M} _ {K A} ^ {\\pi}, M a t M u l \\left[ \\mathbf {M} _ {Q A} ^ {\\pi}, \\mathbf {V} ^ {\\pi} \\right] \\right]\\right), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 326, + 739, + 784, + 758 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where $FFN$ indicates the Feed Forward Network. These steps ensure precise model attention, enhancing the contextual relevance of the resulting tokens.", + "bbox": [ + 212, + 763, + 782, + 792 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Cross-Stream Agent Attention and Agent Token Fusions. Finally, cross-stream agent attention and agent token fusion are proposed, which are explicitly tailored for the visual stream by applying the agent query attention maps and the", + "bbox": [ + 212, + 795, + 782, + 839 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "RAVAR", + "bbox": [ + 674, + 114, + 730, + 126 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "agent tokens from the other two streams. This process involves recalculating the attention maps, thereby ensuring that the final token representation is highly relevant and contextually enriched. We compute the agent query attention map for the visual token stream, where $\\gamma = VT$ , as shown in Eq. 7,", + "bbox": [ + 212, + 146, + 787, + 207 + ], + "page_idx": 9 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {M} _ {Q A} ^ {\\gamma} = \\sigma_ {c} (M a t M u l [ \\alpha * \\mathbf {A} _ {*} ^ {\\gamma}, \\mathbf {Q} ^ {\\gamma} ]). \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 377, + 219, + 785, + 237 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Then, we calculate the cross-stream irrelevance-suppressed attention for the agent query attention of the visual stream through Eq. 8 to achieve cross-stream agent attention fusion to suppress irrelevant information in the agent query attentions of the visual stream, the operations used in the second and third terms are abbreviated as C-ATT and T-ATT in Fig. 3,", + "bbox": [ + 214, + 246, + 787, + 323 + ], + "page_idx": 9 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathbf {M}} _ {Q A} ^ {\\gamma} = A V G \\left[ \\mathbf {M} _ {Q A} ^ {\\gamma}, \\sigma_ {c} (\\sum_ {\\pi} \\mathbf {M} _ {Q A} ^ {\\pi}) * \\mathbf {M} _ {Q A} ^ {\\gamma}, \\sigma_ {t} (\\sum_ {\\pi} \\mathbf {M} _ {Q A} ^ {\\pi}) * \\mathbf {M} _ {Q A} ^ {\\gamma} \\right], \\quad (8)\n$$\n", + "text_format": "latex", + "bbox": [ + 263, + 345, + 785, + 387 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "where $\\sigma_{t}$ denotes the SoftMax operation along the token dimension. $AVG$ indicates the mean operation. We follow the same procedure to calculate the cross-stream irrelevance-suppressed agent tokens before the calculation of the agent key attention, as in Eq. 9 to achieve cross-stream agent token fusion.", + "bbox": [ + 214, + 393, + 787, + 455 + ], + "page_idx": 9 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathbf {A}} _ {*} ^ {\\gamma} = A V G \\left[ \\mathbf {A} _ {*} ^ {\\gamma} + \\sigma_ {c} \\left(\\sum_ {\\pi} \\mathbf {A} _ {*} ^ {\\pi}\\right) * \\mathbf {A} _ {*} ^ {\\gamma} + \\sigma_ {t} \\left(\\sum_ {\\pi} \\mathbf {A} _ {*} ^ {\\pi}\\right) * \\mathbf {A} _ {*} ^ {\\gamma} \\right], \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 299, + 465, + 785, + 508 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "then, we calculate the agent key attention as shown in Eq. 10,", + "bbox": [ + 215, + 512, + 666, + 527 + ], + "page_idx": 9 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathbf {M}} _ {K A} ^ {\\gamma} = \\sigma_ {c} (M a t M u l [ \\alpha * \\hat {\\mathbf {A}} _ {*} ^ {\\gamma}, \\mathbf {K} ^ {\\gamma} ]). \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 372, + 537, + 785, + 564 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "The final aggregated visual tokens can be obtained through the following equation as demonstrated in Eq. 11,", + "bbox": [ + 214, + 571, + 787, + 602 + ], + "page_idx": 9 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {t} _ {*} ^ {\\gamma} = F F N (M a t M u l \\left[ \\hat {\\mathbf {M}} _ {K A} ^ {\\gamma}, M a t M u l \\left[ \\hat {\\mathbf {M}} _ {Q A} ^ {\\gamma}, \\mathbf {V} ^ {\\gamma} \\right] \\right]). \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 313, + 613, + 785, + 640 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Finally, we aggregate all the tokens from three branches by using the mean operation, where $N_{s}$ indicates the number of the stream as shown in Eq. 12,", + "bbox": [ + 214, + 648, + 787, + 679 + ], + "page_idx": 9 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {t} _ {a g g} = \\sum_ {\\phi} \\left[ \\mathbf {t} _ {*} ^ {\\phi} \\right] / N _ {s}. \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 431, + 690, + 785, + 720 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "We construct MLP-based classification and regression heads atop aggregated tokens for center frame bounding box prediction and atomic action recognition.", + "bbox": [ + 214, + 732, + 787, + 763 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.3 Loss Functions", + "text_level": 1, + "bbox": [ + 215, + 785, + 385, + 797 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "We use Binary Cross Entropy (BCE) loss and Mean Squared Error (MSE) loss for the multi-label supervision and the bounding box regression supervision in", + "bbox": [ + 214, + 809, + 787, + 840 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "K. Peng, J. Fu et al.", + "bbox": [ + 271, + 114, + 408, + 128 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "the same way as all the baselines. The equation of the BCE loss is as Eq. 13,", + "bbox": [ + 215, + 146, + 771, + 162 + ], + "page_idx": 10 + }, + { + "type": "equation", + "text": "\n$$\nL _ {B C E} (\\mathbf {y}, \\hat {\\mathbf {y}}) = - \\frac {1}{N _ {c}} \\sum_ {i = 1} ^ {N _ {c}} \\left[ \\mathbf {y} _ {i} \\log \\left(\\hat {\\mathbf {y}} _ {i}\\right) + \\left(1 - \\mathbf {y} _ {i}\\right) \\log \\left(1 - \\hat {\\mathbf {y}} _ {i}\\right) \\right], \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 299, + 174, + 785, + 213 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "where $\\mathbf{y}$ , $\\hat{\\mathbf{y}}$ and $N_{c}$ indicate the one-hot ground truth, the prediction, and the category number. The bounding boxes regression loss is expressed as Eq. 14,", + "bbox": [ + 214, + 224, + 785, + 255 + ], + "page_idx": 10 + }, + { + "type": "equation", + "text": "\n$$\nL _ {M S E} (\\mathbf {b}, \\hat {\\mathbf {b}}) = \\sum_ {j = 1} ^ {4} \\left(\\mathbf {b} _ {j} - \\hat {\\mathbf {b}} _ {j}\\right) ^ {2}, \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 395, + 267, + 785, + 308 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "where $\\mathbf{b}$ indicates the coordinates of the left top and right bottom corners, and $\\hat{\\mathbf{b}}$ represents the predicted boxes. $j$ indicates the coordinate index.", + "bbox": [ + 214, + 319, + 784, + 349 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "5 Experiments", + "text_level": 1, + "bbox": [ + 215, + 373, + 375, + 390 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "5.1 Implementation Details", + "text_level": 1, + "bbox": [ + 215, + 404, + 455, + 419 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "We conduct experiments on four NVIDIA A100 GPUs. We use BertAdam [32] optimizer with learning rate $lr = 1e^{-4}$ , batch size 128, learning rate decay 0.9, and warmup ratio 0.1. The agent number and head number in our RefAtomNet are 4 and 1. Our model has $214M$ trainable parameters. The weight for $L_{MSE}$ and $\\alpha$ are 5 and 0.125. We trained our model for 40 epochs on our dataset. The text encoder is frozen during training. We use multi-label mean Average Precision (mAP), Area Under the Receiver Operating Characteristic curve (AUROC), and the mean Intersection over Union (mIOU) as metrics, of which mIOU has less priority since we prioritize the atomic video action recognition task.", + "bbox": [ + 214, + 429, + 787, + 566 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "5.2 Experimental Results", + "text_level": 1, + "bbox": [ + 215, + 588, + 441, + 603 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Main Results. Results of the main experiments on the proposed RefAVA benchmark are summarized in Tab. 2. We compare our RefAtomNet model with a multitude of published approaches adopted to suit our task (see Section 3.2) stemming from the fields of (1) Atomic Action Localization (AAL); (2) Video Question Answering (VQA); (3) Video-Text Retrieval (VTR); (4) Single Frame (SF) baselines; and (5) Video Object Segmentation (VOS). Baseline methods from the AAL group consistently underperform in spatially localizing the action according to the textual reference (measured as mIOU). Even though the ability of the human spatial localization is not the major expected output, it can still generally illustrate if the model can grasp the correct referring person or not. Among the AAL methods, X3D [16] achieves the highest AUROC score of $59.09\\%$ and $64.51\\%$ on the val and test sets, respectively. The VQA and VTR baselines work better than the AAL baselines since they can reasonably capture the referring person while delivering acceptable atomic human action recognition performances, which benefits from the text-aware pretraining of these two", + "bbox": [ + 214, + 613, + 787, + 840 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "RAVAR", + "bbox": [ + 674, + 114, + 730, + 126 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/2c70ef408284b2411b5140944b7f7f25f927063aebeef20de835f87bc32707f7.jpg", + "table_caption": [ + "Table 2: Experimental results on our RAVAR benchmark." + ], + "table_footnote": [], + "table_body": "
MethodmIOUmAPAUROCmIOUmAPAUROC
ValTest
AALI3D [4]0.0044.0457.770.0044.6462.71
X3D [16]0.2644.4559.090.2746.3464.51
MViTv2-B [47]0.7642.3256.430.6642.6059.30
VideoMAE2-B [81]0.1642.0255.120.2941.8759.10
Hiera-B [69]0.4942.7456.720.6241.1458.70
VQASingularity [38]26.3442.4359.5229.7842.1856.12
AskAnything13B [44]20.0951.4266.1222.3552.2569.35
VTRMeVTR [95]30.7838.4251.0129.7936.2752.45
CLIP4CLIP [55]34.7539.4852.5732.3337.1755.05
XCIIIP [56]35.5442.4656.3031.7940.8258.71
BLIPv2 [42]32.9952.1366.5632.7553.1969.92
SFCLIP [67]+SAM [33]32.9549.7464.5529.8051.3469.11
CLIP [67]+DETR [3]33.9647.5761.8333.8450.9267.29
CLIP [67]+REFCLIP [29]34.4347.7962.5032.2849.9067.44
VOSSu et al. [79]23.7152.1766.6726.0253.2070.19
RefAtomNet (Ours)38.2255.9869.7336.4257.5273.95
", + "bbox": [ + 228, + 171, + 782, + 375 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "(a) Module ablation of the RefAtomNet.", + "bbox": [ + 241, + 422, + 477, + 435 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/a9d525c1085d1b3470b59d1fec739d6c0793de56cf35efc59c1e9e316db6ddb6.jpg", + "table_caption": [ + "Table 3: Experiments for module ablation and comparison with multi-modal fusion." + ], + "table_footnote": [], + "table_body": "
MethodmIOUmAPAUROCmIOUmAPAUROC
ValTest
w/o ALSAF27.3050.7065.3129.0951.2668.20
w/o CAAF36.2155.4369.6635.3855.9072.60
w/o CATF35.0153.8367.7134.5556.9673.40
w/o LSAS31.9055.2169.4731.2555.7372.38
Ours38.2255.9869.7336.4257.5273.95
", + "bbox": [ + 217, + 446, + 483, + 531 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/3128b2ae8e166bc982edeebd994c5b4be662deee189a1a86bfef357adad067e6.jpg", + "table_caption": [ + "(b) Comparison with other fusion approaches." + ], + "table_footnote": [], + "table_body": "
FusionmIOU mAP AUROCmIOU mAP AUROC
ValTest
Addition27.3050.7065.3129.0951.2668.20
Concatenation18.6452.2366.4520.6553.4470.70
Multiplication23.9051.5565.6625.0553.3370.48
AttentionBottleneck [59]33.4750.9765.0733.0254.0271.08
McOmut [98]23.8851.5865.6525.0253.2170.42
Ours38.2255.9869.7336.4257.5273.95
", + "bbox": [ + 501, + 446, + 785, + 531 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "tasks. AskAnything [44] from the VQA group achieves $20.09\\%$ and $22.35\\%$ of mIOU, $51.42\\%$ and $52.25\\%$ of mAP, and $66.12\\%$ and $69.35\\%$ of AUROC on the val and test sets. BLIPv2 [43] from the VTR group delivers $32.99\\%$ and $32.75\\%$ of mIOU, $52.13\\%$ and $53.19\\%$ of mAP, and $66.56\\%$ and $69.92\\%$ of AUROC, on the val and test sets, respectively. Our RefAtomNet achieves state-of-the-art performances, outperforming the best baseline BLIPv2 [42] by $5.23\\%$ , $3.85\\%$ , $3.17\\%$ and $3.67\\%$ , $4.33\\%$ , $4.03\\%$ of mIOU, mAP, and AUROC, on val and test sets.", + "bbox": [ + 212, + 565, + 787, + 671 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Ablations of the Individual Modules. In Tab. 3a, we show the ablation for the components of RefAtomNet by removing each of the proposed designs, where LSAS indicates the location-semantic stream, CAAF indicates the cross-stream agent attention fusion, CATF indicates the cross-stream agent token fusion, and $w / o$ ALSAF indicates by simply using addition to fuse these three streams and without the agent-based location-semantic aware attentional fusion mechanism. Compared with the ablation $w / o$ ALSAF, RefAtomNet achieves promising improvements of $10.92\\%$ , $5.28\\%$ , $4.42\\%$ and $7.33\\%$ , $6.26\\%$ , $5.75\\%$ in terms of mIOU, mAP, and AUROC for the val and test sets, respectively. It indicates that using simple aggregation of the three streams makes the model distract from the important cues for the referring person, illustrated by the large decay of mIOU metric.", + "bbox": [ + 212, + 672, + 787, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "K. Peng, J. Fu et al.", + "bbox": [ + 271, + 114, + 408, + 127 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "(a) Generalizability to test time rephrasing", + "bbox": [ + 233, + 174, + 485, + 186 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/20e42e2cb85ff0303bc5f3bddd69fdb6a22cf36e2d20b100b29cdabc52cede9d.jpg", + "table_caption": [ + "Table 4: Generalizability to different referring styles and encoder architectures." + ], + "table_footnote": [], + "table_body": "
FusionmIOUmAPAUROCmIOUmAPAUROC
ValTest
Singularity [38]18.4541.2758.4720.5441.3955.32
XCLIP [56]31.9541.3554.3529.8440.7458.45
AskAnything [44]19.7451.1165.8321.6051.9669.04
BLIPv2 [42]31.0051.4565.8831.3452.3568.87
Ours34.6555.7569.5233.1457.2373.76
", + "bbox": [ + 215, + 199, + 496, + 280 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/669b5eca744268dc88dc0c7ec4c8c539bf6f7a9ed22de31c47786878ebf09763.jpg", + "table_caption": [ + "(b) Generalizability to different visual-textual encoder architectures." + ], + "table_footnote": [], + "table_body": "
FusionmIOU mAP AUROCmIOU mAP AUROC
ValTest
XCLIP [56]35.5442.4656.3031.7940.8258.71
RefAtomNet (XCLIP)38.5947.4061.2036.6148.5966.47
BLIPv2 [42]32.9952.1366.5632.7553.1969.92
RefAtomNet (BLIPv2)38.2255.9869.7336.4257.5273.95
", + "bbox": [ + 501, + 210, + 784, + 279 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Compared with the ablation $w / o$ LSAS, we find that the location-semantic tokens benefit more for the localization of the correct person in the center frame. The advantages of the LSAS on atomic video action recognition metrics are highlighted in the test set which has higher scenario diversity compared with the val set. Using CAAF and CATF, each of them brings promising benefits by suppressing the irrelevant information in the visual tokens.", + "bbox": [ + 212, + 310, + 782, + 400 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Comparison with Other Fusion Mechanisms. We compare our proposed agent-based location-semantic aware attentional fusion mechanism with late fusion addition, multiplication, concatenation, AttentionBottleNeck [59], and McOmet [98] in Tab. 3b. Our approach outperforms all by suppressing those irrelevant visual tokens on the agent tokens and attention masks.", + "bbox": [ + 212, + 401, + 784, + 474 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Generalizability to Test-time Reference Rephrasing. The reference sentences given by different users may differ in daily life scenarios. To test the generalizability of the model towards different referring styles, we invoke API (gpt-3.5-turbo) of ChatGPT [60] to rephrase the test set description two times and then deliver the averaged performance for RAVAR on the original val and test sets, and the two rephrased val and test sets, where we select 4 most out-performing baselines from our benchmark to construct this ablation study, i.e., Singularity [38], AskAnything [44], XCLIP [56], and BLIPv2 [42], as shown in Tab. 4a. RefAtomNet delivers the best performances. We find out that the test time reference rephrasing will cause performance decay for localization and less decay for the atomic video-based action recognition delivered by our model.", + "bbox": [ + 212, + 477, + 785, + 642 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Generalizability to Different Visual Textual Backbones. Since we use the best-performing baseline BLIPv2 [42] as the visual and textual encoder in our model, it would be interesting to see if the proposed architecture can generalize to different encoder backbones. We thereby demonstrate another ablation study in Tab. 4b, where we equip our RefAtomNet with the XCLIP [56] backbone. Compared with the XCLIP baseline, the RefAtomNet (XCLIP) achieves $3.05\\%$ , $4.94\\%$ , $4.90\\%$ and $4.82\\%$ , $7.77\\%$ , $7.76\\%$ performance improvements in terms of mIOU, mAP, and AUROC on the val and test sets, respectively, showing the great generalizability of our RefAtomNet of the visual and textual encoder.", + "bbox": [ + 212, + 643, + 785, + 779 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Analysis of the Qualitative Results. The qualitative results are delivered in Fig. 4. These examples demonstrate the effectiveness of RefAtomNet compared with the best-performing baseline BLIPv2 [42]. Samples 1 and 8 validate the recognition quality for references without any location indications.", + "bbox": [ + 212, + 779, + 785, + 839 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "RAVAR", + "bbox": [ + 674, + 114, + 730, + 126 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "1 Text reference: The man wearing light green clothes.", + "bbox": [ + 222, + 145, + 444, + 155 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/5509deddc417e1eb431442e202ffcd0047277d50bc829982fa456a8dd455d53f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 223, + 155, + 498, + 227 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "2 Text reference: The man in the center who has gray beard.", + "bbox": [ + 501, + 145, + 748, + 155 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/3d93bf92296f2b83a4e51ab679f12d33af7670d35defd592d8e7224ea981eb67.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 503, + 155, + 779, + 227 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "3 Text reference: The man on the left who is in blue clothes.", + "bbox": [ + 222, + 227, + 467, + 237 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/207de4bb67adb1aa77ee3354b0104f7ccbcd8366ed9c26cdc90b3f59403846e4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 223, + 237, + 498, + 310 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "4 Text reference: The first man on the right.", + "bbox": [ + 501, + 228, + 681, + 238 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/6e9a141ec02e5100c1693d0b670e3ca15aa7134091578f1895fab80c8ca357fc.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 503, + 238, + 779, + 310 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Text reference: The man wearing glasses on the left.", + "bbox": [ + 223, + 310, + 444, + 320 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/77ae1422bb388a1c7b8712011e9962678e3db58fd7a8ff2050470a04794e7f43.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 223, + 320, + 498, + 392 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "6 Text reference: The man on the right back to us.", + "bbox": [ + 503, + 310, + 707, + 320 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/48870fa17142608cb2eaf39c474afb14562293ea15a75a2991bc83d7fff36449.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 503, + 320, + 779, + 392 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Text reference: The woman wearing black clothes and pearls necklace on the right.", + "bbox": [ + 223, + 392, + 478, + 411 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/1d1a36e5d72917f6a1a8a4ff9579bda5481aea59dc6f4168360dc80ca8c4f54b.jpg", + "image_caption": [ + "Fig. 4: An overview of qualitative results. Missed predictions are marked with a red cross, while true positive and false positive predictions are shown in green and red." + ], + "image_footnote": [], + "bbox": [ + 223, + 411, + 498, + 492 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "8 Text reference: The man wearing black suits.", + "bbox": [ + 501, + 392, + 694, + 402 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/28215095954289f8843092093113a56a18904c32160aaa8f6acef65821b6f392.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 503, + 411, + 779, + 492 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Samples 3, 4, 5, and 6 showcase the results for different humans in a shared scene. Samples 2 and 7 show the RAVAR performance for references containing finer details, e.g., the necklace type and the beard color. Our RefAtomNet outperforms the BLIPv2 baseline on selected samples obviously, attributable to its superior capability in eliminating extraneous visual cues and its advanced location-semantic reasoning prowess.", + "bbox": [ + 212, + 561, + 787, + 654 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "6 Conclusions", + "text_level": 1, + "bbox": [ + 215, + 674, + 369, + 690 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In this work, we introduce a novel task called Referring Atomic Video Action Recognition (RAVAR). We establish the RefAVA dataset to address the challenge of identifying atomic actions for individuals of interest in videos based on textual descriptions. Existing methods exhibit poor performance on this new task, prompting the development of RefAtomNet, a vision-language architecture that effectively integrates cross-stream tokens for precise referring atomic video action recognition. Through reference-relevant token enhancement and comprehensive token integration, RefAtomNet achieves impressive results, highlighting its effectiveness in tackling the complexities of the RAVAR task.", + "bbox": [ + 212, + 703, + 787, + 840 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "K. Peng, J. Fu et al.", + "bbox": [ + 271, + 114, + 408, + 128 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 217, + 143, + 401, + 162 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "The project served to prepare the SFB 1574 Circular Factory for the Perpetual Product (project ID: 471687386), approved by the German Research Foundation (DFG, German Research Foundation) with a start date of April 1, 2024. This work was also partially supported in part by the SmartAge project sponsored by the Carl Zeiss Stiftung (P2019-01-003; 2021-2026). This work was performed on the HoreKa supercomputer funded by the Ministry of Science, Research and the Arts Baden-Württemberg and by the Federal Ministry of Education and Research. The authors also acknowledge support by the state of Baden-Württemberg through bwHPC and the German Research Foundation (DFG) through grant INST 35/1597-1 FUGG. This project is also supported by the National Key RD Program under Grant 2022YFB4701400. Lastly, the authors thank for the support of Dr. Sepideh Pashami, the Swedish Innovation Agency VINNOVA, the Digital Futures.", + "bbox": [ + 212, + 176, + 787, + 373 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 215, + 395, + 321, + 410 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Bagad, P., Tapaswi, M., Snoek, C.G.M.: Test of time: Instilling video-language models with a sense of time. In: CVPR (2023)", + "2. Bu, Y., Li, L., Xie, J., Liu, Q., Cai, Y., Huang, Q., Li, Q.: Scene-text oriented referring expression comprehension. TMM (2022)", + "3. Carion, N., Massa, F., Synnaeve, G., Usunier, N., Kirillov, A., Zagoruyko, S.: End-to-end object detection with transformers. In: ECCV (2020)", + "4. Carreira, J., Zisserman, A.: Quo vadis, action recognition? A new model and the kinetics dataset. In: CVPR (2017)", + "5. Castro, S., Deng, N., Huang, P., Burzo, M., Mihalcea, R.: In-the-wild video question answering. In: COLING (2022)", + "6. Chai, W., Guo, X., Wang, G., Lu, Y.: StableVideo: Text-driven consistency-aware diffusion video editing. In: ICCV (2023)", + "7. Chen, J., Zhu, D., Haydarov, K., Li, X., Elhoseiny, M.: Video ChatCaptioner: Towards enriched spatiotemporal descriptions. arXiv preprint arXiv:2304.04227 (2023)", + "8. Chen, X., Fang, H., Lin, T.Y., Vedantam, R., Gupta, S., Dolkar, P., Zitnick, C.L.: Microsoft COCO captions: Data collection and evaluation server. arXiv preprint arXiv:1504.00325 (2015)", + "9. Chen, Y., Wang, J., Lin, L., Qi, Z., Ma, J., Shan, Y.: Tagging before alignment: Integrating multi-modal tags for video-text retrieval. arXiv preprint arXiv:2301.12644 (2023)", + "0. Chen, Z., Ma, L., Luo, W., Wong, K.Y.K.: Weakly-supervised spatio-temporally grounding natural sentence in video. arXiv preprint arXiv:1906.02549 (2019)", + "1. Chung, J., Wu, C.h., Yang, H.r., Tai, Y.W., Tang, C.K.: HAA500: Human-centric atomic action dataset with curated videos. In: ICCV (2021)", + "2. Dang, R., Feng, J., Zhang, H., Ge, C., Song, L., Gong, L., Liu, C., Chen, Q., Zhu, F., Zhao, R., Song, Y.: InstructDET: Diversifying referring object detection with generalized instructions. arXiv preprint arXiv:2310.05136 (2023)", + "3. Deruyttere, T., Vandenhende, S., Grujicic, D., Van Gool, L., Moens, M.F.: Talk2Car: Taking control of your self-driving car. In: EMNLP (2019)" + ], + "bbox": [ + 225, + 425, + 784, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "RAVAR", + "bbox": [ + 674, + 114, + 730, + 126 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "14. Devlin, J., Chang, M.W., Lee, K., Toutanova, K.: BERT: Pre-training of deep bidirectional transformers for language understanding. In: ACL (2019)", + "15. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., Uszkoreit, J., Houlsby, N.: An image is worth 16x16 words: Transformers for image recognition at scale. In: ICLR (2021)", + "16. Feichtenhofer, C.: X3D: Expanding architectures for efficient video recognition. In: CVPR (2020)", + "17. Feichtenhofer, C., Fan, H., Malik, J., He, K.: SlowFast networks for video recognition. In: ICCV (2019)", + "18. Gandhi, M., Gul, M.O., Prakash, E., Grunde-McLaughlin, M., Krishna, R., Agrawala, M.: Measuring compositional consistency for video question answering. In: CVPR (2022)", + "19. Gao, D., Zhou, L., Ji, L., Zhu, L., Yang, Y., Shou, M.Z.: MIST: Multi-modal iterative spatial-temporal transformer for long-form video question answering. In: CVPR (2023)", + "20. Garcia, N., Otani, M., Chu, C., Nakashima, Y.: KnowIT VQA: Answering knowledge-based questions about videos. In: AAAI (2020)", + "21. Gavrilyuk, K., Ghodrati, A., Li, Z., Snoek, C.G.: Actor and action video segmentation from a sentence. In: CVPR (2018)", + "22. Goyal, R., Kahou, S.E., Michalski, V., Materzynska, J., Westphal, S., Kim, H., Haenel, V., Fründ, I., Yianilos, P., Mueller-Freitag, M., Hoppe, F., Thurau, C., Bax, I., Memisevic, R.: The \"something something\" video database for learning and evaluating visual common sense. In: ICCV (2017)", + "23. Gritsenko, A., Xiong, X., Djolonga, J., Dehghani, M., Sun, C., Lucic, M., Schmid, C., Arnab, A.: End-to-end spatio-temporal action localisation with video transformers. arXiv preprint arXiv:2304.12160 (2023)", + "24. Gu, C., Sun, C., Ross, D.A., Vondrick, C., Pantofaru, C., Li, Y., Vijayanarasimhan, S., Toderici, G., Ricco, S., Sukthankar, R., Schmid, C., Malik, J.: AVA: A video dataset of spatio-temporally localized atomic visual actions. In: CVPR (2018)", + "25. Guo, W., Zhang, Y., Yang, J., Yuan, X.: Re-attention for visual question answering. TIP (2021)", + "26. Han, D., Ye, T., Han, Y., Xia, Z., Song, S., Huang, G.: Agent attention: On the integration of softmax and linear attention. arXiv preprint arXiv:2312.08874 (2023)", + "27. Ji, Y., Zhan, Y., Yang, Y., Xu, X., Shen, F., Shen, H.T.: A context knowledge map guided coarse-to-fine action recognition. TIP (2020)", + "28. Jiang, J., Chen, Z., Lin, H., Zhao, X., Gao, Y.: Divide and conquer: Question-guided spatio-temporal contextual attention for video question answering. In: AAAI (2020)", + "29. Jin, L., Luo, G., Zhou, Y., Sun, X., Jiang, G., Shu, A., Ji, R.: Refclip: A universal teacher for weakly supervised referring expression comprehension. In: CVPR (2023)", + "30. Khoreva, A., Rohrbach, A., Schiele, B.: Video object segmentation with language referring expressions. In: ACCV (2019)", + "31. Kim, M., Spinola, F., Benz, P., Kim, T.h.: A*: Atrous spatial temporal action recognition for real time applications. In: WACV (2024)", + "32. Kingma, D.P., Ba, J.: Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980 (2014)", + "33. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., et al.: Segment anything. In: CVPR (2023)", + "34. Kuehne, H., Jhuang, H., Garrote, E., Poggio, T., Serre, T.: HMDB: A large video database for human motion recognition. In: ICCV (2011)" + ], + "bbox": [ + 215, + 146, + 784, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "K. Peng, J. Fu et al.", + "bbox": [ + 271, + 114, + 408, + 127 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "35. Laput, G., Harrison, C.: Sensing fine-grained hand activity with smartwatches. In: CHI (2019)", + "36. Le, T.M., Le, V., Venkatesh, S., Tran, T.: Hierarchical conditional relation networks for video question answering. In: CVPR (2020)", + "37. Lea, C., Vidal, R., Hager, G.D.: Learning convolutional action primitives for fine-grained action recognition. In: ICRA (2016)", + "38. Lei, J., Berg, T.L., Bansal, M.: Revealing single frame bias for video-and-language learning. arXiv preprint arXiv:2206.03428 (2022)", + "39. Lei, J., Li, L., Zhou, L., Gan, Z., Berg, T.L., Bansal, M., Liu, J.: Less is more: ClipBERT for video-and-language learning via sparse sampling. In: CVPR (2021)", + "40. Li, G., Wei, Y., Tian, Y., Xu, C., Wen, J.R., Hu, D.: Learning to answer questions in dynamic audio-visual scenarios. In: CVPR (2022)", + "41. Li, J., Niu, L., Zhang, L.: From representation to reasoning: Towards both evidence and commonsense reasoning for video question-answering. In: CVPR (2022)", + "42. Li, J., Li, D., Savarese, S., Hoi, S.: BLIP-2: Bootstrapping language-image pretraining with frozen image encoders and large language models. In: ICML (2023)", + "43. Li, J., Li, D., Xiong, C., Hoi, S.: BLIP: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In: ICML (2022)", + "44. Li, K., He, Y., Wang, Y., Li, Y., Wang, W., Luo, P., Wang, Y., Wang, L., Qiao, Y.: VideoChat: Chat-centric video understanding. arXiv preprint arXiv:2305.06355 (2023)", + "45. Li, L., Gan, Z., Lin, K., Lin, C.C., Liu, Z., Liu, C., Wang, L.: LAVENDER: Unifying video-language understanding as masked language modeling. In: CVPR (2023)", + "46. Li, R., Li, K., Kuo, Y.C., Shu, M., Qi, X., Shen, X., Jia, J.: Referring image segmentation via recurrent refinement networks. In: CVPR (2018)", + "47. Li, Y., Wu, C.Y., Fan, H., Mangalam, K., Xiong, B., Malik, J., Feichtenhofer, C.: MViTv2: Improved multiscale vision transformers for classification and detection. In: CVPR (2022)", + "48. Lin, J., Chen, J., Peng, K., He, X., Li, Z., Stiefelhagen, R., Yang, K.: EchoTrack: Auditory referring multi-object tracking for autonomous driving. arXiv preprint arXiv:2402.18302 (2024)", + "49. Lin, X., Tiwari, S., Huang, S., Li, M., Shou, M.Z., Ji, H., Chang, S.F.: Towards fast adaptation of pretrained contrastive models for multi-channel video-language retrieval. In: CVPR (2023)", + "50. Liu, J., Wang, L., Yang, M.H.: Referring expression generation and comprehension via attributes. In: ICCV (2017)", + "51. Liu, R., Zhang, J., Peng, K., Zheng, J., Cao, K., Chen, Y., Yang, K., Stiefelhagen, R.: Open scene understanding: Grounded situation recognition meets segment anything for helping people with visual impairments. In: ICCVW (2023)", + "52. Liu, R., Liu, C., Bai, Y., Yuille, A.L.: CLEVR-Ref+: Diagnosing visual reasoning with referring expressions. In: CVPR (2019)", + "53. Liu, S., Hui, T., Huang, S., Wei, Y., Li, B., Li, G.: Cross-modal progressive comprehension for referring segmentation. TPAMI (2021)", + "54. Liu, Y., Li, G., Lin, L.: Cross-modal causal relational reasoning for event-level visual question answering. TPAMI (2023)", + "55. Luo, H., Ji, L., Zhong, M., Chen, Y., Lei, W., Duan, N., Li, T.: CLIP4Clip: An empirical study of CLIP for end to end video clip retrieval and captioning. Neurocomputing (2022)", + "56. Ma, Y., Xu, G., Sun, X., Yan, M., Zhang, J., Ji, R.: X-CLIP: End-to-end multigrained contrastive learning for video-text retrieval. In: MM (2022)" + ], + "bbox": [ + 212, + 146, + 784, + 839 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "RAVAR", + "bbox": [ + 674, + 114, + 730, + 126 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "57. Madasu, A., Aflalo, E., Ben Melech Stan, G., Tseng, S.Y., Bertasius, G., Lal, V.: Improving video retrieval using multilingual knowledge transfer. In: ECIR (2023)", + "58. McIntosh, B., Duarte, K., Rawat, Y.S., Shah, M.: Visual-textual capsule routing for text-based video segmentation. In: CVPR (2020)", + "59. Nagrani, A., Yang, S., Arnab, A., Jansen, A., Schmid, C., Sun, C.: Attention bottlenecks for multimodal fusion. In: NeuIPS (2021)", + "60. OpenAI: ChatGPT: Optimizing language models for dialogue. https://openai.com/ (2022)", + "61. Ordonez, V., Kulkarni, G., Berg, T.: Im2Text: Describing images using 1 million captioned photographs. In: NeurIPS (2011)", + "62. Ou, W., Zhang, J., Peng, K., Yang, K., Jaworek, G., Müller, K., Stiefelhagen, R.: Indoor navigation assistance for visually impaired people via dynamic SLAM and panoptic segmentation with an RGB-D sensor. In: ICCHP (2022)", + "63. Peng, K., Roitberg, A., Yang, K., Zhang, J., Stiefelhagen, R.: TransDARC: Transformer-based driver activity recognition with latent space feature calibration. In: IROS (2022)", + "64. Pramanick, P., Sarkar, C., Paul, S., dev Roychoudhury, R., Bhowmick, B.: DoRO: Disambiguation of referred object for embodied agents. RA-L (2022)", + "65. Pramono, R.R.A., Chen, Y.T., Fang, W.H.: Spatial-temporal action localization with hierarchical self-attention. TMM (2021)", + "66. Qiu, H., Li, H., Wu, Q., Meng, F., Shi, H., Zhao, T., Ngan, K.N.: Language-aware fine-grained object representation for referring expression comprehension. In: MM (2020)", + "67. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., Krueger, G., Sutskever, I.: Learning transferable visual models from natural language supervision. In: ICML (2021)", + "68. Rajasegaran, J., Pavlakos, G., Kanazawa, A., Feichtenhofer, C., Malik, J.: On the benefits of 3D pose and tracking for human action recognition. In: CVPR (2023)", + "69. Ryali, C., Hu, Y., Bolya, D., Wei, C., Fan, H., Huang, P., Aggarwal, V., Chowdhury, A., Poursaeed, O., Hoffman, J., Malik, J., Li, Y., Feichtenhofer, C.: Hiera: A hierarchical vision transformer without the bells-and-whistles. In: ICML (2023)", + "70. Saha, J., Chowdhury, C., Chowdury, I.R., Roy, P.: Fine grained activity recognition using smart handheld. In: ICDCN (2018)", + "71. Seibold, C.M., Reif, S., Kleesiek, J., Stiefelhagen, R.: Reference-guided pseudolabel generation for medical semantic segmentation. In: AAAI (2022)", + "72. Seo, S., Lee, J.Y., Han, B.: URVOS: Unified referring video object segmentation network with a large-scale benchmark. In: ECCV (2020)", + "73. Shao, D., Zhao, Y., Dai, B., Lin, D.: Finegym: A hierarchical video dataset for fine-grained action understanding. In: CVPR (2020)", + "74. Sharma, P., Ding, N., Goodman, S., Soricut, R.: Conceptual captions: A cleaned, hypernymed, image alt-text dataset for automatic image captioning. In: ACL (2018)", + "75. Shi, H., Pan, W., Zhao, Z., Zhang, M., Wu, F.: Unsupervised domain adaptation for referring semantic segmentation. In: MM (2023)", + "76. Shi, H., Li, H., Meng, F., Wu, Q.: Key-word-aware network for referring expression image segmentation. In: ECCV (2018)", + "77. Shi, Y., Xu, H., Yuan, C., Li, B., Hu, W., Zha, Z.J.: Learning video-text aligned representations for video captioning. TOMM (2023)", + "78. Soomro, K., Zamir, A.R., Shah, M.: UCF101: A dataset of 101 human actions classes from videos in the wild. arXiv preprint arXiv:1212.0402 (2012)" + ], + "bbox": [ + 215, + 146, + 784, + 839 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "K. Peng, J. Fu et al.", + "bbox": [ + 271, + 114, + 408, + 127 + ], + "page_idx": 17 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "79. Su, Y., Wang, W., Liu, J., Ma, S., Yang, X.: Sequence as a whole: A unified framework for video action localization with long-range text query. TIP (2023)", + "80. Vasudevan, A.B., Dai, D., Van Gool, L.: Object referring in videos with language and human gaze. In: CVPR (2018)", + "81. Wang, L., Huang, B., Zhao, Z., Tong, Z., He, Y., Wang, Y., Wang, Y., Qiao, Y.: VideoMAE V2: Scaling video masked autoencoders with dual masking. In: CVPR (2023)", + "82. Wang, M., Xing, J., Mei, J., Liu, Y., Jiang, Y.: ActionCLIP: Adapting language-image pretrained models for video action recognition. TNNLS (2023)", + "83. Wang, R., Chen, D., Wu, Z., Chen, Y., Dai, X., Liu, M., Yuan, L., Jiang, Y.G.: Masked video distillation: Rethinking masked feature modeling for self-supervised video representation learning. In: CVPR (2023)", + "84. Wang, S., Yan, R., Huang, P., Dai, G., Song, Y., Shu, X.: Com-STAL: Compositional spatio-temporal action localization. TCSVT (2023)", + "85. Wang, Y., Li, K., Li, Y., He, Y., Huang, B., Zhao, Z., Zhang, H., Xu, J., Liu, Y., Wang, Z., Xing, S., Chen, G., Pan, J., Yu, J., Wang, Y., Wang, L., Qiao, Y.: InternVideo: General video foundation models via generative and discriminative learning. arXiv preprint arXiv:2212.03191 (2022)", + "86. Wu, D., Han, W., Wang, T., Dong, X., Zhang, X., Shen, J.: Referring multi-object tracking. In: CVPR (2023)", + "87. Wu, W., Luo, H., Fang, B., Wang, J., Ouyang, W.: Cap4Video: What can auxiliary captions do for text-video retrieval? In: CVPR (2023)", + "88. Xiao, J., Shang, X., Yao, A., Chua, T.S.: NExT-QA: Next phase of question-answering to explaining temporal actions. In: CVPR (2021)", + "89. Yang, A., Miech, A., Sivic, J., Laptev, I., Schmid, C.: Just ask: Learning to answer questions from millions of narrated videos. In: ICCV (2021)", + "90. Yang, P., Wang, X., Duan, X., Chen, H., Hou, R., Jin, C., Zhu, W.: AVQA: A dataset for audio-visual question answering on videos. In: MM (2022)", + "91. Yu, L., Poirson, P., Yang, S., Berg, A.C., Berg, T.L.: Modeling context in referring expressions. In: ECCV (2016)", + "92. Yu, Z., Xu, D., Yu, J., Yu, T., Zhao, Z., Zhuang, Y., Tao, D.: ActivityNet-QA: A dataset for understanding complex web videos via question answering. In: AAAI (2019)", + "93. Yuan, Z., Yan, X., Liao, Y., Zhang, R., Wang, S., Li, Z., Cui, S.: InstanceRefer: Cooperative holistic understanding for visual grounding on point clouds through instance multi-level contextual referring. In: ICCV (2021)", + "94. Zeng, F., Dong, B., Zhang, Y., Wang, T., Zhang, X., Wei, Y.: MOTR: End-to-end multiple-object tracking with transformer. In: ECCV (2022)", + "95. Zhang, G., Ren, J., Gu, J., Tresp, V.: Multi-event video-text retrieval. In: CVPR (2023)", + "96. Zhang, H., Li, X., Bing, L.: Video-LLaMA: An instruction-tuned audio-visual language model for video understanding. In: EMNLP (2023)", + "97. Zheng, J., Zhang, J., Yang, K., Peng, K., Stiefelhagen, R.: MateRobot: Material recognition in wearable robotics for people with visual impairments. In: ICRA (2024)", + "98. Zong, D., Sun, S.: McOmet: Multimodal fusion transformer for physical audiovisual commonsense reasoning. In: AAAI (2023)" + ], + "bbox": [ + 212, + 146, + 784, + 797 + ], + "page_idx": 18 + }, + { + "type": "header", + "text": "RAVAR", + "bbox": [ + 674, + 114, + 730, + 126 + ], + "page_idx": 18 + }, + { + "type": "page_number", + "text": "19", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 18 + } +] \ No newline at end of file diff --git a/2024/Referring Atomic Video Action Recognition/3c97fc3b-c407-40be-ae83-9d7a4bd31254_model.json b/2024/Referring Atomic Video Action Recognition/3c97fc3b-c407-40be-ae83-9d7a4bd31254_model.json new file mode 100644 index 0000000000000000000000000000000000000000..e885c7561fd31acdf0713af413b252b1eb3ed2e9 --- /dev/null +++ b/2024/Referring Atomic Video Action Recognition/3c97fc3b-c407-40be-ae83-9d7a4bd31254_model.json @@ -0,0 +1,3098 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.251, + 0.142, + 0.753, + 0.163 + ], + "angle": 0, + "content": "Referring Atomic Video Action Recognition" + }, + { + "type": "text", + "bbox": [ + 0.227, + 0.189, + 0.776, + 0.237 + ], + "angle": 0, + "content": "Kunyu Peng\\(^{1,*}\\), Jia Fu\\(^{3,4,*}\\), Kailun Yang\\(^{2,\\dagger}\\), Di Wen\\(^{1}\\), Yufan Chen\\(^{1}\\), Ruiping Liu\\(^{1}\\), Junwei Zheng\\(^{1}\\), Jiaming Zhang\\(^{1}\\), M. Saqib Sarfraz\\(^{1,6}\\), Rainer Stiefelhagen\\(^{1}\\), and Alina Roitberg\\(^{5}\\)" + }, + { + "type": "text", + "bbox": [ + 0.25, + 0.247, + 0.752, + 0.291 + ], + "angle": 0, + "content": "\\(^{1}\\)Karlsruhe Institute of Technology, \\(^{2}\\)Hunan University, \\(^{3}\\)RISE Research Institutes of Sweden, \\(^{4}\\)KTH Royal Institute of Technology, \\(^{5}\\)University of Stuttgart, \\(^{6}\\)Mercedes-Benz Tech Innovation" + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.317, + 0.744, + 0.651 + ], + "angle": 0, + "content": "Abstract. We introduce a new task called Referring Atomic Video Action Recognition (RAVAR), aimed at identifying atomic actions of a particular person based on a textual description and the video data of this person. This task differs from traditional action recognition and localization, where predictions are delivered for all present individuals. In contrast, we focus on recognizing the correct atomic action of a specific individual, guided by text. To explore this task, we present the RefAVA dataset, containing 36,630 instances with manually annotated textual descriptions of the individuals. To establish a strong initial benchmark, we implement and validate baselines from various domains, e.g., atomic action localization, video question answering, and text-video retrieval. Since these existing methods underperform on RAVAR, we introduce RefAtomNet - a novel cross-stream attention-driven method specialized for the unique challenges of RAVAR: the need to interpret a textual referring expression for the targeted individual, utilize this reference to guide the spatial localization and harvest the prediction of the atomic actions for the referring person. The key ingredients are: (1) a multi-stream architecture that connects video, text, and a new location-semantic stream, and (2) cross-stream agent attention fusion and agent token fusion which amplify the most relevant information across these streams and consistently surpasses standard attention-based fusion on RAVAR. Extensive experiments demonstrate the effectiveness of RefAtomNet and its building blocks for recognizing the action of the described individual. The dataset and code will be made publicly available at RAVAR." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.671, + 0.377, + 0.687 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.699, + 0.788, + 0.807 + ], + "angle": 0, + "content": "Referring scene understanding [50, 52, 66, 93] aims to solve the underlying computer vision task for the particular scene element specified via a natural language referring expression. Context-driven queries are vital in information retrieval, assistive systems, and multimedia analysis, but incorporating them into computer vision models is challenging due to a complex entanglement of linguistic descriptions, localization, and visual recognition itself. Several benchmarks for referring scene understanding have been successfully established for multi-object" + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.812, + 0.363, + 0.825 + ], + "angle": 0, + "content": "* Equal contribution" + }, + { + "type": "page_footnote", + "bbox": [ + 0.223, + 0.825, + 0.509, + 0.84 + ], + "angle": 0, + "content": "† Correspondence: kailun.yang@hnu.edu.cn" + }, + { + "type": "list", + "bbox": [ + 0.218, + 0.812, + 0.509, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.41, + 0.129 + ], + "angle": 0, + "content": "K. Peng, J. Fu et al." + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.146, + 0.784, + 0.239 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.277, + 0.252, + 0.725, + 0.266 + ], + "angle": 0, + "content": "Fig. 1: A comparison of AAL Task (left) and RAVAR task (right)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.296, + 0.787, + 0.401 + ], + "angle": 0, + "content": "tracking [86], semantic segmentation [30, 46, 72, 75, 76], medical imaging [71], and object detection [12, 64]. While these efforts focus on object-centric referring expressions, many applications require a human-centric understanding of human actions, e.g., in rehabilitation assistance [35, 70] and human-robot interaction [27, 37]. Our work addresses a new problem - reference-driven recognition of atomic actions performed by the human described via a textual reference, which has been overlooked in the past." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.403, + 0.787, + 0.674 + ], + "angle": 0, + "content": "The field of human action recognition has made exciting advances, from new approaches stemming from the rise of visual transformers [23,47,63,69,81,83,85] to new large-scale datasets [4,22,34,73,78]. Datasets utilized for the prediction of atomic actions often feature recordings with multiple individuals [24]. Nevertheless, the vast majority of previously published works [23,31,68,69,81] either rely on manually window crop of a particular person of interest or automatically generated region of interests to predict the atomic actions for all the individuals in one video, which results in large pre- or post-processing effort to focus on one specific person. This workflow is also very inconvenient in certain applications, such as assistive systems for users with visual impairments [51,62,97], who must understand the state of each present in the scene for effective interactions. Leveraging succinct textual descriptions, which may include broad positional indicators (e.g., left, center, and right), appearance attributes (e.g., hair color and clothing), or gender, to steer the model towards delivering atomic action recognition outcomes for the targeted individual, presents a promising solution to the outlined challenges. Employing these textual cues as a reference, the model can facilitate an end-to-end retrieval of the specified instance across the entire video, subsequently providing precise atomic action recognition results." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.675, + 0.788, + 0.84 + ], + "angle": 0, + "content": "To close this gap, we formalize the new task of Referring Atomic Video Action Recognition (RAVAR). The differences between RAVAR and atomic action localization are shown in Fig. 1. The conventional atomic action recognition in multi-person scenarios is often formalized in an action localization manner, and requires the region of interest to localize the human available a priori, after which atomic action recognition is carried out for each individual. Post-processing is required for individuals of interest. In contrast, RAVAR assumes a textual reference and a video as inputs, delivering the referring individual's atomic actions along with the location. This comparison shows the efficiency of RAVAR, leveraging textual information for less labor-intensive subject identification to assist atomic action analysis for the individual of interest. We further" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.676, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "RAVAR" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.787, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.237 + ], + "angle": 0, + "content": "introduce the new RefAVA dataset for RAVAR. RefAVA is established using the 17,946 video clips of the public AVA dataset [24] initially designed for atomic action localization, which we extend with 36,630 textual descriptions of the corresponding individuals used as the referring expression. The dataset covers diverse settings, e.g., in/out door and day/night time scenarios, and numerous multi-person scenes, constituting an ideal testbed for our task." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.24, + 0.789, + 0.389 + ], + "angle": 0, + "content": "To establish an initial benchmark, we assess 15 well-established approaches from multiple related research domains. Our experiments reveal that none of the benchmarked methods delivers sufficient performance. For example, previous methods within the domains of VQA and VTR typically yield predictions of a coarse analysis regarding actions, whereas strategies in AAL face challenges in spatially localizing the correct referring individual. We attribute this phenomenon to the massive irrelevant information provided by the visual data, distracting the model from the referring location. How to suppress the referring irrelevant information becomes a critical challenge to pursue better RAVAR performances." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.393, + 0.789, + 0.665 + ], + "angle": 0, + "content": "To address this, we introduce the new RefAtomNet approach specialized in simultaneously solving the tasks of understanding textual referring expressions and using them as guides for spatial localization and classification of atomic actions. We begin by computing textual, visual, and newly proposed location-semantic tokens using a large foundation model and well-established object detector. The latter stream is designed to harvest semantic cues of different scene entities: it leverages a pre-trained object detector to localize and identify objects, after which the location-semantic tokens are computed as a fusion of visual bounding box coordinates and text embeddings of the object category. Another novel aspect of our work is the introduction of cross-stream agent attention and agent token fusions, which builds on the concept of agent attention [26] and re-defines it for cross-stream compatibility: agent tokens are reformulated into a 1D sequence format through the use of fully connected layers, bypassing the initial need for 2D pooling, removing the depthwise convolution branch, and forgiving 2D positional encodings. This mechanism enhances the model's ability to filter and emphasize relevant information from multiple data sources effectively. RefAtomNet achieves the highest performance for RAVAR, while a detailed ablation study demonstrates the effectiveness of each component." + }, + { + "type": "text", + "bbox": [ + 0.24, + 0.668, + 0.592, + 0.681 + ], + "angle": 0, + "content": "Our contributions can be summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.7, + 0.791, + 0.745 + ], + "angle": 0, + "content": "- We introduce the new Referring Atomic Video Action Recognition (RAVAR) task and the RefAVA benchmark with 36,630 atomic action instances manually annotated with suitable language expressions." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.748, + 0.787, + 0.792 + ], + "angle": 0, + "content": "- To establish a competitive benchmark, we examine 11 well-established models from different related fields, namely atomic action localization, video question answering, and video-text retrieval." + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.796, + 0.787, + 0.84 + ], + "angle": 0, + "content": "- We propose RefAtomNet, a new RAVAR approach that uses a novel agent-based semantic-location aware attentional fusion to integrate multi-modal tokens, suppressing the irrelevant visual cues. RefAtomNet delivers improve" + }, + { + "type": "list", + "bbox": [ + 0.225, + 0.7, + 0.791, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.41, + 0.129 + ], + "angle": 0, + "content": "K. Peng, J. Fu et al." + }, + { + "type": "text", + "bbox": [ + 0.243, + 0.146, + 0.788, + 0.179 + ], + "angle": 0, + "content": "ments of \\(3.85\\%\\) and \\(3.17\\%\\) of mAP, and \\(4.33\\%\\) and \\(4.03\\%\\) of AUROC, on val and test sets, respectively, compared with the leading baseline, BLIPv2 [42]." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.202, + 0.39, + 0.218 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.236, + 0.789, + 0.446 + ], + "angle": 0, + "content": "Referring Scene Understanding. Referring scene understanding aims to locate parts of interest within images or videos guided by natural language, exhibiting the utility in many computer vision tasks including autonomous driving [86] and video editing [6]. The development of this field cannot be advanced without the contribution of high-quality open-source datasets and benchmarks [2, 12, 30, 48, 72, 80, 86, 91]. For example, the CLEVR-Ref+ benchmark is proposed by Liu et al. [52] to achieve visual reasoning with referring expressions. Li et al. [46] proposed referring image segmentation by using a recurrent refinement network. Wu et al. [86] proposed referring multi-object tracking benchmark. However, there are no referring understanding works focusing on atomic video action analysis, whereupon we first conduct RAVAR benchmarking with our RefAVA dataset. We further clarify that Referring Video Object Segmentation (RVOS) task [21, 53, 58, 79] differs us from multiple perspectives, e.g., RVOS includes action name in input textual reference." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.448, + 0.79, + 0.627 + ], + "angle": 0, + "content": "Video Text Retrieval. Video Text Retrieval (VTR) aims at matching relevant video content with text. Amidst the rise of Vision-Language foundation models like CLIP [67] and BLIP [43], efforts [9,42,49,55-57,77,82,87,95] are being made to apply powerful pre-trained models for promoting their competencies in VTR. XCLIP [56] introduced multi-grained contrastive learning for end-to-end VTR, by counting all the video-sentence, video-word, sentence-frame, and frame-word contrasts. BLIPv2 [42] pre-trained a Querying Transformer (QFormer) to bootstrap vision-and-language representation learning and vision-to-language generative learning. The key difference between VTR and RAVAR lies in their inference tasks: VTR queries text or video to localize elements, while RAVAR integrates textual and visual data for fine-grained subject retrieval and atomic action prediction in videos." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.629, + 0.79, + 0.842 + ], + "angle": 0, + "content": "Video Question Answering. Video Question Answering (VQA) focuses on generating answers to questions posed in natural language for a given video. Depending on the emphasis of the question, factoid VQA [1,5,7,19,20,25,28,36,39, 45,54,88,89,96] straightforward queries visual facts. Lei et al. [38] proposed that a single-frame trained transformer-based model, with large-scale pre-training and a frame ensemble at the inference stage, can perform better than existing multi-frame trained models in factoid VQA tasks. On the other hand, inference VQA [18,41] delves into logical reasoning. Li et al. [44] released a video-centric instruction dataset and leveraged a neural interface to integrate video foundation models and Large Language Models (LLM), showcasing capability in temporal reasoning, causal inference, and event localization. In addition, some multimodal VQA frameworks [20,40,90] explore information-invoking scenarios that incorporate visual, audio, subtitle, and external knowledge. However, compared with the RAVAR task, most of the existing VQA approaches do not particularly focus" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.675, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "RAVAR" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.145, + 0.784, + 0.201 + ], + "angle": 0, + "content": "Table 1: An overview of the referring scene understanding datasets and our RefAVA dataset; Our task is Referring Atomic Video Action Recognition (RAVAR), whereas existing benchmarks focus on Referring Object Detection (ROD), Referring Video-based Object Segmentation (RVOS), and Referring Multi-Object Tracking (RMOT)." + }, + { + "type": "table", + "bbox": [ + 0.216, + 0.213, + 0.784, + 0.258 + ], + "angle": 0, + "content": "
DatasetRefCOCO [91]RefCOCO+ [91]RefCOCOg [91]Talk2Car [13]VID-Sentence [10]Refer-DAVUS17 [30]Refer-YV [94]Refer-KITTI [86]RefAVA
TaskRODRODRODRODRODRVOSRMOTRMOTRAVAR
NFrames26,71119,99226,7119,21759,2384,21993,8696,6501,615,140
NInstance26,71119,99226,71110,5197,6543,9787,451-36,630
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.287, + 0.784, + 0.331 + ], + "angle": 0, + "content": "on atomic action recognition and will only tend to give a coarse textual description, which limits the applications requiring precise prediction, e.g., human-robot assistance [27, 37]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.332, + 0.788, + 0.681 + ], + "angle": 0, + "content": "Atomic Video Action Recognition and Localization. Atomic video-based action recognition [11, 23] and localization [24] involves the identification and analysis of the most fundamental, indivisible actions or movements performed by humans for single and multiple-person scenarios. Compared with the general video action recognition task, atomic video-based action localization is much more fine-grained and is always formulated in a multi-label manner with bounding box predictions. Most of the existing convolutional neural networks (CNN) [4, 16, 17, 81] and the transformer [23, 47, 63, 69, 81, 83, 85] networks for human action recognition are commonly used in the atomic video-based action localization by changing the classification head into multi-label manner, adding additional bounding box prediction head, and integrating region of interest features from human detector. Ryali et al. [69] proposed a hierarchical vision transformer with high efficiency and precision within the realm of the existing methods using video as input. Wang et al. [81] proposed scaling video-masked autoencoders with dual masking. Current methods for predicting individual actions in multi-person scenarios often require manual video cropping for atomic action recognition or generate predictions for all detected individuals, necessitating further human selection and reducing practicality [65, 68, 84]. Most existing AAL methods are not specifically designed for the RAVAR task. We thereby introduce RefAtomNet, a novel method that utilizes location semantics atop predicted location and scene semantic information derived from the scenario together with the textual reference and visual cues, then further uses cross-stream agent attention and agent token fusions to suppress redundant information." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.704, + 0.571, + 0.72 + ], + "angle": 0, + "content": "3 RAVAR: Established Benchmark" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.738, + 0.562, + 0.751 + ], + "angle": 0, + "content": "3.1 Introduction of the RefAVA Dataset" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.784, + 0.84 + ], + "angle": 0, + "content": "Textual Annotations. To acquire precise textual annotations for the individuals of interest, 7 annotators manually provided the textual annotations according to the key frame bounding boxes presented in the AVA dataset [24]. Cross-checking among all the annotators for the annotated individuals is conducted to deliver high textual annotation quality." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.41, + 0.129 + ], + "angle": 0, + "content": "K. Peng, J. Fu et al." + }, + { + "type": "image", + "bbox": [ + 0.227, + 0.149, + 0.511, + 0.292 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.241, + 0.294, + 0.496, + 0.306 + ], + "angle": 0, + "content": "(a) Chord visualization of several keywords." + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.145, + 0.785, + 0.277 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.516, + 0.28, + 0.786, + 0.312 + ], + "angle": 0, + "content": "(b) Annotation statistics: #The number of annotated persons contained in the scenario (x-axis); #The number of instances (y-axis)." + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.323, + 0.788, + 0.365 + ], + "angle": 0, + "content": "Fig. 2: An chord visualization of several keywords from the textual references, shown on the left, and the instance amount in video clips containing different numbers of annotated persons in our RefAVA dataset, shown on the right." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.396, + 0.789, + 0.669 + ], + "angle": 0, + "content": "Dataset. We selected 17,946 video clips from 127 movies of the AVA dataset, preserving the most complex scenarios, and annotated each person on each center frame of every 90s video clip. In total, RefAVA has 36,630 labeled instances, and we split them into 22,658 train instances, 10,916 validation instances, and 3,056 test instances, where the samples from different sets are from different movie scenarios. The textual annotations cover information on approximate age, gender, appearance, relative position in the center frame, etc., and without action description. The chord visualization in Fig. 2a reveals the quantitative causal relations between the most frequent notional words in different categories from the textual references. Further, Fig. 2b shows the statistics of instance amount towards the annotated person number inside the scenario where each instance belongs. We deliver the comparison of the statistics between representative existing referring scene understanding datasets and our RefAVA dataset in Tab. 1. The atomic actions involve 80 categories covering Object Manipulation (OM), Person Interactions (PI), and Person Movement (PM). The videos cover diverse scenarios. The test set is sourced from 26 movies different from those in the train set (67 different movies) and val set (34 different movies) to achieve the evaluation of generalizability." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.692, + 0.611, + 0.706 + ], + "angle": 0, + "content": "3.2 The Baselines for the RAVAR Benchmark" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.719, + 0.788, + 0.842 + ], + "angle": 0, + "content": "We have adapted methodologies from similar fields as our baselines for RAVAR. AAL Baselines. We first reformulate the existing methods from the general atomic video action localization field by integrating the textual reference embeddings into the visual branch. Some action recognition approaches, e.g., I3D [4] and X3D [16], are adapted to AAL following [17]. BERT [14] is used for textual embedding extraction. The selected baselines from this domain can be grouped into CNN-based approaches [4,16] and transformer-based approaches [47,69,81]. All these approaches leverage the pre-trained weight on the Kinetics400 [4]." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.676, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "RAVAR" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.126 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.143, + 0.786, + 0.237 + ], + "angle": 0, + "content": "VQA Baselines. The second group of approaches comes from the video question answering domain, which contains the GPT-based model, i.e., AskAnything [44], and conventional transformer-based model [38]. We input the reference sentence in a questioning manner and add a classification head and a bounding box regression head after the acquisition of the logits. We finetuned the two models for the RAVAR task based on their pre-trained weight on ActivityNetQA [92]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.238, + 0.787, + 0.328 + ], + "angle": 0, + "content": "VTR Baselines. The last group of the foundation model baselines comes from the video-text retrieval task, where XCLIP [56], CLIP4CLIP [55], BLIPv2 [42], and MeVTR [95] are selected. These foundation models are pre-trained on a combination of numerous datasets, incorporating Conceptual Captions [74], SBU Captions [61], and COCO Captions [8], etc. We reformulate these approaches in the same way as the approaches in VQA." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.328, + 0.787, + 0.373 + ], + "angle": 0, + "content": "SF Baselines. Some Single Frame (SF) baselines are adopted, where SAM [33], DETR [3], and REFCLIP [29] are utilized together with the CLIP [67] feature extraction backbone." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.374, + 0.787, + 0.404 + ], + "angle": 0, + "content": "VOS Baseline. We adopt one Video Object Segmentation (VOS) baseline using the encoder of the approach from Su et al. [79] and our prediction head." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.429, + 0.563, + 0.446 + ], + "angle": 0, + "content": "4 RefAtomNet: Proposed Method" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.462, + 0.341, + 0.476 + ], + "angle": 0, + "content": "4.1 Overview" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.488, + 0.787, + 0.655 + ], + "angle": 0, + "content": "We propose a new model, RefAtomNet, with an overview provided in Fig. 3. It leverages three token streams: visual, textual reference, and location-semantic streams. Visual tokens are extracted from the video using ViT [15], while textual reference tokens are obtained from the text using BERT [14]. Both streams are enhanced by QFormer [42]. Additionally, in the location-semantic stream, location-semantic tokens are extracted from the center frame by fusing predicted object coordinates and semantic embeddings of the object categories estimated by a frozen object detector and BERT [14]. To suppress irrelevant information for each stream, we propose the agent-based location-semantic aware attentional fusion to achieve better amplification of relevant information during the cross-stream exchange. The detailed structure is discussed in the following subsections." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.678, + 0.368, + 0.692 + ], + "angle": 0, + "content": "4.2 RefAtomNet" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.704, + 0.787, + 0.81 + ], + "angle": 0, + "content": "Background of QFormer. We rely on BLIPv2 to extract the visual features, where QFormer is the most essential component. Multiple learnable queries are initialized in QFormer as trainable parameters and interact with input data via Transformer attention mechanisms. Each query selectively attends to input parts, capturing task-specific features and updating based on attention outputs. During training, these queries are optimized for specialized information extraction. Queries then directly contribute to generating outputs." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.797, + 0.84 + ], + "angle": 0, + "content": "Extraction of Visual and Textual Reference Tokens. Similar to BLIPv2 [42], we use a pre-trained multimodal model for token extraction from both the video" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.41, + 0.129 + ], + "angle": 0, + "content": "K. Peng, J. Fu et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.266 + ], + "angle": 0, + "content": "and textual reference streams. This model leverages the QFormer [42] architecture to effectively extract and integrate multimodal embeddings from the visual and textual reference data. Specifically, the visual stream employs a ViT [15] backbone encoder to process visual inputs, while the textual branch utilizes BERT [14] to encode textual reference cues. This combined approach facilitates a holistic understanding of both visual content and textual descriptions. The model extracts visual tokens \\((\\mathbf{t}^{VT})\\) and textual reference tokens \\((\\mathbf{t}^{RT})\\) through Eq. 1:" + }, + { + "type": "equation", + "bbox": [ + 0.355, + 0.267, + 0.786, + 0.285 + ], + "angle": 0, + "content": "\\[\n\\mathbf {t} ^ {V T}, \\mathbf {t} ^ {R T} = \\mathcal {V} _ {V L} \\left(\\mathcal {V} _ {V T} \\left(\\mathbf {x} ^ {V T}\\right), \\mathcal {V} _ {R T} \\left(\\mathbf {x} ^ {R T}\\right)\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.291, + 0.788, + 0.367 + ], + "angle": 0, + "content": "where \\(\\mathcal{V}_{VL}\\) represents the visual-textual integration model (i.e., QFormer [42]). \\(\\mathcal{V}_{VT}\\), and \\(\\mathcal{V}_{RT}\\) denote the backbones for extracting visual tokens (i.e., ViT [15]) and textual reference tokens (i.e., BERT [14]), respectively. \\(\\mathbf{x}^{VT}\\) and \\(\\mathbf{x}^{RT}\\) represent the input video and textual reference caption, which are then fed into liner projection layers, i.e., \\(\\mathbf{P}_{VT}\\) and \\(\\mathbf{P}_{RT}\\), respectively." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.367, + 0.788, + 0.457 + ], + "angle": 0, + "content": "Extraction of Location-Semantic Aware Tokens. To integrate more location and semantic information into the token representations, we leverage a well-established object detector DETR [3] to deliver \\( N_{o} \\) detection results based on the input of the center frame of a video clip, noted as the keyframe. Note that, the detections provided by the object detector contain either the human or other objects with a high confidence score according to Eq. 2." + }, + { + "type": "equation", + "bbox": [ + 0.416, + 0.472, + 0.786, + 0.487 + ], + "angle": 0, + "content": "\\[\n\\mathbf {r} _ {\\text {b o x e s}}, \\mathbf {r} _ {\\text {c a t s}} = \\mathcal {V} _ {\\text {d e t s}} (\\mathbf {x} _ {k}), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.498, + 0.788, + 0.619 + ], + "angle": 0, + "content": "where \\(\\mathbf{x}_k\\), \\(\\mathcal{V}_{\\text{dets}}\\) indicates the keyframe and the detection network. \\(\\mathbf{r}_{\\text{boxes}} \\in \\mathbb{R}^{N_o \\times 4}\\) are detected 2D corner coordinates (top left and bottom right corners) of the bounding boxes. \\(\\mathbf{r}_{\\text{cats}}\\) is the predicted category for each bounding box represented in textual format. These category labels of the detected objects are passed to a text encoder (BERT [14]) to extract semantic embeddings. We concatenate these two types of tokens together along the channel dimension and use a single linear projection layer to obtain the aggregated location-semantic aware tokens \\((\\mathbf{t}^{LS})\\) as Eq. 3." + }, + { + "type": "equation", + "bbox": [ + 0.361, + 0.63, + 0.786, + 0.648 + ], + "angle": 0, + "content": "\\[\n\\mathbf {t} ^ {L S} = \\mathbf {P} _ {L S} \\left(\\text {C o n c a t} \\left[ \\mathcal {V} _ {R T} \\left(\\mathbf {r} _ {\\text {c a t s}}\\right), \\mathbf {r} _ {\\text {b o x e s}} \\right]\\right), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.659, + 0.788, + 0.689 + ], + "angle": 0, + "content": "where \\(\\mathbf{P}_{LS}\\) indicates a fully connected layer, \\(\\mathcal{V}_{RT}\\) indicates the language feature extraction backbone, i.e., BERT [14], and Concat indicates the concatenation." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.689, + 0.704, + 0.703 + ], + "angle": 0, + "content": "Agent-Based Location Semantic Aware Attentional Fusion." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.705, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Agent Tokens and Sequential Processing. We use agent tokens, inspired by agent attention [26]. Derived through fully connected layers from input tokens, similar to Query, Key, and Value extraction in transformers, agent tokens aggregate essential information via agent-key and agent-query pairs. To enable agent attention for 1D sequential tokens from our streams, we redefine the 2D agent spatial tokens proposed in [26] into a 1D sequential format for multi-stream fusion. Specifically, we replace 2D pooling with linear projection and eliminate the depthwise convolutional branch and biased position encoding to better suit agent acquisition in our model. These 1D sequential agent tokens serve dual functions:" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.675, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "RAVAR" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "image", + "bbox": [ + 0.219, + 0.146, + 0.784, + 0.342 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.326, + 0.352, + 0.677, + 0.366 + ], + "angle": 0, + "content": "Fig. 3: An overview of the RefAtomNet architecture." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.393, + 0.784, + 0.437 + ], + "angle": 0, + "content": "aggregating crucial information from agent-key and agent-query pairs and efficiently redistributing this information to the original value tokens, therefore, enhancing attention focus and reducing irrelevant cues." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.438, + 0.784, + 0.469 + ], + "angle": 0, + "content": "Agent-Based Attention Mechanism. The agent-based attention mechanism for visual, textual reference, and location-semantic streams is computed as:" + }, + { + "type": "equation", + "bbox": [ + 0.299, + 0.475, + 0.786, + 0.496 + ], + "angle": 0, + "content": "\\[\n\\mathbf {Q} ^ {\\phi}, \\mathbf {K} ^ {\\phi}, \\mathbf {V} ^ {\\phi}, \\mathbf {A} ^ {\\phi} = \\mathbf {W} _ {Q} ^ {\\phi} (\\mathbf {t} ^ {\\phi}), \\mathbf {W} _ {K} ^ {\\phi} (\\mathbf {t} ^ {\\phi}), \\mathbf {W} _ {V} ^ {\\phi} (\\mathbf {t} ^ {\\phi}), \\mathbf {W} _ {A} ^ {\\phi} (\\mathbf {t} ^ {\\phi}), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.501, + 0.788, + 0.713 + ], + "angle": 0, + "content": "where for better readability, we use \\(\\phi\\) to represent \\([RT, VT, LS]\\), where RT indicates the reference tokens, VT indicates the visual tokens, and LS indicates the location-semantic tokens. \\(\\mathbf{Q}^{\\phi}\\), \\(\\mathbf{K}^{\\phi}\\), \\(\\mathbf{V}^{\\phi}\\), and \\(\\mathbf{A}^{\\phi}\\) are the query, key, value, and agent tokens. \\(t^{\\phi}\\) depicts the input. \\(\\mathbf{W}_Q^{\\phi}\\), \\(\\mathbf{W}_K^{\\phi}\\), \\(\\mathbf{W}_V^{\\phi}\\), and \\(\\mathbf{W}_A^{\\phi}\\) are constructed by linear projection layers. We then project the agent tokens using fully connected layers \\(\\mathbf{P}_A^\\phi\\) through \\(\\mathbf{A}_*^{\\phi} = \\mathbf{P}_A^\\phi (\\mathbf{A}^\\phi)\\) as aforementioned. To mitigate redundancy cues of the textual reference and location-semantic streams, we leverage agent attention for both of these two streams. The agent query attention mask \\(\\mathbf{M}_{QA}^{\\pi}\\) and the agent key attention mask \\(\\mathbf{M}_{KA}^{\\pi}\\) are obtained by matrix multiplication (MatMul) and SoftMax operations along the channel dimension (indicated by \\(\\sigma_c\\)) as shown in Eq. 5, where \\(\\pi \\in [RT, LS]\\) and \\(\\alpha\\) indicates a fixed scale factor. The agent attention masks and tokens are computed and refined to ensure that only pertinent information influences the attention mechanism, as shown in Eq. 5 and Eq. 6:" + }, + { + "type": "equation", + "bbox": [ + 0.258, + 0.72, + 0.785, + 0.739 + ], + "angle": 0, + "content": "\\[\n\\mathbf {M} _ {Q A} ^ {\\pi}, \\mathbf {M} _ {K A} ^ {\\pi} = \\sigma_ {c} (M a t M u l [ \\alpha * \\mathbf {A} _ {*} ^ {\\pi}, \\mathbf {Q} ^ {\\pi} ]), \\sigma_ {c} (M a t M u l [ \\alpha * \\mathbf {A} _ {*} ^ {\\pi}, \\mathbf {K} ^ {\\pi} ]), \\tag {5}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.328, + 0.74, + 0.785, + 0.759 + ], + "angle": 0, + "content": "\\[\n\\mathbf {t} _ {*} ^ {\\pi} = F F N \\left(M a t M u l \\left[ \\mathbf {M} _ {K A} ^ {\\pi}, M a t M u l \\left[ \\mathbf {M} _ {Q A} ^ {\\pi}, \\mathbf {V} ^ {\\pi} \\right] \\right]\\right), \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.784, + 0.794 + ], + "angle": 0, + "content": "where \\( FFN \\) indicates the Feed Forward Network. These steps ensure precise model attention, enhancing the contextual relevance of the resulting tokens." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.796, + 0.784, + 0.84 + ], + "angle": 0, + "content": "Cross-Stream Agent Attention and Agent Token Fusions. Finally, cross-stream agent attention and agent token fusion are proposed, which are explicitly tailored for the visual stream by applying the agent query attention maps and the" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.41, + 0.129 + ], + "angle": 0, + "content": "K. Peng, J. Fu et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.208 + ], + "angle": 0, + "content": "agent tokens from the other two streams. This process involves recalculating the attention maps, thereby ensuring that the final token representation is highly relevant and contextually enriched. We compute the agent query attention map for the visual token stream, where \\(\\gamma = VT\\), as shown in Eq. 7," + }, + { + "type": "equation", + "bbox": [ + 0.379, + 0.22, + 0.786, + 0.238 + ], + "angle": 0, + "content": "\\[\n\\mathbf {M} _ {Q A} ^ {\\gamma} = \\sigma_ {c} (M a t M u l [ \\alpha * \\mathbf {A} _ {*} ^ {\\gamma}, \\mathbf {Q} ^ {\\gamma} ]). \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.247, + 0.789, + 0.324 + ], + "angle": 0, + "content": "Then, we calculate the cross-stream irrelevance-suppressed attention for the agent query attention of the visual stream through Eq. 8 to achieve cross-stream agent attention fusion to suppress irrelevant information in the agent query attentions of the visual stream, the operations used in the second and third terms are abbreviated as C-ATT and T-ATT in Fig. 3," + }, + { + "type": "equation", + "bbox": [ + 0.264, + 0.346, + 0.786, + 0.388 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathbf {M}} _ {Q A} ^ {\\gamma} = A V G \\left[ \\mathbf {M} _ {Q A} ^ {\\gamma}, \\sigma_ {c} (\\sum_ {\\pi} \\mathbf {M} _ {Q A} ^ {\\pi}) * \\mathbf {M} _ {Q A} ^ {\\gamma}, \\sigma_ {t} (\\sum_ {\\pi} \\mathbf {M} _ {Q A} ^ {\\pi}) * \\mathbf {M} _ {Q A} ^ {\\gamma} \\right], \\quad (8)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.395, + 0.789, + 0.457 + ], + "angle": 0, + "content": "where \\(\\sigma_{t}\\) denotes the SoftMax operation along the token dimension. \\(AVG\\) indicates the mean operation. We follow the same procedure to calculate the cross-stream irrelevance-suppressed agent tokens before the calculation of the agent key attention, as in Eq. 9 to achieve cross-stream agent token fusion." + }, + { + "type": "equation", + "bbox": [ + 0.3, + 0.467, + 0.786, + 0.509 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathbf {A}} _ {*} ^ {\\gamma} = A V G \\left[ \\mathbf {A} _ {*} ^ {\\gamma} + \\sigma_ {c} \\left(\\sum_ {\\pi} \\mathbf {A} _ {*} ^ {\\pi}\\right) * \\mathbf {A} _ {*} ^ {\\gamma} + \\sigma_ {t} \\left(\\sum_ {\\pi} \\mathbf {A} _ {*} ^ {\\pi}\\right) * \\mathbf {A} _ {*} ^ {\\gamma} \\right], \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.513, + 0.668, + 0.528 + ], + "angle": 0, + "content": "then, we calculate the agent key attention as shown in Eq. 10," + }, + { + "type": "equation", + "bbox": [ + 0.374, + 0.539, + 0.786, + 0.565 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathbf {M}} _ {K A} ^ {\\gamma} = \\sigma_ {c} (M a t M u l [ \\alpha * \\hat {\\mathbf {A}} _ {*} ^ {\\gamma}, \\mathbf {K} ^ {\\gamma} ]). \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.573, + 0.788, + 0.603 + ], + "angle": 0, + "content": "The final aggregated visual tokens can be obtained through the following equation as demonstrated in Eq. 11," + }, + { + "type": "equation", + "bbox": [ + 0.314, + 0.614, + 0.786, + 0.641 + ], + "angle": 0, + "content": "\\[\n\\mathbf {t} _ {*} ^ {\\gamma} = F F N (M a t M u l \\left[ \\hat {\\mathbf {M}} _ {K A} ^ {\\gamma}, M a t M u l \\left[ \\hat {\\mathbf {M}} _ {Q A} ^ {\\gamma}, \\mathbf {V} ^ {\\gamma} \\right] \\right]). \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.649, + 0.788, + 0.68 + ], + "angle": 0, + "content": "Finally, we aggregate all the tokens from three branches by using the mean operation, where \\( N_{s} \\) indicates the number of the stream as shown in Eq. 12," + }, + { + "type": "equation", + "bbox": [ + 0.432, + 0.691, + 0.786, + 0.722 + ], + "angle": 0, + "content": "\\[\n\\mathbf {t} _ {a g g} = \\sum_ {\\phi} \\left[ \\mathbf {t} _ {*} ^ {\\phi} \\right] / N _ {s}. \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.733, + 0.788, + 0.764 + ], + "angle": 0, + "content": "We construct MLP-based classification and regression heads atop aggregated tokens for center frame bounding box prediction and atomic action recognition." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.786, + 0.387, + 0.799 + ], + "angle": 0, + "content": "4.3 Loss Functions" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.81, + 0.789, + 0.841 + ], + "angle": 0, + "content": "We use Binary Cross Entropy (BCE) loss and Mean Squared Error (MSE) loss for the multi-label supervision and the bounding box regression supervision in" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.675, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "RAVAR" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.147, + 0.772, + 0.163 + ], + "angle": 0, + "content": "the same way as all the baselines. The equation of the BCE loss is as Eq. 13," + }, + { + "type": "equation", + "bbox": [ + 0.3, + 0.175, + 0.786, + 0.214 + ], + "angle": 0, + "content": "\\[\nL _ {B C E} (\\mathbf {y}, \\hat {\\mathbf {y}}) = - \\frac {1}{N _ {c}} \\sum_ {i = 1} ^ {N _ {c}} \\left[ \\mathbf {y} _ {i} \\log \\left(\\hat {\\mathbf {y}} _ {i}\\right) + \\left(1 - \\mathbf {y} _ {i}\\right) \\log \\left(1 - \\hat {\\mathbf {y}} _ {i}\\right) \\right], \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.226, + 0.787, + 0.256 + ], + "angle": 0, + "content": "where \\(\\mathbf{y}\\), \\(\\hat{\\mathbf{y}}\\) and \\(N_{c}\\) indicate the one-hot ground truth, the prediction, and the category number. The bounding boxes regression loss is expressed as Eq. 14," + }, + { + "type": "equation", + "bbox": [ + 0.397, + 0.268, + 0.786, + 0.309 + ], + "angle": 0, + "content": "\\[\nL _ {M S E} (\\mathbf {b}, \\hat {\\mathbf {b}}) = \\sum_ {j = 1} ^ {4} \\left(\\mathbf {b} _ {j} - \\hat {\\mathbf {b}} _ {j}\\right) ^ {2}, \\tag {14}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.32, + 0.785, + 0.35 + ], + "angle": 0, + "content": "where \\(\\mathbf{b}\\) indicates the coordinates of the left top and right bottom corners, and \\(\\hat{\\mathbf{b}}\\) represents the predicted boxes. \\(j\\) indicates the coordinate index." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.374, + 0.377, + 0.391 + ], + "angle": 0, + "content": "5 Experiments" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.405, + 0.457, + 0.42 + ], + "angle": 0, + "content": "5.1 Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.43, + 0.788, + 0.567 + ], + "angle": 0, + "content": "We conduct experiments on four NVIDIA A100 GPUs. We use BertAdam [32] optimizer with learning rate \\( lr = 1e^{-4} \\), batch size 128, learning rate decay 0.9, and warmup ratio 0.1. The agent number and head number in our RefAtomNet are 4 and 1. Our model has \\( 214M \\) trainable parameters. The weight for \\( L_{MSE} \\) and \\( \\alpha \\) are 5 and 0.125. We trained our model for 40 epochs on our dataset. The text encoder is frozen during training. We use multi-label mean Average Precision (mAP), Area Under the Receiver Operating Characteristic curve (AUROC), and the mean Intersection over Union (mIOU) as metrics, of which mIOU has less priority since we prioritize the atomic video action recognition task." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.589, + 0.442, + 0.604 + ], + "angle": 0, + "content": "5.2 Experimental Results" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.614, + 0.789, + 0.842 + ], + "angle": 0, + "content": "Main Results. Results of the main experiments on the proposed RefAVA benchmark are summarized in Tab. 2. We compare our RefAtomNet model with a multitude of published approaches adopted to suit our task (see Section 3.2) stemming from the fields of (1) Atomic Action Localization (AAL); (2) Video Question Answering (VQA); (3) Video-Text Retrieval (VTR); (4) Single Frame (SF) baselines; and (5) Video Object Segmentation (VOS). Baseline methods from the AAL group consistently underperform in spatially localizing the action according to the textual reference (measured as mIOU). Even though the ability of the human spatial localization is not the major expected output, it can still generally illustrate if the model can grasp the correct referring person or not. Among the AAL methods, X3D [16] achieves the highest AUROC score of \\(59.09\\%\\) and \\(64.51\\%\\) on the val and test sets, respectively. The VQA and VTR baselines work better than the AAL baselines since they can reasonably capture the referring person while delivering acceptable atomic human action recognition performances, which benefits from the text-aware pretraining of these two" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.409, + 0.128 + ], + "angle": 0, + "content": "K. Peng, J. Fu et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.304, + 0.145, + 0.698, + 0.159 + ], + "angle": 0, + "content": "Table 2: Experimental results on our RAVAR benchmark." + }, + { + "type": "table", + "bbox": [ + 0.229, + 0.172, + 0.784, + 0.375 + ], + "angle": 0, + "content": "
MethodmIOUmAPAUROCmIOUmAPAUROC
ValTest
AALI3D [4]0.0044.0457.770.0044.6462.71
X3D [16]0.2644.4559.090.2746.3464.51
MViTv2-B [47]0.7642.3256.430.6642.6059.30
VideoMAE2-B [81]0.1642.0255.120.2941.8759.10
Hiera-B [69]0.4942.7456.720.6241.1458.70
VQASingularity [38]26.3442.4359.5229.7842.1856.12
AskAnything13B [44]20.0951.4266.1222.3552.2569.35
VTRMeVTR [95]30.7838.4251.0129.7936.2752.45
CLIP4CLIP [55]34.7539.4852.5732.3337.1755.05
XCIIIP [56]35.5442.4656.3031.7940.8258.71
BLIPv2 [42]32.9952.1366.5632.7553.1969.92
SFCLIP [67]+SAM [33]32.9549.7464.5529.8051.3469.11
CLIP [67]+DETR [3]33.9647.5761.8333.8450.9267.29
CLIP [67]+REFCLIP [29]34.4347.7962.5032.2849.9067.44
VOSSu et al. [79]23.7152.1766.6726.0253.2070.19
RefAtomNet (Ours)38.2255.9869.7336.4257.5273.95
" + }, + { + "type": "table_caption", + "bbox": [ + 0.218, + 0.393, + 0.782, + 0.407 + ], + "angle": 0, + "content": "Table 3: Experiments for module ablation and comparison with multi-modal fusion." + }, + { + "type": "table_caption", + "bbox": [ + 0.242, + 0.424, + 0.478, + 0.436 + ], + "angle": 0, + "content": "(a) Module ablation of the RefAtomNet." + }, + { + "type": "table", + "bbox": [ + 0.218, + 0.448, + 0.484, + 0.532 + ], + "angle": 0, + "content": "
MethodmIOUmAPAUROCmIOUmAPAUROC
ValTest
w/o ALSAF27.3050.7065.3129.0951.2668.20
w/o CAAF36.2155.4369.6635.3855.9072.60
w/o CATF35.0153.8367.7134.5556.9673.40
w/o LSAS31.9055.2169.4731.2555.7372.38
Ours38.2255.9869.7336.4257.5273.95
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.424, + 0.779, + 0.437 + ], + "angle": 0, + "content": "(b) Comparison with other fusion approaches." + }, + { + "type": "table", + "bbox": [ + 0.503, + 0.448, + 0.787, + 0.532 + ], + "angle": 0, + "content": "
FusionmIOU mAP AUROCmIOU mAP AUROC
ValTest
Addition27.3050.7065.3129.0951.2668.20
Concatenation18.6452.2366.4520.6553.4470.70
Multiplication23.9051.5565.6625.0553.3370.48
AttentionBottleneck [59]33.4750.9765.0733.0254.0271.08
McOmut [98]23.8851.5865.6525.0253.2170.42
Ours38.2255.9869.7336.4257.5273.95
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.566, + 0.788, + 0.672 + ], + "angle": 0, + "content": "tasks. AskAnything [44] from the VQA group achieves \\(20.09\\%\\) and \\(22.35\\%\\) of mIOU, \\(51.42\\%\\) and \\(52.25\\%\\) of mAP, and \\(66.12\\%\\) and \\(69.35\\%\\) of AUROC on the val and test sets. BLIPv2 [43] from the VTR group delivers \\(32.99\\%\\) and \\(32.75\\%\\) of mIOU, \\(52.13\\%\\) and \\(53.19\\%\\) of mAP, and \\(66.56\\%\\) and \\(69.92\\%\\) of AUROC, on the val and test sets, respectively. Our RefAtomNet achieves state-of-the-art performances, outperforming the best baseline BLIPv2 [42] by \\(5.23\\%\\), \\(3.85\\%\\), \\(3.17\\%\\) and \\(3.67\\%\\), \\(4.33\\%\\), \\(4.03\\%\\) of mIOU, mAP, and AUROC, on val and test sets." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Ablations of the Individual Modules. In Tab. 3a, we show the ablation for the components of RefAtomNet by removing each of the proposed designs, where LSAS indicates the location-semantic stream, CAAF indicates the cross-stream agent attention fusion, CATF indicates the cross-stream agent token fusion, and \\( w / o \\) ALSAF indicates by simply using addition to fuse these three streams and without the agent-based location-semantic aware attentional fusion mechanism. Compared with the ablation \\( w / o \\) ALSAF, RefAtomNet achieves promising improvements of \\( 10.92\\% \\), \\( 5.28\\% \\), \\( 4.42\\% \\) and \\( 7.33\\% \\), \\( 6.26\\% \\), \\( 5.75\\% \\) in terms of mIOU, mAP, and AUROC for the val and test sets, respectively. It indicates that using simple aggregation of the three streams makes the model distract from the important cues for the referring person, illustrated by the large decay of mIOU metric." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.676, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "RAVAR" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "table_caption", + "bbox": [ + 0.235, + 0.145, + 0.767, + 0.159 + ], + "angle": 0, + "content": "Table 4: Generalizability to different referring styles and encoder architectures." + }, + { + "type": "table_caption", + "bbox": [ + 0.235, + 0.175, + 0.486, + 0.188 + ], + "angle": 0, + "content": "(a) Generalizability to test time rephrasing" + }, + { + "type": "table", + "bbox": [ + 0.217, + 0.2, + 0.497, + 0.281 + ], + "angle": 0, + "content": "
FusionmIOUmAPAUROCmIOUmAPAUROC
ValTest
Singularity [38]18.4541.2758.4720.5441.3955.32
XCLIP [56]31.9541.3554.3529.8440.7458.45
AskAnything [44]19.7451.1165.8321.6051.9669.04
BLIPv2 [42]31.0051.4565.8831.3452.3568.87
Ours34.6555.7569.5233.1457.2373.76
" + }, + { + "type": "table_caption", + "bbox": [ + 0.503, + 0.177, + 0.788, + 0.198 + ], + "angle": 0, + "content": "(b) Generalizability to different visual-textual encoder architectures." + }, + { + "type": "table", + "bbox": [ + 0.503, + 0.211, + 0.785, + 0.28 + ], + "angle": 0, + "content": "
FusionmIOU mAP AUROCmIOU mAP AUROC
ValTest
XCLIP [56]35.5442.4656.3031.7940.8258.71
RefAtomNet (XCLIP)38.5947.4061.2036.6148.5966.47
BLIPv2 [42]32.9952.1366.5632.7553.1969.92
RefAtomNet (BLIPv2)38.2255.9869.7336.4257.5273.95
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.311, + 0.784, + 0.401 + ], + "angle": 0, + "content": "Compared with the ablation \\( w / o \\) LSAS, we find that the location-semantic tokens benefit more for the localization of the correct person in the center frame. The advantages of the LSAS on atomic video action recognition metrics are highlighted in the test set which has higher scenario diversity compared with the val set. Using CAAF and CATF, each of them brings promising benefits by suppressing the irrelevant information in the visual tokens." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.402, + 0.785, + 0.476 + ], + "angle": 0, + "content": "Comparison with Other Fusion Mechanisms. We compare our proposed agent-based location-semantic aware attentional fusion mechanism with late fusion addition, multiplication, concatenation, AttentionBottleNeck [59], and McOmet [98] in Tab. 3b. Our approach outperforms all by suppressing those irrelevant visual tokens on the agent tokens and attention masks." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.478, + 0.786, + 0.643 + ], + "angle": 0, + "content": "Generalizability to Test-time Reference Rephrasing. The reference sentences given by different users may differ in daily life scenarios. To test the generalizability of the model towards different referring styles, we invoke API (gpt-3.5-turbo) of ChatGPT [60] to rephrase the test set description two times and then deliver the averaged performance for RAVAR on the original val and test sets, and the two rephrased val and test sets, where we select 4 most out-performing baselines from our benchmark to construct this ablation study, i.e., Singularity [38], AskAnything [44], XCLIP [56], and BLIPv2 [42], as shown in Tab. 4a. RefAtomNet delivers the best performances. We find out that the test time reference rephrasing will cause performance decay for localization and less decay for the atomic video-based action recognition delivered by our model." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.644, + 0.787, + 0.78 + ], + "angle": 0, + "content": "Generalizability to Different Visual Textual Backbones. Since we use the best-performing baseline BLIPv2 [42] as the visual and textual encoder in our model, it would be interesting to see if the proposed architecture can generalize to different encoder backbones. We thereby demonstrate another ablation study in Tab. 4b, where we equip our RefAtomNet with the XCLIP [56] backbone. Compared with the XCLIP baseline, the RefAtomNet (XCLIP) achieves \\(3.05\\%\\), \\(4.94\\%\\), \\(4.90\\%\\) and \\(4.82\\%\\), \\(7.77\\%\\), \\(7.76\\%\\) performance improvements in terms of mIOU, mAP, and AUROC on the val and test sets, respectively, showing the great generalizability of our RefAtomNet of the visual and textual encoder." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Analysis of the Qualitative Results. The qualitative results are delivered in Fig. 4. These examples demonstrate the effectiveness of RefAtomNet compared with the best-performing baseline BLIPv2 [42]. Samples 1 and 8 validate the recognition quality for references without any location indications." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.41, + 0.129 + ], + "angle": 0, + "content": "K. Peng, J. Fu et al." + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.146, + 0.446, + 0.156 + ], + "angle": 0, + "content": "1 Text reference: The man wearing light green clothes." + }, + { + "type": "image", + "bbox": [ + 0.225, + 0.156, + 0.499, + 0.228 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.503, + 0.146, + 0.749, + 0.156 + ], + "angle": 0, + "content": "2 Text reference: The man in the center who has gray beard." + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.156, + 0.78, + 0.228 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.228, + 0.468, + 0.238 + ], + "angle": 0, + "content": "3 Text reference: The man on the left who is in blue clothes." + }, + { + "type": "image", + "bbox": [ + 0.225, + 0.238, + 0.499, + 0.311 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.503, + 0.229, + 0.682, + 0.239 + ], + "angle": 0, + "content": "4 Text reference: The first man on the right." + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.239, + 0.78, + 0.311 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.311, + 0.446, + 0.321 + ], + "angle": 0, + "content": "5 Text reference: The man wearing glasses on the left." + }, + { + "type": "image", + "bbox": [ + 0.225, + 0.321, + 0.499, + 0.393 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.504, + 0.311, + 0.708, + 0.321 + ], + "angle": 0, + "content": "6 Text reference: The man on the right back to us." + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.321, + 0.78, + 0.393 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.393, + 0.48, + 0.412 + ], + "angle": 0, + "content": "Text reference: The woman wearing black clothes and pearls necklace on the right." + }, + { + "type": "image", + "bbox": [ + 0.224, + 0.412, + 0.499, + 0.493 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.503, + 0.393, + 0.696, + 0.404 + ], + "angle": 0, + "content": "8 Text reference: The man wearing black suits." + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.412, + 0.78, + 0.493 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.507, + 0.787, + 0.536 + ], + "angle": 0, + "content": "Fig. 4: An overview of qualitative results. Missed predictions are marked with a red cross, while true positive and false positive predictions are shown in green and red." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.563, + 0.788, + 0.655 + ], + "angle": 0, + "content": "Samples 3, 4, 5, and 6 showcase the results for different humans in a shared scene. Samples 2 and 7 show the RAVAR performance for references containing finer details, e.g., the necklace type and the beard color. Our RefAtomNet outperforms the BLIPv2 baseline on selected samples obviously, attributable to its superior capability in eliminating extraneous visual cues and its advanced location-semantic reasoning prowess." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.675, + 0.37, + 0.691 + ], + "angle": 0, + "content": "6 Conclusions" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.704, + 0.789, + 0.841 + ], + "angle": 0, + "content": "In this work, we introduce a novel task called Referring Atomic Video Action Recognition (RAVAR). We establish the RefAVA dataset to address the challenge of identifying atomic actions for individuals of interest in videos based on textual descriptions. Existing methods exhibit poor performance on this new task, prompting the development of RefAtomNet, a vision-language architecture that effectively integrates cross-stream tokens for precise referring atomic video action recognition. Through reference-relevant token enhancement and comprehensive token integration, RefAtomNet achieves impressive results, highlighting its effectiveness in tackling the complexities of the RAVAR task." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.676, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "RAVAR" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.403, + 0.163 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.177, + 0.788, + 0.374 + ], + "angle": 0, + "content": "The project served to prepare the SFB 1574 Circular Factory for the Perpetual Product (project ID: 471687386), approved by the German Research Foundation (DFG, German Research Foundation) with a start date of April 1, 2024. This work was also partially supported in part by the SmartAge project sponsored by the Carl Zeiss Stiftung (P2019-01-003; 2021-2026). This work was performed on the HoreKa supercomputer funded by the Ministry of Science, Research and the Arts Baden-Württemberg and by the Federal Ministry of Education and Research. The authors also acknowledge support by the state of Baden-Württemberg through bwHPC and the German Research Foundation (DFG) through grant INST 35/1597-1 FUGG. This project is also supported by the National Key RD Program under Grant 2022YFB4701400. Lastly, the authors thank for the support of Dr. Sepideh Pashami, the Swedish Innovation Agency VINNOVA, the Digital Futures." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.396, + 0.323, + 0.411 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.426, + 0.785, + 0.454 + ], + "angle": 0, + "content": "1. Bagad, P., Tapaswi, M., Snoek, C.G.M.: Test of time: Instilling video-language models with a sense of time. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.455, + 0.785, + 0.482 + ], + "angle": 0, + "content": "2. Bu, Y., Li, L., Xie, J., Liu, Q., Cai, Y., Huang, Q., Li, Q.: Scene-text oriented referring expression comprehension. TMM (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.483, + 0.785, + 0.51 + ], + "angle": 0, + "content": "3. Carion, N., Massa, F., Synnaeve, G., Usunier, N., Kirillov, A., Zagoruyko, S.: End-to-end object detection with transformers. In: ECCV (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.51, + 0.785, + 0.537 + ], + "angle": 0, + "content": "4. Carreira, J., Zisserman, A.: Quo vadis, action recognition? A new model and the kinetics dataset. In: CVPR (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.537, + 0.785, + 0.564 + ], + "angle": 0, + "content": "5. Castro, S., Deng, N., Huang, P., Burzo, M., Mihalcea, R.: In-the-wild video question answering. In: COLING (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.565, + 0.785, + 0.592 + ], + "angle": 0, + "content": "6. Chai, W., Guo, X., Wang, G., Lu, Y.: StableVideo: Text-driven consistency-aware diffusion video editing. In: ICCV (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.592, + 0.785, + 0.633 + ], + "angle": 0, + "content": "7. Chen, J., Zhu, D., Haydarov, K., Li, X., Elhoseiny, M.: Video ChatCaptioner: Towards enriched spatiotemporal descriptions. arXiv preprint arXiv:2304.04227 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.633, + 0.785, + 0.675 + ], + "angle": 0, + "content": "8. Chen, X., Fang, H., Lin, T.Y., Vedantam, R., Gupta, S., Dolkar, P., Zitnick, C.L.: Microsoft COCO captions: Data collection and evaluation server. arXiv preprint arXiv:1504.00325 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.675, + 0.785, + 0.716 + ], + "angle": 0, + "content": "9. Chen, Y., Wang, J., Lin, L., Qi, Z., Ma, J., Shan, Y.: Tagging before alignment: Integrating multi-modal tags for video-text retrieval. arXiv preprint arXiv:2301.12644 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.716, + 0.785, + 0.744 + ], + "angle": 0, + "content": "0. Chen, Z., Ma, L., Luo, W., Wong, K.Y.K.: Weakly-supervised spatio-temporally grounding natural sentence in video. arXiv preprint arXiv:1906.02549 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.744, + 0.785, + 0.771 + ], + "angle": 0, + "content": "1. Chung, J., Wu, C.h., Yang, H.r., Tai, Y.W., Tang, C.K.: HAA500: Human-centric atomic action dataset with curated videos. In: ICCV (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.771, + 0.785, + 0.812 + ], + "angle": 0, + "content": "2. Dang, R., Feng, J., Zhang, H., Ge, C., Song, L., Gong, L., Liu, C., Chen, Q., Zhu, F., Zhao, R., Song, Y.: InstructDET: Diversifying referring object detection with generalized instructions. arXiv preprint arXiv:2310.05136 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.812, + 0.785, + 0.84 + ], + "angle": 0, + "content": "3. Deruyttere, T., Vandenhende, S., Grujicic, D., Van Gool, L., Moens, M.F.: Talk2Car: Taking control of your self-driving car. In: EMNLP (2019)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.426, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.409, + 0.128 + ], + "angle": 0, + "content": "K. Peng, J. Fu et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.785, + 0.175 + ], + "angle": 0, + "content": "14. Devlin, J., Chang, M.W., Lee, K., Toutanova, K.: BERT: Pre-training of deep bidirectional transformers for language understanding. In: ACL (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.176, + 0.785, + 0.23 + ], + "angle": 0, + "content": "15. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., Uszkoreit, J., Houlsby, N.: An image is worth 16x16 words: Transformers for image recognition at scale. In: ICLR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.23, + 0.785, + 0.256 + ], + "angle": 0, + "content": "16. Feichtenhofer, C.: X3D: Expanding architectures for efficient video recognition. In: CVPR (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.257, + 0.785, + 0.284 + ], + "angle": 0, + "content": "17. Feichtenhofer, C., Fan, H., Malik, J., He, K.: SlowFast networks for video recognition. In: ICCV (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.285, + 0.785, + 0.325 + ], + "angle": 0, + "content": "18. Gandhi, M., Gul, M.O., Prakash, E., Grunde-McLaughlin, M., Krishna, R., Agrawala, M.: Measuring compositional consistency for video question answering. In: CVPR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.325, + 0.785, + 0.365 + ], + "angle": 0, + "content": "19. Gao, D., Zhou, L., Ji, L., Zhu, L., Yang, Y., Shou, M.Z.: MIST: Multi-modal iterative spatial-temporal transformer for long-form video question answering. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.365, + 0.785, + 0.393 + ], + "angle": 0, + "content": "20. Garcia, N., Otani, M., Chu, C., Nakashima, Y.: KnowIT VQA: Answering knowledge-based questions about videos. In: AAAI (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.393, + 0.785, + 0.419 + ], + "angle": 0, + "content": "21. Gavrilyuk, K., Ghodrati, A., Li, Z., Snoek, C.G.: Actor and action video segmentation from a sentence. In: CVPR (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.42, + 0.785, + 0.475 + ], + "angle": 0, + "content": "22. Goyal, R., Kahou, S.E., Michalski, V., Materzynska, J., Westphal, S., Kim, H., Haenel, V., Fründ, I., Yianilos, P., Mueller-Freitag, M., Hoppe, F., Thurau, C., Bax, I., Memisevic, R.: The \"something something\" video database for learning and evaluating visual common sense. In: ICCV (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.476, + 0.785, + 0.515 + ], + "angle": 0, + "content": "23. Gritsenko, A., Xiong, X., Djolonga, J., Dehghani, M., Sun, C., Lucic, M., Schmid, C., Arnab, A.: End-to-end spatio-temporal action localisation with video transformers. arXiv preprint arXiv:2304.12160 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.516, + 0.785, + 0.556 + ], + "angle": 0, + "content": "24. Gu, C., Sun, C., Ross, D.A., Vondrick, C., Pantofaru, C., Li, Y., Vijayanarasimhan, S., Toderici, G., Ricco, S., Sukthankar, R., Schmid, C., Malik, J.: AVA: A video dataset of spatio-temporally localized atomic visual actions. In: CVPR (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.557, + 0.785, + 0.583 + ], + "angle": 0, + "content": "25. Guo, W., Zhang, Y., Yang, J., Yuan, X.: Re-attention for visual question answering. TIP (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.584, + 0.785, + 0.61 + ], + "angle": 0, + "content": "26. Han, D., Ye, T., Han, Y., Xia, Z., Song, S., Huang, G.: Agent attention: On the integration of softmax and linear attention. arXiv preprint arXiv:2312.08874 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.611, + 0.785, + 0.637 + ], + "angle": 0, + "content": "27. Ji, Y., Zhan, Y., Yang, Y., Xu, X., Shen, F., Shen, H.T.: A context knowledge map guided coarse-to-fine action recognition. TIP (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.637, + 0.785, + 0.677 + ], + "angle": 0, + "content": "28. Jiang, J., Chen, Z., Lin, H., Zhao, X., Gao, Y.: Divide and conquer: Question-guided spatio-temporal contextual attention for video question answering. In: AAAI (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.678, + 0.785, + 0.704 + ], + "angle": 0, + "content": "29. Jin, L., Luo, G., Zhou, Y., Sun, X., Jiang, G., Shu, A., Ji, R.: Refclip: A universal teacher for weakly supervised referring expression comprehension. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.705, + 0.785, + 0.732 + ], + "angle": 0, + "content": "30. Khoreva, A., Rohrbach, A., Schiele, B.: Video object segmentation with language referring expressions. In: ACCV (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.733, + 0.785, + 0.759 + ], + "angle": 0, + "content": "31. Kim, M., Spinola, F., Benz, P., Kim, T.h.: A*: Atrous spatial temporal action recognition for real time applications. In: WACV (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.759, + 0.785, + 0.786 + ], + "angle": 0, + "content": "32. Kingma, D.P., Ba, J.: Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980 (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.786, + 0.785, + 0.813 + ], + "angle": 0, + "content": "33. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., et al.: Segment anything. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.814, + 0.785, + 0.84 + ], + "angle": 0, + "content": "34. Kuehne, H., Jhuang, H., Garrote, E., Poggio, T., Serre, T.: HMDB: A large video database for human motion recognition. In: ICCV (2011)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.676, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "RAVAR" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.785, + 0.175 + ], + "angle": 0, + "content": "35. Laput, G., Harrison, C.: Sensing fine-grained hand activity with smartwatches. In: CHI (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.175, + 0.785, + 0.203 + ], + "angle": 0, + "content": "36. Le, T.M., Le, V., Venkatesh, S., Tran, T.: Hierarchical conditional relation networks for video question answering. In: CVPR (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.203, + 0.785, + 0.231 + ], + "angle": 0, + "content": "37. Lea, C., Vidal, R., Hager, G.D.: Learning convolutional action primitives for fine-grained action recognition. In: ICRA (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.231, + 0.785, + 0.259 + ], + "angle": 0, + "content": "38. Lei, J., Berg, T.L., Bansal, M.: Revealing single frame bias for video-and-language learning. arXiv preprint arXiv:2206.03428 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.259, + 0.785, + 0.286 + ], + "angle": 0, + "content": "39. Lei, J., Li, L., Zhou, L., Gan, Z., Berg, T.L., Bansal, M., Liu, J.: Less is more: ClipBERT for video-and-language learning via sparse sampling. In: CVPR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.286, + 0.785, + 0.314 + ], + "angle": 0, + "content": "40. Li, G., Wei, Y., Tian, Y., Xu, C., Wen, J.R., Hu, D.: Learning to answer questions in dynamic audio-visual scenarios. In: CVPR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.314, + 0.785, + 0.342 + ], + "angle": 0, + "content": "41. Li, J., Niu, L., Zhang, L.: From representation to reasoning: Towards both evidence and commonsense reasoning for video question-answering. In: CVPR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.342, + 0.785, + 0.37 + ], + "angle": 0, + "content": "42. Li, J., Li, D., Savarese, S., Hoi, S.: BLIP-2: Bootstrapping language-image pretraining with frozen image encoders and large language models. In: ICML (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.37, + 0.785, + 0.397 + ], + "angle": 0, + "content": "43. Li, J., Li, D., Xiong, C., Hoi, S.: BLIP: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In: ICML (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.397, + 0.785, + 0.438 + ], + "angle": 0, + "content": "44. Li, K., He, Y., Wang, Y., Li, Y., Wang, W., Luo, P., Wang, Y., Wang, L., Qiao, Y.: VideoChat: Chat-centric video understanding. arXiv preprint arXiv:2305.06355 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.438, + 0.785, + 0.466 + ], + "angle": 0, + "content": "45. Li, L., Gan, Z., Lin, K., Lin, C.C., Liu, Z., Liu, C., Wang, L.: LAVENDER: Unifying video-language understanding as masked language modeling. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.466, + 0.785, + 0.494 + ], + "angle": 0, + "content": "46. Li, R., Li, K., Kuo, Y.C., Shu, M., Qi, X., Shen, X., Jia, J.: Referring image segmentation via recurrent refinement networks. In: CVPR (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.494, + 0.785, + 0.535 + ], + "angle": 0, + "content": "47. Li, Y., Wu, C.Y., Fan, H., Mangalam, K., Xiong, B., Malik, J., Feichtenhofer, C.: MViTv2: Improved multiscale vision transformers for classification and detection. In: CVPR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.535, + 0.785, + 0.577 + ], + "angle": 0, + "content": "48. Lin, J., Chen, J., Peng, K., He, X., Li, Z., Stiefelhagen, R., Yang, K.: EchoTrack: Auditory referring multi-object tracking for autonomous driving. arXiv preprint arXiv:2402.18302 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.577, + 0.785, + 0.618 + ], + "angle": 0, + "content": "49. Lin, X., Tiwari, S., Huang, S., Li, M., Shou, M.Z., Ji, H., Chang, S.F.: Towards fast adaptation of pretrained contrastive models for multi-channel video-language retrieval. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.619, + 0.785, + 0.646 + ], + "angle": 0, + "content": "50. Liu, J., Wang, L., Yang, M.H.: Referring expression generation and comprehension via attributes. In: ICCV (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.646, + 0.785, + 0.688 + ], + "angle": 0, + "content": "51. Liu, R., Zhang, J., Peng, K., Zheng, J., Cao, K., Chen, Y., Yang, K., Stiefelhagen, R.: Open scene understanding: Grounded situation recognition meets segment anything for helping people with visual impairments. In: ICCVW (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.688, + 0.785, + 0.716 + ], + "angle": 0, + "content": "52. Liu, R., Liu, C., Bai, Y., Yuille, A.L.: CLEVR-Ref+: Diagnosing visual reasoning with referring expressions. In: CVPR (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.716, + 0.785, + 0.743 + ], + "angle": 0, + "content": "53. Liu, S., Hui, T., Huang, S., Wei, Y., Li, B., Li, G.: Cross-modal progressive comprehension for referring segmentation. TPAMI (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.743, + 0.785, + 0.771 + ], + "angle": 0, + "content": "54. Liu, Y., Li, G., Lin, L.: Cross-modal causal relational reasoning for event-level visual question answering. TPAMI (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.771, + 0.785, + 0.812 + ], + "angle": 0, + "content": "55. Luo, H., Ji, L., Zhong, M., Chen, Y., Lei, W., Duan, N., Li, T.: CLIP4Clip: An empirical study of CLIP for end to end video clip retrieval and captioning. Neurocomputing (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.812, + 0.785, + 0.84 + ], + "angle": 0, + "content": "56. Ma, Y., Xu, G., Sun, X., Yan, M., Zhang, J., Ji, R.: X-CLIP: End-to-end multigrained contrastive learning for video-text retrieval. In: MM (2022)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.409, + 0.128 + ], + "angle": 0, + "content": "K. Peng, J. Fu et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.175 + ], + "angle": 0, + "content": "57. Madasu, A., Aflalo, E., Ben Melech Stan, G., Tseng, S.Y., Bertasius, G., Lal, V.: Improving video retrieval using multilingual knowledge transfer. In: ECIR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.177, + 0.785, + 0.203 + ], + "angle": 0, + "content": "58. McIntosh, B., Duarte, K., Rawat, Y.S., Shah, M.: Visual-textual capsule routing for text-based video segmentation. In: CVPR (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.204, + 0.785, + 0.231 + ], + "angle": 0, + "content": "59. Nagrani, A., Yang, S., Arnab, A., Jansen, A., Schmid, C., Sun, C.: Attention bottlenecks for multimodal fusion. In: NeuIPS (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.232, + 0.784, + 0.258 + ], + "angle": 0, + "content": "60. OpenAI: ChatGPT: Optimizing language models for dialogue. https://openai.com/ (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.259, + 0.785, + 0.286 + ], + "angle": 0, + "content": "61. Ordonez, V., Kulkarni, G., Berg, T.: Im2Text: Describing images using 1 million captioned photographs. In: NeurIPS (2011)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.287, + 0.785, + 0.327 + ], + "angle": 0, + "content": "62. Ou, W., Zhang, J., Peng, K., Yang, K., Jaworek, G., Müller, K., Stiefelhagen, R.: Indoor navigation assistance for visually impaired people via dynamic SLAM and panoptic segmentation with an RGB-D sensor. In: ICCHP (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.328, + 0.785, + 0.369 + ], + "angle": 0, + "content": "63. Peng, K., Roitberg, A., Yang, K., Zhang, J., Stiefelhagen, R.: TransDARC: Transformer-based driver activity recognition with latent space feature calibration. In: IROS (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.37, + 0.785, + 0.397 + ], + "angle": 0, + "content": "64. Pramanick, P., Sarkar, C., Paul, S., dev Roychoudhury, R., Bhowmick, B.: DoRO: Disambiguation of referred object for embodied agents. RA-L (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.398, + 0.785, + 0.424 + ], + "angle": 0, + "content": "65. Pramono, R.R.A., Chen, Y.T., Fang, W.H.: Spatial-temporal action localization with hierarchical self-attention. TMM (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.425, + 0.785, + 0.466 + ], + "angle": 0, + "content": "66. Qiu, H., Li, H., Wu, Q., Meng, F., Shi, H., Zhao, T., Ngan, K.N.: Language-aware fine-grained object representation for referring expression comprehension. In: MM (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.467, + 0.785, + 0.508 + ], + "angle": 0, + "content": "67. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., Krueger, G., Sutskever, I.: Learning transferable visual models from natural language supervision. In: ICML (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.509, + 0.785, + 0.536 + ], + "angle": 0, + "content": "68. Rajasegaran, J., Pavlakos, G., Kanazawa, A., Feichtenhofer, C., Malik, J.: On the benefits of 3D pose and tracking for human action recognition. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.537, + 0.785, + 0.577 + ], + "angle": 0, + "content": "69. Ryali, C., Hu, Y., Bolya, D., Wei, C., Fan, H., Huang, P., Aggarwal, V., Chowdhury, A., Poursaeed, O., Hoffman, J., Malik, J., Li, Y., Feichtenhofer, C.: Hiera: A hierarchical vision transformer without the bells-and-whistles. In: ICML (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.578, + 0.785, + 0.604 + ], + "angle": 0, + "content": "70. Saha, J., Chowdhury, C., Chowdury, I.R., Roy, P.: Fine grained activity recognition using smart handheld. In: ICDCN (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.605, + 0.785, + 0.633 + ], + "angle": 0, + "content": "71. Seibold, C.M., Reif, S., Kleesiek, J., Stiefelhagen, R.: Reference-guided pseudolabel generation for medical semantic segmentation. In: AAAI (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.634, + 0.785, + 0.66 + ], + "angle": 0, + "content": "72. Seo, S., Lee, J.Y., Han, B.: URVOS: Unified referring video object segmentation network with a large-scale benchmark. In: ECCV (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.661, + 0.785, + 0.688 + ], + "angle": 0, + "content": "73. Shao, D., Zhao, Y., Dai, B., Lin, D.: Finegym: A hierarchical video dataset for fine-grained action understanding. In: CVPR (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.689, + 0.785, + 0.729 + ], + "angle": 0, + "content": "74. Sharma, P., Ding, N., Goodman, S., Soricut, R.: Conceptual captions: A cleaned, hypernymed, image alt-text dataset for automatic image captioning. In: ACL (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.729, + 0.785, + 0.757 + ], + "angle": 0, + "content": "75. Shi, H., Pan, W., Zhao, Z., Zhang, M., Wu, F.: Unsupervised domain adaptation for referring semantic segmentation. In: MM (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.758, + 0.785, + 0.784 + ], + "angle": 0, + "content": "76. Shi, H., Li, H., Meng, F., Wu, Q.: Key-word-aware network for referring expression image segmentation. In: ECCV (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.785, + 0.785, + 0.813 + ], + "angle": 0, + "content": "77. Shi, Y., Xu, H., Yuan, C., Li, B., Hu, W., Zha, Z.J.: Learning video-text aligned representations for video captioning. TOMM (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.814, + 0.785, + 0.84 + ], + "angle": 0, + "content": "78. Soomro, K., Zamir, A.R., Shah, M.: UCF101: A dataset of 101 human actions classes from videos in the wild. arXiv preprint arXiv:1212.0402 (2012)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.676, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "RAVAR" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "19" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.785, + 0.175 + ], + "angle": 0, + "content": "79. Su, Y., Wang, W., Liu, J., Ma, S., Yang, X.: Sequence as a whole: A unified framework for video action localization with long-range text query. TIP (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.177, + 0.785, + 0.203 + ], + "angle": 0, + "content": "80. Vasudevan, A.B., Dai, D., Van Gool, L.: Object referring in videos with language and human gaze. In: CVPR (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.204, + 0.785, + 0.245 + ], + "angle": 0, + "content": "81. Wang, L., Huang, B., Zhao, Z., Tong, Z., He, Y., Wang, Y., Wang, Y., Qiao, Y.: VideoMAE V2: Scaling video masked autoencoders with dual masking. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.246, + 0.785, + 0.273 + ], + "angle": 0, + "content": "82. Wang, M., Xing, J., Mei, J., Liu, Y., Jiang, Y.: ActionCLIP: Adapting language-image pretrained models for video action recognition. TNNLS (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.273, + 0.785, + 0.314 + ], + "angle": 0, + "content": "83. Wang, R., Chen, D., Wu, Z., Chen, Y., Dai, X., Liu, M., Yuan, L., Jiang, Y.G.: Masked video distillation: Rethinking masked feature modeling for self-supervised video representation learning. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.315, + 0.785, + 0.342 + ], + "angle": 0, + "content": "84. Wang, S., Yan, R., Huang, P., Dai, G., Song, Y., Shu, X.: Com-STAL: Compositional spatio-temporal action localization. TCSVT (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.343, + 0.785, + 0.397 + ], + "angle": 0, + "content": "85. Wang, Y., Li, K., Li, Y., He, Y., Huang, B., Zhao, Z., Zhang, H., Xu, J., Liu, Y., Wang, Z., Xing, S., Chen, G., Pan, J., Yu, J., Wang, Y., Wang, L., Qiao, Y.: InternVideo: General video foundation models via generative and discriminative learning. arXiv preprint arXiv:2212.03191 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.398, + 0.785, + 0.424 + ], + "angle": 0, + "content": "86. Wu, D., Han, W., Wang, T., Dong, X., Zhang, X., Shen, J.: Referring multi-object tracking. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.425, + 0.785, + 0.452 + ], + "angle": 0, + "content": "87. Wu, W., Luo, H., Fang, B., Wang, J., Ouyang, W.: Cap4Video: What can auxiliary captions do for text-video retrieval? In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.453, + 0.785, + 0.479 + ], + "angle": 0, + "content": "88. Xiao, J., Shang, X., Yao, A., Chua, T.S.: NExT-QA: Next phase of question-answering to explaining temporal actions. In: CVPR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.481, + 0.785, + 0.508 + ], + "angle": 0, + "content": "89. Yang, A., Miech, A., Sivic, J., Laptev, I., Schmid, C.: Just ask: Learning to answer questions from millions of narrated videos. In: ICCV (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.509, + 0.785, + 0.536 + ], + "angle": 0, + "content": "90. Yang, P., Wang, X., Duan, X., Chen, H., Hou, R., Jin, C., Zhu, W.: AVQA: A dataset for audio-visual question answering on videos. In: MM (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.537, + 0.785, + 0.563 + ], + "angle": 0, + "content": "91. Yu, L., Poirson, P., Yang, S., Berg, A.C., Berg, T.L.: Modeling context in referring expressions. In: ECCV (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.565, + 0.785, + 0.605 + ], + "angle": 0, + "content": "92. Yu, Z., Xu, D., Yu, J., Yu, T., Zhao, Z., Zhuang, Y., Tao, D.: ActivityNet-QA: A dataset for understanding complex web videos via question answering. In: AAAI (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.606, + 0.785, + 0.646 + ], + "angle": 0, + "content": "93. Yuan, Z., Yan, X., Liao, Y., Zhang, R., Wang, S., Li, Z., Cui, S.: InstanceRefer: Cooperative holistic understanding for visual grounding on point clouds through instance multi-level contextual referring. In: ICCV (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.647, + 0.785, + 0.674 + ], + "angle": 0, + "content": "94. Zeng, F., Dong, B., Zhang, Y., Wang, T., Zhang, X., Wei, Y.: MOTR: End-to-end multiple-object tracking with transformer. In: ECCV (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.675, + 0.785, + 0.702 + ], + "angle": 0, + "content": "95. Zhang, G., Ren, J., Gu, J., Tresp, V.: Multi-event video-text retrieval. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.703, + 0.785, + 0.73 + ], + "angle": 0, + "content": "96. Zhang, H., Li, X., Bing, L.: Video-LLaMA: An instruction-tuned audio-visual language model for video understanding. In: EMNLP (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.731, + 0.785, + 0.771 + ], + "angle": 0, + "content": "97. Zheng, J., Zhang, J., Yang, K., Peng, K., Stiefelhagen, R.: MateRobot: Material recognition in wearable robotics for people with visual impairments. In: ICRA (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.772, + 0.785, + 0.799 + ], + "angle": 0, + "content": "98. Zong, D., Sun, S.: McOmet: Multimodal fusion transformer for physical audiovisual commonsense reasoning. In: AAAI (2023)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.785, + 0.799 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/Referring Atomic Video Action Recognition/3c97fc3b-c407-40be-ae83-9d7a4bd31254_origin.pdf b/2024/Referring Atomic Video Action Recognition/3c97fc3b-c407-40be-ae83-9d7a4bd31254_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..f85d3ea3bf4f74166795235aea6b8d9e9d91b8ac --- /dev/null +++ b/2024/Referring Atomic Video Action Recognition/3c97fc3b-c407-40be-ae83-9d7a4bd31254_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:25d26093c283f0a558882f2610debe447dfac42e3536004be6496e5a49c816a0 +size 2690599 diff --git a/2024/Referring Atomic Video Action Recognition/full.md b/2024/Referring Atomic Video Action Recognition/full.md new file mode 100644 index 0000000000000000000000000000000000000000..d7b8be4435a1174a7bfca8237bd28da592966489 --- /dev/null +++ b/2024/Referring Atomic Video Action Recognition/full.md @@ -0,0 +1,392 @@ +# Referring Atomic Video Action Recognition + +Kunyu Peng $^{1,*}$ , Jia Fu $^{3,4,*}$ , Kailun Yang $^{2,\dagger}$ , Di Wen $^{1}$ , Yufan Chen $^{1}$ , Ruiping Liu $^{1}$ , Junwei Zheng $^{1}$ , Jiaming Zhang $^{1}$ , M. Saqib Sarfraz $^{1,6}$ , Rainer Stiefelhagen $^{1}$ , and Alina Roitberg $^{5}$ + +$^{1}$ Karlsruhe Institute of Technology, $^{2}$ Hunan University, $^{3}$ RISE Research Institutes of Sweden, $^{4}$ KTH Royal Institute of Technology, $^{5}$ University of Stuttgart, $^{6}$ Mercedes-Benz Tech Innovation + +Abstract. We introduce a new task called Referring Atomic Video Action Recognition (RAVAR), aimed at identifying atomic actions of a particular person based on a textual description and the video data of this person. This task differs from traditional action recognition and localization, where predictions are delivered for all present individuals. In contrast, we focus on recognizing the correct atomic action of a specific individual, guided by text. To explore this task, we present the RefAVA dataset, containing 36,630 instances with manually annotated textual descriptions of the individuals. To establish a strong initial benchmark, we implement and validate baselines from various domains, e.g., atomic action localization, video question answering, and text-video retrieval. Since these existing methods underperform on RAVAR, we introduce RefAtomNet - a novel cross-stream attention-driven method specialized for the unique challenges of RAVAR: the need to interpret a textual referring expression for the targeted individual, utilize this reference to guide the spatial localization and harvest the prediction of the atomic actions for the referring person. The key ingredients are: (1) a multi-stream architecture that connects video, text, and a new location-semantic stream, and (2) cross-stream agent attention fusion and agent token fusion which amplify the most relevant information across these streams and consistently surpasses standard attention-based fusion on RAVAR. Extensive experiments demonstrate the effectiveness of RefAtomNet and its building blocks for recognizing the action of the described individual. The dataset and code will be made publicly available at RAVAR. + +# 1 Introduction + +Referring scene understanding [50, 52, 66, 93] aims to solve the underlying computer vision task for the particular scene element specified via a natural language referring expression. Context-driven queries are vital in information retrieval, assistive systems, and multimedia analysis, but incorporating them into computer vision models is challenging due to a complex entanglement of linguistic descriptions, localization, and visual recognition itself. Several benchmarks for referring scene understanding have been successfully established for multi-object + +![](images/f616d00f7bfd11a1e4da9bb0839d5a77f557c291b420ae8731c50413dc1fc6fb.jpg) +Fig. 1: A comparison of AAL Task (left) and RAVAR task (right). + +tracking [86], semantic segmentation [30, 46, 72, 75, 76], medical imaging [71], and object detection [12, 64]. While these efforts focus on object-centric referring expressions, many applications require a human-centric understanding of human actions, e.g., in rehabilitation assistance [35, 70] and human-robot interaction [27, 37]. Our work addresses a new problem - reference-driven recognition of atomic actions performed by the human described via a textual reference, which has been overlooked in the past. + +The field of human action recognition has made exciting advances, from new approaches stemming from the rise of visual transformers [23,47,63,69,81,83,85] to new large-scale datasets [4,22,34,73,78]. Datasets utilized for the prediction of atomic actions often feature recordings with multiple individuals [24]. Nevertheless, the vast majority of previously published works [23,31,68,69,81] either rely on manually window crop of a particular person of interest or automatically generated region of interests to predict the atomic actions for all the individuals in one video, which results in large pre- or post-processing effort to focus on one specific person. This workflow is also very inconvenient in certain applications, such as assistive systems for users with visual impairments [51,62,97], who must understand the state of each present in the scene for effective interactions. Leveraging succinct textual descriptions, which may include broad positional indicators (e.g., left, center, and right), appearance attributes (e.g., hair color and clothing), or gender, to steer the model towards delivering atomic action recognition outcomes for the targeted individual, presents a promising solution to the outlined challenges. Employing these textual cues as a reference, the model can facilitate an end-to-end retrieval of the specified instance across the entire video, subsequently providing precise atomic action recognition results. + +To close this gap, we formalize the new task of Referring Atomic Video Action Recognition (RAVAR). The differences between RAVAR and atomic action localization are shown in Fig. 1. The conventional atomic action recognition in multi-person scenarios is often formalized in an action localization manner, and requires the region of interest to localize the human available a priori, after which atomic action recognition is carried out for each individual. Post-processing is required for individuals of interest. In contrast, RAVAR assumes a textual reference and a video as inputs, delivering the referring individual's atomic actions along with the location. This comparison shows the efficiency of RAVAR, leveraging textual information for less labor-intensive subject identification to assist atomic action analysis for the individual of interest. We further + +introduce the new RefAVA dataset for RAVAR. RefAVA is established using the 17,946 video clips of the public AVA dataset [24] initially designed for atomic action localization, which we extend with 36,630 textual descriptions of the corresponding individuals used as the referring expression. The dataset covers diverse settings, e.g., in/out door and day/night time scenarios, and numerous multi-person scenes, constituting an ideal testbed for our task. + +To establish an initial benchmark, we assess 15 well-established approaches from multiple related research domains. Our experiments reveal that none of the benchmarked methods delivers sufficient performance. For example, previous methods within the domains of VQA and VTR typically yield predictions of a coarse analysis regarding actions, whereas strategies in AAL face challenges in spatially localizing the correct referring individual. We attribute this phenomenon to the massive irrelevant information provided by the visual data, distracting the model from the referring location. How to suppress the referring irrelevant information becomes a critical challenge to pursue better RAVAR performances. + +To address this, we introduce the new RefAtomNet approach specialized in simultaneously solving the tasks of understanding textual referring expressions and using them as guides for spatial localization and classification of atomic actions. We begin by computing textual, visual, and newly proposed location-semantic tokens using a large foundation model and well-established object detector. The latter stream is designed to harvest semantic cues of different scene entities: it leverages a pre-trained object detector to localize and identify objects, after which the location-semantic tokens are computed as a fusion of visual bounding box coordinates and text embeddings of the object category. Another novel aspect of our work is the introduction of cross-stream agent attention and agent token fusions, which builds on the concept of agent attention [26] and re-defines it for cross-stream compatibility: agent tokens are reformulated into a 1D sequence format through the use of fully connected layers, bypassing the initial need for 2D pooling, removing the depthwise convolution branch, and forgiving 2D positional encodings. This mechanism enhances the model's ability to filter and emphasize relevant information from multiple data sources effectively. RefAtomNet achieves the highest performance for RAVAR, while a detailed ablation study demonstrates the effectiveness of each component. + +Our contributions can be summarized as follows: + +- We introduce the new Referring Atomic Video Action Recognition (RAVAR) task and the RefAVA benchmark with 36,630 atomic action instances manually annotated with suitable language expressions. +- To establish a competitive benchmark, we examine 11 well-established models from different related fields, namely atomic action localization, video question answering, and video-text retrieval. +- We propose RefAtomNet, a new RAVAR approach that uses a novel agent-based semantic-location aware attentional fusion to integrate multi-modal tokens, suppressing the irrelevant visual cues. RefAtomNet delivers improve + +ments of $3.85\%$ and $3.17\%$ of mAP, and $4.33\%$ and $4.03\%$ of AUROC, on val and test sets, respectively, compared with the leading baseline, BLIPv2 [42]. + +# 2 Related Work + +Referring Scene Understanding. Referring scene understanding aims to locate parts of interest within images or videos guided by natural language, exhibiting the utility in many computer vision tasks including autonomous driving [86] and video editing [6]. The development of this field cannot be advanced without the contribution of high-quality open-source datasets and benchmarks [2, 12, 30, 48, 72, 80, 86, 91]. For example, the CLEVR-Ref+ benchmark is proposed by Liu et al. [52] to achieve visual reasoning with referring expressions. Li et al. [46] proposed referring image segmentation by using a recurrent refinement network. Wu et al. [86] proposed referring multi-object tracking benchmark. However, there are no referring understanding works focusing on atomic video action analysis, whereupon we first conduct RAVAR benchmarking with our RefAVA dataset. We further clarify that Referring Video Object Segmentation (RVOS) task [21, 53, 58, 79] differs us from multiple perspectives, e.g., RVOS includes action name in input textual reference. + +Video Text Retrieval. Video Text Retrieval (VTR) aims at matching relevant video content with text. Amidst the rise of Vision-Language foundation models like CLIP [67] and BLIP [43], efforts [9,42,49,55-57,77,82,87,95] are being made to apply powerful pre-trained models for promoting their competencies in VTR. XCLIP [56] introduced multi-grained contrastive learning for end-to-end VTR, by counting all the video-sentence, video-word, sentence-frame, and frame-word contrasts. BLIPv2 [42] pre-trained a Querying Transformer (QFormer) to bootstrap vision-and-language representation learning and vision-to-language generative learning. The key difference between VTR and RAVAR lies in their inference tasks: VTR queries text or video to localize elements, while RAVAR integrates textual and visual data for fine-grained subject retrieval and atomic action prediction in videos. + +Video Question Answering. Video Question Answering (VQA) focuses on generating answers to questions posed in natural language for a given video. Depending on the emphasis of the question, factoid VQA [1,5,7,19,20,25,28,36,39, 45,54,88,89,96] straightforward queries visual facts. Lei et al. [38] proposed that a single-frame trained transformer-based model, with large-scale pre-training and a frame ensemble at the inference stage, can perform better than existing multi-frame trained models in factoid VQA tasks. On the other hand, inference VQA [18,41] delves into logical reasoning. Li et al. [44] released a video-centric instruction dataset and leveraged a neural interface to integrate video foundation models and Large Language Models (LLM), showcasing capability in temporal reasoning, causal inference, and event localization. In addition, some multimodal VQA frameworks [20,40,90] explore information-invoking scenarios that incorporate visual, audio, subtitle, and external knowledge. However, compared with the RAVAR task, most of the existing VQA approaches do not particularly focus + +Table 1: An overview of the referring scene understanding datasets and our RefAVA dataset; Our task is Referring Atomic Video Action Recognition (RAVAR), whereas existing benchmarks focus on Referring Object Detection (ROD), Referring Video-based Object Segmentation (RVOS), and Referring Multi-Object Tracking (RMOT). + +
DatasetRefCOCO [91]RefCOCO+ [91]RefCOCOg [91]Talk2Car [13]VID-Sentence [10]Refer-DAVUS17 [30]Refer-YV [94]Refer-KITTI [86]RefAVA
TaskRODRODRODRODRODRVOSRMOTRMOTRAVAR
NFrames26,71119,99226,7119,21759,2384,21993,8696,6501,615,140
NInstance26,71119,99226,71110,5197,6543,9787,451-36,630
+ +on atomic action recognition and will only tend to give a coarse textual description, which limits the applications requiring precise prediction, e.g., human-robot assistance [27, 37]. + +Atomic Video Action Recognition and Localization. Atomic video-based action recognition [11, 23] and localization [24] involves the identification and analysis of the most fundamental, indivisible actions or movements performed by humans for single and multiple-person scenarios. Compared with the general video action recognition task, atomic video-based action localization is much more fine-grained and is always formulated in a multi-label manner with bounding box predictions. Most of the existing convolutional neural networks (CNN) [4, 16, 17, 81] and the transformer [23, 47, 63, 69, 81, 83, 85] networks for human action recognition are commonly used in the atomic video-based action localization by changing the classification head into multi-label manner, adding additional bounding box prediction head, and integrating region of interest features from human detector. Ryali et al. [69] proposed a hierarchical vision transformer with high efficiency and precision within the realm of the existing methods using video as input. Wang et al. [81] proposed scaling video-masked autoencoders with dual masking. Current methods for predicting individual actions in multi-person scenarios often require manual video cropping for atomic action recognition or generate predictions for all detected individuals, necessitating further human selection and reducing practicality [65, 68, 84]. Most existing AAL methods are not specifically designed for the RAVAR task. We thereby introduce RefAtomNet, a novel method that utilizes location semantics atop predicted location and scene semantic information derived from the scenario together with the textual reference and visual cues, then further uses cross-stream agent attention and agent token fusions to suppress redundant information. + +# 3 RAVAR: Established Benchmark + +# 3.1 Introduction of the RefAVA Dataset + +Textual Annotations. To acquire precise textual annotations for the individuals of interest, 7 annotators manually provided the textual annotations according to the key frame bounding boxes presented in the AVA dataset [24]. Cross-checking among all the annotators for the annotated individuals is conducted to deliver high textual annotation quality. + +![](images/3e5fe960a155d8fceeda66643690b3c23e8f632cccd672b5859fa1e8c986b73c.jpg) +(a) Chord visualization of several keywords. + +![](images/230b39ea41268350f85d973bcbaa8ac917c594a95c1db74f928394e6e59eff56.jpg) +(b) Annotation statistics: #The number of annotated persons contained in the scenario (x-axis); #The number of instances (y-axis). +Fig. 2: An chord visualization of several keywords from the textual references, shown on the left, and the instance amount in video clips containing different numbers of annotated persons in our RefAVA dataset, shown on the right. + +Dataset. We selected 17,946 video clips from 127 movies of the AVA dataset, preserving the most complex scenarios, and annotated each person on each center frame of every 90s video clip. In total, RefAVA has 36,630 labeled instances, and we split them into 22,658 train instances, 10,916 validation instances, and 3,056 test instances, where the samples from different sets are from different movie scenarios. The textual annotations cover information on approximate age, gender, appearance, relative position in the center frame, etc., and without action description. The chord visualization in Fig. 2a reveals the quantitative causal relations between the most frequent notional words in different categories from the textual references. Further, Fig. 2b shows the statistics of instance amount towards the annotated person number inside the scenario where each instance belongs. We deliver the comparison of the statistics between representative existing referring scene understanding datasets and our RefAVA dataset in Tab. 1. The atomic actions involve 80 categories covering Object Manipulation (OM), Person Interactions (PI), and Person Movement (PM). The videos cover diverse scenarios. The test set is sourced from 26 movies different from those in the train set (67 different movies) and val set (34 different movies) to achieve the evaluation of generalizability. + +# 3.2 The Baselines for the RAVAR Benchmark + +We have adapted methodologies from similar fields as our baselines for RAVAR. AAL Baselines. We first reformulate the existing methods from the general atomic video action localization field by integrating the textual reference embeddings into the visual branch. Some action recognition approaches, e.g., I3D [4] and X3D [16], are adapted to AAL following [17]. BERT [14] is used for textual embedding extraction. The selected baselines from this domain can be grouped into CNN-based approaches [4,16] and transformer-based approaches [47,69,81]. All these approaches leverage the pre-trained weight on the Kinetics400 [4]. + +VQA Baselines. The second group of approaches comes from the video question answering domain, which contains the GPT-based model, i.e., AskAnything [44], and conventional transformer-based model [38]. We input the reference sentence in a questioning manner and add a classification head and a bounding box regression head after the acquisition of the logits. We finetuned the two models for the RAVAR task based on their pre-trained weight on ActivityNetQA [92]. + +VTR Baselines. The last group of the foundation model baselines comes from the video-text retrieval task, where XCLIP [56], CLIP4CLIP [55], BLIPv2 [42], and MeVTR [95] are selected. These foundation models are pre-trained on a combination of numerous datasets, incorporating Conceptual Captions [74], SBU Captions [61], and COCO Captions [8], etc. We reformulate these approaches in the same way as the approaches in VQA. + +SF Baselines. Some Single Frame (SF) baselines are adopted, where SAM [33], DETR [3], and REFCLIP [29] are utilized together with the CLIP [67] feature extraction backbone. + +VOS Baseline. We adopt one Video Object Segmentation (VOS) baseline using the encoder of the approach from Su et al. [79] and our prediction head. + +# 4 RefAtomNet: Proposed Method + +# 4.1 Overview + +We propose a new model, RefAtomNet, with an overview provided in Fig. 3. It leverages three token streams: visual, textual reference, and location-semantic streams. Visual tokens are extracted from the video using ViT [15], while textual reference tokens are obtained from the text using BERT [14]. Both streams are enhanced by QFormer [42]. Additionally, in the location-semantic stream, location-semantic tokens are extracted from the center frame by fusing predicted object coordinates and semantic embeddings of the object categories estimated by a frozen object detector and BERT [14]. To suppress irrelevant information for each stream, we propose the agent-based location-semantic aware attentional fusion to achieve better amplification of relevant information during the cross-stream exchange. The detailed structure is discussed in the following subsections. + +# 4.2 RefAtomNet + +Background of QFormer. We rely on BLIPv2 to extract the visual features, where QFormer is the most essential component. Multiple learnable queries are initialized in QFormer as trainable parameters and interact with input data via Transformer attention mechanisms. Each query selectively attends to input parts, capturing task-specific features and updating based on attention outputs. During training, these queries are optimized for specialized information extraction. Queries then directly contribute to generating outputs. + +Extraction of Visual and Textual Reference Tokens. Similar to BLIPv2 [42], we use a pre-trained multimodal model for token extraction from both the video + +and textual reference streams. This model leverages the QFormer [42] architecture to effectively extract and integrate multimodal embeddings from the visual and textual reference data. Specifically, the visual stream employs a ViT [15] backbone encoder to process visual inputs, while the textual branch utilizes BERT [14] to encode textual reference cues. This combined approach facilitates a holistic understanding of both visual content and textual descriptions. The model extracts visual tokens $(\mathbf{t}^{VT})$ and textual reference tokens $(\mathbf{t}^{RT})$ through Eq. 1: + +$$ +\mathbf {t} ^ {V T}, \mathbf {t} ^ {R T} = \mathcal {V} _ {V L} \left(\mathcal {V} _ {V T} \left(\mathbf {x} ^ {V T}\right), \mathcal {V} _ {R T} \left(\mathbf {x} ^ {R T}\right)\right), \tag {1} +$$ + +where $\mathcal{V}_{VL}$ represents the visual-textual integration model (i.e., QFormer [42]). $\mathcal{V}_{VT}$ , and $\mathcal{V}_{RT}$ denote the backbones for extracting visual tokens (i.e., ViT [15]) and textual reference tokens (i.e., BERT [14]), respectively. $\mathbf{x}^{VT}$ and $\mathbf{x}^{RT}$ represent the input video and textual reference caption, which are then fed into liner projection layers, i.e., $\mathbf{P}_{VT}$ and $\mathbf{P}_{RT}$ , respectively. + +Extraction of Location-Semantic Aware Tokens. To integrate more location and semantic information into the token representations, we leverage a well-established object detector DETR [3] to deliver $N_{o}$ detection results based on the input of the center frame of a video clip, noted as the keyframe. Note that, the detections provided by the object detector contain either the human or other objects with a high confidence score according to Eq. 2. + +$$ +\mathbf {r} _ {\text {b o x e s}}, \mathbf {r} _ {\text {c a t s}} = \mathcal {V} _ {\text {d e t s}} (\mathbf {x} _ {k}), \tag {2} +$$ + +where $\mathbf{x}_k$ , $\mathcal{V}_{\text{dets}}$ indicates the keyframe and the detection network. $\mathbf{r}_{\text{boxes}} \in \mathbb{R}^{N_o \times 4}$ are detected 2D corner coordinates (top left and bottom right corners) of the bounding boxes. $\mathbf{r}_{\text{cats}}$ is the predicted category for each bounding box represented in textual format. These category labels of the detected objects are passed to a text encoder (BERT [14]) to extract semantic embeddings. We concatenate these two types of tokens together along the channel dimension and use a single linear projection layer to obtain the aggregated location-semantic aware tokens $(\mathbf{t}^{LS})$ as Eq. 3. + +$$ +\mathbf {t} ^ {L S} = \mathbf {P} _ {L S} \left(\text {C o n c a t} \left[ \mathcal {V} _ {R T} \left(\mathbf {r} _ {\text {c a t s}}\right), \mathbf {r} _ {\text {b o x e s}} \right]\right), \tag {3} +$$ + +where $\mathbf{P}_{LS}$ indicates a fully connected layer, $\mathcal{V}_{RT}$ indicates the language feature extraction backbone, i.e., BERT [14], and Concat indicates the concatenation. + +# Agent-Based Location Semantic Aware Attentional Fusion. + +Agent Tokens and Sequential Processing. We use agent tokens, inspired by agent attention [26]. Derived through fully connected layers from input tokens, similar to Query, Key, and Value extraction in transformers, agent tokens aggregate essential information via agent-key and agent-query pairs. To enable agent attention for 1D sequential tokens from our streams, we redefine the 2D agent spatial tokens proposed in [26] into a 1D sequential format for multi-stream fusion. Specifically, we replace 2D pooling with linear projection and eliminate the depthwise convolutional branch and biased position encoding to better suit agent acquisition in our model. These 1D sequential agent tokens serve dual functions: + +![](images/5b0adef20eec56d83631eb637cb8d669f8ebd6cf8f797bf7b2af8c7e511379d0.jpg) +Fig. 3: An overview of the RefAtomNet architecture. + +aggregating crucial information from agent-key and agent-query pairs and efficiently redistributing this information to the original value tokens, therefore, enhancing attention focus and reducing irrelevant cues. + +Agent-Based Attention Mechanism. The agent-based attention mechanism for visual, textual reference, and location-semantic streams is computed as: + +$$ +\mathbf {Q} ^ {\phi}, \mathbf {K} ^ {\phi}, \mathbf {V} ^ {\phi}, \mathbf {A} ^ {\phi} = \mathbf {W} _ {Q} ^ {\phi} (\mathbf {t} ^ {\phi}), \mathbf {W} _ {K} ^ {\phi} (\mathbf {t} ^ {\phi}), \mathbf {W} _ {V} ^ {\phi} (\mathbf {t} ^ {\phi}), \mathbf {W} _ {A} ^ {\phi} (\mathbf {t} ^ {\phi}), \tag {4} +$$ + +where for better readability, we use $\phi$ to represent $[RT, VT, LS]$ , where RT indicates the reference tokens, VT indicates the visual tokens, and LS indicates the location-semantic tokens. $\mathbf{Q}^{\phi}$ , $\mathbf{K}^{\phi}$ , $\mathbf{V}^{\phi}$ , and $\mathbf{A}^{\phi}$ are the query, key, value, and agent tokens. $t^{\phi}$ depicts the input. $\mathbf{W}_Q^{\phi}$ , $\mathbf{W}_K^{\phi}$ , $\mathbf{W}_V^{\phi}$ , and $\mathbf{W}_A^{\phi}$ are constructed by linear projection layers. We then project the agent tokens using fully connected layers $\mathbf{P}_A^\phi$ through $\mathbf{A}_*^{\phi} = \mathbf{P}_A^\phi (\mathbf{A}^\phi)$ as aforementioned. To mitigate redundancy cues of the textual reference and location-semantic streams, we leverage agent attention for both of these two streams. The agent query attention mask $\mathbf{M}_{QA}^{\pi}$ and the agent key attention mask $\mathbf{M}_{KA}^{\pi}$ are obtained by matrix multiplication (MatMul) and SoftMax operations along the channel dimension (indicated by $\sigma_c$ ) as shown in Eq. 5, where $\pi \in [RT, LS]$ and $\alpha$ indicates a fixed scale factor. The agent attention masks and tokens are computed and refined to ensure that only pertinent information influences the attention mechanism, as shown in Eq. 5 and Eq. 6: + +$$ +\mathbf {M} _ {Q A} ^ {\pi}, \mathbf {M} _ {K A} ^ {\pi} = \sigma_ {c} (M a t M u l [ \alpha * \mathbf {A} _ {*} ^ {\pi}, \mathbf {Q} ^ {\pi} ]), \sigma_ {c} (M a t M u l [ \alpha * \mathbf {A} _ {*} ^ {\pi}, \mathbf {K} ^ {\pi} ]), \tag {5} +$$ + +$$ +\mathbf {t} _ {*} ^ {\pi} = F F N \left(M a t M u l \left[ \mathbf {M} _ {K A} ^ {\pi}, M a t M u l \left[ \mathbf {M} _ {Q A} ^ {\pi}, \mathbf {V} ^ {\pi} \right] \right]\right), \tag {6} +$$ + +where $FFN$ indicates the Feed Forward Network. These steps ensure precise model attention, enhancing the contextual relevance of the resulting tokens. + +Cross-Stream Agent Attention and Agent Token Fusions. Finally, cross-stream agent attention and agent token fusion are proposed, which are explicitly tailored for the visual stream by applying the agent query attention maps and the + +agent tokens from the other two streams. This process involves recalculating the attention maps, thereby ensuring that the final token representation is highly relevant and contextually enriched. We compute the agent query attention map for the visual token stream, where $\gamma = VT$ , as shown in Eq. 7, + +$$ +\mathbf {M} _ {Q A} ^ {\gamma} = \sigma_ {c} (M a t M u l [ \alpha * \mathbf {A} _ {*} ^ {\gamma}, \mathbf {Q} ^ {\gamma} ]). \tag {7} +$$ + +Then, we calculate the cross-stream irrelevance-suppressed attention for the agent query attention of the visual stream through Eq. 8 to achieve cross-stream agent attention fusion to suppress irrelevant information in the agent query attentions of the visual stream, the operations used in the second and third terms are abbreviated as C-ATT and T-ATT in Fig. 3, + +$$ +\hat {\mathbf {M}} _ {Q A} ^ {\gamma} = A V G \left[ \mathbf {M} _ {Q A} ^ {\gamma}, \sigma_ {c} (\sum_ {\pi} \mathbf {M} _ {Q A} ^ {\pi}) * \mathbf {M} _ {Q A} ^ {\gamma}, \sigma_ {t} (\sum_ {\pi} \mathbf {M} _ {Q A} ^ {\pi}) * \mathbf {M} _ {Q A} ^ {\gamma} \right], \quad (8) +$$ + +where $\sigma_{t}$ denotes the SoftMax operation along the token dimension. $AVG$ indicates the mean operation. We follow the same procedure to calculate the cross-stream irrelevance-suppressed agent tokens before the calculation of the agent key attention, as in Eq. 9 to achieve cross-stream agent token fusion. + +$$ +\hat {\mathbf {A}} _ {*} ^ {\gamma} = A V G \left[ \mathbf {A} _ {*} ^ {\gamma} + \sigma_ {c} \left(\sum_ {\pi} \mathbf {A} _ {*} ^ {\pi}\right) * \mathbf {A} _ {*} ^ {\gamma} + \sigma_ {t} \left(\sum_ {\pi} \mathbf {A} _ {*} ^ {\pi}\right) * \mathbf {A} _ {*} ^ {\gamma} \right], \tag {9} +$$ + +then, we calculate the agent key attention as shown in Eq. 10, + +$$ +\hat {\mathbf {M}} _ {K A} ^ {\gamma} = \sigma_ {c} (M a t M u l [ \alpha * \hat {\mathbf {A}} _ {*} ^ {\gamma}, \mathbf {K} ^ {\gamma} ]). \tag {10} +$$ + +The final aggregated visual tokens can be obtained through the following equation as demonstrated in Eq. 11, + +$$ +\mathbf {t} _ {*} ^ {\gamma} = F F N (M a t M u l \left[ \hat {\mathbf {M}} _ {K A} ^ {\gamma}, M a t M u l \left[ \hat {\mathbf {M}} _ {Q A} ^ {\gamma}, \mathbf {V} ^ {\gamma} \right] \right]). \tag {11} +$$ + +Finally, we aggregate all the tokens from three branches by using the mean operation, where $N_{s}$ indicates the number of the stream as shown in Eq. 12, + +$$ +\mathbf {t} _ {a g g} = \sum_ {\phi} \left[ \mathbf {t} _ {*} ^ {\phi} \right] / N _ {s}. \tag {12} +$$ + +We construct MLP-based classification and regression heads atop aggregated tokens for center frame bounding box prediction and atomic action recognition. + +# 4.3 Loss Functions + +We use Binary Cross Entropy (BCE) loss and Mean Squared Error (MSE) loss for the multi-label supervision and the bounding box regression supervision in + +the same way as all the baselines. The equation of the BCE loss is as Eq. 13, + +$$ +L _ {B C E} (\mathbf {y}, \hat {\mathbf {y}}) = - \frac {1}{N _ {c}} \sum_ {i = 1} ^ {N _ {c}} \left[ \mathbf {y} _ {i} \log \left(\hat {\mathbf {y}} _ {i}\right) + \left(1 - \mathbf {y} _ {i}\right) \log \left(1 - \hat {\mathbf {y}} _ {i}\right) \right], \tag {13} +$$ + +where $\mathbf{y}$ , $\hat{\mathbf{y}}$ and $N_{c}$ indicate the one-hot ground truth, the prediction, and the category number. The bounding boxes regression loss is expressed as Eq. 14, + +$$ +L _ {M S E} (\mathbf {b}, \hat {\mathbf {b}}) = \sum_ {j = 1} ^ {4} \left(\mathbf {b} _ {j} - \hat {\mathbf {b}} _ {j}\right) ^ {2}, \tag {14} +$$ + +where $\mathbf{b}$ indicates the coordinates of the left top and right bottom corners, and $\hat{\mathbf{b}}$ represents the predicted boxes. $j$ indicates the coordinate index. + +# 5 Experiments + +# 5.1 Implementation Details + +We conduct experiments on four NVIDIA A100 GPUs. We use BertAdam [32] optimizer with learning rate $lr = 1e^{-4}$ , batch size 128, learning rate decay 0.9, and warmup ratio 0.1. The agent number and head number in our RefAtomNet are 4 and 1. Our model has $214M$ trainable parameters. The weight for $L_{MSE}$ and $\alpha$ are 5 and 0.125. We trained our model for 40 epochs on our dataset. The text encoder is frozen during training. We use multi-label mean Average Precision (mAP), Area Under the Receiver Operating Characteristic curve (AUROC), and the mean Intersection over Union (mIOU) as metrics, of which mIOU has less priority since we prioritize the atomic video action recognition task. + +# 5.2 Experimental Results + +Main Results. Results of the main experiments on the proposed RefAVA benchmark are summarized in Tab. 2. We compare our RefAtomNet model with a multitude of published approaches adopted to suit our task (see Section 3.2) stemming from the fields of (1) Atomic Action Localization (AAL); (2) Video Question Answering (VQA); (3) Video-Text Retrieval (VTR); (4) Single Frame (SF) baselines; and (5) Video Object Segmentation (VOS). Baseline methods from the AAL group consistently underperform in spatially localizing the action according to the textual reference (measured as mIOU). Even though the ability of the human spatial localization is not the major expected output, it can still generally illustrate if the model can grasp the correct referring person or not. Among the AAL methods, X3D [16] achieves the highest AUROC score of $59.09\%$ and $64.51\%$ on the val and test sets, respectively. The VQA and VTR baselines work better than the AAL baselines since they can reasonably capture the referring person while delivering acceptable atomic human action recognition performances, which benefits from the text-aware pretraining of these two + +Table 2: Experimental results on our RAVAR benchmark. + +
MethodmIOUmAPAUROCmIOUmAPAUROC
ValTest
AALI3D [4]0.0044.0457.770.0044.6462.71
X3D [16]0.2644.4559.090.2746.3464.51
MViTv2-B [47]0.7642.3256.430.6642.6059.30
VideoMAE2-B [81]0.1642.0255.120.2941.8759.10
Hiera-B [69]0.4942.7456.720.6241.1458.70
VQASingularity [38]26.3442.4359.5229.7842.1856.12
AskAnything13B [44]20.0951.4266.1222.3552.2569.35
VTRMeVTR [95]30.7838.4251.0129.7936.2752.45
CLIP4CLIP [55]34.7539.4852.5732.3337.1755.05
XCIIIP [56]35.5442.4656.3031.7940.8258.71
BLIPv2 [42]32.9952.1366.5632.7553.1969.92
SFCLIP [67]+SAM [33]32.9549.7464.5529.8051.3469.11
CLIP [67]+DETR [3]33.9647.5761.8333.8450.9267.29
CLIP [67]+REFCLIP [29]34.4347.7962.5032.2849.9067.44
VOSSu et al. [79]23.7152.1766.6726.0253.2070.19
RefAtomNet (Ours)38.2255.9869.7336.4257.5273.95
+ +(a) Module ablation of the RefAtomNet. + +Table 3: Experiments for module ablation and comparison with multi-modal fusion. + +
MethodmIOUmAPAUROCmIOUmAPAUROC
ValTest
w/o ALSAF27.3050.7065.3129.0951.2668.20
w/o CAAF36.2155.4369.6635.3855.9072.60
w/o CATF35.0153.8367.7134.5556.9673.40
w/o LSAS31.9055.2169.4731.2555.7372.38
Ours38.2255.9869.7336.4257.5273.95
+ +(b) Comparison with other fusion approaches. + +
FusionmIOU mAP AUROCmIOU mAP AUROC
ValTest
Addition27.3050.7065.3129.0951.2668.20
Concatenation18.6452.2366.4520.6553.4470.70
Multiplication23.9051.5565.6625.0553.3370.48
AttentionBottleneck [59]33.4750.9765.0733.0254.0271.08
McOmut [98]23.8851.5865.6525.0253.2170.42
Ours38.2255.9869.7336.4257.5273.95
+ +tasks. AskAnything [44] from the VQA group achieves $20.09\%$ and $22.35\%$ of mIOU, $51.42\%$ and $52.25\%$ of mAP, and $66.12\%$ and $69.35\%$ of AUROC on the val and test sets. BLIPv2 [43] from the VTR group delivers $32.99\%$ and $32.75\%$ of mIOU, $52.13\%$ and $53.19\%$ of mAP, and $66.56\%$ and $69.92\%$ of AUROC, on the val and test sets, respectively. Our RefAtomNet achieves state-of-the-art performances, outperforming the best baseline BLIPv2 [42] by $5.23\%$ , $3.85\%$ , $3.17\%$ and $3.67\%$ , $4.33\%$ , $4.03\%$ of mIOU, mAP, and AUROC, on val and test sets. + +Ablations of the Individual Modules. In Tab. 3a, we show the ablation for the components of RefAtomNet by removing each of the proposed designs, where LSAS indicates the location-semantic stream, CAAF indicates the cross-stream agent attention fusion, CATF indicates the cross-stream agent token fusion, and $w / o$ ALSAF indicates by simply using addition to fuse these three streams and without the agent-based location-semantic aware attentional fusion mechanism. Compared with the ablation $w / o$ ALSAF, RefAtomNet achieves promising improvements of $10.92\%$ , $5.28\%$ , $4.42\%$ and $7.33\%$ , $6.26\%$ , $5.75\%$ in terms of mIOU, mAP, and AUROC for the val and test sets, respectively. It indicates that using simple aggregation of the three streams makes the model distract from the important cues for the referring person, illustrated by the large decay of mIOU metric. + +(a) Generalizability to test time rephrasing + +Table 4: Generalizability to different referring styles and encoder architectures. + +
FusionmIOUmAPAUROCmIOUmAPAUROC
ValTest
Singularity [38]18.4541.2758.4720.5441.3955.32
XCLIP [56]31.9541.3554.3529.8440.7458.45
AskAnything [44]19.7451.1165.8321.6051.9669.04
BLIPv2 [42]31.0051.4565.8831.3452.3568.87
Ours34.6555.7569.5233.1457.2373.76
+ +(b) Generalizability to different visual-textual encoder architectures. + +
FusionmIOU mAP AUROCmIOU mAP AUROC
ValTest
XCLIP [56]35.5442.4656.3031.7940.8258.71
RefAtomNet (XCLIP)38.5947.4061.2036.6148.5966.47
BLIPv2 [42]32.9952.1366.5632.7553.1969.92
RefAtomNet (BLIPv2)38.2255.9869.7336.4257.5273.95
+ +Compared with the ablation $w / o$ LSAS, we find that the location-semantic tokens benefit more for the localization of the correct person in the center frame. The advantages of the LSAS on atomic video action recognition metrics are highlighted in the test set which has higher scenario diversity compared with the val set. Using CAAF and CATF, each of them brings promising benefits by suppressing the irrelevant information in the visual tokens. + +Comparison with Other Fusion Mechanisms. We compare our proposed agent-based location-semantic aware attentional fusion mechanism with late fusion addition, multiplication, concatenation, AttentionBottleNeck [59], and McOmet [98] in Tab. 3b. Our approach outperforms all by suppressing those irrelevant visual tokens on the agent tokens and attention masks. + +Generalizability to Test-time Reference Rephrasing. The reference sentences given by different users may differ in daily life scenarios. To test the generalizability of the model towards different referring styles, we invoke API (gpt-3.5-turbo) of ChatGPT [60] to rephrase the test set description two times and then deliver the averaged performance for RAVAR on the original val and test sets, and the two rephrased val and test sets, where we select 4 most out-performing baselines from our benchmark to construct this ablation study, i.e., Singularity [38], AskAnything [44], XCLIP [56], and BLIPv2 [42], as shown in Tab. 4a. RefAtomNet delivers the best performances. We find out that the test time reference rephrasing will cause performance decay for localization and less decay for the atomic video-based action recognition delivered by our model. + +Generalizability to Different Visual Textual Backbones. Since we use the best-performing baseline BLIPv2 [42] as the visual and textual encoder in our model, it would be interesting to see if the proposed architecture can generalize to different encoder backbones. We thereby demonstrate another ablation study in Tab. 4b, where we equip our RefAtomNet with the XCLIP [56] backbone. Compared with the XCLIP baseline, the RefAtomNet (XCLIP) achieves $3.05\%$ , $4.94\%$ , $4.90\%$ and $4.82\%$ , $7.77\%$ , $7.76\%$ performance improvements in terms of mIOU, mAP, and AUROC on the val and test sets, respectively, showing the great generalizability of our RefAtomNet of the visual and textual encoder. + +Analysis of the Qualitative Results. The qualitative results are delivered in Fig. 4. These examples demonstrate the effectiveness of RefAtomNet compared with the best-performing baseline BLIPv2 [42]. Samples 1 and 8 validate the recognition quality for references without any location indications. + +1 Text reference: The man wearing light green clothes. + +![](images/5509deddc417e1eb431442e202ffcd0047277d50bc829982fa456a8dd455d53f.jpg) + +2 Text reference: The man in the center who has gray beard. + +![](images/3d93bf92296f2b83a4e51ab679f12d33af7670d35defd592d8e7224ea981eb67.jpg) + +3 Text reference: The man on the left who is in blue clothes. + +![](images/207de4bb67adb1aa77ee3354b0104f7ccbcd8366ed9c26cdc90b3f59403846e4.jpg) + +4 Text reference: The first man on the right. + +![](images/6e9a141ec02e5100c1693d0b670e3ca15aa7134091578f1895fab80c8ca357fc.jpg) + +5 Text reference: The man wearing glasses on the left. + +![](images/77ae1422bb388a1c7b8712011e9962678e3db58fd7a8ff2050470a04794e7f43.jpg) + +6 Text reference: The man on the right back to us. + +![](images/48870fa17142608cb2eaf39c474afb14562293ea15a75a2991bc83d7fff36449.jpg) + +Text reference: The woman wearing black clothes and pearls necklace on the right. + +![](images/1d1a36e5d72917f6a1a8a4ff9579bda5481aea59dc6f4168360dc80ca8c4f54b.jpg) +Fig. 4: An overview of qualitative results. Missed predictions are marked with a red cross, while true positive and false positive predictions are shown in green and red. + +8 Text reference: The man wearing black suits. + +![](images/28215095954289f8843092093113a56a18904c32160aaa8f6acef65821b6f392.jpg) + +Samples 3, 4, 5, and 6 showcase the results for different humans in a shared scene. Samples 2 and 7 show the RAVAR performance for references containing finer details, e.g., the necklace type and the beard color. Our RefAtomNet outperforms the BLIPv2 baseline on selected samples obviously, attributable to its superior capability in eliminating extraneous visual cues and its advanced location-semantic reasoning prowess. + +# 6 Conclusions + +In this work, we introduce a novel task called Referring Atomic Video Action Recognition (RAVAR). We establish the RefAVA dataset to address the challenge of identifying atomic actions for individuals of interest in videos based on textual descriptions. Existing methods exhibit poor performance on this new task, prompting the development of RefAtomNet, a vision-language architecture that effectively integrates cross-stream tokens for precise referring atomic video action recognition. Through reference-relevant token enhancement and comprehensive token integration, RefAtomNet achieves impressive results, highlighting its effectiveness in tackling the complexities of the RAVAR task. + +# Acknowledgements + +The project served to prepare the SFB 1574 Circular Factory for the Perpetual Product (project ID: 471687386), approved by the German Research Foundation (DFG, German Research Foundation) with a start date of April 1, 2024. This work was also partially supported in part by the SmartAge project sponsored by the Carl Zeiss Stiftung (P2019-01-003; 2021-2026). This work was performed on the HoreKa supercomputer funded by the Ministry of Science, Research and the Arts Baden-Württemberg and by the Federal Ministry of Education and Research. The authors also acknowledge support by the state of Baden-Württemberg through bwHPC and the German Research Foundation (DFG) through grant INST 35/1597-1 FUGG. This project is also supported by the National Key RD Program under Grant 2022YFB4701400. Lastly, the authors thank for the support of Dr. Sepideh Pashami, the Swedish Innovation Agency VINNOVA, the Digital Futures. + +# References + +1. Bagad, P., Tapaswi, M., Snoek, C.G.M.: Test of time: Instilling video-language models with a sense of time. In: CVPR (2023) +2. Bu, Y., Li, L., Xie, J., Liu, Q., Cai, Y., Huang, Q., Li, Q.: Scene-text oriented referring expression comprehension. TMM (2022) +3. Carion, N., Massa, F., Synnaeve, G., Usunier, N., Kirillov, A., Zagoruyko, S.: End-to-end object detection with transformers. In: ECCV (2020) +4. Carreira, J., Zisserman, A.: Quo vadis, action recognition? A new model and the kinetics dataset. In: CVPR (2017) +5. Castro, S., Deng, N., Huang, P., Burzo, M., Mihalcea, R.: In-the-wild video question answering. In: COLING (2022) +6. Chai, W., Guo, X., Wang, G., Lu, Y.: StableVideo: Text-driven consistency-aware diffusion video editing. In: ICCV (2023) +7. Chen, J., Zhu, D., Haydarov, K., Li, X., Elhoseiny, M.: Video ChatCaptioner: Towards enriched spatiotemporal descriptions. arXiv preprint arXiv:2304.04227 (2023) +8. Chen, X., Fang, H., Lin, T.Y., Vedantam, R., Gupta, S., Dolkar, P., Zitnick, C.L.: Microsoft COCO captions: Data collection and evaluation server. arXiv preprint arXiv:1504.00325 (2015) +9. Chen, Y., Wang, J., Lin, L., Qi, Z., Ma, J., Shan, Y.: Tagging before alignment: Integrating multi-modal tags for video-text retrieval. arXiv preprint arXiv:2301.12644 (2023) +0. Chen, Z., Ma, L., Luo, W., Wong, K.Y.K.: Weakly-supervised spatio-temporally grounding natural sentence in video. arXiv preprint arXiv:1906.02549 (2019) +1. Chung, J., Wu, C.h., Yang, H.r., Tai, Y.W., Tang, C.K.: HAA500: Human-centric atomic action dataset with curated videos. In: ICCV (2021) +2. Dang, R., Feng, J., Zhang, H., Ge, C., Song, L., Gong, L., Liu, C., Chen, Q., Zhu, F., Zhao, R., Song, Y.: InstructDET: Diversifying referring object detection with generalized instructions. arXiv preprint arXiv:2310.05136 (2023) +3. Deruyttere, T., Vandenhende, S., Grujicic, D., Van Gool, L., Moens, M.F.: Talk2Car: Taking control of your self-driving car. In: EMNLP (2019) + +14. Devlin, J., Chang, M.W., Lee, K., Toutanova, K.: BERT: Pre-training of deep bidirectional transformers for language understanding. In: ACL (2019) +15. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., Uszkoreit, J., Houlsby, N.: An image is worth 16x16 words: Transformers for image recognition at scale. In: ICLR (2021) +16. Feichtenhofer, C.: X3D: Expanding architectures for efficient video recognition. In: CVPR (2020) +17. Feichtenhofer, C., Fan, H., Malik, J., He, K.: SlowFast networks for video recognition. In: ICCV (2019) +18. Gandhi, M., Gul, M.O., Prakash, E., Grunde-McLaughlin, M., Krishna, R., Agrawala, M.: Measuring compositional consistency for video question answering. In: CVPR (2022) +19. Gao, D., Zhou, L., Ji, L., Zhu, L., Yang, Y., Shou, M.Z.: MIST: Multi-modal iterative spatial-temporal transformer for long-form video question answering. In: CVPR (2023) +20. Garcia, N., Otani, M., Chu, C., Nakashima, Y.: KnowIT VQA: Answering knowledge-based questions about videos. In: AAAI (2020) +21. Gavrilyuk, K., Ghodrati, A., Li, Z., Snoek, C.G.: Actor and action video segmentation from a sentence. In: CVPR (2018) +22. Goyal, R., Kahou, S.E., Michalski, V., Materzynska, J., Westphal, S., Kim, H., Haenel, V., Fründ, I., Yianilos, P., Mueller-Freitag, M., Hoppe, F., Thurau, C., Bax, I., Memisevic, R.: The "something something" video database for learning and evaluating visual common sense. In: ICCV (2017) +23. Gritsenko, A., Xiong, X., Djolonga, J., Dehghani, M., Sun, C., Lucic, M., Schmid, C., Arnab, A.: End-to-end spatio-temporal action localisation with video transformers. arXiv preprint arXiv:2304.12160 (2023) +24. Gu, C., Sun, C., Ross, D.A., Vondrick, C., Pantofaru, C., Li, Y., Vijayanarasimhan, S., Toderici, G., Ricco, S., Sukthankar, R., Schmid, C., Malik, J.: AVA: A video dataset of spatio-temporally localized atomic visual actions. In: CVPR (2018) +25. Guo, W., Zhang, Y., Yang, J., Yuan, X.: Re-attention for visual question answering. TIP (2021) +26. Han, D., Ye, T., Han, Y., Xia, Z., Song, S., Huang, G.: Agent attention: On the integration of softmax and linear attention. arXiv preprint arXiv:2312.08874 (2023) +27. Ji, Y., Zhan, Y., Yang, Y., Xu, X., Shen, F., Shen, H.T.: A context knowledge map guided coarse-to-fine action recognition. TIP (2020) +28. Jiang, J., Chen, Z., Lin, H., Zhao, X., Gao, Y.: Divide and conquer: Question-guided spatio-temporal contextual attention for video question answering. In: AAAI (2020) +29. Jin, L., Luo, G., Zhou, Y., Sun, X., Jiang, G., Shu, A., Ji, R.: Refclip: A universal teacher for weakly supervised referring expression comprehension. In: CVPR (2023) +30. Khoreva, A., Rohrbach, A., Schiele, B.: Video object segmentation with language referring expressions. In: ACCV (2019) +31. Kim, M., Spinola, F., Benz, P., Kim, T.h.: A*: Atrous spatial temporal action recognition for real time applications. In: WACV (2024) +32. Kingma, D.P., Ba, J.: Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980 (2014) +33. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., et al.: Segment anything. In: CVPR (2023) +34. Kuehne, H., Jhuang, H., Garrote, E., Poggio, T., Serre, T.: HMDB: A large video database for human motion recognition. In: ICCV (2011) + +35. Laput, G., Harrison, C.: Sensing fine-grained hand activity with smartwatches. In: CHI (2019) +36. Le, T.M., Le, V., Venkatesh, S., Tran, T.: Hierarchical conditional relation networks for video question answering. In: CVPR (2020) +37. Lea, C., Vidal, R., Hager, G.D.: Learning convolutional action primitives for fine-grained action recognition. In: ICRA (2016) +38. Lei, J., Berg, T.L., Bansal, M.: Revealing single frame bias for video-and-language learning. arXiv preprint arXiv:2206.03428 (2022) +39. Lei, J., Li, L., Zhou, L., Gan, Z., Berg, T.L., Bansal, M., Liu, J.: Less is more: ClipBERT for video-and-language learning via sparse sampling. In: CVPR (2021) +40. Li, G., Wei, Y., Tian, Y., Xu, C., Wen, J.R., Hu, D.: Learning to answer questions in dynamic audio-visual scenarios. In: CVPR (2022) +41. Li, J., Niu, L., Zhang, L.: From representation to reasoning: Towards both evidence and commonsense reasoning for video question-answering. In: CVPR (2022) +42. Li, J., Li, D., Savarese, S., Hoi, S.: BLIP-2: Bootstrapping language-image pretraining with frozen image encoders and large language models. In: ICML (2023) +43. Li, J., Li, D., Xiong, C., Hoi, S.: BLIP: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In: ICML (2022) +44. Li, K., He, Y., Wang, Y., Li, Y., Wang, W., Luo, P., Wang, Y., Wang, L., Qiao, Y.: VideoChat: Chat-centric video understanding. arXiv preprint arXiv:2305.06355 (2023) +45. Li, L., Gan, Z., Lin, K., Lin, C.C., Liu, Z., Liu, C., Wang, L.: LAVENDER: Unifying video-language understanding as masked language modeling. In: CVPR (2023) +46. Li, R., Li, K., Kuo, Y.C., Shu, M., Qi, X., Shen, X., Jia, J.: Referring image segmentation via recurrent refinement networks. In: CVPR (2018) +47. Li, Y., Wu, C.Y., Fan, H., Mangalam, K., Xiong, B., Malik, J., Feichtenhofer, C.: MViTv2: Improved multiscale vision transformers for classification and detection. In: CVPR (2022) +48. Lin, J., Chen, J., Peng, K., He, X., Li, Z., Stiefelhagen, R., Yang, K.: EchoTrack: Auditory referring multi-object tracking for autonomous driving. arXiv preprint arXiv:2402.18302 (2024) +49. Lin, X., Tiwari, S., Huang, S., Li, M., Shou, M.Z., Ji, H., Chang, S.F.: Towards fast adaptation of pretrained contrastive models for multi-channel video-language retrieval. In: CVPR (2023) +50. Liu, J., Wang, L., Yang, M.H.: Referring expression generation and comprehension via attributes. In: ICCV (2017) +51. Liu, R., Zhang, J., Peng, K., Zheng, J., Cao, K., Chen, Y., Yang, K., Stiefelhagen, R.: Open scene understanding: Grounded situation recognition meets segment anything for helping people with visual impairments. In: ICCVW (2023) +52. Liu, R., Liu, C., Bai, Y., Yuille, A.L.: CLEVR-Ref+: Diagnosing visual reasoning with referring expressions. In: CVPR (2019) +53. Liu, S., Hui, T., Huang, S., Wei, Y., Li, B., Li, G.: Cross-modal progressive comprehension for referring segmentation. TPAMI (2021) +54. Liu, Y., Li, G., Lin, L.: Cross-modal causal relational reasoning for event-level visual question answering. TPAMI (2023) +55. Luo, H., Ji, L., Zhong, M., Chen, Y., Lei, W., Duan, N., Li, T.: CLIP4Clip: An empirical study of CLIP for end to end video clip retrieval and captioning. Neurocomputing (2022) +56. Ma, Y., Xu, G., Sun, X., Yan, M., Zhang, J., Ji, R.: X-CLIP: End-to-end multigrained contrastive learning for video-text retrieval. In: MM (2022) + +57. Madasu, A., Aflalo, E., Ben Melech Stan, G., Tseng, S.Y., Bertasius, G., Lal, V.: Improving video retrieval using multilingual knowledge transfer. In: ECIR (2023) +58. McIntosh, B., Duarte, K., Rawat, Y.S., Shah, M.: Visual-textual capsule routing for text-based video segmentation. In: CVPR (2020) +59. Nagrani, A., Yang, S., Arnab, A., Jansen, A., Schmid, C., Sun, C.: Attention bottlenecks for multimodal fusion. In: NeuIPS (2021) +60. OpenAI: ChatGPT: Optimizing language models for dialogue. https://openai.com/ (2022) +61. Ordonez, V., Kulkarni, G., Berg, T.: Im2Text: Describing images using 1 million captioned photographs. In: NeurIPS (2011) +62. Ou, W., Zhang, J., Peng, K., Yang, K., Jaworek, G., Müller, K., Stiefelhagen, R.: Indoor navigation assistance for visually impaired people via dynamic SLAM and panoptic segmentation with an RGB-D sensor. In: ICCHP (2022) +63. Peng, K., Roitberg, A., Yang, K., Zhang, J., Stiefelhagen, R.: TransDARC: Transformer-based driver activity recognition with latent space feature calibration. In: IROS (2022) +64. Pramanick, P., Sarkar, C., Paul, S., dev Roychoudhury, R., Bhowmick, B.: DoRO: Disambiguation of referred object for embodied agents. RA-L (2022) +65. Pramono, R.R.A., Chen, Y.T., Fang, W.H.: Spatial-temporal action localization with hierarchical self-attention. TMM (2021) +66. Qiu, H., Li, H., Wu, Q., Meng, F., Shi, H., Zhao, T., Ngan, K.N.: Language-aware fine-grained object representation for referring expression comprehension. In: MM (2020) +67. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., Krueger, G., Sutskever, I.: Learning transferable visual models from natural language supervision. In: ICML (2021) +68. Rajasegaran, J., Pavlakos, G., Kanazawa, A., Feichtenhofer, C., Malik, J.: On the benefits of 3D pose and tracking for human action recognition. In: CVPR (2023) +69. Ryali, C., Hu, Y., Bolya, D., Wei, C., Fan, H., Huang, P., Aggarwal, V., Chowdhury, A., Poursaeed, O., Hoffman, J., Malik, J., Li, Y., Feichtenhofer, C.: Hiera: A hierarchical vision transformer without the bells-and-whistles. In: ICML (2023) +70. Saha, J., Chowdhury, C., Chowdury, I.R., Roy, P.: Fine grained activity recognition using smart handheld. In: ICDCN (2018) +71. Seibold, C.M., Reif, S., Kleesiek, J., Stiefelhagen, R.: Reference-guided pseudolabel generation for medical semantic segmentation. In: AAAI (2022) +72. Seo, S., Lee, J.Y., Han, B.: URVOS: Unified referring video object segmentation network with a large-scale benchmark. In: ECCV (2020) +73. Shao, D., Zhao, Y., Dai, B., Lin, D.: Finegym: A hierarchical video dataset for fine-grained action understanding. In: CVPR (2020) +74. Sharma, P., Ding, N., Goodman, S., Soricut, R.: Conceptual captions: A cleaned, hypernymed, image alt-text dataset for automatic image captioning. In: ACL (2018) +75. Shi, H., Pan, W., Zhao, Z., Zhang, M., Wu, F.: Unsupervised domain adaptation for referring semantic segmentation. In: MM (2023) +76. Shi, H., Li, H., Meng, F., Wu, Q.: Key-word-aware network for referring expression image segmentation. In: ECCV (2018) +77. Shi, Y., Xu, H., Yuan, C., Li, B., Hu, W., Zha, Z.J.: Learning video-text aligned representations for video captioning. TOMM (2023) +78. Soomro, K., Zamir, A.R., Shah, M.: UCF101: A dataset of 101 human actions classes from videos in the wild. arXiv preprint arXiv:1212.0402 (2012) + +79. Su, Y., Wang, W., Liu, J., Ma, S., Yang, X.: Sequence as a whole: A unified framework for video action localization with long-range text query. TIP (2023) +80. Vasudevan, A.B., Dai, D., Van Gool, L.: Object referring in videos with language and human gaze. In: CVPR (2018) +81. Wang, L., Huang, B., Zhao, Z., Tong, Z., He, Y., Wang, Y., Wang, Y., Qiao, Y.: VideoMAE V2: Scaling video masked autoencoders with dual masking. In: CVPR (2023) +82. Wang, M., Xing, J., Mei, J., Liu, Y., Jiang, Y.: ActionCLIP: Adapting language-image pretrained models for video action recognition. TNNLS (2023) +83. Wang, R., Chen, D., Wu, Z., Chen, Y., Dai, X., Liu, M., Yuan, L., Jiang, Y.G.: Masked video distillation: Rethinking masked feature modeling for self-supervised video representation learning. In: CVPR (2023) +84. Wang, S., Yan, R., Huang, P., Dai, G., Song, Y., Shu, X.: Com-STAL: Compositional spatio-temporal action localization. TCSVT (2023) +85. Wang, Y., Li, K., Li, Y., He, Y., Huang, B., Zhao, Z., Zhang, H., Xu, J., Liu, Y., Wang, Z., Xing, S., Chen, G., Pan, J., Yu, J., Wang, Y., Wang, L., Qiao, Y.: InternVideo: General video foundation models via generative and discriminative learning. arXiv preprint arXiv:2212.03191 (2022) +86. Wu, D., Han, W., Wang, T., Dong, X., Zhang, X., Shen, J.: Referring multi-object tracking. In: CVPR (2023) +87. Wu, W., Luo, H., Fang, B., Wang, J., Ouyang, W.: Cap4Video: What can auxiliary captions do for text-video retrieval? In: CVPR (2023) +88. Xiao, J., Shang, X., Yao, A., Chua, T.S.: NExT-QA: Next phase of question-answering to explaining temporal actions. In: CVPR (2021) +89. Yang, A., Miech, A., Sivic, J., Laptev, I., Schmid, C.: Just ask: Learning to answer questions from millions of narrated videos. In: ICCV (2021) +90. Yang, P., Wang, X., Duan, X., Chen, H., Hou, R., Jin, C., Zhu, W.: AVQA: A dataset for audio-visual question answering on videos. In: MM (2022) +91. Yu, L., Poirson, P., Yang, S., Berg, A.C., Berg, T.L.: Modeling context in referring expressions. In: ECCV (2016) +92. Yu, Z., Xu, D., Yu, J., Yu, T., Zhao, Z., Zhuang, Y., Tao, D.: ActivityNet-QA: A dataset for understanding complex web videos via question answering. In: AAAI (2019) +93. Yuan, Z., Yan, X., Liao, Y., Zhang, R., Wang, S., Li, Z., Cui, S.: InstanceRefer: Cooperative holistic understanding for visual grounding on point clouds through instance multi-level contextual referring. In: ICCV (2021) +94. Zeng, F., Dong, B., Zhang, Y., Wang, T., Zhang, X., Wei, Y.: MOTR: End-to-end multiple-object tracking with transformer. In: ECCV (2022) +95. Zhang, G., Ren, J., Gu, J., Tresp, V.: Multi-event video-text retrieval. In: CVPR (2023) +96. Zhang, H., Li, X., Bing, L.: Video-LLaMA: An instruction-tuned audio-visual language model for video understanding. In: EMNLP (2023) +97. Zheng, J., Zhang, J., Yang, K., Peng, K., Stiefelhagen, R.: MateRobot: Material recognition in wearable robotics for people with visual impairments. In: ICRA (2024) +98. Zong, D., Sun, S.: McOmet: Multimodal fusion transformer for physical audiovisual commonsense reasoning. In: AAAI (2023) \ No newline at end of file diff --git a/2024/Referring Atomic Video Action Recognition/images.zip b/2024/Referring Atomic Video Action Recognition/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..857f5455f8875e000749b8086694bb72b6057794 --- /dev/null +++ b/2024/Referring Atomic Video Action Recognition/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e7cce582ffdc71379460bb542400b8f4da6802f161a86162bdc43a53e98d9058 +size 580145 diff --git a/2024/Referring Atomic Video Action Recognition/layout.json b/2024/Referring Atomic Video Action Recognition/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..09a49efc2d6ee7f1d4a6000d1569e2f7964fb60b --- /dev/null +++ b/2024/Referring Atomic Video Action Recognition/layout.json @@ -0,0 +1,11578 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 153, + 112, + 460, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 153, + 112, + 460, + 129 + ], + "spans": [ + { + "bbox": [ + 153, + 112, + 460, + 129 + ], + "type": "text", + "content": "Referring Atomic Video Action Recognition" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 138, + 149, + 474, + 187 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 149, + 474, + 187 + ], + "spans": [ + { + "bbox": [ + 138, + 149, + 474, + 187 + ], + "type": "text", + "content": "Kunyu Peng" + }, + { + "bbox": [ + 138, + 149, + 474, + 187 + ], + "type": "inline_equation", + "content": "^{1,*}" + }, + { + "bbox": [ + 138, + 149, + 474, + 187 + ], + "type": "text", + "content": ", Jia Fu" + }, + { + "bbox": [ + 138, + 149, + 474, + 187 + ], + "type": "inline_equation", + "content": "^{3,4,*}" + }, + { + "bbox": [ + 138, + 149, + 474, + 187 + ], + "type": "text", + "content": ", Kailun Yang" + }, + { + "bbox": [ + 138, + 149, + 474, + 187 + ], + "type": "inline_equation", + "content": "^{2,\\dagger}" + }, + { + "bbox": [ + 138, + 149, + 474, + 187 + ], + "type": "text", + "content": ", Di Wen" + }, + { + "bbox": [ + 138, + 149, + 474, + 187 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 138, + 149, + 474, + 187 + ], + "type": "text", + "content": ", Yufan Chen" + }, + { + "bbox": [ + 138, + 149, + 474, + 187 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 138, + 149, + 474, + 187 + ], + "type": "text", + "content": ", Ruiping Liu" + }, + { + "bbox": [ + 138, + 149, + 474, + 187 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 138, + 149, + 474, + 187 + ], + "type": "text", + "content": ", Junwei Zheng" + }, + { + "bbox": [ + 138, + 149, + 474, + 187 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 138, + 149, + 474, + 187 + ], + "type": "text", + "content": ", Jiaming Zhang" + }, + { + "bbox": [ + 138, + 149, + 474, + 187 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 138, + 149, + 474, + 187 + ], + "type": "text", + "content": ", M. Saqib Sarfraz" + }, + { + "bbox": [ + 138, + 149, + 474, + 187 + ], + "type": "inline_equation", + "content": "^{1,6}" + }, + { + "bbox": [ + 138, + 149, + 474, + 187 + ], + "type": "text", + "content": ", Rainer Stiefelhagen" + }, + { + "bbox": [ + 138, + 149, + 474, + 187 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 138, + 149, + 474, + 187 + ], + "type": "text", + "content": ", and Alina Roitberg" + }, + { + "bbox": [ + 138, + 149, + 474, + 187 + ], + "type": "inline_equation", + "content": "^{5}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 153, + 195, + 460, + 230 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 153, + 195, + 460, + 230 + ], + "spans": [ + { + "bbox": [ + 153, + 195, + 460, + 230 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 153, + 195, + 460, + 230 + ], + "type": "text", + "content": "Karlsruhe Institute of Technology, " + }, + { + "bbox": [ + 153, + 195, + 460, + 230 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 153, + 195, + 460, + 230 + ], + "type": "text", + "content": "Hunan University, " + }, + { + "bbox": [ + 153, + 195, + 460, + 230 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 153, + 195, + 460, + 230 + ], + "type": "text", + "content": "RISE Research Institutes of Sweden, " + }, + { + "bbox": [ + 153, + 195, + 460, + 230 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 153, + 195, + 460, + 230 + ], + "type": "text", + "content": "KTH Royal Institute of Technology, " + }, + { + "bbox": [ + 153, + 195, + 460, + 230 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 153, + 195, + 460, + 230 + ], + "type": "text", + "content": "University of Stuttgart, " + }, + { + "bbox": [ + 153, + 195, + 460, + 230 + ], + "type": "inline_equation", + "content": "^{6}" + }, + { + "bbox": [ + 153, + 195, + 460, + 230 + ], + "type": "text", + "content": "Mercedes-Benz Tech Innovation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 160, + 251, + 455, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 251, + 455, + 515 + ], + "spans": [ + { + "bbox": [ + 160, + 251, + 455, + 515 + ], + "type": "text", + "content": "Abstract. We introduce a new task called Referring Atomic Video Action Recognition (RAVAR), aimed at identifying atomic actions of a particular person based on a textual description and the video data of this person. This task differs from traditional action recognition and localization, where predictions are delivered for all present individuals. In contrast, we focus on recognizing the correct atomic action of a specific individual, guided by text. To explore this task, we present the RefAVA dataset, containing 36,630 instances with manually annotated textual descriptions of the individuals. To establish a strong initial benchmark, we implement and validate baselines from various domains, e.g., atomic action localization, video question answering, and text-video retrieval. Since these existing methods underperform on RAVAR, we introduce RefAtomNet - a novel cross-stream attention-driven method specialized for the unique challenges of RAVAR: the need to interpret a textual referring expression for the targeted individual, utilize this reference to guide the spatial localization and harvest the prediction of the atomic actions for the referring person. The key ingredients are: (1) a multi-stream architecture that connects video, text, and a new location-semantic stream, and (2) cross-stream agent attention fusion and agent token fusion which amplify the most relevant information across these streams and consistently surpasses standard attention-based fusion on RAVAR. Extensive experiments demonstrate the effectiveness of RefAtomNet and its building blocks for recognizing the action of the described individual. The dataset and code will be made publicly available at RAVAR." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 531, + 230, + 544 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 531, + 230, + 544 + ], + "spans": [ + { + "bbox": [ + 132, + 531, + 230, + 544 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 553, + 482, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 553, + 482, + 639 + ], + "spans": [ + { + "bbox": [ + 130, + 553, + 482, + 639 + ], + "type": "text", + "content": "Referring scene understanding [50, 52, 66, 93] aims to solve the underlying computer vision task for the particular scene element specified via a natural language referring expression. Context-driven queries are vital in information retrieval, assistive systems, and multimedia analysis, but incorporating them into computer vision models is challenging due to a complex entanglement of linguistic descriptions, localization, and visual recognition itself. Several benchmarks for referring scene understanding have been successfully established for multi-object" + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 643, + 222, + 653 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 643, + 222, + 653 + ], + "spans": [ + { + "bbox": [ + 133, + 643, + 222, + 653 + ], + "type": "text", + "content": "* Equal contribution" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 136, + 653, + 311, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 653, + 311, + 665 + ], + "spans": [ + { + "bbox": [ + 136, + 653, + 311, + 665 + ], + "type": "text", + "content": "† Correspondence: kailun.yang@hnu.edu.cn" + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 136, + 115, + 479, + 189 + ], + "blocks": [ + { + "bbox": [ + 136, + 115, + 479, + 189 + ], + "lines": [ + { + "bbox": [ + 136, + 115, + 479, + 189 + ], + "spans": [ + { + "bbox": [ + 136, + 115, + 479, + 189 + ], + "type": "image", + "image_path": "f616d00f7bfd11a1e4da9bb0839d5a77f557c291b420ae8731c50413dc1fc6fb.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 169, + 199, + 443, + 210 + ], + "lines": [ + { + "bbox": [ + 169, + 199, + 443, + 210 + ], + "spans": [ + { + "bbox": [ + 169, + 199, + 443, + 210 + ], + "type": "text", + "content": "Fig. 1: A comparison of AAL Task (left) and RAVAR task (right)." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 234, + 481, + 317 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 234, + 481, + 317 + ], + "spans": [ + { + "bbox": [ + 130, + 234, + 481, + 317 + ], + "type": "text", + "content": "tracking [86], semantic segmentation [30, 46, 72, 75, 76], medical imaging [71], and object detection [12, 64]. While these efforts focus on object-centric referring expressions, many applications require a human-centric understanding of human actions, e.g., in rehabilitation assistance [35, 70] and human-robot interaction [27, 37]. Our work addresses a new problem - reference-driven recognition of atomic actions performed by the human described via a textual reference, which has been overlooked in the past." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 319, + 481, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 319, + 481, + 533 + ], + "spans": [ + { + "bbox": [ + 130, + 319, + 481, + 533 + ], + "type": "text", + "content": "The field of human action recognition has made exciting advances, from new approaches stemming from the rise of visual transformers [23,47,63,69,81,83,85] to new large-scale datasets [4,22,34,73,78]. Datasets utilized for the prediction of atomic actions often feature recordings with multiple individuals [24]. Nevertheless, the vast majority of previously published works [23,31,68,69,81] either rely on manually window crop of a particular person of interest or automatically generated region of interests to predict the atomic actions for all the individuals in one video, which results in large pre- or post-processing effort to focus on one specific person. This workflow is also very inconvenient in certain applications, such as assistive systems for users with visual impairments [51,62,97], who must understand the state of each present in the scene for effective interactions. Leveraging succinct textual descriptions, which may include broad positional indicators (e.g., left, center, and right), appearance attributes (e.g., hair color and clothing), or gender, to steer the model towards delivering atomic action recognition outcomes for the targeted individual, presents a promising solution to the outlined challenges. Employing these textual cues as a reference, the model can facilitate an end-to-end retrieval of the specified instance across the entire video, subsequently providing precise atomic action recognition results." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 534, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 534, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 534, + 482, + 665 + ], + "type": "text", + "content": "To close this gap, we formalize the new task of Referring Atomic Video Action Recognition (RAVAR). The differences between RAVAR and atomic action localization are shown in Fig. 1. The conventional atomic action recognition in multi-person scenarios is often formalized in an action localization manner, and requires the region of interest to localize the human available a priori, after which atomic action recognition is carried out for each individual. Post-processing is required for individuals of interest. In contrast, RAVAR assumes a textual reference and a video as inputs, delivering the referring individual's atomic actions along with the location. This comparison shows the efficiency of RAVAR, leveraging textual information for less labor-intensive subject identification to assist atomic action analysis for the individual of interest. We further" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 250, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 250, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 250, + 102 + ], + "type": "text", + "content": "K. Peng, J. Fu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 187 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 187 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 187 + ], + "type": "text", + "content": "introduce the new RefAVA dataset for RAVAR. RefAVA is established using the 17,946 video clips of the public AVA dataset [24] initially designed for atomic action localization, which we extend with 36,630 textual descriptions of the corresponding individuals used as the referring expression. The dataset covers diverse settings, e.g., in/out door and day/night time scenarios, and numerous multi-person scenes, constituting an ideal testbed for our task." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 190, + 482, + 308 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 190, + 482, + 308 + ], + "spans": [ + { + "bbox": [ + 130, + 190, + 482, + 308 + ], + "type": "text", + "content": "To establish an initial benchmark, we assess 15 well-established approaches from multiple related research domains. Our experiments reveal that none of the benchmarked methods delivers sufficient performance. For example, previous methods within the domains of VQA and VTR typically yield predictions of a coarse analysis regarding actions, whereas strategies in AAL face challenges in spatially localizing the correct referring individual. We attribute this phenomenon to the massive irrelevant information provided by the visual data, distracting the model from the referring location. How to suppress the referring irrelevant information becomes a critical challenge to pursue better RAVAR performances." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 311, + 482, + 526 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 311, + 482, + 526 + ], + "spans": [ + { + "bbox": [ + 130, + 311, + 482, + 526 + ], + "type": "text", + "content": "To address this, we introduce the new RefAtomNet approach specialized in simultaneously solving the tasks of understanding textual referring expressions and using them as guides for spatial localization and classification of atomic actions. We begin by computing textual, visual, and newly proposed location-semantic tokens using a large foundation model and well-established object detector. The latter stream is designed to harvest semantic cues of different scene entities: it leverages a pre-trained object detector to localize and identify objects, after which the location-semantic tokens are computed as a fusion of visual bounding box coordinates and text embeddings of the object category. Another novel aspect of our work is the introduction of cross-stream agent attention and agent token fusions, which builds on the concept of agent attention [26] and re-defines it for cross-stream compatibility: agent tokens are reformulated into a 1D sequence format through the use of fully connected layers, bypassing the initial need for 2D pooling, removing the depthwise convolution branch, and forgiving 2D positional encodings. This mechanism enhances the model's ability to filter and emphasize relevant information from multiple data sources effectively. RefAtomNet achieves the highest performance for RAVAR, while a detailed ablation study demonstrates the effectiveness of each component." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 146, + 529, + 362, + 539 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 529, + 362, + 539 + ], + "spans": [ + { + "bbox": [ + 146, + 529, + 362, + 539 + ], + "type": "text", + "content": "Our contributions can be summarized as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 137, + 554, + 484, + 665 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 138, + 554, + 484, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 554, + 484, + 590 + ], + "spans": [ + { + "bbox": [ + 138, + 554, + 484, + 590 + ], + "type": "text", + "content": "- We introduce the new Referring Atomic Video Action Recognition (RAVAR) task and the RefAVA benchmark with 36,630 atomic action instances manually annotated with suitable language expressions." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 592, + 481, + 627 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 592, + 481, + 627 + ], + "spans": [ + { + "bbox": [ + 138, + 592, + 481, + 627 + ], + "type": "text", + "content": "- To establish a competitive benchmark, we examine 11 well-established models from different related fields, namely atomic action localization, video question answering, and video-text retrieval." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 137, + 630, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 630, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 137, + 630, + 481, + 665 + ], + "type": "text", + "content": "- We propose RefAtomNet, a new RAVAR approach that uses a novel agent-based semantic-location aware attentional fusion to integrate multi-modal tokens, suppressing the irrelevant visual cues. RefAtomNet delivers improve" + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 413, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 413, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 413, + 91, + 447, + 100 + ], + "type": "text", + "content": "RAVAR" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 148, + 115, + 482, + 141 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 148, + 115, + 482, + 141 + ], + "spans": [ + { + "bbox": [ + 148, + 115, + 482, + 141 + ], + "type": "text", + "content": "ments of " + }, + { + "bbox": [ + 148, + 115, + 482, + 141 + ], + "type": "inline_equation", + "content": "3.85\\%" + }, + { + "bbox": [ + 148, + 115, + 482, + 141 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 148, + 115, + 482, + 141 + ], + "type": "inline_equation", + "content": "3.17\\%" + }, + { + "bbox": [ + 148, + 115, + 482, + 141 + ], + "type": "text", + "content": " of mAP, and " + }, + { + "bbox": [ + 148, + 115, + 482, + 141 + ], + "type": "inline_equation", + "content": "4.33\\%" + }, + { + "bbox": [ + 148, + 115, + 482, + 141 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 148, + 115, + 482, + 141 + ], + "type": "inline_equation", + "content": "4.03\\%" + }, + { + "bbox": [ + 148, + 115, + 482, + 141 + ], + "type": "text", + "content": " of AUROC, on val and test sets, respectively, compared with the leading baseline, BLIPv2 [42]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 159, + 238, + 172 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 159, + 238, + 172 + ], + "spans": [ + { + "bbox": [ + 132, + 159, + 238, + 172 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 186, + 482, + 353 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 186, + 482, + 353 + ], + "spans": [ + { + "bbox": [ + 130, + 186, + 482, + 353 + ], + "type": "text", + "content": "Referring Scene Understanding. Referring scene understanding aims to locate parts of interest within images or videos guided by natural language, exhibiting the utility in many computer vision tasks including autonomous driving [86] and video editing [6]. The development of this field cannot be advanced without the contribution of high-quality open-source datasets and benchmarks [2, 12, 30, 48, 72, 80, 86, 91]. For example, the CLEVR-Ref+ benchmark is proposed by Liu et al. [52] to achieve visual reasoning with referring expressions. Li et al. [46] proposed referring image segmentation by using a recurrent refinement network. Wu et al. [86] proposed referring multi-object tracking benchmark. However, there are no referring understanding works focusing on atomic video action analysis, whereupon we first conduct RAVAR benchmarking with our RefAVA dataset. We further clarify that Referring Video Object Segmentation (RVOS) task [21, 53, 58, 79] differs us from multiple perspectives, e.g., RVOS includes action name in input textual reference." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 354, + 483, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 354, + 483, + 496 + ], + "spans": [ + { + "bbox": [ + 130, + 354, + 483, + 496 + ], + "type": "text", + "content": "Video Text Retrieval. Video Text Retrieval (VTR) aims at matching relevant video content with text. Amidst the rise of Vision-Language foundation models like CLIP [67] and BLIP [43], efforts [9,42,49,55-57,77,82,87,95] are being made to apply powerful pre-trained models for promoting their competencies in VTR. XCLIP [56] introduced multi-grained contrastive learning for end-to-end VTR, by counting all the video-sentence, video-word, sentence-frame, and frame-word contrasts. BLIPv2 [42] pre-trained a Querying Transformer (QFormer) to bootstrap vision-and-language representation learning and vision-to-language generative learning. The key difference between VTR and RAVAR lies in their inference tasks: VTR queries text or video to localize elements, while RAVAR integrates textual and visual data for fine-grained subject retrieval and atomic action prediction in videos." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 498, + 483, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 498, + 483, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 498, + 483, + 666 + ], + "type": "text", + "content": "Video Question Answering. Video Question Answering (VQA) focuses on generating answers to questions posed in natural language for a given video. Depending on the emphasis of the question, factoid VQA [1,5,7,19,20,25,28,36,39, 45,54,88,89,96] straightforward queries visual facts. Lei et al. [38] proposed that a single-frame trained transformer-based model, with large-scale pre-training and a frame ensemble at the inference stage, can perform better than existing multi-frame trained models in factoid VQA tasks. On the other hand, inference VQA [18,41] delves into logical reasoning. Li et al. [44] released a video-centric instruction dataset and leveraged a neural interface to integrate video foundation models and Large Language Models (LLM), showcasing capability in temporal reasoning, causal inference, and event localization. In addition, some multimodal VQA frameworks [20,40,90] explore information-invoking scenarios that incorporate visual, audio, subtitle, and external knowledge. However, compared with the RAVAR task, most of the existing VQA approaches do not particularly focus" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 250, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 250, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 250, + 102 + ], + "type": "text", + "content": "K. Peng, J. Fu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 132, + 168, + 479, + 204 + ], + "blocks": [ + { + "bbox": [ + 130, + 114, + 479, + 159 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 479, + 159 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 479, + 159 + ], + "type": "text", + "content": "Table 1: An overview of the referring scene understanding datasets and our RefAVA dataset; Our task is Referring Atomic Video Action Recognition (RAVAR), whereas existing benchmarks focus on Referring Object Detection (ROD), Referring Video-based Object Segmentation (RVOS), and Referring Multi-Object Tracking (RMOT)." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 132, + 168, + 479, + 204 + ], + "lines": [ + { + "bbox": [ + 132, + 168, + 479, + 204 + ], + "spans": [ + { + "bbox": [ + 132, + 168, + 479, + 204 + ], + "type": "table", + "html": "
DatasetRefCOCO [91]RefCOCO+ [91]RefCOCOg [91]Talk2Car [13]VID-Sentence [10]Refer-DAVUS17 [30]Refer-YV [94]Refer-KITTI [86]RefAVA
TaskRODRODRODRODRODRVOSRMOTRMOTRAVAR
NFrames26,71119,99226,7119,21759,2384,21993,8696,6501,615,140
NInstance26,71119,99226,71110,5197,6543,9787,451-36,630
", + "image_path": "9a26a67e3a8f3b1b1ded71c2923706c13761f81a525867ae06384322ec1bc114.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 227, + 479, + 262 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 227, + 479, + 262 + ], + "spans": [ + { + "bbox": [ + 130, + 227, + 479, + 262 + ], + "type": "text", + "content": "on atomic action recognition and will only tend to give a coarse textual description, which limits the applications requiring precise prediction, e.g., human-robot assistance [27, 37]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 262, + 482, + 539 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 262, + 482, + 539 + ], + "spans": [ + { + "bbox": [ + 130, + 262, + 482, + 539 + ], + "type": "text", + "content": "Atomic Video Action Recognition and Localization. Atomic video-based action recognition [11, 23] and localization [24] involves the identification and analysis of the most fundamental, indivisible actions or movements performed by humans for single and multiple-person scenarios. Compared with the general video action recognition task, atomic video-based action localization is much more fine-grained and is always formulated in a multi-label manner with bounding box predictions. Most of the existing convolutional neural networks (CNN) [4, 16, 17, 81] and the transformer [23, 47, 63, 69, 81, 83, 85] networks for human action recognition are commonly used in the atomic video-based action localization by changing the classification head into multi-label manner, adding additional bounding box prediction head, and integrating region of interest features from human detector. Ryali et al. [69] proposed a hierarchical vision transformer with high efficiency and precision within the realm of the existing methods using video as input. Wang et al. [81] proposed scaling video-masked autoencoders with dual masking. Current methods for predicting individual actions in multi-person scenarios often require manual video cropping for atomic action recognition or generate predictions for all detected individuals, necessitating further human selection and reducing practicality [65, 68, 84]. Most existing AAL methods are not specifically designed for the RAVAR task. We thereby introduce RefAtomNet, a novel method that utilizes location semantics atop predicted location and scene semantic information derived from the scenario together with the textual reference and visual cues, then further uses cross-stream agent attention and agent token fusions to suppress redundant information." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 557, + 349, + 570 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 557, + 349, + 570 + ], + "spans": [ + { + "bbox": [ + 131, + 557, + 349, + 570 + ], + "type": "text", + "content": "3 RAVAR: Established Benchmark" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 584, + 343, + 594 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 584, + 343, + 594 + ], + "spans": [ + { + "bbox": [ + 131, + 584, + 343, + 594 + ], + "type": "text", + "content": "3.1 Introduction of the RefAVA Dataset" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 605, + 479, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 479, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 479, + 665 + ], + "type": "text", + "content": "Textual Annotations. To acquire precise textual annotations for the individuals of interest, 7 annotators manually provided the textual annotations according to the key frame bounding boxes presented in the AVA dataset [24]. Cross-checking among all the annotators for the annotated individuals is conducted to deliver high textual annotation quality." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 413, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 413, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 413, + 91, + 447, + 100 + ], + "type": "text", + "content": "RAVAR" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 138, + 118, + 312, + 231 + ], + "blocks": [ + { + "bbox": [ + 138, + 118, + 312, + 231 + ], + "lines": [ + { + "bbox": [ + 138, + 118, + 312, + 231 + ], + "spans": [ + { + "bbox": [ + 138, + 118, + 312, + 231 + ], + "type": "image", + "image_path": "3e5fe960a155d8fceeda66643690b3c23e8f632cccd672b5859fa1e8c986b73c.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 147, + 232, + 303, + 242 + ], + "lines": [ + { + "bbox": [ + 147, + 232, + 303, + 242 + ], + "spans": [ + { + "bbox": [ + 147, + 232, + 303, + 242 + ], + "type": "text", + "content": "(a) Chord visualization of several keywords." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 316, + 114, + 480, + 219 + ], + "blocks": [ + { + "bbox": [ + 316, + 114, + 480, + 219 + ], + "lines": [ + { + "bbox": [ + 316, + 114, + 480, + 219 + ], + "spans": [ + { + "bbox": [ + 316, + 114, + 480, + 219 + ], + "type": "image", + "image_path": "230b39ea41268350f85d973bcbaa8ac917c594a95c1db74f928394e6e59eff56.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 315, + 221, + 481, + 247 + ], + "lines": [ + { + "bbox": [ + 315, + 221, + 481, + 247 + ], + "spans": [ + { + "bbox": [ + 315, + 221, + 481, + 247 + ], + "type": "text", + "content": "(b) Annotation statistics: #The number of annotated persons contained in the scenario (x-axis); #The number of instances (y-axis)." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 131, + 255, + 482, + 289 + ], + "lines": [ + { + "bbox": [ + 131, + 255, + 482, + 289 + ], + "spans": [ + { + "bbox": [ + 131, + 255, + 482, + 289 + ], + "type": "text", + "content": "Fig. 2: An chord visualization of several keywords from the textual references, shown on the left, and the instance amount in video clips containing different numbers of annotated persons in our RefAVA dataset, shown on the right." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 313, + 482, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 313, + 482, + 529 + ], + "spans": [ + { + "bbox": [ + 130, + 313, + 482, + 529 + ], + "type": "text", + "content": "Dataset. We selected 17,946 video clips from 127 movies of the AVA dataset, preserving the most complex scenarios, and annotated each person on each center frame of every 90s video clip. In total, RefAVA has 36,630 labeled instances, and we split them into 22,658 train instances, 10,916 validation instances, and 3,056 test instances, where the samples from different sets are from different movie scenarios. The textual annotations cover information on approximate age, gender, appearance, relative position in the center frame, etc., and without action description. The chord visualization in Fig. 2a reveals the quantitative causal relations between the most frequent notional words in different categories from the textual references. Further, Fig. 2b shows the statistics of instance amount towards the annotated person number inside the scenario where each instance belongs. We deliver the comparison of the statistics between representative existing referring scene understanding datasets and our RefAVA dataset in Tab. 1. The atomic actions involve 80 categories covering Object Manipulation (OM), Person Interactions (PI), and Person Movement (PM). The videos cover diverse scenarios. The test set is sourced from 26 movies different from those in the train set (67 different movies) and val set (34 different movies) to achieve the evaluation of generalizability." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 548, + 373, + 559 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 548, + 373, + 559 + ], + "spans": [ + { + "bbox": [ + 131, + 548, + 373, + 559 + ], + "type": "text", + "content": "3.2 The Baselines for the RAVAR Benchmark" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "type": "text", + "content": "We have adapted methodologies from similar fields as our baselines for RAVAR. AAL Baselines. We first reformulate the existing methods from the general atomic video action localization field by integrating the textual reference embeddings into the visual branch. Some action recognition approaches, e.g., I3D [4] and X3D [16], are adapted to AAL following [17]. BERT [14] is used for textual embedding extraction. The selected baselines from this domain can be grouped into CNN-based approaches [4,16] and transformer-based approaches [47,69,81]. All these approaches leverage the pre-trained weight on the Kinetics400 [4]." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 250, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 250, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 250, + 102 + ], + "type": "text", + "content": "K. Peng, J. Fu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 113, + 481, + 187 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 113, + 481, + 187 + ], + "spans": [ + { + "bbox": [ + 130, + 113, + 481, + 187 + ], + "type": "text", + "content": "VQA Baselines. The second group of approaches comes from the video question answering domain, which contains the GPT-based model, i.e., AskAnything [44], and conventional transformer-based model [38]. We input the reference sentence in a questioning manner and add a classification head and a bounding box regression head after the acquisition of the logits. We finetuned the two models for the RAVAR task based on their pre-trained weight on ActivityNetQA [92]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 188, + 481, + 259 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 188, + 481, + 259 + ], + "spans": [ + { + "bbox": [ + 130, + 188, + 481, + 259 + ], + "type": "text", + "content": "VTR Baselines. The last group of the foundation model baselines comes from the video-text retrieval task, where XCLIP [56], CLIP4CLIP [55], BLIPv2 [42], and MeVTR [95] are selected. These foundation models are pre-trained on a combination of numerous datasets, incorporating Conceptual Captions [74], SBU Captions [61], and COCO Captions [8], etc. We reformulate these approaches in the same way as the approaches in VQA." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 259, + 481, + 295 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 259, + 481, + 295 + ], + "spans": [ + { + "bbox": [ + 130, + 259, + 481, + 295 + ], + "type": "text", + "content": "SF Baselines. Some Single Frame (SF) baselines are adopted, where SAM [33], DETR [3], and REFCLIP [29] are utilized together with the CLIP [67] feature extraction backbone." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 296, + 481, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 296, + 481, + 319 + ], + "spans": [ + { + "bbox": [ + 131, + 296, + 481, + 319 + ], + "type": "text", + "content": "VOS Baseline. We adopt one Video Object Segmentation (VOS) baseline using the encoder of the approach from Su et al. [79] and our prediction head." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 339, + 344, + 353 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 339, + 344, + 353 + ], + "spans": [ + { + "bbox": [ + 131, + 339, + 344, + 353 + ], + "type": "text", + "content": "4 RefAtomNet: Proposed Method" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 365, + 208, + 376 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 365, + 208, + 376 + ], + "spans": [ + { + "bbox": [ + 131, + 365, + 208, + 376 + ], + "type": "text", + "content": "4.1 Overview" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 386, + 481, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 386, + 481, + 518 + ], + "spans": [ + { + "bbox": [ + 130, + 386, + 481, + 518 + ], + "type": "text", + "content": "We propose a new model, RefAtomNet, with an overview provided in Fig. 3. It leverages three token streams: visual, textual reference, and location-semantic streams. Visual tokens are extracted from the video using ViT [15], while textual reference tokens are obtained from the text using BERT [14]. Both streams are enhanced by QFormer [42]. Additionally, in the location-semantic stream, location-semantic tokens are extracted from the center frame by fusing predicted object coordinates and semantic embeddings of the object categories estimated by a frozen object detector and BERT [14]. To suppress irrelevant information for each stream, we propose the agent-based location-semantic aware attentional fusion to achieve better amplification of relevant information during the cross-stream exchange. The detailed structure is discussed in the following subsections." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 536, + 225, + 548 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 536, + 225, + 548 + ], + "spans": [ + { + "bbox": [ + 131, + 536, + 225, + 548 + ], + "type": "text", + "content": "4.2 RefAtomNet" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 557, + 481, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 557, + 481, + 641 + ], + "spans": [ + { + "bbox": [ + 130, + 557, + 481, + 641 + ], + "type": "text", + "content": "Background of QFormer. We rely on BLIPv2 to extract the visual features, where QFormer is the most essential component. Multiple learnable queries are initialized in QFormer as trainable parameters and interact with input data via Transformer attention mechanisms. Each query selectively attends to input parts, capturing task-specific features and updating based on attention outputs. During training, these queries are optimized for specialized information extraction. Queries then directly contribute to generating outputs." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 641, + 487, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 487, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 487, + 665 + ], + "type": "text", + "content": "Extraction of Visual and Textual Reference Tokens. Similar to BLIPv2 [42], we use a pre-trained multimodal model for token extraction from both the video" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 413, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 413, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 413, + 91, + 447, + 100 + ], + "type": "text", + "content": "RAVAR" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 99 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 99 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 99 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 210 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 210 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 210 + ], + "type": "text", + "content": "and textual reference streams. This model leverages the QFormer [42] architecture to effectively extract and integrate multimodal embeddings from the visual and textual reference data. Specifically, the visual stream employs a ViT [15] backbone encoder to process visual inputs, while the textual branch utilizes BERT [14] to encode textual reference cues. This combined approach facilitates a holistic understanding of both visual content and textual descriptions. The model extracts visual tokens " + }, + { + "bbox": [ + 130, + 116, + 482, + 210 + ], + "type": "inline_equation", + "content": "(\\mathbf{t}^{VT})" + }, + { + "bbox": [ + 130, + 116, + 482, + 210 + ], + "type": "text", + "content": " and textual reference tokens " + }, + { + "bbox": [ + 130, + 116, + 482, + 210 + ], + "type": "inline_equation", + "content": "(\\mathbf{t}^{RT})" + }, + { + "bbox": [ + 130, + 116, + 482, + 210 + ], + "type": "text", + "content": " through Eq. 1:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 217, + 211, + 481, + 225 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 217, + 211, + 481, + 225 + ], + "spans": [ + { + "bbox": [ + 217, + 211, + 481, + 225 + ], + "type": "interline_equation", + "content": "\\mathbf {t} ^ {V T}, \\mathbf {t} ^ {R T} = \\mathcal {V} _ {V L} \\left(\\mathcal {V} _ {V T} \\left(\\mathbf {x} ^ {V T}\\right), \\mathcal {V} _ {R T} \\left(\\mathbf {x} ^ {R T}\\right)\\right), \\tag {1}", + "image_path": "a4a8e321081f4ff8eb02b9bd3b7d5e0d27b6c4e3e11e6163800d15072eb6b8c6.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 230, + 482, + 290 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 230, + 482, + 290 + ], + "spans": [ + { + "bbox": [ + 130, + 230, + 482, + 290 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 230, + 482, + 290 + ], + "type": "inline_equation", + "content": "\\mathcal{V}_{VL}" + }, + { + "bbox": [ + 130, + 230, + 482, + 290 + ], + "type": "text", + "content": " represents the visual-textual integration model (i.e., QFormer [42]). " + }, + { + "bbox": [ + 130, + 230, + 482, + 290 + ], + "type": "inline_equation", + "content": "\\mathcal{V}_{VT}" + }, + { + "bbox": [ + 130, + 230, + 482, + 290 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 230, + 482, + 290 + ], + "type": "inline_equation", + "content": "\\mathcal{V}_{RT}" + }, + { + "bbox": [ + 130, + 230, + 482, + 290 + ], + "type": "text", + "content": " denote the backbones for extracting visual tokens (i.e., ViT [15]) and textual reference tokens (i.e., BERT [14]), respectively. " + }, + { + "bbox": [ + 130, + 230, + 482, + 290 + ], + "type": "inline_equation", + "content": "\\mathbf{x}^{VT}" + }, + { + "bbox": [ + 130, + 230, + 482, + 290 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 230, + 482, + 290 + ], + "type": "inline_equation", + "content": "\\mathbf{x}^{RT}" + }, + { + "bbox": [ + 130, + 230, + 482, + 290 + ], + "type": "text", + "content": " represent the input video and textual reference caption, which are then fed into liner projection layers, i.e., " + }, + { + "bbox": [ + 130, + 230, + 482, + 290 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_{VT}" + }, + { + "bbox": [ + 130, + 230, + 482, + 290 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 230, + 482, + 290 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_{RT}" + }, + { + "bbox": [ + 130, + 230, + 482, + 290 + ], + "type": "text", + "content": ", respectively." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 290, + 482, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 290, + 482, + 361 + ], + "spans": [ + { + "bbox": [ + 130, + 290, + 482, + 361 + ], + "type": "text", + "content": "Extraction of Location-Semantic Aware Tokens. To integrate more location and semantic information into the token representations, we leverage a well-established object detector DETR [3] to deliver " + }, + { + "bbox": [ + 130, + 290, + 482, + 361 + ], + "type": "inline_equation", + "content": "N_{o}" + }, + { + "bbox": [ + 130, + 290, + 482, + 361 + ], + "type": "text", + "content": " detection results based on the input of the center frame of a video clip, noted as the keyframe. Note that, the detections provided by the object detector contain either the human or other objects with a high confidence score according to Eq. 2." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 254, + 373, + 481, + 385 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 254, + 373, + 481, + 385 + ], + "spans": [ + { + "bbox": [ + 254, + 373, + 481, + 385 + ], + "type": "interline_equation", + "content": "\\mathbf {r} _ {\\text {b o x e s}}, \\mathbf {r} _ {\\text {c a t s}} = \\mathcal {V} _ {\\text {d e t s}} (\\mathbf {x} _ {k}), \\tag {2}", + "image_path": "6dac602fa8d80fc1940ca1402e0364f954af353797c4bc484467b96e890e1db8.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 394, + 482, + 490 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 394, + 482, + 490 + ], + "spans": [ + { + "bbox": [ + 130, + 394, + 482, + 490 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 394, + 482, + 490 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_k" + }, + { + "bbox": [ + 130, + 394, + 482, + 490 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 394, + 482, + 490 + ], + "type": "inline_equation", + "content": "\\mathcal{V}_{\\text{dets}}" + }, + { + "bbox": [ + 130, + 394, + 482, + 490 + ], + "type": "text", + "content": " indicates the keyframe and the detection network. " + }, + { + "bbox": [ + 130, + 394, + 482, + 490 + ], + "type": "inline_equation", + "content": "\\mathbf{r}_{\\text{boxes}} \\in \\mathbb{R}^{N_o \\times 4}" + }, + { + "bbox": [ + 130, + 394, + 482, + 490 + ], + "type": "text", + "content": " are detected 2D corner coordinates (top left and bottom right corners) of the bounding boxes. " + }, + { + "bbox": [ + 130, + 394, + 482, + 490 + ], + "type": "inline_equation", + "content": "\\mathbf{r}_{\\text{cats}}" + }, + { + "bbox": [ + 130, + 394, + 482, + 490 + ], + "type": "text", + "content": " is the predicted category for each bounding box represented in textual format. These category labels of the detected objects are passed to a text encoder (BERT [14]) to extract semantic embeddings. We concatenate these two types of tokens together along the channel dimension and use a single linear projection layer to obtain the aggregated location-semantic aware tokens " + }, + { + "bbox": [ + 130, + 394, + 482, + 490 + ], + "type": "inline_equation", + "content": "(\\mathbf{t}^{LS})" + }, + { + "bbox": [ + 130, + 394, + 482, + 490 + ], + "type": "text", + "content": " as Eq. 3." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 220, + 498, + 481, + 513 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 220, + 498, + 481, + 513 + ], + "spans": [ + { + "bbox": [ + 220, + 498, + 481, + 513 + ], + "type": "interline_equation", + "content": "\\mathbf {t} ^ {L S} = \\mathbf {P} _ {L S} \\left(\\text {C o n c a t} \\left[ \\mathcal {V} _ {R T} \\left(\\mathbf {r} _ {\\text {c a t s}}\\right), \\mathbf {r} _ {\\text {b o x e s}} \\right]\\right), \\tag {3}", + "image_path": "4a40ee968ce0d91f728ac2764d4b4714bc0971a5c4b4d806b131ec59848d82a5.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 521, + 482, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 482, + 545 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 482, + 545 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 521, + 482, + 545 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_{LS}" + }, + { + "bbox": [ + 130, + 521, + 482, + 545 + ], + "type": "text", + "content": " indicates a fully connected layer, " + }, + { + "bbox": [ + 130, + 521, + 482, + 545 + ], + "type": "inline_equation", + "content": "\\mathcal{V}_{RT}" + }, + { + "bbox": [ + 130, + 521, + 482, + 545 + ], + "type": "text", + "content": " indicates the language feature extraction backbone, i.e., BERT [14], and Concat indicates the concatenation." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 545, + 430, + 556 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 545, + 430, + 556 + ], + "spans": [ + { + "bbox": [ + 131, + 545, + 430, + 556 + ], + "type": "text", + "content": "Agent-Based Location Semantic Aware Attentional Fusion." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "text", + "content": "Agent Tokens and Sequential Processing. We use agent tokens, inspired by agent attention [26]. Derived through fully connected layers from input tokens, similar to Query, Key, and Value extraction in transformers, agent tokens aggregate essential information via agent-key and agent-query pairs. To enable agent attention for 1D sequential tokens from our streams, we redefine the 2D agent spatial tokens proposed in [26] into a 1D sequential format for multi-stream fusion. Specifically, we replace 2D pooling with linear projection and eliminate the depthwise convolutional branch and biased position encoding to better suit agent acquisition in our model. These 1D sequential agent tokens serve dual functions:" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 250, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 250, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 250, + 102 + ], + "type": "text", + "content": "K. Peng, J. Fu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 134, + 115, + 479, + 270 + ], + "blocks": [ + { + "bbox": [ + 134, + 115, + 479, + 270 + ], + "lines": [ + { + "bbox": [ + 134, + 115, + 479, + 270 + ], + "spans": [ + { + "bbox": [ + 134, + 115, + 479, + 270 + ], + "type": "image", + "image_path": "5b0adef20eec56d83631eb637cb8d669f8ebd6cf8f797bf7b2af8c7e511379d0.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 199, + 278, + 414, + 289 + ], + "lines": [ + { + "bbox": [ + 199, + 278, + 414, + 289 + ], + "spans": [ + { + "bbox": [ + 199, + 278, + 414, + 289 + ], + "type": "text", + "content": "Fig. 3: An overview of the RefAtomNet architecture." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 311, + 479, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 311, + 479, + 346 + ], + "spans": [ + { + "bbox": [ + 130, + 311, + 479, + 346 + ], + "type": "text", + "content": "aggregating crucial information from agent-key and agent-query pairs and efficiently redistributing this information to the original value tokens, therefore, enhancing attention focus and reducing irrelevant cues." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 346, + 479, + 371 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 346, + 479, + 371 + ], + "spans": [ + { + "bbox": [ + 130, + 346, + 479, + 371 + ], + "type": "text", + "content": "Agent-Based Attention Mechanism. The agent-based attention mechanism for visual, textual reference, and location-semantic streams is computed as:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 182, + 376, + 481, + 392 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 182, + 376, + 481, + 392 + ], + "spans": [ + { + "bbox": [ + 182, + 376, + 481, + 392 + ], + "type": "interline_equation", + "content": "\\mathbf {Q} ^ {\\phi}, \\mathbf {K} ^ {\\phi}, \\mathbf {V} ^ {\\phi}, \\mathbf {A} ^ {\\phi} = \\mathbf {W} _ {Q} ^ {\\phi} (\\mathbf {t} ^ {\\phi}), \\mathbf {W} _ {K} ^ {\\phi} (\\mathbf {t} ^ {\\phi}), \\mathbf {W} _ {V} ^ {\\phi} (\\mathbf {t} ^ {\\phi}), \\mathbf {W} _ {A} ^ {\\phi} (\\mathbf {t} ^ {\\phi}), \\tag {4}", + "image_path": "a4b879b8a37e61890e64c7fc59da5ad36fe668d22e58846fc426ed523b0c6789.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 396, + 482, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 396, + 482, + 564 + ], + "spans": [ + { + "bbox": [ + 130, + 396, + 482, + 564 + ], + "type": "text", + "content": "where for better readability, we use " + }, + { + "bbox": [ + 130, + 396, + 482, + 564 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 130, + 396, + 482, + 564 + ], + "type": "text", + "content": " to represent " + }, + { + "bbox": [ + 130, + 396, + 482, + 564 + ], + "type": "inline_equation", + "content": "[RT, VT, LS]" + }, + { + "bbox": [ + 130, + 396, + 482, + 564 + ], + "type": "text", + "content": ", where RT indicates the reference tokens, VT indicates the visual tokens, and LS indicates the location-semantic tokens. " + }, + { + "bbox": [ + 130, + 396, + 482, + 564 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}^{\\phi}" + }, + { + "bbox": [ + 130, + 396, + 482, + 564 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 396, + 482, + 564 + ], + "type": "inline_equation", + "content": "\\mathbf{K}^{\\phi}" + }, + { + "bbox": [ + 130, + 396, + 482, + 564 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 396, + 482, + 564 + ], + "type": "inline_equation", + "content": "\\mathbf{V}^{\\phi}" + }, + { + "bbox": [ + 130, + 396, + 482, + 564 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 396, + 482, + 564 + ], + "type": "inline_equation", + "content": "\\mathbf{A}^{\\phi}" + }, + { + "bbox": [ + 130, + 396, + 482, + 564 + ], + "type": "text", + "content": " are the query, key, value, and agent tokens. " + }, + { + "bbox": [ + 130, + 396, + 482, + 564 + ], + "type": "inline_equation", + "content": "t^{\\phi}" + }, + { + "bbox": [ + 130, + 396, + 482, + 564 + ], + "type": "text", + "content": " depicts the input. " + }, + { + "bbox": [ + 130, + 396, + 482, + 564 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_Q^{\\phi}" + }, + { + "bbox": [ + 130, + 396, + 482, + 564 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 396, + 482, + 564 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_K^{\\phi}" + }, + { + "bbox": [ + 130, + 396, + 482, + 564 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 396, + 482, + 564 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_V^{\\phi}" + }, + { + "bbox": [ + 130, + 396, + 482, + 564 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 396, + 482, + 564 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_A^{\\phi}" + }, + { + "bbox": [ + 130, + 396, + 482, + 564 + ], + "type": "text", + "content": " are constructed by linear projection layers. We then project the agent tokens using fully connected layers " + }, + { + "bbox": [ + 130, + 396, + 482, + 564 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_A^\\phi" + }, + { + "bbox": [ + 130, + 396, + 482, + 564 + ], + "type": "text", + "content": " through " + }, + { + "bbox": [ + 130, + 396, + 482, + 564 + ], + "type": "inline_equation", + "content": "\\mathbf{A}_*^{\\phi} = \\mathbf{P}_A^\\phi (\\mathbf{A}^\\phi)" + }, + { + "bbox": [ + 130, + 396, + 482, + 564 + ], + "type": "text", + "content": " as aforementioned. To mitigate redundancy cues of the textual reference and location-semantic streams, we leverage agent attention for both of these two streams. The agent query attention mask " + }, + { + "bbox": [ + 130, + 396, + 482, + 564 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_{QA}^{\\pi}" + }, + { + "bbox": [ + 130, + 396, + 482, + 564 + ], + "type": "text", + "content": " and the agent key attention mask " + }, + { + "bbox": [ + 130, + 396, + 482, + 564 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_{KA}^{\\pi}" + }, + { + "bbox": [ + 130, + 396, + 482, + 564 + ], + "type": "text", + "content": " are obtained by matrix multiplication (MatMul) and SoftMax operations along the channel dimension (indicated by " + }, + { + "bbox": [ + 130, + 396, + 482, + 564 + ], + "type": "inline_equation", + "content": "\\sigma_c" + }, + { + "bbox": [ + 130, + 396, + 482, + 564 + ], + "type": "text", + "content": ") as shown in Eq. 5, where " + }, + { + "bbox": [ + 130, + 396, + 482, + 564 + ], + "type": "inline_equation", + "content": "\\pi \\in [RT, LS]" + }, + { + "bbox": [ + 130, + 396, + 482, + 564 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 396, + 482, + 564 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 130, + 396, + 482, + 564 + ], + "type": "text", + "content": " indicates a fixed scale factor. The agent attention masks and tokens are computed and refined to ensure that only pertinent information influences the attention mechanism, as shown in Eq. 5 and Eq. 6:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 157, + 570, + 480, + 585 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 157, + 570, + 480, + 585 + ], + "spans": [ + { + "bbox": [ + 157, + 570, + 480, + 585 + ], + "type": "interline_equation", + "content": "\\mathbf {M} _ {Q A} ^ {\\pi}, \\mathbf {M} _ {K A} ^ {\\pi} = \\sigma_ {c} (M a t M u l [ \\alpha * \\mathbf {A} _ {*} ^ {\\pi}, \\mathbf {Q} ^ {\\pi} ]), \\sigma_ {c} (M a t M u l [ \\alpha * \\mathbf {A} _ {*} ^ {\\pi}, \\mathbf {K} ^ {\\pi} ]), \\tag {5}", + "image_path": "b8356e240aa30a1e66c002290db50f25b9f7f46af4174cf78ba09fe1e51e9c45.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 200, + 586, + 480, + 601 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 200, + 586, + 480, + 601 + ], + "spans": [ + { + "bbox": [ + 200, + 586, + 480, + 601 + ], + "type": "interline_equation", + "content": "\\mathbf {t} _ {*} ^ {\\pi} = F F N \\left(M a t M u l \\left[ \\mathbf {M} _ {K A} ^ {\\pi}, M a t M u l \\left[ \\mathbf {M} _ {Q A} ^ {\\pi}, \\mathbf {V} ^ {\\pi} \\right] \\right]\\right), \\tag {6}", + "image_path": "0d6ea7037813e8d71873acd9212846c3739cb08080040399d5da8a8e2669f137.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 605, + 479, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 479, + 628 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 479, + 628 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 605, + 479, + 628 + ], + "type": "inline_equation", + "content": "FFN" + }, + { + "bbox": [ + 130, + 605, + 479, + 628 + ], + "type": "text", + "content": " indicates the Feed Forward Network. These steps ensure precise model attention, enhancing the contextual relevance of the resulting tokens." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 630, + 479, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 630, + 479, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 630, + 479, + 665 + ], + "type": "text", + "content": "Cross-Stream Agent Attention and Agent Token Fusions. Finally, cross-stream agent attention and agent token fusion are proposed, which are explicitly tailored for the visual stream by applying the agent query attention maps and the" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 413, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 413, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 413, + 91, + 447, + 100 + ], + "type": "text", + "content": "RAVAR" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 164 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 164 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 164 + ], + "type": "text", + "content": "agent tokens from the other two streams. This process involves recalculating the attention maps, thereby ensuring that the final token representation is highly relevant and contextually enriched. We compute the agent query attention map for the visual token stream, where " + }, + { + "bbox": [ + 130, + 116, + 482, + 164 + ], + "type": "inline_equation", + "content": "\\gamma = VT" + }, + { + "bbox": [ + 130, + 116, + 482, + 164 + ], + "type": "text", + "content": ", as shown in Eq. 7," + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 231, + 174, + 481, + 188 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 231, + 174, + 481, + 188 + ], + "spans": [ + { + "bbox": [ + 231, + 174, + 481, + 188 + ], + "type": "interline_equation", + "content": "\\mathbf {M} _ {Q A} ^ {\\gamma} = \\sigma_ {c} (M a t M u l [ \\alpha * \\mathbf {A} _ {*} ^ {\\gamma}, \\mathbf {Q} ^ {\\gamma} ]). \\tag {7}", + "image_path": "cca82158ef0d7a690ff6f62dbb8e5faa53a49262c767556a9e7442ea1b915556.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 195, + 482, + 256 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 195, + 482, + 256 + ], + "spans": [ + { + "bbox": [ + 131, + 195, + 482, + 256 + ], + "type": "text", + "content": "Then, we calculate the cross-stream irrelevance-suppressed attention for the agent query attention of the visual stream through Eq. 8 to achieve cross-stream agent attention fusion to suppress irrelevant information in the agent query attentions of the visual stream, the operations used in the second and third terms are abbreviated as C-ATT and T-ATT in Fig. 3," + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 161, + 274, + 481, + 307 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 161, + 274, + 481, + 307 + ], + "spans": [ + { + "bbox": [ + 161, + 274, + 481, + 307 + ], + "type": "interline_equation", + "content": "\\hat {\\mathbf {M}} _ {Q A} ^ {\\gamma} = A V G \\left[ \\mathbf {M} _ {Q A} ^ {\\gamma}, \\sigma_ {c} (\\sum_ {\\pi} \\mathbf {M} _ {Q A} ^ {\\pi}) * \\mathbf {M} _ {Q A} ^ {\\gamma}, \\sigma_ {t} (\\sum_ {\\pi} \\mathbf {M} _ {Q A} ^ {\\pi}) * \\mathbf {M} _ {Q A} ^ {\\gamma} \\right], \\quad (8)", + "image_path": "efae4287192371d7e33cf3ac78cd56ef834898871d7667b4a4092ad5e2ec0dd8.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 312, + 482, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 312, + 482, + 361 + ], + "spans": [ + { + "bbox": [ + 131, + 312, + 482, + 361 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 312, + 482, + 361 + ], + "type": "inline_equation", + "content": "\\sigma_{t}" + }, + { + "bbox": [ + 131, + 312, + 482, + 361 + ], + "type": "text", + "content": " denotes the SoftMax operation along the token dimension. " + }, + { + "bbox": [ + 131, + 312, + 482, + 361 + ], + "type": "inline_equation", + "content": "AVG" + }, + { + "bbox": [ + 131, + 312, + 482, + 361 + ], + "type": "text", + "content": " indicates the mean operation. We follow the same procedure to calculate the cross-stream irrelevance-suppressed agent tokens before the calculation of the agent key attention, as in Eq. 9 to achieve cross-stream agent token fusion." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 183, + 369, + 481, + 403 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 183, + 369, + 481, + 403 + ], + "spans": [ + { + "bbox": [ + 183, + 369, + 481, + 403 + ], + "type": "interline_equation", + "content": "\\hat {\\mathbf {A}} _ {*} ^ {\\gamma} = A V G \\left[ \\mathbf {A} _ {*} ^ {\\gamma} + \\sigma_ {c} \\left(\\sum_ {\\pi} \\mathbf {A} _ {*} ^ {\\pi}\\right) * \\mathbf {A} _ {*} ^ {\\gamma} + \\sigma_ {t} \\left(\\sum_ {\\pi} \\mathbf {A} _ {*} ^ {\\pi}\\right) * \\mathbf {A} _ {*} ^ {\\gamma} \\right], \\tag {9}", + "image_path": "35c7152d1cce09ed99e77360aa41ef1dad9565eef2dd02e94d1a5f982614d835.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 406, + 408, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 406, + 408, + 418 + ], + "spans": [ + { + "bbox": [ + 132, + 406, + 408, + 418 + ], + "type": "text", + "content": "then, we calculate the agent key attention as shown in Eq. 10," + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 228, + 426, + 481, + 447 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 228, + 426, + 481, + 447 + ], + "spans": [ + { + "bbox": [ + 228, + 426, + 481, + 447 + ], + "type": "interline_equation", + "content": "\\hat {\\mathbf {M}} _ {K A} ^ {\\gamma} = \\sigma_ {c} (M a t M u l [ \\alpha * \\hat {\\mathbf {A}} _ {*} ^ {\\gamma}, \\mathbf {K} ^ {\\gamma} ]). \\tag {10}", + "image_path": "5ccc654ab11a4dd7e6211fdc742430a9df610848cdc205c5415b503f16917abd.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 453, + 482, + 477 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 453, + 482, + 477 + ], + "spans": [ + { + "bbox": [ + 131, + 453, + 482, + 477 + ], + "type": "text", + "content": "The final aggregated visual tokens can be obtained through the following equation as demonstrated in Eq. 11," + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 192, + 486, + 481, + 507 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 192, + 486, + 481, + 507 + ], + "spans": [ + { + "bbox": [ + 192, + 486, + 481, + 507 + ], + "type": "interline_equation", + "content": "\\mathbf {t} _ {*} ^ {\\gamma} = F F N (M a t M u l \\left[ \\hat {\\mathbf {M}} _ {K A} ^ {\\gamma}, M a t M u l \\left[ \\hat {\\mathbf {M}} _ {Q A} ^ {\\gamma}, \\mathbf {V} ^ {\\gamma} \\right] \\right]). \\tag {11}", + "image_path": "c0355094af4ee6e7f3e37bc0df73594524f77dde32df2dfde6f9b9826c94ca7d.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 131, + 514, + 482, + 538 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 514, + 482, + 538 + ], + "spans": [ + { + "bbox": [ + 131, + 514, + 482, + 538 + ], + "type": "text", + "content": "Finally, we aggregate all the tokens from three branches by using the mean operation, where " + }, + { + "bbox": [ + 131, + 514, + 482, + 538 + ], + "type": "inline_equation", + "content": "N_{s}" + }, + { + "bbox": [ + 131, + 514, + 482, + 538 + ], + "type": "text", + "content": " indicates the number of the stream as shown in Eq. 12," + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 264, + 547, + 481, + 571 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 264, + 547, + 481, + 571 + ], + "spans": [ + { + "bbox": [ + 264, + 547, + 481, + 571 + ], + "type": "interline_equation", + "content": "\\mathbf {t} _ {a g g} = \\sum_ {\\phi} \\left[ \\mathbf {t} _ {*} ^ {\\phi} \\right] / N _ {s}. \\tag {12}", + "image_path": "d838855b8703a47e9bd9d05e91132299e964305bd1c4ecff995c0df22bf3ddec.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 131, + 580, + 482, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 580, + 482, + 605 + ], + "spans": [ + { + "bbox": [ + 131, + 580, + 482, + 605 + ], + "type": "text", + "content": "We construct MLP-based classification and regression heads atop aggregated tokens for center frame bounding box prediction and atomic action recognition." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 622, + 236, + 632 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 622, + 236, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 622, + 236, + 632 + ], + "type": "text", + "content": "4.3 Loss Functions" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 131, + 641, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 641, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 131, + 641, + 482, + 666 + ], + "type": "text", + "content": "We use Binary Cross Entropy (BCE) loss and Mean Squared Error (MSE) loss for the multi-label supervision and the bounding box regression supervision in" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 250, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 250, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 250, + 102 + ], + "type": "text", + "content": "K. Peng, J. Fu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 472, + 129 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 472, + 129 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 472, + 129 + ], + "type": "text", + "content": "the same way as all the baselines. The equation of the BCE loss is as Eq. 13," + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 183, + 138, + 481, + 169 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 183, + 138, + 481, + 169 + ], + "spans": [ + { + "bbox": [ + 183, + 138, + 481, + 169 + ], + "type": "interline_equation", + "content": "L _ {B C E} (\\mathbf {y}, \\hat {\\mathbf {y}}) = - \\frac {1}{N _ {c}} \\sum_ {i = 1} ^ {N _ {c}} \\left[ \\mathbf {y} _ {i} \\log \\left(\\hat {\\mathbf {y}} _ {i}\\right) + \\left(1 - \\mathbf {y} _ {i}\\right) \\log \\left(1 - \\hat {\\mathbf {y}} _ {i}\\right) \\right], \\tag {13}", + "image_path": "357fb2dc7711a044028842c7cefa3d46bbbeeb60f42adf8eb7434af6584e305a.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 178, + 481, + 202 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 178, + 481, + 202 + ], + "spans": [ + { + "bbox": [ + 131, + 178, + 481, + 202 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 178, + 481, + 202 + ], + "type": "inline_equation", + "content": "\\mathbf{y}" + }, + { + "bbox": [ + 131, + 178, + 481, + 202 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 131, + 178, + 481, + 202 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{y}}" + }, + { + "bbox": [ + 131, + 178, + 481, + 202 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 178, + 481, + 202 + ], + "type": "inline_equation", + "content": "N_{c}" + }, + { + "bbox": [ + 131, + 178, + 481, + 202 + ], + "type": "text", + "content": " indicate the one-hot ground truth, the prediction, and the category number. The bounding boxes regression loss is expressed as Eq. 14," + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 242, + 212, + 481, + 244 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 242, + 212, + 481, + 244 + ], + "spans": [ + { + "bbox": [ + 242, + 212, + 481, + 244 + ], + "type": "interline_equation", + "content": "L _ {M S E} (\\mathbf {b}, \\hat {\\mathbf {b}}) = \\sum_ {j = 1} ^ {4} \\left(\\mathbf {b} _ {j} - \\hat {\\mathbf {b}} _ {j}\\right) ^ {2}, \\tag {14}", + "image_path": "13aff28203153e419a72f613d53f0524b5f6f3b6345aa4f1f5a86e380346909d.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 253, + 480, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 253, + 480, + 277 + ], + "spans": [ + { + "bbox": [ + 131, + 253, + 480, + 277 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 253, + 480, + 277 + ], + "type": "inline_equation", + "content": "\\mathbf{b}" + }, + { + "bbox": [ + 131, + 253, + 480, + 277 + ], + "type": "text", + "content": " indicates the coordinates of the left top and right bottom corners, and " + }, + { + "bbox": [ + 131, + 253, + 480, + 277 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{b}}" + }, + { + "bbox": [ + 131, + 253, + 480, + 277 + ], + "type": "text", + "content": " represents the predicted boxes. " + }, + { + "bbox": [ + 131, + 253, + 480, + 277 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 131, + 253, + 480, + 277 + ], + "type": "text", + "content": " indicates the coordinate index." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 296, + 230, + 309 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 296, + 230, + 309 + ], + "spans": [ + { + "bbox": [ + 132, + 296, + 230, + 309 + ], + "type": "text", + "content": "5 Experiments" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 320, + 279, + 332 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 320, + 279, + 332 + ], + "spans": [ + { + "bbox": [ + 132, + 320, + 279, + 332 + ], + "type": "text", + "content": "5.1 Implementation Details" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 340, + 482, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 340, + 482, + 449 + ], + "spans": [ + { + "bbox": [ + 131, + 340, + 482, + 449 + ], + "type": "text", + "content": "We conduct experiments on four NVIDIA A100 GPUs. We use BertAdam [32] optimizer with learning rate " + }, + { + "bbox": [ + 131, + 340, + 482, + 449 + ], + "type": "inline_equation", + "content": "lr = 1e^{-4}" + }, + { + "bbox": [ + 131, + 340, + 482, + 449 + ], + "type": "text", + "content": ", batch size 128, learning rate decay 0.9, and warmup ratio 0.1. The agent number and head number in our RefAtomNet are 4 and 1. Our model has " + }, + { + "bbox": [ + 131, + 340, + 482, + 449 + ], + "type": "inline_equation", + "content": "214M" + }, + { + "bbox": [ + 131, + 340, + 482, + 449 + ], + "type": "text", + "content": " trainable parameters. The weight for " + }, + { + "bbox": [ + 131, + 340, + 482, + 449 + ], + "type": "inline_equation", + "content": "L_{MSE}" + }, + { + "bbox": [ + 131, + 340, + 482, + 449 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 340, + 482, + 449 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 131, + 340, + 482, + 449 + ], + "type": "text", + "content": " are 5 and 0.125. We trained our model for 40 epochs on our dataset. The text encoder is frozen during training. We use multi-label mean Average Precision (mAP), Area Under the Receiver Operating Characteristic curve (AUROC), and the mean Intersection over Union (mIOU) as metrics, of which mIOU has less priority since we prioritize the atomic video action recognition task." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 466, + 270, + 478 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 466, + 270, + 478 + ], + "spans": [ + { + "bbox": [ + 132, + 466, + 270, + 478 + ], + "type": "text", + "content": "5.2 Experimental Results" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 131, + 486, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 486, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 131, + 486, + 482, + 666 + ], + "type": "text", + "content": "Main Results. Results of the main experiments on the proposed RefAVA benchmark are summarized in Tab. 2. We compare our RefAtomNet model with a multitude of published approaches adopted to suit our task (see Section 3.2) stemming from the fields of (1) Atomic Action Localization (AAL); (2) Video Question Answering (VQA); (3) Video-Text Retrieval (VTR); (4) Single Frame (SF) baselines; and (5) Video Object Segmentation (VOS). Baseline methods from the AAL group consistently underperform in spatially localizing the action according to the textual reference (measured as mIOU). Even though the ability of the human spatial localization is not the major expected output, it can still generally illustrate if the model can grasp the correct referring person or not. Among the AAL methods, X3D [16] achieves the highest AUROC score of " + }, + { + "bbox": [ + 131, + 486, + 482, + 666 + ], + "type": "inline_equation", + "content": "59.09\\%" + }, + { + "bbox": [ + 131, + 486, + 482, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 486, + 482, + 666 + ], + "type": "inline_equation", + "content": "64.51\\%" + }, + { + "bbox": [ + 131, + 486, + 482, + 666 + ], + "type": "text", + "content": " on the val and test sets, respectively. The VQA and VTR baselines work better than the AAL baselines since they can reasonably capture the referring person while delivering acceptable atomic human action recognition performances, which benefits from the text-aware pretraining of these two" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 413, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 413, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 413, + 91, + 447, + 100 + ], + "type": "text", + "content": "RAVAR" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 140, + 136, + 479, + 297 + ], + "blocks": [ + { + "bbox": [ + 186, + 114, + 427, + 125 + ], + "lines": [ + { + "bbox": [ + 186, + 114, + 427, + 125 + ], + "spans": [ + { + "bbox": [ + 186, + 114, + 427, + 125 + ], + "type": "text", + "content": "Table 2: Experimental results on our RAVAR benchmark." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 140, + 136, + 479, + 297 + ], + "lines": [ + { + "bbox": [ + 140, + 136, + 479, + 297 + ], + "spans": [ + { + "bbox": [ + 140, + 136, + 479, + 297 + ], + "type": "table", + "html": "
MethodmIOUmAPAUROCmIOUmAPAUROC
ValTest
AALI3D [4]0.0044.0457.770.0044.6462.71
X3D [16]0.2644.4559.090.2746.3464.51
MViTv2-B [47]0.7642.3256.430.6642.6059.30
VideoMAE2-B [81]0.1642.0255.120.2941.8759.10
Hiera-B [69]0.4942.7456.720.6241.1458.70
VQASingularity [38]26.3442.4359.5229.7842.1856.12
AskAnything13B [44]20.0951.4266.1222.3552.2569.35
VTRMeVTR [95]30.7838.4251.0129.7936.2752.45
CLIP4CLIP [55]34.7539.4852.5732.3337.1755.05
XCIIIP [56]35.5442.4656.3031.7940.8258.71
BLIPv2 [42]32.9952.1366.5632.7553.1969.92
SFCLIP [67]+SAM [33]32.9549.7464.5529.8051.3469.11
CLIP [67]+DETR [3]33.9647.5761.8333.8450.9267.29
CLIP [67]+REFCLIP [29]34.4347.7962.5032.2849.9067.44
VOSSu et al. [79]23.7152.1766.6726.0253.2070.19
RefAtomNet (Ours)38.2255.9869.7336.4257.5273.95
", + "image_path": "2c70ef408284b2411b5140944b7f7f25f927063aebeef20de835f87bc32707f7.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 148, + 335, + 292, + 345 + ], + "lines": [ + { + "bbox": [ + 148, + 335, + 292, + 345 + ], + "spans": [ + { + "bbox": [ + 148, + 335, + 292, + 345 + ], + "type": "text", + "content": "(a) Module ablation of the RefAtomNet." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "text" + }, + { + "type": "table", + "bbox": [ + 133, + 354, + 296, + 421 + ], + "blocks": [ + { + "bbox": [ + 133, + 311, + 478, + 322 + ], + "lines": [ + { + "bbox": [ + 133, + 311, + 478, + 322 + ], + "spans": [ + { + "bbox": [ + 133, + 311, + 478, + 322 + ], + "type": "text", + "content": "Table 3: Experiments for module ablation and comparison with multi-modal fusion." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 133, + 354, + 296, + 421 + ], + "lines": [ + { + "bbox": [ + 133, + 354, + 296, + 421 + ], + "spans": [ + { + "bbox": [ + 133, + 354, + 296, + 421 + ], + "type": "table", + "html": "
MethodmIOUmAPAUROCmIOUmAPAUROC
ValTest
w/o ALSAF27.3050.7065.3129.0951.2668.20
w/o CAAF36.2155.4369.6635.3855.9072.60
w/o CATF35.0153.8367.7134.5556.9673.40
w/o LSAS31.9055.2169.4731.2555.7372.38
Ours38.2255.9869.7336.4257.5273.95
", + "image_path": "a9d525c1085d1b3470b59d1fec739d6c0793de56cf35efc59c1e9e316db6ddb6.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 307, + 354, + 481, + 421 + ], + "blocks": [ + { + "bbox": [ + 313, + 335, + 476, + 346 + ], + "lines": [ + { + "bbox": [ + 313, + 335, + 476, + 346 + ], + "spans": [ + { + "bbox": [ + 313, + 335, + 476, + 346 + ], + "type": "text", + "content": "(b) Comparison with other fusion approaches." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 307, + 354, + 481, + 421 + ], + "lines": [ + { + "bbox": [ + 307, + 354, + 481, + 421 + ], + "spans": [ + { + "bbox": [ + 307, + 354, + 481, + 421 + ], + "type": "table", + "html": "
FusionmIOU mAP AUROCmIOU mAP AUROC
ValTest
Addition27.3050.7065.3129.0951.2668.20
Concatenation18.6452.2366.4520.6553.4470.70
Multiplication23.9051.5565.6625.0553.3370.48
AttentionBottleneck [59]33.4750.9765.0733.0254.0271.08
McOmut [98]23.8851.5865.6525.0253.2170.42
Ours38.2255.9869.7336.4257.5273.95
", + "image_path": "3128b2ae8e166bc982edeebd994c5b4be662deee189a1a86bfef357adad067e6.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 448, + 482, + 532 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 448, + 482, + 532 + ], + "spans": [ + { + "bbox": [ + 130, + 448, + 482, + 532 + ], + "type": "text", + "content": "tasks. AskAnything [44] from the VQA group achieves " + }, + { + "bbox": [ + 130, + 448, + 482, + 532 + ], + "type": "inline_equation", + "content": "20.09\\%" + }, + { + "bbox": [ + 130, + 448, + 482, + 532 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 448, + 482, + 532 + ], + "type": "inline_equation", + "content": "22.35\\%" + }, + { + "bbox": [ + 130, + 448, + 482, + 532 + ], + "type": "text", + "content": " of mIOU, " + }, + { + "bbox": [ + 130, + 448, + 482, + 532 + ], + "type": "inline_equation", + "content": "51.42\\%" + }, + { + "bbox": [ + 130, + 448, + 482, + 532 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 448, + 482, + 532 + ], + "type": "inline_equation", + "content": "52.25\\%" + }, + { + "bbox": [ + 130, + 448, + 482, + 532 + ], + "type": "text", + "content": " of mAP, and " + }, + { + "bbox": [ + 130, + 448, + 482, + 532 + ], + "type": "inline_equation", + "content": "66.12\\%" + }, + { + "bbox": [ + 130, + 448, + 482, + 532 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 448, + 482, + 532 + ], + "type": "inline_equation", + "content": "69.35\\%" + }, + { + "bbox": [ + 130, + 448, + 482, + 532 + ], + "type": "text", + "content": " of AUROC on the val and test sets. BLIPv2 [43] from the VTR group delivers " + }, + { + "bbox": [ + 130, + 448, + 482, + 532 + ], + "type": "inline_equation", + "content": "32.99\\%" + }, + { + "bbox": [ + 130, + 448, + 482, + 532 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 448, + 482, + 532 + ], + "type": "inline_equation", + "content": "32.75\\%" + }, + { + "bbox": [ + 130, + 448, + 482, + 532 + ], + "type": "text", + "content": " of mIOU, " + }, + { + "bbox": [ + 130, + 448, + 482, + 532 + ], + "type": "inline_equation", + "content": "52.13\\%" + }, + { + "bbox": [ + 130, + 448, + 482, + 532 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 448, + 482, + 532 + ], + "type": "inline_equation", + "content": "53.19\\%" + }, + { + "bbox": [ + 130, + 448, + 482, + 532 + ], + "type": "text", + "content": " of mAP, and " + }, + { + "bbox": [ + 130, + 448, + 482, + 532 + ], + "type": "inline_equation", + "content": "66.56\\%" + }, + { + "bbox": [ + 130, + 448, + 482, + 532 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 448, + 482, + 532 + ], + "type": "inline_equation", + "content": "69.92\\%" + }, + { + "bbox": [ + 130, + 448, + 482, + 532 + ], + "type": "text", + "content": " of AUROC, on the val and test sets, respectively. Our RefAtomNet achieves state-of-the-art performances, outperforming the best baseline BLIPv2 [42] by " + }, + { + "bbox": [ + 130, + 448, + 482, + 532 + ], + "type": "inline_equation", + "content": "5.23\\%" + }, + { + "bbox": [ + 130, + 448, + 482, + 532 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 448, + 482, + 532 + ], + "type": "inline_equation", + "content": "3.85\\%" + }, + { + "bbox": [ + 130, + 448, + 482, + 532 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 448, + 482, + 532 + ], + "type": "inline_equation", + "content": "3.17\\%" + }, + { + "bbox": [ + 130, + 448, + 482, + 532 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 448, + 482, + 532 + ], + "type": "inline_equation", + "content": "3.67\\%" + }, + { + "bbox": [ + 130, + 448, + 482, + 532 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 448, + 482, + 532 + ], + "type": "inline_equation", + "content": "4.33\\%" + }, + { + "bbox": [ + 130, + 448, + 482, + 532 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 448, + 482, + 532 + ], + "type": "inline_equation", + "content": "4.03\\%" + }, + { + "bbox": [ + 130, + 448, + 482, + 532 + ], + "type": "text", + "content": " of mIOU, mAP, and AUROC, on val and test sets." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": "Ablations of the Individual Modules. In Tab. 3a, we show the ablation for the components of RefAtomNet by removing each of the proposed designs, where LSAS indicates the location-semantic stream, CAAF indicates the cross-stream agent attention fusion, CATF indicates the cross-stream agent token fusion, and " + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "inline_equation", + "content": "w / o" + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": " ALSAF indicates by simply using addition to fuse these three streams and without the agent-based location-semantic aware attentional fusion mechanism. Compared with the ablation " + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "inline_equation", + "content": "w / o" + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": " ALSAF, RefAtomNet achieves promising improvements of " + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "inline_equation", + "content": "10.92\\%" + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "inline_equation", + "content": "5.28\\%" + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "inline_equation", + "content": "4.42\\%" + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "inline_equation", + "content": "7.33\\%" + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "inline_equation", + "content": "6.26\\%" + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "inline_equation", + "content": "5.75\\%" + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": " in terms of mIOU, mAP, and AUROC for the val and test sets, respectively. It indicates that using simple aggregation of the three streams makes the model distract from the important cues for the referring person, illustrated by the large decay of mIOU metric." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 250, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 250, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 250, + 101 + ], + "type": "text", + "content": "K. Peng, J. Fu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "bbox": [ + 143, + 138, + 297, + 148 + ], + "lines": [ + { + "bbox": [ + 143, + 138, + 297, + 148 + ], + "spans": [ + { + "bbox": [ + 143, + 138, + 297, + 148 + ], + "type": "text", + "content": "(a) Generalizability to test time rephrasing" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "type": "table", + "bbox": [ + 132, + 158, + 304, + 222 + ], + "blocks": [ + { + "bbox": [ + 143, + 114, + 469, + 125 + ], + "lines": [ + { + "bbox": [ + 143, + 114, + 469, + 125 + ], + "spans": [ + { + "bbox": [ + 143, + 114, + 469, + 125 + ], + "type": "text", + "content": "Table 4: Generalizability to different referring styles and encoder architectures." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 132, + 158, + 304, + 222 + ], + "lines": [ + { + "bbox": [ + 132, + 158, + 304, + 222 + ], + "spans": [ + { + "bbox": [ + 132, + 158, + 304, + 222 + ], + "type": "table", + "html": "
FusionmIOUmAPAUROCmIOUmAPAUROC
ValTest
Singularity [38]18.4541.2758.4720.5441.3955.32
XCLIP [56]31.9541.3554.3529.8440.7458.45
AskAnything [44]19.7451.1165.8321.6051.9669.04
BLIPv2 [42]31.0051.4565.8831.3452.3568.87
Ours34.6555.7569.5233.1457.2373.76
", + "image_path": "20e42e2cb85ff0303bc5f3bddd69fdb6a22cf36e2d20b100b29cdabc52cede9d.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 307, + 167, + 480, + 221 + ], + "blocks": [ + { + "bbox": [ + 307, + 140, + 482, + 156 + ], + "lines": [ + { + "bbox": [ + 307, + 140, + 482, + 156 + ], + "spans": [ + { + "bbox": [ + 307, + 140, + 482, + 156 + ], + "type": "text", + "content": "(b) Generalizability to different visual-textual encoder architectures." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 307, + 167, + 480, + 221 + ], + "lines": [ + { + "bbox": [ + 307, + 167, + 480, + 221 + ], + "spans": [ + { + "bbox": [ + 307, + 167, + 480, + 221 + ], + "type": "table", + "html": "
FusionmIOU mAP AUROCmIOU mAP AUROC
ValTest
XCLIP [56]35.5442.4656.3031.7940.8258.71
RefAtomNet (XCLIP)38.5947.4061.2036.6148.5966.47
BLIPv2 [42]32.9952.1366.5632.7553.1969.92
RefAtomNet (BLIPv2)38.2255.9869.7336.4257.5273.95
", + "image_path": "669b5eca744268dc88dc0c7ec4c8c539bf6f7a9ed22de31c47786878ebf09763.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 246, + 479, + 317 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 246, + 479, + 317 + ], + "spans": [ + { + "bbox": [ + 130, + 246, + 479, + 317 + ], + "type": "text", + "content": "Compared with the ablation " + }, + { + "bbox": [ + 130, + 246, + 479, + 317 + ], + "type": "inline_equation", + "content": "w / o" + }, + { + "bbox": [ + 130, + 246, + 479, + 317 + ], + "type": "text", + "content": " LSAS, we find that the location-semantic tokens benefit more for the localization of the correct person in the center frame. The advantages of the LSAS on atomic video action recognition metrics are highlighted in the test set which has higher scenario diversity compared with the val set. Using CAAF and CATF, each of them brings promising benefits by suppressing the irrelevant information in the visual tokens." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 318, + 480, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 318, + 480, + 376 + ], + "spans": [ + { + "bbox": [ + 130, + 318, + 480, + 376 + ], + "type": "text", + "content": "Comparison with Other Fusion Mechanisms. We compare our proposed agent-based location-semantic aware attentional fusion mechanism with late fusion addition, multiplication, concatenation, AttentionBottleNeck [59], and McOmet [98] in Tab. 3b. Our approach outperforms all by suppressing those irrelevant visual tokens on the agent tokens and attention masks." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 378, + 481, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 378, + 481, + 509 + ], + "spans": [ + { + "bbox": [ + 130, + 378, + 481, + 509 + ], + "type": "text", + "content": "Generalizability to Test-time Reference Rephrasing. The reference sentences given by different users may differ in daily life scenarios. To test the generalizability of the model towards different referring styles, we invoke API (gpt-3.5-turbo) of ChatGPT [60] to rephrase the test set description two times and then deliver the averaged performance for RAVAR on the original val and test sets, and the two rephrased val and test sets, where we select 4 most out-performing baselines from our benchmark to construct this ablation study, i.e., Singularity [38], AskAnything [44], XCLIP [56], and BLIPv2 [42], as shown in Tab. 4a. RefAtomNet delivers the best performances. We find out that the test time reference rephrasing will cause performance decay for localization and less decay for the atomic video-based action recognition delivered by our model." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 510, + 481, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 481, + 617 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 481, + 617 + ], + "type": "text", + "content": "Generalizability to Different Visual Textual Backbones. Since we use the best-performing baseline BLIPv2 [42] as the visual and textual encoder in our model, it would be interesting to see if the proposed architecture can generalize to different encoder backbones. We thereby demonstrate another ablation study in Tab. 4b, where we equip our RefAtomNet with the XCLIP [56] backbone. Compared with the XCLIP baseline, the RefAtomNet (XCLIP) achieves " + }, + { + "bbox": [ + 130, + 510, + 481, + 617 + ], + "type": "inline_equation", + "content": "3.05\\%" + }, + { + "bbox": [ + 130, + 510, + 481, + 617 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 510, + 481, + 617 + ], + "type": "inline_equation", + "content": "4.94\\%" + }, + { + "bbox": [ + 130, + 510, + 481, + 617 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 510, + 481, + 617 + ], + "type": "inline_equation", + "content": "4.90\\%" + }, + { + "bbox": [ + 130, + 510, + 481, + 617 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 510, + 481, + 617 + ], + "type": "inline_equation", + "content": "4.82\\%" + }, + { + "bbox": [ + 130, + 510, + 481, + 617 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 510, + 481, + 617 + ], + "type": "inline_equation", + "content": "7.77\\%" + }, + { + "bbox": [ + 130, + 510, + 481, + 617 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 510, + 481, + 617 + ], + "type": "inline_equation", + "content": "7.76\\%" + }, + { + "bbox": [ + 130, + 510, + 481, + 617 + ], + "type": "text", + "content": " performance improvements in terms of mIOU, mAP, and AUROC on the val and test sets, respectively, showing the great generalizability of our RefAtomNet of the visual and textual encoder." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 617, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 481, + 665 + ], + "type": "text", + "content": "Analysis of the Qualitative Results. The qualitative results are delivered in Fig. 4. These examples demonstrate the effectiveness of RefAtomNet compared with the best-performing baseline BLIPv2 [42]. Samples 1 and 8 validate the recognition quality for references without any location indications." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 413, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 413, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 413, + 91, + 447, + 100 + ], + "type": "text", + "content": "RAVAR" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "bbox": [ + 136, + 115, + 272, + 123 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 115, + 272, + 123 + ], + "spans": [ + { + "bbox": [ + 136, + 115, + 272, + 123 + ], + "type": "text", + "content": "1 Text reference: The man wearing light green clothes." + } + ] + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 137, + 123, + 305, + 180 + ], + "blocks": [ + { + "bbox": [ + 137, + 123, + 305, + 180 + ], + "lines": [ + { + "bbox": [ + 137, + 123, + 305, + 180 + ], + "spans": [ + { + "bbox": [ + 137, + 123, + 305, + 180 + ], + "type": "image", + "image_path": "5509deddc417e1eb431442e202ffcd0047277d50bc829982fa456a8dd455d53f.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 307, + 115, + 458, + 123 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 115, + 458, + 123 + ], + "spans": [ + { + "bbox": [ + 307, + 115, + 458, + 123 + ], + "type": "text", + "content": "2 Text reference: The man in the center who has gray beard." + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 308, + 123, + 477, + 180 + ], + "blocks": [ + { + "bbox": [ + 308, + 123, + 477, + 180 + ], + "lines": [ + { + "bbox": [ + 308, + 123, + 477, + 180 + ], + "spans": [ + { + "bbox": [ + 308, + 123, + 477, + 180 + ], + "type": "image", + "image_path": "3d93bf92296f2b83a4e51ab679f12d33af7670d35defd592d8e7224ea981eb67.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 136, + 180, + 286, + 188 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 180, + 286, + 188 + ], + "spans": [ + { + "bbox": [ + 136, + 180, + 286, + 188 + ], + "type": "text", + "content": "3 Text reference: The man on the left who is in blue clothes." + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 137, + 188, + 305, + 246 + ], + "blocks": [ + { + "bbox": [ + 137, + 188, + 305, + 246 + ], + "lines": [ + { + "bbox": [ + 137, + 188, + 305, + 246 + ], + "spans": [ + { + "bbox": [ + 137, + 188, + 305, + 246 + ], + "type": "image", + "image_path": "207de4bb67adb1aa77ee3354b0104f7ccbcd8366ed9c26cdc90b3f59403846e4.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 307, + 181, + 417, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 181, + 417, + 189 + ], + "spans": [ + { + "bbox": [ + 307, + 181, + 417, + 189 + ], + "type": "text", + "content": "4 Text reference: The first man on the right." + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 308, + 189, + 477, + 246 + ], + "blocks": [ + { + "bbox": [ + 308, + 189, + 477, + 246 + ], + "lines": [ + { + "bbox": [ + 308, + 189, + 477, + 246 + ], + "spans": [ + { + "bbox": [ + 308, + 189, + 477, + 246 + ], + "type": "image", + "image_path": "6e9a141ec02e5100c1693d0b670e3ca15aa7134091578f1895fab80c8ca357fc.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 137, + 246, + 272, + 254 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 246, + 272, + 254 + ], + "spans": [ + { + "bbox": [ + 137, + 246, + 272, + 254 + ], + "type": "text", + "content": "5 Text reference: The man wearing glasses on the left." + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 137, + 254, + 305, + 311 + ], + "blocks": [ + { + "bbox": [ + 137, + 254, + 305, + 311 + ], + "lines": [ + { + "bbox": [ + 137, + 254, + 305, + 311 + ], + "spans": [ + { + "bbox": [ + 137, + 254, + 305, + 311 + ], + "type": "image", + "image_path": "77ae1422bb388a1c7b8712011e9962678e3db58fd7a8ff2050470a04794e7f43.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 308, + 246, + 433, + 254 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 246, + 433, + 254 + ], + "spans": [ + { + "bbox": [ + 308, + 246, + 433, + 254 + ], + "type": "text", + "content": "6 Text reference: The man on the right back to us." + } + ] + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 308, + 254, + 477, + 311 + ], + "blocks": [ + { + "bbox": [ + 308, + 254, + 477, + 311 + ], + "lines": [ + { + "bbox": [ + 308, + 254, + 477, + 311 + ], + "spans": [ + { + "bbox": [ + 308, + 254, + 477, + 311 + ], + "type": "image", + "image_path": "48870fa17142608cb2eaf39c474afb14562293ea15a75a2991bc83d7fff36449.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "bbox": [ + 137, + 311, + 293, + 326 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 311, + 293, + 326 + ], + "spans": [ + { + "bbox": [ + 137, + 311, + 293, + 326 + ], + "type": "text", + "content": "Text reference: The woman wearing black clothes and pearls necklace on the right." + } + ] + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 137, + 326, + 305, + 390 + ], + "blocks": [ + { + "bbox": [ + 137, + 326, + 305, + 390 + ], + "lines": [ + { + "bbox": [ + 137, + 326, + 305, + 390 + ], + "spans": [ + { + "bbox": [ + 137, + 326, + 305, + 390 + ], + "type": "image", + "image_path": "1d1a36e5d72917f6a1a8a4ff9579bda5481aea59dc6f4168360dc80ca8c4f54b.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 401, + 481, + 424 + ], + "lines": [ + { + "bbox": [ + 131, + 401, + 481, + 424 + ], + "spans": [ + { + "bbox": [ + 131, + 401, + 481, + 424 + ], + "type": "text", + "content": "Fig. 4: An overview of qualitative results. Missed predictions are marked with a red cross, while true positive and false positive predictions are shown in green and red." + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 311, + 425, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 311, + 425, + 319 + ], + "spans": [ + { + "bbox": [ + 307, + 311, + 425, + 319 + ], + "type": "text", + "content": "8 Text reference: The man wearing black suits." + } + ] + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 308, + 326, + 477, + 390 + ], + "blocks": [ + { + "bbox": [ + 308, + 326, + 477, + 390 + ], + "lines": [ + { + "bbox": [ + 308, + 326, + 477, + 390 + ], + "spans": [ + { + "bbox": [ + 308, + 326, + 477, + 390 + ], + "type": "image", + "image_path": "28215095954289f8843092093113a56a18904c32160aaa8f6acef65821b6f392.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "bbox": [ + 130, + 445, + 482, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 445, + 482, + 518 + ], + "spans": [ + { + "bbox": [ + 130, + 445, + 482, + 518 + ], + "type": "text", + "content": "Samples 3, 4, 5, and 6 showcase the results for different humans in a shared scene. Samples 2 and 7 show the RAVAR performance for references containing finer details, e.g., the necklace type and the beard color. Our RefAtomNet outperforms the BLIPv2 baseline on selected samples obviously, attributable to its superior capability in eliminating extraneous visual cues and its advanced location-semantic reasoning prowess." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 132, + 534, + 226, + 547 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 534, + 226, + 547 + ], + "spans": [ + { + "bbox": [ + 132, + 534, + 226, + 547 + ], + "type": "text", + "content": "6 Conclusions" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "text", + "content": "In this work, we introduce a novel task called Referring Atomic Video Action Recognition (RAVAR). We establish the RefAVA dataset to address the challenge of identifying atomic actions for individuals of interest in videos based on textual descriptions. Existing methods exhibit poor performance on this new task, prompting the development of RefAtomNet, a vision-language architecture that effectively integrates cross-stream tokens for precise referring atomic video action recognition. Through reference-relevant token enhancement and comprehensive token integration, RefAtomNet achieves impressive results, highlighting its effectiveness in tackling the complexities of the RAVAR task." + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 250, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 250, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 250, + 102 + ], + "type": "text", + "content": "K. Peng, J. Fu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 140, + 482, + 296 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 140, + 482, + 296 + ], + "spans": [ + { + "bbox": [ + 130, + 140, + 482, + 296 + ], + "type": "text", + "content": "The project served to prepare the SFB 1574 Circular Factory for the Perpetual Product (project ID: 471687386), approved by the German Research Foundation (DFG, German Research Foundation) with a start date of April 1, 2024. This work was also partially supported in part by the SmartAge project sponsored by the Carl Zeiss Stiftung (P2019-01-003; 2021-2026). This work was performed on the HoreKa supercomputer funded by the Ministry of Science, Research and the Arts Baden-Württemberg and by the Federal Ministry of Education and Research. The authors also acknowledge support by the state of Baden-Württemberg through bwHPC and the German Research Foundation (DFG) through grant INST 35/1597-1 FUGG. This project is also supported by the National Key RD Program under Grant 2022YFB4701400. Lastly, the authors thank for the support of Dr. Sepideh Pashami, the Swedish Innovation Agency VINNOVA, the Digital Futures." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 313, + 197, + 325 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 313, + 197, + 325 + ], + "spans": [ + { + "bbox": [ + 132, + 313, + 197, + 325 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 337, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 138, + 337, + 480, + 359 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 337, + 480, + 359 + ], + "spans": [ + { + "bbox": [ + 138, + 337, + 480, + 359 + ], + "type": "text", + "content": "1. Bagad, P., Tapaswi, M., Snoek, C.G.M.: Test of time: Instilling video-language models with a sense of time. In: CVPR (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 360, + 480, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 360, + 480, + 381 + ], + "spans": [ + { + "bbox": [ + 138, + 360, + 480, + 381 + ], + "type": "text", + "content": "2. Bu, Y., Li, L., Xie, J., Liu, Q., Cai, Y., Huang, Q., Li, Q.: Scene-text oriented referring expression comprehension. TMM (2022)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 382, + 480, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 382, + 480, + 403 + ], + "spans": [ + { + "bbox": [ + 138, + 382, + 480, + 403 + ], + "type": "text", + "content": "3. Carion, N., Massa, F., Synnaeve, G., Usunier, N., Kirillov, A., Zagoruyko, S.: End-to-end object detection with transformers. In: ECCV (2020)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 403, + 480, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 403, + 480, + 425 + ], + "spans": [ + { + "bbox": [ + 138, + 403, + 480, + 425 + ], + "type": "text", + "content": "4. Carreira, J., Zisserman, A.: Quo vadis, action recognition? A new model and the kinetics dataset. In: CVPR (2017)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 425, + 480, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 425, + 480, + 446 + ], + "spans": [ + { + "bbox": [ + 138, + 425, + 480, + 446 + ], + "type": "text", + "content": "5. Castro, S., Deng, N., Huang, P., Burzo, M., Mihalcea, R.: In-the-wild video question answering. In: COLING (2022)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 447, + 480, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 447, + 480, + 468 + ], + "spans": [ + { + "bbox": [ + 138, + 447, + 480, + 468 + ], + "type": "text", + "content": "6. Chai, W., Guo, X., Wang, G., Lu, Y.: StableVideo: Text-driven consistency-aware diffusion video editing. In: ICCV (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 468, + 480, + 501 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 468, + 480, + 501 + ], + "spans": [ + { + "bbox": [ + 138, + 468, + 480, + 501 + ], + "type": "text", + "content": "7. Chen, J., Zhu, D., Haydarov, K., Li, X., Elhoseiny, M.: Video ChatCaptioner: Towards enriched spatiotemporal descriptions. arXiv preprint arXiv:2304.04227 (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 501, + 480, + 534 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 501, + 480, + 534 + ], + "spans": [ + { + "bbox": [ + 138, + 501, + 480, + 534 + ], + "type": "text", + "content": "8. Chen, X., Fang, H., Lin, T.Y., Vedantam, R., Gupta, S., Dolkar, P., Zitnick, C.L.: Microsoft COCO captions: Data collection and evaluation server. arXiv preprint arXiv:1504.00325 (2015)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 534, + 480, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 534, + 480, + 567 + ], + "spans": [ + { + "bbox": [ + 138, + 534, + 480, + 567 + ], + "type": "text", + "content": "9. Chen, Y., Wang, J., Lin, L., Qi, Z., Ma, J., Shan, Y.: Tagging before alignment: Integrating multi-modal tags for video-text retrieval. arXiv preprint arXiv:2301.12644 (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 567, + 480, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 567, + 480, + 589 + ], + "spans": [ + { + "bbox": [ + 138, + 567, + 480, + 589 + ], + "type": "text", + "content": "0. Chen, Z., Ma, L., Luo, W., Wong, K.Y.K.: Weakly-supervised spatio-temporally grounding natural sentence in video. arXiv preprint arXiv:1906.02549 (2019)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 589, + 480, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 589, + 480, + 610 + ], + "spans": [ + { + "bbox": [ + 138, + 589, + 480, + 610 + ], + "type": "text", + "content": "1. Chung, J., Wu, C.h., Yang, H.r., Tai, Y.W., Tang, C.K.: HAA500: Human-centric atomic action dataset with curated videos. In: ICCV (2021)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 610, + 480, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 610, + 480, + 643 + ], + "spans": [ + { + "bbox": [ + 138, + 610, + 480, + 643 + ], + "type": "text", + "content": "2. Dang, R., Feng, J., Zhang, H., Ge, C., Song, L., Gong, L., Liu, C., Chen, Q., Zhu, F., Zhao, R., Song, Y.: InstructDET: Diversifying referring object detection with generalized instructions. arXiv preprint arXiv:2310.05136 (2023)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 643, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 643, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 643, + 480, + 665 + ], + "type": "text", + "content": "3. Deruyttere, T., Vandenhende, S., Grujicic, D., Van Gool, L., Moens, M.F.: Talk2Car: Taking control of your self-driving car. In: EMNLP (2019)" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 413, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 413, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 413, + 91, + 447, + 100 + ], + "type": "text", + "content": "RAVAR" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 23, + "blocks": [ + { + "bbox": [ + 133, + 116, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 480, + 138 + ], + "type": "text", + "content": "14. Devlin, J., Chang, M.W., Lee, K., Toutanova, K.: BERT: Pre-training of deep bidirectional transformers for language understanding. In: ACL (2019)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 139, + 480, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 139, + 480, + 182 + ], + "spans": [ + { + "bbox": [ + 133, + 139, + 480, + 182 + ], + "type": "text", + "content": "15. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., Uszkoreit, J., Houlsby, N.: An image is worth 16x16 words: Transformers for image recognition at scale. In: ICLR (2021)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 182, + 480, + 202 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 182, + 480, + 202 + ], + "spans": [ + { + "bbox": [ + 133, + 182, + 480, + 202 + ], + "type": "text", + "content": "16. Feichtenhofer, C.: X3D: Expanding architectures for efficient video recognition. In: CVPR (2020)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 203, + 480, + 224 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 203, + 480, + 224 + ], + "spans": [ + { + "bbox": [ + 132, + 203, + 480, + 224 + ], + "type": "text", + "content": "17. Feichtenhofer, C., Fan, H., Malik, J., He, K.: SlowFast networks for video recognition. In: ICCV (2019)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 225, + 480, + 257 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 225, + 480, + 257 + ], + "spans": [ + { + "bbox": [ + 132, + 225, + 480, + 257 + ], + "type": "text", + "content": "18. Gandhi, M., Gul, M.O., Prakash, E., Grunde-McLaughlin, M., Krishna, R., Agrawala, M.: Measuring compositional consistency for video question answering. In: CVPR (2022)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 257, + 480, + 289 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 257, + 480, + 289 + ], + "spans": [ + { + "bbox": [ + 132, + 257, + 480, + 289 + ], + "type": "text", + "content": "19. Gao, D., Zhou, L., Ji, L., Zhu, L., Yang, Y., Shou, M.Z.: MIST: Multi-modal iterative spatial-temporal transformer for long-form video question answering. In: CVPR (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 289, + 480, + 311 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 289, + 480, + 311 + ], + "spans": [ + { + "bbox": [ + 132, + 289, + 480, + 311 + ], + "type": "text", + "content": "20. Garcia, N., Otani, M., Chu, C., Nakashima, Y.: KnowIT VQA: Answering knowledge-based questions about videos. In: AAAI (2020)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 311, + 480, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 311, + 480, + 331 + ], + "spans": [ + { + "bbox": [ + 132, + 311, + 480, + 331 + ], + "type": "text", + "content": "21. Gavrilyuk, K., Ghodrati, A., Li, Z., Snoek, C.G.: Actor and action video segmentation from a sentence. In: CVPR (2018)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 332, + 480, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 332, + 480, + 376 + ], + "spans": [ + { + "bbox": [ + 132, + 332, + 480, + 376 + ], + "type": "text", + "content": "22. Goyal, R., Kahou, S.E., Michalski, V., Materzynska, J., Westphal, S., Kim, H., Haenel, V., Fründ, I., Yianilos, P., Mueller-Freitag, M., Hoppe, F., Thurau, C., Bax, I., Memisevic, R.: The \"something something\" video database for learning and evaluating visual common sense. In: ICCV (2017)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 376, + 480, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 376, + 480, + 407 + ], + "spans": [ + { + "bbox": [ + 132, + 376, + 480, + 407 + ], + "type": "text", + "content": "23. Gritsenko, A., Xiong, X., Djolonga, J., Dehghani, M., Sun, C., Lucic, M., Schmid, C., Arnab, A.: End-to-end spatio-temporal action localisation with video transformers. arXiv preprint arXiv:2304.12160 (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 408, + 480, + 440 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 408, + 480, + 440 + ], + "spans": [ + { + "bbox": [ + 132, + 408, + 480, + 440 + ], + "type": "text", + "content": "24. Gu, C., Sun, C., Ross, D.A., Vondrick, C., Pantofaru, C., Li, Y., Vijayanarasimhan, S., Toderici, G., Ricco, S., Sukthankar, R., Schmid, C., Malik, J.: AVA: A video dataset of spatio-temporally localized atomic visual actions. In: CVPR (2018)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 441, + 480, + 461 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 441, + 480, + 461 + ], + "spans": [ + { + "bbox": [ + 132, + 441, + 480, + 461 + ], + "type": "text", + "content": "25. Guo, W., Zhang, Y., Yang, J., Yuan, X.: Re-attention for visual question answering. TIP (2021)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 462, + 480, + 483 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 462, + 480, + 483 + ], + "spans": [ + { + "bbox": [ + 132, + 462, + 480, + 483 + ], + "type": "text", + "content": "26. Han, D., Ye, T., Han, Y., Xia, Z., Song, S., Huang, G.: Agent attention: On the integration of softmax and linear attention. arXiv preprint arXiv:2312.08874 (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 483, + 480, + 504 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 483, + 480, + 504 + ], + "spans": [ + { + "bbox": [ + 132, + 483, + 480, + 504 + ], + "type": "text", + "content": "27. Ji, Y., Zhan, Y., Yang, Y., Xu, X., Shen, F., Shen, H.T.: A context knowledge map guided coarse-to-fine action recognition. TIP (2020)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 504, + 480, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 504, + 480, + 536 + ], + "spans": [ + { + "bbox": [ + 132, + 504, + 480, + 536 + ], + "type": "text", + "content": "28. Jiang, J., Chen, Z., Lin, H., Zhao, X., Gao, Y.: Divide and conquer: Question-guided spatio-temporal contextual attention for video question answering. In: AAAI (2020)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 536, + 480, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 536, + 480, + 557 + ], + "spans": [ + { + "bbox": [ + 132, + 536, + 480, + 557 + ], + "type": "text", + "content": "29. Jin, L., Luo, G., Zhou, Y., Sun, X., Jiang, G., Shu, A., Ji, R.: Refclip: A universal teacher for weakly supervised referring expression comprehension. In: CVPR (2023)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 558, + 480, + 579 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 558, + 480, + 579 + ], + "spans": [ + { + "bbox": [ + 132, + 558, + 480, + 579 + ], + "type": "text", + "content": "30. Khoreva, A., Rohrbach, A., Schiele, B.: Video object segmentation with language referring expressions. In: ACCV (2019)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 580, + 480, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 580, + 480, + 601 + ], + "spans": [ + { + "bbox": [ + 132, + 580, + 480, + 601 + ], + "type": "text", + "content": "31. Kim, M., Spinola, F., Benz, P., Kim, T.h.: A*: Atrous spatial temporal action recognition for real time applications. In: WACV (2024)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 132, + 601, + 480, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 601, + 480, + 622 + ], + "spans": [ + { + "bbox": [ + 132, + 601, + 480, + 622 + ], + "type": "text", + "content": "32. Kingma, D.P., Ba, J.: Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980 (2014)" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 132, + 622, + 480, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 622, + 480, + 643 + ], + "spans": [ + { + "bbox": [ + 132, + 622, + 480, + 643 + ], + "type": "text", + "content": "33. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., et al.: Segment anything. In: CVPR (2023)" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 132, + 644, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 644, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 644, + 480, + 665 + ], + "type": "text", + "content": "34. Kuehne, H., Jhuang, H., Garrote, E., Poggio, T., Serre, T.: HMDB: A large video database for human motion recognition. In: ICCV (2011)" + } + ] + } + ], + "index": 22 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 250, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 250, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 250, + 101 + ], + "type": "text", + "content": "K. Peng, J. Fu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 24, + "blocks": [ + { + "bbox": [ + 130, + 116, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 480, + 138 + ], + "type": "text", + "content": "35. Laput, G., Harrison, C.: Sensing fine-grained hand activity with smartwatches. In: CHI (2019)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 138, + 480, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 138, + 480, + 160 + ], + "spans": [ + { + "bbox": [ + 130, + 138, + 480, + 160 + ], + "type": "text", + "content": "36. Le, T.M., Le, V., Venkatesh, S., Tran, T.: Hierarchical conditional relation networks for video question answering. In: CVPR (2020)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 160, + 480, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 160, + 480, + 182 + ], + "spans": [ + { + "bbox": [ + 132, + 160, + 480, + 182 + ], + "type": "text", + "content": "37. Lea, C., Vidal, R., Hager, G.D.: Learning convolutional action primitives for fine-grained action recognition. In: ICRA (2016)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 182, + 480, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 182, + 480, + 205 + ], + "spans": [ + { + "bbox": [ + 132, + 182, + 480, + 205 + ], + "type": "text", + "content": "38. Lei, J., Berg, T.L., Bansal, M.: Revealing single frame bias for video-and-language learning. arXiv preprint arXiv:2206.03428 (2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 205, + 480, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 205, + 480, + 226 + ], + "spans": [ + { + "bbox": [ + 132, + 205, + 480, + 226 + ], + "type": "text", + "content": "39. Lei, J., Li, L., Zhou, L., Gan, Z., Berg, T.L., Bansal, M., Liu, J.: Less is more: ClipBERT for video-and-language learning via sparse sampling. In: CVPR (2021)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 226, + 480, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 226, + 480, + 248 + ], + "spans": [ + { + "bbox": [ + 132, + 226, + 480, + 248 + ], + "type": "text", + "content": "40. Li, G., Wei, Y., Tian, Y., Xu, C., Wen, J.R., Hu, D.: Learning to answer questions in dynamic audio-visual scenarios. In: CVPR (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 248, + 480, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 248, + 480, + 270 + ], + "spans": [ + { + "bbox": [ + 132, + 248, + 480, + 270 + ], + "type": "text", + "content": "41. Li, J., Niu, L., Zhang, L.: From representation to reasoning: Towards both evidence and commonsense reasoning for video question-answering. In: CVPR (2022)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 270, + 480, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 270, + 480, + 293 + ], + "spans": [ + { + "bbox": [ + 132, + 270, + 480, + 293 + ], + "type": "text", + "content": "42. Li, J., Li, D., Savarese, S., Hoi, S.: BLIP-2: Bootstrapping language-image pretraining with frozen image encoders and large language models. In: ICML (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 293, + 480, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 293, + 480, + 314 + ], + "spans": [ + { + "bbox": [ + 132, + 293, + 480, + 314 + ], + "type": "text", + "content": "43. Li, J., Li, D., Xiong, C., Hoi, S.: BLIP: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In: ICML (2022)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 314, + 480, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 314, + 480, + 346 + ], + "spans": [ + { + "bbox": [ + 132, + 314, + 480, + 346 + ], + "type": "text", + "content": "44. Li, K., He, Y., Wang, Y., Li, Y., Wang, W., Luo, P., Wang, Y., Wang, L., Qiao, Y.: VideoChat: Chat-centric video understanding. arXiv preprint arXiv:2305.06355 (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 346, + 480, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 346, + 480, + 369 + ], + "spans": [ + { + "bbox": [ + 132, + 346, + 480, + 369 + ], + "type": "text", + "content": "45. Li, L., Gan, Z., Lin, K., Lin, C.C., Liu, Z., Liu, C., Wang, L.: LAVENDER: Unifying video-language understanding as masked language modeling. In: CVPR (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 369, + 480, + 391 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 369, + 480, + 391 + ], + "spans": [ + { + "bbox": [ + 132, + 369, + 480, + 391 + ], + "type": "text", + "content": "46. Li, R., Li, K., Kuo, Y.C., Shu, M., Qi, X., Shen, X., Jia, J.: Referring image segmentation via recurrent refinement networks. In: CVPR (2018)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 391, + 480, + 423 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 391, + 480, + 423 + ], + "spans": [ + { + "bbox": [ + 132, + 391, + 480, + 423 + ], + "type": "text", + "content": "47. Li, Y., Wu, C.Y., Fan, H., Mangalam, K., Xiong, B., Malik, J., Feichtenhofer, C.: MViTv2: Improved multiscale vision transformers for classification and detection. In: CVPR (2022)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 423, + 480, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 423, + 480, + 456 + ], + "spans": [ + { + "bbox": [ + 132, + 423, + 480, + 456 + ], + "type": "text", + "content": "48. Lin, J., Chen, J., Peng, K., He, X., Li, Z., Stiefelhagen, R., Yang, K.: EchoTrack: Auditory referring multi-object tracking for autonomous driving. arXiv preprint arXiv:2402.18302 (2024)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 456, + 480, + 489 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 456, + 480, + 489 + ], + "spans": [ + { + "bbox": [ + 132, + 456, + 480, + 489 + ], + "type": "text", + "content": "49. Lin, X., Tiwari, S., Huang, S., Li, M., Shou, M.Z., Ji, H., Chang, S.F.: Towards fast adaptation of pretrained contrastive models for multi-channel video-language retrieval. In: CVPR (2023)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 490, + 480, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 490, + 480, + 511 + ], + "spans": [ + { + "bbox": [ + 132, + 490, + 480, + 511 + ], + "type": "text", + "content": "50. Liu, J., Wang, L., Yang, M.H.: Referring expression generation and comprehension via attributes. In: ICCV (2017)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 511, + 480, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 511, + 480, + 544 + ], + "spans": [ + { + "bbox": [ + 132, + 511, + 480, + 544 + ], + "type": "text", + "content": "51. Liu, R., Zhang, J., Peng, K., Zheng, J., Cao, K., Chen, Y., Yang, K., Stiefelhagen, R.: Open scene understanding: Grounded situation recognition meets segment anything for helping people with visual impairments. In: ICCVW (2023)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 544, + 480, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 544, + 480, + 567 + ], + "spans": [ + { + "bbox": [ + 132, + 544, + 480, + 567 + ], + "type": "text", + "content": "52. Liu, R., Liu, C., Bai, Y., Yuille, A.L.: CLEVR-Ref+: Diagnosing visual reasoning with referring expressions. In: CVPR (2019)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 132, + 567, + 480, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 567, + 480, + 588 + ], + "spans": [ + { + "bbox": [ + 132, + 567, + 480, + 588 + ], + "type": "text", + "content": "53. Liu, S., Hui, T., Huang, S., Wei, Y., Li, B., Li, G.: Cross-modal progressive comprehension for referring segmentation. TPAMI (2021)" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 132, + 588, + 480, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 588, + 480, + 610 + ], + "spans": [ + { + "bbox": [ + 132, + 588, + 480, + 610 + ], + "type": "text", + "content": "54. Liu, Y., Li, G., Lin, L.: Cross-modal causal relational reasoning for event-level visual question answering. TPAMI (2023)" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 132, + 610, + 480, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 610, + 480, + 643 + ], + "spans": [ + { + "bbox": [ + 132, + 610, + 480, + 643 + ], + "type": "text", + "content": "55. Luo, H., Ji, L., Zhong, M., Chen, Y., Lei, W., Duan, N., Li, T.: CLIP4Clip: An empirical study of CLIP for end to end video clip retrieval and captioning. Neurocomputing (2022)" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 132, + 643, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 643, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 643, + 480, + 665 + ], + "type": "text", + "content": "56. Ma, Y., Xu, G., Sun, X., Yan, M., Zhang, J., Ji, R.: X-CLIP: End-to-end multigrained contrastive learning for video-text retrieval. In: MM (2022)" + } + ] + } + ], + "index": 23 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 413, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 413, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 413, + 91, + 447, + 100 + ], + "type": "text", + "content": "RAVAR" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 24, + "blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 480, + 138 + ], + "type": "text", + "content": "57. Madasu, A., Aflalo, E., Ben Melech Stan, G., Tseng, S.Y., Bertasius, G., Lal, V.: Improving video retrieval using multilingual knowledge transfer. In: ECIR (2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 140, + 480, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 140, + 480, + 160 + ], + "spans": [ + { + "bbox": [ + 132, + 140, + 480, + 160 + ], + "type": "text", + "content": "58. McIntosh, B., Duarte, K., Rawat, Y.S., Shah, M.: Visual-textual capsule routing for text-based video segmentation. In: CVPR (2020)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 161, + 480, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 161, + 480, + 182 + ], + "spans": [ + { + "bbox": [ + 133, + 161, + 480, + 182 + ], + "type": "text", + "content": "59. Nagrani, A., Yang, S., Arnab, A., Jansen, A., Schmid, C., Sun, C.: Attention bottlenecks for multimodal fusion. In: NeuIPS (2021)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 183, + 479, + 204 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 183, + 479, + 204 + ], + "spans": [ + { + "bbox": [ + 132, + 183, + 479, + 204 + ], + "type": "text", + "content": "60. OpenAI: ChatGPT: Optimizing language models for dialogue. https://openai.com/ (2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 205, + 480, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 205, + 480, + 226 + ], + "spans": [ + { + "bbox": [ + 132, + 205, + 480, + 226 + ], + "type": "text", + "content": "61. Ordonez, V., Kulkarni, G., Berg, T.: Im2Text: Describing images using 1 million captioned photographs. In: NeurIPS (2011)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 227, + 480, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 227, + 480, + 258 + ], + "spans": [ + { + "bbox": [ + 132, + 227, + 480, + 258 + ], + "type": "text", + "content": "62. Ou, W., Zhang, J., Peng, K., Yang, K., Jaworek, G., Müller, K., Stiefelhagen, R.: Indoor navigation assistance for visually impaired people via dynamic SLAM and panoptic segmentation with an RGB-D sensor. In: ICCHP (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 259, + 480, + 292 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 259, + 480, + 292 + ], + "spans": [ + { + "bbox": [ + 132, + 259, + 480, + 292 + ], + "type": "text", + "content": "63. Peng, K., Roitberg, A., Yang, K., Zhang, J., Stiefelhagen, R.: TransDARC: Transformer-based driver activity recognition with latent space feature calibration. In: IROS (2022)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 293, + 480, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 293, + 480, + 314 + ], + "spans": [ + { + "bbox": [ + 132, + 293, + 480, + 314 + ], + "type": "text", + "content": "64. Pramanick, P., Sarkar, C., Paul, S., dev Roychoudhury, R., Bhowmick, B.: DoRO: Disambiguation of referred object for embodied agents. RA-L (2022)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 315, + 480, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 315, + 480, + 335 + ], + "spans": [ + { + "bbox": [ + 132, + 315, + 480, + 335 + ], + "type": "text", + "content": "65. Pramono, R.R.A., Chen, Y.T., Fang, W.H.: Spatial-temporal action localization with hierarchical self-attention. TMM (2021)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 336, + 480, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 336, + 480, + 369 + ], + "spans": [ + { + "bbox": [ + 132, + 336, + 480, + 369 + ], + "type": "text", + "content": "66. Qiu, H., Li, H., Wu, Q., Meng, F., Shi, H., Zhao, T., Ngan, K.N.: Language-aware fine-grained object representation for referring expression comprehension. In: MM (2020)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 369, + 480, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 369, + 480, + 402 + ], + "spans": [ + { + "bbox": [ + 132, + 369, + 480, + 402 + ], + "type": "text", + "content": "67. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., Krueger, G., Sutskever, I.: Learning transferable visual models from natural language supervision. In: ICML (2021)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 403, + 480, + 424 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 403, + 480, + 424 + ], + "spans": [ + { + "bbox": [ + 132, + 403, + 480, + 424 + ], + "type": "text", + "content": "68. Rajasegaran, J., Pavlakos, G., Kanazawa, A., Feichtenhofer, C., Malik, J.: On the benefits of 3D pose and tracking for human action recognition. In: CVPR (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 425, + 480, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 425, + 480, + 456 + ], + "spans": [ + { + "bbox": [ + 132, + 425, + 480, + 456 + ], + "type": "text", + "content": "69. Ryali, C., Hu, Y., Bolya, D., Wei, C., Fan, H., Huang, P., Aggarwal, V., Chowdhury, A., Poursaeed, O., Hoffman, J., Malik, J., Li, Y., Feichtenhofer, C.: Hiera: A hierarchical vision transformer without the bells-and-whistles. In: ICML (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 457, + 480, + 478 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 457, + 480, + 478 + ], + "spans": [ + { + "bbox": [ + 132, + 457, + 480, + 478 + ], + "type": "text", + "content": "70. Saha, J., Chowdhury, C., Chowdury, I.R., Roy, P.: Fine grained activity recognition using smart handheld. In: ICDCN (2018)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 479, + 480, + 501 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 479, + 480, + 501 + ], + "spans": [ + { + "bbox": [ + 132, + 479, + 480, + 501 + ], + "type": "text", + "content": "71. Seibold, C.M., Reif, S., Kleesiek, J., Stiefelhagen, R.: Reference-guided pseudolabel generation for medical semantic segmentation. In: AAAI (2022)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 502, + 480, + 522 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 502, + 480, + 522 + ], + "spans": [ + { + "bbox": [ + 132, + 502, + 480, + 522 + ], + "type": "text", + "content": "72. Seo, S., Lee, J.Y., Han, B.: URVOS: Unified referring video object segmentation network with a large-scale benchmark. In: ECCV (2020)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 523, + 480, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 523, + 480, + 544 + ], + "spans": [ + { + "bbox": [ + 132, + 523, + 480, + 544 + ], + "type": "text", + "content": "73. Shao, D., Zhao, Y., Dai, B., Lin, D.: Finegym: A hierarchical video dataset for fine-grained action understanding. In: CVPR (2020)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 545, + 480, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 545, + 480, + 577 + ], + "spans": [ + { + "bbox": [ + 132, + 545, + 480, + 577 + ], + "type": "text", + "content": "74. Sharma, P., Ding, N., Goodman, S., Soricut, R.: Conceptual captions: A cleaned, hypernymed, image alt-text dataset for automatic image captioning. In: ACL (2018)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 132, + 577, + 480, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 577, + 480, + 599 + ], + "spans": [ + { + "bbox": [ + 132, + 577, + 480, + 599 + ], + "type": "text", + "content": "75. Shi, H., Pan, W., Zhao, Z., Zhang, M., Wu, F.: Unsupervised domain adaptation for referring semantic segmentation. In: MM (2023)" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 132, + 600, + 480, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 600, + 480, + 620 + ], + "spans": [ + { + "bbox": [ + 132, + 600, + 480, + 620 + ], + "type": "text", + "content": "76. Shi, H., Li, H., Meng, F., Wu, Q.: Key-word-aware network for referring expression image segmentation. In: ECCV (2018)" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 132, + 621, + 480, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 621, + 480, + 643 + ], + "spans": [ + { + "bbox": [ + 132, + 621, + 480, + 643 + ], + "type": "text", + "content": "77. Shi, Y., Xu, H., Yuan, C., Li, B., Hu, W., Zha, Z.J.: Learning video-text aligned representations for video captioning. TOMM (2023)" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 132, + 644, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 644, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 644, + 480, + 665 + ], + "type": "text", + "content": "78. Soomro, K., Zamir, A.R., Shah, M.: UCF101: A dataset of 101 human actions classes from videos in the wild. arXiv preprint arXiv:1212.0402 (2012)" + } + ] + } + ], + "index": 23 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 250, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 250, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 250, + 101 + ], + "type": "text", + "content": "K. Peng, J. Fu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 480, + 632 + ], + "type": "list", + "angle": 0, + "index": 22, + "blocks": [ + { + "bbox": [ + 130, + 116, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 480, + 138 + ], + "type": "text", + "content": "79. Su, Y., Wang, W., Liu, J., Ma, S., Yang, X.: Sequence as a whole: A unified framework for video action localization with long-range text query. TIP (2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 140, + 480, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 140, + 480, + 160 + ], + "spans": [ + { + "bbox": [ + 130, + 140, + 480, + 160 + ], + "type": "text", + "content": "80. Vasudevan, A.B., Dai, D., Van Gool, L.: Object referring in videos with language and human gaze. In: CVPR (2018)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 161, + 480, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 161, + 480, + 194 + ], + "spans": [ + { + "bbox": [ + 132, + 161, + 480, + 194 + ], + "type": "text", + "content": "81. Wang, L., Huang, B., Zhao, Z., Tong, Z., He, Y., Wang, Y., Wang, Y., Qiao, Y.: VideoMAE V2: Scaling video masked autoencoders with dual masking. In: CVPR (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 194, + 480, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 194, + 480, + 216 + ], + "spans": [ + { + "bbox": [ + 132, + 194, + 480, + 216 + ], + "type": "text", + "content": "82. Wang, M., Xing, J., Mei, J., Liu, Y., Jiang, Y.: ActionCLIP: Adapting language-image pretrained models for video action recognition. TNNLS (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 216, + 480, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 216, + 480, + 248 + ], + "spans": [ + { + "bbox": [ + 132, + 216, + 480, + 248 + ], + "type": "text", + "content": "83. Wang, R., Chen, D., Wu, Z., Chen, Y., Dai, X., Liu, M., Yuan, L., Jiang, Y.G.: Masked video distillation: Rethinking masked feature modeling for self-supervised video representation learning. In: CVPR (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 249, + 480, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 249, + 480, + 270 + ], + "spans": [ + { + "bbox": [ + 132, + 249, + 480, + 270 + ], + "type": "text", + "content": "84. Wang, S., Yan, R., Huang, P., Dai, G., Song, Y., Shu, X.: Com-STAL: Compositional spatio-temporal action localization. TCSVT (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 271, + 480, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 271, + 480, + 314 + ], + "spans": [ + { + "bbox": [ + 132, + 271, + 480, + 314 + ], + "type": "text", + "content": "85. Wang, Y., Li, K., Li, Y., He, Y., Huang, B., Zhao, Z., Zhang, H., Xu, J., Liu, Y., Wang, Z., Xing, S., Chen, G., Pan, J., Yu, J., Wang, Y., Wang, L., Qiao, Y.: InternVideo: General video foundation models via generative and discriminative learning. arXiv preprint arXiv:2212.03191 (2022)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 315, + 480, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 315, + 480, + 335 + ], + "spans": [ + { + "bbox": [ + 132, + 315, + 480, + 335 + ], + "type": "text", + "content": "86. Wu, D., Han, W., Wang, T., Dong, X., Zhang, X., Shen, J.: Referring multi-object tracking. In: CVPR (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 336, + 480, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 336, + 480, + 357 + ], + "spans": [ + { + "bbox": [ + 132, + 336, + 480, + 357 + ], + "type": "text", + "content": "87. Wu, W., Luo, H., Fang, B., Wang, J., Ouyang, W.: Cap4Video: What can auxiliary captions do for text-video retrieval? In: CVPR (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 358, + 480, + 379 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 358, + 480, + 379 + ], + "spans": [ + { + "bbox": [ + 132, + 358, + 480, + 379 + ], + "type": "text", + "content": "88. Xiao, J., Shang, X., Yao, A., Chua, T.S.: NExT-QA: Next phase of question-answering to explaining temporal actions. In: CVPR (2021)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 380, + 480, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 380, + 480, + 402 + ], + "spans": [ + { + "bbox": [ + 132, + 380, + 480, + 402 + ], + "type": "text", + "content": "89. Yang, A., Miech, A., Sivic, J., Laptev, I., Schmid, C.: Just ask: Learning to answer questions from millions of narrated videos. In: ICCV (2021)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 403, + 480, + 424 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 403, + 480, + 424 + ], + "spans": [ + { + "bbox": [ + 132, + 403, + 480, + 424 + ], + "type": "text", + "content": "90. Yang, P., Wang, X., Duan, X., Chen, H., Hou, R., Jin, C., Zhu, W.: AVQA: A dataset for audio-visual question answering on videos. In: MM (2022)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 425, + 480, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 425, + 480, + 445 + ], + "spans": [ + { + "bbox": [ + 132, + 425, + 480, + 445 + ], + "type": "text", + "content": "91. Yu, L., Poirson, P., Yang, S., Berg, A.C., Berg, T.L.: Modeling context in referring expressions. In: ECCV (2016)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 447, + 480, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 447, + 480, + 479 + ], + "spans": [ + { + "bbox": [ + 132, + 447, + 480, + 479 + ], + "type": "text", + "content": "92. Yu, Z., Xu, D., Yu, J., Yu, T., Zhao, Z., Zhuang, Y., Tao, D.: ActivityNet-QA: A dataset for understanding complex web videos via question answering. In: AAAI (2019)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 479, + 480, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 479, + 480, + 511 + ], + "spans": [ + { + "bbox": [ + 132, + 479, + 480, + 511 + ], + "type": "text", + "content": "93. Yuan, Z., Yan, X., Liao, Y., Zhang, R., Wang, S., Li, Z., Cui, S.: InstanceRefer: Cooperative holistic understanding for visual grounding on point clouds through instance multi-level contextual referring. In: ICCV (2021)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 512, + 480, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 512, + 480, + 533 + ], + "spans": [ + { + "bbox": [ + 132, + 512, + 480, + 533 + ], + "type": "text", + "content": "94. Zeng, F., Dong, B., Zhang, Y., Wang, T., Zhang, X., Wei, Y.: MOTR: End-to-end multiple-object tracking with transformer. In: ECCV (2022)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 534, + 480, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 534, + 480, + 555 + ], + "spans": [ + { + "bbox": [ + 132, + 534, + 480, + 555 + ], + "type": "text", + "content": "95. Zhang, G., Ren, J., Gu, J., Tresp, V.: Multi-event video-text retrieval. In: CVPR (2023)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 556, + 480, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 556, + 480, + 578 + ], + "spans": [ + { + "bbox": [ + 132, + 556, + 480, + 578 + ], + "type": "text", + "content": "96. Zhang, H., Li, X., Bing, L.: Video-LLaMA: An instruction-tuned audio-visual language model for video understanding. In: EMNLP (2023)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 132, + 578, + 480, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 578, + 480, + 610 + ], + "spans": [ + { + "bbox": [ + 132, + 578, + 480, + 610 + ], + "type": "text", + "content": "97. Zheng, J., Zhang, J., Yang, K., Peng, K., Stiefelhagen, R.: MateRobot: Material recognition in wearable robotics for people with visual impairments. In: ICRA (2024)" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 132, + 611, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 611, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 611, + 480, + 632 + ], + "type": "text", + "content": "98. Zong, D., Sun, S.: McOmet: Multimodal fusion transformer for physical audiovisual commonsense reasoning. In: AAAI (2023)" + } + ] + } + ], + "index": 21 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 413, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 413, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 413, + 91, + 447, + 100 + ], + "type": "text", + "content": "RAVAR" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "19" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 18 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Refine, Discriminate and Align_ Stealing Encoders via Sample-Wise Prototypes and Multi-Relational Extraction/2d1e904f-54f5-49e9-a836-f9f59c16b169_content_list.json b/2024/Refine, Discriminate and Align_ Stealing Encoders via Sample-Wise Prototypes and Multi-Relational Extraction/2d1e904f-54f5-49e9-a836-f9f59c16b169_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..0cf19423df8e4546f50a49de33b6b200ee0a81dc --- /dev/null +++ b/2024/Refine, Discriminate and Align_ Stealing Encoders via Sample-Wise Prototypes and Multi-Relational Extraction/2d1e904f-54f5-49e9-a836-f9f59c16b169_content_list.json @@ -0,0 +1,1802 @@ +[ + { + "type": "text", + "text": "Refine, Discriminate and Align: Stealing Encoders via Sample-Wise Prototypes and Multi-Relational Extraction", + "text_level": 1, + "bbox": [ + 233, + 140, + 772, + 207 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Shuchi $\\mathbf{W u}^{1}$ Chuan Ma $^{2\\text{图}}$ Kang Wei $^{3\\text{图}}$ Xiaogang Xu $^{4}$ Ming Ding $^{5}$ Yuwen Qian $^{1}$ Di Xiao $^{2}$ Tao Xiang $^{2}$ $^{1}$ NJUST $^{2}$ CQU $^{3}$ PolyU $^{4}$ CUHK $^{5}$ Data61, CSIRO", + "bbox": [ + 238, + 233, + 754, + 281 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. This paper introduces RDA, a pioneering approach designed to address two primary deficiencies prevalent in previous endeavors aiming at stealing pre-trained encoders: (1) suboptimal performances attributed to biased optimization objectives, and (2) elevated query costs stemming from the end-to-end paradigm that necessitates querying the target encoder every epoch. Specifically, we initially Refine the representations of the target encoder for each training sample, thereby establishing a less biased optimization objective before the steal-training phase. This is accomplished via a sample-wise prototype, which consolidates the target encoder's representations for a given sample's various perspectives. Demanding exponentially fewer queries compared to the end-to-end approach, prototypes can be instantiated to guide subsequent query-free training. For more potent efficacy, we develop a multi-relational extraction loss that trains the surrogate encoder to Discriminate mismatched embedding-prototype pairs while Aligning those matched ones in terms of both amplitude and angle. In this way, the trained surrogate encoder achieves state-of-the-art results across the board in various downstream datasets with limited queries. Moreover, RDA is shown to be robust to multiple widely-used defenses. Our code is available at https://github.com/ShuchiWu/RDA.", + "bbox": [ + 261, + 318, + 743, + 597 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Model Stealing $\\cdot$ Self-Supervised Learning $\\cdot$ Prototype", + "bbox": [ + 261, + 609, + 712, + 625 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 674, + 375, + 690 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Self-supervised learning (SSL) [4, 5, 9, 11, 12] is endowed with the capability to harness unlabeled data for pre-training a versatile encoder that applicable to a range of downstream tasks, or even showcasing groundbreaking zero-shot performances, e.g., CLIP [32]. However, SSL typically requires a large volume of data and computation resources to achieve a convincing performance, i.e., high-performance encoders are expensive to train [35]. To safeguard the confidentiality and economic worth of these encoders, entities like OpenAI offer their encoders", + "bbox": [ + 212, + 705, + 787, + 813 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "$\\text{忍}$ Corresponding Author: chuan.ma@cqu.edu.cn, adam-kang.wei@polyu.edu.hk", + "bbox": [ + 228, + 824, + 769, + 840 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/102cf00079713b5379c0dbb6a242b65d6ed740f7aa6dffd2a1c596b665e68086.jpg", + "image_caption": [ + "Fig. 1: Illustrations of four stealing methods against SSL. The dotted arrows and text beside interpret how each method optimizes the surrogate encoder. Surrogate encoder branches in (b)-(d) involve data augmentations for training. Both (c) and (d) augment each sample before querying the target encoder but adopting different schemes." + ], + "image_footnote": [], + "bbox": [ + 289, + 143, + 380, + 253 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/bf5a8f6fe9d7486cf9958564a135c5470f0e61d6b0083e049e9045f42bcdca4b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 395, + 146, + 493, + 251 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/30de37fa9d6474e983a08d6720e649e4760880ab9a552717ae81b094d75b0800.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 504, + 146, + 602, + 251 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/c3e937a76416808419e7ca5890fef7608f73334babc6b6527955b087f1ce55da.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 611, + 145, + 712, + 251 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "as a premium service, exclusively unveiling the service API to the public. Users have the privilege of soliciting embeddings for their data, facilitating the training of diverse models tailored for specific downstream tasks. Regrettably, the substantial value of these encoders and their exposure to publicly accessible APIs render them susceptible to model stealing attacks [10,26,34].", + "bbox": [ + 212, + 351, + 784, + 426 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Given a surrogate dataset, model stealing attacks aim to mimic the outputs of a target model to train either a high-accuracy copy of comparable performance or a high-fidelity copy that can serve as a stepping stone to perform further attacks like adversarial examples [31], membership inference attacks [33,36], etc. The goal of this paper is to develop an approach that can train a surrogate encoder with competitive performance on downstream tasks by only accessing the target encoder's output embeddings. Meanwhile, it should require as little cost as possible, which is primarily derived from querying the target encoder.", + "bbox": [ + 212, + 429, + 787, + 549 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We begin with systematically studying existing techniques for stealing pretrained encoders. Specifically, the conventional method [34] and StolenEncoder [26] are similar, and both only need to query the target encoder once with each sample before the training. The output embedding can be viewed as a \"ground truth\" or an \"optimization objective\" to the sample, and the surrogate encoder is optimized to output a similar embedding when the same sample is fed, as illustrated in Figure 1 (a). The sole distinction lies in the optimization of StolenEncoder, which incorporates data augmentations (refer to Figure 1 (b)), as it supposes the target encoder will produce similar embeddings for an image and its augmentations. Nevertheless, we contend that this presumed similarity is of a modest degree, given the disparity between the surrogate and pre-training data. The target encoder's representations for such data's various augmentations may diverge and even be biased to other samples, as visually demonstrated in Figure 2 using a toy experiment. In this sense, only using the embedding of an image's single perspective (regardless of the original or augmented version) as the optimization objective is inadvisable.", + "bbox": [ + 212, + 551, + 787, + 792 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "On the other hand, Cont-Steal [34] augments each sample into two perspectives in each epoch: one is used to query the target encoder while the other is fed to the surrogate encoder. Two embeddings of the same sample from the tar", + "bbox": [ + 212, + 794, + 785, + 839 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "S. Wu et al.", + "bbox": [ + 271, + 114, + 352, + 126 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/1860f8975f0322c0f64a15a77cbd8ad0f96ab1be3e835af5f480b2c40c220fd1.jpg", + "image_caption": [ + "Fig. 2: t-SNE of embeddings belonging to five different images generated by an encoder pre-trained on CIFAR10, with each image augmented into 500 patches and fed into the encoder. Each black marker represents the mean of the 500 embeddings of a certain image, i.e., its prototype. Among the embeddings of an image's various augmentations, some can be diverged or even biased. In contrast, each image's prototype is more distinguishable, i.e., less biased." + ], + "image_footnote": [], + "bbox": [ + 405, + 143, + 598, + 256 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "get and surrogate encoders will be aligned, while those of different samples will be pushed apart, as illustrated in Figure 1 (c). In spite of such an end-to-end training scheme performing better, it suffers from high query costs since each sample is augmented and used to query in each epoch.", + "bbox": [ + 212, + 382, + 787, + 443 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To tackle these issues, we propose RDA. Specifically, we first Refine the target encoder's representations for each sample in the surrogate dataset by averagely aggregating its several (e.g., 10) different augmentations' embeddings, i.e., their mean value, to establish a sample-level prototype for it. Then, the sample-wise prototype is used to guide the surrogate encoder optimization, which can mitigate the impact of biased embeddings, as shown in Figure 2 (see black markers). We have an experiment presented in Figure 10 of our supplementary material to further quantitatively reveal the benefit of prototypes, i.e., significantly more similar with each augmentation patch's embedding, showcasing it is less biased. With a prototype, there is a static optimization objective for each sample across the entire training, and thus, it is done in a query-free manner. This enables RDA to have far less query cost than the end-to-end approach (less than $10\\%$ ). To further enhance the attack efficacy with limited queries, we develop a multi-relational extraction loss that trains the surrogate encoder to Discriminate mismatched embedding-prototype pairs while Aligning those matched ones in terms of both amplitude and angle. The framework and pipeline of RDA are depicted in Figures 1 (d) and 4, respectively.", + "bbox": [ + 212, + 444, + 787, + 700 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Extensive experiments show that RDA acquires a favorable trade-off between the query cost (a.k.a, the money cost) and attack efficacy. As depicted in Figure 3, when the target encoder is pre-trained on CIFAR10, RDA averagely outperforms the state-of-the-art (SOTA), i.e., Cont-Steal [34], by $1.22\\%$ on the stealing efficacy over seven different downstream datasets, with a competitive smaller query cost, i.e., only $1\\%$ of Cont-Steal. Further, this performance gap can be widened to $5.20\\%$ with $10\\%$ query cost of Cont-Steal. In particular, RDA is demonstrated to be robust against multiple prevalent defenses [18,30,44].", + "bbox": [ + 212, + 702, + 787, + 823 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In conclusion, our contributions are three folds:", + "bbox": [ + 238, + 824, + 584, + 839 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "Refine, Discriminate and Align (RDA)", + "bbox": [ + 473, + 114, + 730, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/56a5f18e0e51d7821f4c94d07418fb7820597df2269ce1ae8920a67ff055fcfc.jpg", + "image_caption": [ + "Fig. 3: Performance comparisons between four stealing methods against SSL. The presented results are the mean values achieved by each method over seven different downstream classification tasks, with their corresponding query costs. Our proposed RDA can achieve SOTA results with the least query cost." + ], + "image_footnote": [], + "bbox": [ + 354, + 147, + 643, + 262 + ], + "page_idx": 3 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We comprehensively investigate two critical inadequacies of existing stealing methods against SSL, i.e., suboptimal efficacy and high query costs, and analyze the causes.", + "- We develop a novel approach to train surrogate encoders, namely RDA, with sample-wise prototype guidance and a multi-relational extraction loss.", + "- Extensive experiments are conducted to verify the effectiveness and robustness of RDA." + ], + "bbox": [ + 225, + 362, + 782, + 464 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 503, + 387, + 518 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Prototype Learning. A prototype refers to the mean of embeddings belonging to all images of an identical class, which serves as a proxy of the class [40]. With each class having a prototype, the model is trained to match its output with the corresponding prototype when inputting a sample of a certain class. Prototype learning has proven beneficial for various learning scenarios, e.g., federated learning [8, 16, 41] and few-shot learning [28, 38, 42]. In this paper, the optimization objective (i.e., the mean of embeddings belonging to an identical sample's multiple augmentations) we generate for each training sample is conceptually similar to prototypes, and thus we name it a sample-wise prototype.", + "bbox": [ + 212, + 537, + 782, + 672 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Model Stealing Attacks against SSL. Dziedzic et al. [10] pointed out that the higher dimension of embeddings leaks more information than labels, making SSL more easily stolen. The primary objective of a stealing attack is to train a surrogate encoder to achieve high accuracy on downstream tasks or recreate a high-fidelity copy that can be used to mount further attacks such as adversarial examples [3] and membership inference attacks [25].", + "bbox": [ + 212, + 672, + 782, + 762 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Defenses against Model Stealing Attacks. Existing defenses against model stealing attacks can be categorized based on when they are applied [10]. Perturbation-based defenses, e.g., adding noise [30], top- $k$ [30] and truncating outputs [44], are applied before a stealing attack happens, aiming to limit the information leakage. On the other hand, watermarking defenses embed watermarks or unique", + "bbox": [ + 212, + 763, + 810, + 839 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "S. Wu et al.", + "bbox": [ + 271, + 114, + 352, + 126 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "identifiers into the target model and use them to detect whether it is stolen after a stealing attack happens [17].", + "bbox": [ + 212, + 146, + 782, + 176 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3 Methodology", + "text_level": 1, + "bbox": [ + 214, + 199, + 380, + 217 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1 Threat Model", + "text_level": 1, + "bbox": [ + 214, + 229, + 377, + 243 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Given a target encoder $E_{T}$ , the attacker aims to train a surrogate encoder $E_{S}$ at the lowest possible cost, which can perform competitively on downstream tasks with $E_{T}$ . To achieve this goal, the attacker queries $E_{T}$ for embeddings of an unlabeled surrogated dataset $D_{S} = \\{\\pmb{x}_{1},\\dots,\\pmb{x}_{N}\\}$ consisting of $N$ images, and guide $E_{S}$ to mimic the output of $E_{T}$ for each sample in $D_{S}$ . Specifically, we consider a black-box setting, where the attacker has direct access only to the outputs of $E_{T}$ while remaining unaware of its architecture and training configurations, including pre-training datasets, loss functions, data augmentations schemes, etc.", + "bbox": [ + 212, + 253, + 787, + 376 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2 Sample-Wise Prototypes", + "text_level": 1, + "bbox": [ + 214, + 396, + 465, + 411 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Recalling the discussion in Section 1, we expect to refine the target encoder's representations to find a less biased optimization objective for each sample in $D_{S}$ . Enlightened by prototype learning [40] and Extreme-Multi-Patch-SSL (EMPSL) [43], which establish a prototype/benchmark for each class/sample, we propose a sample-wise prototype generation method for this purpose. In detail, the process of prototype generation can be divided into three steps as follows: 1 Cropping and augmenting each image $\\pmb{x}_i\\in D_S$ into $n$ augmentation patches denoted as $\\{\\pmb{x}_{i,t,c}^{\\prime}\\}_{c = 1}^{n} = \\{\\pmb{x}_{i,t,1}^{\\prime},\\dots,\\pmb{x}_{i,t,n}^{\\prime}\\}$ ; 2 Querying $E_{T}$ with each augmentation patch in $\\{\\pmb{x}_{i,t,c}^{\\prime}\\}_{c = 1}^{n}$ , resulting in a set of embeddings denoted as $\\{E_T(\\pmb{x}_{i,t,c}^{\\prime})\\}_{c = 1}^{n} = \\{E_T(\\pmb{x}_{i,t,1}^{\\prime}),\\dots,E_T(\\pmb{x}_{i,t,n}^{\\prime})\\}$ ; 3 Calculating the mean of $\\{E_T(\\pmb{x}_{i,t,c}^{\\prime})\\}_{c = 1}^{n}$ as the prototype $p_{\\pmb{x}_i}$ for $\\pmb{x}_i$ as follows:", + "bbox": [ + 212, + 419, + 787, + 580 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\np _ {\\boldsymbol {x} _ {i}} = \\frac {1}{n} \\sum_ {c = 1} ^ {n} E _ {T} \\left(\\boldsymbol {x} _ {i, t, c} ^ {\\prime}\\right), \\quad \\boldsymbol {x} _ {i} \\in D _ {s}. \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 387, + 588, + 784, + 604 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Each generated prototype will be stored in a memory bank. These sample-wise prototypes can provide a stable optimization objective for each sample throughout the training process. Consequently, the attacker can accomplish the training in a query-free manner. Notably, setting $n$ to a value that is far less than the training epochs, e.g., 10 vs. 100, is sufficient to yield a favorable performance. The prototype generation process is illustrated in step ① of Figure 4. Supplementary A.2 explains how we build the memory bank in practice.", + "bbox": [ + 212, + 613, + 787, + 720 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3 Multi-Relational Extraction Loss", + "text_level": 1, + "bbox": [ + 214, + 739, + 534, + 753 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As step ② of Figure 4 shows, during training $E_{S}$ , we also perform cropping and augmentation on each image $\\pmb{x}_i \\in D_S$ to create $m$ patches, forming an augmentation patch set denoted as $\\{\\pmb{x}_{i,s,q}^{\\prime}\\}_{q = 1}^{m} = \\{\\pmb{x}_{i,s,1}^{\\prime},\\dots,\\pmb{x}_{i,s,m}^{\\prime}\\}$ . The $m$ does not necessarily need to be set equal to $n$ . Next, we feed each augmentation patch from $\\{\\pmb{x}_{i,s,q}^{\\prime}\\}_{q = 1}^{m}$ to $E_{S}$ , obtaining a set of embeddings denoted as", + "bbox": [ + 212, + 763, + 787, + 842 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "Refine, Discriminate and Align (RDA)", + "bbox": [ + 473, + 114, + 730, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/c1ffdac658369ac2a9dd058449fc982831b16cf26efdf59a466e4c4bad57f29c.jpg", + "image_caption": [ + "Fig. 4: Pipeline of RDA. Prototype generation: augment one sample into $n$ patches and use them to query the target encoder $(E_T)$ . The mean of the $n$ patches's embeddings is defined as a prototype for this sample. Forward encoding: crop one image into $m$ patches and feed them to the surrogate encoder $(E_S)$ for their embeddings. Optimization: align embeddings from the surrogate encoder to their matched prototypes in both angle and amplitude while pushing away those belonging to different samples." + ], + "image_footnote": [], + "bbox": [ + 222, + 147, + 782, + 242 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "$\\{E_S(\\pmb{x}_{i,s,q}^{\\prime})\\}_{q = 1}^{m} = \\{E_S(\\pmb{x}_{i,s,1}^{\\prime}),\\dots ,E_S(\\pmb{x}_{i,s,m}^{\\prime})\\}$ . To optimize $E_{S}$ , we develop a multi-relational extraction loss composed of two parts, i.e., the discriminating loss and the aligning loss.", + "bbox": [ + 212, + 366, + 787, + 412 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Discriminating Loss. The discriminating loss is responsible for training $E_{S}$ to distinguish between different samples, as depicted in step ③ of Figure 4. For this goal, it pushes each embedding in $\\{E_S(\\pmb{x}_{i,s,q}')\\}_{q = 1}^m$ away from mismatched prototypes $p_{\\pmb{x}_j}, i \\neq j$ (denoted as negative pairs). In this regard, contrastive learning [5,6,12,19] offers a solution, as it can pull embeddings of positive pairs close while pushing embeddings of negative pairs apart. Furthermore, since more negative pairs can help improve the performance of contrastive learning [5], we refer to the loss designed by Sha et al. [34] to propose our discriminating loss $\\mathcal{L}_D$ . In particular, $\\mathcal{L}_D$ considers both mismatched prototype-embedding pairs from different samples as well as embeddings from $E_{S}$ for different samples as negative pairs. Formally, $\\mathcal{L}_D$ can be expressed as follows:", + "bbox": [ + 212, + 411, + 787, + 568 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {p o s} \\left(\\boldsymbol {x} _ {i}\\right) = \\frac {1}{m} \\sum_ {q = 1} ^ {m} \\exp \\left(\\operatorname {s i m} \\left(E _ {S} \\left(\\boldsymbol {x} _ {i, s, q} ^ {\\prime}\\right), p _ {\\boldsymbol {x} _ {i}}\\right) / \\tau\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 341, + 574, + 785, + 592 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {n e g} \\left(\\boldsymbol {x} _ {i}\\right) = \\frac {1}{m} \\sum_ {q = 1} ^ {m} \\sum_ {k = 1} ^ {N} \\mathbb {1} _ {[ i \\neq k ]} \\left(\\exp \\left(s i m \\left(E _ {S} \\left(\\boldsymbol {x} _ {i, s, q} ^ {\\prime}\\right), p _ {\\boldsymbol {x} _ {k}}\\right) / \\tau\\right) \\right. \\tag {3} \\\\ + \\exp \\left(\\operatorname {s i m} \\left(E _ {S} \\left(\\boldsymbol {x} _ {i, s, q} ^ {\\prime}\\right), E _ {S} \\left(\\boldsymbol {x} _ {k, s, q} ^ {\\prime}\\right)\\right) / \\tau\\right)), \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 341, + 609, + 785, + 657 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {D} = - \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\log \\frac {\\mathcal {L} _ {p o s} (\\boldsymbol {x} _ {i})}{\\mathcal {L} _ {n e g} (\\boldsymbol {x} _ {i})}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 413, + 662, + 785, + 683 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\operatorname{sim}(u, v)$ represents the cosine similarity between $u$ and $v$ , and $\\tau$ is a temperature parameter.", + "bbox": [ + 212, + 685, + 787, + 714 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Aligning Loss. While $\\mathcal{L}_D$ can separate different samples effectively, it does not fully leverage the potential of prototypes. To further enhance the attack efficacy, we propose to align embeddings from $E_S$ more thoroughly with their matched prototypes from the high-performance $E_T$ , i.e., in terms of both amplitude and angle, as step ③ of Figure 4 shows. To measure the amplitude and angle deviations between two embeddings, we employ the mean square error (MSE) and cosine similarity to quantify them respectively, which can be formulated as follows:", + "bbox": [ + 212, + 714, + 787, + 823 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {a m p} ^ {\\prime} \\left(\\boldsymbol {x} _ {i}\\right) = \\frac {1}{m} \\sum_ {q = 1} ^ {m} \\left\\| E _ {S} \\left(\\boldsymbol {x} _ {i, s, q} ^ {\\prime}\\right) - p _ {\\boldsymbol {x} _ {i}} \\right\\| ^ {2}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 370, + 824, + 785, + 842 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "S. Wu et al.", + "bbox": [ + 271, + 114, + 352, + 126 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {a n g}} ^ {\\prime} \\left(\\boldsymbol {x} _ {i}\\right) = \\frac {1}{m} \\sum_ {q = 1} ^ {m} \\operatorname {s i m} \\left(E _ {S} \\left(\\boldsymbol {x} _ {i, s, q} ^ {\\prime}\\right), p _ {\\boldsymbol {x} _ {i}}\\right). \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 357, + 160, + 785, + 178 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Notice that we will normalize the value of $\\mathcal{L}_{ang}^{\\prime}$ to [0, 1]. However, there is an issue concerning the equal penalty given by $\\mathcal{L}_{amp}^{\\prime}$ and $\\mathcal{L}_{ang}^{\\prime}$ to every deviation increase. For instance, when the MSE increases from 0.8 to 0.9 or from 0.3 to 0.4, both penalties given by $\\mathcal{L}_{amp}^{\\prime}$ are 0.1, which is imprudent. On the contrary, we should have a more rigorous penalty for the increase in MSE from 0.3 to 0.4, as a smaller MSE is more desirable. Likewise, the same penalizing regime should be applied on $1 / \\mathcal{L}_{ang}^{\\prime}$ for the same reason. To this end, we adopt a logarithmic function to redefine the ultimate formulations of $\\mathcal{L}_{amp}$ and $\\mathcal{L}_{ang}$ as follows:", + "bbox": [ + 214, + 180, + 787, + 303 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {a m p}} \\left(\\boldsymbol {x} _ {i}\\right) = \\log \\mathcal {L} _ {\\text {a m p}} ^ {\\prime} \\left(\\boldsymbol {x} _ {i}\\right), \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 416, + 315, + 785, + 330 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {a n g} \\left(\\boldsymbol {x} _ {i}\\right) = \\log \\left(1 / \\mathcal {L} _ {a n g} ^ {\\prime} \\left(\\boldsymbol {x} _ {i}\\right)\\right) = - \\log \\mathcal {L} _ {a n g} ^ {\\prime} \\left(\\boldsymbol {x} _ {i}\\right). \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 357, + 359, + 784, + 375 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We posit the amplitude and angle deviations hold equal importance, and formulate the aligning loss as follows:", + "bbox": [ + 214, + 378, + 782, + 407 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {A} = \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\left(\\mathcal {L} _ {a m p} (\\boldsymbol {x} _ {i}) + \\mathcal {L} _ {a n g} (\\boldsymbol {x} _ {i})\\right) = - \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\log \\frac {\\mathcal {L} _ {a n g} ^ {\\prime} (\\boldsymbol {x} _ {i})}{\\mathcal {L} _ {a m p} ^ {\\prime} (\\boldsymbol {x} _ {i})}. \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 316, + 419, + 785, + 450 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Our experiments in Supplementary A.4 demonstrate the combination of $\\mathcal{L}_{amp}$ and $\\mathcal{L}_{ang}$ offers superior results than using each solely.", + "bbox": [ + 215, + 459, + 782, + 489 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Lastly, the loss function for stealing pre-trained encoders is as follows:", + "bbox": [ + 238, + 489, + 743, + 505 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\lambda_ {1} \\cdot \\mathcal {L} _ {D} + \\lambda_ {2} \\cdot \\mathcal {L} _ {A}, \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 429, + 518, + 784, + 534 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $\\lambda_{1}$ and $\\lambda_{2}$ are preset coefficients to adjust the weight of each part in $\\mathcal{L}$ . To show the superiority of our design, especially on the loss function, we have conducted ablation studies in Section 4.4 and explored several alternative designs in Supplementary A.3.", + "bbox": [ + 214, + 545, + 782, + 604 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Detailed steps of RDA are summarized in Algorithm 1 of Supplementary A.2.", + "bbox": [ + 238, + 606, + 782, + 621 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 215, + 645, + 374, + 662 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.1 Experimental Setup", + "text_level": 1, + "bbox": [ + 215, + 676, + 426, + 693 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Target Encoder Settings. We use SimCLR [5] to pre-train two ResNet18 [13] encoders on CIFAR10 [21] and STL10 [7], respectively, as two medium-scale target encoders. Furthermore, we consider two real-world large-scale ResNet50 encoders as the targets, i.e., the ImageNet encoder pre-trained by Google [5], and the CLIP encoder pre-trained by OpenAI [32]. Besides the ResNet family, RDA also has been demonstrated effective upon other backbones, i.e., VGG19_bn [37], DenseNet121 [15], and MobileNetV2 [14], in Supplementary A.4.", + "bbox": [ + 212, + 703, + 784, + 809 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Attack Settings. Regarding the surrogate dataset, it is derived from Tiny ImageNet [23]. Specifically, we randomly sample 2,500 images and resize them to", + "bbox": [ + 214, + 809, + 785, + 839 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "Refine, Discriminate and Align (RDA)", + "bbox": [ + 473, + 114, + 730, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 114, + 784, + 125 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/1c9d9b33a6eb9cda7a520e664af7331753f4c5f698f9a566e593dcf4d043291f.jpg", + "table_caption": [ + "Table 1: Results of RDA against two Table 2: Results of RDA against the medium-scale encoders. ImageNet Encoder and CLIP." + ], + "table_footnote": [], + "table_body": "
Pre-training DatasetDownstream DatasetTASA\\( \\frac{\\mathrm{{SA}}}{\\mathrm{{TA}}} \\times {100}\\% \\)
CIFAR10MNIST97.2796.6299.33
F-MNIST88.5889.32100.84
GTSRB61.7662.75101.60
SVHN73.7873.7499.95
STL10MNIST96.8496.3199.45
F-MNIST90.0887.9197.59
GTSRB64.4557.7089.53
SVHN61.8570.67114.26
", + "bbox": [ + 254, + 184, + 488, + 287 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/cd311cee8f47c44e8c749bbe8bffccb314d06ddf1be282d6564a485a2eabfa14.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
TargetEncoderDownstreamDatasetTASA\\( \\frac{\\mathrm{{SA}}}{\\mathrm{{TA}}} \\times {100}\\% \\)
ImageNetEncoderMNIST97.3895.8998.47
F-MNIST91.6790.8399.08
GTSRB63.1459.6294.43
SVHN73.2069.2194.55
CLIPMNIST97.9093.9695.98
F-MNIST89.9287.4397.23
GTSRB68.6155.8481.39
SVHN69.5357.5882.81
", + "bbox": [ + 519, + 185, + 741, + 287 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/1b040d2d994bbab0479ac50661105c46e5185032677da2d1426ed08c12753bad.jpg", + "table_caption": [ + "Table 3: Computation resources required by pre-training the two targeted real-world encoders from scratch and RDA to steal them." + ], + "table_footnote": [], + "table_body": "
TargetEncoderPre-trainingRDA
HardwareTime (hrs)HardwareTime (hrs)
ImageNetEncoderTPU v3192RTX A500010.8
CLIPV100 GPU255,74416.2
", + "bbox": [ + 370, + 340, + 630, + 391 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "$32 \\times 32$ for stealing encoders pre-trained on CIFAR10 and STL10. When stealing the ImageNet encoder and CLIP, the number is 40,000 and 60,000, respectively, with each image resized to $224 \\times 224$ . For the surrogate encoder architecture, we adopt a ResNet18 across our experiments. During the attack, we set $n$ in Eq. 1 as 10, $m$ in Eq. 2, 3, 5, and 6 as 5, $\\tau$ in Eq. 2 and 3 as 0.07, and both $\\lambda_{1}$ and $\\lambda_{2}$ in Eq. 10 as 1 unless stated otherwise. For training, the batch size is set as 100, and we employ an Adam optimizer [20] with a learning rate of 0.001.", + "bbox": [ + 212, + 417, + 784, + 523 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Evaluation Settings. We train each surrogate encoder for 100 epochs and test its KNN accuracy [45] on CIFAR10 after each epoch. The best-trained surrogate encoders, as well as target encoders, will be used to train downstream classifiers for the linear probing evaluation. Each downstream classifier will be trained for 100 epochs with an Adam optimizer and a learning rate of 0.0001. Specifically, we totally consider seven downstream datasets, namely MNIST [24], CIFAR10 [21], STL10 [7], GTSRB [39], CIFAR100 [22], SVHN [29], and F-MNIST [46]. Additional small-scale experiments on more complex datasets (e.g., Food 101 [2]) are included in our supplementary material. As done in [26], we use Target Accuracy (TA) to evaluate target encoders, Steal Accuracy (SA) to evaluate surrogate encoders, and $\\frac{\\mathrm{SA}}{\\mathrm{TA}} \\times 100\\%$ to evaluate the efficacy of the stealing attack.", + "bbox": [ + 212, + 523, + 785, + 691 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.2 Effectiveness of RDA", + "text_level": 1, + "bbox": [ + 214, + 710, + 434, + 724 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Stealing Medium-Scale Encoders. Table 1 shows the results achieved by RDA upon two medium-scale target encoders pre-trained on CIFAR10 and STL10, respectively. As the results show, using a surrogate dataset of $5\\%$ the size of CIFAR10 (2,500 / 50,000) and $2.4\\%$ the size of STL10 (2,500 / 105,000), RDA can steal nearly $100\\%$ of the functionality of both target encoders, except for the GTSRB scenario when stealing the encoder pre-trained on STL10. Nonetheless, even in this exceptional case, the obtained $\\frac{\\mathrm{SA}}{\\mathrm{TA}} \\times 100\\%$ is approximately $90\\%$ ,", + "bbox": [ + 212, + 733, + 787, + 843 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "S. Wu et al.", + "bbox": [ + 271, + 114, + 352, + 126 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/f4c9168e25fc959d5a4b810ec6aed0b592ce8c78b8fbef63ffbce383fa0ae723.jpg", + "image_caption": [ + "(a) ImageNet Encoder" + ], + "image_footnote": [], + "bbox": [ + 225, + 146, + 331, + 219 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/826d2dc194a694fd6ef313d3fa012d06054fb999154e96f0a09da433f2a4d946.jpg", + "image_caption": [ + "(b) CLIP" + ], + "image_footnote": [], + "bbox": [ + 338, + 146, + 442, + 219 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/76f52550ead86fafa6ba6471d1896e908607bce1958d55ee71f2d236d97a3d0b.jpg", + "image_caption": [ + "Fig.5: t-SNE of embeddings of 2,000 images sampled from CIFAR10 generated by the ImageNet encoder and CLIP.", + "Fig. 6: The training epochs of RDA are fixed to 100 while those of baselines are prolonged to 300." + ], + "image_footnote": [], + "bbox": [ + 506, + 143, + 727, + 268 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "which exemplifies the remarkable effectiveness of RDA. Furthermore, the surrogate encoder trained through RDA even outperforms the target encoder on multiple datasets. Our interpretation of this is that the target encoder fits the pre-training dataset more strongly compared to the RDA-trained surrogate encoder, thus exhibiting inferior performance on some out-of-distribution datasets.", + "bbox": [ + 212, + 335, + 784, + 411 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Stealing Real-World Large-Scale Encoders. We further evaluate RDA on two real-world pre-trained encoders, i.e., the ImageNet encoder and CLIP. We aim to demonstrate the effectiveness of RDA upon such large-scale encoders trained with a massive amount of data. Specifically, the ImageNet encoder is pre-trained on the ImageNet dataset, which has 1.3 million images. The CLIP encoder is pre-trained on a web-scale dataset that consists of 400 million image-text pairs. During the attack, we set the patch number for training as 1 (i.e., $m$ in Eq. 2, 3, 5, and 6) to save time. Other parameters follow the default setting except the $\\lambda_{1}$ and $\\lambda_{2}$ when stealing CLIP, in which we set $\\lambda_{1}$ and $\\lambda_{2}$ to 1 and 20, respectively. To interpret the reason for this adjustment, we randomly sample 200 images from each class in the testing set of CIFAR10 and visualize the output embeddings from the ImageNet encoder and CLIP using t-SNE [27]. Figure 5 shows that embeddings from the ImageNet encoder are more uniformly distributed compared to those from CLIP. This is because CLIP is a multimodal encoder, which is pre-trained by conducting contrastive learning between image-text pairs. Therefore, the text also shares part of the embedding space and is mutually exclusive with mismatched images. Although $\\mathcal{L}_D$ can push embeddings of different samples apart, a significant weight of it will make all embeddings more uniformly distributed, i.e., not align with CLIP. Therefore, we set a smaller weight to $\\mathcal{L}_D$ when stealing CLIP. This adjustment is practical since the attacker has access to the outputs of the target encoder, and thus can observe its embedding space to adjust its attack settings. Table 2 shows that using a surrogate dataset of only $3.07\\%$ the size of ImageNet (40K / 1.3M) and $0.015\\%$ the size of the training data of CLIP (60K / 400M), RDA can achieve comparable performances with the two encoders across various downstream tasks. The results also reveal that even under the architecture of the surrogate and target encoders being distinct (ResNet18 vs. ResNet50), RDA still exhibits high effec", + "bbox": [ + 212, + 431, + 787, + 843 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "Refine, Discriminate and Align (RDA)", + "bbox": [ + 473, + 114, + 730, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 114, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/323d735577d2edde482abecfc42bc1868e5e7b533a4645a3369c88d5b4badbde.jpg", + "table_caption": [ + "Table 4: Comparisons between RDA and baselines to steal the encoder pre-trained on CIFAR10 under the same surrogate dataset size. We report the $\\frac{\\mathrm{SA}}{\\mathrm{TA}} \\times 100\\%$ achieved by each method and its query cost. The Optimal and suboptimal results are highlighted." + ], + "table_footnote": [], + "table_body": "
MethodCIFAR10CIFAR100MNISTGTSRBSVHNSTL10F-MNISTQueries
Conventional [34]63.3645.4096.1620.3468.2660.2594.522,500
StolenEncoder [26]79.0368.7898.7663.7092.0176.2995.912,500
Cont-Steal [34]80.4373.9498.0395.5596.3980.0895.39250,000
RDA83.8477.4998.9794.2796.7982.2594.692,500
85.8583.0899.33101.6099.9585.56100.8425,000
", + "bbox": [ + 297, + 203, + 699, + 273 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/e1d768f054773423bd290e1c661e41fbee6884c2403c4886b1c1727b883d7a78.jpg", + "table_caption": [ + "Table 5: Comparisons between RDA and StolenEncoder under the same query cost." + ], + "table_footnote": [], + "table_body": "
QueriesMethodsettingSA
2500StolenEncoder2,500 images71.98
RDA500 images×5 patches65.38
625 images×4 patches67.07
1,250 images×2 patches71.44
2,500 images×1 patch76.36
25,000StolenEncoder25,000 images76.23
RDA2,500 images×10 patches78.50
", + "bbox": [ + 380, + 313, + 622, + 392 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "tiveness. Moreover, Table 3 shows that RDA requires much fewer computation resources to steal the two encoders than pre-training them from scratch, owing to the small surrogate dataset and lightweight network architecture.", + "bbox": [ + 212, + 421, + 784, + 467 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.3 Comparison with Existing Methods", + "text_level": 1, + "bbox": [ + 215, + 489, + 560, + 505 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "We consider the conventional method (CVPR 2023, proposed as the baseline by [34]), StolenEncoder (CCS 2022) [26], and Cont-Steal (CVPR 2023) [34] as our baselines. A more detailed explanation of them is in Supplementary A.1. Specifically, we compare RDA with them to steal the encoder pre-trained on CIFAR10 under three different scenarios. For a fair comparison, the surrogate dataset remains identical across all baselines (except for some results presented in Table 5 due to different sizes of surrogate datasets).", + "bbox": [ + 212, + 513, + 787, + 621 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Under the Same Surrogate Dataset Size. In this subsection, we fix the surrogate dataset size to 2,500 and make it identical across all methods. As Table 4 shows, RDA outperforms baselines by a significant margin across all seven downstream datasets with a moderate query cost. We also evaluate RDA under a setting that has the same query cost as the conventional method and StolenEncoder (i.e., 2,500 queries), for which we query with each sample once in its origin version. The results show that RDA still surpasses baselines over five out of seven datasets with the least query cost. In addition, the comparison between the results achieved by RDA with 25,000 and 2,500 queries reveals that more patches for generating prototypes will enhance the attack. The embedding space of the surrogate encoder trained by each method is visualized in Figure 11 of Supplementary A.4, which shows that our RDA can train surrogate encoders to discriminate different samples better. Besides, We have summarized the mean", + "bbox": [ + 212, + 643, + 787, + 840 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "S. Wu et al.", + "bbox": [ + 271, + 114, + 352, + 126 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/68c1f19139c5858039488a50dbd8c8828ced00c55f88def557a15aa2502fb3e0.jpg", + "image_caption": [ + "Fig. 7: Heat map of the achieved SAs (better zoom in). This figure shows the performance of 100 combinations of different training epochs and (a) the surrogate dataset size, (b) the patch number for generating prototypes, and (c) the patch number for training." + ], + "image_footnote": [], + "bbox": [ + 218, + 145, + 379, + 276 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/27b122294343f329ad95e5a5f35b0d1b87729066e691de1baaf4ad42c9a4bbde.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 383, + 146, + 586, + 276 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/33897e5bb8869cd28a205b8bb29ff49b891802d28e985eb34f4d41ff5f411307.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 589, + 146, + 772, + 275 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "value achieved by each method over the seven downstream datasets in Figure 3 to show the superiority of RDA.", + "bbox": [ + 212, + 372, + 784, + 402 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Under the Same Query Cost. To demonstrate the superiority of RDA against baselines under an identical query cost, we consider two fixed values, i.e., 2,500 and 25,000. We take StolenEncoder as the rival since it is the best baseline with the least query cost. We note that Cont-Steal is beyond our consideration because the data volume for it would be too small under such a setting, which is only 25 images for 2,500 queries and 250 images for 25,000 queries. $N$ images $\\times n$ patches in the table indicates that we use $N$ images as the surrogate dataset and crop each image into $n$ patches to generate prototypes. Therefore, the query cost of RDA is $N \\times n$ . As shown in Table 5, under the same query cost of 2,500, RDA can achieve comparable results with StolenEncoder with half the size of its surrogate dataset. With the same 2,500 images, RDA outperforms StolenEncoder by a significant margin, showcasing its superiority. Moreover, under a query cost of 25,000, RDA outperforms StolenEncoder by $2.14\\%$ with only $10\\%$ the size of its surrogate dataset. Besides, comparisons between different configurations of RDA under queries of 2,500 indicate that the training data volume has a dominant impact on the performance.", + "bbox": [ + 212, + 424, + 787, + 667 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Under the Same Time Cost. Since each sample is also cropped and augmented into multiple patches for training (i.e., needing to forward encode each sample multiple times), RDA takes the longest time for training. Table 10 in Supplementary A.4 contains the detailed time cost of each method. To make a fair comparison, we prolong the training for another 200 epochs for baseline methods, i.e., 300 epochs in total, which is sufficient to lead their performances to saturate. The trained surrogate encoders are then evaluated on CIFAR10 as the downstream dataset. Figure 6 reveals that even though baseline methods take longer time to train, RDA still outperforms them with the least query cost, and the performance gap can be further widened by a moderate query cost.", + "bbox": [ + 212, + 688, + 787, + 840 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "Refine, Discriminate and Align (RDA)", + "bbox": [ + 473, + 114, + 730, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/25c6b349b74595b444bbfd8d7e4fe326e62b9c71dab073086cfae645c9b7a5b5.jpg", + "image_caption": [ + "Fig. 8: Ablation studies on the weight of each part in $\\mathcal{L}$ , with $\\lambda_1 = 1$ and $\\lambda_2$ varying." + ], + "image_footnote": [], + "bbox": [ + 266, + 147, + 419, + 252 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/ca8dd28c69d9034c5b3ab2d625676c610c1741385206b9bae714e3a184778157.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 429, + 148, + 576, + 251 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/38b90e2cc3ec07769caada336e93e464ec65ffaac00afc6fa3ee2f821392afeb.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 588, + 148, + 735, + 251 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/8848e4305a5d0169b444b11a18852de117433ed4eeb22b55ddd58ac5fb69607b.jpg", + "table_caption": [ + "Table 6: Ablation studies on loss functions. The presented results are SAs." + ], + "table_footnote": [], + "table_body": "
Loss TypeCIFAR10STL10CIFAR100MNISTF-MNISTGTSRBSVHNAVG
MSE65.6653.8331.0094.6386.1232.7962.9761.00
Cosine Similarity80.2366.5643.8197.0088.7655.5375.9472.54
KL Divergence77.3565.3040.9896.2387.8257.0469.3870.58
InfoNCE73.9961.6839.2495.7687.8658.8268.8869.43
LD75.5463.243.1696.5289.2765.7771.1572.23
LA79.7065.5943.7996.8989.2455.9676.7072.55
LA+LD79.3966.7644.2796.7489.3262.7573.7473.28
", + "bbox": [ + 299, + 316, + 699, + 401 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "4.4 Ablation Studies", + "text_level": 1, + "bbox": [ + 215, + 426, + 400, + 439 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "In this subsection, we conduct ablation studies to investigate the impact of the surrogate dataset size, the patch number for generating prototypes (i.e., $n$ in Eq. 1), the patch number for training (i.e., $m$ in Eq. 2, 3, 5, and 6), and loss functions. All the experiments are conducted to steal the encoder pre-trained on CIFAR10. For evaluations, the trained surrogate encoders are assessed on CIFAR10 in the first three ablation studies. While examining the influence of loss functions, we consider seven distinct downstream datasets, as shown in Table 4, to derive a more comprehensive conclusion. Other configurations remain at their default settings, and we present the resulting SAs.", + "bbox": [ + 212, + 449, + 787, + 585 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Surrogate Dataset Size. Recall our default setting where we randomly sample 2,500 images from Tiny ImageNet to construct the surrogate dataset, which is $5\\%$ the size of CIFAR-10. We vary the ratio from $1\\%$ to $10\\%$ to investigate its impact. Figure 7 (a) shows that a larger surrogate dataset will accelerate the training and improve the attack performance.", + "bbox": [ + 212, + 604, + 787, + 680 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Patch Number for Generating Prototypes. We vary the patch number for generating prototypes from 1 to 100 to investigate its impact. Figure 7 (b) shows that there is a tendency for more patches for generating prototypes will improve the attack performance and accelerate the convergence. However, when the patch number reaches 10, more patches do not further improve the attack performance due to the limited surrogate dataset size.", + "bbox": [ + 212, + 699, + 787, + 789 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Patch Number for Training. We vary the patch number for training from 1 to 10 to investigate its impact. Figure 7 (c) reveals a tendency for more patches", + "bbox": [ + 212, + 809, + 785, + 839 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "S. Wu et al.", + "bbox": [ + 271, + 114, + 352, + 126 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "for training will accelerate the training and achieve better results. However, more patches also indicate longer training time and more computation resources.", + "bbox": [ + 212, + 146, + 782, + 176 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Loss Functions. To demonstrate the superiority of the design of our loss function and the necessity of each part, we conduct ablation studies as shown in Table 6. Table 6 shows that $\\mathcal{L}_D$ largely surpass other loss functions on GTSRB while $\\mathcal{L}_A$ largely surpass other loss functions on SVHN. We hypothesize that $\\mathcal{L}_A$ attains optimal outcomes on SVHN due to the pronounced resemblance between SVHN and the surrogate dataset. It appears that $\\mathcal{L}_A$ inclines towards inducing the surrogate encoder to overfit the surrogate data, thereby elucidating its suboptimal performance on GTSRB, a dataset characterized by less similarity with the surrogate dataset. On the contrary, $\\mathcal{L}_D$ makes the surrogate encoder fit less to the surrogate dataset and has the best result on GTSRB. To further investigate the effect of each part in our loss, we have conducted an ablation experiment about the weight of each part in our loss. As Figure 8 shows, the surrogate encoder will perform better on SVHN while worse on GTSRB when the weight of $\\mathcal{L}_A$ increases. We refer to Supplementary A.4 for a more detailed analysis. Combining $\\mathcal{L}_D$ and $\\mathcal{L}_A$ , our loss is the most robust one, which achieves the best average result over the seven tested datasets. For example, although cosine similarity and $\\mathcal{L}_A$ performs well on SVHN (about $76\\%$ ), their performance on GTSRB is terrible (about $56\\%$ ). Our loss design can improve their performance on GTSRB by about $7\\%$ in the cost of less than $3\\%$ decrease on SVHN.", + "bbox": [ + 212, + 198, + 787, + 484 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "4.5 Robustness to Defenses", + "text_level": 1, + "bbox": [ + 214, + 506, + 455, + 520 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "In this section, we evaluate the robustness of RDA against three perturbation-based defenses that aim to limit information leakage and one watermarking defense that aims to detect whether a suspected encoder is stolen. The pretraining and downstream datasets both are set to CIFAR10.", + "bbox": [ + 212, + 531, + 782, + 590 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Perturbation-Based Defense. We evaluate three common practices of this type of defense, i.e., adding noise [30], top- $k$ [30] and rounding [44].", + "bbox": [ + 212, + 613, + 782, + 643 + ], + "page_idx": 12 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Adding noise: Adding noise means that the defender will introduce noise to the original outputs of the target encoder. Following [10], we set the mean of the noise to 0 and vary the standard deviations to control the noise level.", + "- Top- $k$ : Top- $k$ means that the defender will only output the first $k$ largest number of each embedding from the target encoder and set the rest as 0. We vary the value of $k$ to simulate different perturbation levels.", + "- Rounding: Rounding truncates each value in the embedding to a specific precision, which we vary to simulate different perturbation levels." + ], + "bbox": [ + 212, + 643, + 784, + 761 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "The results of the three perturbation-based defense methods are summarized in (a), (b), and (c) of Figure 9, respectively. We can observe in Figure 9 (a) that while adding noise can mitigate the model stealing attack, it decreases the utility of the target encoder more significantly. Moreover, Figure 9 (b) shows that though top- $k$ can decrease the efficacy of RDA as the defender lowers the", + "bbox": [ + 212, + 765, + 782, + 839 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "Refine, Discriminate and Align (RDA)", + "bbox": [ + 473, + 114, + 730, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/4835adab1adabc6cc1d857acb85922308e7178cc04cf586b498ef38f5396a1a6.jpg", + "image_caption": [ + "(a) Adding noise" + ], + "image_footnote": [], + "bbox": [ + 220, + 145, + 357, + 223 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/358420117ec56ecb2e34b4f230719cb3d958959c7017bdf62cadfec3efff4fbc.jpg", + "image_caption": [ + "(b) Top-k" + ], + "image_footnote": [], + "bbox": [ + 359, + 143, + 498, + 223 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/006437a89d4af20256750db59303260d6c8017deeda1896b6b9d2d638a9e45f5.jpg", + "image_caption": [ + "(c) Rounding", + "Fig. 9: The performance of different defense methods on CIFAR10." + ], + "image_footnote": [], + "bbox": [ + 501, + 143, + 638, + 223 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/6659c498ab0c0bba8ea463f127386d1f0f61792d668e6c9d4e21de0e37050eab.jpg", + "image_caption": [ + "(d) Watermarking" + ], + "image_footnote": [], + "bbox": [ + 643, + 148, + 781, + 223 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "value of $k$ , it also severely deteriorates the target encoder's performance. On the other hand, Figure 9 (c) reveals that rounding has a minimal effect on the performance of both the target encoder and attack. Our experimental results demonstrate the robustness of RDA against perturbation-based defenses.", + "bbox": [ + 212, + 294, + 784, + 354 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Watermarking Defense. As shown by Adi et al. [1], backdoors can be used as watermarks to claim the ownership of a model. In this sense, we follow [34] and leverage BadEncoder [18] to embed a backdoor into the target encoder as the watermark. Ideally, if the watermarked encoder is stolen, the surrogate encoder trained via stealing should also be triggered by the defender-specified trigger to exhibit certain behaviors, and thus the ownership can be claimed. Fortunately, Figure 9 (d) shows that RDA can steal the functionality of the target encoder while unlearning the watermark. More specifically, the watermark rate (WR) of the target encoder is $99.87\\%$ but only $17.04\\%$ for the surrogate encoder when the training converges. This observation indicates that the watermark embedded in the target encoder cannot be preserved by RDA.", + "bbox": [ + 212, + 377, + 787, + 544 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 566, + 359, + 583 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In this paper, we have proposed a novel model stealing method against SSL named RDA, which stands for refine, discriminate, and align. Compared with previous methods [26,34], RDA establishes a less biased optimization objective for each training sample and strives to extract more abundant functionality of the target encoder. This empowers RDA to exhibit tremendous effectiveness, efficiency, and robustness across diverse datasets and settings.", + "bbox": [ + 212, + 598, + 785, + 688 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Ethical Concerns and Possible Defenses. We underscore that the misuse of model stealing techniques can jeopardize the privacy and economic rights of embedding service providers. We hope that our work will inspire the development of more sophisticated defense mechanisms to thwart model stealing attacks. A possible defense involves rejection schemes for suspicious queries that appear highly similar, which we have more detailed discussions about in the Supplementary.", + "bbox": [ + 212, + 688, + 787, + 779 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Limitation and Future Work. The query cost of RDA is not yet optimal and can be further reduced by techniques like clustering to establish an identical prototype for multiple similar images. Moreover, more unique (e.g., more efficient) designs of RDA for transformer-based huge models remain largely unexplored.", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "S. Wu et al.", + "bbox": [ + 271, + 114, + 352, + 126 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgements. This work is supported by the National Key R&D Program of China under Grant 2022YFB3103500, the National Natural Science Foundation of China under Grant U20A20176 and 62072062, the National Natural Science Foundation of Chongqing under Grant cstc2022ycjh-bgzxm0031, and the Chongqing Research Program of Basic Research and Frontier Technology (Chongqing Talent) under Grant cstc2024ycjh-bgzxm0048.", + "bbox": [ + 212, + 146, + 787, + 238 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 215, + 258, + 325, + 275 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Adi, Y., Baum, C., Cisse, M., Pinkas, B., Keshet, J.: Turning your weakness into a strength: Watermarking deep neural networks by backdoorsing. In: 27th USENIX Security Symposium (USENIX Security 18). pp. 1615-1631 (2018)", + "2. Bossard, L., Guillaumeun, M., Van Gool, L.: Food-101-mining discriminative components with random forests. In: Computer vision-ECCV 2014: 13th European conference, zurich, Switzerland, September 6-12, 2014, proceedings, part VI 13. pp. 446-461. Springer (2014)", + "3. Carlini, N., Wagner, D.: Towards evaluating the robustness of neural networks. In: 2017 IEEE symposium on security and privacy (sp). pp. 39-57. IEEE (2017)", + "4. Caron, M., Touvron, H., Misra, I., Jégou, H., Mairal, J., Bojanowski, P., Joulin, A.: Emerging properties in self-supervised vision transformers. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 9650-9660 (2021)", + "5. Chen, T., Kornblith, S., Norouzi, M., Hinton, G.: A simple framework for contrastive learning of visual representations. In: International conference on machine learning. pp. 1597-1607. PMLR (2020)", + "6. Chen, X., Fan, H., Girshick, R., He, K.: Improved baselines with momentum contrastive learning. arXiv preprint arXiv:2003.04297 (2020)", + "7. Coates, A., Ng, A., Lee, H.: An analysis of single-layer networks in unsupervised feature learning. In: Proceedings of the fourteenth international conference on artificial intelligence and statistics. pp. 215-223. JMLR Workshop and Conference Proceedings (2011)", + "8. Dai, Y., Chen, Z., Li, J., Heinecke, S., Sun, L., Xu, R.: Tackling data heterogeneity in federated learning with class prototypes. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 37, pp. 7314-7322 (2023)", + "9. Devlin, J., Chang, M.W., Lee, K., Toutanova, K.: Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805 (2018)", + "10. Dziedzic, A., Dhawan, N., Kaleem, M.A., Guan, J., Papernot, N.: On the difficulty of defending self-supervised learning against model extraction. In: International Conference on Machine Learning. pp. 5757-5776. PMLR (2022)", + "11. Grill, J.B., Strub, F., Altché, F., Tallec, C., Richemond, P., Buchatskaya, E., Doersch, C., Avila Pires, B., Guo, Z., Gheshlaghi Azar, M., et al.: Bootstrap your own latent-a new approach to self-supervised learning. Advances in neural information processing systems 33, 21271-21284 (2020)", + "12. He, K., Fan, H., Wu, Y., Xie, S., Girshick, R.: Momentum contrast for unsupervised visual representation learning. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 9729-9738 (2020)", + "13. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 770-778 (2016)" + ], + "bbox": [ + 218, + 287, + 785, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "Refine, Discriminate and Align (RDA)", + "bbox": [ + 473, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 785, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "14. Howard, A.G., Zhu, M., Chen, B., Kalenichenko, D., Wang, W., Weyand, T., Andreetto, M., Adam, H.: Mobilenets: Efficient convolutional neural networks for mobile vision applications. arXiv preprint arXiv:1704.04861 (2017)", + "15. Huang, G., Liu, Z., Van Der Maaten, L., Weinberger, K.Q.: Densely connected convolutional networks. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 4700-4708 (2017)", + "16. Huang, W., Ye, M., Shi, Z., Li, H., Du, B.: Rethinking federated learning with domain shift: A prototype view. In: 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 16312-16322. IEEE (2023)", + "17. Jia, H., Choquette-Choo, C.A., Chandrasekaran, V., Papernot, N.: Entangled watermarks as a defense against model extraction. In: 30th USENIX Security Symposium (USENIX Security 21). pp. 1937-1954 (2021)", + "18. Jia, J., Liu, Y., Gong, N.Z.: Badencoder: Backdoor attacks to pre-trained encoders in self-supervised learning. In: 2022 IEEE Symposium on Security and Privacy (SP). pp. 2043-2059. IEEE (2022)", + "19. Khosla, P., Teterwak, P., Wang, C., Sarna, A., Tian, Y., Isola, P., Maschinot, A., Liu, C., Krishnan, D.: Supervised contrastive learning. Advances in neural information processing systems 33, 18661-18673 (2020)", + "20. Kingma, D.P., Ba, J.: Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980 (2014)", + "21. Krizhevsky, A., Hinton, G., et al.: Learning multiple layers of features from tiny images (2009)", + "22. Krizhevsky, A., Nair, V., Hinton, G.: Cifar-10 (canadian institute for advanced research). URL http://www.cs.toronto.edu/kriz/cifar.html 5(4), 1 (2010)", + "23. Le, Y., Yang, X.: Tiny imagenet visual recognition challenge. CS 231N 7(7), 3 (2015)", + "24. LeCun, Y., Cortes, C.: MNIST handwritten digit database (2010), http://yann.lecun.com/exdb/mnist/", + "25. Liu, H., Jia, J., Qu, W., Gong, N.Z.: Encodermi: Membership inference against pre-trained encoders in contrastive learning. In: Proceedings of the 2021 ACM SIGSAC Conference on Computer and Communications Security. pp. 2081-2095 (2021)", + "26. Liu, Y., Jia, J., Liu, H., Gong, N.Z.: Stolenencoder: stealing pre-trained encoders in self-supervised learning. In: Proceedings of the 2022 ACM SIGSAC Conference on Computer and Communications Security. pp. 2115-2128 (2022)", + "27. Van der Maaten, L., Hinton, G.: Visualizing data using t-sne. Journal of machine learning research 9(11) (2008)", + "28. Mettes, P., Van der Pol, E., Snoek, C.: Hyperspherical prototype networks. Advances in neural information processing systems 32 (2019)", + "29. Netzer, Y., Wang, T., Coates, A., Bissacco, A., Wu, B., Ng, A.Y.: Reading digits in natural images with unsupervised feature learning (2011)", + "30. Orekondy, T., Schiele, B., Fritz, M.: Knockoff nets: Stealing functionality of black-box models. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 4954-4963 (2019)", + "31. Papernot, N., McDaniel, P., Goodfellow, I., Jha, S., Celik, Z.B., Swami, A.: Practical black-box attacks against machine learning. In: Proceedings of the 2017 ACM on Asia conference on computer and communications security. pp. 506-519 (2017)", + "32. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021)" + ], + "bbox": [ + 215, + 146, + 785, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "S. Wu et al.", + "bbox": [ + 271, + 114, + 351, + 126 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "33. Salem, A., Zhang, Y., Humbert, M., Berrang, P., Fritz, M., Backes, M.: Ml-leaks: Model and data independent membership inference attacks and defenses on machine learning models. arXiv preprint arXiv:1806.01246 (2018)", + "34. Sha, Z., He, X., Yu, N., Backes, M., Zhang, Y.: Can't steal? cont-steal! contrastive stealing attacks against image encoders. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16373-16383 (2023)", + "35. Sharir, O., Peleg, B., Shoham, Y.: The cost of training nlp models: A concise overview. arXiv preprint arXiv:2004.08900 (2020)", + "36. Shokri, R., Stronati, M., Song, C., Shmatikov, V.: Membership inference attacks against machine learning models. In: 2017 IEEE symposium on security and privacy (SP). pp. 3-18. IEEE (2017)", + "37. Simonyan, K., Zisserman, A.: Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556 (2014)", + "38. Snell, J., Swersky, K., Zemel, R.: Prototypical networks for few-shot learning. Advances in neural information processing systems 30 (2017)", + "39. Stallkamp, J., Schlipsing, M., Salmen, J., Igel, C.: Man vs. computer: Benchmarking machine learning algorithms for traffic sign recognition. Neural networks 32, 323-332 (2012)", + "40. Tan, Y., Long, G., Liu, L., Zhou, T., Lu, Q., Jiang, J., Zhang, C.: Fedproto: Federated prototype learning across heterogeneous clients. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 36, pp. 8432-8440 (2022)", + "41. Tan, Y., Long, G., Ma, J., Liu, L., Zhou, T., Jiang, J.: Federated learning from pretrained models: A contrastive learning approach. Advances in Neural Information Processing Systems 35, 19332-19344 (2022)", + "42. Tian, Y., Wang, Y., Krishnan, D., Tenenbaum, J.B., Isola, P.: Rethinking few-shot image classification: a good embedding is all you need? In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XIV 16. pp. 266-282. Springer (2020)", + "43. Tong, S., Chen, Y., Ma, Y., Lecun, Y.: Emp-ssl: Towards self-supervised learning in one training epoch. arXiv preprint arXiv:2304.03977 (2023)", + "44. Tramér, F., Zhang, F., Juels, A., Reiter, M.K., Ristenpart, T.: Stealing machine learning models via prediction {APIs}. In: 25th USENIX security symposium (USENIX Security 16). pp. 601-618 (2016)", + "45. Wu, Z., Xiong, Y., Yu, S.X., Lin, D.: Unsupervised feature learning via nonparametric instance discrimination. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 3733-3742 (2018)", + "46. Xiao, H., Rasul, K., Vollgraf, R.: Fashion-mnist: a novel image dataset for benchmarking machine learning algorithms. arXiv preprint arXiv:1708.07747 (2017)" + ], + "bbox": [ + 212, + 146, + 787, + 674 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "Refine, Discriminate and Align (RDA)", + "bbox": [ + 473, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + } +] \ No newline at end of file diff --git a/2024/Refine, Discriminate and Align_ Stealing Encoders via Sample-Wise Prototypes and Multi-Relational Extraction/2d1e904f-54f5-49e9-a836-f9f59c16b169_model.json b/2024/Refine, Discriminate and Align_ Stealing Encoders via Sample-Wise Prototypes and Multi-Relational Extraction/2d1e904f-54f5-49e9-a836-f9f59c16b169_model.json new file mode 100644 index 0000000000000000000000000000000000000000..7cc329e37ac280561d2289f4a8f7b3c8ca5a2567 --- /dev/null +++ b/2024/Refine, Discriminate and Align_ Stealing Encoders via Sample-Wise Prototypes and Multi-Relational Extraction/2d1e904f-54f5-49e9-a836-f9f59c16b169_model.json @@ -0,0 +1,2445 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.235, + 0.141, + 0.773, + 0.208 + ], + "angle": 0, + "content": "Refine, Discriminate and Align: Stealing Encoders via Sample-Wise Prototypes and Multi-Relational Extraction" + }, + { + "type": "text", + "bbox": [ + 0.24, + 0.234, + 0.756, + 0.282 + ], + "angle": 0, + "content": "Shuchi \\(\\mathbf{W u}^{1}\\) Chuan Ma\\(^{2\\text{图}}\\) Kang Wei\\(^{3\\text{图}}\\) Xiaogang Xu\\(^{4}\\) Ming Ding\\(^{5}\\) Yuwen Qian\\(^{1}\\) Di Xiao\\(^{2}\\) Tao Xiang\\(^{2}\\) \n\\(^{1}\\) NJUST \\(^{2}\\) CQU \\(^{3}\\) PolyU \\(^{4}\\) CUHK \\(^{5}\\) Data61, CSIRO" + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.319, + 0.744, + 0.598 + ], + "angle": 0, + "content": "Abstract. This paper introduces RDA, a pioneering approach designed to address two primary deficiencies prevalent in previous endeavors aiming at stealing pre-trained encoders: (1) suboptimal performances attributed to biased optimization objectives, and (2) elevated query costs stemming from the end-to-end paradigm that necessitates querying the target encoder every epoch. Specifically, we initially Refine the representations of the target encoder for each training sample, thereby establishing a less biased optimization objective before the steal-training phase. This is accomplished via a sample-wise prototype, which consolidates the target encoder's representations for a given sample's various perspectives. Demanding exponentially fewer queries compared to the end-to-end approach, prototypes can be instantiated to guide subsequent query-free training. For more potent efficacy, we develop a multi-relational extraction loss that trains the surrogate encoder to Discriminate mismatched embedding-prototype pairs while Aligning those matched ones in terms of both amplitude and angle. In this way, the trained surrogate encoder achieves state-of-the-art results across the board in various downstream datasets with limited queries. Moreover, RDA is shown to be robust to multiple widely-used defenses. Our code is available at https://github.com/ShuchiWu/RDA." + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.611, + 0.713, + 0.625 + ], + "angle": 0, + "content": "Keywords: Model Stealing \\(\\cdot\\) Self-Supervised Learning \\(\\cdot\\) Prototype" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.675, + 0.377, + 0.691 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.707, + 0.788, + 0.814 + ], + "angle": 0, + "content": "Self-supervised learning (SSL) [4, 5, 9, 11, 12] is endowed with the capability to harness unlabeled data for pre-training a versatile encoder that applicable to a range of downstream tasks, or even showcasing groundbreaking zero-shot performances, e.g., CLIP [32]. However, SSL typically requires a large volume of data and computation resources to achieve a convincing performance, i.e., high-performance encoders are expensive to train [35]. To safeguard the confidentiality and economic worth of these encoders, entities like OpenAI offer their encoders" + }, + { + "type": "page_footnote", + "bbox": [ + 0.23, + 0.825, + 0.771, + 0.841 + ], + "angle": 0, + "content": "\\(\\text{忍}\\) Corresponding Author: chuan.ma@cqu.edu.cn, adam-kang.wei@polyu.edu.hk" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.353, + 0.127 + ], + "angle": 0, + "content": "S. Wu et al." + }, + { + "type": "image", + "bbox": [ + 0.29, + 0.145, + 0.382, + 0.254 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.396, + 0.147, + 0.495, + 0.252 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.506, + 0.147, + 0.603, + 0.252 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.612, + 0.146, + 0.714, + 0.252 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.262, + 0.788, + 0.318 + ], + "angle": 0, + "content": "Fig. 1: Illustrations of four stealing methods against SSL. The dotted arrows and text beside interpret how each method optimizes the surrogate encoder. Surrogate encoder branches in (b)-(d) involve data augmentations for training. Both (c) and (d) augment each sample before querying the target encoder but adopting different schemes." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.352, + 0.785, + 0.428 + ], + "angle": 0, + "content": "as a premium service, exclusively unveiling the service API to the public. Users have the privilege of soliciting embeddings for their data, facilitating the training of diverse models tailored for specific downstream tasks. Regrettably, the substantial value of these encoders and their exposure to publicly accessible APIs render them susceptible to model stealing attacks [10,26,34]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.43, + 0.788, + 0.55 + ], + "angle": 0, + "content": "Given a surrogate dataset, model stealing attacks aim to mimic the outputs of a target model to train either a high-accuracy copy of comparable performance or a high-fidelity copy that can serve as a stepping stone to perform further attacks like adversarial examples [31], membership inference attacks [33,36], etc. The goal of this paper is to develop an approach that can train a surrogate encoder with competitive performance on downstream tasks by only accessing the target encoder's output embeddings. Meanwhile, it should require as little cost as possible, which is primarily derived from querying the target encoder." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.552, + 0.788, + 0.793 + ], + "angle": 0, + "content": "We begin with systematically studying existing techniques for stealing pretrained encoders. Specifically, the conventional method [34] and StolenEncoder [26] are similar, and both only need to query the target encoder once with each sample before the training. The output embedding can be viewed as a \"ground truth\" or an \"optimization objective\" to the sample, and the surrogate encoder is optimized to output a similar embedding when the same sample is fed, as illustrated in Figure 1 (a). The sole distinction lies in the optimization of StolenEncoder, which incorporates data augmentations (refer to Figure 1 (b)), as it supposes the target encoder will produce similar embeddings for an image and its augmentations. Nevertheless, we contend that this presumed similarity is of a modest degree, given the disparity between the surrogate and pre-training data. The target encoder's representations for such data's various augmentations may diverge and even be biased to other samples, as visually demonstrated in Figure 2 using a toy experiment. In this sense, only using the embedding of an image's single perspective (regardless of the original or augmented version) as the optimization objective is inadvisable." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.787, + 0.84 + ], + "angle": 0, + "content": "On the other hand, Cont-Steal [34] augments each sample into two perspectives in each epoch: one is used to query the target encoder while the other is fed to the surrogate encoder. Two embeddings of the same sample from the tar" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.474, + 0.115, + 0.731, + 0.129 + ], + "angle": 0, + "content": "Refine, Discriminate and Align (RDA)" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "image", + "bbox": [ + 0.406, + 0.144, + 0.599, + 0.257 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.266, + 0.788, + 0.351 + ], + "angle": 0, + "content": "Fig. 2: t-SNE of embeddings belonging to five different images generated by an encoder pre-trained on CIFAR10, with each image augmented into 500 patches and fed into the encoder. Each black marker represents the mean of the 500 embeddings of a certain image, i.e., its prototype. Among the embeddings of an image's various augmentations, some can be diverged or even biased. In contrast, each image's prototype is more distinguishable, i.e., less biased." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.383, + 0.788, + 0.444 + ], + "angle": 0, + "content": "get and surrogate encoders will be aligned, while those of different samples will be pushed apart, as illustrated in Figure 1 (c). In spite of such an end-to-end training scheme performing better, it suffers from high query costs since each sample is augmented and used to query in each epoch." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.445, + 0.788, + 0.702 + ], + "angle": 0, + "content": "To tackle these issues, we propose RDA. Specifically, we first Refine the target encoder's representations for each sample in the surrogate dataset by averagely aggregating its several (e.g., 10) different augmentations' embeddings, i.e., their mean value, to establish a sample-level prototype for it. Then, the sample-wise prototype is used to guide the surrogate encoder optimization, which can mitigate the impact of biased embeddings, as shown in Figure 2 (see black markers). We have an experiment presented in Figure 10 of our supplementary material to further quantitatively reveal the benefit of prototypes, i.e., significantly more similar with each augmentation patch's embedding, showcasing it is less biased. With a prototype, there is a static optimization objective for each sample across the entire training, and thus, it is done in a query-free manner. This enables RDA to have far less query cost than the end-to-end approach (less than \\(10\\%\\)). To further enhance the attack efficacy with limited queries, we develop a multi-relational extraction loss that trains the surrogate encoder to Discriminate mismatched embedding-prototype pairs while Aligning those matched ones in terms of both amplitude and angle. The framework and pipeline of RDA are depicted in Figures 1 (d) and 4, respectively." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.703, + 0.788, + 0.824 + ], + "angle": 0, + "content": "Extensive experiments show that RDA acquires a favorable trade-off between the query cost (a.k.a, the money cost) and attack efficacy. As depicted in Figure 3, when the target encoder is pre-trained on CIFAR10, RDA averagely outperforms the state-of-the-art (SOTA), i.e., Cont-Steal [34], by \\(1.22\\%\\) on the stealing efficacy over seven different downstream datasets, with a competitive smaller query cost, i.e., only \\(1\\%\\) of Cont-Steal. Further, this performance gap can be widened to \\(5.20\\%\\) with \\(10\\%\\) query cost of Cont-Steal. In particular, RDA is demonstrated to be robust against multiple prevalent defenses [18,30,44]." + }, + { + "type": "text", + "bbox": [ + 0.239, + 0.825, + 0.585, + 0.84 + ], + "angle": 0, + "content": "In conclusion, our contributions are three folds:" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.353, + 0.127 + ], + "angle": 0, + "content": "S. Wu et al." + }, + { + "type": "image", + "bbox": [ + 0.356, + 0.148, + 0.644, + 0.263 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.276, + 0.788, + 0.334 + ], + "angle": 0, + "content": "Fig. 3: Performance comparisons between four stealing methods against SSL. The presented results are the mean values achieved by each method over seven different downstream classification tasks, with their corresponding query costs. Our proposed RDA can achieve SOTA results with the least query cost." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.363, + 0.784, + 0.405 + ], + "angle": 0, + "content": "- We comprehensively investigate two critical inadequacies of existing stealing methods against SSL, i.e., suboptimal efficacy and high query costs, and analyze the causes." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.408, + 0.784, + 0.435 + ], + "angle": 0, + "content": "- We develop a novel approach to train surrogate encoders, namely RDA, with sample-wise prototype guidance and a multi-relational extraction loss." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.439, + 0.784, + 0.465 + ], + "angle": 0, + "content": "- Extensive experiments are conducted to verify the effectiveness and robustness of RDA." + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.363, + 0.784, + 0.465 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.504, + 0.388, + 0.52 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.538, + 0.784, + 0.673 + ], + "angle": 0, + "content": "Prototype Learning. A prototype refers to the mean of embeddings belonging to all images of an identical class, which serves as a proxy of the class [40]. With each class having a prototype, the model is trained to match its output with the corresponding prototype when inputting a sample of a certain class. Prototype learning has proven beneficial for various learning scenarios, e.g., federated learning [8, 16, 41] and few-shot learning [28, 38, 42]. In this paper, the optimization objective (i.e., the mean of embeddings belonging to an identical sample's multiple augmentations) we generate for each training sample is conceptually similar to prototypes, and thus we name it a sample-wise prototype." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.784, + 0.763 + ], + "angle": 0, + "content": "Model Stealing Attacks against SSL. Dziedzic et al. [10] pointed out that the higher dimension of embeddings leaks more information than labels, making SSL more easily stolen. The primary objective of a stealing attack is to train a surrogate encoder to achieve high accuracy on downstream tasks or recreate a high-fidelity copy that can be used to mount further attacks such as adversarial examples [3] and membership inference attacks [25]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.811, + 0.84 + ], + "angle": 0, + "content": "Defenses against Model Stealing Attacks. Existing defenses against model stealing attacks can be categorized based on when they are applied [10]. Perturbation-based defenses, e.g., adding noise [30], top-\\(k\\) [30] and truncating outputs [44], are applied before a stealing attack happens, aiming to limit the information leakage. On the other hand, watermarking defenses embed watermarks or unique" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.474, + 0.115, + 0.731, + 0.129 + ], + "angle": 0, + "content": "Refine, Discriminate and Align (RDA)" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.178 + ], + "angle": 0, + "content": "identifiers into the target model and use them to detect whether it is stolen after a stealing attack happens [17]." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.2, + 0.381, + 0.218 + ], + "angle": 0, + "content": "3 Methodology" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.23, + 0.379, + 0.244 + ], + "angle": 0, + "content": "3.1 Threat Model" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.255, + 0.788, + 0.377 + ], + "angle": 0, + "content": "Given a target encoder \\( E_{T} \\), the attacker aims to train a surrogate encoder \\( E_{S} \\) at the lowest possible cost, which can perform competitively on downstream tasks with \\( E_{T} \\). To achieve this goal, the attacker queries \\( E_{T} \\) for embeddings of an unlabeled surrogated dataset \\( D_{S} = \\{\\pmb{x}_{1},\\dots,\\pmb{x}_{N}\\} \\) consisting of \\( N \\) images, and guide \\( E_{S} \\) to mimic the output of \\( E_{T} \\) for each sample in \\( D_{S} \\). Specifically, we consider a black-box setting, where the attacker has direct access only to the outputs of \\( E_{T} \\) while remaining unaware of its architecture and training configurations, including pre-training datasets, loss functions, data augmentations schemes, etc." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.397, + 0.467, + 0.412 + ], + "angle": 0, + "content": "3.2 Sample-Wise Prototypes" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.42, + 0.789, + 0.581 + ], + "angle": 0, + "content": "Recalling the discussion in Section 1, we expect to refine the target encoder's representations to find a less biased optimization objective for each sample in \\( D_{S} \\). Enlightened by prototype learning [40] and Extreme-Multi-Patch-SSL (EMPSL) [43], which establish a prototype/benchmark for each class/sample, we propose a sample-wise prototype generation method for this purpose. In detail, the process of prototype generation can be divided into three steps as follows: 1 Cropping and augmenting each image \\( \\pmb{x}_i\\in D_S \\) into \\( n \\) augmentation patches denoted as \\( \\{\\pmb{x}_{i,t,c}^{\\prime}\\}_{c = 1}^{n} = \\{\\pmb{x}_{i,t,1}^{\\prime},\\dots,\\pmb{x}_{i,t,n}^{\\prime}\\} \\); 2 Querying \\( E_{T} \\) with each augmentation patch in \\( \\{\\pmb{x}_{i,t,c}^{\\prime}\\}_{c = 1}^{n} \\), resulting in a set of embeddings denoted as \\( \\{E_T(\\pmb{x}_{i,t,c}^{\\prime})\\}_{c = 1}^{n} = \\{E_T(\\pmb{x}_{i,t,1}^{\\prime}),\\dots,E_T(\\pmb{x}_{i,t,n}^{\\prime})\\} \\); 3 Calculating the mean of \\( \\{E_T(\\pmb{x}_{i,t,c}^{\\prime})\\}_{c = 1}^{n} \\) as the prototype \\( p_{\\pmb{x}_i} \\) for \\( \\pmb{x}_i \\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.388, + 0.589, + 0.785, + 0.605 + ], + "angle": 0, + "content": "\\[\np _ {\\boldsymbol {x} _ {i}} = \\frac {1}{n} \\sum_ {c = 1} ^ {n} E _ {T} \\left(\\boldsymbol {x} _ {i, t, c} ^ {\\prime}\\right), \\quad \\boldsymbol {x} _ {i} \\in D _ {s}. \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.614, + 0.788, + 0.721 + ], + "angle": 0, + "content": "Each generated prototype will be stored in a memory bank. These sample-wise prototypes can provide a stable optimization objective for each sample throughout the training process. Consequently, the attacker can accomplish the training in a query-free manner. Notably, setting \\( n \\) to a value that is far less than the training epochs, e.g., 10 vs. 100, is sufficient to yield a favorable performance. The prototype generation process is illustrated in step ① of Figure 4. Supplementary A.2 explains how we build the memory bank in practice." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.741, + 0.535, + 0.755 + ], + "angle": 0, + "content": "3.3 Multi-Relational Extraction Loss" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.788, + 0.843 + ], + "angle": 0, + "content": "As step ② of Figure 4 shows, during training \\( E_{S} \\), we also perform cropping and augmentation on each image \\( \\pmb{x}_i \\in D_S \\) to create \\( m \\) patches, forming an augmentation patch set denoted as \\( \\{\\pmb{x}_{i,s,q}^{\\prime}\\}_{q = 1}^{m} = \\{\\pmb{x}_{i,s,1}^{\\prime},\\dots,\\pmb{x}_{i,s,m}^{\\prime}\\} \\). The \\( m \\) does not necessarily need to be set equal to \\( n \\). Next, we feed each augmentation patch from \\( \\{\\pmb{x}_{i,s,q}^{\\prime}\\}_{q = 1}^{m} \\) to \\( E_{S} \\), obtaining a set of embeddings denoted as" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.353, + 0.127 + ], + "angle": 0, + "content": "S. Wu et al." + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.148, + 0.784, + 0.243 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.256, + 0.788, + 0.34 + ], + "angle": 0, + "content": "Fig. 4: Pipeline of RDA. Prototype generation: augment one sample into \\( n \\) patches and use them to query the target encoder \\( (E_T) \\). The mean of the \\( n \\) patches's embeddings is defined as a prototype for this sample. Forward encoding: crop one image into \\( m \\) patches and feed them to the surrogate encoder \\( (E_S) \\) for their embeddings. Optimization: align embeddings from the surrogate encoder to their matched prototypes in both angle and amplitude while pushing away those belonging to different samples." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.367, + 0.789, + 0.413 + ], + "angle": 0, + "content": "\\(\\{E_S(\\pmb{x}_{i,s,q}^{\\prime})\\}_{q = 1}^{m} = \\{E_S(\\pmb{x}_{i,s,1}^{\\prime}),\\dots ,E_S(\\pmb{x}_{i,s,m}^{\\prime})\\}\\). To optimize \\(E_{S}\\), we develop a multi-relational extraction loss composed of two parts, i.e., the discriminating loss and the aligning loss." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.412, + 0.789, + 0.569 + ], + "angle": 0, + "content": "Discriminating Loss. The discriminating loss is responsible for training \\( E_{S} \\) to distinguish between different samples, as depicted in step ③ of Figure 4. For this goal, it pushes each embedding in \\( \\{E_S(\\pmb{x}_{i,s,q}')\\}_{q = 1}^m \\) away from mismatched prototypes \\( p_{\\pmb{x}_j}, i \\neq j \\) (denoted as negative pairs). In this regard, contrastive learning [5,6,12,19] offers a solution, as it can pull embeddings of positive pairs close while pushing embeddings of negative pairs apart. Furthermore, since more negative pairs can help improve the performance of contrastive learning [5], we refer to the loss designed by Sha et al. [34] to propose our discriminating loss \\( \\mathcal{L}_D \\). In particular, \\( \\mathcal{L}_D \\) considers both mismatched prototype-embedding pairs from different samples as well as embeddings from \\( E_{S} \\) for different samples as negative pairs. Formally, \\( \\mathcal{L}_D \\) can be expressed as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.342, + 0.575, + 0.787, + 0.593 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {p o s} \\left(\\boldsymbol {x} _ {i}\\right) = \\frac {1}{m} \\sum_ {q = 1} ^ {m} \\exp \\left(\\operatorname {s i m} \\left(E _ {S} \\left(\\boldsymbol {x} _ {i, s, q} ^ {\\prime}\\right), p _ {\\boldsymbol {x} _ {i}}\\right) / \\tau\\right), \\tag {2}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.342, + 0.61, + 0.786, + 0.658 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {n e g} \\left(\\boldsymbol {x} _ {i}\\right) = \\frac {1}{m} \\sum_ {q = 1} ^ {m} \\sum_ {k = 1} ^ {N} \\mathbb {1} _ {[ i \\neq k ]} \\left(\\exp \\left(s i m \\left(E _ {S} \\left(\\boldsymbol {x} _ {i, s, q} ^ {\\prime}\\right), p _ {\\boldsymbol {x} _ {k}}\\right) / \\tau\\right) \\right. \\tag {3} \\\\ + \\exp \\left(\\operatorname {s i m} \\left(E _ {S} \\left(\\boldsymbol {x} _ {i, s, q} ^ {\\prime}\\right), E _ {S} \\left(\\boldsymbol {x} _ {k, s, q} ^ {\\prime}\\right)\\right) / \\tau\\right)), \\\\ \\end{array}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.414, + 0.664, + 0.786, + 0.684 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {D} = - \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\log \\frac {\\mathcal {L} _ {p o s} (\\boldsymbol {x} _ {i})}{\\mathcal {L} _ {n e g} (\\boldsymbol {x} _ {i})}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.686, + 0.788, + 0.715 + ], + "angle": 0, + "content": "where \\( \\operatorname{sim}(u, v) \\) represents the cosine similarity between \\( u \\) and \\( v \\), and \\( \\tau \\) is a temperature parameter." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.715, + 0.788, + 0.824 + ], + "angle": 0, + "content": "Aligning Loss. While \\(\\mathcal{L}_D\\) can separate different samples effectively, it does not fully leverage the potential of prototypes. To further enhance the attack efficacy, we propose to align embeddings from \\(E_S\\) more thoroughly with their matched prototypes from the high-performance \\(E_T\\), i.e., in terms of both amplitude and angle, as step ③ of Figure 4 shows. To measure the amplitude and angle deviations between two embeddings, we employ the mean square error (MSE) and cosine similarity to quantify them respectively, which can be formulated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.371, + 0.825, + 0.786, + 0.843 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {a m p} ^ {\\prime} \\left(\\boldsymbol {x} _ {i}\\right) = \\frac {1}{m} \\sum_ {q = 1} ^ {m} \\left\\| E _ {S} \\left(\\boldsymbol {x} _ {i, s, q} ^ {\\prime}\\right) - p _ {\\boldsymbol {x} _ {i}} \\right\\| ^ {2}, \\tag {5}\n\\]" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.474, + 0.115, + 0.731, + 0.129 + ], + "angle": 0, + "content": "Refine, Discriminate and Align (RDA)" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.126 + ], + "angle": 0, + "content": "7" + }, + { + "type": "equation", + "bbox": [ + 0.359, + 0.161, + 0.786, + 0.179 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {a n g}} ^ {\\prime} \\left(\\boldsymbol {x} _ {i}\\right) = \\frac {1}{m} \\sum_ {q = 1} ^ {m} \\operatorname {s i m} \\left(E _ {S} \\left(\\boldsymbol {x} _ {i, s, q} ^ {\\prime}\\right), p _ {\\boldsymbol {x} _ {i}}\\right). \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.181, + 0.788, + 0.304 + ], + "angle": 0, + "content": "Notice that we will normalize the value of \\(\\mathcal{L}_{ang}^{\\prime}\\) to [0, 1]. However, there is an issue concerning the equal penalty given by \\(\\mathcal{L}_{amp}^{\\prime}\\) and \\(\\mathcal{L}_{ang}^{\\prime}\\) to every deviation increase. For instance, when the MSE increases from 0.8 to 0.9 or from 0.3 to 0.4, both penalties given by \\(\\mathcal{L}_{amp}^{\\prime}\\) are 0.1, which is imprudent. On the contrary, we should have a more rigorous penalty for the increase in MSE from 0.3 to 0.4, as a smaller MSE is more desirable. Likewise, the same penalizing regime should be applied on \\(1 / \\mathcal{L}_{ang}^{\\prime}\\) for the same reason. To this end, we adopt a logarithmic function to redefine the ultimate formulations of \\(\\mathcal{L}_{amp}\\) and \\(\\mathcal{L}_{ang}\\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.417, + 0.316, + 0.786, + 0.332 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {a m p}} \\left(\\boldsymbol {x} _ {i}\\right) = \\log \\mathcal {L} _ {\\text {a m p}} ^ {\\prime} \\left(\\boldsymbol {x} _ {i}\\right), \\tag {7}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.359, + 0.36, + 0.785, + 0.375 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {a n g} \\left(\\boldsymbol {x} _ {i}\\right) = \\log \\left(1 / \\mathcal {L} _ {a n g} ^ {\\prime} \\left(\\boldsymbol {x} _ {i}\\right)\\right) = - \\log \\mathcal {L} _ {a n g} ^ {\\prime} \\left(\\boldsymbol {x} _ {i}\\right). \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.38, + 0.784, + 0.409 + ], + "angle": 0, + "content": "We posit the amplitude and angle deviations hold equal importance, and formulate the aligning loss as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.318, + 0.42, + 0.786, + 0.452 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {A} = \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\left(\\mathcal {L} _ {a m p} (\\boldsymbol {x} _ {i}) + \\mathcal {L} _ {a n g} (\\boldsymbol {x} _ {i})\\right) = - \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\log \\frac {\\mathcal {L} _ {a n g} ^ {\\prime} (\\boldsymbol {x} _ {i})}{\\mathcal {L} _ {a m p} ^ {\\prime} (\\boldsymbol {x} _ {i})}. \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.46, + 0.784, + 0.491 + ], + "angle": 0, + "content": "Our experiments in Supplementary A.4 demonstrate the combination of \\(\\mathcal{L}_{amp}\\) and \\(\\mathcal{L}_{ang}\\) offers superior results than using each solely." + }, + { + "type": "text", + "bbox": [ + 0.239, + 0.491, + 0.744, + 0.506 + ], + "angle": 0, + "content": "Lastly, the loss function for stealing pre-trained encoders is as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.431, + 0.52, + 0.785, + 0.535 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\lambda_ {1} \\cdot \\mathcal {L} _ {D} + \\lambda_ {2} \\cdot \\mathcal {L} _ {A}, \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.546, + 0.784, + 0.606 + ], + "angle": 0, + "content": "where \\(\\lambda_{1}\\) and \\(\\lambda_{2}\\) are preset coefficients to adjust the weight of each part in \\(\\mathcal{L}\\). To show the superiority of our design, especially on the loss function, we have conducted ablation studies in Section 4.4 and explored several alternative designs in Supplementary A.3." + }, + { + "type": "text", + "bbox": [ + 0.239, + 0.607, + 0.784, + 0.622 + ], + "angle": 0, + "content": "Detailed steps of RDA are summarized in Algorithm 1 of Supplementary A.2." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.646, + 0.375, + 0.663 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.678, + 0.427, + 0.694 + ], + "angle": 0, + "content": "4.1 Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.704, + 0.785, + 0.81 + ], + "angle": 0, + "content": "Target Encoder Settings. We use SimCLR [5] to pre-train two ResNet18 [13] encoders on CIFAR10 [21] and STL10 [7], respectively, as two medium-scale target encoders. Furthermore, we consider two real-world large-scale ResNet50 encoders as the targets, i.e., the ImageNet encoder pre-trained by Google [5], and the CLIP encoder pre-trained by OpenAI [32]. Besides the ResNet family, RDA also has been demonstrated effective upon other backbones, i.e., VGG19_bn [37], DenseNet121 [15], and MobileNetV2 [14], in Supplementary A.4." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.81, + 0.786, + 0.84 + ], + "angle": 0, + "content": "Attack Settings. Regarding the surrogate dataset, it is derived from Tiny ImageNet [23]. Specifically, we randomly sample 2,500 images and resize them to" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.353, + 0.127 + ], + "angle": 0, + "content": "S. Wu et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.241, + 0.145, + 0.761, + 0.173 + ], + "angle": 0, + "content": "Table 1: Results of RDA against two Table 2: Results of RDA against the medium-scale encoders. ImageNet Encoder and CLIP." + }, + { + "type": "table", + "bbox": [ + 0.256, + 0.185, + 0.489, + 0.289 + ], + "angle": 0, + "content": "
Pre-training DatasetDownstream DatasetTASA\\( \\frac{\\mathrm{{SA}}}{\\mathrm{{TA}}} \\times {100}\\% \\)
CIFAR10MNIST97.2796.6299.33
F-MNIST88.5889.32100.84
GTSRB61.7662.75101.60
SVHN73.7873.7499.95
STL10MNIST96.8496.3199.45
F-MNIST90.0887.9197.59
GTSRB64.4557.7089.53
SVHN61.8570.67114.26
" + }, + { + "type": "table", + "bbox": [ + 0.521, + 0.186, + 0.742, + 0.289 + ], + "angle": 0, + "content": "
TargetEncoderDownstreamDatasetTASA\\( \\frac{\\mathrm{{SA}}}{\\mathrm{{TA}}} \\times {100}\\% \\)
ImageNetEncoderMNIST97.3895.8998.47
F-MNIST91.6790.8399.08
GTSRB63.1459.6294.43
SVHN73.2069.2194.55
CLIPMNIST97.9093.9695.98
F-MNIST89.9287.4397.23
GTSRB68.6155.8481.39
SVHN69.5357.5882.81
" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.3, + 0.785, + 0.328 + ], + "angle": 0, + "content": "Table 3: Computation resources required by pre-training the two targeted real-world encoders from scratch and RDA to steal them." + }, + { + "type": "table", + "bbox": [ + 0.372, + 0.341, + 0.632, + 0.392 + ], + "angle": 0, + "content": "
TargetEncoderPre-trainingRDA
HardwareTime (hrs)HardwareTime (hrs)
ImageNetEncoderTPU v3192RTX A500010.8
CLIPV100 GPU255,74416.2
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.418, + 0.785, + 0.524 + ], + "angle": 0, + "content": "\\(32 \\times 32\\) for stealing encoders pre-trained on CIFAR10 and STL10. When stealing the ImageNet encoder and CLIP, the number is 40,000 and 60,000, respectively, with each image resized to \\(224 \\times 224\\). For the surrogate encoder architecture, we adopt a ResNet18 across our experiments. During the attack, we set \\(n\\) in Eq. 1 as 10, \\(m\\) in Eq. 2, 3, 5, and 6 as 5, \\(\\tau\\) in Eq. 2 and 3 as 0.07, and both \\(\\lambda_{1}\\) and \\(\\lambda_{2}\\) in Eq. 10 as 1 unless stated otherwise. For training, the batch size is set as 100, and we employ an Adam optimizer [20] with a learning rate of 0.001." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.525, + 0.787, + 0.693 + ], + "angle": 0, + "content": "Evaluation Settings. We train each surrogate encoder for 100 epochs and test its KNN accuracy [45] on CIFAR10 after each epoch. The best-trained surrogate encoders, as well as target encoders, will be used to train downstream classifiers for the linear probing evaluation. Each downstream classifier will be trained for 100 epochs with an Adam optimizer and a learning rate of 0.0001. Specifically, we totally consider seven downstream datasets, namely MNIST [24], CIFAR10 [21], STL10 [7], GTSRB [39], CIFAR100 [22], SVHN [29], and F-MNIST [46]. Additional small-scale experiments on more complex datasets (e.g., Food 101 [2]) are included in our supplementary material. As done in [26], we use Target Accuracy (TA) to evaluate target encoders, Steal Accuracy (SA) to evaluate surrogate encoders, and \\(\\frac{\\mathrm{SA}}{\\mathrm{TA}} \\times 100\\%\\) to evaluate the efficacy of the stealing attack." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.711, + 0.436, + 0.725 + ], + "angle": 0, + "content": "4.2 Effectiveness of RDA" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.734, + 0.788, + 0.844 + ], + "angle": 0, + "content": "Stealing Medium-Scale Encoders. Table 1 shows the results achieved by RDA upon two medium-scale target encoders pre-trained on CIFAR10 and STL10, respectively. As the results show, using a surrogate dataset of \\(5\\%\\) the size of CIFAR10 (2,500 / 50,000) and \\(2.4\\%\\) the size of STL10 (2,500 / 105,000), RDA can steal nearly \\(100\\%\\) of the functionality of both target encoders, except for the GTSRB scenario when stealing the encoder pre-trained on STL10. Nonetheless, even in this exceptional case, the obtained \\(\\frac{\\mathrm{SA}}{\\mathrm{TA}} \\times 100\\%\\) is approximately \\(90\\%\\)," + } + ], + [ + { + "type": "header", + "bbox": [ + 0.474, + 0.115, + 0.731, + 0.129 + ], + "angle": 0, + "content": "Refine, Discriminate and Align (RDA)" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "image", + "bbox": [ + 0.227, + 0.147, + 0.332, + 0.22 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.225, + 0.226, + 0.336, + 0.237 + ], + "angle": 0, + "content": "(a) ImageNet Encoder" + }, + { + "type": "image", + "bbox": [ + 0.339, + 0.147, + 0.444, + 0.22 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.371, + 0.228, + 0.412, + 0.239 + ], + "angle": 0, + "content": "(b) CLIP" + }, + { + "type": "image_caption", + "bbox": [ + 0.218, + 0.25, + 0.452, + 0.305 + ], + "angle": 0, + "content": "Fig.5: t-SNE of embeddings of 2,000 images sampled from CIFAR10 generated by the ImageNet encoder and CLIP." + }, + { + "type": "image", + "bbox": [ + 0.507, + 0.144, + 0.728, + 0.27 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.451, + 0.28, + 0.784, + 0.309 + ], + "angle": 0, + "content": "Fig. 6: The training epochs of RDA are fixed to 100 while those of baselines are prolonged to 300." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.337, + 0.785, + 0.412 + ], + "angle": 0, + "content": "which exemplifies the remarkable effectiveness of RDA. Furthermore, the surrogate encoder trained through RDA even outperforms the target encoder on multiple datasets. Our interpretation of this is that the target encoder fits the pre-training dataset more strongly compared to the RDA-trained surrogate encoder, thus exhibiting inferior performance on some out-of-distribution datasets." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.433, + 0.788, + 0.844 + ], + "angle": 0, + "content": "Stealing Real-World Large-Scale Encoders. We further evaluate RDA on two real-world pre-trained encoders, i.e., the ImageNet encoder and CLIP. We aim to demonstrate the effectiveness of RDA upon such large-scale encoders trained with a massive amount of data. Specifically, the ImageNet encoder is pre-trained on the ImageNet dataset, which has 1.3 million images. The CLIP encoder is pre-trained on a web-scale dataset that consists of 400 million image-text pairs. During the attack, we set the patch number for training as 1 (i.e., \\( m \\) in Eq. 2, 3, 5, and 6) to save time. Other parameters follow the default setting except the \\( \\lambda_{1} \\) and \\( \\lambda_{2} \\) when stealing CLIP, in which we set \\( \\lambda_{1} \\) and \\( \\lambda_{2} \\) to 1 and 20, respectively. To interpret the reason for this adjustment, we randomly sample 200 images from each class in the testing set of CIFAR10 and visualize the output embeddings from the ImageNet encoder and CLIP using t-SNE [27]. Figure 5 shows that embeddings from the ImageNet encoder are more uniformly distributed compared to those from CLIP. This is because CLIP is a multimodal encoder, which is pre-trained by conducting contrastive learning between image-text pairs. Therefore, the text also shares part of the embedding space and is mutually exclusive with mismatched images. Although \\( \\mathcal{L}_D \\) can push embeddings of different samples apart, a significant weight of it will make all embeddings more uniformly distributed, i.e., not align with CLIP. Therefore, we set a smaller weight to \\( \\mathcal{L}_D \\) when stealing CLIP. This adjustment is practical since the attacker has access to the outputs of the target encoder, and thus can observe its embedding space to adjust its attack settings. Table 2 shows that using a surrogate dataset of only \\( 3.07\\% \\) the size of ImageNet (40K / 1.3M) and \\( 0.015\\% \\) the size of the training data of CLIP (60K / 400M), RDA can achieve comparable performances with the two encoders across various downstream tasks. The results also reveal that even under the architecture of the surrogate and target encoders being distinct (ResNet18 vs. ResNet50), RDA still exhibits high effec" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.353, + 0.127 + ], + "angle": 0, + "content": "S. Wu et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.788, + 0.192 + ], + "angle": 0, + "content": "Table 4: Comparisons between RDA and baselines to steal the encoder pre-trained on CIFAR10 under the same surrogate dataset size. We report the \\(\\frac{\\mathrm{SA}}{\\mathrm{TA}} \\times 100\\%\\) achieved by each method and its query cost. The Optimal and suboptimal results are highlighted." + }, + { + "type": "table", + "bbox": [ + 0.299, + 0.204, + 0.7, + 0.274 + ], + "angle": 0, + "content": "
MethodCIFAR10CIFAR100MNISTGTSRBSVHNSTL10F-MNISTQueries
Conventional [34]63.3645.4096.1620.3468.2660.2594.522,500
StolenEncoder [26]79.0368.7898.7663.7092.0176.2995.912,500
Cont-Steal [34]80.4373.9498.0395.5596.3980.0895.39250,000
RDA83.8477.4998.9794.2796.7982.2594.692,500
85.8583.0899.33101.6099.9585.56100.8425,000
" + }, + { + "type": "table_caption", + "bbox": [ + 0.218, + 0.287, + 0.784, + 0.302 + ], + "angle": 0, + "content": "Table 5: Comparisons between RDA and StolenEncoder under the same query cost." + }, + { + "type": "table", + "bbox": [ + 0.382, + 0.314, + 0.623, + 0.393 + ], + "angle": 0, + "content": "
QueriesMethodsettingSA
2500StolenEncoder2,500 images71.98
RDA500 images×5 patches65.38
625 images×4 patches67.07
1,250 images×2 patches71.44
2,500 images×1 patch76.36
25,000StolenEncoder25,000 images76.23
RDA2,500 images×10 patches78.50
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.422, + 0.785, + 0.468 + ], + "angle": 0, + "content": "tiveness. Moreover, Table 3 shows that RDA requires much fewer computation resources to steal the two encoders than pre-training them from scratch, owing to the small surrogate dataset and lightweight network architecture." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.49, + 0.561, + 0.506 + ], + "angle": 0, + "content": "4.3 Comparison with Existing Methods" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.515, + 0.788, + 0.622 + ], + "angle": 0, + "content": "We consider the conventional method (CVPR 2023, proposed as the baseline by [34]), StolenEncoder (CCS 2022) [26], and Cont-Steal (CVPR 2023) [34] as our baselines. A more detailed explanation of them is in Supplementary A.1. Specifically, we compare RDA with them to steal the encoder pre-trained on CIFAR10 under three different scenarios. For a fair comparison, the surrogate dataset remains identical across all baselines (except for some results presented in Table 5 due to different sizes of surrogate datasets)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.644, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Under the Same Surrogate Dataset Size. In this subsection, we fix the surrogate dataset size to 2,500 and make it identical across all methods. As Table 4 shows, RDA outperforms baselines by a significant margin across all seven downstream datasets with a moderate query cost. We also evaluate RDA under a setting that has the same query cost as the conventional method and StolenEncoder (i.e., 2,500 queries), for which we query with each sample once in its origin version. The results show that RDA still surpasses baselines over five out of seven datasets with the least query cost. In addition, the comparison between the results achieved by RDA with 25,000 and 2,500 queries reveals that more patches for generating prototypes will enhance the attack. The embedding space of the surrogate encoder trained by each method is visualized in Figure 11 of Supplementary A.4, which shows that our RDA can train surrogate encoders to discriminate different samples better. Besides, We have summarized the mean" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.474, + 0.115, + 0.731, + 0.129 + ], + "angle": 0, + "content": "Refine, Discriminate and Align (RDA)" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "image", + "bbox": [ + 0.219, + 0.146, + 0.38, + 0.277 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.384, + 0.147, + 0.588, + 0.277 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.591, + 0.147, + 0.774, + 0.276 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.287, + 0.785, + 0.343 + ], + "angle": 0, + "content": "Fig. 7: Heat map of the achieved SAs (better zoom in). This figure shows the performance of 100 combinations of different training epochs and (a) the surrogate dataset size, (b) the patch number for generating prototypes, and (c) the patch number for training." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.373, + 0.785, + 0.403 + ], + "angle": 0, + "content": "value achieved by each method over the seven downstream datasets in Figure 3 to show the superiority of RDA." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.425, + 0.788, + 0.668 + ], + "angle": 0, + "content": "Under the Same Query Cost. To demonstrate the superiority of RDA against baselines under an identical query cost, we consider two fixed values, i.e., 2,500 and 25,000. We take StolenEncoder as the rival since it is the best baseline with the least query cost. We note that Cont-Steal is beyond our consideration because the data volume for it would be too small under such a setting, which is only 25 images for 2,500 queries and 250 images for 25,000 queries. \\(N\\) images \\(\\times n\\) patches in the table indicates that we use \\(N\\) images as the surrogate dataset and crop each image into \\(n\\) patches to generate prototypes. Therefore, the query cost of RDA is \\(N \\times n\\). As shown in Table 5, under the same query cost of 2,500, RDA can achieve comparable results with StolenEncoder with half the size of its surrogate dataset. With the same 2,500 images, RDA outperforms StolenEncoder by a significant margin, showcasing its superiority. Moreover, under a query cost of 25,000, RDA outperforms StolenEncoder by \\(2.14\\%\\) with only \\(10\\%\\) the size of its surrogate dataset. Besides, comparisons between different configurations of RDA under queries of 2,500 indicate that the training data volume has a dominant impact on the performance." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Under the Same Time Cost. Since each sample is also cropped and augmented into multiple patches for training (i.e., needing to forward encode each sample multiple times), RDA takes the longest time for training. Table 10 in Supplementary A.4 contains the detailed time cost of each method. To make a fair comparison, we prolong the training for another 200 epochs for baseline methods, i.e., 300 epochs in total, which is sufficient to lead their performances to saturate. The trained surrogate encoders are then evaluated on CIFAR10 as the downstream dataset. Figure 6 reveals that even though baseline methods take longer time to train, RDA still outperforms them with the least query cost, and the performance gap can be further widened by a moderate query cost." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.353, + 0.127 + ], + "angle": 0, + "content": "S. Wu et al." + }, + { + "type": "image", + "bbox": [ + 0.267, + 0.148, + 0.421, + 0.253 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.43, + 0.149, + 0.578, + 0.252 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.589, + 0.15, + 0.736, + 0.252 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.263, + 0.784, + 0.278 + ], + "angle": 0, + "content": "Fig. 8: Ablation studies on the weight of each part in \\(\\mathcal{L}\\), with \\(\\lambda_1 = 1\\) and \\(\\lambda_2\\) varying." + }, + { + "type": "table_caption", + "bbox": [ + 0.249, + 0.291, + 0.751, + 0.305 + ], + "angle": 0, + "content": "Table 6: Ablation studies on loss functions. The presented results are SAs." + }, + { + "type": "table", + "bbox": [ + 0.3, + 0.317, + 0.7, + 0.402 + ], + "angle": 0, + "content": "
Loss TypeCIFAR10STL10CIFAR100MNISTF-MNISTGTSRBSVHNAVG
MSE65.6653.8331.0094.6386.1232.7962.9761.00
Cosine Similarity80.2366.5643.8197.0088.7655.5375.9472.54
KL Divergence77.3565.3040.9896.2387.8257.0469.3870.58
InfoNCE73.9961.6839.2495.7687.8658.8268.8869.43
LD75.5463.243.1696.5289.2765.7771.1572.23
LA79.7065.5943.7996.8989.2455.9676.7072.55
LA+LD79.3966.7644.2796.7489.3262.7573.7473.28
" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.427, + 0.401, + 0.44 + ], + "angle": 0, + "content": "4.4 Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.45, + 0.788, + 0.587 + ], + "angle": 0, + "content": "In this subsection, we conduct ablation studies to investigate the impact of the surrogate dataset size, the patch number for generating prototypes (i.e., \\( n \\) in Eq. 1), the patch number for training (i.e., \\( m \\) in Eq. 2, 3, 5, and 6), and loss functions. All the experiments are conducted to steal the encoder pre-trained on CIFAR10. For evaluations, the trained surrogate encoders are assessed on CIFAR10 in the first three ablation studies. While examining the influence of loss functions, we consider seven distinct downstream datasets, as shown in Table 4, to derive a more comprehensive conclusion. Other configurations remain at their default settings, and we present the resulting SAs." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.605, + 0.788, + 0.681 + ], + "angle": 0, + "content": "Surrogate Dataset Size. Recall our default setting where we randomly sample 2,500 images from Tiny ImageNet to construct the surrogate dataset, which is \\(5\\%\\) the size of CIFAR-10. We vary the ratio from \\(1\\%\\) to \\(10\\%\\) to investigate its impact. Figure 7 (a) shows that a larger surrogate dataset will accelerate the training and improve the attack performance." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.7, + 0.788, + 0.79 + ], + "angle": 0, + "content": "Patch Number for Generating Prototypes. We vary the patch number for generating prototypes from 1 to 100 to investigate its impact. Figure 7 (b) shows that there is a tendency for more patches for generating prototypes will improve the attack performance and accelerate the convergence. However, when the patch number reaches 10, more patches do not further improve the attack performance due to the limited surrogate dataset size." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Patch Number for Training. We vary the patch number for training from 1 to 10 to investigate its impact. Figure 7 (c) reveals a tendency for more patches" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.474, + 0.115, + 0.732, + 0.13 + ], + "angle": 0, + "content": "Refine, Discriminate and Align (RDA)" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.178 + ], + "angle": 0, + "content": "for training will accelerate the training and achieve better results. However, more patches also indicate longer training time and more computation resources." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.199, + 0.789, + 0.486 + ], + "angle": 0, + "content": "Loss Functions. To demonstrate the superiority of the design of our loss function and the necessity of each part, we conduct ablation studies as shown in Table 6. Table 6 shows that \\(\\mathcal{L}_D\\) largely surpass other loss functions on GTSRB while \\(\\mathcal{L}_A\\) largely surpass other loss functions on SVHN. We hypothesize that \\(\\mathcal{L}_A\\) attains optimal outcomes on SVHN due to the pronounced resemblance between SVHN and the surrogate dataset. It appears that \\(\\mathcal{L}_A\\) inclines towards inducing the surrogate encoder to overfit the surrogate data, thereby elucidating its suboptimal performance on GTSRB, a dataset characterized by less similarity with the surrogate dataset. On the contrary, \\(\\mathcal{L}_D\\) makes the surrogate encoder fit less to the surrogate dataset and has the best result on GTSRB. To further investigate the effect of each part in our loss, we have conducted an ablation experiment about the weight of each part in our loss. As Figure 8 shows, the surrogate encoder will perform better on SVHN while worse on GTSRB when the weight of \\(\\mathcal{L}_A\\) increases. We refer to Supplementary A.4 for a more detailed analysis. Combining \\(\\mathcal{L}_D\\) and \\(\\mathcal{L}_A\\), our loss is the most robust one, which achieves the best average result over the seven tested datasets. For example, although cosine similarity and \\(\\mathcal{L}_A\\) performs well on SVHN (about \\(76\\%\\)), their performance on GTSRB is terrible (about \\(56\\%\\)). Our loss design can improve their performance on GTSRB by about \\(7\\%\\) in the cost of less than \\(3\\%\\) decrease on SVHN." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.507, + 0.457, + 0.521 + ], + "angle": 0, + "content": "4.5 Robustness to Defenses" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.532, + 0.784, + 0.592 + ], + "angle": 0, + "content": "In this section, we evaluate the robustness of RDA against three perturbation-based defenses that aim to limit information leakage and one watermarking defense that aims to detect whether a suspected encoder is stolen. The pretraining and downstream datasets both are set to CIFAR10." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.614, + 0.784, + 0.645 + ], + "angle": 0, + "content": "Perturbation-Based Defense. We evaluate three common practices of this type of defense, i.e., adding noise [30], top-\\(k\\) [30] and rounding [44]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.645, + 0.785, + 0.687 + ], + "angle": 0, + "content": "- Adding noise: Adding noise means that the defender will introduce noise to the original outputs of the target encoder. Following [10], we set the mean of the noise to 0 and vary the standard deviations to control the noise level." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.69, + 0.785, + 0.732 + ], + "angle": 0, + "content": "- Top-\\(k\\): Top-\\(k\\) means that the defender will only output the first \\(k\\) largest number of each embedding from the target encoder and set the rest as 0. We vary the value of \\(k\\) to simulate different perturbation levels." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.784, + 0.762 + ], + "angle": 0, + "content": "- Rounding: Rounding truncates each value in the embedding to a specific precision, which we vary to simulate different perturbation levels." + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.645, + 0.785, + 0.762 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.766, + 0.784, + 0.84 + ], + "angle": 0, + "content": "The results of the three perturbation-based defense methods are summarized in (a), (b), and (c) of Figure 9, respectively. We can observe in Figure 9 (a) that while adding noise can mitigate the model stealing attack, it decreases the utility of the target encoder more significantly. Moreover, Figure 9 (b) shows that though top-\\(k\\) can decrease the efficacy of RDA as the defender lowers the" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.353, + 0.127 + ], + "angle": 0, + "content": "S. Wu et al." + }, + { + "type": "image", + "bbox": [ + 0.221, + 0.146, + 0.359, + 0.224 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.272, + 0.228, + 0.325, + 0.238 + ], + "angle": 0, + "content": "(a) Adding noise" + }, + { + "type": "image", + "bbox": [ + 0.361, + 0.145, + 0.499, + 0.224 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.425, + 0.228, + 0.455, + 0.237 + ], + "angle": 0, + "content": "(b) Top-k" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.145, + 0.639, + 0.224 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.56, + 0.228, + 0.602, + 0.238 + ], + "angle": 0, + "content": "(c) Rounding" + }, + { + "type": "image", + "bbox": [ + 0.645, + 0.149, + 0.782, + 0.224 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.694, + 0.228, + 0.75, + 0.238 + ], + "angle": 0, + "content": "(d) Watermarking" + }, + { + "type": "image_caption", + "bbox": [ + 0.275, + 0.25, + 0.726, + 0.265 + ], + "angle": 0, + "content": "Fig. 9: The performance of different defense methods on CIFAR10." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.295, + 0.785, + 0.356 + ], + "angle": 0, + "content": "value of \\( k \\), it also severely deteriorates the target encoder's performance. On the other hand, Figure 9 (c) reveals that rounding has a minimal effect on the performance of both the target encoder and attack. Our experimental results demonstrate the robustness of RDA against perturbation-based defenses." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.378, + 0.788, + 0.545 + ], + "angle": 0, + "content": "Watermarking Defense. As shown by Adi et al. [1], backdoors can be used as watermarks to claim the ownership of a model. In this sense, we follow [34] and leverage BadEncoder [18] to embed a backdoor into the target encoder as the watermark. Ideally, if the watermarked encoder is stolen, the surrogate encoder trained via stealing should also be triggered by the defender-specified trigger to exhibit certain behaviors, and thus the ownership can be claimed. Fortunately, Figure 9 (d) shows that RDA can steal the functionality of the target encoder while unlearning the watermark. More specifically, the watermark rate (WR) of the target encoder is \\(99.87\\%\\) but only \\(17.04\\%\\) for the surrogate encoder when the training converges. This observation indicates that the watermark embedded in the target encoder cannot be preserved by RDA." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.567, + 0.36, + 0.584 + ], + "angle": 0, + "content": "5 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.599, + 0.787, + 0.689 + ], + "angle": 0, + "content": "In this paper, we have proposed a novel model stealing method against SSL named RDA, which stands for refine, discriminate, and align. Compared with previous methods [26,34], RDA establishes a less biased optimization objective for each training sample and strives to extract more abundant functionality of the target encoder. This empowers RDA to exhibit tremendous effectiveness, efficiency, and robustness across diverse datasets and settings." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.788, + 0.78 + ], + "angle": 0, + "content": "Ethical Concerns and Possible Defenses. We underscore that the misuse of model stealing techniques can jeopardize the privacy and economic rights of embedding service providers. We hope that our work will inspire the development of more sophisticated defense mechanisms to thwart model stealing attacks. A possible defense involves rejection schemes for suspicious queries that appear highly similar, which we have more detailed discussions about in the Supplementary." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Limitation and Future Work. The query cost of RDA is not yet optimal and can be further reduced by techniques like clustering to establish an identical prototype for multiple similar images. Moreover, more unique (e.g., more efficient) designs of RDA for transformer-based huge models remain largely unexplored." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.474, + 0.115, + 0.732, + 0.13 + ], + "angle": 0, + "content": "Refine, Discriminate and Align (RDA)" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.239 + ], + "angle": 0, + "content": "Acknowledgements. This work is supported by the National Key R&D Program of China under Grant 2022YFB3103500, the National Natural Science Foundation of China under Grant U20A20176 and 62072062, the National Natural Science Foundation of Chongqing under Grant cstc2022ycjh-bgzxm0031, and the Chongqing Research Program of Basic Research and Frontier Technology (Chongqing Talent) under Grant cstc2024ycjh-bgzxm0048." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.26, + 0.326, + 0.276 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.289, + 0.785, + 0.331 + ], + "angle": 0, + "content": "1. Adi, Y., Baum, C., Cisse, M., Pinkas, B., Keshet, J.: Turning your weakness into a strength: Watermarking deep neural networks by backdoorsing. In: 27th USENIX Security Symposium (USENIX Security 18). pp. 1615-1631 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.333, + 0.785, + 0.387 + ], + "angle": 0, + "content": "2. Bossard, L., Guillaumeun, M., Van Gool, L.: Food-101-mining discriminative components with random forests. In: Computer vision-ECCV 2014: 13th European conference, zurich, Switzerland, September 6-12, 2014, proceedings, part VI 13. pp. 446-461. Springer (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.388, + 0.785, + 0.414 + ], + "angle": 0, + "content": "3. Carlini, N., Wagner, D.: Towards evaluating the robustness of neural networks. In: 2017 IEEE symposium on security and privacy (sp). pp. 39-57. IEEE (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.415, + 0.787, + 0.455 + ], + "angle": 0, + "content": "4. Caron, M., Touvron, H., Misra, I., Jégou, H., Mairal, J., Bojanowski, P., Joulin, A.: Emerging properties in self-supervised vision transformers. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 9650-9660 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.457, + 0.785, + 0.496 + ], + "angle": 0, + "content": "5. Chen, T., Kornblith, S., Norouzi, M., Hinton, G.: A simple framework for contrastive learning of visual representations. In: International conference on machine learning. pp. 1597-1607. PMLR (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.497, + 0.785, + 0.524 + ], + "angle": 0, + "content": "6. Chen, X., Fan, H., Girshick, R., He, K.: Improved baselines with momentum contrastive learning. arXiv preprint arXiv:2003.04297 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.525, + 0.785, + 0.579 + ], + "angle": 0, + "content": "7. Coates, A., Ng, A., Lee, H.: An analysis of single-layer networks in unsupervised feature learning. In: Proceedings of the fourteenth international conference on artificial intelligence and statistics. pp. 215-223. JMLR Workshop and Conference Proceedings (2011)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.58, + 0.785, + 0.62 + ], + "angle": 0, + "content": "8. Dai, Y., Chen, Z., Li, J., Heinecke, S., Sun, L., Xu, R.: Tackling data heterogeneity in federated learning with class prototypes. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 37, pp. 7314-7322 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.621, + 0.785, + 0.661 + ], + "angle": 0, + "content": "9. Devlin, J., Chang, M.W., Lee, K., Toutanova, K.: Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.662, + 0.785, + 0.702 + ], + "angle": 0, + "content": "10. Dziedzic, A., Dhawan, N., Kaleem, M.A., Guan, J., Papernot, N.: On the difficulty of defending self-supervised learning against model extraction. In: International Conference on Machine Learning. pp. 5757-5776. PMLR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.703, + 0.785, + 0.758 + ], + "angle": 0, + "content": "11. Grill, J.B., Strub, F., Altché, F., Tallec, C., Richemond, P., Buchatskaya, E., Doersch, C., Avila Pires, B., Guo, Z., Gheshlaghi Azar, M., et al.: Bootstrap your own latent-a new approach to self-supervised learning. Advances in neural information processing systems 33, 21271-21284 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.759, + 0.785, + 0.799 + ], + "angle": 0, + "content": "12. He, K., Fan, H., Wu, Y., Xie, S., Girshick, R.: Momentum contrast for unsupervised visual representation learning. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 9729-9738 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.8, + 0.785, + 0.84 + ], + "angle": 0, + "content": "13. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 770-778 (2016)" + }, + { + "type": "list", + "bbox": [ + 0.22, + 0.289, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.352, + 0.127 + ], + "angle": 0, + "content": "S. Wu et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.787, + 0.189 + ], + "angle": 0, + "content": "14. Howard, A.G., Zhu, M., Chen, B., Kalenichenko, D., Wang, W., Weyand, T., Andreetto, M., Adam, H.: Mobilenets: Efficient convolutional neural networks for mobile vision applications. arXiv preprint arXiv:1704.04861 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.189, + 0.787, + 0.23 + ], + "angle": 0, + "content": "15. Huang, G., Liu, Z., Van Der Maaten, L., Weinberger, K.Q.: Densely connected convolutional networks. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 4700-4708 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.23, + 0.787, + 0.271 + ], + "angle": 0, + "content": "16. Huang, W., Ye, M., Shi, Z., Li, H., Du, B.: Rethinking federated learning with domain shift: A prototype view. In: 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 16312-16322. IEEE (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.271, + 0.787, + 0.312 + ], + "angle": 0, + "content": "17. Jia, H., Choquette-Choo, C.A., Chandrasekaran, V., Papernot, N.: Entangled watermarks as a defense against model extraction. In: 30th USENIX Security Symposium (USENIX Security 21). pp. 1937-1954 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.312, + 0.787, + 0.353 + ], + "angle": 0, + "content": "18. Jia, J., Liu, Y., Gong, N.Z.: Badencoder: Backdoor attacks to pre-trained encoders in self-supervised learning. In: 2022 IEEE Symposium on Security and Privacy (SP). pp. 2043-2059. IEEE (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.353, + 0.787, + 0.393 + ], + "angle": 0, + "content": "19. Khosla, P., Teterwak, P., Wang, C., Sarna, A., Tian, Y., Isola, P., Maschinot, A., Liu, C., Krishnan, D.: Supervised contrastive learning. Advances in neural information processing systems 33, 18661-18673 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.393, + 0.787, + 0.421 + ], + "angle": 0, + "content": "20. Kingma, D.P., Ba, J.: Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980 (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.421, + 0.787, + 0.447 + ], + "angle": 0, + "content": "21. Krizhevsky, A., Hinton, G., et al.: Learning multiple layers of features from tiny images (2009)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.447, + 0.787, + 0.473 + ], + "angle": 0, + "content": "22. Krizhevsky, A., Nair, V., Hinton, G.: Cifar-10 (canadian institute for advanced research). URL http://www.cs.toronto.edu/kriz/cifar.html 5(4), 1 (2010)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.473, + 0.787, + 0.501 + ], + "angle": 0, + "content": "23. Le, Y., Yang, X.: Tiny imagenet visual recognition challenge. CS 231N 7(7), 3 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.501, + 0.787, + 0.527 + ], + "angle": 0, + "content": "24. LeCun, Y., Cortes, C.: MNIST handwritten digit database (2010), http://yann.lecun.com/exdb/mnist/" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.527, + 0.787, + 0.582 + ], + "angle": 0, + "content": "25. Liu, H., Jia, J., Qu, W., Gong, N.Z.: Encodermi: Membership inference against pre-trained encoders in contrastive learning. In: Proceedings of the 2021 ACM SIGSAC Conference on Computer and Communications Security. pp. 2081-2095 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.582, + 0.787, + 0.623 + ], + "angle": 0, + "content": "26. Liu, Y., Jia, J., Liu, H., Gong, N.Z.: Stolenencoder: stealing pre-trained encoders in self-supervised learning. In: Proceedings of the 2022 ACM SIGSAC Conference on Computer and Communications Security. pp. 2115-2128 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.623, + 0.787, + 0.651 + ], + "angle": 0, + "content": "27. Van der Maaten, L., Hinton, G.: Visualizing data using t-sne. Journal of machine learning research 9(11) (2008)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.651, + 0.787, + 0.677 + ], + "angle": 0, + "content": "28. Mettes, P., Van der Pol, E., Snoek, C.: Hyperspherical prototype networks. Advances in neural information processing systems 32 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.677, + 0.787, + 0.704 + ], + "angle": 0, + "content": "29. Netzer, Y., Wang, T., Coates, A., Bissacco, A., Wu, B., Ng, A.Y.: Reading digits in natural images with unsupervised feature learning (2011)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.704, + 0.787, + 0.744 + ], + "angle": 0, + "content": "30. Orekondy, T., Schiele, B., Fritz, M.: Knockoff nets: Stealing functionality of black-box models. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 4954-4963 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.744, + 0.787, + 0.786 + ], + "angle": 0, + "content": "31. Papernot, N., McDaniel, P., Goodfellow, I., Jha, S., Celik, Z.B., Swami, A.: Practical black-box attacks against machine learning. In: Proceedings of the 2017 ACM on Asia conference on computer and communications security. pp. 506-519 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.786, + 0.787, + 0.84 + ], + "angle": 0, + "content": "32. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.474, + 0.115, + 0.732, + 0.13 + ], + "angle": 0, + "content": "Refine, Discriminate and Align (RDA)" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.189 + ], + "angle": 0, + "content": "33. Salem, A., Zhang, Y., Humbert, M., Berrang, P., Fritz, M., Backes, M.: Ml-leaks: Model and data independent membership inference attacks and defenses on machine learning models. arXiv preprint arXiv:1806.01246 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.19, + 0.788, + 0.232 + ], + "angle": 0, + "content": "34. Sha, Z., He, X., Yu, N., Backes, M., Zhang, Y.: Can't steal? cont-steal! contrastive stealing attacks against image encoders. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16373-16383 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.232, + 0.788, + 0.259 + ], + "angle": 0, + "content": "35. Sharir, O., Peleg, B., Shoham, Y.: The cost of training nlp models: A concise overview. arXiv preprint arXiv:2004.08900 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.26, + 0.788, + 0.3 + ], + "angle": 0, + "content": "36. Shokri, R., Stronati, M., Song, C., Shmatikov, V.: Membership inference attacks against machine learning models. In: 2017 IEEE symposium on security and privacy (SP). pp. 3-18. IEEE (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.301, + 0.788, + 0.328 + ], + "angle": 0, + "content": "37. Simonyan, K., Zisserman, A.: Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556 (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.328, + 0.788, + 0.356 + ], + "angle": 0, + "content": "38. Snell, J., Swersky, K., Zemel, R.: Prototypical networks for few-shot learning. Advances in neural information processing systems 30 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.357, + 0.788, + 0.397 + ], + "angle": 0, + "content": "39. Stallkamp, J., Schlipsing, M., Salmen, J., Igel, C.: Man vs. computer: Benchmarking machine learning algorithms for traffic sign recognition. Neural networks 32, 323-332 (2012)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.398, + 0.788, + 0.439 + ], + "angle": 0, + "content": "40. Tan, Y., Long, G., Liu, L., Zhou, T., Lu, Q., Jiang, J., Zhang, C.: Fedproto: Federated prototype learning across heterogeneous clients. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 36, pp. 8432-8440 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.44, + 0.788, + 0.481 + ], + "angle": 0, + "content": "41. Tan, Y., Long, G., Ma, J., Liu, L., Zhou, T., Jiang, J.: Federated learning from pretrained models: A contrastive learning approach. Advances in Neural Information Processing Systems 35, 19332-19344 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.481, + 0.788, + 0.536 + ], + "angle": 0, + "content": "42. Tian, Y., Wang, Y., Krishnan, D., Tenenbaum, J.B., Isola, P.: Rethinking few-shot image classification: a good embedding is all you need? In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XIV 16. pp. 266-282. Springer (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.536, + 0.788, + 0.564 + ], + "angle": 0, + "content": "43. Tong, S., Chen, Y., Ma, Y., Lecun, Y.: Emp-ssl: Towards self-supervised learning in one training epoch. arXiv preprint arXiv:2304.03977 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.565, + 0.788, + 0.605 + ], + "angle": 0, + "content": "44. Tramér, F., Zhang, F., Juels, A., Reiter, M.K., Ristenpart, T.: Stealing machine learning models via prediction {APIs}. In: 25th USENIX security symposium (USENIX Security 16). pp. 601-618 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.606, + 0.788, + 0.647 + ], + "angle": 0, + "content": "45. Wu, Z., Xiong, Y., Yu, S.X., Lin, D.: Unsupervised feature learning via nonparametric instance discrimination. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 3733-3742 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.647, + 0.788, + 0.675 + ], + "angle": 0, + "content": "46. Xiao, H., Rasul, K., Vollgraf, R.: Fashion-mnist: a novel image dataset for benchmarking machine learning algorithms. arXiv preprint arXiv:1708.07747 (2017)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.675 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/Refine, Discriminate and Align_ Stealing Encoders via Sample-Wise Prototypes and Multi-Relational Extraction/2d1e904f-54f5-49e9-a836-f9f59c16b169_origin.pdf b/2024/Refine, Discriminate and Align_ Stealing Encoders via Sample-Wise Prototypes and Multi-Relational Extraction/2d1e904f-54f5-49e9-a836-f9f59c16b169_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..8190ff2c0aa6164adb75156ac4e122a9d864f4ca --- /dev/null +++ b/2024/Refine, Discriminate and Align_ Stealing Encoders via Sample-Wise Prototypes and Multi-Relational Extraction/2d1e904f-54f5-49e9-a836-f9f59c16b169_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ad2b6e44239abd6526aa4eebaca45ef28fa2d90b6548409cf0ac41430daeddc2 +size 3001604 diff --git a/2024/Refine, Discriminate and Align_ Stealing Encoders via Sample-Wise Prototypes and Multi-Relational Extraction/full.md b/2024/Refine, Discriminate and Align_ Stealing Encoders via Sample-Wise Prototypes and Multi-Relational Extraction/full.md new file mode 100644 index 0000000000000000000000000000000000000000..525e43c81ea7f821fdbc86cfe3011f805c3c5ece --- /dev/null +++ b/2024/Refine, Discriminate and Align_ Stealing Encoders via Sample-Wise Prototypes and Multi-Relational Extraction/full.md @@ -0,0 +1,319 @@ +# Refine, Discriminate and Align: Stealing Encoders via Sample-Wise Prototypes and Multi-Relational Extraction + +Shuchi $\mathbf{W u}^{1}$ Chuan Ma $^{2\text{图}}$ Kang Wei $^{3\text{图}}$ Xiaogang Xu $^{4}$ Ming Ding $^{5}$ Yuwen Qian $^{1}$ Di Xiao $^{2}$ Tao Xiang $^{2}$ $^{1}$ NJUST $^{2}$ CQU $^{3}$ PolyU $^{4}$ CUHK $^{5}$ Data61, CSIRO + +Abstract. This paper introduces RDA, a pioneering approach designed to address two primary deficiencies prevalent in previous endeavors aiming at stealing pre-trained encoders: (1) suboptimal performances attributed to biased optimization objectives, and (2) elevated query costs stemming from the end-to-end paradigm that necessitates querying the target encoder every epoch. Specifically, we initially Refine the representations of the target encoder for each training sample, thereby establishing a less biased optimization objective before the steal-training phase. This is accomplished via a sample-wise prototype, which consolidates the target encoder's representations for a given sample's various perspectives. Demanding exponentially fewer queries compared to the end-to-end approach, prototypes can be instantiated to guide subsequent query-free training. For more potent efficacy, we develop a multi-relational extraction loss that trains the surrogate encoder to Discriminate mismatched embedding-prototype pairs while Aligning those matched ones in terms of both amplitude and angle. In this way, the trained surrogate encoder achieves state-of-the-art results across the board in various downstream datasets with limited queries. Moreover, RDA is shown to be robust to multiple widely-used defenses. Our code is available at https://github.com/ShuchiWu/RDA. + +Keywords: Model Stealing $\cdot$ Self-Supervised Learning $\cdot$ Prototype + +# 1 Introduction + +Self-supervised learning (SSL) [4, 5, 9, 11, 12] is endowed with the capability to harness unlabeled data for pre-training a versatile encoder that applicable to a range of downstream tasks, or even showcasing groundbreaking zero-shot performances, e.g., CLIP [32]. However, SSL typically requires a large volume of data and computation resources to achieve a convincing performance, i.e., high-performance encoders are expensive to train [35]. To safeguard the confidentiality and economic worth of these encoders, entities like OpenAI offer their encoders + +![](images/102cf00079713b5379c0dbb6a242b65d6ed740f7aa6dffd2a1c596b665e68086.jpg) +Fig. 1: Illustrations of four stealing methods against SSL. The dotted arrows and text beside interpret how each method optimizes the surrogate encoder. Surrogate encoder branches in (b)-(d) involve data augmentations for training. Both (c) and (d) augment each sample before querying the target encoder but adopting different schemes. + +![](images/bf5a8f6fe9d7486cf9958564a135c5470f0e61d6b0083e049e9045f42bcdca4b.jpg) + +![](images/30de37fa9d6474e983a08d6720e649e4760880ab9a552717ae81b094d75b0800.jpg) + +![](images/c3e937a76416808419e7ca5890fef7608f73334babc6b6527955b087f1ce55da.jpg) + +as a premium service, exclusively unveiling the service API to the public. Users have the privilege of soliciting embeddings for their data, facilitating the training of diverse models tailored for specific downstream tasks. Regrettably, the substantial value of these encoders and their exposure to publicly accessible APIs render them susceptible to model stealing attacks [10,26,34]. + +Given a surrogate dataset, model stealing attacks aim to mimic the outputs of a target model to train either a high-accuracy copy of comparable performance or a high-fidelity copy that can serve as a stepping stone to perform further attacks like adversarial examples [31], membership inference attacks [33,36], etc. The goal of this paper is to develop an approach that can train a surrogate encoder with competitive performance on downstream tasks by only accessing the target encoder's output embeddings. Meanwhile, it should require as little cost as possible, which is primarily derived from querying the target encoder. + +We begin with systematically studying existing techniques for stealing pretrained encoders. Specifically, the conventional method [34] and StolenEncoder [26] are similar, and both only need to query the target encoder once with each sample before the training. The output embedding can be viewed as a "ground truth" or an "optimization objective" to the sample, and the surrogate encoder is optimized to output a similar embedding when the same sample is fed, as illustrated in Figure 1 (a). The sole distinction lies in the optimization of StolenEncoder, which incorporates data augmentations (refer to Figure 1 (b)), as it supposes the target encoder will produce similar embeddings for an image and its augmentations. Nevertheless, we contend that this presumed similarity is of a modest degree, given the disparity between the surrogate and pre-training data. The target encoder's representations for such data's various augmentations may diverge and even be biased to other samples, as visually demonstrated in Figure 2 using a toy experiment. In this sense, only using the embedding of an image's single perspective (regardless of the original or augmented version) as the optimization objective is inadvisable. + +On the other hand, Cont-Steal [34] augments each sample into two perspectives in each epoch: one is used to query the target encoder while the other is fed to the surrogate encoder. Two embeddings of the same sample from the tar + +![](images/1860f8975f0322c0f64a15a77cbd8ad0f96ab1be3e835af5f480b2c40c220fd1.jpg) +Fig. 2: t-SNE of embeddings belonging to five different images generated by an encoder pre-trained on CIFAR10, with each image augmented into 500 patches and fed into the encoder. Each black marker represents the mean of the 500 embeddings of a certain image, i.e., its prototype. Among the embeddings of an image's various augmentations, some can be diverged or even biased. In contrast, each image's prototype is more distinguishable, i.e., less biased. + +get and surrogate encoders will be aligned, while those of different samples will be pushed apart, as illustrated in Figure 1 (c). In spite of such an end-to-end training scheme performing better, it suffers from high query costs since each sample is augmented and used to query in each epoch. + +To tackle these issues, we propose RDA. Specifically, we first Refine the target encoder's representations for each sample in the surrogate dataset by averagely aggregating its several (e.g., 10) different augmentations' embeddings, i.e., their mean value, to establish a sample-level prototype for it. Then, the sample-wise prototype is used to guide the surrogate encoder optimization, which can mitigate the impact of biased embeddings, as shown in Figure 2 (see black markers). We have an experiment presented in Figure 10 of our supplementary material to further quantitatively reveal the benefit of prototypes, i.e., significantly more similar with each augmentation patch's embedding, showcasing it is less biased. With a prototype, there is a static optimization objective for each sample across the entire training, and thus, it is done in a query-free manner. This enables RDA to have far less query cost than the end-to-end approach (less than $10\%$ ). To further enhance the attack efficacy with limited queries, we develop a multi-relational extraction loss that trains the surrogate encoder to Discriminate mismatched embedding-prototype pairs while Aligning those matched ones in terms of both amplitude and angle. The framework and pipeline of RDA are depicted in Figures 1 (d) and 4, respectively. + +Extensive experiments show that RDA acquires a favorable trade-off between the query cost (a.k.a, the money cost) and attack efficacy. As depicted in Figure 3, when the target encoder is pre-trained on CIFAR10, RDA averagely outperforms the state-of-the-art (SOTA), i.e., Cont-Steal [34], by $1.22\%$ on the stealing efficacy over seven different downstream datasets, with a competitive smaller query cost, i.e., only $1\%$ of Cont-Steal. Further, this performance gap can be widened to $5.20\%$ with $10\%$ query cost of Cont-Steal. In particular, RDA is demonstrated to be robust against multiple prevalent defenses [18,30,44]. + +In conclusion, our contributions are three folds: + +![](images/56a5f18e0e51d7821f4c94d07418fb7820597df2269ce1ae8920a67ff055fcfc.jpg) +Fig. 3: Performance comparisons between four stealing methods against SSL. The presented results are the mean values achieved by each method over seven different downstream classification tasks, with their corresponding query costs. Our proposed RDA can achieve SOTA results with the least query cost. + +- We comprehensively investigate two critical inadequacies of existing stealing methods against SSL, i.e., suboptimal efficacy and high query costs, and analyze the causes. +- We develop a novel approach to train surrogate encoders, namely RDA, with sample-wise prototype guidance and a multi-relational extraction loss. +- Extensive experiments are conducted to verify the effectiveness and robustness of RDA. + +# 2 Related Work + +Prototype Learning. A prototype refers to the mean of embeddings belonging to all images of an identical class, which serves as a proxy of the class [40]. With each class having a prototype, the model is trained to match its output with the corresponding prototype when inputting a sample of a certain class. Prototype learning has proven beneficial for various learning scenarios, e.g., federated learning [8, 16, 41] and few-shot learning [28, 38, 42]. In this paper, the optimization objective (i.e., the mean of embeddings belonging to an identical sample's multiple augmentations) we generate for each training sample is conceptually similar to prototypes, and thus we name it a sample-wise prototype. + +Model Stealing Attacks against SSL. Dziedzic et al. [10] pointed out that the higher dimension of embeddings leaks more information than labels, making SSL more easily stolen. The primary objective of a stealing attack is to train a surrogate encoder to achieve high accuracy on downstream tasks or recreate a high-fidelity copy that can be used to mount further attacks such as adversarial examples [3] and membership inference attacks [25]. + +Defenses against Model Stealing Attacks. Existing defenses against model stealing attacks can be categorized based on when they are applied [10]. Perturbation-based defenses, e.g., adding noise [30], top- $k$ [30] and truncating outputs [44], are applied before a stealing attack happens, aiming to limit the information leakage. On the other hand, watermarking defenses embed watermarks or unique + +identifiers into the target model and use them to detect whether it is stolen after a stealing attack happens [17]. + +# 3 Methodology + +# 3.1 Threat Model + +Given a target encoder $E_{T}$ , the attacker aims to train a surrogate encoder $E_{S}$ at the lowest possible cost, which can perform competitively on downstream tasks with $E_{T}$ . To achieve this goal, the attacker queries $E_{T}$ for embeddings of an unlabeled surrogated dataset $D_{S} = \{\pmb{x}_{1},\dots,\pmb{x}_{N}\}$ consisting of $N$ images, and guide $E_{S}$ to mimic the output of $E_{T}$ for each sample in $D_{S}$ . Specifically, we consider a black-box setting, where the attacker has direct access only to the outputs of $E_{T}$ while remaining unaware of its architecture and training configurations, including pre-training datasets, loss functions, data augmentations schemes, etc. + +# 3.2 Sample-Wise Prototypes + +Recalling the discussion in Section 1, we expect to refine the target encoder's representations to find a less biased optimization objective for each sample in $D_{S}$ . Enlightened by prototype learning [40] and Extreme-Multi-Patch-SSL (EMPSL) [43], which establish a prototype/benchmark for each class/sample, we propose a sample-wise prototype generation method for this purpose. In detail, the process of prototype generation can be divided into three steps as follows: 1 Cropping and augmenting each image $\pmb{x}_i\in D_S$ into $n$ augmentation patches denoted as $\{\pmb{x}_{i,t,c}^{\prime}\}_{c = 1}^{n} = \{\pmb{x}_{i,t,1}^{\prime},\dots,\pmb{x}_{i,t,n}^{\prime}\}$ ; 2 Querying $E_{T}$ with each augmentation patch in $\{\pmb{x}_{i,t,c}^{\prime}\}_{c = 1}^{n}$ , resulting in a set of embeddings denoted as $\{E_T(\pmb{x}_{i,t,c}^{\prime})\}_{c = 1}^{n} = \{E_T(\pmb{x}_{i,t,1}^{\prime}),\dots,E_T(\pmb{x}_{i,t,n}^{\prime})\}$ ; 3 Calculating the mean of $\{E_T(\pmb{x}_{i,t,c}^{\prime})\}_{c = 1}^{n}$ as the prototype $p_{\pmb{x}_i}$ for $\pmb{x}_i$ as follows: + +$$ +p _ {\boldsymbol {x} _ {i}} = \frac {1}{n} \sum_ {c = 1} ^ {n} E _ {T} \left(\boldsymbol {x} _ {i, t, c} ^ {\prime}\right), \quad \boldsymbol {x} _ {i} \in D _ {s}. \tag {1} +$$ + +Each generated prototype will be stored in a memory bank. These sample-wise prototypes can provide a stable optimization objective for each sample throughout the training process. Consequently, the attacker can accomplish the training in a query-free manner. Notably, setting $n$ to a value that is far less than the training epochs, e.g., 10 vs. 100, is sufficient to yield a favorable performance. The prototype generation process is illustrated in step ① of Figure 4. Supplementary A.2 explains how we build the memory bank in practice. + +# 3.3 Multi-Relational Extraction Loss + +As step ② of Figure 4 shows, during training $E_{S}$ , we also perform cropping and augmentation on each image $\pmb{x}_i \in D_S$ to create $m$ patches, forming an augmentation patch set denoted as $\{\pmb{x}_{i,s,q}^{\prime}\}_{q = 1}^{m} = \{\pmb{x}_{i,s,1}^{\prime},\dots,\pmb{x}_{i,s,m}^{\prime}\}$ . The $m$ does not necessarily need to be set equal to $n$ . Next, we feed each augmentation patch from $\{\pmb{x}_{i,s,q}^{\prime}\}_{q = 1}^{m}$ to $E_{S}$ , obtaining a set of embeddings denoted as + +![](images/c1ffdac658369ac2a9dd058449fc982831b16cf26efdf59a466e4c4bad57f29c.jpg) +Fig. 4: Pipeline of RDA. Prototype generation: augment one sample into $n$ patches and use them to query the target encoder $(E_T)$ . The mean of the $n$ patches's embeddings is defined as a prototype for this sample. Forward encoding: crop one image into $m$ patches and feed them to the surrogate encoder $(E_S)$ for their embeddings. Optimization: align embeddings from the surrogate encoder to their matched prototypes in both angle and amplitude while pushing away those belonging to different samples. + +$\{E_S(\pmb{x}_{i,s,q}^{\prime})\}_{q = 1}^{m} = \{E_S(\pmb{x}_{i,s,1}^{\prime}),\dots ,E_S(\pmb{x}_{i,s,m}^{\prime})\}$ . To optimize $E_{S}$ , we develop a multi-relational extraction loss composed of two parts, i.e., the discriminating loss and the aligning loss. + +Discriminating Loss. The discriminating loss is responsible for training $E_{S}$ to distinguish between different samples, as depicted in step ③ of Figure 4. For this goal, it pushes each embedding in $\{E_S(\pmb{x}_{i,s,q}')\}_{q = 1}^m$ away from mismatched prototypes $p_{\pmb{x}_j}, i \neq j$ (denoted as negative pairs). In this regard, contrastive learning [5,6,12,19] offers a solution, as it can pull embeddings of positive pairs close while pushing embeddings of negative pairs apart. Furthermore, since more negative pairs can help improve the performance of contrastive learning [5], we refer to the loss designed by Sha et al. [34] to propose our discriminating loss $\mathcal{L}_D$ . In particular, $\mathcal{L}_D$ considers both mismatched prototype-embedding pairs from different samples as well as embeddings from $E_{S}$ for different samples as negative pairs. Formally, $\mathcal{L}_D$ can be expressed as follows: + +$$ +\mathcal {L} _ {p o s} \left(\boldsymbol {x} _ {i}\right) = \frac {1}{m} \sum_ {q = 1} ^ {m} \exp \left(\operatorname {s i m} \left(E _ {S} \left(\boldsymbol {x} _ {i, s, q} ^ {\prime}\right), p _ {\boldsymbol {x} _ {i}}\right) / \tau\right), \tag {2} +$$ + +$$ +\begin{array}{l} \mathcal {L} _ {n e g} \left(\boldsymbol {x} _ {i}\right) = \frac {1}{m} \sum_ {q = 1} ^ {m} \sum_ {k = 1} ^ {N} \mathbb {1} _ {[ i \neq k ]} \left(\exp \left(s i m \left(E _ {S} \left(\boldsymbol {x} _ {i, s, q} ^ {\prime}\right), p _ {\boldsymbol {x} _ {k}}\right) / \tau\right) \right. \tag {3} \\ + \exp \left(\operatorname {s i m} \left(E _ {S} \left(\boldsymbol {x} _ {i, s, q} ^ {\prime}\right), E _ {S} \left(\boldsymbol {x} _ {k, s, q} ^ {\prime}\right)\right) / \tau\right)), \\ \end{array} +$$ + +$$ +\mathcal {L} _ {D} = - \frac {1}{N} \sum_ {i = 1} ^ {N} \log \frac {\mathcal {L} _ {p o s} (\boldsymbol {x} _ {i})}{\mathcal {L} _ {n e g} (\boldsymbol {x} _ {i})}, \tag {4} +$$ + +where $\operatorname{sim}(u, v)$ represents the cosine similarity between $u$ and $v$ , and $\tau$ is a temperature parameter. + +Aligning Loss. While $\mathcal{L}_D$ can separate different samples effectively, it does not fully leverage the potential of prototypes. To further enhance the attack efficacy, we propose to align embeddings from $E_S$ more thoroughly with their matched prototypes from the high-performance $E_T$ , i.e., in terms of both amplitude and angle, as step ③ of Figure 4 shows. To measure the amplitude and angle deviations between two embeddings, we employ the mean square error (MSE) and cosine similarity to quantify them respectively, which can be formulated as follows: + +$$ +\mathcal {L} _ {a m p} ^ {\prime} \left(\boldsymbol {x} _ {i}\right) = \frac {1}{m} \sum_ {q = 1} ^ {m} \left\| E _ {S} \left(\boldsymbol {x} _ {i, s, q} ^ {\prime}\right) - p _ {\boldsymbol {x} _ {i}} \right\| ^ {2}, \tag {5} +$$ + +$$ +\mathcal {L} _ {\text {a n g}} ^ {\prime} \left(\boldsymbol {x} _ {i}\right) = \frac {1}{m} \sum_ {q = 1} ^ {m} \operatorname {s i m} \left(E _ {S} \left(\boldsymbol {x} _ {i, s, q} ^ {\prime}\right), p _ {\boldsymbol {x} _ {i}}\right). \tag {6} +$$ + +Notice that we will normalize the value of $\mathcal{L}_{ang}^{\prime}$ to [0, 1]. However, there is an issue concerning the equal penalty given by $\mathcal{L}_{amp}^{\prime}$ and $\mathcal{L}_{ang}^{\prime}$ to every deviation increase. For instance, when the MSE increases from 0.8 to 0.9 or from 0.3 to 0.4, both penalties given by $\mathcal{L}_{amp}^{\prime}$ are 0.1, which is imprudent. On the contrary, we should have a more rigorous penalty for the increase in MSE from 0.3 to 0.4, as a smaller MSE is more desirable. Likewise, the same penalizing regime should be applied on $1 / \mathcal{L}_{ang}^{\prime}$ for the same reason. To this end, we adopt a logarithmic function to redefine the ultimate formulations of $\mathcal{L}_{amp}$ and $\mathcal{L}_{ang}$ as follows: + +$$ +\mathcal {L} _ {\text {a m p}} \left(\boldsymbol {x} _ {i}\right) = \log \mathcal {L} _ {\text {a m p}} ^ {\prime} \left(\boldsymbol {x} _ {i}\right), \tag {7} +$$ + +$$ +\mathcal {L} _ {a n g} \left(\boldsymbol {x} _ {i}\right) = \log \left(1 / \mathcal {L} _ {a n g} ^ {\prime} \left(\boldsymbol {x} _ {i}\right)\right) = - \log \mathcal {L} _ {a n g} ^ {\prime} \left(\boldsymbol {x} _ {i}\right). \tag {8} +$$ + +We posit the amplitude and angle deviations hold equal importance, and formulate the aligning loss as follows: + +$$ +\mathcal {L} _ {A} = \frac {1}{N} \sum_ {i = 1} ^ {N} \left(\mathcal {L} _ {a m p} (\boldsymbol {x} _ {i}) + \mathcal {L} _ {a n g} (\boldsymbol {x} _ {i})\right) = - \frac {1}{N} \sum_ {i = 1} ^ {N} \log \frac {\mathcal {L} _ {a n g} ^ {\prime} (\boldsymbol {x} _ {i})}{\mathcal {L} _ {a m p} ^ {\prime} (\boldsymbol {x} _ {i})}. \tag {9} +$$ + +Our experiments in Supplementary A.4 demonstrate the combination of $\mathcal{L}_{amp}$ and $\mathcal{L}_{ang}$ offers superior results than using each solely. + +Lastly, the loss function for stealing pre-trained encoders is as follows: + +$$ +\mathcal {L} = \lambda_ {1} \cdot \mathcal {L} _ {D} + \lambda_ {2} \cdot \mathcal {L} _ {A}, \tag {10} +$$ + +where $\lambda_{1}$ and $\lambda_{2}$ are preset coefficients to adjust the weight of each part in $\mathcal{L}$ . To show the superiority of our design, especially on the loss function, we have conducted ablation studies in Section 4.4 and explored several alternative designs in Supplementary A.3. + +Detailed steps of RDA are summarized in Algorithm 1 of Supplementary A.2. + +# 4 Experiments + +# 4.1 Experimental Setup + +Target Encoder Settings. We use SimCLR [5] to pre-train two ResNet18 [13] encoders on CIFAR10 [21] and STL10 [7], respectively, as two medium-scale target encoders. Furthermore, we consider two real-world large-scale ResNet50 encoders as the targets, i.e., the ImageNet encoder pre-trained by Google [5], and the CLIP encoder pre-trained by OpenAI [32]. Besides the ResNet family, RDA also has been demonstrated effective upon other backbones, i.e., VGG19_bn [37], DenseNet121 [15], and MobileNetV2 [14], in Supplementary A.4. + +Attack Settings. Regarding the surrogate dataset, it is derived from Tiny ImageNet [23]. Specifically, we randomly sample 2,500 images and resize them to + +Table 1: Results of RDA against two Table 2: Results of RDA against the medium-scale encoders. ImageNet Encoder and CLIP. + +
Pre-training DatasetDownstream DatasetTASA\( \frac{\mathrm{{SA}}}{\mathrm{{TA}}} \times {100}\% \)
CIFAR10MNIST97.2796.6299.33
F-MNIST88.5889.32100.84
GTSRB61.7662.75101.60
SVHN73.7873.7499.95
STL10MNIST96.8496.3199.45
F-MNIST90.0887.9197.59
GTSRB64.4557.7089.53
SVHN61.8570.67114.26
+ +
TargetEncoderDownstreamDatasetTASA\( \frac{\mathrm{{SA}}}{\mathrm{{TA}}} \times {100}\% \)
ImageNetEncoderMNIST97.3895.8998.47
F-MNIST91.6790.8399.08
GTSRB63.1459.6294.43
SVHN73.2069.2194.55
CLIPMNIST97.9093.9695.98
F-MNIST89.9287.4397.23
GTSRB68.6155.8481.39
SVHN69.5357.5882.81
+ +Table 3: Computation resources required by pre-training the two targeted real-world encoders from scratch and RDA to steal them. + +
TargetEncoderPre-trainingRDA
HardwareTime (hrs)HardwareTime (hrs)
ImageNetEncoderTPU v3192RTX A500010.8
CLIPV100 GPU255,74416.2
+ +$32 \times 32$ for stealing encoders pre-trained on CIFAR10 and STL10. When stealing the ImageNet encoder and CLIP, the number is 40,000 and 60,000, respectively, with each image resized to $224 \times 224$ . For the surrogate encoder architecture, we adopt a ResNet18 across our experiments. During the attack, we set $n$ in Eq. 1 as 10, $m$ in Eq. 2, 3, 5, and 6 as 5, $\tau$ in Eq. 2 and 3 as 0.07, and both $\lambda_{1}$ and $\lambda_{2}$ in Eq. 10 as 1 unless stated otherwise. For training, the batch size is set as 100, and we employ an Adam optimizer [20] with a learning rate of 0.001. + +Evaluation Settings. We train each surrogate encoder for 100 epochs and test its KNN accuracy [45] on CIFAR10 after each epoch. The best-trained surrogate encoders, as well as target encoders, will be used to train downstream classifiers for the linear probing evaluation. Each downstream classifier will be trained for 100 epochs with an Adam optimizer and a learning rate of 0.0001. Specifically, we totally consider seven downstream datasets, namely MNIST [24], CIFAR10 [21], STL10 [7], GTSRB [39], CIFAR100 [22], SVHN [29], and F-MNIST [46]. Additional small-scale experiments on more complex datasets (e.g., Food 101 [2]) are included in our supplementary material. As done in [26], we use Target Accuracy (TA) to evaluate target encoders, Steal Accuracy (SA) to evaluate surrogate encoders, and $\frac{\mathrm{SA}}{\mathrm{TA}} \times 100\%$ to evaluate the efficacy of the stealing attack. + +# 4.2 Effectiveness of RDA + +Stealing Medium-Scale Encoders. Table 1 shows the results achieved by RDA upon two medium-scale target encoders pre-trained on CIFAR10 and STL10, respectively. As the results show, using a surrogate dataset of $5\%$ the size of CIFAR10 (2,500 / 50,000) and $2.4\%$ the size of STL10 (2,500 / 105,000), RDA can steal nearly $100\%$ of the functionality of both target encoders, except for the GTSRB scenario when stealing the encoder pre-trained on STL10. Nonetheless, even in this exceptional case, the obtained $\frac{\mathrm{SA}}{\mathrm{TA}} \times 100\%$ is approximately $90\%$ , + +![](images/f4c9168e25fc959d5a4b810ec6aed0b592ce8c78b8fbef63ffbce383fa0ae723.jpg) +(a) ImageNet Encoder + +![](images/826d2dc194a694fd6ef313d3fa012d06054fb999154e96f0a09da433f2a4d946.jpg) +(b) CLIP + +![](images/76f52550ead86fafa6ba6471d1896e908607bce1958d55ee71f2d236d97a3d0b.jpg) +Fig.5: t-SNE of embeddings of 2,000 images sampled from CIFAR10 generated by the ImageNet encoder and CLIP. +Fig. 6: The training epochs of RDA are fixed to 100 while those of baselines are prolonged to 300. + +which exemplifies the remarkable effectiveness of RDA. Furthermore, the surrogate encoder trained through RDA even outperforms the target encoder on multiple datasets. Our interpretation of this is that the target encoder fits the pre-training dataset more strongly compared to the RDA-trained surrogate encoder, thus exhibiting inferior performance on some out-of-distribution datasets. + +Stealing Real-World Large-Scale Encoders. We further evaluate RDA on two real-world pre-trained encoders, i.e., the ImageNet encoder and CLIP. We aim to demonstrate the effectiveness of RDA upon such large-scale encoders trained with a massive amount of data. Specifically, the ImageNet encoder is pre-trained on the ImageNet dataset, which has 1.3 million images. The CLIP encoder is pre-trained on a web-scale dataset that consists of 400 million image-text pairs. During the attack, we set the patch number for training as 1 (i.e., $m$ in Eq. 2, 3, 5, and 6) to save time. Other parameters follow the default setting except the $\lambda_{1}$ and $\lambda_{2}$ when stealing CLIP, in which we set $\lambda_{1}$ and $\lambda_{2}$ to 1 and 20, respectively. To interpret the reason for this adjustment, we randomly sample 200 images from each class in the testing set of CIFAR10 and visualize the output embeddings from the ImageNet encoder and CLIP using t-SNE [27]. Figure 5 shows that embeddings from the ImageNet encoder are more uniformly distributed compared to those from CLIP. This is because CLIP is a multimodal encoder, which is pre-trained by conducting contrastive learning between image-text pairs. Therefore, the text also shares part of the embedding space and is mutually exclusive with mismatched images. Although $\mathcal{L}_D$ can push embeddings of different samples apart, a significant weight of it will make all embeddings more uniformly distributed, i.e., not align with CLIP. Therefore, we set a smaller weight to $\mathcal{L}_D$ when stealing CLIP. This adjustment is practical since the attacker has access to the outputs of the target encoder, and thus can observe its embedding space to adjust its attack settings. Table 2 shows that using a surrogate dataset of only $3.07\%$ the size of ImageNet (40K / 1.3M) and $0.015\%$ the size of the training data of CLIP (60K / 400M), RDA can achieve comparable performances with the two encoders across various downstream tasks. The results also reveal that even under the architecture of the surrogate and target encoders being distinct (ResNet18 vs. ResNet50), RDA still exhibits high effec + +Table 4: Comparisons between RDA and baselines to steal the encoder pre-trained on CIFAR10 under the same surrogate dataset size. We report the $\frac{\mathrm{SA}}{\mathrm{TA}} \times 100\%$ achieved by each method and its query cost. The Optimal and suboptimal results are highlighted. + +
MethodCIFAR10CIFAR100MNISTGTSRBSVHNSTL10F-MNISTQueries
Conventional [34]63.3645.4096.1620.3468.2660.2594.522,500
StolenEncoder [26]79.0368.7898.7663.7092.0176.2995.912,500
Cont-Steal [34]80.4373.9498.0395.5596.3980.0895.39250,000
RDA83.8477.4998.9794.2796.7982.2594.692,500
85.8583.0899.33101.6099.9585.56100.8425,000
+ +Table 5: Comparisons between RDA and StolenEncoder under the same query cost. + +
QueriesMethodsettingSA
2500StolenEncoder2,500 images71.98
RDA500 images×5 patches65.38
625 images×4 patches67.07
1,250 images×2 patches71.44
2,500 images×1 patch76.36
25,000StolenEncoder25,000 images76.23
RDA2,500 images×10 patches78.50
+ +tiveness. Moreover, Table 3 shows that RDA requires much fewer computation resources to steal the two encoders than pre-training them from scratch, owing to the small surrogate dataset and lightweight network architecture. + +# 4.3 Comparison with Existing Methods + +We consider the conventional method (CVPR 2023, proposed as the baseline by [34]), StolenEncoder (CCS 2022) [26], and Cont-Steal (CVPR 2023) [34] as our baselines. A more detailed explanation of them is in Supplementary A.1. Specifically, we compare RDA with them to steal the encoder pre-trained on CIFAR10 under three different scenarios. For a fair comparison, the surrogate dataset remains identical across all baselines (except for some results presented in Table 5 due to different sizes of surrogate datasets). + +Under the Same Surrogate Dataset Size. In this subsection, we fix the surrogate dataset size to 2,500 and make it identical across all methods. As Table 4 shows, RDA outperforms baselines by a significant margin across all seven downstream datasets with a moderate query cost. We also evaluate RDA under a setting that has the same query cost as the conventional method and StolenEncoder (i.e., 2,500 queries), for which we query with each sample once in its origin version. The results show that RDA still surpasses baselines over five out of seven datasets with the least query cost. In addition, the comparison between the results achieved by RDA with 25,000 and 2,500 queries reveals that more patches for generating prototypes will enhance the attack. The embedding space of the surrogate encoder trained by each method is visualized in Figure 11 of Supplementary A.4, which shows that our RDA can train surrogate encoders to discriminate different samples better. Besides, We have summarized the mean + +![](images/68c1f19139c5858039488a50dbd8c8828ced00c55f88def557a15aa2502fb3e0.jpg) +Fig. 7: Heat map of the achieved SAs (better zoom in). This figure shows the performance of 100 combinations of different training epochs and (a) the surrogate dataset size, (b) the patch number for generating prototypes, and (c) the patch number for training. + +![](images/27b122294343f329ad95e5a5f35b0d1b87729066e691de1baaf4ad42c9a4bbde.jpg) + +![](images/33897e5bb8869cd28a205b8bb29ff49b891802d28e985eb34f4d41ff5f411307.jpg) + +value achieved by each method over the seven downstream datasets in Figure 3 to show the superiority of RDA. + +Under the Same Query Cost. To demonstrate the superiority of RDA against baselines under an identical query cost, we consider two fixed values, i.e., 2,500 and 25,000. We take StolenEncoder as the rival since it is the best baseline with the least query cost. We note that Cont-Steal is beyond our consideration because the data volume for it would be too small under such a setting, which is only 25 images for 2,500 queries and 250 images for 25,000 queries. $N$ images $\times n$ patches in the table indicates that we use $N$ images as the surrogate dataset and crop each image into $n$ patches to generate prototypes. Therefore, the query cost of RDA is $N \times n$ . As shown in Table 5, under the same query cost of 2,500, RDA can achieve comparable results with StolenEncoder with half the size of its surrogate dataset. With the same 2,500 images, RDA outperforms StolenEncoder by a significant margin, showcasing its superiority. Moreover, under a query cost of 25,000, RDA outperforms StolenEncoder by $2.14\%$ with only $10\%$ the size of its surrogate dataset. Besides, comparisons between different configurations of RDA under queries of 2,500 indicate that the training data volume has a dominant impact on the performance. + +Under the Same Time Cost. Since each sample is also cropped and augmented into multiple patches for training (i.e., needing to forward encode each sample multiple times), RDA takes the longest time for training. Table 10 in Supplementary A.4 contains the detailed time cost of each method. To make a fair comparison, we prolong the training for another 200 epochs for baseline methods, i.e., 300 epochs in total, which is sufficient to lead their performances to saturate. The trained surrogate encoders are then evaluated on CIFAR10 as the downstream dataset. Figure 6 reveals that even though baseline methods take longer time to train, RDA still outperforms them with the least query cost, and the performance gap can be further widened by a moderate query cost. + +![](images/25c6b349b74595b444bbfd8d7e4fe326e62b9c71dab073086cfae645c9b7a5b5.jpg) +Fig. 8: Ablation studies on the weight of each part in $\mathcal{L}$ , with $\lambda_1 = 1$ and $\lambda_2$ varying. + +![](images/ca8dd28c69d9034c5b3ab2d625676c610c1741385206b9bae714e3a184778157.jpg) + +![](images/38b90e2cc3ec07769caada336e93e464ec65ffaac00afc6fa3ee2f821392afeb.jpg) + +Table 6: Ablation studies on loss functions. The presented results are SAs. + +
Loss TypeCIFAR10STL10CIFAR100MNISTF-MNISTGTSRBSVHNAVG
MSE65.6653.8331.0094.6386.1232.7962.9761.00
Cosine Similarity80.2366.5643.8197.0088.7655.5375.9472.54
KL Divergence77.3565.3040.9896.2387.8257.0469.3870.58
InfoNCE73.9961.6839.2495.7687.8658.8268.8869.43
LD75.5463.243.1696.5289.2765.7771.1572.23
LA79.7065.5943.7996.8989.2455.9676.7072.55
LA+LD79.3966.7644.2796.7489.3262.7573.7473.28
+ +# 4.4 Ablation Studies + +In this subsection, we conduct ablation studies to investigate the impact of the surrogate dataset size, the patch number for generating prototypes (i.e., $n$ in Eq. 1), the patch number for training (i.e., $m$ in Eq. 2, 3, 5, and 6), and loss functions. All the experiments are conducted to steal the encoder pre-trained on CIFAR10. For evaluations, the trained surrogate encoders are assessed on CIFAR10 in the first three ablation studies. While examining the influence of loss functions, we consider seven distinct downstream datasets, as shown in Table 4, to derive a more comprehensive conclusion. Other configurations remain at their default settings, and we present the resulting SAs. + +Surrogate Dataset Size. Recall our default setting where we randomly sample 2,500 images from Tiny ImageNet to construct the surrogate dataset, which is $5\%$ the size of CIFAR-10. We vary the ratio from $1\%$ to $10\%$ to investigate its impact. Figure 7 (a) shows that a larger surrogate dataset will accelerate the training and improve the attack performance. + +Patch Number for Generating Prototypes. We vary the patch number for generating prototypes from 1 to 100 to investigate its impact. Figure 7 (b) shows that there is a tendency for more patches for generating prototypes will improve the attack performance and accelerate the convergence. However, when the patch number reaches 10, more patches do not further improve the attack performance due to the limited surrogate dataset size. + +Patch Number for Training. We vary the patch number for training from 1 to 10 to investigate its impact. Figure 7 (c) reveals a tendency for more patches + +for training will accelerate the training and achieve better results. However, more patches also indicate longer training time and more computation resources. + +Loss Functions. To demonstrate the superiority of the design of our loss function and the necessity of each part, we conduct ablation studies as shown in Table 6. Table 6 shows that $\mathcal{L}_D$ largely surpass other loss functions on GTSRB while $\mathcal{L}_A$ largely surpass other loss functions on SVHN. We hypothesize that $\mathcal{L}_A$ attains optimal outcomes on SVHN due to the pronounced resemblance between SVHN and the surrogate dataset. It appears that $\mathcal{L}_A$ inclines towards inducing the surrogate encoder to overfit the surrogate data, thereby elucidating its suboptimal performance on GTSRB, a dataset characterized by less similarity with the surrogate dataset. On the contrary, $\mathcal{L}_D$ makes the surrogate encoder fit less to the surrogate dataset and has the best result on GTSRB. To further investigate the effect of each part in our loss, we have conducted an ablation experiment about the weight of each part in our loss. As Figure 8 shows, the surrogate encoder will perform better on SVHN while worse on GTSRB when the weight of $\mathcal{L}_A$ increases. We refer to Supplementary A.4 for a more detailed analysis. Combining $\mathcal{L}_D$ and $\mathcal{L}_A$ , our loss is the most robust one, which achieves the best average result over the seven tested datasets. For example, although cosine similarity and $\mathcal{L}_A$ performs well on SVHN (about $76\%$ ), their performance on GTSRB is terrible (about $56\%$ ). Our loss design can improve their performance on GTSRB by about $7\%$ in the cost of less than $3\%$ decrease on SVHN. + +# 4.5 Robustness to Defenses + +In this section, we evaluate the robustness of RDA against three perturbation-based defenses that aim to limit information leakage and one watermarking defense that aims to detect whether a suspected encoder is stolen. The pretraining and downstream datasets both are set to CIFAR10. + +Perturbation-Based Defense. We evaluate three common practices of this type of defense, i.e., adding noise [30], top- $k$ [30] and rounding [44]. + +- Adding noise: Adding noise means that the defender will introduce noise to the original outputs of the target encoder. Following [10], we set the mean of the noise to 0 and vary the standard deviations to control the noise level. +- Top- $k$ : Top- $k$ means that the defender will only output the first $k$ largest number of each embedding from the target encoder and set the rest as 0. We vary the value of $k$ to simulate different perturbation levels. +- Rounding: Rounding truncates each value in the embedding to a specific precision, which we vary to simulate different perturbation levels. + +The results of the three perturbation-based defense methods are summarized in (a), (b), and (c) of Figure 9, respectively. We can observe in Figure 9 (a) that while adding noise can mitigate the model stealing attack, it decreases the utility of the target encoder more significantly. Moreover, Figure 9 (b) shows that though top- $k$ can decrease the efficacy of RDA as the defender lowers the + +![](images/4835adab1adabc6cc1d857acb85922308e7178cc04cf586b498ef38f5396a1a6.jpg) +(a) Adding noise + +![](images/358420117ec56ecb2e34b4f230719cb3d958959c7017bdf62cadfec3efff4fbc.jpg) +(b) Top-k + +![](images/006437a89d4af20256750db59303260d6c8017deeda1896b6b9d2d638a9e45f5.jpg) +(c) Rounding +Fig. 9: The performance of different defense methods on CIFAR10. + +![](images/6659c498ab0c0bba8ea463f127386d1f0f61792d668e6c9d4e21de0e37050eab.jpg) +(d) Watermarking + +value of $k$ , it also severely deteriorates the target encoder's performance. On the other hand, Figure 9 (c) reveals that rounding has a minimal effect on the performance of both the target encoder and attack. Our experimental results demonstrate the robustness of RDA against perturbation-based defenses. + +Watermarking Defense. As shown by Adi et al. [1], backdoors can be used as watermarks to claim the ownership of a model. In this sense, we follow [34] and leverage BadEncoder [18] to embed a backdoor into the target encoder as the watermark. Ideally, if the watermarked encoder is stolen, the surrogate encoder trained via stealing should also be triggered by the defender-specified trigger to exhibit certain behaviors, and thus the ownership can be claimed. Fortunately, Figure 9 (d) shows that RDA can steal the functionality of the target encoder while unlearning the watermark. More specifically, the watermark rate (WR) of the target encoder is $99.87\%$ but only $17.04\%$ for the surrogate encoder when the training converges. This observation indicates that the watermark embedded in the target encoder cannot be preserved by RDA. + +# 5 Conclusion + +In this paper, we have proposed a novel model stealing method against SSL named RDA, which stands for refine, discriminate, and align. Compared with previous methods [26,34], RDA establishes a less biased optimization objective for each training sample and strives to extract more abundant functionality of the target encoder. This empowers RDA to exhibit tremendous effectiveness, efficiency, and robustness across diverse datasets and settings. + +Ethical Concerns and Possible Defenses. We underscore that the misuse of model stealing techniques can jeopardize the privacy and economic rights of embedding service providers. We hope that our work will inspire the development of more sophisticated defense mechanisms to thwart model stealing attacks. A possible defense involves rejection schemes for suspicious queries that appear highly similar, which we have more detailed discussions about in the Supplementary. + +Limitation and Future Work. The query cost of RDA is not yet optimal and can be further reduced by techniques like clustering to establish an identical prototype for multiple similar images. Moreover, more unique (e.g., more efficient) designs of RDA for transformer-based huge models remain largely unexplored. + +Acknowledgements. This work is supported by the National Key R&D Program of China under Grant 2022YFB3103500, the National Natural Science Foundation of China under Grant U20A20176 and 62072062, the National Natural Science Foundation of Chongqing under Grant cstc2022ycjh-bgzxm0031, and the Chongqing Research Program of Basic Research and Frontier Technology (Chongqing Talent) under Grant cstc2024ycjh-bgzxm0048. + +# References + +1. Adi, Y., Baum, C., Cisse, M., Pinkas, B., Keshet, J.: Turning your weakness into a strength: Watermarking deep neural networks by backdoorsing. In: 27th USENIX Security Symposium (USENIX Security 18). pp. 1615-1631 (2018) +2. Bossard, L., Guillaumeun, M., Van Gool, L.: Food-101-mining discriminative components with random forests. In: Computer vision-ECCV 2014: 13th European conference, zurich, Switzerland, September 6-12, 2014, proceedings, part VI 13. pp. 446-461. Springer (2014) +3. Carlini, N., Wagner, D.: Towards evaluating the robustness of neural networks. In: 2017 IEEE symposium on security and privacy (sp). pp. 39-57. IEEE (2017) +4. Caron, M., Touvron, H., Misra, I., Jégou, H., Mairal, J., Bojanowski, P., Joulin, A.: Emerging properties in self-supervised vision transformers. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 9650-9660 (2021) +5. Chen, T., Kornblith, S., Norouzi, M., Hinton, G.: A simple framework for contrastive learning of visual representations. In: International conference on machine learning. pp. 1597-1607. PMLR (2020) +6. Chen, X., Fan, H., Girshick, R., He, K.: Improved baselines with momentum contrastive learning. arXiv preprint arXiv:2003.04297 (2020) +7. Coates, A., Ng, A., Lee, H.: An analysis of single-layer networks in unsupervised feature learning. In: Proceedings of the fourteenth international conference on artificial intelligence and statistics. pp. 215-223. JMLR Workshop and Conference Proceedings (2011) +8. Dai, Y., Chen, Z., Li, J., Heinecke, S., Sun, L., Xu, R.: Tackling data heterogeneity in federated learning with class prototypes. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 37, pp. 7314-7322 (2023) +9. Devlin, J., Chang, M.W., Lee, K., Toutanova, K.: Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805 (2018) +10. Dziedzic, A., Dhawan, N., Kaleem, M.A., Guan, J., Papernot, N.: On the difficulty of defending self-supervised learning against model extraction. In: International Conference on Machine Learning. pp. 5757-5776. PMLR (2022) +11. Grill, J.B., Strub, F., Altché, F., Tallec, C., Richemond, P., Buchatskaya, E., Doersch, C., Avila Pires, B., Guo, Z., Gheshlaghi Azar, M., et al.: Bootstrap your own latent-a new approach to self-supervised learning. Advances in neural information processing systems 33, 21271-21284 (2020) +12. He, K., Fan, H., Wu, Y., Xie, S., Girshick, R.: Momentum contrast for unsupervised visual representation learning. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 9729-9738 (2020) +13. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 770-778 (2016) + +14. Howard, A.G., Zhu, M., Chen, B., Kalenichenko, D., Wang, W., Weyand, T., Andreetto, M., Adam, H.: Mobilenets: Efficient convolutional neural networks for mobile vision applications. arXiv preprint arXiv:1704.04861 (2017) +15. Huang, G., Liu, Z., Van Der Maaten, L., Weinberger, K.Q.: Densely connected convolutional networks. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 4700-4708 (2017) +16. Huang, W., Ye, M., Shi, Z., Li, H., Du, B.: Rethinking federated learning with domain shift: A prototype view. In: 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 16312-16322. IEEE (2023) +17. Jia, H., Choquette-Choo, C.A., Chandrasekaran, V., Papernot, N.: Entangled watermarks as a defense against model extraction. In: 30th USENIX Security Symposium (USENIX Security 21). pp. 1937-1954 (2021) +18. Jia, J., Liu, Y., Gong, N.Z.: Badencoder: Backdoor attacks to pre-trained encoders in self-supervised learning. In: 2022 IEEE Symposium on Security and Privacy (SP). pp. 2043-2059. IEEE (2022) +19. Khosla, P., Teterwak, P., Wang, C., Sarna, A., Tian, Y., Isola, P., Maschinot, A., Liu, C., Krishnan, D.: Supervised contrastive learning. Advances in neural information processing systems 33, 18661-18673 (2020) +20. Kingma, D.P., Ba, J.: Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980 (2014) +21. Krizhevsky, A., Hinton, G., et al.: Learning multiple layers of features from tiny images (2009) +22. Krizhevsky, A., Nair, V., Hinton, G.: Cifar-10 (canadian institute for advanced research). URL http://www.cs.toronto.edu/kriz/cifar.html 5(4), 1 (2010) +23. Le, Y., Yang, X.: Tiny imagenet visual recognition challenge. CS 231N 7(7), 3 (2015) +24. LeCun, Y., Cortes, C.: MNIST handwritten digit database (2010), http://yann.lecun.com/exdb/mnist/ +25. Liu, H., Jia, J., Qu, W., Gong, N.Z.: Encodermi: Membership inference against pre-trained encoders in contrastive learning. In: Proceedings of the 2021 ACM SIGSAC Conference on Computer and Communications Security. pp. 2081-2095 (2021) +26. Liu, Y., Jia, J., Liu, H., Gong, N.Z.: Stolenencoder: stealing pre-trained encoders in self-supervised learning. In: Proceedings of the 2022 ACM SIGSAC Conference on Computer and Communications Security. pp. 2115-2128 (2022) +27. Van der Maaten, L., Hinton, G.: Visualizing data using t-sne. Journal of machine learning research 9(11) (2008) +28. Mettes, P., Van der Pol, E., Snoek, C.: Hyperspherical prototype networks. Advances in neural information processing systems 32 (2019) +29. Netzer, Y., Wang, T., Coates, A., Bissacco, A., Wu, B., Ng, A.Y.: Reading digits in natural images with unsupervised feature learning (2011) +30. Orekondy, T., Schiele, B., Fritz, M.: Knockoff nets: Stealing functionality of black-box models. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 4954-4963 (2019) +31. Papernot, N., McDaniel, P., Goodfellow, I., Jha, S., Celik, Z.B., Swami, A.: Practical black-box attacks against machine learning. In: Proceedings of the 2017 ACM on Asia conference on computer and communications security. pp. 506-519 (2017) +32. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021) + +33. Salem, A., Zhang, Y., Humbert, M., Berrang, P., Fritz, M., Backes, M.: Ml-leaks: Model and data independent membership inference attacks and defenses on machine learning models. arXiv preprint arXiv:1806.01246 (2018) +34. Sha, Z., He, X., Yu, N., Backes, M., Zhang, Y.: Can't steal? cont-steal! contrastive stealing attacks against image encoders. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16373-16383 (2023) +35. Sharir, O., Peleg, B., Shoham, Y.: The cost of training nlp models: A concise overview. arXiv preprint arXiv:2004.08900 (2020) +36. Shokri, R., Stronati, M., Song, C., Shmatikov, V.: Membership inference attacks against machine learning models. In: 2017 IEEE symposium on security and privacy (SP). pp. 3-18. IEEE (2017) +37. Simonyan, K., Zisserman, A.: Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556 (2014) +38. Snell, J., Swersky, K., Zemel, R.: Prototypical networks for few-shot learning. Advances in neural information processing systems 30 (2017) +39. Stallkamp, J., Schlipsing, M., Salmen, J., Igel, C.: Man vs. computer: Benchmarking machine learning algorithms for traffic sign recognition. Neural networks 32, 323-332 (2012) +40. Tan, Y., Long, G., Liu, L., Zhou, T., Lu, Q., Jiang, J., Zhang, C.: Fedproto: Federated prototype learning across heterogeneous clients. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 36, pp. 8432-8440 (2022) +41. Tan, Y., Long, G., Ma, J., Liu, L., Zhou, T., Jiang, J.: Federated learning from pretrained models: A contrastive learning approach. Advances in Neural Information Processing Systems 35, 19332-19344 (2022) +42. Tian, Y., Wang, Y., Krishnan, D., Tenenbaum, J.B., Isola, P.: Rethinking few-shot image classification: a good embedding is all you need? In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XIV 16. pp. 266-282. Springer (2020) +43. Tong, S., Chen, Y., Ma, Y., Lecun, Y.: Emp-ssl: Towards self-supervised learning in one training epoch. arXiv preprint arXiv:2304.03977 (2023) +44. Tramér, F., Zhang, F., Juels, A., Reiter, M.K., Ristenpart, T.: Stealing machine learning models via prediction {APIs}. In: 25th USENIX security symposium (USENIX Security 16). pp. 601-618 (2016) +45. Wu, Z., Xiong, Y., Yu, S.X., Lin, D.: Unsupervised feature learning via nonparametric instance discrimination. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 3733-3742 (2018) +46. Xiao, H., Rasul, K., Vollgraf, R.: Fashion-mnist: a novel image dataset for benchmarking machine learning algorithms. arXiv preprint arXiv:1708.07747 (2017) \ No newline at end of file diff --git a/2024/Refine, Discriminate and Align_ Stealing Encoders via Sample-Wise Prototypes and Multi-Relational Extraction/images.zip b/2024/Refine, Discriminate and Align_ Stealing Encoders via Sample-Wise Prototypes and Multi-Relational Extraction/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..2f94e63e14ae5b3f44e2b05e05f22e4631d292dc --- /dev/null +++ b/2024/Refine, Discriminate and Align_ Stealing Encoders via Sample-Wise Prototypes and Multi-Relational Extraction/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f31db15d1861104b3a43c8356eb4915fcb3cfe21e518cfc7b2ee936c20db3aee +size 420561 diff --git a/2024/Refine, Discriminate and Align_ Stealing Encoders via Sample-Wise Prototypes and Multi-Relational Extraction/layout.json b/2024/Refine, Discriminate and Align_ Stealing Encoders via Sample-Wise Prototypes and Multi-Relational Extraction/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..9b7952b27e1a029a3c7969af22a56877f6e3efcc --- /dev/null +++ b/2024/Refine, Discriminate and Align_ Stealing Encoders via Sample-Wise Prototypes and Multi-Relational Extraction/layout.json @@ -0,0 +1,10493 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 143, + 111, + 473, + 164 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 111, + 473, + 164 + ], + "spans": [ + { + "bbox": [ + 143, + 111, + 473, + 164 + ], + "type": "text", + "content": "Refine, Discriminate and Align: Stealing Encoders via Sample-Wise Prototypes and Multi-Relational Extraction" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 185, + 462, + 223 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 185, + 462, + 223 + ], + "spans": [ + { + "bbox": [ + 146, + 185, + 462, + 223 + ], + "type": "text", + "content": "Shuchi " + }, + { + "bbox": [ + 146, + 185, + 462, + 223 + ], + "type": "inline_equation", + "content": "\\mathbf{W u}^{1}" + }, + { + "bbox": [ + 146, + 185, + 462, + 223 + ], + "type": "text", + "content": " Chuan Ma" + }, + { + "bbox": [ + 146, + 185, + 462, + 223 + ], + "type": "inline_equation", + "content": "^{2\\text{图}}" + }, + { + "bbox": [ + 146, + 185, + 462, + 223 + ], + "type": "text", + "content": " Kang Wei" + }, + { + "bbox": [ + 146, + 185, + 462, + 223 + ], + "type": "inline_equation", + "content": "^{3\\text{图}}" + }, + { + "bbox": [ + 146, + 185, + 462, + 223 + ], + "type": "text", + "content": " Xiaogang Xu" + }, + { + "bbox": [ + 146, + 185, + 462, + 223 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 146, + 185, + 462, + 223 + ], + "type": "text", + "content": " Ming Ding" + }, + { + "bbox": [ + 146, + 185, + 462, + 223 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 146, + 185, + 462, + 223 + ], + "type": "text", + "content": " Yuwen Qian" + }, + { + "bbox": [ + 146, + 185, + 462, + 223 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 146, + 185, + 462, + 223 + ], + "type": "text", + "content": " Di Xiao" + }, + { + "bbox": [ + 146, + 185, + 462, + 223 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 146, + 185, + 462, + 223 + ], + "type": "text", + "content": " Tao Xiang" + }, + { + "bbox": [ + 146, + 185, + 462, + 223 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 146, + 185, + 462, + 223 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 146, + 185, + 462, + 223 + ], + "type": "text", + "content": " NJUST " + }, + { + "bbox": [ + 146, + 185, + 462, + 223 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 146, + 185, + 462, + 223 + ], + "type": "text", + "content": " CQU " + }, + { + "bbox": [ + 146, + 185, + 462, + 223 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 146, + 185, + 462, + 223 + ], + "type": "text", + "content": " PolyU " + }, + { + "bbox": [ + 146, + 185, + 462, + 223 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 146, + 185, + 462, + 223 + ], + "type": "text", + "content": " CUHK " + }, + { + "bbox": [ + 146, + 185, + 462, + 223 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 146, + 185, + 462, + 223 + ], + "type": "text", + "content": " Data61, CSIRO" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 160, + 252, + 455, + 473 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 252, + 455, + 473 + ], + "spans": [ + { + "bbox": [ + 160, + 252, + 455, + 473 + ], + "type": "text", + "content": "Abstract. This paper introduces RDA, a pioneering approach designed to address two primary deficiencies prevalent in previous endeavors aiming at stealing pre-trained encoders: (1) suboptimal performances attributed to biased optimization objectives, and (2) elevated query costs stemming from the end-to-end paradigm that necessitates querying the target encoder every epoch. Specifically, we initially Refine the representations of the target encoder for each training sample, thereby establishing a less biased optimization objective before the steal-training phase. This is accomplished via a sample-wise prototype, which consolidates the target encoder's representations for a given sample's various perspectives. Demanding exponentially fewer queries compared to the end-to-end approach, prototypes can be instantiated to guide subsequent query-free training. For more potent efficacy, we develop a multi-relational extraction loss that trains the surrogate encoder to Discriminate mismatched embedding-prototype pairs while Aligning those matched ones in terms of both amplitude and angle. In this way, the trained surrogate encoder achieves state-of-the-art results across the board in various downstream datasets with limited queries. Moreover, RDA is shown to be robust to multiple widely-used defenses. Our code is available at https://github.com/ShuchiWu/RDA." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 160, + 483, + 436, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 483, + 436, + 495 + ], + "spans": [ + { + "bbox": [ + 160, + 483, + 436, + 495 + ], + "type": "text", + "content": "Keywords: Model Stealing " + }, + { + "bbox": [ + 160, + 483, + 436, + 495 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 483, + 436, + 495 + ], + "type": "text", + "content": " Self-Supervised Learning " + }, + { + "bbox": [ + 160, + 483, + 436, + 495 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 483, + 436, + 495 + ], + "type": "text", + "content": " Prototype" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 534, + 230, + 547 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 534, + 230, + 547 + ], + "spans": [ + { + "bbox": [ + 132, + 534, + 230, + 547 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 559, + 482, + 644 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 559, + 482, + 644 + ], + "spans": [ + { + "bbox": [ + 130, + 559, + 482, + 644 + ], + "type": "text", + "content": "Self-supervised learning (SSL) [4, 5, 9, 11, 12] is endowed with the capability to harness unlabeled data for pre-training a versatile encoder that applicable to a range of downstream tasks, or even showcasing groundbreaking zero-shot performances, e.g., CLIP [32]. However, SSL typically requires a large volume of data and computation resources to achieve a convincing performance, i.e., high-performance encoders are expensive to train [35]. To safeguard the confidentiality and economic worth of these encoders, entities like OpenAI offer their encoders" + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 140, + 653, + 471, + 666 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 140, + 653, + 471, + 666 + ], + "spans": [ + { + "bbox": [ + 140, + 653, + 471, + 666 + ], + "type": "inline_equation", + "content": "\\text{忍}" + }, + { + "bbox": [ + 140, + 653, + 471, + 666 + ], + "type": "text", + "content": " Corresponding Author: chuan.ma@cqu.edu.cn, adam-kang.wei@polyu.edu.hk" + } + ] + } + ], + "index": 6 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 177, + 114, + 233, + 201 + ], + "blocks": [ + { + "bbox": [ + 177, + 114, + 233, + 201 + ], + "lines": [ + { + "bbox": [ + 177, + 114, + 233, + 201 + ], + "spans": [ + { + "bbox": [ + 177, + 114, + 233, + 201 + ], + "type": "image", + "image_path": "102cf00079713b5379c0dbb6a242b65d6ed740f7aa6dffd2a1c596b665e68086.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 207, + 482, + 251 + ], + "lines": [ + { + "bbox": [ + 130, + 207, + 482, + 251 + ], + "spans": [ + { + "bbox": [ + 130, + 207, + 482, + 251 + ], + "type": "text", + "content": "Fig. 1: Illustrations of four stealing methods against SSL. The dotted arrows and text beside interpret how each method optimizes the surrogate encoder. Surrogate encoder branches in (b)-(d) involve data augmentations for training. Both (c) and (d) augment each sample before querying the target encoder but adopting different schemes." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 242, + 116, + 302, + 199 + ], + "blocks": [ + { + "bbox": [ + 242, + 116, + 302, + 199 + ], + "lines": [ + { + "bbox": [ + 242, + 116, + 302, + 199 + ], + "spans": [ + { + "bbox": [ + 242, + 116, + 302, + 199 + ], + "type": "image", + "image_path": "bf5a8f6fe9d7486cf9958564a135c5470f0e61d6b0083e049e9045f42bcdca4b.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 309, + 116, + 369, + 199 + ], + "blocks": [ + { + "bbox": [ + 309, + 116, + 369, + 199 + ], + "lines": [ + { + "bbox": [ + 309, + 116, + 369, + 199 + ], + "spans": [ + { + "bbox": [ + 309, + 116, + 369, + 199 + ], + "type": "image", + "image_path": "30de37fa9d6474e983a08d6720e649e4760880ab9a552717ae81b094d75b0800.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 374, + 115, + 436, + 199 + ], + "blocks": [ + { + "bbox": [ + 374, + 115, + 436, + 199 + ], + "lines": [ + { + "bbox": [ + 374, + 115, + 436, + 199 + ], + "spans": [ + { + "bbox": [ + 374, + 115, + 436, + 199 + ], + "type": "image", + "image_path": "c3e937a76416808419e7ca5890fef7608f73334babc6b6527955b087f1ce55da.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 278, + 480, + 338 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 278, + 480, + 338 + ], + "spans": [ + { + "bbox": [ + 130, + 278, + 480, + 338 + ], + "type": "text", + "content": "as a premium service, exclusively unveiling the service API to the public. Users have the privilege of soliciting embeddings for their data, facilitating the training of diverse models tailored for specific downstream tasks. Regrettably, the substantial value of these encoders and their exposure to publicly accessible APIs render them susceptible to model stealing attacks [10,26,34]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 340, + 482, + 435 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 340, + 482, + 435 + ], + "spans": [ + { + "bbox": [ + 130, + 340, + 482, + 435 + ], + "type": "text", + "content": "Given a surrogate dataset, model stealing attacks aim to mimic the outputs of a target model to train either a high-accuracy copy of comparable performance or a high-fidelity copy that can serve as a stepping stone to perform further attacks like adversarial examples [31], membership inference attacks [33,36], etc. The goal of this paper is to develop an approach that can train a surrogate encoder with competitive performance on downstream tasks by only accessing the target encoder's output embeddings. Meanwhile, it should require as little cost as possible, which is primarily derived from querying the target encoder." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 437, + 482, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 437, + 482, + 628 + ], + "spans": [ + { + "bbox": [ + 130, + 437, + 482, + 628 + ], + "type": "text", + "content": "We begin with systematically studying existing techniques for stealing pretrained encoders. Specifically, the conventional method [34] and StolenEncoder [26] are similar, and both only need to query the target encoder once with each sample before the training. The output embedding can be viewed as a \"ground truth\" or an \"optimization objective\" to the sample, and the surrogate encoder is optimized to output a similar embedding when the same sample is fed, as illustrated in Figure 1 (a). The sole distinction lies in the optimization of StolenEncoder, which incorporates data augmentations (refer to Figure 1 (b)), as it supposes the target encoder will produce similar embeddings for an image and its augmentations. Nevertheless, we contend that this presumed similarity is of a modest degree, given the disparity between the surrogate and pre-training data. The target encoder's representations for such data's various augmentations may diverge and even be biased to other samples, as visually demonstrated in Figure 2 using a toy experiment. In this sense, only using the embedding of an image's single perspective (regardless of the original or augmented version) as the optimization objective is inadvisable." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 629, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 481, + 665 + ], + "type": "text", + "content": "On the other hand, Cont-Steal [34] augments each sample into two perspectives in each epoch: one is used to query the target encoder while the other is fed to the surrogate encoder. Two embeddings of the same sample from the tar" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 216, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 216, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 216, + 100 + ], + "type": "text", + "content": "S. Wu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 248, + 114, + 366, + 203 + ], + "blocks": [ + { + "bbox": [ + 248, + 114, + 366, + 203 + ], + "lines": [ + { + "bbox": [ + 248, + 114, + 366, + 203 + ], + "spans": [ + { + "bbox": [ + 248, + 114, + 366, + 203 + ], + "type": "image", + "image_path": "1860f8975f0322c0f64a15a77cbd8ad0f96ab1be3e835af5f480b2c40c220fd1.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 210, + 482, + 277 + ], + "lines": [ + { + "bbox": [ + 130, + 210, + 482, + 277 + ], + "spans": [ + { + "bbox": [ + 130, + 210, + 482, + 277 + ], + "type": "text", + "content": "Fig. 2: t-SNE of embeddings belonging to five different images generated by an encoder pre-trained on CIFAR10, with each image augmented into 500 patches and fed into the encoder. Each black marker represents the mean of the 500 embeddings of a certain image, i.e., its prototype. Among the embeddings of an image's various augmentations, some can be diverged or even biased. In contrast, each image's prototype is more distinguishable, i.e., less biased." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 303, + 482, + 351 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 303, + 482, + 351 + ], + "spans": [ + { + "bbox": [ + 130, + 303, + 482, + 351 + ], + "type": "text", + "content": "get and surrogate encoders will be aligned, while those of different samples will be pushed apart, as illustrated in Figure 1 (c). In spite of such an end-to-end training scheme performing better, it suffers from high query costs since each sample is augmented and used to query in each epoch." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 352, + 482, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 352, + 482, + 555 + ], + "spans": [ + { + "bbox": [ + 130, + 352, + 482, + 555 + ], + "type": "text", + "content": "To tackle these issues, we propose RDA. Specifically, we first Refine the target encoder's representations for each sample in the surrogate dataset by averagely aggregating its several (e.g., 10) different augmentations' embeddings, i.e., their mean value, to establish a sample-level prototype for it. Then, the sample-wise prototype is used to guide the surrogate encoder optimization, which can mitigate the impact of biased embeddings, as shown in Figure 2 (see black markers). We have an experiment presented in Figure 10 of our supplementary material to further quantitatively reveal the benefit of prototypes, i.e., significantly more similar with each augmentation patch's embedding, showcasing it is less biased. With a prototype, there is a static optimization objective for each sample across the entire training, and thus, it is done in a query-free manner. This enables RDA to have far less query cost than the end-to-end approach (less than " + }, + { + "bbox": [ + 130, + 352, + 482, + 555 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 130, + 352, + 482, + 555 + ], + "type": "text", + "content": "). To further enhance the attack efficacy with limited queries, we develop a multi-relational extraction loss that trains the surrogate encoder to Discriminate mismatched embedding-prototype pairs while Aligning those matched ones in terms of both amplitude and angle. The framework and pipeline of RDA are depicted in Figures 1 (d) and 4, respectively." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 556, + 482, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 556, + 482, + 652 + ], + "spans": [ + { + "bbox": [ + 130, + 556, + 482, + 652 + ], + "type": "text", + "content": "Extensive experiments show that RDA acquires a favorable trade-off between the query cost (a.k.a, the money cost) and attack efficacy. As depicted in Figure 3, when the target encoder is pre-trained on CIFAR10, RDA averagely outperforms the state-of-the-art (SOTA), i.e., Cont-Steal [34], by " + }, + { + "bbox": [ + 130, + 556, + 482, + 652 + ], + "type": "inline_equation", + "content": "1.22\\%" + }, + { + "bbox": [ + 130, + 556, + 482, + 652 + ], + "type": "text", + "content": " on the stealing efficacy over seven different downstream datasets, with a competitive smaller query cost, i.e., only " + }, + { + "bbox": [ + 130, + 556, + 482, + 652 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 130, + 556, + 482, + 652 + ], + "type": "text", + "content": " of Cont-Steal. Further, this performance gap can be widened to " + }, + { + "bbox": [ + 130, + 556, + 482, + 652 + ], + "type": "inline_equation", + "content": "5.20\\%" + }, + { + "bbox": [ + 130, + 556, + 482, + 652 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 130, + 556, + 482, + 652 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 130, + 556, + 482, + 652 + ], + "type": "text", + "content": " query cost of Cont-Steal. In particular, RDA is demonstrated to be robust against multiple prevalent defenses [18,30,44]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 146, + 653, + 358, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 653, + 358, + 665 + ], + "spans": [ + { + "bbox": [ + 146, + 653, + 358, + 665 + ], + "type": "text", + "content": "In conclusion, our contributions are three folds:" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 290, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 290, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 290, + 91, + 447, + 102 + ], + "type": "text", + "content": "Refine, Discriminate and Align (RDA)" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 217, + 117, + 394, + 208 + ], + "blocks": [ + { + "bbox": [ + 217, + 117, + 394, + 208 + ], + "lines": [ + { + "bbox": [ + 217, + 117, + 394, + 208 + ], + "spans": [ + { + "bbox": [ + 217, + 117, + 394, + 208 + ], + "type": "image", + "image_path": "56a5f18e0e51d7821f4c94d07418fb7820597df2269ce1ae8920a67ff055fcfc.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 218, + 482, + 264 + ], + "lines": [ + { + "bbox": [ + 130, + 218, + 482, + 264 + ], + "spans": [ + { + "bbox": [ + 130, + 218, + 482, + 264 + ], + "type": "text", + "content": "Fig. 3: Performance comparisons between four stealing methods against SSL. The presented results are the mean values achieved by each method over seven different downstream classification tasks, with their corresponding query costs. Our proposed RDA can achieve SOTA results with the least query cost." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 287, + 479, + 368 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 138, + 287, + 479, + 320 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 287, + 479, + 320 + ], + "spans": [ + { + "bbox": [ + 138, + 287, + 479, + 320 + ], + "type": "text", + "content": "- We comprehensively investigate two critical inadequacies of existing stealing methods against SSL, i.e., suboptimal efficacy and high query costs, and analyze the causes." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 323, + 479, + 344 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 323, + 479, + 344 + ], + "spans": [ + { + "bbox": [ + 138, + 323, + 479, + 344 + ], + "type": "text", + "content": "- We develop a novel approach to train surrogate encoders, namely RDA, with sample-wise prototype guidance and a multi-relational extraction loss." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 347, + 479, + 368 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 347, + 479, + 368 + ], + "spans": [ + { + "bbox": [ + 138, + 347, + 479, + 368 + ], + "type": "text", + "content": "- Extensive experiments are conducted to verify the effectiveness and robustness of RDA." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 132, + 399, + 237, + 411 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 399, + 237, + 411 + ], + "spans": [ + { + "bbox": [ + 132, + 399, + 237, + 411 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 426, + 479, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 426, + 479, + 533 + ], + "spans": [ + { + "bbox": [ + 130, + 426, + 479, + 533 + ], + "type": "text", + "content": "Prototype Learning. A prototype refers to the mean of embeddings belonging to all images of an identical class, which serves as a proxy of the class [40]. With each class having a prototype, the model is trained to match its output with the corresponding prototype when inputting a sample of a certain class. Prototype learning has proven beneficial for various learning scenarios, e.g., federated learning [8, 16, 41] and few-shot learning [28, 38, 42]. In this paper, the optimization objective (i.e., the mean of embeddings belonging to an identical sample's multiple augmentations) we generate for each training sample is conceptually similar to prototypes, and thus we name it a sample-wise prototype." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 533, + 479, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 479, + 604 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 479, + 604 + ], + "type": "text", + "content": "Model Stealing Attacks against SSL. Dziedzic et al. [10] pointed out that the higher dimension of embeddings leaks more information than labels, making SSL more easily stolen. The primary objective of a stealing attack is to train a surrogate encoder to achieve high accuracy on downstream tasks or recreate a high-fidelity copy that can be used to mount further attacks such as adversarial examples [3] and membership inference attacks [25]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 605, + 496, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 496, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 496, + 665 + ], + "type": "text", + "content": "Defenses against Model Stealing Attacks. Existing defenses against model stealing attacks can be categorized based on when they are applied [10]. Perturbation-based defenses, e.g., adding noise [30], top-" + }, + { + "bbox": [ + 130, + 605, + 496, + 665 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 605, + 496, + 665 + ], + "type": "text", + "content": " [30] and truncating outputs [44], are applied before a stealing attack happens, aiming to limit the information leakage. On the other hand, watermarking defenses embed watermarks or unique" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 216, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 216, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 216, + 100 + ], + "type": "text", + "content": "S. Wu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": "identifiers into the target model and use them to detect whether it is stolen after a stealing attack happens [17]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 158, + 233, + 172 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 158, + 233, + 172 + ], + "spans": [ + { + "bbox": [ + 131, + 158, + 233, + 172 + ], + "type": "text", + "content": "3 Methodology" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 182, + 231, + 193 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 182, + 231, + 193 + ], + "spans": [ + { + "bbox": [ + 131, + 182, + 231, + 193 + ], + "type": "text", + "content": "3.1 Threat Model" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 201, + 482, + 298 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 201, + 482, + 298 + ], + "spans": [ + { + "bbox": [ + 130, + 201, + 482, + 298 + ], + "type": "text", + "content": "Given a target encoder " + }, + { + "bbox": [ + 130, + 201, + 482, + 298 + ], + "type": "inline_equation", + "content": "E_{T}" + }, + { + "bbox": [ + 130, + 201, + 482, + 298 + ], + "type": "text", + "content": ", the attacker aims to train a surrogate encoder " + }, + { + "bbox": [ + 130, + 201, + 482, + 298 + ], + "type": "inline_equation", + "content": "E_{S}" + }, + { + "bbox": [ + 130, + 201, + 482, + 298 + ], + "type": "text", + "content": " at the lowest possible cost, which can perform competitively on downstream tasks with " + }, + { + "bbox": [ + 130, + 201, + 482, + 298 + ], + "type": "inline_equation", + "content": "E_{T}" + }, + { + "bbox": [ + 130, + 201, + 482, + 298 + ], + "type": "text", + "content": ". To achieve this goal, the attacker queries " + }, + { + "bbox": [ + 130, + 201, + 482, + 298 + ], + "type": "inline_equation", + "content": "E_{T}" + }, + { + "bbox": [ + 130, + 201, + 482, + 298 + ], + "type": "text", + "content": " for embeddings of an unlabeled surrogated dataset " + }, + { + "bbox": [ + 130, + 201, + 482, + 298 + ], + "type": "inline_equation", + "content": "D_{S} = \\{\\pmb{x}_{1},\\dots,\\pmb{x}_{N}\\}" + }, + { + "bbox": [ + 130, + 201, + 482, + 298 + ], + "type": "text", + "content": " consisting of " + }, + { + "bbox": [ + 130, + 201, + 482, + 298 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 201, + 482, + 298 + ], + "type": "text", + "content": " images, and guide " + }, + { + "bbox": [ + 130, + 201, + 482, + 298 + ], + "type": "inline_equation", + "content": "E_{S}" + }, + { + "bbox": [ + 130, + 201, + 482, + 298 + ], + "type": "text", + "content": " to mimic the output of " + }, + { + "bbox": [ + 130, + 201, + 482, + 298 + ], + "type": "inline_equation", + "content": "E_{T}" + }, + { + "bbox": [ + 130, + 201, + 482, + 298 + ], + "type": "text", + "content": " for each sample in " + }, + { + "bbox": [ + 130, + 201, + 482, + 298 + ], + "type": "inline_equation", + "content": "D_{S}" + }, + { + "bbox": [ + 130, + 201, + 482, + 298 + ], + "type": "text", + "content": ". Specifically, we consider a black-box setting, where the attacker has direct access only to the outputs of " + }, + { + "bbox": [ + 130, + 201, + 482, + 298 + ], + "type": "inline_equation", + "content": "E_{T}" + }, + { + "bbox": [ + 130, + 201, + 482, + 298 + ], + "type": "text", + "content": " while remaining unaware of its architecture and training configurations, including pre-training datasets, loss functions, data augmentations schemes, etc." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 314, + 285, + 326 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 314, + 285, + 326 + ], + "spans": [ + { + "bbox": [ + 131, + 314, + 285, + 326 + ], + "type": "text", + "content": "3.2 Sample-Wise Prototypes" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 332, + 482, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 332, + 482, + 460 + ], + "spans": [ + { + "bbox": [ + 130, + 332, + 482, + 460 + ], + "type": "text", + "content": "Recalling the discussion in Section 1, we expect to refine the target encoder's representations to find a less biased optimization objective for each sample in " + }, + { + "bbox": [ + 130, + 332, + 482, + 460 + ], + "type": "inline_equation", + "content": "D_{S}" + }, + { + "bbox": [ + 130, + 332, + 482, + 460 + ], + "type": "text", + "content": ". Enlightened by prototype learning [40] and Extreme-Multi-Patch-SSL (EMPSL) [43], which establish a prototype/benchmark for each class/sample, we propose a sample-wise prototype generation method for this purpose. In detail, the process of prototype generation can be divided into three steps as follows: 1 Cropping and augmenting each image " + }, + { + "bbox": [ + 130, + 332, + 482, + 460 + ], + "type": "inline_equation", + "content": "\\pmb{x}_i\\in D_S" + }, + { + "bbox": [ + 130, + 332, + 482, + 460 + ], + "type": "text", + "content": " into " + }, + { + "bbox": [ + 130, + 332, + 482, + 460 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 130, + 332, + 482, + 460 + ], + "type": "text", + "content": " augmentation patches denoted as " + }, + { + "bbox": [ + 130, + 332, + 482, + 460 + ], + "type": "inline_equation", + "content": "\\{\\pmb{x}_{i,t,c}^{\\prime}\\}_{c = 1}^{n} = \\{\\pmb{x}_{i,t,1}^{\\prime},\\dots,\\pmb{x}_{i,t,n}^{\\prime}\\}" + }, + { + "bbox": [ + 130, + 332, + 482, + 460 + ], + "type": "text", + "content": "; 2 Querying " + }, + { + "bbox": [ + 130, + 332, + 482, + 460 + ], + "type": "inline_equation", + "content": "E_{T}" + }, + { + "bbox": [ + 130, + 332, + 482, + 460 + ], + "type": "text", + "content": " with each augmentation patch in " + }, + { + "bbox": [ + 130, + 332, + 482, + 460 + ], + "type": "inline_equation", + "content": "\\{\\pmb{x}_{i,t,c}^{\\prime}\\}_{c = 1}^{n}" + }, + { + "bbox": [ + 130, + 332, + 482, + 460 + ], + "type": "text", + "content": ", resulting in a set of embeddings denoted as " + }, + { + "bbox": [ + 130, + 332, + 482, + 460 + ], + "type": "inline_equation", + "content": "\\{E_T(\\pmb{x}_{i,t,c}^{\\prime})\\}_{c = 1}^{n} = \\{E_T(\\pmb{x}_{i,t,1}^{\\prime}),\\dots,E_T(\\pmb{x}_{i,t,n}^{\\prime})\\}" + }, + { + "bbox": [ + 130, + 332, + 482, + 460 + ], + "type": "text", + "content": "; 3 Calculating the mean of " + }, + { + "bbox": [ + 130, + 332, + 482, + 460 + ], + "type": "inline_equation", + "content": "\\{E_T(\\pmb{x}_{i,t,c}^{\\prime})\\}_{c = 1}^{n}" + }, + { + "bbox": [ + 130, + 332, + 482, + 460 + ], + "type": "text", + "content": " as the prototype " + }, + { + "bbox": [ + 130, + 332, + 482, + 460 + ], + "type": "inline_equation", + "content": "p_{\\pmb{x}_i}" + }, + { + "bbox": [ + 130, + 332, + 482, + 460 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 130, + 332, + 482, + 460 + ], + "type": "inline_equation", + "content": "\\pmb{x}_i" + }, + { + "bbox": [ + 130, + 332, + 482, + 460 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 237, + 466, + 480, + 479 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 237, + 466, + 480, + 479 + ], + "spans": [ + { + "bbox": [ + 237, + 466, + 480, + 479 + ], + "type": "interline_equation", + "content": "p _ {\\boldsymbol {x} _ {i}} = \\frac {1}{n} \\sum_ {c = 1} ^ {n} E _ {T} \\left(\\boldsymbol {x} _ {i, t, c} ^ {\\prime}\\right), \\quad \\boldsymbol {x} _ {i} \\in D _ {s}. \\tag {1}", + "image_path": "67d84cb602161c5503ecdd0e6f4e5d51e99adbfd9b77e9333df2a0ac47e90ade.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 486, + 482, + 571 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 486, + 482, + 571 + ], + "spans": [ + { + "bbox": [ + 130, + 486, + 482, + 571 + ], + "type": "text", + "content": "Each generated prototype will be stored in a memory bank. These sample-wise prototypes can provide a stable optimization objective for each sample throughout the training process. Consequently, the attacker can accomplish the training in a query-free manner. Notably, setting " + }, + { + "bbox": [ + 130, + 486, + 482, + 571 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 130, + 486, + 482, + 571 + ], + "type": "text", + "content": " to a value that is far less than the training epochs, e.g., 10 vs. 100, is sufficient to yield a favorable performance. The prototype generation process is illustrated in step ① of Figure 4. Supplementary A.2 explains how we build the memory bank in practice." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 586, + 327, + 597 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 586, + 327, + 597 + ], + "spans": [ + { + "bbox": [ + 131, + 586, + 327, + 597 + ], + "type": "text", + "content": "3.3 Multi-Relational Extraction Loss" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "text", + "content": "As step ② of Figure 4 shows, during training " + }, + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "inline_equation", + "content": "E_{S}" + }, + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "text", + "content": ", we also perform cropping and augmentation on each image " + }, + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\pmb{x}_i \\in D_S" + }, + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "text", + "content": " to create " + }, + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "text", + "content": " patches, forming an augmentation patch set denoted as " + }, + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\{\\pmb{x}_{i,s,q}^{\\prime}\\}_{q = 1}^{m} = \\{\\pmb{x}_{i,s,1}^{\\prime},\\dots,\\pmb{x}_{i,s,m}^{\\prime}\\}" + }, + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "text", + "content": ". The " + }, + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "text", + "content": " does not necessarily need to be set equal to " + }, + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "text", + "content": ". Next, we feed each augmentation patch from " + }, + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\{\\pmb{x}_{i,s,q}^{\\prime}\\}_{q = 1}^{m}" + }, + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "inline_equation", + "content": "E_{S}" + }, + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "text", + "content": ", obtaining a set of embeddings denoted as" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 290, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 290, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 290, + 91, + 447, + 102 + ], + "type": "text", + "content": "Refine, Discriminate and Align (RDA)" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 136, + 117, + 479, + 192 + ], + "blocks": [ + { + "bbox": [ + 136, + 117, + 479, + 192 + ], + "lines": [ + { + "bbox": [ + 136, + 117, + 479, + 192 + ], + "spans": [ + { + "bbox": [ + 136, + 117, + 479, + 192 + ], + "type": "image", + "image_path": "c1ffdac658369ac2a9dd058449fc982831b16cf26efdf59a466e4c4bad57f29c.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 202, + 482, + 269 + ], + "lines": [ + { + "bbox": [ + 130, + 202, + 482, + 269 + ], + "spans": [ + { + "bbox": [ + 130, + 202, + 482, + 269 + ], + "type": "text", + "content": "Fig. 4: Pipeline of RDA. Prototype generation: augment one sample into " + }, + { + "bbox": [ + 130, + 202, + 482, + 269 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 130, + 202, + 482, + 269 + ], + "type": "text", + "content": " patches and use them to query the target encoder " + }, + { + "bbox": [ + 130, + 202, + 482, + 269 + ], + "type": "inline_equation", + "content": "(E_T)" + }, + { + "bbox": [ + 130, + 202, + 482, + 269 + ], + "type": "text", + "content": ". The mean of the " + }, + { + "bbox": [ + 130, + 202, + 482, + 269 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 130, + 202, + 482, + 269 + ], + "type": "text", + "content": " patches's embeddings is defined as a prototype for this sample. Forward encoding: crop one image into " + }, + { + "bbox": [ + 130, + 202, + 482, + 269 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 130, + 202, + 482, + 269 + ], + "type": "text", + "content": " patches and feed them to the surrogate encoder " + }, + { + "bbox": [ + 130, + 202, + 482, + 269 + ], + "type": "inline_equation", + "content": "(E_S)" + }, + { + "bbox": [ + 130, + 202, + 482, + 269 + ], + "type": "text", + "content": " for their embeddings. Optimization: align embeddings from the surrogate encoder to their matched prototypes in both angle and amplitude while pushing away those belonging to different samples." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 290, + 482, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 290, + 482, + 327 + ], + "spans": [ + { + "bbox": [ + 130, + 290, + 482, + 327 + ], + "type": "inline_equation", + "content": "\\{E_S(\\pmb{x}_{i,s,q}^{\\prime})\\}_{q = 1}^{m} = \\{E_S(\\pmb{x}_{i,s,1}^{\\prime}),\\dots ,E_S(\\pmb{x}_{i,s,m}^{\\prime})\\}" + }, + { + "bbox": [ + 130, + 290, + 482, + 327 + ], + "type": "text", + "content": ". To optimize " + }, + { + "bbox": [ + 130, + 290, + 482, + 327 + ], + "type": "inline_equation", + "content": "E_{S}" + }, + { + "bbox": [ + 130, + 290, + 482, + 327 + ], + "type": "text", + "content": ", we develop a multi-relational extraction loss composed of two parts, i.e., the discriminating loss and the aligning loss." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 326, + 482, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 326, + 482, + 450 + ], + "spans": [ + { + "bbox": [ + 130, + 326, + 482, + 450 + ], + "type": "text", + "content": "Discriminating Loss. The discriminating loss is responsible for training " + }, + { + "bbox": [ + 130, + 326, + 482, + 450 + ], + "type": "inline_equation", + "content": "E_{S}" + }, + { + "bbox": [ + 130, + 326, + 482, + 450 + ], + "type": "text", + "content": " to distinguish between different samples, as depicted in step ③ of Figure 4. For this goal, it pushes each embedding in " + }, + { + "bbox": [ + 130, + 326, + 482, + 450 + ], + "type": "inline_equation", + "content": "\\{E_S(\\pmb{x}_{i,s,q}')\\}_{q = 1}^m" + }, + { + "bbox": [ + 130, + 326, + 482, + 450 + ], + "type": "text", + "content": " away from mismatched prototypes " + }, + { + "bbox": [ + 130, + 326, + 482, + 450 + ], + "type": "inline_equation", + "content": "p_{\\pmb{x}_j}, i \\neq j" + }, + { + "bbox": [ + 130, + 326, + 482, + 450 + ], + "type": "text", + "content": " (denoted as negative pairs). In this regard, contrastive learning [5,6,12,19] offers a solution, as it can pull embeddings of positive pairs close while pushing embeddings of negative pairs apart. Furthermore, since more negative pairs can help improve the performance of contrastive learning [5], we refer to the loss designed by Sha et al. [34] to propose our discriminating loss " + }, + { + "bbox": [ + 130, + 326, + 482, + 450 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_D" + }, + { + "bbox": [ + 130, + 326, + 482, + 450 + ], + "type": "text", + "content": ". In particular, " + }, + { + "bbox": [ + 130, + 326, + 482, + 450 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_D" + }, + { + "bbox": [ + 130, + 326, + 482, + 450 + ], + "type": "text", + "content": " considers both mismatched prototype-embedding pairs from different samples as well as embeddings from " + }, + { + "bbox": [ + 130, + 326, + 482, + 450 + ], + "type": "inline_equation", + "content": "E_{S}" + }, + { + "bbox": [ + 130, + 326, + 482, + 450 + ], + "type": "text", + "content": " for different samples as negative pairs. Formally, " + }, + { + "bbox": [ + 130, + 326, + 482, + 450 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_D" + }, + { + "bbox": [ + 130, + 326, + 482, + 450 + ], + "type": "text", + "content": " can be expressed as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 209, + 455, + 481, + 469 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 209, + 455, + 481, + 469 + ], + "spans": [ + { + "bbox": [ + 209, + 455, + 481, + 469 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {p o s} \\left(\\boldsymbol {x} _ {i}\\right) = \\frac {1}{m} \\sum_ {q = 1} ^ {m} \\exp \\left(\\operatorname {s i m} \\left(E _ {S} \\left(\\boldsymbol {x} _ {i, s, q} ^ {\\prime}\\right), p _ {\\boldsymbol {x} _ {i}}\\right) / \\tau\\right), \\tag {2}", + "image_path": "b766d866b74f07f7687c677820342cdc4ae4b1e73d5003603365aa3b795bdbe6.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 209, + 483, + 481, + 521 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 209, + 483, + 481, + 521 + ], + "spans": [ + { + "bbox": [ + 209, + 483, + 481, + 521 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {n e g} \\left(\\boldsymbol {x} _ {i}\\right) = \\frac {1}{m} \\sum_ {q = 1} ^ {m} \\sum_ {k = 1} ^ {N} \\mathbb {1} _ {[ i \\neq k ]} \\left(\\exp \\left(s i m \\left(E _ {S} \\left(\\boldsymbol {x} _ {i, s, q} ^ {\\prime}\\right), p _ {\\boldsymbol {x} _ {k}}\\right) / \\tau\\right) \\right. \\tag {3} \\\\ + \\exp \\left(\\operatorname {s i m} \\left(E _ {S} \\left(\\boldsymbol {x} _ {i, s, q} ^ {\\prime}\\right), E _ {S} \\left(\\boldsymbol {x} _ {k, s, q} ^ {\\prime}\\right)\\right) / \\tau\\right)), \\\\ \\end{array}", + "image_path": "92acd032459e3fa71e2280212c14c4ca2cd219cbc2f5305f055d46cc00eb2ffd.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 253, + 525, + 481, + 541 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 253, + 525, + 481, + 541 + ], + "spans": [ + { + "bbox": [ + 253, + 525, + 481, + 541 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {D} = - \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\log \\frac {\\mathcal {L} _ {p o s} (\\boldsymbol {x} _ {i})}{\\mathcal {L} _ {n e g} (\\boldsymbol {x} _ {i})}, \\tag {4}", + "image_path": "4f5dd55ca9c69aa7e66433f3539e561651978562db6cbb63e6786f327be90938.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 543, + 482, + 566 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 543, + 482, + 566 + ], + "spans": [ + { + "bbox": [ + 130, + 543, + 482, + 566 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 543, + 482, + 566 + ], + "type": "inline_equation", + "content": "\\operatorname{sim}(u, v)" + }, + { + "bbox": [ + 130, + 543, + 482, + 566 + ], + "type": "text", + "content": " represents the cosine similarity between " + }, + { + "bbox": [ + 130, + 543, + 482, + 566 + ], + "type": "inline_equation", + "content": "u" + }, + { + "bbox": [ + 130, + 543, + 482, + 566 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 543, + 482, + 566 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 130, + 543, + 482, + 566 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 543, + 482, + 566 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 130, + 543, + 482, + 566 + ], + "type": "text", + "content": " is a temperature parameter." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 566, + 482, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 566, + 482, + 652 + ], + "spans": [ + { + "bbox": [ + 130, + 566, + 482, + 652 + ], + "type": "text", + "content": "Aligning Loss. While " + }, + { + "bbox": [ + 130, + 566, + 482, + 652 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_D" + }, + { + "bbox": [ + 130, + 566, + 482, + 652 + ], + "type": "text", + "content": " can separate different samples effectively, it does not fully leverage the potential of prototypes. To further enhance the attack efficacy, we propose to align embeddings from " + }, + { + "bbox": [ + 130, + 566, + 482, + 652 + ], + "type": "inline_equation", + "content": "E_S" + }, + { + "bbox": [ + 130, + 566, + 482, + 652 + ], + "type": "text", + "content": " more thoroughly with their matched prototypes from the high-performance " + }, + { + "bbox": [ + 130, + 566, + 482, + 652 + ], + "type": "inline_equation", + "content": "E_T" + }, + { + "bbox": [ + 130, + 566, + 482, + 652 + ], + "type": "text", + "content": ", i.e., in terms of both amplitude and angle, as step ③ of Figure 4 shows. To measure the amplitude and angle deviations between two embeddings, we employ the mean square error (MSE) and cosine similarity to quantify them respectively, which can be formulated as follows:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 227, + 653, + 481, + 667 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 227, + 653, + 481, + 667 + ], + "spans": [ + { + "bbox": [ + 227, + 653, + 481, + 667 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {a m p} ^ {\\prime} \\left(\\boldsymbol {x} _ {i}\\right) = \\frac {1}{m} \\sum_ {q = 1} ^ {m} \\left\\| E _ {S} \\left(\\boldsymbol {x} _ {i, s, q} ^ {\\prime}\\right) - p _ {\\boldsymbol {x} _ {i}} \\right\\| ^ {2}, \\tag {5}", + "image_path": "ab847b25819e67596ee68c8f6a57e3d03744575e3c5af62e811734bb59811c6c.jpg" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 216, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 216, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 216, + 100 + ], + "type": "text", + "content": "S. Wu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 219, + 127, + 481, + 141 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 219, + 127, + 481, + 141 + ], + "spans": [ + { + "bbox": [ + 219, + 127, + 481, + 141 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {a n g}} ^ {\\prime} \\left(\\boldsymbol {x} _ {i}\\right) = \\frac {1}{m} \\sum_ {q = 1} ^ {m} \\operatorname {s i m} \\left(E _ {S} \\left(\\boldsymbol {x} _ {i, s, q} ^ {\\prime}\\right), p _ {\\boldsymbol {x} _ {i}}\\right). \\tag {6}", + "image_path": "e4b7475863e175ec89be04620f9a9f062d4fd2f457128ee3026532b0da87b1c1.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 143, + 482, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 143, + 482, + 240 + ], + "spans": [ + { + "bbox": [ + 131, + 143, + 482, + 240 + ], + "type": "text", + "content": "Notice that we will normalize the value of " + }, + { + "bbox": [ + 131, + 143, + 482, + 240 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{ang}^{\\prime}" + }, + { + "bbox": [ + 131, + 143, + 482, + 240 + ], + "type": "text", + "content": " to [0, 1]. However, there is an issue concerning the equal penalty given by " + }, + { + "bbox": [ + 131, + 143, + 482, + 240 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{amp}^{\\prime}" + }, + { + "bbox": [ + 131, + 143, + 482, + 240 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 143, + 482, + 240 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{ang}^{\\prime}" + }, + { + "bbox": [ + 131, + 143, + 482, + 240 + ], + "type": "text", + "content": " to every deviation increase. For instance, when the MSE increases from 0.8 to 0.9 or from 0.3 to 0.4, both penalties given by " + }, + { + "bbox": [ + 131, + 143, + 482, + 240 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{amp}^{\\prime}" + }, + { + "bbox": [ + 131, + 143, + 482, + 240 + ], + "type": "text", + "content": " are 0.1, which is imprudent. On the contrary, we should have a more rigorous penalty for the increase in MSE from 0.3 to 0.4, as a smaller MSE is more desirable. Likewise, the same penalizing regime should be applied on " + }, + { + "bbox": [ + 131, + 143, + 482, + 240 + ], + "type": "inline_equation", + "content": "1 / \\mathcal{L}_{ang}^{\\prime}" + }, + { + "bbox": [ + 131, + 143, + 482, + 240 + ], + "type": "text", + "content": " for the same reason. To this end, we adopt a logarithmic function to redefine the ultimate formulations of " + }, + { + "bbox": [ + 131, + 143, + 482, + 240 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{amp}" + }, + { + "bbox": [ + 131, + 143, + 482, + 240 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 143, + 482, + 240 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{ang}" + }, + { + "bbox": [ + 131, + 143, + 482, + 240 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 255, + 250, + 481, + 262 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 255, + 250, + 481, + 262 + ], + "spans": [ + { + "bbox": [ + 255, + 250, + 481, + 262 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {a m p}} \\left(\\boldsymbol {x} _ {i}\\right) = \\log \\mathcal {L} _ {\\text {a m p}} ^ {\\prime} \\left(\\boldsymbol {x} _ {i}\\right), \\tag {7}", + "image_path": "e9062342d510db5cda5e587be38c2eb1201486739f31b360394f30225709c4b4.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 219, + 285, + 480, + 297 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 219, + 285, + 480, + 297 + ], + "spans": [ + { + "bbox": [ + 219, + 285, + 480, + 297 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {a n g} \\left(\\boldsymbol {x} _ {i}\\right) = \\log \\left(1 / \\mathcal {L} _ {a n g} ^ {\\prime} \\left(\\boldsymbol {x} _ {i}\\right)\\right) = - \\log \\mathcal {L} _ {a n g} ^ {\\prime} \\left(\\boldsymbol {x} _ {i}\\right). \\tag {8}", + "image_path": "e2d997859e9b4b22448a5fe7d959653e6f8998bca51dab3473aaeec1f8d4f413.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 300, + 479, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 300, + 479, + 323 + ], + "spans": [ + { + "bbox": [ + 131, + 300, + 479, + 323 + ], + "type": "text", + "content": "We posit the amplitude and angle deviations hold equal importance, and formulate the aligning loss as follows:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 194, + 332, + 481, + 357 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 194, + 332, + 481, + 357 + ], + "spans": [ + { + "bbox": [ + 194, + 332, + 481, + 357 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {A} = \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\left(\\mathcal {L} _ {a m p} (\\boldsymbol {x} _ {i}) + \\mathcal {L} _ {a n g} (\\boldsymbol {x} _ {i})\\right) = - \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\log \\frac {\\mathcal {L} _ {a n g} ^ {\\prime} (\\boldsymbol {x} _ {i})}{\\mathcal {L} _ {a m p} ^ {\\prime} (\\boldsymbol {x} _ {i})}. \\tag {9}", + "image_path": "710e4d643a5a50e8a120668a41239d561ae07ab60707dd0544b993714fe48231.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 364, + 479, + 388 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 364, + 479, + 388 + ], + "spans": [ + { + "bbox": [ + 132, + 364, + 479, + 388 + ], + "type": "text", + "content": "Our experiments in Supplementary A.4 demonstrate the combination of " + }, + { + "bbox": [ + 132, + 364, + 479, + 388 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{amp}" + }, + { + "bbox": [ + 132, + 364, + 479, + 388 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 132, + 364, + 479, + 388 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{ang}" + }, + { + "bbox": [ + 132, + 364, + 479, + 388 + ], + "type": "text", + "content": " offers superior results than using each solely." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 146, + 388, + 455, + 400 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 388, + 455, + 400 + ], + "spans": [ + { + "bbox": [ + 146, + 388, + 455, + 400 + ], + "type": "text", + "content": "Lastly, the loss function for stealing pre-trained encoders is as follows:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 263, + 411, + 480, + 423 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 263, + 411, + 480, + 423 + ], + "spans": [ + { + "bbox": [ + 263, + 411, + 480, + 423 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\lambda_ {1} \\cdot \\mathcal {L} _ {D} + \\lambda_ {2} \\cdot \\mathcal {L} _ {A}, \\tag {10}", + "image_path": "dea898230bd3c87f55456a1de63d82b52b8a688e8f0ed877337df3fc7f83b181.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 131, + 432, + 479, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 432, + 479, + 479 + ], + "spans": [ + { + "bbox": [ + 131, + 432, + 479, + 479 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 432, + 479, + 479 + ], + "type": "inline_equation", + "content": "\\lambda_{1}" + }, + { + "bbox": [ + 131, + 432, + 479, + 479 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 432, + 479, + 479 + ], + "type": "inline_equation", + "content": "\\lambda_{2}" + }, + { + "bbox": [ + 131, + 432, + 479, + 479 + ], + "type": "text", + "content": " are preset coefficients to adjust the weight of each part in " + }, + { + "bbox": [ + 131, + 432, + 479, + 479 + ], + "type": "inline_equation", + "content": "\\mathcal{L}" + }, + { + "bbox": [ + 131, + 432, + 479, + 479 + ], + "type": "text", + "content": ". To show the superiority of our design, especially on the loss function, we have conducted ablation studies in Section 4.4 and explored several alternative designs in Supplementary A.3." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 146, + 480, + 479, + 492 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 480, + 479, + 492 + ], + "spans": [ + { + "bbox": [ + 146, + 480, + 479, + 492 + ], + "type": "text", + "content": "Detailed steps of RDA are summarized in Algorithm 1 of Supplementary A.2." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 511, + 229, + 525 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 511, + 229, + 525 + ], + "spans": [ + { + "bbox": [ + 132, + 511, + 229, + 525 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 536, + 261, + 549 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 536, + 261, + 549 + ], + "spans": [ + { + "bbox": [ + 132, + 536, + 261, + 549 + ], + "type": "text", + "content": "4.1 Experimental Setup" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 557, + 480, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 557, + 480, + 641 + ], + "spans": [ + { + "bbox": [ + 130, + 557, + 480, + 641 + ], + "type": "text", + "content": "Target Encoder Settings. We use SimCLR [5] to pre-train two ResNet18 [13] encoders on CIFAR10 [21] and STL10 [7], respectively, as two medium-scale target encoders. Furthermore, we consider two real-world large-scale ResNet50 encoders as the targets, i.e., the ImageNet encoder pre-trained by Google [5], and the CLIP encoder pre-trained by OpenAI [32]. Besides the ResNet family, RDA also has been demonstrated effective upon other backbones, i.e., VGG19_bn [37], DenseNet121 [15], and MobileNetV2 [14], in Supplementary A.4." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 131, + 641, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 641, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 131, + 641, + 481, + 665 + ], + "type": "text", + "content": "Attack Settings. Regarding the surrogate dataset, it is derived from Tiny ImageNet [23]. Specifically, we randomly sample 2,500 images and resize them to" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 290, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 290, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 290, + 91, + 447, + 102 + ], + "type": "text", + "content": "Refine, Discriminate and Align (RDA)" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 99 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 99 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 99 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 156, + 146, + 299, + 228 + ], + "blocks": [ + { + "bbox": [ + 147, + 114, + 465, + 137 + ], + "lines": [ + { + "bbox": [ + 147, + 114, + 465, + 137 + ], + "spans": [ + { + "bbox": [ + 147, + 114, + 465, + 137 + ], + "type": "text", + "content": "Table 1: Results of RDA against two Table 2: Results of RDA against the medium-scale encoders. ImageNet Encoder and CLIP." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 156, + 146, + 299, + 228 + ], + "lines": [ + { + "bbox": [ + 156, + 146, + 299, + 228 + ], + "spans": [ + { + "bbox": [ + 156, + 146, + 299, + 228 + ], + "type": "table", + "html": "
Pre-training DatasetDownstream DatasetTASA\\( \\frac{\\mathrm{{SA}}}{\\mathrm{{TA}}} \\times {100}\\% \\)
CIFAR10MNIST97.2796.6299.33
F-MNIST88.5889.32100.84
GTSRB61.7662.75101.60
SVHN73.7873.7499.95
STL10MNIST96.8496.3199.45
F-MNIST90.0887.9197.59
GTSRB64.4557.7089.53
SVHN61.8570.67114.26
", + "image_path": "1c9d9b33a6eb9cda7a520e664af7331753f4c5f698f9a566e593dcf4d043291f.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 318, + 147, + 454, + 228 + ], + "blocks": [ + { + "bbox": [ + 318, + 147, + 454, + 228 + ], + "lines": [ + { + "bbox": [ + 318, + 147, + 454, + 228 + ], + "spans": [ + { + "bbox": [ + 318, + 147, + 454, + 228 + ], + "type": "table", + "html": "
TargetEncoderDownstreamDatasetTASA\\( \\frac{\\mathrm{{SA}}}{\\mathrm{{TA}}} \\times {100}\\% \\)
ImageNetEncoderMNIST97.3895.8998.47
F-MNIST91.6790.8399.08
GTSRB63.1459.6294.43
SVHN73.2069.2194.55
CLIPMNIST97.9093.9695.98
F-MNIST89.9287.4397.23
GTSRB68.6155.8481.39
SVHN69.5357.5882.81
", + "image_path": "cd311cee8f47c44e8c749bbe8bffccb314d06ddf1be282d6564a485a2eabfa14.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 227, + 270, + 386, + 310 + ], + "blocks": [ + { + "bbox": [ + 130, + 237, + 480, + 259 + ], + "lines": [ + { + "bbox": [ + 130, + 237, + 480, + 259 + ], + "spans": [ + { + "bbox": [ + 130, + 237, + 480, + 259 + ], + "type": "text", + "content": "Table 3: Computation resources required by pre-training the two targeted real-world encoders from scratch and RDA to steal them." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 227, + 270, + 386, + 310 + ], + "lines": [ + { + "bbox": [ + 227, + 270, + 386, + 310 + ], + "spans": [ + { + "bbox": [ + 227, + 270, + 386, + 310 + ], + "type": "table", + "html": "
TargetEncoderPre-trainingRDA
HardwareTime (hrs)HardwareTime (hrs)
ImageNetEncoderTPU v3192RTX A500010.8
CLIPV100 GPU255,74416.2
", + "image_path": "1b040d2d994bbab0479ac50661105c46e5185032677da2d1426ed08c12753bad.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 331, + 480, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 331, + 480, + 415 + ], + "spans": [ + { + "bbox": [ + 130, + 331, + 480, + 415 + ], + "type": "inline_equation", + "content": "32 \\times 32" + }, + { + "bbox": [ + 130, + 331, + 480, + 415 + ], + "type": "text", + "content": " for stealing encoders pre-trained on CIFAR10 and STL10. When stealing the ImageNet encoder and CLIP, the number is 40,000 and 60,000, respectively, with each image resized to " + }, + { + "bbox": [ + 130, + 331, + 480, + 415 + ], + "type": "inline_equation", + "content": "224 \\times 224" + }, + { + "bbox": [ + 130, + 331, + 480, + 415 + ], + "type": "text", + "content": ". For the surrogate encoder architecture, we adopt a ResNet18 across our experiments. During the attack, we set " + }, + { + "bbox": [ + 130, + 331, + 480, + 415 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 130, + 331, + 480, + 415 + ], + "type": "text", + "content": " in Eq. 1 as 10, " + }, + { + "bbox": [ + 130, + 331, + 480, + 415 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 130, + 331, + 480, + 415 + ], + "type": "text", + "content": " in Eq. 2, 3, 5, and 6 as 5, " + }, + { + "bbox": [ + 130, + 331, + 480, + 415 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 130, + 331, + 480, + 415 + ], + "type": "text", + "content": " in Eq. 2 and 3 as 0.07, and both " + }, + { + "bbox": [ + 130, + 331, + 480, + 415 + ], + "type": "inline_equation", + "content": "\\lambda_{1}" + }, + { + "bbox": [ + 130, + 331, + 480, + 415 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 331, + 480, + 415 + ], + "type": "inline_equation", + "content": "\\lambda_{2}" + }, + { + "bbox": [ + 130, + 331, + 480, + 415 + ], + "type": "text", + "content": " in Eq. 10 as 1 unless stated otherwise. For training, the batch size is set as 100, and we employ an Adam optimizer [20] with a learning rate of 0.001." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 415, + 481, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 415, + 481, + 548 + ], + "spans": [ + { + "bbox": [ + 130, + 415, + 481, + 548 + ], + "type": "text", + "content": "Evaluation Settings. We train each surrogate encoder for 100 epochs and test its KNN accuracy [45] on CIFAR10 after each epoch. The best-trained surrogate encoders, as well as target encoders, will be used to train downstream classifiers for the linear probing evaluation. Each downstream classifier will be trained for 100 epochs with an Adam optimizer and a learning rate of 0.0001. Specifically, we totally consider seven downstream datasets, namely MNIST [24], CIFAR10 [21], STL10 [7], GTSRB [39], CIFAR100 [22], SVHN [29], and F-MNIST [46]. Additional small-scale experiments on more complex datasets (e.g., Food 101 [2]) are included in our supplementary material. As done in [26], we use Target Accuracy (TA) to evaluate target encoders, Steal Accuracy (SA) to evaluate surrogate encoders, and " + }, + { + "bbox": [ + 130, + 415, + 481, + 548 + ], + "type": "inline_equation", + "content": "\\frac{\\mathrm{SA}}{\\mathrm{TA}} \\times 100\\%" + }, + { + "bbox": [ + 130, + 415, + 481, + 548 + ], + "type": "text", + "content": " to evaluate the efficacy of the stealing attack." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 563, + 266, + 574 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 563, + 266, + 574 + ], + "spans": [ + { + "bbox": [ + 131, + 563, + 266, + 574 + ], + "type": "text", + "content": "4.2 Effectiveness of RDA" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 581, + 482, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 581, + 482, + 668 + ], + "spans": [ + { + "bbox": [ + 130, + 581, + 482, + 668 + ], + "type": "text", + "content": "Stealing Medium-Scale Encoders. Table 1 shows the results achieved by RDA upon two medium-scale target encoders pre-trained on CIFAR10 and STL10, respectively. As the results show, using a surrogate dataset of " + }, + { + "bbox": [ + 130, + 581, + 482, + 668 + ], + "type": "inline_equation", + "content": "5\\%" + }, + { + "bbox": [ + 130, + 581, + 482, + 668 + ], + "type": "text", + "content": " the size of CIFAR10 (2,500 / 50,000) and " + }, + { + "bbox": [ + 130, + 581, + 482, + 668 + ], + "type": "inline_equation", + "content": "2.4\\%" + }, + { + "bbox": [ + 130, + 581, + 482, + 668 + ], + "type": "text", + "content": " the size of STL10 (2,500 / 105,000), RDA can steal nearly " + }, + { + "bbox": [ + 130, + 581, + 482, + 668 + ], + "type": "inline_equation", + "content": "100\\%" + }, + { + "bbox": [ + 130, + 581, + 482, + 668 + ], + "type": "text", + "content": " of the functionality of both target encoders, except for the GTSRB scenario when stealing the encoder pre-trained on STL10. Nonetheless, even in this exceptional case, the obtained " + }, + { + "bbox": [ + 130, + 581, + 482, + 668 + ], + "type": "inline_equation", + "content": "\\frac{\\mathrm{SA}}{\\mathrm{TA}} \\times 100\\%" + }, + { + "bbox": [ + 130, + 581, + 482, + 668 + ], + "type": "text", + "content": " is approximately " + }, + { + "bbox": [ + 130, + 581, + 482, + 668 + ], + "type": "inline_equation", + "content": "90\\%" + }, + { + "bbox": [ + 130, + 581, + 482, + 668 + ], + "type": "text", + "content": "," + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 216, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 216, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 216, + 100 + ], + "type": "text", + "content": "S. Wu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 138, + 116, + 203, + 174 + ], + "blocks": [ + { + "bbox": [ + 138, + 116, + 203, + 174 + ], + "lines": [ + { + "bbox": [ + 138, + 116, + 203, + 174 + ], + "spans": [ + { + "bbox": [ + 138, + 116, + 203, + 174 + ], + "type": "image", + "image_path": "f4c9168e25fc959d5a4b810ec6aed0b592ce8c78b8fbef63ffbce383fa0ae723.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 137, + 178, + 205, + 187 + ], + "lines": [ + { + "bbox": [ + 137, + 178, + 205, + 187 + ], + "spans": [ + { + "bbox": [ + 137, + 178, + 205, + 187 + ], + "type": "text", + "content": "(a) ImageNet Encoder" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 207, + 116, + 271, + 174 + ], + "blocks": [ + { + "bbox": [ + 207, + 116, + 271, + 174 + ], + "lines": [ + { + "bbox": [ + 207, + 116, + 271, + 174 + ], + "spans": [ + { + "bbox": [ + 207, + 116, + 271, + 174 + ], + "type": "image", + "image_path": "826d2dc194a694fd6ef313d3fa012d06054fb999154e96f0a09da433f2a4d946.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 227, + 180, + 252, + 189 + ], + "lines": [ + { + "bbox": [ + 227, + 180, + 252, + 189 + ], + "spans": [ + { + "bbox": [ + 227, + 180, + 252, + 189 + ], + "type": "text", + "content": "(b) CLIP" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 310, + 114, + 445, + 213 + ], + "blocks": [ + { + "bbox": [ + 133, + 198, + 276, + 241 + ], + "lines": [ + { + "bbox": [ + 133, + 198, + 276, + 241 + ], + "spans": [ + { + "bbox": [ + 133, + 198, + 276, + 241 + ], + "type": "text", + "content": "Fig.5: t-SNE of embeddings of 2,000 images sampled from CIFAR10 generated by the ImageNet encoder and CLIP." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 310, + 114, + 445, + 213 + ], + "lines": [ + { + "bbox": [ + 310, + 114, + 445, + 213 + ], + "spans": [ + { + "bbox": [ + 310, + 114, + 445, + 213 + ], + "type": "image", + "image_path": "76f52550ead86fafa6ba6471d1896e908607bce1958d55ee71f2d236d97a3d0b.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 276, + 221, + 479, + 244 + ], + "lines": [ + { + "bbox": [ + 276, + 221, + 479, + 244 + ], + "spans": [ + { + "bbox": [ + 276, + 221, + 479, + 244 + ], + "type": "text", + "content": "Fig. 6: The training epochs of RDA are fixed to 100 while those of baselines are prolonged to 300." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 266, + 480, + 326 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 266, + 480, + 326 + ], + "spans": [ + { + "bbox": [ + 130, + 266, + 480, + 326 + ], + "type": "text", + "content": "which exemplifies the remarkable effectiveness of RDA. Furthermore, the surrogate encoder trained through RDA even outperforms the target encoder on multiple datasets. Our interpretation of this is that the target encoder fits the pre-training dataset more strongly compared to the RDA-trained surrogate encoder, thus exhibiting inferior performance on some out-of-distribution datasets." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 342, + 482, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 342, + 482, + 668 + ], + "spans": [ + { + "bbox": [ + 130, + 342, + 482, + 668 + ], + "type": "text", + "content": "Stealing Real-World Large-Scale Encoders. We further evaluate RDA on two real-world pre-trained encoders, i.e., the ImageNet encoder and CLIP. We aim to demonstrate the effectiveness of RDA upon such large-scale encoders trained with a massive amount of data. Specifically, the ImageNet encoder is pre-trained on the ImageNet dataset, which has 1.3 million images. The CLIP encoder is pre-trained on a web-scale dataset that consists of 400 million image-text pairs. During the attack, we set the patch number for training as 1 (i.e., " + }, + { + "bbox": [ + 130, + 342, + 482, + 668 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 130, + 342, + 482, + 668 + ], + "type": "text", + "content": " in Eq. 2, 3, 5, and 6) to save time. Other parameters follow the default setting except the " + }, + { + "bbox": [ + 130, + 342, + 482, + 668 + ], + "type": "inline_equation", + "content": "\\lambda_{1}" + }, + { + "bbox": [ + 130, + 342, + 482, + 668 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 342, + 482, + 668 + ], + "type": "inline_equation", + "content": "\\lambda_{2}" + }, + { + "bbox": [ + 130, + 342, + 482, + 668 + ], + "type": "text", + "content": " when stealing CLIP, in which we set " + }, + { + "bbox": [ + 130, + 342, + 482, + 668 + ], + "type": "inline_equation", + "content": "\\lambda_{1}" + }, + { + "bbox": [ + 130, + 342, + 482, + 668 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 342, + 482, + 668 + ], + "type": "inline_equation", + "content": "\\lambda_{2}" + }, + { + "bbox": [ + 130, + 342, + 482, + 668 + ], + "type": "text", + "content": " to 1 and 20, respectively. To interpret the reason for this adjustment, we randomly sample 200 images from each class in the testing set of CIFAR10 and visualize the output embeddings from the ImageNet encoder and CLIP using t-SNE [27]. Figure 5 shows that embeddings from the ImageNet encoder are more uniformly distributed compared to those from CLIP. This is because CLIP is a multimodal encoder, which is pre-trained by conducting contrastive learning between image-text pairs. Therefore, the text also shares part of the embedding space and is mutually exclusive with mismatched images. Although " + }, + { + "bbox": [ + 130, + 342, + 482, + 668 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_D" + }, + { + "bbox": [ + 130, + 342, + 482, + 668 + ], + "type": "text", + "content": " can push embeddings of different samples apart, a significant weight of it will make all embeddings more uniformly distributed, i.e., not align with CLIP. Therefore, we set a smaller weight to " + }, + { + "bbox": [ + 130, + 342, + 482, + 668 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_D" + }, + { + "bbox": [ + 130, + 342, + 482, + 668 + ], + "type": "text", + "content": " when stealing CLIP. This adjustment is practical since the attacker has access to the outputs of the target encoder, and thus can observe its embedding space to adjust its attack settings. Table 2 shows that using a surrogate dataset of only " + }, + { + "bbox": [ + 130, + 342, + 482, + 668 + ], + "type": "inline_equation", + "content": "3.07\\%" + }, + { + "bbox": [ + 130, + 342, + 482, + 668 + ], + "type": "text", + "content": " the size of ImageNet (40K / 1.3M) and " + }, + { + "bbox": [ + 130, + 342, + 482, + 668 + ], + "type": "inline_equation", + "content": "0.015\\%" + }, + { + "bbox": [ + 130, + 342, + 482, + 668 + ], + "type": "text", + "content": " the size of the training data of CLIP (60K / 400M), RDA can achieve comparable performances with the two encoders across various downstream tasks. The results also reveal that even under the architecture of the surrogate and target encoders being distinct (ResNet18 vs. ResNet50), RDA still exhibits high effec" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 290, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 290, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 290, + 91, + 447, + 102 + ], + "type": "text", + "content": "Refine, Discriminate and Align (RDA)" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 182, + 161, + 428, + 217 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 482, + 152 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 482, + 152 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 482, + 152 + ], + "type": "text", + "content": "Table 4: Comparisons between RDA and baselines to steal the encoder pre-trained on CIFAR10 under the same surrogate dataset size. We report the " + }, + { + "bbox": [ + 132, + 114, + 482, + 152 + ], + "type": "inline_equation", + "content": "\\frac{\\mathrm{SA}}{\\mathrm{TA}} \\times 100\\%" + }, + { + "bbox": [ + 132, + 114, + 482, + 152 + ], + "type": "text", + "content": " achieved by each method and its query cost. The Optimal and suboptimal results are highlighted." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 182, + 161, + 428, + 217 + ], + "lines": [ + { + "bbox": [ + 182, + 161, + 428, + 217 + ], + "spans": [ + { + "bbox": [ + 182, + 161, + 428, + 217 + ], + "type": "table", + "html": "
MethodCIFAR10CIFAR100MNISTGTSRBSVHNSTL10F-MNISTQueries
Conventional [34]63.3645.4096.1620.3468.2660.2594.522,500
StolenEncoder [26]79.0368.7898.7663.7092.0176.2995.912,500
Cont-Steal [34]80.4373.9498.0395.5596.3980.0895.39250,000
RDA83.8477.4998.9794.2796.7982.2594.692,500
85.8583.0899.33101.6099.9585.56100.8425,000
", + "image_path": "323d735577d2edde482abecfc42bc1868e5e7b533a4645a3369c88d5b4badbde.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 233, + 248, + 381, + 311 + ], + "blocks": [ + { + "bbox": [ + 133, + 227, + 479, + 239 + ], + "lines": [ + { + "bbox": [ + 133, + 227, + 479, + 239 + ], + "spans": [ + { + "bbox": [ + 133, + 227, + 479, + 239 + ], + "type": "text", + "content": "Table 5: Comparisons between RDA and StolenEncoder under the same query cost." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 233, + 248, + 381, + 311 + ], + "lines": [ + { + "bbox": [ + 233, + 248, + 381, + 311 + ], + "spans": [ + { + "bbox": [ + 233, + 248, + 381, + 311 + ], + "type": "table", + "html": "
QueriesMethodsettingSA
2500StolenEncoder2,500 images71.98
RDA500 images×5 patches65.38
625 images×4 patches67.07
1,250 images×2 patches71.44
2,500 images×1 patch76.36
25,000StolenEncoder25,000 images76.23
RDA2,500 images×10 patches78.50
", + "image_path": "e1d768f054773423bd290e1c661e41fbee6884c2403c4886b1c1727b883d7a78.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 334, + 480, + 370 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 334, + 480, + 370 + ], + "spans": [ + { + "bbox": [ + 130, + 334, + 480, + 370 + ], + "type": "text", + "content": "tiveness. Moreover, Table 3 shows that RDA requires much fewer computation resources to steal the two encoders than pre-training them from scratch, owing to the small surrogate dataset and lightweight network architecture." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 388, + 343, + 400 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 388, + 343, + 400 + ], + "spans": [ + { + "bbox": [ + 132, + 388, + 343, + 400 + ], + "type": "text", + "content": "4.3 Comparison with Existing Methods" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 407, + 482, + 492 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 407, + 482, + 492 + ], + "spans": [ + { + "bbox": [ + 130, + 407, + 482, + 492 + ], + "type": "text", + "content": "We consider the conventional method (CVPR 2023, proposed as the baseline by [34]), StolenEncoder (CCS 2022) [26], and Cont-Steal (CVPR 2023) [34] as our baselines. A more detailed explanation of them is in Supplementary A.1. Specifically, we compare RDA with them to steal the encoder pre-trained on CIFAR10 under three different scenarios. For a fair comparison, the surrogate dataset remains identical across all baselines (except for some results presented in Table 5 due to different sizes of surrogate datasets)." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "content": "Under the Same Surrogate Dataset Size. In this subsection, we fix the surrogate dataset size to 2,500 and make it identical across all methods. As Table 4 shows, RDA outperforms baselines by a significant margin across all seven downstream datasets with a moderate query cost. We also evaluate RDA under a setting that has the same query cost as the conventional method and StolenEncoder (i.e., 2,500 queries), for which we query with each sample once in its origin version. The results show that RDA still surpasses baselines over five out of seven datasets with the least query cost. In addition, the comparison between the results achieved by RDA with 25,000 and 2,500 queries reveals that more patches for generating prototypes will enhance the attack. The embedding space of the surrogate encoder trained by each method is visualized in Figure 11 of Supplementary A.4, which shows that our RDA can train surrogate encoders to discriminate different samples better. Besides, We have summarized the mean" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 216, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 216, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 216, + 100 + ], + "type": "text", + "content": "S. Wu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 134, + 115, + 232, + 219 + ], + "blocks": [ + { + "bbox": [ + 134, + 115, + 232, + 219 + ], + "lines": [ + { + "bbox": [ + 134, + 115, + 232, + 219 + ], + "spans": [ + { + "bbox": [ + 134, + 115, + 232, + 219 + ], + "type": "image", + "image_path": "68c1f19139c5858039488a50dbd8c8828ced00c55f88def557a15aa2502fb3e0.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 227, + 480, + 271 + ], + "lines": [ + { + "bbox": [ + 130, + 227, + 480, + 271 + ], + "spans": [ + { + "bbox": [ + 130, + 227, + 480, + 271 + ], + "type": "text", + "content": "Fig. 7: Heat map of the achieved SAs (better zoom in). This figure shows the performance of 100 combinations of different training epochs and (a) the surrogate dataset size, (b) the patch number for generating prototypes, and (c) the patch number for training." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 235, + 116, + 359, + 219 + ], + "blocks": [ + { + "bbox": [ + 235, + 116, + 359, + 219 + ], + "lines": [ + { + "bbox": [ + 235, + 116, + 359, + 219 + ], + "spans": [ + { + "bbox": [ + 235, + 116, + 359, + 219 + ], + "type": "image", + "image_path": "27b122294343f329ad95e5a5f35b0d1b87729066e691de1baaf4ad42c9a4bbde.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 361, + 116, + 473, + 218 + ], + "blocks": [ + { + "bbox": [ + 361, + 116, + 473, + 218 + ], + "lines": [ + { + "bbox": [ + 361, + 116, + 473, + 218 + ], + "spans": [ + { + "bbox": [ + 361, + 116, + 473, + 218 + ], + "type": "image", + "image_path": "33897e5bb8869cd28a205b8bb29ff49b891802d28e985eb34f4d41ff5f411307.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 295, + 480, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 295, + 480, + 319 + ], + "spans": [ + { + "bbox": [ + 130, + 295, + 480, + 319 + ], + "type": "text", + "content": "value achieved by each method over the seven downstream datasets in Figure 3 to show the superiority of RDA." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 336, + 482, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 336, + 482, + 529 + ], + "spans": [ + { + "bbox": [ + 130, + 336, + 482, + 529 + ], + "type": "text", + "content": "Under the Same Query Cost. To demonstrate the superiority of RDA against baselines under an identical query cost, we consider two fixed values, i.e., 2,500 and 25,000. We take StolenEncoder as the rival since it is the best baseline with the least query cost. We note that Cont-Steal is beyond our consideration because the data volume for it would be too small under such a setting, which is only 25 images for 2,500 queries and 250 images for 25,000 queries. " + }, + { + "bbox": [ + 130, + 336, + 482, + 529 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 336, + 482, + 529 + ], + "type": "text", + "content": " images " + }, + { + "bbox": [ + 130, + 336, + 482, + 529 + ], + "type": "inline_equation", + "content": "\\times n" + }, + { + "bbox": [ + 130, + 336, + 482, + 529 + ], + "type": "text", + "content": " patches in the table indicates that we use " + }, + { + "bbox": [ + 130, + 336, + 482, + 529 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 336, + 482, + 529 + ], + "type": "text", + "content": " images as the surrogate dataset and crop each image into " + }, + { + "bbox": [ + 130, + 336, + 482, + 529 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 130, + 336, + 482, + 529 + ], + "type": "text", + "content": " patches to generate prototypes. Therefore, the query cost of RDA is " + }, + { + "bbox": [ + 130, + 336, + 482, + 529 + ], + "type": "inline_equation", + "content": "N \\times n" + }, + { + "bbox": [ + 130, + 336, + 482, + 529 + ], + "type": "text", + "content": ". As shown in Table 5, under the same query cost of 2,500, RDA can achieve comparable results with StolenEncoder with half the size of its surrogate dataset. With the same 2,500 images, RDA outperforms StolenEncoder by a significant margin, showcasing its superiority. Moreover, under a query cost of 25,000, RDA outperforms StolenEncoder by " + }, + { + "bbox": [ + 130, + 336, + 482, + 529 + ], + "type": "inline_equation", + "content": "2.14\\%" + }, + { + "bbox": [ + 130, + 336, + 482, + 529 + ], + "type": "text", + "content": " with only " + }, + { + "bbox": [ + 130, + 336, + 482, + 529 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 130, + 336, + 482, + 529 + ], + "type": "text", + "content": " the size of its surrogate dataset. Besides, comparisons between different configurations of RDA under queries of 2,500 indicate that the training data volume has a dominant impact on the performance." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": "Under the Same Time Cost. Since each sample is also cropped and augmented into multiple patches for training (i.e., needing to forward encode each sample multiple times), RDA takes the longest time for training. Table 10 in Supplementary A.4 contains the detailed time cost of each method. To make a fair comparison, we prolong the training for another 200 epochs for baseline methods, i.e., 300 epochs in total, which is sufficient to lead their performances to saturate. The trained surrogate encoders are then evaluated on CIFAR10 as the downstream dataset. Figure 6 reveals that even though baseline methods take longer time to train, RDA still outperforms them with the least query cost, and the performance gap can be further widened by a moderate query cost." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 290, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 290, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 290, + 91, + 447, + 102 + ], + "type": "text", + "content": "Refine, Discriminate and Align (RDA)" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 163, + 117, + 257, + 200 + ], + "blocks": [ + { + "bbox": [ + 163, + 117, + 257, + 200 + ], + "lines": [ + { + "bbox": [ + 163, + 117, + 257, + 200 + ], + "spans": [ + { + "bbox": [ + 163, + 117, + 257, + 200 + ], + "type": "image", + "image_path": "25c6b349b74595b444bbfd8d7e4fe326e62b9c71dab073086cfae645c9b7a5b5.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 208, + 479, + 220 + ], + "lines": [ + { + "bbox": [ + 132, + 208, + 479, + 220 + ], + "spans": [ + { + "bbox": [ + 132, + 208, + 479, + 220 + ], + "type": "text", + "content": "Fig. 8: Ablation studies on the weight of each part in " + }, + { + "bbox": [ + 132, + 208, + 479, + 220 + ], + "type": "inline_equation", + "content": "\\mathcal{L}" + }, + { + "bbox": [ + 132, + 208, + 479, + 220 + ], + "type": "text", + "content": ", with " + }, + { + "bbox": [ + 132, + 208, + 479, + 220 + ], + "type": "inline_equation", + "content": "\\lambda_1 = 1" + }, + { + "bbox": [ + 132, + 208, + 479, + 220 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 132, + 208, + 479, + 220 + ], + "type": "inline_equation", + "content": "\\lambda_2" + }, + { + "bbox": [ + 132, + 208, + 479, + 220 + ], + "type": "text", + "content": " varying." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 263, + 118, + 353, + 199 + ], + "blocks": [ + { + "bbox": [ + 263, + 118, + 353, + 199 + ], + "lines": [ + { + "bbox": [ + 263, + 118, + 353, + 199 + ], + "spans": [ + { + "bbox": [ + 263, + 118, + 353, + 199 + ], + "type": "image", + "image_path": "ca8dd28c69d9034c5b3ab2d625676c610c1741385206b9bae714e3a184778157.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 360, + 118, + 450, + 199 + ], + "blocks": [ + { + "bbox": [ + 360, + 118, + 450, + 199 + ], + "lines": [ + { + "bbox": [ + 360, + 118, + 450, + 199 + ], + "spans": [ + { + "bbox": [ + 360, + 118, + 450, + 199 + ], + "type": "image", + "image_path": "38b90e2cc3ec07769caada336e93e464ec65ffaac00afc6fa3ee2f821392afeb.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 183, + 251, + 428, + 318 + ], + "blocks": [ + { + "bbox": [ + 152, + 230, + 459, + 241 + ], + "lines": [ + { + "bbox": [ + 152, + 230, + 459, + 241 + ], + "spans": [ + { + "bbox": [ + 152, + 230, + 459, + 241 + ], + "type": "text", + "content": "Table 6: Ablation studies on loss functions. The presented results are SAs." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 183, + 251, + 428, + 318 + ], + "lines": [ + { + "bbox": [ + 183, + 251, + 428, + 318 + ], + "spans": [ + { + "bbox": [ + 183, + 251, + 428, + 318 + ], + "type": "table", + "html": "
Loss TypeCIFAR10STL10CIFAR100MNISTF-MNISTGTSRBSVHNAVG
MSE65.6653.8331.0094.6386.1232.7962.9761.00
Cosine Similarity80.2366.5643.8197.0088.7655.5375.9472.54
KL Divergence77.3565.3040.9896.2387.8257.0469.3870.58
InfoNCE73.9961.6839.2495.7687.8658.8268.8869.43
LD75.5463.243.1696.5289.2765.7771.1572.23
LA79.7065.5943.7996.8989.2455.9676.7072.55
LA+LD79.3966.7644.2796.7489.3262.7573.7473.28
", + "image_path": "8848e4305a5d0169b444b11a18852de117433ed4eeb22b55ddd58ac5fb69607b.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 338, + 245, + 348 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 338, + 245, + 348 + ], + "spans": [ + { + "bbox": [ + 132, + 338, + 245, + 348 + ], + "type": "text", + "content": "4.4 Ablation Studies" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 356, + 482, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 356, + 482, + 464 + ], + "spans": [ + { + "bbox": [ + 130, + 356, + 482, + 464 + ], + "type": "text", + "content": "In this subsection, we conduct ablation studies to investigate the impact of the surrogate dataset size, the patch number for generating prototypes (i.e., " + }, + { + "bbox": [ + 130, + 356, + 482, + 464 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 130, + 356, + 482, + 464 + ], + "type": "text", + "content": " in Eq. 1), the patch number for training (i.e., " + }, + { + "bbox": [ + 130, + 356, + 482, + 464 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 130, + 356, + 482, + 464 + ], + "type": "text", + "content": " in Eq. 2, 3, 5, and 6), and loss functions. All the experiments are conducted to steal the encoder pre-trained on CIFAR10. For evaluations, the trained surrogate encoders are assessed on CIFAR10 in the first three ablation studies. While examining the influence of loss functions, we consider seven distinct downstream datasets, as shown in Table 4, to derive a more comprehensive conclusion. Other configurations remain at their default settings, and we present the resulting SAs." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 479, + 482, + 539 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 479, + 482, + 539 + ], + "spans": [ + { + "bbox": [ + 130, + 479, + 482, + 539 + ], + "type": "text", + "content": "Surrogate Dataset Size. Recall our default setting where we randomly sample 2,500 images from Tiny ImageNet to construct the surrogate dataset, which is " + }, + { + "bbox": [ + 130, + 479, + 482, + 539 + ], + "type": "inline_equation", + "content": "5\\%" + }, + { + "bbox": [ + 130, + 479, + 482, + 539 + ], + "type": "text", + "content": " the size of CIFAR-10. We vary the ratio from " + }, + { + "bbox": [ + 130, + 479, + 482, + 539 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 130, + 479, + 482, + 539 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 130, + 479, + 482, + 539 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 130, + 479, + 482, + 539 + ], + "type": "text", + "content": " to investigate its impact. Figure 7 (a) shows that a larger surrogate dataset will accelerate the training and improve the attack performance." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 554, + 482, + 625 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 554, + 482, + 625 + ], + "spans": [ + { + "bbox": [ + 130, + 554, + 482, + 625 + ], + "type": "text", + "content": "Patch Number for Generating Prototypes. We vary the patch number for generating prototypes from 1 to 100 to investigate its impact. Figure 7 (b) shows that there is a tendency for more patches for generating prototypes will improve the attack performance and accelerate the convergence. However, when the patch number reaches 10, more patches do not further improve the attack performance due to the limited surrogate dataset size." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "content": "Patch Number for Training. We vary the patch number for training from 1 to 10 to investigate its impact. Figure 7 (c) reveals a tendency for more patches" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 216, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 216, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 216, + 100 + ], + "type": "text", + "content": "S. Wu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": "for training will accelerate the training and achieve better results. However, more patches also indicate longer training time and more computation resources." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 157, + 482, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 157, + 482, + 384 + ], + "spans": [ + { + "bbox": [ + 130, + 157, + 482, + 384 + ], + "type": "text", + "content": "Loss Functions. To demonstrate the superiority of the design of our loss function and the necessity of each part, we conduct ablation studies as shown in Table 6. Table 6 shows that " + }, + { + "bbox": [ + 130, + 157, + 482, + 384 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_D" + }, + { + "bbox": [ + 130, + 157, + 482, + 384 + ], + "type": "text", + "content": " largely surpass other loss functions on GTSRB while " + }, + { + "bbox": [ + 130, + 157, + 482, + 384 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_A" + }, + { + "bbox": [ + 130, + 157, + 482, + 384 + ], + "type": "text", + "content": " largely surpass other loss functions on SVHN. We hypothesize that " + }, + { + "bbox": [ + 130, + 157, + 482, + 384 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_A" + }, + { + "bbox": [ + 130, + 157, + 482, + 384 + ], + "type": "text", + "content": " attains optimal outcomes on SVHN due to the pronounced resemblance between SVHN and the surrogate dataset. It appears that " + }, + { + "bbox": [ + 130, + 157, + 482, + 384 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_A" + }, + { + "bbox": [ + 130, + 157, + 482, + 384 + ], + "type": "text", + "content": " inclines towards inducing the surrogate encoder to overfit the surrogate data, thereby elucidating its suboptimal performance on GTSRB, a dataset characterized by less similarity with the surrogate dataset. On the contrary, " + }, + { + "bbox": [ + 130, + 157, + 482, + 384 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_D" + }, + { + "bbox": [ + 130, + 157, + 482, + 384 + ], + "type": "text", + "content": " makes the surrogate encoder fit less to the surrogate dataset and has the best result on GTSRB. To further investigate the effect of each part in our loss, we have conducted an ablation experiment about the weight of each part in our loss. As Figure 8 shows, the surrogate encoder will perform better on SVHN while worse on GTSRB when the weight of " + }, + { + "bbox": [ + 130, + 157, + 482, + 384 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_A" + }, + { + "bbox": [ + 130, + 157, + 482, + 384 + ], + "type": "text", + "content": " increases. We refer to Supplementary A.4 for a more detailed analysis. Combining " + }, + { + "bbox": [ + 130, + 157, + 482, + 384 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_D" + }, + { + "bbox": [ + 130, + 157, + 482, + 384 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 157, + 482, + 384 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_A" + }, + { + "bbox": [ + 130, + 157, + 482, + 384 + ], + "type": "text", + "content": ", our loss is the most robust one, which achieves the best average result over the seven tested datasets. For example, although cosine similarity and " + }, + { + "bbox": [ + 130, + 157, + 482, + 384 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_A" + }, + { + "bbox": [ + 130, + 157, + 482, + 384 + ], + "type": "text", + "content": " performs well on SVHN (about " + }, + { + "bbox": [ + 130, + 157, + 482, + 384 + ], + "type": "inline_equation", + "content": "76\\%" + }, + { + "bbox": [ + 130, + 157, + 482, + 384 + ], + "type": "text", + "content": "), their performance on GTSRB is terrible (about " + }, + { + "bbox": [ + 130, + 157, + 482, + 384 + ], + "type": "inline_equation", + "content": "56\\%" + }, + { + "bbox": [ + 130, + 157, + 482, + 384 + ], + "type": "text", + "content": "). Our loss design can improve their performance on GTSRB by about " + }, + { + "bbox": [ + 130, + 157, + 482, + 384 + ], + "type": "inline_equation", + "content": "7\\%" + }, + { + "bbox": [ + 130, + 157, + 482, + 384 + ], + "type": "text", + "content": " in the cost of less than " + }, + { + "bbox": [ + 130, + 157, + 482, + 384 + ], + "type": "inline_equation", + "content": "3\\%" + }, + { + "bbox": [ + 130, + 157, + 482, + 384 + ], + "type": "text", + "content": " decrease on SVHN." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 401, + 279, + 412 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 401, + 279, + 412 + ], + "spans": [ + { + "bbox": [ + 131, + 401, + 279, + 412 + ], + "type": "text", + "content": "4.5 Robustness to Defenses" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 421, + 479, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 421, + 479, + 468 + ], + "spans": [ + { + "bbox": [ + 130, + 421, + 479, + 468 + ], + "type": "text", + "content": "In this section, we evaluate the robustness of RDA against three perturbation-based defenses that aim to limit information leakage and one watermarking defense that aims to detect whether a suspected encoder is stolen. The pretraining and downstream datasets both are set to CIFAR10." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 486, + 479, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 486, + 479, + 510 + ], + "spans": [ + { + "bbox": [ + 130, + 486, + 479, + 510 + ], + "type": "text", + "content": "Perturbation-Based Defense. We evaluate three common practices of this type of defense, i.e., adding noise [30], top-" + }, + { + "bbox": [ + 130, + 486, + 479, + 510 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 486, + 479, + 510 + ], + "type": "text", + "content": " [30] and rounding [44]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 510, + 480, + 603 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 130, + 510, + 480, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 480, + 544 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 480, + 544 + ], + "type": "text", + "content": "- Adding noise: Adding noise means that the defender will introduce noise to the original outputs of the target encoder. Following [10], we set the mean of the noise to 0 and vary the standard deviations to control the noise level." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 546, + 480, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 546, + 480, + 579 + ], + "spans": [ + { + "bbox": [ + 130, + 546, + 480, + 579 + ], + "type": "text", + "content": "- Top-" + }, + { + "bbox": [ + 130, + 546, + 480, + 579 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 546, + 480, + 579 + ], + "type": "text", + "content": ": Top-" + }, + { + "bbox": [ + 130, + 546, + 480, + 579 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 546, + 480, + 579 + ], + "type": "text", + "content": " means that the defender will only output the first " + }, + { + "bbox": [ + 130, + 546, + 480, + 579 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 546, + 480, + 579 + ], + "type": "text", + "content": " largest number of each embedding from the target encoder and set the rest as 0. We vary the value of " + }, + { + "bbox": [ + 130, + 546, + 480, + 579 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 546, + 480, + 579 + ], + "type": "text", + "content": " to simulate different perturbation levels." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 582, + 479, + 603 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 479, + 603 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 479, + 603 + ], + "type": "text", + "content": "- Rounding: Rounding truncates each value in the embedding to a specific precision, which we vary to simulate different perturbation levels." + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 130, + 606, + 479, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 606, + 479, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 606, + 479, + 665 + ], + "type": "text", + "content": "The results of the three perturbation-based defense methods are summarized in (a), (b), and (c) of Figure 9, respectively. We can observe in Figure 9 (a) that while adding noise can mitigate the model stealing attack, it decreases the utility of the target encoder more significantly. Moreover, Figure 9 (b) shows that though top-" + }, + { + "bbox": [ + 130, + 606, + 479, + 665 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 606, + 479, + 665 + ], + "type": "text", + "content": " can decrease the efficacy of RDA as the defender lowers the" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 290, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 290, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 290, + 91, + 447, + 102 + ], + "type": "text", + "content": "Refine, Discriminate and Align (RDA)" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 135, + 115, + 219, + 177 + ], + "blocks": [ + { + "bbox": [ + 135, + 115, + 219, + 177 + ], + "lines": [ + { + "bbox": [ + 135, + 115, + 219, + 177 + ], + "spans": [ + { + "bbox": [ + 135, + 115, + 219, + 177 + ], + "type": "image", + "image_path": "4835adab1adabc6cc1d857acb85922308e7178cc04cf586b498ef38f5396a1a6.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 166, + 180, + 198, + 188 + ], + "lines": [ + { + "bbox": [ + 166, + 180, + 198, + 188 + ], + "spans": [ + { + "bbox": [ + 166, + 180, + 198, + 188 + ], + "type": "text", + "content": "(a) Adding noise" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 220, + 114, + 305, + 177 + ], + "blocks": [ + { + "bbox": [ + 220, + 114, + 305, + 177 + ], + "lines": [ + { + "bbox": [ + 220, + 114, + 305, + 177 + ], + "spans": [ + { + "bbox": [ + 220, + 114, + 305, + 177 + ], + "type": "image", + "image_path": "358420117ec56ecb2e34b4f230719cb3d958959c7017bdf62cadfec3efff4fbc.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 260, + 180, + 278, + 187 + ], + "lines": [ + { + "bbox": [ + 260, + 180, + 278, + 187 + ], + "spans": [ + { + "bbox": [ + 260, + 180, + 278, + 187 + ], + "type": "text", + "content": "(b) Top-k" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 307, + 114, + 391, + 177 + ], + "blocks": [ + { + "bbox": [ + 307, + 114, + 391, + 177 + ], + "lines": [ + { + "bbox": [ + 307, + 114, + 391, + 177 + ], + "spans": [ + { + "bbox": [ + 307, + 114, + 391, + 177 + ], + "type": "image", + "image_path": "006437a89d4af20256750db59303260d6c8017deeda1896b6b9d2d638a9e45f5.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 342, + 180, + 368, + 188 + ], + "lines": [ + { + "bbox": [ + 342, + 180, + 368, + 188 + ], + "spans": [ + { + "bbox": [ + 342, + 180, + 368, + 188 + ], + "type": "text", + "content": "(c) Rounding" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 168, + 198, + 444, + 209 + ], + "lines": [ + { + "bbox": [ + 168, + 198, + 444, + 209 + ], + "spans": [ + { + "bbox": [ + 168, + 198, + 444, + 209 + ], + "type": "text", + "content": "Fig. 9: The performance of different defense methods on CIFAR10." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 394, + 118, + 478, + 177 + ], + "blocks": [ + { + "bbox": [ + 394, + 118, + 478, + 177 + ], + "lines": [ + { + "bbox": [ + 394, + 118, + 478, + 177 + ], + "spans": [ + { + "bbox": [ + 394, + 118, + 478, + 177 + ], + "type": "image", + "image_path": "6659c498ab0c0bba8ea463f127386d1f0f61792d668e6c9d4e21de0e37050eab.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 424, + 180, + 459, + 188 + ], + "lines": [ + { + "bbox": [ + 424, + 180, + 459, + 188 + ], + "spans": [ + { + "bbox": [ + 424, + 180, + 459, + 188 + ], + "type": "text", + "content": "(d) Watermarking" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 233, + 480, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 233, + 480, + 281 + ], + "spans": [ + { + "bbox": [ + 130, + 233, + 480, + 281 + ], + "type": "text", + "content": "value of " + }, + { + "bbox": [ + 130, + 233, + 480, + 281 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 233, + 480, + 281 + ], + "type": "text", + "content": ", it also severely deteriorates the target encoder's performance. On the other hand, Figure 9 (c) reveals that rounding has a minimal effect on the performance of both the target encoder and attack. Our experimental results demonstrate the robustness of RDA against perturbation-based defenses." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 299, + 482, + 431 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 299, + 482, + 431 + ], + "spans": [ + { + "bbox": [ + 130, + 299, + 482, + 431 + ], + "type": "text", + "content": "Watermarking Defense. As shown by Adi et al. [1], backdoors can be used as watermarks to claim the ownership of a model. In this sense, we follow [34] and leverage BadEncoder [18] to embed a backdoor into the target encoder as the watermark. Ideally, if the watermarked encoder is stolen, the surrogate encoder trained via stealing should also be triggered by the defender-specified trigger to exhibit certain behaviors, and thus the ownership can be claimed. Fortunately, Figure 9 (d) shows that RDA can steal the functionality of the target encoder while unlearning the watermark. More specifically, the watermark rate (WR) of the target encoder is " + }, + { + "bbox": [ + 130, + 299, + 482, + 431 + ], + "type": "inline_equation", + "content": "99.87\\%" + }, + { + "bbox": [ + 130, + 299, + 482, + 431 + ], + "type": "text", + "content": " but only " + }, + { + "bbox": [ + 130, + 299, + 482, + 431 + ], + "type": "inline_equation", + "content": "17.04\\%" + }, + { + "bbox": [ + 130, + 299, + 482, + 431 + ], + "type": "text", + "content": " for the surrogate encoder when the training converges. This observation indicates that the watermark embedded in the target encoder cannot be preserved by RDA." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 449, + 220, + 462 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 449, + 220, + 462 + ], + "spans": [ + { + "bbox": [ + 132, + 449, + 220, + 462 + ], + "type": "text", + "content": "5 Conclusion" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 474, + 481, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 474, + 481, + 545 + ], + "spans": [ + { + "bbox": [ + 130, + 474, + 481, + 545 + ], + "type": "text", + "content": "In this paper, we have proposed a novel model stealing method against SSL named RDA, which stands for refine, discriminate, and align. Compared with previous methods [26,34], RDA establishes a less biased optimization objective for each training sample and strives to extract more abundant functionality of the target encoder. This empowers RDA to exhibit tremendous effectiveness, efficiency, and robustness across diverse datasets and settings." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 545, + 482, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 482, + 617 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 482, + 617 + ], + "type": "text", + "content": "Ethical Concerns and Possible Defenses. We underscore that the misuse of model stealing techniques can jeopardize the privacy and economic rights of embedding service providers. We hope that our work will inspire the development of more sophisticated defense mechanisms to thwart model stealing attacks. A possible defense involves rejection schemes for suspicious queries that appear highly similar, which we have more detailed discussions about in the Supplementary." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "Limitation and Future Work. The query cost of RDA is not yet optimal and can be further reduced by techniques like clustering to establish an identical prototype for multiple similar images. Moreover, more unique (e.g., more efficient) designs of RDA for transformer-based huge models remain largely unexplored." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 216, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 216, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 216, + 100 + ], + "type": "text", + "content": "S. Wu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "text", + "content": "Acknowledgements. This work is supported by the National Key R&D Program of China under Grant 2022YFB3103500, the National Natural Science Foundation of China under Grant U20A20176 and 62072062, the National Natural Science Foundation of Chongqing under Grant cstc2022ycjh-bgzxm0031, and the Chongqing Research Program of Basic Research and Frontier Technology (Chongqing Talent) under Grant cstc2024ycjh-bgzxm0048." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 205, + 199, + 218 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 205, + 199, + 218 + ], + "spans": [ + { + "bbox": [ + 132, + 205, + 199, + 218 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 134, + 228, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 138, + 228, + 480, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 228, + 480, + 262 + ], + "spans": [ + { + "bbox": [ + 138, + 228, + 480, + 262 + ], + "type": "text", + "content": "1. Adi, Y., Baum, C., Cisse, M., Pinkas, B., Keshet, J.: Turning your weakness into a strength: Watermarking deep neural networks by backdoorsing. In: 27th USENIX Security Symposium (USENIX Security 18). pp. 1615-1631 (2018)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 263, + 480, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 263, + 480, + 306 + ], + "spans": [ + { + "bbox": [ + 138, + 263, + 480, + 306 + ], + "type": "text", + "content": "2. Bossard, L., Guillaumeun, M., Van Gool, L.: Food-101-mining discriminative components with random forests. In: Computer vision-ECCV 2014: 13th European conference, zurich, Switzerland, September 6-12, 2014, proceedings, part VI 13. pp. 446-461. Springer (2014)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 307, + 480, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 307, + 480, + 327 + ], + "spans": [ + { + "bbox": [ + 138, + 307, + 480, + 327 + ], + "type": "text", + "content": "3. Carlini, N., Wagner, D.: Towards evaluating the robustness of neural networks. In: 2017 IEEE symposium on security and privacy (sp). pp. 39-57. IEEE (2017)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 328, + 481, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 328, + 481, + 360 + ], + "spans": [ + { + "bbox": [ + 138, + 328, + 481, + 360 + ], + "type": "text", + "content": "4. Caron, M., Touvron, H., Misra, I., Jégou, H., Mairal, J., Bojanowski, P., Joulin, A.: Emerging properties in self-supervised vision transformers. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 9650-9660 (2021)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 361, + 480, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 361, + 480, + 392 + ], + "spans": [ + { + "bbox": [ + 138, + 361, + 480, + 392 + ], + "type": "text", + "content": "5. Chen, T., Kornblith, S., Norouzi, M., Hinton, G.: A simple framework for contrastive learning of visual representations. In: International conference on machine learning. pp. 1597-1607. PMLR (2020)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 393, + 480, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 393, + 480, + 415 + ], + "spans": [ + { + "bbox": [ + 138, + 393, + 480, + 415 + ], + "type": "text", + "content": "6. Chen, X., Fan, H., Girshick, R., He, K.: Improved baselines with momentum contrastive learning. arXiv preprint arXiv:2003.04297 (2020)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 415, + 480, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 415, + 480, + 458 + ], + "spans": [ + { + "bbox": [ + 138, + 415, + 480, + 458 + ], + "type": "text", + "content": "7. Coates, A., Ng, A., Lee, H.: An analysis of single-layer networks in unsupervised feature learning. In: Proceedings of the fourteenth international conference on artificial intelligence and statistics. pp. 215-223. JMLR Workshop and Conference Proceedings (2011)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 459, + 480, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 459, + 480, + 491 + ], + "spans": [ + { + "bbox": [ + 138, + 459, + 480, + 491 + ], + "type": "text", + "content": "8. Dai, Y., Chen, Z., Li, J., Heinecke, S., Sun, L., Xu, R.: Tackling data heterogeneity in federated learning with class prototypes. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 37, pp. 7314-7322 (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 491, + 480, + 523 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 491, + 480, + 523 + ], + "spans": [ + { + "bbox": [ + 138, + 491, + 480, + 523 + ], + "type": "text", + "content": "9. Devlin, J., Chang, M.W., Lee, K., Toutanova, K.: Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805 (2018)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 134, + 524, + 480, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 524, + 480, + 555 + ], + "spans": [ + { + "bbox": [ + 134, + 524, + 480, + 555 + ], + "type": "text", + "content": "10. Dziedzic, A., Dhawan, N., Kaleem, M.A., Guan, J., Papernot, N.: On the difficulty of defending self-supervised learning against model extraction. In: International Conference on Machine Learning. pp. 5757-5776. PMLR (2022)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 134, + 556, + 480, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 556, + 480, + 600 + ], + "spans": [ + { + "bbox": [ + 134, + 556, + 480, + 600 + ], + "type": "text", + "content": "11. Grill, J.B., Strub, F., Altché, F., Tallec, C., Richemond, P., Buchatskaya, E., Doersch, C., Avila Pires, B., Guo, Z., Gheshlaghi Azar, M., et al.: Bootstrap your own latent-a new approach to self-supervised learning. Advances in neural information processing systems 33, 21271-21284 (2020)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 134, + 601, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 601, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 134, + 601, + 480, + 632 + ], + "type": "text", + "content": "12. He, K., Fan, H., Wu, Y., Xie, S., Girshick, R.: Momentum contrast for unsupervised visual representation learning. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 9729-9738 (2020)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 134, + 633, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 633, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 134, + 633, + 480, + 665 + ], + "type": "text", + "content": "13. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 770-778 (2016)" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 290, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 290, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 290, + 91, + 447, + 102 + ], + "type": "text", + "content": "Refine, Discriminate and Align (RDA)" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "type": "text", + "content": "14. Howard, A.G., Zhu, M., Chen, B., Kalenichenko, D., Wang, W., Weyand, T., Andreetto, M., Adam, H.: Mobilenets: Efficient convolutional neural networks for mobile vision applications. arXiv preprint arXiv:1704.04861 (2017)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 149, + 481, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 149, + 481, + 182 + ], + "spans": [ + { + "bbox": [ + 133, + 149, + 481, + 182 + ], + "type": "text", + "content": "15. Huang, G., Liu, Z., Van Der Maaten, L., Weinberger, K.Q.: Densely connected convolutional networks. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 4700-4708 (2017)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 182, + 481, + 214 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 182, + 481, + 214 + ], + "spans": [ + { + "bbox": [ + 133, + 182, + 481, + 214 + ], + "type": "text", + "content": "16. Huang, W., Ye, M., Shi, Z., Li, H., Du, B.: Rethinking federated learning with domain shift: A prototype view. In: 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 16312-16322. IEEE (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 214, + 481, + 247 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 214, + 481, + 247 + ], + "spans": [ + { + "bbox": [ + 133, + 214, + 481, + 247 + ], + "type": "text", + "content": "17. Jia, H., Choquette-Choo, C.A., Chandrasekaran, V., Papernot, N.: Entangled watermarks as a defense against model extraction. In: 30th USENIX Security Symposium (USENIX Security 21). pp. 1937-1954 (2021)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 133, + 247, + 481, + 279 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 247, + 481, + 279 + ], + "spans": [ + { + "bbox": [ + 133, + 247, + 481, + 279 + ], + "type": "text", + "content": "18. Jia, J., Liu, Y., Gong, N.Z.: Badencoder: Backdoor attacks to pre-trained encoders in self-supervised learning. In: 2022 IEEE Symposium on Security and Privacy (SP). pp. 2043-2059. IEEE (2022)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 133, + 279, + 481, + 311 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 279, + 481, + 311 + ], + "spans": [ + { + "bbox": [ + 133, + 279, + 481, + 311 + ], + "type": "text", + "content": "19. Khosla, P., Teterwak, P., Wang, C., Sarna, A., Tian, Y., Isola, P., Maschinot, A., Liu, C., Krishnan, D.: Supervised contrastive learning. Advances in neural information processing systems 33, 18661-18673 (2020)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 311, + 481, + 333 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 311, + 481, + 333 + ], + "spans": [ + { + "bbox": [ + 132, + 311, + 481, + 333 + ], + "type": "text", + "content": "20. Kingma, D.P., Ba, J.: Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980 (2014)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 333, + 481, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 333, + 481, + 354 + ], + "spans": [ + { + "bbox": [ + 132, + 333, + 481, + 354 + ], + "type": "text", + "content": "21. Krizhevsky, A., Hinton, G., et al.: Learning multiple layers of features from tiny images (2009)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 354, + 481, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 354, + 481, + 374 + ], + "spans": [ + { + "bbox": [ + 132, + 354, + 481, + 374 + ], + "type": "text", + "content": "22. Krizhevsky, A., Nair, V., Hinton, G.: Cifar-10 (canadian institute for advanced research). URL http://www.cs.toronto.edu/kriz/cifar.html 5(4), 1 (2010)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 374, + 481, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 374, + 481, + 396 + ], + "spans": [ + { + "bbox": [ + 132, + 374, + 481, + 396 + ], + "type": "text", + "content": "23. Le, Y., Yang, X.: Tiny imagenet visual recognition challenge. CS 231N 7(7), 3 (2015)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 396, + 481, + 417 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 396, + 481, + 417 + ], + "spans": [ + { + "bbox": [ + 132, + 396, + 481, + 417 + ], + "type": "text", + "content": "24. LeCun, Y., Cortes, C.: MNIST handwritten digit database (2010), http://yann.lecun.com/exdb/mnist/" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 417, + 481, + 460 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 417, + 481, + 460 + ], + "spans": [ + { + "bbox": [ + 132, + 417, + 481, + 460 + ], + "type": "text", + "content": "25. Liu, H., Jia, J., Qu, W., Gong, N.Z.: Encodermi: Membership inference against pre-trained encoders in contrastive learning. In: Proceedings of the 2021 ACM SIGSAC Conference on Computer and Communications Security. pp. 2081-2095 (2021)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 460, + 481, + 493 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 460, + 481, + 493 + ], + "spans": [ + { + "bbox": [ + 132, + 460, + 481, + 493 + ], + "type": "text", + "content": "26. Liu, Y., Jia, J., Liu, H., Gong, N.Z.: Stolenencoder: stealing pre-trained encoders in self-supervised learning. In: Proceedings of the 2022 ACM SIGSAC Conference on Computer and Communications Security. pp. 2115-2128 (2022)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 493, + 481, + 515 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 493, + 481, + 515 + ], + "spans": [ + { + "bbox": [ + 132, + 493, + 481, + 515 + ], + "type": "text", + "content": "27. Van der Maaten, L., Hinton, G.: Visualizing data using t-sne. Journal of machine learning research 9(11) (2008)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 515, + 481, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 515, + 481, + 536 + ], + "spans": [ + { + "bbox": [ + 132, + 515, + 481, + 536 + ], + "type": "text", + "content": "28. Mettes, P., Van der Pol, E., Snoek, C.: Hyperspherical prototype networks. Advances in neural information processing systems 32 (2019)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 536, + 481, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 536, + 481, + 557 + ], + "spans": [ + { + "bbox": [ + 132, + 536, + 481, + 557 + ], + "type": "text", + "content": "29. Netzer, Y., Wang, T., Coates, A., Bissacco, A., Wu, B., Ng, A.Y.: Reading digits in natural images with unsupervised feature learning (2011)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 557, + 481, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 557, + 481, + 589 + ], + "spans": [ + { + "bbox": [ + 132, + 557, + 481, + 589 + ], + "type": "text", + "content": "30. Orekondy, T., Schiele, B., Fritz, M.: Knockoff nets: Stealing functionality of black-box models. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 4954-4963 (2019)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 589, + 481, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 589, + 481, + 622 + ], + "spans": [ + { + "bbox": [ + 132, + 589, + 481, + 622 + ], + "type": "text", + "content": "31. Papernot, N., McDaniel, P., Goodfellow, I., Jha, S., Celik, Z.B., Swami, A.: Practical black-box attacks against machine learning. In: Proceedings of the 2017 ACM on Asia conference on computer and communications security. pp. 506-519 (2017)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 132, + 622, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 622, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 622, + 481, + 665 + ], + "type": "text", + "content": "32. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021)" + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 215, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 215, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 215, + 100 + ], + "type": "text", + "content": "S. Wu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 534 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "text", + "content": "33. Salem, A., Zhang, Y., Humbert, M., Berrang, P., Fritz, M., Backes, M.: Ml-leaks: Model and data independent membership inference attacks and defenses on machine learning models. arXiv preprint arXiv:1806.01246 (2018)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 150, + 482, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 150, + 482, + 183 + ], + "spans": [ + { + "bbox": [ + 130, + 150, + 482, + 183 + ], + "type": "text", + "content": "34. Sha, Z., He, X., Yu, N., Backes, M., Zhang, Y.: Can't steal? cont-steal! contrastive stealing attacks against image encoders. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16373-16383 (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 183, + 482, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 183, + 482, + 205 + ], + "spans": [ + { + "bbox": [ + 130, + 183, + 482, + 205 + ], + "type": "text", + "content": "35. Sharir, O., Peleg, B., Shoham, Y.: The cost of training nlp models: A concise overview. arXiv preprint arXiv:2004.08900 (2020)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 205, + 482, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 205, + 482, + 237 + ], + "spans": [ + { + "bbox": [ + 130, + 205, + 482, + 237 + ], + "type": "text", + "content": "36. Shokri, R., Stronati, M., Song, C., Shmatikov, V.: Membership inference attacks against machine learning models. In: 2017 IEEE symposium on security and privacy (SP). pp. 3-18. IEEE (2017)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 238, + 482, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 238, + 482, + 259 + ], + "spans": [ + { + "bbox": [ + 130, + 238, + 482, + 259 + ], + "type": "text", + "content": "37. Simonyan, K., Zisserman, A.: Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556 (2014)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 259, + 482, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 259, + 482, + 281 + ], + "spans": [ + { + "bbox": [ + 130, + 259, + 482, + 281 + ], + "type": "text", + "content": "38. Snell, J., Swersky, K., Zemel, R.: Prototypical networks for few-shot learning. Advances in neural information processing systems 30 (2017)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 282, + 482, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 282, + 482, + 314 + ], + "spans": [ + { + "bbox": [ + 130, + 282, + 482, + 314 + ], + "type": "text", + "content": "39. Stallkamp, J., Schlipsing, M., Salmen, J., Igel, C.: Man vs. computer: Benchmarking machine learning algorithms for traffic sign recognition. Neural networks 32, 323-332 (2012)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 315, + 482, + 347 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 315, + 482, + 347 + ], + "spans": [ + { + "bbox": [ + 130, + 315, + 482, + 347 + ], + "type": "text", + "content": "40. Tan, Y., Long, G., Liu, L., Zhou, T., Lu, Q., Jiang, J., Zhang, C.: Fedproto: Federated prototype learning across heterogeneous clients. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 36, pp. 8432-8440 (2022)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 348, + 482, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 348, + 482, + 380 + ], + "spans": [ + { + "bbox": [ + 130, + 348, + 482, + 380 + ], + "type": "text", + "content": "41. Tan, Y., Long, G., Ma, J., Liu, L., Zhou, T., Jiang, J.: Federated learning from pretrained models: A contrastive learning approach. Advances in Neural Information Processing Systems 35, 19332-19344 (2022)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 380, + 482, + 424 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 380, + 482, + 424 + ], + "spans": [ + { + "bbox": [ + 130, + 380, + 482, + 424 + ], + "type": "text", + "content": "42. Tian, Y., Wang, Y., Krishnan, D., Tenenbaum, J.B., Isola, P.: Rethinking few-shot image classification: a good embedding is all you need? In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XIV 16. pp. 266-282. Springer (2020)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 424, + 482, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 424, + 482, + 446 + ], + "spans": [ + { + "bbox": [ + 130, + 424, + 482, + 446 + ], + "type": "text", + "content": "43. Tong, S., Chen, Y., Ma, Y., Lecun, Y.: Emp-ssl: Towards self-supervised learning in one training epoch. arXiv preprint arXiv:2304.03977 (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 447, + 482, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 447, + 482, + 479 + ], + "spans": [ + { + "bbox": [ + 130, + 447, + 482, + 479 + ], + "type": "text", + "content": "44. Tramér, F., Zhang, F., Juels, A., Reiter, M.K., Ristenpart, T.: Stealing machine learning models via prediction {APIs}. In: 25th USENIX security symposium (USENIX Security 16). pp. 601-618 (2016)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 479, + 482, + 512 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 479, + 482, + 512 + ], + "spans": [ + { + "bbox": [ + 130, + 479, + 482, + 512 + ], + "type": "text", + "content": "45. Wu, Z., Xiong, Y., Yu, S.X., Lin, D.: Unsupervised feature learning via nonparametric instance discrimination. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 3733-3742 (2018)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 512, + 482, + 534 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 512, + 482, + 534 + ], + "spans": [ + { + "bbox": [ + 130, + 512, + 482, + 534 + ], + "type": "text", + "content": "46. Xiao, H., Rasul, K., Vollgraf, R.: Fashion-mnist: a novel image dataset for benchmarking machine learning algorithms. arXiv preprint arXiv:1708.07747 (2017)" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 290, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 290, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 290, + 91, + 447, + 102 + ], + "type": "text", + "content": "Refine, Discriminate and Align (RDA)" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Reflective Instruction Tuning_ Mitigating Hallucinations in Large Vision-Language Models/8b065bde-2d53-44c1-a90a-d68b84da76bf_content_list.json b/2024/Reflective Instruction Tuning_ Mitigating Hallucinations in Large Vision-Language Models/8b065bde-2d53-44c1-a90a-d68b84da76bf_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..2c668bbc9f6e8a2f569e207d1294b808f06e9826 --- /dev/null +++ b/2024/Reflective Instruction Tuning_ Mitigating Hallucinations in Large Vision-Language Models/8b065bde-2d53-44c1-a90a-d68b84da76bf_content_list.json @@ -0,0 +1,1407 @@ +[ + { + "type": "text", + "text": "Reflective Instruction Tuning: Mitigating Hallucinations in Large Vision-Language Models", + "text_level": 1, + "bbox": [ + 223, + 141, + 779, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jinrui Zhang $^{1}$ , Teng Wang $^{1,2}$ , Haigang Zhang $^{3}$ , Ping Lu $^{4}$ , Feng Zheng $^{1,5}$", + "bbox": [ + 215, + 210, + 785, + 228 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Southern University of Science and Technology", + "bbox": [ + 336, + 238, + 666, + 253 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "2 The University of Hong Kong", + "bbox": [ + 393, + 253, + 609, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3 Shenzhen Polytechnic University", + "bbox": [ + 383, + 268, + 617, + 281 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "4 The Cloud Computing and IT Institute of ZTE Corporation", + "bbox": [ + 294, + 281, + 709, + 295 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{5}$ Research Institute of Multiple Agents and Embodied Intelligence, Peng Cheng Laboratory, Shenzhen, China", + "bbox": [ + 274, + 295, + 725, + 323 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Large vision-language models (LVLMs) have shown promising performance on a variety of vision-language tasks. However, they remain susceptible to hallucinations, generating outputs misaligned with visual content or instructions. While various mitigation strategies have been proposed, they often neglect a key contributor to hallucinations: lack of fine-grained reasoning supervision during training. Without intermediate reasoning steps, models may establish superficial shortcuts between instructions and responses, failing to internalize the inherent reasoning logic. To address this challenge, we propose reflective instruction tuning, which integrates rationale learning into visual instruction tuning. Unlike previous methods that learning from responses only, our approach entails the model predicting rationales justifying why responses are correct or incorrect. This fosters a deeper engagement with the fine-grained reasoning underlying each response, thus enhancing the model's reasoning proficiency. To facilitate this approach, we propose REVERIE, the first large-scale instruction-tuning dataset with ReflEctiVE Rationale annotations. REVERIE comprises 115k machine-generated reasoning instructions, each meticulously annotated with a corresponding pair of correct and confusing responses, alongside comprehensive rationales elucidating the justification behind the correctness or erroneousness of each response. Experimental results on multiple LVLM benchmarks reveal that reflective instruction tuning with the REVERIE dataset yields noticeable performance gain over the baseline model, demonstrating the effectiveness of reflecting from the rationales. Project page is at https://zjr2000.github.io/projects/reverie", + "bbox": [ + 261, + 351, + 743, + 698 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Large Vision Language Models $\\cdot$ Visual Instruction Tuning $\\cdot$ Hallucination Mitigation", + "bbox": [ + 261, + 709, + 743, + 739 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 760, + 375, + 776 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recently, large vision-language models (LVLMs) [1, 5, 10, 29, 30, 50, 58, 63] have garnered significant interest within the research community. By leveraging large", + "bbox": [ + 212, + 787, + 787, + 820 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Corresponding author.", + "bbox": [ + 218, + 825, + 385, + 840 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/2fa31c18577b9bef1435711bf43fb3f98dcb8835e0202f63c351b6ba8714afda.jpg", + "image_caption": [ + "Fig. 1: Difference between vanilla instruction tuning and the proposed reflective instruction tuning. Vanilla instruction tuning only trains LVLMs solely for response generation, lacking of supervising the learning of fine-grained reasoning details. Reflective instruction tuning additionally trains the model to reflect the rationale underlying the response, which provides more fine-grained supervision (e.g., the key visual evidence and facts to reach the response, highlighted in red), facilitating the model learning to capture more critical information." + ], + "image_footnote": [], + "bbox": [ + 243, + 145, + 754, + 276 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "language models (LLMs) [8,9,33,42,43,62] as a unified task interface and employing visual instruction tuning [30] for enhanced task-level generalization, LVLMs have showcased remarkable reasoning performance and robustness in following user instructions. These breakthroughs offer substantial promise for the evolution of artificial general intelligence. Despite these advancements, LVLMs continue to be impeded by the phenomenon of hallucination [12, 24], wherein the models generate outputs that are inconsistent with visual content, established world knowledge, or the instructions provided. This significantly undermines the reliability of such models in real-world scenarios. Extensive research has investigated the origins of hallucinations in LVLMs, identifying factors such as noisy and biased training data [13, 15, 28], limited visual perception [1, 7], misalignment between vision and language modalities [18], and linguistic bias [12, 21]. To mitigate these hallucination sources, various mitigation methods have been proposed. These include: creating clean and balanced datasets [28, 55], scaling up visual resolution [1, 29] or utilizing additional perceptual models [51], employing alignment modules with increased parameters [7, 29], and designing decoding strategies that reduce over-reliance on language priors [21].", + "bbox": [ + 212, + 426, + 787, + 684 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Despite significant progress, existing approaches often overlook a crucial factor contributing to hallucinations: the lack of fine-grained reasoning supervision. When learning to solve problems, humans often improve themselves by continuously reflecting on each step involved in the reasoning process. However, most previous LVLMs perform instruction tuning on datasets containing only instructions and direct responses, omitting access to the underlying rationale, i.e., intermediate reasoning steps. Specifically, Fig. 1 (left) shows the training paradigm of vanilla instruction tuning applied to a recognition task. The model is trained solely to predict the animal category name, without explicit supervision on the reasoning details from the discriminating information (e.g., the defining", + "bbox": [ + 212, + 689, + 789, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "J. Zhang et al.", + "bbox": [ + 271, + 114, + 370, + 127 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "attributes of the animal) to the judgment. This strategy risks training the model to establish superficial shortcuts between instructions and responses, rather than understanding the tasks and internalizing the underlying logic of the reasoning process. Consequently, the model becomes susceptible to producing incorrect inferences and hallucinatory content misaligned with visual content.", + "bbox": [ + 212, + 146, + 782, + 220 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Inspired by the human ability to improve reasoning through reflection on intermediate steps [2,38], we propose reflective instruction tuning to address the aforementioned issue. Unlike vanilla instruction tuning, which only trains the model to predict the response, we incorporate the learning of underlying rationales, providing the model with fine-grained reasoning supervision. Specifically, reflective instruction tuning trains the model to generate positive rationales (justifying correct responses) and negative rationales (explaining errors in incorrect responses). The positive rationale explicitly provides more fine-grained visual content, the relationship between multiple visual concepts, and step-by-step reasoning. This facilitates the model to capture finer visual details and better understand the task definition underlying instruction. The negative rationale provides more illustration of the core characteristics linked to the erroneous categorization of negative answers, helping the model concentrate on crucial concepts within the decision-making process. This improves the model's ability to discriminate between correct and potential hallucinatory responses. Fig. 1 (right) presents an example of positive and negative rationale, the positive rationale provides a more detailed justification of the correct answer, and the negative rationale highlights errors in the incorrect answer, aiding discrimination. Reflective instruction tuning mirrors the human cognitive process of reflection, in which individuals analyze step-by-step reasoning and errors to gain a deeper understanding. Training the model with both positive and negative rationales enhances several core cognitive capabilities: it improves visual perception through more visual information, facilitates instruction understanding with coherent reasoning logic, and promotes error awareness by highlighting critical information. Consequently, this training paradigm mitigates hallucinatory outputs.", + "bbox": [ + 212, + 226, + 787, + 604 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "While various visual instruction tuning datasets [4,28,30,44,52,60] exist, most of them consist solely of instruction-response pairs, omitting the annotation of rationales. To facilitate reflective instruction tuning that learns from both positive and negative rationales, we introduce REVERIE, the first large-scale visual instruction tuning dataset with ReflEctiVE rationalE annotations. REVERIE builds upon existing VQA datasets [31,39], and is further extended with natural images from Visual Genome [19]. The positive and negative rationales are annotated by Gemini family multimodal models [41]. REVERIE comprises 115k instructions encompassing four task types. Each instruction includes a positive response, an average of 1.2 negative responses, and accompanying rationales that justify the correctness or incorrectness of each response. This yields a total of 254k training instances, i.e., (instruction, response, rationale) tuples.", + "bbox": [ + 212, + 608, + 787, + 790 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To enhance the adaptability of reflective instruction tuning to diverse tasks, we formulate the rationale data as a conversation format and position rationale learning subsequent to the model's prediction of responses, thereby decoupling", + "bbox": [ + 212, + 795, + 785, + 839 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "Reflective Instruction Tuning", + "bbox": [ + 535, + 114, + 730, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "rationale generation from the context of response prediction. We perform reflective instruction tuning on two LVLMs with the proposed REVERIE dataset and further conduct experiments on multiple commonly used LVLM evaluation benchmarks. The results show that the proposed reflective tuning with REVERIE could yield substantial performance gain over the baselines. We also observed that utilizing the negative rationale could bring additional improvement than merely using the positive rationale. In summary, our contributions are three-fold: (a) We propose reflective instruction tuning, a novel approach that incorporates learning from both positive and negative rationales into visual instruction tuning to mitigate hallucination. (b) We propose REVERIE, the first large-scale instruction tuning dataset with both positive and negative fine-grained rationales. We hope this dataset could facilitate future research on rationale learning in LVLMs. (c) Extensive experiments show that reflective instruction tuning with the proposed dataset could bring substantial performance gain across multiple LVLM evaluation benchmarks, demonstrating the effectiveness of incorporating reflective rationales for mitigating hallucinations.", + "bbox": [ + 212, + 146, + 787, + 388 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 412, + 387, + 429 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Large vision-language models. The remarkable generalization capabilities of LLMs [8,9,33,35,42,43,57,62] have inspired the LVLMs [1,5,10,29,30,50,58,63]. LVLMs aim to extend the reasoning abilities of LLMs into the visual domain, potentially enabling more sophisticated and multimodal understanding. Typically, the architecture of previous LVLMs incorporates three fundamental components: a pre-trained visual encoder (e.g., CLIP [34]) for visual perception, a pre-trained LLM (e.g., LLaMA [42]) serving as a language understanding and generation interface, and a connector (e.g., Q-former [10, 23], linear projector [29, 30, 63]) facilitating vision-language alignment. A common training paradigm for these models involves vision-language pre-training using large-scale vision-text datasets [27, 30, 37, 63] and visual instruction tuning [30] with various visual task datasets [14, 29, 30, 58]. To enhance the generalization capabilities of models, previous methods have explored several techniques, such as: 1) employing stronger perception models [17, 64] or LLMs [36, 56]; 2) designing improved connectors [3, 7, 20]; 3) utilizing a mixture-of-expert design [6, 26]. Given the importance of visual instruction tuning in building general-purpose visual assistants, many studies have also focused on improving performance through the creation of high-quality instruction tuning datasets. These datasets target a wide range of visual understanding capabilities, encompassing text-rich image comprehension [58], region-level analysis [61], chart interpretation [14], and multi-image reasoning [16]. The proposed REVIRIE dataset also aims to improve the model performance during visual instruction tuning, but we focus on enhancing the reasoning capability by offering fine-grained reasoning supervision.", + "bbox": [ + 212, + 446, + 787, + 792 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Learning from reasoning steps. Several studies [11, 25, 31, 46-48] highlight the effectiveness of learning from detailed reasoning steps. Wei et al. [48] propose chain-of-thought prompting, where LLMs are explicitly prompted to think", + "bbox": [ + 212, + 795, + 787, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "J. Zhang et al.", + "bbox": [ + 271, + 114, + 370, + 128 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "step-by-step during inference, leading to substantial gains on reasoning tasks. Furthermore, recent work suggests that training with explicit rationales can improve model reasoning capabilities [32, 54]. Additionally, learning from answer explanations has shown promise in enhancing performance on vision-language reasoning tasks [31, 49, 59]. Despite these advancements, learning from reasoning steps in visual instruction tuning remains under-explored due to the scarcity of large-scale rationale-annotated visual instruction tuning datasets. LLaVA-Instruct-158k [30] contains a portion of annotations for complex reasoning tasks, and the responses have step-by-step reasoning process annotation that is similar to the concept of rationale. However, certain limitations remain: 1) these rationales are all positive coupled with correct responses, neglecting negative rationales crucial for discerning incorrect responses, and 2) they are restricted to open-ended reasoning tasks, offering limited benefits to the diverse range of tasks. We address these gaps by proposing REVERIE, the first large-scale visual instruction tuning dataset that contains both positive and negative rationales on various visual tasks, facilitating the future research on rationale learning in visual instruction tuning.", + "bbox": [ + 212, + 146, + 787, + 402 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Hallucination in LVLMs. Previous research on hallucination in LVLMs has focused on two key areas: 1) the development of benchmarks for evaluating hallucinations [12, 24, 28, 40], and 2) the exploration of techniques to mitigate hallucinations [18, 22, 28, 45]. Li et al. [24] identify the object hallucination issue in LVLMs, and they propose the POPE benchmark for its evaluation. Guan et al. [12] divide the hallucination into language hallucination and visual hallucination and further introduce HALLUSIONBENCH to evaluate and diagnose each hallucination. Various studies address hallucination mitigation. Liu et al. [28] reduce hallucinations with a balanced visual instruction tuning dataset to better handle negative instructions. Jiang et al. [18] employ contrastive learning to distinguish hallucinated and non-hallucinated representations in latent space. Wang et al. [45] and Leng et al. [22] design contrastive decoding strategies in inference to mitigate the hallucination caused by language bias. Similar to [28], our work also focuses on mitigating hallucinations through improved visual instruction datasets. However, we emphasize fine-grained reasoning supervision to enhance consistency throughout the reasoning process.", + "bbox": [ + 212, + 404, + 787, + 645 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3 The REVERIE Dataset", + "text_level": 1, + "bbox": [ + 215, + 670, + 478, + 686 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Despite previous works have provided many diverse visual instruction tuning datasets (e.g., LLaVA-Instruct-158k [30], LRV-Instruction [28]), only a small fraction of samples possess rationale annotations for positive responses. Furthermore, these datasets often omit annotations for hard negative responses and their underlying rationales, rendering them hard to provide fine-grained and discriminative supervision for addressing hallucinations. To facilitate model \"reflect\" from both positive and negative responses, we propose REVERIE, the first large-scale visual instruction tuning dataset with detailed reflective rationale annotation, serving as a basis of the proposed reflective instruction tuning strategy", + "bbox": [ + 212, + 704, + 787, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "Reflective Instruction Tuning", + "bbox": [ + 535, + 114, + 730, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/169c9c074bad62e02529e118d075dac872dd0d5d8b03d07a1d76ffd7c42a9c20.jpg", + "image_caption": [ + "Fig. 2: Overview of the REVERIE dataset's data collection pipeline. We first employ Gemini-Vision-Pro to annotate the instructions, responses, and rationales for each image. Gemini-Pro is then used to check the consistency between positive and negative rationales. Inconsistent samples are filtered to maintain dataset quality." + ], + "image_footnote": [], + "bbox": [ + 246, + 143, + 756, + 333 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "(introduced in Sec. 4). As shown in Table 1, REVERIE provides hard negative response annotation and contains more extensive rationale annotations compared with the existing datasets. Fig. 2 shows the overall pipeline of our data collection strategy, we first annotate instructions, responses, and positive/negative rationales for each collected image, followed by filtering out noisy samples by the consistency-based filtering strategy. The details are as follows (The prompts for data generation are available in supplementary materials).", + "bbox": [ + 212, + 434, + 787, + 541 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/55ad36dec3cef6981313bcb76e7bce9e60a02d7d969c1619e33d90ca0521902d.jpg", + "table_caption": [ + "Table 1: Comparison of the REVERIE with the data used by other LVLMs." + ], + "table_footnote": [], + "table_body": "
REVERIE LLaVA-1.5 InstructBLIP MiniGPT-4 LRV-Instruction
# Training Instances254k665k1.2M3.5k400k
Hard Negative Responses?✓XXXX
# Positive Rationales115k77k77kXX
# Negative Rationales138kXXXX
", + "bbox": [ + 228, + 597, + 769, + 674 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Data collection and annotation. To ensure the dataset encompasses diverse visual concepts and broad domain coverage, we sample a subset of images from the Visual Genome dataset [19]. Subsequently, we annotate these images with instructions and corresponding responses. Given the high cost of manual labelling, we leverage a strong LVLM, Gemini-Vision-Pro, as our annotator. Considering that the rationale for addressing simple query instruction may contain limited intermediate steps and benefit less to learning fine-grained reasoning, we explicitly prompt Gemini to generate instructions that need multi-step to solve and avoid generating simple queries. Additionally, it is prompted to generate both", + "bbox": [ + 212, + 703, + 787, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "J. Zhang et al.", + "bbox": [ + 271, + 114, + 370, + 127 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "positive and negative responses in conjunction with each instruction. Intuitively, a more confusing negative response demands more discriminative information to distinguish it from the positive answer. Thus, we explicitly prompt Gemini to generate negative responses that are easy to confuse.", + "bbox": [ + 212, + 146, + 782, + 205 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "After generating the instructions and responses, we prompt Gemini-Vision-Pro to generate rationales underlying the responses. Considering that positive and negative rationales have different emphases, we design two prompts and generate them separately. The positive rationales are expected to help the model learn the inherent knowledge and visual concepts underlying each problem-solving step, thereby facilitating fine-grained vision-language alignment and understanding of task definitions. It should include 1) analysis of visual content and instructions, 2) identification of core visual concepts relevant to the instructions, and 3) step-by-step reasoning grounded on the visual information and required knowledge. This is achieved by explicitly incorporating words such as \"explanation\", \"details\", and \"knowledge\" into the prompt to facilitate Gemini-Vision-Pro to include key image details and necessary outside knowledge in the rationale. The negative rationales are expected to provide discriminative information that aids in distinguishing between valid and hallucinatory outputs. Thus, we let Gemini-Vision-Pro highlight discriminative details in the negative rationale.", + "bbox": [ + 212, + 210, + 784, + 436 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Consider A-OK-VQA [39] and ScienceQA [31] already have some human annotated QA pairs and positive rationales, we include them into the REVERIE dataset to increase task diversity. Since both datasets provide questions in a multiple-choice format, we incorporate all incorrect answer choices and generate an additional negative rationale for each of them with the same pipeline.", + "bbox": [ + 212, + 440, + 785, + 516 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Consistency-based data filtering. While Gemini-Vision-Pro demonstrates impressive capabilities, it can still generate hallucinatory content when tasked to generate responses and rationales. Previous work [53] has shown that noise training data is one of the sources of the hallucinations. To guarantee the quality of generated data, we propose a method that leverages the consistency between positive and negative rationales to filter out noisy examples. Since both positive and negative rationales explicate the reasoning process behind the instruction but emphasize different aspects, they should contain information that is mutually consistent. When one rationale includes noisy information, this inherent consistency will be disrupted. Inspired by that, we employ a large language model, i.e., Gemini-Pro, to detect inconsistencies between these two types of rationales and only keep the samples that contain intra-consistent information (See supplementary materials for detailed prompt).", + "bbox": [ + 212, + 518, + 785, + 715 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Data statistics. After the removal of noisy samples, the REVERIE dataset comprises 71,558 natural images. This includes 50,938 images sourced from Visual Genome [19], 15,706 from the COCO [27] and 4,914 images from ScienceQA [31]. REVERIE contains 115,280 instructions paired with corresponding positive responses and 138,897 negative responses, where each response is supplemented with a reflective rationale, rendering a total of 254,177 training instances. According to the output formats, REVERIE covers four types of vision-language tasks, including multiple-choice QA, short-answer QA, open-", + "bbox": [ + 212, + 719, + 785, + 839 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "Reflective Instruction Tuning", + "bbox": [ + 535, + 114, + 730, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "ended QA, and Yes/No questions, and their distributions are visualized in Fig. 3 (a). We also analyzed the distribution of rationale length and the frequency of noun occurrence. As shown in Fig. 3 (b) and (c), over $50\\%$ of rationales exceed 25 words in length and contain more than 8 nouns, suggesting that rationales provide a substantial amount of information and knowledge.", + "bbox": [ + 212, + 146, + 787, + 223 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/257896d69dfee4efb0526e08ad5b9e28d9e791df9bb68a3e216a6055cfeaf6fe.jpg", + "image_caption": [ + "(a) Distribution of Different Task" + ], + "image_footnote": [], + "bbox": [ + 256, + 253, + 426, + 335 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/df35c8037a93b05ed1023ccb1dab4d8539b5885612f079474304d88278292497.jpg", + "image_caption": [ + "(b) Distribution of Rationale Length" + ], + "image_footnote": [], + "bbox": [ + 419, + 250, + 581, + 339 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/59648b7be52428021db84a4a1cd804423db5192d912b395c8a4d7c8f407d1bb0.jpg", + "image_caption": [ + "(c) Distribution of noun occurrence", + "Fig. 3: Statistics of the REVERIE dataset." + ], + "image_footnote": [], + "bbox": [ + 584, + 250, + 756, + 339 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4 Reflective Instruction Tuning", + "text_level": 1, + "bbox": [ + 215, + 438, + 537, + 455 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Reflection plays a crucial role in the human learning process, allowing individuals to extract meaningful insights, learn from their mistakes, and make informed choices in the future. However, vanilla instruction tuning solely focuses on training models to predict responses, thus omitting access to the underlying rationale. This limitation impedes the model's ability to effectively internalize and logically process instruction-relevant information, increasing the likelihood of generating hallucinatory outputs. Inspired by humans' ability to improve problem-solving skills through reflection on explanations for both correct and incorrect answers [2,38], we propose reflective instruction tuning. This approach facilitates learning from the rationale behind responses, extending the training process to predict both positive and negative rationales rather than merely generating correct responses to instructions.", + "bbox": [ + 212, + 474, + 787, + 655 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Overall architecture. The objective of visual instruction tuning is to improve the instruction-following capability of the LVLM by leveraging diverse visual task data. Typically, LVLMs consist of a visual encoder, a vision-language connector, and an LLM. Given an image sample $I$ and user instruction $X$ , the visual encoder extracts image token features, which are then projected into the language space by the modality connector. The projected visual tokens, along with the textual instructions, serve as the prefix in the LLM. An auto-regressive LLM is trained to generate response $A$ that aligns with both the instructions and the image.", + "bbox": [ + 212, + 657, + 787, + 777 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Multi-turn conversation. Directly training a model to concurrently predict rationales alongside responses in a single-turn prediction might hinder inference results. This is because some tasks do not necessitate explicit rationale generation, limiting its real-world applicability. To address this, we decouple rationale", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "J. Zhang et al.", + "bbox": [ + 271, + 114, + 370, + 128 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "learning from response prediction to improve the adaptability of reflective instruction tuning to diverse tasks. We achieve this by formulating the rationale data in a multi-turn conversational format. The first turn incorporates $I$ and $X$ as inputs, prompting the model to generate $A$ . Subsequently, the second turn introduces rationale generation prompts to let the model predict rationale. By putting the rationale learning behind, response prediction relies solely on the context of $I$ and $X$ , mitigating interference to the inference results. To make the model understand this rationale generation task, we explicitly prompt the model to generate knowledge and facts that are related to instruction. Specifically, in positive rationale learning, the model receives the prompt \"Explain why. Including any necessary facts or knowledge\" following its response. In negative rationale learning, an incorrect answer is provided, followed by the prompt \"Explain why this answer is wrong: {incorrect answer}. Including any necessary facts or knowledge.\"", + "bbox": [ + 212, + 146, + 787, + 358 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "5 Experiments", + "text_level": 1, + "bbox": [ + 215, + 378, + 375, + 395 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We demonstrate the effectiveness of learning from rationales by applying reflective instruction tuning with the REVERIE dataset to two baseline LVLMs: (1) LLaVA-1.0-7b-lora [30]: We initialize the model pre-trained with LAION-CC-SBU-558k, and perform instruction tuning with the combination of REVERIE and LLaVA-Instruct-80k. (2) LLaVA-1.5-7b-lora [29]: Similar to LLaVA-1.0, we perform instruction tuning with the combination of REVERIE and the original 665k instructions. The resulting models are denoted as REVERIE-1.0-7b-lora and REVERIE-1.5-7b-lora, respectively. Implementation details are provided in the supplementary materials.", + "bbox": [ + 212, + 406, + 789, + 542 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "5.1 Main Results", + "text_level": 1, + "bbox": [ + 215, + 561, + 375, + 575 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Performance comparison with baselines on six benchmarks. We compare the performance of the model reflective tuned with the proposed REVERIE dataset. As shown in Table 2, training with REVERIE yields clear performance gain on most of the benchmarks across 2 baselines. Notably, we achieve a score of 83.8 on the POPE benchmark, reflecting a significant improvement of 12.7 points over the baseline LLaVA-1.0-7b-lora. This suggests that the REVERIE dataset provides more fine-grained supervision, enabling the model to make more accurate judgments regarding object existence and significantly alleviating hallucination issues. Additionally, we observe substantial performance gains on two comprehensive LVLM evaluation benchmarks, MME and MMBench. Specifically, REVERIE-1.0-7b-lora achieves improvements of 321.2 and 27.2 points on MME and MMBench, respectively, compared to the baseline. This suggests that 1) the REVERIE dataset encompasses a wider range of visual concepts and knowledge compared to the original LLaVA-Instruct-80k dataset, demonstrably benefiting performance on diverse visual perception tasks evaluated in MME; 2) incorporating rationales during training empowers the model to develop a deeper understanding of task definitions and effectively distinguish positive answers from", + "bbox": [ + 212, + 583, + 787, + 840 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "Reflective Instruction Tuning", + "bbox": [ + 535, + 114, + 730, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/017a105a32025198033b6b5e2c1ba66ea88142355c14a339b85edb77aa411834.jpg", + "table_caption": [ + "Table 2: Comparison with baselines on six benchmarks. * means training images of the dataset are used during training. † means rerun using their released code." + ], + "table_footnote": [], + "table_body": "
MethodScienceQA1POPEMMEMMBenchMM-VetGQA
InstructBLIP (vicuna-7b)60.5--3626.249.2
Qwen-VL-7b67.1--38.2-59.3
Qwen-VL-Chat-7b68.2-1487.560.6-57.5*
LLaVA-1.5-7b66.885.91510.764.330.562.0*
LLaVA-1.5-7b-lora68.486.41476.966.130.263.0*
LLaVA-1.5-7b-lora†67.786.11471.065.132.562.8*
LLaVA-1.0-7b-lora†42.771.1819.827.230.07.10
REVERIE-1.0-7b-lora70.1*83.81168.155.427.836.5
LLaVA-1.5-7b-lora (w/ ScienceQA)†76.3*86.61439.267.931.160.7*
REVERIE-1.5-7b-lora80.5*86.41474.967.330.861.8*
", + "bbox": [ + 236, + 184, + 759, + 345 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/36dda41332f5a7a8e5808640062a3c827d07a1010095e5d668c4ff49a9467133.jpg", + "table_caption": [ + "Table 3: Results of using different rationales on POPE." + ], + "table_footnote": [], + "table_body": "
SubsetModelAccuracyPrecisionRecall|F1 Score
RandomResponse only86.8789.4284.53
w/ pos rationales85.6087.9383.53
w/ neg rationales86.2587.9884.93
w/ pos + neg rationales87.1590.2184.20
PopularResponse only82.5781.3384.53
w/ pos rationales81.8080.7383.53
w/ neg rationales83.3082.2584.93
w/ pos + neg rationales84.0083.8684.20
AdversarialResponse only76.8373.2584.53
w/ pos rationales77.2374.1983.53
w/ neg rationales77.7374.2484.93
w/ pos + neg rationales79.3776.7884.20
", + "bbox": [ + 245, + 386, + 753, + 595 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "provided negative answers, thereby achieving superior performance on multiple-choice question benchmarks like MMBench. Notably, for a fair comparison, we trained the baseline, LLaVA-1.5-7b-lora, using the combination of its original instruction tuning data and the VQA data from ScienceQA (only data with image context). We observed that simply introducing the ScienceQA QA data led to a performance decline on MME and GQA. This suggests that learning from response data alone may establish unintended shortcuts between instructions and responses, particularly for the complex reasoning instructions in ScienceQA. Our REVERIE-1.5-7b-lora demonstrates significant performance gains over its baseline on ScienceQA, MME, and GQA. Meanwhile, it achieves competitive performance on other benchmarks. This suggests that the rationale from our dataset may help alleviate the shortcut issue.", + "bbox": [ + 212, + 627, + 787, + 808 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Effectiveness of positive/negative rationale. We conduct experiments to analyze the role of different rationales used in reflective instruction tuning. Ta-", + "bbox": [ + 212, + 809, + 785, + 839 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "J. Zhang et al.", + "bbox": [ + 271, + 114, + 370, + 128 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/94bb53c457128eb3a0c60ca71bc89905fc77e3bf058ee5113f324e903d33b4a1.jpg", + "table_caption": [ + "Table 4: Results of using different rationales on MMBench. We report L2-ability performance, where AR, CP, FP-S, FP-C, LR, and RR denote attribute reasoning, coarse perception, fine-grained perception (single-instance), fine-grained perception (cross-instance), logic reasoning, and relation reasoning, respectively." + ], + "table_footnote": [], + "table_body": "
ModelARCPFP-SFP-CLRRRAccuracy
Response only61.6960.0950.5142.7626.6742.6151.70
w/ pos rationales63.1867.1144.1449.1624.1740.8752.13
w/ neg rationales64.6864.4350.8444.8322.5043.4852.30
w/ pos + neg rationales65.1468.4657.2451.5219.1750.4355.44
", + "bbox": [ + 236, + 212, + 759, + 297 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/ab37b76df90d14cf9fc9dde18eaa82f81f17620fd9e77f24398b129767f8339a.jpg", + "table_caption": [ + "Table 5: Results of using different rationales on GQA. We report the semantic-level accuracy to analyze the impact of rationale on questions with different subjects." + ], + "table_footnote": [], + "table_body": "
ModelAttributeObjectRelationGlobal CategoryAccuracy
Response only47.1759.9022.2131.2119.67
w/ pos rationales46.7857.4623.4429.3022.89
w/ neg rationales48.0156.6824.7731.8528.63
w/ pos + neg rationales47.3258.3524.7631.8528.11
36.53
", + "bbox": [ + 236, + 351, + 759, + 436 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "ble 3, Table 4 and Table 5 show the results of using different rationales on POPE, MMBench, and GQA, respectively. The results show that using both positive and negative rationale yields the best performance on most metrics. The results on POPE show that using both two rationales brings the highest F1 score, suggesting that training with rationale may alleviate the instruction-response shortcut and further alleviate the object hallucination issue. The results on MMBench show that training exclusively with positive rationales yields higher perception scores (in terms of CP and FP-C). This indicates that positive rationales enhance the capture of visual details. Conversely, training solely with negative rationales results in superior reasoning performance (in terms of AR and RR), suggesting that negative rationales provide critical clues that improve discriminative ability. We also infer that training with negative rationales could help the model better distinguish the correct answer from the negative response, benefiting more to the task with candidate answers provided, e.g., multiple choice. Our results on GQA further show that both positive and negative rationales contribute to performance gains in questions with category as their subject. This suggests that rationales provide more fine-grained information that enables the model to better understand the defining characteristics of a category.", + "bbox": [ + 215, + 463, + 784, + 734 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "5.2 Ablation Studies", + "text_level": 1, + "bbox": [ + 217, + 756, + 398, + 768 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "We investigate several key designs of reflective instruction tuning and REVERIE: 1) the design of rationale generation prompt; 2) the effectiveness of consistency-based data filtering; 3) the design of conversational context; All experiments are conducted with REVERIE-1.0-7b-lora.", + "bbox": [ + 215, + 779, + 782, + 838 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "Reflective Instruction Tuning", + "bbox": [ + 535, + 114, + 730, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/75c0fee9eb24358348d8939e5fa1cd6dfc5405a36028a78930941f04f55cf593.jpg", + "table_caption": [ + "Table 6: Performance comparison of different rationale prompts." + ], + "table_footnote": [], + "table_body": "
Explicit GuidanceType-SpecificLength ControlMMBPOPEGQAMME
(a)55.483.836.81128.6
(b)✓✓55.283.236.01111.9
(c)✓✓✓54.383.835.81141.0
(d)✓55.483.836.51168.1
", + "bbox": [ + 230, + 171, + 767, + 258 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Rationale generation prompt. We conducted ablation studies to analyze the design of the rationale generation prompt employed in reflective instruction tuning. Specifically, we test the variants of rationale generation prompt from three perspectives: (1) the use of explicit guidance, e.g., \"Including any necessary facts or knowledge\"; 2) the utilization of separate prompts for positive and negative rationales; 3) the use of length-specific prompts. Table 6 summarizes the performance of these prompt variants. Our final choice, represented by Line (d), includes only explicit guidance and yields the best overall performance across four benchmarks. Line (a) demonstrates that using the prompt without explicit guidance leads to a significant performance drop on the MME benchmark. This suggests that providing clear guidance for clarifying the rationale generation task is crucial for aligning rationales with visual content. Moreover, Lines (c) and (d) indicate that overly complex prompts containing excessive constraints may impede the model's understanding of the rationale generation task, adversely affecting performance across most benchmarks.", + "bbox": [ + 212, + 287, + 785, + 513 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Consistency-based data filtering. We conducted experiments on the noise-infused rationale data to investigate the effectiveness of the proposed consistency-based data filtering strategy. Specifically, we train the model with the rationale data before filtering. As shown in Table 7, training with noised data that are not intra-consistent resulted in performance degradation across most benchmarks, suggesting that using the consistency-based filter strategy to maintain data quality is necessary for learning from rationale.", + "bbox": [ + 212, + 515, + 785, + 622 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Table 7: Effect of rationale data filtering on model performance.", + "bbox": [ + 223, + 654, + 501, + 683 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/b4673171eb5d19d86b92ab6b97a067b14fcc67fe27325cde95b1d1c5f6b30386.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelMMBPOPEGQAMME
w/o filtering55.283.536.31132.3
w/ filtering55.483.836.51168.1
", + "bbox": [ + 241, + 695, + 480, + 744 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/42014aac02dabed315f5b726add83b325c67b37b42e5d0468c611e9273a2c032.jpg", + "table_caption": [ + "Table 8: Performance comparison of different conversational contexts." + ], + "table_footnote": [], + "table_body": "
ModelMMBPOPEGQAMME
pos rationale first49.379.536.21144.7
neg rationale first48.682.436.41160.2
separate context55.483.836.51168.1
", + "bbox": [ + 511, + 691, + 781, + 750 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Conversational context of reflective instruction tuning. We conducted ablation studies on several training data format variants, each incorporating different conversational contexts. We test the performance of three variants: 1) positive rationale first, where all negative rationales are learned to predict after", + "bbox": [ + 212, + 779, + 785, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "J. Zhang et al.", + "bbox": [ + 271, + 114, + 370, + 127 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "positive rationale turns; 2) negative rationale first, the inverse order of the previous one; 3) separate context, where the positive and negative rationales are separate into two distinct samples. The previous two variants put the positive and negative rationales for one instruction into a single conversation, making the model see previous rationales when learning to predict the current one. Table 8 shows the performance of these three variants. Our final choice that using separate contexts consistently outperforms the other two variants across four benchmarks. This suggests that providing the model with access to additional rationales during the current rationale generation step might lead to information leakage. In such a scenario, the model may tend to leverage pre-existing information from other rationales rather than learning to reason according to visual information and acquired knowledge. This hinders the model benefited from the fine-grained supervision from rationales.", + "bbox": [ + 212, + 146, + 787, + 343 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "5.3 Qualitative Analysis", + "text_level": 1, + "bbox": [ + 215, + 366, + 429, + 382 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/0e96a6377ce4e56b9b578d60c423e9cfb9125439e30e377ca64b966f595e5ff5.jpg", + "image_caption": [ + "Fig. 4: Visualization of the generation of positive rationales." + ], + "image_footnote": [], + "bbox": [ + 263, + 421, + 738, + 674 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "We conducted a case analysis to assess whether reflective instruction tuning improves fine-grained reasoning and mitigates hallucination. We compared baseline models with our proposed approach by prompting both to generate explicit rationales for their responses, aiming to verify if the models' logic aligns with correct reasoning. Fig. 4 visualizes the output rationales of different models. LLaVA directly generates rationale along with the response. However, the response is incorrect, and the rationale contains hallucinated content (highlighted in red).", + "bbox": [ + 212, + 734, + 787, + 840 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "Reflective Instruction Tuning", + "bbox": [ + 535, + 114, + 730, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 784, + 126 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/d3d8975ded3131c01a0efc384b10725b34d85ae65dca9acd64bd81ebacaff77b.jpg", + "image_caption": [ + "Fig. 5: Visualization of the generation of negative rationales." + ], + "image_footnote": [], + "bbox": [ + 261, + 147, + 738, + 402 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "REVERIE (w/o R) (i.e., the model trained without rationales) and REVERIE both predicted the correct answer. However, REVERIE (w/o R) generates a rationale inconsistent with the response, suggesting that training without fine-grained supervision may lead to the shortcut between instruction and response rather than solving the problem from the correct reasoning paths. In contrast, the rationale generated by REVERIE presents more information that is necessary (highlighted in green) for addressing the instruction. We further evaluate the models' capability to generate negative rationales. As shown in Fig. 5, REVERIE correctly identifies crucial information (e.g., all presented colors) from images, while the other two models are misled by the context of two sub-images. This suggests that incorporating rationale as supervision enhances the model's fine-grained understanding of visual context, reducing hallucination.", + "bbox": [ + 212, + 458, + 787, + 642 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "6 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 660, + 359, + 676 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In this work, we identify a key contributor to hallucinations in existing LVLMs: the lack of fine-grained reasoning supervision during visual instruction tuning. To address this, we introduce reflective instruction tuning, which draws inspiration from human reflective learning processes by training the model on the rationales of positive and negative answers. Additionally, we propose REVERIE, the first large-scale visual instruction tuning dataset featuring comprehensive annotations for both positive and negative rationales. Experiments demonstrate that reflective instruction tuning with REVERIE significantly enhances LVLM performance, emphasizing the importance of fine-grained reasoning supervision in mitigating hallucinations.", + "bbox": [ + 212, + 688, + 787, + 840 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "J. Zhang et al.", + "bbox": [ + 271, + 114, + 370, + 128 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgement", + "text_level": 1, + "bbox": [ + 217, + 143, + 392, + 162 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "This work is supported by the National Natural Science Foundation of China (Grant NO.62122035).", + "bbox": [ + 215, + 178, + 784, + 209 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 233, + 321, + 250 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Bai, J., Bai, S., Yang, S., Wang, S., Tan, S., Wang, P., Lin, J., Zhou, C., Zhou, J.: Qwen-vl: A frontier large vision-language model with versatile abilities. arXiv preprint arXiv:2308.12966 (2023)", + "2. Boud, D., Keogh, R., Walker, D.: Reflection: Turning experience into learning. Routledge (2013)", + "3. Cha, J., Kang, W., Mun, J., Roh, B.: Honeybee: Locality-enhanced projector for multimodal llm. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13817-13827 (2024)", + "4. Chen, D., Liu, J., Dai, W., Wang, B.: Visual instruction tuning with polite flamingo. arXiv preprint arXiv:2307.01003 (2023)", + "5. Chen, K., Zhang, Z., Zeng, W., Zhang, R., Zhu, F., Zhao, R.: Shikra: Unleashing multimodal llm's referential dialogue magic. arXiv preprint arXiv:2306.15195 (2023)", + "6. Chen, S., Jie, Z., Ma, L.: Llava-mole: Sparse mixture of lora experts for mitigating data conflicts in instruction finetuning mllms. arXiv preprint arXiv:2401.16160 (2024)", + "7. Chen, Z., Wu, J., Wang, W., Su, W., Chen, G., Xing, S., Muyan, Z., Zhang, Q., Zhu, X., Lu, L., et al.: Internvl: Scaling up vision foundation models and aligning for generic visual-linguistic tasks. arXiv preprint arXiv:2312.14238 (2023)", + "8. Chowdhery, A., Narang, S., Devlin, J., Bosma, M., Mishra, G., Roberts, A., Barham, P., Chung, H.W., Sutton, C., Gehrmann, S., et al.: Palm: Scaling language modeling with pathways. Journal of Machine Learning Research 24(240), 1-113 (2023)", + "9. Chung, H.W., Hou, L., Longpre, S., Zoph, B., Tay, Y., Fedus, W., Li, Y., Wang, X., Dehghani, M., Brahma, S., et al.: Scaling instruction-finetuned language models. arXiv preprint arXiv:2210.11416 (2022)", + "0. Dai, W., Li, J., Li, D., Tiong, A.M.H., Zhao, J., Wang, W., Li, B., Fung, P., Hoi, S.: Instructblip: Towards general-purpose vision-language models with instruction tuning (2023)", + "1. Ge, J., Luo, H., Qian, S., Gan, Y., Fu, J., Zhan, S.: Chain of thought prompt tuning in vision language models. arXiv preprint arXiv:2304.07919 (2023)", + "2. Guan, T., Liu, F., Li, X.W.R.X.Z., Wang, X.L.X., Yacoob, L.C.F.H.Y., Zhou, D.M.T.: Hallusionbench: An advanced diagnostic suite for entangled language hallucination & visual illusion in large vision-language models", + "3. Gunjal, A., Yin, J., Bas, E.: Detecting and preventing hallucinations in large vision language models. arXiv preprint arXiv:2308.06394 (2023)", + "4. Han, Y., Zhang, C., Chen, X., Yang, X., Wang, Z., Yu, G., Fu, B., Zhang, H.: Chartllama: A multimodal llm for chart understanding and generation. arXiv preprint arXiv:2311.16483 (2023)", + "5. Hu, H., Zhang, J., Zhao, M., Sun, Z.: Ciem: Contrastive instruction evaluation method for better instruction tuning. arXiv preprint arXiv:2309.02301 (2023)" + ], + "bbox": [ + 225, + 266, + 785, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "Reflective Instruction Tuning", + "bbox": [ + 535, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "16. Huang, Y., Meng, Z., Liu, F., Su, Y., Collier, N., Lu, Y.: Sparkles: Unlocking chats across multiple images for multimodal instruction-following models. arXiv preprint arXiv:2308.16463 (2023)", + "17. Jain, J., Yang, J., Shi, H.: Vcoder: Versatile vision encoders for multimodal large language models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 27992-28002 (2024)", + "18. Jiang, C., Xu, H., Dong, M., Chen, J., Ye, W., Yan, M., Ye, Q., Zhang, J., Huang, F., Zhang, S.: Hallucination augmented contrastive learning for multimodal large language model. arXiv preprint arXiv:2312.06968 (2023)", + "19. Krishna, R., Zhu, Y., Groth, O., Johnson, J., Hata, K., Kravitz, J., Chen, S., Kalantidis, Y., Li, L.J., Shamma, D.A., et al.: Visual genome: Connecting language and vision using crowdsourced dense image annotations. International journal of computer vision 123, 32-73 (2017)", + "20. Lee, B.K., Kim, C.W., Park, B., Ro, Y.M.: Meteor: Mamba-based traversal of rationale for large language and vision models. arXiv preprint arXiv:2405.15574 (2024)", + "21. Leng, S., Zhang, H., Chen, G., Li, X., Lu, S., Miao, C., Bing, L.: Mitigating object hallucinations in large vision-language models through visual contrastive decoding. arXiv preprint arXiv:2311.16922 (2023)", + "22. Leng, S., Zhang, H., Chen, G., Li, X., Lu, S., Miao, C., Bing, L.: Mitigating object hallucinations in large vision-language models through visual contrastive decoding. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13872-13882 (2024)", + "23. Li, J., Li, D., Savarese, S., Hoi, S.: Blip-2: Bootstrapping language-image pretraining with frozen image encoders and large language models. arXiv preprint arXiv:2301.12597 (2023)", + "24. Li, Y., Du, Y., Zhou, K., Wang, J., Zhao, W.X., Wen, J.R.: Evaluating object hallucination in large vision-language models. arXiv preprint arXiv:2305.10355 (2023)", + "25. Lightman, H., Kosaraju, V., Burda, Y., Edwards, H., Baker, B., Lee, T., Leike, J., Schulman, J., Sutskever, I., Cobbe, K.: Let's verify step by step. arXiv preprint arXiv:2305.20050 (2023)", + "26. Lin, B., Tang, Z., Ye, Y., Cui, J., Zhu, B., Jin, P., Zhang, J., Ning, M., Yuan, L.: Moe-llava: Mixture of experts for large vision-language models. arXiv preprint arXiv:2401.15947 (2024)", + "27. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dólar, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13. pp. 740-755. Springer (2014)", + "28. Liu, F., Lin, K., Li, L., Wang, J., Yacoob, Y., Wang, L.: Mitigating hallucination in large multi-modal models via robust instruction tuning. In: The Twelfth International Conference on Learning Representations (2023)", + "29. Liu, H., Li, C., Li, Y., Lee, Y.J.: Improved baselines with visual instruction tuning. arXiv preprint arXiv:2310.03744 (2023)", + "30. Liu, H., Li, C., Wu, Q., Lee, Y.J.: Visual instruction tuning. Advances in neural information processing systems 36 (2024)", + "31. Lu, P., Mishra, S., Xia, T., Qiu, L., Chang, K.W., Zhu, S.C., Tafjord, O., Clark, P., Kalyan, A.: Learn to explain: Multimodal reasoning via thought chains for science question answering. Advances in Neural Information Processing Systems 35, 2507-2521 (2022)", + "32. Magister, L.C., Mallinson, J., Adamek, J., Malmi, E., Severyn, A.: Teaching small language models to reason. arXiv preprint arXiv:2212.08410 (2022)" + ], + "bbox": [ + 215, + 146, + 784, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "J. Zhang et al.", + "bbox": [ + 271, + 114, + 370, + 127 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "33. Ouyang, L., Wu, J., Jiang, X., Almeida, D., Wainwright, C., Mishkin, P., Zhang, C., Agarwal, S., Slama, K., Ray, A., et al.: Training language models to follow instructions with human feedback. Advances in Neural Information Processing Systems 35, 27730-27744 (2022)", + "34. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021)", + "35. Raffel, C., Shazeer, N., Roberts, A., Lee, K., Narang, S., Matena, M., Zhou, Y., Li, W., Liu, P.J.: Exploring the limits of transfer learning with a unified text-to-text transformer. The Journal of Machine Learning Research 21(1), 5485-5551 (2020)", + "36. Rasheed, H., Maaz, M., Khan, S., Khan, F.S.: Llava++: Extending visual capabilities with llama-3 and phi-3 (2024), https://github.com/mbzuaai-oryx/LLaVA-pp", + "37. Schuhmann, C., Beaumont, R., Vencu, R., Gordon, C., Wightman, R., Cherti, M., Coombes, T., Katta, A., Mullis, C., Wortsman, M., et al.: Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems 35, 25278-25294 (2022)", + "38. Schunk, D.H., Zimmerman, B.J.: Self-regulated learning: From teaching to self-reflective practice. Guilford Press (1998)", + "39. Schwenk, D., Khandelwal, A., Clark, C., Marino, K., Mottaghi, R.: A-okvqa: A benchmark for visual question answering using world knowledge. In: European Conference on Computer Vision. pp. 146-162. Springer (2022)", + "40. Sun, Z., Shen, S., Cao, S., Liu, H., Li, C., Shen, Y., Gan, C., Gui, L.Y., Wang, Y.X., Yang, Y., et al.: Aligning large multimodal models with factually augmented rlhf. arXiv preprint arXiv:2309.14525 (2023)", + "41. Team, G., Anil, R., Borgeaud, S., Wu, Y., Alayrac, J.B., Yu, J., Soricut, R., Schalkwyk, J., Dai, A.M., Hauth, A., et al.: Gemini: a family of highly capable multimodal models. arXiv preprint arXiv:2312.11805 (2023)", + "42. Touvron, H., Lavril, T., Izacard, G., Martinet, X., Lachaux, M.A., Lacroix, T., Rozière, B., Goyal, N., Hambro, E., Azhar, F., Rodriguez, A., Joulin, A., Grave, E., Lample, G.: Llama: Open and efficient foundation language models (2023)", + "43. Touvron, H., Martin, L., Stone, K., Albert, P., Almahairi, A., Babaei, Y., Bashlykov, N., Batra, S., Bhargava, P., Bhosale, S., et al.: Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288 (2023)", + "44. Wang, J., Meng, L., Weng, Z., He, B., Wu, Z., Jiang, Y.G.: To see is to believe: Prompting gpt-4v for better visual instruction tuning. arXiv preprint arXiv:2311.07574 (2023)", + "45. Wang, X., Pan, J., Ding, L., Biemann, C.: Mitigating hallucinations in large vision-language models with instruction contrastive decoding. arXiv preprint arXiv:2403.18715 (2024)", + "46. Wang, X., Wei, J., Schuurmans, D., Le, Q., Chi, E., Narang, S., Chowdhery, A., Zhou, D.: Self-consistency improves chain of thought reasoning in language models. arXiv preprint arXiv:2203.11171 (2022)", + "47. Wang, X., Wei, J., Schuurmans, D., Le, Q., Chi, E., Zhou, D.: Rationale-augmented ensembles in language models. arXiv preprint arXiv:2207.00747 (2022)", + "48. Wei, J., Wang, X., Schuurmans, D., Bosma, M., Xia, F., Chi, E., Le, Q.V., Zhou, D., et al.: Chain-of-thought prompting elicits reasoning in large language models. Advances in Neural Information Processing Systems 35, 24824-24837 (2022)", + "49. Wu, Y., Zhang, P., Xiong, W., Oguz, B., Gee, J.C., Nie, Y.: The role of chain-of-thought in complex vision-language reasoning task. arXiv preprint arXiv:2311.09193 (2023)" + ], + "bbox": [ + 212, + 146, + 787, + 839 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "Reflective Instruction Tuning", + "bbox": [ + 535, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 784, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "50. Ye, Q., Xu, H., Xu, G., Ye, J., Yan, M., Zhou, Y., Wang, J., Hu, A., Shi, P., Shi, Y., et al.: mplug-owl: Modularization empowers large language models with multimodality. arXiv preprint arXiv:2304.14178 (2023)", + "51. Yin, S., Fu, C., Zhao, S., Xu, T., Wang, H., Sui, D., Shen, Y., Li, K., Sun, X., Chen, E.: Woodpecker: Hallucination correction for multimodal large language models. arXiv preprint arXiv:2310.16045 (2023)", + "52. Yin, Z., Wang, J., Cao, J., Shi, Z., Liu, D., Li, M., Huang, X., Wang, Z., Sheng, L., Bai, L., et al.: Lamm: Language-assisted multi-modal instruction-tuning dataset, framework, and benchmark. Advances in Neural Information Processing Systems 36 (2024)", + "53. Yu, Q., Li, J., Wei, L., Pang, L., Ye, W., Qin, B., Tang, S., Tian, Q., Zhuang, Y.: Hallucidoctor: Mitigating hallucinatory toxicity in visual instruction data (2023)", + "54. Zelikman, E., Mu, J., Goodman, N.D., Wu, Y.T.: Star: Self-taught reasoner bootstrapping reasoning with reasoning (2022)", + "55. Zhai, B., Yang, S., Zhao, X., Xu, C., Shen, S., Zhao, D., Keutzer, K., Li, M., Yan, T., Fan, X.: Halle-switch: Rethinking and controlling object existence hallucinations in large vision language models for detailed caption. arXiv preprint arXiv:2310.01779 (2023)", + "56. Zhang, P., Dong, X., Wang, B., Cao, Y., Xu, C., Ouyang, L., Zhao, Z., Ding, S., Zhang, S., Duan, H., Zhang, W., Yan, H., Zhang, X., Li, W., Li, J., Chen, K., He, C., Zhang, X., Qiao, Y., Lin, D., Wang, J.: Internl m-xcomposer: A vision-language large model for advanced text-image comprehension and composition. arXiv preprint arXiv:2309.15112 (2023)", + "57. Zhang, S., Roller, S., Goyal, N., Artetxe, M., Chen, M., Chen, S., Dewan, C., Diab, M., Li, X., Lin, X.V., et al.: Opt: Open pre-trained transformer language models. arXiv preprint arXiv:2205.01068 (2022)", + "58. Zhang, Y., Zhang, R., Gu, J., Zhou, Y., Lipka, N., Yang, D., Sun, T.: Llavar: Enhanced visual instruction tuning for text-rich image understanding. arXiv preprint arXiv:2306.17107 (2023)", + "59. Zhang, Z., Zhang, A., Li, M., Zhao, H., Karypis, G., Smola, A.: Multimodal chain-of-thought reasoning in language models. arXiv preprint arXiv:2302.00923 (2023)", + "60. Zhao, B., Wu, B., Huang, T.: Svit: Scaling up visual instruction tuning. arXiv preprint arXiv:2307.04087 (2023)", + "61. Zhao, L., Yu, E., Ge, Z., Yang, J., Wei, H., Zhou, H., Sun, J., Peng, Y., Dong, R., Han, C., et al.: Chatspot: Bootstrapping multimodal llms via precise referring instruction tuning. arXiv preprint arXiv:2307.09474 (2023)", + "62. Zheng, L., Chiang, W.L., Sheng, Y., Zhuang, S., Wu, Z., Zhuang, Y., Lin, Z., Li, Z., Li, D., Xing, E.P., Zhang, H., Gonzalez, J.E., Stoica, I.: Judging lvm-as-a-judge with mt-bench and chatbot arena (2023)", + "63. Zhu, D., Chen, J., Shen, X., Li, X., Elhoseiny, M.: Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592 (2023)", + "64. Zong, Z., Ma, B., Shen, D., Song, G., Shao, H., Jiang, D., Li, H., Liu, Y.: Mova: Adapting mixture of vision experts to multimodal context. arXiv preprint arXiv:2404.13046 (2024)" + ], + "bbox": [ + 215, + 146, + 784, + 770 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "J. Zhang et al.", + "bbox": [ + 271, + 114, + 370, + 128 + ], + "page_idx": 17 + } +] \ No newline at end of file diff --git a/2024/Reflective Instruction Tuning_ Mitigating Hallucinations in Large Vision-Language Models/8b065bde-2d53-44c1-a90a-d68b84da76bf_model.json b/2024/Reflective Instruction Tuning_ Mitigating Hallucinations in Large Vision-Language Models/8b065bde-2d53-44c1-a90a-d68b84da76bf_model.json new file mode 100644 index 0000000000000000000000000000000000000000..2aa53a0d2d06a39af13272e784ec1c337cb0eeab --- /dev/null +++ b/2024/Reflective Instruction Tuning_ Mitigating Hallucinations in Large Vision-Language Models/8b065bde-2d53-44c1-a90a-d68b84da76bf_model.json @@ -0,0 +1,2161 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.224, + 0.142, + 0.78, + 0.187 + ], + "angle": 0, + "content": "Reflective Instruction Tuning: Mitigating Hallucinations in Large Vision-Language Models" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.212, + 0.786, + 0.229 + ], + "angle": 0, + "content": "Jinrui Zhang\\(^{1}\\), Teng Wang\\(^{1,2}\\), Haigang Zhang\\(^{3}\\), Ping Lu\\(^{4}\\), Feng Zheng\\(^{1,5}\\)" + }, + { + "type": "text", + "bbox": [ + 0.337, + 0.239, + 0.667, + 0.254 + ], + "angle": 0, + "content": "\\(^{1}\\) Southern University of Science and Technology" + }, + { + "type": "text", + "bbox": [ + 0.394, + 0.254, + 0.61, + 0.269 + ], + "angle": 0, + "content": "2 The University of Hong Kong" + }, + { + "type": "text", + "bbox": [ + 0.384, + 0.269, + 0.619, + 0.282 + ], + "angle": 0, + "content": "3 Shenzhen Polytechnic University" + }, + { + "type": "text", + "bbox": [ + 0.295, + 0.282, + 0.71, + 0.296 + ], + "angle": 0, + "content": "4 The Cloud Computing and IT Institute of ZTE Corporation" + }, + { + "type": "text", + "bbox": [ + 0.276, + 0.296, + 0.726, + 0.324 + ], + "angle": 0, + "content": "\\(^{5}\\) Research Institute of Multiple Agents and Embodied Intelligence, Peng Cheng Laboratory, Shenzhen, China" + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.352, + 0.744, + 0.699 + ], + "angle": 0, + "content": "Abstract. Large vision-language models (LVLMs) have shown promising performance on a variety of vision-language tasks. However, they remain susceptible to hallucinations, generating outputs misaligned with visual content or instructions. While various mitigation strategies have been proposed, they often neglect a key contributor to hallucinations: lack of fine-grained reasoning supervision during training. Without intermediate reasoning steps, models may establish superficial shortcuts between instructions and responses, failing to internalize the inherent reasoning logic. To address this challenge, we propose reflective instruction tuning, which integrates rationale learning into visual instruction tuning. Unlike previous methods that learning from responses only, our approach entails the model predicting rationales justifying why responses are correct or incorrect. This fosters a deeper engagement with the fine-grained reasoning underlying each response, thus enhancing the model's reasoning proficiency. To facilitate this approach, we propose REVERIE, the first large-scale instruction-tuning dataset with ReflEctiVE Rationale annotations. REVERIE comprises 115k machine-generated reasoning instructions, each meticulously annotated with a corresponding pair of correct and confusing responses, alongside comprehensive rationales elucidating the justification behind the correctness or erroneousness of each response. Experimental results on multiple LVLM benchmarks reveal that reflective instruction tuning with the REVERIE dataset yields noticeable performance gain over the baseline model, demonstrating the effectiveness of reflecting from the rationales. Project page is at https://zjr2000.github.io/projects/reverie" + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.71, + 0.744, + 0.74 + ], + "angle": 0, + "content": "Keywords: Large Vision Language Models \\(\\cdot\\) Visual Instruction Tuning \\(\\cdot\\) Hallucination Mitigation" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.761, + 0.377, + 0.777 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.788, + 0.788, + 0.821 + ], + "angle": 0, + "content": "Recently, large vision-language models (LVLMs) [1, 5, 10, 29, 30, 50, 58, 63] have garnered significant interest within the research community. By leveraging large" + }, + { + "type": "page_footnote", + "bbox": [ + 0.22, + 0.826, + 0.386, + 0.841 + ], + "angle": 0, + "content": "* Corresponding author." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.371, + 0.128 + ], + "angle": 0, + "content": "J. Zhang et al." + }, + { + "type": "image", + "bbox": [ + 0.244, + 0.146, + 0.756, + 0.277 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.29, + 0.789, + 0.389 + ], + "angle": 0, + "content": "Fig. 1: Difference between vanilla instruction tuning and the proposed reflective instruction tuning. Vanilla instruction tuning only trains LVLMs solely for response generation, lacking of supervising the learning of fine-grained reasoning details. Reflective instruction tuning additionally trains the model to reflect the rationale underlying the response, which provides more fine-grained supervision (e.g., the key visual evidence and facts to reach the response, highlighted in red), facilitating the model learning to capture more critical information." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.427, + 0.789, + 0.685 + ], + "angle": 0, + "content": "language models (LLMs) [8,9,33,42,43,62] as a unified task interface and employing visual instruction tuning [30] for enhanced task-level generalization, LVLMs have showcased remarkable reasoning performance and robustness in following user instructions. These breakthroughs offer substantial promise for the evolution of artificial general intelligence. Despite these advancements, LVLMs continue to be impeded by the phenomenon of hallucination [12, 24], wherein the models generate outputs that are inconsistent with visual content, established world knowledge, or the instructions provided. This significantly undermines the reliability of such models in real-world scenarios. Extensive research has investigated the origins of hallucinations in LVLMs, identifying factors such as noisy and biased training data [13, 15, 28], limited visual perception [1, 7], misalignment between vision and language modalities [18], and linguistic bias [12, 21]. To mitigate these hallucination sources, various mitigation methods have been proposed. These include: creating clean and balanced datasets [28, 55], scaling up visual resolution [1, 29] or utilizing additional perceptual models [51], employing alignment modules with increased parameters [7, 29], and designing decoding strategies that reduce over-reliance on language priors [21]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.69, + 0.79, + 0.842 + ], + "angle": 0, + "content": "Despite significant progress, existing approaches often overlook a crucial factor contributing to hallucinations: the lack of fine-grained reasoning supervision. When learning to solve problems, humans often improve themselves by continuously reflecting on each step involved in the reasoning process. However, most previous LVLMs perform instruction tuning on datasets containing only instructions and direct responses, omitting access to the underlying rationale, i.e., intermediate reasoning steps. Specifically, Fig. 1 (left) shows the training paradigm of vanilla instruction tuning applied to a recognition task. The model is trained solely to predict the animal category name, without explicit supervision on the reasoning details from the discriminating information (e.g., the defining" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.536, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Reflective Instruction Tuning" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.222 + ], + "angle": 0, + "content": "attributes of the animal) to the judgment. This strategy risks training the model to establish superficial shortcuts between instructions and responses, rather than understanding the tasks and internalizing the underlying logic of the reasoning process. Consequently, the model becomes susceptible to producing incorrect inferences and hallucinatory content misaligned with visual content." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.227, + 0.788, + 0.605 + ], + "angle": 0, + "content": "Inspired by the human ability to improve reasoning through reflection on intermediate steps [2,38], we propose reflective instruction tuning to address the aforementioned issue. Unlike vanilla instruction tuning, which only trains the model to predict the response, we incorporate the learning of underlying rationales, providing the model with fine-grained reasoning supervision. Specifically, reflective instruction tuning trains the model to generate positive rationales (justifying correct responses) and negative rationales (explaining errors in incorrect responses). The positive rationale explicitly provides more fine-grained visual content, the relationship between multiple visual concepts, and step-by-step reasoning. This facilitates the model to capture finer visual details and better understand the task definition underlying instruction. The negative rationale provides more illustration of the core characteristics linked to the erroneous categorization of negative answers, helping the model concentrate on crucial concepts within the decision-making process. This improves the model's ability to discriminate between correct and potential hallucinatory responses. Fig. 1 (right) presents an example of positive and negative rationale, the positive rationale provides a more detailed justification of the correct answer, and the negative rationale highlights errors in the incorrect answer, aiding discrimination. Reflective instruction tuning mirrors the human cognitive process of reflection, in which individuals analyze step-by-step reasoning and errors to gain a deeper understanding. Training the model with both positive and negative rationales enhances several core cognitive capabilities: it improves visual perception through more visual information, facilitates instruction understanding with coherent reasoning logic, and promotes error awareness by highlighting critical information. Consequently, this training paradigm mitigates hallucinatory outputs." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.609, + 0.788, + 0.791 + ], + "angle": 0, + "content": "While various visual instruction tuning datasets [4,28,30,44,52,60] exist, most of them consist solely of instruction-response pairs, omitting the annotation of rationales. To facilitate reflective instruction tuning that learns from both positive and negative rationales, we introduce REVERIE, the first large-scale visual instruction tuning dataset with ReflEctiVE rationalE annotations. REVERIE builds upon existing VQA datasets [31,39], and is further extended with natural images from Visual Genome [19]. The positive and negative rationales are annotated by Gemini family multimodal models [41]. REVERIE comprises 115k instructions encompassing four task types. Each instruction includes a positive response, an average of 1.2 negative responses, and accompanying rationales that justify the correctness or incorrectness of each response. This yields a total of 254k training instances, i.e., (instruction, response, rationale) tuples." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.796, + 0.787, + 0.84 + ], + "angle": 0, + "content": "To enhance the adaptability of reflective instruction tuning to diverse tasks, we formulate the rationale data as a conversation format and position rationale learning subsequent to the model's prediction of responses, thereby decoupling" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.371, + 0.129 + ], + "angle": 0, + "content": "J. Zhang et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.39 + ], + "angle": 0, + "content": "rationale generation from the context of response prediction. We perform reflective instruction tuning on two LVLMs with the proposed REVERIE dataset and further conduct experiments on multiple commonly used LVLM evaluation benchmarks. The results show that the proposed reflective tuning with REVERIE could yield substantial performance gain over the baselines. We also observed that utilizing the negative rationale could bring additional improvement than merely using the positive rationale. In summary, our contributions are three-fold: (a) We propose reflective instruction tuning, a novel approach that incorporates learning from both positive and negative rationales into visual instruction tuning to mitigate hallucination. (b) We propose REVERIE, the first large-scale instruction tuning dataset with both positive and negative fine-grained rationales. We hope this dataset could facilitate future research on rationale learning in LVLMs. (c) Extensive experiments show that reflective instruction tuning with the proposed dataset could bring substantial performance gain across multiple LVLM evaluation benchmarks, demonstrating the effectiveness of incorporating reflective rationales for mitigating hallucinations." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.414, + 0.388, + 0.43 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.447, + 0.789, + 0.794 + ], + "angle": 0, + "content": "Large vision-language models. The remarkable generalization capabilities of LLMs [8,9,33,35,42,43,57,62] have inspired the LVLMs [1,5,10,29,30,50,58,63]. LVLMs aim to extend the reasoning abilities of LLMs into the visual domain, potentially enabling more sophisticated and multimodal understanding. Typically, the architecture of previous LVLMs incorporates three fundamental components: a pre-trained visual encoder (e.g., CLIP [34]) for visual perception, a pre-trained LLM (e.g., LLaMA [42]) serving as a language understanding and generation interface, and a connector (e.g., Q-former [10, 23], linear projector [29, 30, 63]) facilitating vision-language alignment. A common training paradigm for these models involves vision-language pre-training using large-scale vision-text datasets [27, 30, 37, 63] and visual instruction tuning [30] with various visual task datasets [14, 29, 30, 58]. To enhance the generalization capabilities of models, previous methods have explored several techniques, such as: 1) employing stronger perception models [17, 64] or LLMs [36, 56]; 2) designing improved connectors [3, 7, 20]; 3) utilizing a mixture-of-expert design [6, 26]. Given the importance of visual instruction tuning in building general-purpose visual assistants, many studies have also focused on improving performance through the creation of high-quality instruction tuning datasets. These datasets target a wide range of visual understanding capabilities, encompassing text-rich image comprehension [58], region-level analysis [61], chart interpretation [14], and multi-image reasoning [16]. The proposed REVIRIE dataset also aims to improve the model performance during visual instruction tuning, but we focus on enhancing the reasoning capability by offering fine-grained reasoning supervision." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.796, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Learning from reasoning steps. Several studies [11, 25, 31, 46-48] highlight the effectiveness of learning from detailed reasoning steps. Wei et al. [48] propose chain-of-thought prompting, where LLMs are explicitly prompted to think" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.536, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Reflective Instruction Tuning" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.403 + ], + "angle": 0, + "content": "step-by-step during inference, leading to substantial gains on reasoning tasks. Furthermore, recent work suggests that training with explicit rationales can improve model reasoning capabilities [32, 54]. Additionally, learning from answer explanations has shown promise in enhancing performance on vision-language reasoning tasks [31, 49, 59]. Despite these advancements, learning from reasoning steps in visual instruction tuning remains under-explored due to the scarcity of large-scale rationale-annotated visual instruction tuning datasets. LLaVA-Instruct-158k [30] contains a portion of annotations for complex reasoning tasks, and the responses have step-by-step reasoning process annotation that is similar to the concept of rationale. However, certain limitations remain: 1) these rationales are all positive coupled with correct responses, neglecting negative rationales crucial for discerning incorrect responses, and 2) they are restricted to open-ended reasoning tasks, offering limited benefits to the diverse range of tasks. We address these gaps by proposing REVERIE, the first large-scale visual instruction tuning dataset that contains both positive and negative rationales on various visual tasks, facilitating the future research on rationale learning in visual instruction tuning." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.405, + 0.788, + 0.646 + ], + "angle": 0, + "content": "Hallucination in LVLMs. Previous research on hallucination in LVLMs has focused on two key areas: 1) the development of benchmarks for evaluating hallucinations [12, 24, 28, 40], and 2) the exploration of techniques to mitigate hallucinations [18, 22, 28, 45]. Li et al. [24] identify the object hallucination issue in LVLMs, and they propose the POPE benchmark for its evaluation. Guan et al. [12] divide the hallucination into language hallucination and visual hallucination and further introduce HALLUSIONBENCH to evaluate and diagnose each hallucination. Various studies address hallucination mitigation. Liu et al. [28] reduce hallucinations with a balanced visual instruction tuning dataset to better handle negative instructions. Jiang et al. [18] employ contrastive learning to distinguish hallucinated and non-hallucinated representations in latent space. Wang et al. [45] and Leng et al. [22] design contrastive decoding strategies in inference to mitigate the hallucination caused by language bias. Similar to [28], our work also focuses on mitigating hallucinations through improved visual instruction datasets. However, we emphasize fine-grained reasoning supervision to enhance consistency throughout the reasoning process." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.671, + 0.48, + 0.687 + ], + "angle": 0, + "content": "3 The REVERIE Dataset" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.705, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Despite previous works have provided many diverse visual instruction tuning datasets (e.g., LLaVA-Instruct-158k [30], LRV-Instruction [28]), only a small fraction of samples possess rationale annotations for positive responses. Furthermore, these datasets often omit annotations for hard negative responses and their underlying rationales, rendering them hard to provide fine-grained and discriminative supervision for addressing hallucinations. To facilitate model \"reflect\" from both positive and negative responses, we propose REVERIE, the first large-scale visual instruction tuning dataset with detailed reflective rationale annotation, serving as a basis of the proposed reflective instruction tuning strategy" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.371, + 0.128 + ], + "angle": 0, + "content": "J. Zhang et al." + }, + { + "type": "image", + "bbox": [ + 0.248, + 0.145, + 0.758, + 0.334 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.345, + 0.788, + 0.404 + ], + "angle": 0, + "content": "Fig. 2: Overview of the REVERIE dataset's data collection pipeline. We first employ Gemini-Vision-Pro to annotate the instructions, responses, and rationales for each image. Gemini-Pro is then used to check the consistency between positive and negative rationales. Inconsistent samples are filtered to maintain dataset quality." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.435, + 0.788, + 0.542 + ], + "angle": 0, + "content": "(introduced in Sec. 4). As shown in Table 1, REVERIE provides hard negative response annotation and contains more extensive rationale annotations compared with the existing datasets. Fig. 2 shows the overall pipeline of our data collection strategy, we first annotate instructions, responses, and positive/negative rationales for each collected image, followed by filtering out noisy samples by the consistency-based filtering strategy. The details are as follows (The prompts for data generation are available in supplementary materials)." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.57, + 0.786, + 0.586 + ], + "angle": 0, + "content": "Table 1: Comparison of the REVERIE with the data used by other LVLMs." + }, + { + "type": "table", + "bbox": [ + 0.229, + 0.598, + 0.77, + 0.675 + ], + "angle": 0, + "content": "
REVERIE LLaVA-1.5 InstructBLIP MiniGPT-4 LRV-Instruction
# Training Instances254k665k1.2M3.5k400k
Hard Negative Responses?✓XXXX
# Positive Rationales115k77k77kXX
# Negative Rationales138kXXXX
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.704, + 0.789, + 0.842 + ], + "angle": 0, + "content": "Data collection and annotation. To ensure the dataset encompasses diverse visual concepts and broad domain coverage, we sample a subset of images from the Visual Genome dataset [19]. Subsequently, we annotate these images with instructions and corresponding responses. Given the high cost of manual labelling, we leverage a strong LVLM, Gemini-Vision-Pro, as our annotator. Considering that the rationale for addressing simple query instruction may contain limited intermediate steps and benefit less to learning fine-grained reasoning, we explicitly prompt Gemini to generate instructions that need multi-step to solve and avoid generating simple queries. Additionally, it is prompted to generate both" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.536, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Reflective Instruction Tuning" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.207 + ], + "angle": 0, + "content": "positive and negative responses in conjunction with each instruction. Intuitively, a more confusing negative response demands more discriminative information to distinguish it from the positive answer. Thus, we explicitly prompt Gemini to generate negative responses that are easy to confuse." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.211, + 0.785, + 0.437 + ], + "angle": 0, + "content": "After generating the instructions and responses, we prompt Gemini-Vision-Pro to generate rationales underlying the responses. Considering that positive and negative rationales have different emphases, we design two prompts and generate them separately. The positive rationales are expected to help the model learn the inherent knowledge and visual concepts underlying each problem-solving step, thereby facilitating fine-grained vision-language alignment and understanding of task definitions. It should include 1) analysis of visual content and instructions, 2) identification of core visual concepts relevant to the instructions, and 3) step-by-step reasoning grounded on the visual information and required knowledge. This is achieved by explicitly incorporating words such as \"explanation\", \"details\", and \"knowledge\" into the prompt to facilitate Gemini-Vision-Pro to include key image details and necessary outside knowledge in the rationale. The negative rationales are expected to provide discriminative information that aids in distinguishing between valid and hallucinatory outputs. Thus, we let Gemini-Vision-Pro highlight discriminative details in the negative rationale." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.441, + 0.787, + 0.517 + ], + "angle": 0, + "content": "Consider A-OK-VQA [39] and ScienceQA [31] already have some human annotated QA pairs and positive rationales, we include them into the REVERIE dataset to increase task diversity. Since both datasets provide questions in a multiple-choice format, we incorporate all incorrect answer choices and generate an additional negative rationale for each of them with the same pipeline." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.52, + 0.787, + 0.716 + ], + "angle": 0, + "content": "Consistency-based data filtering. While Gemini-Vision-Pro demonstrates impressive capabilities, it can still generate hallucinatory content when tasked to generate responses and rationales. Previous work [53] has shown that noise training data is one of the sources of the hallucinations. To guarantee the quality of generated data, we propose a method that leverages the consistency between positive and negative rationales to filter out noisy examples. Since both positive and negative rationales explicate the reasoning process behind the instruction but emphasize different aspects, they should contain information that is mutually consistent. When one rationale includes noisy information, this inherent consistency will be disrupted. Inspired by that, we employ a large language model, i.e., Gemini-Pro, to detect inconsistencies between these two types of rationales and only keep the samples that contain intra-consistent information (See supplementary materials for detailed prompt)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Data statistics. After the removal of noisy samples, the REVERIE dataset comprises 71,558 natural images. This includes 50,938 images sourced from Visual Genome [19], 15,706 from the COCO [27] and 4,914 images from ScienceQA [31]. REVERIE contains 115,280 instructions paired with corresponding positive responses and 138,897 negative responses, where each response is supplemented with a reflective rationale, rendering a total of 254,177 training instances. According to the output formats, REVERIE covers four types of vision-language tasks, including multiple-choice QA, short-answer QA, open-" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.371, + 0.129 + ], + "angle": 0, + "content": "J. Zhang et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.224 + ], + "angle": 0, + "content": "ended QA, and Yes/No questions, and their distributions are visualized in Fig. 3 (a). We also analyzed the distribution of rationale length and the frequency of noun occurrence. As shown in Fig. 3 (b) and (c), over \\(50\\%\\) of rationales exceed 25 words in length and contain more than 8 nouns, suggesting that rationales provide a substantial amount of information and knowledge." + }, + { + "type": "image", + "bbox": [ + 0.257, + 0.255, + 0.427, + 0.337 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.258, + 0.342, + 0.407, + 0.353 + ], + "angle": 0, + "content": "(a) Distribution of Different Task" + }, + { + "type": "image", + "bbox": [ + 0.421, + 0.251, + 0.583, + 0.34 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.421, + 0.342, + 0.583, + 0.353 + ], + "angle": 0, + "content": "(b) Distribution of Rationale Length" + }, + { + "type": "image", + "bbox": [ + 0.585, + 0.251, + 0.757, + 0.34 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.595, + 0.342, + 0.756, + 0.353 + ], + "angle": 0, + "content": "(c) Distribution of noun occurrence" + }, + { + "type": "image_caption", + "bbox": [ + 0.337, + 0.365, + 0.666, + 0.379 + ], + "angle": 0, + "content": "Fig. 3: Statistics of the REVERIE dataset." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.439, + 0.538, + 0.457 + ], + "angle": 0, + "content": "4 Reflective Instruction Tuning" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.476, + 0.788, + 0.656 + ], + "angle": 0, + "content": "Reflection plays a crucial role in the human learning process, allowing individuals to extract meaningful insights, learn from their mistakes, and make informed choices in the future. However, vanilla instruction tuning solely focuses on training models to predict responses, thus omitting access to the underlying rationale. This limitation impedes the model's ability to effectively internalize and logically process instruction-relevant information, increasing the likelihood of generating hallucinatory outputs. Inspired by humans' ability to improve problem-solving skills through reflection on explanations for both correct and incorrect answers [2,38], we propose reflective instruction tuning. This approach facilitates learning from the rationale behind responses, extending the training process to predict both positive and negative rationales rather than merely generating correct responses to instructions." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.658, + 0.788, + 0.779 + ], + "angle": 0, + "content": "Overall architecture. The objective of visual instruction tuning is to improve the instruction-following capability of the LVLM by leveraging diverse visual task data. Typically, LVLMs consist of a visual encoder, a vision-language connector, and an LLM. Given an image sample \\( I \\) and user instruction \\( X \\), the visual encoder extracts image token features, which are then projected into the language space by the modality connector. The projected visual tokens, along with the textual instructions, serve as the prefix in the LLM. An auto-regressive LLM is trained to generate response \\( A \\) that aligns with both the instructions and the image." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Multi-turn conversation. Directly training a model to concurrently predict rationales alongside responses in a single-turn prediction might hinder inference results. This is because some tasks do not necessitate explicit rationale generation, limiting its real-world applicability. To address this, we decouple rationale" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.536, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Reflective Instruction Tuning" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.359 + ], + "angle": 0, + "content": "learning from response prediction to improve the adaptability of reflective instruction tuning to diverse tasks. We achieve this by formulating the rationale data in a multi-turn conversational format. The first turn incorporates \\( I \\) and \\( X \\) as inputs, prompting the model to generate \\( A \\). Subsequently, the second turn introduces rationale generation prompts to let the model predict rationale. By putting the rationale learning behind, response prediction relies solely on the context of \\( I \\) and \\( X \\), mitigating interference to the inference results. To make the model understand this rationale generation task, we explicitly prompt the model to generate knowledge and facts that are related to instruction. Specifically, in positive rationale learning, the model receives the prompt \"Explain why. Including any necessary facts or knowledge\" following its response. In negative rationale learning, an incorrect answer is provided, followed by the prompt \"Explain why this answer is wrong: {incorrect answer}. Including any necessary facts or knowledge.\"" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.379, + 0.377, + 0.396 + ], + "angle": 0, + "content": "5 Experiments" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.407, + 0.79, + 0.543 + ], + "angle": 0, + "content": "We demonstrate the effectiveness of learning from rationales by applying reflective instruction tuning with the REVERIE dataset to two baseline LVLMs: (1) LLaVA-1.0-7b-lora [30]: We initialize the model pre-trained with LAION-CC-SBU-558k, and perform instruction tuning with the combination of REVERIE and LLaVA-Instruct-80k. (2) LLaVA-1.5-7b-lora [29]: Similar to LLaVA-1.0, we perform instruction tuning with the combination of REVERIE and the original 665k instructions. The resulting models are denoted as REVERIE-1.0-7b-lora and REVERIE-1.5-7b-lora, respectively. Implementation details are provided in the supplementary materials." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.562, + 0.377, + 0.576 + ], + "angle": 0, + "content": "5.1 Main Results" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.584, + 0.789, + 0.841 + ], + "angle": 0, + "content": "Performance comparison with baselines on six benchmarks. We compare the performance of the model reflective tuned with the proposed REVERIE dataset. As shown in Table 2, training with REVERIE yields clear performance gain on most of the benchmarks across 2 baselines. Notably, we achieve a score of 83.8 on the POPE benchmark, reflecting a significant improvement of 12.7 points over the baseline LLaVA-1.0-7b-lora. This suggests that the REVERIE dataset provides more fine-grained supervision, enabling the model to make more accurate judgments regarding object existence and significantly alleviating hallucination issues. Additionally, we observe substantial performance gains on two comprehensive LVLM evaluation benchmarks, MME and MMBench. Specifically, REVERIE-1.0-7b-lora achieves improvements of 321.2 and 27.2 points on MME and MMBench, respectively, compared to the baseline. This suggests that 1) the REVERIE dataset encompasses a wider range of visual concepts and knowledge compared to the original LLaVA-Instruct-80k dataset, demonstrably benefiting performance on diverse visual perception tasks evaluated in MME; 2) incorporating rationales during training empowers the model to develop a deeper understanding of task definitions and effectively distinguish positive answers from" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.371, + 0.129 + ], + "angle": 0, + "content": "J. Zhang et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.788, + 0.174 + ], + "angle": 0, + "content": "Table 2: Comparison with baselines on six benchmarks. * means training images of the dataset are used during training. † means rerun using their released code." + }, + { + "type": "table", + "bbox": [ + 0.238, + 0.185, + 0.761, + 0.346 + ], + "angle": 0, + "content": "
MethodScienceQA1POPEMMEMMBenchMM-VetGQA
InstructBLIP (vicuna-7b)60.5--3626.249.2
Qwen-VL-7b67.1--38.2-59.3
Qwen-VL-Chat-7b68.2-1487.560.6-57.5*
LLaVA-1.5-7b66.885.91510.764.330.562.0*
LLaVA-1.5-7b-lora68.486.41476.966.130.263.0*
LLaVA-1.5-7b-lora†67.786.11471.065.132.562.8*
LLaVA-1.0-7b-lora†42.771.1819.827.230.07.10
REVERIE-1.0-7b-lora70.1*83.81168.155.427.836.5
LLaVA-1.5-7b-lora (w/ ScienceQA)†76.3*86.61439.267.931.160.7*
REVERIE-1.5-7b-lora80.5*86.41474.967.330.861.8*
" + }, + { + "type": "table_caption", + "bbox": [ + 0.289, + 0.361, + 0.713, + 0.375 + ], + "angle": 0, + "content": "Table 3: Results of using different rationales on POPE." + }, + { + "type": "table", + "bbox": [ + 0.246, + 0.387, + 0.754, + 0.597 + ], + "angle": 0, + "content": "
SubsetModelAccuracyPrecisionRecall|F1 Score
RandomResponse only86.8789.4284.53
w/ pos rationales85.6087.9383.53
w/ neg rationales86.2587.9884.93
w/ pos + neg rationales87.1590.2184.20
PopularResponse only82.5781.3384.53
w/ pos rationales81.8080.7383.53
w/ neg rationales83.3082.2584.93
w/ pos + neg rationales84.0083.8684.20
AdversarialResponse only76.8373.2584.53
w/ pos rationales77.2374.1983.53
w/ neg rationales77.7374.2484.93
w/ pos + neg rationales79.3776.7884.20
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.628, + 0.788, + 0.809 + ], + "angle": 0, + "content": "provided negative answers, thereby achieving superior performance on multiple-choice question benchmarks like MMBench. Notably, for a fair comparison, we trained the baseline, LLaVA-1.5-7b-lora, using the combination of its original instruction tuning data and the VQA data from ScienceQA (only data with image context). We observed that simply introducing the ScienceQA QA data led to a performance decline on MME and GQA. This suggests that learning from response data alone may establish unintended shortcuts between instructions and responses, particularly for the complex reasoning instructions in ScienceQA. Our REVERIE-1.5-7b-lora demonstrates significant performance gains over its baseline on ScienceQA, MME, and GQA. Meanwhile, it achieves competitive performance on other benchmarks. This suggests that the rationale from our dataset may help alleviate the shortcut issue." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Effectiveness of positive/negative rationale. We conduct experiments to analyze the role of different rationales used in reflective instruction tuning. Ta-" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.536, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Reflective Instruction Tuning" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "table_caption", + "bbox": [ + 0.217, + 0.145, + 0.785, + 0.201 + ], + "angle": 0, + "content": "Table 4: Results of using different rationales on MMBench. We report L2-ability performance, where AR, CP, FP-S, FP-C, LR, and RR denote attribute reasoning, coarse perception, fine-grained perception (single-instance), fine-grained perception (cross-instance), logic reasoning, and relation reasoning, respectively." + }, + { + "type": "table", + "bbox": [ + 0.237, + 0.213, + 0.761, + 0.298 + ], + "angle": 0, + "content": "
ModelARCPFP-SFP-CLRRRAccuracy
Response only61.6960.0950.5142.7626.6742.6151.70
w/ pos rationales63.1867.1144.1449.1624.1740.8752.13
w/ neg rationales64.6864.4350.8444.8322.5043.4852.30
w/ pos + neg rationales65.1468.4657.2451.5219.1750.4355.44
" + }, + { + "type": "table_caption", + "bbox": [ + 0.217, + 0.311, + 0.784, + 0.34 + ], + "angle": 0, + "content": "Table 5: Results of using different rationales on GQA. We report the semantic-level accuracy to analyze the impact of rationale on questions with different subjects." + }, + { + "type": "table", + "bbox": [ + 0.238, + 0.352, + 0.761, + 0.438 + ], + "angle": 0, + "content": "
ModelAttributeObjectRelationGlobal CategoryAccuracy
Response only47.1759.9022.2131.2119.67
w/ pos rationales46.7857.4623.4429.3022.89
w/ neg rationales48.0156.6824.7731.8528.63
w/ pos + neg rationales47.3258.3524.7631.8528.11
36.53
" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.464, + 0.785, + 0.736 + ], + "angle": 0, + "content": "ble 3, Table 4 and Table 5 show the results of using different rationales on POPE, MMBench, and GQA, respectively. The results show that using both positive and negative rationale yields the best performance on most metrics. The results on POPE show that using both two rationales brings the highest F1 score, suggesting that training with rationale may alleviate the instruction-response shortcut and further alleviate the object hallucination issue. The results on MMBench show that training exclusively with positive rationales yields higher perception scores (in terms of CP and FP-C). This indicates that positive rationales enhance the capture of visual details. Conversely, training solely with negative rationales results in superior reasoning performance (in terms of AR and RR), suggesting that negative rationales provide critical clues that improve discriminative ability. We also infer that training with negative rationales could help the model better distinguish the correct answer from the negative response, benefiting more to the task with candidate answers provided, e.g., multiple choice. Our results on GQA further show that both positive and negative rationales contribute to performance gains in questions with category as their subject. This suggests that rationales provide more fine-grained information that enables the model to better understand the defining characteristics of a category." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.757, + 0.4, + 0.77 + ], + "angle": 0, + "content": "5.2 Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.78, + 0.784, + 0.839 + ], + "angle": 0, + "content": "We investigate several key designs of reflective instruction tuning and REVERIE: 1) the design of rationale generation prompt; 2) the effectiveness of consistency-based data filtering; 3) the design of conversational context; All experiments are conducted with REVERIE-1.0-7b-lora." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.371, + 0.128 + ], + "angle": 0, + "content": "J. Zhang et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.254, + 0.145, + 0.748, + 0.16 + ], + "angle": 0, + "content": "Table 6: Performance comparison of different rationale prompts." + }, + { + "type": "table", + "bbox": [ + 0.232, + 0.172, + 0.769, + 0.259 + ], + "angle": 0, + "content": "
Explicit GuidanceType-SpecificLength ControlMMBPOPEGQAMME
(a)55.483.836.81128.6
(b)✓✓55.283.236.01111.9
(c)✓✓✓54.383.835.81141.0
(d)✓55.483.836.51168.1
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.289, + 0.787, + 0.515 + ], + "angle": 0, + "content": "Rationale generation prompt. We conducted ablation studies to analyze the design of the rationale generation prompt employed in reflective instruction tuning. Specifically, we test the variants of rationale generation prompt from three perspectives: (1) the use of explicit guidance, e.g., \"Including any necessary facts or knowledge\"; 2) the utilization of separate prompts for positive and negative rationales; 3) the use of length-specific prompts. Table 6 summarizes the performance of these prompt variants. Our final choice, represented by Line (d), includes only explicit guidance and yields the best overall performance across four benchmarks. Line (a) demonstrates that using the prompt without explicit guidance leads to a significant performance drop on the MME benchmark. This suggests that providing clear guidance for clarifying the rationale generation task is crucial for aligning rationales with visual content. Moreover, Lines (c) and (d) indicate that overly complex prompts containing excessive constraints may impede the model's understanding of the rationale generation task, adversely affecting performance across most benchmarks." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.516, + 0.787, + 0.623 + ], + "angle": 0, + "content": "Consistency-based data filtering. We conducted experiments on the noise-infused rationale data to investigate the effectiveness of the proposed consistency-based data filtering strategy. Specifically, we train the model with the rationale data before filtering. As shown in Table 7, training with noised data that are not intra-consistent resulted in performance degradation across most benchmarks, suggesting that using the consistency-based filter strategy to maintain data quality is necessary for learning from rationale." + }, + { + "type": "table_caption", + "bbox": [ + 0.225, + 0.655, + 0.502, + 0.684 + ], + "angle": 0, + "content": "Table 7: Effect of rationale data filtering on model performance." + }, + { + "type": "table", + "bbox": [ + 0.242, + 0.696, + 0.482, + 0.745 + ], + "angle": 0, + "content": "
ModelMMBPOPEGQAMME
w/o filtering55.283.536.31132.3
w/ filtering55.483.836.51168.1
" + }, + { + "type": "table_caption", + "bbox": [ + 0.509, + 0.651, + 0.788, + 0.679 + ], + "angle": 0, + "content": "Table 8: Performance comparison of different conversational contexts." + }, + { + "type": "table", + "bbox": [ + 0.512, + 0.692, + 0.782, + 0.75 + ], + "angle": 0, + "content": "
ModelMMBPOPEGQAMME
pos rationale first49.379.536.21144.7
neg rationale first48.682.436.41160.2
separate context55.483.836.51168.1
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.787, + 0.842 + ], + "angle": 0, + "content": "Conversational context of reflective instruction tuning. We conducted ablation studies on several training data format variants, each incorporating different conversational contexts. We test the performance of three variants: 1) positive rationale first, where all negative rationales are learned to predict after" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.536, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Reflective Instruction Tuning" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.344 + ], + "angle": 0, + "content": "positive rationale turns; 2) negative rationale first, the inverse order of the previous one; 3) separate context, where the positive and negative rationales are separate into two distinct samples. The previous two variants put the positive and negative rationales for one instruction into a single conversation, making the model see previous rationales when learning to predict the current one. Table 8 shows the performance of these three variants. Our final choice that using separate contexts consistently outperforms the other two variants across four benchmarks. This suggests that providing the model with access to additional rationales during the current rationale generation step might lead to information leakage. In such a scenario, the model may tend to leverage pre-existing information from other rationales rather than learning to reason according to visual information and acquired knowledge. This hinders the model benefited from the fine-grained supervision from rationales." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.367, + 0.43, + 0.383 + ], + "angle": 0, + "content": "5.3 Qualitative Analysis" + }, + { + "type": "image", + "bbox": [ + 0.264, + 0.422, + 0.74, + 0.675 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.272, + 0.689, + 0.729, + 0.704 + ], + "angle": 0, + "content": "Fig. 4: Visualization of the generation of positive rationales." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.788, + 0.841 + ], + "angle": 0, + "content": "We conducted a case analysis to assess whether reflective instruction tuning improves fine-grained reasoning and mitigates hallucination. We compared baseline models with our proposed approach by prompting both to generate explicit rationales for their responses, aiming to verify if the models' logic aligns with correct reasoning. Fig. 4 visualizes the output rationales of different models. LLaVA directly generates rationale along with the response. However, the response is incorrect, and the rationale contains hallucinated content (highlighted in red)." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.371, + 0.129 + ], + "angle": 0, + "content": "J. Zhang et al." + }, + { + "type": "image", + "bbox": [ + 0.263, + 0.148, + 0.74, + 0.404 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.27, + 0.418, + 0.733, + 0.433 + ], + "angle": 0, + "content": "Fig. 5: Visualization of the generation of negative rationales." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.459, + 0.788, + 0.643 + ], + "angle": 0, + "content": "REVERIE (w/o R) (i.e., the model trained without rationales) and REVERIE both predicted the correct answer. However, REVERIE (w/o R) generates a rationale inconsistent with the response, suggesting that training without fine-grained supervision may lead to the shortcut between instruction and response rather than solving the problem from the correct reasoning paths. In contrast, the rationale generated by REVERIE presents more information that is necessary (highlighted in green) for addressing the instruction. We further evaluate the models' capability to generate negative rationales. As shown in Fig. 5, REVERIE correctly identifies crucial information (e.g., all presented colors) from images, while the other two models are misled by the context of two sub-images. This suggests that incorporating rationale as supervision enhances the model's fine-grained understanding of visual context, reducing hallucination." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.661, + 0.36, + 0.677 + ], + "angle": 0, + "content": "6 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.789, + 0.841 + ], + "angle": 0, + "content": "In this work, we identify a key contributor to hallucinations in existing LVLMs: the lack of fine-grained reasoning supervision during visual instruction tuning. To address this, we introduce reflective instruction tuning, which draws inspiration from human reflective learning processes by training the model on the rationales of positive and negative answers. Additionally, we propose REVERIE, the first large-scale visual instruction tuning dataset featuring comprehensive annotations for both positive and negative rationales. Experiments demonstrate that reflective instruction tuning with REVERIE significantly enhances LVLM performance, emphasizing the importance of fine-grained reasoning supervision in mitigating hallucinations." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.536, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Reflective Instruction Tuning" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.393, + 0.163 + ], + "angle": 0, + "content": "Acknowledgement" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.179, + 0.785, + 0.21 + ], + "angle": 0, + "content": "This work is supported by the National Natural Science Foundation of China (Grant NO.62122035)." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.234, + 0.323, + 0.25 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.267, + 0.786, + 0.309 + ], + "angle": 0, + "content": "1. Bai, J., Bai, S., Yang, S., Wang, S., Tan, S., Wang, P., Lin, J., Zhou, C., Zhou, J.: Qwen-vl: A frontier large vision-language model with versatile abilities. arXiv preprint arXiv:2308.12966 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.311, + 0.786, + 0.338 + ], + "angle": 0, + "content": "2. Boud, D., Keogh, R., Walker, D.: Reflection: Turning experience into learning. Routledge (2013)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.34, + 0.786, + 0.38 + ], + "angle": 0, + "content": "3. Cha, J., Kang, W., Mun, J., Roh, B.: Honeybee: Locality-enhanced projector for multimodal llm. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13817-13827 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.381, + 0.786, + 0.407 + ], + "angle": 0, + "content": "4. Chen, D., Liu, J., Dai, W., Wang, B.: Visual instruction tuning with polite flamingo. arXiv preprint arXiv:2307.01003 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.409, + 0.786, + 0.449 + ], + "angle": 0, + "content": "5. Chen, K., Zhang, Z., Zeng, W., Zhang, R., Zhu, F., Zhao, R.: Shikra: Unleashing multimodal llm's referential dialogue magic. arXiv preprint arXiv:2306.15195 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.451, + 0.786, + 0.49 + ], + "angle": 0, + "content": "6. Chen, S., Jie, Z., Ma, L.: Llava-mole: Sparse mixture of lora experts for mitigating data conflicts in instruction finetuning mllms. arXiv preprint arXiv:2401.16160 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.492, + 0.786, + 0.533 + ], + "angle": 0, + "content": "7. Chen, Z., Wu, J., Wang, W., Su, W., Chen, G., Xing, S., Muyan, Z., Zhang, Q., Zhu, X., Lu, L., et al.: Internvl: Scaling up vision foundation models and aligning for generic visual-linguistic tasks. arXiv preprint arXiv:2312.14238 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.535, + 0.786, + 0.588 + ], + "angle": 0, + "content": "8. Chowdhery, A., Narang, S., Devlin, J., Bosma, M., Mishra, G., Roberts, A., Barham, P., Chung, H.W., Sutton, C., Gehrmann, S., et al.: Palm: Scaling language modeling with pathways. Journal of Machine Learning Research 24(240), 1-113 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.59, + 0.786, + 0.63 + ], + "angle": 0, + "content": "9. Chung, H.W., Hou, L., Longpre, S., Zoph, B., Tay, Y., Fedus, W., Li, Y., Wang, X., Dehghani, M., Brahma, S., et al.: Scaling instruction-finetuned language models. arXiv preprint arXiv:2210.11416 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.632, + 0.786, + 0.672 + ], + "angle": 0, + "content": "0. Dai, W., Li, J., Li, D., Tiong, A.M.H., Zhao, J., Wang, W., Li, B., Fung, P., Hoi, S.: Instructblip: Towards general-purpose vision-language models with instruction tuning (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.674, + 0.786, + 0.701 + ], + "angle": 0, + "content": "1. Ge, J., Luo, H., Qian, S., Gan, Y., Fu, J., Zhan, S.: Chain of thought prompt tuning in vision language models. arXiv preprint arXiv:2304.07919 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.702, + 0.786, + 0.742 + ], + "angle": 0, + "content": "2. Guan, T., Liu, F., Li, X.W.R.X.Z., Wang, X.L.X., Yacoob, L.C.F.H.Y., Zhou, D.M.T.: Hallusionbench: An advanced diagnostic suite for entangled language hallucination & visual illusion in large vision-language models" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.744, + 0.786, + 0.771 + ], + "angle": 0, + "content": "3. Gunjal, A., Yin, J., Bas, E.: Detecting and preventing hallucinations in large vision language models. arXiv preprint arXiv:2308.06394 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.772, + 0.786, + 0.813 + ], + "angle": 0, + "content": "4. Han, Y., Zhang, C., Chen, X., Yang, X., Wang, Z., Yu, G., Fu, B., Zhang, H.: Chartllama: A multimodal llm for chart understanding and generation. arXiv preprint arXiv:2311.16483 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.814, + 0.786, + 0.84 + ], + "angle": 0, + "content": "5. Hu, H., Zhang, J., Zhao, M., Sun, Z.: Ciem: Contrastive instruction evaluation method for better instruction tuning. arXiv preprint arXiv:2309.02301 (2023)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.267, + 0.786, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.371, + 0.128 + ], + "angle": 0, + "content": "J. Zhang et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.785, + 0.188 + ], + "angle": 0, + "content": "16. Huang, Y., Meng, Z., Liu, F., Su, Y., Collier, N., Lu, Y.: Sparkles: Unlocking chats across multiple images for multimodal instruction-following models. arXiv preprint arXiv:2308.16463 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.189, + 0.785, + 0.229 + ], + "angle": 0, + "content": "17. Jain, J., Yang, J., Shi, H.: Vcoder: Versatile vision encoders for multimodal large language models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 27992-28002 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.23, + 0.785, + 0.27 + ], + "angle": 0, + "content": "18. Jiang, C., Xu, H., Dong, M., Chen, J., Ye, W., Yan, M., Ye, Q., Zhang, J., Huang, F., Zhang, S.: Hallucination augmented contrastive learning for multimodal large language model. arXiv preprint arXiv:2312.06968 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.271, + 0.785, + 0.325 + ], + "angle": 0, + "content": "19. Krishna, R., Zhu, Y., Groth, O., Johnson, J., Hata, K., Kravitz, J., Chen, S., Kalantidis, Y., Li, L.J., Shamma, D.A., et al.: Visual genome: Connecting language and vision using crowdsourced dense image annotations. International journal of computer vision 123, 32-73 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.326, + 0.785, + 0.366 + ], + "angle": 0, + "content": "20. Lee, B.K., Kim, C.W., Park, B., Ro, Y.M.: Meteor: Mamba-based traversal of rationale for large language and vision models. arXiv preprint arXiv:2405.15574 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.367, + 0.785, + 0.407 + ], + "angle": 0, + "content": "21. Leng, S., Zhang, H., Chen, G., Li, X., Lu, S., Miao, C., Bing, L.: Mitigating object hallucinations in large vision-language models through visual contrastive decoding. arXiv preprint arXiv:2311.16922 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.407, + 0.785, + 0.461 + ], + "angle": 0, + "content": "22. Leng, S., Zhang, H., Chen, G., Li, X., Lu, S., Miao, C., Bing, L.: Mitigating object hallucinations in large vision-language models through visual contrastive decoding. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13872-13882 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.461, + 0.785, + 0.501 + ], + "angle": 0, + "content": "23. Li, J., Li, D., Savarese, S., Hoi, S.: Blip-2: Bootstrapping language-image pretraining with frozen image encoders and large language models. arXiv preprint arXiv:2301.12597 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.502, + 0.785, + 0.529 + ], + "angle": 0, + "content": "24. Li, Y., Du, Y., Zhou, K., Wang, J., Zhao, W.X., Wen, J.R.: Evaluating object hallucination in large vision-language models. arXiv preprint arXiv:2305.10355 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.53, + 0.785, + 0.569 + ], + "angle": 0, + "content": "25. Lightman, H., Kosaraju, V., Burda, Y., Edwards, H., Baker, B., Lee, T., Leike, J., Schulman, J., Sutskever, I., Cobbe, K.: Let's verify step by step. arXiv preprint arXiv:2305.20050 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.569, + 0.785, + 0.609 + ], + "angle": 0, + "content": "26. Lin, B., Tang, Z., Ye, Y., Cui, J., Zhu, B., Jin, P., Zhang, J., Ning, M., Yuan, L.: Moe-llava: Mixture of experts for large vision-language models. arXiv preprint arXiv:2401.15947 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.61, + 0.785, + 0.664 + ], + "angle": 0, + "content": "27. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dólar, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13. pp. 740-755. Springer (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.665, + 0.785, + 0.705 + ], + "angle": 0, + "content": "28. Liu, F., Lin, K., Li, L., Wang, J., Yacoob, Y., Wang, L.: Mitigating hallucination in large multi-modal models via robust instruction tuning. In: The Twelfth International Conference on Learning Representations (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.706, + 0.785, + 0.733 + ], + "angle": 0, + "content": "29. Liu, H., Li, C., Li, Y., Lee, Y.J.: Improved baselines with visual instruction tuning. arXiv preprint arXiv:2310.03744 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.733, + 0.785, + 0.759 + ], + "angle": 0, + "content": "30. Liu, H., Li, C., Wu, Q., Lee, Y.J.: Visual instruction tuning. Advances in neural information processing systems 36 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.759, + 0.785, + 0.813 + ], + "angle": 0, + "content": "31. Lu, P., Mishra, S., Xia, T., Qiu, L., Chang, K.W., Zhu, S.C., Tafjord, O., Clark, P., Kalyan, A.: Learn to explain: Multimodal reasoning via thought chains for science question answering. Advances in Neural Information Processing Systems 35, 2507-2521 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.814, + 0.785, + 0.84 + ], + "angle": 0, + "content": "32. Magister, L.C., Mallinson, J., Adamek, J., Malmi, E., Severyn, A.: Teaching small language models to reason. arXiv preprint arXiv:2212.08410 (2022)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.536, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Reflective Instruction Tuning" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.203 + ], + "angle": 0, + "content": "33. Ouyang, L., Wu, J., Jiang, X., Almeida, D., Wainwright, C., Mishkin, P., Zhang, C., Agarwal, S., Slama, K., Ray, A., et al.: Training language models to follow instructions with human feedback. Advances in Neural Information Processing Systems 35, 27730-27744 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.203, + 0.787, + 0.257 + ], + "angle": 0, + "content": "34. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.257, + 0.787, + 0.299 + ], + "angle": 0, + "content": "35. Raffel, C., Shazeer, N., Roberts, A., Lee, K., Narang, S., Matena, M., Zhou, Y., Li, W., Liu, P.J.: Exploring the limits of transfer learning with a unified text-to-text transformer. The Journal of Machine Learning Research 21(1), 5485-5551 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.299, + 0.787, + 0.326 + ], + "angle": 0, + "content": "36. Rasheed, H., Maaz, M., Khan, S., Khan, F.S.: Llava++: Extending visual capabilities with llama-3 and phi-3 (2024), https://github.com/mbzuaai-oryx/LLaVA-pp" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.326, + 0.787, + 0.379 + ], + "angle": 0, + "content": "37. Schuhmann, C., Beaumont, R., Vencu, R., Gordon, C., Wightman, R., Cherti, M., Coombes, T., Katta, A., Mullis, C., Wortsman, M., et al.: Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems 35, 25278-25294 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.379, + 0.787, + 0.406 + ], + "angle": 0, + "content": "38. Schunk, D.H., Zimmerman, B.J.: Self-regulated learning: From teaching to self-reflective practice. Guilford Press (1998)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.406, + 0.787, + 0.447 + ], + "angle": 0, + "content": "39. Schwenk, D., Khandelwal, A., Clark, C., Marino, K., Mottaghi, R.: A-okvqa: A benchmark for visual question answering using world knowledge. In: European Conference on Computer Vision. pp. 146-162. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.447, + 0.787, + 0.488 + ], + "angle": 0, + "content": "40. Sun, Z., Shen, S., Cao, S., Liu, H., Li, C., Shen, Y., Gan, C., Gui, L.Y., Wang, Y.X., Yang, Y., et al.: Aligning large multimodal models with factually augmented rlhf. arXiv preprint arXiv:2309.14525 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.488, + 0.787, + 0.528 + ], + "angle": 0, + "content": "41. Team, G., Anil, R., Borgeaud, S., Wu, Y., Alayrac, J.B., Yu, J., Soricut, R., Schalkwyk, J., Dai, A.M., Hauth, A., et al.: Gemini: a family of highly capable multimodal models. arXiv preprint arXiv:2312.11805 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.528, + 0.787, + 0.57 + ], + "angle": 0, + "content": "42. Touvron, H., Lavril, T., Izacard, G., Martinet, X., Lachaux, M.A., Lacroix, T., Rozière, B., Goyal, N., Hambro, E., Azhar, F., Rodriguez, A., Joulin, A., Grave, E., Lample, G.: Llama: Open and efficient foundation language models (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.569, + 0.787, + 0.61 + ], + "angle": 0, + "content": "43. Touvron, H., Martin, L., Stone, K., Albert, P., Almahairi, A., Babaei, Y., Bashlykov, N., Batra, S., Bhargava, P., Bhosale, S., et al.: Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.61, + 0.787, + 0.651 + ], + "angle": 0, + "content": "44. Wang, J., Meng, L., Weng, Z., He, B., Wu, Z., Jiang, Y.G.: To see is to believe: Prompting gpt-4v for better visual instruction tuning. arXiv preprint arXiv:2311.07574 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.651, + 0.787, + 0.691 + ], + "angle": 0, + "content": "45. Wang, X., Pan, J., Ding, L., Biemann, C.: Mitigating hallucinations in large vision-language models with instruction contrastive decoding. arXiv preprint arXiv:2403.18715 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.691, + 0.787, + 0.732 + ], + "angle": 0, + "content": "46. Wang, X., Wei, J., Schuurmans, D., Le, Q., Chi, E., Narang, S., Chowdhery, A., Zhou, D.: Self-consistency improves chain of thought reasoning in language models. arXiv preprint arXiv:2203.11171 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.732, + 0.787, + 0.759 + ], + "angle": 0, + "content": "47. Wang, X., Wei, J., Schuurmans, D., Le, Q., Chi, E., Zhou, D.: Rationale-augmented ensembles in language models. arXiv preprint arXiv:2207.00747 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.759, + 0.787, + 0.8 + ], + "angle": 0, + "content": "48. Wei, J., Wang, X., Schuurmans, D., Bosma, M., Xia, F., Chi, E., Le, Q.V., Zhou, D., et al.: Chain-of-thought prompting elicits reasoning in large language models. Advances in Neural Information Processing Systems 35, 24824-24837 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.8, + 0.787, + 0.84 + ], + "angle": 0, + "content": "49. Wu, Y., Zhang, P., Xiong, W., Oguz, B., Gee, J.C., Nie, Y.: The role of chain-of-thought in complex vision-language reasoning task. arXiv preprint arXiv:2311.09193 (2023)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.371, + 0.129 + ], + "angle": 0, + "content": "J. Zhang et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.189 + ], + "angle": 0, + "content": "50. Ye, Q., Xu, H., Xu, G., Ye, J., Yan, M., Zhou, Y., Wang, J., Hu, A., Shi, P., Shi, Y., et al.: mplug-owl: Modularization empowers large language models with multimodality. arXiv preprint arXiv:2304.14178 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.19, + 0.785, + 0.232 + ], + "angle": 0, + "content": "51. Yin, S., Fu, C., Zhao, S., Xu, T., Wang, H., Sui, D., Shen, Y., Li, K., Sun, X., Chen, E.: Woodpecker: Hallucination correction for multimodal large language models. arXiv preprint arXiv:2310.16045 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.232, + 0.785, + 0.286 + ], + "angle": 0, + "content": "52. Yin, Z., Wang, J., Cao, J., Shi, Z., Liu, D., Li, M., Huang, X., Wang, Z., Sheng, L., Bai, L., et al.: Lamm: Language-assisted multi-modal instruction-tuning dataset, framework, and benchmark. Advances in Neural Information Processing Systems 36 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.287, + 0.785, + 0.314 + ], + "angle": 0, + "content": "53. Yu, Q., Li, J., Wei, L., Pang, L., Ye, W., Qin, B., Tang, S., Tian, Q., Zhuang, Y.: Hallucidoctor: Mitigating hallucinatory toxicity in visual instruction data (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.315, + 0.785, + 0.342 + ], + "angle": 0, + "content": "54. Zelikman, E., Mu, J., Goodman, N.D., Wu, Y.T.: Star: Self-taught reasoner bootstrapping reasoning with reasoning (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.343, + 0.785, + 0.396 + ], + "angle": 0, + "content": "55. Zhai, B., Yang, S., Zhao, X., Xu, C., Shen, S., Zhao, D., Keutzer, K., Li, M., Yan, T., Fan, X.: Halle-switch: Rethinking and controlling object existence hallucinations in large vision language models for detailed caption. arXiv preprint arXiv:2310.01779 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.398, + 0.785, + 0.466 + ], + "angle": 0, + "content": "56. Zhang, P., Dong, X., Wang, B., Cao, Y., Xu, C., Ouyang, L., Zhao, Z., Ding, S., Zhang, S., Duan, H., Zhang, W., Yan, H., Zhang, X., Li, W., Li, J., Chen, K., He, C., Zhang, X., Qiao, Y., Lin, D., Wang, J.: Internl m-xcomposer: A vision-language large model for advanced text-image comprehension and composition. arXiv preprint arXiv:2309.15112 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.467, + 0.785, + 0.508 + ], + "angle": 0, + "content": "57. Zhang, S., Roller, S., Goyal, N., Artetxe, M., Chen, M., Chen, S., Dewan, C., Diab, M., Li, X., Lin, X.V., et al.: Opt: Open pre-trained transformer language models. arXiv preprint arXiv:2205.01068 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.509, + 0.785, + 0.55 + ], + "angle": 0, + "content": "58. Zhang, Y., Zhang, R., Gu, J., Zhou, Y., Lipka, N., Yang, D., Sun, T.: Llavar: Enhanced visual instruction tuning for text-rich image understanding. arXiv preprint arXiv:2306.17107 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.55, + 0.785, + 0.578 + ], + "angle": 0, + "content": "59. Zhang, Z., Zhang, A., Li, M., Zhao, H., Karypis, G., Smola, A.: Multimodal chain-of-thought reasoning in language models. arXiv preprint arXiv:2302.00923 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.578, + 0.785, + 0.605 + ], + "angle": 0, + "content": "60. Zhao, B., Wu, B., Huang, T.: Svit: Scaling up visual instruction tuning. arXiv preprint arXiv:2307.04087 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.606, + 0.785, + 0.646 + ], + "angle": 0, + "content": "61. Zhao, L., Yu, E., Ge, Z., Yang, J., Wei, H., Zhou, H., Sun, J., Peng, Y., Dong, R., Han, C., et al.: Chatspot: Bootstrapping multimodal llms via precise referring instruction tuning. arXiv preprint arXiv:2307.09474 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.647, + 0.785, + 0.688 + ], + "angle": 0, + "content": "62. Zheng, L., Chiang, W.L., Sheng, Y., Zhuang, S., Wu, Z., Zhuang, Y., Lin, Z., Li, Z., Li, D., Xing, E.P., Zhang, H., Gonzalez, J.E., Stoica, I.: Judging lvm-as-a-judge with mt-bench and chatbot arena (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.689, + 0.785, + 0.73 + ], + "angle": 0, + "content": "63. Zhu, D., Chen, J., Shen, X., Li, X., Elhoseiny, M.: Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.73, + 0.785, + 0.771 + ], + "angle": 0, + "content": "64. Zong, Z., Ma, B., Shen, D., Song, G., Shao, H., Jiang, D., Li, H., Liu, Y.: Mova: Adapting mixture of vision experts to multimodal context. arXiv preprint arXiv:2404.13046 (2024)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.771 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/Reflective Instruction Tuning_ Mitigating Hallucinations in Large Vision-Language Models/8b065bde-2d53-44c1-a90a-d68b84da76bf_origin.pdf b/2024/Reflective Instruction Tuning_ Mitigating Hallucinations in Large Vision-Language Models/8b065bde-2d53-44c1-a90a-d68b84da76bf_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..f918fb546fd3d7e24fa3ff0029b757bdefbbf2be --- /dev/null +++ b/2024/Reflective Instruction Tuning_ Mitigating Hallucinations in Large Vision-Language Models/8b065bde-2d53-44c1-a90a-d68b84da76bf_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:19cc76d024cf1fa595d799550b04acd3c9026b307c662b007550dffff449f733 +size 968687 diff --git a/2024/Reflective Instruction Tuning_ Mitigating Hallucinations in Large Vision-Language Models/full.md b/2024/Reflective Instruction Tuning_ Mitigating Hallucinations in Large Vision-Language Models/full.md new file mode 100644 index 0000000000000000000000000000000000000000..080ba0cde2824348eba7cdd5a5b39ee112ca51d1 --- /dev/null +++ b/2024/Reflective Instruction Tuning_ Mitigating Hallucinations in Large Vision-Language Models/full.md @@ -0,0 +1,239 @@ +# Reflective Instruction Tuning: Mitigating Hallucinations in Large Vision-Language Models + +Jinrui Zhang $^{1}$ , Teng Wang $^{1,2}$ , Haigang Zhang $^{3}$ , Ping Lu $^{4}$ , Feng Zheng $^{1,5}$ + +$^{1}$ Southern University of Science and Technology + +2 The University of Hong Kong + +3 Shenzhen Polytechnic University + +4 The Cloud Computing and IT Institute of ZTE Corporation + +$^{5}$ Research Institute of Multiple Agents and Embodied Intelligence, Peng Cheng Laboratory, Shenzhen, China + +Abstract. Large vision-language models (LVLMs) have shown promising performance on a variety of vision-language tasks. However, they remain susceptible to hallucinations, generating outputs misaligned with visual content or instructions. While various mitigation strategies have been proposed, they often neglect a key contributor to hallucinations: lack of fine-grained reasoning supervision during training. Without intermediate reasoning steps, models may establish superficial shortcuts between instructions and responses, failing to internalize the inherent reasoning logic. To address this challenge, we propose reflective instruction tuning, which integrates rationale learning into visual instruction tuning. Unlike previous methods that learning from responses only, our approach entails the model predicting rationales justifying why responses are correct or incorrect. This fosters a deeper engagement with the fine-grained reasoning underlying each response, thus enhancing the model's reasoning proficiency. To facilitate this approach, we propose REVERIE, the first large-scale instruction-tuning dataset with ReflEctiVE Rationale annotations. REVERIE comprises 115k machine-generated reasoning instructions, each meticulously annotated with a corresponding pair of correct and confusing responses, alongside comprehensive rationales elucidating the justification behind the correctness or erroneousness of each response. Experimental results on multiple LVLM benchmarks reveal that reflective instruction tuning with the REVERIE dataset yields noticeable performance gain over the baseline model, demonstrating the effectiveness of reflecting from the rationales. Project page is at https://zjr2000.github.io/projects/reverie + +Keywords: Large Vision Language Models $\cdot$ Visual Instruction Tuning $\cdot$ Hallucination Mitigation + +# 1 Introduction + +Recently, large vision-language models (LVLMs) [1, 5, 10, 29, 30, 50, 58, 63] have garnered significant interest within the research community. By leveraging large + +![](images/2fa31c18577b9bef1435711bf43fb3f98dcb8835e0202f63c351b6ba8714afda.jpg) +Fig. 1: Difference between vanilla instruction tuning and the proposed reflective instruction tuning. Vanilla instruction tuning only trains LVLMs solely for response generation, lacking of supervising the learning of fine-grained reasoning details. Reflective instruction tuning additionally trains the model to reflect the rationale underlying the response, which provides more fine-grained supervision (e.g., the key visual evidence and facts to reach the response, highlighted in red), facilitating the model learning to capture more critical information. + +language models (LLMs) [8,9,33,42,43,62] as a unified task interface and employing visual instruction tuning [30] for enhanced task-level generalization, LVLMs have showcased remarkable reasoning performance and robustness in following user instructions. These breakthroughs offer substantial promise for the evolution of artificial general intelligence. Despite these advancements, LVLMs continue to be impeded by the phenomenon of hallucination [12, 24], wherein the models generate outputs that are inconsistent with visual content, established world knowledge, or the instructions provided. This significantly undermines the reliability of such models in real-world scenarios. Extensive research has investigated the origins of hallucinations in LVLMs, identifying factors such as noisy and biased training data [13, 15, 28], limited visual perception [1, 7], misalignment between vision and language modalities [18], and linguistic bias [12, 21]. To mitigate these hallucination sources, various mitigation methods have been proposed. These include: creating clean and balanced datasets [28, 55], scaling up visual resolution [1, 29] or utilizing additional perceptual models [51], employing alignment modules with increased parameters [7, 29], and designing decoding strategies that reduce over-reliance on language priors [21]. + +Despite significant progress, existing approaches often overlook a crucial factor contributing to hallucinations: the lack of fine-grained reasoning supervision. When learning to solve problems, humans often improve themselves by continuously reflecting on each step involved in the reasoning process. However, most previous LVLMs perform instruction tuning on datasets containing only instructions and direct responses, omitting access to the underlying rationale, i.e., intermediate reasoning steps. Specifically, Fig. 1 (left) shows the training paradigm of vanilla instruction tuning applied to a recognition task. The model is trained solely to predict the animal category name, without explicit supervision on the reasoning details from the discriminating information (e.g., the defining + +attributes of the animal) to the judgment. This strategy risks training the model to establish superficial shortcuts between instructions and responses, rather than understanding the tasks and internalizing the underlying logic of the reasoning process. Consequently, the model becomes susceptible to producing incorrect inferences and hallucinatory content misaligned with visual content. + +Inspired by the human ability to improve reasoning through reflection on intermediate steps [2,38], we propose reflective instruction tuning to address the aforementioned issue. Unlike vanilla instruction tuning, which only trains the model to predict the response, we incorporate the learning of underlying rationales, providing the model with fine-grained reasoning supervision. Specifically, reflective instruction tuning trains the model to generate positive rationales (justifying correct responses) and negative rationales (explaining errors in incorrect responses). The positive rationale explicitly provides more fine-grained visual content, the relationship between multiple visual concepts, and step-by-step reasoning. This facilitates the model to capture finer visual details and better understand the task definition underlying instruction. The negative rationale provides more illustration of the core characteristics linked to the erroneous categorization of negative answers, helping the model concentrate on crucial concepts within the decision-making process. This improves the model's ability to discriminate between correct and potential hallucinatory responses. Fig. 1 (right) presents an example of positive and negative rationale, the positive rationale provides a more detailed justification of the correct answer, and the negative rationale highlights errors in the incorrect answer, aiding discrimination. Reflective instruction tuning mirrors the human cognitive process of reflection, in which individuals analyze step-by-step reasoning and errors to gain a deeper understanding. Training the model with both positive and negative rationales enhances several core cognitive capabilities: it improves visual perception through more visual information, facilitates instruction understanding with coherent reasoning logic, and promotes error awareness by highlighting critical information. Consequently, this training paradigm mitigates hallucinatory outputs. + +While various visual instruction tuning datasets [4,28,30,44,52,60] exist, most of them consist solely of instruction-response pairs, omitting the annotation of rationales. To facilitate reflective instruction tuning that learns from both positive and negative rationales, we introduce REVERIE, the first large-scale visual instruction tuning dataset with ReflEctiVE rationalE annotations. REVERIE builds upon existing VQA datasets [31,39], and is further extended with natural images from Visual Genome [19]. The positive and negative rationales are annotated by Gemini family multimodal models [41]. REVERIE comprises 115k instructions encompassing four task types. Each instruction includes a positive response, an average of 1.2 negative responses, and accompanying rationales that justify the correctness or incorrectness of each response. This yields a total of 254k training instances, i.e., (instruction, response, rationale) tuples. + +To enhance the adaptability of reflective instruction tuning to diverse tasks, we formulate the rationale data as a conversation format and position rationale learning subsequent to the model's prediction of responses, thereby decoupling + +rationale generation from the context of response prediction. We perform reflective instruction tuning on two LVLMs with the proposed REVERIE dataset and further conduct experiments on multiple commonly used LVLM evaluation benchmarks. The results show that the proposed reflective tuning with REVERIE could yield substantial performance gain over the baselines. We also observed that utilizing the negative rationale could bring additional improvement than merely using the positive rationale. In summary, our contributions are three-fold: (a) We propose reflective instruction tuning, a novel approach that incorporates learning from both positive and negative rationales into visual instruction tuning to mitigate hallucination. (b) We propose REVERIE, the first large-scale instruction tuning dataset with both positive and negative fine-grained rationales. We hope this dataset could facilitate future research on rationale learning in LVLMs. (c) Extensive experiments show that reflective instruction tuning with the proposed dataset could bring substantial performance gain across multiple LVLM evaluation benchmarks, demonstrating the effectiveness of incorporating reflective rationales for mitigating hallucinations. + +# 2 Related Work + +Large vision-language models. The remarkable generalization capabilities of LLMs [8,9,33,35,42,43,57,62] have inspired the LVLMs [1,5,10,29,30,50,58,63]. LVLMs aim to extend the reasoning abilities of LLMs into the visual domain, potentially enabling more sophisticated and multimodal understanding. Typically, the architecture of previous LVLMs incorporates three fundamental components: a pre-trained visual encoder (e.g., CLIP [34]) for visual perception, a pre-trained LLM (e.g., LLaMA [42]) serving as a language understanding and generation interface, and a connector (e.g., Q-former [10, 23], linear projector [29, 30, 63]) facilitating vision-language alignment. A common training paradigm for these models involves vision-language pre-training using large-scale vision-text datasets [27, 30, 37, 63] and visual instruction tuning [30] with various visual task datasets [14, 29, 30, 58]. To enhance the generalization capabilities of models, previous methods have explored several techniques, such as: 1) employing stronger perception models [17, 64] or LLMs [36, 56]; 2) designing improved connectors [3, 7, 20]; 3) utilizing a mixture-of-expert design [6, 26]. Given the importance of visual instruction tuning in building general-purpose visual assistants, many studies have also focused on improving performance through the creation of high-quality instruction tuning datasets. These datasets target a wide range of visual understanding capabilities, encompassing text-rich image comprehension [58], region-level analysis [61], chart interpretation [14], and multi-image reasoning [16]. The proposed REVIRIE dataset also aims to improve the model performance during visual instruction tuning, but we focus on enhancing the reasoning capability by offering fine-grained reasoning supervision. + +Learning from reasoning steps. Several studies [11, 25, 31, 46-48] highlight the effectiveness of learning from detailed reasoning steps. Wei et al. [48] propose chain-of-thought prompting, where LLMs are explicitly prompted to think + +step-by-step during inference, leading to substantial gains on reasoning tasks. Furthermore, recent work suggests that training with explicit rationales can improve model reasoning capabilities [32, 54]. Additionally, learning from answer explanations has shown promise in enhancing performance on vision-language reasoning tasks [31, 49, 59]. Despite these advancements, learning from reasoning steps in visual instruction tuning remains under-explored due to the scarcity of large-scale rationale-annotated visual instruction tuning datasets. LLaVA-Instruct-158k [30] contains a portion of annotations for complex reasoning tasks, and the responses have step-by-step reasoning process annotation that is similar to the concept of rationale. However, certain limitations remain: 1) these rationales are all positive coupled with correct responses, neglecting negative rationales crucial for discerning incorrect responses, and 2) they are restricted to open-ended reasoning tasks, offering limited benefits to the diverse range of tasks. We address these gaps by proposing REVERIE, the first large-scale visual instruction tuning dataset that contains both positive and negative rationales on various visual tasks, facilitating the future research on rationale learning in visual instruction tuning. + +Hallucination in LVLMs. Previous research on hallucination in LVLMs has focused on two key areas: 1) the development of benchmarks for evaluating hallucinations [12, 24, 28, 40], and 2) the exploration of techniques to mitigate hallucinations [18, 22, 28, 45]. Li et al. [24] identify the object hallucination issue in LVLMs, and they propose the POPE benchmark for its evaluation. Guan et al. [12] divide the hallucination into language hallucination and visual hallucination and further introduce HALLUSIONBENCH to evaluate and diagnose each hallucination. Various studies address hallucination mitigation. Liu et al. [28] reduce hallucinations with a balanced visual instruction tuning dataset to better handle negative instructions. Jiang et al. [18] employ contrastive learning to distinguish hallucinated and non-hallucinated representations in latent space. Wang et al. [45] and Leng et al. [22] design contrastive decoding strategies in inference to mitigate the hallucination caused by language bias. Similar to [28], our work also focuses on mitigating hallucinations through improved visual instruction datasets. However, we emphasize fine-grained reasoning supervision to enhance consistency throughout the reasoning process. + +# 3 The REVERIE Dataset + +Despite previous works have provided many diverse visual instruction tuning datasets (e.g., LLaVA-Instruct-158k [30], LRV-Instruction [28]), only a small fraction of samples possess rationale annotations for positive responses. Furthermore, these datasets often omit annotations for hard negative responses and their underlying rationales, rendering them hard to provide fine-grained and discriminative supervision for addressing hallucinations. To facilitate model "reflect" from both positive and negative responses, we propose REVERIE, the first large-scale visual instruction tuning dataset with detailed reflective rationale annotation, serving as a basis of the proposed reflective instruction tuning strategy + +![](images/169c9c074bad62e02529e118d075dac872dd0d5d8b03d07a1d76ffd7c42a9c20.jpg) +Fig. 2: Overview of the REVERIE dataset's data collection pipeline. We first employ Gemini-Vision-Pro to annotate the instructions, responses, and rationales for each image. Gemini-Pro is then used to check the consistency between positive and negative rationales. Inconsistent samples are filtered to maintain dataset quality. + +(introduced in Sec. 4). As shown in Table 1, REVERIE provides hard negative response annotation and contains more extensive rationale annotations compared with the existing datasets. Fig. 2 shows the overall pipeline of our data collection strategy, we first annotate instructions, responses, and positive/negative rationales for each collected image, followed by filtering out noisy samples by the consistency-based filtering strategy. The details are as follows (The prompts for data generation are available in supplementary materials). + +Table 1: Comparison of the REVERIE with the data used by other LVLMs. + +
REVERIE LLaVA-1.5 InstructBLIP MiniGPT-4 LRV-Instruction
# Training Instances254k665k1.2M3.5k400k
Hard Negative Responses?✓XXXX
# Positive Rationales115k77k77kXX
# Negative Rationales138kXXXX
+ +Data collection and annotation. To ensure the dataset encompasses diverse visual concepts and broad domain coverage, we sample a subset of images from the Visual Genome dataset [19]. Subsequently, we annotate these images with instructions and corresponding responses. Given the high cost of manual labelling, we leverage a strong LVLM, Gemini-Vision-Pro, as our annotator. Considering that the rationale for addressing simple query instruction may contain limited intermediate steps and benefit less to learning fine-grained reasoning, we explicitly prompt Gemini to generate instructions that need multi-step to solve and avoid generating simple queries. Additionally, it is prompted to generate both + +positive and negative responses in conjunction with each instruction. Intuitively, a more confusing negative response demands more discriminative information to distinguish it from the positive answer. Thus, we explicitly prompt Gemini to generate negative responses that are easy to confuse. + +After generating the instructions and responses, we prompt Gemini-Vision-Pro to generate rationales underlying the responses. Considering that positive and negative rationales have different emphases, we design two prompts and generate them separately. The positive rationales are expected to help the model learn the inherent knowledge and visual concepts underlying each problem-solving step, thereby facilitating fine-grained vision-language alignment and understanding of task definitions. It should include 1) analysis of visual content and instructions, 2) identification of core visual concepts relevant to the instructions, and 3) step-by-step reasoning grounded on the visual information and required knowledge. This is achieved by explicitly incorporating words such as "explanation", "details", and "knowledge" into the prompt to facilitate Gemini-Vision-Pro to include key image details and necessary outside knowledge in the rationale. The negative rationales are expected to provide discriminative information that aids in distinguishing between valid and hallucinatory outputs. Thus, we let Gemini-Vision-Pro highlight discriminative details in the negative rationale. + +Consider A-OK-VQA [39] and ScienceQA [31] already have some human annotated QA pairs and positive rationales, we include them into the REVERIE dataset to increase task diversity. Since both datasets provide questions in a multiple-choice format, we incorporate all incorrect answer choices and generate an additional negative rationale for each of them with the same pipeline. + +Consistency-based data filtering. While Gemini-Vision-Pro demonstrates impressive capabilities, it can still generate hallucinatory content when tasked to generate responses and rationales. Previous work [53] has shown that noise training data is one of the sources of the hallucinations. To guarantee the quality of generated data, we propose a method that leverages the consistency between positive and negative rationales to filter out noisy examples. Since both positive and negative rationales explicate the reasoning process behind the instruction but emphasize different aspects, they should contain information that is mutually consistent. When one rationale includes noisy information, this inherent consistency will be disrupted. Inspired by that, we employ a large language model, i.e., Gemini-Pro, to detect inconsistencies between these two types of rationales and only keep the samples that contain intra-consistent information (See supplementary materials for detailed prompt). + +Data statistics. After the removal of noisy samples, the REVERIE dataset comprises 71,558 natural images. This includes 50,938 images sourced from Visual Genome [19], 15,706 from the COCO [27] and 4,914 images from ScienceQA [31]. REVERIE contains 115,280 instructions paired with corresponding positive responses and 138,897 negative responses, where each response is supplemented with a reflective rationale, rendering a total of 254,177 training instances. According to the output formats, REVERIE covers four types of vision-language tasks, including multiple-choice QA, short-answer QA, open- + +ended QA, and Yes/No questions, and their distributions are visualized in Fig. 3 (a). We also analyzed the distribution of rationale length and the frequency of noun occurrence. As shown in Fig. 3 (b) and (c), over $50\%$ of rationales exceed 25 words in length and contain more than 8 nouns, suggesting that rationales provide a substantial amount of information and knowledge. + +![](images/257896d69dfee4efb0526e08ad5b9e28d9e791df9bb68a3e216a6055cfeaf6fe.jpg) +(a) Distribution of Different Task + +![](images/df35c8037a93b05ed1023ccb1dab4d8539b5885612f079474304d88278292497.jpg) +(b) Distribution of Rationale Length + +![](images/59648b7be52428021db84a4a1cd804423db5192d912b395c8a4d7c8f407d1bb0.jpg) +(c) Distribution of noun occurrence +Fig. 3: Statistics of the REVERIE dataset. + +# 4 Reflective Instruction Tuning + +Reflection plays a crucial role in the human learning process, allowing individuals to extract meaningful insights, learn from their mistakes, and make informed choices in the future. However, vanilla instruction tuning solely focuses on training models to predict responses, thus omitting access to the underlying rationale. This limitation impedes the model's ability to effectively internalize and logically process instruction-relevant information, increasing the likelihood of generating hallucinatory outputs. Inspired by humans' ability to improve problem-solving skills through reflection on explanations for both correct and incorrect answers [2,38], we propose reflective instruction tuning. This approach facilitates learning from the rationale behind responses, extending the training process to predict both positive and negative rationales rather than merely generating correct responses to instructions. + +Overall architecture. The objective of visual instruction tuning is to improve the instruction-following capability of the LVLM by leveraging diverse visual task data. Typically, LVLMs consist of a visual encoder, a vision-language connector, and an LLM. Given an image sample $I$ and user instruction $X$ , the visual encoder extracts image token features, which are then projected into the language space by the modality connector. The projected visual tokens, along with the textual instructions, serve as the prefix in the LLM. An auto-regressive LLM is trained to generate response $A$ that aligns with both the instructions and the image. + +Multi-turn conversation. Directly training a model to concurrently predict rationales alongside responses in a single-turn prediction might hinder inference results. This is because some tasks do not necessitate explicit rationale generation, limiting its real-world applicability. To address this, we decouple rationale + +learning from response prediction to improve the adaptability of reflective instruction tuning to diverse tasks. We achieve this by formulating the rationale data in a multi-turn conversational format. The first turn incorporates $I$ and $X$ as inputs, prompting the model to generate $A$ . Subsequently, the second turn introduces rationale generation prompts to let the model predict rationale. By putting the rationale learning behind, response prediction relies solely on the context of $I$ and $X$ , mitigating interference to the inference results. To make the model understand this rationale generation task, we explicitly prompt the model to generate knowledge and facts that are related to instruction. Specifically, in positive rationale learning, the model receives the prompt "Explain why. Including any necessary facts or knowledge" following its response. In negative rationale learning, an incorrect answer is provided, followed by the prompt "Explain why this answer is wrong: {incorrect answer}. Including any necessary facts or knowledge." + +# 5 Experiments + +We demonstrate the effectiveness of learning from rationales by applying reflective instruction tuning with the REVERIE dataset to two baseline LVLMs: (1) LLaVA-1.0-7b-lora [30]: We initialize the model pre-trained with LAION-CC-SBU-558k, and perform instruction tuning with the combination of REVERIE and LLaVA-Instruct-80k. (2) LLaVA-1.5-7b-lora [29]: Similar to LLaVA-1.0, we perform instruction tuning with the combination of REVERIE and the original 665k instructions. The resulting models are denoted as REVERIE-1.0-7b-lora and REVERIE-1.5-7b-lora, respectively. Implementation details are provided in the supplementary materials. + +# 5.1 Main Results + +Performance comparison with baselines on six benchmarks. We compare the performance of the model reflective tuned with the proposed REVERIE dataset. As shown in Table 2, training with REVERIE yields clear performance gain on most of the benchmarks across 2 baselines. Notably, we achieve a score of 83.8 on the POPE benchmark, reflecting a significant improvement of 12.7 points over the baseline LLaVA-1.0-7b-lora. This suggests that the REVERIE dataset provides more fine-grained supervision, enabling the model to make more accurate judgments regarding object existence and significantly alleviating hallucination issues. Additionally, we observe substantial performance gains on two comprehensive LVLM evaluation benchmarks, MME and MMBench. Specifically, REVERIE-1.0-7b-lora achieves improvements of 321.2 and 27.2 points on MME and MMBench, respectively, compared to the baseline. This suggests that 1) the REVERIE dataset encompasses a wider range of visual concepts and knowledge compared to the original LLaVA-Instruct-80k dataset, demonstrably benefiting performance on diverse visual perception tasks evaluated in MME; 2) incorporating rationales during training empowers the model to develop a deeper understanding of task definitions and effectively distinguish positive answers from + +Table 2: Comparison with baselines on six benchmarks. * means training images of the dataset are used during training. † means rerun using their released code. + +
MethodScienceQA1POPEMMEMMBenchMM-VetGQA
InstructBLIP (vicuna-7b)60.5--3626.249.2
Qwen-VL-7b67.1--38.2-59.3
Qwen-VL-Chat-7b68.2-1487.560.6-57.5*
LLaVA-1.5-7b66.885.91510.764.330.562.0*
LLaVA-1.5-7b-lora68.486.41476.966.130.263.0*
LLaVA-1.5-7b-lora†67.786.11471.065.132.562.8*
LLaVA-1.0-7b-lora†42.771.1819.827.230.07.10
REVERIE-1.0-7b-lora70.1*83.81168.155.427.836.5
LLaVA-1.5-7b-lora (w/ ScienceQA)†76.3*86.61439.267.931.160.7*
REVERIE-1.5-7b-lora80.5*86.41474.967.330.861.8*
+ +Table 3: Results of using different rationales on POPE. + +
SubsetModelAccuracyPrecisionRecall|F1 Score
RandomResponse only86.8789.4284.53
w/ pos rationales85.6087.9383.53
w/ neg rationales86.2587.9884.93
w/ pos + neg rationales87.1590.2184.20
PopularResponse only82.5781.3384.53
w/ pos rationales81.8080.7383.53
w/ neg rationales83.3082.2584.93
w/ pos + neg rationales84.0083.8684.20
AdversarialResponse only76.8373.2584.53
w/ pos rationales77.2374.1983.53
w/ neg rationales77.7374.2484.93
w/ pos + neg rationales79.3776.7884.20
+ +provided negative answers, thereby achieving superior performance on multiple-choice question benchmarks like MMBench. Notably, for a fair comparison, we trained the baseline, LLaVA-1.5-7b-lora, using the combination of its original instruction tuning data and the VQA data from ScienceQA (only data with image context). We observed that simply introducing the ScienceQA QA data led to a performance decline on MME and GQA. This suggests that learning from response data alone may establish unintended shortcuts between instructions and responses, particularly for the complex reasoning instructions in ScienceQA. Our REVERIE-1.5-7b-lora demonstrates significant performance gains over its baseline on ScienceQA, MME, and GQA. Meanwhile, it achieves competitive performance on other benchmarks. This suggests that the rationale from our dataset may help alleviate the shortcut issue. + +Effectiveness of positive/negative rationale. We conduct experiments to analyze the role of different rationales used in reflective instruction tuning. Ta- + +Table 4: Results of using different rationales on MMBench. We report L2-ability performance, where AR, CP, FP-S, FP-C, LR, and RR denote attribute reasoning, coarse perception, fine-grained perception (single-instance), fine-grained perception (cross-instance), logic reasoning, and relation reasoning, respectively. + +
ModelARCPFP-SFP-CLRRRAccuracy
Response only61.6960.0950.5142.7626.6742.6151.70
w/ pos rationales63.1867.1144.1449.1624.1740.8752.13
w/ neg rationales64.6864.4350.8444.8322.5043.4852.30
w/ pos + neg rationales65.1468.4657.2451.5219.1750.4355.44
+ +Table 5: Results of using different rationales on GQA. We report the semantic-level accuracy to analyze the impact of rationale on questions with different subjects. + +
ModelAttributeObjectRelationGlobal CategoryAccuracy
Response only47.1759.9022.2131.2119.67
w/ pos rationales46.7857.4623.4429.3022.89
w/ neg rationales48.0156.6824.7731.8528.63
w/ pos + neg rationales47.3258.3524.7631.8528.11
36.53
+ +ble 3, Table 4 and Table 5 show the results of using different rationales on POPE, MMBench, and GQA, respectively. The results show that using both positive and negative rationale yields the best performance on most metrics. The results on POPE show that using both two rationales brings the highest F1 score, suggesting that training with rationale may alleviate the instruction-response shortcut and further alleviate the object hallucination issue. The results on MMBench show that training exclusively with positive rationales yields higher perception scores (in terms of CP and FP-C). This indicates that positive rationales enhance the capture of visual details. Conversely, training solely with negative rationales results in superior reasoning performance (in terms of AR and RR), suggesting that negative rationales provide critical clues that improve discriminative ability. We also infer that training with negative rationales could help the model better distinguish the correct answer from the negative response, benefiting more to the task with candidate answers provided, e.g., multiple choice. Our results on GQA further show that both positive and negative rationales contribute to performance gains in questions with category as their subject. This suggests that rationales provide more fine-grained information that enables the model to better understand the defining characteristics of a category. + +# 5.2 Ablation Studies + +We investigate several key designs of reflective instruction tuning and REVERIE: 1) the design of rationale generation prompt; 2) the effectiveness of consistency-based data filtering; 3) the design of conversational context; All experiments are conducted with REVERIE-1.0-7b-lora. + +Table 6: Performance comparison of different rationale prompts. + +
Explicit GuidanceType-SpecificLength ControlMMBPOPEGQAMME
(a)55.483.836.81128.6
(b)✓✓55.283.236.01111.9
(c)✓✓✓54.383.835.81141.0
(d)✓55.483.836.51168.1
+ +Rationale generation prompt. We conducted ablation studies to analyze the design of the rationale generation prompt employed in reflective instruction tuning. Specifically, we test the variants of rationale generation prompt from three perspectives: (1) the use of explicit guidance, e.g., "Including any necessary facts or knowledge"; 2) the utilization of separate prompts for positive and negative rationales; 3) the use of length-specific prompts. Table 6 summarizes the performance of these prompt variants. Our final choice, represented by Line (d), includes only explicit guidance and yields the best overall performance across four benchmarks. Line (a) demonstrates that using the prompt without explicit guidance leads to a significant performance drop on the MME benchmark. This suggests that providing clear guidance for clarifying the rationale generation task is crucial for aligning rationales with visual content. Moreover, Lines (c) and (d) indicate that overly complex prompts containing excessive constraints may impede the model's understanding of the rationale generation task, adversely affecting performance across most benchmarks. + +Consistency-based data filtering. We conducted experiments on the noise-infused rationale data to investigate the effectiveness of the proposed consistency-based data filtering strategy. Specifically, we train the model with the rationale data before filtering. As shown in Table 7, training with noised data that are not intra-consistent resulted in performance degradation across most benchmarks, suggesting that using the consistency-based filter strategy to maintain data quality is necessary for learning from rationale. + +Table 7: Effect of rationale data filtering on model performance. + +
ModelMMBPOPEGQAMME
w/o filtering55.283.536.31132.3
w/ filtering55.483.836.51168.1
+ +Table 8: Performance comparison of different conversational contexts. + +
ModelMMBPOPEGQAMME
pos rationale first49.379.536.21144.7
neg rationale first48.682.436.41160.2
separate context55.483.836.51168.1
+ +Conversational context of reflective instruction tuning. We conducted ablation studies on several training data format variants, each incorporating different conversational contexts. We test the performance of three variants: 1) positive rationale first, where all negative rationales are learned to predict after + +positive rationale turns; 2) negative rationale first, the inverse order of the previous one; 3) separate context, where the positive and negative rationales are separate into two distinct samples. The previous two variants put the positive and negative rationales for one instruction into a single conversation, making the model see previous rationales when learning to predict the current one. Table 8 shows the performance of these three variants. Our final choice that using separate contexts consistently outperforms the other two variants across four benchmarks. This suggests that providing the model with access to additional rationales during the current rationale generation step might lead to information leakage. In such a scenario, the model may tend to leverage pre-existing information from other rationales rather than learning to reason according to visual information and acquired knowledge. This hinders the model benefited from the fine-grained supervision from rationales. + +# 5.3 Qualitative Analysis + +![](images/0e96a6377ce4e56b9b578d60c423e9cfb9125439e30e377ca64b966f595e5ff5.jpg) +Fig. 4: Visualization of the generation of positive rationales. + +We conducted a case analysis to assess whether reflective instruction tuning improves fine-grained reasoning and mitigates hallucination. We compared baseline models with our proposed approach by prompting both to generate explicit rationales for their responses, aiming to verify if the models' logic aligns with correct reasoning. Fig. 4 visualizes the output rationales of different models. LLaVA directly generates rationale along with the response. However, the response is incorrect, and the rationale contains hallucinated content (highlighted in red). + +![](images/d3d8975ded3131c01a0efc384b10725b34d85ae65dca9acd64bd81ebacaff77b.jpg) +Fig. 5: Visualization of the generation of negative rationales. + +REVERIE (w/o R) (i.e., the model trained without rationales) and REVERIE both predicted the correct answer. However, REVERIE (w/o R) generates a rationale inconsistent with the response, suggesting that training without fine-grained supervision may lead to the shortcut between instruction and response rather than solving the problem from the correct reasoning paths. In contrast, the rationale generated by REVERIE presents more information that is necessary (highlighted in green) for addressing the instruction. We further evaluate the models' capability to generate negative rationales. As shown in Fig. 5, REVERIE correctly identifies crucial information (e.g., all presented colors) from images, while the other two models are misled by the context of two sub-images. This suggests that incorporating rationale as supervision enhances the model's fine-grained understanding of visual context, reducing hallucination. + +# 6 Conclusion + +In this work, we identify a key contributor to hallucinations in existing LVLMs: the lack of fine-grained reasoning supervision during visual instruction tuning. To address this, we introduce reflective instruction tuning, which draws inspiration from human reflective learning processes by training the model on the rationales of positive and negative answers. Additionally, we propose REVERIE, the first large-scale visual instruction tuning dataset featuring comprehensive annotations for both positive and negative rationales. Experiments demonstrate that reflective instruction tuning with REVERIE significantly enhances LVLM performance, emphasizing the importance of fine-grained reasoning supervision in mitigating hallucinations. + +# Acknowledgement + +This work is supported by the National Natural Science Foundation of China (Grant NO.62122035). + +# References + +1. Bai, J., Bai, S., Yang, S., Wang, S., Tan, S., Wang, P., Lin, J., Zhou, C., Zhou, J.: Qwen-vl: A frontier large vision-language model with versatile abilities. arXiv preprint arXiv:2308.12966 (2023) +2. Boud, D., Keogh, R., Walker, D.: Reflection: Turning experience into learning. Routledge (2013) +3. Cha, J., Kang, W., Mun, J., Roh, B.: Honeybee: Locality-enhanced projector for multimodal llm. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13817-13827 (2024) +4. Chen, D., Liu, J., Dai, W., Wang, B.: Visual instruction tuning with polite flamingo. arXiv preprint arXiv:2307.01003 (2023) +5. Chen, K., Zhang, Z., Zeng, W., Zhang, R., Zhu, F., Zhao, R.: Shikra: Unleashing multimodal llm's referential dialogue magic. arXiv preprint arXiv:2306.15195 (2023) +6. Chen, S., Jie, Z., Ma, L.: Llava-mole: Sparse mixture of lora experts for mitigating data conflicts in instruction finetuning mllms. arXiv preprint arXiv:2401.16160 (2024) +7. Chen, Z., Wu, J., Wang, W., Su, W., Chen, G., Xing, S., Muyan, Z., Zhang, Q., Zhu, X., Lu, L., et al.: Internvl: Scaling up vision foundation models and aligning for generic visual-linguistic tasks. arXiv preprint arXiv:2312.14238 (2023) +8. Chowdhery, A., Narang, S., Devlin, J., Bosma, M., Mishra, G., Roberts, A., Barham, P., Chung, H.W., Sutton, C., Gehrmann, S., et al.: Palm: Scaling language modeling with pathways. Journal of Machine Learning Research 24(240), 1-113 (2023) +9. Chung, H.W., Hou, L., Longpre, S., Zoph, B., Tay, Y., Fedus, W., Li, Y., Wang, X., Dehghani, M., Brahma, S., et al.: Scaling instruction-finetuned language models. arXiv preprint arXiv:2210.11416 (2022) +0. Dai, W., Li, J., Li, D., Tiong, A.M.H., Zhao, J., Wang, W., Li, B., Fung, P., Hoi, S.: Instructblip: Towards general-purpose vision-language models with instruction tuning (2023) +1. Ge, J., Luo, H., Qian, S., Gan, Y., Fu, J., Zhan, S.: Chain of thought prompt tuning in vision language models. arXiv preprint arXiv:2304.07919 (2023) +2. Guan, T., Liu, F., Li, X.W.R.X.Z., Wang, X.L.X., Yacoob, L.C.F.H.Y., Zhou, D.M.T.: Hallusionbench: An advanced diagnostic suite for entangled language hallucination & visual illusion in large vision-language models +3. Gunjal, A., Yin, J., Bas, E.: Detecting and preventing hallucinations in large vision language models. arXiv preprint arXiv:2308.06394 (2023) +4. Han, Y., Zhang, C., Chen, X., Yang, X., Wang, Z., Yu, G., Fu, B., Zhang, H.: Chartllama: A multimodal llm for chart understanding and generation. arXiv preprint arXiv:2311.16483 (2023) +5. Hu, H., Zhang, J., Zhao, M., Sun, Z.: Ciem: Contrastive instruction evaluation method for better instruction tuning. arXiv preprint arXiv:2309.02301 (2023) + +16. Huang, Y., Meng, Z., Liu, F., Su, Y., Collier, N., Lu, Y.: Sparkles: Unlocking chats across multiple images for multimodal instruction-following models. arXiv preprint arXiv:2308.16463 (2023) +17. Jain, J., Yang, J., Shi, H.: Vcoder: Versatile vision encoders for multimodal large language models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 27992-28002 (2024) +18. Jiang, C., Xu, H., Dong, M., Chen, J., Ye, W., Yan, M., Ye, Q., Zhang, J., Huang, F., Zhang, S.: Hallucination augmented contrastive learning for multimodal large language model. arXiv preprint arXiv:2312.06968 (2023) +19. Krishna, R., Zhu, Y., Groth, O., Johnson, J., Hata, K., Kravitz, J., Chen, S., Kalantidis, Y., Li, L.J., Shamma, D.A., et al.: Visual genome: Connecting language and vision using crowdsourced dense image annotations. International journal of computer vision 123, 32-73 (2017) +20. Lee, B.K., Kim, C.W., Park, B., Ro, Y.M.: Meteor: Mamba-based traversal of rationale for large language and vision models. arXiv preprint arXiv:2405.15574 (2024) +21. Leng, S., Zhang, H., Chen, G., Li, X., Lu, S., Miao, C., Bing, L.: Mitigating object hallucinations in large vision-language models through visual contrastive decoding. arXiv preprint arXiv:2311.16922 (2023) +22. Leng, S., Zhang, H., Chen, G., Li, X., Lu, S., Miao, C., Bing, L.: Mitigating object hallucinations in large vision-language models through visual contrastive decoding. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13872-13882 (2024) +23. Li, J., Li, D., Savarese, S., Hoi, S.: Blip-2: Bootstrapping language-image pretraining with frozen image encoders and large language models. arXiv preprint arXiv:2301.12597 (2023) +24. Li, Y., Du, Y., Zhou, K., Wang, J., Zhao, W.X., Wen, J.R.: Evaluating object hallucination in large vision-language models. arXiv preprint arXiv:2305.10355 (2023) +25. Lightman, H., Kosaraju, V., Burda, Y., Edwards, H., Baker, B., Lee, T., Leike, J., Schulman, J., Sutskever, I., Cobbe, K.: Let's verify step by step. arXiv preprint arXiv:2305.20050 (2023) +26. Lin, B., Tang, Z., Ye, Y., Cui, J., Zhu, B., Jin, P., Zhang, J., Ning, M., Yuan, L.: Moe-llava: Mixture of experts for large vision-language models. arXiv preprint arXiv:2401.15947 (2024) +27. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dólar, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13. pp. 740-755. Springer (2014) +28. Liu, F., Lin, K., Li, L., Wang, J., Yacoob, Y., Wang, L.: Mitigating hallucination in large multi-modal models via robust instruction tuning. In: The Twelfth International Conference on Learning Representations (2023) +29. Liu, H., Li, C., Li, Y., Lee, Y.J.: Improved baselines with visual instruction tuning. arXiv preprint arXiv:2310.03744 (2023) +30. Liu, H., Li, C., Wu, Q., Lee, Y.J.: Visual instruction tuning. Advances in neural information processing systems 36 (2024) +31. Lu, P., Mishra, S., Xia, T., Qiu, L., Chang, K.W., Zhu, S.C., Tafjord, O., Clark, P., Kalyan, A.: Learn to explain: Multimodal reasoning via thought chains for science question answering. Advances in Neural Information Processing Systems 35, 2507-2521 (2022) +32. Magister, L.C., Mallinson, J., Adamek, J., Malmi, E., Severyn, A.: Teaching small language models to reason. arXiv preprint arXiv:2212.08410 (2022) + +33. Ouyang, L., Wu, J., Jiang, X., Almeida, D., Wainwright, C., Mishkin, P., Zhang, C., Agarwal, S., Slama, K., Ray, A., et al.: Training language models to follow instructions with human feedback. Advances in Neural Information Processing Systems 35, 27730-27744 (2022) +34. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021) +35. Raffel, C., Shazeer, N., Roberts, A., Lee, K., Narang, S., Matena, M., Zhou, Y., Li, W., Liu, P.J.: Exploring the limits of transfer learning with a unified text-to-text transformer. The Journal of Machine Learning Research 21(1), 5485-5551 (2020) +36. Rasheed, H., Maaz, M., Khan, S., Khan, F.S.: Llava++: Extending visual capabilities with llama-3 and phi-3 (2024), https://github.com/mbzuaai-oryx/LLaVA-pp +37. Schuhmann, C., Beaumont, R., Vencu, R., Gordon, C., Wightman, R., Cherti, M., Coombes, T., Katta, A., Mullis, C., Wortsman, M., et al.: Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems 35, 25278-25294 (2022) +38. Schunk, D.H., Zimmerman, B.J.: Self-regulated learning: From teaching to self-reflective practice. Guilford Press (1998) +39. Schwenk, D., Khandelwal, A., Clark, C., Marino, K., Mottaghi, R.: A-okvqa: A benchmark for visual question answering using world knowledge. In: European Conference on Computer Vision. pp. 146-162. Springer (2022) +40. Sun, Z., Shen, S., Cao, S., Liu, H., Li, C., Shen, Y., Gan, C., Gui, L.Y., Wang, Y.X., Yang, Y., et al.: Aligning large multimodal models with factually augmented rlhf. arXiv preprint arXiv:2309.14525 (2023) +41. Team, G., Anil, R., Borgeaud, S., Wu, Y., Alayrac, J.B., Yu, J., Soricut, R., Schalkwyk, J., Dai, A.M., Hauth, A., et al.: Gemini: a family of highly capable multimodal models. arXiv preprint arXiv:2312.11805 (2023) +42. Touvron, H., Lavril, T., Izacard, G., Martinet, X., Lachaux, M.A., Lacroix, T., Rozière, B., Goyal, N., Hambro, E., Azhar, F., Rodriguez, A., Joulin, A., Grave, E., Lample, G.: Llama: Open and efficient foundation language models (2023) +43. Touvron, H., Martin, L., Stone, K., Albert, P., Almahairi, A., Babaei, Y., Bashlykov, N., Batra, S., Bhargava, P., Bhosale, S., et al.: Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288 (2023) +44. Wang, J., Meng, L., Weng, Z., He, B., Wu, Z., Jiang, Y.G.: To see is to believe: Prompting gpt-4v for better visual instruction tuning. arXiv preprint arXiv:2311.07574 (2023) +45. Wang, X., Pan, J., Ding, L., Biemann, C.: Mitigating hallucinations in large vision-language models with instruction contrastive decoding. arXiv preprint arXiv:2403.18715 (2024) +46. Wang, X., Wei, J., Schuurmans, D., Le, Q., Chi, E., Narang, S., Chowdhery, A., Zhou, D.: Self-consistency improves chain of thought reasoning in language models. arXiv preprint arXiv:2203.11171 (2022) +47. Wang, X., Wei, J., Schuurmans, D., Le, Q., Chi, E., Zhou, D.: Rationale-augmented ensembles in language models. arXiv preprint arXiv:2207.00747 (2022) +48. Wei, J., Wang, X., Schuurmans, D., Bosma, M., Xia, F., Chi, E., Le, Q.V., Zhou, D., et al.: Chain-of-thought prompting elicits reasoning in large language models. Advances in Neural Information Processing Systems 35, 24824-24837 (2022) +49. Wu, Y., Zhang, P., Xiong, W., Oguz, B., Gee, J.C., Nie, Y.: The role of chain-of-thought in complex vision-language reasoning task. arXiv preprint arXiv:2311.09193 (2023) + +50. Ye, Q., Xu, H., Xu, G., Ye, J., Yan, M., Zhou, Y., Wang, J., Hu, A., Shi, P., Shi, Y., et al.: mplug-owl: Modularization empowers large language models with multimodality. arXiv preprint arXiv:2304.14178 (2023) +51. Yin, S., Fu, C., Zhao, S., Xu, T., Wang, H., Sui, D., Shen, Y., Li, K., Sun, X., Chen, E.: Woodpecker: Hallucination correction for multimodal large language models. arXiv preprint arXiv:2310.16045 (2023) +52. Yin, Z., Wang, J., Cao, J., Shi, Z., Liu, D., Li, M., Huang, X., Wang, Z., Sheng, L., Bai, L., et al.: Lamm: Language-assisted multi-modal instruction-tuning dataset, framework, and benchmark. Advances in Neural Information Processing Systems 36 (2024) +53. Yu, Q., Li, J., Wei, L., Pang, L., Ye, W., Qin, B., Tang, S., Tian, Q., Zhuang, Y.: Hallucidoctor: Mitigating hallucinatory toxicity in visual instruction data (2023) +54. Zelikman, E., Mu, J., Goodman, N.D., Wu, Y.T.: Star: Self-taught reasoner bootstrapping reasoning with reasoning (2022) +55. Zhai, B., Yang, S., Zhao, X., Xu, C., Shen, S., Zhao, D., Keutzer, K., Li, M., Yan, T., Fan, X.: Halle-switch: Rethinking and controlling object existence hallucinations in large vision language models for detailed caption. arXiv preprint arXiv:2310.01779 (2023) +56. Zhang, P., Dong, X., Wang, B., Cao, Y., Xu, C., Ouyang, L., Zhao, Z., Ding, S., Zhang, S., Duan, H., Zhang, W., Yan, H., Zhang, X., Li, W., Li, J., Chen, K., He, C., Zhang, X., Qiao, Y., Lin, D., Wang, J.: Internl m-xcomposer: A vision-language large model for advanced text-image comprehension and composition. arXiv preprint arXiv:2309.15112 (2023) +57. Zhang, S., Roller, S., Goyal, N., Artetxe, M., Chen, M., Chen, S., Dewan, C., Diab, M., Li, X., Lin, X.V., et al.: Opt: Open pre-trained transformer language models. arXiv preprint arXiv:2205.01068 (2022) +58. Zhang, Y., Zhang, R., Gu, J., Zhou, Y., Lipka, N., Yang, D., Sun, T.: Llavar: Enhanced visual instruction tuning for text-rich image understanding. arXiv preprint arXiv:2306.17107 (2023) +59. Zhang, Z., Zhang, A., Li, M., Zhao, H., Karypis, G., Smola, A.: Multimodal chain-of-thought reasoning in language models. arXiv preprint arXiv:2302.00923 (2023) +60. Zhao, B., Wu, B., Huang, T.: Svit: Scaling up visual instruction tuning. arXiv preprint arXiv:2307.04087 (2023) +61. Zhao, L., Yu, E., Ge, Z., Yang, J., Wei, H., Zhou, H., Sun, J., Peng, Y., Dong, R., Han, C., et al.: Chatspot: Bootstrapping multimodal llms via precise referring instruction tuning. arXiv preprint arXiv:2307.09474 (2023) +62. Zheng, L., Chiang, W.L., Sheng, Y., Zhuang, S., Wu, Z., Zhuang, Y., Lin, Z., Li, Z., Li, D., Xing, E.P., Zhang, H., Gonzalez, J.E., Stoica, I.: Judging lvm-as-a-judge with mt-bench and chatbot arena (2023) +63. Zhu, D., Chen, J., Shen, X., Li, X., Elhoseiny, M.: Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592 (2023) +64. Zong, Z., Ma, B., Shen, D., Song, G., Shao, H., Jiang, D., Li, H., Liu, Y.: Mova: Adapting mixture of vision experts to multimodal context. arXiv preprint arXiv:2404.13046 (2024) \ No newline at end of file diff --git a/2024/Reflective Instruction Tuning_ Mitigating Hallucinations in Large Vision-Language Models/images.zip b/2024/Reflective Instruction Tuning_ Mitigating Hallucinations in Large Vision-Language Models/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..4f6ac48bc4acd0fd793668ae083bdb896cf29c0b --- /dev/null +++ b/2024/Reflective Instruction Tuning_ Mitigating Hallucinations in Large Vision-Language Models/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:625da67381adefe354c2832c891f2998dc1b2e45e01d3185bb44ea05ef3d31b0 +size 573083 diff --git a/2024/Reflective Instruction Tuning_ Mitigating Hallucinations in Large Vision-Language Models/layout.json b/2024/Reflective Instruction Tuning_ Mitigating Hallucinations in Large Vision-Language Models/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..f70deb7681aeef7b0dd5ed03f39091034d7d0386 --- /dev/null +++ b/2024/Reflective Instruction Tuning_ Mitigating Hallucinations in Large Vision-Language Models/layout.json @@ -0,0 +1,7015 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 137, + 112, + 477, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 112, + 477, + 148 + ], + "spans": [ + { + "bbox": [ + 137, + 112, + 477, + 148 + ], + "type": "text", + "content": "Reflective Instruction Tuning: Mitigating Hallucinations in Large Vision-Language Models" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 132, + 167, + 481, + 181 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 167, + 481, + 181 + ], + "spans": [ + { + "bbox": [ + 132, + 167, + 481, + 181 + ], + "type": "text", + "content": "Jinrui Zhang" + }, + { + "bbox": [ + 132, + 167, + 481, + 181 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 132, + 167, + 481, + 181 + ], + "type": "text", + "content": ", Teng Wang" + }, + { + "bbox": [ + 132, + 167, + 481, + 181 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 132, + 167, + 481, + 181 + ], + "type": "text", + "content": ", Haigang Zhang" + }, + { + "bbox": [ + 132, + 167, + 481, + 181 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 132, + 167, + 481, + 181 + ], + "type": "text", + "content": ", Ping Lu" + }, + { + "bbox": [ + 132, + 167, + 481, + 181 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 132, + 167, + 481, + 181 + ], + "type": "text", + "content": ", Feng Zheng" + }, + { + "bbox": [ + 132, + 167, + 481, + 181 + ], + "type": "inline_equation", + "content": "^{1,5}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 206, + 189, + 408, + 201 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 206, + 189, + 408, + 201 + ], + "spans": [ + { + "bbox": [ + 206, + 189, + 408, + 201 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 206, + 189, + 408, + 201 + ], + "type": "text", + "content": " Southern University of Science and Technology" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 241, + 201, + 373, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 241, + 201, + 373, + 213 + ], + "spans": [ + { + "bbox": [ + 241, + 201, + 373, + 213 + ], + "type": "text", + "content": "2 The University of Hong Kong" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 235, + 213, + 378, + 223 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 235, + 213, + 378, + 223 + ], + "spans": [ + { + "bbox": [ + 235, + 213, + 378, + 223 + ], + "type": "text", + "content": "3 Shenzhen Polytechnic University" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 180, + 223, + 434, + 234 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 180, + 223, + 434, + 234 + ], + "spans": [ + { + "bbox": [ + 180, + 223, + 434, + 234 + ], + "type": "text", + "content": "4 The Cloud Computing and IT Institute of ZTE Corporation" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 168, + 234, + 444, + 256 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 168, + 234, + 444, + 256 + ], + "spans": [ + { + "bbox": [ + 168, + 234, + 444, + 256 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 168, + 234, + 444, + 256 + ], + "type": "text", + "content": " Research Institute of Multiple Agents and Embodied Intelligence, Peng Cheng Laboratory, Shenzhen, China" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 160, + 278, + 455, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 278, + 455, + 553 + ], + "spans": [ + { + "bbox": [ + 160, + 278, + 455, + 553 + ], + "type": "text", + "content": "Abstract. Large vision-language models (LVLMs) have shown promising performance on a variety of vision-language tasks. However, they remain susceptible to hallucinations, generating outputs misaligned with visual content or instructions. While various mitigation strategies have been proposed, they often neglect a key contributor to hallucinations: lack of fine-grained reasoning supervision during training. Without intermediate reasoning steps, models may establish superficial shortcuts between instructions and responses, failing to internalize the inherent reasoning logic. To address this challenge, we propose reflective instruction tuning, which integrates rationale learning into visual instruction tuning. Unlike previous methods that learning from responses only, our approach entails the model predicting rationales justifying why responses are correct or incorrect. This fosters a deeper engagement with the fine-grained reasoning underlying each response, thus enhancing the model's reasoning proficiency. To facilitate this approach, we propose REVERIE, the first large-scale instruction-tuning dataset with ReflEctiVE Rationale annotations. REVERIE comprises 115k machine-generated reasoning instructions, each meticulously annotated with a corresponding pair of correct and confusing responses, alongside comprehensive rationales elucidating the justification behind the correctness or erroneousness of each response. Experimental results on multiple LVLM benchmarks reveal that reflective instruction tuning with the REVERIE dataset yields noticeable performance gain over the baseline model, demonstrating the effectiveness of reflecting from the rationales. Project page is at https://zjr2000.github.io/projects/reverie" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 160, + 562, + 455, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 562, + 455, + 586 + ], + "spans": [ + { + "bbox": [ + 160, + 562, + 455, + 586 + ], + "type": "text", + "content": "Keywords: Large Vision Language Models " + }, + { + "bbox": [ + 160, + 562, + 455, + 586 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 562, + 455, + 586 + ], + "type": "text", + "content": " Visual Instruction Tuning " + }, + { + "bbox": [ + 160, + 562, + 455, + 586 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 562, + 455, + 586 + ], + "type": "text", + "content": " Hallucination Mitigation" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 602, + 230, + 615 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 602, + 230, + 615 + ], + "spans": [ + { + "bbox": [ + 132, + 602, + 230, + 615 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 624, + 482, + 650 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 624, + 482, + 650 + ], + "spans": [ + { + "bbox": [ + 130, + 624, + 482, + 650 + ], + "type": "text", + "content": "Recently, large vision-language models (LVLMs) [1, 5, 10, 29, 30, 50, 58, 63] have garnered significant interest within the research community. By leveraging large" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 134, + 654, + 236, + 666 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 654, + 236, + 666 + ], + "spans": [ + { + "bbox": [ + 134, + 654, + 236, + 666 + ], + "type": "text", + "content": "* Corresponding author." + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 149, + 115, + 462, + 219 + ], + "blocks": [ + { + "bbox": [ + 149, + 115, + 462, + 219 + ], + "lines": [ + { + "bbox": [ + 149, + 115, + 462, + 219 + ], + "spans": [ + { + "bbox": [ + 149, + 115, + 462, + 219 + ], + "type": "image", + "image_path": "2fa31c18577b9bef1435711bf43fb3f98dcb8835e0202f63c351b6ba8714afda.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 229, + 482, + 308 + ], + "lines": [ + { + "bbox": [ + 130, + 229, + 482, + 308 + ], + "spans": [ + { + "bbox": [ + 130, + 229, + 482, + 308 + ], + "type": "text", + "content": "Fig. 1: Difference between vanilla instruction tuning and the proposed reflective instruction tuning. Vanilla instruction tuning only trains LVLMs solely for response generation, lacking of supervising the learning of fine-grained reasoning details. Reflective instruction tuning additionally trains the model to reflect the rationale underlying the response, which provides more fine-grained supervision (e.g., the key visual evidence and facts to reach the response, highlighted in red), facilitating the model learning to capture more critical information." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 338, + 482, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 338, + 482, + 542 + ], + "spans": [ + { + "bbox": [ + 130, + 338, + 482, + 542 + ], + "type": "text", + "content": "language models (LLMs) [8,9,33,42,43,62] as a unified task interface and employing visual instruction tuning [30] for enhanced task-level generalization, LVLMs have showcased remarkable reasoning performance and robustness in following user instructions. These breakthroughs offer substantial promise for the evolution of artificial general intelligence. Despite these advancements, LVLMs continue to be impeded by the phenomenon of hallucination [12, 24], wherein the models generate outputs that are inconsistent with visual content, established world knowledge, or the instructions provided. This significantly undermines the reliability of such models in real-world scenarios. Extensive research has investigated the origins of hallucinations in LVLMs, identifying factors such as noisy and biased training data [13, 15, 28], limited visual perception [1, 7], misalignment between vision and language modalities [18], and linguistic bias [12, 21]. To mitigate these hallucination sources, various mitigation methods have been proposed. These include: creating clean and balanced datasets [28, 55], scaling up visual resolution [1, 29] or utilizing additional perceptual models [51], employing alignment modules with increased parameters [7, 29], and designing decoding strategies that reduce over-reliance on language priors [21]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 546, + 483, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 546, + 483, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 546, + 483, + 666 + ], + "type": "text", + "content": "Despite significant progress, existing approaches often overlook a crucial factor contributing to hallucinations: the lack of fine-grained reasoning supervision. When learning to solve problems, humans often improve themselves by continuously reflecting on each step involved in the reasoning process. However, most previous LVLMs perform instruction tuning on datasets containing only instructions and direct responses, omitting access to the underlying rationale, i.e., intermediate reasoning steps. Specifically, Fig. 1 (left) shows the training paradigm of vanilla instruction tuning applied to a recognition task. The model is trained solely to predict the animal category name, without explicit supervision on the reasoning details from the discriminating information (e.g., the defining" + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 227, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 227, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 227, + 101 + ], + "type": "text", + "content": "J. Zhang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 175 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 175 + ], + "type": "text", + "content": "attributes of the animal) to the judgment. This strategy risks training the model to establish superficial shortcuts between instructions and responses, rather than understanding the tasks and internalizing the underlying logic of the reasoning process. Consequently, the model becomes susceptible to producing incorrect inferences and hallucinatory content misaligned with visual content." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 179, + 482, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 179, + 482, + 479 + ], + "spans": [ + { + "bbox": [ + 130, + 179, + 482, + 479 + ], + "type": "text", + "content": "Inspired by the human ability to improve reasoning through reflection on intermediate steps [2,38], we propose reflective instruction tuning to address the aforementioned issue. Unlike vanilla instruction tuning, which only trains the model to predict the response, we incorporate the learning of underlying rationales, providing the model with fine-grained reasoning supervision. Specifically, reflective instruction tuning trains the model to generate positive rationales (justifying correct responses) and negative rationales (explaining errors in incorrect responses). The positive rationale explicitly provides more fine-grained visual content, the relationship between multiple visual concepts, and step-by-step reasoning. This facilitates the model to capture finer visual details and better understand the task definition underlying instruction. The negative rationale provides more illustration of the core characteristics linked to the erroneous categorization of negative answers, helping the model concentrate on crucial concepts within the decision-making process. This improves the model's ability to discriminate between correct and potential hallucinatory responses. Fig. 1 (right) presents an example of positive and negative rationale, the positive rationale provides a more detailed justification of the correct answer, and the negative rationale highlights errors in the incorrect answer, aiding discrimination. Reflective instruction tuning mirrors the human cognitive process of reflection, in which individuals analyze step-by-step reasoning and errors to gain a deeper understanding. Training the model with both positive and negative rationales enhances several core cognitive capabilities: it improves visual perception through more visual information, facilitates instruction understanding with coherent reasoning logic, and promotes error awareness by highlighting critical information. Consequently, this training paradigm mitigates hallucinatory outputs." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 482, + 482, + 626 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 482, + 482, + 626 + ], + "spans": [ + { + "bbox": [ + 130, + 482, + 482, + 626 + ], + "type": "text", + "content": "While various visual instruction tuning datasets [4,28,30,44,52,60] exist, most of them consist solely of instruction-response pairs, omitting the annotation of rationales. To facilitate reflective instruction tuning that learns from both positive and negative rationales, we introduce REVERIE, the first large-scale visual instruction tuning dataset with ReflEctiVE rationalE annotations. REVERIE builds upon existing VQA datasets [31,39], and is further extended with natural images from Visual Genome [19]. The positive and negative rationales are annotated by Gemini family multimodal models [41]. REVERIE comprises 115k instructions encompassing four task types. Each instruction includes a positive response, an average of 1.2 negative responses, and accompanying rationales that justify the correctness or incorrectness of each response. This yields a total of 254k training instances, i.e., (instruction, response, rationale) tuples." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 630, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 630, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 630, + 481, + 665 + ], + "type": "text", + "content": "To enhance the adaptability of reflective instruction tuning to diverse tasks, we formulate the rationale data as a conversation format and position rationale learning subsequent to the model's prediction of responses, thereby decoupling" + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 328, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 328, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 328, + 91, + 447, + 102 + ], + "type": "text", + "content": "Reflective Instruction Tuning" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 308 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 308 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 308 + ], + "type": "text", + "content": "rationale generation from the context of response prediction. We perform reflective instruction tuning on two LVLMs with the proposed REVERIE dataset and further conduct experiments on multiple commonly used LVLM evaluation benchmarks. The results show that the proposed reflective tuning with REVERIE could yield substantial performance gain over the baselines. We also observed that utilizing the negative rationale could bring additional improvement than merely using the positive rationale. In summary, our contributions are three-fold: (a) We propose reflective instruction tuning, a novel approach that incorporates learning from both positive and negative rationales into visual instruction tuning to mitigate hallucination. (b) We propose REVERIE, the first large-scale instruction tuning dataset with both positive and negative fine-grained rationales. We hope this dataset could facilitate future research on rationale learning in LVLMs. (c) Extensive experiments show that reflective instruction tuning with the proposed dataset could bring substantial performance gain across multiple LVLM evaluation benchmarks, demonstrating the effectiveness of incorporating reflective rationales for mitigating hallucinations." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 327, + 237, + 340 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 327, + 237, + 340 + ], + "spans": [ + { + "bbox": [ + 132, + 327, + 237, + 340 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 354, + 482, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 354, + 482, + 628 + ], + "spans": [ + { + "bbox": [ + 130, + 354, + 482, + 628 + ], + "type": "text", + "content": "Large vision-language models. The remarkable generalization capabilities of LLMs [8,9,33,35,42,43,57,62] have inspired the LVLMs [1,5,10,29,30,50,58,63]. LVLMs aim to extend the reasoning abilities of LLMs into the visual domain, potentially enabling more sophisticated and multimodal understanding. Typically, the architecture of previous LVLMs incorporates three fundamental components: a pre-trained visual encoder (e.g., CLIP [34]) for visual perception, a pre-trained LLM (e.g., LLaMA [42]) serving as a language understanding and generation interface, and a connector (e.g., Q-former [10, 23], linear projector [29, 30, 63]) facilitating vision-language alignment. A common training paradigm for these models involves vision-language pre-training using large-scale vision-text datasets [27, 30, 37, 63] and visual instruction tuning [30] with various visual task datasets [14, 29, 30, 58]. To enhance the generalization capabilities of models, previous methods have explored several techniques, such as: 1) employing stronger perception models [17, 64] or LLMs [36, 56]; 2) designing improved connectors [3, 7, 20]; 3) utilizing a mixture-of-expert design [6, 26]. Given the importance of visual instruction tuning in building general-purpose visual assistants, many studies have also focused on improving performance through the creation of high-quality instruction tuning datasets. These datasets target a wide range of visual understanding capabilities, encompassing text-rich image comprehension [58], region-level analysis [61], chart interpretation [14], and multi-image reasoning [16]. The proposed REVIRIE dataset also aims to improve the model performance during visual instruction tuning, but we focus on enhancing the reasoning capability by offering fine-grained reasoning supervision." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 630, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 630, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 630, + 482, + 666 + ], + "type": "text", + "content": "Learning from reasoning steps. Several studies [11, 25, 31, 46-48] highlight the effectiveness of learning from detailed reasoning steps. Wei et al. [48] propose chain-of-thought prompting, where LLMs are explicitly prompted to think" + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "type": "text", + "content": "J. Zhang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 319 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 319 + ], + "type": "text", + "content": "step-by-step during inference, leading to substantial gains on reasoning tasks. Furthermore, recent work suggests that training with explicit rationales can improve model reasoning capabilities [32, 54]. Additionally, learning from answer explanations has shown promise in enhancing performance on vision-language reasoning tasks [31, 49, 59]. Despite these advancements, learning from reasoning steps in visual instruction tuning remains under-explored due to the scarcity of large-scale rationale-annotated visual instruction tuning datasets. LLaVA-Instruct-158k [30] contains a portion of annotations for complex reasoning tasks, and the responses have step-by-step reasoning process annotation that is similar to the concept of rationale. However, certain limitations remain: 1) these rationales are all positive coupled with correct responses, neglecting negative rationales crucial for discerning incorrect responses, and 2) they are restricted to open-ended reasoning tasks, offering limited benefits to the diverse range of tasks. We address these gaps by proposing REVERIE, the first large-scale visual instruction tuning dataset that contains both positive and negative rationales on various visual tasks, facilitating the future research on rationale learning in visual instruction tuning." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 320, + 482, + 511 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 320, + 482, + 511 + ], + "spans": [ + { + "bbox": [ + 130, + 320, + 482, + 511 + ], + "type": "text", + "content": "Hallucination in LVLMs. Previous research on hallucination in LVLMs has focused on two key areas: 1) the development of benchmarks for evaluating hallucinations [12, 24, 28, 40], and 2) the exploration of techniques to mitigate hallucinations [18, 22, 28, 45]. Li et al. [24] identify the object hallucination issue in LVLMs, and they propose the POPE benchmark for its evaluation. Guan et al. [12] divide the hallucination into language hallucination and visual hallucination and further introduce HALLUSIONBENCH to evaluate and diagnose each hallucination. Various studies address hallucination mitigation. Liu et al. [28] reduce hallucinations with a balanced visual instruction tuning dataset to better handle negative instructions. Jiang et al. [18] employ contrastive learning to distinguish hallucinated and non-hallucinated representations in latent space. Wang et al. [45] and Leng et al. [22] design contrastive decoding strategies in inference to mitigate the hallucination caused by language bias. Similar to [28], our work also focuses on mitigating hallucinations through improved visual instruction datasets. However, we emphasize fine-grained reasoning supervision to enhance consistency throughout the reasoning process." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 531, + 293, + 544 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 531, + 293, + 544 + ], + "spans": [ + { + "bbox": [ + 132, + 531, + 293, + 544 + ], + "type": "text", + "content": "3 The REVERIE Dataset" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "text", + "content": "Despite previous works have provided many diverse visual instruction tuning datasets (e.g., LLaVA-Instruct-158k [30], LRV-Instruction [28]), only a small fraction of samples possess rationale annotations for positive responses. Furthermore, these datasets often omit annotations for hard negative responses and their underlying rationales, rendering them hard to provide fine-grained and discriminative supervision for addressing hallucinations. To facilitate model \"reflect\" from both positive and negative responses, we propose REVERIE, the first large-scale visual instruction tuning dataset with detailed reflective rationale annotation, serving as a basis of the proposed reflective instruction tuning strategy" + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 328, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 328, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 328, + 91, + 447, + 102 + ], + "type": "text", + "content": "Reflective Instruction Tuning" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 151, + 114, + 463, + 264 + ], + "blocks": [ + { + "bbox": [ + 151, + 114, + 463, + 264 + ], + "lines": [ + { + "bbox": [ + 151, + 114, + 463, + 264 + ], + "spans": [ + { + "bbox": [ + 151, + 114, + 463, + 264 + ], + "type": "image", + "image_path": "169c9c074bad62e02529e118d075dac872dd0d5d8b03d07a1d76ffd7c42a9c20.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 273, + 482, + 319 + ], + "lines": [ + { + "bbox": [ + 130, + 273, + 482, + 319 + ], + "spans": [ + { + "bbox": [ + 130, + 273, + 482, + 319 + ], + "type": "text", + "content": "Fig. 2: Overview of the REVERIE dataset's data collection pipeline. We first employ Gemini-Vision-Pro to annotate the instructions, responses, and rationales for each image. Gemini-Pro is then used to check the consistency between positive and negative rationales. Inconsistent samples are filtered to maintain dataset quality." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 344, + 482, + 429 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 344, + 482, + 429 + ], + "spans": [ + { + "bbox": [ + 130, + 344, + 482, + 429 + ], + "type": "text", + "content": "(introduced in Sec. 4). As shown in Table 1, REVERIE provides hard negative response annotation and contains more extensive rationale annotations compared with the existing datasets. Fig. 2 shows the overall pipeline of our data collection strategy, we first annotate instructions, responses, and positive/negative rationales for each collected image, followed by filtering out noisy samples by the consistency-based filtering strategy. The details are as follows (The prompts for data generation are available in supplementary materials)." + } + ] + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 140, + 473, + 471, + 534 + ], + "blocks": [ + { + "bbox": [ + 132, + 451, + 481, + 464 + ], + "lines": [ + { + "bbox": [ + 132, + 451, + 481, + 464 + ], + "spans": [ + { + "bbox": [ + 132, + 451, + 481, + 464 + ], + "type": "text", + "content": "Table 1: Comparison of the REVERIE with the data used by other LVLMs." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 140, + 473, + 471, + 534 + ], + "lines": [ + { + "bbox": [ + 140, + 473, + 471, + 534 + ], + "spans": [ + { + "bbox": [ + 140, + 473, + 471, + 534 + ], + "type": "table", + "html": "
REVERIE LLaVA-1.5 InstructBLIP MiniGPT-4 LRV-Instruction
# Training Instances254k665k1.2M3.5k400k
Hard Negative Responses?✓XXXX
# Positive Rationales115k77k77kXX
# Negative Rationales138kXXXX
", + "image_path": "55ad36dec3cef6981313bcb76e7bce9e60a02d7d969c1619e33d90ca0521902d.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "text", + "content": "Data collection and annotation. To ensure the dataset encompasses diverse visual concepts and broad domain coverage, we sample a subset of images from the Visual Genome dataset [19]. Subsequently, we annotate these images with instructions and corresponding responses. Given the high cost of manual labelling, we leverage a strong LVLM, Gemini-Vision-Pro, as our annotator. Considering that the rationale for addressing simple query instruction may contain limited intermediate steps and benefit less to learning fine-grained reasoning, we explicitly prompt Gemini to generate instructions that need multi-step to solve and avoid generating simple queries. Additionally, it is prompted to generate both" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 227, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 227, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 227, + 101 + ], + "type": "text", + "content": "J. Zhang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "text", + "content": "positive and negative responses in conjunction with each instruction. Intuitively, a more confusing negative response demands more discriminative information to distinguish it from the positive answer. Thus, we explicitly prompt Gemini to generate negative responses that are easy to confuse." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 167, + 480, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 167, + 480, + 346 + ], + "spans": [ + { + "bbox": [ + 130, + 167, + 480, + 346 + ], + "type": "text", + "content": "After generating the instructions and responses, we prompt Gemini-Vision-Pro to generate rationales underlying the responses. Considering that positive and negative rationales have different emphases, we design two prompts and generate them separately. The positive rationales are expected to help the model learn the inherent knowledge and visual concepts underlying each problem-solving step, thereby facilitating fine-grained vision-language alignment and understanding of task definitions. It should include 1) analysis of visual content and instructions, 2) identification of core visual concepts relevant to the instructions, and 3) step-by-step reasoning grounded on the visual information and required knowledge. This is achieved by explicitly incorporating words such as \"explanation\", \"details\", and \"knowledge\" into the prompt to facilitate Gemini-Vision-Pro to include key image details and necessary outside knowledge in the rationale. The negative rationales are expected to provide discriminative information that aids in distinguishing between valid and hallucinatory outputs. Thus, we let Gemini-Vision-Pro highlight discriminative details in the negative rationale." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 349, + 481, + 409 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 349, + 481, + 409 + ], + "spans": [ + { + "bbox": [ + 130, + 349, + 481, + 409 + ], + "type": "text", + "content": "Consider A-OK-VQA [39] and ScienceQA [31] already have some human annotated QA pairs and positive rationales, we include them into the REVERIE dataset to increase task diversity. Since both datasets provide questions in a multiple-choice format, we incorporate all incorrect answer choices and generate an additional negative rationale for each of them with the same pipeline." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 411, + 481, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 411, + 481, + 567 + ], + "spans": [ + { + "bbox": [ + 130, + 411, + 481, + 567 + ], + "type": "text", + "content": "Consistency-based data filtering. While Gemini-Vision-Pro demonstrates impressive capabilities, it can still generate hallucinatory content when tasked to generate responses and rationales. Previous work [53] has shown that noise training data is one of the sources of the hallucinations. To guarantee the quality of generated data, we propose a method that leverages the consistency between positive and negative rationales to filter out noisy examples. Since both positive and negative rationales explicate the reasoning process behind the instruction but emphasize different aspects, they should contain information that is mutually consistent. When one rationale includes noisy information, this inherent consistency will be disrupted. Inspired by that, we employ a large language model, i.e., Gemini-Pro, to detect inconsistencies between these two types of rationales and only keep the samples that contain intra-consistent information (See supplementary materials for detailed prompt)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 570, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 481, + 665 + ], + "type": "text", + "content": "Data statistics. After the removal of noisy samples, the REVERIE dataset comprises 71,558 natural images. This includes 50,938 images sourced from Visual Genome [19], 15,706 from the COCO [27] and 4,914 images from ScienceQA [31]. REVERIE contains 115,280 instructions paired with corresponding positive responses and 138,897 negative responses, where each response is supplemented with a reflective rationale, rendering a total of 254,177 training instances. According to the output formats, REVERIE covers four types of vision-language tasks, including multiple-choice QA, short-answer QA, open-" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 328, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 328, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 328, + 91, + 447, + 102 + ], + "type": "text", + "content": "Reflective Instruction Tuning" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "type": "text", + "content": "ended QA, and Yes/No questions, and their distributions are visualized in Fig. 3 (a). We also analyzed the distribution of rationale length and the frequency of noun occurrence. As shown in Fig. 3 (b) and (c), over " + }, + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "type": "text", + "content": " of rationales exceed 25 words in length and contain more than 8 nouns, suggesting that rationales provide a substantial amount of information and knowledge." + } + ] + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 157, + 201, + 261, + 266 + ], + "blocks": [ + { + "bbox": [ + 157, + 201, + 261, + 266 + ], + "lines": [ + { + "bbox": [ + 157, + 201, + 261, + 266 + ], + "spans": [ + { + "bbox": [ + 157, + 201, + 261, + 266 + ], + "type": "image", + "image_path": "257896d69dfee4efb0526e08ad5b9e28d9e791df9bb68a3e216a6055cfeaf6fe.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 157, + 270, + 249, + 279 + ], + "lines": [ + { + "bbox": [ + 157, + 270, + 249, + 279 + ], + "spans": [ + { + "bbox": [ + 157, + 270, + 249, + 279 + ], + "type": "text", + "content": "(a) Distribution of Different Task" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 257, + 198, + 356, + 269 + ], + "blocks": [ + { + "bbox": [ + 257, + 198, + 356, + 269 + ], + "lines": [ + { + "bbox": [ + 257, + 198, + 356, + 269 + ], + "spans": [ + { + "bbox": [ + 257, + 198, + 356, + 269 + ], + "type": "image", + "image_path": "df35c8037a93b05ed1023ccb1dab4d8539b5885612f079474304d88278292497.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 257, + 270, + 356, + 279 + ], + "lines": [ + { + "bbox": [ + 257, + 270, + 356, + 279 + ], + "spans": [ + { + "bbox": [ + 257, + 270, + 356, + 279 + ], + "type": "text", + "content": "(b) Distribution of Rationale Length" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 358, + 198, + 463, + 269 + ], + "blocks": [ + { + "bbox": [ + 358, + 198, + 463, + 269 + ], + "lines": [ + { + "bbox": [ + 358, + 198, + 463, + 269 + ], + "spans": [ + { + "bbox": [ + 358, + 198, + 463, + 269 + ], + "type": "image", + "image_path": "59648b7be52428021db84a4a1cd804423db5192d912b395c8a4d7c8f407d1bb0.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 364, + 270, + 462, + 279 + ], + "lines": [ + { + "bbox": [ + 364, + 270, + 462, + 279 + ], + "spans": [ + { + "bbox": [ + 364, + 270, + 462, + 279 + ], + "type": "text", + "content": "(c) Distribution of noun occurrence" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 206, + 289, + 407, + 300 + ], + "lines": [ + { + "bbox": [ + 206, + 289, + 407, + 300 + ], + "spans": [ + { + "bbox": [ + 206, + 289, + 407, + 300 + ], + "type": "text", + "content": "Fig. 3: Statistics of the REVERIE dataset." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 347, + 329, + 361 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 347, + 329, + 361 + ], + "spans": [ + { + "bbox": [ + 132, + 347, + 329, + 361 + ], + "type": "text", + "content": "4 Reflective Instruction Tuning" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 376, + 482, + 519 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 376, + 482, + 519 + ], + "spans": [ + { + "bbox": [ + 130, + 376, + 482, + 519 + ], + "type": "text", + "content": "Reflection plays a crucial role in the human learning process, allowing individuals to extract meaningful insights, learn from their mistakes, and make informed choices in the future. However, vanilla instruction tuning solely focuses on training models to predict responses, thus omitting access to the underlying rationale. This limitation impedes the model's ability to effectively internalize and logically process instruction-relevant information, increasing the likelihood of generating hallucinatory outputs. Inspired by humans' ability to improve problem-solving skills through reflection on explanations for both correct and incorrect answers [2,38], we propose reflective instruction tuning. This approach facilitates learning from the rationale behind responses, extending the training process to predict both positive and negative rationales rather than merely generating correct responses to instructions." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 521, + 482, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 482, + 616 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 482, + 616 + ], + "type": "text", + "content": "Overall architecture. The objective of visual instruction tuning is to improve the instruction-following capability of the LVLM by leveraging diverse visual task data. Typically, LVLMs consist of a visual encoder, a vision-language connector, and an LLM. Given an image sample " + }, + { + "bbox": [ + 130, + 521, + 482, + 616 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 130, + 521, + 482, + 616 + ], + "type": "text", + "content": " and user instruction " + }, + { + "bbox": [ + 130, + 521, + 482, + 616 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 130, + 521, + 482, + 616 + ], + "type": "text", + "content": ", the visual encoder extracts image token features, which are then projected into the language space by the modality connector. The projected visual tokens, along with the textual instructions, serve as the prefix in the LLM. An auto-regressive LLM is trained to generate response " + }, + { + "bbox": [ + 130, + 521, + 482, + 616 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 130, + 521, + 482, + 616 + ], + "type": "text", + "content": " that aligns with both the instructions and the image." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "Multi-turn conversation. Directly training a model to concurrently predict rationales alongside responses in a single-turn prediction might hinder inference results. This is because some tasks do not necessitate explicit rationale generation, limiting its real-world applicability. To address this, we decouple rationale" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "type": "text", + "content": "J. Zhang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 284 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 284 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 284 + ], + "type": "text", + "content": "learning from response prediction to improve the adaptability of reflective instruction tuning to diverse tasks. We achieve this by formulating the rationale data in a multi-turn conversational format. The first turn incorporates " + }, + { + "bbox": [ + 130, + 116, + 482, + 284 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 130, + 116, + 482, + 284 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 116, + 482, + 284 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 130, + 116, + 482, + 284 + ], + "type": "text", + "content": " as inputs, prompting the model to generate " + }, + { + "bbox": [ + 130, + 116, + 482, + 284 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 130, + 116, + 482, + 284 + ], + "type": "text", + "content": ". Subsequently, the second turn introduces rationale generation prompts to let the model predict rationale. By putting the rationale learning behind, response prediction relies solely on the context of " + }, + { + "bbox": [ + 130, + 116, + 482, + 284 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 130, + 116, + 482, + 284 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 116, + 482, + 284 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 130, + 116, + 482, + 284 + ], + "type": "text", + "content": ", mitigating interference to the inference results. To make the model understand this rationale generation task, we explicitly prompt the model to generate knowledge and facts that are related to instruction. Specifically, in positive rationale learning, the model receives the prompt \"Explain why. Including any necessary facts or knowledge\" following its response. In negative rationale learning, an incorrect answer is provided, followed by the prompt \"Explain why this answer is wrong: {incorrect answer}. Including any necessary facts or knowledge.\"" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 300, + 230, + 313 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 300, + 230, + 313 + ], + "spans": [ + { + "bbox": [ + 132, + 300, + 230, + 313 + ], + "type": "text", + "content": "5 Experiments" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 322, + 483, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 322, + 483, + 430 + ], + "spans": [ + { + "bbox": [ + 130, + 322, + 483, + 430 + ], + "type": "text", + "content": "We demonstrate the effectiveness of learning from rationales by applying reflective instruction tuning with the REVERIE dataset to two baseline LVLMs: (1) LLaVA-1.0-7b-lora [30]: We initialize the model pre-trained with LAION-CC-SBU-558k, and perform instruction tuning with the combination of REVERIE and LLaVA-Instruct-80k. (2) LLaVA-1.5-7b-lora [29]: Similar to LLaVA-1.0, we perform instruction tuning with the combination of REVERIE and the original 665k instructions. The resulting models are denoted as REVERIE-1.0-7b-lora and REVERIE-1.5-7b-lora, respectively. Implementation details are provided in the supplementary materials." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 445, + 230, + 456 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 445, + 230, + 456 + ], + "spans": [ + { + "bbox": [ + 132, + 445, + 230, + 456 + ], + "type": "text", + "content": "5.1 Main Results" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 462, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 462, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 462, + 482, + 666 + ], + "type": "text", + "content": "Performance comparison with baselines on six benchmarks. We compare the performance of the model reflective tuned with the proposed REVERIE dataset. As shown in Table 2, training with REVERIE yields clear performance gain on most of the benchmarks across 2 baselines. Notably, we achieve a score of 83.8 on the POPE benchmark, reflecting a significant improvement of 12.7 points over the baseline LLaVA-1.0-7b-lora. This suggests that the REVERIE dataset provides more fine-grained supervision, enabling the model to make more accurate judgments regarding object existence and significantly alleviating hallucination issues. Additionally, we observe substantial performance gains on two comprehensive LVLM evaluation benchmarks, MME and MMBench. Specifically, REVERIE-1.0-7b-lora achieves improvements of 321.2 and 27.2 points on MME and MMBench, respectively, compared to the baseline. This suggests that 1) the REVERIE dataset encompasses a wider range of visual concepts and knowledge compared to the original LLaVA-Instruct-80k dataset, demonstrably benefiting performance on diverse visual perception tasks evaluated in MME; 2) incorporating rationales during training empowers the model to develop a deeper understanding of task definitions and effectively distinguish positive answers from" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 328, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 328, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 328, + 91, + 447, + 102 + ], + "type": "text", + "content": "Reflective Instruction Tuning" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 145, + 146, + 465, + 274 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 482, + 137 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 482, + 137 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 482, + 137 + ], + "type": "text", + "content": "Table 2: Comparison with baselines on six benchmarks. * means training images of the dataset are used during training. † means rerun using their released code." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 145, + 146, + 465, + 274 + ], + "lines": [ + { + "bbox": [ + 145, + 146, + 465, + 274 + ], + "spans": [ + { + "bbox": [ + 145, + 146, + 465, + 274 + ], + "type": "table", + "html": "
MethodScienceQA1POPEMMEMMBenchMM-VetGQA
InstructBLIP (vicuna-7b)60.5--3626.249.2
Qwen-VL-7b67.1--38.2-59.3
Qwen-VL-Chat-7b68.2-1487.560.6-57.5*
LLaVA-1.5-7b66.885.91510.764.330.562.0*
LLaVA-1.5-7b-lora68.486.41476.966.130.263.0*
LLaVA-1.5-7b-lora†67.786.11471.065.132.562.8*
LLaVA-1.0-7b-lora†42.771.1819.827.230.07.10
REVERIE-1.0-7b-lora70.1*83.81168.155.427.836.5
LLaVA-1.5-7b-lora (w/ ScienceQA)†76.3*86.61439.267.931.160.7*
REVERIE-1.5-7b-lora80.5*86.41474.967.330.861.8*
", + "image_path": "017a105a32025198033b6b5e2c1ba66ea88142355c14a339b85edb77aa411834.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 150, + 306, + 461, + 472 + ], + "blocks": [ + { + "bbox": [ + 176, + 285, + 436, + 297 + ], + "lines": [ + { + "bbox": [ + 176, + 285, + 436, + 297 + ], + "spans": [ + { + "bbox": [ + 176, + 285, + 436, + 297 + ], + "type": "text", + "content": "Table 3: Results of using different rationales on POPE." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 150, + 306, + 461, + 472 + ], + "lines": [ + { + "bbox": [ + 150, + 306, + 461, + 472 + ], + "spans": [ + { + "bbox": [ + 150, + 306, + 461, + 472 + ], + "type": "table", + "html": "
SubsetModelAccuracyPrecisionRecall|F1 Score
RandomResponse only86.8789.4284.53
w/ pos rationales85.6087.9383.53
w/ neg rationales86.2587.9884.93
w/ pos + neg rationales87.1590.2184.20
PopularResponse only82.5781.3384.53
w/ pos rationales81.8080.7383.53
w/ neg rationales83.3082.2584.93
w/ pos + neg rationales84.0083.8684.20
AdversarialResponse only76.8373.2584.53
w/ pos rationales77.2374.1983.53
w/ neg rationales77.7374.2484.93
w/ pos + neg rationales79.3776.7884.20
", + "image_path": "36dda41332f5a7a8e5808640062a3c827d07a1010095e5d668c4ff49a9467133.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 497, + 482, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 497, + 482, + 640 + ], + "spans": [ + { + "bbox": [ + 130, + 497, + 482, + 640 + ], + "type": "text", + "content": "provided negative answers, thereby achieving superior performance on multiple-choice question benchmarks like MMBench. Notably, for a fair comparison, we trained the baseline, LLaVA-1.5-7b-lora, using the combination of its original instruction tuning data and the VQA data from ScienceQA (only data with image context). We observed that simply introducing the ScienceQA QA data led to a performance decline on MME and GQA. This suggests that learning from response data alone may establish unintended shortcuts between instructions and responses, particularly for the complex reasoning instructions in ScienceQA. Our REVERIE-1.5-7b-lora demonstrates significant performance gains over its baseline on ScienceQA, MME, and GQA. Meanwhile, it achieves competitive performance on other benchmarks. This suggests that the rationale from our dataset may help alleviate the shortcut issue." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "content": "Effectiveness of positive/negative rationale. We conduct experiments to analyze the role of different rationales used in reflective instruction tuning. Ta-" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "type": "text", + "content": "J. Zhang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 145, + 168, + 465, + 236 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 480, + 159 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 480, + 159 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 480, + 159 + ], + "type": "text", + "content": "Table 4: Results of using different rationales on MMBench. We report L2-ability performance, where AR, CP, FP-S, FP-C, LR, and RR denote attribute reasoning, coarse perception, fine-grained perception (single-instance), fine-grained perception (cross-instance), logic reasoning, and relation reasoning, respectively." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 145, + 168, + 465, + 236 + ], + "lines": [ + { + "bbox": [ + 145, + 168, + 465, + 236 + ], + "spans": [ + { + "bbox": [ + 145, + 168, + 465, + 236 + ], + "type": "table", + "html": "
ModelARCPFP-SFP-CLRRRAccuracy
Response only61.6960.0950.5142.7626.6742.6151.70
w/ pos rationales63.1867.1144.1449.1624.1740.8752.13
w/ neg rationales64.6864.4350.8444.8322.5043.4852.30
w/ pos + neg rationales65.1468.4657.2451.5219.1750.4355.44
", + "image_path": "94bb53c457128eb3a0c60ca71bc89905fc77e3bf058ee5113f324e903d33b4a1.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 145, + 278, + 465, + 346 + ], + "blocks": [ + { + "bbox": [ + 132, + 246, + 479, + 269 + ], + "lines": [ + { + "bbox": [ + 132, + 246, + 479, + 269 + ], + "spans": [ + { + "bbox": [ + 132, + 246, + 479, + 269 + ], + "type": "text", + "content": "Table 5: Results of using different rationales on GQA. We report the semantic-level accuracy to analyze the impact of rationale on questions with different subjects." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 145, + 278, + 465, + 346 + ], + "lines": [ + { + "bbox": [ + 145, + 278, + 465, + 346 + ], + "spans": [ + { + "bbox": [ + 145, + 278, + 465, + 346 + ], + "type": "table", + "html": "
ModelAttributeObjectRelationGlobal CategoryAccuracy
Response only47.1759.9022.2131.2119.67
w/ pos rationales46.7857.4623.4429.3022.89
w/ neg rationales48.0156.6824.7731.8528.63
w/ pos + neg rationales47.3258.3524.7631.8528.11
36.53
", + "image_path": "ab37b76df90d14cf9fc9dde18eaa82f81f17620fd9e77f24398b129767f8339a.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 367, + 480, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 367, + 480, + 582 + ], + "spans": [ + { + "bbox": [ + 132, + 367, + 480, + 582 + ], + "type": "text", + "content": "ble 3, Table 4 and Table 5 show the results of using different rationales on POPE, MMBench, and GQA, respectively. The results show that using both positive and negative rationale yields the best performance on most metrics. The results on POPE show that using both two rationales brings the highest F1 score, suggesting that training with rationale may alleviate the instruction-response shortcut and further alleviate the object hallucination issue. The results on MMBench show that training exclusively with positive rationales yields higher perception scores (in terms of CP and FP-C). This indicates that positive rationales enhance the capture of visual details. Conversely, training solely with negative rationales results in superior reasoning performance (in terms of AR and RR), suggesting that negative rationales provide critical clues that improve discriminative ability. We also infer that training with negative rationales could help the model better distinguish the correct answer from the negative response, benefiting more to the task with candidate answers provided, e.g., multiple choice. Our results on GQA further show that both positive and negative rationales contribute to performance gains in questions with category as their subject. This suggests that rationales provide more fine-grained information that enables the model to better understand the defining characteristics of a category." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 133, + 599, + 244, + 609 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 599, + 244, + 609 + ], + "spans": [ + { + "bbox": [ + 133, + 599, + 244, + 609 + ], + "type": "text", + "content": "5.2 Ablation Studies" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 617, + 479, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 617, + 479, + 664 + ], + "spans": [ + { + "bbox": [ + 132, + 617, + 479, + 664 + ], + "type": "text", + "content": "We investigate several key designs of reflective instruction tuning and REVERIE: 1) the design of rationale generation prompt; 2) the effectiveness of consistency-based data filtering; 3) the design of conversational context; All experiments are conducted with REVERIE-1.0-7b-lora." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 328, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 328, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 328, + 91, + 447, + 102 + ], + "type": "text", + "content": "Reflective Instruction Tuning" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 141, + 136, + 470, + 205 + ], + "blocks": [ + { + "bbox": [ + 155, + 114, + 457, + 126 + ], + "lines": [ + { + "bbox": [ + 155, + 114, + 457, + 126 + ], + "spans": [ + { + "bbox": [ + 155, + 114, + 457, + 126 + ], + "type": "text", + "content": "Table 6: Performance comparison of different rationale prompts." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 141, + 136, + 470, + 205 + ], + "lines": [ + { + "bbox": [ + 141, + 136, + 470, + 205 + ], + "spans": [ + { + "bbox": [ + 141, + 136, + 470, + 205 + ], + "type": "table", + "html": "
Explicit GuidanceType-SpecificLength ControlMMBPOPEGQAMME
(a)55.483.836.81128.6
(b)✓✓55.283.236.01111.9
(c)✓✓✓54.383.835.81141.0
(d)✓55.483.836.51168.1
", + "image_path": "75c0fee9eb24358348d8939e5fa1cd6dfc5405a36028a78930941f04f55cf593.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 228, + 481, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 228, + 481, + 407 + ], + "spans": [ + { + "bbox": [ + 130, + 228, + 481, + 407 + ], + "type": "text", + "content": "Rationale generation prompt. We conducted ablation studies to analyze the design of the rationale generation prompt employed in reflective instruction tuning. Specifically, we test the variants of rationale generation prompt from three perspectives: (1) the use of explicit guidance, e.g., \"Including any necessary facts or knowledge\"; 2) the utilization of separate prompts for positive and negative rationales; 3) the use of length-specific prompts. Table 6 summarizes the performance of these prompt variants. Our final choice, represented by Line (d), includes only explicit guidance and yields the best overall performance across four benchmarks. Line (a) demonstrates that using the prompt without explicit guidance leads to a significant performance drop on the MME benchmark. This suggests that providing clear guidance for clarifying the rationale generation task is crucial for aligning rationales with visual content. Moreover, Lines (c) and (d) indicate that overly complex prompts containing excessive constraints may impede the model's understanding of the rationale generation task, adversely affecting performance across most benchmarks." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 408, + 481, + 493 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 408, + 481, + 493 + ], + "spans": [ + { + "bbox": [ + 130, + 408, + 481, + 493 + ], + "type": "text", + "content": "Consistency-based data filtering. We conducted experiments on the noise-infused rationale data to investigate the effectiveness of the proposed consistency-based data filtering strategy. Specifically, we train the model with the rationale data before filtering. As shown in Table 7, training with noised data that are not intra-consistent resulted in performance degradation across most benchmarks, suggesting that using the consistency-based filter strategy to maintain data quality is necessary for learning from rationale." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 137, + 518, + 307, + 541 + ], + "lines": [ + { + "bbox": [ + 137, + 518, + 307, + 541 + ], + "spans": [ + { + "bbox": [ + 137, + 518, + 307, + 541 + ], + "type": "text", + "content": "Table 7: Effect of rationale data filtering on model performance." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "text" + }, + { + "type": "table", + "bbox": [ + 148, + 551, + 294, + 590 + ], + "blocks": [ + { + "bbox": [ + 148, + 551, + 294, + 590 + ], + "lines": [ + { + "bbox": [ + 148, + 551, + 294, + 590 + ], + "spans": [ + { + "bbox": [ + 148, + 551, + 294, + 590 + ], + "type": "table", + "html": "
ModelMMBPOPEGQAMME
w/o filtering55.283.536.31132.3
w/ filtering55.483.836.51168.1
", + "image_path": "b4673171eb5d19d86b92ab6b97a067b14fcc67fe27325cde95b1d1c5f6b30386.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 313, + 548, + 478, + 594 + ], + "blocks": [ + { + "bbox": [ + 311, + 515, + 482, + 537 + ], + "lines": [ + { + "bbox": [ + 311, + 515, + 482, + 537 + ], + "spans": [ + { + "bbox": [ + 311, + 515, + 482, + 537 + ], + "type": "text", + "content": "Table 8: Performance comparison of different conversational contexts." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 313, + 548, + 478, + 594 + ], + "lines": [ + { + "bbox": [ + 313, + 548, + 478, + 594 + ], + "spans": [ + { + "bbox": [ + 313, + 548, + 478, + 594 + ], + "type": "table", + "html": "
ModelMMBPOPEGQAMME
pos rationale first49.379.536.21144.7
neg rationale first48.682.436.41160.2
separate context55.483.836.51168.1
", + "image_path": "42014aac02dabed315f5b726add83b325c67b37b42e5d0468c611e9273a2c032.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "type": "text", + "content": "Conversational context of reflective instruction tuning. We conducted ablation studies on several training data format variants, each incorporating different conversational contexts. We test the performance of three variants: 1) positive rationale first, where all negative rationales are learned to predict after" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 227, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 227, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 227, + 101 + ], + "type": "text", + "content": "J. Zhang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 272 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 272 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 272 + ], + "type": "text", + "content": "positive rationale turns; 2) negative rationale first, the inverse order of the previous one; 3) separate context, where the positive and negative rationales are separate into two distinct samples. The previous two variants put the positive and negative rationales for one instruction into a single conversation, making the model see previous rationales when learning to predict the current one. Table 8 shows the performance of these three variants. Our final choice that using separate contexts consistently outperforms the other two variants across four benchmarks. This suggests that providing the model with access to additional rationales during the current rationale generation step might lead to information leakage. In such a scenario, the model may tend to leverage pre-existing information from other rationales rather than learning to reason according to visual information and acquired knowledge. This hinders the model benefited from the fine-grained supervision from rationales." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 290, + 263, + 303 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 290, + 263, + 303 + ], + "spans": [ + { + "bbox": [ + 132, + 290, + 263, + 303 + ], + "type": "text", + "content": "5.3 Qualitative Analysis" + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 161, + 334, + 452, + 534 + ], + "blocks": [ + { + "bbox": [ + 161, + 334, + 452, + 534 + ], + "lines": [ + { + "bbox": [ + 161, + 334, + 452, + 534 + ], + "spans": [ + { + "bbox": [ + 161, + 334, + 452, + 534 + ], + "type": "image", + "image_path": "0e96a6377ce4e56b9b578d60c423e9cfb9125439e30e377ca64b966f595e5ff5.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 166, + 545, + 446, + 557 + ], + "lines": [ + { + "bbox": [ + 166, + 545, + 446, + 557 + ], + "spans": [ + { + "bbox": [ + 166, + 545, + 446, + 557 + ], + "type": "text", + "content": "Fig. 4: Visualization of the generation of positive rationales." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "content": "We conducted a case analysis to assess whether reflective instruction tuning improves fine-grained reasoning and mitigates hallucination. We compared baseline models with our proposed approach by prompting both to generate explicit rationales for their responses, aiming to verify if the models' logic aligns with correct reasoning. Fig. 4 visualizes the output rationales of different models. LLaVA directly generates rationale along with the response. However, the response is incorrect, and the rationale contains hallucinated content (highlighted in red)." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 328, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 328, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 328, + 91, + 447, + 102 + ], + "type": "text", + "content": "Reflective Instruction Tuning" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 160, + 117, + 452, + 319 + ], + "blocks": [ + { + "bbox": [ + 160, + 117, + 452, + 319 + ], + "lines": [ + { + "bbox": [ + 160, + 117, + 452, + 319 + ], + "spans": [ + { + "bbox": [ + 160, + 117, + 452, + 319 + ], + "type": "image", + "image_path": "d3d8975ded3131c01a0efc384b10725b34d85ae65dca9acd64bd81ebacaff77b.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 165, + 331, + 448, + 342 + ], + "lines": [ + { + "bbox": [ + 165, + 331, + 448, + 342 + ], + "spans": [ + { + "bbox": [ + 165, + 331, + 448, + 342 + ], + "type": "text", + "content": "Fig. 5: Visualization of the generation of negative rationales." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 363, + 482, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 363, + 482, + 509 + ], + "spans": [ + { + "bbox": [ + 130, + 363, + 482, + 509 + ], + "type": "text", + "content": "REVERIE (w/o R) (i.e., the model trained without rationales) and REVERIE both predicted the correct answer. However, REVERIE (w/o R) generates a rationale inconsistent with the response, suggesting that training without fine-grained supervision may lead to the shortcut between instruction and response rather than solving the problem from the correct reasoning paths. In contrast, the rationale generated by REVERIE presents more information that is necessary (highlighted in green) for addressing the instruction. We further evaluate the models' capability to generate negative rationales. As shown in Fig. 5, REVERIE correctly identifies crucial information (e.g., all presented colors) from images, while the other two models are misled by the context of two sub-images. This suggests that incorporating rationale as supervision enhances the model's fine-grained understanding of visual context, reducing hallucination." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 523, + 220, + 536 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 523, + 220, + 536 + ], + "spans": [ + { + "bbox": [ + 132, + 523, + 220, + 536 + ], + "type": "text", + "content": "6 Conclusion" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": "In this work, we identify a key contributor to hallucinations in existing LVLMs: the lack of fine-grained reasoning supervision during visual instruction tuning. To address this, we introduce reflective instruction tuning, which draws inspiration from human reflective learning processes by training the model on the rationales of positive and negative answers. Additionally, we propose REVERIE, the first large-scale visual instruction tuning dataset featuring comprehensive annotations for both positive and negative rationales. Experiments demonstrate that reflective instruction tuning with REVERIE significantly enhances LVLM performance, emphasizing the importance of fine-grained reasoning supervision in mitigating hallucinations." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "type": "text", + "content": "J. Zhang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 240, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 240, + 129 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 240, + 129 + ], + "type": "text", + "content": "Acknowledgement" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 141, + 480, + 166 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 141, + 480, + 166 + ], + "spans": [ + { + "bbox": [ + 132, + 141, + 480, + 166 + ], + "type": "text", + "content": "This work is supported by the National Natural Science Foundation of China (Grant NO.62122035)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 185, + 197, + 198 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 185, + 197, + 198 + ], + "spans": [ + { + "bbox": [ + 133, + 185, + 197, + 198 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 211, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 138, + 211, + 481, + 244 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 211, + 481, + 244 + ], + "spans": [ + { + "bbox": [ + 138, + 211, + 481, + 244 + ], + "type": "text", + "content": "1. Bai, J., Bai, S., Yang, S., Wang, S., Tan, S., Wang, P., Lin, J., Zhou, C., Zhou, J.: Qwen-vl: A frontier large vision-language model with versatile abilities. arXiv preprint arXiv:2308.12966 (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 246, + 481, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 246, + 481, + 267 + ], + "spans": [ + { + "bbox": [ + 138, + 246, + 481, + 267 + ], + "type": "text", + "content": "2. Boud, D., Keogh, R., Walker, D.: Reflection: Turning experience into learning. Routledge (2013)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 269, + 481, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 269, + 481, + 300 + ], + "spans": [ + { + "bbox": [ + 138, + 269, + 481, + 300 + ], + "type": "text", + "content": "3. Cha, J., Kang, W., Mun, J., Roh, B.: Honeybee: Locality-enhanced projector for multimodal llm. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13817-13827 (2024)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 301, + 481, + 322 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 301, + 481, + 322 + ], + "spans": [ + { + "bbox": [ + 138, + 301, + 481, + 322 + ], + "type": "text", + "content": "4. Chen, D., Liu, J., Dai, W., Wang, B.: Visual instruction tuning with polite flamingo. arXiv preprint arXiv:2307.01003 (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 323, + 481, + 355 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 323, + 481, + 355 + ], + "spans": [ + { + "bbox": [ + 138, + 323, + 481, + 355 + ], + "type": "text", + "content": "5. Chen, K., Zhang, Z., Zeng, W., Zhang, R., Zhu, F., Zhao, R.: Shikra: Unleashing multimodal llm's referential dialogue magic. arXiv preprint arXiv:2306.15195 (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 357, + 481, + 388 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 357, + 481, + 388 + ], + "spans": [ + { + "bbox": [ + 138, + 357, + 481, + 388 + ], + "type": "text", + "content": "6. Chen, S., Jie, Z., Ma, L.: Llava-mole: Sparse mixture of lora experts for mitigating data conflicts in instruction finetuning mllms. arXiv preprint arXiv:2401.16160 (2024)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 389, + 481, + 422 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 389, + 481, + 422 + ], + "spans": [ + { + "bbox": [ + 138, + 389, + 481, + 422 + ], + "type": "text", + "content": "7. Chen, Z., Wu, J., Wang, W., Su, W., Chen, G., Xing, S., Muyan, Z., Zhang, Q., Zhu, X., Lu, L., et al.: Internvl: Scaling up vision foundation models and aligning for generic visual-linguistic tasks. arXiv preprint arXiv:2312.14238 (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 423, + 481, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 423, + 481, + 465 + ], + "spans": [ + { + "bbox": [ + 138, + 423, + 481, + 465 + ], + "type": "text", + "content": "8. Chowdhery, A., Narang, S., Devlin, J., Bosma, M., Mishra, G., Roberts, A., Barham, P., Chung, H.W., Sutton, C., Gehrmann, S., et al.: Palm: Scaling language modeling with pathways. Journal of Machine Learning Research 24(240), 1-113 (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 467, + 481, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 467, + 481, + 498 + ], + "spans": [ + { + "bbox": [ + 138, + 467, + 481, + 498 + ], + "type": "text", + "content": "9. Chung, H.W., Hou, L., Longpre, S., Zoph, B., Tay, Y., Fedus, W., Li, Y., Wang, X., Dehghani, M., Brahma, S., et al.: Scaling instruction-finetuned language models. arXiv preprint arXiv:2210.11416 (2022)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 500, + 481, + 532 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 500, + 481, + 532 + ], + "spans": [ + { + "bbox": [ + 138, + 500, + 481, + 532 + ], + "type": "text", + "content": "0. Dai, W., Li, J., Li, D., Tiong, A.M.H., Zhao, J., Wang, W., Li, B., Fung, P., Hoi, S.: Instructblip: Towards general-purpose vision-language models with instruction tuning (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 533, + 481, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 533, + 481, + 555 + ], + "spans": [ + { + "bbox": [ + 138, + 533, + 481, + 555 + ], + "type": "text", + "content": "1. Ge, J., Luo, H., Qian, S., Gan, Y., Fu, J., Zhan, S.: Chain of thought prompt tuning in vision language models. arXiv preprint arXiv:2304.07919 (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 555, + 481, + 587 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 555, + 481, + 587 + ], + "spans": [ + { + "bbox": [ + 138, + 555, + 481, + 587 + ], + "type": "text", + "content": "2. Guan, T., Liu, F., Li, X.W.R.X.Z., Wang, X.L.X., Yacoob, L.C.F.H.Y., Zhou, D.M.T.: Hallusionbench: An advanced diagnostic suite for entangled language hallucination & visual illusion in large vision-language models" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 589, + 481, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 589, + 481, + 610 + ], + "spans": [ + { + "bbox": [ + 138, + 589, + 481, + 610 + ], + "type": "text", + "content": "3. Gunjal, A., Yin, J., Bas, E.: Detecting and preventing hallucinations in large vision language models. arXiv preprint arXiv:2308.06394 (2023)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 611, + 481, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 611, + 481, + 643 + ], + "spans": [ + { + "bbox": [ + 138, + 611, + 481, + 643 + ], + "type": "text", + "content": "4. Han, Y., Zhang, C., Chen, X., Yang, X., Wang, Z., Yu, G., Fu, B., Zhang, H.: Chartllama: A multimodal llm for chart understanding and generation. arXiv preprint arXiv:2311.16483 (2023)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 138, + 644, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 644, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 644, + 481, + 665 + ], + "type": "text", + "content": "5. Hu, H., Zhang, J., Zhao, M., Sun, Z.: Ciem: Contrastive instruction evaluation method for better instruction tuning. arXiv preprint arXiv:2309.02301 (2023)" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 328, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 328, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 328, + 91, + 447, + 102 + ], + "type": "text", + "content": "Reflective Instruction Tuning" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 133, + 116, + 480, + 148 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 480, + 148 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 480, + 148 + ], + "type": "text", + "content": "16. Huang, Y., Meng, Z., Liu, F., Su, Y., Collier, N., Lu, Y.: Sparkles: Unlocking chats across multiple images for multimodal instruction-following models. arXiv preprint arXiv:2308.16463 (2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 149, + 480, + 181 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 149, + 480, + 181 + ], + "spans": [ + { + "bbox": [ + 133, + 149, + 480, + 181 + ], + "type": "text", + "content": "17. Jain, J., Yang, J., Shi, H.: Vcoder: Versatile vision encoders for multimodal large language models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 27992-28002 (2024)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 182, + 480, + 213 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 182, + 480, + 213 + ], + "spans": [ + { + "bbox": [ + 133, + 182, + 480, + 213 + ], + "type": "text", + "content": "18. Jiang, C., Xu, H., Dong, M., Chen, J., Ye, W., Yan, M., Ye, Q., Zhang, J., Huang, F., Zhang, S.: Hallucination augmented contrastive learning for multimodal large language model. arXiv preprint arXiv:2312.06968 (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 214, + 480, + 257 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 214, + 480, + 257 + ], + "spans": [ + { + "bbox": [ + 133, + 214, + 480, + 257 + ], + "type": "text", + "content": "19. Krishna, R., Zhu, Y., Groth, O., Johnson, J., Hata, K., Kravitz, J., Chen, S., Kalantidis, Y., Li, L.J., Shamma, D.A., et al.: Visual genome: Connecting language and vision using crowdsourced dense image annotations. International journal of computer vision 123, 32-73 (2017)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 133, + 258, + 480, + 289 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 258, + 480, + 289 + ], + "spans": [ + { + "bbox": [ + 133, + 258, + 480, + 289 + ], + "type": "text", + "content": "20. Lee, B.K., Kim, C.W., Park, B., Ro, Y.M.: Meteor: Mamba-based traversal of rationale for large language and vision models. arXiv preprint arXiv:2405.15574 (2024)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 290, + 480, + 322 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 290, + 480, + 322 + ], + "spans": [ + { + "bbox": [ + 132, + 290, + 480, + 322 + ], + "type": "text", + "content": "21. Leng, S., Zhang, H., Chen, G., Li, X., Lu, S., Miao, C., Bing, L.: Mitigating object hallucinations in large vision-language models through visual contrastive decoding. arXiv preprint arXiv:2311.16922 (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 322, + 480, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 322, + 480, + 365 + ], + "spans": [ + { + "bbox": [ + 132, + 322, + 480, + 365 + ], + "type": "text", + "content": "22. Leng, S., Zhang, H., Chen, G., Li, X., Lu, S., Miao, C., Bing, L.: Mitigating object hallucinations in large vision-language models through visual contrastive decoding. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13872-13882 (2024)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 365, + 480, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 365, + 480, + 396 + ], + "spans": [ + { + "bbox": [ + 132, + 365, + 480, + 396 + ], + "type": "text", + "content": "23. Li, J., Li, D., Savarese, S., Hoi, S.: Blip-2: Bootstrapping language-image pretraining with frozen image encoders and large language models. arXiv preprint arXiv:2301.12597 (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 397, + 480, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 397, + 480, + 418 + ], + "spans": [ + { + "bbox": [ + 132, + 397, + 480, + 418 + ], + "type": "text", + "content": "24. Li, Y., Du, Y., Zhou, K., Wang, J., Zhao, W.X., Wen, J.R.: Evaluating object hallucination in large vision-language models. arXiv preprint arXiv:2305.10355 (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 133, + 419, + 480, + 450 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 419, + 480, + 450 + ], + "spans": [ + { + "bbox": [ + 133, + 419, + 480, + 450 + ], + "type": "text", + "content": "25. Lightman, H., Kosaraju, V., Burda, Y., Edwards, H., Baker, B., Lee, T., Leike, J., Schulman, J., Sutskever, I., Cobbe, K.: Let's verify step by step. arXiv preprint arXiv:2305.20050 (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 133, + 450, + 480, + 482 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 450, + 480, + 482 + ], + "spans": [ + { + "bbox": [ + 133, + 450, + 480, + 482 + ], + "type": "text", + "content": "26. Lin, B., Tang, Z., Ye, Y., Cui, J., Zhu, B., Jin, P., Zhang, J., Ning, M., Yuan, L.: Moe-llava: Mixture of experts for large vision-language models. arXiv preprint arXiv:2401.15947 (2024)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 133, + 483, + 480, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 483, + 480, + 525 + ], + "spans": [ + { + "bbox": [ + 133, + 483, + 480, + 525 + ], + "type": "text", + "content": "27. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dólar, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13. pp. 740-755. Springer (2014)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 133, + 526, + 480, + 558 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 526, + 480, + 558 + ], + "spans": [ + { + "bbox": [ + 133, + 526, + 480, + 558 + ], + "type": "text", + "content": "28. Liu, F., Lin, K., Li, L., Wang, J., Yacoob, Y., Wang, L.: Mitigating hallucination in large multi-modal models via robust instruction tuning. In: The Twelfth International Conference on Learning Representations (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 133, + 559, + 480, + 580 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 559, + 480, + 580 + ], + "spans": [ + { + "bbox": [ + 133, + 559, + 480, + 580 + ], + "type": "text", + "content": "29. Liu, H., Li, C., Li, Y., Lee, Y.J.: Improved baselines with visual instruction tuning. arXiv preprint arXiv:2310.03744 (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 133, + 580, + 480, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 580, + 480, + 601 + ], + "spans": [ + { + "bbox": [ + 133, + 580, + 480, + 601 + ], + "type": "text", + "content": "30. Liu, H., Li, C., Wu, Q., Lee, Y.J.: Visual instruction tuning. Advances in neural information processing systems 36 (2024)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 133, + 601, + 480, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 601, + 480, + 643 + ], + "spans": [ + { + "bbox": [ + 133, + 601, + 480, + 643 + ], + "type": "text", + "content": "31. Lu, P., Mishra, S., Xia, T., Qiu, L., Chang, K.W., Zhu, S.C., Tafjord, O., Clark, P., Kalyan, A.: Learn to explain: Multimodal reasoning via thought chains for science question answering. Advances in Neural Information Processing Systems 35, 2507-2521 (2022)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 133, + 644, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 644, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 644, + 480, + 665 + ], + "type": "text", + "content": "32. Magister, L.C., Mallinson, J., Adamek, J., Malmi, E., Severyn, A.: Teaching small language models to reason. arXiv preprint arXiv:2212.08410 (2022)" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 227, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 227, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 227, + 101 + ], + "type": "text", + "content": "J. Zhang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 665 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 160 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 160 + ], + "type": "text", + "content": "33. Ouyang, L., Wu, J., Jiang, X., Almeida, D., Wainwright, C., Mishkin, P., Zhang, C., Agarwal, S., Slama, K., Ray, A., et al.: Training language models to follow instructions with human feedback. Advances in Neural Information Processing Systems 35, 27730-27744 (2022)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 160, + 481, + 203 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 160, + 481, + 203 + ], + "spans": [ + { + "bbox": [ + 130, + 160, + 481, + 203 + ], + "type": "text", + "content": "34. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 203, + 481, + 236 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 203, + 481, + 236 + ], + "spans": [ + { + "bbox": [ + 130, + 203, + 481, + 236 + ], + "type": "text", + "content": "35. Raffel, C., Shazeer, N., Roberts, A., Lee, K., Narang, S., Matena, M., Zhou, Y., Li, W., Liu, P.J.: Exploring the limits of transfer learning with a unified text-to-text transformer. The Journal of Machine Learning Research 21(1), 5485-5551 (2020)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 236, + 481, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 236, + 481, + 258 + ], + "spans": [ + { + "bbox": [ + 130, + 236, + 481, + 258 + ], + "type": "text", + "content": "36. Rasheed, H., Maaz, M., Khan, S., Khan, F.S.: Llava++: Extending visual capabilities with llama-3 and phi-3 (2024), https://github.com/mbzuaai-oryx/LLaVA-pp" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 258, + 481, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 258, + 481, + 300 + ], + "spans": [ + { + "bbox": [ + 130, + 258, + 481, + 300 + ], + "type": "text", + "content": "37. Schuhmann, C., Beaumont, R., Vencu, R., Gordon, C., Wightman, R., Cherti, M., Coombes, T., Katta, A., Mullis, C., Wortsman, M., et al.: Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems 35, 25278-25294 (2022)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 300, + 481, + 321 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 300, + 481, + 321 + ], + "spans": [ + { + "bbox": [ + 130, + 300, + 481, + 321 + ], + "type": "text", + "content": "38. Schunk, D.H., Zimmerman, B.J.: Self-regulated learning: From teaching to self-reflective practice. Guilford Press (1998)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 321, + 481, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 321, + 481, + 354 + ], + "spans": [ + { + "bbox": [ + 130, + 321, + 481, + 354 + ], + "type": "text", + "content": "39. Schwenk, D., Khandelwal, A., Clark, C., Marino, K., Mottaghi, R.: A-okvqa: A benchmark for visual question answering using world knowledge. In: European Conference on Computer Vision. pp. 146-162. Springer (2022)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 354, + 481, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 354, + 481, + 386 + ], + "spans": [ + { + "bbox": [ + 130, + 354, + 481, + 386 + ], + "type": "text", + "content": "40. Sun, Z., Shen, S., Cao, S., Liu, H., Li, C., Shen, Y., Gan, C., Gui, L.Y., Wang, Y.X., Yang, Y., et al.: Aligning large multimodal models with factually augmented rlhf. arXiv preprint arXiv:2309.14525 (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 386, + 481, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 386, + 481, + 418 + ], + "spans": [ + { + "bbox": [ + 130, + 386, + 481, + 418 + ], + "type": "text", + "content": "41. Team, G., Anil, R., Borgeaud, S., Wu, Y., Alayrac, J.B., Yu, J., Soricut, R., Schalkwyk, J., Dai, A.M., Hauth, A., et al.: Gemini: a family of highly capable multimodal models. arXiv preprint arXiv:2312.11805 (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 418, + 481, + 451 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 418, + 481, + 451 + ], + "spans": [ + { + "bbox": [ + 130, + 418, + 481, + 451 + ], + "type": "text", + "content": "42. Touvron, H., Lavril, T., Izacard, G., Martinet, X., Lachaux, M.A., Lacroix, T., Rozière, B., Goyal, N., Hambro, E., Azhar, F., Rodriguez, A., Joulin, A., Grave, E., Lample, G.: Llama: Open and efficient foundation language models (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 450, + 481, + 483 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 450, + 481, + 483 + ], + "spans": [ + { + "bbox": [ + 130, + 450, + 481, + 483 + ], + "type": "text", + "content": "43. Touvron, H., Martin, L., Stone, K., Albert, P., Almahairi, A., Babaei, Y., Bashlykov, N., Batra, S., Bhargava, P., Bhosale, S., et al.: Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288 (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 483, + 481, + 515 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 483, + 481, + 515 + ], + "spans": [ + { + "bbox": [ + 130, + 483, + 481, + 515 + ], + "type": "text", + "content": "44. Wang, J., Meng, L., Weng, Z., He, B., Wu, Z., Jiang, Y.G.: To see is to believe: Prompting gpt-4v for better visual instruction tuning. arXiv preprint arXiv:2311.07574 (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 515, + 481, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 515, + 481, + 547 + ], + "spans": [ + { + "bbox": [ + 130, + 515, + 481, + 547 + ], + "type": "text", + "content": "45. Wang, X., Pan, J., Ding, L., Biemann, C.: Mitigating hallucinations in large vision-language models with instruction contrastive decoding. arXiv preprint arXiv:2403.18715 (2024)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 547, + 481, + 579 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 547, + 481, + 579 + ], + "spans": [ + { + "bbox": [ + 130, + 547, + 481, + 579 + ], + "type": "text", + "content": "46. Wang, X., Wei, J., Schuurmans, D., Le, Q., Chi, E., Narang, S., Chowdhery, A., Zhou, D.: Self-consistency improves chain of thought reasoning in language models. arXiv preprint arXiv:2203.11171 (2022)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 579, + 481, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 579, + 481, + 601 + ], + "spans": [ + { + "bbox": [ + 130, + 579, + 481, + 601 + ], + "type": "text", + "content": "47. Wang, X., Wei, J., Schuurmans, D., Le, Q., Chi, E., Zhou, D.: Rationale-augmented ensembles in language models. arXiv preprint arXiv:2207.00747 (2022)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 130, + 601, + 481, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 601, + 481, + 633 + ], + "spans": [ + { + "bbox": [ + 130, + 601, + 481, + 633 + ], + "type": "text", + "content": "48. Wei, J., Wang, X., Schuurmans, D., Bosma, M., Xia, F., Chi, E., Le, Q.V., Zhou, D., et al.: Chain-of-thought prompting elicits reasoning in large language models. Advances in Neural Information Processing Systems 35, 24824-24837 (2022)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 130, + 633, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 633, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 633, + 481, + 665 + ], + "type": "text", + "content": "49. Wu, Y., Zhang, P., Xiong, W., Oguz, B., Gee, J.C., Nie, Y.: The role of chain-of-thought in complex vision-language reasoning task. arXiv preprint arXiv:2311.09193 (2023)" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 328, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 328, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 328, + 91, + 447, + 102 + ], + "type": "text", + "content": "Reflective Instruction Tuning" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 610 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 480, + 149 + ], + "type": "text", + "content": "50. Ye, Q., Xu, H., Xu, G., Ye, J., Yan, M., Zhou, Y., Wang, J., Hu, A., Shi, P., Shi, Y., et al.: mplug-owl: Modularization empowers large language models with multimodality. arXiv preprint arXiv:2304.14178 (2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 150, + 480, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 150, + 480, + 183 + ], + "spans": [ + { + "bbox": [ + 132, + 150, + 480, + 183 + ], + "type": "text", + "content": "51. Yin, S., Fu, C., Zhao, S., Xu, T., Wang, H., Sui, D., Shen, Y., Li, K., Sun, X., Chen, E.: Woodpecker: Hallucination correction for multimodal large language models. arXiv preprint arXiv:2310.16045 (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 183, + 480, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 183, + 480, + 226 + ], + "spans": [ + { + "bbox": [ + 132, + 183, + 480, + 226 + ], + "type": "text", + "content": "52. Yin, Z., Wang, J., Cao, J., Shi, Z., Liu, D., Li, M., Huang, X., Wang, Z., Sheng, L., Bai, L., et al.: Lamm: Language-assisted multi-modal instruction-tuning dataset, framework, and benchmark. Advances in Neural Information Processing Systems 36 (2024)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 227, + 480, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 227, + 480, + 248 + ], + "spans": [ + { + "bbox": [ + 132, + 227, + 480, + 248 + ], + "type": "text", + "content": "53. Yu, Q., Li, J., Wei, L., Pang, L., Ye, W., Qin, B., Tang, S., Tian, Q., Zhuang, Y.: Hallucidoctor: Mitigating hallucinatory toxicity in visual instruction data (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 249, + 480, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 249, + 480, + 270 + ], + "spans": [ + { + "bbox": [ + 132, + 249, + 480, + 270 + ], + "type": "text", + "content": "54. Zelikman, E., Mu, J., Goodman, N.D., Wu, Y.T.: Star: Self-taught reasoner bootstrapping reasoning with reasoning (2022)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 271, + 480, + 313 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 271, + 480, + 313 + ], + "spans": [ + { + "bbox": [ + 132, + 271, + 480, + 313 + ], + "type": "text", + "content": "55. Zhai, B., Yang, S., Zhao, X., Xu, C., Shen, S., Zhao, D., Keutzer, K., Li, M., Yan, T., Fan, X.: Halle-switch: Rethinking and controlling object existence hallucinations in large vision language models for detailed caption. arXiv preprint arXiv:2310.01779 (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 315, + 480, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 315, + 480, + 369 + ], + "spans": [ + { + "bbox": [ + 132, + 315, + 480, + 369 + ], + "type": "text", + "content": "56. Zhang, P., Dong, X., Wang, B., Cao, Y., Xu, C., Ouyang, L., Zhao, Z., Ding, S., Zhang, S., Duan, H., Zhang, W., Yan, H., Zhang, X., Li, W., Li, J., Chen, K., He, C., Zhang, X., Qiao, Y., Lin, D., Wang, J.: Internl m-xcomposer: A vision-language large model for advanced text-image comprehension and composition. arXiv preprint arXiv:2309.15112 (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 369, + 480, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 369, + 480, + 402 + ], + "spans": [ + { + "bbox": [ + 132, + 369, + 480, + 402 + ], + "type": "text", + "content": "57. Zhang, S., Roller, S., Goyal, N., Artetxe, M., Chen, M., Chen, S., Dewan, C., Diab, M., Li, X., Lin, X.V., et al.: Opt: Open pre-trained transformer language models. arXiv preprint arXiv:2205.01068 (2022)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 403, + 480, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 403, + 480, + 435 + ], + "spans": [ + { + "bbox": [ + 132, + 403, + 480, + 435 + ], + "type": "text", + "content": "58. Zhang, Y., Zhang, R., Gu, J., Zhou, Y., Lipka, N., Yang, D., Sun, T.: Llavar: Enhanced visual instruction tuning for text-rich image understanding. arXiv preprint arXiv:2306.17107 (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 435, + 480, + 457 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 435, + 480, + 457 + ], + "spans": [ + { + "bbox": [ + 132, + 435, + 480, + 457 + ], + "type": "text", + "content": "59. Zhang, Z., Zhang, A., Li, M., Zhao, H., Karypis, G., Smola, A.: Multimodal chain-of-thought reasoning in language models. arXiv preprint arXiv:2302.00923 (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 457, + 480, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 457, + 480, + 479 + ], + "spans": [ + { + "bbox": [ + 132, + 457, + 480, + 479 + ], + "type": "text", + "content": "60. Zhao, B., Wu, B., Huang, T.: Svit: Scaling up visual instruction tuning. arXiv preprint arXiv:2307.04087 (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 479, + 480, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 479, + 480, + 511 + ], + "spans": [ + { + "bbox": [ + 132, + 479, + 480, + 511 + ], + "type": "text", + "content": "61. Zhao, L., Yu, E., Ge, Z., Yang, J., Wei, H., Zhou, H., Sun, J., Peng, Y., Dong, R., Han, C., et al.: Chatspot: Bootstrapping multimodal llms via precise referring instruction tuning. arXiv preprint arXiv:2307.09474 (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 512, + 480, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 512, + 480, + 544 + ], + "spans": [ + { + "bbox": [ + 132, + 512, + 480, + 544 + ], + "type": "text", + "content": "62. Zheng, L., Chiang, W.L., Sheng, Y., Zhuang, S., Wu, Z., Zhuang, Y., Lin, Z., Li, Z., Li, D., Xing, E.P., Zhang, H., Gonzalez, J.E., Stoica, I.: Judging lvm-as-a-judge with mt-bench and chatbot arena (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 545, + 480, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 545, + 480, + 578 + ], + "spans": [ + { + "bbox": [ + 132, + 545, + 480, + 578 + ], + "type": "text", + "content": "63. Zhu, D., Chen, J., Shen, X., Li, X., Elhoseiny, M.: Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592 (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 578, + 480, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 578, + 480, + 610 + ], + "spans": [ + { + "bbox": [ + 132, + 578, + 480, + 610 + ], + "type": "text", + "content": "64. Zong, Z., Ma, B., Shen, D., Song, G., Shao, H., Jiang, D., Li, H., Liu, Y.: Mova: Adapting mixture of vision experts to multimodal context. arXiv preprint arXiv:2404.13046 (2024)" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 227, + 102 + ], + "type": "text", + "content": "J. Zhang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Region-Adaptive Transform with Segmentation Prior for Image Compression/7ddd6f0f-062f-4fcd-9635-2169878bc97b_content_list.json b/2024/Region-Adaptive Transform with Segmentation Prior for Image Compression/7ddd6f0f-062f-4fcd-9635-2169878bc97b_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..05dd32ae1e8b5803107b38892a8a901282843463 --- /dev/null +++ b/2024/Region-Adaptive Transform with Segmentation Prior for Image Compression/7ddd6f0f-062f-4fcd-9635-2169878bc97b_content_list.json @@ -0,0 +1,1909 @@ +[ + { + "type": "text", + "text": "Region-Adaptive Transform with Segmentation Prior for Image Compression", + "text_level": 1, + "bbox": [ + 232, + 140, + 772, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yuxi Liu $^{1,2}$ , Wenhan Yang $^{3}$ , Huihui Bai $^{1,2}$ , Yunchao Wei $^{1,2,3}$ , and Yao Zhao $^{\\star 1,2,3}$", + "bbox": [ + 240, + 210, + 759, + 242 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Institute of Information Science, Beijing Jiaotong University, Beijing, China \n $^{2}$ Visual Intelligence + X International Joint Laboratory of the Ministry of Education \n $^{3}$ PengCheng Laboratory, Shenzhen, China \n{yuxiliu, hhbai, yunchao.wei, yzhao}@bjtu.edu.cn, yangwh@pcl.ac.cn", + "bbox": [ + 215, + 253, + 785, + 311 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Learned Image Compression (LIC) has shown remarkable progress in recent years. Existing works commonly employ CNN-based or Transformer-based modules as transform methods for compression. However, there is no prior research on neural transform that focuses on specific regions. In response, we introduce the class-agnostic segmentation masks (i.e. semantic masks without category labels) for extracting region-adaptive contextual information. Our proposed module, Region-Adaptive Transform, applies adaptive convolutions on different regions guided by the masks. Additionally, we introduce a plug-and-play module named Scale Affine Layer to incorporate rich contexts from various regions. While there have been prior image compression efforts that involve segmentation masks as additional intermediate inputs, our approach differs significantly from them. Our advantages lie in that, to avoid extra bitrate overhead, we treat these masks as privilege information, which is accessible during the model training stage but not required during the inference phase. To the best of our knowledge, we are the first to employ class-agnostic masks as privilege information and achieve superior performance in pixel-fidelity metrics, such as Peak Signal to Noise Ratio (PSNR). The experimental results demonstrate our improvement compared to previously well-performing methods, with about $8.2\\%$ bitrate saving compared to VTM-17.0. The source code is available at https://github.com/GityuxiLiu/SegPIC-for-Image-Compression.", + "bbox": [ + 261, + 348, + 738, + 654 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Learned Image Compression $\\cdot$ Segmentation Prior", + "bbox": [ + 261, + 667, + 683, + 681 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 708, + 375, + 724 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "With the advent of the multimedia era, an ever-growing volume of image data has emerged on the internet. These vast volumes of image data demand larger transmission bandwidth and storage capacity. In this context, the importance of image compression technology has become even more pronounced, evoking greater efforts in efficient image compression techniques. Lossy image compres", + "bbox": [ + 212, + 739, + 787, + 816 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Corresponding author", + "bbox": [ + 217, + 824, + 382, + 839 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/0e4b5631c4dae19c2369313434f77ba00505965d04a1f677608d5c2e2f704eaf.jpg", + "image_caption": [ + "(a) Convictional LIC framework" + ], + "image_footnote": [], + "bbox": [ + 338, + 143, + 666, + 213 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/c0c76c0d31e5f465ee91214d749e0aaf0cf49a4e22189e98f43822733d9c0928.jpg", + "image_caption": [ + "(b) Data-dependent LIC framework" + ], + "image_footnote": [], + "bbox": [ + 338, + 234, + 665, + 299 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/c10f29b94376026738b0c62e9af84326aafc84815ee25a2a130a861a6da5fdc6.jpg", + "image_caption": [ + "(c) Proposed LIC framework", + "Fig. 1: Paradigms of end-to-end learned image compression (LIC) methods. (a) Conventional LIC with fixed learned kernels. (b) Data-dependent LIC with image-specific kernels [43]. (c) The proposed method is capable of generating region-adaptive kernels from different regions, providing more fine-grained transforms." + ], + "image_footnote": [], + "bbox": [ + 339, + 319, + 663, + 407 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "sion technology can reduce the data size to a lower level while maintaining good visual quality. In the past decades, the classical lossy image compression methods such as JPEG [42], JPEG2000 [38], BPG [7], and VVC [24], used to have excellent performance and were widely adopted. These methods include modules of transform, quantization, and entropy coding. The transform process aims to convert images into a more compact coefficient by removing pixel correlations as much as possible. The commonly employed transforms include the discrete cosine transform (DCT) [3] and wavelet transform [14]. These transformations are linear, characterized by simplicity and efficiency. However, their fixed forms are unable to adapt to the contexts of complex signals. To address this issue, latter efforts propose data-driven transformation approaches, e.g. the Karhunen-Loève transform (KLT) [39]. Furthermore, the subsequent works are dedicated to constructing complex patterns for the adaptive selection of local transforms, improving adaptability and overall performance.", + "bbox": [ + 212, + 522, + 787, + 733 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Benfit from the advancement of neural network technology, many learned image compression (LIC) methods have been proposed, which shows a promising alternative for designing transform: learning from data instead of handcraft design. Balle et al. [5, 6] proposed the first well-known end-to-end framework. They utilize variational auto-encoders (VAE) [26] as a transform module and optimize the parameters of the kernels to better align with the natural data distribution. To further improve the representation capabilities of the transform", + "bbox": [ + 212, + 734, + 787, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "Y. Liu et al.", + "bbox": [ + 271, + 114, + 354, + 126 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "module, some works [12, 30, 43, 48, 49] design advanced modules to introduce richer contextual information, e.g. residual blocks [20] and self-attention mechanism [15, 32]. These works optimize learnable kernels to fit the data distribution to attain a more powerful transform. However, their kernel parameters are fixed after training (see Fig.1 (a)). Wang et al. [43] introduce a neural data-dependent transform, which extracts dynamic convolution kernels based on the context of the entire image for more adaptive transform (see Fig. 1 (b)). However, even within a single image, the context could vary significantly across different regions, e.g. stone and sheep regions in Fig. 1. It limits the description of the signals to a more fine-grained level using such image-level transforms.", + "bbox": [ + 212, + 146, + 787, + 297 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To address the issue, this paper makes the first efforts to design learnable region-adaptive transform to capture the mapping regularity in different regions. In detail, we employ class-agnostic4 segmentation masks to extract region-specific information for transform (see Fig. 1 (c)). Diffs from previous works [4, 11, 16, 21], our approach stands out with two characteristics in mask utilization. First, our masks are \"class-agnostic\", where the semantic labels are abandoned but a more effective compression-driven semantic prior will be learned by our model in a data-driven manner. Second, to avoid incurring additional costs in both bitrate and computational resources, we treat these masks as privilege information, which are available during training but remain inaccessible during the inference phase [29]. Specifically, we replace the masks with uniform grid partitions (see Fig. 9 (c)) in inference, and the experimental result shows approximate performance. It implies that our model has learned useful semantic knowledge from the masks in training, incorporating this knowledge into the model (i.e. the network's parameters). Hence, it remains capable of mining contextual information even from relatively coarse partition regions, i.e., the given uniform grid partitions. This simple operation can effectively maintain the consistency of our model's architecture and parameters, whether or not masks are used at different phases. The more details about the difference between our method and previous semantic-mask-based works will be explained at Sec. 2.", + "bbox": [ + 212, + 297, + 787, + 599 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Based on the above motivations, we propose a novel image compression framework called Segmentation-Prior-Guided Image Compression (SegPIC), equipped with two key modules: Region-Adaptive Transform (RAT) and Scale Affine Layer (SAL). The RAT adopts region-adaptive kernels for transform, performing the convolution whose weights are generated from the prototypes and the context of latent features under the guidance of masks. The SAL affines the latent features for better mining of semantic contexts in the Encoder and Decoder. Experimental results show that SegPIC is superior to previously well-performing image compression methods. Our main contributions are as follows:", + "bbox": [ + 212, + 599, + 787, + 736 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "- We propose a novel image compression framework called Segmentation-Prior-Guided Image Compression (SegPIC) to generate more effective region-adaptive transforms in a data-driven manner.", + "bbox": [ + 225, + 744, + 787, + 789 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "Region-Adaptive Transform with Segmentation Prior for Image Compression", + "bbox": [ + 218, + 114, + 730, + 130 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "page_footnote", + "text": "${}^{4}$ \"class-agnostic\" means distinguishing different objects in an image regardless of what those objects are. Such masks can be regarded as a generalized representation of partition maps widely used in conventional CODECs.", + "bbox": [ + 220, + 796, + 787, + 839 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Our SegPIC is equipped with two proposed modules: Region-Adaptive Transform (RAT) and Scale Affine Layer (SAL). The RAT adopts region-adaptive kernels for transform, while the SAL affines the latent features for better mining of semantic contexts.", + "- We utilize the class-agnostic masks as privilege information to assist the training of SegPIC. The \"class-agnostic\" property promotes the model to learn a compression-driven semantic prior, while the \"privilege\" way avoids additional bitrate and complexity for masks in inference.", + "- Experimental results demonstrate that our SegPIC is superior to previously well-performing image compression methods. We also present sufficient ablation experiments to confirm the necessity of the privilege information." + ], + "bbox": [ + 223, + 146, + 784, + 311 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2 Related Works", + "text_level": 1, + "bbox": [ + 215, + 337, + 395, + 353 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Learned Image Compression. Image compression can be divided into three main components: transform, quantization, and entropy coding. In brief, the transformation module maps the image to a latent feature space. Subsequently, quantization is performed after this phase, resulting in some quantization distortions but achieving lower bitrate. The entropy encoding module estimates the data distribution of the latent features and inputs it to the Arithmetic Coding (AC) module, which generates the bitstream for transmission.", + "bbox": [ + 212, + 371, + 782, + 474 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To enhance transform module, some works [30,43,48,49] adopt more powerful modules like self-attention mechanism [32, 41] to explore broader context. [31] proposed a transform block that mixes CNN and Transformer to combine their advantages. As for the entropy estimation module, the studies [12, 19, 33, 34] usually focus on exploring novel auto-regressive methods to split the latent features into several portions to be decoded in order. The portions having been decoded will assist in decoding the next portion, through the spatial or channel correlation of these portions. This auto-regressive mechanism may introduce additional latency but achieve better performance than Hyper Prior [6]. Recently, [23] integrates different types of correlations present in latent representation, e.g. channel-wise, local spatial, and global spatial correlations. Their proposed novel auto-regressive entropy module achieve powerful performance.", + "bbox": [ + 212, + 477, + 782, + 656 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Thanks to these great works, the performance of LIC has become so strong that achieving further enhancements becomes progressively challenging. Therefore, some new directions in compression have emerged, which is also meaningful on certain occasions. For example, some semantic-based compression methods [10, 11] adopt GAN [17] to compress images at extremely low bitrate. These methods maintain a good human visual perception effect but usually, poor pixel-fidelity metrics, e.g. Peak Signal to Noise Ratio (PSNR) and Multi-Scale Structural Similarity Index (MS-SSIM) [46]. Specifically, these methods could introduce some synthetic artifacts with low fidelity. For example, a semantic compression method could reconstruct an image of a bird with different breeds from the original image because the training category labels do not distinguish the bird's breed. Another direction is compression for machines, which means compress", + "bbox": [ + 212, + 659, + 784, + 839 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "Y. Liu et al.", + "bbox": [ + 271, + 114, + 354, + 127 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/9cc7ffa734bebf9d81b607c1f20b9fe6ce6033197cf401b6541bbec60fa0f014.jpg", + "image_caption": [ + "(a) The overall framework" + ], + "image_footnote": [], + "bbox": [ + 267, + 145, + 733, + 333 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/17cd25dd45547abe1efde816681db3f4bf91e02afb712ae2719c419e87bd3f0b.jpg", + "image_caption": [ + "(b)Downsample block (out channel C)" + ], + "image_footnote": [], + "bbox": [ + 256, + 369, + 493, + 454 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/5bf1028832daf3fd0fbde7ec4964a20e6844baeb5d5bff68d20b0470d47fadfd.jpg", + "image_caption": [ + "(c) Upsample block (out channel C)", + "Fig. 2: (a) Overall framework of our SegPIC. The blue flow lines represent the decoding process on the decoder side. The dashed lines represent the process of extracting the prototypes $p'$ , $p$ (see Eq. 5). The compression of $p'$ , $p$ is not fully depicted for simplicity, please refer to Fig. 4 for details. $N$ is set to 192, and $M$ is 320 as in [49]. RAT is the proposed Region-Adaptive Transform. SAL is the proposed Scale Affine Layer. WAM is Window Attention Module [49]. ChARM is Channel-wise Auto-Regressive Model [34]. FM is a Factorized Model [6]. GDN is Generalized Divisive Normalization [5]. (b) is Downsample Block. \"Downsample, N\" in (a) means the block with out-channel N. \"Conv C,5,2\" means out-channel C, kernel size 5, stride 2. (c) is Upsample Block, and \"TConv\" is Transposed Convolution." + ], + "image_footnote": [], + "bbox": [ + 508, + 369, + 746, + 453 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "ing images for downstream tasks [16, 40]. These methods focus on balancing the bitrate and performance of downstream tasks, not just the pixel-fidelity metrics.", + "bbox": [ + 212, + 656, + 785, + 686 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Although some works [10, 11, 16] about the two directions also adopt masks, our usage is significantly different. The main differences are as follows: 1) We utilize the masks as privilege information, without the mask information during the inference phase, while usually, the masks are indispensable for other methods. 2) We employ class-agnostic masks to guide the model to extract rich compression-friendly contexts among semantically related regions, regardless of the category that might degrade the fidelity as mentioned above. 3) The proposed model is optimized towards better pixel-fidelity metrics, e.g. PSNR and MS-SSIM.", + "bbox": [ + 212, + 688, + 787, + 808 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Region-adaptive Convolution. Due to its capacity to model the degradation levels in different regions of images, region-adaptive convolution is applied to", + "bbox": [ + 212, + 809, + 787, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "Region-Adaptive Transform with Segmentation Prior for Image Compression", + "bbox": [ + 220, + 114, + 730, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "some low-level vision tasks such as super-resolution and denoising [37, 47]. In terms of compression, the reconstruction quality varies across different regions. For instance, regions with complex textures are more challenging compared to those with simpler patterns. Therefore, specific processing can be applied to different regions based on their characteristics. The regions can be divided using segmentation masks or partitions based on certain regular patterns. To the best of our knowledge, we are the first to introduce this mechanism on image compression for optimizing pixel-fidelity metrics.", + "bbox": [ + 212, + 146, + 787, + 268 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3 Proposed SegPIC", + "text_level": 1, + "bbox": [ + 214, + 289, + 423, + 306 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.1 Motivation", + "text_level": 1, + "bbox": [ + 214, + 320, + 354, + 335 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "As mentioned above, previous transforms either adopt convolutions with fixed parameters or image-specific convolutions. Our motivation is to construct a region-adaptive transform via the following three properties:", + "bbox": [ + 212, + 344, + 785, + 390 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Region-dependent: we propose a neural network with region-adaptive transforms guided by the segmentation masks.", + "- Class-Agnostic: the adopted mask does not include a category label, allowing the model to be free to learn compression-friendly semantic priors.", + "- Privilege Information: the masks are expected as privilege information, not being used in inference for saving bitrate and computation resources." + ], + "bbox": [ + 223, + 398, + 784, + 487 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Based on these considerations, we propose our SegPIC in the VAE framework incorporated with two proposed modules: SAL and RAT.", + "bbox": [ + 214, + 497, + 785, + 527 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "- SAL: First of all, we extract a prototype from each region, which will be compressed and transmitted to be accessible on both the encoder and decoder sides. The prototypes are expected to extract high-level information from regions to ensure the representative ability as much as possible. However, the encoder in compression frameworks often extracts features that contain low-level information. Therefore, a module is highly desirable for communicating information across different levels. SAL is designed for extracting more abundant semantics and contexts, enriching the prototypes, and connecting features from two different levels.", + "bbox": [ + 223, + 537, + 785, + 672 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "- RAT: performing the Depth&Point-wise Separable Convolution (DPSCov), a novel convolution whose weights are generated from the prototypes and the context of latent features under the guidance of masks. Note that, we only utilize the masks to guide RAT to learn more compression-friendly semantic priors in training, while during the inference phase, we instead adopt uniform grid partitions as coarse alternatives.", + "bbox": [ + 223, + 672, + 785, + 763 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.2 Formulation", + "text_level": 1, + "bbox": [ + 214, + 785, + 362, + 800 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this section, we will introduce the problem and our method in mathematical form. Given an image $x$ , the encoder $En$ maps it to latent features $y$ . After", + "bbox": [ + 212, + 809, + 785, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "Y. Liu et al.", + "bbox": [ + 271, + 114, + 354, + 127 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "quantization $Q$ , $\\hat{y}$ will be compressed to the bitstream by the entropy coding, such as arithmetic coding (AC). Then the receiver converts the bitstream back into $\\hat{y}$ by entropy decoding. The above process can be formulated as follows:", + "bbox": [ + 212, + 146, + 782, + 191 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\ny = E n (x), \\quad \\hat {y} = Q (y),\n$$\n", + "text_format": "latex", + "bbox": [ + 387, + 202, + 557, + 218 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Transmitter: $bit_{\\hat{y}} = AE(\\hat{y} |p_{\\hat{y}})$ (1)", + "bbox": [ + 387, + 220, + 784, + 237 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Receiver: $\\hat{y} = AD(bit_{\\hat{y}}|p_{\\hat{y}})$", + "bbox": [ + 387, + 239, + 614, + 256 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $p_{\\hat{y}}$ is the estimated distribution of $\\hat{y}$ , $\\text{bit}_{\\hat{y}}$ is the binary bitstream of $\\hat{y}$ , and $AE$ and $AD$ are arithmetic encoding and decoding. We use a single Gaussian distribution to estimate $\\hat{y} \\sim \\mathcal{N}(\\mu, \\sigma^2)$ , and the mean and variance value $\\mu, \\sigma^2$ will be yielded through an auto-regressive mechanism. The formulation is:", + "bbox": [ + 214, + 265, + 784, + 325 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\nz = E n _ {z} (y), \\hat {z} = Q (z), \\quad \\mu_ {i}, \\sigma_ {i} = F \\left(D e _ {z} (\\hat {z}), \\hat {y} _ {< i}\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 325, + 335, + 784, + 353 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $En_{z}, De_{z}$ is hyper encoder and decoder, $\\mu_{i}, \\sigma_{i}$ are the mean and variance of $\\hat{y}_{i} \\sim \\mathcal{N}(\\mu_{i}, \\sigma_{i}^{2})$ , $\\hat{y}_{i}$ is the currently decoded portion of $\\hat{y}$ , $\\hat{y}_{< i}$ is the already decoded portion to assist decoding $\\hat{y}_{i}$ , and $F$ is the auto-regressive entropy model. The AC process of $z$ in Eq. 2 is omitted for brevity. Finally, the receiver adopts decoder $De$ to reconstruct the image $\\hat{x}$ , which is formulated as $\\hat{x} = De(\\hat{y})$ .", + "bbox": [ + 214, + 362, + 784, + 452 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The loss function $\\mathcal{L}$ to optimize this problem is:", + "bbox": [ + 238, + 454, + 584, + 468 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = R + \\lambda \\cdot D = R (\\hat {y}) + R (\\hat {z}) + \\lambda \\cdot \\mathbb {E} _ {x \\sim p _ {x}} [ d (x, \\hat {x}) ], \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 318, + 479, + 784, + 496 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $R(\\hat{y}), R(\\hat{z})$ are the bitrate of $\\hat{y}, \\hat{z}$ , $d(x, \\hat{x})$ is the distortion between $x$ and $\\hat{x}$ , $\\lambda$ denotes different rate-distortion trade-offs. Here we use Mean Square Error (MSE) or (1-MS-SSIM) as the distortion function.", + "bbox": [ + 214, + 505, + 782, + 549 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In our implementation, we use class-agnostic semantic masks (denoted as $m$ ) as additional inputs of the encoding $En$ during the training phase. Moreover, we extract a prototype from each region at the middle layer of $En$ , then compress and transmit them (denoted as $\\hat{p}$ ). In inference, we replace the masks with uniform grid partitions, so there is no additional bitstream for masks, while the bitrates of the prototypes also need to be taken into account. The $En$ and $De$ will transform the image with the assistance of the prototypes. The formulations requiring adjustment are as follows:", + "bbox": [ + 212, + 551, + 784, + 672 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\ny = E n (x, \\hat {p} | m), \\quad \\hat {x} = D e (\\hat {y}, \\hat {p} | m),\n$$\n", + "text_format": "latex", + "bbox": [ + 377, + 681, + 625, + 698 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mu_ {i}, \\sigma_ {i} = F (D e _ {z} (\\hat {z}), \\hat {y} _ {< i}, \\hat {p} | m), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 377, + 700, + 784, + 717 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\nR = R (\\hat {p}) + R (\\hat {y}) + R (\\hat {z}).\n$$\n", + "text_format": "latex", + "bbox": [ + 377, + 719, + 558, + 734 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.3 Network Design", + "text_level": 1, + "bbox": [ + 215, + 755, + 393, + 770 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this section, we will introduce the concrete process and our network design. We use [49] as the backbone. The overall architecture is in Fig. 2.", + "bbox": [ + 212, + 779, + 782, + 809 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Scale Affine Layer. The Encoder and Decoder are responsible for image transform and reconstruction, each comprising 4 downsampling or upsampling blocks.", + "bbox": [ + 212, + 810, + 782, + 839 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "Region-Adaptive Transform with Segmentation Prior for Image Compression", + "bbox": [ + 218, + 114, + 732, + 130 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The previous sampling blocks consisted of a single sample convolution layer, aimed at modifying the data size. However, this may not be sufficient to extract adequate semantics and contexts. Inspired by Spatial Feature Transform (SFT) [45], which can generate visually pleasing textures in segmentation-guided low-level vision, we propose to add a Scale Affine Layer (SAL) between two adjacent convolution layers (see Fig. 2 (b) (c)). It consists of two $1 \\times 1$ convolution layers with a GELU activation function in between, which is formulated as:", + "bbox": [ + 215, + 146, + 787, + 252 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} X _ {s} = \\operatorname {S A L} _ {s} (X), \\quad X _ {b} = \\operatorname {S A L} _ {b} (X), \\\\ Y = X _ {s} \\otimes X + X _ {b}, \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 375, + 260, + 625, + 296 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $X$ and $Y$ represent the input and output, $X_{s}$ and $X_{b}$ denote the scales and biases generated by SAL, $\\otimes$ denotes the element-wise product.", + "bbox": [ + 212, + 304, + 785, + 333 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Extract and Transmit the Prototypes. The Encoder first converts the original image $x \\in \\mathbb{R}^{3 \\times H \\times W}$ into latent features $y' \\in \\mathbb{R}^{N \\times H/4 \\times W/4}$ at the middle layer. Then we conduct Masked Average Pooling (MAP) on $y'$ through the class-agnostic masks $m \\in \\{0,1\\}^{n \\times H \\times W}$ (see Fig. 3 (a)), to acquire the prototypes $p' \\in \\mathbb{R}^{N \\times n}$ , where $n$ is the number of the masks. Thanks to the proposed SAL, the Encoder becomes deeper, allowing $p'$ to incorporate more semantics and contexts from the corresponding mask. The formulation of MAP is as follows:", + "bbox": [ + 212, + 334, + 787, + 440 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\np ^ {\\prime} = \\operatorname {M A P} \\left(y ^ {\\prime}, m\\right) = \\left\\{\\frac {1}{\\sum m ^ {(i)}} \\sum y ^ {\\prime} \\otimes m ^ {(i)}, \\dots \\right\\} _ {i \\in [ 1, n ]}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 310, + 446, + 785, + 479 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $y'$ are the mid-level features (see Fig. 2 (a)), $\\sum$ denotes summation along the spatial dimension. $p'$ will be compressed and transmitted to the receiver (see Fig. 4), and the reconstructed $\\hat{p}'$ is shared between the Encoder and Decoder. The channel $\\{C_1, C_2, C_3, C_4\\}$ is {192, 128, 96, 96} for $p'$ (on Encoder and Decoder) and {320, 192, 128, 96} for $p$ (on Entropy Model).", + "bbox": [ + 212, + 486, + 787, + 563 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Region-Adaptive Transform. Here we propose a novel convolution called Depth&Point-wise Separable Convolution (DPSCov), in which each element has an individual convolution kernel, and there is no channel-wise mutual communication for computational efficiency.", + "bbox": [ + 212, + 563, + 787, + 622 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Before introducing it, we will start with standard convolution briefly. The input of standard convolution can be denoted as $X \\in \\mathbb{R}^{C \\times H \\times W}$ , the output is $Y \\in \\mathbb{R}^{O \\times H \\times W}$ , and the kernel is $W \\in \\mathbb{R}^{O \\times C \\times k \\times k}$ , where $k$ is kernel size. The formulation of standard convolution is as follows:", + "bbox": [ + 212, + 623, + 787, + 683 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\nY _ {o, h, w} = \\sum_ {c = 1} ^ {C} X _ {c, h, w} * W _ {o, c} = \\sum_ {c = 1} ^ {C} \\sum_ {i, j} X _ {c, h + i, w + j} \\cdot W _ {o, c, i, j},\n$$\n", + "text_format": "latex", + "bbox": [ + 297, + 689, + 705, + 732 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $*$ is 2D convolution, $i,j\\in [-\\lfloor k / 2\\rfloor ,\\lfloor k / 2\\rfloor ]$ , $k$ is kernel size, and $\\lfloor \\cdot \\rfloor$ denotes floor rounding.", + "bbox": [ + 212, + 739, + 785, + 770 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "DPSCov can be seen as a combination of Depthwise Separable Convolution (DWConv) [13] and Dynamic Convolution [22]. The formulation is as follows:", + "bbox": [ + 214, + 771, + 785, + 801 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\nY _ {c, h, w} = X _ {c, h, w} * W _ {c, h, w} = \\sum_ {i, j} X _ {c, h + i, w + j} \\cdot W _ {c, h, w, i, j}. \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 308, + 809, + 785, + 844 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "Y. Liu et al.", + "bbox": [ + 271, + 114, + 354, + 127 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/8cee4801ee1779f81fd5971f4525b7ddbd395b93ef9a6ae76d5ce8a6a4fef5d5.jpg", + "image_caption": [ + "(a) Extract the prototypes through Masked Average Pooling" + ], + "image_footnote": [], + "bbox": [ + 217, + 141, + 480, + 292 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/5e12f059022ed5e0379e315379cb45c246c1bec341489f0cee517a70babe7e81.jpg", + "image_caption": [ + "(b)The Region-Adaptive Transform (RAT) with masks and prototypes", + "Fig. 3: The diagram of extracting the prototypes and the proposed RAT module. DPSCov is proposed Depth&Point-wise Separable Convolution (see Eq. 6). CTL is Channel Transform Layers, DKG is the DPSCov Kernel Generator, and CAG is the Channel Attention Generator. DPSCov is Depth&Point-wise Separable Convolution. The detailed modules are represented in Fig. 5." + ], + "image_footnote": [], + "bbox": [ + 501, + 142, + 777, + 292 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "The computational complexity of standard convolution is $\\mathrm{O}(O \\times C \\times H \\times W \\times k^2)$ , while DPSCov is $\\mathrm{O}(C \\times H \\times W \\times k^2)$ , same as DWConv.", + "bbox": [ + 212, + 434, + 784, + 465 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "The overall design of RAT is depicted in Fig. 3 (b). We expand the prototypes based on the positional information of the masks, filling in the corresponding copies of the prototypes at each pixel position. Then, we feed-forward the input through Channel Transform Layers (CTL) and concatenate it with the expanded features. The resulting variable, which contains both semantics and contexts, is used for subsequent filter generation. The DPSCConv Kernels Generator (DKG) is designed for generating the filters. In the first layer of DKG, we employ a 1x1 convolution to attempt the fusion of the prototype and the pixel's intrinsic attribute. We opt for grouped convolution in the final layer of the DKG to reduce the computation and parameters. Then Channel Attention Generator (CAG) generates the channel attention to be multiplied with the output of DPSCConv along the channel dimension. Finally, the output passes through a 1x1 convolution to merge channel information. We apply the RAT in both the Encoder and Decoder, as well as the entropy encoding module.", + "bbox": [ + 212, + 470, + 787, + 683 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Entropy Encoding Module. We use the Channel-wise Auto-Regressive Model (ChARM) [34] as our entropy estimation module, and the specific hyperparameter configuration follows [49]. In this model, the latent variable $y$ is partitioned into 10 groups along the channel dimension. The Gaussian prior distribution $p(y_{i}|\\hat{z},\\hat{y} < i)$ is estimated in an auto-regressive manner, where the mean and scale of $y_{i}$ depend on the quantized latent in the preceding groups $\\hat{y}_{< i}$ (see Eq. 2). To further enhance the estimation ability, we also deploy RAT in it. The RAT here can transmit key region prototypes through a separate bitstream, assisting in more efficient encoding and decoding. The prototypes $p$ could be seen as segmentation-prior, a supplement of the hyper-prior information $z$ .", + "bbox": [ + 212, + 688, + 787, + 840 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "Region-Adaptive Transform with Segmentation Prior for Image Compression", + "bbox": [ + 220, + 114, + 730, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/eb4a3ab36c142d7645b2c53cf5111bf45853c23d9711ea164bb525b232e7a4a5.jpg", + "image_caption": [ + "Fig.4: The architecture of the Prototype Encoder and Decoder for compressing and transmitting the prototypes. \"Linear $\\mathrm{C_1,C_2}$ \" means linear layer with in-channel $\\mathrm{C_1}$ and out-channel $\\mathrm{C_2}$ ." + ], + "image_footnote": [], + "bbox": [ + 217, + 146, + 444, + 276 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/21a49e6f4ff40e1fe858fa64914610b261b5d57059e6895d3c2eeb78266c0394.jpg", + "image_caption": [ + "Fig. 5: The detailed modules in RAT, including Channel Transform layers (CTL), DPSCConv Kernels Generator (DKG), and Channel Attention Generator (CAG). LReLU is Leaky ReLU. GAP is Global Average Pooling. \"GConv k²C,3,1,C\" means Grouped Convolution with out-channel k²C, kernel size 3, stride 1 and group C." + ], + "image_footnote": [], + "bbox": [ + 496, + 145, + 754, + 262 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/c19946c3964124eeb05f949fa7c9eb96ed02ebe093abb87c070906ff3ed2f767.jpg", + "table_caption": [ + "Table 1: BD-rate(%) performance on Kodak and CLIC Professional Valid (CLIC for short). The anchor is VTM-17.0." + ], + "table_footnote": [], + "table_body": "
ModelKodak \nBD-rate↓CLIC \nBD-rate↓ModelKodak \nBD-rate↓CLIC \nBD-rate↓
BPG [7]+31.5+48.1EVC [44]+0.27-
Cheng2020 [12]+7.46+11.7Feng2023 [16]-2.21-
Agustsson2023 [2]+6.15-ELIC [18]-3.08-3.58
ChARM [34]+3.58-Contextformer [28]-5.38-
Wang2022 [43]+1.10-0.99Pan2022 [35]-5.42-
BL-S+15.5+16.6Zou2022 (BL) [49]-2.89+0.14
Ours-S+3.87+5.26Ours-8.18-8.26
", + "bbox": [ + 228, + 425, + 769, + 575 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 215, + 601, + 375, + 619 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.1 Experimental Setup", + "text_level": 1, + "bbox": [ + 215, + 633, + 428, + 648 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Training. We use COCO-Stuff [9] as the training dataset, which is composed of 118k images with panoptic semantic masks. We just use the class-agnostic masks without the category labels. We randomly crop the images with the size of $256 \\times 256$ . The models are trained for 270 epochs (about 1M iterations) using the Adam optimizer [25] with a batch size of 32. The initial learning rate is set to $1 \\times 10^{-4}$ for 170 epochs, and reduced by half every 30 epochs. Our models are optimized by the rate-distortion trade-off loss function (see Eq. 3 and Eq. 4). The $\\lambda$ belongs to $\\{0.0018, 0.0035, 0.0067, 0.0130, 0.0250, 0.0483\\}$ for MSE as the distortion, and $\\{2.4, 4.58, 8.73, 16.64, 31.73, 60.50\\}$ for MS-SSIM [46].", + "bbox": [ + 212, + 657, + 787, + 794 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Evaluation. We evaluate our model on two commonly used datasets, Kodak [27] with the size of $768 \\times 512$ , and CLIC Professional Validation [1] with the size of 2K. Performance is evaluated using both bitrate and distortions. Distortion", + "bbox": [ + 212, + 795, + 787, + 839 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "Y. Liu et al.", + "bbox": [ + 271, + 114, + 354, + 126 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/3fab81b18d85b9af2da515dccb4e08f471e002a5c5255227512be96c01cd4a43.jpg", + "image_caption": [ + "Fig.6: Rate-Distortion Performance on Kodak [27] and CLIC Professional Valid dataset [1]. Our SegPIC introduces proposed RAT and SAL based on Zou2022 [49]. We also implement a simple baseline model without self-attention and auto-regressive mechanism (BL-S), and incorporate our proposed modules on it (Ours-S)." + ], + "image_footnote": [], + "bbox": [ + 215, + 143, + 500, + 314 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/b15adf7dcb86af8218c86e4850e35b8412a5f81623f3c28950b6ab6d1ba4dc6a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 504, + 142, + 787, + 314 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "measurement involves PSNR and MS-SSIM. And bitrates are assessed in bits per pixel (bpp). In this phase, we replace the masks with $4 \\times 4$ grid partitions. Thus, there is no bitrate cost related to the masks, while the bitrate of prototypes $R(p)$ and $R(p^{\\prime})$ is still taken into account.", + "bbox": [ + 212, + 407, + 787, + 470 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4.2 Rate-Distortion Performance", + "text_level": 1, + "bbox": [ + 215, + 488, + 503, + 503 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "We compare our SegPIC with previously well-performing methods, including discretized Gaussian mixture likelihoods and attention modules [12] (denoted as Cheng2020), ChARM [34] (denoted as Minnen2022), ChARM with WAM [49] (denoted as Zou2022), image-specific data-dependent transform model [43] (denoted as Wang2022), the best-performing traditional method VTM-17.0 [24] and influential traditional method BPG [7]. Furthermore, we introduce two simplified models without self-attention and auto-regressive mechanism, denoted as baseline-small (BL-S) and the one incorporating our proposed SAL and RAT (Ours-S), to demonstrate the generality of our modules. Details about the architecture of these two models will be provided in the appendix.", + "bbox": [ + 212, + 510, + 787, + 661 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "As illustrated in Fig. 6 and Tab. 1, compared to VTM-17.0, our SegPIC could save $8.18\\%$ BD-rate [8] on Kodak and $8.26\\%$ on CLIC, which outperforms all the LIC and traditional methods mentioned above. To further verify the efficiency of the proposed RAT and SAL, We also compare SegPIC with the baseline Zou2022, which can save $5.45\\%$ BD-rate on Kodak, and $8.50\\%$ on CLIC (see Tab. 2). Moreover, Ours-S even achieves closely matched performance compared to ChARM. The MS-SSIM results are provided in the appendix.", + "bbox": [ + 212, + 661, + 787, + 768 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4.3 Ablation Study", + "text_level": 1, + "bbox": [ + 215, + 787, + 390, + 801 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "To assess the contributions of our RAT and SAL modules individually, we independently incorporate these two modules into the baseline Zou2022. Conse", + "bbox": [ + 212, + 809, + 785, + 840 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "Region-Adaptive Transform with Segmentation Prior for Image Compression", + "bbox": [ + 220, + 114, + 730, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/037fcdb385ee81d514b5e0afb34bc99719581bf3e0cfe5d0e1c0893b5a3de7ac.jpg", + "image_caption": [ + "Original kodim04" + ], + "image_footnote": [], + "bbox": [ + 218, + 143, + 328, + 271 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/88f8f732716dcc89c4bd9a52b6ca000370f7b794d55f65ec43922b84eedd3817.jpg", + "image_caption": [ + "Ground Truth" + ], + "image_footnote": [], + "bbox": [ + 334, + 143, + 411, + 203 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/3a515b43679d80026bbc295f4a634476d04ff5438809fd4c1800edac89ff2fa9.jpg", + "image_caption": [ + "VTM (31.02/0.115)" + ], + "image_footnote": [], + "bbox": [ + 418, + 145, + 493, + 202 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/957f2fa822d29f159d60a8f265647b5825cc3fcc7e1b48cf532de5f76e37520b.jpg", + "image_caption": [ + "Zou2022 (31.46/0.119)" + ], + "image_footnote": [], + "bbox": [ + 336, + 212, + 411, + 271 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/ed79477c33635b9b68f99243f22619e22f37dd47832808243f2fef3f99acbf92.jpg", + "image_caption": [ + "Ours (31.53/0.109)" + ], + "image_footnote": [], + "bbox": [ + 418, + 212, + 493, + 271 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/93a90fa3b48708c694116e52b568ce4ba6ab0125dbd855173dd89607b9d29914.jpg", + "image_caption": [ + "Original kodim24", + "Fig. 7: Visualization of the reconstructed images $\\text{kodim}04$ and $\\text{kodim}24$ in Kodak. The metrics are (PNSR↑/bpp↓). It shows that our SegPIC can distinguish the objects' contours more accurately, making the edges sharper with less bitrate." + ], + "image_footnote": [], + "bbox": [ + 506, + 145, + 617, + 272 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/a22d4a9b97a192cefbba941436eb88f77f6ae650adfa4d91dded3d1219e1d562.jpg", + "image_caption": [ + "Ground Truth" + ], + "image_footnote": [], + "bbox": [ + 624, + 145, + 700, + 203 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/d5f179f9575f7e2ab0b96a0ac72120bd471c4b7534ff81294554c0e61edc9d3f.jpg", + "image_caption": [ + "VTM (28.96/0.325)" + ], + "image_footnote": [], + "bbox": [ + 707, + 145, + 782, + 203 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/2f291c2d0f6b37ee88a44033e03a0064a7d949295b200a1a320a86779e756397.jpg", + "image_caption": [ + "Zou2022 (28.67/0.292)" + ], + "image_footnote": [], + "bbox": [ + 624, + 212, + 700, + 271 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/c750ea1eaaf842512ddaed095f47d8cda2a9668fb97b1a22b0015b3eb01db6c1.jpg", + "image_caption": [ + "Ours (28.98/0.289)" + ], + "image_footnote": [], + "bbox": [ + 707, + 212, + 782, + 271 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/a6686ae11827cfdd13446e77efab1a4aaad8be0515fcf9c851d80293a0b3d924.jpg", + "table_caption": [ + "Table 2: Ablation Study on Kodak and CLIC. The Baseline (BL) is Zou2022 [49]. The speed is measured on Kodak with one Nvidia RTX 3090 GPU." + ], + "table_footnote": [], + "table_body": "
ModelKodakCLIC
ParamsBD-rateBD-rateEncDec Time
Zou2022 (BL)75.2 M0.00.0120 ms124 ms
BL w/ SAL76.7 M-2.12-2.93--
BL w/ RAT82.0 M-2.75-4.35--
Ours83.5 M-5.45-8.50142 ms130 ms
ChARM48.8 M+6.52-107 ms108 ms
BL-S21.7 M+19.0+16.3--
Ours-S29.2 M+6.96+5.1199 ms46 ms
", + "bbox": [ + 282, + 410, + 715, + 551 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "sequently, we construct two models: \"baseline w/ SAL\" and \"baseline w/ RAT.\" As illustrated in Tab. 2, the parameter counts introduced by RAT and SAL separately are $1.5\\mathrm{M}$ and $5.8\\mathrm{M}$ , which occupy only a small fraction of the total parameter count, about $1.7\\%$ and $6.9\\%$ . And they can contribute to BD-rate saving by $2.93\\%$ and $4.35\\%$ on CLIC Professional Valid, respectively. It shows that the proposed RAT and SAL are lightweight and efficient. When we combine these two modules, SAL will provide stronger prototypes for RAT, leading to better results, i.e., a saving of $8.50\\%$ BD-rate.", + "bbox": [ + 212, + 602, + 787, + 724 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "With our proposed modules, Ours-S achieves high performance with a significantly smaller parameter count. Compared to ChARM [34], Ours-S saves nearly $40\\%$ parameters and has faster speed while achieving comparable performance. Ours-S is a CNN-based model without any self-attention module. And the entropy estimation is just the hyperprior-style [6], without the efficient but time-consuming auto-regressive architecture. It indicates that the proposed modules RAT and SAL are highly effective and lightweight.", + "bbox": [ + 212, + 733, + 787, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "Y. Liu et al.", + "bbox": [ + 271, + 114, + 354, + 126 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/71dfb814646e3399dad66a4939b48e0bdf3255c1113c3f9db5f5a7e82a7f7870.jpg", + "image_caption": [ + "(a) Kodak" + ], + "image_footnote": [], + "bbox": [ + 217, + 143, + 408, + 246 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/b44d25f77fba26cfab222d0247693025870a3fe5e2c0e9dde208324f9f6ce3f6.jpg", + "image_caption": [ + "(b) COCO-sample", + "Fig. 8: We measure the impact of the masks separately in both the training and testing phases. \"test w/ masks\" means we adopt the masks in both training and testing, \"test w/o masks\" means we adopt the masks only in training, while \"train w/o masks\" means we do not use masks all the time. Naturally, if the masks are not used, $4 \\times 4$ grid partitions will be used as alternatives." + ], + "image_footnote": [], + "bbox": [ + 411, + 145, + 596, + 247 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/ecafbcf43bd572a8595e8556218b96639595e10c0dae1ebbb538e7dbb1685a46.jpg", + "image_caption": [ + "(c) Validation loss in training" + ], + "image_footnote": [], + "bbox": [ + 599, + 145, + 781, + 247 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "4.4 Subjective Results", + "text_level": 1, + "bbox": [ + 215, + 369, + 415, + 383 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Fig. 7 shows the subjective results compared with VTM and Zou2022. As we can see, with the help of the segmentation masks in training, our SegPIC can distinguish the objects' contours more accurately, making the edges sharper with less bitrate. In $kodim04$ , we can see the wool reconstructed by SegPIC is more consistent. In $kodim24$ , the bird patterns decoded by SegPIC are more clear, and the blurrings are less than in other methods. These results are based on grid partitions used as alternatives to masks. It demonstrates that SegPIC has acquired a more robust representative capability for various objects. Consequently, SegPIC can discern objects in complex textures more effectively.", + "bbox": [ + 212, + 396, + 787, + 532 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "4.5 Analysis of the Role of Segmentation Masks", + "text_level": 1, + "bbox": [ + 215, + 556, + 625, + 571 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Quantitative Results. To validate the role of the masks, we make experiments on two datasets, Kodak and COCO-sample. Due to Kodak not being a segmentation dataset, the corresponding masks are generated using the segmentation method [36]. Otherwise, we randomly sample 20 images and their masks from COCO Valid 2017 [9] to create the COCO-sample dataset.", + "bbox": [ + 212, + 582, + 787, + 657 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "As illustrated in Fig. 8 (a) (b), during the testing (inference) phase, the differences caused by masks are not so evident. However, for training, the situation is entirely different. With the assistance of the masks, the model achieves significantly better performance, especially at low bitrates. It could imply that, due to bitrate loss constraints during training, the information in the prototypes is implicitly transferred to the learnable parameters in RAT, allowing limited contextual information to still be effective during the inference phase. At low bitrates, since the lack of sufficient high-frequency information, pixels in the region exhibit stronger similarity, enhancing the representational ability of features extracted by prototypes. Fig. 8 (c) shows the early stage of model training.", + "bbox": [ + 212, + 657, + 787, + 809 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "As shown in Tab. 3, the model training w/o masks (the last column) leads to noticeable degradation. However, training w/ masks has a relatively minor", + "bbox": [ + 214, + 809, + 785, + 839 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "Region-Adaptive Transform with Segmentation Prior for Image Compression", + "bbox": [ + 220, + 114, + 730, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 784, + 126 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/77ae98955415a15f9706ec26f4656280dbe084927ad793772e397990b7abe755.jpg", + "image_caption": [ + "Fig. 9: Visualization of the kernels generated by DKG in RAT at the Encoder (see Fig. 3), in the cases with and without the masks. Fig. (d)(e)(f) are generated by the same pre-trained network weights. During testing, they go through processes with or without masks, respectively. The distortions in (d) can be formulated as $(x - \\hat{x})^2$ . We take the maximum absolute values of the tensors along the channels and apply 2D max pooling on them for better visualization." + ], + "image_footnote": [], + "bbox": [ + 218, + 143, + 787, + 233 + ], + "page_idx": 13 + }, + { + "type": "table", + "img_path": "images/9abe95cb5dde4a2107205fb3695116451bb574925e6d73184c6a2817128ac72b.jpg", + "table_caption": [ + "Table 3: The effect of the masks and the grid number on BD-rate on COCO-sample. \"n×\" denotes n×n grid partitions to replace masks." + ], + "table_footnote": [], + "table_body": "
Traintrain w/ masks4×
Testtest w/ masks2×4×8×4×
BD-rate ↓-0.11+0.210(default)+0.46+2.10
", + "bbox": [ + 287, + 383, + 715, + 444 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "impact on the pattern and number in testing. So we can replace the masks in testing to further save bitrate but should take the masks as privileged information during training. The hyperparameter $4 \\times$ in testing is an empirical trade-off between the effect of RAT (i.e. image reconstruction precision) and the quantity of the prototypes (i.e. the additional bitrate they introduced).", + "bbox": [ + 212, + 473, + 787, + 547 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Qualitative Analysis. Fig. 9 shows the visualization of the kernels generated by DKG in RAT at the Encoder (see Fig. 2). Although there is not a noticeable difference in Fig. 9 (e), it is obvious that the gradients of $w / \\text{mask}$ are more focused on the semantic objects (see Fig. 9 (f)), which is complex and prone to causing distortions. This attribute will assist in optimizing the models, enabling it to better reconstruct various objects. These observations indicate that the masks do play a necessary role during the training phase.", + "bbox": [ + 212, + 550, + 787, + 656 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 683, + 359, + 699 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "We propose a novel image compression framework called Segmentation-Prior-Guided Image Compression (SegPIC), equipped with two proposed modules: Region-Adaptive Transform (RAT) and Scale Affine Layer (SAL). We adopt the class-agnostic masks as privilege information to assist the training of SegPIC. Experimental results demonstrate that our SegPIC is superior to the previous works. Otherwise, we analyze the role of the masks. Experiments and visualization results show that the masks can guide the model to focus more on semantic objects in training, leading to higher pixel-fidelity performance.", + "bbox": [ + 212, + 719, + 787, + 840 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "Y. Liu et al.", + "bbox": [ + 271, + 114, + 354, + 127 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 217, + 143, + 401, + 162 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "This work was supported in part by the National Key R&D Program of China (No.2021ZD0112100), National NSF of China (No.62120106009), Guangdong Basic and Applied Basic Research Foundation (2024A1515010454), the Basic and Frontier Research Project of PCL, and the Major Key Project of PCL.", + "bbox": [ + 212, + 178, + 787, + 239 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 215, + 263, + 323, + 277 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Workshop and challenge on learned image compression (clic2020) (2020), http://www.compression.cc 10, 11", + "2. Agustsson, E., Minnen, D., Toderici, G., Mentzer, F.: Multi-realism image compression with a conditional generator. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 22324-22333 (2023) 10", + "3. Ahmed, N., Natarajan, T., Rao, K.R.: Discrete cosine transform. IEEE Transactions on Computers 100(1), 90-93 (1974) 2", + "4. Akbari, M., Liang, J., Han, J.: Dsslic: Deep semantic segmentation-based layered image compression. In: ICASSP. pp. 2042-2046. IEEE (2019) 3", + "5. Balle, J., Laparra, V., Simoncelli, E.P.: End-to-end optimized image compression. In: Int. Conf. Learn. Represent. (2017) 2, 5", + "6. Balle, J., Minnen, D., Singh, S., Hwang, S.J., Johnston, N.: Variational image compression with a scale hyperprior. In: Int. Conf. Learn. Represent. (2018) 2, 4, 5, 12", + "7. Bellard, F.: Bpg image format (2014), https://bellard.org/bpg/ 2, 10, 11", + "8. Bjontegaard, G.: Calculation of average psnr differences between rd-curves. ITU SG16 Doc. VCEG-M33 (2001) 11", + "9. Caesar, H., Uijlings, J., Ferrari, V.: COCO-Stuff: Thing and stuff classes in context. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 1209-1218 (2018) 10, 13", + "0. Chang, J., Mao, Q., Zhao, Z., Wang, S., Wang, S., Zhu, H., Ma, S.: Layered conceptual image compression via deep semantic synthesis. In: IEEE Int. Conf. Image Process. pp. 694-698 (2019) 4, 5", + "11. Chang, J., Zhao, Z., Yang, L., Jia, C., Zhang, J., Ma, S.: Thousand to one: Semantic prior modeling for conceptual coding. In: Int. Conf. Multimedia and Expo. pp. 1-6. IEEE (2021) 3, 4, 5", + "12. Cheng, Z., Sun, H., Takeuchi, M., Katto, J.: Learned image compression with discretized gaussian mixture likelihoods and attention modules. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 7939-7948 (2020) 3, 4, 10, 11", + "13. Chollet, F.: Xception: Deep learning with depthwise separable convolutions. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 1251-1258 (2017) 8", + "14. Daubechies, I.: The wavelet transform, time-frequency localization and signal analysis. IEEE Transactions on Information Theory 36(5), 961-1005 (1990) 2", + "15. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., Uszkoreit, J., Houlsby, N.: An image is worth 16x16 words: Transformers for image recognition at scale. In: Int. Conf. Learn. Represent. (2021) 3", + "16. Feng, R., Gao, Y., Jin, X., Feng, R., Chen, Z.: Semantically structured image compression via irregular group-based decoupling. In: Int. Conf. Comput. Vis. pp. 17237-17247 (2023) 3, 5, 10" + ], + "bbox": [ + 218, + 295, + 784, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "Region-Adaptive Transform with Segmentation Prior for Image Compression", + "bbox": [ + 220, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "17. Goodfellow, I., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., Courville, A., Bengio, Y.: Generative adversarial nets. In: Ghahramani, Z., Welling, M., Cortes, C., Lawrence, N., Weinberger, K. (eds.) Adv. Neural Inform. Process. Syst. vol. 27. Curran Associates, Inc. (2014) 4", + "18. He, D., Yang, Z., Peng, W., Ma, R., Qin, H., Wang, Y.: Elic: Efficient learned image compression with unevenly grouped space-channel contextual adaptive coding. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 5718-5727 (2022) 10", + "19. He, D., Zheng, Y., Sun, B., Wang, Y., Qin, H.: Checkerboard context model for efficient learned image compression. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 14771-14780 (2021) 4", + "20. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 770-778 (2016) 3", + "21. Hoang, T.M., Zhou, J., Fan, Y.: Image compression with encoder-decoder matched semantic segmentation. In: IEEE Conf. Comput. Vis. Pattern Recog. Worksh. pp. 160-161 (2020) 3", + "22. Jia, X., De Brabandere, B., Tuytelaars, T., Gool, L.V.: Dynamic filter networks. Adv. Neural Inform. Process. Syst. 29 (2016) 8", + "23. Jiang, W., Yang, J., Zhai, Y., Ning, P., Gao, F., Wang, R.: Mlic: Multi-reference entropy model for learned image compression. In: ACM Int. Conf. Multimedia. pp. 7618-7627 (2023) 4", + "24. (JVET), J.V.E.T.: Versatile video coding (2021), https://jvet.hhi.fraunhofer.de/ 2, 11", + "25. Kingma, D.P., Ba, J.: Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980 (2014) 10", + "26. Kingma, D.P., Welling, M.: Auto-encoding variational bayes. In: Int. Conf. Learn. Represent. (2014) 2", + "27. Kodak, E.: Kodak lossless true color image suite (photocd pcd0992) (1993), http://r0k.us/graphics/kodak/ 10, 11", + "28. Koyuncu, A.B., Gao, H., Boev, A., Gaikov, G., Alshina, E., Steinbach, E.: Contextformer: A transformer with spatio-channel attention for context modeling in learned image compression. In: Eur. Conf. Comput. Vis. pp. 447-463. Springer (2022) 10", + "29. Li, F., Zhang, L., Liu, Z., Lei, J., Li, Z.: Multi-frequency representation enhancement with privilege information for video super-resolution. In: Int. Conf. Comput. Vis. pp. 12814-12825 (2023) 3", + "30. Liu, J., Lu, G., Hu, Z., Xu, D.: A unified end-to-end framework for efficient deep image compression. arXiv preprint arXiv:2002.03370 (2020) 3, 4", + "31. Liu, J., Sun, H., Katto, J.: Learned image compression with mixed transformer-cnn architectures. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 14388-14397 (2023) 4", + "32. Liu, Z., Lin, Y., Cao, Y., Hu, H., Wei, Y., Zhang, Z., Lin, S., Guo, B.: Swin transformer: Hierarchical vision transformer using shifted windows. In: Int. Conf. Comput. Vis. pp. 10012-10022 (2021) 3, 4", + "33. Minnen, D., Balle, J., Toderici, G.D.: Joint autoregressive and hierarchical priors for learned image compression. Adv. Neural Inform. Process. Syst. 31 (2018) 4", + "34. Minnen, D., Singh, S.: Channel-wise autoregressive entropy models for learned image compression. In: IEEE Int. Conf. Image Process. pp. 3339-3343. IEEE (2020) 4, 5, 9, 10, 11, 12", + "35. Pan, G., Lu, G., Hu, Z., Xu, D.: Content adaptive latents and decoder for neural image compression. In: Eur. Conf. Comput. Vis. pp. 556-573. Springer (2022) 10" + ], + "bbox": [ + 215, + 146, + 785, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "Y. Liu et al.", + "bbox": [ + 271, + 114, + 354, + 127 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "36. Qi, L., Kuen, J., Wang, Y., Gu, J., Zhao, H., Torr, P., Lin, Z., Jia, J.: Open world entity segmentation. IEEE Trans. Pattern Anal. Mach. Intell. (2022) 13", + "37. Shen, H., Zhao, Z.Q., Zhang, W.: Adaptive dynamic filtering network for image denoising. In: AAAI. vol. 37, pp. 2227-2235 (2023) 6", + "38. Skodras, A., Christopoulos, C., Ebrahimi, T.: The JPEG 2000 still image compression standard. IEEE Signal Processing Magazine 18(5), 36-58 (2001) 2", + "39. Stark, H., Woods, J.W.: Probability, random processes, and estimation theory for engineers. Prentice-Hall, Inc. (1986) 2", + "40. Sun, S., He, T., Chen, Z.: Semantic structured image coding framework for multiple intelligent applications. IEEE Trans. Circuit Syst. Video Technol. 31(9), 3631-3642 (2020) 5", + "41. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. Adv. Neural Inform. Process. Syst. 30 (2017) 4", + "42. Wallace, G.K.: The JPEG still picture compression standard. Communications of the ACM 34(4), 30-44 (1991) 2", + "43. Wang, D., Yang, W., Hu, Y., Liu, J.: Neural data-dependent transform for learned image compression. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 17379-17388 (2022) 2, 3, 4, 10, 11", + "44. Wang, G.H., Li, J., Li, B., Lu, Y.: Evc: Towards real-time neural image compression with mask decay. In: Int. Conf. Learn. Represent. (2023) 10", + "45. Wang, X., Yu, K., Dong, C., Loy, C.C.: Recovering realistic texture in image superresolution by deep spatial feature transform. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 606-615 (2018) 8", + "46. Wang, Z., Simoncelli, E.P., Bovik, A.C.: Multiscale structural similarity for image quality assessment. In: The Thirty-Seventh Asilomar Conference on Signals, Systems & Computers, 2003. vol. 2, pp. 1398-1402. IEEE (2003) 4, 10", + "47. Xu, Y.S., Tseng, S.Y.R., Tseng, Y., Kuo, H.K., Tsai, Y.M.: Unified dynamic convolutional network for super-resolution with variational degradations. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 12496-12505 (2020) 6", + "48. Zhu, Y., Yang, Y., Cohen, T.: Transformer-based transform coding. In: Int. Conf. Learn. Represent. (2021) 3, 4", + "49. Zou, R., Song, C., Zhang, Z.: The devil is in the details: Window-based attention for image compression. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 17492-17501 (2022) 3, 4, 5, 7, 9, 10, 11, 12" + ], + "bbox": [ + 215, + 146, + 785, + 632 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "Region-Adaptive Transform with Segmentation Prior for Image Compression", + "bbox": [ + 220, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + } +] \ No newline at end of file diff --git a/2024/Region-Adaptive Transform with Segmentation Prior for Image Compression/7ddd6f0f-062f-4fcd-9635-2169878bc97b_model.json b/2024/Region-Adaptive Transform with Segmentation Prior for Image Compression/7ddd6f0f-062f-4fcd-9635-2169878bc97b_model.json new file mode 100644 index 0000000000000000000000000000000000000000..2ecd73c65e3ba51a46e70e82843577c460932dfd --- /dev/null +++ b/2024/Region-Adaptive Transform with Segmentation Prior for Image Compression/7ddd6f0f-062f-4fcd-9635-2169878bc97b_model.json @@ -0,0 +1,2698 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.233, + 0.141, + 0.773, + 0.187 + ], + "angle": 0, + "content": "Region-Adaptive Transform with Segmentation Prior for Image Compression" + }, + { + "type": "text", + "bbox": [ + 0.241, + 0.212, + 0.761, + 0.243 + ], + "angle": 0, + "content": "Yuxi Liu\\(^{1,2}\\), Wenhan Yang\\(^{3}\\), Huihui Bai\\(^{1,2}\\), Yunchao Wei\\(^{1,2,3}\\), and Yao Zhao\\(^{\\star 1,2,3}\\)" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.255, + 0.786, + 0.313 + ], + "angle": 0, + "content": "\\(^{1}\\) Institute of Information Science, Beijing Jiaotong University, Beijing, China \n\\(^{2}\\) Visual Intelligence + X International Joint Laboratory of the Ministry of Education \n\\(^{3}\\) PengCheng Laboratory, Shenzhen, China \n{yuxiliu, hhbai, yunchao.wei, yzhao}@bjtu.edu.cn, yangwh@pcl.ac.cn" + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.349, + 0.74, + 0.655 + ], + "angle": 0, + "content": "Abstract. Learned Image Compression (LIC) has shown remarkable progress in recent years. Existing works commonly employ CNN-based or Transformer-based modules as transform methods for compression. However, there is no prior research on neural transform that focuses on specific regions. In response, we introduce the class-agnostic segmentation masks (i.e. semantic masks without category labels) for extracting region-adaptive contextual information. Our proposed module, Region-Adaptive Transform, applies adaptive convolutions on different regions guided by the masks. Additionally, we introduce a plug-and-play module named Scale Affine Layer to incorporate rich contexts from various regions. While there have been prior image compression efforts that involve segmentation masks as additional intermediate inputs, our approach differs significantly from them. Our advantages lie in that, to avoid extra bitrate overhead, we treat these masks as privilege information, which is accessible during the model training stage but not required during the inference phase. To the best of our knowledge, we are the first to employ class-agnostic masks as privilege information and achieve superior performance in pixel-fidelity metrics, such as Peak Signal to Noise Ratio (PSNR). The experimental results demonstrate our improvement compared to previously well-performing methods, with about \\(8.2\\%\\) bitrate saving compared to VTM-17.0. The source code is available at https://github.com/GityuxiLiu/SegPIC-for-Image-Compression." + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.668, + 0.684, + 0.683 + ], + "angle": 0, + "content": "Keywords: Learned Image Compression \\(\\cdot\\) Segmentation Prior" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.709, + 0.377, + 0.725 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.741, + 0.788, + 0.817 + ], + "angle": 0, + "content": "With the advent of the multimedia era, an ever-growing volume of image data has emerged on the internet. These vast volumes of image data demand larger transmission bandwidth and storage capacity. In this context, the importance of image compression technology has become even more pronounced, evoking greater efforts in efficient image compression techniques. Lossy image compres" + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.825, + 0.383, + 0.84 + ], + "angle": 0, + "content": "* Corresponding author" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.356, + 0.127 + ], + "angle": 0, + "content": "Y. Liu et al." + }, + { + "type": "image", + "bbox": [ + 0.339, + 0.145, + 0.667, + 0.214 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.428, + 0.216, + 0.576, + 0.227 + ], + "angle": 0, + "content": "(a) Convictional LIC framework" + }, + { + "type": "image", + "bbox": [ + 0.339, + 0.235, + 0.666, + 0.3 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.419, + 0.302, + 0.584, + 0.313 + ], + "angle": 0, + "content": "(b) Data-dependent LIC framework" + }, + { + "type": "image", + "bbox": [ + 0.34, + 0.32, + 0.665, + 0.409 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.435, + 0.413, + 0.569, + 0.424 + ], + "angle": 0, + "content": "(c) Proposed LIC framework" + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.435, + 0.788, + 0.492 + ], + "angle": 0, + "content": "Fig. 1: Paradigms of end-to-end learned image compression (LIC) methods. (a) Conventional LIC with fixed learned kernels. (b) Data-dependent LIC with image-specific kernels [43]. (c) The proposed method is capable of generating region-adaptive kernels from different regions, providing more fine-grained transforms." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.523, + 0.788, + 0.734 + ], + "angle": 0, + "content": "sion technology can reduce the data size to a lower level while maintaining good visual quality. In the past decades, the classical lossy image compression methods such as JPEG [42], JPEG2000 [38], BPG [7], and VVC [24], used to have excellent performance and were widely adopted. These methods include modules of transform, quantization, and entropy coding. The transform process aims to convert images into a more compact coefficient by removing pixel correlations as much as possible. The commonly employed transforms include the discrete cosine transform (DCT) [3] and wavelet transform [14]. These transformations are linear, characterized by simplicity and efficiency. However, their fixed forms are unable to adapt to the contexts of complex signals. To address this issue, latter efforts propose data-driven transformation approaches, e.g. the Karhunen-Loève transform (KLT) [39]. Furthermore, the subsequent works are dedicated to constructing complex patterns for the adaptive selection of local transforms, improving adaptability and overall performance." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Benfit from the advancement of neural network technology, many learned image compression (LIC) methods have been proposed, which shows a promising alternative for designing transform: learning from data instead of handcraft design. Balle et al. [5, 6] proposed the first well-known end-to-end framework. They utilize variational auto-encoders (VAE) [26] as a transform module and optimize the parameters of the kernels to better align with the natural data distribution. To further improve the representation capabilities of the transform" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.22, + 0.115, + 0.732, + 0.131 + ], + "angle": 0, + "content": "Region-Adaptive Transform with Segmentation Prior for Image Compression" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.298 + ], + "angle": 0, + "content": "module, some works [12, 30, 43, 48, 49] design advanced modules to introduce richer contextual information, e.g. residual blocks [20] and self-attention mechanism [15, 32]. These works optimize learnable kernels to fit the data distribution to attain a more powerful transform. However, their kernel parameters are fixed after training (see Fig.1 (a)). Wang et al. [43] introduce a neural data-dependent transform, which extracts dynamic convolution kernels based on the context of the entire image for more adaptive transform (see Fig. 1 (b)). However, even within a single image, the context could vary significantly across different regions, e.g. stone and sheep regions in Fig. 1. It limits the description of the signals to a more fine-grained level using such image-level transforms." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.298, + 0.789, + 0.6 + ], + "angle": 0, + "content": "To address the issue, this paper makes the first efforts to design learnable region-adaptive transform to capture the mapping regularity in different regions. In detail, we employ class-agnostic4 segmentation masks to extract region-specific information for transform (see Fig. 1 (c)). Diffs from previous works [4, 11, 16, 21], our approach stands out with two characteristics in mask utilization. First, our masks are \"class-agnostic\", where the semantic labels are abandoned but a more effective compression-driven semantic prior will be learned by our model in a data-driven manner. Second, to avoid incurring additional costs in both bitrate and computational resources, we treat these masks as privilege information, which are available during training but remain inaccessible during the inference phase [29]. Specifically, we replace the masks with uniform grid partitions (see Fig. 9 (c)) in inference, and the experimental result shows approximate performance. It implies that our model has learned useful semantic knowledge from the masks in training, incorporating this knowledge into the model (i.e. the network's parameters). Hence, it remains capable of mining contextual information even from relatively coarse partition regions, i.e., the given uniform grid partitions. This simple operation can effectively maintain the consistency of our model's architecture and parameters, whether or not masks are used at different phases. The more details about the difference between our method and previous semantic-mask-based works will be explained at Sec. 2." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.601, + 0.789, + 0.737 + ], + "angle": 0, + "content": "Based on the above motivations, we propose a novel image compression framework called Segmentation-Prior-Guided Image Compression (SegPIC), equipped with two key modules: Region-Adaptive Transform (RAT) and Scale Affine Layer (SAL). The RAT adopts region-adaptive kernels for transform, performing the convolution whose weights are generated from the prototypes and the context of latent features under the guidance of masks. The SAL affines the latent features for better mining of semantic contexts in the Encoder and Decoder. Experimental results show that SegPIC is superior to previously well-performing image compression methods. Our main contributions are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.745, + 0.788, + 0.79 + ], + "angle": 0, + "content": "- We propose a novel image compression framework called Segmentation-Prior-Guided Image Compression (SegPIC) to generate more effective region-adaptive transforms in a data-driven manner." + }, + { + "type": "page_footnote", + "bbox": [ + 0.221, + 0.797, + 0.789, + 0.84 + ], + "angle": 0, + "content": "\\( {}^{4} \\) \"class-agnostic\" means distinguishing different objects in an image regardless of what those objects are. Such masks can be regarded as a generalized representation of partition maps widely used in conventional CODECs." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.356, + 0.128 + ], + "angle": 0, + "content": "Y. Liu et al." + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.147, + 0.784, + 0.206 + ], + "angle": 0, + "content": "- Our SegPIC is equipped with two proposed modules: Region-Adaptive Transform (RAT) and Scale Affine Layer (SAL). The RAT adopts region-adaptive kernels for transform, while the SAL affines the latent features for better mining of semantic contexts." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.208, + 0.785, + 0.265 + ], + "angle": 0, + "content": "- We utilize the class-agnostic masks as privilege information to assist the training of SegPIC. The \"class-agnostic\" property promotes the model to learn a compression-driven semantic prior, while the \"privilege\" way avoids additional bitrate and complexity for masks in inference." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.269, + 0.785, + 0.312 + ], + "angle": 0, + "content": "- Experimental results demonstrate that our SegPIC is superior to previously well-performing image compression methods. We also present sufficient ablation experiments to confirm the necessity of the privilege information." + }, + { + "type": "list", + "bbox": [ + 0.225, + 0.147, + 0.785, + 0.312 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.338, + 0.396, + 0.354 + ], + "angle": 0, + "content": "2 Related Works" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.372, + 0.783, + 0.476 + ], + "angle": 0, + "content": "Learned Image Compression. Image compression can be divided into three main components: transform, quantization, and entropy coding. In brief, the transformation module maps the image to a latent feature space. Subsequently, quantization is performed after this phase, resulting in some quantization distortions but achieving lower bitrate. The entropy encoding module estimates the data distribution of the latent features and inputs it to the Arithmetic Coding (AC) module, which generates the bitstream for transmission." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.478, + 0.784, + 0.657 + ], + "angle": 0, + "content": "To enhance transform module, some works [30,43,48,49] adopt more powerful modules like self-attention mechanism [32, 41] to explore broader context. [31] proposed a transform block that mixes CNN and Transformer to combine their advantages. As for the entropy estimation module, the studies [12, 19, 33, 34] usually focus on exploring novel auto-regressive methods to split the latent features into several portions to be decoded in order. The portions having been decoded will assist in decoding the next portion, through the spatial or channel correlation of these portions. This auto-regressive mechanism may introduce additional latency but achieve better performance than Hyper Prior [6]. Recently, [23] integrates different types of correlations present in latent representation, e.g. channel-wise, local spatial, and global spatial correlations. Their proposed novel auto-regressive entropy module achieve powerful performance." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.66, + 0.785, + 0.84 + ], + "angle": 0, + "content": "Thanks to these great works, the performance of LIC has become so strong that achieving further enhancements becomes progressively challenging. Therefore, some new directions in compression have emerged, which is also meaningful on certain occasions. For example, some semantic-based compression methods [10, 11] adopt GAN [17] to compress images at extremely low bitrate. These methods maintain a good human visual perception effect but usually, poor pixel-fidelity metrics, e.g. Peak Signal to Noise Ratio (PSNR) and Multi-Scale Structural Similarity Index (MS-SSIM) [46]. Specifically, these methods could introduce some synthetic artifacts with low fidelity. For example, a semantic compression method could reconstruct an image of a bird with different breeds from the original image because the training category labels do not distinguish the bird's breed. Another direction is compression for machines, which means compress" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.221, + 0.115, + 0.732, + 0.13 + ], + "angle": 0, + "content": "Region-Adaptive Transform with Segmentation Prior for Image Compression" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "image", + "bbox": [ + 0.269, + 0.146, + 0.734, + 0.334 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.421, + 0.34, + 0.561, + 0.353 + ], + "angle": 0, + "content": "(a) The overall framework" + }, + { + "type": "image", + "bbox": [ + 0.257, + 0.37, + 0.495, + 0.455 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.272, + 0.463, + 0.478, + 0.476 + ], + "angle": 0, + "content": "(b)Downsample block (out channel C)" + }, + { + "type": "image", + "bbox": [ + 0.509, + 0.37, + 0.747, + 0.454 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.524, + 0.463, + 0.713, + 0.476 + ], + "angle": 0, + "content": "(c) Upsample block (out channel C)" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.487, + 0.788, + 0.626 + ], + "angle": 0, + "content": "Fig. 2: (a) Overall framework of our SegPIC. The blue flow lines represent the decoding process on the decoder side. The dashed lines represent the process of extracting the prototypes \\( p' \\), \\( p \\) (see Eq. 5). The compression of \\( p' \\), \\( p \\) is not fully depicted for simplicity, please refer to Fig. 4 for details. \\( N \\) is set to 192, and \\( M \\) is 320 as in [49]. RAT is the proposed Region-Adaptive Transform. SAL is the proposed Scale Affine Layer. WAM is Window Attention Module [49]. ChARM is Channel-wise Auto-Regressive Model [34]. FM is a Factorized Model [6]. GDN is Generalized Divisive Normalization [5]. (b) is Downsample Block. \"Downsample, N\" in (a) means the block with out-channel N. \"Conv C,5,2\" means out-channel C, kernel size 5, stride 2. (c) is Upsample Block, and \"TConv\" is Transposed Convolution." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.657, + 0.787, + 0.687 + ], + "angle": 0, + "content": "ing images for downstream tasks [16, 40]. These methods focus on balancing the bitrate and performance of downstream tasks, not just the pixel-fidelity metrics." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.788, + 0.809 + ], + "angle": 0, + "content": "Although some works [10, 11, 16] about the two directions also adopt masks, our usage is significantly different. The main differences are as follows: 1) We utilize the masks as privilege information, without the mask information during the inference phase, while usually, the masks are indispensable for other methods. 2) We employ class-agnostic masks to guide the model to extract rich compression-friendly contexts among semantically related regions, regardless of the category that might degrade the fidelity as mentioned above. 3) The proposed model is optimized towards better pixel-fidelity metrics, e.g. PSNR and MS-SSIM." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Region-adaptive Convolution. Due to its capacity to model the degradation levels in different regions of images, region-adaptive convolution is applied to" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.356, + 0.128 + ], + "angle": 0, + "content": "Y. Liu et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.269 + ], + "angle": 0, + "content": "some low-level vision tasks such as super-resolution and denoising [37, 47]. In terms of compression, the reconstruction quality varies across different regions. For instance, regions with complex textures are more challenging compared to those with simpler patterns. Therefore, specific processing can be applied to different regions based on their characteristics. The regions can be divided using segmentation masks or partitions based on certain regular patterns. To the best of our knowledge, we are the first to introduce this mechanism on image compression for optimizing pixel-fidelity metrics." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.29, + 0.424, + 0.308 + ], + "angle": 0, + "content": "3 Proposed SegPIC" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.321, + 0.356, + 0.336 + ], + "angle": 0, + "content": "3.1 Motivation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.345, + 0.787, + 0.391 + ], + "angle": 0, + "content": "As mentioned above, previous transforms either adopt convolutions with fixed parameters or image-specific convolutions. Our motivation is to construct a region-adaptive transform via the following three properties:" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.399, + 0.785, + 0.428 + ], + "angle": 0, + "content": "- Region-dependent: we propose a neural network with region-adaptive transforms guided by the segmentation masks." + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.429, + 0.785, + 0.459 + ], + "angle": 0, + "content": "- Class-Agnostic: the adopted mask does not include a category label, allowing the model to be free to learn compression-friendly semantic priors." + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.459, + 0.785, + 0.488 + ], + "angle": 0, + "content": "- Privilege Information: the masks are expected as privilege information, not being used in inference for saving bitrate and computation resources." + }, + { + "type": "list", + "bbox": [ + 0.225, + 0.399, + 0.785, + 0.488 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.498, + 0.787, + 0.529 + ], + "angle": 0, + "content": "Based on these considerations, we propose our SegPIC in the VAE framework incorporated with two proposed modules: SAL and RAT." + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.539, + 0.787, + 0.674 + ], + "angle": 0, + "content": "- SAL: First of all, we extract a prototype from each region, which will be compressed and transmitted to be accessible on both the encoder and decoder sides. The prototypes are expected to extract high-level information from regions to ensure the representative ability as much as possible. However, the encoder in compression frameworks often extracts features that contain low-level information. Therefore, a module is highly desirable for communicating information across different levels. SAL is designed for extracting more abundant semantics and contexts, enriching the prototypes, and connecting features from two different levels." + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.674, + 0.787, + 0.765 + ], + "angle": 0, + "content": "- RAT: performing the Depth&Point-wise Separable Convolution (DPSCov), a novel convolution whose weights are generated from the prototypes and the context of latent features under the guidance of masks. Note that, we only utilize the masks to guide RAT to learn more compression-friendly semantic priors in training, while during the inference phase, we instead adopt uniform grid partitions as coarse alternatives." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.786, + 0.364, + 0.801 + ], + "angle": 0, + "content": "3.2 Formulation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.841 + ], + "angle": 0, + "content": "In this section, we will introduce the problem and our method in mathematical form. Given an image \\( x \\), the encoder \\( En \\) maps it to latent features \\( y \\). After" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.22, + 0.115, + 0.733, + 0.131 + ], + "angle": 0, + "content": "Region-Adaptive Transform with Segmentation Prior for Image Compression" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.193 + ], + "angle": 0, + "content": "quantization \\( Q \\), \\( \\hat{y} \\) will be compressed to the bitstream by the entropy coding, such as arithmetic coding (AC). Then the receiver converts the bitstream back into \\( \\hat{y} \\) by entropy decoding. The above process can be formulated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.388, + 0.203, + 0.558, + 0.219 + ], + "angle": 0, + "content": "\\[\ny = E n (x), \\quad \\hat {y} = Q (y),\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.388, + 0.221, + 0.785, + 0.238 + ], + "angle": 0, + "content": "Transmitter: \\(bit_{\\hat{y}} = AE(\\hat{y} |p_{\\hat{y}})\\) (1)" + }, + { + "type": "text", + "bbox": [ + 0.388, + 0.24, + 0.615, + 0.257 + ], + "angle": 0, + "content": "Receiver: \\(\\hat{y} = AD(bit_{\\hat{y}}|p_{\\hat{y}})\\)" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.266, + 0.785, + 0.326 + ], + "angle": 0, + "content": "where \\( p_{\\hat{y}} \\) is the estimated distribution of \\( \\hat{y} \\), \\( \\text{bit}_{\\hat{y}} \\) is the binary bitstream of \\( \\hat{y} \\), and \\( AE \\) and \\( AD \\) are arithmetic encoding and decoding. We use a single Gaussian distribution to estimate \\( \\hat{y} \\sim \\mathcal{N}(\\mu, \\sigma^2) \\), and the mean and variance value \\( \\mu, \\sigma^2 \\) will be yielded through an auto-regressive mechanism. The formulation is:" + }, + { + "type": "equation", + "bbox": [ + 0.326, + 0.337, + 0.785, + 0.354 + ], + "angle": 0, + "content": "\\[\nz = E n _ {z} (y), \\hat {z} = Q (z), \\quad \\mu_ {i}, \\sigma_ {i} = F \\left(D e _ {z} (\\hat {z}), \\hat {y} _ {< i}\\right), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.363, + 0.785, + 0.453 + ], + "angle": 0, + "content": "where \\( En_{z}, De_{z} \\) is hyper encoder and decoder, \\( \\mu_{i}, \\sigma_{i} \\) are the mean and variance of \\( \\hat{y}_{i} \\sim \\mathcal{N}(\\mu_{i}, \\sigma_{i}^{2}) \\), \\( \\hat{y}_{i} \\) is the currently decoded portion of \\( \\hat{y} \\), \\( \\hat{y}_{< i} \\) is the already decoded portion to assist decoding \\( \\hat{y}_{i} \\), and \\( F \\) is the auto-regressive entropy model. The AC process of \\( z \\) in Eq. 2 is omitted for brevity. Finally, the receiver adopts decoder \\( De \\) to reconstruct the image \\( \\hat{x} \\), which is formulated as \\( \\hat{x} = De(\\hat{y}) \\)." + }, + { + "type": "text", + "bbox": [ + 0.239, + 0.455, + 0.586, + 0.469 + ], + "angle": 0, + "content": "The loss function \\(\\mathcal{L}\\) to optimize this problem is:" + }, + { + "type": "equation", + "bbox": [ + 0.319, + 0.48, + 0.785, + 0.497 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = R + \\lambda \\cdot D = R (\\hat {y}) + R (\\hat {z}) + \\lambda \\cdot \\mathbb {E} _ {x \\sim p _ {x}} [ d (x, \\hat {x}) ], \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.506, + 0.784, + 0.55 + ], + "angle": 0, + "content": "where \\( R(\\hat{y}), R(\\hat{z}) \\) are the bitrate of \\( \\hat{y}, \\hat{z} \\), \\( d(x, \\hat{x}) \\) is the distortion between \\( x \\) and \\( \\hat{x} \\), \\( \\lambda \\) denotes different rate-distortion trade-offs. Here we use Mean Square Error (MSE) or (1-MS-SSIM) as the distortion function." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.552, + 0.785, + 0.673 + ], + "angle": 0, + "content": "In our implementation, we use class-agnostic semantic masks (denoted as \\( m \\)) as additional inputs of the encoding \\( En \\) during the training phase. Moreover, we extract a prototype from each region at the middle layer of \\( En \\), then compress and transmit them (denoted as \\( \\hat{p} \\)). In inference, we replace the masks with uniform grid partitions, so there is no additional bitstream for masks, while the bitrates of the prototypes also need to be taken into account. The \\( En \\) and \\( De \\) will transform the image with the assistance of the prototypes. The formulations requiring adjustment are as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.379, + 0.683, + 0.626, + 0.699 + ], + "angle": 0, + "content": "\\[\ny = E n (x, \\hat {p} | m), \\quad \\hat {x} = D e (\\hat {y}, \\hat {p} | m),\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.378, + 0.702, + 0.785, + 0.718 + ], + "angle": 0, + "content": "\\[\n\\mu_ {i}, \\sigma_ {i} = F (D e _ {z} (\\hat {z}), \\hat {y} _ {< i}, \\hat {p} | m), \\tag {4}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.379, + 0.72, + 0.559, + 0.736 + ], + "angle": 0, + "content": "\\[\nR = R (\\hat {p}) + R (\\hat {y}) + R (\\hat {z}).\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.756, + 0.395, + 0.771 + ], + "angle": 0, + "content": "3.3 Network Design" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.783, + 0.81 + ], + "angle": 0, + "content": "In this section, we will introduce the concrete process and our network design. We use [49] as the backbone. The overall architecture is in Fig. 2." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.811, + 0.784, + 0.84 + ], + "angle": 0, + "content": "Scale Affine Layer. The Encoder and Decoder are responsible for image transform and reconstruction, each comprising 4 downsampling or upsampling blocks." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.356, + 0.128 + ], + "angle": 0, + "content": "Y. Liu et al." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.147, + 0.788, + 0.253 + ], + "angle": 0, + "content": "The previous sampling blocks consisted of a single sample convolution layer, aimed at modifying the data size. However, this may not be sufficient to extract adequate semantics and contexts. Inspired by Spatial Feature Transform (SFT) [45], which can generate visually pleasing textures in segmentation-guided low-level vision, we propose to add a Scale Affine Layer (SAL) between two adjacent convolution layers (see Fig. 2 (b) (c)). It consists of two \\(1 \\times 1\\) convolution layers with a GELU activation function in between, which is formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.377, + 0.261, + 0.626, + 0.297 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} X _ {s} = \\operatorname {S A L} _ {s} (X), \\quad X _ {b} = \\operatorname {S A L} _ {b} (X), \\\\ Y = X _ {s} \\otimes X + X _ {b}, \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.305, + 0.786, + 0.334 + ], + "angle": 0, + "content": "where \\( X \\) and \\( Y \\) represent the input and output, \\( X_{s} \\) and \\( X_{b} \\) denote the scales and biases generated by SAL, \\( \\otimes \\) denotes the element-wise product." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.335, + 0.788, + 0.441 + ], + "angle": 0, + "content": "Extract and Transmit the Prototypes. The Encoder first converts the original image \\( x \\in \\mathbb{R}^{3 \\times H \\times W} \\) into latent features \\( y' \\in \\mathbb{R}^{N \\times H/4 \\times W/4} \\) at the middle layer. Then we conduct Masked Average Pooling (MAP) on \\( y' \\) through the class-agnostic masks \\( m \\in \\{0,1\\}^{n \\times H \\times W} \\) (see Fig. 3 (a)), to acquire the prototypes \\( p' \\in \\mathbb{R}^{N \\times n} \\), where \\( n \\) is the number of the masks. Thanks to the proposed SAL, the Encoder becomes deeper, allowing \\( p' \\) to incorporate more semantics and contexts from the corresponding mask. The formulation of MAP is as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.311, + 0.448, + 0.787, + 0.481 + ], + "angle": 0, + "content": "\\[\np ^ {\\prime} = \\operatorname {M A P} \\left(y ^ {\\prime}, m\\right) = \\left\\{\\frac {1}{\\sum m ^ {(i)}} \\sum y ^ {\\prime} \\otimes m ^ {(i)}, \\dots \\right\\} _ {i \\in [ 1, n ]}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.487, + 0.788, + 0.564 + ], + "angle": 0, + "content": "where \\( y' \\) are the mid-level features (see Fig. 2 (a)), \\( \\sum \\) denotes summation along the spatial dimension. \\( p' \\) will be compressed and transmitted to the receiver (see Fig. 4), and the reconstructed \\( \\hat{p}' \\) is shared between the Encoder and Decoder. The channel \\( \\{C_1, C_2, C_3, C_4\\} \\) is {192, 128, 96, 96} for \\( p' \\) (on Encoder and Decoder) and {320, 192, 128, 96} for \\( p \\) (on Entropy Model)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.564, + 0.788, + 0.623 + ], + "angle": 0, + "content": "Region-Adaptive Transform. Here we propose a novel convolution called Depth&Point-wise Separable Convolution (DPSCov), in which each element has an individual convolution kernel, and there is no channel-wise mutual communication for computational efficiency." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.624, + 0.788, + 0.684 + ], + "angle": 0, + "content": "Before introducing it, we will start with standard convolution briefly. The input of standard convolution can be denoted as \\( X \\in \\mathbb{R}^{C \\times H \\times W} \\), the output is \\( Y \\in \\mathbb{R}^{O \\times H \\times W} \\), and the kernel is \\( W \\in \\mathbb{R}^{O \\times C \\times k \\times k} \\), where \\( k \\) is kernel size. The formulation of standard convolution is as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.298, + 0.69, + 0.706, + 0.733 + ], + "angle": 0, + "content": "\\[\nY _ {o, h, w} = \\sum_ {c = 1} ^ {C} X _ {c, h, w} * W _ {o, c} = \\sum_ {c = 1} ^ {C} \\sum_ {i, j} X _ {c, h + i, w + j} \\cdot W _ {o, c, i, j},\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.741, + 0.787, + 0.771 + ], + "angle": 0, + "content": "where \\( * \\) is 2D convolution, \\( i,j\\in [-\\lfloor k / 2\\rfloor ,\\lfloor k / 2\\rfloor ] \\), \\( k \\) is kernel size, and \\( \\lfloor \\cdot \\rfloor \\) denotes floor rounding." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.772, + 0.787, + 0.802 + ], + "angle": 0, + "content": "DPSCov can be seen as a combination of Depthwise Separable Convolution (DWConv) [13] and Dynamic Convolution [22]. The formulation is as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.309, + 0.81, + 0.787, + 0.845 + ], + "angle": 0, + "content": "\\[\nY _ {c, h, w} = X _ {c, h, w} * W _ {c, h, w} = \\sum_ {i, j} X _ {c, h + i, w + j} \\cdot W _ {c, h, w, i, j}. \\tag {6}\n\\]" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.221, + 0.115, + 0.732, + 0.13 + ], + "angle": 0, + "content": "Region-Adaptive Transform with Segmentation Prior for Image Compression" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.142, + 0.481, + 0.294 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.217, + 0.302, + 0.473, + 0.314 + ], + "angle": 0, + "content": "(a) Extract the prototypes through Masked Average Pooling" + }, + { + "type": "image", + "bbox": [ + 0.502, + 0.143, + 0.779, + 0.293 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.486, + 0.302, + 0.786, + 0.314 + ], + "angle": 0, + "content": "(b)The Region-Adaptive Transform (RAT) with masks and prototypes" + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.324, + 0.788, + 0.395 + ], + "angle": 0, + "content": "Fig. 3: The diagram of extracting the prototypes and the proposed RAT module. DPSCov is proposed Depth&Point-wise Separable Convolution (see Eq. 6). CTL is Channel Transform Layers, DKG is the DPSCov Kernel Generator, and CAG is the Channel Attention Generator. DPSCov is Depth&Point-wise Separable Convolution. The detailed modules are represented in Fig. 5." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.435, + 0.785, + 0.467 + ], + "angle": 0, + "content": "The computational complexity of standard convolution is \\( \\mathrm{O}(O \\times C \\times H \\times W \\times k^2) \\), while DPSCov is \\( \\mathrm{O}(C \\times H \\times W \\times k^2) \\), same as DWConv." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.472, + 0.789, + 0.684 + ], + "angle": 0, + "content": "The overall design of RAT is depicted in Fig. 3 (b). We expand the prototypes based on the positional information of the masks, filling in the corresponding copies of the prototypes at each pixel position. Then, we feed-forward the input through Channel Transform Layers (CTL) and concatenate it with the expanded features. The resulting variable, which contains both semantics and contexts, is used for subsequent filter generation. The DPSCConv Kernels Generator (DKG) is designed for generating the filters. In the first layer of DKG, we employ a 1x1 convolution to attempt the fusion of the prototype and the pixel's intrinsic attribute. We opt for grouped convolution in the final layer of the DKG to reduce the computation and parameters. Then Channel Attention Generator (CAG) generates the channel attention to be multiplied with the output of DPSCConv along the channel dimension. Finally, the output passes through a 1x1 convolution to merge channel information. We apply the RAT in both the Encoder and Decoder, as well as the entropy encoding module." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.789, + 0.841 + ], + "angle": 0, + "content": "Entropy Encoding Module. We use the Channel-wise Auto-Regressive Model (ChARM) [34] as our entropy estimation module, and the specific hyperparameter configuration follows [49]. In this model, the latent variable \\( y \\) is partitioned into 10 groups along the channel dimension. The Gaussian prior distribution \\( p(y_{i}|\\hat{z},\\hat{y} < i) \\) is estimated in an auto-regressive manner, where the mean and scale of \\( y_{i} \\) depend on the quantized latent in the preceding groups \\( \\hat{y}_{< i} \\) (see Eq. 2). To further enhance the estimation ability, we also deploy RAT in it. The RAT here can transmit key region prototypes through a separate bitstream, assisting in more efficient encoding and decoding. The prototypes \\( p \\) could be seen as segmentation-prior, a supplement of the hyper-prior information \\( z \\)." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.356, + 0.127 + ], + "angle": 0, + "content": "Y. Liu et al." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.147, + 0.445, + 0.277 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.287, + 0.449, + 0.371 + ], + "angle": 0, + "content": "Fig.4: The architecture of the Prototype Encoder and Decoder for compressing and transmitting the prototypes. \"Linear \\( \\mathrm{C_1,C_2} \\)\" means linear layer with in-channel \\( \\mathrm{C_1} \\) and out-channel \\( \\mathrm{C_2} \\)." + }, + { + "type": "image", + "bbox": [ + 0.498, + 0.146, + 0.756, + 0.263 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.461, + 0.274, + 0.792, + 0.371 + ], + "angle": 0, + "content": "Fig. 5: The detailed modules in RAT, including Channel Transform layers (CTL), DPSCConv Kernels Generator (DKG), and Channel Attention Generator (CAG). LReLU is Leaky ReLU. GAP is Global Average Pooling. \"GConv k²C,3,1,C\" means Grouped Convolution with out-channel k²C, kernel size 3, stride 1 and group C." + }, + { + "type": "table_caption", + "bbox": [ + 0.215, + 0.385, + 0.789, + 0.414 + ], + "angle": 0, + "content": "Table 1: BD-rate(%) performance on Kodak and CLIC Professional Valid (CLIC for short). The anchor is VTM-17.0." + }, + { + "type": "table", + "bbox": [ + 0.23, + 0.426, + 0.77, + 0.577 + ], + "angle": 0, + "content": "
ModelKodak \nBD-rate↓CLIC \nBD-rate↓ModelKodak \nBD-rate↓CLIC \nBD-rate↓
BPG [7]+31.5+48.1EVC [44]+0.27-
Cheng2020 [12]+7.46+11.7Feng2023 [16]-2.21-
Agustsson2023 [2]+6.15-ELIC [18]-3.08-3.58
ChARM [34]+3.58-Contextformer [28]-5.38-
Wang2022 [43]+1.10-0.99Pan2022 [35]-5.42-
BL-S+15.5+16.6Zou2022 (BL) [49]-2.89+0.14
Ours-S+3.87+5.26Ours-8.18-8.26
" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.602, + 0.377, + 0.62 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.634, + 0.429, + 0.649 + ], + "angle": 0, + "content": "4.1 Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.659, + 0.788, + 0.795 + ], + "angle": 0, + "content": "Training. We use COCO-Stuff [9] as the training dataset, which is composed of 118k images with panoptic semantic masks. We just use the class-agnostic masks without the category labels. We randomly crop the images with the size of \\(256 \\times 256\\). The models are trained for 270 epochs (about 1M iterations) using the Adam optimizer [25] with a batch size of 32. The initial learning rate is set to \\(1 \\times 10^{-4}\\) for 170 epochs, and reduced by half every 30 epochs. Our models are optimized by the rate-distortion trade-off loss function (see Eq. 3 and Eq. 4). The \\(\\lambda\\) belongs to \\(\\{0.0018, 0.0035, 0.0067, 0.0130, 0.0250, 0.0483\\}\\) for MSE as the distortion, and \\(\\{2.4, 4.58, 8.73, 16.64, 31.73, 60.50\\}\\) for MS-SSIM [46]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.796, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Evaluation. We evaluate our model on two commonly used datasets, Kodak [27] with the size of \\( 768 \\times 512 \\), and CLIC Professional Validation [1] with the size of 2K. Performance is evaluated using both bitrate and distortions. Distortion" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.221, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Region-Adaptive Transform with Segmentation Prior for Image Compression" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "image", + "bbox": [ + 0.216, + 0.144, + 0.501, + 0.315 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.506, + 0.143, + 0.788, + 0.315 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.325, + 0.788, + 0.385 + ], + "angle": 0, + "content": "Fig.6: Rate-Distortion Performance on Kodak [27] and CLIC Professional Valid dataset [1]. Our SegPIC introduces proposed RAT and SAL based on Zou2022 [49]. We also implement a simple baseline model without self-attention and auto-regressive mechanism (BL-S), and incorporate our proposed modules on it (Ours-S)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.408, + 0.788, + 0.472 + ], + "angle": 0, + "content": "measurement involves PSNR and MS-SSIM. And bitrates are assessed in bits per pixel (bpp). In this phase, we replace the masks with \\(4 \\times 4\\) grid partitions. Thus, there is no bitrate cost related to the masks, while the bitrate of prototypes \\(R(p)\\) and \\(R(p^{\\prime})\\) is still taken into account." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.489, + 0.504, + 0.505 + ], + "angle": 0, + "content": "4.2 Rate-Distortion Performance" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.511, + 0.788, + 0.662 + ], + "angle": 0, + "content": "We compare our SegPIC with previously well-performing methods, including discretized Gaussian mixture likelihoods and attention modules [12] (denoted as Cheng2020), ChARM [34] (denoted as Minnen2022), ChARM with WAM [49] (denoted as Zou2022), image-specific data-dependent transform model [43] (denoted as Wang2022), the best-performing traditional method VTM-17.0 [24] and influential traditional method BPG [7]. Furthermore, we introduce two simplified models without self-attention and auto-regressive mechanism, denoted as baseline-small (BL-S) and the one incorporating our proposed SAL and RAT (Ours-S), to demonstrate the generality of our modules. Details about the architecture of these two models will be provided in the appendix." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.662, + 0.788, + 0.77 + ], + "angle": 0, + "content": "As illustrated in Fig. 6 and Tab. 1, compared to VTM-17.0, our SegPIC could save \\(8.18\\%\\) BD-rate [8] on Kodak and \\(8.26\\%\\) on CLIC, which outperforms all the LIC and traditional methods mentioned above. To further verify the efficiency of the proposed RAT and SAL, We also compare SegPIC with the baseline Zou2022, which can save \\(5.45\\%\\) BD-rate on Kodak, and \\(8.50\\%\\) on CLIC (see Tab. 2). Moreover, Ours-S even achieves closely matched performance compared to ChARM. The MS-SSIM results are provided in the appendix." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.788, + 0.391, + 0.803 + ], + "angle": 0, + "content": "4.3 Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.842 + ], + "angle": 0, + "content": "To assess the contributions of our RAT and SAL modules individually, we independently incorporate these two modules into the baseline Zou2022. Conse" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.356, + 0.127 + ], + "angle": 0, + "content": "Y. Liu et al." + }, + { + "type": "image", + "bbox": [ + 0.22, + 0.145, + 0.33, + 0.272 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.246, + 0.272, + 0.305, + 0.281 + ], + "angle": 0, + "content": "Original kodim04" + }, + { + "type": "image", + "bbox": [ + 0.336, + 0.145, + 0.413, + 0.204 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.351, + 0.204, + 0.396, + 0.211 + ], + "angle": 0, + "content": "Ground Truth" + }, + { + "type": "image", + "bbox": [ + 0.419, + 0.146, + 0.495, + 0.203 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.425, + 0.204, + 0.49, + 0.211 + ], + "angle": 0, + "content": "VTM (31.02/0.115)" + }, + { + "type": "image", + "bbox": [ + 0.337, + 0.213, + 0.413, + 0.272 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.338, + 0.272, + 0.412, + 0.281 + ], + "angle": 0, + "content": "Zou2022 (31.46/0.119)" + }, + { + "type": "image", + "bbox": [ + 0.419, + 0.213, + 0.495, + 0.272 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.425, + 0.272, + 0.488, + 0.28 + ], + "angle": 0, + "content": "Ours (31.53/0.109)" + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.146, + 0.618, + 0.273 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.536, + 0.272, + 0.594, + 0.28 + ], + "angle": 0, + "content": "Original kodim24" + }, + { + "type": "image", + "bbox": [ + 0.625, + 0.146, + 0.701, + 0.204 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.641, + 0.204, + 0.686, + 0.211 + ], + "angle": 0, + "content": "Ground Truth" + }, + { + "type": "image", + "bbox": [ + 0.708, + 0.146, + 0.783, + 0.204 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.714, + 0.204, + 0.78, + 0.211 + ], + "angle": 0, + "content": "VTM (28.96/0.325)" + }, + { + "type": "image", + "bbox": [ + 0.625, + 0.213, + 0.701, + 0.272 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.627, + 0.272, + 0.702, + 0.28 + ], + "angle": 0, + "content": "Zou2022 (28.67/0.292)" + }, + { + "type": "image", + "bbox": [ + 0.708, + 0.213, + 0.783, + 0.272 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.714, + 0.272, + 0.777, + 0.28 + ], + "angle": 0, + "content": "Ours (28.98/0.289)" + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.292, + 0.787, + 0.335 + ], + "angle": 0, + "content": "Fig. 7: Visualization of the reconstructed images \\( \\text{kodim}04 \\) and \\( \\text{kodim}24 \\) in Kodak. The metrics are (PNSR↑/bpp↓). It shows that our SegPIC can distinguish the objects' contours more accurately, making the edges sharper with less bitrate." + }, + { + "type": "table_caption", + "bbox": [ + 0.215, + 0.37, + 0.788, + 0.399 + ], + "angle": 0, + "content": "Table 2: Ablation Study on Kodak and CLIC. The Baseline (BL) is Zou2022 [49]. The speed is measured on Kodak with one Nvidia RTX 3090 GPU." + }, + { + "type": "table", + "bbox": [ + 0.283, + 0.411, + 0.716, + 0.553 + ], + "angle": 0, + "content": "
ModelKodakCLIC
ParamsBD-rateBD-rateEncDec Time
Zou2022 (BL)75.2 M0.00.0120 ms124 ms
BL w/ SAL76.7 M-2.12-2.93--
BL w/ RAT82.0 M-2.75-4.35--
Ours83.5 M-5.45-8.50142 ms130 ms
ChARM48.8 M+6.52-107 ms108 ms
BL-S21.7 M+19.0+16.3--
Ours-S29.2 M+6.96+5.1199 ms46 ms
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.603, + 0.788, + 0.725 + ], + "angle": 0, + "content": "sequently, we construct two models: \"baseline w/ SAL\" and \"baseline w/ RAT.\" As illustrated in Tab. 2, the parameter counts introduced by RAT and SAL separately are \\(1.5\\mathrm{M}\\) and \\(5.8\\mathrm{M}\\), which occupy only a small fraction of the total parameter count, about \\(1.7\\%\\) and \\(6.9\\%\\). And they can contribute to BD-rate saving by \\(2.93\\%\\) and \\(4.35\\%\\) on CLIC Professional Valid, respectively. It shows that the proposed RAT and SAL are lightweight and efficient. When we combine these two modules, SAL will provide stronger prototypes for RAT, leading to better results, i.e., a saving of \\(8.50\\%\\) BD-rate." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.734, + 0.788, + 0.841 + ], + "angle": 0, + "content": "With our proposed modules, Ours-S achieves high performance with a significantly smaller parameter count. Compared to ChARM [34], Ours-S saves nearly \\(40\\%\\) parameters and has faster speed while achieving comparable performance. Ours-S is a CNN-based model without any self-attention module. And the entropy estimation is just the hyperprior-style [6], without the efficient but time-consuming auto-regressive architecture. It indicates that the proposed modules RAT and SAL are highly effective and lightweight." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.221, + 0.115, + 0.732, + 0.13 + ], + "angle": 0, + "content": "Region-Adaptive Transform with Segmentation Prior for Image Compression" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.145, + 0.41, + 0.247 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.295, + 0.248, + 0.338, + 0.257 + ], + "angle": 0, + "content": "(a) Kodak" + }, + { + "type": "image", + "bbox": [ + 0.412, + 0.146, + 0.597, + 0.248 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.475, + 0.248, + 0.551, + 0.258 + ], + "angle": 0, + "content": "(b) COCO-sample" + }, + { + "type": "image", + "bbox": [ + 0.6, + 0.146, + 0.782, + 0.248 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.641, + 0.248, + 0.762, + 0.258 + ], + "angle": 0, + "content": "(c) Validation loss in training" + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.27, + 0.788, + 0.34 + ], + "angle": 0, + "content": "Fig. 8: We measure the impact of the masks separately in both the training and testing phases. \"test w/ masks\" means we adopt the masks in both training and testing, \"test w/o masks\" means we adopt the masks only in training, while \"train w/o masks\" means we do not use masks all the time. Naturally, if the masks are not used, \\(4 \\times 4\\) grid partitions will be used as alternatives." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.371, + 0.416, + 0.385 + ], + "angle": 0, + "content": "4.4 Subjective Results" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.397, + 0.789, + 0.534 + ], + "angle": 0, + "content": "Fig. 7 shows the subjective results compared with VTM and Zou2022. As we can see, with the help of the segmentation masks in training, our SegPIC can distinguish the objects' contours more accurately, making the edges sharper with less bitrate. In \\(kodim04\\), we can see the wool reconstructed by SegPIC is more consistent. In \\(kodim24\\), the bird patterns decoded by SegPIC are more clear, and the blurrings are less than in other methods. These results are based on grid partitions used as alternatives to masks. It demonstrates that SegPIC has acquired a more robust representative capability for various objects. Consequently, SegPIC can discern objects in complex textures more effectively." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.557, + 0.627, + 0.572 + ], + "angle": 0, + "content": "4.5 Analysis of the Role of Segmentation Masks" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.583, + 0.788, + 0.658 + ], + "angle": 0, + "content": "Quantitative Results. To validate the role of the masks, we make experiments on two datasets, Kodak and COCO-sample. Due to Kodak not being a segmentation dataset, the corresponding masks are generated using the segmentation method [36]. Otherwise, we randomly sample 20 images and their masks from COCO Valid 2017 [9] to create the COCO-sample dataset." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.659, + 0.788, + 0.81 + ], + "angle": 0, + "content": "As illustrated in Fig. 8 (a) (b), during the testing (inference) phase, the differences caused by masks are not so evident. However, for training, the situation is entirely different. With the assistance of the masks, the model achieves significantly better performance, especially at low bitrates. It could imply that, due to bitrate loss constraints during training, the information in the prototypes is implicitly transferred to the learnable parameters in RAT, allowing limited contextual information to still be effective during the inference phase. At low bitrates, since the lack of sufficient high-frequency information, pixels in the region exhibit stronger similarity, enhancing the representational ability of features extracted by prototypes. Fig. 8 (c) shows the early stage of model training." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.81, + 0.787, + 0.84 + ], + "angle": 0, + "content": "As shown in Tab. 3, the model training w/o masks (the last column) leads to noticeable degradation. However, training w/ masks has a relatively minor" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.356, + 0.128 + ], + "angle": 0, + "content": "Y. Liu et al." + }, + { + "type": "image", + "bbox": [ + 0.22, + 0.144, + 0.788, + 0.234 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.243, + 0.788, + 0.327 + ], + "angle": 0, + "content": "Fig. 9: Visualization of the kernels generated by DKG in RAT at the Encoder (see Fig. 3), in the cases with and without the masks. Fig. (d)(e)(f) are generated by the same pre-trained network weights. During testing, they go through processes with or without masks, respectively. The distortions in (d) can be formulated as \\((x - \\hat{x})^2\\). We take the maximum absolute values of the tensors along the channels and apply 2D max pooling on them for better visualization." + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.343, + 0.784, + 0.371 + ], + "angle": 0, + "content": "Table 3: The effect of the masks and the grid number on BD-rate on COCO-sample. \"n×\" denotes n×n grid partitions to replace masks." + }, + { + "type": "table", + "bbox": [ + 0.289, + 0.384, + 0.716, + 0.445 + ], + "angle": 0, + "content": "
Traintrain w/ masks4×
Testtest w/ masks2×4×8×4×
BD-rate ↓-0.11+0.210(default)+0.46+2.10
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.474, + 0.788, + 0.549 + ], + "angle": 0, + "content": "impact on the pattern and number in testing. So we can replace the masks in testing to further save bitrate but should take the masks as privileged information during training. The hyperparameter \\( 4 \\times \\) in testing is an empirical trade-off between the effect of RAT (i.e. image reconstruction precision) and the quantity of the prototypes (i.e. the additional bitrate they introduced)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.551, + 0.788, + 0.657 + ], + "angle": 0, + "content": "Qualitative Analysis. Fig. 9 shows the visualization of the kernels generated by DKG in RAT at the Encoder (see Fig. 2). Although there is not a noticeable difference in Fig. 9 (e), it is obvious that the gradients of \\( w / \\text{mask} \\) are more focused on the semantic objects (see Fig. 9 (f)), which is complex and prone to causing distortions. This attribute will assist in optimizing the models, enabling it to better reconstruct various objects. These observations indicate that the masks do play a necessary role during the training phase." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.684, + 0.36, + 0.7 + ], + "angle": 0, + "content": "5 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.788, + 0.841 + ], + "angle": 0, + "content": "We propose a novel image compression framework called Segmentation-Prior-Guided Image Compression (SegPIC), equipped with two proposed modules: Region-Adaptive Transform (RAT) and Scale Affine Layer (SAL). We adopt the class-agnostic masks as privilege information to assist the training of SegPIC. Experimental results demonstrate that our SegPIC is superior to the previous works. Otherwise, we analyze the role of the masks. Experiments and visualization results show that the masks can guide the model to focus more on semantic objects in training, leading to higher pixel-fidelity performance." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.221, + 0.115, + 0.732, + 0.13 + ], + "angle": 0, + "content": "Region-Adaptive Transform with Segmentation Prior for Image Compression" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.403, + 0.163 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.179, + 0.788, + 0.24 + ], + "angle": 0, + "content": "This work was supported in part by the National Key R&D Program of China (No.2021ZD0112100), National NSF of China (No.62120106009), Guangdong Basic and Applied Basic Research Foundation (2024A1515010454), the Basic and Frontier Research Project of PCL, and the Major Key Project of PCL." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.264, + 0.324, + 0.279 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.296, + 0.784, + 0.325 + ], + "angle": 0, + "content": "1. Workshop and challenge on learned image compression (clic2020) (2020), http://www.compression.cc 10, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.326, + 0.785, + 0.367 + ], + "angle": 0, + "content": "2. Agustsson, E., Minnen, D., Toderici, G., Mentzer, F.: Multi-realism image compression with a conditional generator. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 22324-22333 (2023) 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.368, + 0.785, + 0.394 + ], + "angle": 0, + "content": "3. Ahmed, N., Natarajan, T., Rao, K.R.: Discrete cosine transform. IEEE Transactions on Computers 100(1), 90-93 (1974) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.395, + 0.785, + 0.422 + ], + "angle": 0, + "content": "4. Akbari, M., Liang, J., Han, J.: Dsslic: Deep semantic segmentation-based layered image compression. In: ICASSP. pp. 2042-2046. IEEE (2019) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.423, + 0.785, + 0.45 + ], + "angle": 0, + "content": "5. Balle, J., Laparra, V., Simoncelli, E.P.: End-to-end optimized image compression. In: Int. Conf. Learn. Represent. (2017) 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.451, + 0.785, + 0.49 + ], + "angle": 0, + "content": "6. Balle, J., Minnen, D., Singh, S., Hwang, S.J., Johnston, N.: Variational image compression with a scale hyperprior. In: Int. Conf. Learn. Represent. (2018) 2, 4, 5, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.492, + 0.751, + 0.507 + ], + "angle": 0, + "content": "7. Bellard, F.: Bpg image format (2014), https://bellard.org/bpg/ 2, 10, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.508, + 0.785, + 0.534 + ], + "angle": 0, + "content": "8. Bjontegaard, G.: Calculation of average psnr differences between rd-curves. ITU SG16 Doc. VCEG-M33 (2001) 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.535, + 0.785, + 0.562 + ], + "angle": 0, + "content": "9. Caesar, H., Uijlings, J., Ferrari, V.: COCO-Stuff: Thing and stuff classes in context. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 1209-1218 (2018) 10, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.563, + 0.785, + 0.603 + ], + "angle": 0, + "content": "0. Chang, J., Mao, Q., Zhao, Z., Wang, S., Wang, S., Zhu, H., Ma, S.: Layered conceptual image compression via deep semantic synthesis. In: IEEE Int. Conf. Image Process. pp. 694-698 (2019) 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.604, + 0.785, + 0.644 + ], + "angle": 0, + "content": "11. Chang, J., Zhao, Z., Yang, L., Jia, C., Zhang, J., Ma, S.: Thousand to one: Semantic prior modeling for conceptual coding. In: Int. Conf. Multimedia and Expo. pp. 1-6. IEEE (2021) 3, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.645, + 0.785, + 0.687 + ], + "angle": 0, + "content": "12. Cheng, Z., Sun, H., Takeuchi, M., Katto, J.: Learned image compression with discretized gaussian mixture likelihoods and attention modules. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 7939-7948 (2020) 3, 4, 10, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.688, + 0.785, + 0.715 + ], + "angle": 0, + "content": "13. Chollet, F.: Xception: Deep learning with depthwise separable convolutions. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 1251-1258 (2017) 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.716, + 0.785, + 0.743 + ], + "angle": 0, + "content": "14. Daubechies, I.: The wavelet transform, time-frequency localization and signal analysis. IEEE Transactions on Information Theory 36(5), 961-1005 (1990) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.744, + 0.785, + 0.799 + ], + "angle": 0, + "content": "15. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., Uszkoreit, J., Houlsby, N.: An image is worth 16x16 words: Transformers for image recognition at scale. In: Int. Conf. Learn. Represent. (2021) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.799, + 0.785, + 0.84 + ], + "angle": 0, + "content": "16. Feng, R., Gao, Y., Jin, X., Feng, R., Chen, Z.: Semantically structured image compression via irregular group-based decoupling. In: Int. Conf. Comput. Vis. pp. 17237-17247 (2023) 3, 5, 10" + }, + { + "type": "list", + "bbox": [ + 0.22, + 0.296, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.356, + 0.128 + ], + "angle": 0, + "content": "Y. Liu et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.787, + 0.203 + ], + "angle": 0, + "content": "17. Goodfellow, I., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., Courville, A., Bengio, Y.: Generative adversarial nets. In: Ghahramani, Z., Welling, M., Cortes, C., Lawrence, N., Weinberger, K. (eds.) Adv. Neural Inform. Process. Syst. vol. 27. Curran Associates, Inc. (2014) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.204, + 0.787, + 0.245 + ], + "angle": 0, + "content": "18. He, D., Yang, Z., Peng, W., Ma, R., Qin, H., Wang, Y.: Elic: Efficient learned image compression with unevenly grouped space-channel contextual adaptive coding. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 5718-5727 (2022) 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.246, + 0.787, + 0.286 + ], + "angle": 0, + "content": "19. He, D., Zheng, Y., Sun, B., Wang, Y., Qin, H.: Checkerboard context model for efficient learned image compression. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 14771-14780 (2021) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.287, + 0.787, + 0.314 + ], + "angle": 0, + "content": "20. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 770-778 (2016) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.315, + 0.787, + 0.355 + ], + "angle": 0, + "content": "21. Hoang, T.M., Zhou, J., Fan, Y.: Image compression with encoder-decoder matched semantic segmentation. In: IEEE Conf. Comput. Vis. Pattern Recog. Worksh. pp. 160-161 (2020) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.356, + 0.787, + 0.384 + ], + "angle": 0, + "content": "22. Jia, X., De Brabandere, B., Tuytelaars, T., Gool, L.V.: Dynamic filter networks. Adv. Neural Inform. Process. Syst. 29 (2016) 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.384, + 0.787, + 0.424 + ], + "angle": 0, + "content": "23. Jiang, W., Yang, J., Zhai, Y., Ning, P., Gao, F., Wang, R.: Mlic: Multi-reference entropy model for learned image compression. In: ACM Int. Conf. Multimedia. pp. 7618-7627 (2023) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.425, + 0.787, + 0.452 + ], + "angle": 0, + "content": "24. (JVET), J.V.E.T.: Versatile video coding (2021), https://jvet.hhi.fraunhofer.de/ 2, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.453, + 0.787, + 0.479 + ], + "angle": 0, + "content": "25. Kingma, D.P., Ba, J.: Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980 (2014) 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.48, + 0.787, + 0.508 + ], + "angle": 0, + "content": "26. Kingma, D.P., Welling, M.: Auto-encoding variational bayes. In: Int. Conf. Learn. Represent. (2014) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.509, + 0.787, + 0.536 + ], + "angle": 0, + "content": "27. Kodak, E.: Kodak lossless true color image suite (photocd pcd0992) (1993), http://r0k.us/graphics/kodak/ 10, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.537, + 0.787, + 0.59 + ], + "angle": 0, + "content": "28. Koyuncu, A.B., Gao, H., Boev, A., Gaikov, G., Alshina, E., Steinbach, E.: Contextformer: A transformer with spatio-channel attention for context modeling in learned image compression. In: Eur. Conf. Comput. Vis. pp. 447-463. Springer (2022) 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.591, + 0.787, + 0.632 + ], + "angle": 0, + "content": "29. Li, F., Zhang, L., Liu, Z., Lei, J., Li, Z.: Multi-frequency representation enhancement with privilege information for video super-resolution. In: Int. Conf. Comput. Vis. pp. 12814-12825 (2023) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.633, + 0.787, + 0.661 + ], + "angle": 0, + "content": "30. Liu, J., Lu, G., Hu, Z., Xu, D.: A unified end-to-end framework for efficient deep image compression. arXiv preprint arXiv:2002.03370 (2020) 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.661, + 0.787, + 0.702 + ], + "angle": 0, + "content": "31. Liu, J., Sun, H., Katto, J.: Learned image compression with mixed transformer-cnn architectures. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 14388-14397 (2023) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.702, + 0.787, + 0.743 + ], + "angle": 0, + "content": "32. Liu, Z., Lin, Y., Cao, Y., Hu, H., Wei, Y., Zhang, Z., Lin, S., Guo, B.: Swin transformer: Hierarchical vision transformer using shifted windows. In: Int. Conf. Comput. Vis. pp. 10012-10022 (2021) 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.744, + 0.787, + 0.771 + ], + "angle": 0, + "content": "33. Minnen, D., Balle, J., Toderici, G.D.: Joint autoregressive and hierarchical priors for learned image compression. Adv. Neural Inform. Process. Syst. 31 (2018) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.772, + 0.787, + 0.812 + ], + "angle": 0, + "content": "34. Minnen, D., Singh, S.: Channel-wise autoregressive entropy models for learned image compression. In: IEEE Int. Conf. Image Process. pp. 3339-3343. IEEE (2020) 4, 5, 9, 10, 11, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.813, + 0.787, + 0.84 + ], + "angle": 0, + "content": "35. Pan, G., Lu, G., Hu, Z., Xu, D.: Content adaptive latents and decoder for neural image compression. In: Eur. Conf. Comput. Vis. pp. 556-573. Springer (2022) 10" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.221, + 0.115, + 0.732, + 0.13 + ], + "angle": 0, + "content": "Region-Adaptive Transform with Segmentation Prior for Image Compression" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.175 + ], + "angle": 0, + "content": "36. Qi, L., Kuen, J., Wang, Y., Gu, J., Zhao, H., Torr, P., Lin, Z., Jia, J.: Open world entity segmentation. IEEE Trans. Pattern Anal. Mach. Intell. (2022) 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.177, + 0.785, + 0.203 + ], + "angle": 0, + "content": "37. Shen, H., Zhao, Z.Q., Zhang, W.: Adaptive dynamic filtering network for image denoising. In: AAAI. vol. 37, pp. 2227-2235 (2023) 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.204, + 0.785, + 0.231 + ], + "angle": 0, + "content": "38. Skodras, A., Christopoulos, C., Ebrahimi, T.: The JPEG 2000 still image compression standard. IEEE Signal Processing Magazine 18(5), 36-58 (2001) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.232, + 0.785, + 0.259 + ], + "angle": 0, + "content": "39. Stark, H., Woods, J.W.: Probability, random processes, and estimation theory for engineers. Prentice-Hall, Inc. (1986) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.26, + 0.785, + 0.299 + ], + "angle": 0, + "content": "40. Sun, S., He, T., Chen, Z.: Semantic structured image coding framework for multiple intelligent applications. IEEE Trans. Circuit Syst. Video Technol. 31(9), 3631-3642 (2020) 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.3, + 0.785, + 0.342 + ], + "angle": 0, + "content": "41. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. Adv. Neural Inform. Process. Syst. 30 (2017) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.343, + 0.787, + 0.37 + ], + "angle": 0, + "content": "42. Wallace, G.K.: The JPEG still picture compression standard. Communications of the ACM 34(4), 30-44 (1991) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.371, + 0.785, + 0.41 + ], + "angle": 0, + "content": "43. Wang, D., Yang, W., Hu, Y., Liu, J.: Neural data-dependent transform for learned image compression. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 17379-17388 (2022) 2, 3, 4, 10, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.411, + 0.785, + 0.439 + ], + "angle": 0, + "content": "44. Wang, G.H., Li, J., Li, B., Lu, Y.: Evc: Towards real-time neural image compression with mask decay. In: Int. Conf. Learn. Represent. (2023) 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.44, + 0.785, + 0.48 + ], + "angle": 0, + "content": "45. Wang, X., Yu, K., Dong, C., Loy, C.C.: Recovering realistic texture in image superresolution by deep spatial feature transform. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 606-615 (2018) 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.481, + 0.785, + 0.522 + ], + "angle": 0, + "content": "46. Wang, Z., Simoncelli, E.P., Bovik, A.C.: Multiscale structural similarity for image quality assessment. In: The Thirty-Seventh Asilomar Conference on Signals, Systems & Computers, 2003. vol. 2, pp. 1398-1402. IEEE (2003) 4, 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.523, + 0.785, + 0.564 + ], + "angle": 0, + "content": "47. Xu, Y.S., Tseng, S.Y.R., Tseng, Y., Kuo, H.K., Tsai, Y.M.: Unified dynamic convolutional network for super-resolution with variational degradations. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 12496-12505 (2020) 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.565, + 0.785, + 0.592 + ], + "angle": 0, + "content": "48. Zhu, Y., Yang, Y., Cohen, T.: Transformer-based transform coding. In: Int. Conf. Learn. Represent. (2021) 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.592, + 0.785, + 0.633 + ], + "angle": 0, + "content": "49. Zou, R., Song, C., Zhang, Z.: The devil is in the details: Window-based attention for image compression. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 17492-17501 (2022) 3, 4, 5, 7, 9, 10, 11, 12" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.633 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/Region-Adaptive Transform with Segmentation Prior for Image Compression/7ddd6f0f-062f-4fcd-9635-2169878bc97b_origin.pdf b/2024/Region-Adaptive Transform with Segmentation Prior for Image Compression/7ddd6f0f-062f-4fcd-9635-2169878bc97b_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..89a6a43f1c62d50f1b971e861095b6325677fbbe --- /dev/null +++ b/2024/Region-Adaptive Transform with Segmentation Prior for Image Compression/7ddd6f0f-062f-4fcd-9635-2169878bc97b_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:720cc2657acbf85d7630663edbd3ec13d59bd0ed90406276e2a3f2e783f2c7e9 +size 1372622 diff --git a/2024/Region-Adaptive Transform with Segmentation Prior for Image Compression/full.md b/2024/Region-Adaptive Transform with Segmentation Prior for Image Compression/full.md new file mode 100644 index 0000000000000000000000000000000000000000..5d76e40b69f0e95f3b460e46ea3bdd3188a50300 --- /dev/null +++ b/2024/Region-Adaptive Transform with Segmentation Prior for Image Compression/full.md @@ -0,0 +1,352 @@ +# Region-Adaptive Transform with Segmentation Prior for Image Compression + +Yuxi Liu $^{1,2}$ , Wenhan Yang $^{3}$ , Huihui Bai $^{1,2}$ , Yunchao Wei $^{1,2,3}$ , and Yao Zhao $^{\star 1,2,3}$ + +$^{1}$ Institute of Information Science, Beijing Jiaotong University, Beijing, China + $^{2}$ Visual Intelligence + X International Joint Laboratory of the Ministry of Education + $^{3}$ PengCheng Laboratory, Shenzhen, China +{yuxiliu, hhbai, yunchao.wei, yzhao}@bjtu.edu.cn, yangwh@pcl.ac.cn + +Abstract. Learned Image Compression (LIC) has shown remarkable progress in recent years. Existing works commonly employ CNN-based or Transformer-based modules as transform methods for compression. However, there is no prior research on neural transform that focuses on specific regions. In response, we introduce the class-agnostic segmentation masks (i.e. semantic masks without category labels) for extracting region-adaptive contextual information. Our proposed module, Region-Adaptive Transform, applies adaptive convolutions on different regions guided by the masks. Additionally, we introduce a plug-and-play module named Scale Affine Layer to incorporate rich contexts from various regions. While there have been prior image compression efforts that involve segmentation masks as additional intermediate inputs, our approach differs significantly from them. Our advantages lie in that, to avoid extra bitrate overhead, we treat these masks as privilege information, which is accessible during the model training stage but not required during the inference phase. To the best of our knowledge, we are the first to employ class-agnostic masks as privilege information and achieve superior performance in pixel-fidelity metrics, such as Peak Signal to Noise Ratio (PSNR). The experimental results demonstrate our improvement compared to previously well-performing methods, with about $8.2\%$ bitrate saving compared to VTM-17.0. The source code is available at https://github.com/GityuxiLiu/SegPIC-for-Image-Compression. + +Keywords: Learned Image Compression $\cdot$ Segmentation Prior + +# 1 Introduction + +With the advent of the multimedia era, an ever-growing volume of image data has emerged on the internet. These vast volumes of image data demand larger transmission bandwidth and storage capacity. In this context, the importance of image compression technology has become even more pronounced, evoking greater efforts in efficient image compression techniques. Lossy image compres + +![](images/0e4b5631c4dae19c2369313434f77ba00505965d04a1f677608d5c2e2f704eaf.jpg) +(a) Convictional LIC framework + +![](images/c0c76c0d31e5f465ee91214d749e0aaf0cf49a4e22189e98f43822733d9c0928.jpg) +(b) Data-dependent LIC framework + +![](images/c10f29b94376026738b0c62e9af84326aafc84815ee25a2a130a861a6da5fdc6.jpg) +(c) Proposed LIC framework +Fig. 1: Paradigms of end-to-end learned image compression (LIC) methods. (a) Conventional LIC with fixed learned kernels. (b) Data-dependent LIC with image-specific kernels [43]. (c) The proposed method is capable of generating region-adaptive kernels from different regions, providing more fine-grained transforms. + +sion technology can reduce the data size to a lower level while maintaining good visual quality. In the past decades, the classical lossy image compression methods such as JPEG [42], JPEG2000 [38], BPG [7], and VVC [24], used to have excellent performance and were widely adopted. These methods include modules of transform, quantization, and entropy coding. The transform process aims to convert images into a more compact coefficient by removing pixel correlations as much as possible. The commonly employed transforms include the discrete cosine transform (DCT) [3] and wavelet transform [14]. These transformations are linear, characterized by simplicity and efficiency. However, their fixed forms are unable to adapt to the contexts of complex signals. To address this issue, latter efforts propose data-driven transformation approaches, e.g. the Karhunen-Loève transform (KLT) [39]. Furthermore, the subsequent works are dedicated to constructing complex patterns for the adaptive selection of local transforms, improving adaptability and overall performance. + +Benfit from the advancement of neural network technology, many learned image compression (LIC) methods have been proposed, which shows a promising alternative for designing transform: learning from data instead of handcraft design. Balle et al. [5, 6] proposed the first well-known end-to-end framework. They utilize variational auto-encoders (VAE) [26] as a transform module and optimize the parameters of the kernels to better align with the natural data distribution. To further improve the representation capabilities of the transform + +module, some works [12, 30, 43, 48, 49] design advanced modules to introduce richer contextual information, e.g. residual blocks [20] and self-attention mechanism [15, 32]. These works optimize learnable kernels to fit the data distribution to attain a more powerful transform. However, their kernel parameters are fixed after training (see Fig.1 (a)). Wang et al. [43] introduce a neural data-dependent transform, which extracts dynamic convolution kernels based on the context of the entire image for more adaptive transform (see Fig. 1 (b)). However, even within a single image, the context could vary significantly across different regions, e.g. stone and sheep regions in Fig. 1. It limits the description of the signals to a more fine-grained level using such image-level transforms. + +To address the issue, this paper makes the first efforts to design learnable region-adaptive transform to capture the mapping regularity in different regions. In detail, we employ class-agnostic4 segmentation masks to extract region-specific information for transform (see Fig. 1 (c)). Diffs from previous works [4, 11, 16, 21], our approach stands out with two characteristics in mask utilization. First, our masks are "class-agnostic", where the semantic labels are abandoned but a more effective compression-driven semantic prior will be learned by our model in a data-driven manner. Second, to avoid incurring additional costs in both bitrate and computational resources, we treat these masks as privilege information, which are available during training but remain inaccessible during the inference phase [29]. Specifically, we replace the masks with uniform grid partitions (see Fig. 9 (c)) in inference, and the experimental result shows approximate performance. It implies that our model has learned useful semantic knowledge from the masks in training, incorporating this knowledge into the model (i.e. the network's parameters). Hence, it remains capable of mining contextual information even from relatively coarse partition regions, i.e., the given uniform grid partitions. This simple operation can effectively maintain the consistency of our model's architecture and parameters, whether or not masks are used at different phases. The more details about the difference between our method and previous semantic-mask-based works will be explained at Sec. 2. + +Based on the above motivations, we propose a novel image compression framework called Segmentation-Prior-Guided Image Compression (SegPIC), equipped with two key modules: Region-Adaptive Transform (RAT) and Scale Affine Layer (SAL). The RAT adopts region-adaptive kernels for transform, performing the convolution whose weights are generated from the prototypes and the context of latent features under the guidance of masks. The SAL affines the latent features for better mining of semantic contexts in the Encoder and Decoder. Experimental results show that SegPIC is superior to previously well-performing image compression methods. Our main contributions are as follows: + +- We propose a novel image compression framework called Segmentation-Prior-Guided Image Compression (SegPIC) to generate more effective region-adaptive transforms in a data-driven manner. + +- Our SegPIC is equipped with two proposed modules: Region-Adaptive Transform (RAT) and Scale Affine Layer (SAL). The RAT adopts region-adaptive kernels for transform, while the SAL affines the latent features for better mining of semantic contexts. +- We utilize the class-agnostic masks as privilege information to assist the training of SegPIC. The "class-agnostic" property promotes the model to learn a compression-driven semantic prior, while the "privilege" way avoids additional bitrate and complexity for masks in inference. +- Experimental results demonstrate that our SegPIC is superior to previously well-performing image compression methods. We also present sufficient ablation experiments to confirm the necessity of the privilege information. + +# 2 Related Works + +Learned Image Compression. Image compression can be divided into three main components: transform, quantization, and entropy coding. In brief, the transformation module maps the image to a latent feature space. Subsequently, quantization is performed after this phase, resulting in some quantization distortions but achieving lower bitrate. The entropy encoding module estimates the data distribution of the latent features and inputs it to the Arithmetic Coding (AC) module, which generates the bitstream for transmission. + +To enhance transform module, some works [30,43,48,49] adopt more powerful modules like self-attention mechanism [32, 41] to explore broader context. [31] proposed a transform block that mixes CNN and Transformer to combine their advantages. As for the entropy estimation module, the studies [12, 19, 33, 34] usually focus on exploring novel auto-regressive methods to split the latent features into several portions to be decoded in order. The portions having been decoded will assist in decoding the next portion, through the spatial or channel correlation of these portions. This auto-regressive mechanism may introduce additional latency but achieve better performance than Hyper Prior [6]. Recently, [23] integrates different types of correlations present in latent representation, e.g. channel-wise, local spatial, and global spatial correlations. Their proposed novel auto-regressive entropy module achieve powerful performance. + +Thanks to these great works, the performance of LIC has become so strong that achieving further enhancements becomes progressively challenging. Therefore, some new directions in compression have emerged, which is also meaningful on certain occasions. For example, some semantic-based compression methods [10, 11] adopt GAN [17] to compress images at extremely low bitrate. These methods maintain a good human visual perception effect but usually, poor pixel-fidelity metrics, e.g. Peak Signal to Noise Ratio (PSNR) and Multi-Scale Structural Similarity Index (MS-SSIM) [46]. Specifically, these methods could introduce some synthetic artifacts with low fidelity. For example, a semantic compression method could reconstruct an image of a bird with different breeds from the original image because the training category labels do not distinguish the bird's breed. Another direction is compression for machines, which means compress + +![](images/9cc7ffa734bebf9d81b607c1f20b9fe6ce6033197cf401b6541bbec60fa0f014.jpg) +(a) The overall framework + +![](images/17cd25dd45547abe1efde816681db3f4bf91e02afb712ae2719c419e87bd3f0b.jpg) +(b)Downsample block (out channel C) + +![](images/5bf1028832daf3fd0fbde7ec4964a20e6844baeb5d5bff68d20b0470d47fadfd.jpg) +(c) Upsample block (out channel C) +Fig. 2: (a) Overall framework of our SegPIC. The blue flow lines represent the decoding process on the decoder side. The dashed lines represent the process of extracting the prototypes $p'$ , $p$ (see Eq. 5). The compression of $p'$ , $p$ is not fully depicted for simplicity, please refer to Fig. 4 for details. $N$ is set to 192, and $M$ is 320 as in [49]. RAT is the proposed Region-Adaptive Transform. SAL is the proposed Scale Affine Layer. WAM is Window Attention Module [49]. ChARM is Channel-wise Auto-Regressive Model [34]. FM is a Factorized Model [6]. GDN is Generalized Divisive Normalization [5]. (b) is Downsample Block. "Downsample, N" in (a) means the block with out-channel N. "Conv C,5,2" means out-channel C, kernel size 5, stride 2. (c) is Upsample Block, and "TConv" is Transposed Convolution. + +ing images for downstream tasks [16, 40]. These methods focus on balancing the bitrate and performance of downstream tasks, not just the pixel-fidelity metrics. + +Although some works [10, 11, 16] about the two directions also adopt masks, our usage is significantly different. The main differences are as follows: 1) We utilize the masks as privilege information, without the mask information during the inference phase, while usually, the masks are indispensable for other methods. 2) We employ class-agnostic masks to guide the model to extract rich compression-friendly contexts among semantically related regions, regardless of the category that might degrade the fidelity as mentioned above. 3) The proposed model is optimized towards better pixel-fidelity metrics, e.g. PSNR and MS-SSIM. + +Region-adaptive Convolution. Due to its capacity to model the degradation levels in different regions of images, region-adaptive convolution is applied to + +some low-level vision tasks such as super-resolution and denoising [37, 47]. In terms of compression, the reconstruction quality varies across different regions. For instance, regions with complex textures are more challenging compared to those with simpler patterns. Therefore, specific processing can be applied to different regions based on their characteristics. The regions can be divided using segmentation masks or partitions based on certain regular patterns. To the best of our knowledge, we are the first to introduce this mechanism on image compression for optimizing pixel-fidelity metrics. + +# 3 Proposed SegPIC + +# 3.1 Motivation + +As mentioned above, previous transforms either adopt convolutions with fixed parameters or image-specific convolutions. Our motivation is to construct a region-adaptive transform via the following three properties: + +- Region-dependent: we propose a neural network with region-adaptive transforms guided by the segmentation masks. +- Class-Agnostic: the adopted mask does not include a category label, allowing the model to be free to learn compression-friendly semantic priors. +- Privilege Information: the masks are expected as privilege information, not being used in inference for saving bitrate and computation resources. + +Based on these considerations, we propose our SegPIC in the VAE framework incorporated with two proposed modules: SAL and RAT. + +- SAL: First of all, we extract a prototype from each region, which will be compressed and transmitted to be accessible on both the encoder and decoder sides. The prototypes are expected to extract high-level information from regions to ensure the representative ability as much as possible. However, the encoder in compression frameworks often extracts features that contain low-level information. Therefore, a module is highly desirable for communicating information across different levels. SAL is designed for extracting more abundant semantics and contexts, enriching the prototypes, and connecting features from two different levels. + +- RAT: performing the Depth&Point-wise Separable Convolution (DPSCov), a novel convolution whose weights are generated from the prototypes and the context of latent features under the guidance of masks. Note that, we only utilize the masks to guide RAT to learn more compression-friendly semantic priors in training, while during the inference phase, we instead adopt uniform grid partitions as coarse alternatives. + +# 3.2 Formulation + +In this section, we will introduce the problem and our method in mathematical form. Given an image $x$ , the encoder $En$ maps it to latent features $y$ . After + +quantization $Q$ , $\hat{y}$ will be compressed to the bitstream by the entropy coding, such as arithmetic coding (AC). Then the receiver converts the bitstream back into $\hat{y}$ by entropy decoding. The above process can be formulated as follows: + +$$ +y = E n (x), \quad \hat {y} = Q (y), +$$ + +Transmitter: $bit_{\hat{y}} = AE(\hat{y} |p_{\hat{y}})$ (1) + +Receiver: $\hat{y} = AD(bit_{\hat{y}}|p_{\hat{y}})$ + +where $p_{\hat{y}}$ is the estimated distribution of $\hat{y}$ , $\text{bit}_{\hat{y}}$ is the binary bitstream of $\hat{y}$ , and $AE$ and $AD$ are arithmetic encoding and decoding. We use a single Gaussian distribution to estimate $\hat{y} \sim \mathcal{N}(\mu, \sigma^2)$ , and the mean and variance value $\mu, \sigma^2$ will be yielded through an auto-regressive mechanism. The formulation is: + +$$ +z = E n _ {z} (y), \hat {z} = Q (z), \quad \mu_ {i}, \sigma_ {i} = F \left(D e _ {z} (\hat {z}), \hat {y} _ {< i}\right), \tag {2} +$$ + +where $En_{z}, De_{z}$ is hyper encoder and decoder, $\mu_{i}, \sigma_{i}$ are the mean and variance of $\hat{y}_{i} \sim \mathcal{N}(\mu_{i}, \sigma_{i}^{2})$ , $\hat{y}_{i}$ is the currently decoded portion of $\hat{y}$ , $\hat{y}_{< i}$ is the already decoded portion to assist decoding $\hat{y}_{i}$ , and $F$ is the auto-regressive entropy model. The AC process of $z$ in Eq. 2 is omitted for brevity. Finally, the receiver adopts decoder $De$ to reconstruct the image $\hat{x}$ , which is formulated as $\hat{x} = De(\hat{y})$ . + +The loss function $\mathcal{L}$ to optimize this problem is: + +$$ +\mathcal {L} = R + \lambda \cdot D = R (\hat {y}) + R (\hat {z}) + \lambda \cdot \mathbb {E} _ {x \sim p _ {x}} [ d (x, \hat {x}) ], \tag {3} +$$ + +where $R(\hat{y}), R(\hat{z})$ are the bitrate of $\hat{y}, \hat{z}$ , $d(x, \hat{x})$ is the distortion between $x$ and $\hat{x}$ , $\lambda$ denotes different rate-distortion trade-offs. Here we use Mean Square Error (MSE) or (1-MS-SSIM) as the distortion function. + +In our implementation, we use class-agnostic semantic masks (denoted as $m$ ) as additional inputs of the encoding $En$ during the training phase. Moreover, we extract a prototype from each region at the middle layer of $En$ , then compress and transmit them (denoted as $\hat{p}$ ). In inference, we replace the masks with uniform grid partitions, so there is no additional bitstream for masks, while the bitrates of the prototypes also need to be taken into account. The $En$ and $De$ will transform the image with the assistance of the prototypes. The formulations requiring adjustment are as follows: + +$$ +y = E n (x, \hat {p} | m), \quad \hat {x} = D e (\hat {y}, \hat {p} | m), +$$ + +$$ +\mu_ {i}, \sigma_ {i} = F (D e _ {z} (\hat {z}), \hat {y} _ {< i}, \hat {p} | m), \tag {4} +$$ + +$$ +R = R (\hat {p}) + R (\hat {y}) + R (\hat {z}). +$$ + +# 3.3 Network Design + +In this section, we will introduce the concrete process and our network design. We use [49] as the backbone. The overall architecture is in Fig. 2. + +Scale Affine Layer. The Encoder and Decoder are responsible for image transform and reconstruction, each comprising 4 downsampling or upsampling blocks. + +The previous sampling blocks consisted of a single sample convolution layer, aimed at modifying the data size. However, this may not be sufficient to extract adequate semantics and contexts. Inspired by Spatial Feature Transform (SFT) [45], which can generate visually pleasing textures in segmentation-guided low-level vision, we propose to add a Scale Affine Layer (SAL) between two adjacent convolution layers (see Fig. 2 (b) (c)). It consists of two $1 \times 1$ convolution layers with a GELU activation function in between, which is formulated as: + +$$ +\begin{array}{l} X _ {s} = \operatorname {S A L} _ {s} (X), \quad X _ {b} = \operatorname {S A L} _ {b} (X), \\ Y = X _ {s} \otimes X + X _ {b}, \\ \end{array} +$$ + +where $X$ and $Y$ represent the input and output, $X_{s}$ and $X_{b}$ denote the scales and biases generated by SAL, $\otimes$ denotes the element-wise product. + +Extract and Transmit the Prototypes. The Encoder first converts the original image $x \in \mathbb{R}^{3 \times H \times W}$ into latent features $y' \in \mathbb{R}^{N \times H/4 \times W/4}$ at the middle layer. Then we conduct Masked Average Pooling (MAP) on $y'$ through the class-agnostic masks $m \in \{0,1\}^{n \times H \times W}$ (see Fig. 3 (a)), to acquire the prototypes $p' \in \mathbb{R}^{N \times n}$ , where $n$ is the number of the masks. Thanks to the proposed SAL, the Encoder becomes deeper, allowing $p'$ to incorporate more semantics and contexts from the corresponding mask. The formulation of MAP is as follows: + +$$ +p ^ {\prime} = \operatorname {M A P} \left(y ^ {\prime}, m\right) = \left\{\frac {1}{\sum m ^ {(i)}} \sum y ^ {\prime} \otimes m ^ {(i)}, \dots \right\} _ {i \in [ 1, n ]}, \tag {5} +$$ + +where $y'$ are the mid-level features (see Fig. 2 (a)), $\sum$ denotes summation along the spatial dimension. $p'$ will be compressed and transmitted to the receiver (see Fig. 4), and the reconstructed $\hat{p}'$ is shared between the Encoder and Decoder. The channel $\{C_1, C_2, C_3, C_4\}$ is {192, 128, 96, 96} for $p'$ (on Encoder and Decoder) and {320, 192, 128, 96} for $p$ (on Entropy Model). + +Region-Adaptive Transform. Here we propose a novel convolution called Depth&Point-wise Separable Convolution (DPSCov), in which each element has an individual convolution kernel, and there is no channel-wise mutual communication for computational efficiency. + +Before introducing it, we will start with standard convolution briefly. The input of standard convolution can be denoted as $X \in \mathbb{R}^{C \times H \times W}$ , the output is $Y \in \mathbb{R}^{O \times H \times W}$ , and the kernel is $W \in \mathbb{R}^{O \times C \times k \times k}$ , where $k$ is kernel size. The formulation of standard convolution is as follows: + +$$ +Y _ {o, h, w} = \sum_ {c = 1} ^ {C} X _ {c, h, w} * W _ {o, c} = \sum_ {c = 1} ^ {C} \sum_ {i, j} X _ {c, h + i, w + j} \cdot W _ {o, c, i, j}, +$$ + +where $*$ is 2D convolution, $i,j\in [-\lfloor k / 2\rfloor ,\lfloor k / 2\rfloor ]$ , $k$ is kernel size, and $\lfloor \cdot \rfloor$ denotes floor rounding. + +DPSCov can be seen as a combination of Depthwise Separable Convolution (DWConv) [13] and Dynamic Convolution [22]. The formulation is as follows: + +$$ +Y _ {c, h, w} = X _ {c, h, w} * W _ {c, h, w} = \sum_ {i, j} X _ {c, h + i, w + j} \cdot W _ {c, h, w, i, j}. \tag {6} +$$ + +![](images/8cee4801ee1779f81fd5971f4525b7ddbd395b93ef9a6ae76d5ce8a6a4fef5d5.jpg) +(a) Extract the prototypes through Masked Average Pooling + +![](images/5e12f059022ed5e0379e315379cb45c246c1bec341489f0cee517a70babe7e81.jpg) +(b)The Region-Adaptive Transform (RAT) with masks and prototypes +Fig. 3: The diagram of extracting the prototypes and the proposed RAT module. DPSCov is proposed Depth&Point-wise Separable Convolution (see Eq. 6). CTL is Channel Transform Layers, DKG is the DPSCov Kernel Generator, and CAG is the Channel Attention Generator. DPSCov is Depth&Point-wise Separable Convolution. The detailed modules are represented in Fig. 5. + +The computational complexity of standard convolution is $\mathrm{O}(O \times C \times H \times W \times k^2)$ , while DPSCov is $\mathrm{O}(C \times H \times W \times k^2)$ , same as DWConv. + +The overall design of RAT is depicted in Fig. 3 (b). We expand the prototypes based on the positional information of the masks, filling in the corresponding copies of the prototypes at each pixel position. Then, we feed-forward the input through Channel Transform Layers (CTL) and concatenate it with the expanded features. The resulting variable, which contains both semantics and contexts, is used for subsequent filter generation. The DPSCConv Kernels Generator (DKG) is designed for generating the filters. In the first layer of DKG, we employ a 1x1 convolution to attempt the fusion of the prototype and the pixel's intrinsic attribute. We opt for grouped convolution in the final layer of the DKG to reduce the computation and parameters. Then Channel Attention Generator (CAG) generates the channel attention to be multiplied with the output of DPSCConv along the channel dimension. Finally, the output passes through a 1x1 convolution to merge channel information. We apply the RAT in both the Encoder and Decoder, as well as the entropy encoding module. + +Entropy Encoding Module. We use the Channel-wise Auto-Regressive Model (ChARM) [34] as our entropy estimation module, and the specific hyperparameter configuration follows [49]. In this model, the latent variable $y$ is partitioned into 10 groups along the channel dimension. The Gaussian prior distribution $p(y_{i}|\hat{z},\hat{y} < i)$ is estimated in an auto-regressive manner, where the mean and scale of $y_{i}$ depend on the quantized latent in the preceding groups $\hat{y}_{< i}$ (see Eq. 2). To further enhance the estimation ability, we also deploy RAT in it. The RAT here can transmit key region prototypes through a separate bitstream, assisting in more efficient encoding and decoding. The prototypes $p$ could be seen as segmentation-prior, a supplement of the hyper-prior information $z$ . + +![](images/eb4a3ab36c142d7645b2c53cf5111bf45853c23d9711ea164bb525b232e7a4a5.jpg) +Fig.4: The architecture of the Prototype Encoder and Decoder for compressing and transmitting the prototypes. "Linear $\mathrm{C_1,C_2}$ " means linear layer with in-channel $\mathrm{C_1}$ and out-channel $\mathrm{C_2}$ . + +![](images/21a49e6f4ff40e1fe858fa64914610b261b5d57059e6895d3c2eeb78266c0394.jpg) +Fig. 5: The detailed modules in RAT, including Channel Transform layers (CTL), DPSCConv Kernels Generator (DKG), and Channel Attention Generator (CAG). LReLU is Leaky ReLU. GAP is Global Average Pooling. "GConv k²C,3,1,C" means Grouped Convolution with out-channel k²C, kernel size 3, stride 1 and group C. + +Table 1: BD-rate(%) performance on Kodak and CLIC Professional Valid (CLIC for short). The anchor is VTM-17.0. + +
ModelKodak +BD-rate↓CLIC +BD-rate↓ModelKodak +BD-rate↓CLIC +BD-rate↓
BPG [7]+31.5+48.1EVC [44]+0.27-
Cheng2020 [12]+7.46+11.7Feng2023 [16]-2.21-
Agustsson2023 [2]+6.15-ELIC [18]-3.08-3.58
ChARM [34]+3.58-Contextformer [28]-5.38-
Wang2022 [43]+1.10-0.99Pan2022 [35]-5.42-
BL-S+15.5+16.6Zou2022 (BL) [49]-2.89+0.14
Ours-S+3.87+5.26Ours-8.18-8.26
+ +# 4 Experiments + +# 4.1 Experimental Setup + +Training. We use COCO-Stuff [9] as the training dataset, which is composed of 118k images with panoptic semantic masks. We just use the class-agnostic masks without the category labels. We randomly crop the images with the size of $256 \times 256$ . The models are trained for 270 epochs (about 1M iterations) using the Adam optimizer [25] with a batch size of 32. The initial learning rate is set to $1 \times 10^{-4}$ for 170 epochs, and reduced by half every 30 epochs. Our models are optimized by the rate-distortion trade-off loss function (see Eq. 3 and Eq. 4). The $\lambda$ belongs to $\{0.0018, 0.0035, 0.0067, 0.0130, 0.0250, 0.0483\}$ for MSE as the distortion, and $\{2.4, 4.58, 8.73, 16.64, 31.73, 60.50\}$ for MS-SSIM [46]. + +Evaluation. We evaluate our model on two commonly used datasets, Kodak [27] with the size of $768 \times 512$ , and CLIC Professional Validation [1] with the size of 2K. Performance is evaluated using both bitrate and distortions. Distortion + +![](images/3fab81b18d85b9af2da515dccb4e08f471e002a5c5255227512be96c01cd4a43.jpg) +Fig.6: Rate-Distortion Performance on Kodak [27] and CLIC Professional Valid dataset [1]. Our SegPIC introduces proposed RAT and SAL based on Zou2022 [49]. We also implement a simple baseline model without self-attention and auto-regressive mechanism (BL-S), and incorporate our proposed modules on it (Ours-S). + +![](images/b15adf7dcb86af8218c86e4850e35b8412a5f81623f3c28950b6ab6d1ba4dc6a.jpg) + +measurement involves PSNR and MS-SSIM. And bitrates are assessed in bits per pixel (bpp). In this phase, we replace the masks with $4 \times 4$ grid partitions. Thus, there is no bitrate cost related to the masks, while the bitrate of prototypes $R(p)$ and $R(p^{\prime})$ is still taken into account. + +# 4.2 Rate-Distortion Performance + +We compare our SegPIC with previously well-performing methods, including discretized Gaussian mixture likelihoods and attention modules [12] (denoted as Cheng2020), ChARM [34] (denoted as Minnen2022), ChARM with WAM [49] (denoted as Zou2022), image-specific data-dependent transform model [43] (denoted as Wang2022), the best-performing traditional method VTM-17.0 [24] and influential traditional method BPG [7]. Furthermore, we introduce two simplified models without self-attention and auto-regressive mechanism, denoted as baseline-small (BL-S) and the one incorporating our proposed SAL and RAT (Ours-S), to demonstrate the generality of our modules. Details about the architecture of these two models will be provided in the appendix. + +As illustrated in Fig. 6 and Tab. 1, compared to VTM-17.0, our SegPIC could save $8.18\%$ BD-rate [8] on Kodak and $8.26\%$ on CLIC, which outperforms all the LIC and traditional methods mentioned above. To further verify the efficiency of the proposed RAT and SAL, We also compare SegPIC with the baseline Zou2022, which can save $5.45\%$ BD-rate on Kodak, and $8.50\%$ on CLIC (see Tab. 2). Moreover, Ours-S even achieves closely matched performance compared to ChARM. The MS-SSIM results are provided in the appendix. + +# 4.3 Ablation Study + +To assess the contributions of our RAT and SAL modules individually, we independently incorporate these two modules into the baseline Zou2022. Conse + +![](images/037fcdb385ee81d514b5e0afb34bc99719581bf3e0cfe5d0e1c0893b5a3de7ac.jpg) +Original kodim04 + +![](images/88f8f732716dcc89c4bd9a52b6ca000370f7b794d55f65ec43922b84eedd3817.jpg) +Ground Truth + +![](images/3a515b43679d80026bbc295f4a634476d04ff5438809fd4c1800edac89ff2fa9.jpg) +VTM (31.02/0.115) + +![](images/957f2fa822d29f159d60a8f265647b5825cc3fcc7e1b48cf532de5f76e37520b.jpg) +Zou2022 (31.46/0.119) + +![](images/ed79477c33635b9b68f99243f22619e22f37dd47832808243f2fef3f99acbf92.jpg) +Ours (31.53/0.109) + +![](images/93a90fa3b48708c694116e52b568ce4ba6ab0125dbd855173dd89607b9d29914.jpg) +Original kodim24 +Fig. 7: Visualization of the reconstructed images $\text{kodim}04$ and $\text{kodim}24$ in Kodak. The metrics are (PNSR↑/bpp↓). It shows that our SegPIC can distinguish the objects' contours more accurately, making the edges sharper with less bitrate. + +![](images/a22d4a9b97a192cefbba941436eb88f77f6ae650adfa4d91dded3d1219e1d562.jpg) +Ground Truth + +![](images/d5f179f9575f7e2ab0b96a0ac72120bd471c4b7534ff81294554c0e61edc9d3f.jpg) +VTM (28.96/0.325) + +![](images/2f291c2d0f6b37ee88a44033e03a0064a7d949295b200a1a320a86779e756397.jpg) +Zou2022 (28.67/0.292) + +![](images/c750ea1eaaf842512ddaed095f47d8cda2a9668fb97b1a22b0015b3eb01db6c1.jpg) +Ours (28.98/0.289) + +Table 2: Ablation Study on Kodak and CLIC. The Baseline (BL) is Zou2022 [49]. The speed is measured on Kodak with one Nvidia RTX 3090 GPU. + +
ModelKodakCLIC
ParamsBD-rateBD-rateEncDec Time
Zou2022 (BL)75.2 M0.00.0120 ms124 ms
BL w/ SAL76.7 M-2.12-2.93--
BL w/ RAT82.0 M-2.75-4.35--
Ours83.5 M-5.45-8.50142 ms130 ms
ChARM48.8 M+6.52-107 ms108 ms
BL-S21.7 M+19.0+16.3--
Ours-S29.2 M+6.96+5.1199 ms46 ms
+ +sequently, we construct two models: "baseline w/ SAL" and "baseline w/ RAT." As illustrated in Tab. 2, the parameter counts introduced by RAT and SAL separately are $1.5\mathrm{M}$ and $5.8\mathrm{M}$ , which occupy only a small fraction of the total parameter count, about $1.7\%$ and $6.9\%$ . And they can contribute to BD-rate saving by $2.93\%$ and $4.35\%$ on CLIC Professional Valid, respectively. It shows that the proposed RAT and SAL are lightweight and efficient. When we combine these two modules, SAL will provide stronger prototypes for RAT, leading to better results, i.e., a saving of $8.50\%$ BD-rate. + +With our proposed modules, Ours-S achieves high performance with a significantly smaller parameter count. Compared to ChARM [34], Ours-S saves nearly $40\%$ parameters and has faster speed while achieving comparable performance. Ours-S is a CNN-based model without any self-attention module. And the entropy estimation is just the hyperprior-style [6], without the efficient but time-consuming auto-regressive architecture. It indicates that the proposed modules RAT and SAL are highly effective and lightweight. + +![](images/71dfb814646e3399dad66a4939b48e0bdf3255c1113c3f9db5f5a7e82a7f7870.jpg) +(a) Kodak + +![](images/b44d25f77fba26cfab222d0247693025870a3fe5e2c0e9dde208324f9f6ce3f6.jpg) +(b) COCO-sample +Fig. 8: We measure the impact of the masks separately in both the training and testing phases. "test w/ masks" means we adopt the masks in both training and testing, "test w/o masks" means we adopt the masks only in training, while "train w/o masks" means we do not use masks all the time. Naturally, if the masks are not used, $4 \times 4$ grid partitions will be used as alternatives. + +![](images/ecafbcf43bd572a8595e8556218b96639595e10c0dae1ebbb538e7dbb1685a46.jpg) +(c) Validation loss in training + +# 4.4 Subjective Results + +Fig. 7 shows the subjective results compared with VTM and Zou2022. As we can see, with the help of the segmentation masks in training, our SegPIC can distinguish the objects' contours more accurately, making the edges sharper with less bitrate. In $kodim04$ , we can see the wool reconstructed by SegPIC is more consistent. In $kodim24$ , the bird patterns decoded by SegPIC are more clear, and the blurrings are less than in other methods. These results are based on grid partitions used as alternatives to masks. It demonstrates that SegPIC has acquired a more robust representative capability for various objects. Consequently, SegPIC can discern objects in complex textures more effectively. + +# 4.5 Analysis of the Role of Segmentation Masks + +Quantitative Results. To validate the role of the masks, we make experiments on two datasets, Kodak and COCO-sample. Due to Kodak not being a segmentation dataset, the corresponding masks are generated using the segmentation method [36]. Otherwise, we randomly sample 20 images and their masks from COCO Valid 2017 [9] to create the COCO-sample dataset. + +As illustrated in Fig. 8 (a) (b), during the testing (inference) phase, the differences caused by masks are not so evident. However, for training, the situation is entirely different. With the assistance of the masks, the model achieves significantly better performance, especially at low bitrates. It could imply that, due to bitrate loss constraints during training, the information in the prototypes is implicitly transferred to the learnable parameters in RAT, allowing limited contextual information to still be effective during the inference phase. At low bitrates, since the lack of sufficient high-frequency information, pixels in the region exhibit stronger similarity, enhancing the representational ability of features extracted by prototypes. Fig. 8 (c) shows the early stage of model training. + +As shown in Tab. 3, the model training w/o masks (the last column) leads to noticeable degradation. However, training w/ masks has a relatively minor + +![](images/77ae98955415a15f9706ec26f4656280dbe084927ad793772e397990b7abe755.jpg) +Fig. 9: Visualization of the kernels generated by DKG in RAT at the Encoder (see Fig. 3), in the cases with and without the masks. Fig. (d)(e)(f) are generated by the same pre-trained network weights. During testing, they go through processes with or without masks, respectively. The distortions in (d) can be formulated as $(x - \hat{x})^2$ . We take the maximum absolute values of the tensors along the channels and apply 2D max pooling on them for better visualization. + +Table 3: The effect of the masks and the grid number on BD-rate on COCO-sample. "n×" denotes n×n grid partitions to replace masks. + +
Traintrain w/ masks4×
Testtest w/ masks2×4×8×4×
BD-rate ↓-0.11+0.210(default)+0.46+2.10
+ +impact on the pattern and number in testing. So we can replace the masks in testing to further save bitrate but should take the masks as privileged information during training. The hyperparameter $4 \times$ in testing is an empirical trade-off between the effect of RAT (i.e. image reconstruction precision) and the quantity of the prototypes (i.e. the additional bitrate they introduced). + +Qualitative Analysis. Fig. 9 shows the visualization of the kernels generated by DKG in RAT at the Encoder (see Fig. 2). Although there is not a noticeable difference in Fig. 9 (e), it is obvious that the gradients of $w / \text{mask}$ are more focused on the semantic objects (see Fig. 9 (f)), which is complex and prone to causing distortions. This attribute will assist in optimizing the models, enabling it to better reconstruct various objects. These observations indicate that the masks do play a necessary role during the training phase. + +# 5 Conclusion + +We propose a novel image compression framework called Segmentation-Prior-Guided Image Compression (SegPIC), equipped with two proposed modules: Region-Adaptive Transform (RAT) and Scale Affine Layer (SAL). We adopt the class-agnostic masks as privilege information to assist the training of SegPIC. Experimental results demonstrate that our SegPIC is superior to the previous works. Otherwise, we analyze the role of the masks. Experiments and visualization results show that the masks can guide the model to focus more on semantic objects in training, leading to higher pixel-fidelity performance. + +# Acknowledgements + +This work was supported in part by the National Key R&D Program of China (No.2021ZD0112100), National NSF of China (No.62120106009), Guangdong Basic and Applied Basic Research Foundation (2024A1515010454), the Basic and Frontier Research Project of PCL, and the Major Key Project of PCL. + +# References + +1. Workshop and challenge on learned image compression (clic2020) (2020), http://www.compression.cc 10, 11 +2. Agustsson, E., Minnen, D., Toderici, G., Mentzer, F.: Multi-realism image compression with a conditional generator. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 22324-22333 (2023) 10 +3. Ahmed, N., Natarajan, T., Rao, K.R.: Discrete cosine transform. IEEE Transactions on Computers 100(1), 90-93 (1974) 2 +4. Akbari, M., Liang, J., Han, J.: Dsslic: Deep semantic segmentation-based layered image compression. In: ICASSP. pp. 2042-2046. IEEE (2019) 3 +5. Balle, J., Laparra, V., Simoncelli, E.P.: End-to-end optimized image compression. In: Int. Conf. Learn. Represent. (2017) 2, 5 +6. Balle, J., Minnen, D., Singh, S., Hwang, S.J., Johnston, N.: Variational image compression with a scale hyperprior. In: Int. Conf. Learn. Represent. (2018) 2, 4, 5, 12 +7. Bellard, F.: Bpg image format (2014), https://bellard.org/bpg/ 2, 10, 11 +8. Bjontegaard, G.: Calculation of average psnr differences between rd-curves. ITU SG16 Doc. VCEG-M33 (2001) 11 +9. Caesar, H., Uijlings, J., Ferrari, V.: COCO-Stuff: Thing and stuff classes in context. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 1209-1218 (2018) 10, 13 +0. Chang, J., Mao, Q., Zhao, Z., Wang, S., Wang, S., Zhu, H., Ma, S.: Layered conceptual image compression via deep semantic synthesis. In: IEEE Int. Conf. Image Process. pp. 694-698 (2019) 4, 5 +11. Chang, J., Zhao, Z., Yang, L., Jia, C., Zhang, J., Ma, S.: Thousand to one: Semantic prior modeling for conceptual coding. In: Int. Conf. Multimedia and Expo. pp. 1-6. IEEE (2021) 3, 4, 5 +12. Cheng, Z., Sun, H., Takeuchi, M., Katto, J.: Learned image compression with discretized gaussian mixture likelihoods and attention modules. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 7939-7948 (2020) 3, 4, 10, 11 +13. Chollet, F.: Xception: Deep learning with depthwise separable convolutions. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 1251-1258 (2017) 8 +14. Daubechies, I.: The wavelet transform, time-frequency localization and signal analysis. IEEE Transactions on Information Theory 36(5), 961-1005 (1990) 2 +15. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., Uszkoreit, J., Houlsby, N.: An image is worth 16x16 words: Transformers for image recognition at scale. In: Int. Conf. Learn. Represent. (2021) 3 +16. Feng, R., Gao, Y., Jin, X., Feng, R., Chen, Z.: Semantically structured image compression via irregular group-based decoupling. In: Int. Conf. Comput. Vis. pp. 17237-17247 (2023) 3, 5, 10 + +17. Goodfellow, I., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., Courville, A., Bengio, Y.: Generative adversarial nets. In: Ghahramani, Z., Welling, M., Cortes, C., Lawrence, N., Weinberger, K. (eds.) Adv. Neural Inform. Process. Syst. vol. 27. Curran Associates, Inc. (2014) 4 +18. He, D., Yang, Z., Peng, W., Ma, R., Qin, H., Wang, Y.: Elic: Efficient learned image compression with unevenly grouped space-channel contextual adaptive coding. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 5718-5727 (2022) 10 +19. He, D., Zheng, Y., Sun, B., Wang, Y., Qin, H.: Checkerboard context model for efficient learned image compression. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 14771-14780 (2021) 4 +20. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 770-778 (2016) 3 +21. Hoang, T.M., Zhou, J., Fan, Y.: Image compression with encoder-decoder matched semantic segmentation. In: IEEE Conf. Comput. Vis. Pattern Recog. Worksh. pp. 160-161 (2020) 3 +22. Jia, X., De Brabandere, B., Tuytelaars, T., Gool, L.V.: Dynamic filter networks. Adv. Neural Inform. Process. Syst. 29 (2016) 8 +23. Jiang, W., Yang, J., Zhai, Y., Ning, P., Gao, F., Wang, R.: Mlic: Multi-reference entropy model for learned image compression. In: ACM Int. Conf. Multimedia. pp. 7618-7627 (2023) 4 +24. (JVET), J.V.E.T.: Versatile video coding (2021), https://jvet.hhi.fraunhofer.de/ 2, 11 +25. Kingma, D.P., Ba, J.: Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980 (2014) 10 +26. Kingma, D.P., Welling, M.: Auto-encoding variational bayes. In: Int. Conf. Learn. Represent. (2014) 2 +27. Kodak, E.: Kodak lossless true color image suite (photocd pcd0992) (1993), http://r0k.us/graphics/kodak/ 10, 11 +28. Koyuncu, A.B., Gao, H., Boev, A., Gaikov, G., Alshina, E., Steinbach, E.: Contextformer: A transformer with spatio-channel attention for context modeling in learned image compression. In: Eur. Conf. Comput. Vis. pp. 447-463. Springer (2022) 10 +29. Li, F., Zhang, L., Liu, Z., Lei, J., Li, Z.: Multi-frequency representation enhancement with privilege information for video super-resolution. In: Int. Conf. Comput. Vis. pp. 12814-12825 (2023) 3 +30. Liu, J., Lu, G., Hu, Z., Xu, D.: A unified end-to-end framework for efficient deep image compression. arXiv preprint arXiv:2002.03370 (2020) 3, 4 +31. Liu, J., Sun, H., Katto, J.: Learned image compression with mixed transformer-cnn architectures. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 14388-14397 (2023) 4 +32. Liu, Z., Lin, Y., Cao, Y., Hu, H., Wei, Y., Zhang, Z., Lin, S., Guo, B.: Swin transformer: Hierarchical vision transformer using shifted windows. In: Int. Conf. Comput. Vis. pp. 10012-10022 (2021) 3, 4 +33. Minnen, D., Balle, J., Toderici, G.D.: Joint autoregressive and hierarchical priors for learned image compression. Adv. Neural Inform. Process. Syst. 31 (2018) 4 +34. Minnen, D., Singh, S.: Channel-wise autoregressive entropy models for learned image compression. In: IEEE Int. Conf. Image Process. pp. 3339-3343. IEEE (2020) 4, 5, 9, 10, 11, 12 +35. Pan, G., Lu, G., Hu, Z., Xu, D.: Content adaptive latents and decoder for neural image compression. In: Eur. Conf. Comput. Vis. pp. 556-573. Springer (2022) 10 + +36. Qi, L., Kuen, J., Wang, Y., Gu, J., Zhao, H., Torr, P., Lin, Z., Jia, J.: Open world entity segmentation. IEEE Trans. Pattern Anal. Mach. Intell. (2022) 13 +37. Shen, H., Zhao, Z.Q., Zhang, W.: Adaptive dynamic filtering network for image denoising. In: AAAI. vol. 37, pp. 2227-2235 (2023) 6 +38. Skodras, A., Christopoulos, C., Ebrahimi, T.: The JPEG 2000 still image compression standard. IEEE Signal Processing Magazine 18(5), 36-58 (2001) 2 +39. Stark, H., Woods, J.W.: Probability, random processes, and estimation theory for engineers. Prentice-Hall, Inc. (1986) 2 +40. Sun, S., He, T., Chen, Z.: Semantic structured image coding framework for multiple intelligent applications. IEEE Trans. Circuit Syst. Video Technol. 31(9), 3631-3642 (2020) 5 +41. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. Adv. Neural Inform. Process. Syst. 30 (2017) 4 +42. Wallace, G.K.: The JPEG still picture compression standard. Communications of the ACM 34(4), 30-44 (1991) 2 +43. Wang, D., Yang, W., Hu, Y., Liu, J.: Neural data-dependent transform for learned image compression. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 17379-17388 (2022) 2, 3, 4, 10, 11 +44. Wang, G.H., Li, J., Li, B., Lu, Y.: Evc: Towards real-time neural image compression with mask decay. In: Int. Conf. Learn. Represent. (2023) 10 +45. Wang, X., Yu, K., Dong, C., Loy, C.C.: Recovering realistic texture in image superresolution by deep spatial feature transform. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 606-615 (2018) 8 +46. Wang, Z., Simoncelli, E.P., Bovik, A.C.: Multiscale structural similarity for image quality assessment. In: The Thirty-Seventh Asilomar Conference on Signals, Systems & Computers, 2003. vol. 2, pp. 1398-1402. IEEE (2003) 4, 10 +47. Xu, Y.S., Tseng, S.Y.R., Tseng, Y., Kuo, H.K., Tsai, Y.M.: Unified dynamic convolutional network for super-resolution with variational degradations. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 12496-12505 (2020) 6 +48. Zhu, Y., Yang, Y., Cohen, T.: Transformer-based transform coding. In: Int. Conf. Learn. Represent. (2021) 3, 4 +49. Zou, R., Song, C., Zhang, Z.: The devil is in the details: Window-based attention for image compression. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 17492-17501 (2022) 3, 4, 5, 7, 9, 10, 11, 12 \ No newline at end of file diff --git a/2024/Region-Adaptive Transform with Segmentation Prior for Image Compression/images.zip b/2024/Region-Adaptive Transform with Segmentation Prior for Image Compression/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..f162663e23a18dfa44809a0c858a45764c954b82 --- /dev/null +++ b/2024/Region-Adaptive Transform with Segmentation Prior for Image Compression/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ca37c4a91972f207938151e79d3dc3573f9f605028d38706f131d301d29ee16e +size 466233 diff --git a/2024/Region-Adaptive Transform with Segmentation Prior for Image Compression/layout.json b/2024/Region-Adaptive Transform with Segmentation Prior for Image Compression/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..a298d81984dabc9bb1ea8630fa4a330cca62c9bf --- /dev/null +++ b/2024/Region-Adaptive Transform with Segmentation Prior for Image Compression/layout.json @@ -0,0 +1,10925 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 142, + 111, + 473, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 142, + 111, + 473, + 148 + ], + "spans": [ + { + "bbox": [ + 142, + 111, + 473, + 148 + ], + "type": "text", + "content": "Region-Adaptive Transform with Segmentation Prior for Image Compression" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 147, + 167, + 465, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 147, + 167, + 465, + 192 + ], + "spans": [ + { + "bbox": [ + 147, + 167, + 465, + 192 + ], + "type": "text", + "content": "Yuxi Liu" + }, + { + "bbox": [ + 147, + 167, + 465, + 192 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 147, + 167, + 465, + 192 + ], + "type": "text", + "content": ", Wenhan Yang" + }, + { + "bbox": [ + 147, + 167, + 465, + 192 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 147, + 167, + 465, + 192 + ], + "type": "text", + "content": ", Huihui Bai" + }, + { + "bbox": [ + 147, + 167, + 465, + 192 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 147, + 167, + 465, + 192 + ], + "type": "text", + "content": ", Yunchao Wei" + }, + { + "bbox": [ + 147, + 167, + 465, + 192 + ], + "type": "inline_equation", + "content": "^{1,2,3}" + }, + { + "bbox": [ + 147, + 167, + 465, + 192 + ], + "type": "text", + "content": ", and Yao Zhao" + }, + { + "bbox": [ + 147, + 167, + 465, + 192 + ], + "type": "inline_equation", + "content": "^{\\star 1,2,3}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 132, + 201, + 481, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 201, + 481, + 247 + ], + "spans": [ + { + "bbox": [ + 132, + 201, + 481, + 247 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 132, + 201, + 481, + 247 + ], + "type": "text", + "content": " Institute of Information Science, Beijing Jiaotong University, Beijing, China \n" + }, + { + "bbox": [ + 132, + 201, + 481, + 247 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 132, + 201, + 481, + 247 + ], + "type": "text", + "content": " Visual Intelligence + X International Joint Laboratory of the Ministry of Education \n" + }, + { + "bbox": [ + 132, + 201, + 481, + 247 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 132, + 201, + 481, + 247 + ], + "type": "text", + "content": " PengCheng Laboratory, Shenzhen, China \n{yuxiliu, hhbai, yunchao.wei, yzhao}@bjtu.edu.cn, yangwh@pcl.ac.cn" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 160, + 276, + 452, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 276, + 452, + 518 + ], + "spans": [ + { + "bbox": [ + 160, + 276, + 452, + 518 + ], + "type": "text", + "content": "Abstract. Learned Image Compression (LIC) has shown remarkable progress in recent years. Existing works commonly employ CNN-based or Transformer-based modules as transform methods for compression. However, there is no prior research on neural transform that focuses on specific regions. In response, we introduce the class-agnostic segmentation masks (i.e. semantic masks without category labels) for extracting region-adaptive contextual information. Our proposed module, Region-Adaptive Transform, applies adaptive convolutions on different regions guided by the masks. Additionally, we introduce a plug-and-play module named Scale Affine Layer to incorporate rich contexts from various regions. While there have been prior image compression efforts that involve segmentation masks as additional intermediate inputs, our approach differs significantly from them. Our advantages lie in that, to avoid extra bitrate overhead, we treat these masks as privilege information, which is accessible during the model training stage but not required during the inference phase. To the best of our knowledge, we are the first to employ class-agnostic masks as privilege information and achieve superior performance in pixel-fidelity metrics, such as Peak Signal to Noise Ratio (PSNR). The experimental results demonstrate our improvement compared to previously well-performing methods, with about " + }, + { + "bbox": [ + 160, + 276, + 452, + 518 + ], + "type": "inline_equation", + "content": "8.2\\%" + }, + { + "bbox": [ + 160, + 276, + 452, + 518 + ], + "type": "text", + "content": " bitrate saving compared to VTM-17.0. The source code is available at https://github.com/GityuxiLiu/SegPIC-for-Image-Compression." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 160, + 529, + 418, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 529, + 418, + 540 + ], + "spans": [ + { + "bbox": [ + 160, + 529, + 418, + 540 + ], + "type": "text", + "content": "Keywords: Learned Image Compression " + }, + { + "bbox": [ + 160, + 529, + 418, + 540 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 529, + 418, + 540 + ], + "type": "text", + "content": " Segmentation Prior" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 561, + 230, + 574 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 561, + 230, + 574 + ], + "spans": [ + { + "bbox": [ + 132, + 561, + 230, + 574 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 586, + 482, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 586, + 482, + 647 + ], + "spans": [ + { + "bbox": [ + 130, + 586, + 482, + 647 + ], + "type": "text", + "content": "With the advent of the multimedia era, an ever-growing volume of image data has emerged on the internet. These vast volumes of image data demand larger transmission bandwidth and storage capacity. In this context, the importance of image compression technology has become even more pronounced, evoking greater efforts in efficient image compression techniques. Lossy image compres" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 653, + 234, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 653, + 234, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 653, + 234, + 665 + ], + "type": "text", + "content": "* Corresponding author" + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 207, + 114, + 408, + 169 + ], + "blocks": [ + { + "bbox": [ + 207, + 114, + 408, + 169 + ], + "lines": [ + { + "bbox": [ + 207, + 114, + 408, + 169 + ], + "spans": [ + { + "bbox": [ + 207, + 114, + 408, + 169 + ], + "type": "image", + "image_path": "0e4b5631c4dae19c2369313434f77ba00505965d04a1f677608d5c2e2f704eaf.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 261, + 171, + 352, + 179 + ], + "lines": [ + { + "bbox": [ + 261, + 171, + 352, + 179 + ], + "spans": [ + { + "bbox": [ + 261, + 171, + 352, + 179 + ], + "type": "text", + "content": "(a) Convictional LIC framework" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 207, + 186, + 407, + 237 + ], + "blocks": [ + { + "bbox": [ + 207, + 186, + 407, + 237 + ], + "lines": [ + { + "bbox": [ + 207, + 186, + 407, + 237 + ], + "spans": [ + { + "bbox": [ + 207, + 186, + 407, + 237 + ], + "type": "image", + "image_path": "c0c76c0d31e5f465ee91214d749e0aaf0cf49a4e22189e98f43822733d9c0928.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 256, + 239, + 357, + 247 + ], + "lines": [ + { + "bbox": [ + 256, + 239, + 357, + 247 + ], + "spans": [ + { + "bbox": [ + 256, + 239, + 357, + 247 + ], + "type": "text", + "content": "(b) Data-dependent LIC framework" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 208, + 253, + 406, + 323 + ], + "blocks": [ + { + "bbox": [ + 208, + 253, + 406, + 323 + ], + "lines": [ + { + "bbox": [ + 208, + 253, + 406, + 323 + ], + "spans": [ + { + "bbox": [ + 208, + 253, + 406, + 323 + ], + "type": "image", + "image_path": "c10f29b94376026738b0c62e9af84326aafc84815ee25a2a130a861a6da5fdc6.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 266, + 327, + 348, + 335 + ], + "lines": [ + { + "bbox": [ + 266, + 327, + 348, + 335 + ], + "spans": [ + { + "bbox": [ + 266, + 327, + 348, + 335 + ], + "type": "text", + "content": "(c) Proposed LIC framework" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 131, + 344, + 482, + 389 + ], + "lines": [ + { + "bbox": [ + 131, + 344, + 482, + 389 + ], + "spans": [ + { + "bbox": [ + 131, + 344, + 482, + 389 + ], + "type": "text", + "content": "Fig. 1: Paradigms of end-to-end learned image compression (LIC) methods. (a) Conventional LIC with fixed learned kernels. (b) Data-dependent LIC with image-specific kernels [43]. (c) The proposed method is capable of generating region-adaptive kernels from different regions, providing more fine-grained transforms." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 414, + 482, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 414, + 482, + 581 + ], + "spans": [ + { + "bbox": [ + 130, + 414, + 482, + 581 + ], + "type": "text", + "content": "sion technology can reduce the data size to a lower level while maintaining good visual quality. In the past decades, the classical lossy image compression methods such as JPEG [42], JPEG2000 [38], BPG [7], and VVC [24], used to have excellent performance and were widely adopted. These methods include modules of transform, quantization, and entropy coding. The transform process aims to convert images into a more compact coefficient by removing pixel correlations as much as possible. The commonly employed transforms include the discrete cosine transform (DCT) [3] and wavelet transform [14]. These transformations are linear, characterized by simplicity and efficiency. However, their fixed forms are unable to adapt to the contexts of complex signals. To address this issue, latter efforts propose data-driven transformation approaches, e.g. the Karhunen-Loève transform (KLT) [39]. Furthermore, the subsequent works are dedicated to constructing complex patterns for the adaptive selection of local transforms, improving adaptability and overall performance." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "content": "Benfit from the advancement of neural network technology, many learned image compression (LIC) methods have been proposed, which shows a promising alternative for designing transform: learning from data instead of handcraft design. Balle et al. [5, 6] proposed the first well-known end-to-end framework. They utilize variational auto-encoders (VAE) [26] as a transform module and optimize the parameters of the kernels to better align with the natural data distribution. To further improve the representation capabilities of the transform" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "text", + "content": "Y. Liu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "text", + "content": "module, some works [12, 30, 43, 48, 49] design advanced modules to introduce richer contextual information, e.g. residual blocks [20] and self-attention mechanism [15, 32]. These works optimize learnable kernels to fit the data distribution to attain a more powerful transform. However, their kernel parameters are fixed after training (see Fig.1 (a)). Wang et al. [43] introduce a neural data-dependent transform, which extracts dynamic convolution kernels based on the context of the entire image for more adaptive transform (see Fig. 1 (b)). However, even within a single image, the context could vary significantly across different regions, e.g. stone and sheep regions in Fig. 1. It limits the description of the signals to a more fine-grained level using such image-level transforms." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 236, + 482, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 236, + 482, + 475 + ], + "spans": [ + { + "bbox": [ + 130, + 236, + 482, + 475 + ], + "type": "text", + "content": "To address the issue, this paper makes the first efforts to design learnable region-adaptive transform to capture the mapping regularity in different regions. In detail, we employ class-agnostic4 segmentation masks to extract region-specific information for transform (see Fig. 1 (c)). Diffs from previous works [4, 11, 16, 21], our approach stands out with two characteristics in mask utilization. First, our masks are \"class-agnostic\", where the semantic labels are abandoned but a more effective compression-driven semantic prior will be learned by our model in a data-driven manner. Second, to avoid incurring additional costs in both bitrate and computational resources, we treat these masks as privilege information, which are available during training but remain inaccessible during the inference phase [29]. Specifically, we replace the masks with uniform grid partitions (see Fig. 9 (c)) in inference, and the experimental result shows approximate performance. It implies that our model has learned useful semantic knowledge from the masks in training, incorporating this knowledge into the model (i.e. the network's parameters). Hence, it remains capable of mining contextual information even from relatively coarse partition regions, i.e., the given uniform grid partitions. This simple operation can effectively maintain the consistency of our model's architecture and parameters, whether or not masks are used at different phases. The more details about the difference between our method and previous semantic-mask-based works will be explained at Sec. 2." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 475, + 482, + 583 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 475, + 482, + 583 + ], + "spans": [ + { + "bbox": [ + 130, + 475, + 482, + 583 + ], + "type": "text", + "content": "Based on the above motivations, we propose a novel image compression framework called Segmentation-Prior-Guided Image Compression (SegPIC), equipped with two key modules: Region-Adaptive Transform (RAT) and Scale Affine Layer (SAL). The RAT adopts region-adaptive kernels for transform, performing the convolution whose weights are generated from the prototypes and the context of latent features under the guidance of masks. The SAL affines the latent features for better mining of semantic contexts in the Encoder and Decoder. Experimental results show that SegPIC is superior to previously well-performing image compression methods. Our main contributions are as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 590, + 482, + 625 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 590, + 482, + 625 + ], + "spans": [ + { + "bbox": [ + 138, + 590, + 482, + 625 + ], + "type": "text", + "content": "- We propose a novel image compression framework called Segmentation-Prior-Guided Image Compression (SegPIC) to generate more effective region-adaptive transforms in a data-driven manner." + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 134, + 91, + 447, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 91, + 447, + 103 + ], + "spans": [ + { + "bbox": [ + 134, + 91, + 447, + 103 + ], + "type": "text", + "content": "Region-Adaptive Transform with Segmentation Prior for Image Compression" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 135, + 631, + 482, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 631, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 135, + 631, + 482, + 665 + ], + "type": "inline_equation", + "content": "{}^{4}" + }, + { + "bbox": [ + 135, + 631, + 482, + 665 + ], + "type": "text", + "content": " \"class-agnostic\" means distinguishing different objects in an image regardless of what those objects are. Such masks can be regarded as a generalized representation of partition maps widely used in conventional CODECs." + } + ] + } + ], + "index": 6 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 137, + 116, + 480, + 247 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 137, + 116, + 479, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 116, + 479, + 163 + ], + "spans": [ + { + "bbox": [ + 137, + 116, + 479, + 163 + ], + "type": "text", + "content": "- Our SegPIC is equipped with two proposed modules: Region-Adaptive Transform (RAT) and Scale Affine Layer (SAL). The RAT adopts region-adaptive kernels for transform, while the SAL affines the latent features for better mining of semantic contexts." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 164, + 480, + 209 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 164, + 480, + 209 + ], + "spans": [ + { + "bbox": [ + 138, + 164, + 480, + 209 + ], + "type": "text", + "content": "- We utilize the class-agnostic masks as privilege information to assist the training of SegPIC. The \"class-agnostic\" property promotes the model to learn a compression-driven semantic prior, while the \"privilege\" way avoids additional bitrate and complexity for masks in inference." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 213, + 480, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 213, + 480, + 247 + ], + "spans": [ + { + "bbox": [ + 138, + 213, + 480, + 247 + ], + "type": "text", + "content": "- Experimental results demonstrate that our SegPIC is superior to previously well-performing image compression methods. We also present sufficient ablation experiments to confirm the necessity of the privilege information." + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 132, + 267, + 242, + 280 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 267, + 242, + 280 + ], + "spans": [ + { + "bbox": [ + 132, + 267, + 242, + 280 + ], + "type": "text", + "content": "2 Related Works" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 294, + 479, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 294, + 479, + 376 + ], + "spans": [ + { + "bbox": [ + 130, + 294, + 479, + 376 + ], + "type": "text", + "content": "Learned Image Compression. Image compression can be divided into three main components: transform, quantization, and entropy coding. In brief, the transformation module maps the image to a latent feature space. Subsequently, quantization is performed after this phase, resulting in some quantization distortions but achieving lower bitrate. The entropy encoding module estimates the data distribution of the latent features and inputs it to the Arithmetic Coding (AC) module, which generates the bitstream for transmission." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 378, + 479, + 520 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 378, + 479, + 520 + ], + "spans": [ + { + "bbox": [ + 130, + 378, + 479, + 520 + ], + "type": "text", + "content": "To enhance transform module, some works [30,43,48,49] adopt more powerful modules like self-attention mechanism [32, 41] to explore broader context. [31] proposed a transform block that mixes CNN and Transformer to combine their advantages. As for the entropy estimation module, the studies [12, 19, 33, 34] usually focus on exploring novel auto-regressive methods to split the latent features into several portions to be decoded in order. The portions having been decoded will assist in decoding the next portion, through the spatial or channel correlation of these portions. This auto-regressive mechanism may introduce additional latency but achieve better performance than Hyper Prior [6]. Recently, [23] integrates different types of correlations present in latent representation, e.g. channel-wise, local spatial, and global spatial correlations. Their proposed novel auto-regressive entropy module achieve powerful performance." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 522, + 480, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 522, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 522, + 480, + 665 + ], + "type": "text", + "content": "Thanks to these great works, the performance of LIC has become so strong that achieving further enhancements becomes progressively challenging. Therefore, some new directions in compression have emerged, which is also meaningful on certain occasions. For example, some semantic-based compression methods [10, 11] adopt GAN [17] to compress images at extremely low bitrate. These methods maintain a good human visual perception effect but usually, poor pixel-fidelity metrics, e.g. Peak Signal to Noise Ratio (PSNR) and Multi-Scale Structural Similarity Index (MS-SSIM) [46]. Specifically, these methods could introduce some synthetic artifacts with low fidelity. For example, a semantic compression method could reconstruct an image of a bird with different breeds from the original image because the training category labels do not distinguish the bird's breed. Another direction is compression for machines, which means compress" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "type": "text", + "content": "Y. Liu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 164, + 115, + 449, + 264 + ], + "blocks": [ + { + "bbox": [ + 164, + 115, + 449, + 264 + ], + "lines": [ + { + "bbox": [ + 164, + 115, + 449, + 264 + ], + "spans": [ + { + "bbox": [ + 164, + 115, + 449, + 264 + ], + "type": "image", + "image_path": "9cc7ffa734bebf9d81b607c1f20b9fe6ce6033197cf401b6541bbec60fa0f014.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 257, + 269, + 343, + 279 + ], + "lines": [ + { + "bbox": [ + 257, + 269, + 343, + 279 + ], + "spans": [ + { + "bbox": [ + 257, + 269, + 343, + 279 + ], + "type": "text", + "content": "(a) The overall framework" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 157, + 293, + 302, + 360 + ], + "blocks": [ + { + "bbox": [ + 157, + 293, + 302, + 360 + ], + "lines": [ + { + "bbox": [ + 157, + 293, + 302, + 360 + ], + "spans": [ + { + "bbox": [ + 157, + 293, + 302, + 360 + ], + "type": "image", + "image_path": "17cd25dd45547abe1efde816681db3f4bf91e02afb712ae2719c419e87bd3f0b.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 166, + 366, + 292, + 376 + ], + "lines": [ + { + "bbox": [ + 166, + 366, + 292, + 376 + ], + "spans": [ + { + "bbox": [ + 166, + 366, + 292, + 376 + ], + "type": "text", + "content": "(b)Downsample block (out channel C)" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 311, + 293, + 457, + 359 + ], + "blocks": [ + { + "bbox": [ + 311, + 293, + 457, + 359 + ], + "lines": [ + { + "bbox": [ + 311, + 293, + 457, + 359 + ], + "spans": [ + { + "bbox": [ + 311, + 293, + 457, + 359 + ], + "type": "image", + "image_path": "5bf1028832daf3fd0fbde7ec4964a20e6844baeb5d5bff68d20b0470d47fadfd.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 320, + 366, + 436, + 376 + ], + "lines": [ + { + "bbox": [ + 320, + 366, + 436, + 376 + ], + "spans": [ + { + "bbox": [ + 320, + 366, + 436, + 376 + ], + "type": "text", + "content": "(c) Upsample block (out channel C)" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 385, + 482, + 495 + ], + "lines": [ + { + "bbox": [ + 130, + 385, + 482, + 495 + ], + "spans": [ + { + "bbox": [ + 130, + 385, + 482, + 495 + ], + "type": "text", + "content": "Fig. 2: (a) Overall framework of our SegPIC. The blue flow lines represent the decoding process on the decoder side. The dashed lines represent the process of extracting the prototypes " + }, + { + "bbox": [ + 130, + 385, + 482, + 495 + ], + "type": "inline_equation", + "content": "p'" + }, + { + "bbox": [ + 130, + 385, + 482, + 495 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 385, + 482, + 495 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 130, + 385, + 482, + 495 + ], + "type": "text", + "content": " (see Eq. 5). The compression of " + }, + { + "bbox": [ + 130, + 385, + 482, + 495 + ], + "type": "inline_equation", + "content": "p'" + }, + { + "bbox": [ + 130, + 385, + 482, + 495 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 385, + 482, + 495 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 130, + 385, + 482, + 495 + ], + "type": "text", + "content": " is not fully depicted for simplicity, please refer to Fig. 4 for details. " + }, + { + "bbox": [ + 130, + 385, + 482, + 495 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 385, + 482, + 495 + ], + "type": "text", + "content": " is set to 192, and " + }, + { + "bbox": [ + 130, + 385, + 482, + 495 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 130, + 385, + 482, + 495 + ], + "type": "text", + "content": " is 320 as in [49]. RAT is the proposed Region-Adaptive Transform. SAL is the proposed Scale Affine Layer. WAM is Window Attention Module [49]. ChARM is Channel-wise Auto-Regressive Model [34]. FM is a Factorized Model [6]. GDN is Generalized Divisive Normalization [5]. (b) is Downsample Block. \"Downsample, N\" in (a) means the block with out-channel N. \"Conv C,5,2\" means out-channel C, kernel size 5, stride 2. (c) is Upsample Block, and \"TConv\" is Transposed Convolution." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 520, + 481, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 520, + 481, + 544 + ], + "spans": [ + { + "bbox": [ + 130, + 520, + 481, + 544 + ], + "type": "text", + "content": "ing images for downstream tasks [16, 40]. These methods focus on balancing the bitrate and performance of downstream tasks, not just the pixel-fidelity metrics." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 545, + 482, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 482, + 640 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 482, + 640 + ], + "type": "text", + "content": "Although some works [10, 11, 16] about the two directions also adopt masks, our usage is significantly different. The main differences are as follows: 1) We utilize the masks as privilege information, without the mask information during the inference phase, while usually, the masks are indispensable for other methods. 2) We employ class-agnostic masks to guide the model to extract rich compression-friendly contexts among semantically related regions, regardless of the category that might degrade the fidelity as mentioned above. 3) The proposed model is optimized towards better pixel-fidelity metrics, e.g. PSNR and MS-SSIM." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "content": "Region-adaptive Convolution. Due to its capacity to model the degradation levels in different regions of images, region-adaptive convolution is applied to" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 135, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 135, + 91, + 447, + 102 + ], + "type": "text", + "content": "Region-Adaptive Transform with Segmentation Prior for Image Compression" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "text", + "content": "some low-level vision tasks such as super-resolution and denoising [37, 47]. In terms of compression, the reconstruction quality varies across different regions. For instance, regions with complex textures are more challenging compared to those with simpler patterns. Therefore, specific processing can be applied to different regions based on their characteristics. The regions can be divided using segmentation masks or partitions based on certain regular patterns. To the best of our knowledge, we are the first to introduce this mechanism on image compression for optimizing pixel-fidelity metrics." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 229, + 259, + 243 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 229, + 259, + 243 + ], + "spans": [ + { + "bbox": [ + 131, + 229, + 259, + 243 + ], + "type": "text", + "content": "3 Proposed SegPIC" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 254, + 217, + 266 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 254, + 217, + 266 + ], + "spans": [ + { + "bbox": [ + 131, + 254, + 217, + 266 + ], + "type": "text", + "content": "3.1 Motivation" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 273, + 481, + 309 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 273, + 481, + 309 + ], + "spans": [ + { + "bbox": [ + 130, + 273, + 481, + 309 + ], + "type": "text", + "content": "As mentioned above, previous transforms either adopt convolutions with fixed parameters or image-specific convolutions. Our motivation is to construct a region-adaptive transform via the following three properties:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 137, + 316, + 480, + 386 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 138, + 316, + 480, + 338 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 316, + 480, + 338 + ], + "spans": [ + { + "bbox": [ + 138, + 316, + 480, + 338 + ], + "type": "text", + "content": "- Region-dependent: we propose a neural network with region-adaptive transforms guided by the segmentation masks." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 137, + 339, + 480, + 363 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 339, + 480, + 363 + ], + "spans": [ + { + "bbox": [ + 137, + 339, + 480, + 363 + ], + "type": "text", + "content": "- Class-Agnostic: the adopted mask does not include a category label, allowing the model to be free to learn compression-friendly semantic priors." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 137, + 363, + 480, + 386 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 363, + 480, + 386 + ], + "spans": [ + { + "bbox": [ + 137, + 363, + 480, + 386 + ], + "type": "text", + "content": "- Privilege Information: the masks are expected as privilege information, not being used in inference for saving bitrate and computation resources." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 131, + 394, + 481, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 394, + 481, + 418 + ], + "spans": [ + { + "bbox": [ + 131, + 394, + 481, + 418 + ], + "type": "text", + "content": "Based on these considerations, we propose our SegPIC in the VAE framework incorporated with two proposed modules: SAL and RAT." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 137, + 426, + 481, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 426, + 481, + 533 + ], + "spans": [ + { + "bbox": [ + 137, + 426, + 481, + 533 + ], + "type": "text", + "content": "- SAL: First of all, we extract a prototype from each region, which will be compressed and transmitted to be accessible on both the encoder and decoder sides. The prototypes are expected to extract high-level information from regions to ensure the representative ability as much as possible. However, the encoder in compression frameworks often extracts features that contain low-level information. Therefore, a module is highly desirable for communicating information across different levels. SAL is designed for extracting more abundant semantics and contexts, enriching the prototypes, and connecting features from two different levels." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 137, + 533, + 481, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 533, + 481, + 605 + ], + "spans": [ + { + "bbox": [ + 137, + 533, + 481, + 605 + ], + "type": "text", + "content": "- RAT: performing the Depth&Point-wise Separable Convolution (DPSCov), a novel convolution whose weights are generated from the prototypes and the context of latent features under the guidance of masks. Note that, we only utilize the masks to guide RAT to learn more compression-friendly semantic priors in training, while during the inference phase, we instead adopt uniform grid partitions as coarse alternatives." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 131, + 622, + 222, + 634 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 622, + 222, + 634 + ], + "spans": [ + { + "bbox": [ + 131, + 622, + 222, + 634 + ], + "type": "text", + "content": "3.2 Formulation" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "content": "In this section, we will introduce the problem and our method in mathematical form. Given an image " + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "content": ", the encoder " + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "inline_equation", + "content": "En" + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "content": " maps it to latent features " + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "content": ". After" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "type": "text", + "content": "Y. Liu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "content": "quantization " + }, + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "inline_equation", + "content": "\\hat{y}" + }, + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "content": " will be compressed to the bitstream by the entropy coding, such as arithmetic coding (AC). Then the receiver converts the bitstream back into " + }, + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "inline_equation", + "content": "\\hat{y}" + }, + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "content": " by entropy decoding. The above process can be formulated as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 237, + 160, + 341, + 173 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 237, + 160, + 341, + 173 + ], + "spans": [ + { + "bbox": [ + 237, + 160, + 341, + 173 + ], + "type": "interline_equation", + "content": "y = E n (x), \\quad \\hat {y} = Q (y),", + "image_path": "7c1b5284e6cd6181daf0b0f8c683de125370f139c9032d500de27e3fe2af2f3b.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 237, + 175, + 480, + 188 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 237, + 175, + 480, + 188 + ], + "spans": [ + { + "bbox": [ + 237, + 175, + 480, + 188 + ], + "type": "text", + "content": "Transmitter: " + }, + { + "bbox": [ + 237, + 175, + 480, + 188 + ], + "type": "inline_equation", + "content": "bit_{\\hat{y}} = AE(\\hat{y} |p_{\\hat{y}})" + }, + { + "bbox": [ + 237, + 175, + 480, + 188 + ], + "type": "text", + "content": " (1)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 237, + 190, + 376, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 237, + 190, + 376, + 203 + ], + "spans": [ + { + "bbox": [ + 237, + 190, + 376, + 203 + ], + "type": "text", + "content": "Receiver: " + }, + { + "bbox": [ + 237, + 190, + 376, + 203 + ], + "type": "inline_equation", + "content": "\\hat{y} = AD(bit_{\\hat{y}}|p_{\\hat{y}})" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 210, + 480, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 210, + 480, + 258 + ], + "spans": [ + { + "bbox": [ + 131, + 210, + 480, + 258 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 210, + 480, + 258 + ], + "type": "inline_equation", + "content": "p_{\\hat{y}}" + }, + { + "bbox": [ + 131, + 210, + 480, + 258 + ], + "type": "text", + "content": " is the estimated distribution of " + }, + { + "bbox": [ + 131, + 210, + 480, + 258 + ], + "type": "inline_equation", + "content": "\\hat{y}" + }, + { + "bbox": [ + 131, + 210, + 480, + 258 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 131, + 210, + 480, + 258 + ], + "type": "inline_equation", + "content": "\\text{bit}_{\\hat{y}}" + }, + { + "bbox": [ + 131, + 210, + 480, + 258 + ], + "type": "text", + "content": " is the binary bitstream of " + }, + { + "bbox": [ + 131, + 210, + 480, + 258 + ], + "type": "inline_equation", + "content": "\\hat{y}" + }, + { + "bbox": [ + 131, + 210, + 480, + 258 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 131, + 210, + 480, + 258 + ], + "type": "inline_equation", + "content": "AE" + }, + { + "bbox": [ + 131, + 210, + 480, + 258 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 210, + 480, + 258 + ], + "type": "inline_equation", + "content": "AD" + }, + { + "bbox": [ + 131, + 210, + 480, + 258 + ], + "type": "text", + "content": " are arithmetic encoding and decoding. We use a single Gaussian distribution to estimate " + }, + { + "bbox": [ + 131, + 210, + 480, + 258 + ], + "type": "inline_equation", + "content": "\\hat{y} \\sim \\mathcal{N}(\\mu, \\sigma^2)" + }, + { + "bbox": [ + 131, + 210, + 480, + 258 + ], + "type": "text", + "content": ", and the mean and variance value " + }, + { + "bbox": [ + 131, + 210, + 480, + 258 + ], + "type": "inline_equation", + "content": "\\mu, \\sigma^2" + }, + { + "bbox": [ + 131, + 210, + 480, + 258 + ], + "type": "text", + "content": " will be yielded through an auto-regressive mechanism. The formulation is:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 199, + 266, + 480, + 280 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 199, + 266, + 480, + 280 + ], + "spans": [ + { + "bbox": [ + 199, + 266, + 480, + 280 + ], + "type": "interline_equation", + "content": "z = E n _ {z} (y), \\hat {z} = Q (z), \\quad \\mu_ {i}, \\sigma_ {i} = F \\left(D e _ {z} (\\hat {z}), \\hat {y} _ {< i}\\right), \\tag {2}", + "image_path": "6e4de29c5bfd575e9fe98ce9fea9be65111180d455b609132cd69afdb76c0270.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 287, + 480, + 358 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 287, + 480, + 358 + ], + "spans": [ + { + "bbox": [ + 131, + 287, + 480, + 358 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 287, + 480, + 358 + ], + "type": "inline_equation", + "content": "En_{z}, De_{z}" + }, + { + "bbox": [ + 131, + 287, + 480, + 358 + ], + "type": "text", + "content": " is hyper encoder and decoder, " + }, + { + "bbox": [ + 131, + 287, + 480, + 358 + ], + "type": "inline_equation", + "content": "\\mu_{i}, \\sigma_{i}" + }, + { + "bbox": [ + 131, + 287, + 480, + 358 + ], + "type": "text", + "content": " are the mean and variance of " + }, + { + "bbox": [ + 131, + 287, + 480, + 358 + ], + "type": "inline_equation", + "content": "\\hat{y}_{i} \\sim \\mathcal{N}(\\mu_{i}, \\sigma_{i}^{2})" + }, + { + "bbox": [ + 131, + 287, + 480, + 358 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 131, + 287, + 480, + 358 + ], + "type": "inline_equation", + "content": "\\hat{y}_{i}" + }, + { + "bbox": [ + 131, + 287, + 480, + 358 + ], + "type": "text", + "content": " is the currently decoded portion of " + }, + { + "bbox": [ + 131, + 287, + 480, + 358 + ], + "type": "inline_equation", + "content": "\\hat{y}" + }, + { + "bbox": [ + 131, + 287, + 480, + 358 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 131, + 287, + 480, + 358 + ], + "type": "inline_equation", + "content": "\\hat{y}_{< i}" + }, + { + "bbox": [ + 131, + 287, + 480, + 358 + ], + "type": "text", + "content": " is the already decoded portion to assist decoding " + }, + { + "bbox": [ + 131, + 287, + 480, + 358 + ], + "type": "inline_equation", + "content": "\\hat{y}_{i}" + }, + { + "bbox": [ + 131, + 287, + 480, + 358 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 131, + 287, + 480, + 358 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 131, + 287, + 480, + 358 + ], + "type": "text", + "content": " is the auto-regressive entropy model. The AC process of " + }, + { + "bbox": [ + 131, + 287, + 480, + 358 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 131, + 287, + 480, + 358 + ], + "type": "text", + "content": " in Eq. 2 is omitted for brevity. Finally, the receiver adopts decoder " + }, + { + "bbox": [ + 131, + 287, + 480, + 358 + ], + "type": "inline_equation", + "content": "De" + }, + { + "bbox": [ + 131, + 287, + 480, + 358 + ], + "type": "text", + "content": " to reconstruct the image " + }, + { + "bbox": [ + 131, + 287, + 480, + 358 + ], + "type": "inline_equation", + "content": "\\hat{x}" + }, + { + "bbox": [ + 131, + 287, + 480, + 358 + ], + "type": "text", + "content": ", which is formulated as " + }, + { + "bbox": [ + 131, + 287, + 480, + 358 + ], + "type": "inline_equation", + "content": "\\hat{x} = De(\\hat{y})" + }, + { + "bbox": [ + 131, + 287, + 480, + 358 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 146, + 360, + 358, + 371 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 360, + 358, + 371 + ], + "spans": [ + { + "bbox": [ + 146, + 360, + 358, + 371 + ], + "type": "text", + "content": "The loss function " + }, + { + "bbox": [ + 146, + 360, + 358, + 371 + ], + "type": "inline_equation", + "content": "\\mathcal{L}" + }, + { + "bbox": [ + 146, + 360, + 358, + 371 + ], + "type": "text", + "content": " to optimize this problem is:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 195, + 380, + 480, + 393 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 195, + 380, + 480, + 393 + ], + "spans": [ + { + "bbox": [ + 195, + 380, + 480, + 393 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = R + \\lambda \\cdot D = R (\\hat {y}) + R (\\hat {z}) + \\lambda \\cdot \\mathbb {E} _ {x \\sim p _ {x}} [ d (x, \\hat {x}) ], \\tag {3}", + "image_path": "e1511766e0c8bdaa1f389b9b5c8487185fee6dbb9b14254fa2b1b59bdd99b76a.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 131, + 400, + 479, + 435 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 400, + 479, + 435 + ], + "spans": [ + { + "bbox": [ + 131, + 400, + 479, + 435 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 400, + 479, + 435 + ], + "type": "inline_equation", + "content": "R(\\hat{y}), R(\\hat{z})" + }, + { + "bbox": [ + 131, + 400, + 479, + 435 + ], + "type": "text", + "content": " are the bitrate of " + }, + { + "bbox": [ + 131, + 400, + 479, + 435 + ], + "type": "inline_equation", + "content": "\\hat{y}, \\hat{z}" + }, + { + "bbox": [ + 131, + 400, + 479, + 435 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 131, + 400, + 479, + 435 + ], + "type": "inline_equation", + "content": "d(x, \\hat{x})" + }, + { + "bbox": [ + 131, + 400, + 479, + 435 + ], + "type": "text", + "content": " is the distortion between " + }, + { + "bbox": [ + 131, + 400, + 479, + 435 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 131, + 400, + 479, + 435 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 400, + 479, + 435 + ], + "type": "inline_equation", + "content": "\\hat{x}" + }, + { + "bbox": [ + 131, + 400, + 479, + 435 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 131, + 400, + 479, + 435 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 131, + 400, + 479, + 435 + ], + "type": "text", + "content": " denotes different rate-distortion trade-offs. Here we use Mean Square Error (MSE) or (1-MS-SSIM) as the distortion function." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 437, + 480, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 437, + 480, + 533 + ], + "spans": [ + { + "bbox": [ + 130, + 437, + 480, + 533 + ], + "type": "text", + "content": "In our implementation, we use class-agnostic semantic masks (denoted as " + }, + { + "bbox": [ + 130, + 437, + 480, + 533 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 130, + 437, + 480, + 533 + ], + "type": "text", + "content": ") as additional inputs of the encoding " + }, + { + "bbox": [ + 130, + 437, + 480, + 533 + ], + "type": "inline_equation", + "content": "En" + }, + { + "bbox": [ + 130, + 437, + 480, + 533 + ], + "type": "text", + "content": " during the training phase. Moreover, we extract a prototype from each region at the middle layer of " + }, + { + "bbox": [ + 130, + 437, + 480, + 533 + ], + "type": "inline_equation", + "content": "En" + }, + { + "bbox": [ + 130, + 437, + 480, + 533 + ], + "type": "text", + "content": ", then compress and transmit them (denoted as " + }, + { + "bbox": [ + 130, + 437, + 480, + 533 + ], + "type": "inline_equation", + "content": "\\hat{p}" + }, + { + "bbox": [ + 130, + 437, + 480, + 533 + ], + "type": "text", + "content": "). In inference, we replace the masks with uniform grid partitions, so there is no additional bitstream for masks, while the bitrates of the prototypes also need to be taken into account. The " + }, + { + "bbox": [ + 130, + 437, + 480, + 533 + ], + "type": "inline_equation", + "content": "En" + }, + { + "bbox": [ + 130, + 437, + 480, + 533 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 437, + 480, + 533 + ], + "type": "inline_equation", + "content": "De" + }, + { + "bbox": [ + 130, + 437, + 480, + 533 + ], + "type": "text", + "content": " will transform the image with the assistance of the prototypes. The formulations requiring adjustment are as follows:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 231, + 540, + 383, + 553 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 231, + 540, + 383, + 553 + ], + "spans": [ + { + "bbox": [ + 231, + 540, + 383, + 553 + ], + "type": "interline_equation", + "content": "y = E n (x, \\hat {p} | m), \\quad \\hat {x} = D e (\\hat {y}, \\hat {p} | m),", + "image_path": "b2bc13d05d8550df427efc35dba6e49fe70ca133598f61f5cadebfc9c5d9356d.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 231, + 555, + 480, + 568 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 231, + 555, + 480, + 568 + ], + "spans": [ + { + "bbox": [ + 231, + 555, + 480, + 568 + ], + "type": "interline_equation", + "content": "\\mu_ {i}, \\sigma_ {i} = F (D e _ {z} (\\hat {z}), \\hat {y} _ {< i}, \\hat {p} | m), \\tag {4}", + "image_path": "491d35835f068c15d02fb0ccf16d3706063dd2c2f4bcc240039cea2eb2761779.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 231, + 570, + 342, + 582 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 231, + 570, + 342, + 582 + ], + "spans": [ + { + "bbox": [ + 231, + 570, + 342, + 582 + ], + "type": "interline_equation", + "content": "R = R (\\hat {p}) + R (\\hat {y}) + R (\\hat {z}).", + "image_path": "1893262b4c3b2fd179bcd73664725ba0e141aa672602c5708e3efb7ade66702e.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 598, + 241, + 610 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 598, + 241, + 610 + ], + "spans": [ + { + "bbox": [ + 132, + 598, + 241, + 610 + ], + "type": "text", + "content": "3.3 Network Design" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 130, + 617, + 479, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 479, + 641 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 479, + 641 + ], + "type": "text", + "content": "In this section, we will introduce the concrete process and our network design. We use [49] as the backbone. The overall architecture is in Fig. 2." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 130, + 642, + 479, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 642, + 479, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 642, + 479, + 665 + ], + "type": "text", + "content": "Scale Affine Layer. The Encoder and Decoder are responsible for image transform and reconstruction, each comprising 4 downsampling or upsampling blocks." + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 134, + 91, + 448, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 91, + 448, + 103 + ], + "spans": [ + { + "bbox": [ + 134, + 91, + 448, + 103 + ], + "type": "text", + "content": "Region-Adaptive Transform with Segmentation Prior for Image Compression" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 482, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 482, + 200 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 482, + 200 + ], + "type": "text", + "content": "The previous sampling blocks consisted of a single sample convolution layer, aimed at modifying the data size. However, this may not be sufficient to extract adequate semantics and contexts. Inspired by Spatial Feature Transform (SFT) [45], which can generate visually pleasing textures in segmentation-guided low-level vision, we propose to add a Scale Affine Layer (SAL) between two adjacent convolution layers (see Fig. 2 (b) (c)). It consists of two " + }, + { + "bbox": [ + 132, + 116, + 482, + 200 + ], + "type": "inline_equation", + "content": "1 \\times 1" + }, + { + "bbox": [ + 132, + 116, + 482, + 200 + ], + "type": "text", + "content": " convolution layers with a GELU activation function in between, which is formulated as:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 230, + 206, + 383, + 235 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 230, + 206, + 383, + 235 + ], + "spans": [ + { + "bbox": [ + 230, + 206, + 383, + 235 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} X _ {s} = \\operatorname {S A L} _ {s} (X), \\quad X _ {b} = \\operatorname {S A L} _ {b} (X), \\\\ Y = X _ {s} \\otimes X + X _ {b}, \\\\ \\end{array}", + "image_path": "d848d45115ed9b98f908b40c9e88b3d7c9d9725b81a9e1f004dae451a0ac210c.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 241, + 481, + 264 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 241, + 481, + 264 + ], + "spans": [ + { + "bbox": [ + 130, + 241, + 481, + 264 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 241, + 481, + 264 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 130, + 241, + 481, + 264 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 241, + 481, + 264 + ], + "type": "inline_equation", + "content": "Y" + }, + { + "bbox": [ + 130, + 241, + 481, + 264 + ], + "type": "text", + "content": " represent the input and output, " + }, + { + "bbox": [ + 130, + 241, + 481, + 264 + ], + "type": "inline_equation", + "content": "X_{s}" + }, + { + "bbox": [ + 130, + 241, + 481, + 264 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 241, + 481, + 264 + ], + "type": "inline_equation", + "content": "X_{b}" + }, + { + "bbox": [ + 130, + 241, + 481, + 264 + ], + "type": "text", + "content": " denote the scales and biases generated by SAL, " + }, + { + "bbox": [ + 130, + 241, + 481, + 264 + ], + "type": "inline_equation", + "content": "\\otimes" + }, + { + "bbox": [ + 130, + 241, + 481, + 264 + ], + "type": "text", + "content": " denotes the element-wise product." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 265, + 482, + 349 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 265, + 482, + 349 + ], + "spans": [ + { + "bbox": [ + 130, + 265, + 482, + 349 + ], + "type": "text", + "content": "Extract and Transmit the Prototypes. The Encoder first converts the original image " + }, + { + "bbox": [ + 130, + 265, + 482, + 349 + ], + "type": "inline_equation", + "content": "x \\in \\mathbb{R}^{3 \\times H \\times W}" + }, + { + "bbox": [ + 130, + 265, + 482, + 349 + ], + "type": "text", + "content": " into latent features " + }, + { + "bbox": [ + 130, + 265, + 482, + 349 + ], + "type": "inline_equation", + "content": "y' \\in \\mathbb{R}^{N \\times H/4 \\times W/4}" + }, + { + "bbox": [ + 130, + 265, + 482, + 349 + ], + "type": "text", + "content": " at the middle layer. Then we conduct Masked Average Pooling (MAP) on " + }, + { + "bbox": [ + 130, + 265, + 482, + 349 + ], + "type": "inline_equation", + "content": "y'" + }, + { + "bbox": [ + 130, + 265, + 482, + 349 + ], + "type": "text", + "content": " through the class-agnostic masks " + }, + { + "bbox": [ + 130, + 265, + 482, + 349 + ], + "type": "inline_equation", + "content": "m \\in \\{0,1\\}^{n \\times H \\times W}" + }, + { + "bbox": [ + 130, + 265, + 482, + 349 + ], + "type": "text", + "content": " (see Fig. 3 (a)), to acquire the prototypes " + }, + { + "bbox": [ + 130, + 265, + 482, + 349 + ], + "type": "inline_equation", + "content": "p' \\in \\mathbb{R}^{N \\times n}" + }, + { + "bbox": [ + 130, + 265, + 482, + 349 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 265, + 482, + 349 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 130, + 265, + 482, + 349 + ], + "type": "text", + "content": " is the number of the masks. Thanks to the proposed SAL, the Encoder becomes deeper, allowing " + }, + { + "bbox": [ + 130, + 265, + 482, + 349 + ], + "type": "inline_equation", + "content": "p'" + }, + { + "bbox": [ + 130, + 265, + 482, + 349 + ], + "type": "text", + "content": " to incorporate more semantics and contexts from the corresponding mask. The formulation of MAP is as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 190, + 354, + 481, + 380 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 190, + 354, + 481, + 380 + ], + "spans": [ + { + "bbox": [ + 190, + 354, + 481, + 380 + ], + "type": "interline_equation", + "content": "p ^ {\\prime} = \\operatorname {M A P} \\left(y ^ {\\prime}, m\\right) = \\left\\{\\frac {1}{\\sum m ^ {(i)}} \\sum y ^ {\\prime} \\otimes m ^ {(i)}, \\dots \\right\\} _ {i \\in [ 1, n ]}, \\tag {5}", + "image_path": "6bc9a27f32ad7ef39f0a2cbab153e5db2a88df0e814952237e90faa337b0d1d6.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 385, + 482, + 446 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 385, + 482, + 446 + ], + "spans": [ + { + "bbox": [ + 130, + 385, + 482, + 446 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 385, + 482, + 446 + ], + "type": "inline_equation", + "content": "y'" + }, + { + "bbox": [ + 130, + 385, + 482, + 446 + ], + "type": "text", + "content": " are the mid-level features (see Fig. 2 (a)), " + }, + { + "bbox": [ + 130, + 385, + 482, + 446 + ], + "type": "inline_equation", + "content": "\\sum" + }, + { + "bbox": [ + 130, + 385, + 482, + 446 + ], + "type": "text", + "content": " denotes summation along the spatial dimension. " + }, + { + "bbox": [ + 130, + 385, + 482, + 446 + ], + "type": "inline_equation", + "content": "p'" + }, + { + "bbox": [ + 130, + 385, + 482, + 446 + ], + "type": "text", + "content": " will be compressed and transmitted to the receiver (see Fig. 4), and the reconstructed " + }, + { + "bbox": [ + 130, + 385, + 482, + 446 + ], + "type": "inline_equation", + "content": "\\hat{p}'" + }, + { + "bbox": [ + 130, + 385, + 482, + 446 + ], + "type": "text", + "content": " is shared between the Encoder and Decoder. The channel " + }, + { + "bbox": [ + 130, + 385, + 482, + 446 + ], + "type": "inline_equation", + "content": "\\{C_1, C_2, C_3, C_4\\}" + }, + { + "bbox": [ + 130, + 385, + 482, + 446 + ], + "type": "text", + "content": " is {192, 128, 96, 96} for " + }, + { + "bbox": [ + 130, + 385, + 482, + 446 + ], + "type": "inline_equation", + "content": "p'" + }, + { + "bbox": [ + 130, + 385, + 482, + 446 + ], + "type": "text", + "content": " (on Encoder and Decoder) and {320, 192, 128, 96} for " + }, + { + "bbox": [ + 130, + 385, + 482, + 446 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 130, + 385, + 482, + 446 + ], + "type": "text", + "content": " (on Entropy Model)." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 446, + 482, + 493 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 446, + 482, + 493 + ], + "spans": [ + { + "bbox": [ + 130, + 446, + 482, + 493 + ], + "type": "text", + "content": "Region-Adaptive Transform. Here we propose a novel convolution called Depth&Point-wise Separable Convolution (DPSCov), in which each element has an individual convolution kernel, and there is no channel-wise mutual communication for computational efficiency." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 494, + 482, + 541 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 494, + 482, + 541 + ], + "spans": [ + { + "bbox": [ + 130, + 494, + 482, + 541 + ], + "type": "text", + "content": "Before introducing it, we will start with standard convolution briefly. The input of standard convolution can be denoted as " + }, + { + "bbox": [ + 130, + 494, + 482, + 541 + ], + "type": "inline_equation", + "content": "X \\in \\mathbb{R}^{C \\times H \\times W}" + }, + { + "bbox": [ + 130, + 494, + 482, + 541 + ], + "type": "text", + "content": ", the output is " + }, + { + "bbox": [ + 130, + 494, + 482, + 541 + ], + "type": "inline_equation", + "content": "Y \\in \\mathbb{R}^{O \\times H \\times W}" + }, + { + "bbox": [ + 130, + 494, + 482, + 541 + ], + "type": "text", + "content": ", and the kernel is " + }, + { + "bbox": [ + 130, + 494, + 482, + 541 + ], + "type": "inline_equation", + "content": "W \\in \\mathbb{R}^{O \\times C \\times k \\times k}" + }, + { + "bbox": [ + 130, + 494, + 482, + 541 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 494, + 482, + 541 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 494, + 482, + 541 + ], + "type": "text", + "content": " is kernel size. The formulation of standard convolution is as follows:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 182, + 546, + 432, + 580 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 182, + 546, + 432, + 580 + ], + "spans": [ + { + "bbox": [ + 182, + 546, + 432, + 580 + ], + "type": "interline_equation", + "content": "Y _ {o, h, w} = \\sum_ {c = 1} ^ {C} X _ {c, h, w} * W _ {o, c} = \\sum_ {c = 1} ^ {C} \\sum_ {i, j} X _ {c, h + i, w + j} \\cdot W _ {o, c, i, j},", + "image_path": "22bbdfcbac8143586d699aabf0b525d3da4cbcc9b5aa819b64548551e16a8b97.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 586, + 481, + 610 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 586, + 481, + 610 + ], + "spans": [ + { + "bbox": [ + 130, + 586, + 481, + 610 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 586, + 481, + 610 + ], + "type": "inline_equation", + "content": "*" + }, + { + "bbox": [ + 130, + 586, + 481, + 610 + ], + "type": "text", + "content": " is 2D convolution, " + }, + { + "bbox": [ + 130, + 586, + 481, + 610 + ], + "type": "inline_equation", + "content": "i,j\\in [-\\lfloor k / 2\\rfloor ,\\lfloor k / 2\\rfloor ]" + }, + { + "bbox": [ + 130, + 586, + 481, + 610 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 586, + 481, + 610 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 586, + 481, + 610 + ], + "type": "text", + "content": " is kernel size, and " + }, + { + "bbox": [ + 130, + 586, + 481, + 610 + ], + "type": "inline_equation", + "content": "\\lfloor \\cdot \\rfloor" + }, + { + "bbox": [ + 130, + 586, + 481, + 610 + ], + "type": "text", + "content": " denotes floor rounding." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 131, + 611, + 481, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 611, + 481, + 635 + ], + "spans": [ + { + "bbox": [ + 131, + 611, + 481, + 635 + ], + "type": "text", + "content": "DPSCov can be seen as a combination of Depthwise Separable Convolution (DWConv) [13] and Dynamic Convolution [22]. The formulation is as follows:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 189, + 641, + 481, + 669 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 189, + 641, + 481, + 669 + ], + "spans": [ + { + "bbox": [ + 189, + 641, + 481, + 669 + ], + "type": "interline_equation", + "content": "Y _ {c, h, w} = X _ {c, h, w} * W _ {c, h, w} = \\sum_ {i, j} X _ {c, h + i, w + j} \\cdot W _ {c, h, w, i, j}. \\tag {6}", + "image_path": "38809fb900e1fbe084c3ed67653c3c780af35ffe1f513aeff735b06e69869ed6.jpg" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "type": "text", + "content": "Y. Liu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 112, + 294, + 232 + ], + "blocks": [ + { + "bbox": [ + 133, + 112, + 294, + 232 + ], + "lines": [ + { + "bbox": [ + 133, + 112, + 294, + 232 + ], + "spans": [ + { + "bbox": [ + 133, + 112, + 294, + 232 + ], + "type": "image", + "image_path": "8cee4801ee1779f81fd5971f4525b7ddbd395b93ef9a6ae76d5ce8a6a4fef5d5.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 239, + 289, + 248 + ], + "lines": [ + { + "bbox": [ + 132, + 239, + 289, + 248 + ], + "spans": [ + { + "bbox": [ + 132, + 239, + 289, + 248 + ], + "type": "text", + "content": "(a) Extract the prototypes through Masked Average Pooling" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 307, + 113, + 476, + 232 + ], + "blocks": [ + { + "bbox": [ + 307, + 113, + 476, + 232 + ], + "lines": [ + { + "bbox": [ + 307, + 113, + 476, + 232 + ], + "spans": [ + { + "bbox": [ + 307, + 113, + 476, + 232 + ], + "type": "image", + "image_path": "5e12f059022ed5e0379e315379cb45c246c1bec341489f0cee517a70babe7e81.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 297, + 239, + 481, + 248 + ], + "lines": [ + { + "bbox": [ + 297, + 239, + 481, + 248 + ], + "spans": [ + { + "bbox": [ + 297, + 239, + 481, + 248 + ], + "type": "text", + "content": "(b)The Region-Adaptive Transform (RAT) with masks and prototypes" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 132, + 256, + 482, + 312 + ], + "lines": [ + { + "bbox": [ + 132, + 256, + 482, + 312 + ], + "spans": [ + { + "bbox": [ + 132, + 256, + 482, + 312 + ], + "type": "text", + "content": "Fig. 3: The diagram of extracting the prototypes and the proposed RAT module. DPSCov is proposed Depth&Point-wise Separable Convolution (see Eq. 6). CTL is Channel Transform Layers, DKG is the DPSCov Kernel Generator, and CAG is the Channel Attention Generator. DPSCov is Depth&Point-wise Separable Convolution. The detailed modules are represented in Fig. 5." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 344, + 480, + 369 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 344, + 480, + 369 + ], + "spans": [ + { + "bbox": [ + 130, + 344, + 480, + 369 + ], + "type": "text", + "content": "The computational complexity of standard convolution is " + }, + { + "bbox": [ + 130, + 344, + 480, + 369 + ], + "type": "inline_equation", + "content": "\\mathrm{O}(O \\times C \\times H \\times W \\times k^2)" + }, + { + "bbox": [ + 130, + 344, + 480, + 369 + ], + "type": "text", + "content": ", while DPSCov is " + }, + { + "bbox": [ + 130, + 344, + 480, + 369 + ], + "type": "inline_equation", + "content": "\\mathrm{O}(C \\times H \\times W \\times k^2)" + }, + { + "bbox": [ + 130, + 344, + 480, + 369 + ], + "type": "text", + "content": ", same as DWConv." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 373, + 482, + 541 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 373, + 482, + 541 + ], + "spans": [ + { + "bbox": [ + 130, + 373, + 482, + 541 + ], + "type": "text", + "content": "The overall design of RAT is depicted in Fig. 3 (b). We expand the prototypes based on the positional information of the masks, filling in the corresponding copies of the prototypes at each pixel position. Then, we feed-forward the input through Channel Transform Layers (CTL) and concatenate it with the expanded features. The resulting variable, which contains both semantics and contexts, is used for subsequent filter generation. The DPSCConv Kernels Generator (DKG) is designed for generating the filters. In the first layer of DKG, we employ a 1x1 convolution to attempt the fusion of the prototype and the pixel's intrinsic attribute. We opt for grouped convolution in the final layer of the DKG to reduce the computation and parameters. Then Channel Attention Generator (CAG) generates the channel attention to be multiplied with the output of DPSCConv along the channel dimension. Finally, the output passes through a 1x1 convolution to merge channel information. We apply the RAT in both the Encoder and Decoder, as well as the entropy encoding module." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": "Entropy Encoding Module. We use the Channel-wise Auto-Regressive Model (ChARM) [34] as our entropy estimation module, and the specific hyperparameter configuration follows [49]. In this model, the latent variable " + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": " is partitioned into 10 groups along the channel dimension. The Gaussian prior distribution " + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "inline_equation", + "content": "p(y_{i}|\\hat{z},\\hat{y} < i)" + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": " is estimated in an auto-regressive manner, where the mean and scale of " + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "inline_equation", + "content": "y_{i}" + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": " depend on the quantized latent in the preceding groups " + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\hat{y}_{< i}" + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": " (see Eq. 2). To further enhance the estimation ability, we also deploy RAT in it. The RAT here can transmit key region prototypes through a separate bitstream, assisting in more efficient encoding and decoding. The prototypes " + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": " could be seen as segmentation-prior, a supplement of the hyper-prior information " + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 135, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 135, + 91, + 447, + 102 + ], + "type": "text", + "content": "Region-Adaptive Transform with Segmentation Prior for Image Compression" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 116, + 272, + 219 + ], + "blocks": [ + { + "bbox": [ + 133, + 116, + 272, + 219 + ], + "lines": [ + { + "bbox": [ + 133, + 116, + 272, + 219 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 272, + 219 + ], + "type": "image", + "image_path": "eb4a3ab36c142d7645b2c53cf5111bf45853c23d9711ea164bb525b232e7a4a5.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 227, + 274, + 293 + ], + "lines": [ + { + "bbox": [ + 131, + 227, + 274, + 293 + ], + "spans": [ + { + "bbox": [ + 131, + 227, + 274, + 293 + ], + "type": "text", + "content": "Fig.4: The architecture of the Prototype Encoder and Decoder for compressing and transmitting the prototypes. \"Linear " + }, + { + "bbox": [ + 131, + 227, + 274, + 293 + ], + "type": "inline_equation", + "content": "\\mathrm{C_1,C_2}" + }, + { + "bbox": [ + 131, + 227, + 274, + 293 + ], + "type": "text", + "content": "\" means linear layer with in-channel " + }, + { + "bbox": [ + 131, + 227, + 274, + 293 + ], + "type": "inline_equation", + "content": "\\mathrm{C_1}" + }, + { + "bbox": [ + 131, + 227, + 274, + 293 + ], + "type": "text", + "content": " and out-channel " + }, + { + "bbox": [ + 131, + 227, + 274, + 293 + ], + "type": "inline_equation", + "content": "\\mathrm{C_2}" + }, + { + "bbox": [ + 131, + 227, + 274, + 293 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 304, + 115, + 462, + 208 + ], + "blocks": [ + { + "bbox": [ + 304, + 115, + 462, + 208 + ], + "lines": [ + { + "bbox": [ + 304, + 115, + 462, + 208 + ], + "spans": [ + { + "bbox": [ + 304, + 115, + 462, + 208 + ], + "type": "image", + "image_path": "21a49e6f4ff40e1fe858fa64914610b261b5d57059e6895d3c2eeb78266c0394.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 282, + 217, + 484, + 293 + ], + "lines": [ + { + "bbox": [ + 282, + 217, + 484, + 293 + ], + "spans": [ + { + "bbox": [ + 282, + 217, + 484, + 293 + ], + "type": "text", + "content": "Fig. 5: The detailed modules in RAT, including Channel Transform layers (CTL), DPSCConv Kernels Generator (DKG), and Channel Attention Generator (CAG). LReLU is Leaky ReLU. GAP is Global Average Pooling. \"GConv k²C,3,1,C\" means Grouped Convolution with out-channel k²C, kernel size 3, stride 1 and group C." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 140, + 337, + 471, + 456 + ], + "blocks": [ + { + "bbox": [ + 131, + 304, + 482, + 327 + ], + "lines": [ + { + "bbox": [ + 131, + 304, + 482, + 327 + ], + "spans": [ + { + "bbox": [ + 131, + 304, + 482, + 327 + ], + "type": "text", + "content": "Table 1: BD-rate(%) performance on Kodak and CLIC Professional Valid (CLIC for short). The anchor is VTM-17.0." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 140, + 337, + 471, + 456 + ], + "lines": [ + { + "bbox": [ + 140, + 337, + 471, + 456 + ], + "spans": [ + { + "bbox": [ + 140, + 337, + 471, + 456 + ], + "type": "table", + "html": "
ModelKodak \nBD-rate↓CLIC \nBD-rate↓ModelKodak \nBD-rate↓CLIC \nBD-rate↓
BPG [7]+31.5+48.1EVC [44]+0.27-
Cheng2020 [12]+7.46+11.7Feng2023 [16]-2.21-
Agustsson2023 [2]+6.15-ELIC [18]-3.08-3.58
ChARM [34]+3.58-Contextformer [28]-5.38-
Wang2022 [43]+1.10-0.99Pan2022 [35]-5.42-
BL-S+15.5+16.6Zou2022 (BL) [49]-2.89+0.14
Ours-S+3.87+5.26Ours-8.18-8.26
", + "image_path": "c19946c3964124eeb05f949fa7c9eb96ed02ebe093abb87c070906ff3ed2f767.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 476, + 230, + 491 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 476, + 230, + 491 + ], + "spans": [ + { + "bbox": [ + 132, + 476, + 230, + 491 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 502, + 262, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 502, + 262, + 514 + ], + "spans": [ + { + "bbox": [ + 132, + 502, + 262, + 514 + ], + "type": "text", + "content": "4.1 Experimental Setup" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 521, + 482, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 482, + 629 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 482, + 629 + ], + "type": "text", + "content": "Training. We use COCO-Stuff [9] as the training dataset, which is composed of 118k images with panoptic semantic masks. We just use the class-agnostic masks without the category labels. We randomly crop the images with the size of " + }, + { + "bbox": [ + 130, + 521, + 482, + 629 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 130, + 521, + 482, + 629 + ], + "type": "text", + "content": ". The models are trained for 270 epochs (about 1M iterations) using the Adam optimizer [25] with a batch size of 32. The initial learning rate is set to " + }, + { + "bbox": [ + 130, + 521, + 482, + 629 + ], + "type": "inline_equation", + "content": "1 \\times 10^{-4}" + }, + { + "bbox": [ + 130, + 521, + 482, + 629 + ], + "type": "text", + "content": " for 170 epochs, and reduced by half every 30 epochs. Our models are optimized by the rate-distortion trade-off loss function (see Eq. 3 and Eq. 4). The " + }, + { + "bbox": [ + 130, + 521, + 482, + 629 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 130, + 521, + 482, + 629 + ], + "type": "text", + "content": " belongs to " + }, + { + "bbox": [ + 130, + 521, + 482, + 629 + ], + "type": "inline_equation", + "content": "\\{0.0018, 0.0035, 0.0067, 0.0130, 0.0250, 0.0483\\}" + }, + { + "bbox": [ + 130, + 521, + 482, + 629 + ], + "type": "text", + "content": " for MSE as the distortion, and " + }, + { + "bbox": [ + 130, + 521, + 482, + 629 + ], + "type": "inline_equation", + "content": "\\{2.4, 4.58, 8.73, 16.64, 31.73, 60.50\\}" + }, + { + "bbox": [ + 130, + 521, + 482, + 629 + ], + "type": "text", + "content": " for MS-SSIM [46]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 630, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 630, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 630, + 482, + 665 + ], + "type": "text", + "content": "Evaluation. We evaluate our model on two commonly used datasets, Kodak [27] with the size of " + }, + { + "bbox": [ + 130, + 630, + 482, + 665 + ], + "type": "inline_equation", + "content": "768 \\times 512" + }, + { + "bbox": [ + 130, + 630, + 482, + 665 + ], + "type": "text", + "content": ", and CLIC Professional Validation [1] with the size of 2K. Performance is evaluated using both bitrate and distortions. Distortion" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "text", + "content": "Y. Liu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 132, + 114, + 306, + 249 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 306, + 249 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 306, + 249 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 306, + 249 + ], + "type": "image", + "image_path": "3fab81b18d85b9af2da515dccb4e08f471e002a5c5255227512be96c01cd4a43.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 257, + 482, + 304 + ], + "lines": [ + { + "bbox": [ + 130, + 257, + 482, + 304 + ], + "spans": [ + { + "bbox": [ + 130, + 257, + 482, + 304 + ], + "type": "text", + "content": "Fig.6: Rate-Distortion Performance on Kodak [27] and CLIC Professional Valid dataset [1]. Our SegPIC introduces proposed RAT and SAL based on Zou2022 [49]. We also implement a simple baseline model without self-attention and auto-regressive mechanism (BL-S), and incorporate our proposed modules on it (Ours-S)." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 309, + 113, + 482, + 249 + ], + "blocks": [ + { + "bbox": [ + 309, + 113, + 482, + 249 + ], + "lines": [ + { + "bbox": [ + 309, + 113, + 482, + 249 + ], + "spans": [ + { + "bbox": [ + 309, + 113, + 482, + 249 + ], + "type": "image", + "image_path": "b15adf7dcb86af8218c86e4850e35b8412a5f81623f3c28950b6ab6d1ba4dc6a.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 323, + 482, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 323, + 482, + 373 + ], + "spans": [ + { + "bbox": [ + 130, + 323, + 482, + 373 + ], + "type": "text", + "content": "measurement involves PSNR and MS-SSIM. And bitrates are assessed in bits per pixel (bpp). In this phase, we replace the masks with " + }, + { + "bbox": [ + 130, + 323, + 482, + 373 + ], + "type": "inline_equation", + "content": "4 \\times 4" + }, + { + "bbox": [ + 130, + 323, + 482, + 373 + ], + "type": "text", + "content": " grid partitions. Thus, there is no bitrate cost related to the masks, while the bitrate of prototypes " + }, + { + "bbox": [ + 130, + 323, + 482, + 373 + ], + "type": "inline_equation", + "content": "R(p)" + }, + { + "bbox": [ + 130, + 323, + 482, + 373 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 323, + 482, + 373 + ], + "type": "inline_equation", + "content": "R(p^{\\prime})" + }, + { + "bbox": [ + 130, + 323, + 482, + 373 + ], + "type": "text", + "content": " is still taken into account." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 387, + 308, + 399 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 387, + 308, + 399 + ], + "spans": [ + { + "bbox": [ + 132, + 387, + 308, + 399 + ], + "type": "text", + "content": "4.2 Rate-Distortion Performance" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 404, + 482, + 524 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 404, + 482, + 524 + ], + "spans": [ + { + "bbox": [ + 130, + 404, + 482, + 524 + ], + "type": "text", + "content": "We compare our SegPIC with previously well-performing methods, including discretized Gaussian mixture likelihoods and attention modules [12] (denoted as Cheng2020), ChARM [34] (denoted as Minnen2022), ChARM with WAM [49] (denoted as Zou2022), image-specific data-dependent transform model [43] (denoted as Wang2022), the best-performing traditional method VTM-17.0 [24] and influential traditional method BPG [7]. Furthermore, we introduce two simplified models without self-attention and auto-regressive mechanism, denoted as baseline-small (BL-S) and the one incorporating our proposed SAL and RAT (Ours-S), to demonstrate the generality of our modules. Details about the architecture of these two models will be provided in the appendix." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 524, + 482, + 609 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 524, + 482, + 609 + ], + "spans": [ + { + "bbox": [ + 130, + 524, + 482, + 609 + ], + "type": "text", + "content": "As illustrated in Fig. 6 and Tab. 1, compared to VTM-17.0, our SegPIC could save " + }, + { + "bbox": [ + 130, + 524, + 482, + 609 + ], + "type": "inline_equation", + "content": "8.18\\%" + }, + { + "bbox": [ + 130, + 524, + 482, + 609 + ], + "type": "text", + "content": " BD-rate [8] on Kodak and " + }, + { + "bbox": [ + 130, + 524, + 482, + 609 + ], + "type": "inline_equation", + "content": "8.26\\%" + }, + { + "bbox": [ + 130, + 524, + 482, + 609 + ], + "type": "text", + "content": " on CLIC, which outperforms all the LIC and traditional methods mentioned above. To further verify the efficiency of the proposed RAT and SAL, We also compare SegPIC with the baseline Zou2022, which can save " + }, + { + "bbox": [ + 130, + 524, + 482, + 609 + ], + "type": "inline_equation", + "content": "5.45\\%" + }, + { + "bbox": [ + 130, + 524, + 482, + 609 + ], + "type": "text", + "content": " BD-rate on Kodak, and " + }, + { + "bbox": [ + 130, + 524, + 482, + 609 + ], + "type": "inline_equation", + "content": "8.50\\%" + }, + { + "bbox": [ + 130, + 524, + 482, + 609 + ], + "type": "text", + "content": " on CLIC (see Tab. 2). Moreover, Ours-S even achieves closely matched performance compared to ChARM. The MS-SSIM results are provided in the appendix." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 624, + 239, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 624, + 239, + 635 + ], + "spans": [ + { + "bbox": [ + 132, + 624, + 239, + 635 + ], + "type": "text", + "content": "4.3 Ablation Study" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "content": "To assess the contributions of our RAT and SAL modules individually, we independently incorporate these two modules into the baseline Zou2022. Conse" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 135, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 135, + 91, + 447, + 102 + ], + "type": "text", + "content": "Region-Adaptive Transform with Segmentation Prior for Image Compression" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 134, + 114, + 201, + 215 + ], + "blocks": [ + { + "bbox": [ + 134, + 114, + 201, + 215 + ], + "lines": [ + { + "bbox": [ + 134, + 114, + 201, + 215 + ], + "spans": [ + { + "bbox": [ + 134, + 114, + 201, + 215 + ], + "type": "image", + "image_path": "037fcdb385ee81d514b5e0afb34bc99719581bf3e0cfe5d0e1c0893b5a3de7ac.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 150, + 215, + 186, + 222 + ], + "lines": [ + { + "bbox": [ + 150, + 215, + 186, + 222 + ], + "spans": [ + { + "bbox": [ + 150, + 215, + 186, + 222 + ], + "type": "text", + "content": "Original kodim04" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 205, + 114, + 252, + 161 + ], + "blocks": [ + { + "bbox": [ + 205, + 114, + 252, + 161 + ], + "lines": [ + { + "bbox": [ + 205, + 114, + 252, + 161 + ], + "spans": [ + { + "bbox": [ + 205, + 114, + 252, + 161 + ], + "type": "image", + "image_path": "88f8f732716dcc89c4bd9a52b6ca000370f7b794d55f65ec43922b84eedd3817.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 214, + 161, + 242, + 167 + ], + "lines": [ + { + "bbox": [ + 214, + 161, + 242, + 167 + ], + "spans": [ + { + "bbox": [ + 214, + 161, + 242, + 167 + ], + "type": "text", + "content": "Ground Truth" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 256, + 115, + 302, + 160 + ], + "blocks": [ + { + "bbox": [ + 256, + 115, + 302, + 160 + ], + "lines": [ + { + "bbox": [ + 256, + 115, + 302, + 160 + ], + "spans": [ + { + "bbox": [ + 256, + 115, + 302, + 160 + ], + "type": "image", + "image_path": "3a515b43679d80026bbc295f4a634476d04ff5438809fd4c1800edac89ff2fa9.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 260, + 161, + 299, + 167 + ], + "lines": [ + { + "bbox": [ + 260, + 161, + 299, + 167 + ], + "spans": [ + { + "bbox": [ + 260, + 161, + 299, + 167 + ], + "type": "text", + "content": "VTM (31.02/0.115)" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 206, + 168, + 252, + 215 + ], + "blocks": [ + { + "bbox": [ + 206, + 168, + 252, + 215 + ], + "lines": [ + { + "bbox": [ + 206, + 168, + 252, + 215 + ], + "spans": [ + { + "bbox": [ + 206, + 168, + 252, + 215 + ], + "type": "image", + "image_path": "957f2fa822d29f159d60a8f265647b5825cc3fcc7e1b48cf532de5f76e37520b.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 206, + 215, + 252, + 222 + ], + "lines": [ + { + "bbox": [ + 206, + 215, + 252, + 222 + ], + "spans": [ + { + "bbox": [ + 206, + 215, + 252, + 222 + ], + "type": "text", + "content": "Zou2022 (31.46/0.119)" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 256, + 168, + 302, + 215 + ], + "blocks": [ + { + "bbox": [ + 256, + 168, + 302, + 215 + ], + "lines": [ + { + "bbox": [ + 256, + 168, + 302, + 215 + ], + "spans": [ + { + "bbox": [ + 256, + 168, + 302, + 215 + ], + "type": "image", + "image_path": "ed79477c33635b9b68f99243f22619e22f37dd47832808243f2fef3f99acbf92.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 260, + 215, + 298, + 221 + ], + "lines": [ + { + "bbox": [ + 260, + 215, + 298, + 221 + ], + "spans": [ + { + "bbox": [ + 260, + 215, + 298, + 221 + ], + "type": "text", + "content": "Ours (31.53/0.109)" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 310, + 115, + 378, + 216 + ], + "blocks": [ + { + "bbox": [ + 310, + 115, + 378, + 216 + ], + "lines": [ + { + "bbox": [ + 310, + 115, + 378, + 216 + ], + "spans": [ + { + "bbox": [ + 310, + 115, + 378, + 216 + ], + "type": "image", + "image_path": "93a90fa3b48708c694116e52b568ce4ba6ab0125dbd855173dd89607b9d29914.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 328, + 215, + 363, + 221 + ], + "lines": [ + { + "bbox": [ + 328, + 215, + 363, + 221 + ], + "spans": [ + { + "bbox": [ + 328, + 215, + 363, + 221 + ], + "type": "text", + "content": "Original kodim24" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 131, + 231, + 481, + 265 + ], + "lines": [ + { + "bbox": [ + 131, + 231, + 481, + 265 + ], + "spans": [ + { + "bbox": [ + 131, + 231, + 481, + 265 + ], + "type": "text", + "content": "Fig. 7: Visualization of the reconstructed images " + }, + { + "bbox": [ + 131, + 231, + 481, + 265 + ], + "type": "inline_equation", + "content": "\\text{kodim}04" + }, + { + "bbox": [ + 131, + 231, + 481, + 265 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 231, + 481, + 265 + ], + "type": "inline_equation", + "content": "\\text{kodim}24" + }, + { + "bbox": [ + 131, + 231, + 481, + 265 + ], + "type": "text", + "content": " in Kodak. The metrics are (PNSR↑/bpp↓). It shows that our SegPIC can distinguish the objects' contours more accurately, making the edges sharper with less bitrate." + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 382, + 115, + 429, + 161 + ], + "blocks": [ + { + "bbox": [ + 382, + 115, + 429, + 161 + ], + "lines": [ + { + "bbox": [ + 382, + 115, + 429, + 161 + ], + "spans": [ + { + "bbox": [ + 382, + 115, + 429, + 161 + ], + "type": "image", + "image_path": "a22d4a9b97a192cefbba941436eb88f77f6ae650adfa4d91dded3d1219e1d562.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 392, + 161, + 419, + 167 + ], + "lines": [ + { + "bbox": [ + 392, + 161, + 419, + 167 + ], + "spans": [ + { + "bbox": [ + 392, + 161, + 419, + 167 + ], + "type": "text", + "content": "Ground Truth" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 433, + 115, + 479, + 161 + ], + "blocks": [ + { + "bbox": [ + 433, + 115, + 479, + 161 + ], + "lines": [ + { + "bbox": [ + 433, + 115, + 479, + 161 + ], + "spans": [ + { + "bbox": [ + 433, + 115, + 479, + 161 + ], + "type": "image", + "image_path": "d5f179f9575f7e2ab0b96a0ac72120bd471c4b7534ff81294554c0e61edc9d3f.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 436, + 161, + 477, + 167 + ], + "lines": [ + { + "bbox": [ + 436, + 161, + 477, + 167 + ], + "spans": [ + { + "bbox": [ + 436, + 161, + 477, + 167 + ], + "type": "text", + "content": "VTM (28.96/0.325)" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 382, + 168, + 429, + 215 + ], + "blocks": [ + { + "bbox": [ + 382, + 168, + 429, + 215 + ], + "lines": [ + { + "bbox": [ + 382, + 168, + 429, + 215 + ], + "spans": [ + { + "bbox": [ + 382, + 168, + 429, + 215 + ], + "type": "image", + "image_path": "2f291c2d0f6b37ee88a44033e03a0064a7d949295b200a1a320a86779e756397.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 383, + 215, + 429, + 221 + ], + "lines": [ + { + "bbox": [ + 383, + 215, + 429, + 221 + ], + "spans": [ + { + "bbox": [ + 383, + 215, + 429, + 221 + ], + "type": "text", + "content": "Zou2022 (28.67/0.292)" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 433, + 168, + 479, + 215 + ], + "blocks": [ + { + "bbox": [ + 433, + 168, + 479, + 215 + ], + "lines": [ + { + "bbox": [ + 433, + 168, + 479, + 215 + ], + "spans": [ + { + "bbox": [ + 433, + 168, + 479, + 215 + ], + "type": "image", + "image_path": "c750ea1eaaf842512ddaed095f47d8cda2a9668fb97b1a22b0015b3eb01db6c1.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 436, + 215, + 475, + 221 + ], + "lines": [ + { + "bbox": [ + 436, + 215, + 475, + 221 + ], + "spans": [ + { + "bbox": [ + 436, + 215, + 475, + 221 + ], + "type": "text", + "content": "Ours (28.98/0.289)" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "type": "table", + "bbox": [ + 173, + 325, + 438, + 437 + ], + "blocks": [ + { + "bbox": [ + 131, + 293, + 482, + 316 + ], + "lines": [ + { + "bbox": [ + 131, + 293, + 482, + 316 + ], + "spans": [ + { + "bbox": [ + 131, + 293, + 482, + 316 + ], + "type": "text", + "content": "Table 2: Ablation Study on Kodak and CLIC. The Baseline (BL) is Zou2022 [49]. The speed is measured on Kodak with one Nvidia RTX 3090 GPU." + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 173, + 325, + 438, + 437 + ], + "lines": [ + { + "bbox": [ + 173, + 325, + 438, + 437 + ], + "spans": [ + { + "bbox": [ + 173, + 325, + 438, + 437 + ], + "type": "table", + "html": "
ModelKodakCLIC
ParamsBD-rateBD-rateEncDec Time
Zou2022 (BL)75.2 M0.00.0120 ms124 ms
BL w/ SAL76.7 M-2.12-2.93--
BL w/ RAT82.0 M-2.75-4.35--
Ours83.5 M-5.45-8.50142 ms130 ms
ChARM48.8 M+6.52-107 ms108 ms
BL-S21.7 M+19.0+16.3--
Ours-S29.2 M+6.96+5.1199 ms46 ms
", + "image_path": "a6686ae11827cfdd13446e77efab1a4aaad8be0515fcf9c851d80293a0b3d924.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "table_body" + } + ], + "index": 24 + }, + { + "bbox": [ + 130, + 477, + 482, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 477, + 482, + 574 + ], + "spans": [ + { + "bbox": [ + 130, + 477, + 482, + 574 + ], + "type": "text", + "content": "sequently, we construct two models: \"baseline w/ SAL\" and \"baseline w/ RAT.\" As illustrated in Tab. 2, the parameter counts introduced by RAT and SAL separately are " + }, + { + "bbox": [ + 130, + 477, + 482, + 574 + ], + "type": "inline_equation", + "content": "1.5\\mathrm{M}" + }, + { + "bbox": [ + 130, + 477, + 482, + 574 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 477, + 482, + 574 + ], + "type": "inline_equation", + "content": "5.8\\mathrm{M}" + }, + { + "bbox": [ + 130, + 477, + 482, + 574 + ], + "type": "text", + "content": ", which occupy only a small fraction of the total parameter count, about " + }, + { + "bbox": [ + 130, + 477, + 482, + 574 + ], + "type": "inline_equation", + "content": "1.7\\%" + }, + { + "bbox": [ + 130, + 477, + 482, + 574 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 477, + 482, + 574 + ], + "type": "inline_equation", + "content": "6.9\\%" + }, + { + "bbox": [ + 130, + 477, + 482, + 574 + ], + "type": "text", + "content": ". And they can contribute to BD-rate saving by " + }, + { + "bbox": [ + 130, + 477, + 482, + 574 + ], + "type": "inline_equation", + "content": "2.93\\%" + }, + { + "bbox": [ + 130, + 477, + 482, + 574 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 477, + 482, + 574 + ], + "type": "inline_equation", + "content": "4.35\\%" + }, + { + "bbox": [ + 130, + 477, + 482, + 574 + ], + "type": "text", + "content": " on CLIC Professional Valid, respectively. It shows that the proposed RAT and SAL are lightweight and efficient. When we combine these two modules, SAL will provide stronger prototypes for RAT, leading to better results, i.e., a saving of " + }, + { + "bbox": [ + 130, + 477, + 482, + 574 + ], + "type": "inline_equation", + "content": "8.50\\%" + }, + { + "bbox": [ + 130, + 477, + 482, + 574 + ], + "type": "text", + "content": " BD-rate." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "content": "With our proposed modules, Ours-S achieves high performance with a significantly smaller parameter count. Compared to ChARM [34], Ours-S saves nearly " + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "inline_equation", + "content": "40\\%" + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "content": " parameters and has faster speed while achieving comparable performance. Ours-S is a CNN-based model without any self-attention module. And the entropy estimation is just the hyperprior-style [6], without the efficient but time-consuming auto-regressive architecture. It indicates that the proposed modules RAT and SAL are highly effective and lightweight." + } + ] + } + ], + "index": 26 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "text", + "content": "Y. Liu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 114, + 250, + 195 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 250, + 195 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 250, + 195 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 250, + 195 + ], + "type": "image", + "image_path": "71dfb814646e3399dad66a4939b48e0bdf3255c1113c3f9db5f5a7e82a7f7870.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 180, + 196, + 206, + 203 + ], + "lines": [ + { + "bbox": [ + 180, + 196, + 206, + 203 + ], + "spans": [ + { + "bbox": [ + 180, + 196, + 206, + 203 + ], + "type": "text", + "content": "(a) Kodak" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 252, + 115, + 365, + 196 + ], + "blocks": [ + { + "bbox": [ + 252, + 115, + 365, + 196 + ], + "lines": [ + { + "bbox": [ + 252, + 115, + 365, + 196 + ], + "spans": [ + { + "bbox": [ + 252, + 115, + 365, + 196 + ], + "type": "image", + "image_path": "b44d25f77fba26cfab222d0247693025870a3fe5e2c0e9dde208324f9f6ce3f6.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 290, + 196, + 337, + 204 + ], + "lines": [ + { + "bbox": [ + 290, + 196, + 337, + 204 + ], + "spans": [ + { + "bbox": [ + 290, + 196, + 337, + 204 + ], + "type": "text", + "content": "(b) COCO-sample" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 131, + 213, + 482, + 269 + ], + "lines": [ + { + "bbox": [ + 131, + 213, + 482, + 269 + ], + "spans": [ + { + "bbox": [ + 131, + 213, + 482, + 269 + ], + "type": "text", + "content": "Fig. 8: We measure the impact of the masks separately in both the training and testing phases. \"test w/ masks\" means we adopt the masks in both training and testing, \"test w/o masks\" means we adopt the masks only in training, while \"train w/o masks\" means we do not use masks all the time. Naturally, if the masks are not used, " + }, + { + "bbox": [ + 131, + 213, + 482, + 269 + ], + "type": "inline_equation", + "content": "4 \\times 4" + }, + { + "bbox": [ + 131, + 213, + 482, + 269 + ], + "type": "text", + "content": " grid partitions will be used as alternatives." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 367, + 115, + 478, + 196 + ], + "blocks": [ + { + "bbox": [ + 367, + 115, + 478, + 196 + ], + "lines": [ + { + "bbox": [ + 367, + 115, + 478, + 196 + ], + "spans": [ + { + "bbox": [ + 367, + 115, + 478, + 196 + ], + "type": "image", + "image_path": "ecafbcf43bd572a8595e8556218b96639595e10c0dae1ebbb538e7dbb1685a46.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 392, + 196, + 466, + 204 + ], + "lines": [ + { + "bbox": [ + 392, + 196, + 466, + 204 + ], + "spans": [ + { + "bbox": [ + 392, + 196, + 466, + 204 + ], + "type": "text", + "content": "(c) Validation loss in training" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 293, + 254, + 304 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 293, + 254, + 304 + ], + "spans": [ + { + "bbox": [ + 132, + 293, + 254, + 304 + ], + "type": "text", + "content": "4.4 Subjective Results" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 314, + 482, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 314, + 482, + 422 + ], + "spans": [ + { + "bbox": [ + 130, + 314, + 482, + 422 + ], + "type": "text", + "content": "Fig. 7 shows the subjective results compared with VTM and Zou2022. As we can see, with the help of the segmentation masks in training, our SegPIC can distinguish the objects' contours more accurately, making the edges sharper with less bitrate. In " + }, + { + "bbox": [ + 130, + 314, + 482, + 422 + ], + "type": "inline_equation", + "content": "kodim04" + }, + { + "bbox": [ + 130, + 314, + 482, + 422 + ], + "type": "text", + "content": ", we can see the wool reconstructed by SegPIC is more consistent. In " + }, + { + "bbox": [ + 130, + 314, + 482, + 422 + ], + "type": "inline_equation", + "content": "kodim24" + }, + { + "bbox": [ + 130, + 314, + 482, + 422 + ], + "type": "text", + "content": ", the bird patterns decoded by SegPIC are more clear, and the blurrings are less than in other methods. These results are based on grid partitions used as alternatives to masks. It demonstrates that SegPIC has acquired a more robust representative capability for various objects. Consequently, SegPIC can discern objects in complex textures more effectively." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 441, + 383, + 453 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 441, + 383, + 453 + ], + "spans": [ + { + "bbox": [ + 132, + 441, + 383, + 453 + ], + "type": "text", + "content": "4.5 Analysis of the Role of Segmentation Masks" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 461, + 482, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 461, + 482, + 521 + ], + "spans": [ + { + "bbox": [ + 130, + 461, + 482, + 521 + ], + "type": "text", + "content": "Quantitative Results. To validate the role of the masks, we make experiments on two datasets, Kodak and COCO-sample. Due to Kodak not being a segmentation dataset, the corresponding masks are generated using the segmentation method [36]. Otherwise, we randomly sample 20 images and their masks from COCO Valid 2017 [9] to create the COCO-sample dataset." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 521, + 482, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 482, + 641 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 482, + 641 + ], + "type": "text", + "content": "As illustrated in Fig. 8 (a) (b), during the testing (inference) phase, the differences caused by masks are not so evident. However, for training, the situation is entirely different. With the assistance of the masks, the model achieves significantly better performance, especially at low bitrates. It could imply that, due to bitrate loss constraints during training, the information in the prototypes is implicitly transferred to the learnable parameters in RAT, allowing limited contextual information to still be effective during the inference phase. At low bitrates, since the lack of sufficient high-frequency information, pixels in the region exhibit stronger similarity, enhancing the representational ability of features extracted by prototypes. Fig. 8 (c) shows the early stage of model training." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 131, + 641, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 641, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 131, + 641, + 481, + 665 + ], + "type": "text", + "content": "As shown in Tab. 3, the model training w/o masks (the last column) leads to noticeable degradation. However, training w/ masks has a relatively minor" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 135, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 135, + 91, + 447, + 102 + ], + "type": "text", + "content": "Region-Adaptive Transform with Segmentation Prior for Image Compression" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 134, + 114, + 482, + 185 + ], + "blocks": [ + { + "bbox": [ + 134, + 114, + 482, + 185 + ], + "lines": [ + { + "bbox": [ + 134, + 114, + 482, + 185 + ], + "spans": [ + { + "bbox": [ + 134, + 114, + 482, + 185 + ], + "type": "image", + "image_path": "77ae98955415a15f9706ec26f4656280dbe084927ad793772e397990b7abe755.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 192, + 482, + 258 + ], + "lines": [ + { + "bbox": [ + 130, + 192, + 482, + 258 + ], + "spans": [ + { + "bbox": [ + 130, + 192, + 482, + 258 + ], + "type": "text", + "content": "Fig. 9: Visualization of the kernels generated by DKG in RAT at the Encoder (see Fig. 3), in the cases with and without the masks. Fig. (d)(e)(f) are generated by the same pre-trained network weights. During testing, they go through processes with or without masks, respectively. The distortions in (d) can be formulated as " + }, + { + "bbox": [ + 130, + 192, + 482, + 258 + ], + "type": "inline_equation", + "content": "(x - \\hat{x})^2" + }, + { + "bbox": [ + 130, + 192, + 482, + 258 + ], + "type": "text", + "content": ". We take the maximum absolute values of the tensors along the channels and apply 2D max pooling on them for better visualization." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 176, + 304, + 438, + 352 + ], + "blocks": [ + { + "bbox": [ + 130, + 271, + 479, + 293 + ], + "lines": [ + { + "bbox": [ + 130, + 271, + 479, + 293 + ], + "spans": [ + { + "bbox": [ + 130, + 271, + 479, + 293 + ], + "type": "text", + "content": "Table 3: The effect of the masks and the grid number on BD-rate on COCO-sample. \"n×\" denotes n×n grid partitions to replace masks." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 176, + 304, + 438, + 352 + ], + "lines": [ + { + "bbox": [ + 176, + 304, + 438, + 352 + ], + "spans": [ + { + "bbox": [ + 176, + 304, + 438, + 352 + ], + "type": "table", + "html": "
Traintrain w/ masks4×
Testtest w/ masks2×4×8×4×
BD-rate ↓-0.11+0.210(default)+0.46+2.10
", + "image_path": "9abe95cb5dde4a2107205fb3695116451bb574925e6d73184c6a2817128ac72b.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 375, + 482, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 375, + 482, + 434 + ], + "spans": [ + { + "bbox": [ + 130, + 375, + 482, + 434 + ], + "type": "text", + "content": "impact on the pattern and number in testing. So we can replace the masks in testing to further save bitrate but should take the masks as privileged information during training. The hyperparameter " + }, + { + "bbox": [ + 130, + 375, + 482, + 434 + ], + "type": "inline_equation", + "content": "4 \\times" + }, + { + "bbox": [ + 130, + 375, + 482, + 434 + ], + "type": "text", + "content": " in testing is an empirical trade-off between the effect of RAT (i.e. image reconstruction precision) and the quantity of the prototypes (i.e. the additional bitrate they introduced)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 436, + 482, + 520 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 436, + 482, + 520 + ], + "spans": [ + { + "bbox": [ + 130, + 436, + 482, + 520 + ], + "type": "text", + "content": "Qualitative Analysis. Fig. 9 shows the visualization of the kernels generated by DKG in RAT at the Encoder (see Fig. 2). Although there is not a noticeable difference in Fig. 9 (e), it is obvious that the gradients of " + }, + { + "bbox": [ + 130, + 436, + 482, + 520 + ], + "type": "inline_equation", + "content": "w / \\text{mask}" + }, + { + "bbox": [ + 130, + 436, + 482, + 520 + ], + "type": "text", + "content": " are more focused on the semantic objects (see Fig. 9 (f)), which is complex and prone to causing distortions. This attribute will assist in optimizing the models, enabling it to better reconstruct various objects. These observations indicate that the masks do play a necessary role during the training phase." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 541, + 220, + 554 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 541, + 220, + 554 + ], + "spans": [ + { + "bbox": [ + 132, + 541, + 220, + 554 + ], + "type": "text", + "content": "5 Conclusion" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": "We propose a novel image compression framework called Segmentation-Prior-Guided Image Compression (SegPIC), equipped with two proposed modules: Region-Adaptive Transform (RAT) and Scale Affine Layer (SAL). We adopt the class-agnostic masks as privilege information to assist the training of SegPIC. Experimental results demonstrate that our SegPIC is superior to the previous works. Otherwise, we analyze the role of the masks. Experiments and visualization results show that the masks can guide the model to focus more on semantic objects in training, leading to higher pixel-fidelity performance." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "type": "text", + "content": "Y. Liu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 141, + 482, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 141, + 482, + 190 + ], + "spans": [ + { + "bbox": [ + 130, + 141, + 482, + 190 + ], + "type": "text", + "content": "This work was supported in part by the National Key R&D Program of China (No.2021ZD0112100), National NSF of China (No.62120106009), Guangdong Basic and Applied Basic Research Foundation (2024A1515010454), the Basic and Frontier Research Project of PCL, and the Major Key Project of PCL." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 209, + 198, + 220 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 209, + 198, + 220 + ], + "spans": [ + { + "bbox": [ + 132, + 209, + 198, + 220 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 134, + 234, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 138, + 234, + 479, + 257 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 234, + 479, + 257 + ], + "spans": [ + { + "bbox": [ + 138, + 234, + 479, + 257 + ], + "type": "text", + "content": "1. Workshop and challenge on learned image compression (clic2020) (2020), http://www.compression.cc 10, 11" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 258, + 480, + 290 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 258, + 480, + 290 + ], + "spans": [ + { + "bbox": [ + 138, + 258, + 480, + 290 + ], + "type": "text", + "content": "2. Agustsson, E., Minnen, D., Toderici, G., Mentzer, F.: Multi-realism image compression with a conditional generator. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 22324-22333 (2023) 10" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 291, + 480, + 312 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 291, + 480, + 312 + ], + "spans": [ + { + "bbox": [ + 138, + 291, + 480, + 312 + ], + "type": "text", + "content": "3. Ahmed, N., Natarajan, T., Rao, K.R.: Discrete cosine transform. IEEE Transactions on Computers 100(1), 90-93 (1974) 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 312, + 480, + 334 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 312, + 480, + 334 + ], + "spans": [ + { + "bbox": [ + 138, + 312, + 480, + 334 + ], + "type": "text", + "content": "4. Akbari, M., Liang, J., Han, J.: Dsslic: Deep semantic segmentation-based layered image compression. In: ICASSP. pp. 2042-2046. IEEE (2019) 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 335, + 480, + 356 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 335, + 480, + 356 + ], + "spans": [ + { + "bbox": [ + 138, + 335, + 480, + 356 + ], + "type": "text", + "content": "5. Balle, J., Laparra, V., Simoncelli, E.P.: End-to-end optimized image compression. In: Int. Conf. Learn. Represent. (2017) 2, 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 357, + 480, + 388 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 357, + 480, + 388 + ], + "spans": [ + { + "bbox": [ + 138, + 357, + 480, + 388 + ], + "type": "text", + "content": "6. Balle, J., Minnen, D., Singh, S., Hwang, S.J., Johnston, N.: Variational image compression with a scale hyperprior. In: Int. Conf. Learn. Represent. (2018) 2, 4, 5, 12" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 389, + 459, + 401 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 389, + 459, + 401 + ], + "spans": [ + { + "bbox": [ + 138, + 389, + 459, + 401 + ], + "type": "text", + "content": "7. Bellard, F.: Bpg image format (2014), https://bellard.org/bpg/ 2, 10, 11" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 402, + 480, + 422 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 402, + 480, + 422 + ], + "spans": [ + { + "bbox": [ + 138, + 402, + 480, + 422 + ], + "type": "text", + "content": "8. Bjontegaard, G.: Calculation of average psnr differences between rd-curves. ITU SG16 Doc. VCEG-M33 (2001) 11" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 423, + 480, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 423, + 480, + 445 + ], + "spans": [ + { + "bbox": [ + 138, + 423, + 480, + 445 + ], + "type": "text", + "content": "9. Caesar, H., Uijlings, J., Ferrari, V.: COCO-Stuff: Thing and stuff classes in context. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 1209-1218 (2018) 10, 13" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 445, + 480, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 445, + 480, + 477 + ], + "spans": [ + { + "bbox": [ + 138, + 445, + 480, + 477 + ], + "type": "text", + "content": "0. Chang, J., Mao, Q., Zhao, Z., Wang, S., Wang, S., Zhu, H., Ma, S.: Layered conceptual image compression via deep semantic synthesis. In: IEEE Int. Conf. Image Process. pp. 694-698 (2019) 4, 5" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 134, + 478, + 480, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 478, + 480, + 510 + ], + "spans": [ + { + "bbox": [ + 134, + 478, + 480, + 510 + ], + "type": "text", + "content": "11. Chang, J., Zhao, Z., Yang, L., Jia, C., Zhang, J., Ma, S.: Thousand to one: Semantic prior modeling for conceptual coding. In: Int. Conf. Multimedia and Expo. pp. 1-6. IEEE (2021) 3, 4, 5" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 134, + 510, + 480, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 510, + 480, + 544 + ], + "spans": [ + { + "bbox": [ + 134, + 510, + 480, + 544 + ], + "type": "text", + "content": "12. Cheng, Z., Sun, H., Takeuchi, M., Katto, J.: Learned image compression with discretized gaussian mixture likelihoods and attention modules. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 7939-7948 (2020) 3, 4, 10, 11" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 134, + 544, + 480, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 544, + 480, + 566 + ], + "spans": [ + { + "bbox": [ + 134, + 544, + 480, + 566 + ], + "type": "text", + "content": "13. Chollet, F.: Xception: Deep learning with depthwise separable convolutions. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 1251-1258 (2017) 8" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 134, + 567, + 480, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 567, + 480, + 588 + ], + "spans": [ + { + "bbox": [ + 134, + 567, + 480, + 588 + ], + "type": "text", + "content": "14. Daubechies, I.: The wavelet transform, time-frequency localization and signal analysis. IEEE Transactions on Information Theory 36(5), 961-1005 (1990) 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 134, + 589, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 589, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 134, + 589, + 480, + 632 + ], + "type": "text", + "content": "15. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., Uszkoreit, J., Houlsby, N.: An image is worth 16x16 words: Transformers for image recognition at scale. In: Int. Conf. Learn. Represent. (2021) 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 134, + 632, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 632, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 134, + 632, + 480, + 665 + ], + "type": "text", + "content": "16. Feng, R., Gao, Y., Jin, X., Feng, R., Chen, Z.: Semantically structured image compression via irregular group-based decoupling. In: Int. Conf. Comput. Vis. pp. 17237-17247 (2023) 3, 5, 10" + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 135, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 135, + 91, + 447, + 102 + ], + "type": "text", + "content": "Region-Adaptive Transform with Segmentation Prior for Image Compression" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 481, + 160 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 481, + 160 + ], + "type": "text", + "content": "17. Goodfellow, I., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., Courville, A., Bengio, Y.: Generative adversarial nets. In: Ghahramani, Z., Welling, M., Cortes, C., Lawrence, N., Weinberger, K. (eds.) Adv. Neural Inform. Process. Syst. vol. 27. Curran Associates, Inc. (2014) 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 161, + 481, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 161, + 481, + 194 + ], + "spans": [ + { + "bbox": [ + 132, + 161, + 481, + 194 + ], + "type": "text", + "content": "18. He, D., Yang, Z., Peng, W., Ma, R., Qin, H., Wang, Y.: Elic: Efficient learned image compression with unevenly grouped space-channel contextual adaptive coding. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 5718-5727 (2022) 10" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 194, + 481, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 194, + 481, + 226 + ], + "spans": [ + { + "bbox": [ + 132, + 194, + 481, + 226 + ], + "type": "text", + "content": "19. He, D., Zheng, Y., Sun, B., Wang, Y., Qin, H.: Checkerboard context model for efficient learned image compression. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 14771-14780 (2021) 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 227, + 481, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 227, + 481, + 248 + ], + "spans": [ + { + "bbox": [ + 132, + 227, + 481, + 248 + ], + "type": "text", + "content": "20. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 770-778 (2016) 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 249, + 481, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 249, + 481, + 281 + ], + "spans": [ + { + "bbox": [ + 132, + 249, + 481, + 281 + ], + "type": "text", + "content": "21. Hoang, T.M., Zhou, J., Fan, Y.: Image compression with encoder-decoder matched semantic segmentation. In: IEEE Conf. Comput. Vis. Pattern Recog. Worksh. pp. 160-161 (2020) 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 281, + 481, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 281, + 481, + 304 + ], + "spans": [ + { + "bbox": [ + 132, + 281, + 481, + 304 + ], + "type": "text", + "content": "22. Jia, X., De Brabandere, B., Tuytelaars, T., Gool, L.V.: Dynamic filter networks. Adv. Neural Inform. Process. Syst. 29 (2016) 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 304, + 481, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 304, + 481, + 335 + ], + "spans": [ + { + "bbox": [ + 132, + 304, + 481, + 335 + ], + "type": "text", + "content": "23. Jiang, W., Yang, J., Zhai, Y., Ning, P., Gao, F., Wang, R.: Mlic: Multi-reference entropy model for learned image compression. In: ACM Int. Conf. Multimedia. pp. 7618-7627 (2023) 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 336, + 481, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 336, + 481, + 357 + ], + "spans": [ + { + "bbox": [ + 132, + 336, + 481, + 357 + ], + "type": "text", + "content": "24. (JVET), J.V.E.T.: Versatile video coding (2021), https://jvet.hhi.fraunhofer.de/ 2, 11" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 358, + 481, + 379 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 358, + 481, + 379 + ], + "spans": [ + { + "bbox": [ + 132, + 358, + 481, + 379 + ], + "type": "text", + "content": "25. Kingma, D.P., Ba, J.: Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980 (2014) 10" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 380, + 481, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 380, + 481, + 402 + ], + "spans": [ + { + "bbox": [ + 132, + 380, + 481, + 402 + ], + "type": "text", + "content": "26. Kingma, D.P., Welling, M.: Auto-encoding variational bayes. In: Int. Conf. Learn. Represent. (2014) 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 403, + 481, + 424 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 403, + 481, + 424 + ], + "spans": [ + { + "bbox": [ + 132, + 403, + 481, + 424 + ], + "type": "text", + "content": "27. Kodak, E.: Kodak lossless true color image suite (photocd pcd0992) (1993), http://r0k.us/graphics/kodak/ 10, 11" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 425, + 481, + 467 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 425, + 481, + 467 + ], + "spans": [ + { + "bbox": [ + 132, + 425, + 481, + 467 + ], + "type": "text", + "content": "28. Koyuncu, A.B., Gao, H., Boev, A., Gaikov, G., Alshina, E., Steinbach, E.: Contextformer: A transformer with spatio-channel attention for context modeling in learned image compression. In: Eur. Conf. Comput. Vis. pp. 447-463. Springer (2022) 10" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 468, + 481, + 500 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 468, + 481, + 500 + ], + "spans": [ + { + "bbox": [ + 132, + 468, + 481, + 500 + ], + "type": "text", + "content": "29. Li, F., Zhang, L., Liu, Z., Lei, J., Li, Z.: Multi-frequency representation enhancement with privilege information for video super-resolution. In: Int. Conf. Comput. Vis. pp. 12814-12825 (2023) 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 501, + 481, + 523 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 501, + 481, + 523 + ], + "spans": [ + { + "bbox": [ + 132, + 501, + 481, + 523 + ], + "type": "text", + "content": "30. Liu, J., Lu, G., Hu, Z., Xu, D.: A unified end-to-end framework for efficient deep image compression. arXiv preprint arXiv:2002.03370 (2020) 3, 4" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 523, + 481, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 523, + 481, + 555 + ], + "spans": [ + { + "bbox": [ + 132, + 523, + 481, + 555 + ], + "type": "text", + "content": "31. Liu, J., Sun, H., Katto, J.: Learned image compression with mixed transformer-cnn architectures. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 14388-14397 (2023) 4" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 555, + 481, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 555, + 481, + 588 + ], + "spans": [ + { + "bbox": [ + 132, + 555, + 481, + 588 + ], + "type": "text", + "content": "32. Liu, Z., Lin, Y., Cao, Y., Hu, H., Wei, Y., Zhang, Z., Lin, S., Guo, B.: Swin transformer: Hierarchical vision transformer using shifted windows. In: Int. Conf. Comput. Vis. pp. 10012-10022 (2021) 3, 4" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 589, + 481, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 589, + 481, + 610 + ], + "spans": [ + { + "bbox": [ + 132, + 589, + 481, + 610 + ], + "type": "text", + "content": "33. Minnen, D., Balle, J., Toderici, G.D.: Joint autoregressive and hierarchical priors for learned image compression. Adv. Neural Inform. Process. Syst. 31 (2018) 4" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 611, + 481, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 611, + 481, + 643 + ], + "spans": [ + { + "bbox": [ + 132, + 611, + 481, + 643 + ], + "type": "text", + "content": "34. Minnen, D., Singh, S.: Channel-wise autoregressive entropy models for learned image compression. In: IEEE Int. Conf. Image Process. pp. 3339-3343. IEEE (2020) 4, 5, 9, 10, 11, 12" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 132, + 643, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 643, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 643, + 481, + 665 + ], + "type": "text", + "content": "35. Pan, G., Lu, G., Hu, Z., Xu, D.: Content adaptive latents and decoder for neural image compression. In: Eur. Conf. Comput. Vis. pp. 556-573. Springer (2022) 10" + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "type": "text", + "content": "Y. Liu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 501 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 480, + 138 + ], + "type": "text", + "content": "36. Qi, L., Kuen, J., Wang, Y., Gu, J., Zhao, H., Torr, P., Lin, Z., Jia, J.: Open world entity segmentation. IEEE Trans. Pattern Anal. Mach. Intell. (2022) 13" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 140, + 480, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 140, + 480, + 160 + ], + "spans": [ + { + "bbox": [ + 132, + 140, + 480, + 160 + ], + "type": "text", + "content": "37. Shen, H., Zhao, Z.Q., Zhang, W.: Adaptive dynamic filtering network for image denoising. In: AAAI. vol. 37, pp. 2227-2235 (2023) 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 161, + 480, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 161, + 480, + 182 + ], + "spans": [ + { + "bbox": [ + 133, + 161, + 480, + 182 + ], + "type": "text", + "content": "38. Skodras, A., Christopoulos, C., Ebrahimi, T.: The JPEG 2000 still image compression standard. IEEE Signal Processing Magazine 18(5), 36-58 (2001) 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 183, + 480, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 183, + 480, + 205 + ], + "spans": [ + { + "bbox": [ + 133, + 183, + 480, + 205 + ], + "type": "text", + "content": "39. Stark, H., Woods, J.W.: Probability, random processes, and estimation theory for engineers. Prentice-Hall, Inc. (1986) 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 134, + 205, + 480, + 236 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 205, + 480, + 236 + ], + "spans": [ + { + "bbox": [ + 134, + 205, + 480, + 236 + ], + "type": "text", + "content": "40. Sun, S., He, T., Chen, Z.: Semantic structured image coding framework for multiple intelligent applications. IEEE Trans. Circuit Syst. Video Technol. 31(9), 3631-3642 (2020) 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 133, + 237, + 480, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 237, + 480, + 270 + ], + "spans": [ + { + "bbox": [ + 133, + 237, + 480, + 270 + ], + "type": "text", + "content": "41. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. Adv. Neural Inform. Process. Syst. 30 (2017) 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 133, + 271, + 481, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 271, + 481, + 293 + ], + "spans": [ + { + "bbox": [ + 133, + 271, + 481, + 293 + ], + "type": "text", + "content": "42. Wallace, G.K.: The JPEG still picture compression standard. Communications of the ACM 34(4), 30-44 (1991) 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 134, + 293, + 480, + 324 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 293, + 480, + 324 + ], + "spans": [ + { + "bbox": [ + 134, + 293, + 480, + 324 + ], + "type": "text", + "content": "43. Wang, D., Yang, W., Hu, Y., Liu, J.: Neural data-dependent transform for learned image compression. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 17379-17388 (2022) 2, 3, 4, 10, 11" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 133, + 325, + 480, + 347 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 325, + 480, + 347 + ], + "spans": [ + { + "bbox": [ + 133, + 325, + 480, + 347 + ], + "type": "text", + "content": "44. Wang, G.H., Li, J., Li, B., Lu, Y.: Evc: Towards real-time neural image compression with mask decay. In: Int. Conf. Learn. Represent. (2023) 10" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 134, + 348, + 480, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 348, + 480, + 380 + ], + "spans": [ + { + "bbox": [ + 134, + 348, + 480, + 380 + ], + "type": "text", + "content": "45. Wang, X., Yu, K., Dong, C., Loy, C.C.: Recovering realistic texture in image superresolution by deep spatial feature transform. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 606-615 (2018) 8" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 134, + 380, + 480, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 380, + 480, + 413 + ], + "spans": [ + { + "bbox": [ + 134, + 380, + 480, + 413 + ], + "type": "text", + "content": "46. Wang, Z., Simoncelli, E.P., Bovik, A.C.: Multiscale structural similarity for image quality assessment. In: The Thirty-Seventh Asilomar Conference on Signals, Systems & Computers, 2003. vol. 2, pp. 1398-1402. IEEE (2003) 4, 10" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 134, + 414, + 480, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 414, + 480, + 446 + ], + "spans": [ + { + "bbox": [ + 134, + 414, + 480, + 446 + ], + "type": "text", + "content": "47. Xu, Y.S., Tseng, S.Y.R., Tseng, Y., Kuo, H.K., Tsai, Y.M.: Unified dynamic convolutional network for super-resolution with variational degradations. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 12496-12505 (2020) 6" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 134, + 447, + 480, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 447, + 480, + 468 + ], + "spans": [ + { + "bbox": [ + 134, + 447, + 480, + 468 + ], + "type": "text", + "content": "48. Zhu, Y., Yang, Y., Cohen, T.: Transformer-based transform coding. In: Int. Conf. Learn. Represent. (2021) 3, 4" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 134, + 468, + 480, + 501 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 468, + 480, + 501 + ], + "spans": [ + { + "bbox": [ + 134, + 468, + 480, + 501 + ], + "type": "text", + "content": "49. Zou, R., Song, C., Zhang, Z.: The devil is in the details: Window-based attention for image compression. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 17492-17501 (2022) 3, 4, 5, 7, 9, 10, 11, 12" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 135, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 135, + 91, + 447, + 102 + ], + "type": "text", + "content": "Region-Adaptive Transform with Segmentation Prior for Image Compression" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Region-Aware Sequence-to-Sequence Learning for Hyperspectral Denoising/fb5f87c2-230f-4062-9b75-422a250a1410_content_list.json b/2024/Region-Aware Sequence-to-Sequence Learning for Hyperspectral Denoising/fb5f87c2-230f-4062-9b75-422a250a1410_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..adf4e04f913fdbbda5b0f2b2df83e11e8d708270 --- /dev/null +++ b/2024/Region-Aware Sequence-to-Sequence Learning for Hyperspectral Denoising/fb5f87c2-230f-4062-9b75-422a250a1410_content_list.json @@ -0,0 +1,1928 @@ +[ + { + "type": "text", + "text": "Region-Aware Sequence-to-Sequence Learning for Hyperspectral Denoising", + "text_level": 1, + "bbox": [ + 218, + 140, + 782, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jiahua Xiao, Yang Liu, and Xing Wei*", + "bbox": [ + 343, + 212, + 658, + 227 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "School of Software Engineering, Xi'an Jiaotong University {xjh847286495,1228754216}@stu.xjtu.edu.cn, weixing@mail.xjtu.edu.cn", + "bbox": [ + 240, + 239, + 759, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Proper spectral modeling within hyperspectral image (HSI) is critical yet highly challenging for HSI denoising. In contrast to existing methods that struggle between effectiveness and flexibility in modeling long-range spectral dependencies and directly explore spatial-spectral information without region discrimination, we introduce RAS2S—a simple yet effective sequence-to-sequence (Seq2Seq) learning framework for better HSI denoising. RAS2S treats HSI denoising as a Seq2Seq translation problem, which converts the noisy spectral sequence to its clean ones in an autoregressive fashion. In addition, spatial-spectral information exploration without region discrimination contradicts the intrinsic spatial-spectral diversity of HSIs, leading to negative interference from spatial-spectral unrelated regions. Thus we propose a novel spatial-spectral region-aware module to distinctively perceive the semantic regions with different spatial-spectral representations, maximizing the spectral modeling potential of Seq2Seq learning. With such an improved Seq2Seq learning paradigm, RAS2S not only shows huge potential in capturing long-range spectral dependencies, but also maintains the flexibility to handle diverse HSIs. Extensive experiments demonstrate that RAS2S outperforms existing state-of-the-art methods quantitatively and qualitatively with a minimal model size, merely 0.08M. Source code is available at https://github.com/MIV-XJTU/RAS2S.", + "bbox": [ + 261, + 299, + 743, + 592 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Hyperspectral Denoising $\\cdot$ Spectral Dependency $\\cdot$ Sequence-to-Sequence Learning $\\cdot$ Region Discrimination", + "bbox": [ + 261, + 603, + 740, + 632 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 654, + 375, + 671 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Unlike conventional images (e.g., grayscale or RGB), hyperspectral images (HSIs) capture the spectral attributes of various objects and materials within a static spatial scenario, widely applied in medical diagnosis [31], material identification [26, 48] and detection [59, 60]. However, various noise types introduced in the imaging process severely disrupt both the visual perception of HSIs and their aforementioned applications, making HSI denoising a crucial yet challenging task. Thanks to the unique spectral information inherited in HSIs [56], which reflects the inner properties of objects, combining spatial and spectral learning allows for more effective noise identification and HSI restoration.", + "bbox": [ + 212, + 683, + 787, + 820 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Corresponding author: Xing Wei.", + "bbox": [ + 218, + 825, + 455, + 840 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In the deep learning era, mainstream CNN-based methods [46, 47, 51, 52] jointly extract spatial-spectral information using 2D convolutional filters over a short window of adjacent spectral bands. However, the limited spectral receptive field severely constrains the utilization of global spectral information. While 3D convolution [13, 41, 43] provides a broader spatial-spectral perspective, it introduces significantly increased complexity. More recently, transformer-based", + "bbox": [ + 212, + 146, + 472, + 357 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/c9cd5ce64a57d1b1bcdb8d00aa2e31ce0e0c191d40ebafc78053cf0060559bb8.jpg", + "image_caption": [ + "Fig. 1: Denoising performance vs. inference time (CPU) under CAVE [34]. The circle areas are proportional to the model size." + ], + "image_footnote": [], + "bbox": [ + 485, + 152, + 781, + 290 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "methods [5,27,28] leverage self-attention mechanisms to capture complex spatial-spectral relations and effectively remove noise within HSIs. However, they are unable to handle HSIs with arbitrary spectral numbers and need dataset-specific training. As a result, there seems to be a dilemma between the sufficiency and flexibility of exploring long-range spectral dependencies.", + "bbox": [ + 212, + 358, + 787, + 434 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To solve this dilemma, we consider continuous spectral bands as a sequence and introduce a novel perspective of treating HSI denoising as a sequence translation problem. We attempt to leverage the proven capability of sequence-to-sequence (Seq2Seq) models in sequence modeling [2,21,37] to tackle this problem, which effectively capture long-range dependencies with fewer parameters (shown in Fig. 1) and spectral flexibility. However, directly applying existing Seq2Seq frameworks to spectral sequences poses challenges: (i) Existing Seq2Seq models generate sequences in a forward order (e.g., text and video sequences), hindering their application to spectral sequences, which can utilize both preceding and succeeding bands. To address this, we employ a bidirectional encoder-decoder architecture tailored for spatial-spectral modeling, where the encoder processes noisy bands to hidden states, and the decoder utilizes these states to dynamically generate denoised spectral sequences autoregressively, capturing long-range contextual dependencies bidirectionally. (ii) Existing Seq2Seq models ignore the information gap between adjacent hidden states and the current input, hindering effective hidden state aggregation. We mitigate this by introducing a hidden state compensation module that adjusts adjacent hidden states with cross-attention, alleviating the adverse effects of the information gap.", + "bbox": [ + 212, + 440, + 787, + 712 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In addition, HSIs inherently contain diverse semantic variations across spatial and spectral domains. However, current mainstream convolutional methods perform information extraction using filter sharing across the spatial-spectral domain, treating the feature of each position without discrimination, which will introduces negative interference from semantically unrelated regions during parameter updates. To address this, leveraging insights from spatial non-local self-similarity [32, 58] and spectral distinctiveness (see Fig. 2), we propose discriminative semantic region learning implemented through a Spatial Spectral", + "bbox": [ + 212, + 719, + 787, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "J. Xiao et al.", + "bbox": [ + 271, + 114, + 359, + 127 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Region-Aware (SSRA) module for more focused exploration of spatial-spectral characteristics. SSRA dynamically allocates learnable filters to corresponding semantic regions by utilizing encoded hidden states rich in spatial-spectral aggregation information to generate region mask predictions in both spatial and spectral branches.", + "bbox": [ + 212, + 146, + 787, + 222 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The region-guided masks divide the input into distinct regions, with each region assigned a dedicated filter, eliminating irrelevant spatial-spectral interference and maximizing spectral modeling potential of Seq2Seq learning. Thus, the Region-Aware Sequence-to-Sequence (RAS2S) learning framework is proposed for HSI denoising task. Benefiting from this tailored Seq2Seq architecture, our approach effectively captures long-range spectral dependencies for arbitrary spectral numbers, and eliminates the need for dataset-specific training. Overall, this work makes the following contributions:", + "bbox": [ + 212, + 223, + 447, + 537 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/beed88671b24c2a3fa22a12c742fb300be28e85b03ceaddee9a7147fcf21b36e.jpg", + "image_caption": [ + "Fig.2: Illustration of spectral distinctiveness. Spatial positions with similar materials have similar spectral reflection patterns (e.g., Position A and B), while those with different materials show divergent patterns (e.g., Position C and D). This motivates assigning learnable filters to corresponding regions with similar spatial-spectral semantic representations, enabling each region to more effectively focus on its own vital spatial-spectral characteristics. We term this discriminative semantic region learning." + ], + "image_footnote": [], + "bbox": [ + 459, + 227, + 784, + 373 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We introduce RAS2S, the first framework to reformulate the HSI denoising problem from a sequence-to-sequence learning perspective, tailored for spectral sequence modeling.", + "- We first propose the discriminative semantic region learning within spatial-spectral domains and implement it through a novel SSRA module to maximize the denoising potential of the Seq2Seq framework.", + "- Extensive experiments on prevailing benchmarks demonstrate the superiority and generality of RAS2S, achieving state-of-the-art performance with few parameters." + ], + "bbox": [ + 225, + 554, + 784, + 691 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 717, + 387, + 732 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.1 Hyperspectral Image Denoising", + "text_level": 1, + "bbox": [ + 215, + 751, + 522, + 767 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Traditional methods rely on handcrafted priors based on hyperspectral image (HSI) statistics and often struggle when faced with noise distributions inconsistent with prior assumptions. For instance, approaches like [8, 17, 40] capture global spatial-spectral correlation, effectively leveraging low-rank properties and", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "Sequence-to-Sequence Hyperspectral Denoising", + "bbox": [ + 416, + 114, + 732, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "inherent smoothness in HSIs. Others, such as [7,35], utilize total variation priors to preserve marginal features effectively. In the era of deep learning, 2D CNN-based methods like [47,51,52] jointly extract spatial-spectral information through a short window of adjacent spectral bands, constrained by their spectral receptive field. Meanwhile, 3D CNN-based methods like [36,43,44] capture spatial-spectral dependencies but with increased computational cost. RNN-based approaches like [15] explore sequential spectrum analysis but struggle with long-range spatial-spectral dependencies. Recently, transformer-based models such as [9,27,28], leveraging self-attention mechanisms, show promising denoising performance. However, the extensive interaction among global tokens and the need for fixed-length spectral bands pose computational and flexibility challenges for modeling long-range spectral dependencies. In contrast, our model, a simple serial encoder-decoder architecture, effectively captures long-range spectral dependencies and handles HSIs with varying spectral numbers at a minimal model size.", + "bbox": [ + 212, + 146, + 787, + 372 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.2 Sequence-to-Sequence Learning", + "text_level": 1, + "bbox": [ + 214, + 395, + 522, + 411 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Sequence-to-Sequence (Seq2Seq) learning [37] is a pivotal framework in natural language processing (NLP) and has shown significant success in various tasks like machine translation [39] and text summarization [50]. These models use an encoder to process input sequences and a decoder to generate output sequences. They evolve with the introduction of attention mechanisms, allowing them to focus on specific parts of the input when generating output. Recently, there are a growing trend in leveraging Seq2Seq models for various computer vision tasks, such as video restoration [29] and pose estimation [54]. Unlike natural language paragraphs and video frame sequences, which follow a temporal order, hyperspectral bands can utilize both preceding and succeeding bands. Tailored modifications are necessary to achieve effective spatial-spectral modeling and maximize the potential of Seq2Seq models for HSI denoising.", + "bbox": [ + 212, + 420, + 787, + 602 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3 Method", + "text_level": 1, + "bbox": [ + 214, + 625, + 330, + 640 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In the subsequent sections, we represent the pairs of noisy and clean HSI as $Y = \\{y_{i}\\}_{1}^{B}$ and $X = \\{x_{i}\\}_{1}^{B}$ , where $y_{i}$ and $x_{i}$ denote individual spectral bands, and $Y, X \\in \\mathbb{R}^{H \\times W \\times B}$ . Here, $H$ and $W$ represent the spatial height and width, respectively, while $B$ indicates the number of spectral bands in the HSI. Therefore, the task of HSI denoising is formulated to maximize the conditional probability $P(X|Y)$ .", + "bbox": [ + 212, + 656, + 787, + 748 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1 HSI Denoising as Sequence Translation", + "text_level": 1, + "bbox": [ + 214, + 768, + 586, + 784 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We treat HSI denoising as a sequence translation problem and introduce RAS2S. However, existing Seq2Seq models cannot be directly transferred to HSI processing tasks. To effectively models the long-range spectral dependencies, our model", + "bbox": [ + 212, + 794, + 785, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "J. Xiao et al.", + "bbox": [ + 271, + 114, + 359, + 127 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/7a2dbca4de80fa93412b54dcbf40c4194c754ef5f4c01780dc403d8582bc0cf8.jpg", + "image_caption": [ + "Fig. 3: RAS2S architecture. The workflow of the proposed RAS2S is depicted in (a), which models long-range spectral dependencies with a tailored bidirectional encoder-decoder architecture. The modules at the bottom display the internal details of (b) hidden state compensation (HFComp) and (c) adjacent attention mechanism (AAM)." + ], + "image_footnote": [], + "bbox": [ + 228, + 143, + 769, + 378 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "uses the idea of bidirectional RNN to capture dependencies in both preceding bands $\\{y_{i}\\}_{1}^{i - 1}$ and succeeding bands $\\{y_{i}\\}_{i + 1}^{B}$ , and then outputs the bidirectional propagated hidden state sets $\\left(\\left\\{h_i^{fw}\\right\\}_1^B\\right.$ and $\\left\\{h_i^{bw}\\right\\}_1^B$ . For brevity and clarity, we take the forward propagation process as an example to elaborate the workflow of RAS2S in the following section, which can be easily extended to a bidirectional process. As presented in Fig. 3(a), RAS2S consists of the following key components.", + "bbox": [ + 212, + 476, + 787, + 594 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Encoder. In forward propagation process, the goal of encoder is to transform the shallow feature $\\{f_i\\}_1^B$ of each band into the hidden state set $\\left\\{h_i^{fw - e}\\right\\}_1^B \\in \\mathbb{R}^{H \\times W \\times C}$ . In order to alleviate the information gap between preceding hidden features $h_{i-1}^{fw - e}$ and current visual features $f_i$ , we propose to make the Hidden State Compensation (HFComp) for $h_{i-1}^{fw - e}$ to obtain an adjusted hidden state $h_{i-1}^{fw - ec}$ . The HFComp will be introduced immediately. In implementation, we utilize the ConvGRU cell [3] to facilitate the learning of long-range spatial-spectral information, which combines $f_i$ and $h_{i-1}^{fw - ec}$ to generate the final output $h_i^{fw - e}$ . Thus, the workflow of encoder for each spectral step can be formulated as:", + "bbox": [ + 212, + 607, + 787, + 782 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nh _ {i - 1} ^ {f w} ^ {- e c} = \\mathrm {H F C o m p} (f _ {i}, h _ {i - 1} ^ {f w} ^ {- e}), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 387, + 782, + 785, + 804 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nh _ {i} ^ {f w - e} = \\mathrm {C o n v G R U} (f _ {i}, h _ {i - 1} ^ {f w - e c}), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 383, + 820, + 785, + 842 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "Sequence-to-Sequence Hyperspectral Denoising", + "bbox": [ + 416, + 114, + 732, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where the backward propagation process also produces the hidden state $h_i^{bw - e}$ . The final encoded hidden state $h_i^e$ can be generated by using a 2D convolution layer to fuse the forward and backward hidden states ( $h_i^{fw - e}$ and $h_i^{bw - e}$ ).", + "bbox": [ + 212, + 143, + 782, + 196 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "HFComp. In forward Encoder, HFComp (see Fig. 3(b)) adaptively adjusts $h_{i-1}^{f w - e}$ through calculating the spectral cross-attention map [12] $\\operatorname{Att}(Q, K, V) \\in \\mathbb{R}^{C \\times C}$ with $f_{i}$ . The whole procedure of HFComp is formulated as:", + "bbox": [ + 214, + 205, + 782, + 255 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {A t t} (Q, K, V) = V \\cdot \\operatorname {S o f t m a x} (K \\cdot Q), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 367, + 267, + 785, + 285 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nh _ {i - 1} ^ {f w - e c} = \\operatorname {R B} (\\operatorname {A t t} (Q, K, V)) + h _ {i - 1} ^ {f w - e}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 362, + 309, + 785, + 330 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $Q \\in \\mathbb{R}^{C \\times HW}$ , $V \\in \\mathbb{R}^{HW \\times C}$ are linearly projected from $h_{i-1}^{fw}^{-e}$ and $K \\in \\mathbb{R}^{HW \\times C}$ is projected from $f_i$ . To improve the feature processing ability, the adjusted hidden features $h_{i-1}^{fw}^{-ec} \\in \\mathbb{R}^{H \\times W \\times C}$ is obtained from the concatenated residual blocks (RBs [16]).", + "bbox": [ + 212, + 338, + 782, + 404 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "SSRA. The SSRA aim to generate the region-aware encoded hidden state $h_i^e - r$ based on discriminative semantic region learning. The motivation and implementation of discriminative semantic region learning will be elaborated in subsection 3.2.", + "bbox": [ + 214, + 417, + 782, + 477 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "AAM. The attention mechanism is widely used in the Seq2Seq model for addressing the challenges of handling long sequences [30], facilitating the effective capture of crucial information in the input sequence. In the context of HSI denoising, we introduce the Adjacent Attention Mechanism (AAM, see Fig. 3(c)) to capture the spatial-spectral context information from the $2n + 1$ adjusted adjacent encoded hidden states $\\{h_j^e\\}_{i - n}^{i + n}$ according to generated decoded hidden state $h_{i - 1}^{fw - d}$ . Thus, the context feature features $c_{i}$ can be represented by the weighted sum of adjacent encoded hidden states:", + "bbox": [ + 214, + 493, + 784, + 622 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nc _ {i} = \\sum_ {j = i - n} ^ {i + n} \\lambda_ {i j} h _ {j} ^ {e}; \\quad \\left\\{\\lambda_ {i j} \\right\\} _ {i - n} ^ {i + n} = \\operatorname {R S} \\left(h _ {i - 1} ^ {f w} - ^ {d}, \\left\\{h _ {j} ^ {e} \\right\\} _ {i - n} ^ {i + n}\\right), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 310, + 633, + 784, + 676 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\lambda$ is the weight of adjacent encoded hidden states. $\\mathrm{RS}(\\cdot ,\\cdot)$ means to calculate the relevant scores between two input features. In implementation, we use a 3D convolution layer for parameter learning as follows:", + "bbox": [ + 214, + 688, + 782, + 734 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\left\\{\\lambda_ {i _ {j}} \\right\\} _ {i - n} ^ {i + n} = \\operatorname {S o f t m a x} \\left(\\operatorname {T a n h} \\left(\\left[ h _ {i - 1} ^ {f w} - ^ {d}, \\left\\{h _ {j} ^ {e} \\right\\} _ {i - n} ^ {i + n} \\right]\\right)\\right) \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 292, + 744, + 784, + 767 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $[\\cdot ,\\cdot ]$ represents concatenation along channel dimension.", + "bbox": [ + 214, + 779, + 656, + 795 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Decoder. For denoised sequence generation, We employ the identical structure of encoder in the decoder. It serially generate the subsequent output $x_{i}$", + "bbox": [ + 214, + 809, + 782, + 839 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "J. Xiao et al.", + "bbox": [ + 271, + 114, + 359, + 127 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/97b7ea5f00580126092794ab51b9430fa7ab3bcea17fcf69e5d0a6089682ca9b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 233, + 143, + 509, + 268 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b0f58510b1391f68e127259b9965a5d548e91e95e4d51efb01d1c734155af736.jpg", + "image_caption": [ + "Fig. 4: Illustrations of (a) spatial-spectral Fig. 5: Visualizations of spatial and region-aware module (SSRA) and (b) Spec- spectral guided masks on various HSIs. DynConv." + ], + "image_footnote": [], + "bbox": [ + 233, + 268, + 509, + 340 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/27dc62866771648d55cf11df06f271c336786477d20c73d01b97f87830c3f520.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 527, + 143, + 606, + 205 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a305c74d972c12ec2f2d84b2c4d25a380fdbbb1cce197c84c77b5d09fbd54747.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 609, + 143, + 687, + 207 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/9dd2788841eae821f33ea948f992111de7e48c48d10e0ae4d3ea227fac9e19d3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 691, + 143, + 767, + 207 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/56be98780f07f198ea96eac8378dc2a60ecf10034c06ae8e5ebe384a7146949e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 527, + 208, + 606, + 268 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/4c9d49dd654f9a8e34fd61437fbf79f5854e89d830f3b3805794389c23c44ef1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 609, + 208, + 687, + 268 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/2f90c03edf6a84fa7181b8462e33249a1b88ca2edf9f6532e2712b1911c566e0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 691, + 208, + 767, + 268 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/6bc6e7f408cffd22f9af8114c71b0a9f64299d5dac8ff9f39584cb6b32933097.jpg", + "image_caption": [ + "Pseudo Images" + ], + "image_footnote": [], + "bbox": [ + 527, + 271, + 606, + 330 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/137960403d0b18262ed9e80c6f8152bff9b9702c574fcca6feca919dd1cdbf83.jpg", + "image_caption": [ + "Spatial Masks" + ], + "image_footnote": [], + "bbox": [ + 609, + 271, + 687, + 330 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/377cd8775444d35532f836c13b12b2a7e6424eb59ef555584e9fcc387504937f.jpg", + "image_caption": [ + "Spectral Masks" + ], + "image_footnote": [], + "bbox": [ + 691, + 271, + 767, + 330 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "with residual learning [55] in an autoregressive manner, conditioned on context features $c_{i}$ , the decoded hidden feature $h_{i}^{d}$ and generated output $x_{i-1}$ :", + "bbox": [ + 214, + 425, + 782, + 455 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\nh _ {i} ^ {d} = \\operatorname {D e c o d e r} \\left(c _ {i}, x _ {i - 1}\\right), \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 413, + 465, + 784, + 483 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\nx _ {i} = \\operatorname {C o n v 2 D} \\left(h _ {i} ^ {d}\\right) + y _ {i}, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 416, + 508, + 784, + 526 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $c_{i}$ is generated by AAM. The $h_{i}^{d}$ is fused from its forward and backward states $(h_{i - 1}^{fw - d}$ and $h_{i - 1}^{bw - d}$ ) and decoded feature $f_{i}^{d}$ , where $f_{i}^{d}$ is the concatenation feature of generated output $x_{i - 1}$ and context $c_{i}$ . Thus, the inference of $\\{x_{i}\\}_{1}^{B}$ is to maximizes $P(\\{x_{i}\\}_{1}^{B} \\mid \\{h_{i}^{e}\\}_{1}^{B})$ . Based on the chain rule, the goal of Decoder can be reformulated as:", + "bbox": [ + 212, + 530, + 787, + 613 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\arg \\max \\prod_ {i = 1} ^ {B} P \\left(x _ {i} \\mid x _ {1}, \\dots , x _ {i - 1}; h _ {1} ^ {e}, \\dots , h _ {B} ^ {e}\\right). \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 349, + 622, + 784, + 662 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.2 Discriminative Semantic Regions Learning", + "text_level": 1, + "bbox": [ + 214, + 694, + 612, + 709 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Taking a deep insight into the intrinsic properties of HSI (see Fig. 2), we propose the SSRA module (see Fig. 4(a)) to automatically assign learnable filters to their corresponding semantic regions for modeling diverse spatial-spectral variations. Inspired by DRConv [10] (proposed for high-level vision tasks like classification and detection) that divides shallow features into several regions based on local spatial semantics, which is not infeasible and reliable in the regions severely contaminated by noises for HSI, SSRA utilizes HSI-specific spatial-spectral properties to improve semantic region segmentation. In detail, thanks to the Seq2Seq", + "bbox": [ + 212, + 719, + 787, + 840 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "Sequence-to-Sequence Hyperspectral Denoising", + "bbox": [ + 416, + 114, + 732, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "framework, the SSRA includes spatial and spectral mask prediction branches that operate on the hidden state list $\\{h_i^e\\}_1^B$ to guide the allocation of filters, which effectively aggregates spatial-spectral representation.", + "bbox": [ + 212, + 146, + 782, + 191 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In the spatial branch, we propose SpecDynConv (see Fig. 4(b)) that empowers the spatial convolutions with spectral modeling abilities to address the limitations of local spatial semantics and generates more precise spatial mask prediction $M_{i}^{spa} \\in \\mathbb{R}^{H \\times W \\times r}$ ( $r$ means output region number). Specifically, the filter of SpecDynConv $w^{dyn} \\in \\mathbb{R}^{k^2 \\times r \\times C}$ for hidden state $h_{i}^{e}$ is decomposed into the product of a base filter $w^{b} \\in \\mathbb{R}^{k^2 \\times r \\times C}$ and a calibration weight $\\alpha_{i} \\in \\mathbb{R}^{1 \\times 1 \\times C}$ :", + "bbox": [ + 212, + 191, + 784, + 287 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\nw ^ {d y n} = \\alpha_ {i} \\cdot w ^ {b} + 1; \\quad \\alpha_ {i} = \\operatorname {C o n v 1 D} \\left(\\operatorname {G A P} \\left(h _ {i} ^ {e}\\right)\\right), \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 334, + 301, + 784, + 318 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where the $\\alpha_{i}$ is dynamically generated from $h_i^e$ through the global average pooling over the spatial dimension and two-layer 1D convolutions, which effectively model the spectral dynamics for all channels in $w^{b}$ . Different from SENet [19], the calibration weight generated by the input feature can better adapt to itself as illustrated in [20]. Thus, the spatial mask $M_{i}^{spa}$ is calculated as:", + "bbox": [ + 212, + 325, + 782, + 401 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\nM _ {i} ^ {s p a} = \\operatorname {S c a t t e r} \\left(\\operatorname {A r g m a x} \\left(h _ {i} ^ {e} * w ^ {d y n}\\right)\\right) \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 370, + 414, + 784, + 430 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $*$ is 2D convolution operation. $\\operatorname{Argmax}(\\cdot)$ is the index of the maximum value along the channel dimension, indicating which position belongs to which region. Subsequently, the result of $\\operatorname{Argmax}(\\cdot)$ is scattered to its one-hot form.", + "bbox": [ + 212, + 441, + 782, + 487 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In the spectral branch, based on the nature of HSI as illustrated in Fig 2, we propose the SpecClassifier to adaptively learn the spectral inconsistency of each spatial position for semantic region segmentation. Specifically, the SpecClassifier efficiently performs spectral mask prediction using a single-layer linear network. It takes the reshaped hidden state $h_i^e \\in \\mathbb{R}^{HW \\times C}$ as input and produces the spectral mask $M_i^{spec} \\in \\mathbb{R}^{HW \\times r}$ , which is then reshaped to match the dimensions of $M_i^{spa}$ and executes the same index scattering operation as in the spatial branch:", + "bbox": [ + 212, + 488, + 784, + 606 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\nM _ {i} ^ {\\text {s p e c}} = \\operatorname {S c a t t e r} \\left(\\operatorname {A r g m a x} \\left(\\operatorname {L i n e a r} \\left(h _ {i} ^ {e}\\right)\\right)\\right). \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 361, + 609, + 784, + 626 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Through spatial and spectral branches, the encoded state $h_i^e$ is divided into $r$ distinct regions, with each region sharing only one filter [38] in $w = \\{w_1, \\ldots, w_r\\}$ . The summation is applied to obtain the enhanced features $\\{h_i^{spa}, h_i^{spec}\\}$ from $\\{M_i^{spa}, M_i^{spec}\\}$ as below:", + "bbox": [ + 212, + 633, + 782, + 694 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\left\\{h _ {i} ^ {s p a}, h _ {i} ^ {s p e c} \\right\\} = \\sum_ {r} \\left(h _ {i} ^ {e} * w _ {i}\\right) \\otimes \\left\\{M _ {i} ^ {s p a}, M _ {i} ^ {s p e c} \\right\\}, \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 338, + 705, + 784, + 736 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $\\otimes$ denotes the pixel-wise multiplications. Finally, $\\{h_i^{spa}, h_i^{spec}\\}$ are fused to generate the final result $h_i^e - r \\in \\mathbb{R}^{H \\times W \\times C}$ using a 2D convolution layer. To better visualize the region-aware capability of the SSRA module, we use shallow feature $f_i$ as input instead of deep encoded hidden feature $h_i^e$ , Fig. 5 presents the predicted spatial and spectral guided masks ( $r = 2$ ) generated by the SSRA module, demonstrating that our method successfully assigns filters to regions", + "bbox": [ + 212, + 748, + 784, + 839 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "J. Xiao et al.", + "bbox": [ + 271, + 114, + 359, + 127 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "with similar spatial-spectral semantics. It needs to be noted that the spatial and spectral semantic segmentation results for the same position do not always align. The spatial branch emphasizes texture edges for partitioning, and the spectral branch focuses on target spectral characteristics. Therefore, SSRA combines both advantages and predictions for a more comprehensive feature learning.", + "bbox": [ + 212, + 146, + 787, + 223 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 214, + 244, + 375, + 263 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.1 Implementation Details", + "text_level": 1, + "bbox": [ + 214, + 277, + 455, + 292 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Datasets. Following [28], the RAS2S is trained on ICVL [1]. To evaluate the proposed method, seven public benchmark datasets are adopted, including ICVL [1], CAVE [34], KAIST [11], RealHSI [57], and remotely sensed datasets Pavia centre [14], WDC Mall [45], and Urban [33] for testing. Evaluation metrics include PSNR, SSIM [42] and SAM [53].", + "bbox": [ + 212, + 301, + 785, + 380 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Noise Settings. To simulate real-world noise interference, we introduce four types of noise (i.e., non-i.i.d Gaussian, stripe, deadline, and impulse noise) and combine them to create five different noise cases [43]. Among them, the mixture noise case is not introduced in the training process. To assess the generalization of models beyond training noise, we conduct additional validation on i.i.d Gaussian, speckle, and Poisson noise. More details and results are provided in the supplementary material.", + "bbox": [ + 212, + 387, + 785, + 494 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Network Settings. We optimize the model with Adam [22]. The batch size is set to 8 with 90 epochs of training. The learning rate is initialized as 1e-3 and gradually decreased to 1e-6. We train the model on single NVIDIA GTX3090 GPU. Unless otherwise specified, all the quantitative and qualitative comparisons are conducted with 3 learnable regions within SSRA, and the sequence attention length within AAM is set to 3.", + "bbox": [ + 212, + 502, + 785, + 594 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.2 Main Results", + "text_level": 1, + "bbox": [ + 214, + 616, + 372, + 630 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "In all experiments, the RAS2S trained on ICVL [1] is directly adopted to denoise HSI datasets with variant-numbers of bands, without necessitating dataset-specific training and fine-tuning. In contrast, methods such as GRNet [6], T3SC [4], SST [27] and SERT [28], which lack flexibility with a high number of spectral bands, we adopt a sliding window manner for denoising and calculate the average values as the ultimate results. More results are provided in the supplementary material.", + "bbox": [ + 212, + 641, + 795, + 748 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Simulation Scenarios Denoising", + "text_level": 1, + "bbox": [ + 214, + 768, + 478, + 786 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Quantitative Comparison. (i) Noise robustness evaluations on the KAIST [11] dataset are presented in Tab. 1. RAS2S exhibits superior denoising performance across all noise cases, particularly outperforming the state-of-the-art method by", + "bbox": [ + 212, + 794, + 799, + 842 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "Sequence-to-Sequence Hyperspectral Denoising", + "bbox": [ + 416, + 114, + 730, + 130 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/5f71f9bb2afa51582989d212589447fc001dde6d822405c68147996cd5819f7e.jpg", + "table_caption": [ + "Table 1: Average quantitative results under various noise cases on KAIST [11] dataset. The best and second-best values are highlighted." + ], + "table_footnote": [], + "table_body": "
MethodsNon-i.i.d GaussianGaussian+StripeGaussian+DeadlineGaussian+ImpulseGaussian+Mixture
PSNR↑SSIM↑SAM↓PSNR↑SSIM↑SAM↓PSNR↑SSIM↑SAM↓PSNR↑SSIM↑SAM↓PSNR↑SSIM↑SAM↓
Noisy18.260.61451.074118.100.61361.074617.230.57701.096015.000.46251.096513.700.42161.1133
LLRT [8]33.360.93390.166530.850.87250.635130.060.89420.549527.140.71260.740024.740.66850.7764
LRTDTV [40]37.330.95100.219537.160.94980.211336.490.94130.235437.030.95050.244535.600.93550.2503
LLRGTV [18]36.660.91090.367836.680.91520.351034.760.87610.426833.610.85020.510932.140.81660.5218
NGMeet [17]32.900.90380.275532.640.90270.302232.060.90020.296625.790.7230.739524.540.70420.7612
QRNN3D [43]39.420.98740.208539.240.98770.208638.990.98680.216738.150.98530.250537.100.98310.2934
GRNet [6]30.580.91310.523530.410.90930.529830.220.90440.528028.770.88700.599628.290.88470.5780
MAC-Net [49]34.330.95370.479533.370.95200.473332.820.94280.471831.020.92050.592528.410.90410.5840
T3SC [4]37.790.99150.160337.680.99150.167837.040.99030.188834.870.98080.256133.620.97500.2958
GRUnet [24]39.390.98630.223239.210.98690.211338.940.98500.229438.110.98340.281136.690.97410.3525
MST [5]32.680.95450.394232.530.95720.396532.570.95300.405034.850.95690.374734.120.95130.3884
MAN [23]38.760.98870.207638.540.98890.210238.160.98810.211736.380.97790.322035.270.97040.3818
SST [27]36.790.97740.226736.590.97790.229436.340.97620.232335.620.97260.262234.650.96380.3050
BERT [28]37.770.98510.191837.570.98490.195637.350.98360.197636.580.98030.226235.710.97610.2459
HSDT-S [25]40.180.99050.141340.070.99090.140639.760.99200.146539.470.99130.153738.310.99040.1582
RAS2S(Ours)40.700.99160.137340.740.99190.136440.520.99360.129340.140.99330.136239.480.99200.1532
", + "bbox": [ + 218, + 184, + 782, + 354 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/1069fda4e46179645be84635288f02c53f814850e55ee6120e952a5fc19ceda8.jpg", + "table_caption": [ + "Table 2: Average quantitative results under mixture noise case on different datasets. The best and second-best values are highlighted." + ], + "table_footnote": [], + "table_body": "
MethodsICVL [1]CAVE [34]Pavia Centre [14]WDC [45]
PSNR↑SSIM↑SAM↓PSNR↑SSIM↑SAM↓PSNR↑SSIM↑SAM↓PSNR↑SSIM↑SAM↓
Noisy13.970.33920.898714.170.41881.137113.890.34000.974613.980.21481.0120
LLRT [8]24.790.62940.488924.590.57950.817116.630.48950.843722.800.50830.5718
LRTDTV [40]34.460.91840.112733.820.90850.293829.650.89630.244536.330.85970.2148
LLRGTV [18]31.390.87560.253827.120.72210.656728.410.89230.314234.320.82600.3150
NGMeet [17]26.590.77870.401623.200.62290.787422.590.66580.483923.260.50310.4541
QRNN3D [43]39.220.99040.080936.550.98250.424432.930.96980.157033.960.87440.1344
GRNet [6]31.670.95570.143128.440.88990.632926.570.85360.281525.140.76820.3255
MAC-Net [49]30.750.93320.267328.530.89200.623427.340.88130.353030.740.77400.5371
T3SC [4]35.680.97900.138933.610.97280.413731.390.95230.231430.490.90640.1972
GRUNet [24]38.630.99130.065135.980.97640.404931.860.95920.159831.370.80180.1650
MST [5]39.120.99140.058233.970.94730.465731.210.93550.197133.490.85570.2289
MAN [23]35.850.96910.145835.150.97120.499032.940.97110.182032.770.80650.1638
SST [27]39.580.99280.048034.890.96160.409532.850.95880.155530.590.89240.1860
BERT [28]40.440.99410.047035.860.97370.340333.280.96530.145131.330.90980.1609
HSDT-S [25]40.760.99400.050538.520.98920.236234.250.97680.129537.790.94650.1049
RAS2S(Ours)41.330.99530.046739.030.99130.222136.840.98620.095840.920.97600.0735
", + "bbox": [ + 220, + 409, + 782, + 614 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "3.6 dB in untrained mixture noise cases. This highlights the impressive denoising robustness of RAS2S. (ii) Generalization results on multiple datasets are shown in Tab. 2. RAS2S outperforms other methods on ICVL [1] and CAVE [34] (with the same spectral number as the training set). Additionally, RAS2S exhibits stronger generalization abilities on remotely sensed datasets, Pavia Centre [14] and WDC [45], with a larger spectral number than the training set. RAS2S outperforms the state-of-the-art denoiser on all indicators by a large margin, demonstrating robust generalization and insensitivity to spectral categories.", + "bbox": [ + 212, + 645, + 787, + 768 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Qualitative Comparison. Qualitative comparisons between RAS2S and other models in Fig. 6 reveal the following: (i) For ICVL [1] and CAVE [34] with the same spectral number as the training set, other methods generate inaccurate textures and exhibit inferior noise removal performance. (ii) For the KAIST [11]", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "J. Xiao et al.", + "bbox": [ + 271, + 114, + 359, + 126 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/28282d7ce9ad1aafd8fcafcbae489991f765e56d0e41f1ce6e50e0762401588f.jpg", + "image_caption": [ + "Fig. 6: Visual comparisons on prevailing paired datasets. Except for RealHSI [57], which involves real-world noise, the other datasets are contaminated with Non-i.i.d Gaussian+Mixture noise. GT denotes the reference ground-truth." + ], + "image_footnote": [], + "bbox": [ + 222, + 143, + 781, + 364 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/7e4a50cd7719f2109d905c35e0893761fb24e7b35f857f40471b14dc90243426.jpg", + "image_caption": [ + "Fig. 7: Visual comparison on the real-world dataset Urban." + ], + "image_footnote": [], + "bbox": [ + 228, + 431, + 777, + 623 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "dataset with high spatial resolution, RAS2S effectively removes complex noises while retaining detailed textures and edges. In contrast, other models introduce color discrepancies and detail loss (T3SC [4] and SERT [28]) or leave behind residual noise (MACNet [49] and SST [27]). (iii) For the WDC [45] and Pavia Centre datasets with a large spectral number, RAS2S exhibits outstanding noise removal while preserving main structures with fewer artifacts.", + "bbox": [ + 212, + 678, + 787, + 768 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Real-World Scenarios Since the ultimate goal is to deal with complicated real-world noises and produce visually pleasing HSIs, we conduct experiments on the real-world datasets, i.e., RealHSI [57] and Urban [33].", + "bbox": [ + 212, + 794, + 787, + 842 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "Sequence-to-Sequence Hyperspectral Denoising", + "bbox": [ + 416, + 114, + 732, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 116, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/ebeac14cd7b9d64462aefbd24d863586d4e0f6217691980eb302fb4bdfe3b125.jpg", + "table_caption": [ + "Table 3: Efficiency results on CAVE [34] and denoising results on RealHSI [57]." + ], + "table_footnote": [], + "table_body": "
MethodsEfficiencyRealHSI [57]
Params(M)CPU(s)↓GPU(s)↓PSNR(dB)↑SAM↓
LLRT[8]-807.21-26.600.1004
LRTDTV[40]-97.65-26.580.0620
LLRGTV[18]-171.72-26.140.0626
NGMeet[17]-86.03-27.000.0592
QRNN3D[43]0.8640.060.3226.600.0629
GRNet[6]41.442.660.1226.800.1168
MACNet[49]0.434.952.5324.250.1480
T3SC[4]0.8320.730.5126.720.0576
GRUNet[24]14.2875.580.5026.600.0587
MAN230.8727.110.4526.930.0703
SST274.1035.722.0626.950.0711
BERT281.9118.690.5326.510.0850
HSDT-S250.1317.490.3127.250.0544
RAS2S(Ours)0.088.160.5827.390.0535
", + "bbox": [ + 218, + 184, + 516, + 323 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/d476c88369642ad3e64b8375a022645a44a1638c449b2d49b0751d18a2304675.jpg", + "table_caption": [ + "Table 4: Ablation studies to verify module effectiveness." + ], + "table_footnote": [], + "table_body": "
AAMHFCompSSRAParams(KB)↓Time(s)↓PSNR(dB)↑
---79.117.1337.12
✓--79.437.7437.56
✓✓-99.628.3238.45
✓✓✓84.118.1639.03
", + "bbox": [ + 535, + 184, + 779, + 224 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/1383050284ffa89f4a145128eb041f35ee104e557c4d6a69882749ef47f815ab.jpg", + "table_caption": [ + "Table 5: Evaluation of component effectiveness within SSRA." + ], + "table_footnote": [], + "table_body": "
MethodsParams(KB)↓Time(s)↓PSNR(dB)↑
w/oSSRA76.467.6237.51
w/Conv2D82.877.9237.67
w/SpecDynConv84.068.0437.92
w/SpecClassifier82.537.9638.23
w/SSRA (Ours)84.118.1639.03
", + "bbox": [ + 535, + 266, + 782, + 323 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Quantitative Comparison. The quantitative comparison of different methods on the RealHSI [57] is presented in Tab. 3. Model parameters and inference time also can be observed. The CPU and GPU inference time are measured on Intel(R) Xeon(R) Gold 6226R and NVIDIA GTX3090, respectively. As can be seen, our RAS2S outperforms SERT [28] by a significant margin, achieving up to 0.88 dB PSNR scores, while employing even fewer parameters (only $4.18\\%$ of SERT [28] parameters) and runtime. The results indicate that despite the simplicity, our RAS2S present effectiveness in dealing with real-world noise.", + "bbox": [ + 212, + 351, + 787, + 472 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Qualitative Comparison. For visual comparison of denoised Urban [33] in Fig. 7, the competing methods have obvious residual stripe noise and blurring. In contrast, RAS2S still generates more realistic and visually pleasing textures without residual noises and artifacts, demonstrating the robustness of our method in real-world applications.", + "bbox": [ + 212, + 482, + 787, + 559 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "4.3 Ablation Studies", + "text_level": 1, + "bbox": [ + 215, + 594, + 401, + 608 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Unless otherwise specified, we apply the CAVE dataset under the mixture noise for ablation studies. We modify different baselines to a similar model size by adjusting the feature channels for fair comparison.", + "bbox": [ + 212, + 617, + 785, + 664 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Component Analysis. Tab. 4 summarizes the impact of different modules. The base Seq2Seq model achieves competitive performance with minimal model size, confirming the effectiveness of Seq2Seq learning. AAM improves denoising by $0.44\\mathrm{dB}$ , HFComp enhances results by addressing information gaps between hidden states and current spectral features, while SSRA significantly boosts performance by $0.58\\mathrm{dB}$ in PSNR, demonstrating the effectiveness of semantic region learning for spatial-spectral information exploration. Tab. 5 compares SSRA designs, with SpecDynConv enhancing denoising by $0.41\\mathrm{dB}$ and SpecClassifier achieving a $0.72\\mathrm{dB}$ PSNR gain by leveraging spectral properties. SSRA combines these advantages for comprehensive feature learning, resulting in substantial improvement.", + "bbox": [ + 212, + 672, + 787, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "J. Xiao et al.", + "bbox": [ + 271, + 114, + 359, + 126 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/ad31c5d0232d4454a77349411a80f22098fa48901994146667059e5a0871eb06.jpg", + "image_caption": [ + "(a) Attention length" + ], + "image_footnote": [], + "bbox": [ + 225, + 146, + 362, + 190 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/1dbf0846adae0e75d2b94ddef511c826781063d47744584935117cf504e62719.jpg", + "image_caption": [ + "(b) Region number $(r)$" + ], + "image_footnote": [], + "bbox": [ + 364, + 146, + 498, + 191 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/aff925b60b746ea967f0881b202af7ed875af27d42f69d245971141844e0e973.jpg", + "image_caption": [ + "(a) Loss", + "Fig.8: Performance vs. the hyperparameters with different number. The $\\rightarrow$ ,and curves denote the denoising performance on CAVE, ICVL and WDC datasets, respectively.", + "Fig.9: Convergence curve. We show the average training loss (a) and training batch PSNR (b) of the final model." + ], + "image_footnote": [], + "bbox": [ + 526, + 148, + 645, + 195 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/a1ec2a93c7d85cf440b42e89a447241337820757e4988a34ba4489919070408e.jpg", + "image_caption": [ + "(b) PSNR" + ], + "image_footnote": [], + "bbox": [ + 645, + 148, + 767, + 195 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Hyperparameter Analysis. We assess our proposed model under various settings for the sequence attention length in AAM and the learnable-region number in SSRA across multiple datasets. Fig. 8(a) shows that the model achieves optimal performance across all datasets when the sequence attention length in AAM is set to 3, suggesting that the nearest hidden state provides maximum assistance in capturing spatial-spectral relations for the current state. Longer attention lengths introduce excessive redundant spatial-spectral information, adversely affecting denoising performance. Fig. 8(b) indicates optimal performance when the learnable-region number in SSRA is 2 or 3, as the simplified SSRA design limits capacity for more regions, and increasing regions may lead to imprecise segmentation, degrading model performance.", + "bbox": [ + 212, + 315, + 787, + 482 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "4.4 Discussion", + "text_level": 1, + "bbox": [ + 215, + 502, + 349, + 516 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Model Complexity. Although RAS2S includes the quadratic attention module (HFComp), it is simpler than Transformer-based methods as shown in Tab. 6.", + "bbox": [ + 212, + 526, + 784, + 556 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "(i) Fewer parameters: Transformer methods often stack numerous attention layers, while RAS2S utilizes only a single layer of cross-attention. (ii) Lower computational complexity: In most denoising transformers", + "bbox": [ + 212, + 556, + 527, + 632 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "(e.g., SST, HSDT), the spectral feature channels $C$ with quadratic computational overhead are typically more than the spectral bands $B$ . In contrast, the channels $C$ of RAS2S is more less than bands $B$ , resulting in lower model size and computational complexity. Moreover, the smaller parameter space increases the likelihood of finding the global or local optimum during optimization, making it easier for the final model to converge. The convergence curve is shown in the Fig. 9.", + "bbox": [ + 212, + 632, + 785, + 738 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/0aacddae905e3643632d8542cf0de3fd2baff723e906df77b18d54a61b1c1ceb.jpg", + "table_caption": [ + "Table 6: Complexity analysis." + ], + "table_footnote": [], + "table_body": "
ModelPSNR(dB)Params(M)Complexity
SST34.894.10HW(M2C + C2)
HSDT-S38.520.13HWC2B
Ours39.030.08HWC2B
", + "bbox": [ + 529, + 577, + 782, + 630 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Unidirection vs. Bidirection. Existing Seq2Seq models follow unidirectional forward temporal modeling, which is unsuitable for hyperspectral spectral sequences. We validate forward, backward, and bidirectional spectral modeling approaches. Tab. 7 shows that forward and backward spectral modeling yield similar denoising performance, as spectral sequence are not sensitive to the order of spectral bands. In contrast, the bidirectional modeling framework enhances", + "bbox": [ + 212, + 748, + 787, + 840 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "Sequence-to-Sequence Hyperspectral Denoising", + "bbox": [ + 416, + 114, + 732, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 784, + 126 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/c7db54cb8a9e073b417b4c416f0fd1a4be95219a57cc7e19f09093fb9eb5801c.jpg", + "table_caption": [ + "Table 7: Ablation studies to verify bidirection framework." + ], + "table_footnote": [], + "table_body": "
ModelParams(KB)↓Times(s)↓PSNR(dB)↑
Forward91.487.6738.42
Backward91.487.6738.29
Bidirection84.118.1639.03
", + "bbox": [ + 218, + 184, + 460, + 232 + ], + "page_idx": 13 + }, + { + "type": "table", + "img_path": "images/edb541af4d24e6c2135e61451cf5a48bc05ec0a115d39cab5e2195980d019d4c.jpg", + "table_caption": [ + "Table 8: Ablation studies to verify autoregressive paradigm." + ], + "table_footnote": [], + "table_body": "
ModelParams(KB)↓Times(s)↓PSNR(dB)↑
w/o AR80.567.8338.62
w/ AR84.118.1639.03
", + "bbox": [ + 488, + 185, + 763, + 232 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "denoising performance by integrating preceding and succeeding spectral information, effectively capturing long-range spectral dependencies.", + "bbox": [ + 212, + 258, + 782, + 291 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Autoregressive Paradigm. Current per-band denoising methods restore each band independently using spectral information from adjacent bands, but overlook leveraging spectral similarity constraints. Our autoregressive framework addresses this by incorporating information from previously generated bands to enhance the reconstruction of the current band, enabling more effective learning, maintaining consistency across bands, and reducing noise impact. Without the autoregressive paradigm, Equation 9 simplifies to:", + "bbox": [ + 212, + 299, + 784, + 406 + ], + "page_idx": 13 + }, + { + "type": "equation", + "text": "\n$$\n\\arg \\max \\prod_ {i = 1} ^ {B} P \\left(x _ {i} \\mid h _ {1} ^ {e}, \\dots , h _ {B} ^ {e}\\right). \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 395, + 415, + 785, + 455 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "We make an ablation study in Tab. 8. It is evident that adopting the autoregressive paradigm lead to overall better denoising performance for our model.", + "bbox": [ + 212, + 463, + 785, + 494 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 515, + 359, + 532 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "We propose RAS2S, a simple yet effective framework for HSI denoising, leveraging the power of Seq2Seq learning. We regard HSI denoising as a noisy-to-clean sequence translation problem, and maximize the denoising potential of the Seq2Seq model through discriminative semantic region learning, utilizing insights from the intrinsic spatial-spectral properties of HSI. RAS2S uses a simple bidirectional encoder-decoder architecture and serially models spatial-spectral correlations without the need for dataset-specific training, presenting the huge potential for capturing long-range dependencies. More about, RAS2S effectively handles the diverse spatial-spectral semantic variations, delivering more precise denoising results. Extensive experiments prove RAS2S outperforms other mainstream methods and achieves state-of-the-art on prevailing benchmark datasets with few parameters.", + "bbox": [ + 212, + 546, + 787, + 728 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 215, + 750, + 403, + 767 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "This work was supported by National Key R&D Program of China under Grant No. 2021ZD0110400, the Fundamental Research Funds for the Central Universities No. xxj032023020, and sponsored by the CAAI-MindSpore Open Fund, developed on OpenI Community.", + "bbox": [ + 212, + 779, + 785, + 840 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "J. Xiao et al.", + "bbox": [ + 271, + 114, + 359, + 127 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 215, + 143, + 321, + 159 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Arad, B., Ben-Shahar, O.: Sparse recovery of hyperspectral signal from natural rgb images. In: ECCV (2016) 9, 10", + "2. Bai, Y., Zhao, Z., Gong, Y., Wei, X.: Artrackv2: Prompting autoregressive tracker where to look and how to describe. In: CVPR (2024) 2", + "3. Ballas, N., Yao, L., Pal, C., Courville, A.: Delving deeper into convolutional networks for learning video representations. In: ICLR (2016) 5", + "4. Bodrito, T., Zouaoui, A., Chanussot, J., Mairal, J.: A trainable spectral-spatial sparse coding model for hyperspectral image restoration. In: NeurIPS (2021) 9, 10, 11, 12", + "5. Cai, Y., Lin, J., Hu, X., Wang, H., Yuan, X., Zhang, Y., Timofte, R., Van Gool, L.: Mask-guided spectral-wise transformer for efficient hyperspectral image reconstruction. In: CVPR (2022) 2, 10", + "6. Cao, X., Fu, X., Xu, C., Meng, D.: Deep spatial-spectral global reasoning network for hyperspectral image denoising. IEEE TGRS 60, 1-14 (2021) 9, 10, 12", + "7. Chang, Y., Yan, L., Fang, H., Luo, C.: Anisotropic spectral-spatial total variation model for multispectral remote sensing image destriping. IEEE TIP 24(6), 1852-1866 (2015) 4", + "8. Chang, Y., Yan, L., Zhong, S.: Hyper-laplacian regularized unidirectional low-rank tensor recovery for multispectral image denoising. In: CVPR (2017) 3, 10, 12", + "9. Chen, H., Zhao, W., Xu, T., Shi, G., Zhou, S., Liu, P., Li, J.: Spectral-wise implicit neural representation for hyperspectral image reconstruction. IEEE TCSVT (2023) 4", + "0. Chen, J., Wang, X., Guo, Z., Zhang, X., Sun, J.: Dynamic region-aware convolution. In: CVPR (2021) 7", + "1. Choi, I., Kim, M., Gutierrez, D., Jeon, D., Nam, G.: High-quality hyperspectral reconstruction using a spectral prior. ACM TOG 36, 13 (2017) 9, 10", + "2. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. In: ICLR (2020) 6", + "3. Fu, G., Xiong, F., Lu, J., Zhou, J., Qian, Y.: Nonlocal spatial-spectral neural network for hyperspectral image denoising. IEEE TGRS 60, 1-16 (2022) 2", + "4. Gamba, P.: A collection of data for urban area characterization. In: IGARSS (2004) 9, 10", + "5. Guan, J., Lai, R., Li, H., Yang, Y., Gu, L.: Dnrcnn: Deep recurrent convolutional neural network for hsi destriping. IEEE TNNLS (2022) 4", + "6. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: CVPR (2016) 6", + "7. He, W., Yao, Q., Li, C., Yokoya, N., Zhao, Q.: Non-local meets global: An integrated paradigm for hyperspectral denoising. In: CVPR (2019) 3, 10, 12", + "8. He, W., Zhang, H., Shen, H., Zhang, L.: Hyperspectral image denoising using local low-rank matrix recovery and global spatial-spectral total variation. IEEE J-STARS 11(3), 713-729 (2018) 10, 12", + "9. Hu, J., Shen, L., Sun, G.: Squeeze-and-excitation networks. In: CVPR (2018) 8", + "20. Huang, Z., Zhang, S., Pan, L., Qing, Z., Tang, M., Liu, Z., Ang Jr, M.H.: Tada! temporally-adaptive convolutions for video understanding. In: ICLR (2022) 8", + "21. Keneshloo, Y., Shi, T., Ramakrishnan, N., Reddy, C.K.: Deep reinforcement learning for sequence-to-sequence models. IEEE TNNLS 31(7), 2469-2489 (2019) 2" + ], + "bbox": [ + 225, + 178, + 784, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "Sequence-to-Sequence Hyperspectral Denoising", + "bbox": [ + 416, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "22. Kingma, D.P., Ba, J.: Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980 (2014) 9", + "23. Lai, Z., Fu, Y.: Mixed attention network for hyperspectral image denoising. arXiv preprint arXiv:2301.11525 (2023) 10, 12", + "24. Lai, Z., Wei, K., Fu, Y.: Deep plug-and-play prior for hyperspectral image restoration. Neurocomputing 481, 281-293 (2022) 10, 12", + "25. Lai, Z., Yan, C., Fu, Y.: Hybrid spectral denoising transformer with guided attention. In: ICCV (2023) 10, 12", + "26. Li, J., Zhang, J., Chen, F., Zhao, K., Zeng, D.: Adaptive material matching for hyperspectral imagery destringing. IEEE TGRS 60, 1-20 (2022) 1", + "27. Li, M., Fu, Y., Zhang, Y.: Spatial-spectral transformer for hyperspectral image denoising. In: AAAI (2023) 2, 4, 9, 10, 11, 12", + "28. Li, M., Liu, J., Fu, Y., Zhang, Y., Dou, D.: Spectral enhanced rectangle transformer for hyperspectral image denoising. In: CVPR (2023) 2, 4, 9, 10, 11, 12", + "29. Lin, J., Hu, X., Cai, Y., Wang, H., Yan, Y., Zou, X., Zhang, Y., Van Gool, L.: Unsupervised flow-aligned sequence-to-sequence learning for video restoration. In: ICML (2022) 4", + "30. Liu, T., Wang, K., Sha, L., Chang, B., Sui, Z.: Table-to-text generation by structure-aware seq2seq learning. In: AAAI (2018) 6", + "31. Lu, G., Fei, B.: Medical hyperspectral imaging: a review. Journal of biomedical optics 19(1), 010901-010901 (2014) 1", + "32. Maggioni, M., Katkovnik, V., Egiazarian, K., Foi, A.: Nonlocal transform-domain filter for volumetric data denoising and reconstruction. IEEE TIP 22(1), 119-133 (2012) 2", + "33. Mnih, V., Hinton, G.E.: Learning to detect roads in high-resolution aerial images. In: ECCV (2010) 9, 11, 12", + "34. Park, J.I., Lee, M.H., Grossberg, M.D., Nayar, S.K.: Multispectral imaging using multiplexed illumination. In: ICCV (2007) 2, 9, 10, 12", + "35. Peng, J., Xie, Q., Zhao, Q., Wang, Y., Yee, L., Meng, D.: Enhanced 3dtv regularization and its applications on hsi denoising and compressed sensing. IEEE TIP 29, 7889-7903 (2020) 4", + "36. Shi, Q., Tang, X., Yang, T., Liu, R., Zhang, L.: Hyperspectral image denoising using a 3-d attention denoising network. IEEE TGRS 59(12), 10348-10363 (2021) 4", + "37. Sutskever, I., Vinyals, O., Le, Q.V.: Sequence to sequence learning with neural networks. In: NeurIPS (2014) 2, 4", + "38. Wang, H., Xu, K., Lau, R.W.: Local color distributions prior for image enhancement. In: ECCV (2022) 8", + "39. Wang, W., Jiao, W., Hao, Y., Wang, X., Shi, S., Tu, Z., Lyu, M.: Understanding and improving sequence-to-sequence pretraining for neural machine translation. ACL (2022) 4", + "40. Wang, Y., Peng, J., Zhao, Q., Leung, Y., Zhao, X.L., Meng, D.: Hyperspectral image restoration via total variation regularized low-rank tensor decomposition. IEEE J-STARS 11(4), 1227-1243 (2017) 3, 10, 12", + "41. Wang, Z., Ng, M.K., Zhuang, L., Gao, L., Zhang, B.: Nonlocal self-similarity-based hyperspectral remote sensing image denoising with 3-d convolutional neural network. IEEE TGRS 60, 1-17 (2022) 2", + "42. Wang, Z., Bovik, A.C., Sheikh, H.R., Simoncelli, E.P.: Image quality assessment: from error visibility to structural similarity. IEEE TIP 13(4), 600-612 (2004) 9", + "43. Wei, K., Fu, Y., Huang, H.: 3-d quasi-recurrent neural network for hyperspectral image denoising. IEEE TNNLS 32(1), 363-375 (2020) 2, 4, 9, 10, 12" + ], + "bbox": [ + 215, + 146, + 784, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "J. Xiao et al.", + "bbox": [ + 271, + 114, + 359, + 126 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "44. Wei, X., Xiao, J., Gong, Y.: Blind hyperspectral image denoising with degradation information learning. Remote Sensing 15(2), 490 (2023) 4", + "45. Wu, Z., Wang, Q., Wu, Z., Shen, Y.: Total variation-regularized weighted nuclear norm minimization for hyperspectral image mixed denoising. JEE 25(1), 013037-013037 (2016) 9, 10, 11", + "46. Xiao, J., Ji, Y., Wei, X.: Hyperspectral image denoising with spectrum alignment. In: ACMMM. pp. 5495-5503 (2023) 2", + "47. Xiao, J., Wei, X.: Hyperspectral image denoising using uncertainty-aware adjustor. In: IJCAI (2023) 2, 4", + "48. Xiong, F., Zhou, J., Qian, Y.: Material based object tracking in hyperspectral videos. IEEE TGRS 29, 3719-3733 (2020) 1", + "49. Xiong, F., Zhou, J., Zhao, Q., Lu, J., Qian, Y.: Mac-net: Model-aided nonlocal neural network for hyperspectral image denoising. IEEE TGRS 60, 1-14 (2021) 10, 11, 12", + "50. Xu, S., Zhang, X., Wu, Y., Wei, F.: Sequence level contrastive learning for text summarization. In: AAAI (2022) 4", + "51. Yuan, Q., Zhang, Q., Li, J., Shen, H., Zhang, L.: Hyperspectral image denoising employing a spatial-spectral deep residual convolutional neural network. IEEE TGRS 57(2), 1205-1218 (2018) 2, 4", + "52. Yuan, Y., Ma, H., Liu, G.: Partial-dnet: A novel blind denoising model with noise intensity estimation for hsi. IEEE TGRS 60, 1-13 (2021) 2, 4", + "53. Yuhas, R.H., Boardman, J.W., Goetz, A.F.: Determination of semi-arid landscape endmembers and seasonal trends using convex geometry spectral unmixing techniques. In: JPL (1993) 9", + "54. Zhang, J., Tu, Z., Yang, J., Chen, Y., Yuan, J.: Mixste: Seq2seq mixed spatiotemporal encoder for 3d human pose estimation in video. In: CVPR (2022) 4", + "55. Zhang, K., Zuo, W., Chen, Y., Meng, D., Zhang, L.: Beyond a gaussian denoiser: Residual learning of deep cnn for image denoising. IEEE TIP 26(7), 3142-3155 (2017) 7", + "56. Zhang, Q., Zheng, Y., Yuan, Q., Song, M., Yu, H., Xiao, Y.: Hyperspectral image denoising: From model-driven, data-driven, to model-data-driven. IEEE TNNLS (2023) 1", + "57. Zhang, T., Fu, Y., Li, C.: Hyperspectral image denoising with realistic data. In: ICCV (2021) 9, 11, 12", + "58. Zhou, S., Zhang, J., Zuo, W., Loy, C.C.: Cross-scale internal graph neural network for image super-resolution. In: NeurIPS. vol. 33, pp. 3499-3509 (2020) 2", + "59. Zhu, D., Du, B., Zhang, L.: Two-stream convolutional networks for hyperspectral target detection. IEEE TGRS 59(8), 6907-6921 (2020) 1", + "60. Zhuang, L., Gao, L., Zhang, B., Fu, X., Bioucas-Dias, J.M.: Hyperspectral image denoising and anomaly detection based on low-rank and sparse representations. IEEE TGRS 60, 1-17 (2020) 1" + ], + "bbox": [ + 212, + 146, + 787, + 715 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "Sequence-to-Sequence Hyperspectral Denoising", + "bbox": [ + 416, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + } +] \ No newline at end of file diff --git a/2024/Region-Aware Sequence-to-Sequence Learning for Hyperspectral Denoising/fb5f87c2-230f-4062-9b75-422a250a1410_model.json b/2024/Region-Aware Sequence-to-Sequence Learning for Hyperspectral Denoising/fb5f87c2-230f-4062-9b75-422a250a1410_model.json new file mode 100644 index 0000000000000000000000000000000000000000..96399641104b7171b1fbeece4fcf06a94fa92148 --- /dev/null +++ b/2024/Region-Aware Sequence-to-Sequence Learning for Hyperspectral Denoising/fb5f87c2-230f-4062-9b75-422a250a1410_model.json @@ -0,0 +1,2698 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.22, + 0.141, + 0.784, + 0.187 + ], + "angle": 0, + "content": "Region-Aware Sequence-to-Sequence Learning for Hyperspectral Denoising" + }, + { + "type": "text", + "bbox": [ + 0.344, + 0.213, + 0.66, + 0.228 + ], + "angle": 0, + "content": "Jiahua Xiao, Yang Liu, and Xing Wei*" + }, + { + "type": "text", + "bbox": [ + 0.241, + 0.24, + 0.761, + 0.269 + ], + "angle": 0, + "content": "School of Software Engineering, Xi'an Jiaotong University {xjh847286495,1228754216}@stu.xjtu.edu.cn, weixing@mail.xjtu.edu.cn" + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.3, + 0.744, + 0.593 + ], + "angle": 0, + "content": "Abstract. Proper spectral modeling within hyperspectral image (HSI) is critical yet highly challenging for HSI denoising. In contrast to existing methods that struggle between effectiveness and flexibility in modeling long-range spectral dependencies and directly explore spatial-spectral information without region discrimination, we introduce RAS2S—a simple yet effective sequence-to-sequence (Seq2Seq) learning framework for better HSI denoising. RAS2S treats HSI denoising as a Seq2Seq translation problem, which converts the noisy spectral sequence to its clean ones in an autoregressive fashion. In addition, spatial-spectral information exploration without region discrimination contradicts the intrinsic spatial-spectral diversity of HSIs, leading to negative interference from spatial-spectral unrelated regions. Thus we propose a novel spatial-spectral region-aware module to distinctively perceive the semantic regions with different spatial-spectral representations, maximizing the spectral modeling potential of Seq2Seq learning. With such an improved Seq2Seq learning paradigm, RAS2S not only shows huge potential in capturing long-range spectral dependencies, but also maintains the flexibility to handle diverse HSIs. Extensive experiments demonstrate that RAS2S outperforms existing state-of-the-art methods quantitatively and qualitatively with a minimal model size, merely 0.08M. Source code is available at https://github.com/MIV-XJTU/RAS2S." + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.604, + 0.741, + 0.633 + ], + "angle": 0, + "content": "Keywords: Hyperspectral Denoising \\(\\cdot\\) Spectral Dependency \\(\\cdot\\) Sequence-to-Sequence Learning \\(\\cdot\\) Region Discrimination" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.655, + 0.377, + 0.672 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.684, + 0.788, + 0.821 + ], + "angle": 0, + "content": "Unlike conventional images (e.g., grayscale or RGB), hyperspectral images (HSIs) capture the spectral attributes of various objects and materials within a static spatial scenario, widely applied in medical diagnosis [31], material identification [26, 48] and detection [59, 60]. However, various noise types introduced in the imaging process severely disrupt both the visual perception of HSIs and their aforementioned applications, making HSI denoising a crucial yet challenging task. Thanks to the unique spectral information inherited in HSIs [56], which reflects the inner properties of objects, combining spatial and spectral learning allows for more effective noise identification and HSI restoration." + }, + { + "type": "page_footnote", + "bbox": [ + 0.22, + 0.826, + 0.457, + 0.841 + ], + "angle": 0, + "content": "* Corresponding author: Xing Wei." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.128 + ], + "angle": 0, + "content": "J. Xiao et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.473, + 0.358 + ], + "angle": 0, + "content": "In the deep learning era, mainstream CNN-based methods [46, 47, 51, 52] jointly extract spatial-spectral information using 2D convolutional filters over a short window of adjacent spectral bands. However, the limited spectral receptive field severely constrains the utilization of global spectral information. While 3D convolution [13, 41, 43] provides a broader spatial-spectral perspective, it introduces significantly increased complexity. More recently, transformer-based" + }, + { + "type": "image", + "bbox": [ + 0.486, + 0.153, + 0.782, + 0.291 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.479, + 0.303, + 0.788, + 0.346 + ], + "angle": 0, + "content": "Fig. 1: Denoising performance vs. inference time (CPU) under CAVE [34]. The circle areas are proportional to the model size." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.359, + 0.788, + 0.435 + ], + "angle": 0, + "content": "methods [5,27,28] leverage self-attention mechanisms to capture complex spatial-spectral relations and effectively remove noise within HSIs. However, they are unable to handle HSIs with arbitrary spectral numbers and need dataset-specific training. As a result, there seems to be a dilemma between the sufficiency and flexibility of exploring long-range spectral dependencies." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.441, + 0.788, + 0.713 + ], + "angle": 0, + "content": "To solve this dilemma, we consider continuous spectral bands as a sequence and introduce a novel perspective of treating HSI denoising as a sequence translation problem. We attempt to leverage the proven capability of sequence-to-sequence (Seq2Seq) models in sequence modeling [2,21,37] to tackle this problem, which effectively capture long-range dependencies with fewer parameters (shown in Fig. 1) and spectral flexibility. However, directly applying existing Seq2Seq frameworks to spectral sequences poses challenges: (i) Existing Seq2Seq models generate sequences in a forward order (e.g., text and video sequences), hindering their application to spectral sequences, which can utilize both preceding and succeeding bands. To address this, we employ a bidirectional encoder-decoder architecture tailored for spatial-spectral modeling, where the encoder processes noisy bands to hidden states, and the decoder utilizes these states to dynamically generate denoised spectral sequences autoregressively, capturing long-range contextual dependencies bidirectionally. (ii) Existing Seq2Seq models ignore the information gap between adjacent hidden states and the current input, hindering effective hidden state aggregation. We mitigate this by introducing a hidden state compensation module that adjusts adjacent hidden states with cross-attention, alleviating the adverse effects of the information gap." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.788, + 0.842 + ], + "angle": 0, + "content": "In addition, HSIs inherently contain diverse semantic variations across spatial and spectral domains. However, current mainstream convolutional methods perform information extraction using filter sharing across the spatial-spectral domain, treating the feature of each position without discrimination, which will introduces negative interference from semantically unrelated regions during parameter updates. To address this, leveraging insights from spatial non-local self-similarity [32, 58] and spectral distinctiveness (see Fig. 2), we propose discriminative semantic region learning implemented through a Spatial Spectral" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.418, + 0.115, + 0.733, + 0.13 + ], + "angle": 0, + "content": "Sequence-to-Sequence Hyperspectral Denoising" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.223 + ], + "angle": 0, + "content": "Region-Aware (SSRA) module for more focused exploration of spatial-spectral characteristics. SSRA dynamically allocates learnable filters to corresponding semantic regions by utilizing encoded hidden states rich in spatial-spectral aggregation information to generate region mask predictions in both spatial and spectral branches." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.224, + 0.449, + 0.539 + ], + "angle": 0, + "content": "The region-guided masks divide the input into distinct regions, with each region assigned a dedicated filter, eliminating irrelevant spatial-spectral interference and maximizing spectral modeling potential of Seq2Seq learning. Thus, the Region-Aware Sequence-to-Sequence (RAS2S) learning framework is proposed for HSI denoising task. Benefiting from this tailored Seq2Seq architecture, our approach effectively captures long-range spectral dependencies for arbitrary spectral numbers, and eliminates the need for dataset-specific training. Overall, this work makes the following contributions:" + }, + { + "type": "image", + "bbox": [ + 0.46, + 0.228, + 0.785, + 0.374 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.456, + 0.383, + 0.789, + 0.537 + ], + "angle": 0, + "content": "Fig.2: Illustration of spectral distinctiveness. Spatial positions with similar materials have similar spectral reflection patterns (e.g., Position A and B), while those with different materials show divergent patterns (e.g., Position C and D). This motivates assigning learnable filters to corresponding regions with similar spatial-spectral semantic representations, enabling each region to more effectively focus on its own vital spatial-spectral characteristics. We term this discriminative semantic region learning." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.555, + 0.784, + 0.601 + ], + "angle": 0, + "content": "- We introduce RAS2S, the first framework to reformulate the HSI denoising problem from a sequence-to-sequence learning perspective, tailored for spectral sequence modeling." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.601, + 0.785, + 0.647 + ], + "angle": 0, + "content": "- We first propose the discriminative semantic region learning within spatial-spectral domains and implement it through a novel SSRA module to maximize the denoising potential of the Seq2Seq framework." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.647, + 0.785, + 0.692 + ], + "angle": 0, + "content": "- Extensive experiments on prevailing benchmarks demonstrate the superiority and generality of RAS2S, achieving state-of-the-art performance with few parameters." + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.555, + 0.785, + 0.692 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.718, + 0.388, + 0.733 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.752, + 0.524, + 0.768 + ], + "angle": 0, + "content": "2.1 Hyperspectral Image Denoising" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Traditional methods rely on handcrafted priors based on hyperspectral image (HSI) statistics and often struggle when faced with noise distributions inconsistent with prior assumptions. For instance, approaches like [8, 17, 40] capture global spatial-spectral correlation, effectively leveraging low-rank properties and" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.128 + ], + "angle": 0, + "content": "J. Xiao et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.373 + ], + "angle": 0, + "content": "inherent smoothness in HSIs. Others, such as [7,35], utilize total variation priors to preserve marginal features effectively. In the era of deep learning, 2D CNN-based methods like [47,51,52] jointly extract spatial-spectral information through a short window of adjacent spectral bands, constrained by their spectral receptive field. Meanwhile, 3D CNN-based methods like [36,43,44] capture spatial-spectral dependencies but with increased computational cost. RNN-based approaches like [15] explore sequential spectrum analysis but struggle with long-range spatial-spectral dependencies. Recently, transformer-based models such as [9,27,28], leveraging self-attention mechanisms, show promising denoising performance. However, the extensive interaction among global tokens and the need for fixed-length spectral bands pose computational and flexibility challenges for modeling long-range spectral dependencies. In contrast, our model, a simple serial encoder-decoder architecture, effectively captures long-range spectral dependencies and handles HSIs with varying spectral numbers at a minimal model size." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.396, + 0.523, + 0.412 + ], + "angle": 0, + "content": "2.2 Sequence-to-Sequence Learning" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.421, + 0.789, + 0.603 + ], + "angle": 0, + "content": "Sequence-to-Sequence (Seq2Seq) learning [37] is a pivotal framework in natural language processing (NLP) and has shown significant success in various tasks like machine translation [39] and text summarization [50]. These models use an encoder to process input sequences and a decoder to generate output sequences. They evolve with the introduction of attention mechanisms, allowing them to focus on specific parts of the input when generating output. Recently, there are a growing trend in leveraging Seq2Seq models for various computer vision tasks, such as video restoration [29] and pose estimation [54]. Unlike natural language paragraphs and video frame sequences, which follow a temporal order, hyperspectral bands can utilize both preceding and succeeding bands. Tailored modifications are necessary to achieve effective spatial-spectral modeling and maximize the potential of Seq2Seq models for HSI denoising." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.625, + 0.331, + 0.641 + ], + "angle": 0, + "content": "3 Method" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.657, + 0.788, + 0.749 + ], + "angle": 0, + "content": "In the subsequent sections, we represent the pairs of noisy and clean HSI as \\( Y = \\{y_{i}\\}_{1}^{B} \\) and \\( X = \\{x_{i}\\}_{1}^{B} \\), where \\( y_{i} \\) and \\( x_{i} \\) denote individual spectral bands, and \\( Y, X \\in \\mathbb{R}^{H \\times W \\times B} \\). Here, \\( H \\) and \\( W \\) represent the spatial height and width, respectively, while \\( B \\) indicates the number of spectral bands in the HSI. Therefore, the task of HSI denoising is formulated to maximize the conditional probability \\( P(X|Y) \\)." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.77, + 0.587, + 0.785 + ], + "angle": 0, + "content": "3.1 HSI Denoising as Sequence Translation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.787, + 0.842 + ], + "angle": 0, + "content": "We treat HSI denoising as a sequence translation problem and introduce RAS2S. However, existing Seq2Seq models cannot be directly transferred to HSI processing tasks. To effectively models the long-range spectral dependencies, our model" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.418, + 0.115, + 0.733, + 0.13 + ], + "angle": 0, + "content": "Sequence-to-Sequence Hyperspectral Denoising" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "image", + "bbox": [ + 0.229, + 0.144, + 0.77, + 0.379 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.39, + 0.785, + 0.447 + ], + "angle": 0, + "content": "Fig. 3: RAS2S architecture. The workflow of the proposed RAS2S is depicted in (a), which models long-range spectral dependencies with a tailored bidirectional encoder-decoder architecture. The modules at the bottom display the internal details of (b) hidden state compensation (HFComp) and (c) adjacent attention mechanism (AAM)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.477, + 0.788, + 0.595 + ], + "angle": 0, + "content": "uses the idea of bidirectional RNN to capture dependencies in both preceding bands \\(\\{y_{i}\\}_{1}^{i - 1}\\) and succeeding bands \\(\\{y_{i}\\}_{i + 1}^{B}\\), and then outputs the bidirectional propagated hidden state sets \\(\\left(\\left\\{h_i^{fw}\\right\\}_1^B\\right.\\) and \\(\\left\\{h_i^{bw}\\right\\}_1^B\\). For brevity and clarity, we take the forward propagation process as an example to elaborate the workflow of RAS2S in the following section, which can be easily extended to a bidirectional process. As presented in Fig. 3(a), RAS2S consists of the following key components." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.608, + 0.788, + 0.783 + ], + "angle": 0, + "content": "Encoder. In forward propagation process, the goal of encoder is to transform the shallow feature \\(\\{f_i\\}_1^B\\) of each band into the hidden state set \\(\\left\\{h_i^{fw - e}\\right\\}_1^B \\in \\mathbb{R}^{H \\times W \\times C}\\). In order to alleviate the information gap between preceding hidden features \\(h_{i-1}^{fw - e}\\) and current visual features \\(f_i\\), we propose to make the Hidden State Compensation (HFComp) for \\(h_{i-1}^{fw - e}\\) to obtain an adjusted hidden state \\(h_{i-1}^{fw - ec}\\). The HFComp will be introduced immediately. In implementation, we utilize the ConvGRU cell [3] to facilitate the learning of long-range spatial-spectral information, which combines \\(f_i\\) and \\(h_{i-1}^{fw - ec}\\) to generate the final output \\(h_i^{fw - e}\\). Thus, the workflow of encoder for each spectral step can be formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.388, + 0.783, + 0.787, + 0.805 + ], + "angle": 0, + "content": "\\[\nh _ {i - 1} ^ {f w} ^ {- e c} = \\mathrm {H F C o m p} (f _ {i}, h _ {i - 1} ^ {f w} ^ {- e}), \\tag {1}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.384, + 0.821, + 0.787, + 0.843 + ], + "angle": 0, + "content": "\\[\nh _ {i} ^ {f w - e} = \\mathrm {C o n v G R U} (f _ {i}, h _ {i - 1} ^ {f w - e c}), \\tag {2}\n\\]" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.128 + ], + "angle": 0, + "content": "J. Xiao et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.145, + 0.784, + 0.197 + ], + "angle": 0, + "content": "where the backward propagation process also produces the hidden state \\( h_i^{bw - e} \\). The final encoded hidden state \\( h_i^e \\) can be generated by using a 2D convolution layer to fuse the forward and backward hidden states (\\( h_i^{fw - e} \\) and \\( h_i^{bw - e} \\))." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.206, + 0.784, + 0.256 + ], + "angle": 0, + "content": "HFComp. In forward Encoder, HFComp (see Fig. 3(b)) adaptively adjusts \\( h_{i-1}^{f w - e} \\) through calculating the spectral cross-attention map [12] \\( \\operatorname{Att}(Q, K, V) \\in \\mathbb{R}^{C \\times C} \\) with \\( f_{i} \\). The whole procedure of HFComp is formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.368, + 0.268, + 0.786, + 0.286 + ], + "angle": 0, + "content": "\\[\n\\operatorname {A t t} (Q, K, V) = V \\cdot \\operatorname {S o f t m a x} (K \\cdot Q), \\tag {3}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.364, + 0.31, + 0.786, + 0.331 + ], + "angle": 0, + "content": "\\[\nh _ {i - 1} ^ {f w - e c} = \\operatorname {R B} (\\operatorname {A t t} (Q, K, V)) + h _ {i - 1} ^ {f w - e}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.339, + 0.784, + 0.405 + ], + "angle": 0, + "content": "where \\(Q \\in \\mathbb{R}^{C \\times HW}\\), \\(V \\in \\mathbb{R}^{HW \\times C}\\) are linearly projected from \\(h_{i-1}^{fw}^{-e}\\) and \\(K \\in \\mathbb{R}^{HW \\times C}\\) is projected from \\(f_i\\). To improve the feature processing ability, the adjusted hidden features \\(h_{i-1}^{fw}^{-ec} \\in \\mathbb{R}^{H \\times W \\times C}\\) is obtained from the concatenated residual blocks (RBs [16])." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.419, + 0.784, + 0.478 + ], + "angle": 0, + "content": "SSRA. The SSRA aim to generate the region-aware encoded hidden state \\( h_i^e - r \\) based on discriminative semantic region learning. The motivation and implementation of discriminative semantic region learning will be elaborated in subsection 3.2." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.494, + 0.785, + 0.623 + ], + "angle": 0, + "content": "AAM. The attention mechanism is widely used in the Seq2Seq model for addressing the challenges of handling long sequences [30], facilitating the effective capture of crucial information in the input sequence. In the context of HSI denoising, we introduce the Adjacent Attention Mechanism (AAM, see Fig. 3(c)) to capture the spatial-spectral context information from the \\(2n + 1\\) adjusted adjacent encoded hidden states \\(\\{h_j^e\\}_{i - n}^{i + n}\\) according to generated decoded hidden state \\(h_{i - 1}^{fw - d}\\). Thus, the context feature features \\(c_{i}\\) can be represented by the weighted sum of adjacent encoded hidden states:" + }, + { + "type": "equation", + "bbox": [ + 0.311, + 0.635, + 0.785, + 0.677 + ], + "angle": 0, + "content": "\\[\nc _ {i} = \\sum_ {j = i - n} ^ {i + n} \\lambda_ {i j} h _ {j} ^ {e}; \\quad \\left\\{\\lambda_ {i j} \\right\\} _ {i - n} ^ {i + n} = \\operatorname {R S} \\left(h _ {i - 1} ^ {f w} - ^ {d}, \\left\\{h _ {j} ^ {e} \\right\\} _ {i - n} ^ {i + n}\\right), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.689, + 0.784, + 0.735 + ], + "angle": 0, + "content": "where \\(\\lambda\\) is the weight of adjacent encoded hidden states. \\(\\mathrm{RS}(\\cdot ,\\cdot)\\) means to calculate the relevant scores between two input features. In implementation, we use a 3D convolution layer for parameter learning as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.293, + 0.746, + 0.785, + 0.768 + ], + "angle": 0, + "content": "\\[\n\\left\\{\\lambda_ {i _ {j}} \\right\\} _ {i - n} ^ {i + n} = \\operatorname {S o f t m a x} \\left(\\operatorname {T a n h} \\left(\\left[ h _ {i - 1} ^ {f w} - ^ {d}, \\left\\{h _ {j} ^ {e} \\right\\} _ {i - n} ^ {i + n} \\right]\\right)\\right) \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.78, + 0.657, + 0.796 + ], + "angle": 0, + "content": "where \\([\\cdot ,\\cdot ]\\) represents concatenation along channel dimension." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.81, + 0.784, + 0.84 + ], + "angle": 0, + "content": "Decoder. For denoised sequence generation, We employ the identical structure of encoder in the decoder. It serially generate the subsequent output \\( x_{i} \\)" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.418, + 0.115, + 0.733, + 0.13 + ], + "angle": 0, + "content": "Sequence-to-Sequence Hyperspectral Denoising" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "image", + "bbox": [ + 0.234, + 0.145, + 0.511, + 0.269 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.234, + 0.27, + 0.51, + 0.341 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.529, + 0.145, + 0.607, + 0.207 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.61, + 0.145, + 0.688, + 0.208 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.692, + 0.145, + 0.769, + 0.208 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.529, + 0.209, + 0.607, + 0.269 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.61, + 0.209, + 0.688, + 0.269 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.692, + 0.209, + 0.769, + 0.269 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.529, + 0.272, + 0.607, + 0.332 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.61, + 0.272, + 0.688, + 0.332 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.692, + 0.272, + 0.769, + 0.332 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.532, + 0.334, + 0.604, + 0.341 + ], + "angle": 0, + "content": "Pseudo Images" + }, + { + "type": "image_caption", + "bbox": [ + 0.614, + 0.334, + 0.683, + 0.341 + ], + "angle": 0, + "content": "Spatial Masks" + }, + { + "type": "image_caption", + "bbox": [ + 0.693, + 0.334, + 0.768, + 0.341 + ], + "angle": 0, + "content": "Spectral Masks" + }, + { + "type": "image_caption", + "bbox": [ + 0.22, + 0.355, + 0.781, + 0.397 + ], + "angle": 0, + "content": "Fig. 4: Illustrations of (a) spatial-spectral Fig. 5: Visualizations of spatial and region-aware module (SSRA) and (b) Spec- spectral guided masks on various HSIs. DynConv." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.426, + 0.784, + 0.457 + ], + "angle": 0, + "content": "with residual learning [55] in an autoregressive manner, conditioned on context features \\( c_{i} \\), the decoded hidden feature \\( h_{i}^{d} \\) and generated output \\( x_{i-1} \\):" + }, + { + "type": "equation", + "bbox": [ + 0.415, + 0.466, + 0.785, + 0.484 + ], + "angle": 0, + "content": "\\[\nh _ {i} ^ {d} = \\operatorname {D e c o d e r} \\left(c _ {i}, x _ {i - 1}\\right), \\tag {7}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.418, + 0.509, + 0.785, + 0.527 + ], + "angle": 0, + "content": "\\[\nx _ {i} = \\operatorname {C o n v 2 D} \\left(h _ {i} ^ {d}\\right) + y _ {i}, \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.531, + 0.788, + 0.614 + ], + "angle": 0, + "content": "where \\( c_{i} \\) is generated by AAM. The \\( h_{i}^{d} \\) is fused from its forward and backward states \\( (h_{i - 1}^{fw - d} \\) and \\( h_{i - 1}^{bw - d} \\)) and decoded feature \\( f_{i}^{d} \\), where \\( f_{i}^{d} \\) is the concatenation feature of generated output \\( x_{i - 1} \\) and context \\( c_{i} \\). Thus, the inference of \\( \\{x_{i}\\}_{1}^{B} \\) is to maximizes \\( P(\\{x_{i}\\}_{1}^{B} \\mid \\{h_{i}^{e}\\}_{1}^{B}) \\). Based on the chain rule, the goal of Decoder can be reformulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.35, + 0.623, + 0.785, + 0.663 + ], + "angle": 0, + "content": "\\[\n\\arg \\max \\prod_ {i = 1} ^ {B} P \\left(x _ {i} \\mid x _ {1}, \\dots , x _ {i - 1}; h _ {1} ^ {e}, \\dots , h _ {B} ^ {e}\\right). \\tag {9}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.695, + 0.613, + 0.71 + ], + "angle": 0, + "content": "3.2 Discriminative Semantic Regions Learning" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Taking a deep insight into the intrinsic properties of HSI (see Fig. 2), we propose the SSRA module (see Fig. 4(a)) to automatically assign learnable filters to their corresponding semantic regions for modeling diverse spatial-spectral variations. Inspired by DRConv [10] (proposed for high-level vision tasks like classification and detection) that divides shallow features into several regions based on local spatial semantics, which is not infeasible and reliable in the regions severely contaminated by noises for HSI, SSRA utilizes HSI-specific spatial-spectral properties to improve semantic region segmentation. In detail, thanks to the Seq2Seq" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.128 + ], + "angle": 0, + "content": "J. Xiao et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.193 + ], + "angle": 0, + "content": "framework, the SSRA includes spatial and spectral mask prediction branches that operate on the hidden state list \\(\\{h_i^e\\}_1^B\\) to guide the allocation of filters, which effectively aggregates spatial-spectral representation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.193, + 0.785, + 0.288 + ], + "angle": 0, + "content": "In the spatial branch, we propose SpecDynConv (see Fig. 4(b)) that empowers the spatial convolutions with spectral modeling abilities to address the limitations of local spatial semantics and generates more precise spatial mask prediction \\( M_{i}^{spa} \\in \\mathbb{R}^{H \\times W \\times r} \\) (\\( r \\) means output region number). Specifically, the filter of SpecDynConv \\( w^{dyn} \\in \\mathbb{R}^{k^2 \\times r \\times C} \\) for hidden state \\( h_{i}^{e} \\) is decomposed into the product of a base filter \\( w^{b} \\in \\mathbb{R}^{k^2 \\times r \\times C} \\) and a calibration weight \\( \\alpha_{i} \\in \\mathbb{R}^{1 \\times 1 \\times C} \\):" + }, + { + "type": "equation", + "bbox": [ + 0.335, + 0.302, + 0.785, + 0.319 + ], + "angle": 0, + "content": "\\[\nw ^ {d y n} = \\alpha_ {i} \\cdot w ^ {b} + 1; \\quad \\alpha_ {i} = \\operatorname {C o n v 1 D} \\left(\\operatorname {G A P} \\left(h _ {i} ^ {e}\\right)\\right), \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.326, + 0.784, + 0.402 + ], + "angle": 0, + "content": "where the \\(\\alpha_{i}\\) is dynamically generated from \\(h_i^e\\) through the global average pooling over the spatial dimension and two-layer 1D convolutions, which effectively model the spectral dynamics for all channels in \\(w^{b}\\). Different from SENet [19], the calibration weight generated by the input feature can better adapt to itself as illustrated in [20]. Thus, the spatial mask \\(M_{i}^{spa}\\) is calculated as:" + }, + { + "type": "equation", + "bbox": [ + 0.371, + 0.415, + 0.785, + 0.431 + ], + "angle": 0, + "content": "\\[\nM _ {i} ^ {s p a} = \\operatorname {S c a t t e r} \\left(\\operatorname {A r g m a x} \\left(h _ {i} ^ {e} * w ^ {d y n}\\right)\\right) \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.443, + 0.784, + 0.488 + ], + "angle": 0, + "content": "where \\( * \\) is 2D convolution operation. \\( \\operatorname{Argmax}(\\cdot) \\) is the index of the maximum value along the channel dimension, indicating which position belongs to which region. Subsequently, the result of \\( \\operatorname{Argmax}(\\cdot) \\) is scattered to its one-hot form." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.489, + 0.785, + 0.607 + ], + "angle": 0, + "content": "In the spectral branch, based on the nature of HSI as illustrated in Fig 2, we propose the SpecClassifier to adaptively learn the spectral inconsistency of each spatial position for semantic region segmentation. Specifically, the SpecClassifier efficiently performs spectral mask prediction using a single-layer linear network. It takes the reshaped hidden state \\( h_i^e \\in \\mathbb{R}^{HW \\times C} \\) as input and produces the spectral mask \\( M_i^{spec} \\in \\mathbb{R}^{HW \\times r} \\), which is then reshaped to match the dimensions of \\( M_i^{spa} \\) and executes the same index scattering operation as in the spatial branch:" + }, + { + "type": "equation", + "bbox": [ + 0.362, + 0.61, + 0.785, + 0.627 + ], + "angle": 0, + "content": "\\[\nM _ {i} ^ {\\text {s p e c}} = \\operatorname {S c a t t e r} \\left(\\operatorname {A r g m a x} \\left(\\operatorname {L i n e a r} \\left(h _ {i} ^ {e}\\right)\\right)\\right). \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.634, + 0.784, + 0.695 + ], + "angle": 0, + "content": "Through spatial and spectral branches, the encoded state \\( h_i^e \\) is divided into \\( r \\) distinct regions, with each region sharing only one filter [38] in \\( w = \\{w_1, \\ldots, w_r\\} \\). The summation is applied to obtain the enhanced features \\( \\{h_i^{spa}, h_i^{spec}\\} \\) from \\( \\{M_i^{spa}, M_i^{spec}\\} \\) as below:" + }, + { + "type": "equation", + "bbox": [ + 0.339, + 0.707, + 0.785, + 0.737 + ], + "angle": 0, + "content": "\\[\n\\left\\{h _ {i} ^ {s p a}, h _ {i} ^ {s p e c} \\right\\} = \\sum_ {r} \\left(h _ {i} ^ {e} * w _ {i}\\right) \\otimes \\left\\{M _ {i} ^ {s p a}, M _ {i} ^ {s p e c} \\right\\}, \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.749, + 0.785, + 0.84 + ], + "angle": 0, + "content": "where \\(\\otimes\\) denotes the pixel-wise multiplications. Finally, \\(\\{h_i^{spa}, h_i^{spec}\\}\\) are fused to generate the final result \\(h_i^e - r \\in \\mathbb{R}^{H \\times W \\times C}\\) using a 2D convolution layer. To better visualize the region-aware capability of the SSRA module, we use shallow feature \\(f_i\\) as input instead of deep encoded hidden feature \\(h_i^e\\), Fig. 5 presents the predicted spatial and spectral guided masks (\\(r = 2\\)) generated by the SSRA module, demonstrating that our method successfully assigns filters to regions" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.418, + 0.115, + 0.732, + 0.131 + ], + "angle": 0, + "content": "Sequence-to-Sequence Hyperspectral Denoising" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.224 + ], + "angle": 0, + "content": "with similar spatial-spectral semantics. It needs to be noted that the spatial and spectral semantic segmentation results for the same position do not always align. The spatial branch emphasizes texture edges for partitioning, and the spectral branch focuses on target spectral characteristics. Therefore, SSRA combines both advantages and predictions for a more comprehensive feature learning." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.246, + 0.376, + 0.264 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.278, + 0.457, + 0.294 + ], + "angle": 0, + "content": "4.1 Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.302, + 0.787, + 0.381 + ], + "angle": 0, + "content": "Datasets. Following [28], the RAS2S is trained on ICVL [1]. To evaluate the proposed method, seven public benchmark datasets are adopted, including ICVL [1], CAVE [34], KAIST [11], RealHSI [57], and remotely sensed datasets Pavia centre [14], WDC Mall [45], and Urban [33] for testing. Evaluation metrics include PSNR, SSIM [42] and SAM [53]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.388, + 0.787, + 0.495 + ], + "angle": 0, + "content": "Noise Settings. To simulate real-world noise interference, we introduce four types of noise (i.e., non-i.i.d Gaussian, stripe, deadline, and impulse noise) and combine them to create five different noise cases [43]. Among them, the mixture noise case is not introduced in the training process. To assess the generalization of models beyond training noise, we conduct additional validation on i.i.d Gaussian, speckle, and Poisson noise. More details and results are provided in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.503, + 0.787, + 0.595 + ], + "angle": 0, + "content": "Network Settings. We optimize the model with Adam [22]. The batch size is set to 8 with 90 epochs of training. The learning rate is initialized as 1e-3 and gradually decreased to 1e-6. We train the model on single NVIDIA GTX3090 GPU. Unless otherwise specified, all the quantitative and qualitative comparisons are conducted with 3 learnable regions within SSRA, and the sequence attention length within AAM is set to 3." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.617, + 0.373, + 0.631 + ], + "angle": 0, + "content": "4.2 Main Results" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.642, + 0.797, + 0.749 + ], + "angle": 0, + "content": "In all experiments, the RAS2S trained on ICVL [1] is directly adopted to denoise HSI datasets with variant-numbers of bands, without necessitating dataset-specific training and fine-tuning. In contrast, methods such as GRNet [6], T3SC [4], SST [27] and SERT [28], which lack flexibility with a high number of spectral bands, we adopt a sliding window manner for denoising and calculate the average values as the ultimate results. More results are provided in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.77, + 0.48, + 0.787 + ], + "angle": 0, + "content": "Simulation Scenarios Denoising" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.8, + 0.843 + ], + "angle": 0, + "content": "Quantitative Comparison. (i) Noise robustness evaluations on the KAIST [11] dataset are presented in Tab. 1. RAS2S exhibits superior denoising performance across all noise cases, particularly outperforming the state-of-the-art method by" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.127 + ], + "angle": 0, + "content": "J. Xiao et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.217, + 0.145, + 0.784, + 0.174 + ], + "angle": 0, + "content": "Table 1: Average quantitative results under various noise cases on KAIST [11] dataset. The best and second-best values are highlighted." + }, + { + "type": "table", + "bbox": [ + 0.22, + 0.185, + 0.784, + 0.355 + ], + "angle": 0, + "content": "
MethodsNon-i.i.d GaussianGaussian+StripeGaussian+DeadlineGaussian+ImpulseGaussian+Mixture
PSNR↑SSIM↑SAM↓PSNR↑SSIM↑SAM↓PSNR↑SSIM↑SAM↓PSNR↑SSIM↑SAM↓PSNR↑SSIM↑SAM↓
Noisy18.260.61451.074118.100.61361.074617.230.57701.096015.000.46251.096513.700.42161.1133
LLRT [8]33.360.93390.166530.850.87250.635130.060.89420.549527.140.71260.740024.740.66850.7764
LRTDTV [40]37.330.95100.219537.160.94980.211336.490.94130.235437.030.95050.244535.600.93550.2503
LLRGTV [18]36.660.91090.367836.680.91520.351034.760.87610.426833.610.85020.510932.140.81660.5218
NGMeet [17]32.900.90380.275532.640.90270.302232.060.90020.296625.790.7230.739524.540.70420.7612
QRNN3D [43]39.420.98740.208539.240.98770.208638.990.98680.216738.150.98530.250537.100.98310.2934
GRNet [6]30.580.91310.523530.410.90930.529830.220.90440.528028.770.88700.599628.290.88470.5780
MAC-Net [49]34.330.95370.479533.370.95200.473332.820.94280.471831.020.92050.592528.410.90410.5840
T3SC [4]37.790.99150.160337.680.99150.167837.040.99030.188834.870.98080.256133.620.97500.2958
GRUnet [24]39.390.98630.223239.210.98690.211338.940.98500.229438.110.98340.281136.690.97410.3525
MST [5]32.680.95450.394232.530.95720.396532.570.95300.405034.850.95690.374734.120.95130.3884
MAN [23]38.760.98870.207638.540.98890.210238.160.98810.211736.380.97790.322035.270.97040.3818
SST [27]36.790.97740.226736.590.97790.229436.340.97620.232335.620.97260.262234.650.96380.3050
BERT [28]37.770.98510.191837.570.98490.195637.350.98360.197636.580.98030.226235.710.97610.2459
HSDT-S [25]40.180.99050.141340.070.99090.140639.760.99200.146539.470.99130.153738.310.99040.1582
RAS2S(Ours)40.700.99160.137340.740.99190.136440.520.99360.129340.140.99330.136239.480.99200.1532
" + }, + { + "type": "table_caption", + "bbox": [ + 0.217, + 0.37, + 0.784, + 0.398 + ], + "angle": 0, + "content": "Table 2: Average quantitative results under mixture noise case on different datasets. The best and second-best values are highlighted." + }, + { + "type": "table", + "bbox": [ + 0.221, + 0.41, + 0.784, + 0.615 + ], + "angle": 0, + "content": "
MethodsICVL [1]CAVE [34]Pavia Centre [14]WDC [45]
PSNR↑SSIM↑SAM↓PSNR↑SSIM↑SAM↓PSNR↑SSIM↑SAM↓PSNR↑SSIM↑SAM↓
Noisy13.970.33920.898714.170.41881.137113.890.34000.974613.980.21481.0120
LLRT [8]24.790.62940.488924.590.57950.817116.630.48950.843722.800.50830.5718
LRTDTV [40]34.460.91840.112733.820.90850.293829.650.89630.244536.330.85970.2148
LLRGTV [18]31.390.87560.253827.120.72210.656728.410.89230.314234.320.82600.3150
NGMeet [17]26.590.77870.401623.200.62290.787422.590.66580.483923.260.50310.4541
QRNN3D [43]39.220.99040.080936.550.98250.424432.930.96980.157033.960.87440.1344
GRNet [6]31.670.95570.143128.440.88990.632926.570.85360.281525.140.76820.3255
MAC-Net [49]30.750.93320.267328.530.89200.623427.340.88130.353030.740.77400.5371
T3SC [4]35.680.97900.138933.610.97280.413731.390.95230.231430.490.90640.1972
GRUNet [24]38.630.99130.065135.980.97640.404931.860.95920.159831.370.80180.1650
MST [5]39.120.99140.058233.970.94730.465731.210.93550.197133.490.85570.2289
MAN [23]35.850.96910.145835.150.97120.499032.940.97110.182032.770.80650.1638
SST [27]39.580.99280.048034.890.96160.409532.850.95880.155530.590.89240.1860
BERT [28]40.440.99410.047035.860.97370.340333.280.96530.145131.330.90980.1609
HSDT-S [25]40.760.99400.050538.520.98920.236234.250.97680.129537.790.94650.1049
RAS2S(Ours)41.330.99530.046739.030.99130.222136.840.98620.095840.920.97600.0735
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.646, + 0.788, + 0.769 + ], + "angle": 0, + "content": "3.6 dB in untrained mixture noise cases. This highlights the impressive denoising robustness of RAS2S. (ii) Generalization results on multiple datasets are shown in Tab. 2. RAS2S outperforms other methods on ICVL [1] and CAVE [34] (with the same spectral number as the training set). Additionally, RAS2S exhibits stronger generalization abilities on remotely sensed datasets, Pavia Centre [14] and WDC [45], with a larger spectral number than the training set. RAS2S outperforms the state-of-the-art denoiser on all indicators by a large margin, demonstrating robust generalization and insensitivity to spectral categories." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Qualitative Comparison. Qualitative comparisons between RAS2S and other models in Fig. 6 reveal the following: (i) For ICVL [1] and CAVE [34] with the same spectral number as the training set, other methods generate inaccurate textures and exhibit inferior noise removal performance. (ii) For the KAIST [11]" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.418, + 0.115, + 0.733, + 0.13 + ], + "angle": 0, + "content": "Sequence-to-Sequence Hyperspectral Denoising" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.144, + 0.782, + 0.365 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.376, + 0.788, + 0.418 + ], + "angle": 0, + "content": "Fig. 6: Visual comparisons on prevailing paired datasets. Except for RealHSI [57], which involves real-world noise, the other datasets are contaminated with Non-i.i.d Gaussian+Mixture noise. GT denotes the reference ground-truth." + }, + { + "type": "image", + "bbox": [ + 0.229, + 0.433, + 0.779, + 0.624 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.301, + 0.633, + 0.7, + 0.647 + ], + "angle": 0, + "content": "Fig. 7: Visual comparison on the real-world dataset Urban." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.679, + 0.788, + 0.77 + ], + "angle": 0, + "content": "dataset with high spatial resolution, RAS2S effectively removes complex noises while retaining detailed textures and edges. In contrast, other models introduce color discrepancies and detail loss (T3SC [4] and SERT [28]) or leave behind residual noise (MACNet [49] and SST [27]). (iii) For the WDC [45] and Pavia Centre datasets with a large spectral number, RAS2S exhibits outstanding noise removal while preserving main structures with fewer artifacts." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.788, + 0.843 + ], + "angle": 0, + "content": "Real-World Scenarios Since the ultimate goal is to deal with complicated real-world noises and produce visually pleasing HSIs, we conduct experiments on the real-world datasets, i.e., RealHSI [57] and Urban [33]." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.127 + ], + "angle": 0, + "content": "J. Xiao et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.52, + 0.174 + ], + "angle": 0, + "content": "Table 3: Efficiency results on CAVE [34] and denoising results on RealHSI [57]." + }, + { + "type": "table", + "bbox": [ + 0.22, + 0.185, + 0.517, + 0.324 + ], + "angle": 0, + "content": "
MethodsEfficiencyRealHSI [57]
Params(M)CPU(s)↓GPU(s)↓PSNR(dB)↑SAM↓
LLRT[8]-807.21-26.600.1004
LRTDTV[40]-97.65-26.580.0620
LLRGTV[18]-171.72-26.140.0626
NGMeet[17]-86.03-27.000.0592
QRNN3D[43]0.8640.060.3226.600.0629
GRNet[6]41.442.660.1226.800.1168
MACNet[49]0.434.952.5324.250.1480
T3SC[4]0.8320.730.5126.720.0576
GRUNet[24]14.2875.580.5026.600.0587
MAN230.8727.110.4526.930.0703
SST274.1035.722.0626.950.0711
BERT281.9118.690.5326.510.0850
HSDT-S250.1317.490.3127.250.0544
RAS2S(Ours)0.088.160.5827.390.0535
" + }, + { + "type": "table_caption", + "bbox": [ + 0.532, + 0.145, + 0.782, + 0.172 + ], + "angle": 0, + "content": "Table 4: Ablation studies to verify module effectiveness." + }, + { + "type": "table", + "bbox": [ + 0.536, + 0.185, + 0.78, + 0.226 + ], + "angle": 0, + "content": "
AAMHFCompSSRAParams(KB)↓Time(s)↓PSNR(dB)↑
---79.117.1337.12
✓--79.437.7437.56
✓✓-99.628.3238.45
✓✓✓84.118.1639.03
" + }, + { + "type": "table_caption", + "bbox": [ + 0.532, + 0.227, + 0.782, + 0.254 + ], + "angle": 0, + "content": "Table 5: Evaluation of component effectiveness within SSRA." + }, + { + "type": "table", + "bbox": [ + 0.537, + 0.267, + 0.783, + 0.324 + ], + "angle": 0, + "content": "
MethodsParams(KB)↓Time(s)↓PSNR(dB)↑
w/oSSRA76.467.6237.51
w/Conv2D82.877.9237.67
w/SpecDynConv84.068.0437.92
w/SpecClassifier82.537.9638.23
w/SSRA (Ours)84.118.1639.03
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.352, + 0.788, + 0.473 + ], + "angle": 0, + "content": "Quantitative Comparison. The quantitative comparison of different methods on the RealHSI [57] is presented in Tab. 3. Model parameters and inference time also can be observed. The CPU and GPU inference time are measured on Intel(R) Xeon(R) Gold 6226R and NVIDIA GTX3090, respectively. As can be seen, our RAS2S outperforms SERT [28] by a significant margin, achieving up to 0.88 dB PSNR scores, while employing even fewer parameters (only \\(4.18\\%\\) of SERT [28] parameters) and runtime. The results indicate that despite the simplicity, our RAS2S present effectiveness in dealing with real-world noise." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.483, + 0.788, + 0.56 + ], + "angle": 0, + "content": "Qualitative Comparison. For visual comparison of denoised Urban [33] in Fig. 7, the competing methods have obvious residual stripe noise and blurring. In contrast, RAS2S still generates more realistic and visually pleasing textures without residual noises and artifacts, demonstrating the robustness of our method in real-world applications." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.595, + 0.402, + 0.609 + ], + "angle": 0, + "content": "4.3 Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.618, + 0.787, + 0.665 + ], + "angle": 0, + "content": "Unless otherwise specified, we apply the CAVE dataset under the mixture noise for ablation studies. We modify different baselines to a similar model size by adjusting the feature channels for fair comparison." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Component Analysis. Tab. 4 summarizes the impact of different modules. The base Seq2Seq model achieves competitive performance with minimal model size, confirming the effectiveness of Seq2Seq learning. AAM improves denoising by \\(0.44\\mathrm{dB}\\), HFComp enhances results by addressing information gaps between hidden states and current spectral features, while SSRA significantly boosts performance by \\(0.58\\mathrm{dB}\\) in PSNR, demonstrating the effectiveness of semantic region learning for spatial-spectral information exploration. Tab. 5 compares SSRA designs, with SpecDynConv enhancing denoising by \\(0.41\\mathrm{dB}\\) and SpecClassifier achieving a \\(0.72\\mathrm{dB}\\) PSNR gain by leveraging spectral properties. SSRA combines these advantages for comprehensive feature learning, resulting in substantial improvement." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.418, + 0.115, + 0.733, + 0.13 + ], + "angle": 0, + "content": "Sequence-to-Sequence Hyperspectral Denoising" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "image", + "bbox": [ + 0.227, + 0.147, + 0.363, + 0.191 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.252, + 0.193, + 0.352, + 0.202 + ], + "angle": 0, + "content": "(a) Attention length" + }, + { + "type": "image", + "bbox": [ + 0.365, + 0.147, + 0.499, + 0.192 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.383, + 0.194, + 0.496, + 0.203 + ], + "angle": 0, + "content": "(b) Region number \\((r)\\)" + }, + { + "type": "image", + "bbox": [ + 0.527, + 0.149, + 0.646, + 0.196 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.571, + 0.197, + 0.605, + 0.206 + ], + "angle": 0, + "content": "(a) Loss" + }, + { + "type": "image", + "bbox": [ + 0.647, + 0.15, + 0.768, + 0.196 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.686, + 0.197, + 0.725, + 0.206 + ], + "angle": 0, + "content": "(b) PSNR" + }, + { + "type": "image_caption", + "bbox": [ + 0.221, + 0.217, + 0.51, + 0.287 + ], + "angle": 0, + "content": "Fig.8: Performance vs. the hyperparameters with different number. The \\(\\rightarrow\\) ,and curves denote the denoising performance on CAVE, ICVL and WDC datasets, respectively." + }, + { + "type": "image_caption", + "bbox": [ + 0.52, + 0.218, + 0.781, + 0.273 + ], + "angle": 0, + "content": "Fig.9: Convergence curve. We show the average training loss (a) and training batch PSNR (b) of the final model." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.316, + 0.788, + 0.483 + ], + "angle": 0, + "content": "Hyperparameter Analysis. We assess our proposed model under various settings for the sequence attention length in AAM and the learnable-region number in SSRA across multiple datasets. Fig. 8(a) shows that the model achieves optimal performance across all datasets when the sequence attention length in AAM is set to 3, suggesting that the nearest hidden state provides maximum assistance in capturing spatial-spectral relations for the current state. Longer attention lengths introduce excessive redundant spatial-spectral information, adversely affecting denoising performance. Fig. 8(b) indicates optimal performance when the learnable-region number in SSRA is 2 or 3, as the simplified SSRA design limits capacity for more regions, and increasing regions may lead to imprecise segmentation, degrading model performance." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.503, + 0.35, + 0.517 + ], + "angle": 0, + "content": "4.4 Discussion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.527, + 0.785, + 0.557 + ], + "angle": 0, + "content": "Model Complexity. Although RAS2S includes the quadratic attention module (HFComp), it is simpler than Transformer-based methods as shown in Tab. 6." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.558, + 0.528, + 0.633 + ], + "angle": 0, + "content": "(i) Fewer parameters: Transformer methods often stack numerous attention layers, while RAS2S utilizes only a single layer of cross-attention. (ii) Lower computational complexity: In most denoising transformers" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.633, + 0.787, + 0.739 + ], + "angle": 0, + "content": "(e.g., SST, HSDT), the spectral feature channels \\( C \\) with quadratic computational overhead are typically more than the spectral bands \\( B \\). In contrast, the channels \\( C \\) of RAS2S is more less than bands \\( B \\), resulting in lower model size and computational complexity. Moreover, the smaller parameter space increases the likelihood of finding the global or local optimum during optimization, making it easier for the final model to converge. The convergence curve is shown in the Fig. 9." + }, + { + "type": "table_caption", + "bbox": [ + 0.553, + 0.565, + 0.76, + 0.578 + ], + "angle": 0, + "content": "Table 6: Complexity analysis." + }, + { + "type": "table", + "bbox": [ + 0.531, + 0.578, + 0.784, + 0.631 + ], + "angle": 0, + "content": "
ModelPSNR(dB)Params(M)Complexity
SST34.894.10HW(M2C + C2)
HSDT-S38.520.13HWC2B
Ours39.030.08HWC2B
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.749, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Unidirection vs. Bidirection. Existing Seq2Seq models follow unidirectional forward temporal modeling, which is unsuitable for hyperspectral spectral sequences. We validate forward, backward, and bidirectional spectral modeling approaches. Tab. 7 shows that forward and backward spectral modeling yield similar denoising performance, as spectral sequence are not sensitive to the order of spectral bands. In contrast, the bidirectional modeling framework enhances" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.128 + ], + "angle": 0, + "content": "J. Xiao et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.465, + 0.173 + ], + "angle": 0, + "content": "Table 7: Ablation studies to verify bidirection framework." + }, + { + "type": "table", + "bbox": [ + 0.22, + 0.185, + 0.462, + 0.233 + ], + "angle": 0, + "content": "
ModelParams(KB)↓Times(s)↓PSNR(dB)↑
Forward91.487.6738.42
Backward91.487.6738.29
Bidirection84.118.1639.03
" + }, + { + "type": "table_caption", + "bbox": [ + 0.485, + 0.145, + 0.768, + 0.174 + ], + "angle": 0, + "content": "Table 8: Ablation studies to verify autoregressive paradigm." + }, + { + "type": "table", + "bbox": [ + 0.49, + 0.186, + 0.764, + 0.233 + ], + "angle": 0, + "content": "
ModelParams(KB)↓Times(s)↓PSNR(dB)↑
w/o AR80.567.8338.62
w/ AR84.118.1639.03
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.26, + 0.784, + 0.292 + ], + "angle": 0, + "content": "denoising performance by integrating preceding and succeeding spectral information, effectively capturing long-range spectral dependencies." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.3, + 0.785, + 0.407 + ], + "angle": 0, + "content": "Autoregressive Paradigm. Current per-band denoising methods restore each band independently using spectral information from adjacent bands, but overlook leveraging spectral similarity constraints. Our autoregressive framework addresses this by incorporating information from previously generated bands to enhance the reconstruction of the current band, enabling more effective learning, maintaining consistency across bands, and reducing noise impact. Without the autoregressive paradigm, Equation 9 simplifies to:" + }, + { + "type": "equation", + "bbox": [ + 0.396, + 0.416, + 0.786, + 0.456 + ], + "angle": 0, + "content": "\\[\n\\arg \\max \\prod_ {i = 1} ^ {B} P \\left(x _ {i} \\mid h _ {1} ^ {e}, \\dots , h _ {B} ^ {e}\\right). \\tag {14}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.464, + 0.787, + 0.496 + ], + "angle": 0, + "content": "We make an ablation study in Tab. 8. It is evident that adopting the autoregressive paradigm lead to overall better denoising performance for our model." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.516, + 0.36, + 0.533 + ], + "angle": 0, + "content": "5 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.547, + 0.788, + 0.729 + ], + "angle": 0, + "content": "We propose RAS2S, a simple yet effective framework for HSI denoising, leveraging the power of Seq2Seq learning. We regard HSI denoising as a noisy-to-clean sequence translation problem, and maximize the denoising potential of the Seq2Seq model through discriminative semantic region learning, utilizing insights from the intrinsic spatial-spectral properties of HSI. RAS2S uses a simple bidirectional encoder-decoder architecture and serially models spatial-spectral correlations without the need for dataset-specific training, presenting the huge potential for capturing long-range dependencies. More about, RAS2S effectively handles the diverse spatial-spectral semantic variations, delivering more precise denoising results. Extensive experiments prove RAS2S outperforms other mainstream methods and achieves state-of-the-art on prevailing benchmark datasets with few parameters." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.75, + 0.405, + 0.768 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.787, + 0.842 + ], + "angle": 0, + "content": "This work was supported by National Key R&D Program of China under Grant No. 2021ZD0110400, the Fundamental Research Funds for the Central Universities No. xxj032023020, and sponsored by the CAAI-MindSpore Open Fund, developed on OpenI Community." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.418, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Sequence-to-Sequence Hyperspectral Denoising" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.145, + 0.323, + 0.16 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.179, + 0.785, + 0.207 + ], + "angle": 0, + "content": "1. Arad, B., Ben-Shahar, O.: Sparse recovery of hyperspectral signal from natural rgb images. In: ECCV (2016) 9, 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.209, + 0.785, + 0.235 + ], + "angle": 0, + "content": "2. Bai, Y., Zhao, Z., Gong, Y., Wei, X.: Artrackv2: Prompting autoregressive tracker where to look and how to describe. In: CVPR (2024) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.237, + 0.785, + 0.263 + ], + "angle": 0, + "content": "3. Ballas, N., Yao, L., Pal, C., Courville, A.: Delving deeper into convolutional networks for learning video representations. In: ICLR (2016) 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.265, + 0.785, + 0.304 + ], + "angle": 0, + "content": "4. Bodrito, T., Zouaoui, A., Chanussot, J., Mairal, J.: A trainable spectral-spatial sparse coding model for hyperspectral image restoration. In: NeurIPS (2021) 9, 10, 11, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.307, + 0.785, + 0.347 + ], + "angle": 0, + "content": "5. Cai, Y., Lin, J., Hu, X., Wang, H., Yuan, X., Zhang, Y., Timofte, R., Van Gool, L.: Mask-guided spectral-wise transformer for efficient hyperspectral image reconstruction. In: CVPR (2022) 2, 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.349, + 0.785, + 0.376 + ], + "angle": 0, + "content": "6. Cao, X., Fu, X., Xu, C., Meng, D.: Deep spatial-spectral global reasoning network for hyperspectral image denoising. IEEE TGRS 60, 1-14 (2021) 9, 10, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.378, + 0.785, + 0.417 + ], + "angle": 0, + "content": "7. Chang, Y., Yan, L., Fang, H., Luo, C.: Anisotropic spectral-spatial total variation model for multispectral remote sensing image destriping. IEEE TIP 24(6), 1852-1866 (2015) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.419, + 0.785, + 0.446 + ], + "angle": 0, + "content": "8. Chang, Y., Yan, L., Zhong, S.: Hyper-laplacian regularized unidirectional low-rank tensor recovery for multispectral image denoising. In: CVPR (2017) 3, 10, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.448, + 0.785, + 0.486 + ], + "angle": 0, + "content": "9. Chen, H., Zhao, W., Xu, T., Shi, G., Zhou, S., Liu, P., Li, J.: Spectral-wise implicit neural representation for hyperspectral image reconstruction. IEEE TCSVT (2023) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.489, + 0.785, + 0.516 + ], + "angle": 0, + "content": "0. Chen, J., Wang, X., Guo, Z., Zhang, X., Sun, J.: Dynamic region-aware convolution. In: CVPR (2021) 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.518, + 0.785, + 0.544 + ], + "angle": 0, + "content": "1. Choi, I., Kim, M., Gutierrez, D., Jeon, D., Nam, G.: High-quality hyperspectral reconstruction using a spectral prior. ACM TOG 36, 13 (2017) 9, 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.546, + 0.785, + 0.586 + ], + "angle": 0, + "content": "2. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. In: ICLR (2020) 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.588, + 0.785, + 0.614 + ], + "angle": 0, + "content": "3. Fu, G., Xiong, F., Lu, J., Zhou, J., Qian, Y.: Nonlocal spatial-spectral neural network for hyperspectral image denoising. IEEE TGRS 60, 1-16 (2022) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.616, + 0.785, + 0.641 + ], + "angle": 0, + "content": "4. Gamba, P.: A collection of data for urban area characterization. In: IGARSS (2004) 9, 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.644, + 0.785, + 0.671 + ], + "angle": 0, + "content": "5. Guan, J., Lai, R., Li, H., Yang, Y., Gu, L.: Dnrcnn: Deep recurrent convolutional neural network for hsi destriping. IEEE TNNLS (2022) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.673, + 0.785, + 0.699 + ], + "angle": 0, + "content": "6. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: CVPR (2016) 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.701, + 0.785, + 0.727 + ], + "angle": 0, + "content": "7. He, W., Yao, Q., Li, C., Yokoya, N., Zhao, Q.: Non-local meets global: An integrated paradigm for hyperspectral denoising. In: CVPR (2019) 3, 10, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.729, + 0.785, + 0.769 + ], + "angle": 0, + "content": "8. He, W., Zhang, H., Shen, H., Zhang, L.: Hyperspectral image denoising using local low-rank matrix recovery and global spatial-spectral total variation. IEEE J-STARS 11(3), 713-729 (2018) 10, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.771, + 0.772, + 0.784 + ], + "angle": 0, + "content": "9. Hu, J., Shen, L., Sun, G.: Squeeze-and-excitation networks. In: CVPR (2018) 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.786, + 0.785, + 0.812 + ], + "angle": 0, + "content": "20. Huang, Z., Zhang, S., Pan, L., Qing, Z., Tang, M., Liu, Z., Ang Jr, M.H.: Tada! temporally-adaptive convolutions for video understanding. In: ICLR (2022) 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.814, + 0.785, + 0.84 + ], + "angle": 0, + "content": "21. Keneshloo, Y., Shi, T., Ramakrishnan, N., Reddy, C.K.: Deep reinforcement learning for sequence-to-sequence models. IEEE TNNLS 31(7), 2469-2489 (2019) 2" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.179, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.127 + ], + "angle": 0, + "content": "J. Xiao et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.175 + ], + "angle": 0, + "content": "22. Kingma, D.P., Ba, J.: Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980 (2014) 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.176, + 0.785, + 0.202 + ], + "angle": 0, + "content": "23. Lai, Z., Fu, Y.: Mixed attention network for hyperspectral image denoising. arXiv preprint arXiv:2301.11525 (2023) 10, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.203, + 0.785, + 0.229 + ], + "angle": 0, + "content": "24. Lai, Z., Wei, K., Fu, Y.: Deep plug-and-play prior for hyperspectral image restoration. Neurocomputing 481, 281-293 (2022) 10, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.23, + 0.785, + 0.256 + ], + "angle": 0, + "content": "25. Lai, Z., Yan, C., Fu, Y.: Hybrid spectral denoising transformer with guided attention. In: ICCV (2023) 10, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.257, + 0.785, + 0.284 + ], + "angle": 0, + "content": "26. Li, J., Zhang, J., Chen, F., Zhao, K., Zeng, D.: Adaptive material matching for hyperspectral imagery destringing. IEEE TGRS 60, 1-20 (2022) 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.285, + 0.785, + 0.31 + ], + "angle": 0, + "content": "27. Li, M., Fu, Y., Zhang, Y.: Spatial-spectral transformer for hyperspectral image denoising. In: AAAI (2023) 2, 4, 9, 10, 11, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.311, + 0.785, + 0.338 + ], + "angle": 0, + "content": "28. Li, M., Liu, J., Fu, Y., Zhang, Y., Dou, D.: Spectral enhanced rectangle transformer for hyperspectral image denoising. In: CVPR (2023) 2, 4, 9, 10, 11, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.339, + 0.785, + 0.378 + ], + "angle": 0, + "content": "29. Lin, J., Hu, X., Cai, Y., Wang, H., Yan, Y., Zou, X., Zhang, Y., Van Gool, L.: Unsupervised flow-aligned sequence-to-sequence learning for video restoration. In: ICML (2022) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.379, + 0.785, + 0.405 + ], + "angle": 0, + "content": "30. Liu, T., Wang, K., Sha, L., Chang, B., Sui, Z.: Table-to-text generation by structure-aware seq2seq learning. In: AAAI (2018) 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.406, + 0.785, + 0.432 + ], + "angle": 0, + "content": "31. Lu, G., Fei, B.: Medical hyperspectral imaging: a review. Journal of biomedical optics 19(1), 010901-010901 (2014) 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.433, + 0.785, + 0.472 + ], + "angle": 0, + "content": "32. Maggioni, M., Katkovnik, V., Egiazarian, K., Foi, A.: Nonlocal transform-domain filter for volumetric data denoising and reconstruction. IEEE TIP 22(1), 119-133 (2012) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.473, + 0.785, + 0.5 + ], + "angle": 0, + "content": "33. Mnih, V., Hinton, G.E.: Learning to detect roads in high-resolution aerial images. In: ECCV (2010) 9, 11, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.501, + 0.785, + 0.527 + ], + "angle": 0, + "content": "34. Park, J.I., Lee, M.H., Grossberg, M.D., Nayar, S.K.: Multispectral imaging using multiplexed illumination. In: ICCV (2007) 2, 9, 10, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.528, + 0.785, + 0.568 + ], + "angle": 0, + "content": "35. Peng, J., Xie, Q., Zhao, Q., Wang, Y., Yee, L., Meng, D.: Enhanced 3dtv regularization and its applications on hsi denoising and compressed sensing. IEEE TIP 29, 7889-7903 (2020) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.569, + 0.785, + 0.607 + ], + "angle": 0, + "content": "36. Shi, Q., Tang, X., Yang, T., Liu, R., Zhang, L.: Hyperspectral image denoising using a 3-d attention denoising network. IEEE TGRS 59(12), 10348-10363 (2021) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.609, + 0.785, + 0.636 + ], + "angle": 0, + "content": "37. Sutskever, I., Vinyals, O., Le, Q.V.: Sequence to sequence learning with neural networks. In: NeurIPS (2014) 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.637, + 0.785, + 0.663 + ], + "angle": 0, + "content": "38. Wang, H., Xu, K., Lau, R.W.: Local color distributions prior for image enhancement. In: ECCV (2022) 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.664, + 0.785, + 0.703 + ], + "angle": 0, + "content": "39. Wang, W., Jiao, W., Hao, Y., Wang, X., Shi, S., Tu, Z., Lyu, M.: Understanding and improving sequence-to-sequence pretraining for neural machine translation. ACL (2022) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.704, + 0.785, + 0.745 + ], + "angle": 0, + "content": "40. Wang, Y., Peng, J., Zhao, Q., Leung, Y., Zhao, X.L., Meng, D.: Hyperspectral image restoration via total variation regularized low-rank tensor decomposition. IEEE J-STARS 11(4), 1227-1243 (2017) 3, 10, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.746, + 0.785, + 0.786 + ], + "angle": 0, + "content": "41. Wang, Z., Ng, M.K., Zhuang, L., Gao, L., Zhang, B.: Nonlocal self-similarity-based hyperspectral remote sensing image denoising with 3-d convolutional neural network. IEEE TGRS 60, 1-17 (2022) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.787, + 0.785, + 0.813 + ], + "angle": 0, + "content": "42. Wang, Z., Bovik, A.C., Sheikh, H.R., Simoncelli, E.P.: Image quality assessment: from error visibility to structural similarity. IEEE TIP 13(4), 600-612 (2004) 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.814, + 0.785, + 0.84 + ], + "angle": 0, + "content": "43. Wei, K., Fu, Y., Huang, H.: 3-d quasi-recurrent neural network for hyperspectral image denoising. IEEE TNNLS 32(1), 363-375 (2020) 2, 4, 9, 10, 12" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.418, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Sequence-to-Sequence Hyperspectral Denoising" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.175 + ], + "angle": 0, + "content": "44. Wei, X., Xiao, J., Gong, Y.: Blind hyperspectral image denoising with degradation information learning. Remote Sensing 15(2), 490 (2023) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.176, + 0.788, + 0.217 + ], + "angle": 0, + "content": "45. Wu, Z., Wang, Q., Wu, Z., Shen, Y.: Total variation-regularized weighted nuclear norm minimization for hyperspectral image mixed denoising. JEE 25(1), 013037-013037 (2016) 9, 10, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.218, + 0.786, + 0.244 + ], + "angle": 0, + "content": "46. Xiao, J., Ji, Y., Wei, X.: Hyperspectral image denoising with spectrum alignment. In: ACMMM. pp. 5495-5503 (2023) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.245, + 0.785, + 0.272 + ], + "angle": 0, + "content": "47. Xiao, J., Wei, X.: Hyperspectral image denoising using uncertainty-aware adjustor. In: IJCAI (2023) 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.273, + 0.785, + 0.299 + ], + "angle": 0, + "content": "48. Xiong, F., Zhou, J., Qian, Y.: Material based object tracking in hyperspectral videos. IEEE TGRS 29, 3719-3733 (2020) 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.3, + 0.785, + 0.34 + ], + "angle": 0, + "content": "49. Xiong, F., Zhou, J., Zhao, Q., Lu, J., Qian, Y.: Mac-net: Model-aided nonlocal neural network for hyperspectral image denoising. IEEE TGRS 60, 1-14 (2021) 10, 11, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.342, + 0.785, + 0.369 + ], + "angle": 0, + "content": "50. Xu, S., Zhang, X., Wu, Y., Wei, F.: Sequence level contrastive learning for text summarization. In: AAAI (2022) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.37, + 0.785, + 0.41 + ], + "angle": 0, + "content": "51. Yuan, Q., Zhang, Q., Li, J., Shen, H., Zhang, L.: Hyperspectral image denoising employing a spatial-spectral deep residual convolutional neural network. IEEE TGRS 57(2), 1205-1218 (2018) 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.411, + 0.785, + 0.438 + ], + "angle": 0, + "content": "52. Yuan, Y., Ma, H., Liu, G.: Partial-dnet: A novel blind denoising model with noise intensity estimation for hsi. IEEE TGRS 60, 1-13 (2021) 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.439, + 0.785, + 0.48 + ], + "angle": 0, + "content": "53. Yuhas, R.H., Boardman, J.W., Goetz, A.F.: Determination of semi-arid landscape endmembers and seasonal trends using convex geometry spectral unmixing techniques. In: JPL (1993) 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.481, + 0.785, + 0.508 + ], + "angle": 0, + "content": "54. Zhang, J., Tu, Z., Yang, J., Chen, Y., Yuan, J.: Mixste: Seq2seq mixed spatiotemporal encoder for 3d human pose estimation in video. In: CVPR (2022) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.508, + 0.785, + 0.549 + ], + "angle": 0, + "content": "55. Zhang, K., Zuo, W., Chen, Y., Meng, D., Zhang, L.: Beyond a gaussian denoiser: Residual learning of deep cnn for image denoising. IEEE TIP 26(7), 3142-3155 (2017) 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.55, + 0.785, + 0.59 + ], + "angle": 0, + "content": "56. Zhang, Q., Zheng, Y., Yuan, Q., Song, M., Yu, H., Xiao, Y.: Hyperspectral image denoising: From model-driven, data-driven, to model-data-driven. IEEE TNNLS (2023) 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.591, + 0.785, + 0.618 + ], + "angle": 0, + "content": "57. Zhang, T., Fu, Y., Li, C.: Hyperspectral image denoising with realistic data. In: ICCV (2021) 9, 11, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.619, + 0.785, + 0.646 + ], + "angle": 0, + "content": "58. Zhou, S., Zhang, J., Zuo, W., Loy, C.C.: Cross-scale internal graph neural network for image super-resolution. In: NeurIPS. vol. 33, pp. 3499-3509 (2020) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.647, + 0.785, + 0.674 + ], + "angle": 0, + "content": "59. Zhu, D., Du, B., Zhang, L.: Two-stream convolutional networks for hyperspectral target detection. IEEE TGRS 59(8), 6907-6921 (2020) 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.675, + 0.785, + 0.716 + ], + "angle": 0, + "content": "60. Zhuang, L., Gao, L., Zhang, B., Fu, X., Bioucas-Dias, J.M.: Hyperspectral image denoising and anomaly detection based on low-rank and sparse representations. IEEE TGRS 60, 1-17 (2020) 1" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.716 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/Region-Aware Sequence-to-Sequence Learning for Hyperspectral Denoising/fb5f87c2-230f-4062-9b75-422a250a1410_origin.pdf b/2024/Region-Aware Sequence-to-Sequence Learning for Hyperspectral Denoising/fb5f87c2-230f-4062-9b75-422a250a1410_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..261167f718189f6b33e7abd0a514cc186f74fb6f --- /dev/null +++ b/2024/Region-Aware Sequence-to-Sequence Learning for Hyperspectral Denoising/fb5f87c2-230f-4062-9b75-422a250a1410_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:fe2869626d0eeecbbfe7344c382595afd2d10286f3a3aae604038bf91ef8e7e9 +size 9721367 diff --git a/2024/Region-Aware Sequence-to-Sequence Learning for Hyperspectral Denoising/full.md b/2024/Region-Aware Sequence-to-Sequence Learning for Hyperspectral Denoising/full.md new file mode 100644 index 0000000000000000000000000000000000000000..b25b5c91a8c434c662717c740c13332fb1cbce80 --- /dev/null +++ b/2024/Region-Aware Sequence-to-Sequence Learning for Hyperspectral Denoising/full.md @@ -0,0 +1,363 @@ +# Region-Aware Sequence-to-Sequence Learning for Hyperspectral Denoising + +Jiahua Xiao, Yang Liu, and Xing Wei* + +School of Software Engineering, Xi'an Jiaotong University {xjh847286495,1228754216}@stu.xjtu.edu.cn, weixing@mail.xjtu.edu.cn + +Abstract. Proper spectral modeling within hyperspectral image (HSI) is critical yet highly challenging for HSI denoising. In contrast to existing methods that struggle between effectiveness and flexibility in modeling long-range spectral dependencies and directly explore spatial-spectral information without region discrimination, we introduce RAS2S—a simple yet effective sequence-to-sequence (Seq2Seq) learning framework for better HSI denoising. RAS2S treats HSI denoising as a Seq2Seq translation problem, which converts the noisy spectral sequence to its clean ones in an autoregressive fashion. In addition, spatial-spectral information exploration without region discrimination contradicts the intrinsic spatial-spectral diversity of HSIs, leading to negative interference from spatial-spectral unrelated regions. Thus we propose a novel spatial-spectral region-aware module to distinctively perceive the semantic regions with different spatial-spectral representations, maximizing the spectral modeling potential of Seq2Seq learning. With such an improved Seq2Seq learning paradigm, RAS2S not only shows huge potential in capturing long-range spectral dependencies, but also maintains the flexibility to handle diverse HSIs. Extensive experiments demonstrate that RAS2S outperforms existing state-of-the-art methods quantitatively and qualitatively with a minimal model size, merely 0.08M. Source code is available at https://github.com/MIV-XJTU/RAS2S. + +Keywords: Hyperspectral Denoising $\cdot$ Spectral Dependency $\cdot$ Sequence-to-Sequence Learning $\cdot$ Region Discrimination + +# 1 Introduction + +Unlike conventional images (e.g., grayscale or RGB), hyperspectral images (HSIs) capture the spectral attributes of various objects and materials within a static spatial scenario, widely applied in medical diagnosis [31], material identification [26, 48] and detection [59, 60]. However, various noise types introduced in the imaging process severely disrupt both the visual perception of HSIs and their aforementioned applications, making HSI denoising a crucial yet challenging task. Thanks to the unique spectral information inherited in HSIs [56], which reflects the inner properties of objects, combining spatial and spectral learning allows for more effective noise identification and HSI restoration. + +In the deep learning era, mainstream CNN-based methods [46, 47, 51, 52] jointly extract spatial-spectral information using 2D convolutional filters over a short window of adjacent spectral bands. However, the limited spectral receptive field severely constrains the utilization of global spectral information. While 3D convolution [13, 41, 43] provides a broader spatial-spectral perspective, it introduces significantly increased complexity. More recently, transformer-based + +![](images/c9cd5ce64a57d1b1bcdb8d00aa2e31ce0e0c191d40ebafc78053cf0060559bb8.jpg) +Fig. 1: Denoising performance vs. inference time (CPU) under CAVE [34]. The circle areas are proportional to the model size. + +methods [5,27,28] leverage self-attention mechanisms to capture complex spatial-spectral relations and effectively remove noise within HSIs. However, they are unable to handle HSIs with arbitrary spectral numbers and need dataset-specific training. As a result, there seems to be a dilemma between the sufficiency and flexibility of exploring long-range spectral dependencies. + +To solve this dilemma, we consider continuous spectral bands as a sequence and introduce a novel perspective of treating HSI denoising as a sequence translation problem. We attempt to leverage the proven capability of sequence-to-sequence (Seq2Seq) models in sequence modeling [2,21,37] to tackle this problem, which effectively capture long-range dependencies with fewer parameters (shown in Fig. 1) and spectral flexibility. However, directly applying existing Seq2Seq frameworks to spectral sequences poses challenges: (i) Existing Seq2Seq models generate sequences in a forward order (e.g., text and video sequences), hindering their application to spectral sequences, which can utilize both preceding and succeeding bands. To address this, we employ a bidirectional encoder-decoder architecture tailored for spatial-spectral modeling, where the encoder processes noisy bands to hidden states, and the decoder utilizes these states to dynamically generate denoised spectral sequences autoregressively, capturing long-range contextual dependencies bidirectionally. (ii) Existing Seq2Seq models ignore the information gap between adjacent hidden states and the current input, hindering effective hidden state aggregation. We mitigate this by introducing a hidden state compensation module that adjusts adjacent hidden states with cross-attention, alleviating the adverse effects of the information gap. + +In addition, HSIs inherently contain diverse semantic variations across spatial and spectral domains. However, current mainstream convolutional methods perform information extraction using filter sharing across the spatial-spectral domain, treating the feature of each position without discrimination, which will introduces negative interference from semantically unrelated regions during parameter updates. To address this, leveraging insights from spatial non-local self-similarity [32, 58] and spectral distinctiveness (see Fig. 2), we propose discriminative semantic region learning implemented through a Spatial Spectral + +Region-Aware (SSRA) module for more focused exploration of spatial-spectral characteristics. SSRA dynamically allocates learnable filters to corresponding semantic regions by utilizing encoded hidden states rich in spatial-spectral aggregation information to generate region mask predictions in both spatial and spectral branches. + +The region-guided masks divide the input into distinct regions, with each region assigned a dedicated filter, eliminating irrelevant spatial-spectral interference and maximizing spectral modeling potential of Seq2Seq learning. Thus, the Region-Aware Sequence-to-Sequence (RAS2S) learning framework is proposed for HSI denoising task. Benefiting from this tailored Seq2Seq architecture, our approach effectively captures long-range spectral dependencies for arbitrary spectral numbers, and eliminates the need for dataset-specific training. Overall, this work makes the following contributions: + +![](images/beed88671b24c2a3fa22a12c742fb300be28e85b03ceaddee9a7147fcf21b36e.jpg) +Fig.2: Illustration of spectral distinctiveness. Spatial positions with similar materials have similar spectral reflection patterns (e.g., Position A and B), while those with different materials show divergent patterns (e.g., Position C and D). This motivates assigning learnable filters to corresponding regions with similar spatial-spectral semantic representations, enabling each region to more effectively focus on its own vital spatial-spectral characteristics. We term this discriminative semantic region learning. + +- We introduce RAS2S, the first framework to reformulate the HSI denoising problem from a sequence-to-sequence learning perspective, tailored for spectral sequence modeling. +- We first propose the discriminative semantic region learning within spatial-spectral domains and implement it through a novel SSRA module to maximize the denoising potential of the Seq2Seq framework. +- Extensive experiments on prevailing benchmarks demonstrate the superiority and generality of RAS2S, achieving state-of-the-art performance with few parameters. + +# 2 Related Work + +# 2.1 Hyperspectral Image Denoising + +Traditional methods rely on handcrafted priors based on hyperspectral image (HSI) statistics and often struggle when faced with noise distributions inconsistent with prior assumptions. For instance, approaches like [8, 17, 40] capture global spatial-spectral correlation, effectively leveraging low-rank properties and + +inherent smoothness in HSIs. Others, such as [7,35], utilize total variation priors to preserve marginal features effectively. In the era of deep learning, 2D CNN-based methods like [47,51,52] jointly extract spatial-spectral information through a short window of adjacent spectral bands, constrained by their spectral receptive field. Meanwhile, 3D CNN-based methods like [36,43,44] capture spatial-spectral dependencies but with increased computational cost. RNN-based approaches like [15] explore sequential spectrum analysis but struggle with long-range spatial-spectral dependencies. Recently, transformer-based models such as [9,27,28], leveraging self-attention mechanisms, show promising denoising performance. However, the extensive interaction among global tokens and the need for fixed-length spectral bands pose computational and flexibility challenges for modeling long-range spectral dependencies. In contrast, our model, a simple serial encoder-decoder architecture, effectively captures long-range spectral dependencies and handles HSIs with varying spectral numbers at a minimal model size. + +# 2.2 Sequence-to-Sequence Learning + +Sequence-to-Sequence (Seq2Seq) learning [37] is a pivotal framework in natural language processing (NLP) and has shown significant success in various tasks like machine translation [39] and text summarization [50]. These models use an encoder to process input sequences and a decoder to generate output sequences. They evolve with the introduction of attention mechanisms, allowing them to focus on specific parts of the input when generating output. Recently, there are a growing trend in leveraging Seq2Seq models for various computer vision tasks, such as video restoration [29] and pose estimation [54]. Unlike natural language paragraphs and video frame sequences, which follow a temporal order, hyperspectral bands can utilize both preceding and succeeding bands. Tailored modifications are necessary to achieve effective spatial-spectral modeling and maximize the potential of Seq2Seq models for HSI denoising. + +# 3 Method + +In the subsequent sections, we represent the pairs of noisy and clean HSI as $Y = \{y_{i}\}_{1}^{B}$ and $X = \{x_{i}\}_{1}^{B}$ , where $y_{i}$ and $x_{i}$ denote individual spectral bands, and $Y, X \in \mathbb{R}^{H \times W \times B}$ . Here, $H$ and $W$ represent the spatial height and width, respectively, while $B$ indicates the number of spectral bands in the HSI. Therefore, the task of HSI denoising is formulated to maximize the conditional probability $P(X|Y)$ . + +# 3.1 HSI Denoising as Sequence Translation + +We treat HSI denoising as a sequence translation problem and introduce RAS2S. However, existing Seq2Seq models cannot be directly transferred to HSI processing tasks. To effectively models the long-range spectral dependencies, our model + +![](images/7a2dbca4de80fa93412b54dcbf40c4194c754ef5f4c01780dc403d8582bc0cf8.jpg) +Fig. 3: RAS2S architecture. The workflow of the proposed RAS2S is depicted in (a), which models long-range spectral dependencies with a tailored bidirectional encoder-decoder architecture. The modules at the bottom display the internal details of (b) hidden state compensation (HFComp) and (c) adjacent attention mechanism (AAM). + +uses the idea of bidirectional RNN to capture dependencies in both preceding bands $\{y_{i}\}_{1}^{i - 1}$ and succeeding bands $\{y_{i}\}_{i + 1}^{B}$ , and then outputs the bidirectional propagated hidden state sets $\left(\left\{h_i^{fw}\right\}_1^B\right.$ and $\left\{h_i^{bw}\right\}_1^B$ . For brevity and clarity, we take the forward propagation process as an example to elaborate the workflow of RAS2S in the following section, which can be easily extended to a bidirectional process. As presented in Fig. 3(a), RAS2S consists of the following key components. + +Encoder. In forward propagation process, the goal of encoder is to transform the shallow feature $\{f_i\}_1^B$ of each band into the hidden state set $\left\{h_i^{fw - e}\right\}_1^B \in \mathbb{R}^{H \times W \times C}$ . In order to alleviate the information gap between preceding hidden features $h_{i-1}^{fw - e}$ and current visual features $f_i$ , we propose to make the Hidden State Compensation (HFComp) for $h_{i-1}^{fw - e}$ to obtain an adjusted hidden state $h_{i-1}^{fw - ec}$ . The HFComp will be introduced immediately. In implementation, we utilize the ConvGRU cell [3] to facilitate the learning of long-range spatial-spectral information, which combines $f_i$ and $h_{i-1}^{fw - ec}$ to generate the final output $h_i^{fw - e}$ . Thus, the workflow of encoder for each spectral step can be formulated as: + +$$ +h _ {i - 1} ^ {f w} ^ {- e c} = \mathrm {H F C o m p} (f _ {i}, h _ {i - 1} ^ {f w} ^ {- e}), \tag {1} +$$ + +$$ +h _ {i} ^ {f w - e} = \mathrm {C o n v G R U} (f _ {i}, h _ {i - 1} ^ {f w - e c}), \tag {2} +$$ + +where the backward propagation process also produces the hidden state $h_i^{bw - e}$ . The final encoded hidden state $h_i^e$ can be generated by using a 2D convolution layer to fuse the forward and backward hidden states ( $h_i^{fw - e}$ and $h_i^{bw - e}$ ). + +HFComp. In forward Encoder, HFComp (see Fig. 3(b)) adaptively adjusts $h_{i-1}^{f w - e}$ through calculating the spectral cross-attention map [12] $\operatorname{Att}(Q, K, V) \in \mathbb{R}^{C \times C}$ with $f_{i}$ . The whole procedure of HFComp is formulated as: + +$$ +\operatorname {A t t} (Q, K, V) = V \cdot \operatorname {S o f t m a x} (K \cdot Q), \tag {3} +$$ + +$$ +h _ {i - 1} ^ {f w - e c} = \operatorname {R B} (\operatorname {A t t} (Q, K, V)) + h _ {i - 1} ^ {f w - e}, \tag {4} +$$ + +where $Q \in \mathbb{R}^{C \times HW}$ , $V \in \mathbb{R}^{HW \times C}$ are linearly projected from $h_{i-1}^{fw}^{-e}$ and $K \in \mathbb{R}^{HW \times C}$ is projected from $f_i$ . To improve the feature processing ability, the adjusted hidden features $h_{i-1}^{fw}^{-ec} \in \mathbb{R}^{H \times W \times C}$ is obtained from the concatenated residual blocks (RBs [16]). + +SSRA. The SSRA aim to generate the region-aware encoded hidden state $h_i^e - r$ based on discriminative semantic region learning. The motivation and implementation of discriminative semantic region learning will be elaborated in subsection 3.2. + +AAM. The attention mechanism is widely used in the Seq2Seq model for addressing the challenges of handling long sequences [30], facilitating the effective capture of crucial information in the input sequence. In the context of HSI denoising, we introduce the Adjacent Attention Mechanism (AAM, see Fig. 3(c)) to capture the spatial-spectral context information from the $2n + 1$ adjusted adjacent encoded hidden states $\{h_j^e\}_{i - n}^{i + n}$ according to generated decoded hidden state $h_{i - 1}^{fw - d}$ . Thus, the context feature features $c_{i}$ can be represented by the weighted sum of adjacent encoded hidden states: + +$$ +c _ {i} = \sum_ {j = i - n} ^ {i + n} \lambda_ {i j} h _ {j} ^ {e}; \quad \left\{\lambda_ {i j} \right\} _ {i - n} ^ {i + n} = \operatorname {R S} \left(h _ {i - 1} ^ {f w} - ^ {d}, \left\{h _ {j} ^ {e} \right\} _ {i - n} ^ {i + n}\right), \tag {5} +$$ + +where $\lambda$ is the weight of adjacent encoded hidden states. $\mathrm{RS}(\cdot ,\cdot)$ means to calculate the relevant scores between two input features. In implementation, we use a 3D convolution layer for parameter learning as follows: + +$$ +\left\{\lambda_ {i _ {j}} \right\} _ {i - n} ^ {i + n} = \operatorname {S o f t m a x} \left(\operatorname {T a n h} \left(\left[ h _ {i - 1} ^ {f w} - ^ {d}, \left\{h _ {j} ^ {e} \right\} _ {i - n} ^ {i + n} \right]\right)\right) \tag {6} +$$ + +where $[\cdot ,\cdot ]$ represents concatenation along channel dimension. + +Decoder. For denoised sequence generation, We employ the identical structure of encoder in the decoder. It serially generate the subsequent output $x_{i}$ + +![](images/97b7ea5f00580126092794ab51b9430fa7ab3bcea17fcf69e5d0a6089682ca9b.jpg) + +![](images/b0f58510b1391f68e127259b9965a5d548e91e95e4d51efb01d1c734155af736.jpg) +Fig. 4: Illustrations of (a) spatial-spectral Fig. 5: Visualizations of spatial and region-aware module (SSRA) and (b) Spec- spectral guided masks on various HSIs. DynConv. + +![](images/27dc62866771648d55cf11df06f271c336786477d20c73d01b97f87830c3f520.jpg) + +![](images/a305c74d972c12ec2f2d84b2c4d25a380fdbbb1cce197c84c77b5d09fbd54747.jpg) + +![](images/9dd2788841eae821f33ea948f992111de7e48c48d10e0ae4d3ea227fac9e19d3.jpg) + +![](images/56be98780f07f198ea96eac8378dc2a60ecf10034c06ae8e5ebe384a7146949e.jpg) + +![](images/4c9d49dd654f9a8e34fd61437fbf79f5854e89d830f3b3805794389c23c44ef1.jpg) + +![](images/2f90c03edf6a84fa7181b8462e33249a1b88ca2edf9f6532e2712b1911c566e0.jpg) + +![](images/6bc6e7f408cffd22f9af8114c71b0a9f64299d5dac8ff9f39584cb6b32933097.jpg) +Pseudo Images + +![](images/137960403d0b18262ed9e80c6f8152bff9b9702c574fcca6feca919dd1cdbf83.jpg) +Spatial Masks + +![](images/377cd8775444d35532f836c13b12b2a7e6424eb59ef555584e9fcc387504937f.jpg) +Spectral Masks + +with residual learning [55] in an autoregressive manner, conditioned on context features $c_{i}$ , the decoded hidden feature $h_{i}^{d}$ and generated output $x_{i-1}$ : + +$$ +h _ {i} ^ {d} = \operatorname {D e c o d e r} \left(c _ {i}, x _ {i - 1}\right), \tag {7} +$$ + +$$ +x _ {i} = \operatorname {C o n v 2 D} \left(h _ {i} ^ {d}\right) + y _ {i}, \tag {8} +$$ + +where $c_{i}$ is generated by AAM. The $h_{i}^{d}$ is fused from its forward and backward states $(h_{i - 1}^{fw - d}$ and $h_{i - 1}^{bw - d}$ ) and decoded feature $f_{i}^{d}$ , where $f_{i}^{d}$ is the concatenation feature of generated output $x_{i - 1}$ and context $c_{i}$ . Thus, the inference of $\{x_{i}\}_{1}^{B}$ is to maximizes $P(\{x_{i}\}_{1}^{B} \mid \{h_{i}^{e}\}_{1}^{B})$ . Based on the chain rule, the goal of Decoder can be reformulated as: + +$$ +\arg \max \prod_ {i = 1} ^ {B} P \left(x _ {i} \mid x _ {1}, \dots , x _ {i - 1}; h _ {1} ^ {e}, \dots , h _ {B} ^ {e}\right). \tag {9} +$$ + +# 3.2 Discriminative Semantic Regions Learning + +Taking a deep insight into the intrinsic properties of HSI (see Fig. 2), we propose the SSRA module (see Fig. 4(a)) to automatically assign learnable filters to their corresponding semantic regions for modeling diverse spatial-spectral variations. Inspired by DRConv [10] (proposed for high-level vision tasks like classification and detection) that divides shallow features into several regions based on local spatial semantics, which is not infeasible and reliable in the regions severely contaminated by noises for HSI, SSRA utilizes HSI-specific spatial-spectral properties to improve semantic region segmentation. In detail, thanks to the Seq2Seq + +framework, the SSRA includes spatial and spectral mask prediction branches that operate on the hidden state list $\{h_i^e\}_1^B$ to guide the allocation of filters, which effectively aggregates spatial-spectral representation. + +In the spatial branch, we propose SpecDynConv (see Fig. 4(b)) that empowers the spatial convolutions with spectral modeling abilities to address the limitations of local spatial semantics and generates more precise spatial mask prediction $M_{i}^{spa} \in \mathbb{R}^{H \times W \times r}$ ( $r$ means output region number). Specifically, the filter of SpecDynConv $w^{dyn} \in \mathbb{R}^{k^2 \times r \times C}$ for hidden state $h_{i}^{e}$ is decomposed into the product of a base filter $w^{b} \in \mathbb{R}^{k^2 \times r \times C}$ and a calibration weight $\alpha_{i} \in \mathbb{R}^{1 \times 1 \times C}$ : + +$$ +w ^ {d y n} = \alpha_ {i} \cdot w ^ {b} + 1; \quad \alpha_ {i} = \operatorname {C o n v 1 D} \left(\operatorname {G A P} \left(h _ {i} ^ {e}\right)\right), \tag {10} +$$ + +where the $\alpha_{i}$ is dynamically generated from $h_i^e$ through the global average pooling over the spatial dimension and two-layer 1D convolutions, which effectively model the spectral dynamics for all channels in $w^{b}$ . Different from SENet [19], the calibration weight generated by the input feature can better adapt to itself as illustrated in [20]. Thus, the spatial mask $M_{i}^{spa}$ is calculated as: + +$$ +M _ {i} ^ {s p a} = \operatorname {S c a t t e r} \left(\operatorname {A r g m a x} \left(h _ {i} ^ {e} * w ^ {d y n}\right)\right) \tag {11} +$$ + +where $*$ is 2D convolution operation. $\operatorname{Argmax}(\cdot)$ is the index of the maximum value along the channel dimension, indicating which position belongs to which region. Subsequently, the result of $\operatorname{Argmax}(\cdot)$ is scattered to its one-hot form. + +In the spectral branch, based on the nature of HSI as illustrated in Fig 2, we propose the SpecClassifier to adaptively learn the spectral inconsistency of each spatial position for semantic region segmentation. Specifically, the SpecClassifier efficiently performs spectral mask prediction using a single-layer linear network. It takes the reshaped hidden state $h_i^e \in \mathbb{R}^{HW \times C}$ as input and produces the spectral mask $M_i^{spec} \in \mathbb{R}^{HW \times r}$ , which is then reshaped to match the dimensions of $M_i^{spa}$ and executes the same index scattering operation as in the spatial branch: + +$$ +M _ {i} ^ {\text {s p e c}} = \operatorname {S c a t t e r} \left(\operatorname {A r g m a x} \left(\operatorname {L i n e a r} \left(h _ {i} ^ {e}\right)\right)\right). \tag {12} +$$ + +Through spatial and spectral branches, the encoded state $h_i^e$ is divided into $r$ distinct regions, with each region sharing only one filter [38] in $w = \{w_1, \ldots, w_r\}$ . The summation is applied to obtain the enhanced features $\{h_i^{spa}, h_i^{spec}\}$ from $\{M_i^{spa}, M_i^{spec}\}$ as below: + +$$ +\left\{h _ {i} ^ {s p a}, h _ {i} ^ {s p e c} \right\} = \sum_ {r} \left(h _ {i} ^ {e} * w _ {i}\right) \otimes \left\{M _ {i} ^ {s p a}, M _ {i} ^ {s p e c} \right\}, \tag {13} +$$ + +where $\otimes$ denotes the pixel-wise multiplications. Finally, $\{h_i^{spa}, h_i^{spec}\}$ are fused to generate the final result $h_i^e - r \in \mathbb{R}^{H \times W \times C}$ using a 2D convolution layer. To better visualize the region-aware capability of the SSRA module, we use shallow feature $f_i$ as input instead of deep encoded hidden feature $h_i^e$ , Fig. 5 presents the predicted spatial and spectral guided masks ( $r = 2$ ) generated by the SSRA module, demonstrating that our method successfully assigns filters to regions + +with similar spatial-spectral semantics. It needs to be noted that the spatial and spectral semantic segmentation results for the same position do not always align. The spatial branch emphasizes texture edges for partitioning, and the spectral branch focuses on target spectral characteristics. Therefore, SSRA combines both advantages and predictions for a more comprehensive feature learning. + +# 4 Experiments + +# 4.1 Implementation Details + +Datasets. Following [28], the RAS2S is trained on ICVL [1]. To evaluate the proposed method, seven public benchmark datasets are adopted, including ICVL [1], CAVE [34], KAIST [11], RealHSI [57], and remotely sensed datasets Pavia centre [14], WDC Mall [45], and Urban [33] for testing. Evaluation metrics include PSNR, SSIM [42] and SAM [53]. + +Noise Settings. To simulate real-world noise interference, we introduce four types of noise (i.e., non-i.i.d Gaussian, stripe, deadline, and impulse noise) and combine them to create five different noise cases [43]. Among them, the mixture noise case is not introduced in the training process. To assess the generalization of models beyond training noise, we conduct additional validation on i.i.d Gaussian, speckle, and Poisson noise. More details and results are provided in the supplementary material. + +Network Settings. We optimize the model with Adam [22]. The batch size is set to 8 with 90 epochs of training. The learning rate is initialized as 1e-3 and gradually decreased to 1e-6. We train the model on single NVIDIA GTX3090 GPU. Unless otherwise specified, all the quantitative and qualitative comparisons are conducted with 3 learnable regions within SSRA, and the sequence attention length within AAM is set to 3. + +# 4.2 Main Results + +In all experiments, the RAS2S trained on ICVL [1] is directly adopted to denoise HSI datasets with variant-numbers of bands, without necessitating dataset-specific training and fine-tuning. In contrast, methods such as GRNet [6], T3SC [4], SST [27] and SERT [28], which lack flexibility with a high number of spectral bands, we adopt a sliding window manner for denoising and calculate the average values as the ultimate results. More results are provided in the supplementary material. + +# Simulation Scenarios Denoising + +Quantitative Comparison. (i) Noise robustness evaluations on the KAIST [11] dataset are presented in Tab. 1. RAS2S exhibits superior denoising performance across all noise cases, particularly outperforming the state-of-the-art method by + +Table 1: Average quantitative results under various noise cases on KAIST [11] dataset. The best and second-best values are highlighted. + +
MethodsNon-i.i.d GaussianGaussian+StripeGaussian+DeadlineGaussian+ImpulseGaussian+Mixture
PSNR↑SSIM↑SAM↓PSNR↑SSIM↑SAM↓PSNR↑SSIM↑SAM↓PSNR↑SSIM↑SAM↓PSNR↑SSIM↑SAM↓
Noisy18.260.61451.074118.100.61361.074617.230.57701.096015.000.46251.096513.700.42161.1133
LLRT [8]33.360.93390.166530.850.87250.635130.060.89420.549527.140.71260.740024.740.66850.7764
LRTDTV [40]37.330.95100.219537.160.94980.211336.490.94130.235437.030.95050.244535.600.93550.2503
LLRGTV [18]36.660.91090.367836.680.91520.351034.760.87610.426833.610.85020.510932.140.81660.5218
NGMeet [17]32.900.90380.275532.640.90270.302232.060.90020.296625.790.7230.739524.540.70420.7612
QRNN3D [43]39.420.98740.208539.240.98770.208638.990.98680.216738.150.98530.250537.100.98310.2934
GRNet [6]30.580.91310.523530.410.90930.529830.220.90440.528028.770.88700.599628.290.88470.5780
MAC-Net [49]34.330.95370.479533.370.95200.473332.820.94280.471831.020.92050.592528.410.90410.5840
T3SC [4]37.790.99150.160337.680.99150.167837.040.99030.188834.870.98080.256133.620.97500.2958
GRUnet [24]39.390.98630.223239.210.98690.211338.940.98500.229438.110.98340.281136.690.97410.3525
MST [5]32.680.95450.394232.530.95720.396532.570.95300.405034.850.95690.374734.120.95130.3884
MAN [23]38.760.98870.207638.540.98890.210238.160.98810.211736.380.97790.322035.270.97040.3818
SST [27]36.790.97740.226736.590.97790.229436.340.97620.232335.620.97260.262234.650.96380.3050
BERT [28]37.770.98510.191837.570.98490.195637.350.98360.197636.580.98030.226235.710.97610.2459
HSDT-S [25]40.180.99050.141340.070.99090.140639.760.99200.146539.470.99130.153738.310.99040.1582
RAS2S(Ours)40.700.99160.137340.740.99190.136440.520.99360.129340.140.99330.136239.480.99200.1532
+ +Table 2: Average quantitative results under mixture noise case on different datasets. The best and second-best values are highlighted. + +
MethodsICVL [1]CAVE [34]Pavia Centre [14]WDC [45]
PSNR↑SSIM↑SAM↓PSNR↑SSIM↑SAM↓PSNR↑SSIM↑SAM↓PSNR↑SSIM↑SAM↓
Noisy13.970.33920.898714.170.41881.137113.890.34000.974613.980.21481.0120
LLRT [8]24.790.62940.488924.590.57950.817116.630.48950.843722.800.50830.5718
LRTDTV [40]34.460.91840.112733.820.90850.293829.650.89630.244536.330.85970.2148
LLRGTV [18]31.390.87560.253827.120.72210.656728.410.89230.314234.320.82600.3150
NGMeet [17]26.590.77870.401623.200.62290.787422.590.66580.483923.260.50310.4541
QRNN3D [43]39.220.99040.080936.550.98250.424432.930.96980.157033.960.87440.1344
GRNet [6]31.670.95570.143128.440.88990.632926.570.85360.281525.140.76820.3255
MAC-Net [49]30.750.93320.267328.530.89200.623427.340.88130.353030.740.77400.5371
T3SC [4]35.680.97900.138933.610.97280.413731.390.95230.231430.490.90640.1972
GRUNet [24]38.630.99130.065135.980.97640.404931.860.95920.159831.370.80180.1650
MST [5]39.120.99140.058233.970.94730.465731.210.93550.197133.490.85570.2289
MAN [23]35.850.96910.145835.150.97120.499032.940.97110.182032.770.80650.1638
SST [27]39.580.99280.048034.890.96160.409532.850.95880.155530.590.89240.1860
BERT [28]40.440.99410.047035.860.97370.340333.280.96530.145131.330.90980.1609
HSDT-S [25]40.760.99400.050538.520.98920.236234.250.97680.129537.790.94650.1049
RAS2S(Ours)41.330.99530.046739.030.99130.222136.840.98620.095840.920.97600.0735
+ +3.6 dB in untrained mixture noise cases. This highlights the impressive denoising robustness of RAS2S. (ii) Generalization results on multiple datasets are shown in Tab. 2. RAS2S outperforms other methods on ICVL [1] and CAVE [34] (with the same spectral number as the training set). Additionally, RAS2S exhibits stronger generalization abilities on remotely sensed datasets, Pavia Centre [14] and WDC [45], with a larger spectral number than the training set. RAS2S outperforms the state-of-the-art denoiser on all indicators by a large margin, demonstrating robust generalization and insensitivity to spectral categories. + +Qualitative Comparison. Qualitative comparisons between RAS2S and other models in Fig. 6 reveal the following: (i) For ICVL [1] and CAVE [34] with the same spectral number as the training set, other methods generate inaccurate textures and exhibit inferior noise removal performance. (ii) For the KAIST [11] + +![](images/28282d7ce9ad1aafd8fcafcbae489991f765e56d0e41f1ce6e50e0762401588f.jpg) +Fig. 6: Visual comparisons on prevailing paired datasets. Except for RealHSI [57], which involves real-world noise, the other datasets are contaminated with Non-i.i.d Gaussian+Mixture noise. GT denotes the reference ground-truth. + +![](images/7e4a50cd7719f2109d905c35e0893761fb24e7b35f857f40471b14dc90243426.jpg) +Fig. 7: Visual comparison on the real-world dataset Urban. + +dataset with high spatial resolution, RAS2S effectively removes complex noises while retaining detailed textures and edges. In contrast, other models introduce color discrepancies and detail loss (T3SC [4] and SERT [28]) or leave behind residual noise (MACNet [49] and SST [27]). (iii) For the WDC [45] and Pavia Centre datasets with a large spectral number, RAS2S exhibits outstanding noise removal while preserving main structures with fewer artifacts. + +Real-World Scenarios Since the ultimate goal is to deal with complicated real-world noises and produce visually pleasing HSIs, we conduct experiments on the real-world datasets, i.e., RealHSI [57] and Urban [33]. + +Table 3: Efficiency results on CAVE [34] and denoising results on RealHSI [57]. + +
MethodsEfficiencyRealHSI [57]
Params(M)CPU(s)↓GPU(s)↓PSNR(dB)↑SAM↓
LLRT[8]-807.21-26.600.1004
LRTDTV[40]-97.65-26.580.0620
LLRGTV[18]-171.72-26.140.0626
NGMeet[17]-86.03-27.000.0592
QRNN3D[43]0.8640.060.3226.600.0629
GRNet[6]41.442.660.1226.800.1168
MACNet[49]0.434.952.5324.250.1480
T3SC[4]0.8320.730.5126.720.0576
GRUNet[24]14.2875.580.5026.600.0587
MAN230.8727.110.4526.930.0703
SST274.1035.722.0626.950.0711
BERT281.9118.690.5326.510.0850
HSDT-S250.1317.490.3127.250.0544
RAS2S(Ours)0.088.160.5827.390.0535
+ +Table 4: Ablation studies to verify module effectiveness. + +
AAMHFCompSSRAParams(KB)↓Time(s)↓PSNR(dB)↑
---79.117.1337.12
✓--79.437.7437.56
✓✓-99.628.3238.45
✓✓✓84.118.1639.03
+ +Table 5: Evaluation of component effectiveness within SSRA. + +
MethodsParams(KB)↓Time(s)↓PSNR(dB)↑
w/oSSRA76.467.6237.51
w/Conv2D82.877.9237.67
w/SpecDynConv84.068.0437.92
w/SpecClassifier82.537.9638.23
w/SSRA (Ours)84.118.1639.03
+ +Quantitative Comparison. The quantitative comparison of different methods on the RealHSI [57] is presented in Tab. 3. Model parameters and inference time also can be observed. The CPU and GPU inference time are measured on Intel(R) Xeon(R) Gold 6226R and NVIDIA GTX3090, respectively. As can be seen, our RAS2S outperforms SERT [28] by a significant margin, achieving up to 0.88 dB PSNR scores, while employing even fewer parameters (only $4.18\%$ of SERT [28] parameters) and runtime. The results indicate that despite the simplicity, our RAS2S present effectiveness in dealing with real-world noise. + +Qualitative Comparison. For visual comparison of denoised Urban [33] in Fig. 7, the competing methods have obvious residual stripe noise and blurring. In contrast, RAS2S still generates more realistic and visually pleasing textures without residual noises and artifacts, demonstrating the robustness of our method in real-world applications. + +# 4.3 Ablation Studies + +Unless otherwise specified, we apply the CAVE dataset under the mixture noise for ablation studies. We modify different baselines to a similar model size by adjusting the feature channels for fair comparison. + +Component Analysis. Tab. 4 summarizes the impact of different modules. The base Seq2Seq model achieves competitive performance with minimal model size, confirming the effectiveness of Seq2Seq learning. AAM improves denoising by $0.44\mathrm{dB}$ , HFComp enhances results by addressing information gaps between hidden states and current spectral features, while SSRA significantly boosts performance by $0.58\mathrm{dB}$ in PSNR, demonstrating the effectiveness of semantic region learning for spatial-spectral information exploration. Tab. 5 compares SSRA designs, with SpecDynConv enhancing denoising by $0.41\mathrm{dB}$ and SpecClassifier achieving a $0.72\mathrm{dB}$ PSNR gain by leveraging spectral properties. SSRA combines these advantages for comprehensive feature learning, resulting in substantial improvement. + +![](images/ad31c5d0232d4454a77349411a80f22098fa48901994146667059e5a0871eb06.jpg) +(a) Attention length + +![](images/1dbf0846adae0e75d2b94ddef511c826781063d47744584935117cf504e62719.jpg) +(b) Region number $(r)$ + +![](images/aff925b60b746ea967f0881b202af7ed875af27d42f69d245971141844e0e973.jpg) +(a) Loss +Fig.8: Performance vs. the hyperparameters with different number. The $\rightarrow$ ,and curves denote the denoising performance on CAVE, ICVL and WDC datasets, respectively. +Fig.9: Convergence curve. We show the average training loss (a) and training batch PSNR (b) of the final model. + +![](images/a1ec2a93c7d85cf440b42e89a447241337820757e4988a34ba4489919070408e.jpg) +(b) PSNR + +Hyperparameter Analysis. We assess our proposed model under various settings for the sequence attention length in AAM and the learnable-region number in SSRA across multiple datasets. Fig. 8(a) shows that the model achieves optimal performance across all datasets when the sequence attention length in AAM is set to 3, suggesting that the nearest hidden state provides maximum assistance in capturing spatial-spectral relations for the current state. Longer attention lengths introduce excessive redundant spatial-spectral information, adversely affecting denoising performance. Fig. 8(b) indicates optimal performance when the learnable-region number in SSRA is 2 or 3, as the simplified SSRA design limits capacity for more regions, and increasing regions may lead to imprecise segmentation, degrading model performance. + +# 4.4 Discussion + +Model Complexity. Although RAS2S includes the quadratic attention module (HFComp), it is simpler than Transformer-based methods as shown in Tab. 6. + +(i) Fewer parameters: Transformer methods often stack numerous attention layers, while RAS2S utilizes only a single layer of cross-attention. (ii) Lower computational complexity: In most denoising transformers + +(e.g., SST, HSDT), the spectral feature channels $C$ with quadratic computational overhead are typically more than the spectral bands $B$ . In contrast, the channels $C$ of RAS2S is more less than bands $B$ , resulting in lower model size and computational complexity. Moreover, the smaller parameter space increases the likelihood of finding the global or local optimum during optimization, making it easier for the final model to converge. The convergence curve is shown in the Fig. 9. + +Table 6: Complexity analysis. + +
ModelPSNR(dB)Params(M)Complexity
SST34.894.10HW(M2C + C2)
HSDT-S38.520.13HWC2B
Ours39.030.08HWC2B
+ +Unidirection vs. Bidirection. Existing Seq2Seq models follow unidirectional forward temporal modeling, which is unsuitable for hyperspectral spectral sequences. We validate forward, backward, and bidirectional spectral modeling approaches. Tab. 7 shows that forward and backward spectral modeling yield similar denoising performance, as spectral sequence are not sensitive to the order of spectral bands. In contrast, the bidirectional modeling framework enhances + +Table 7: Ablation studies to verify bidirection framework. + +
ModelParams(KB)↓Times(s)↓PSNR(dB)↑
Forward91.487.6738.42
Backward91.487.6738.29
Bidirection84.118.1639.03
+ +Table 8: Ablation studies to verify autoregressive paradigm. + +
ModelParams(KB)↓Times(s)↓PSNR(dB)↑
w/o AR80.567.8338.62
w/ AR84.118.1639.03
+ +denoising performance by integrating preceding and succeeding spectral information, effectively capturing long-range spectral dependencies. + +Autoregressive Paradigm. Current per-band denoising methods restore each band independently using spectral information from adjacent bands, but overlook leveraging spectral similarity constraints. Our autoregressive framework addresses this by incorporating information from previously generated bands to enhance the reconstruction of the current band, enabling more effective learning, maintaining consistency across bands, and reducing noise impact. Without the autoregressive paradigm, Equation 9 simplifies to: + +$$ +\arg \max \prod_ {i = 1} ^ {B} P \left(x _ {i} \mid h _ {1} ^ {e}, \dots , h _ {B} ^ {e}\right). \tag {14} +$$ + +We make an ablation study in Tab. 8. It is evident that adopting the autoregressive paradigm lead to overall better denoising performance for our model. + +# 5 Conclusion + +We propose RAS2S, a simple yet effective framework for HSI denoising, leveraging the power of Seq2Seq learning. We regard HSI denoising as a noisy-to-clean sequence translation problem, and maximize the denoising potential of the Seq2Seq model through discriminative semantic region learning, utilizing insights from the intrinsic spatial-spectral properties of HSI. RAS2S uses a simple bidirectional encoder-decoder architecture and serially models spatial-spectral correlations without the need for dataset-specific training, presenting the huge potential for capturing long-range dependencies. More about, RAS2S effectively handles the diverse spatial-spectral semantic variations, delivering more precise denoising results. Extensive experiments prove RAS2S outperforms other mainstream methods and achieves state-of-the-art on prevailing benchmark datasets with few parameters. + +# Acknowledgements + +This work was supported by National Key R&D Program of China under Grant No. 2021ZD0110400, the Fundamental Research Funds for the Central Universities No. xxj032023020, and sponsored by the CAAI-MindSpore Open Fund, developed on OpenI Community. + +# References + +1. Arad, B., Ben-Shahar, O.: Sparse recovery of hyperspectral signal from natural rgb images. In: ECCV (2016) 9, 10 +2. Bai, Y., Zhao, Z., Gong, Y., Wei, X.: Artrackv2: Prompting autoregressive tracker where to look and how to describe. In: CVPR (2024) 2 +3. Ballas, N., Yao, L., Pal, C., Courville, A.: Delving deeper into convolutional networks for learning video representations. In: ICLR (2016) 5 +4. Bodrito, T., Zouaoui, A., Chanussot, J., Mairal, J.: A trainable spectral-spatial sparse coding model for hyperspectral image restoration. In: NeurIPS (2021) 9, 10, 11, 12 +5. Cai, Y., Lin, J., Hu, X., Wang, H., Yuan, X., Zhang, Y., Timofte, R., Van Gool, L.: Mask-guided spectral-wise transformer for efficient hyperspectral image reconstruction. In: CVPR (2022) 2, 10 +6. Cao, X., Fu, X., Xu, C., Meng, D.: Deep spatial-spectral global reasoning network for hyperspectral image denoising. IEEE TGRS 60, 1-14 (2021) 9, 10, 12 +7. Chang, Y., Yan, L., Fang, H., Luo, C.: Anisotropic spectral-spatial total variation model for multispectral remote sensing image destriping. IEEE TIP 24(6), 1852-1866 (2015) 4 +8. Chang, Y., Yan, L., Zhong, S.: Hyper-laplacian regularized unidirectional low-rank tensor recovery for multispectral image denoising. In: CVPR (2017) 3, 10, 12 +9. Chen, H., Zhao, W., Xu, T., Shi, G., Zhou, S., Liu, P., Li, J.: Spectral-wise implicit neural representation for hyperspectral image reconstruction. IEEE TCSVT (2023) 4 +0. Chen, J., Wang, X., Guo, Z., Zhang, X., Sun, J.: Dynamic region-aware convolution. In: CVPR (2021) 7 +1. Choi, I., Kim, M., Gutierrez, D., Jeon, D., Nam, G.: High-quality hyperspectral reconstruction using a spectral prior. ACM TOG 36, 13 (2017) 9, 10 +2. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. In: ICLR (2020) 6 +3. Fu, G., Xiong, F., Lu, J., Zhou, J., Qian, Y.: Nonlocal spatial-spectral neural network for hyperspectral image denoising. IEEE TGRS 60, 1-16 (2022) 2 +4. Gamba, P.: A collection of data for urban area characterization. In: IGARSS (2004) 9, 10 +5. Guan, J., Lai, R., Li, H., Yang, Y., Gu, L.: Dnrcnn: Deep recurrent convolutional neural network for hsi destriping. IEEE TNNLS (2022) 4 +6. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: CVPR (2016) 6 +7. He, W., Yao, Q., Li, C., Yokoya, N., Zhao, Q.: Non-local meets global: An integrated paradigm for hyperspectral denoising. In: CVPR (2019) 3, 10, 12 +8. He, W., Zhang, H., Shen, H., Zhang, L.: Hyperspectral image denoising using local low-rank matrix recovery and global spatial-spectral total variation. IEEE J-STARS 11(3), 713-729 (2018) 10, 12 +9. Hu, J., Shen, L., Sun, G.: Squeeze-and-excitation networks. In: CVPR (2018) 8 +20. Huang, Z., Zhang, S., Pan, L., Qing, Z., Tang, M., Liu, Z., Ang Jr, M.H.: Tada! temporally-adaptive convolutions for video understanding. In: ICLR (2022) 8 +21. Keneshloo, Y., Shi, T., Ramakrishnan, N., Reddy, C.K.: Deep reinforcement learning for sequence-to-sequence models. IEEE TNNLS 31(7), 2469-2489 (2019) 2 + +22. Kingma, D.P., Ba, J.: Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980 (2014) 9 +23. Lai, Z., Fu, Y.: Mixed attention network for hyperspectral image denoising. arXiv preprint arXiv:2301.11525 (2023) 10, 12 +24. Lai, Z., Wei, K., Fu, Y.: Deep plug-and-play prior for hyperspectral image restoration. Neurocomputing 481, 281-293 (2022) 10, 12 +25. Lai, Z., Yan, C., Fu, Y.: Hybrid spectral denoising transformer with guided attention. In: ICCV (2023) 10, 12 +26. Li, J., Zhang, J., Chen, F., Zhao, K., Zeng, D.: Adaptive material matching for hyperspectral imagery destringing. IEEE TGRS 60, 1-20 (2022) 1 +27. Li, M., Fu, Y., Zhang, Y.: Spatial-spectral transformer for hyperspectral image denoising. In: AAAI (2023) 2, 4, 9, 10, 11, 12 +28. Li, M., Liu, J., Fu, Y., Zhang, Y., Dou, D.: Spectral enhanced rectangle transformer for hyperspectral image denoising. In: CVPR (2023) 2, 4, 9, 10, 11, 12 +29. Lin, J., Hu, X., Cai, Y., Wang, H., Yan, Y., Zou, X., Zhang, Y., Van Gool, L.: Unsupervised flow-aligned sequence-to-sequence learning for video restoration. In: ICML (2022) 4 +30. Liu, T., Wang, K., Sha, L., Chang, B., Sui, Z.: Table-to-text generation by structure-aware seq2seq learning. In: AAAI (2018) 6 +31. Lu, G., Fei, B.: Medical hyperspectral imaging: a review. Journal of biomedical optics 19(1), 010901-010901 (2014) 1 +32. Maggioni, M., Katkovnik, V., Egiazarian, K., Foi, A.: Nonlocal transform-domain filter for volumetric data denoising and reconstruction. IEEE TIP 22(1), 119-133 (2012) 2 +33. Mnih, V., Hinton, G.E.: Learning to detect roads in high-resolution aerial images. In: ECCV (2010) 9, 11, 12 +34. Park, J.I., Lee, M.H., Grossberg, M.D., Nayar, S.K.: Multispectral imaging using multiplexed illumination. In: ICCV (2007) 2, 9, 10, 12 +35. Peng, J., Xie, Q., Zhao, Q., Wang, Y., Yee, L., Meng, D.: Enhanced 3dtv regularization and its applications on hsi denoising and compressed sensing. IEEE TIP 29, 7889-7903 (2020) 4 +36. Shi, Q., Tang, X., Yang, T., Liu, R., Zhang, L.: Hyperspectral image denoising using a 3-d attention denoising network. IEEE TGRS 59(12), 10348-10363 (2021) 4 +37. Sutskever, I., Vinyals, O., Le, Q.V.: Sequence to sequence learning with neural networks. In: NeurIPS (2014) 2, 4 +38. Wang, H., Xu, K., Lau, R.W.: Local color distributions prior for image enhancement. In: ECCV (2022) 8 +39. Wang, W., Jiao, W., Hao, Y., Wang, X., Shi, S., Tu, Z., Lyu, M.: Understanding and improving sequence-to-sequence pretraining for neural machine translation. ACL (2022) 4 +40. Wang, Y., Peng, J., Zhao, Q., Leung, Y., Zhao, X.L., Meng, D.: Hyperspectral image restoration via total variation regularized low-rank tensor decomposition. IEEE J-STARS 11(4), 1227-1243 (2017) 3, 10, 12 +41. Wang, Z., Ng, M.K., Zhuang, L., Gao, L., Zhang, B.: Nonlocal self-similarity-based hyperspectral remote sensing image denoising with 3-d convolutional neural network. IEEE TGRS 60, 1-17 (2022) 2 +42. Wang, Z., Bovik, A.C., Sheikh, H.R., Simoncelli, E.P.: Image quality assessment: from error visibility to structural similarity. IEEE TIP 13(4), 600-612 (2004) 9 +43. Wei, K., Fu, Y., Huang, H.: 3-d quasi-recurrent neural network for hyperspectral image denoising. IEEE TNNLS 32(1), 363-375 (2020) 2, 4, 9, 10, 12 + +44. Wei, X., Xiao, J., Gong, Y.: Blind hyperspectral image denoising with degradation information learning. Remote Sensing 15(2), 490 (2023) 4 +45. Wu, Z., Wang, Q., Wu, Z., Shen, Y.: Total variation-regularized weighted nuclear norm minimization for hyperspectral image mixed denoising. JEE 25(1), 013037-013037 (2016) 9, 10, 11 +46. Xiao, J., Ji, Y., Wei, X.: Hyperspectral image denoising with spectrum alignment. In: ACMMM. pp. 5495-5503 (2023) 2 +47. Xiao, J., Wei, X.: Hyperspectral image denoising using uncertainty-aware adjustor. In: IJCAI (2023) 2, 4 +48. Xiong, F., Zhou, J., Qian, Y.: Material based object tracking in hyperspectral videos. IEEE TGRS 29, 3719-3733 (2020) 1 +49. Xiong, F., Zhou, J., Zhao, Q., Lu, J., Qian, Y.: Mac-net: Model-aided nonlocal neural network for hyperspectral image denoising. IEEE TGRS 60, 1-14 (2021) 10, 11, 12 +50. Xu, S., Zhang, X., Wu, Y., Wei, F.: Sequence level contrastive learning for text summarization. In: AAAI (2022) 4 +51. Yuan, Q., Zhang, Q., Li, J., Shen, H., Zhang, L.: Hyperspectral image denoising employing a spatial-spectral deep residual convolutional neural network. IEEE TGRS 57(2), 1205-1218 (2018) 2, 4 +52. Yuan, Y., Ma, H., Liu, G.: Partial-dnet: A novel blind denoising model with noise intensity estimation for hsi. IEEE TGRS 60, 1-13 (2021) 2, 4 +53. Yuhas, R.H., Boardman, J.W., Goetz, A.F.: Determination of semi-arid landscape endmembers and seasonal trends using convex geometry spectral unmixing techniques. In: JPL (1993) 9 +54. Zhang, J., Tu, Z., Yang, J., Chen, Y., Yuan, J.: Mixste: Seq2seq mixed spatiotemporal encoder for 3d human pose estimation in video. In: CVPR (2022) 4 +55. Zhang, K., Zuo, W., Chen, Y., Meng, D., Zhang, L.: Beyond a gaussian denoiser: Residual learning of deep cnn for image denoising. IEEE TIP 26(7), 3142-3155 (2017) 7 +56. Zhang, Q., Zheng, Y., Yuan, Q., Song, M., Yu, H., Xiao, Y.: Hyperspectral image denoising: From model-driven, data-driven, to model-data-driven. IEEE TNNLS (2023) 1 +57. Zhang, T., Fu, Y., Li, C.: Hyperspectral image denoising with realistic data. In: ICCV (2021) 9, 11, 12 +58. Zhou, S., Zhang, J., Zuo, W., Loy, C.C.: Cross-scale internal graph neural network for image super-resolution. In: NeurIPS. vol. 33, pp. 3499-3509 (2020) 2 +59. Zhu, D., Du, B., Zhang, L.: Two-stream convolutional networks for hyperspectral target detection. IEEE TGRS 59(8), 6907-6921 (2020) 1 +60. Zhuang, L., Gao, L., Zhang, B., Fu, X., Bioucas-Dias, J.M.: Hyperspectral image denoising and anomaly detection based on low-rank and sparse representations. IEEE TGRS 60, 1-17 (2020) 1 \ No newline at end of file diff --git a/2024/Region-Aware Sequence-to-Sequence Learning for Hyperspectral Denoising/images.zip b/2024/Region-Aware Sequence-to-Sequence Learning for Hyperspectral Denoising/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..a5fa035a202c98e37b9cd17e085656e698f8609e --- /dev/null +++ b/2024/Region-Aware Sequence-to-Sequence Learning for Hyperspectral Denoising/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:27db59861adbfc2dcb6744d00fe2285930535035fb8f25ba56b81df0418a8953 +size 838619 diff --git a/2024/Region-Aware Sequence-to-Sequence Learning for Hyperspectral Denoising/layout.json b/2024/Region-Aware Sequence-to-Sequence Learning for Hyperspectral Denoising/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..9f84cf5e59ab8afa52018c117ebdc98aad2c9754 --- /dev/null +++ b/2024/Region-Aware Sequence-to-Sequence Learning for Hyperspectral Denoising/layout.json @@ -0,0 +1,10421 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 134, + 111, + 479, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 111, + 479, + 148 + ], + "spans": [ + { + "bbox": [ + 134, + 111, + 479, + 148 + ], + "type": "text", + "content": "Region-Aware Sequence-to-Sequence Learning for Hyperspectral Denoising" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 210, + 168, + 403, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 210, + 168, + 403, + 180 + ], + "spans": [ + { + "bbox": [ + 210, + 168, + 403, + 180 + ], + "type": "text", + "content": "Jiahua Xiao, Yang Liu, and Xing Wei*" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 147, + 190, + 465, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 147, + 190, + 465, + 213 + ], + "spans": [ + { + "bbox": [ + 147, + 190, + 465, + 213 + ], + "type": "text", + "content": "School of Software Engineering, Xi'an Jiaotong University {xjh847286495,1228754216}@stu.xjtu.edu.cn, weixing@mail.xjtu.edu.cn" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 160, + 237, + 455, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 237, + 455, + 469 + ], + "spans": [ + { + "bbox": [ + 160, + 237, + 455, + 469 + ], + "type": "text", + "content": "Abstract. Proper spectral modeling within hyperspectral image (HSI) is critical yet highly challenging for HSI denoising. In contrast to existing methods that struggle between effectiveness and flexibility in modeling long-range spectral dependencies and directly explore spatial-spectral information without region discrimination, we introduce RAS2S—a simple yet effective sequence-to-sequence (Seq2Seq) learning framework for better HSI denoising. RAS2S treats HSI denoising as a Seq2Seq translation problem, which converts the noisy spectral sequence to its clean ones in an autoregressive fashion. In addition, spatial-spectral information exploration without region discrimination contradicts the intrinsic spatial-spectral diversity of HSIs, leading to negative interference from spatial-spectral unrelated regions. Thus we propose a novel spatial-spectral region-aware module to distinctively perceive the semantic regions with different spatial-spectral representations, maximizing the spectral modeling potential of Seq2Seq learning. With such an improved Seq2Seq learning paradigm, RAS2S not only shows huge potential in capturing long-range spectral dependencies, but also maintains the flexibility to handle diverse HSIs. Extensive experiments demonstrate that RAS2S outperforms existing state-of-the-art methods quantitatively and qualitatively with a minimal model size, merely 0.08M. Source code is available at https://github.com/MIV-XJTU/RAS2S." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 160, + 478, + 453, + 501 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 478, + 453, + 501 + ], + "spans": [ + { + "bbox": [ + 160, + 478, + 453, + 501 + ], + "type": "text", + "content": "Keywords: Hyperspectral Denoising " + }, + { + "bbox": [ + 160, + 478, + 453, + 501 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 478, + 453, + 501 + ], + "type": "text", + "content": " Spectral Dependency " + }, + { + "bbox": [ + 160, + 478, + 453, + 501 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 478, + 453, + 501 + ], + "type": "text", + "content": " Sequence-to-Sequence Learning " + }, + { + "bbox": [ + 160, + 478, + 453, + 501 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 478, + 453, + 501 + ], + "type": "text", + "content": " Region Discrimination" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 518, + 230, + 532 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 518, + 230, + 532 + ], + "spans": [ + { + "bbox": [ + 132, + 518, + 230, + 532 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 541, + 482, + 650 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 541, + 482, + 650 + ], + "spans": [ + { + "bbox": [ + 130, + 541, + 482, + 650 + ], + "type": "text", + "content": "Unlike conventional images (e.g., grayscale or RGB), hyperspectral images (HSIs) capture the spectral attributes of various objects and materials within a static spatial scenario, widely applied in medical diagnosis [31], material identification [26, 48] and detection [59, 60]. However, various noise types introduced in the imaging process severely disrupt both the visual perception of HSIs and their aforementioned applications, making HSI denoising a crucial yet challenging task. Thanks to the unique spectral information inherited in HSIs [56], which reflects the inner properties of objects, combining spatial and spectral learning allows for more effective noise identification and HSI restoration." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 134, + 654, + 279, + 666 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 654, + 279, + 666 + ], + "spans": [ + { + "bbox": [ + 134, + 654, + 279, + 666 + ], + "type": "text", + "content": "* Corresponding author: Xing Wei." + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 289, + 283 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 289, + 283 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 289, + 283 + ], + "type": "text", + "content": "In the deep learning era, mainstream CNN-based methods [46, 47, 51, 52] jointly extract spatial-spectral information using 2D convolutional filters over a short window of adjacent spectral bands. However, the limited spectral receptive field severely constrains the utilization of global spectral information. While 3D convolution [13, 41, 43] provides a broader spatial-spectral perspective, it introduces significantly increased complexity. More recently, transformer-based" + } + ] + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 297, + 121, + 478, + 230 + ], + "blocks": [ + { + "bbox": [ + 297, + 121, + 478, + 230 + ], + "lines": [ + { + "bbox": [ + 297, + 121, + 478, + 230 + ], + "spans": [ + { + "bbox": [ + 297, + 121, + 478, + 230 + ], + "type": "image", + "image_path": "c9cd5ce64a57d1b1bcdb8d00aa2e31ce0e0c191d40ebafc78053cf0060559bb8.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 293, + 239, + 482, + 274 + ], + "lines": [ + { + "bbox": [ + 293, + 239, + 482, + 274 + ], + "spans": [ + { + "bbox": [ + 293, + 239, + 482, + 274 + ], + "type": "text", + "content": "Fig. 1: Denoising performance vs. inference time (CPU) under CAVE [34]. The circle areas are proportional to the model size." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 284, + 482, + 344 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 284, + 482, + 344 + ], + "spans": [ + { + "bbox": [ + 130, + 284, + 482, + 344 + ], + "type": "text", + "content": "methods [5,27,28] leverage self-attention mechanisms to capture complex spatial-spectral relations and effectively remove noise within HSIs. However, they are unable to handle HSIs with arbitrary spectral numbers and need dataset-specific training. As a result, there seems to be a dilemma between the sufficiency and flexibility of exploring long-range spectral dependencies." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 349, + 482, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 349, + 482, + 564 + ], + "spans": [ + { + "bbox": [ + 130, + 349, + 482, + 564 + ], + "type": "text", + "content": "To solve this dilemma, we consider continuous spectral bands as a sequence and introduce a novel perspective of treating HSI denoising as a sequence translation problem. We attempt to leverage the proven capability of sequence-to-sequence (Seq2Seq) models in sequence modeling [2,21,37] to tackle this problem, which effectively capture long-range dependencies with fewer parameters (shown in Fig. 1) and spectral flexibility. However, directly applying existing Seq2Seq frameworks to spectral sequences poses challenges: (i) Existing Seq2Seq models generate sequences in a forward order (e.g., text and video sequences), hindering their application to spectral sequences, which can utilize both preceding and succeeding bands. To address this, we employ a bidirectional encoder-decoder architecture tailored for spatial-spectral modeling, where the encoder processes noisy bands to hidden states, and the decoder utilizes these states to dynamically generate denoised spectral sequences autoregressively, capturing long-range contextual dependencies bidirectionally. (ii) Existing Seq2Seq models ignore the information gap between adjacent hidden states and the current input, hindering effective hidden state aggregation. We mitigate this by introducing a hidden state compensation module that adjusts adjacent hidden states with cross-attention, alleviating the adverse effects of the information gap." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": "In addition, HSIs inherently contain diverse semantic variations across spatial and spectral domains. However, current mainstream convolutional methods perform information extraction using filter sharing across the spatial-spectral domain, treating the feature of each position without discrimination, which will introduces negative interference from semantically unrelated regions during parameter updates. To address this, leveraging insights from spatial non-local self-similarity [32, 58] and spectral distinctiveness (see Fig. 2), we propose discriminative semantic region learning implemented through a Spatial Spectral" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "type": "text", + "content": "J. Xiao et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 176 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 176 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 176 + ], + "type": "text", + "content": "Region-Aware (SSRA) module for more focused exploration of spatial-spectral characteristics. SSRA dynamically allocates learnable filters to corresponding semantic regions by utilizing encoded hidden states rich in spatial-spectral aggregation information to generate region mask predictions in both spatial and spectral branches." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 177, + 274, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 177, + 274, + 426 + ], + "spans": [ + { + "bbox": [ + 130, + 177, + 274, + 426 + ], + "type": "text", + "content": "The region-guided masks divide the input into distinct regions, with each region assigned a dedicated filter, eliminating irrelevant spatial-spectral interference and maximizing spectral modeling potential of Seq2Seq learning. Thus, the Region-Aware Sequence-to-Sequence (RAS2S) learning framework is proposed for HSI denoising task. Benefiting from this tailored Seq2Seq architecture, our approach effectively captures long-range spectral dependencies for arbitrary spectral numbers, and eliminates the need for dataset-specific training. Overall, this work makes the following contributions:" + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 281, + 180, + 480, + 296 + ], + "blocks": [ + { + "bbox": [ + 281, + 180, + 480, + 296 + ], + "lines": [ + { + "bbox": [ + 281, + 180, + 480, + 296 + ], + "spans": [ + { + "bbox": [ + 281, + 180, + 480, + 296 + ], + "type": "image", + "image_path": "beed88671b24c2a3fa22a12c742fb300be28e85b03ceaddee9a7147fcf21b36e.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 279, + 303, + 482, + 425 + ], + "lines": [ + { + "bbox": [ + 279, + 303, + 482, + 425 + ], + "spans": [ + { + "bbox": [ + 279, + 303, + 482, + 425 + ], + "type": "text", + "content": "Fig.2: Illustration of spectral distinctiveness. Spatial positions with similar materials have similar spectral reflection patterns (e.g., Position A and B), while those with different materials show divergent patterns (e.g., Position C and D). This motivates assigning learnable filters to corresponding regions with similar spatial-spectral semantic representations, enabling each region to more effectively focus on its own vital spatial-spectral characteristics. We term this discriminative semantic region learning." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 439, + 480, + 548 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 138, + 439, + 479, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 439, + 479, + 475 + ], + "spans": [ + { + "bbox": [ + 138, + 439, + 479, + 475 + ], + "type": "text", + "content": "- We introduce RAS2S, the first framework to reformulate the HSI denoising problem from a sequence-to-sequence learning perspective, tailored for spectral sequence modeling." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 475, + 480, + 512 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 475, + 480, + 512 + ], + "spans": [ + { + "bbox": [ + 138, + 475, + 480, + 512 + ], + "type": "text", + "content": "- We first propose the discriminative semantic region learning within spatial-spectral domains and implement it through a novel SSRA module to maximize the denoising potential of the Seq2Seq framework." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 512, + 480, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 512, + 480, + 548 + ], + "spans": [ + { + "bbox": [ + 138, + 512, + 480, + 548 + ], + "type": "text", + "content": "- Extensive experiments on prevailing benchmarks demonstrate the superiority and generality of RAS2S, achieving state-of-the-art performance with few parameters." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 132, + 568, + 237, + 580 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 568, + 237, + 580 + ], + "spans": [ + { + "bbox": [ + 132, + 568, + 237, + 580 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 595, + 320, + 608 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 595, + 320, + 608 + ], + "spans": [ + { + "bbox": [ + 132, + 595, + 320, + 608 + ], + "type": "text", + "content": "2.1 Hyperspectral Image Denoising" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "Traditional methods rely on handcrafted priors based on hyperspectral image (HSI) statistics and often struggle when faced with noise distributions inconsistent with prior assumptions. For instance, approaches like [8, 17, 40] capture global spatial-spectral correlation, effectively leveraging low-rank properties and" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 255, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 255, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 255, + 91, + 448, + 102 + ], + "type": "text", + "content": "Sequence-to-Sequence Hyperspectral Denoising" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 295 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 295 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 295 + ], + "type": "text", + "content": "inherent smoothness in HSIs. Others, such as [7,35], utilize total variation priors to preserve marginal features effectively. In the era of deep learning, 2D CNN-based methods like [47,51,52] jointly extract spatial-spectral information through a short window of adjacent spectral bands, constrained by their spectral receptive field. Meanwhile, 3D CNN-based methods like [36,43,44] capture spatial-spectral dependencies but with increased computational cost. RNN-based approaches like [15] explore sequential spectrum analysis but struggle with long-range spatial-spectral dependencies. Recently, transformer-based models such as [9,27,28], leveraging self-attention mechanisms, show promising denoising performance. However, the extensive interaction among global tokens and the need for fixed-length spectral bands pose computational and flexibility challenges for modeling long-range spectral dependencies. In contrast, our model, a simple serial encoder-decoder architecture, effectively captures long-range spectral dependencies and handles HSIs with varying spectral numbers at a minimal model size." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 313, + 320, + 326 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 313, + 320, + 326 + ], + "spans": [ + { + "bbox": [ + 131, + 313, + 320, + 326 + ], + "type": "text", + "content": "2.2 Sequence-to-Sequence Learning" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 333, + 482, + 477 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 333, + 482, + 477 + ], + "spans": [ + { + "bbox": [ + 130, + 333, + 482, + 477 + ], + "type": "text", + "content": "Sequence-to-Sequence (Seq2Seq) learning [37] is a pivotal framework in natural language processing (NLP) and has shown significant success in various tasks like machine translation [39] and text summarization [50]. These models use an encoder to process input sequences and a decoder to generate output sequences. They evolve with the introduction of attention mechanisms, allowing them to focus on specific parts of the input when generating output. Recently, there are a growing trend in leveraging Seq2Seq models for various computer vision tasks, such as video restoration [29] and pose estimation [54]. Unlike natural language paragraphs and video frame sequences, which follow a temporal order, hyperspectral bands can utilize both preceding and succeeding bands. Tailored modifications are necessary to achieve effective spatial-spectral modeling and maximize the potential of Seq2Seq models for HSI denoising." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 495, + 202, + 507 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 495, + 202, + 507 + ], + "spans": [ + { + "bbox": [ + 131, + 495, + 202, + 507 + ], + "type": "text", + "content": "3 Method" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 520, + 482, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 520, + 482, + 593 + ], + "spans": [ + { + "bbox": [ + 130, + 520, + 482, + 593 + ], + "type": "text", + "content": "In the subsequent sections, we represent the pairs of noisy and clean HSI as " + }, + { + "bbox": [ + 130, + 520, + 482, + 593 + ], + "type": "inline_equation", + "content": "Y = \\{y_{i}\\}_{1}^{B}" + }, + { + "bbox": [ + 130, + 520, + 482, + 593 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 520, + 482, + 593 + ], + "type": "inline_equation", + "content": "X = \\{x_{i}\\}_{1}^{B}" + }, + { + "bbox": [ + 130, + 520, + 482, + 593 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 520, + 482, + 593 + ], + "type": "inline_equation", + "content": "y_{i}" + }, + { + "bbox": [ + 130, + 520, + 482, + 593 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 520, + 482, + 593 + ], + "type": "inline_equation", + "content": "x_{i}" + }, + { + "bbox": [ + 130, + 520, + 482, + 593 + ], + "type": "text", + "content": " denote individual spectral bands, and " + }, + { + "bbox": [ + 130, + 520, + 482, + 593 + ], + "type": "inline_equation", + "content": "Y, X \\in \\mathbb{R}^{H \\times W \\times B}" + }, + { + "bbox": [ + 130, + 520, + 482, + 593 + ], + "type": "text", + "content": ". Here, " + }, + { + "bbox": [ + 130, + 520, + 482, + 593 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 130, + 520, + 482, + 593 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 520, + 482, + 593 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 130, + 520, + 482, + 593 + ], + "type": "text", + "content": " represent the spatial height and width, respectively, while " + }, + { + "bbox": [ + 130, + 520, + 482, + 593 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 130, + 520, + 482, + 593 + ], + "type": "text", + "content": " indicates the number of spectral bands in the HSI. Therefore, the task of HSI denoising is formulated to maximize the conditional probability " + }, + { + "bbox": [ + 130, + 520, + 482, + 593 + ], + "type": "inline_equation", + "content": "P(X|Y)" + }, + { + "bbox": [ + 130, + 520, + 482, + 593 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 609, + 359, + 621 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 609, + 359, + 621 + ], + "spans": [ + { + "bbox": [ + 131, + 609, + 359, + 621 + ], + "type": "text", + "content": "3.1 HSI Denoising as Sequence Translation" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "type": "text", + "content": "We treat HSI denoising as a sequence translation problem and introduce RAS2S. However, existing Seq2Seq models cannot be directly transferred to HSI processing tasks. To effectively models the long-range spectral dependencies, our model" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "type": "text", + "content": "J. Xiao et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 140, + 114, + 471, + 300 + ], + "blocks": [ + { + "bbox": [ + 140, + 114, + 471, + 300 + ], + "lines": [ + { + "bbox": [ + 140, + 114, + 471, + 300 + ], + "spans": [ + { + "bbox": [ + 140, + 114, + 471, + 300 + ], + "type": "image", + "image_path": "7a2dbca4de80fa93412b54dcbf40c4194c754ef5f4c01780dc403d8582bc0cf8.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 308, + 480, + 354 + ], + "lines": [ + { + "bbox": [ + 131, + 308, + 480, + 354 + ], + "spans": [ + { + "bbox": [ + 131, + 308, + 480, + 354 + ], + "type": "text", + "content": "Fig. 3: RAS2S architecture. The workflow of the proposed RAS2S is depicted in (a), which models long-range spectral dependencies with a tailored bidirectional encoder-decoder architecture. The modules at the bottom display the internal details of (b) hidden state compensation (HFComp) and (c) adjacent attention mechanism (AAM)." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 377, + 482, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 377, + 482, + 471 + ], + "spans": [ + { + "bbox": [ + 130, + 377, + 482, + 471 + ], + "type": "text", + "content": "uses the idea of bidirectional RNN to capture dependencies in both preceding bands " + }, + { + "bbox": [ + 130, + 377, + 482, + 471 + ], + "type": "inline_equation", + "content": "\\{y_{i}\\}_{1}^{i - 1}" + }, + { + "bbox": [ + 130, + 377, + 482, + 471 + ], + "type": "text", + "content": " and succeeding bands " + }, + { + "bbox": [ + 130, + 377, + 482, + 471 + ], + "type": "inline_equation", + "content": "\\{y_{i}\\}_{i + 1}^{B}" + }, + { + "bbox": [ + 130, + 377, + 482, + 471 + ], + "type": "text", + "content": ", and then outputs the bidirectional propagated hidden state sets " + }, + { + "bbox": [ + 130, + 377, + 482, + 471 + ], + "type": "inline_equation", + "content": "\\left(\\left\\{h_i^{fw}\\right\\}_1^B\\right." + }, + { + "bbox": [ + 130, + 377, + 482, + 471 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 377, + 482, + 471 + ], + "type": "inline_equation", + "content": "\\left\\{h_i^{bw}\\right\\}_1^B" + }, + { + "bbox": [ + 130, + 377, + 482, + 471 + ], + "type": "text", + "content": ". For brevity and clarity, we take the forward propagation process as an example to elaborate the workflow of RAS2S in the following section, which can be easily extended to a bidirectional process. As presented in Fig. 3(a), RAS2S consists of the following key components." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 481, + 482, + 620 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 481, + 482, + 620 + ], + "spans": [ + { + "bbox": [ + 130, + 481, + 482, + 620 + ], + "type": "text", + "content": "Encoder. In forward propagation process, the goal of encoder is to transform the shallow feature " + }, + { + "bbox": [ + 130, + 481, + 482, + 620 + ], + "type": "inline_equation", + "content": "\\{f_i\\}_1^B" + }, + { + "bbox": [ + 130, + 481, + 482, + 620 + ], + "type": "text", + "content": " of each band into the hidden state set " + }, + { + "bbox": [ + 130, + 481, + 482, + 620 + ], + "type": "inline_equation", + "content": "\\left\\{h_i^{fw - e}\\right\\}_1^B \\in \\mathbb{R}^{H \\times W \\times C}" + }, + { + "bbox": [ + 130, + 481, + 482, + 620 + ], + "type": "text", + "content": ". In order to alleviate the information gap between preceding hidden features " + }, + { + "bbox": [ + 130, + 481, + 482, + 620 + ], + "type": "inline_equation", + "content": "h_{i-1}^{fw - e}" + }, + { + "bbox": [ + 130, + 481, + 482, + 620 + ], + "type": "text", + "content": " and current visual features " + }, + { + "bbox": [ + 130, + 481, + 482, + 620 + ], + "type": "inline_equation", + "content": "f_i" + }, + { + "bbox": [ + 130, + 481, + 482, + 620 + ], + "type": "text", + "content": ", we propose to make the Hidden State Compensation (HFComp) for " + }, + { + "bbox": [ + 130, + 481, + 482, + 620 + ], + "type": "inline_equation", + "content": "h_{i-1}^{fw - e}" + }, + { + "bbox": [ + 130, + 481, + 482, + 620 + ], + "type": "text", + "content": " to obtain an adjusted hidden state " + }, + { + "bbox": [ + 130, + 481, + 482, + 620 + ], + "type": "inline_equation", + "content": "h_{i-1}^{fw - ec}" + }, + { + "bbox": [ + 130, + 481, + 482, + 620 + ], + "type": "text", + "content": ". The HFComp will be introduced immediately. In implementation, we utilize the ConvGRU cell [3] to facilitate the learning of long-range spatial-spectral information, which combines " + }, + { + "bbox": [ + 130, + 481, + 482, + 620 + ], + "type": "inline_equation", + "content": "f_i" + }, + { + "bbox": [ + 130, + 481, + 482, + 620 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 481, + 482, + 620 + ], + "type": "inline_equation", + "content": "h_{i-1}^{fw - ec}" + }, + { + "bbox": [ + 130, + 481, + 482, + 620 + ], + "type": "text", + "content": " to generate the final output " + }, + { + "bbox": [ + 130, + 481, + 482, + 620 + ], + "type": "inline_equation", + "content": "h_i^{fw - e}" + }, + { + "bbox": [ + 130, + 481, + 482, + 620 + ], + "type": "text", + "content": ". Thus, the workflow of encoder for each spectral step can be formulated as:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 237, + 620, + 481, + 637 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 237, + 620, + 481, + 637 + ], + "spans": [ + { + "bbox": [ + 237, + 620, + 481, + 637 + ], + "type": "interline_equation", + "content": "h _ {i - 1} ^ {f w} ^ {- e c} = \\mathrm {H F C o m p} (f _ {i}, h _ {i - 1} ^ {f w} ^ {- e}), \\tag {1}", + "image_path": "dec8e0fa533666a4dd661d81efa49c0aec206dddc0b42a9e819421745d534324.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 235, + 650, + 481, + 667 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 235, + 650, + 481, + 667 + ], + "spans": [ + { + "bbox": [ + 235, + 650, + 481, + 667 + ], + "type": "interline_equation", + "content": "h _ {i} ^ {f w - e} = \\mathrm {C o n v G R U} (f _ {i}, h _ {i - 1} ^ {f w - e c}), \\tag {2}", + "image_path": "ef05fb83a90e700cca046d55454b84c28552a9a81c0ef3d5e0faff2caa5e9192.jpg" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 255, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 255, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 255, + 91, + 448, + 102 + ], + "type": "text", + "content": "Sequence-to-Sequence Hyperspectral Denoising" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 114, + 479, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 114, + 479, + 156 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 479, + 156 + ], + "type": "text", + "content": "where the backward propagation process also produces the hidden state " + }, + { + "bbox": [ + 130, + 114, + 479, + 156 + ], + "type": "inline_equation", + "content": "h_i^{bw - e}" + }, + { + "bbox": [ + 130, + 114, + 479, + 156 + ], + "type": "text", + "content": ". The final encoded hidden state " + }, + { + "bbox": [ + 130, + 114, + 479, + 156 + ], + "type": "inline_equation", + "content": "h_i^e" + }, + { + "bbox": [ + 130, + 114, + 479, + 156 + ], + "type": "text", + "content": " can be generated by using a 2D convolution layer to fuse the forward and backward hidden states (" + }, + { + "bbox": [ + 130, + 114, + 479, + 156 + ], + "type": "inline_equation", + "content": "h_i^{fw - e}" + }, + { + "bbox": [ + 130, + 114, + 479, + 156 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 114, + 479, + 156 + ], + "type": "inline_equation", + "content": "h_i^{bw - e}" + }, + { + "bbox": [ + 130, + 114, + 479, + 156 + ], + "type": "text", + "content": ")." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 163, + 479, + 202 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 163, + 479, + 202 + ], + "spans": [ + { + "bbox": [ + 131, + 163, + 479, + 202 + ], + "type": "text", + "content": "HFComp. In forward Encoder, HFComp (see Fig. 3(b)) adaptively adjusts " + }, + { + "bbox": [ + 131, + 163, + 479, + 202 + ], + "type": "inline_equation", + "content": "h_{i-1}^{f w - e}" + }, + { + "bbox": [ + 131, + 163, + 479, + 202 + ], + "type": "text", + "content": " through calculating the spectral cross-attention map [12] " + }, + { + "bbox": [ + 131, + 163, + 479, + 202 + ], + "type": "inline_equation", + "content": "\\operatorname{Att}(Q, K, V) \\in \\mathbb{R}^{C \\times C}" + }, + { + "bbox": [ + 131, + 163, + 479, + 202 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 131, + 163, + 479, + 202 + ], + "type": "inline_equation", + "content": "f_{i}" + }, + { + "bbox": [ + 131, + 163, + 479, + 202 + ], + "type": "text", + "content": ". The whole procedure of HFComp is formulated as:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 225, + 212, + 481, + 226 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 225, + 212, + 481, + 226 + ], + "spans": [ + { + "bbox": [ + 225, + 212, + 481, + 226 + ], + "type": "interline_equation", + "content": "\\operatorname {A t t} (Q, K, V) = V \\cdot \\operatorname {S o f t m a x} (K \\cdot Q), \\tag {3}", + "image_path": "19c9adadadd7e046cccedba82615554f79c6a736b63efbc5a99e1fb03a37f1b1.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 222, + 245, + 481, + 262 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 222, + 245, + 481, + 262 + ], + "spans": [ + { + "bbox": [ + 222, + 245, + 481, + 262 + ], + "type": "interline_equation", + "content": "h _ {i - 1} ^ {f w - e c} = \\operatorname {R B} (\\operatorname {A t t} (Q, K, V)) + h _ {i - 1} ^ {f w - e}, \\tag {4}", + "image_path": "048bdca94ef6fb4ce0749cc8520c610dff28b570eed1299b63dc78446c775a1e.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 268, + 479, + 320 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 268, + 479, + 320 + ], + "spans": [ + { + "bbox": [ + 130, + 268, + 479, + 320 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 268, + 479, + 320 + ], + "type": "inline_equation", + "content": "Q \\in \\mathbb{R}^{C \\times HW}" + }, + { + "bbox": [ + 130, + 268, + 479, + 320 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 268, + 479, + 320 + ], + "type": "inline_equation", + "content": "V \\in \\mathbb{R}^{HW \\times C}" + }, + { + "bbox": [ + 130, + 268, + 479, + 320 + ], + "type": "text", + "content": " are linearly projected from " + }, + { + "bbox": [ + 130, + 268, + 479, + 320 + ], + "type": "inline_equation", + "content": "h_{i-1}^{fw}^{-e}" + }, + { + "bbox": [ + 130, + 268, + 479, + 320 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 268, + 479, + 320 + ], + "type": "inline_equation", + "content": "K \\in \\mathbb{R}^{HW \\times C}" + }, + { + "bbox": [ + 130, + 268, + 479, + 320 + ], + "type": "text", + "content": " is projected from " + }, + { + "bbox": [ + 130, + 268, + 479, + 320 + ], + "type": "inline_equation", + "content": "f_i" + }, + { + "bbox": [ + 130, + 268, + 479, + 320 + ], + "type": "text", + "content": ". To improve the feature processing ability, the adjusted hidden features " + }, + { + "bbox": [ + 130, + 268, + 479, + 320 + ], + "type": "inline_equation", + "content": "h_{i-1}^{fw}^{-ec} \\in \\mathbb{R}^{H \\times W \\times C}" + }, + { + "bbox": [ + 130, + 268, + 479, + 320 + ], + "type": "text", + "content": " is obtained from the concatenated residual blocks (RBs [16])." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 331, + 479, + 378 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 331, + 479, + 378 + ], + "spans": [ + { + "bbox": [ + 131, + 331, + 479, + 378 + ], + "type": "text", + "content": "SSRA. The SSRA aim to generate the region-aware encoded hidden state " + }, + { + "bbox": [ + 131, + 331, + 479, + 378 + ], + "type": "inline_equation", + "content": "h_i^e - r" + }, + { + "bbox": [ + 131, + 331, + 479, + 378 + ], + "type": "text", + "content": " based on discriminative semantic region learning. The motivation and implementation of discriminative semantic region learning will be elaborated in subsection 3.2." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 391, + 480, + 493 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 391, + 480, + 493 + ], + "spans": [ + { + "bbox": [ + 131, + 391, + 480, + 493 + ], + "type": "text", + "content": "AAM. The attention mechanism is widely used in the Seq2Seq model for addressing the challenges of handling long sequences [30], facilitating the effective capture of crucial information in the input sequence. In the context of HSI denoising, we introduce the Adjacent Attention Mechanism (AAM, see Fig. 3(c)) to capture the spatial-spectral context information from the " + }, + { + "bbox": [ + 131, + 391, + 480, + 493 + ], + "type": "inline_equation", + "content": "2n + 1" + }, + { + "bbox": [ + 131, + 391, + 480, + 493 + ], + "type": "text", + "content": " adjusted adjacent encoded hidden states " + }, + { + "bbox": [ + 131, + 391, + 480, + 493 + ], + "type": "inline_equation", + "content": "\\{h_j^e\\}_{i - n}^{i + n}" + }, + { + "bbox": [ + 131, + 391, + 480, + 493 + ], + "type": "text", + "content": " according to generated decoded hidden state " + }, + { + "bbox": [ + 131, + 391, + 480, + 493 + ], + "type": "inline_equation", + "content": "h_{i - 1}^{fw - d}" + }, + { + "bbox": [ + 131, + 391, + 480, + 493 + ], + "type": "text", + "content": ". Thus, the context feature features " + }, + { + "bbox": [ + 131, + 391, + 480, + 493 + ], + "type": "inline_equation", + "content": "c_{i}" + }, + { + "bbox": [ + 131, + 391, + 480, + 493 + ], + "type": "text", + "content": " can be represented by the weighted sum of adjacent encoded hidden states:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 190, + 502, + 480, + 536 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 190, + 502, + 480, + 536 + ], + "spans": [ + { + "bbox": [ + 190, + 502, + 480, + 536 + ], + "type": "interline_equation", + "content": "c _ {i} = \\sum_ {j = i - n} ^ {i + n} \\lambda_ {i j} h _ {j} ^ {e}; \\quad \\left\\{\\lambda_ {i j} \\right\\} _ {i - n} ^ {i + n} = \\operatorname {R S} \\left(h _ {i - 1} ^ {f w} - ^ {d}, \\left\\{h _ {j} ^ {e} \\right\\} _ {i - n} ^ {i + n}\\right), \\tag {5}", + "image_path": "682ac5e694a14cad27ec7c96f93cdf5e3e8601a633ce0bf492976082d0564bec.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 545, + 479, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 545, + 479, + 582 + ], + "spans": [ + { + "bbox": [ + 131, + 545, + 479, + 582 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 545, + 479, + 582 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 131, + 545, + 479, + 582 + ], + "type": "text", + "content": " is the weight of adjacent encoded hidden states. " + }, + { + "bbox": [ + 131, + 545, + 479, + 582 + ], + "type": "inline_equation", + "content": "\\mathrm{RS}(\\cdot ,\\cdot)" + }, + { + "bbox": [ + 131, + 545, + 479, + 582 + ], + "type": "text", + "content": " means to calculate the relevant scores between two input features. In implementation, we use a 3D convolution layer for parameter learning as follows:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 179, + 590, + 480, + 608 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 179, + 590, + 480, + 608 + ], + "spans": [ + { + "bbox": [ + 179, + 590, + 480, + 608 + ], + "type": "interline_equation", + "content": "\\left\\{\\lambda_ {i _ {j}} \\right\\} _ {i - n} ^ {i + n} = \\operatorname {S o f t m a x} \\left(\\operatorname {T a n h} \\left(\\left[ h _ {i - 1} ^ {f w} - ^ {d}, \\left\\{h _ {j} ^ {e} \\right\\} _ {i - n} ^ {i + n} \\right]\\right)\\right) \\tag {6}", + "image_path": "5218840d1e8ac167d38b7264f9101cb5a55415a482581e1748c83233104c825b.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 131, + 617, + 402, + 630 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 617, + 402, + 630 + ], + "spans": [ + { + "bbox": [ + 131, + 617, + 402, + 630 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 617, + 402, + 630 + ], + "type": "inline_equation", + "content": "[\\cdot ,\\cdot ]" + }, + { + "bbox": [ + 131, + 617, + 402, + 630 + ], + "type": "text", + "content": " represents concatenation along channel dimension." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 131, + 641, + 479, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 641, + 479, + 665 + ], + "spans": [ + { + "bbox": [ + 131, + 641, + 479, + 665 + ], + "type": "text", + "content": "Decoder. For denoised sequence generation, We employ the identical structure of encoder in the decoder. It serially generate the subsequent output " + }, + { + "bbox": [ + 131, + 641, + 479, + 665 + ], + "type": "inline_equation", + "content": "x_{i}" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "type": "text", + "content": "J. Xiao et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 143, + 114, + 312, + 213 + ], + "blocks": [ + { + "bbox": [ + 143, + 114, + 312, + 213 + ], + "lines": [ + { + "bbox": [ + 143, + 114, + 312, + 213 + ], + "spans": [ + { + "bbox": [ + 143, + 114, + 312, + 213 + ], + "type": "image", + "image_path": "97b7ea5f00580126092794ab51b9430fa7ab3bcea17fcf69e5d0a6089682ca9b.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 143, + 213, + 312, + 270 + ], + "blocks": [ + { + "bbox": [ + 143, + 213, + 312, + 270 + ], + "lines": [ + { + "bbox": [ + 143, + 213, + 312, + 270 + ], + "spans": [ + { + "bbox": [ + 143, + 213, + 312, + 270 + ], + "type": "image", + "image_path": "b0f58510b1391f68e127259b9965a5d548e91e95e4d51efb01d1c734155af736.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 134, + 281, + 477, + 314 + ], + "lines": [ + { + "bbox": [ + 134, + 281, + 477, + 314 + ], + "spans": [ + { + "bbox": [ + 134, + 281, + 477, + 314 + ], + "type": "text", + "content": "Fig. 4: Illustrations of (a) spatial-spectral Fig. 5: Visualizations of spatial and region-aware module (SSRA) and (b) Spec- spectral guided masks on various HSIs. DynConv." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 323, + 114, + 371, + 163 + ], + "blocks": [ + { + "bbox": [ + 323, + 114, + 371, + 163 + ], + "lines": [ + { + "bbox": [ + 323, + 114, + 371, + 163 + ], + "spans": [ + { + "bbox": [ + 323, + 114, + 371, + 163 + ], + "type": "image", + "image_path": "27dc62866771648d55cf11df06f271c336786477d20c73d01b97f87830c3f520.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 373, + 114, + 421, + 164 + ], + "blocks": [ + { + "bbox": [ + 373, + 114, + 421, + 164 + ], + "lines": [ + { + "bbox": [ + 373, + 114, + 421, + 164 + ], + "spans": [ + { + "bbox": [ + 373, + 114, + 421, + 164 + ], + "type": "image", + "image_path": "a305c74d972c12ec2f2d84b2c4d25a380fdbbb1cce197c84c77b5d09fbd54747.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 423, + 114, + 470, + 164 + ], + "blocks": [ + { + "bbox": [ + 423, + 114, + 470, + 164 + ], + "lines": [ + { + "bbox": [ + 423, + 114, + 470, + 164 + ], + "spans": [ + { + "bbox": [ + 423, + 114, + 470, + 164 + ], + "type": "image", + "image_path": "9dd2788841eae821f33ea948f992111de7e48c48d10e0ae4d3ea227fac9e19d3.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 323, + 165, + 371, + 213 + ], + "blocks": [ + { + "bbox": [ + 323, + 165, + 371, + 213 + ], + "lines": [ + { + "bbox": [ + 323, + 165, + 371, + 213 + ], + "spans": [ + { + "bbox": [ + 323, + 165, + 371, + 213 + ], + "type": "image", + "image_path": "56be98780f07f198ea96eac8378dc2a60ecf10034c06ae8e5ebe384a7146949e.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 373, + 165, + 421, + 213 + ], + "blocks": [ + { + "bbox": [ + 373, + 165, + 421, + 213 + ], + "lines": [ + { + "bbox": [ + 373, + 165, + 421, + 213 + ], + "spans": [ + { + "bbox": [ + 373, + 165, + 421, + 213 + ], + "type": "image", + "image_path": "4c9d49dd654f9a8e34fd61437fbf79f5854e89d830f3b3805794389c23c44ef1.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 423, + 165, + 470, + 213 + ], + "blocks": [ + { + "bbox": [ + 423, + 165, + 470, + 213 + ], + "lines": [ + { + "bbox": [ + 423, + 165, + 470, + 213 + ], + "spans": [ + { + "bbox": [ + 423, + 165, + 470, + 213 + ], + "type": "image", + "image_path": "2f90c03edf6a84fa7181b8462e33249a1b88ca2edf9f6532e2712b1911c566e0.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 323, + 215, + 371, + 262 + ], + "blocks": [ + { + "bbox": [ + 323, + 215, + 371, + 262 + ], + "lines": [ + { + "bbox": [ + 323, + 215, + 371, + 262 + ], + "spans": [ + { + "bbox": [ + 323, + 215, + 371, + 262 + ], + "type": "image", + "image_path": "6bc6e7f408cffd22f9af8114c71b0a9f64299d5dac8ff9f39584cb6b32933097.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 325, + 264, + 369, + 270 + ], + "lines": [ + { + "bbox": [ + 325, + 264, + 369, + 270 + ], + "spans": [ + { + "bbox": [ + 325, + 264, + 369, + 270 + ], + "type": "text", + "content": "Pseudo Images" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 373, + 215, + 421, + 262 + ], + "blocks": [ + { + "bbox": [ + 373, + 215, + 421, + 262 + ], + "lines": [ + { + "bbox": [ + 373, + 215, + 421, + 262 + ], + "spans": [ + { + "bbox": [ + 373, + 215, + 421, + 262 + ], + "type": "image", + "image_path": "137960403d0b18262ed9e80c6f8152bff9b9702c574fcca6feca919dd1cdbf83.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 375, + 264, + 417, + 270 + ], + "lines": [ + { + "bbox": [ + 375, + 264, + 417, + 270 + ], + "spans": [ + { + "bbox": [ + 375, + 264, + 417, + 270 + ], + "type": "text", + "content": "Spatial Masks" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 423, + 215, + 470, + 262 + ], + "blocks": [ + { + "bbox": [ + 423, + 215, + 470, + 262 + ], + "lines": [ + { + "bbox": [ + 423, + 215, + 470, + 262 + ], + "spans": [ + { + "bbox": [ + 423, + 215, + 470, + 262 + ], + "type": "image", + "image_path": "377cd8775444d35532f836c13b12b2a7e6424eb59ef555584e9fcc387504937f.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 424, + 264, + 470, + 270 + ], + "lines": [ + { + "bbox": [ + 424, + 264, + 470, + 270 + ], + "spans": [ + { + "bbox": [ + 424, + 264, + 470, + 270 + ], + "type": "text", + "content": "Spectral Masks" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 131, + 337, + 479, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 337, + 479, + 361 + ], + "spans": [ + { + "bbox": [ + 131, + 337, + 479, + 361 + ], + "type": "text", + "content": "with residual learning [55] in an autoregressive manner, conditioned on context features " + }, + { + "bbox": [ + 131, + 337, + 479, + 361 + ], + "type": "inline_equation", + "content": "c_{i}" + }, + { + "bbox": [ + 131, + 337, + 479, + 361 + ], + "type": "text", + "content": ", the decoded hidden feature " + }, + { + "bbox": [ + 131, + 337, + 479, + 361 + ], + "type": "inline_equation", + "content": "h_{i}^{d}" + }, + { + "bbox": [ + 131, + 337, + 479, + 361 + ], + "type": "text", + "content": " and generated output " + }, + { + "bbox": [ + 131, + 337, + 479, + 361 + ], + "type": "inline_equation", + "content": "x_{i-1}" + }, + { + "bbox": [ + 131, + 337, + 479, + 361 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 253, + 369, + 480, + 383 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 253, + 369, + 480, + 383 + ], + "spans": [ + { + "bbox": [ + 253, + 369, + 480, + 383 + ], + "type": "interline_equation", + "content": "h _ {i} ^ {d} = \\operatorname {D e c o d e r} \\left(c _ {i}, x _ {i - 1}\\right), \\tag {7}", + "image_path": "5ea87bc18175c6b180ccf3d9babb278f8e495f84b06ed0b70aafc614057416a2.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 255, + 403, + 480, + 417 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 255, + 403, + 480, + 417 + ], + "spans": [ + { + "bbox": [ + 255, + 403, + 480, + 417 + ], + "type": "interline_equation", + "content": "x _ {i} = \\operatorname {C o n v 2 D} \\left(h _ {i} ^ {d}\\right) + y _ {i}, \\tag {8}", + "image_path": "869a91bcea98e13e4dd934d63d4e152b0ce85f6730d9145833235f342a1b1151.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 130, + 420, + 482, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 420, + 482, + 486 + ], + "spans": [ + { + "bbox": [ + 130, + 420, + 482, + 486 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 420, + 482, + 486 + ], + "type": "inline_equation", + "content": "c_{i}" + }, + { + "bbox": [ + 130, + 420, + 482, + 486 + ], + "type": "text", + "content": " is generated by AAM. The " + }, + { + "bbox": [ + 130, + 420, + 482, + 486 + ], + "type": "inline_equation", + "content": "h_{i}^{d}" + }, + { + "bbox": [ + 130, + 420, + 482, + 486 + ], + "type": "text", + "content": " is fused from its forward and backward states " + }, + { + "bbox": [ + 130, + 420, + 482, + 486 + ], + "type": "inline_equation", + "content": "(h_{i - 1}^{fw - d}" + }, + { + "bbox": [ + 130, + 420, + 482, + 486 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 420, + 482, + 486 + ], + "type": "inline_equation", + "content": "h_{i - 1}^{bw - d}" + }, + { + "bbox": [ + 130, + 420, + 482, + 486 + ], + "type": "text", + "content": ") and decoded feature " + }, + { + "bbox": [ + 130, + 420, + 482, + 486 + ], + "type": "inline_equation", + "content": "f_{i}^{d}" + }, + { + "bbox": [ + 130, + 420, + 482, + 486 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 420, + 482, + 486 + ], + "type": "inline_equation", + "content": "f_{i}^{d}" + }, + { + "bbox": [ + 130, + 420, + 482, + 486 + ], + "type": "text", + "content": " is the concatenation feature of generated output " + }, + { + "bbox": [ + 130, + 420, + 482, + 486 + ], + "type": "inline_equation", + "content": "x_{i - 1}" + }, + { + "bbox": [ + 130, + 420, + 482, + 486 + ], + "type": "text", + "content": " and context " + }, + { + "bbox": [ + 130, + 420, + 482, + 486 + ], + "type": "inline_equation", + "content": "c_{i}" + }, + { + "bbox": [ + 130, + 420, + 482, + 486 + ], + "type": "text", + "content": ". Thus, the inference of " + }, + { + "bbox": [ + 130, + 420, + 482, + 486 + ], + "type": "inline_equation", + "content": "\\{x_{i}\\}_{1}^{B}" + }, + { + "bbox": [ + 130, + 420, + 482, + 486 + ], + "type": "text", + "content": " is to maximizes " + }, + { + "bbox": [ + 130, + 420, + 482, + 486 + ], + "type": "inline_equation", + "content": "P(\\{x_{i}\\}_{1}^{B} \\mid \\{h_{i}^{e}\\}_{1}^{B})" + }, + { + "bbox": [ + 130, + 420, + 482, + 486 + ], + "type": "text", + "content": ". Based on the chain rule, the goal of Decoder can be reformulated as:" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 214, + 493, + 480, + 525 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 214, + 493, + 480, + 525 + ], + "spans": [ + { + "bbox": [ + 214, + 493, + 480, + 525 + ], + "type": "interline_equation", + "content": "\\arg \\max \\prod_ {i = 1} ^ {B} P \\left(x _ {i} \\mid x _ {1}, \\dots , x _ {i - 1}; h _ {1} ^ {e}, \\dots , h _ {B} ^ {e}\\right). \\tag {9}", + "image_path": "55ef4333225a4bff219ed561a925b925443b4806f09c80d050ae3a92464559a6.jpg" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 131, + 550, + 375, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 550, + 375, + 562 + ], + "spans": [ + { + "bbox": [ + 131, + 550, + 375, + 562 + ], + "type": "text", + "content": "3.2 Discriminative Semantic Regions Learning" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": "Taking a deep insight into the intrinsic properties of HSI (see Fig. 2), we propose the SSRA module (see Fig. 4(a)) to automatically assign learnable filters to their corresponding semantic regions for modeling diverse spatial-spectral variations. Inspired by DRConv [10] (proposed for high-level vision tasks like classification and detection) that divides shallow features into several regions based on local spatial semantics, which is not infeasible and reliable in the regions severely contaminated by noises for HSI, SSRA utilizes HSI-specific spatial-spectral properties to improve semantic region segmentation. In detail, thanks to the Seq2Seq" + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 255, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 255, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 255, + 91, + 448, + 102 + ], + "type": "text", + "content": "Sequence-to-Sequence Hyperspectral Denoising" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "content": "framework, the SSRA includes spatial and spectral mask prediction branches that operate on the hidden state list " + }, + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "inline_equation", + "content": "\\{h_i^e\\}_1^B" + }, + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "content": " to guide the allocation of filters, which effectively aggregates spatial-spectral representation." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 152, + 480, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 152, + 480, + 228 + ], + "spans": [ + { + "bbox": [ + 130, + 152, + 480, + 228 + ], + "type": "text", + "content": "In the spatial branch, we propose SpecDynConv (see Fig. 4(b)) that empowers the spatial convolutions with spectral modeling abilities to address the limitations of local spatial semantics and generates more precise spatial mask prediction " + }, + { + "bbox": [ + 130, + 152, + 480, + 228 + ], + "type": "inline_equation", + "content": "M_{i}^{spa} \\in \\mathbb{R}^{H \\times W \\times r}" + }, + { + "bbox": [ + 130, + 152, + 480, + 228 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 130, + 152, + 480, + 228 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 130, + 152, + 480, + 228 + ], + "type": "text", + "content": " means output region number). Specifically, the filter of SpecDynConv " + }, + { + "bbox": [ + 130, + 152, + 480, + 228 + ], + "type": "inline_equation", + "content": "w^{dyn} \\in \\mathbb{R}^{k^2 \\times r \\times C}" + }, + { + "bbox": [ + 130, + 152, + 480, + 228 + ], + "type": "text", + "content": " for hidden state " + }, + { + "bbox": [ + 130, + 152, + 480, + 228 + ], + "type": "inline_equation", + "content": "h_{i}^{e}" + }, + { + "bbox": [ + 130, + 152, + 480, + 228 + ], + "type": "text", + "content": " is decomposed into the product of a base filter " + }, + { + "bbox": [ + 130, + 152, + 480, + 228 + ], + "type": "inline_equation", + "content": "w^{b} \\in \\mathbb{R}^{k^2 \\times r \\times C}" + }, + { + "bbox": [ + 130, + 152, + 480, + 228 + ], + "type": "text", + "content": " and a calibration weight " + }, + { + "bbox": [ + 130, + 152, + 480, + 228 + ], + "type": "inline_equation", + "content": "\\alpha_{i} \\in \\mathbb{R}^{1 \\times 1 \\times C}" + }, + { + "bbox": [ + 130, + 152, + 480, + 228 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 205, + 239, + 480, + 252 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 205, + 239, + 480, + 252 + ], + "spans": [ + { + "bbox": [ + 205, + 239, + 480, + 252 + ], + "type": "interline_equation", + "content": "w ^ {d y n} = \\alpha_ {i} \\cdot w ^ {b} + 1; \\quad \\alpha_ {i} = \\operatorname {C o n v 1 D} \\left(\\operatorname {G A P} \\left(h _ {i} ^ {e}\\right)\\right), \\tag {10}", + "image_path": "64f8e18a5012944a2fb8a3825e982e4e43d781d273767a879b491e8118ddbe6f.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 258, + 479, + 318 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 258, + 479, + 318 + ], + "spans": [ + { + "bbox": [ + 130, + 258, + 479, + 318 + ], + "type": "text", + "content": "where the " + }, + { + "bbox": [ + 130, + 258, + 479, + 318 + ], + "type": "inline_equation", + "content": "\\alpha_{i}" + }, + { + "bbox": [ + 130, + 258, + 479, + 318 + ], + "type": "text", + "content": " is dynamically generated from " + }, + { + "bbox": [ + 130, + 258, + 479, + 318 + ], + "type": "inline_equation", + "content": "h_i^e" + }, + { + "bbox": [ + 130, + 258, + 479, + 318 + ], + "type": "text", + "content": " through the global average pooling over the spatial dimension and two-layer 1D convolutions, which effectively model the spectral dynamics for all channels in " + }, + { + "bbox": [ + 130, + 258, + 479, + 318 + ], + "type": "inline_equation", + "content": "w^{b}" + }, + { + "bbox": [ + 130, + 258, + 479, + 318 + ], + "type": "text", + "content": ". Different from SENet [19], the calibration weight generated by the input feature can better adapt to itself as illustrated in [20]. Thus, the spatial mask " + }, + { + "bbox": [ + 130, + 258, + 479, + 318 + ], + "type": "inline_equation", + "content": "M_{i}^{spa}" + }, + { + "bbox": [ + 130, + 258, + 479, + 318 + ], + "type": "text", + "content": " is calculated as:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 227, + 328, + 480, + 341 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 227, + 328, + 480, + 341 + ], + "spans": [ + { + "bbox": [ + 227, + 328, + 480, + 341 + ], + "type": "interline_equation", + "content": "M _ {i} ^ {s p a} = \\operatorname {S c a t t e r} \\left(\\operatorname {A r g m a x} \\left(h _ {i} ^ {e} * w ^ {d y n}\\right)\\right) \\tag {11}", + "image_path": "778e0dde160d5ad3e2d88bdd904df02b4cfec369360f292407215e4edf32783d.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 350, + 479, + 386 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 350, + 479, + 386 + ], + "spans": [ + { + "bbox": [ + 130, + 350, + 479, + 386 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 350, + 479, + 386 + ], + "type": "inline_equation", + "content": "*" + }, + { + "bbox": [ + 130, + 350, + 479, + 386 + ], + "type": "text", + "content": " is 2D convolution operation. " + }, + { + "bbox": [ + 130, + 350, + 479, + 386 + ], + "type": "inline_equation", + "content": "\\operatorname{Argmax}(\\cdot)" + }, + { + "bbox": [ + 130, + 350, + 479, + 386 + ], + "type": "text", + "content": " is the index of the maximum value along the channel dimension, indicating which position belongs to which region. Subsequently, the result of " + }, + { + "bbox": [ + 130, + 350, + 479, + 386 + ], + "type": "inline_equation", + "content": "\\operatorname{Argmax}(\\cdot)" + }, + { + "bbox": [ + 130, + 350, + 479, + 386 + ], + "type": "text", + "content": " is scattered to its one-hot form." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 387, + 480, + 480 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 387, + 480, + 480 + ], + "spans": [ + { + "bbox": [ + 130, + 387, + 480, + 480 + ], + "type": "text", + "content": "In the spectral branch, based on the nature of HSI as illustrated in Fig 2, we propose the SpecClassifier to adaptively learn the spectral inconsistency of each spatial position for semantic region segmentation. Specifically, the SpecClassifier efficiently performs spectral mask prediction using a single-layer linear network. It takes the reshaped hidden state " + }, + { + "bbox": [ + 130, + 387, + 480, + 480 + ], + "type": "inline_equation", + "content": "h_i^e \\in \\mathbb{R}^{HW \\times C}" + }, + { + "bbox": [ + 130, + 387, + 480, + 480 + ], + "type": "text", + "content": " as input and produces the spectral mask " + }, + { + "bbox": [ + 130, + 387, + 480, + 480 + ], + "type": "inline_equation", + "content": "M_i^{spec} \\in \\mathbb{R}^{HW \\times r}" + }, + { + "bbox": [ + 130, + 387, + 480, + 480 + ], + "type": "text", + "content": ", which is then reshaped to match the dimensions of " + }, + { + "bbox": [ + 130, + 387, + 480, + 480 + ], + "type": "inline_equation", + "content": "M_i^{spa}" + }, + { + "bbox": [ + 130, + 387, + 480, + 480 + ], + "type": "text", + "content": " and executes the same index scattering operation as in the spatial branch:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 221, + 483, + 480, + 496 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 221, + 483, + 480, + 496 + ], + "spans": [ + { + "bbox": [ + 221, + 483, + 480, + 496 + ], + "type": "interline_equation", + "content": "M _ {i} ^ {\\text {s p e c}} = \\operatorname {S c a t t e r} \\left(\\operatorname {A r g m a x} \\left(\\operatorname {L i n e a r} \\left(h _ {i} ^ {e}\\right)\\right)\\right). \\tag {12}", + "image_path": "7adf53751db9c4923f3c365c78565027106008d8743fa792007e3be63a9bedaa.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 502, + 479, + 550 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 502, + 479, + 550 + ], + "spans": [ + { + "bbox": [ + 130, + 502, + 479, + 550 + ], + "type": "text", + "content": "Through spatial and spectral branches, the encoded state " + }, + { + "bbox": [ + 130, + 502, + 479, + 550 + ], + "type": "inline_equation", + "content": "h_i^e" + }, + { + "bbox": [ + 130, + 502, + 479, + 550 + ], + "type": "text", + "content": " is divided into " + }, + { + "bbox": [ + 130, + 502, + 479, + 550 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 130, + 502, + 479, + 550 + ], + "type": "text", + "content": " distinct regions, with each region sharing only one filter [38] in " + }, + { + "bbox": [ + 130, + 502, + 479, + 550 + ], + "type": "inline_equation", + "content": "w = \\{w_1, \\ldots, w_r\\}" + }, + { + "bbox": [ + 130, + 502, + 479, + 550 + ], + "type": "text", + "content": ". The summation is applied to obtain the enhanced features " + }, + { + "bbox": [ + 130, + 502, + 479, + 550 + ], + "type": "inline_equation", + "content": "\\{h_i^{spa}, h_i^{spec}\\}" + }, + { + "bbox": [ + 130, + 502, + 479, + 550 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 130, + 502, + 479, + 550 + ], + "type": "inline_equation", + "content": "\\{M_i^{spa}, M_i^{spec}\\}" + }, + { + "bbox": [ + 130, + 502, + 479, + 550 + ], + "type": "text", + "content": " as below:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 207, + 559, + 480, + 583 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 207, + 559, + 480, + 583 + ], + "spans": [ + { + "bbox": [ + 207, + 559, + 480, + 583 + ], + "type": "interline_equation", + "content": "\\left\\{h _ {i} ^ {s p a}, h _ {i} ^ {s p e c} \\right\\} = \\sum_ {r} \\left(h _ {i} ^ {e} * w _ {i}\\right) \\otimes \\left\\{M _ {i} ^ {s p a}, M _ {i} ^ {s p e c} \\right\\}, \\tag {13}", + "image_path": "0b2e9ee9aa998db1f5faee3a9d803fa2a0e6f013213628c63527d93e2f0d8603.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 593, + 480, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 593, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 593, + 480, + 665 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 593, + 480, + 665 + ], + "type": "inline_equation", + "content": "\\otimes" + }, + { + "bbox": [ + 130, + 593, + 480, + 665 + ], + "type": "text", + "content": " denotes the pixel-wise multiplications. Finally, " + }, + { + "bbox": [ + 130, + 593, + 480, + 665 + ], + "type": "inline_equation", + "content": "\\{h_i^{spa}, h_i^{spec}\\}" + }, + { + "bbox": [ + 130, + 593, + 480, + 665 + ], + "type": "text", + "content": " are fused to generate the final result " + }, + { + "bbox": [ + 130, + 593, + 480, + 665 + ], + "type": "inline_equation", + "content": "h_i^e - r \\in \\mathbb{R}^{H \\times W \\times C}" + }, + { + "bbox": [ + 130, + 593, + 480, + 665 + ], + "type": "text", + "content": " using a 2D convolution layer. To better visualize the region-aware capability of the SSRA module, we use shallow feature " + }, + { + "bbox": [ + 130, + 593, + 480, + 665 + ], + "type": "inline_equation", + "content": "f_i" + }, + { + "bbox": [ + 130, + 593, + 480, + 665 + ], + "type": "text", + "content": " as input instead of deep encoded hidden feature " + }, + { + "bbox": [ + 130, + 593, + 480, + 665 + ], + "type": "inline_equation", + "content": "h_i^e" + }, + { + "bbox": [ + 130, + 593, + 480, + 665 + ], + "type": "text", + "content": ", Fig. 5 presents the predicted spatial and spectral guided masks (" + }, + { + "bbox": [ + 130, + 593, + 480, + 665 + ], + "type": "inline_equation", + "content": "r = 2" + }, + { + "bbox": [ + 130, + 593, + 480, + 665 + ], + "type": "text", + "content": ") generated by the SSRA module, demonstrating that our method successfully assigns filters to regions" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "type": "text", + "content": "J. Xiao et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "type": "text", + "content": "with similar spatial-spectral semantics. It needs to be noted that the spatial and spectral semantic segmentation results for the same position do not always align. The spatial branch emphasizes texture edges for partitioning, and the spectral branch focuses on target spectral characteristics. Therefore, SSRA combines both advantages and predictions for a more comprehensive feature learning." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 194, + 230, + 209 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 194, + 230, + 209 + ], + "spans": [ + { + "bbox": [ + 131, + 194, + 230, + 209 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 220, + 279, + 232 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 220, + 279, + 232 + ], + "spans": [ + { + "bbox": [ + 131, + 220, + 279, + 232 + ], + "type": "text", + "content": "4.1 Implementation Details" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 239, + 481, + 301 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 239, + 481, + 301 + ], + "spans": [ + { + "bbox": [ + 130, + 239, + 481, + 301 + ], + "type": "text", + "content": "Datasets. Following [28], the RAS2S is trained on ICVL [1]. To evaluate the proposed method, seven public benchmark datasets are adopted, including ICVL [1], CAVE [34], KAIST [11], RealHSI [57], and remotely sensed datasets Pavia centre [14], WDC Mall [45], and Urban [33] for testing. Evaluation metrics include PSNR, SSIM [42] and SAM [53]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 307, + 481, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 307, + 481, + 392 + ], + "spans": [ + { + "bbox": [ + 130, + 307, + 481, + 392 + ], + "type": "text", + "content": "Noise Settings. To simulate real-world noise interference, we introduce four types of noise (i.e., non-i.i.d Gaussian, stripe, deadline, and impulse noise) and combine them to create five different noise cases [43]. Among them, the mixture noise case is not introduced in the training process. To assess the generalization of models beyond training noise, we conduct additional validation on i.i.d Gaussian, speckle, and Poisson noise. More details and results are provided in the supplementary material." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 398, + 481, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 398, + 481, + 471 + ], + "spans": [ + { + "bbox": [ + 130, + 398, + 481, + 471 + ], + "type": "text", + "content": "Network Settings. We optimize the model with Adam [22]. The batch size is set to 8 with 90 epochs of training. The learning rate is initialized as 1e-3 and gradually decreased to 1e-6. We train the model on single NVIDIA GTX3090 GPU. Unless otherwise specified, all the quantitative and qualitative comparisons are conducted with 3 learnable regions within SSRA, and the sequence attention length within AAM is set to 3." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 488, + 228, + 499 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 488, + 228, + 499 + ], + "spans": [ + { + "bbox": [ + 131, + 488, + 228, + 499 + ], + "type": "text", + "content": "4.2 Main Results" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 508, + 487, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 508, + 487, + 593 + ], + "spans": [ + { + "bbox": [ + 130, + 508, + 487, + 593 + ], + "type": "text", + "content": "In all experiments, the RAS2S trained on ICVL [1] is directly adopted to denoise HSI datasets with variant-numbers of bands, without necessitating dataset-specific training and fine-tuning. In contrast, methods such as GRNet [6], T3SC [4], SST [27] and SERT [28], which lack flexibility with a high number of spectral bands, we adopt a sliding window manner for denoising and calculate the average values as the ultimate results. More results are provided in the supplementary material." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 609, + 293, + 623 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 609, + 293, + 623 + ], + "spans": [ + { + "bbox": [ + 131, + 609, + 293, + 623 + ], + "type": "text", + "content": "Simulation Scenarios Denoising" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 629, + 489, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 489, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 489, + 667 + ], + "type": "text", + "content": "Quantitative Comparison. (i) Noise robustness evaluations on the KAIST [11] dataset are presented in Tab. 1. RAS2S exhibits superior denoising performance across all noise cases, particularly outperforming the state-of-the-art method by" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 255, + 91, + 447, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 255, + 91, + 447, + 103 + ], + "spans": [ + { + "bbox": [ + 255, + 91, + 447, + 103 + ], + "type": "text", + "content": "Sequence-to-Sequence Hyperspectral Denoising" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 134, + 146, + 479, + 281 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 479, + 137 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 479, + 137 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 479, + 137 + ], + "type": "text", + "content": "Table 1: Average quantitative results under various noise cases on KAIST [11] dataset. The best and second-best values are highlighted." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 134, + 146, + 479, + 281 + ], + "lines": [ + { + "bbox": [ + 134, + 146, + 479, + 281 + ], + "spans": [ + { + "bbox": [ + 134, + 146, + 479, + 281 + ], + "type": "table", + "html": "
MethodsNon-i.i.d GaussianGaussian+StripeGaussian+DeadlineGaussian+ImpulseGaussian+Mixture
PSNR↑SSIM↑SAM↓PSNR↑SSIM↑SAM↓PSNR↑SSIM↑SAM↓PSNR↑SSIM↑SAM↓PSNR↑SSIM↑SAM↓
Noisy18.260.61451.074118.100.61361.074617.230.57701.096015.000.46251.096513.700.42161.1133
LLRT [8]33.360.93390.166530.850.87250.635130.060.89420.549527.140.71260.740024.740.66850.7764
LRTDTV [40]37.330.95100.219537.160.94980.211336.490.94130.235437.030.95050.244535.600.93550.2503
LLRGTV [18]36.660.91090.367836.680.91520.351034.760.87610.426833.610.85020.510932.140.81660.5218
NGMeet [17]32.900.90380.275532.640.90270.302232.060.90020.296625.790.7230.739524.540.70420.7612
QRNN3D [43]39.420.98740.208539.240.98770.208638.990.98680.216738.150.98530.250537.100.98310.2934
GRNet [6]30.580.91310.523530.410.90930.529830.220.90440.528028.770.88700.599628.290.88470.5780
MAC-Net [49]34.330.95370.479533.370.95200.473332.820.94280.471831.020.92050.592528.410.90410.5840
T3SC [4]37.790.99150.160337.680.99150.167837.040.99030.188834.870.98080.256133.620.97500.2958
GRUnet [24]39.390.98630.223239.210.98690.211338.940.98500.229438.110.98340.281136.690.97410.3525
MST [5]32.680.95450.394232.530.95720.396532.570.95300.405034.850.95690.374734.120.95130.3884
MAN [23]38.760.98870.207638.540.98890.210238.160.98810.211736.380.97790.322035.270.97040.3818
SST [27]36.790.97740.226736.590.97790.229436.340.97620.232335.620.97260.262234.650.96380.3050
BERT [28]37.770.98510.191837.570.98490.195637.350.98360.197636.580.98030.226235.710.97610.2459
HSDT-S [25]40.180.99050.141340.070.99090.140639.760.99200.146539.470.99130.153738.310.99040.1582
RAS2S(Ours)40.700.99160.137340.740.99190.136440.520.99360.129340.140.99330.136239.480.99200.1532
", + "image_path": "5f71f9bb2afa51582989d212589447fc001dde6d822405c68147996cd5819f7e.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 135, + 324, + 479, + 487 + ], + "blocks": [ + { + "bbox": [ + 132, + 293, + 479, + 315 + ], + "lines": [ + { + "bbox": [ + 132, + 293, + 479, + 315 + ], + "spans": [ + { + "bbox": [ + 132, + 293, + 479, + 315 + ], + "type": "text", + "content": "Table 2: Average quantitative results under mixture noise case on different datasets. The best and second-best values are highlighted." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 135, + 324, + 479, + 487 + ], + "lines": [ + { + "bbox": [ + 135, + 324, + 479, + 487 + ], + "spans": [ + { + "bbox": [ + 135, + 324, + 479, + 487 + ], + "type": "table", + "html": "
MethodsICVL [1]CAVE [34]Pavia Centre [14]WDC [45]
PSNR↑SSIM↑SAM↓PSNR↑SSIM↑SAM↓PSNR↑SSIM↑SAM↓PSNR↑SSIM↑SAM↓
Noisy13.970.33920.898714.170.41881.137113.890.34000.974613.980.21481.0120
LLRT [8]24.790.62940.488924.590.57950.817116.630.48950.843722.800.50830.5718
LRTDTV [40]34.460.91840.112733.820.90850.293829.650.89630.244536.330.85970.2148
LLRGTV [18]31.390.87560.253827.120.72210.656728.410.89230.314234.320.82600.3150
NGMeet [17]26.590.77870.401623.200.62290.787422.590.66580.483923.260.50310.4541
QRNN3D [43]39.220.99040.080936.550.98250.424432.930.96980.157033.960.87440.1344
GRNet [6]31.670.95570.143128.440.88990.632926.570.85360.281525.140.76820.3255
MAC-Net [49]30.750.93320.267328.530.89200.623427.340.88130.353030.740.77400.5371
T3SC [4]35.680.97900.138933.610.97280.413731.390.95230.231430.490.90640.1972
GRUNet [24]38.630.99130.065135.980.97640.404931.860.95920.159831.370.80180.1650
MST [5]39.120.99140.058233.970.94730.465731.210.93550.197133.490.85570.2289
MAN [23]35.850.96910.145835.150.97120.499032.940.97110.182032.770.80650.1638
SST [27]39.580.99280.048034.890.96160.409532.850.95880.155530.590.89240.1860
BERT [28]40.440.99410.047035.860.97370.340333.280.96530.145131.330.90980.1609
HSDT-S [25]40.760.99400.050538.520.98920.236234.250.97680.129537.790.94650.1049
RAS2S(Ours)41.330.99530.046739.030.99130.222136.840.98620.095840.920.97600.0735
", + "image_path": "1069fda4e46179645be84635288f02c53f814850e55ee6120e952a5fc19ceda8.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 511, + 482, + 609 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 511, + 482, + 609 + ], + "spans": [ + { + "bbox": [ + 130, + 511, + 482, + 609 + ], + "type": "text", + "content": "3.6 dB in untrained mixture noise cases. This highlights the impressive denoising robustness of RAS2S. (ii) Generalization results on multiple datasets are shown in Tab. 2. RAS2S outperforms other methods on ICVL [1] and CAVE [34] (with the same spectral number as the training set). Additionally, RAS2S exhibits stronger generalization abilities on remotely sensed datasets, Pavia Centre [14] and WDC [45], with a larger spectral number than the training set. RAS2S outperforms the state-of-the-art denoiser on all indicators by a large margin, demonstrating robust generalization and insensitivity to spectral categories." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "Qualitative Comparison. Qualitative comparisons between RAS2S and other models in Fig. 6 reveal the following: (i) For ICVL [1] and CAVE [34] with the same spectral number as the training set, other methods generate inaccurate textures and exhibit inferior noise removal performance. (ii) For the KAIST [11]" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "text", + "content": "J. Xiao et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 136, + 114, + 478, + 289 + ], + "blocks": [ + { + "bbox": [ + 136, + 114, + 478, + 289 + ], + "lines": [ + { + "bbox": [ + 136, + 114, + 478, + 289 + ], + "spans": [ + { + "bbox": [ + 136, + 114, + 478, + 289 + ], + "type": "image", + "image_path": "28282d7ce9ad1aafd8fcafcbae489991f765e56d0e41f1ce6e50e0762401588f.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 297, + 482, + 331 + ], + "lines": [ + { + "bbox": [ + 132, + 297, + 482, + 331 + ], + "spans": [ + { + "bbox": [ + 132, + 297, + 482, + 331 + ], + "type": "text", + "content": "Fig. 6: Visual comparisons on prevailing paired datasets. Except for RealHSI [57], which involves real-world noise, the other datasets are contaminated with Non-i.i.d Gaussian+Mixture noise. GT denotes the reference ground-truth." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 140, + 342, + 476, + 494 + ], + "blocks": [ + { + "bbox": [ + 140, + 342, + 476, + 494 + ], + "lines": [ + { + "bbox": [ + 140, + 342, + 476, + 494 + ], + "spans": [ + { + "bbox": [ + 140, + 342, + 476, + 494 + ], + "type": "image", + "image_path": "7e4a50cd7719f2109d905c35e0893761fb24e7b35f857f40471b14dc90243426.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 184, + 501, + 428, + 512 + ], + "lines": [ + { + "bbox": [ + 184, + 501, + 428, + 512 + ], + "spans": [ + { + "bbox": [ + 184, + 501, + 428, + 512 + ], + "type": "text", + "content": "Fig. 7: Visual comparison on the real-world dataset Urban." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 537, + 482, + 609 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 537, + 482, + 609 + ], + "spans": [ + { + "bbox": [ + 130, + 537, + 482, + 609 + ], + "type": "text", + "content": "dataset with high spatial resolution, RAS2S effectively removes complex noises while retaining detailed textures and edges. In contrast, other models introduce color discrepancies and detail loss (T3SC [4] and SERT [28]) or leave behind residual noise (MACNet [49] and SST [27]). (iii) For the WDC [45] and Pavia Centre datasets with a large spectral number, RAS2S exhibits outstanding noise removal while preserving main structures with fewer artifacts." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "text", + "content": "Real-World Scenarios Since the ultimate goal is to deal with complicated real-world noises and produce visually pleasing HSIs, we conduct experiments on the real-world datasets, i.e., RealHSI [57] and Urban [33]." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 255, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 255, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 255, + 91, + 448, + 102 + ], + "type": "text", + "content": "Sequence-to-Sequence Hyperspectral Denoising" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 134, + 146, + 316, + 256 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 318, + 137 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 318, + 137 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 318, + 137 + ], + "type": "text", + "content": "Table 3: Efficiency results on CAVE [34] and denoising results on RealHSI [57]." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 134, + 146, + 316, + 256 + ], + "lines": [ + { + "bbox": [ + 134, + 146, + 316, + 256 + ], + "spans": [ + { + "bbox": [ + 134, + 146, + 316, + 256 + ], + "type": "table", + "html": "
MethodsEfficiencyRealHSI [57]
Params(M)CPU(s)↓GPU(s)↓PSNR(dB)↑SAM↓
LLRT[8]-807.21-26.600.1004
LRTDTV[40]-97.65-26.580.0620
LLRGTV[18]-171.72-26.140.0626
NGMeet[17]-86.03-27.000.0592
QRNN3D[43]0.8640.060.3226.600.0629
GRNet[6]41.442.660.1226.800.1168
MACNet[49]0.434.952.5324.250.1480
T3SC[4]0.8320.730.5126.720.0576
GRUNet[24]14.2875.580.5026.600.0587
MAN230.8727.110.4526.930.0703
SST274.1035.722.0626.950.0711
BERT281.9118.690.5326.510.0850
HSDT-S250.1317.490.3127.250.0544
RAS2S(Ours)0.088.160.5827.390.0535
", + "image_path": "ebeac14cd7b9d64462aefbd24d863586d4e0f6217691980eb302fb4bdfe3b125.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 328, + 146, + 477, + 178 + ], + "blocks": [ + { + "bbox": [ + 325, + 114, + 478, + 136 + ], + "lines": [ + { + "bbox": [ + 325, + 114, + 478, + 136 + ], + "spans": [ + { + "bbox": [ + 325, + 114, + 478, + 136 + ], + "type": "text", + "content": "Table 4: Ablation studies to verify module effectiveness." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 328, + 146, + 477, + 178 + ], + "lines": [ + { + "bbox": [ + 328, + 146, + 477, + 178 + ], + "spans": [ + { + "bbox": [ + 328, + 146, + 477, + 178 + ], + "type": "table", + "html": "
AAMHFCompSSRAParams(KB)↓Time(s)↓PSNR(dB)↑
---79.117.1337.12
✓--79.437.7437.56
✓✓-99.628.3238.45
✓✓✓84.118.1639.03
", + "image_path": "d476c88369642ad3e64b8375a022645a44a1638c449b2d49b0751d18a2304675.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 328, + 211, + 479, + 256 + ], + "blocks": [ + { + "bbox": [ + 325, + 179, + 478, + 201 + ], + "lines": [ + { + "bbox": [ + 325, + 179, + 478, + 201 + ], + "spans": [ + { + "bbox": [ + 325, + 179, + 478, + 201 + ], + "type": "text", + "content": "Table 5: Evaluation of component effectiveness within SSRA." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 328, + 211, + 479, + 256 + ], + "lines": [ + { + "bbox": [ + 328, + 211, + 479, + 256 + ], + "spans": [ + { + "bbox": [ + 328, + 211, + 479, + 256 + ], + "type": "table", + "html": "
MethodsParams(KB)↓Time(s)↓PSNR(dB)↑
w/oSSRA76.467.6237.51
w/Conv2D82.877.9237.67
w/SpecDynConv84.068.0437.92
w/SpecClassifier82.537.9638.23
w/SSRA (Ours)84.118.1639.03
", + "image_path": "1383050284ffa89f4a145128eb041f35ee104e557c4d6a69882749ef47f815ab.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 278, + 482, + 374 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 278, + 482, + 374 + ], + "spans": [ + { + "bbox": [ + 130, + 278, + 482, + 374 + ], + "type": "text", + "content": "Quantitative Comparison. The quantitative comparison of different methods on the RealHSI [57] is presented in Tab. 3. Model parameters and inference time also can be observed. The CPU and GPU inference time are measured on Intel(R) Xeon(R) Gold 6226R and NVIDIA GTX3090, respectively. As can be seen, our RAS2S outperforms SERT [28] by a significant margin, achieving up to 0.88 dB PSNR scores, while employing even fewer parameters (only " + }, + { + "bbox": [ + 130, + 278, + 482, + 374 + ], + "type": "inline_equation", + "content": "4.18\\%" + }, + { + "bbox": [ + 130, + 278, + 482, + 374 + ], + "type": "text", + "content": " of SERT [28] parameters) and runtime. The results indicate that despite the simplicity, our RAS2S present effectiveness in dealing with real-world noise." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 382, + 482, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 382, + 482, + 443 + ], + "spans": [ + { + "bbox": [ + 130, + 382, + 482, + 443 + ], + "type": "text", + "content": "Qualitative Comparison. For visual comparison of denoised Urban [33] in Fig. 7, the competing methods have obvious residual stripe noise and blurring. In contrast, RAS2S still generates more realistic and visually pleasing textures without residual noises and artifacts, demonstrating the robustness of our method in real-world applications." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 471, + 246, + 482 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 471, + 246, + 482 + ], + "spans": [ + { + "bbox": [ + 132, + 471, + 246, + 482 + ], + "type": "text", + "content": "4.3 Ablation Studies" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 489, + 481, + 526 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 489, + 481, + 526 + ], + "spans": [ + { + "bbox": [ + 130, + 489, + 481, + 526 + ], + "type": "text", + "content": "Unless otherwise specified, we apply the CAVE dataset under the mixture noise for ablation studies. We modify different baselines to a similar model size by adjusting the feature channels for fair comparison." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": "Component Analysis. Tab. 4 summarizes the impact of different modules. The base Seq2Seq model achieves competitive performance with minimal model size, confirming the effectiveness of Seq2Seq learning. AAM improves denoising by " + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "inline_equation", + "content": "0.44\\mathrm{dB}" + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": ", HFComp enhances results by addressing information gaps between hidden states and current spectral features, while SSRA significantly boosts performance by " + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "inline_equation", + "content": "0.58\\mathrm{dB}" + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": " in PSNR, demonstrating the effectiveness of semantic region learning for spatial-spectral information exploration. Tab. 5 compares SSRA designs, with SpecDynConv enhancing denoising by " + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "inline_equation", + "content": "0.41\\mathrm{dB}" + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": " and SpecClassifier achieving a " + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "inline_equation", + "content": "0.72\\mathrm{dB}" + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": " PSNR gain by leveraging spectral properties. SSRA combines these advantages for comprehensive feature learning, resulting in substantial improvement." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "text", + "content": "J. Xiao et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 138, + 116, + 222, + 151 + ], + "blocks": [ + { + "bbox": [ + 138, + 116, + 222, + 151 + ], + "lines": [ + { + "bbox": [ + 138, + 116, + 222, + 151 + ], + "spans": [ + { + "bbox": [ + 138, + 116, + 222, + 151 + ], + "type": "image", + "image_path": "ad31c5d0232d4454a77349411a80f22098fa48901994146667059e5a0871eb06.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 154, + 152, + 215, + 159 + ], + "lines": [ + { + "bbox": [ + 154, + 152, + 215, + 159 + ], + "spans": [ + { + "bbox": [ + 154, + 152, + 215, + 159 + ], + "type": "text", + "content": "(a) Attention length" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 223, + 116, + 305, + 152 + ], + "blocks": [ + { + "bbox": [ + 223, + 116, + 305, + 152 + ], + "lines": [ + { + "bbox": [ + 223, + 116, + 305, + 152 + ], + "spans": [ + { + "bbox": [ + 223, + 116, + 305, + 152 + ], + "type": "image", + "image_path": "1dbf0846adae0e75d2b94ddef511c826781063d47744584935117cf504e62719.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 234, + 153, + 303, + 160 + ], + "lines": [ + { + "bbox": [ + 234, + 153, + 303, + 160 + ], + "spans": [ + { + "bbox": [ + 234, + 153, + 303, + 160 + ], + "type": "text", + "content": "(b) Region number " + }, + { + "bbox": [ + 234, + 153, + 303, + 160 + ], + "type": "inline_equation", + "content": "(r)" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 322, + 118, + 395, + 155 + ], + "blocks": [ + { + "bbox": [ + 322, + 118, + 395, + 155 + ], + "lines": [ + { + "bbox": [ + 322, + 118, + 395, + 155 + ], + "spans": [ + { + "bbox": [ + 322, + 118, + 395, + 155 + ], + "type": "image", + "image_path": "aff925b60b746ea967f0881b202af7ed875af27d42f69d245971141844e0e973.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 349, + 156, + 370, + 163 + ], + "lines": [ + { + "bbox": [ + 349, + 156, + 370, + 163 + ], + "spans": [ + { + "bbox": [ + 349, + 156, + 370, + 163 + ], + "type": "text", + "content": "(a) Loss" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 135, + 171, + 312, + 227 + ], + "lines": [ + { + "bbox": [ + 135, + 171, + 312, + 227 + ], + "spans": [ + { + "bbox": [ + 135, + 171, + 312, + 227 + ], + "type": "text", + "content": "Fig.8: Performance vs. the hyperparameters with different number. The " + }, + { + "bbox": [ + 135, + 171, + 312, + 227 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 135, + 171, + 312, + 227 + ], + "type": "text", + "content": " ,and curves denote the denoising performance on CAVE, ICVL and WDC datasets, respectively." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 318, + 172, + 477, + 216 + ], + "lines": [ + { + "bbox": [ + 318, + 172, + 477, + 216 + ], + "spans": [ + { + "bbox": [ + 318, + 172, + 477, + 216 + ], + "type": "text", + "content": "Fig.9: Convergence curve. We show the average training loss (a) and training batch PSNR (b) of the final model." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 395, + 118, + 470, + 155 + ], + "blocks": [ + { + "bbox": [ + 395, + 118, + 470, + 155 + ], + "lines": [ + { + "bbox": [ + 395, + 118, + 470, + 155 + ], + "spans": [ + { + "bbox": [ + 395, + 118, + 470, + 155 + ], + "type": "image", + "image_path": "a1ec2a93c7d85cf440b42e89a447241337820757e4988a34ba4489919070408e.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 419, + 156, + 443, + 163 + ], + "lines": [ + { + "bbox": [ + 419, + 156, + 443, + 163 + ], + "spans": [ + { + "bbox": [ + 419, + 156, + 443, + 163 + ], + "type": "text", + "content": "(b) PSNR" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 250, + 482, + 382 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 250, + 482, + 382 + ], + "spans": [ + { + "bbox": [ + 130, + 250, + 482, + 382 + ], + "type": "text", + "content": "Hyperparameter Analysis. We assess our proposed model under various settings for the sequence attention length in AAM and the learnable-region number in SSRA across multiple datasets. Fig. 8(a) shows that the model achieves optimal performance across all datasets when the sequence attention length in AAM is set to 3, suggesting that the nearest hidden state provides maximum assistance in capturing spatial-spectral relations for the current state. Longer attention lengths introduce excessive redundant spatial-spectral information, adversely affecting denoising performance. Fig. 8(b) indicates optimal performance when the learnable-region number in SSRA is 2 or 3, as the simplified SSRA design limits capacity for more regions, and increasing regions may lead to imprecise segmentation, degrading model performance." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 398, + 214, + 409 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 398, + 214, + 409 + ], + "spans": [ + { + "bbox": [ + 132, + 398, + 214, + 409 + ], + "type": "text", + "content": "4.4 Discussion" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 417, + 480, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 417, + 480, + 441 + ], + "spans": [ + { + "bbox": [ + 130, + 417, + 480, + 441 + ], + "type": "text", + "content": "Model Complexity. Although RAS2S includes the quadratic attention module (HFComp), it is simpler than Transformer-based methods as shown in Tab. 6." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 441, + 323, + 501 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 441, + 323, + 501 + ], + "spans": [ + { + "bbox": [ + 130, + 441, + 323, + 501 + ], + "type": "text", + "content": "(i) Fewer parameters: Transformer methods often stack numerous attention layers, while RAS2S utilizes only a single layer of cross-attention. (ii) Lower computational complexity: In most denoising transformers" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 501, + 481, + 585 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 501, + 481, + 585 + ], + "spans": [ + { + "bbox": [ + 130, + 501, + 481, + 585 + ], + "type": "text", + "content": "(e.g., SST, HSDT), the spectral feature channels " + }, + { + "bbox": [ + 130, + 501, + 481, + 585 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 130, + 501, + 481, + 585 + ], + "type": "text", + "content": " with quadratic computational overhead are typically more than the spectral bands " + }, + { + "bbox": [ + 130, + 501, + 481, + 585 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 130, + 501, + 481, + 585 + ], + "type": "text", + "content": ". In contrast, the channels " + }, + { + "bbox": [ + 130, + 501, + 481, + 585 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 130, + 501, + 481, + 585 + ], + "type": "text", + "content": " of RAS2S is more less than bands " + }, + { + "bbox": [ + 130, + 501, + 481, + 585 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 130, + 501, + 481, + 585 + ], + "type": "text", + "content": ", resulting in lower model size and computational complexity. Moreover, the smaller parameter space increases the likelihood of finding the global or local optimum during optimization, making it easier for the final model to converge. The convergence curve is shown in the Fig. 9." + } + ] + } + ], + "index": 16 + }, + { + "type": "table", + "bbox": [ + 324, + 457, + 479, + 499 + ], + "blocks": [ + { + "bbox": [ + 338, + 447, + 465, + 457 + ], + "lines": [ + { + "bbox": [ + 338, + 447, + 465, + 457 + ], + "spans": [ + { + "bbox": [ + 338, + 447, + 465, + 457 + ], + "type": "text", + "content": "Table 6: Complexity analysis." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 324, + 457, + 479, + 499 + ], + "lines": [ + { + "bbox": [ + 324, + 457, + 479, + 499 + ], + "spans": [ + { + "bbox": [ + 324, + 457, + 479, + 499 + ], + "type": "table", + "html": "
ModelPSNR(dB)Params(M)Complexity
SST34.894.10HW(M2C + C2)
HSDT-S38.520.13HWC2B
Ours39.030.08HWC2B
", + "image_path": "0aacddae905e3643632d8542cf0de3fd2baff723e906df77b18d54a61b1c1ceb.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "table_body" + } + ], + "index": 18 + }, + { + "bbox": [ + 130, + 593, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 593, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 593, + 482, + 666 + ], + "type": "text", + "content": "Unidirection vs. Bidirection. Existing Seq2Seq models follow unidirectional forward temporal modeling, which is unsuitable for hyperspectral spectral sequences. We validate forward, backward, and bidirectional spectral modeling approaches. Tab. 7 shows that forward and backward spectral modeling yield similar denoising performance, as spectral sequence are not sensitive to the order of spectral bands. In contrast, the bidirectional modeling framework enhances" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 255, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 255, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 255, + 91, + 448, + 102 + ], + "type": "text", + "content": "Sequence-to-Sequence Hyperspectral Denoising" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 134, + 146, + 282, + 184 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 284, + 137 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 284, + 137 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 284, + 137 + ], + "type": "text", + "content": "Table 7: Ablation studies to verify bidirection framework." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 134, + 146, + 282, + 184 + ], + "lines": [ + { + "bbox": [ + 134, + 146, + 282, + 184 + ], + "spans": [ + { + "bbox": [ + 134, + 146, + 282, + 184 + ], + "type": "table", + "html": "
ModelParams(KB)↓Times(s)↓PSNR(dB)↑
Forward91.487.6738.42
Backward91.487.6738.29
Bidirection84.118.1639.03
", + "image_path": "c7db54cb8a9e073b417b4c416f0fd1a4be95219a57cc7e19f09093fb9eb5801c.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 299, + 147, + 467, + 184 + ], + "blocks": [ + { + "bbox": [ + 296, + 114, + 470, + 137 + ], + "lines": [ + { + "bbox": [ + 296, + 114, + 470, + 137 + ], + "spans": [ + { + "bbox": [ + 296, + 114, + 470, + 137 + ], + "type": "text", + "content": "Table 8: Ablation studies to verify autoregressive paradigm." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 299, + 147, + 467, + 184 + ], + "lines": [ + { + "bbox": [ + 299, + 147, + 467, + 184 + ], + "spans": [ + { + "bbox": [ + 299, + 147, + 467, + 184 + ], + "type": "table", + "html": "
ModelParams(KB)↓Times(s)↓PSNR(dB)↑
w/o AR80.567.8338.62
w/ AR84.118.1639.03
", + "image_path": "edb541af4d24e6c2135e61451cf5a48bc05ec0a115d39cab5e2195980d019d4c.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 205, + 479, + 231 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 205, + 479, + 231 + ], + "spans": [ + { + "bbox": [ + 130, + 205, + 479, + 231 + ], + "type": "text", + "content": "denoising performance by integrating preceding and succeeding spectral information, effectively capturing long-range spectral dependencies." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 237, + 480, + 322 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 237, + 480, + 322 + ], + "spans": [ + { + "bbox": [ + 130, + 237, + 480, + 322 + ], + "type": "text", + "content": "Autoregressive Paradigm. Current per-band denoising methods restore each band independently using spectral information from adjacent bands, but overlook leveraging spectral similarity constraints. Our autoregressive framework addresses this by incorporating information from previously generated bands to enhance the reconstruction of the current band, enabling more effective learning, maintaining consistency across bands, and reducing noise impact. Without the autoregressive paradigm, Equation 9 simplifies to:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 242, + 329, + 481, + 361 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 242, + 329, + 481, + 361 + ], + "spans": [ + { + "bbox": [ + 242, + 329, + 481, + 361 + ], + "type": "interline_equation", + "content": "\\arg \\max \\prod_ {i = 1} ^ {B} P \\left(x _ {i} \\mid h _ {1} ^ {e}, \\dots , h _ {B} ^ {e}\\right). \\tag {14}", + "image_path": "396dfd7b2f38ee1e8bc7094157283ee5e7e15096496cdb47298d0551023163b0.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 367, + 481, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 367, + 481, + 392 + ], + "spans": [ + { + "bbox": [ + 130, + 367, + 481, + 392 + ], + "type": "text", + "content": "We make an ablation study in Tab. 8. It is evident that adopting the autoregressive paradigm lead to overall better denoising performance for our model." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 408, + 220, + 422 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 408, + 220, + 422 + ], + "spans": [ + { + "bbox": [ + 132, + 408, + 220, + 422 + ], + "type": "text", + "content": "5 Conclusion" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 433, + 482, + 577 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 433, + 482, + 577 + ], + "spans": [ + { + "bbox": [ + 130, + 433, + 482, + 577 + ], + "type": "text", + "content": "We propose RAS2S, a simple yet effective framework for HSI denoising, leveraging the power of Seq2Seq learning. We regard HSI denoising as a noisy-to-clean sequence translation problem, and maximize the denoising potential of the Seq2Seq model through discriminative semantic region learning, utilizing insights from the intrinsic spatial-spectral properties of HSI. RAS2S uses a simple bidirectional encoder-decoder architecture and serially models spatial-spectral correlations without the need for dataset-specific training, presenting the huge potential for capturing long-range dependencies. More about, RAS2S effectively handles the diverse spatial-spectral semantic variations, delivering more precise denoising results. Extensive experiments prove RAS2S outperforms other mainstream methods and achieves state-of-the-art on prevailing benchmark datasets with few parameters." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 594, + 247, + 608 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 594, + 247, + 608 + ], + "spans": [ + { + "bbox": [ + 132, + 594, + 247, + 608 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "type": "text", + "content": "This work was supported by National Key R&D Program of China under Grant No. 2021ZD0110400, the Fundamental Research Funds for the Central Universities No. xxj032023020, and sponsored by the CAAI-MindSpore Open Fund, developed on OpenI Community." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "type": "text", + "content": "J. Xiao et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 197, + 126 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 197, + 126 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 197, + 126 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 141, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 24, + "blocks": [ + { + "bbox": [ + 138, + 141, + 480, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 141, + 480, + 163 + ], + "spans": [ + { + "bbox": [ + 138, + 141, + 480, + 163 + ], + "type": "text", + "content": "1. Arad, B., Ben-Shahar, O.: Sparse recovery of hyperspectral signal from natural rgb images. In: ECCV (2016) 9, 10" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 165, + 480, + 186 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 165, + 480, + 186 + ], + "spans": [ + { + "bbox": [ + 138, + 165, + 480, + 186 + ], + "type": "text", + "content": "2. Bai, Y., Zhao, Z., Gong, Y., Wei, X.: Artrackv2: Prompting autoregressive tracker where to look and how to describe. In: CVPR (2024) 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 187, + 480, + 208 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 187, + 480, + 208 + ], + "spans": [ + { + "bbox": [ + 138, + 187, + 480, + 208 + ], + "type": "text", + "content": "3. Ballas, N., Yao, L., Pal, C., Courville, A.: Delving deeper into convolutional networks for learning video representations. In: ICLR (2016) 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 209, + 480, + 240 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 209, + 480, + 240 + ], + "spans": [ + { + "bbox": [ + 138, + 209, + 480, + 240 + ], + "type": "text", + "content": "4. Bodrito, T., Zouaoui, A., Chanussot, J., Mairal, J.: A trainable spectral-spatial sparse coding model for hyperspectral image restoration. In: NeurIPS (2021) 9, 10, 11, 12" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 243, + 480, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 243, + 480, + 274 + ], + "spans": [ + { + "bbox": [ + 138, + 243, + 480, + 274 + ], + "type": "text", + "content": "5. Cai, Y., Lin, J., Hu, X., Wang, H., Yuan, X., Zhang, Y., Timofte, R., Van Gool, L.: Mask-guided spectral-wise transformer for efficient hyperspectral image reconstruction. In: CVPR (2022) 2, 10" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 276, + 480, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 276, + 480, + 297 + ], + "spans": [ + { + "bbox": [ + 138, + 276, + 480, + 297 + ], + "type": "text", + "content": "6. Cao, X., Fu, X., Xu, C., Meng, D.: Deep spatial-spectral global reasoning network for hyperspectral image denoising. IEEE TGRS 60, 1-14 (2021) 9, 10, 12" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 299, + 480, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 299, + 480, + 330 + ], + "spans": [ + { + "bbox": [ + 138, + 299, + 480, + 330 + ], + "type": "text", + "content": "7. Chang, Y., Yan, L., Fang, H., Luo, C.: Anisotropic spectral-spatial total variation model for multispectral remote sensing image destriping. IEEE TIP 24(6), 1852-1866 (2015) 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 331, + 480, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 331, + 480, + 353 + ], + "spans": [ + { + "bbox": [ + 138, + 331, + 480, + 353 + ], + "type": "text", + "content": "8. Chang, Y., Yan, L., Zhong, S.: Hyper-laplacian regularized unidirectional low-rank tensor recovery for multispectral image denoising. In: CVPR (2017) 3, 10, 12" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 354, + 480, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 354, + 480, + 384 + ], + "spans": [ + { + "bbox": [ + 138, + 354, + 480, + 384 + ], + "type": "text", + "content": "9. Chen, H., Zhao, W., Xu, T., Shi, G., Zhou, S., Liu, P., Li, J.: Spectral-wise implicit neural representation for hyperspectral image reconstruction. IEEE TCSVT (2023) 4" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 387, + 480, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 387, + 480, + 408 + ], + "spans": [ + { + "bbox": [ + 138, + 387, + 480, + 408 + ], + "type": "text", + "content": "0. Chen, J., Wang, X., Guo, Z., Zhang, X., Sun, J.: Dynamic region-aware convolution. In: CVPR (2021) 7" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 410, + 480, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 410, + 480, + 430 + ], + "spans": [ + { + "bbox": [ + 138, + 410, + 480, + 430 + ], + "type": "text", + "content": "1. Choi, I., Kim, M., Gutierrez, D., Jeon, D., Nam, G.: High-quality hyperspectral reconstruction using a spectral prior. ACM TOG 36, 13 (2017) 9, 10" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 432, + 480, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 432, + 480, + 464 + ], + "spans": [ + { + "bbox": [ + 138, + 432, + 480, + 464 + ], + "type": "text", + "content": "2. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. In: ICLR (2020) 6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 465, + 480, + 486 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 465, + 480, + 486 + ], + "spans": [ + { + "bbox": [ + 138, + 465, + 480, + 486 + ], + "type": "text", + "content": "3. Fu, G., Xiong, F., Lu, J., Zhou, J., Qian, Y.: Nonlocal spatial-spectral neural network for hyperspectral image denoising. IEEE TGRS 60, 1-16 (2022) 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 487, + 480, + 507 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 487, + 480, + 507 + ], + "spans": [ + { + "bbox": [ + 138, + 487, + 480, + 507 + ], + "type": "text", + "content": "4. Gamba, P.: A collection of data for urban area characterization. In: IGARSS (2004) 9, 10" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 510, + 480, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 510, + 480, + 531 + ], + "spans": [ + { + "bbox": [ + 138, + 510, + 480, + 531 + ], + "type": "text", + "content": "5. Guan, J., Lai, R., Li, H., Yang, Y., Gu, L.: Dnrcnn: Deep recurrent convolutional neural network for hsi destriping. IEEE TNNLS (2022) 4" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 533, + 480, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 533, + 480, + 553 + ], + "spans": [ + { + "bbox": [ + 138, + 533, + 480, + 553 + ], + "type": "text", + "content": "6. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: CVPR (2016) 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 138, + 555, + 480, + 575 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 555, + 480, + 575 + ], + "spans": [ + { + "bbox": [ + 138, + 555, + 480, + 575 + ], + "type": "text", + "content": "7. He, W., Yao, Q., Li, C., Yokoya, N., Zhao, Q.: Non-local meets global: An integrated paradigm for hyperspectral denoising. In: CVPR (2019) 3, 10, 12" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 138, + 577, + 480, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 577, + 480, + 609 + ], + "spans": [ + { + "bbox": [ + 138, + 577, + 480, + 609 + ], + "type": "text", + "content": "8. He, W., Zhang, H., Shen, H., Zhang, L.: Hyperspectral image denoising using local low-rank matrix recovery and global spatial-spectral total variation. IEEE J-STARS 11(3), 713-729 (2018) 10, 12" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 138, + 610, + 472, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 610, + 472, + 620 + ], + "spans": [ + { + "bbox": [ + 138, + 610, + 472, + 620 + ], + "type": "text", + "content": "9. Hu, J., Shen, L., Sun, G.: Squeeze-and-excitation networks. In: CVPR (2018) 8" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 138, + 622, + 480, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 622, + 480, + 643 + ], + "spans": [ + { + "bbox": [ + 138, + 622, + 480, + 643 + ], + "type": "text", + "content": "20. Huang, Z., Zhang, S., Pan, L., Qing, Z., Tang, M., Liu, Z., Ang Jr, M.H.: Tada! temporally-adaptive convolutions for video understanding. In: ICLR (2022) 8" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 138, + 644, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 644, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 644, + 480, + 665 + ], + "type": "text", + "content": "21. Keneshloo, Y., Shi, T., Ramakrishnan, N., Reddy, C.K.: Deep reinforcement learning for sequence-to-sequence models. IEEE TNNLS 31(7), 2469-2489 (2019) 2" + } + ] + } + ], + "index": 23 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 255, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 255, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 255, + 91, + 447, + 102 + ], + "type": "text", + "content": "Sequence-to-Sequence Hyperspectral Denoising" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 24, + "blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 480, + 138 + ], + "type": "text", + "content": "22. Kingma, D.P., Ba, J.: Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980 (2014) 9" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 139, + 480, + 159 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 139, + 480, + 159 + ], + "spans": [ + { + "bbox": [ + 132, + 139, + 480, + 159 + ], + "type": "text", + "content": "23. Lai, Z., Fu, Y.: Mixed attention network for hyperspectral image denoising. arXiv preprint arXiv:2301.11525 (2023) 10, 12" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 160, + 480, + 181 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 160, + 480, + 181 + ], + "spans": [ + { + "bbox": [ + 132, + 160, + 480, + 181 + ], + "type": "text", + "content": "24. Lai, Z., Wei, K., Fu, Y.: Deep plug-and-play prior for hyperspectral image restoration. Neurocomputing 481, 281-293 (2022) 10, 12" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 182, + 480, + 202 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 182, + 480, + 202 + ], + "spans": [ + { + "bbox": [ + 132, + 182, + 480, + 202 + ], + "type": "text", + "content": "25. Lai, Z., Yan, C., Fu, Y.: Hybrid spectral denoising transformer with guided attention. In: ICCV (2023) 10, 12" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 203, + 480, + 224 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 203, + 480, + 224 + ], + "spans": [ + { + "bbox": [ + 132, + 203, + 480, + 224 + ], + "type": "text", + "content": "26. Li, J., Zhang, J., Chen, F., Zhao, K., Zeng, D.: Adaptive material matching for hyperspectral imagery destringing. IEEE TGRS 60, 1-20 (2022) 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 225, + 480, + 245 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 225, + 480, + 245 + ], + "spans": [ + { + "bbox": [ + 132, + 225, + 480, + 245 + ], + "type": "text", + "content": "27. Li, M., Fu, Y., Zhang, Y.: Spatial-spectral transformer for hyperspectral image denoising. In: AAAI (2023) 2, 4, 9, 10, 11, 12" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 246, + 480, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 246, + 480, + 267 + ], + "spans": [ + { + "bbox": [ + 132, + 246, + 480, + 267 + ], + "type": "text", + "content": "28. Li, M., Liu, J., Fu, Y., Zhang, Y., Dou, D.: Spectral enhanced rectangle transformer for hyperspectral image denoising. In: CVPR (2023) 2, 4, 9, 10, 11, 12" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 268, + 480, + 299 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 268, + 480, + 299 + ], + "spans": [ + { + "bbox": [ + 132, + 268, + 480, + 299 + ], + "type": "text", + "content": "29. Lin, J., Hu, X., Cai, Y., Wang, H., Yan, Y., Zou, X., Zhang, Y., Van Gool, L.: Unsupervised flow-aligned sequence-to-sequence learning for video restoration. In: ICML (2022) 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 300, + 480, + 320 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 300, + 480, + 320 + ], + "spans": [ + { + "bbox": [ + 132, + 300, + 480, + 320 + ], + "type": "text", + "content": "30. Liu, T., Wang, K., Sha, L., Chang, B., Sui, Z.: Table-to-text generation by structure-aware seq2seq learning. In: AAAI (2018) 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 321, + 480, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 321, + 480, + 342 + ], + "spans": [ + { + "bbox": [ + 132, + 321, + 480, + 342 + ], + "type": "text", + "content": "31. Lu, G., Fei, B.: Medical hyperspectral imaging: a review. Journal of biomedical optics 19(1), 010901-010901 (2014) 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 342, + 480, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 342, + 480, + 373 + ], + "spans": [ + { + "bbox": [ + 132, + 342, + 480, + 373 + ], + "type": "text", + "content": "32. Maggioni, M., Katkovnik, V., Egiazarian, K., Foi, A.: Nonlocal transform-domain filter for volumetric data denoising and reconstruction. IEEE TIP 22(1), 119-133 (2012) 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 374, + 480, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 374, + 480, + 396 + ], + "spans": [ + { + "bbox": [ + 132, + 374, + 480, + 396 + ], + "type": "text", + "content": "33. Mnih, V., Hinton, G.E.: Learning to detect roads in high-resolution aerial images. In: ECCV (2010) 9, 11, 12" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 396, + 480, + 417 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 396, + 480, + 417 + ], + "spans": [ + { + "bbox": [ + 132, + 396, + 480, + 417 + ], + "type": "text", + "content": "34. Park, J.I., Lee, M.H., Grossberg, M.D., Nayar, S.K.: Multispectral imaging using multiplexed illumination. In: ICCV (2007) 2, 9, 10, 12" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 418, + 480, + 449 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 418, + 480, + 449 + ], + "spans": [ + { + "bbox": [ + 132, + 418, + 480, + 449 + ], + "type": "text", + "content": "35. Peng, J., Xie, Q., Zhao, Q., Wang, Y., Yee, L., Meng, D.: Enhanced 3dtv regularization and its applications on hsi denoising and compressed sensing. IEEE TIP 29, 7889-7903 (2020) 4" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 450, + 480, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 450, + 480, + 480 + ], + "spans": [ + { + "bbox": [ + 132, + 450, + 480, + 480 + ], + "type": "text", + "content": "36. Shi, Q., Tang, X., Yang, T., Liu, R., Zhang, L.: Hyperspectral image denoising using a 3-d attention denoising network. IEEE TGRS 59(12), 10348-10363 (2021) 4" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 482, + 480, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 482, + 480, + 503 + ], + "spans": [ + { + "bbox": [ + 132, + 482, + 480, + 503 + ], + "type": "text", + "content": "37. Sutskever, I., Vinyals, O., Le, Q.V.: Sequence to sequence learning with neural networks. In: NeurIPS (2014) 2, 4" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 504, + 480, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 504, + 480, + 525 + ], + "spans": [ + { + "bbox": [ + 132, + 504, + 480, + 525 + ], + "type": "text", + "content": "38. Wang, H., Xu, K., Lau, R.W.: Local color distributions prior for image enhancement. In: ECCV (2022) 8" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 525, + 480, + 556 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 525, + 480, + 556 + ], + "spans": [ + { + "bbox": [ + 132, + 525, + 480, + 556 + ], + "type": "text", + "content": "39. Wang, W., Jiao, W., Hao, Y., Wang, X., Shi, S., Tu, Z., Lyu, M.: Understanding and improving sequence-to-sequence pretraining for neural machine translation. ACL (2022) 4" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 132, + 557, + 480, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 557, + 480, + 590 + ], + "spans": [ + { + "bbox": [ + 132, + 557, + 480, + 590 + ], + "type": "text", + "content": "40. Wang, Y., Peng, J., Zhao, Q., Leung, Y., Zhao, X.L., Meng, D.: Hyperspectral image restoration via total variation regularized low-rank tensor decomposition. IEEE J-STARS 11(4), 1227-1243 (2017) 3, 10, 12" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 132, + 590, + 480, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 590, + 480, + 622 + ], + "spans": [ + { + "bbox": [ + 132, + 590, + 480, + 622 + ], + "type": "text", + "content": "41. Wang, Z., Ng, M.K., Zhuang, L., Gao, L., Zhang, B.: Nonlocal self-similarity-based hyperspectral remote sensing image denoising with 3-d convolutional neural network. IEEE TGRS 60, 1-17 (2022) 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 132, + 623, + 480, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 623, + 480, + 643 + ], + "spans": [ + { + "bbox": [ + 132, + 623, + 480, + 643 + ], + "type": "text", + "content": "42. Wang, Z., Bovik, A.C., Sheikh, H.R., Simoncelli, E.P.: Image quality assessment: from error visibility to structural similarity. IEEE TIP 13(4), 600-612 (2004) 9" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 132, + 644, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 644, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 644, + 480, + 665 + ], + "type": "text", + "content": "43. Wei, K., Fu, Y., Huang, H.: 3-d quasi-recurrent neural network for hyperspectral image denoising. IEEE TNNLS 32(1), 363-375 (2020) 2, 4, 9, 10, 12" + } + ] + } + ], + "index": 23 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "text", + "content": "J. Xiao et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 567 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 138 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 138 + ], + "type": "text", + "content": "44. Wei, X., Xiao, J., Gong, Y.: Blind hyperspectral image denoising with degradation information learning. Remote Sensing 15(2), 490 (2023) 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 139, + 482, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 139, + 482, + 171 + ], + "spans": [ + { + "bbox": [ + 130, + 139, + 482, + 171 + ], + "type": "text", + "content": "45. Wu, Z., Wang, Q., Wu, Z., Shen, Y.: Total variation-regularized weighted nuclear norm minimization for hyperspectral image mixed denoising. JEE 25(1), 013037-013037 (2016) 9, 10, 11" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 172, + 481, + 193 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 172, + 481, + 193 + ], + "spans": [ + { + "bbox": [ + 132, + 172, + 481, + 193 + ], + "type": "text", + "content": "46. Xiao, J., Ji, Y., Wei, X.: Hyperspectral image denoising with spectrum alignment. In: ACMMM. pp. 5495-5503 (2023) 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 194, + 480, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 194, + 480, + 215 + ], + "spans": [ + { + "bbox": [ + 132, + 194, + 480, + 215 + ], + "type": "text", + "content": "47. Xiao, J., Wei, X.: Hyperspectral image denoising using uncertainty-aware adjustor. In: IJCAI (2023) 2, 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 216, + 480, + 236 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 216, + 480, + 236 + ], + "spans": [ + { + "bbox": [ + 132, + 216, + 480, + 236 + ], + "type": "text", + "content": "48. Xiong, F., Zhou, J., Qian, Y.: Material based object tracking in hyperspectral videos. IEEE TGRS 29, 3719-3733 (2020) 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 237, + 480, + 269 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 237, + 480, + 269 + ], + "spans": [ + { + "bbox": [ + 132, + 237, + 480, + 269 + ], + "type": "text", + "content": "49. Xiong, F., Zhou, J., Zhao, Q., Lu, J., Qian, Y.: Mac-net: Model-aided nonlocal neural network for hyperspectral image denoising. IEEE TGRS 60, 1-14 (2021) 10, 11, 12" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 270, + 480, + 292 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 270, + 480, + 292 + ], + "spans": [ + { + "bbox": [ + 132, + 270, + 480, + 292 + ], + "type": "text", + "content": "50. Xu, S., Zhang, X., Wu, Y., Wei, F.: Sequence level contrastive learning for text summarization. In: AAAI (2022) 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 293, + 480, + 324 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 293, + 480, + 324 + ], + "spans": [ + { + "bbox": [ + 132, + 293, + 480, + 324 + ], + "type": "text", + "content": "51. Yuan, Q., Zhang, Q., Li, J., Shen, H., Zhang, L.: Hyperspectral image denoising employing a spatial-spectral deep residual convolutional neural network. IEEE TGRS 57(2), 1205-1218 (2018) 2, 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 325, + 480, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 325, + 480, + 346 + ], + "spans": [ + { + "bbox": [ + 132, + 325, + 480, + 346 + ], + "type": "text", + "content": "52. Yuan, Y., Ma, H., Liu, G.: Partial-dnet: A novel blind denoising model with noise intensity estimation for hsi. IEEE TGRS 60, 1-13 (2021) 2, 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 347, + 480, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 347, + 480, + 380 + ], + "spans": [ + { + "bbox": [ + 132, + 347, + 480, + 380 + ], + "type": "text", + "content": "53. Yuhas, R.H., Boardman, J.W., Goetz, A.F.: Determination of semi-arid landscape endmembers and seasonal trends using convex geometry spectral unmixing techniques. In: JPL (1993) 9" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 380, + 480, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 380, + 480, + 402 + ], + "spans": [ + { + "bbox": [ + 132, + 380, + 480, + 402 + ], + "type": "text", + "content": "54. Zhang, J., Tu, Z., Yang, J., Chen, Y., Yuan, J.: Mixste: Seq2seq mixed spatiotemporal encoder for 3d human pose estimation in video. In: CVPR (2022) 4" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 402, + 480, + 434 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 402, + 480, + 434 + ], + "spans": [ + { + "bbox": [ + 132, + 402, + 480, + 434 + ], + "type": "text", + "content": "55. Zhang, K., Zuo, W., Chen, Y., Meng, D., Zhang, L.: Beyond a gaussian denoiser: Residual learning of deep cnn for image denoising. IEEE TIP 26(7), 3142-3155 (2017) 7" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 435, + 480, + 467 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 435, + 480, + 467 + ], + "spans": [ + { + "bbox": [ + 132, + 435, + 480, + 467 + ], + "type": "text", + "content": "56. Zhang, Q., Zheng, Y., Yuan, Q., Song, M., Yu, H., Xiao, Y.: Hyperspectral image denoising: From model-driven, data-driven, to model-data-driven. IEEE TNNLS (2023) 1" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 468, + 480, + 489 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 468, + 480, + 489 + ], + "spans": [ + { + "bbox": [ + 132, + 468, + 480, + 489 + ], + "type": "text", + "content": "57. Zhang, T., Fu, Y., Li, C.: Hyperspectral image denoising with realistic data. In: ICCV (2021) 9, 11, 12" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 490, + 480, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 490, + 480, + 511 + ], + "spans": [ + { + "bbox": [ + 132, + 490, + 480, + 511 + ], + "type": "text", + "content": "58. Zhou, S., Zhang, J., Zuo, W., Loy, C.C.: Cross-scale internal graph neural network for image super-resolution. In: NeurIPS. vol. 33, pp. 3499-3509 (2020) 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 512, + 480, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 512, + 480, + 533 + ], + "spans": [ + { + "bbox": [ + 132, + 512, + 480, + 533 + ], + "type": "text", + "content": "59. Zhu, D., Du, B., Zhang, L.: Two-stream convolutional networks for hyperspectral target detection. IEEE TGRS 59(8), 6907-6921 (2020) 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 534, + 480, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 534, + 480, + 567 + ], + "spans": [ + { + "bbox": [ + 132, + 534, + 480, + 567 + ], + "type": "text", + "content": "60. Zhuang, L., Gao, L., Zhang, B., Fu, X., Bioucas-Dias, J.M.: Hyperspectral image denoising and anomaly detection based on low-rank and sparse representations. IEEE TGRS 60, 1-17 (2020) 1" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 255, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 255, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 255, + 91, + 447, + 102 + ], + "type": "text", + "content": "Sequence-to-Sequence Hyperspectral Denoising" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Region-Native Visual Tokenization/a10ef64d-036f-4197-bc89-c920baeb7773_content_list.json b/2024/Region-Native Visual Tokenization/a10ef64d-036f-4197-bc89-c920baeb7773_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..2691759f266187d04d32d1f29588e5b83c5f014b --- /dev/null +++ b/2024/Region-Native Visual Tokenization/a10ef64d-036f-4197-bc89-c920baeb7773_content_list.json @@ -0,0 +1,2213 @@ +[ + { + "type": "text", + "text": "Region-Native Visual Tokenization", + "text_level": 1, + "bbox": [ + 300, + 141, + 702, + 162 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Mengyu Wang $^{1,5(\\dagger)}$ $\\odot$ , Yuyao Huang $^{1,5(\\dagger,\\ddagger)}$ $\\odot$ , Henghui Ding $^{4}$ $\\odot$ , Xinlong Wang $^{2}$ $\\odot$ , Tiejun Huang $^{2}$ , Yao Zhao $^{1,5,6}$ $\\odot$ , Yunchao Wei $^{1,5,6(\\boxtimes)}$ $\\odot$ , and Shuicheng Yan $^{3}$ $\\odot$", + "bbox": [ + 277, + 188, + 725, + 236 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Institute of Information Science, Beijing Jiaotong University, Beijing, China \n $^{2}$ Beijing Academy of Artificial Intelligence \n $^{3}$ Skywork AI", + "bbox": [ + 240, + 244, + 759, + 289 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{4}$ Institute of Big Data, Fudan University, Shanghai, China \n $^{5}$ Visual Intelligence + X International Joint Laboratory of the Ministry of Education \n $^{6}$ PengCheng Laboratory, Shenzhen, China", + "bbox": [ + 215, + 289, + 785, + 330 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. We explore an innovative region-based visual token representation and present the REgion-native AutoencoDER (Reader). In contrast to the majority of previous methods, which represent each image as a grid-shaped tokens map, Reader perceives each image into sequential region-based tokens, with each token corresponding to an object or one part of an object in the image. Specifically, Reader comprises both an encoder and a decoder. The encoder can partition each image into an adaptive number of arbitrary-shaped regions and encode each region into a token. Subsequently, the decoder utilizes this adaptive-length token sequence to reconstruct the original image. Experimental results demonstrate that such region-based token representation possesses two main notable characteristics. Firstly, it achieves highly efficient image encoding. Reader can adaptively use more regions to represent complex areas and fewer regions in simpler ones, thus avoiding information redundancy. Consequently, it achieves superior reconstruction fidelity compared to previous methods, despite using significantly fewer tokens for each image. Secondly, the region-based manner enables manipulation on a local region without causing global changes. As a result, Reader inherently supports diverse image editing operations, including erasing, adding, replacing, and modifying shapes on the objects, and achieves excellent performance in the image editing benchmark of smile transferring. Code is provided at https://github.com/MengyuWang826/Reader.git.", + "bbox": [ + 259, + 359, + 743, + 664 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 686, + 375, + 702 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Tokenization, originating from Natural Language Processing (NLP), involves breaking a textual data stream into several elements called tokens. It plays an important role for converting raw data into a form that is predictable and analyzable by deep learning systems [13, 42]. With the gradual maturity of tokenization technique in the NLP field, it has also been adapted to the Computer", + "bbox": [ + 212, + 714, + 787, + 792 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Internship at BAAI.", + "bbox": [ + 230, + 797, + 387, + 811 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "$\\ddagger$ Internship at Skywork AI.", + "bbox": [ + 230, + 811, + 424, + 825 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "Corresponding author.", + "bbox": [ + 230, + 825, + 406, + 839 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/d4c748bfbeb5fe9442a095a84505b3facf700546a43524df61b7e95f775eab59.jpg", + "image_caption": [ + "(a) Image" + ], + "image_footnote": [], + "bbox": [ + 228, + 143, + 356, + 265 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/38b04be8ce5feeda681d079903de032e8014071e4115c059ba7054a19bb5c483.jpg", + "image_caption": [ + "(b) Grid-based" + ], + "image_footnote": [], + "bbox": [ + 367, + 143, + 495, + 265 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/922722d1bdf5da29da6d2e8202bbd3376d8e91679c64782064ef824e68817403.jpg", + "image_caption": [ + "(c) Object-based", + "Fig.1: Illustration of different types of visual tokenization methods. Grid-based methods [16,56] represent images as grid-shaped tokens map. Object-based methods [1,59] use detection models to represent objects as tokens. While our region-based method represents arbitrarily shaped regions as tokens, which offers more flexibility and retains the naturally existed boundaries presented in images." + ], + "image_footnote": [], + "bbox": [ + 506, + 143, + 633, + 265 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/1250be1d8897bed585b3d0e2a6ac2e6db7f4cb06a5cdfe6367efc60dbccc1b4d.jpg", + "image_caption": [ + "(d) Region-based" + ], + "image_footnote": [], + "bbox": [ + 645, + 143, + 772, + 265 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Vision (CV) domain, known as visual tokenization, and has become the foundation for many tasks such as image/video generation [16, 27, 56] and multimodal generation/perception [11, 14, 18, 38].", + "bbox": [ + 212, + 378, + 784, + 425 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Existing visual tokenization methods can be categorized into two main types, which we refer to as grid-based and object-based. Grid-based methods typically employ a CNN [23] or ViT [15] backbone to encode the original image into a downsampled grid-shaped feature map, followed by quantization of the feature map to get the image tokens. However, in such methods, each token corresponds to a patch in the original image (see Fig. 1b), which does not adhere to the following characteristics of the image prior. Firstly, different areas in images have varying information densities. While in these methods, patches are treated equally, leading to information loss in densely informative patches and redundancy in sparsely informative ones. Secondly, areas with different semantics within an image have natural boundaries. While employing patches as the units of image tokens disregards the natural boundaries, causing local information within the original image to be distributed and coupled across multiple tokens, limiting the extraction and utilization of local features. Object-based methods [1,35,51,53,59], as an alternative and more natural approach, typically appear in vision-language pretraining. These methods employ detection or segmentation models (see Fig. 1c) to encode the objects within images into tokens, which are then interacted with or aligned with language information. However, these methods only focus on extracting high-level semantic information, neglecting low-level visual details in the image. Consequently, they cannot be used for tasks that require image outputs, such as image synthesis and editing.", + "bbox": [ + 212, + 428, + 787, + 746 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "These challenges suggest that, an ideal visual token representation should take into account the crucial visual-prior characteristics, including the uneven information density and the inherent semantic boundaries, and further achieve local-information decoupling and global-information compression. However, previous practices either ignore the visual-prior characteristics or, due to their coarse granularity, result in significant information loss. To address these problems, we", + "bbox": [ + 212, + 750, + 787, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "Wang et al.", + "bbox": [ + 271, + 114, + 349, + 128 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/5270a8fc731de1bbd8f72d290eee3e5e0d29ec9c7742f92e389c4989eaaf4407.jpg", + "image_caption": [ + "Erasing the two people" + ], + "image_footnote": [], + "bbox": [ + 217, + 143, + 313, + 199 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/b20413719837a8e504e08eb997d723e574f4c1f86223e040aca868cbe77903b7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 313, + 143, + 406, + 199 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/411afb2488cddc08e64f14916b58abcad901e89f1cca17e5d72a9c294bc4162d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 406, + 143, + 501, + 199 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/2257e285b40fe8cc01d06188e4793d0df4fb9de90df4cb121966ac65c0f4ef56.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 501, + 143, + 594, + 199 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/e7620b9cc2feef062ab16ea4037ece79f3a495db031d6a6098830c823e54717b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 596, + 143, + 687, + 199 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/295ce50e3e49d94eb6a0f37f6950586437262c29321bb0b53e1670090571b9c6.jpg", + "image_caption": [ + "mushroom" + ], + "image_footnote": [], + "bbox": [ + 689, + 143, + 784, + 199 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/e5874e47f54b65f77e6652ee29469d17e8d8717534f2fc772e9776b5be12935f.jpg", + "image_caption": [ + "Slimming the jawline and enlarging the eyes", + "Fig. 2: Examples of the proposed Reader performing various editing operations. Reader showcases its strengths in flexibly editing multiple attributes of existing real images under several interactive operations (see Sec. 3.4). In each pair, the left displays the original image, the right displays the edited result and the text describes the corresponding editing objective." + ], + "image_footnote": [], + "bbox": [ + 217, + 208, + 313, + 280 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/23e9da7a1aa570467727496c9feb26855ddb55174e60ecd306ff5bd41c203ef9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 313, + 208, + 406, + 280 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/af48cf96e9542c7e1854a45f9fa75787598b9a26d4c3b26b0c77d4d2b1c55e7a.jpg", + "image_caption": [ + "Non-smiling expressions to smiling expressions" + ], + "image_footnote": [], + "bbox": [ + 408, + 208, + 501, + 280 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/54825d6133390800e1a0171abe75761903599596fb1703bb6d12d0e5247d1a7f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 501, + 208, + 594, + 280 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/6b73c694b020ec4ff185f59699e65788f7e71a913010901b7c57739583ff7329.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 596, + 208, + 689, + 280 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "propose the REGion-native AutoencoDER (Reader), which takes arbitrarily shaped and sized regions (see Fig. 1d) as the fundamental units for visual tokens, providing more flexibility and aligning better with the visual-prior characteristics. The Reader framework conceptualizes each image as a composite of several regions, with each region corresponding to either a small-sized object or a portion of a regular or large-sized object.", + "bbox": [ + 212, + 383, + 784, + 474 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Specifically, the Reader architecture comprises an encoder and a decoder, as illustrated in Fig. 3. The encoder first partitions an image into an adaptive-number of non-overlapping regions, where each region is represented as a binary mask and an embedding. Following this, the encoder employs a novel region embedder to integrate the geometric information from the mask with the visual features from the embedding, and produce the desired token for each region. Subsequently, a novel cross-attention-based decoder is employed to reconstruct the original image from these tokens.", + "bbox": [ + 212, + 476, + 784, + 595 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We further conduct both qualitative and quantitative experiments and demonstrate the following two main notable characteristics of Reader:", + "bbox": [ + 212, + 597, + 787, + 626 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Efficiency: Reader achieves highly efficient encoding, yielding reconstruction fidelity superior to that of previous methods, despite employing a significantly smaller number of tokens. This is intuitive since the region-based manner enables Reader to naturally represent complex areas with more regions, while representing simpler areas with fewer regions. This effectively avoids information redundancy, thus possessing inherent efficiency.", + "bbox": [ + 212, + 628, + 784, + 717 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Disentanglement: Representing non-overlapping regions as tokens guarantees their disentanglement from one another, and further empower Reader with the ability to perform image editing. Taking object erasure as an example, the disentanglement permits an object in the image to be represented solely by the tokens corresponding to its interior. Through removing these tokens and then reconstructing the image, object erasing can be naturally achieved. Consequently, without any additional modules or training data, Reader facilitates diverse object-level editing operations (see Fig. 2), such as erasing, adding, re", + "bbox": [ + 212, + 719, + 785, + 839 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "Region-Native Autoencoder", + "bbox": [ + 544, + 114, + 730, + 127 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "placing, and modifying shapes on objects, and demonstrates excellent performance in the image editing benchmark of smile transferring [31].", + "bbox": [ + 212, + 146, + 782, + 176 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We believe that these experiments effectively demonstrate the inherent advantages of region-based tokenization method. We hope this will bring some novel insights to the community about visual tokenization and unlock its value in various downstream tasks in the future, such as interaction with LLMs and multimodal perception, generation, and beyond.", + "bbox": [ + 212, + 178, + 787, + 253 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 284, + 387, + 301 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Grid-based Tokenization The majority of grid-based methods follows the feature extraction, quantization, and reconstruction paradigm, initially introduced by VQVAE [39]. These methods represent image tokens as a grid-shaped code map. Expanding upon this framework, VQGAN [16] incorporates perceptual loss [58] and GAN loss [28] to augment the original L1 reconstruction loss, with the aim of improving both reconstruction fidelity and codebook learning. RQ-VAE [32] introduces residual quantization to further enhance reconstruction fidelity, while leading to the cost of significantly higher number of tokens. ViT-VQGAN [56] replaces the previously employed CNN-based architecture with a ViT-based one. Sharing a similar motivation to ours, DQ-VAE [27] proposes dynamically selecting patch granularity between two or three pre-defined strides to achieve dynamic-length coding. However, this approach remains constrained by the grid-based paradigm. In contrast, our Reader takes arbitrarily shaped regions as image token units and offers greater flexibility.", + "bbox": [ + 212, + 328, + 787, + 541 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Object-based Tokenization Methods of this kind typically find applications in vision-language pretraining and multimodal perception. For instance, [1] uses a Faster R-CNN [44] for proposals prediction and encodes the proposals into tokens to subsequently perform image captioning or visual question answering. Similarly, [59] also employs a Faster R-CNN to extract object tokens and maps these tokens with a Graph Neural Network [57]. GroupViT [53] proposes a learnable clustering architecture based on ViT [15] to extract object information, and aligns it with text embedding akin to CLIP [41]. However, these methods focus on extracting semantic information at object-level, leading to excessively coarse-grained and overlooking on detailed visual information. In contrast, our Reader employs an autoencoder architecture to extract dynamic-grained local regions as the unit of image tokens, which better preserves detailed information and showcases superior performance in image reconstruction and editing.", + "bbox": [ + 212, + 546, + 787, + 743 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Scene Decomposition The scene decomposition task [4,37,55], also referred to as object-centric representation, has the most similar form to our Reader. This task aims to decompose a scene into multiple objects and learn the representation of each object. However, these works focus on application in some synthetic datasets containing only basic 3D shapes [20,29,43]. In contrast, our Reader can be applied to a wide range of natural images.", + "bbox": [ + 212, + 750, + 787, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 230, + 127 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "Wang et al.", + "bbox": [ + 271, + 114, + 349, + 128 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Image Editing Most early image editing methods are GAN-based [21]. Some of these approaches [22, 34, 40, 60] rely on the controllability of the GAN's latent space, employing supervised or unsupervised approaches to achieve desired transformations of GAN's latent variable. Other works [8, 31, 52] utilize GANs' conditional generation ability and synthesize the edited images based on the input conditions. The remarkable success of diffusion models [25, 48] in image synthesis has also led to the emergence of numerous diffusion-based editing methods [3, 10, 19, 47, 54]. These approaches largely build upon the robust semantic understanding and rendering capabilities of large pre-trained diffusion models [45].", + "bbox": [ + 212, + 146, + 782, + 282 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The semantic-map based methods [19, 31, 34] can produce the most similar effects to our Reader. However, compared to these works, our Reader benefits from its region-based representation, enabling a wider variety of editing operations without requiring additional input, with details elaborated in Sec. 3.4.", + "bbox": [ + 212, + 282, + 784, + 342 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3 Methodology", + "text_level": 1, + "bbox": [ + 215, + 367, + 380, + 386 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We propose Reader (region-native autoencoder) in this work, which performs region-based visual tokenization and enables diverse controllable image editing operations. In the following, we will first provide detailed descriptions about the model architecture of the Reader's encoder and decoder in Sec. 3.1 and Sec. 3.2, respectively. Sec. 3.3 outlines the objective function utilized, while Sec. 3.4 delves into the detailed implementation for image editing. The model architecture is illustrated in Fig. 3.", + "bbox": [ + 212, + 400, + 784, + 505 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1 Encoder Architecture", + "text_level": 1, + "bbox": [ + 215, + 529, + 441, + 544 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The encoder of Reader needs the capability to first segment an image into numerous regions and then encode each region into a token. Thus, we propose a encoder with three core components: an image feature extractor, a region decoder and a region embedder.", + "bbox": [ + 212, + 555, + 782, + 614 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Feature Extractor To extract information from the original image for subsequent image segmentation, tokenization, and reconstruction steps, we employ a standard ViT [15] along with a simple FPN [33] as our feature extractor. Given the importance of fine-grained details for high-fidelity reconstruction, we configure the patch size of the ViT to be $8 \\times 8$ , a smaller size compared to conventional settings, yet effective in preserving local structure. The strides of the simple FPN are set to $(1, \\frac{1}{2})$ , thus the output features are at resolutions of $\\frac{1}{8}$ and $\\frac{1}{4}$ of the original image.", + "bbox": [ + 212, + 622, + 782, + 743 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Region Decoder We employ a region decoder to learn the region representation and segment image into regions. Specifically, we employ a standard Transformer decoder [50], which is similar to previous Transformer-based segmentation works [5-7,36], augmented with an additional bidirectional Transformer block [30].", + "bbox": [ + 212, + 750, + 782, + 809 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In the Transformer decoder, $N$ learnable embeddings are used as input queries to extract the region information, termed region queries. Within each", + "bbox": [ + 212, + 809, + 782, + 839 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "Region-Native Autoencoder", + "bbox": [ + 544, + 114, + 730, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "layer of the Transformer decoder, one of the two-sized image features alternately serves as the key and value. Through cross-attention, the region queries can extract region-level information from the image features. Subsequently, the bidirectional Transformer block takes the updated region queries and the larger-sized image feature as input to perform both region-to-image and image-to-region attention and update them. This process yields the region embeddings $\\mathcal{E}_r \\in \\mathbb{R}^{N \\times c}$ and a region-aware image feature $\\mathcal{E}_{img} \\in \\mathbb{R}^{c \\times h \\times w}$ . Here, $(h, w)$ represents the resolution of the image feature, which is $\\frac{1}{4}$ of the original image size, and $c$ denotes the channel dimension.", + "bbox": [ + 212, + 146, + 787, + 281 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We then compute the similarity matrix between the region embeddings and the image feature via a dot product, formulated as:", + "bbox": [ + 214, + 282, + 785, + 313 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nS = \\mathcal {E} _ {r} \\odot \\mathcal {E} _ {i m g} \\in \\mathcal {R} ^ {N \\times h \\times w}. \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 406, + 320, + 785, + 339 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We use this similarity matrix to get the desired regions. To prevent information redundancy caused by overlapping, we apply an argmax operation along the first dimension of the similarity matrix. This process means matching each pixel with one embedding that has the highest similarity to the pixel's feature, which can be formulated as:", + "bbox": [ + 214, + 345, + 787, + 421 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nS _ {\\text {h a r d}} = \\operatorname {o n e - h o t} (\\operatorname {a r g m a x} (S)) \\in \\{0, 1 \\} ^ {N \\times h \\times w}. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 344, + 429, + 785, + 446 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Consequently, the $S_{\\mathrm{hard}}$ in Eq. (2) can also be seen as $N$ binary masks, where the foreground of each mask indicates which pixels belong to that specific region. However, not each one among the $N$ masks contains foreground. In fact, in some simple images, numerous pixels may be matched with a single region embedding, while other region embeddings remain unmatched with any pixel (illustrated as \"used\" and \"unused\" in Fig. 3). Therefore, while $N$ is a fixed hyper-parameter, the actual number of regions corresponding to each image is dynamic, determined by the complexity of the image.", + "bbox": [ + 212, + 455, + 787, + 575 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Region Embedder Through the region decoder, the original image is represented as several regions. Each region comprises a region embedding $\\mathcal{E}_r[i]$ representing the visual feature and a binary mask $S_{\\mathrm{hard}}[i]$ representing the geometric information (e.g., shape, position and scale). The region embedder is further employed to merge the above two components and produce the desired tokens.", + "bbox": [ + 212, + 582, + 787, + 657 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Following GroupViT [53], we first aggregate the feature of each pixel into the corresponding region embedding. To make the entire process differentiable, enabling the gradients of image reconstruction to guide the granularity of regions, we employ the straight-through trick [39] here. This process is formulated as:", + "bbox": [ + 212, + 657, + 787, + 719 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nS = \\operatorname {s o f t m a x} (S, \\dim = 0), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 382, + 727, + 784, + 743 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {S} = S _ {\\text {h a r d}} + S - \\operatorname {s g} \\cdot (S), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 382, + 748, + 784, + 765 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathcal {E}} _ {r} [ i ] = \\mathcal {E} _ {r} [ i ] + \\sum_ {j, k} \\hat {S} [ i, j, k ] \\cdot \\mathcal {E} _ {i m g} [:, j, k ], \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 362, + 768, + 784, + 801 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "in which the $\\mathrm{sg}\\cdot ()$ refers to stop gradient. While the pixel features aggregation process shown in Eq. (5) is actually a pooling process, leading to over-smoothing", + "bbox": [ + 214, + 809, + 785, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "Wang et al.", + "bbox": [ + 271, + 114, + 349, + 128 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/64d85e7b2d3ad9641b2068c725a153eea26732b31626e306a51131a4f52d987c.jpg", + "image_caption": [ + "Fig. 3: An overview of the proposed Reader. It comprises an encoder and a decoder. The encoder can segment each image into an adaptive number of regions, which is illustrated as the \"used\" part, and embeds each region into a token with the Region Embedder. The decoder then reconstructs the original image from these tokens with the learnable patch queries." + ], + "image_footnote": [], + "bbox": [ + 215, + 145, + 785, + 303 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "on local details. We additionally employ a masked cross-attention layer [6] to enhance the visual feature extraction. This layer uses the region embeddings $\\hat{\\mathcal{E}}_r$ as queries to perform attention to the image feature $\\mathcal{E}_{img}$ only in the foreground of each region in order to extract more local details.", + "bbox": [ + 212, + 395, + 782, + 455 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Subsequently, we incorporate a module to explicitly encode the geometric information in the masks and integrate it into the region embeddings. This approach serves two primary purposes: Firstly, the inclusion of geometric information enhances the delineation of boundaries between regions, thereby facilitating image reconstruction. Secondly, such explicit encoding also enables more editing operations in subsequent steps. Specifically, we flatten each mask and utilize a linear layer to project it into a latent space. Additionally, the region embeddings are also projected with an MLP and are then fused with the former projection through element-wise addition. Following a normalization and another MLP, the region embedder produces the desired region-based image tokens.", + "bbox": [ + 212, + 455, + 787, + 608 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.2 Decoder Architecture", + "text_level": 1, + "bbox": [ + 214, + 631, + 441, + 645 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The encoder of Reader represents each image as an adaptive-length sequence. Consequently, the decoder needs the capacity to decode this adaptive-length sequence into a uniformly-sized image. This poses a challenge of mapping variable-length input to a fixed-length output. To address this, the decoder architecture of Reader is based on the Transformer decoder, leveraging the cross-attention mechanism to resolve the aforementioned problem.", + "bbox": [ + 212, + 657, + 784, + 748 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Specifically, we introduce $M$ learnable embeddings as the input queries for the Transformer decoder. Each embedding is responsible for reconstructing a local patch in the output image and is referred to as a patch query. The queries are also incorporated with the positional embeddings of the center coordinate of the patches. To prevent information loss and enable the Transformer decoder to focus on different information at each layer, we employ a separate MLP for each", + "bbox": [ + 212, + 750, + 787, + 840 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "Region-Native Autoencoder", + "bbox": [ + 544, + 114, + 730, + 127 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 114, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "layer to project the image tokens. These projections then serve as the key and value for the corresponding layer.", + "bbox": [ + 212, + 146, + 782, + 176 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Through the Transformer decoder, the patch queries interact with the image tokens through cross-attention, gathering information related to the corresponding patches and updating their values. After the Transformer decoder, we integrate an linear layer to finally map the output patch queries to the pixel space, which is a reverse process of the patch embedder of ViT.", + "bbox": [ + 212, + 176, + 784, + 252 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "3.3 Objective Function", + "text_level": 1, + "bbox": [ + 214, + 268, + 419, + 285 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The objective functions related to the image reconstruction are kept consistent with VQGAN [16], which is a combination of the L1 loss, perceptual loss [58] and a patch-based GAN loss [28], denoted as $\\mathcal{L}_{img}$ .", + "bbox": [ + 212, + 290, + 782, + 335 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "However, in our observation, relying solely on the image reconstruction loss is insufficient for training the encoder to produce semantically meaningful regions. In practice, the regions easily collapse into uniformly distributed patches. Therefore, additional supervision on the output regions becomes necessary. We employ either the pseudo-annotations produced by SAM [30] or manual annotations to supervise the output regions in different experiments, with detailed information provided in Sec. 4.", + "bbox": [ + 212, + 335, + 784, + 439 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In order to train the similarity matrix in Eq. (3) to resemble a one-hot form at each pixel, we employ a cross-entropy loss here. Unlike the tasks, where cross-entropy typically used with defined category IDs, there is no fixed correspondence between the predicted regions and the masks of annotations here. Following [5,7], we employ a Hungarian-algorithm based bipartite matching to determine the matching relationship, with more details provided in supplementary material. The overall objective function is formulated as:", + "bbox": [ + 212, + 441, + 784, + 546 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathcal {L} _ {i m g} + \\mathcal {L} _ {c e}. \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 442, + 554, + 784, + 570 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "3.4 Image Editing Implementation", + "text_level": 1, + "bbox": [ + 214, + 585, + 516, + 602 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The proposed Reader inherently facilitates diverse image editing operations. Without any additional modules, this is achieved by simply manipulating the image tokens outputted by the encoder and then inputting these manipulated tokens into the decoder to get the edited image.", + "bbox": [ + 212, + 607, + 782, + 666 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Specifically, before the region embedder of Reader, each region is represented by two components: the geometric information (represented as a binary mask) and the visual information (represented as the region embedding). This enables independent manipulations on both two parts and achieves different editing effects. Therefore, we define the following three-level image editing operations. A schematic diagram is shown in Fig. 4 to provide an intuitive understanding.", + "bbox": [ + 212, + 667, + 784, + 758 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Geometry Level These operations refer to only manipulating the masks through affine transformations or interactive brushing by users, while maintaining the embeddings unaltered. Geometry-level operations allows to perform moving, scaling, and shape modification on objects. As an example, the left branch in Fig. 4 alters the shape of the lips without changing their color.", + "bbox": [ + 214, + 763, + 787, + 839 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "Wang et al.", + "bbox": [ + 271, + 114, + 349, + 128 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/befe1cac5f5f0d4bf3fdf961ce3fae5bcca9d14164161864e437a0a5d07f55e0.jpg", + "image_caption": [ + "Fig. 4: Illustration of editing operations. Different operations on the embeddings and masks leads to different editing effects. The left, middle and right branches correspond to the three-levels operations described in Sec. 3.4, respectively." + ], + "image_footnote": [], + "bbox": [ + 215, + 143, + 787, + 282 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Feature Level Feature-level operations maintain the mask unchanged while replacing the corresponding embedding with one from another object or another image. This facilitates local feature transferring without impacting geometric characteristics. An example is illustrated in the middle branch of Fig. 4, where the color of lips is changed by replacing the corresponding two embeddings.", + "bbox": [ + 212, + 349, + 782, + 426 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Token Level The erasure, addition and replacement of objects are categorized as this level. Erasure can be achieved by removing tokens corresponding to a specific object. Conversely, addition can be accomplished by appending tokens from another image to the tokens of the current image. Object replacement can then be achieved by combining both operations. Furthermore, these operations can also be combined with the geometry-level manipulations to achieve addition and replacement on objects while controlling their shapes, scale and position, etc. An example is illustrated in the right branch of Fig. 4, where the mouth and lips are replaced by those from another image.", + "bbox": [ + 212, + 431, + 787, + 568 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 215, + 592, + 375, + 609 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.1 Experiment on ImageNet", + "text_level": 1, + "bbox": [ + 215, + 625, + 472, + 641 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Herein we train Reader on ImageNet training set and report qualitative results in image editing and quantitative comparisons on reconstruction fidelity.", + "bbox": [ + 212, + 652, + 782, + 683 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Implementation As outlined in Sec. 3.3, additional supervision is essential to prevent the collapse of producing regions. Consequently, in this experiment, we employ SAM [30] to produce pseudo-annotations on the ImageNet training set. Specifically, we keep the original settings of the segment anything mode of SAM, which effectively segments all prominent objects or local parts in images. Since there is substantial overlapping in the original results of SAM, while the cross-entropy loss needs non-overlapping annotations, we implement the following two post-processing steps. Firstly, removing all larger masks that encompass other ones, retaining only smaller, more localized masks. Secondly, regions not encompassed by the pseudo annotations do not contribute to the objective function.", + "bbox": [ + 212, + 688, + 787, + 840 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "Region-Native Autoencoder", + "bbox": [ + 544, + 114, + 730, + 127 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 114, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/9e205f9a1b59472ba7e1d89678a225844b2ae1d826cda4eb442548f3cc89b592.jpg", + "image_caption": [ + "Fig. 5: Qualitative results in ImageNet validation set and Internet images. The editing objective is labeled with text under image." + ], + "image_footnote": [], + "bbox": [ + 215, + 142, + 787, + 368 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "The input and output sizes of images are set to $(256 \\times 256)$ in this experiment. Both the ViT in the encoder and the Transformer in the decoder follow the configuration of ViT-tiny, which utilizes 192 channels and 3 heads in the multi-head attention. The dimension of image tokens is set to 256 to maintain consistency with the previous methods. The number $N$ of learnable region queries in the encoder is set to 256, which is the minimal tokens number of each image in previous works and serves as the maximum regions number of our model. The number $M$ of learnable patch queries in the decoder is set to 1024, with each query responsible for reconstructing a $8 \\times 8$ patch in the output image. For more implementation details, please refer to our supplementary material.", + "bbox": [ + 212, + 433, + 787, + 585 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Results We provide several image editing examples in Fig. 5 corresponding to all the three-level editing operations described in Sec. 3.4. For example, \"sticking out the tongue\" needs only manipulation on the tongue's mask, \"replacing the color\" is achieved by replacing the region embedding of the strawberry and \"replacing with another hat\" needs deleting the tokens of the original hat and replacing them with new ones from another image.", + "bbox": [ + 212, + 590, + 787, + 681 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "For quantitative comparison of the reconstruction fidelity, we employ both the distribution-level metric rFID [24] and the image-patches-level metrics PSNR, SSIM [26], LPIPS [58] to evaluate the reconstructed images of ImageNet validation set. We employ the VQGAN [16], RQ-VAE [32] and DQ-VAE [32] as baselines. The former two employ fixed tokens number of each image, while the latter uses dynamic tokens number through picking one granularity for each patch from three pre-defined strides $\\left(\\frac{1}{8},\\frac{1}{16},\\frac{1}{32}\\right)$ . The tokens number of DQ-VAE is computed by the ratio of different strides provided in their paper. Tokens number of our Reader corresponds to the average number observed across the entire ImageNet validation set.", + "bbox": [ + 212, + 689, + 787, + 840 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "Wang et al.", + "bbox": [ + 271, + 114, + 349, + 128 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/e161aa6587dc685d67210ad941d2bf0363993b4fb9122e16219a2b2179990b93.jpg", + "table_caption": [ + "Table 1: Quantitative results for the reconstruction fidelity on the ImageNet validation set. \"Tokens\" refers to the number of image tokens utilized for each image. The dimension of image tokens is 256 for all methods. The \"+\" denotes data excerpted from the original paper. Top 1 and top 2 results are highlighted in red and blue, respectively." + ], + "table_footnote": [], + "table_body": "
MethodTokensPSNR ↑SSIM ↑LPIPS ↓rFID ↓Params(M)FLOPs(G)
VQGAN(1024)16 × 1619.360.480.318.0572.14196.33
VQGAN(16384)16 × 1619.830.490.284.9276.07197.33
RQ-VAE(16384)†8 × 8 × 210.77
RQ-VAE(16384)8 × 8 × 420.030.500.285.24104.40193.64
DQ-VAE(16384)†2564.08 - 7.32
Reader(ours)79.0924.750.720.202.4326.6040.24
", + "bbox": [ + 223, + 212, + 769, + 325 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Results shown in Tab. 1 indicate that, in comparison to the baselines, our Reader attains superior reconstruction fidelity despite having the smallest tokens number, model parameters and computational complexity. While compared to DQ-VAE, which also employs dynamic length encoding, our method also demonstrates significant superiority. This finding is intuitive that, compared to both the fixed- and dynamic-sized patches, arbitrarily shaped regions can better handle the uneven information density in images and avoid information redundancy, thus achieving higher reconstruction fidelity with fewer tokens.", + "bbox": [ + 212, + 353, + 787, + 474 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4.2 Experiment on Human Face", + "text_level": 1, + "bbox": [ + 214, + 498, + 495, + 513 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "To further evaluate the editing capabilities of our Reader, we transfer it to images of human faces, which are commonly employed in previous image editing works, allowing to provide quantitative comparisons.", + "bbox": [ + 212, + 523, + 785, + 571 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Implementation We fine-tune the Reader in this experiment, which is pretrained as outlined in Sec. 4.1. We utilize the CelebAMask-HQ dataset [31], a face parsing dataset with manual annotations. The input and output sizes of images for this experiment are configured as $(512 \\times 512)$ to align with the previous methods [31, 34]. In order to preserving the knowledge already learned by the pre-trained model, we keep the majority of Reader architecture unchanged and make the following modifications.", + "bbox": [ + 212, + 575, + 787, + 681 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "(1) Adding a downsampling-and-convolution layer with a residual connection in the region embedder and initializing the new convolution's output to zero. This is employed to make the double-sized masks meet the input length of previous mask projection layer. (2) Increasing the number of learnable patch queries from 1024 to 4096 in order to get double-sized reconstructed images. While loading the pre-trained weights, we perform bicubic interpolation on the weights of the positional embeddings and the patch queries to match the new size.", + "bbox": [ + 212, + 683, + 787, + 789 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Results In Fig. 6, we present the qualitative results corresponding to the three-level editing operations. In the top and middle rows, we performed overlay editing, gradually changing one part of the geometric and visual properties at a", + "bbox": [ + 212, + 794, + 787, + 842 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "Region-Native Autoencoder", + "bbox": [ + 544, + 114, + 730, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/0843069c29102297a94cb9bdcfce2ac88a13bf57ab5834d64be5d15020008956.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 217, + 143, + 312, + 217 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/85c806a5f9d63341ca97c53c51c8940f4e65c13cb4ac7b427a727ebfbd0e349b.jpg", + "image_caption": [ + "+ Close Mouth" + ], + "image_footnote": [], + "bbox": [ + 313, + 143, + 406, + 217 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/9901f7026a6bd0ed496a2d301877969a2496363d27789a93412244ce7d048c79.jpg", + "image_caption": [ + "+ Sharp Jawline" + ], + "image_footnote": [], + "bbox": [ + 408, + 143, + 500, + 217 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/564cc1003e0345133e3002ac21294de28a7502234c630d9d7915746ede7adb58.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 501, + 143, + 594, + 217 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/b2fb8084a0e25d74b0507884c48fb19e6164a3ef0ebfe272cfddf1a72dcc2164.jpg", + "image_caption": [ + "+ Enlarge Nose" + ], + "image_footnote": [], + "bbox": [ + 596, + 143, + 689, + 217 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/91d5fb0201234a8e9846d18ecfdbbe05dd6993b4cdceec07397f2c9fed2257fb.jpg", + "image_caption": [ + "+Wide Jawline" + ], + "image_footnote": [], + "bbox": [ + 692, + 143, + 785, + 217 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/7977c98a981c09c9856ecc980d374a632993124a4b2e063c558ca73763ef1627.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 217, + 224, + 313, + 297 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/0e2567278f8901560a6dc53f6c429b20a27c9f4a57e77780827287b6e3e1d39a.jpg", + "image_caption": [ + "+ Eyes Feature" + ], + "image_footnote": [], + "bbox": [ + 313, + 224, + 406, + 297 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/2d916f4abdc3cdb915cdbbf64078d7b166603006fe77c01ead5e250e1654db83.jpg", + "image_caption": [ + "+ Lips Featue" + ], + "image_footnote": [], + "bbox": [ + 408, + 224, + 500, + 297 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/2c1d0ec5a0a5073222d5b857650115b8c2fcb1a8b40b5ccf5522e4b588d7434b.jpg", + "image_caption": [ + "+ Eyes Feature" + ], + "image_footnote": [], + "bbox": [ + 501, + 224, + 594, + 297 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/1318f74c05214260178b920946013c4e9190701061041328e809781210225f53.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 596, + 224, + 689, + 297 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/22775f61615154e2fba5987b10510adb5b1ebf9a20f32efa001b497131235eb5.jpg", + "image_caption": [ + "+ Lips Featue" + ], + "image_footnote": [], + "bbox": [ + 692, + 224, + 785, + 297 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/b887752c7c9b846802e32352e1b3971500911bf57a603d765b4626266817d3dc.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 217, + 308, + 312, + 380 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/da5a91604be712f05143c8028fb883de9127e87770ed56d7941eae1c66d815ee.jpg", + "image_caption": [ + "+Smiling" + ], + "image_footnote": [], + "bbox": [ + 313, + 308, + 406, + 380 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/eb9e09e5a26800d3ef79050ee06904dcc3f1380e93b1e752f57910fdc04bc6b5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 408, + 306, + 500, + 380 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/645e4dbf9be656c983b39f5d438f841d8086cc72b4de513acaaed89a09870448.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 500, + 306, + 591, + 380 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/7d89dad10c48dbcfa0c07f65866bb69638af4746c59c75b382aaee427e90bfe2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 596, + 306, + 689, + 380 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/d58c5f568a5d00bfa6884e4c9dc1ea43241b564c020ab262652f8eb76a254976.jpg", + "image_caption": [ + "+Smiling" + ], + "image_footnote": [], + "bbox": [ + 692, + 306, + 785, + 380 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/377ad898242c7c411bb7dc163c84e923225bf5150e9e1d59bee996e05dc4f3f5.jpg", + "image_caption": [ + "Fig. 6: Qualitative results in huaman faces. The editing objective is labeled with text under corresponding image.", + "Original Image", + "Fig. 7: Qualitative comparisons in the smile transferring benchmark. Our Reader can selectively manipulate only the target attributes without affecting other features or regions. The red boxes highlight the local details that should remain unchanged after editing. Please kindly zoom in for a better view." + ], + "image_footnote": [], + "bbox": [ + 217, + 426, + 310, + 500 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/3fd414cec35e5f5ef03a59f22a8b0ec6fd7ca011b28cbc52dfbc3f2fab03349c.jpg", + "image_caption": [ + "MaskGAN" + ], + "image_footnote": [], + "bbox": [ + 313, + 426, + 406, + 500 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/9757fecff33a3c356261405a6648fc98a3d69d84f70b30b9863245cd9e790551.jpg", + "image_caption": [ + "Ours" + ], + "image_footnote": [], + "bbox": [ + 408, + 426, + 500, + 500 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/db08d64c90318e1d6809f5705a42eb384e85cb713e68763988d8a106e6ae1166.jpg", + "image_caption": [ + "Original Image" + ], + "image_footnote": [], + "bbox": [ + 501, + 426, + 594, + 500 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/a370999f17a3d1eee7d93661f343367582bdd7434fd1c5a22f2729da55f6ade2.jpg", + "image_caption": [ + "MaskGAN" + ], + "image_footnote": [], + "bbox": [ + 596, + 426, + 689, + 500 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/3e9c9200f8e376f5881b55f4833de76afbfdb55a8dbae8b65785ad6f849a624d.jpg", + "image_caption": [ + "Ours" + ], + "image_footnote": [], + "bbox": [ + 692, + 426, + 785, + 500 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "time. In the bottom row, we simultaneously manipulated both the masks and embeddings, facilitating significant change in expression.", + "bbox": [ + 212, + 594, + 784, + 625 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "To quantitatively measure Reader's image editing capabilities, we utilize the smile transferring benchmark introduced by MaskGAN [31], which requires converting the faces with neutral expressions into smiling faces. We compare our Reader with the MaskGAN [31], LocalEditing [9], InterFaceGAN [46] and EditGAN [34], which also provide results in this benchmark. We employ the following metrics to keep consistent with [34].", + "bbox": [ + 212, + 625, + 784, + 715 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "(1) Semantic Correctness: Training a binary classifier to recognize the smile attribute and evaluating whether the edited faces exhibit a smiling expression. (2) Distribution-level Image Quality: Utilizing FID [24] and KID [2] to assess the quality of edited images. (3) Identity Preservation: Leveraging a pretrained ArcFace feature extraction network [12] to evaluate whether the face's identity is maintained post-editing.", + "bbox": [ + 212, + 717, + 784, + 808 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "To automate smile transferring without manual adjustments, we employ a similar approach to MaskGAN. Specifically, we select 400 pairs of neutral-", + "bbox": [ + 212, + 809, + 784, + 839 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "Wang et al.", + "bbox": [ + 271, + 114, + 349, + 128 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/a4e0340e93eec76c469c53c0702be9fc604c49e1a587a1eb87ff7d70d39ea843.jpg", + "table_caption": [ + "Table 2: Quantitative results in the smile transferring benchmark. Top 1 and top 2 results are highlighted in red and blue, respectively. Our Reader significantly outperforms previous methods in the FID and ID Score metrics by a large margin." + ], + "table_footnote": [], + "table_body": "
MetricAttribute Acc. (%)↑FID↓KID↓ID Score↑
MaskGAN [31]77.346.840.0200.4611
LocalEditing [9]26.041.260.0120.5823
LocalEditing - Encoding4Editing [49]41.848.280.0160.7295
InterFaceGAN [46]83.539.420.0100.7295
EditGAN [34]91.541.740.0130.7047
EditGAN+30 [34]85.840.830.0120.7452
Reader (ours)84.237.910.0110.8470
", + "bbox": [ + 228, + 198, + 769, + 327 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "expression and smiling-expression faces with the closest head position as the source and target images. Subsequently, we replace both the tokens corresponding to mouth and lips of the source image with those from the target image.", + "bbox": [ + 212, + 357, + 782, + 402 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Results shown in Tab. 2 demonstrate that our Reader achieves excellent performance across all metrics. Specifically, our Reader significantly outperforms previous methods in the FID and ID Score metrics by a large margin. This indicates that the edited images produced by Reader exhibit higher quality while better preserving the ID features of the original images. This observation effectively highlights the inherent advantage of our method. By representing the image through a series of decoupled tokens, Reader can selectively manipulate only the target attributes without affecting other regions. To provide a more intuitive perception, we illustrate some examples of the original and edited image pairs in Fig. 7. It can be observed that our method effectively prevents changes in the areas which are not the editing target, while achieving editing goals. In contrast, the GAN-based baseline [31] is prone to inducing global changes.", + "bbox": [ + 212, + 402, + 787, + 585 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "4.3 Ablation Study", + "text_level": 1, + "bbox": [ + 214, + 612, + 388, + 627 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "In Tab. 3a, we conducted ablation studies to evaluate the effectiveness of some optional hyper-parameters and modules in the model architecture, where different configurations represent: Patch Size: The patch size of the ViT in Reader's encoder. Bidirectional: Whether to use bidirectional transformer block in the region decoder. Masked-Attn: Whether to employ the masked cross-attention layer in the region embedder. Results indicate that a ViT with a smaller patch size paired with the additional modules produces the best image fidelity. This demonstrates that our default settings can better preserve the fine-grained details in the image, thereby enhancing the quality of the output image.", + "bbox": [ + 212, + 642, + 787, + 777 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "In Tab. 3b, we conducted ablation experiment to determine the granularity of pseudo-labels utilized in training. In this experiment, during preprocessing, we only removed the overlapping parts from the pseudo-labels and retained masks of all different sizes to train our Reader. Subsequently, we calculated the average", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "Region-Native Autoencoder", + "bbox": [ + 544, + 114, + 730, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/7b434653f866d862758eacc0d979b5f9d75a1cda8b6b1a816381e3e5af577619.jpg", + "table_caption": [ + "Table 3: Ablation studies on both model architecture, hyper parameter and data preprocessing. Our default setting is marked with gray." + ], + "table_footnote": [ + "(a) Module configurations." + ], + "table_body": "
Patch SizeBidirectionalMasked-AttnrFID ↓
8 × 8XX98.39
✓X37.50
✓✓11.42
16 × 16✓✓61.64
", + "bbox": [ + 250, + 184, + 571, + 258 + ], + "page_idx": 13 + }, + { + "type": "table", + "img_path": "images/9c474bb8f9c8c00466395645759ed12454fcf2597093883b403b5a1ccea1b49d.jpg", + "table_caption": [], + "table_footnote": [ + "(b) Annotation configurations." + ], + "table_body": "
Area ProportionL1 ↓
< 1/321.44
1/32 ~ 1/161.56
1/16 ~ 1/82.53
> 1/83.46
", + "bbox": [ + 598, + 184, + 754, + 258 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "L1 distance of different-sized regions between the original images and the reconstructed images. The \"Area Proportion\" in Tab. 3b represents the proportion of the region's area relative to the entire image. The results indicate that when the region's area proportion is less than $\\frac{1}{16}$ , the L1 distance stabilizes with changes in area. Therefore, in the formal experiments, we only retained pseudo-labels with area proportion smaller than $\\frac{1}{16}$ for training.", + "bbox": [ + 212, + 292, + 785, + 386 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Conclusion and Discussion", + "text_level": 1, + "bbox": [ + 215, + 412, + 508, + 428 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In conclusion, this paper explores an innovative region-based visual token representation and presents Reader. Reader employs arbitrarily shaped and sized regions as the units of image tokens, which aligns better with the visual prior: the non-uniformity of information density and the naturally existing semantic boundaries. Through experiments, we demonstrate the following inherent advantages of this region-based tokenization method. Efficiency: This method has a high compression ratio and can achieve superior reconstruction fidelity while employing a significantly smaller number of tokens for each image. Disentanglement: The region-based tokens guarantees their disentanglement from one another, and further empower Reader with the ability to perform image editing.", + "bbox": [ + 212, + 449, + 787, + 601 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Limitation and Feature Work As this paper focuses on exploring the feasibility and effectiveness of region-based tokenization, we adopted a minimal configuration in our experiments, manifested as employing the tiny-version transformer and training model on relatively small-sized images $(256\\times 256)$ . Although Reader can produce already satisfactory results under such configurations, the parameter size and data resolution still limit its broader application. For example, some high-resolution images may exhibit a degree of blurring after editing.", + "bbox": [ + 212, + 606, + 787, + 712 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In the future work, we will pursue scaling up both model parameters and the volume of data. Additionally, we aim to explore its application in various scenarios, such as interaction with LLMs and multimodal perception, generation.", + "bbox": [ + 212, + 712, + 787, + 758 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Broader Impact Reader possesses the capability to serve as an image editor, yielding numerous interesting outcomes, but also opens up the possibility of real photo manipulation for nefarious purposes. However, with the recent advancements in watermarking techniques and manipulated images identifying techniques [17], this risk is avoidable.", + "bbox": [ + 212, + 763, + 787, + 840 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "Wang et al.", + "bbox": [ + 271, + 114, + 349, + 128 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 217, + 143, + 401, + 162 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "This research was funded by the Fundamental Research Funds for the Central Universities (2024XKRC082) and the National NSF of China (No.U23A20314).", + "bbox": [ + 215, + 180, + 784, + 210 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 239, + 321, + 253 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Anderson, P., He, X., Buehler, C., Teney, D., Johnson, M., Gould, S., Zhang, L.: Bottom-up and top-down attention for image captioning and visual question answering. In: CVPR. pp. 6077-6086 (2018)", + "2. Binkowski, M., Sutherland, D.J., Arbel, M., Gretton, A.: Demystifying MMD gans. In: ICLR. OpenReview.net (2018)", + "3. Brooks, T., Holynski, A., Efros, A.A.: Instructpix2pix: Learning to follow image editing instructions. In: CVPR. pp. 18392-18402. IEEE (2023)", + "4. Burgess, C.P., Matthew, L., Watters, N., Kabra, R., Higgins, I., Botvinick, M.M., Lerchner, A.: Monet: Unsupervised scene decomposition and representation. CoRR abs/1901.11390 (2019)", + "5. Carion, N., Massa, F., Synnaeve, G., Usunier, N., Kirillov, A., Zagoruyko, S.: End-to-end object detection with transformers. In: Vedaldi, A., Bischof, H., Brox, T., Frahm, J. (eds.) ECCV. Lecture Notes in Computer Science, vol. 12346, pp. 213-229. Springer (2020)", + "6. Cheng, B., Misra, I., Schwing, A.G., Kirillov, A., Girdhar, R.: Masked-attention mask transformer for universal image segmentation. In: CVPR. pp. 1290–1299 (2022)", + "7. Cheng, B., Schwing, A.G., Kirillov, A.: Per-pixel classification is not all you need for semantic segmentation. In: Ranzato, M., Beygelzimer, A., Dauphin, Y.N., Liang, P., Vaughan, J.W. (eds.) NeurIPS. pp. 17864-17875 (2021)", + "8. Choi, Y., Choi, M., Kim, M., Ha, J., Kim, S., Choo, J.: Stargan: Unified generative adversarial networks for multi-domain image-to-image translation. In: CVPR. pp. 8789-8797. Computer Vision Foundation / IEEE Computer Society (2018)", + "9. Collins, E., Bala, R., Price, B., Süssstrunk, S.: Editing in style: Uncovering the local semantics of gans. In: CVPR. pp. 5770-5779. Computer Vision Foundation / IEEE (2020)", + "0. Couairon, G., Verbeek, J., Schwenk, H., Cord, M.: Diffedit: Diffusion-based semantic image editing with mask guidance. In: ICLR. OpenReview.net (2023)", + "1. Crowson, K., Biderman, S., Kornis, D., Stander, D., Hallahan, E., Castricato, L., Raff, E.: Vqgan-clip: Open domain image generation and editing with natural language guidance. In: ECCV. pp. 88–105. Springer (2022)", + "2. Deng, J., Guo, J., Xue, N., Zafeiriou, S.: Arcface: Additive angular margin loss for deep face recognition. In: CVPR. pp. 4690-4699. Computer Vision Foundation / IEEE (2019)", + "3. Devlin, J., Chang, M., Lee, K., Toutanova, K.: BERT: pre-training of deep bidirectional transformers for language understanding. In: Burstein, J., Doran, C., Solorio, T. (eds.) NAACL-HLT. pp. 4171-4186. Association for Computational Linguistics (2019)", + "4. Ding, H., Liu, C., He, S., Jiang, X., Loy, C.C.: MeViS: A large-scale benchmark for video segmentation with motion expressions. In: ICCV (2023)" + ], + "bbox": [ + 225, + 273, + 785, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "Region-Native Autoencoder", + "bbox": [ + 544, + 114, + 730, + 127 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "15. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., Uszkoreit, J., Houlsby, N.: An image is worth 16x16 words: Transformers for image recognition at scale. In: ICLR. OpenReview.net (2021)", + "16. Esser, P., Rombach, R., Ommer, B.: Taming transformers for high-resolution image synthesis. In: CVPR. pp. 12873-12883. Computer Vision Foundation / IEEE (2021)", + "17. Fernandez, P., Couairon, G., Jégou, H., Douze, M., Furon, T.: The stable signature: Rooting watermarks in latent diffusion models. CoRR abs/2303.15435 (2023)", + "18. Gafni, O., Polyak, A., Ashual, O., Sheynin, S., Parikh, D., Taigman, Y.: Make-a-scene: Scene-based text-to-image generation with human priors. In: ECCV. pp. 89-106. Springer (2022)", + "19. Goel, V., Peruzzo, E., Jiang, Y., Xu, D., Sebe, N., Darrell, T., Wang, Z., Shi, H.: Pair-diffusion: Object-level image editing with structure-and-appearance paired diffusion models. CoRR abs/2303.17546 (2023)", + "20. Gondal, M.W., Wuthrich, M., Miladinovic, D., Locatello, F., Breidt, M., Volchkov, V., Akpo, J., Bachem, O., Scholkopf, B., Bauer, S.: On the transfer of inductive bias from simulation to the real world: a new disentanglement dataset. In: Wallach, H.M., Larochelle, H., Beygelzimer, A., d'Alché-Buc, F., Fox, E.B., Garnett, R. (eds.) NeurIPS. pp. 15714-15725 (2019)", + "21. Goodfellow, I.J., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., Courville, A.C., Bengio, Y.: Generative adversarial nets. In: Ghahramani, Z., Welling, M., Cortes, C., Lawrence, N.D., Weinberger, K.Q. (eds.) NeurIPS (2014)", + "22. Härkönen, E., Hertzmann, A., Lehtinen, J., Paris, S.: Ganspace: Discovering interpretable GAN controls. In: Larochelle, H., Ranzato, M., Hadsell, R., Balcan, M., Lin, H. (eds.) NeurIPS (2020)", + "23. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: CVPR. pp. 770-778. IEEE Computer Society (2016)", + "24. Heusel, M., Ramsauer, H., Unterthiner, T., Nessler, B., Hochreiter, S.: Gans trained by a two time-scale update rule converge to a local nash equilibrium. In: Guyon, I., von Luxburg, U., Bengio, S., Wallach, H.M., Fergus, R., Vishwanathan, S.V.N., Garnett, R. (eds.) NeurIPS. pp. 6626-6637 (2017)", + "25. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. In: Larochelle, H., Ranzato, M., Hadsell, R., Balcan, M., Lin, H. (eds.) NeurIPS (2020)", + "26. Horé, A., Ziou, D.: Image quality metrics: PSNR vs. SSIM. In: ICPR. pp. 2366-2369. IEEE Computer Society (2010)", + "27. Huang, M., Mao, Z., Chen, Z., Zhang, Y.: Towards accurate image coding: Improved autoregressive image generation with dynamic vector quantization. In: CVPR. pp. 22596-22605 (2023)", + "28. Isola, P., Zhu, J., Zhou, T., Efros, A.A.: Image-to-image translation with conditional adversarial networks. In: CVPR. pp. 5967-5976. IEEE Computer Society (2017)", + "29. Kim, H., Mnih, A.: Disentangling by factorising. In: Dy, J.G., Krause, A. (eds.) ICML. Proceedings of Machine Learning Research, vol. 80, pp. 2654-2663. PMLR (2018)", + "30. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W., Dollár, P., Girshick, R.B.: Segment anything. CoRR abs/2304.02643 (2023)", + "31. Lee, C., Liu, Z., Wu, L., Luo, P.: Maskgan: Towards diverse and interactive facial image manipulation. In: CVPR. pp. 5548-5557. Computer Vision Foundation / IEEE (2020)" + ], + "bbox": [ + 215, + 146, + 785, + 840 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "Wang et al.", + "bbox": [ + 271, + 114, + 349, + 128 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "32. Lee, D., Kim, C., Kim, S., Cho, M., Han, W.: Autoregressive image generation using residual quantization. In: CVPR. pp. 11513-11522. IEEE (2022). https://doi.org/10.1109/CVPR52688.2022.01123, https://doi.org/10.1109/CVPR52688.2022.01123", + "33. Li, Y., Mao, H., Girshick, R., He, K.: Exploring plain vision transformer backbones for object detection. In: ECCV. pp. 280-296. Springer (2022)", + "34. Ling, H., Kreis, K., Li, D., Kim, S.W., Torralba, A., Fidler, S.: Editgan: High-precision semantic image editing. In: Ranzato, M., Beygelzimer, A., Dauphin, Y.N., Liang, P., Vaughan, J.W. (eds.) NeurIPS. pp. 16331-16345 (2021)", + "35. Liu, C., Ding, H., Jiang, X.: GRES: Generalized referring expression segmentation. In: CVPR (2023)", + "36. Liu, C., Jiang, X., Ding, H.: Primitivenet: decomposing the global constraints for referring segmentation. Visual Intelligence 2(1), 16 (2024)", + "37. Locatello, F., Weissenborn, D., Unterthiner, T., Mahendran, A., Heigold, G., Uszkoreit, J., Dosovitskiy, A., Kipf, T.: Object-centric learning with slot attention. In: Larochelle, H., Ranzato, M., Hadsell, R., Balcan, M., Lin, H. (eds.) NeurIPS (2020)", + "38. Ning, J., Li, C., Zhang, Z., Wang, C., Geng, Z., Dai, Q., He, K., Hu, H.: All in tokens: Unifying output space of visual tasks via soft token. In: ICCV. pp. 19900-19910 (2023)", + "39. van den Oord, A., Vinyals, O., Kavukcuoglu, K.: Neural discrete representation learning. In: Guyon, I., von Luxburg, U., Bengio, S., Wallach, H.M., Fergus, R., Vishwanathan, S.V.N., Garnett, R. (eds.) NeurIPS. pp. 6306-6315 (2017)", + "40. Pan, X., Tewari, A., Leimkuhler, T., Liu, L., Meka, A., Theobalt, C.: Drag your GAN: interactive point-based manipulation on the generative image manifold. In: Brunvand, E., Sheffer, A., Wimmer, M. (eds.) SIGGRAPH. pp. 78:1-78:11. ACM (2023)", + "41. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: ICML. pp. 8748-8763. PMLR (2021)", + "42. Radford, A., Narasimhan, K., Salimans, T., Sutskever, I., et al.: Improving language understanding by generative pre-training (2018)", + "43. Reed, S.E., Zhang, Y., Zhang, Y., Lee, H.: Deep visual analogy-making. In: Cortes, C., Lawrence, N.D., Lee, D.D., Sugiyama, M., Garnett, R. (eds.) NeurIPS. pp. 1252-1260 (2015)", + "44. Ren, S., He, K., Girshick, R., Sun, J.: Faster r-cnn: Towards real-time object detection with region proposal networks. NeurIPS 28 (2015)", + "45. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: CVPR. pp. 10674-10685. IEEE (2022)", + "46. Shen, Y., Yang, C., Tang, X., Zhou, B.: Interfacegan: Interpreting the disentangled face representation learned by gans. IEEE TPAMI 44(4), 2004-2018 (2022)", + "47. Shuai, X., Ding, H., Ma, X., Tu, R., Jiang, Y.G., Tao, D.: A survey of multimodal-guided image editing with text-to-image diffusion models. arXiv preprint arXiv:2406.14555 (2024)", + "48. Sohl-Dickstein, J., Weiss, E.A., Maheswaranathan, N., Ganguli, S.: Deep unsupervised learning using nonequilibrium thermodynamics. In: Bach, F.R., Blei, D.M. (eds.) ICML. JMLR Workshop and Conference Proceedings, vol. 37, pp. 2256-2265. JMLR.org (2015)", + "49. Tov, O., Alaluf, Y., Nitzan, Y., Patashnik, O., Cohen-Or, D.: Designing an encoder for stylegan image manipulation. ACM TOG 40(4), 133:1-133:14 (2021)" + ], + "bbox": [ + 212, + 146, + 787, + 840 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "Region-Native Autoencoder", + "bbox": [ + 544, + 114, + 730, + 127 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "50. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. In: Guyon, I., von Luxburg, U., Bengio, S., Wallach, H.M., Fergus, R., Vishwanathan, S.V.N., Garnett, R. (eds.) NeurIPS. pp. 5998-6008 (2017)", + "51. Wu, B., Xu, C., Dai, X., Wan, A., Zhang, P., Yan, Z., Tomizuka, M., Gonzalez, J., Keutzer, K., Vajda, P.: Visual transformers: Token-based image representation and processing for computer vision. arXiv preprint arXiv:2006.03677 (2020)", + "52. Wu, R., Zhang, G., Lu, S., Chen, T.: Cascade EF-GAN: progressive facial expression editing with local focuses. In: CVPR. pp. 5020-5029. Computer Vision Foundation / IEEE (2020)", + "53. Xu, J., Mello, S.D., Liu, S., Byeon, W., Breuel, T.M., Kautz, J., Wang, X.: Groupvit: Semantic segmentation emerges from text supervision. In: CVPR. pp. 18113-18123. IEEE (2022)", + "54. Yang, B., Gu, S., Zhang, B., Zhang, T., Chen, X., Sun, X., Chen, D., Wen, F.: Paint by example: Exemplar-based image editing with diffusion models. In: CVPR. pp. 18381-18391. IEEE (2023). https://doi.org/10.1109/CVPR52729.2023.01763, https://doi.org/10.1109/CVPR52729.2023.01763", + "55. Yang, T., Wang, Y., Lu, Y., Zheng, N.: Visual concepts tokenization. In: NeurIPS (2022)", + "56. Yu, J., Li, X., Koh, J.Y., Zhang, H., Pang, R., Qin, J., Ku, A., Xu, Y., Baldridge, J., Wu, Y.: Vector-quantized image modeling with improved VQGAN. In: ICLR. OpenReview.net (2022)", + "57. Zellers, R., Yatskar, M., Thomson, S., Choi, Y.: Neural motifs: Scene graph parsing with global context. In: CVPR. pp. 5831-5840 (2018)", + "58. Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual metric. In: CVPR. pp. 586-595. Computer Vision Foundation / IEEE Computer Society (2018)", + "59. Zhong, Y., Wang, L., Chen, J., Yu, D., Li, Y.: Comprehensive image captioning via scene graph decomposition. In: ECCV. pp. 211-229. Springer (2020)", + "60. Zhu, J., Yang, C., Shen, Y., Shi, Z., Zhao, D., Chen, Q.: Linkgan: Linking GAN latents to pixels for controllable image synthesis. CoRR abs/2301.04604 (2023)" + ], + "bbox": [ + 215, + 146, + 785, + 575 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "Wang et al.", + "bbox": [ + 271, + 114, + 349, + 128 + ], + "page_idx": 17 + } +] \ No newline at end of file diff --git a/2024/Region-Native Visual Tokenization/a10ef64d-036f-4197-bc89-c920baeb7773_model.json b/2024/Region-Native Visual Tokenization/a10ef64d-036f-4197-bc89-c920baeb7773_model.json new file mode 100644 index 0000000000000000000000000000000000000000..5d8225421a318557ea424b8a0ab7b4caac500124 --- /dev/null +++ b/2024/Region-Native Visual Tokenization/a10ef64d-036f-4197-bc89-c920baeb7773_model.json @@ -0,0 +1,3063 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.302, + 0.142, + 0.703, + 0.163 + ], + "angle": 0, + "content": "Region-Native Visual Tokenization" + }, + { + "type": "text", + "bbox": [ + 0.278, + 0.189, + 0.726, + 0.237 + ], + "angle": 0, + "content": "Mengyu Wang\\(^{1,5(\\dagger)}\\) \\(\\odot\\), Yuyao Huang\\(^{1,5(\\dagger,\\ddagger)}\\) \\(\\odot\\), Henghui Ding\\(^{4}\\) \\(\\odot\\), Xinlong Wang\\(^{2}\\) \\(\\odot\\), Tiejun Huang\\(^{2}\\), Yao Zhao\\(^{1,5,6}\\) \\(\\odot\\), Yunchao Wei\\(^{1,5,6(\\boxtimes)}\\) \\(\\odot\\), and Shuicheng Yan\\(^{3}\\) \\(\\odot\\)" + }, + { + "type": "text", + "bbox": [ + 0.241, + 0.246, + 0.761, + 0.29 + ], + "angle": 0, + "content": "\\(^{1}\\) Institute of Information Science, Beijing Jiaotong University, Beijing, China \n\\(^{2}\\) Beijing Academy of Artificial Intelligence \n\\(^{3}\\) Skywork AI" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.29, + 0.786, + 0.332 + ], + "angle": 0, + "content": "\\(^{4}\\) Institute of Big Data, Fudan University, Shanghai, China \n\\(^{5}\\) Visual Intelligence + X International Joint Laboratory of the Ministry of Education \n\\(^{6}\\) PengCheng Laboratory, Shenzhen, China" + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.36, + 0.744, + 0.665 + ], + "angle": 0, + "content": "Abstract. We explore an innovative region-based visual token representation and present the REgion-native AutoencoDER (Reader). In contrast to the majority of previous methods, which represent each image as a grid-shaped tokens map, Reader perceives each image into sequential region-based tokens, with each token corresponding to an object or one part of an object in the image. Specifically, Reader comprises both an encoder and a decoder. The encoder can partition each image into an adaptive number of arbitrary-shaped regions and encode each region into a token. Subsequently, the decoder utilizes this adaptive-length token sequence to reconstruct the original image. Experimental results demonstrate that such region-based token representation possesses two main notable characteristics. Firstly, it achieves highly efficient image encoding. Reader can adaptively use more regions to represent complex areas and fewer regions in simpler ones, thus avoiding information redundancy. Consequently, it achieves superior reconstruction fidelity compared to previous methods, despite using significantly fewer tokens for each image. Secondly, the region-based manner enables manipulation on a local region without causing global changes. As a result, Reader inherently supports diverse image editing operations, including erasing, adding, replacing, and modifying shapes on the objects, and achieves excellent performance in the image editing benchmark of smile transferring. Code is provided at https://github.com/MengyuWang826/Reader.git." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.687, + 0.377, + 0.703 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.715, + 0.788, + 0.793 + ], + "angle": 0, + "content": "Tokenization, originating from Natural Language Processing (NLP), involves breaking a textual data stream into several elements called tokens. It plays an important role for converting raw data into a form that is predictable and analyzable by deep learning systems [13, 42]. With the gradual maturity of tokenization technique in the NLP field, it has also been adapted to the Computer" + }, + { + "type": "page_footnote", + "bbox": [ + 0.232, + 0.798, + 0.388, + 0.812 + ], + "angle": 0, + "content": "† Internship at BAAI." + }, + { + "type": "page_footnote", + "bbox": [ + 0.232, + 0.813, + 0.425, + 0.826 + ], + "angle": 0, + "content": "\\(\\ddagger\\) Internship at Skywork AI." + }, + { + "type": "page_footnote", + "bbox": [ + 0.232, + 0.827, + 0.408, + 0.84 + ], + "angle": 0, + "content": "Corresponding author." + }, + { + "type": "list", + "bbox": [ + 0.232, + 0.798, + 0.425, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.351, + 0.129 + ], + "angle": 0, + "content": "Wang et al." + }, + { + "type": "image", + "bbox": [ + 0.229, + 0.145, + 0.357, + 0.266 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.264, + 0.268, + 0.323, + 0.28 + ], + "angle": 0, + "content": "(a) Image" + }, + { + "type": "image", + "bbox": [ + 0.368, + 0.144, + 0.496, + 0.266 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.386, + 0.268, + 0.474, + 0.279 + ], + "angle": 0, + "content": "(b) Grid-based" + }, + { + "type": "image", + "bbox": [ + 0.507, + 0.144, + 0.635, + 0.266 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.523, + 0.268, + 0.621, + 0.28 + ], + "angle": 0, + "content": "(c) Object-based" + }, + { + "type": "image", + "bbox": [ + 0.647, + 0.145, + 0.774, + 0.266 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.661, + 0.268, + 0.762, + 0.28 + ], + "angle": 0, + "content": "(d) Region-based" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.292, + 0.785, + 0.362 + ], + "angle": 0, + "content": "Fig.1: Illustration of different types of visual tokenization methods. Grid-based methods [16,56] represent images as grid-shaped tokens map. Object-based methods [1,59] use detection models to represent objects as tokens. While our region-based method represents arbitrarily shaped regions as tokens, which offers more flexibility and retains the naturally existed boundaries presented in images." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.38, + 0.785, + 0.426 + ], + "angle": 0, + "content": "Vision (CV) domain, known as visual tokenization, and has become the foundation for many tasks such as image/video generation [16, 27, 56] and multimodal generation/perception [11, 14, 18, 38]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.429, + 0.788, + 0.747 + ], + "angle": 0, + "content": "Existing visual tokenization methods can be categorized into two main types, which we refer to as grid-based and object-based. Grid-based methods typically employ a CNN [23] or ViT [15] backbone to encode the original image into a downsampled grid-shaped feature map, followed by quantization of the feature map to get the image tokens. However, in such methods, each token corresponds to a patch in the original image (see Fig. 1b), which does not adhere to the following characteristics of the image prior. Firstly, different areas in images have varying information densities. While in these methods, patches are treated equally, leading to information loss in densely informative patches and redundancy in sparsely informative ones. Secondly, areas with different semantics within an image have natural boundaries. While employing patches as the units of image tokens disregards the natural boundaries, causing local information within the original image to be distributed and coupled across multiple tokens, limiting the extraction and utilization of local features. Object-based methods [1,35,51,53,59], as an alternative and more natural approach, typically appear in vision-language pretraining. These methods employ detection or segmentation models (see Fig. 1c) to encode the objects within images into tokens, which are then interacted with or aligned with language information. However, these methods only focus on extracting high-level semantic information, neglecting low-level visual details in the image. Consequently, they cannot be used for tasks that require image outputs, such as image synthesis and editing." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.841 + ], + "angle": 0, + "content": "These challenges suggest that, an ideal visual token representation should take into account the crucial visual-prior characteristics, including the uneven information density and the inherent semantic boundaries, and further achieve local-information decoupling and global-information compression. However, previous practices either ignore the visual-prior characteristics or, due to their coarse granularity, result in significant information loss. To address these problems, we" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.545, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "Region-Native Autoencoder" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.145, + 0.314, + 0.2 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.314, + 0.145, + 0.407, + 0.2 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.408, + 0.145, + 0.503, + 0.2 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.145, + 0.596, + 0.2 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.597, + 0.145, + 0.689, + 0.2 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.691, + 0.145, + 0.785, + 0.2 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.269, + 0.2, + 0.354, + 0.207 + ], + "angle": 0, + "content": "Erasing the two people" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.209, + 0.314, + 0.281 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.233, + 0.281, + 0.391, + 0.289 + ], + "angle": 0, + "content": "Slimming the jawline and enlarging the eyes" + }, + { + "type": "image", + "bbox": [ + 0.315, + 0.209, + 0.408, + 0.281 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.41, + 0.209, + 0.503, + 0.281 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.209, + 0.595, + 0.281 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.597, + 0.209, + 0.69, + 0.281 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.692, + 0.2, + 0.732, + 0.207 + ], + "angle": 0, + "content": "mushroom" + }, + { + "type": "image_caption", + "bbox": [ + 0.413, + 0.281, + 0.585, + 0.289 + ], + "angle": 0, + "content": "Non-smiling expressions to smiling expressions" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.295, + 0.787, + 0.364 + ], + "angle": 0, + "content": "Fig. 2: Examples of the proposed Reader performing various editing operations. Reader showcases its strengths in flexibly editing multiple attributes of existing real images under several interactive operations (see Sec. 3.4). In each pair, the left displays the original image, the right displays the edited result and the text describes the corresponding editing objective." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.385, + 0.785, + 0.476 + ], + "angle": 0, + "content": "propose the REGion-native AutoencoDER (Reader), which takes arbitrarily shaped and sized regions (see Fig. 1d) as the fundamental units for visual tokens, providing more flexibility and aligning better with the visual-prior characteristics. The Reader framework conceptualizes each image as a composite of several regions, with each region corresponding to either a small-sized object or a portion of a regular or large-sized object." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.477, + 0.785, + 0.596 + ], + "angle": 0, + "content": "Specifically, the Reader architecture comprises an encoder and a decoder, as illustrated in Fig. 3. The encoder first partitions an image into an adaptive-number of non-overlapping regions, where each region is represented as a binary mask and an embedding. Following this, the encoder employs a novel region embedder to integrate the geometric information from the mask with the visual features from the embedding, and produce the desired token for each region. Subsequently, a novel cross-attention-based decoder is employed to reconstruct the original image from these tokens." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.598, + 0.789, + 0.627 + ], + "angle": 0, + "content": "We further conduct both qualitative and quantitative experiments and demonstrate the following two main notable characteristics of Reader:" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.629, + 0.785, + 0.718 + ], + "angle": 0, + "content": "Efficiency: Reader achieves highly efficient encoding, yielding reconstruction fidelity superior to that of previous methods, despite employing a significantly smaller number of tokens. This is intuitive since the region-based manner enables Reader to naturally represent complex areas with more regions, while representing simpler areas with fewer regions. This effectively avoids information redundancy, thus possessing inherent efficiency." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.786, + 0.84 + ], + "angle": 0, + "content": "Disentanglement: Representing non-overlapping regions as tokens guarantees their disentanglement from one another, and further empower Reader with the ability to perform image editing. Taking object erasure as an example, the disentanglement permits an object in the image to be represented solely by the tokens corresponding to its interior. Through removing these tokens and then reconstructing the image, object erasing can be naturally achieved. Consequently, without any additional modules or training data, Reader facilitates diverse object-level editing operations (see Fig. 2), such as erasing, adding, re" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.232, + 0.128 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.351, + 0.129 + ], + "angle": 0, + "content": "Wang et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.178 + ], + "angle": 0, + "content": "placing, and modifying shapes on objects, and demonstrates excellent performance in the image editing benchmark of smile transferring [31]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.179, + 0.788, + 0.255 + ], + "angle": 0, + "content": "We believe that these experiments effectively demonstrate the inherent advantages of region-based tokenization method. We hope this will bring some novel insights to the community about visual tokenization and unlock its value in various downstream tasks in the future, such as interaction with LLMs and multimodal perception, generation, and beyond." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.285, + 0.388, + 0.302 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.329, + 0.788, + 0.542 + ], + "angle": 0, + "content": "Grid-based Tokenization The majority of grid-based methods follows the feature extraction, quantization, and reconstruction paradigm, initially introduced by VQVAE [39]. These methods represent image tokens as a grid-shaped code map. Expanding upon this framework, VQGAN [16] incorporates perceptual loss [58] and GAN loss [28] to augment the original L1 reconstruction loss, with the aim of improving both reconstruction fidelity and codebook learning. RQ-VAE [32] introduces residual quantization to further enhance reconstruction fidelity, while leading to the cost of significantly higher number of tokens. ViT-VQGAN [56] replaces the previously employed CNN-based architecture with a ViT-based one. Sharing a similar motivation to ours, DQ-VAE [27] proposes dynamically selecting patch granularity between two or three pre-defined strides to achieve dynamic-length coding. However, this approach remains constrained by the grid-based paradigm. In contrast, our Reader takes arbitrarily shaped regions as image token units and offers greater flexibility." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.547, + 0.788, + 0.744 + ], + "angle": 0, + "content": "Object-based Tokenization Methods of this kind typically find applications in vision-language pretraining and multimodal perception. For instance, [1] uses a Faster R-CNN [44] for proposals prediction and encodes the proposals into tokens to subsequently perform image captioning or visual question answering. Similarly, [59] also employs a Faster R-CNN to extract object tokens and maps these tokens with a Graph Neural Network [57]. GroupViT [53] proposes a learnable clustering architecture based on ViT [15] to extract object information, and aligns it with text embedding akin to CLIP [41]. However, these methods focus on extracting semantic information at object-level, leading to excessively coarse-grained and overlooking on detailed visual information. In contrast, our Reader employs an autoencoder architecture to extract dynamic-grained local regions as the unit of image tokens, which better preserves detailed information and showcases superior performance in image reconstruction and editing." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Scene Decomposition The scene decomposition task [4,37,55], also referred to as object-centric representation, has the most similar form to our Reader. This task aims to decompose a scene into multiple objects and learn the representation of each object. However, these works focus on application in some synthetic datasets containing only basic 3D shapes [20,29,43]. In contrast, our Reader can be applied to a wide range of natural images." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.545, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Region-Native Autoencoder" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.283 + ], + "angle": 0, + "content": "Image Editing Most early image editing methods are GAN-based [21]. Some of these approaches [22, 34, 40, 60] rely on the controllability of the GAN's latent space, employing supervised or unsupervised approaches to achieve desired transformations of GAN's latent variable. Other works [8, 31, 52] utilize GANs' conditional generation ability and synthesize the edited images based on the input conditions. The remarkable success of diffusion models [25, 48] in image synthesis has also led to the emergence of numerous diffusion-based editing methods [3, 10, 19, 47, 54]. These approaches largely build upon the robust semantic understanding and rendering capabilities of large pre-trained diffusion models [45]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.284, + 0.785, + 0.343 + ], + "angle": 0, + "content": "The semantic-map based methods [19, 31, 34] can produce the most similar effects to our Reader. However, compared to these works, our Reader benefits from its region-based representation, enabling a wider variety of editing operations without requiring additional input, with details elaborated in Sec. 3.4." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.368, + 0.381, + 0.387 + ], + "angle": 0, + "content": "3 Methodology" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.401, + 0.785, + 0.506 + ], + "angle": 0, + "content": "We propose Reader (region-native autoencoder) in this work, which performs region-based visual tokenization and enables diverse controllable image editing operations. In the following, we will first provide detailed descriptions about the model architecture of the Reader's encoder and decoder in Sec. 3.1 and Sec. 3.2, respectively. Sec. 3.3 outlines the objective function utilized, while Sec. 3.4 delves into the detailed implementation for image editing. The model architecture is illustrated in Fig. 3." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.53, + 0.442, + 0.545 + ], + "angle": 0, + "content": "3.1 Encoder Architecture" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.556, + 0.784, + 0.616 + ], + "angle": 0, + "content": "The encoder of Reader needs the capability to first segment an image into numerous regions and then encode each region into a token. Thus, we propose a encoder with three core components: an image feature extractor, a region decoder and a region embedder." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.623, + 0.784, + 0.744 + ], + "angle": 0, + "content": "Feature Extractor To extract information from the original image for subsequent image segmentation, tokenization, and reconstruction steps, we employ a standard ViT [15] along with a simple FPN [33] as our feature extractor. Given the importance of fine-grained details for high-fidelity reconstruction, we configure the patch size of the ViT to be \\(8 \\times 8\\), a smaller size compared to conventional settings, yet effective in preserving local structure. The strides of the simple FPN are set to \\((1, \\frac{1}{2})\\), thus the output features are at resolutions of \\(\\frac{1}{8}\\) and \\(\\frac{1}{4}\\) of the original image." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.784, + 0.81 + ], + "angle": 0, + "content": "Region Decoder We employ a region decoder to learn the region representation and segment image into regions. Specifically, we employ a standard Transformer decoder [50], which is similar to previous Transformer-based segmentation works [5-7,36], augmented with an additional bidirectional Transformer block [30]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.784, + 0.84 + ], + "angle": 0, + "content": "In the Transformer decoder, \\(N\\) learnable embeddings are used as input queries to extract the region information, termed region queries. Within each" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.351, + 0.129 + ], + "angle": 0, + "content": "Wang et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.282 + ], + "angle": 0, + "content": "layer of the Transformer decoder, one of the two-sized image features alternately serves as the key and value. Through cross-attention, the region queries can extract region-level information from the image features. Subsequently, the bidirectional Transformer block takes the updated region queries and the larger-sized image feature as input to perform both region-to-image and image-to-region attention and update them. This process yields the region embeddings \\(\\mathcal{E}_r \\in \\mathbb{R}^{N \\times c}\\) and a region-aware image feature \\(\\mathcal{E}_{img} \\in \\mathbb{R}^{c \\times h \\times w}\\). Here, \\((h, w)\\) represents the resolution of the image feature, which is \\(\\frac{1}{4}\\) of the original image size, and \\(c\\) denotes the channel dimension." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.283, + 0.787, + 0.314 + ], + "angle": 0, + "content": "We then compute the similarity matrix between the region embeddings and the image feature via a dot product, formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.408, + 0.321, + 0.786, + 0.34 + ], + "angle": 0, + "content": "\\[\nS = \\mathcal {E} _ {r} \\odot \\mathcal {E} _ {i m g} \\in \\mathcal {R} ^ {N \\times h \\times w}. \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.347, + 0.788, + 0.422 + ], + "angle": 0, + "content": "We use this similarity matrix to get the desired regions. To prevent information redundancy caused by overlapping, we apply an argmax operation along the first dimension of the similarity matrix. This process means matching each pixel with one embedding that has the highest similarity to the pixel's feature, which can be formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.345, + 0.43, + 0.786, + 0.448 + ], + "angle": 0, + "content": "\\[\nS _ {\\text {h a r d}} = \\operatorname {o n e - h o t} (\\operatorname {a r g m a x} (S)) \\in \\{0, 1 \\} ^ {N \\times h \\times w}. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.456, + 0.788, + 0.577 + ], + "angle": 0, + "content": "Consequently, the \\( S_{\\mathrm{hard}} \\) in Eq. (2) can also be seen as \\( N \\) binary masks, where the foreground of each mask indicates which pixels belong to that specific region. However, not each one among the \\( N \\) masks contains foreground. In fact, in some simple images, numerous pixels may be matched with a single region embedding, while other region embeddings remain unmatched with any pixel (illustrated as \"used\" and \"unused\" in Fig. 3). Therefore, while \\( N \\) is a fixed hyper-parameter, the actual number of regions corresponding to each image is dynamic, determined by the complexity of the image." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.583, + 0.788, + 0.658 + ], + "angle": 0, + "content": "Region Embedder Through the region decoder, the original image is represented as several regions. Each region comprises a region embedding \\(\\mathcal{E}_r[i]\\) representing the visual feature and a binary mask \\(S_{\\mathrm{hard}}[i]\\) representing the geometric information (e.g., shape, position and scale). The region embedder is further employed to merge the above two components and produce the desired tokens." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.658, + 0.788, + 0.72 + ], + "angle": 0, + "content": "Following GroupViT [53], we first aggregate the feature of each pixel into the corresponding region embedding. To make the entire process differentiable, enabling the gradients of image reconstruction to guide the granularity of regions, we employ the straight-through trick [39] here. This process is formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.383, + 0.728, + 0.785, + 0.744 + ], + "angle": 0, + "content": "\\[\nS = \\operatorname {s o f t m a x} (S, \\dim = 0), \\tag {3}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.383, + 0.749, + 0.785, + 0.766 + ], + "angle": 0, + "content": "\\[\n\\tilde {S} = S _ {\\text {h a r d}} + S - \\operatorname {s g} \\cdot (S), \\tag {4}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.363, + 0.77, + 0.785, + 0.802 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathcal {E}} _ {r} [ i ] = \\mathcal {E} _ {r} [ i ] + \\sum_ {j, k} \\hat {S} [ i, j, k ] \\cdot \\mathcal {E} _ {i m g} [:, j, k ], \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.81, + 0.787, + 0.842 + ], + "angle": 0, + "content": "in which the \\(\\mathrm{sg}\\cdot ()\\) refers to stop gradient. While the pixel features aggregation process shown in Eq. (5) is actually a pooling process, leading to over-smoothing" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.545, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "Region-Native Autoencoder" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "image", + "bbox": [ + 0.217, + 0.146, + 0.787, + 0.304 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.307, + 0.785, + 0.377 + ], + "angle": 0, + "content": "Fig. 3: An overview of the proposed Reader. It comprises an encoder and a decoder. The encoder can segment each image into an adaptive number of regions, which is illustrated as the \"used\" part, and embeds each region into a token with the Region Embedder. The decoder then reconstructs the original image from these tokens with the learnable patch queries." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.396, + 0.784, + 0.456 + ], + "angle": 0, + "content": "on local details. We additionally employ a masked cross-attention layer [6] to enhance the visual feature extraction. This layer uses the region embeddings \\(\\hat{\\mathcal{E}}_r\\) as queries to perform attention to the image feature \\(\\mathcal{E}_{img}\\) only in the foreground of each region in order to extract more local details." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.457, + 0.788, + 0.609 + ], + "angle": 0, + "content": "Subsequently, we incorporate a module to explicitly encode the geometric information in the masks and integrate it into the region embeddings. This approach serves two primary purposes: Firstly, the inclusion of geometric information enhances the delineation of boundaries between regions, thereby facilitating image reconstruction. Secondly, such explicit encoding also enables more editing operations in subsequent steps. Specifically, we flatten each mask and utilize a linear layer to project it into a latent space. Additionally, the region embeddings are also projected with an MLP and are then fused with the former projection through element-wise addition. Following a normalization and another MLP, the region embedder produces the desired region-based image tokens." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.632, + 0.442, + 0.646 + ], + "angle": 0, + "content": "3.2 Decoder Architecture" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.659, + 0.785, + 0.749 + ], + "angle": 0, + "content": "The encoder of Reader represents each image as an adaptive-length sequence. Consequently, the decoder needs the capacity to decode this adaptive-length sequence into a uniformly-sized image. This poses a challenge of mapping variable-length input to a fixed-length output. To address this, the decoder architecture of Reader is based on the Transformer decoder, leveraging the cross-attention mechanism to resolve the aforementioned problem." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Specifically, we introduce \\( M \\) learnable embeddings as the input queries for the Transformer decoder. Each embedding is responsible for reconstructing a local patch in the output image and is referred to as a patch query. The queries are also incorporated with the positional embeddings of the center coordinate of the patches. To prevent information loss and enable the Transformer decoder to focus on different information at each layer, we employ a separate MLP for each" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.351, + 0.129 + ], + "angle": 0, + "content": "Wang et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.177 + ], + "angle": 0, + "content": "layer to project the image tokens. These projections then serve as the key and value for the corresponding layer." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.178, + 0.785, + 0.253 + ], + "angle": 0, + "content": "Through the Transformer decoder, the patch queries interact with the image tokens through cross-attention, gathering information related to the corresponding patches and updating their values. After the Transformer decoder, we integrate an linear layer to finally map the output patch queries to the pixel space, which is a reverse process of the patch embedder of ViT." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.27, + 0.421, + 0.286 + ], + "angle": 0, + "content": "3.3 Objective Function" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.291, + 0.784, + 0.337 + ], + "angle": 0, + "content": "The objective functions related to the image reconstruction are kept consistent with VQGAN [16], which is a combination of the L1 loss, perceptual loss [58] and a patch-based GAN loss [28], denoted as \\(\\mathcal{L}_{img}\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.337, + 0.785, + 0.44 + ], + "angle": 0, + "content": "However, in our observation, relying solely on the image reconstruction loss is insufficient for training the encoder to produce semantically meaningful regions. In practice, the regions easily collapse into uniformly distributed patches. Therefore, additional supervision on the output regions becomes necessary. We employ either the pseudo-annotations produced by SAM [30] or manual annotations to supervise the output regions in different experiments, with detailed information provided in Sec. 4." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.442, + 0.785, + 0.547 + ], + "angle": 0, + "content": "In order to train the similarity matrix in Eq. (3) to resemble a one-hot form at each pixel, we employ a cross-entropy loss here. Unlike the tasks, where cross-entropy typically used with defined category IDs, there is no fixed correspondence between the predicted regions and the masks of annotations here. Following [5,7], we employ a Hungarian-algorithm based bipartite matching to determine the matching relationship, with more details provided in supplementary material. The overall objective function is formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.444, + 0.555, + 0.785, + 0.571 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathcal {L} _ {i m g} + \\mathcal {L} _ {c e}. \\tag {6}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.587, + 0.517, + 0.603 + ], + "angle": 0, + "content": "3.4 Image Editing Implementation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.608, + 0.784, + 0.667 + ], + "angle": 0, + "content": "The proposed Reader inherently facilitates diverse image editing operations. Without any additional modules, this is achieved by simply manipulating the image tokens outputted by the encoder and then inputting these manipulated tokens into the decoder to get the edited image." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.669, + 0.785, + 0.759 + ], + "angle": 0, + "content": "Specifically, before the region embedder of Reader, each region is represented by two components: the geometric information (represented as a binary mask) and the visual information (represented as the region embedding). This enables independent manipulations on both two parts and achieves different editing effects. Therefore, we define the following three-level image editing operations. A schematic diagram is shown in Fig. 4 to provide an intuitive understanding." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.765, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Geometry Level These operations refer to only manipulating the masks through affine transformations or interactive brushing by users, while maintaining the embeddings unaltered. Geometry-level operations allows to perform moving, scaling, and shape modification on objects. As an example, the left branch in Fig. 4 alters the shape of the lips without changing their color." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.545, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "Region-Native Autoencoder" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "image", + "bbox": [ + 0.216, + 0.144, + 0.788, + 0.284 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.289, + 0.788, + 0.332 + ], + "angle": 0, + "content": "Fig. 4: Illustration of editing operations. Different operations on the embeddings and masks leads to different editing effects. The left, middle and right branches correspond to the three-levels operations described in Sec. 3.4, respectively." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.351, + 0.784, + 0.427 + ], + "angle": 0, + "content": "Feature Level Feature-level operations maintain the mask unchanged while replacing the corresponding embedding with one from another object or another image. This facilitates local feature transferring without impacting geometric characteristics. An example is illustrated in the middle branch of Fig. 4, where the color of lips is changed by replacing the corresponding two embeddings." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.433, + 0.788, + 0.569 + ], + "angle": 0, + "content": "Token Level The erasure, addition and replacement of objects are categorized as this level. Erasure can be achieved by removing tokens corresponding to a specific object. Conversely, addition can be accomplished by appending tokens from another image to the tokens of the current image. Object replacement can then be achieved by combining both operations. Furthermore, these operations can also be combined with the geometry-level manipulations to achieve addition and replacement on objects while controlling their shapes, scale and position, etc. An example is illustrated in the right branch of Fig. 4, where the mouth and lips are replaced by those from another image." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.593, + 0.376, + 0.611 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.626, + 0.473, + 0.642 + ], + "angle": 0, + "content": "4.1 Experiment on ImageNet" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.653, + 0.784, + 0.684 + ], + "angle": 0, + "content": "Herein we train Reader on ImageNet training set and report qualitative results in image editing and quantitative comparisons on reconstruction fidelity." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Implementation As outlined in Sec. 3.3, additional supervision is essential to prevent the collapse of producing regions. Consequently, in this experiment, we employ SAM [30] to produce pseudo-annotations on the ImageNet training set. Specifically, we keep the original settings of the segment anything mode of SAM, which effectively segments all prominent objects or local parts in images. Since there is substantial overlapping in the original results of SAM, while the cross-entropy loss needs non-overlapping annotations, we implement the following two post-processing steps. Firstly, removing all larger masks that encompass other ones, retaining only smaller, more localized masks. Secondly, regions not encompassed by the pseudo annotations do not contribute to the objective function." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.351, + 0.129 + ], + "angle": 0, + "content": "Wang et al." + }, + { + "type": "image", + "bbox": [ + 0.217, + 0.143, + 0.788, + 0.369 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.372, + 0.788, + 0.401 + ], + "angle": 0, + "content": "Fig. 5: Qualitative results in ImageNet validation set and Internet images. The editing objective is labeled with text under image." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.434, + 0.788, + 0.587 + ], + "angle": 0, + "content": "The input and output sizes of images are set to \\((256 \\times 256)\\) in this experiment. Both the ViT in the encoder and the Transformer in the decoder follow the configuration of ViT-tiny, which utilizes 192 channels and 3 heads in the multi-head attention. The dimension of image tokens is set to 256 to maintain consistency with the previous methods. The number \\(N\\) of learnable region queries in the encoder is set to 256, which is the minimal tokens number of each image in previous works and serves as the maximum regions number of our model. The number \\(M\\) of learnable patch queries in the decoder is set to 1024, with each query responsible for reconstructing a \\(8 \\times 8\\) patch in the output image. For more implementation details, please refer to our supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.591, + 0.788, + 0.683 + ], + "angle": 0, + "content": "Results We provide several image editing examples in Fig. 5 corresponding to all the three-level editing operations described in Sec. 3.4. For example, \"sticking out the tongue\" needs only manipulation on the tongue's mask, \"replacing the color\" is achieved by replacing the region embedding of the strawberry and \"replacing with another hat\" needs deleting the tokens of the original hat and replacing them with new ones from another image." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.69, + 0.788, + 0.841 + ], + "angle": 0, + "content": "For quantitative comparison of the reconstruction fidelity, we employ both the distribution-level metric rFID [24] and the image-patches-level metrics PSNR, SSIM [26], LPIPS [58] to evaluate the reconstructed images of ImageNet validation set. We employ the VQGAN [16], RQ-VAE [32] and DQ-VAE [32] as baselines. The former two employ fixed tokens number of each image, while the latter uses dynamic tokens number through picking one granularity for each patch from three pre-defined strides \\(\\left(\\frac{1}{8},\\frac{1}{16},\\frac{1}{32}\\right)\\). The tokens number of DQ-VAE is computed by the ratio of different strides provided in their paper. Tokens number of our Reader corresponds to the average number observed across the entire ImageNet validation set." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.545, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Region-Native Autoencoder" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.145, + 0.788, + 0.203 + ], + "angle": 0, + "content": "Table 1: Quantitative results for the reconstruction fidelity on the ImageNet validation set. \"Tokens\" refers to the number of image tokens utilized for each image. The dimension of image tokens is 256 for all methods. The \"+\" denotes data excerpted from the original paper. Top 1 and top 2 results are highlighted in red and blue, respectively." + }, + { + "type": "table", + "bbox": [ + 0.224, + 0.213, + 0.77, + 0.326 + ], + "angle": 0, + "content": "
MethodTokensPSNR ↑SSIM ↑LPIPS ↓rFID ↓Params(M)FLOPs(G)
VQGAN(1024)16 × 1619.360.480.318.0572.14196.33
VQGAN(16384)16 × 1619.830.490.284.9276.07197.33
RQ-VAE(16384)†8 × 8 × 210.77
RQ-VAE(16384)8 × 8 × 420.030.500.285.24104.40193.64
DQ-VAE(16384)†2564.08 - 7.32
Reader(ours)79.0924.750.720.202.4326.6040.24
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.354, + 0.788, + 0.475 + ], + "angle": 0, + "content": "Results shown in Tab. 1 indicate that, in comparison to the baselines, our Reader attains superior reconstruction fidelity despite having the smallest tokens number, model parameters and computational complexity. While compared to DQ-VAE, which also employs dynamic length encoding, our method also demonstrates significant superiority. This finding is intuitive that, compared to both the fixed- and dynamic-sized patches, arbitrarily shaped regions can better handle the uneven information density in images and avoid information redundancy, thus achieving higher reconstruction fidelity with fewer tokens." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.499, + 0.496, + 0.515 + ], + "angle": 0, + "content": "4.2 Experiment on Human Face" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.525, + 0.787, + 0.572 + ], + "angle": 0, + "content": "To further evaluate the editing capabilities of our Reader, we transfer it to images of human faces, which are commonly employed in previous image editing works, allowing to provide quantitative comparisons." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.577, + 0.788, + 0.682 + ], + "angle": 0, + "content": "Implementation We fine-tune the Reader in this experiment, which is pretrained as outlined in Sec. 4.1. We utilize the CelebAMask-HQ dataset [31], a face parsing dataset with manual annotations. The input and output sizes of images for this experiment are configured as \\((512 \\times 512)\\) to align with the previous methods [31, 34]. In order to preserving the knowledge already learned by the pre-trained model, we keep the majority of Reader architecture unchanged and make the following modifications." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.684, + 0.788, + 0.79 + ], + "angle": 0, + "content": "(1) Adding a downsampling-and-convolution layer with a residual connection in the region embedder and initializing the new convolution's output to zero. This is employed to make the double-sized masks meet the input length of previous mask projection layer. (2) Increasing the number of learnable patch queries from 1024 to 4096 in order to get double-sized reconstructed images. While loading the pre-trained weights, we perform bicubic interpolation on the weights of the positional embeddings and the patch queries to match the new size." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.788, + 0.843 + ], + "angle": 0, + "content": "Results In Fig. 6, we present the qualitative results corresponding to the three-level editing operations. In the top and middle rows, we performed overlay editing, gradually changing one part of the geometric and visual properties at a" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.351, + 0.129 + ], + "angle": 0, + "content": "Wang et al." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.144, + 0.313, + 0.218 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.315, + 0.144, + 0.407, + 0.218 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.324, + 0.218, + 0.398, + 0.226 + ], + "angle": 0, + "content": "+ Close Mouth" + }, + { + "type": "image", + "bbox": [ + 0.41, + 0.144, + 0.501, + 0.218 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.418, + 0.218, + 0.495, + 0.226 + ], + "angle": 0, + "content": "+ Sharp Jawline" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.144, + 0.596, + 0.218 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.598, + 0.144, + 0.69, + 0.218 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.607, + 0.218, + 0.683, + 0.226 + ], + "angle": 0, + "content": "+ Enlarge Nose" + }, + { + "type": "image", + "bbox": [ + 0.694, + 0.144, + 0.786, + 0.218 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.702, + 0.218, + 0.776, + 0.226 + ], + "angle": 0, + "content": "+Wide Jawline" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.226, + 0.314, + 0.299 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.315, + 0.226, + 0.408, + 0.299 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.41, + 0.226, + 0.501, + 0.299 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.226, + 0.596, + 0.299 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.598, + 0.226, + 0.69, + 0.299 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.693, + 0.226, + 0.786, + 0.299 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.324, + 0.299, + 0.397, + 0.308 + ], + "angle": 0, + "content": "+ Eyes Feature" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.309, + 0.313, + 0.381 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.315, + 0.309, + 0.408, + 0.381 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.419, + 0.299, + 0.488, + 0.308 + ], + "angle": 0, + "content": "+ Lips Featue" + }, + { + "type": "image", + "bbox": [ + 0.41, + 0.308, + 0.5, + 0.381 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.308, + 0.593, + 0.381 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.595, + 0.299, + 0.681, + 0.308 + ], + "angle": 0, + "content": "+ Eyes Feature" + }, + { + "type": "image", + "bbox": [ + 0.597, + 0.308, + 0.691, + 0.381 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.705, + 0.299, + 0.772, + 0.308 + ], + "angle": 0, + "content": "+ Lips Featue" + }, + { + "type": "image", + "bbox": [ + 0.693, + 0.308, + 0.786, + 0.381 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.333, + 0.381, + 0.381, + 0.391 + ], + "angle": 0, + "content": "+Smiling" + }, + { + "type": "image_caption", + "bbox": [ + 0.713, + 0.381, + 0.76, + 0.391 + ], + "angle": 0, + "content": "+Smiling" + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.394, + 0.785, + 0.422 + ], + "angle": 0, + "content": "Fig. 6: Qualitative results in huaman faces. The editing objective is labeled with text under corresponding image." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.428, + 0.312, + 0.5 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.229, + 0.5, + 0.3, + 0.51 + ], + "angle": 0, + "content": "Original Image" + }, + { + "type": "image", + "bbox": [ + 0.314, + 0.428, + 0.407, + 0.5 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.335, + 0.5, + 0.384, + 0.51 + ], + "angle": 0, + "content": "MaskGAN" + }, + { + "type": "image", + "bbox": [ + 0.41, + 0.428, + 0.501, + 0.5 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.442, + 0.5, + 0.467, + 0.51 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.428, + 0.596, + 0.5 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.5, + 0.585, + 0.51 + ], + "angle": 0, + "content": "Original Image" + }, + { + "type": "image", + "bbox": [ + 0.598, + 0.428, + 0.691, + 0.5 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.619, + 0.5, + 0.668, + 0.51 + ], + "angle": 0, + "content": "MaskGAN" + }, + { + "type": "image", + "bbox": [ + 0.693, + 0.428, + 0.786, + 0.5 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.726, + 0.5, + 0.751, + 0.509 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.516, + 0.785, + 0.572 + ], + "angle": 0, + "content": "Fig. 7: Qualitative comparisons in the smile transferring benchmark. Our Reader can selectively manipulate only the target attributes without affecting other features or regions. The red boxes highlight the local details that should remain unchanged after editing. Please kindly zoom in for a better view." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.595, + 0.785, + 0.625 + ], + "angle": 0, + "content": "time. In the bottom row, we simultaneously manipulated both the masks and embeddings, facilitating significant change in expression." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.626, + 0.785, + 0.717 + ], + "angle": 0, + "content": "To quantitatively measure Reader's image editing capabilities, we utilize the smile transferring benchmark introduced by MaskGAN [31], which requires converting the faces with neutral expressions into smiling faces. We compare our Reader with the MaskGAN [31], LocalEditing [9], InterFaceGAN [46] and EditGAN [34], which also provide results in this benchmark. We employ the following metrics to keep consistent with [34]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.718, + 0.785, + 0.809 + ], + "angle": 0, + "content": "(1) Semantic Correctness: Training a binary classifier to recognize the smile attribute and evaluating whether the edited faces exhibit a smiling expression. (2) Distribution-level Image Quality: Utilizing FID [24] and KID [2] to assess the quality of edited images. (3) Identity Preservation: Leveraging a pretrained ArcFace feature extraction network [12] to evaluate whether the face's identity is maintained post-editing." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.785, + 0.84 + ], + "angle": 0, + "content": "To automate smile transferring without manual adjustments, we employ a similar approach to MaskGAN. Specifically, we select 400 pairs of neutral-" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.545, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Region-Native Autoencoder" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "table_caption", + "bbox": [ + 0.215, + 0.145, + 0.788, + 0.187 + ], + "angle": 0, + "content": "Table 2: Quantitative results in the smile transferring benchmark. Top 1 and top 2 results are highlighted in red and blue, respectively. Our Reader significantly outperforms previous methods in the FID and ID Score metrics by a large margin." + }, + { + "type": "table", + "bbox": [ + 0.229, + 0.199, + 0.77, + 0.328 + ], + "angle": 0, + "content": "
MetricAttribute Acc. (%)↑FID↓KID↓ID Score↑
MaskGAN [31]77.346.840.0200.4611
LocalEditing [9]26.041.260.0120.5823
LocalEditing - Encoding4Editing [49]41.848.280.0160.7295
InterFaceGAN [46]83.539.420.0100.7295
EditGAN [34]91.541.740.0130.7047
EditGAN+30 [34]85.840.830.0120.7452
Reader (ours)84.237.910.0110.8470
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.358, + 0.784, + 0.403 + ], + "angle": 0, + "content": "expression and smiling-expression faces with the closest head position as the source and target images. Subsequently, we replace both the tokens corresponding to mouth and lips of the source image with those from the target image." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.404, + 0.788, + 0.587 + ], + "angle": 0, + "content": "Results shown in Tab. 2 demonstrate that our Reader achieves excellent performance across all metrics. Specifically, our Reader significantly outperforms previous methods in the FID and ID Score metrics by a large margin. This indicates that the edited images produced by Reader exhibit higher quality while better preserving the ID features of the original images. This observation effectively highlights the inherent advantage of our method. By representing the image through a series of decoupled tokens, Reader can selectively manipulate only the target attributes without affecting other regions. To provide a more intuitive perception, we illustrate some examples of the original and edited image pairs in Fig. 7. It can be observed that our method effectively prevents changes in the areas which are not the editing target, while achieving editing goals. In contrast, the GAN-based baseline [31] is prone to inducing global changes." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.613, + 0.39, + 0.628 + ], + "angle": 0, + "content": "4.3 Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.643, + 0.788, + 0.779 + ], + "angle": 0, + "content": "In Tab. 3a, we conducted ablation studies to evaluate the effectiveness of some optional hyper-parameters and modules in the model architecture, where different configurations represent: Patch Size: The patch size of the ViT in Reader's encoder. Bidirectional: Whether to use bidirectional transformer block in the region decoder. Masked-Attn: Whether to employ the masked cross-attention layer in the region embedder. Results indicate that a ViT with a smaller patch size paired with the additional modules produces the best image fidelity. This demonstrates that our default settings can better preserve the fine-grained details in the image, thereby enhancing the quality of the output image." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.841 + ], + "angle": 0, + "content": "In Tab. 3b, we conducted ablation experiment to determine the granularity of pseudo-labels utilized in training. In this experiment, during preprocessing, we only removed the overlapping parts from the pseudo-labels and retained masks of all different sizes to train our Reader. Subsequently, we calculated the average" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.351, + 0.129 + ], + "angle": 0, + "content": "Wang et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.784, + 0.174 + ], + "angle": 0, + "content": "Table 3: Ablation studies on both model architecture, hyper parameter and data preprocessing. Our default setting is marked with gray." + }, + { + "type": "table", + "bbox": [ + 0.25, + 0.185, + 0.572, + 0.26 + ], + "angle": 0, + "content": "
Patch SizeBidirectionalMasked-AttnrFID ↓
8 × 8XX98.39
✓X37.50
✓✓11.42
16 × 16✓✓61.64
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.329, + 0.262, + 0.483, + 0.274 + ], + "angle": 0, + "content": "(a) Module configurations." + }, + { + "type": "table", + "bbox": [ + 0.599, + 0.185, + 0.755, + 0.26 + ], + "angle": 0, + "content": "
Area ProportionL1 ↓
< 1/321.44
1/32 ~ 1/161.56
1/16 ~ 1/82.53
> 1/83.46
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.576, + 0.262, + 0.754, + 0.274 + ], + "angle": 0, + "content": "(b) Annotation configurations." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.294, + 0.787, + 0.387 + ], + "angle": 0, + "content": "L1 distance of different-sized regions between the original images and the reconstructed images. The \"Area Proportion\" in Tab. 3b represents the proportion of the region's area relative to the entire image. The results indicate that when the region's area proportion is less than \\(\\frac{1}{16}\\), the L1 distance stabilizes with changes in area. Therefore, in the formal experiments, we only retained pseudo-labels with area proportion smaller than \\(\\frac{1}{16}\\) for training." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.413, + 0.509, + 0.429 + ], + "angle": 0, + "content": "5 Conclusion and Discussion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.45, + 0.788, + 0.602 + ], + "angle": 0, + "content": "In conclusion, this paper explores an innovative region-based visual token representation and presents Reader. Reader employs arbitrarily shaped and sized regions as the units of image tokens, which aligns better with the visual prior: the non-uniformity of information density and the naturally existing semantic boundaries. Through experiments, we demonstrate the following inherent advantages of this region-based tokenization method. Efficiency: This method has a high compression ratio and can achieve superior reconstruction fidelity while employing a significantly smaller number of tokens for each image. Disentanglement: The region-based tokens guarantees their disentanglement from one another, and further empower Reader with the ability to perform image editing." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.607, + 0.788, + 0.713 + ], + "angle": 0, + "content": "Limitation and Feature Work As this paper focuses on exploring the feasibility and effectiveness of region-based tokenization, we adopted a minimal configuration in our experiments, manifested as employing the tiny-version transformer and training model on relatively small-sized images \\((256\\times 256)\\). Although Reader can produce already satisfactory results under such configurations, the parameter size and data resolution still limit its broader application. For example, some high-resolution images may exhibit a degree of blurring after editing." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.713, + 0.788, + 0.76 + ], + "angle": 0, + "content": "In the future work, we will pursue scaling up both model parameters and the volume of data. Additionally, we aim to explore its application in various scenarios, such as interaction with LLMs and multimodal perception, generation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Broader Impact Reader possesses the capability to serve as an image editor, yielding numerous interesting outcomes, but also opens up the possibility of real photo manipulation for nefarious purposes. However, with the recent advancements in watermarking techniques and manipulated images identifying techniques [17], this risk is avoidable." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.545, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "Region-Native Autoencoder" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.403, + 0.163 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.181, + 0.785, + 0.212 + ], + "angle": 0, + "content": "This research was funded by the Fundamental Research Funds for the Central Universities (2024XKRC082) and the National NSF of China (No.U23A20314)." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.24, + 0.323, + 0.255 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.274, + 0.786, + 0.317 + ], + "angle": 0, + "content": "1. Anderson, P., He, X., Buehler, C., Teney, D., Johnson, M., Gould, S., Zhang, L.: Bottom-up and top-down attention for image captioning and visual question answering. In: CVPR. pp. 6077-6086 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.318, + 0.786, + 0.345 + ], + "angle": 0, + "content": "2. Binkowski, M., Sutherland, D.J., Arbel, M., Gretton, A.: Demystifying MMD gans. In: ICLR. OpenReview.net (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.347, + 0.786, + 0.374 + ], + "angle": 0, + "content": "3. Brooks, T., Holynski, A., Efros, A.A.: Instructpix2pix: Learning to follow image editing instructions. In: CVPR. pp. 18392-18402. IEEE (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.375, + 0.786, + 0.416 + ], + "angle": 0, + "content": "4. Burgess, C.P., Matthew, L., Watters, N., Kabra, R., Higgins, I., Botvinick, M.M., Lerchner, A.: Monet: Unsupervised scene decomposition and representation. CoRR abs/1901.11390 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.418, + 0.786, + 0.472 + ], + "angle": 0, + "content": "5. Carion, N., Massa, F., Synnaeve, G., Usunier, N., Kirillov, A., Zagoruyko, S.: End-to-end object detection with transformers. In: Vedaldi, A., Bischof, H., Brox, T., Frahm, J. (eds.) ECCV. Lecture Notes in Computer Science, vol. 12346, pp. 213-229. Springer (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.474, + 0.786, + 0.514 + ], + "angle": 0, + "content": "6. Cheng, B., Misra, I., Schwing, A.G., Kirillov, A., Girdhar, R.: Masked-attention mask transformer for universal image segmentation. In: CVPR. pp. 1290–1299 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.516, + 0.786, + 0.557 + ], + "angle": 0, + "content": "7. Cheng, B., Schwing, A.G., Kirillov, A.: Per-pixel classification is not all you need for semantic segmentation. In: Ranzato, M., Beygelzimer, A., Dauphin, Y.N., Liang, P., Vaughan, J.W. (eds.) NeurIPS. pp. 17864-17875 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.559, + 0.786, + 0.6 + ], + "angle": 0, + "content": "8. Choi, Y., Choi, M., Kim, M., Ha, J., Kim, S., Choo, J.: Stargan: Unified generative adversarial networks for multi-domain image-to-image translation. In: CVPR. pp. 8789-8797. Computer Vision Foundation / IEEE Computer Society (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.601, + 0.786, + 0.642 + ], + "angle": 0, + "content": "9. Collins, E., Bala, R., Price, B., Süssstrunk, S.: Editing in style: Uncovering the local semantics of gans. In: CVPR. pp. 5770-5779. Computer Vision Foundation / IEEE (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.643, + 0.786, + 0.671 + ], + "angle": 0, + "content": "0. Couairon, G., Verbeek, J., Schwenk, H., Cord, M.: Diffedit: Diffusion-based semantic image editing with mask guidance. In: ICLR. OpenReview.net (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.672, + 0.786, + 0.713 + ], + "angle": 0, + "content": "1. Crowson, K., Biderman, S., Kornis, D., Stander, D., Hallahan, E., Castricato, L., Raff, E.: Vqgan-clip: Open domain image generation and editing with natural language guidance. In: ECCV. pp. 88–105. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.714, + 0.786, + 0.755 + ], + "angle": 0, + "content": "2. Deng, J., Guo, J., Xue, N., Zafeiriou, S.: Arcface: Additive angular margin loss for deep face recognition. In: CVPR. pp. 4690-4699. Computer Vision Foundation / IEEE (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.757, + 0.786, + 0.811 + ], + "angle": 0, + "content": "3. Devlin, J., Chang, M., Lee, K., Toutanova, K.: BERT: pre-training of deep bidirectional transformers for language understanding. In: Burstein, J., Doran, C., Solorio, T. (eds.) NAACL-HLT. pp. 4171-4186. Association for Computational Linguistics (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.812, + 0.786, + 0.84 + ], + "angle": 0, + "content": "4. Ding, H., Liu, C., He, S., Jiang, X., Loy, C.C.: MeViS: A large-scale benchmark for video segmentation with motion expressions. In: ICCV (2023)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.274, + 0.786, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.351, + 0.129 + ], + "angle": 0, + "content": "Wang et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.787, + 0.203 + ], + "angle": 0, + "content": "15. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., Uszkoreit, J., Houlsby, N.: An image is worth 16x16 words: Transformers for image recognition at scale. In: ICLR. OpenReview.net (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.203, + 0.787, + 0.243 + ], + "angle": 0, + "content": "16. Esser, P., Rombach, R., Ommer, B.: Taming transformers for high-resolution image synthesis. In: CVPR. pp. 12873-12883. Computer Vision Foundation / IEEE (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.243, + 0.787, + 0.271 + ], + "angle": 0, + "content": "17. Fernandez, P., Couairon, G., Jégou, H., Douze, M., Furon, T.: The stable signature: Rooting watermarks in latent diffusion models. CoRR abs/2303.15435 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.271, + 0.787, + 0.311 + ], + "angle": 0, + "content": "18. Gafni, O., Polyak, A., Ashual, O., Sheynin, S., Parikh, D., Taigman, Y.: Make-a-scene: Scene-based text-to-image generation with human priors. In: ECCV. pp. 89-106. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.311, + 0.787, + 0.352 + ], + "angle": 0, + "content": "19. Goel, V., Peruzzo, E., Jiang, Y., Xu, D., Sebe, N., Darrell, T., Wang, Z., Shi, H.: Pair-diffusion: Object-level image editing with structure-and-appearance paired diffusion models. CoRR abs/2303.17546 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.352, + 0.787, + 0.42 + ], + "angle": 0, + "content": "20. Gondal, M.W., Wuthrich, M., Miladinovic, D., Locatello, F., Breidt, M., Volchkov, V., Akpo, J., Bachem, O., Scholkopf, B., Bauer, S.: On the transfer of inductive bias from simulation to the real world: a new disentanglement dataset. In: Wallach, H.M., Larochelle, H., Beygelzimer, A., d'Alché-Buc, F., Fox, E.B., Garnett, R. (eds.) NeurIPS. pp. 15714-15725 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.42, + 0.787, + 0.461 + ], + "angle": 0, + "content": "21. Goodfellow, I.J., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., Courville, A.C., Bengio, Y.: Generative adversarial nets. In: Ghahramani, Z., Welling, M., Cortes, C., Lawrence, N.D., Weinberger, K.Q. (eds.) NeurIPS (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.461, + 0.787, + 0.502 + ], + "angle": 0, + "content": "22. Härkönen, E., Hertzmann, A., Lehtinen, J., Paris, S.: Ganspace: Discovering interpretable GAN controls. In: Larochelle, H., Ranzato, M., Hadsell, R., Balcan, M., Lin, H. (eds.) NeurIPS (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.502, + 0.787, + 0.529 + ], + "angle": 0, + "content": "23. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: CVPR. pp. 770-778. IEEE Computer Society (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.529, + 0.787, + 0.583 + ], + "angle": 0, + "content": "24. Heusel, M., Ramsauer, H., Unterthiner, T., Nessler, B., Hochreiter, S.: Gans trained by a two time-scale update rule converge to a local nash equilibrium. In: Guyon, I., von Luxburg, U., Bengio, S., Wallach, H.M., Fergus, R., Vishwanathan, S.V.N., Garnett, R. (eds.) NeurIPS. pp. 6626-6637 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.583, + 0.787, + 0.61 + ], + "angle": 0, + "content": "25. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. In: Larochelle, H., Ranzato, M., Hadsell, R., Balcan, M., Lin, H. (eds.) NeurIPS (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.61, + 0.787, + 0.637 + ], + "angle": 0, + "content": "26. Horé, A., Ziou, D.: Image quality metrics: PSNR vs. SSIM. In: ICPR. pp. 2366-2369. IEEE Computer Society (2010)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.637, + 0.787, + 0.678 + ], + "angle": 0, + "content": "27. Huang, M., Mao, Z., Chen, Z., Zhang, Y.: Towards accurate image coding: Improved autoregressive image generation with dynamic vector quantization. In: CVPR. pp. 22596-22605 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.678, + 0.787, + 0.718 + ], + "angle": 0, + "content": "28. Isola, P., Zhu, J., Zhou, T., Efros, A.A.: Image-to-image translation with conditional adversarial networks. In: CVPR. pp. 5967-5976. IEEE Computer Society (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.718, + 0.787, + 0.759 + ], + "angle": 0, + "content": "29. Kim, H., Mnih, A.: Disentangling by factorising. In: Dy, J.G., Krause, A. (eds.) ICML. Proceedings of Machine Learning Research, vol. 80, pp. 2654-2663. PMLR (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.759, + 0.787, + 0.8 + ], + "angle": 0, + "content": "30. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W., Dollár, P., Girshick, R.B.: Segment anything. CoRR abs/2304.02643 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.8, + 0.787, + 0.841 + ], + "angle": 0, + "content": "31. Lee, C., Liu, Z., Wu, L., Luo, P.: Maskgan: Towards diverse and interactive facial image manipulation. In: CVPR. pp. 5548-5557. Computer Vision Foundation / IEEE (2020)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.545, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "Region-Native Autoencoder" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.202 + ], + "angle": 0, + "content": "32. Lee, D., Kim, C., Kim, S., Cho, M., Han, W.: Autoregressive image generation using residual quantization. In: CVPR. pp. 11513-11522. IEEE (2022). https://doi.org/10.1109/CVPR52688.2022.01123, https://doi.org/10.1109/CVPR52688.2022.01123" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.203, + 0.787, + 0.23 + ], + "angle": 0, + "content": "33. Li, Y., Mao, H., Girshick, R., He, K.: Exploring plain vision transformer backbones for object detection. In: ECCV. pp. 280-296. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.23, + 0.786, + 0.271 + ], + "angle": 0, + "content": "34. Ling, H., Kreis, K., Li, D., Kim, S.W., Torralba, A., Fidler, S.: Editgan: High-precision semantic image editing. In: Ranzato, M., Beygelzimer, A., Dauphin, Y.N., Liang, P., Vaughan, J.W. (eds.) NeurIPS. pp. 16331-16345 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.271, + 0.786, + 0.298 + ], + "angle": 0, + "content": "35. Liu, C., Ding, H., Jiang, X.: GRES: Generalized referring expression segmentation. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.298, + 0.787, + 0.325 + ], + "angle": 0, + "content": "36. Liu, C., Jiang, X., Ding, H.: Primitivenet: decomposing the global constraints for referring segmentation. Visual Intelligence 2(1), 16 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.325, + 0.787, + 0.379 + ], + "angle": 0, + "content": "37. Locatello, F., Weissenborn, D., Unterthiner, T., Mahendran, A., Heigold, G., Uszkoreit, J., Dosovitskiy, A., Kipf, T.: Object-centric learning with slot attention. In: Larochelle, H., Ranzato, M., Hadsell, R., Balcan, M., Lin, H. (eds.) NeurIPS (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.379, + 0.787, + 0.42 + ], + "angle": 0, + "content": "38. Ning, J., Li, C., Zhang, Z., Wang, C., Geng, Z., Dai, Q., He, K., Hu, H.: All in tokens: Unifying output space of visual tasks via soft token. In: ICCV. pp. 19900-19910 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.42, + 0.787, + 0.461 + ], + "angle": 0, + "content": "39. van den Oord, A., Vinyals, O., Kavukcuoglu, K.: Neural discrete representation learning. In: Guyon, I., von Luxburg, U., Bengio, S., Wallach, H.M., Fergus, R., Vishwanathan, S.V.N., Garnett, R. (eds.) NeurIPS. pp. 6306-6315 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.461, + 0.787, + 0.515 + ], + "angle": 0, + "content": "40. Pan, X., Tewari, A., Leimkuhler, T., Liu, L., Meka, A., Theobalt, C.: Drag your GAN: interactive point-based manipulation on the generative image manifold. In: Brunvand, E., Sheffer, A., Wimmer, M. (eds.) SIGGRAPH. pp. 78:1-78:11. ACM (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.515, + 0.787, + 0.556 + ], + "angle": 0, + "content": "41. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: ICML. pp. 8748-8763. PMLR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.556, + 0.787, + 0.583 + ], + "angle": 0, + "content": "42. Radford, A., Narasimhan, K., Salimans, T., Sutskever, I., et al.: Improving language understanding by generative pre-training (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.583, + 0.787, + 0.624 + ], + "angle": 0, + "content": "43. Reed, S.E., Zhang, Y., Zhang, Y., Lee, H.: Deep visual analogy-making. In: Cortes, C., Lawrence, N.D., Lee, D.D., Sugiyama, M., Garnett, R. (eds.) NeurIPS. pp. 1252-1260 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.624, + 0.787, + 0.651 + ], + "angle": 0, + "content": "44. Ren, S., He, K., Girshick, R., Sun, J.: Faster r-cnn: Towards real-time object detection with region proposal networks. NeurIPS 28 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.651, + 0.787, + 0.691 + ], + "angle": 0, + "content": "45. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: CVPR. pp. 10674-10685. IEEE (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.691, + 0.787, + 0.718 + ], + "angle": 0, + "content": "46. Shen, Y., Yang, C., Tang, X., Zhou, B.: Interfacegan: Interpreting the disentangled face representation learned by gans. IEEE TPAMI 44(4), 2004-2018 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.718, + 0.787, + 0.759 + ], + "angle": 0, + "content": "47. Shuai, X., Ding, H., Ma, X., Tu, R., Jiang, Y.G., Tao, D.: A survey of multimodal-guided image editing with text-to-image diffusion models. arXiv preprint arXiv:2406.14555 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.759, + 0.787, + 0.813 + ], + "angle": 0, + "content": "48. Sohl-Dickstein, J., Weiss, E.A., Maheswaranathan, N., Ganguli, S.: Deep unsupervised learning using nonequilibrium thermodynamics. In: Bach, F.R., Blei, D.M. (eds.) ICML. JMLR Workshop and Conference Proceedings, vol. 37, pp. 2256-2265. JMLR.org (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.813, + 0.787, + 0.841 + ], + "angle": 0, + "content": "49. Tov, O., Alaluf, Y., Nitzan, Y., Patashnik, O., Cohen-Or, D.: Designing an encoder for stylegan image manipulation. ACM TOG 40(4), 133:1-133:14 (2021)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.351, + 0.129 + ], + "angle": 0, + "content": "Wang et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.204 + ], + "angle": 0, + "content": "50. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. In: Guyon, I., von Luxburg, U., Bengio, S., Wallach, H.M., Fergus, R., Vishwanathan, S.V.N., Garnett, R. (eds.) NeurIPS. pp. 5998-6008 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.205, + 0.787, + 0.245 + ], + "angle": 0, + "content": "51. Wu, B., Xu, C., Dai, X., Wan, A., Zhang, P., Yan, Z., Tomizuka, M., Gonzalez, J., Keutzer, K., Vajda, P.: Visual transformers: Token-based image representation and processing for computer vision. arXiv preprint arXiv:2006.03677 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.246, + 0.787, + 0.286 + ], + "angle": 0, + "content": "52. Wu, R., Zhang, G., Lu, S., Chen, T.: Cascade EF-GAN: progressive facial expression editing with local focuses. In: CVPR. pp. 5020-5029. Computer Vision Foundation / IEEE (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.287, + 0.787, + 0.327 + ], + "angle": 0, + "content": "53. Xu, J., Mello, S.D., Liu, S., Byeon, W., Breuel, T.M., Kautz, J., Wang, X.: Groupvit: Semantic segmentation emerges from text supervision. In: CVPR. pp. 18113-18123. IEEE (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.328, + 0.787, + 0.383 + ], + "angle": 0, + "content": "54. Yang, B., Gu, S., Zhang, B., Zhang, T., Chen, X., Sun, X., Chen, D., Wen, F.: Paint by example: Exemplar-based image editing with diffusion models. In: CVPR. pp. 18381-18391. IEEE (2023). https://doi.org/10.1109/CVPR52729.2023.01763, https://doi.org/10.1109/CVPR52729.2023.01763" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.384, + 0.787, + 0.411 + ], + "angle": 0, + "content": "55. Yang, T., Wang, Y., Lu, Y., Zheng, N.: Visual concepts tokenization. In: NeurIPS (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.412, + 0.787, + 0.452 + ], + "angle": 0, + "content": "56. Yu, J., Li, X., Koh, J.Y., Zhang, H., Pang, R., Qin, J., Ku, A., Xu, Y., Baldridge, J., Wu, Y.: Vector-quantized image modeling with improved VQGAN. In: ICLR. OpenReview.net (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.453, + 0.787, + 0.479 + ], + "angle": 0, + "content": "57. Zellers, R., Yatskar, M., Thomson, S., Choi, Y.: Neural motifs: Scene graph parsing with global context. In: CVPR. pp. 5831-5840 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.48, + 0.787, + 0.522 + ], + "angle": 0, + "content": "58. Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual metric. In: CVPR. pp. 586-595. Computer Vision Foundation / IEEE Computer Society (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.523, + 0.787, + 0.55 + ], + "angle": 0, + "content": "59. Zhong, Y., Wang, L., Chen, J., Yu, D., Li, Y.: Comprehensive image captioning via scene graph decomposition. In: ECCV. pp. 211-229. Springer (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.55, + 0.787, + 0.577 + ], + "angle": 0, + "content": "60. Zhu, J., Yang, C., Shen, Y., Shi, Z., Zhao, D., Chen, Q.: Linkgan: Linking GAN latents to pixels for controllable image synthesis. CoRR abs/2301.04604 (2023)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.577 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/Region-Native Visual Tokenization/a10ef64d-036f-4197-bc89-c920baeb7773_origin.pdf b/2024/Region-Native Visual Tokenization/a10ef64d-036f-4197-bc89-c920baeb7773_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..9c9de1a3e85faf380582e2ed45f1372e2b4fd6e4 --- /dev/null +++ b/2024/Region-Native Visual Tokenization/a10ef64d-036f-4197-bc89-c920baeb7773_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:62256863a3a7a54093df8e8f71545ad9b8711fd22af2123cc07d1a988970a490 +size 6196673 diff --git a/2024/Region-Native Visual Tokenization/full.md b/2024/Region-Native Visual Tokenization/full.md new file mode 100644 index 0000000000000000000000000000000000000000..8c2efd2ebaf542b91011848aac213d4db4f02a2e --- /dev/null +++ b/2024/Region-Native Visual Tokenization/full.md @@ -0,0 +1,394 @@ +# Region-Native Visual Tokenization + +Mengyu Wang $^{1,5(\dagger)}$ $\odot$ , Yuyao Huang $^{1,5(\dagger,\ddagger)}$ $\odot$ , Henghui Ding $^{4}$ $\odot$ , Xinlong Wang $^{2}$ $\odot$ , Tiejun Huang $^{2}$ , Yao Zhao $^{1,5,6}$ $\odot$ , Yunchao Wei $^{1,5,6(\boxtimes)}$ $\odot$ , and Shuicheng Yan $^{3}$ $\odot$ + +$^{1}$ Institute of Information Science, Beijing Jiaotong University, Beijing, China + $^{2}$ Beijing Academy of Artificial Intelligence + $^{3}$ Skywork AI + +$^{4}$ Institute of Big Data, Fudan University, Shanghai, China + $^{5}$ Visual Intelligence + X International Joint Laboratory of the Ministry of Education + $^{6}$ PengCheng Laboratory, Shenzhen, China + +Abstract. We explore an innovative region-based visual token representation and present the REgion-native AutoencoDER (Reader). In contrast to the majority of previous methods, which represent each image as a grid-shaped tokens map, Reader perceives each image into sequential region-based tokens, with each token corresponding to an object or one part of an object in the image. Specifically, Reader comprises both an encoder and a decoder. The encoder can partition each image into an adaptive number of arbitrary-shaped regions and encode each region into a token. Subsequently, the decoder utilizes this adaptive-length token sequence to reconstruct the original image. Experimental results demonstrate that such region-based token representation possesses two main notable characteristics. Firstly, it achieves highly efficient image encoding. Reader can adaptively use more regions to represent complex areas and fewer regions in simpler ones, thus avoiding information redundancy. Consequently, it achieves superior reconstruction fidelity compared to previous methods, despite using significantly fewer tokens for each image. Secondly, the region-based manner enables manipulation on a local region without causing global changes. As a result, Reader inherently supports diverse image editing operations, including erasing, adding, replacing, and modifying shapes on the objects, and achieves excellent performance in the image editing benchmark of smile transferring. Code is provided at https://github.com/MengyuWang826/Reader.git. + +# 1 Introduction + +Tokenization, originating from Natural Language Processing (NLP), involves breaking a textual data stream into several elements called tokens. It plays an important role for converting raw data into a form that is predictable and analyzable by deep learning systems [13, 42]. With the gradual maturity of tokenization technique in the NLP field, it has also been adapted to the Computer + +![](images/d4c748bfbeb5fe9442a095a84505b3facf700546a43524df61b7e95f775eab59.jpg) +(a) Image + +![](images/38b04be8ce5feeda681d079903de032e8014071e4115c059ba7054a19bb5c483.jpg) +(b) Grid-based + +![](images/922722d1bdf5da29da6d2e8202bbd3376d8e91679c64782064ef824e68817403.jpg) +(c) Object-based +Fig.1: Illustration of different types of visual tokenization methods. Grid-based methods [16,56] represent images as grid-shaped tokens map. Object-based methods [1,59] use detection models to represent objects as tokens. While our region-based method represents arbitrarily shaped regions as tokens, which offers more flexibility and retains the naturally existed boundaries presented in images. + +![](images/1250be1d8897bed585b3d0e2a6ac2e6db7f4cb06a5cdfe6367efc60dbccc1b4d.jpg) +(d) Region-based + +Vision (CV) domain, known as visual tokenization, and has become the foundation for many tasks such as image/video generation [16, 27, 56] and multimodal generation/perception [11, 14, 18, 38]. + +Existing visual tokenization methods can be categorized into two main types, which we refer to as grid-based and object-based. Grid-based methods typically employ a CNN [23] or ViT [15] backbone to encode the original image into a downsampled grid-shaped feature map, followed by quantization of the feature map to get the image tokens. However, in such methods, each token corresponds to a patch in the original image (see Fig. 1b), which does not adhere to the following characteristics of the image prior. Firstly, different areas in images have varying information densities. While in these methods, patches are treated equally, leading to information loss in densely informative patches and redundancy in sparsely informative ones. Secondly, areas with different semantics within an image have natural boundaries. While employing patches as the units of image tokens disregards the natural boundaries, causing local information within the original image to be distributed and coupled across multiple tokens, limiting the extraction and utilization of local features. Object-based methods [1,35,51,53,59], as an alternative and more natural approach, typically appear in vision-language pretraining. These methods employ detection or segmentation models (see Fig. 1c) to encode the objects within images into tokens, which are then interacted with or aligned with language information. However, these methods only focus on extracting high-level semantic information, neglecting low-level visual details in the image. Consequently, they cannot be used for tasks that require image outputs, such as image synthesis and editing. + +These challenges suggest that, an ideal visual token representation should take into account the crucial visual-prior characteristics, including the uneven information density and the inherent semantic boundaries, and further achieve local-information decoupling and global-information compression. However, previous practices either ignore the visual-prior characteristics or, due to their coarse granularity, result in significant information loss. To address these problems, we + +![](images/5270a8fc731de1bbd8f72d290eee3e5e0d29ec9c7742f92e389c4989eaaf4407.jpg) +Erasing the two people + +![](images/b20413719837a8e504e08eb997d723e574f4c1f86223e040aca868cbe77903b7.jpg) + +![](images/411afb2488cddc08e64f14916b58abcad901e89f1cca17e5d72a9c294bc4162d.jpg) + +![](images/2257e285b40fe8cc01d06188e4793d0df4fb9de90df4cb121966ac65c0f4ef56.jpg) + +![](images/e7620b9cc2feef062ab16ea4037ece79f3a495db031d6a6098830c823e54717b.jpg) + +![](images/295ce50e3e49d94eb6a0f37f6950586437262c29321bb0b53e1670090571b9c6.jpg) +mushroom + +![](images/e5874e47f54b65f77e6652ee29469d17e8d8717534f2fc772e9776b5be12935f.jpg) +Slimming the jawline and enlarging the eyes +Fig. 2: Examples of the proposed Reader performing various editing operations. Reader showcases its strengths in flexibly editing multiple attributes of existing real images under several interactive operations (see Sec. 3.4). In each pair, the left displays the original image, the right displays the edited result and the text describes the corresponding editing objective. + +![](images/23e9da7a1aa570467727496c9feb26855ddb55174e60ecd306ff5bd41c203ef9.jpg) + +![](images/af48cf96e9542c7e1854a45f9fa75787598b9a26d4c3b26b0c77d4d2b1c55e7a.jpg) +Non-smiling expressions to smiling expressions + +![](images/54825d6133390800e1a0171abe75761903599596fb1703bb6d12d0e5247d1a7f.jpg) + +![](images/6b73c694b020ec4ff185f59699e65788f7e71a913010901b7c57739583ff7329.jpg) + +propose the REGion-native AutoencoDER (Reader), which takes arbitrarily shaped and sized regions (see Fig. 1d) as the fundamental units for visual tokens, providing more flexibility and aligning better with the visual-prior characteristics. The Reader framework conceptualizes each image as a composite of several regions, with each region corresponding to either a small-sized object or a portion of a regular or large-sized object. + +Specifically, the Reader architecture comprises an encoder and a decoder, as illustrated in Fig. 3. The encoder first partitions an image into an adaptive-number of non-overlapping regions, where each region is represented as a binary mask and an embedding. Following this, the encoder employs a novel region embedder to integrate the geometric information from the mask with the visual features from the embedding, and produce the desired token for each region. Subsequently, a novel cross-attention-based decoder is employed to reconstruct the original image from these tokens. + +We further conduct both qualitative and quantitative experiments and demonstrate the following two main notable characteristics of Reader: + +Efficiency: Reader achieves highly efficient encoding, yielding reconstruction fidelity superior to that of previous methods, despite employing a significantly smaller number of tokens. This is intuitive since the region-based manner enables Reader to naturally represent complex areas with more regions, while representing simpler areas with fewer regions. This effectively avoids information redundancy, thus possessing inherent efficiency. + +Disentanglement: Representing non-overlapping regions as tokens guarantees their disentanglement from one another, and further empower Reader with the ability to perform image editing. Taking object erasure as an example, the disentanglement permits an object in the image to be represented solely by the tokens corresponding to its interior. Through removing these tokens and then reconstructing the image, object erasing can be naturally achieved. Consequently, without any additional modules or training data, Reader facilitates diverse object-level editing operations (see Fig. 2), such as erasing, adding, re + +placing, and modifying shapes on objects, and demonstrates excellent performance in the image editing benchmark of smile transferring [31]. + +We believe that these experiments effectively demonstrate the inherent advantages of region-based tokenization method. We hope this will bring some novel insights to the community about visual tokenization and unlock its value in various downstream tasks in the future, such as interaction with LLMs and multimodal perception, generation, and beyond. + +# 2 Related Work + +Grid-based Tokenization The majority of grid-based methods follows the feature extraction, quantization, and reconstruction paradigm, initially introduced by VQVAE [39]. These methods represent image tokens as a grid-shaped code map. Expanding upon this framework, VQGAN [16] incorporates perceptual loss [58] and GAN loss [28] to augment the original L1 reconstruction loss, with the aim of improving both reconstruction fidelity and codebook learning. RQ-VAE [32] introduces residual quantization to further enhance reconstruction fidelity, while leading to the cost of significantly higher number of tokens. ViT-VQGAN [56] replaces the previously employed CNN-based architecture with a ViT-based one. Sharing a similar motivation to ours, DQ-VAE [27] proposes dynamically selecting patch granularity between two or three pre-defined strides to achieve dynamic-length coding. However, this approach remains constrained by the grid-based paradigm. In contrast, our Reader takes arbitrarily shaped regions as image token units and offers greater flexibility. + +Object-based Tokenization Methods of this kind typically find applications in vision-language pretraining and multimodal perception. For instance, [1] uses a Faster R-CNN [44] for proposals prediction and encodes the proposals into tokens to subsequently perform image captioning or visual question answering. Similarly, [59] also employs a Faster R-CNN to extract object tokens and maps these tokens with a Graph Neural Network [57]. GroupViT [53] proposes a learnable clustering architecture based on ViT [15] to extract object information, and aligns it with text embedding akin to CLIP [41]. However, these methods focus on extracting semantic information at object-level, leading to excessively coarse-grained and overlooking on detailed visual information. In contrast, our Reader employs an autoencoder architecture to extract dynamic-grained local regions as the unit of image tokens, which better preserves detailed information and showcases superior performance in image reconstruction and editing. + +Scene Decomposition The scene decomposition task [4,37,55], also referred to as object-centric representation, has the most similar form to our Reader. This task aims to decompose a scene into multiple objects and learn the representation of each object. However, these works focus on application in some synthetic datasets containing only basic 3D shapes [20,29,43]. In contrast, our Reader can be applied to a wide range of natural images. + +Image Editing Most early image editing methods are GAN-based [21]. Some of these approaches [22, 34, 40, 60] rely on the controllability of the GAN's latent space, employing supervised or unsupervised approaches to achieve desired transformations of GAN's latent variable. Other works [8, 31, 52] utilize GANs' conditional generation ability and synthesize the edited images based on the input conditions. The remarkable success of diffusion models [25, 48] in image synthesis has also led to the emergence of numerous diffusion-based editing methods [3, 10, 19, 47, 54]. These approaches largely build upon the robust semantic understanding and rendering capabilities of large pre-trained diffusion models [45]. + +The semantic-map based methods [19, 31, 34] can produce the most similar effects to our Reader. However, compared to these works, our Reader benefits from its region-based representation, enabling a wider variety of editing operations without requiring additional input, with details elaborated in Sec. 3.4. + +# 3 Methodology + +We propose Reader (region-native autoencoder) in this work, which performs region-based visual tokenization and enables diverse controllable image editing operations. In the following, we will first provide detailed descriptions about the model architecture of the Reader's encoder and decoder in Sec. 3.1 and Sec. 3.2, respectively. Sec. 3.3 outlines the objective function utilized, while Sec. 3.4 delves into the detailed implementation for image editing. The model architecture is illustrated in Fig. 3. + +# 3.1 Encoder Architecture + +The encoder of Reader needs the capability to first segment an image into numerous regions and then encode each region into a token. Thus, we propose a encoder with three core components: an image feature extractor, a region decoder and a region embedder. + +Feature Extractor To extract information from the original image for subsequent image segmentation, tokenization, and reconstruction steps, we employ a standard ViT [15] along with a simple FPN [33] as our feature extractor. Given the importance of fine-grained details for high-fidelity reconstruction, we configure the patch size of the ViT to be $8 \times 8$ , a smaller size compared to conventional settings, yet effective in preserving local structure. The strides of the simple FPN are set to $(1, \frac{1}{2})$ , thus the output features are at resolutions of $\frac{1}{8}$ and $\frac{1}{4}$ of the original image. + +Region Decoder We employ a region decoder to learn the region representation and segment image into regions. Specifically, we employ a standard Transformer decoder [50], which is similar to previous Transformer-based segmentation works [5-7,36], augmented with an additional bidirectional Transformer block [30]. + +In the Transformer decoder, $N$ learnable embeddings are used as input queries to extract the region information, termed region queries. Within each + +layer of the Transformer decoder, one of the two-sized image features alternately serves as the key and value. Through cross-attention, the region queries can extract region-level information from the image features. Subsequently, the bidirectional Transformer block takes the updated region queries and the larger-sized image feature as input to perform both region-to-image and image-to-region attention and update them. This process yields the region embeddings $\mathcal{E}_r \in \mathbb{R}^{N \times c}$ and a region-aware image feature $\mathcal{E}_{img} \in \mathbb{R}^{c \times h \times w}$ . Here, $(h, w)$ represents the resolution of the image feature, which is $\frac{1}{4}$ of the original image size, and $c$ denotes the channel dimension. + +We then compute the similarity matrix between the region embeddings and the image feature via a dot product, formulated as: + +$$ +S = \mathcal {E} _ {r} \odot \mathcal {E} _ {i m g} \in \mathcal {R} ^ {N \times h \times w}. \tag {1} +$$ + +We use this similarity matrix to get the desired regions. To prevent information redundancy caused by overlapping, we apply an argmax operation along the first dimension of the similarity matrix. This process means matching each pixel with one embedding that has the highest similarity to the pixel's feature, which can be formulated as: + +$$ +S _ {\text {h a r d}} = \operatorname {o n e - h o t} (\operatorname {a r g m a x} (S)) \in \{0, 1 \} ^ {N \times h \times w}. \tag {2} +$$ + +Consequently, the $S_{\mathrm{hard}}$ in Eq. (2) can also be seen as $N$ binary masks, where the foreground of each mask indicates which pixels belong to that specific region. However, not each one among the $N$ masks contains foreground. In fact, in some simple images, numerous pixels may be matched with a single region embedding, while other region embeddings remain unmatched with any pixel (illustrated as "used" and "unused" in Fig. 3). Therefore, while $N$ is a fixed hyper-parameter, the actual number of regions corresponding to each image is dynamic, determined by the complexity of the image. + +Region Embedder Through the region decoder, the original image is represented as several regions. Each region comprises a region embedding $\mathcal{E}_r[i]$ representing the visual feature and a binary mask $S_{\mathrm{hard}}[i]$ representing the geometric information (e.g., shape, position and scale). The region embedder is further employed to merge the above two components and produce the desired tokens. + +Following GroupViT [53], we first aggregate the feature of each pixel into the corresponding region embedding. To make the entire process differentiable, enabling the gradients of image reconstruction to guide the granularity of regions, we employ the straight-through trick [39] here. This process is formulated as: + +$$ +S = \operatorname {s o f t m a x} (S, \dim = 0), \tag {3} +$$ + +$$ +\tilde {S} = S _ {\text {h a r d}} + S - \operatorname {s g} \cdot (S), \tag {4} +$$ + +$$ +\hat {\mathcal {E}} _ {r} [ i ] = \mathcal {E} _ {r} [ i ] + \sum_ {j, k} \hat {S} [ i, j, k ] \cdot \mathcal {E} _ {i m g} [:, j, k ], \tag {5} +$$ + +in which the $\mathrm{sg}\cdot ()$ refers to stop gradient. While the pixel features aggregation process shown in Eq. (5) is actually a pooling process, leading to over-smoothing + +![](images/64d85e7b2d3ad9641b2068c725a153eea26732b31626e306a51131a4f52d987c.jpg) +Fig. 3: An overview of the proposed Reader. It comprises an encoder and a decoder. The encoder can segment each image into an adaptive number of regions, which is illustrated as the "used" part, and embeds each region into a token with the Region Embedder. The decoder then reconstructs the original image from these tokens with the learnable patch queries. + +on local details. We additionally employ a masked cross-attention layer [6] to enhance the visual feature extraction. This layer uses the region embeddings $\hat{\mathcal{E}}_r$ as queries to perform attention to the image feature $\mathcal{E}_{img}$ only in the foreground of each region in order to extract more local details. + +Subsequently, we incorporate a module to explicitly encode the geometric information in the masks and integrate it into the region embeddings. This approach serves two primary purposes: Firstly, the inclusion of geometric information enhances the delineation of boundaries between regions, thereby facilitating image reconstruction. Secondly, such explicit encoding also enables more editing operations in subsequent steps. Specifically, we flatten each mask and utilize a linear layer to project it into a latent space. Additionally, the region embeddings are also projected with an MLP and are then fused with the former projection through element-wise addition. Following a normalization and another MLP, the region embedder produces the desired region-based image tokens. + +# 3.2 Decoder Architecture + +The encoder of Reader represents each image as an adaptive-length sequence. Consequently, the decoder needs the capacity to decode this adaptive-length sequence into a uniformly-sized image. This poses a challenge of mapping variable-length input to a fixed-length output. To address this, the decoder architecture of Reader is based on the Transformer decoder, leveraging the cross-attention mechanism to resolve the aforementioned problem. + +Specifically, we introduce $M$ learnable embeddings as the input queries for the Transformer decoder. Each embedding is responsible for reconstructing a local patch in the output image and is referred to as a patch query. The queries are also incorporated with the positional embeddings of the center coordinate of the patches. To prevent information loss and enable the Transformer decoder to focus on different information at each layer, we employ a separate MLP for each + +layer to project the image tokens. These projections then serve as the key and value for the corresponding layer. + +Through the Transformer decoder, the patch queries interact with the image tokens through cross-attention, gathering information related to the corresponding patches and updating their values. After the Transformer decoder, we integrate an linear layer to finally map the output patch queries to the pixel space, which is a reverse process of the patch embedder of ViT. + +# 3.3 Objective Function + +The objective functions related to the image reconstruction are kept consistent with VQGAN [16], which is a combination of the L1 loss, perceptual loss [58] and a patch-based GAN loss [28], denoted as $\mathcal{L}_{img}$ . + +However, in our observation, relying solely on the image reconstruction loss is insufficient for training the encoder to produce semantically meaningful regions. In practice, the regions easily collapse into uniformly distributed patches. Therefore, additional supervision on the output regions becomes necessary. We employ either the pseudo-annotations produced by SAM [30] or manual annotations to supervise the output regions in different experiments, with detailed information provided in Sec. 4. + +In order to train the similarity matrix in Eq. (3) to resemble a one-hot form at each pixel, we employ a cross-entropy loss here. Unlike the tasks, where cross-entropy typically used with defined category IDs, there is no fixed correspondence between the predicted regions and the masks of annotations here. Following [5,7], we employ a Hungarian-algorithm based bipartite matching to determine the matching relationship, with more details provided in supplementary material. The overall objective function is formulated as: + +$$ +\mathcal {L} = \mathcal {L} _ {i m g} + \mathcal {L} _ {c e}. \tag {6} +$$ + +# 3.4 Image Editing Implementation + +The proposed Reader inherently facilitates diverse image editing operations. Without any additional modules, this is achieved by simply manipulating the image tokens outputted by the encoder and then inputting these manipulated tokens into the decoder to get the edited image. + +Specifically, before the region embedder of Reader, each region is represented by two components: the geometric information (represented as a binary mask) and the visual information (represented as the region embedding). This enables independent manipulations on both two parts and achieves different editing effects. Therefore, we define the following three-level image editing operations. A schematic diagram is shown in Fig. 4 to provide an intuitive understanding. + +Geometry Level These operations refer to only manipulating the masks through affine transformations or interactive brushing by users, while maintaining the embeddings unaltered. Geometry-level operations allows to perform moving, scaling, and shape modification on objects. As an example, the left branch in Fig. 4 alters the shape of the lips without changing their color. + +![](images/befe1cac5f5f0d4bf3fdf961ce3fae5bcca9d14164161864e437a0a5d07f55e0.jpg) +Fig. 4: Illustration of editing operations. Different operations on the embeddings and masks leads to different editing effects. The left, middle and right branches correspond to the three-levels operations described in Sec. 3.4, respectively. + +Feature Level Feature-level operations maintain the mask unchanged while replacing the corresponding embedding with one from another object or another image. This facilitates local feature transferring without impacting geometric characteristics. An example is illustrated in the middle branch of Fig. 4, where the color of lips is changed by replacing the corresponding two embeddings. + +Token Level The erasure, addition and replacement of objects are categorized as this level. Erasure can be achieved by removing tokens corresponding to a specific object. Conversely, addition can be accomplished by appending tokens from another image to the tokens of the current image. Object replacement can then be achieved by combining both operations. Furthermore, these operations can also be combined with the geometry-level manipulations to achieve addition and replacement on objects while controlling their shapes, scale and position, etc. An example is illustrated in the right branch of Fig. 4, where the mouth and lips are replaced by those from another image. + +# 4 Experiments + +# 4.1 Experiment on ImageNet + +Herein we train Reader on ImageNet training set and report qualitative results in image editing and quantitative comparisons on reconstruction fidelity. + +Implementation As outlined in Sec. 3.3, additional supervision is essential to prevent the collapse of producing regions. Consequently, in this experiment, we employ SAM [30] to produce pseudo-annotations on the ImageNet training set. Specifically, we keep the original settings of the segment anything mode of SAM, which effectively segments all prominent objects or local parts in images. Since there is substantial overlapping in the original results of SAM, while the cross-entropy loss needs non-overlapping annotations, we implement the following two post-processing steps. Firstly, removing all larger masks that encompass other ones, retaining only smaller, more localized masks. Secondly, regions not encompassed by the pseudo annotations do not contribute to the objective function. + +![](images/9e205f9a1b59472ba7e1d89678a225844b2ae1d826cda4eb442548f3cc89b592.jpg) +Fig. 5: Qualitative results in ImageNet validation set and Internet images. The editing objective is labeled with text under image. + +The input and output sizes of images are set to $(256 \times 256)$ in this experiment. Both the ViT in the encoder and the Transformer in the decoder follow the configuration of ViT-tiny, which utilizes 192 channels and 3 heads in the multi-head attention. The dimension of image tokens is set to 256 to maintain consistency with the previous methods. The number $N$ of learnable region queries in the encoder is set to 256, which is the minimal tokens number of each image in previous works and serves as the maximum regions number of our model. The number $M$ of learnable patch queries in the decoder is set to 1024, with each query responsible for reconstructing a $8 \times 8$ patch in the output image. For more implementation details, please refer to our supplementary material. + +Results We provide several image editing examples in Fig. 5 corresponding to all the three-level editing operations described in Sec. 3.4. For example, "sticking out the tongue" needs only manipulation on the tongue's mask, "replacing the color" is achieved by replacing the region embedding of the strawberry and "replacing with another hat" needs deleting the tokens of the original hat and replacing them with new ones from another image. + +For quantitative comparison of the reconstruction fidelity, we employ both the distribution-level metric rFID [24] and the image-patches-level metrics PSNR, SSIM [26], LPIPS [58] to evaluate the reconstructed images of ImageNet validation set. We employ the VQGAN [16], RQ-VAE [32] and DQ-VAE [32] as baselines. The former two employ fixed tokens number of each image, while the latter uses dynamic tokens number through picking one granularity for each patch from three pre-defined strides $\left(\frac{1}{8},\frac{1}{16},\frac{1}{32}\right)$ . The tokens number of DQ-VAE is computed by the ratio of different strides provided in their paper. Tokens number of our Reader corresponds to the average number observed across the entire ImageNet validation set. + +Table 1: Quantitative results for the reconstruction fidelity on the ImageNet validation set. "Tokens" refers to the number of image tokens utilized for each image. The dimension of image tokens is 256 for all methods. The "+" denotes data excerpted from the original paper. Top 1 and top 2 results are highlighted in red and blue, respectively. + +
MethodTokensPSNR ↑SSIM ↑LPIPS ↓rFID ↓Params(M)FLOPs(G)
VQGAN(1024)16 × 1619.360.480.318.0572.14196.33
VQGAN(16384)16 × 1619.830.490.284.9276.07197.33
RQ-VAE(16384)†8 × 8 × 210.77
RQ-VAE(16384)8 × 8 × 420.030.500.285.24104.40193.64
DQ-VAE(16384)†2564.08 - 7.32
Reader(ours)79.0924.750.720.202.4326.6040.24
+ +Results shown in Tab. 1 indicate that, in comparison to the baselines, our Reader attains superior reconstruction fidelity despite having the smallest tokens number, model parameters and computational complexity. While compared to DQ-VAE, which also employs dynamic length encoding, our method also demonstrates significant superiority. This finding is intuitive that, compared to both the fixed- and dynamic-sized patches, arbitrarily shaped regions can better handle the uneven information density in images and avoid information redundancy, thus achieving higher reconstruction fidelity with fewer tokens. + +# 4.2 Experiment on Human Face + +To further evaluate the editing capabilities of our Reader, we transfer it to images of human faces, which are commonly employed in previous image editing works, allowing to provide quantitative comparisons. + +Implementation We fine-tune the Reader in this experiment, which is pretrained as outlined in Sec. 4.1. We utilize the CelebAMask-HQ dataset [31], a face parsing dataset with manual annotations. The input and output sizes of images for this experiment are configured as $(512 \times 512)$ to align with the previous methods [31, 34]. In order to preserving the knowledge already learned by the pre-trained model, we keep the majority of Reader architecture unchanged and make the following modifications. + +(1) Adding a downsampling-and-convolution layer with a residual connection in the region embedder and initializing the new convolution's output to zero. This is employed to make the double-sized masks meet the input length of previous mask projection layer. (2) Increasing the number of learnable patch queries from 1024 to 4096 in order to get double-sized reconstructed images. While loading the pre-trained weights, we perform bicubic interpolation on the weights of the positional embeddings and the patch queries to match the new size. + +Results In Fig. 6, we present the qualitative results corresponding to the three-level editing operations. In the top and middle rows, we performed overlay editing, gradually changing one part of the geometric and visual properties at a + +![](images/0843069c29102297a94cb9bdcfce2ac88a13bf57ab5834d64be5d15020008956.jpg) + +![](images/85c806a5f9d63341ca97c53c51c8940f4e65c13cb4ac7b427a727ebfbd0e349b.jpg) ++ Close Mouth + +![](images/9901f7026a6bd0ed496a2d301877969a2496363d27789a93412244ce7d048c79.jpg) ++ Sharp Jawline + +![](images/564cc1003e0345133e3002ac21294de28a7502234c630d9d7915746ede7adb58.jpg) + +![](images/b2fb8084a0e25d74b0507884c48fb19e6164a3ef0ebfe272cfddf1a72dcc2164.jpg) ++ Enlarge Nose + +![](images/91d5fb0201234a8e9846d18ecfdbbe05dd6993b4cdceec07397f2c9fed2257fb.jpg) ++Wide Jawline + +![](images/7977c98a981c09c9856ecc980d374a632993124a4b2e063c558ca73763ef1627.jpg) + +![](images/0e2567278f8901560a6dc53f6c429b20a27c9f4a57e77780827287b6e3e1d39a.jpg) ++ Eyes Feature + +![](images/2d916f4abdc3cdb915cdbbf64078d7b166603006fe77c01ead5e250e1654db83.jpg) ++ Lips Featue + +![](images/2c1d0ec5a0a5073222d5b857650115b8c2fcb1a8b40b5ccf5522e4b588d7434b.jpg) ++ Eyes Feature + +![](images/1318f74c05214260178b920946013c4e9190701061041328e809781210225f53.jpg) + +![](images/22775f61615154e2fba5987b10510adb5b1ebf9a20f32efa001b497131235eb5.jpg) ++ Lips Featue + +![](images/b887752c7c9b846802e32352e1b3971500911bf57a603d765b4626266817d3dc.jpg) + +![](images/da5a91604be712f05143c8028fb883de9127e87770ed56d7941eae1c66d815ee.jpg) ++Smiling + +![](images/eb9e09e5a26800d3ef79050ee06904dcc3f1380e93b1e752f57910fdc04bc6b5.jpg) + +![](images/645e4dbf9be656c983b39f5d438f841d8086cc72b4de513acaaed89a09870448.jpg) + +![](images/7d89dad10c48dbcfa0c07f65866bb69638af4746c59c75b382aaee427e90bfe2.jpg) + +![](images/d58c5f568a5d00bfa6884e4c9dc1ea43241b564c020ab262652f8eb76a254976.jpg) ++Smiling + +![](images/377ad898242c7c411bb7dc163c84e923225bf5150e9e1d59bee996e05dc4f3f5.jpg) +Fig. 6: Qualitative results in huaman faces. The editing objective is labeled with text under corresponding image. +Original Image +Fig. 7: Qualitative comparisons in the smile transferring benchmark. Our Reader can selectively manipulate only the target attributes without affecting other features or regions. The red boxes highlight the local details that should remain unchanged after editing. Please kindly zoom in for a better view. + +![](images/3fd414cec35e5f5ef03a59f22a8b0ec6fd7ca011b28cbc52dfbc3f2fab03349c.jpg) +MaskGAN + +![](images/9757fecff33a3c356261405a6648fc98a3d69d84f70b30b9863245cd9e790551.jpg) +Ours + +![](images/db08d64c90318e1d6809f5705a42eb384e85cb713e68763988d8a106e6ae1166.jpg) +Original Image + +![](images/a370999f17a3d1eee7d93661f343367582bdd7434fd1c5a22f2729da55f6ade2.jpg) +MaskGAN + +![](images/3e9c9200f8e376f5881b55f4833de76afbfdb55a8dbae8b65785ad6f849a624d.jpg) +Ours + +time. In the bottom row, we simultaneously manipulated both the masks and embeddings, facilitating significant change in expression. + +To quantitatively measure Reader's image editing capabilities, we utilize the smile transferring benchmark introduced by MaskGAN [31], which requires converting the faces with neutral expressions into smiling faces. We compare our Reader with the MaskGAN [31], LocalEditing [9], InterFaceGAN [46] and EditGAN [34], which also provide results in this benchmark. We employ the following metrics to keep consistent with [34]. + +(1) Semantic Correctness: Training a binary classifier to recognize the smile attribute and evaluating whether the edited faces exhibit a smiling expression. (2) Distribution-level Image Quality: Utilizing FID [24] and KID [2] to assess the quality of edited images. (3) Identity Preservation: Leveraging a pretrained ArcFace feature extraction network [12] to evaluate whether the face's identity is maintained post-editing. + +To automate smile transferring without manual adjustments, we employ a similar approach to MaskGAN. Specifically, we select 400 pairs of neutral- + +Table 2: Quantitative results in the smile transferring benchmark. Top 1 and top 2 results are highlighted in red and blue, respectively. Our Reader significantly outperforms previous methods in the FID and ID Score metrics by a large margin. + +
MetricAttribute Acc. (%)↑FID↓KID↓ID Score↑
MaskGAN [31]77.346.840.0200.4611
LocalEditing [9]26.041.260.0120.5823
LocalEditing - Encoding4Editing [49]41.848.280.0160.7295
InterFaceGAN [46]83.539.420.0100.7295
EditGAN [34]91.541.740.0130.7047
EditGAN+30 [34]85.840.830.0120.7452
Reader (ours)84.237.910.0110.8470
+ +expression and smiling-expression faces with the closest head position as the source and target images. Subsequently, we replace both the tokens corresponding to mouth and lips of the source image with those from the target image. + +Results shown in Tab. 2 demonstrate that our Reader achieves excellent performance across all metrics. Specifically, our Reader significantly outperforms previous methods in the FID and ID Score metrics by a large margin. This indicates that the edited images produced by Reader exhibit higher quality while better preserving the ID features of the original images. This observation effectively highlights the inherent advantage of our method. By representing the image through a series of decoupled tokens, Reader can selectively manipulate only the target attributes without affecting other regions. To provide a more intuitive perception, we illustrate some examples of the original and edited image pairs in Fig. 7. It can be observed that our method effectively prevents changes in the areas which are not the editing target, while achieving editing goals. In contrast, the GAN-based baseline [31] is prone to inducing global changes. + +# 4.3 Ablation Study + +In Tab. 3a, we conducted ablation studies to evaluate the effectiveness of some optional hyper-parameters and modules in the model architecture, where different configurations represent: Patch Size: The patch size of the ViT in Reader's encoder. Bidirectional: Whether to use bidirectional transformer block in the region decoder. Masked-Attn: Whether to employ the masked cross-attention layer in the region embedder. Results indicate that a ViT with a smaller patch size paired with the additional modules produces the best image fidelity. This demonstrates that our default settings can better preserve the fine-grained details in the image, thereby enhancing the quality of the output image. + +In Tab. 3b, we conducted ablation experiment to determine the granularity of pseudo-labels utilized in training. In this experiment, during preprocessing, we only removed the overlapping parts from the pseudo-labels and retained masks of all different sizes to train our Reader. Subsequently, we calculated the average + +Table 3: Ablation studies on both model architecture, hyper parameter and data preprocessing. Our default setting is marked with gray. + +
Patch SizeBidirectionalMasked-AttnrFID ↓
8 × 8XX98.39
✓X37.50
✓✓11.42
16 × 16✓✓61.64
+ +(a) Module configurations. + +
Area ProportionL1 ↓
< 1/321.44
1/32 ~ 1/161.56
1/16 ~ 1/82.53
> 1/83.46
+ +(b) Annotation configurations. + +L1 distance of different-sized regions between the original images and the reconstructed images. The "Area Proportion" in Tab. 3b represents the proportion of the region's area relative to the entire image. The results indicate that when the region's area proportion is less than $\frac{1}{16}$ , the L1 distance stabilizes with changes in area. Therefore, in the formal experiments, we only retained pseudo-labels with area proportion smaller than $\frac{1}{16}$ for training. + +# 5 Conclusion and Discussion + +In conclusion, this paper explores an innovative region-based visual token representation and presents Reader. Reader employs arbitrarily shaped and sized regions as the units of image tokens, which aligns better with the visual prior: the non-uniformity of information density and the naturally existing semantic boundaries. Through experiments, we demonstrate the following inherent advantages of this region-based tokenization method. Efficiency: This method has a high compression ratio and can achieve superior reconstruction fidelity while employing a significantly smaller number of tokens for each image. Disentanglement: The region-based tokens guarantees their disentanglement from one another, and further empower Reader with the ability to perform image editing. + +Limitation and Feature Work As this paper focuses on exploring the feasibility and effectiveness of region-based tokenization, we adopted a minimal configuration in our experiments, manifested as employing the tiny-version transformer and training model on relatively small-sized images $(256\times 256)$ . Although Reader can produce already satisfactory results under such configurations, the parameter size and data resolution still limit its broader application. For example, some high-resolution images may exhibit a degree of blurring after editing. + +In the future work, we will pursue scaling up both model parameters and the volume of data. Additionally, we aim to explore its application in various scenarios, such as interaction with LLMs and multimodal perception, generation. + +Broader Impact Reader possesses the capability to serve as an image editor, yielding numerous interesting outcomes, but also opens up the possibility of real photo manipulation for nefarious purposes. However, with the recent advancements in watermarking techniques and manipulated images identifying techniques [17], this risk is avoidable. + +# Acknowledgements + +This research was funded by the Fundamental Research Funds for the Central Universities (2024XKRC082) and the National NSF of China (No.U23A20314). + +# References + +1. Anderson, P., He, X., Buehler, C., Teney, D., Johnson, M., Gould, S., Zhang, L.: Bottom-up and top-down attention for image captioning and visual question answering. In: CVPR. pp. 6077-6086 (2018) +2. Binkowski, M., Sutherland, D.J., Arbel, M., Gretton, A.: Demystifying MMD gans. In: ICLR. OpenReview.net (2018) +3. Brooks, T., Holynski, A., Efros, A.A.: Instructpix2pix: Learning to follow image editing instructions. In: CVPR. pp. 18392-18402. IEEE (2023) +4. Burgess, C.P., Matthew, L., Watters, N., Kabra, R., Higgins, I., Botvinick, M.M., Lerchner, A.: Monet: Unsupervised scene decomposition and representation. CoRR abs/1901.11390 (2019) +5. Carion, N., Massa, F., Synnaeve, G., Usunier, N., Kirillov, A., Zagoruyko, S.: End-to-end object detection with transformers. In: Vedaldi, A., Bischof, H., Brox, T., Frahm, J. (eds.) ECCV. Lecture Notes in Computer Science, vol. 12346, pp. 213-229. Springer (2020) +6. Cheng, B., Misra, I., Schwing, A.G., Kirillov, A., Girdhar, R.: Masked-attention mask transformer for universal image segmentation. In: CVPR. pp. 1290–1299 (2022) +7. Cheng, B., Schwing, A.G., Kirillov, A.: Per-pixel classification is not all you need for semantic segmentation. In: Ranzato, M., Beygelzimer, A., Dauphin, Y.N., Liang, P., Vaughan, J.W. (eds.) NeurIPS. pp. 17864-17875 (2021) +8. Choi, Y., Choi, M., Kim, M., Ha, J., Kim, S., Choo, J.: Stargan: Unified generative adversarial networks for multi-domain image-to-image translation. In: CVPR. pp. 8789-8797. Computer Vision Foundation / IEEE Computer Society (2018) +9. Collins, E., Bala, R., Price, B., Süssstrunk, S.: Editing in style: Uncovering the local semantics of gans. In: CVPR. pp. 5770-5779. Computer Vision Foundation / IEEE (2020) +0. Couairon, G., Verbeek, J., Schwenk, H., Cord, M.: Diffedit: Diffusion-based semantic image editing with mask guidance. In: ICLR. OpenReview.net (2023) +1. Crowson, K., Biderman, S., Kornis, D., Stander, D., Hallahan, E., Castricato, L., Raff, E.: Vqgan-clip: Open domain image generation and editing with natural language guidance. In: ECCV. pp. 88–105. Springer (2022) +2. Deng, J., Guo, J., Xue, N., Zafeiriou, S.: Arcface: Additive angular margin loss for deep face recognition. In: CVPR. pp. 4690-4699. Computer Vision Foundation / IEEE (2019) +3. Devlin, J., Chang, M., Lee, K., Toutanova, K.: BERT: pre-training of deep bidirectional transformers for language understanding. In: Burstein, J., Doran, C., Solorio, T. (eds.) NAACL-HLT. pp. 4171-4186. Association for Computational Linguistics (2019) +4. Ding, H., Liu, C., He, S., Jiang, X., Loy, C.C.: MeViS: A large-scale benchmark for video segmentation with motion expressions. In: ICCV (2023) + +15. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., Uszkoreit, J., Houlsby, N.: An image is worth 16x16 words: Transformers for image recognition at scale. In: ICLR. OpenReview.net (2021) +16. Esser, P., Rombach, R., Ommer, B.: Taming transformers for high-resolution image synthesis. In: CVPR. pp. 12873-12883. Computer Vision Foundation / IEEE (2021) +17. Fernandez, P., Couairon, G., Jégou, H., Douze, M., Furon, T.: The stable signature: Rooting watermarks in latent diffusion models. CoRR abs/2303.15435 (2023) +18. Gafni, O., Polyak, A., Ashual, O., Sheynin, S., Parikh, D., Taigman, Y.: Make-a-scene: Scene-based text-to-image generation with human priors. In: ECCV. pp. 89-106. Springer (2022) +19. Goel, V., Peruzzo, E., Jiang, Y., Xu, D., Sebe, N., Darrell, T., Wang, Z., Shi, H.: Pair-diffusion: Object-level image editing with structure-and-appearance paired diffusion models. CoRR abs/2303.17546 (2023) +20. Gondal, M.W., Wuthrich, M., Miladinovic, D., Locatello, F., Breidt, M., Volchkov, V., Akpo, J., Bachem, O., Scholkopf, B., Bauer, S.: On the transfer of inductive bias from simulation to the real world: a new disentanglement dataset. In: Wallach, H.M., Larochelle, H., Beygelzimer, A., d'Alché-Buc, F., Fox, E.B., Garnett, R. (eds.) NeurIPS. pp. 15714-15725 (2019) +21. Goodfellow, I.J., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., Courville, A.C., Bengio, Y.: Generative adversarial nets. In: Ghahramani, Z., Welling, M., Cortes, C., Lawrence, N.D., Weinberger, K.Q. (eds.) NeurIPS (2014) +22. Härkönen, E., Hertzmann, A., Lehtinen, J., Paris, S.: Ganspace: Discovering interpretable GAN controls. In: Larochelle, H., Ranzato, M., Hadsell, R., Balcan, M., Lin, H. (eds.) NeurIPS (2020) +23. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: CVPR. pp. 770-778. IEEE Computer Society (2016) +24. Heusel, M., Ramsauer, H., Unterthiner, T., Nessler, B., Hochreiter, S.: Gans trained by a two time-scale update rule converge to a local nash equilibrium. In: Guyon, I., von Luxburg, U., Bengio, S., Wallach, H.M., Fergus, R., Vishwanathan, S.V.N., Garnett, R. (eds.) NeurIPS. pp. 6626-6637 (2017) +25. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. In: Larochelle, H., Ranzato, M., Hadsell, R., Balcan, M., Lin, H. (eds.) NeurIPS (2020) +26. Horé, A., Ziou, D.: Image quality metrics: PSNR vs. SSIM. In: ICPR. pp. 2366-2369. IEEE Computer Society (2010) +27. Huang, M., Mao, Z., Chen, Z., Zhang, Y.: Towards accurate image coding: Improved autoregressive image generation with dynamic vector quantization. In: CVPR. pp. 22596-22605 (2023) +28. Isola, P., Zhu, J., Zhou, T., Efros, A.A.: Image-to-image translation with conditional adversarial networks. In: CVPR. pp. 5967-5976. IEEE Computer Society (2017) +29. Kim, H., Mnih, A.: Disentangling by factorising. In: Dy, J.G., Krause, A. (eds.) ICML. Proceedings of Machine Learning Research, vol. 80, pp. 2654-2663. PMLR (2018) +30. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W., Dollár, P., Girshick, R.B.: Segment anything. CoRR abs/2304.02643 (2023) +31. Lee, C., Liu, Z., Wu, L., Luo, P.: Maskgan: Towards diverse and interactive facial image manipulation. In: CVPR. pp. 5548-5557. Computer Vision Foundation / IEEE (2020) + +32. Lee, D., Kim, C., Kim, S., Cho, M., Han, W.: Autoregressive image generation using residual quantization. In: CVPR. pp. 11513-11522. IEEE (2022). https://doi.org/10.1109/CVPR52688.2022.01123, https://doi.org/10.1109/CVPR52688.2022.01123 +33. Li, Y., Mao, H., Girshick, R., He, K.: Exploring plain vision transformer backbones for object detection. In: ECCV. pp. 280-296. Springer (2022) +34. Ling, H., Kreis, K., Li, D., Kim, S.W., Torralba, A., Fidler, S.: Editgan: High-precision semantic image editing. In: Ranzato, M., Beygelzimer, A., Dauphin, Y.N., Liang, P., Vaughan, J.W. (eds.) NeurIPS. pp. 16331-16345 (2021) +35. Liu, C., Ding, H., Jiang, X.: GRES: Generalized referring expression segmentation. In: CVPR (2023) +36. Liu, C., Jiang, X., Ding, H.: Primitivenet: decomposing the global constraints for referring segmentation. Visual Intelligence 2(1), 16 (2024) +37. Locatello, F., Weissenborn, D., Unterthiner, T., Mahendran, A., Heigold, G., Uszkoreit, J., Dosovitskiy, A., Kipf, T.: Object-centric learning with slot attention. In: Larochelle, H., Ranzato, M., Hadsell, R., Balcan, M., Lin, H. (eds.) NeurIPS (2020) +38. Ning, J., Li, C., Zhang, Z., Wang, C., Geng, Z., Dai, Q., He, K., Hu, H.: All in tokens: Unifying output space of visual tasks via soft token. In: ICCV. pp. 19900-19910 (2023) +39. van den Oord, A., Vinyals, O., Kavukcuoglu, K.: Neural discrete representation learning. In: Guyon, I., von Luxburg, U., Bengio, S., Wallach, H.M., Fergus, R., Vishwanathan, S.V.N., Garnett, R. (eds.) NeurIPS. pp. 6306-6315 (2017) +40. Pan, X., Tewari, A., Leimkuhler, T., Liu, L., Meka, A., Theobalt, C.: Drag your GAN: interactive point-based manipulation on the generative image manifold. In: Brunvand, E., Sheffer, A., Wimmer, M. (eds.) SIGGRAPH. pp. 78:1-78:11. ACM (2023) +41. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: ICML. pp. 8748-8763. PMLR (2021) +42. Radford, A., Narasimhan, K., Salimans, T., Sutskever, I., et al.: Improving language understanding by generative pre-training (2018) +43. Reed, S.E., Zhang, Y., Zhang, Y., Lee, H.: Deep visual analogy-making. In: Cortes, C., Lawrence, N.D., Lee, D.D., Sugiyama, M., Garnett, R. (eds.) NeurIPS. pp. 1252-1260 (2015) +44. Ren, S., He, K., Girshick, R., Sun, J.: Faster r-cnn: Towards real-time object detection with region proposal networks. NeurIPS 28 (2015) +45. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: CVPR. pp. 10674-10685. IEEE (2022) +46. Shen, Y., Yang, C., Tang, X., Zhou, B.: Interfacegan: Interpreting the disentangled face representation learned by gans. IEEE TPAMI 44(4), 2004-2018 (2022) +47. Shuai, X., Ding, H., Ma, X., Tu, R., Jiang, Y.G., Tao, D.: A survey of multimodal-guided image editing with text-to-image diffusion models. arXiv preprint arXiv:2406.14555 (2024) +48. Sohl-Dickstein, J., Weiss, E.A., Maheswaranathan, N., Ganguli, S.: Deep unsupervised learning using nonequilibrium thermodynamics. In: Bach, F.R., Blei, D.M. (eds.) ICML. JMLR Workshop and Conference Proceedings, vol. 37, pp. 2256-2265. JMLR.org (2015) +49. Tov, O., Alaluf, Y., Nitzan, Y., Patashnik, O., Cohen-Or, D.: Designing an encoder for stylegan image manipulation. ACM TOG 40(4), 133:1-133:14 (2021) + +50. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. In: Guyon, I., von Luxburg, U., Bengio, S., Wallach, H.M., Fergus, R., Vishwanathan, S.V.N., Garnett, R. (eds.) NeurIPS. pp. 5998-6008 (2017) +51. Wu, B., Xu, C., Dai, X., Wan, A., Zhang, P., Yan, Z., Tomizuka, M., Gonzalez, J., Keutzer, K., Vajda, P.: Visual transformers: Token-based image representation and processing for computer vision. arXiv preprint arXiv:2006.03677 (2020) +52. Wu, R., Zhang, G., Lu, S., Chen, T.: Cascade EF-GAN: progressive facial expression editing with local focuses. In: CVPR. pp. 5020-5029. Computer Vision Foundation / IEEE (2020) +53. Xu, J., Mello, S.D., Liu, S., Byeon, W., Breuel, T.M., Kautz, J., Wang, X.: Groupvit: Semantic segmentation emerges from text supervision. In: CVPR. pp. 18113-18123. IEEE (2022) +54. Yang, B., Gu, S., Zhang, B., Zhang, T., Chen, X., Sun, X., Chen, D., Wen, F.: Paint by example: Exemplar-based image editing with diffusion models. In: CVPR. pp. 18381-18391. IEEE (2023). https://doi.org/10.1109/CVPR52729.2023.01763, https://doi.org/10.1109/CVPR52729.2023.01763 +55. Yang, T., Wang, Y., Lu, Y., Zheng, N.: Visual concepts tokenization. In: NeurIPS (2022) +56. Yu, J., Li, X., Koh, J.Y., Zhang, H., Pang, R., Qin, J., Ku, A., Xu, Y., Baldridge, J., Wu, Y.: Vector-quantized image modeling with improved VQGAN. In: ICLR. OpenReview.net (2022) +57. Zellers, R., Yatskar, M., Thomson, S., Choi, Y.: Neural motifs: Scene graph parsing with global context. In: CVPR. pp. 5831-5840 (2018) +58. Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual metric. In: CVPR. pp. 586-595. Computer Vision Foundation / IEEE Computer Society (2018) +59. Zhong, Y., Wang, L., Chen, J., Yu, D., Li, Y.: Comprehensive image captioning via scene graph decomposition. In: ECCV. pp. 211-229. Springer (2020) +60. Zhu, J., Yang, C., Shen, Y., Shi, Z., Zhao, D., Chen, Q.: Linkgan: Linking GAN latents to pixels for controllable image synthesis. CoRR abs/2301.04604 (2023) \ No newline at end of file diff --git a/2024/Region-Native Visual Tokenization/images.zip b/2024/Region-Native Visual Tokenization/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..e71e6580caf943a49e589b4e2db4194e5ad1b928 --- /dev/null +++ b/2024/Region-Native Visual Tokenization/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:30ed873363e910c9e3e591ff0dd1b3c6239935a870fa5cf235c75eb3b964f288 +size 557909 diff --git a/2024/Region-Native Visual Tokenization/layout.json b/2024/Region-Native Visual Tokenization/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..56a70a5b350d0f08c2e3af119cf45f1d050aafe7 --- /dev/null +++ b/2024/Region-Native Visual Tokenization/layout.json @@ -0,0 +1,10662 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 184, + 112, + 430, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 184, + 112, + 430, + 129 + ], + "spans": [ + { + "bbox": [ + 184, + 112, + 430, + 129 + ], + "type": "text", + "content": "Region-Native Visual Tokenization" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 170, + 149, + 444, + 187 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 170, + 149, + 444, + 187 + ], + "spans": [ + { + "bbox": [ + 170, + 149, + 444, + 187 + ], + "type": "text", + "content": "Mengyu Wang" + }, + { + "bbox": [ + 170, + 149, + 444, + 187 + ], + "type": "inline_equation", + "content": "^{1,5(\\dagger)}" + }, + { + "bbox": [ + 170, + 149, + 444, + 187 + ], + "type": "inline_equation", + "content": "\\odot" + }, + { + "bbox": [ + 170, + 149, + 444, + 187 + ], + "type": "text", + "content": ", Yuyao Huang" + }, + { + "bbox": [ + 170, + 149, + 444, + 187 + ], + "type": "inline_equation", + "content": "^{1,5(\\dagger,\\ddagger)}" + }, + { + "bbox": [ + 170, + 149, + 444, + 187 + ], + "type": "inline_equation", + "content": "\\odot" + }, + { + "bbox": [ + 170, + 149, + 444, + 187 + ], + "type": "text", + "content": ", Henghui Ding" + }, + { + "bbox": [ + 170, + 149, + 444, + 187 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 170, + 149, + 444, + 187 + ], + "type": "inline_equation", + "content": "\\odot" + }, + { + "bbox": [ + 170, + 149, + 444, + 187 + ], + "type": "text", + "content": ", Xinlong Wang" + }, + { + "bbox": [ + 170, + 149, + 444, + 187 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 170, + 149, + 444, + 187 + ], + "type": "inline_equation", + "content": "\\odot" + }, + { + "bbox": [ + 170, + 149, + 444, + 187 + ], + "type": "text", + "content": ", Tiejun Huang" + }, + { + "bbox": [ + 170, + 149, + 444, + 187 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 170, + 149, + 444, + 187 + ], + "type": "text", + "content": ", Yao Zhao" + }, + { + "bbox": [ + 170, + 149, + 444, + 187 + ], + "type": "inline_equation", + "content": "^{1,5,6}" + }, + { + "bbox": [ + 170, + 149, + 444, + 187 + ], + "type": "inline_equation", + "content": "\\odot" + }, + { + "bbox": [ + 170, + 149, + 444, + 187 + ], + "type": "text", + "content": ", Yunchao Wei" + }, + { + "bbox": [ + 170, + 149, + 444, + 187 + ], + "type": "inline_equation", + "content": "^{1,5,6(\\boxtimes)}" + }, + { + "bbox": [ + 170, + 149, + 444, + 187 + ], + "type": "inline_equation", + "content": "\\odot" + }, + { + "bbox": [ + 170, + 149, + 444, + 187 + ], + "type": "text", + "content": ", and Shuicheng Yan" + }, + { + "bbox": [ + 170, + 149, + 444, + 187 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 170, + 149, + 444, + 187 + ], + "type": "inline_equation", + "content": "\\odot" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 147, + 194, + 465, + 229 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 147, + 194, + 465, + 229 + ], + "spans": [ + { + "bbox": [ + 147, + 194, + 465, + 229 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 147, + 194, + 465, + 229 + ], + "type": "text", + "content": " Institute of Information Science, Beijing Jiaotong University, Beijing, China \n" + }, + { + "bbox": [ + 147, + 194, + 465, + 229 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 147, + 194, + 465, + 229 + ], + "type": "text", + "content": " Beijing Academy of Artificial Intelligence \n" + }, + { + "bbox": [ + 147, + 194, + 465, + 229 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 147, + 194, + 465, + 229 + ], + "type": "text", + "content": " Skywork AI" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 229, + 481, + 262 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 229, + 481, + 262 + ], + "spans": [ + { + "bbox": [ + 132, + 229, + 481, + 262 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 132, + 229, + 481, + 262 + ], + "type": "text", + "content": " Institute of Big Data, Fudan University, Shanghai, China \n" + }, + { + "bbox": [ + 132, + 229, + 481, + 262 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 132, + 229, + 481, + 262 + ], + "type": "text", + "content": " Visual Intelligence + X International Joint Laboratory of the Ministry of Education \n" + }, + { + "bbox": [ + 132, + 229, + 481, + 262 + ], + "type": "inline_equation", + "content": "^{6}" + }, + { + "bbox": [ + 132, + 229, + 481, + 262 + ], + "type": "text", + "content": " PengCheng Laboratory, Shenzhen, China" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 159, + 285, + 455, + 526 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 285, + 455, + 526 + ], + "spans": [ + { + "bbox": [ + 159, + 285, + 455, + 526 + ], + "type": "text", + "content": "Abstract. We explore an innovative region-based visual token representation and present the REgion-native AutoencoDER (Reader). In contrast to the majority of previous methods, which represent each image as a grid-shaped tokens map, Reader perceives each image into sequential region-based tokens, with each token corresponding to an object or one part of an object in the image. Specifically, Reader comprises both an encoder and a decoder. The encoder can partition each image into an adaptive number of arbitrary-shaped regions and encode each region into a token. Subsequently, the decoder utilizes this adaptive-length token sequence to reconstruct the original image. Experimental results demonstrate that such region-based token representation possesses two main notable characteristics. Firstly, it achieves highly efficient image encoding. Reader can adaptively use more regions to represent complex areas and fewer regions in simpler ones, thus avoiding information redundancy. Consequently, it achieves superior reconstruction fidelity compared to previous methods, despite using significantly fewer tokens for each image. Secondly, the region-based manner enables manipulation on a local region without causing global changes. As a result, Reader inherently supports diverse image editing operations, including erasing, adding, replacing, and modifying shapes on the objects, and achieves excellent performance in the image editing benchmark of smile transferring. Code is provided at https://github.com/MengyuWang826/Reader.git." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 544, + 230, + 556 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 544, + 230, + 556 + ], + "spans": [ + { + "bbox": [ + 132, + 544, + 230, + 556 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 566, + 482, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 566, + 482, + 628 + ], + "spans": [ + { + "bbox": [ + 130, + 566, + 482, + 628 + ], + "type": "text", + "content": "Tokenization, originating from Natural Language Processing (NLP), involves breaking a textual data stream into several elements called tokens. It plays an important role for converting raw data into a form that is predictable and analyzable by deep learning systems [13, 42]. With the gradual maturity of tokenization technique in the NLP field, it has also been adapted to the Computer" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 141, + 632, + 237, + 643 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 632, + 237, + 643 + ], + "spans": [ + { + "bbox": [ + 141, + 632, + 237, + 643 + ], + "type": "text", + "content": "† Internship at BAAI." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 141, + 643, + 260, + 654 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 643, + 260, + 654 + ], + "spans": [ + { + "bbox": [ + 141, + 643, + 260, + 654 + ], + "type": "inline_equation", + "content": "\\ddagger" + }, + { + "bbox": [ + 141, + 643, + 260, + 654 + ], + "type": "text", + "content": " Internship at Skywork AI." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 141, + 654, + 249, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 654, + 249, + 665 + ], + "spans": [ + { + "bbox": [ + 141, + 654, + 249, + 665 + ], + "type": "text", + "content": "Corresponding author." + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 140, + 114, + 218, + 210 + ], + "blocks": [ + { + "bbox": [ + 140, + 114, + 218, + 210 + ], + "lines": [ + { + "bbox": [ + 140, + 114, + 218, + 210 + ], + "spans": [ + { + "bbox": [ + 140, + 114, + 218, + 210 + ], + "type": "image", + "image_path": "d4c748bfbeb5fe9442a095a84505b3facf700546a43524df61b7e95f775eab59.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 161, + 212, + 197, + 221 + ], + "lines": [ + { + "bbox": [ + 161, + 212, + 197, + 221 + ], + "spans": [ + { + "bbox": [ + 161, + 212, + 197, + 221 + ], + "type": "text", + "content": "(a) Image" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 225, + 114, + 303, + 210 + ], + "blocks": [ + { + "bbox": [ + 225, + 114, + 303, + 210 + ], + "lines": [ + { + "bbox": [ + 225, + 114, + 303, + 210 + ], + "spans": [ + { + "bbox": [ + 225, + 114, + 303, + 210 + ], + "type": "image", + "image_path": "38b04be8ce5feeda681d079903de032e8014071e4115c059ba7054a19bb5c483.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 236, + 212, + 290, + 220 + ], + "lines": [ + { + "bbox": [ + 236, + 212, + 290, + 220 + ], + "spans": [ + { + "bbox": [ + 236, + 212, + 290, + 220 + ], + "type": "text", + "content": "(b) Grid-based" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 310, + 114, + 388, + 210 + ], + "blocks": [ + { + "bbox": [ + 310, + 114, + 388, + 210 + ], + "lines": [ + { + "bbox": [ + 310, + 114, + 388, + 210 + ], + "spans": [ + { + "bbox": [ + 310, + 114, + 388, + 210 + ], + "type": "image", + "image_path": "922722d1bdf5da29da6d2e8202bbd3376d8e91679c64782064ef824e68817403.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 320, + 212, + 380, + 221 + ], + "lines": [ + { + "bbox": [ + 320, + 212, + 380, + 221 + ], + "spans": [ + { + "bbox": [ + 320, + 212, + 380, + 221 + ], + "type": "text", + "content": "(c) Object-based" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 231, + 480, + 286 + ], + "lines": [ + { + "bbox": [ + 130, + 231, + 480, + 286 + ], + "spans": [ + { + "bbox": [ + 130, + 231, + 480, + 286 + ], + "type": "text", + "content": "Fig.1: Illustration of different types of visual tokenization methods. Grid-based methods [16,56] represent images as grid-shaped tokens map. Object-based methods [1,59] use detection models to represent objects as tokens. While our region-based method represents arbitrarily shaped regions as tokens, which offers more flexibility and retains the naturally existed boundaries presented in images." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 395, + 114, + 473, + 210 + ], + "blocks": [ + { + "bbox": [ + 395, + 114, + 473, + 210 + ], + "lines": [ + { + "bbox": [ + 395, + 114, + 473, + 210 + ], + "spans": [ + { + "bbox": [ + 395, + 114, + 473, + 210 + ], + "type": "image", + "image_path": "1250be1d8897bed585b3d0e2a6ac2e6db7f4cb06a5cdfe6367efc60dbccc1b4d.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 404, + 212, + 466, + 221 + ], + "lines": [ + { + "bbox": [ + 404, + 212, + 466, + 221 + ], + "spans": [ + { + "bbox": [ + 404, + 212, + 466, + 221 + ], + "type": "text", + "content": "(d) Region-based" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 300, + 480, + 337 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 300, + 480, + 337 + ], + "spans": [ + { + "bbox": [ + 130, + 300, + 480, + 337 + ], + "type": "text", + "content": "Vision (CV) domain, known as visual tokenization, and has become the foundation for many tasks such as image/video generation [16, 27, 56] and multimodal generation/perception [11, 14, 18, 38]." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 339, + 482, + 591 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 339, + 482, + 591 + ], + "spans": [ + { + "bbox": [ + 130, + 339, + 482, + 591 + ], + "type": "text", + "content": "Existing visual tokenization methods can be categorized into two main types, which we refer to as grid-based and object-based. Grid-based methods typically employ a CNN [23] or ViT [15] backbone to encode the original image into a downsampled grid-shaped feature map, followed by quantization of the feature map to get the image tokens. However, in such methods, each token corresponds to a patch in the original image (see Fig. 1b), which does not adhere to the following characteristics of the image prior. Firstly, different areas in images have varying information densities. While in these methods, patches are treated equally, leading to information loss in densely informative patches and redundancy in sparsely informative ones. Secondly, areas with different semantics within an image have natural boundaries. While employing patches as the units of image tokens disregards the natural boundaries, causing local information within the original image to be distributed and coupled across multiple tokens, limiting the extraction and utilization of local features. Object-based methods [1,35,51,53,59], as an alternative and more natural approach, typically appear in vision-language pretraining. These methods employ detection or segmentation models (see Fig. 1c) to encode the objects within images into tokens, which are then interacted with or aligned with language information. However, these methods only focus on extracting high-level semantic information, neglecting low-level visual details in the image. Consequently, they cannot be used for tasks that require image outputs, such as image synthesis and editing." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": "These challenges suggest that, an ideal visual token representation should take into account the crucial visual-prior characteristics, including the uneven information density and the inherent semantic boundaries, and further achieve local-information decoupling and global-information compression. However, previous practices either ignore the visual-prior characteristics or, due to their coarse granularity, result in significant information loss. To address these problems, we" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 214, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 214, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 214, + 102 + ], + "type": "text", + "content": "Wang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 114, + 192, + 158 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 192, + 158 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 192, + 158 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 192, + 158 + ], + "type": "image", + "image_path": "5270a8fc731de1bbd8f72d290eee3e5e0d29ec9c7742f92e389c4989eaaf4407.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 164, + 158, + 216, + 163 + ], + "lines": [ + { + "bbox": [ + 164, + 158, + 216, + 163 + ], + "spans": [ + { + "bbox": [ + 164, + 158, + 216, + 163 + ], + "type": "text", + "content": "Erasing the two people" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 192, + 114, + 249, + 158 + ], + "blocks": [ + { + "bbox": [ + 192, + 114, + 249, + 158 + ], + "lines": [ + { + "bbox": [ + 192, + 114, + 249, + 158 + ], + "spans": [ + { + "bbox": [ + 192, + 114, + 249, + 158 + ], + "type": "image", + "image_path": "b20413719837a8e504e08eb997d723e574f4c1f86223e040aca868cbe77903b7.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 249, + 114, + 307, + 158 + ], + "blocks": [ + { + "bbox": [ + 249, + 114, + 307, + 158 + ], + "lines": [ + { + "bbox": [ + 249, + 114, + 307, + 158 + ], + "spans": [ + { + "bbox": [ + 249, + 114, + 307, + 158 + ], + "type": "image", + "image_path": "411afb2488cddc08e64f14916b58abcad901e89f1cca17e5d72a9c294bc4162d.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 307, + 114, + 364, + 158 + ], + "blocks": [ + { + "bbox": [ + 307, + 114, + 364, + 158 + ], + "lines": [ + { + "bbox": [ + 307, + 114, + 364, + 158 + ], + "spans": [ + { + "bbox": [ + 307, + 114, + 364, + 158 + ], + "type": "image", + "image_path": "2257e285b40fe8cc01d06188e4793d0df4fb9de90df4cb121966ac65c0f4ef56.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 365, + 114, + 421, + 158 + ], + "blocks": [ + { + "bbox": [ + 365, + 114, + 421, + 158 + ], + "lines": [ + { + "bbox": [ + 365, + 114, + 421, + 158 + ], + "spans": [ + { + "bbox": [ + 365, + 114, + 421, + 158 + ], + "type": "image", + "image_path": "e7620b9cc2feef062ab16ea4037ece79f3a495db031d6a6098830c823e54717b.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 422, + 114, + 480, + 158 + ], + "blocks": [ + { + "bbox": [ + 422, + 114, + 480, + 158 + ], + "lines": [ + { + "bbox": [ + 422, + 114, + 480, + 158 + ], + "spans": [ + { + "bbox": [ + 422, + 114, + 480, + 158 + ], + "type": "image", + "image_path": "295ce50e3e49d94eb6a0f37f6950586437262c29321bb0b53e1670090571b9c6.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 423, + 158, + 447, + 163 + ], + "lines": [ + { + "bbox": [ + 423, + 158, + 447, + 163 + ], + "spans": [ + { + "bbox": [ + 423, + 158, + 447, + 163 + ], + "type": "text", + "content": "mushroom" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 133, + 165, + 192, + 222 + ], + "blocks": [ + { + "bbox": [ + 133, + 165, + 192, + 222 + ], + "lines": [ + { + "bbox": [ + 133, + 165, + 192, + 222 + ], + "spans": [ + { + "bbox": [ + 133, + 165, + 192, + 222 + ], + "type": "image", + "image_path": "e5874e47f54b65f77e6652ee29469d17e8d8717534f2fc772e9776b5be12935f.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 142, + 222, + 239, + 228 + ], + "lines": [ + { + "bbox": [ + 142, + 222, + 239, + 228 + ], + "spans": [ + { + "bbox": [ + 142, + 222, + 239, + 228 + ], + "type": "text", + "content": "Slimming the jawline and enlarging the eyes" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 233, + 481, + 288 + ], + "lines": [ + { + "bbox": [ + 130, + 233, + 481, + 288 + ], + "spans": [ + { + "bbox": [ + 130, + 233, + 481, + 288 + ], + "type": "text", + "content": "Fig. 2: Examples of the proposed Reader performing various editing operations. Reader showcases its strengths in flexibly editing multiple attributes of existing real images under several interactive operations (see Sec. 3.4). In each pair, the left displays the original image, the right displays the edited result and the text describes the corresponding editing objective." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 192, + 165, + 249, + 222 + ], + "blocks": [ + { + "bbox": [ + 192, + 165, + 249, + 222 + ], + "lines": [ + { + "bbox": [ + 192, + 165, + 249, + 222 + ], + "spans": [ + { + "bbox": [ + 192, + 165, + 249, + 222 + ], + "type": "image", + "image_path": "23e9da7a1aa570467727496c9feb26855ddb55174e60ecd306ff5bd41c203ef9.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 250, + 165, + 307, + 222 + ], + "blocks": [ + { + "bbox": [ + 250, + 165, + 307, + 222 + ], + "lines": [ + { + "bbox": [ + 250, + 165, + 307, + 222 + ], + "spans": [ + { + "bbox": [ + 250, + 165, + 307, + 222 + ], + "type": "image", + "image_path": "af48cf96e9542c7e1854a45f9fa75787598b9a26d4c3b26b0c77d4d2b1c55e7a.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 252, + 222, + 358, + 228 + ], + "lines": [ + { + "bbox": [ + 252, + 222, + 358, + 228 + ], + "spans": [ + { + "bbox": [ + 252, + 222, + 358, + 228 + ], + "type": "text", + "content": "Non-smiling expressions to smiling expressions" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 307, + 165, + 364, + 222 + ], + "blocks": [ + { + "bbox": [ + 307, + 165, + 364, + 222 + ], + "lines": [ + { + "bbox": [ + 307, + 165, + 364, + 222 + ], + "spans": [ + { + "bbox": [ + 307, + 165, + 364, + 222 + ], + "type": "image", + "image_path": "54825d6133390800e1a0171abe75761903599596fb1703bb6d12d0e5247d1a7f.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 365, + 165, + 422, + 222 + ], + "blocks": [ + { + "bbox": [ + 365, + 165, + 422, + 222 + ], + "lines": [ + { + "bbox": [ + 365, + 165, + 422, + 222 + ], + "spans": [ + { + "bbox": [ + 365, + 165, + 422, + 222 + ], + "type": "image", + "image_path": "6b73c694b020ec4ff185f59699e65788f7e71a913010901b7c57739583ff7329.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 304, + 480, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 304, + 480, + 376 + ], + "spans": [ + { + "bbox": [ + 130, + 304, + 480, + 376 + ], + "type": "text", + "content": "propose the REGion-native AutoencoDER (Reader), which takes arbitrarily shaped and sized regions (see Fig. 1d) as the fundamental units for visual tokens, providing more flexibility and aligning better with the visual-prior characteristics. The Reader framework conceptualizes each image as a composite of several regions, with each region corresponding to either a small-sized object or a portion of a regular or large-sized object." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 130, + 377, + 480, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 377, + 480, + 472 + ], + "spans": [ + { + "bbox": [ + 130, + 377, + 480, + 472 + ], + "type": "text", + "content": "Specifically, the Reader architecture comprises an encoder and a decoder, as illustrated in Fig. 3. The encoder first partitions an image into an adaptive-number of non-overlapping regions, where each region is represented as a binary mask and an embedding. Following this, the encoder employs a novel region embedder to integrate the geometric information from the mask with the visual features from the embedding, and produce the desired token for each region. Subsequently, a novel cross-attention-based decoder is employed to reconstruct the original image from these tokens." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 130, + 473, + 482, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 473, + 482, + 496 + ], + "spans": [ + { + "bbox": [ + 130, + 473, + 482, + 496 + ], + "type": "text", + "content": "We further conduct both qualitative and quantitative experiments and demonstrate the following two main notable characteristics of Reader:" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 130, + 498, + 480, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 498, + 480, + 568 + ], + "spans": [ + { + "bbox": [ + 130, + 498, + 480, + 568 + ], + "type": "text", + "content": "Efficiency: Reader achieves highly efficient encoding, yielding reconstruction fidelity superior to that of previous methods, despite employing a significantly smaller number of tokens. This is intuitive since the region-based manner enables Reader to naturally represent complex areas with more regions, while representing simpler areas with fewer regions. This effectively avoids information redundancy, thus possessing inherent efficiency." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 130, + 570, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 481, + 665 + ], + "type": "text", + "content": "Disentanglement: Representing non-overlapping regions as tokens guarantees their disentanglement from one another, and further empower Reader with the ability to perform image editing. Taking object erasure as an example, the disentanglement permits an object in the image to be represented solely by the tokens corresponding to its interior. Through removing these tokens and then reconstructing the image, object erasing can be naturally achieved. Consequently, without any additional modules or training data, Reader facilitates diverse object-level editing operations (see Fig. 2), such as erasing, adding, re" + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 333, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 333, + 91, + 447, + 101 + ], + "type": "text", + "content": "Region-Native Autoencoder" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": "placing, and modifying shapes on objects, and demonstrates excellent performance in the image editing benchmark of smile transferring [31]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 141, + 482, + 201 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 141, + 482, + 201 + ], + "spans": [ + { + "bbox": [ + 130, + 141, + 482, + 201 + ], + "type": "text", + "content": "We believe that these experiments effectively demonstrate the inherent advantages of region-based tokenization method. We hope this will bring some novel insights to the community about visual tokenization and unlock its value in various downstream tasks in the future, such as interaction with LLMs and multimodal perception, generation, and beyond." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 225, + 237, + 239 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 225, + 237, + 239 + ], + "spans": [ + { + "bbox": [ + 132, + 225, + 237, + 239 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 260, + 482, + 429 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 260, + 482, + 429 + ], + "spans": [ + { + "bbox": [ + 130, + 260, + 482, + 429 + ], + "type": "text", + "content": "Grid-based Tokenization The majority of grid-based methods follows the feature extraction, quantization, and reconstruction paradigm, initially introduced by VQVAE [39]. These methods represent image tokens as a grid-shaped code map. Expanding upon this framework, VQGAN [16] incorporates perceptual loss [58] and GAN loss [28] to augment the original L1 reconstruction loss, with the aim of improving both reconstruction fidelity and codebook learning. RQ-VAE [32] introduces residual quantization to further enhance reconstruction fidelity, while leading to the cost of significantly higher number of tokens. ViT-VQGAN [56] replaces the previously employed CNN-based architecture with a ViT-based one. Sharing a similar motivation to ours, DQ-VAE [27] proposes dynamically selecting patch granularity between two or three pre-defined strides to achieve dynamic-length coding. However, this approach remains constrained by the grid-based paradigm. In contrast, our Reader takes arbitrarily shaped regions as image token units and offers greater flexibility." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 433, + 482, + 589 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 433, + 482, + 589 + ], + "spans": [ + { + "bbox": [ + 130, + 433, + 482, + 589 + ], + "type": "text", + "content": "Object-based Tokenization Methods of this kind typically find applications in vision-language pretraining and multimodal perception. For instance, [1] uses a Faster R-CNN [44] for proposals prediction and encodes the proposals into tokens to subsequently perform image captioning or visual question answering. Similarly, [59] also employs a Faster R-CNN to extract object tokens and maps these tokens with a Graph Neural Network [57]. GroupViT [53] proposes a learnable clustering architecture based on ViT [15] to extract object information, and aligns it with text embedding akin to CLIP [41]. However, these methods focus on extracting semantic information at object-level, leading to excessively coarse-grained and overlooking on detailed visual information. In contrast, our Reader employs an autoencoder architecture to extract dynamic-grained local regions as the unit of image tokens, which better preserves detailed information and showcases superior performance in image reconstruction and editing." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": "Scene Decomposition The scene decomposition task [4,37,55], also referred to as object-centric representation, has the most similar form to our Reader. This task aims to decompose a scene into multiple objects and learn the representation of each object. However, these works focus on application in some synthetic datasets containing only basic 3D shapes [20,29,43]. In contrast, our Reader can be applied to a wide range of natural images." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 141, + 101 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 141, + 101 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 141, + 101 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 214, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 214, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 214, + 102 + ], + "type": "text", + "content": "Wang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 224 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 224 + ], + "type": "text", + "content": "Image Editing Most early image editing methods are GAN-based [21]. Some of these approaches [22, 34, 40, 60] rely on the controllability of the GAN's latent space, employing supervised or unsupervised approaches to achieve desired transformations of GAN's latent variable. Other works [8, 31, 52] utilize GANs' conditional generation ability and synthesize the edited images based on the input conditions. The remarkable success of diffusion models [25, 48] in image synthesis has also led to the emergence of numerous diffusion-based editing methods [3, 10, 19, 47, 54]. These approaches largely build upon the robust semantic understanding and rendering capabilities of large pre-trained diffusion models [45]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 224, + 480, + 271 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 224, + 480, + 271 + ], + "spans": [ + { + "bbox": [ + 130, + 224, + 480, + 271 + ], + "type": "text", + "content": "The semantic-map based methods [19, 31, 34] can produce the most similar effects to our Reader. However, compared to these works, our Reader benefits from its region-based representation, enabling a wider variety of editing operations without requiring additional input, with details elaborated in Sec. 3.4." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 291, + 233, + 306 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 291, + 233, + 306 + ], + "spans": [ + { + "bbox": [ + 132, + 291, + 233, + 306 + ], + "type": "text", + "content": "3 Methodology" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 317, + 480, + 400 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 317, + 480, + 400 + ], + "spans": [ + { + "bbox": [ + 130, + 317, + 480, + 400 + ], + "type": "text", + "content": "We propose Reader (region-native autoencoder) in this work, which performs region-based visual tokenization and enables diverse controllable image editing operations. In the following, we will first provide detailed descriptions about the model architecture of the Reader's encoder and decoder in Sec. 3.1 and Sec. 3.2, respectively. Sec. 3.3 outlines the objective function utilized, while Sec. 3.4 delves into the detailed implementation for image editing. The model architecture is illustrated in Fig. 3." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 419, + 270, + 431 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 419, + 270, + 431 + ], + "spans": [ + { + "bbox": [ + 132, + 419, + 270, + 431 + ], + "type": "text", + "content": "3.1 Encoder Architecture" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 440, + 479, + 487 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 440, + 479, + 487 + ], + "spans": [ + { + "bbox": [ + 130, + 440, + 479, + 487 + ], + "type": "text", + "content": "The encoder of Reader needs the capability to first segment an image into numerous regions and then encode each region into a token. Thus, we propose a encoder with three core components: an image feature extractor, a region decoder and a region embedder." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 493, + 479, + 589 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 493, + 479, + 589 + ], + "spans": [ + { + "bbox": [ + 130, + 493, + 479, + 589 + ], + "type": "text", + "content": "Feature Extractor To extract information from the original image for subsequent image segmentation, tokenization, and reconstruction steps, we employ a standard ViT [15] along with a simple FPN [33] as our feature extractor. Given the importance of fine-grained details for high-fidelity reconstruction, we configure the patch size of the ViT to be " + }, + { + "bbox": [ + 130, + 493, + 479, + 589 + ], + "type": "inline_equation", + "content": "8 \\times 8" + }, + { + "bbox": [ + 130, + 493, + 479, + 589 + ], + "type": "text", + "content": ", a smaller size compared to conventional settings, yet effective in preserving local structure. The strides of the simple FPN are set to " + }, + { + "bbox": [ + 130, + 493, + 479, + 589 + ], + "type": "inline_equation", + "content": "(1, \\frac{1}{2})" + }, + { + "bbox": [ + 130, + 493, + 479, + 589 + ], + "type": "text", + "content": ", thus the output features are at resolutions of " + }, + { + "bbox": [ + 130, + 493, + 479, + 589 + ], + "type": "inline_equation", + "content": "\\frac{1}{8}" + }, + { + "bbox": [ + 130, + 493, + 479, + 589 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 493, + 479, + 589 + ], + "type": "inline_equation", + "content": "\\frac{1}{4}" + }, + { + "bbox": [ + 130, + 493, + 479, + 589 + ], + "type": "text", + "content": " of the original image." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 594, + 479, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 479, + 641 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 479, + 641 + ], + "type": "text", + "content": "Region Decoder We employ a region decoder to learn the region representation and segment image into regions. Specifically, we employ a standard Transformer decoder [50], which is similar to previous Transformer-based segmentation works [5-7,36], augmented with an additional bidirectional Transformer block [30]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 641, + 479, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 479, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 479, + 665 + ], + "type": "text", + "content": "In the Transformer decoder, " + }, + { + "bbox": [ + 130, + 641, + 479, + 665 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 641, + 479, + 665 + ], + "type": "text", + "content": " learnable embeddings are used as input queries to extract the region information, termed region queries. Within each" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 333, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 333, + 91, + 447, + 102 + ], + "type": "text", + "content": "Region-Native Autoencoder" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 223 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 223 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 223 + ], + "type": "text", + "content": "layer of the Transformer decoder, one of the two-sized image features alternately serves as the key and value. Through cross-attention, the region queries can extract region-level information from the image features. Subsequently, the bidirectional Transformer block takes the updated region queries and the larger-sized image feature as input to perform both region-to-image and image-to-region attention and update them. This process yields the region embeddings " + }, + { + "bbox": [ + 130, + 116, + 482, + 223 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_r \\in \\mathbb{R}^{N \\times c}" + }, + { + "bbox": [ + 130, + 116, + 482, + 223 + ], + "type": "text", + "content": " and a region-aware image feature " + }, + { + "bbox": [ + 130, + 116, + 482, + 223 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_{img} \\in \\mathbb{R}^{c \\times h \\times w}" + }, + { + "bbox": [ + 130, + 116, + 482, + 223 + ], + "type": "text", + "content": ". Here, " + }, + { + "bbox": [ + 130, + 116, + 482, + 223 + ], + "type": "inline_equation", + "content": "(h, w)" + }, + { + "bbox": [ + 130, + 116, + 482, + 223 + ], + "type": "text", + "content": " represents the resolution of the image feature, which is " + }, + { + "bbox": [ + 130, + 116, + 482, + 223 + ], + "type": "inline_equation", + "content": "\\frac{1}{4}" + }, + { + "bbox": [ + 130, + 116, + 482, + 223 + ], + "type": "text", + "content": " of the original image size, and " + }, + { + "bbox": [ + 130, + 116, + 482, + 223 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 130, + 116, + 482, + 223 + ], + "type": "text", + "content": " denotes the channel dimension." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 224, + 481, + 248 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 224, + 481, + 248 + ], + "spans": [ + { + "bbox": [ + 131, + 224, + 481, + 248 + ], + "type": "text", + "content": "We then compute the similarity matrix between the region embeddings and the image feature via a dot product, formulated as:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 249, + 254, + 481, + 269 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 249, + 254, + 481, + 269 + ], + "spans": [ + { + "bbox": [ + 249, + 254, + 481, + 269 + ], + "type": "interline_equation", + "content": "S = \\mathcal {E} _ {r} \\odot \\mathcal {E} _ {i m g} \\in \\mathcal {R} ^ {N \\times h \\times w}. \\tag {1}", + "image_path": "0662f03775e957db0f3a186465384720e2258e22f3a973e8b574983023753cf1.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 274, + 482, + 334 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 274, + 482, + 334 + ], + "spans": [ + { + "bbox": [ + 131, + 274, + 482, + 334 + ], + "type": "text", + "content": "We use this similarity matrix to get the desired regions. To prevent information redundancy caused by overlapping, we apply an argmax operation along the first dimension of the similarity matrix. This process means matching each pixel with one embedding that has the highest similarity to the pixel's feature, which can be formulated as:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 211, + 340, + 481, + 354 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 211, + 340, + 481, + 354 + ], + "spans": [ + { + "bbox": [ + 211, + 340, + 481, + 354 + ], + "type": "interline_equation", + "content": "S _ {\\text {h a r d}} = \\operatorname {o n e - h o t} (\\operatorname {a r g m a x} (S)) \\in \\{0, 1 \\} ^ {N \\times h \\times w}. \\tag {2}", + "image_path": "e7a2fff69167598ce9f01020350beb990a42b5aaff10dd14bd30c4e913b4117e.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 361, + 482, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 361, + 482, + 456 + ], + "spans": [ + { + "bbox": [ + 130, + 361, + 482, + 456 + ], + "type": "text", + "content": "Consequently, the " + }, + { + "bbox": [ + 130, + 361, + 482, + 456 + ], + "type": "inline_equation", + "content": "S_{\\mathrm{hard}}" + }, + { + "bbox": [ + 130, + 361, + 482, + 456 + ], + "type": "text", + "content": " in Eq. (2) can also be seen as " + }, + { + "bbox": [ + 130, + 361, + 482, + 456 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 361, + 482, + 456 + ], + "type": "text", + "content": " binary masks, where the foreground of each mask indicates which pixels belong to that specific region. However, not each one among the " + }, + { + "bbox": [ + 130, + 361, + 482, + 456 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 361, + 482, + 456 + ], + "type": "text", + "content": " masks contains foreground. In fact, in some simple images, numerous pixels may be matched with a single region embedding, while other region embeddings remain unmatched with any pixel (illustrated as \"used\" and \"unused\" in Fig. 3). Therefore, while " + }, + { + "bbox": [ + 130, + 361, + 482, + 456 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 361, + 482, + 456 + ], + "type": "text", + "content": " is a fixed hyper-parameter, the actual number of regions corresponding to each image is dynamic, determined by the complexity of the image." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 461, + 482, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 461, + 482, + 521 + ], + "spans": [ + { + "bbox": [ + 130, + 461, + 482, + 521 + ], + "type": "text", + "content": "Region Embedder Through the region decoder, the original image is represented as several regions. Each region comprises a region embedding " + }, + { + "bbox": [ + 130, + 461, + 482, + 521 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_r[i]" + }, + { + "bbox": [ + 130, + 461, + 482, + 521 + ], + "type": "text", + "content": " representing the visual feature and a binary mask " + }, + { + "bbox": [ + 130, + 461, + 482, + 521 + ], + "type": "inline_equation", + "content": "S_{\\mathrm{hard}}[i]" + }, + { + "bbox": [ + 130, + 461, + 482, + 521 + ], + "type": "text", + "content": " representing the geometric information (e.g., shape, position and scale). The region embedder is further employed to merge the above two components and produce the desired tokens." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 521, + 482, + 570 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 482, + 570 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 482, + 570 + ], + "type": "text", + "content": "Following GroupViT [53], we first aggregate the feature of each pixel into the corresponding region embedding. To make the entire process differentiable, enabling the gradients of image reconstruction to guide the granularity of regions, we employ the straight-through trick [39] here. This process is formulated as:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 234, + 576, + 480, + 589 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 234, + 576, + 480, + 589 + ], + "spans": [ + { + "bbox": [ + 234, + 576, + 480, + 589 + ], + "type": "interline_equation", + "content": "S = \\operatorname {s o f t m a x} (S, \\dim = 0), \\tag {3}", + "image_path": "af2e896d430d023b60ed2c18f315c8be597a6fcf3cd645fea2ab9c42cb294a40.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 234, + 593, + 480, + 606 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 234, + 593, + 480, + 606 + ], + "spans": [ + { + "bbox": [ + 234, + 593, + 480, + 606 + ], + "type": "interline_equation", + "content": "\\tilde {S} = S _ {\\text {h a r d}} + S - \\operatorname {s g} \\cdot (S), \\tag {4}", + "image_path": "06d450c37e986ed5c6f25290bd23965ed9c85b0e3f6154144b6328b4b1463562.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 222, + 609, + 480, + 635 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 222, + 609, + 480, + 635 + ], + "spans": [ + { + "bbox": [ + 222, + 609, + 480, + 635 + ], + "type": "interline_equation", + "content": "\\hat {\\mathcal {E}} _ {r} [ i ] = \\mathcal {E} _ {r} [ i ] + \\sum_ {j, k} \\hat {S} [ i, j, k ] \\cdot \\mathcal {E} _ {i m g} [:, j, k ], \\tag {5}", + "image_path": "4373d176f3de01c437cc4759c8a26818b6322c601085e3ae8537efa8f6dfb945.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 131, + 641, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 641, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 131, + 641, + 481, + 666 + ], + "type": "text", + "content": "in which the " + }, + { + "bbox": [ + 131, + 641, + 481, + 666 + ], + "type": "inline_equation", + "content": "\\mathrm{sg}\\cdot ()" + }, + { + "bbox": [ + 131, + 641, + 481, + 666 + ], + "type": "text", + "content": " refers to stop gradient. While the pixel features aggregation process shown in Eq. (5) is actually a pooling process, leading to over-smoothing" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 214, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 214, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 214, + 102 + ], + "type": "text", + "content": "Wang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 132, + 115, + 481, + 240 + ], + "blocks": [ + { + "bbox": [ + 132, + 115, + 481, + 240 + ], + "lines": [ + { + "bbox": [ + 132, + 115, + 481, + 240 + ], + "spans": [ + { + "bbox": [ + 132, + 115, + 481, + 240 + ], + "type": "image", + "image_path": "64d85e7b2d3ad9641b2068c725a153eea26732b31626e306a51131a4f52d987c.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 243, + 480, + 298 + ], + "lines": [ + { + "bbox": [ + 130, + 243, + 480, + 298 + ], + "spans": [ + { + "bbox": [ + 130, + 243, + 480, + 298 + ], + "type": "text", + "content": "Fig. 3: An overview of the proposed Reader. It comprises an encoder and a decoder. The encoder can segment each image into an adaptive number of regions, which is illustrated as the \"used\" part, and embeds each region into a token with the Region Embedder. The decoder then reconstructs the original image from these tokens with the learnable patch queries." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 313, + 479, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 313, + 479, + 361 + ], + "spans": [ + { + "bbox": [ + 130, + 313, + 479, + 361 + ], + "type": "text", + "content": "on local details. We additionally employ a masked cross-attention layer [6] to enhance the visual feature extraction. This layer uses the region embeddings " + }, + { + "bbox": [ + 130, + 313, + 479, + 361 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{E}}_r" + }, + { + "bbox": [ + 130, + 313, + 479, + 361 + ], + "type": "text", + "content": " as queries to perform attention to the image feature " + }, + { + "bbox": [ + 130, + 313, + 479, + 361 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_{img}" + }, + { + "bbox": [ + 130, + 313, + 479, + 361 + ], + "type": "text", + "content": " only in the foreground of each region in order to extract more local details." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 361, + 482, + 482 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 361, + 482, + 482 + ], + "spans": [ + { + "bbox": [ + 130, + 361, + 482, + 482 + ], + "type": "text", + "content": "Subsequently, we incorporate a module to explicitly encode the geometric information in the masks and integrate it into the region embeddings. This approach serves two primary purposes: Firstly, the inclusion of geometric information enhances the delineation of boundaries between regions, thereby facilitating image reconstruction. Secondly, such explicit encoding also enables more editing operations in subsequent steps. Specifically, we flatten each mask and utilize a linear layer to project it into a latent space. Additionally, the region embeddings are also projected with an MLP and are then fused with the former projection through element-wise addition. Following a normalization and another MLP, the region embedder produces the desired region-based image tokens." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 500, + 270, + 511 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 500, + 270, + 511 + ], + "spans": [ + { + "bbox": [ + 131, + 500, + 270, + 511 + ], + "type": "text", + "content": "3.2 Decoder Architecture" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 521, + 480, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 480, + 593 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 480, + 593 + ], + "type": "text", + "content": "The encoder of Reader represents each image as an adaptive-length sequence. Consequently, the decoder needs the capacity to decode this adaptive-length sequence into a uniformly-sized image. This poses a challenge of mapping variable-length input to a fixed-length output. To address this, the decoder architecture of Reader is based on the Transformer decoder, leveraging the cross-attention mechanism to resolve the aforementioned problem." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": "Specifically, we introduce " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": " learnable embeddings as the input queries for the Transformer decoder. Each embedding is responsible for reconstructing a local patch in the output image and is referred to as a patch query. The queries are also incorporated with the positional embeddings of the center coordinate of the patches. To prevent information loss and enable the Transformer decoder to focus on different information at each layer, we employ a separate MLP for each" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 333, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 333, + 91, + 447, + 101 + ], + "type": "text", + "content": "Region-Native Autoencoder" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": "layer to project the image tokens. These projections then serve as the key and value for the corresponding layer." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 140, + 480, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 140, + 480, + 200 + ], + "spans": [ + { + "bbox": [ + 130, + 140, + 480, + 200 + ], + "type": "text", + "content": "Through the Transformer decoder, the patch queries interact with the image tokens through cross-attention, gathering information related to the corresponding patches and updating their values. After the Transformer decoder, we integrate an linear layer to finally map the output patch queries to the pixel space, which is a reverse process of the patch embedder of ViT." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 213, + 257, + 226 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 213, + 257, + 226 + ], + "spans": [ + { + "bbox": [ + 131, + 213, + 257, + 226 + ], + "type": "text", + "content": "3.3 Objective Function" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 230, + 479, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 230, + 479, + 266 + ], + "spans": [ + { + "bbox": [ + 130, + 230, + 479, + 266 + ], + "type": "text", + "content": "The objective functions related to the image reconstruction are kept consistent with VQGAN [16], which is a combination of the L1 loss, perceptual loss [58] and a patch-based GAN loss [28], denoted as " + }, + { + "bbox": [ + 130, + 230, + 479, + 266 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{img}" + }, + { + "bbox": [ + 130, + 230, + 479, + 266 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 266, + 480, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 266, + 480, + 348 + ], + "spans": [ + { + "bbox": [ + 130, + 266, + 480, + 348 + ], + "type": "text", + "content": "However, in our observation, relying solely on the image reconstruction loss is insufficient for training the encoder to produce semantically meaningful regions. In practice, the regions easily collapse into uniformly distributed patches. Therefore, additional supervision on the output regions becomes necessary. We employ either the pseudo-annotations produced by SAM [30] or manual annotations to supervise the output regions in different experiments, with detailed information provided in Sec. 4." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 350, + 480, + 433 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 350, + 480, + 433 + ], + "spans": [ + { + "bbox": [ + 130, + 350, + 480, + 433 + ], + "type": "text", + "content": "In order to train the similarity matrix in Eq. (3) to resemble a one-hot form at each pixel, we employ a cross-entropy loss here. Unlike the tasks, where cross-entropy typically used with defined category IDs, there is no fixed correspondence between the predicted regions and the masks of annotations here. Following [5,7], we employ a Hungarian-algorithm based bipartite matching to determine the matching relationship, with more details provided in supplementary material. The overall objective function is formulated as:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 271, + 439, + 480, + 452 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 271, + 439, + 480, + 452 + ], + "spans": [ + { + "bbox": [ + 271, + 439, + 480, + 452 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathcal {L} _ {i m g} + \\mathcal {L} _ {c e}. \\tag {6}", + "image_path": "b1d4d1102138d0f62ea2bc099b5ddd96ad276d4bf3475f93be2c5f7832668c4f.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 464, + 316, + 477 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 464, + 316, + 477 + ], + "spans": [ + { + "bbox": [ + 131, + 464, + 316, + 477 + ], + "type": "text", + "content": "3.4 Image Editing Implementation" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 481, + 479, + 528 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 481, + 479, + 528 + ], + "spans": [ + { + "bbox": [ + 130, + 481, + 479, + 528 + ], + "type": "text", + "content": "The proposed Reader inherently facilitates diverse image editing operations. Without any additional modules, this is achieved by simply manipulating the image tokens outputted by the encoder and then inputting these manipulated tokens into the decoder to get the edited image." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 529, + 480, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 529, + 480, + 601 + ], + "spans": [ + { + "bbox": [ + 130, + 529, + 480, + 601 + ], + "type": "text", + "content": "Specifically, before the region embedder of Reader, each region is represented by two components: the geometric information (represented as a binary mask) and the visual information (represented as the region embedding). This enables independent manipulations on both two parts and achieves different editing effects. Therefore, we define the following three-level image editing operations. A schematic diagram is shown in Fig. 4 to provide an intuitive understanding." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 131, + 605, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 605, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 131, + 605, + 482, + 665 + ], + "type": "text", + "content": "Geometry Level These operations refer to only manipulating the masks through affine transformations or interactive brushing by users, while maintaining the embeddings unaltered. Geometry-level operations allows to perform moving, scaling, and shape modification on objects. As an example, the left branch in Fig. 4 alters the shape of the lips without changing their color." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 214, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 214, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 214, + 102 + ], + "type": "text", + "content": "Wang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 132, + 114, + 482, + 224 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 482, + 224 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 482, + 224 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 482, + 224 + ], + "type": "image", + "image_path": "befe1cac5f5f0d4bf3fdf961ce3fae5bcca9d14164161864e437a0a5d07f55e0.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 228, + 482, + 262 + ], + "lines": [ + { + "bbox": [ + 130, + 228, + 482, + 262 + ], + "spans": [ + { + "bbox": [ + 130, + 228, + 482, + 262 + ], + "type": "text", + "content": "Fig. 4: Illustration of editing operations. Different operations on the embeddings and masks leads to different editing effects. The left, middle and right branches correspond to the three-levels operations described in Sec. 3.4, respectively." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 277, + 479, + 338 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 277, + 479, + 338 + ], + "spans": [ + { + "bbox": [ + 130, + 277, + 479, + 338 + ], + "type": "text", + "content": "Feature Level Feature-level operations maintain the mask unchanged while replacing the corresponding embedding with one from another object or another image. This facilitates local feature transferring without impacting geometric characteristics. An example is illustrated in the middle branch of Fig. 4, where the color of lips is changed by replacing the corresponding two embeddings." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 342, + 482, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 342, + 482, + 450 + ], + "spans": [ + { + "bbox": [ + 130, + 342, + 482, + 450 + ], + "type": "text", + "content": "Token Level The erasure, addition and replacement of objects are categorized as this level. Erasure can be achieved by removing tokens corresponding to a specific object. Conversely, addition can be accomplished by appending tokens from another image to the tokens of the current image. Object replacement can then be achieved by combining both operations. Furthermore, these operations can also be combined with the geometry-level manipulations to achieve addition and replacement on objects while controlling their shapes, scale and position, etc. An example is illustrated in the right branch of Fig. 4, where the mouth and lips are replaced by those from another image." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 469, + 230, + 483 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 469, + 230, + 483 + ], + "spans": [ + { + "bbox": [ + 132, + 469, + 230, + 483 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 495, + 289, + 508 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 495, + 289, + 508 + ], + "spans": [ + { + "bbox": [ + 132, + 495, + 289, + 508 + ], + "type": "text", + "content": "4.1 Experiment on ImageNet" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 517, + 479, + 541 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 517, + 479, + 541 + ], + "spans": [ + { + "bbox": [ + 130, + 517, + 479, + 541 + ], + "type": "text", + "content": "Herein we train Reader on ImageNet training set and report qualitative results in image editing and quantitative comparisons on reconstruction fidelity." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": "Implementation As outlined in Sec. 3.3, additional supervision is essential to prevent the collapse of producing regions. Consequently, in this experiment, we employ SAM [30] to produce pseudo-annotations on the ImageNet training set. Specifically, we keep the original settings of the segment anything mode of SAM, which effectively segments all prominent objects or local parts in images. Since there is substantial overlapping in the original results of SAM, while the cross-entropy loss needs non-overlapping annotations, we implement the following two post-processing steps. Firstly, removing all larger masks that encompass other ones, retaining only smaller, more localized masks. Secondly, regions not encompassed by the pseudo annotations do not contribute to the objective function." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 333, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 333, + 91, + 447, + 101 + ], + "type": "text", + "content": "Region-Native Autoencoder" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 132, + 113, + 482, + 292 + ], + "blocks": [ + { + "bbox": [ + 132, + 113, + 482, + 292 + ], + "lines": [ + { + "bbox": [ + 132, + 113, + 482, + 292 + ], + "spans": [ + { + "bbox": [ + 132, + 113, + 482, + 292 + ], + "type": "image", + "image_path": "9e205f9a1b59472ba7e1d89678a225844b2ae1d826cda4eb442548f3cc89b592.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 294, + 482, + 317 + ], + "lines": [ + { + "bbox": [ + 130, + 294, + 482, + 317 + ], + "spans": [ + { + "bbox": [ + 130, + 294, + 482, + 317 + ], + "type": "text", + "content": "Fig. 5: Qualitative results in ImageNet validation set and Internet images. The editing objective is labeled with text under image." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 343, + 482, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 343, + 482, + 464 + ], + "spans": [ + { + "bbox": [ + 130, + 343, + 482, + 464 + ], + "type": "text", + "content": "The input and output sizes of images are set to " + }, + { + "bbox": [ + 130, + 343, + 482, + 464 + ], + "type": "inline_equation", + "content": "(256 \\times 256)" + }, + { + "bbox": [ + 130, + 343, + 482, + 464 + ], + "type": "text", + "content": " in this experiment. Both the ViT in the encoder and the Transformer in the decoder follow the configuration of ViT-tiny, which utilizes 192 channels and 3 heads in the multi-head attention. The dimension of image tokens is set to 256 to maintain consistency with the previous methods. The number " + }, + { + "bbox": [ + 130, + 343, + 482, + 464 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 343, + 482, + 464 + ], + "type": "text", + "content": " of learnable region queries in the encoder is set to 256, which is the minimal tokens number of each image in previous works and serves as the maximum regions number of our model. The number " + }, + { + "bbox": [ + 130, + 343, + 482, + 464 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 130, + 343, + 482, + 464 + ], + "type": "text", + "content": " of learnable patch queries in the decoder is set to 1024, with each query responsible for reconstructing a " + }, + { + "bbox": [ + 130, + 343, + 482, + 464 + ], + "type": "inline_equation", + "content": "8 \\times 8" + }, + { + "bbox": [ + 130, + 343, + 482, + 464 + ], + "type": "text", + "content": " patch in the output image. For more implementation details, please refer to our supplementary material." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 468, + 482, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 468, + 482, + 540 + ], + "spans": [ + { + "bbox": [ + 130, + 468, + 482, + 540 + ], + "type": "text", + "content": "Results We provide several image editing examples in Fig. 5 corresponding to all the three-level editing operations described in Sec. 3.4. For example, \"sticking out the tongue\" needs only manipulation on the tongue's mask, \"replacing the color\" is achieved by replacing the region embedding of the strawberry and \"replacing with another hat\" needs deleting the tokens of the original hat and replacing them with new ones from another image." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "type": "text", + "content": "For quantitative comparison of the reconstruction fidelity, we employ both the distribution-level metric rFID [24] and the image-patches-level metrics PSNR, SSIM [26], LPIPS [58] to evaluate the reconstructed images of ImageNet validation set. We employ the VQGAN [16], RQ-VAE [32] and DQ-VAE [32] as baselines. The former two employ fixed tokens number of each image, while the latter uses dynamic tokens number through picking one granularity for each patch from three pre-defined strides " + }, + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\left(\\frac{1}{8},\\frac{1}{16},\\frac{1}{32}\\right)" + }, + { + "bbox": [ + 130, + 546, + 482, + 666 + ], + "type": "text", + "content": ". The tokens number of DQ-VAE is computed by the ratio of different strides provided in their paper. Tokens number of our Reader corresponds to the average number observed across the entire ImageNet validation set." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 214, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 214, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 214, + 102 + ], + "type": "text", + "content": "Wang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 137, + 168, + 471, + 258 + ], + "blocks": [ + { + "bbox": [ + 130, + 114, + 482, + 160 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 482, + 160 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 482, + 160 + ], + "type": "text", + "content": "Table 1: Quantitative results for the reconstruction fidelity on the ImageNet validation set. \"Tokens\" refers to the number of image tokens utilized for each image. The dimension of image tokens is 256 for all methods. The \"+\" denotes data excerpted from the original paper. Top 1 and top 2 results are highlighted in red and blue, respectively." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 137, + 168, + 471, + 258 + ], + "lines": [ + { + "bbox": [ + 137, + 168, + 471, + 258 + ], + "spans": [ + { + "bbox": [ + 137, + 168, + 471, + 258 + ], + "type": "table", + "html": "
MethodTokensPSNR ↑SSIM ↑LPIPS ↓rFID ↓Params(M)FLOPs(G)
VQGAN(1024)16 × 1619.360.480.318.0572.14196.33
VQGAN(16384)16 × 1619.830.490.284.9276.07197.33
RQ-VAE(16384)†8 × 8 × 210.77
RQ-VAE(16384)8 × 8 × 420.030.500.285.24104.40193.64
DQ-VAE(16384)†2564.08 - 7.32
Reader(ours)79.0924.750.720.202.4326.6040.24
", + "image_path": "e161aa6587dc685d67210ad941d2bf0363993b4fb9122e16219a2b2179990b93.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 280, + 482, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 280, + 482, + 376 + ], + "spans": [ + { + "bbox": [ + 130, + 280, + 482, + 376 + ], + "type": "text", + "content": "Results shown in Tab. 1 indicate that, in comparison to the baselines, our Reader attains superior reconstruction fidelity despite having the smallest tokens number, model parameters and computational complexity. While compared to DQ-VAE, which also employs dynamic length encoding, our method also demonstrates significant superiority. This finding is intuitive that, compared to both the fixed- and dynamic-sized patches, arbitrarily shaped regions can better handle the uneven information density in images and avoid information redundancy, thus achieving higher reconstruction fidelity with fewer tokens." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 395, + 303, + 407 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 395, + 303, + 407 + ], + "spans": [ + { + "bbox": [ + 131, + 395, + 303, + 407 + ], + "type": "text", + "content": "4.2 Experiment on Human Face" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 415, + 481, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 415, + 481, + 453 + ], + "spans": [ + { + "bbox": [ + 130, + 415, + 481, + 453 + ], + "type": "text", + "content": "To further evaluate the editing capabilities of our Reader, we transfer it to images of human faces, which are commonly employed in previous image editing works, allowing to provide quantitative comparisons." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 456, + 482, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 456, + 482, + 540 + ], + "spans": [ + { + "bbox": [ + 130, + 456, + 482, + 540 + ], + "type": "text", + "content": "Implementation We fine-tune the Reader in this experiment, which is pretrained as outlined in Sec. 4.1. We utilize the CelebAMask-HQ dataset [31], a face parsing dataset with manual annotations. The input and output sizes of images for this experiment are configured as " + }, + { + "bbox": [ + 130, + 456, + 482, + 540 + ], + "type": "inline_equation", + "content": "(512 \\times 512)" + }, + { + "bbox": [ + 130, + 456, + 482, + 540 + ], + "type": "text", + "content": " to align with the previous methods [31, 34]. In order to preserving the knowledge already learned by the pre-trained model, we keep the majority of Reader architecture unchanged and make the following modifications." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 541, + 482, + 625 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 541, + 482, + 625 + ], + "spans": [ + { + "bbox": [ + 130, + 541, + 482, + 625 + ], + "type": "text", + "content": "(1) Adding a downsampling-and-convolution layer with a residual connection in the region embedder and initializing the new convolution's output to zero. This is employed to make the double-sized masks meet the input length of previous mask projection layer. (2) Increasing the number of learnable patch queries from 1024 to 4096 in order to get double-sized reconstructed images. While loading the pre-trained weights, we perform bicubic interpolation on the weights of the positional embeddings and the patch queries to match the new size." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "text", + "content": "Results In Fig. 6, we present the qualitative results corresponding to the three-level editing operations. In the top and middle rows, we performed overlay editing, gradually changing one part of the geometric and visual properties at a" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 333, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 333, + 91, + 447, + 102 + ], + "type": "text", + "content": "Region-Native Autoencoder" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 114, + 191, + 172 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 191, + 172 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 191, + 172 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 191, + 172 + ], + "type": "image", + "image_path": "0843069c29102297a94cb9bdcfce2ac88a13bf57ab5834d64be5d15020008956.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 192, + 114, + 249, + 172 + ], + "blocks": [ + { + "bbox": [ + 192, + 114, + 249, + 172 + ], + "lines": [ + { + "bbox": [ + 192, + 114, + 249, + 172 + ], + "spans": [ + { + "bbox": [ + 192, + 114, + 249, + 172 + ], + "type": "image", + "image_path": "85c806a5f9d63341ca97c53c51c8940f4e65c13cb4ac7b427a727ebfbd0e349b.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 198, + 172, + 243, + 178 + ], + "lines": [ + { + "bbox": [ + 198, + 172, + 243, + 178 + ], + "spans": [ + { + "bbox": [ + 198, + 172, + 243, + 178 + ], + "type": "text", + "content": "+ Close Mouth" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 250, + 114, + 306, + 172 + ], + "blocks": [ + { + "bbox": [ + 250, + 114, + 306, + 172 + ], + "lines": [ + { + "bbox": [ + 250, + 114, + 306, + 172 + ], + "spans": [ + { + "bbox": [ + 250, + 114, + 306, + 172 + ], + "type": "image", + "image_path": "9901f7026a6bd0ed496a2d301877969a2496363d27789a93412244ce7d048c79.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 255, + 172, + 302, + 178 + ], + "lines": [ + { + "bbox": [ + 255, + 172, + 302, + 178 + ], + "spans": [ + { + "bbox": [ + 255, + 172, + 302, + 178 + ], + "type": "text", + "content": "+ Sharp Jawline" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 307, + 114, + 364, + 172 + ], + "blocks": [ + { + "bbox": [ + 307, + 114, + 364, + 172 + ], + "lines": [ + { + "bbox": [ + 307, + 114, + 364, + 172 + ], + "spans": [ + { + "bbox": [ + 307, + 114, + 364, + 172 + ], + "type": "image", + "image_path": "564cc1003e0345133e3002ac21294de28a7502234c630d9d7915746ede7adb58.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 365, + 114, + 422, + 172 + ], + "blocks": [ + { + "bbox": [ + 365, + 114, + 422, + 172 + ], + "lines": [ + { + "bbox": [ + 365, + 114, + 422, + 172 + ], + "spans": [ + { + "bbox": [ + 365, + 114, + 422, + 172 + ], + "type": "image", + "image_path": "b2fb8084a0e25d74b0507884c48fb19e6164a3ef0ebfe272cfddf1a72dcc2164.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 371, + 172, + 417, + 178 + ], + "lines": [ + { + "bbox": [ + 371, + 172, + 417, + 178 + ], + "spans": [ + { + "bbox": [ + 371, + 172, + 417, + 178 + ], + "type": "text", + "content": "+ Enlarge Nose" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 424, + 114, + 481, + 172 + ], + "blocks": [ + { + "bbox": [ + 424, + 114, + 481, + 172 + ], + "lines": [ + { + "bbox": [ + 424, + 114, + 481, + 172 + ], + "spans": [ + { + "bbox": [ + 424, + 114, + 481, + 172 + ], + "type": "image", + "image_path": "91d5fb0201234a8e9846d18ecfdbbe05dd6993b4cdceec07397f2c9fed2257fb.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 429, + 172, + 474, + 178 + ], + "lines": [ + { + "bbox": [ + 429, + 172, + 474, + 178 + ], + "spans": [ + { + "bbox": [ + 429, + 172, + 474, + 178 + ], + "type": "text", + "content": "+Wide Jawline" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 133, + 178, + 192, + 236 + ], + "blocks": [ + { + "bbox": [ + 133, + 178, + 192, + 236 + ], + "lines": [ + { + "bbox": [ + 133, + 178, + 192, + 236 + ], + "spans": [ + { + "bbox": [ + 133, + 178, + 192, + 236 + ], + "type": "image", + "image_path": "7977c98a981c09c9856ecc980d374a632993124a4b2e063c558ca73763ef1627.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 192, + 178, + 249, + 236 + ], + "blocks": [ + { + "bbox": [ + 192, + 178, + 249, + 236 + ], + "lines": [ + { + "bbox": [ + 192, + 178, + 249, + 236 + ], + "spans": [ + { + "bbox": [ + 192, + 178, + 249, + 236 + ], + "type": "image", + "image_path": "0e2567278f8901560a6dc53f6c429b20a27c9f4a57e77780827287b6e3e1d39a.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 198, + 236, + 242, + 243 + ], + "lines": [ + { + "bbox": [ + 198, + 236, + 242, + 243 + ], + "spans": [ + { + "bbox": [ + 198, + 236, + 242, + 243 + ], + "type": "text", + "content": "+ Eyes Feature" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 250, + 178, + 306, + 236 + ], + "blocks": [ + { + "bbox": [ + 250, + 178, + 306, + 236 + ], + "lines": [ + { + "bbox": [ + 250, + 178, + 306, + 236 + ], + "spans": [ + { + "bbox": [ + 250, + 178, + 306, + 236 + ], + "type": "image", + "image_path": "2d916f4abdc3cdb915cdbbf64078d7b166603006fe77c01ead5e250e1654db83.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 256, + 236, + 298, + 243 + ], + "lines": [ + { + "bbox": [ + 256, + 236, + 298, + 243 + ], + "spans": [ + { + "bbox": [ + 256, + 236, + 298, + 243 + ], + "type": "text", + "content": "+ Lips Featue" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 307, + 178, + 364, + 236 + ], + "blocks": [ + { + "bbox": [ + 307, + 178, + 364, + 236 + ], + "lines": [ + { + "bbox": [ + 307, + 178, + 364, + 236 + ], + "spans": [ + { + "bbox": [ + 307, + 178, + 364, + 236 + ], + "type": "image", + "image_path": "2c1d0ec5a0a5073222d5b857650115b8c2fcb1a8b40b5ccf5522e4b588d7434b.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 364, + 236, + 416, + 243 + ], + "lines": [ + { + "bbox": [ + 364, + 236, + 416, + 243 + ], + "spans": [ + { + "bbox": [ + 364, + 236, + 416, + 243 + ], + "type": "text", + "content": "+ Eyes Feature" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 365, + 178, + 422, + 236 + ], + "blocks": [ + { + "bbox": [ + 365, + 178, + 422, + 236 + ], + "lines": [ + { + "bbox": [ + 365, + 178, + 422, + 236 + ], + "spans": [ + { + "bbox": [ + 365, + 178, + 422, + 236 + ], + "type": "image", + "image_path": "1318f74c05214260178b920946013c4e9190701061041328e809781210225f53.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 424, + 178, + 481, + 236 + ], + "blocks": [ + { + "bbox": [ + 424, + 178, + 481, + 236 + ], + "lines": [ + { + "bbox": [ + 424, + 178, + 481, + 236 + ], + "spans": [ + { + "bbox": [ + 424, + 178, + 481, + 236 + ], + "type": "image", + "image_path": "22775f61615154e2fba5987b10510adb5b1ebf9a20f32efa001b497131235eb5.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 431, + 236, + 472, + 243 + ], + "lines": [ + { + "bbox": [ + 431, + 236, + 472, + 243 + ], + "spans": [ + { + "bbox": [ + 431, + 236, + 472, + 243 + ], + "type": "text", + "content": "+ Lips Featue" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 133, + 244, + 191, + 301 + ], + "blocks": [ + { + "bbox": [ + 133, + 244, + 191, + 301 + ], + "lines": [ + { + "bbox": [ + 133, + 244, + 191, + 301 + ], + "spans": [ + { + "bbox": [ + 133, + 244, + 191, + 301 + ], + "type": "image", + "image_path": "b887752c7c9b846802e32352e1b3971500911bf57a603d765b4626266817d3dc.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 192, + 244, + 249, + 301 + ], + "blocks": [ + { + "bbox": [ + 192, + 244, + 249, + 301 + ], + "lines": [ + { + "bbox": [ + 192, + 244, + 249, + 301 + ], + "spans": [ + { + "bbox": [ + 192, + 244, + 249, + 301 + ], + "type": "image", + "image_path": "da5a91604be712f05143c8028fb883de9127e87770ed56d7941eae1c66d815ee.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 203, + 301, + 233, + 309 + ], + "lines": [ + { + "bbox": [ + 203, + 301, + 233, + 309 + ], + "spans": [ + { + "bbox": [ + 203, + 301, + 233, + 309 + ], + "type": "text", + "content": "+Smiling" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 250, + 243, + 306, + 301 + ], + "blocks": [ + { + "bbox": [ + 250, + 243, + 306, + 301 + ], + "lines": [ + { + "bbox": [ + 250, + 243, + 306, + 301 + ], + "spans": [ + { + "bbox": [ + 250, + 243, + 306, + 301 + ], + "type": "image", + "image_path": "eb9e09e5a26800d3ef79050ee06904dcc3f1380e93b1e752f57910fdc04bc6b5.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 306, + 243, + 362, + 301 + ], + "blocks": [ + { + "bbox": [ + 306, + 243, + 362, + 301 + ], + "lines": [ + { + "bbox": [ + 306, + 243, + 362, + 301 + ], + "spans": [ + { + "bbox": [ + 306, + 243, + 362, + 301 + ], + "type": "image", + "image_path": "645e4dbf9be656c983b39f5d438f841d8086cc72b4de513acaaed89a09870448.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 365, + 243, + 422, + 301 + ], + "blocks": [ + { + "bbox": [ + 365, + 243, + 422, + 301 + ], + "lines": [ + { + "bbox": [ + 365, + 243, + 422, + 301 + ], + "spans": [ + { + "bbox": [ + 365, + 243, + 422, + 301 + ], + "type": "image", + "image_path": "7d89dad10c48dbcfa0c07f65866bb69638af4746c59c75b382aaee427e90bfe2.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 424, + 243, + 481, + 301 + ], + "blocks": [ + { + "bbox": [ + 424, + 243, + 481, + 301 + ], + "lines": [ + { + "bbox": [ + 424, + 243, + 481, + 301 + ], + "spans": [ + { + "bbox": [ + 424, + 243, + 481, + 301 + ], + "type": "image", + "image_path": "d58c5f568a5d00bfa6884e4c9dc1ea43241b564c020ab262652f8eb76a254976.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 436, + 301, + 465, + 309 + ], + "lines": [ + { + "bbox": [ + 436, + 301, + 465, + 309 + ], + "spans": [ + { + "bbox": [ + 436, + 301, + 465, + 309 + ], + "type": "text", + "content": "+Smiling" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_caption" + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 133, + 338, + 190, + 396 + ], + "blocks": [ + { + "bbox": [ + 132, + 312, + 480, + 334 + ], + "lines": [ + { + "bbox": [ + 132, + 312, + 480, + 334 + ], + "spans": [ + { + "bbox": [ + 132, + 312, + 480, + 334 + ], + "type": "text", + "content": "Fig. 6: Qualitative results in huaman faces. The editing objective is labeled with text under corresponding image." + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 133, + 338, + 190, + 396 + ], + "lines": [ + { + "bbox": [ + 133, + 338, + 190, + 396 + ], + "spans": [ + { + "bbox": [ + 133, + 338, + 190, + 396 + ], + "type": "image", + "image_path": "377ad898242c7c411bb7dc163c84e923225bf5150e9e1d59bee996e05dc4f3f5.jpg" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 140, + 396, + 183, + 403 + ], + "lines": [ + { + "bbox": [ + 140, + 396, + 183, + 403 + ], + "spans": [ + { + "bbox": [ + 140, + 396, + 183, + 403 + ], + "type": "text", + "content": "Original Image" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 408, + 480, + 453 + ], + "lines": [ + { + "bbox": [ + 130, + 408, + 480, + 453 + ], + "spans": [ + { + "bbox": [ + 130, + 408, + 480, + 453 + ], + "type": "text", + "content": "Fig. 7: Qualitative comparisons in the smile transferring benchmark. Our Reader can selectively manipulate only the target attributes without affecting other features or regions. The red boxes highlight the local details that should remain unchanged after editing. Please kindly zoom in for a better view." + } + ] + } + ], + "index": 43, + "angle": 0, + "type": "image_caption" + } + ], + "index": 31 + }, + { + "type": "image", + "bbox": [ + 192, + 338, + 249, + 396 + ], + "blocks": [ + { + "bbox": [ + 192, + 338, + 249, + 396 + ], + "lines": [ + { + "bbox": [ + 192, + 338, + 249, + 396 + ], + "spans": [ + { + "bbox": [ + 192, + 338, + 249, + 396 + ], + "type": "image", + "image_path": "3fd414cec35e5f5ef03a59f22a8b0ec6fd7ca011b28cbc52dfbc3f2fab03349c.jpg" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 205, + 396, + 235, + 403 + ], + "lines": [ + { + "bbox": [ + 205, + 396, + 235, + 403 + ], + "spans": [ + { + "bbox": [ + 205, + 396, + 235, + 403 + ], + "type": "text", + "content": "MaskGAN" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_caption" + } + ], + "index": 33 + }, + { + "type": "image", + "bbox": [ + 250, + 338, + 306, + 396 + ], + "blocks": [ + { + "bbox": [ + 250, + 338, + 306, + 396 + ], + "lines": [ + { + "bbox": [ + 250, + 338, + 306, + 396 + ], + "spans": [ + { + "bbox": [ + 250, + 338, + 306, + 396 + ], + "type": "image", + "image_path": "9757fecff33a3c356261405a6648fc98a3d69d84f70b30b9863245cd9e790551.jpg" + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 270, + 396, + 285, + 403 + ], + "lines": [ + { + "bbox": [ + 270, + 396, + 285, + 403 + ], + "spans": [ + { + "bbox": [ + 270, + 396, + 285, + 403 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_caption" + } + ], + "index": 35 + }, + { + "type": "image", + "bbox": [ + 307, + 338, + 364, + 396 + ], + "blocks": [ + { + "bbox": [ + 307, + 338, + 364, + 396 + ], + "lines": [ + { + "bbox": [ + 307, + 338, + 364, + 396 + ], + "spans": [ + { + "bbox": [ + 307, + 338, + 364, + 396 + ], + "type": "image", + "image_path": "db08d64c90318e1d6809f5705a42eb384e85cb713e68763988d8a106e6ae1166.jpg" + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 396, + 358, + 403 + ], + "lines": [ + { + "bbox": [ + 313, + 396, + 358, + 403 + ], + "spans": [ + { + "bbox": [ + 313, + 396, + 358, + 403 + ], + "type": "text", + "content": "Original Image" + } + ] + } + ], + "index": 38, + "angle": 0, + "type": "image_caption" + } + ], + "index": 37 + }, + { + "type": "image", + "bbox": [ + 365, + 338, + 422, + 396 + ], + "blocks": [ + { + "bbox": [ + 365, + 338, + 422, + 396 + ], + "lines": [ + { + "bbox": [ + 365, + 338, + 422, + 396 + ], + "spans": [ + { + "bbox": [ + 365, + 338, + 422, + 396 + ], + "type": "image", + "image_path": "a370999f17a3d1eee7d93661f343367582bdd7434fd1c5a22f2729da55f6ade2.jpg" + } + ] + } + ], + "index": 39, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 378, + 396, + 408, + 403 + ], + "lines": [ + { + "bbox": [ + 378, + 396, + 408, + 403 + ], + "spans": [ + { + "bbox": [ + 378, + 396, + 408, + 403 + ], + "type": "text", + "content": "MaskGAN" + } + ] + } + ], + "index": 40, + "angle": 0, + "type": "image_caption" + } + ], + "index": 39 + }, + { + "type": "image", + "bbox": [ + 424, + 338, + 481, + 396 + ], + "blocks": [ + { + "bbox": [ + 424, + 338, + 481, + 396 + ], + "lines": [ + { + "bbox": [ + 424, + 338, + 481, + 396 + ], + "spans": [ + { + "bbox": [ + 424, + 338, + 481, + 396 + ], + "type": "image", + "image_path": "3e9c9200f8e376f5881b55f4833de76afbfdb55a8dbae8b65785ad6f849a624d.jpg" + } + ] + } + ], + "index": 41, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 444, + 396, + 459, + 403 + ], + "lines": [ + { + "bbox": [ + 444, + 396, + 459, + 403 + ], + "spans": [ + { + "bbox": [ + 444, + 396, + 459, + 403 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 42, + "angle": 0, + "type": "image_caption" + } + ], + "index": 41 + }, + { + "bbox": [ + 130, + 471, + 480, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 471, + 480, + 495 + ], + "spans": [ + { + "bbox": [ + 130, + 471, + 480, + 495 + ], + "type": "text", + "content": "time. In the bottom row, we simultaneously manipulated both the masks and embeddings, facilitating significant change in expression." + } + ] + } + ], + "index": 44 + }, + { + "bbox": [ + 130, + 495, + 480, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 495, + 480, + 567 + ], + "spans": [ + { + "bbox": [ + 130, + 495, + 480, + 567 + ], + "type": "text", + "content": "To quantitatively measure Reader's image editing capabilities, we utilize the smile transferring benchmark introduced by MaskGAN [31], which requires converting the faces with neutral expressions into smiling faces. We compare our Reader with the MaskGAN [31], LocalEditing [9], InterFaceGAN [46] and EditGAN [34], which also provide results in this benchmark. We employ the following metrics to keep consistent with [34]." + } + ] + } + ], + "index": 45 + }, + { + "bbox": [ + 130, + 568, + 480, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 568, + 480, + 640 + ], + "spans": [ + { + "bbox": [ + 130, + 568, + 480, + 640 + ], + "type": "text", + "content": "(1) Semantic Correctness: Training a binary classifier to recognize the smile attribute and evaluating whether the edited faces exhibit a smiling expression. (2) Distribution-level Image Quality: Utilizing FID [24] and KID [2] to assess the quality of edited images. (3) Identity Preservation: Leveraging a pretrained ArcFace feature extraction network [12] to evaluate whether the face's identity is maintained post-editing." + } + ] + } + ], + "index": 46 + }, + { + "bbox": [ + 130, + 641, + 480, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 480, + 665 + ], + "type": "text", + "content": "To automate smile transferring without manual adjustments, we employ a similar approach to MaskGAN. Specifically, we select 400 pairs of neutral-" + } + ] + } + ], + "index": 47 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 214, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 214, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 214, + 102 + ], + "type": "text", + "content": "Wang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 140, + 157, + 471, + 259 + ], + "blocks": [ + { + "bbox": [ + 131, + 114, + 482, + 148 + ], + "lines": [ + { + "bbox": [ + 131, + 114, + 482, + 148 + ], + "spans": [ + { + "bbox": [ + 131, + 114, + 482, + 148 + ], + "type": "text", + "content": "Table 2: Quantitative results in the smile transferring benchmark. Top 1 and top 2 results are highlighted in red and blue, respectively. Our Reader significantly outperforms previous methods in the FID and ID Score metrics by a large margin." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 140, + 157, + 471, + 259 + ], + "lines": [ + { + "bbox": [ + 140, + 157, + 471, + 259 + ], + "spans": [ + { + "bbox": [ + 140, + 157, + 471, + 259 + ], + "type": "table", + "html": "
MetricAttribute Acc. (%)↑FID↓KID↓ID Score↑
MaskGAN [31]77.346.840.0200.4611
LocalEditing [9]26.041.260.0120.5823
LocalEditing - Encoding4Editing [49]41.848.280.0160.7295
InterFaceGAN [46]83.539.420.0100.7295
EditGAN [34]91.541.740.0130.7047
EditGAN+30 [34]85.840.830.0120.7452
Reader (ours)84.237.910.0110.8470
", + "image_path": "a4e0340e93eec76c469c53c0702be9fc604c49e1a587a1eb87ff7d70d39ea843.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 283, + 479, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 283, + 479, + 319 + ], + "spans": [ + { + "bbox": [ + 130, + 283, + 479, + 319 + ], + "type": "text", + "content": "expression and smiling-expression faces with the closest head position as the source and target images. Subsequently, we replace both the tokens corresponding to mouth and lips of the source image with those from the target image." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 319, + 482, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 319, + 482, + 464 + ], + "spans": [ + { + "bbox": [ + 130, + 319, + 482, + 464 + ], + "type": "text", + "content": "Results shown in Tab. 2 demonstrate that our Reader achieves excellent performance across all metrics. Specifically, our Reader significantly outperforms previous methods in the FID and ID Score metrics by a large margin. This indicates that the edited images produced by Reader exhibit higher quality while better preserving the ID features of the original images. This observation effectively highlights the inherent advantage of our method. By representing the image through a series of decoupled tokens, Reader can selectively manipulate only the target attributes without affecting other regions. To provide a more intuitive perception, we illustrate some examples of the original and edited image pairs in Fig. 7. It can be observed that our method effectively prevents changes in the areas which are not the editing target, while achieving editing goals. In contrast, the GAN-based baseline [31] is prone to inducing global changes." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 485, + 238, + 497 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 485, + 238, + 497 + ], + "spans": [ + { + "bbox": [ + 131, + 485, + 238, + 497 + ], + "type": "text", + "content": "4.3 Ablation Study" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 509, + 482, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 509, + 482, + 616 + ], + "spans": [ + { + "bbox": [ + 130, + 509, + 482, + 616 + ], + "type": "text", + "content": "In Tab. 3a, we conducted ablation studies to evaluate the effectiveness of some optional hyper-parameters and modules in the model architecture, where different configurations represent: Patch Size: The patch size of the ViT in Reader's encoder. Bidirectional: Whether to use bidirectional transformer block in the region decoder. Masked-Attn: Whether to employ the masked cross-attention layer in the region embedder. Results indicate that a ViT with a smaller patch size paired with the additional modules produces the best image fidelity. This demonstrates that our default settings can better preserve the fine-grained details in the image, thereby enhancing the quality of the output image." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "In Tab. 3b, we conducted ablation experiment to determine the granularity of pseudo-labels utilized in training. In this experiment, during preprocessing, we only removed the overlapping parts from the pseudo-labels and retained masks of all different sizes to train our Reader. Subsequently, we calculated the average" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 333, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 333, + 91, + 447, + 102 + ], + "type": "text", + "content": "Region-Native Autoencoder" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 153, + 146, + 350, + 205 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 479, + 137 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 479, + 137 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 479, + 137 + ], + "type": "text", + "content": "Table 3: Ablation studies on both model architecture, hyper parameter and data preprocessing. Our default setting is marked with gray." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 153, + 146, + 350, + 205 + ], + "lines": [ + { + "bbox": [ + 153, + 146, + 350, + 205 + ], + "spans": [ + { + "bbox": [ + 153, + 146, + 350, + 205 + ], + "type": "table", + "html": "
Patch SizeBidirectionalMasked-AttnrFID ↓
8 × 8XX98.39
✓X37.50
✓✓11.42
16 × 16✓✓61.64
", + "image_path": "7b434653f866d862758eacc0d979b5f9d75a1cda8b6b1a816381e3e5af577619.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 201, + 207, + 295, + 217 + ], + "lines": [ + { + "bbox": [ + 201, + 207, + 295, + 217 + ], + "spans": [ + { + "bbox": [ + 201, + 207, + 295, + 217 + ], + "type": "text", + "content": "(a) Module configurations." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 366, + 146, + 462, + 205 + ], + "blocks": [ + { + "bbox": [ + 366, + 146, + 462, + 205 + ], + "lines": [ + { + "bbox": [ + 366, + 146, + 462, + 205 + ], + "spans": [ + { + "bbox": [ + 366, + 146, + 462, + 205 + ], + "type": "table", + "html": "
Area ProportionL1 ↓
< 1/321.44
1/32 ~ 1/161.56
1/16 ~ 1/82.53
> 1/83.46
", + "image_path": "9c474bb8f9c8c00466395645759ed12454fcf2597093883b403b5a1ccea1b49d.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 352, + 207, + 461, + 217 + ], + "lines": [ + { + "bbox": [ + 352, + 207, + 461, + 217 + ], + "spans": [ + { + "bbox": [ + 352, + 207, + 461, + 217 + ], + "type": "text", + "content": "(b) Annotation configurations." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 232, + 481, + 306 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 232, + 481, + 306 + ], + "spans": [ + { + "bbox": [ + 130, + 232, + 481, + 306 + ], + "type": "text", + "content": "L1 distance of different-sized regions between the original images and the reconstructed images. The \"Area Proportion\" in Tab. 3b represents the proportion of the region's area relative to the entire image. The results indicate that when the region's area proportion is less than " + }, + { + "bbox": [ + 130, + 232, + 481, + 306 + ], + "type": "inline_equation", + "content": "\\frac{1}{16}" + }, + { + "bbox": [ + 130, + 232, + 481, + 306 + ], + "type": "text", + "content": ", the L1 distance stabilizes with changes in area. Therefore, in the formal experiments, we only retained pseudo-labels with area proportion smaller than " + }, + { + "bbox": [ + 130, + 232, + 481, + 306 + ], + "type": "inline_equation", + "content": "\\frac{1}{16}" + }, + { + "bbox": [ + 130, + 232, + 481, + 306 + ], + "type": "text", + "content": " for training." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 327, + 311, + 339 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 327, + 311, + 339 + ], + "spans": [ + { + "bbox": [ + 132, + 327, + 311, + 339 + ], + "type": "text", + "content": "5 Conclusion and Discussion" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 356, + 482, + 476 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 356, + 482, + 476 + ], + "spans": [ + { + "bbox": [ + 130, + 356, + 482, + 476 + ], + "type": "text", + "content": "In conclusion, this paper explores an innovative region-based visual token representation and presents Reader. Reader employs arbitrarily shaped and sized regions as the units of image tokens, which aligns better with the visual prior: the non-uniformity of information density and the naturally existing semantic boundaries. Through experiments, we demonstrate the following inherent advantages of this region-based tokenization method. Efficiency: This method has a high compression ratio and can achieve superior reconstruction fidelity while employing a significantly smaller number of tokens for each image. Disentanglement: The region-based tokens guarantees their disentanglement from one another, and further empower Reader with the ability to perform image editing." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 480, + 482, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 480, + 482, + 564 + ], + "spans": [ + { + "bbox": [ + 130, + 480, + 482, + 564 + ], + "type": "text", + "content": "Limitation and Feature Work As this paper focuses on exploring the feasibility and effectiveness of region-based tokenization, we adopted a minimal configuration in our experiments, manifested as employing the tiny-version transformer and training model on relatively small-sized images " + }, + { + "bbox": [ + 130, + 480, + 482, + 564 + ], + "type": "inline_equation", + "content": "(256\\times 256)" + }, + { + "bbox": [ + 130, + 480, + 482, + 564 + ], + "type": "text", + "content": ". Although Reader can produce already satisfactory results under such configurations, the parameter size and data resolution still limit its broader application. For example, some high-resolution images may exhibit a degree of blurring after editing." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 564, + 482, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 564, + 482, + 601 + ], + "spans": [ + { + "bbox": [ + 130, + 564, + 482, + 601 + ], + "type": "text", + "content": "In the future work, we will pursue scaling up both model parameters and the volume of data. Additionally, we aim to explore its application in various scenarios, such as interaction with LLMs and multimodal perception, generation." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": "Broader Impact Reader possesses the capability to serve as an image editor, yielding numerous interesting outcomes, but also opens up the possibility of real photo manipulation for nefarious purposes. However, with the recent advancements in watermarking techniques and manipulated images identifying techniques [17], this risk is avoidable." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 214, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 214, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 214, + 102 + ], + "type": "text", + "content": "Wang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 143, + 480, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 143, + 480, + 167 + ], + "spans": [ + { + "bbox": [ + 132, + 143, + 480, + 167 + ], + "type": "text", + "content": "This research was funded by the Fundamental Research Funds for the Central Universities (2024XKRC082) and the National NSF of China (No.U23A20314)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 190, + 197, + 201 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 190, + 197, + 201 + ], + "spans": [ + { + "bbox": [ + 133, + 190, + 197, + 201 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 217, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 138, + 217, + 481, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 217, + 481, + 251 + ], + "spans": [ + { + "bbox": [ + 138, + 217, + 481, + 251 + ], + "type": "text", + "content": "1. Anderson, P., He, X., Buehler, C., Teney, D., Johnson, M., Gould, S., Zhang, L.: Bottom-up and top-down attention for image captioning and visual question answering. In: CVPR. pp. 6077-6086 (2018)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 251, + 481, + 273 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 251, + 481, + 273 + ], + "spans": [ + { + "bbox": [ + 138, + 251, + 481, + 273 + ], + "type": "text", + "content": "2. Binkowski, M., Sutherland, D.J., Arbel, M., Gretton, A.: Demystifying MMD gans. In: ICLR. OpenReview.net (2018)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 274, + 481, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 274, + 481, + 296 + ], + "spans": [ + { + "bbox": [ + 138, + 274, + 481, + 296 + ], + "type": "text", + "content": "3. Brooks, T., Holynski, A., Efros, A.A.: Instructpix2pix: Learning to follow image editing instructions. In: CVPR. pp. 18392-18402. IEEE (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 297, + 481, + 329 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 297, + 481, + 329 + ], + "spans": [ + { + "bbox": [ + 138, + 297, + 481, + 329 + ], + "type": "text", + "content": "4. Burgess, C.P., Matthew, L., Watters, N., Kabra, R., Higgins, I., Botvinick, M.M., Lerchner, A.: Monet: Unsupervised scene decomposition and representation. CoRR abs/1901.11390 (2019)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 331, + 481, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 331, + 481, + 373 + ], + "spans": [ + { + "bbox": [ + 138, + 331, + 481, + 373 + ], + "type": "text", + "content": "5. Carion, N., Massa, F., Synnaeve, G., Usunier, N., Kirillov, A., Zagoruyko, S.: End-to-end object detection with transformers. In: Vedaldi, A., Bischof, H., Brox, T., Frahm, J. (eds.) ECCV. Lecture Notes in Computer Science, vol. 12346, pp. 213-229. Springer (2020)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 375, + 481, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 375, + 481, + 407 + ], + "spans": [ + { + "bbox": [ + 138, + 375, + 481, + 407 + ], + "type": "text", + "content": "6. Cheng, B., Misra, I., Schwing, A.G., Kirillov, A., Girdhar, R.: Masked-attention mask transformer for universal image segmentation. In: CVPR. pp. 1290–1299 (2022)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 408, + 481, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 408, + 481, + 441 + ], + "spans": [ + { + "bbox": [ + 138, + 408, + 481, + 441 + ], + "type": "text", + "content": "7. Cheng, B., Schwing, A.G., Kirillov, A.: Per-pixel classification is not all you need for semantic segmentation. In: Ranzato, M., Beygelzimer, A., Dauphin, Y.N., Liang, P., Vaughan, J.W. (eds.) NeurIPS. pp. 17864-17875 (2021)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 442, + 481, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 442, + 481, + 475 + ], + "spans": [ + { + "bbox": [ + 138, + 442, + 481, + 475 + ], + "type": "text", + "content": "8. Choi, Y., Choi, M., Kim, M., Ha, J., Kim, S., Choo, J.: Stargan: Unified generative adversarial networks for multi-domain image-to-image translation. In: CVPR. pp. 8789-8797. Computer Vision Foundation / IEEE Computer Society (2018)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 475, + 481, + 508 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 475, + 481, + 508 + ], + "spans": [ + { + "bbox": [ + 138, + 475, + 481, + 508 + ], + "type": "text", + "content": "9. Collins, E., Bala, R., Price, B., Süssstrunk, S.: Editing in style: Uncovering the local semantics of gans. In: CVPR. pp. 5770-5779. Computer Vision Foundation / IEEE (2020)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 509, + 481, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 509, + 481, + 531 + ], + "spans": [ + { + "bbox": [ + 138, + 509, + 481, + 531 + ], + "type": "text", + "content": "0. Couairon, G., Verbeek, J., Schwenk, H., Cord, M.: Diffedit: Diffusion-based semantic image editing with mask guidance. In: ICLR. OpenReview.net (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 532, + 481, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 532, + 481, + 564 + ], + "spans": [ + { + "bbox": [ + 138, + 532, + 481, + 564 + ], + "type": "text", + "content": "1. Crowson, K., Biderman, S., Kornis, D., Stander, D., Hallahan, E., Castricato, L., Raff, E.: Vqgan-clip: Open domain image generation and editing with natural language guidance. In: ECCV. pp. 88–105. Springer (2022)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 565, + 481, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 565, + 481, + 597 + ], + "spans": [ + { + "bbox": [ + 138, + 565, + 481, + 597 + ], + "type": "text", + "content": "2. Deng, J., Guo, J., Xue, N., Zafeiriou, S.: Arcface: Additive angular margin loss for deep face recognition. In: CVPR. pp. 4690-4699. Computer Vision Foundation / IEEE (2019)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 599, + 481, + 642 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 599, + 481, + 642 + ], + "spans": [ + { + "bbox": [ + 138, + 599, + 481, + 642 + ], + "type": "text", + "content": "3. Devlin, J., Chang, M., Lee, K., Toutanova, K.: BERT: pre-training of deep bidirectional transformers for language understanding. In: Burstein, J., Doran, C., Solorio, T. (eds.) NAACL-HLT. pp. 4171-4186. Association for Computational Linguistics (2019)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 643, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 643, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 643, + 481, + 665 + ], + "type": "text", + "content": "4. Ding, H., Liu, C., He, S., Jiang, X., Loy, C.C.: MeViS: A large-scale benchmark for video segmentation with motion expressions. In: ICCV (2023)" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 333, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 333, + 91, + 447, + 101 + ], + "type": "text", + "content": "Region-Native Autoencoder" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 666 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 481, + 160 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 481, + 160 + ], + "type": "text", + "content": "15. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., Uszkoreit, J., Houlsby, N.: An image is worth 16x16 words: Transformers for image recognition at scale. In: ICLR. OpenReview.net (2021)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 160, + 481, + 192 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 160, + 481, + 192 + ], + "spans": [ + { + "bbox": [ + 133, + 160, + 481, + 192 + ], + "type": "text", + "content": "16. Esser, P., Rombach, R., Ommer, B.: Taming transformers for high-resolution image synthesis. In: CVPR. pp. 12873-12883. Computer Vision Foundation / IEEE (2021)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 192, + 481, + 214 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 192, + 481, + 214 + ], + "spans": [ + { + "bbox": [ + 132, + 192, + 481, + 214 + ], + "type": "text", + "content": "17. Fernandez, P., Couairon, G., Jégou, H., Douze, M., Furon, T.: The stable signature: Rooting watermarks in latent diffusion models. CoRR abs/2303.15435 (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 214, + 481, + 246 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 214, + 481, + 246 + ], + "spans": [ + { + "bbox": [ + 132, + 214, + 481, + 246 + ], + "type": "text", + "content": "18. Gafni, O., Polyak, A., Ashual, O., Sheynin, S., Parikh, D., Taigman, Y.: Make-a-scene: Scene-based text-to-image generation with human priors. In: ECCV. pp. 89-106. Springer (2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 246, + 481, + 278 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 246, + 481, + 278 + ], + "spans": [ + { + "bbox": [ + 132, + 246, + 481, + 278 + ], + "type": "text", + "content": "19. Goel, V., Peruzzo, E., Jiang, Y., Xu, D., Sebe, N., Darrell, T., Wang, Z., Shi, H.: Pair-diffusion: Object-level image editing with structure-and-appearance paired diffusion models. CoRR abs/2303.17546 (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 278, + 481, + 332 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 278, + 481, + 332 + ], + "spans": [ + { + "bbox": [ + 132, + 278, + 481, + 332 + ], + "type": "text", + "content": "20. Gondal, M.W., Wuthrich, M., Miladinovic, D., Locatello, F., Breidt, M., Volchkov, V., Akpo, J., Bachem, O., Scholkopf, B., Bauer, S.: On the transfer of inductive bias from simulation to the real world: a new disentanglement dataset. In: Wallach, H.M., Larochelle, H., Beygelzimer, A., d'Alché-Buc, F., Fox, E.B., Garnett, R. (eds.) NeurIPS. pp. 15714-15725 (2019)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 332, + 481, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 332, + 481, + 365 + ], + "spans": [ + { + "bbox": [ + 132, + 332, + 481, + 365 + ], + "type": "text", + "content": "21. Goodfellow, I.J., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., Courville, A.C., Bengio, Y.: Generative adversarial nets. In: Ghahramani, Z., Welling, M., Cortes, C., Lawrence, N.D., Weinberger, K.Q. (eds.) NeurIPS (2014)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 365, + 481, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 365, + 481, + 397 + ], + "spans": [ + { + "bbox": [ + 132, + 365, + 481, + 397 + ], + "type": "text", + "content": "22. Härkönen, E., Hertzmann, A., Lehtinen, J., Paris, S.: Ganspace: Discovering interpretable GAN controls. In: Larochelle, H., Ranzato, M., Hadsell, R., Balcan, M., Lin, H. (eds.) NeurIPS (2020)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 397, + 481, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 397, + 481, + 418 + ], + "spans": [ + { + "bbox": [ + 132, + 397, + 481, + 418 + ], + "type": "text", + "content": "23. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: CVPR. pp. 770-778. IEEE Computer Society (2016)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 418, + 481, + 461 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 418, + 481, + 461 + ], + "spans": [ + { + "bbox": [ + 132, + 418, + 481, + 461 + ], + "type": "text", + "content": "24. Heusel, M., Ramsauer, H., Unterthiner, T., Nessler, B., Hochreiter, S.: Gans trained by a two time-scale update rule converge to a local nash equilibrium. In: Guyon, I., von Luxburg, U., Bengio, S., Wallach, H.M., Fergus, R., Vishwanathan, S.V.N., Garnett, R. (eds.) NeurIPS. pp. 6626-6637 (2017)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 461, + 481, + 483 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 461, + 481, + 483 + ], + "spans": [ + { + "bbox": [ + 132, + 461, + 481, + 483 + ], + "type": "text", + "content": "25. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. In: Larochelle, H., Ranzato, M., Hadsell, R., Balcan, M., Lin, H. (eds.) NeurIPS (2020)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 483, + 481, + 504 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 483, + 481, + 504 + ], + "spans": [ + { + "bbox": [ + 132, + 483, + 481, + 504 + ], + "type": "text", + "content": "26. Horé, A., Ziou, D.: Image quality metrics: PSNR vs. SSIM. In: ICPR. pp. 2366-2369. IEEE Computer Society (2010)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 504, + 481, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 504, + 481, + 536 + ], + "spans": [ + { + "bbox": [ + 132, + 504, + 481, + 536 + ], + "type": "text", + "content": "27. Huang, M., Mao, Z., Chen, Z., Zhang, Y.: Towards accurate image coding: Improved autoregressive image generation with dynamic vector quantization. In: CVPR. pp. 22596-22605 (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 536, + 481, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 536, + 481, + 568 + ], + "spans": [ + { + "bbox": [ + 132, + 536, + 481, + 568 + ], + "type": "text", + "content": "28. Isola, P., Zhu, J., Zhou, T., Efros, A.A.: Image-to-image translation with conditional adversarial networks. In: CVPR. pp. 5967-5976. IEEE Computer Society (2017)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 568, + 481, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 568, + 481, + 601 + ], + "spans": [ + { + "bbox": [ + 132, + 568, + 481, + 601 + ], + "type": "text", + "content": "29. Kim, H., Mnih, A.: Disentangling by factorising. In: Dy, J.G., Krause, A. (eds.) ICML. Proceedings of Machine Learning Research, vol. 80, pp. 2654-2663. PMLR (2018)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 601, + 481, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 601, + 481, + 633 + ], + "spans": [ + { + "bbox": [ + 132, + 601, + 481, + 633 + ], + "type": "text", + "content": "30. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W., Dollár, P., Girshick, R.B.: Segment anything. CoRR abs/2304.02643 (2023)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 633, + 481, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 633, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 633, + 481, + 666 + ], + "type": "text", + "content": "31. Lee, C., Liu, Z., Wu, L., Luo, P.: Maskgan: Towards diverse and interactive facial image manipulation. In: CVPR. pp. 5548-5557. Computer Vision Foundation / IEEE (2020)" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 214, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 214, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 214, + 102 + ], + "type": "text", + "content": "Wang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 666 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 159 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 159 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 159 + ], + "type": "text", + "content": "32. Lee, D., Kim, C., Kim, S., Cho, M., Han, W.: Autoregressive image generation using residual quantization. In: CVPR. pp. 11513-11522. IEEE (2022). https://doi.org/10.1109/CVPR52688.2022.01123, https://doi.org/10.1109/CVPR52688.2022.01123" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 160, + 481, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 160, + 481, + 182 + ], + "spans": [ + { + "bbox": [ + 130, + 160, + 481, + 182 + ], + "type": "text", + "content": "33. Li, Y., Mao, H., Girshick, R., He, K.: Exploring plain vision transformer backbones for object detection. In: ECCV. pp. 280-296. Springer (2022)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 182, + 481, + 214 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 182, + 481, + 214 + ], + "spans": [ + { + "bbox": [ + 132, + 182, + 481, + 214 + ], + "type": "text", + "content": "34. Ling, H., Kreis, K., Li, D., Kim, S.W., Torralba, A., Fidler, S.: Editgan: High-precision semantic image editing. In: Ranzato, M., Beygelzimer, A., Dauphin, Y.N., Liang, P., Vaughan, J.W. (eds.) NeurIPS. pp. 16331-16345 (2021)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 214, + 481, + 236 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 214, + 481, + 236 + ], + "spans": [ + { + "bbox": [ + 132, + 214, + 481, + 236 + ], + "type": "text", + "content": "35. Liu, C., Ding, H., Jiang, X.: GRES: Generalized referring expression segmentation. In: CVPR (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 236, + 481, + 257 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 236, + 481, + 257 + ], + "spans": [ + { + "bbox": [ + 132, + 236, + 481, + 257 + ], + "type": "text", + "content": "36. Liu, C., Jiang, X., Ding, H.: Primitivenet: decomposing the global constraints for referring segmentation. Visual Intelligence 2(1), 16 (2024)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 257, + 481, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 257, + 481, + 300 + ], + "spans": [ + { + "bbox": [ + 132, + 257, + 481, + 300 + ], + "type": "text", + "content": "37. Locatello, F., Weissenborn, D., Unterthiner, T., Mahendran, A., Heigold, G., Uszkoreit, J., Dosovitskiy, A., Kipf, T.: Object-centric learning with slot attention. In: Larochelle, H., Ranzato, M., Hadsell, R., Balcan, M., Lin, H. (eds.) NeurIPS (2020)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 300, + 481, + 332 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 300, + 481, + 332 + ], + "spans": [ + { + "bbox": [ + 132, + 300, + 481, + 332 + ], + "type": "text", + "content": "38. Ning, J., Li, C., Zhang, Z., Wang, C., Geng, Z., Dai, Q., He, K., Hu, H.: All in tokens: Unifying output space of visual tasks via soft token. In: ICCV. pp. 19900-19910 (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 332, + 481, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 332, + 481, + 365 + ], + "spans": [ + { + "bbox": [ + 132, + 332, + 481, + 365 + ], + "type": "text", + "content": "39. van den Oord, A., Vinyals, O., Kavukcuoglu, K.: Neural discrete representation learning. In: Guyon, I., von Luxburg, U., Bengio, S., Wallach, H.M., Fergus, R., Vishwanathan, S.V.N., Garnett, R. (eds.) NeurIPS. pp. 6306-6315 (2017)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 365, + 481, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 365, + 481, + 407 + ], + "spans": [ + { + "bbox": [ + 132, + 365, + 481, + 407 + ], + "type": "text", + "content": "40. Pan, X., Tewari, A., Leimkuhler, T., Liu, L., Meka, A., Theobalt, C.: Drag your GAN: interactive point-based manipulation on the generative image manifold. In: Brunvand, E., Sheffer, A., Wimmer, M. (eds.) SIGGRAPH. pp. 78:1-78:11. ACM (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 407, + 481, + 440 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 407, + 481, + 440 + ], + "spans": [ + { + "bbox": [ + 132, + 407, + 481, + 440 + ], + "type": "text", + "content": "41. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: ICML. pp. 8748-8763. PMLR (2021)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 440, + 481, + 461 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 440, + 481, + 461 + ], + "spans": [ + { + "bbox": [ + 132, + 440, + 481, + 461 + ], + "type": "text", + "content": "42. Radford, A., Narasimhan, K., Salimans, T., Sutskever, I., et al.: Improving language understanding by generative pre-training (2018)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 461, + 481, + 494 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 461, + 481, + 494 + ], + "spans": [ + { + "bbox": [ + 132, + 461, + 481, + 494 + ], + "type": "text", + "content": "43. Reed, S.E., Zhang, Y., Zhang, Y., Lee, H.: Deep visual analogy-making. In: Cortes, C., Lawrence, N.D., Lee, D.D., Sugiyama, M., Garnett, R. (eds.) NeurIPS. pp. 1252-1260 (2015)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 494, + 481, + 515 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 494, + 481, + 515 + ], + "spans": [ + { + "bbox": [ + 132, + 494, + 481, + 515 + ], + "type": "text", + "content": "44. Ren, S., He, K., Girshick, R., Sun, J.: Faster r-cnn: Towards real-time object detection with region proposal networks. NeurIPS 28 (2015)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 515, + 481, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 515, + 481, + 547 + ], + "spans": [ + { + "bbox": [ + 132, + 515, + 481, + 547 + ], + "type": "text", + "content": "45. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: CVPR. pp. 10674-10685. IEEE (2022)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 547, + 481, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 547, + 481, + 568 + ], + "spans": [ + { + "bbox": [ + 132, + 547, + 481, + 568 + ], + "type": "text", + "content": "46. Shen, Y., Yang, C., Tang, X., Zhou, B.: Interfacegan: Interpreting the disentangled face representation learned by gans. IEEE TPAMI 44(4), 2004-2018 (2022)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 568, + 481, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 568, + 481, + 601 + ], + "spans": [ + { + "bbox": [ + 132, + 568, + 481, + 601 + ], + "type": "text", + "content": "47. Shuai, X., Ding, H., Ma, X., Tu, R., Jiang, Y.G., Tao, D.: A survey of multimodal-guided image editing with text-to-image diffusion models. arXiv preprint arXiv:2406.14555 (2024)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 601, + 481, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 601, + 481, + 643 + ], + "spans": [ + { + "bbox": [ + 132, + 601, + 481, + 643 + ], + "type": "text", + "content": "48. Sohl-Dickstein, J., Weiss, E.A., Maheswaranathan, N., Ganguli, S.: Deep unsupervised learning using nonequilibrium thermodynamics. In: Bach, F.R., Blei, D.M. (eds.) ICML. JMLR Workshop and Conference Proceedings, vol. 37, pp. 2256-2265. JMLR.org (2015)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 643, + 481, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 643, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 643, + 481, + 666 + ], + "type": "text", + "content": "49. Tov, O., Alaluf, Y., Nitzan, Y., Patashnik, O., Cohen-Or, D.: Designing an encoder for stylegan image manipulation. ACM TOG 40(4), 133:1-133:14 (2021)" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 333, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 333, + 91, + 447, + 101 + ], + "type": "text", + "content": "Region-Native Autoencoder" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 456 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 481, + 161 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 481, + 161 + ], + "type": "text", + "content": "50. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. In: Guyon, I., von Luxburg, U., Bengio, S., Wallach, H.M., Fergus, R., Vishwanathan, S.V.N., Garnett, R. (eds.) NeurIPS. pp. 5998-6008 (2017)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 162, + 481, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 162, + 481, + 194 + ], + "spans": [ + { + "bbox": [ + 132, + 162, + 481, + 194 + ], + "type": "text", + "content": "51. Wu, B., Xu, C., Dai, X., Wan, A., Zhang, P., Yan, Z., Tomizuka, M., Gonzalez, J., Keutzer, K., Vajda, P.: Visual transformers: Token-based image representation and processing for computer vision. arXiv preprint arXiv:2006.03677 (2020)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 194, + 481, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 194, + 481, + 226 + ], + "spans": [ + { + "bbox": [ + 132, + 194, + 481, + 226 + ], + "type": "text", + "content": "52. Wu, R., Zhang, G., Lu, S., Chen, T.: Cascade EF-GAN: progressive facial expression editing with local focuses. In: CVPR. pp. 5020-5029. Computer Vision Foundation / IEEE (2020)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 227, + 481, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 227, + 481, + 258 + ], + "spans": [ + { + "bbox": [ + 132, + 227, + 481, + 258 + ], + "type": "text", + "content": "53. Xu, J., Mello, S.D., Liu, S., Byeon, W., Breuel, T.M., Kautz, J., Wang, X.: Groupvit: Semantic segmentation emerges from text supervision. In: CVPR. pp. 18113-18123. IEEE (2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 259, + 481, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 259, + 481, + 303 + ], + "spans": [ + { + "bbox": [ + 132, + 259, + 481, + 303 + ], + "type": "text", + "content": "54. Yang, B., Gu, S., Zhang, B., Zhang, T., Chen, X., Sun, X., Chen, D., Wen, F.: Paint by example: Exemplar-based image editing with diffusion models. In: CVPR. pp. 18381-18391. IEEE (2023). https://doi.org/10.1109/CVPR52729.2023.01763, https://doi.org/10.1109/CVPR52729.2023.01763" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 304, + 481, + 325 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 304, + 481, + 325 + ], + "spans": [ + { + "bbox": [ + 132, + 304, + 481, + 325 + ], + "type": "text", + "content": "55. Yang, T., Wang, Y., Lu, Y., Zheng, N.: Visual concepts tokenization. In: NeurIPS (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 326, + 481, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 326, + 481, + 357 + ], + "spans": [ + { + "bbox": [ + 132, + 326, + 481, + 357 + ], + "type": "text", + "content": "56. Yu, J., Li, X., Koh, J.Y., Zhang, H., Pang, R., Qin, J., Ku, A., Xu, Y., Baldridge, J., Wu, Y.: Vector-quantized image modeling with improved VQGAN. In: ICLR. OpenReview.net (2022)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 358, + 481, + 379 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 358, + 481, + 379 + ], + "spans": [ + { + "bbox": [ + 132, + 358, + 481, + 379 + ], + "type": "text", + "content": "57. Zellers, R., Yatskar, M., Thomson, S., Choi, Y.: Neural motifs: Scene graph parsing with global context. In: CVPR. pp. 5831-5840 (2018)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 380, + 481, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 380, + 481, + 413 + ], + "spans": [ + { + "bbox": [ + 132, + 380, + 481, + 413 + ], + "type": "text", + "content": "58. Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual metric. In: CVPR. pp. 586-595. Computer Vision Foundation / IEEE Computer Society (2018)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 414, + 481, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 414, + 481, + 435 + ], + "spans": [ + { + "bbox": [ + 132, + 414, + 481, + 435 + ], + "type": "text", + "content": "59. Zhong, Y., Wang, L., Chen, J., Yu, D., Li, Y.: Comprehensive image captioning via scene graph decomposition. In: ECCV. pp. 211-229. Springer (2020)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 435, + 481, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 435, + 481, + 456 + ], + "spans": [ + { + "bbox": [ + 132, + 435, + 481, + 456 + ], + "type": "text", + "content": "60. Zhu, J., Yang, C., Shen, Y., Shi, Z., Zhao, D., Chen, Q.: Linkgan: Linking GAN latents to pixels for controllable image synthesis. CoRR abs/2301.04604 (2023)" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 214, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 214, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 214, + 102 + ], + "type": "text", + "content": "Wang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Region-aware Distribution Contrast_ A Novel Approach to Multi-Task Partially Supervised Learning/b09b7220-34e4-48d5-bc47-6cde62cb650d_content_list.json b/2024/Region-aware Distribution Contrast_ A Novel Approach to Multi-Task Partially Supervised Learning/b09b7220-34e4-48d5-bc47-6cde62cb650d_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..bc9f15ab406190211f978d2f762ff9d6007d7bef --- /dev/null +++ b/2024/Region-aware Distribution Contrast_ A Novel Approach to Multi-Task Partially Supervised Learning/b09b7220-34e4-48d5-bc47-6cde62cb650d_content_list.json @@ -0,0 +1,1552 @@ +[ + { + "type": "text", + "text": "Region-aware Distribution Contrast: A Novel Approach to Multi-Task Partially Supervised Learning", + "text_level": 1, + "bbox": [ + 259, + 141, + 741, + 209 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Meixuan Li, Tianyu Li, Guoqing Wang*, Peng Wang, Yang Yang, and Jie Zou", + "bbox": [ + 354, + 237, + 648, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "University of Electronic Science and Technology of China (UESTC)", + "bbox": [ + 276, + 279, + 725, + 292 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "limeixuan0801@126.com, {cosmos.yu, p.wang6}@@hotmail.com,", + "bbox": [ + 282, + 295, + 715, + 306 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{gqwang0420, yang.yang, jie.zou}@uestc.edu.cn", + "bbox": [ + 328, + 309, + 676, + 321 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. In this study, we address the intricate challenge of multi-task dense prediction, encompassing tasks such as semantic segmentation, depth estimation, and surface normal estimation, particularly when dealing with partially annotated data (MTPSL). The complexity arises from the absence of complete task labels for each training image. Given the inter-related nature of these pixel-wise dense tasks, our focus is on mining and capturing cross-task relationships. Existing solutions typically rely on learning global image representations for global cross-task image matching, imposing constraints that, unfortunately, sacrifice the finer structures within the images. Attempting local matching as a remedy faces hurdles due to the lack of precise region supervision, making local alignment a challenging endeavor. The introduction of Segment Anything Model (SAM) sheds light on addressing local alignment challenges by providing free and high-quality solutions for region detection. Leveraging SAM-detected regions, the subsequent challenge lies in aligning the representations within these regions. Diverging from conventional methods that directly learn a monolithic image representation, our proposal involves modeling region-wise representations using Gaussian Distributions. Aligning these distributions between corresponding regions from different tasks imparts higher flexibility and capacity to capture intra-region structures, accommodating a broader range of tasks. This innovative approach significantly enhances our ability to effectively capture cross-task relationships, resulting in improved overall performance in partially supervised multi-task dense prediction scenarios. Extensive experiments conducted on two widely used benchmarks underscore the superior effectiveness of our proposed method, showcasing state-of-the-art performance even when compared to fully supervised methods.", + "bbox": [ + 261, + 378, + 738, + 751 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "https://github.com/HereNowL/Region-aware-Distribution-Contrast", + "bbox": [ + 261, + 752, + 743, + 763 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Multi-task Learning $\\cdot$ Partially Supervised Learning $\\cdot$ Scene Understanding $\\cdot$ Contrastive Learning", + "bbox": [ + 261, + 781, + 740, + 809 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Corresponding author.", + "bbox": [ + 218, + 825, + 383, + 839 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/3e8368e9e8842c8887bc6a48ae776131aed8eca9183b0ff8e41cb6dc82697476.jpg", + "image_caption": [ + "(a) Contrastive Multiview Coding" + ], + "image_footnote": [], + "bbox": [ + 261, + 143, + 491, + 250 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/e41de6a7d875d566b8ffb340ef64c57013b9eaeab9ee834916b9fe09545c8a6e.jpg", + "image_caption": [ + "(b) Region-wise Distribution Contrast", + "Fig. 1: Multiview consistency and region consistency. (a) Illustration of contrastive multiview consistency [38]. (b) Illustration of region consistency, where $p_{i,s}$ and $p_{j,s}$ represent the region distribution of semantic segmentation, $p_{i,d}$ and $p_{j,d}$ denotes the region distribution of depth estimation." + ], + "image_footnote": [], + "bbox": [ + 501, + 143, + 743, + 247 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 357, + 375, + 372 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "With the rapid advancement of deep learning, dense prediction tasks, including semantic segmentation, depth estimation, and surface normal estimation, have witnessed remarkable progress [11, 16, 30, 34, 47]. Considering the inherent interdependence among these dense prediction tasks [45, 46], there has been a growing interest in employing unified multi-task learning networks to jointly tackle different dense prediction tasks [7, 22, 28, 29, 31, 32, 39, 43, 50]. In contrast to inefficient single-task learning networks, multi-task learning networks effectively acquire shared features, enabling them to capture more generalizable information across various tasks while avoiding redundant training [39, 43, 50].", + "bbox": [ + 212, + 386, + 787, + 523 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recent advancements in multi-task dense prediction methods have primarily focused on two crucial aspects. The first aspect is dedicated to designing intricate network architectures to achieve effective multi-task learning [29, 32, 39, 43, 50], while the second aspect is devoted to devising optimal balancing strategies for loss functions in multi-task learning [7, 14, 22, 28, 29], aiming to mitigate the negative transfer across tasks. These directions mainly focus on fully supervised multi-task learning, where complete task labels are available for all training samples. However, acquiring pixel-level labels for dense prediction tasks is resource-intensive, particularly with multiple visual tasks. It is common to encounter scenarios where labels for certain tasks are missing or unreliable in some training samples [25], leading to a new research direction known as multi-task partially supervised learning (MTPSL).", + "bbox": [ + 212, + 523, + 787, + 704 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Since not all task labels are available for each training image in MTPSL, it is natural to formulate a multi-task learning framework which can leverage the implicit relationship between tasks with and without explicit labels. Dense prediction tasks, such as semantic segmentation, depth estimation, and normal prediction, exhibit a complementary co-existence relationship [37,45,46], where each task can serve as auxiliary information for others, offering valuable insights reciprocally. While existing approaches map predictions into a joint vector-wise semantic space for task alignment [25], we argue that this overlooks the underlying structural information inherent in the original image. A global vector", + "bbox": [ + 212, + 704, + 787, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "M. Li et al.", + "bbox": [ + 271, + 114, + 349, + 126 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "might be insufficient to characterize the information in the entire scene, limiting its effectiveness in constraining cross-task relationships. Local matching, as shown in Fig. 1, with its intrinsic local properties, serves as a potential remedy. Nonetheless, the absence of precise region supervision presents a fundamental obstacle, thereby intensifying the challenges associated with its implementation.", + "bbox": [ + 212, + 146, + 784, + 222 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To tackle the outlined challenges, we introduce region-based cross-task alignment by leveraging the robust region segmentation capabilities inherent in Segment Anything Model (SAM) models. Utilizing well-segmented image regions facilitates the abstraction of region-level information, enabling the addition of cross-task consistency constraints. However, a subsequent crucial challenge arises in effectively representing these regional features to ensure the consistency and accuracy of cross-task alignment. In contrast to conventional methods that directly learn a unified image representation, we employ Gaussian distributions to model region representations, enhancing flexibility and improving the ability to capture intra-region structures. By adapting Gaussian distributions between corresponding regions from different tasks, our approach can be more broadly applicable to diverse task types, such as pixel-level classification tasks (e.g., semantic segmentation) and pixel-level regression tasks (e.g., depth estimation). Subsequently, we introduce distributional contrastive learning to explicitly perform region-wise cross-task alignment. This approach serves as a robust constraint, offering a meaningful way to mine relationships between tasks and enhance dense prediction results, particularly for those tasks lacking pixel-wise labels.", + "bbox": [ + 212, + 222, + 785, + 478 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In summary, our work masks the following contributions:", + "bbox": [ + 238, + 479, + 651, + 493 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We tackle the challenge of multi-task partially supervised (MTPSL) dense prediction from a fresh perspective. Our approach involves extracting cross-task local alignment through the utilization of SAM's easily obtainable local regions. This innovative strategy has demonstrated notable in alleviating label shortages.", + "- We propose a novel region distribution contrast method for local alignment, which offers increased flexibility, robustness, and wide applicability across multiple tasks.", + "- We validate the effectiveness of our proposed method through extensive experiments conducted on two widely used benchmark datasets. Notably, our approach achieves state-of-the-art performance in partially supervised learning, while also showing great potential in fully supervised learning." + ], + "bbox": [ + 223, + 501, + 782, + 679 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 703, + 387, + 719 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Multi-task Supervised Learning. Multi-task learning utilizes a single model to simultaneously handle multiple visual tasks, offering advantages such as faster inference speed and more efficient utilization of input data. Some approaches [13, 32, 48] focus on designing the encoder of multi-task network, while other approaches pay more attention to designing complex decoders to generate task-specific predictions by utilizing shared features [2, 39, 43, 50]. Uncertainty [22] obtain task loss weights by considering the heteroscedastic uncertainty", + "bbox": [ + 212, + 734, + 785, + 840 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "Region-aware Distribution Contrast", + "bbox": [ + 491, + 114, + 730, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "of each task, Grad-Norm [7] directly adjusts gradients to balance task losses, and DWA [14] learns time-varying average task weights by considering the loss variation rate of each task. It is important to note that these methods operate under a significant and strong assumption, namely that they are developed and studied based on the availability of labels for all tasks across all images.", + "bbox": [ + 212, + 146, + 787, + 223 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Multi-task Semi-supervised Learning and Partially-supervised Learning. Learning multi-task models on fully annotated data requires a large-scale labeled dataset, and the cost of collecting sufficient labeled data can be high. Therefore, two common scenarios are worthy of note in multi-task learning. One is semi-supervised multi-task learning, where the dataset consists of limited annotation information for all tasks and a large amount of unlabeled data. A bunch of methods have been proposed for semi-supervised multi-task learning [8,18,19,24,27]. In recent works [8,18,19,24], regularization terms are applied to unlabeled samples from each task to encourage consistent predictions when the input is perturbed. Another scenario involves partial-supervision multi-task learning, where the dataset lacks sufficient labels for each task, and not all tasks have labels for every image. Li et al. [25] maps tasks into high-dimensional vectors for cross-task alignment. However, vector representations are insufficient and do not capture local-level alignment. In this paper, we propose leveraging distributions as a substitute for vector representations of task features and employing contrastive learning to achieve region-level alignment across tasks.", + "bbox": [ + 212, + 226, + 802, + 468 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Contrastive Learning for Vision. Recently, contrastive learning has made significant advancements in unsupervised learning [5,6,15,33,42]. DenseCL [40] and RegionContrast [17] have been utilized for semantic segmentation tasks using contrastive learning at the pixel and region levels respectively, achieving excellent results. For depth estimation, WCL [12] transformed depth from continuous values to discrete values and constructed contrastive losses based on windows to form positive and negative sample pairs. CMC [38] was the first to propose that different viewpoints of the same image should be mapped to nearby positions in a high-dimensional space, while different images should be mapped to distant positions, thus ensuring multiview consistency. Inspired by CMC, we believe that multiple viewpoints within the same local region should also exhibit consistency, while differing from other regions. In this paper, we leverage this idea to achieve region-level contrast for MTPSL.", + "bbox": [ + 212, + 473, + 789, + 670 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Gaussian Distribution. Recently, Gaussian distribution has been widely employed in computer vision to characterize the distribution of features [3, 4, 20, 26, 41, 49]. GMMSeg [26] utilized the Expectation-Maximization algorithm to construct Gaussian Mixture Models for each class in semantic segmentation, capturing class-conditional densities. AGMM [41] constructed an adaptive Gaussian Mixture Model for sparse annotation in semantic segmentation by incorporating labeled pixels and their similar unlabeled counterparts. DUL [3] employed mean and variance to characterize the distribution of faces for face recognition. Gaussian distributions have the capability to probabilistically describe the distribution characteristics of different visual tasks. However, to the best of our knowledge, Gaussian distributions have not been applied in MTPSL.", + "bbox": [ + 212, + 672, + 789, + 843 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "M. Li et al.", + "bbox": [ + 271, + 114, + 349, + 127 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/e0634edcee03b66869a58f843e1e3be0b743524f41a4e2cfab06a731939c7ceb.jpg", + "image_caption": [ + "Fig. 2: Illustration of region-aware distribution contrast learning method for MTPSL. During training, supervised constraints $L_{Sup}$ are applied to the annotated task $s$ . For task $s$ and unlabelled task $t$ , the $a_{\\theta}$ map the true label $y_{s}$ and the prediction $\\hat{y}_{t}$ in the high-dimensional space respectively and then model the region of task-specific features extracted using SAM as a Gaussian distribution. Contrastive learning is then employed to minimize the distance between distributions of the same region across different tasks while maximizing the distance to distributions of other regions." + ], + "image_footnote": [], + "bbox": [ + 263, + 142, + 707, + 316 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3 Method", + "text_level": 1, + "bbox": [ + 215, + 452, + 330, + 467 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1 Preliminaries", + "text_level": 1, + "bbox": [ + 215, + 482, + 372, + 496 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Problem Setup. Considering the problem of multi-task learning involving $K$ tasks, where $K \\geq 2$ . A training dataset $S = \\{x_{i}\\}_{i=1}^{N}$ is given with $N$ partially labelled samples, indicating that for each training sample $x_{i}$ , only a subset of the $K$ tasks are provided true labels. Let $P_{i}$ represents the number of labeled tasks for $x_{i}$ and $Q_{i}$ represents the number of unlabeled tasks, such that $P_{i} + Q_{i} = K$ . When $P_{i}$ equals $K$ for all $x_{i} \\in S$ , it indicates fully supervised multi-task learning. Conversely, when $Q_{i}$ equals $K$ for all $x_{i} \\in S$ , it implies that no task labels are available, leading to unsupervised multi-task learning. In this paper, we focus on addressing partially supervised multi-task learning, where each training image $x_{i}$ can obtain labels for at least one task ( $P_{i} \\geq 1$ ).", + "bbox": [ + 212, + 507, + 787, + 659 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2 Overview of Our Method", + "text_level": 1, + "bbox": [ + 215, + 680, + 470, + 694 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As shown in Fig. 2, our method focuses on utilizing Gaussian distributions to represent features in local regions for contrastive learning, achieving better distribution consistency across tasks at the region-level for MTPSL. Assuming that task $s$ has label while task $t$ does not, our objective is to achieve alignment between the true label $y_{s}$ of task $s$ and the prediction $\\hat{y}_t$ of task $t$ at the region level using contrastive learning.", + "bbox": [ + 212, + 703, + 787, + 794 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In the training stage, to accurately and efficiently obtain region-level features for prediction, we simultaneously feed the input image to both the multi-task learning (MTL) network and the pre-trained SAM. Predictions are obtained", + "bbox": [ + 212, + 795, + 787, + 839 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "Region-aware Distribution Contrast", + "bbox": [ + 491, + 114, + 730, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/fbb93ca70c827e550da39fb85b3f29a8a2c26074e15c8d06cddb311332949774.jpg", + "image_caption": [ + "Fig. 3: Illustration of region-level cross-task consistency. Initially, SAM predictions are employed to extract regions from the features of the particular task. Following that, these regions are modeled as Gaussian distributions. Finally, contrastive learning is utilized to minimize the distance between regions of the same region across different tasks and maximize the distance to regions of other regions." + ], + "image_footnote": [], + "bbox": [ + 277, + 143, + 728, + 300 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "from both MTL network and SAM. For tasks with labeled data, we apply fully-supervised loss $L_{Sup}$ as explicit training supervision. For tasks without label, an auxiliary network $a_{\\theta}$ maps the unlabelled task prediction $\\hat{y}_t$ and the annotated task label $y_s$ to the same high-dimensional feature map space, and use $L_{RC}$ to achieve cross-task consistency at the region level. Specifically, We first extract region-wise features in $\\hat{y}_t$ and $y_s$ based on the segmented regions identified by SAM, then model them as Gaussian distributions. After acquiring the Gaussian distributions for features in each region of each task, we utilize contrastive learning to minimize the distance between the Gaussian distributions of features in the same region across different tasks and maximize the distance from the Gaussian distributions of features from different regions. Finally, the optimization of our method can be defined as:", + "bbox": [ + 212, + 414, + 787, + 594 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nL o s s = L _ {S u p} + L _ {R C}. \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 424, + 609, + 785, + 626 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In the inference stage, as illustrated in Fig. 2, the input image is fed only into the MTL network to generate predictions for multiple tasks, notably, with SAM no longer being utilized.", + "bbox": [ + 212, + 637, + 787, + 683 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.3 Region-aware Contrastive Learning", + "text_level": 1, + "bbox": [ + 214, + 707, + 553, + 722 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this section, we explain the implementation details to achieve region-level cross-task consistency.", + "bbox": [ + 212, + 733, + 785, + 762 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Region-level Features Extraction. To realize the region-level cross-task consistency, it is essential to informatively choose regions for feature extraction. Considering the robust capabilities of SAM for image segmentation [23], we employ it to generate the regions for each training sample, leveraging its ability to provide valuable prior knowledge. As shown in Fig. 2, to obtain fine-grained", + "bbox": [ + 212, + 763, + 787, + 839 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "M. Li et al.", + "bbox": [ + 271, + 114, + 349, + 126 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "semantics and fine-grained edges of regions, we fed the image into MTL network and pre-trained SAM simultaneously. SAM performs fine-grained segmentation of each object in the image, and its output masks contain the ID of each area and its corresponding position in the image. Inspired by [21], for convenient region-wise feature extraction in task-specific feature maps, we compute the predicted masks of SAM as a grayscale image, where each pixel value corresponds to the area ID returned by SAM, indicating its category. Leveraging the grayscale masks of SAM, we split the image into different regions and extract region-wise features within feature maps of different tasks.", + "bbox": [ + 212, + 146, + 782, + 281 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Gaussian Distribution Modeling. After obtaining the region-wise features for each task, we consider modeling the region-wise features as Gaussian distributions. The utilization of Gaussian distributions represents the features of a region in a probabilistic manner, which provides a more comprehensive depiction of the region's variations and compensates for the shortcomings of pixel-level hard alignment. Specifically, we model each region $r_i$ from the task-specific feature map $f_t$ as a Gaussian distribution:", + "bbox": [ + 212, + 282, + 784, + 388 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\np _ {i, t} = p \\left(r _ {i} \\mid f _ {t}\\right) = N \\left(r _ {i}; \\boldsymbol {\\mu} _ {i, t}, \\boldsymbol {\\Sigma} _ {i, t}\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 380, + 398, + 784, + 415 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $\\pmb{\\mu}_{i,t}$ and $\\pmb{\\Sigma}_{i,t}$ represent mean and covariance matrix respectively, calculated from the region $r_i$ in the feature map $f_{t}$ , as shown in Fig. 3.", + "bbox": [ + 212, + 424, + 782, + 454 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Region Distribution Contrast. Once we have obtained the Gaussian distributions for each region, our objective is to utilize contrastive learning to minimize the distance between Gaussian distributions of the same region across different tasks and maximize the distance between Gaussian distributions of different regions. Hence, the region-wise Gaussian distribution contrastive loss for the region distribution $p_{i,t}$ can be defined as follows:", + "bbox": [ + 212, + 455, + 784, + 544 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\nL _ {i, t} ^ {N C E} = - \\sum_ {k _ {+}} \\log \\frac {\\operatorname {s i m} \\left(p _ {i , t} , k _ {+}\\right)}{\\operatorname {s i m} \\left(p _ {i , t} , k _ {+}\\right) + \\sum_ {k _ {-}} \\operatorname {s i m} \\left(p _ {i , t} , k _ {-}\\right)}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 307, + 556, + 784, + 589 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {s i m} \\left(p _ {i, t}, k\\right) = \\exp \\left(- W _ {\\text {d i s t a n c e}} \\left(p _ {i, t}, k\\right) / \\tau\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 357, + 606, + 784, + 619 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $k_{+} \\in \\{p_{i,s}\\}_{s \\neq t}$ , $k_{-} \\in \\{p_{j,*}\\}_{j \\neq i}$ , * denotes any task, sim denotes the exponential equation of the Wasserstein distance, and $\\tau$ is the temperature parameter. Wasserstein distance [35] is a metric used to quantify the dissimilarity between two probability distributions and well-suited for accurately measuring the distance between discrete and continuous distributions, as it accounts for the underlying structure and geometry of the data. It measures the minimum amount of work required to transform one distribution into another and can be calculated as follows:", + "bbox": [ + 212, + 625, + 782, + 744 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} W _ {\\text {d i s t a n c e}} \\left(p _ {i, t}, k\\right) = \\left\\| \\boldsymbol {\\mu} _ {i, t} - \\boldsymbol {\\mu} _ {k} \\right\\| ^ {2} + \\operatorname {T r} \\left(\\boldsymbol {\\Sigma} _ {i, t}\\right) \\\\ + \\operatorname {T r} \\left(\\boldsymbol {\\Sigma} _ {k}\\right) - 2 \\operatorname {T r} \\left(\\left(\\boldsymbol {\\Sigma} _ {i, t} \\boldsymbol {\\Sigma} _ {k}\\right) ^ {1 / 2}\\right), \\tag {5} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 325, + 753, + 784, + 801 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $\\mathrm{Tr}$ represents the trace of the matrix, $\\pmb{\\mu}_k$ and $\\pmb{\\Sigma}_k$ represent the mean and covariance matrix of sample $k$ . Based on $L^{NCE}$ , the region-wise contrastive loss", + "bbox": [ + 212, + 809, + 782, + 839 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "Region-aware Distribution Contrast", + "bbox": [ + 491, + 114, + 730, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 114, + 785, + 126 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "$L_{RC}$ can be defined as follows:", + "bbox": [ + 215, + 146, + 439, + 161 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\nL _ {R C} = \\frac {1}{K M} \\sum_ {t} ^ {K} \\sum_ {i} ^ {M} L _ {i, t} ^ {N C E}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 403, + 172, + 785, + 213 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $M$ is the number of regions in the feature map $t$ .", + "bbox": [ + 215, + 224, + 616, + 239 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "By minimizing $L_{RC}$ , the network learns to contrast the distributions from the same region with those from different regions, thereby enabling better cross-task consistency at the region level.", + "bbox": [ + 212, + 239, + 784, + 282 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Alternative Local Extraction Strategy. Here we explore an alternative local extraction strategy to investigate local cross-task consistency. To extract local information from the feature map, the simplest approach is to divide the map into patches, where patches at the same position serve as positive samples, and their distances are minimized. Conversely, patches from other positions act as negative samples, and their distances are maximized. We refer to this local extraction strategy as patch-aware contrast, which will be compared to the region-aware contrast method in Sec. 4.", + "bbox": [ + 212, + 285, + 784, + 404 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Alternative Local Contrast Strategy. Alternatively, the method of modeling regions as Gaussian distributions and performing contrastive learning can be replaced with other contrastive strategies. One strategy is pixel-to-pixel level region contrast, where each pixel within the region is pulled closer to the corresponding pixel in the positive sample region and pushed far apart from each pixel in the negative sample region. We refer to this as pixel contrast. Another strategy is to map the region to a vector and bring it closer to the vector of the positive sample region while pushing it away from the vector of the negative sample region. This strategy is referred to as vector contrast. We compare these alternative methods in Sec. 4.", + "bbox": [ + 212, + 406, + 785, + 555 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 215, + 580, + 375, + 598 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Datasets. We evaluated our method on two standard dense prediction multi-task datasets: NYU-V2 [36] and Cityscapes [9]. NYU-V2 is an indoor dataset of natural scenes, including three dense prediction tasks: semantic segmentation, depth estimation, and surface normal estimation. The semantic segmentation task in NYU-V2 contains 13 categories, depth estimation is provided by Microsoft Kinect, and surface normal estimation is provided by [10]. As [25, 29], all images were resized to $288 \\times 384$ for training and evaluation purposes. Cityscapes is a well-known dataset for autonomous driving, consisting of outdoor street scenes with two tasks: semantic segmentation and depth estimation. Following [25, 29], we used 7-class semantic segmentation annotations for evaluation and resized all images to $128 \\times 256$ to improve training speed.", + "bbox": [ + 212, + 612, + 785, + 779 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Experimental Setting. Following [25], we evaluated our method in two partially supervised multi-task settings: the one label setting and the random label setting. In the one label setting, each image is assigned a random task with a label. In the random label setting, each image is assigned random labels for", + "bbox": [ + 212, + 779, + 785, + 839 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "M. Li et al.", + "bbox": [ + 271, + 114, + 349, + 127 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/f707a989f859cc1faab9be31016cf997a70f8382b698e7b52640f15ee490b08f.jpg", + "table_caption": [ + "Table 1: Quantitative comparison on the NYU-V2 dataset. $\\uparrow (\\downarrow)$ denotes that, larger (smaller) values lead to better quality. * represents the available partially supervised SOTA. * represents the available fully supervised SOTA. + denotes the percentage improvement over the SOTA performance, corresponding to $\\uparrow (\\downarrow)$ . The bold denotes the best." + ], + "table_footnote": [], + "table_body": "
Label settingMethodBackboneExtraction strategyTasks
PatchRegionSemantic(mIoU)↑Depth(aErr)↓Normals(mErr)↓
fullTaskExpert* [44]ViT-Large57.580.37305.90
onelabelLi et al. [25]SegNet30.360.608832.08
onelabelDejaVu* [1]SegNet31.020.595932.15
onelabelGaussian Contrast(Ours)SegNet✓31.79 (+2.48)0.5835 (+2.25)30.38 (+5.51)
onelabelLi et al. * [25]HRNet1839.420.507114.79
onelabelDejaVu [1]HRNet18———
onelabelGaussian ContrastHRNet18✓41.56 (+5.43)0.4884 (+3.69)7.71 (+47.87)
onelabelGaussian Contrast(Ours)HRNet18✓42.28 (+7.26)0.4641 (+8.48)4.86 (+67.14)
onelabelTaskExpert* [44]ViT-Large49.530.430511.12
onelabelTaskExpert [44]+OursViT-Large✓55.81 (+12.68)0.4092 (+4.95)8.46 (+23.92)
random labelLi et al. [25]SegNet34.260.578731.06
random labelDejaVu* [1]SegNet35.720.566529.82
random labelGaussian Contrast(Ours)SegNet✓37.40 (+4.70)0.5428 (+4.18)28.97 (+2.85)
random labelLi et al. * [25]HRNet1841.350.484514.34
random labelDejaVu [1]HRNet18———
random labelGaussian ContrastHRNet18✓45.79 (+10.74)0.4619 (+4.66)7.37 (+48.61)
random labelGaussian Contrast(Ours)HRNet18✓46.21 (+11.75)0.4482 (+7.49)4.49 (+68.69)
random labelTaskExpert* [44]ViT-Large56.320.42826.47
random labelTaskExpert [44]+OursViT-Large✓57.14 (+1.46)0.3977 (+7.12)4.78 (+26.12)
", + "bbox": [ + 256, + 226, + 746, + 422 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "multiple tasks, with at least one task having a label and at most $K - 1$ tasks having labels, where $K$ represents the total number of tasks in the multi-task setup.", + "bbox": [ + 212, + 448, + 784, + 492 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Implementation Details and Evaluation Metrics. All our experiments were conducted on the NVIDIA A100 GPU. For NYU-V2, we trained the models for 20 epochs, while for Cityscapes, we trained them for 100 epochs. Following [25, 29], we employed cross-entropy loss for semantic segmentation, L1 loss for depth estimation, and cosine similarity loss for surface normal estimation. We utilized the exact same evaluation metrics as mentioned in [25, 29]. For the semantic segmentation task, we employed the mean intersection over union (mIoU) metric. The absolute error (aErr) was used to evaluate the depth estimation task, and the mean error (mErr) was employed for surface normal estimation.", + "bbox": [ + 212, + 493, + 787, + 643 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.1 Quantitative Evaluation", + "text_level": 1, + "bbox": [ + 214, + 665, + 460, + 680 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Quantitative Results on NYU-V2 Dataset. The evaluation results are shown in Tab. 1. Due to the absence of pre-trained parameters in Li et al. [25], the training process exhibits slower convergence. To expedite the convergence speed, we replaced the backbone of Li et al. [25] with HRNet18, which possesses a smaller training parameter size but utilizes pre-trained weights. The comparisons between Li et al. [25] with SegNet as the backbone and Li et al. [25] with HRNet18 as the backbone are presented in the first row and the fourth row, respectively, for each label setting. Since the source code of DejaVu [1] is not publicly available, we are unable to replace its backbone with HRNet18 for a fair comparison. To resolve this issue, we provide the results of our method based", + "bbox": [ + 212, + 688, + 787, + 840 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "Region-aware Distribution Contrast", + "bbox": [ + 491, + 114, + 730, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 114, + 785, + 126 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/d86848cd02852be31722a1afe7ed7afacb932e99eb137d6b2202a1f3ef454631.jpg", + "table_caption": [ + "Table 2: Quantitative comparison on the Cityscapes dataset. $\\uparrow (\\downarrow)$ denotes that, larger (smaller) values lead to better quality. * represents the available SOTA. + denotes the percentage improvement over the SOTA performance, corresponding to $\\uparrow$ ( $\\downarrow$ ). The bold denotes the best." + ], + "table_footnote": [], + "table_body": "
Label settingMethodBackboneContrast strategyTasks
Vector Pixel GaussianSemantic(mIoU)↑Depth(aErr) ↓
onelabelLi et al. [25]SegNet74.900.0161
onelabelLi et al. * [25]HRNet1881.730.0157
onelabelRegion-aware ContrastHRNet18✓82.76 (+1.26)0.0145 (+7.64)
onelabelRegion-aware ContrastHRNet1883.20 (+1.80)0.0141 (+10.19)
onelabelRegion-aware Contrast (Ours)HRNet1883.92 (+2.68)0.0121 (+22.73)
", + "bbox": [ + 272, + 212, + 730, + 284 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/2e0cdd819264641e445b4509671add86b934f965596b54c63de9e792d51c5c8f.jpg", + "table_caption": [ + "Table 3: Ablation study on the NYU-V2 dataset to explore the performance of different local extraction strategies and different local contrast strategies." + ], + "table_footnote": [], + "table_body": "
Label settingExtraction strategyContrast strategyTasks
PatchRegionVector Pixel GaussianSemantic(mIoU)↑Depth(aErr)↓Normals(mErr)↓
onelabel✓✓40.240.4998
onelabel✓✓41.500.4934
onelabel✓✓41.560.4884
onelabel✓✓41.840.4727
onelabel✓✓42.050.4663
onelabel✓✓42.280.4641
random label✓✓43.980.4795
random label✓✓44.620.4777
random label✓✓45.790.4619
random label✓✓45.830.4630
random label✓✓46.410.4518
random label✓✓46.210.4482
", + "bbox": [ + 272, + 340, + 730, + 474 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/a439ac06aa26d330254df83d7af8e02e21e825d90577a20bf0d36605342e9a44.jpg", + "image_caption": [ + "Fig. 4: Qualitative results of onelabel setting on NYU-V2. The first row shows the input image, the second row represents the ground-truth or predictions of semantic segmentation, the third row plots the ground-truth or predictions of depth estimation, and the final row presents the ground-truth or predictions of surface normal estimation." + ], + "image_footnote": [], + "bbox": [ + 261, + 492, + 743, + 691 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "on SegNet, which utilizes the same backbone. In all the methods compared using the same backbone, our approach demonstrates significant advantages. In the one label setting, the patch-aware Gaussian contrast method demonstrates", + "bbox": [ + 215, + 794, + 787, + 840 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "M. Li et al.", + "bbox": [ + 271, + 114, + 349, + 127 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/e030ba3de3ea14eca303112f5385202e20fe63ec6a7820c2772c12b33d83c7a1.jpg", + "image_caption": [ + "Fig. 5: Qualitative results of random label setting on NYU-V2. The first row shows the input image, the second row represents the ground-truth or predictions of semantic segmentation, the third row plots the ground-truth or predictions of depth estimation, and the final row presents the ground-truth or predictions of surface normal estimation." + ], + "image_footnote": [], + "bbox": [ + 261, + 143, + 743, + 339 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "significant improvements across all three tasks compared to Li et al. [25] across all three tasks, as it achieves a certain degree of local-level cross-task alignment. Furthermore, the region-aware Gaussian contrast method further enhances performance, achieving improvements of $2.48\\%$ , $2.25\\%$ and $5.51\\%$ over the SOTA [1] across the three tasks, owing to the achievement of cross-task consistency at the object level. In the random label setting, the superiority of the region-aware Gaussian contrast method can be observed, as it outperforms other methods significantly across all three tasks, yielding improvements of $4.70\\%$ , $4.18\\%$ and $2.85\\%$ , respectively, over the SOTA [1].", + "bbox": [ + 212, + 468, + 787, + 604 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Furthermore, we provide a comparison with the state-of-the-art fully supervised method, TaskExpert [44]. Tab. 1 clearly shows that our method significantly improves the performance of TaskExpert in partially supervised scenarios. Particularly in the random setting, when combined with our method, TaskExpert achieves even better performance than the fully supervised TaskExpert in the surface normal task. This once again demonstrates the effectiveness and wide applicability of our approach.", + "bbox": [ + 212, + 608, + 787, + 715 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Quantitative Results on Cityscapes Dataset. The evaluation results are describe in Tab. 2. Similar to NYU-V2, we first compared Li et al. [25] with HRNet18 as the backbone. Subsequently, we discussed the results of the region-aware contrast method under different contrast strategies. It is observed that Gaussian contrast demonstrates improvements in semantic segmentation and depth estimation compared to vector-based methods, indicating that representing regions using Gaussian distributions provides a more comprehensive and accurate representation than using vectors. In terms of computational speed, the", + "bbox": [ + 212, + 719, + 787, + 840 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "Region-aware Distribution Contrast", + "bbox": [ + 491, + 114, + 730, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/1294417756ecfaffc98f834f88a5edf2ba30c88e07d527ff9fa7844bc8879f22.jpg", + "table_caption": [ + "Table 4: Ablation study on the NYU-V2 dataset to investigate the contribution of SAM in the fully supervised setting." + ], + "table_footnote": [], + "table_body": "
MethodVector GaussianSemantic(mIoU)↑Depth(aErr)↓Normals(mErr)↓
Region-aware Contrast✓49.000.42068.1878
Region-aware Contrast (Ours)✓51.140.39824.8108
", + "bbox": [ + 300, + 184, + 702, + 219 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Gaussian contrast method is faster and more efficient, while still performing on par with pixel-based contrast methods.", + "bbox": [ + 212, + 247, + 782, + 277 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "4.2 Qualitative Evaluation", + "text_level": 1, + "bbox": [ + 214, + 300, + 447, + 315 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "The qualitative evaluation of onelabel setting, random label setting on the NYU-V2 dataset, and onelabel setting on the Cityscapes dataset are shown in Fig. 4, Fig. 5 and Fig. 6, respectively. We present a comparative analysis of the results from two sets of images for each setting of each dataset. As indicated by Fig. 4 and Fig. 5, it can be observed that our method exhibits superior accuracy and smoothness in semantic segmentation, which is particularly evident in Fig. 5. Regarding depth estimation, our method, which is based on region-aware Gaussian distribution contrast, achieves more accurate predictions when objects undergo changes compared to Li et al. [25]. The latter fails to capture such variations effectively. As for surface normal estimation, our method outperforms Li et al. [25] overall and demonstrates finer contour handling. These observations highlight the advantages of our approach over Li et al. [25] in terms of classification accuracy, smoothness in semantic segmentation, accurate depth estimation in the presence of object changes, and finer handling of surface normals.", + "bbox": [ + 212, + 325, + 787, + 547 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "We further exhibit the multi-task prediction results on the Cityscapes dataset in Fig. 6. Benefited from our proposed region-aware distribution contrast learning method, both the semantic segmentation and depth estimation tasks exhibit improved capability in capturing object edges and fine-grained details compared to Li et al. [25]. Additionally, there are significant overall performance improvement at a global level.", + "bbox": [ + 212, + 551, + 787, + 642 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "4.3 Ablation Study", + "text_level": 1, + "bbox": [ + 214, + 664, + 390, + 679 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "We conduct ablation study on NYU-V2 dataset to prove the effectiveness of our method from various aspects.", + "bbox": [ + 212, + 689, + 785, + 718 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Local Extraction Strategy. Tab. 3 provides a comparison between different local extraction strategies. All methods are compared based on HRNet18 as the backbone. It can be observed that in both the onelabel setting and random label setting, regardless of the contrast strategy employed, the region-aware contrast method significantly outperforms the patch-aware contrast method. This is because the region-aware method takes into account the objects present in the image and performs contrast based on objects, rather than simply comparing patches that may contain multiple objects.", + "bbox": [ + 212, + 719, + 787, + 839 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "M. Li et al.", + "bbox": [ + 271, + 114, + 349, + 127 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/87aec302bd7afacf1602e4c7516c97e91d078af20e021e6b2fde90c6fb645094.jpg", + "image_caption": [ + "Fig.6: Qualitative results of onelabel setting on Cityscapes. The first row shows the input image, the second row represents the ground-truth or predictions of semantic segmentation, and the final row plots the ground-truth or predictions of depth estimation." + ], + "image_footnote": [], + "bbox": [ + 261, + 142, + 743, + 345 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/348fd584912ca036c312facb44d88746323e958ab3d5fa512631126c0ca6e30e.jpg", + "table_caption": [ + "Table 5: Ablation study on the NYU-V2 dataset to investigate the performance of various distance measurements." + ], + "table_footnote": [], + "table_body": "
Label settingMethodTasks
Semantic(mIoU)↑Depth(aErr)↓Normals(mErr)↓
onelabelKL divergence41.150.47394.94
onelabelWasserstein distance (Ours)42.280.46414.86
random labelKL divergence45.530.45395.04
random labelWasserstein distance (Ours)46.210.44824.49
", + "bbox": [ + 287, + 474, + 718, + 547 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Local Contrast Strategy. Tab. 3 also provides a comparison between different local contrast strategies. All methods are similarly compared based on HRNet18. It can be observed that regardless of the label setting and the local extraction strategy employed, the Gaussian distribution contrast consistently outperforms the vector contrast and approaches or even surpasses the pixel contrast. This is because multi-task learning scenarios involve discrete and continuous dense predictions, utilizing Gaussian distributions to represent regions for contrast allows for better cross-task alignment in the overall distribution. In contrast, the hard alignment strategy of pixel contrast may lead to inaccuracies.", + "bbox": [ + 212, + 580, + 787, + 715 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Contribution of SAM. As shown in Tab. 2, Tab. 3, and Tab. 6 in our article, although the vector contrast approach also incorporates SAM, it fails to yield satisfactory results. In contrast, the adoption of distribution contrast exhibits a significant improvement over the vector contrast approach. To further demonstrate this point, we provide a comparison of results using semantic segmentation labels instead of SAM in fully supervised setting, as shown in Tab. 4. This further emphasizes that solely introducing fine-grained segmentation results is insufficient, highlighting the importance of the distribution contrast method.", + "bbox": [ + 212, + 718, + 787, + 840 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "Region-aware Distribution Contrast", + "bbox": [ + 491, + 114, + 730, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 784, + 126 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/ab9b042b7756ec114a842b482fa89ce7884e087b9ac65e7b02e6c4bab120f5bb.jpg", + "table_caption": [ + "Table 6: Ablation study on the NYU-V2 dataset to investigate the performance of various methods in the fully supervised setting." + ], + "table_footnote": [], + "table_body": "
MethodContrast strategyTasks
Vector Pixel GaussianSemantic(mIoU)↑Depth(aErr)↓Normals(mErr)↓
MTL46.380.46606.7298
Li et al. [25]47.200.44775.1192
Region-aware Contrast✓47.640.42315.1072
Region-aware Contrast✓50.060.41684.5318
Region-aware Contrast (Ours)✓50.830.40513.2878
", + "bbox": [ + 287, + 184, + 718, + 256 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Distance Measurement for Distributions. Tab. 5 presents a comparison of different distance measurements. While KL divergence is commonly used to calculate the distance between Gaussian distributions, it fails to accurately reflect the distance between discrete and continuous distributions, which is crucial in the context of multi-task alignment. On the other hand, Wasserstein distance has the advantage of measuring the distance between two non-overlapping distributions. The experimental results in Tab. 5 demonstrate that Wasserstein distance is better suited for evaluating the distance between different task distributions. More detailed information can be found in the supplementary materials.", + "bbox": [ + 212, + 282, + 785, + 417 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Fully Supervised Setting. Tab. 6 presents the comparison of our method in the fully supervised setting. All methods are based on HRNet18. Compared to other methods, the region-aware distribution contrast method once again demonstrates superiority, outperforming significantly in all three tasks. This is because Gaussian distributions can characterize the distribution characteristics of different tasks and facilitate accurate and efficient alignment with both vector and pixel-based approaches.", + "bbox": [ + 212, + 419, + 785, + 525 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 549, + 359, + 566 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In this paper, we propose a novel region-aware distribution contrast learning method for MTPSL. To facilitate cross-task alignment at the region level, predictions from different tasks are initially mapped into a joint feature map space. Simultaneously, SAM is employed to generate regions for each training sample, leveraging its ability to provide valuable prior knowledge. Subsequently, region-wise features are modeled by fitting a Gaussian distribution. The alignment among regions is achieved by minimizing the disparity among distributions from different tasks within the same region and maximizing the distance between distributions from different regions. The effectiveness of our method is demonstrated on two standard multi-task datasets through extensive experiments, outperforming the state-of-the-art by a significant margin. Although our method is designed for MTPSL, it can be applied to general multi-task learning problems to leverage the intrinsic relationships across different tasks, as demonstrated in the fully supervised experiments in Sec. 4.3. Moreover, we believe that the proposed region-aware distribution contrast learning method provides a promising way for solving more general multi-modal problems, and we will investigate this in our future work.", + "bbox": [ + 212, + 583, + 785, + 839 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "M. Li et al.", + "bbox": [ + 271, + 114, + 349, + 127 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgements: This work was supported in part by the National Natural Science Foundation of China under grant U23B2011, 62102069, U20B2063 and 62220106008, the Key R&D Program of Zhejiang under grant 2024SSYS0091, the Sichuan Science and Technology Program under Grant 2024NSFTD0034.", + "bbox": [ + 212, + 146, + 787, + 207 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 215, + 232, + 323, + 248 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Borse, S., Das, D., Park, H., Cai, H., Garrepalli, R., Porikli, F.: Dejavu: Conditional regenerative learning to enhance dense prediction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 19466-19477 (2023)", + "2. Brüggemann, D., Kanakis, M., Obukhov, A., Georgoulis, S., Van Gool, L.: Exploring relational context for multi-task dense prediction. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 15869-15878 (2021)", + "3. Chang, J., Lan, Z., Cheng, C., Wei, Y.: Data uncertainty learning in face recognition. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5710-5719 (2020)", + "4. Chen, S., Xie, G., Liu, Y., Peng, Q., Sun, B., Li, H., You, X., Shao, L.: Hsva: Hierarchical semantic-visual adaptation for zero-shot learning. Advances in Neural Information Processing Systems 34, 16622-16634 (2021)", + "5. Chen, T., Kornblith, S., Norouzi, M., Hinton, G.: A simple framework for contrastive learning of visual representations. In: International Conference on Machine Learning. pp. 1597-1607. PMLR (2020)", + "6. Chen, X., Fan, H., Girshick, R., He, K.: Improved baselines with momentum contrastive learning. arXiv preprint arXiv:2003.04297 (2020)", + "7. Chen, Z., Badrinarayanan, V., Lee, C.Y., Rabinovich, A.: GradNorm: Gradient normalization for adaptive loss balancing in deep multi-task networks. In: International Conference on Machine Learning. pp. 794-803. PMLR (2018)", + "8. Chen, Z., Zhu, L., Wan, L., Wang, S., Feng, W., Heng, P.A.: A multi-task mean teacher for semi-supervised shadow detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5611-5620 (2020)", + "9. Cordts, M., Omran, M., Ramos, S., Rehfeld, T., Enzweiler, M., Benenson, R., Franke, U., Roth, S., Schiele, B.: The cityscapes dataset for semantic urban scene understanding. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 3213-3223 (2016)", + "0. Eigen, D., Fergus, R.: Predicting depth, surface normals and semantic labels with a common multi-scale convolutional architecture. In: Proceedings of the IEEE International Conference on Computer Vision. pp. 2650-2658 (2015)", + "1. Eigen, D., Puhrsch, C., Fergus, R.: Depth map prediction from a single image using a multi-scale deep network. Advances in Neural Information Processing Systems 27 (2014)", + "2. Fan, R., Poggi, M., Mattoccia, S.: Contrastive learning for depth prediction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3225-3236 (2023)", + "3. Gao, Y., Ma, J., Zhao, M., Liu, W., Yuille, A.L.: Nddr-cnn: Layerwise feature fusing in multi-task cnns by neural discriminative dimensionality reduction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3205-3214 (2019)" + ], + "bbox": [ + 225, + 266, + 785, + 840 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "Region-aware Distribution Contrast", + "bbox": [ + 491, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "14. Guo, M., Haque, A., Huang, D.A., Yeung, S., Fei-Fei, L.: Dynamic task prioritization for multitask learning. In: Proceedings of the European Conference on Computer Vision. pp. 270-287 (2018)", + "15. He, K., Fan, H., Wu, Y., Xie, S., Girshick, R.: Momentum contrast for unsupervised visual representation learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9729-9738 (2020)", + "16. He, K., Gkioxari, G., Dólár, P., Girshick, R.: Mask r-cnn. In: Proceedings of the IEEE International Conference on Computer Vision. pp. 2961-2969 (2017)", + "17. Hu, H., Cui, J., Wang, L.: Region-aware contrastive learning for semantic segmentation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 16291-16301 (2021)", + "18. Huang, C., Tang, H., Fan, W., Xiao, Y., Hao, D., Qian, Z., Terzopoulos, D., et al.: Partly supervised multi-task learning. In: 2020 19th IEEE International Conference on Machine Learning and Applications. pp. 769-774. IEEE (2020)", + "19. Imran, A.A.Z., Terzopoulos, D.: Semi-supervised multi-task learning with chest x-ray images. In: Machine Learning in Medical Imaging: 10th International Workshop, MLMI 2019, Held in Conjunction with MICCAI 2019, Shenzhen, China, October 13, 2019, Proceedings 10. pp. 151-159. Springer (2019)", + "20. Jin, X., Lan, C., Zeng, W., Chen, Z.: Global distance-distributions separation for unsupervised person re-identification. In: Proceedings of the European Conference on Computer Vision. pp. 735–751. Springer (2020)", + "21. Jin, Z., Chen, S., Chen, Y., Xu, Z., Feng, H.: Let segment anything help image dehaze. arXiv preprint arXiv:2306.15870 (2023)", + "22. Kendall, A., Gal, Y., Cipolla, R.: Multi-task learning using uncertainty to weigh losses for scene geometry and semantics. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 7482-7491 (2018)", + "23. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., Dollar, P., Girshick, R.: Segment anything. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 4015-4026 (October 2023)", + "24. Latif, S., Rana, R., Khalifa, S., Jurdak, R., Epps, J., Schuller, B.W.: Multi-task semi-supervised adversarial autoencoding for speech emotion recognition. IEEE Transactions on Affective Computing 13(2), 992-1004 (2020)", + "25. Li, W.H., Liu, X., Bilen, H.: Learning multiple dense prediction tasks from partially annotated data. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18879-18889 (2022)", + "26. Liang, C., Wang, W., Miao, J., Yang, Y.: Gmmseg: Gaussian mixture based generative semantic segmentation models. Advances in Neural Information Processing Systems 35, 31360-31375 (2022)", + "27. Liu, Q., Liao, X., Carin, L.: Semi-supervised multitask learning. Advances in Neural Information Processing Systems 20 (2007)", + "28. Liu, S., James, S., Davison, A., Johns, E.: Auto-lambda: Disentangling dynamic task relationships. Transactions on Machine Learning Research (2022)", + "29. Liu, S., Johns, E., Davison, A.J.: End-to-end multi-task learning with attention. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1871-1880 (2019)", + "30. Long, J., Shelhamer, E., Darrell, T.: Fully convolutional networks for semantic segmentation. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 3431-3440 (2015)" + ], + "bbox": [ + 215, + 146, + 785, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "M. Li et al.", + "bbox": [ + 271, + 114, + 349, + 126 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "31. Lu, Y., Kumar, A., Zhai, S., Cheng, Y., Javidi, T., Feris, R.: Fully-adaptive feature sharing in multi-task networks with applications in person attribute classification. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 5334-5343 (2017)", + "32. Misra, I., Shrivastava, A., Gupta, A., Hebert, M.: Cross-Stitch networks for multitask learning. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 3994-4003 (2016)", + "33. Oord, A.v.d., Li, Y., Vinyals, O.: Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748 (2018)", + "34. Poggi, M., Aleotti, F., Tosi, F., Mattoccia, S.: On the uncertainty of self-supervised monocular depth estimation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3227-3237 (2020)", + "35. Rüschendorf, L.: The Wasserstein distance and approximation theorems. Probability Theory and Related Fields 70(1), 117-129 (1985)", + "36. Silberman, N., Hoiem, D., Kohli, P., Fergus, R.: Indoor segmentation and support inference from rgb images. In: Proceedings of the European Conference on Computer Vision. pp. 746-760. Springer (2012)", + "37. Standley, T., Zamir, A., Chen, D., Guibas, L., Malik, J., Savarese, S.: Which tasks should be learned together in multi-task learning? In: International Conference on Machine Learning. pp. 9120-9132. PMLR (2020)", + "38. Tian, Y., Krishnan, D., Isola, P.: Contrastive multiview coding. In: Proceedings of the European Conference on Computer Vision. pp. 776-794. Springer (2020)", + "39. Vandenhende, S., Georgoulis, S., Van Gool, L.: MTI-Net: Multi-scale task interaction networks for multi-task learning. In: Proceedings of the European Conference on Computer Vision. pp. 527-543. Springer (2020)", + "40. Wang, X., Zhang, R., Shen, C., Kong, T., Li, L.: Dense contrastive learning for self-supervised visual pre-training. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3024-3033 (2021)", + "41. Wu, L., Zhong, Z., Fang, L., He, X., Liu, Q., Ma, J., Chen, H.: Sparsely annotated semantic segmentation with adaptive gaussian mixtures. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15454-15464 (2023)", + "42. Wu, Z., Xiong, Y., Yu, S.X., Lin, D.: Unsupervised feature learning via nonparametric instance discrimination. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 3733-3742 (2018)", + "43. Xu, D., Ouyang, W., Wang, X., Sebe, N.: PAD-Net: Multi-tasks guided prediction-and-distillation network for simultaneous depth estimation and scene parsing. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 675-684 (2018)", + "44. Ye, H., Xu, D.: Taskexpert: Dynamically assembling multi-task representations with memorial mixture-of-experts. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 21828-21837 (2023)", + "45. Zamir, A.R., Sax, A., Cheerla, N., Suri, R., Cao, Z., Malik, J., Guibas, L.J.: Robust learning through cross-task consistency. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11197-11206 (2020)", + "46. Zamir, A.R., Sax, A., Shen, W., Guibas, L.J., Malik, J., Savarese, S.: Taskonomy: Disentangling task transfer learning. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 3712-3722 (2018)", + "47. Zhang, J., Fan, D.P., Dai, Y., Anwar, S., Saleh, F.S., Zhang, T., Barnes, N.: Uc-net: Uncertainty inspired rgb-d saliency detection via conditional variational autoen" + ], + "bbox": [ + 212, + 146, + 787, + 839 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "Region-aware Distribution Contrast", + "bbox": [ + 491, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "coders. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8582-8591 (2020)", + "48. Zhang, L., Liu, X., Guan, H.: Automtl: A programming framework for automating efficient multi-task learning. Advances in Neural Information Processing Systems 35, 34216-34228 (2022)", + "49. Zhang, M., Zhao, X., Yao, J., Yuan, C., Huang, W.: When noisy labels meet long tail dilemmas: A representation calibration method. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 15890-15900 (2023)", + "50. Zhang, Z., Cui, Z., Xu, C., Yan, Y., Sebe, N., Yang, J.: Pattern-affinitive propagation across depth, surface normal and semantic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4106-4115 (2019)" + ], + "bbox": [ + 215, + 146, + 785, + 313 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "M. Li et al.", + "bbox": [ + 271, + 114, + 349, + 126 + ], + "page_idx": 17 + } +] \ No newline at end of file diff --git a/2024/Region-aware Distribution Contrast_ A Novel Approach to Multi-Task Partially Supervised Learning/b09b7220-34e4-48d5-bc47-6cde62cb650d_model.json b/2024/Region-aware Distribution Contrast_ A Novel Approach to Multi-Task Partially Supervised Learning/b09b7220-34e4-48d5-bc47-6cde62cb650d_model.json new file mode 100644 index 0000000000000000000000000000000000000000..366fc43cb6b1ac032651523f91a92ebc9dcb8ab8 --- /dev/null +++ b/2024/Region-aware Distribution Contrast_ A Novel Approach to Multi-Task Partially Supervised Learning/b09b7220-34e4-48d5-bc47-6cde62cb650d_model.json @@ -0,0 +1,2194 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.26, + 0.142, + 0.743, + 0.21 + ], + "angle": 0, + "content": "Region-aware Distribution Contrast: A Novel Approach to Multi-Task Partially Supervised Learning" + }, + { + "type": "text", + "bbox": [ + 0.356, + 0.238, + 0.65, + 0.27 + ], + "angle": 0, + "content": "Meixuan Li, Tianyu Li, Guoqing Wang*, Peng Wang, Yang Yang, and Jie Zou" + }, + { + "type": "text", + "bbox": [ + 0.277, + 0.28, + 0.726, + 0.294 + ], + "angle": 0, + "content": "University of Electronic Science and Technology of China (UESTC)" + }, + { + "type": "text", + "bbox": [ + 0.284, + 0.296, + 0.717, + 0.308 + ], + "angle": 0, + "content": "limeixuan0801@126.com, {cosmos.yu, p.wang6}@@hotmail.com," + }, + { + "type": "text", + "bbox": [ + 0.329, + 0.31, + 0.677, + 0.322 + ], + "angle": 0, + "content": "{gqwang0420, yang.yang, jie.zou}@uestc.edu.cn" + }, + { + "type": "text", + "bbox": [ + 0.263, + 0.379, + 0.74, + 0.752 + ], + "angle": 0, + "content": "Abstract. In this study, we address the intricate challenge of multi-task dense prediction, encompassing tasks such as semantic segmentation, depth estimation, and surface normal estimation, particularly when dealing with partially annotated data (MTPSL). The complexity arises from the absence of complete task labels for each training image. Given the inter-related nature of these pixel-wise dense tasks, our focus is on mining and capturing cross-task relationships. Existing solutions typically rely on learning global image representations for global cross-task image matching, imposing constraints that, unfortunately, sacrifice the finer structures within the images. Attempting local matching as a remedy faces hurdles due to the lack of precise region supervision, making local alignment a challenging endeavor. The introduction of Segment Anything Model (SAM) sheds light on addressing local alignment challenges by providing free and high-quality solutions for region detection. Leveraging SAM-detected regions, the subsequent challenge lies in aligning the representations within these regions. Diverging from conventional methods that directly learn a monolithic image representation, our proposal involves modeling region-wise representations using Gaussian Distributions. Aligning these distributions between corresponding regions from different tasks imparts higher flexibility and capacity to capture intra-region structures, accommodating a broader range of tasks. This innovative approach significantly enhances our ability to effectively capture cross-task relationships, resulting in improved overall performance in partially supervised multi-task dense prediction scenarios. Extensive experiments conducted on two widely used benchmarks underscore the superior effectiveness of our proposed method, showcasing state-of-the-art performance even when compared to fully supervised methods." + }, + { + "type": "text", + "bbox": [ + 0.263, + 0.753, + 0.745, + 0.765 + ], + "angle": 0, + "content": "https://github.com/HereNowL/Region-aware-Distribution-Contrast" + }, + { + "type": "text", + "bbox": [ + 0.263, + 0.782, + 0.741, + 0.81 + ], + "angle": 0, + "content": "Keywords: Multi-task Learning \\(\\cdot\\) Partially Supervised Learning \\(\\cdot\\) Scene Understanding \\(\\cdot\\) Contrastive Learning" + }, + { + "type": "page_footnote", + "bbox": [ + 0.22, + 0.826, + 0.385, + 0.84 + ], + "angle": 0, + "content": "* Corresponding author." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.351, + 0.127 + ], + "angle": 0, + "content": "M. Li et al." + }, + { + "type": "image", + "bbox": [ + 0.262, + 0.144, + 0.493, + 0.25 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.319, + 0.252, + 0.438, + 0.261 + ], + "angle": 0, + "content": "(a) Contrastive Multiview Coding" + }, + { + "type": "image", + "bbox": [ + 0.502, + 0.145, + 0.744, + 0.248 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.556, + 0.251, + 0.689, + 0.261 + ], + "angle": 0, + "content": "(b) Region-wise Distribution Contrast" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.276, + 0.788, + 0.332 + ], + "angle": 0, + "content": "Fig. 1: Multiview consistency and region consistency. (a) Illustration of contrastive multiview consistency [38]. (b) Illustration of region consistency, where \\( p_{i,s} \\) and \\( p_{j,s} \\) represent the region distribution of semantic segmentation, \\( p_{i,d} \\) and \\( p_{j,d} \\) denotes the region distribution of depth estimation." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.358, + 0.376, + 0.373 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.387, + 0.788, + 0.524 + ], + "angle": 0, + "content": "With the rapid advancement of deep learning, dense prediction tasks, including semantic segmentation, depth estimation, and surface normal estimation, have witnessed remarkable progress [11, 16, 30, 34, 47]. Considering the inherent interdependence among these dense prediction tasks [45, 46], there has been a growing interest in employing unified multi-task learning networks to jointly tackle different dense prediction tasks [7, 22, 28, 29, 31, 32, 39, 43, 50]. In contrast to inefficient single-task learning networks, multi-task learning networks effectively acquire shared features, enabling them to capture more generalizable information across various tasks while avoiding redundant training [39, 43, 50]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.524, + 0.788, + 0.705 + ], + "angle": 0, + "content": "Recent advancements in multi-task dense prediction methods have primarily focused on two crucial aspects. The first aspect is dedicated to designing intricate network architectures to achieve effective multi-task learning [29, 32, 39, 43, 50], while the second aspect is devoted to devising optimal balancing strategies for loss functions in multi-task learning [7, 14, 22, 28, 29], aiming to mitigate the negative transfer across tasks. These directions mainly focus on fully supervised multi-task learning, where complete task labels are available for all training samples. However, acquiring pixel-level labels for dense prediction tasks is resource-intensive, particularly with multiple visual tasks. It is common to encounter scenarios where labels for certain tasks are missing or unreliable in some training samples [25], leading to a new research direction known as multi-task partially supervised learning (MTPSL)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.705, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Since not all task labels are available for each training image in MTPSL, it is natural to formulate a multi-task learning framework which can leverage the implicit relationship between tasks with and without explicit labels. Dense prediction tasks, such as semantic segmentation, depth estimation, and normal prediction, exhibit a complementary co-existence relationship [37,45,46], where each task can serve as auxiliary information for others, offering valuable insights reciprocally. While existing approaches map predictions into a joint vector-wise semantic space for task alignment [25], we argue that this overlooks the underlying structural information inherent in the original image. A global vector" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.492, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Region-aware Distribution Contrast" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.785, + 0.223 + ], + "angle": 0, + "content": "might be insufficient to characterize the information in the entire scene, limiting its effectiveness in constraining cross-task relationships. Local matching, as shown in Fig. 1, with its intrinsic local properties, serves as a potential remedy. Nonetheless, the absence of precise region supervision presents a fundamental obstacle, thereby intensifying the challenges associated with its implementation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.223, + 0.786, + 0.479 + ], + "angle": 0, + "content": "To tackle the outlined challenges, we introduce region-based cross-task alignment by leveraging the robust region segmentation capabilities inherent in Segment Anything Model (SAM) models. Utilizing well-segmented image regions facilitates the abstraction of region-level information, enabling the addition of cross-task consistency constraints. However, a subsequent crucial challenge arises in effectively representing these regional features to ensure the consistency and accuracy of cross-task alignment. In contrast to conventional methods that directly learn a unified image representation, we employ Gaussian distributions to model region representations, enhancing flexibility and improving the ability to capture intra-region structures. By adapting Gaussian distributions between corresponding regions from different tasks, our approach can be more broadly applicable to diverse task types, such as pixel-level classification tasks (e.g., semantic segmentation) and pixel-level regression tasks (e.g., depth estimation). Subsequently, we introduce distributional contrastive learning to explicitly perform region-wise cross-task alignment. This approach serves as a robust constraint, offering a meaningful way to mine relationships between tasks and enhance dense prediction results, particularly for those tasks lacking pixel-wise labels." + }, + { + "type": "text", + "bbox": [ + 0.239, + 0.48, + 0.652, + 0.494 + ], + "angle": 0, + "content": "In summary, our work masks the following contributions:" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.502, + 0.784, + 0.575 + ], + "angle": 0, + "content": "- We tackle the challenge of multi-task partially supervised (MTPSL) dense prediction from a fresh perspective. Our approach involves extracting cross-task local alignment through the utilization of SAM's easily obtainable local regions. This innovative strategy has demonstrated notable in alleviating label shortages." + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.578, + 0.784, + 0.619 + ], + "angle": 0, + "content": "- We propose a novel region distribution contrast method for local alignment, which offers increased flexibility, robustness, and wide applicability across multiple tasks." + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.623, + 0.784, + 0.68 + ], + "angle": 0, + "content": "- We validate the effectiveness of our proposed method through extensive experiments conducted on two widely used benchmark datasets. Notably, our approach achieves state-of-the-art performance in partially supervised learning, while also showing great potential in fully supervised learning." + }, + { + "type": "list", + "bbox": [ + 0.225, + 0.502, + 0.784, + 0.68 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.704, + 0.388, + 0.72 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.786, + 0.841 + ], + "angle": 0, + "content": "Multi-task Supervised Learning. Multi-task learning utilizes a single model to simultaneously handle multiple visual tasks, offering advantages such as faster inference speed and more efficient utilization of input data. Some approaches [13, 32, 48] focus on designing the encoder of multi-task network, while other approaches pay more attention to designing complex decoders to generate task-specific predictions by utilizing shared features [2, 39, 43, 50]. Uncertainty [22] obtain task loss weights by considering the heteroscedastic uncertainty" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.35, + 0.128 + ], + "angle": 0, + "content": "M. Li et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.224 + ], + "angle": 0, + "content": "of each task, Grad-Norm [7] directly adjusts gradients to balance task losses, and DWA [14] learns time-varying average task weights by considering the loss variation rate of each task. It is important to note that these methods operate under a significant and strong assumption, namely that they are developed and studied based on the availability of labels for all tasks across all images." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.227, + 0.803, + 0.469 + ], + "angle": 0, + "content": "Multi-task Semi-supervised Learning and Partially-supervised Learning. Learning multi-task models on fully annotated data requires a large-scale labeled dataset, and the cost of collecting sufficient labeled data can be high. Therefore, two common scenarios are worthy of note in multi-task learning. One is semi-supervised multi-task learning, where the dataset consists of limited annotation information for all tasks and a large amount of unlabeled data. A bunch of methods have been proposed for semi-supervised multi-task learning [8,18,19,24,27]. In recent works [8,18,19,24], regularization terms are applied to unlabeled samples from each task to encourage consistent predictions when the input is perturbed. Another scenario involves partial-supervision multi-task learning, where the dataset lacks sufficient labels for each task, and not all tasks have labels for every image. Li et al. [25] maps tasks into high-dimensional vectors for cross-task alignment. However, vector representations are insufficient and do not capture local-level alignment. In this paper, we propose leveraging distributions as a substitute for vector representations of task features and employing contrastive learning to achieve region-level alignment across tasks." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.474, + 0.79, + 0.671 + ], + "angle": 0, + "content": "Contrastive Learning for Vision. Recently, contrastive learning has made significant advancements in unsupervised learning [5,6,15,33,42]. DenseCL [40] and RegionContrast [17] have been utilized for semantic segmentation tasks using contrastive learning at the pixel and region levels respectively, achieving excellent results. For depth estimation, WCL [12] transformed depth from continuous values to discrete values and constructed contrastive losses based on windows to form positive and negative sample pairs. CMC [38] was the first to propose that different viewpoints of the same image should be mapped to nearby positions in a high-dimensional space, while different images should be mapped to distant positions, thus ensuring multiview consistency. Inspired by CMC, we believe that multiple viewpoints within the same local region should also exhibit consistency, while differing from other regions. In this paper, we leverage this idea to achieve region-level contrast for MTPSL." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.79, + 0.844 + ], + "angle": 0, + "content": "Gaussian Distribution. Recently, Gaussian distribution has been widely employed in computer vision to characterize the distribution of features [3, 4, 20, 26, 41, 49]. GMMSeg [26] utilized the Expectation-Maximization algorithm to construct Gaussian Mixture Models for each class in semantic segmentation, capturing class-conditional densities. AGMM [41] constructed an adaptive Gaussian Mixture Model for sparse annotation in semantic segmentation by incorporating labeled pixels and their similar unlabeled counterparts. DUL [3] employed mean and variance to characterize the distribution of faces for face recognition. Gaussian distributions have the capability to probabilistically describe the distribution characteristics of different visual tasks. However, to the best of our knowledge, Gaussian distributions have not been applied in MTPSL." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.493, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Region-aware Distribution Contrast" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "image", + "bbox": [ + 0.264, + 0.143, + 0.709, + 0.317 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.328, + 0.788, + 0.427 + ], + "angle": 0, + "content": "Fig. 2: Illustration of region-aware distribution contrast learning method for MTPSL. During training, supervised constraints \\( L_{Sup} \\) are applied to the annotated task \\( s \\). For task \\( s \\) and unlabelled task \\( t \\), the \\( a_{\\theta} \\) map the true label \\( y_{s} \\) and the prediction \\( \\hat{y}_{t} \\) in the high-dimensional space respectively and then model the region of task-specific features extracted using SAM as a Gaussian distribution. Contrastive learning is then employed to minimize the distance between distributions of the same region across different tasks while maximizing the distance to distributions of other regions." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.453, + 0.331, + 0.468 + ], + "angle": 0, + "content": "3 Method" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.483, + 0.373, + 0.497 + ], + "angle": 0, + "content": "3.1 Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.508, + 0.789, + 0.66 + ], + "angle": 0, + "content": "Problem Setup. Considering the problem of multi-task learning involving \\( K \\) tasks, where \\( K \\geq 2 \\). A training dataset \\( S = \\{x_{i}\\}_{i=1}^{N} \\) is given with \\( N \\) partially labelled samples, indicating that for each training sample \\( x_{i} \\), only a subset of the \\( K \\) tasks are provided true labels. Let \\( P_{i} \\) represents the number of labeled tasks for \\( x_{i} \\) and \\( Q_{i} \\) represents the number of unlabeled tasks, such that \\( P_{i} + Q_{i} = K \\). When \\( P_{i} \\) equals \\( K \\) for all \\( x_{i} \\in S \\), it indicates fully supervised multi-task learning. Conversely, when \\( Q_{i} \\) equals \\( K \\) for all \\( x_{i} \\in S \\), it implies that no task labels are available, leading to unsupervised multi-task learning. In this paper, we focus on addressing partially supervised multi-task learning, where each training image \\( x_{i} \\) can obtain labels for at least one task (\\( P_{i} \\geq 1 \\))." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.681, + 0.472, + 0.695 + ], + "angle": 0, + "content": "3.2 Overview of Our Method" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.704, + 0.788, + 0.795 + ], + "angle": 0, + "content": "As shown in Fig. 2, our method focuses on utilizing Gaussian distributions to represent features in local regions for contrastive learning, achieving better distribution consistency across tasks at the region-level for MTPSL. Assuming that task \\( s \\) has label while task \\( t \\) does not, our objective is to achieve alignment between the true label \\( y_{s} \\) of task \\( s \\) and the prediction \\( \\hat{y}_t \\) of task \\( t \\) at the region level using contrastive learning." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.796, + 0.788, + 0.84 + ], + "angle": 0, + "content": "In the training stage, to accurately and efficiently obtain region-level features for prediction, we simultaneously feed the input image to both the multi-task learning (MTL) network and the pre-trained SAM. Predictions are obtained" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.351, + 0.127 + ], + "angle": 0, + "content": "M. Li et al." + }, + { + "type": "image", + "bbox": [ + 0.279, + 0.144, + 0.73, + 0.301 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.314, + 0.788, + 0.384 + ], + "angle": 0, + "content": "Fig. 3: Illustration of region-level cross-task consistency. Initially, SAM predictions are employed to extract regions from the features of the particular task. Following that, these regions are modeled as Gaussian distributions. Finally, contrastive learning is utilized to minimize the distance between regions of the same region across different tasks and maximize the distance to regions of other regions." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.415, + 0.789, + 0.595 + ], + "angle": 0, + "content": "from both MTL network and SAM. For tasks with labeled data, we apply fully-supervised loss \\( L_{Sup} \\) as explicit training supervision. For tasks without label, an auxiliary network \\( a_{\\theta} \\) maps the unlabelled task prediction \\( \\hat{y}_t \\) and the annotated task label \\( y_s \\) to the same high-dimensional feature map space, and use \\( L_{RC} \\) to achieve cross-task consistency at the region level. Specifically, We first extract region-wise features in \\( \\hat{y}_t \\) and \\( y_s \\) based on the segmented regions identified by SAM, then model them as Gaussian distributions. After acquiring the Gaussian distributions for features in each region of each task, we utilize contrastive learning to minimize the distance between the Gaussian distributions of features in the same region across different tasks and maximize the distance from the Gaussian distributions of features from different regions. Finally, the optimization of our method can be defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.425, + 0.61, + 0.787, + 0.627 + ], + "angle": 0, + "content": "\\[\nL o s s = L _ {S u p} + L _ {R C}. \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.638, + 0.788, + 0.684 + ], + "angle": 0, + "content": "In the inference stage, as illustrated in Fig. 2, the input image is fed only into the MTL network to generate predictions for multiple tasks, notably, with SAM no longer being utilized." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.708, + 0.555, + 0.723 + ], + "angle": 0, + "content": "3.3 Region-aware Contrastive Learning" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.734, + 0.787, + 0.763 + ], + "angle": 0, + "content": "In this section, we explain the implementation details to achieve region-level cross-task consistency." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Region-level Features Extraction. To realize the region-level cross-task consistency, it is essential to informatively choose regions for feature extraction. Considering the robust capabilities of SAM for image segmentation [23], we employ it to generate the regions for each training sample, leveraging its ability to provide valuable prior knowledge. As shown in Fig. 2, to obtain fine-grained" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.492, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Region-aware Distribution Contrast" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.787, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.282 + ], + "angle": 0, + "content": "semantics and fine-grained edges of regions, we fed the image into MTL network and pre-trained SAM simultaneously. SAM performs fine-grained segmentation of each object in the image, and its output masks contain the ID of each area and its corresponding position in the image. Inspired by [21], for convenient region-wise feature extraction in task-specific feature maps, we compute the predicted masks of SAM as a grayscale image, where each pixel value corresponds to the area ID returned by SAM, indicating its category. Leveraging the grayscale masks of SAM, we split the image into different regions and extract region-wise features within feature maps of different tasks." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.283, + 0.785, + 0.389 + ], + "angle": 0, + "content": "Gaussian Distribution Modeling. After obtaining the region-wise features for each task, we consider modeling the region-wise features as Gaussian distributions. The utilization of Gaussian distributions represents the features of a region in a probabilistic manner, which provides a more comprehensive depiction of the region's variations and compensates for the shortcomings of pixel-level hard alignment. Specifically, we model each region \\( r_i \\) from the task-specific feature map \\( f_t \\) as a Gaussian distribution:" + }, + { + "type": "equation", + "bbox": [ + 0.381, + 0.4, + 0.785, + 0.416 + ], + "angle": 0, + "content": "\\[\np _ {i, t} = p \\left(r _ {i} \\mid f _ {t}\\right) = N \\left(r _ {i}; \\boldsymbol {\\mu} _ {i, t}, \\boldsymbol {\\Sigma} _ {i, t}\\right), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.425, + 0.784, + 0.455 + ], + "angle": 0, + "content": "where \\(\\pmb{\\mu}_{i,t}\\) and \\(\\pmb{\\Sigma}_{i,t}\\) represent mean and covariance matrix respectively, calculated from the region \\(r_i\\) in the feature map \\(f_{t}\\), as shown in Fig. 3." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.456, + 0.785, + 0.545 + ], + "angle": 0, + "content": "Region Distribution Contrast. Once we have obtained the Gaussian distributions for each region, our objective is to utilize contrastive learning to minimize the distance between Gaussian distributions of the same region across different tasks and maximize the distance between Gaussian distributions of different regions. Hence, the region-wise Gaussian distribution contrastive loss for the region distribution \\( p_{i,t} \\) can be defined as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.308, + 0.557, + 0.785, + 0.59 + ], + "angle": 0, + "content": "\\[\nL _ {i, t} ^ {N C E} = - \\sum_ {k _ {+}} \\log \\frac {\\operatorname {s i m} \\left(p _ {i , t} , k _ {+}\\right)}{\\operatorname {s i m} \\left(p _ {i , t} , k _ {+}\\right) + \\sum_ {k _ {-}} \\operatorname {s i m} \\left(p _ {i , t} , k _ {-}\\right)}, \\tag {3}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.358, + 0.607, + 0.785, + 0.621 + ], + "angle": 0, + "content": "\\[\n\\operatorname {s i m} \\left(p _ {i, t}, k\\right) = \\exp \\left(- W _ {\\text {d i s t a n c e}} \\left(p _ {i, t}, k\\right) / \\tau\\right), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.626, + 0.784, + 0.746 + ], + "angle": 0, + "content": "where \\( k_{+} \\in \\{p_{i,s}\\}_{s \\neq t} \\), \\( k_{-} \\in \\{p_{j,*}\\}_{j \\neq i} \\), * denotes any task, sim denotes the exponential equation of the Wasserstein distance, and \\( \\tau \\) is the temperature parameter. Wasserstein distance [35] is a metric used to quantify the dissimilarity between two probability distributions and well-suited for accurately measuring the distance between discrete and continuous distributions, as it accounts for the underlying structure and geometry of the data. It measures the minimum amount of work required to transform one distribution into another and can be calculated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.326, + 0.755, + 0.785, + 0.802 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} W _ {\\text {d i s t a n c e}} \\left(p _ {i, t}, k\\right) = \\left\\| \\boldsymbol {\\mu} _ {i, t} - \\boldsymbol {\\mu} _ {k} \\right\\| ^ {2} + \\operatorname {T r} \\left(\\boldsymbol {\\Sigma} _ {i, t}\\right) \\\\ + \\operatorname {T r} \\left(\\boldsymbol {\\Sigma} _ {k}\\right) - 2 \\operatorname {T r} \\left(\\left(\\boldsymbol {\\Sigma} _ {i, t} \\boldsymbol {\\Sigma} _ {k}\\right) ^ {1 / 2}\\right), \\tag {5} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.784, + 0.84 + ], + "angle": 0, + "content": "where \\(\\mathrm{Tr}\\) represents the trace of the matrix, \\(\\pmb{\\mu}_k\\) and \\(\\pmb{\\Sigma}_k\\) represent the mean and covariance matrix of sample \\(k\\). Based on \\(L^{NCE}\\), the region-wise contrastive loss" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.35, + 0.128 + ], + "angle": 0, + "content": "M. Li et al." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.147, + 0.441, + 0.162 + ], + "angle": 0, + "content": "\\(L_{RC}\\) can be defined as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.405, + 0.174, + 0.786, + 0.214 + ], + "angle": 0, + "content": "\\[\nL _ {R C} = \\frac {1}{K M} \\sum_ {t} ^ {K} \\sum_ {i} ^ {M} L _ {i, t} ^ {N C E}, \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.225, + 0.617, + 0.24 + ], + "angle": 0, + "content": "where \\(M\\) is the number of regions in the feature map \\(t\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.24, + 0.785, + 0.284 + ], + "angle": 0, + "content": "By minimizing \\( L_{RC} \\), the network learns to contrast the distributions from the same region with those from different regions, thereby enabling better cross-task consistency at the region level." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.286, + 0.785, + 0.405 + ], + "angle": 0, + "content": "Alternative Local Extraction Strategy. Here we explore an alternative local extraction strategy to investigate local cross-task consistency. To extract local information from the feature map, the simplest approach is to divide the map into patches, where patches at the same position serve as positive samples, and their distances are minimized. Conversely, patches from other positions act as negative samples, and their distances are maximized. We refer to this local extraction strategy as patch-aware contrast, which will be compared to the region-aware contrast method in Sec. 4." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.407, + 0.787, + 0.556 + ], + "angle": 0, + "content": "Alternative Local Contrast Strategy. Alternatively, the method of modeling regions as Gaussian distributions and performing contrastive learning can be replaced with other contrastive strategies. One strategy is pixel-to-pixel level region contrast, where each pixel within the region is pulled closer to the corresponding pixel in the positive sample region and pushed far apart from each pixel in the negative sample region. We refer to this as pixel contrast. Another strategy is to map the region to a vector and bring it closer to the vector of the positive sample region while pushing it away from the vector of the negative sample region. This strategy is referred to as vector contrast. We compare these alternative methods in Sec. 4." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.582, + 0.376, + 0.599 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.613, + 0.787, + 0.78 + ], + "angle": 0, + "content": "Datasets. We evaluated our method on two standard dense prediction multi-task datasets: NYU-V2 [36] and Cityscapes [9]. NYU-V2 is an indoor dataset of natural scenes, including three dense prediction tasks: semantic segmentation, depth estimation, and surface normal estimation. The semantic segmentation task in NYU-V2 contains 13 categories, depth estimation is provided by Microsoft Kinect, and surface normal estimation is provided by [10]. As [25, 29], all images were resized to \\(288 \\times 384\\) for training and evaluation purposes. Cityscapes is a well-known dataset for autonomous driving, consisting of outdoor street scenes with two tasks: semantic segmentation and depth estimation. Following [25, 29], we used 7-class semantic segmentation annotations for evaluation and resized all images to \\(128 \\times 256\\) to improve training speed." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Experimental Setting. Following [25], we evaluated our method in two partially supervised multi-task settings: the one label setting and the random label setting. In the one label setting, each image is assigned a random task with a label. In the random label setting, each image is assigned random labels for" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.492, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Region-aware Distribution Contrast" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.787, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.145, + 0.788, + 0.214 + ], + "angle": 0, + "content": "Table 1: Quantitative comparison on the NYU-V2 dataset. \\(\\uparrow (\\downarrow)\\) denotes that, larger (smaller) values lead to better quality. * represents the available partially supervised SOTA. * represents the available fully supervised SOTA. + denotes the percentage improvement over the SOTA performance, corresponding to \\(\\uparrow (\\downarrow)\\). The bold denotes the best." + }, + { + "type": "table", + "bbox": [ + 0.258, + 0.227, + 0.747, + 0.424 + ], + "angle": 0, + "content": "
Label settingMethodBackboneExtraction strategyTasks
PatchRegionSemantic(mIoU)↑Depth(aErr)↓Normals(mErr)↓
fullTaskExpert* [44]ViT-Large57.580.37305.90
onelabelLi et al. [25]SegNet30.360.608832.08
onelabelDejaVu* [1]SegNet31.020.595932.15
onelabelGaussian Contrast(Ours)SegNet✓31.79 (+2.48)0.5835 (+2.25)30.38 (+5.51)
onelabelLi et al. * [25]HRNet1839.420.507114.79
onelabelDejaVu [1]HRNet18———
onelabelGaussian ContrastHRNet18✓41.56 (+5.43)0.4884 (+3.69)7.71 (+47.87)
onelabelGaussian Contrast(Ours)HRNet18✓42.28 (+7.26)0.4641 (+8.48)4.86 (+67.14)
onelabelTaskExpert* [44]ViT-Large49.530.430511.12
onelabelTaskExpert [44]+OursViT-Large✓55.81 (+12.68)0.4092 (+4.95)8.46 (+23.92)
random labelLi et al. [25]SegNet34.260.578731.06
random labelDejaVu* [1]SegNet35.720.566529.82
random labelGaussian Contrast(Ours)SegNet✓37.40 (+4.70)0.5428 (+4.18)28.97 (+2.85)
random labelLi et al. * [25]HRNet1841.350.484514.34
random labelDejaVu [1]HRNet18———
random labelGaussian ContrastHRNet18✓45.79 (+10.74)0.4619 (+4.66)7.37 (+48.61)
random labelGaussian Contrast(Ours)HRNet18✓46.21 (+11.75)0.4482 (+7.49)4.49 (+68.69)
random labelTaskExpert* [44]ViT-Large56.320.42826.47
random labelTaskExpert [44]+OursViT-Large✓57.14 (+1.46)0.3977 (+7.12)4.78 (+26.12)
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.449, + 0.785, + 0.493 + ], + "angle": 0, + "content": "multiple tasks, with at least one task having a label and at most \\( K - 1 \\) tasks having labels, where \\( K \\) represents the total number of tasks in the multi-task setup." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.494, + 0.788, + 0.645 + ], + "angle": 0, + "content": "Implementation Details and Evaluation Metrics. All our experiments were conducted on the NVIDIA A100 GPU. For NYU-V2, we trained the models for 20 epochs, while for Cityscapes, we trained them for 100 epochs. Following [25, 29], we employed cross-entropy loss for semantic segmentation, L1 loss for depth estimation, and cosine similarity loss for surface normal estimation. We utilized the exact same evaluation metrics as mentioned in [25, 29]. For the semantic segmentation task, we employed the mean intersection over union (mIoU) metric. The absolute error (aErr) was used to evaluate the depth estimation task, and the mean error (mErr) was employed for surface normal estimation." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.666, + 0.461, + 0.681 + ], + "angle": 0, + "content": "4.1 Quantitative Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.789, + 0.842 + ], + "angle": 0, + "content": "Quantitative Results on NYU-V2 Dataset. The evaluation results are shown in Tab. 1. Due to the absence of pre-trained parameters in Li et al. [25], the training process exhibits slower convergence. To expedite the convergence speed, we replaced the backbone of Li et al. [25] with HRNet18, which possesses a smaller training parameter size but utilizes pre-trained weights. The comparisons between Li et al. [25] with SegNet as the backbone and Li et al. [25] with HRNet18 as the backbone are presented in the first row and the fourth row, respectively, for each label setting. Since the source code of DejaVu [1] is not publicly available, we are unable to replace its backbone with HRNet18 for a fair comparison. To resolve this issue, we provide the results of our method based" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.351, + 0.128 + ], + "angle": 0, + "content": "M. Li et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.788, + 0.2 + ], + "angle": 0, + "content": "Table 2: Quantitative comparison on the Cityscapes dataset. \\(\\uparrow (\\downarrow)\\) denotes that, larger (smaller) values lead to better quality. * represents the available SOTA. + denotes the percentage improvement over the SOTA performance, corresponding to \\(\\uparrow\\) (\\(\\downarrow\\)). The bold denotes the best." + }, + { + "type": "table", + "bbox": [ + 0.273, + 0.213, + 0.732, + 0.285 + ], + "angle": 0, + "content": "
Label settingMethodBackboneContrast strategyTasks
Vector Pixel GaussianSemantic(mIoU)↑Depth(aErr) ↓
onelabelLi et al. [25]SegNet74.900.0161
onelabelLi et al. * [25]HRNet1881.730.0157
onelabelRegion-aware ContrastHRNet18✓82.76 (+1.26)0.0145 (+7.64)
onelabelRegion-aware ContrastHRNet1883.20 (+1.80)0.0141 (+10.19)
onelabelRegion-aware Contrast (Ours)HRNet1883.92 (+2.68)0.0121 (+22.73)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.301, + 0.785, + 0.33 + ], + "angle": 0, + "content": "Table 3: Ablation study on the NYU-V2 dataset to explore the performance of different local extraction strategies and different local contrast strategies." + }, + { + "type": "table", + "bbox": [ + 0.273, + 0.341, + 0.732, + 0.476 + ], + "angle": 0, + "content": "
Label settingExtraction strategyContrast strategyTasks
PatchRegionVector Pixel GaussianSemantic(mIoU)↑Depth(aErr)↓Normals(mErr)↓
onelabel✓✓40.240.4998
onelabel✓✓41.500.4934
onelabel✓✓41.560.4884
onelabel✓✓41.840.4727
onelabel✓✓42.050.4663
onelabel✓✓42.280.4641
random label✓✓43.980.4795
random label✓✓44.620.4777
random label✓✓45.790.4619
random label✓✓45.830.4630
random label✓✓46.410.4518
random label✓✓46.210.4482
" + }, + { + "type": "image", + "bbox": [ + 0.263, + 0.493, + 0.744, + 0.692 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.705, + 0.788, + 0.763 + ], + "angle": 0, + "content": "Fig. 4: Qualitative results of onelabel setting on NYU-V2. The first row shows the input image, the second row represents the ground-truth or predictions of semantic segmentation, the third row plots the ground-truth or predictions of depth estimation, and the final row presents the ground-truth or predictions of surface normal estimation." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.795, + 0.788, + 0.841 + ], + "angle": 0, + "content": "on SegNet, which utilizes the same backbone. In all the methods compared using the same backbone, our approach demonstrates significant advantages. In the one label setting, the patch-aware Gaussian contrast method demonstrates" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.493, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Region-aware Distribution Contrast" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "image", + "bbox": [ + 0.262, + 0.144, + 0.744, + 0.34 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.359, + 0.788, + 0.429 + ], + "angle": 0, + "content": "Fig. 5: Qualitative results of random label setting on NYU-V2. The first row shows the input image, the second row represents the ground-truth or predictions of semantic segmentation, the third row plots the ground-truth or predictions of depth estimation, and the final row presents the ground-truth or predictions of surface normal estimation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.469, + 0.788, + 0.606 + ], + "angle": 0, + "content": "significant improvements across all three tasks compared to Li et al. [25] across all three tasks, as it achieves a certain degree of local-level cross-task alignment. Furthermore, the region-aware Gaussian contrast method further enhances performance, achieving improvements of \\(2.48\\%\\), \\(2.25\\%\\) and \\(5.51\\%\\) over the SOTA [1] across the three tasks, owing to the achievement of cross-task consistency at the object level. In the random label setting, the superiority of the region-aware Gaussian contrast method can be observed, as it outperforms other methods significantly across all three tasks, yielding improvements of \\(4.70\\%\\), \\(4.18\\%\\) and \\(2.85\\%\\), respectively, over the SOTA [1]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.609, + 0.788, + 0.716 + ], + "angle": 0, + "content": "Furthermore, we provide a comparison with the state-of-the-art fully supervised method, TaskExpert [44]. Tab. 1 clearly shows that our method significantly improves the performance of TaskExpert in partially supervised scenarios. Particularly in the random setting, when combined with our method, TaskExpert achieves even better performance than the fully supervised TaskExpert in the surface normal task. This once again demonstrates the effectiveness and wide applicability of our approach." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Quantitative Results on Cityscapes Dataset. The evaluation results are describe in Tab. 2. Similar to NYU-V2, we first compared Li et al. [25] with HRNet18 as the backbone. Subsequently, we discussed the results of the region-aware contrast method under different contrast strategies. It is observed that Gaussian contrast demonstrates improvements in semantic segmentation and depth estimation compared to vector-based methods, indicating that representing regions using Gaussian distributions provides a more comprehensive and accurate representation than using vectors. In terms of computational speed, the" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.351, + 0.128 + ], + "angle": 0, + "content": "M. Li et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.215, + 0.145, + 0.788, + 0.174 + ], + "angle": 0, + "content": "Table 4: Ablation study on the NYU-V2 dataset to investigate the contribution of SAM in the fully supervised setting." + }, + { + "type": "table", + "bbox": [ + 0.302, + 0.185, + 0.703, + 0.22 + ], + "angle": 0, + "content": "
MethodVector GaussianSemantic(mIoU)↑Depth(aErr)↓Normals(mErr)↓
Region-aware Contrast✓49.000.42068.1878
Region-aware Contrast (Ours)✓51.140.39824.8108
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.248, + 0.784, + 0.278 + ], + "angle": 0, + "content": "Gaussian contrast method is faster and more efficient, while still performing on par with pixel-based contrast methods." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.301, + 0.449, + 0.316 + ], + "angle": 0, + "content": "4.2 Qualitative Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.326, + 0.788, + 0.549 + ], + "angle": 0, + "content": "The qualitative evaluation of onelabel setting, random label setting on the NYU-V2 dataset, and onelabel setting on the Cityscapes dataset are shown in Fig. 4, Fig. 5 and Fig. 6, respectively. We present a comparative analysis of the results from two sets of images for each setting of each dataset. As indicated by Fig. 4 and Fig. 5, it can be observed that our method exhibits superior accuracy and smoothness in semantic segmentation, which is particularly evident in Fig. 5. Regarding depth estimation, our method, which is based on region-aware Gaussian distribution contrast, achieves more accurate predictions when objects undergo changes compared to Li et al. [25]. The latter fails to capture such variations effectively. As for surface normal estimation, our method outperforms Li et al. [25] overall and demonstrates finer contour handling. These observations highlight the advantages of our approach over Li et al. [25] in terms of classification accuracy, smoothness in semantic segmentation, accurate depth estimation in the presence of object changes, and finer handling of surface normals." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.552, + 0.788, + 0.643 + ], + "angle": 0, + "content": "We further exhibit the multi-task prediction results on the Cityscapes dataset in Fig. 6. Benefited from our proposed region-aware distribution contrast learning method, both the semantic segmentation and depth estimation tasks exhibit improved capability in capturing object edges and fine-grained details compared to Li et al. [25]. Additionally, there are significant overall performance improvement at a global level." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.665, + 0.391, + 0.68 + ], + "angle": 0, + "content": "4.3 Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.69, + 0.787, + 0.719 + ], + "angle": 0, + "content": "We conduct ablation study on NYU-V2 dataset to prove the effectiveness of our method from various aspects." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Local Extraction Strategy. Tab. 3 provides a comparison between different local extraction strategies. All methods are compared based on HRNet18 as the backbone. It can be observed that in both the onelabel setting and random label setting, regardless of the contrast strategy employed, the region-aware contrast method significantly outperforms the patch-aware contrast method. This is because the region-aware method takes into account the objects present in the image and performs contrast based on objects, rather than simply comparing patches that may contain multiple objects." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.493, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Region-aware Distribution Contrast" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "image", + "bbox": [ + 0.262, + 0.143, + 0.744, + 0.347 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.359, + 0.788, + 0.414 + ], + "angle": 0, + "content": "Fig.6: Qualitative results of onelabel setting on Cityscapes. The first row shows the input image, the second row represents the ground-truth or predictions of semantic segmentation, and the final row plots the ground-truth or predictions of depth estimation." + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.434, + 0.789, + 0.463 + ], + "angle": 0, + "content": "Table 5: Ablation study on the NYU-V2 dataset to investigate the performance of various distance measurements." + }, + { + "type": "table", + "bbox": [ + 0.288, + 0.475, + 0.719, + 0.549 + ], + "angle": 0, + "content": "
Label settingMethodTasks
Semantic(mIoU)↑Depth(aErr)↓Normals(mErr)↓
onelabelKL divergence41.150.47394.94
onelabelWasserstein distance (Ours)42.280.46414.86
random labelKL divergence45.530.45395.04
random labelWasserstein distance (Ours)46.210.44824.49
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.581, + 0.788, + 0.717 + ], + "angle": 0, + "content": "Local Contrast Strategy. Tab. 3 also provides a comparison between different local contrast strategies. All methods are similarly compared based on HRNet18. It can be observed that regardless of the label setting and the local extraction strategy employed, the Gaussian distribution contrast consistently outperforms the vector contrast and approaches or even surpasses the pixel contrast. This is because multi-task learning scenarios involve discrete and continuous dense predictions, utilizing Gaussian distributions to represent regions for contrast allows for better cross-task alignment in the overall distribution. In contrast, the hard alignment strategy of pixel contrast may lead to inaccuracies." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.719, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Contribution of SAM. As shown in Tab. 2, Tab. 3, and Tab. 6 in our article, although the vector contrast approach also incorporates SAM, it fails to yield satisfactory results. In contrast, the adoption of distribution contrast exhibits a significant improvement over the vector contrast approach. To further demonstrate this point, we provide a comparison of results using semantic segmentation labels instead of SAM in fully supervised setting, as shown in Tab. 4. This further emphasizes that solely introducing fine-grained segmentation results is insufficient, highlighting the importance of the distribution contrast method." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.351, + 0.128 + ], + "angle": 0, + "content": "M. Li et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.788, + 0.174 + ], + "angle": 0, + "content": "Table 6: Ablation study on the NYU-V2 dataset to investigate the performance of various methods in the fully supervised setting." + }, + { + "type": "table", + "bbox": [ + 0.288, + 0.185, + 0.719, + 0.257 + ], + "angle": 0, + "content": "
MethodContrast strategyTasks
Vector Pixel GaussianSemantic(mIoU)↑Depth(aErr)↓Normals(mErr)↓
MTL46.380.46606.7298
Li et al. [25]47.200.44775.1192
Region-aware Contrast✓47.640.42315.1072
Region-aware Contrast✓50.060.41684.5318
Region-aware Contrast (Ours)✓50.830.40513.2878
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.284, + 0.787, + 0.419 + ], + "angle": 0, + "content": "Distance Measurement for Distributions. Tab. 5 presents a comparison of different distance measurements. While KL divergence is commonly used to calculate the distance between Gaussian distributions, it fails to accurately reflect the distance between discrete and continuous distributions, which is crucial in the context of multi-task alignment. On the other hand, Wasserstein distance has the advantage of measuring the distance between two non-overlapping distributions. The experimental results in Tab. 5 demonstrate that Wasserstein distance is better suited for evaluating the distance between different task distributions. More detailed information can be found in the supplementary materials." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.42, + 0.787, + 0.526 + ], + "angle": 0, + "content": "Fully Supervised Setting. Tab. 6 presents the comparison of our method in the fully supervised setting. All methods are based on HRNet18. Compared to other methods, the region-aware distribution contrast method once again demonstrates superiority, outperforming significantly in all three tasks. This is because Gaussian distributions can characterize the distribution characteristics of different tasks and facilitate accurate and efficient alignment with both vector and pixel-based approaches." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.55, + 0.36, + 0.567 + ], + "angle": 0, + "content": "5 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.584, + 0.787, + 0.84 + ], + "angle": 0, + "content": "In this paper, we propose a novel region-aware distribution contrast learning method for MTPSL. To facilitate cross-task alignment at the region level, predictions from different tasks are initially mapped into a joint feature map space. Simultaneously, SAM is employed to generate regions for each training sample, leveraging its ability to provide valuable prior knowledge. Subsequently, region-wise features are modeled by fitting a Gaussian distribution. The alignment among regions is achieved by minimizing the disparity among distributions from different tasks within the same region and maximizing the distance between distributions from different regions. The effectiveness of our method is demonstrated on two standard multi-task datasets through extensive experiments, outperforming the state-of-the-art by a significant margin. Although our method is designed for MTPSL, it can be applied to general multi-task learning problems to leverage the intrinsic relationships across different tasks, as demonstrated in the fully supervised experiments in Sec. 4.3. Moreover, we believe that the proposed region-aware distribution contrast learning method provides a promising way for solving more general multi-modal problems, and we will investigate this in our future work." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.492, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Region-aware Distribution Contrast" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.208 + ], + "angle": 0, + "content": "Acknowledgements: This work was supported in part by the National Natural Science Foundation of China under grant U23B2011, 62102069, U20B2063 and 62220106008, the Key R&D Program of Zhejiang under grant 2024SSYS0091, the Sichuan Science and Technology Program under Grant 2024NSFTD0034." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.233, + 0.324, + 0.249 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.267, + 0.786, + 0.322 + ], + "angle": 0, + "content": "1. Borse, S., Das, D., Park, H., Cai, H., Garrepalli, R., Porikli, F.: Dejavu: Conditional regenerative learning to enhance dense prediction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 19466-19477 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.324, + 0.786, + 0.365 + ], + "angle": 0, + "content": "2. Brüggemann, D., Kanakis, M., Obukhov, A., Georgoulis, S., Van Gool, L.: Exploring relational context for multi-task dense prediction. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 15869-15878 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.366, + 0.786, + 0.407 + ], + "angle": 0, + "content": "3. Chang, J., Lan, Z., Cheng, C., Wei, Y.: Data uncertainty learning in face recognition. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5710-5719 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.407, + 0.786, + 0.449 + ], + "angle": 0, + "content": "4. Chen, S., Xie, G., Liu, Y., Peng, Q., Sun, B., Li, H., You, X., Shao, L.: Hsva: Hierarchical semantic-visual adaptation for zero-shot learning. Advances in Neural Information Processing Systems 34, 16622-16634 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.449, + 0.786, + 0.491 + ], + "angle": 0, + "content": "5. Chen, T., Kornblith, S., Norouzi, M., Hinton, G.: A simple framework for contrastive learning of visual representations. In: International Conference on Machine Learning. pp. 1597-1607. PMLR (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.491, + 0.786, + 0.519 + ], + "angle": 0, + "content": "6. Chen, X., Fan, H., Girshick, R., He, K.: Improved baselines with momentum contrastive learning. arXiv preprint arXiv:2003.04297 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.519, + 0.786, + 0.561 + ], + "angle": 0, + "content": "7. Chen, Z., Badrinarayanan, V., Lee, C.Y., Rabinovich, A.: GradNorm: Gradient normalization for adaptive loss balancing in deep multi-task networks. In: International Conference on Machine Learning. pp. 794-803. PMLR (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.561, + 0.786, + 0.603 + ], + "angle": 0, + "content": "8. Chen, Z., Zhu, L., Wan, L., Wang, S., Feng, W., Heng, P.A.: A multi-task mean teacher for semi-supervised shadow detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5611-5620 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.603, + 0.786, + 0.659 + ], + "angle": 0, + "content": "9. Cordts, M., Omran, M., Ramos, S., Rehfeld, T., Enzweiler, M., Benenson, R., Franke, U., Roth, S., Schiele, B.: The cityscapes dataset for semantic urban scene understanding. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 3213-3223 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.659, + 0.786, + 0.701 + ], + "angle": 0, + "content": "0. Eigen, D., Fergus, R.: Predicting depth, surface normals and semantic labels with a common multi-scale convolutional architecture. In: Proceedings of the IEEE International Conference on Computer Vision. pp. 2650-2658 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.701, + 0.786, + 0.742 + ], + "angle": 0, + "content": "1. Eigen, D., Puhrsch, C., Fergus, R.: Depth map prediction from a single image using a multi-scale deep network. Advances in Neural Information Processing Systems 27 (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.742, + 0.786, + 0.785 + ], + "angle": 0, + "content": "2. Fan, R., Poggi, M., Mattoccia, S.: Contrastive learning for depth prediction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3225-3236 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.785, + 0.786, + 0.841 + ], + "angle": 0, + "content": "3. Gao, Y., Ma, J., Zhao, M., Liu, W., Yuille, A.L.: Nddr-cnn: Layerwise feature fusing in multi-task cnns by neural discriminative dimensionality reduction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3205-3214 (2019)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.267, + 0.786, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.35, + 0.127 + ], + "angle": 0, + "content": "M. Li et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.785, + 0.189 + ], + "angle": 0, + "content": "14. Guo, M., Haque, A., Huang, D.A., Yeung, S., Fei-Fei, L.: Dynamic task prioritization for multitask learning. In: Proceedings of the European Conference on Computer Vision. pp. 270-287 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.19, + 0.787, + 0.233 + ], + "angle": 0, + "content": "15. He, K., Fan, H., Wu, Y., Xie, S., Girshick, R.: Momentum contrast for unsupervised visual representation learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9729-9738 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.234, + 0.785, + 0.261 + ], + "angle": 0, + "content": "16. He, K., Gkioxari, G., Dólár, P., Girshick, R.: Mask r-cnn. In: Proceedings of the IEEE International Conference on Computer Vision. pp. 2961-2969 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.262, + 0.785, + 0.303 + ], + "angle": 0, + "content": "17. Hu, H., Cui, J., Wang, L.: Region-aware contrastive learning for semantic segmentation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 16291-16301 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.304, + 0.785, + 0.346 + ], + "angle": 0, + "content": "18. Huang, C., Tang, H., Fan, W., Xiao, Y., Hao, D., Qian, Z., Terzopoulos, D., et al.: Partly supervised multi-task learning. In: 2020 19th IEEE International Conference on Machine Learning and Applications. pp. 769-774. IEEE (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.346, + 0.785, + 0.403 + ], + "angle": 0, + "content": "19. Imran, A.A.Z., Terzopoulos, D.: Semi-supervised multi-task learning with chest x-ray images. In: Machine Learning in Medical Imaging: 10th International Workshop, MLMI 2019, Held in Conjunction with MICCAI 2019, Shenzhen, China, October 13, 2019, Proceedings 10. pp. 151-159. Springer (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.403, + 0.785, + 0.445 + ], + "angle": 0, + "content": "20. Jin, X., Lan, C., Zeng, W., Chen, Z.: Global distance-distributions separation for unsupervised person re-identification. In: Proceedings of the European Conference on Computer Vision. pp. 735–751. Springer (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.445, + 0.785, + 0.472 + ], + "angle": 0, + "content": "21. Jin, Z., Chen, S., Chen, Y., Xu, Z., Feng, H.: Let segment anything help image dehaze. arXiv preprint arXiv:2306.15870 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.473, + 0.785, + 0.515 + ], + "angle": 0, + "content": "22. Kendall, A., Gal, Y., Cipolla, R.: Multi-task learning using uncertainty to weigh losses for scene geometry and semantics. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 7482-7491 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.516, + 0.785, + 0.572 + ], + "angle": 0, + "content": "23. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., Dollar, P., Girshick, R.: Segment anything. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 4015-4026 (October 2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.572, + 0.785, + 0.614 + ], + "angle": 0, + "content": "24. Latif, S., Rana, R., Khalifa, S., Jurdak, R., Epps, J., Schuller, B.W.: Multi-task semi-supervised adversarial autoencoding for speech emotion recognition. IEEE Transactions on Affective Computing 13(2), 992-1004 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.614, + 0.785, + 0.655 + ], + "angle": 0, + "content": "25. Li, W.H., Liu, X., Bilen, H.: Learning multiple dense prediction tasks from partially annotated data. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18879-18889 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.657, + 0.785, + 0.699 + ], + "angle": 0, + "content": "26. Liang, C., Wang, W., Miao, J., Yang, Y.: Gmmseg: Gaussian mixture based generative semantic segmentation models. Advances in Neural Information Processing Systems 35, 31360-31375 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.7, + 0.785, + 0.727 + ], + "angle": 0, + "content": "27. Liu, Q., Liao, X., Carin, L.: Semi-supervised multitask learning. Advances in Neural Information Processing Systems 20 (2007)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.728, + 0.785, + 0.755 + ], + "angle": 0, + "content": "28. Liu, S., James, S., Davison, A., Johns, E.: Auto-lambda: Disentangling dynamic task relationships. Transactions on Machine Learning Research (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.757, + 0.785, + 0.798 + ], + "angle": 0, + "content": "29. Liu, S., Johns, E., Davison, A.J.: End-to-end multi-task learning with attention. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1871-1880 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.799, + 0.785, + 0.84 + ], + "angle": 0, + "content": "30. Long, J., Shelhamer, E., Darrell, T.: Fully convolutional networks for semantic segmentation. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 3431-3440 (2015)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.493, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Region-aware Distribution Contrast" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.203 + ], + "angle": 0, + "content": "31. Lu, Y., Kumar, A., Zhai, S., Cheng, Y., Javidi, T., Feris, R.: Fully-adaptive feature sharing in multi-task networks with applications in person attribute classification. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 5334-5343 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.204, + 0.787, + 0.245 + ], + "angle": 0, + "content": "32. Misra, I., Shrivastava, A., Gupta, A., Hebert, M.: Cross-Stitch networks for multitask learning. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 3994-4003 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.246, + 0.786, + 0.273 + ], + "angle": 0, + "content": "33. Oord, A.v.d., Li, Y., Vinyals, O.: Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.273, + 0.787, + 0.314 + ], + "angle": 0, + "content": "34. Poggi, M., Aleotti, F., Tosi, F., Mattoccia, S.: On the uncertainty of self-supervised monocular depth estimation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3227-3237 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.315, + 0.787, + 0.342 + ], + "angle": 0, + "content": "35. Rüschendorf, L.: The Wasserstein distance and approximation theorems. Probability Theory and Related Fields 70(1), 117-129 (1985)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.343, + 0.787, + 0.383 + ], + "angle": 0, + "content": "36. Silberman, N., Hoiem, D., Kohli, P., Fergus, R.: Indoor segmentation and support inference from rgb images. In: Proceedings of the European Conference on Computer Vision. pp. 746-760. Springer (2012)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.384, + 0.787, + 0.424 + ], + "angle": 0, + "content": "37. Standley, T., Zamir, A., Chen, D., Guibas, L., Malik, J., Savarese, S.: Which tasks should be learned together in multi-task learning? In: International Conference on Machine Learning. pp. 9120-9132. PMLR (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.425, + 0.787, + 0.452 + ], + "angle": 0, + "content": "38. Tian, Y., Krishnan, D., Isola, P.: Contrastive multiview coding. In: Proceedings of the European Conference on Computer Vision. pp. 776-794. Springer (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.453, + 0.787, + 0.493 + ], + "angle": 0, + "content": "39. Vandenhende, S., Georgoulis, S., Van Gool, L.: MTI-Net: Multi-scale task interaction networks for multi-task learning. In: Proceedings of the European Conference on Computer Vision. pp. 527-543. Springer (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.494, + 0.787, + 0.536 + ], + "angle": 0, + "content": "40. Wang, X., Zhang, R., Shen, C., Kong, T., Li, L.: Dense contrastive learning for self-supervised visual pre-training. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3024-3033 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.536, + 0.787, + 0.59 + ], + "angle": 0, + "content": "41. Wu, L., Zhong, Z., Fang, L., He, X., Liu, Q., Ma, J., Chen, H.: Sparsely annotated semantic segmentation with adaptive gaussian mixtures. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15454-15464 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.591, + 0.787, + 0.633 + ], + "angle": 0, + "content": "42. Wu, Z., Xiong, Y., Yu, S.X., Lin, D.: Unsupervised feature learning via nonparametric instance discrimination. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 3733-3742 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.633, + 0.787, + 0.687 + ], + "angle": 0, + "content": "43. Xu, D., Ouyang, W., Wang, X., Sebe, N.: PAD-Net: Multi-tasks guided prediction-and-distillation network for simultaneous depth estimation and scene parsing. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 675-684 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.688, + 0.787, + 0.729 + ], + "angle": 0, + "content": "44. Ye, H., Xu, D.: Taskexpert: Dynamically assembling multi-task representations with memorial mixture-of-experts. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 21828-21837 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.729, + 0.787, + 0.771 + ], + "angle": 0, + "content": "45. Zamir, A.R., Sax, A., Cheerla, N., Suri, R., Cao, Z., Malik, J., Guibas, L.J.: Robust learning through cross-task consistency. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11197-11206 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.772, + 0.787, + 0.812 + ], + "angle": 0, + "content": "46. Zamir, A.R., Sax, A., Shen, W., Guibas, L.J., Malik, J., Savarese, S.: Taskonomy: Disentangling task transfer learning. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 3712-3722 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.813, + 0.787, + 0.84 + ], + "angle": 0, + "content": "47. Zhang, J., Fan, D.P., Dai, Y., Anwar, S., Saleh, F.S., Zhang, T., Barnes, N.: Uc-net: Uncertainty inspired rgb-d saliency detection via conditional variational autoen" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.35, + 0.127 + ], + "angle": 0, + "content": "M. Li et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.241, + 0.147, + 0.786, + 0.175 + ], + "angle": 0, + "content": "coders. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8582-8591 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.177, + 0.787, + 0.217 + ], + "angle": 0, + "content": "48. Zhang, L., Liu, X., Guan, H.: Automtl: A programming framework for automating efficient multi-task learning. Advances in Neural Information Processing Systems 35, 34216-34228 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.218, + 0.787, + 0.259 + ], + "angle": 0, + "content": "49. Zhang, M., Zhao, X., Yao, J., Yuan, C., Huang, W.: When noisy labels meet long tail dilemmas: A representation calibration method. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 15890-15900 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.26, + 0.787, + 0.314 + ], + "angle": 0, + "content": "50. Zhang, Z., Cui, Z., Xu, C., Yan, Y., Sebe, N., Yang, J.: Pattern-affinitive propagation across depth, surface normal and semantic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4106-4115 (2019)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.314 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/Region-aware Distribution Contrast_ A Novel Approach to Multi-Task Partially Supervised Learning/b09b7220-34e4-48d5-bc47-6cde62cb650d_origin.pdf b/2024/Region-aware Distribution Contrast_ A Novel Approach to Multi-Task Partially Supervised Learning/b09b7220-34e4-48d5-bc47-6cde62cb650d_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..1073db2cd10d067072edfbe3b1949417f938e810 --- /dev/null +++ b/2024/Region-aware Distribution Contrast_ A Novel Approach to Multi-Task Partially Supervised Learning/b09b7220-34e4-48d5-bc47-6cde62cb650d_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2f29f4268c70ba2629c9ec09a27c516d00eff8f57dee9dc9eed9e38089df60f7 +size 14014988 diff --git a/2024/Region-aware Distribution Contrast_ A Novel Approach to Multi-Task Partially Supervised Learning/full.md b/2024/Region-aware Distribution Contrast_ A Novel Approach to Multi-Task Partially Supervised Learning/full.md new file mode 100644 index 0000000000000000000000000000000000000000..f0173f403fbfbd5cdef4d0fcc203e3d8075983e7 --- /dev/null +++ b/2024/Region-aware Distribution Contrast_ A Novel Approach to Multi-Task Partially Supervised Learning/full.md @@ -0,0 +1,266 @@ +# Region-aware Distribution Contrast: A Novel Approach to Multi-Task Partially Supervised Learning + +Meixuan Li, Tianyu Li, Guoqing Wang*, Peng Wang, Yang Yang, and Jie Zou + +University of Electronic Science and Technology of China (UESTC) + +limeixuan0801@126.com, {cosmos.yu, p.wang6}@@hotmail.com, + +{gqwang0420, yang.yang, jie.zou}@uestc.edu.cn + +Abstract. In this study, we address the intricate challenge of multi-task dense prediction, encompassing tasks such as semantic segmentation, depth estimation, and surface normal estimation, particularly when dealing with partially annotated data (MTPSL). The complexity arises from the absence of complete task labels for each training image. Given the inter-related nature of these pixel-wise dense tasks, our focus is on mining and capturing cross-task relationships. Existing solutions typically rely on learning global image representations for global cross-task image matching, imposing constraints that, unfortunately, sacrifice the finer structures within the images. Attempting local matching as a remedy faces hurdles due to the lack of precise region supervision, making local alignment a challenging endeavor. The introduction of Segment Anything Model (SAM) sheds light on addressing local alignment challenges by providing free and high-quality solutions for region detection. Leveraging SAM-detected regions, the subsequent challenge lies in aligning the representations within these regions. Diverging from conventional methods that directly learn a monolithic image representation, our proposal involves modeling region-wise representations using Gaussian Distributions. Aligning these distributions between corresponding regions from different tasks imparts higher flexibility and capacity to capture intra-region structures, accommodating a broader range of tasks. This innovative approach significantly enhances our ability to effectively capture cross-task relationships, resulting in improved overall performance in partially supervised multi-task dense prediction scenarios. Extensive experiments conducted on two widely used benchmarks underscore the superior effectiveness of our proposed method, showcasing state-of-the-art performance even when compared to fully supervised methods. + +https://github.com/HereNowL/Region-aware-Distribution-Contrast + +Keywords: Multi-task Learning $\cdot$ Partially Supervised Learning $\cdot$ Scene Understanding $\cdot$ Contrastive Learning + +![](images/3e8368e9e8842c8887bc6a48ae776131aed8eca9183b0ff8e41cb6dc82697476.jpg) +(a) Contrastive Multiview Coding + +![](images/e41de6a7d875d566b8ffb340ef64c57013b9eaeab9ee834916b9fe09545c8a6e.jpg) +(b) Region-wise Distribution Contrast +Fig. 1: Multiview consistency and region consistency. (a) Illustration of contrastive multiview consistency [38]. (b) Illustration of region consistency, where $p_{i,s}$ and $p_{j,s}$ represent the region distribution of semantic segmentation, $p_{i,d}$ and $p_{j,d}$ denotes the region distribution of depth estimation. + +# 1 Introduction + +With the rapid advancement of deep learning, dense prediction tasks, including semantic segmentation, depth estimation, and surface normal estimation, have witnessed remarkable progress [11, 16, 30, 34, 47]. Considering the inherent interdependence among these dense prediction tasks [45, 46], there has been a growing interest in employing unified multi-task learning networks to jointly tackle different dense prediction tasks [7, 22, 28, 29, 31, 32, 39, 43, 50]. In contrast to inefficient single-task learning networks, multi-task learning networks effectively acquire shared features, enabling them to capture more generalizable information across various tasks while avoiding redundant training [39, 43, 50]. + +Recent advancements in multi-task dense prediction methods have primarily focused on two crucial aspects. The first aspect is dedicated to designing intricate network architectures to achieve effective multi-task learning [29, 32, 39, 43, 50], while the second aspect is devoted to devising optimal balancing strategies for loss functions in multi-task learning [7, 14, 22, 28, 29], aiming to mitigate the negative transfer across tasks. These directions mainly focus on fully supervised multi-task learning, where complete task labels are available for all training samples. However, acquiring pixel-level labels for dense prediction tasks is resource-intensive, particularly with multiple visual tasks. It is common to encounter scenarios where labels for certain tasks are missing or unreliable in some training samples [25], leading to a new research direction known as multi-task partially supervised learning (MTPSL). + +Since not all task labels are available for each training image in MTPSL, it is natural to formulate a multi-task learning framework which can leverage the implicit relationship between tasks with and without explicit labels. Dense prediction tasks, such as semantic segmentation, depth estimation, and normal prediction, exhibit a complementary co-existence relationship [37,45,46], where each task can serve as auxiliary information for others, offering valuable insights reciprocally. While existing approaches map predictions into a joint vector-wise semantic space for task alignment [25], we argue that this overlooks the underlying structural information inherent in the original image. A global vector + +might be insufficient to characterize the information in the entire scene, limiting its effectiveness in constraining cross-task relationships. Local matching, as shown in Fig. 1, with its intrinsic local properties, serves as a potential remedy. Nonetheless, the absence of precise region supervision presents a fundamental obstacle, thereby intensifying the challenges associated with its implementation. + +To tackle the outlined challenges, we introduce region-based cross-task alignment by leveraging the robust region segmentation capabilities inherent in Segment Anything Model (SAM) models. Utilizing well-segmented image regions facilitates the abstraction of region-level information, enabling the addition of cross-task consistency constraints. However, a subsequent crucial challenge arises in effectively representing these regional features to ensure the consistency and accuracy of cross-task alignment. In contrast to conventional methods that directly learn a unified image representation, we employ Gaussian distributions to model region representations, enhancing flexibility and improving the ability to capture intra-region structures. By adapting Gaussian distributions between corresponding regions from different tasks, our approach can be more broadly applicable to diverse task types, such as pixel-level classification tasks (e.g., semantic segmentation) and pixel-level regression tasks (e.g., depth estimation). Subsequently, we introduce distributional contrastive learning to explicitly perform region-wise cross-task alignment. This approach serves as a robust constraint, offering a meaningful way to mine relationships between tasks and enhance dense prediction results, particularly for those tasks lacking pixel-wise labels. + +In summary, our work masks the following contributions: + +- We tackle the challenge of multi-task partially supervised (MTPSL) dense prediction from a fresh perspective. Our approach involves extracting cross-task local alignment through the utilization of SAM's easily obtainable local regions. This innovative strategy has demonstrated notable in alleviating label shortages. +- We propose a novel region distribution contrast method for local alignment, which offers increased flexibility, robustness, and wide applicability across multiple tasks. +- We validate the effectiveness of our proposed method through extensive experiments conducted on two widely used benchmark datasets. Notably, our approach achieves state-of-the-art performance in partially supervised learning, while also showing great potential in fully supervised learning. + +# 2 Related Work + +Multi-task Supervised Learning. Multi-task learning utilizes a single model to simultaneously handle multiple visual tasks, offering advantages such as faster inference speed and more efficient utilization of input data. Some approaches [13, 32, 48] focus on designing the encoder of multi-task network, while other approaches pay more attention to designing complex decoders to generate task-specific predictions by utilizing shared features [2, 39, 43, 50]. Uncertainty [22] obtain task loss weights by considering the heteroscedastic uncertainty + +of each task, Grad-Norm [7] directly adjusts gradients to balance task losses, and DWA [14] learns time-varying average task weights by considering the loss variation rate of each task. It is important to note that these methods operate under a significant and strong assumption, namely that they are developed and studied based on the availability of labels for all tasks across all images. + +Multi-task Semi-supervised Learning and Partially-supervised Learning. Learning multi-task models on fully annotated data requires a large-scale labeled dataset, and the cost of collecting sufficient labeled data can be high. Therefore, two common scenarios are worthy of note in multi-task learning. One is semi-supervised multi-task learning, where the dataset consists of limited annotation information for all tasks and a large amount of unlabeled data. A bunch of methods have been proposed for semi-supervised multi-task learning [8,18,19,24,27]. In recent works [8,18,19,24], regularization terms are applied to unlabeled samples from each task to encourage consistent predictions when the input is perturbed. Another scenario involves partial-supervision multi-task learning, where the dataset lacks sufficient labels for each task, and not all tasks have labels for every image. Li et al. [25] maps tasks into high-dimensional vectors for cross-task alignment. However, vector representations are insufficient and do not capture local-level alignment. In this paper, we propose leveraging distributions as a substitute for vector representations of task features and employing contrastive learning to achieve region-level alignment across tasks. + +Contrastive Learning for Vision. Recently, contrastive learning has made significant advancements in unsupervised learning [5,6,15,33,42]. DenseCL [40] and RegionContrast [17] have been utilized for semantic segmentation tasks using contrastive learning at the pixel and region levels respectively, achieving excellent results. For depth estimation, WCL [12] transformed depth from continuous values to discrete values and constructed contrastive losses based on windows to form positive and negative sample pairs. CMC [38] was the first to propose that different viewpoints of the same image should be mapped to nearby positions in a high-dimensional space, while different images should be mapped to distant positions, thus ensuring multiview consistency. Inspired by CMC, we believe that multiple viewpoints within the same local region should also exhibit consistency, while differing from other regions. In this paper, we leverage this idea to achieve region-level contrast for MTPSL. + +Gaussian Distribution. Recently, Gaussian distribution has been widely employed in computer vision to characterize the distribution of features [3, 4, 20, 26, 41, 49]. GMMSeg [26] utilized the Expectation-Maximization algorithm to construct Gaussian Mixture Models for each class in semantic segmentation, capturing class-conditional densities. AGMM [41] constructed an adaptive Gaussian Mixture Model for sparse annotation in semantic segmentation by incorporating labeled pixels and their similar unlabeled counterparts. DUL [3] employed mean and variance to characterize the distribution of faces for face recognition. Gaussian distributions have the capability to probabilistically describe the distribution characteristics of different visual tasks. However, to the best of our knowledge, Gaussian distributions have not been applied in MTPSL. + +![](images/e0634edcee03b66869a58f843e1e3be0b743524f41a4e2cfab06a731939c7ceb.jpg) +Fig. 2: Illustration of region-aware distribution contrast learning method for MTPSL. During training, supervised constraints $L_{Sup}$ are applied to the annotated task $s$ . For task $s$ and unlabelled task $t$ , the $a_{\theta}$ map the true label $y_{s}$ and the prediction $\hat{y}_{t}$ in the high-dimensional space respectively and then model the region of task-specific features extracted using SAM as a Gaussian distribution. Contrastive learning is then employed to minimize the distance between distributions of the same region across different tasks while maximizing the distance to distributions of other regions. + +# 3 Method + +# 3.1 Preliminaries + +Problem Setup. Considering the problem of multi-task learning involving $K$ tasks, where $K \geq 2$ . A training dataset $S = \{x_{i}\}_{i=1}^{N}$ is given with $N$ partially labelled samples, indicating that for each training sample $x_{i}$ , only a subset of the $K$ tasks are provided true labels. Let $P_{i}$ represents the number of labeled tasks for $x_{i}$ and $Q_{i}$ represents the number of unlabeled tasks, such that $P_{i} + Q_{i} = K$ . When $P_{i}$ equals $K$ for all $x_{i} \in S$ , it indicates fully supervised multi-task learning. Conversely, when $Q_{i}$ equals $K$ for all $x_{i} \in S$ , it implies that no task labels are available, leading to unsupervised multi-task learning. In this paper, we focus on addressing partially supervised multi-task learning, where each training image $x_{i}$ can obtain labels for at least one task ( $P_{i} \geq 1$ ). + +# 3.2 Overview of Our Method + +As shown in Fig. 2, our method focuses on utilizing Gaussian distributions to represent features in local regions for contrastive learning, achieving better distribution consistency across tasks at the region-level for MTPSL. Assuming that task $s$ has label while task $t$ does not, our objective is to achieve alignment between the true label $y_{s}$ of task $s$ and the prediction $\hat{y}_t$ of task $t$ at the region level using contrastive learning. + +In the training stage, to accurately and efficiently obtain region-level features for prediction, we simultaneously feed the input image to both the multi-task learning (MTL) network and the pre-trained SAM. Predictions are obtained + +![](images/fbb93ca70c827e550da39fb85b3f29a8a2c26074e15c8d06cddb311332949774.jpg) +Fig. 3: Illustration of region-level cross-task consistency. Initially, SAM predictions are employed to extract regions from the features of the particular task. Following that, these regions are modeled as Gaussian distributions. Finally, contrastive learning is utilized to minimize the distance between regions of the same region across different tasks and maximize the distance to regions of other regions. + +from both MTL network and SAM. For tasks with labeled data, we apply fully-supervised loss $L_{Sup}$ as explicit training supervision. For tasks without label, an auxiliary network $a_{\theta}$ maps the unlabelled task prediction $\hat{y}_t$ and the annotated task label $y_s$ to the same high-dimensional feature map space, and use $L_{RC}$ to achieve cross-task consistency at the region level. Specifically, We first extract region-wise features in $\hat{y}_t$ and $y_s$ based on the segmented regions identified by SAM, then model them as Gaussian distributions. After acquiring the Gaussian distributions for features in each region of each task, we utilize contrastive learning to minimize the distance between the Gaussian distributions of features in the same region across different tasks and maximize the distance from the Gaussian distributions of features from different regions. Finally, the optimization of our method can be defined as: + +$$ +L o s s = L _ {S u p} + L _ {R C}. \tag {1} +$$ + +In the inference stage, as illustrated in Fig. 2, the input image is fed only into the MTL network to generate predictions for multiple tasks, notably, with SAM no longer being utilized. + +# 3.3 Region-aware Contrastive Learning + +In this section, we explain the implementation details to achieve region-level cross-task consistency. + +Region-level Features Extraction. To realize the region-level cross-task consistency, it is essential to informatively choose regions for feature extraction. Considering the robust capabilities of SAM for image segmentation [23], we employ it to generate the regions for each training sample, leveraging its ability to provide valuable prior knowledge. As shown in Fig. 2, to obtain fine-grained + +semantics and fine-grained edges of regions, we fed the image into MTL network and pre-trained SAM simultaneously. SAM performs fine-grained segmentation of each object in the image, and its output masks contain the ID of each area and its corresponding position in the image. Inspired by [21], for convenient region-wise feature extraction in task-specific feature maps, we compute the predicted masks of SAM as a grayscale image, where each pixel value corresponds to the area ID returned by SAM, indicating its category. Leveraging the grayscale masks of SAM, we split the image into different regions and extract region-wise features within feature maps of different tasks. + +Gaussian Distribution Modeling. After obtaining the region-wise features for each task, we consider modeling the region-wise features as Gaussian distributions. The utilization of Gaussian distributions represents the features of a region in a probabilistic manner, which provides a more comprehensive depiction of the region's variations and compensates for the shortcomings of pixel-level hard alignment. Specifically, we model each region $r_i$ from the task-specific feature map $f_t$ as a Gaussian distribution: + +$$ +p _ {i, t} = p \left(r _ {i} \mid f _ {t}\right) = N \left(r _ {i}; \boldsymbol {\mu} _ {i, t}, \boldsymbol {\Sigma} _ {i, t}\right), \tag {2} +$$ + +where $\pmb{\mu}_{i,t}$ and $\pmb{\Sigma}_{i,t}$ represent mean and covariance matrix respectively, calculated from the region $r_i$ in the feature map $f_{t}$ , as shown in Fig. 3. + +Region Distribution Contrast. Once we have obtained the Gaussian distributions for each region, our objective is to utilize contrastive learning to minimize the distance between Gaussian distributions of the same region across different tasks and maximize the distance between Gaussian distributions of different regions. Hence, the region-wise Gaussian distribution contrastive loss for the region distribution $p_{i,t}$ can be defined as follows: + +$$ +L _ {i, t} ^ {N C E} = - \sum_ {k _ {+}} \log \frac {\operatorname {s i m} \left(p _ {i , t} , k _ {+}\right)}{\operatorname {s i m} \left(p _ {i , t} , k _ {+}\right) + \sum_ {k _ {-}} \operatorname {s i m} \left(p _ {i , t} , k _ {-}\right)}, \tag {3} +$$ + +$$ +\operatorname {s i m} \left(p _ {i, t}, k\right) = \exp \left(- W _ {\text {d i s t a n c e}} \left(p _ {i, t}, k\right) / \tau\right), \tag {4} +$$ + +where $k_{+} \in \{p_{i,s}\}_{s \neq t}$ , $k_{-} \in \{p_{j,*}\}_{j \neq i}$ , * denotes any task, sim denotes the exponential equation of the Wasserstein distance, and $\tau$ is the temperature parameter. Wasserstein distance [35] is a metric used to quantify the dissimilarity between two probability distributions and well-suited for accurately measuring the distance between discrete and continuous distributions, as it accounts for the underlying structure and geometry of the data. It measures the minimum amount of work required to transform one distribution into another and can be calculated as follows: + +$$ +\begin{array}{l} W _ {\text {d i s t a n c e}} \left(p _ {i, t}, k\right) = \left\| \boldsymbol {\mu} _ {i, t} - \boldsymbol {\mu} _ {k} \right\| ^ {2} + \operatorname {T r} \left(\boldsymbol {\Sigma} _ {i, t}\right) \\ + \operatorname {T r} \left(\boldsymbol {\Sigma} _ {k}\right) - 2 \operatorname {T r} \left(\left(\boldsymbol {\Sigma} _ {i, t} \boldsymbol {\Sigma} _ {k}\right) ^ {1 / 2}\right), \tag {5} \\ \end{array} +$$ + +where $\mathrm{Tr}$ represents the trace of the matrix, $\pmb{\mu}_k$ and $\pmb{\Sigma}_k$ represent the mean and covariance matrix of sample $k$ . Based on $L^{NCE}$ , the region-wise contrastive loss + +$L_{RC}$ can be defined as follows: + +$$ +L _ {R C} = \frac {1}{K M} \sum_ {t} ^ {K} \sum_ {i} ^ {M} L _ {i, t} ^ {N C E}, \tag {6} +$$ + +where $M$ is the number of regions in the feature map $t$ . + +By minimizing $L_{RC}$ , the network learns to contrast the distributions from the same region with those from different regions, thereby enabling better cross-task consistency at the region level. + +Alternative Local Extraction Strategy. Here we explore an alternative local extraction strategy to investigate local cross-task consistency. To extract local information from the feature map, the simplest approach is to divide the map into patches, where patches at the same position serve as positive samples, and their distances are minimized. Conversely, patches from other positions act as negative samples, and their distances are maximized. We refer to this local extraction strategy as patch-aware contrast, which will be compared to the region-aware contrast method in Sec. 4. + +Alternative Local Contrast Strategy. Alternatively, the method of modeling regions as Gaussian distributions and performing contrastive learning can be replaced with other contrastive strategies. One strategy is pixel-to-pixel level region contrast, where each pixel within the region is pulled closer to the corresponding pixel in the positive sample region and pushed far apart from each pixel in the negative sample region. We refer to this as pixel contrast. Another strategy is to map the region to a vector and bring it closer to the vector of the positive sample region while pushing it away from the vector of the negative sample region. This strategy is referred to as vector contrast. We compare these alternative methods in Sec. 4. + +# 4 Experiments + +Datasets. We evaluated our method on two standard dense prediction multi-task datasets: NYU-V2 [36] and Cityscapes [9]. NYU-V2 is an indoor dataset of natural scenes, including three dense prediction tasks: semantic segmentation, depth estimation, and surface normal estimation. The semantic segmentation task in NYU-V2 contains 13 categories, depth estimation is provided by Microsoft Kinect, and surface normal estimation is provided by [10]. As [25, 29], all images were resized to $288 \times 384$ for training and evaluation purposes. Cityscapes is a well-known dataset for autonomous driving, consisting of outdoor street scenes with two tasks: semantic segmentation and depth estimation. Following [25, 29], we used 7-class semantic segmentation annotations for evaluation and resized all images to $128 \times 256$ to improve training speed. + +Experimental Setting. Following [25], we evaluated our method in two partially supervised multi-task settings: the one label setting and the random label setting. In the one label setting, each image is assigned a random task with a label. In the random label setting, each image is assigned random labels for + +Table 1: Quantitative comparison on the NYU-V2 dataset. $\uparrow (\downarrow)$ denotes that, larger (smaller) values lead to better quality. * represents the available partially supervised SOTA. * represents the available fully supervised SOTA. + denotes the percentage improvement over the SOTA performance, corresponding to $\uparrow (\downarrow)$ . The bold denotes the best. + +
Label settingMethodBackboneExtraction strategyTasks
PatchRegionSemantic(mIoU)↑Depth(aErr)↓Normals(mErr)↓
fullTaskExpert* [44]ViT-Large57.580.37305.90
onelabelLi et al. [25]SegNet30.360.608832.08
onelabelDejaVu* [1]SegNet31.020.595932.15
onelabelGaussian Contrast(Ours)SegNet✓31.79 (+2.48)0.5835 (+2.25)30.38 (+5.51)
onelabelLi et al. * [25]HRNet1839.420.507114.79
onelabelDejaVu [1]HRNet18———
onelabelGaussian ContrastHRNet18✓41.56 (+5.43)0.4884 (+3.69)7.71 (+47.87)
onelabelGaussian Contrast(Ours)HRNet18✓42.28 (+7.26)0.4641 (+8.48)4.86 (+67.14)
onelabelTaskExpert* [44]ViT-Large49.530.430511.12
onelabelTaskExpert [44]+OursViT-Large✓55.81 (+12.68)0.4092 (+4.95)8.46 (+23.92)
random labelLi et al. [25]SegNet34.260.578731.06
random labelDejaVu* [1]SegNet35.720.566529.82
random labelGaussian Contrast(Ours)SegNet✓37.40 (+4.70)0.5428 (+4.18)28.97 (+2.85)
random labelLi et al. * [25]HRNet1841.350.484514.34
random labelDejaVu [1]HRNet18———
random labelGaussian ContrastHRNet18✓45.79 (+10.74)0.4619 (+4.66)7.37 (+48.61)
random labelGaussian Contrast(Ours)HRNet18✓46.21 (+11.75)0.4482 (+7.49)4.49 (+68.69)
random labelTaskExpert* [44]ViT-Large56.320.42826.47
random labelTaskExpert [44]+OursViT-Large✓57.14 (+1.46)0.3977 (+7.12)4.78 (+26.12)
+ +multiple tasks, with at least one task having a label and at most $K - 1$ tasks having labels, where $K$ represents the total number of tasks in the multi-task setup. + +Implementation Details and Evaluation Metrics. All our experiments were conducted on the NVIDIA A100 GPU. For NYU-V2, we trained the models for 20 epochs, while for Cityscapes, we trained them for 100 epochs. Following [25, 29], we employed cross-entropy loss for semantic segmentation, L1 loss for depth estimation, and cosine similarity loss for surface normal estimation. We utilized the exact same evaluation metrics as mentioned in [25, 29]. For the semantic segmentation task, we employed the mean intersection over union (mIoU) metric. The absolute error (aErr) was used to evaluate the depth estimation task, and the mean error (mErr) was employed for surface normal estimation. + +# 4.1 Quantitative Evaluation + +Quantitative Results on NYU-V2 Dataset. The evaluation results are shown in Tab. 1. Due to the absence of pre-trained parameters in Li et al. [25], the training process exhibits slower convergence. To expedite the convergence speed, we replaced the backbone of Li et al. [25] with HRNet18, which possesses a smaller training parameter size but utilizes pre-trained weights. The comparisons between Li et al. [25] with SegNet as the backbone and Li et al. [25] with HRNet18 as the backbone are presented in the first row and the fourth row, respectively, for each label setting. Since the source code of DejaVu [1] is not publicly available, we are unable to replace its backbone with HRNet18 for a fair comparison. To resolve this issue, we provide the results of our method based + +Table 2: Quantitative comparison on the Cityscapes dataset. $\uparrow (\downarrow)$ denotes that, larger (smaller) values lead to better quality. * represents the available SOTA. + denotes the percentage improvement over the SOTA performance, corresponding to $\uparrow$ ( $\downarrow$ ). The bold denotes the best. + +
Label settingMethodBackboneContrast strategyTasks
Vector Pixel GaussianSemantic(mIoU)↑Depth(aErr) ↓
onelabelLi et al. [25]SegNet74.900.0161
onelabelLi et al. * [25]HRNet1881.730.0157
onelabelRegion-aware ContrastHRNet18✓82.76 (+1.26)0.0145 (+7.64)
onelabelRegion-aware ContrastHRNet1883.20 (+1.80)0.0141 (+10.19)
onelabelRegion-aware Contrast (Ours)HRNet1883.92 (+2.68)0.0121 (+22.73)
+ +Table 3: Ablation study on the NYU-V2 dataset to explore the performance of different local extraction strategies and different local contrast strategies. + +
Label settingExtraction strategyContrast strategyTasks
PatchRegionVector Pixel GaussianSemantic(mIoU)↑Depth(aErr)↓Normals(mErr)↓
onelabel✓✓40.240.4998
onelabel✓✓41.500.4934
onelabel✓✓41.560.4884
onelabel✓✓41.840.4727
onelabel✓✓42.050.4663
onelabel✓✓42.280.4641
random label✓✓43.980.4795
random label✓✓44.620.4777
random label✓✓45.790.4619
random label✓✓45.830.4630
random label✓✓46.410.4518
random label✓✓46.210.4482
+ +![](images/a439ac06aa26d330254df83d7af8e02e21e825d90577a20bf0d36605342e9a44.jpg) +Fig. 4: Qualitative results of onelabel setting on NYU-V2. The first row shows the input image, the second row represents the ground-truth or predictions of semantic segmentation, the third row plots the ground-truth or predictions of depth estimation, and the final row presents the ground-truth or predictions of surface normal estimation. + +on SegNet, which utilizes the same backbone. In all the methods compared using the same backbone, our approach demonstrates significant advantages. In the one label setting, the patch-aware Gaussian contrast method demonstrates + +![](images/e030ba3de3ea14eca303112f5385202e20fe63ec6a7820c2772c12b33d83c7a1.jpg) +Fig. 5: Qualitative results of random label setting on NYU-V2. The first row shows the input image, the second row represents the ground-truth or predictions of semantic segmentation, the third row plots the ground-truth or predictions of depth estimation, and the final row presents the ground-truth or predictions of surface normal estimation. + +significant improvements across all three tasks compared to Li et al. [25] across all three tasks, as it achieves a certain degree of local-level cross-task alignment. Furthermore, the region-aware Gaussian contrast method further enhances performance, achieving improvements of $2.48\%$ , $2.25\%$ and $5.51\%$ over the SOTA [1] across the three tasks, owing to the achievement of cross-task consistency at the object level. In the random label setting, the superiority of the region-aware Gaussian contrast method can be observed, as it outperforms other methods significantly across all three tasks, yielding improvements of $4.70\%$ , $4.18\%$ and $2.85\%$ , respectively, over the SOTA [1]. + +Furthermore, we provide a comparison with the state-of-the-art fully supervised method, TaskExpert [44]. Tab. 1 clearly shows that our method significantly improves the performance of TaskExpert in partially supervised scenarios. Particularly in the random setting, when combined with our method, TaskExpert achieves even better performance than the fully supervised TaskExpert in the surface normal task. This once again demonstrates the effectiveness and wide applicability of our approach. + +Quantitative Results on Cityscapes Dataset. The evaluation results are describe in Tab. 2. Similar to NYU-V2, we first compared Li et al. [25] with HRNet18 as the backbone. Subsequently, we discussed the results of the region-aware contrast method under different contrast strategies. It is observed that Gaussian contrast demonstrates improvements in semantic segmentation and depth estimation compared to vector-based methods, indicating that representing regions using Gaussian distributions provides a more comprehensive and accurate representation than using vectors. In terms of computational speed, the + +Table 4: Ablation study on the NYU-V2 dataset to investigate the contribution of SAM in the fully supervised setting. + +
MethodVector GaussianSemantic(mIoU)↑Depth(aErr)↓Normals(mErr)↓
Region-aware Contrast✓49.000.42068.1878
Region-aware Contrast (Ours)✓51.140.39824.8108
+ +Gaussian contrast method is faster and more efficient, while still performing on par with pixel-based contrast methods. + +# 4.2 Qualitative Evaluation + +The qualitative evaluation of onelabel setting, random label setting on the NYU-V2 dataset, and onelabel setting on the Cityscapes dataset are shown in Fig. 4, Fig. 5 and Fig. 6, respectively. We present a comparative analysis of the results from two sets of images for each setting of each dataset. As indicated by Fig. 4 and Fig. 5, it can be observed that our method exhibits superior accuracy and smoothness in semantic segmentation, which is particularly evident in Fig. 5. Regarding depth estimation, our method, which is based on region-aware Gaussian distribution contrast, achieves more accurate predictions when objects undergo changes compared to Li et al. [25]. The latter fails to capture such variations effectively. As for surface normal estimation, our method outperforms Li et al. [25] overall and demonstrates finer contour handling. These observations highlight the advantages of our approach over Li et al. [25] in terms of classification accuracy, smoothness in semantic segmentation, accurate depth estimation in the presence of object changes, and finer handling of surface normals. + +We further exhibit the multi-task prediction results on the Cityscapes dataset in Fig. 6. Benefited from our proposed region-aware distribution contrast learning method, both the semantic segmentation and depth estimation tasks exhibit improved capability in capturing object edges and fine-grained details compared to Li et al. [25]. Additionally, there are significant overall performance improvement at a global level. + +# 4.3 Ablation Study + +We conduct ablation study on NYU-V2 dataset to prove the effectiveness of our method from various aspects. + +Local Extraction Strategy. Tab. 3 provides a comparison between different local extraction strategies. All methods are compared based on HRNet18 as the backbone. It can be observed that in both the onelabel setting and random label setting, regardless of the contrast strategy employed, the region-aware contrast method significantly outperforms the patch-aware contrast method. This is because the region-aware method takes into account the objects present in the image and performs contrast based on objects, rather than simply comparing patches that may contain multiple objects. + +![](images/87aec302bd7afacf1602e4c7516c97e91d078af20e021e6b2fde90c6fb645094.jpg) +Fig.6: Qualitative results of onelabel setting on Cityscapes. The first row shows the input image, the second row represents the ground-truth or predictions of semantic segmentation, and the final row plots the ground-truth or predictions of depth estimation. + +Table 5: Ablation study on the NYU-V2 dataset to investigate the performance of various distance measurements. + +
Label settingMethodTasks
Semantic(mIoU)↑Depth(aErr)↓Normals(mErr)↓
onelabelKL divergence41.150.47394.94
onelabelWasserstein distance (Ours)42.280.46414.86
random labelKL divergence45.530.45395.04
random labelWasserstein distance (Ours)46.210.44824.49
+ +Local Contrast Strategy. Tab. 3 also provides a comparison between different local contrast strategies. All methods are similarly compared based on HRNet18. It can be observed that regardless of the label setting and the local extraction strategy employed, the Gaussian distribution contrast consistently outperforms the vector contrast and approaches or even surpasses the pixel contrast. This is because multi-task learning scenarios involve discrete and continuous dense predictions, utilizing Gaussian distributions to represent regions for contrast allows for better cross-task alignment in the overall distribution. In contrast, the hard alignment strategy of pixel contrast may lead to inaccuracies. + +Contribution of SAM. As shown in Tab. 2, Tab. 3, and Tab. 6 in our article, although the vector contrast approach also incorporates SAM, it fails to yield satisfactory results. In contrast, the adoption of distribution contrast exhibits a significant improvement over the vector contrast approach. To further demonstrate this point, we provide a comparison of results using semantic segmentation labels instead of SAM in fully supervised setting, as shown in Tab. 4. This further emphasizes that solely introducing fine-grained segmentation results is insufficient, highlighting the importance of the distribution contrast method. + +Table 6: Ablation study on the NYU-V2 dataset to investigate the performance of various methods in the fully supervised setting. + +
MethodContrast strategyTasks
Vector Pixel GaussianSemantic(mIoU)↑Depth(aErr)↓Normals(mErr)↓
MTL46.380.46606.7298
Li et al. [25]47.200.44775.1192
Region-aware Contrast✓47.640.42315.1072
Region-aware Contrast✓50.060.41684.5318
Region-aware Contrast (Ours)✓50.830.40513.2878
+ +Distance Measurement for Distributions. Tab. 5 presents a comparison of different distance measurements. While KL divergence is commonly used to calculate the distance between Gaussian distributions, it fails to accurately reflect the distance between discrete and continuous distributions, which is crucial in the context of multi-task alignment. On the other hand, Wasserstein distance has the advantage of measuring the distance between two non-overlapping distributions. The experimental results in Tab. 5 demonstrate that Wasserstein distance is better suited for evaluating the distance between different task distributions. More detailed information can be found in the supplementary materials. + +Fully Supervised Setting. Tab. 6 presents the comparison of our method in the fully supervised setting. All methods are based on HRNet18. Compared to other methods, the region-aware distribution contrast method once again demonstrates superiority, outperforming significantly in all three tasks. This is because Gaussian distributions can characterize the distribution characteristics of different tasks and facilitate accurate and efficient alignment with both vector and pixel-based approaches. + +# 5 Conclusion + +In this paper, we propose a novel region-aware distribution contrast learning method for MTPSL. To facilitate cross-task alignment at the region level, predictions from different tasks are initially mapped into a joint feature map space. Simultaneously, SAM is employed to generate regions for each training sample, leveraging its ability to provide valuable prior knowledge. Subsequently, region-wise features are modeled by fitting a Gaussian distribution. The alignment among regions is achieved by minimizing the disparity among distributions from different tasks within the same region and maximizing the distance between distributions from different regions. The effectiveness of our method is demonstrated on two standard multi-task datasets through extensive experiments, outperforming the state-of-the-art by a significant margin. Although our method is designed for MTPSL, it can be applied to general multi-task learning problems to leverage the intrinsic relationships across different tasks, as demonstrated in the fully supervised experiments in Sec. 4.3. Moreover, we believe that the proposed region-aware distribution contrast learning method provides a promising way for solving more general multi-modal problems, and we will investigate this in our future work. + +Acknowledgements: This work was supported in part by the National Natural Science Foundation of China under grant U23B2011, 62102069, U20B2063 and 62220106008, the Key R&D Program of Zhejiang under grant 2024SSYS0091, the Sichuan Science and Technology Program under Grant 2024NSFTD0034. + +# References + +1. Borse, S., Das, D., Park, H., Cai, H., Garrepalli, R., Porikli, F.: Dejavu: Conditional regenerative learning to enhance dense prediction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 19466-19477 (2023) +2. Brüggemann, D., Kanakis, M., Obukhov, A., Georgoulis, S., Van Gool, L.: Exploring relational context for multi-task dense prediction. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 15869-15878 (2021) +3. Chang, J., Lan, Z., Cheng, C., Wei, Y.: Data uncertainty learning in face recognition. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5710-5719 (2020) +4. Chen, S., Xie, G., Liu, Y., Peng, Q., Sun, B., Li, H., You, X., Shao, L.: Hsva: Hierarchical semantic-visual adaptation for zero-shot learning. Advances in Neural Information Processing Systems 34, 16622-16634 (2021) +5. Chen, T., Kornblith, S., Norouzi, M., Hinton, G.: A simple framework for contrastive learning of visual representations. In: International Conference on Machine Learning. pp. 1597-1607. PMLR (2020) +6. Chen, X., Fan, H., Girshick, R., He, K.: Improved baselines with momentum contrastive learning. arXiv preprint arXiv:2003.04297 (2020) +7. Chen, Z., Badrinarayanan, V., Lee, C.Y., Rabinovich, A.: GradNorm: Gradient normalization for adaptive loss balancing in deep multi-task networks. In: International Conference on Machine Learning. pp. 794-803. PMLR (2018) +8. Chen, Z., Zhu, L., Wan, L., Wang, S., Feng, W., Heng, P.A.: A multi-task mean teacher for semi-supervised shadow detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5611-5620 (2020) +9. Cordts, M., Omran, M., Ramos, S., Rehfeld, T., Enzweiler, M., Benenson, R., Franke, U., Roth, S., Schiele, B.: The cityscapes dataset for semantic urban scene understanding. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 3213-3223 (2016) +0. Eigen, D., Fergus, R.: Predicting depth, surface normals and semantic labels with a common multi-scale convolutional architecture. In: Proceedings of the IEEE International Conference on Computer Vision. pp. 2650-2658 (2015) +1. Eigen, D., Puhrsch, C., Fergus, R.: Depth map prediction from a single image using a multi-scale deep network. Advances in Neural Information Processing Systems 27 (2014) +2. Fan, R., Poggi, M., Mattoccia, S.: Contrastive learning for depth prediction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3225-3236 (2023) +3. Gao, Y., Ma, J., Zhao, M., Liu, W., Yuille, A.L.: Nddr-cnn: Layerwise feature fusing in multi-task cnns by neural discriminative dimensionality reduction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3205-3214 (2019) + +14. Guo, M., Haque, A., Huang, D.A., Yeung, S., Fei-Fei, L.: Dynamic task prioritization for multitask learning. In: Proceedings of the European Conference on Computer Vision. pp. 270-287 (2018) +15. He, K., Fan, H., Wu, Y., Xie, S., Girshick, R.: Momentum contrast for unsupervised visual representation learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9729-9738 (2020) +16. He, K., Gkioxari, G., Dólár, P., Girshick, R.: Mask r-cnn. In: Proceedings of the IEEE International Conference on Computer Vision. pp. 2961-2969 (2017) +17. Hu, H., Cui, J., Wang, L.: Region-aware contrastive learning for semantic segmentation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 16291-16301 (2021) +18. Huang, C., Tang, H., Fan, W., Xiao, Y., Hao, D., Qian, Z., Terzopoulos, D., et al.: Partly supervised multi-task learning. In: 2020 19th IEEE International Conference on Machine Learning and Applications. pp. 769-774. IEEE (2020) +19. Imran, A.A.Z., Terzopoulos, D.: Semi-supervised multi-task learning with chest x-ray images. In: Machine Learning in Medical Imaging: 10th International Workshop, MLMI 2019, Held in Conjunction with MICCAI 2019, Shenzhen, China, October 13, 2019, Proceedings 10. pp. 151-159. Springer (2019) +20. Jin, X., Lan, C., Zeng, W., Chen, Z.: Global distance-distributions separation for unsupervised person re-identification. In: Proceedings of the European Conference on Computer Vision. pp. 735–751. Springer (2020) +21. Jin, Z., Chen, S., Chen, Y., Xu, Z., Feng, H.: Let segment anything help image dehaze. arXiv preprint arXiv:2306.15870 (2023) +22. Kendall, A., Gal, Y., Cipolla, R.: Multi-task learning using uncertainty to weigh losses for scene geometry and semantics. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 7482-7491 (2018) +23. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., Dollar, P., Girshick, R.: Segment anything. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 4015-4026 (October 2023) +24. Latif, S., Rana, R., Khalifa, S., Jurdak, R., Epps, J., Schuller, B.W.: Multi-task semi-supervised adversarial autoencoding for speech emotion recognition. IEEE Transactions on Affective Computing 13(2), 992-1004 (2020) +25. Li, W.H., Liu, X., Bilen, H.: Learning multiple dense prediction tasks from partially annotated data. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18879-18889 (2022) +26. Liang, C., Wang, W., Miao, J., Yang, Y.: Gmmseg: Gaussian mixture based generative semantic segmentation models. Advances in Neural Information Processing Systems 35, 31360-31375 (2022) +27. Liu, Q., Liao, X., Carin, L.: Semi-supervised multitask learning. Advances in Neural Information Processing Systems 20 (2007) +28. Liu, S., James, S., Davison, A., Johns, E.: Auto-lambda: Disentangling dynamic task relationships. Transactions on Machine Learning Research (2022) +29. Liu, S., Johns, E., Davison, A.J.: End-to-end multi-task learning with attention. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1871-1880 (2019) +30. Long, J., Shelhamer, E., Darrell, T.: Fully convolutional networks for semantic segmentation. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 3431-3440 (2015) + +31. Lu, Y., Kumar, A., Zhai, S., Cheng, Y., Javidi, T., Feris, R.: Fully-adaptive feature sharing in multi-task networks with applications in person attribute classification. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 5334-5343 (2017) +32. Misra, I., Shrivastava, A., Gupta, A., Hebert, M.: Cross-Stitch networks for multitask learning. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 3994-4003 (2016) +33. Oord, A.v.d., Li, Y., Vinyals, O.: Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748 (2018) +34. Poggi, M., Aleotti, F., Tosi, F., Mattoccia, S.: On the uncertainty of self-supervised monocular depth estimation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3227-3237 (2020) +35. Rüschendorf, L.: The Wasserstein distance and approximation theorems. Probability Theory and Related Fields 70(1), 117-129 (1985) +36. Silberman, N., Hoiem, D., Kohli, P., Fergus, R.: Indoor segmentation and support inference from rgb images. In: Proceedings of the European Conference on Computer Vision. pp. 746-760. Springer (2012) +37. Standley, T., Zamir, A., Chen, D., Guibas, L., Malik, J., Savarese, S.: Which tasks should be learned together in multi-task learning? In: International Conference on Machine Learning. pp. 9120-9132. PMLR (2020) +38. Tian, Y., Krishnan, D., Isola, P.: Contrastive multiview coding. In: Proceedings of the European Conference on Computer Vision. pp. 776-794. Springer (2020) +39. Vandenhende, S., Georgoulis, S., Van Gool, L.: MTI-Net: Multi-scale task interaction networks for multi-task learning. In: Proceedings of the European Conference on Computer Vision. pp. 527-543. Springer (2020) +40. Wang, X., Zhang, R., Shen, C., Kong, T., Li, L.: Dense contrastive learning for self-supervised visual pre-training. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3024-3033 (2021) +41. Wu, L., Zhong, Z., Fang, L., He, X., Liu, Q., Ma, J., Chen, H.: Sparsely annotated semantic segmentation with adaptive gaussian mixtures. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15454-15464 (2023) +42. Wu, Z., Xiong, Y., Yu, S.X., Lin, D.: Unsupervised feature learning via nonparametric instance discrimination. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 3733-3742 (2018) +43. Xu, D., Ouyang, W., Wang, X., Sebe, N.: PAD-Net: Multi-tasks guided prediction-and-distillation network for simultaneous depth estimation and scene parsing. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 675-684 (2018) +44. Ye, H., Xu, D.: Taskexpert: Dynamically assembling multi-task representations with memorial mixture-of-experts. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 21828-21837 (2023) +45. Zamir, A.R., Sax, A., Cheerla, N., Suri, R., Cao, Z., Malik, J., Guibas, L.J.: Robust learning through cross-task consistency. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11197-11206 (2020) +46. Zamir, A.R., Sax, A., Shen, W., Guibas, L.J., Malik, J., Savarese, S.: Taskonomy: Disentangling task transfer learning. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 3712-3722 (2018) +47. Zhang, J., Fan, D.P., Dai, Y., Anwar, S., Saleh, F.S., Zhang, T., Barnes, N.: Uc-net: Uncertainty inspired rgb-d saliency detection via conditional variational autoen + +coders. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8582-8591 (2020) +48. Zhang, L., Liu, X., Guan, H.: Automtl: A programming framework for automating efficient multi-task learning. Advances in Neural Information Processing Systems 35, 34216-34228 (2022) +49. Zhang, M., Zhao, X., Yao, J., Yuan, C., Huang, W.: When noisy labels meet long tail dilemmas: A representation calibration method. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 15890-15900 (2023) +50. Zhang, Z., Cui, Z., Xu, C., Yan, Y., Sebe, N., Yang, J.: Pattern-affinitive propagation across depth, surface normal and semantic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4106-4115 (2019) \ No newline at end of file diff --git a/2024/Region-aware Distribution Contrast_ A Novel Approach to Multi-Task Partially Supervised Learning/images.zip b/2024/Region-aware Distribution Contrast_ A Novel Approach to Multi-Task Partially Supervised Learning/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..e3bec6521f61fd635307d48041023cb59cbcf6a1 --- /dev/null +++ b/2024/Region-aware Distribution Contrast_ A Novel Approach to Multi-Task Partially Supervised Learning/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5281ab4ae97f78949841090db1d8198be6b5437e0baa10f3b8b36115ba22e81a +size 530479 diff --git a/2024/Region-aware Distribution Contrast_ A Novel Approach to Multi-Task Partially Supervised Learning/layout.json b/2024/Region-aware Distribution Contrast_ A Novel Approach to Multi-Task Partially Supervised Learning/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..6c5b6d373b1954211a73793972ef88f6a119d01a --- /dev/null +++ b/2024/Region-aware Distribution Contrast_ A Novel Approach to Multi-Task Partially Supervised Learning/layout.json @@ -0,0 +1,8315 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 159, + 112, + 454, + 166 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 112, + 454, + 166 + ], + "spans": [ + { + "bbox": [ + 159, + 112, + 454, + 166 + ], + "type": "text", + "content": "Region-aware Distribution Contrast: A Novel Approach to Multi-Task Partially Supervised Learning" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 217, + 188, + 397, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 217, + 188, + 397, + 213 + ], + "spans": [ + { + "bbox": [ + 217, + 188, + 397, + 213 + ], + "type": "text", + "content": "Meixuan Li, Tianyu Li, Guoqing Wang*, Peng Wang, Yang Yang, and Jie Zou" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 169, + 221, + 444, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 169, + 221, + 444, + 232 + ], + "spans": [ + { + "bbox": [ + 169, + 221, + 444, + 232 + ], + "type": "text", + "content": "University of Electronic Science and Technology of China (UESTC)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 173, + 234, + 438, + 243 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 173, + 234, + 438, + 243 + ], + "spans": [ + { + "bbox": [ + 173, + 234, + 438, + 243 + ], + "type": "text", + "content": "limeixuan0801@126.com, {cosmos.yu, p.wang6}@@hotmail.com," + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 201, + 245, + 414, + 255 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 201, + 245, + 414, + 255 + ], + "spans": [ + { + "bbox": [ + 201, + 245, + 414, + 255 + ], + "type": "text", + "content": "{gqwang0420, yang.yang, jie.zou}@uestc.edu.cn" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 160, + 300, + 452, + 595 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 300, + 452, + 595 + ], + "spans": [ + { + "bbox": [ + 160, + 300, + 452, + 595 + ], + "type": "text", + "content": "Abstract. In this study, we address the intricate challenge of multi-task dense prediction, encompassing tasks such as semantic segmentation, depth estimation, and surface normal estimation, particularly when dealing with partially annotated data (MTPSL). The complexity arises from the absence of complete task labels for each training image. Given the inter-related nature of these pixel-wise dense tasks, our focus is on mining and capturing cross-task relationships. Existing solutions typically rely on learning global image representations for global cross-task image matching, imposing constraints that, unfortunately, sacrifice the finer structures within the images. Attempting local matching as a remedy faces hurdles due to the lack of precise region supervision, making local alignment a challenging endeavor. The introduction of Segment Anything Model (SAM) sheds light on addressing local alignment challenges by providing free and high-quality solutions for region detection. Leveraging SAM-detected regions, the subsequent challenge lies in aligning the representations within these regions. Diverging from conventional methods that directly learn a monolithic image representation, our proposal involves modeling region-wise representations using Gaussian Distributions. Aligning these distributions between corresponding regions from different tasks imparts higher flexibility and capacity to capture intra-region structures, accommodating a broader range of tasks. This innovative approach significantly enhances our ability to effectively capture cross-task relationships, resulting in improved overall performance in partially supervised multi-task dense prediction scenarios. Extensive experiments conducted on two widely used benchmarks underscore the superior effectiveness of our proposed method, showcasing state-of-the-art performance even when compared to fully supervised methods." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 160, + 596, + 455, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 596, + 455, + 605 + ], + "spans": [ + { + "bbox": [ + 160, + 596, + 455, + 605 + ], + "type": "text", + "content": "https://github.com/HereNowL/Region-aware-Distribution-Contrast" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 160, + 619, + 453, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 619, + 453, + 641 + ], + "spans": [ + { + "bbox": [ + 160, + 619, + 453, + 641 + ], + "type": "text", + "content": "Keywords: Multi-task Learning " + }, + { + "bbox": [ + 160, + 619, + 453, + 641 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 619, + 453, + 641 + ], + "type": "text", + "content": " Partially Supervised Learning " + }, + { + "bbox": [ + 160, + 619, + 453, + 641 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 619, + 453, + 641 + ], + "type": "text", + "content": " Scene Understanding " + }, + { + "bbox": [ + 160, + 619, + 453, + 641 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 619, + 453, + 641 + ], + "type": "text", + "content": " Contrastive Learning" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 134, + 654, + 235, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 654, + 235, + 665 + ], + "spans": [ + { + "bbox": [ + 134, + 654, + 235, + 665 + ], + "type": "text", + "content": "* Corresponding author." + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 160, + 114, + 301, + 198 + ], + "blocks": [ + { + "bbox": [ + 160, + 114, + 301, + 198 + ], + "lines": [ + { + "bbox": [ + 160, + 114, + 301, + 198 + ], + "spans": [ + { + "bbox": [ + 160, + 114, + 301, + 198 + ], + "type": "image", + "image_path": "3e8368e9e8842c8887bc6a48ae776131aed8eca9183b0ff8e41cb6dc82697476.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 195, + 199, + 268, + 206 + ], + "lines": [ + { + "bbox": [ + 195, + 199, + 268, + 206 + ], + "spans": [ + { + "bbox": [ + 195, + 199, + 268, + 206 + ], + "type": "text", + "content": "(a) Contrastive Multiview Coding" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 307, + 114, + 455, + 196 + ], + "blocks": [ + { + "bbox": [ + 307, + 114, + 455, + 196 + ], + "lines": [ + { + "bbox": [ + 307, + 114, + 455, + 196 + ], + "spans": [ + { + "bbox": [ + 307, + 114, + 455, + 196 + ], + "type": "image", + "image_path": "e41de6a7d875d566b8ffb340ef64c57013b9eaeab9ee834916b9fe09545c8a6e.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 340, + 198, + 421, + 206 + ], + "lines": [ + { + "bbox": [ + 340, + 198, + 421, + 206 + ], + "spans": [ + { + "bbox": [ + 340, + 198, + 421, + 206 + ], + "type": "text", + "content": "(b) Region-wise Distribution Contrast" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 218, + 482, + 262 + ], + "lines": [ + { + "bbox": [ + 130, + 218, + 482, + 262 + ], + "spans": [ + { + "bbox": [ + 130, + 218, + 482, + 262 + ], + "type": "text", + "content": "Fig. 1: Multiview consistency and region consistency. (a) Illustration of contrastive multiview consistency [38]. (b) Illustration of region consistency, where " + }, + { + "bbox": [ + 130, + 218, + 482, + 262 + ], + "type": "inline_equation", + "content": "p_{i,s}" + }, + { + "bbox": [ + 130, + 218, + 482, + 262 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 218, + 482, + 262 + ], + "type": "inline_equation", + "content": "p_{j,s}" + }, + { + "bbox": [ + 130, + 218, + 482, + 262 + ], + "type": "text", + "content": " represent the region distribution of semantic segmentation, " + }, + { + "bbox": [ + 130, + 218, + 482, + 262 + ], + "type": "inline_equation", + "content": "p_{i,d}" + }, + { + "bbox": [ + 130, + 218, + 482, + 262 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 218, + 482, + 262 + ], + "type": "inline_equation", + "content": "p_{j,d}" + }, + { + "bbox": [ + 130, + 218, + 482, + 262 + ], + "type": "text", + "content": " denotes the region distribution of depth estimation." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 283, + 230, + 295 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 283, + 230, + 295 + ], + "spans": [ + { + "bbox": [ + 132, + 283, + 230, + 295 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 306, + 482, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 306, + 482, + 415 + ], + "spans": [ + { + "bbox": [ + 130, + 306, + 482, + 415 + ], + "type": "text", + "content": "With the rapid advancement of deep learning, dense prediction tasks, including semantic segmentation, depth estimation, and surface normal estimation, have witnessed remarkable progress [11, 16, 30, 34, 47]. Considering the inherent interdependence among these dense prediction tasks [45, 46], there has been a growing interest in employing unified multi-task learning networks to jointly tackle different dense prediction tasks [7, 22, 28, 29, 31, 32, 39, 43, 50]. In contrast to inefficient single-task learning networks, multi-task learning networks effectively acquire shared features, enabling them to capture more generalizable information across various tasks while avoiding redundant training [39, 43, 50]." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 415, + 482, + 558 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 415, + 482, + 558 + ], + "spans": [ + { + "bbox": [ + 130, + 415, + 482, + 558 + ], + "type": "text", + "content": "Recent advancements in multi-task dense prediction methods have primarily focused on two crucial aspects. The first aspect is dedicated to designing intricate network architectures to achieve effective multi-task learning [29, 32, 39, 43, 50], while the second aspect is devoted to devising optimal balancing strategies for loss functions in multi-task learning [7, 14, 22, 28, 29], aiming to mitigate the negative transfer across tasks. These directions mainly focus on fully supervised multi-task learning, where complete task labels are available for all training samples. However, acquiring pixel-level labels for dense prediction tasks is resource-intensive, particularly with multiple visual tasks. It is common to encounter scenarios where labels for certain tasks are missing or unreliable in some training samples [25], leading to a new research direction known as multi-task partially supervised learning (MTPSL)." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "text", + "content": "Since not all task labels are available for each training image in MTPSL, it is natural to formulate a multi-task learning framework which can leverage the implicit relationship between tasks with and without explicit labels. Dense prediction tasks, such as semantic segmentation, depth estimation, and normal prediction, exhibit a complementary co-existence relationship [37,45,46], where each task can serve as auxiliary information for others, offering valuable insights reciprocally. While existing approaches map predictions into a joint vector-wise semantic space for task alignment [25], we argue that this overlooks the underlying structural information inherent in the original image. A global vector" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 214, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 214, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 214, + 100 + ], + "type": "text", + "content": "M. Li et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 480, + 176 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 480, + 176 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 480, + 176 + ], + "type": "text", + "content": "might be insufficient to characterize the information in the entire scene, limiting its effectiveness in constraining cross-task relationships. Local matching, as shown in Fig. 1, with its intrinsic local properties, serves as a potential remedy. Nonetheless, the absence of precise region supervision presents a fundamental obstacle, thereby intensifying the challenges associated with its implementation." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 176, + 481, + 379 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 176, + 481, + 379 + ], + "spans": [ + { + "bbox": [ + 130, + 176, + 481, + 379 + ], + "type": "text", + "content": "To tackle the outlined challenges, we introduce region-based cross-task alignment by leveraging the robust region segmentation capabilities inherent in Segment Anything Model (SAM) models. Utilizing well-segmented image regions facilitates the abstraction of region-level information, enabling the addition of cross-task consistency constraints. However, a subsequent crucial challenge arises in effectively representing these regional features to ensure the consistency and accuracy of cross-task alignment. In contrast to conventional methods that directly learn a unified image representation, we employ Gaussian distributions to model region representations, enhancing flexibility and improving the ability to capture intra-region structures. By adapting Gaussian distributions between corresponding regions from different tasks, our approach can be more broadly applicable to diverse task types, such as pixel-level classification tasks (e.g., semantic segmentation) and pixel-level regression tasks (e.g., depth estimation). Subsequently, we introduce distributional contrastive learning to explicitly perform region-wise cross-task alignment. This approach serves as a robust constraint, offering a meaningful way to mine relationships between tasks and enhance dense prediction results, particularly for those tasks lacking pixel-wise labels." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 146, + 380, + 399, + 391 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 380, + 399, + 391 + ], + "spans": [ + { + "bbox": [ + 146, + 380, + 399, + 391 + ], + "type": "text", + "content": "In summary, our work masks the following contributions:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 137, + 397, + 479, + 538 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 138, + 397, + 479, + 455 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 397, + 479, + 455 + ], + "spans": [ + { + "bbox": [ + 138, + 397, + 479, + 455 + ], + "type": "text", + "content": "- We tackle the challenge of multi-task partially supervised (MTPSL) dense prediction from a fresh perspective. Our approach involves extracting cross-task local alignment through the utilization of SAM's easily obtainable local regions. This innovative strategy has demonstrated notable in alleviating label shortages." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 137, + 457, + 479, + 490 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 457, + 479, + 490 + ], + "spans": [ + { + "bbox": [ + 137, + 457, + 479, + 490 + ], + "type": "text", + "content": "- We propose a novel region distribution contrast method for local alignment, which offers increased flexibility, robustness, and wide applicability across multiple tasks." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 137, + 493, + 479, + 538 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 493, + 479, + 538 + ], + "spans": [ + { + "bbox": [ + 137, + 493, + 479, + 538 + ], + "type": "text", + "content": "- We validate the effectiveness of our proposed method through extensive experiments conducted on two widely used benchmark datasets. Notably, our approach achieves state-of-the-art performance in partially supervised learning, while also showing great potential in fully supervised learning." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 132, + 557, + 237, + 570 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 557, + 237, + 570 + ], + "spans": [ + { + "bbox": [ + 132, + 557, + 237, + 570 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 582, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 481, + 666 + ], + "type": "text", + "content": "Multi-task Supervised Learning. Multi-task learning utilizes a single model to simultaneously handle multiple visual tasks, offering advantages such as faster inference speed and more efficient utilization of input data. Some approaches [13, 32, 48] focus on designing the encoder of multi-task network, while other approaches pay more attention to designing complex decoders to generate task-specific predictions by utilizing shared features [2, 39, 43, 50]. Uncertainty [22] obtain task loss weights by considering the heteroscedastic uncertainty" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 301, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 301, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 301, + 91, + 447, + 102 + ], + "type": "text", + "content": "Region-aware Distribution Contrast" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "type": "text", + "content": "of each task, Grad-Norm [7] directly adjusts gradients to balance task losses, and DWA [14] learns time-varying average task weights by considering the loss variation rate of each task. It is important to note that these methods operate under a significant and strong assumption, namely that they are developed and studied based on the availability of labels for all tasks across all images." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 179, + 491, + 371 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 179, + 491, + 371 + ], + "spans": [ + { + "bbox": [ + 130, + 179, + 491, + 371 + ], + "type": "text", + "content": "Multi-task Semi-supervised Learning and Partially-supervised Learning. Learning multi-task models on fully annotated data requires a large-scale labeled dataset, and the cost of collecting sufficient labeled data can be high. Therefore, two common scenarios are worthy of note in multi-task learning. One is semi-supervised multi-task learning, where the dataset consists of limited annotation information for all tasks and a large amount of unlabeled data. A bunch of methods have been proposed for semi-supervised multi-task learning [8,18,19,24,27]. In recent works [8,18,19,24], regularization terms are applied to unlabeled samples from each task to encourage consistent predictions when the input is perturbed. Another scenario involves partial-supervision multi-task learning, where the dataset lacks sufficient labels for each task, and not all tasks have labels for every image. Li et al. [25] maps tasks into high-dimensional vectors for cross-task alignment. However, vector representations are insufficient and do not capture local-level alignment. In this paper, we propose leveraging distributions as a substitute for vector representations of task features and employing contrastive learning to achieve region-level alignment across tasks." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 375, + 483, + 531 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 375, + 483, + 531 + ], + "spans": [ + { + "bbox": [ + 130, + 375, + 483, + 531 + ], + "type": "text", + "content": "Contrastive Learning for Vision. Recently, contrastive learning has made significant advancements in unsupervised learning [5,6,15,33,42]. DenseCL [40] and RegionContrast [17] have been utilized for semantic segmentation tasks using contrastive learning at the pixel and region levels respectively, achieving excellent results. For depth estimation, WCL [12] transformed depth from continuous values to discrete values and constructed contrastive losses based on windows to form positive and negative sample pairs. CMC [38] was the first to propose that different viewpoints of the same image should be mapped to nearby positions in a high-dimensional space, while different images should be mapped to distant positions, thus ensuring multiview consistency. Inspired by CMC, we believe that multiple viewpoints within the same local region should also exhibit consistency, while differing from other regions. In this paper, we leverage this idea to achieve region-level contrast for MTPSL." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 533, + 483, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 483, + 668 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 483, + 668 + ], + "type": "text", + "content": "Gaussian Distribution. Recently, Gaussian distribution has been widely employed in computer vision to characterize the distribution of features [3, 4, 20, 26, 41, 49]. GMMSeg [26] utilized the Expectation-Maximization algorithm to construct Gaussian Mixture Models for each class in semantic segmentation, capturing class-conditional densities. AGMM [41] constructed an adaptive Gaussian Mixture Model for sparse annotation in semantic segmentation by incorporating labeled pixels and their similar unlabeled counterparts. DUL [3] employed mean and variance to characterize the distribution of faces for face recognition. Gaussian distributions have the capability to probabilistically describe the distribution characteristics of different visual tasks. However, to the best of our knowledge, Gaussian distributions have not been applied in MTPSL." + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 214, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 214, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 214, + 101 + ], + "type": "text", + "content": "M. Li et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 161, + 113, + 433, + 251 + ], + "blocks": [ + { + "bbox": [ + 161, + 113, + 433, + 251 + ], + "lines": [ + { + "bbox": [ + 161, + 113, + 433, + 251 + ], + "spans": [ + { + "bbox": [ + 161, + 113, + 433, + 251 + ], + "type": "image", + "image_path": "e0634edcee03b66869a58f843e1e3be0b743524f41a4e2cfab06a731939c7ceb.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 259, + 482, + 338 + ], + "lines": [ + { + "bbox": [ + 130, + 259, + 482, + 338 + ], + "spans": [ + { + "bbox": [ + 130, + 259, + 482, + 338 + ], + "type": "text", + "content": "Fig. 2: Illustration of region-aware distribution contrast learning method for MTPSL. During training, supervised constraints " + }, + { + "bbox": [ + 130, + 259, + 482, + 338 + ], + "type": "inline_equation", + "content": "L_{Sup}" + }, + { + "bbox": [ + 130, + 259, + 482, + 338 + ], + "type": "text", + "content": " are applied to the annotated task " + }, + { + "bbox": [ + 130, + 259, + 482, + 338 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 130, + 259, + 482, + 338 + ], + "type": "text", + "content": ". For task " + }, + { + "bbox": [ + 130, + 259, + 482, + 338 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 130, + 259, + 482, + 338 + ], + "type": "text", + "content": " and unlabelled task " + }, + { + "bbox": [ + 130, + 259, + 482, + 338 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 259, + 482, + 338 + ], + "type": "text", + "content": ", the " + }, + { + "bbox": [ + 130, + 259, + 482, + 338 + ], + "type": "inline_equation", + "content": "a_{\\theta}" + }, + { + "bbox": [ + 130, + 259, + 482, + 338 + ], + "type": "text", + "content": " map the true label " + }, + { + "bbox": [ + 130, + 259, + 482, + 338 + ], + "type": "inline_equation", + "content": "y_{s}" + }, + { + "bbox": [ + 130, + 259, + 482, + 338 + ], + "type": "text", + "content": " and the prediction " + }, + { + "bbox": [ + 130, + 259, + 482, + 338 + ], + "type": "inline_equation", + "content": "\\hat{y}_{t}" + }, + { + "bbox": [ + 130, + 259, + 482, + 338 + ], + "type": "text", + "content": " in the high-dimensional space respectively and then model the region of task-specific features extracted using SAM as a Gaussian distribution. Contrastive learning is then employed to minimize the distance between distributions of the same region across different tasks while maximizing the distance to distributions of other regions." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 358, + 202, + 370 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 358, + 202, + 370 + ], + "spans": [ + { + "bbox": [ + 132, + 358, + 202, + 370 + ], + "type": "text", + "content": "3 Method" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 382, + 228, + 393 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 382, + 228, + 393 + ], + "spans": [ + { + "bbox": [ + 132, + 382, + 228, + 393 + ], + "type": "text", + "content": "3.1 Preliminaries" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 402, + 482, + 522 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 402, + 482, + 522 + ], + "spans": [ + { + "bbox": [ + 130, + 402, + 482, + 522 + ], + "type": "text", + "content": "Problem Setup. Considering the problem of multi-task learning involving " + }, + { + "bbox": [ + 130, + 402, + 482, + 522 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 402, + 482, + 522 + ], + "type": "text", + "content": " tasks, where " + }, + { + "bbox": [ + 130, + 402, + 482, + 522 + ], + "type": "inline_equation", + "content": "K \\geq 2" + }, + { + "bbox": [ + 130, + 402, + 482, + 522 + ], + "type": "text", + "content": ". A training dataset " + }, + { + "bbox": [ + 130, + 402, + 482, + 522 + ], + "type": "inline_equation", + "content": "S = \\{x_{i}\\}_{i=1}^{N}" + }, + { + "bbox": [ + 130, + 402, + 482, + 522 + ], + "type": "text", + "content": " is given with " + }, + { + "bbox": [ + 130, + 402, + 482, + 522 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 402, + 482, + 522 + ], + "type": "text", + "content": " partially labelled samples, indicating that for each training sample " + }, + { + "bbox": [ + 130, + 402, + 482, + 522 + ], + "type": "inline_equation", + "content": "x_{i}" + }, + { + "bbox": [ + 130, + 402, + 482, + 522 + ], + "type": "text", + "content": ", only a subset of the " + }, + { + "bbox": [ + 130, + 402, + 482, + 522 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 402, + 482, + 522 + ], + "type": "text", + "content": " tasks are provided true labels. Let " + }, + { + "bbox": [ + 130, + 402, + 482, + 522 + ], + "type": "inline_equation", + "content": "P_{i}" + }, + { + "bbox": [ + 130, + 402, + 482, + 522 + ], + "type": "text", + "content": " represents the number of labeled tasks for " + }, + { + "bbox": [ + 130, + 402, + 482, + 522 + ], + "type": "inline_equation", + "content": "x_{i}" + }, + { + "bbox": [ + 130, + 402, + 482, + 522 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 402, + 482, + 522 + ], + "type": "inline_equation", + "content": "Q_{i}" + }, + { + "bbox": [ + 130, + 402, + 482, + 522 + ], + "type": "text", + "content": " represents the number of unlabeled tasks, such that " + }, + { + "bbox": [ + 130, + 402, + 482, + 522 + ], + "type": "inline_equation", + "content": "P_{i} + Q_{i} = K" + }, + { + "bbox": [ + 130, + 402, + 482, + 522 + ], + "type": "text", + "content": ". When " + }, + { + "bbox": [ + 130, + 402, + 482, + 522 + ], + "type": "inline_equation", + "content": "P_{i}" + }, + { + "bbox": [ + 130, + 402, + 482, + 522 + ], + "type": "text", + "content": " equals " + }, + { + "bbox": [ + 130, + 402, + 482, + 522 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 402, + 482, + 522 + ], + "type": "text", + "content": " for all " + }, + { + "bbox": [ + 130, + 402, + 482, + 522 + ], + "type": "inline_equation", + "content": "x_{i} \\in S" + }, + { + "bbox": [ + 130, + 402, + 482, + 522 + ], + "type": "text", + "content": ", it indicates fully supervised multi-task learning. Conversely, when " + }, + { + "bbox": [ + 130, + 402, + 482, + 522 + ], + "type": "inline_equation", + "content": "Q_{i}" + }, + { + "bbox": [ + 130, + 402, + 482, + 522 + ], + "type": "text", + "content": " equals " + }, + { + "bbox": [ + 130, + 402, + 482, + 522 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 402, + 482, + 522 + ], + "type": "text", + "content": " for all " + }, + { + "bbox": [ + 130, + 402, + 482, + 522 + ], + "type": "inline_equation", + "content": "x_{i} \\in S" + }, + { + "bbox": [ + 130, + 402, + 482, + 522 + ], + "type": "text", + "content": ", it implies that no task labels are available, leading to unsupervised multi-task learning. In this paper, we focus on addressing partially supervised multi-task learning, where each training image " + }, + { + "bbox": [ + 130, + 402, + 482, + 522 + ], + "type": "inline_equation", + "content": "x_{i}" + }, + { + "bbox": [ + 130, + 402, + 482, + 522 + ], + "type": "text", + "content": " can obtain labels for at least one task (" + }, + { + "bbox": [ + 130, + 402, + 482, + 522 + ], + "type": "inline_equation", + "content": "P_{i} \\geq 1" + }, + { + "bbox": [ + 130, + 402, + 482, + 522 + ], + "type": "text", + "content": ")." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 539, + 288, + 550 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 539, + 288, + 550 + ], + "spans": [ + { + "bbox": [ + 132, + 539, + 288, + 550 + ], + "type": "text", + "content": "3.2 Overview of Our Method" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 557, + 482, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 557, + 482, + 629 + ], + "spans": [ + { + "bbox": [ + 130, + 557, + 482, + 629 + ], + "type": "text", + "content": "As shown in Fig. 2, our method focuses on utilizing Gaussian distributions to represent features in local regions for contrastive learning, achieving better distribution consistency across tasks at the region-level for MTPSL. Assuming that task " + }, + { + "bbox": [ + 130, + 557, + 482, + 629 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 130, + 557, + 482, + 629 + ], + "type": "text", + "content": " has label while task " + }, + { + "bbox": [ + 130, + 557, + 482, + 629 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 557, + 482, + 629 + ], + "type": "text", + "content": " does not, our objective is to achieve alignment between the true label " + }, + { + "bbox": [ + 130, + 557, + 482, + 629 + ], + "type": "inline_equation", + "content": "y_{s}" + }, + { + "bbox": [ + 130, + 557, + 482, + 629 + ], + "type": "text", + "content": " of task " + }, + { + "bbox": [ + 130, + 557, + 482, + 629 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 130, + 557, + 482, + 629 + ], + "type": "text", + "content": " and the prediction " + }, + { + "bbox": [ + 130, + 557, + 482, + 629 + ], + "type": "inline_equation", + "content": "\\hat{y}_t" + }, + { + "bbox": [ + 130, + 557, + 482, + 629 + ], + "type": "text", + "content": " of task " + }, + { + "bbox": [ + 130, + 557, + 482, + 629 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 557, + 482, + 629 + ], + "type": "text", + "content": " at the region level using contrastive learning." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 630, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 630, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 630, + 482, + 665 + ], + "type": "text", + "content": "In the training stage, to accurately and efficiently obtain region-level features for prediction, we simultaneously feed the input image to both the multi-task learning (MTL) network and the pre-trained SAM. Predictions are obtained" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 301, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 301, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 301, + 91, + 447, + 102 + ], + "type": "text", + "content": "Region-aware Distribution Contrast" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 170, + 114, + 446, + 238 + ], + "blocks": [ + { + "bbox": [ + 170, + 114, + 446, + 238 + ], + "lines": [ + { + "bbox": [ + 170, + 114, + 446, + 238 + ], + "spans": [ + { + "bbox": [ + 170, + 114, + 446, + 238 + ], + "type": "image", + "image_path": "fbb93ca70c827e550da39fb85b3f29a8a2c26074e15c8d06cddb311332949774.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 248, + 482, + 304 + ], + "lines": [ + { + "bbox": [ + 130, + 248, + 482, + 304 + ], + "spans": [ + { + "bbox": [ + 130, + 248, + 482, + 304 + ], + "type": "text", + "content": "Fig. 3: Illustration of region-level cross-task consistency. Initially, SAM predictions are employed to extract regions from the features of the particular task. Following that, these regions are modeled as Gaussian distributions. Finally, contrastive learning is utilized to minimize the distance between regions of the same region across different tasks and maximize the distance to regions of other regions." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 328, + 482, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 328, + 482, + 471 + ], + "spans": [ + { + "bbox": [ + 130, + 328, + 482, + 471 + ], + "type": "text", + "content": "from both MTL network and SAM. For tasks with labeled data, we apply fully-supervised loss " + }, + { + "bbox": [ + 130, + 328, + 482, + 471 + ], + "type": "inline_equation", + "content": "L_{Sup}" + }, + { + "bbox": [ + 130, + 328, + 482, + 471 + ], + "type": "text", + "content": " as explicit training supervision. For tasks without label, an auxiliary network " + }, + { + "bbox": [ + 130, + 328, + 482, + 471 + ], + "type": "inline_equation", + "content": "a_{\\theta}" + }, + { + "bbox": [ + 130, + 328, + 482, + 471 + ], + "type": "text", + "content": " maps the unlabelled task prediction " + }, + { + "bbox": [ + 130, + 328, + 482, + 471 + ], + "type": "inline_equation", + "content": "\\hat{y}_t" + }, + { + "bbox": [ + 130, + 328, + 482, + 471 + ], + "type": "text", + "content": " and the annotated task label " + }, + { + "bbox": [ + 130, + 328, + 482, + 471 + ], + "type": "inline_equation", + "content": "y_s" + }, + { + "bbox": [ + 130, + 328, + 482, + 471 + ], + "type": "text", + "content": " to the same high-dimensional feature map space, and use " + }, + { + "bbox": [ + 130, + 328, + 482, + 471 + ], + "type": "inline_equation", + "content": "L_{RC}" + }, + { + "bbox": [ + 130, + 328, + 482, + 471 + ], + "type": "text", + "content": " to achieve cross-task consistency at the region level. Specifically, We first extract region-wise features in " + }, + { + "bbox": [ + 130, + 328, + 482, + 471 + ], + "type": "inline_equation", + "content": "\\hat{y}_t" + }, + { + "bbox": [ + 130, + 328, + 482, + 471 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 328, + 482, + 471 + ], + "type": "inline_equation", + "content": "y_s" + }, + { + "bbox": [ + 130, + 328, + 482, + 471 + ], + "type": "text", + "content": " based on the segmented regions identified by SAM, then model them as Gaussian distributions. After acquiring the Gaussian distributions for features in each region of each task, we utilize contrastive learning to minimize the distance between the Gaussian distributions of features in the same region across different tasks and maximize the distance from the Gaussian distributions of features from different regions. Finally, the optimization of our method can be defined as:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 260, + 483, + 481, + 496 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 260, + 483, + 481, + 496 + ], + "spans": [ + { + "bbox": [ + 260, + 483, + 481, + 496 + ], + "type": "interline_equation", + "content": "L o s s = L _ {S u p} + L _ {R C}. \\tag {1}", + "image_path": "1b51f1dbbf4a9fbd9c25c12ae88c1688bdeaa73b1203fccca4b44bcc07c87f99.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 505, + 482, + 541 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 505, + 482, + 541 + ], + "spans": [ + { + "bbox": [ + 130, + 505, + 482, + 541 + ], + "type": "text", + "content": "In the inference stage, as illustrated in Fig. 2, the input image is fed only into the MTL network to generate predictions for multiple tasks, notably, with SAM no longer being utilized." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 560, + 339, + 572 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 560, + 339, + 572 + ], + "spans": [ + { + "bbox": [ + 131, + 560, + 339, + 572 + ], + "type": "text", + "content": "3.3 Region-aware Contrastive Learning" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 581, + 481, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 581, + 481, + 604 + ], + "spans": [ + { + "bbox": [ + 130, + 581, + 481, + 604 + ], + "type": "text", + "content": "In this section, we explain the implementation details to achieve region-level cross-task consistency." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "type": "text", + "content": "Region-level Features Extraction. To realize the region-level cross-task consistency, it is essential to informatively choose regions for feature extraction. Considering the robust capabilities of SAM for image segmentation [23], we employ it to generate the regions for each training sample, leveraging its ability to provide valuable prior knowledge. As shown in Fig. 2, to obtain fine-grained" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 214, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 214, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 214, + 100 + ], + "type": "text", + "content": "M. Li et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 223 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 223 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 223 + ], + "type": "text", + "content": "semantics and fine-grained edges of regions, we fed the image into MTL network and pre-trained SAM simultaneously. SAM performs fine-grained segmentation of each object in the image, and its output masks contain the ID of each area and its corresponding position in the image. Inspired by [21], for convenient region-wise feature extraction in task-specific feature maps, we compute the predicted masks of SAM as a grayscale image, where each pixel value corresponds to the area ID returned by SAM, indicating its category. Leveraging the grayscale masks of SAM, we split the image into different regions and extract region-wise features within feature maps of different tasks." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 224, + 480, + 308 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 224, + 480, + 308 + ], + "spans": [ + { + "bbox": [ + 130, + 224, + 480, + 308 + ], + "type": "text", + "content": "Gaussian Distribution Modeling. After obtaining the region-wise features for each task, we consider modeling the region-wise features as Gaussian distributions. The utilization of Gaussian distributions represents the features of a region in a probabilistic manner, which provides a more comprehensive depiction of the region's variations and compensates for the shortcomings of pixel-level hard alignment. Specifically, we model each region " + }, + { + "bbox": [ + 130, + 224, + 480, + 308 + ], + "type": "inline_equation", + "content": "r_i" + }, + { + "bbox": [ + 130, + 224, + 480, + 308 + ], + "type": "text", + "content": " from the task-specific feature map " + }, + { + "bbox": [ + 130, + 224, + 480, + 308 + ], + "type": "inline_equation", + "content": "f_t" + }, + { + "bbox": [ + 130, + 224, + 480, + 308 + ], + "type": "text", + "content": " as a Gaussian distribution:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 233, + 316, + 480, + 329 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 233, + 316, + 480, + 329 + ], + "spans": [ + { + "bbox": [ + 233, + 316, + 480, + 329 + ], + "type": "interline_equation", + "content": "p _ {i, t} = p \\left(r _ {i} \\mid f _ {t}\\right) = N \\left(r _ {i}; \\boldsymbol {\\mu} _ {i, t}, \\boldsymbol {\\Sigma} _ {i, t}\\right), \\tag {2}", + "image_path": "2afb8eb6b0c74ce14ad8ebc63e7732ff93623f43e3b05f6403b655404ad15988.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 336, + 479, + 360 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 336, + 479, + 360 + ], + "spans": [ + { + "bbox": [ + 130, + 336, + 479, + 360 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 336, + 479, + 360 + ], + "type": "inline_equation", + "content": "\\pmb{\\mu}_{i,t}" + }, + { + "bbox": [ + 130, + 336, + 479, + 360 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 336, + 479, + 360 + ], + "type": "inline_equation", + "content": "\\pmb{\\Sigma}_{i,t}" + }, + { + "bbox": [ + 130, + 336, + 479, + 360 + ], + "type": "text", + "content": " represent mean and covariance matrix respectively, calculated from the region " + }, + { + "bbox": [ + 130, + 336, + 479, + 360 + ], + "type": "inline_equation", + "content": "r_i" + }, + { + "bbox": [ + 130, + 336, + 479, + 360 + ], + "type": "text", + "content": " in the feature map " + }, + { + "bbox": [ + 130, + 336, + 479, + 360 + ], + "type": "inline_equation", + "content": "f_{t}" + }, + { + "bbox": [ + 130, + 336, + 479, + 360 + ], + "type": "text", + "content": ", as shown in Fig. 3." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 361, + 480, + 431 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 361, + 480, + 431 + ], + "spans": [ + { + "bbox": [ + 130, + 361, + 480, + 431 + ], + "type": "text", + "content": "Region Distribution Contrast. Once we have obtained the Gaussian distributions for each region, our objective is to utilize contrastive learning to minimize the distance between Gaussian distributions of the same region across different tasks and maximize the distance between Gaussian distributions of different regions. Hence, the region-wise Gaussian distribution contrastive loss for the region distribution " + }, + { + "bbox": [ + 130, + 361, + 480, + 431 + ], + "type": "inline_equation", + "content": "p_{i,t}" + }, + { + "bbox": [ + 130, + 361, + 480, + 431 + ], + "type": "text", + "content": " can be defined as follows:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 188, + 441, + 480, + 467 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 188, + 441, + 480, + 467 + ], + "spans": [ + { + "bbox": [ + 188, + 441, + 480, + 467 + ], + "type": "interline_equation", + "content": "L _ {i, t} ^ {N C E} = - \\sum_ {k _ {+}} \\log \\frac {\\operatorname {s i m} \\left(p _ {i , t} , k _ {+}\\right)}{\\operatorname {s i m} \\left(p _ {i , t} , k _ {+}\\right) + \\sum_ {k _ {-}} \\operatorname {s i m} \\left(p _ {i , t} , k _ {-}\\right)}, \\tag {3}", + "image_path": "0b8417e9de4121fc380f800601f144c09b1584d311db9cc9b6a47bf0224c518a.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 219, + 480, + 480, + 491 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 219, + 480, + 480, + 491 + ], + "spans": [ + { + "bbox": [ + 219, + 480, + 480, + 491 + ], + "type": "interline_equation", + "content": "\\operatorname {s i m} \\left(p _ {i, t}, k\\right) = \\exp \\left(- W _ {\\text {d i s t a n c e}} \\left(p _ {i, t}, k\\right) / \\tau\\right), \\tag {4}", + "image_path": "f036070bef41faffac0c1a7a7b387674fce6825646c85b44164333b7ae241bad.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 495, + 479, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 495, + 479, + 590 + ], + "spans": [ + { + "bbox": [ + 130, + 495, + 479, + 590 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 495, + 479, + 590 + ], + "type": "inline_equation", + "content": "k_{+} \\in \\{p_{i,s}\\}_{s \\neq t}" + }, + { + "bbox": [ + 130, + 495, + 479, + 590 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 495, + 479, + 590 + ], + "type": "inline_equation", + "content": "k_{-} \\in \\{p_{j,*}\\}_{j \\neq i}" + }, + { + "bbox": [ + 130, + 495, + 479, + 590 + ], + "type": "text", + "content": ", * denotes any task, sim denotes the exponential equation of the Wasserstein distance, and " + }, + { + "bbox": [ + 130, + 495, + 479, + 590 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 130, + 495, + 479, + 590 + ], + "type": "text", + "content": " is the temperature parameter. Wasserstein distance [35] is a metric used to quantify the dissimilarity between two probability distributions and well-suited for accurately measuring the distance between discrete and continuous distributions, as it accounts for the underlying structure and geometry of the data. It measures the minimum amount of work required to transform one distribution into another and can be calculated as follows:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 199, + 597, + 480, + 635 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 199, + 597, + 480, + 635 + ], + "spans": [ + { + "bbox": [ + 199, + 597, + 480, + 635 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} W _ {\\text {d i s t a n c e}} \\left(p _ {i, t}, k\\right) = \\left\\| \\boldsymbol {\\mu} _ {i, t} - \\boldsymbol {\\mu} _ {k} \\right\\| ^ {2} + \\operatorname {T r} \\left(\\boldsymbol {\\Sigma} _ {i, t}\\right) \\\\ + \\operatorname {T r} \\left(\\boldsymbol {\\Sigma} _ {k}\\right) - 2 \\operatorname {T r} \\left(\\left(\\boldsymbol {\\Sigma} _ {i, t} \\boldsymbol {\\Sigma} _ {k}\\right) ^ {1 / 2}\\right), \\tag {5} \\\\ \\end{array}", + "image_path": "fdfb4b51524da86e45ed6b02b63754f4942ebe8615824ab468477fc72ff35db6.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 641, + 479, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 479, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 479, + 665 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 641, + 479, + 665 + ], + "type": "inline_equation", + "content": "\\mathrm{Tr}" + }, + { + "bbox": [ + 130, + 641, + 479, + 665 + ], + "type": "text", + "content": " represents the trace of the matrix, " + }, + { + "bbox": [ + 130, + 641, + 479, + 665 + ], + "type": "inline_equation", + "content": "\\pmb{\\mu}_k" + }, + { + "bbox": [ + 130, + 641, + 479, + 665 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 641, + 479, + 665 + ], + "type": "inline_equation", + "content": "\\pmb{\\Sigma}_k" + }, + { + "bbox": [ + 130, + 641, + 479, + 665 + ], + "type": "text", + "content": " represent the mean and covariance matrix of sample " + }, + { + "bbox": [ + 130, + 641, + 479, + 665 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 641, + 479, + 665 + ], + "type": "text", + "content": ". Based on " + }, + { + "bbox": [ + 130, + 641, + 479, + 665 + ], + "type": "inline_equation", + "content": "L^{NCE}" + }, + { + "bbox": [ + 130, + 641, + 479, + 665 + ], + "type": "text", + "content": ", the region-wise contrastive loss" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 301, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 301, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 301, + 91, + 447, + 102 + ], + "type": "text", + "content": "Region-aware Distribution Contrast" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 269, + 128 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 269, + 128 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 269, + 128 + ], + "type": "inline_equation", + "content": "L_{RC}" + }, + { + "bbox": [ + 132, + 116, + 269, + 128 + ], + "type": "text", + "content": " can be defined as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 247, + 137, + 481, + 169 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 247, + 137, + 481, + 169 + ], + "spans": [ + { + "bbox": [ + 247, + 137, + 481, + 169 + ], + "type": "interline_equation", + "content": "L _ {R C} = \\frac {1}{K M} \\sum_ {t} ^ {K} \\sum_ {i} ^ {M} L _ {i, t} ^ {N C E}, \\tag {6}", + "image_path": "4ef1bc01df756728375a257b472c76928c7f59d815a4de94f1e857279b509804.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 178, + 377, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 178, + 377, + 190 + ], + "spans": [ + { + "bbox": [ + 132, + 178, + 377, + 190 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 132, + 178, + 377, + 190 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 132, + 178, + 377, + 190 + ], + "type": "text", + "content": " is the number of regions in the feature map " + }, + { + "bbox": [ + 132, + 178, + 377, + 190 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 132, + 178, + 377, + 190 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 190, + 480, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 190, + 480, + 224 + ], + "spans": [ + { + "bbox": [ + 130, + 190, + 480, + 224 + ], + "type": "text", + "content": "By minimizing " + }, + { + "bbox": [ + 130, + 190, + 480, + 224 + ], + "type": "inline_equation", + "content": "L_{RC}" + }, + { + "bbox": [ + 130, + 190, + 480, + 224 + ], + "type": "text", + "content": ", the network learns to contrast the distributions from the same region with those from different regions, thereby enabling better cross-task consistency at the region level." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 226, + 480, + 320 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 226, + 480, + 320 + ], + "spans": [ + { + "bbox": [ + 130, + 226, + 480, + 320 + ], + "type": "text", + "content": "Alternative Local Extraction Strategy. Here we explore an alternative local extraction strategy to investigate local cross-task consistency. To extract local information from the feature map, the simplest approach is to divide the map into patches, where patches at the same position serve as positive samples, and their distances are minimized. Conversely, patches from other positions act as negative samples, and their distances are maximized. We refer to this local extraction strategy as patch-aware contrast, which will be compared to the region-aware contrast method in Sec. 4." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 322, + 481, + 440 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 322, + 481, + 440 + ], + "spans": [ + { + "bbox": [ + 130, + 322, + 481, + 440 + ], + "type": "text", + "content": "Alternative Local Contrast Strategy. Alternatively, the method of modeling regions as Gaussian distributions and performing contrastive learning can be replaced with other contrastive strategies. One strategy is pixel-to-pixel level region contrast, where each pixel within the region is pulled closer to the corresponding pixel in the positive sample region and pushed far apart from each pixel in the negative sample region. We refer to this as pixel contrast. Another strategy is to map the region to a vector and bring it closer to the vector of the positive sample region while pushing it away from the vector of the negative sample region. This strategy is referred to as vector contrast. We compare these alternative methods in Sec. 4." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 460, + 230, + 474 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 460, + 230, + 474 + ], + "spans": [ + { + "bbox": [ + 132, + 460, + 230, + 474 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 485, + 481, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 485, + 481, + 617 + ], + "spans": [ + { + "bbox": [ + 130, + 485, + 481, + 617 + ], + "type": "text", + "content": "Datasets. We evaluated our method on two standard dense prediction multi-task datasets: NYU-V2 [36] and Cityscapes [9]. NYU-V2 is an indoor dataset of natural scenes, including three dense prediction tasks: semantic segmentation, depth estimation, and surface normal estimation. The semantic segmentation task in NYU-V2 contains 13 categories, depth estimation is provided by Microsoft Kinect, and surface normal estimation is provided by [10]. As [25, 29], all images were resized to " + }, + { + "bbox": [ + 130, + 485, + 481, + 617 + ], + "type": "inline_equation", + "content": "288 \\times 384" + }, + { + "bbox": [ + 130, + 485, + 481, + 617 + ], + "type": "text", + "content": " for training and evaluation purposes. Cityscapes is a well-known dataset for autonomous driving, consisting of outdoor street scenes with two tasks: semantic segmentation and depth estimation. Following [25, 29], we used 7-class semantic segmentation annotations for evaluation and resized all images to " + }, + { + "bbox": [ + 130, + 485, + 481, + 617 + ], + "type": "inline_equation", + "content": "128 \\times 256" + }, + { + "bbox": [ + 130, + 485, + 481, + 617 + ], + "type": "text", + "content": " to improve training speed." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 617, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 481, + 665 + ], + "type": "text", + "content": "Experimental Setting. Following [25], we evaluated our method in two partially supervised multi-task settings: the one label setting and the random label setting. In the one label setting, each image is assigned a random task with a label. In the random label setting, each image is assigned random labels for" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 214, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 214, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 214, + 101 + ], + "type": "text", + "content": "M. Li et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 157, + 179, + 457, + 335 + ], + "blocks": [ + { + "bbox": [ + 130, + 114, + 482, + 169 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 482, + 169 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 482, + 169 + ], + "type": "text", + "content": "Table 1: Quantitative comparison on the NYU-V2 dataset. " + }, + { + "bbox": [ + 130, + 114, + 482, + 169 + ], + "type": "inline_equation", + "content": "\\uparrow (\\downarrow)" + }, + { + "bbox": [ + 130, + 114, + 482, + 169 + ], + "type": "text", + "content": " denotes that, larger (smaller) values lead to better quality. * represents the available partially supervised SOTA. * represents the available fully supervised SOTA. + denotes the percentage improvement over the SOTA performance, corresponding to " + }, + { + "bbox": [ + 130, + 114, + 482, + 169 + ], + "type": "inline_equation", + "content": "\\uparrow (\\downarrow)" + }, + { + "bbox": [ + 130, + 114, + 482, + 169 + ], + "type": "text", + "content": ". The bold denotes the best." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 157, + 179, + 457, + 335 + ], + "lines": [ + { + "bbox": [ + 157, + 179, + 457, + 335 + ], + "spans": [ + { + "bbox": [ + 157, + 179, + 457, + 335 + ], + "type": "table", + "html": "
Label settingMethodBackboneExtraction strategyTasks
PatchRegionSemantic(mIoU)↑Depth(aErr)↓Normals(mErr)↓
fullTaskExpert* [44]ViT-Large57.580.37305.90
onelabelLi et al. [25]SegNet30.360.608832.08
onelabelDejaVu* [1]SegNet31.020.595932.15
onelabelGaussian Contrast(Ours)SegNet✓31.79 (+2.48)0.5835 (+2.25)30.38 (+5.51)
onelabelLi et al. * [25]HRNet1839.420.507114.79
onelabelDejaVu [1]HRNet18———
onelabelGaussian ContrastHRNet18✓41.56 (+5.43)0.4884 (+3.69)7.71 (+47.87)
onelabelGaussian Contrast(Ours)HRNet18✓42.28 (+7.26)0.4641 (+8.48)4.86 (+67.14)
onelabelTaskExpert* [44]ViT-Large49.530.430511.12
onelabelTaskExpert [44]+OursViT-Large✓55.81 (+12.68)0.4092 (+4.95)8.46 (+23.92)
random labelLi et al. [25]SegNet34.260.578731.06
random labelDejaVu* [1]SegNet35.720.566529.82
random labelGaussian Contrast(Ours)SegNet✓37.40 (+4.70)0.5428 (+4.18)28.97 (+2.85)
random labelLi et al. * [25]HRNet1841.350.484514.34
random labelDejaVu [1]HRNet18———
random labelGaussian ContrastHRNet18✓45.79 (+10.74)0.4619 (+4.66)7.37 (+48.61)
random labelGaussian Contrast(Ours)HRNet18✓46.21 (+11.75)0.4482 (+7.49)4.49 (+68.69)
random labelTaskExpert* [44]ViT-Large56.320.42826.47
random labelTaskExpert [44]+OursViT-Large✓57.14 (+1.46)0.3977 (+7.12)4.78 (+26.12)
", + "image_path": "f707a989f859cc1faab9be31016cf997a70f8382b698e7b52640f15ee490b08f.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 355, + 480, + 390 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 355, + 480, + 390 + ], + "spans": [ + { + "bbox": [ + 130, + 355, + 480, + 390 + ], + "type": "text", + "content": "multiple tasks, with at least one task having a label and at most " + }, + { + "bbox": [ + 130, + 355, + 480, + 390 + ], + "type": "inline_equation", + "content": "K - 1" + }, + { + "bbox": [ + 130, + 355, + 480, + 390 + ], + "type": "text", + "content": " tasks having labels, where " + }, + { + "bbox": [ + 130, + 355, + 480, + 390 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 355, + 480, + 390 + ], + "type": "text", + "content": " represents the total number of tasks in the multi-task setup." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 391, + 482, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 391, + 482, + 510 + ], + "spans": [ + { + "bbox": [ + 130, + 391, + 482, + 510 + ], + "type": "text", + "content": "Implementation Details and Evaluation Metrics. All our experiments were conducted on the NVIDIA A100 GPU. For NYU-V2, we trained the models for 20 epochs, while for Cityscapes, we trained them for 100 epochs. Following [25, 29], we employed cross-entropy loss for semantic segmentation, L1 loss for depth estimation, and cosine similarity loss for surface normal estimation. We utilized the exact same evaluation metrics as mentioned in [25, 29]. For the semantic segmentation task, we employed the mean intersection over union (mIoU) metric. The absolute error (aErr) was used to evaluate the depth estimation task, and the mean error (mErr) was employed for surface normal estimation." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 527, + 282, + 539 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 527, + 282, + 539 + ], + "spans": [ + { + "bbox": [ + 131, + 527, + 282, + 539 + ], + "type": "text", + "content": "4.1 Quantitative Evaluation" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": "Quantitative Results on NYU-V2 Dataset. The evaluation results are shown in Tab. 1. Due to the absence of pre-trained parameters in Li et al. [25], the training process exhibits slower convergence. To expedite the convergence speed, we replaced the backbone of Li et al. [25] with HRNet18, which possesses a smaller training parameter size but utilizes pre-trained weights. The comparisons between Li et al. [25] with SegNet as the backbone and Li et al. [25] with HRNet18 as the backbone are presented in the first row and the fourth row, respectively, for each label setting. Since the source code of DejaVu [1] is not publicly available, we are unable to replace its backbone with HRNet18 for a fair comparison. To resolve this issue, we provide the results of our method based" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 301, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 301, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 301, + 91, + 447, + 102 + ], + "type": "text", + "content": "Region-aware Distribution Contrast" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 167, + 168, + 447, + 225 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 482, + 158 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 482, + 158 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 482, + 158 + ], + "type": "text", + "content": "Table 2: Quantitative comparison on the Cityscapes dataset. " + }, + { + "bbox": [ + 132, + 114, + 482, + 158 + ], + "type": "inline_equation", + "content": "\\uparrow (\\downarrow)" + }, + { + "bbox": [ + 132, + 114, + 482, + 158 + ], + "type": "text", + "content": " denotes that, larger (smaller) values lead to better quality. * represents the available SOTA. + denotes the percentage improvement over the SOTA performance, corresponding to " + }, + { + "bbox": [ + 132, + 114, + 482, + 158 + ], + "type": "inline_equation", + "content": "\\uparrow" + }, + { + "bbox": [ + 132, + 114, + 482, + 158 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 132, + 114, + 482, + 158 + ], + "type": "inline_equation", + "content": "\\downarrow" + }, + { + "bbox": [ + 132, + 114, + 482, + 158 + ], + "type": "text", + "content": "). The bold denotes the best." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 167, + 168, + 447, + 225 + ], + "lines": [ + { + "bbox": [ + 167, + 168, + 447, + 225 + ], + "spans": [ + { + "bbox": [ + 167, + 168, + 447, + 225 + ], + "type": "table", + "html": "
Label settingMethodBackboneContrast strategyTasks
Vector Pixel GaussianSemantic(mIoU)↑Depth(aErr) ↓
onelabelLi et al. [25]SegNet74.900.0161
onelabelLi et al. * [25]HRNet1881.730.0157
onelabelRegion-aware ContrastHRNet18✓82.76 (+1.26)0.0145 (+7.64)
onelabelRegion-aware ContrastHRNet1883.20 (+1.80)0.0141 (+10.19)
onelabelRegion-aware Contrast (Ours)HRNet1883.92 (+2.68)0.0121 (+22.73)
", + "image_path": "d86848cd02852be31722a1afe7ed7afacb932e99eb137d6b2202a1f3ef454631.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 167, + 270, + 447, + 376 + ], + "blocks": [ + { + "bbox": [ + 132, + 238, + 480, + 261 + ], + "lines": [ + { + "bbox": [ + 132, + 238, + 480, + 261 + ], + "spans": [ + { + "bbox": [ + 132, + 238, + 480, + 261 + ], + "type": "text", + "content": "Table 3: Ablation study on the NYU-V2 dataset to explore the performance of different local extraction strategies and different local contrast strategies." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 167, + 270, + 447, + 376 + ], + "lines": [ + { + "bbox": [ + 167, + 270, + 447, + 376 + ], + "spans": [ + { + "bbox": [ + 167, + 270, + 447, + 376 + ], + "type": "table", + "html": "
Label settingExtraction strategyContrast strategyTasks
PatchRegionVector Pixel GaussianSemantic(mIoU)↑Depth(aErr)↓Normals(mErr)↓
onelabel✓✓40.240.4998
onelabel✓✓41.500.4934
onelabel✓✓41.560.4884
onelabel✓✓41.840.4727
onelabel✓✓42.050.4663
onelabel✓✓42.280.4641
random label✓✓43.980.4795
random label✓✓44.620.4777
random label✓✓45.790.4619
random label✓✓45.830.4630
random label✓✓46.410.4518
random label✓✓46.210.4482
", + "image_path": "2e0cdd819264641e445b4509671add86b934f965596b54c63de9e792d51c5c8f.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 160, + 390, + 455, + 548 + ], + "blocks": [ + { + "bbox": [ + 160, + 390, + 455, + 548 + ], + "lines": [ + { + "bbox": [ + 160, + 390, + 455, + 548 + ], + "spans": [ + { + "bbox": [ + 160, + 390, + 455, + 548 + ], + "type": "image", + "image_path": "a439ac06aa26d330254df83d7af8e02e21e825d90577a20bf0d36605342e9a44.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 558, + 482, + 604 + ], + "lines": [ + { + "bbox": [ + 132, + 558, + 482, + 604 + ], + "spans": [ + { + "bbox": [ + 132, + 558, + 482, + 604 + ], + "type": "text", + "content": "Fig. 4: Qualitative results of onelabel setting on NYU-V2. The first row shows the input image, the second row represents the ground-truth or predictions of semantic segmentation, the third row plots the ground-truth or predictions of depth estimation, and the final row presents the ground-truth or predictions of surface normal estimation." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 629, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 629, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 629, + 482, + 666 + ], + "type": "text", + "content": "on SegNet, which utilizes the same backbone. In all the methods compared using the same backbone, our approach demonstrates significant advantages. In the one label setting, the patch-aware Gaussian contrast method demonstrates" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 214, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 214, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 214, + 101 + ], + "type": "text", + "content": "M. Li et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 160, + 114, + 455, + 269 + ], + "blocks": [ + { + "bbox": [ + 160, + 114, + 455, + 269 + ], + "lines": [ + { + "bbox": [ + 160, + 114, + 455, + 269 + ], + "spans": [ + { + "bbox": [ + 160, + 114, + 455, + 269 + ], + "type": "image", + "image_path": "e030ba3de3ea14eca303112f5385202e20fe63ec6a7820c2772c12b33d83c7a1.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 284, + 482, + 339 + ], + "lines": [ + { + "bbox": [ + 130, + 284, + 482, + 339 + ], + "spans": [ + { + "bbox": [ + 130, + 284, + 482, + 339 + ], + "type": "text", + "content": "Fig. 5: Qualitative results of random label setting on NYU-V2. The first row shows the input image, the second row represents the ground-truth or predictions of semantic segmentation, the third row plots the ground-truth or predictions of depth estimation, and the final row presents the ground-truth or predictions of surface normal estimation." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 371, + 482, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 371, + 482, + 479 + ], + "spans": [ + { + "bbox": [ + 130, + 371, + 482, + 479 + ], + "type": "text", + "content": "significant improvements across all three tasks compared to Li et al. [25] across all three tasks, as it achieves a certain degree of local-level cross-task alignment. Furthermore, the region-aware Gaussian contrast method further enhances performance, achieving improvements of " + }, + { + "bbox": [ + 130, + 371, + 482, + 479 + ], + "type": "inline_equation", + "content": "2.48\\%" + }, + { + "bbox": [ + 130, + 371, + 482, + 479 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 371, + 482, + 479 + ], + "type": "inline_equation", + "content": "2.25\\%" + }, + { + "bbox": [ + 130, + 371, + 482, + 479 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 371, + 482, + 479 + ], + "type": "inline_equation", + "content": "5.51\\%" + }, + { + "bbox": [ + 130, + 371, + 482, + 479 + ], + "type": "text", + "content": " over the SOTA [1] across the three tasks, owing to the achievement of cross-task consistency at the object level. In the random label setting, the superiority of the region-aware Gaussian contrast method can be observed, as it outperforms other methods significantly across all three tasks, yielding improvements of " + }, + { + "bbox": [ + 130, + 371, + 482, + 479 + ], + "type": "inline_equation", + "content": "4.70\\%" + }, + { + "bbox": [ + 130, + 371, + 482, + 479 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 371, + 482, + 479 + ], + "type": "inline_equation", + "content": "4.18\\%" + }, + { + "bbox": [ + 130, + 371, + 482, + 479 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 371, + 482, + 479 + ], + "type": "inline_equation", + "content": "2.85\\%" + }, + { + "bbox": [ + 130, + 371, + 482, + 479 + ], + "type": "text", + "content": ", respectively, over the SOTA [1]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 482, + 482, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 482, + 482, + 567 + ], + "spans": [ + { + "bbox": [ + 130, + 482, + 482, + 567 + ], + "type": "text", + "content": "Furthermore, we provide a comparison with the state-of-the-art fully supervised method, TaskExpert [44]. Tab. 1 clearly shows that our method significantly improves the performance of TaskExpert in partially supervised scenarios. Particularly in the random setting, when combined with our method, TaskExpert achieves even better performance than the fully supervised TaskExpert in the surface normal task. This once again demonstrates the effectiveness and wide applicability of our approach." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": "Quantitative Results on Cityscapes Dataset. The evaluation results are describe in Tab. 2. Similar to NYU-V2, we first compared Li et al. [25] with HRNet18 as the backbone. Subsequently, we discussed the results of the region-aware contrast method under different contrast strategies. It is observed that Gaussian contrast demonstrates improvements in semantic segmentation and depth estimation compared to vector-based methods, indicating that representing regions using Gaussian distributions provides a more comprehensive and accurate representation than using vectors. In terms of computational speed, the" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 301, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 301, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 301, + 91, + 447, + 102 + ], + "type": "text", + "content": "Region-aware Distribution Contrast" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 184, + 146, + 430, + 174 + ], + "blocks": [ + { + "bbox": [ + 131, + 114, + 482, + 137 + ], + "lines": [ + { + "bbox": [ + 131, + 114, + 482, + 137 + ], + "spans": [ + { + "bbox": [ + 131, + 114, + 482, + 137 + ], + "type": "text", + "content": "Table 4: Ablation study on the NYU-V2 dataset to investigate the contribution of SAM in the fully supervised setting." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 184, + 146, + 430, + 174 + ], + "lines": [ + { + "bbox": [ + 184, + 146, + 430, + 174 + ], + "spans": [ + { + "bbox": [ + 184, + 146, + 430, + 174 + ], + "type": "table", + "html": "
MethodVector GaussianSemantic(mIoU)↑Depth(aErr)↓Normals(mErr)↓
Region-aware Contrast✓49.000.42068.1878
Region-aware Contrast (Ours)✓51.140.39824.8108
", + "image_path": "1294417756ecfaffc98f834f88a5edf2ba30c88e07d527ff9fa7844bc8879f22.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 196, + 479, + 220 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 196, + 479, + 220 + ], + "spans": [ + { + "bbox": [ + 130, + 196, + 479, + 220 + ], + "type": "text", + "content": "Gaussian contrast method is faster and more efficient, while still performing on par with pixel-based contrast methods." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 238, + 274, + 250 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 238, + 274, + 250 + ], + "spans": [ + { + "bbox": [ + 131, + 238, + 274, + 250 + ], + "type": "text", + "content": "4.2 Qualitative Evaluation" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 258, + 482, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 258, + 482, + 434 + ], + "spans": [ + { + "bbox": [ + 130, + 258, + 482, + 434 + ], + "type": "text", + "content": "The qualitative evaluation of onelabel setting, random label setting on the NYU-V2 dataset, and onelabel setting on the Cityscapes dataset are shown in Fig. 4, Fig. 5 and Fig. 6, respectively. We present a comparative analysis of the results from two sets of images for each setting of each dataset. As indicated by Fig. 4 and Fig. 5, it can be observed that our method exhibits superior accuracy and smoothness in semantic segmentation, which is particularly evident in Fig. 5. Regarding depth estimation, our method, which is based on region-aware Gaussian distribution contrast, achieves more accurate predictions when objects undergo changes compared to Li et al. [25]. The latter fails to capture such variations effectively. As for surface normal estimation, our method outperforms Li et al. [25] overall and demonstrates finer contour handling. These observations highlight the advantages of our approach over Li et al. [25] in terms of classification accuracy, smoothness in semantic segmentation, accurate depth estimation in the presence of object changes, and finer handling of surface normals." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 437, + 482, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 437, + 482, + 509 + ], + "spans": [ + { + "bbox": [ + 130, + 437, + 482, + 509 + ], + "type": "text", + "content": "We further exhibit the multi-task prediction results on the Cityscapes dataset in Fig. 6. Benefited from our proposed region-aware distribution contrast learning method, both the semantic segmentation and depth estimation tasks exhibit improved capability in capturing object edges and fine-grained details compared to Li et al. [25]. Additionally, there are significant overall performance improvement at a global level." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 526, + 239, + 538 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 526, + 239, + 538 + ], + "spans": [ + { + "bbox": [ + 131, + 526, + 239, + 538 + ], + "type": "text", + "content": "4.3 Ablation Study" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 546, + 481, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 546, + 481, + 569 + ], + "spans": [ + { + "bbox": [ + 130, + 546, + 481, + 569 + ], + "type": "text", + "content": "We conduct ablation study on NYU-V2 dataset to prove the effectiveness of our method from various aspects." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "text", + "content": "Local Extraction Strategy. Tab. 3 provides a comparison between different local extraction strategies. All methods are compared based on HRNet18 as the backbone. It can be observed that in both the onelabel setting and random label setting, regardless of the contrast strategy employed, the region-aware contrast method significantly outperforms the patch-aware contrast method. This is because the region-aware method takes into account the objects present in the image and performs contrast based on objects, rather than simply comparing patches that may contain multiple objects." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 214, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 214, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 214, + 101 + ], + "type": "text", + "content": "M. Li et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 160, + 113, + 455, + 274 + ], + "blocks": [ + { + "bbox": [ + 160, + 113, + 455, + 274 + ], + "lines": [ + { + "bbox": [ + 160, + 113, + 455, + 274 + ], + "spans": [ + { + "bbox": [ + 160, + 113, + 455, + 274 + ], + "type": "image", + "image_path": "87aec302bd7afacf1602e4c7516c97e91d078af20e021e6b2fde90c6fb645094.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 284, + 482, + 327 + ], + "lines": [ + { + "bbox": [ + 130, + 284, + 482, + 327 + ], + "spans": [ + { + "bbox": [ + 130, + 284, + 482, + 327 + ], + "type": "text", + "content": "Fig.6: Qualitative results of onelabel setting on Cityscapes. The first row shows the input image, the second row represents the ground-truth or predictions of semantic segmentation, and the final row plots the ground-truth or predictions of depth estimation." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 176, + 376, + 440, + 434 + ], + "blocks": [ + { + "bbox": [ + 130, + 343, + 482, + 366 + ], + "lines": [ + { + "bbox": [ + 130, + 343, + 482, + 366 + ], + "spans": [ + { + "bbox": [ + 130, + 343, + 482, + 366 + ], + "type": "text", + "content": "Table 5: Ablation study on the NYU-V2 dataset to investigate the performance of various distance measurements." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 176, + 376, + 440, + 434 + ], + "lines": [ + { + "bbox": [ + 176, + 376, + 440, + 434 + ], + "spans": [ + { + "bbox": [ + 176, + 376, + 440, + 434 + ], + "type": "table", + "html": "
Label settingMethodTasks
Semantic(mIoU)↑Depth(aErr)↓Normals(mErr)↓
onelabelKL divergence41.150.47394.94
onelabelWasserstein distance (Ours)42.280.46414.86
random labelKL divergence45.530.45395.04
random labelWasserstein distance (Ours)46.210.44824.49
", + "image_path": "348fd584912ca036c312facb44d88746323e958ab3d5fa512631126c0ca6e30e.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 460, + 482, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 460, + 482, + 567 + ], + "spans": [ + { + "bbox": [ + 130, + 460, + 482, + 567 + ], + "type": "text", + "content": "Local Contrast Strategy. Tab. 3 also provides a comparison between different local contrast strategies. All methods are similarly compared based on HRNet18. It can be observed that regardless of the label setting and the local extraction strategy employed, the Gaussian distribution contrast consistently outperforms the vector contrast and approaches or even surpasses the pixel contrast. This is because multi-task learning scenarios involve discrete and continuous dense predictions, utilizing Gaussian distributions to represent regions for contrast allows for better cross-task alignment in the overall distribution. In contrast, the hard alignment strategy of pixel contrast may lead to inaccuracies." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "type": "text", + "content": "Contribution of SAM. As shown in Tab. 2, Tab. 3, and Tab. 6 in our article, although the vector contrast approach also incorporates SAM, it fails to yield satisfactory results. In contrast, the adoption of distribution contrast exhibits a significant improvement over the vector contrast approach. To further demonstrate this point, we provide a comparison of results using semantic segmentation labels instead of SAM in fully supervised setting, as shown in Tab. 4. This further emphasizes that solely introducing fine-grained segmentation results is insufficient, highlighting the importance of the distribution contrast method." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 301, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 301, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 301, + 91, + 447, + 102 + ], + "type": "text", + "content": "Region-aware Distribution Contrast" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 176, + 146, + 440, + 203 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 482, + 137 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 482, + 137 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 482, + 137 + ], + "type": "text", + "content": "Table 6: Ablation study on the NYU-V2 dataset to investigate the performance of various methods in the fully supervised setting." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 176, + 146, + 440, + 203 + ], + "lines": [ + { + "bbox": [ + 176, + 146, + 440, + 203 + ], + "spans": [ + { + "bbox": [ + 176, + 146, + 440, + 203 + ], + "type": "table", + "html": "
MethodContrast strategyTasks
Vector Pixel GaussianSemantic(mIoU)↑Depth(aErr)↓Normals(mErr)↓
MTL46.380.46606.7298
Li et al. [25]47.200.44775.1192
Region-aware Contrast✓47.640.42315.1072
Region-aware Contrast✓50.060.41684.5318
Region-aware Contrast (Ours)✓50.830.40513.2878
", + "image_path": "ab9b042b7756ec114a842b482fa89ce7884e087b9ac65e7b02e6c4bab120f5bb.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 224, + 481, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 224, + 481, + 331 + ], + "spans": [ + { + "bbox": [ + 130, + 224, + 481, + 331 + ], + "type": "text", + "content": "Distance Measurement for Distributions. Tab. 5 presents a comparison of different distance measurements. While KL divergence is commonly used to calculate the distance between Gaussian distributions, it fails to accurately reflect the distance between discrete and continuous distributions, which is crucial in the context of multi-task alignment. On the other hand, Wasserstein distance has the advantage of measuring the distance between two non-overlapping distributions. The experimental results in Tab. 5 demonstrate that Wasserstein distance is better suited for evaluating the distance between different task distributions. More detailed information can be found in the supplementary materials." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 332, + 481, + 416 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 332, + 481, + 416 + ], + "spans": [ + { + "bbox": [ + 130, + 332, + 481, + 416 + ], + "type": "text", + "content": "Fully Supervised Setting. Tab. 6 presents the comparison of our method in the fully supervised setting. All methods are based on HRNet18. Compared to other methods, the region-aware distribution contrast method once again demonstrates superiority, outperforming significantly in all three tasks. This is because Gaussian distributions can characterize the distribution characteristics of different tasks and facilitate accurate and efficient alignment with both vector and pixel-based approaches." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 435, + 220, + 449 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 435, + 220, + 449 + ], + "spans": [ + { + "bbox": [ + 132, + 435, + 220, + 449 + ], + "type": "text", + "content": "5 Conclusion" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 462, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 462, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 462, + 481, + 665 + ], + "type": "text", + "content": "In this paper, we propose a novel region-aware distribution contrast learning method for MTPSL. To facilitate cross-task alignment at the region level, predictions from different tasks are initially mapped into a joint feature map space. Simultaneously, SAM is employed to generate regions for each training sample, leveraging its ability to provide valuable prior knowledge. Subsequently, region-wise features are modeled by fitting a Gaussian distribution. The alignment among regions is achieved by minimizing the disparity among distributions from different tasks within the same region and maximizing the distance between distributions from different regions. The effectiveness of our method is demonstrated on two standard multi-task datasets through extensive experiments, outperforming the state-of-the-art by a significant margin. Although our method is designed for MTPSL, it can be applied to general multi-task learning problems to leverage the intrinsic relationships across different tasks, as demonstrated in the fully supervised experiments in Sec. 4.3. Moreover, we believe that the proposed region-aware distribution contrast learning method provides a promising way for solving more general multi-modal problems, and we will investigate this in our future work." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 214, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 214, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 214, + 101 + ], + "type": "text", + "content": "M. Li et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 164 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 164 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 164 + ], + "type": "text", + "content": "Acknowledgements: This work was supported in part by the National Natural Science Foundation of China under grant U23B2011, 62102069, U20B2063 and 62220106008, the Key R&D Program of Zhejiang under grant 2024SSYS0091, the Sichuan Science and Technology Program under Grant 2024NSFTD0034." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 184, + 198, + 197 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 184, + 198, + 197 + ], + "spans": [ + { + "bbox": [ + 132, + 184, + 198, + 197 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 211, + 481, + 666 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 138, + 211, + 481, + 255 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 211, + 481, + 255 + ], + "spans": [ + { + "bbox": [ + 138, + 211, + 481, + 255 + ], + "type": "text", + "content": "1. Borse, S., Das, D., Park, H., Cai, H., Garrepalli, R., Porikli, F.: Dejavu: Conditional regenerative learning to enhance dense prediction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 19466-19477 (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 256, + 481, + 289 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 256, + 481, + 289 + ], + "spans": [ + { + "bbox": [ + 138, + 256, + 481, + 289 + ], + "type": "text", + "content": "2. Brüggemann, D., Kanakis, M., Obukhov, A., Georgoulis, S., Van Gool, L.: Exploring relational context for multi-task dense prediction. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 15869-15878 (2021)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 289, + 481, + 322 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 289, + 481, + 322 + ], + "spans": [ + { + "bbox": [ + 138, + 289, + 481, + 322 + ], + "type": "text", + "content": "3. Chang, J., Lan, Z., Cheng, C., Wei, Y.: Data uncertainty learning in face recognition. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5710-5719 (2020)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 322, + 481, + 355 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 322, + 481, + 355 + ], + "spans": [ + { + "bbox": [ + 138, + 322, + 481, + 355 + ], + "type": "text", + "content": "4. Chen, S., Xie, G., Liu, Y., Peng, Q., Sun, B., Li, H., You, X., Shao, L.: Hsva: Hierarchical semantic-visual adaptation for zero-shot learning. Advances in Neural Information Processing Systems 34, 16622-16634 (2021)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 355, + 481, + 388 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 355, + 481, + 388 + ], + "spans": [ + { + "bbox": [ + 138, + 355, + 481, + 388 + ], + "type": "text", + "content": "5. Chen, T., Kornblith, S., Norouzi, M., Hinton, G.: A simple framework for contrastive learning of visual representations. In: International Conference on Machine Learning. pp. 1597-1607. PMLR (2020)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 388, + 481, + 411 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 388, + 481, + 411 + ], + "spans": [ + { + "bbox": [ + 138, + 388, + 481, + 411 + ], + "type": "text", + "content": "6. Chen, X., Fan, H., Girshick, R., He, K.: Improved baselines with momentum contrastive learning. arXiv preprint arXiv:2003.04297 (2020)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 411, + 481, + 444 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 411, + 481, + 444 + ], + "spans": [ + { + "bbox": [ + 138, + 411, + 481, + 444 + ], + "type": "text", + "content": "7. Chen, Z., Badrinarayanan, V., Lee, C.Y., Rabinovich, A.: GradNorm: Gradient normalization for adaptive loss balancing in deep multi-task networks. In: International Conference on Machine Learning. pp. 794-803. PMLR (2018)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 444, + 481, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 444, + 481, + 477 + ], + "spans": [ + { + "bbox": [ + 138, + 444, + 481, + 477 + ], + "type": "text", + "content": "8. Chen, Z., Zhu, L., Wan, L., Wang, S., Feng, W., Heng, P.A.: A multi-task mean teacher for semi-supervised shadow detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5611-5620 (2020)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 477, + 481, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 477, + 481, + 521 + ], + "spans": [ + { + "bbox": [ + 138, + 477, + 481, + 521 + ], + "type": "text", + "content": "9. Cordts, M., Omran, M., Ramos, S., Rehfeld, T., Enzweiler, M., Benenson, R., Franke, U., Roth, S., Schiele, B.: The cityscapes dataset for semantic urban scene understanding. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 3213-3223 (2016)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 521, + 481, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 521, + 481, + 555 + ], + "spans": [ + { + "bbox": [ + 138, + 521, + 481, + 555 + ], + "type": "text", + "content": "0. Eigen, D., Fergus, R.: Predicting depth, surface normals and semantic labels with a common multi-scale convolutional architecture. In: Proceedings of the IEEE International Conference on Computer Vision. pp. 2650-2658 (2015)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 555, + 481, + 587 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 555, + 481, + 587 + ], + "spans": [ + { + "bbox": [ + 138, + 555, + 481, + 587 + ], + "type": "text", + "content": "1. Eigen, D., Puhrsch, C., Fergus, R.: Depth map prediction from a single image using a multi-scale deep network. Advances in Neural Information Processing Systems 27 (2014)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 587, + 481, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 587, + 481, + 621 + ], + "spans": [ + { + "bbox": [ + 138, + 587, + 481, + 621 + ], + "type": "text", + "content": "2. Fan, R., Poggi, M., Mattoccia, S.: Contrastive learning for depth prediction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3225-3236 (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 621, + 481, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 621, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 138, + 621, + 481, + 666 + ], + "type": "text", + "content": "3. Gao, Y., Ma, J., Zhao, M., Liu, W., Yuille, A.L.: Nddr-cnn: Layerwise feature fusing in multi-task cnns by neural discriminative dimensionality reduction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3205-3214 (2019)" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 301, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 301, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 301, + 91, + 447, + 102 + ], + "type": "text", + "content": "Region-aware Distribution Contrast" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "type": "text", + "content": "14. Guo, M., Haque, A., Huang, D.A., Yeung, S., Fei-Fei, L.: Dynamic task prioritization for multitask learning. In: Proceedings of the European Conference on Computer Vision. pp. 270-287 (2018)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 150, + 481, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 150, + 481, + 184 + ], + "spans": [ + { + "bbox": [ + 132, + 150, + 481, + 184 + ], + "type": "text", + "content": "15. He, K., Fan, H., Wu, Y., Xie, S., Girshick, R.: Momentum contrast for unsupervised visual representation learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9729-9738 (2020)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 185, + 480, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 185, + 480, + 206 + ], + "spans": [ + { + "bbox": [ + 132, + 185, + 480, + 206 + ], + "type": "text", + "content": "16. He, K., Gkioxari, G., Dólár, P., Girshick, R.: Mask r-cnn. In: Proceedings of the IEEE International Conference on Computer Vision. pp. 2961-2969 (2017)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 207, + 480, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 207, + 480, + 239 + ], + "spans": [ + { + "bbox": [ + 132, + 207, + 480, + 239 + ], + "type": "text", + "content": "17. Hu, H., Cui, J., Wang, L.: Region-aware contrastive learning for semantic segmentation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 16291-16301 (2021)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 240, + 480, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 240, + 480, + 274 + ], + "spans": [ + { + "bbox": [ + 132, + 240, + 480, + 274 + ], + "type": "text", + "content": "18. Huang, C., Tang, H., Fan, W., Xiao, Y., Hao, D., Qian, Z., Terzopoulos, D., et al.: Partly supervised multi-task learning. In: 2020 19th IEEE International Conference on Machine Learning and Applications. pp. 769-774. IEEE (2020)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 274, + 480, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 274, + 480, + 319 + ], + "spans": [ + { + "bbox": [ + 132, + 274, + 480, + 319 + ], + "type": "text", + "content": "19. Imran, A.A.Z., Terzopoulos, D.: Semi-supervised multi-task learning with chest x-ray images. In: Machine Learning in Medical Imaging: 10th International Workshop, MLMI 2019, Held in Conjunction with MICCAI 2019, Shenzhen, China, October 13, 2019, Proceedings 10. pp. 151-159. Springer (2019)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 319, + 480, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 319, + 480, + 352 + ], + "spans": [ + { + "bbox": [ + 132, + 319, + 480, + 352 + ], + "type": "text", + "content": "20. Jin, X., Lan, C., Zeng, W., Chen, Z.: Global distance-distributions separation for unsupervised person re-identification. In: Proceedings of the European Conference on Computer Vision. pp. 735–751. Springer (2020)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 352, + 480, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 352, + 480, + 373 + ], + "spans": [ + { + "bbox": [ + 132, + 352, + 480, + 373 + ], + "type": "text", + "content": "21. Jin, Z., Chen, S., Chen, Y., Xu, Z., Feng, H.: Let segment anything help image dehaze. arXiv preprint arXiv:2306.15870 (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 374, + 480, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 374, + 480, + 407 + ], + "spans": [ + { + "bbox": [ + 132, + 374, + 480, + 407 + ], + "type": "text", + "content": "22. Kendall, A., Gal, Y., Cipolla, R.: Multi-task learning using uncertainty to weigh losses for scene geometry and semantics. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 7482-7491 (2018)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 408, + 480, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 408, + 480, + 453 + ], + "spans": [ + { + "bbox": [ + 132, + 408, + 480, + 453 + ], + "type": "text", + "content": "23. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., Dollar, P., Girshick, R.: Segment anything. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 4015-4026 (October 2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 453, + 480, + 486 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 453, + 480, + 486 + ], + "spans": [ + { + "bbox": [ + 132, + 453, + 480, + 486 + ], + "type": "text", + "content": "24. Latif, S., Rana, R., Khalifa, S., Jurdak, R., Epps, J., Schuller, B.W.: Multi-task semi-supervised adversarial autoencoding for speech emotion recognition. IEEE Transactions on Affective Computing 13(2), 992-1004 (2020)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 486, + 480, + 518 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 486, + 480, + 518 + ], + "spans": [ + { + "bbox": [ + 132, + 486, + 480, + 518 + ], + "type": "text", + "content": "25. Li, W.H., Liu, X., Bilen, H.: Learning multiple dense prediction tasks from partially annotated data. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18879-18889 (2022)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 520, + 480, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 520, + 480, + 553 + ], + "spans": [ + { + "bbox": [ + 132, + 520, + 480, + 553 + ], + "type": "text", + "content": "26. Liang, C., Wang, W., Miao, J., Yang, Y.: Gmmseg: Gaussian mixture based generative semantic segmentation models. Advances in Neural Information Processing Systems 35, 31360-31375 (2022)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 554, + 480, + 575 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 554, + 480, + 575 + ], + "spans": [ + { + "bbox": [ + 132, + 554, + 480, + 575 + ], + "type": "text", + "content": "27. Liu, Q., Liao, X., Carin, L.: Semi-supervised multitask learning. Advances in Neural Information Processing Systems 20 (2007)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 576, + 480, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 576, + 480, + 597 + ], + "spans": [ + { + "bbox": [ + 132, + 576, + 480, + 597 + ], + "type": "text", + "content": "28. Liu, S., James, S., Davison, A., Johns, E.: Auto-lambda: Disentangling dynamic task relationships. Transactions on Machine Learning Research (2022)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 599, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 599, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 599, + 480, + 632 + ], + "type": "text", + "content": "29. Liu, S., Johns, E., Davison, A.J.: End-to-end multi-task learning with attention. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1871-1880 (2019)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "type": "text", + "content": "30. Long, J., Shelhamer, E., Darrell, T.: Fully convolutional networks for semantic segmentation. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 3431-3440 (2015)" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 214, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 214, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 214, + 100 + ], + "type": "text", + "content": "M. Li et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 665 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 160 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 160 + ], + "type": "text", + "content": "31. Lu, Y., Kumar, A., Zhai, S., Cheng, Y., Javidi, T., Feris, R.: Fully-adaptive feature sharing in multi-task networks with applications in person attribute classification. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 5334-5343 (2017)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 161, + 481, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 161, + 481, + 194 + ], + "spans": [ + { + "bbox": [ + 130, + 161, + 481, + 194 + ], + "type": "text", + "content": "32. Misra, I., Shrivastava, A., Gupta, A., Hebert, M.: Cross-Stitch networks for multitask learning. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 3994-4003 (2016)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 194, + 481, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 194, + 481, + 216 + ], + "spans": [ + { + "bbox": [ + 132, + 194, + 481, + 216 + ], + "type": "text", + "content": "33. Oord, A.v.d., Li, Y., Vinyals, O.: Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748 (2018)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 216, + 481, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 216, + 481, + 248 + ], + "spans": [ + { + "bbox": [ + 132, + 216, + 481, + 248 + ], + "type": "text", + "content": "34. Poggi, M., Aleotti, F., Tosi, F., Mattoccia, S.: On the uncertainty of self-supervised monocular depth estimation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3227-3237 (2020)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 249, + 481, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 249, + 481, + 270 + ], + "spans": [ + { + "bbox": [ + 132, + 249, + 481, + 270 + ], + "type": "text", + "content": "35. Rüschendorf, L.: The Wasserstein distance and approximation theorems. Probability Theory and Related Fields 70(1), 117-129 (1985)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 271, + 481, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 271, + 481, + 303 + ], + "spans": [ + { + "bbox": [ + 132, + 271, + 481, + 303 + ], + "type": "text", + "content": "36. Silberman, N., Hoiem, D., Kohli, P., Fergus, R.: Indoor segmentation and support inference from rgb images. In: Proceedings of the European Conference on Computer Vision. pp. 746-760. Springer (2012)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 304, + 481, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 304, + 481, + 335 + ], + "spans": [ + { + "bbox": [ + 132, + 304, + 481, + 335 + ], + "type": "text", + "content": "37. Standley, T., Zamir, A., Chen, D., Guibas, L., Malik, J., Savarese, S.: Which tasks should be learned together in multi-task learning? In: International Conference on Machine Learning. pp. 9120-9132. PMLR (2020)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 336, + 481, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 336, + 481, + 357 + ], + "spans": [ + { + "bbox": [ + 132, + 336, + 481, + 357 + ], + "type": "text", + "content": "38. Tian, Y., Krishnan, D., Isola, P.: Contrastive multiview coding. In: Proceedings of the European Conference on Computer Vision. pp. 776-794. Springer (2020)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 358, + 481, + 390 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 358, + 481, + 390 + ], + "spans": [ + { + "bbox": [ + 132, + 358, + 481, + 390 + ], + "type": "text", + "content": "39. Vandenhende, S., Georgoulis, S., Van Gool, L.: MTI-Net: Multi-scale task interaction networks for multi-task learning. In: Proceedings of the European Conference on Computer Vision. pp. 527-543. Springer (2020)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 391, + 481, + 424 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 391, + 481, + 424 + ], + "spans": [ + { + "bbox": [ + 132, + 391, + 481, + 424 + ], + "type": "text", + "content": "40. Wang, X., Zhang, R., Shen, C., Kong, T., Li, L.: Dense contrastive learning for self-supervised visual pre-training. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3024-3033 (2021)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 424, + 481, + 467 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 424, + 481, + 467 + ], + "spans": [ + { + "bbox": [ + 132, + 424, + 481, + 467 + ], + "type": "text", + "content": "41. Wu, L., Zhong, Z., Fang, L., He, X., Liu, Q., Ma, J., Chen, H.: Sparsely annotated semantic segmentation with adaptive gaussian mixtures. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15454-15464 (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 468, + 481, + 501 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 468, + 481, + 501 + ], + "spans": [ + { + "bbox": [ + 132, + 468, + 481, + 501 + ], + "type": "text", + "content": "42. Wu, Z., Xiong, Y., Yu, S.X., Lin, D.: Unsupervised feature learning via nonparametric instance discrimination. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 3733-3742 (2018)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 501, + 481, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 501, + 481, + 544 + ], + "spans": [ + { + "bbox": [ + 132, + 501, + 481, + 544 + ], + "type": "text", + "content": "43. Xu, D., Ouyang, W., Wang, X., Sebe, N.: PAD-Net: Multi-tasks guided prediction-and-distillation network for simultaneous depth estimation and scene parsing. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 675-684 (2018)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 544, + 481, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 544, + 481, + 577 + ], + "spans": [ + { + "bbox": [ + 132, + 544, + 481, + 577 + ], + "type": "text", + "content": "44. Ye, H., Xu, D.: Taskexpert: Dynamically assembling multi-task representations with memorial mixture-of-experts. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 21828-21837 (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 577, + 481, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 577, + 481, + 610 + ], + "spans": [ + { + "bbox": [ + 132, + 577, + 481, + 610 + ], + "type": "text", + "content": "45. Zamir, A.R., Sax, A., Cheerla, N., Suri, R., Cao, Z., Malik, J., Guibas, L.J.: Robust learning through cross-task consistency. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11197-11206 (2020)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 611, + 481, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 611, + 481, + 643 + ], + "spans": [ + { + "bbox": [ + 132, + 611, + 481, + 643 + ], + "type": "text", + "content": "46. Zamir, A.R., Sax, A., Shen, W., Guibas, L.J., Malik, J., Savarese, S.: Taskonomy: Disentangling task transfer learning. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 3712-3722 (2018)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 643, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 643, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 643, + 481, + 665 + ], + "type": "text", + "content": "47. Zhang, J., Fan, D.P., Dai, Y., Anwar, S., Saleh, F.S., Zhang, T., Barnes, N.: Uc-net: Uncertainty inspired rgb-d saliency detection via conditional variational autoen" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 301, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 301, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 301, + 91, + 447, + 102 + ], + "type": "text", + "content": "Region-aware Distribution Contrast" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 248 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 147, + 116, + 481, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 147, + 116, + 481, + 138 + ], + "spans": [ + { + "bbox": [ + 147, + 116, + 481, + 138 + ], + "type": "text", + "content": "coders. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8582-8591 (2020)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 140, + 481, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 140, + 481, + 171 + ], + "spans": [ + { + "bbox": [ + 133, + 140, + 481, + 171 + ], + "type": "text", + "content": "48. Zhang, L., Liu, X., Guan, H.: Automtl: A programming framework for automating efficient multi-task learning. Advances in Neural Information Processing Systems 35, 34216-34228 (2022)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 172, + 481, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 172, + 481, + 205 + ], + "spans": [ + { + "bbox": [ + 132, + 172, + 481, + 205 + ], + "type": "text", + "content": "49. Zhang, M., Zhao, X., Yao, J., Yuan, C., Huang, W.: When noisy labels meet long tail dilemmas: A representation calibration method. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 15890-15900 (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 205, + 481, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 205, + 481, + 248 + ], + "spans": [ + { + "bbox": [ + 132, + 205, + 481, + 248 + ], + "type": "text", + "content": "50. Zhang, Z., Cui, Z., Xu, C., Yan, Y., Sebe, N., Yang, J.: Pattern-affinitive propagation across depth, surface normal and semantic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4106-4115 (2019)" + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 214, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 214, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 214, + 100 + ], + "type": "text", + "content": "M. Li et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Region-centric Image-Language Pretraining for Open-Vocabulary Detection/be279718-23a8-4b0a-ac0c-2bca133ca6e7_content_list.json b/2024/Region-centric Image-Language Pretraining for Open-Vocabulary Detection/be279718-23a8-4b0a-ac0c-2bca133ca6e7_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..9c3be77f4ce6f8800ea41556651ae6f8abdfc0c2 --- /dev/null +++ b/2024/Region-centric Image-Language Pretraining for Open-Vocabulary Detection/be279718-23a8-4b0a-ac0c-2bca133ca6e7_content_list.json @@ -0,0 +1,1633 @@ +[ + { + "type": "text", + "text": "Region-centric Image-Language Pretraining for Open-Vocabulary Detection", + "text_level": 1, + "bbox": [ + 233, + 140, + 772, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Dahun Kim Anelia Angelova Weicheng Kuo", + "bbox": [ + 326, + 212, + 676, + 228 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Google DeepMind", + "bbox": [ + 439, + 239, + 562, + 253 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. We present a new open-vocabulary detection approach based on region-centric image-language pretraining to bridge the gap between image-level pretraining and open-vocabulary object detection. At the pretraining phase, we incorporate the detector architecture on top of the classification backbone, which better serves the region-level recognition needs of detection by enabling the detector heads to learn from large-scale image-text pairs. Using only standard contrastive loss and no pseudo-labeling, our approach is a simple yet effective extension of the contrastive learning method to learn emergent object-semantic cues. In addition, we propose a shifted-window learning approach upon window attention to make the backbone representation more robust, translation-invariant, and less biased by the window pattern. On the popular LVIS open-vocabulary detection benchmark, our approach sets a new state of the art of 37.6 mask $\\mathrm{AP}_r$ using the common ViT-L backbone and public LAION dataset, and 40.5 mask $\\mathrm{AP}_r$ using the DataComp-1B dataset, significantly outperforming the best existing approach by $+3.7$ mask $\\mathrm{AP}_r$ . On the COCO benchmark, we achieve very competitive 39.6 novel AP without pseudo labeling or weak supervision. In addition, we evaluate our approach on the transfer detection setup, where it demonstrates notable improvement over the baseline. Visualization reveals emerging object locality from the pretraining recipes compared to the baseline. $^1$", + "bbox": [ + 261, + 292, + 743, + 584 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 608, + 375, + 625 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Understanding and localizing all objects in the visual world is a key problem in computer vision, crucial for applications like self-driving cars. Traditional object detectors rely on costly human annotations, which limits scalability. Open-vocabulary detection (OVD) [41] addresses this by pretraining on large-scale image-text data (e.g. CLIP [28]), enabling detectors to identify objects based on user-provided text queries unavailable during training.", + "bbox": [ + 212, + 633, + 787, + 724 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Most existing open-vocabulary detection works assume the pretrained CLIP backbone is given, and focus on techniques such as knowledge distillation [5, 10], weak supervision [46], pseudo labeling [13, 29, 43, 44], and frozen backbone application [18], using the pretrained backbone. Consequently, during detection finetuning, the detector heads often need to be trained from scratch. This tends to result in sub-optimal generalization because the detector heads are trained on", + "bbox": [ + 212, + 724, + 787, + 816 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "1 project page: github.com/google-research/google-research/tree/master/fvlm/dito", + "bbox": [ + 217, + 824, + 772, + 840 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/1bb6a16bc741c0f47086655e1b9ec48d486a66a2f29c88a22e6848703434a665.jpg", + "image_caption": [ + "Fig. 1: DITO overview (left): Existing image-text pretraining methods for open-vocabulary detection [16,17] update only the vision transformer backbone and finetune detector heads from scratch. We propose to pretrain both backbone and detector heads directly from the large-scale image-text paired data, without a need for pseudo labeling or box annotations. In open-vocabulary detection finetuning, we introduce a simple shifted-window learning method to produce more robust representations from the pretrained vision transformer. DITO prediction (right): LVIS results only show the novel categories (e.g., bulldozer, fishbowl, subwoofer, heron). While Ego4D is a real-world and out-of-distribution data, many unseen objects are detected (e.g., steel lid, sticker on the wall, recycle bin). Best viewed with zoom in." + ], + "image_footnote": [], + "bbox": [ + 256, + 143, + 483, + 321 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/62a73a1fd804c918a98fc141e92f52ef8bb1637939eb2abae175fe7e9522da84.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 517, + 143, + 630, + 321 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/c27ddce40d35c2ede58ada8a9d98d6c2b2d4ad3852e169f6b5fe8c778d0492c4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 638, + 143, + 756, + 321 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "the limited vocabulary of detection datasets, while only the backbone contains the knowledge of open-vocabulary concepts.", + "bbox": [ + 212, + 489, + 785, + 520 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Several studies have integrated detection models into CLIP pretraining. For instance, RegionCLIP [44] employs an off-the-shelf detector during CLIP pretraining to generate pseudo region-text labels. However, their pretraining only updates the image backbone, with the detector heads exclusively trained on detection data. These pseudo labeling-based pretraining methods [25, 44] require multi-stage training and handcrafted processing of pseudo labels, increasing complexity and training costs. Similarly, approaches such as GLIP [19, 42], Grounding DINO [22], CoDet [24] and DetCLIP [38, 39] integrate detector architectures in CLIP training, but their joint training demands additional detection and visual grounding datasets and complex multitask learning setups.", + "bbox": [ + 212, + 521, + 787, + 672 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we propose a simple yet effective solution, Region-centric Pretraining approach, which incorporates detector modules into CLIP pretraining without the need for pseudo labeling or box annotations. This method involves generating random box regions across feature pyramid levels, followed by feature pooling over these regions. Subsequently, an image-text contrastive loss is applied, encouraging text-aligned region features to contribute more to the whole image representation. Our approach not only facilitates the warm-starting of detector heads in finetuning, but also leads to emergent representations with more localized semantic information compared to the baseline CLIP backbone, as demonstrated in our experiments. Compared to pseudo-labeling techniques [35,44-46], our approach can be viewed as an extension of contrastive", + "bbox": [ + 212, + 674, + 787, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "D. Kim et al.", + "bbox": [ + 271, + 114, + 362, + 127 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "models to bypass the need for offline object proposal generation, which could be cumbersome for large-scale image-text data.", + "bbox": [ + 212, + 146, + 782, + 176 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In addition, we focus on the Vision Transformer (ViT) based CLIP models. These CLIP ViTs excel in zero-shot recognition, but are less commonly integrated into open-vocabulary detectors compared to ConvNets [18, 36, 44]. Pretraining ViT backbones on lower resolutions and then adapting them to higher-resolution detection images poses challenges, including increased computational demands and potential disruption of locality in pretrained features. Although the windowed attention technique of ViTDet [17, 20] reduces computation and preserves locality, it introduces bias from fixed window patterns. To address this, we propose Shifted-Window Learning (SWL) approach to enhance information mixing across fixed windows and mitigate grid pattern bias. Unlike the Swin Transformer [23], which applies shifted window layer by layer, SWL applies shifted windowed attention as a separate forward pass using the same ViT backbone. This simple strategy enhances the windowed attention representation when applied low-res trained ViT backbone to high-res images, ensuring compatibility with vanilla ViT backbones pretrained without shifted windows. SWL also better preserves the open-vocabulary knowledge of pretrained features compared to full-attention ViT perhaps due to its emphasis on local cues.", + "bbox": [ + 212, + 176, + 785, + 434 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Incorporating region-centric pretraining and shifted-window learning, our approach DITO (Detection-aligned Image-Text pretraining for Open-vocabulary detection) serves to narrow the gap between image-text pretraining and open-vocabulary detection, improving generalization. The best DITO model achieves 37.6 mask $\\mathrm{AP}_r$ on LVIS benchmark, surpassing the previous best by $+3.7\\mathrm{AP}_r$ . It achieves the state-of-the-art 40.5 mask $\\mathrm{AP}_r$ when pretrained on the DataComp1B dataset. On the COCO benchmark, DITO achieves a very competitive 39.6 novel AP without relying on pseudo-labels or joint training.", + "bbox": [ + 212, + 434, + 787, + 555 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 214, + 564, + 387, + 580 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Open-vocabulary detection. Conventional closed-set object detectors perform well but are limited by vocabulary size. Open-vocabulary detection, inspired by the strong zero-shot capabilities of Vision-Language Models (VLMs) like CLIP [28], has made notable progress. Various techniques have been explored, including knowledge distillation [10], prompt optimization [5], and the use of pretrained CLIP backbones with new detection heads [17,18,26,36]. Many top-performing methods [6,8,19,25,44] rely on pseudo labeling to prevent catastrophic forgetting during detection finetuning, but these methods often require multi-stage training [35,44] or extra steps to generate high-quality labels [25]. In contrast, our approach focuses on both upstream image-text pretraining and downstream open-vocabulary detection without using pseudo labeling, leading to significant improvements.", + "bbox": [ + 212, + 594, + 787, + 776 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Region-aligned Vision-Language Models. Building on advances in aligning image-text representations [28, 31], several studies have integrated region-level alignment into CLIP pretraining. For example, RegionCLIP [44] learns region-word alignment through pseudo region-text pairs but relies on an off-", + "bbox": [ + 212, + 779, + 785, + 840 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "Region-centric Image-Language Pretraining for Open-Vocabulary Detection", + "bbox": [ + 228, + 114, + 730, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "the-shelf detector for pseudo labeling and focuses only on training the image backbone. GLIP [19, 42], Grounding DINO [22], DetCLIP [38, 39], CoDet [24] integrate detector architectures into CLIP training to align regions with words, but they require additional visual grounding annotations, leading to complex multitask learning. To our knowledge, our approach is the first to incorporate detector modules in image-text pretraining without relying on box annotations.", + "bbox": [ + 212, + 146, + 787, + 238 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Vision Transformers in open-vocabulary detection. While ViT-based CLIP models excel in zero-shot recognition, their adaptation to open-vocabulary detection is less explored compared to ConvNet-based CLIP models. For instance, OWL-ViT [25, 26] finetunes ViT on higher-resolution detection images with global attentions, but the full attention is computationally intensive. ROViT [17] introduces region-aware positional embeddings to aid in detection finetuning and uses windowed attention from ViTDet [20]. Our paper proposes shifted-window learning to enhance information mixing across fixed windows and reduce window-induced bias while preserving the locality structure of pretrained features, thereby improving open-vocabulary detection.", + "bbox": [ + 212, + 243, + 789, + 396 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Self-supervised pretraining for visual tasks. Self-supervised learning has emerged as a promising paradigm to learn object features for complex visual tasks such as detection, given the challenge of scaling up human annotations. Most relevant direction is contrastive learning, where the contrastive samples can take the forms of augmented images [3], sliding windows [37], object proposals [34], or point samples [1]. Some alternative strategies like pseudo-labeling [45] and pixel reconstruction [11] have also proven effective. While the majority of these methods have focused on learning from images without textual context, and applying to closed-vocabulary detection, DITO leverages large image-text data to tackle the more demanding open-vocabulary detection task, without a need for offline proposal generation [34, 45].", + "bbox": [ + 212, + 401, + 789, + 569 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3 Method", + "text_level": 1, + "bbox": [ + 215, + 599, + 331, + 616 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We address the problem of open-vocabulary object detection. At training time, the model can access the class and box labels of base categories $(C_B)$ . At test time, the model is tasked with detecting objects from a set of novel categories $(C_N)$ not present in the training set. To achieve this, we leverage pretrained vision and language models (VLMs) as in prior studies [10, 17, 18, 44].", + "bbox": [ + 212, + 640, + 787, + 717 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "More specifically, we leverage recent advances in ViT-based detectors [17,20] for their promising results. However, instead of solely taking the pretrained ViT backbone, we demonstrate how to enhance the VLMs by integrating detector heads into the CLIP pretraining process, referred to as Region-centric image-text Pretraining (REP). Additionally, for detection finetuning, we propose a Shifted-Window Learning (SWL) to enhance the adaptation of the pretrained ViT to open-vocabulary detection task. By combining these approaches, DITO achieves significant improvements in open-vocabulary detection over prior arts.", + "bbox": [ + 212, + 719, + 787, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "D. Kim et al.", + "bbox": [ + 271, + 114, + 362, + 127 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/76ad26ff24ea4bd2aaa44277f85c2e510eecedb0d4d4f4f0427b90ddc48d92b8.jpg", + "image_caption": [ + "Fig. 2: DITO method. Region-centric image-text pretraining (left): We train the detector heads (e.g. FPN [20, 21], Faster RCNN head [30], and RPN [30]) upon a ViT encoder backbone with multi-level image-text contrastive loss to bridge the gap between image-text pretraining and open-vocabulary detection. Shifting-window learning for detection (right): We roll the image and combines the shifted features with the original features to mitigate the bias of windowed attention grid [20], and produce more robust semantic representation." + ], + "image_footnote": [], + "bbox": [ + 222, + 145, + 607, + 349 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/f852f89776229d8b0874d93a5ec7b29fa46ed9f6108793b84d76a61a739017e4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 617, + 152, + 769, + 347 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1 Preliminaries", + "text_level": 1, + "bbox": [ + 214, + 467, + 372, + 481 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Baseline. Our baseline approach is RO-ViT [17], a state-of-the-art ViT-based method for open-vocabulary detection. RO-ViT introduces CLIP pretraining with a new positional embedding scheme called cropped positional embedding (CPE). CPE involves randomly cropping and resizing the standard whole-image positional embedding during pretraining to enhance the ViT's generalization onto region-level recognition task and higher resolution detection inputs downstream. For detection finetuning, it adopts ViTDet [20] architecture, initialized with the pretrained ViT backbone. In the following, we describe the image-text pretraining and the downstream open-vocabulary detection.", + "bbox": [ + 212, + 487, + 787, + 625 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Image-text pretraining. We adopt dual-encoder CLIP pretraining widely used in existing works [28, 31]. The image embeddings $\\{v\\}$ and text embeddings $\\{l\\}$ are the average-pooled outputs from the image and text encoders, respectively. As in previous works, we compute the dot product of the embeddings in batch $B$ , and scale it by a learnable temperature $\\tau$ before applying the InfoNCE loss [27, 28]. Mathematically, the image-to-text (I2T) loss can be expressed as:", + "bbox": [ + 214, + 627, + 787, + 719 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL _ {\\mathrm {I 2 T}} = - \\frac {1}{B} \\sum_ {i = 1} ^ {B} \\log \\left(\\frac {\\exp \\left(v _ {i} l _ {i} / \\tau\\right)}{\\sum_ {j = 1} ^ {B} \\exp \\left(v _ {i} l _ {j} / \\tau\\right)}\\right). \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 357, + 728, + 785, + 771 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The text-to-image (T2I) loss is symmetrical by exchanging the inner/outer summation loops. The total contrastive loss $L_{con}$ is obtained by $L_{con} = (L_{\\mathrm{I2T}} + L_{\\mathrm{T2I}}) / 2$ . As mentioned above, we adopt the cropped positional embeddings (CPE) following [17].", + "bbox": [ + 214, + 779, + 787, + 842 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "Region-centric Image-Language Pretraining for Open-Vocabulary Detection", + "bbox": [ + 228, + 114, + 730, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 782, + 126 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Open-vocabulary detection finetuning. At the fine-tuning stage, our detection finetuning recipe follows previous studies [10, 17, 18, 41]. During the training phase, we use the RoI-Align [12] feature as the detection embedding for each detected region. We replace the fixed-size classifier layer with the text embeddings of base categories $(C_B)$ . The detection score $p_i$ is determined by calculating the cosine similarity between the region embedding $r_i$ and text embeddings of base categories $(C_B)$ followed by a softmax operation. We prepend an additional background class embedding to $C_B$ and use the term \"background\" to represent the background category. Any proposals that do not match to any base category annotations are treated as background during training. It is important that the text embeddings are computed from the same text encoder as from the image-text pretraining. During testing, we expand the text embeddings to include the novel categories $(C_B \\cup C_N)$ , resulting in $(C_B \\cup C_N + 1)$ categories including the background. We calculate the detection scores $(p_i)$ as the cosine similarity between the region embeddings $(r_i)$ and the expanded text embeddings. Apart from the detection embedding $(r_i)$ , we extract the VLM embedding [18] of region $i$ by RoI-Align at the last ViT backbone feature map. The VLM score $(z_i)$ is calculated as the cosine similarity with the text embeddings of the combined categories $(C_B \\cup C_N)$ .", + "bbox": [ + 212, + 146, + 787, + 431 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To compute the open-vocabulary detection score $(s_i^{\\mathrm{ens}})$ , we ensemble the detection and VLM scores by geometric means [10,18]. The formula is as follows:", + "bbox": [ + 212, + 433, + 787, + 464 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\ns _ {i} ^ {\\text {e n s}} = \\left\\{ \\begin{array}{l l} z _ {i} ^ {(1 - \\alpha)} \\cdot p _ {i} ^ {\\alpha} & \\text {i f} i \\in C _ {B} \\\\ z _ {i} ^ {(1 - \\beta)} \\cdot p _ {i} ^ {\\beta} & \\text {i f} i \\in C _ {N} \\end{array} \\right. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 385, + 474, + 785, + 516 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Here, $\\alpha, \\beta$ are floats $\\in [0,1]$ that control the weighting of base versus novel categories. The background score comes from the detection score $(p_i)$ alone, because we observe the VLM score of \"background\" class is often less reliable.", + "bbox": [ + 212, + 523, + 787, + 569 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.2 Region-centric Image-Text Pretraining", + "text_level": 1, + "bbox": [ + 214, + 585, + 581, + 602 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Standard image-text pretraining uses classification architectures (e.g. ViT backbone followed by global pooling) as the language supervision occurs at the image level rather than the region level. Subsequently, for downstream detection, new detection heads are introduced and trained from scratch on a limited set of detection categories [16-18]. To fully utilize the knowledge embedded in large-scale image-text data, we propose Region-centric Pretraining (REP) which integrates the detector modules during the CLIP pretraining phase. Specifically, given access to image-text paired data but lacking box labels, our pretraining focuses on the region-recognition pathway of a detector, encompassing components like the backbone, FPN [20,21], RoI-Align [12], RPN-objectness [30], and Faster RCNN-classifier [30]. Consequently, the detector heads can be warm-started from the knowledge of large image-text data, thereby improving the generalization capability. To our knowledge, we are the first to integrate detector modules in image-text pretraining without box labels, and our experiments demonstrate clear benefits of our approach in open-vocabulary detection.", + "bbox": [ + 212, + 613, + 787, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "D. Kim et al.", + "bbox": [ + 271, + 114, + 362, + 127 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Detector head learning from random regions. Fig. 2 (left) illustrates our region-centric pretraining system. Following existing works [16, 17], we adopt SimpleFPN [20] and Faster R-CNN [30] models to remove the detector differences and study the benefits of our region-centric pretraining. Specifically, the multilevel feature pyramid is computed from the ViT backbone. Then, RoI-Align [12] and Faster R-CNN classifier head are applied to these feature maps to match the classification pathway in pretraining with the region-recognition pathway in detection finetuning (see Table 4b for ablations).", + "bbox": [ + 212, + 146, + 782, + 263 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "For each level $i$ of the feature pyramid, we randomly generate $n_i$ box regions uniformly over the image by sampling the box size $h, w \\sim \\text{Uniform}(0.2, 0.5)$ and aspect ratio $h / w \\sim \\text{Uniform}(0.5, 2.0)$ . The $n_i$ value is set proportional to the size of the $i$ -th feature map so that larger feature map would be covered by more regions. We extract the RoI-features of each region by RoI-Align, and feed them through the region classifier head [30] to obtain the RoI embeddings.", + "bbox": [ + 212, + 267, + 784, + 357 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Multi-level image-text supervision. After computing the RoI embeddings across pyramid levels, we perform a max pooling over the RoI embeddings per-level to obtain an image embedding for each pyramid level. Intuitively, max pooling allows the representation to focus on salient regions and discriminative features, thereby learning region-level information without explicit supervision. Then we apply the standard image-text contrastive loss (see Eqn. (1)) on each feature level separately, which aids the learning of rich semantic and object knowledge within every feature map (see Table 4c for ablations). The losses from all levels are weighted equally and summed together. Without explicit region-level supervision, the max pooling over regions encourages the more salient, text-aligned region features to contribute more to the whole image representation in the contrastive loss. Our experiments show emergent region-text alignments from the multi-level training, where the feature maps possess more localized semantic information compared to the baseline CLIP [17] backbone (see Fig. 3).", + "bbox": [ + 212, + 361, + 782, + 571 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Region proposal network learning. Fig. 3 shows that the learnt multi-level representations exhibit localized semantics well-aligned with the text query, i.e. the salient regions have higher similarity with the text relative to the background. Motivated by this, we employ the multi-level visual-text similarity as a supervisory signal for training the Region Proposal Network (RPN) [30]. Specifically, we compute the cosine similarity between each RoI embedding and the text, and use it as the target RPN score for the center location of each RoI. Any negative dot product value is mapped to zero to keep the target score in range [0,1]. We use L1 regression loss and set the loss weight equal to the multi-level contrastive loss (see supplementary). The losses are propagated only through the RoI centers and other pixels are ignored. Note that the box regression of the RPN is not trained here but learnt later through the detection finetuning, as we only use the image-text paired data without any box annotations.", + "bbox": [ + 212, + 577, + 782, + 771 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.3 Shifted-Window Learning for Detection", + "text_level": 1, + "bbox": [ + 214, + 790, + 586, + 804 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The CLIP ViT backbones are typically pretrained on lower resolutions (e.g. $224 \\times 224$ ) and later adapted to higher resolution detection images (e.g. $1024 \\times 1024$ ).", + "bbox": [ + 212, + 809, + 800, + 839 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "Region-centric Image-Language Pretraining for Open-Vocabulary Detection", + "bbox": [ + 228, + 114, + 730, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "code", + "sub_type": "algorithm", + "code_caption": [ + "Algorithm 1 Shifted Window backbone" + ], + "code_body": "x: image patch tokens + positional embeddings. M: attention window size. s = M // 2 $y =$ forward_vit_with_win_attn(x, M) $x^{\\prime} =$ np-roll(x, shift=[s, s], axis=[0, 1]) # shift $y^\\prime =$ forward_vit_with_win_attn $(x^{\\prime},M)$ $y'' =$ np-roll(y', shift=[-s, -s], axis=[0, 1]) # unshift return $(y + y'') / 2$ # SWL backbone output", + "bbox": [ + 215, + 164, + 784, + 258 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "While detection tasks benefit from global information, directly applying the pretrained ViT to high-resolution inputs is computationally intensive and can disrupt the locality structure of the pretrained lower-res features. Windowed attention in ViTDet [20] helps reduce computation but introduces bias from the fixed-size grid pattern. To address this, we propose the Shifted-Window Learning (SWL) approach, which improves information mixing across fixed windows and mitigates grid pattern bias.", + "bbox": [ + 212, + 299, + 787, + 407 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Network architecture. Fig. 2 (right) and Algorithm 1 describe the SWL algorithm. The standard ViT consists of a patchifying layer and a set of transformer layers. After feeding the image through the patchifying layer, we obtain a feature map $x$ of shape $(h, w, d)$ . This feature map $x$ is fed through the rest of the ViT with windowed attention layers on a grid $K \\times K$ , and $L$ global attention layers evenly spaced throughout the ViT (where $L = 4$ ) following [20], resulting in output $y$ . In parallel, we create another copy of $x$ , which is rolled along both $h$ and $w$ axes by $s$ pixels. The elements that roll beyond the last position are reintroduced from the first. We carefully design the attention masks such that the rolled around patches would not attend to the patches on the other side of the image (see the right figure of Algorithm 1). The shift size $s$ is set as the half of the attention window size $M$ (i.e. $s = M // 2$ ). Empirically, the window size $M = 16$ equals the image size (e.g. 1024) divided by the product of patch size (e.g. $P = 16$ ) and the grid size (e.g. $K = 4$ ). The shifted feature map $x'$ is then processed through the rest of the ViT in the same manner, resulting in output $y'$ in the same shape $(h, w, d)$ as $y$ . We then unshift $y'$ and combine it with $y$ by averaging. We apply the above shifted window operations in detection finetuning and inference times.", + "bbox": [ + 212, + 414, + 789, + 688 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Comparison with Swin Transformer. Compared to the Swin Transformer [23], we apply the shifted-window ideas as separate forward passes, while Swin Transformer applies similar ideas in an alternating manner layer by layer. Our approach requires no change to the vanilla transformer architecture and is compatible with any ViT backbones pretrained without shifted windows (e.g. [28]), whereas Swin Transformer requires specialized pretraining on the same architecture. Compared to the full-attention ViT [4], we observe that shifted-window ViT taps more effectively into the semantic knowledge of pretrained backbone than full global attention, perhaps because the window helps the model focus more on local cues and ignore the noises farther away.", + "bbox": [ + 212, + 696, + 808, + 849 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "D. Kim et al.", + "bbox": [ + 271, + 114, + 362, + 127 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "3.4 Distillation from Frozen ViT Backbone", + "text_level": 1, + "bbox": [ + 215, + 146, + 583, + 160 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "While the ViT backbone adapts to the detection tasks, it may lose some of its pretrained open-vocabulary knowledge. Therefore, we propose a simple distillation approach which uses a separate frozen ViT to teach the finetuned ViT backbone during the detection finetuning. We use a cosine distance loss that aligns the RoI-Align embeddings extracted from the feature maps of both backbones. The cosine distance is computed for each RoI then averaged over all RoIs. Empirically, we find it advantageous to add a $1 \\times 1$ Conv projection layer to the finetuned ViT backbone before the RoI-Align, allowing some flexibility in distillation when jointly trained with other detection objectives. The auxiliary distillation loss is then added to other detection losses, with a loss weight $\\gamma = 1$ (see supplementary). At inference, the ViT backbone features after the projection are used to compute the region VLM score $z_{i}$ (Sec. 3.1). It is worth noting that the frozen ViT backbone is only used during training for distillation purposes and is removed at inference.", + "bbox": [ + 212, + 167, + 787, + 378 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "While previous studies [2, 10] have utilized knowledge distillation from the CLIP models, their teacher CLIP models mostly operate on image crops in an offline process, thus needing multiple forward passes through the backbone for all RoIs. In contrast, our distillation operates efficiently on RoIs cropped from dense feature maps in a single forward pass during the detection finetuning with minimal overhead.", + "bbox": [ + 212, + 378, + 787, + 469 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4 Experimental Results", + "text_level": 1, + "bbox": [ + 215, + 484, + 460, + 502 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Baseline reproduction. As discussed in Sec. 3.1, our baseline method is RO-ViT [17], a leading ViT-based approach for open-vocabulary detection. RO-ViT introduces cropped positional embedding (CPE) in CLIP pretraining to enhance the generalization of the pretrained ViT onto the downstream detection. Additionally, it adopts global average pooling of the ViT features instead of CLs-token pooling, which simplifies the adaptation onto higher resolution inputs, and the extraction of region features (e.g. RoI-Align) at the final layer. We reproduce the CLIP pretraining of RO-ViT [17] using the widely adopted LAION2B dataset [31], as the OpenAI CLIP's WIT [28] and ALIGN [14] datasets are not publicly available. We trained the CLIP models from scratch, following the same pretraining protocol and hyperparameters, including 500k iterations and 16k batch size (i.e. 8B samples seen in training), $224 \\times 224$ image size, global average pooling, and cropped positional embedding. We use the standard InfoNCE loss (Eqn. (1)) instead of their focal constrastive loss [17]. The following compares our reproduced CLIP with OpenAI CLIP [28] and OpenCLIP [31]:", + "bbox": [ + 212, + 516, + 787, + 743 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/6db5dad00c99d7b10139b48db6a5ad8cd2240f4ec7962841b0c494160f3c887b.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
methoddataset#samples seenbackboneImageNet Top-1 Acc.
CLIP [28]WIT-400M13BViT-L/1475.5
OpenCLIP [31]LAION-2B32BViT-L/1475.2
RO-ViT CLIP [17] our repro.LAION-2B8BViT-L/1673.9
", + "bbox": [ + 218, + 751, + 782, + 801 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "On the zero-shot ImageNet classification benchmark, the LAION-2B pretraining matches or slightly underperforms the OpenAI CLIP with WIT-400M.", + "bbox": [ + 215, + 809, + 785, + 839 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "Region-centric Image-Language Pretraining for Open-Vocabulary Detection", + "bbox": [ + 228, + 114, + 730, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/0ff3bd1951075c3d9d838938e25e2cb244493d5fcdcbbf99f83d150a0d2c92bd.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
methodpretraining modelpretraining datadetector backbonew/ pseudo box labelsmask APrmask AP
ConvNet based:
OV-DETR [40]ViT-B/32CLIP-400MR-50-17.426.6
Kaul et al. [15]ViT-B/32CLIP-400MR-50-19.330.6
DetPro-Cascade [5]ViT-B/32CLIP-400MR-50-20.027.0
Rasheed [29]ViT-B/32CLIP-400MR-50-21.125.9
BARON [35]ViT-B/32CLIP-400MR-50-22.627.6
CoDet [24]R-50CLIP-400M + CC3MR-50-23.430.7
EdaDet [32]R-50CLIP-400MR-50-23.727.5
VL-PLM [43]ViT-B/32CLIP-400MR-50✓17.227.0
PromptDet [6]ViT-B/32CLIP-400MR-50✓21.425.3
OADB [33]ViT-B/32CLIP-400MR-50✓21.726.6
RegionCLIP [44]R-50x4CLIP-400M + CC3MR-50x4✓22.032.3
CORA [36]R-50x4CLIP-400MR-50x4✓22.2bx-
Detric-CN2 [46]ViT-B/32CLIP-400M + INet-21KR-50WS24.632.4
ViLD-Ens [10]EffNet-B7ALIGN-1.8BEffNet-B7-26.329.3
F-VLM [18]R-50x64CLIP-400MR-50x64-32.834.9
ViT based:
OWL-ViT [26]O+VGViT-L/14CLIP-400MViT-L/14-25.6bx34.7bx
OWL-ViT v2 [25]O+VGViT-L/14WebLI-10BViT-L/14✓45.9bx50.4bx
RO-ViT [17]ViT-B/16ALIGN-1.8BViT-B/16-28.030.2
RO-ViT [17]†ViT-L/16LAION-2BViT-L/16-32.432.9
CFM-ViT [16]ViT-B/16ALIGN-1.8BViT-B/16-28.832.0
CFM-ViT [16]ViT-L/16ALIGN-1.8BViT-L/16-33.936.6
CFM-ViT [16]*ViT-L/16LAION-2BViT-L/16-33.836.4
DITO (ours)ViT-S/16LAION-2BViT-S/16-26.228.8
DITO (ours)ViT-B/16LAION-2BViT-B/16-31.532.4
DITO (ours)ViT-L/16LAION-2BViT-L/16-37.636.2
DITO (ours)ViT-L/16DataComp-1BViT-L/16-40.538.0
", + "bbox": [ + 218, + 143, + 784, + 450 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Table 1: LVIS open-vocabulary detection (mask AP). DITO outperforms the best existing approach by $+3.7$ mask $\\mathrm{AP}_r$ . WS: uses weak supervision from ImageNet-21K. †: reports LAION-2B results in arXiv version. *: our reproduced result using LAION-2B. O+VG: uses extra Objects365 and Visual Genome data.", + "bbox": [ + 212, + 454, + 787, + 511 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Pretraining setup. After the above-mentioned baseline CLIP training, we apply our region-centric pretraining (REP) where we freeze the image and text encoders trained in the first phase and introduce the detector heads. We use the Simple FPN [20], and classification layers of Faster R-CNN and RPN [30], where we replace the batch normalization with layer normalization. At the $i$ -th pyramid level $i \\in \\{2, 3, 4, 5\\}$ , we randomly sample $n_i \\in \\{400, 200, 100, 50\\}$ box regions and compute their RoI-Align features. We use a short training cycle of $30k$ iterations, $4k$ batch size, $256 \\times 256$ image size, AdamW optimizer with an initial learning rate of 1e-4 with linear decay. For both phases of image-text pretraining, we use the publicly available LAION-2B [31] dataset.", + "bbox": [ + 212, + 532, + 787, + 685 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Detection finetuning setup. As noted in Sec. 3.2, we adopt the ViTDet [20] with SimpleFPN as our detector following the baseline works [16,17]. We follow the same finetuning settings of [17]. Specifically, we train the detector with image size $1024 \\times 1024$ and use windowed attention in the backbone with grid size $4 \\times 4$ . The learning rate for the backbone is set lower as $0.6 \\times$ of the detector head layers. We use $\\alpha = 0.3$ , $\\beta = 0.65$ for score combination in Eqn. (2). The text embedding of each category is calculated as the average over the CLIP prompt templates. We use the batch size 128, the SGD optimizer with momentum 0.9. The initial learning rate and iterations are set to 0.18 and $36.8\\mathrm{k}$ for LVIS, and 0.02 and $11.3\\mathrm{k}$ for COCO datasets.", + "bbox": [ + 212, + 688, + 787, + 839 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "D. Kim et al.", + "bbox": [ + 271, + 114, + 361, + 126 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/0ddbbbbf2fc186fe14ce5112e2ec3ef78dbae512aec39cefe750b8b7f60b0d22.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
methodpretraining modelpretraining datadetector backbonew/ pseudo box labelsnovel APAP
ConvNet based:
OVR-CNN [41]R-50CLIP-400M + COCO-CapR-50-22.839.9
ViLD [10]ViT-B/32CLIP-400MR-50-27.651.3
F-VLM [18]R-50CLIP-400MR-50-28.039.6
OV-DETR [40]ViT-B/32CLIP-400MR-50-29.452.7
CoDet [24]R-50CLIP-400M + COCO-CapR-50-30.646.6
PromptDet [6]ViT-B/32CLIP-400MR-50✓26.650.6
XPM [13]R-50CLIP-400MR-50✓27.041.2
OADB [33]ViT-B/32CLIP-400M + COCO-CapR-50✓30.047.2
VL-PLM [43]ViT-B/32CLIP-400MR-50✓34.453.5
RegionCLIP [44]R-50x4CLIP-400M + CC3M + COCO-CapR-50x4✓39.355.7
EdaDet [32]R-50CLIP-400MR-50✓40.252.5
CORA [36]R-50x4CLIP-400MR-50x4✓41.743.8
Detic-CN2 [46]ViT-B/32CLIP-400M + INet-21KR-50WS27.845.0
ViT based:
RO-ViT [17]ViT-B/16ALIGN-1.8BViT-B/16-30.241.5
RO-ViT [17]ViT-L/16ALIGN-1.8BViT-L/16-33.047.7
RO-ViT [17]*ViT-L/16LAION-2BViT-L/16-33.347.9
CFM-ViT [16]ViT-B/16ALIGN-1.8BViT-B/16-30.842.4
CFM-ViT [16]ViT-L/16ALIGN-1.8BViT-L/16-34.146.0
CFM-ViT [16]*ViT-L/16LAION-2BViT-L/16-34.346.4
DITO (ours)ViT-S/16LAION-2BViT-S/16-32.344.4
DITO (ours)ViT-B/16LAION-2BViT-B/16-36.648.8
DITO (ours)ViT-L/16LAION-2BViT-L/16-39.654.4
DITO (ours)ViT-L/16DataComp-1BViT-L/16-40.254.6
", + "bbox": [ + 215, + 143, + 785, + 428 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/c4e8c95bf2692509d5e782149ae9f6e6c43f70224e596b97b7a4158de9ade568.jpg", + "table_caption": [ + "Table 2: COCO open-vocabulary detection (box AP50). DITO demonstrates a very competitive novel category AP without using pseudo labeling or weak supervision (WS). *: our reproduced result using LAION-2B." + ], + "table_footnote": [], + "table_body": "
methodbackboneAP\\( AP_{50} \\)\\( AP_{75} \\)
Supervised [10]R-5025.638.628.0
ViLD [10]R-5011.818.212.6
DetPro [5]R-5012.118.812.9
BARON [35]R-5013.621.014.5
F-VLM [18]R-50x1616.225.317.5
F-VLM [18]R-50x6417.727.419.1
RO-ViT [17]ViT-L/1617.126.918.5
CFM-ViT [16]ViT-L/1618.728.920.3
DITO (ours)ViT-L/1620.031.821.5
", + "bbox": [ + 315, + 479, + 679, + 594 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Table 3: Zero-shot transfer detection from LVISbase to Objects365 (box AP). All models are tested on Objects365 dataset following the setup of [10].", + "bbox": [ + 215, + 595, + 784, + 626 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4.1 Main Results", + "text_level": 1, + "bbox": [ + 215, + 647, + 372, + 662 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "LVIS Benchmark. In Table 1, we report the comparison with existing methods on the challenging LVIS benchmark. The 'frequent' and 'common' classes of the dataset belong to the base categories $\\mathrm{C}_B$ , and the 'rare' classes are the novel categories $\\mathrm{C}_N$ . The primary metric is the mask AP on rare classes (mask $\\mathrm{AP}_r$ ). The DITO model achieves the performance of 37.6 mask $\\mathrm{AP}_r$ , which significantly outperforms the state-of-the-art approach RO-ViT [17] with the same ViT-L backbone by +5.1 points using the same pretraining data LAION-2B [31]. We also outperform the state-of-the-art CFM-ViT [16] by +3.6 points. Our best performance sets a new state-of-the-art 40.9 mask $\\mathrm{AP}_r$ when using DataComp1B [7] in pretraining. With the ViT-B backbone, DITO maintains a healthy margin of around +2.5 $\\mathrm{AP}_r$ above existing ViT-B based approaches.", + "bbox": [ + 212, + 672, + 787, + 840 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "Region-centric Image-Language Pretraining for Open-Vocabulary Detection", + "bbox": [ + 228, + 114, + 730, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 116, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/97a291560d6c641a960d42662b803a288f17bcf2a1a4bcc0e761864a47d8a49a.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
methodAPrAP
RO-ViT [17]32.432.9
RO-ViT our repro.32.233.0
w/REP34.8 (+2.6)34.9
w/SWL35.0 (+2.8)35.2
w/REP+SWL36.3 (+4.1)35.8
w/REP+SWL+FD37.6 (+5.4)36.2
", + "bbox": [ + 228, + 143, + 447, + 220 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/d71b46da26f3a0d11794c4159f0040e8b007243a84690db33b0789026d19645a.jpg", + "table_caption": [ + "(a) DITO framework." + ], + "table_footnote": [], + "table_body": "
pretraining methodAPr
RO-ViT our repro.32.2
w/FPN33.2
w/FPN+Head34.2
w/FPN+Head+RPN34.8
", + "bbox": [ + 460, + 143, + 620, + 200 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/ce98476c5ff5fc6f8412e92914b3252df8065936988fe8fcc48653b345a1889f.jpg", + "table_caption": [ + "(b) Detector components." + ], + "table_footnote": [], + "table_body": "
RoI embeddingAPr
global avg/lvl34.0
global max/lvl33.7
multi RoIs, avg/lvl33.8
multi RoIs, max/lvl34.8
multi RoIs, max all34.1
", + "bbox": [ + 635, + 143, + 785, + 210 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/04c52f179a58b973e4aebabf03425eeca308712554b61792c61b14df7601e4e4.jpg", + "table_caption": [ + "(c) RoI embedding.", + "Table 4: Ablation on overall DITO framework and Region-centric Pretraining (Sec. 3.2). REP: Region-centric Pretraining, SWL: Shifted-Window Learning, FD: Frozen backbone Distillation. Best setting is in gray." + ], + "table_footnote": [], + "table_body": "
backboneAPrAP
fully global attn.33.433.8
baseline window attn. [17]32.233.0
Swin [23] style31.333.1
shifted window35.035.2
", + "bbox": [ + 243, + 292, + 493, + 349 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/7fda14b5668654059d4d03137e6b8d593e44ad6e3193ad047450f17bf93e9666.jpg", + "table_caption": [ + "(a) Shifted-window learning." + ], + "table_footnote": [], + "table_body": "
# global attn. layerbasew/ SWL
030.734.6 (+3.9)
432.235.0 (+2.8)
1232.434.2 (+1.8)
24 (all layers)33.433.4 (+0.0)
", + "bbox": [ + 516, + 292, + 772, + 349 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "(b) SWL w.r.t. # global attention layers.", + "bbox": [ + 504, + 351, + 772, + 363 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Table 5: Ablation on Shifted-Window Learning (SWL - Sec. 3.3).", + "bbox": [ + 250, + 371, + 750, + 385 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "COCO Benchmark. We present the comparison on COCO benchmark in Table 2. The main metric is AP50 of novel categories (novel AP). Without using pseudo labeling [6,13,43,44], weak supervision [46], or externally trained detector modules [29, 35], our model demonstrates competitive results of 39.6 novel AP with LAION-2B and 40.2 with DataComp-1B. Among the ViT-based methods, DITO outperforms recent works RO-ViT [17] and CFM-ViT [16] by a clear margin of $+6.3$ and $+5.3$ points, respectively.", + "bbox": [ + 212, + 401, + 784, + 507 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Transfer detection. We further evaluate DITO in the transfer detection setting, where the open-vocabulary detector trained on one dataset $(\\mathrm{LVIS}_{base})$ is tested on another dataset (Objects365) without any finetuning. By simply replacing the text embeddings, Table 3 shows that DITO achieves 20.0 AP, outperforming previous methods using ConvNet or ViT backbones of similar size.", + "bbox": [ + 212, + 511, + 784, + 585 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "4.2 Ablation Studies", + "text_level": 1, + "bbox": [ + 214, + 604, + 400, + 617 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "For ablation study, we use the ViT-L/16 model pretrained with LAION-2B, and evaluate on the LVIS benchmark and report mask $\\mathrm{AP}_r$ .", + "bbox": [ + 212, + 621, + 784, + 651 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "DITO overall framework. Table 4a summarizes the benefits of each DITO components. The region-centric pretraining (REP) improves the contrastive model baseline by $+2.6\\mathrm{AP}_r$ and shifted window learning (SWL) improves the baseline by $+2.8$ points. Combining both strategies brings a significant gain of $+4.1\\mathrm{AP}_r$ . Lastly, we observe that incorporating the frozen backbone distillation leads to an additional boost of $+1.2$ points. In the following, we provide ablations for each components.", + "bbox": [ + 212, + 654, + 792, + 760 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Region-centric Pretraining. In Table 4b, we ablate our region-centric image-text pretraining by progressively adding the FPN, Faster R-CNN head, and RPN into the pretraining. Our 'baseline' is the contrastive image-text pretraining with cropped positional embedding [17]. On top of this, 'w/ FPN' introduces the FPN into the pretraining, where each pyramid level (whole image) map is mean-pooled", + "bbox": [ + 212, + 763, + 785, + 839 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "D. Kim et al.", + "bbox": [ + 271, + 114, + 361, + 126 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/0eef19f6046d3530c0ccd16ae6b852078f672903c4e08bfa6a6a8012182b13ab.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
backboneAPAPrAPcAPf
GT boxes given (region classification):
before det. training48.756.851.542.1
after det. training53.754.253.753.5
after det. training w/FD54.857.754.353.2
", + "bbox": [ + 233, + 143, + 514, + 200 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/76e41f5397de3b2a2041f9ff38eb3a918e4f1ff57df9a8c3ea8947e7c1b5ae66.jpg", + "table_caption": [ + "(a) GT region recognition." + ], + "table_footnote": [], + "table_body": "
methodAPrAP
RO-ViT our repro.32.233.0
RO-ViT our repro.+FD34.0 (+1.8)33.6
w/ REP+SWL36.335.8
w/ REP+SWL+FD37.6 (+1.3)36.2
", + "bbox": [ + 534, + 143, + 782, + 200 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "(b) Effect of FD on OVD.", + "bbox": [ + 568, + 202, + 736, + 214 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Table 6: Ablation on frozen backbone distillation (FD - Sec. 3.4).", + "bbox": [ + 250, + 222, + 750, + 237 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "into an image embedding, followed by the image-text contrastive loss per level. It improves the baseline by $+1.0\\mathrm{AP}_r$ . Adding the Faster R-CNN head further improves the alignment between the pretraining and detection finetuning, showing a gain of $+2.0\\mathrm{AP}_r$ . Incorporating all components i.e. the FPN, Faster R-CNN and RPN heads achieves the best 34.8 $\\mathrm{AP}_r$ , a significant gain of +2.6 points over the baseline.", + "bbox": [ + 212, + 257, + 784, + 347 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Table 4c shows that both global avg- and max-pooling are sub-optimal due to the lack of saliency map (avg-pool), and the limited capacity of a single pixel to represent semantic concepts for contrastive learning (max-pool), respectively. Our approach combines the best of both worlds, by first avg-pooling within each RoI, and then max-pooling over these RoI embeddings. Each embedding represents a proper RoI and the global representation captures the saliency map through max-pooling. Pooling per pyramid level (i.e. multi-level image-text supervision) outperforms pooling over all levels.", + "bbox": [ + 212, + 348, + 784, + 469 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/b96117cfb9376fb716d03fff8f0e0ca5da34dadf5fd46874d5c062fe64de13ea.jpg", + "image_caption": [ + "\"am riding a cart pulled by a donkey" + ], + "image_footnote": [], + "bbox": [ + 217, + 478, + 403, + 527 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/1e73822b46fb7338b5e7b76eded1625ba323c2d7f427c289a958c2135eaeff20.jpg", + "image_caption": [ + "\"a crane operates amidst piles of rubble\"" + ], + "image_footnote": [], + "bbox": [ + 406, + 478, + 594, + 527 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/10698e6b13c6dee5cf535e0541da93d88e221089a891af5d2144a501987044eb.jpg", + "image_caption": [ + "\"aperson wearing blue rubber gloves is loading a syringe with vegetables\"" + ], + "image_footnote": [], + "bbox": [ + 596, + 478, + 784, + 527 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/441f38d48142acc4a2426621641d897860befe2659f79a5f7497c3af68af3a3b.jpg", + "image_caption": [ + "\"amotherand her two children sit on a green bench\"", + "Fig. 3: Visual-text similarity map. For each example, we show the paired image (left) and text (bottom) input, and the visual-text similarity map using the contrastive model baseline [17] backbone features (middle) or our region-centric pretraining features (right). We use Flickr30K (top row) and COCO Captions (bottom row) datasets." + ], + "image_footnote": [], + "bbox": [ + 217, + 546, + 403, + 595 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/12b3edad64336eddf3980b8fe35f84ac2e66a45fcd2d24a7f8bd8043510b9074.jpg", + "image_caption": [ + "\"Cows on the street next to people that are on the sidewalk\"" + ], + "image_footnote": [], + "bbox": [ + 406, + 546, + 594, + 595 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/1f9acc1833c0be07060bdfd5457cfb76accdc3fca7c2867d70a93a7c93765476.jpg", + "image_caption": [ + "\"small white bird sits on the back of a horse in an open field\"" + ], + "image_footnote": [], + "bbox": [ + 596, + 546, + 784, + 595 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Shifted-Window Learning for detection. The CLIP ViT backbones are initially pretrained on lower resolutions and then adapted to higher resolution detection images. In Table 5a, we assess the efficacy of the shifted-window backbone in open-vocabulary detection training. Although the fully global attention model improves the detection task compared to the baseline windowed attention model [17], directly applying the pretrained ViT on high-resolution inputs may not be optimal, potentially compromising the locality structure of the pretrained lower-resolution features. Our shifted-window learning approach achieves $35.0\\mathrm{AP}_r$ by preserving the locality structure from windowed attention as well as integrating information across fixed windows, outperforming the fully global", + "bbox": [ + 212, + 688, + 787, + 840 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "Region-centric Image-Language Pretraining for Open-Vocabulary Detection", + "bbox": [ + 228, + 114, + 730, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 784, + 126 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "attention model (33.4 $\\mathrm{AP}_r$ ) which is computationally intensive. Notably, OWL-ViT [25,26] adopts the fully global attention model. In addition, naively applying the layer-alternating shifted window as in Swin [23] leads to a performance drop (see Sec. 3.3). Table 5b delves deeper into the behavior of SWL. The advantage of SWL diminishes steadily with increasing number of global attention layers in the windowed attention backbone, validating its better information mixing enabled by the SWL.", + "bbox": [ + 212, + 146, + 787, + 252 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Frozen backbone distillation. Table 6a studies the region classification capability of the CLIP ViT backbone before and after the detection finetuning. We use the ground truth boxes and measures AP scores to evaluate the zero-shot region classification of the base ('frequent' + 'common') and novel ('rare') categories. We observe that the finetuned backbone indeed shows a notable drop of $-2.6\\mathrm{AP}_r$ on novel classes, while overfitting to the base classes. The frozen backbone distillation ('w/ FD') leads to a significant improvement of $+3.5\\mathrm{AP}_r$ even surpassing the frozen pretrained backbone (before detection finetuning), while maintaining performance on the base classes ( $\\mathrm{AP}_c$ and $\\mathrm{AP}_f$ ). Table 6b presents the open-vocabulary detection results where the frozen backbone distillation brings a clear gain of $+1.3\\sim 1.8\\mathrm{AP}_r$ . These results highlight the efficacy of our frozen backbone distillation in detection finetuning (Sec. 3.4), as it effectively preserves the pretrained open-vocabulary knowledge while acquiring explicit region-text alignment through detection supervision.", + "bbox": [ + 212, + 255, + 787, + 467 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "4.3 Visualization", + "text_level": 1, + "bbox": [ + 214, + 481, + 372, + 494 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In Fig. 3, we visualize the similarity maps between image features and query text embeddings. We compare the baseline [17] contrastive model's backbone features with our region-centric pretrained features. We select pyramid level 4 which has the same resolution as the backbone features and apply the Faster R-CNN head in a sliding window manner to obtain the dense feature map. We observe that region-centric pretraining captures more localized semantic", + "bbox": [ + 212, + 498, + 787, + 588 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Fig. 1 visualizes DITO outputs on LVIS novel categories and real-world out-of-distribution data from Ego4D [9]. Using the same DITO detector trained on $\\mathrm{LVIS}_{\\text {base }}$ , we find that DITO effectively identifies many novel and unseen objects, even under significant domain shifts.", + "bbox": [ + 212, + 589, + 787, + 650 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Conclusion", + "text_level": 1, + "bbox": [ + 214, + 665, + 361, + 681 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "We introduce DITO, a region-centric approach for open-vocabulary detection using large-scale image-text pairs. By integrating detection architecture onto the image backbone in CLIP pretraining, it learns locality-sensitive information without requiring pseudo labeling or box annotations. Furthermore, we propose a shifted-window learning method to mitigate the bias of the window attention pattern in CLIP ViT detectors. Experiments show that DITO outperforms the state-of-the-art by large margins on the LVIS benchmark, and is very competitive on the COCO benchmark and transfer detection. We hope this work would inspire the community to explore region-centric image-language pretraining for open-vocabulary localization tasks.", + "bbox": [ + 212, + 688, + 787, + 840 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "D. Kim et al.", + "bbox": [ + 271, + 114, + 361, + 127 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 215, + 143, + 323, + 160 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Bai, Y., Chen, X., Kirillov, A., Yuille, A., Berg, A.C.: Point-level region contrast for object detection pre-training. In: CVPR. pp. 16061-16070 (June 2022) 4", + "2. Chen, J., Zhu, D., Qian, G., Ghanem, B., Yan, Z., Zhu, C., Xiao, F., Elhoseiny, M., Culatana, S.C.: Exploring open-vocabulary semantic segmentation without human labels. arXiv preprint arXiv:2306.00450 (2023) 9", + "3. Chen, T., Kornblith, S., Norouzi, M., Hinton, G.: A simple framework for contrastive learning of visual representations. arXiv preprint arXiv:2002.05709 (2020) 4", + "4. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929 (2020) 8", + "5. Du, Y., Wei, F., Zhang, Z., Shi, M., Gao, Y., Li, G.: Learning to prompt for open-vocabulary object detection with vision-language model. In: CVPR (2022) 1, 3, 10, 11", + "6. Feng, C., Zhong, Y., Jie, Z., Chu, X., Ren, H., Wei, X., Xie, W., Ma, L.: Promptdet: Towards open-vocabulary detection using uncurated images. In: European Conference on Computer Vision. pp. 701-717. Springer (2022) 3, 10, 11, 12", + "7. Gadre, S.Y., Ilharco, G., Fang, A., Hayase, J., Smyrnis, G., Nguyen, T., Marten, R., Wortsman, M., Ghosh, D., Zhang, J., et al.: Datacomp: In search of the next generation of multimodal datasets. arXiv preprint arXiv:2304.14108 (2023) 11", + "8. Gao, M., Xing, C., Niebles, J.C., Li, J., Xu, R., Liu, W., Xiong, C.: Open vocabulary object detection with pseudo bounding-box labels. In: ECCV (2022) 3", + "9. Grauman, K., Westbury, A., Byrne, E., Chavis, Z., Furnari, A., Girdhar, R., Hamburger, J., Jiang, H., Liu, M., Liu, X., et al.: Ego4d: Around the world in 3,000 hours of egocentric video. In: CVPR. pp. 18995-19012 (2022) 14", + "0. Gu, X., Lin, T.Y., Kuo, W., Cui, Y.: Open-vocabulary object detection via vision and language knowledge distillation. In: ICLR (2022) 1, 3, 4, 6, 9, 10, 11", + "1. He, K., Chen, X., Xie, S., Li, Y., Dólar, P., Girshick, R.: Masked autoencoders are scalable vision learners. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16000-16009 (2022) 4", + "2. He, K., Gkioxari, G., Dollár, P., Girshick, R.: Mask r-cnn. In: ICCV (2017) 6, 7", + "3. Huynh, D., Kuen, J., Lin, Z., Gu, J., Elhamifar, E.: Open-vocabulary instance segmentation via robust cross-modal pseudo-labeling. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7020-7031 (2022) 1, 11, 12", + "4. Jia, C., Yang, Y., Xia, Y., Chen, Y.T., Parekh, Z., Pham, H., Le, Q.V., Sung, Y., Li, Z., Duerig, T.: Scaling up visual and vision-language representation learning with noisy text supervision. In: ICML (2021) 9", + "5. Kaul, P., Xie, W., Zisserman, A.: Multi-modal classifiers for open-vocabulary object detection. arXiv preprint arXiv:2306.05493 (2023) 10", + "6. Kim, D., Angelova, A., Kuo, W.: Contrastive feature masking open-vocabulary vision transformer. ICCV (2023) 2, 6, 7, 10, 11, 12", + "7. Kim, D., Angelova, A., Kuo, W.: Region-aware pretraining for open-vocabulary object detection with vision transformers. In: Conference on Computer Vision and Pattern Recognition (CVPR) (2023) 2, 3, 4, 5, 6, 7, 9, 10, 11, 12, 13, 14", + "8. Kuo, W., Cui, Y., Gu, X., Piergiovanni, A., Angelova, A.: F-vlm: Open-vocabulary object detection upon frozen vision and language models. ICLR (2023) 1, 3, 4, 6, 10, 11" + ], + "bbox": [ + 225, + 172, + 784, + 838 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "Region-centric Image-Language Pretraining for Open-Vocabulary Detection", + "bbox": [ + 228, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "19. Li, L.H., Zhang, P., Zhang, H., Yang, J., Li, C., Zhong, Y., Wang, L., Yuan, L., Zhang, L., Hwang, J.N., Chang, K.W., Gao, J.: Grounded language-image pretraining. In: CVPR (2022) 2, 3, 4", + "20. Li, Y., Mao, H., Girshick, R., He, K.: Exploring plain vision transformer backbones for object detection. In: ECCV (2022) 3, 4, 5, 6, 7, 8, 10", + "21. Lin, T.Y., Dollar, P., Girshick, R., He, K., Hariharan, B., Belongie, S.: Feature pyramid networks for object detection. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 2117-2125 (2017) 5, 6", + "22. Liu, S., Zeng, Z., Ren, T., Li, F., Zhang, H., Yang, J., Li, C., Yang, J., Su, H., Zhu, J., et al.: Grounding dino: Marrying dino with grounded pre-training for open-set object detection. arXiv preprint arXiv:2303.05499 (2023) 2, 4", + "23. Liu, Z., Lin, Y., Cao, Y., Hu, H., Wei, Y., Zhang, Z., Lin, S., Guo, B.: Swin transformer: Hierarchical vision transformer using shifted windows. CVPR (2021) 3, 8, 12, 14", + "24. Ma, C., Jiang, Y., Wen, X., Yuan, Z., Qi, X.: Codet: Co-occurrence guided region-word alignment for open-vocabulary object detection. Advances in Neural Information Processing Systems 36 (2024) 2, 4, 10, 11", + "25. Minderer, M., Gritsenko, A., Houlsby, N.: Scaling open-vocabulary object detection. Advances in Neural Information Processing Systems 36 (2023) 2, 3, 4, 10, 14", + "26. Minderer, M., Gritsenko, A., Stone, A., Neumann, M., Weissenborn, D., Dosovitskiy, A., Mahendran, A., Arnab, A., Dehghani, M., Shen, Z., Wang, X., Zhai, X., Kipf, T., Houlsby, N.: Simple open-vocabulary object detection with vision transformers. In: ECCV (2022) 3, 4, 10, 14", + "27. Oord, A.v.d., Li, Y., Vinyals, O.: Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748 (2018) 5", + "28. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., Krueger, G., Sutskever, I.: Learning transferable visual models from natural language supervision. In: ICML (2021) 1, 3, 5, 8, 9", + "29. Rasheed, H., Maaz, M., Khattak, M.U., Khan, S., Khan, F.S.: Bridging the gap between object and image-level representations for open-vocabulary detection. arXiv preprint arXiv:2207.03482 (2022) 1, 10, 12", + "30. Ren, S., He, K., Girshick, R.B., Sun, J.: Faster r-cnn: Towards real-time object detection with region proposal networks. In: NeurIPS (2015) 5, 6, 7, 10", + "31. Schuhmann, C., Vencu, R., Beaumont, R., Kaczmarczyk, R., Mullis, C., Katta, A., Coombes, T., Jitsev, J., Komatsuzaki, A.: Laion-400m: Open dataset of clip-filtered 400 million image-text pairs. arXiv preprint arXiv:2111.02114 (2021) 3, 5, 9, 10, 11", + "32. Shi, C., Yang, S.: Edadet: Open-vocabulary object detection using early dense alignment. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 15724-15734 (2023) 10, 11", + "33. Wang, L., Liu, Y., Du, P., Ding, Z., Liao, Y., Qi, Q., Chen, B., Liu, S.: Object-aware distillation pyramid for open-vocabulary object detection. In: CVPR (2023) 10, 11", + "34. Wei, F., Gao, Y., Wu, Z., Hu, H., Lin, S.: Aligning pretraining for detection via object-level contrastive learning. In: NeurIPS (2021) 4", + "35. Wu, S., Zhang, W., Jin, S., Liu, W., Loy, C.C.: Aligning bag of regions for open-vocabulary object detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15254-15264 (2023) 2, 3, 10, 11, 12" + ], + "bbox": [ + 215, + 146, + 784, + 837 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "D. Kim et al.", + "bbox": [ + 271, + 114, + 361, + 126 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "36. Wu, X., Zhu, F., Zhao, R., Li, H.: Cora: Adapting clip for open-vocabulary detection with region prompting and anchor pre-matching. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7031-7040 (2023) 3, 10, 11", + "37. Xiao, T., Reed, C.J., Wang, X., Keutzer, K., Darrell, T.: Region similarity representation learning. In: ICCV. pp. 10539-10548 (October 2021) 4", + "38. Yao, L., Han, J., Liang, X., Xu, D., Zhang, W., Li, Z., Xu, H.: Detclipv2: Scalable open-vocabulary object detection pre-training via word-region alignment. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 23497-23506 (2023) 2, 4", + "39. Yao, L., Han, J., Wen, Y., Liang, X., Xu, D., Zhang, W., Li, Z., Xu, C., , Xu, H.: Detclip: Dictionary-enriched visual-concept paralleled pre-training for open-world detection. In: arXiv:2209.09407 (2022) 2, 4", + "40. Zang, Y., Li, W., Zhou, K., Huang, C., Loy, C.C.: Open-vocabulary DETR with conditional matching. In: ECCV. pp. 106-122. Springer (2022) 10, 11", + "41. Zareian, A., Rosa, K.D., Hu, D.H., Chang, S.F.: Open-vocabulary object detection using captions. In: CVPR (2021) 1, 6, 11", + "42. Zhang, H., Zhang, P., Hu, X., Chen, Y.C., Li, L., Dai, X., Wang, L., Yuan, L., Hwang, J.N., Gao, J.: Glipv2: Unifying localization and vision-language understanding. Advances in Neural Information Processing Systems 35, 36067-36080 (2022) 2, 4", + "43. Zhao, S., Zhang, Z., Schulter, S., Zhao, L., Stathopoulos, A., Chandraker, M., Metaxas, D., et al.: Exploiting unlabeled data with vision and language models for object detection. In: ECCV (2022) 1, 10, 11, 12", + "44. Zhong, Y., Yang, J., Zhang, P., Li, C., Codella, N., Li, L.H., Zhou, L., Dai, X., Yuan, L., Li, Y., Gao, J.: Regionclip: Region-based language-image pretraining. In: CVPR (2022) 1, 2, 3, 4, 10, 11, 12", + "45. Zhong, Y., Wang, J., Wang, L., Peng, J., Wang, Y.X., Zhang, L.: Dap: Detection-aware pre-training with weak supervision. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 4537-4546 (2021) 2, 4", + "46. Zhou, X., Girdhar, R., Joulin, A., Krahenbihl, P., Misra, I.: Detecting twenty-thousand classes using image-level supervision. In: ECCV (2022) 1, 2, 10, 11, 12" + ], + "bbox": [ + 215, + 146, + 787, + 603 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "Region-centric Image-Language Pretraining for Open-Vocabulary Detection", + "bbox": [ + 228, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + } +] \ No newline at end of file diff --git a/2024/Region-centric Image-Language Pretraining for Open-Vocabulary Detection/be279718-23a8-4b0a-ac0c-2bca133ca6e7_model.json b/2024/Region-centric Image-Language Pretraining for Open-Vocabulary Detection/be279718-23a8-4b0a-ac0c-2bca133ca6e7_model.json new file mode 100644 index 0000000000000000000000000000000000000000..17738d167b86fdcb5f3ead3f22f4cb0110bebcf5 --- /dev/null +++ b/2024/Region-centric Image-Language Pretraining for Open-Vocabulary Detection/be279718-23a8-4b0a-ac0c-2bca133ca6e7_model.json @@ -0,0 +1,2203 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.234, + 0.141, + 0.773, + 0.187 + ], + "angle": 0, + "content": "Region-centric Image-Language Pretraining for Open-Vocabulary Detection" + }, + { + "type": "text", + "bbox": [ + 0.327, + 0.213, + 0.677, + 0.229 + ], + "angle": 0, + "content": "Dahun Kim Anelia Angelova Weicheng Kuo" + }, + { + "type": "text", + "bbox": [ + 0.44, + 0.24, + 0.563, + 0.255 + ], + "angle": 0, + "content": "Google DeepMind" + }, + { + "type": "text", + "bbox": [ + 0.263, + 0.293, + 0.744, + 0.585 + ], + "angle": 0, + "content": "Abstract. We present a new open-vocabulary detection approach based on region-centric image-language pretraining to bridge the gap between image-level pretraining and open-vocabulary object detection. At the pretraining phase, we incorporate the detector architecture on top of the classification backbone, which better serves the region-level recognition needs of detection by enabling the detector heads to learn from large-scale image-text pairs. Using only standard contrastive loss and no pseudo-labeling, our approach is a simple yet effective extension of the contrastive learning method to learn emergent object-semantic cues. In addition, we propose a shifted-window learning approach upon window attention to make the backbone representation more robust, translation-invariant, and less biased by the window pattern. On the popular LVIS open-vocabulary detection benchmark, our approach sets a new state of the art of 37.6 mask \\(\\mathrm{AP}_r\\) using the common ViT-L backbone and public LAION dataset, and 40.5 mask \\(\\mathrm{AP}_r\\) using the DataComp-1B dataset, significantly outperforming the best existing approach by \\(+3.7\\) mask \\(\\mathrm{AP}_r\\). On the COCO benchmark, we achieve very competitive 39.6 novel AP without pseudo labeling or weak supervision. In addition, we evaluate our approach on the transfer detection setup, where it demonstrates notable improvement over the baseline. Visualization reveals emerging object locality from the pretraining recipes compared to the baseline.\\(^1\\)" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.609, + 0.377, + 0.625 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.635, + 0.788, + 0.725 + ], + "angle": 0, + "content": "Understanding and localizing all objects in the visual world is a key problem in computer vision, crucial for applications like self-driving cars. Traditional object detectors rely on costly human annotations, which limits scalability. Open-vocabulary detection (OVD) [41] addresses this by pretraining on large-scale image-text data (e.g. CLIP [28]), enabling detectors to identify objects based on user-provided text queries unavailable during training." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.726, + 0.788, + 0.817 + ], + "angle": 0, + "content": "Most existing open-vocabulary detection works assume the pretrained CLIP backbone is given, and focus on techniques such as knowledge distillation [5, 10], weak supervision [46], pseudo labeling [13, 29, 43, 44], and frozen backbone application [18], using the pretrained backbone. Consequently, during detection finetuning, the detector heads often need to be trained from scratch. This tends to result in sub-optimal generalization because the detector heads are trained on" + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.825, + 0.773, + 0.841 + ], + "angle": 0, + "content": "1 project page: github.com/google-research/google-research/tree/master/fvlm/dito" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.364, + 0.128 + ], + "angle": 0, + "content": "D. Kim et al." + }, + { + "type": "image", + "bbox": [ + 0.258, + 0.145, + 0.484, + 0.322 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.145, + 0.631, + 0.323 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.639, + 0.145, + 0.757, + 0.323 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.333, + 0.788, + 0.472 + ], + "angle": 0, + "content": "Fig. 1: DITO overview (left): Existing image-text pretraining methods for open-vocabulary detection [16,17] update only the vision transformer backbone and finetune detector heads from scratch. We propose to pretrain both backbone and detector heads directly from the large-scale image-text paired data, without a need for pseudo labeling or box annotations. In open-vocabulary detection finetuning, we introduce a simple shifted-window learning method to produce more robust representations from the pretrained vision transformer. DITO prediction (right): LVIS results only show the novel categories (e.g., bulldozer, fishbowl, subwoofer, heron). While Ego4D is a real-world and out-of-distribution data, many unseen objects are detected (e.g., steel lid, sticker on the wall, recycle bin). Best viewed with zoom in." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.491, + 0.787, + 0.521 + ], + "angle": 0, + "content": "the limited vocabulary of detection datasets, while only the backbone contains the knowledge of open-vocabulary concepts." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.522, + 0.788, + 0.674 + ], + "angle": 0, + "content": "Several studies have integrated detection models into CLIP pretraining. For instance, RegionCLIP [44] employs an off-the-shelf detector during CLIP pretraining to generate pseudo region-text labels. However, their pretraining only updates the image backbone, with the detector heads exclusively trained on detection data. These pseudo labeling-based pretraining methods [25, 44] require multi-stage training and handcrafted processing of pseudo labels, increasing complexity and training costs. Similarly, approaches such as GLIP [19, 42], Grounding DINO [22], CoDet [24] and DetCLIP [38, 39] integrate detector architectures in CLIP training, but their joint training demands additional detection and visual grounding datasets and complex multitask learning setups." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.675, + 0.789, + 0.841 + ], + "angle": 0, + "content": "In this paper, we propose a simple yet effective solution, Region-centric Pretraining approach, which incorporates detector modules into CLIP pretraining without the need for pseudo labeling or box annotations. This method involves generating random box regions across feature pyramid levels, followed by feature pooling over these regions. Subsequently, an image-text contrastive loss is applied, encouraging text-aligned region features to contribute more to the whole image representation. Our approach not only facilitates the warm-starting of detector heads in finetuning, but also leads to emergent representations with more localized semantic information compared to the baseline CLIP backbone, as demonstrated in our experiments. Compared to pseudo-labeling techniques [35,44-46], our approach can be viewed as an extension of contrastive" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.23, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Region-centric Image-Language Pretraining for Open-Vocabulary Detection" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.177 + ], + "angle": 0, + "content": "models to bypass the need for offline object proposal generation, which could be cumbersome for large-scale image-text data." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.178, + 0.787, + 0.435 + ], + "angle": 0, + "content": "In addition, we focus on the Vision Transformer (ViT) based CLIP models. These CLIP ViTs excel in zero-shot recognition, but are less commonly integrated into open-vocabulary detectors compared to ConvNets [18, 36, 44]. Pretraining ViT backbones on lower resolutions and then adapting them to higher-resolution detection images poses challenges, including increased computational demands and potential disruption of locality in pretrained features. Although the windowed attention technique of ViTDet [17, 20] reduces computation and preserves locality, it introduces bias from fixed window patterns. To address this, we propose Shifted-Window Learning (SWL) approach to enhance information mixing across fixed windows and mitigate grid pattern bias. Unlike the Swin Transformer [23], which applies shifted window layer by layer, SWL applies shifted windowed attention as a separate forward pass using the same ViT backbone. This simple strategy enhances the windowed attention representation when applied low-res trained ViT backbone to high-res images, ensuring compatibility with vanilla ViT backbones pretrained without shifted windows. SWL also better preserves the open-vocabulary knowledge of pretrained features compared to full-attention ViT perhaps due to its emphasis on local cues." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.435, + 0.788, + 0.556 + ], + "angle": 0, + "content": "Incorporating region-centric pretraining and shifted-window learning, our approach DITO (Detection-aligned Image-Text pretraining for Open-vocabulary detection) serves to narrow the gap between image-text pretraining and open-vocabulary detection, improving generalization. The best DITO model achieves 37.6 mask \\(\\mathrm{AP}_r\\) on LVIS benchmark, surpassing the previous best by \\(+3.7\\mathrm{AP}_r\\). It achieves the state-of-the-art 40.5 mask \\(\\mathrm{AP}_r\\) when pretrained on the DataComp1B dataset. On the COCO benchmark, DITO achieves a very competitive 39.6 novel AP without relying on pseudo-labels or joint training." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.565, + 0.388, + 0.581 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.595, + 0.788, + 0.777 + ], + "angle": 0, + "content": "Open-vocabulary detection. Conventional closed-set object detectors perform well but are limited by vocabulary size. Open-vocabulary detection, inspired by the strong zero-shot capabilities of Vision-Language Models (VLMs) like CLIP [28], has made notable progress. Various techniques have been explored, including knowledge distillation [10], prompt optimization [5], and the use of pretrained CLIP backbones with new detection heads [17,18,26,36]. Many top-performing methods [6,8,19,25,44] rely on pseudo labeling to prevent catastrophic forgetting during detection finetuning, but these methods often require multi-stage training [35,44] or extra steps to generate high-quality labels [25]. In contrast, our approach focuses on both upstream image-text pretraining and downstream open-vocabulary detection without using pseudo labeling, leading to significant improvements." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.787, + 0.841 + ], + "angle": 0, + "content": "Region-aligned Vision-Language Models. Building on advances in aligning image-text representations [28, 31], several studies have integrated region-level alignment into CLIP pretraining. For example, RegionCLIP [44] learns region-word alignment through pseudo region-text pairs but relies on an off-" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.363, + 0.128 + ], + "angle": 0, + "content": "D. Kim et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.239 + ], + "angle": 0, + "content": "the-shelf detector for pseudo labeling and focuses only on training the image backbone. GLIP [19, 42], Grounding DINO [22], DetCLIP [38, 39], CoDet [24] integrate detector architectures into CLIP training to align regions with words, but they require additional visual grounding annotations, leading to complex multitask learning. To our knowledge, our approach is the first to incorporate detector modules in image-text pretraining without relying on box annotations." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.244, + 0.79, + 0.397 + ], + "angle": 0, + "content": "Vision Transformers in open-vocabulary detection. While ViT-based CLIP models excel in zero-shot recognition, their adaptation to open-vocabulary detection is less explored compared to ConvNet-based CLIP models. For instance, OWL-ViT [25, 26] finetunes ViT on higher-resolution detection images with global attentions, but the full attention is computationally intensive. ROViT [17] introduces region-aware positional embeddings to aid in detection finetuning and uses windowed attention from ViTDet [20]. Our paper proposes shifted-window learning to enhance information mixing across fixed windows and reduce window-induced bias while preserving the locality structure of pretrained features, thereby improving open-vocabulary detection." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.402, + 0.79, + 0.57 + ], + "angle": 0, + "content": "Self-supervised pretraining for visual tasks. Self-supervised learning has emerged as a promising paradigm to learn object features for complex visual tasks such as detection, given the challenge of scaling up human annotations. Most relevant direction is contrastive learning, where the contrastive samples can take the forms of augmented images [3], sliding windows [37], object proposals [34], or point samples [1]. Some alternative strategies like pseudo-labeling [45] and pixel reconstruction [11] have also proven effective. While the majority of these methods have focused on learning from images without textual context, and applying to closed-vocabulary detection, DITO leverages large image-text data to tackle the more demanding open-vocabulary detection task, without a need for offline proposal generation [34, 45]." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.6, + 0.333, + 0.617 + ], + "angle": 0, + "content": "3 Method" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.641, + 0.788, + 0.718 + ], + "angle": 0, + "content": "We address the problem of open-vocabulary object detection. At training time, the model can access the class and box labels of base categories \\((C_B)\\). At test time, the model is tasked with detecting objects from a set of novel categories \\((C_N)\\) not present in the training set. To achieve this, we leverage pretrained vision and language models (VLMs) as in prior studies [10, 17, 18, 44]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.788, + 0.842 + ], + "angle": 0, + "content": "More specifically, we leverage recent advances in ViT-based detectors [17,20] for their promising results. However, instead of solely taking the pretrained ViT backbone, we demonstrate how to enhance the VLMs by integrating detector heads into the CLIP pretraining process, referred to as Region-centric image-text Pretraining (REP). Additionally, for detection finetuning, we propose a Shifted-Window Learning (SWL) to enhance the adaptation of the pretrained ViT to open-vocabulary detection task. By combining these approaches, DITO achieves significant improvements in open-vocabulary detection over prior arts." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.229, + 0.115, + 0.731, + 0.129 + ], + "angle": 0, + "content": "Region-centric Image-Language Pretraining for Open-Vocabulary Detection" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.784, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.146, + 0.608, + 0.35 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.619, + 0.153, + 0.77, + 0.348 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.359, + 0.788, + 0.457 + ], + "angle": 0, + "content": "Fig. 2: DITO method. Region-centric image-text pretraining (left): We train the detector heads (e.g. FPN [20, 21], Faster RCNN head [30], and RPN [30]) upon a ViT encoder backbone with multi-level image-text contrastive loss to bridge the gap between image-text pretraining and open-vocabulary detection. Shifting-window learning for detection (right): We roll the image and combines the shifted features with the original features to mitigate the bias of windowed attention grid [20], and produce more robust semantic representation." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.468, + 0.373, + 0.482 + ], + "angle": 0, + "content": "3.1 Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.488, + 0.788, + 0.625 + ], + "angle": 0, + "content": "Baseline. Our baseline approach is RO-ViT [17], a state-of-the-art ViT-based method for open-vocabulary detection. RO-ViT introduces CLIP pretraining with a new positional embedding scheme called cropped positional embedding (CPE). CPE involves randomly cropping and resizing the standard whole-image positional embedding during pretraining to enhance the ViT's generalization onto region-level recognition task and higher resolution detection inputs downstream. For detection finetuning, it adopts ViTDet [20] architecture, initialized with the pretrained ViT backbone. In the following, we describe the image-text pretraining and the downstream open-vocabulary detection." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.628, + 0.788, + 0.72 + ], + "angle": 0, + "content": "Image-text pretraining. We adopt dual-encoder CLIP pretraining widely used in existing works [28, 31]. The image embeddings \\(\\{v\\}\\) and text embeddings \\(\\{l\\}\\) are the average-pooled outputs from the image and text encoders, respectively. As in previous works, we compute the dot product of the embeddings in batch \\(B\\), and scale it by a learnable temperature \\(\\tau\\) before applying the InfoNCE loss [27, 28]. Mathematically, the image-to-text (I2T) loss can be expressed as:" + }, + { + "type": "equation", + "bbox": [ + 0.359, + 0.729, + 0.787, + 0.772 + ], + "angle": 0, + "content": "\\[\nL _ {\\mathrm {I 2 T}} = - \\frac {1}{B} \\sum_ {i = 1} ^ {B} \\log \\left(\\frac {\\exp \\left(v _ {i} l _ {i} / \\tau\\right)}{\\sum_ {j = 1} ^ {B} \\exp \\left(v _ {i} l _ {j} / \\tau\\right)}\\right). \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.78, + 0.788, + 0.843 + ], + "angle": 0, + "content": "The text-to-image (T2I) loss is symmetrical by exchanging the inner/outer summation loops. The total contrastive loss \\( L_{con} \\) is obtained by \\( L_{con} = (L_{\\mathrm{I2T}} + L_{\\mathrm{T2I}}) / 2 \\). As mentioned above, we adopt the cropped positional embeddings (CPE) following [17]." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.363, + 0.128 + ], + "angle": 0, + "content": "D. Kim et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.433 + ], + "angle": 0, + "content": "Open-vocabulary detection finetuning. At the fine-tuning stage, our detection finetuning recipe follows previous studies [10, 17, 18, 41]. During the training phase, we use the RoI-Align [12] feature as the detection embedding for each detected region. We replace the fixed-size classifier layer with the text embeddings of base categories \\((C_B)\\). The detection score \\(p_i\\) is determined by calculating the cosine similarity between the region embedding \\(r_i\\) and text embeddings of base categories \\((C_B)\\) followed by a softmax operation. We prepend an additional background class embedding to \\(C_B\\) and use the term \"background\" to represent the background category. Any proposals that do not match to any base category annotations are treated as background during training. It is important that the text embeddings are computed from the same text encoder as from the image-text pretraining. During testing, we expand the text embeddings to include the novel categories \\((C_B \\cup C_N)\\), resulting in \\((C_B \\cup C_N + 1)\\) categories including the background. We calculate the detection scores \\((p_i)\\) as the cosine similarity between the region embeddings \\((r_i)\\) and the expanded text embeddings. Apart from the detection embedding \\((r_i)\\), we extract the VLM embedding [18] of region \\(i\\) by RoI-Align at the last ViT backbone feature map. The VLM score \\((z_i)\\) is calculated as the cosine similarity with the text embeddings of the combined categories \\((C_B \\cup C_N)\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.434, + 0.788, + 0.465 + ], + "angle": 0, + "content": "To compute the open-vocabulary detection score \\((s_i^{\\mathrm{ens}})\\), we ensemble the detection and VLM scores by geometric means [10,18]. The formula is as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.386, + 0.476, + 0.786, + 0.517 + ], + "angle": 0, + "content": "\\[\ns _ {i} ^ {\\text {e n s}} = \\left\\{ \\begin{array}{l l} z _ {i} ^ {(1 - \\alpha)} \\cdot p _ {i} ^ {\\alpha} & \\text {i f} i \\in C _ {B} \\\\ z _ {i} ^ {(1 - \\beta)} \\cdot p _ {i} ^ {\\beta} & \\text {i f} i \\in C _ {N} \\end{array} \\right. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.524, + 0.788, + 0.57 + ], + "angle": 0, + "content": "Here, \\(\\alpha, \\beta\\) are floats \\(\\in [0,1]\\) that control the weighting of base versus novel categories. The background score comes from the detection score \\((p_i)\\) alone, because we observe the VLM score of \"background\" class is often less reliable." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.587, + 0.583, + 0.603 + ], + "angle": 0, + "content": "3.2 Region-centric Image-Text Pretraining" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.614, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Standard image-text pretraining uses classification architectures (e.g. ViT backbone followed by global pooling) as the language supervision occurs at the image level rather than the region level. Subsequently, for downstream detection, new detection heads are introduced and trained from scratch on a limited set of detection categories [16-18]. To fully utilize the knowledge embedded in large-scale image-text data, we propose Region-centric Pretraining (REP) which integrates the detector modules during the CLIP pretraining phase. Specifically, given access to image-text paired data but lacking box labels, our pretraining focuses on the region-recognition pathway of a detector, encompassing components like the backbone, FPN [20,21], RoI-Align [12], RPN-objectness [30], and Faster RCNN-classifier [30]. Consequently, the detector heads can be warm-started from the knowledge of large image-text data, thereby improving the generalization capability. To our knowledge, we are the first to integrate detector modules in image-text pretraining without box labels, and our experiments demonstrate clear benefits of our approach in open-vocabulary detection." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.229, + 0.115, + 0.731, + 0.129 + ], + "angle": 0, + "content": "Region-centric Image-Language Pretraining for Open-Vocabulary Detection" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.265 + ], + "angle": 0, + "content": "Detector head learning from random regions. Fig. 2 (left) illustrates our region-centric pretraining system. Following existing works [16, 17], we adopt SimpleFPN [20] and Faster R-CNN [30] models to remove the detector differences and study the benefits of our region-centric pretraining. Specifically, the multilevel feature pyramid is computed from the ViT backbone. Then, RoI-Align [12] and Faster R-CNN classifier head are applied to these feature maps to match the classification pathway in pretraining with the region-recognition pathway in detection finetuning (see Table 4b for ablations)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.268, + 0.785, + 0.358 + ], + "angle": 0, + "content": "For each level \\( i \\) of the feature pyramid, we randomly generate \\( n_i \\) box regions uniformly over the image by sampling the box size \\( h, w \\sim \\text{Uniform}(0.2, 0.5) \\) and aspect ratio \\( h / w \\sim \\text{Uniform}(0.5, 2.0) \\). The \\( n_i \\) value is set proportional to the size of the \\( i \\)-th feature map so that larger feature map would be covered by more regions. We extract the RoI-features of each region by RoI-Align, and feed them through the region classifier head [30] to obtain the RoI embeddings." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.362, + 0.784, + 0.573 + ], + "angle": 0, + "content": "Multi-level image-text supervision. After computing the RoI embeddings across pyramid levels, we perform a max pooling over the RoI embeddings per-level to obtain an image embedding for each pyramid level. Intuitively, max pooling allows the representation to focus on salient regions and discriminative features, thereby learning region-level information without explicit supervision. Then we apply the standard image-text contrastive loss (see Eqn. (1)) on each feature level separately, which aids the learning of rich semantic and object knowledge within every feature map (see Table 4c for ablations). The losses from all levels are weighted equally and summed together. Without explicit region-level supervision, the max pooling over regions encourages the more salient, text-aligned region features to contribute more to the whole image representation in the contrastive loss. Our experiments show emergent region-text alignments from the multi-level training, where the feature maps possess more localized semantic information compared to the baseline CLIP [17] backbone (see Fig. 3)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.578, + 0.784, + 0.772 + ], + "angle": 0, + "content": "Region proposal network learning. Fig. 3 shows that the learnt multi-level representations exhibit localized semantics well-aligned with the text query, i.e. the salient regions have higher similarity with the text relative to the background. Motivated by this, we employ the multi-level visual-text similarity as a supervisory signal for training the Region Proposal Network (RPN) [30]. Specifically, we compute the cosine similarity between each RoI embedding and the text, and use it as the target RPN score for the center location of each RoI. Any negative dot product value is mapped to zero to keep the target score in range [0,1]. We use L1 regression loss and set the loss weight equal to the multi-level contrastive loss (see supplementary). The losses are propagated only through the RoI centers and other pixels are ignored. Note that the box regression of the RPN is not trained here but learnt later through the detection finetuning, as we only use the image-text paired data without any box annotations." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.791, + 0.588, + 0.805 + ], + "angle": 0, + "content": "3.3 Shifted-Window Learning for Detection" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.801, + 0.84 + ], + "angle": 0, + "content": "The CLIP ViT backbones are typically pretrained on lower resolutions (e.g. \\(224 \\times 224\\)) and later adapted to higher resolution detection images (e.g. \\(1024 \\times 1024\\))." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.364, + 0.128 + ], + "angle": 0, + "content": "D. Kim et al." + }, + { + "type": "code_caption", + "bbox": [ + 0.218, + 0.146, + 0.514, + 0.162 + ], + "angle": 0, + "content": "Algorithm 1 Shifted Window backbone" + }, + { + "type": "algorithm", + "bbox": [ + 0.217, + 0.165, + 0.785, + 0.259 + ], + "angle": 0, + "content": "x: image patch tokens + positional embeddings. M: attention window size. s = M // 2 \\(y =\\) forward_vit_with_win_attn(x, M) \\(x^{\\prime} =\\) np-roll(x, shift=[s, s], axis=[0, 1]) # shift \\(y^\\prime =\\) forward_vit_with_win_attn \\((x^{\\prime},M)\\) \\(y'' =\\) np-roll(y', shift=[-s, -s], axis=[0, 1]) # unshift return \\((y + y'') / 2\\) # SWL backbone output" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.3, + 0.789, + 0.409 + ], + "angle": 0, + "content": "While detection tasks benefit from global information, directly applying the pretrained ViT to high-resolution inputs is computationally intensive and can disrupt the locality structure of the pretrained lower-res features. Windowed attention in ViTDet [20] helps reduce computation but introduces bias from the fixed-size grid pattern. To address this, we propose the Shifted-Window Learning (SWL) approach, which improves information mixing across fixed windows and mitigates grid pattern bias." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.415, + 0.79, + 0.689 + ], + "angle": 0, + "content": "Network architecture. Fig. 2 (right) and Algorithm 1 describe the SWL algorithm. The standard ViT consists of a patchifying layer and a set of transformer layers. After feeding the image through the patchifying layer, we obtain a feature map \\( x \\) of shape \\( (h, w, d) \\). This feature map \\( x \\) is fed through the rest of the ViT with windowed attention layers on a grid \\( K \\times K \\), and \\( L \\) global attention layers evenly spaced throughout the ViT (where \\( L = 4 \\)) following [20], resulting in output \\( y \\). In parallel, we create another copy of \\( x \\), which is rolled along both \\( h \\) and \\( w \\) axes by \\( s \\) pixels. The elements that roll beyond the last position are reintroduced from the first. We carefully design the attention masks such that the rolled around patches would not attend to the patches on the other side of the image (see the right figure of Algorithm 1). The shift size \\( s \\) is set as the half of the attention window size \\( M \\) (i.e. \\( s = M // 2 \\)). Empirically, the window size \\( M = 16 \\) equals the image size (e.g. 1024) divided by the product of patch size (e.g. \\( P = 16 \\)) and the grid size (e.g. \\( K = 4 \\)). The shifted feature map \\( x' \\) is then processed through the rest of the ViT in the same manner, resulting in output \\( y' \\) in the same shape \\( (h, w, d) \\) as \\( y \\). We then unshift \\( y' \\) and combine it with \\( y \\) by averaging. We apply the above shifted window operations in detection finetuning and inference times." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.697, + 0.81, + 0.851 + ], + "angle": 0, + "content": "Comparison with Swin Transformer. Compared to the Swin Transformer [23], we apply the shifted-window ideas as separate forward passes, while Swin Transformer applies similar ideas in an alternating manner layer by layer. Our approach requires no change to the vanilla transformer architecture and is compatible with any ViT backbones pretrained without shifted windows (e.g. [28]), whereas Swin Transformer requires specialized pretraining on the same architecture. Compared to the full-attention ViT [4], we observe that shifted-window ViT taps more effectively into the semantic knowledge of pretrained backbone than full global attention, perhaps because the window helps the model focus more on local cues and ignore the noises farther away." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.229, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Region-centric Image-Language Pretraining for Open-Vocabulary Detection" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.147, + 0.584, + 0.161 + ], + "angle": 0, + "content": "3.4 Distillation from Frozen ViT Backbone" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.168, + 0.788, + 0.379 + ], + "angle": 0, + "content": "While the ViT backbone adapts to the detection tasks, it may lose some of its pretrained open-vocabulary knowledge. Therefore, we propose a simple distillation approach which uses a separate frozen ViT to teach the finetuned ViT backbone during the detection finetuning. We use a cosine distance loss that aligns the RoI-Align embeddings extracted from the feature maps of both backbones. The cosine distance is computed for each RoI then averaged over all RoIs. Empirically, we find it advantageous to add a \\(1 \\times 1\\) Conv projection layer to the finetuned ViT backbone before the RoI-Align, allowing some flexibility in distillation when jointly trained with other detection objectives. The auxiliary distillation loss is then added to other detection losses, with a loss weight \\(\\gamma = 1\\) (see supplementary). At inference, the ViT backbone features after the projection are used to compute the region VLM score \\(z_{i}\\) (Sec. 3.1). It is worth noting that the frozen ViT backbone is only used during training for distillation purposes and is removed at inference." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.38, + 0.788, + 0.47 + ], + "angle": 0, + "content": "While previous studies [2, 10] have utilized knowledge distillation from the CLIP models, their teacher CLIP models mostly operate on image crops in an offline process, thus needing multiple forward passes through the backbone for all RoIs. In contrast, our distillation operates efficiently on RoIs cropped from dense feature maps in a single forward pass during the detection finetuning with minimal overhead." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.486, + 0.462, + 0.503 + ], + "angle": 0, + "content": "4 Experimental Results" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.517, + 0.789, + 0.744 + ], + "angle": 0, + "content": "Baseline reproduction. As discussed in Sec. 3.1, our baseline method is RO-ViT [17], a leading ViT-based approach for open-vocabulary detection. RO-ViT introduces cropped positional embedding (CPE) in CLIP pretraining to enhance the generalization of the pretrained ViT onto the downstream detection. Additionally, it adopts global average pooling of the ViT features instead of CLs-token pooling, which simplifies the adaptation onto higher resolution inputs, and the extraction of region features (e.g. RoI-Align) at the final layer. We reproduce the CLIP pretraining of RO-ViT [17] using the widely adopted LAION2B dataset [31], as the OpenAI CLIP's WIT [28] and ALIGN [14] datasets are not publicly available. We trained the CLIP models from scratch, following the same pretraining protocol and hyperparameters, including 500k iterations and 16k batch size (i.e. 8B samples seen in training), \\(224 \\times 224\\) image size, global average pooling, and cropped positional embedding. We use the standard InfoNCE loss (Eqn. (1)) instead of their focal constrastive loss [17]. The following compares our reproduced CLIP with OpenAI CLIP [28] and OpenCLIP [31]:" + }, + { + "type": "table", + "bbox": [ + 0.219, + 0.752, + 0.784, + 0.802 + ], + "angle": 0, + "content": "
methoddataset#samples seenbackboneImageNet Top-1 Acc.
CLIP [28]WIT-400M13BViT-L/1475.5
OpenCLIP [31]LAION-2B32BViT-L/1475.2
RO-ViT CLIP [17] our repro.LAION-2B8BViT-L/1673.9
" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.81, + 0.787, + 0.84 + ], + "angle": 0, + "content": "On the zero-shot ImageNet classification benchmark, the LAION-2B pretraining matches or slightly underperforms the OpenAI CLIP with WIT-400M." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.362, + 0.127 + ], + "angle": 0, + "content": "D. Kim et al." + }, + { + "type": "table", + "bbox": [ + 0.219, + 0.144, + 0.785, + 0.452 + ], + "angle": 0, + "content": "
methodpretraining modelpretraining datadetector backbonew/ pseudo box labelsmask APrmask AP
ConvNet based:
OV-DETR [40]ViT-B/32CLIP-400MR-50-17.426.6
Kaul et al. [15]ViT-B/32CLIP-400MR-50-19.330.6
DetPro-Cascade [5]ViT-B/32CLIP-400MR-50-20.027.0
Rasheed [29]ViT-B/32CLIP-400MR-50-21.125.9
BARON [35]ViT-B/32CLIP-400MR-50-22.627.6
CoDet [24]R-50CLIP-400M + CC3MR-50-23.430.7
EdaDet [32]R-50CLIP-400MR-50-23.727.5
VL-PLM [43]ViT-B/32CLIP-400MR-50✓17.227.0
PromptDet [6]ViT-B/32CLIP-400MR-50✓21.425.3
OADB [33]ViT-B/32CLIP-400MR-50✓21.726.6
RegionCLIP [44]R-50x4CLIP-400M + CC3MR-50x4✓22.032.3
CORA [36]R-50x4CLIP-400MR-50x4✓22.2bx-
Detric-CN2 [46]ViT-B/32CLIP-400M + INet-21KR-50WS24.632.4
ViLD-Ens [10]EffNet-B7ALIGN-1.8BEffNet-B7-26.329.3
F-VLM [18]R-50x64CLIP-400MR-50x64-32.834.9
ViT based:
OWL-ViT [26]O+VGViT-L/14CLIP-400MViT-L/14-25.6bx34.7bx
OWL-ViT v2 [25]O+VGViT-L/14WebLI-10BViT-L/14✓45.9bx50.4bx
RO-ViT [17]ViT-B/16ALIGN-1.8BViT-B/16-28.030.2
RO-ViT [17]†ViT-L/16LAION-2BViT-L/16-32.432.9
CFM-ViT [16]ViT-B/16ALIGN-1.8BViT-B/16-28.832.0
CFM-ViT [16]ViT-L/16ALIGN-1.8BViT-L/16-33.936.6
CFM-ViT [16]*ViT-L/16LAION-2BViT-L/16-33.836.4
DITO (ours)ViT-S/16LAION-2BViT-S/16-26.228.8
DITO (ours)ViT-B/16LAION-2BViT-B/16-31.532.4
DITO (ours)ViT-L/16LAION-2BViT-L/16-37.636.2
DITO (ours)ViT-L/16DataComp-1BViT-L/16-40.538.0
" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.455, + 0.788, + 0.512 + ], + "angle": 0, + "content": "Table 1: LVIS open-vocabulary detection (mask AP). DITO outperforms the best existing approach by \\(+3.7\\) mask \\(\\mathrm{AP}_r\\). WS: uses weak supervision from ImageNet-21K. †: reports LAION-2B results in arXiv version. *: our reproduced result using LAION-2B. O+VG: uses extra Objects365 and Visual Genome data." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.534, + 0.788, + 0.686 + ], + "angle": 0, + "content": "Pretraining setup. After the above-mentioned baseline CLIP training, we apply our region-centric pretraining (REP) where we freeze the image and text encoders trained in the first phase and introduce the detector heads. We use the Simple FPN [20], and classification layers of Faster R-CNN and RPN [30], where we replace the batch normalization with layer normalization. At the \\(i\\)-th pyramid level \\(i \\in \\{2, 3, 4, 5\\}\\), we randomly sample \\(n_i \\in \\{400, 200, 100, 50\\}\\) box regions and compute their RoI-Align features. We use a short training cycle of \\(30k\\) iterations, \\(4k\\) batch size, \\(256 \\times 256\\) image size, AdamW optimizer with an initial learning rate of 1e-4 with linear decay. For both phases of image-text pretraining, we use the publicly available LAION-2B [31] dataset." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Detection finetuning setup. As noted in Sec. 3.2, we adopt the ViTDet [20] with SimpleFPN as our detector following the baseline works [16,17]. We follow the same finetuning settings of [17]. Specifically, we train the detector with image size \\(1024 \\times 1024\\) and use windowed attention in the backbone with grid size \\(4 \\times 4\\). The learning rate for the backbone is set lower as \\(0.6 \\times\\) of the detector head layers. We use \\(\\alpha = 0.3\\), \\(\\beta = 0.65\\) for score combination in Eqn. (2). The text embedding of each category is calculated as the average over the CLIP prompt templates. We use the batch size 128, the SGD optimizer with momentum 0.9. The initial learning rate and iterations are set to 0.18 and \\(36.8\\mathrm{k}\\) for LVIS, and 0.02 and \\(11.3\\mathrm{k}\\) for COCO datasets." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.23, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Region-centric Image-Language Pretraining for Open-Vocabulary Detection" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.117, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "table", + "bbox": [ + 0.217, + 0.144, + 0.787, + 0.429 + ], + "angle": 0, + "content": "
methodpretraining modelpretraining datadetector backbonew/ pseudo box labelsnovel APAP
ConvNet based:
OVR-CNN [41]R-50CLIP-400M + COCO-CapR-50-22.839.9
ViLD [10]ViT-B/32CLIP-400MR-50-27.651.3
F-VLM [18]R-50CLIP-400MR-50-28.039.6
OV-DETR [40]ViT-B/32CLIP-400MR-50-29.452.7
CoDet [24]R-50CLIP-400M + COCO-CapR-50-30.646.6
PromptDet [6]ViT-B/32CLIP-400MR-50✓26.650.6
XPM [13]R-50CLIP-400MR-50✓27.041.2
OADB [33]ViT-B/32CLIP-400M + COCO-CapR-50✓30.047.2
VL-PLM [43]ViT-B/32CLIP-400MR-50✓34.453.5
RegionCLIP [44]R-50x4CLIP-400M + CC3M + COCO-CapR-50x4✓39.355.7
EdaDet [32]R-50CLIP-400MR-50✓40.252.5
CORA [36]R-50x4CLIP-400MR-50x4✓41.743.8
Detic-CN2 [46]ViT-B/32CLIP-400M + INet-21KR-50WS27.845.0
ViT based:
RO-ViT [17]ViT-B/16ALIGN-1.8BViT-B/16-30.241.5
RO-ViT [17]ViT-L/16ALIGN-1.8BViT-L/16-33.047.7
RO-ViT [17]*ViT-L/16LAION-2BViT-L/16-33.347.9
CFM-ViT [16]ViT-B/16ALIGN-1.8BViT-B/16-30.842.4
CFM-ViT [16]ViT-L/16ALIGN-1.8BViT-L/16-34.146.0
CFM-ViT [16]*ViT-L/16LAION-2BViT-L/16-34.346.4
DITO (ours)ViT-S/16LAION-2BViT-S/16-32.344.4
DITO (ours)ViT-B/16LAION-2BViT-B/16-36.648.8
DITO (ours)ViT-L/16LAION-2BViT-L/16-39.654.4
DITO (ours)ViT-L/16DataComp-1BViT-L/16-40.254.6
" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.432, + 0.788, + 0.473 + ], + "angle": 0, + "content": "Table 2: COCO open-vocabulary detection (box AP50). DITO demonstrates a very competitive novel category AP without using pseudo labeling or weak supervision (WS). *: our reproduced result using LAION-2B." + }, + { + "type": "table", + "bbox": [ + 0.316, + 0.48, + 0.681, + 0.595 + ], + "angle": 0, + "content": "
methodbackboneAP\\( AP_{50} \\)\\( AP_{75} \\)
Supervised [10]R-5025.638.628.0
ViLD [10]R-5011.818.212.6
DetPro [5]R-5012.118.812.9
BARON [35]R-5013.621.014.5
F-VLM [18]R-50x1616.225.317.5
F-VLM [18]R-50x6417.727.419.1
RO-ViT [17]ViT-L/1617.126.918.5
CFM-ViT [16]ViT-L/1618.728.920.3
DITO (ours)ViT-L/1620.031.821.5
" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.597, + 0.785, + 0.627 + ], + "angle": 0, + "content": "Table 3: Zero-shot transfer detection from LVISbase to Objects365 (box AP). All models are tested on Objects365 dataset following the setup of [10]." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.648, + 0.373, + 0.663 + ], + "angle": 0, + "content": "4.1 Main Results" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.788, + 0.841 + ], + "angle": 0, + "content": "LVIS Benchmark. In Table 1, we report the comparison with existing methods on the challenging LVIS benchmark. The 'frequent' and 'common' classes of the dataset belong to the base categories \\( \\mathrm{C}_B \\), and the 'rare' classes are the novel categories \\( \\mathrm{C}_N \\). The primary metric is the mask AP on rare classes (mask \\( \\mathrm{AP}_r \\)). The DITO model achieves the performance of 37.6 mask \\( \\mathrm{AP}_r \\), which significantly outperforms the state-of-the-art approach RO-ViT [17] with the same ViT-L backbone by +5.1 points using the same pretraining data LAION-2B [31]. We also outperform the state-of-the-art CFM-ViT [16] by +3.6 points. Our best performance sets a new state-of-the-art 40.9 mask \\( \\mathrm{AP}_r \\) when using DataComp1B [7] in pretraining. With the ViT-B backbone, DITO maintains a healthy margin of around +2.5 \\( \\mathrm{AP}_r \\) above existing ViT-B based approaches." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.362, + 0.127 + ], + "angle": 0, + "content": "D. Kim et al." + }, + { + "type": "table", + "bbox": [ + 0.229, + 0.144, + 0.449, + 0.222 + ], + "angle": 0, + "content": "
methodAPrAP
RO-ViT [17]32.432.9
RO-ViT our repro.32.233.0
w/REP34.8 (+2.6)34.9
w/SWL35.0 (+2.8)35.2
w/REP+SWL36.3 (+4.1)35.8
w/REP+SWL+FD37.6 (+5.4)36.2
" + }, + { + "type": "table_caption", + "bbox": [ + 0.262, + 0.224, + 0.402, + 0.236 + ], + "angle": 0, + "content": "(a) DITO framework." + }, + { + "type": "table", + "bbox": [ + 0.462, + 0.145, + 0.621, + 0.202 + ], + "angle": 0, + "content": "
pretraining methodAPr
RO-ViT our repro.32.2
w/FPN33.2
w/FPN+Head34.2
w/FPN+Head+RPN34.8
" + }, + { + "type": "table_caption", + "bbox": [ + 0.453, + 0.223, + 0.62, + 0.236 + ], + "angle": 0, + "content": "(b) Detector components." + }, + { + "type": "table", + "bbox": [ + 0.636, + 0.145, + 0.787, + 0.211 + ], + "angle": 0, + "content": "
RoI embeddingAPr
global avg/lvl34.0
global max/lvl33.7
multi RoIs, avg/lvl33.8
multi RoIs, max/lvl34.8
multi RoIs, max all34.1
" + }, + { + "type": "table_caption", + "bbox": [ + 0.639, + 0.223, + 0.768, + 0.236 + ], + "angle": 0, + "content": "(c) RoI embedding." + }, + { + "type": "table_caption", + "bbox": [ + 0.215, + 0.243, + 0.785, + 0.285 + ], + "angle": 0, + "content": "Table 4: Ablation on overall DITO framework and Region-centric Pretraining (Sec. 3.2). REP: Region-centric Pretraining, SWL: Shifted-Window Learning, FD: Frozen backbone Distillation. Best setting is in gray." + }, + { + "type": "table", + "bbox": [ + 0.245, + 0.293, + 0.495, + 0.351 + ], + "angle": 0, + "content": "
backboneAPrAP
fully global attn.33.433.8
baseline window attn. [17]32.233.0
Swin [23] style31.333.1
shifted window35.035.2
" + }, + { + "type": "table_caption", + "bbox": [ + 0.269, + 0.352, + 0.454, + 0.364 + ], + "angle": 0, + "content": "(a) Shifted-window learning." + }, + { + "type": "table", + "bbox": [ + 0.517, + 0.293, + 0.774, + 0.351 + ], + "angle": 0, + "content": "
# global attn. layerbasew/ SWL
030.734.6 (+3.9)
432.235.0 (+2.8)
1232.434.2 (+1.8)
24 (all layers)33.433.4 (+0.0)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.506, + 0.352, + 0.773, + 0.364 + ], + "angle": 0, + "content": "(b) SWL w.r.t. # global attention layers." + }, + { + "type": "table_caption", + "bbox": [ + 0.251, + 0.372, + 0.75, + 0.386 + ], + "angle": 0, + "content": "Table 5: Ablation on Shifted-Window Learning (SWL - Sec. 3.3)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.402, + 0.785, + 0.508 + ], + "angle": 0, + "content": "COCO Benchmark. We present the comparison on COCO benchmark in Table 2. The main metric is AP50 of novel categories (novel AP). Without using pseudo labeling [6,13,43,44], weak supervision [46], or externally trained detector modules [29, 35], our model demonstrates competitive results of 39.6 novel AP with LAION-2B and 40.2 with DataComp-1B. Among the ViT-based methods, DITO outperforms recent works RO-ViT [17] and CFM-ViT [16] by a clear margin of \\(+6.3\\) and \\(+5.3\\) points, respectively." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.512, + 0.785, + 0.587 + ], + "angle": 0, + "content": "Transfer detection. We further evaluate DITO in the transfer detection setting, where the open-vocabulary detector trained on one dataset \\((\\mathrm{LVIS}_{base})\\) is tested on another dataset (Objects365) without any finetuning. By simply replacing the text embeddings, Table 3 shows that DITO achieves 20.0 AP, outperforming previous methods using ConvNet or ViT backbones of similar size." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.605, + 0.401, + 0.618 + ], + "angle": 0, + "content": "4.2 Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.622, + 0.785, + 0.652 + ], + "angle": 0, + "content": "For ablation study, we use the ViT-L/16 model pretrained with LAION-2B, and evaluate on the LVIS benchmark and report mask \\(\\mathrm{AP}_r\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.655, + 0.793, + 0.761 + ], + "angle": 0, + "content": "DITO overall framework. Table 4a summarizes the benefits of each DITO components. The region-centric pretraining (REP) improves the contrastive model baseline by \\(+2.6\\mathrm{AP}_r\\) and shifted window learning (SWL) improves the baseline by \\(+2.8\\) points. Combining both strategies brings a significant gain of \\(+4.1\\mathrm{AP}_r\\). Lastly, we observe that incorporating the frozen backbone distillation leads to an additional boost of \\(+1.2\\) points. In the following, we provide ablations for each components." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Region-centric Pretraining. In Table 4b, we ablate our region-centric image-text pretraining by progressively adding the FPN, Faster R-CNN head, and RPN into the pretraining. Our 'baseline' is the contrastive image-text pretraining with cropped positional embedding [17]. On top of this, 'w/ FPN' introduces the FPN into the pretraining, where each pyramid level (whole image) map is mean-pooled" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.229, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Region-centric Image-Language Pretraining for Open-Vocabulary Detection" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "table", + "bbox": [ + 0.234, + 0.144, + 0.516, + 0.201 + ], + "angle": 0, + "content": "
backboneAPAPrAPcAPf
GT boxes given (region classification):
before det. training48.756.851.542.1
after det. training53.754.253.753.5
after det. training w/FD54.857.754.353.2
" + }, + { + "type": "table_caption", + "bbox": [ + 0.281, + 0.203, + 0.454, + 0.215 + ], + "angle": 0, + "content": "(a) GT region recognition." + }, + { + "type": "table", + "bbox": [ + 0.535, + 0.144, + 0.784, + 0.202 + ], + "angle": 0, + "content": "
methodAPrAP
RO-ViT our repro.32.233.0
RO-ViT our repro.+FD34.0 (+1.8)33.6
w/ REP+SWL36.335.8
w/ REP+SWL+FD37.6 (+1.3)36.2
" + }, + { + "type": "table_caption", + "bbox": [ + 0.57, + 0.203, + 0.738, + 0.215 + ], + "angle": 0, + "content": "(b) Effect of FD on OVD." + }, + { + "type": "table_caption", + "bbox": [ + 0.251, + 0.223, + 0.75, + 0.238 + ], + "angle": 0, + "content": "Table 6: Ablation on frozen backbone distillation (FD - Sec. 3.4)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.258, + 0.785, + 0.348 + ], + "angle": 0, + "content": "into an image embedding, followed by the image-text contrastive loss per level. It improves the baseline by \\(+1.0\\mathrm{AP}_r\\). Adding the Faster R-CNN head further improves the alignment between the pretraining and detection finetuning, showing a gain of \\(+2.0\\mathrm{AP}_r\\). Incorporating all components i.e. the FPN, Faster R-CNN and RPN heads achieves the best 34.8 \\(\\mathrm{AP}_r\\), a significant gain of +2.6 points over the baseline." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.349, + 0.785, + 0.47 + ], + "angle": 0, + "content": "Table 4c shows that both global avg- and max-pooling are sub-optimal due to the lack of saliency map (avg-pool), and the limited capacity of a single pixel to represent semantic concepts for contrastive learning (max-pool), respectively. Our approach combines the best of both worlds, by first avg-pooling within each RoI, and then max-pooling over these RoI embeddings. Each embedding represents a proper RoI and the global representation captures the saliency map through max-pooling. Pooling per pyramid level (i.e. multi-level image-text supervision) outperforms pooling over all levels." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.479, + 0.405, + 0.529 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.23, + 0.531, + 0.394, + 0.541 + ], + "angle": 0, + "content": "\"am riding a cart pulled by a donkey" + }, + { + "type": "image", + "bbox": [ + 0.408, + 0.479, + 0.596, + 0.529 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.418, + 0.531, + 0.587, + 0.541 + ], + "angle": 0, + "content": "\"a crane operates amidst piles of rubble\"" + }, + { + "type": "image", + "bbox": [ + 0.598, + 0.479, + 0.785, + 0.529 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.611, + 0.529, + 0.774, + 0.546 + ], + "angle": 0, + "content": "\"aperson wearing blue rubber gloves is loading a syringe with vegetables\"" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.547, + 0.405, + 0.596 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.232, + 0.597, + 0.393, + 0.614 + ], + "angle": 0, + "content": "\"amotherand her two children sit on a green bench\"" + }, + { + "type": "image", + "bbox": [ + 0.408, + 0.547, + 0.596, + 0.596 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.421, + 0.597, + 0.583, + 0.613 + ], + "angle": 0, + "content": "\"Cows on the street next to people that are on the sidewalk\"" + }, + { + "type": "image", + "bbox": [ + 0.598, + 0.547, + 0.785, + 0.596 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.616, + 0.597, + 0.769, + 0.613 + ], + "angle": 0, + "content": "\"small white bird sits on the back of a horse in an open field\"" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.619, + 0.785, + 0.675 + ], + "angle": 0, + "content": "Fig. 3: Visual-text similarity map. For each example, we show the paired image (left) and text (bottom) input, and the visual-text similarity map using the contrastive model baseline [17] backbone features (middle) or our region-centric pretraining features (right). We use Flickr30K (top row) and COCO Captions (bottom row) datasets." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Shifted-Window Learning for detection. The CLIP ViT backbones are initially pretrained on lower resolutions and then adapted to higher resolution detection images. In Table 5a, we assess the efficacy of the shifted-window backbone in open-vocabulary detection training. Although the fully global attention model improves the detection task compared to the baseline windowed attention model [17], directly applying the pretrained ViT on high-resolution inputs may not be optimal, potentially compromising the locality structure of the pretrained lower-resolution features. Our shifted-window learning approach achieves \\(35.0\\mathrm{AP}_r\\) by preserving the locality structure from windowed attention as well as integrating information across fixed windows, outperforming the fully global" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.362, + 0.128 + ], + "angle": 0, + "content": "D. Kim et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.253 + ], + "angle": 0, + "content": "attention model (33.4 \\(\\mathrm{AP}_r\\)) which is computationally intensive. Notably, OWL-ViT [25,26] adopts the fully global attention model. In addition, naively applying the layer-alternating shifted window as in Swin [23] leads to a performance drop (see Sec. 3.3). Table 5b delves deeper into the behavior of SWL. The advantage of SWL diminishes steadily with increasing number of global attention layers in the windowed attention backbone, validating its better information mixing enabled by the SWL." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.256, + 0.788, + 0.468 + ], + "angle": 0, + "content": "Frozen backbone distillation. Table 6a studies the region classification capability of the CLIP ViT backbone before and after the detection finetuning. We use the ground truth boxes and measures AP scores to evaluate the zero-shot region classification of the base ('frequent' + 'common') and novel ('rare') categories. We observe that the finetuned backbone indeed shows a notable drop of \\(-2.6\\mathrm{AP}_r\\) on novel classes, while overfitting to the base classes. The frozen backbone distillation ('w/ FD') leads to a significant improvement of \\(+3.5\\mathrm{AP}_r\\) even surpassing the frozen pretrained backbone (before detection finetuning), while maintaining performance on the base classes (\\(\\mathrm{AP}_c\\) and \\(\\mathrm{AP}_f\\)). Table 6b presents the open-vocabulary detection results where the frozen backbone distillation brings a clear gain of \\(+1.3\\sim 1.8\\mathrm{AP}_r\\). These results highlight the efficacy of our frozen backbone distillation in detection finetuning (Sec. 3.4), as it effectively preserves the pretrained open-vocabulary knowledge while acquiring explicit region-text alignment through detection supervision." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.482, + 0.373, + 0.496 + ], + "angle": 0, + "content": "4.3 Visualization" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.499, + 0.788, + 0.589 + ], + "angle": 0, + "content": "In Fig. 3, we visualize the similarity maps between image features and query text embeddings. We compare the baseline [17] contrastive model's backbone features with our region-centric pretrained features. We select pyramid level 4 which has the same resolution as the backbone features and apply the Faster R-CNN head in a sliding window manner to obtain the dense feature map. We observe that region-centric pretraining captures more localized semantic" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.59, + 0.788, + 0.651 + ], + "angle": 0, + "content": "Fig. 1 visualizes DITO outputs on LVIS novel categories and real-world out-of-distribution data from Ego4D [9]. Using the same DITO detector trained on \\(\\mathrm{LVIS}_{\\text {base }}\\), we find that DITO effectively identifies many novel and unseen objects, even under significant domain shifts." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.666, + 0.362, + 0.683 + ], + "angle": 0, + "content": "5 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.788, + 0.841 + ], + "angle": 0, + "content": "We introduce DITO, a region-centric approach for open-vocabulary detection using large-scale image-text pairs. By integrating detection architecture onto the image backbone in CLIP pretraining, it learns locality-sensitive information without requiring pseudo labeling or box annotations. Furthermore, we propose a shifted-window learning method to mitigate the bias of the window attention pattern in CLIP ViT detectors. Experiments show that DITO outperforms the state-of-the-art by large margins on the LVIS benchmark, and is very competitive on the COCO benchmark and transfer detection. We hope this work would inspire the community to explore region-centric image-language pretraining for open-vocabulary localization tasks." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.229, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Region-centric Image-Language Pretraining for Open-Vocabulary Detection" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.145, + 0.324, + 0.161 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.173, + 0.785, + 0.2 + ], + "angle": 0, + "content": "1. Bai, Y., Chen, X., Kirillov, A., Yuille, A., Berg, A.C.: Point-level region contrast for object detection pre-training. In: CVPR. pp. 16061-16070 (June 2022) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.201, + 0.785, + 0.241 + ], + "angle": 0, + "content": "2. Chen, J., Zhu, D., Qian, G., Ghanem, B., Yan, Z., Zhu, C., Xiao, F., Elhoseiny, M., Culatana, S.C.: Exploring open-vocabulary semantic segmentation without human labels. arXiv preprint arXiv:2306.00450 (2023) 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.242, + 0.785, + 0.281 + ], + "angle": 0, + "content": "3. Chen, T., Kornblith, S., Norouzi, M., Hinton, G.: A simple framework for contrastive learning of visual representations. arXiv preprint arXiv:2002.05709 (2020) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.282, + 0.785, + 0.337 + ], + "angle": 0, + "content": "4. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929 (2020) 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.338, + 0.785, + 0.377 + ], + "angle": 0, + "content": "5. Du, Y., Wei, F., Zhang, Z., Shi, M., Gao, Y., Li, G.: Learning to prompt for open-vocabulary object detection with vision-language model. In: CVPR (2022) 1, 3, 10, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.378, + 0.785, + 0.419 + ], + "angle": 0, + "content": "6. Feng, C., Zhong, Y., Jie, Z., Chu, X., Ren, H., Wei, X., Xie, W., Ma, L.: Promptdet: Towards open-vocabulary detection using uncurated images. In: European Conference on Computer Vision. pp. 701-717. Springer (2022) 3, 10, 11, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.42, + 0.785, + 0.46 + ], + "angle": 0, + "content": "7. Gadre, S.Y., Ilharco, G., Fang, A., Hayase, J., Smyrnis, G., Nguyen, T., Marten, R., Wortsman, M., Ghosh, D., Zhang, J., et al.: Datacomp: In search of the next generation of multimodal datasets. arXiv preprint arXiv:2304.14108 (2023) 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.46, + 0.785, + 0.487 + ], + "angle": 0, + "content": "8. Gao, M., Xing, C., Niebles, J.C., Li, J., Xu, R., Liu, W., Xiong, C.: Open vocabulary object detection with pseudo bounding-box labels. In: ECCV (2022) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.488, + 0.785, + 0.528 + ], + "angle": 0, + "content": "9. Grauman, K., Westbury, A., Byrne, E., Chavis, Z., Furnari, A., Girdhar, R., Hamburger, J., Jiang, H., Liu, M., Liu, X., et al.: Ego4d: Around the world in 3,000 hours of egocentric video. In: CVPR. pp. 18995-19012 (2022) 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.529, + 0.785, + 0.554 + ], + "angle": 0, + "content": "0. Gu, X., Lin, T.Y., Kuo, W., Cui, Y.: Open-vocabulary object detection via vision and language knowledge distillation. In: ICLR (2022) 1, 3, 4, 6, 9, 10, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.555, + 0.785, + 0.595 + ], + "angle": 0, + "content": "1. He, K., Chen, X., Xie, S., Li, Y., Dólar, P., Girshick, R.: Masked autoencoders are scalable vision learners. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16000-16009 (2022) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.596, + 0.776, + 0.609 + ], + "angle": 0, + "content": "2. He, K., Gkioxari, G., Dollár, P., Girshick, R.: Mask r-cnn. In: ICCV (2017) 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.61, + 0.785, + 0.663 + ], + "angle": 0, + "content": "3. Huynh, D., Kuen, J., Lin, Z., Gu, J., Elhamifar, E.: Open-vocabulary instance segmentation via robust cross-modal pseudo-labeling. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7020-7031 (2022) 1, 11, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.664, + 0.785, + 0.704 + ], + "angle": 0, + "content": "4. Jia, C., Yang, Y., Xia, Y., Chen, Y.T., Parekh, Z., Pham, H., Le, Q.V., Sung, Y., Li, Z., Duerig, T.: Scaling up visual and vision-language representation learning with noisy text supervision. In: ICML (2021) 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.705, + 0.785, + 0.731 + ], + "angle": 0, + "content": "5. Kaul, P., Xie, W., Zisserman, A.: Multi-modal classifiers for open-vocabulary object detection. arXiv preprint arXiv:2306.05493 (2023) 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.732, + 0.785, + 0.758 + ], + "angle": 0, + "content": "6. Kim, D., Angelova, A., Kuo, W.: Contrastive feature masking open-vocabulary vision transformer. ICCV (2023) 2, 6, 7, 10, 11, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.759, + 0.785, + 0.799 + ], + "angle": 0, + "content": "7. Kim, D., Angelova, A., Kuo, W.: Region-aware pretraining for open-vocabulary object detection with vision transformers. In: Conference on Computer Vision and Pattern Recognition (CVPR) (2023) 2, 3, 4, 5, 6, 7, 9, 10, 11, 12, 13, 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.8, + 0.785, + 0.839 + ], + "angle": 0, + "content": "8. Kuo, W., Cui, Y., Gu, X., Piergiovanni, A., Angelova, A.: F-vlm: Open-vocabulary object detection upon frozen vision and language models. ICLR (2023) 1, 3, 4, 6, 10, 11" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.173, + 0.785, + 0.839 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.362, + 0.127 + ], + "angle": 0, + "content": "D. Kim et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.785, + 0.189 + ], + "angle": 0, + "content": "19. Li, L.H., Zhang, P., Zhang, H., Yang, J., Li, C., Zhong, Y., Wang, L., Yuan, L., Zhang, L., Hwang, J.N., Chang, K.W., Gao, J.: Grounded language-image pretraining. In: CVPR (2022) 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.19, + 0.785, + 0.217 + ], + "angle": 0, + "content": "20. Li, Y., Mao, H., Girshick, R., He, K.: Exploring plain vision transformer backbones for object detection. In: ECCV (2022) 3, 4, 5, 6, 7, 8, 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.218, + 0.785, + 0.259 + ], + "angle": 0, + "content": "21. Lin, T.Y., Dollar, P., Girshick, R., He, K., Hariharan, B., Belongie, S.: Feature pyramid networks for object detection. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 2117-2125 (2017) 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.26, + 0.785, + 0.3 + ], + "angle": 0, + "content": "22. Liu, S., Zeng, Z., Ren, T., Li, F., Zhang, H., Yang, J., Li, C., Yang, J., Su, H., Zhu, J., et al.: Grounding dino: Marrying dino with grounded pre-training for open-set object detection. arXiv preprint arXiv:2303.05499 (2023) 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.301, + 0.785, + 0.34 + ], + "angle": 0, + "content": "23. Liu, Z., Lin, Y., Cao, Y., Hu, H., Wei, Y., Zhang, Z., Lin, S., Guo, B.: Swin transformer: Hierarchical vision transformer using shifted windows. CVPR (2021) 3, 8, 12, 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.342, + 0.785, + 0.383 + ], + "angle": 0, + "content": "24. Ma, C., Jiang, Y., Wen, X., Yuan, Z., Qi, X.: Codet: Co-occurrence guided region-word alignment for open-vocabulary object detection. Advances in Neural Information Processing Systems 36 (2024) 2, 4, 10, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.384, + 0.785, + 0.423 + ], + "angle": 0, + "content": "25. Minderer, M., Gritsenko, A., Houlsby, N.: Scaling open-vocabulary object detection. Advances in Neural Information Processing Systems 36 (2023) 2, 3, 4, 10, 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.425, + 0.785, + 0.48 + ], + "angle": 0, + "content": "26. Minderer, M., Gritsenko, A., Stone, A., Neumann, M., Weissenborn, D., Dosovitskiy, A., Mahendran, A., Arnab, A., Dehghani, M., Shen, Z., Wang, X., Zhai, X., Kipf, T., Houlsby, N.: Simple open-vocabulary object detection with vision transformers. In: ECCV (2022) 3, 4, 10, 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.481, + 0.785, + 0.508 + ], + "angle": 0, + "content": "27. Oord, A.v.d., Li, Y., Vinyals, O.: Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748 (2018) 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.509, + 0.785, + 0.55 + ], + "angle": 0, + "content": "28. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., Krueger, G., Sutskever, I.: Learning transferable visual models from natural language supervision. In: ICML (2021) 1, 3, 5, 8, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.551, + 0.785, + 0.591 + ], + "angle": 0, + "content": "29. Rasheed, H., Maaz, M., Khattak, M.U., Khan, S., Khan, F.S.: Bridging the gap between object and image-level representations for open-vocabulary detection. arXiv preprint arXiv:2207.03482 (2022) 1, 10, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.592, + 0.785, + 0.619 + ], + "angle": 0, + "content": "30. Ren, S., He, K., Girshick, R.B., Sun, J.: Faster r-cnn: Towards real-time object detection with region proposal networks. In: NeurIPS (2015) 5, 6, 7, 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.62, + 0.785, + 0.673 + ], + "angle": 0, + "content": "31. Schuhmann, C., Vencu, R., Beaumont, R., Kaczmarczyk, R., Mullis, C., Katta, A., Coombes, T., Jitsev, J., Komatsuzaki, A.: Laion-400m: Open dataset of clip-filtered 400 million image-text pairs. arXiv preprint arXiv:2111.02114 (2021) 3, 5, 9, 10, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.675, + 0.785, + 0.715 + ], + "angle": 0, + "content": "32. Shi, C., Yang, S.: Edadet: Open-vocabulary object detection using early dense alignment. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 15724-15734 (2023) 10, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.716, + 0.785, + 0.757 + ], + "angle": 0, + "content": "33. Wang, L., Liu, Y., Du, P., Ding, Z., Liao, Y., Qi, Q., Chen, B., Liu, S.: Object-aware distillation pyramid for open-vocabulary object detection. In: CVPR (2023) 10, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.758, + 0.785, + 0.785 + ], + "angle": 0, + "content": "34. Wei, F., Gao, Y., Wu, Z., Hu, H., Lin, S.: Aligning pretraining for detection via object-level contrastive learning. In: NeurIPS (2021) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.786, + 0.785, + 0.838 + ], + "angle": 0, + "content": "35. Wu, S., Zhang, W., Jin, S., Liu, W., Loy, C.C.: Aligning bag of regions for open-vocabulary object detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15254-15264 (2023) 2, 3, 10, 11, 12" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.838 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.229, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Region-centric Image-Language Pretraining for Open-Vocabulary Detection" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.788, + 0.203 + ], + "angle": 0, + "content": "36. Wu, X., Zhu, F., Zhao, R., Li, H.: Cora: Adapting clip for open-vocabulary detection with region prompting and anchor pre-matching. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7031-7040 (2023) 3, 10, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.204, + 0.787, + 0.231 + ], + "angle": 0, + "content": "37. Xiao, T., Reed, C.J., Wang, X., Keutzer, K., Darrell, T.: Region similarity representation learning. In: ICCV. pp. 10539-10548 (October 2021) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.232, + 0.787, + 0.286 + ], + "angle": 0, + "content": "38. Yao, L., Han, J., Liang, X., Xu, D., Zhang, W., Li, Z., Xu, H.: Detclipv2: Scalable open-vocabulary object detection pre-training via word-region alignment. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 23497-23506 (2023) 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.287, + 0.787, + 0.327 + ], + "angle": 0, + "content": "39. Yao, L., Han, J., Wen, Y., Liang, X., Xu, D., Zhang, W., Li, Z., Xu, C., , Xu, H.: Detclip: Dictionary-enriched visual-concept paralleled pre-training for open-world detection. In: arXiv:2209.09407 (2022) 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.328, + 0.787, + 0.356 + ], + "angle": 0, + "content": "40. Zang, Y., Li, W., Zhou, K., Huang, C., Loy, C.C.: Open-vocabulary DETR with conditional matching. In: ECCV. pp. 106-122. Springer (2022) 10, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.357, + 0.787, + 0.383 + ], + "angle": 0, + "content": "41. Zareian, A., Rosa, K.D., Hu, D.H., Chang, S.F.: Open-vocabulary object detection using captions. In: CVPR (2021) 1, 6, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.384, + 0.787, + 0.437 + ], + "angle": 0, + "content": "42. Zhang, H., Zhang, P., Hu, X., Chen, Y.C., Li, L., Dai, X., Wang, L., Yuan, L., Hwang, J.N., Gao, J.: Glipv2: Unifying localization and vision-language understanding. Advances in Neural Information Processing Systems 35, 36067-36080 (2022) 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.439, + 0.787, + 0.48 + ], + "angle": 0, + "content": "43. Zhao, S., Zhang, Z., Schulter, S., Zhao, L., Stathopoulos, A., Chandraker, M., Metaxas, D., et al.: Exploiting unlabeled data with vision and language models for object detection. In: ECCV (2022) 1, 10, 11, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.481, + 0.787, + 0.521 + ], + "angle": 0, + "content": "44. Zhong, Y., Yang, J., Zhang, P., Li, C., Codella, N., Li, L.H., Zhou, L., Dai, X., Yuan, L., Li, Y., Gao, J.: Regionclip: Region-based language-image pretraining. In: CVPR (2022) 1, 2, 3, 4, 10, 11, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.522, + 0.787, + 0.563 + ], + "angle": 0, + "content": "45. Zhong, Y., Wang, J., Wang, L., Peng, J., Wang, Y.X., Zhang, L.: Dap: Detection-aware pre-training with weak supervision. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 4537-4546 (2021) 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.564, + 0.787, + 0.604 + ], + "angle": 0, + "content": "46. Zhou, X., Girdhar, R., Joulin, A., Krahenbihl, P., Misra, I.: Detecting twenty-thousand classes using image-level supervision. In: ECCV (2022) 1, 2, 10, 11, 12" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.788, + 0.604 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/Region-centric Image-Language Pretraining for Open-Vocabulary Detection/be279718-23a8-4b0a-ac0c-2bca133ca6e7_origin.pdf b/2024/Region-centric Image-Language Pretraining for Open-Vocabulary Detection/be279718-23a8-4b0a-ac0c-2bca133ca6e7_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..583aac96d08f410154aba4e13131a9932a19c148 --- /dev/null +++ b/2024/Region-centric Image-Language Pretraining for Open-Vocabulary Detection/be279718-23a8-4b0a-ac0c-2bca133ca6e7_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8df85ce2e363acdd7fc0fc31e40f70326ba01db877917ac89726bb849d0a5b09 +size 4095678 diff --git a/2024/Region-centric Image-Language Pretraining for Open-Vocabulary Detection/full.md b/2024/Region-centric Image-Language Pretraining for Open-Vocabulary Detection/full.md new file mode 100644 index 0000000000000000000000000000000000000000..ec9e8820b89e74e1669c5dd03cfe07647de3d49b --- /dev/null +++ b/2024/Region-centric Image-Language Pretraining for Open-Vocabulary Detection/full.md @@ -0,0 +1,271 @@ +# Region-centric Image-Language Pretraining for Open-Vocabulary Detection + +Dahun Kim Anelia Angelova Weicheng Kuo + +Google DeepMind + +Abstract. We present a new open-vocabulary detection approach based on region-centric image-language pretraining to bridge the gap between image-level pretraining and open-vocabulary object detection. At the pretraining phase, we incorporate the detector architecture on top of the classification backbone, which better serves the region-level recognition needs of detection by enabling the detector heads to learn from large-scale image-text pairs. Using only standard contrastive loss and no pseudo-labeling, our approach is a simple yet effective extension of the contrastive learning method to learn emergent object-semantic cues. In addition, we propose a shifted-window learning approach upon window attention to make the backbone representation more robust, translation-invariant, and less biased by the window pattern. On the popular LVIS open-vocabulary detection benchmark, our approach sets a new state of the art of 37.6 mask $\mathrm{AP}_r$ using the common ViT-L backbone and public LAION dataset, and 40.5 mask $\mathrm{AP}_r$ using the DataComp-1B dataset, significantly outperforming the best existing approach by $+3.7$ mask $\mathrm{AP}_r$ . On the COCO benchmark, we achieve very competitive 39.6 novel AP without pseudo labeling or weak supervision. In addition, we evaluate our approach on the transfer detection setup, where it demonstrates notable improvement over the baseline. Visualization reveals emerging object locality from the pretraining recipes compared to the baseline. $^1$ + +# 1 Introduction + +Understanding and localizing all objects in the visual world is a key problem in computer vision, crucial for applications like self-driving cars. Traditional object detectors rely on costly human annotations, which limits scalability. Open-vocabulary detection (OVD) [41] addresses this by pretraining on large-scale image-text data (e.g. CLIP [28]), enabling detectors to identify objects based on user-provided text queries unavailable during training. + +Most existing open-vocabulary detection works assume the pretrained CLIP backbone is given, and focus on techniques such as knowledge distillation [5, 10], weak supervision [46], pseudo labeling [13, 29, 43, 44], and frozen backbone application [18], using the pretrained backbone. Consequently, during detection finetuning, the detector heads often need to be trained from scratch. This tends to result in sub-optimal generalization because the detector heads are trained on + +![](images/1bb6a16bc741c0f47086655e1b9ec48d486a66a2f29c88a22e6848703434a665.jpg) +Fig. 1: DITO overview (left): Existing image-text pretraining methods for open-vocabulary detection [16,17] update only the vision transformer backbone and finetune detector heads from scratch. We propose to pretrain both backbone and detector heads directly from the large-scale image-text paired data, without a need for pseudo labeling or box annotations. In open-vocabulary detection finetuning, we introduce a simple shifted-window learning method to produce more robust representations from the pretrained vision transformer. DITO prediction (right): LVIS results only show the novel categories (e.g., bulldozer, fishbowl, subwoofer, heron). While Ego4D is a real-world and out-of-distribution data, many unseen objects are detected (e.g., steel lid, sticker on the wall, recycle bin). Best viewed with zoom in. + +![](images/62a73a1fd804c918a98fc141e92f52ef8bb1637939eb2abae175fe7e9522da84.jpg) + +![](images/c27ddce40d35c2ede58ada8a9d98d6c2b2d4ad3852e169f6b5fe8c778d0492c4.jpg) + +the limited vocabulary of detection datasets, while only the backbone contains the knowledge of open-vocabulary concepts. + +Several studies have integrated detection models into CLIP pretraining. For instance, RegionCLIP [44] employs an off-the-shelf detector during CLIP pretraining to generate pseudo region-text labels. However, their pretraining only updates the image backbone, with the detector heads exclusively trained on detection data. These pseudo labeling-based pretraining methods [25, 44] require multi-stage training and handcrafted processing of pseudo labels, increasing complexity and training costs. Similarly, approaches such as GLIP [19, 42], Grounding DINO [22], CoDet [24] and DetCLIP [38, 39] integrate detector architectures in CLIP training, but their joint training demands additional detection and visual grounding datasets and complex multitask learning setups. + +In this paper, we propose a simple yet effective solution, Region-centric Pretraining approach, which incorporates detector modules into CLIP pretraining without the need for pseudo labeling or box annotations. This method involves generating random box regions across feature pyramid levels, followed by feature pooling over these regions. Subsequently, an image-text contrastive loss is applied, encouraging text-aligned region features to contribute more to the whole image representation. Our approach not only facilitates the warm-starting of detector heads in finetuning, but also leads to emergent representations with more localized semantic information compared to the baseline CLIP backbone, as demonstrated in our experiments. Compared to pseudo-labeling techniques [35,44-46], our approach can be viewed as an extension of contrastive + +models to bypass the need for offline object proposal generation, which could be cumbersome for large-scale image-text data. + +In addition, we focus on the Vision Transformer (ViT) based CLIP models. These CLIP ViTs excel in zero-shot recognition, but are less commonly integrated into open-vocabulary detectors compared to ConvNets [18, 36, 44]. Pretraining ViT backbones on lower resolutions and then adapting them to higher-resolution detection images poses challenges, including increased computational demands and potential disruption of locality in pretrained features. Although the windowed attention technique of ViTDet [17, 20] reduces computation and preserves locality, it introduces bias from fixed window patterns. To address this, we propose Shifted-Window Learning (SWL) approach to enhance information mixing across fixed windows and mitigate grid pattern bias. Unlike the Swin Transformer [23], which applies shifted window layer by layer, SWL applies shifted windowed attention as a separate forward pass using the same ViT backbone. This simple strategy enhances the windowed attention representation when applied low-res trained ViT backbone to high-res images, ensuring compatibility with vanilla ViT backbones pretrained without shifted windows. SWL also better preserves the open-vocabulary knowledge of pretrained features compared to full-attention ViT perhaps due to its emphasis on local cues. + +Incorporating region-centric pretraining and shifted-window learning, our approach DITO (Detection-aligned Image-Text pretraining for Open-vocabulary detection) serves to narrow the gap between image-text pretraining and open-vocabulary detection, improving generalization. The best DITO model achieves 37.6 mask $\mathrm{AP}_r$ on LVIS benchmark, surpassing the previous best by $+3.7\mathrm{AP}_r$ . It achieves the state-of-the-art 40.5 mask $\mathrm{AP}_r$ when pretrained on the DataComp1B dataset. On the COCO benchmark, DITO achieves a very competitive 39.6 novel AP without relying on pseudo-labels or joint training. + +# 2 Related Work + +Open-vocabulary detection. Conventional closed-set object detectors perform well but are limited by vocabulary size. Open-vocabulary detection, inspired by the strong zero-shot capabilities of Vision-Language Models (VLMs) like CLIP [28], has made notable progress. Various techniques have been explored, including knowledge distillation [10], prompt optimization [5], and the use of pretrained CLIP backbones with new detection heads [17,18,26,36]. Many top-performing methods [6,8,19,25,44] rely on pseudo labeling to prevent catastrophic forgetting during detection finetuning, but these methods often require multi-stage training [35,44] or extra steps to generate high-quality labels [25]. In contrast, our approach focuses on both upstream image-text pretraining and downstream open-vocabulary detection without using pseudo labeling, leading to significant improvements. + +Region-aligned Vision-Language Models. Building on advances in aligning image-text representations [28, 31], several studies have integrated region-level alignment into CLIP pretraining. For example, RegionCLIP [44] learns region-word alignment through pseudo region-text pairs but relies on an off- + +the-shelf detector for pseudo labeling and focuses only on training the image backbone. GLIP [19, 42], Grounding DINO [22], DetCLIP [38, 39], CoDet [24] integrate detector architectures into CLIP training to align regions with words, but they require additional visual grounding annotations, leading to complex multitask learning. To our knowledge, our approach is the first to incorporate detector modules in image-text pretraining without relying on box annotations. + +Vision Transformers in open-vocabulary detection. While ViT-based CLIP models excel in zero-shot recognition, their adaptation to open-vocabulary detection is less explored compared to ConvNet-based CLIP models. For instance, OWL-ViT [25, 26] finetunes ViT on higher-resolution detection images with global attentions, but the full attention is computationally intensive. ROViT [17] introduces region-aware positional embeddings to aid in detection finetuning and uses windowed attention from ViTDet [20]. Our paper proposes shifted-window learning to enhance information mixing across fixed windows and reduce window-induced bias while preserving the locality structure of pretrained features, thereby improving open-vocabulary detection. + +Self-supervised pretraining for visual tasks. Self-supervised learning has emerged as a promising paradigm to learn object features for complex visual tasks such as detection, given the challenge of scaling up human annotations. Most relevant direction is contrastive learning, where the contrastive samples can take the forms of augmented images [3], sliding windows [37], object proposals [34], or point samples [1]. Some alternative strategies like pseudo-labeling [45] and pixel reconstruction [11] have also proven effective. While the majority of these methods have focused on learning from images without textual context, and applying to closed-vocabulary detection, DITO leverages large image-text data to tackle the more demanding open-vocabulary detection task, without a need for offline proposal generation [34, 45]. + +# 3 Method + +We address the problem of open-vocabulary object detection. At training time, the model can access the class and box labels of base categories $(C_B)$ . At test time, the model is tasked with detecting objects from a set of novel categories $(C_N)$ not present in the training set. To achieve this, we leverage pretrained vision and language models (VLMs) as in prior studies [10, 17, 18, 44]. + +More specifically, we leverage recent advances in ViT-based detectors [17,20] for their promising results. However, instead of solely taking the pretrained ViT backbone, we demonstrate how to enhance the VLMs by integrating detector heads into the CLIP pretraining process, referred to as Region-centric image-text Pretraining (REP). Additionally, for detection finetuning, we propose a Shifted-Window Learning (SWL) to enhance the adaptation of the pretrained ViT to open-vocabulary detection task. By combining these approaches, DITO achieves significant improvements in open-vocabulary detection over prior arts. + +![](images/76ad26ff24ea4bd2aaa44277f85c2e510eecedb0d4d4f4f0427b90ddc48d92b8.jpg) +Fig. 2: DITO method. Region-centric image-text pretraining (left): We train the detector heads (e.g. FPN [20, 21], Faster RCNN head [30], and RPN [30]) upon a ViT encoder backbone with multi-level image-text contrastive loss to bridge the gap between image-text pretraining and open-vocabulary detection. Shifting-window learning for detection (right): We roll the image and combines the shifted features with the original features to mitigate the bias of windowed attention grid [20], and produce more robust semantic representation. + +![](images/f852f89776229d8b0874d93a5ec7b29fa46ed9f6108793b84d76a61a739017e4.jpg) + +# 3.1 Preliminaries + +Baseline. Our baseline approach is RO-ViT [17], a state-of-the-art ViT-based method for open-vocabulary detection. RO-ViT introduces CLIP pretraining with a new positional embedding scheme called cropped positional embedding (CPE). CPE involves randomly cropping and resizing the standard whole-image positional embedding during pretraining to enhance the ViT's generalization onto region-level recognition task and higher resolution detection inputs downstream. For detection finetuning, it adopts ViTDet [20] architecture, initialized with the pretrained ViT backbone. In the following, we describe the image-text pretraining and the downstream open-vocabulary detection. + +Image-text pretraining. We adopt dual-encoder CLIP pretraining widely used in existing works [28, 31]. The image embeddings $\{v\}$ and text embeddings $\{l\}$ are the average-pooled outputs from the image and text encoders, respectively. As in previous works, we compute the dot product of the embeddings in batch $B$ , and scale it by a learnable temperature $\tau$ before applying the InfoNCE loss [27, 28]. Mathematically, the image-to-text (I2T) loss can be expressed as: + +$$ +L _ {\mathrm {I 2 T}} = - \frac {1}{B} \sum_ {i = 1} ^ {B} \log \left(\frac {\exp \left(v _ {i} l _ {i} / \tau\right)}{\sum_ {j = 1} ^ {B} \exp \left(v _ {i} l _ {j} / \tau\right)}\right). \tag {1} +$$ + +The text-to-image (T2I) loss is symmetrical by exchanging the inner/outer summation loops. The total contrastive loss $L_{con}$ is obtained by $L_{con} = (L_{\mathrm{I2T}} + L_{\mathrm{T2I}}) / 2$ . As mentioned above, we adopt the cropped positional embeddings (CPE) following [17]. + +Open-vocabulary detection finetuning. At the fine-tuning stage, our detection finetuning recipe follows previous studies [10, 17, 18, 41]. During the training phase, we use the RoI-Align [12] feature as the detection embedding for each detected region. We replace the fixed-size classifier layer with the text embeddings of base categories $(C_B)$ . The detection score $p_i$ is determined by calculating the cosine similarity between the region embedding $r_i$ and text embeddings of base categories $(C_B)$ followed by a softmax operation. We prepend an additional background class embedding to $C_B$ and use the term "background" to represent the background category. Any proposals that do not match to any base category annotations are treated as background during training. It is important that the text embeddings are computed from the same text encoder as from the image-text pretraining. During testing, we expand the text embeddings to include the novel categories $(C_B \cup C_N)$ , resulting in $(C_B \cup C_N + 1)$ categories including the background. We calculate the detection scores $(p_i)$ as the cosine similarity between the region embeddings $(r_i)$ and the expanded text embeddings. Apart from the detection embedding $(r_i)$ , we extract the VLM embedding [18] of region $i$ by RoI-Align at the last ViT backbone feature map. The VLM score $(z_i)$ is calculated as the cosine similarity with the text embeddings of the combined categories $(C_B \cup C_N)$ . + +To compute the open-vocabulary detection score $(s_i^{\mathrm{ens}})$ , we ensemble the detection and VLM scores by geometric means [10,18]. The formula is as follows: + +$$ +s _ {i} ^ {\text {e n s}} = \left\{ \begin{array}{l l} z _ {i} ^ {(1 - \alpha)} \cdot p _ {i} ^ {\alpha} & \text {i f} i \in C _ {B} \\ z _ {i} ^ {(1 - \beta)} \cdot p _ {i} ^ {\beta} & \text {i f} i \in C _ {N} \end{array} \right. \tag {2} +$$ + +Here, $\alpha, \beta$ are floats $\in [0,1]$ that control the weighting of base versus novel categories. The background score comes from the detection score $(p_i)$ alone, because we observe the VLM score of "background" class is often less reliable. + +# 3.2 Region-centric Image-Text Pretraining + +Standard image-text pretraining uses classification architectures (e.g. ViT backbone followed by global pooling) as the language supervision occurs at the image level rather than the region level. Subsequently, for downstream detection, new detection heads are introduced and trained from scratch on a limited set of detection categories [16-18]. To fully utilize the knowledge embedded in large-scale image-text data, we propose Region-centric Pretraining (REP) which integrates the detector modules during the CLIP pretraining phase. Specifically, given access to image-text paired data but lacking box labels, our pretraining focuses on the region-recognition pathway of a detector, encompassing components like the backbone, FPN [20,21], RoI-Align [12], RPN-objectness [30], and Faster RCNN-classifier [30]. Consequently, the detector heads can be warm-started from the knowledge of large image-text data, thereby improving the generalization capability. To our knowledge, we are the first to integrate detector modules in image-text pretraining without box labels, and our experiments demonstrate clear benefits of our approach in open-vocabulary detection. + +Detector head learning from random regions. Fig. 2 (left) illustrates our region-centric pretraining system. Following existing works [16, 17], we adopt SimpleFPN [20] and Faster R-CNN [30] models to remove the detector differences and study the benefits of our region-centric pretraining. Specifically, the multilevel feature pyramid is computed from the ViT backbone. Then, RoI-Align [12] and Faster R-CNN classifier head are applied to these feature maps to match the classification pathway in pretraining with the region-recognition pathway in detection finetuning (see Table 4b for ablations). + +For each level $i$ of the feature pyramid, we randomly generate $n_i$ box regions uniformly over the image by sampling the box size $h, w \sim \text{Uniform}(0.2, 0.5)$ and aspect ratio $h / w \sim \text{Uniform}(0.5, 2.0)$ . The $n_i$ value is set proportional to the size of the $i$ -th feature map so that larger feature map would be covered by more regions. We extract the RoI-features of each region by RoI-Align, and feed them through the region classifier head [30] to obtain the RoI embeddings. + +Multi-level image-text supervision. After computing the RoI embeddings across pyramid levels, we perform a max pooling over the RoI embeddings per-level to obtain an image embedding for each pyramid level. Intuitively, max pooling allows the representation to focus on salient regions and discriminative features, thereby learning region-level information without explicit supervision. Then we apply the standard image-text contrastive loss (see Eqn. (1)) on each feature level separately, which aids the learning of rich semantic and object knowledge within every feature map (see Table 4c for ablations). The losses from all levels are weighted equally and summed together. Without explicit region-level supervision, the max pooling over regions encourages the more salient, text-aligned region features to contribute more to the whole image representation in the contrastive loss. Our experiments show emergent region-text alignments from the multi-level training, where the feature maps possess more localized semantic information compared to the baseline CLIP [17] backbone (see Fig. 3). + +Region proposal network learning. Fig. 3 shows that the learnt multi-level representations exhibit localized semantics well-aligned with the text query, i.e. the salient regions have higher similarity with the text relative to the background. Motivated by this, we employ the multi-level visual-text similarity as a supervisory signal for training the Region Proposal Network (RPN) [30]. Specifically, we compute the cosine similarity between each RoI embedding and the text, and use it as the target RPN score for the center location of each RoI. Any negative dot product value is mapped to zero to keep the target score in range [0,1]. We use L1 regression loss and set the loss weight equal to the multi-level contrastive loss (see supplementary). The losses are propagated only through the RoI centers and other pixels are ignored. Note that the box regression of the RPN is not trained here but learnt later through the detection finetuning, as we only use the image-text paired data without any box annotations. + +# 3.3 Shifted-Window Learning for Detection + +The CLIP ViT backbones are typically pretrained on lower resolutions (e.g. $224 \times 224$ ) and later adapted to higher resolution detection images (e.g. $1024 \times 1024$ ). + +Algorithm 1 Shifted Window backbone +x: image patch tokens + positional embeddings. M: attention window size. s = M // 2 $y =$ forward_vit_with_win_attn(x, M) $x^{\prime} =$ np-roll(x, shift=[s, s], axis=[0, 1]) # shift $y^\prime =$ forward_vit_with_win_attn $(x^{\prime},M)$ $y'' =$ np-roll(y', shift=[-s, -s], axis=[0, 1]) # unshift return $(y + y'') / 2$ # SWL backbone output + +While detection tasks benefit from global information, directly applying the pretrained ViT to high-resolution inputs is computationally intensive and can disrupt the locality structure of the pretrained lower-res features. Windowed attention in ViTDet [20] helps reduce computation but introduces bias from the fixed-size grid pattern. To address this, we propose the Shifted-Window Learning (SWL) approach, which improves information mixing across fixed windows and mitigates grid pattern bias. + +Network architecture. Fig. 2 (right) and Algorithm 1 describe the SWL algorithm. The standard ViT consists of a patchifying layer and a set of transformer layers. After feeding the image through the patchifying layer, we obtain a feature map $x$ of shape $(h, w, d)$ . This feature map $x$ is fed through the rest of the ViT with windowed attention layers on a grid $K \times K$ , and $L$ global attention layers evenly spaced throughout the ViT (where $L = 4$ ) following [20], resulting in output $y$ . In parallel, we create another copy of $x$ , which is rolled along both $h$ and $w$ axes by $s$ pixels. The elements that roll beyond the last position are reintroduced from the first. We carefully design the attention masks such that the rolled around patches would not attend to the patches on the other side of the image (see the right figure of Algorithm 1). The shift size $s$ is set as the half of the attention window size $M$ (i.e. $s = M // 2$ ). Empirically, the window size $M = 16$ equals the image size (e.g. 1024) divided by the product of patch size (e.g. $P = 16$ ) and the grid size (e.g. $K = 4$ ). The shifted feature map $x'$ is then processed through the rest of the ViT in the same manner, resulting in output $y'$ in the same shape $(h, w, d)$ as $y$ . We then unshift $y'$ and combine it with $y$ by averaging. We apply the above shifted window operations in detection finetuning and inference times. + +Comparison with Swin Transformer. Compared to the Swin Transformer [23], we apply the shifted-window ideas as separate forward passes, while Swin Transformer applies similar ideas in an alternating manner layer by layer. Our approach requires no change to the vanilla transformer architecture and is compatible with any ViT backbones pretrained without shifted windows (e.g. [28]), whereas Swin Transformer requires specialized pretraining on the same architecture. Compared to the full-attention ViT [4], we observe that shifted-window ViT taps more effectively into the semantic knowledge of pretrained backbone than full global attention, perhaps because the window helps the model focus more on local cues and ignore the noises farther away. + +# 3.4 Distillation from Frozen ViT Backbone + +While the ViT backbone adapts to the detection tasks, it may lose some of its pretrained open-vocabulary knowledge. Therefore, we propose a simple distillation approach which uses a separate frozen ViT to teach the finetuned ViT backbone during the detection finetuning. We use a cosine distance loss that aligns the RoI-Align embeddings extracted from the feature maps of both backbones. The cosine distance is computed for each RoI then averaged over all RoIs. Empirically, we find it advantageous to add a $1 \times 1$ Conv projection layer to the finetuned ViT backbone before the RoI-Align, allowing some flexibility in distillation when jointly trained with other detection objectives. The auxiliary distillation loss is then added to other detection losses, with a loss weight $\gamma = 1$ (see supplementary). At inference, the ViT backbone features after the projection are used to compute the region VLM score $z_{i}$ (Sec. 3.1). It is worth noting that the frozen ViT backbone is only used during training for distillation purposes and is removed at inference. + +While previous studies [2, 10] have utilized knowledge distillation from the CLIP models, their teacher CLIP models mostly operate on image crops in an offline process, thus needing multiple forward passes through the backbone for all RoIs. In contrast, our distillation operates efficiently on RoIs cropped from dense feature maps in a single forward pass during the detection finetuning with minimal overhead. + +# 4 Experimental Results + +Baseline reproduction. As discussed in Sec. 3.1, our baseline method is RO-ViT [17], a leading ViT-based approach for open-vocabulary detection. RO-ViT introduces cropped positional embedding (CPE) in CLIP pretraining to enhance the generalization of the pretrained ViT onto the downstream detection. Additionally, it adopts global average pooling of the ViT features instead of CLs-token pooling, which simplifies the adaptation onto higher resolution inputs, and the extraction of region features (e.g. RoI-Align) at the final layer. We reproduce the CLIP pretraining of RO-ViT [17] using the widely adopted LAION2B dataset [31], as the OpenAI CLIP's WIT [28] and ALIGN [14] datasets are not publicly available. We trained the CLIP models from scratch, following the same pretraining protocol and hyperparameters, including 500k iterations and 16k batch size (i.e. 8B samples seen in training), $224 \times 224$ image size, global average pooling, and cropped positional embedding. We use the standard InfoNCE loss (Eqn. (1)) instead of their focal constrastive loss [17]. The following compares our reproduced CLIP with OpenAI CLIP [28] and OpenCLIP [31]: + +
methoddataset#samples seenbackboneImageNet Top-1 Acc.
CLIP [28]WIT-400M13BViT-L/1475.5
OpenCLIP [31]LAION-2B32BViT-L/1475.2
RO-ViT CLIP [17] our repro.LAION-2B8BViT-L/1673.9
+ +On the zero-shot ImageNet classification benchmark, the LAION-2B pretraining matches or slightly underperforms the OpenAI CLIP with WIT-400M. + +
methodpretraining modelpretraining datadetector backbonew/ pseudo box labelsmask APrmask AP
ConvNet based:
OV-DETR [40]ViT-B/32CLIP-400MR-50-17.426.6
Kaul et al. [15]ViT-B/32CLIP-400MR-50-19.330.6
DetPro-Cascade [5]ViT-B/32CLIP-400MR-50-20.027.0
Rasheed [29]ViT-B/32CLIP-400MR-50-21.125.9
BARON [35]ViT-B/32CLIP-400MR-50-22.627.6
CoDet [24]R-50CLIP-400M + CC3MR-50-23.430.7
EdaDet [32]R-50CLIP-400MR-50-23.727.5
VL-PLM [43]ViT-B/32CLIP-400MR-50✓17.227.0
PromptDet [6]ViT-B/32CLIP-400MR-50✓21.425.3
OADB [33]ViT-B/32CLIP-400MR-50✓21.726.6
RegionCLIP [44]R-50x4CLIP-400M + CC3MR-50x4✓22.032.3
CORA [36]R-50x4CLIP-400MR-50x4✓22.2bx-
Detric-CN2 [46]ViT-B/32CLIP-400M + INet-21KR-50WS24.632.4
ViLD-Ens [10]EffNet-B7ALIGN-1.8BEffNet-B7-26.329.3
F-VLM [18]R-50x64CLIP-400MR-50x64-32.834.9
ViT based:
OWL-ViT [26]O+VGViT-L/14CLIP-400MViT-L/14-25.6bx34.7bx
OWL-ViT v2 [25]O+VGViT-L/14WebLI-10BViT-L/14✓45.9bx50.4bx
RO-ViT [17]ViT-B/16ALIGN-1.8BViT-B/16-28.030.2
RO-ViT [17]†ViT-L/16LAION-2BViT-L/16-32.432.9
CFM-ViT [16]ViT-B/16ALIGN-1.8BViT-B/16-28.832.0
CFM-ViT [16]ViT-L/16ALIGN-1.8BViT-L/16-33.936.6
CFM-ViT [16]*ViT-L/16LAION-2BViT-L/16-33.836.4
DITO (ours)ViT-S/16LAION-2BViT-S/16-26.228.8
DITO (ours)ViT-B/16LAION-2BViT-B/16-31.532.4
DITO (ours)ViT-L/16LAION-2BViT-L/16-37.636.2
DITO (ours)ViT-L/16DataComp-1BViT-L/16-40.538.0
+ +Table 1: LVIS open-vocabulary detection (mask AP). DITO outperforms the best existing approach by $+3.7$ mask $\mathrm{AP}_r$ . WS: uses weak supervision from ImageNet-21K. †: reports LAION-2B results in arXiv version. *: our reproduced result using LAION-2B. O+VG: uses extra Objects365 and Visual Genome data. + +Pretraining setup. After the above-mentioned baseline CLIP training, we apply our region-centric pretraining (REP) where we freeze the image and text encoders trained in the first phase and introduce the detector heads. We use the Simple FPN [20], and classification layers of Faster R-CNN and RPN [30], where we replace the batch normalization with layer normalization. At the $i$ -th pyramid level $i \in \{2, 3, 4, 5\}$ , we randomly sample $n_i \in \{400, 200, 100, 50\}$ box regions and compute their RoI-Align features. We use a short training cycle of $30k$ iterations, $4k$ batch size, $256 \times 256$ image size, AdamW optimizer with an initial learning rate of 1e-4 with linear decay. For both phases of image-text pretraining, we use the publicly available LAION-2B [31] dataset. + +Detection finetuning setup. As noted in Sec. 3.2, we adopt the ViTDet [20] with SimpleFPN as our detector following the baseline works [16,17]. We follow the same finetuning settings of [17]. Specifically, we train the detector with image size $1024 \times 1024$ and use windowed attention in the backbone with grid size $4 \times 4$ . The learning rate for the backbone is set lower as $0.6 \times$ of the detector head layers. We use $\alpha = 0.3$ , $\beta = 0.65$ for score combination in Eqn. (2). The text embedding of each category is calculated as the average over the CLIP prompt templates. We use the batch size 128, the SGD optimizer with momentum 0.9. The initial learning rate and iterations are set to 0.18 and $36.8\mathrm{k}$ for LVIS, and 0.02 and $11.3\mathrm{k}$ for COCO datasets. + +
methodpretraining modelpretraining datadetector backbonew/ pseudo box labelsnovel APAP
ConvNet based:
OVR-CNN [41]R-50CLIP-400M + COCO-CapR-50-22.839.9
ViLD [10]ViT-B/32CLIP-400MR-50-27.651.3
F-VLM [18]R-50CLIP-400MR-50-28.039.6
OV-DETR [40]ViT-B/32CLIP-400MR-50-29.452.7
CoDet [24]R-50CLIP-400M + COCO-CapR-50-30.646.6
PromptDet [6]ViT-B/32CLIP-400MR-50✓26.650.6
XPM [13]R-50CLIP-400MR-50✓27.041.2
OADB [33]ViT-B/32CLIP-400M + COCO-CapR-50✓30.047.2
VL-PLM [43]ViT-B/32CLIP-400MR-50✓34.453.5
RegionCLIP [44]R-50x4CLIP-400M + CC3M + COCO-CapR-50x4✓39.355.7
EdaDet [32]R-50CLIP-400MR-50✓40.252.5
CORA [36]R-50x4CLIP-400MR-50x4✓41.743.8
Detic-CN2 [46]ViT-B/32CLIP-400M + INet-21KR-50WS27.845.0
ViT based:
RO-ViT [17]ViT-B/16ALIGN-1.8BViT-B/16-30.241.5
RO-ViT [17]ViT-L/16ALIGN-1.8BViT-L/16-33.047.7
RO-ViT [17]*ViT-L/16LAION-2BViT-L/16-33.347.9
CFM-ViT [16]ViT-B/16ALIGN-1.8BViT-B/16-30.842.4
CFM-ViT [16]ViT-L/16ALIGN-1.8BViT-L/16-34.146.0
CFM-ViT [16]*ViT-L/16LAION-2BViT-L/16-34.346.4
DITO (ours)ViT-S/16LAION-2BViT-S/16-32.344.4
DITO (ours)ViT-B/16LAION-2BViT-B/16-36.648.8
DITO (ours)ViT-L/16LAION-2BViT-L/16-39.654.4
DITO (ours)ViT-L/16DataComp-1BViT-L/16-40.254.6
+ +Table 2: COCO open-vocabulary detection (box AP50). DITO demonstrates a very competitive novel category AP without using pseudo labeling or weak supervision (WS). *: our reproduced result using LAION-2B. + +
methodbackboneAP\( AP_{50} \)\( AP_{75} \)
Supervised [10]R-5025.638.628.0
ViLD [10]R-5011.818.212.6
DetPro [5]R-5012.118.812.9
BARON [35]R-5013.621.014.5
F-VLM [18]R-50x1616.225.317.5
F-VLM [18]R-50x6417.727.419.1
RO-ViT [17]ViT-L/1617.126.918.5
CFM-ViT [16]ViT-L/1618.728.920.3
DITO (ours)ViT-L/1620.031.821.5
+ +Table 3: Zero-shot transfer detection from LVISbase to Objects365 (box AP). All models are tested on Objects365 dataset following the setup of [10]. + +# 4.1 Main Results + +LVIS Benchmark. In Table 1, we report the comparison with existing methods on the challenging LVIS benchmark. The 'frequent' and 'common' classes of the dataset belong to the base categories $\mathrm{C}_B$ , and the 'rare' classes are the novel categories $\mathrm{C}_N$ . The primary metric is the mask AP on rare classes (mask $\mathrm{AP}_r$ ). The DITO model achieves the performance of 37.6 mask $\mathrm{AP}_r$ , which significantly outperforms the state-of-the-art approach RO-ViT [17] with the same ViT-L backbone by +5.1 points using the same pretraining data LAION-2B [31]. We also outperform the state-of-the-art CFM-ViT [16] by +3.6 points. Our best performance sets a new state-of-the-art 40.9 mask $\mathrm{AP}_r$ when using DataComp1B [7] in pretraining. With the ViT-B backbone, DITO maintains a healthy margin of around +2.5 $\mathrm{AP}_r$ above existing ViT-B based approaches. + +
methodAPrAP
RO-ViT [17]32.432.9
RO-ViT our repro.32.233.0
w/REP34.8 (+2.6)34.9
w/SWL35.0 (+2.8)35.2
w/REP+SWL36.3 (+4.1)35.8
w/REP+SWL+FD37.6 (+5.4)36.2
+ +(a) DITO framework. + +
pretraining methodAPr
RO-ViT our repro.32.2
w/FPN33.2
w/FPN+Head34.2
w/FPN+Head+RPN34.8
+ +(b) Detector components. + +
RoI embeddingAPr
global avg/lvl34.0
global max/lvl33.7
multi RoIs, avg/lvl33.8
multi RoIs, max/lvl34.8
multi RoIs, max all34.1
+ +(c) RoI embedding. +Table 4: Ablation on overall DITO framework and Region-centric Pretraining (Sec. 3.2). REP: Region-centric Pretraining, SWL: Shifted-Window Learning, FD: Frozen backbone Distillation. Best setting is in gray. + +
backboneAPrAP
fully global attn.33.433.8
baseline window attn. [17]32.233.0
Swin [23] style31.333.1
shifted window35.035.2
+ +(a) Shifted-window learning. + +
# global attn. layerbasew/ SWL
030.734.6 (+3.9)
432.235.0 (+2.8)
1232.434.2 (+1.8)
24 (all layers)33.433.4 (+0.0)
+ +(b) SWL w.r.t. # global attention layers. + +Table 5: Ablation on Shifted-Window Learning (SWL - Sec. 3.3). + +COCO Benchmark. We present the comparison on COCO benchmark in Table 2. The main metric is AP50 of novel categories (novel AP). Without using pseudo labeling [6,13,43,44], weak supervision [46], or externally trained detector modules [29, 35], our model demonstrates competitive results of 39.6 novel AP with LAION-2B and 40.2 with DataComp-1B. Among the ViT-based methods, DITO outperforms recent works RO-ViT [17] and CFM-ViT [16] by a clear margin of $+6.3$ and $+5.3$ points, respectively. + +Transfer detection. We further evaluate DITO in the transfer detection setting, where the open-vocabulary detector trained on one dataset $(\mathrm{LVIS}_{base})$ is tested on another dataset (Objects365) without any finetuning. By simply replacing the text embeddings, Table 3 shows that DITO achieves 20.0 AP, outperforming previous methods using ConvNet or ViT backbones of similar size. + +# 4.2 Ablation Studies + +For ablation study, we use the ViT-L/16 model pretrained with LAION-2B, and evaluate on the LVIS benchmark and report mask $\mathrm{AP}_r$ . + +DITO overall framework. Table 4a summarizes the benefits of each DITO components. The region-centric pretraining (REP) improves the contrastive model baseline by $+2.6\mathrm{AP}_r$ and shifted window learning (SWL) improves the baseline by $+2.8$ points. Combining both strategies brings a significant gain of $+4.1\mathrm{AP}_r$ . Lastly, we observe that incorporating the frozen backbone distillation leads to an additional boost of $+1.2$ points. In the following, we provide ablations for each components. + +Region-centric Pretraining. In Table 4b, we ablate our region-centric image-text pretraining by progressively adding the FPN, Faster R-CNN head, and RPN into the pretraining. Our 'baseline' is the contrastive image-text pretraining with cropped positional embedding [17]. On top of this, 'w/ FPN' introduces the FPN into the pretraining, where each pyramid level (whole image) map is mean-pooled + +
backboneAPAPrAPcAPf
GT boxes given (region classification):
before det. training48.756.851.542.1
after det. training53.754.253.753.5
after det. training w/FD54.857.754.353.2
+ +(a) GT region recognition. + +
methodAPrAP
RO-ViT our repro.32.233.0
RO-ViT our repro.+FD34.0 (+1.8)33.6
w/ REP+SWL36.335.8
w/ REP+SWL+FD37.6 (+1.3)36.2
+ +(b) Effect of FD on OVD. + +Table 6: Ablation on frozen backbone distillation (FD - Sec. 3.4). + +into an image embedding, followed by the image-text contrastive loss per level. It improves the baseline by $+1.0\mathrm{AP}_r$ . Adding the Faster R-CNN head further improves the alignment between the pretraining and detection finetuning, showing a gain of $+2.0\mathrm{AP}_r$ . Incorporating all components i.e. the FPN, Faster R-CNN and RPN heads achieves the best 34.8 $\mathrm{AP}_r$ , a significant gain of +2.6 points over the baseline. + +Table 4c shows that both global avg- and max-pooling are sub-optimal due to the lack of saliency map (avg-pool), and the limited capacity of a single pixel to represent semantic concepts for contrastive learning (max-pool), respectively. Our approach combines the best of both worlds, by first avg-pooling within each RoI, and then max-pooling over these RoI embeddings. Each embedding represents a proper RoI and the global representation captures the saliency map through max-pooling. Pooling per pyramid level (i.e. multi-level image-text supervision) outperforms pooling over all levels. + +![](images/b96117cfb9376fb716d03fff8f0e0ca5da34dadf5fd46874d5c062fe64de13ea.jpg) +"am riding a cart pulled by a donkey + +![](images/1e73822b46fb7338b5e7b76eded1625ba323c2d7f427c289a958c2135eaeff20.jpg) +"a crane operates amidst piles of rubble" + +![](images/10698e6b13c6dee5cf535e0541da93d88e221089a891af5d2144a501987044eb.jpg) +"aperson wearing blue rubber gloves is loading a syringe with vegetables" + +![](images/441f38d48142acc4a2426621641d897860befe2659f79a5f7497c3af68af3a3b.jpg) +"amotherand her two children sit on a green bench" +Fig. 3: Visual-text similarity map. For each example, we show the paired image (left) and text (bottom) input, and the visual-text similarity map using the contrastive model baseline [17] backbone features (middle) or our region-centric pretraining features (right). We use Flickr30K (top row) and COCO Captions (bottom row) datasets. + +![](images/12b3edad64336eddf3980b8fe35f84ac2e66a45fcd2d24a7f8bd8043510b9074.jpg) +"Cows on the street next to people that are on the sidewalk" + +![](images/1f9acc1833c0be07060bdfd5457cfb76accdc3fca7c2867d70a93a7c93765476.jpg) +"small white bird sits on the back of a horse in an open field" + +Shifted-Window Learning for detection. The CLIP ViT backbones are initially pretrained on lower resolutions and then adapted to higher resolution detection images. In Table 5a, we assess the efficacy of the shifted-window backbone in open-vocabulary detection training. Although the fully global attention model improves the detection task compared to the baseline windowed attention model [17], directly applying the pretrained ViT on high-resolution inputs may not be optimal, potentially compromising the locality structure of the pretrained lower-resolution features. Our shifted-window learning approach achieves $35.0\mathrm{AP}_r$ by preserving the locality structure from windowed attention as well as integrating information across fixed windows, outperforming the fully global + +attention model (33.4 $\mathrm{AP}_r$ ) which is computationally intensive. Notably, OWL-ViT [25,26] adopts the fully global attention model. In addition, naively applying the layer-alternating shifted window as in Swin [23] leads to a performance drop (see Sec. 3.3). Table 5b delves deeper into the behavior of SWL. The advantage of SWL diminishes steadily with increasing number of global attention layers in the windowed attention backbone, validating its better information mixing enabled by the SWL. + +Frozen backbone distillation. Table 6a studies the region classification capability of the CLIP ViT backbone before and after the detection finetuning. We use the ground truth boxes and measures AP scores to evaluate the zero-shot region classification of the base ('frequent' + 'common') and novel ('rare') categories. We observe that the finetuned backbone indeed shows a notable drop of $-2.6\mathrm{AP}_r$ on novel classes, while overfitting to the base classes. The frozen backbone distillation ('w/ FD') leads to a significant improvement of $+3.5\mathrm{AP}_r$ even surpassing the frozen pretrained backbone (before detection finetuning), while maintaining performance on the base classes ( $\mathrm{AP}_c$ and $\mathrm{AP}_f$ ). Table 6b presents the open-vocabulary detection results where the frozen backbone distillation brings a clear gain of $+1.3\sim 1.8\mathrm{AP}_r$ . These results highlight the efficacy of our frozen backbone distillation in detection finetuning (Sec. 3.4), as it effectively preserves the pretrained open-vocabulary knowledge while acquiring explicit region-text alignment through detection supervision. + +# 4.3 Visualization + +In Fig. 3, we visualize the similarity maps between image features and query text embeddings. We compare the baseline [17] contrastive model's backbone features with our region-centric pretrained features. We select pyramid level 4 which has the same resolution as the backbone features and apply the Faster R-CNN head in a sliding window manner to obtain the dense feature map. We observe that region-centric pretraining captures more localized semantic + +Fig. 1 visualizes DITO outputs on LVIS novel categories and real-world out-of-distribution data from Ego4D [9]. Using the same DITO detector trained on $\mathrm{LVIS}_{\text {base }}$ , we find that DITO effectively identifies many novel and unseen objects, even under significant domain shifts. + +# 5 Conclusion + +We introduce DITO, a region-centric approach for open-vocabulary detection using large-scale image-text pairs. By integrating detection architecture onto the image backbone in CLIP pretraining, it learns locality-sensitive information without requiring pseudo labeling or box annotations. Furthermore, we propose a shifted-window learning method to mitigate the bias of the window attention pattern in CLIP ViT detectors. Experiments show that DITO outperforms the state-of-the-art by large margins on the LVIS benchmark, and is very competitive on the COCO benchmark and transfer detection. We hope this work would inspire the community to explore region-centric image-language pretraining for open-vocabulary localization tasks. + +# References + +1. Bai, Y., Chen, X., Kirillov, A., Yuille, A., Berg, A.C.: Point-level region contrast for object detection pre-training. In: CVPR. pp. 16061-16070 (June 2022) 4 +2. Chen, J., Zhu, D., Qian, G., Ghanem, B., Yan, Z., Zhu, C., Xiao, F., Elhoseiny, M., Culatana, S.C.: Exploring open-vocabulary semantic segmentation without human labels. arXiv preprint arXiv:2306.00450 (2023) 9 +3. Chen, T., Kornblith, S., Norouzi, M., Hinton, G.: A simple framework for contrastive learning of visual representations. arXiv preprint arXiv:2002.05709 (2020) 4 +4. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929 (2020) 8 +5. Du, Y., Wei, F., Zhang, Z., Shi, M., Gao, Y., Li, G.: Learning to prompt for open-vocabulary object detection with vision-language model. In: CVPR (2022) 1, 3, 10, 11 +6. Feng, C., Zhong, Y., Jie, Z., Chu, X., Ren, H., Wei, X., Xie, W., Ma, L.: Promptdet: Towards open-vocabulary detection using uncurated images. In: European Conference on Computer Vision. pp. 701-717. Springer (2022) 3, 10, 11, 12 +7. Gadre, S.Y., Ilharco, G., Fang, A., Hayase, J., Smyrnis, G., Nguyen, T., Marten, R., Wortsman, M., Ghosh, D., Zhang, J., et al.: Datacomp: In search of the next generation of multimodal datasets. arXiv preprint arXiv:2304.14108 (2023) 11 +8. Gao, M., Xing, C., Niebles, J.C., Li, J., Xu, R., Liu, W., Xiong, C.: Open vocabulary object detection with pseudo bounding-box labels. In: ECCV (2022) 3 +9. Grauman, K., Westbury, A., Byrne, E., Chavis, Z., Furnari, A., Girdhar, R., Hamburger, J., Jiang, H., Liu, M., Liu, X., et al.: Ego4d: Around the world in 3,000 hours of egocentric video. In: CVPR. pp. 18995-19012 (2022) 14 +0. Gu, X., Lin, T.Y., Kuo, W., Cui, Y.: Open-vocabulary object detection via vision and language knowledge distillation. In: ICLR (2022) 1, 3, 4, 6, 9, 10, 11 +1. He, K., Chen, X., Xie, S., Li, Y., Dólar, P., Girshick, R.: Masked autoencoders are scalable vision learners. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16000-16009 (2022) 4 +2. He, K., Gkioxari, G., Dollár, P., Girshick, R.: Mask r-cnn. In: ICCV (2017) 6, 7 +3. Huynh, D., Kuen, J., Lin, Z., Gu, J., Elhamifar, E.: Open-vocabulary instance segmentation via robust cross-modal pseudo-labeling. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7020-7031 (2022) 1, 11, 12 +4. Jia, C., Yang, Y., Xia, Y., Chen, Y.T., Parekh, Z., Pham, H., Le, Q.V., Sung, Y., Li, Z., Duerig, T.: Scaling up visual and vision-language representation learning with noisy text supervision. In: ICML (2021) 9 +5. Kaul, P., Xie, W., Zisserman, A.: Multi-modal classifiers for open-vocabulary object detection. arXiv preprint arXiv:2306.05493 (2023) 10 +6. Kim, D., Angelova, A., Kuo, W.: Contrastive feature masking open-vocabulary vision transformer. ICCV (2023) 2, 6, 7, 10, 11, 12 +7. Kim, D., Angelova, A., Kuo, W.: Region-aware pretraining for open-vocabulary object detection with vision transformers. In: Conference on Computer Vision and Pattern Recognition (CVPR) (2023) 2, 3, 4, 5, 6, 7, 9, 10, 11, 12, 13, 14 +8. Kuo, W., Cui, Y., Gu, X., Piergiovanni, A., Angelova, A.: F-vlm: Open-vocabulary object detection upon frozen vision and language models. ICLR (2023) 1, 3, 4, 6, 10, 11 + +19. Li, L.H., Zhang, P., Zhang, H., Yang, J., Li, C., Zhong, Y., Wang, L., Yuan, L., Zhang, L., Hwang, J.N., Chang, K.W., Gao, J.: Grounded language-image pretraining. In: CVPR (2022) 2, 3, 4 +20. Li, Y., Mao, H., Girshick, R., He, K.: Exploring plain vision transformer backbones for object detection. In: ECCV (2022) 3, 4, 5, 6, 7, 8, 10 +21. Lin, T.Y., Dollar, P., Girshick, R., He, K., Hariharan, B., Belongie, S.: Feature pyramid networks for object detection. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 2117-2125 (2017) 5, 6 +22. Liu, S., Zeng, Z., Ren, T., Li, F., Zhang, H., Yang, J., Li, C., Yang, J., Su, H., Zhu, J., et al.: Grounding dino: Marrying dino with grounded pre-training for open-set object detection. arXiv preprint arXiv:2303.05499 (2023) 2, 4 +23. Liu, Z., Lin, Y., Cao, Y., Hu, H., Wei, Y., Zhang, Z., Lin, S., Guo, B.: Swin transformer: Hierarchical vision transformer using shifted windows. CVPR (2021) 3, 8, 12, 14 +24. Ma, C., Jiang, Y., Wen, X., Yuan, Z., Qi, X.: Codet: Co-occurrence guided region-word alignment for open-vocabulary object detection. Advances in Neural Information Processing Systems 36 (2024) 2, 4, 10, 11 +25. Minderer, M., Gritsenko, A., Houlsby, N.: Scaling open-vocabulary object detection. Advances in Neural Information Processing Systems 36 (2023) 2, 3, 4, 10, 14 +26. Minderer, M., Gritsenko, A., Stone, A., Neumann, M., Weissenborn, D., Dosovitskiy, A., Mahendran, A., Arnab, A., Dehghani, M., Shen, Z., Wang, X., Zhai, X., Kipf, T., Houlsby, N.: Simple open-vocabulary object detection with vision transformers. In: ECCV (2022) 3, 4, 10, 14 +27. Oord, A.v.d., Li, Y., Vinyals, O.: Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748 (2018) 5 +28. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., Krueger, G., Sutskever, I.: Learning transferable visual models from natural language supervision. In: ICML (2021) 1, 3, 5, 8, 9 +29. Rasheed, H., Maaz, M., Khattak, M.U., Khan, S., Khan, F.S.: Bridging the gap between object and image-level representations for open-vocabulary detection. arXiv preprint arXiv:2207.03482 (2022) 1, 10, 12 +30. Ren, S., He, K., Girshick, R.B., Sun, J.: Faster r-cnn: Towards real-time object detection with region proposal networks. In: NeurIPS (2015) 5, 6, 7, 10 +31. Schuhmann, C., Vencu, R., Beaumont, R., Kaczmarczyk, R., Mullis, C., Katta, A., Coombes, T., Jitsev, J., Komatsuzaki, A.: Laion-400m: Open dataset of clip-filtered 400 million image-text pairs. arXiv preprint arXiv:2111.02114 (2021) 3, 5, 9, 10, 11 +32. Shi, C., Yang, S.: Edadet: Open-vocabulary object detection using early dense alignment. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 15724-15734 (2023) 10, 11 +33. Wang, L., Liu, Y., Du, P., Ding, Z., Liao, Y., Qi, Q., Chen, B., Liu, S.: Object-aware distillation pyramid for open-vocabulary object detection. In: CVPR (2023) 10, 11 +34. Wei, F., Gao, Y., Wu, Z., Hu, H., Lin, S.: Aligning pretraining for detection via object-level contrastive learning. In: NeurIPS (2021) 4 +35. Wu, S., Zhang, W., Jin, S., Liu, W., Loy, C.C.: Aligning bag of regions for open-vocabulary object detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15254-15264 (2023) 2, 3, 10, 11, 12 + +36. Wu, X., Zhu, F., Zhao, R., Li, H.: Cora: Adapting clip for open-vocabulary detection with region prompting and anchor pre-matching. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7031-7040 (2023) 3, 10, 11 +37. Xiao, T., Reed, C.J., Wang, X., Keutzer, K., Darrell, T.: Region similarity representation learning. In: ICCV. pp. 10539-10548 (October 2021) 4 +38. Yao, L., Han, J., Liang, X., Xu, D., Zhang, W., Li, Z., Xu, H.: Detclipv2: Scalable open-vocabulary object detection pre-training via word-region alignment. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 23497-23506 (2023) 2, 4 +39. Yao, L., Han, J., Wen, Y., Liang, X., Xu, D., Zhang, W., Li, Z., Xu, C., , Xu, H.: Detclip: Dictionary-enriched visual-concept paralleled pre-training for open-world detection. In: arXiv:2209.09407 (2022) 2, 4 +40. Zang, Y., Li, W., Zhou, K., Huang, C., Loy, C.C.: Open-vocabulary DETR with conditional matching. In: ECCV. pp. 106-122. Springer (2022) 10, 11 +41. Zareian, A., Rosa, K.D., Hu, D.H., Chang, S.F.: Open-vocabulary object detection using captions. In: CVPR (2021) 1, 6, 11 +42. Zhang, H., Zhang, P., Hu, X., Chen, Y.C., Li, L., Dai, X., Wang, L., Yuan, L., Hwang, J.N., Gao, J.: Glipv2: Unifying localization and vision-language understanding. Advances in Neural Information Processing Systems 35, 36067-36080 (2022) 2, 4 +43. Zhao, S., Zhang, Z., Schulter, S., Zhao, L., Stathopoulos, A., Chandraker, M., Metaxas, D., et al.: Exploiting unlabeled data with vision and language models for object detection. In: ECCV (2022) 1, 10, 11, 12 +44. Zhong, Y., Yang, J., Zhang, P., Li, C., Codella, N., Li, L.H., Zhou, L., Dai, X., Yuan, L., Li, Y., Gao, J.: Regionclip: Region-based language-image pretraining. In: CVPR (2022) 1, 2, 3, 4, 10, 11, 12 +45. Zhong, Y., Wang, J., Wang, L., Peng, J., Wang, Y.X., Zhang, L.: Dap: Detection-aware pre-training with weak supervision. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 4537-4546 (2021) 2, 4 +46. Zhou, X., Girdhar, R., Joulin, A., Krahenbihl, P., Misra, I.: Detecting twenty-thousand classes using image-level supervision. In: ECCV (2022) 1, 2, 10, 11, 12 \ No newline at end of file diff --git a/2024/Region-centric Image-Language Pretraining for Open-Vocabulary Detection/images.zip b/2024/Region-centric Image-Language Pretraining for Open-Vocabulary Detection/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..eaf16dd601d551b38189761490f277ff215e6ae2 --- /dev/null +++ b/2024/Region-centric Image-Language Pretraining for Open-Vocabulary Detection/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:aa73e089d6671e578f40bf9b458db69fb24a0b53765ec3f33f57bbcd8b8b4da0 +size 544549 diff --git a/2024/Region-centric Image-Language Pretraining for Open-Vocabulary Detection/layout.json b/2024/Region-centric Image-Language Pretraining for Open-Vocabulary Detection/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..e20bc2bbd3a06aee6303fef1537e946cc26f2f80 --- /dev/null +++ b/2024/Region-centric Image-Language Pretraining for Open-Vocabulary Detection/layout.json @@ -0,0 +1,9187 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 143, + 111, + 473, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 111, + 473, + 148 + ], + "spans": [ + { + "bbox": [ + 143, + 111, + 473, + 148 + ], + "type": "text", + "content": "Region-centric Image-Language Pretraining for Open-Vocabulary Detection" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 200, + 168, + 414, + 181 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 200, + 168, + 414, + 181 + ], + "spans": [ + { + "bbox": [ + 200, + 168, + 414, + 181 + ], + "type": "text", + "content": "Dahun Kim Anelia Angelova Weicheng Kuo" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 269, + 190, + 344, + 201 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 269, + 190, + 344, + 201 + ], + "spans": [ + { + "bbox": [ + 269, + 190, + 344, + 201 + ], + "type": "text", + "content": "Google DeepMind" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 160, + 232, + 455, + 463 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 232, + 455, + 463 + ], + "spans": [ + { + "bbox": [ + 160, + 232, + 455, + 463 + ], + "type": "text", + "content": "Abstract. We present a new open-vocabulary detection approach based on region-centric image-language pretraining to bridge the gap between image-level pretraining and open-vocabulary object detection. At the pretraining phase, we incorporate the detector architecture on top of the classification backbone, which better serves the region-level recognition needs of detection by enabling the detector heads to learn from large-scale image-text pairs. Using only standard contrastive loss and no pseudo-labeling, our approach is a simple yet effective extension of the contrastive learning method to learn emergent object-semantic cues. In addition, we propose a shifted-window learning approach upon window attention to make the backbone representation more robust, translation-invariant, and less biased by the window pattern. On the popular LVIS open-vocabulary detection benchmark, our approach sets a new state of the art of 37.6 mask " + }, + { + "bbox": [ + 160, + 232, + 455, + 463 + ], + "type": "inline_equation", + "content": "\\mathrm{AP}_r" + }, + { + "bbox": [ + 160, + 232, + 455, + 463 + ], + "type": "text", + "content": " using the common ViT-L backbone and public LAION dataset, and 40.5 mask " + }, + { + "bbox": [ + 160, + 232, + 455, + 463 + ], + "type": "inline_equation", + "content": "\\mathrm{AP}_r" + }, + { + "bbox": [ + 160, + 232, + 455, + 463 + ], + "type": "text", + "content": " using the DataComp-1B dataset, significantly outperforming the best existing approach by " + }, + { + "bbox": [ + 160, + 232, + 455, + 463 + ], + "type": "inline_equation", + "content": "+3.7" + }, + { + "bbox": [ + 160, + 232, + 455, + 463 + ], + "type": "text", + "content": " mask " + }, + { + "bbox": [ + 160, + 232, + 455, + 463 + ], + "type": "inline_equation", + "content": "\\mathrm{AP}_r" + }, + { + "bbox": [ + 160, + 232, + 455, + 463 + ], + "type": "text", + "content": ". On the COCO benchmark, we achieve very competitive 39.6 novel AP without pseudo labeling or weak supervision. In addition, we evaluate our approach on the transfer detection setup, where it demonstrates notable improvement over the baseline. Visualization reveals emerging object locality from the pretraining recipes compared to the baseline." + }, + { + "bbox": [ + 160, + 232, + 455, + 463 + ], + "type": "inline_equation", + "content": "^1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 482, + 230, + 495 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 482, + 230, + 495 + ], + "spans": [ + { + "bbox": [ + 132, + 482, + 230, + 495 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 502, + 482, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 502, + 482, + 574 + ], + "spans": [ + { + "bbox": [ + 130, + 502, + 482, + 574 + ], + "type": "text", + "content": "Understanding and localizing all objects in the visual world is a key problem in computer vision, crucial for applications like self-driving cars. Traditional object detectors rely on costly human annotations, which limits scalability. Open-vocabulary detection (OVD) [41] addresses this by pretraining on large-scale image-text data (e.g. CLIP [28]), enabling detectors to identify objects based on user-provided text queries unavailable during training." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 574, + 482, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 574, + 482, + 647 + ], + "spans": [ + { + "bbox": [ + 130, + 574, + 482, + 647 + ], + "type": "text", + "content": "Most existing open-vocabulary detection works assume the pretrained CLIP backbone is given, and focus on techniques such as knowledge distillation [5, 10], weak supervision [46], pseudo labeling [13, 29, 43, 44], and frozen backbone application [18], using the pretrained backbone. Consequently, during detection finetuning, the detector heads often need to be trained from scratch. This tends to result in sub-optimal generalization because the detector heads are trained on" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 653, + 473, + 666 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 653, + 473, + 666 + ], + "spans": [ + { + "bbox": [ + 133, + 653, + 473, + 666 + ], + "type": "text", + "content": "1 project page: github.com/google-research/google-research/tree/master/fvlm/dito" + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 157, + 114, + 296, + 255 + ], + "blocks": [ + { + "bbox": [ + 157, + 114, + 296, + 255 + ], + "lines": [ + { + "bbox": [ + 157, + 114, + 296, + 255 + ], + "spans": [ + { + "bbox": [ + 157, + 114, + 296, + 255 + ], + "type": "image", + "image_path": "1bb6a16bc741c0f47086655e1b9ec48d486a66a2f29c88a22e6848703434a665.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 263, + 482, + 373 + ], + "lines": [ + { + "bbox": [ + 130, + 263, + 482, + 373 + ], + "spans": [ + { + "bbox": [ + 130, + 263, + 482, + 373 + ], + "type": "text", + "content": "Fig. 1: DITO overview (left): Existing image-text pretraining methods for open-vocabulary detection [16,17] update only the vision transformer backbone and finetune detector heads from scratch. We propose to pretrain both backbone and detector heads directly from the large-scale image-text paired data, without a need for pseudo labeling or box annotations. In open-vocabulary detection finetuning, we introduce a simple shifted-window learning method to produce more robust representations from the pretrained vision transformer. DITO prediction (right): LVIS results only show the novel categories (e.g., bulldozer, fishbowl, subwoofer, heron). While Ego4D is a real-world and out-of-distribution data, many unseen objects are detected (e.g., steel lid, sticker on the wall, recycle bin). Best viewed with zoom in." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 317, + 114, + 386, + 255 + ], + "blocks": [ + { + "bbox": [ + 317, + 114, + 386, + 255 + ], + "lines": [ + { + "bbox": [ + 317, + 114, + 386, + 255 + ], + "spans": [ + { + "bbox": [ + 317, + 114, + 386, + 255 + ], + "type": "image", + "image_path": "62a73a1fd804c918a98fc141e92f52ef8bb1637939eb2abae175fe7e9522da84.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 391, + 114, + 463, + 255 + ], + "blocks": [ + { + "bbox": [ + 391, + 114, + 463, + 255 + ], + "lines": [ + { + "bbox": [ + 391, + 114, + 463, + 255 + ], + "spans": [ + { + "bbox": [ + 391, + 114, + 463, + 255 + ], + "type": "image", + "image_path": "c27ddce40d35c2ede58ada8a9d98d6c2b2d4ad3852e169f6b5fe8c778d0492c4.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 388, + 481, + 412 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 388, + 481, + 412 + ], + "spans": [ + { + "bbox": [ + 130, + 388, + 481, + 412 + ], + "type": "text", + "content": "the limited vocabulary of detection datasets, while only the backbone contains the knowledge of open-vocabulary concepts." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 413, + 482, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 413, + 482, + 533 + ], + "spans": [ + { + "bbox": [ + 130, + 413, + 482, + 533 + ], + "type": "text", + "content": "Several studies have integrated detection models into CLIP pretraining. For instance, RegionCLIP [44] employs an off-the-shelf detector during CLIP pretraining to generate pseudo region-text labels. However, their pretraining only updates the image backbone, with the detector heads exclusively trained on detection data. These pseudo labeling-based pretraining methods [25, 44] require multi-stage training and handcrafted processing of pseudo labels, increasing complexity and training costs. Similarly, approaches such as GLIP [19, 42], Grounding DINO [22], CoDet [24] and DetCLIP [38, 39] integrate detector architectures in CLIP training, but their joint training demands additional detection and visual grounding datasets and complex multitask learning setups." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 534, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 534, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 534, + 482, + 666 + ], + "type": "text", + "content": "In this paper, we propose a simple yet effective solution, Region-centric Pretraining approach, which incorporates detector modules into CLIP pretraining without the need for pseudo labeling or box annotations. This method involves generating random box regions across feature pyramid levels, followed by feature pooling over these regions. Subsequently, an image-text contrastive loss is applied, encouraging text-aligned region features to contribute more to the whole image representation. Our approach not only facilitates the warm-starting of detector heads in finetuning, but also leads to emergent representations with more localized semantic information compared to the baseline CLIP backbone, as demonstrated in our experiments. Compared to pseudo-labeling techniques [35,44-46], our approach can be viewed as an extension of contrastive" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 222, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 222, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 222, + 101 + ], + "type": "text", + "content": "D. Kim et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": "models to bypass the need for offline object proposal generation, which could be cumbersome for large-scale image-text data." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 140, + 481, + 344 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 140, + 481, + 344 + ], + "spans": [ + { + "bbox": [ + 130, + 140, + 481, + 344 + ], + "type": "text", + "content": "In addition, we focus on the Vision Transformer (ViT) based CLIP models. These CLIP ViTs excel in zero-shot recognition, but are less commonly integrated into open-vocabulary detectors compared to ConvNets [18, 36, 44]. Pretraining ViT backbones on lower resolutions and then adapting them to higher-resolution detection images poses challenges, including increased computational demands and potential disruption of locality in pretrained features. Although the windowed attention technique of ViTDet [17, 20] reduces computation and preserves locality, it introduces bias from fixed window patterns. To address this, we propose Shifted-Window Learning (SWL) approach to enhance information mixing across fixed windows and mitigate grid pattern bias. Unlike the Swin Transformer [23], which applies shifted window layer by layer, SWL applies shifted windowed attention as a separate forward pass using the same ViT backbone. This simple strategy enhances the windowed attention representation when applied low-res trained ViT backbone to high-res images, ensuring compatibility with vanilla ViT backbones pretrained without shifted windows. SWL also better preserves the open-vocabulary knowledge of pretrained features compared to full-attention ViT perhaps due to its emphasis on local cues." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 344, + 482, + 440 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 344, + 482, + 440 + ], + "spans": [ + { + "bbox": [ + 130, + 344, + 482, + 440 + ], + "type": "text", + "content": "Incorporating region-centric pretraining and shifted-window learning, our approach DITO (Detection-aligned Image-Text pretraining for Open-vocabulary detection) serves to narrow the gap between image-text pretraining and open-vocabulary detection, improving generalization. The best DITO model achieves 37.6 mask " + }, + { + "bbox": [ + 130, + 344, + 482, + 440 + ], + "type": "inline_equation", + "content": "\\mathrm{AP}_r" + }, + { + "bbox": [ + 130, + 344, + 482, + 440 + ], + "type": "text", + "content": " on LVIS benchmark, surpassing the previous best by " + }, + { + "bbox": [ + 130, + 344, + 482, + 440 + ], + "type": "inline_equation", + "content": "+3.7\\mathrm{AP}_r" + }, + { + "bbox": [ + 130, + 344, + 482, + 440 + ], + "type": "text", + "content": ". It achieves the state-of-the-art 40.5 mask " + }, + { + "bbox": [ + 130, + 344, + 482, + 440 + ], + "type": "inline_equation", + "content": "\\mathrm{AP}_r" + }, + { + "bbox": [ + 130, + 344, + 482, + 440 + ], + "type": "text", + "content": " when pretrained on the DataComp1B dataset. On the COCO benchmark, DITO achieves a very competitive 39.6 novel AP without relying on pseudo-labels or joint training." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 447, + 237, + 460 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 447, + 237, + 460 + ], + "spans": [ + { + "bbox": [ + 131, + 447, + 237, + 460 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 471, + 482, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 471, + 482, + 615 + ], + "spans": [ + { + "bbox": [ + 130, + 471, + 482, + 615 + ], + "type": "text", + "content": "Open-vocabulary detection. Conventional closed-set object detectors perform well but are limited by vocabulary size. Open-vocabulary detection, inspired by the strong zero-shot capabilities of Vision-Language Models (VLMs) like CLIP [28], has made notable progress. Various techniques have been explored, including knowledge distillation [10], prompt optimization [5], and the use of pretrained CLIP backbones with new detection heads [17,18,26,36]. Many top-performing methods [6,8,19,25,44] rely on pseudo labeling to prevent catastrophic forgetting during detection finetuning, but these methods often require multi-stage training [35,44] or extra steps to generate high-quality labels [25]. In contrast, our approach focuses on both upstream image-text pretraining and downstream open-vocabulary detection without using pseudo labeling, leading to significant improvements." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "type": "text", + "content": "Region-aligned Vision-Language Models. Building on advances in aligning image-text representations [28, 31], several studies have integrated region-level alignment into CLIP pretraining. For example, RegionCLIP [44] learns region-word alignment through pseudo region-text pairs but relies on an off-" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 140, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 140, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 140, + 91, + 447, + 102 + ], + "type": "text", + "content": "Region-centric Image-Language Pretraining for Open-Vocabulary Detection" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "text", + "content": "the-shelf detector for pseudo labeling and focuses only on training the image backbone. GLIP [19, 42], Grounding DINO [22], DetCLIP [38, 39], CoDet [24] integrate detector architectures into CLIP training to align regions with words, but they require additional visual grounding annotations, leading to complex multitask learning. To our knowledge, our approach is the first to incorporate detector modules in image-text pretraining without relying on box annotations." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 193, + 483, + 314 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 193, + 483, + 314 + ], + "spans": [ + { + "bbox": [ + 130, + 193, + 483, + 314 + ], + "type": "text", + "content": "Vision Transformers in open-vocabulary detection. While ViT-based CLIP models excel in zero-shot recognition, their adaptation to open-vocabulary detection is less explored compared to ConvNet-based CLIP models. For instance, OWL-ViT [25, 26] finetunes ViT on higher-resolution detection images with global attentions, but the full attention is computationally intensive. ROViT [17] introduces region-aware positional embeddings to aid in detection finetuning and uses windowed attention from ViTDet [20]. Our paper proposes shifted-window learning to enhance information mixing across fixed windows and reduce window-induced bias while preserving the locality structure of pretrained features, thereby improving open-vocabulary detection." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 318, + 483, + 451 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 318, + 483, + 451 + ], + "spans": [ + { + "bbox": [ + 130, + 318, + 483, + 451 + ], + "type": "text", + "content": "Self-supervised pretraining for visual tasks. Self-supervised learning has emerged as a promising paradigm to learn object features for complex visual tasks such as detection, given the challenge of scaling up human annotations. Most relevant direction is contrastive learning, where the contrastive samples can take the forms of augmented images [3], sliding windows [37], object proposals [34], or point samples [1]. Some alternative strategies like pseudo-labeling [45] and pixel reconstruction [11] have also proven effective. While the majority of these methods have focused on learning from images without textual context, and applying to closed-vocabulary detection, DITO leverages large image-text data to tackle the more demanding open-vocabulary detection task, without a need for offline proposal generation [34, 45]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 475, + 203, + 488 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 475, + 203, + 488 + ], + "spans": [ + { + "bbox": [ + 132, + 475, + 203, + 488 + ], + "type": "text", + "content": "3 Method" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 507, + 482, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 507, + 482, + 568 + ], + "spans": [ + { + "bbox": [ + 130, + 507, + 482, + 568 + ], + "type": "text", + "content": "We address the problem of open-vocabulary object detection. At training time, the model can access the class and box labels of base categories " + }, + { + "bbox": [ + 130, + 507, + 482, + 568 + ], + "type": "inline_equation", + "content": "(C_B)" + }, + { + "bbox": [ + 130, + 507, + 482, + 568 + ], + "type": "text", + "content": ". At test time, the model is tasked with detecting objects from a set of novel categories " + }, + { + "bbox": [ + 130, + 507, + 482, + 568 + ], + "type": "inline_equation", + "content": "(C_N)" + }, + { + "bbox": [ + 130, + 507, + 482, + 568 + ], + "type": "text", + "content": " not present in the training set. To achieve this, we leverage pretrained vision and language models (VLMs) as in prior studies [10, 17, 18, 44]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": "More specifically, we leverage recent advances in ViT-based detectors [17,20] for their promising results. However, instead of solely taking the pretrained ViT backbone, we demonstrate how to enhance the VLMs by integrating detector heads into the CLIP pretraining process, referred to as Region-centric image-text Pretraining (REP). Additionally, for detection finetuning, we propose a Shifted-Window Learning (SWL) to enhance the adaptation of the pretrained ViT to open-vocabulary detection task. By combining these approaches, DITO achieves significant improvements in open-vocabulary detection over prior arts." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 222, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 222, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 222, + 101 + ], + "type": "text", + "content": "D. Kim et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 136, + 115, + 372, + 277 + ], + "blocks": [ + { + "bbox": [ + 136, + 115, + 372, + 277 + ], + "lines": [ + { + "bbox": [ + 136, + 115, + 372, + 277 + ], + "spans": [ + { + "bbox": [ + 136, + 115, + 372, + 277 + ], + "type": "image", + "image_path": "76ad26ff24ea4bd2aaa44277f85c2e510eecedb0d4d4f4f0427b90ddc48d92b8.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 284, + 482, + 361 + ], + "lines": [ + { + "bbox": [ + 130, + 284, + 482, + 361 + ], + "spans": [ + { + "bbox": [ + 130, + 284, + 482, + 361 + ], + "type": "text", + "content": "Fig. 2: DITO method. Region-centric image-text pretraining (left): We train the detector heads (e.g. FPN [20, 21], Faster RCNN head [30], and RPN [30]) upon a ViT encoder backbone with multi-level image-text contrastive loss to bridge the gap between image-text pretraining and open-vocabulary detection. Shifting-window learning for detection (right): We roll the image and combines the shifted features with the original features to mitigate the bias of windowed attention grid [20], and produce more robust semantic representation." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 378, + 121, + 471, + 275 + ], + "blocks": [ + { + "bbox": [ + 378, + 121, + 471, + 275 + ], + "lines": [ + { + "bbox": [ + 378, + 121, + 471, + 275 + ], + "spans": [ + { + "bbox": [ + 378, + 121, + 471, + 275 + ], + "type": "image", + "image_path": "f852f89776229d8b0874d93a5ec7b29fa46ed9f6108793b84d76a61a739017e4.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 370, + 228, + 381 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 370, + 228, + 381 + ], + "spans": [ + { + "bbox": [ + 131, + 370, + 228, + 381 + ], + "type": "text", + "content": "3.1 Preliminaries" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 386, + 482, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 386, + 482, + 495 + ], + "spans": [ + { + "bbox": [ + 130, + 386, + 482, + 495 + ], + "type": "text", + "content": "Baseline. Our baseline approach is RO-ViT [17], a state-of-the-art ViT-based method for open-vocabulary detection. RO-ViT introduces CLIP pretraining with a new positional embedding scheme called cropped positional embedding (CPE). CPE involves randomly cropping and resizing the standard whole-image positional embedding during pretraining to enhance the ViT's generalization onto region-level recognition task and higher resolution detection inputs downstream. For detection finetuning, it adopts ViTDet [20] architecture, initialized with the pretrained ViT backbone. In the following, we describe the image-text pretraining and the downstream open-vocabulary detection." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 497, + 482, + 570 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 497, + 482, + 570 + ], + "spans": [ + { + "bbox": [ + 131, + 497, + 482, + 570 + ], + "type": "text", + "content": "Image-text pretraining. We adopt dual-encoder CLIP pretraining widely used in existing works [28, 31]. The image embeddings " + }, + { + "bbox": [ + 131, + 497, + 482, + 570 + ], + "type": "inline_equation", + "content": "\\{v\\}" + }, + { + "bbox": [ + 131, + 497, + 482, + 570 + ], + "type": "text", + "content": " and text embeddings " + }, + { + "bbox": [ + 131, + 497, + 482, + 570 + ], + "type": "inline_equation", + "content": "\\{l\\}" + }, + { + "bbox": [ + 131, + 497, + 482, + 570 + ], + "type": "text", + "content": " are the average-pooled outputs from the image and text encoders, respectively. As in previous works, we compute the dot product of the embeddings in batch " + }, + { + "bbox": [ + 131, + 497, + 482, + 570 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 131, + 497, + 482, + 570 + ], + "type": "text", + "content": ", and scale it by a learnable temperature " + }, + { + "bbox": [ + 131, + 497, + 482, + 570 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 131, + 497, + 482, + 570 + ], + "type": "text", + "content": " before applying the InfoNCE loss [27, 28]. Mathematically, the image-to-text (I2T) loss can be expressed as:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 219, + 577, + 481, + 611 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 219, + 577, + 481, + 611 + ], + "spans": [ + { + "bbox": [ + 219, + 577, + 481, + 611 + ], + "type": "interline_equation", + "content": "L _ {\\mathrm {I 2 T}} = - \\frac {1}{B} \\sum_ {i = 1} ^ {B} \\log \\left(\\frac {\\exp \\left(v _ {i} l _ {i} / \\tau\\right)}{\\sum_ {j = 1} ^ {B} \\exp \\left(v _ {i} l _ {j} / \\tau\\right)}\\right). \\tag {1}", + "image_path": "ded2f916ced546246b430c60d188087272e5c8f0da573c5cec9825368ae2abdb.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 617, + 482, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 617, + 482, + 667 + ], + "spans": [ + { + "bbox": [ + 131, + 617, + 482, + 667 + ], + "type": "text", + "content": "The text-to-image (T2I) loss is symmetrical by exchanging the inner/outer summation loops. The total contrastive loss " + }, + { + "bbox": [ + 131, + 617, + 482, + 667 + ], + "type": "inline_equation", + "content": "L_{con}" + }, + { + "bbox": [ + 131, + 617, + 482, + 667 + ], + "type": "text", + "content": " is obtained by " + }, + { + "bbox": [ + 131, + 617, + 482, + 667 + ], + "type": "inline_equation", + "content": "L_{con} = (L_{\\mathrm{I2T}} + L_{\\mathrm{T2I}}) / 2" + }, + { + "bbox": [ + 131, + 617, + 482, + 667 + ], + "type": "text", + "content": ". As mentioned above, we adopt the cropped positional embeddings (CPE) following [17]." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 140, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 140, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 140, + 91, + 447, + 102 + ], + "type": "text", + "content": "Region-centric Image-Language Pretraining for Open-Vocabulary Detection" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 479, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 342 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 342 + ], + "type": "text", + "content": "Open-vocabulary detection finetuning. At the fine-tuning stage, our detection finetuning recipe follows previous studies [10, 17, 18, 41]. During the training phase, we use the RoI-Align [12] feature as the detection embedding for each detected region. We replace the fixed-size classifier layer with the text embeddings of base categories " + }, + { + "bbox": [ + 130, + 116, + 482, + 342 + ], + "type": "inline_equation", + "content": "(C_B)" + }, + { + "bbox": [ + 130, + 116, + 482, + 342 + ], + "type": "text", + "content": ". The detection score " + }, + { + "bbox": [ + 130, + 116, + 482, + 342 + ], + "type": "inline_equation", + "content": "p_i" + }, + { + "bbox": [ + 130, + 116, + 482, + 342 + ], + "type": "text", + "content": " is determined by calculating the cosine similarity between the region embedding " + }, + { + "bbox": [ + 130, + 116, + 482, + 342 + ], + "type": "inline_equation", + "content": "r_i" + }, + { + "bbox": [ + 130, + 116, + 482, + 342 + ], + "type": "text", + "content": " and text embeddings of base categories " + }, + { + "bbox": [ + 130, + 116, + 482, + 342 + ], + "type": "inline_equation", + "content": "(C_B)" + }, + { + "bbox": [ + 130, + 116, + 482, + 342 + ], + "type": "text", + "content": " followed by a softmax operation. We prepend an additional background class embedding to " + }, + { + "bbox": [ + 130, + 116, + 482, + 342 + ], + "type": "inline_equation", + "content": "C_B" + }, + { + "bbox": [ + 130, + 116, + 482, + 342 + ], + "type": "text", + "content": " and use the term \"background\" to represent the background category. Any proposals that do not match to any base category annotations are treated as background during training. It is important that the text embeddings are computed from the same text encoder as from the image-text pretraining. During testing, we expand the text embeddings to include the novel categories " + }, + { + "bbox": [ + 130, + 116, + 482, + 342 + ], + "type": "inline_equation", + "content": "(C_B \\cup C_N)" + }, + { + "bbox": [ + 130, + 116, + 482, + 342 + ], + "type": "text", + "content": ", resulting in " + }, + { + "bbox": [ + 130, + 116, + 482, + 342 + ], + "type": "inline_equation", + "content": "(C_B \\cup C_N + 1)" + }, + { + "bbox": [ + 130, + 116, + 482, + 342 + ], + "type": "text", + "content": " categories including the background. We calculate the detection scores " + }, + { + "bbox": [ + 130, + 116, + 482, + 342 + ], + "type": "inline_equation", + "content": "(p_i)" + }, + { + "bbox": [ + 130, + 116, + 482, + 342 + ], + "type": "text", + "content": " as the cosine similarity between the region embeddings " + }, + { + "bbox": [ + 130, + 116, + 482, + 342 + ], + "type": "inline_equation", + "content": "(r_i)" + }, + { + "bbox": [ + 130, + 116, + 482, + 342 + ], + "type": "text", + "content": " and the expanded text embeddings. Apart from the detection embedding " + }, + { + "bbox": [ + 130, + 116, + 482, + 342 + ], + "type": "inline_equation", + "content": "(r_i)" + }, + { + "bbox": [ + 130, + 116, + 482, + 342 + ], + "type": "text", + "content": ", we extract the VLM embedding [18] of region " + }, + { + "bbox": [ + 130, + 116, + 482, + 342 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 116, + 482, + 342 + ], + "type": "text", + "content": " by RoI-Align at the last ViT backbone feature map. The VLM score " + }, + { + "bbox": [ + 130, + 116, + 482, + 342 + ], + "type": "inline_equation", + "content": "(z_i)" + }, + { + "bbox": [ + 130, + 116, + 482, + 342 + ], + "type": "text", + "content": " is calculated as the cosine similarity with the text embeddings of the combined categories " + }, + { + "bbox": [ + 130, + 116, + 482, + 342 + ], + "type": "inline_equation", + "content": "(C_B \\cup C_N)" + }, + { + "bbox": [ + 130, + 116, + 482, + 342 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 343, + 482, + 368 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 343, + 482, + 368 + ], + "spans": [ + { + "bbox": [ + 130, + 343, + 482, + 368 + ], + "type": "text", + "content": "To compute the open-vocabulary detection score " + }, + { + "bbox": [ + 130, + 343, + 482, + 368 + ], + "type": "inline_equation", + "content": "(s_i^{\\mathrm{ens}})" + }, + { + "bbox": [ + 130, + 343, + 482, + 368 + ], + "type": "text", + "content": ", we ensemble the detection and VLM scores by geometric means [10,18]. The formula is as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 236, + 376, + 481, + 409 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 236, + 376, + 481, + 409 + ], + "spans": [ + { + "bbox": [ + 236, + 376, + 481, + 409 + ], + "type": "interline_equation", + "content": "s _ {i} ^ {\\text {e n s}} = \\left\\{ \\begin{array}{l l} z _ {i} ^ {(1 - \\alpha)} \\cdot p _ {i} ^ {\\alpha} & \\text {i f} i \\in C _ {B} \\\\ z _ {i} ^ {(1 - \\beta)} \\cdot p _ {i} ^ {\\beta} & \\text {i f} i \\in C _ {N} \\end{array} \\right. \\tag {2}", + "image_path": "77078872e53f86bd07d4408f618f5e6d863d29e517e910bfc82cd2ffdf101be4.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 415, + 482, + 451 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 415, + 482, + 451 + ], + "spans": [ + { + "bbox": [ + 130, + 415, + 482, + 451 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 130, + 415, + 482, + 451 + ], + "type": "inline_equation", + "content": "\\alpha, \\beta" + }, + { + "bbox": [ + 130, + 415, + 482, + 451 + ], + "type": "text", + "content": " are floats " + }, + { + "bbox": [ + 130, + 415, + 482, + 451 + ], + "type": "inline_equation", + "content": "\\in [0,1]" + }, + { + "bbox": [ + 130, + 415, + 482, + 451 + ], + "type": "text", + "content": " that control the weighting of base versus novel categories. The background score comes from the detection score " + }, + { + "bbox": [ + 130, + 415, + 482, + 451 + ], + "type": "inline_equation", + "content": "(p_i)" + }, + { + "bbox": [ + 130, + 415, + 482, + 451 + ], + "type": "text", + "content": " alone, because we observe the VLM score of \"background\" class is often less reliable." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 464, + 356, + 477 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 464, + 356, + 477 + ], + "spans": [ + { + "bbox": [ + 131, + 464, + 356, + 477 + ], + "type": "text", + "content": "3.2 Region-centric Image-Text Pretraining" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "type": "text", + "content": "Standard image-text pretraining uses classification architectures (e.g. ViT backbone followed by global pooling) as the language supervision occurs at the image level rather than the region level. Subsequently, for downstream detection, new detection heads are introduced and trained from scratch on a limited set of detection categories [16-18]. To fully utilize the knowledge embedded in large-scale image-text data, we propose Region-centric Pretraining (REP) which integrates the detector modules during the CLIP pretraining phase. Specifically, given access to image-text paired data but lacking box labels, our pretraining focuses on the region-recognition pathway of a detector, encompassing components like the backbone, FPN [20,21], RoI-Align [12], RPN-objectness [30], and Faster RCNN-classifier [30]. Consequently, the detector heads can be warm-started from the knowledge of large image-text data, thereby improving the generalization capability. To our knowledge, we are the first to integrate detector modules in image-text pretraining without box labels, and our experiments demonstrate clear benefits of our approach in open-vocabulary detection." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 222, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 222, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 222, + 101 + ], + "type": "text", + "content": "D. Kim et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 209 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 209 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 209 + ], + "type": "text", + "content": "Detector head learning from random regions. Fig. 2 (left) illustrates our region-centric pretraining system. Following existing works [16, 17], we adopt SimpleFPN [20] and Faster R-CNN [30] models to remove the detector differences and study the benefits of our region-centric pretraining. Specifically, the multilevel feature pyramid is computed from the ViT backbone. Then, RoI-Align [12] and Faster R-CNN classifier head are applied to these feature maps to match the classification pathway in pretraining with the region-recognition pathway in detection finetuning (see Table 4b for ablations)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 212, + 480, + 283 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 212, + 480, + 283 + ], + "spans": [ + { + "bbox": [ + 130, + 212, + 480, + 283 + ], + "type": "text", + "content": "For each level " + }, + { + "bbox": [ + 130, + 212, + 480, + 283 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 212, + 480, + 283 + ], + "type": "text", + "content": " of the feature pyramid, we randomly generate " + }, + { + "bbox": [ + 130, + 212, + 480, + 283 + ], + "type": "inline_equation", + "content": "n_i" + }, + { + "bbox": [ + 130, + 212, + 480, + 283 + ], + "type": "text", + "content": " box regions uniformly over the image by sampling the box size " + }, + { + "bbox": [ + 130, + 212, + 480, + 283 + ], + "type": "inline_equation", + "content": "h, w \\sim \\text{Uniform}(0.2, 0.5)" + }, + { + "bbox": [ + 130, + 212, + 480, + 283 + ], + "type": "text", + "content": " and aspect ratio " + }, + { + "bbox": [ + 130, + 212, + 480, + 283 + ], + "type": "inline_equation", + "content": "h / w \\sim \\text{Uniform}(0.5, 2.0)" + }, + { + "bbox": [ + 130, + 212, + 480, + 283 + ], + "type": "text", + "content": ". The " + }, + { + "bbox": [ + 130, + 212, + 480, + 283 + ], + "type": "inline_equation", + "content": "n_i" + }, + { + "bbox": [ + 130, + 212, + 480, + 283 + ], + "type": "text", + "content": " value is set proportional to the size of the " + }, + { + "bbox": [ + 130, + 212, + 480, + 283 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 212, + 480, + 283 + ], + "type": "text", + "content": "-th feature map so that larger feature map would be covered by more regions. We extract the RoI-features of each region by RoI-Align, and feed them through the region classifier head [30] to obtain the RoI embeddings." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 286, + 479, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 286, + 479, + 453 + ], + "spans": [ + { + "bbox": [ + 130, + 286, + 479, + 453 + ], + "type": "text", + "content": "Multi-level image-text supervision. After computing the RoI embeddings across pyramid levels, we perform a max pooling over the RoI embeddings per-level to obtain an image embedding for each pyramid level. Intuitively, max pooling allows the representation to focus on salient regions and discriminative features, thereby learning region-level information without explicit supervision. Then we apply the standard image-text contrastive loss (see Eqn. (1)) on each feature level separately, which aids the learning of rich semantic and object knowledge within every feature map (see Table 4c for ablations). The losses from all levels are weighted equally and summed together. Without explicit region-level supervision, the max pooling over regions encourages the more salient, text-aligned region features to contribute more to the whole image representation in the contrastive loss. Our experiments show emergent region-text alignments from the multi-level training, where the feature maps possess more localized semantic information compared to the baseline CLIP [17] backbone (see Fig. 3)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 457, + 479, + 611 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 457, + 479, + 611 + ], + "spans": [ + { + "bbox": [ + 130, + 457, + 479, + 611 + ], + "type": "text", + "content": "Region proposal network learning. Fig. 3 shows that the learnt multi-level representations exhibit localized semantics well-aligned with the text query, i.e. the salient regions have higher similarity with the text relative to the background. Motivated by this, we employ the multi-level visual-text similarity as a supervisory signal for training the Region Proposal Network (RPN) [30]. Specifically, we compute the cosine similarity between each RoI embedding and the text, and use it as the target RPN score for the center location of each RoI. Any negative dot product value is mapped to zero to keep the target score in range [0,1]. We use L1 regression loss and set the loss weight equal to the multi-level contrastive loss (see supplementary). The losses are propagated only through the RoI centers and other pixels are ignored. Note that the box regression of the RPN is not trained here but learnt later through the detection finetuning, as we only use the image-text paired data without any box annotations." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 626, + 359, + 637 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 626, + 359, + 637 + ], + "spans": [ + { + "bbox": [ + 131, + 626, + 359, + 637 + ], + "type": "text", + "content": "3.3 Shifted-Window Learning for Detection" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 641, + 490, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 490, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 490, + 665 + ], + "type": "text", + "content": "The CLIP ViT backbones are typically pretrained on lower resolutions (e.g. " + }, + { + "bbox": [ + 130, + 641, + 490, + 665 + ], + "type": "inline_equation", + "content": "224 \\times 224" + }, + { + "bbox": [ + 130, + 641, + 490, + 665 + ], + "type": "text", + "content": ") and later adapted to higher resolution detection images (e.g. " + }, + { + "bbox": [ + 130, + 641, + 490, + 665 + ], + "type": "inline_equation", + "content": "1024 \\times 1024" + }, + { + "bbox": [ + 130, + 641, + 490, + 665 + ], + "type": "text", + "content": ")." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 140, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 140, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 140, + 91, + 447, + 102 + ], + "type": "text", + "content": "Region-centric Image-Language Pretraining for Open-Vocabulary Detection" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "code", + "bbox": [ + 132, + 130, + 480, + 205 + ], + "blocks": [ + { + "bbox": [ + 133, + 115, + 314, + 128 + ], + "lines": [ + { + "bbox": [ + 133, + 115, + 314, + 128 + ], + "spans": [ + { + "bbox": [ + 133, + 115, + 314, + 128 + ], + "type": "text", + "content": "Algorithm 1 Shifted Window backbone" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "code_caption" + }, + { + "bbox": [ + 132, + 130, + 480, + 205 + ], + "lines": [ + { + "bbox": [ + 132, + 130, + 480, + 205 + ], + "spans": [ + { + "bbox": [ + 132, + 130, + 480, + 205 + ], + "type": "text", + "content": "x: image patch tokens + positional embeddings. M: attention window size. s = M // 2 " + }, + { + "bbox": [ + 132, + 130, + 480, + 205 + ], + "type": "inline_equation", + "content": "y =" + }, + { + "bbox": [ + 132, + 130, + 480, + 205 + ], + "type": "text", + "content": " forward_vit_with_win_attn(x, M) " + }, + { + "bbox": [ + 132, + 130, + 480, + 205 + ], + "type": "inline_equation", + "content": "x^{\\prime} =" + }, + { + "bbox": [ + 132, + 130, + 480, + 205 + ], + "type": "text", + "content": " np-roll(x, shift=[s, s], axis=[0, 1]) # shift " + }, + { + "bbox": [ + 132, + 130, + 480, + 205 + ], + "type": "inline_equation", + "content": "y^\\prime =" + }, + { + "bbox": [ + 132, + 130, + 480, + 205 + ], + "type": "text", + "content": " forward_vit_with_win_attn " + }, + { + "bbox": [ + 132, + 130, + 480, + 205 + ], + "type": "inline_equation", + "content": "(x^{\\prime},M)" + }, + { + "bbox": [ + 132, + 130, + 480, + 205 + ], + "type": "inline_equation", + "content": "y'' =" + }, + { + "bbox": [ + 132, + 130, + 480, + 205 + ], + "type": "text", + "content": " np-roll(y', shift=[-s, -s], axis=[0, 1]) # unshift return " + }, + { + "bbox": [ + 132, + 130, + 480, + 205 + ], + "type": "inline_equation", + "content": "(y + y'') / 2" + }, + { + "bbox": [ + 132, + 130, + 480, + 205 + ], + "type": "text", + "content": " # SWL backbone output" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "code_body" + } + ], + "index": 3, + "sub_type": "algorithm" + }, + { + "bbox": [ + 130, + 237, + 482, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 237, + 482, + 323 + ], + "spans": [ + { + "bbox": [ + 130, + 237, + 482, + 323 + ], + "type": "text", + "content": "While detection tasks benefit from global information, directly applying the pretrained ViT to high-resolution inputs is computationally intensive and can disrupt the locality structure of the pretrained lower-res features. Windowed attention in ViTDet [20] helps reduce computation but introduces bias from the fixed-size grid pattern. To address this, we propose the Shifted-Window Learning (SWL) approach, which improves information mixing across fixed windows and mitigates grid pattern bias." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 328, + 483, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 328, + 483, + 545 + ], + "spans": [ + { + "bbox": [ + 130, + 328, + 483, + 545 + ], + "type": "text", + "content": "Network architecture. Fig. 2 (right) and Algorithm 1 describe the SWL algorithm. The standard ViT consists of a patchifying layer and a set of transformer layers. After feeding the image through the patchifying layer, we obtain a feature map " + }, + { + "bbox": [ + 130, + 328, + 483, + 545 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 130, + 328, + 483, + 545 + ], + "type": "text", + "content": " of shape " + }, + { + "bbox": [ + 130, + 328, + 483, + 545 + ], + "type": "inline_equation", + "content": "(h, w, d)" + }, + { + "bbox": [ + 130, + 328, + 483, + 545 + ], + "type": "text", + "content": ". This feature map " + }, + { + "bbox": [ + 130, + 328, + 483, + 545 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 130, + 328, + 483, + 545 + ], + "type": "text", + "content": " is fed through the rest of the ViT with windowed attention layers on a grid " + }, + { + "bbox": [ + 130, + 328, + 483, + 545 + ], + "type": "inline_equation", + "content": "K \\times K" + }, + { + "bbox": [ + 130, + 328, + 483, + 545 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 328, + 483, + 545 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 130, + 328, + 483, + 545 + ], + "type": "text", + "content": " global attention layers evenly spaced throughout the ViT (where " + }, + { + "bbox": [ + 130, + 328, + 483, + 545 + ], + "type": "inline_equation", + "content": "L = 4" + }, + { + "bbox": [ + 130, + 328, + 483, + 545 + ], + "type": "text", + "content": ") following [20], resulting in output " + }, + { + "bbox": [ + 130, + 328, + 483, + 545 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 130, + 328, + 483, + 545 + ], + "type": "text", + "content": ". In parallel, we create another copy of " + }, + { + "bbox": [ + 130, + 328, + 483, + 545 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 130, + 328, + 483, + 545 + ], + "type": "text", + "content": ", which is rolled along both " + }, + { + "bbox": [ + 130, + 328, + 483, + 545 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 130, + 328, + 483, + 545 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 328, + 483, + 545 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 130, + 328, + 483, + 545 + ], + "type": "text", + "content": " axes by " + }, + { + "bbox": [ + 130, + 328, + 483, + 545 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 130, + 328, + 483, + 545 + ], + "type": "text", + "content": " pixels. The elements that roll beyond the last position are reintroduced from the first. We carefully design the attention masks such that the rolled around patches would not attend to the patches on the other side of the image (see the right figure of Algorithm 1). The shift size " + }, + { + "bbox": [ + 130, + 328, + 483, + 545 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 130, + 328, + 483, + 545 + ], + "type": "text", + "content": " is set as the half of the attention window size " + }, + { + "bbox": [ + 130, + 328, + 483, + 545 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 130, + 328, + 483, + 545 + ], + "type": "text", + "content": " (i.e. " + }, + { + "bbox": [ + 130, + 328, + 483, + 545 + ], + "type": "inline_equation", + "content": "s = M // 2" + }, + { + "bbox": [ + 130, + 328, + 483, + 545 + ], + "type": "text", + "content": "). Empirically, the window size " + }, + { + "bbox": [ + 130, + 328, + 483, + 545 + ], + "type": "inline_equation", + "content": "M = 16" + }, + { + "bbox": [ + 130, + 328, + 483, + 545 + ], + "type": "text", + "content": " equals the image size (e.g. 1024) divided by the product of patch size (e.g. " + }, + { + "bbox": [ + 130, + 328, + 483, + 545 + ], + "type": "inline_equation", + "content": "P = 16" + }, + { + "bbox": [ + 130, + 328, + 483, + 545 + ], + "type": "text", + "content": ") and the grid size (e.g. " + }, + { + "bbox": [ + 130, + 328, + 483, + 545 + ], + "type": "inline_equation", + "content": "K = 4" + }, + { + "bbox": [ + 130, + 328, + 483, + 545 + ], + "type": "text", + "content": "). The shifted feature map " + }, + { + "bbox": [ + 130, + 328, + 483, + 545 + ], + "type": "inline_equation", + "content": "x'" + }, + { + "bbox": [ + 130, + 328, + 483, + 545 + ], + "type": "text", + "content": " is then processed through the rest of the ViT in the same manner, resulting in output " + }, + { + "bbox": [ + 130, + 328, + 483, + 545 + ], + "type": "inline_equation", + "content": "y'" + }, + { + "bbox": [ + 130, + 328, + 483, + 545 + ], + "type": "text", + "content": " in the same shape " + }, + { + "bbox": [ + 130, + 328, + 483, + 545 + ], + "type": "inline_equation", + "content": "(h, w, d)" + }, + { + "bbox": [ + 130, + 328, + 483, + 545 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 130, + 328, + 483, + 545 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 130, + 328, + 483, + 545 + ], + "type": "text", + "content": ". We then unshift " + }, + { + "bbox": [ + 130, + 328, + 483, + 545 + ], + "type": "inline_equation", + "content": "y'" + }, + { + "bbox": [ + 130, + 328, + 483, + 545 + ], + "type": "text", + "content": " and combine it with " + }, + { + "bbox": [ + 130, + 328, + 483, + 545 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 130, + 328, + 483, + 545 + ], + "type": "text", + "content": " by averaging. We apply the above shifted window operations in detection finetuning and inference times." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 552, + 495, + 673 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 552, + 495, + 673 + ], + "spans": [ + { + "bbox": [ + 130, + 552, + 495, + 673 + ], + "type": "text", + "content": "Comparison with Swin Transformer. Compared to the Swin Transformer [23], we apply the shifted-window ideas as separate forward passes, while Swin Transformer applies similar ideas in an alternating manner layer by layer. Our approach requires no change to the vanilla transformer architecture and is compatible with any ViT backbones pretrained without shifted windows (e.g. [28]), whereas Swin Transformer requires specialized pretraining on the same architecture. Compared to the full-attention ViT [4], we observe that shifted-window ViT taps more effectively into the semantic knowledge of pretrained backbone than full global attention, perhaps because the window helps the model focus more on local cues and ignore the noises farther away." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 222, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 222, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 222, + 101 + ], + "type": "text", + "content": "D. Kim et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 357, + 127 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 357, + 127 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 357, + 127 + ], + "type": "text", + "content": "3.4 Distillation from Frozen ViT Backbone" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 133, + 482, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 133, + 482, + 300 + ], + "spans": [ + { + "bbox": [ + 130, + 133, + 482, + 300 + ], + "type": "text", + "content": "While the ViT backbone adapts to the detection tasks, it may lose some of its pretrained open-vocabulary knowledge. Therefore, we propose a simple distillation approach which uses a separate frozen ViT to teach the finetuned ViT backbone during the detection finetuning. We use a cosine distance loss that aligns the RoI-Align embeddings extracted from the feature maps of both backbones. The cosine distance is computed for each RoI then averaged over all RoIs. Empirically, we find it advantageous to add a " + }, + { + "bbox": [ + 130, + 133, + 482, + 300 + ], + "type": "inline_equation", + "content": "1 \\times 1" + }, + { + "bbox": [ + 130, + 133, + 482, + 300 + ], + "type": "text", + "content": " Conv projection layer to the finetuned ViT backbone before the RoI-Align, allowing some flexibility in distillation when jointly trained with other detection objectives. The auxiliary distillation loss is then added to other detection losses, with a loss weight " + }, + { + "bbox": [ + 130, + 133, + 482, + 300 + ], + "type": "inline_equation", + "content": "\\gamma = 1" + }, + { + "bbox": [ + 130, + 133, + 482, + 300 + ], + "type": "text", + "content": " (see supplementary). At inference, the ViT backbone features after the projection are used to compute the region VLM score " + }, + { + "bbox": [ + 130, + 133, + 482, + 300 + ], + "type": "inline_equation", + "content": "z_{i}" + }, + { + "bbox": [ + 130, + 133, + 482, + 300 + ], + "type": "text", + "content": " (Sec. 3.1). It is worth noting that the frozen ViT backbone is only used during training for distillation purposes and is removed at inference." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 300, + 482, + 372 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 300, + 482, + 372 + ], + "spans": [ + { + "bbox": [ + 130, + 300, + 482, + 372 + ], + "type": "text", + "content": "While previous studies [2, 10] have utilized knowledge distillation from the CLIP models, their teacher CLIP models mostly operate on image crops in an offline process, thus needing multiple forward passes through the backbone for all RoIs. In contrast, our distillation operates efficiently on RoIs cropped from dense feature maps in a single forward pass during the detection finetuning with minimal overhead." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 384, + 282, + 398 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 384, + 282, + 398 + ], + "spans": [ + { + "bbox": [ + 132, + 384, + 282, + 398 + ], + "type": "text", + "content": "4 Experimental Results" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 409, + 482, + 589 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 409, + 482, + 589 + ], + "spans": [ + { + "bbox": [ + 130, + 409, + 482, + 589 + ], + "type": "text", + "content": "Baseline reproduction. As discussed in Sec. 3.1, our baseline method is RO-ViT [17], a leading ViT-based approach for open-vocabulary detection. RO-ViT introduces cropped positional embedding (CPE) in CLIP pretraining to enhance the generalization of the pretrained ViT onto the downstream detection. Additionally, it adopts global average pooling of the ViT features instead of CLs-token pooling, which simplifies the adaptation onto higher resolution inputs, and the extraction of region features (e.g. RoI-Align) at the final layer. We reproduce the CLIP pretraining of RO-ViT [17] using the widely adopted LAION2B dataset [31], as the OpenAI CLIP's WIT [28] and ALIGN [14] datasets are not publicly available. We trained the CLIP models from scratch, following the same pretraining protocol and hyperparameters, including 500k iterations and 16k batch size (i.e. 8B samples seen in training), " + }, + { + "bbox": [ + 130, + 409, + 482, + 589 + ], + "type": "inline_equation", + "content": "224 \\times 224" + }, + { + "bbox": [ + 130, + 409, + 482, + 589 + ], + "type": "text", + "content": " image size, global average pooling, and cropped positional embedding. We use the standard InfoNCE loss (Eqn. (1)) instead of their focal constrastive loss [17]. The following compares our reproduced CLIP with OpenAI CLIP [28] and OpenCLIP [31]:" + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 134, + 595, + 479, + 635 + ], + "blocks": [ + { + "bbox": [ + 134, + 595, + 479, + 635 + ], + "lines": [ + { + "bbox": [ + 134, + 595, + 479, + 635 + ], + "spans": [ + { + "bbox": [ + 134, + 595, + 479, + 635 + ], + "type": "table", + "html": "
methoddataset#samples seenbackboneImageNet Top-1 Acc.
CLIP [28]WIT-400M13BViT-L/1475.5
OpenCLIP [31]LAION-2B32BViT-L/1475.2
RO-ViT CLIP [17] our repro.LAION-2B8BViT-L/1673.9
", + "image_path": "6db5dad00c99d7b10139b48db6a5ad8cd2240f4ec7962841b0c494160f3c887b.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 641, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 641, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 641, + 481, + 665 + ], + "type": "text", + "content": "On the zero-shot ImageNet classification benchmark, the LAION-2B pretraining matches or slightly underperforms the OpenAI CLIP with WIT-400M." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 140, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 140, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 140, + 91, + 447, + 102 + ], + "type": "text", + "content": "Region-centric Image-Language Pretraining for Open-Vocabulary Detection" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 134, + 114, + 480, + 357 + ], + "blocks": [ + { + "bbox": [ + 134, + 114, + 480, + 357 + ], + "lines": [ + { + "bbox": [ + 134, + 114, + 480, + 357 + ], + "spans": [ + { + "bbox": [ + 134, + 114, + 480, + 357 + ], + "type": "table", + "html": "
methodpretraining modelpretraining datadetector backbonew/ pseudo box labelsmask APrmask AP
ConvNet based:
OV-DETR [40]ViT-B/32CLIP-400MR-50-17.426.6
Kaul et al. [15]ViT-B/32CLIP-400MR-50-19.330.6
DetPro-Cascade [5]ViT-B/32CLIP-400MR-50-20.027.0
Rasheed [29]ViT-B/32CLIP-400MR-50-21.125.9
BARON [35]ViT-B/32CLIP-400MR-50-22.627.6
CoDet [24]R-50CLIP-400M + CC3MR-50-23.430.7
EdaDet [32]R-50CLIP-400MR-50-23.727.5
VL-PLM [43]ViT-B/32CLIP-400MR-50✓17.227.0
PromptDet [6]ViT-B/32CLIP-400MR-50✓21.425.3
OADB [33]ViT-B/32CLIP-400MR-50✓21.726.6
RegionCLIP [44]R-50x4CLIP-400M + CC3MR-50x4✓22.032.3
CORA [36]R-50x4CLIP-400MR-50x4✓22.2bx-
Detric-CN2 [46]ViT-B/32CLIP-400M + INet-21KR-50WS24.632.4
ViLD-Ens [10]EffNet-B7ALIGN-1.8BEffNet-B7-26.329.3
F-VLM [18]R-50x64CLIP-400MR-50x64-32.834.9
ViT based:
OWL-ViT [26]O+VGViT-L/14CLIP-400MViT-L/14-25.6bx34.7bx
OWL-ViT v2 [25]O+VGViT-L/14WebLI-10BViT-L/14✓45.9bx50.4bx
RO-ViT [17]ViT-B/16ALIGN-1.8BViT-B/16-28.030.2
RO-ViT [17]†ViT-L/16LAION-2BViT-L/16-32.432.9
CFM-ViT [16]ViT-B/16ALIGN-1.8BViT-B/16-28.832.0
CFM-ViT [16]ViT-L/16ALIGN-1.8BViT-L/16-33.936.6
CFM-ViT [16]*ViT-L/16LAION-2BViT-L/16-33.836.4
DITO (ours)ViT-S/16LAION-2BViT-S/16-26.228.8
DITO (ours)ViT-B/16LAION-2BViT-B/16-31.532.4
DITO (ours)ViT-L/16LAION-2BViT-L/16-37.636.2
DITO (ours)ViT-L/16DataComp-1BViT-L/16-40.538.0
", + "image_path": "0ff3bd1951075c3d9d838938e25e2cb244493d5fcdcbbf99f83d150a0d2c92bd.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 360, + 482, + 405 + ], + "lines": [ + { + "bbox": [ + 130, + 360, + 482, + 405 + ], + "spans": [ + { + "bbox": [ + 130, + 360, + 482, + 405 + ], + "type": "text", + "content": "Table 1: LVIS open-vocabulary detection (mask AP). DITO outperforms the best existing approach by " + }, + { + "bbox": [ + 130, + 360, + 482, + 405 + ], + "type": "inline_equation", + "content": "+3.7" + }, + { + "bbox": [ + 130, + 360, + 482, + 405 + ], + "type": "text", + "content": " mask " + }, + { + "bbox": [ + 130, + 360, + 482, + 405 + ], + "type": "inline_equation", + "content": "\\mathrm{AP}_r" + }, + { + "bbox": [ + 130, + 360, + 482, + 405 + ], + "type": "text", + "content": ". WS: uses weak supervision from ImageNet-21K. †: reports LAION-2B results in arXiv version. *: our reproduced result using LAION-2B. O+VG: uses extra Objects365 and Visual Genome data." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 130, + 422, + 482, + 543 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 422, + 482, + 543 + ], + "spans": [ + { + "bbox": [ + 130, + 422, + 482, + 543 + ], + "type": "text", + "content": "Pretraining setup. After the above-mentioned baseline CLIP training, we apply our region-centric pretraining (REP) where we freeze the image and text encoders trained in the first phase and introduce the detector heads. We use the Simple FPN [20], and classification layers of Faster R-CNN and RPN [30], where we replace the batch normalization with layer normalization. At the " + }, + { + "bbox": [ + 130, + 422, + 482, + 543 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 422, + 482, + 543 + ], + "type": "text", + "content": "-th pyramid level " + }, + { + "bbox": [ + 130, + 422, + 482, + 543 + ], + "type": "inline_equation", + "content": "i \\in \\{2, 3, 4, 5\\}" + }, + { + "bbox": [ + 130, + 422, + 482, + 543 + ], + "type": "text", + "content": ", we randomly sample " + }, + { + "bbox": [ + 130, + 422, + 482, + 543 + ], + "type": "inline_equation", + "content": "n_i \\in \\{400, 200, 100, 50\\}" + }, + { + "bbox": [ + 130, + 422, + 482, + 543 + ], + "type": "text", + "content": " box regions and compute their RoI-Align features. We use a short training cycle of " + }, + { + "bbox": [ + 130, + 422, + 482, + 543 + ], + "type": "inline_equation", + "content": "30k" + }, + { + "bbox": [ + 130, + 422, + 482, + 543 + ], + "type": "text", + "content": " iterations, " + }, + { + "bbox": [ + 130, + 422, + 482, + 543 + ], + "type": "inline_equation", + "content": "4k" + }, + { + "bbox": [ + 130, + 422, + 482, + 543 + ], + "type": "text", + "content": " batch size, " + }, + { + "bbox": [ + 130, + 422, + 482, + 543 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 130, + 422, + 482, + 543 + ], + "type": "text", + "content": " image size, AdamW optimizer with an initial learning rate of 1e-4 with linear decay. For both phases of image-text pretraining, we use the publicly available LAION-2B [31] dataset." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 545, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 482, + 665 + ], + "type": "text", + "content": "Detection finetuning setup. As noted in Sec. 3.2, we adopt the ViTDet [20] with SimpleFPN as our detector following the baseline works [16,17]. We follow the same finetuning settings of [17]. Specifically, we train the detector with image size " + }, + { + "bbox": [ + 130, + 545, + 482, + 665 + ], + "type": "inline_equation", + "content": "1024 \\times 1024" + }, + { + "bbox": [ + 130, + 545, + 482, + 665 + ], + "type": "text", + "content": " and use windowed attention in the backbone with grid size " + }, + { + "bbox": [ + 130, + 545, + 482, + 665 + ], + "type": "inline_equation", + "content": "4 \\times 4" + }, + { + "bbox": [ + 130, + 545, + 482, + 665 + ], + "type": "text", + "content": ". The learning rate for the backbone is set lower as " + }, + { + "bbox": [ + 130, + 545, + 482, + 665 + ], + "type": "inline_equation", + "content": "0.6 \\times" + }, + { + "bbox": [ + 130, + 545, + 482, + 665 + ], + "type": "text", + "content": " of the detector head layers. We use " + }, + { + "bbox": [ + 130, + 545, + 482, + 665 + ], + "type": "inline_equation", + "content": "\\alpha = 0.3" + }, + { + "bbox": [ + 130, + 545, + 482, + 665 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 545, + 482, + 665 + ], + "type": "inline_equation", + "content": "\\beta = 0.65" + }, + { + "bbox": [ + 130, + 545, + 482, + 665 + ], + "type": "text", + "content": " for score combination in Eqn. (2). The text embedding of each category is calculated as the average over the CLIP prompt templates. We use the batch size 128, the SGD optimizer with momentum 0.9. The initial learning rate and iterations are set to 0.18 and " + }, + { + "bbox": [ + 130, + 545, + 482, + 665 + ], + "type": "inline_equation", + "content": "36.8\\mathrm{k}" + }, + { + "bbox": [ + 130, + 545, + 482, + 665 + ], + "type": "text", + "content": " for LVIS, and 0.02 and " + }, + { + "bbox": [ + 130, + 545, + 482, + 665 + ], + "type": "inline_equation", + "content": "11.3\\mathrm{k}" + }, + { + "bbox": [ + 130, + 545, + 482, + 665 + ], + "type": "text", + "content": " for COCO datasets." + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 221, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 221, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 221, + 100 + ], + "type": "text", + "content": "D. Kim et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 132, + 114, + 481, + 339 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 481, + 339 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 481, + 339 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 481, + 339 + ], + "type": "table", + "html": "
methodpretraining modelpretraining datadetector backbonew/ pseudo box labelsnovel APAP
ConvNet based:
OVR-CNN [41]R-50CLIP-400M + COCO-CapR-50-22.839.9
ViLD [10]ViT-B/32CLIP-400MR-50-27.651.3
F-VLM [18]R-50CLIP-400MR-50-28.039.6
OV-DETR [40]ViT-B/32CLIP-400MR-50-29.452.7
CoDet [24]R-50CLIP-400M + COCO-CapR-50-30.646.6
PromptDet [6]ViT-B/32CLIP-400MR-50✓26.650.6
XPM [13]R-50CLIP-400MR-50✓27.041.2
OADB [33]ViT-B/32CLIP-400M + COCO-CapR-50✓30.047.2
VL-PLM [43]ViT-B/32CLIP-400MR-50✓34.453.5
RegionCLIP [44]R-50x4CLIP-400M + CC3M + COCO-CapR-50x4✓39.355.7
EdaDet [32]R-50CLIP-400MR-50✓40.252.5
CORA [36]R-50x4CLIP-400MR-50x4✓41.743.8
Detic-CN2 [46]ViT-B/32CLIP-400M + INet-21KR-50WS27.845.0
ViT based:
RO-ViT [17]ViT-B/16ALIGN-1.8BViT-B/16-30.241.5
RO-ViT [17]ViT-L/16ALIGN-1.8BViT-L/16-33.047.7
RO-ViT [17]*ViT-L/16LAION-2BViT-L/16-33.347.9
CFM-ViT [16]ViT-B/16ALIGN-1.8BViT-B/16-30.842.4
CFM-ViT [16]ViT-L/16ALIGN-1.8BViT-L/16-34.146.0
CFM-ViT [16]*ViT-L/16LAION-2BViT-L/16-34.346.4
DITO (ours)ViT-S/16LAION-2BViT-S/16-32.344.4
DITO (ours)ViT-B/16LAION-2BViT-B/16-36.648.8
DITO (ours)ViT-L/16LAION-2BViT-L/16-39.654.4
DITO (ours)ViT-L/16DataComp-1BViT-L/16-40.254.6
", + "image_path": "0ddbbbbf2fc186fe14ce5112e2ec3ef78dbae512aec39cefe750b8b7f60b0d22.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 193, + 380, + 416, + 471 + ], + "blocks": [ + { + "bbox": [ + 132, + 342, + 482, + 374 + ], + "lines": [ + { + "bbox": [ + 132, + 342, + 482, + 374 + ], + "spans": [ + { + "bbox": [ + 132, + 342, + 482, + 374 + ], + "type": "text", + "content": "Table 2: COCO open-vocabulary detection (box AP50). DITO demonstrates a very competitive novel category AP without using pseudo labeling or weak supervision (WS). *: our reproduced result using LAION-2B." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 193, + 380, + 416, + 471 + ], + "lines": [ + { + "bbox": [ + 193, + 380, + 416, + 471 + ], + "spans": [ + { + "bbox": [ + 193, + 380, + 416, + 471 + ], + "type": "table", + "html": "
methodbackboneAP\\( AP_{50} \\)\\( AP_{75} \\)
Supervised [10]R-5025.638.628.0
ViLD [10]R-5011.818.212.6
DetPro [5]R-5012.118.812.9
BARON [35]R-5013.621.014.5
F-VLM [18]R-50x1616.225.317.5
F-VLM [18]R-50x6417.727.419.1
RO-ViT [17]ViT-L/1617.126.918.5
CFM-ViT [16]ViT-L/1618.728.920.3
DITO (ours)ViT-L/1620.031.821.5
", + "image_path": "c4e8c95bf2692509d5e782149ae9f6e6c43f70224e596b97b7a4158de9ade568.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 472, + 480, + 496 + ], + "lines": [ + { + "bbox": [ + 132, + 472, + 480, + 496 + ], + "spans": [ + { + "bbox": [ + 132, + 472, + 480, + 496 + ], + "type": "text", + "content": "Table 3: Zero-shot transfer detection from LVISbase to Objects365 (box AP). All models are tested on Objects365 dataset following the setup of [10]." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 132, + 513, + 228, + 525 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 513, + 228, + 525 + ], + "spans": [ + { + "bbox": [ + 132, + 513, + 228, + 525 + ], + "type": "text", + "content": "4.1 Main Results" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": "LVIS Benchmark. In Table 1, we report the comparison with existing methods on the challenging LVIS benchmark. The 'frequent' and 'common' classes of the dataset belong to the base categories " + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathrm{C}_B" + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": ", and the 'rare' classes are the novel categories " + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathrm{C}_N" + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": ". The primary metric is the mask AP on rare classes (mask " + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathrm{AP}_r" + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": "). The DITO model achieves the performance of 37.6 mask " + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathrm{AP}_r" + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": ", which significantly outperforms the state-of-the-art approach RO-ViT [17] with the same ViT-L backbone by +5.1 points using the same pretraining data LAION-2B [31]. We also outperform the state-of-the-art CFM-ViT [16] by +3.6 points. Our best performance sets a new state-of-the-art 40.9 mask " + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathrm{AP}_r" + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": " when using DataComp1B [7] in pretraining. With the ViT-B backbone, DITO maintains a healthy margin of around +2.5 " + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathrm{AP}_r" + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": " above existing ViT-B based approaches." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 140, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 140, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 140, + 91, + 447, + 102 + ], + "type": "text", + "content": "Region-centric Image-Language Pretraining for Open-Vocabulary Detection" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 140, + 114, + 274, + 175 + ], + "blocks": [ + { + "bbox": [ + 140, + 114, + 274, + 175 + ], + "lines": [ + { + "bbox": [ + 140, + 114, + 274, + 175 + ], + "spans": [ + { + "bbox": [ + 140, + 114, + 274, + 175 + ], + "type": "table", + "html": "
methodAPrAP
RO-ViT [17]32.432.9
RO-ViT our repro.32.233.0
w/REP34.8 (+2.6)34.9
w/SWL35.0 (+2.8)35.2
w/REP+SWL36.3 (+4.1)35.8
w/REP+SWL+FD37.6 (+5.4)36.2
", + "image_path": "97a291560d6c641a960d42662b803a288f17bcf2a1a4bcc0e761864a47d8a49a.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 282, + 114, + 380, + 159 + ], + "blocks": [ + { + "bbox": [ + 160, + 177, + 246, + 186 + ], + "lines": [ + { + "bbox": [ + 160, + 177, + 246, + 186 + ], + "spans": [ + { + "bbox": [ + 160, + 177, + 246, + 186 + ], + "type": "text", + "content": "(a) DITO framework." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 282, + 114, + 380, + 159 + ], + "lines": [ + { + "bbox": [ + 282, + 114, + 380, + 159 + ], + "spans": [ + { + "bbox": [ + 282, + 114, + 380, + 159 + ], + "type": "table", + "html": "
pretraining methodAPr
RO-ViT our repro.32.2
w/FPN33.2
w/FPN+Head34.2
w/FPN+Head+RPN34.8
", + "image_path": "d71b46da26f3a0d11794c4159f0040e8b007243a84690db33b0789026d19645a.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 389, + 114, + 481, + 167 + ], + "blocks": [ + { + "bbox": [ + 277, + 176, + 379, + 186 + ], + "lines": [ + { + "bbox": [ + 277, + 176, + 379, + 186 + ], + "spans": [ + { + "bbox": [ + 277, + 176, + 379, + 186 + ], + "type": "text", + "content": "(b) Detector components." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 389, + 114, + 481, + 167 + ], + "lines": [ + { + "bbox": [ + 389, + 114, + 481, + 167 + ], + "spans": [ + { + "bbox": [ + 389, + 114, + 481, + 167 + ], + "type": "table", + "html": "
RoI embeddingAPr
global avg/lvl34.0
global max/lvl33.7
multi RoIs, avg/lvl33.8
multi RoIs, max/lvl34.8
multi RoIs, max all34.1
", + "image_path": "ce98476c5ff5fc6f8412e92914b3252df8065936988fe8fcc48653b345a1889f.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 149, + 232, + 302, + 277 + ], + "blocks": [ + { + "bbox": [ + 391, + 176, + 470, + 186 + ], + "lines": [ + { + "bbox": [ + 391, + 176, + 470, + 186 + ], + "spans": [ + { + "bbox": [ + 391, + 176, + 470, + 186 + ], + "type": "text", + "content": "(c) RoI embedding." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 131, + 192, + 480, + 225 + ], + "lines": [ + { + "bbox": [ + 131, + 192, + 480, + 225 + ], + "spans": [ + { + "bbox": [ + 131, + 192, + 480, + 225 + ], + "type": "text", + "content": "Table 4: Ablation on overall DITO framework and Region-centric Pretraining (Sec. 3.2). REP: Region-centric Pretraining, SWL: Shifted-Window Learning, FD: Frozen backbone Distillation. Best setting is in gray." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 149, + 232, + 302, + 277 + ], + "lines": [ + { + "bbox": [ + 149, + 232, + 302, + 277 + ], + "spans": [ + { + "bbox": [ + 149, + 232, + 302, + 277 + ], + "type": "table", + "html": "
backboneAPrAP
fully global attn.33.433.8
baseline window attn. [17]32.233.0
Swin [23] style31.333.1
shifted window35.035.2
", + "image_path": "04c52f179a58b973e4aebabf03425eeca308712554b61792c61b14df7601e4e4.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 316, + 232, + 473, + 277 + ], + "blocks": [ + { + "bbox": [ + 164, + 278, + 277, + 288 + ], + "lines": [ + { + "bbox": [ + 164, + 278, + 277, + 288 + ], + "spans": [ + { + "bbox": [ + 164, + 278, + 277, + 288 + ], + "type": "text", + "content": "(a) Shifted-window learning." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 316, + 232, + 473, + 277 + ], + "lines": [ + { + "bbox": [ + 316, + 232, + 473, + 277 + ], + "spans": [ + { + "bbox": [ + 316, + 232, + 473, + 277 + ], + "type": "table", + "html": "
# global attn. layerbasew/ SWL
030.734.6 (+3.9)
432.235.0 (+2.8)
1232.434.2 (+1.8)
24 (all layers)33.433.4 (+0.0)
", + "image_path": "7fda14b5668654059d4d03137e6b8d593e44ad6e3193ad047450f17bf93e9666.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 309, + 278, + 473, + 288 + ], + "lines": [ + { + "bbox": [ + 309, + 278, + 473, + 288 + ], + "spans": [ + { + "bbox": [ + 309, + 278, + 473, + 288 + ], + "type": "text", + "content": "(b) SWL w.r.t. # global attention layers." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 153, + 294, + 459, + 305 + ], + "lines": [ + { + "bbox": [ + 153, + 294, + 459, + 305 + ], + "spans": [ + { + "bbox": [ + 153, + 294, + 459, + 305 + ], + "type": "text", + "content": "Table 5: Ablation on Shifted-Window Learning (SWL - Sec. 3.3)." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 130, + 318, + 480, + 402 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 318, + 480, + 402 + ], + "spans": [ + { + "bbox": [ + 130, + 318, + 480, + 402 + ], + "type": "text", + "content": "COCO Benchmark. We present the comparison on COCO benchmark in Table 2. The main metric is AP50 of novel categories (novel AP). Without using pseudo labeling [6,13,43,44], weak supervision [46], or externally trained detector modules [29, 35], our model demonstrates competitive results of 39.6 novel AP with LAION-2B and 40.2 with DataComp-1B. Among the ViT-based methods, DITO outperforms recent works RO-ViT [17] and CFM-ViT [16] by a clear margin of " + }, + { + "bbox": [ + 130, + 318, + 480, + 402 + ], + "type": "inline_equation", + "content": "+6.3" + }, + { + "bbox": [ + 130, + 318, + 480, + 402 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 318, + 480, + 402 + ], + "type": "inline_equation", + "content": "+5.3" + }, + { + "bbox": [ + 130, + 318, + 480, + 402 + ], + "type": "text", + "content": " points, respectively." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 405, + 480, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 405, + 480, + 464 + ], + "spans": [ + { + "bbox": [ + 130, + 405, + 480, + 464 + ], + "type": "text", + "content": "Transfer detection. We further evaluate DITO in the transfer detection setting, where the open-vocabulary detector trained on one dataset " + }, + { + "bbox": [ + 130, + 405, + 480, + 464 + ], + "type": "inline_equation", + "content": "(\\mathrm{LVIS}_{base})" + }, + { + "bbox": [ + 130, + 405, + 480, + 464 + ], + "type": "text", + "content": " is tested on another dataset (Objects365) without any finetuning. By simply replacing the text embeddings, Table 3 shows that DITO achieves 20.0 AP, outperforming previous methods using ConvNet or ViT backbones of similar size." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 131, + 479, + 245, + 489 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 479, + 245, + 489 + ], + "spans": [ + { + "bbox": [ + 131, + 479, + 245, + 489 + ], + "type": "text", + "content": "4.2 Ablation Studies" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 130, + 492, + 480, + 516 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 492, + 480, + 516 + ], + "spans": [ + { + "bbox": [ + 130, + 492, + 480, + 516 + ], + "type": "text", + "content": "For ablation study, we use the ViT-L/16 model pretrained with LAION-2B, and evaluate on the LVIS benchmark and report mask " + }, + { + "bbox": [ + 130, + 492, + 480, + 516 + ], + "type": "inline_equation", + "content": "\\mathrm{AP}_r" + }, + { + "bbox": [ + 130, + 492, + 480, + 516 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 130, + 518, + 485, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 518, + 485, + 602 + ], + "spans": [ + { + "bbox": [ + 130, + 518, + 485, + 602 + ], + "type": "text", + "content": "DITO overall framework. Table 4a summarizes the benefits of each DITO components. The region-centric pretraining (REP) improves the contrastive model baseline by " + }, + { + "bbox": [ + 130, + 518, + 485, + 602 + ], + "type": "inline_equation", + "content": "+2.6\\mathrm{AP}_r" + }, + { + "bbox": [ + 130, + 518, + 485, + 602 + ], + "type": "text", + "content": " and shifted window learning (SWL) improves the baseline by " + }, + { + "bbox": [ + 130, + 518, + 485, + 602 + ], + "type": "inline_equation", + "content": "+2.8" + }, + { + "bbox": [ + 130, + 518, + 485, + 602 + ], + "type": "text", + "content": " points. Combining both strategies brings a significant gain of " + }, + { + "bbox": [ + 130, + 518, + 485, + 602 + ], + "type": "inline_equation", + "content": "+4.1\\mathrm{AP}_r" + }, + { + "bbox": [ + 130, + 518, + 485, + 602 + ], + "type": "text", + "content": ". Lastly, we observe that incorporating the frozen backbone distillation leads to an additional boost of " + }, + { + "bbox": [ + 130, + 518, + 485, + 602 + ], + "type": "inline_equation", + "content": "+1.2" + }, + { + "bbox": [ + 130, + 518, + 485, + 602 + ], + "type": "text", + "content": " points. In the following, we provide ablations for each components." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 130, + 605, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 481, + 665 + ], + "type": "text", + "content": "Region-centric Pretraining. In Table 4b, we ablate our region-centric image-text pretraining by progressively adding the FPN, Faster R-CNN head, and RPN into the pretraining. Our 'baseline' is the contrastive image-text pretraining with cropped positional embedding [17]. On top of this, 'w/ FPN' introduces the FPN into the pretraining, where each pyramid level (whole image) map is mean-pooled" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 221, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 221, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 221, + 100 + ], + "type": "text", + "content": "D. Kim et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 143, + 114, + 315, + 159 + ], + "blocks": [ + { + "bbox": [ + 143, + 114, + 315, + 159 + ], + "lines": [ + { + "bbox": [ + 143, + 114, + 315, + 159 + ], + "spans": [ + { + "bbox": [ + 143, + 114, + 315, + 159 + ], + "type": "table", + "html": "
backboneAPAPrAPcAPf
GT boxes given (region classification):
before det. training48.756.851.542.1
after det. training53.754.253.753.5
after det. training w/FD54.857.754.353.2
", + "image_path": "0eef19f6046d3530c0ccd16ae6b852078f672903c4e08bfa6a6a8012182b13ab.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 327, + 114, + 479, + 159 + ], + "blocks": [ + { + "bbox": [ + 171, + 160, + 277, + 170 + ], + "lines": [ + { + "bbox": [ + 171, + 160, + 277, + 170 + ], + "spans": [ + { + "bbox": [ + 171, + 160, + 277, + 170 + ], + "type": "text", + "content": "(a) GT region recognition." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 327, + 114, + 479, + 159 + ], + "lines": [ + { + "bbox": [ + 327, + 114, + 479, + 159 + ], + "spans": [ + { + "bbox": [ + 327, + 114, + 479, + 159 + ], + "type": "table", + "html": "
methodAPrAP
RO-ViT our repro.32.233.0
RO-ViT our repro.+FD34.0 (+1.8)33.6
w/ REP+SWL36.335.8
w/ REP+SWL+FD37.6 (+1.3)36.2
", + "image_path": "76e41f5397de3b2a2041f9ff38eb3a918e4f1ff57df9a8c3ea8947e7c1b5ae66.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 348, + 160, + 451, + 170 + ], + "lines": [ + { + "bbox": [ + 348, + 160, + 451, + 170 + ], + "spans": [ + { + "bbox": [ + 348, + 160, + 451, + 170 + ], + "type": "text", + "content": "(b) Effect of FD on OVD." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 153, + 176, + 459, + 188 + ], + "lines": [ + { + "bbox": [ + 153, + 176, + 459, + 188 + ], + "spans": [ + { + "bbox": [ + 153, + 176, + 459, + 188 + ], + "type": "text", + "content": "Table 6: Ablation on frozen backbone distillation (FD - Sec. 3.4)." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 130, + 204, + 480, + 275 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 204, + 480, + 275 + ], + "spans": [ + { + "bbox": [ + 130, + 204, + 480, + 275 + ], + "type": "text", + "content": "into an image embedding, followed by the image-text contrastive loss per level. It improves the baseline by " + }, + { + "bbox": [ + 130, + 204, + 480, + 275 + ], + "type": "inline_equation", + "content": "+1.0\\mathrm{AP}_r" + }, + { + "bbox": [ + 130, + 204, + 480, + 275 + ], + "type": "text", + "content": ". Adding the Faster R-CNN head further improves the alignment between the pretraining and detection finetuning, showing a gain of " + }, + { + "bbox": [ + 130, + 204, + 480, + 275 + ], + "type": "inline_equation", + "content": "+2.0\\mathrm{AP}_r" + }, + { + "bbox": [ + 130, + 204, + 480, + 275 + ], + "type": "text", + "content": ". Incorporating all components i.e. the FPN, Faster R-CNN and RPN heads achieves the best 34.8 " + }, + { + "bbox": [ + 130, + 204, + 480, + 275 + ], + "type": "inline_equation", + "content": "\\mathrm{AP}_r" + }, + { + "bbox": [ + 130, + 204, + 480, + 275 + ], + "type": "text", + "content": ", a significant gain of +2.6 points over the baseline." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 276, + 480, + 372 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 276, + 480, + 372 + ], + "spans": [ + { + "bbox": [ + 130, + 276, + 480, + 372 + ], + "type": "text", + "content": "Table 4c shows that both global avg- and max-pooling are sub-optimal due to the lack of saliency map (avg-pool), and the limited capacity of a single pixel to represent semantic concepts for contrastive learning (max-pool), respectively. Our approach combines the best of both worlds, by first avg-pooling within each RoI, and then max-pooling over these RoI embeddings. Each embedding represents a proper RoI and the global representation captures the saliency map through max-pooling. Pooling per pyramid level (i.e. multi-level image-text supervision) outperforms pooling over all levels." + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 133, + 379, + 247, + 418 + ], + "blocks": [ + { + "bbox": [ + 133, + 379, + 247, + 418 + ], + "lines": [ + { + "bbox": [ + 133, + 379, + 247, + 418 + ], + "spans": [ + { + "bbox": [ + 133, + 379, + 247, + 418 + ], + "type": "image", + "image_path": "b96117cfb9376fb716d03fff8f0e0ca5da34dadf5fd46874d5c062fe64de13ea.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 140, + 420, + 241, + 428 + ], + "lines": [ + { + "bbox": [ + 140, + 420, + 241, + 428 + ], + "spans": [ + { + "bbox": [ + 140, + 420, + 241, + 428 + ], + "type": "text", + "content": "\"am riding a cart pulled by a donkey" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 249, + 379, + 364, + 418 + ], + "blocks": [ + { + "bbox": [ + 249, + 379, + 364, + 418 + ], + "lines": [ + { + "bbox": [ + 249, + 379, + 364, + 418 + ], + "spans": [ + { + "bbox": [ + 249, + 379, + 364, + 418 + ], + "type": "image", + "image_path": "1e73822b46fb7338b5e7b76eded1625ba323c2d7f427c289a958c2135eaeff20.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 255, + 420, + 359, + 428 + ], + "lines": [ + { + "bbox": [ + 255, + 420, + 359, + 428 + ], + "spans": [ + { + "bbox": [ + 255, + 420, + 359, + 428 + ], + "type": "text", + "content": "\"a crane operates amidst piles of rubble\"" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 365, + 379, + 480, + 418 + ], + "blocks": [ + { + "bbox": [ + 365, + 379, + 480, + 418 + ], + "lines": [ + { + "bbox": [ + 365, + 379, + 480, + 418 + ], + "spans": [ + { + "bbox": [ + 365, + 379, + 480, + 418 + ], + "type": "image", + "image_path": "10698e6b13c6dee5cf535e0541da93d88e221089a891af5d2144a501987044eb.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 373, + 418, + 473, + 432 + ], + "lines": [ + { + "bbox": [ + 373, + 418, + 473, + 432 + ], + "spans": [ + { + "bbox": [ + 373, + 418, + 473, + 432 + ], + "type": "text", + "content": "\"aperson wearing blue rubber gloves is loading a syringe with vegetables\"" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 133, + 433, + 247, + 472 + ], + "blocks": [ + { + "bbox": [ + 133, + 433, + 247, + 472 + ], + "lines": [ + { + "bbox": [ + 133, + 433, + 247, + 472 + ], + "spans": [ + { + "bbox": [ + 133, + 433, + 247, + 472 + ], + "type": "image", + "image_path": "441f38d48142acc4a2426621641d897860befe2659f79a5f7497c3af68af3a3b.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 141, + 472, + 240, + 486 + ], + "lines": [ + { + "bbox": [ + 141, + 472, + 240, + 486 + ], + "spans": [ + { + "bbox": [ + 141, + 472, + 240, + 486 + ], + "type": "text", + "content": "\"amotherand her two children sit on a green bench\"" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 490, + 480, + 534 + ], + "lines": [ + { + "bbox": [ + 130, + 490, + 480, + 534 + ], + "spans": [ + { + "bbox": [ + 130, + 490, + 480, + 534 + ], + "type": "text", + "content": "Fig. 3: Visual-text similarity map. For each example, we show the paired image (left) and text (bottom) input, and the visual-text similarity map using the contrastive model baseline [17] backbone features (middle) or our region-centric pretraining features (right). We use Flickr30K (top row) and COCO Captions (bottom row) datasets." + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 249, + 433, + 364, + 472 + ], + "blocks": [ + { + "bbox": [ + 249, + 433, + 364, + 472 + ], + "lines": [ + { + "bbox": [ + 249, + 433, + 364, + 472 + ], + "spans": [ + { + "bbox": [ + 249, + 433, + 364, + 472 + ], + "type": "image", + "image_path": "12b3edad64336eddf3980b8fe35f84ac2e66a45fcd2d24a7f8bd8043510b9074.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 257, + 472, + 356, + 485 + ], + "lines": [ + { + "bbox": [ + 257, + 472, + 356, + 485 + ], + "spans": [ + { + "bbox": [ + 257, + 472, + 356, + 485 + ], + "type": "text", + "content": "\"Cows on the street next to people that are on the sidewalk\"" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 365, + 433, + 480, + 472 + ], + "blocks": [ + { + "bbox": [ + 365, + 433, + 480, + 472 + ], + "lines": [ + { + "bbox": [ + 365, + 433, + 480, + 472 + ], + "spans": [ + { + "bbox": [ + 365, + 433, + 480, + 472 + ], + "type": "image", + "image_path": "1f9acc1833c0be07060bdfd5457cfb76accdc3fca7c2867d70a93a7c93765476.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 376, + 472, + 470, + 485 + ], + "lines": [ + { + "bbox": [ + 376, + 472, + 470, + 485 + ], + "spans": [ + { + "bbox": [ + 376, + 472, + 470, + 485 + ], + "type": "text", + "content": "\"small white bird sits on the back of a horse in an open field\"" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": "Shifted-Window Learning for detection. The CLIP ViT backbones are initially pretrained on lower resolutions and then adapted to higher resolution detection images. In Table 5a, we assess the efficacy of the shifted-window backbone in open-vocabulary detection training. Although the fully global attention model improves the detection task compared to the baseline windowed attention model [17], directly applying the pretrained ViT on high-resolution inputs may not be optimal, potentially compromising the locality structure of the pretrained lower-resolution features. Our shifted-window learning approach achieves " + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "inline_equation", + "content": "35.0\\mathrm{AP}_r" + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": " by preserving the locality structure from windowed attention as well as integrating information across fixed windows, outperforming the fully global" + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 140, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 140, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 140, + 91, + 447, + 102 + ], + "type": "text", + "content": "Region-centric Image-Language Pretraining for Open-Vocabulary Detection" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "text", + "content": "attention model (33.4 " + }, + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "inline_equation", + "content": "\\mathrm{AP}_r" + }, + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "text", + "content": ") which is computationally intensive. Notably, OWL-ViT [25,26] adopts the fully global attention model. In addition, naively applying the layer-alternating shifted window as in Swin [23] leads to a performance drop (see Sec. 3.3). Table 5b delves deeper into the behavior of SWL. The advantage of SWL diminishes steadily with increasing number of global attention layers in the windowed attention backbone, validating its better information mixing enabled by the SWL." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 202, + 482, + 370 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 202, + 482, + 370 + ], + "spans": [ + { + "bbox": [ + 130, + 202, + 482, + 370 + ], + "type": "text", + "content": "Frozen backbone distillation. Table 6a studies the region classification capability of the CLIP ViT backbone before and after the detection finetuning. We use the ground truth boxes and measures AP scores to evaluate the zero-shot region classification of the base ('frequent' + 'common') and novel ('rare') categories. We observe that the finetuned backbone indeed shows a notable drop of " + }, + { + "bbox": [ + 130, + 202, + 482, + 370 + ], + "type": "inline_equation", + "content": "-2.6\\mathrm{AP}_r" + }, + { + "bbox": [ + 130, + 202, + 482, + 370 + ], + "type": "text", + "content": " on novel classes, while overfitting to the base classes. The frozen backbone distillation ('w/ FD') leads to a significant improvement of " + }, + { + "bbox": [ + 130, + 202, + 482, + 370 + ], + "type": "inline_equation", + "content": "+3.5\\mathrm{AP}_r" + }, + { + "bbox": [ + 130, + 202, + 482, + 370 + ], + "type": "text", + "content": " even surpassing the frozen pretrained backbone (before detection finetuning), while maintaining performance on the base classes (" + }, + { + "bbox": [ + 130, + 202, + 482, + 370 + ], + "type": "inline_equation", + "content": "\\mathrm{AP}_c" + }, + { + "bbox": [ + 130, + 202, + 482, + 370 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 202, + 482, + 370 + ], + "type": "inline_equation", + "content": "\\mathrm{AP}_f" + }, + { + "bbox": [ + 130, + 202, + 482, + 370 + ], + "type": "text", + "content": "). Table 6b presents the open-vocabulary detection results where the frozen backbone distillation brings a clear gain of " + }, + { + "bbox": [ + 130, + 202, + 482, + 370 + ], + "type": "inline_equation", + "content": "+1.3\\sim 1.8\\mathrm{AP}_r" + }, + { + "bbox": [ + 130, + 202, + 482, + 370 + ], + "type": "text", + "content": ". These results highlight the efficacy of our frozen backbone distillation in detection finetuning (Sec. 3.4), as it effectively preserves the pretrained open-vocabulary knowledge while acquiring explicit region-text alignment through detection supervision." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 381, + 228, + 392 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 381, + 228, + 392 + ], + "spans": [ + { + "bbox": [ + 131, + 381, + 228, + 392 + ], + "type": "text", + "content": "4.3 Visualization" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 395, + 482, + 466 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 395, + 482, + 466 + ], + "spans": [ + { + "bbox": [ + 130, + 395, + 482, + 466 + ], + "type": "text", + "content": "In Fig. 3, we visualize the similarity maps between image features and query text embeddings. We compare the baseline [17] contrastive model's backbone features with our region-centric pretrained features. We select pyramid level 4 which has the same resolution as the backbone features and apply the Faster R-CNN head in a sliding window manner to obtain the dense feature map. We observe that region-centric pretraining captures more localized semantic" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 467, + 482, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 467, + 482, + 515 + ], + "spans": [ + { + "bbox": [ + 130, + 467, + 482, + 515 + ], + "type": "text", + "content": "Fig. 1 visualizes DITO outputs on LVIS novel categories and real-world out-of-distribution data from Ego4D [9]. Using the same DITO detector trained on " + }, + { + "bbox": [ + 130, + 467, + 482, + 515 + ], + "type": "inline_equation", + "content": "\\mathrm{LVIS}_{\\text {base }}" + }, + { + "bbox": [ + 130, + 467, + 482, + 515 + ], + "type": "text", + "content": ", we find that DITO effectively identifies many novel and unseen objects, even under significant domain shifts." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 527, + 221, + 540 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 527, + 221, + 540 + ], + "spans": [ + { + "bbox": [ + 131, + 527, + 221, + 540 + ], + "type": "text", + "content": "5 Conclusion" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": "We introduce DITO, a region-centric approach for open-vocabulary detection using large-scale image-text pairs. By integrating detection architecture onto the image backbone in CLIP pretraining, it learns locality-sensitive information without requiring pseudo labeling or box annotations. Furthermore, we propose a shifted-window learning method to mitigate the bias of the window attention pattern in CLIP ViT detectors. Experiments show that DITO outperforms the state-of-the-art by large margins on the LVIS benchmark, and is very competitive on the COCO benchmark and transfer detection. We hope this work would inspire the community to explore region-centric image-language pretraining for open-vocabulary localization tasks." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 221, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 221, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 221, + 101 + ], + "type": "text", + "content": "D. Kim et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 198, + 127 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 198, + 127 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 198, + 127 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 137, + 480, + 664 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 138, + 137, + 480, + 158 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 137, + 480, + 158 + ], + "spans": [ + { + "bbox": [ + 138, + 137, + 480, + 158 + ], + "type": "text", + "content": "1. Bai, Y., Chen, X., Kirillov, A., Yuille, A., Berg, A.C.: Point-level region contrast for object detection pre-training. In: CVPR. pp. 16061-16070 (June 2022) 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 159, + 480, + 190 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 159, + 480, + 190 + ], + "spans": [ + { + "bbox": [ + 138, + 159, + 480, + 190 + ], + "type": "text", + "content": "2. Chen, J., Zhu, D., Qian, G., Ghanem, B., Yan, Z., Zhu, C., Xiao, F., Elhoseiny, M., Culatana, S.C.: Exploring open-vocabulary semantic segmentation without human labels. arXiv preprint arXiv:2306.00450 (2023) 9" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 191, + 480, + 222 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 191, + 480, + 222 + ], + "spans": [ + { + "bbox": [ + 138, + 191, + 480, + 222 + ], + "type": "text", + "content": "3. Chen, T., Kornblith, S., Norouzi, M., Hinton, G.: A simple framework for contrastive learning of visual representations. arXiv preprint arXiv:2002.05709 (2020) 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 223, + 480, + 266 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 223, + 480, + 266 + ], + "spans": [ + { + "bbox": [ + 138, + 223, + 480, + 266 + ], + "type": "text", + "content": "4. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929 (2020) 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 267, + 480, + 298 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 267, + 480, + 298 + ], + "spans": [ + { + "bbox": [ + 138, + 267, + 480, + 298 + ], + "type": "text", + "content": "5. Du, Y., Wei, F., Zhang, Z., Shi, M., Gao, Y., Li, G.: Learning to prompt for open-vocabulary object detection with vision-language model. In: CVPR (2022) 1, 3, 10, 11" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 299, + 480, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 299, + 480, + 331 + ], + "spans": [ + { + "bbox": [ + 138, + 299, + 480, + 331 + ], + "type": "text", + "content": "6. Feng, C., Zhong, Y., Jie, Z., Chu, X., Ren, H., Wei, X., Xie, W., Ma, L.: Promptdet: Towards open-vocabulary detection using uncurated images. In: European Conference on Computer Vision. pp. 701-717. Springer (2022) 3, 10, 11, 12" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 332, + 480, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 332, + 480, + 364 + ], + "spans": [ + { + "bbox": [ + 138, + 332, + 480, + 364 + ], + "type": "text", + "content": "7. Gadre, S.Y., Ilharco, G., Fang, A., Hayase, J., Smyrnis, G., Nguyen, T., Marten, R., Wortsman, M., Ghosh, D., Zhang, J., et al.: Datacomp: In search of the next generation of multimodal datasets. arXiv preprint arXiv:2304.14108 (2023) 11" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 364, + 480, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 364, + 480, + 385 + ], + "spans": [ + { + "bbox": [ + 138, + 364, + 480, + 385 + ], + "type": "text", + "content": "8. Gao, M., Xing, C., Niebles, J.C., Li, J., Xu, R., Liu, W., Xiong, C.: Open vocabulary object detection with pseudo bounding-box labels. In: ECCV (2022) 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 386, + 480, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 386, + 480, + 418 + ], + "spans": [ + { + "bbox": [ + 138, + 386, + 480, + 418 + ], + "type": "text", + "content": "9. Grauman, K., Westbury, A., Byrne, E., Chavis, Z., Furnari, A., Girdhar, R., Hamburger, J., Jiang, H., Liu, M., Liu, X., et al.: Ego4d: Around the world in 3,000 hours of egocentric video. In: CVPR. pp. 18995-19012 (2022) 14" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 418, + 480, + 438 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 418, + 480, + 438 + ], + "spans": [ + { + "bbox": [ + 138, + 418, + 480, + 438 + ], + "type": "text", + "content": "0. Gu, X., Lin, T.Y., Kuo, W., Cui, Y.: Open-vocabulary object detection via vision and language knowledge distillation. In: ICLR (2022) 1, 3, 4, 6, 9, 10, 11" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 439, + 480, + 471 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 439, + 480, + 471 + ], + "spans": [ + { + "bbox": [ + 138, + 439, + 480, + 471 + ], + "type": "text", + "content": "1. He, K., Chen, X., Xie, S., Li, Y., Dólar, P., Girshick, R.: Masked autoencoders are scalable vision learners. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16000-16009 (2022) 4" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 472, + 474, + 482 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 472, + 474, + 482 + ], + "spans": [ + { + "bbox": [ + 138, + 472, + 474, + 482 + ], + "type": "text", + "content": "2. He, K., Gkioxari, G., Dollár, P., Girshick, R.: Mask r-cnn. In: ICCV (2017) 6, 7" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 483, + 480, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 483, + 480, + 525 + ], + "spans": [ + { + "bbox": [ + 138, + 483, + 480, + 525 + ], + "type": "text", + "content": "3. Huynh, D., Kuen, J., Lin, Z., Gu, J., Elhamifar, E.: Open-vocabulary instance segmentation via robust cross-modal pseudo-labeling. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7020-7031 (2022) 1, 11, 12" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 525, + 480, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 525, + 480, + 557 + ], + "spans": [ + { + "bbox": [ + 138, + 525, + 480, + 557 + ], + "type": "text", + "content": "4. Jia, C., Yang, Y., Xia, Y., Chen, Y.T., Parekh, Z., Pham, H., Le, Q.V., Sung, Y., Li, Z., Duerig, T.: Scaling up visual and vision-language representation learning with noisy text supervision. In: ICML (2021) 9" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 558, + 480, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 558, + 480, + 578 + ], + "spans": [ + { + "bbox": [ + 138, + 558, + 480, + 578 + ], + "type": "text", + "content": "5. Kaul, P., Xie, W., Zisserman, A.: Multi-modal classifiers for open-vocabulary object detection. arXiv preprint arXiv:2306.05493 (2023) 10" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 579, + 480, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 579, + 480, + 600 + ], + "spans": [ + { + "bbox": [ + 138, + 579, + 480, + 600 + ], + "type": "text", + "content": "6. Kim, D., Angelova, A., Kuo, W.: Contrastive feature masking open-vocabulary vision transformer. ICCV (2023) 2, 6, 7, 10, 11, 12" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 138, + 601, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 601, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 138, + 601, + 480, + 632 + ], + "type": "text", + "content": "7. Kim, D., Angelova, A., Kuo, W.: Region-aware pretraining for open-vocabulary object detection with vision transformers. In: Conference on Computer Vision and Pattern Recognition (CVPR) (2023) 2, 3, 4, 5, 6, 7, 9, 10, 11, 12, 13, 14" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 138, + 633, + 480, + 664 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 633, + 480, + 664 + ], + "spans": [ + { + "bbox": [ + 138, + 633, + 480, + 664 + ], + "type": "text", + "content": "8. Kuo, W., Cui, Y., Gu, X., Piergiovanni, A., Angelova, A.: F-vlm: Open-vocabulary object detection upon frozen vision and language models. ICLR (2023) 1, 3, 4, 6, 10, 11" + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 140, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 140, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 140, + 91, + 447, + 102 + ], + "type": "text", + "content": "Region-centric Image-Language Pretraining for Open-Vocabulary Detection" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 663 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "type": "text", + "content": "19. Li, L.H., Zhang, P., Zhang, H., Yang, J., Li, C., Zhong, Y., Wang, L., Yuan, L., Zhang, L., Hwang, J.N., Chang, K.W., Gao, J.: Grounded language-image pretraining. In: CVPR (2022) 2, 3, 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 150, + 480, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 150, + 480, + 171 + ], + "spans": [ + { + "bbox": [ + 132, + 150, + 480, + 171 + ], + "type": "text", + "content": "20. Li, Y., Mao, H., Girshick, R., He, K.: Exploring plain vision transformer backbones for object detection. In: ECCV (2022) 3, 4, 5, 6, 7, 8, 10" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 172, + 480, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 172, + 480, + 205 + ], + "spans": [ + { + "bbox": [ + 132, + 172, + 480, + 205 + ], + "type": "text", + "content": "21. Lin, T.Y., Dollar, P., Girshick, R., He, K., Hariharan, B., Belongie, S.: Feature pyramid networks for object detection. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 2117-2125 (2017) 5, 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 205, + 480, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 205, + 480, + 237 + ], + "spans": [ + { + "bbox": [ + 132, + 205, + 480, + 237 + ], + "type": "text", + "content": "22. Liu, S., Zeng, Z., Ren, T., Li, F., Zhang, H., Yang, J., Li, C., Yang, J., Su, H., Zhu, J., et al.: Grounding dino: Marrying dino with grounded pre-training for open-set object detection. arXiv preprint arXiv:2303.05499 (2023) 2, 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 238, + 480, + 269 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 238, + 480, + 269 + ], + "spans": [ + { + "bbox": [ + 132, + 238, + 480, + 269 + ], + "type": "text", + "content": "23. Liu, Z., Lin, Y., Cao, Y., Hu, H., Wei, Y., Zhang, Z., Lin, S., Guo, B.: Swin transformer: Hierarchical vision transformer using shifted windows. CVPR (2021) 3, 8, 12, 14" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 270, + 480, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 270, + 480, + 303 + ], + "spans": [ + { + "bbox": [ + 132, + 270, + 480, + 303 + ], + "type": "text", + "content": "24. Ma, C., Jiang, Y., Wen, X., Yuan, Z., Qi, X.: Codet: Co-occurrence guided region-word alignment for open-vocabulary object detection. Advances in Neural Information Processing Systems 36 (2024) 2, 4, 10, 11" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 304, + 480, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 304, + 480, + 335 + ], + "spans": [ + { + "bbox": [ + 132, + 304, + 480, + 335 + ], + "type": "text", + "content": "25. Minderer, M., Gritsenko, A., Houlsby, N.: Scaling open-vocabulary object detection. Advances in Neural Information Processing Systems 36 (2023) 2, 3, 4, 10, 14" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 336, + 480, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 336, + 480, + 380 + ], + "spans": [ + { + "bbox": [ + 132, + 336, + 480, + 380 + ], + "type": "text", + "content": "26. Minderer, M., Gritsenko, A., Stone, A., Neumann, M., Weissenborn, D., Dosovitskiy, A., Mahendran, A., Arnab, A., Dehghani, M., Shen, Z., Wang, X., Zhai, X., Kipf, T., Houlsby, N.: Simple open-vocabulary object detection with vision transformers. In: ECCV (2022) 3, 4, 10, 14" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 380, + 480, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 380, + 480, + 402 + ], + "spans": [ + { + "bbox": [ + 132, + 380, + 480, + 402 + ], + "type": "text", + "content": "27. Oord, A.v.d., Li, Y., Vinyals, O.: Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748 (2018) 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 403, + 480, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 403, + 480, + 435 + ], + "spans": [ + { + "bbox": [ + 132, + 403, + 480, + 435 + ], + "type": "text", + "content": "28. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., Krueger, G., Sutskever, I.: Learning transferable visual models from natural language supervision. In: ICML (2021) 1, 3, 5, 8, 9" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 436, + 480, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 436, + 480, + 468 + ], + "spans": [ + { + "bbox": [ + 132, + 436, + 480, + 468 + ], + "type": "text", + "content": "29. Rasheed, H., Maaz, M., Khattak, M.U., Khan, S., Khan, F.S.: Bridging the gap between object and image-level representations for open-vocabulary detection. arXiv preprint arXiv:2207.03482 (2022) 1, 10, 12" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 468, + 480, + 490 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 468, + 480, + 490 + ], + "spans": [ + { + "bbox": [ + 132, + 468, + 480, + 490 + ], + "type": "text", + "content": "30. Ren, S., He, K., Girshick, R.B., Sun, J.: Faster r-cnn: Towards real-time object detection with region proposal networks. In: NeurIPS (2015) 5, 6, 7, 10" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 491, + 480, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 491, + 480, + 533 + ], + "spans": [ + { + "bbox": [ + 132, + 491, + 480, + 533 + ], + "type": "text", + "content": "31. Schuhmann, C., Vencu, R., Beaumont, R., Kaczmarczyk, R., Mullis, C., Katta, A., Coombes, T., Jitsev, J., Komatsuzaki, A.: Laion-400m: Open dataset of clip-filtered 400 million image-text pairs. arXiv preprint arXiv:2111.02114 (2021) 3, 5, 9, 10, 11" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 534, + 480, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 534, + 480, + 566 + ], + "spans": [ + { + "bbox": [ + 132, + 534, + 480, + 566 + ], + "type": "text", + "content": "32. Shi, C., Yang, S.: Edadet: Open-vocabulary object detection using early dense alignment. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 15724-15734 (2023) 10, 11" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 567, + 480, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 567, + 480, + 599 + ], + "spans": [ + { + "bbox": [ + 132, + 567, + 480, + 599 + ], + "type": "text", + "content": "33. Wang, L., Liu, Y., Du, P., Ding, Z., Liao, Y., Qi, Q., Chen, B., Liu, S.: Object-aware distillation pyramid for open-vocabulary object detection. In: CVPR (2023) 10, 11" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 600, + 480, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 600, + 480, + 621 + ], + "spans": [ + { + "bbox": [ + 132, + 600, + 480, + 621 + ], + "type": "text", + "content": "34. Wei, F., Gao, Y., Wu, Z., Hu, H., Lin, S.: Aligning pretraining for detection via object-level contrastive learning. In: NeurIPS (2021) 4" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 622, + 480, + 663 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 622, + 480, + 663 + ], + "spans": [ + { + "bbox": [ + 132, + 622, + 480, + 663 + ], + "type": "text", + "content": "35. Wu, S., Zhang, W., Jin, S., Liu, W., Loy, C.C.: Aligning bag of regions for open-vocabulary object detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15254-15264 (2023) 2, 3, 10, 11, 12" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 221, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 221, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 221, + 100 + ], + "type": "text", + "content": "D. Kim et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 482, + 478 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 132, + 116, + 482, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 482, + 160 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 482, + 160 + ], + "type": "text", + "content": "36. Wu, X., Zhu, F., Zhao, R., Li, H.: Cora: Adapting clip for open-vocabulary detection with region prompting and anchor pre-matching. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7031-7040 (2023) 3, 10, 11" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 161, + 481, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 161, + 481, + 182 + ], + "spans": [ + { + "bbox": [ + 132, + 161, + 481, + 182 + ], + "type": "text", + "content": "37. Xiao, T., Reed, C.J., Wang, X., Keutzer, K., Darrell, T.: Region similarity representation learning. In: ICCV. pp. 10539-10548 (October 2021) 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 183, + 481, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 183, + 481, + 226 + ], + "spans": [ + { + "bbox": [ + 132, + 183, + 481, + 226 + ], + "type": "text", + "content": "38. Yao, L., Han, J., Liang, X., Xu, D., Zhang, W., Li, Z., Xu, H.: Detclipv2: Scalable open-vocabulary object detection pre-training via word-region alignment. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 23497-23506 (2023) 2, 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 227, + 481, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 227, + 481, + 258 + ], + "spans": [ + { + "bbox": [ + 132, + 227, + 481, + 258 + ], + "type": "text", + "content": "39. Yao, L., Han, J., Wen, Y., Liang, X., Xu, D., Zhang, W., Li, Z., Xu, C., , Xu, H.: Detclip: Dictionary-enriched visual-concept paralleled pre-training for open-world detection. In: arXiv:2209.09407 (2022) 2, 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 259, + 481, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 259, + 481, + 281 + ], + "spans": [ + { + "bbox": [ + 132, + 259, + 481, + 281 + ], + "type": "text", + "content": "40. Zang, Y., Li, W., Zhou, K., Huang, C., Loy, C.C.: Open-vocabulary DETR with conditional matching. In: ECCV. pp. 106-122. Springer (2022) 10, 11" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 282, + 481, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 282, + 481, + 303 + ], + "spans": [ + { + "bbox": [ + 132, + 282, + 481, + 303 + ], + "type": "text", + "content": "41. Zareian, A., Rosa, K.D., Hu, D.H., Chang, S.F.: Open-vocabulary object detection using captions. In: CVPR (2021) 1, 6, 11" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 304, + 481, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 304, + 481, + 346 + ], + "spans": [ + { + "bbox": [ + 132, + 304, + 481, + 346 + ], + "type": "text", + "content": "42. Zhang, H., Zhang, P., Hu, X., Chen, Y.C., Li, L., Dai, X., Wang, L., Yuan, L., Hwang, J.N., Gao, J.: Glipv2: Unifying localization and vision-language understanding. Advances in Neural Information Processing Systems 35, 36067-36080 (2022) 2, 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 347, + 481, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 347, + 481, + 380 + ], + "spans": [ + { + "bbox": [ + 132, + 347, + 481, + 380 + ], + "type": "text", + "content": "43. Zhao, S., Zhang, Z., Schulter, S., Zhao, L., Stathopoulos, A., Chandraker, M., Metaxas, D., et al.: Exploiting unlabeled data with vision and language models for object detection. In: ECCV (2022) 1, 10, 11, 12" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 380, + 481, + 412 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 380, + 481, + 412 + ], + "spans": [ + { + "bbox": [ + 132, + 380, + 481, + 412 + ], + "type": "text", + "content": "44. Zhong, Y., Yang, J., Zhang, P., Li, C., Codella, N., Li, L.H., Zhou, L., Dai, X., Yuan, L., Li, Y., Gao, J.: Regionclip: Region-based language-image pretraining. In: CVPR (2022) 1, 2, 3, 4, 10, 11, 12" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 413, + 481, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 413, + 481, + 445 + ], + "spans": [ + { + "bbox": [ + 132, + 413, + 481, + 445 + ], + "type": "text", + "content": "45. Zhong, Y., Wang, J., Wang, L., Peng, J., Wang, Y.X., Zhang, L.: Dap: Detection-aware pre-training with weak supervision. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 4537-4546 (2021) 2, 4" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 446, + 481, + 478 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 446, + 481, + 478 + ], + "spans": [ + { + "bbox": [ + 132, + 446, + 481, + 478 + ], + "type": "text", + "content": "46. Zhou, X., Girdhar, R., Joulin, A., Krahenbihl, P., Misra, I.: Detecting twenty-thousand classes using image-level supervision. In: ECCV (2022) 1, 2, 10, 11, 12" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 140, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 140, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 140, + 91, + 447, + 102 + ], + "type": "text", + "content": "Region-centric Image-Language Pretraining for Open-Vocabulary Detection" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/RegionDrag_ Fast Region-Based Image Editing with Diffusion Models/7e845ddf-5fd4-480c-a68c-61d4b586b518_content_list.json b/2024/RegionDrag_ Fast Region-Based Image Editing with Diffusion Models/7e845ddf-5fd4-480c-a68c-61d4b586b518_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..1c3a114a54ea7d774b1ca1775fa0c8ec570c7b83 --- /dev/null +++ b/2024/RegionDrag_ Fast Region-Based Image Editing with Diffusion Models/7e845ddf-5fd4-480c-a68c-61d4b586b518_content_list.json @@ -0,0 +1,1502 @@ +[ + { + "type": "text", + "text": "RegionDrag: Fast Region-Based Image Editing with Diffusion Models", + "text_level": 1, + "bbox": [ + 235, + 140, + 769, + 184 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jingyi Lu1, Xinghui Li2, and Kai Han1†", + "bbox": [ + 339, + 212, + 661, + 228 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1The University of Hong Kong 2University of Oxford", + "bbox": [ + 316, + 239, + 684, + 253 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "lujingyi@connect.hku.hk, xinghui@robots.ox.ac.uk, kaihanx@hku.hk", + "bbox": [ + 253, + 255, + 750, + 268 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/41b6b77d8b627d6b993d834af71986038a62eae5f00fc1e7ffbd63a6a2c74a6b.jpg", + "image_caption": [ + "Fig. 1: Comparison of editing results and latency between point-drag-based methods and our region-drag-based method. Our gradient-free, region-based framework reduces editing time from approximately one minute to about 1.5 seconds for $512 \\times 512$ resolution images, while producing results that better align with users' intentions." + ], + "image_footnote": [], + "bbox": [ + 243, + 273, + 756, + 444 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Point-drag-based image editing methods, like DragDiffusion, have attracted significant attention. However, point-drag-based approaches suffer from computational overhead and misinterpretation of user intentions, due to the sparsity of point-based editing instructions. In this paper, we propose a region-based copy-and-paste dragging method, RegionDrag, to overcome these limitations. RegionDrag allows users to express their editing instructions in the form of handle and target regions, enabling more precise control and alleviating ambiguity. In addition, region-based operations complete editing in one iteration and are much faster than point-drag-based methods. We also incorporate the attention-swapping technique for enhanced stability during editing. To validate our approach, we extend existing point-drag-based datasets with region-based dragging instructions. Experimental results demonstrate that RegionDrag outperforms existing point-drag-based approaches in terms of speed, accuracy, and alignment with user intentions. Remarkably, RegionDrag completes the edit on an image with a resolution of $512 \\times 512$ in less than 2 seconds, which is more than $100 \\times$ faster than DragDiffusion, while achieving better performance. Project page: https://visual-ai.github.io/regiondrag.", + "bbox": [ + 259, + 526, + 740, + 790 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Region-Based Dragging $\\cdot$ Image Editing $\\cdot$ Diffusion Models", + "bbox": [ + 261, + 803, + 740, + 816 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Corresponding author.", + "bbox": [ + 220, + 824, + 383, + 839 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 217, + 143, + 374, + 160 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Stable Diffusion (SD) [23] is a widely adopted text-to-image generative model known for its efficacy in producing high-fidelity images. As it is trained with millions of images, the vast amount of knowledge learned about images enables the model to edit existing images in a zero-shot manner. A recent line of work that has gained attention in the community is point-drag-based image editing using SD [15, 18, 24]. Point-drag-based methods first allow users to designate several points and drag them to desired positions. They then invert the latent representation of the input image to a particular timestep in the diffusion process and edit the image by enforcing similarities between local latent representations at the initial and final positions of the points. This is done either through optimization [15-17, 24] or direct copy-and-paste [18]. The edited latent representation is finally denoised and decoded to the edited image.", + "bbox": [ + 212, + 174, + 785, + 354 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Although point-drag-based methods demonstrate encouraging results, they exhibit several limitations. First, as the editing solely relies on dragging sparse points, they have to interpolate dozens of intermediate points along the dragging directions and edit the image iteratively to avoid editing failures. This significantly slows down the speed of editing. Second, dragging points cannot always faithfully reflect the desired effect. This form of instruction is prone to being misinterpreted by the model, so the editing results may not fully align with users' actual intentions. For example, moving an object to the left and expanding the object to the left can equally be represented by dragging points to the left.", + "bbox": [ + 212, + 357, + 785, + 492 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Therefore, we propose an alternative form of dragging to address the above issues. Instead of relying on points, we propose a region-based copy-and-paste dragging method, RegionDrag. Specifically, we first invert the latent representation of the unedited image to a specific time step using DDPM inversion [28], an inverse process to the stochastic DDPM sampling [25] which benefits the image editing [18]. Users then draw a handle region and a target region (see Fig. 1), where the former is the part that users intend to drag and the latter illustrates desired positions that users intend to achieve. After that, we establish a dense mapping between the two regions using our proposed Region-to-Point Mapping algorithm, and the latent representation covered by the handle region is mapped to the target region according to the dense mapping. Finally, the edited latent representation goes through the DDPM sampling process, and the mapping operation is repeated at multiple time steps. In the meantime, we also employ the attention-swapping technique proposed in [2] to stabilize the editing. To evaluate RegionDrag, we expand two existing point-drag-based image editing datasets [18,24] by adding equivalent region-based dragging instructions to each image. Experiments demonstrate that RegionDrag is significantly faster than point-drag-based methods as it completes editing in a single iteration. In addition, the paste region offers clearer constraints than point inputs and reduces chances of misinterpretation, leading to more faithful editing results.", + "bbox": [ + 212, + 492, + 785, + 792 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our contributions can be summarized as follows: (1) We introduce a region-based image editing method to overcome the limitations of point-drag-based approaches, utilizing richer input context to better align the editing results with", + "bbox": [ + 212, + 795, + 785, + 839 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "Lu et al.", + "bbox": [ + 271, + 114, + 330, + 126 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "the users' intentions. (2) By employing a gradient-free copy-paste operation, our region-based image editing becomes significantly faster than existing methods (see Fig. 1), completing the dragging in one single iteration. (3) We extend two point-drag-based datasets with region-based dragging instructions to validate RegionDrag's effectiveness and benchmark region-based editing methods.", + "bbox": [ + 215, + 146, + 787, + 220 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 247, + 387, + 263 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Generative Models for Image Editing. Early advancements in image generation have been driven by generative adversarial networks (GANs) [6, 9-11]. However, their practical application in real-world image editing is limited by the diversity of GAN training data and the effectiveness of GAN inversion techniques [3, 4, 21, 27, 29]. The emergence of text-to-image diffusion models leads to novel image editing techniques that utilize text prompts to modify high-level characteristics such as style, motion, or object categories [1, 2, 5, 12]. Nevertheless, image editing methods based on user text prompts inherently struggle to manipulate images at the pixel level. Dragging methods [4, 14-17, 19, 24] aim to address this limitation by controlling the overall posture and shape of objects through iterative movement and tracking of one or multiple key points. RegionDrag introduces a region-based approach as a superior alternative, offering increased stability and efficiency for fine-grained image editing tasks.", + "bbox": [ + 215, + 280, + 785, + 476 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Image Editing by Dragging Points. When employing drag-based methods for image editing, users can manipulate images by designating pairs of handle points and target points. The point-drag-based methods are expected to produce an image that meets two criteria: (1) features at the handle points are relocated to the target points, and (2) the original identity of the edited object is maintained. DragGAN [19] first enables editing on GAN-generated images involving multiple dragging point pairs. Specifically, DragGAN decomposes the editing process into several iterations, alternating between motion supervision and point tracking. In the motion supervision phase, DragGAN optimizes the StyleGAN's latent code [10,11] using the distance loss between the initial and target positions of the handle points. Following each motion supervision iteration, DragGAN updates the positions of the handle points by employing a point-tracking technique. FreeDrag [14] switches the dragging from pixel space to feature space, enabling more stable and precise manipulation. DragDiffusion [24] extends the supervise-and-track framework of DragGAN [19] to diffusion models. It inverts the latent representation of the unedited image to a partially noisy status at a selected time step, drags point features by optimizing the latent representation, and denoises it. RotationDrag [15] further refines the DragDiffusion framework [24] for cases of rotating objects. Another line of research, including DragonDiffusion [16] and DiffEditor [17], draws inspiration from classifier-guided generation, extending feature dragging throughout the entire denoising process rather than confining it to a single timestep. A method similar to ours, named SDE-Drag [18], eliminates memory-intensive backpropagation found in [16, 17, 24] by copying and pasting points within the diffusion latent space. Although SDE-Drag demon-", + "bbox": [ + 215, + 478, + 785, + 840 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "RegionDrag: Fast Region-Based Image Editing with Diffusion Models", + "bbox": [ + 269, + 114, + 730, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "strates promising results, it remains time-consuming due to its reliance on interpolating intermediate points along the dragging path to stabilize the editing process. In this paper, we introduce a simple yet effective region-based framework that utilizes the rich context inherent in region pairs to complete all edits in just one editing step.", + "bbox": [ + 212, + 146, + 782, + 222 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Appearance Consistency for Drag Editing. SD-based image editing shifts the image's latent distribution to incorporate user-specified structural, layout, and shape changes. However, this process can also introduce undesirable changes and artifacts. Mitigating these side effects remains challenging. Existing methods [15,18,24] often require training the SD model with LoRA (Low-Rank Adaptation) [8] for each unedited image to maintain style and appearance consistency. While LoRA is parameter-efficient through the use of low-rank matrix decomposition, it requires extensive preparation time and may overly constrain the image's appearance, limiting further edits. Additionally, point-drag-based methods often necessitate masking the editing area to reduce input ambiguity and enhance image consistency. In contrast, in this paper, we incorporate a training-free approach called mutual self-attention control [2], which preserves the image's identity by leveraging keys and values in the self-attention blocks of the model. This approach is seamlessly integrated into the editing pipeline, eliminating the need for an additional module to maintain the image's identity. Furthermore, unlike point-drag-based methods that rely on a mask to confine the editing area, RegionDrag eliminates this step to provide a more convenient editing experience.", + "bbox": [ + 212, + 222, + 787, + 479 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3 RegionDrag", + "text_level": 1, + "bbox": [ + 215, + 502, + 370, + 520 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "RegionDrag enables users to input handle and target region pairs, which are then used for editing through two primary steps: (1) copying latent representations covered by handle regions and storing self-attention features during inversion, and (2) pasting copied latent representations to target positions and inserting stored self-attention features during denoising.", + "bbox": [ + 212, + 534, + 782, + 609 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "This section begins by reviewing the diffusion-based image editing pipeline and point-drag-based methods in Sec. 3.1. We then discuss the limitations of point-drag-based methods and introduce how our region-based input addresses these limitations in Sec. 3.2. Finally, Sec. 3.3 presents our editing pipeline used to process region-based inputs.", + "bbox": [ + 212, + 609, + 787, + 686 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1 Preliminary", + "text_level": 1, + "bbox": [ + 215, + 708, + 362, + 724 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Diffusion-based image editing involves two main stages: inversion and denoising. Unedited images are first gradually inverted to a specific timestep in the diffusion process. The editing is then performed on the images at this timestep and the images are finally denoised back to their original image space. The transition between timesteps in the diffusion process is governed by a sampling scheduler. A commonly used one is the DDIM scheduler [25]. When an image latent $z_{s}$ at timestep $s$ transitions to $z_{t}$ at timestep $t$ , it goes through:", + "bbox": [ + 212, + 733, + 787, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "Lu et al.", + "bbox": [ + 271, + 114, + 330, + 126 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} z _ {t} = f _ {s \\rightarrow t} (z _ {s}) \\\\ = \\sqrt {\\alpha_ {t}} \\left(\\frac {z _ {s} - \\sqrt {1 - \\alpha_ {s}} \\epsilon_ {\\theta} (z _ {s} , s , C)}{\\sqrt {\\alpha_ {s}}}\\right) + \\sqrt {1 - \\alpha_ {t} - \\sigma_ {s} ^ {2}} \\epsilon_ {\\theta} (z _ {s}, s, C) + \\sigma_ {s} w _ {s}, \\quad (1) \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 232, + 171, + 784, + 224 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\epsilon_{\\theta}(z_s,s,C)$ is the noise predicted by the diffusion model, $\\alpha_{t}$ is a monotonically decreasing noise scheduling function dependent on $t$ , $w_{s}$ is Gaussian noise, and $\\sigma_{s} = \\eta \\sqrt{(1 - \\alpha_{t}) / (1 - \\alpha_{s})}\\sqrt{1 - \\alpha_{s} / \\alpha_{t}}$ with $\\eta = 0$ . Since $\\eta$ is set to 0, the transition is a deterministic process and it is widely used in many methods [16,17,23,24]. If $\\eta > 0$ , the sampling becomes a stochastic process known as the DDPM sampling [7]. We choose DDPM over DDIM as [18] shows that DDPM's randomness reduces divergence between unedited and edited image distributions.", + "bbox": [ + 212, + 234, + 782, + 354 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Point-drag-based methods, such as DragDiffusion and SDE-Drag [18,24], edit an image by relocating its latent representations at user-designated $n$ handle points $h_{1:n}$ to corresponding target positions $t_{1:n}$ . Due to the limited context available in sparse points, point-drag-based methods divide drag editing into $K$ sub-steps. At sub-step $k$ , the algorithm drags latent representations from handle points $h_{k,1:n}$ to target positions $t_{k,1:n}$ , which are found by direct interpolation or an extra point matching step. DragDiffusion optimizes $z_t^k$ by minimizing the $\\ell_1$ -distance between UNet upblock features $F(z_t^k)$ at handle $h_{k,1:n}$ and target $t_{k,1:n}$ points, respectively denoted as $F(z_t^k)[h_{k,1:n}]$ and $F(z_t^k)[t_{k,1:n}]$ :", + "bbox": [ + 212, + 354, + 784, + 492 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nz _ {t} ^ {k + 1} = z _ {t} ^ {k} - \\eta \\cdot \\nabla_ {z _ {t} ^ {k}} \\| F \\left(z _ {t} ^ {k}\\right) [ h _ {k, 1: n} ] - F \\left(z _ {t} ^ {k}\\right) [ t _ {k, 1: n} ] \\| _ {1}. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 318, + 503, + 784, + 523 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "DragonDiffusion [16] and DiffEditor [17] are DragDiffusion variants with $K = 1$ and extend optimization across multiple diffusion timesteps. On the other hand, SDE-Drag copies the latent representations from handle points to target points using the copy-paste function CP (Eq. (3)), where $z[h]$ and $z[t]$ represent the latent code at positions $h$ and $t$ . This is then followed by a denoising and inversion cycle (Eq. (4)).", + "bbox": [ + 212, + 527, + 782, + 619 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {C P} \\left(z _ {1}, z _ {2}, h, t\\right) = \\left(z _ {2} [ t ] \\leftarrow z _ {1} [ h ]\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 385, + 633, + 784, + 651 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nz _ {t} ^ {k + 1} = f _ {0 \\rightarrow t} \\left(f _ {t \\rightarrow 0} \\left(\\mathrm {C P} \\left(z _ {t} ^ {k}, z _ {t} ^ {k}, h _ {k, 1: n}, t _ {k, 1: n}\\right)\\right)\\right). \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 336, + 670, + 784, + 688 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2 From Point-Based to Region-Based Dragging", + "text_level": 1, + "bbox": [ + 214, + 709, + 633, + 724 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Although point-drag-based methods offer an intuitive means of user input, the limited information derived from sparse points poses challenges for models during the editing process. Specifically, point instructions can result in two primary issues: input ambiguity and slow inference. First, point instructions are inherently ambiguous. One dragging action could correspond to multiple plausible editing effects. Consider a user attempting to elongate a bird's beak in an image, as depicted in Fig. 2. The user selects a point on the beak and drags it", + "bbox": [ + 212, + 734, + 782, + 839 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "RegionDrag: Fast Region-Based Image Editing with Diffusion Models", + "bbox": [ + 267, + 114, + 732, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/94c9037661634a68167ffe25c0e7bed943628ce3bbdf9ed3d2f5161584b96483.jpg", + "image_caption": [ + "Fig. 2: Overall comparison of point-based editing and region-based editing, exemplified by manipulating a bird's beak. The region-based approach is shown to provide a more user-friendly and less ambiguous editing experience." + ], + "image_footnote": [], + "bbox": [ + 245, + 143, + 377, + 271 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/c93ee77287bed885ab9db6fefd2d037c79f4e0a9f4d63032cb841232b2ba81db.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 388, + 143, + 460, + 267 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/fa8f65a9a38dc77d4547cf8f56d9bfe272c25561cda12d1047fbd19e4f4d3815.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 462, + 145, + 596, + 267 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/3045a75b20c1569f949fb5ecc3042adc562b3788ac87bd05d88bb3c71f5f7d81.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 607, + 146, + 681, + 268 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/f7bd6c89106aa01e0fc9439be1a3d792e0c0c72e993cf364695948334f45ab35.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 684, + 178, + 756, + 239 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "towards the upper-left corner. Point-drag-based methods, however, might misinterpret the user's goal as enlarging the beak or moving the entire bird to the left, rather than extending the beak as intended, leading to a misalignment between the user's intention and the model's output.", + "bbox": [ + 212, + 354, + 784, + 415 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Second, the complexity involved in point-drag-based editing requires considerable computational overhead. Point-drag-based editing is challenging because the model must deduce changes across the entire image from the motion of a single or a few points. To carry out this complex drag operation while preserving the object's identity, point-drag-based methods heavily rely on two computationally intensive steps: training a unique LoRA [8] for each image and breaking down the dragging process into a series of sub-steps. Particularly, LoRA helps the model maintain the original image's identity and step-by-step dragging boosts the chance of achieving desired editing effects. Otherwise, the editing results may suffer from significant identity distortion or void editing, as shown in Fig. 2. The root of this problem lies in that sparse points do not impose sufficient constraints on editing, so the model has to rely on LoRA to prevent distortion and iterative editing to provide some degree of additional supervision along the path of the dragging. Consequently, most point-drag-based methods require several minutes to edit one image, rendering them impractical for real-world applications.", + "bbox": [ + 212, + 417, + 787, + 657 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The simplest solution to these problems is to encourage users to provide a sufficient number of points. However, such an approach would result in users spending too much time on designating and dragging points. Therefore, we design a form of editing that is not only user-friendly but also provides more informative context to the model, thus avoiding instruction ambiguity, slow inference, and exhaustive efforts from users. Instead of relying on dragging points, we propose to use region-based operations, where users assign a handle region $H$ to indicate the area they wish to drag and a target region $T$ to illustrate the desired position they would like to achieve. We then establish a dense mapping between the two regions using our Region-to-Point Mapping algorithm and complete the editing by directly copying the latent representation covered by the handle region to the target region in one inversion and denoising cycle. Despite", + "bbox": [ + 212, + 659, + 787, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "Lu et al.", + "bbox": [ + 271, + 114, + 330, + 126 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/c2a87b96473ac1ef4358ea0c1963b73c3286a2c1cbc3f01cf8469647ceeb1d57.jpg", + "image_caption": [ + "Handle Region Target Region $\\longrightarrow$ Copy-Paste Inference", + "Fig. 3: General pipeline of our method. Rich context provided by the region pairs enables users to complete accurate edits in one inversion and denoising cycle. The latent representation of the handle region is copied and pasted onto the target region throughout multiple timesteps for drag editing. Keys and values within the self-attention blocks are reused to ensure image consistency." + ], + "image_footnote": [], + "bbox": [ + 246, + 162, + 756, + 308 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "the simplicity of this operation, it addresses ambiguity and overhead from two perspectives: (1) Region-based operation is more expressive and accurate than dragging points and would significantly alleviate the ambiguity. As demonstrated in Fig. 2, we express extending the bird's beak by simply drawing a longer beak, hence reducing ambiguity presented in point-drag-based inputs. (2) Each region corresponds to a large number of points after dense mapping, so it provides stronger constraints on editing results than sparse points. As a result, we do not have to interpolate intermediate points along the dragging path and crave extra supervision, and editing can be completed in one editing step. Moreover, the handle and target regions can vary in size and take arbitrary shapes allowing users to define them conveniently.", + "bbox": [ + 212, + 417, + 787, + 583 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.3 Editing Pipeline", + "text_level": 1, + "bbox": [ + 214, + 604, + 397, + 619 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We elaborate on our editing pipeline in this section. We first introduce the our region-based user input, followed by our Region-to-Point Mapping algorithm, and finally the main working pipeline.", + "bbox": [ + 212, + 628, + 784, + 672 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "User Input. The handle and target regions can be defined in two ways: (1) by entering vertices to form a polygon (e.g., a triangle or quadrilateral), or (2) by brushing out a flexible region using a brush tool. The choice of input form largely depends on the user's preferences. Vertices are ideal for editing well-defined shapes, like moving a window on a building; a brush tool is better suited for irregular shapes, like a curved road or human hair.", + "bbox": [ + 212, + 672, + 785, + 763 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Region-to-Point Mapping. To preserve the original spatial information when copying and pasting latent representation, we need to establish a dense mapping between the handle and target region. If the regions are confined to triangular or quadrilateral shapes, we can compute a transformation matrix using affine or perspective mappings. However, finding a similar transformation for brush-out", + "bbox": [ + 212, + 763, + 785, + 839 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "RegionDrag: Fast Region-Based Image Editing with Diffusion Models", + "bbox": [ + 269, + 114, + 732, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "code", + "sub_type": "algorithm", + "code_caption": [ + "Algorithm 1 Region-to-Point Mapping" + ], + "code_body": "Input: Handle region $H$ with $N_{H}$ handle points $\\{(x_i^H,y_i^H)\\}_{i = 1}^{N_H}$ , target region $T$ with $N_{T}$ target points $\\{(x_i^T,y_i^T)\\}_{i = 1}^{N_T}$ . \nOutput: List of mapped point pairs $P$ . \nRequire: Initialize empty list $P$ . \n1: for $(x,y)\\in \\{(x_i^T,y_i^T)\\}_{i = 1}^{N_T}$ do \n2: Horizontal scaling: \n3: $x^{\\prime}\\gets (x - \\min (x_{1:N_T}^T)) / (\\max (x_{1:N_T}^T) - \\min (x_{1:N_T}^T))$ \n4: $x^{\\prime}\\gets \\lfloor x^{\\prime}\\cdot (\\max (x_{1:N_H}^H) - \\min (x_{1:N_H}^H)) + \\min (x_{1:N_H}^H)\\rfloor$ \n5: Column-by-column vertical scaling: \n6: $y^\\prime \\gets (y - \\min (y_i^T\\mid x_i^T = x)) / (\\max (y_i^T\\mid x_i^T = x) - \\min (y_i^T\\mid x_i^T = x))$ \n7: $y^\\prime \\gets \\lfloor y^\\prime \\cdot (\\max (y_i^H\\mid x_i^H = x^\\prime) - \\min (y_i^H\\mid x_i^H = x^\\prime)) + \\min (y_i^H\\mid x_i^H = x^\\prime)\\rfloor$ \n8: Add point pair $((x',y'),(x,y))$ to $P$ \n9: end for \n10: return $P$", + "bbox": [ + 215, + 164, + 785, + 364 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "regions with arbitrary shapes is challenging. To address this issue, we propose an algorithm to numerically find the mapping between two regions.", + "bbox": [ + 212, + 400, + 784, + 430 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Let $\\{(x_i^H, y_i^H)\\}_{i=1}^{N_H}$ and $\\{(x_j^T, y_j^T)\\}_{j=1}^{N_T}$ be the sets of pixels in handle and target regions respectively. We begin by linearly scaling the target region's width to match the handle region's width. For each pixel $(x, y) \\in \\{(x_j^T, y_j^T)\\}_{j=1}^{N_T}$ in the target region, $x$ is adjusted by first subtracting the x-coordinate lower boundary $\\min(x_{1:N_T}^T)$ and is normalized using the target region's x-coordinate range $\\max(x_{1:N_T}^T) - \\min(x_{1:N_T}^T)$ . It is then scaled to the handle region's x-coordinate range $\\max(x_{1:N_H}^H) - \\min(x_{1:N_H}^H)$ , resulting in $x'$ .", + "bbox": [ + 212, + 431, + 784, + 542 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This step ensures that both regions have the same number of columns of pixels. Next, we map each target region's vertical pixel column to its corresponding handle region's column. Each point's $y$ in the target region is adjusted by subtracting the lower boundary of its vertical column $\\min(y_i^T \\mid x_i^T = x)$ and is normalized by its column's y-coordinate range $\\max(y_i^T \\mid x_i^T = x) - \\min(y_i^T \\mid x_i^T = x)$ , followed by being scaled to the corresponding handle region's column's y-coordinate range $\\max(y_i^H \\mid x_i^H = x') - \\min(y_i^H \\mid x_i^H = x')$ . The notation $f(\\cdot \\mid \\cdot)$ represents a function $f$ (e.g., min or max) that takes a variable and a condition, and returns the function's value for the variable among points satisfying the condition. The whole process is summarised in Algorithm 1.", + "bbox": [ + 212, + 544, + 785, + 696 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Main Pipeline. As illustrated in Fig. 3, RegionDrag utilizes the image editing pipeline mentioned in Sec. 3.1. Initially, the latent representation of the image $z_0$ is inverted to $z_{t'}$ , where $t'$ is a selected timestep prior to maximum timestep $T$ . Each intermediate step $z_0, z_1, \\ldots, z_{t'}$ is cached during inversion for future use. We then duplicate $z_{t'}$ and denote the copy as $z_{t'}'$ . The handle regions of $z_{t'}'$ are blended with Gaussian noise $\\varepsilon$ according to the blending function $r_{\\alpha}(z, H)$ defined in Eq. (5):", + "bbox": [ + 212, + 698, + 785, + 805 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\nr _ {\\alpha} (z, H) = (1 - H) \\cdot z + H \\cdot (\\sqrt {1 - \\alpha^ {2}} \\cdot z + \\alpha \\cdot \\varepsilon), \\quad \\varepsilon \\sim \\mathcal {N} (0, I), \\qquad (5)\n$$\n", + "text_format": "latex", + "bbox": [ + 267, + 821, + 785, + 842 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "Lu et al.", + "bbox": [ + 271, + 114, + 330, + 127 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $H$ is a binary mask, with the handle region assigned a value of 1 and $\\alpha$ is a blending coefficient that ranges from 0 to 1, which governs the strength of the blending effect. The noise $\\varepsilon$ is drawn from a Gaussian distribution with mean 0 and identity covariance matrix, denoted by $\\mathcal{N}(0,I)$ . If $\\alpha$ is set to a lower value, less noise is added, which preserves more of the original image's features and details in the handle regions. After blending with Gaussian noise, $z_{t'}'$ goes through the denoising process using the DDPM sampler, and the dragging is conducted in a copy-paste manner. At a denoising timestep $t$ , we extract latent representation within the handle region of $z_t$ and map it to the target region of $z_{t}'$ according to the dense mapping computed by either geometric transformation or Algorithm 1. In the meantime, we employ mutual self-attention control [2] to help maintain the identity of images. In brief, when passing through the self-attention module of the UNet, the key and value $(k_t', v_t')$ used when denoising $z_t'$ are replaced with those $(k_t, v_t)$ from $z_t$ . This allows the edited image to keep the layout and identity of the original image, thereby stabilizing the editing process. After $z_{t}'$ is gradually denoised to $z_0'$ (see Fig. 3), it is decoded to the edited image $x_0'$ . Although RegionDrag is introduced with a single handle-target region pair, it supports multiple pairs of input $\\{(H_i, T_i)\\}_{i=1}^n$ , allowing users to specify several modifications in a single editing session. Dense mapping is individually constructed for each region pair but collectively used in latent copy-paste operations. The entire pipeline is summarised in Algorithm 2.", + "bbox": [ + 212, + 146, + 787, + 465 + ], + "page_idx": 8 + }, + { + "type": "code", + "sub_type": "algorithm", + "code_caption": [ + "Algorithm 2 Editing Pipeline" + ], + "code_body": "Input: Image $x_0$ , handle and target region pairs $\\{(H_i, T_i)\\}_{i=1}^n$ . \nOutput: Edited image $x_0'$ . \nRequire: Diffusion function $f_{s \\to t}$ Eq. (1), VAE encoder $E$ , VAE decoder $D$ , region-to-point mapping function $m$ , copy-paste function $CP$ Eq. (3), handle resampling function $r_\\alpha$ Eq. (5), copy-paste time interval $(t', t'')$ . \n1: Preparation: \n2: $\\{(h_i, t_i)\\}_{i=1}^N \\gets m(\\{(H_i, T_i)\\}_{i=1}^n)$ \n3: $z_0 \\gets E(x_0)$ \n4: Inversion stage: \n5: for $t = 0$ to $t' - 1$ do \n6: $z_{t+1}, k_{t+1}, v_{t+1} \\gets f_{t \\to t+1}(z_t)$ \n7: end for \n8: $z_{t'}' \\gets r_\\alpha(z_{t'}, \\cup_{i=1}^n H_i)$ \n9: Denoising stage: \n10: for $t = t'$ down to 1 do \n11: if $t \\geq t''$ then \n12: $z_t' \\gets \\mathrm{CP}(z_t, h_{1:N}, z_t', t_{1:N})$ \n13: end if \n14: $z_{t-1}' \\gets f_{t \\to t-1}(z_t', k_t' \\gets k_t, v_t' \\gets v_t)$ \n15: end for \n16: $x_0' \\gets D(z_0')$ \n17: return $x_0'$", + "bbox": [ + 215, + 525, + 787, + 834 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "RegionDrag: Fast Region-Based Image Editing with Diffusion Models", + "bbox": [ + 269, + 114, + 730, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4 Experimental Results", + "text_level": 1, + "bbox": [ + 215, + 143, + 460, + 162 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.1 Datasets", + "text_level": 1, + "bbox": [ + 215, + 180, + 333, + 194 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "SDE-Drag [18] and DragDiffusion [24] have each introduced a dataset named DragBench to evaluate the performance of point-drag-based methods. For clarity, we refer to the one by SDE-Drag as DragBench-S, which contains 100 samples, and to the one by DragDiffusion as DragBench-D, which includes 205 samples. Each sample comprises an image, a descriptive prompt, an optional mask delineating the editing region, and at least one pair of points that reflects the user's intention. Empirically, DragBench-S offers relatively more detailed prompts and includes a mask in 56 of its samples, while DragBench-D is characterized by more complex intentions and features carefully crafted masks in each of its samples.", + "bbox": [ + 212, + 208, + 785, + 345 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "To evaluate region-based editing, we introduce two new benchmarks: DragBench-SR and DragBench-DR (R is short for 'Region'), which are modified versions of DragBench-S and DragBench-D, respectively. These benchmarks are consistent with their point-drag-based counterparts in terms of images, prompts, and masks but differ by reflecting the user's intention through regions instead of points. DragBench-SR and DragBench-DR consist of 8 and 23 polygon-annotated samples, respectively, with the remainders of each dataset using brushes. For a handle-target region pair, the median number of equivalent point pairs is 267 in DragBench-SR and 201 in DragBench-DR. We visualize the distribution of the number of transformed point pairs in Fig. 4.", + "bbox": [ + 212, + 345, + 785, + 496 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "During the annotation process, we ensure that the region pairs are aligned with the point pairs to achieve the same editing effect. To simulate how real users would typically draw handle-target region pairs, we do not meticulously draw each region. According to our observations, each region-based sample is annotated in approximately 10 seconds. This approach allows for a more realistic evaluation of the region-based editing methods while maintaining consistency with the point-based datasets.", + "bbox": [ + 212, + 497, + 785, + 603 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.2 Evaluation Metrics", + "text_level": 1, + "bbox": [ + 215, + 628, + 419, + 643 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "LPIPS: We follow [24] and use Learned Perceptual Image Patch Similarity (LPIPS) v0.1 [30] to measure the identity similarity between the edited image and the original image. LPIPS computes the AlexNet [13] feature distances between the image pairs. A high LPIPS score indicates that unexpected identity changes or artifacts occur due to editing. A lower LPIPS score suggests that the object's identity has been well-preserved during editing; however, it does not necessarily imply a better drag edit, as two identical images would yield an LPIPS score of 0.", + "bbox": [ + 212, + 657, + 785, + 777 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Mean Distance (MD): DragDiffusion [24] introduces the MD metric to assess how well an approach moves the handle points' content to the target points. To identify where the handle points have been moved to by the method, DragDiffusion employs DIFT [26] to find the most similar points to the handle points", + "bbox": [ + 212, + 779, + 785, + 840 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "Lu et al.", + "bbox": [ + 271, + 114, + 330, + 126 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/2c225e96570194b0602d6fb448e941cbee5e817a4df55a5019505fc936766d1c.jpg", + "image_caption": [ + "Fig.4: Log-transformed frequency distribution of equivalent point pair counts in DragBench-SR and DragBench-DR." + ], + "image_footnote": [], + "bbox": [ + 235, + 143, + 464, + 255 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/0c0d5ffacdd813f4c7936a20b1d678be40b7874d3126658456d0878e64320b3d.jpg", + "image_caption": [ + "Fig. 5: Illustration of MD computation. Dragged handle points will be searched within the search mask." + ], + "image_footnote": [], + "bbox": [ + 504, + 143, + 767, + 252 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "$h_{1:n}$ in the entire edited image and denotes them as $h_{1:n}^{\\prime}$ . It then uses the normalized Euclidean distance between DIFT-matched points and the true target positions $t_{1:n}$ as the metric. However, we believe that we only have to search the area around the handle points and their corresponding true target points for $h_{1:n}^{\\prime}$ instead of the entire image, to avoid DIFT mistakenly identifying irrelevant points in the images as $h^\\prime$ and excessively penalizing certain methods. Formally, given a dragging point pair $\\{h,t\\}$ , we define its searching mask as", + "bbox": [ + 212, + 353, + 784, + 460 + ], + "page_idx": 10 + }, + { + "type": "equation", + "text": "\n$$\nM (x) = \\left\\{ \\begin{array}{l l} 1 & \\text {i f} \\min (d (x, h), d (x, t)) < \\frac {d (h , t)}{\\sqrt {2}} \\\\ 0 & \\text {o t h e r w i s e} \\end{array} \\right., \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 333, + 470, + 784, + 510 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "where $\\mathrm{d}(\\cdot, \\cdot)$ calculates the normalized Euclidean distance between two points $(x_1, y_1)$ and $(x_2, y_2)$ by", + "bbox": [ + 214, + 521, + 782, + 554 + ], + "page_idx": 10 + }, + { + "type": "equation", + "text": "\n$$\nd \\left(\\left(x _ {1}, y _ {1}\\right), \\left(x _ {2}, y _ {2}\\right)\\right) = \\sqrt {\\left(\\frac {x _ {2} - x _ {1}}{W}\\right) ^ {2} + \\left(\\frac {y _ {2} - y _ {1}}{H}\\right) ^ {2}}, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 313, + 563, + 784, + 604 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "where $W$ and $H$ are width and height of the image. We provide examples of searching masks in Fig. 5. MD is then defined as the average of the normalized distances $\\bar{d}$ between the target points and the DIFT-matched points across all $n$ points:", + "bbox": [ + 212, + 613, + 787, + 672 + ], + "page_idx": 10 + }, + { + "type": "equation", + "text": "\n$$\n\\bar {d} = \\frac {1}{n} \\sum_ {i = 1} ^ {n} d \\left(t _ {i}, h _ {i} ^ {\\prime}\\right). \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 431, + 672, + 784, + 712 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4.3 Implementation Details", + "text_level": 1, + "bbox": [ + 215, + 739, + 455, + 753 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Our method is implemented in Python using the HuggingFace [22] and PyTorch [20] libraries. We employ Stable Diffusion v1-5 as our diffusion model with an image size of $512 \\times 512$ , being consistent with previous diffusion-based dragging methods. A DDPM sampler is utilized for both the inversion and denoising sampling processes, configured to use a total of 20 steps. The latent", + "bbox": [ + 212, + 763, + 787, + 840 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "RegionDrag: Fast Region-Based Image Editing with Diffusion Models", + "bbox": [ + 269, + 114, + 732, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/9a2c254e51053ef29ebc2c30affdbef9834d9e7059eba659716a863aad06bb29.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodDragBench-S(R)DragBench-D(R)
Time (↓)MD (↓)LPIPS (↓)MD (↓)LPIPS (↓)
SDE-Drag [18]126.17.512.48.114.9
DragDiffusion [24]177.77.018.06.711.5
DiffEditor [17]43.123.617.622.110.9
Ours1.56.49.96.69.2
", + "bbox": [ + 274, + 143, + 722, + 243 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Table 1: Comparisons of our method with baseline methods using MD(×100) and LPIPS(×100) metrics on DragBench-S(R) and DragBench-D(R) datasets. The time is measured in seconds and averaged across both datasets. The image size is 512×512.", + "bbox": [ + 217, + 253, + 782, + 295 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "representation is inverted to timestep $t' = 500$ out of the total 1000 steps in SD1-5. Consequently, we perform 10 inversion steps and 10 denoising steps. Mutual self-attention control is enabled throughout all timesteps, and the latent copy-paste operation is terminated at $t'' = 200$ . The noise weight $\\alpha$ is set to 1. All experimental results are obtained on an NVIDIA Tesla V100 GPU.", + "bbox": [ + 212, + 328, + 784, + 402 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "4.4Baselines", + "text_level": 1, + "bbox": [ + 215, + 431, + 338, + 446 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "We compare RegionDrag with point-drag-based diffusion methods, including DragDiffusion [24], SDE-Drag [18], and DiffEditor [17]. GAN-based methods are excluded due to their limitations in editing the diverse images in the DragBench-S and DragBench-D datasets, as they require domain-specific StyleGAN checkpoints. Diffusion-based methods, which outperform GAN-based methods in editing tasks, are better suited for our evaluation. Execution times are averaged across both datasets, and all methods are tested on the same device using publicly released code.", + "bbox": [ + 212, + 463, + 784, + 583 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "4.5 Quantitative Evaluation", + "text_level": 1, + "bbox": [ + 215, + 612, + 459, + 628 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "To quantitatively evaluate the editing performance of the methods, we employ LPIPS and Mean Distance as metrics, multiplying both by 100 for illustration purposes. As demonstrated in Table 1, RegionDrag significantly outperforms those computationally expensive point-based methods on both DragBench-S(R) and DragBench-D(R) datasets. These results highlight RegionDrag's superior performance in maintaining image consistency while achieving competitive editing results across different datasets. In addition to its effectiveness, RegionDrag also excels in terms of efficiency. RegionDrag achieves fast inference speed, requiring approximately 1.5 seconds to edit a $512 \\times 512$ image, which is 20 times faster than the second-fastest method and 100 times faster than DragDiffusion [24]. The inference time for RegionDrag is comparable to generating an image with 20 steps using SD1-5, given that the copy-paste operations introduce negligible computational overhead.", + "bbox": [ + 212, + 643, + 785, + 839 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "Lu et al.", + "bbox": [ + 271, + 114, + 330, + 126 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/2a9629d4d04440744af6eccd89f8ad7634322d4fe4891251a15f095350ffaab4.jpg", + "image_caption": [ + "Fig. 6: Qualitative comparisons with baseline methods. Handle regions and target regions are respectively denoted by red and blue masks." + ], + "image_footnote": [], + "bbox": [ + 218, + 143, + 784, + 472 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "4.6 Qualitative Results", + "text_level": 1, + "bbox": [ + 215, + 542, + 419, + 556 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Figure 6 compares examples of point-drag-based and region-based editing inputs and their corresponding results, demonstrating the effectiveness of Region-Drag. Our region-based method utilizes the comprehensive context provided by annotated regions to target desired modifications while preserving the overall coherence of the image, outperforming point-drag-based editing methods.", + "bbox": [ + 214, + 566, + 787, + 643 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "4.7 Ablation Study", + "text_level": 1, + "bbox": [ + 215, + 664, + 390, + 679 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "We argue that the sparsity of point inputs leads to inferior editing results. To quantitatively demonstrate this, we conducted tests on the DragBench-DR dataset by randomly selecting subsets of the equivalent transformed points within each sample and performing inference using these subsets. We gradually reduced the percentage of selected points to observe the impact on the MD metric. As illustrated in Fig. 7, the results exhibit a clear upward trend in MD as the percentage of utilized points decreases. This suggests that sparse point inputs provide weaker constraints on the output compared to region-based inputs, leading to unsatisfactory editing results. It confirms the benefits of employing region-based inputs in RegionDrag.", + "bbox": [ + 214, + 689, + 787, + 840 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "RegionDrag: Fast Region-Based Image Editing with Diffusion Models", + "bbox": [ + 269, + 114, + 732, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/0d3d76007f3f0d760792441ff9132a6ebc5ee77756ecffecacb995d38764d0a4.jpg", + "image_caption": [ + "Fig.7: Ablation study on the impact of the percentage of inputted transformed points." + ], + "image_footnote": [], + "bbox": [ + 217, + 146, + 433, + 300 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/936a0cecf9b3b355d6aebe1f627bff86121ea6a3bea97e90d75f1314a602f113.jpg", + "image_caption": [ + "Fig. 8: Qualitative examples illustrating the impact of multi-step copy-paste." + ], + "image_footnote": [], + "bbox": [ + 460, + 143, + 575, + 289 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/961aeb8ba451318f9b7a5996afb508550ce2094fddad7b33e4535d3a853f2d0c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 580, + 143, + 784, + 289 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "We copy and paste image's latent representation over a time interval during denoising. To validate this design, we compare it to copy-paste only at the initial denoising timestep. Figure 8 shows that editing just at the initial step can yield unpredictable results because the edits may be lost in subsequent denoising phases. Multi-step copy-paste solves this by providing extra guidance at smaller timesteps while preserving image fidelity.", + "bbox": [ + 212, + 386, + 787, + 479 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 500, + 359, + 516 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In this paper, we have introduced an efficient and effective region-based editing framework, RegionDrag, for high-fidelity image editing. Unlike existing approaches that utilize point-drag-based editing, RegionDrag reconsiders the editing problem from the region perspective. RegionDrag allows for editing in a single step through copying and pasting the latent representation and self-attention features of the image, which not only provides excellent efficiency but also achieves superior editing performance. Furthermore, we have introduced two new benchmarks, DragBench-SR and DragBench-DR based on existing datasets, for the evaluation of region-based editing. Experimental results have consistently demonstrated the superior efficiency and editing performance of our method.", + "bbox": [ + 212, + 532, + 787, + 684 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgments", + "text_level": 1, + "bbox": [ + 215, + 707, + 393, + 724 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Kai Han and Jingyi Lu would like to acknowledge the support of the Hong Kong Research Grants Council - General Research Fund (Grant No.: 17211024).", + "bbox": [ + 214, + 738, + 785, + 768 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "Lu et al.", + "bbox": [ + 271, + 114, + 330, + 126 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 215, + 143, + 321, + 159 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Brooks, T., Holynski, A., Efros, A.A.: Instructpix2pix: Learning to follow image editing instructions. In: CVPR (2023)", + "2. Cao, M., Wang, X., Qi, Z., Shan, Y., Qie, X., Zheng, Y.: Masactrl: Tuning-free mutual self-attention control for consistent image synthesis and editing. arXiv preprint arXiv:2304.08465 (2023)", + "3. Creswell, A., Bharath, A.A.: Inverting the generator of a generative adversarial network. IEEE TNNLS (2018)", + "4. Endo, Y.: User-controllable latent transformer for stylegan image layout editing. In: CGF (2022)", + "5. Epstein, D., Jabri, A., Poole, B., Efros, A., Holynski, A.: Diffusion self-guidance for controllable image generation. In: NeurIPS (2024)", + "6. Goodfellow, I., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., Courville, A., Bengio, Y.: Generative adversarial nets. In: NeurIPS (2014)", + "7. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. In: NeurIPS (2020)", + "8. Hu, E.J., Shen, Y., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., Chen, W.: Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685 (2021)", + "9. Kang, M., Zhu, J.Y., Zhang, R., Park, J., Shechtman, E., Paris, S., Park, T.: Scaling up gans for text-to-image synthesis. In: CVPR (2023)", + "0. Karras, T., Aittala, M., Laine, S., Härkönen, E., Hellsten, J., Lehtinen, J., Aila, T.: Alias-free generative adversarial networks. In: NeurIPS (2021)", + "1. Karras, T., Laine, S., Aittala, M., Hellsten, J., Lehtinen, J., Aila, T.: Analyzing and improving the image quality of stylegan. In: CVPR (2020)", + "2. Kawar, B., Zada, S., Lang, O., Tov, O., Chang, H., Dekel, T., Mosseri, I., Irani, M.: Imagic: Text-based real image editing with diffusion models. In: CVPR (2023)", + "3. Krizhevsky, A., Sutskever, I., Hinton, G.E.: Imagenet classification with deep convolutional neural networks. NeurIPS (2012)", + "4. Ling, P., Chen, L., Zhang, P., Chen, H., Jin, Y.: Freedrag: Point tracking is not you need for interactive point-based image editing. arXiv preprint arXiv:2307.04684 (2023)", + "5. Luo, M., Cheng, W., Yang, J.: Rotationdrag: Point-based image editing with rotated diffusion features. arXiv preprint arXiv:2401.06442 (2024)", + "6. Mou, C., Wang, X., Song, J., Shan, Y., Zhang, J.: Dragondiffusion: Enabling drag-style manipulation on diffusion models. arXiv preprint arXiv:2307.02421 (2023)", + "7. Mou, C., Wang, X., Song, J., Shan, Y., Zhang, J.: Diffeditor: Boosting accuracy and flexibility on diffusion-based image editing. arXiv preprint arXiv:2402.02583 (2024)", + "8. Nie, S., Guo, H.A., Lu, C., Zhou, Y., Zheng, C., Li, C.: The blessing of randomness: Sde beats ode in general diffusion-based image editing. arXiv preprint arXiv:2311.01410 (2023)", + "9. Pan, X., Tewari, A., Leimkuhler, T., Liu, L., Meka, A., Theobalt, C.: Drag your gan: Interactive point-based manipulation on the generative image manifold. In: ACM SIGGRAPH (2023)", + "20. Paszke, A., Gross, S., Massa, F., Lerer, A., Bradbury, J., Chanan, G., Killeen, T., Lin, Z., Gimelshein, N., Antiga, L., et al.: Pytorch: An imperative style, high-performance deep learning library. In: NeurIPS (2019)" + ], + "bbox": [ + 225, + 178, + 784, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "RegionDrag: Fast Region-Based Image Editing with Diffusion Models", + "bbox": [ + 269, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "21. Patashnik, O., Wu, Z., Shechtman, E., Cohen-Or, D., Lischinski, D.: Styleclip: Text-driven manipulation of stylegan imagery. In: ICCV (2021)", + "22. von Platen, P., Patil, S., Lozhkov, A., Cuenca, P., Lambert, N., Rasul, K., Davaadorj, M., Wolf, T.: Diffusers: State-of-the-art diffusion models. https://github.com/huggingface/diffusers (2022)", + "23. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: CVPR (2022)", + "24. Shi, Y., Xue, C., Pan, J., Zhang, W., Tan, V.Y., Bai, S.: Dragdiffusion: Harnessing diffusion models for interactive point-based image editing. arXiv preprint arXiv:2306.14435 (2023)", + "25. Song, J., Meng, C., Ermon, S.: Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502 (2020)", + "26. Tang, L., Jia, M., Wang, Q., Phoo, C.P., Hariharan, B.: Emergent correspondence from image diffusion. In: NeurIPS (2024)", + "27. Wang, T., Zhang, Y., Fan, Y., Wang, J., Chen, Q.: High-fidelity gan inversion for image attribute editing. In: CVPR (2022)", + "28. Wu, C.H., De la Torre, F.: Unifying diffusion models' latent space, with applications to cyclediffusion and guidance. arXiv preprint arXiv:2210.05559 (2022)", + "29. Xia, W., Zhang, Y., Yang, Y., Xue, J.H., Zhou, B., Yang, M.H.: Gan inversion: A survey. IEEE TPAMI (2022)", + "30. Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual metric. In: CVPR (2018)" + ], + "bbox": [ + 215, + 146, + 785, + 452 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "Lu et al.", + "bbox": [ + 271, + 114, + 330, + 126 + ], + "page_idx": 15 + } +] \ No newline at end of file diff --git a/2024/RegionDrag_ Fast Region-Based Image Editing with Diffusion Models/7e845ddf-5fd4-480c-a68c-61d4b586b518_model.json b/2024/RegionDrag_ Fast Region-Based Image Editing with Diffusion Models/7e845ddf-5fd4-480c-a68c-61d4b586b518_model.json new file mode 100644 index 0000000000000000000000000000000000000000..b418c5835570a35a3838b1cc88cb032c44aded75 --- /dev/null +++ b/2024/RegionDrag_ Fast Region-Based Image Editing with Diffusion Models/7e845ddf-5fd4-480c-a68c-61d4b586b518_model.json @@ -0,0 +1,1871 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.236, + 0.141, + 0.77, + 0.185 + ], + "angle": 0, + "content": "RegionDrag: Fast Region-Based Image Editing with Diffusion Models" + }, + { + "type": "text", + "bbox": [ + 0.34, + 0.213, + 0.663, + 0.229 + ], + "angle": 0, + "content": "Jingyi Lu1, Xinghui Li2, and Kai Han1†" + }, + { + "type": "text", + "bbox": [ + 0.318, + 0.24, + 0.686, + 0.255 + ], + "angle": 0, + "content": "1The University of Hong Kong 2University of Oxford" + }, + { + "type": "text", + "bbox": [ + 0.254, + 0.256, + 0.75, + 0.269 + ], + "angle": 0, + "content": "lujingyi@connect.hku.hk, xinghui@robots.ox.ac.uk, kaihanx@hku.hk" + }, + { + "type": "image", + "bbox": [ + 0.245, + 0.275, + 0.758, + 0.445 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.454, + 0.788, + 0.51 + ], + "angle": 0, + "content": "Fig. 1: Comparison of editing results and latency between point-drag-based methods and our region-drag-based method. Our gradient-free, region-based framework reduces editing time from approximately one minute to about 1.5 seconds for \\(512 \\times 512\\) resolution images, while producing results that better align with users' intentions." + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.527, + 0.741, + 0.791 + ], + "angle": 0, + "content": "Abstract. Point-drag-based image editing methods, like DragDiffusion, have attracted significant attention. However, point-drag-based approaches suffer from computational overhead and misinterpretation of user intentions, due to the sparsity of point-based editing instructions. In this paper, we propose a region-based copy-and-paste dragging method, RegionDrag, to overcome these limitations. RegionDrag allows users to express their editing instructions in the form of handle and target regions, enabling more precise control and alleviating ambiguity. In addition, region-based operations complete editing in one iteration and are much faster than point-drag-based methods. We also incorporate the attention-swapping technique for enhanced stability during editing. To validate our approach, we extend existing point-drag-based datasets with region-based dragging instructions. Experimental results demonstrate that RegionDrag outperforms existing point-drag-based approaches in terms of speed, accuracy, and alignment with user intentions. Remarkably, RegionDrag completes the edit on an image with a resolution of \\(512 \\times 512\\) in less than 2 seconds, which is more than \\(100 \\times\\) faster than DragDiffusion, while achieving better performance. Project page: https://visual-ai.github.io/regiondrag." + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.804, + 0.741, + 0.818 + ], + "angle": 0, + "content": "Keywords: Region-Based Dragging \\(\\cdot\\) Image Editing \\(\\cdot\\) Diffusion Models" + }, + { + "type": "page_footnote", + "bbox": [ + 0.221, + 0.825, + 0.384, + 0.84 + ], + "angle": 0, + "content": "† Corresponding author." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.116, + 0.331, + 0.127 + ], + "angle": 0, + "content": "Lu et al." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.375, + 0.161 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.175, + 0.787, + 0.356 + ], + "angle": 0, + "content": "Stable Diffusion (SD) [23] is a widely adopted text-to-image generative model known for its efficacy in producing high-fidelity images. As it is trained with millions of images, the vast amount of knowledge learned about images enables the model to edit existing images in a zero-shot manner. A recent line of work that has gained attention in the community is point-drag-based image editing using SD [15, 18, 24]. Point-drag-based methods first allow users to designate several points and drag them to desired positions. They then invert the latent representation of the input image to a particular timestep in the diffusion process and edit the image by enforcing similarities between local latent representations at the initial and final positions of the points. This is done either through optimization [15-17, 24] or direct copy-and-paste [18]. The edited latent representation is finally denoised and decoded to the edited image." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.358, + 0.787, + 0.493 + ], + "angle": 0, + "content": "Although point-drag-based methods demonstrate encouraging results, they exhibit several limitations. First, as the editing solely relies on dragging sparse points, they have to interpolate dozens of intermediate points along the dragging directions and edit the image iteratively to avoid editing failures. This significantly slows down the speed of editing. Second, dragging points cannot always faithfully reflect the desired effect. This form of instruction is prone to being misinterpreted by the model, so the editing results may not fully align with users' actual intentions. For example, moving an object to the left and expanding the object to the left can equally be represented by dragging points to the left." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.493, + 0.787, + 0.794 + ], + "angle": 0, + "content": "Therefore, we propose an alternative form of dragging to address the above issues. Instead of relying on points, we propose a region-based copy-and-paste dragging method, RegionDrag. Specifically, we first invert the latent representation of the unedited image to a specific time step using DDPM inversion [28], an inverse process to the stochastic DDPM sampling [25] which benefits the image editing [18]. Users then draw a handle region and a target region (see Fig. 1), where the former is the part that users intend to drag and the latter illustrates desired positions that users intend to achieve. After that, we establish a dense mapping between the two regions using our proposed Region-to-Point Mapping algorithm, and the latent representation covered by the handle region is mapped to the target region according to the dense mapping. Finally, the edited latent representation goes through the DDPM sampling process, and the mapping operation is repeated at multiple time steps. In the meantime, we also employ the attention-swapping technique proposed in [2] to stabilize the editing. To evaluate RegionDrag, we expand two existing point-drag-based image editing datasets [18,24] by adding equivalent region-based dragging instructions to each image. Experiments demonstrate that RegionDrag is significantly faster than point-drag-based methods as it completes editing in a single iteration. In addition, the paste region offers clearer constraints than point inputs and reduces chances of misinterpretation, leading to more faithful editing results." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.796, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Our contributions can be summarized as follows: (1) We introduce a region-based image editing method to overcome the limitations of point-drag-based approaches, utilizing richer input context to better align the editing results with" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.27, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "RegionDrag: Fast Region-Based Image Editing with Diffusion Models" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.147, + 0.788, + 0.222 + ], + "angle": 0, + "content": "the users' intentions. (2) By employing a gradient-free copy-paste operation, our region-based image editing becomes significantly faster than existing methods (see Fig. 1), completing the dragging in one single iteration. (3) We extend two point-drag-based datasets with region-based dragging instructions to validate RegionDrag's effectiveness and benchmark region-based editing methods." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.248, + 0.388, + 0.264 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.281, + 0.787, + 0.477 + ], + "angle": 0, + "content": "Generative Models for Image Editing. Early advancements in image generation have been driven by generative adversarial networks (GANs) [6, 9-11]. However, their practical application in real-world image editing is limited by the diversity of GAN training data and the effectiveness of GAN inversion techniques [3, 4, 21, 27, 29]. The emergence of text-to-image diffusion models leads to novel image editing techniques that utilize text prompts to modify high-level characteristics such as style, motion, or object categories [1, 2, 5, 12]. Nevertheless, image editing methods based on user text prompts inherently struggle to manipulate images at the pixel level. Dragging methods [4, 14-17, 19, 24] aim to address this limitation by controlling the overall posture and shape of objects through iterative movement and tracking of one or multiple key points. RegionDrag introduces a region-based approach as a superior alternative, offering increased stability and efficiency for fine-grained image editing tasks." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.479, + 0.787, + 0.841 + ], + "angle": 0, + "content": "Image Editing by Dragging Points. When employing drag-based methods for image editing, users can manipulate images by designating pairs of handle points and target points. The point-drag-based methods are expected to produce an image that meets two criteria: (1) features at the handle points are relocated to the target points, and (2) the original identity of the edited object is maintained. DragGAN [19] first enables editing on GAN-generated images involving multiple dragging point pairs. Specifically, DragGAN decomposes the editing process into several iterations, alternating between motion supervision and point tracking. In the motion supervision phase, DragGAN optimizes the StyleGAN's latent code [10,11] using the distance loss between the initial and target positions of the handle points. Following each motion supervision iteration, DragGAN updates the positions of the handle points by employing a point-tracking technique. FreeDrag [14] switches the dragging from pixel space to feature space, enabling more stable and precise manipulation. DragDiffusion [24] extends the supervise-and-track framework of DragGAN [19] to diffusion models. It inverts the latent representation of the unedited image to a partially noisy status at a selected time step, drags point features by optimizing the latent representation, and denoises it. RotationDrag [15] further refines the DragDiffusion framework [24] for cases of rotating objects. Another line of research, including DragonDiffusion [16] and DiffEditor [17], draws inspiration from classifier-guided generation, extending feature dragging throughout the entire denoising process rather than confining it to a single timestep. A method similar to ours, named SDE-Drag [18], eliminates memory-intensive backpropagation found in [16, 17, 24] by copying and pasting points within the diffusion latent space. Although SDE-Drag demon-" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.116, + 0.331, + 0.127 + ], + "angle": 0, + "content": "Lu et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.223 + ], + "angle": 0, + "content": "strates promising results, it remains time-consuming due to its reliance on interpolating intermediate points along the dragging path to stabilize the editing process. In this paper, we introduce a simple yet effective region-based framework that utilizes the rich context inherent in region pairs to complete all edits in just one editing step." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.223, + 0.788, + 0.481 + ], + "angle": 0, + "content": "Appearance Consistency for Drag Editing. SD-based image editing shifts the image's latent distribution to incorporate user-specified structural, layout, and shape changes. However, this process can also introduce undesirable changes and artifacts. Mitigating these side effects remains challenging. Existing methods [15,18,24] often require training the SD model with LoRA (Low-Rank Adaptation) [8] for each unedited image to maintain style and appearance consistency. While LoRA is parameter-efficient through the use of low-rank matrix decomposition, it requires extensive preparation time and may overly constrain the image's appearance, limiting further edits. Additionally, point-drag-based methods often necessitate masking the editing area to reduce input ambiguity and enhance image consistency. In contrast, in this paper, we incorporate a training-free approach called mutual self-attention control [2], which preserves the image's identity by leveraging keys and values in the self-attention blocks of the model. This approach is seamlessly integrated into the editing pipeline, eliminating the need for an additional module to maintain the image's identity. Furthermore, unlike point-drag-based methods that rely on a mask to confine the editing area, RegionDrag eliminates this step to provide a more convenient editing experience." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.503, + 0.371, + 0.521 + ], + "angle": 0, + "content": "3 RegionDrag" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.535, + 0.784, + 0.61 + ], + "angle": 0, + "content": "RegionDrag enables users to input handle and target region pairs, which are then used for editing through two primary steps: (1) copying latent representations covered by handle regions and storing self-attention features during inversion, and (2) pasting copied latent representations to target positions and inserting stored self-attention features during denoising." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.611, + 0.788, + 0.687 + ], + "angle": 0, + "content": "This section begins by reviewing the diffusion-based image editing pipeline and point-drag-based methods in Sec. 3.1. We then discuss the limitations of point-drag-based methods and introduce how our region-based input addresses these limitations in Sec. 3.2. Finally, Sec. 3.3 presents our editing pipeline used to process region-based inputs." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.709, + 0.363, + 0.725 + ], + "angle": 0, + "content": "3.1 Preliminary" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.734, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Diffusion-based image editing involves two main stages: inversion and denoising. Unedited images are first gradually inverted to a specific timestep in the diffusion process. The editing is then performed on the images at this timestep and the images are finally denoised back to their original image space. The transition between timesteps in the diffusion process is governed by a sampling scheduler. A commonly used one is the DDIM scheduler [25]. When an image latent \\( z_{s} \\) at timestep \\( s \\) transitions to \\( z_{t} \\) at timestep \\( t \\), it goes through:" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.269, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "RegionDrag: Fast Region-Based Image Editing with Diffusion Models" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "equation", + "bbox": [ + 0.233, + 0.172, + 0.785, + 0.226 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} z _ {t} = f _ {s \\rightarrow t} (z _ {s}) \\\\ = \\sqrt {\\alpha_ {t}} \\left(\\frac {z _ {s} - \\sqrt {1 - \\alpha_ {s}} \\epsilon_ {\\theta} (z _ {s} , s , C)}{\\sqrt {\\alpha_ {s}}}\\right) + \\sqrt {1 - \\alpha_ {t} - \\sigma_ {s} ^ {2}} \\epsilon_ {\\theta} (z _ {s}, s, C) + \\sigma_ {s} w _ {s}, \\quad (1) \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.235, + 0.784, + 0.355 + ], + "angle": 0, + "content": "where \\(\\epsilon_{\\theta}(z_s,s,C)\\) is the noise predicted by the diffusion model, \\(\\alpha_{t}\\) is a monotonically decreasing noise scheduling function dependent on \\(t\\), \\(w_{s}\\) is Gaussian noise, and \\(\\sigma_{s} = \\eta \\sqrt{(1 - \\alpha_{t}) / (1 - \\alpha_{s})}\\sqrt{1 - \\alpha_{s} / \\alpha_{t}}\\) with \\(\\eta = 0\\). Since \\(\\eta\\) is set to 0, the transition is a deterministic process and it is widely used in many methods [16,17,23,24]. If \\(\\eta > 0\\), the sampling becomes a stochastic process known as the DDPM sampling [7]. We choose DDPM over DDIM as [18] shows that DDPM's randomness reduces divergence between unedited and edited image distributions." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.356, + 0.785, + 0.493 + ], + "angle": 0, + "content": "Point-drag-based methods, such as DragDiffusion and SDE-Drag [18,24], edit an image by relocating its latent representations at user-designated \\( n \\) handle points \\( h_{1:n} \\) to corresponding target positions \\( t_{1:n} \\). Due to the limited context available in sparse points, point-drag-based methods divide drag editing into \\( K \\) sub-steps. At sub-step \\( k \\), the algorithm drags latent representations from handle points \\( h_{k,1:n} \\) to target positions \\( t_{k,1:n} \\), which are found by direct interpolation or an extra point matching step. DragDiffusion optimizes \\( z_t^k \\) by minimizing the \\( \\ell_1 \\)-distance between UNet upblock features \\( F(z_t^k) \\) at handle \\( h_{k,1:n} \\) and target \\( t_{k,1:n} \\) points, respectively denoted as \\( F(z_t^k)[h_{k,1:n}] \\) and \\( F(z_t^k)[t_{k,1:n}] \\):" + }, + { + "type": "equation", + "bbox": [ + 0.32, + 0.505, + 0.785, + 0.524 + ], + "angle": 0, + "content": "\\[\nz _ {t} ^ {k + 1} = z _ {t} ^ {k} - \\eta \\cdot \\nabla_ {z _ {t} ^ {k}} \\| F \\left(z _ {t} ^ {k}\\right) [ h _ {k, 1: n} ] - F \\left(z _ {t} ^ {k}\\right) [ t _ {k, 1: n} ] \\| _ {1}. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.529, + 0.784, + 0.62 + ], + "angle": 0, + "content": "DragonDiffusion [16] and DiffEditor [17] are DragDiffusion variants with \\( K = 1 \\) and extend optimization across multiple diffusion timesteps. On the other hand, SDE-Drag copies the latent representations from handle points to target points using the copy-paste function CP (Eq. (3)), where \\( z[h] \\) and \\( z[t] \\) represent the latent code at positions \\( h \\) and \\( t \\). This is then followed by a denoising and inversion cycle (Eq. (4))." + }, + { + "type": "equation", + "bbox": [ + 0.387, + 0.635, + 0.785, + 0.652 + ], + "angle": 0, + "content": "\\[\n\\mathrm {C P} \\left(z _ {1}, z _ {2}, h, t\\right) = \\left(z _ {2} [ t ] \\leftarrow z _ {1} [ h ]\\right), \\tag {3}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.338, + 0.671, + 0.785, + 0.689 + ], + "angle": 0, + "content": "\\[\nz _ {t} ^ {k + 1} = f _ {0 \\rightarrow t} \\left(f _ {t \\rightarrow 0} \\left(\\mathrm {C P} \\left(z _ {t} ^ {k}, z _ {t} ^ {k}, h _ {k, 1: n}, t _ {k, 1: n}\\right)\\right)\\right). \\tag {4}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.71, + 0.634, + 0.726 + ], + "angle": 0, + "content": "3.2 From Point-Based to Region-Based Dragging" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.784, + 0.84 + ], + "angle": 0, + "content": "Although point-drag-based methods offer an intuitive means of user input, the limited information derived from sparse points poses challenges for models during the editing process. Specifically, point instructions can result in two primary issues: input ambiguity and slow inference. First, point instructions are inherently ambiguous. One dragging action could correspond to multiple plausible editing effects. Consider a user attempting to elongate a bird's beak in an image, as depicted in Fig. 2. The user selects a point on the beak and drags it" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.331, + 0.127 + ], + "angle": 0, + "content": "Lu et al." + }, + { + "type": "image", + "bbox": [ + 0.246, + 0.144, + 0.379, + 0.272 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.389, + 0.145, + 0.462, + 0.268 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.464, + 0.146, + 0.597, + 0.268 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.609, + 0.147, + 0.683, + 0.269 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.685, + 0.179, + 0.757, + 0.24 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.283, + 0.788, + 0.325 + ], + "angle": 0, + "content": "Fig. 2: Overall comparison of point-based editing and region-based editing, exemplified by manipulating a bird's beak. The region-based approach is shown to provide a more user-friendly and less ambiguous editing experience." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.356, + 0.785, + 0.416 + ], + "angle": 0, + "content": "towards the upper-left corner. Point-drag-based methods, however, might misinterpret the user's goal as enlarging the beak or moving the entire bird to the left, rather than extending the beak as intended, leading to a misalignment between the user's intention and the model's output." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.418, + 0.788, + 0.658 + ], + "angle": 0, + "content": "Second, the complexity involved in point-drag-based editing requires considerable computational overhead. Point-drag-based editing is challenging because the model must deduce changes across the entire image from the motion of a single or a few points. To carry out this complex drag operation while preserving the object's identity, point-drag-based methods heavily rely on two computationally intensive steps: training a unique LoRA [8] for each image and breaking down the dragging process into a series of sub-steps. Particularly, LoRA helps the model maintain the original image's identity and step-by-step dragging boosts the chance of achieving desired editing effects. Otherwise, the editing results may suffer from significant identity distortion or void editing, as shown in Fig. 2. The root of this problem lies in that sparse points do not impose sufficient constraints on editing, so the model has to rely on LoRA to prevent distortion and iterative editing to provide some degree of additional supervision along the path of the dragging. Consequently, most point-drag-based methods require several minutes to edit one image, rendering them impractical for real-world applications." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.66, + 0.788, + 0.841 + ], + "angle": 0, + "content": "The simplest solution to these problems is to encourage users to provide a sufficient number of points. However, such an approach would result in users spending too much time on designating and dragging points. Therefore, we design a form of editing that is not only user-friendly but also provides more informative context to the model, thus avoiding instruction ambiguity, slow inference, and exhaustive efforts from users. Instead of relying on dragging points, we propose to use region-based operations, where users assign a handle region \\( H \\) to indicate the area they wish to drag and a target region \\( T \\) to illustrate the desired position they would like to achieve. We then establish a dense mapping between the two regions using our Region-to-Point Mapping algorithm and complete the editing by directly copying the latent representation covered by the handle region to the target region in one inversion and denoising cycle. Despite" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.27, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "RegionDrag: Fast Region-Based Image Editing with Diffusion Models" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "image_caption", + "bbox": [ + 0.325, + 0.145, + 0.668, + 0.157 + ], + "angle": 0, + "content": "Handle Region Target Region \\(\\longrightarrow\\) Copy-Paste Inference" + }, + { + "type": "image", + "bbox": [ + 0.248, + 0.163, + 0.758, + 0.309 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.319, + 0.785, + 0.389 + ], + "angle": 0, + "content": "Fig. 3: General pipeline of our method. Rich context provided by the region pairs enables users to complete accurate edits in one inversion and denoising cycle. The latent representation of the handle region is copied and pasted onto the target region throughout multiple timesteps for drag editing. Keys and values within the self-attention blocks are reused to ensure image consistency." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.418, + 0.788, + 0.584 + ], + "angle": 0, + "content": "the simplicity of this operation, it addresses ambiguity and overhead from two perspectives: (1) Region-based operation is more expressive and accurate than dragging points and would significantly alleviate the ambiguity. As demonstrated in Fig. 2, we express extending the bird's beak by simply drawing a longer beak, hence reducing ambiguity presented in point-drag-based inputs. (2) Each region corresponds to a large number of points after dense mapping, so it provides stronger constraints on editing results than sparse points. As a result, we do not have to interpolate intermediate points along the dragging path and crave extra supervision, and editing can be completed in one editing step. Moreover, the handle and target regions can vary in size and take arbitrary shapes allowing users to define them conveniently." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.605, + 0.398, + 0.621 + ], + "angle": 0, + "content": "3.3 Editing Pipeline" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.629, + 0.785, + 0.674 + ], + "angle": 0, + "content": "We elaborate on our editing pipeline in this section. We first introduce the our region-based user input, followed by our Region-to-Point Mapping algorithm, and finally the main working pipeline." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.786, + 0.764 + ], + "angle": 0, + "content": "User Input. The handle and target regions can be defined in two ways: (1) by entering vertices to form a polygon (e.g., a triangle or quadrilateral), or (2) by brushing out a flexible region using a brush tool. The choice of input form largely depends on the user's preferences. Vertices are ideal for editing well-defined shapes, like moving a window on a building; a brush tool is better suited for irregular shapes, like a curved road or human hair." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Region-to-Point Mapping. To preserve the original spatial information when copying and pasting latent representation, we need to establish a dense mapping between the handle and target region. If the regions are confined to triangular or quadrilateral shapes, we can compute a transformation matrix using affine or perspective mappings. However, finding a similar transformation for brush-out" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.331, + 0.128 + ], + "angle": 0, + "content": "Lu et al." + }, + { + "type": "code_caption", + "bbox": [ + 0.218, + 0.147, + 0.51, + 0.162 + ], + "angle": 0, + "content": "Algorithm 1 Region-to-Point Mapping" + }, + { + "type": "algorithm", + "bbox": [ + 0.217, + 0.165, + 0.787, + 0.365 + ], + "angle": 0, + "content": "Input: Handle region \\(H\\) with \\(N_{H}\\) handle points \\(\\{(x_i^H,y_i^H)\\}_{i = 1}^{N_H}\\), target region \\(T\\) with \\(N_{T}\\) target points \\(\\{(x_i^T,y_i^T)\\}_{i = 1}^{N_T}\\). \nOutput: List of mapped point pairs \\(P\\). \nRequire: Initialize empty list \\(P\\). \n1: for \\((x,y)\\in \\{(x_i^T,y_i^T)\\}_{i = 1}^{N_T}\\) do \n2: Horizontal scaling: \n3: \\(x^{\\prime}\\gets (x - \\min (x_{1:N_T}^T)) / (\\max (x_{1:N_T}^T) - \\min (x_{1:N_T}^T))\\) \n4: \\(x^{\\prime}\\gets \\lfloor x^{\\prime}\\cdot (\\max (x_{1:N_H}^H) - \\min (x_{1:N_H}^H)) + \\min (x_{1:N_H}^H)\\rfloor\\) \n5: Column-by-column vertical scaling: \n6: \\(y^\\prime \\gets (y - \\min (y_i^T\\mid x_i^T = x)) / (\\max (y_i^T\\mid x_i^T = x) - \\min (y_i^T\\mid x_i^T = x))\\) \n7: \\(y^\\prime \\gets \\lfloor y^\\prime \\cdot (\\max (y_i^H\\mid x_i^H = x^\\prime) - \\min (y_i^H\\mid x_i^H = x^\\prime)) + \\min (y_i^H\\mid x_i^H = x^\\prime)\\rfloor\\) \n8: Add point pair \\(((x',y'),(x,y))\\) to \\(P\\) \n9: end for \n10: return \\(P\\)" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.401, + 0.785, + 0.431 + ], + "angle": 0, + "content": "regions with arbitrary shapes is challenging. To address this issue, we propose an algorithm to numerically find the mapping between two regions." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.432, + 0.785, + 0.544 + ], + "angle": 0, + "content": "Let \\(\\{(x_i^H, y_i^H)\\}_{i=1}^{N_H}\\) and \\(\\{(x_j^T, y_j^T)\\}_{j=1}^{N_T}\\) be the sets of pixels in handle and target regions respectively. We begin by linearly scaling the target region's width to match the handle region's width. For each pixel \\((x, y) \\in \\{(x_j^T, y_j^T)\\}_{j=1}^{N_T}\\) in the target region, \\(x\\) is adjusted by first subtracting the x-coordinate lower boundary \\(\\min(x_{1:N_T}^T)\\) and is normalized using the target region's x-coordinate range \\(\\max(x_{1:N_T}^T) - \\min(x_{1:N_T}^T)\\). It is then scaled to the handle region's x-coordinate range \\(\\max(x_{1:N_H}^H) - \\min(x_{1:N_H}^H)\\), resulting in \\(x'\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.545, + 0.787, + 0.697 + ], + "angle": 0, + "content": "This step ensures that both regions have the same number of columns of pixels. Next, we map each target region's vertical pixel column to its corresponding handle region's column. Each point's \\(y\\) in the target region is adjusted by subtracting the lower boundary of its vertical column \\(\\min(y_i^T \\mid x_i^T = x)\\) and is normalized by its column's y-coordinate range \\(\\max(y_i^T \\mid x_i^T = x) - \\min(y_i^T \\mid x_i^T = x)\\), followed by being scaled to the corresponding handle region's column's y-coordinate range \\(\\max(y_i^H \\mid x_i^H = x') - \\min(y_i^H \\mid x_i^H = x')\\). The notation \\(f(\\cdot \\mid \\cdot)\\) represents a function \\(f\\) (e.g., min or max) that takes a variable and a condition, and returns the function's value for the variable among points satisfying the condition. The whole process is summarised in Algorithm 1." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.699, + 0.787, + 0.806 + ], + "angle": 0, + "content": "Main Pipeline. As illustrated in Fig. 3, RegionDrag utilizes the image editing pipeline mentioned in Sec. 3.1. Initially, the latent representation of the image \\( z_0 \\) is inverted to \\( z_{t'} \\), where \\( t' \\) is a selected timestep prior to maximum timestep \\( T \\). Each intermediate step \\( z_0, z_1, \\ldots, z_{t'} \\) is cached during inversion for future use. We then duplicate \\( z_{t'} \\) and denote the copy as \\( z_{t'}' \\). The handle regions of \\( z_{t'}' \\) are blended with Gaussian noise \\( \\varepsilon \\) according to the blending function \\( r_{\\alpha}(z, H) \\) defined in Eq. (5):" + }, + { + "type": "equation", + "bbox": [ + 0.268, + 0.822, + 0.786, + 0.843 + ], + "angle": 0, + "content": "\\[\nr _ {\\alpha} (z, H) = (1 - H) \\cdot z + H \\cdot (\\sqrt {1 - \\alpha^ {2}} \\cdot z + \\alpha \\cdot \\varepsilon), \\quad \\varepsilon \\sim \\mathcal {N} (0, I), \\qquad (5)\n\\]" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.27, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "RegionDrag: Fast Region-Based Image Editing with Diffusion Models" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.466 + ], + "angle": 0, + "content": "where \\( H \\) is a binary mask, with the handle region assigned a value of 1 and \\( \\alpha \\) is a blending coefficient that ranges from 0 to 1, which governs the strength of the blending effect. The noise \\( \\varepsilon \\) is drawn from a Gaussian distribution with mean 0 and identity covariance matrix, denoted by \\( \\mathcal{N}(0,I) \\). If \\( \\alpha \\) is set to a lower value, less noise is added, which preserves more of the original image's features and details in the handle regions. After blending with Gaussian noise, \\( z_{t'}' \\) goes through the denoising process using the DDPM sampler, and the dragging is conducted in a copy-paste manner. At a denoising timestep \\( t \\), we extract latent representation within the handle region of \\( z_t \\) and map it to the target region of \\( z_{t}' \\) according to the dense mapping computed by either geometric transformation or Algorithm 1. In the meantime, we employ mutual self-attention control [2] to help maintain the identity of images. In brief, when passing through the self-attention module of the UNet, the key and value \\( (k_t', v_t') \\) used when denoising \\( z_t' \\) are replaced with those \\( (k_t, v_t) \\) from \\( z_t \\). This allows the edited image to keep the layout and identity of the original image, thereby stabilizing the editing process. After \\( z_{t}' \\) is gradually denoised to \\( z_0' \\) (see Fig. 3), it is decoded to the edited image \\( x_0' \\). Although RegionDrag is introduced with a single handle-target region pair, it supports multiple pairs of input \\( \\{(H_i, T_i)\\}_{i=1}^n \\), allowing users to specify several modifications in a single editing session. Dense mapping is individually constructed for each region pair but collectively used in latent copy-paste operations. The entire pipeline is summarised in Algorithm 2." + }, + { + "type": "code_caption", + "bbox": [ + 0.218, + 0.508, + 0.443, + 0.524 + ], + "angle": 0, + "content": "Algorithm 2 Editing Pipeline" + }, + { + "type": "algorithm", + "bbox": [ + 0.217, + 0.526, + 0.789, + 0.835 + ], + "angle": 0, + "content": "Input: Image \\(x_0\\), handle and target region pairs \\(\\{(H_i, T_i)\\}_{i=1}^n\\). \nOutput: Edited image \\(x_0'\\). \nRequire: Diffusion function \\(f_{s \\to t}\\) Eq. (1), VAE encoder \\(E\\), VAE decoder \\(D\\), region-to-point mapping function \\(m\\), copy-paste function \\(CP\\) Eq. (3), handle resampling function \\(r_\\alpha\\) Eq. (5), copy-paste time interval \\((t', t'')\\). \n1: Preparation: \n2: \\(\\{(h_i, t_i)\\}_{i=1}^N \\gets m(\\{(H_i, T_i)\\}_{i=1}^n)\\) \n3: \\(z_0 \\gets E(x_0)\\) \n4: Inversion stage: \n5: for \\(t = 0\\) to \\(t' - 1\\) do \n6: \\(z_{t+1}, k_{t+1}, v_{t+1} \\gets f_{t \\to t+1}(z_t)\\) \n7: end for \n8: \\(z_{t'}' \\gets r_\\alpha(z_{t'}, \\cup_{i=1}^n H_i)\\) \n9: Denoising stage: \n10: for \\(t = t'\\) down to 1 do \n11: if \\(t \\geq t''\\) then \n12: \\(z_t' \\gets \\mathrm{CP}(z_t, h_{1:N}, z_t', t_{1:N})\\) \n13: end if \n14: \\(z_{t-1}' \\gets f_{t \\to t-1}(z_t', k_t' \\gets k_t, v_t' \\gets v_t)\\) \n15: end for \n16: \\(x_0' \\gets D(z_0')\\) \n17: return \\(x_0'\\)" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.116, + 0.331, + 0.127 + ], + "angle": 0, + "content": "Lu et al." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.145, + 0.461, + 0.163 + ], + "angle": 0, + "content": "4 Experimental Results" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.181, + 0.334, + 0.195 + ], + "angle": 0, + "content": "4.1 Datasets" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.209, + 0.787, + 0.346 + ], + "angle": 0, + "content": "SDE-Drag [18] and DragDiffusion [24] have each introduced a dataset named DragBench to evaluate the performance of point-drag-based methods. For clarity, we refer to the one by SDE-Drag as DragBench-S, which contains 100 samples, and to the one by DragDiffusion as DragBench-D, which includes 205 samples. Each sample comprises an image, a descriptive prompt, an optional mask delineating the editing region, and at least one pair of points that reflects the user's intention. Empirically, DragBench-S offers relatively more detailed prompts and includes a mask in 56 of its samples, while DragBench-D is characterized by more complex intentions and features carefully crafted masks in each of its samples." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.347, + 0.787, + 0.497 + ], + "angle": 0, + "content": "To evaluate region-based editing, we introduce two new benchmarks: DragBench-SR and DragBench-DR (R is short for 'Region'), which are modified versions of DragBench-S and DragBench-D, respectively. These benchmarks are consistent with their point-drag-based counterparts in terms of images, prompts, and masks but differ by reflecting the user's intention through regions instead of points. DragBench-SR and DragBench-DR consist of 8 and 23 polygon-annotated samples, respectively, with the remainders of each dataset using brushes. For a handle-target region pair, the median number of equivalent point pairs is 267 in DragBench-SR and 201 in DragBench-DR. We visualize the distribution of the number of transformed point pairs in Fig. 4." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.498, + 0.787, + 0.604 + ], + "angle": 0, + "content": "During the annotation process, we ensure that the region pairs are aligned with the point pairs to achieve the same editing effect. To simulate how real users would typically draw handle-target region pairs, we do not meticulously draw each region. According to our observations, each region-based sample is annotated in approximately 10 seconds. This approach allows for a more realistic evaluation of the region-based editing methods while maintaining consistency with the point-based datasets." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.629, + 0.421, + 0.644 + ], + "angle": 0, + "content": "4.2 Evaluation Metrics" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.658, + 0.787, + 0.779 + ], + "angle": 0, + "content": "LPIPS: We follow [24] and use Learned Perceptual Image Patch Similarity (LPIPS) v0.1 [30] to measure the identity similarity between the edited image and the original image. LPIPS computes the AlexNet [13] feature distances between the image pairs. A high LPIPS score indicates that unexpected identity changes or artifacts occur due to editing. A lower LPIPS score suggests that the object's identity has been well-preserved during editing; however, it does not necessarily imply a better drag edit, as two identical images would yield an LPIPS score of 0." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.787, + 0.841 + ], + "angle": 0, + "content": "Mean Distance (MD): DragDiffusion [24] introduces the MD metric to assess how well an approach moves the handle points' content to the target points. To identify where the handle points have been moved to by the method, DragDiffusion employs DIFT [26] to find the most similar points to the handle points" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.27, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "RegionDrag: Fast Region-Based Image Editing with Diffusion Models" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "image", + "bbox": [ + 0.236, + 0.145, + 0.465, + 0.256 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.232, + 0.269, + 0.493, + 0.325 + ], + "angle": 0, + "content": "Fig.4: Log-transformed frequency distribution of equivalent point pair counts in DragBench-SR and DragBench-DR." + }, + { + "type": "image", + "bbox": [ + 0.506, + 0.145, + 0.768, + 0.253 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.503, + 0.27, + 0.769, + 0.313 + ], + "angle": 0, + "content": "Fig. 5: Illustration of MD computation. Dragged handle points will be searched within the search mask." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.354, + 0.785, + 0.462 + ], + "angle": 0, + "content": "\\(h_{1:n}\\) in the entire edited image and denotes them as \\(h_{1:n}^{\\prime}\\). It then uses the normalized Euclidean distance between DIFT-matched points and the true target positions \\(t_{1:n}\\) as the metric. However, we believe that we only have to search the area around the handle points and their corresponding true target points for \\(h_{1:n}^{\\prime}\\) instead of the entire image, to avoid DIFT mistakenly identifying irrelevant points in the images as \\(h^\\prime\\) and excessively penalizing certain methods. Formally, given a dragging point pair \\(\\{h,t\\}\\), we define its searching mask as" + }, + { + "type": "equation", + "bbox": [ + 0.334, + 0.471, + 0.785, + 0.511 + ], + "angle": 0, + "content": "\\[\nM (x) = \\left\\{ \\begin{array}{l l} 1 & \\text {i f} \\min (d (x, h), d (x, t)) < \\frac {d (h , t)}{\\sqrt {2}} \\\\ 0 & \\text {o t h e r w i s e} \\end{array} \\right., \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.522, + 0.784, + 0.555 + ], + "angle": 0, + "content": "where \\( \\mathrm{d}(\\cdot, \\cdot) \\) calculates the normalized Euclidean distance between two points \\( (x_1, y_1) \\) and \\( (x_2, y_2) \\) by" + }, + { + "type": "equation", + "bbox": [ + 0.314, + 0.564, + 0.785, + 0.605 + ], + "angle": 0, + "content": "\\[\nd \\left(\\left(x _ {1}, y _ {1}\\right), \\left(x _ {2}, y _ {2}\\right)\\right) = \\sqrt {\\left(\\frac {x _ {2} - x _ {1}}{W}\\right) ^ {2} + \\left(\\frac {y _ {2} - y _ {1}}{H}\\right) ^ {2}}, \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.614, + 0.788, + 0.674 + ], + "angle": 0, + "content": "where \\( W \\) and \\( H \\) are width and height of the image. We provide examples of searching masks in Fig. 5. MD is then defined as the average of the normalized distances \\( \\bar{d} \\) between the target points and the DIFT-matched points across all \\( n \\) points:" + }, + { + "type": "equation", + "bbox": [ + 0.432, + 0.673, + 0.785, + 0.713 + ], + "angle": 0, + "content": "\\[\n\\bar {d} = \\frac {1}{n} \\sum_ {i = 1} ^ {n} d \\left(t _ {i}, h _ {i} ^ {\\prime}\\right). \\tag {8}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.74, + 0.457, + 0.755 + ], + "angle": 0, + "content": "4.3 Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Our method is implemented in Python using the HuggingFace [22] and PyTorch [20] libraries. We employ Stable Diffusion v1-5 as our diffusion model with an image size of \\( 512 \\times 512 \\), being consistent with previous diffusion-based dragging methods. A DDPM sampler is utilized for both the inversion and denoising sampling processes, configured to use a total of 20 steps. The latent" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.331, + 0.127 + ], + "angle": 0, + "content": "Lu et al." + }, + { + "type": "table", + "bbox": [ + 0.276, + 0.144, + 0.723, + 0.244 + ], + "angle": 0, + "content": "
MethodDragBench-S(R)DragBench-D(R)
Time (↓)MD (↓)LPIPS (↓)MD (↓)LPIPS (↓)
SDE-Drag [18]126.17.512.48.114.9
DragDiffusion [24]177.77.018.06.711.5
DiffEditor [17]43.123.617.622.110.9
Ours1.56.49.96.69.2
" + }, + { + "type": "table_caption", + "bbox": [ + 0.218, + 0.254, + 0.784, + 0.296 + ], + "angle": 0, + "content": "Table 1: Comparisons of our method with baseline methods using MD(×100) and LPIPS(×100) metrics on DragBench-S(R) and DragBench-D(R) datasets. The time is measured in seconds and averaged across both datasets. The image size is 512×512." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.329, + 0.785, + 0.404 + ], + "angle": 0, + "content": "representation is inverted to timestep \\( t' = 500 \\) out of the total 1000 steps in SD1-5. Consequently, we perform 10 inversion steps and 10 denoising steps. Mutual self-attention control is enabled throughout all timesteps, and the latent copy-paste operation is terminated at \\( t'' = 200 \\). The noise weight \\( \\alpha \\) is set to 1. All experimental results are obtained on an NVIDIA Tesla V100 GPU." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.433, + 0.339, + 0.447 + ], + "angle": 0, + "content": "4.4Baselines" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.464, + 0.785, + 0.584 + ], + "angle": 0, + "content": "We compare RegionDrag with point-drag-based diffusion methods, including DragDiffusion [24], SDE-Drag [18], and DiffEditor [17]. GAN-based methods are excluded due to their limitations in editing the diverse images in the DragBench-S and DragBench-D datasets, as they require domain-specific StyleGAN checkpoints. Diffusion-based methods, which outperform GAN-based methods in editing tasks, are better suited for our evaluation. Execution times are averaged across both datasets, and all methods are tested on the same device using publicly released code." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.613, + 0.46, + 0.629 + ], + "angle": 0, + "content": "4.5 Quantitative Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.644, + 0.787, + 0.84 + ], + "angle": 0, + "content": "To quantitatively evaluate the editing performance of the methods, we employ LPIPS and Mean Distance as metrics, multiplying both by 100 for illustration purposes. As demonstrated in Table 1, RegionDrag significantly outperforms those computationally expensive point-based methods on both DragBench-S(R) and DragBench-D(R) datasets. These results highlight RegionDrag's superior performance in maintaining image consistency while achieving competitive editing results across different datasets. In addition to its effectiveness, RegionDrag also excels in terms of efficiency. RegionDrag achieves fast inference speed, requiring approximately 1.5 seconds to edit a \\(512 \\times 512\\) image, which is 20 times faster than the second-fastest method and 100 times faster than DragDiffusion [24]. The inference time for RegionDrag is comparable to generating an image with 20 steps using SD1-5, given that the copy-paste operations introduce negligible computational overhead." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.27, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "RegionDrag: Fast Region-Based Image Editing with Diffusion Models" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "image", + "bbox": [ + 0.219, + 0.145, + 0.785, + 0.473 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.486, + 0.785, + 0.514 + ], + "angle": 0, + "content": "Fig. 6: Qualitative comparisons with baseline methods. Handle regions and target regions are respectively denoted by red and blue masks." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.543, + 0.421, + 0.558 + ], + "angle": 0, + "content": "4.6 Qualitative Results" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.568, + 0.788, + 0.645 + ], + "angle": 0, + "content": "Figure 6 compares examples of point-drag-based and region-based editing inputs and their corresponding results, demonstrating the effectiveness of Region-Drag. Our region-based method utilizes the comprehensive context provided by annotated regions to target desired modifications while preserving the overall coherence of the image, outperforming point-drag-based editing methods." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.665, + 0.391, + 0.68 + ], + "angle": 0, + "content": "4.7 Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.69, + 0.789, + 0.841 + ], + "angle": 0, + "content": "We argue that the sparsity of point inputs leads to inferior editing results. To quantitatively demonstrate this, we conducted tests on the DragBench-DR dataset by randomly selecting subsets of the equivalent transformed points within each sample and performing inference using these subsets. We gradually reduced the percentage of selected points to observe the impact on the MD metric. As illustrated in Fig. 7, the results exhibit a clear upward trend in MD as the percentage of utilized points decreases. This suggests that sparse point inputs provide weaker constraints on the output compared to region-based inputs, leading to unsatisfactory editing results. It confirms the benefits of employing region-based inputs in RegionDrag." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.116, + 0.331, + 0.127 + ], + "angle": 0, + "content": "Lu et al." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.147, + 0.434, + 0.301 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.315, + 0.436, + 0.357 + ], + "angle": 0, + "content": "Fig.7: Ablation study on the impact of the percentage of inputted transformed points." + }, + { + "type": "image", + "bbox": [ + 0.461, + 0.144, + 0.576, + 0.29 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.581, + 0.144, + 0.785, + 0.29 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.459, + 0.317, + 0.785, + 0.346 + ], + "angle": 0, + "content": "Fig. 8: Qualitative examples illustrating the impact of multi-step copy-paste." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.387, + 0.788, + 0.48 + ], + "angle": 0, + "content": "We copy and paste image's latent representation over a time interval during denoising. To validate this design, we compare it to copy-paste only at the initial denoising timestep. Figure 8 shows that editing just at the initial step can yield unpredictable results because the edits may be lost in subsequent denoising phases. Multi-step copy-paste solves this by providing extra guidance at smaller timesteps while preserving image fidelity." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.501, + 0.36, + 0.517 + ], + "angle": 0, + "content": "5 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.533, + 0.788, + 0.685 + ], + "angle": 0, + "content": "In this paper, we have introduced an efficient and effective region-based editing framework, RegionDrag, for high-fidelity image editing. Unlike existing approaches that utilize point-drag-based editing, RegionDrag reconsiders the editing problem from the region perspective. RegionDrag allows for editing in a single step through copying and pasting the latent representation and self-attention features of the image, which not only provides excellent efficiency but also achieves superior editing performance. Furthermore, we have introduced two new benchmarks, DragBench-SR and DragBench-DR based on existing datasets, for the evaluation of region-based editing. Experimental results have consistently demonstrated the superior efficiency and editing performance of our method." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.708, + 0.395, + 0.725 + ], + "angle": 0, + "content": "Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.739, + 0.787, + 0.77 + ], + "angle": 0, + "content": "Kai Han and Jingyi Lu would like to acknowledge the support of the Hong Kong Research Grants Council - General Research Fund (Grant No.: 17211024)." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.27, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "RegionDrag: Fast Region-Based Image Editing with Diffusion Models" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.145, + 0.323, + 0.16 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.179, + 0.785, + 0.207 + ], + "angle": 0, + "content": "1. Brooks, T., Holynski, A., Efros, A.A.: Instructpix2pix: Learning to follow image editing instructions. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.209, + 0.785, + 0.248 + ], + "angle": 0, + "content": "2. Cao, M., Wang, X., Qi, Z., Shan, Y., Qie, X., Zheng, Y.: Masactrl: Tuning-free mutual self-attention control for consistent image synthesis and editing. arXiv preprint arXiv:2304.08465 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.251, + 0.785, + 0.277 + ], + "angle": 0, + "content": "3. Creswell, A., Bharath, A.A.: Inverting the generator of a generative adversarial network. IEEE TNNLS (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.279, + 0.785, + 0.305 + ], + "angle": 0, + "content": "4. Endo, Y.: User-controllable latent transformer for stylegan image layout editing. In: CGF (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.308, + 0.785, + 0.334 + ], + "angle": 0, + "content": "5. Epstein, D., Jabri, A., Poole, B., Efros, A., Holynski, A.: Diffusion self-guidance for controllable image generation. In: NeurIPS (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.336, + 0.785, + 0.362 + ], + "angle": 0, + "content": "6. Goodfellow, I., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., Courville, A., Bengio, Y.: Generative adversarial nets. In: NeurIPS (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.364, + 0.785, + 0.39 + ], + "angle": 0, + "content": "7. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. In: NeurIPS (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.392, + 0.785, + 0.432 + ], + "angle": 0, + "content": "8. Hu, E.J., Shen, Y., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., Chen, W.: Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.434, + 0.785, + 0.46 + ], + "angle": 0, + "content": "9. Kang, M., Zhu, J.Y., Zhang, R., Park, J., Shechtman, E., Paris, S., Park, T.: Scaling up gans for text-to-image synthesis. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.462, + 0.785, + 0.489 + ], + "angle": 0, + "content": "0. Karras, T., Aittala, M., Laine, S., Härkönen, E., Hellsten, J., Lehtinen, J., Aila, T.: Alias-free generative adversarial networks. In: NeurIPS (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.491, + 0.785, + 0.517 + ], + "angle": 0, + "content": "1. Karras, T., Laine, S., Aittala, M., Hellsten, J., Lehtinen, J., Aila, T.: Analyzing and improving the image quality of stylegan. In: CVPR (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.519, + 0.785, + 0.545 + ], + "angle": 0, + "content": "2. Kawar, B., Zada, S., Lang, O., Tov, O., Chang, H., Dekel, T., Mosseri, I., Irani, M.: Imagic: Text-based real image editing with diffusion models. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.547, + 0.785, + 0.572 + ], + "angle": 0, + "content": "3. Krizhevsky, A., Sutskever, I., Hinton, G.E.: Imagenet classification with deep convolutional neural networks. NeurIPS (2012)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.575, + 0.785, + 0.615 + ], + "angle": 0, + "content": "4. Ling, P., Chen, L., Zhang, P., Chen, H., Jin, Y.: Freedrag: Point tracking is not you need for interactive point-based image editing. arXiv preprint arXiv:2307.04684 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.617, + 0.785, + 0.643 + ], + "angle": 0, + "content": "5. Luo, M., Cheng, W., Yang, J.: Rotationdrag: Point-based image editing with rotated diffusion features. arXiv preprint arXiv:2401.06442 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.646, + 0.785, + 0.672 + ], + "angle": 0, + "content": "6. Mou, C., Wang, X., Song, J., Shan, Y., Zhang, J.: Dragondiffusion: Enabling drag-style manipulation on diffusion models. arXiv preprint arXiv:2307.02421 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.674, + 0.785, + 0.713 + ], + "angle": 0, + "content": "7. Mou, C., Wang, X., Song, J., Shan, Y., Zhang, J.: Diffeditor: Boosting accuracy and flexibility on diffusion-based image editing. arXiv preprint arXiv:2402.02583 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.716, + 0.785, + 0.755 + ], + "angle": 0, + "content": "8. Nie, S., Guo, H.A., Lu, C., Zhou, Y., Zheng, C., Li, C.: The blessing of randomness: Sde beats ode in general diffusion-based image editing. arXiv preprint arXiv:2311.01410 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.758, + 0.785, + 0.797 + ], + "angle": 0, + "content": "9. Pan, X., Tewari, A., Leimkuhler, T., Liu, L., Meka, A., Theobalt, C.: Drag your gan: Interactive point-based manipulation on the generative image manifold. In: ACM SIGGRAPH (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.8, + 0.785, + 0.84 + ], + "angle": 0, + "content": "20. Paszke, A., Gross, S., Massa, F., Lerer, A., Bradbury, J., Chanan, G., Killeen, T., Lin, Z., Gimelshein, N., Antiga, L., et al.: Pytorch: An imperative style, high-performance deep learning library. In: NeurIPS (2019)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.179, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.116, + 0.331, + 0.127 + ], + "angle": 0, + "content": "Lu et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.175 + ], + "angle": 0, + "content": "21. Patashnik, O., Wu, Z., Shechtman, E., Cohen-Or, D., Lischinski, D.: Styleclip: Text-driven manipulation of stylegan imagery. In: ICCV (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.176, + 0.787, + 0.217 + ], + "angle": 0, + "content": "22. von Platen, P., Patil, S., Lozhkov, A., Cuenca, P., Lambert, N., Rasul, K., Davaadorj, M., Wolf, T.: Diffusers: State-of-the-art diffusion models. https://github.com/huggingface/diffusers (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.218, + 0.785, + 0.244 + ], + "angle": 0, + "content": "23. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: CVPR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.245, + 0.785, + 0.285 + ], + "angle": 0, + "content": "24. Shi, Y., Xue, C., Pan, J., Zhang, W., Tan, V.Y., Bai, S.: Dragdiffusion: Harnessing diffusion models for interactive point-based image editing. arXiv preprint arXiv:2306.14435 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.287, + 0.785, + 0.313 + ], + "angle": 0, + "content": "25. Song, J., Meng, C., Ermon, S.: Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.315, + 0.785, + 0.341 + ], + "angle": 0, + "content": "26. Tang, L., Jia, M., Wang, Q., Phoo, C.P., Hariharan, B.: Emergent correspondence from image diffusion. In: NeurIPS (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.342, + 0.785, + 0.37 + ], + "angle": 0, + "content": "27. Wang, T., Zhang, Y., Fan, Y., Wang, J., Chen, Q.: High-fidelity gan inversion for image attribute editing. In: CVPR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.371, + 0.785, + 0.397 + ], + "angle": 0, + "content": "28. Wu, C.H., De la Torre, F.: Unifying diffusion models' latent space, with applications to cyclediffusion and guidance. arXiv preprint arXiv:2210.05559 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.398, + 0.785, + 0.424 + ], + "angle": 0, + "content": "29. Xia, W., Zhang, Y., Yang, Y., Xue, J.H., Zhou, B., Yang, M.H.: Gan inversion: A survey. IEEE TPAMI (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.425, + 0.785, + 0.453 + ], + "angle": 0, + "content": "30. Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual metric. In: CVPR (2018)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.453 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/RegionDrag_ Fast Region-Based Image Editing with Diffusion Models/7e845ddf-5fd4-480c-a68c-61d4b586b518_origin.pdf b/2024/RegionDrag_ Fast Region-Based Image Editing with Diffusion Models/7e845ddf-5fd4-480c-a68c-61d4b586b518_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..804690b3f163956669db79eaa6a99230fc87e33f --- /dev/null +++ b/2024/RegionDrag_ Fast Region-Based Image Editing with Diffusion Models/7e845ddf-5fd4-480c-a68c-61d4b586b518_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5524af7aa33dbac1520d7579a996287bc9050fd076aa94085913e632f2634896 +size 1894625 diff --git a/2024/RegionDrag_ Fast Region-Based Image Editing with Diffusion Models/full.md b/2024/RegionDrag_ Fast Region-Based Image Editing with Diffusion Models/full.md new file mode 100644 index 0000000000000000000000000000000000000000..e88445dffb4d934006ecd2261e3fe35b43f059e7 --- /dev/null +++ b/2024/RegionDrag_ Fast Region-Based Image Editing with Diffusion Models/full.md @@ -0,0 +1,275 @@ +# RegionDrag: Fast Region-Based Image Editing with Diffusion Models + +Jingyi Lu1, Xinghui Li2, and Kai Han1† + +1The University of Hong Kong 2University of Oxford + +lujingyi@connect.hku.hk, xinghui@robots.ox.ac.uk, kaihanx@hku.hk + +![](images/41b6b77d8b627d6b993d834af71986038a62eae5f00fc1e7ffbd63a6a2c74a6b.jpg) +Fig. 1: Comparison of editing results and latency between point-drag-based methods and our region-drag-based method. Our gradient-free, region-based framework reduces editing time from approximately one minute to about 1.5 seconds for $512 \times 512$ resolution images, while producing results that better align with users' intentions. + +Abstract. Point-drag-based image editing methods, like DragDiffusion, have attracted significant attention. However, point-drag-based approaches suffer from computational overhead and misinterpretation of user intentions, due to the sparsity of point-based editing instructions. In this paper, we propose a region-based copy-and-paste dragging method, RegionDrag, to overcome these limitations. RegionDrag allows users to express their editing instructions in the form of handle and target regions, enabling more precise control and alleviating ambiguity. In addition, region-based operations complete editing in one iteration and are much faster than point-drag-based methods. We also incorporate the attention-swapping technique for enhanced stability during editing. To validate our approach, we extend existing point-drag-based datasets with region-based dragging instructions. Experimental results demonstrate that RegionDrag outperforms existing point-drag-based approaches in terms of speed, accuracy, and alignment with user intentions. Remarkably, RegionDrag completes the edit on an image with a resolution of $512 \times 512$ in less than 2 seconds, which is more than $100 \times$ faster than DragDiffusion, while achieving better performance. Project page: https://visual-ai.github.io/regiondrag. + +Keywords: Region-Based Dragging $\cdot$ Image Editing $\cdot$ Diffusion Models + +# 1 Introduction + +Stable Diffusion (SD) [23] is a widely adopted text-to-image generative model known for its efficacy in producing high-fidelity images. As it is trained with millions of images, the vast amount of knowledge learned about images enables the model to edit existing images in a zero-shot manner. A recent line of work that has gained attention in the community is point-drag-based image editing using SD [15, 18, 24]. Point-drag-based methods first allow users to designate several points and drag them to desired positions. They then invert the latent representation of the input image to a particular timestep in the diffusion process and edit the image by enforcing similarities between local latent representations at the initial and final positions of the points. This is done either through optimization [15-17, 24] or direct copy-and-paste [18]. The edited latent representation is finally denoised and decoded to the edited image. + +Although point-drag-based methods demonstrate encouraging results, they exhibit several limitations. First, as the editing solely relies on dragging sparse points, they have to interpolate dozens of intermediate points along the dragging directions and edit the image iteratively to avoid editing failures. This significantly slows down the speed of editing. Second, dragging points cannot always faithfully reflect the desired effect. This form of instruction is prone to being misinterpreted by the model, so the editing results may not fully align with users' actual intentions. For example, moving an object to the left and expanding the object to the left can equally be represented by dragging points to the left. + +Therefore, we propose an alternative form of dragging to address the above issues. Instead of relying on points, we propose a region-based copy-and-paste dragging method, RegionDrag. Specifically, we first invert the latent representation of the unedited image to a specific time step using DDPM inversion [28], an inverse process to the stochastic DDPM sampling [25] which benefits the image editing [18]. Users then draw a handle region and a target region (see Fig. 1), where the former is the part that users intend to drag and the latter illustrates desired positions that users intend to achieve. After that, we establish a dense mapping between the two regions using our proposed Region-to-Point Mapping algorithm, and the latent representation covered by the handle region is mapped to the target region according to the dense mapping. Finally, the edited latent representation goes through the DDPM sampling process, and the mapping operation is repeated at multiple time steps. In the meantime, we also employ the attention-swapping technique proposed in [2] to stabilize the editing. To evaluate RegionDrag, we expand two existing point-drag-based image editing datasets [18,24] by adding equivalent region-based dragging instructions to each image. Experiments demonstrate that RegionDrag is significantly faster than point-drag-based methods as it completes editing in a single iteration. In addition, the paste region offers clearer constraints than point inputs and reduces chances of misinterpretation, leading to more faithful editing results. + +Our contributions can be summarized as follows: (1) We introduce a region-based image editing method to overcome the limitations of point-drag-based approaches, utilizing richer input context to better align the editing results with + +the users' intentions. (2) By employing a gradient-free copy-paste operation, our region-based image editing becomes significantly faster than existing methods (see Fig. 1), completing the dragging in one single iteration. (3) We extend two point-drag-based datasets with region-based dragging instructions to validate RegionDrag's effectiveness and benchmark region-based editing methods. + +# 2 Related Work + +Generative Models for Image Editing. Early advancements in image generation have been driven by generative adversarial networks (GANs) [6, 9-11]. However, their practical application in real-world image editing is limited by the diversity of GAN training data and the effectiveness of GAN inversion techniques [3, 4, 21, 27, 29]. The emergence of text-to-image diffusion models leads to novel image editing techniques that utilize text prompts to modify high-level characteristics such as style, motion, or object categories [1, 2, 5, 12]. Nevertheless, image editing methods based on user text prompts inherently struggle to manipulate images at the pixel level. Dragging methods [4, 14-17, 19, 24] aim to address this limitation by controlling the overall posture and shape of objects through iterative movement and tracking of one or multiple key points. RegionDrag introduces a region-based approach as a superior alternative, offering increased stability and efficiency for fine-grained image editing tasks. + +Image Editing by Dragging Points. When employing drag-based methods for image editing, users can manipulate images by designating pairs of handle points and target points. The point-drag-based methods are expected to produce an image that meets two criteria: (1) features at the handle points are relocated to the target points, and (2) the original identity of the edited object is maintained. DragGAN [19] first enables editing on GAN-generated images involving multiple dragging point pairs. Specifically, DragGAN decomposes the editing process into several iterations, alternating between motion supervision and point tracking. In the motion supervision phase, DragGAN optimizes the StyleGAN's latent code [10,11] using the distance loss between the initial and target positions of the handle points. Following each motion supervision iteration, DragGAN updates the positions of the handle points by employing a point-tracking technique. FreeDrag [14] switches the dragging from pixel space to feature space, enabling more stable and precise manipulation. DragDiffusion [24] extends the supervise-and-track framework of DragGAN [19] to diffusion models. It inverts the latent representation of the unedited image to a partially noisy status at a selected time step, drags point features by optimizing the latent representation, and denoises it. RotationDrag [15] further refines the DragDiffusion framework [24] for cases of rotating objects. Another line of research, including DragonDiffusion [16] and DiffEditor [17], draws inspiration from classifier-guided generation, extending feature dragging throughout the entire denoising process rather than confining it to a single timestep. A method similar to ours, named SDE-Drag [18], eliminates memory-intensive backpropagation found in [16, 17, 24] by copying and pasting points within the diffusion latent space. Although SDE-Drag demon- + +strates promising results, it remains time-consuming due to its reliance on interpolating intermediate points along the dragging path to stabilize the editing process. In this paper, we introduce a simple yet effective region-based framework that utilizes the rich context inherent in region pairs to complete all edits in just one editing step. + +Appearance Consistency for Drag Editing. SD-based image editing shifts the image's latent distribution to incorporate user-specified structural, layout, and shape changes. However, this process can also introduce undesirable changes and artifacts. Mitigating these side effects remains challenging. Existing methods [15,18,24] often require training the SD model with LoRA (Low-Rank Adaptation) [8] for each unedited image to maintain style and appearance consistency. While LoRA is parameter-efficient through the use of low-rank matrix decomposition, it requires extensive preparation time and may overly constrain the image's appearance, limiting further edits. Additionally, point-drag-based methods often necessitate masking the editing area to reduce input ambiguity and enhance image consistency. In contrast, in this paper, we incorporate a training-free approach called mutual self-attention control [2], which preserves the image's identity by leveraging keys and values in the self-attention blocks of the model. This approach is seamlessly integrated into the editing pipeline, eliminating the need for an additional module to maintain the image's identity. Furthermore, unlike point-drag-based methods that rely on a mask to confine the editing area, RegionDrag eliminates this step to provide a more convenient editing experience. + +# 3 RegionDrag + +RegionDrag enables users to input handle and target region pairs, which are then used for editing through two primary steps: (1) copying latent representations covered by handle regions and storing self-attention features during inversion, and (2) pasting copied latent representations to target positions and inserting stored self-attention features during denoising. + +This section begins by reviewing the diffusion-based image editing pipeline and point-drag-based methods in Sec. 3.1. We then discuss the limitations of point-drag-based methods and introduce how our region-based input addresses these limitations in Sec. 3.2. Finally, Sec. 3.3 presents our editing pipeline used to process region-based inputs. + +# 3.1 Preliminary + +Diffusion-based image editing involves two main stages: inversion and denoising. Unedited images are first gradually inverted to a specific timestep in the diffusion process. The editing is then performed on the images at this timestep and the images are finally denoised back to their original image space. The transition between timesteps in the diffusion process is governed by a sampling scheduler. A commonly used one is the DDIM scheduler [25]. When an image latent $z_{s}$ at timestep $s$ transitions to $z_{t}$ at timestep $t$ , it goes through: + +$$ +\begin{array}{l} z _ {t} = f _ {s \rightarrow t} (z _ {s}) \\ = \sqrt {\alpha_ {t}} \left(\frac {z _ {s} - \sqrt {1 - \alpha_ {s}} \epsilon_ {\theta} (z _ {s} , s , C)}{\sqrt {\alpha_ {s}}}\right) + \sqrt {1 - \alpha_ {t} - \sigma_ {s} ^ {2}} \epsilon_ {\theta} (z _ {s}, s, C) + \sigma_ {s} w _ {s}, \quad (1) \\ \end{array} +$$ + +where $\epsilon_{\theta}(z_s,s,C)$ is the noise predicted by the diffusion model, $\alpha_{t}$ is a monotonically decreasing noise scheduling function dependent on $t$ , $w_{s}$ is Gaussian noise, and $\sigma_{s} = \eta \sqrt{(1 - \alpha_{t}) / (1 - \alpha_{s})}\sqrt{1 - \alpha_{s} / \alpha_{t}}$ with $\eta = 0$ . Since $\eta$ is set to 0, the transition is a deterministic process and it is widely used in many methods [16,17,23,24]. If $\eta > 0$ , the sampling becomes a stochastic process known as the DDPM sampling [7]. We choose DDPM over DDIM as [18] shows that DDPM's randomness reduces divergence between unedited and edited image distributions. + +Point-drag-based methods, such as DragDiffusion and SDE-Drag [18,24], edit an image by relocating its latent representations at user-designated $n$ handle points $h_{1:n}$ to corresponding target positions $t_{1:n}$ . Due to the limited context available in sparse points, point-drag-based methods divide drag editing into $K$ sub-steps. At sub-step $k$ , the algorithm drags latent representations from handle points $h_{k,1:n}$ to target positions $t_{k,1:n}$ , which are found by direct interpolation or an extra point matching step. DragDiffusion optimizes $z_t^k$ by minimizing the $\ell_1$ -distance between UNet upblock features $F(z_t^k)$ at handle $h_{k,1:n}$ and target $t_{k,1:n}$ points, respectively denoted as $F(z_t^k)[h_{k,1:n}]$ and $F(z_t^k)[t_{k,1:n}]$ : + +$$ +z _ {t} ^ {k + 1} = z _ {t} ^ {k} - \eta \cdot \nabla_ {z _ {t} ^ {k}} \| F \left(z _ {t} ^ {k}\right) [ h _ {k, 1: n} ] - F \left(z _ {t} ^ {k}\right) [ t _ {k, 1: n} ] \| _ {1}. \tag {2} +$$ + +DragonDiffusion [16] and DiffEditor [17] are DragDiffusion variants with $K = 1$ and extend optimization across multiple diffusion timesteps. On the other hand, SDE-Drag copies the latent representations from handle points to target points using the copy-paste function CP (Eq. (3)), where $z[h]$ and $z[t]$ represent the latent code at positions $h$ and $t$ . This is then followed by a denoising and inversion cycle (Eq. (4)). + +$$ +\mathrm {C P} \left(z _ {1}, z _ {2}, h, t\right) = \left(z _ {2} [ t ] \leftarrow z _ {1} [ h ]\right), \tag {3} +$$ + +$$ +z _ {t} ^ {k + 1} = f _ {0 \rightarrow t} \left(f _ {t \rightarrow 0} \left(\mathrm {C P} \left(z _ {t} ^ {k}, z _ {t} ^ {k}, h _ {k, 1: n}, t _ {k, 1: n}\right)\right)\right). \tag {4} +$$ + +# 3.2 From Point-Based to Region-Based Dragging + +Although point-drag-based methods offer an intuitive means of user input, the limited information derived from sparse points poses challenges for models during the editing process. Specifically, point instructions can result in two primary issues: input ambiguity and slow inference. First, point instructions are inherently ambiguous. One dragging action could correspond to multiple plausible editing effects. Consider a user attempting to elongate a bird's beak in an image, as depicted in Fig. 2. The user selects a point on the beak and drags it + +![](images/94c9037661634a68167ffe25c0e7bed943628ce3bbdf9ed3d2f5161584b96483.jpg) +Fig. 2: Overall comparison of point-based editing and region-based editing, exemplified by manipulating a bird's beak. The region-based approach is shown to provide a more user-friendly and less ambiguous editing experience. + +![](images/c93ee77287bed885ab9db6fefd2d037c79f4e0a9f4d63032cb841232b2ba81db.jpg) + +![](images/fa8f65a9a38dc77d4547cf8f56d9bfe272c25561cda12d1047fbd19e4f4d3815.jpg) + +![](images/3045a75b20c1569f949fb5ecc3042adc562b3788ac87bd05d88bb3c71f5f7d81.jpg) + +![](images/f7bd6c89106aa01e0fc9439be1a3d792e0c0c72e993cf364695948334f45ab35.jpg) + +towards the upper-left corner. Point-drag-based methods, however, might misinterpret the user's goal as enlarging the beak or moving the entire bird to the left, rather than extending the beak as intended, leading to a misalignment between the user's intention and the model's output. + +Second, the complexity involved in point-drag-based editing requires considerable computational overhead. Point-drag-based editing is challenging because the model must deduce changes across the entire image from the motion of a single or a few points. To carry out this complex drag operation while preserving the object's identity, point-drag-based methods heavily rely on two computationally intensive steps: training a unique LoRA [8] for each image and breaking down the dragging process into a series of sub-steps. Particularly, LoRA helps the model maintain the original image's identity and step-by-step dragging boosts the chance of achieving desired editing effects. Otherwise, the editing results may suffer from significant identity distortion or void editing, as shown in Fig. 2. The root of this problem lies in that sparse points do not impose sufficient constraints on editing, so the model has to rely on LoRA to prevent distortion and iterative editing to provide some degree of additional supervision along the path of the dragging. Consequently, most point-drag-based methods require several minutes to edit one image, rendering them impractical for real-world applications. + +The simplest solution to these problems is to encourage users to provide a sufficient number of points. However, such an approach would result in users spending too much time on designating and dragging points. Therefore, we design a form of editing that is not only user-friendly but also provides more informative context to the model, thus avoiding instruction ambiguity, slow inference, and exhaustive efforts from users. Instead of relying on dragging points, we propose to use region-based operations, where users assign a handle region $H$ to indicate the area they wish to drag and a target region $T$ to illustrate the desired position they would like to achieve. We then establish a dense mapping between the two regions using our Region-to-Point Mapping algorithm and complete the editing by directly copying the latent representation covered by the handle region to the target region in one inversion and denoising cycle. Despite + +![](images/c2a87b96473ac1ef4358ea0c1963b73c3286a2c1cbc3f01cf8469647ceeb1d57.jpg) +Handle Region Target Region $\longrightarrow$ Copy-Paste Inference +Fig. 3: General pipeline of our method. Rich context provided by the region pairs enables users to complete accurate edits in one inversion and denoising cycle. The latent representation of the handle region is copied and pasted onto the target region throughout multiple timesteps for drag editing. Keys and values within the self-attention blocks are reused to ensure image consistency. + +the simplicity of this operation, it addresses ambiguity and overhead from two perspectives: (1) Region-based operation is more expressive and accurate than dragging points and would significantly alleviate the ambiguity. As demonstrated in Fig. 2, we express extending the bird's beak by simply drawing a longer beak, hence reducing ambiguity presented in point-drag-based inputs. (2) Each region corresponds to a large number of points after dense mapping, so it provides stronger constraints on editing results than sparse points. As a result, we do not have to interpolate intermediate points along the dragging path and crave extra supervision, and editing can be completed in one editing step. Moreover, the handle and target regions can vary in size and take arbitrary shapes allowing users to define them conveniently. + +# 3.3 Editing Pipeline + +We elaborate on our editing pipeline in this section. We first introduce the our region-based user input, followed by our Region-to-Point Mapping algorithm, and finally the main working pipeline. + +User Input. The handle and target regions can be defined in two ways: (1) by entering vertices to form a polygon (e.g., a triangle or quadrilateral), or (2) by brushing out a flexible region using a brush tool. The choice of input form largely depends on the user's preferences. Vertices are ideal for editing well-defined shapes, like moving a window on a building; a brush tool is better suited for irregular shapes, like a curved road or human hair. + +Region-to-Point Mapping. To preserve the original spatial information when copying and pasting latent representation, we need to establish a dense mapping between the handle and target region. If the regions are confined to triangular or quadrilateral shapes, we can compute a transformation matrix using affine or perspective mappings. However, finding a similar transformation for brush-out + +Algorithm 1 Region-to-Point Mapping +Input: Handle region $H$ with $N_{H}$ handle points $\{(x_i^H,y_i^H)\}_{i = 1}^{N_H}$ , target region $T$ with $N_{T}$ target points $\{(x_i^T,y_i^T)\}_{i = 1}^{N_T}$ . +Output: List of mapped point pairs $P$ . +Require: Initialize empty list $P$ . +1: for $(x,y)\in \{(x_i^T,y_i^T)\}_{i = 1}^{N_T}$ do +2: Horizontal scaling: +3: $x^{\prime}\gets (x - \min (x_{1:N_T}^T)) / (\max (x_{1:N_T}^T) - \min (x_{1:N_T}^T))$ +4: $x^{\prime}\gets \lfloor x^{\prime}\cdot (\max (x_{1:N_H}^H) - \min (x_{1:N_H}^H)) + \min (x_{1:N_H}^H)\rfloor$ +5: Column-by-column vertical scaling: +6: $y^\prime \gets (y - \min (y_i^T\mid x_i^T = x)) / (\max (y_i^T\mid x_i^T = x) - \min (y_i^T\mid x_i^T = x))$ +7: $y^\prime \gets \lfloor y^\prime \cdot (\max (y_i^H\mid x_i^H = x^\prime) - \min (y_i^H\mid x_i^H = x^\prime)) + \min (y_i^H\mid x_i^H = x^\prime)\rfloor$ +8: Add point pair $((x',y'),(x,y))$ to $P$ +9: end for +10: return $P$ + +regions with arbitrary shapes is challenging. To address this issue, we propose an algorithm to numerically find the mapping between two regions. + +Let $\{(x_i^H, y_i^H)\}_{i=1}^{N_H}$ and $\{(x_j^T, y_j^T)\}_{j=1}^{N_T}$ be the sets of pixels in handle and target regions respectively. We begin by linearly scaling the target region's width to match the handle region's width. For each pixel $(x, y) \in \{(x_j^T, y_j^T)\}_{j=1}^{N_T}$ in the target region, $x$ is adjusted by first subtracting the x-coordinate lower boundary $\min(x_{1:N_T}^T)$ and is normalized using the target region's x-coordinate range $\max(x_{1:N_T}^T) - \min(x_{1:N_T}^T)$ . It is then scaled to the handle region's x-coordinate range $\max(x_{1:N_H}^H) - \min(x_{1:N_H}^H)$ , resulting in $x'$ . + +This step ensures that both regions have the same number of columns of pixels. Next, we map each target region's vertical pixel column to its corresponding handle region's column. Each point's $y$ in the target region is adjusted by subtracting the lower boundary of its vertical column $\min(y_i^T \mid x_i^T = x)$ and is normalized by its column's y-coordinate range $\max(y_i^T \mid x_i^T = x) - \min(y_i^T \mid x_i^T = x)$ , followed by being scaled to the corresponding handle region's column's y-coordinate range $\max(y_i^H \mid x_i^H = x') - \min(y_i^H \mid x_i^H = x')$ . The notation $f(\cdot \mid \cdot)$ represents a function $f$ (e.g., min or max) that takes a variable and a condition, and returns the function's value for the variable among points satisfying the condition. The whole process is summarised in Algorithm 1. + +Main Pipeline. As illustrated in Fig. 3, RegionDrag utilizes the image editing pipeline mentioned in Sec. 3.1. Initially, the latent representation of the image $z_0$ is inverted to $z_{t'}$ , where $t'$ is a selected timestep prior to maximum timestep $T$ . Each intermediate step $z_0, z_1, \ldots, z_{t'}$ is cached during inversion for future use. We then duplicate $z_{t'}$ and denote the copy as $z_{t'}'$ . The handle regions of $z_{t'}'$ are blended with Gaussian noise $\varepsilon$ according to the blending function $r_{\alpha}(z, H)$ defined in Eq. (5): + +$$ +r _ {\alpha} (z, H) = (1 - H) \cdot z + H \cdot (\sqrt {1 - \alpha^ {2}} \cdot z + \alpha \cdot \varepsilon), \quad \varepsilon \sim \mathcal {N} (0, I), \qquad (5) +$$ + +where $H$ is a binary mask, with the handle region assigned a value of 1 and $\alpha$ is a blending coefficient that ranges from 0 to 1, which governs the strength of the blending effect. The noise $\varepsilon$ is drawn from a Gaussian distribution with mean 0 and identity covariance matrix, denoted by $\mathcal{N}(0,I)$ . If $\alpha$ is set to a lower value, less noise is added, which preserves more of the original image's features and details in the handle regions. After blending with Gaussian noise, $z_{t'}'$ goes through the denoising process using the DDPM sampler, and the dragging is conducted in a copy-paste manner. At a denoising timestep $t$ , we extract latent representation within the handle region of $z_t$ and map it to the target region of $z_{t}'$ according to the dense mapping computed by either geometric transformation or Algorithm 1. In the meantime, we employ mutual self-attention control [2] to help maintain the identity of images. In brief, when passing through the self-attention module of the UNet, the key and value $(k_t', v_t')$ used when denoising $z_t'$ are replaced with those $(k_t, v_t)$ from $z_t$ . This allows the edited image to keep the layout and identity of the original image, thereby stabilizing the editing process. After $z_{t}'$ is gradually denoised to $z_0'$ (see Fig. 3), it is decoded to the edited image $x_0'$ . Although RegionDrag is introduced with a single handle-target region pair, it supports multiple pairs of input $\{(H_i, T_i)\}_{i=1}^n$ , allowing users to specify several modifications in a single editing session. Dense mapping is individually constructed for each region pair but collectively used in latent copy-paste operations. The entire pipeline is summarised in Algorithm 2. + +Algorithm 2 Editing Pipeline +Input: Image $x_0$ , handle and target region pairs $\{(H_i, T_i)\}_{i=1}^n$ . +Output: Edited image $x_0'$ . +Require: Diffusion function $f_{s \to t}$ Eq. (1), VAE encoder $E$ , VAE decoder $D$ , region-to-point mapping function $m$ , copy-paste function $CP$ Eq. (3), handle resampling function $r_\alpha$ Eq. (5), copy-paste time interval $(t', t'')$ . +1: Preparation: +2: $\{(h_i, t_i)\}_{i=1}^N \gets m(\{(H_i, T_i)\}_{i=1}^n)$ +3: $z_0 \gets E(x_0)$ +4: Inversion stage: +5: for $t = 0$ to $t' - 1$ do +6: $z_{t+1}, k_{t+1}, v_{t+1} \gets f_{t \to t+1}(z_t)$ +7: end for +8: $z_{t'}' \gets r_\alpha(z_{t'}, \cup_{i=1}^n H_i)$ +9: Denoising stage: +10: for $t = t'$ down to 1 do +11: if $t \geq t''$ then +12: $z_t' \gets \mathrm{CP}(z_t, h_{1:N}, z_t', t_{1:N})$ +13: end if +14: $z_{t-1}' \gets f_{t \to t-1}(z_t', k_t' \gets k_t, v_t' \gets v_t)$ +15: end for +16: $x_0' \gets D(z_0')$ +17: return $x_0'$ + +# 4 Experimental Results + +# 4.1 Datasets + +SDE-Drag [18] and DragDiffusion [24] have each introduced a dataset named DragBench to evaluate the performance of point-drag-based methods. For clarity, we refer to the one by SDE-Drag as DragBench-S, which contains 100 samples, and to the one by DragDiffusion as DragBench-D, which includes 205 samples. Each sample comprises an image, a descriptive prompt, an optional mask delineating the editing region, and at least one pair of points that reflects the user's intention. Empirically, DragBench-S offers relatively more detailed prompts and includes a mask in 56 of its samples, while DragBench-D is characterized by more complex intentions and features carefully crafted masks in each of its samples. + +To evaluate region-based editing, we introduce two new benchmarks: DragBench-SR and DragBench-DR (R is short for 'Region'), which are modified versions of DragBench-S and DragBench-D, respectively. These benchmarks are consistent with their point-drag-based counterparts in terms of images, prompts, and masks but differ by reflecting the user's intention through regions instead of points. DragBench-SR and DragBench-DR consist of 8 and 23 polygon-annotated samples, respectively, with the remainders of each dataset using brushes. For a handle-target region pair, the median number of equivalent point pairs is 267 in DragBench-SR and 201 in DragBench-DR. We visualize the distribution of the number of transformed point pairs in Fig. 4. + +During the annotation process, we ensure that the region pairs are aligned with the point pairs to achieve the same editing effect. To simulate how real users would typically draw handle-target region pairs, we do not meticulously draw each region. According to our observations, each region-based sample is annotated in approximately 10 seconds. This approach allows for a more realistic evaluation of the region-based editing methods while maintaining consistency with the point-based datasets. + +# 4.2 Evaluation Metrics + +LPIPS: We follow [24] and use Learned Perceptual Image Patch Similarity (LPIPS) v0.1 [30] to measure the identity similarity between the edited image and the original image. LPIPS computes the AlexNet [13] feature distances between the image pairs. A high LPIPS score indicates that unexpected identity changes or artifacts occur due to editing. A lower LPIPS score suggests that the object's identity has been well-preserved during editing; however, it does not necessarily imply a better drag edit, as two identical images would yield an LPIPS score of 0. + +Mean Distance (MD): DragDiffusion [24] introduces the MD metric to assess how well an approach moves the handle points' content to the target points. To identify where the handle points have been moved to by the method, DragDiffusion employs DIFT [26] to find the most similar points to the handle points + +![](images/2c225e96570194b0602d6fb448e941cbee5e817a4df55a5019505fc936766d1c.jpg) +Fig.4: Log-transformed frequency distribution of equivalent point pair counts in DragBench-SR and DragBench-DR. + +![](images/0c0d5ffacdd813f4c7936a20b1d678be40b7874d3126658456d0878e64320b3d.jpg) +Fig. 5: Illustration of MD computation. Dragged handle points will be searched within the search mask. + +$h_{1:n}$ in the entire edited image and denotes them as $h_{1:n}^{\prime}$ . It then uses the normalized Euclidean distance between DIFT-matched points and the true target positions $t_{1:n}$ as the metric. However, we believe that we only have to search the area around the handle points and their corresponding true target points for $h_{1:n}^{\prime}$ instead of the entire image, to avoid DIFT mistakenly identifying irrelevant points in the images as $h^\prime$ and excessively penalizing certain methods. Formally, given a dragging point pair $\{h,t\}$ , we define its searching mask as + +$$ +M (x) = \left\{ \begin{array}{l l} 1 & \text {i f} \min (d (x, h), d (x, t)) < \frac {d (h , t)}{\sqrt {2}} \\ 0 & \text {o t h e r w i s e} \end{array} \right., \tag {6} +$$ + +where $\mathrm{d}(\cdot, \cdot)$ calculates the normalized Euclidean distance between two points $(x_1, y_1)$ and $(x_2, y_2)$ by + +$$ +d \left(\left(x _ {1}, y _ {1}\right), \left(x _ {2}, y _ {2}\right)\right) = \sqrt {\left(\frac {x _ {2} - x _ {1}}{W}\right) ^ {2} + \left(\frac {y _ {2} - y _ {1}}{H}\right) ^ {2}}, \tag {7} +$$ + +where $W$ and $H$ are width and height of the image. We provide examples of searching masks in Fig. 5. MD is then defined as the average of the normalized distances $\bar{d}$ between the target points and the DIFT-matched points across all $n$ points: + +$$ +\bar {d} = \frac {1}{n} \sum_ {i = 1} ^ {n} d \left(t _ {i}, h _ {i} ^ {\prime}\right). \tag {8} +$$ + +# 4.3 Implementation Details + +Our method is implemented in Python using the HuggingFace [22] and PyTorch [20] libraries. We employ Stable Diffusion v1-5 as our diffusion model with an image size of $512 \times 512$ , being consistent with previous diffusion-based dragging methods. A DDPM sampler is utilized for both the inversion and denoising sampling processes, configured to use a total of 20 steps. The latent + +
MethodDragBench-S(R)DragBench-D(R)
Time (↓)MD (↓)LPIPS (↓)MD (↓)LPIPS (↓)
SDE-Drag [18]126.17.512.48.114.9
DragDiffusion [24]177.77.018.06.711.5
DiffEditor [17]43.123.617.622.110.9
Ours1.56.49.96.69.2
+ +Table 1: Comparisons of our method with baseline methods using MD(×100) and LPIPS(×100) metrics on DragBench-S(R) and DragBench-D(R) datasets. The time is measured in seconds and averaged across both datasets. The image size is 512×512. + +representation is inverted to timestep $t' = 500$ out of the total 1000 steps in SD1-5. Consequently, we perform 10 inversion steps and 10 denoising steps. Mutual self-attention control is enabled throughout all timesteps, and the latent copy-paste operation is terminated at $t'' = 200$ . The noise weight $\alpha$ is set to 1. All experimental results are obtained on an NVIDIA Tesla V100 GPU. + +# 4.4Baselines + +We compare RegionDrag with point-drag-based diffusion methods, including DragDiffusion [24], SDE-Drag [18], and DiffEditor [17]. GAN-based methods are excluded due to their limitations in editing the diverse images in the DragBench-S and DragBench-D datasets, as they require domain-specific StyleGAN checkpoints. Diffusion-based methods, which outperform GAN-based methods in editing tasks, are better suited for our evaluation. Execution times are averaged across both datasets, and all methods are tested on the same device using publicly released code. + +# 4.5 Quantitative Evaluation + +To quantitatively evaluate the editing performance of the methods, we employ LPIPS and Mean Distance as metrics, multiplying both by 100 for illustration purposes. As demonstrated in Table 1, RegionDrag significantly outperforms those computationally expensive point-based methods on both DragBench-S(R) and DragBench-D(R) datasets. These results highlight RegionDrag's superior performance in maintaining image consistency while achieving competitive editing results across different datasets. In addition to its effectiveness, RegionDrag also excels in terms of efficiency. RegionDrag achieves fast inference speed, requiring approximately 1.5 seconds to edit a $512 \times 512$ image, which is 20 times faster than the second-fastest method and 100 times faster than DragDiffusion [24]. The inference time for RegionDrag is comparable to generating an image with 20 steps using SD1-5, given that the copy-paste operations introduce negligible computational overhead. + +![](images/2a9629d4d04440744af6eccd89f8ad7634322d4fe4891251a15f095350ffaab4.jpg) +Fig. 6: Qualitative comparisons with baseline methods. Handle regions and target regions are respectively denoted by red and blue masks. + +# 4.6 Qualitative Results + +Figure 6 compares examples of point-drag-based and region-based editing inputs and their corresponding results, demonstrating the effectiveness of Region-Drag. Our region-based method utilizes the comprehensive context provided by annotated regions to target desired modifications while preserving the overall coherence of the image, outperforming point-drag-based editing methods. + +# 4.7 Ablation Study + +We argue that the sparsity of point inputs leads to inferior editing results. To quantitatively demonstrate this, we conducted tests on the DragBench-DR dataset by randomly selecting subsets of the equivalent transformed points within each sample and performing inference using these subsets. We gradually reduced the percentage of selected points to observe the impact on the MD metric. As illustrated in Fig. 7, the results exhibit a clear upward trend in MD as the percentage of utilized points decreases. This suggests that sparse point inputs provide weaker constraints on the output compared to region-based inputs, leading to unsatisfactory editing results. It confirms the benefits of employing region-based inputs in RegionDrag. + +![](images/0d3d76007f3f0d760792441ff9132a6ebc5ee77756ecffecacb995d38764d0a4.jpg) +Fig.7: Ablation study on the impact of the percentage of inputted transformed points. + +![](images/936a0cecf9b3b355d6aebe1f627bff86121ea6a3bea97e90d75f1314a602f113.jpg) +Fig. 8: Qualitative examples illustrating the impact of multi-step copy-paste. + +![](images/961aeb8ba451318f9b7a5996afb508550ce2094fddad7b33e4535d3a853f2d0c.jpg) + +We copy and paste image's latent representation over a time interval during denoising. To validate this design, we compare it to copy-paste only at the initial denoising timestep. Figure 8 shows that editing just at the initial step can yield unpredictable results because the edits may be lost in subsequent denoising phases. Multi-step copy-paste solves this by providing extra guidance at smaller timesteps while preserving image fidelity. + +# 5 Conclusion + +In this paper, we have introduced an efficient and effective region-based editing framework, RegionDrag, for high-fidelity image editing. Unlike existing approaches that utilize point-drag-based editing, RegionDrag reconsiders the editing problem from the region perspective. RegionDrag allows for editing in a single step through copying and pasting the latent representation and self-attention features of the image, which not only provides excellent efficiency but also achieves superior editing performance. Furthermore, we have introduced two new benchmarks, DragBench-SR and DragBench-DR based on existing datasets, for the evaluation of region-based editing. Experimental results have consistently demonstrated the superior efficiency and editing performance of our method. + +# Acknowledgments + +Kai Han and Jingyi Lu would like to acknowledge the support of the Hong Kong Research Grants Council - General Research Fund (Grant No.: 17211024). + +# References + +1. Brooks, T., Holynski, A., Efros, A.A.: Instructpix2pix: Learning to follow image editing instructions. In: CVPR (2023) +2. Cao, M., Wang, X., Qi, Z., Shan, Y., Qie, X., Zheng, Y.: Masactrl: Tuning-free mutual self-attention control for consistent image synthesis and editing. arXiv preprint arXiv:2304.08465 (2023) +3. Creswell, A., Bharath, A.A.: Inverting the generator of a generative adversarial network. IEEE TNNLS (2018) +4. Endo, Y.: User-controllable latent transformer for stylegan image layout editing. In: CGF (2022) +5. Epstein, D., Jabri, A., Poole, B., Efros, A., Holynski, A.: Diffusion self-guidance for controllable image generation. In: NeurIPS (2024) +6. Goodfellow, I., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., Courville, A., Bengio, Y.: Generative adversarial nets. In: NeurIPS (2014) +7. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. In: NeurIPS (2020) +8. Hu, E.J., Shen, Y., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., Chen, W.: Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685 (2021) +9. Kang, M., Zhu, J.Y., Zhang, R., Park, J., Shechtman, E., Paris, S., Park, T.: Scaling up gans for text-to-image synthesis. In: CVPR (2023) +0. Karras, T., Aittala, M., Laine, S., Härkönen, E., Hellsten, J., Lehtinen, J., Aila, T.: Alias-free generative adversarial networks. In: NeurIPS (2021) +1. Karras, T., Laine, S., Aittala, M., Hellsten, J., Lehtinen, J., Aila, T.: Analyzing and improving the image quality of stylegan. In: CVPR (2020) +2. Kawar, B., Zada, S., Lang, O., Tov, O., Chang, H., Dekel, T., Mosseri, I., Irani, M.: Imagic: Text-based real image editing with diffusion models. In: CVPR (2023) +3. Krizhevsky, A., Sutskever, I., Hinton, G.E.: Imagenet classification with deep convolutional neural networks. NeurIPS (2012) +4. Ling, P., Chen, L., Zhang, P., Chen, H., Jin, Y.: Freedrag: Point tracking is not you need for interactive point-based image editing. arXiv preprint arXiv:2307.04684 (2023) +5. Luo, M., Cheng, W., Yang, J.: Rotationdrag: Point-based image editing with rotated diffusion features. arXiv preprint arXiv:2401.06442 (2024) +6. Mou, C., Wang, X., Song, J., Shan, Y., Zhang, J.: Dragondiffusion: Enabling drag-style manipulation on diffusion models. arXiv preprint arXiv:2307.02421 (2023) +7. Mou, C., Wang, X., Song, J., Shan, Y., Zhang, J.: Diffeditor: Boosting accuracy and flexibility on diffusion-based image editing. arXiv preprint arXiv:2402.02583 (2024) +8. Nie, S., Guo, H.A., Lu, C., Zhou, Y., Zheng, C., Li, C.: The blessing of randomness: Sde beats ode in general diffusion-based image editing. arXiv preprint arXiv:2311.01410 (2023) +9. Pan, X., Tewari, A., Leimkuhler, T., Liu, L., Meka, A., Theobalt, C.: Drag your gan: Interactive point-based manipulation on the generative image manifold. In: ACM SIGGRAPH (2023) +20. Paszke, A., Gross, S., Massa, F., Lerer, A., Bradbury, J., Chanan, G., Killeen, T., Lin, Z., Gimelshein, N., Antiga, L., et al.: Pytorch: An imperative style, high-performance deep learning library. In: NeurIPS (2019) + +21. Patashnik, O., Wu, Z., Shechtman, E., Cohen-Or, D., Lischinski, D.: Styleclip: Text-driven manipulation of stylegan imagery. In: ICCV (2021) +22. von Platen, P., Patil, S., Lozhkov, A., Cuenca, P., Lambert, N., Rasul, K., Davaadorj, M., Wolf, T.: Diffusers: State-of-the-art diffusion models. https://github.com/huggingface/diffusers (2022) +23. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: CVPR (2022) +24. Shi, Y., Xue, C., Pan, J., Zhang, W., Tan, V.Y., Bai, S.: Dragdiffusion: Harnessing diffusion models for interactive point-based image editing. arXiv preprint arXiv:2306.14435 (2023) +25. Song, J., Meng, C., Ermon, S.: Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502 (2020) +26. Tang, L., Jia, M., Wang, Q., Phoo, C.P., Hariharan, B.: Emergent correspondence from image diffusion. In: NeurIPS (2024) +27. Wang, T., Zhang, Y., Fan, Y., Wang, J., Chen, Q.: High-fidelity gan inversion for image attribute editing. In: CVPR (2022) +28. Wu, C.H., De la Torre, F.: Unifying diffusion models' latent space, with applications to cyclediffusion and guidance. arXiv preprint arXiv:2210.05559 (2022) +29. Xia, W., Zhang, Y., Yang, Y., Xue, J.H., Zhou, B., Yang, M.H.: Gan inversion: A survey. IEEE TPAMI (2022) +30. Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual metric. In: CVPR (2018) \ No newline at end of file diff --git a/2024/RegionDrag_ Fast Region-Based Image Editing with Diffusion Models/images.zip b/2024/RegionDrag_ Fast Region-Based Image Editing with Diffusion Models/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..450ef24f117e2f3428bff74438c6c8d5a4fccaf9 --- /dev/null +++ b/2024/RegionDrag_ Fast Region-Based Image Editing with Diffusion Models/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8c15d659cc72d7904c6b814eb2409e3194f5110cfb453f6dfac949994267bd55 +size 452958 diff --git a/2024/RegionDrag_ Fast Region-Based Image Editing with Diffusion Models/layout.json b/2024/RegionDrag_ Fast Region-Based Image Editing with Diffusion Models/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..26c8ecd5d972a70befd957635daadb32f3c4c633 --- /dev/null +++ b/2024/RegionDrag_ Fast Region-Based Image Editing with Diffusion Models/layout.json @@ -0,0 +1,8548 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 144, + 111, + 471, + 146 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 144, + 111, + 471, + 146 + ], + "spans": [ + { + "bbox": [ + 144, + 111, + 471, + 146 + ], + "type": "text", + "content": "RegionDrag: Fast Region-Based Image Editing with Diffusion Models" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 208, + 168, + 405, + 181 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 208, + 168, + 405, + 181 + ], + "spans": [ + { + "bbox": [ + 208, + 168, + 405, + 181 + ], + "type": "text", + "content": "Jingyi Lu1, Xinghui Li2, and Kai Han1†" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 194, + 190, + 419, + 201 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 194, + 190, + 419, + 201 + ], + "spans": [ + { + "bbox": [ + 194, + 190, + 419, + 201 + ], + "type": "text", + "content": "1The University of Hong Kong 2University of Oxford" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 155, + 202, + 459, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 155, + 202, + 459, + 213 + ], + "spans": [ + { + "bbox": [ + 155, + 202, + 459, + 213 + ], + "type": "text", + "content": "lujingyi@connect.hku.hk, xinghui@robots.ox.ac.uk, kaihanx@hku.hk" + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 149, + 217, + 463, + 352 + ], + "blocks": [ + { + "bbox": [ + 149, + 217, + 463, + 352 + ], + "lines": [ + { + "bbox": [ + 149, + 217, + 463, + 352 + ], + "spans": [ + { + "bbox": [ + 149, + 217, + 463, + 352 + ], + "type": "image", + "image_path": "41b6b77d8b627d6b993d834af71986038a62eae5f00fc1e7ffbd63a6a2c74a6b.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 359, + 482, + 403 + ], + "lines": [ + { + "bbox": [ + 130, + 359, + 482, + 403 + ], + "spans": [ + { + "bbox": [ + 130, + 359, + 482, + 403 + ], + "type": "text", + "content": "Fig. 1: Comparison of editing results and latency between point-drag-based methods and our region-drag-based method. Our gradient-free, region-based framework reduces editing time from approximately one minute to about 1.5 seconds for " + }, + { + "bbox": [ + 130, + 359, + 482, + 403 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 130, + 359, + 482, + 403 + ], + "type": "text", + "content": " resolution images, while producing results that better align with users' intentions." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 159, + 417, + 453, + 626 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 417, + 453, + 626 + ], + "spans": [ + { + "bbox": [ + 159, + 417, + 453, + 626 + ], + "type": "text", + "content": "Abstract. Point-drag-based image editing methods, like DragDiffusion, have attracted significant attention. However, point-drag-based approaches suffer from computational overhead and misinterpretation of user intentions, due to the sparsity of point-based editing instructions. In this paper, we propose a region-based copy-and-paste dragging method, RegionDrag, to overcome these limitations. RegionDrag allows users to express their editing instructions in the form of handle and target regions, enabling more precise control and alleviating ambiguity. In addition, region-based operations complete editing in one iteration and are much faster than point-drag-based methods. We also incorporate the attention-swapping technique for enhanced stability during editing. To validate our approach, we extend existing point-drag-based datasets with region-based dragging instructions. Experimental results demonstrate that RegionDrag outperforms existing point-drag-based approaches in terms of speed, accuracy, and alignment with user intentions. Remarkably, RegionDrag completes the edit on an image with a resolution of " + }, + { + "bbox": [ + 159, + 417, + 453, + 626 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 159, + 417, + 453, + 626 + ], + "type": "text", + "content": " in less than 2 seconds, which is more than " + }, + { + "bbox": [ + 159, + 417, + 453, + 626 + ], + "type": "inline_equation", + "content": "100 \\times" + }, + { + "bbox": [ + 159, + 417, + 453, + 626 + ], + "type": "text", + "content": " faster than DragDiffusion, while achieving better performance. Project page: https://visual-ai.github.io/regiondrag." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 160, + 636, + 453, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 636, + 453, + 647 + ], + "spans": [ + { + "bbox": [ + 160, + 636, + 453, + 647 + ], + "type": "text", + "content": "Keywords: Region-Based Dragging " + }, + { + "bbox": [ + 160, + 636, + 453, + 647 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 636, + 453, + 647 + ], + "type": "text", + "content": " Image Editing " + }, + { + "bbox": [ + 160, + 636, + 453, + 647 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 636, + 453, + 647 + ], + "type": "text", + "content": " Diffusion Models" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 135, + 653, + 235, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 653, + 235, + 665 + ], + "spans": [ + { + "bbox": [ + 135, + 653, + 235, + 665 + ], + "type": "text", + "content": "† Corresponding author." + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 229, + 127 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 229, + 127 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 229, + 127 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 138, + 481, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 138, + 481, + 281 + ], + "spans": [ + { + "bbox": [ + 130, + 138, + 481, + 281 + ], + "type": "text", + "content": "Stable Diffusion (SD) [23] is a widely adopted text-to-image generative model known for its efficacy in producing high-fidelity images. As it is trained with millions of images, the vast amount of knowledge learned about images enables the model to edit existing images in a zero-shot manner. A recent line of work that has gained attention in the community is point-drag-based image editing using SD [15, 18, 24]. Point-drag-based methods first allow users to designate several points and drag them to desired positions. They then invert the latent representation of the input image to a particular timestep in the diffusion process and edit the image by enforcing similarities between local latent representations at the initial and final positions of the points. This is done either through optimization [15-17, 24] or direct copy-and-paste [18]. The edited latent representation is finally denoised and decoded to the edited image." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 283, + 481, + 390 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 283, + 481, + 390 + ], + "spans": [ + { + "bbox": [ + 130, + 283, + 481, + 390 + ], + "type": "text", + "content": "Although point-drag-based methods demonstrate encouraging results, they exhibit several limitations. First, as the editing solely relies on dragging sparse points, they have to interpolate dozens of intermediate points along the dragging directions and edit the image iteratively to avoid editing failures. This significantly slows down the speed of editing. Second, dragging points cannot always faithfully reflect the desired effect. This form of instruction is prone to being misinterpreted by the model, so the editing results may not fully align with users' actual intentions. For example, moving an object to the left and expanding the object to the left can equally be represented by dragging points to the left." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 390, + 481, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 390, + 481, + 628 + ], + "spans": [ + { + "bbox": [ + 130, + 390, + 481, + 628 + ], + "type": "text", + "content": "Therefore, we propose an alternative form of dragging to address the above issues. Instead of relying on points, we propose a region-based copy-and-paste dragging method, RegionDrag. Specifically, we first invert the latent representation of the unedited image to a specific time step using DDPM inversion [28], an inverse process to the stochastic DDPM sampling [25] which benefits the image editing [18]. Users then draw a handle region and a target region (see Fig. 1), where the former is the part that users intend to drag and the latter illustrates desired positions that users intend to achieve. After that, we establish a dense mapping between the two regions using our proposed Region-to-Point Mapping algorithm, and the latent representation covered by the handle region is mapped to the target region according to the dense mapping. Finally, the edited latent representation goes through the DDPM sampling process, and the mapping operation is repeated at multiple time steps. In the meantime, we also employ the attention-swapping technique proposed in [2] to stabilize the editing. To evaluate RegionDrag, we expand two existing point-drag-based image editing datasets [18,24] by adding equivalent region-based dragging instructions to each image. Experiments demonstrate that RegionDrag is significantly faster than point-drag-based methods as it completes editing in a single iteration. In addition, the paste region offers clearer constraints than point inputs and reduces chances of misinterpretation, leading to more faithful editing results." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 630, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 630, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 630, + 481, + 665 + ], + "type": "text", + "content": "Our contributions can be summarized as follows: (1) We introduce a region-based image editing method to overcome the limitations of point-drag-based approaches, utilizing richer input context to better align the editing results with" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 202, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 202, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 202, + 100 + ], + "type": "text", + "content": "Lu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 482, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 482, + 175 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 482, + 175 + ], + "type": "text", + "content": "the users' intentions. (2) By employing a gradient-free copy-paste operation, our region-based image editing becomes significantly faster than existing methods (see Fig. 1), completing the dragging in one single iteration. (3) We extend two point-drag-based datasets with region-based dragging instructions to validate RegionDrag's effectiveness and benchmark region-based editing methods." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 196, + 237, + 209 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 196, + 237, + 209 + ], + "spans": [ + { + "bbox": [ + 132, + 196, + 237, + 209 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 222, + 481, + 377 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 222, + 481, + 377 + ], + "spans": [ + { + "bbox": [ + 132, + 222, + 481, + 377 + ], + "type": "text", + "content": "Generative Models for Image Editing. Early advancements in image generation have been driven by generative adversarial networks (GANs) [6, 9-11]. However, their practical application in real-world image editing is limited by the diversity of GAN training data and the effectiveness of GAN inversion techniques [3, 4, 21, 27, 29]. The emergence of text-to-image diffusion models leads to novel image editing techniques that utilize text prompts to modify high-level characteristics such as style, motion, or object categories [1, 2, 5, 12]. Nevertheless, image editing methods based on user text prompts inherently struggle to manipulate images at the pixel level. Dragging methods [4, 14-17, 19, 24] aim to address this limitation by controlling the overall posture and shape of objects through iterative movement and tracking of one or multiple key points. RegionDrag introduces a region-based approach as a superior alternative, offering increased stability and efficiency for fine-grained image editing tasks." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 379, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 379, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 379, + 481, + 666 + ], + "type": "text", + "content": "Image Editing by Dragging Points. When employing drag-based methods for image editing, users can manipulate images by designating pairs of handle points and target points. The point-drag-based methods are expected to produce an image that meets two criteria: (1) features at the handle points are relocated to the target points, and (2) the original identity of the edited object is maintained. DragGAN [19] first enables editing on GAN-generated images involving multiple dragging point pairs. Specifically, DragGAN decomposes the editing process into several iterations, alternating between motion supervision and point tracking. In the motion supervision phase, DragGAN optimizes the StyleGAN's latent code [10,11] using the distance loss between the initial and target positions of the handle points. Following each motion supervision iteration, DragGAN updates the positions of the handle points by employing a point-tracking technique. FreeDrag [14] switches the dragging from pixel space to feature space, enabling more stable and precise manipulation. DragDiffusion [24] extends the supervise-and-track framework of DragGAN [19] to diffusion models. It inverts the latent representation of the unedited image to a partially noisy status at a selected time step, drags point features by optimizing the latent representation, and denoises it. RotationDrag [15] further refines the DragDiffusion framework [24] for cases of rotating objects. Another line of research, including DragonDiffusion [16] and DiffEditor [17], draws inspiration from classifier-guided generation, extending feature dragging throughout the entire denoising process rather than confining it to a single timestep. A method similar to ours, named SDE-Drag [18], eliminates memory-intensive backpropagation found in [16, 17, 24] by copying and pasting points within the diffusion latent space. Although SDE-Drag demon-" + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 165, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 165, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 165, + 91, + 447, + 102 + ], + "type": "text", + "content": "RegionDrag: Fast Region-Based Image Editing with Diffusion Models" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 176 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 176 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 176 + ], + "type": "text", + "content": "strates promising results, it remains time-consuming due to its reliance on interpolating intermediate points along the dragging path to stabilize the editing process. In this paper, we introduce a simple yet effective region-based framework that utilizes the rich context inherent in region pairs to complete all edits in just one editing step." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 176, + 482, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 176, + 482, + 380 + ], + "spans": [ + { + "bbox": [ + 130, + 176, + 482, + 380 + ], + "type": "text", + "content": "Appearance Consistency for Drag Editing. SD-based image editing shifts the image's latent distribution to incorporate user-specified structural, layout, and shape changes. However, this process can also introduce undesirable changes and artifacts. Mitigating these side effects remains challenging. Existing methods [15,18,24] often require training the SD model with LoRA (Low-Rank Adaptation) [8] for each unedited image to maintain style and appearance consistency. While LoRA is parameter-efficient through the use of low-rank matrix decomposition, it requires extensive preparation time and may overly constrain the image's appearance, limiting further edits. Additionally, point-drag-based methods often necessitate masking the editing area to reduce input ambiguity and enhance image consistency. In contrast, in this paper, we incorporate a training-free approach called mutual self-attention control [2], which preserves the image's identity by leveraging keys and values in the self-attention blocks of the model. This approach is seamlessly integrated into the editing pipeline, eliminating the need for an additional module to maintain the image's identity. Furthermore, unlike point-drag-based methods that rely on a mask to confine the editing area, RegionDrag eliminates this step to provide a more convenient editing experience." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 398, + 227, + 412 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 398, + 227, + 412 + ], + "spans": [ + { + "bbox": [ + 132, + 398, + 227, + 412 + ], + "type": "text", + "content": "3 RegionDrag" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 423, + 479, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 423, + 479, + 483 + ], + "spans": [ + { + "bbox": [ + 130, + 423, + 479, + 483 + ], + "type": "text", + "content": "RegionDrag enables users to input handle and target region pairs, which are then used for editing through two primary steps: (1) copying latent representations covered by handle regions and storing self-attention features during inversion, and (2) pasting copied latent representations to target positions and inserting stored self-attention features during denoising." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 483, + 482, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 483, + 482, + 544 + ], + "spans": [ + { + "bbox": [ + 130, + 483, + 482, + 544 + ], + "type": "text", + "content": "This section begins by reviewing the diffusion-based image editing pipeline and point-drag-based methods in Sec. 3.1. We then discuss the limitations of point-drag-based methods and introduce how our region-based input addresses these limitations in Sec. 3.2. Finally, Sec. 3.3 presents our editing pipeline used to process region-based inputs." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 561, + 222, + 574 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 561, + 222, + 574 + ], + "spans": [ + { + "bbox": [ + 132, + 561, + 222, + 574 + ], + "type": "text", + "content": "3.1 Preliminary" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "content": "Diffusion-based image editing involves two main stages: inversion and denoising. Unedited images are first gradually inverted to a specific timestep in the diffusion process. The editing is then performed on the images at this timestep and the images are finally denoised back to their original image space. The transition between timesteps in the diffusion process is governed by a sampling scheduler. A commonly used one is the DDIM scheduler [25]. When an image latent " + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "inline_equation", + "content": "z_{s}" + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "content": " at timestep " + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "content": " transitions to " + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "inline_equation", + "content": "z_{t}" + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "content": " at timestep " + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "content": ", it goes through:" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 202, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 202, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 202, + 100 + ], + "type": "text", + "content": "Lu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 142, + 136, + 480, + 178 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 142, + 136, + 480, + 178 + ], + "spans": [ + { + "bbox": [ + 142, + 136, + 480, + 178 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} z _ {t} = f _ {s \\rightarrow t} (z _ {s}) \\\\ = \\sqrt {\\alpha_ {t}} \\left(\\frac {z _ {s} - \\sqrt {1 - \\alpha_ {s}} \\epsilon_ {\\theta} (z _ {s} , s , C)}{\\sqrt {\\alpha_ {s}}}\\right) + \\sqrt {1 - \\alpha_ {t} - \\sigma_ {s} ^ {2}} \\epsilon_ {\\theta} (z _ {s}, s, C) + \\sigma_ {s} w _ {s}, \\quad (1) \\\\ \\end{array}", + "image_path": "cbddb8cc25414c9d467a99a46f3555d37bf8cbd9cc764eb7cc519d92e26e0cd7.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 186, + 479, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 186, + 479, + 281 + ], + "spans": [ + { + "bbox": [ + 130, + 186, + 479, + 281 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 186, + 479, + 281 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}(z_s,s,C)" + }, + { + "bbox": [ + 130, + 186, + 479, + 281 + ], + "type": "text", + "content": " is the noise predicted by the diffusion model, " + }, + { + "bbox": [ + 130, + 186, + 479, + 281 + ], + "type": "inline_equation", + "content": "\\alpha_{t}" + }, + { + "bbox": [ + 130, + 186, + 479, + 281 + ], + "type": "text", + "content": " is a monotonically decreasing noise scheduling function dependent on " + }, + { + "bbox": [ + 130, + 186, + 479, + 281 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 186, + 479, + 281 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 186, + 479, + 281 + ], + "type": "inline_equation", + "content": "w_{s}" + }, + { + "bbox": [ + 130, + 186, + 479, + 281 + ], + "type": "text", + "content": " is Gaussian noise, and " + }, + { + "bbox": [ + 130, + 186, + 479, + 281 + ], + "type": "inline_equation", + "content": "\\sigma_{s} = \\eta \\sqrt{(1 - \\alpha_{t}) / (1 - \\alpha_{s})}\\sqrt{1 - \\alpha_{s} / \\alpha_{t}}" + }, + { + "bbox": [ + 130, + 186, + 479, + 281 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 130, + 186, + 479, + 281 + ], + "type": "inline_equation", + "content": "\\eta = 0" + }, + { + "bbox": [ + 130, + 186, + 479, + 281 + ], + "type": "text", + "content": ". Since " + }, + { + "bbox": [ + 130, + 186, + 479, + 281 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 130, + 186, + 479, + 281 + ], + "type": "text", + "content": " is set to 0, the transition is a deterministic process and it is widely used in many methods [16,17,23,24]. If " + }, + { + "bbox": [ + 130, + 186, + 479, + 281 + ], + "type": "inline_equation", + "content": "\\eta > 0" + }, + { + "bbox": [ + 130, + 186, + 479, + 281 + ], + "type": "text", + "content": ", the sampling becomes a stochastic process known as the DDPM sampling [7]. We choose DDPM over DDIM as [18] shows that DDPM's randomness reduces divergence between unedited and edited image distributions." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 281, + 480, + 390 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 281, + 480, + 390 + ], + "spans": [ + { + "bbox": [ + 130, + 281, + 480, + 390 + ], + "type": "text", + "content": "Point-drag-based methods, such as DragDiffusion and SDE-Drag [18,24], edit an image by relocating its latent representations at user-designated " + }, + { + "bbox": [ + 130, + 281, + 480, + 390 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 130, + 281, + 480, + 390 + ], + "type": "text", + "content": " handle points " + }, + { + "bbox": [ + 130, + 281, + 480, + 390 + ], + "type": "inline_equation", + "content": "h_{1:n}" + }, + { + "bbox": [ + 130, + 281, + 480, + 390 + ], + "type": "text", + "content": " to corresponding target positions " + }, + { + "bbox": [ + 130, + 281, + 480, + 390 + ], + "type": "inline_equation", + "content": "t_{1:n}" + }, + { + "bbox": [ + 130, + 281, + 480, + 390 + ], + "type": "text", + "content": ". Due to the limited context available in sparse points, point-drag-based methods divide drag editing into " + }, + { + "bbox": [ + 130, + 281, + 480, + 390 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 281, + 480, + 390 + ], + "type": "text", + "content": " sub-steps. At sub-step " + }, + { + "bbox": [ + 130, + 281, + 480, + 390 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 281, + 480, + 390 + ], + "type": "text", + "content": ", the algorithm drags latent representations from handle points " + }, + { + "bbox": [ + 130, + 281, + 480, + 390 + ], + "type": "inline_equation", + "content": "h_{k,1:n}" + }, + { + "bbox": [ + 130, + 281, + 480, + 390 + ], + "type": "text", + "content": " to target positions " + }, + { + "bbox": [ + 130, + 281, + 480, + 390 + ], + "type": "inline_equation", + "content": "t_{k,1:n}" + }, + { + "bbox": [ + 130, + 281, + 480, + 390 + ], + "type": "text", + "content": ", which are found by direct interpolation or an extra point matching step. DragDiffusion optimizes " + }, + { + "bbox": [ + 130, + 281, + 480, + 390 + ], + "type": "inline_equation", + "content": "z_t^k" + }, + { + "bbox": [ + 130, + 281, + 480, + 390 + ], + "type": "text", + "content": " by minimizing the " + }, + { + "bbox": [ + 130, + 281, + 480, + 390 + ], + "type": "inline_equation", + "content": "\\ell_1" + }, + { + "bbox": [ + 130, + 281, + 480, + 390 + ], + "type": "text", + "content": "-distance between UNet upblock features " + }, + { + "bbox": [ + 130, + 281, + 480, + 390 + ], + "type": "inline_equation", + "content": "F(z_t^k)" + }, + { + "bbox": [ + 130, + 281, + 480, + 390 + ], + "type": "text", + "content": " at handle " + }, + { + "bbox": [ + 130, + 281, + 480, + 390 + ], + "type": "inline_equation", + "content": "h_{k,1:n}" + }, + { + "bbox": [ + 130, + 281, + 480, + 390 + ], + "type": "text", + "content": " and target " + }, + { + "bbox": [ + 130, + 281, + 480, + 390 + ], + "type": "inline_equation", + "content": "t_{k,1:n}" + }, + { + "bbox": [ + 130, + 281, + 480, + 390 + ], + "type": "text", + "content": " points, respectively denoted as " + }, + { + "bbox": [ + 130, + 281, + 480, + 390 + ], + "type": "inline_equation", + "content": "F(z_t^k)[h_{k,1:n}]" + }, + { + "bbox": [ + 130, + 281, + 480, + 390 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 281, + 480, + 390 + ], + "type": "inline_equation", + "content": "F(z_t^k)[t_{k,1:n}]" + }, + { + "bbox": [ + 130, + 281, + 480, + 390 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 195, + 399, + 480, + 415 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 195, + 399, + 480, + 415 + ], + "spans": [ + { + "bbox": [ + 195, + 399, + 480, + 415 + ], + "type": "interline_equation", + "content": "z _ {t} ^ {k + 1} = z _ {t} ^ {k} - \\eta \\cdot \\nabla_ {z _ {t} ^ {k}} \\| F \\left(z _ {t} ^ {k}\\right) [ h _ {k, 1: n} ] - F \\left(z _ {t} ^ {k}\\right) [ t _ {k, 1: n} ] \\| _ {1}. \\tag {2}", + "image_path": "21cb623fc815921bbd56d1eea62317c13383e7b5989e64d0538f4b2f33c1da8a.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 418, + 479, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 418, + 479, + 491 + ], + "spans": [ + { + "bbox": [ + 130, + 418, + 479, + 491 + ], + "type": "text", + "content": "DragonDiffusion [16] and DiffEditor [17] are DragDiffusion variants with " + }, + { + "bbox": [ + 130, + 418, + 479, + 491 + ], + "type": "inline_equation", + "content": "K = 1" + }, + { + "bbox": [ + 130, + 418, + 479, + 491 + ], + "type": "text", + "content": " and extend optimization across multiple diffusion timesteps. On the other hand, SDE-Drag copies the latent representations from handle points to target points using the copy-paste function CP (Eq. (3)), where " + }, + { + "bbox": [ + 130, + 418, + 479, + 491 + ], + "type": "inline_equation", + "content": "z[h]" + }, + { + "bbox": [ + 130, + 418, + 479, + 491 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 418, + 479, + 491 + ], + "type": "inline_equation", + "content": "z[t]" + }, + { + "bbox": [ + 130, + 418, + 479, + 491 + ], + "type": "text", + "content": " represent the latent code at positions " + }, + { + "bbox": [ + 130, + 418, + 479, + 491 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 130, + 418, + 479, + 491 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 418, + 479, + 491 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 418, + 479, + 491 + ], + "type": "text", + "content": ". This is then followed by a denoising and inversion cycle (Eq. (4))." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 236, + 502, + 480, + 516 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 236, + 502, + 480, + 516 + ], + "spans": [ + { + "bbox": [ + 236, + 502, + 480, + 516 + ], + "type": "interline_equation", + "content": "\\mathrm {C P} \\left(z _ {1}, z _ {2}, h, t\\right) = \\left(z _ {2} [ t ] \\leftarrow z _ {1} [ h ]\\right), \\tag {3}", + "image_path": "592105cec8a020e6894ddbbfb5f9b93803894bb1e0a2bbb6de979bcb2376c193.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 206, + 531, + 480, + 545 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 206, + 531, + 480, + 545 + ], + "spans": [ + { + "bbox": [ + 206, + 531, + 480, + 545 + ], + "type": "interline_equation", + "content": "z _ {t} ^ {k + 1} = f _ {0 \\rightarrow t} \\left(f _ {t \\rightarrow 0} \\left(\\mathrm {C P} \\left(z _ {t} ^ {k}, z _ {t} ^ {k}, h _ {k, 1: n}, t _ {k, 1: n}\\right)\\right)\\right). \\tag {4}", + "image_path": "38df46b9cbcdd33b8daee3aa0ccb5d928dcb6a73392baca25e06749772c623ee.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 562, + 388, + 574 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 562, + 388, + 574 + ], + "spans": [ + { + "bbox": [ + 131, + 562, + 388, + 574 + ], + "type": "text", + "content": "3.2 From Point-Based to Region-Based Dragging" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 582, + 479, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 479, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 479, + 665 + ], + "type": "text", + "content": "Although point-drag-based methods offer an intuitive means of user input, the limited information derived from sparse points poses challenges for models during the editing process. Specifically, point instructions can result in two primary issues: input ambiguity and slow inference. First, point instructions are inherently ambiguous. One dragging action could correspond to multiple plausible editing effects. Consider a user attempting to elongate a bird's beak in an image, as depicted in Fig. 2. The user selects a point on the beak and drags it" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 164, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 164, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 164, + 91, + 448, + 102 + ], + "type": "text", + "content": "RegionDrag: Fast Region-Based Image Editing with Diffusion Models" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 150, + 114, + 231, + 215 + ], + "blocks": [ + { + "bbox": [ + 150, + 114, + 231, + 215 + ], + "lines": [ + { + "bbox": [ + 150, + 114, + 231, + 215 + ], + "spans": [ + { + "bbox": [ + 150, + 114, + 231, + 215 + ], + "type": "image", + "image_path": "94c9037661634a68167ffe25c0e7bed943628ce3bbdf9ed3d2f5161584b96483.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 224, + 482, + 257 + ], + "lines": [ + { + "bbox": [ + 130, + 224, + 482, + 257 + ], + "spans": [ + { + "bbox": [ + 130, + 224, + 482, + 257 + ], + "type": "text", + "content": "Fig. 2: Overall comparison of point-based editing and region-based editing, exemplified by manipulating a bird's beak. The region-based approach is shown to provide a more user-friendly and less ambiguous editing experience." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 238, + 114, + 282, + 212 + ], + "blocks": [ + { + "bbox": [ + 238, + 114, + 282, + 212 + ], + "lines": [ + { + "bbox": [ + 238, + 114, + 282, + 212 + ], + "spans": [ + { + "bbox": [ + 238, + 114, + 282, + 212 + ], + "type": "image", + "image_path": "c93ee77287bed885ab9db6fefd2d037c79f4e0a9f4d63032cb841232b2ba81db.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 283, + 115, + 365, + 212 + ], + "blocks": [ + { + "bbox": [ + 283, + 115, + 365, + 212 + ], + "lines": [ + { + "bbox": [ + 283, + 115, + 365, + 212 + ], + "spans": [ + { + "bbox": [ + 283, + 115, + 365, + 212 + ], + "type": "image", + "image_path": "fa8f65a9a38dc77d4547cf8f56d9bfe272c25561cda12d1047fbd19e4f4d3815.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 372, + 116, + 417, + 213 + ], + "blocks": [ + { + "bbox": [ + 372, + 116, + 417, + 213 + ], + "lines": [ + { + "bbox": [ + 372, + 116, + 417, + 213 + ], + "spans": [ + { + "bbox": [ + 372, + 116, + 417, + 213 + ], + "type": "image", + "image_path": "3045a75b20c1569f949fb5ecc3042adc562b3788ac87bd05d88bb3c71f5f7d81.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 419, + 141, + 463, + 190 + ], + "blocks": [ + { + "bbox": [ + 419, + 141, + 463, + 190 + ], + "lines": [ + { + "bbox": [ + 419, + 141, + 463, + 190 + ], + "spans": [ + { + "bbox": [ + 419, + 141, + 463, + 190 + ], + "type": "image", + "image_path": "f7bd6c89106aa01e0fc9439be1a3d792e0c0c72e993cf364695948334f45ab35.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 281, + 480, + 329 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 281, + 480, + 329 + ], + "spans": [ + { + "bbox": [ + 130, + 281, + 480, + 329 + ], + "type": "text", + "content": "towards the upper-left corner. Point-drag-based methods, however, might misinterpret the user's goal as enlarging the beak or moving the entire bird to the left, rather than extending the beak as intended, leading to a misalignment between the user's intention and the model's output." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 331, + 482, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 331, + 482, + 521 + ], + "spans": [ + { + "bbox": [ + 130, + 331, + 482, + 521 + ], + "type": "text", + "content": "Second, the complexity involved in point-drag-based editing requires considerable computational overhead. Point-drag-based editing is challenging because the model must deduce changes across the entire image from the motion of a single or a few points. To carry out this complex drag operation while preserving the object's identity, point-drag-based methods heavily rely on two computationally intensive steps: training a unique LoRA [8] for each image and breaking down the dragging process into a series of sub-steps. Particularly, LoRA helps the model maintain the original image's identity and step-by-step dragging boosts the chance of achieving desired editing effects. Otherwise, the editing results may suffer from significant identity distortion or void editing, as shown in Fig. 2. The root of this problem lies in that sparse points do not impose sufficient constraints on editing, so the model has to rely on LoRA to prevent distortion and iterative editing to provide some degree of additional supervision along the path of the dragging. Consequently, most point-drag-based methods require several minutes to edit one image, rendering them impractical for real-world applications." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 522, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 522, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 522, + 482, + 666 + ], + "type": "text", + "content": "The simplest solution to these problems is to encourage users to provide a sufficient number of points. However, such an approach would result in users spending too much time on designating and dragging points. Therefore, we design a form of editing that is not only user-friendly but also provides more informative context to the model, thus avoiding instruction ambiguity, slow inference, and exhaustive efforts from users. Instead of relying on dragging points, we propose to use region-based operations, where users assign a handle region " + }, + { + "bbox": [ + 130, + 522, + 482, + 666 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 130, + 522, + 482, + 666 + ], + "type": "text", + "content": " to indicate the area they wish to drag and a target region " + }, + { + "bbox": [ + 130, + 522, + 482, + 666 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 130, + 522, + 482, + 666 + ], + "type": "text", + "content": " to illustrate the desired position they would like to achieve. We then establish a dense mapping between the two regions using our Region-to-Point Mapping algorithm and complete the editing by directly copying the latent representation covered by the handle region to the target region in one inversion and denoising cycle. Despite" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 202, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 202, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 202, + 100 + ], + "type": "text", + "content": "Lu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 151, + 129, + 463, + 244 + ], + "blocks": [ + { + "bbox": [ + 198, + 114, + 408, + 124 + ], + "lines": [ + { + "bbox": [ + 198, + 114, + 408, + 124 + ], + "spans": [ + { + "bbox": [ + 198, + 114, + 408, + 124 + ], + "type": "text", + "content": "Handle Region Target Region " + }, + { + "bbox": [ + 198, + 114, + 408, + 124 + ], + "type": "inline_equation", + "content": "\\longrightarrow" + }, + { + "bbox": [ + 198, + 114, + 408, + 124 + ], + "type": "text", + "content": " Copy-Paste Inference" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 151, + 129, + 463, + 244 + ], + "lines": [ + { + "bbox": [ + 151, + 129, + 463, + 244 + ], + "spans": [ + { + "bbox": [ + 151, + 129, + 463, + 244 + ], + "type": "image", + "image_path": "c2a87b96473ac1ef4358ea0c1963b73c3286a2c1cbc3f01cf8469647ceeb1d57.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 252, + 480, + 308 + ], + "lines": [ + { + "bbox": [ + 130, + 252, + 480, + 308 + ], + "spans": [ + { + "bbox": [ + 130, + 252, + 480, + 308 + ], + "type": "text", + "content": "Fig. 3: General pipeline of our method. Rich context provided by the region pairs enables users to complete accurate edits in one inversion and denoising cycle. The latent representation of the handle region is copied and pasted onto the target region throughout multiple timesteps for drag editing. Keys and values within the self-attention blocks are reused to ensure image consistency." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 331, + 482, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 331, + 482, + 462 + ], + "spans": [ + { + "bbox": [ + 130, + 331, + 482, + 462 + ], + "type": "text", + "content": "the simplicity of this operation, it addresses ambiguity and overhead from two perspectives: (1) Region-based operation is more expressive and accurate than dragging points and would significantly alleviate the ambiguity. As demonstrated in Fig. 2, we express extending the bird's beak by simply drawing a longer beak, hence reducing ambiguity presented in point-drag-based inputs. (2) Each region corresponds to a large number of points after dense mapping, so it provides stronger constraints on editing results than sparse points. As a result, we do not have to interpolate intermediate points along the dragging path and crave extra supervision, and editing can be completed in one editing step. Moreover, the handle and target regions can vary in size and take arbitrary shapes allowing users to define them conveniently." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 479, + 243, + 491 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 479, + 243, + 491 + ], + "spans": [ + { + "bbox": [ + 131, + 479, + 243, + 491 + ], + "type": "text", + "content": "3.3 Editing Pipeline" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 498, + 480, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 498, + 480, + 533 + ], + "spans": [ + { + "bbox": [ + 130, + 498, + 480, + 533 + ], + "type": "text", + "content": "We elaborate on our editing pipeline in this section. We first introduce the our region-based user input, followed by our Region-to-Point Mapping algorithm, and finally the main working pipeline." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 533, + 481, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 481, + 605 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 481, + 605 + ], + "type": "text", + "content": "User Input. The handle and target regions can be defined in two ways: (1) by entering vertices to form a polygon (e.g., a triangle or quadrilateral), or (2) by brushing out a flexible region using a brush tool. The choice of input form largely depends on the user's preferences. Vertices are ideal for editing well-defined shapes, like moving a window on a building; a brush tool is better suited for irregular shapes, like a curved road or human hair." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 605, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 481, + 665 + ], + "type": "text", + "content": "Region-to-Point Mapping. To preserve the original spatial information when copying and pasting latent representation, we need to establish a dense mapping between the handle and target region. If the regions are confined to triangular or quadrilateral shapes, we can compute a transformation matrix using affine or perspective mappings. However, finding a similar transformation for brush-out" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 165, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 165, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 165, + 91, + 448, + 102 + ], + "type": "text", + "content": "RegionDrag: Fast Region-Based Image Editing with Diffusion Models" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "code", + "bbox": [ + 132, + 130, + 481, + 289 + ], + "blocks": [ + { + "bbox": [ + 133, + 116, + 312, + 128 + ], + "lines": [ + { + "bbox": [ + 133, + 116, + 312, + 128 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 312, + 128 + ], + "type": "text", + "content": "Algorithm 1 Region-to-Point Mapping" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "code_caption" + }, + { + "bbox": [ + 132, + 130, + 481, + 289 + ], + "lines": [ + { + "bbox": [ + 132, + 130, + 481, + 289 + ], + "spans": [ + { + "bbox": [ + 132, + 130, + 481, + 289 + ], + "type": "text", + "content": "Input: Handle region " + }, + { + "bbox": [ + 132, + 130, + 481, + 289 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 132, + 130, + 481, + 289 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 132, + 130, + 481, + 289 + ], + "type": "inline_equation", + "content": "N_{H}" + }, + { + "bbox": [ + 132, + 130, + 481, + 289 + ], + "type": "text", + "content": " handle points " + }, + { + "bbox": [ + 132, + 130, + 481, + 289 + ], + "type": "inline_equation", + "content": "\\{(x_i^H,y_i^H)\\}_{i = 1}^{N_H}" + }, + { + "bbox": [ + 132, + 130, + 481, + 289 + ], + "type": "text", + "content": ", target region " + }, + { + "bbox": [ + 132, + 130, + 481, + 289 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 132, + 130, + 481, + 289 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 132, + 130, + 481, + 289 + ], + "type": "inline_equation", + "content": "N_{T}" + }, + { + "bbox": [ + 132, + 130, + 481, + 289 + ], + "type": "text", + "content": " target points " + }, + { + "bbox": [ + 132, + 130, + 481, + 289 + ], + "type": "inline_equation", + "content": "\\{(x_i^T,y_i^T)\\}_{i = 1}^{N_T}" + }, + { + "bbox": [ + 132, + 130, + 481, + 289 + ], + "type": "text", + "content": ". \nOutput: List of mapped point pairs " + }, + { + "bbox": [ + 132, + 130, + 481, + 289 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 132, + 130, + 481, + 289 + ], + "type": "text", + "content": ". \nRequire: Initialize empty list " + }, + { + "bbox": [ + 132, + 130, + 481, + 289 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 132, + 130, + 481, + 289 + ], + "type": "text", + "content": ". \n1: for " + }, + { + "bbox": [ + 132, + 130, + 481, + 289 + ], + "type": "inline_equation", + "content": "(x,y)\\in \\{(x_i^T,y_i^T)\\}_{i = 1}^{N_T}" + }, + { + "bbox": [ + 132, + 130, + 481, + 289 + ], + "type": "text", + "content": " do \n2: Horizontal scaling: \n3: " + }, + { + "bbox": [ + 132, + 130, + 481, + 289 + ], + "type": "inline_equation", + "content": "x^{\\prime}\\gets (x - \\min (x_{1:N_T}^T)) / (\\max (x_{1:N_T}^T) - \\min (x_{1:N_T}^T))" + }, + { + "bbox": [ + 132, + 130, + 481, + 289 + ], + "type": "text", + "content": " \n4: " + }, + { + "bbox": [ + 132, + 130, + 481, + 289 + ], + "type": "inline_equation", + "content": "x^{\\prime}\\gets \\lfloor x^{\\prime}\\cdot (\\max (x_{1:N_H}^H) - \\min (x_{1:N_H}^H)) + \\min (x_{1:N_H}^H)\\rfloor" + }, + { + "bbox": [ + 132, + 130, + 481, + 289 + ], + "type": "text", + "content": " \n5: Column-by-column vertical scaling: \n6: " + }, + { + "bbox": [ + 132, + 130, + 481, + 289 + ], + "type": "inline_equation", + "content": "y^\\prime \\gets (y - \\min (y_i^T\\mid x_i^T = x)) / (\\max (y_i^T\\mid x_i^T = x) - \\min (y_i^T\\mid x_i^T = x))" + }, + { + "bbox": [ + 132, + 130, + 481, + 289 + ], + "type": "text", + "content": " \n7: " + }, + { + "bbox": [ + 132, + 130, + 481, + 289 + ], + "type": "inline_equation", + "content": "y^\\prime \\gets \\lfloor y^\\prime \\cdot (\\max (y_i^H\\mid x_i^H = x^\\prime) - \\min (y_i^H\\mid x_i^H = x^\\prime)) + \\min (y_i^H\\mid x_i^H = x^\\prime)\\rfloor" + }, + { + "bbox": [ + 132, + 130, + 481, + 289 + ], + "type": "text", + "content": " \n8: Add point pair " + }, + { + "bbox": [ + 132, + 130, + 481, + 289 + ], + "type": "inline_equation", + "content": "((x',y'),(x,y))" + }, + { + "bbox": [ + 132, + 130, + 481, + 289 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 132, + 130, + 481, + 289 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 132, + 130, + 481, + 289 + ], + "type": "text", + "content": " \n9: end for \n10: return " + }, + { + "bbox": [ + 132, + 130, + 481, + 289 + ], + "type": "inline_equation", + "content": "P" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "code_body" + } + ], + "index": 3, + "sub_type": "algorithm" + }, + { + "bbox": [ + 130, + 317, + 480, + 341 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 317, + 480, + 341 + ], + "spans": [ + { + "bbox": [ + 130, + 317, + 480, + 341 + ], + "type": "text", + "content": "regions with arbitrary shapes is challenging. To address this issue, we propose an algorithm to numerically find the mapping between two regions." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 342, + 480, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 342, + 480, + 430 + ], + "spans": [ + { + "bbox": [ + 130, + 342, + 480, + 430 + ], + "type": "text", + "content": "Let " + }, + { + "bbox": [ + 130, + 342, + 480, + 430 + ], + "type": "inline_equation", + "content": "\\{(x_i^H, y_i^H)\\}_{i=1}^{N_H}" + }, + { + "bbox": [ + 130, + 342, + 480, + 430 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 342, + 480, + 430 + ], + "type": "inline_equation", + "content": "\\{(x_j^T, y_j^T)\\}_{j=1}^{N_T}" + }, + { + "bbox": [ + 130, + 342, + 480, + 430 + ], + "type": "text", + "content": " be the sets of pixels in handle and target regions respectively. We begin by linearly scaling the target region's width to match the handle region's width. For each pixel " + }, + { + "bbox": [ + 130, + 342, + 480, + 430 + ], + "type": "inline_equation", + "content": "(x, y) \\in \\{(x_j^T, y_j^T)\\}_{j=1}^{N_T}" + }, + { + "bbox": [ + 130, + 342, + 480, + 430 + ], + "type": "text", + "content": " in the target region, " + }, + { + "bbox": [ + 130, + 342, + 480, + 430 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 130, + 342, + 480, + 430 + ], + "type": "text", + "content": " is adjusted by first subtracting the x-coordinate lower boundary " + }, + { + "bbox": [ + 130, + 342, + 480, + 430 + ], + "type": "inline_equation", + "content": "\\min(x_{1:N_T}^T)" + }, + { + "bbox": [ + 130, + 342, + 480, + 430 + ], + "type": "text", + "content": " and is normalized using the target region's x-coordinate range " + }, + { + "bbox": [ + 130, + 342, + 480, + 430 + ], + "type": "inline_equation", + "content": "\\max(x_{1:N_T}^T) - \\min(x_{1:N_T}^T)" + }, + { + "bbox": [ + 130, + 342, + 480, + 430 + ], + "type": "text", + "content": ". It is then scaled to the handle region's x-coordinate range " + }, + { + "bbox": [ + 130, + 342, + 480, + 430 + ], + "type": "inline_equation", + "content": "\\max(x_{1:N_H}^H) - \\min(x_{1:N_H}^H)" + }, + { + "bbox": [ + 130, + 342, + 480, + 430 + ], + "type": "text", + "content": ", resulting in " + }, + { + "bbox": [ + 130, + 342, + 480, + 430 + ], + "type": "inline_equation", + "content": "x'" + }, + { + "bbox": [ + 130, + 342, + 480, + 430 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 431, + 481, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 431, + 481, + 552 + ], + "spans": [ + { + "bbox": [ + 130, + 431, + 481, + 552 + ], + "type": "text", + "content": "This step ensures that both regions have the same number of columns of pixels. Next, we map each target region's vertical pixel column to its corresponding handle region's column. Each point's " + }, + { + "bbox": [ + 130, + 431, + 481, + 552 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 130, + 431, + 481, + 552 + ], + "type": "text", + "content": " in the target region is adjusted by subtracting the lower boundary of its vertical column " + }, + { + "bbox": [ + 130, + 431, + 481, + 552 + ], + "type": "inline_equation", + "content": "\\min(y_i^T \\mid x_i^T = x)" + }, + { + "bbox": [ + 130, + 431, + 481, + 552 + ], + "type": "text", + "content": " and is normalized by its column's y-coordinate range " + }, + { + "bbox": [ + 130, + 431, + 481, + 552 + ], + "type": "inline_equation", + "content": "\\max(y_i^T \\mid x_i^T = x) - \\min(y_i^T \\mid x_i^T = x)" + }, + { + "bbox": [ + 130, + 431, + 481, + 552 + ], + "type": "text", + "content": ", followed by being scaled to the corresponding handle region's column's y-coordinate range " + }, + { + "bbox": [ + 130, + 431, + 481, + 552 + ], + "type": "inline_equation", + "content": "\\max(y_i^H \\mid x_i^H = x') - \\min(y_i^H \\mid x_i^H = x')" + }, + { + "bbox": [ + 130, + 431, + 481, + 552 + ], + "type": "text", + "content": ". The notation " + }, + { + "bbox": [ + 130, + 431, + 481, + 552 + ], + "type": "inline_equation", + "content": "f(\\cdot \\mid \\cdot)" + }, + { + "bbox": [ + 130, + 431, + 481, + 552 + ], + "type": "text", + "content": " represents a function " + }, + { + "bbox": [ + 130, + 431, + 481, + 552 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 130, + 431, + 481, + 552 + ], + "type": "text", + "content": " (e.g., min or max) that takes a variable and a condition, and returns the function's value for the variable among points satisfying the condition. The whole process is summarised in Algorithm 1." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 553, + 481, + 638 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 553, + 481, + 638 + ], + "spans": [ + { + "bbox": [ + 130, + 553, + 481, + 638 + ], + "type": "text", + "content": "Main Pipeline. As illustrated in Fig. 3, RegionDrag utilizes the image editing pipeline mentioned in Sec. 3.1. Initially, the latent representation of the image " + }, + { + "bbox": [ + 130, + 553, + 481, + 638 + ], + "type": "inline_equation", + "content": "z_0" + }, + { + "bbox": [ + 130, + 553, + 481, + 638 + ], + "type": "text", + "content": " is inverted to " + }, + { + "bbox": [ + 130, + 553, + 481, + 638 + ], + "type": "inline_equation", + "content": "z_{t'}" + }, + { + "bbox": [ + 130, + 553, + 481, + 638 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 553, + 481, + 638 + ], + "type": "inline_equation", + "content": "t'" + }, + { + "bbox": [ + 130, + 553, + 481, + 638 + ], + "type": "text", + "content": " is a selected timestep prior to maximum timestep " + }, + { + "bbox": [ + 130, + 553, + 481, + 638 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 130, + 553, + 481, + 638 + ], + "type": "text", + "content": ". Each intermediate step " + }, + { + "bbox": [ + 130, + 553, + 481, + 638 + ], + "type": "inline_equation", + "content": "z_0, z_1, \\ldots, z_{t'}" + }, + { + "bbox": [ + 130, + 553, + 481, + 638 + ], + "type": "text", + "content": " is cached during inversion for future use. We then duplicate " + }, + { + "bbox": [ + 130, + 553, + 481, + 638 + ], + "type": "inline_equation", + "content": "z_{t'}" + }, + { + "bbox": [ + 130, + 553, + 481, + 638 + ], + "type": "text", + "content": " and denote the copy as " + }, + { + "bbox": [ + 130, + 553, + 481, + 638 + ], + "type": "inline_equation", + "content": "z_{t'}'" + }, + { + "bbox": [ + 130, + 553, + 481, + 638 + ], + "type": "text", + "content": ". The handle regions of " + }, + { + "bbox": [ + 130, + 553, + 481, + 638 + ], + "type": "inline_equation", + "content": "z_{t'}'" + }, + { + "bbox": [ + 130, + 553, + 481, + 638 + ], + "type": "text", + "content": " are blended with Gaussian noise " + }, + { + "bbox": [ + 130, + 553, + 481, + 638 + ], + "type": "inline_equation", + "content": "\\varepsilon" + }, + { + "bbox": [ + 130, + 553, + 481, + 638 + ], + "type": "text", + "content": " according to the blending function " + }, + { + "bbox": [ + 130, + 553, + 481, + 638 + ], + "type": "inline_equation", + "content": "r_{\\alpha}(z, H)" + }, + { + "bbox": [ + 130, + 553, + 481, + 638 + ], + "type": "text", + "content": " defined in Eq. (5):" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 164, + 651, + 481, + 667 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 164, + 651, + 481, + 667 + ], + "spans": [ + { + "bbox": [ + 164, + 651, + 481, + 667 + ], + "type": "interline_equation", + "content": "r _ {\\alpha} (z, H) = (1 - H) \\cdot z + H \\cdot (\\sqrt {1 - \\alpha^ {2}} \\cdot z + \\alpha \\cdot \\varepsilon), \\quad \\varepsilon \\sim \\mathcal {N} (0, I), \\qquad (5)", + "image_path": "fc2e185c12d87584ce3e781751fe77d5575a84d4f21f4d9f12b00830fcd3a731.jpg" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 202, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 202, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 202, + 101 + ], + "type": "text", + "content": "Lu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 369 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 369 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 369 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 116, + 482, + 369 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 130, + 116, + 482, + 369 + ], + "type": "text", + "content": " is a binary mask, with the handle region assigned a value of 1 and " + }, + { + "bbox": [ + 130, + 116, + 482, + 369 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 130, + 116, + 482, + 369 + ], + "type": "text", + "content": " is a blending coefficient that ranges from 0 to 1, which governs the strength of the blending effect. The noise " + }, + { + "bbox": [ + 130, + 116, + 482, + 369 + ], + "type": "inline_equation", + "content": "\\varepsilon" + }, + { + "bbox": [ + 130, + 116, + 482, + 369 + ], + "type": "text", + "content": " is drawn from a Gaussian distribution with mean 0 and identity covariance matrix, denoted by " + }, + { + "bbox": [ + 130, + 116, + 482, + 369 + ], + "type": "inline_equation", + "content": "\\mathcal{N}(0,I)" + }, + { + "bbox": [ + 130, + 116, + 482, + 369 + ], + "type": "text", + "content": ". If " + }, + { + "bbox": [ + 130, + 116, + 482, + 369 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 130, + 116, + 482, + 369 + ], + "type": "text", + "content": " is set to a lower value, less noise is added, which preserves more of the original image's features and details in the handle regions. After blending with Gaussian noise, " + }, + { + "bbox": [ + 130, + 116, + 482, + 369 + ], + "type": "inline_equation", + "content": "z_{t'}'" + }, + { + "bbox": [ + 130, + 116, + 482, + 369 + ], + "type": "text", + "content": " goes through the denoising process using the DDPM sampler, and the dragging is conducted in a copy-paste manner. At a denoising timestep " + }, + { + "bbox": [ + 130, + 116, + 482, + 369 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 116, + 482, + 369 + ], + "type": "text", + "content": ", we extract latent representation within the handle region of " + }, + { + "bbox": [ + 130, + 116, + 482, + 369 + ], + "type": "inline_equation", + "content": "z_t" + }, + { + "bbox": [ + 130, + 116, + 482, + 369 + ], + "type": "text", + "content": " and map it to the target region of " + }, + { + "bbox": [ + 130, + 116, + 482, + 369 + ], + "type": "inline_equation", + "content": "z_{t}'" + }, + { + "bbox": [ + 130, + 116, + 482, + 369 + ], + "type": "text", + "content": " according to the dense mapping computed by either geometric transformation or Algorithm 1. In the meantime, we employ mutual self-attention control [2] to help maintain the identity of images. In brief, when passing through the self-attention module of the UNet, the key and value " + }, + { + "bbox": [ + 130, + 116, + 482, + 369 + ], + "type": "inline_equation", + "content": "(k_t', v_t')" + }, + { + "bbox": [ + 130, + 116, + 482, + 369 + ], + "type": "text", + "content": " used when denoising " + }, + { + "bbox": [ + 130, + 116, + 482, + 369 + ], + "type": "inline_equation", + "content": "z_t'" + }, + { + "bbox": [ + 130, + 116, + 482, + 369 + ], + "type": "text", + "content": " are replaced with those " + }, + { + "bbox": [ + 130, + 116, + 482, + 369 + ], + "type": "inline_equation", + "content": "(k_t, v_t)" + }, + { + "bbox": [ + 130, + 116, + 482, + 369 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 130, + 116, + 482, + 369 + ], + "type": "inline_equation", + "content": "z_t" + }, + { + "bbox": [ + 130, + 116, + 482, + 369 + ], + "type": "text", + "content": ". This allows the edited image to keep the layout and identity of the original image, thereby stabilizing the editing process. After " + }, + { + "bbox": [ + 130, + 116, + 482, + 369 + ], + "type": "inline_equation", + "content": "z_{t}'" + }, + { + "bbox": [ + 130, + 116, + 482, + 369 + ], + "type": "text", + "content": " is gradually denoised to " + }, + { + "bbox": [ + 130, + 116, + 482, + 369 + ], + "type": "inline_equation", + "content": "z_0'" + }, + { + "bbox": [ + 130, + 116, + 482, + 369 + ], + "type": "text", + "content": " (see Fig. 3), it is decoded to the edited image " + }, + { + "bbox": [ + 130, + 116, + 482, + 369 + ], + "type": "inline_equation", + "content": "x_0'" + }, + { + "bbox": [ + 130, + 116, + 482, + 369 + ], + "type": "text", + "content": ". Although RegionDrag is introduced with a single handle-target region pair, it supports multiple pairs of input " + }, + { + "bbox": [ + 130, + 116, + 482, + 369 + ], + "type": "inline_equation", + "content": "\\{(H_i, T_i)\\}_{i=1}^n" + }, + { + "bbox": [ + 130, + 116, + 482, + 369 + ], + "type": "text", + "content": ", allowing users to specify several modifications in a single editing session. Dense mapping is individually constructed for each region pair but collectively used in latent copy-paste operations. The entire pipeline is summarised in Algorithm 2." + } + ] + } + ], + "index": 2 + }, + { + "type": "code", + "bbox": [ + 132, + 416, + 482, + 661 + ], + "blocks": [ + { + "bbox": [ + 133, + 402, + 271, + 415 + ], + "lines": [ + { + "bbox": [ + 133, + 402, + 271, + 415 + ], + "spans": [ + { + "bbox": [ + 133, + 402, + 271, + 415 + ], + "type": "text", + "content": "Algorithm 2 Editing Pipeline" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "code_caption" + }, + { + "bbox": [ + 132, + 416, + 482, + 661 + ], + "lines": [ + { + "bbox": [ + 132, + 416, + 482, + 661 + ], + "spans": [ + { + "bbox": [ + 132, + 416, + 482, + 661 + ], + "type": "text", + "content": "Input: Image " + }, + { + "bbox": [ + 132, + 416, + 482, + 661 + ], + "type": "inline_equation", + "content": "x_0" + }, + { + "bbox": [ + 132, + 416, + 482, + 661 + ], + "type": "text", + "content": ", handle and target region pairs " + }, + { + "bbox": [ + 132, + 416, + 482, + 661 + ], + "type": "inline_equation", + "content": "\\{(H_i, T_i)\\}_{i=1}^n" + }, + { + "bbox": [ + 132, + 416, + 482, + 661 + ], + "type": "text", + "content": ". \nOutput: Edited image " + }, + { + "bbox": [ + 132, + 416, + 482, + 661 + ], + "type": "inline_equation", + "content": "x_0'" + }, + { + "bbox": [ + 132, + 416, + 482, + 661 + ], + "type": "text", + "content": ". \nRequire: Diffusion function " + }, + { + "bbox": [ + 132, + 416, + 482, + 661 + ], + "type": "inline_equation", + "content": "f_{s \\to t}" + }, + { + "bbox": [ + 132, + 416, + 482, + 661 + ], + "type": "text", + "content": " Eq. (1), VAE encoder " + }, + { + "bbox": [ + 132, + 416, + 482, + 661 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 132, + 416, + 482, + 661 + ], + "type": "text", + "content": ", VAE decoder " + }, + { + "bbox": [ + 132, + 416, + 482, + 661 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 132, + 416, + 482, + 661 + ], + "type": "text", + "content": ", region-to-point mapping function " + }, + { + "bbox": [ + 132, + 416, + 482, + 661 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 132, + 416, + 482, + 661 + ], + "type": "text", + "content": ", copy-paste function " + }, + { + "bbox": [ + 132, + 416, + 482, + 661 + ], + "type": "inline_equation", + "content": "CP" + }, + { + "bbox": [ + 132, + 416, + 482, + 661 + ], + "type": "text", + "content": " Eq. (3), handle resampling function " + }, + { + "bbox": [ + 132, + 416, + 482, + 661 + ], + "type": "inline_equation", + "content": "r_\\alpha" + }, + { + "bbox": [ + 132, + 416, + 482, + 661 + ], + "type": "text", + "content": " Eq. (5), copy-paste time interval " + }, + { + "bbox": [ + 132, + 416, + 482, + 661 + ], + "type": "inline_equation", + "content": "(t', t'')" + }, + { + "bbox": [ + 132, + 416, + 482, + 661 + ], + "type": "text", + "content": ". \n1: Preparation: \n2: " + }, + { + "bbox": [ + 132, + 416, + 482, + 661 + ], + "type": "inline_equation", + "content": "\\{(h_i, t_i)\\}_{i=1}^N \\gets m(\\{(H_i, T_i)\\}_{i=1}^n)" + }, + { + "bbox": [ + 132, + 416, + 482, + 661 + ], + "type": "text", + "content": " \n3: " + }, + { + "bbox": [ + 132, + 416, + 482, + 661 + ], + "type": "inline_equation", + "content": "z_0 \\gets E(x_0)" + }, + { + "bbox": [ + 132, + 416, + 482, + 661 + ], + "type": "text", + "content": " \n4: Inversion stage: \n5: for " + }, + { + "bbox": [ + 132, + 416, + 482, + 661 + ], + "type": "inline_equation", + "content": "t = 0" + }, + { + "bbox": [ + 132, + 416, + 482, + 661 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 132, + 416, + 482, + 661 + ], + "type": "inline_equation", + "content": "t' - 1" + }, + { + "bbox": [ + 132, + 416, + 482, + 661 + ], + "type": "text", + "content": " do \n6: " + }, + { + "bbox": [ + 132, + 416, + 482, + 661 + ], + "type": "inline_equation", + "content": "z_{t+1}, k_{t+1}, v_{t+1} \\gets f_{t \\to t+1}(z_t)" + }, + { + "bbox": [ + 132, + 416, + 482, + 661 + ], + "type": "text", + "content": " \n7: end for \n8: " + }, + { + "bbox": [ + 132, + 416, + 482, + 661 + ], + "type": "inline_equation", + "content": "z_{t'}' \\gets r_\\alpha(z_{t'}, \\cup_{i=1}^n H_i)" + }, + { + "bbox": [ + 132, + 416, + 482, + 661 + ], + "type": "text", + "content": " \n9: Denoising stage: \n10: for " + }, + { + "bbox": [ + 132, + 416, + 482, + 661 + ], + "type": "inline_equation", + "content": "t = t'" + }, + { + "bbox": [ + 132, + 416, + 482, + 661 + ], + "type": "text", + "content": " down to 1 do \n11: if " + }, + { + "bbox": [ + 132, + 416, + 482, + 661 + ], + "type": "inline_equation", + "content": "t \\geq t''" + }, + { + "bbox": [ + 132, + 416, + 482, + 661 + ], + "type": "text", + "content": " then \n12: " + }, + { + "bbox": [ + 132, + 416, + 482, + 661 + ], + "type": "inline_equation", + "content": "z_t' \\gets \\mathrm{CP}(z_t, h_{1:N}, z_t', t_{1:N})" + }, + { + "bbox": [ + 132, + 416, + 482, + 661 + ], + "type": "text", + "content": " \n13: end if \n14: " + }, + { + "bbox": [ + 132, + 416, + 482, + 661 + ], + "type": "inline_equation", + "content": "z_{t-1}' \\gets f_{t \\to t-1}(z_t', k_t' \\gets k_t, v_t' \\gets v_t)" + }, + { + "bbox": [ + 132, + 416, + 482, + 661 + ], + "type": "text", + "content": " \n15: end for \n16: " + }, + { + "bbox": [ + 132, + 416, + 482, + 661 + ], + "type": "inline_equation", + "content": "x_0' \\gets D(z_0')" + }, + { + "bbox": [ + 132, + 416, + 482, + 661 + ], + "type": "text", + "content": " \n17: return " + }, + { + "bbox": [ + 132, + 416, + 482, + 661 + ], + "type": "inline_equation", + "content": "x_0'" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "code_body" + } + ], + "index": 4, + "sub_type": "algorithm" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 165, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 165, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 165, + 91, + 447, + 102 + ], + "type": "text", + "content": "RegionDrag: Fast Region-Based Image Editing with Diffusion Models" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 282, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 282, + 129 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 282, + 129 + ], + "type": "text", + "content": "4 Experimental Results" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 143, + 204, + 154 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 143, + 204, + 154 + ], + "spans": [ + { + "bbox": [ + 132, + 143, + 204, + 154 + ], + "type": "text", + "content": "4.1 Datasets" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 165, + 481, + 274 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 165, + 481, + 274 + ], + "spans": [ + { + "bbox": [ + 130, + 165, + 481, + 274 + ], + "type": "text", + "content": "SDE-Drag [18] and DragDiffusion [24] have each introduced a dataset named DragBench to evaluate the performance of point-drag-based methods. For clarity, we refer to the one by SDE-Drag as DragBench-S, which contains 100 samples, and to the one by DragDiffusion as DragBench-D, which includes 205 samples. Each sample comprises an image, a descriptive prompt, an optional mask delineating the editing region, and at least one pair of points that reflects the user's intention. Empirically, DragBench-S offers relatively more detailed prompts and includes a mask in 56 of its samples, while DragBench-D is characterized by more complex intentions and features carefully crafted masks in each of its samples." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 274, + 481, + 393 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 274, + 481, + 393 + ], + "spans": [ + { + "bbox": [ + 130, + 274, + 481, + 393 + ], + "type": "text", + "content": "To evaluate region-based editing, we introduce two new benchmarks: DragBench-SR and DragBench-DR (R is short for 'Region'), which are modified versions of DragBench-S and DragBench-D, respectively. These benchmarks are consistent with their point-drag-based counterparts in terms of images, prompts, and masks but differ by reflecting the user's intention through regions instead of points. DragBench-SR and DragBench-DR consist of 8 and 23 polygon-annotated samples, respectively, with the remainders of each dataset using brushes. For a handle-target region pair, the median number of equivalent point pairs is 267 in DragBench-SR and 201 in DragBench-DR. We visualize the distribution of the number of transformed point pairs in Fig. 4." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 394, + 481, + 478 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 394, + 481, + 478 + ], + "spans": [ + { + "bbox": [ + 130, + 394, + 481, + 478 + ], + "type": "text", + "content": "During the annotation process, we ensure that the region pairs are aligned with the point pairs to achieve the same editing effect. To simulate how real users would typically draw handle-target region pairs, we do not meticulously draw each region. According to our observations, each region-based sample is annotated in approximately 10 seconds. This approach allows for a more realistic evaluation of the region-based editing methods while maintaining consistency with the point-based datasets." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 498, + 257, + 510 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 498, + 257, + 510 + ], + "spans": [ + { + "bbox": [ + 132, + 498, + 257, + 510 + ], + "type": "text", + "content": "4.2 Evaluation Metrics" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 521, + 481, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 481, + 616 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 481, + 616 + ], + "type": "text", + "content": "LPIPS: We follow [24] and use Learned Perceptual Image Patch Similarity (LPIPS) v0.1 [30] to measure the identity similarity between the edited image and the original image. LPIPS computes the AlexNet [13] feature distances between the image pairs. A high LPIPS score indicates that unexpected identity changes or artifacts occur due to editing. A lower LPIPS score suggests that the object's identity has been well-preserved during editing; however, it does not necessarily imply a better drag edit, as two identical images would yield an LPIPS score of 0." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "type": "text", + "content": "Mean Distance (MD): DragDiffusion [24] introduces the MD metric to assess how well an approach moves the handle points' content to the target points. To identify where the handle points have been moved to by the method, DragDiffusion employs DIFT [26] to find the most similar points to the handle points" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 202, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 202, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 202, + 100 + ], + "type": "text", + "content": "Lu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 144, + 114, + 284, + 202 + ], + "blocks": [ + { + "bbox": [ + 144, + 114, + 284, + 202 + ], + "lines": [ + { + "bbox": [ + 144, + 114, + 284, + 202 + ], + "spans": [ + { + "bbox": [ + 144, + 114, + 284, + 202 + ], + "type": "image", + "image_path": "2c225e96570194b0602d6fb448e941cbee5e817a4df55a5019505fc936766d1c.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 141, + 213, + 301, + 257 + ], + "lines": [ + { + "bbox": [ + 141, + 213, + 301, + 257 + ], + "spans": [ + { + "bbox": [ + 141, + 213, + 301, + 257 + ], + "type": "text", + "content": "Fig.4: Log-transformed frequency distribution of equivalent point pair counts in DragBench-SR and DragBench-DR." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 309, + 114, + 470, + 200 + ], + "blocks": [ + { + "bbox": [ + 309, + 114, + 470, + 200 + ], + "lines": [ + { + "bbox": [ + 309, + 114, + 470, + 200 + ], + "spans": [ + { + "bbox": [ + 309, + 114, + 470, + 200 + ], + "type": "image", + "image_path": "0c0d5ffacdd813f4c7936a20b1d678be40b7874d3126658456d0878e64320b3d.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 307, + 213, + 470, + 247 + ], + "lines": [ + { + "bbox": [ + 307, + 213, + 470, + 247 + ], + "spans": [ + { + "bbox": [ + 307, + 213, + 470, + 247 + ], + "type": "text", + "content": "Fig. 5: Illustration of MD computation. Dragged handle points will be searched within the search mask." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 280, + 480, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 280, + 480, + 365 + ], + "spans": [ + { + "bbox": [ + 130, + 280, + 480, + 365 + ], + "type": "inline_equation", + "content": "h_{1:n}" + }, + { + "bbox": [ + 130, + 280, + 480, + 365 + ], + "type": "text", + "content": " in the entire edited image and denotes them as " + }, + { + "bbox": [ + 130, + 280, + 480, + 365 + ], + "type": "inline_equation", + "content": "h_{1:n}^{\\prime}" + }, + { + "bbox": [ + 130, + 280, + 480, + 365 + ], + "type": "text", + "content": ". It then uses the normalized Euclidean distance between DIFT-matched points and the true target positions " + }, + { + "bbox": [ + 130, + 280, + 480, + 365 + ], + "type": "inline_equation", + "content": "t_{1:n}" + }, + { + "bbox": [ + 130, + 280, + 480, + 365 + ], + "type": "text", + "content": " as the metric. However, we believe that we only have to search the area around the handle points and their corresponding true target points for " + }, + { + "bbox": [ + 130, + 280, + 480, + 365 + ], + "type": "inline_equation", + "content": "h_{1:n}^{\\prime}" + }, + { + "bbox": [ + 130, + 280, + 480, + 365 + ], + "type": "text", + "content": " instead of the entire image, to avoid DIFT mistakenly identifying irrelevant points in the images as " + }, + { + "bbox": [ + 130, + 280, + 480, + 365 + ], + "type": "inline_equation", + "content": "h^\\prime" + }, + { + "bbox": [ + 130, + 280, + 480, + 365 + ], + "type": "text", + "content": " and excessively penalizing certain methods. Formally, given a dragging point pair " + }, + { + "bbox": [ + 130, + 280, + 480, + 365 + ], + "type": "inline_equation", + "content": "\\{h,t\\}" + }, + { + "bbox": [ + 130, + 280, + 480, + 365 + ], + "type": "text", + "content": ", we define its searching mask as" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 204, + 373, + 480, + 404 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 204, + 373, + 480, + 404 + ], + "spans": [ + { + "bbox": [ + 204, + 373, + 480, + 404 + ], + "type": "interline_equation", + "content": "M (x) = \\left\\{ \\begin{array}{l l} 1 & \\text {i f} \\min (d (x, h), d (x, t)) < \\frac {d (h , t)}{\\sqrt {2}} \\\\ 0 & \\text {o t h e r w i s e} \\end{array} \\right., \\tag {6}", + "image_path": "4ced856b2d74bc0a5fd35a5d4ffd4c71e7933ebec99859a26c90c637194a7227.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 413, + 479, + 439 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 413, + 479, + 439 + ], + "spans": [ + { + "bbox": [ + 131, + 413, + 479, + 439 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 413, + 479, + 439 + ], + "type": "inline_equation", + "content": "\\mathrm{d}(\\cdot, \\cdot)" + }, + { + "bbox": [ + 131, + 413, + 479, + 439 + ], + "type": "text", + "content": " calculates the normalized Euclidean distance between two points " + }, + { + "bbox": [ + 131, + 413, + 479, + 439 + ], + "type": "inline_equation", + "content": "(x_1, y_1)" + }, + { + "bbox": [ + 131, + 413, + 479, + 439 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 413, + 479, + 439 + ], + "type": "inline_equation", + "content": "(x_2, y_2)" + }, + { + "bbox": [ + 131, + 413, + 479, + 439 + ], + "type": "text", + "content": " by" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 192, + 446, + 480, + 479 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 192, + 446, + 480, + 479 + ], + "spans": [ + { + "bbox": [ + 192, + 446, + 480, + 479 + ], + "type": "interline_equation", + "content": "d \\left(\\left(x _ {1}, y _ {1}\\right), \\left(x _ {2}, y _ {2}\\right)\\right) = \\sqrt {\\left(\\frac {x _ {2} - x _ {1}}{W}\\right) ^ {2} + \\left(\\frac {y _ {2} - y _ {1}}{H}\\right) ^ {2}}, \\tag {7}", + "image_path": "4df7208d306fc2e25f43367f939734c351bc6e979fa5b496d612e982ba1415fd.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 486, + 482, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 486, + 482, + 533 + ], + "spans": [ + { + "bbox": [ + 130, + 486, + 482, + 533 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 486, + 482, + 533 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 130, + 486, + 482, + 533 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 486, + 482, + 533 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 130, + 486, + 482, + 533 + ], + "type": "text", + "content": " are width and height of the image. We provide examples of searching masks in Fig. 5. MD is then defined as the average of the normalized distances " + }, + { + "bbox": [ + 130, + 486, + 482, + 533 + ], + "type": "inline_equation", + "content": "\\bar{d}" + }, + { + "bbox": [ + 130, + 486, + 482, + 533 + ], + "type": "text", + "content": " between the target points and the DIFT-matched points across all " + }, + { + "bbox": [ + 130, + 486, + 482, + 533 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 130, + 486, + 482, + 533 + ], + "type": "text", + "content": " points:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 264, + 533, + 480, + 564 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 264, + 533, + 480, + 564 + ], + "spans": [ + { + "bbox": [ + 264, + 533, + 480, + 564 + ], + "type": "interline_equation", + "content": "\\bar {d} = \\frac {1}{n} \\sum_ {i = 1} ^ {n} d \\left(t _ {i}, h _ {i} ^ {\\prime}\\right). \\tag {8}", + "image_path": "82391ca3a57ef77c809d183af5c313c790cd8702dc5f8340c68912243124eb81.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 586, + 279, + 597 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 586, + 279, + 597 + ], + "spans": [ + { + "bbox": [ + 132, + 586, + 279, + 597 + ], + "type": "text", + "content": "4.3 Implementation Details" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": "Our method is implemented in Python using the HuggingFace [22] and PyTorch [20] libraries. We employ Stable Diffusion v1-5 as our diffusion model with an image size of " + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": ", being consistent with previous diffusion-based dragging methods. A DDPM sampler is utilized for both the inversion and denoising sampling processes, configured to use a total of 20 steps. The latent" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 165, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 165, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 165, + 91, + 448, + 102 + ], + "type": "text", + "content": "RegionDrag: Fast Region-Based Image Editing with Diffusion Models" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 168, + 114, + 442, + 193 + ], + "blocks": [ + { + "bbox": [ + 168, + 114, + 442, + 193 + ], + "lines": [ + { + "bbox": [ + 168, + 114, + 442, + 193 + ], + "spans": [ + { + "bbox": [ + 168, + 114, + 442, + 193 + ], + "type": "table", + "html": "
MethodDragBench-S(R)DragBench-D(R)
Time (↓)MD (↓)LPIPS (↓)MD (↓)LPIPS (↓)
SDE-Drag [18]126.17.512.48.114.9
DragDiffusion [24]177.77.018.06.711.5
DiffEditor [17]43.123.617.622.110.9
Ours1.56.49.96.69.2
", + "image_path": "9a2c254e51053ef29ebc2c30affdbef9834d9e7059eba659716a863aad06bb29.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 201, + 479, + 234 + ], + "lines": [ + { + "bbox": [ + 133, + 201, + 479, + 234 + ], + "spans": [ + { + "bbox": [ + 133, + 201, + 479, + 234 + ], + "type": "text", + "content": "Table 1: Comparisons of our method with baseline methods using MD(×100) and LPIPS(×100) metrics on DragBench-S(R) and DragBench-D(R) datasets. The time is measured in seconds and averaged across both datasets. The image size is 512×512." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 130, + 260, + 480, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 260, + 480, + 319 + ], + "spans": [ + { + "bbox": [ + 130, + 260, + 480, + 319 + ], + "type": "text", + "content": "representation is inverted to timestep " + }, + { + "bbox": [ + 130, + 260, + 480, + 319 + ], + "type": "inline_equation", + "content": "t' = 500" + }, + { + "bbox": [ + 130, + 260, + 480, + 319 + ], + "type": "text", + "content": " out of the total 1000 steps in SD1-5. Consequently, we perform 10 inversion steps and 10 denoising steps. Mutual self-attention control is enabled throughout all timesteps, and the latent copy-paste operation is terminated at " + }, + { + "bbox": [ + 130, + 260, + 480, + 319 + ], + "type": "inline_equation", + "content": "t'' = 200" + }, + { + "bbox": [ + 130, + 260, + 480, + 319 + ], + "type": "text", + "content": ". The noise weight " + }, + { + "bbox": [ + 130, + 260, + 480, + 319 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 130, + 260, + 480, + 319 + ], + "type": "text", + "content": " is set to 1. All experimental results are obtained on an NVIDIA Tesla V100 GPU." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 342, + 207, + 354 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 342, + 207, + 354 + ], + "spans": [ + { + "bbox": [ + 132, + 342, + 207, + 354 + ], + "type": "text", + "content": "4.4Baselines" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 367, + 480, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 367, + 480, + 462 + ], + "spans": [ + { + "bbox": [ + 130, + 367, + 480, + 462 + ], + "type": "text", + "content": "We compare RegionDrag with point-drag-based diffusion methods, including DragDiffusion [24], SDE-Drag [18], and DiffEditor [17]. GAN-based methods are excluded due to their limitations in editing the diverse images in the DragBench-S and DragBench-D datasets, as they require domain-specific StyleGAN checkpoints. Diffusion-based methods, which outperform GAN-based methods in editing tasks, are better suited for our evaluation. Execution times are averaged across both datasets, and all methods are tested on the same device using publicly released code." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 485, + 281, + 498 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 485, + 281, + 498 + ], + "spans": [ + { + "bbox": [ + 132, + 485, + 281, + 498 + ], + "type": "text", + "content": "4.5 Quantitative Evaluation" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 510, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 481, + 665 + ], + "type": "text", + "content": "To quantitatively evaluate the editing performance of the methods, we employ LPIPS and Mean Distance as metrics, multiplying both by 100 for illustration purposes. As demonstrated in Table 1, RegionDrag significantly outperforms those computationally expensive point-based methods on both DragBench-S(R) and DragBench-D(R) datasets. These results highlight RegionDrag's superior performance in maintaining image consistency while achieving competitive editing results across different datasets. In addition to its effectiveness, RegionDrag also excels in terms of efficiency. RegionDrag achieves fast inference speed, requiring approximately 1.5 seconds to edit a " + }, + { + "bbox": [ + 130, + 510, + 481, + 665 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 130, + 510, + 481, + 665 + ], + "type": "text", + "content": " image, which is 20 times faster than the second-fastest method and 100 times faster than DragDiffusion [24]. The inference time for RegionDrag is comparable to generating an image with 20 steps using SD1-5, given that the copy-paste operations introduce negligible computational overhead." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 202, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 202, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 202, + 100 + ], + "type": "text", + "content": "Lu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 134, + 114, + 480, + 374 + ], + "blocks": [ + { + "bbox": [ + 134, + 114, + 480, + 374 + ], + "lines": [ + { + "bbox": [ + 134, + 114, + 480, + 374 + ], + "spans": [ + { + "bbox": [ + 134, + 114, + 480, + 374 + ], + "type": "image", + "image_path": "2a9629d4d04440744af6eccd89f8ad7634322d4fe4891251a15f095350ffaab4.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 384, + 480, + 407 + ], + "lines": [ + { + "bbox": [ + 132, + 384, + 480, + 407 + ], + "spans": [ + { + "bbox": [ + 132, + 384, + 480, + 407 + ], + "type": "text", + "content": "Fig. 6: Qualitative comparisons with baseline methods. Handle regions and target regions are respectively denoted by red and blue masks." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 430, + 257, + 441 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 430, + 257, + 441 + ], + "spans": [ + { + "bbox": [ + 132, + 430, + 257, + 441 + ], + "type": "text", + "content": "4.6 Qualitative Results" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 449, + 482, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 449, + 482, + 510 + ], + "spans": [ + { + "bbox": [ + 131, + 449, + 482, + 510 + ], + "type": "text", + "content": "Figure 6 compares examples of point-drag-based and region-based editing inputs and their corresponding results, demonstrating the effectiveness of Region-Drag. Our region-based method utilizes the comprehensive context provided by annotated regions to target desired modifications while preserving the overall coherence of the image, outperforming point-drag-based editing methods." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 526, + 239, + 538 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 526, + 239, + 538 + ], + "spans": [ + { + "bbox": [ + 132, + 526, + 239, + 538 + ], + "type": "text", + "content": "4.7 Ablation Study" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 546, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 546, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 131, + 546, + 482, + 666 + ], + "type": "text", + "content": "We argue that the sparsity of point inputs leads to inferior editing results. To quantitatively demonstrate this, we conducted tests on the DragBench-DR dataset by randomly selecting subsets of the equivalent transformed points within each sample and performing inference using these subsets. We gradually reduced the percentage of selected points to observe the impact on the MD metric. As illustrated in Fig. 7, the results exhibit a clear upward trend in MD as the percentage of utilized points decreases. This suggests that sparse point inputs provide weaker constraints on the output compared to region-based inputs, leading to unsatisfactory editing results. It confirms the benefits of employing region-based inputs in RegionDrag." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 165, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 165, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 165, + 91, + 448, + 102 + ], + "type": "text", + "content": "RegionDrag: Fast Region-Based Image Editing with Diffusion Models" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 116, + 265, + 238 + ], + "blocks": [ + { + "bbox": [ + 133, + 116, + 265, + 238 + ], + "lines": [ + { + "bbox": [ + 133, + 116, + 265, + 238 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 265, + 238 + ], + "type": "image", + "image_path": "0d3d76007f3f0d760792441ff9132a6ebc5ee77756ecffecacb995d38764d0a4.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 249, + 266, + 282 + ], + "lines": [ + { + "bbox": [ + 131, + 249, + 266, + 282 + ], + "spans": [ + { + "bbox": [ + 131, + 249, + 266, + 282 + ], + "type": "text", + "content": "Fig.7: Ablation study on the impact of the percentage of inputted transformed points." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 282, + 114, + 352, + 229 + ], + "blocks": [ + { + "bbox": [ + 282, + 114, + 352, + 229 + ], + "lines": [ + { + "bbox": [ + 282, + 114, + 352, + 229 + ], + "spans": [ + { + "bbox": [ + 282, + 114, + 352, + 229 + ], + "type": "image", + "image_path": "936a0cecf9b3b355d6aebe1f627bff86121ea6a3bea97e90d75f1314a602f113.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 280, + 251, + 480, + 274 + ], + "lines": [ + { + "bbox": [ + 280, + 251, + 480, + 274 + ], + "spans": [ + { + "bbox": [ + 280, + 251, + 480, + 274 + ], + "type": "text", + "content": "Fig. 8: Qualitative examples illustrating the impact of multi-step copy-paste." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 355, + 114, + 480, + 229 + ], + "blocks": [ + { + "bbox": [ + 355, + 114, + 480, + 229 + ], + "lines": [ + { + "bbox": [ + 355, + 114, + 480, + 229 + ], + "spans": [ + { + "bbox": [ + 355, + 114, + 480, + 229 + ], + "type": "image", + "image_path": "961aeb8ba451318f9b7a5996afb508550ce2094fddad7b33e4535d3a853f2d0c.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 306, + 482, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 306, + 482, + 380 + ], + "spans": [ + { + "bbox": [ + 130, + 306, + 482, + 380 + ], + "type": "text", + "content": "We copy and paste image's latent representation over a time interval during denoising. To validate this design, we compare it to copy-paste only at the initial denoising timestep. Figure 8 shows that editing just at the initial step can yield unpredictable results because the edits may be lost in subsequent denoising phases. Multi-step copy-paste solves this by providing extra guidance at smaller timesteps while preserving image fidelity." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 396, + 220, + 409 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 396, + 220, + 409 + ], + "spans": [ + { + "bbox": [ + 132, + 396, + 220, + 409 + ], + "type": "text", + "content": "5 Conclusion" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 422, + 482, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 422, + 482, + 542 + ], + "spans": [ + { + "bbox": [ + 130, + 422, + 482, + 542 + ], + "type": "text", + "content": "In this paper, we have introduced an efficient and effective region-based editing framework, RegionDrag, for high-fidelity image editing. Unlike existing approaches that utilize point-drag-based editing, RegionDrag reconsiders the editing problem from the region perspective. RegionDrag allows for editing in a single step through copying and pasting the latent representation and self-attention features of the image, which not only provides excellent efficiency but also achieves superior editing performance. Furthermore, we have introduced two new benchmarks, DragBench-SR and DragBench-DR based on existing datasets, for the evaluation of region-based editing. Experimental results have consistently demonstrated the superior efficiency and editing performance of our method." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 560, + 241, + 574 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 560, + 241, + 574 + ], + "spans": [ + { + "bbox": [ + 132, + 560, + 241, + 574 + ], + "type": "text", + "content": "Acknowledgments" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 131, + 585, + 481, + 609 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 585, + 481, + 609 + ], + "spans": [ + { + "bbox": [ + 131, + 585, + 481, + 609 + ], + "type": "text", + "content": "Kai Han and Jingyi Lu would like to acknowledge the support of the Hong Kong Research Grants Council - General Research Fund (Grant No.: 17211024)." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 202, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 202, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 202, + 100 + ], + "type": "text", + "content": "Lu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 197, + 126 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 197, + 126 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 197, + 126 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 141, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 23, + "blocks": [ + { + "bbox": [ + 138, + 141, + 480, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 141, + 480, + 163 + ], + "spans": [ + { + "bbox": [ + 138, + 141, + 480, + 163 + ], + "type": "text", + "content": "1. Brooks, T., Holynski, A., Efros, A.A.: Instructpix2pix: Learning to follow image editing instructions. In: CVPR (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 165, + 480, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 165, + 480, + 196 + ], + "spans": [ + { + "bbox": [ + 138, + 165, + 480, + 196 + ], + "type": "text", + "content": "2. Cao, M., Wang, X., Qi, Z., Shan, Y., Qie, X., Zheng, Y.: Masactrl: Tuning-free mutual self-attention control for consistent image synthesis and editing. arXiv preprint arXiv:2304.08465 (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 198, + 480, + 219 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 198, + 480, + 219 + ], + "spans": [ + { + "bbox": [ + 138, + 198, + 480, + 219 + ], + "type": "text", + "content": "3. Creswell, A., Bharath, A.A.: Inverting the generator of a generative adversarial network. IEEE TNNLS (2018)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 220, + 480, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 220, + 480, + 241 + ], + "spans": [ + { + "bbox": [ + 138, + 220, + 480, + 241 + ], + "type": "text", + "content": "4. Endo, Y.: User-controllable latent transformer for stylegan image layout editing. In: CGF (2022)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 243, + 480, + 264 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 243, + 480, + 264 + ], + "spans": [ + { + "bbox": [ + 138, + 243, + 480, + 264 + ], + "type": "text", + "content": "5. Epstein, D., Jabri, A., Poole, B., Efros, A., Holynski, A.: Diffusion self-guidance for controllable image generation. In: NeurIPS (2024)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 266, + 480, + 286 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 266, + 480, + 286 + ], + "spans": [ + { + "bbox": [ + 138, + 266, + 480, + 286 + ], + "type": "text", + "content": "6. Goodfellow, I., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., Courville, A., Bengio, Y.: Generative adversarial nets. In: NeurIPS (2014)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 288, + 480, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 288, + 480, + 308 + ], + "spans": [ + { + "bbox": [ + 138, + 288, + 480, + 308 + ], + "type": "text", + "content": "7. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. In: NeurIPS (2020)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 310, + 480, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 310, + 480, + 342 + ], + "spans": [ + { + "bbox": [ + 138, + 310, + 480, + 342 + ], + "type": "text", + "content": "8. Hu, E.J., Shen, Y., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., Chen, W.: Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685 (2021)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 343, + 480, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 343, + 480, + 364 + ], + "spans": [ + { + "bbox": [ + 138, + 343, + 480, + 364 + ], + "type": "text", + "content": "9. Kang, M., Zhu, J.Y., Zhang, R., Park, J., Shechtman, E., Paris, S., Park, T.: Scaling up gans for text-to-image synthesis. In: CVPR (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 365, + 480, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 365, + 480, + 387 + ], + "spans": [ + { + "bbox": [ + 138, + 365, + 480, + 387 + ], + "type": "text", + "content": "0. Karras, T., Aittala, M., Laine, S., Härkönen, E., Hellsten, J., Lehtinen, J., Aila, T.: Alias-free generative adversarial networks. In: NeurIPS (2021)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 388, + 480, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 388, + 480, + 409 + ], + "spans": [ + { + "bbox": [ + 138, + 388, + 480, + 409 + ], + "type": "text", + "content": "1. Karras, T., Laine, S., Aittala, M., Hellsten, J., Lehtinen, J., Aila, T.: Analyzing and improving the image quality of stylegan. In: CVPR (2020)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 411, + 480, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 411, + 480, + 431 + ], + "spans": [ + { + "bbox": [ + 138, + 411, + 480, + 431 + ], + "type": "text", + "content": "2. Kawar, B., Zada, S., Lang, O., Tov, O., Chang, H., Dekel, T., Mosseri, I., Irani, M.: Imagic: Text-based real image editing with diffusion models. In: CVPR (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 433, + 480, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 433, + 480, + 453 + ], + "spans": [ + { + "bbox": [ + 138, + 433, + 480, + 453 + ], + "type": "text", + "content": "3. Krizhevsky, A., Sutskever, I., Hinton, G.E.: Imagenet classification with deep convolutional neural networks. NeurIPS (2012)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 455, + 480, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 455, + 480, + 487 + ], + "spans": [ + { + "bbox": [ + 138, + 455, + 480, + 487 + ], + "type": "text", + "content": "4. Ling, P., Chen, L., Zhang, P., Chen, H., Jin, Y.: Freedrag: Point tracking is not you need for interactive point-based image editing. arXiv preprint arXiv:2307.04684 (2023)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 488, + 480, + 509 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 488, + 480, + 509 + ], + "spans": [ + { + "bbox": [ + 138, + 488, + 480, + 509 + ], + "type": "text", + "content": "5. Luo, M., Cheng, W., Yang, J.: Rotationdrag: Point-based image editing with rotated diffusion features. arXiv preprint arXiv:2401.06442 (2024)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 511, + 480, + 532 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 511, + 480, + 532 + ], + "spans": [ + { + "bbox": [ + 138, + 511, + 480, + 532 + ], + "type": "text", + "content": "6. Mou, C., Wang, X., Song, J., Shan, Y., Zhang, J.: Dragondiffusion: Enabling drag-style manipulation on diffusion models. arXiv preprint arXiv:2307.02421 (2023)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 138, + 533, + 480, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 533, + 480, + 564 + ], + "spans": [ + { + "bbox": [ + 138, + 533, + 480, + 564 + ], + "type": "text", + "content": "7. Mou, C., Wang, X., Song, J., Shan, Y., Zhang, J.: Diffeditor: Boosting accuracy and flexibility on diffusion-based image editing. arXiv preprint arXiv:2402.02583 (2024)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 138, + 567, + 480, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 567, + 480, + 597 + ], + "spans": [ + { + "bbox": [ + 138, + 567, + 480, + 597 + ], + "type": "text", + "content": "8. Nie, S., Guo, H.A., Lu, C., Zhou, Y., Zheng, C., Li, C.: The blessing of randomness: Sde beats ode in general diffusion-based image editing. arXiv preprint arXiv:2311.01410 (2023)" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 138, + 600, + 480, + 631 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 600, + 480, + 631 + ], + "spans": [ + { + "bbox": [ + 138, + 600, + 480, + 631 + ], + "type": "text", + "content": "9. Pan, X., Tewari, A., Leimkuhler, T., Liu, L., Meka, A., Theobalt, C.: Drag your gan: Interactive point-based manipulation on the generative image manifold. In: ACM SIGGRAPH (2023)" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 138, + 633, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 633, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 633, + 480, + 665 + ], + "type": "text", + "content": "20. Paszke, A., Gross, S., Massa, F., Lerer, A., Bradbury, J., Chanan, G., Killeen, T., Lin, Z., Gimelshein, N., Antiga, L., et al.: Pytorch: An imperative style, high-performance deep learning library. In: NeurIPS (2019)" + } + ] + } + ], + "index": 22 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 165, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 165, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 165, + 91, + 447, + 102 + ], + "type": "text", + "content": "RegionDrag: Fast Region-Based Image Editing with Diffusion Models" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 358 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 480, + 138 + ], + "type": "text", + "content": "21. Patashnik, O., Wu, Z., Shechtman, E., Cohen-Or, D., Lischinski, D.: Styleclip: Text-driven manipulation of stylegan imagery. In: ICCV (2021)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 139, + 481, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 139, + 481, + 171 + ], + "spans": [ + { + "bbox": [ + 132, + 139, + 481, + 171 + ], + "type": "text", + "content": "22. von Platen, P., Patil, S., Lozhkov, A., Cuenca, P., Lambert, N., Rasul, K., Davaadorj, M., Wolf, T.: Diffusers: State-of-the-art diffusion models. https://github.com/huggingface/diffusers (2022)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 172, + 480, + 193 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 172, + 480, + 193 + ], + "spans": [ + { + "bbox": [ + 132, + 172, + 480, + 193 + ], + "type": "text", + "content": "23. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: CVPR (2022)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 194, + 480, + 225 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 194, + 480, + 225 + ], + "spans": [ + { + "bbox": [ + 132, + 194, + 480, + 225 + ], + "type": "text", + "content": "24. Shi, Y., Xue, C., Pan, J., Zhang, W., Tan, V.Y., Bai, S.: Dragdiffusion: Harnessing diffusion models for interactive point-based image editing. arXiv preprint arXiv:2306.14435 (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 227, + 480, + 247 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 227, + 480, + 247 + ], + "spans": [ + { + "bbox": [ + 132, + 227, + 480, + 247 + ], + "type": "text", + "content": "25. Song, J., Meng, C., Ermon, S.: Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502 (2020)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 249, + 480, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 249, + 480, + 270 + ], + "spans": [ + { + "bbox": [ + 132, + 249, + 480, + 270 + ], + "type": "text", + "content": "26. Tang, L., Jia, M., Wang, Q., Phoo, C.P., Hariharan, B.: Emergent correspondence from image diffusion. In: NeurIPS (2024)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 270, + 480, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 270, + 480, + 293 + ], + "spans": [ + { + "bbox": [ + 132, + 270, + 480, + 293 + ], + "type": "text", + "content": "27. Wang, T., Zhang, Y., Fan, Y., Wang, J., Chen, Q.: High-fidelity gan inversion for image attribute editing. In: CVPR (2022)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 293, + 480, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 293, + 480, + 314 + ], + "spans": [ + { + "bbox": [ + 132, + 293, + 480, + 314 + ], + "type": "text", + "content": "28. Wu, C.H., De la Torre, F.: Unifying diffusion models' latent space, with applications to cyclediffusion and guidance. arXiv preprint arXiv:2210.05559 (2022)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 315, + 480, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 315, + 480, + 335 + ], + "spans": [ + { + "bbox": [ + 132, + 315, + 480, + 335 + ], + "type": "text", + "content": "29. Xia, W., Zhang, Y., Yang, Y., Xue, J.H., Zhou, B., Yang, M.H.: Gan inversion: A survey. IEEE TPAMI (2022)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 336, + 480, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 336, + 480, + 358 + ], + "spans": [ + { + "bbox": [ + 132, + 336, + 480, + 358 + ], + "type": "text", + "content": "30. Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual metric. In: CVPR (2018)" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 202, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 202, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 202, + 100 + ], + "type": "text", + "content": "Lu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Regularizing Dynamic Radiance Fields with Kinematic Fields/79573c22-a91b-4733-9726-bd12faf03469_content_list.json b/2024/Regularizing Dynamic Radiance Fields with Kinematic Fields/79573c22-a91b-4733-9726-bd12faf03469_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..8638016d90d6def431b44668e5132b601f773952 --- /dev/null +++ b/2024/Regularizing Dynamic Radiance Fields with Kinematic Fields/79573c22-a91b-4733-9726-bd12faf03469_content_list.json @@ -0,0 +1,1897 @@ +[ + { + "type": "text", + "text": "Regularizing Dynamic Radiance Fields with Kinematic Fields", + "text_level": 1, + "bbox": [ + 250, + 140, + 751, + 184 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Woobin $\\mathrm{Im}^{\\star 1,2}$ , Geonho Cha $^3$ , Sebin Lee $^1$ , Jumin Lee $^1$ , Juhyeong Seon $^1$ , Dongyoon Wee $^3$ , and Sung-Eui Yoon $^1$", + "bbox": [ + 240, + 210, + 759, + 244 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ KAIST {iwbn,seb.lee,jmlee,munuwazzi,sungeui}@kaist.ac.kr", + "bbox": [ + 307, + 253, + 694, + 281 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ Samsung Advanced Institute of Technology (SAIT)", + "bbox": [ + 323, + 282, + 676, + 296 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3 Naver Cloud", + "bbox": [ + 442, + 296, + 558, + 309 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{geonho.cha,dongyoon.wei}@navercorp.com", + "bbox": [ + 349, + 311, + 651, + 324 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. This paper presents a novel approach for reconstructing dynamic radiance fields from monocular videos. We integrate kinematics with dynamic radiance fields, bridging the gap between the sparse nature of monocular videos and the real-world physics. Our method introduces the kinematic field, capturing motion through kinematic quantities: velocity, acceleration, and jerk. The kinematic field is jointly learned with the dynamic radiance field by minimizing the photometric loss without motion ground truth. We further augment our method with physics-driven regularizers grounded in kinematics. We propose physics-driven regularizers that ensure the physical validity of predicted kinematic quantities, including advective acceleration and jerk. Additionally, we control the motion trajectory based on rigidity equations formed with the predicted kinematic quantities. In experiments, our method outperforms the state-of-the-arts by capturing physical motion patterns within challenging real-world monocular videos.", + "bbox": [ + 259, + 359, + 743, + 566 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Dynamic radiance fields $\\cdot$ Kinematic fields $\\cdot$ 3D video", + "bbox": [ + 261, + 580, + 705, + 595 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 619, + 375, + 635 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recognizing motion is crucial to how we visually perceive the world [13], and video is the medium that carries the motion to viewers in a remote place or at a different time. Through the lens of video, we watch and enjoy movies, user-generated contents, and remote video chats. While videos depict motion in a 2D format, transforming them into three-dimensional experiences can elevate our perception of the content to co-presence via applications such as teleportation [24].", + "bbox": [ + 212, + 650, + 787, + 755 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Dynamic radiance fields [5,9,11,16,19,27] have been studied as a promising representation for reconstructing 3D dynamic scenes; it enables realistic novel view and time synthesis, which can upgrade 2D videos into 3D videos. Since multi-view cameras can give more information about the geometry in a scene,", + "bbox": [ + 212, + 756, + 787, + 816 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* This work was done during an internship at Naver Cloud.", + "bbox": [ + 217, + 824, + 633, + 839 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "existing works have utilized camera rigs with dozens of cameras to capture multi-view video datasets [1, 5, 16]. Although high-quality multi-view videos result in undeniably better quality than monocular videos do, the multi-view capturing devices and environments are prohibitive in most cases.", + "bbox": [ + 212, + 146, + 782, + 205 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Thus, there have been efforts to reconstruct dynamic radiance fields from monocular videos [10, 18, 20]. Compared to multi-view videos, monocular videos provide sparser information, which makes the reconstruction more challenging. For instance, there is scale ambiguity in monocular video, making it difficult to accurately locate an object within a space [34]. Thus, dynamic scene reconstruction from monocular videos has been considered an ill-posed problem.", + "bbox": [ + 212, + 207, + 784, + 296 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we incorporate kinematics to fill the gap between sparse monocular videos and real-world physics, potentially enhancing novel view and time synthesis. Our core idea is that motion within the real-world is governed by the principles of kinematics in physics, implying that representations of the world, such as videos, are inherently anchored in kinematics. Building on this foundation, we generalize conventional displacement fields – often called scene-flow fields [18] – to kinematic fields. In prior work, motion in radiance fields has been controlled via the regularization in the frequency domain [28, 33], displacement fields [18], and deformation fields [25].", + "bbox": [ + 212, + 297, + 784, + 431 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Different from these approaches, our kinematic field regulates motion in radiance fields using kinematics; it innately creates continuous trajectories and enables us to impose regularization based on kinematics. We integrate physics-based regularizers to make the kinematic field physically plausible and thus improve the dynamic radiance field accordingly. For instance, the rigidity regularizer based on the first and the second invariant of the strain rate is applied on the kinematic field, suppressing non-rigid motion in the field. We quantitatively and qualitatively validate the proposed regularizers on the NVIDIA Dynamic Scenes (NDVS) dataset, showing the effectiveness of our method over the existing models. Lastly, the kinematic field uncovers the kinematics within a scene without a ground truth; it estimates not only point-wise 3D displacement within a scene, but also underlying velocity, acceleration, and jerk.", + "bbox": [ + 212, + 433, + 785, + 614 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 636, + 387, + 651 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Neural radiance fields. In recent years, neural radiance fields [4,23,32] have emerged as a promising approach for learning an implicit neural representation of a 3D scene. Combined with the differentiable volume rendering technique, it has enabled realistic novel-view synthesis [4,23,32,37].", + "bbox": [ + 212, + 666, + 782, + 727 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In essence, a radiance field $\\mathcal{F}_{\\mathrm{ST}}$ maps a 5D vector $(x,y,z,\\theta ,\\phi)$ to a 4D output vector $(\\mathbf{c}_s,\\sigma_s)$ :", + "bbox": [ + 212, + 728, + 782, + 757 + ], + "page_idx": 1 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {F} _ {\\mathrm {S T}}: (x, y, z, \\theta , \\phi) \\rightarrow (\\mathbf {c} _ {s}, \\sigma_ {s}), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 395, + 758, + 784, + 773 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "where the view-dependent color $\\mathbf{c}_s = (r,g,b)$ and density $\\sigma_s$ are learned in relation to the spatial position $(x,y,z)$ and the viewing direction $(\\theta ,\\phi)$ . With the volumetric representation learned via an iterative optimization, volume rendering techniques [15, 22, 23] are employed to aggregate colors and densities along", + "bbox": [ + 212, + 779, + 785, + 839 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "W. Im et al.", + "bbox": [ + 271, + 114, + 354, + 127 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "a cast ray. Consequently, this leads to realistic novel-view synthesis from the inferred radiance, after being trained on dozens of images.", + "bbox": [ + 212, + 146, + 782, + 176 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Dynamic radiance fields. Radiance fields also facilitate the novel-view synthesis of scenes with moving objects [1,5,16,25,26]. While this is more challenging than the static case, dynamic radiance fields incorporating motion allow for the rendering of a frame at a desired time and viewpoint. As a result, it can be utilized to make visual effects such as bullet time, stabilized, dolly zoomed, and stereo videos. Commonly, the time-dependent radiance fields are referred to as dynamic radiance fields, a term we will adopt throughout this paper.", + "bbox": [ + 212, + 176, + 784, + 282 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In dynamic radiance fields, radiance is dependent on a varying time $t$ . This modifies Eq. 1 by introducing a time-dependent term, resulting in:", + "bbox": [ + 212, + 282, + 785, + 314 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {F} _ {\\mathrm {D Y}}: (x, y, z, t, \\theta , \\phi) \\rightarrow (\\mathbf {c} _ {d}, \\sigma_ {d}), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 387, + 327, + 785, + 343 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where the color and density are conditioned on time $t$ . In this formulation, moving geometries of objects (e.g., translation and deformation) and changing radiance (e.g., moving shadows) can be learned via optimization.", + "bbox": [ + 212, + 354, + 782, + 400 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Learning a dynamic representation presents inherent challenges due to the introduction of temporal dimension. To optimize this representation, a high-quality multi-view dataset can be employed to obtain fine rendering results [1, 5, 16]. Meanwhile, realistically captured imagery of a scene tends to be nonsynchronized and is likely to be obtained with a limited number of cameras and restricted displacements; this often results in a monocular setup [11, 18, 25, 26]. Previous researches have delved into deformable models for monocular videos [20, 25, 26, 31]. In these methods, a deformation field maps a spatio-temporal coordinate to a reference canonical space. While the canonical space aids in maintaining radiance consistency across global time frames, it tends to exhibit reduced effectiveness in generic monocular videos, where new objects might appear within a video [28]. In this study, we embrace the concept of scene flow fields [18], which is constructed without canonical space and deformation networks.", + "bbox": [ + 212, + 401, + 785, + 595 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Regularization with kinematic priors. Monocular videos present challenges in reconstruction due to the constrained viewpoints of moving objects. To address this, integrating auxiliary flow fields has proven to be effective [8, 10, 18, 19]. Existing studies employ either scene flow fields [10, 18, 19] or velocity fields [8] to capture spatial displacements between consecutive frames. Since the scene flows can give spatial correspondences of the dynamic radiance field, it can be used to promote photometric consistency between the renderings of a given ray and its time-deformed counterpart.", + "bbox": [ + 212, + 598, + 785, + 717 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Physical priors incorporated with scene flow fields play a pivotal role in regulating the radiance field. Regularization of motion has been applied to the dynamic field, bridging the gap between the sparse visual data and plausible real-world motion. For instance, a smoothness regularizer ensures that the displacements within close proximity are similar to each other, and kinetic energy regularizer promotes consistency between forward and backward flows, resulting in smoother displacement fields [18]. More deeply integrated with physics, the elastic energy regularizer minimizes the deformation between the canonical space", + "bbox": [ + 212, + 719, + 785, + 839 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "Regularizing Dynamic Radiance Fields with Kinematic Fields", + "bbox": [ + 320, + 114, + 732, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/5d61a5cc06096f3d2093cbc73e139d1cfd6ef4d69d9cb995e6080f498377603b.jpg", + "image_caption": [ + "Fig. 1: Radiance and kinematic fields. This figure summarizes the three fields our method utilizes. The static and dynamic radiance fields are used for rendering, and the kinematic field is used in the training phase, regularizing the dynamic radiance field." + ], + "image_footnote": [], + "bbox": [ + 233, + 148, + 774, + 321 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "and the deformed space [25]. In fluid reconstruction works [2, 7], velocity fields grounded in physics have been shown to be critical in accurate 4D reconstruction. For instance, Navier-Stokes and transport equations can be integrated in reconstructing smoke [7] and fluid surfaces [2] with dynamic radiance fields to reconstruct more physically plausible structures.", + "bbox": [ + 212, + 401, + 782, + 474 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We integrate kinematics to reconstruct dynamic radiance fields from generic monocular videos. We introduce the kinematic field, which captures motion through kinematic quantities such as velocity, acceleration, and jerk. Different from existing velocity field-based methods [7,8,34], our method has higher-order terms (i.e., acceleration and jerk) as latent of spatio-temporal correspondences. Further enhancing our approach, we incorporate physics-driven regularizers based on a transport equation [7] and the strain rate tensor, addressing dynamics of moving particles and rigid motion.", + "bbox": [ + 212, + 477, + 787, + 599 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3 Method", + "text_level": 1, + "bbox": [ + 215, + 619, + 330, + 635 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1 Overview", + "text_level": 1, + "bbox": [ + 215, + 650, + 341, + 664 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our system takes as input a video sequence, denoted as $(I_i, t_i, P_i)_{i=1}^N$ , where $N$ is the number of frames. For each frame, we have an image $I_i \\in \\mathbb{R}^{H \\times W \\times 3}$ , a timestamp $t_i \\in \\mathbb{R}$ , and a precomputed camera pose $P_i \\in SE(3)$ and intrinsic matrix $K_i \\in \\mathbb{R}^{3 \\times 3}$ . From this input, our primary objective is to synthesize an image, denoted as $\\hat{I}_k$ , corresponding to a novel time and viewpoint $(t_k, P_k, K_k)$ . The parameters $t_k$ , $P_k$ , and $K_k$ are not restricted to those found in the original video input, allowing our system to render images from previously unobserved times and viewpoints.", + "bbox": [ + 212, + 672, + 787, + 794 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To achieve this image synthesis for novel times and views, we leverage the radiance-field-based volume rendering [23]. Inspired by existing methods [18,35], we train two distinct radiance fields (Fig. 1): a dynamic radiance field, denoted", + "bbox": [ + 212, + 795, + 787, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "W. Im et al.", + "bbox": [ + 271, + 114, + 354, + 126 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "as $\\mathcal{F}_{\\mathrm{DY}}$ (Eq. 2), and a static radiance field, denoted as $\\mathcal{F}_{\\mathrm{ST}}$ (Eq. 1). To blend the two fields, we utilize a technique from D²-NeRF [35], where rendering weights for dynamic and static samples are derived from the volume densities.", + "bbox": [ + 212, + 146, + 782, + 191 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Addressing the challenge of achieving physically plausible spatio-temporal structures, we introduce the novel kinematic field. The kinematic field accounts for the kinematic quantities (e.g., velocity and acceleration). These quantities can be converted into displacement (Eq. 8), enabling us to use the photometric consistency loss (Eq. 14) for spatio-temporal regularization, as used in previous work [18]. With these kinematic estimations, we apply physics-based regularization techniques. These techniques involve density transport (Eq. 12) and motion rigidity through the strain rate tensor (Eq. 11). Combining the photometric loss (Eq. 14) with these techniques enhances the reconstruction of radiance and geometry, leading to improved novel time and view synthesis performance from underdetermined data such as a monocular video.", + "bbox": [ + 212, + 191, + 787, + 357 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2 Dynamic and Static Radiance Fields", + "text_level": 1, + "bbox": [ + 215, + 378, + 562, + 395 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In monocular dynamic radiance field methods, it is common to decompose the radiance fields into dynamic and static fields [11, 18, 20] to prevent undesirable motion in the static area. We employ tensor-decomposition-based representations [5, 6] for both dynamic and static fields. Specifically, for the dynamic radiance field $\\mathcal{F}_{\\mathrm{DY}}$ , we use HexPlane [5], and for the static radiance field $\\mathcal{F}_{\\mathrm{ST}}$ , we use TensoRF [6]. Further details on the tensor decomposition can be found in the supplementary material.", + "bbox": [ + 212, + 404, + 784, + 508 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Definition. The dynamic and static radiance fields are defined in Eq. 1-2 ( $\\mathcal{F}_{\\mathrm{ST}}$ and $\\mathcal{F}_{\\mathrm{DY}}$ ). Here, the inputs correspond to the position $(x,y,z)$ , the timestamp $t$ , and the viewing direction $(\\theta, \\phi)$ from the camera origin. Each field yields the color $\\mathbf{c}$ and the volume density $\\sigma$ . Note that the static field is independent of time $t$ , ensuring a consistent structure throughout the whole time span.", + "bbox": [ + 212, + 510, + 785, + 585 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Composite rendering. We use a composite rendering method to obtain the final rendering with the two radiance fields. It is to be noted that individual components (i.e., dynamic-only and static-only) can be rendered using the standard volume rendering technique [23]. To combine both fields, we employ an additive composition method in which each field influences the transmittance of a ray [35]. Let's consider a camera ray $\\mathbf{r}(\\tau) = \\mathbf{o} + \\tau \\mathbf{u}$ , where $\\mathbf{o}$ is the origin and $\\mathbf{u}$ is the direction. With a predefined range for the ray distance $[\\tau_n, \\tau_f]$ , the rendering function is formulated as:", + "bbox": [ + 212, + 585, + 787, + 705 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {C} (\\mathbf {r}, t _ {i}) = \\int_ {\\tau_ {n}} ^ {\\tau_ {f}} T (\\tau) \\left(\\sigma_ {s} (\\tau) \\cdot \\mathbf {c} _ {s} (\\tau) + \\sigma_ {d} (\\tau , t _ {i}) \\cdot \\mathbf {c} _ {d} (\\tau , t _ {i})\\right) d \\tau , \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 295, + 715, + 785, + 750 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\text {w h e r e} T (\\tau) = \\exp \\left(- \\int_ {\\tau_ {n}} ^ {\\tau} \\left(\\sigma_ {s} (u) + \\sigma_ {d} (u, t _ {i})\\right) d u\\right). \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 357, + 752, + 785, + 785 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Here, each notation for the sample point is simplified for concise representation (e.g., $\\sigma_s(\\tau) = \\sigma_s(\\mathbf{r}(\\tau))$ ). We visualize the example of the composite rendering and separate rendering in the first and second rows in Fig. 2.", + "bbox": [ + 212, + 794, + 787, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "Regularizing Dynamic Radiance Fields with Kinematic Fields", + "bbox": [ + 320, + 114, + 732, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/00fbf48959c2f2140a078b90b27309e78b499fe24e1e1da8d7a13a664a28ca19.jpg", + "image_caption": [ + "Fig. 2: Visualization of each predicted component. The top row presents synthesized views at different times. In (a) and (b), the dynamic and static components of the scene at time $t_2$ are depicted separately. (b) In the case of an inobservable static area in the entire sequence (e.g., the space behind the jumping people), radiance might not be correct. The displacement field (c) can be computed by Taylor approximation (Eq. 8) with motion fields (d-f): velocity, acceleration, and jerk. Each field is visualized through reprojecting each field to the camera view. The standard HSV visualization [3] is used to colorize arrows." + ], + "image_footnote": [], + "bbox": [ + 290, + 146, + 723, + 356 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Dynamic and static separation. To factorize a scene into static and dynamic parts, we use a metric indicating the likelihood that a point belongs to the dynamic fields [35]. This metric is given by:", + "bbox": [ + 212, + 511, + 787, + 556 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\np _ {d} = \\frac {\\sigma_ {d}}{\\sigma_ {d} + \\sigma_ {s}}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 446, + 569, + 785, + 597 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\sigma_{d}$ and $\\sigma_{s}$ are density values sampled from a spatio-temporal point.", + "bbox": [ + 212, + 607, + 746, + 622 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To facilitate the separation of the radiance fields, we employ the skewed binary entropy loss, as proposed by [35]. The entropy loss is defined as:", + "bbox": [ + 212, + 623, + 785, + 654 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {b} = - \\left(p _ {d} ^ {k} \\cdot \\log \\left(p _ {d} ^ {k}\\right) + \\left(1 - p _ {d} ^ {k}\\right) \\cdot \\log \\left(1 - p _ {d} ^ {k}\\right)\\right), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 339, + 665, + 785, + 685 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where the exponent $k \\geq 1$ is the skew factor, controlling a bias toward static explanation. This entropy-based loss can encourage the radiance fields to have a clearer distinction between the static and dynamic components.", + "bbox": [ + 212, + 696, + 787, + 742 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.3 Kinematic Field", + "text_level": 1, + "bbox": [ + 214, + 766, + 395, + 780 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The kinematic field has a spatio-temporal structure to facilitate spatial and temporal queries. We use the tensor-decomposition-based representation [5] for the kinematic field. As depicted in Fig. 1, the kinematic field serves to regulate", + "bbox": [ + 212, + 794, + 785, + 839 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "W. Im et al.", + "bbox": [ + 271, + 114, + 354, + 126 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/ddb5f4f6daa9e7693eb568e64ac48585a260925bf180acb76283f19bda938471.jpg", + "image_caption": [ + "(a) Without kinematic regularization" + ], + "image_footnote": [], + "bbox": [ + 217, + 142, + 493, + 220 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/5d14acfbe7ef51cf0b59c6b338cc36d26cd2c12e21bcb29cccd19e62ab2a8a6f.jpg", + "image_caption": [ + "(b) With kinematic regularization", + "Fig. 3: Effect of kinematic regularization. We visualize the rendered RGB and motion of each field. Without kinematic regularization, motion fields tend to show granular patterns. Our kinematic fields not only make the field smoother but also satisfy the kinematic property, i.e., $\\mathbf{a} = \\partial \\mathbf{v} / \\partial t + \\mathbf{v} \\cdot \\nabla \\mathbf{v}$ . We abbreviate the advective equation to 'Adv.' in the figure." + ], + "image_footnote": [], + "bbox": [ + 504, + 142, + 782, + 220 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "the radiance field, ensuring that the motion in the radiance field is physically plausible.", + "bbox": [ + 212, + 347, + 782, + 378 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Definition. The kinematic field is defined as:", + "bbox": [ + 215, + 378, + 545, + 392 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {F} _ {\\mathrm {K}}: (x, y, z, t) \\rightarrow (\\mathbf {v}, \\mathbf {a}, \\mathbf {j}, \\dots), \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 388, + 402, + 785, + 420 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where each vector represents velocity, acceleration, and jerk, such that $\\mathbf{v} = (v_x, v_y, v_z)$ , $\\mathbf{a} = (a_x, a_y, a_z)$ , and $\\mathbf{j} = (j_x, j_y, j_z)$ . Note that we can extend the field into higher-order quantities such as snap (4th), crackle (5th), and so on.", + "bbox": [ + 212, + 430, + 782, + 474 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Displacement. Using the provided kinematic quantities, we can approximate a particle's displacement at position $(x,y,z)$ and time $t + \\Delta t$ via the Taylor approximation, assuming that the kinematic quantities are consistent within a local temporal range $\\Delta t$ . The displacement $d(x,y,z,t,\\Delta t)$ can be approximated by:", + "bbox": [ + 212, + 476, + 784, + 547 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\nd (x, y, z, t, \\Delta t) \\approx \\frac {(\\Delta t) \\mathbf {v}}{1 !} + \\frac {(\\Delta t) ^ {2} \\mathbf {a}}{2 !} + \\frac {(\\Delta t) ^ {3} \\mathbf {j}}{3 !} + \\dots . \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 316, + 549, + 784, + 580 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Computing the displacement of a point within a temporal range allows for computing the photometric consistency loss (Eq. 14) and scene flows from the kinematic field.", + "bbox": [ + 212, + 583, + 782, + 628 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.4 Kinematic Regularization", + "text_level": 1, + "bbox": [ + 214, + 650, + 472, + 666 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Given the kinematic field, it is important to make sure the learned motion dynamics are physically plausible. As Fig. 3 indicates, learning kinematic fields without adequate regularization may result in a sub-optimal kinematic field. The subsequent sections elucidate regularization strategies.", + "bbox": [ + 212, + 672, + 782, + 733 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Integrity of kinematic fields. Learning the kinematic field $\\mathcal{F}_{\\mathrm{K}}$ naively does not guarantee the integrity of the kinematic quantities. During training, the quantities are supervised via the photometric consistency loss (Eq. 14), which is applied upon the approximated displacement (Eq. 8). It is important to note that even with supervision by the displacement, the physical relationships among the kinematic quantities may not be correct, since the higher-order terms in the displacement equation (Eq. 8) make the system underdetermined.", + "bbox": [ + 212, + 734, + 784, + 839 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "Regularizing Dynamic Radiance Fields with Kinematic Fields", + "bbox": [ + 320, + 114, + 732, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Given a velocity field, we can derive the acceleration field using the principle of advective acceleration. Given acceleration $\\mathbf{a}$ and velocity $\\mathbf{v}$ at a spatiotemporal coordinate, the partial derivative and advection yield $\\mathbf{a} = \\partial \\mathbf{v} / \\partial t + \\mathbf{v} \\cdot \\nabla \\mathbf{v}$ . Similarly, we can formulate the relationship between $\\mathbf{v}$ , $\\mathbf{a}$ and $\\mathbf{j}$ as: $\\mathbf{j} = \\partial \\mathbf{a} / \\partial t + \\mathbf{v} \\cdot \\nabla \\mathbf{a}$ .", + "bbox": [ + 212, + 146, + 782, + 220 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To ensure kinematic consistency, we introduce the kinematic loss $\\mathcal{L}_{\\mathrm{K}}$ defined as:", + "bbox": [ + 212, + 222, + 782, + 250 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {K}} = \\left\\| \\mathbf {a} - \\frac {\\partial \\mathbf {v}}{\\partial t} - \\mathbf {v} \\cdot \\nabla \\mathbf {v} \\right\\| _ {2} ^ {2} + \\left\\| \\mathbf {j} - \\frac {\\partial \\mathbf {a}}{\\partial t} - \\mathbf {v} \\cdot \\nabla \\mathbf {a} \\right\\| _ {2} ^ {2} + \\dots , \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 308, + 247, + 784, + 284 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $\\nabla$ is the Jacobian operation. To compute the partial derivatives and Jacobian matrices, we use a numerical method as presented in a neural surface reconstruction work [17].", + "bbox": [ + 212, + 287, + 782, + 333 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Physics-informed regularization. In the context of regularizing radiance fields through the use of an auxiliary kinematic field, it is important to acknowledge the highly underdetermined nature of the problem. Essentially, there exists a broad spectrum of possible potential kinematic field solutions that could perfectly fit the observed training views. This complexity necessitates the introduction of a regularization strategy that can guide the motion representation toward more accurately reflecting real-world physics.", + "bbox": [ + 212, + 333, + 784, + 436 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "While not universally ideal, rigidity [25] and incompressibility [2,7] have been identified as effective metrics for encapsulating real-world dynamics; in videos that we normally encounter, objects in a scene have nearly consistent density and are mostly rigid in a short time range.", + "bbox": [ + 212, + 438, + 782, + 500 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Rigidity. One benefit of having a velocity field is that we can measure the rate of deformation by its derivatives. A strain rate tensor $D$ defines a kinematic property of a deformation; it captures the rate at which a material undergoes deformation, and it is crucial for ensuring that objects maintain their structural integrity during motion.", + "bbox": [ + 212, + 500, + 782, + 573 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we propose using the invariants of strain rate tensor [12, 29] as a measure of rigidity. The velocity gradient $\\nabla \\mathbf{v}$ sampled from the kinematic field, leads to the strain rate tensor $D$ by:", + "bbox": [ + 212, + 574, + 782, + 619 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\nD = \\frac {1}{2} (\\nabla \\mathbf {v} + (\\nabla \\mathbf {v}) ^ {T}). \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 419, + 628, + 784, + 656 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The rigidity loss function $\\mathcal{L}_R$ is formulated based on the first and second invariants of the strain rate tensor:", + "bbox": [ + 212, + 662, + 782, + 693 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {R} = \\underbrace {\\lambda_ {\\mathrm {d i v}} (\\nabla \\cdot \\mathbf {v}) ^ {2}} _ {\\text {1 s t i n v a r i a n t}} + \\underbrace {\\left(\\frac {1}{2} (\\operatorname {t r} (D) ^ {2} - \\operatorname {t r} (D \\cdot D)))\\right) ^ {2}} _ {\\text {2 n d i n v a r i a n t}}, \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 336, + 700, + 784, + 746 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $\\operatorname{tr}(\\cdot)$ denotes the trace operator and $\\lambda_{\\mathrm{div}}$ controls the balance between the two terms. Note that the first invariant of strain rate measures dilational change (i.e., volumetric change) with respect to time and the second invariant of strain rate measures distortional change.", + "bbox": [ + 212, + 755, + 782, + 816 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "W. Im et al.", + "bbox": [ + 271, + 114, + 354, + 127 + ], + "page_idx": 7 + }, + { + "type": "page_footnote", + "text": "$\\overline{4f'(x)} = \\frac{f(x + \\epsilon) - f(x - \\epsilon)}{2\\epsilon}$", + "bbox": [ + 217, + 821, + 377, + 842 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/1551b001fadd2ddc6de0d6978ba34f979017653f9dfb6947b58194b66e89d8e6.jpg", + "image_caption": [ + "Fig.4: Visualization of density variation by a spatial coordinate $x$ . The plot displays the gradient $-\\partial \\mathcal{L}_{\\mathrm{T}} / \\partial v_{\\sigma_d}$ with arrows. With the velocity field directed to the right (i.e., $\\mathbf{v}_x = 0.3$ ), we can compute the gradient of the transport regularization $\\mathcal{L}_{\\mathrm{T}}$ w.r.t. the rate of density change $v_{\\sigma_d} = \\partial \\sigma_d / \\partial t$ . Minimizing $\\mathcal{L}_{\\mathrm{T}}$ allows us to render the density at $t + \\Delta t$ aligned with the flow field." + ], + "image_footnote": [], + "bbox": [ + 217, + 148, + 460, + 318 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/cb83f9a1a3f40142d3ed601b8bede8889a07f14c8d49fafd0a30f20542b6983f.jpg", + "image_caption": [ + "Fig. 5: Photometric consistency loss. During optimization, we deform each ray from a reference time and pose $(t,P_{t})$ to different timestamps $\\gamma$ and $i$ . Here, we sample $i$ from the neighboring frame times, and we consequently sample an intermediate timestamp $\\gamma \\sim \\mathcal{U}(t,i)$ . Given each timestamp, we can deform a ray using Eq. 8. The photometric loss $\\mathcal{L}_{\\mathrm{photo}}$ is computed based on the color consistency between the deformed ray and the original color, enhancing temporal consistency." + ], + "image_footnote": [], + "bbox": [ + 485, + 148, + 782, + 321 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Transport. A transport equation has been adopted in physics-informed neural networks [7]. By simplifying the Navier-Stokes equation, a transport loss can be defined as: $\\frac{\\partial\\sigma}{\\partial t} +\\mathbf{v}\\cdot \\nabla \\sigma = 0$ . We leverage the physics-informed regularizers based on the transport equation and the divergence of the velocity:", + "bbox": [ + 212, + 510, + 787, + 571 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {T}} = \\left(\\frac {\\partial \\sigma_ {d}}{\\partial t} + \\mathbf {v} \\cdot \\nabla \\sigma_ {d}\\right) ^ {2}, \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 406, + 575, + 785, + 612 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where $\\sigma_{d}$ is the volume density used for volume rendering. This aligns the velocity field with the density field (i.e., Fig. 4), preventing a sudden disappearance or appearance of objects in the dynamic radiance field.", + "bbox": [ + 212, + 618, + 787, + 666 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "3.5 Learning Objectives", + "text_level": 1, + "bbox": [ + 214, + 683, + 428, + 699 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Cycle consistency. The cycle consistency is essential to ensure that the trajectories derived from the kinematic field do not introduce misalignments [18]. We use cycle constraints among four timestamps $t$ , $i$ , $j$ , and $\\gamma$ . Given a reference time $t$ , we randomly sample time $i$ from the neighboring frames within a threshold. We define $j = t + 2(i - t)$ , making it twice the distance from $t$ as $i$ is from $t$ . $\\gamma \\sim \\mathcal{U}(t,i)$ is an intermediate time, uniformly sampled between $t$ and $i$ . With the time variables, we formulate our loss:", + "bbox": [ + 212, + 705, + 787, + 811 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {C}} = \\rho (\\mathbf {d} _ {t \\rightarrow i} + \\mathbf {d} _ {i \\rightarrow t}) + \\rho (\\mathbf {d} _ {t \\rightarrow i \\rightarrow j} - \\mathbf {d} _ {t \\rightarrow j}) + \\rho (\\mathbf {d} _ {t \\rightarrow \\gamma \\rightarrow i} - \\mathbf {d} _ {t \\rightarrow i}), \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 254, + 819, + 785, + 837 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "Regularizing Dynamic Radiance Fields with Kinematic Fields", + "bbox": [ + 321, + 114, + 732, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where $\\rho$ denotes the generalized chabonnier loss [30]. Here, we abbreviate the displacement $d(\\cdot)$ (Eq. 8) for conciseness: $\\mathbf{d}_{t\\to i}$ is shorthand for $d(\\mathbf{x},t,i - t)$ , and $\\mathbf{d}_{t\\to i\\to j}$ simplifies $d(\\mathbf{x} + \\mathbf{d}_{t\\to i},i,j - i)$ .", + "bbox": [ + 212, + 146, + 782, + 191 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Photometric consistency loss. To enhance temporal coherency, we deform rays into nearby frame times and enhance the consistency among them (Fig. 5). Specifically, given a reference ray $\\mathbf{r}_t$ at time $t$ , a pixel color $\\hat{C}(\\mathbf{r}_t, t)$ can be obtained using the volume rendering (Eq. 4). In addition to the reference ray, we apply ray deformation based on the predicted kinematic quantities. As in cycle consistency (Eq. 13), we sample time $i$ from neighboring frames, and time $\\gamma \\sim \\mathcal{U}(t, i)$ between times $t$ and $i$ . Different from existing approaches [10, 18], we deform rays with kinematic quantities using Taylor approximation. The deforming process results in two deformed rays $\\mathbf{r}_{t \\to i}$ and $\\mathbf{r}_{t \\to \\gamma}$ , and two rendered colors $\\hat{C}(\\mathbf{r}_{t \\to i}, i)$ and $\\hat{C}(\\mathbf{r}_{t \\to \\gamma}, \\gamma)$ , correspondingly. Based on the deformed rays, we apply the photometric consistency loss, comparing the rendered color with the ground-truth color. The loss function is formulated as:", + "bbox": [ + 212, + 191, + 784, + 375 + ], + "page_idx": 9 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {\\text {p h o t o}} = \\| \\hat {C} (\\mathbf {r} _ {t}, t) - C (\\mathbf {r} _ {t}, t) \\| _ {2} ^ {2} \\tag {14} \\\\ + \\lambda \\alpha (\\mathbf {r} _ {t}) \\rho (\\hat {C} (\\mathbf {r} _ {t \\rightarrow i}, i) - C (\\mathbf {r} _ {t}, t)) + \\lambda \\alpha (\\mathbf {r} _ {t}) \\rho (\\hat {C} (\\mathbf {r} _ {t \\rightarrow \\gamma}, \\gamma) - C (\\mathbf {r} _ {t}, t)), \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 230, + 385, + 784, + 422 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "where $\\lambda$ is a loss weight for the temporal consistency, and $\\alpha (\\mathbf{r}_t)$ is the sum of rendering weights of the dynamic densities, representing the dynamic likelihood of the ray $\\mathbf{r}_t$ .", + "bbox": [ + 212, + 433, + 782, + 477 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Final objective. Our final objective function, denoted as $\\mathcal{L}$ , is the weighted sum of the various loss terms discussed previously. To keep things simple with the variety of loss functions, we symbolize the losses to $\\mathcal{L}_{\\mathrm{photo}}$ , $\\mathcal{L}_{\\mathrm{kinematic}}$ , and $\\mathcal{L}_{\\mathrm{reg}}$ . This results in the final loss:", + "bbox": [ + 212, + 479, + 784, + 539 + ], + "page_idx": 9 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathbb {E} _ {\\mathbf {r} _ {t}} \\left[ \\mathcal {L} _ {\\text {p h o t o}} + \\mathcal {L} _ {\\text {k i n e m a t i c}} + \\mathcal {L} _ {\\text {r e g}} \\right]. \\tag {15}\n$$\n", + "text_format": "latex", + "bbox": [ + 374, + 550, + 782, + 566 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Here, $\\mathcal{L}_{\\mathrm{kinematic}}$ is the weighted sum of the kinematic integrity $\\mathcal{L}_{\\mathrm{K}}$ , transport $\\mathcal{L}_{\\mathrm{T}}$ , and rigidity $\\mathcal{L}_{\\mathrm{R}}$ . Note that $\\mathbb{E}_{\\mathbf{r}_t}[\\mathcal{L}_{\\mathrm{kinematic}}]$ is computed with randomly sampled points along the ray $\\mathbf{r}_t$ from each minibatch. The additional regularization $\\mathcal{L}_{\\mathrm{reg}}$ includes the skewed entropy $\\mathcal{L}_b$ , the cycle consistency $\\mathcal{L}_{\\mathrm{C}}$ , total-variation loss [5,6], distortion loss [4], monocular depth supervision [18], normal regularization [14,32] and optical flow supervision [18].", + "bbox": [ + 214, + 575, + 782, + 667 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4 Experiment", + "text_level": 1, + "bbox": [ + 214, + 688, + 366, + 705 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Dataset. We use the NVIDIA dynamic view synthesis dataset (NDVS) [36] to evaluate the effectiveness of our model. For the monocular setup, we adopt the preprocessing steps proposed by NSFF [18] and DynamicNeRF [10]. This dataset has 8 scenes and is captured with 12 synchronized cameras. When used for a monocular video reconstruction task, the dataset shows relatively low effective multi-view factors (EMFs) [11], compared to other monocular datasets; this means that the dataset is more challenging to reconstruct due to the less camera motion and faster-moving objects.", + "bbox": [ + 212, + 718, + 782, + 839 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "W. Im et al.", + "bbox": [ + 271, + 114, + 354, + 126 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "There are two widely used monocular NDVS protocols: 12-frames [10] and 24-frames [18]. Most papers opt for either protocol; however, we include both protocols for rigorous evaluation. Meanwhile, 24-frames-sparse is proposed by us, as neither existing protocol provides novel time evaluation; we leave out odd-numbered frames at unseen times for testing. For the 'Dynamic Only' marks in tables, we use the dynamic mask to evaluate only in terms of dynamic pixels; please refer to [10,18] for how to generate dynamic masks. We detail the protocols below.", + "bbox": [ + 212, + 146, + 787, + 265 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- 24-frames: this sequence is configured with 24 training frames, and 264 test frames from unseen multi-views. For a fair comparison, we utilize the dynamic mask, optical flow, and monocular depth, extracted by the same model with the prior work [18].", + "- 24-frames-sparse: we reduce the 24-frames sequence into 12 frames by sampling the even-numbered frames for training. This sparse sampling allows for testing novel view synthesis at both seen times and novel times.", + "- 12-frames: this sequence has 12 training and 12 testing samples. We use the preprocessed dynamic mask, optical flow, and monocular depth downloadable from the code repository of DynamicNeRF [10]." + ], + "bbox": [ + 225, + 279, + 782, + 429 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4.1 Implementation Details", + "text_level": 1, + "bbox": [ + 215, + 452, + 455, + 467 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Fields configuration. Our radiance and kinematic fields are configured with HexPlane [5] and TensoRF [6]. After features are sampled from the point queries, small MLP decoders process the features to yield colors and kinematic outputs. For density outputs, we use a linear layer following the practices in [5,6]. Further details on field configurations are available in the supplementary material.", + "bbox": [ + 212, + 477, + 782, + 549 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Training. We use 70,000 steps to optimize our models, which takes about 7 hours with one NVIDIA V100 or A100 GPU. We initialize the static field for 5,000 steps, taking approximately 5 minutes. After finishing the initialization of the static field, we jointly train the fields: dynamic, static, and kinematic fields. During the joint training, we lower the learning rate of the static field by the factor of 0.1 at predefined iterations to prevent dynamic objects from being portrayed as static objects. Additional training details are provided in the supplementary material.", + "bbox": [ + 212, + 553, + 782, + 669 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Evaluation metrics. We report peak signal-to-noise ratio (PSNR), structural similarity (SSIM), and perceptual similarity (LPIPS). These three metrics have been used as standard in prior studies to evaluate the quality of renderings.", + "bbox": [ + 212, + 672, + 782, + 715 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Base architecture (HexPlane). Since our model is based on tensor-decomposed radiance fields, we compare our models with HexPlane [5]. The baseline model is the dynamic radiance field $\\mathcal{F}_{\\mathrm{DY}}$ , without the static radiance field $\\mathcal{F}_{\\mathrm{ST}}$ and the kinematic field $\\mathcal{F}_{\\mathrm{K}}$ . For a fair comparison, we use monocular depth supervision both to the baseline model and our models, although the original HexPlane model does not support the supervision from monocular depth. Since the base HexPlane model does not have an auxiliary flow field, we do not use the optical flow supervision and the photometric consistency loss.", + "bbox": [ + 212, + 719, + 792, + 837 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "Regularizing Dynamic Radiance Fields with Kinematic Fields", + "bbox": [ + 320, + 114, + 730, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 116, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/3a9c0b5e98e8687225d9e745d60a4d13ebf39ec4c9c3cbd75c7550a0e2412b3f.jpg", + "table_caption": [ + "Table 1: Ablation study on Balloon1 scene of the NDVS dataset (24-frames-sparse). Please refer to the supplementary material for ablation results on other NDVS scenes." + ], + "table_footnote": [], + "table_body": "
MethodsNovel TimesSeen Times
FullDynamic OnlyFullDynamic Only
PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓
HexPlane [5]19.840.6680.14518.120.4190.25118.930.6490.13717.740.4290.202
Ours25.670.8650.06421.160.6250.17025.800.8670.05021.350.6320.121
w/o Higher order motion24.360.8420.07719.930.5560.20024.630.8460.06120.220.5720.145
w/o Lkinematic24.960.8530.06820.150.5680.18525.300.8580.05220.650.5940.129
w/o Kinematic integrity25.460.8620.06620.780.6070.17925.570.8650.05121.020.6200.126
w/o Rigidity25.110.8560.06620.400.5820.18125.430.8600.05120.820.6030.123
", + "bbox": [ + 218, + 184, + 782, + 271 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/16ac34041a8548cbf376d994f7713df7efa133d9447f29fe78e0eca2908a444d.jpg", + "image_caption": [ + "Fig. 6: Comparative visualization of kinematic fields. The first row shows RGB values rendered from our dynamic radiance field, showing PSNR values from a ground truth RGB image. From the second to the last row, we show flow (i.e., displacement), velocity, and acceleration." + ], + "image_footnote": [], + "bbox": [ + 228, + 284, + 774, + 540 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "4.2 Evaluation of Kinematic Fields", + "text_level": 1, + "bbox": [ + 215, + 641, + 517, + 655 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Ablation of kinematic losses. To show the effectiveness of our approach, we conduct a comprehensive ablation study on the Balloon1 scene of the NDVS dataset, focusing on kinematic loss components. In Table 1, we show ablative results from seen and unseen frame times. Our findings indicate that the model achieves optimal performance when all components are integrated.", + "bbox": [ + 212, + 672, + 787, + 748 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Fig. 6 delves into the kinematic field analysis. Omitting higher order motion restricts the model to utilizing only the velocity field, which proves insufficient for capturing non-linear motion. This limitation hinders the model's ability to minimize kinematic losses, leading to a suboptimal flow representation. In the absence of $\\mathcal{L}_{\\mathrm{kinematic}}$ , the radiance field suffers from reduced PSNR scores, which can be attributed to the physically incorrect motion and geometry. Compared", + "bbox": [ + 212, + 750, + 787, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "W. Im et al.", + "bbox": [ + 271, + 114, + 354, + 127 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/a9c9a9ded169c82671c72721386a643c1ca4977fcaf5df7036ef34668ee0c877.jpg", + "image_caption": [ + "Fig. 7: Motion Order. Reconstruction Accuracy for the Balloon1 Scene (24-framesparse): The $x$ -axis of each plot categorizes maximum motion order as follows: velocity (1), acceleration (2), jerk (3), snap (4), and crackle (5). The metric utilized for each plot is specified at the top." + ], + "image_footnote": [], + "bbox": [ + 222, + 148, + 356, + 244 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/86c577485099033c2c2c217354d6aafa70ae76281e8ab4026870c8709c3324f2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 362, + 148, + 496, + 244 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/267eb84c3692aa30abeb538a785ae608f57c0287c9af9fe2a578ec64ace0b96a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 504, + 148, + 635, + 244 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/11267c13c409d4a187ce802bb658ffac45c97ceb3f45404e71b75d9049f67ff8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 647, + 148, + 777, + 244 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/6cf082eaf1def2a19ca2a7f4393819329df9c0960da3a1b3b8f6275536c6f334.jpg", + "table_caption": [ + "Table 2: Quantitative comparison on NDVS. This table reports average results on all scenes. Please refer to the supplementary material for per-scene results." + ], + "table_footnote": [], + "table_body": "
(a) 12-frames
MethodsFull
PSNR↑LPIPS↓
NeRF [23]*21.990.250
NSFF [18]*24.330.199
DynamicNeRF [10]*26.100.082
HyperNeRF [26]*17.600.367
RoDynRF [20]*25.890.065
Ours26.330.071
*Results reported in [20]
", + "bbox": [ + 220, + 375, + 426, + 494 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/1c4093f1fa36d0acc85aef7d8772e2a7fd8dd6611340e6b1d811bb6c1c7ee30d.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
(b) 24-frames
MethodsFullDynamic Only
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
NeRF [23]*24.900.8930.09816.980.5320.314
Luo et al. [21]*21.370.7460.14116.970.5300.207
Yoon et al. [36]*21.780.7610.12717.340.5470.200
NSFF [18]*28.190.9280.04521.910.7580.097
NSFF [18]28.040.9270.04521.340.7400.111
Ours28.780.9360.04222.130.7740.094
*Results reported in [18]
", + "bbox": [ + 429, + 375, + 782, + 493 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "to the full model, the model without the rigidity loss results in noisy, non-rigid motion, resulting in incorrect motion and the worse reconstruction accuracy.", + "bbox": [ + 212, + 522, + 784, + 551 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Motion order. Our study represents motion through various kinematic quantities. Adjusting the kinematic field to include different orders of motion, we can model our system using just velocity [7, 8], or extend to higher orders – fourth (snap) and fifth (crackle) – by extending the Taylor series (Eq. 8) and the kinematic integrity (Eq. 9) equations. Fig. 7 illustrates the influence of motion order on reconstruction accuracy for the NDVS dataset. As motion order increases, the reconstruction accuracy first improves, peaking at jerk, but further orders degrades the accuracy. We conjecture that the augmentation of model complexity with higher orders of kinematic quantities may render the model more susceptible to overfitting.", + "bbox": [ + 212, + 551, + 787, + 703 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "4.3 Comparison with SOTA models", + "text_level": 1, + "bbox": [ + 215, + 724, + 522, + 739 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Quantitative results. NSFF [18] is a dynamic neural radiance field model, which utilizes scene-flow fields for the regularization of a dynamic radiance field. Since NSFF is a neural model, the network training takes about 2 days on two V100 GPUs (96 GPU hours), using the official code by the authors. Note that optimizing a scene using our method takes about 7 hours per scene, which makes ours 12 times faster than NSFF in terms of GPU hours. We compare our results", + "bbox": [ + 212, + 750, + 787, + 839 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "Regularizing Dynamic Radiance Fields with Kinematic Fields", + "bbox": [ + 321, + 114, + 732, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 784, + 126 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/a0b747be6e9416d394293e76f6342c55d31ec714da4f924512e4d287da88f976.jpg", + "image_caption": [ + "Fig. 8: Qualitative results on NDVS (12-frames). The second and fourth rows illustrate the overlay of the synthesized and ground truth views of the scene." + ], + "image_footnote": [], + "bbox": [ + 217, + 142, + 782, + 328 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "with NSFF [18] in Table 2b. In the table, our method shows better accuracy in terms of all metrics. In addition, we report results on NDVS (12-frames) in Table 2a.", + "bbox": [ + 212, + 395, + 784, + 439 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Qualitative results. Fig. 8 illustrates qualitative results, including the synthesized images, ground truth, and their overlaid images. As shown in the first and third rows, our method improves the visual quality of the rendered views. Additionally, the overlaid images demonstrate the geometrical accuracy of our model, showcasing its better alignment with the ground truth. This enhanced performance is attributed to the kinematics and physics incorporated in our methodology, which is fundamental to the improvement.", + "bbox": [ + 212, + 440, + 784, + 546 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 566, + 357, + 583 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In this paper, we have presented a novel method to learn dynamic radiance fields with kinematic fields. We propose the kinematic field, where kinematic quantities are learned by joint learning with the radiance fields. Using the kinematic quantities from the kinematic fields, we propose physics-based loss functions that underpin our method. Lastly, we quantitatively and qualitatively show the effectiveness of our method on the monocular dataset. We believe that this work not only introduces the potential of incorporating kinematics into the learning of radiance fields but also sets a foundation for future direction in physics-based dynamic scene understanding.", + "bbox": [ + 212, + 595, + 784, + 732 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgment. This project was partly supported by the Naver Cloud Corporation. Additionally, this work received support from the Institute of Information & communications Technology Planning & Evaluation (IITP) grant (RS-2023-00237965) and the National Research Foundation of Korea (NRF) grant (No. RS-2023-00208506(2024)), funded by the Korea government (MSIT). Sung-Eui Yoon is a corresponding author.", + "bbox": [ + 212, + 748, + 785, + 839 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "W. Im et al.", + "bbox": [ + 271, + 114, + 354, + 126 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 143, + 321, + 159 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Attal, B., Huang, J.B., Richardt, C., Zollhoefer, M., Kopf, J., O'Toole, M., Kim, C.: Hyperreel: High-fidelity 6-dof video with ray-conditioned sampling. arXiv preprint arXiv:2301.02238 (2023)", + "2. Baieri, D., Esposito, S., Maggioli, F., Rodola, E.: Fluid dynamics network: Topology-agnostic 4d reconstruction via fluid dynamics priors. arXiv preprint arXiv:2303.09871 (2023)", + "3. Baker, S., Scharstein, D., Lewis, J., Roth, S., Black, M.J., Szeliski, R.: A database and evaluation methodology for optical flow. International journal of computer vision 92, 1-31 (2011)", + "4. Barron, J.T., Mildenhall, B., Tancik, M., Hedman, P., Martin-Brualla, R., Srinivasan, P.P.: Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5855-5864 (2021)", + "5. Cao, A., Johnson, J.: Hexplane: A fast representation for dynamic scenes. arXiv preprint arXiv:2301.09632 (2023)", + "6. Chen, A., Xu, Z., Geiger, A., Yu, J., Su, H.: Tensorf: Tensorial radiance fields. In: Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXXII. pp. 333-350. Springer (2022)", + "7. Chu, M., Liu, L., Zheng, Q., Franz, E., Seidel, H.P., Theobalt, C., Zayer, R.: Physics informed neural fields for smoke reconstruction with sparse data. ACM Transactions on Graphics (TOG) 41(4), 1-14 (2022)", + "8. Du, Y., Zhang, Y., Yu, H.X., Tenenbaum, J.B., Wu, J.: Neural radiance flow for 4d view synthesis and video processing. In: 2021 IEEE/CVF International Conference on Computer Vision (ICCV). pp. 14304-14314. IEEE Computer Society (2021)", + "9. Fang, J., Yi, T., Wang, X., Xie, L., Zhang, X., Liu, W., Niešner, M., Tian, Q.: Fast dynamic radiance fields with time-aware neural voxels. In: SIGGRAPH Asia 2022 Conference Papers. pp. 1-9 (2022)", + "10. Gao, C., Saraf, A., Kopf, J., Huang, J.B.: Dynamic view synthesis from dynamic monocular video. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 5712-5721 (October 2021)", + "11. Gao, H., Li, R., Tulsiani, S., Russell, B., Kanazawa, A.: Monocular dynamic view synthesis: A reality check. In: Oh, A.H., Agarwal, A., Belgrave, D., Cho, K. (eds.) Advances in Neural Information Processing Systems (2022), https://openreview.net/forum?id=pCrB8orUkSq", + "12. Hart-Smith, L.: Application of the strain invariant failure theory (sift) to metals and fiber-polymer composites. Philosophical Magazine 90(31-32), 4263-4331 (2010)", + "13. Hess, R., Baker, C., Zihl, J.: The\" motion-blind\" patient: low-level spatial and temporal filters. Journal of Neuroscience 9(5), 1628-1640 (1989)", + "14. Jin, H., Liu, I., Xu, P., Zhang, X., Han, S., Bi, S., Zhou, X., Xu, Z., Su, H.: Tensoroir: Tensorial inverse rendering. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2023)", + "15. Kajiya, J.T., Von Herzen, B.P.: Ray tracing volume densities. In: Proceedings of the 11th Annual Conference on Computer Graphics and Interactive Techniques. p. 165-174 (1984)", + "16. Li, T., Slavcheva, M., Zollhoefer, M., Green, S., Lassner, C., Kim, C., Schmidt, T., Lovegrove, S., Goesele, M., Newcombe, R., et al.: Neural 3d video synthesis from multi-view video. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5521-5531 (2022)" + ], + "bbox": [ + 218, + 172, + 784, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "Regularizing Dynamic Radiance Fields with Kinematic Fields", + "bbox": [ + 321, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "17. Li, Z., Müller, T., Evans, A., Taylor, R.H., Unberath, M., Liu, M.Y., Lin, C.H.: Neuralangelo: High-fidelity neural surface reconstruction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8456-8465 (2023)", + "18. Li, Z., Niklaus, S., Snavely, N., Wang, O.: Neural scene flow fields for space-time view synthesis of dynamic scenes. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6498-6508 (2021)", + "19. Li, Z., Wang, Q., Cole, F., Tucker, R., Snavely, N.: Dynibar: Neural dynamic image-based rendering. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4273-4284 (2023)", + "20. Liu, Y.L., Gao, C., Meuleman, A., Tseng, H.Y., Saraf, A., Kim, C., Chuang, Y.Y., Kopf, J., Huang, J.B.: Robust dynamic radiance fields. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 13-23 (June 2023)", + "21. Luo, X., Huang, J.B., Szeliski, R., Matzen, K., Kopf, J.: Consistent video depth estimation. ACM Transactions on Graphics (ToG) 39(4), 71-1 (2020)", + "22. Max, N., Chen, M.: Local and global illumination in the volume rendering integral. Tech. rep., Lawrence Livermore National Lab.(LLNL), Livermore, CA (United States) (2005)", + "23. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. In: ECCV (2020)", + "24. Orts-Escalano, S., Rhemann, C., Fanello, S., Chang, W., Kowdle, A., Degtyarev, Y., Kim, D., Davidson, P.L., Khamis, S., Dou, M., et al.: Holoportation: Virtual 3d teleportation in real-time. In: Proceedings of the 29th annual symposium on user interface software and technology. pp. 741-754 (2016)", + "25. Park, K., Sinha, U., Barron, J.T., Bouaziz, S., Goldman, D.B., Seitz, S.M., Martin-Brualla, R.: Nerfies: Deformable neural radiance fields. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5865-5874 (2021)", + "26. Park, K., Sinha, U., Hedman, P., Barron, J.T., Bouaziz, S., Goldman, D.B., Martin-Brualla, R., Seitz, S.M.: Hypernerf: A higher-dimensional representation for topologically varying neural radiance fields. arXiv preprint arXiv:2106.13228 (2021)", + "27. Pumarola, A., Corona, E., Pons-Moll, G., Moreno-Noguer, F.: D-nerf: Neural radiance fields for dynamic scenes. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10318-10327 (2021)", + "28. Ramasinghe, S., Shevchenko, V., Avraham, G., Hengel, A.V.D.: Bali-rf: Bandlimited radiance fields for dynamic scene modeling. arXiv preprint arXiv:2302.13543 (2023)", + "29. Soria, J., Sondergaard, R., Cantwell, B., Chong, M., Perry, A.: A study of the fine-scale motions of incompressible time-developing mixing layers. Physics of Fluids 6(2), 871-884 (1994)", + "30. Sun, D., Roth, S., Black, M.J.: A quantitative analysis of current practices in optical flow estimation and the principles behind them. International Journal of Computer Vision 106, 115-137 (2014)", + "31. Tretschk, E., Tewari, A., Golyanik, V., Zollhöfer, M., Lassner, C., Theobalt, C.: Non-rigid neural radiance fields: Reconstruction and novel view synthesis of a dynamic scene from monocular video. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 12959-12970 (2021)", + "32. Verbin, D., Hedman, P., Mildenhall, B., Zickler, T., Barron, J.T., Srinivasan, P.P.: Ref-nerf: Structured view-dependent appearance for neural radiance fields. In: 2022" + ], + "bbox": [ + 215, + 146, + 785, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "W. Im et al.", + "bbox": [ + 271, + 114, + 354, + 126 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 5481-5490. IEEE (2022)", + "33. Wang, C., Eckart, B., Lucey, S., Gallo, O.: Neural trajectory fields for dynamic novel view synthesis. arXiv preprint arXiv:2105.05994 (2021)", + "34. Wang, C., MacDonald, L.E., Jeni, L.A., Lucey, S.: Flow supervision for deformable nerf. arXiv preprint arXiv:2303.16333 (2023)", + "35. Wu, T., Zhong, F., Tagliasacchi, A., Cole, F., Oztireli, C.: D\\~ 2nerf: Self-supervised decoupling of dynamic and static objects from a monocular video. Advances in Neural Information Processing Systems 35, 32653-32666 (2022)", + "36. Yoon, J.S., Kim, K., Gallo, O., Park, H.S., Kautz, J.: Novel view synthesis of dynamic scenes with globally coherent depths from a monocular camera. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5336-5345 (2020)", + "37. Zhang, K., Riegler, G., Snavely, N., Koltun, V.: Nerf++: Analyzing and improving neural radiance fields. arXiv preprint arXiv:2010.07492 (2020)" + ], + "bbox": [ + 215, + 146, + 784, + 354 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "Regularizing Dynamic Radiance Fields with Kinematic Fields", + "bbox": [ + 321, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 16 + } +] \ No newline at end of file diff --git a/2024/Regularizing Dynamic Radiance Fields with Kinematic Fields/79573c22-a91b-4733-9726-bd12faf03469_model.json b/2024/Regularizing Dynamic Radiance Fields with Kinematic Fields/79573c22-a91b-4733-9726-bd12faf03469_model.json new file mode 100644 index 0000000000000000000000000000000000000000..aa937a91576a68f2086feee78569c7e45420daf6 --- /dev/null +++ b/2024/Regularizing Dynamic Radiance Fields with Kinematic Fields/79573c22-a91b-4733-9726-bd12faf03469_model.json @@ -0,0 +1,2379 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.251, + 0.141, + 0.753, + 0.185 + ], + "angle": 0, + "content": "Regularizing Dynamic Radiance Fields with Kinematic Fields" + }, + { + "type": "text", + "bbox": [ + 0.241, + 0.212, + 0.761, + 0.245 + ], + "angle": 0, + "content": "Woobin \\(\\mathrm{Im}^{\\star 1,2}\\), Geonho Cha\\(^3\\), Sebin Lee\\(^1\\), Jumin Lee\\(^1\\), Juhyeong Seon\\(^1\\), Dongyoon Wee\\(^3\\), and Sung-Eui Yoon\\(^1\\)" + }, + { + "type": "text", + "bbox": [ + 0.308, + 0.254, + 0.696, + 0.282 + ], + "angle": 0, + "content": "\\(^{1}\\) KAIST {iwbn,seb.lee,jmlee,munuwazzi,sungeui}@kaist.ac.kr" + }, + { + "type": "text", + "bbox": [ + 0.325, + 0.283, + 0.678, + 0.297 + ], + "angle": 0, + "content": "\\(^{2}\\) Samsung Advanced Institute of Technology (SAIT)" + }, + { + "type": "text", + "bbox": [ + 0.444, + 0.297, + 0.559, + 0.31 + ], + "angle": 0, + "content": "3 Naver Cloud" + }, + { + "type": "text", + "bbox": [ + 0.351, + 0.312, + 0.653, + 0.325 + ], + "angle": 0, + "content": "{geonho.cha,dongyoon.wei}@navercorp.com" + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.36, + 0.744, + 0.568 + ], + "angle": 0, + "content": "Abstract. This paper presents a novel approach for reconstructing dynamic radiance fields from monocular videos. We integrate kinematics with dynamic radiance fields, bridging the gap between the sparse nature of monocular videos and the real-world physics. Our method introduces the kinematic field, capturing motion through kinematic quantities: velocity, acceleration, and jerk. The kinematic field is jointly learned with the dynamic radiance field by minimizing the photometric loss without motion ground truth. We further augment our method with physics-driven regularizers grounded in kinematics. We propose physics-driven regularizers that ensure the physical validity of predicted kinematic quantities, including advective acceleration and jerk. Additionally, we control the motion trajectory based on rigidity equations formed with the predicted kinematic quantities. In experiments, our method outperforms the state-of-the-arts by capturing physical motion patterns within challenging real-world monocular videos." + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.581, + 0.707, + 0.596 + ], + "angle": 0, + "content": "Keywords: Dynamic radiance fields \\(\\cdot\\) Kinematic fields \\(\\cdot\\) 3D video" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.62, + 0.377, + 0.636 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.651, + 0.788, + 0.756 + ], + "angle": 0, + "content": "Recognizing motion is crucial to how we visually perceive the world [13], and video is the medium that carries the motion to viewers in a remote place or at a different time. Through the lens of video, we watch and enjoy movies, user-generated contents, and remote video chats. While videos depict motion in a 2D format, transforming them into three-dimensional experiences can elevate our perception of the content to co-presence via applications such as teleportation [24]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.757, + 0.788, + 0.817 + ], + "angle": 0, + "content": "Dynamic radiance fields [5,9,11,16,19,27] have been studied as a promising representation for reconstructing 3D dynamic scenes; it enables realistic novel view and time synthesis, which can upgrade 2D videos into 3D videos. Since multi-view cameras can give more information about the geometry in a scene," + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.825, + 0.634, + 0.84 + ], + "angle": 0, + "content": "* This work was done during an internship at Naver Cloud." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.356, + 0.128 + ], + "angle": 0, + "content": "W. Im et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.207 + ], + "angle": 0, + "content": "existing works have utilized camera rigs with dozens of cameras to capture multi-view video datasets [1, 5, 16]. Although high-quality multi-view videos result in undeniably better quality than monocular videos do, the multi-view capturing devices and environments are prohibitive in most cases." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.208, + 0.785, + 0.297 + ], + "angle": 0, + "content": "Thus, there have been efforts to reconstruct dynamic radiance fields from monocular videos [10, 18, 20]. Compared to multi-view videos, monocular videos provide sparser information, which makes the reconstruction more challenging. For instance, there is scale ambiguity in monocular video, making it difficult to accurately locate an object within a space [34]. Thus, dynamic scene reconstruction from monocular videos has been considered an ill-posed problem." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.299, + 0.785, + 0.433 + ], + "angle": 0, + "content": "In this paper, we incorporate kinematics to fill the gap between sparse monocular videos and real-world physics, potentially enhancing novel view and time synthesis. Our core idea is that motion within the real-world is governed by the principles of kinematics in physics, implying that representations of the world, such as videos, are inherently anchored in kinematics. Building on this foundation, we generalize conventional displacement fields – often called scene-flow fields [18] – to kinematic fields. In prior work, motion in radiance fields has been controlled via the regularization in the frequency domain [28, 33], displacement fields [18], and deformation fields [25]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.434, + 0.787, + 0.616 + ], + "angle": 0, + "content": "Different from these approaches, our kinematic field regulates motion in radiance fields using kinematics; it innately creates continuous trajectories and enables us to impose regularization based on kinematics. We integrate physics-based regularizers to make the kinematic field physically plausible and thus improve the dynamic radiance field accordingly. For instance, the rigidity regularizer based on the first and the second invariant of the strain rate is applied on the kinematic field, suppressing non-rigid motion in the field. We quantitatively and qualitatively validate the proposed regularizers on the NVIDIA Dynamic Scenes (NDVS) dataset, showing the effectiveness of our method over the existing models. Lastly, the kinematic field uncovers the kinematics within a scene without a ground truth; it estimates not only point-wise 3D displacement within a scene, but also underlying velocity, acceleration, and jerk." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.637, + 0.388, + 0.652 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.667, + 0.784, + 0.728 + ], + "angle": 0, + "content": "Neural radiance fields. In recent years, neural radiance fields [4,23,32] have emerged as a promising approach for learning an implicit neural representation of a 3D scene. Combined with the differentiable volume rendering technique, it has enabled realistic novel-view synthesis [4,23,32,37]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.729, + 0.784, + 0.758 + ], + "angle": 0, + "content": "In essence, a radiance field \\(\\mathcal{F}_{\\mathrm{ST}}\\) maps a 5D vector \\((x,y,z,\\theta ,\\phi)\\) to a 4D output vector \\((\\mathbf{c}_s,\\sigma_s)\\):" + }, + { + "type": "equation", + "bbox": [ + 0.397, + 0.759, + 0.785, + 0.774 + ], + "angle": 0, + "content": "\\[\n\\mathcal {F} _ {\\mathrm {S T}}: (x, y, z, \\theta , \\phi) \\rightarrow (\\mathbf {c} _ {s}, \\sigma_ {s}), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.786, + 0.84 + ], + "angle": 0, + "content": "where the view-dependent color \\(\\mathbf{c}_s = (r,g,b)\\) and density \\(\\sigma_s\\) are learned in relation to the spatial position \\((x,y,z)\\) and the viewing direction \\((\\theta ,\\phi)\\). With the volumetric representation learned via an iterative optimization, volume rendering techniques [15, 22, 23] are employed to aggregate colors and densities along" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.321, + 0.115, + 0.733, + 0.13 + ], + "angle": 0, + "content": "Regularizing Dynamic Radiance Fields with Kinematic Fields" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.177 + ], + "angle": 0, + "content": "a cast ray. Consequently, this leads to realistic novel-view synthesis from the inferred radiance, after being trained on dozens of images." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.178, + 0.785, + 0.283 + ], + "angle": 0, + "content": "Dynamic radiance fields. Radiance fields also facilitate the novel-view synthesis of scenes with moving objects [1,5,16,25,26]. While this is more challenging than the static case, dynamic radiance fields incorporating motion allow for the rendering of a frame at a desired time and viewpoint. As a result, it can be utilized to make visual effects such as bullet time, stabilized, dolly zoomed, and stereo videos. Commonly, the time-dependent radiance fields are referred to as dynamic radiance fields, a term we will adopt throughout this paper." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.284, + 0.787, + 0.315 + ], + "angle": 0, + "content": "In dynamic radiance fields, radiance is dependent on a varying time \\( t \\). This modifies Eq. 1 by introducing a time-dependent term, resulting in:" + }, + { + "type": "equation", + "bbox": [ + 0.388, + 0.328, + 0.786, + 0.344 + ], + "angle": 0, + "content": "\\[\n\\mathcal {F} _ {\\mathrm {D Y}}: (x, y, z, t, \\theta , \\phi) \\rightarrow (\\mathbf {c} _ {d}, \\sigma_ {d}), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.356, + 0.784, + 0.401 + ], + "angle": 0, + "content": "where the color and density are conditioned on time \\( t \\). In this formulation, moving geometries of objects (e.g., translation and deformation) and changing radiance (e.g., moving shadows) can be learned via optimization." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.402, + 0.786, + 0.597 + ], + "angle": 0, + "content": "Learning a dynamic representation presents inherent challenges due to the introduction of temporal dimension. To optimize this representation, a high-quality multi-view dataset can be employed to obtain fine rendering results [1, 5, 16]. Meanwhile, realistically captured imagery of a scene tends to be nonsynchronized and is likely to be obtained with a limited number of cameras and restricted displacements; this often results in a monocular setup [11, 18, 25, 26]. Previous researches have delved into deformable models for monocular videos [20, 25, 26, 31]. In these methods, a deformation field maps a spatio-temporal coordinate to a reference canonical space. While the canonical space aids in maintaining radiance consistency across global time frames, it tends to exhibit reduced effectiveness in generic monocular videos, where new objects might appear within a video [28]. In this study, we embrace the concept of scene flow fields [18], which is constructed without canonical space and deformation networks." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.599, + 0.786, + 0.718 + ], + "angle": 0, + "content": "Regularization with kinematic priors. Monocular videos present challenges in reconstruction due to the constrained viewpoints of moving objects. To address this, integrating auxiliary flow fields has proven to be effective [8, 10, 18, 19]. Existing studies employ either scene flow fields [10, 18, 19] or velocity fields [8] to capture spatial displacements between consecutive frames. Since the scene flows can give spatial correspondences of the dynamic radiance field, it can be used to promote photometric consistency between the renderings of a given ray and its time-deformed counterpart." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Physical priors incorporated with scene flow fields play a pivotal role in regulating the radiance field. Regularization of motion has been applied to the dynamic field, bridging the gap between the sparse visual data and plausible real-world motion. For instance, a smoothness regularizer ensures that the displacements within close proximity are similar to each other, and kinetic energy regularizer promotes consistency between forward and backward flows, resulting in smoother displacement fields [18]. More deeply integrated with physics, the elastic energy regularizer minimizes the deformation between the canonical space" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.356, + 0.127 + ], + "angle": 0, + "content": "W. Im et al." + }, + { + "type": "image", + "bbox": [ + 0.234, + 0.149, + 0.776, + 0.322 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.331, + 0.788, + 0.375 + ], + "angle": 0, + "content": "Fig. 1: Radiance and kinematic fields. This figure summarizes the three fields our method utilizes. The static and dynamic radiance fields are used for rendering, and the kinematic field is used in the training phase, regularizing the dynamic radiance field." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.402, + 0.784, + 0.476 + ], + "angle": 0, + "content": "and the deformed space [25]. In fluid reconstruction works [2, 7], velocity fields grounded in physics have been shown to be critical in accurate 4D reconstruction. For instance, Navier-Stokes and transport equations can be integrated in reconstructing smoke [7] and fluid surfaces [2] with dynamic radiance fields to reconstruct more physically plausible structures." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.478, + 0.788, + 0.6 + ], + "angle": 0, + "content": "We integrate kinematics to reconstruct dynamic radiance fields from generic monocular videos. We introduce the kinematic field, which captures motion through kinematic quantities such as velocity, acceleration, and jerk. Different from existing velocity field-based methods [7,8,34], our method has higher-order terms (i.e., acceleration and jerk) as latent of spatio-temporal correspondences. Further enhancing our approach, we incorporate physics-driven regularizers based on a transport equation [7] and the strain rate tensor, addressing dynamics of moving particles and rigid motion." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.62, + 0.331, + 0.636 + ], + "angle": 0, + "content": "3 Method" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.651, + 0.342, + 0.665 + ], + "angle": 0, + "content": "3.1 Overview" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.673, + 0.788, + 0.795 + ], + "angle": 0, + "content": "Our system takes as input a video sequence, denoted as \\((I_i, t_i, P_i)_{i=1}^N\\), where \\(N\\) is the number of frames. For each frame, we have an image \\(I_i \\in \\mathbb{R}^{H \\times W \\times 3}\\), a timestamp \\(t_i \\in \\mathbb{R}\\), and a precomputed camera pose \\(P_i \\in SE(3)\\) and intrinsic matrix \\(K_i \\in \\mathbb{R}^{3 \\times 3}\\). From this input, our primary objective is to synthesize an image, denoted as \\(\\hat{I}_k\\), corresponding to a novel time and viewpoint \\((t_k, P_k, K_k)\\). The parameters \\(t_k\\), \\(P_k\\), and \\(K_k\\) are not restricted to those found in the original video input, allowing our system to render images from previously unobserved times and viewpoints." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.796, + 0.788, + 0.842 + ], + "angle": 0, + "content": "To achieve this image synthesis for novel times and views, we leverage the radiance-field-based volume rendering [23]. Inspired by existing methods [18,35], we train two distinct radiance fields (Fig. 1): a dynamic radiance field, denoted" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.321, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "Regularizing Dynamic Radiance Fields with Kinematic Fields" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.192 + ], + "angle": 0, + "content": "as \\(\\mathcal{F}_{\\mathrm{DY}}\\) (Eq. 2), and a static radiance field, denoted as \\(\\mathcal{F}_{\\mathrm{ST}}\\) (Eq. 1). To blend the two fields, we utilize a technique from D²-NeRF [35], where rendering weights for dynamic and static samples are derived from the volume densities." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.192, + 0.788, + 0.358 + ], + "angle": 0, + "content": "Addressing the challenge of achieving physically plausible spatio-temporal structures, we introduce the novel kinematic field. The kinematic field accounts for the kinematic quantities (e.g., velocity and acceleration). These quantities can be converted into displacement (Eq. 8), enabling us to use the photometric consistency loss (Eq. 14) for spatio-temporal regularization, as used in previous work [18]. With these kinematic estimations, we apply physics-based regularization techniques. These techniques involve density transport (Eq. 12) and motion rigidity through the strain rate tensor (Eq. 11). Combining the photometric loss (Eq. 14) with these techniques enhances the reconstruction of radiance and geometry, leading to improved novel time and view synthesis performance from underdetermined data such as a monocular video." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.38, + 0.563, + 0.396 + ], + "angle": 0, + "content": "3.2 Dynamic and Static Radiance Fields" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.405, + 0.785, + 0.51 + ], + "angle": 0, + "content": "In monocular dynamic radiance field methods, it is common to decompose the radiance fields into dynamic and static fields [11, 18, 20] to prevent undesirable motion in the static area. We employ tensor-decomposition-based representations [5, 6] for both dynamic and static fields. Specifically, for the dynamic radiance field \\(\\mathcal{F}_{\\mathrm{DY}}\\), we use HexPlane [5], and for the static radiance field \\(\\mathcal{F}_{\\mathrm{ST}}\\), we use TensoRF [6]. Further details on the tensor decomposition can be found in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.511, + 0.787, + 0.586 + ], + "angle": 0, + "content": "Definition. The dynamic and static radiance fields are defined in Eq. 1-2 (\\( \\mathcal{F}_{\\mathrm{ST}} \\) and \\( \\mathcal{F}_{\\mathrm{DY}} \\)). Here, the inputs correspond to the position \\( (x,y,z) \\), the timestamp \\( t \\), and the viewing direction \\( (\\theta, \\phi) \\) from the camera origin. Each field yields the color \\( \\mathbf{c} \\) and the volume density \\( \\sigma \\). Note that the static field is independent of time \\( t \\), ensuring a consistent structure throughout the whole time span." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.587, + 0.788, + 0.707 + ], + "angle": 0, + "content": "Composite rendering. We use a composite rendering method to obtain the final rendering with the two radiance fields. It is to be noted that individual components (i.e., dynamic-only and static-only) can be rendered using the standard volume rendering technique [23]. To combine both fields, we employ an additive composition method in which each field influences the transmittance of a ray [35]. Let's consider a camera ray \\(\\mathbf{r}(\\tau) = \\mathbf{o} + \\tau \\mathbf{u}\\), where \\(\\mathbf{o}\\) is the origin and \\(\\mathbf{u}\\) is the direction. With a predefined range for the ray distance \\([\\tau_n, \\tau_f]\\), the rendering function is formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.297, + 0.717, + 0.786, + 0.75 + ], + "angle": 0, + "content": "\\[\n\\hat {C} (\\mathbf {r}, t _ {i}) = \\int_ {\\tau_ {n}} ^ {\\tau_ {f}} T (\\tau) \\left(\\sigma_ {s} (\\tau) \\cdot \\mathbf {c} _ {s} (\\tau) + \\sigma_ {d} (\\tau , t _ {i}) \\cdot \\mathbf {c} _ {d} (\\tau , t _ {i})\\right) d \\tau , \\tag {3}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.359, + 0.753, + 0.786, + 0.786 + ], + "angle": 0, + "content": "\\[\n\\text {w h e r e} T (\\tau) = \\exp \\left(- \\int_ {\\tau_ {n}} ^ {\\tau} \\left(\\sigma_ {s} (u) + \\sigma_ {d} (u, t _ {i})\\right) d u\\right). \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Here, each notation for the sample point is simplified for concise representation (e.g., \\(\\sigma_s(\\tau) = \\sigma_s(\\mathbf{r}(\\tau))\\)). We visualize the example of the composite rendering and separate rendering in the first and second rows in Fig. 2." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.356, + 0.127 + ], + "angle": 0, + "content": "W. Im et al." + }, + { + "type": "image", + "bbox": [ + 0.292, + 0.147, + 0.725, + 0.357 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.37, + 0.788, + 0.48 + ], + "angle": 0, + "content": "Fig. 2: Visualization of each predicted component. The top row presents synthesized views at different times. In (a) and (b), the dynamic and static components of the scene at time \\( t_2 \\) are depicted separately. (b) In the case of an inobservable static area in the entire sequence (e.g., the space behind the jumping people), radiance might not be correct. The displacement field (c) can be computed by Taylor approximation (Eq. 8) with motion fields (d-f): velocity, acceleration, and jerk. Each field is visualized through reprojecting each field to the camera view. The standard HSV visualization [3] is used to colorize arrows." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.512, + 0.788, + 0.558 + ], + "angle": 0, + "content": "Dynamic and static separation. To factorize a scene into static and dynamic parts, we use a metric indicating the likelihood that a point belongs to the dynamic fields [35]. This metric is given by:" + }, + { + "type": "equation", + "bbox": [ + 0.447, + 0.57, + 0.786, + 0.598 + ], + "angle": 0, + "content": "\\[\np _ {d} = \\frac {\\sigma_ {d}}{\\sigma_ {d} + \\sigma_ {s}}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.608, + 0.748, + 0.623 + ], + "angle": 0, + "content": "where \\(\\sigma_{d}\\) and \\(\\sigma_{s}\\) are density values sampled from a spatio-temporal point." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.624, + 0.787, + 0.655 + ], + "angle": 0, + "content": "To facilitate the separation of the radiance fields, we employ the skewed binary entropy loss, as proposed by [35]. The entropy loss is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.341, + 0.666, + 0.787, + 0.686 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {b} = - \\left(p _ {d} ^ {k} \\cdot \\log \\left(p _ {d} ^ {k}\\right) + \\left(1 - p _ {d} ^ {k}\\right) \\cdot \\log \\left(1 - p _ {d} ^ {k}\\right)\\right), \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.697, + 0.788, + 0.743 + ], + "angle": 0, + "content": "where the exponent \\( k \\geq 1 \\) is the skew factor, controlling a bias toward static explanation. This entropy-based loss can encourage the radiance fields to have a clearer distinction between the static and dynamic components." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.767, + 0.396, + 0.781 + ], + "angle": 0, + "content": "3.3 Kinematic Field" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.787, + 0.84 + ], + "angle": 0, + "content": "The kinematic field has a spatio-temporal structure to facilitate spatial and temporal queries. We use the tensor-decomposition-based representation [5] for the kinematic field. As depicted in Fig. 1, the kinematic field serves to regulate" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.321, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "Regularizing Dynamic Radiance Fields with Kinematic Fields" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.143, + 0.495, + 0.221 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.25, + 0.224, + 0.466, + 0.236 + ], + "angle": 0, + "content": "(a) Without kinematic regularization" + }, + { + "type": "image", + "bbox": [ + 0.506, + 0.143, + 0.784, + 0.222 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.547, + 0.224, + 0.746, + 0.236 + ], + "angle": 0, + "content": "(b) With kinematic regularization" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.25, + 0.788, + 0.32 + ], + "angle": 0, + "content": "Fig. 3: Effect of kinematic regularization. We visualize the rendered RGB and motion of each field. Without kinematic regularization, motion fields tend to show granular patterns. Our kinematic fields not only make the field smoother but also satisfy the kinematic property, i.e., \\(\\mathbf{a} = \\partial \\mathbf{v} / \\partial t + \\mathbf{v} \\cdot \\nabla \\mathbf{v}\\). We abbreviate the advective equation to 'Adv.' in the figure." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.348, + 0.784, + 0.379 + ], + "angle": 0, + "content": "the radiance field, ensuring that the motion in the radiance field is physically plausible." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.379, + 0.547, + 0.393 + ], + "angle": 0, + "content": "Definition. The kinematic field is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.39, + 0.404, + 0.786, + 0.421 + ], + "angle": 0, + "content": "\\[\n\\mathcal {F} _ {\\mathrm {K}}: (x, y, z, t) \\rightarrow (\\mathbf {v}, \\mathbf {a}, \\mathbf {j}, \\dots), \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.431, + 0.784, + 0.476 + ], + "angle": 0, + "content": "where each vector represents velocity, acceleration, and jerk, such that \\(\\mathbf{v} = (v_x, v_y, v_z)\\), \\(\\mathbf{a} = (a_x, a_y, a_z)\\), and \\(\\mathbf{j} = (j_x, j_y, j_z)\\). Note that we can extend the field into higher-order quantities such as snap (4th), crackle (5th), and so on." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.477, + 0.785, + 0.549 + ], + "angle": 0, + "content": "Displacement. Using the provided kinematic quantities, we can approximate a particle's displacement at position \\((x,y,z)\\) and time \\(t + \\Delta t\\) via the Taylor approximation, assuming that the kinematic quantities are consistent within a local temporal range \\(\\Delta t\\). The displacement \\(d(x,y,z,t,\\Delta t)\\) can be approximated by:" + }, + { + "type": "equation", + "bbox": [ + 0.318, + 0.55, + 0.785, + 0.581 + ], + "angle": 0, + "content": "\\[\nd (x, y, z, t, \\Delta t) \\approx \\frac {(\\Delta t) \\mathbf {v}}{1 !} + \\frac {(\\Delta t) ^ {2} \\mathbf {a}}{2 !} + \\frac {(\\Delta t) ^ {3} \\mathbf {j}}{3 !} + \\dots . \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.584, + 0.784, + 0.629 + ], + "angle": 0, + "content": "Computing the displacement of a point within a temporal range allows for computing the photometric consistency loss (Eq. 14) and scene flows from the kinematic field." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.651, + 0.473, + 0.667 + ], + "angle": 0, + "content": "3.4 Kinematic Regularization" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.784, + 0.734 + ], + "angle": 0, + "content": "Given the kinematic field, it is important to make sure the learned motion dynamics are physically plausible. As Fig. 3 indicates, learning kinematic fields without adequate regularization may result in a sub-optimal kinematic field. The subsequent sections elucidate regularization strategies." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.785, + 0.84 + ], + "angle": 0, + "content": "Integrity of kinematic fields. Learning the kinematic field \\(\\mathcal{F}_{\\mathrm{K}}\\) naively does not guarantee the integrity of the kinematic quantities. During training, the quantities are supervised via the photometric consistency loss (Eq. 14), which is applied upon the approximated displacement (Eq. 8). It is important to note that even with supervision by the displacement, the physical relationships among the kinematic quantities may not be correct, since the higher-order terms in the displacement equation (Eq. 8) make the system underdetermined." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.356, + 0.128 + ], + "angle": 0, + "content": "W. Im et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.221 + ], + "angle": 0, + "content": "Given a velocity field, we can derive the acceleration field using the principle of advective acceleration. Given acceleration \\(\\mathbf{a}\\) and velocity \\(\\mathbf{v}\\) at a spatiotemporal coordinate, the partial derivative and advection yield \\(\\mathbf{a} = \\partial \\mathbf{v} / \\partial t + \\mathbf{v} \\cdot \\nabla \\mathbf{v}\\). Similarly, we can formulate the relationship between \\(\\mathbf{v}\\), \\(\\mathbf{a}\\) and \\(\\mathbf{j}\\) as: \\(\\mathbf{j} = \\partial \\mathbf{a} / \\partial t + \\mathbf{v} \\cdot \\nabla \\mathbf{a}\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.223, + 0.784, + 0.25 + ], + "angle": 0, + "content": "To ensure kinematic consistency, we introduce the kinematic loss \\(\\mathcal{L}_{\\mathrm{K}}\\) defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.31, + 0.248, + 0.785, + 0.285 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {K}} = \\left\\| \\mathbf {a} - \\frac {\\partial \\mathbf {v}}{\\partial t} - \\mathbf {v} \\cdot \\nabla \\mathbf {v} \\right\\| _ {2} ^ {2} + \\left\\| \\mathbf {j} - \\frac {\\partial \\mathbf {a}}{\\partial t} - \\mathbf {v} \\cdot \\nabla \\mathbf {a} \\right\\| _ {2} ^ {2} + \\dots , \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.288, + 0.784, + 0.334 + ], + "angle": 0, + "content": "where \\(\\nabla\\) is the Jacobian operation. To compute the partial derivatives and Jacobian matrices, we use a numerical method as presented in a neural surface reconstruction work [17]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.334, + 0.785, + 0.438 + ], + "angle": 0, + "content": "Physics-informed regularization. In the context of regularizing radiance fields through the use of an auxiliary kinematic field, it is important to acknowledge the highly underdetermined nature of the problem. Essentially, there exists a broad spectrum of possible potential kinematic field solutions that could perfectly fit the observed training views. This complexity necessitates the introduction of a regularization strategy that can guide the motion representation toward more accurately reflecting real-world physics." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.439, + 0.784, + 0.5 + ], + "angle": 0, + "content": "While not universally ideal, rigidity [25] and incompressibility [2,7] have been identified as effective metrics for encapsulating real-world dynamics; in videos that we normally encounter, objects in a scene have nearly consistent density and are mostly rigid in a short time range." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.5, + 0.784, + 0.574 + ], + "angle": 0, + "content": "Rigidity. One benefit of having a velocity field is that we can measure the rate of deformation by its derivatives. A strain rate tensor \\( D \\) defines a kinematic property of a deformation; it captures the rate at which a material undergoes deformation, and it is crucial for ensuring that objects maintain their structural integrity during motion." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.575, + 0.784, + 0.62 + ], + "angle": 0, + "content": "In this paper, we propose using the invariants of strain rate tensor [12, 29] as a measure of rigidity. The velocity gradient \\(\\nabla \\mathbf{v}\\) sampled from the kinematic field, leads to the strain rate tensor \\(D\\) by:" + }, + { + "type": "equation", + "bbox": [ + 0.42, + 0.629, + 0.785, + 0.657 + ], + "angle": 0, + "content": "\\[\nD = \\frac {1}{2} (\\nabla \\mathbf {v} + (\\nabla \\mathbf {v}) ^ {T}). \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.664, + 0.783, + 0.694 + ], + "angle": 0, + "content": "The rigidity loss function \\(\\mathcal{L}_R\\) is formulated based on the first and second invariants of the strain rate tensor:" + }, + { + "type": "equation", + "bbox": [ + 0.337, + 0.702, + 0.785, + 0.747 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {R} = \\underbrace {\\lambda_ {\\mathrm {d i v}} (\\nabla \\cdot \\mathbf {v}) ^ {2}} _ {\\text {1 s t i n v a r i a n t}} + \\underbrace {\\left(\\frac {1}{2} (\\operatorname {t r} (D) ^ {2} - \\operatorname {t r} (D \\cdot D)))\\right) ^ {2}} _ {\\text {2 n d i n v a r i a n t}}, \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.756, + 0.784, + 0.817 + ], + "angle": 0, + "content": "where \\(\\operatorname{tr}(\\cdot)\\) denotes the trace operator and \\(\\lambda_{\\mathrm{div}}\\) controls the balance between the two terms. Note that the first invariant of strain rate measures dilational change (i.e., volumetric change) with respect to time and the second invariant of strain rate measures distortional change." + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.822, + 0.379, + 0.843 + ], + "angle": 0, + "content": "\\(\\overline{4f'(x)} = \\frac{f(x + \\epsilon) - f(x - \\epsilon)}{2\\epsilon}\\)" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.322, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "Regularizing Dynamic Radiance Fields with Kinematic Fields" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.15, + 0.462, + 0.319 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.329, + 0.466, + 0.482 + ], + "angle": 0, + "content": "Fig.4: Visualization of density variation by a spatial coordinate \\( x \\). The plot displays the gradient \\( -\\partial \\mathcal{L}_{\\mathrm{T}} / \\partial v_{\\sigma_d} \\) with arrows. With the velocity field directed to the right (i.e., \\( \\mathbf{v}_x = 0.3 \\)), we can compute the gradient of the transport regularization \\( \\mathcal{L}_{\\mathrm{T}} \\) w.r.t. the rate of density change \\( v_{\\sigma_d} = \\partial \\sigma_d / \\partial t \\). Minimizing \\( \\mathcal{L}_{\\mathrm{T}} \\) allows us to render the density at \\( t + \\Delta t \\) aligned with the flow field." + }, + { + "type": "image", + "bbox": [ + 0.486, + 0.149, + 0.784, + 0.322 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.474, + 0.33, + 0.788, + 0.484 + ], + "angle": 0, + "content": "Fig. 5: Photometric consistency loss. During optimization, we deform each ray from a reference time and pose \\((t,P_{t})\\) to different timestamps \\(\\gamma\\) and \\(i\\). Here, we sample \\(i\\) from the neighboring frame times, and we consequently sample an intermediate timestamp \\(\\gamma \\sim \\mathcal{U}(t,i)\\). Given each timestamp, we can deform a ray using Eq. 8. The photometric loss \\(\\mathcal{L}_{\\mathrm{photo}}\\) is computed based on the color consistency between the deformed ray and the original color, enhancing temporal consistency." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.511, + 0.788, + 0.572 + ], + "angle": 0, + "content": "Transport. A transport equation has been adopted in physics-informed neural networks [7]. By simplifying the Navier-Stokes equation, a transport loss can be defined as: \\(\\frac{\\partial\\sigma}{\\partial t} +\\mathbf{v}\\cdot \\nabla \\sigma = 0\\). We leverage the physics-informed regularizers based on the transport equation and the divergence of the velocity:" + }, + { + "type": "equation", + "bbox": [ + 0.407, + 0.577, + 0.786, + 0.613 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {T}} = \\left(\\frac {\\partial \\sigma_ {d}}{\\partial t} + \\mathbf {v} \\cdot \\nabla \\sigma_ {d}\\right) ^ {2}, \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.619, + 0.788, + 0.667 + ], + "angle": 0, + "content": "where \\(\\sigma_{d}\\) is the volume density used for volume rendering. This aligns the velocity field with the density field (i.e., Fig. 4), preventing a sudden disappearance or appearance of objects in the dynamic radiance field." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.684, + 0.429, + 0.7 + ], + "angle": 0, + "content": "3.5 Learning Objectives" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.706, + 0.788, + 0.812 + ], + "angle": 0, + "content": "Cycle consistency. The cycle consistency is essential to ensure that the trajectories derived from the kinematic field do not introduce misalignments [18]. We use cycle constraints among four timestamps \\( t \\), \\( i \\), \\( j \\), and \\( \\gamma \\). Given a reference time \\( t \\), we randomly sample time \\( i \\) from the neighboring frames within a threshold. We define \\( j = t + 2(i - t) \\), making it twice the distance from \\( t \\) as \\( i \\) is from \\( t \\). \\( \\gamma \\sim \\mathcal{U}(t,i) \\) is an intermediate time, uniformly sampled between \\( t \\) and \\( i \\). With the time variables, we formulate our loss:" + }, + { + "type": "equation", + "bbox": [ + 0.256, + 0.82, + 0.787, + 0.838 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {C}} = \\rho (\\mathbf {d} _ {t \\rightarrow i} + \\mathbf {d} _ {i \\rightarrow t}) + \\rho (\\mathbf {d} _ {t \\rightarrow i \\rightarrow j} - \\mathbf {d} _ {t \\rightarrow j}) + \\rho (\\mathbf {d} _ {t \\rightarrow \\gamma \\rightarrow i} - \\mathbf {d} _ {t \\rightarrow i}), \\tag {13}\n\\]" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.355, + 0.127 + ], + "angle": 0, + "content": "W. Im et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.192 + ], + "angle": 0, + "content": "where \\(\\rho\\) denotes the generalized chabonnier loss [30]. Here, we abbreviate the displacement \\(d(\\cdot)\\) (Eq. 8) for conciseness: \\(\\mathbf{d}_{t\\to i}\\) is shorthand for \\(d(\\mathbf{x},t,i - t)\\), and \\(\\mathbf{d}_{t\\to i\\to j}\\) simplifies \\(d(\\mathbf{x} + \\mathbf{d}_{t\\to i},i,j - i)\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.193, + 0.785, + 0.375 + ], + "angle": 0, + "content": "Photometric consistency loss. To enhance temporal coherency, we deform rays into nearby frame times and enhance the consistency among them (Fig. 5). Specifically, given a reference ray \\(\\mathbf{r}_t\\) at time \\(t\\), a pixel color \\(\\hat{C}(\\mathbf{r}_t, t)\\) can be obtained using the volume rendering (Eq. 4). In addition to the reference ray, we apply ray deformation based on the predicted kinematic quantities. As in cycle consistency (Eq. 13), we sample time \\(i\\) from neighboring frames, and time \\(\\gamma \\sim \\mathcal{U}(t, i)\\) between times \\(t\\) and \\(i\\). Different from existing approaches [10, 18], we deform rays with kinematic quantities using Taylor approximation. The deforming process results in two deformed rays \\(\\mathbf{r}_{t \\to i}\\) and \\(\\mathbf{r}_{t \\to \\gamma}\\), and two rendered colors \\(\\hat{C}(\\mathbf{r}_{t \\to i}, i)\\) and \\(\\hat{C}(\\mathbf{r}_{t \\to \\gamma}, \\gamma)\\), correspondingly. Based on the deformed rays, we apply the photometric consistency loss, comparing the rendered color with the ground-truth color. The loss function is formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.232, + 0.386, + 0.785, + 0.424 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {\\text {p h o t o}} = \\| \\hat {C} (\\mathbf {r} _ {t}, t) - C (\\mathbf {r} _ {t}, t) \\| _ {2} ^ {2} \\tag {14} \\\\ + \\lambda \\alpha (\\mathbf {r} _ {t}) \\rho (\\hat {C} (\\mathbf {r} _ {t \\rightarrow i}, i) - C (\\mathbf {r} _ {t}, t)) + \\lambda \\alpha (\\mathbf {r} _ {t}) \\rho (\\hat {C} (\\mathbf {r} _ {t \\rightarrow \\gamma}, \\gamma) - C (\\mathbf {r} _ {t}, t)), \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.434, + 0.784, + 0.478 + ], + "angle": 0, + "content": "where \\(\\lambda\\) is a loss weight for the temporal consistency, and \\(\\alpha (\\mathbf{r}_t)\\) is the sum of rendering weights of the dynamic densities, representing the dynamic likelihood of the ray \\(\\mathbf{r}_t\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.48, + 0.785, + 0.54 + ], + "angle": 0, + "content": "Final objective. Our final objective function, denoted as \\(\\mathcal{L}\\), is the weighted sum of the various loss terms discussed previously. To keep things simple with the variety of loss functions, we symbolize the losses to \\(\\mathcal{L}_{\\mathrm{photo}}\\), \\(\\mathcal{L}_{\\mathrm{kinematic}}\\), and \\(\\mathcal{L}_{\\mathrm{reg}}\\). This results in the final loss:" + }, + { + "type": "equation", + "bbox": [ + 0.375, + 0.551, + 0.784, + 0.567 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathbb {E} _ {\\mathbf {r} _ {t}} \\left[ \\mathcal {L} _ {\\text {p h o t o}} + \\mathcal {L} _ {\\text {k i n e m a t i c}} + \\mathcal {L} _ {\\text {r e g}} \\right]. \\tag {15}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.577, + 0.784, + 0.668 + ], + "angle": 0, + "content": "Here, \\(\\mathcal{L}_{\\mathrm{kinematic}}\\) is the weighted sum of the kinematic integrity \\(\\mathcal{L}_{\\mathrm{K}}\\), transport \\(\\mathcal{L}_{\\mathrm{T}}\\), and rigidity \\(\\mathcal{L}_{\\mathrm{R}}\\). Note that \\(\\mathbb{E}_{\\mathbf{r}_t}[\\mathcal{L}_{\\mathrm{kinematic}}]\\) is computed with randomly sampled points along the ray \\(\\mathbf{r}_t\\) from each minibatch. The additional regularization \\(\\mathcal{L}_{\\mathrm{reg}}\\) includes the skewed entropy \\(\\mathcal{L}_b\\), the cycle consistency \\(\\mathcal{L}_{\\mathrm{C}}\\), total-variation loss [5,6], distortion loss [4], monocular depth supervision [18], normal regularization [14,32] and optical flow supervision [18]." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.689, + 0.367, + 0.707 + ], + "angle": 0, + "content": "4 Experiment" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.719, + 0.784, + 0.84 + ], + "angle": 0, + "content": "Dataset. We use the NVIDIA dynamic view synthesis dataset (NDVS) [36] to evaluate the effectiveness of our model. For the monocular setup, we adopt the preprocessing steps proposed by NSFF [18] and DynamicNeRF [10]. This dataset has 8 scenes and is captured with 12 synchronized cameras. When used for a monocular video reconstruction task, the dataset shows relatively low effective multi-view factors (EMFs) [11], compared to other monocular datasets; this means that the dataset is more challenging to reconstruct due to the less camera motion and faster-moving objects." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.321, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Regularizing Dynamic Radiance Fields with Kinematic Fields" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.266 + ], + "angle": 0, + "content": "There are two widely used monocular NDVS protocols: 12-frames [10] and 24-frames [18]. Most papers opt for either protocol; however, we include both protocols for rigorous evaluation. Meanwhile, 24-frames-sparse is proposed by us, as neither existing protocol provides novel time evaluation; we leave out odd-numbered frames at unseen times for testing. For the 'Dynamic Only' marks in tables, we use the dynamic mask to evaluate only in terms of dynamic pixels; please refer to [10,18] for how to generate dynamic masks. We detail the protocols below." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.28, + 0.784, + 0.34 + ], + "angle": 0, + "content": "- 24-frames: this sequence is configured with 24 training frames, and 264 test frames from unseen multi-views. For a fair comparison, we utilize the dynamic mask, optical flow, and monocular depth, extracted by the same model with the prior work [18]." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.34, + 0.784, + 0.384 + ], + "angle": 0, + "content": "- 24-frames-sparse: we reduce the 24-frames sequence into 12 frames by sampling the even-numbered frames for training. This sparse sampling allows for testing novel view synthesis at both seen times and novel times." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.386, + 0.784, + 0.43 + ], + "angle": 0, + "content": "- 12-frames: this sequence has 12 training and 12 testing samples. We use the preprocessed dynamic mask, optical flow, and monocular depth downloadable from the code repository of DynamicNeRF [10]." + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.28, + 0.784, + 0.43 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.453, + 0.457, + 0.468 + ], + "angle": 0, + "content": "4.1 Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.478, + 0.784, + 0.55 + ], + "angle": 0, + "content": "Fields configuration. Our radiance and kinematic fields are configured with HexPlane [5] and TensoRF [6]. After features are sampled from the point queries, small MLP decoders process the features to yield colors and kinematic outputs. For density outputs, we use a linear layer following the practices in [5,6]. Further details on field configurations are available in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.554, + 0.784, + 0.67 + ], + "angle": 0, + "content": "Training. We use 70,000 steps to optimize our models, which takes about 7 hours with one NVIDIA V100 or A100 GPU. We initialize the static field for 5,000 steps, taking approximately 5 minutes. After finishing the initialization of the static field, we jointly train the fields: dynamic, static, and kinematic fields. During the joint training, we lower the learning rate of the static field by the factor of 0.1 at predefined iterations to prevent dynamic objects from being portrayed as static objects. Additional training details are provided in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.784, + 0.717 + ], + "angle": 0, + "content": "Evaluation metrics. We report peak signal-to-noise ratio (PSNR), structural similarity (SSIM), and perceptual similarity (LPIPS). These three metrics have been used as standard in prior studies to evaluate the quality of renderings." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.793, + 0.838 + ], + "angle": 0, + "content": "Base architecture (HexPlane). Since our model is based on tensor-decomposed radiance fields, we compare our models with HexPlane [5]. The baseline model is the dynamic radiance field \\(\\mathcal{F}_{\\mathrm{DY}}\\), without the static radiance field \\(\\mathcal{F}_{\\mathrm{ST}}\\) and the kinematic field \\(\\mathcal{F}_{\\mathrm{K}}\\). For a fair comparison, we use monocular depth supervision both to the baseline model and our models, although the original HexPlane model does not support the supervision from monocular depth. Since the base HexPlane model does not have an auxiliary flow field, we do not use the optical flow supervision and the photometric consistency loss." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.356, + 0.128 + ], + "angle": 0, + "content": "W. Im et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.784, + 0.173 + ], + "angle": 0, + "content": "Table 1: Ablation study on Balloon1 scene of the NDVS dataset (24-frames-sparse). Please refer to the supplementary material for ablation results on other NDVS scenes." + }, + { + "type": "table", + "bbox": [ + 0.22, + 0.185, + 0.784, + 0.272 + ], + "angle": 0, + "content": "
MethodsNovel TimesSeen Times
FullDynamic OnlyFullDynamic Only
PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓
HexPlane [5]19.840.6680.14518.120.4190.25118.930.6490.13717.740.4290.202
Ours25.670.8650.06421.160.6250.17025.800.8670.05021.350.6320.121
w/o Higher order motion24.360.8420.07719.930.5560.20024.630.8460.06120.220.5720.145
w/o Lkinematic24.960.8530.06820.150.5680.18525.300.8580.05220.650.5940.129
w/o Kinematic integrity25.460.8620.06620.780.6070.17925.570.8650.05121.020.6200.126
w/o Rigidity25.110.8560.06620.400.5820.18125.430.8600.05120.820.6030.123
" + }, + { + "type": "image", + "bbox": [ + 0.23, + 0.285, + 0.775, + 0.541 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.553, + 0.788, + 0.608 + ], + "angle": 0, + "content": "Fig. 6: Comparative visualization of kinematic fields. The first row shows RGB values rendered from our dynamic radiance field, showing PSNR values from a ground truth RGB image. From the second to the last row, we show flow (i.e., displacement), velocity, and acceleration." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.642, + 0.519, + 0.656 + ], + "angle": 0, + "content": "4.2 Evaluation of Kinematic Fields" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.673, + 0.788, + 0.749 + ], + "angle": 0, + "content": "Ablation of kinematic losses. To show the effectiveness of our approach, we conduct a comprehensive ablation study on the Balloon1 scene of the NDVS dataset, focusing on kinematic loss components. In Table 1, we show ablative results from seen and unseen frame times. Our findings indicate that the model achieves optimal performance when all components are integrated." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Fig. 6 delves into the kinematic field analysis. Omitting higher order motion restricts the model to utilizing only the velocity field, which proves insufficient for capturing non-linear motion. This limitation hinders the model's ability to minimize kinematic losses, leading to a suboptimal flow representation. In the absence of \\(\\mathcal{L}_{\\mathrm{kinematic}}\\), the radiance field suffers from reduced PSNR scores, which can be attributed to the physically incorrect motion and geometry. Compared" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.322, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "Regularizing Dynamic Radiance Fields with Kinematic Fields" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.149, + 0.357, + 0.246 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.363, + 0.149, + 0.497, + 0.246 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.506, + 0.149, + 0.637, + 0.246 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.648, + 0.149, + 0.779, + 0.246 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.262, + 0.785, + 0.318 + ], + "angle": 0, + "content": "Fig. 7: Motion Order. Reconstruction Accuracy for the Balloon1 Scene (24-framesparse): The \\( x \\)-axis of each plot categorizes maximum motion order as follows: velocity (1), acceleration (2), jerk (3), snap (4), and crackle (5). The metric utilized for each plot is specified at the top." + }, + { + "type": "table_caption", + "bbox": [ + 0.215, + 0.332, + 0.784, + 0.361 + ], + "angle": 0, + "content": "Table 2: Quantitative comparison on NDVS. This table reports average results on all scenes. Please refer to the supplementary material for per-scene results." + }, + { + "type": "table", + "bbox": [ + 0.221, + 0.375, + 0.427, + 0.495 + ], + "angle": 0, + "content": "
(a) 12-frames
MethodsFull
PSNR↑LPIPS↓
NeRF [23]*21.990.250
NSFF [18]*24.330.199
DynamicNeRF [10]*26.100.082
HyperNeRF [26]*17.600.367
RoDynRF [20]*25.890.065
Ours26.330.071
*Results reported in [20]
" + }, + { + "type": "table", + "bbox": [ + 0.431, + 0.375, + 0.784, + 0.494 + ], + "angle": 0, + "content": "
(b) 24-frames
MethodsFullDynamic Only
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
NeRF [23]*24.900.8930.09816.980.5320.314
Luo et al. [21]*21.370.7460.14116.970.5300.207
Yoon et al. [36]*21.780.7610.12717.340.5470.200
NSFF [18]*28.190.9280.04521.910.7580.097
NSFF [18]28.040.9270.04521.340.7400.111
Ours28.780.9360.04222.130.7740.094
*Results reported in [18]
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.523, + 0.785, + 0.553 + ], + "angle": 0, + "content": "to the full model, the model without the rigidity loss results in noisy, non-rigid motion, resulting in incorrect motion and the worse reconstruction accuracy." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.553, + 0.788, + 0.704 + ], + "angle": 0, + "content": "Motion order. Our study represents motion through various kinematic quantities. Adjusting the kinematic field to include different orders of motion, we can model our system using just velocity [7, 8], or extend to higher orders – fourth (snap) and fifth (crackle) – by extending the Taylor series (Eq. 8) and the kinematic integrity (Eq. 9) equations. Fig. 7 illustrates the influence of motion order on reconstruction accuracy for the NDVS dataset. As motion order increases, the reconstruction accuracy first improves, peaking at jerk, but further orders degrades the accuracy. We conjecture that the augmentation of model complexity with higher orders of kinematic quantities may render the model more susceptible to overfitting." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.726, + 0.524, + 0.74 + ], + "angle": 0, + "content": "4.3 Comparison with SOTA models" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Quantitative results. NSFF [18] is a dynamic neural radiance field model, which utilizes scene-flow fields for the regularization of a dynamic radiance field. Since NSFF is a neural model, the network training takes about 2 days on two V100 GPUs (96 GPU hours), using the official code by the authors. Note that optimizing a scene using our method takes about 7 hours per scene, which makes ours 12 times faster than NSFF in terms of GPU hours. We compare our results" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.356, + 0.127 + ], + "angle": 0, + "content": "W. Im et al." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.143, + 0.784, + 0.329 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.34, + 0.785, + 0.369 + ], + "angle": 0, + "content": "Fig. 8: Qualitative results on NDVS (12-frames). The second and fourth rows illustrate the overlay of the synthesized and ground truth views of the scene." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.396, + 0.785, + 0.44 + ], + "angle": 0, + "content": "with NSFF [18] in Table 2b. In the table, our method shows better accuracy in terms of all metrics. In addition, we report results on NDVS (12-frames) in Table 2a." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.441, + 0.785, + 0.547 + ], + "angle": 0, + "content": "Qualitative results. Fig. 8 illustrates qualitative results, including the synthesized images, ground truth, and their overlaid images. As shown in the first and third rows, our method improves the visual quality of the rendered views. Additionally, the overlaid images demonstrate the geometrical accuracy of our model, showcasing its better alignment with the ground truth. This enhanced performance is attributed to the kinematics and physics incorporated in our methodology, which is fundamental to the improvement." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.568, + 0.359, + 0.584 + ], + "angle": 0, + "content": "5 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.597, + 0.785, + 0.733 + ], + "angle": 0, + "content": "In this paper, we have presented a novel method to learn dynamic radiance fields with kinematic fields. We propose the kinematic field, where kinematic quantities are learned by joint learning with the radiance fields. Using the kinematic quantities from the kinematic fields, we propose physics-based loss functions that underpin our method. Lastly, we quantitatively and qualitatively show the effectiveness of our method on the monocular dataset. We believe that this work not only introduces the potential of incorporating kinematics into the learning of radiance fields but also sets a foundation for future direction in physics-based dynamic scene understanding." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.749, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Acknowledgment. This project was partly supported by the Naver Cloud Corporation. Additionally, this work received support from the Institute of Information & communications Technology Planning & Evaluation (IITP) grant (RS-2023-00237965) and the National Research Foundation of Korea (NRF) grant (No. RS-2023-00208506(2024)), funded by the Korea government (MSIT). Sung-Eui Yoon is a corresponding author." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.322, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Regularizing Dynamic Radiance Fields with Kinematic Fields" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.323, + 0.16 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.173, + 0.785, + 0.214 + ], + "angle": 0, + "content": "1. Attal, B., Huang, J.B., Richardt, C., Zollhoefer, M., Kopf, J., O'Toole, M., Kim, C.: Hyperreel: High-fidelity 6-dof video with ray-conditioned sampling. arXiv preprint arXiv:2301.02238 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.215, + 0.785, + 0.255 + ], + "angle": 0, + "content": "2. Baieri, D., Esposito, S., Maggioli, F., Rodola, E.: Fluid dynamics network: Topology-agnostic 4d reconstruction via fluid dynamics priors. arXiv preprint arXiv:2303.09871 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.256, + 0.785, + 0.295 + ], + "angle": 0, + "content": "3. Baker, S., Scharstein, D., Lewis, J., Roth, S., Black, M.J., Szeliski, R.: A database and evaluation methodology for optical flow. International journal of computer vision 92, 1-31 (2011)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.296, + 0.785, + 0.35 + ], + "angle": 0, + "content": "4. Barron, J.T., Mildenhall, B., Tancik, M., Hedman, P., Martin-Brualla, R., Srinivasan, P.P.: Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5855-5864 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.351, + 0.785, + 0.377 + ], + "angle": 0, + "content": "5. Cao, A., Johnson, J.: Hexplane: A fast representation for dynamic scenes. arXiv preprint arXiv:2301.09632 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.378, + 0.785, + 0.418 + ], + "angle": 0, + "content": "6. Chen, A., Xu, Z., Geiger, A., Yu, J., Su, H.: Tensorf: Tensorial radiance fields. In: Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXXII. pp. 333-350. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.419, + 0.785, + 0.459 + ], + "angle": 0, + "content": "7. Chu, M., Liu, L., Zheng, Q., Franz, E., Seidel, H.P., Theobalt, C., Zayer, R.: Physics informed neural fields for smoke reconstruction with sparse data. ACM Transactions on Graphics (TOG) 41(4), 1-14 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.46, + 0.785, + 0.5 + ], + "angle": 0, + "content": "8. Du, Y., Zhang, Y., Yu, H.X., Tenenbaum, J.B., Wu, J.: Neural radiance flow for 4d view synthesis and video processing. In: 2021 IEEE/CVF International Conference on Computer Vision (ICCV). pp. 14304-14314. IEEE Computer Society (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.501, + 0.785, + 0.54 + ], + "angle": 0, + "content": "9. Fang, J., Yi, T., Wang, X., Xie, L., Zhang, X., Liu, W., Niešner, M., Tian, Q.: Fast dynamic radiance fields with time-aware neural voxels. In: SIGGRAPH Asia 2022 Conference Papers. pp. 1-9 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.541, + 0.785, + 0.581 + ], + "angle": 0, + "content": "10. Gao, C., Saraf, A., Kopf, J., Huang, J.B.: Dynamic view synthesis from dynamic monocular video. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 5712-5721 (October 2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.582, + 0.785, + 0.636 + ], + "angle": 0, + "content": "11. Gao, H., Li, R., Tulsiani, S., Russell, B., Kanazawa, A.: Monocular dynamic view synthesis: A reality check. In: Oh, A.H., Agarwal, A., Belgrave, D., Cho, K. (eds.) Advances in Neural Information Processing Systems (2022), https://openreview.net/forum?id=pCrB8orUkSq" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.637, + 0.785, + 0.677 + ], + "angle": 0, + "content": "12. Hart-Smith, L.: Application of the strain invariant failure theory (sift) to metals and fiber-polymer composites. Philosophical Magazine 90(31-32), 4263-4331 (2010)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.678, + 0.785, + 0.704 + ], + "angle": 0, + "content": "13. Hess, R., Baker, C., Zihl, J.: The\" motion-blind\" patient: low-level spatial and temporal filters. Journal of Neuroscience 9(5), 1628-1640 (1989)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.705, + 0.785, + 0.745 + ], + "angle": 0, + "content": "14. Jin, H., Liu, I., Xu, P., Zhang, X., Han, S., Bi, S., Zhou, X., Xu, Z., Su, H.: Tensoroir: Tensorial inverse rendering. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.746, + 0.785, + 0.785 + ], + "angle": 0, + "content": "15. Kajiya, J.T., Von Herzen, B.P.: Ray tracing volume densities. In: Proceedings of the 11th Annual Conference on Computer Graphics and Interactive Techniques. p. 165-174 (1984)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.786, + 0.785, + 0.84 + ], + "angle": 0, + "content": "16. Li, T., Slavcheva, M., Zollhoefer, M., Green, S., Lassner, C., Kim, C., Schmidt, T., Lovegrove, S., Goesele, M., Newcombe, R., et al.: Neural 3d video synthesis from multi-view video. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5521-5531 (2022)" + }, + { + "type": "list", + "bbox": [ + 0.22, + 0.173, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.356, + 0.127 + ], + "angle": 0, + "content": "W. Im et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.787, + 0.203 + ], + "angle": 0, + "content": "17. Li, Z., Müller, T., Evans, A., Taylor, R.H., Unberath, M., Liu, M.Y., Lin, C.H.: Neuralangelo: High-fidelity neural surface reconstruction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8456-8465 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.204, + 0.787, + 0.245 + ], + "angle": 0, + "content": "18. Li, Z., Niklaus, S., Snavely, N., Wang, O.: Neural scene flow fields for space-time view synthesis of dynamic scenes. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6498-6508 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.246, + 0.786, + 0.286 + ], + "angle": 0, + "content": "19. Li, Z., Wang, Q., Cole, F., Tucker, R., Snavely, N.: Dynibar: Neural dynamic image-based rendering. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4273-4284 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.287, + 0.787, + 0.342 + ], + "angle": 0, + "content": "20. Liu, Y.L., Gao, C., Meuleman, A., Tseng, H.Y., Saraf, A., Kim, C., Chuang, Y.Y., Kopf, J., Huang, J.B.: Robust dynamic radiance fields. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 13-23 (June 2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.343, + 0.787, + 0.37 + ], + "angle": 0, + "content": "21. Luo, X., Huang, J.B., Szeliski, R., Matzen, K., Kopf, J.: Consistent video depth estimation. ACM Transactions on Graphics (ToG) 39(4), 71-1 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.371, + 0.787, + 0.411 + ], + "angle": 0, + "content": "22. Max, N., Chen, M.: Local and global illumination in the volume rendering integral. Tech. rep., Lawrence Livermore National Lab.(LLNL), Livermore, CA (United States) (2005)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.411, + 0.787, + 0.452 + ], + "angle": 0, + "content": "23. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. In: ECCV (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.453, + 0.787, + 0.508 + ], + "angle": 0, + "content": "24. Orts-Escalano, S., Rhemann, C., Fanello, S., Chang, W., Kowdle, A., Degtyarev, Y., Kim, D., Davidson, P.L., Khamis, S., Dou, M., et al.: Holoportation: Virtual 3d teleportation in real-time. In: Proceedings of the 29th annual symposium on user interface software and technology. pp. 741-754 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.509, + 0.787, + 0.55 + ], + "angle": 0, + "content": "25. Park, K., Sinha, U., Barron, J.T., Bouaziz, S., Goldman, D.B., Seitz, S.M., Martin-Brualla, R.: Nerfies: Deformable neural radiance fields. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5865-5874 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.55, + 0.787, + 0.591 + ], + "angle": 0, + "content": "26. Park, K., Sinha, U., Hedman, P., Barron, J.T., Bouaziz, S., Goldman, D.B., Martin-Brualla, R., Seitz, S.M.: Hypernerf: A higher-dimensional representation for topologically varying neural radiance fields. arXiv preprint arXiv:2106.13228 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.592, + 0.787, + 0.633 + ], + "angle": 0, + "content": "27. Pumarola, A., Corona, E., Pons-Moll, G., Moreno-Noguer, F.: D-nerf: Neural radiance fields for dynamic scenes. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10318-10327 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.633, + 0.787, + 0.674 + ], + "angle": 0, + "content": "28. Ramasinghe, S., Shevchenko, V., Avraham, G., Hengel, A.V.D.: Bali-rf: Bandlimited radiance fields for dynamic scene modeling. arXiv preprint arXiv:2302.13543 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.675, + 0.787, + 0.715 + ], + "angle": 0, + "content": "29. Soria, J., Sondergaard, R., Cantwell, B., Chong, M., Perry, A.: A study of the fine-scale motions of incompressible time-developing mixing layers. Physics of Fluids 6(2), 871-884 (1994)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.716, + 0.787, + 0.757 + ], + "angle": 0, + "content": "30. Sun, D., Roth, S., Black, M.J.: A quantitative analysis of current practices in optical flow estimation and the principles behind them. International Journal of Computer Vision 106, 115-137 (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.758, + 0.787, + 0.813 + ], + "angle": 0, + "content": "31. Tretschk, E., Tewari, A., Golyanik, V., Zollhöfer, M., Lassner, C., Theobalt, C.: Non-rigid neural radiance fields: Reconstruction and novel view synthesis of a dynamic scene from monocular video. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 12959-12970 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.813, + 0.787, + 0.84 + ], + "angle": 0, + "content": "32. Verbin, D., Hedman, P., Mildenhall, B., Zickler, T., Barron, J.T., Srinivasan, P.P.: Ref-nerf: Structured view-dependent appearance for neural radiance fields. In: 2022" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.322, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Regularizing Dynamic Radiance Fields with Kinematic Fields" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.245, + 0.147, + 0.785, + 0.175 + ], + "angle": 0, + "content": "IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 5481-5490. IEEE (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.176, + 0.785, + 0.203 + ], + "angle": 0, + "content": "33. Wang, C., Eckart, B., Lucey, S., Gallo, O.: Neural trajectory fields for dynamic novel view synthesis. arXiv preprint arXiv:2105.05994 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.204, + 0.785, + 0.231 + ], + "angle": 0, + "content": "34. Wang, C., MacDonald, L.E., Jeni, L.A., Lucey, S.: Flow supervision for deformable nerf. arXiv preprint arXiv:2303.16333 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.232, + 0.785, + 0.272 + ], + "angle": 0, + "content": "35. Wu, T., Zhong, F., Tagliasacchi, A., Cole, F., Oztireli, C.: D\\~ 2nerf: Self-supervised decoupling of dynamic and static objects from a monocular video. Advances in Neural Information Processing Systems 35, 32653-32666 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.273, + 0.785, + 0.327 + ], + "angle": 0, + "content": "36. Yoon, J.S., Kim, K., Gallo, O., Park, H.S., Kautz, J.: Novel view synthesis of dynamic scenes with globally coherent depths from a monocular camera. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5336-5345 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.328, + 0.785, + 0.356 + ], + "angle": 0, + "content": "37. Zhang, K., Riegler, G., Snavely, N., Koltun, V.: Nerf++: Analyzing and improving neural radiance fields. arXiv preprint arXiv:2010.07492 (2020)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.356 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/Regularizing Dynamic Radiance Fields with Kinematic Fields/79573c22-a91b-4733-9726-bd12faf03469_origin.pdf b/2024/Regularizing Dynamic Radiance Fields with Kinematic Fields/79573c22-a91b-4733-9726-bd12faf03469_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..083cf68ad0cc9da7c75ea909fb7d095db342a295 --- /dev/null +++ b/2024/Regularizing Dynamic Radiance Fields with Kinematic Fields/79573c22-a91b-4733-9726-bd12faf03469_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4578b95e6ccb258ced3aa76d88d01c81c3adfa657e6b08857790ebb94dde2401 +size 8475435 diff --git a/2024/Regularizing Dynamic Radiance Fields with Kinematic Fields/full.md b/2024/Regularizing Dynamic Radiance Fields with Kinematic Fields/full.md new file mode 100644 index 0000000000000000000000000000000000000000..53dc1d3ac687c1c06914b574a6847ded7798374f --- /dev/null +++ b/2024/Regularizing Dynamic Radiance Fields with Kinematic Fields/full.md @@ -0,0 +1,332 @@ +# Regularizing Dynamic Radiance Fields with Kinematic Fields + +Woobin $\mathrm{Im}^{\star 1,2}$ , Geonho Cha $^3$ , Sebin Lee $^1$ , Jumin Lee $^1$ , Juhyeong Seon $^1$ , Dongyoon Wee $^3$ , and Sung-Eui Yoon $^1$ + +$^{1}$ KAIST {iwbn,seb.lee,jmlee,munuwazzi,sungeui}@kaist.ac.kr + +$^{2}$ Samsung Advanced Institute of Technology (SAIT) + +3 Naver Cloud + +{geonho.cha,dongyoon.wei}@navercorp.com + +Abstract. This paper presents a novel approach for reconstructing dynamic radiance fields from monocular videos. We integrate kinematics with dynamic radiance fields, bridging the gap between the sparse nature of monocular videos and the real-world physics. Our method introduces the kinematic field, capturing motion through kinematic quantities: velocity, acceleration, and jerk. The kinematic field is jointly learned with the dynamic radiance field by minimizing the photometric loss without motion ground truth. We further augment our method with physics-driven regularizers grounded in kinematics. We propose physics-driven regularizers that ensure the physical validity of predicted kinematic quantities, including advective acceleration and jerk. Additionally, we control the motion trajectory based on rigidity equations formed with the predicted kinematic quantities. In experiments, our method outperforms the state-of-the-arts by capturing physical motion patterns within challenging real-world monocular videos. + +Keywords: Dynamic radiance fields $\cdot$ Kinematic fields $\cdot$ 3D video + +# 1 Introduction + +Recognizing motion is crucial to how we visually perceive the world [13], and video is the medium that carries the motion to viewers in a remote place or at a different time. Through the lens of video, we watch and enjoy movies, user-generated contents, and remote video chats. While videos depict motion in a 2D format, transforming them into three-dimensional experiences can elevate our perception of the content to co-presence via applications such as teleportation [24]. + +Dynamic radiance fields [5,9,11,16,19,27] have been studied as a promising representation for reconstructing 3D dynamic scenes; it enables realistic novel view and time synthesis, which can upgrade 2D videos into 3D videos. Since multi-view cameras can give more information about the geometry in a scene, + +existing works have utilized camera rigs with dozens of cameras to capture multi-view video datasets [1, 5, 16]. Although high-quality multi-view videos result in undeniably better quality than monocular videos do, the multi-view capturing devices and environments are prohibitive in most cases. + +Thus, there have been efforts to reconstruct dynamic radiance fields from monocular videos [10, 18, 20]. Compared to multi-view videos, monocular videos provide sparser information, which makes the reconstruction more challenging. For instance, there is scale ambiguity in monocular video, making it difficult to accurately locate an object within a space [34]. Thus, dynamic scene reconstruction from monocular videos has been considered an ill-posed problem. + +In this paper, we incorporate kinematics to fill the gap between sparse monocular videos and real-world physics, potentially enhancing novel view and time synthesis. Our core idea is that motion within the real-world is governed by the principles of kinematics in physics, implying that representations of the world, such as videos, are inherently anchored in kinematics. Building on this foundation, we generalize conventional displacement fields – often called scene-flow fields [18] – to kinematic fields. In prior work, motion in radiance fields has been controlled via the regularization in the frequency domain [28, 33], displacement fields [18], and deformation fields [25]. + +Different from these approaches, our kinematic field regulates motion in radiance fields using kinematics; it innately creates continuous trajectories and enables us to impose regularization based on kinematics. We integrate physics-based regularizers to make the kinematic field physically plausible and thus improve the dynamic radiance field accordingly. For instance, the rigidity regularizer based on the first and the second invariant of the strain rate is applied on the kinematic field, suppressing non-rigid motion in the field. We quantitatively and qualitatively validate the proposed regularizers on the NVIDIA Dynamic Scenes (NDVS) dataset, showing the effectiveness of our method over the existing models. Lastly, the kinematic field uncovers the kinematics within a scene without a ground truth; it estimates not only point-wise 3D displacement within a scene, but also underlying velocity, acceleration, and jerk. + +# 2 Related Work + +Neural radiance fields. In recent years, neural radiance fields [4,23,32] have emerged as a promising approach for learning an implicit neural representation of a 3D scene. Combined with the differentiable volume rendering technique, it has enabled realistic novel-view synthesis [4,23,32,37]. + +In essence, a radiance field $\mathcal{F}_{\mathrm{ST}}$ maps a 5D vector $(x,y,z,\theta ,\phi)$ to a 4D output vector $(\mathbf{c}_s,\sigma_s)$ : + +$$ +\mathcal {F} _ {\mathrm {S T}}: (x, y, z, \theta , \phi) \rightarrow (\mathbf {c} _ {s}, \sigma_ {s}), \tag {1} +$$ + +where the view-dependent color $\mathbf{c}_s = (r,g,b)$ and density $\sigma_s$ are learned in relation to the spatial position $(x,y,z)$ and the viewing direction $(\theta ,\phi)$ . With the volumetric representation learned via an iterative optimization, volume rendering techniques [15, 22, 23] are employed to aggregate colors and densities along + +a cast ray. Consequently, this leads to realistic novel-view synthesis from the inferred radiance, after being trained on dozens of images. + +Dynamic radiance fields. Radiance fields also facilitate the novel-view synthesis of scenes with moving objects [1,5,16,25,26]. While this is more challenging than the static case, dynamic radiance fields incorporating motion allow for the rendering of a frame at a desired time and viewpoint. As a result, it can be utilized to make visual effects such as bullet time, stabilized, dolly zoomed, and stereo videos. Commonly, the time-dependent radiance fields are referred to as dynamic radiance fields, a term we will adopt throughout this paper. + +In dynamic radiance fields, radiance is dependent on a varying time $t$ . This modifies Eq. 1 by introducing a time-dependent term, resulting in: + +$$ +\mathcal {F} _ {\mathrm {D Y}}: (x, y, z, t, \theta , \phi) \rightarrow (\mathbf {c} _ {d}, \sigma_ {d}), \tag {2} +$$ + +where the color and density are conditioned on time $t$ . In this formulation, moving geometries of objects (e.g., translation and deformation) and changing radiance (e.g., moving shadows) can be learned via optimization. + +Learning a dynamic representation presents inherent challenges due to the introduction of temporal dimension. To optimize this representation, a high-quality multi-view dataset can be employed to obtain fine rendering results [1, 5, 16]. Meanwhile, realistically captured imagery of a scene tends to be nonsynchronized and is likely to be obtained with a limited number of cameras and restricted displacements; this often results in a monocular setup [11, 18, 25, 26]. Previous researches have delved into deformable models for monocular videos [20, 25, 26, 31]. In these methods, a deformation field maps a spatio-temporal coordinate to a reference canonical space. While the canonical space aids in maintaining radiance consistency across global time frames, it tends to exhibit reduced effectiveness in generic monocular videos, where new objects might appear within a video [28]. In this study, we embrace the concept of scene flow fields [18], which is constructed without canonical space and deformation networks. + +Regularization with kinematic priors. Monocular videos present challenges in reconstruction due to the constrained viewpoints of moving objects. To address this, integrating auxiliary flow fields has proven to be effective [8, 10, 18, 19]. Existing studies employ either scene flow fields [10, 18, 19] or velocity fields [8] to capture spatial displacements between consecutive frames. Since the scene flows can give spatial correspondences of the dynamic radiance field, it can be used to promote photometric consistency between the renderings of a given ray and its time-deformed counterpart. + +Physical priors incorporated with scene flow fields play a pivotal role in regulating the radiance field. Regularization of motion has been applied to the dynamic field, bridging the gap between the sparse visual data and plausible real-world motion. For instance, a smoothness regularizer ensures that the displacements within close proximity are similar to each other, and kinetic energy regularizer promotes consistency between forward and backward flows, resulting in smoother displacement fields [18]. More deeply integrated with physics, the elastic energy regularizer minimizes the deformation between the canonical space + +![](images/5d61a5cc06096f3d2093cbc73e139d1cfd6ef4d69d9cb995e6080f498377603b.jpg) +Fig. 1: Radiance and kinematic fields. This figure summarizes the three fields our method utilizes. The static and dynamic radiance fields are used for rendering, and the kinematic field is used in the training phase, regularizing the dynamic radiance field. + +and the deformed space [25]. In fluid reconstruction works [2, 7], velocity fields grounded in physics have been shown to be critical in accurate 4D reconstruction. For instance, Navier-Stokes and transport equations can be integrated in reconstructing smoke [7] and fluid surfaces [2] with dynamic radiance fields to reconstruct more physically plausible structures. + +We integrate kinematics to reconstruct dynamic radiance fields from generic monocular videos. We introduce the kinematic field, which captures motion through kinematic quantities such as velocity, acceleration, and jerk. Different from existing velocity field-based methods [7,8,34], our method has higher-order terms (i.e., acceleration and jerk) as latent of spatio-temporal correspondences. Further enhancing our approach, we incorporate physics-driven regularizers based on a transport equation [7] and the strain rate tensor, addressing dynamics of moving particles and rigid motion. + +# 3 Method + +# 3.1 Overview + +Our system takes as input a video sequence, denoted as $(I_i, t_i, P_i)_{i=1}^N$ , where $N$ is the number of frames. For each frame, we have an image $I_i \in \mathbb{R}^{H \times W \times 3}$ , a timestamp $t_i \in \mathbb{R}$ , and a precomputed camera pose $P_i \in SE(3)$ and intrinsic matrix $K_i \in \mathbb{R}^{3 \times 3}$ . From this input, our primary objective is to synthesize an image, denoted as $\hat{I}_k$ , corresponding to a novel time and viewpoint $(t_k, P_k, K_k)$ . The parameters $t_k$ , $P_k$ , and $K_k$ are not restricted to those found in the original video input, allowing our system to render images from previously unobserved times and viewpoints. + +To achieve this image synthesis for novel times and views, we leverage the radiance-field-based volume rendering [23]. Inspired by existing methods [18,35], we train two distinct radiance fields (Fig. 1): a dynamic radiance field, denoted + +as $\mathcal{F}_{\mathrm{DY}}$ (Eq. 2), and a static radiance field, denoted as $\mathcal{F}_{\mathrm{ST}}$ (Eq. 1). To blend the two fields, we utilize a technique from D²-NeRF [35], where rendering weights for dynamic and static samples are derived from the volume densities. + +Addressing the challenge of achieving physically plausible spatio-temporal structures, we introduce the novel kinematic field. The kinematic field accounts for the kinematic quantities (e.g., velocity and acceleration). These quantities can be converted into displacement (Eq. 8), enabling us to use the photometric consistency loss (Eq. 14) for spatio-temporal regularization, as used in previous work [18]. With these kinematic estimations, we apply physics-based regularization techniques. These techniques involve density transport (Eq. 12) and motion rigidity through the strain rate tensor (Eq. 11). Combining the photometric loss (Eq. 14) with these techniques enhances the reconstruction of radiance and geometry, leading to improved novel time and view synthesis performance from underdetermined data such as a monocular video. + +# 3.2 Dynamic and Static Radiance Fields + +In monocular dynamic radiance field methods, it is common to decompose the radiance fields into dynamic and static fields [11, 18, 20] to prevent undesirable motion in the static area. We employ tensor-decomposition-based representations [5, 6] for both dynamic and static fields. Specifically, for the dynamic radiance field $\mathcal{F}_{\mathrm{DY}}$ , we use HexPlane [5], and for the static radiance field $\mathcal{F}_{\mathrm{ST}}$ , we use TensoRF [6]. Further details on the tensor decomposition can be found in the supplementary material. + +Definition. The dynamic and static radiance fields are defined in Eq. 1-2 ( $\mathcal{F}_{\mathrm{ST}}$ and $\mathcal{F}_{\mathrm{DY}}$ ). Here, the inputs correspond to the position $(x,y,z)$ , the timestamp $t$ , and the viewing direction $(\theta, \phi)$ from the camera origin. Each field yields the color $\mathbf{c}$ and the volume density $\sigma$ . Note that the static field is independent of time $t$ , ensuring a consistent structure throughout the whole time span. + +Composite rendering. We use a composite rendering method to obtain the final rendering with the two radiance fields. It is to be noted that individual components (i.e., dynamic-only and static-only) can be rendered using the standard volume rendering technique [23]. To combine both fields, we employ an additive composition method in which each field influences the transmittance of a ray [35]. Let's consider a camera ray $\mathbf{r}(\tau) = \mathbf{o} + \tau \mathbf{u}$ , where $\mathbf{o}$ is the origin and $\mathbf{u}$ is the direction. With a predefined range for the ray distance $[\tau_n, \tau_f]$ , the rendering function is formulated as: + +$$ +\hat {C} (\mathbf {r}, t _ {i}) = \int_ {\tau_ {n}} ^ {\tau_ {f}} T (\tau) \left(\sigma_ {s} (\tau) \cdot \mathbf {c} _ {s} (\tau) + \sigma_ {d} (\tau , t _ {i}) \cdot \mathbf {c} _ {d} (\tau , t _ {i})\right) d \tau , \tag {3} +$$ + +$$ +\text {w h e r e} T (\tau) = \exp \left(- \int_ {\tau_ {n}} ^ {\tau} \left(\sigma_ {s} (u) + \sigma_ {d} (u, t _ {i})\right) d u\right). \tag {4} +$$ + +Here, each notation for the sample point is simplified for concise representation (e.g., $\sigma_s(\tau) = \sigma_s(\mathbf{r}(\tau))$ ). We visualize the example of the composite rendering and separate rendering in the first and second rows in Fig. 2. + +![](images/00fbf48959c2f2140a078b90b27309e78b499fe24e1e1da8d7a13a664a28ca19.jpg) +Fig. 2: Visualization of each predicted component. The top row presents synthesized views at different times. In (a) and (b), the dynamic and static components of the scene at time $t_2$ are depicted separately. (b) In the case of an inobservable static area in the entire sequence (e.g., the space behind the jumping people), radiance might not be correct. The displacement field (c) can be computed by Taylor approximation (Eq. 8) with motion fields (d-f): velocity, acceleration, and jerk. Each field is visualized through reprojecting each field to the camera view. The standard HSV visualization [3] is used to colorize arrows. + +Dynamic and static separation. To factorize a scene into static and dynamic parts, we use a metric indicating the likelihood that a point belongs to the dynamic fields [35]. This metric is given by: + +$$ +p _ {d} = \frac {\sigma_ {d}}{\sigma_ {d} + \sigma_ {s}}, \tag {5} +$$ + +where $\sigma_{d}$ and $\sigma_{s}$ are density values sampled from a spatio-temporal point. + +To facilitate the separation of the radiance fields, we employ the skewed binary entropy loss, as proposed by [35]. The entropy loss is defined as: + +$$ +\mathcal {L} _ {b} = - \left(p _ {d} ^ {k} \cdot \log \left(p _ {d} ^ {k}\right) + \left(1 - p _ {d} ^ {k}\right) \cdot \log \left(1 - p _ {d} ^ {k}\right)\right), \tag {6} +$$ + +where the exponent $k \geq 1$ is the skew factor, controlling a bias toward static explanation. This entropy-based loss can encourage the radiance fields to have a clearer distinction between the static and dynamic components. + +# 3.3 Kinematic Field + +The kinematic field has a spatio-temporal structure to facilitate spatial and temporal queries. We use the tensor-decomposition-based representation [5] for the kinematic field. As depicted in Fig. 1, the kinematic field serves to regulate + +![](images/ddb5f4f6daa9e7693eb568e64ac48585a260925bf180acb76283f19bda938471.jpg) +(a) Without kinematic regularization + +![](images/5d14acfbe7ef51cf0b59c6b338cc36d26cd2c12e21bcb29cccd19e62ab2a8a6f.jpg) +(b) With kinematic regularization +Fig. 3: Effect of kinematic regularization. We visualize the rendered RGB and motion of each field. Without kinematic regularization, motion fields tend to show granular patterns. Our kinematic fields not only make the field smoother but also satisfy the kinematic property, i.e., $\mathbf{a} = \partial \mathbf{v} / \partial t + \mathbf{v} \cdot \nabla \mathbf{v}$ . We abbreviate the advective equation to 'Adv.' in the figure. + +the radiance field, ensuring that the motion in the radiance field is physically plausible. + +Definition. The kinematic field is defined as: + +$$ +\mathcal {F} _ {\mathrm {K}}: (x, y, z, t) \rightarrow (\mathbf {v}, \mathbf {a}, \mathbf {j}, \dots), \tag {7} +$$ + +where each vector represents velocity, acceleration, and jerk, such that $\mathbf{v} = (v_x, v_y, v_z)$ , $\mathbf{a} = (a_x, a_y, a_z)$ , and $\mathbf{j} = (j_x, j_y, j_z)$ . Note that we can extend the field into higher-order quantities such as snap (4th), crackle (5th), and so on. + +Displacement. Using the provided kinematic quantities, we can approximate a particle's displacement at position $(x,y,z)$ and time $t + \Delta t$ via the Taylor approximation, assuming that the kinematic quantities are consistent within a local temporal range $\Delta t$ . The displacement $d(x,y,z,t,\Delta t)$ can be approximated by: + +$$ +d (x, y, z, t, \Delta t) \approx \frac {(\Delta t) \mathbf {v}}{1 !} + \frac {(\Delta t) ^ {2} \mathbf {a}}{2 !} + \frac {(\Delta t) ^ {3} \mathbf {j}}{3 !} + \dots . \tag {8} +$$ + +Computing the displacement of a point within a temporal range allows for computing the photometric consistency loss (Eq. 14) and scene flows from the kinematic field. + +# 3.4 Kinematic Regularization + +Given the kinematic field, it is important to make sure the learned motion dynamics are physically plausible. As Fig. 3 indicates, learning kinematic fields without adequate regularization may result in a sub-optimal kinematic field. The subsequent sections elucidate regularization strategies. + +Integrity of kinematic fields. Learning the kinematic field $\mathcal{F}_{\mathrm{K}}$ naively does not guarantee the integrity of the kinematic quantities. During training, the quantities are supervised via the photometric consistency loss (Eq. 14), which is applied upon the approximated displacement (Eq. 8). It is important to note that even with supervision by the displacement, the physical relationships among the kinematic quantities may not be correct, since the higher-order terms in the displacement equation (Eq. 8) make the system underdetermined. + +Given a velocity field, we can derive the acceleration field using the principle of advective acceleration. Given acceleration $\mathbf{a}$ and velocity $\mathbf{v}$ at a spatiotemporal coordinate, the partial derivative and advection yield $\mathbf{a} = \partial \mathbf{v} / \partial t + \mathbf{v} \cdot \nabla \mathbf{v}$ . Similarly, we can formulate the relationship between $\mathbf{v}$ , $\mathbf{a}$ and $\mathbf{j}$ as: $\mathbf{j} = \partial \mathbf{a} / \partial t + \mathbf{v} \cdot \nabla \mathbf{a}$ . + +To ensure kinematic consistency, we introduce the kinematic loss $\mathcal{L}_{\mathrm{K}}$ defined as: + +$$ +\mathcal {L} _ {\mathrm {K}} = \left\| \mathbf {a} - \frac {\partial \mathbf {v}}{\partial t} - \mathbf {v} \cdot \nabla \mathbf {v} \right\| _ {2} ^ {2} + \left\| \mathbf {j} - \frac {\partial \mathbf {a}}{\partial t} - \mathbf {v} \cdot \nabla \mathbf {a} \right\| _ {2} ^ {2} + \dots , \tag {9} +$$ + +where $\nabla$ is the Jacobian operation. To compute the partial derivatives and Jacobian matrices, we use a numerical method as presented in a neural surface reconstruction work [17]. + +Physics-informed regularization. In the context of regularizing radiance fields through the use of an auxiliary kinematic field, it is important to acknowledge the highly underdetermined nature of the problem. Essentially, there exists a broad spectrum of possible potential kinematic field solutions that could perfectly fit the observed training views. This complexity necessitates the introduction of a regularization strategy that can guide the motion representation toward more accurately reflecting real-world physics. + +While not universally ideal, rigidity [25] and incompressibility [2,7] have been identified as effective metrics for encapsulating real-world dynamics; in videos that we normally encounter, objects in a scene have nearly consistent density and are mostly rigid in a short time range. + +Rigidity. One benefit of having a velocity field is that we can measure the rate of deformation by its derivatives. A strain rate tensor $D$ defines a kinematic property of a deformation; it captures the rate at which a material undergoes deformation, and it is crucial for ensuring that objects maintain their structural integrity during motion. + +In this paper, we propose using the invariants of strain rate tensor [12, 29] as a measure of rigidity. The velocity gradient $\nabla \mathbf{v}$ sampled from the kinematic field, leads to the strain rate tensor $D$ by: + +$$ +D = \frac {1}{2} (\nabla \mathbf {v} + (\nabla \mathbf {v}) ^ {T}). \tag {10} +$$ + +The rigidity loss function $\mathcal{L}_R$ is formulated based on the first and second invariants of the strain rate tensor: + +$$ +\mathcal {L} _ {R} = \underbrace {\lambda_ {\mathrm {d i v}} (\nabla \cdot \mathbf {v}) ^ {2}} _ {\text {1 s t i n v a r i a n t}} + \underbrace {\left(\frac {1}{2} (\operatorname {t r} (D) ^ {2} - \operatorname {t r} (D \cdot D)))\right) ^ {2}} _ {\text {2 n d i n v a r i a n t}}, \tag {11} +$$ + +where $\operatorname{tr}(\cdot)$ denotes the trace operator and $\lambda_{\mathrm{div}}$ controls the balance between the two terms. Note that the first invariant of strain rate measures dilational change (i.e., volumetric change) with respect to time and the second invariant of strain rate measures distortional change. + +![](images/1551b001fadd2ddc6de0d6978ba34f979017653f9dfb6947b58194b66e89d8e6.jpg) +Fig.4: Visualization of density variation by a spatial coordinate $x$ . The plot displays the gradient $-\partial \mathcal{L}_{\mathrm{T}} / \partial v_{\sigma_d}$ with arrows. With the velocity field directed to the right (i.e., $\mathbf{v}_x = 0.3$ ), we can compute the gradient of the transport regularization $\mathcal{L}_{\mathrm{T}}$ w.r.t. the rate of density change $v_{\sigma_d} = \partial \sigma_d / \partial t$ . Minimizing $\mathcal{L}_{\mathrm{T}}$ allows us to render the density at $t + \Delta t$ aligned with the flow field. + +![](images/cb83f9a1a3f40142d3ed601b8bede8889a07f14c8d49fafd0a30f20542b6983f.jpg) +Fig. 5: Photometric consistency loss. During optimization, we deform each ray from a reference time and pose $(t,P_{t})$ to different timestamps $\gamma$ and $i$ . Here, we sample $i$ from the neighboring frame times, and we consequently sample an intermediate timestamp $\gamma \sim \mathcal{U}(t,i)$ . Given each timestamp, we can deform a ray using Eq. 8. The photometric loss $\mathcal{L}_{\mathrm{photo}}$ is computed based on the color consistency between the deformed ray and the original color, enhancing temporal consistency. + +Transport. A transport equation has been adopted in physics-informed neural networks [7]. By simplifying the Navier-Stokes equation, a transport loss can be defined as: $\frac{\partial\sigma}{\partial t} +\mathbf{v}\cdot \nabla \sigma = 0$ . We leverage the physics-informed regularizers based on the transport equation and the divergence of the velocity: + +$$ +\mathcal {L} _ {\mathrm {T}} = \left(\frac {\partial \sigma_ {d}}{\partial t} + \mathbf {v} \cdot \nabla \sigma_ {d}\right) ^ {2}, \tag {12} +$$ + +where $\sigma_{d}$ is the volume density used for volume rendering. This aligns the velocity field with the density field (i.e., Fig. 4), preventing a sudden disappearance or appearance of objects in the dynamic radiance field. + +# 3.5 Learning Objectives + +Cycle consistency. The cycle consistency is essential to ensure that the trajectories derived from the kinematic field do not introduce misalignments [18]. We use cycle constraints among four timestamps $t$ , $i$ , $j$ , and $\gamma$ . Given a reference time $t$ , we randomly sample time $i$ from the neighboring frames within a threshold. We define $j = t + 2(i - t)$ , making it twice the distance from $t$ as $i$ is from $t$ . $\gamma \sim \mathcal{U}(t,i)$ is an intermediate time, uniformly sampled between $t$ and $i$ . With the time variables, we formulate our loss: + +$$ +\mathcal {L} _ {\mathrm {C}} = \rho (\mathbf {d} _ {t \rightarrow i} + \mathbf {d} _ {i \rightarrow t}) + \rho (\mathbf {d} _ {t \rightarrow i \rightarrow j} - \mathbf {d} _ {t \rightarrow j}) + \rho (\mathbf {d} _ {t \rightarrow \gamma \rightarrow i} - \mathbf {d} _ {t \rightarrow i}), \tag {13} +$$ + +where $\rho$ denotes the generalized chabonnier loss [30]. Here, we abbreviate the displacement $d(\cdot)$ (Eq. 8) for conciseness: $\mathbf{d}_{t\to i}$ is shorthand for $d(\mathbf{x},t,i - t)$ , and $\mathbf{d}_{t\to i\to j}$ simplifies $d(\mathbf{x} + \mathbf{d}_{t\to i},i,j - i)$ . + +Photometric consistency loss. To enhance temporal coherency, we deform rays into nearby frame times and enhance the consistency among them (Fig. 5). Specifically, given a reference ray $\mathbf{r}_t$ at time $t$ , a pixel color $\hat{C}(\mathbf{r}_t, t)$ can be obtained using the volume rendering (Eq. 4). In addition to the reference ray, we apply ray deformation based on the predicted kinematic quantities. As in cycle consistency (Eq. 13), we sample time $i$ from neighboring frames, and time $\gamma \sim \mathcal{U}(t, i)$ between times $t$ and $i$ . Different from existing approaches [10, 18], we deform rays with kinematic quantities using Taylor approximation. The deforming process results in two deformed rays $\mathbf{r}_{t \to i}$ and $\mathbf{r}_{t \to \gamma}$ , and two rendered colors $\hat{C}(\mathbf{r}_{t \to i}, i)$ and $\hat{C}(\mathbf{r}_{t \to \gamma}, \gamma)$ , correspondingly. Based on the deformed rays, we apply the photometric consistency loss, comparing the rendered color with the ground-truth color. The loss function is formulated as: + +$$ +\begin{array}{l} \mathcal {L} _ {\text {p h o t o}} = \| \hat {C} (\mathbf {r} _ {t}, t) - C (\mathbf {r} _ {t}, t) \| _ {2} ^ {2} \tag {14} \\ + \lambda \alpha (\mathbf {r} _ {t}) \rho (\hat {C} (\mathbf {r} _ {t \rightarrow i}, i) - C (\mathbf {r} _ {t}, t)) + \lambda \alpha (\mathbf {r} _ {t}) \rho (\hat {C} (\mathbf {r} _ {t \rightarrow \gamma}, \gamma) - C (\mathbf {r} _ {t}, t)), \\ \end{array} +$$ + +where $\lambda$ is a loss weight for the temporal consistency, and $\alpha (\mathbf{r}_t)$ is the sum of rendering weights of the dynamic densities, representing the dynamic likelihood of the ray $\mathbf{r}_t$ . + +Final objective. Our final objective function, denoted as $\mathcal{L}$ , is the weighted sum of the various loss terms discussed previously. To keep things simple with the variety of loss functions, we symbolize the losses to $\mathcal{L}_{\mathrm{photo}}$ , $\mathcal{L}_{\mathrm{kinematic}}$ , and $\mathcal{L}_{\mathrm{reg}}$ . This results in the final loss: + +$$ +\mathcal {L} = \mathbb {E} _ {\mathbf {r} _ {t}} \left[ \mathcal {L} _ {\text {p h o t o}} + \mathcal {L} _ {\text {k i n e m a t i c}} + \mathcal {L} _ {\text {r e g}} \right]. \tag {15} +$$ + +Here, $\mathcal{L}_{\mathrm{kinematic}}$ is the weighted sum of the kinematic integrity $\mathcal{L}_{\mathrm{K}}$ , transport $\mathcal{L}_{\mathrm{T}}$ , and rigidity $\mathcal{L}_{\mathrm{R}}$ . Note that $\mathbb{E}_{\mathbf{r}_t}[\mathcal{L}_{\mathrm{kinematic}}]$ is computed with randomly sampled points along the ray $\mathbf{r}_t$ from each minibatch. The additional regularization $\mathcal{L}_{\mathrm{reg}}$ includes the skewed entropy $\mathcal{L}_b$ , the cycle consistency $\mathcal{L}_{\mathrm{C}}$ , total-variation loss [5,6], distortion loss [4], monocular depth supervision [18], normal regularization [14,32] and optical flow supervision [18]. + +# 4 Experiment + +Dataset. We use the NVIDIA dynamic view synthesis dataset (NDVS) [36] to evaluate the effectiveness of our model. For the monocular setup, we adopt the preprocessing steps proposed by NSFF [18] and DynamicNeRF [10]. This dataset has 8 scenes and is captured with 12 synchronized cameras. When used for a monocular video reconstruction task, the dataset shows relatively low effective multi-view factors (EMFs) [11], compared to other monocular datasets; this means that the dataset is more challenging to reconstruct due to the less camera motion and faster-moving objects. + +There are two widely used monocular NDVS protocols: 12-frames [10] and 24-frames [18]. Most papers opt for either protocol; however, we include both protocols for rigorous evaluation. Meanwhile, 24-frames-sparse is proposed by us, as neither existing protocol provides novel time evaluation; we leave out odd-numbered frames at unseen times for testing. For the 'Dynamic Only' marks in tables, we use the dynamic mask to evaluate only in terms of dynamic pixels; please refer to [10,18] for how to generate dynamic masks. We detail the protocols below. + +- 24-frames: this sequence is configured with 24 training frames, and 264 test frames from unseen multi-views. For a fair comparison, we utilize the dynamic mask, optical flow, and monocular depth, extracted by the same model with the prior work [18]. +- 24-frames-sparse: we reduce the 24-frames sequence into 12 frames by sampling the even-numbered frames for training. This sparse sampling allows for testing novel view synthesis at both seen times and novel times. +- 12-frames: this sequence has 12 training and 12 testing samples. We use the preprocessed dynamic mask, optical flow, and monocular depth downloadable from the code repository of DynamicNeRF [10]. + +# 4.1 Implementation Details + +Fields configuration. Our radiance and kinematic fields are configured with HexPlane [5] and TensoRF [6]. After features are sampled from the point queries, small MLP decoders process the features to yield colors and kinematic outputs. For density outputs, we use a linear layer following the practices in [5,6]. Further details on field configurations are available in the supplementary material. + +Training. We use 70,000 steps to optimize our models, which takes about 7 hours with one NVIDIA V100 or A100 GPU. We initialize the static field for 5,000 steps, taking approximately 5 minutes. After finishing the initialization of the static field, we jointly train the fields: dynamic, static, and kinematic fields. During the joint training, we lower the learning rate of the static field by the factor of 0.1 at predefined iterations to prevent dynamic objects from being portrayed as static objects. Additional training details are provided in the supplementary material. + +Evaluation metrics. We report peak signal-to-noise ratio (PSNR), structural similarity (SSIM), and perceptual similarity (LPIPS). These three metrics have been used as standard in prior studies to evaluate the quality of renderings. + +Base architecture (HexPlane). Since our model is based on tensor-decomposed radiance fields, we compare our models with HexPlane [5]. The baseline model is the dynamic radiance field $\mathcal{F}_{\mathrm{DY}}$ , without the static radiance field $\mathcal{F}_{\mathrm{ST}}$ and the kinematic field $\mathcal{F}_{\mathrm{K}}$ . For a fair comparison, we use monocular depth supervision both to the baseline model and our models, although the original HexPlane model does not support the supervision from monocular depth. Since the base HexPlane model does not have an auxiliary flow field, we do not use the optical flow supervision and the photometric consistency loss. + +Table 1: Ablation study on Balloon1 scene of the NDVS dataset (24-frames-sparse). Please refer to the supplementary material for ablation results on other NDVS scenes. + +
MethodsNovel TimesSeen Times
FullDynamic OnlyFullDynamic Only
PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓
HexPlane [5]19.840.6680.14518.120.4190.25118.930.6490.13717.740.4290.202
Ours25.670.8650.06421.160.6250.17025.800.8670.05021.350.6320.121
w/o Higher order motion24.360.8420.07719.930.5560.20024.630.8460.06120.220.5720.145
w/o Lkinematic24.960.8530.06820.150.5680.18525.300.8580.05220.650.5940.129
w/o Kinematic integrity25.460.8620.06620.780.6070.17925.570.8650.05121.020.6200.126
w/o Rigidity25.110.8560.06620.400.5820.18125.430.8600.05120.820.6030.123
+ +![](images/16ac34041a8548cbf376d994f7713df7efa133d9447f29fe78e0eca2908a444d.jpg) +Fig. 6: Comparative visualization of kinematic fields. The first row shows RGB values rendered from our dynamic radiance field, showing PSNR values from a ground truth RGB image. From the second to the last row, we show flow (i.e., displacement), velocity, and acceleration. + +# 4.2 Evaluation of Kinematic Fields + +Ablation of kinematic losses. To show the effectiveness of our approach, we conduct a comprehensive ablation study on the Balloon1 scene of the NDVS dataset, focusing on kinematic loss components. In Table 1, we show ablative results from seen and unseen frame times. Our findings indicate that the model achieves optimal performance when all components are integrated. + +Fig. 6 delves into the kinematic field analysis. Omitting higher order motion restricts the model to utilizing only the velocity field, which proves insufficient for capturing non-linear motion. This limitation hinders the model's ability to minimize kinematic losses, leading to a suboptimal flow representation. In the absence of $\mathcal{L}_{\mathrm{kinematic}}$ , the radiance field suffers from reduced PSNR scores, which can be attributed to the physically incorrect motion and geometry. Compared + +![](images/a9c9a9ded169c82671c72721386a643c1ca4977fcaf5df7036ef34668ee0c877.jpg) +Fig. 7: Motion Order. Reconstruction Accuracy for the Balloon1 Scene (24-framesparse): The $x$ -axis of each plot categorizes maximum motion order as follows: velocity (1), acceleration (2), jerk (3), snap (4), and crackle (5). The metric utilized for each plot is specified at the top. + +![](images/86c577485099033c2c2c217354d6aafa70ae76281e8ab4026870c8709c3324f2.jpg) + +![](images/267eb84c3692aa30abeb538a785ae608f57c0287c9af9fe2a578ec64ace0b96a.jpg) + +![](images/11267c13c409d4a187ce802bb658ffac45c97ceb3f45404e71b75d9049f67ff8.jpg) + +Table 2: Quantitative comparison on NDVS. This table reports average results on all scenes. Please refer to the supplementary material for per-scene results. + +
(a) 12-frames
MethodsFull
PSNR↑LPIPS↓
NeRF [23]*21.990.250
NSFF [18]*24.330.199
DynamicNeRF [10]*26.100.082
HyperNeRF [26]*17.600.367
RoDynRF [20]*25.890.065
Ours26.330.071
*Results reported in [20]
+ +
(b) 24-frames
MethodsFullDynamic Only
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
NeRF [23]*24.900.8930.09816.980.5320.314
Luo et al. [21]*21.370.7460.14116.970.5300.207
Yoon et al. [36]*21.780.7610.12717.340.5470.200
NSFF [18]*28.190.9280.04521.910.7580.097
NSFF [18]28.040.9270.04521.340.7400.111
Ours28.780.9360.04222.130.7740.094
*Results reported in [18]
+ +to the full model, the model without the rigidity loss results in noisy, non-rigid motion, resulting in incorrect motion and the worse reconstruction accuracy. + +Motion order. Our study represents motion through various kinematic quantities. Adjusting the kinematic field to include different orders of motion, we can model our system using just velocity [7, 8], or extend to higher orders – fourth (snap) and fifth (crackle) – by extending the Taylor series (Eq. 8) and the kinematic integrity (Eq. 9) equations. Fig. 7 illustrates the influence of motion order on reconstruction accuracy for the NDVS dataset. As motion order increases, the reconstruction accuracy first improves, peaking at jerk, but further orders degrades the accuracy. We conjecture that the augmentation of model complexity with higher orders of kinematic quantities may render the model more susceptible to overfitting. + +# 4.3 Comparison with SOTA models + +Quantitative results. NSFF [18] is a dynamic neural radiance field model, which utilizes scene-flow fields for the regularization of a dynamic radiance field. Since NSFF is a neural model, the network training takes about 2 days on two V100 GPUs (96 GPU hours), using the official code by the authors. Note that optimizing a scene using our method takes about 7 hours per scene, which makes ours 12 times faster than NSFF in terms of GPU hours. We compare our results + +![](images/a0b747be6e9416d394293e76f6342c55d31ec714da4f924512e4d287da88f976.jpg) +Fig. 8: Qualitative results on NDVS (12-frames). The second and fourth rows illustrate the overlay of the synthesized and ground truth views of the scene. + +with NSFF [18] in Table 2b. In the table, our method shows better accuracy in terms of all metrics. In addition, we report results on NDVS (12-frames) in Table 2a. + +Qualitative results. Fig. 8 illustrates qualitative results, including the synthesized images, ground truth, and their overlaid images. As shown in the first and third rows, our method improves the visual quality of the rendered views. Additionally, the overlaid images demonstrate the geometrical accuracy of our model, showcasing its better alignment with the ground truth. This enhanced performance is attributed to the kinematics and physics incorporated in our methodology, which is fundamental to the improvement. + +# 5 Conclusion + +In this paper, we have presented a novel method to learn dynamic radiance fields with kinematic fields. We propose the kinematic field, where kinematic quantities are learned by joint learning with the radiance fields. Using the kinematic quantities from the kinematic fields, we propose physics-based loss functions that underpin our method. Lastly, we quantitatively and qualitatively show the effectiveness of our method on the monocular dataset. We believe that this work not only introduces the potential of incorporating kinematics into the learning of radiance fields but also sets a foundation for future direction in physics-based dynamic scene understanding. + +Acknowledgment. This project was partly supported by the Naver Cloud Corporation. Additionally, this work received support from the Institute of Information & communications Technology Planning & Evaluation (IITP) grant (RS-2023-00237965) and the National Research Foundation of Korea (NRF) grant (No. RS-2023-00208506(2024)), funded by the Korea government (MSIT). Sung-Eui Yoon is a corresponding author. + +# References + +1. Attal, B., Huang, J.B., Richardt, C., Zollhoefer, M., Kopf, J., O'Toole, M., Kim, C.: Hyperreel: High-fidelity 6-dof video with ray-conditioned sampling. arXiv preprint arXiv:2301.02238 (2023) +2. Baieri, D., Esposito, S., Maggioli, F., Rodola, E.: Fluid dynamics network: Topology-agnostic 4d reconstruction via fluid dynamics priors. arXiv preprint arXiv:2303.09871 (2023) +3. Baker, S., Scharstein, D., Lewis, J., Roth, S., Black, M.J., Szeliski, R.: A database and evaluation methodology for optical flow. International journal of computer vision 92, 1-31 (2011) +4. Barron, J.T., Mildenhall, B., Tancik, M., Hedman, P., Martin-Brualla, R., Srinivasan, P.P.: Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5855-5864 (2021) +5. Cao, A., Johnson, J.: Hexplane: A fast representation for dynamic scenes. arXiv preprint arXiv:2301.09632 (2023) +6. Chen, A., Xu, Z., Geiger, A., Yu, J., Su, H.: Tensorf: Tensorial radiance fields. In: Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXXII. pp. 333-350. Springer (2022) +7. Chu, M., Liu, L., Zheng, Q., Franz, E., Seidel, H.P., Theobalt, C., Zayer, R.: Physics informed neural fields for smoke reconstruction with sparse data. ACM Transactions on Graphics (TOG) 41(4), 1-14 (2022) +8. Du, Y., Zhang, Y., Yu, H.X., Tenenbaum, J.B., Wu, J.: Neural radiance flow for 4d view synthesis and video processing. In: 2021 IEEE/CVF International Conference on Computer Vision (ICCV). pp. 14304-14314. IEEE Computer Society (2021) +9. Fang, J., Yi, T., Wang, X., Xie, L., Zhang, X., Liu, W., Niešner, M., Tian, Q.: Fast dynamic radiance fields with time-aware neural voxels. In: SIGGRAPH Asia 2022 Conference Papers. pp. 1-9 (2022) +10. Gao, C., Saraf, A., Kopf, J., Huang, J.B.: Dynamic view synthesis from dynamic monocular video. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 5712-5721 (October 2021) +11. Gao, H., Li, R., Tulsiani, S., Russell, B., Kanazawa, A.: Monocular dynamic view synthesis: A reality check. In: Oh, A.H., Agarwal, A., Belgrave, D., Cho, K. (eds.) Advances in Neural Information Processing Systems (2022), https://openreview.net/forum?id=pCrB8orUkSq +12. Hart-Smith, L.: Application of the strain invariant failure theory (sift) to metals and fiber-polymer composites. Philosophical Magazine 90(31-32), 4263-4331 (2010) +13. Hess, R., Baker, C., Zihl, J.: The" motion-blind" patient: low-level spatial and temporal filters. Journal of Neuroscience 9(5), 1628-1640 (1989) +14. Jin, H., Liu, I., Xu, P., Zhang, X., Han, S., Bi, S., Zhou, X., Xu, Z., Su, H.: Tensoroir: Tensorial inverse rendering. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2023) +15. Kajiya, J.T., Von Herzen, B.P.: Ray tracing volume densities. In: Proceedings of the 11th Annual Conference on Computer Graphics and Interactive Techniques. p. 165-174 (1984) +16. Li, T., Slavcheva, M., Zollhoefer, M., Green, S., Lassner, C., Kim, C., Schmidt, T., Lovegrove, S., Goesele, M., Newcombe, R., et al.: Neural 3d video synthesis from multi-view video. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5521-5531 (2022) + +17. Li, Z., Müller, T., Evans, A., Taylor, R.H., Unberath, M., Liu, M.Y., Lin, C.H.: Neuralangelo: High-fidelity neural surface reconstruction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8456-8465 (2023) +18. Li, Z., Niklaus, S., Snavely, N., Wang, O.: Neural scene flow fields for space-time view synthesis of dynamic scenes. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6498-6508 (2021) +19. Li, Z., Wang, Q., Cole, F., Tucker, R., Snavely, N.: Dynibar: Neural dynamic image-based rendering. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4273-4284 (2023) +20. Liu, Y.L., Gao, C., Meuleman, A., Tseng, H.Y., Saraf, A., Kim, C., Chuang, Y.Y., Kopf, J., Huang, J.B.: Robust dynamic radiance fields. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 13-23 (June 2023) +21. Luo, X., Huang, J.B., Szeliski, R., Matzen, K., Kopf, J.: Consistent video depth estimation. ACM Transactions on Graphics (ToG) 39(4), 71-1 (2020) +22. Max, N., Chen, M.: Local and global illumination in the volume rendering integral. Tech. rep., Lawrence Livermore National Lab.(LLNL), Livermore, CA (United States) (2005) +23. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. In: ECCV (2020) +24. Orts-Escalano, S., Rhemann, C., Fanello, S., Chang, W., Kowdle, A., Degtyarev, Y., Kim, D., Davidson, P.L., Khamis, S., Dou, M., et al.: Holoportation: Virtual 3d teleportation in real-time. In: Proceedings of the 29th annual symposium on user interface software and technology. pp. 741-754 (2016) +25. Park, K., Sinha, U., Barron, J.T., Bouaziz, S., Goldman, D.B., Seitz, S.M., Martin-Brualla, R.: Nerfies: Deformable neural radiance fields. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5865-5874 (2021) +26. Park, K., Sinha, U., Hedman, P., Barron, J.T., Bouaziz, S., Goldman, D.B., Martin-Brualla, R., Seitz, S.M.: Hypernerf: A higher-dimensional representation for topologically varying neural radiance fields. arXiv preprint arXiv:2106.13228 (2021) +27. Pumarola, A., Corona, E., Pons-Moll, G., Moreno-Noguer, F.: D-nerf: Neural radiance fields for dynamic scenes. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10318-10327 (2021) +28. Ramasinghe, S., Shevchenko, V., Avraham, G., Hengel, A.V.D.: Bali-rf: Bandlimited radiance fields for dynamic scene modeling. arXiv preprint arXiv:2302.13543 (2023) +29. Soria, J., Sondergaard, R., Cantwell, B., Chong, M., Perry, A.: A study of the fine-scale motions of incompressible time-developing mixing layers. Physics of Fluids 6(2), 871-884 (1994) +30. Sun, D., Roth, S., Black, M.J.: A quantitative analysis of current practices in optical flow estimation and the principles behind them. International Journal of Computer Vision 106, 115-137 (2014) +31. Tretschk, E., Tewari, A., Golyanik, V., Zollhöfer, M., Lassner, C., Theobalt, C.: Non-rigid neural radiance fields: Reconstruction and novel view synthesis of a dynamic scene from monocular video. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 12959-12970 (2021) +32. Verbin, D., Hedman, P., Mildenhall, B., Zickler, T., Barron, J.T., Srinivasan, P.P.: Ref-nerf: Structured view-dependent appearance for neural radiance fields. In: 2022 + +IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 5481-5490. IEEE (2022) +33. Wang, C., Eckart, B., Lucey, S., Gallo, O.: Neural trajectory fields for dynamic novel view synthesis. arXiv preprint arXiv:2105.05994 (2021) +34. Wang, C., MacDonald, L.E., Jeni, L.A., Lucey, S.: Flow supervision for deformable nerf. arXiv preprint arXiv:2303.16333 (2023) +35. Wu, T., Zhong, F., Tagliasacchi, A., Cole, F., Oztireli, C.: D\~ 2nerf: Self-supervised decoupling of dynamic and static objects from a monocular video. Advances in Neural Information Processing Systems 35, 32653-32666 (2022) +36. Yoon, J.S., Kim, K., Gallo, O., Park, H.S., Kautz, J.: Novel view synthesis of dynamic scenes with globally coherent depths from a monocular camera. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5336-5345 (2020) +37. Zhang, K., Riegler, G., Snavely, N., Koltun, V.: Nerf++: Analyzing and improving neural radiance fields. arXiv preprint arXiv:2010.07492 (2020) \ No newline at end of file diff --git a/2024/Regularizing Dynamic Radiance Fields with Kinematic Fields/images.zip b/2024/Regularizing Dynamic Radiance Fields with Kinematic Fields/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..52eca570ed6f9940c1dbcd7c42cb923420093a4e --- /dev/null +++ b/2024/Regularizing Dynamic Radiance Fields with Kinematic Fields/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:72955a672eadb59671e1c3d90e80fb453305c4a32683639a1eb312d9114b99c4 +size 541486 diff --git a/2024/Regularizing Dynamic Radiance Fields with Kinematic Fields/layout.json b/2024/Regularizing Dynamic Radiance Fields with Kinematic Fields/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..6ebc65af646366739db3e41715dd7806241ae985 --- /dev/null +++ b/2024/Regularizing Dynamic Radiance Fields with Kinematic Fields/layout.json @@ -0,0 +1,10154 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 153, + 111, + 460, + 146 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 153, + 111, + 460, + 146 + ], + "spans": [ + { + "bbox": [ + 153, + 111, + 460, + 146 + ], + "type": "text", + "content": "Regularizing Dynamic Radiance Fields with Kinematic Fields" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 147, + 167, + 465, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 147, + 167, + 465, + 194 + ], + "spans": [ + { + "bbox": [ + 147, + 167, + 465, + 194 + ], + "type": "text", + "content": "Woobin " + }, + { + "bbox": [ + 147, + 167, + 465, + 194 + ], + "type": "inline_equation", + "content": "\\mathrm{Im}^{\\star 1,2}" + }, + { + "bbox": [ + 147, + 167, + 465, + 194 + ], + "type": "text", + "content": ", Geonho Cha" + }, + { + "bbox": [ + 147, + 167, + 465, + 194 + ], + "type": "inline_equation", + "content": "^3" + }, + { + "bbox": [ + 147, + 167, + 465, + 194 + ], + "type": "text", + "content": ", Sebin Lee" + }, + { + "bbox": [ + 147, + 167, + 465, + 194 + ], + "type": "inline_equation", + "content": "^1" + }, + { + "bbox": [ + 147, + 167, + 465, + 194 + ], + "type": "text", + "content": ", Jumin Lee" + }, + { + "bbox": [ + 147, + 167, + 465, + 194 + ], + "type": "inline_equation", + "content": "^1" + }, + { + "bbox": [ + 147, + 167, + 465, + 194 + ], + "type": "text", + "content": ", Juhyeong Seon" + }, + { + "bbox": [ + 147, + 167, + 465, + 194 + ], + "type": "inline_equation", + "content": "^1" + }, + { + "bbox": [ + 147, + 167, + 465, + 194 + ], + "type": "text", + "content": ", Dongyoon Wee" + }, + { + "bbox": [ + 147, + 167, + 465, + 194 + ], + "type": "inline_equation", + "content": "^3" + }, + { + "bbox": [ + 147, + 167, + 465, + 194 + ], + "type": "text", + "content": ", and Sung-Eui Yoon" + }, + { + "bbox": [ + 147, + 167, + 465, + 194 + ], + "type": "inline_equation", + "content": "^1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 188, + 201, + 425, + 223 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 188, + 201, + 425, + 223 + ], + "spans": [ + { + "bbox": [ + 188, + 201, + 425, + 223 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 188, + 201, + 425, + 223 + ], + "type": "text", + "content": " KAIST {iwbn,seb.lee,jmlee,munuwazzi,sungeui}@kaist.ac.kr" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 198, + 224, + 414, + 235 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 198, + 224, + 414, + 235 + ], + "spans": [ + { + "bbox": [ + 198, + 224, + 414, + 235 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 198, + 224, + 414, + 235 + ], + "type": "text", + "content": " Samsung Advanced Institute of Technology (SAIT)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 271, + 235, + 342, + 245 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 271, + 235, + 342, + 245 + ], + "spans": [ + { + "bbox": [ + 271, + 235, + 342, + 245 + ], + "type": "text", + "content": "3 Naver Cloud" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 214, + 247, + 399, + 257 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 214, + 247, + 399, + 257 + ], + "spans": [ + { + "bbox": [ + 214, + 247, + 399, + 257 + ], + "type": "text", + "content": "{geonho.cha,dongyoon.wei}@navercorp.com" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 159, + 285, + 455, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 285, + 455, + 449 + ], + "spans": [ + { + "bbox": [ + 159, + 285, + 455, + 449 + ], + "type": "text", + "content": "Abstract. This paper presents a novel approach for reconstructing dynamic radiance fields from monocular videos. We integrate kinematics with dynamic radiance fields, bridging the gap between the sparse nature of monocular videos and the real-world physics. Our method introduces the kinematic field, capturing motion through kinematic quantities: velocity, acceleration, and jerk. The kinematic field is jointly learned with the dynamic radiance field by minimizing the photometric loss without motion ground truth. We further augment our method with physics-driven regularizers grounded in kinematics. We propose physics-driven regularizers that ensure the physical validity of predicted kinematic quantities, including advective acceleration and jerk. Additionally, we control the motion trajectory based on rigidity equations formed with the predicted kinematic quantities. In experiments, our method outperforms the state-of-the-arts by capturing physical motion patterns within challenging real-world monocular videos." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 160, + 460, + 432, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 460, + 432, + 472 + ], + "spans": [ + { + "bbox": [ + 160, + 460, + 432, + 472 + ], + "type": "text", + "content": "Keywords: Dynamic radiance fields " + }, + { + "bbox": [ + 160, + 460, + 432, + 472 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 460, + 432, + 472 + ], + "type": "text", + "content": " Kinematic fields " + }, + { + "bbox": [ + 160, + 460, + 432, + 472 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 460, + 432, + 472 + ], + "type": "text", + "content": " 3D video" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 491, + 230, + 503 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 491, + 230, + 503 + ], + "spans": [ + { + "bbox": [ + 132, + 491, + 230, + 503 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 515, + 482, + 598 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 515, + 482, + 598 + ], + "spans": [ + { + "bbox": [ + 130, + 515, + 482, + 598 + ], + "type": "text", + "content": "Recognizing motion is crucial to how we visually perceive the world [13], and video is the medium that carries the motion to viewers in a remote place or at a different time. Through the lens of video, we watch and enjoy movies, user-generated contents, and remote video chats. While videos depict motion in a 2D format, transforming them into three-dimensional experiences can elevate our perception of the content to co-presence via applications such as teleportation [24]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 599, + 482, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 599, + 482, + 647 + ], + "spans": [ + { + "bbox": [ + 130, + 599, + 482, + 647 + ], + "type": "text", + "content": "Dynamic radiance fields [5,9,11,16,19,27] have been studied as a promising representation for reconstructing 3D dynamic scenes; it enables realistic novel view and time synthesis, which can upgrade 2D videos into 3D videos. Since multi-view cameras can give more information about the geometry in a scene," + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 653, + 388, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 653, + 388, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 653, + 388, + 665 + ], + "type": "text", + "content": "* This work was done during an internship at Naver Cloud." + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "text", + "content": "existing works have utilized camera rigs with dozens of cameras to capture multi-view video datasets [1, 5, 16]. Although high-quality multi-view videos result in undeniably better quality than monocular videos do, the multi-view capturing devices and environments are prohibitive in most cases." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 164, + 480, + 235 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 164, + 480, + 235 + ], + "spans": [ + { + "bbox": [ + 130, + 164, + 480, + 235 + ], + "type": "text", + "content": "Thus, there have been efforts to reconstruct dynamic radiance fields from monocular videos [10, 18, 20]. Compared to multi-view videos, monocular videos provide sparser information, which makes the reconstruction more challenging. For instance, there is scale ambiguity in monocular video, making it difficult to accurately locate an object within a space [34]. Thus, dynamic scene reconstruction from monocular videos has been considered an ill-posed problem." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 236, + 480, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 236, + 480, + 342 + ], + "spans": [ + { + "bbox": [ + 130, + 236, + 480, + 342 + ], + "type": "text", + "content": "In this paper, we incorporate kinematics to fill the gap between sparse monocular videos and real-world physics, potentially enhancing novel view and time synthesis. Our core idea is that motion within the real-world is governed by the principles of kinematics in physics, implying that representations of the world, such as videos, are inherently anchored in kinematics. Building on this foundation, we generalize conventional displacement fields – often called scene-flow fields [18] – to kinematic fields. In prior work, motion in radiance fields has been controlled via the regularization in the frequency domain [28, 33], displacement fields [18], and deformation fields [25]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 343, + 481, + 487 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 343, + 481, + 487 + ], + "spans": [ + { + "bbox": [ + 130, + 343, + 481, + 487 + ], + "type": "text", + "content": "Different from these approaches, our kinematic field regulates motion in radiance fields using kinematics; it innately creates continuous trajectories and enables us to impose regularization based on kinematics. We integrate physics-based regularizers to make the kinematic field physically plausible and thus improve the dynamic radiance field accordingly. For instance, the rigidity regularizer based on the first and the second invariant of the strain rate is applied on the kinematic field, suppressing non-rigid motion in the field. We quantitatively and qualitatively validate the proposed regularizers on the NVIDIA Dynamic Scenes (NDVS) dataset, showing the effectiveness of our method over the existing models. Lastly, the kinematic field uncovers the kinematics within a scene without a ground truth; it estimates not only point-wise 3D displacement within a scene, but also underlying velocity, acceleration, and jerk." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 504, + 237, + 516 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 504, + 237, + 516 + ], + "spans": [ + { + "bbox": [ + 132, + 504, + 237, + 516 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 528, + 479, + 576 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 528, + 479, + 576 + ], + "spans": [ + { + "bbox": [ + 130, + 528, + 479, + 576 + ], + "type": "text", + "content": "Neural radiance fields. In recent years, neural radiance fields [4,23,32] have emerged as a promising approach for learning an implicit neural representation of a 3D scene. Combined with the differentiable volume rendering technique, it has enabled realistic novel-view synthesis [4,23,32,37]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 577, + 479, + 600 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 577, + 479, + 600 + ], + "spans": [ + { + "bbox": [ + 130, + 577, + 479, + 600 + ], + "type": "text", + "content": "In essence, a radiance field " + }, + { + "bbox": [ + 130, + 577, + 479, + 600 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\mathrm{ST}}" + }, + { + "bbox": [ + 130, + 577, + 479, + 600 + ], + "type": "text", + "content": " maps a 5D vector " + }, + { + "bbox": [ + 130, + 577, + 479, + 600 + ], + "type": "inline_equation", + "content": "(x,y,z,\\theta ,\\phi)" + }, + { + "bbox": [ + 130, + 577, + 479, + 600 + ], + "type": "text", + "content": " to a 4D output vector " + }, + { + "bbox": [ + 130, + 577, + 479, + 600 + ], + "type": "inline_equation", + "content": "(\\mathbf{c}_s,\\sigma_s)" + }, + { + "bbox": [ + 130, + 577, + 479, + 600 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 242, + 601, + 480, + 613 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 242, + 601, + 480, + 613 + ], + "spans": [ + { + "bbox": [ + 242, + 601, + 480, + 613 + ], + "type": "interline_equation", + "content": "\\mathcal {F} _ {\\mathrm {S T}}: (x, y, z, \\theta , \\phi) \\rightarrow (\\mathbf {c} _ {s}, \\sigma_ {s}), \\tag {1}", + "image_path": "efe154e86a1ceb4c735d1776ed6cf1df1c2178e349268e221e064aa2f43f4891.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 617, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 481, + 665 + ], + "type": "text", + "content": "where the view-dependent color " + }, + { + "bbox": [ + 130, + 617, + 481, + 665 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_s = (r,g,b)" + }, + { + "bbox": [ + 130, + 617, + 481, + 665 + ], + "type": "text", + "content": " and density " + }, + { + "bbox": [ + 130, + 617, + 481, + 665 + ], + "type": "inline_equation", + "content": "\\sigma_s" + }, + { + "bbox": [ + 130, + 617, + 481, + 665 + ], + "type": "text", + "content": " are learned in relation to the spatial position " + }, + { + "bbox": [ + 130, + 617, + 481, + 665 + ], + "type": "inline_equation", + "content": "(x,y,z)" + }, + { + "bbox": [ + 130, + 617, + 481, + 665 + ], + "type": "text", + "content": " and the viewing direction " + }, + { + "bbox": [ + 130, + 617, + 481, + 665 + ], + "type": "inline_equation", + "content": "(\\theta ,\\phi)" + }, + { + "bbox": [ + 130, + 617, + 481, + 665 + ], + "type": "text", + "content": ". With the volumetric representation learned via an iterative optimization, volume rendering techniques [15, 22, 23] are employed to aggregate colors and densities along" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "type": "text", + "content": "W. Im et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": "a cast ray. Consequently, this leads to realistic novel-view synthesis from the inferred radiance, after being trained on dozens of images." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 140, + 480, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 140, + 480, + 224 + ], + "spans": [ + { + "bbox": [ + 130, + 140, + 480, + 224 + ], + "type": "text", + "content": "Dynamic radiance fields. Radiance fields also facilitate the novel-view synthesis of scenes with moving objects [1,5,16,25,26]. While this is more challenging than the static case, dynamic radiance fields incorporating motion allow for the rendering of a frame at a desired time and viewpoint. As a result, it can be utilized to make visual effects such as bullet time, stabilized, dolly zoomed, and stereo videos. Commonly, the time-dependent radiance fields are referred to as dynamic radiance fields, a term we will adopt throughout this paper." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 224, + 481, + 249 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 224, + 481, + 249 + ], + "spans": [ + { + "bbox": [ + 130, + 224, + 481, + 249 + ], + "type": "text", + "content": "In dynamic radiance fields, radiance is dependent on a varying time " + }, + { + "bbox": [ + 130, + 224, + 481, + 249 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 224, + 481, + 249 + ], + "type": "text", + "content": ". This modifies Eq. 1 by introducing a time-dependent term, resulting in:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 237, + 259, + 481, + 272 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 237, + 259, + 481, + 272 + ], + "spans": [ + { + "bbox": [ + 237, + 259, + 481, + 272 + ], + "type": "interline_equation", + "content": "\\mathcal {F} _ {\\mathrm {D Y}}: (x, y, z, t, \\theta , \\phi) \\rightarrow (\\mathbf {c} _ {d}, \\sigma_ {d}), \\tag {2}", + "image_path": "66b909522d655ed98d2d3111dbadac40ff71a3b5b3cb319f5dbf08afd046fa0c.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 281, + 479, + 317 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 281, + 479, + 317 + ], + "spans": [ + { + "bbox": [ + 130, + 281, + 479, + 317 + ], + "type": "text", + "content": "where the color and density are conditioned on time " + }, + { + "bbox": [ + 130, + 281, + 479, + 317 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 281, + 479, + 317 + ], + "type": "text", + "content": ". In this formulation, moving geometries of objects (e.g., translation and deformation) and changing radiance (e.g., moving shadows) can be learned via optimization." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 318, + 481, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 318, + 481, + 472 + ], + "spans": [ + { + "bbox": [ + 130, + 318, + 481, + 472 + ], + "type": "text", + "content": "Learning a dynamic representation presents inherent challenges due to the introduction of temporal dimension. To optimize this representation, a high-quality multi-view dataset can be employed to obtain fine rendering results [1, 5, 16]. Meanwhile, realistically captured imagery of a scene tends to be nonsynchronized and is likely to be obtained with a limited number of cameras and restricted displacements; this often results in a monocular setup [11, 18, 25, 26]. Previous researches have delved into deformable models for monocular videos [20, 25, 26, 31]. In these methods, a deformation field maps a spatio-temporal coordinate to a reference canonical space. While the canonical space aids in maintaining radiance consistency across global time frames, it tends to exhibit reduced effectiveness in generic monocular videos, where new objects might appear within a video [28]. In this study, we embrace the concept of scene flow fields [18], which is constructed without canonical space and deformation networks." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 474, + 481, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 474, + 481, + 568 + ], + "spans": [ + { + "bbox": [ + 130, + 474, + 481, + 568 + ], + "type": "text", + "content": "Regularization with kinematic priors. Monocular videos present challenges in reconstruction due to the constrained viewpoints of moving objects. To address this, integrating auxiliary flow fields has proven to be effective [8, 10, 18, 19]. Existing studies employ either scene flow fields [10, 18, 19] or velocity fields [8] to capture spatial displacements between consecutive frames. Since the scene flows can give spatial correspondences of the dynamic radiance field, it can be used to promote photometric consistency between the renderings of a given ray and its time-deformed counterpart." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 570, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 481, + 665 + ], + "type": "text", + "content": "Physical priors incorporated with scene flow fields play a pivotal role in regulating the radiance field. Regularization of motion has been applied to the dynamic field, bridging the gap between the sparse visual data and plausible real-world motion. For instance, a smoothness regularizer ensures that the displacements within close proximity are similar to each other, and kinetic energy regularizer promotes consistency between forward and backward flows, resulting in smoother displacement fields [18]. More deeply integrated with physics, the elastic energy regularizer minimizes the deformation between the canonical space" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 196, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 196, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 196, + 91, + 448, + 102 + ], + "type": "text", + "content": "Regularizing Dynamic Radiance Fields with Kinematic Fields" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 143, + 118, + 474, + 255 + ], + "blocks": [ + { + "bbox": [ + 143, + 118, + 474, + 255 + ], + "lines": [ + { + "bbox": [ + 143, + 118, + 474, + 255 + ], + "spans": [ + { + "bbox": [ + 143, + 118, + 474, + 255 + ], + "type": "image", + "image_path": "5d61a5cc06096f3d2093cbc73e139d1cfd6ef4d69d9cb995e6080f498377603b.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 262, + 482, + 297 + ], + "lines": [ + { + "bbox": [ + 131, + 262, + 482, + 297 + ], + "spans": [ + { + "bbox": [ + 131, + 262, + 482, + 297 + ], + "type": "text", + "content": "Fig. 1: Radiance and kinematic fields. This figure summarizes the three fields our method utilizes. The static and dynamic radiance fields are used for rendering, and the kinematic field is used in the training phase, regularizing the dynamic radiance field." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 318, + 479, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 318, + 479, + 376 + ], + "spans": [ + { + "bbox": [ + 130, + 318, + 479, + 376 + ], + "type": "text", + "content": "and the deformed space [25]. In fluid reconstruction works [2, 7], velocity fields grounded in physics have been shown to be critical in accurate 4D reconstruction. For instance, Navier-Stokes and transport equations can be integrated in reconstructing smoke [7] and fluid surfaces [2] with dynamic radiance fields to reconstruct more physically plausible structures." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 378, + 482, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 378, + 482, + 475 + ], + "spans": [ + { + "bbox": [ + 130, + 378, + 482, + 475 + ], + "type": "text", + "content": "We integrate kinematics to reconstruct dynamic radiance fields from generic monocular videos. We introduce the kinematic field, which captures motion through kinematic quantities such as velocity, acceleration, and jerk. Different from existing velocity field-based methods [7,8,34], our method has higher-order terms (i.e., acceleration and jerk) as latent of spatio-temporal correspondences. Further enhancing our approach, we incorporate physics-driven regularizers based on a transport equation [7] and the strain rate tensor, addressing dynamics of moving particles and rigid motion." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 491, + 202, + 503 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 491, + 202, + 503 + ], + "spans": [ + { + "bbox": [ + 132, + 491, + 202, + 503 + ], + "type": "text", + "content": "3 Method" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 515, + 209, + 526 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 515, + 209, + 526 + ], + "spans": [ + { + "bbox": [ + 132, + 515, + 209, + 526 + ], + "type": "text", + "content": "3.1 Overview" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 533, + 482, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 482, + 629 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 482, + 629 + ], + "type": "text", + "content": "Our system takes as input a video sequence, denoted as " + }, + { + "bbox": [ + 130, + 533, + 482, + 629 + ], + "type": "inline_equation", + "content": "(I_i, t_i, P_i)_{i=1}^N" + }, + { + "bbox": [ + 130, + 533, + 482, + 629 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 533, + 482, + 629 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 533, + 482, + 629 + ], + "type": "text", + "content": " is the number of frames. For each frame, we have an image " + }, + { + "bbox": [ + 130, + 533, + 482, + 629 + ], + "type": "inline_equation", + "content": "I_i \\in \\mathbb{R}^{H \\times W \\times 3}" + }, + { + "bbox": [ + 130, + 533, + 482, + 629 + ], + "type": "text", + "content": ", a timestamp " + }, + { + "bbox": [ + 130, + 533, + 482, + 629 + ], + "type": "inline_equation", + "content": "t_i \\in \\mathbb{R}" + }, + { + "bbox": [ + 130, + 533, + 482, + 629 + ], + "type": "text", + "content": ", and a precomputed camera pose " + }, + { + "bbox": [ + 130, + 533, + 482, + 629 + ], + "type": "inline_equation", + "content": "P_i \\in SE(3)" + }, + { + "bbox": [ + 130, + 533, + 482, + 629 + ], + "type": "text", + "content": " and intrinsic matrix " + }, + { + "bbox": [ + 130, + 533, + 482, + 629 + ], + "type": "inline_equation", + "content": "K_i \\in \\mathbb{R}^{3 \\times 3}" + }, + { + "bbox": [ + 130, + 533, + 482, + 629 + ], + "type": "text", + "content": ". From this input, our primary objective is to synthesize an image, denoted as " + }, + { + "bbox": [ + 130, + 533, + 482, + 629 + ], + "type": "inline_equation", + "content": "\\hat{I}_k" + }, + { + "bbox": [ + 130, + 533, + 482, + 629 + ], + "type": "text", + "content": ", corresponding to a novel time and viewpoint " + }, + { + "bbox": [ + 130, + 533, + 482, + 629 + ], + "type": "inline_equation", + "content": "(t_k, P_k, K_k)" + }, + { + "bbox": [ + 130, + 533, + 482, + 629 + ], + "type": "text", + "content": ". The parameters " + }, + { + "bbox": [ + 130, + 533, + 482, + 629 + ], + "type": "inline_equation", + "content": "t_k" + }, + { + "bbox": [ + 130, + 533, + 482, + 629 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 533, + 482, + 629 + ], + "type": "inline_equation", + "content": "P_k" + }, + { + "bbox": [ + 130, + 533, + 482, + 629 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 533, + 482, + 629 + ], + "type": "inline_equation", + "content": "K_k" + }, + { + "bbox": [ + 130, + 533, + 482, + 629 + ], + "type": "text", + "content": " are not restricted to those found in the original video input, allowing our system to render images from previously unobserved times and viewpoints." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 630, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 630, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 630, + 482, + 666 + ], + "type": "text", + "content": "To achieve this image synthesis for novel times and views, we leverage the radiance-field-based volume rendering [23]. Inspired by existing methods [18,35], we train two distinct radiance fields (Fig. 1): a dynamic radiance field, denoted" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "text", + "content": "W. Im et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "content": "as " + }, + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\mathrm{DY}}" + }, + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "content": " (Eq. 2), and a static radiance field, denoted as " + }, + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\mathrm{ST}}" + }, + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "content": " (Eq. 1). To blend the two fields, we utilize a technique from D²-NeRF [35], where rendering weights for dynamic and static samples are derived from the volume densities." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 152, + 482, + 283 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 152, + 482, + 283 + ], + "spans": [ + { + "bbox": [ + 130, + 152, + 482, + 283 + ], + "type": "text", + "content": "Addressing the challenge of achieving physically plausible spatio-temporal structures, we introduce the novel kinematic field. The kinematic field accounts for the kinematic quantities (e.g., velocity and acceleration). These quantities can be converted into displacement (Eq. 8), enabling us to use the photometric consistency loss (Eq. 14) for spatio-temporal regularization, as used in previous work [18]. With these kinematic estimations, we apply physics-based regularization techniques. These techniques involve density transport (Eq. 12) and motion rigidity through the strain rate tensor (Eq. 11). Combining the photometric loss (Eq. 14) with these techniques enhances the reconstruction of radiance and geometry, leading to improved novel time and view synthesis performance from underdetermined data such as a monocular video." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 300, + 344, + 313 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 300, + 344, + 313 + ], + "spans": [ + { + "bbox": [ + 132, + 300, + 344, + 313 + ], + "type": "text", + "content": "3.2 Dynamic and Static Radiance Fields" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 320, + 480, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 320, + 480, + 403 + ], + "spans": [ + { + "bbox": [ + 130, + 320, + 480, + 403 + ], + "type": "text", + "content": "In monocular dynamic radiance field methods, it is common to decompose the radiance fields into dynamic and static fields [11, 18, 20] to prevent undesirable motion in the static area. We employ tensor-decomposition-based representations [5, 6] for both dynamic and static fields. Specifically, for the dynamic radiance field " + }, + { + "bbox": [ + 130, + 320, + 480, + 403 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\mathrm{DY}}" + }, + { + "bbox": [ + 130, + 320, + 480, + 403 + ], + "type": "text", + "content": ", we use HexPlane [5], and for the static radiance field " + }, + { + "bbox": [ + 130, + 320, + 480, + 403 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\mathrm{ST}}" + }, + { + "bbox": [ + 130, + 320, + 480, + 403 + ], + "type": "text", + "content": ", we use TensoRF [6]. Further details on the tensor decomposition can be found in the supplementary material." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 404, + 481, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 404, + 481, + 464 + ], + "spans": [ + { + "bbox": [ + 130, + 404, + 481, + 464 + ], + "type": "text", + "content": "Definition. The dynamic and static radiance fields are defined in Eq. 1-2 (" + }, + { + "bbox": [ + 130, + 404, + 481, + 464 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\mathrm{ST}}" + }, + { + "bbox": [ + 130, + 404, + 481, + 464 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 404, + 481, + 464 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\mathrm{DY}}" + }, + { + "bbox": [ + 130, + 404, + 481, + 464 + ], + "type": "text", + "content": "). Here, the inputs correspond to the position " + }, + { + "bbox": [ + 130, + 404, + 481, + 464 + ], + "type": "inline_equation", + "content": "(x,y,z)" + }, + { + "bbox": [ + 130, + 404, + 481, + 464 + ], + "type": "text", + "content": ", the timestamp " + }, + { + "bbox": [ + 130, + 404, + 481, + 464 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 404, + 481, + 464 + ], + "type": "text", + "content": ", and the viewing direction " + }, + { + "bbox": [ + 130, + 404, + 481, + 464 + ], + "type": "inline_equation", + "content": "(\\theta, \\phi)" + }, + { + "bbox": [ + 130, + 404, + 481, + 464 + ], + "type": "text", + "content": " from the camera origin. Each field yields the color " + }, + { + "bbox": [ + 130, + 404, + 481, + 464 + ], + "type": "inline_equation", + "content": "\\mathbf{c}" + }, + { + "bbox": [ + 130, + 404, + 481, + 464 + ], + "type": "text", + "content": " and the volume density " + }, + { + "bbox": [ + 130, + 404, + 481, + 464 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 130, + 404, + 481, + 464 + ], + "type": "text", + "content": ". Note that the static field is independent of time " + }, + { + "bbox": [ + 130, + 404, + 481, + 464 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 404, + 481, + 464 + ], + "type": "text", + "content": ", ensuring a consistent structure throughout the whole time span." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 464, + 482, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 464, + 482, + 559 + ], + "spans": [ + { + "bbox": [ + 130, + 464, + 482, + 559 + ], + "type": "text", + "content": "Composite rendering. We use a composite rendering method to obtain the final rendering with the two radiance fields. It is to be noted that individual components (i.e., dynamic-only and static-only) can be rendered using the standard volume rendering technique [23]. To combine both fields, we employ an additive composition method in which each field influences the transmittance of a ray [35]. Let's consider a camera ray " + }, + { + "bbox": [ + 130, + 464, + 482, + 559 + ], + "type": "inline_equation", + "content": "\\mathbf{r}(\\tau) = \\mathbf{o} + \\tau \\mathbf{u}" + }, + { + "bbox": [ + 130, + 464, + 482, + 559 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 464, + 482, + 559 + ], + "type": "inline_equation", + "content": "\\mathbf{o}" + }, + { + "bbox": [ + 130, + 464, + 482, + 559 + ], + "type": "text", + "content": " is the origin and " + }, + { + "bbox": [ + 130, + 464, + 482, + 559 + ], + "type": "inline_equation", + "content": "\\mathbf{u}" + }, + { + "bbox": [ + 130, + 464, + 482, + 559 + ], + "type": "text", + "content": " is the direction. With a predefined range for the ray distance " + }, + { + "bbox": [ + 130, + 464, + 482, + 559 + ], + "type": "inline_equation", + "content": "[\\tau_n, \\tau_f]" + }, + { + "bbox": [ + 130, + 464, + 482, + 559 + ], + "type": "text", + "content": ", the rendering function is formulated as:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 181, + 567, + 481, + 594 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 181, + 567, + 481, + 594 + ], + "spans": [ + { + "bbox": [ + 181, + 567, + 481, + 594 + ], + "type": "interline_equation", + "content": "\\hat {C} (\\mathbf {r}, t _ {i}) = \\int_ {\\tau_ {n}} ^ {\\tau_ {f}} T (\\tau) \\left(\\sigma_ {s} (\\tau) \\cdot \\mathbf {c} _ {s} (\\tau) + \\sigma_ {d} (\\tau , t _ {i}) \\cdot \\mathbf {c} _ {d} (\\tau , t _ {i})\\right) d \\tau , \\tag {3}", + "image_path": "b77845811a695b27b53975f59c36e858f171254a06a70d0bc94e9387e2897c11.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 219, + 596, + 481, + 622 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 219, + 596, + 481, + 622 + ], + "spans": [ + { + "bbox": [ + 219, + 596, + 481, + 622 + ], + "type": "interline_equation", + "content": "\\text {w h e r e} T (\\tau) = \\exp \\left(- \\int_ {\\tau_ {n}} ^ {\\tau} \\left(\\sigma_ {s} (u) + \\sigma_ {d} (u, t _ {i})\\right) d u\\right). \\tag {4}", + "image_path": "aae763c20225e191a0bcf87c8b548c8e891793b8c2fcfdde88ff4be124496a05.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "text", + "content": "Here, each notation for the sample point is simplified for concise representation (e.g., " + }, + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\sigma_s(\\tau) = \\sigma_s(\\mathbf{r}(\\tau))" + }, + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "text", + "content": "). We visualize the example of the composite rendering and separate rendering in the first and second rows in Fig. 2." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 196, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 196, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 196, + 91, + 448, + 102 + ], + "type": "text", + "content": "Regularizing Dynamic Radiance Fields with Kinematic Fields" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 178, + 116, + 443, + 282 + ], + "blocks": [ + { + "bbox": [ + 178, + 116, + 443, + 282 + ], + "lines": [ + { + "bbox": [ + 178, + 116, + 443, + 282 + ], + "spans": [ + { + "bbox": [ + 178, + 116, + 443, + 282 + ], + "type": "image", + "image_path": "00fbf48959c2f2140a078b90b27309e78b499fe24e1e1da8d7a13a664a28ca19.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 293, + 482, + 380 + ], + "lines": [ + { + "bbox": [ + 130, + 293, + 482, + 380 + ], + "spans": [ + { + "bbox": [ + 130, + 293, + 482, + 380 + ], + "type": "text", + "content": "Fig. 2: Visualization of each predicted component. The top row presents synthesized views at different times. In (a) and (b), the dynamic and static components of the scene at time " + }, + { + "bbox": [ + 130, + 293, + 482, + 380 + ], + "type": "inline_equation", + "content": "t_2" + }, + { + "bbox": [ + 130, + 293, + 482, + 380 + ], + "type": "text", + "content": " are depicted separately. (b) In the case of an inobservable static area in the entire sequence (e.g., the space behind the jumping people), radiance might not be correct. The displacement field (c) can be computed by Taylor approximation (Eq. 8) with motion fields (d-f): velocity, acceleration, and jerk. Each field is visualized through reprojecting each field to the camera view. The standard HSV visualization [3] is used to colorize arrows." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 405, + 482, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 405, + 482, + 441 + ], + "spans": [ + { + "bbox": [ + 130, + 405, + 482, + 441 + ], + "type": "text", + "content": "Dynamic and static separation. To factorize a scene into static and dynamic parts, we use a metric indicating the likelihood that a point belongs to the dynamic fields [35]. This metric is given by:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 273, + 451, + 481, + 473 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 273, + 451, + 481, + 473 + ], + "spans": [ + { + "bbox": [ + 273, + 451, + 481, + 473 + ], + "type": "interline_equation", + "content": "p _ {d} = \\frac {\\sigma_ {d}}{\\sigma_ {d} + \\sigma_ {s}}, \\tag {5}", + "image_path": "e98066e9902717c36d959cad713ff3affa9842003adcee7c24d78a95acc5a924.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 481, + 457, + 493 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 481, + 457, + 493 + ], + "spans": [ + { + "bbox": [ + 130, + 481, + 457, + 493 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 481, + 457, + 493 + ], + "type": "inline_equation", + "content": "\\sigma_{d}" + }, + { + "bbox": [ + 130, + 481, + 457, + 493 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 481, + 457, + 493 + ], + "type": "inline_equation", + "content": "\\sigma_{s}" + }, + { + "bbox": [ + 130, + 481, + 457, + 493 + ], + "type": "text", + "content": " are density values sampled from a spatio-temporal point." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 494, + 481, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 494, + 481, + 518 + ], + "spans": [ + { + "bbox": [ + 130, + 494, + 481, + 518 + ], + "type": "text", + "content": "To facilitate the separation of the radiance fields, we employ the skewed binary entropy loss, as proposed by [35]. The entropy loss is defined as:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 208, + 527, + 481, + 543 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 208, + 527, + 481, + 543 + ], + "spans": [ + { + "bbox": [ + 208, + 527, + 481, + 543 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {b} = - \\left(p _ {d} ^ {k} \\cdot \\log \\left(p _ {d} ^ {k}\\right) + \\left(1 - p _ {d} ^ {k}\\right) \\cdot \\log \\left(1 - p _ {d} ^ {k}\\right)\\right), \\tag {6}", + "image_path": "9f9b87c077765f957e272b70f9969d0dcdb2def685717f29005cfc7bd13983ef.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 552, + 482, + 588 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 552, + 482, + 588 + ], + "spans": [ + { + "bbox": [ + 130, + 552, + 482, + 588 + ], + "type": "text", + "content": "where the exponent " + }, + { + "bbox": [ + 130, + 552, + 482, + 588 + ], + "type": "inline_equation", + "content": "k \\geq 1" + }, + { + "bbox": [ + 130, + 552, + 482, + 588 + ], + "type": "text", + "content": " is the skew factor, controlling a bias toward static explanation. This entropy-based loss can encourage the radiance fields to have a clearer distinction between the static and dynamic components." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 607, + 242, + 618 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 607, + 242, + 618 + ], + "spans": [ + { + "bbox": [ + 131, + 607, + 242, + 618 + ], + "type": "text", + "content": "3.3 Kinematic Field" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 629, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 481, + 665 + ], + "type": "text", + "content": "The kinematic field has a spatio-temporal structure to facilitate spatial and temporal queries. We use the tensor-decomposition-based representation [5] for the kinematic field. As depicted in Fig. 1, the kinematic field serves to regulate" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "text", + "content": "W. Im et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 113, + 302, + 175 + ], + "blocks": [ + { + "bbox": [ + 133, + 113, + 302, + 175 + ], + "lines": [ + { + "bbox": [ + 133, + 113, + 302, + 175 + ], + "spans": [ + { + "bbox": [ + 133, + 113, + 302, + 175 + ], + "type": "image", + "image_path": "ddb5f4f6daa9e7693eb568e64ac48585a260925bf180acb76283f19bda938471.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 153, + 177, + 285, + 186 + ], + "lines": [ + { + "bbox": [ + 153, + 177, + 285, + 186 + ], + "spans": [ + { + "bbox": [ + 153, + 177, + 285, + 186 + ], + "type": "text", + "content": "(a) Without kinematic regularization" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 309, + 113, + 479, + 175 + ], + "blocks": [ + { + "bbox": [ + 309, + 113, + 479, + 175 + ], + "lines": [ + { + "bbox": [ + 309, + 113, + 479, + 175 + ], + "spans": [ + { + "bbox": [ + 309, + 113, + 479, + 175 + ], + "type": "image", + "image_path": "5d14acfbe7ef51cf0b59c6b338cc36d26cd2c12e21bcb29cccd19e62ab2a8a6f.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 334, + 177, + 456, + 186 + ], + "lines": [ + { + "bbox": [ + 334, + 177, + 456, + 186 + ], + "spans": [ + { + "bbox": [ + 334, + 177, + 456, + 186 + ], + "type": "text", + "content": "(b) With kinematic regularization" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 198, + 482, + 253 + ], + "lines": [ + { + "bbox": [ + 130, + 198, + 482, + 253 + ], + "spans": [ + { + "bbox": [ + 130, + 198, + 482, + 253 + ], + "type": "text", + "content": "Fig. 3: Effect of kinematic regularization. We visualize the rendered RGB and motion of each field. Without kinematic regularization, motion fields tend to show granular patterns. Our kinematic fields not only make the field smoother but also satisfy the kinematic property, i.e., " + }, + { + "bbox": [ + 130, + 198, + 482, + 253 + ], + "type": "inline_equation", + "content": "\\mathbf{a} = \\partial \\mathbf{v} / \\partial t + \\mathbf{v} \\cdot \\nabla \\mathbf{v}" + }, + { + "bbox": [ + 130, + 198, + 482, + 253 + ], + "type": "text", + "content": ". We abbreviate the advective equation to 'Adv.' in the figure." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 275, + 479, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 275, + 479, + 300 + ], + "spans": [ + { + "bbox": [ + 130, + 275, + 479, + 300 + ], + "type": "text", + "content": "the radiance field, ensuring that the motion in the radiance field is physically plausible." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 300, + 334, + 311 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 300, + 334, + 311 + ], + "spans": [ + { + "bbox": [ + 132, + 300, + 334, + 311 + ], + "type": "text", + "content": "Definition. The kinematic field is defined as:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 238, + 319, + 481, + 333 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 238, + 319, + 481, + 333 + ], + "spans": [ + { + "bbox": [ + 238, + 319, + 481, + 333 + ], + "type": "interline_equation", + "content": "\\mathcal {F} _ {\\mathrm {K}}: (x, y, z, t) \\rightarrow (\\mathbf {v}, \\mathbf {a}, \\mathbf {j}, \\dots), \\tag {7}", + "image_path": "3bc26f4f748ecf0a8fe5ee6e37380df6172ef194b04c24cac1aea84a4ab0b84a.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 341, + 479, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 341, + 479, + 376 + ], + "spans": [ + { + "bbox": [ + 130, + 341, + 479, + 376 + ], + "type": "text", + "content": "where each vector represents velocity, acceleration, and jerk, such that " + }, + { + "bbox": [ + 130, + 341, + 479, + 376 + ], + "type": "inline_equation", + "content": "\\mathbf{v} = (v_x, v_y, v_z)" + }, + { + "bbox": [ + 130, + 341, + 479, + 376 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 341, + 479, + 376 + ], + "type": "inline_equation", + "content": "\\mathbf{a} = (a_x, a_y, a_z)" + }, + { + "bbox": [ + 130, + 341, + 479, + 376 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 341, + 479, + 376 + ], + "type": "inline_equation", + "content": "\\mathbf{j} = (j_x, j_y, j_z)" + }, + { + "bbox": [ + 130, + 341, + 479, + 376 + ], + "type": "text", + "content": ". Note that we can extend the field into higher-order quantities such as snap (4th), crackle (5th), and so on." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 377, + 480, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 377, + 480, + 434 + ], + "spans": [ + { + "bbox": [ + 130, + 377, + 480, + 434 + ], + "type": "text", + "content": "Displacement. Using the provided kinematic quantities, we can approximate a particle's displacement at position " + }, + { + "bbox": [ + 130, + 377, + 480, + 434 + ], + "type": "inline_equation", + "content": "(x,y,z)" + }, + { + "bbox": [ + 130, + 377, + 480, + 434 + ], + "type": "text", + "content": " and time " + }, + { + "bbox": [ + 130, + 377, + 480, + 434 + ], + "type": "inline_equation", + "content": "t + \\Delta t" + }, + { + "bbox": [ + 130, + 377, + 480, + 434 + ], + "type": "text", + "content": " via the Taylor approximation, assuming that the kinematic quantities are consistent within a local temporal range " + }, + { + "bbox": [ + 130, + 377, + 480, + 434 + ], + "type": "inline_equation", + "content": "\\Delta t" + }, + { + "bbox": [ + 130, + 377, + 480, + 434 + ], + "type": "text", + "content": ". The displacement " + }, + { + "bbox": [ + 130, + 377, + 480, + 434 + ], + "type": "inline_equation", + "content": "d(x,y,z,t,\\Delta t)" + }, + { + "bbox": [ + 130, + 377, + 480, + 434 + ], + "type": "text", + "content": " can be approximated by:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 194, + 435, + 480, + 460 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 194, + 435, + 480, + 460 + ], + "spans": [ + { + "bbox": [ + 194, + 435, + 480, + 460 + ], + "type": "interline_equation", + "content": "d (x, y, z, t, \\Delta t) \\approx \\frac {(\\Delta t) \\mathbf {v}}{1 !} + \\frac {(\\Delta t) ^ {2} \\mathbf {a}}{2 !} + \\frac {(\\Delta t) ^ {3} \\mathbf {j}}{3 !} + \\dots . \\tag {8}", + "image_path": "f9c546e30e39629db2244aa90188043bffd7d98754540268c485476849beb0ab.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 462, + 479, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 462, + 479, + 498 + ], + "spans": [ + { + "bbox": [ + 130, + 462, + 479, + 498 + ], + "type": "text", + "content": "Computing the displacement of a point within a temporal range allows for computing the photometric consistency loss (Eq. 14) and scene flows from the kinematic field." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 131, + 515, + 289, + 528 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 515, + 289, + 528 + ], + "spans": [ + { + "bbox": [ + 131, + 515, + 289, + 528 + ], + "type": "text", + "content": "3.4 Kinematic Regularization" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 533, + 479, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 479, + 581 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 479, + 581 + ], + "type": "text", + "content": "Given the kinematic field, it is important to make sure the learned motion dynamics are physically plausible. As Fig. 3 indicates, learning kinematic fields without adequate regularization may result in a sub-optimal kinematic field. The subsequent sections elucidate regularization strategies." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 582, + 480, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 480, + 665 + ], + "type": "text", + "content": "Integrity of kinematic fields. Learning the kinematic field " + }, + { + "bbox": [ + 130, + 582, + 480, + 665 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\mathrm{K}}" + }, + { + "bbox": [ + 130, + 582, + 480, + 665 + ], + "type": "text", + "content": " naively does not guarantee the integrity of the kinematic quantities. During training, the quantities are supervised via the photometric consistency loss (Eq. 14), which is applied upon the approximated displacement (Eq. 8). It is important to note that even with supervision by the displacement, the physical relationships among the kinematic quantities may not be correct, since the higher-order terms in the displacement equation (Eq. 8) make the system underdetermined." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 196, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 196, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 196, + 91, + 448, + 102 + ], + "type": "text", + "content": "Regularizing Dynamic Radiance Fields with Kinematic Fields" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 175 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 175 + ], + "type": "text", + "content": "Given a velocity field, we can derive the acceleration field using the principle of advective acceleration. Given acceleration " + }, + { + "bbox": [ + 130, + 116, + 479, + 175 + ], + "type": "inline_equation", + "content": "\\mathbf{a}" + }, + { + "bbox": [ + 130, + 116, + 479, + 175 + ], + "type": "text", + "content": " and velocity " + }, + { + "bbox": [ + 130, + 116, + 479, + 175 + ], + "type": "inline_equation", + "content": "\\mathbf{v}" + }, + { + "bbox": [ + 130, + 116, + 479, + 175 + ], + "type": "text", + "content": " at a spatiotemporal coordinate, the partial derivative and advection yield " + }, + { + "bbox": [ + 130, + 116, + 479, + 175 + ], + "type": "inline_equation", + "content": "\\mathbf{a} = \\partial \\mathbf{v} / \\partial t + \\mathbf{v} \\cdot \\nabla \\mathbf{v}" + }, + { + "bbox": [ + 130, + 116, + 479, + 175 + ], + "type": "text", + "content": ". Similarly, we can formulate the relationship between " + }, + { + "bbox": [ + 130, + 116, + 479, + 175 + ], + "type": "inline_equation", + "content": "\\mathbf{v}" + }, + { + "bbox": [ + 130, + 116, + 479, + 175 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 116, + 479, + 175 + ], + "type": "inline_equation", + "content": "\\mathbf{a}" + }, + { + "bbox": [ + 130, + 116, + 479, + 175 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 116, + 479, + 175 + ], + "type": "inline_equation", + "content": "\\mathbf{j}" + }, + { + "bbox": [ + 130, + 116, + 479, + 175 + ], + "type": "text", + "content": " as: " + }, + { + "bbox": [ + 130, + 116, + 479, + 175 + ], + "type": "inline_equation", + "content": "\\mathbf{j} = \\partial \\mathbf{a} / \\partial t + \\mathbf{v} \\cdot \\nabla \\mathbf{a}" + }, + { + "bbox": [ + 130, + 116, + 479, + 175 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 176, + 479, + 198 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 176, + 479, + 198 + ], + "spans": [ + { + "bbox": [ + 130, + 176, + 479, + 198 + ], + "type": "text", + "content": "To ensure kinematic consistency, we introduce the kinematic loss " + }, + { + "bbox": [ + 130, + 176, + 479, + 198 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{K}}" + }, + { + "bbox": [ + 130, + 176, + 479, + 198 + ], + "type": "text", + "content": " defined as:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 189, + 196, + 480, + 225 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 189, + 196, + 480, + 225 + ], + "spans": [ + { + "bbox": [ + 189, + 196, + 480, + 225 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {K}} = \\left\\| \\mathbf {a} - \\frac {\\partial \\mathbf {v}}{\\partial t} - \\mathbf {v} \\cdot \\nabla \\mathbf {v} \\right\\| _ {2} ^ {2} + \\left\\| \\mathbf {j} - \\frac {\\partial \\mathbf {a}}{\\partial t} - \\mathbf {v} \\cdot \\nabla \\mathbf {a} \\right\\| _ {2} ^ {2} + \\dots , \\tag {9}", + "image_path": "67a32ffe5bbc7a91f5a1e3e2babece954f625491ca9fabe6273a451f8a42c854.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 228, + 479, + 264 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 228, + 479, + 264 + ], + "spans": [ + { + "bbox": [ + 130, + 228, + 479, + 264 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 228, + 479, + 264 + ], + "type": "inline_equation", + "content": "\\nabla" + }, + { + "bbox": [ + 130, + 228, + 479, + 264 + ], + "type": "text", + "content": " is the Jacobian operation. To compute the partial derivatives and Jacobian matrices, we use a numerical method as presented in a neural surface reconstruction work [17]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 264, + 480, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 264, + 480, + 346 + ], + "spans": [ + { + "bbox": [ + 130, + 264, + 480, + 346 + ], + "type": "text", + "content": "Physics-informed regularization. In the context of regularizing radiance fields through the use of an auxiliary kinematic field, it is important to acknowledge the highly underdetermined nature of the problem. Essentially, there exists a broad spectrum of possible potential kinematic field solutions that could perfectly fit the observed training views. This complexity necessitates the introduction of a regularization strategy that can guide the motion representation toward more accurately reflecting real-world physics." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 347, + 479, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 347, + 479, + 396 + ], + "spans": [ + { + "bbox": [ + 130, + 347, + 479, + 396 + ], + "type": "text", + "content": "While not universally ideal, rigidity [25] and incompressibility [2,7] have been identified as effective metrics for encapsulating real-world dynamics; in videos that we normally encounter, objects in a scene have nearly consistent density and are mostly rigid in a short time range." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 396, + 479, + 454 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 396, + 479, + 454 + ], + "spans": [ + { + "bbox": [ + 130, + 396, + 479, + 454 + ], + "type": "text", + "content": "Rigidity. One benefit of having a velocity field is that we can measure the rate of deformation by its derivatives. A strain rate tensor " + }, + { + "bbox": [ + 130, + 396, + 479, + 454 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 130, + 396, + 479, + 454 + ], + "type": "text", + "content": " defines a kinematic property of a deformation; it captures the rate at which a material undergoes deformation, and it is crucial for ensuring that objects maintain their structural integrity during motion." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 455, + 479, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 455, + 479, + 491 + ], + "spans": [ + { + "bbox": [ + 130, + 455, + 479, + 491 + ], + "type": "text", + "content": "In this paper, we propose using the invariants of strain rate tensor [12, 29] as a measure of rigidity. The velocity gradient " + }, + { + "bbox": [ + 130, + 455, + 479, + 491 + ], + "type": "inline_equation", + "content": "\\nabla \\mathbf{v}" + }, + { + "bbox": [ + 130, + 455, + 479, + 491 + ], + "type": "text", + "content": " sampled from the kinematic field, leads to the strain rate tensor " + }, + { + "bbox": [ + 130, + 455, + 479, + 491 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 130, + 455, + 479, + 491 + ], + "type": "text", + "content": " by:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 257, + 498, + 480, + 520 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 257, + 498, + 480, + 520 + ], + "spans": [ + { + "bbox": [ + 257, + 498, + 480, + 520 + ], + "type": "interline_equation", + "content": "D = \\frac {1}{2} (\\nabla \\mathbf {v} + (\\nabla \\mathbf {v}) ^ {T}). \\tag {10}", + "image_path": "ca3088ab9d75ec4e69d13735263ecded992d489cc651a685a65909931490d6ce.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 525, + 479, + 549 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 525, + 479, + 549 + ], + "spans": [ + { + "bbox": [ + 130, + 525, + 479, + 549 + ], + "type": "text", + "content": "The rigidity loss function " + }, + { + "bbox": [ + 130, + 525, + 479, + 549 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_R" + }, + { + "bbox": [ + 130, + 525, + 479, + 549 + ], + "type": "text", + "content": " is formulated based on the first and second invariants of the strain rate tensor:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 206, + 555, + 480, + 591 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 206, + 555, + 480, + 591 + ], + "spans": [ + { + "bbox": [ + 206, + 555, + 480, + 591 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {R} = \\underbrace {\\lambda_ {\\mathrm {d i v}} (\\nabla \\cdot \\mathbf {v}) ^ {2}} _ {\\text {1 s t i n v a r i a n t}} + \\underbrace {\\left(\\frac {1}{2} (\\operatorname {t r} (D) ^ {2} - \\operatorname {t r} (D \\cdot D)))\\right) ^ {2}} _ {\\text {2 n d i n v a r i a n t}}, \\tag {11}", + "image_path": "50d03d8c596424262f9609911bdd9e4bc27492ba6e29ac90dfcf8943c083d18b.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 598, + 479, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 598, + 479, + 647 + ], + "spans": [ + { + "bbox": [ + 130, + 598, + 479, + 647 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 598, + 479, + 647 + ], + "type": "inline_equation", + "content": "\\operatorname{tr}(\\cdot)" + }, + { + "bbox": [ + 130, + 598, + 479, + 647 + ], + "type": "text", + "content": " denotes the trace operator and " + }, + { + "bbox": [ + 130, + 598, + 479, + 647 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{div}}" + }, + { + "bbox": [ + 130, + 598, + 479, + 647 + ], + "type": "text", + "content": " controls the balance between the two terms. Note that the first invariant of strain rate measures dilational change (i.e., volumetric change) with respect to time and the second invariant of strain rate measures distortional change." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "type": "text", + "content": "W. Im et al." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 133, + 651, + 231, + 667 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 651, + 231, + 667 + ], + "spans": [ + { + "bbox": [ + 133, + 651, + 231, + 667 + ], + "type": "inline_equation", + "content": "\\overline{4f'(x)} = \\frac{f(x + \\epsilon) - f(x - \\epsilon)}{2\\epsilon}" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 118, + 282, + 252 + ], + "blocks": [ + { + "bbox": [ + 133, + 118, + 282, + 252 + ], + "lines": [ + { + "bbox": [ + 133, + 118, + 282, + 252 + ], + "spans": [ + { + "bbox": [ + 133, + 118, + 282, + 252 + ], + "type": "image", + "image_path": "1551b001fadd2ddc6de0d6978ba34f979017653f9dfb6947b58194b66e89d8e6.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 260, + 285, + 381 + ], + "lines": [ + { + "bbox": [ + 130, + 260, + 285, + 381 + ], + "spans": [ + { + "bbox": [ + 130, + 260, + 285, + 381 + ], + "type": "text", + "content": "Fig.4: Visualization of density variation by a spatial coordinate " + }, + { + "bbox": [ + 130, + 260, + 285, + 381 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 130, + 260, + 285, + 381 + ], + "type": "text", + "content": ". The plot displays the gradient " + }, + { + "bbox": [ + 130, + 260, + 285, + 381 + ], + "type": "inline_equation", + "content": "-\\partial \\mathcal{L}_{\\mathrm{T}} / \\partial v_{\\sigma_d}" + }, + { + "bbox": [ + 130, + 260, + 285, + 381 + ], + "type": "text", + "content": " with arrows. With the velocity field directed to the right (i.e., " + }, + { + "bbox": [ + 130, + 260, + 285, + 381 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_x = 0.3" + }, + { + "bbox": [ + 130, + 260, + 285, + 381 + ], + "type": "text", + "content": "), we can compute the gradient of the transport regularization " + }, + { + "bbox": [ + 130, + 260, + 285, + 381 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{T}}" + }, + { + "bbox": [ + 130, + 260, + 285, + 381 + ], + "type": "text", + "content": " w.r.t. the rate of density change " + }, + { + "bbox": [ + 130, + 260, + 285, + 381 + ], + "type": "inline_equation", + "content": "v_{\\sigma_d} = \\partial \\sigma_d / \\partial t" + }, + { + "bbox": [ + 130, + 260, + 285, + 381 + ], + "type": "text", + "content": ". Minimizing " + }, + { + "bbox": [ + 130, + 260, + 285, + 381 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{T}}" + }, + { + "bbox": [ + 130, + 260, + 285, + 381 + ], + "type": "text", + "content": " allows us to render the density at " + }, + { + "bbox": [ + 130, + 260, + 285, + 381 + ], + "type": "inline_equation", + "content": "t + \\Delta t" + }, + { + "bbox": [ + 130, + 260, + 285, + 381 + ], + "type": "text", + "content": " aligned with the flow field." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 297, + 118, + 479, + 255 + ], + "blocks": [ + { + "bbox": [ + 297, + 118, + 479, + 255 + ], + "lines": [ + { + "bbox": [ + 297, + 118, + 479, + 255 + ], + "spans": [ + { + "bbox": [ + 297, + 118, + 479, + 255 + ], + "type": "image", + "image_path": "cb83f9a1a3f40142d3ed601b8bede8889a07f14c8d49fafd0a30f20542b6983f.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 290, + 261, + 482, + 383 + ], + "lines": [ + { + "bbox": [ + 290, + 261, + 482, + 383 + ], + "spans": [ + { + "bbox": [ + 290, + 261, + 482, + 383 + ], + "type": "text", + "content": "Fig. 5: Photometric consistency loss. During optimization, we deform each ray from a reference time and pose " + }, + { + "bbox": [ + 290, + 261, + 482, + 383 + ], + "type": "inline_equation", + "content": "(t,P_{t})" + }, + { + "bbox": [ + 290, + 261, + 482, + 383 + ], + "type": "text", + "content": " to different timestamps " + }, + { + "bbox": [ + 290, + 261, + 482, + 383 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 290, + 261, + 482, + 383 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 290, + 261, + 482, + 383 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 290, + 261, + 482, + 383 + ], + "type": "text", + "content": ". Here, we sample " + }, + { + "bbox": [ + 290, + 261, + 482, + 383 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 290, + 261, + 482, + 383 + ], + "type": "text", + "content": " from the neighboring frame times, and we consequently sample an intermediate timestamp " + }, + { + "bbox": [ + 290, + 261, + 482, + 383 + ], + "type": "inline_equation", + "content": "\\gamma \\sim \\mathcal{U}(t,i)" + }, + { + "bbox": [ + 290, + 261, + 482, + 383 + ], + "type": "text", + "content": ". Given each timestamp, we can deform a ray using Eq. 8. The photometric loss " + }, + { + "bbox": [ + 290, + 261, + 482, + 383 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{photo}}" + }, + { + "bbox": [ + 290, + 261, + 482, + 383 + ], + "type": "text", + "content": " is computed based on the color consistency between the deformed ray and the original color, enhancing temporal consistency." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 404, + 482, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 404, + 482, + 453 + ], + "spans": [ + { + "bbox": [ + 130, + 404, + 482, + 453 + ], + "type": "text", + "content": "Transport. A transport equation has been adopted in physics-informed neural networks [7]. By simplifying the Navier-Stokes equation, a transport loss can be defined as: " + }, + { + "bbox": [ + 130, + 404, + 482, + 453 + ], + "type": "inline_equation", + "content": "\\frac{\\partial\\sigma}{\\partial t} +\\mathbf{v}\\cdot \\nabla \\sigma = 0" + }, + { + "bbox": [ + 130, + 404, + 482, + 453 + ], + "type": "text", + "content": ". We leverage the physics-informed regularizers based on the transport equation and the divergence of the velocity:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 249, + 456, + 481, + 485 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 249, + 456, + 481, + 485 + ], + "spans": [ + { + "bbox": [ + 249, + 456, + 481, + 485 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {T}} = \\left(\\frac {\\partial \\sigma_ {d}}{\\partial t} + \\mathbf {v} \\cdot \\nabla \\sigma_ {d}\\right) ^ {2}, \\tag {12}", + "image_path": "1d99e2f48501f43f6f0103551bc963144226d928a717e10635546d02abbbac82.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 490, + 482, + 528 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 490, + 482, + 528 + ], + "spans": [ + { + "bbox": [ + 130, + 490, + 482, + 528 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 490, + 482, + 528 + ], + "type": "inline_equation", + "content": "\\sigma_{d}" + }, + { + "bbox": [ + 130, + 490, + 482, + 528 + ], + "type": "text", + "content": " is the volume density used for volume rendering. This aligns the velocity field with the density field (i.e., Fig. 4), preventing a sudden disappearance or appearance of objects in the dynamic radiance field." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 541, + 262, + 554 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 541, + 262, + 554 + ], + "spans": [ + { + "bbox": [ + 131, + 541, + 262, + 554 + ], + "type": "text", + "content": "3.5 Learning Objectives" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 559, + 482, + 643 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 559, + 482, + 643 + ], + "spans": [ + { + "bbox": [ + 130, + 559, + 482, + 643 + ], + "type": "text", + "content": "Cycle consistency. The cycle consistency is essential to ensure that the trajectories derived from the kinematic field do not introduce misalignments [18]. We use cycle constraints among four timestamps " + }, + { + "bbox": [ + 130, + 559, + 482, + 643 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 559, + 482, + 643 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 559, + 482, + 643 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 559, + 482, + 643 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 559, + 482, + 643 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 130, + 559, + 482, + 643 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 559, + 482, + 643 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 130, + 559, + 482, + 643 + ], + "type": "text", + "content": ". Given a reference time " + }, + { + "bbox": [ + 130, + 559, + 482, + 643 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 559, + 482, + 643 + ], + "type": "text", + "content": ", we randomly sample time " + }, + { + "bbox": [ + 130, + 559, + 482, + 643 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 559, + 482, + 643 + ], + "type": "text", + "content": " from the neighboring frames within a threshold. We define " + }, + { + "bbox": [ + 130, + 559, + 482, + 643 + ], + "type": "inline_equation", + "content": "j = t + 2(i - t)" + }, + { + "bbox": [ + 130, + 559, + 482, + 643 + ], + "type": "text", + "content": ", making it twice the distance from " + }, + { + "bbox": [ + 130, + 559, + 482, + 643 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 559, + 482, + 643 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 130, + 559, + 482, + 643 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 559, + 482, + 643 + ], + "type": "text", + "content": " is from " + }, + { + "bbox": [ + 130, + 559, + 482, + 643 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 559, + 482, + 643 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 130, + 559, + 482, + 643 + ], + "type": "inline_equation", + "content": "\\gamma \\sim \\mathcal{U}(t,i)" + }, + { + "bbox": [ + 130, + 559, + 482, + 643 + ], + "type": "text", + "content": " is an intermediate time, uniformly sampled between " + }, + { + "bbox": [ + 130, + 559, + 482, + 643 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 559, + 482, + 643 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 559, + 482, + 643 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 559, + 482, + 643 + ], + "type": "text", + "content": ". With the time variables, we formulate our loss:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 156, + 649, + 481, + 663 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 156, + 649, + 481, + 663 + ], + "spans": [ + { + "bbox": [ + 156, + 649, + 481, + 663 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {C}} = \\rho (\\mathbf {d} _ {t \\rightarrow i} + \\mathbf {d} _ {i \\rightarrow t}) + \\rho (\\mathbf {d} _ {t \\rightarrow i \\rightarrow j} - \\mathbf {d} _ {t \\rightarrow j}) + \\rho (\\mathbf {d} _ {t \\rightarrow \\gamma \\rightarrow i} - \\mathbf {d} _ {t \\rightarrow i}), \\tag {13}", + "image_path": "4db93c03d8e4f6f381c1dd5be926e1b1a8ce86e5aac900df9fc9d6c0e0b74946.jpg" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 197, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 197, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 197, + 91, + 448, + 102 + ], + "type": "text", + "content": "Regularizing Dynamic Radiance Fields with Kinematic Fields" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "inline_equation", + "content": "\\rho" + }, + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "content": " denotes the generalized chabonnier loss [30]. Here, we abbreviate the displacement " + }, + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "inline_equation", + "content": "d(\\cdot)" + }, + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "content": " (Eq. 8) for conciseness: " + }, + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "inline_equation", + "content": "\\mathbf{d}_{t\\to i}" + }, + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "content": " is shorthand for " + }, + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "inline_equation", + "content": "d(\\mathbf{x},t,i - t)" + }, + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "inline_equation", + "content": "\\mathbf{d}_{t\\to i\\to j}" + }, + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "content": " simplifies " + }, + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "inline_equation", + "content": "d(\\mathbf{x} + \\mathbf{d}_{t\\to i},i,j - i)" + }, + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 152, + 480, + 297 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 152, + 480, + 297 + ], + "spans": [ + { + "bbox": [ + 130, + 152, + 480, + 297 + ], + "type": "text", + "content": "Photometric consistency loss. To enhance temporal coherency, we deform rays into nearby frame times and enhance the consistency among them (Fig. 5). Specifically, given a reference ray " + }, + { + "bbox": [ + 130, + 152, + 480, + 297 + ], + "type": "inline_equation", + "content": "\\mathbf{r}_t" + }, + { + "bbox": [ + 130, + 152, + 480, + 297 + ], + "type": "text", + "content": " at time " + }, + { + "bbox": [ + 130, + 152, + 480, + 297 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 152, + 480, + 297 + ], + "type": "text", + "content": ", a pixel color " + }, + { + "bbox": [ + 130, + 152, + 480, + 297 + ], + "type": "inline_equation", + "content": "\\hat{C}(\\mathbf{r}_t, t)" + }, + { + "bbox": [ + 130, + 152, + 480, + 297 + ], + "type": "text", + "content": " can be obtained using the volume rendering (Eq. 4). In addition to the reference ray, we apply ray deformation based on the predicted kinematic quantities. As in cycle consistency (Eq. 13), we sample time " + }, + { + "bbox": [ + 130, + 152, + 480, + 297 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 152, + 480, + 297 + ], + "type": "text", + "content": " from neighboring frames, and time " + }, + { + "bbox": [ + 130, + 152, + 480, + 297 + ], + "type": "inline_equation", + "content": "\\gamma \\sim \\mathcal{U}(t, i)" + }, + { + "bbox": [ + 130, + 152, + 480, + 297 + ], + "type": "text", + "content": " between times " + }, + { + "bbox": [ + 130, + 152, + 480, + 297 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 152, + 480, + 297 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 152, + 480, + 297 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 152, + 480, + 297 + ], + "type": "text", + "content": ". Different from existing approaches [10, 18], we deform rays with kinematic quantities using Taylor approximation. The deforming process results in two deformed rays " + }, + { + "bbox": [ + 130, + 152, + 480, + 297 + ], + "type": "inline_equation", + "content": "\\mathbf{r}_{t \\to i}" + }, + { + "bbox": [ + 130, + 152, + 480, + 297 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 152, + 480, + 297 + ], + "type": "inline_equation", + "content": "\\mathbf{r}_{t \\to \\gamma}" + }, + { + "bbox": [ + 130, + 152, + 480, + 297 + ], + "type": "text", + "content": ", and two rendered colors " + }, + { + "bbox": [ + 130, + 152, + 480, + 297 + ], + "type": "inline_equation", + "content": "\\hat{C}(\\mathbf{r}_{t \\to i}, i)" + }, + { + "bbox": [ + 130, + 152, + 480, + 297 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 152, + 480, + 297 + ], + "type": "inline_equation", + "content": "\\hat{C}(\\mathbf{r}_{t \\to \\gamma}, \\gamma)" + }, + { + "bbox": [ + 130, + 152, + 480, + 297 + ], + "type": "text", + "content": ", correspondingly. Based on the deformed rays, we apply the photometric consistency loss, comparing the rendered color with the ground-truth color. The loss function is formulated as:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 141, + 305, + 480, + 335 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 305, + 480, + 335 + ], + "spans": [ + { + "bbox": [ + 141, + 305, + 480, + 335 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {\\text {p h o t o}} = \\| \\hat {C} (\\mathbf {r} _ {t}, t) - C (\\mathbf {r} _ {t}, t) \\| _ {2} ^ {2} \\tag {14} \\\\ + \\lambda \\alpha (\\mathbf {r} _ {t}) \\rho (\\hat {C} (\\mathbf {r} _ {t \\rightarrow i}, i) - C (\\mathbf {r} _ {t}, t)) + \\lambda \\alpha (\\mathbf {r} _ {t}) \\rho (\\hat {C} (\\mathbf {r} _ {t \\rightarrow \\gamma}, \\gamma) - C (\\mathbf {r} _ {t}, t)), \\\\ \\end{array}", + "image_path": "cd0867f95cbd88d3babedf196fc51df5be52e88f3f3f6d81eefb7400950fab46.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 343, + 479, + 378 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 343, + 479, + 378 + ], + "spans": [ + { + "bbox": [ + 130, + 343, + 479, + 378 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 343, + 479, + 378 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 130, + 343, + 479, + 378 + ], + "type": "text", + "content": " is a loss weight for the temporal consistency, and " + }, + { + "bbox": [ + 130, + 343, + 479, + 378 + ], + "type": "inline_equation", + "content": "\\alpha (\\mathbf{r}_t)" + }, + { + "bbox": [ + 130, + 343, + 479, + 378 + ], + "type": "text", + "content": " is the sum of rendering weights of the dynamic densities, representing the dynamic likelihood of the ray " + }, + { + "bbox": [ + 130, + 343, + 479, + 378 + ], + "type": "inline_equation", + "content": "\\mathbf{r}_t" + }, + { + "bbox": [ + 130, + 343, + 479, + 378 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 380, + 480, + 427 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 380, + 480, + 427 + ], + "spans": [ + { + "bbox": [ + 130, + 380, + 480, + 427 + ], + "type": "text", + "content": "Final objective. Our final objective function, denoted as " + }, + { + "bbox": [ + 130, + 380, + 480, + 427 + ], + "type": "inline_equation", + "content": "\\mathcal{L}" + }, + { + "bbox": [ + 130, + 380, + 480, + 427 + ], + "type": "text", + "content": ", is the weighted sum of the various loss terms discussed previously. To keep things simple with the variety of loss functions, we symbolize the losses to " + }, + { + "bbox": [ + 130, + 380, + 480, + 427 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{photo}}" + }, + { + "bbox": [ + 130, + 380, + 480, + 427 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 380, + 480, + 427 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{kinematic}}" + }, + { + "bbox": [ + 130, + 380, + 480, + 427 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 380, + 480, + 427 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{reg}}" + }, + { + "bbox": [ + 130, + 380, + 480, + 427 + ], + "type": "text", + "content": ". This results in the final loss:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 229, + 436, + 479, + 449 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 229, + 436, + 479, + 449 + ], + "spans": [ + { + "bbox": [ + 229, + 436, + 479, + 449 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathbb {E} _ {\\mathbf {r} _ {t}} \\left[ \\mathcal {L} _ {\\text {p h o t o}} + \\mathcal {L} _ {\\text {k i n e m a t i c}} + \\mathcal {L} _ {\\text {r e g}} \\right]. \\tag {15}", + "image_path": "22a2d901692ba906940f7cf635948b19c0d039c4fbfe400a93c0dc0a1b817b34.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 456, + 479, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 456, + 479, + 529 + ], + "spans": [ + { + "bbox": [ + 131, + 456, + 479, + 529 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 131, + 456, + 479, + 529 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{kinematic}}" + }, + { + "bbox": [ + 131, + 456, + 479, + 529 + ], + "type": "text", + "content": " is the weighted sum of the kinematic integrity " + }, + { + "bbox": [ + 131, + 456, + 479, + 529 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{K}}" + }, + { + "bbox": [ + 131, + 456, + 479, + 529 + ], + "type": "text", + "content": ", transport " + }, + { + "bbox": [ + 131, + 456, + 479, + 529 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{T}}" + }, + { + "bbox": [ + 131, + 456, + 479, + 529 + ], + "type": "text", + "content": ", and rigidity " + }, + { + "bbox": [ + 131, + 456, + 479, + 529 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{R}}" + }, + { + "bbox": [ + 131, + 456, + 479, + 529 + ], + "type": "text", + "content": ". Note that " + }, + { + "bbox": [ + 131, + 456, + 479, + 529 + ], + "type": "inline_equation", + "content": "\\mathbb{E}_{\\mathbf{r}_t}[\\mathcal{L}_{\\mathrm{kinematic}}]" + }, + { + "bbox": [ + 131, + 456, + 479, + 529 + ], + "type": "text", + "content": " is computed with randomly sampled points along the ray " + }, + { + "bbox": [ + 131, + 456, + 479, + 529 + ], + "type": "inline_equation", + "content": "\\mathbf{r}_t" + }, + { + "bbox": [ + 131, + 456, + 479, + 529 + ], + "type": "text", + "content": " from each minibatch. The additional regularization " + }, + { + "bbox": [ + 131, + 456, + 479, + 529 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{reg}}" + }, + { + "bbox": [ + 131, + 456, + 479, + 529 + ], + "type": "text", + "content": " includes the skewed entropy " + }, + { + "bbox": [ + 131, + 456, + 479, + 529 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_b" + }, + { + "bbox": [ + 131, + 456, + 479, + 529 + ], + "type": "text", + "content": ", the cycle consistency " + }, + { + "bbox": [ + 131, + 456, + 479, + 529 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{C}}" + }, + { + "bbox": [ + 131, + 456, + 479, + 529 + ], + "type": "text", + "content": ", total-variation loss [5,6], distortion loss [4], monocular depth supervision [18], normal regularization [14,32] and optical flow supervision [18]." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 545, + 224, + 559 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 545, + 224, + 559 + ], + "spans": [ + { + "bbox": [ + 131, + 545, + 224, + 559 + ], + "type": "text", + "content": "4 Experiment" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 569, + 479, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 569, + 479, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 569, + 479, + 665 + ], + "type": "text", + "content": "Dataset. We use the NVIDIA dynamic view synthesis dataset (NDVS) [36] to evaluate the effectiveness of our model. For the monocular setup, we adopt the preprocessing steps proposed by NSFF [18] and DynamicNeRF [10]. This dataset has 8 scenes and is captured with 12 synchronized cameras. When used for a monocular video reconstruction task, the dataset shows relatively low effective multi-view factors (EMFs) [11], compared to other monocular datasets; this means that the dataset is more challenging to reconstruct due to the less camera motion and faster-moving objects." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "text", + "content": "W. Im et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 210 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 210 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 210 + ], + "type": "text", + "content": "There are two widely used monocular NDVS protocols: 12-frames [10] and 24-frames [18]. Most papers opt for either protocol; however, we include both protocols for rigorous evaluation. Meanwhile, 24-frames-sparse is proposed by us, as neither existing protocol provides novel time evaluation; we leave out odd-numbered frames at unseen times for testing. For the 'Dynamic Only' marks in tables, we use the dynamic mask to evaluate only in terms of dynamic pixels; please refer to [10,18] for how to generate dynamic masks. We detail the protocols below." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 221, + 479, + 340 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 138, + 221, + 479, + 269 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 221, + 479, + 269 + ], + "spans": [ + { + "bbox": [ + 138, + 221, + 479, + 269 + ], + "type": "text", + "content": "- 24-frames: this sequence is configured with 24 training frames, and 264 test frames from unseen multi-views. For a fair comparison, we utilize the dynamic mask, optical flow, and monocular depth, extracted by the same model with the prior work [18]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 269, + 479, + 304 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 269, + 479, + 304 + ], + "spans": [ + { + "bbox": [ + 138, + 269, + 479, + 304 + ], + "type": "text", + "content": "- 24-frames-sparse: we reduce the 24-frames sequence into 12 frames by sampling the even-numbered frames for training. This sparse sampling allows for testing novel view synthesis at both seen times and novel times." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 305, + 479, + 340 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 305, + 479, + 340 + ], + "spans": [ + { + "bbox": [ + 138, + 305, + 479, + 340 + ], + "type": "text", + "content": "- 12-frames: this sequence has 12 training and 12 testing samples. We use the preprocessed dynamic mask, optical flow, and monocular depth downloadable from the code repository of DynamicNeRF [10]." + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 132, + 358, + 279, + 370 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 358, + 279, + 370 + ], + "spans": [ + { + "bbox": [ + 132, + 358, + 279, + 370 + ], + "type": "text", + "content": "4.1 Implementation Details" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 378, + 479, + 435 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 378, + 479, + 435 + ], + "spans": [ + { + "bbox": [ + 130, + 378, + 479, + 435 + ], + "type": "text", + "content": "Fields configuration. Our radiance and kinematic fields are configured with HexPlane [5] and TensoRF [6]. After features are sampled from the point queries, small MLP decoders process the features to yield colors and kinematic outputs. For density outputs, we use a linear layer following the practices in [5,6]. Further details on field configurations are available in the supplementary material." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 438, + 479, + 530 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 438, + 479, + 530 + ], + "spans": [ + { + "bbox": [ + 130, + 438, + 479, + 530 + ], + "type": "text", + "content": "Training. We use 70,000 steps to optimize our models, which takes about 7 hours with one NVIDIA V100 or A100 GPU. We initialize the static field for 5,000 steps, taking approximately 5 minutes. After finishing the initialization of the static field, we jointly train the fields: dynamic, static, and kinematic fields. During the joint training, we lower the learning rate of the static field by the factor of 0.1 at predefined iterations to prevent dynamic objects from being portrayed as static objects. Additional training details are provided in the supplementary material." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 533, + 479, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 479, + 567 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 479, + 567 + ], + "type": "text", + "content": "Evaluation metrics. We report peak signal-to-noise ratio (PSNR), structural similarity (SSIM), and perceptual similarity (LPIPS). These three metrics have been used as standard in prior studies to evaluate the quality of renderings." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 570, + 485, + 663 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 485, + 663 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 485, + 663 + ], + "type": "text", + "content": "Base architecture (HexPlane). Since our model is based on tensor-decomposed radiance fields, we compare our models with HexPlane [5]. The baseline model is the dynamic radiance field " + }, + { + "bbox": [ + 130, + 570, + 485, + 663 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\mathrm{DY}}" + }, + { + "bbox": [ + 130, + 570, + 485, + 663 + ], + "type": "text", + "content": ", without the static radiance field " + }, + { + "bbox": [ + 130, + 570, + 485, + 663 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\mathrm{ST}}" + }, + { + "bbox": [ + 130, + 570, + 485, + 663 + ], + "type": "text", + "content": " and the kinematic field " + }, + { + "bbox": [ + 130, + 570, + 485, + 663 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\mathrm{K}}" + }, + { + "bbox": [ + 130, + 570, + 485, + 663 + ], + "type": "text", + "content": ". For a fair comparison, we use monocular depth supervision both to the baseline model and our models, although the original HexPlane model does not support the supervision from monocular depth. Since the base HexPlane model does not have an auxiliary flow field, we do not use the optical flow supervision and the photometric consistency loss." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 196, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 196, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 196, + 91, + 447, + 102 + ], + "type": "text", + "content": "Regularizing Dynamic Radiance Fields with Kinematic Fields" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 134, + 146, + 479, + 215 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 479, + 137 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 479, + 137 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 479, + 137 + ], + "type": "text", + "content": "Table 1: Ablation study on Balloon1 scene of the NDVS dataset (24-frames-sparse). Please refer to the supplementary material for ablation results on other NDVS scenes." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 134, + 146, + 479, + 215 + ], + "lines": [ + { + "bbox": [ + 134, + 146, + 479, + 215 + ], + "spans": [ + { + "bbox": [ + 134, + 146, + 479, + 215 + ], + "type": "table", + "html": "
MethodsNovel TimesSeen Times
FullDynamic OnlyFullDynamic Only
PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓
HexPlane [5]19.840.6680.14518.120.4190.25118.930.6490.13717.740.4290.202
Ours25.670.8650.06421.160.6250.17025.800.8670.05021.350.6320.121
w/o Higher order motion24.360.8420.07719.930.5560.20024.630.8460.06120.220.5720.145
w/o Lkinematic24.960.8530.06820.150.5680.18525.300.8580.05220.650.5940.129
w/o Kinematic integrity25.460.8620.06620.780.6070.17925.570.8650.05121.020.6200.126
w/o Rigidity25.110.8560.06620.400.5820.18125.430.8600.05120.820.6030.123
", + "image_path": "3a9c0b5e98e8687225d9e745d60a4d13ebf39ec4c9c3cbd75c7550a0e2412b3f.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 140, + 225, + 474, + 428 + ], + "blocks": [ + { + "bbox": [ + 140, + 225, + 474, + 428 + ], + "lines": [ + { + "bbox": [ + 140, + 225, + 474, + 428 + ], + "spans": [ + { + "bbox": [ + 140, + 225, + 474, + 428 + ], + "type": "image", + "image_path": "16ac34041a8548cbf376d994f7713df7efa133d9447f29fe78e0eca2908a444d.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 437, + 482, + 481 + ], + "lines": [ + { + "bbox": [ + 130, + 437, + 482, + 481 + ], + "spans": [ + { + "bbox": [ + 130, + 437, + 482, + 481 + ], + "type": "text", + "content": "Fig. 6: Comparative visualization of kinematic fields. The first row shows RGB values rendered from our dynamic radiance field, showing PSNR values from a ground truth RGB image. From the second to the last row, we show flow (i.e., displacement), velocity, and acceleration." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 508, + 317, + 519 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 508, + 317, + 519 + ], + "spans": [ + { + "bbox": [ + 132, + 508, + 317, + 519 + ], + "type": "text", + "content": "4.2 Evaluation of Kinematic Fields" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 533, + 482, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 482, + 593 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 482, + 593 + ], + "type": "text", + "content": "Ablation of kinematic losses. To show the effectiveness of our approach, we conduct a comprehensive ablation study on the Balloon1 scene of the NDVS dataset, focusing on kinematic loss components. In Table 1, we show ablative results from seen and unseen frame times. Our findings indicate that the model achieves optimal performance when all components are integrated." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": "Fig. 6 delves into the kinematic field analysis. Omitting higher order motion restricts the model to utilizing only the velocity field, which proves insufficient for capturing non-linear motion. This limitation hinders the model's ability to minimize kinematic losses, leading to a suboptimal flow representation. In the absence of " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{kinematic}}" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": ", the radiance field suffers from reduced PSNR scores, which can be attributed to the physically incorrect motion and geometry. Compared" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "type": "text", + "content": "W. Im et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 136, + 118, + 218, + 194 + ], + "blocks": [ + { + "bbox": [ + 136, + 118, + 218, + 194 + ], + "lines": [ + { + "bbox": [ + 136, + 118, + 218, + 194 + ], + "spans": [ + { + "bbox": [ + 136, + 118, + 218, + 194 + ], + "type": "image", + "image_path": "a9c9a9ded169c82671c72721386a643c1ca4977fcaf5df7036ef34668ee0c877.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 207, + 480, + 251 + ], + "lines": [ + { + "bbox": [ + 130, + 207, + 480, + 251 + ], + "spans": [ + { + "bbox": [ + 130, + 207, + 480, + 251 + ], + "type": "text", + "content": "Fig. 7: Motion Order. Reconstruction Accuracy for the Balloon1 Scene (24-framesparse): The " + }, + { + "bbox": [ + 130, + 207, + 480, + 251 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 130, + 207, + 480, + 251 + ], + "type": "text", + "content": "-axis of each plot categorizes maximum motion order as follows: velocity (1), acceleration (2), jerk (3), snap (4), and crackle (5). The metric utilized for each plot is specified at the top." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 222, + 118, + 304, + 194 + ], + "blocks": [ + { + "bbox": [ + 222, + 118, + 304, + 194 + ], + "lines": [ + { + "bbox": [ + 222, + 118, + 304, + 194 + ], + "spans": [ + { + "bbox": [ + 222, + 118, + 304, + 194 + ], + "type": "image", + "image_path": "86c577485099033c2c2c217354d6aafa70ae76281e8ab4026870c8709c3324f2.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 309, + 118, + 389, + 194 + ], + "blocks": [ + { + "bbox": [ + 309, + 118, + 389, + 194 + ], + "lines": [ + { + "bbox": [ + 309, + 118, + 389, + 194 + ], + "spans": [ + { + "bbox": [ + 309, + 118, + 389, + 194 + ], + "type": "image", + "image_path": "267eb84c3692aa30abeb538a785ae608f57c0287c9af9fe2a578ec64ace0b96a.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 396, + 118, + 476, + 194 + ], + "blocks": [ + { + "bbox": [ + 396, + 118, + 476, + 194 + ], + "lines": [ + { + "bbox": [ + 396, + 118, + 476, + 194 + ], + "spans": [ + { + "bbox": [ + 396, + 118, + 476, + 194 + ], + "type": "image", + "image_path": "11267c13c409d4a187ce802bb658ffac45c97ceb3f45404e71b75d9049f67ff8.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 135, + 297, + 261, + 392 + ], + "blocks": [ + { + "bbox": [ + 131, + 262, + 479, + 285 + ], + "lines": [ + { + "bbox": [ + 131, + 262, + 479, + 285 + ], + "spans": [ + { + "bbox": [ + 131, + 262, + 479, + 285 + ], + "type": "text", + "content": "Table 2: Quantitative comparison on NDVS. This table reports average results on all scenes. Please refer to the supplementary material for per-scene results." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 135, + 297, + 261, + 392 + ], + "lines": [ + { + "bbox": [ + 135, + 297, + 261, + 392 + ], + "spans": [ + { + "bbox": [ + 135, + 297, + 261, + 392 + ], + "type": "table", + "html": "
(a) 12-frames
MethodsFull
PSNR↑LPIPS↓
NeRF [23]*21.990.250
NSFF [18]*24.330.199
DynamicNeRF [10]*26.100.082
HyperNeRF [26]*17.600.367
RoDynRF [20]*25.890.065
Ours26.330.071
*Results reported in [20]
", + "image_path": "6cf082eaf1def2a19ca2a7f4393819329df9c0960da3a1b3b8f6275536c6f334.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 263, + 297, + 479, + 391 + ], + "blocks": [ + { + "bbox": [ + 263, + 297, + 479, + 391 + ], + "lines": [ + { + "bbox": [ + 263, + 297, + 479, + 391 + ], + "spans": [ + { + "bbox": [ + 263, + 297, + 479, + 391 + ], + "type": "table", + "html": "
(b) 24-frames
MethodsFullDynamic Only
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
NeRF [23]*24.900.8930.09816.980.5320.314
Luo et al. [21]*21.370.7460.14116.970.5300.207
Yoon et al. [36]*21.780.7610.12717.340.5470.200
NSFF [18]*28.190.9280.04521.910.7580.097
NSFF [18]28.040.9270.04521.340.7400.111
Ours28.780.9360.04222.130.7740.094
*Results reported in [18]
", + "image_path": "1c4093f1fa36d0acc85aef7d8772e2a7fd8dd6611340e6b1d811bb6c1c7ee30d.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 414, + 480, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 414, + 480, + 437 + ], + "spans": [ + { + "bbox": [ + 130, + 414, + 480, + 437 + ], + "type": "text", + "content": "to the full model, the model without the rigidity loss results in noisy, non-rigid motion, resulting in incorrect motion and the worse reconstruction accuracy." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 437, + 482, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 437, + 482, + 557 + ], + "spans": [ + { + "bbox": [ + 130, + 437, + 482, + 557 + ], + "type": "text", + "content": "Motion order. Our study represents motion through various kinematic quantities. Adjusting the kinematic field to include different orders of motion, we can model our system using just velocity [7, 8], or extend to higher orders – fourth (snap) and fifth (crackle) – by extending the Taylor series (Eq. 8) and the kinematic integrity (Eq. 9) equations. Fig. 7 illustrates the influence of motion order on reconstruction accuracy for the NDVS dataset. As motion order increases, the reconstruction accuracy first improves, peaking at jerk, but further orders degrades the accuracy. We conjecture that the augmentation of model complexity with higher orders of kinematic quantities may render the model more susceptible to overfitting." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 574, + 320, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 574, + 320, + 586 + ], + "spans": [ + { + "bbox": [ + 132, + 574, + 320, + 586 + ], + "type": "text", + "content": "4.3 Comparison with SOTA models" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 594, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 665 + ], + "type": "text", + "content": "Quantitative results. NSFF [18] is a dynamic neural radiance field model, which utilizes scene-flow fields for the regularization of a dynamic radiance field. Since NSFF is a neural model, the network training takes about 2 days on two V100 GPUs (96 GPU hours), using the official code by the authors. Note that optimizing a scene using our method takes about 7 hours per scene, which makes ours 12 times faster than NSFF in terms of GPU hours. We compare our results" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 197, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 197, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 197, + 91, + 448, + 102 + ], + "type": "text", + "content": "Regularizing Dynamic Radiance Fields with Kinematic Fields" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 113, + 479, + 260 + ], + "blocks": [ + { + "bbox": [ + 133, + 113, + 479, + 260 + ], + "lines": [ + { + "bbox": [ + 133, + 113, + 479, + 260 + ], + "spans": [ + { + "bbox": [ + 133, + 113, + 479, + 260 + ], + "type": "image", + "image_path": "a0b747be6e9416d394293e76f6342c55d31ec714da4f924512e4d287da88f976.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 269, + 480, + 292 + ], + "lines": [ + { + "bbox": [ + 132, + 269, + 480, + 292 + ], + "spans": [ + { + "bbox": [ + 132, + 269, + 480, + 292 + ], + "type": "text", + "content": "Fig. 8: Qualitative results on NDVS (12-frames). The second and fourth rows illustrate the overlay of the synthesized and ground truth views of the scene." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 313, + 480, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 313, + 480, + 348 + ], + "spans": [ + { + "bbox": [ + 130, + 313, + 480, + 348 + ], + "type": "text", + "content": "with NSFF [18] in Table 2b. In the table, our method shows better accuracy in terms of all metrics. In addition, we report results on NDVS (12-frames) in Table 2a." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 349, + 480, + 433 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 349, + 480, + 433 + ], + "spans": [ + { + "bbox": [ + 130, + 349, + 480, + 433 + ], + "type": "text", + "content": "Qualitative results. Fig. 8 illustrates qualitative results, including the synthesized images, ground truth, and their overlaid images. As shown in the first and third rows, our method improves the visual quality of the rendered views. Additionally, the overlaid images demonstrate the geometrical accuracy of our model, showcasing its better alignment with the ground truth. This enhanced performance is attributed to the kinematics and physics incorporated in our methodology, which is fundamental to the improvement." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 449, + 219, + 462 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 449, + 219, + 462 + ], + "spans": [ + { + "bbox": [ + 132, + 449, + 219, + 462 + ], + "type": "text", + "content": "5 Conclusion" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 472, + 480, + 580 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 472, + 480, + 580 + ], + "spans": [ + { + "bbox": [ + 130, + 472, + 480, + 580 + ], + "type": "text", + "content": "In this paper, we have presented a novel method to learn dynamic radiance fields with kinematic fields. We propose the kinematic field, where kinematic quantities are learned by joint learning with the radiance fields. Using the kinematic quantities from the kinematic fields, we propose physics-based loss functions that underpin our method. Lastly, we quantitatively and qualitatively show the effectiveness of our method on the monocular dataset. We believe that this work not only introduces the potential of incorporating kinematics into the learning of radiance fields but also sets a foundation for future direction in physics-based dynamic scene understanding." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 593, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 593, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 593, + 481, + 665 + ], + "type": "text", + "content": "Acknowledgment. This project was partly supported by the Naver Cloud Corporation. Additionally, this work received support from the Institute of Information & communications Technology Planning & Evaluation (IITP) grant (RS-2023-00237965) and the National Research Foundation of Korea (NRF) grant (No. RS-2023-00208506(2024)), funded by the Korea government (MSIT). Sung-Eui Yoon is a corresponding author." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "text", + "content": "W. Im et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 197, + 126 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 197, + 126 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 197, + 126 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 134, + 137, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 138, + 137, + 480, + 169 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 137, + 480, + 169 + ], + "spans": [ + { + "bbox": [ + 138, + 137, + 480, + 169 + ], + "type": "text", + "content": "1. Attal, B., Huang, J.B., Richardt, C., Zollhoefer, M., Kopf, J., O'Toole, M., Kim, C.: Hyperreel: High-fidelity 6-dof video with ray-conditioned sampling. arXiv preprint arXiv:2301.02238 (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 170, + 480, + 201 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 170, + 480, + 201 + ], + "spans": [ + { + "bbox": [ + 138, + 170, + 480, + 201 + ], + "type": "text", + "content": "2. Baieri, D., Esposito, S., Maggioli, F., Rodola, E.: Fluid dynamics network: Topology-agnostic 4d reconstruction via fluid dynamics priors. arXiv preprint arXiv:2303.09871 (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 202, + 480, + 233 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 202, + 480, + 233 + ], + "spans": [ + { + "bbox": [ + 138, + 202, + 480, + 233 + ], + "type": "text", + "content": "3. Baker, S., Scharstein, D., Lewis, J., Roth, S., Black, M.J., Szeliski, R.: A database and evaluation methodology for optical flow. International journal of computer vision 92, 1-31 (2011)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 234, + 480, + 277 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 234, + 480, + 277 + ], + "spans": [ + { + "bbox": [ + 138, + 234, + 480, + 277 + ], + "type": "text", + "content": "4. Barron, J.T., Mildenhall, B., Tancik, M., Hedman, P., Martin-Brualla, R., Srinivasan, P.P.: Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5855-5864 (2021)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 277, + 480, + 298 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 277, + 480, + 298 + ], + "spans": [ + { + "bbox": [ + 138, + 277, + 480, + 298 + ], + "type": "text", + "content": "5. Cao, A., Johnson, J.: Hexplane: A fast representation for dynamic scenes. arXiv preprint arXiv:2301.09632 (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 299, + 480, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 299, + 480, + 331 + ], + "spans": [ + { + "bbox": [ + 138, + 299, + 480, + 331 + ], + "type": "text", + "content": "6. Chen, A., Xu, Z., Geiger, A., Yu, J., Su, H.: Tensorf: Tensorial radiance fields. In: Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXXII. pp. 333-350. Springer (2022)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 331, + 480, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 331, + 480, + 363 + ], + "spans": [ + { + "bbox": [ + 138, + 331, + 480, + 363 + ], + "type": "text", + "content": "7. Chu, M., Liu, L., Zheng, Q., Franz, E., Seidel, H.P., Theobalt, C., Zayer, R.: Physics informed neural fields for smoke reconstruction with sparse data. ACM Transactions on Graphics (TOG) 41(4), 1-14 (2022)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 364, + 480, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 364, + 480, + 396 + ], + "spans": [ + { + "bbox": [ + 138, + 364, + 480, + 396 + ], + "type": "text", + "content": "8. Du, Y., Zhang, Y., Yu, H.X., Tenenbaum, J.B., Wu, J.: Neural radiance flow for 4d view synthesis and video processing. In: 2021 IEEE/CVF International Conference on Computer Vision (ICCV). pp. 14304-14314. IEEE Computer Society (2021)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 396, + 480, + 427 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 396, + 480, + 427 + ], + "spans": [ + { + "bbox": [ + 138, + 396, + 480, + 427 + ], + "type": "text", + "content": "9. Fang, J., Yi, T., Wang, X., Xie, L., Zhang, X., Liu, W., Niešner, M., Tian, Q.: Fast dynamic radiance fields with time-aware neural voxels. In: SIGGRAPH Asia 2022 Conference Papers. pp. 1-9 (2022)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 428, + 480, + 460 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 428, + 480, + 460 + ], + "spans": [ + { + "bbox": [ + 138, + 428, + 480, + 460 + ], + "type": "text", + "content": "10. Gao, C., Saraf, A., Kopf, J., Huang, J.B.: Dynamic view synthesis from dynamic monocular video. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 5712-5721 (October 2021)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 134, + 460, + 480, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 460, + 480, + 503 + ], + "spans": [ + { + "bbox": [ + 134, + 460, + 480, + 503 + ], + "type": "text", + "content": "11. Gao, H., Li, R., Tulsiani, S., Russell, B., Kanazawa, A.: Monocular dynamic view synthesis: A reality check. In: Oh, A.H., Agarwal, A., Belgrave, D., Cho, K. (eds.) Advances in Neural Information Processing Systems (2022), https://openreview.net/forum?id=pCrB8orUkSq" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 134, + 504, + 480, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 504, + 480, + 536 + ], + "spans": [ + { + "bbox": [ + 134, + 504, + 480, + 536 + ], + "type": "text", + "content": "12. Hart-Smith, L.: Application of the strain invariant failure theory (sift) to metals and fiber-polymer composites. Philosophical Magazine 90(31-32), 4263-4331 (2010)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 134, + 536, + 480, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 536, + 480, + 557 + ], + "spans": [ + { + "bbox": [ + 134, + 536, + 480, + 557 + ], + "type": "text", + "content": "13. Hess, R., Baker, C., Zihl, J.: The\" motion-blind\" patient: low-level spatial and temporal filters. Journal of Neuroscience 9(5), 1628-1640 (1989)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 134, + 558, + 480, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 558, + 480, + 590 + ], + "spans": [ + { + "bbox": [ + 134, + 558, + 480, + 590 + ], + "type": "text", + "content": "14. Jin, H., Liu, I., Xu, P., Zhang, X., Han, S., Bi, S., Zhou, X., Xu, Z., Su, H.: Tensoroir: Tensorial inverse rendering. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2023)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 134, + 590, + 480, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 590, + 480, + 621 + ], + "spans": [ + { + "bbox": [ + 134, + 590, + 480, + 621 + ], + "type": "text", + "content": "15. Kajiya, J.T., Von Herzen, B.P.: Ray tracing volume densities. In: Proceedings of the 11th Annual Conference on Computer Graphics and Interactive Techniques. p. 165-174 (1984)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 134, + 622, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 622, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 134, + 622, + 480, + 665 + ], + "type": "text", + "content": "16. Li, T., Slavcheva, M., Zollhoefer, M., Green, S., Lassner, C., Kim, C., Schmidt, T., Lovegrove, S., Goesele, M., Newcombe, R., et al.: Neural 3d video synthesis from multi-view video. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5521-5531 (2022)" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 197, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 197, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 197, + 91, + 447, + 102 + ], + "type": "text", + "content": "Regularizing Dynamic Radiance Fields with Kinematic Fields" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 481, + 160 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 481, + 160 + ], + "type": "text", + "content": "17. Li, Z., Müller, T., Evans, A., Taylor, R.H., Unberath, M., Liu, M.Y., Lin, C.H.: Neuralangelo: High-fidelity neural surface reconstruction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8456-8465 (2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 161, + 481, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 161, + 481, + 194 + ], + "spans": [ + { + "bbox": [ + 132, + 161, + 481, + 194 + ], + "type": "text", + "content": "18. Li, Z., Niklaus, S., Snavely, N., Wang, O.: Neural scene flow fields for space-time view synthesis of dynamic scenes. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6498-6508 (2021)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 194, + 481, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 194, + 481, + 226 + ], + "spans": [ + { + "bbox": [ + 132, + 194, + 481, + 226 + ], + "type": "text", + "content": "19. Li, Z., Wang, Q., Cole, F., Tucker, R., Snavely, N.: Dynibar: Neural dynamic image-based rendering. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4273-4284 (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 227, + 481, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 227, + 481, + 270 + ], + "spans": [ + { + "bbox": [ + 132, + 227, + 481, + 270 + ], + "type": "text", + "content": "20. Liu, Y.L., Gao, C., Meuleman, A., Tseng, H.Y., Saraf, A., Kim, C., Chuang, Y.Y., Kopf, J., Huang, J.B.: Robust dynamic radiance fields. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 13-23 (June 2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 271, + 481, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 271, + 481, + 293 + ], + "spans": [ + { + "bbox": [ + 132, + 271, + 481, + 293 + ], + "type": "text", + "content": "21. Luo, X., Huang, J.B., Szeliski, R., Matzen, K., Kopf, J.: Consistent video depth estimation. ACM Transactions on Graphics (ToG) 39(4), 71-1 (2020)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 293, + 481, + 325 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 293, + 481, + 325 + ], + "spans": [ + { + "bbox": [ + 132, + 293, + 481, + 325 + ], + "type": "text", + "content": "22. Max, N., Chen, M.: Local and global illumination in the volume rendering integral. Tech. rep., Lawrence Livermore National Lab.(LLNL), Livermore, CA (United States) (2005)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 325, + 481, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 325, + 481, + 357 + ], + "spans": [ + { + "bbox": [ + 132, + 325, + 481, + 357 + ], + "type": "text", + "content": "23. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. In: ECCV (2020)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 358, + 481, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 358, + 481, + 402 + ], + "spans": [ + { + "bbox": [ + 132, + 358, + 481, + 402 + ], + "type": "text", + "content": "24. Orts-Escalano, S., Rhemann, C., Fanello, S., Chang, W., Kowdle, A., Degtyarev, Y., Kim, D., Davidson, P.L., Khamis, S., Dou, M., et al.: Holoportation: Virtual 3d teleportation in real-time. In: Proceedings of the 29th annual symposium on user interface software and technology. pp. 741-754 (2016)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 403, + 481, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 403, + 481, + 435 + ], + "spans": [ + { + "bbox": [ + 132, + 403, + 481, + 435 + ], + "type": "text", + "content": "25. Park, K., Sinha, U., Barron, J.T., Bouaziz, S., Goldman, D.B., Seitz, S.M., Martin-Brualla, R.: Nerfies: Deformable neural radiance fields. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5865-5874 (2021)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 435, + 481, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 435, + 481, + 468 + ], + "spans": [ + { + "bbox": [ + 132, + 435, + 481, + 468 + ], + "type": "text", + "content": "26. Park, K., Sinha, U., Hedman, P., Barron, J.T., Bouaziz, S., Goldman, D.B., Martin-Brualla, R., Seitz, S.M.: Hypernerf: A higher-dimensional representation for topologically varying neural radiance fields. arXiv preprint arXiv:2106.13228 (2021)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 468, + 481, + 501 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 468, + 481, + 501 + ], + "spans": [ + { + "bbox": [ + 132, + 468, + 481, + 501 + ], + "type": "text", + "content": "27. Pumarola, A., Corona, E., Pons-Moll, G., Moreno-Noguer, F.: D-nerf: Neural radiance fields for dynamic scenes. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10318-10327 (2021)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 501, + 481, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 501, + 481, + 533 + ], + "spans": [ + { + "bbox": [ + 132, + 501, + 481, + 533 + ], + "type": "text", + "content": "28. Ramasinghe, S., Shevchenko, V., Avraham, G., Hengel, A.V.D.: Bali-rf: Bandlimited radiance fields for dynamic scene modeling. arXiv preprint arXiv:2302.13543 (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 534, + 481, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 534, + 481, + 566 + ], + "spans": [ + { + "bbox": [ + 132, + 534, + 481, + 566 + ], + "type": "text", + "content": "29. Soria, J., Sondergaard, R., Cantwell, B., Chong, M., Perry, A.: A study of the fine-scale motions of incompressible time-developing mixing layers. Physics of Fluids 6(2), 871-884 (1994)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 567, + 481, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 567, + 481, + 599 + ], + "spans": [ + { + "bbox": [ + 132, + 567, + 481, + 599 + ], + "type": "text", + "content": "30. Sun, D., Roth, S., Black, M.J.: A quantitative analysis of current practices in optical flow estimation and the principles behind them. International Journal of Computer Vision 106, 115-137 (2014)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 600, + 481, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 600, + 481, + 643 + ], + "spans": [ + { + "bbox": [ + 132, + 600, + 481, + 643 + ], + "type": "text", + "content": "31. Tretschk, E., Tewari, A., Golyanik, V., Zollhöfer, M., Lassner, C., Theobalt, C.: Non-rigid neural radiance fields: Reconstruction and novel view synthesis of a dynamic scene from monocular video. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 12959-12970 (2021)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 643, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 643, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 643, + 481, + 665 + ], + "type": "text", + "content": "32. Verbin, D., Hedman, P., Mildenhall, B., Zickler, T., Barron, J.T., Srinivasan, P.P.: Ref-nerf: Structured view-dependent appearance for neural radiance fields. In: 2022" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "text", + "content": "W. Im et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 281 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 149, + 116, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 116, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 149, + 116, + 480, + 138 + ], + "type": "text", + "content": "IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 5481-5490. IEEE (2022)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 139, + 480, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 139, + 480, + 160 + ], + "spans": [ + { + "bbox": [ + 133, + 139, + 480, + 160 + ], + "type": "text", + "content": "33. Wang, C., Eckart, B., Lucey, S., Gallo, O.: Neural trajectory fields for dynamic novel view synthesis. arXiv preprint arXiv:2105.05994 (2021)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 161, + 480, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 161, + 480, + 182 + ], + "spans": [ + { + "bbox": [ + 133, + 161, + 480, + 182 + ], + "type": "text", + "content": "34. Wang, C., MacDonald, L.E., Jeni, L.A., Lucey, S.: Flow supervision for deformable nerf. arXiv preprint arXiv:2303.16333 (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 183, + 480, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 183, + 480, + 215 + ], + "spans": [ + { + "bbox": [ + 132, + 183, + 480, + 215 + ], + "type": "text", + "content": "35. Wu, T., Zhong, F., Tagliasacchi, A., Cole, F., Oztireli, C.: D\\~ 2nerf: Self-supervised decoupling of dynamic and static objects from a monocular video. Advances in Neural Information Processing Systems 35, 32653-32666 (2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 216, + 480, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 216, + 480, + 258 + ], + "spans": [ + { + "bbox": [ + 132, + 216, + 480, + 258 + ], + "type": "text", + "content": "36. Yoon, J.S., Kim, K., Gallo, O., Park, H.S., Kautz, J.: Novel view synthesis of dynamic scenes with globally coherent depths from a monocular camera. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5336-5345 (2020)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 259, + 480, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 259, + 480, + 281 + ], + "spans": [ + { + "bbox": [ + 132, + 259, + 480, + 281 + ], + "type": "text", + "content": "37. Zhang, K., Riegler, G., Snavely, N., Koltun, V.: Nerf++: Analyzing and improving neural radiance fields. arXiv preprint arXiv:2010.07492 (2020)" + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 197, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 197, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 197, + 91, + 447, + 102 + ], + "type": "text", + "content": "Regularizing Dynamic Radiance Fields with Kinematic Fields" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Regulating Model Reliance on Non-Robust Features by Smoothing Input Marginal Density/1125a5c6-c219-425e-8e55-ef1e3d508033_content_list.json b/2024/Regulating Model Reliance on Non-Robust Features by Smoothing Input Marginal Density/1125a5c6-c219-425e-8e55-ef1e3d508033_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..1be3b88393a8348427604cc5011aa884e21e4909 --- /dev/null +++ b/2024/Regulating Model Reliance on Non-Robust Features by Smoothing Input Marginal Density/1125a5c6-c219-425e-8e55-ef1e3d508033_content_list.json @@ -0,0 +1,1858 @@ +[ + { + "type": "text", + "text": "Regulating Model Reliance on Non-Robust Features by Smoothing Input Marginal Density", + "text_level": 1, + "bbox": [ + 230, + 140, + 772, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Peiyu Yang1, Naveed Akhtar2, Mubarak Shah3, and Ajmal Mian1", + "bbox": [ + 240, + 210, + 759, + 227 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 The University of Western Australia, Perth, Australia", + "bbox": [ + 313, + 238, + 687, + 252 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "2 The University of Melbourne, Melbourne, Australia", + "bbox": [ + 320, + 253, + 679, + 266 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3 University of Central Florida, Orlando, America", + "bbox": [ + 333, + 266, + 668, + 280 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "peiyu.yang@research.uwa.edu.au, naveed.akhtar1@unimelb.edu.au, shah@crcv.ucf.edu, ajmal.mian@uwa.edu.au", + "bbox": [ + 263, + 282, + 741, + 308 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Trustworthy machine learning necessitates meticulous regulation of model reliance on non-robust features. We propose a framework to delineate and regulate such features by attributing model predictions to the input. Within our approach, robust feature attributions exhibit a certain consistency, while non-robust feature attributions are susceptible to fluctuations. This behavior allows identification of correlation between model reliance on non-robust features and smoothness of marginal density of the input samples. Hence, we uniquely regularize the gradients of the marginal density w.r.t. the input features for robustness. We also devise an efficient implementation of our regularization4 to address the potential numerical instability of the underlying optimization process. Moreover, we analytically reveal that, as opposed to our marginal density smoothing, the prevalent input gradient regularization smoothens conditional or joint density of the input, which can cause limited robustness. Our experiments validate the effectiveness of the proposed method, providing clear evidence of its capability to address the feature leakage problem and mitigate spurious correlations. Extensive results further establish that our technique enables the model to exhibit robustness against perturbations in pixel values, input gradients, and density.", + "bbox": [ + 259, + 345, + 743, + 609 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Robust features $\\cdot$ Regularization $\\cdot$ Feature attributions", + "bbox": [ + 261, + 623, + 715, + 637 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 664, + 375, + 679 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Research on mitigating model reliance on non-robust input features has recently gained increasing attention due to high-stake machine learning applications [12, 18,38,49]. In this paper, we advance this direction by introducing a regularization technique that promotes a smooth marginal probability density function of the input to regulate the model's reliance on non-robust features.", + "bbox": [ + 212, + 695, + 785, + 768 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To distinguish between robust and non-robust features, we leverage the notion of attributions [16, 52, 58]. For a model $f$ parameterized by $\\theta$ , attributions characterize the importance of the $i$ -th feature $x_{i}$ of the input $x$ for the model", + "bbox": [ + 212, + 771, + 787, + 816 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "4 Our code is available at https://github.com/typeiyu/Input_density_reg.", + "bbox": [ + 217, + 824, + 732, + 840 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/9cca57a90c893a17217a53e9d82aeb00daa20dac0f98771c14b116809af6dcb9.jpg", + "image_caption": [ + "Fig. 1: Attribution maps [45] and insertion game scores [34] for samples from (a) BlockMNIST and (b) CelebA-Hair datasets. As compared to input gradient regularization, our regularization leads to lower feature leakage while also achieving higher AUC for the insertion game." + ], + "image_footnote": [], + "bbox": [ + 217, + 143, + 787, + 279 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "prediction by quantifying the output change between $f(x; \\theta)$ and $f(x_{[x_i = 0]}; \\theta)$ . Since robust input features contribute to model predictions equally well across slight condition variations, their attributions exhibit a certain consistency. On the other hand, non-robust feature attributions fluctuate under such variations. This potentially identifies a correlation between the model's reliance on non-robust features and the smoothness of the marginal probability density function of the input samples $p_\\theta(x)$ . This is an important insight for our contribution. For robustness, it offers a possibility of model regularization using the gradients of the marginal density with respect to the input $\\nabla_x p_\\theta(x)$ . We propose to regularize $\\nabla_x p_\\theta(x)$ to encourage the model to prioritize the use of robust features and regulate its reliance on non-robust features. However, this can also lead to numerical instability during model optimization. To address that, we further introduce a stable and efficient implementation to estimate the gradient of marginal density.", + "bbox": [ + 212, + 354, + 787, + 551 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We also investigate input gradient norm regularization [13, 36, 37] and reveal that input gradients can be interpreted as input gradients of the log-conditional density $\\nabla_{x}\\log p_{\\theta}(x|y)$ or log-joint density $\\nabla_{x}\\log p_{\\theta}(x,y)$ . Input gradient regularization mitigates the model's reliance on non-robust features specific to the class label $y = i$ , leading to unintentional blindness to class-specific non-robust features where $y \\neq i$ . In contrast, our regularization encourages smoothness of the marginal density $p_{\\theta}(x)$ without imposing unintended constraints, providing a comprehensive regularization of the non-robust features. In Fig. 1, we employ attribution maps [45] and insertion game scores [34] to compare the robustness of vanilla models, input gradient regularized models and models trained with our regularization on BlockMNIST [44] and CelebA-Hair [30] datasets. As the representative examples show, our method suppresses both feature leakage and feature spurious correlation to improve model robustness and interpretability.", + "bbox": [ + 212, + 551, + 787, + 748 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The effectiveness of our approach is extensively established through a series of experiments. First, using BlockMNIST dataset [44], we demonstrate that the model trained with our regularization considerably mitigates the problem of feature leakage. This problem occurs when a model wrongly attributes importance to irrelevant but persistent features in the data to achieve better accuracy, e.g., the null block (☑) in Fig. 1. Addressing feature leakage helps suppress", + "bbox": [ + 212, + 750, + 787, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "P. Yang et al.", + "bbox": [ + 271, + 114, + 364, + 128 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "spurious correlation between the input features and model predictions [2, 39]. Moreover, we establish the robustness of the models trained with our regularization against perturbations from adversarial attacks [17, 31], input pixels [41] and input density, demonstrating the broad applicability of our approach. Our main contributions are summarized as follows.", + "bbox": [ + 212, + 146, + 787, + 220 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. We identify robust and non-robust features by leveraging attributions, and establish the correlation between model reliance on non-robust features and the smoothness of data marginal density.", + "2. We propose an efficient technique for calculating the gradients of log density, also addressing the numerical instability of the underlying optimization.", + "3. Through extensive experiments, we demonstrate the effectiveness of the proposed regularization and additionally establish that our approach exhibits general robustness against perturbations." + ], + "bbox": [ + 220, + 236, + 782, + 354 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 383, + 387, + 398 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Regularization for Interpretability Robustness: Despite advancements in model transparency [26, 27, 40, 57], current deep neural networks still lack interpretability in their decision-making process, which is exacerbated by their reliance on non-robust input features. Prior studies, such as [1, 2, 44], identify that standard models are prone to relying on irrelevant or spuriously correlated features. To address that, several regularization techniques are proposed to improve model interpretability. In [14, 42], the authors incorporated prior knowledge into the model training process to regularize the model behavior. Dombrowski et al. [11, 12] found that regularizing the input Hessian using SoftPlus activations or weight decay can boost resilience against manipulated inputs. In [18], a joint energy-based model is trained as a discriminative model for improved robustness. Srinivas and Fleuret [49] enhanced the model interpretability by improving the alignment between the implicit density and the data distribution.", + "bbox": [ + 212, + 416, + 787, + 611 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Regularization for Adversarial Robustness: In addition to the other sources of prediction unreliability, adversarial attacks can manipulate model outputs with imperceptible perturbations to inputs [17, 31]. To address this, adversarial training through data augmentation with adversarial samples is widely employed [9, 31, 43]. Certified adversarial robustness through regularizations [4, 36, 46] is another branch of methods to defend against adversarial perturbations. Inspired by the classic double backpropagation [13], Ross and Doshi-Velez [36] regularized the input gradient norm for adversarial robustness. Etmann [15] further explored different variants of double backpropagation regularizations for various real-world scenarios. Moosavi-Dezfooli et al. [32] also proposed regularization to encourage a low curvature for adversarial robustness. To improve adversarial robustness, Chen et al. [10] computed the norm of attributions integrated from clean samples to adversarial samples as the regularization term. Ilyas et al. [24] proposed a disentangling method to distinguish feature robustness for explaining adversarial examples. However, they do not focus on general feature robustness.", + "bbox": [ + 212, + 613, + 787, + 839 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "Regulating Model Reliance on Non-Robust Features", + "bbox": [ + 383, + 114, + 730, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "From their perspective, non-robust features are highly predictive; yet adversarially brittle. In contrast, our work focuses on general robustness derived directly from natural images.", + "bbox": [ + 212, + 146, + 782, + 191 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Attribution Methods: Feature attribution methods are used to estimate the importance of input features for a model's prediction and can be categorized as either perturbation-based or back-propagation-based techniques. Perturbation-based methods [16,58,61] calculate the attribution scores by repeatedly perturbing the input features and analyzing the resulting effects on the model prediction. These methods are also extended for evaluating the reliability of the computed attributions [3,34,41,55]. The back-propagation-based techniques [47,52,54,56] estimate the attribution scores by computing the gradients or integrated gradients with respect to the input features in the backward propagation process. In contrast to perturbation-based techniques, back-propagation-based attribution methods offer notable advantages in terms of computational efficiency and reliability. Given the transparency of attributions, our work leverages the attribution framework to distinguish between robust and non-robust features, which enables us to systematically analyze the robustness of input features.", + "bbox": [ + 212, + 191, + 787, + 402 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3 Feature Robustness by Attributions", + "text_level": 1, + "bbox": [ + 214, + 426, + 599, + 443 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We first provide a framework for distinguishing between robust and non-robust features by analyzing their attributions. Herein, attribution inconsistencies among the features with distinct degrees of robustness identify a correlation between the model's reliance on non-robust features and the smoothness of output logits.", + "bbox": [ + 212, + 455, + 790, + 516 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Let us consider an input sample $x \\in \\mathbb{R}^n$ with label $y \\in \\mathbb{R}^c$ from a dataset $\\mathcal{D}$ , and a classifier $f: \\mathbb{R}^n \\to \\mathbb{R}^c$ parameterized by $\\theta$ . We denote robust and non-robust features within the input $x$ as $x_{rob}, x_{nrob} \\subseteq x$ . Consider an attribution method $\\phi: \\mathbb{R}^c \\to \\mathbb{R}^n$ attributing model predictions to input features by estimating their importance, resulting in an attribution map $M = \\phi(f(x; \\theta))$ . Inspired by the success of attributions in model explanation, we identify robust and non-robust features by leveraging their attributions.", + "bbox": [ + 212, + 517, + 787, + 622 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Without loss of generality, we assume that attributions $M$ of the features can be estimated by calculating the change in output logits when these features are removed from the input, following perturbation-based methods [35,58]: $M_{x_{rob}} = f(x;\\theta) - f(x_{[x_{rob} = 0]};\\theta)$ and $M_{x_{nrob}} = f(x;\\theta) - f(x_{[x_{nrob} = 0]};\\theta)$ . For ease of understanding, we alternatively use $f(x_{nrob};\\theta)$ and $f(x_{rob};\\theta)$ to represent attributions $M_{rob}$ and $M_{nrob}$ in the text to follow. We define robust features within the attribution framework as follows.", + "bbox": [ + 212, + 623, + 787, + 728 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Definition 1. A feature $x_{feat}$ shared among different input instances under its domain $\\Delta_{x_{feat}}$ is robust if, for a randomly chosen class $y = i$ , its attribution $M_{x_{feat}}$ is bounded by a small constant $h$ under a metric $c(\\cdot)$ , i.e., $c(f(x_{feat};\\theta) - f(\\tilde{x}_{feat};\\theta)) \\leq h : \\tilde{x}_{feat} \\in \\Delta_{x_{feat}}$ .", + "bbox": [ + 214, + 739, + 787, + 801 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Under Definition 1, robust features are expected to contribute consistently to the model's prediction across different input samples. Non-robust features, on", + "bbox": [ + 212, + 809, + 785, + 839 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "P. Yang et al.", + "bbox": [ + 271, + 114, + 362, + 128 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "the other hand, are those that contribute to the prediction score inconsistently or only under specific conditions. Our focus here is on distinguishing between robust and non-robust features, without requiring to specify a particular metric. Definition 1 emphasizes attribution consistency for robust features rather than attribution positivity, thereby allowing for robust features that can also make a negative contribution to the model's prediction. Building further upon the above definition, we make the following remark.", + "bbox": [ + 212, + 146, + 782, + 252 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Remark 1. Robust features are largely condition-invariant in that they retain similar attributions despite slight changes to the input. In contrast, non-robust features are condition-specific in that their attributions either vary drastically with slightly varying input conditions, or behave robustly only under specific conditions.", + "bbox": [ + 214, + 265, + 787, + 340 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Robust features exhibit stable behavior across the input space, which is observable through consistent output logits $f(x_{rob};\\theta)$ in classifiers regardless of the input instance or class $y$ . In contrast, non-robust features rely on specific conditions to exhibit a particular behavior, which is tailored to a specific class $y = i$ or the input instance. Robust modeling aims for a stronger reliance on robust features for prediction. Due to the consistency of output logits $f(x;\\theta)$ for robust features, a smooth $f(x;\\theta)$ is a desirable property for model robustness.", + "bbox": [ + 212, + 353, + 784, + 459 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4 Smoothing Marginal Density of Input", + "text_level": 1, + "bbox": [ + 214, + 482, + 617, + 501 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Here, we establish the relation between model robustness and gradients of the input marginal density. Then, a robust regularization is proposed for regulating model reliance on non-robust features by smoothing marginal density.", + "bbox": [ + 212, + 513, + 782, + 559 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We commence our analytical analysis with probability density, following Bridle [7]. Given a class $y = i$ , a joint probability density function over the input with the output logit $f_{i}(x;\\theta)$ is defined as", + "bbox": [ + 212, + 560, + 782, + 604 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\np _ {\\theta} (x, y = i) = e ^ {f _ {i} (x; \\theta)} / Z _ {\\theta}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 410, + 616, + 784, + 633 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where the constant $Z_{\\theta} = \\int e^{f_i(x;\\theta)}dx$ is the partition function. $Z_{\\theta}$ normalizes the input $x$ to a probability density by integrating over all possible input points $x$ in the input space via the model $f$ . By applying Bayes' rule, we eliminate the condition $y = i$ , resulting in the marginal density being defined solely on the input $x$ : $p_{\\theta}(x) = p_{\\theta}(y = i,x) / p_{\\theta}(y = i|x)$ . The conditional density function $p_{\\theta}(y = i|x)$ can be further defined as", + "bbox": [ + 212, + 643, + 784, + 736 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\np _ {\\theta} (y = i | x) = e ^ {f _ {i} (x; \\theta)} / Z _ {f (x; \\theta)}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 398, + 747, + 784, + 765 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $Z_{f(x;\\theta)} = \\sum_{i=1}^{C} e^{f_i(x;\\theta)}$ is the partition function for the output logits $f_i(x;\\theta)$ defined on all the $C$ classes. To simplify the notation, we use $Z_{f(x)}$ to represent $Z_{f(x;\\theta)}$ in the subsequent discussion. Exploiting the symmetry property of the joint density defined in Eq. 1, i.e., $p_\\theta(x,y=i) = p_\\theta(y=i,x)$ , the", + "bbox": [ + 212, + 777, + 787, + 842 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "Regulating Model Reliance on Non-Robust Features", + "bbox": [ + 383, + 114, + 732, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "marginal density $p_{\\theta}(x)$ can be expressed as", + "bbox": [ + 215, + 146, + 527, + 161 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\np _ {\\theta} (x) = \\frac {e ^ {f _ {i} (x ; \\theta)} / Z _ {\\theta}}{e ^ {f _ {i} (x ; \\theta)} / Z _ {f (x)}} = \\frac {Z _ {f (x)}}{Z _ {\\theta}}. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 388, + 174, + 785, + 210 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "As identified in the previous section, smoothness of output logits across inputs encourages the use of robust features by the model. Hence, we consider the marginal density $p_{\\theta}(x)$ defined on the output logits $f(x; \\theta)$ across the input space. Promoting a smaller gradient of the marginal density with respect to the input $x$ , denoted as $\\nabla_x p_{\\theta}(x)$ , contributes to the smoothness of the output logits. Thus, a positive correlation can be established between the use of robust features and the smoothness of the probability marginal density $p_{\\theta}(x)$ . In particular, the smooth output logits of robust features across input samples suggest that these features will have relatively small gradients of the density $p_{\\theta}(x)$ with respect to the input values. On the other hand, non-robust features with fluctuating output logits will have large gradients of the density that need to be suppressed during the training process for model robustness. Therefore, we can conclude with the following remark.", + "bbox": [ + 212, + 220, + 787, + 417 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Remark 2. Model reliance on non-robust features $x_{nrob}$ can be regulated by regularizing the gradients of the marginal density $p_{\\theta}(x)$ w.r.t. $x$ , and this regularization can be achieved through optimizing the model parameters.", + "bbox": [ + 215, + 431, + 789, + 479 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In the light of Remark 2., we propose a regularization term for minimizing the gradients of marginal density. However, computing the gradients of the marginal density $\\nabla_{x}p_{\\theta}(x)$ is not feasible because the partition function defined on the entire input space is intractable. To avoid the estimation of the intractable $Z_{\\theta}$ , we instead compute the gradients with respect to the log density as $\\nabla_{x}\\log p_{\\theta}(x) = \\nabla_{x}\\log Z_{f(x)}$ . This is possible because $Z_{\\theta}$ solely depends on the model parameter $\\theta$ and not the input $x$ . Expanding the partition function $Z_{f(x)}$ in $\\nabla_{x}\\log p_{\\theta}(x)$ , we obtain $\\nabla_{x}\\log p_{\\theta}(x) = \\sum_{i=1}^{C}\\nabla_{x}e^{f_{i}(x;\\theta)} / \\sum_{i=1}^{C}e^{f_{i}(x;\\theta)}$ . The $p$ -norm of this gradient is computed as the regularization term. In the optimization process, the goal is to find the optimal parameter $\\theta^{*}$ , cf. Remark 2., by minimizing the loss $\\ell$ as", + "bbox": [ + 212, + 492, + 787, + 657 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\theta^ {*} = \\min _ {\\theta} \\ell (f (x; \\theta), y) + \\lambda | | \\frac {\\sum_ {i = 1} ^ {C} \\nabla_ {x} e ^ {f _ {i} (x ; \\theta)}}{\\sum_ {i = 1} ^ {C} e ^ {f _ {i} (x ; \\theta)}} | | _ {p}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 338, + 659, + 785, + 698 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\lambda$ indicates the magnitude of the coefficient for controlling the strength of the regularization.", + "bbox": [ + 212, + 703, + 787, + 733 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To regulate model reliance on non-robust features, our regularization encourages the smoothness of marginal density by regularizing its gradients. Since the output logit change in the log-marginal density $\\log p_{\\theta}(x) = \\log Z_{f(x)} - \\log Z_{\\theta}$ , and $Z_{\\theta}$ is independent of the input $x$ , we can only focus on the first term $Z_{f(x)} = \\sum_{i=1}^{C} e^{f_i(x;\\theta)}$ . Recall, Definition 1 of robust features. Assuming the robust input feature $x_{rob}$ exists in a random input $x$ , the corresponding output logit $f_i(x_{rob})$ will consistently attribute to the model prediction. This property", + "bbox": [ + 212, + 734, + 787, + 842 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "P. Yang et al.", + "bbox": [ + 271, + 114, + 362, + 128 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "of $x_{rob}$ leads to the smoothness of output change $\\sum_{i=1}^{C} e^{f_i(x_{rob};\\theta)}$ across different input samples and class labels. In contrast, non-robust input features $x_{nrob}$ show relatively high attributions for the output logit $f_i(x_{nrob})$ for a given class $i$ . However, they cannot maintain consistency in the attributions across different inputs or labels, leading to fluctuations in the output change $\\sum_{i=1}^{C} e^{f_i(x_{nrob};\\theta)}$ . In Fig. 1(a) and Fig. 13 in Supp. 17, non-robust features in the Null block exhibit fluctuating attributions across different samples on the standard trained model. It is demonstrated that the magnitude of gradients for input features in the marginal density $p_\\theta(x)$ reflects the model's sensitivity to those features. We leverage this relation to mitigate model reliance on the non-robust features by smoothing the marginal density of the input samples.", + "bbox": [ + 212, + 143, + 787, + 313 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5 Stable and Efficient Implementation for Regularization", + "text_level": 1, + "bbox": [ + 215, + 337, + 781, + 356 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "From the implementation perspective, the gradient computation of marginal density involves multiple exponential operations in both the numerator and denominator of Eq. 4 which can introduce numerical instability in the optimization process, leading to gradient vanishing and explosion problems. Such issues can potentially hinder the application of our regularization to large non-linear models or wide-distribution data. For instance, batch normalization (BN) layers [25] solving internal covariate shifts with learnable scaling and shifting parameters can amplify the errors caused by the exponential operations during the backpropagation. In Fig. 2(a), yellow and red curves indicate original and inverse implementations for minimizing the gradient of log density $\\nabla_{x}\\log Z_{f(x)}$ . It is illustrated that the Frobenius gradient norm of both implementations undergoes rapid numerical overflow as the number of training iterations increases, revealing the serious issue of exploding and vanishing gradients using the proposed regularization. Therefore, it is crucial to address these through careful implementation to ensure the feasibility of our regularization.", + "bbox": [ + 212, + 371, + 787, + 598 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To address this challenge, we transform the computation from the summation of exponential operations to softmax. By subtracting a constant value $\\eta$ from the logits before exponentiation, softmax can prevent numerical overflow during the exponential operations, i.e., $e^{f_i(x;\\theta)} / \\sum_j e^{f_j(x;\\theta)} = e^{f_i(x;\\theta) - \\eta} / \\sum_j e^{f_j(x;\\theta) - \\eta}$ . Thus, we incorporate the softmax function into the computation of log density", + "bbox": [ + 212, + 598, + 787, + 674 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/4883dab3090e5047ad2a3d730cd81760d88a43f83de7a679acb12b363c6048f8.jpg", + "image_caption": [ + "(a) Numerical Stability." + ], + "image_footnote": [], + "bbox": [ + 256, + 681, + 406, + 777 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/df2bee27ab0b88dec8a8b67ecd54d5d671cc737f30fec82477cb9670f6b7db52.jpg", + "image_caption": [ + "(b) Training Efficiency." + ], + "image_footnote": [], + "bbox": [ + 429, + 681, + 581, + 776 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/421827394b8016dee901706db3a42f0c89dd6b46b3628ba42c3a297682adb77e.jpg", + "image_caption": [ + "(c) Training Time.", + "Fig. 2: The comparison of numerical stability, training efficiency and training time across different ResNet-34 models." + ], + "image_footnote": [], + "bbox": [ + 594, + 683, + 743, + 781 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "Regulating Model Reliance on Non-Robust Features", + "bbox": [ + 383, + 114, + 732, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "gradient $\\nabla_{x}logz_{f(x)}$ as", + "bbox": [ + 215, + 146, + 382, + 162 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\nabla_ {x} l o g z _ {f (x)} = \\nabla_ {x} f _ {i} (x; \\theta) - \\nabla_ {x} (l o g e ^ {f _ {i} (x; \\theta)} - l o g z _ {f (x)}), \\\\ = \\nabla_ {x} f _ {i} (x; \\theta) - \\nabla_ {x} \\left(\\log \\left(f _ {i} (x; \\theta) / z _ {f (x)}\\right)\\right), \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 318, + 175, + 683, + 212 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "(5)", + "bbox": [ + 761, + 186, + 782, + 200 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $f_{i}(x)$ indicates a logit of a random $i$ -th class, and $z_{f(x)}$ equals $\\sum_{j} e^{f_{j}(x; \\theta)}$ . Thus, the gradients with respect to the log-marginal density $p_{\\theta}(x)$ can be replaced by computing the difference between the gradient $\\nabla_{x} f_{i}(x; \\theta)$ and the gradient of a log-softmax output $\\nabla_{x} \\log(e^{f_{i}(x; \\theta)} / Z_{f(x)})$ .", + "bbox": [ + 212, + 224, + 784, + 286 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Our technique improves upon the common approach [5, 6] to achieve numerical stability in log exponential sum calculations, which typically employs the formula $\\log(\\sum_{i \\in \\{1, \\ldots, n\\}} e^{x_i}) = \\eta + \\log(\\sum_{i \\in \\{1, \\ldots, n\\}} e^{x_i - \\eta})$ , with $\\eta$ being the maximum value of inputs $\\{x_1, \\ldots, x_n\\}$ . We address the numerical instability by employing softmax, avoiding computationally expensive comparisons of the maximum gradient values. Specifically, the basic stable implementation involves finding the maximum gradient within the gradients of various classes, leading to a time complexity of $O(n)$ , where $n$ represents the number of classes. In contrast, our method eliminates numerical instabilities associated with a randomly selected class, eliminating the need for maximum value comparisons and leading to a more favorable time complexity of $O(1)$ .", + "bbox": [ + 212, + 286, + 785, + 453 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "While our solution helps avoid numerical instability, it requires twice the gradient computations as compared to the sole calculation of density gradients. Hence, we propose an efficient mechanism for estimating the difference in the gradients. Specifically, we compute the gradient of the difference between two outputs to approximate the difference between the two gradients of outputs using Taylor series as", + "bbox": [ + 212, + 455, + 787, + 544 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\nabla_ {x} f _ {i} (x; \\theta) - \\nabla_ {x} \\log \\left(e ^ {f _ {i} (x; \\theta)} / Z _ {f (x)}\\right) \\approx \\nabla_ {x} \\left(f _ {i} (x; \\theta) - \\log \\left(e ^ {f _ {i} (x; \\theta)} / Z _ {f (x)}\\right)\\right). \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 243, + 556, + 784, + 574 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The proof of Eq. 6 is provided in Supp. 9. As such, the proposed approach enables stable and efficient model optimization. The blue curve in Fig. 2(a) shows that our efficient implementation can effectively avoid numerical instability in the gradient computation. Fig. 2(b) provides a comparison of training efficiency on CIFAR-10 [28] between the basic numerically stable implementation and our two derived alternatives. More results regarding the comparison of numerical stability are provided in Supp. 15. In Fig. 2(c), we delve into a specific examination of the training times of models trained across the three datasets including BlockMNIST, CelebA and CIFAR-10. It is evident that our methods substantially improve the per-epoch training time. Moreover, our methods can yield further beneficial trade-offs as the number of classes increases.", + "bbox": [ + 212, + 585, + 787, + 752 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6 Limited Robustness in Input Gradient Regularization", + "text_level": 1, + "bbox": [ + 212, + 776, + 769, + 794 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Input gradient regularization (InputGrad Reg.) [29,36] computes the Frobenius norm of input gradients $||\\nabla_x f_i(x;\\theta)||_F$ for a given class label $y = i$ , which is a", + "bbox": [ + 212, + 809, + 785, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "P. Yang et al.", + "bbox": [ + 271, + 114, + 364, + 128 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "baseline robust regularization for model optimization. Existing works [29,36,37] explain InputGrad Reg. as a prediction stability technique for robustness against perturbations. Here, we reveal that the input gradient norm potentially regularizes the gradients of implicit data density. Moreover, we provide an understanding of how InputGrad Reg. encourages robustness as well as its limitations.", + "bbox": [ + 212, + 146, + 782, + 222 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Suppose all the classes have equal probability, i.e., $p_{\\theta}(y = i) = 1 / C$ . We can express the class-conditional density $p_{\\theta}(x|y)$ by using Bayes's rule as", + "bbox": [ + 212, + 222, + 782, + 252 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\np _ {\\theta} (x \\mid y = i) = \\frac {p _ {\\theta} (x , y = i)}{p _ {\\theta} (y = i)} = \\frac {e ^ {f _ {i} (x ; \\theta)}}{Z _ {\\theta} / C}. \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 369, + 265, + 785, + 297 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Now, we compute the gradients of the log density defined in Eq. 7 with respect to the input $x$ as", + "bbox": [ + 214, + 308, + 782, + 338 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\nabla_ {x} \\log p _ {\\theta} (x, y = i) = \\nabla_ {x} \\log p _ {\\theta} (x | y = i) = \\nabla_ {x} f _ {i} (x; \\theta). \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 313, + 353, + 784, + 368 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Eq. 8 demonstrates that the input gradients can be interpreted as the gradients of either the log joint density or the log conditional density with respect to the input $x$ . This formulation highlights that InputGrad Reg. encourages consistent attributions of input features for the model's predictions. However, it is important to note that InputGrad Reg. is formulated under the specific condition $y = i$ , which limits its effectiveness in resolving inconsistencies when predicting a different class $y = j$ , where $j \\neq i$ . Although input features consistently contribute to the model's prediction under a given class, InputGrad Reg. fails to consider inconsistent attributions of these features across different classes, thereby allowing model non-robust behavior to exist. Consequently, a model trained with input gradient regularization may still exhibit spurious robustness relying on specific conditions. Fig. 1(a) illustrates an example where the model trained with InputGrad Reg. is incapable of suppressing non-robust features in Null block.", + "bbox": [ + 212, + 380, + 787, + 590 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Remark 3. Input gradient regularization smoothens the joint and conditional density of the input $x$ under a specific label $y = i$ , compromising its ability to resist the class-specific non-robust features.", + "bbox": [ + 214, + 604, + 787, + 650 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "In contrast to the existing works [36, 49] that highlight the reasons for Input-Grad Reg. efficacy, we reveal a weakness of this technique, cf. Remark 3.. Unlike regularizing gradients based on joint or conditional densities, our approach allows for more effective regularization without imposing the condition $y = i$ . Our method focuses on regularizing the gradients of marginal density $\\nabla_{x}p_{\\theta}(x)$ , thereby smoothing the output logits across the input samples.", + "bbox": [ + 212, + 662, + 787, + 753 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "7 Experiments", + "text_level": 1, + "bbox": [ + 215, + 777, + 375, + 795 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "In this section, we perform extensive experiments to validate the efficacy of our regularization and the newly established correlation between the smooth", + "bbox": [ + 212, + 809, + 785, + 839 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "Regulating Model Reliance on Non-Robust Features", + "bbox": [ + 383, + 114, + 732, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "marginal density and model reliance on non-robust features. Specifically, we present measurement results for addressing feature leakage and mitigating spurious correlations. In addition to specific applications, further results on the robustness against perturbations in input pixels, gradients and density are presented. Additional details about the datasets and the models used in our experiments can be found in Supp. 19.", + "bbox": [ + 212, + 146, + 787, + 238 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "7.1 Efficacy against Feature Leakage and Adversarial Attacks", + "text_level": 1, + "bbox": [ + 214, + 266, + 733, + 282 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "In [2, 44], it is demonstrated that deep models end up assigning importance to irrelevant input features. Shah et al. [44] used BlockMNIST in their experiments, which is a synthetic dataset extended from MNIST [29]. To each MNIST sample, BlockMNIST attaches a null block (an irrelevant pattern) randomly at the top or bottom of the image, as shown in Fig. 3(a). Shah et al. [44] observed that the explaining tool InputGrad [47] attributes importance to both the informative number block and the uninformative null block in the standard trained model. This phenomenon is termed as feature leakage by the authors. Interestingly, the issue is mitigated in adversarially trained models. Since this dataset allows for a controlled robustness assessment, we first evaluate our method on BlockMNIST to analyze the model's reliance on irrelevant features.", + "bbox": [ + 212, + 296, + 787, + 462 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Reproducibility and Quantitative Measurement of Feature Leakage.", + "bbox": [ + 214, + 464, + 782, + 479 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Owing to the unreliability of InputGrad caused by model saturation [45], we employ Integrated Gradients (IG) [52], an axiomatic explanation tool, to faithfully reinvestigate the feature leakage phenomenon. In Fig. 3(b)-(c), we show that the attributions computed by IG can reproduce different leakage phenomena from the informative block region to the null block region on standard and adversarially trained models. Feature leakage is an important phenomenon in the context of model robustness. However, there is a lack of a quantitative metric in the current literature to quantify its extent. We use integrated gradients to define the metric $M_{leakage}$ to address this gap. Mathematically,", + "bbox": [ + 212, + 479, + 472, + 797 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/5db74f99dcb8a59b9ca5a2787382f37ed2a0821c46b25f20de72903915974618.jpg", + "image_caption": [ + "(a) Input samples of BlockMNIST" + ], + "image_footnote": [], + "bbox": [ + 486, + 488, + 779, + 573 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/07d33db9873d2df2789915643336c04869b09dda61165952f19dc25b5460305e.jpg", + "image_caption": [ + "(b) Integrated Gradients on (c) Integrated Gradients on standard trained model adversarially trained model", + "Fig.3: BlockMNIST samples and feature leakage problem. (a) BlockMNIST randomly appends a null block at the top or bottom of MNIST samples. (b&c) Attribution maps are calculated by IG on the standard and adversarially trained models." + ], + "image_footnote": [], + "bbox": [ + 482, + 590, + 777, + 646 + ], + "page_idx": 9 + }, + { + "type": "equation", + "text": "\n$$\nM _ {\\text {l e a k a g e}} = \\mathbb {E} _ {x _ {n r o b} \\sim D} | | x _ {n r o b} \\times \\int_ {\\alpha = 0} ^ {1} \\frac {\\partial f (\\alpha \\cdot x _ {n r o b} ; \\theta))}{\\partial x _ {n r o b}} d \\alpha | | _ {2}, \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 295, + 808, + 785, + 844 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "P. Yang et al.", + "bbox": [ + 271, + 114, + 362, + 128 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "where $\\alpha$ is the step from the absence to the presence of the features, and $x_{nrob}$ specifies the non-robust features in the null block. Since the attributions of $x_{nrob}$ for the model's prediction should ideally be zero, we use the $L_{2}$ norm of attributions to quantify the extent of feature leakage.", + "bbox": [ + 212, + 146, + 787, + 207 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Robustness against Feature Leakage. Table 1 presents the experimental results on the BlockMNIST dataset. We compare our method with other robust regularizations and techniques including InputGrad [36], IG-SUM [10], SoftPlus activations [11] and Hessian [12]. InputGrad and Hessian regularize the first-order and second-order gradients w.r.t. the input. Models trained with SoftPlus activations and Hessian regularization fail to suppress the leakage problem, which indicates that feature leakage is not caused by the geometry of the model output manifold or high curvature [11, 59]. InputGrad regularization demonstrates robustness against both $L_{2}$ and $L_{\\infty}$ adversarial attacks, yet it still fails to address the leakage problem. The result aligns well with Remark 3, which highlights the allowance of non-robust features across different classes in InputGrad regularization. In addition, the use of IG in IG-Norm regularization that accumulates input gradients as a regularization term results in behavior similar to InputGrad regularization, leading to compromised robustness. These results further reveal that adversarial robustness is not a sufficient condition for suppressing feature leakage. Our method demonstrates a considerable improvement over other techniques for feature leakage, while also maintaining superiority in adversarial robustness. In Tab. 1, we use $L_{2}$ norm for the compared regularization terms for fair benchmarking. We show in Supp. 10 that optimal norm exploration can yield an even more favorable trade-off for our technique for model robustness.", + "bbox": [ + 212, + 210, + 787, + 513 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Feature Leakage in Adversarily Robust Models. A FGSM adversarily trained model [17] augments the training samples by adversarial examples $x + \\epsilon \\cdot sign(\\nabla_x f_i(x; \\theta))$ . Notably minimizing the loss of the perturbed input $x + \\epsilon \\cdot sign(\\nabla_x f_i(x; \\theta))$ is similar to the InputGrad regularization. Thus, training with FGSM is still limited in its ability to suppress the leakage problem. In contrast, PGD attack [31] weakens the effect of the condition $y = i$ by iteratively searching for the perturbations from a random starting point. This process leads to a substantial enhancement in suppressing feature leakage, see Tab. 1.", + "bbox": [ + 212, + 516, + 789, + 638 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/4ff636f4d5661e11fb8722e8ae8cbae85fb390a6e7cbc10c46d13fb463a9dbe2.jpg", + "table_caption": [ + "Table 1: Results on BlockMNIST. Feature leakage, standard accuracy and adversarial accuracy under $L_{2}$ and $L_{\\infty}$ PGD-20 attacks are reported. ST & AT: Standard and Adversarial Training." + ], + "table_footnote": [], + "table_body": "
MethodFeature Leakage ↓PGD-20 (L2) ↑PGD-20 (L∞) ↑Accu. ↑
AT (FGSM)3.32487.480.0099.02
AT (PGD)2.31392.7528.0598.97
ST3.65773.570.0099.12
+ SoftPlus Acts.3.53367.950.0298.52
+ Hessian Reg.4.25880.060.0098.48
+ InputGrad Reg.3.46183.4621.1494.56
+ IG-SUM Reg.3.49782.1122.7392.87
+ Our Reg.2.25985.4129.3693.05
", + "bbox": [ + 215, + 715, + 792, + 854 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "Regulating Model Reliance on Non-Robust Features", + "bbox": [ + 383, + 114, + 730, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 116, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/7e1934dfae16c366802ad5283e745e806f06518ced504975b62d9a4042b22a52.jpg", + "image_caption": [ + "(a) Feature Leakage." + ], + "image_footnote": [], + "bbox": [ + 215, + 145, + 352, + 224 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/165ac84da9f9c1381686ae165c5c1b522eb2606fba7946221a49bb74a8cefc11.jpg", + "image_caption": [ + "(b) $L_{\\infty}$ Attack.", + "Fig. 4: Performance comparison between our method and InputGrad regularization under varying regularization coefficient for (a) Feature leakage, (b)-(c) Adversarial Accuracy under $L_{\\infty}$ and $L_{2}$ PGD-20 attacks, and (d) Accuracy." + ], + "image_footnote": [], + "bbox": [ + 361, + 143, + 496, + 224 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/b8cd40a4f5e1ade19ebe033e88077d01c8263368860ddbdbbb9cc01086363987.jpg", + "image_caption": [ + "(c) $L_{2}$ Attack." + ], + "image_footnote": [], + "bbox": [ + 506, + 143, + 640, + 224 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/5a9d1441cb3f3bbf7e71b1dbb4d8e2057bd84eeffb98bd7115c45bdb03f51e22.jpg", + "image_caption": [ + "(d) Accuracy." + ], + "image_footnote": [], + "bbox": [ + 651, + 143, + 785, + 224 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Magnitude of Coefficient for Regularization. Figs. 4(a)-(d) present a comparison of results for feature leakage and adversarial accuracy under PGD attacks, as well as the standard accuracy across varying magnitudes for the regularization strength. The results affirm that our method effectively regulates feature leakage by imposing a penalty on non-robust features. Moreover, our regularization enables the model to defend against both $L_{2}$ and $L_{\\infty}$ attacks while maintaining high accuracy, showing an outstanding trade-off across four criteria. More adversarial robustness comparisons on CIFAR dataset [28] are reported in Supp. 11.", + "bbox": [ + 212, + 305, + 784, + 441 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "7.2 Efficacy for Spurious Correlation", + "text_level": 1, + "bbox": [ + 215, + 465, + 532, + 479 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Recent research has highlighted the susceptibility of neural models in learning spurious correlations that enhance performance on a given data but fail to generalize [2,8,39]. For instance, in CelebA-Hair dataset [30], which commonly consists of samples containing female celebrities with blond hair and male celebrities with dark hair, models heavily rely on the spuriously correlated gender feature to predict the target hair color [19, 39]. Consequently, accuracy tends to be lower for the samples containing", + "bbox": [ + 212, + 492, + 486, + 703 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "male celebrities with blond hair, see Fig. 5.", + "bbox": [ + 215, + 704, + 527, + 718 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "To mitigate spurious correlations, distributional robust optimization (DRO) techniques have been proposed to re-weight the training loss of input samples from different groups [23,39]. In our regularization, an additional penalty is imposed to penalize the model's reliance on these spuriously correlated features because of their inconsistent attributions. This encourages the use of robust features while suppressing the model's reliance on spuriously correlated features. Reliable quantification of the model's robustness on natural images for spurious correlation suppression is an unresolved issue in the literature. We employ", + "bbox": [ + 212, + 719, + 784, + 839 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/00aa977129fc7b52da8eb0c90bfa240334d443e45459bb1086ea76ef90c2ea3a.jpg", + "image_caption": [ + "Common training samples:", + "Label: Blond hair", + "Spur. Feat.: Female", + "Accuracy: $92.1\\%$" + ], + "image_footnote": [], + "bbox": [ + 500, + 513, + 588, + 580 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/449779662b8899761d692b6e86ee95d564b3e117dfd4d58035d352ff5eeba19d.jpg", + "image_caption": [ + "Label: Dark hair", + "Spur. Feat.: Male", + "Accuracy: $99.7\\%$" + ], + "image_footnote": [], + "bbox": [ + 594, + 513, + 679, + 580 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/37c14f47adcb222d4da384c9736006a1eeb053e0319fd902d3f904720f19592c.jpg", + "image_caption": [ + "Test sample:", + "Label: Blond hair", + "Spur.Feat.:Male", + "Accuracy: $23.6\\%$" + ], + "image_footnote": [], + "bbox": [ + 697, + 513, + 782, + 579 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "P. Yang et al.", + "bbox": [ + 271, + 114, + 362, + 128 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/b81f59740d7fb43ffc6e795bfa460d39853adb7e490bd80ea790ee40f0bb855c.jpg", + "table_caption": [ + "Table 2: Worst-group accuracy and overall accuracy comparisons between Vanilla ResNet-34, group DRO, Score-Matching, InputGrad and our regularization on CelebA-Hair and Waterbirds datasets." + ], + "table_footnote": [], + "table_body": "
MethodWorst-Group Accuracy (%)Overall Accuracy (%)
CelebA-HairWaterbirdsCelebA-HairWaterbirds
Vanilla Model49.90 ± 8.6962.90 ± 0.1094.90 ± 0.3987.70 ± 0.08
Group DRO59.44 ± 5.9863.60 ± 0.1794.96 ± 0.2187.60 ± 0.05
Score-Matching Reg.59.78 ± 7.5658.19 ± 1.5593.46 ± 0.7185.64 ± 0.38
InputGrad Reg.82.66 ± 3.6358.18 ± 1.2292.12 ± 2.1885.50 ± 0.27
Our Reg.85.62 ± 5.3663.78 ± 2.8392.30 ± 1.3886.48 ± 0.38
Group DRO + Ours82.98 ± 4.6973.82 ± 2.2793.62 ± 0.7490.52 ± 0.17
", + "bbox": [ + 215, + 198, + 787, + 329 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "attribute maps [52] and insertion game [34] to demonstrate the effectiveness of our regularization in suppressing the use of the spuriously correlated gender feature and promoting the use of the robust hair feature, as shown in Fig. 1(b). This superiority leads to performance improvements on worst-case samples in the model trained with our regularization.", + "bbox": [ + 212, + 352, + 785, + 428 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Table 2 presents comparison of accuracy on worst-group samples and overall samples from the CelebA-Hair and Waterbirds datasets [30,39]. The Waterbirds dataset consists of synthetic bird images from CUB-200-2011 [53] and Places [60] datasets, incorporating spurious background features, such as land and water scenes, to confuse true labels of bird categories. We compare the proposed regularization method with InputGrad and Score-Matching regularizations [49], as well as Group DRO [39]. Score-Matching regularization is proposed to enhance the interpretability of the model by improving the alignment of implicit density models. Experimental results clearly demonstrate the effectiveness of our method in enhancing worst-group accuracy while maintaining overall sample accuracy. Furthermore, the performance gains can be further enhanced by incorporating our regularization technique into Group DRO. This enhancement highlights the efficacy of our regularization in terms of its practicality. More results using attribution maps and the insertion game on both CelebA-Hair and Waterbirds datasets are provided in Supp. 17. Moreover, out-of-distribution detection [21] is also performed on CIFAR [28] and SVHN [33] in Supp. 14.", + "bbox": [ + 212, + 429, + 787, + 672 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "7.3 Efficacy against Pixels, Gradients and Density Perturbations", + "text_level": 1, + "bbox": [ + 215, + 693, + 763, + 709 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "We first employ pixel perturbation [41, 55] to quantitatively compare the robustness of different models following Srinivas and Fleuret [49] who iteratively removed the most important input pixels identified by attribution maps for model robustness evaluation. Robust models are expected to exhibit increased sensitivity when removing the most important pixels and decreased sensitivity when removing the least important ones. We assess the difference in fractional output logit change between the images with the top and bottom $k\\%$ most salient pixels using SmoothGrad [48] on ResNet-18 [20] trained on CIFAR-10", + "bbox": [ + 212, + 719, + 787, + 840 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "Regulating Model Reliance on Non-Robust Features", + "bbox": [ + 383, + 114, + 730, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 784, + 126 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/4a556c4dc6a686a71038b588c4cf2cb3128f7e07eb4dc4fe0adb53238a35c9db.jpg", + "image_caption": [ + "(a) Pixel Perturbation on CIFAR-10 and CIFAR-100." + ], + "image_footnote": [], + "bbox": [ + 217, + 143, + 356, + 232 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/d063c23504d6e04835e0415096766816b2189cd8b350fa9825a055ef65593ab6.jpg", + "image_caption": [ + "Fig. 6: Robustness comparison. (a) Pixel perturbation experimental results on ResNet-18 trained on CIFAR-10 and CIFAR-100. Higher curves indicate better results. (b) Robustness of relative gradients and absolute density on CIFAR-10. Lower curves indicate better results." + ], + "image_footnote": [], + "bbox": [ + 357, + 143, + 496, + 232 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/57ccbb079a2189b48c2834c2a0f237da9f841be08faf3239d05910a3d8139e79.jpg", + "image_caption": [ + "(b) Robustness of Gradients and Density." + ], + "image_footnote": [], + "bbox": [ + 496, + 143, + 640, + 232 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/1c87bd4c79667abbf7b60f57532a6bfabf57e32c68bd9dd48ef483fd5d4367ae.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 640, + 143, + 785, + 232 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "and CIFAR-100 [28], as depicted in Fig. 6(a). Given the recognized challenge of reference image ambiguity in IG for explaining natural images [14, 51], we opt to use SmoothGrad for explaining predictions of images in the CIFAR dataset. SmoothGrad has demonstrated to outperform IG in such conditions [22, 50]. Recognizing their respective strengths, we use different attribution methods in appropriate conditions for reliable results. It can be observed that models trained with our regularization significantly outperform different robust regularizations including Score-Matching, InputGrad and CURE regularizations [32].", + "bbox": [ + 212, + 335, + 784, + 455 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "We also test the robustness of the relative input gradients $||\\nabla_{x}f(x + \\delta) - \\nabla_{x}f(x)||_{2} / ||\\nabla_{x}f(x)||_{2}$ and the absolute density through $\\sum_{i = 1}^{C}e^{f_i(x + \\delta) - f_i(x)}$ on the input with Gaussian noise $\\delta$ in increasing standard deviation on CIFAR-10 dataset, as shown in Fig. 6(b). We can observe that our regularization leads to competitive robustness for the relative gradient in comparison with regularizing the Hessian norm in CURE. Moreover, our regularization naturally leads to density robustness, which is associated with a strong generative ability of the models [49]. More tests on CIFAR-100 are provided in Supp. 13. Supp. 16 provides the visualizations calculated by activation optimization. These results affirm that our regularization improves both the discriminative and generative abilities of models.", + "bbox": [ + 212, + 457, + 785, + 625 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "8 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 652, + 359, + 669 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In this paper, we define robust and non-robust features from a feature attribution perspective, and establish a correlation between the smoothness of input marginal density and model reliance on non-robust features. This connection motivates us to propose a regularization that targets the gradients of the marginal density, aiming to regulate the reliance on non-robust features. Extensive experiments demonstrate the effectiveness of our regularization in boosting model robustness across a wide range of applications. We emphasize that our approach does not advocate for the complete removal of model reliance on non-robust features, but instead seeks to achieve a balance between model performance and robustness through appropriate regularization strength.", + "bbox": [ + 212, + 688, + 787, + 840 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "P. Yang et al.", + "bbox": [ + 271, + 114, + 362, + 128 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgments", + "text_level": 1, + "bbox": [ + 217, + 143, + 392, + 162 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "This research was supported by ARC Discovery Grant 190102443 and by Google Research under Google Research Scholar Program. Dr. Naveed Akhtar is a recipient of the Australian Research Council Discovery Early Career Researcher Award (project number DE230101058) funded by the Australian Government. Professor Ajmal Mian is the recipient of an Australian Research Council Future Fellowship Award (project number FT210100268) funded by the Australian Government.", + "bbox": [ + 212, + 176, + 787, + 282 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 306, + 321, + 321 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Adebayo, J., Muelly, M., Abelson, H., Kim, B.: Post hoc explanations may be ineffective for detecting unknown spurious correlation. In: International Conference on Learning Representations, ICLR (2022)", + "2. Adebayo, J., Muelly, M., Liccardi, I., Kim, B.: Debugging tests for model explanations. In: Advances in Neural Information Processing Systems, NeurIPS (2020)", + "3. Ancona, M., Ceolini, E., Öztireli, C., Gross, M.: Towards better understanding of gradient-based attribution methods for deep neural networks. In: International Conference on Learning Representations, ICLR (2018)", + "4. Anil, C., Lucas, J., Grosse, R.: Sorting out lipschitz function approximation. In: International Conference on Machine Learning, ICML (2019)", + "5. Blanchard, P., Higham, D.J., Higham, N.J.: Accurately computing the log-sum-exp and softmax functions. IMA Journal of Numerical Analysis 41(4), 2311-2330 (2021)", + "6. Bridle, J.: Training stochastic model recognition algorithms as networks can lead to maximum mutual information estimation of parameters. Advances in Neural Information Processing Systems, NeurIPS (1989)", + "7. Bridle, J.S.: Probabilistic interpretation of feedforward classification network outputs, with relationships to statistical pattern recognition. In: Neurocomputing: Algorithms, Architectures and Applications, pp. 227-236. Springer (1990)", + "8. Buolamwini, J., Gebru, T.: Gender shades: Intersectional accuracy disparities in commercial gender classification. In: ACM Conference on Fairness, Accountability, and Transparency, FAccT (2018)", + "9. Carlini, N., Wagner, D.: Towards evaluating the robustness of neural networks. In: IEEE Symposium on Security and Privacy, EuroS&P (2017)", + "10. Chen, J., Wu, X., Rastogi, V., Liang, Y., Jha, S.: Robust attribution regularization. In: Advances in Neural Information Processing Systems, NeurIPS (2019)", + "11. Dombrowski, A.K., Alber, M., Anders, C., Ackermann, M., Müller, K.R., Kessel, P.: Explanations can be manipulated and geometry is to blame. In: Advances in Neural Information Processing Systems, NeurIPS (2019)", + "12. Dombrowski, A.K., Anders, C.J., Müller, K.R., Kessel, P.: Towards robust explanations for deep neural networks. Pattern Recognition 121, 108194 (2022)", + "13. Drucker, H., LeCun, Y.: Improving generalization performance using double backpropagation. IEEE Trans. Neural Networks (1992)", + "14. Erion, G.G., Janizek, J.D., Sturmfels, P., Lundberg, S.M., Lee, S.: Improving performance of deep learning models with axiomatic attribution priors and expected gradients. Nature Machine Intelligence (2021)" + ], + "bbox": [ + 218, + 339, + 784, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "Regulating Model Reliance on Non-Robust Features", + "bbox": [ + 383, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "15. Etmann, C.: A closer look at double backpropagation. CoRR (2019)", + "16. Fong, R.C., Vedaldi, A.: Interpretable explanations of black boxes by meaningful perturbation. In: International Conference on Computer Vision, ICCV (2017)", + "17. Goodfellow, I.J., Shlens, J., Szegedy, C.: Explaining and harnessing adversarial examples. In: International Conference on Learning Representations, ICLR (2015)", + "18. Grathwohl, W., Wang, K.C., Jacobsen, J.H., Duvenaud, D., Norouzi, M., Swersky, K.: Your classifier is secretly an energy based model and you should treat it like one. In: International Conference on Learning Representations, ICLR (2020)", + "19. Hashimoto, T., Srivastava, M., Namkoong, H., Liang, P.: Fairness without demographics in repeated loss minimization. In: International Conference on Machine Learning, ICML (2018)", + "20. He, K., Zhang, X., Ren, S., Sun, J.: Identity mappings in deep residual networks. In: European Conference on Computer Vision, European Conference on Computer Vision, ECCV (2016)", + "21. Hendrycks, D., Gimpel, K.: A baseline for detecting misclassified and out-of-distribution examples in neural networks. CoRR (2016)", + "22. Hooker, S., Erhan, D., Kindermans, P., Kim, B.: A benchmark for interpretability methods in deep neural networks. In: Advances in Neural Information Processing Systems, NeurIPS (2019)", + "23. Hu, W., Niu, G., Sato, I., Sugiyama, M.: Does distributionally robust supervised learning give robust classifiers? In: International Conference on Machine Learning, ICML (2018)", + "24. Ilyas, A., Santurkar, S., Tsipras, D., Engstrom, L., Tran, B., Madry, A.: Adversarial examples are not bugs, they are features. Advances in neural information processing systems 32 (2019)", + "25. Ioffe, S., Szegedy, C.: Batch normalization: Accelerating deep network training by reducing internal covariate shift. In: International Conference on Machine Learning, ICML (2015)", + "26. Jiang, J., Wen, Z., Mansoor, A., Mian, A.: Efficient hyperparameter optimization with adaptive fidelity identification. In: IEEE Conference on Computer Vision and Pattern Recognition (2024)", + "27. Jiang, J., Wen, Z., Yang, P., Mansoor, A., Mian, A.: Fast-pgm: Fast probabilistic graphical model learning and inference. arXiv preprint arXiv:2405.15605 (2024)", + "28. Krizhevsky, A., Hinton, G., et al.: Learning multiple layers of features from tiny images. Technical Report, University of Toronto (2009)", + "29. LeCun, Y., Bottou, L., Bengio, Y., Haffner, P.: Gradient-based learning applied to document recognition. Proc. IEEE (1998)", + "30. Liu, Z., Luo, P., Wang, X., Tang, X.: Deep learning face attributes in the wild. In: International Conference on Computer Vision, ICCV (2015)", + "31. Madry, A., Makelov, A., Schmidt, L., Tsipras, D., Vladu, A.: Towards deep learning models resistant to adversarial attacks. In: International Conference on Learning Representations, ICLR (2018)", + "32. Moosavi-Dezfooli, S.M., Fawzi, A., Uesato, J., Frossard, P.: Robustness via curvature regularization, and vice versa. In: IEEE Conference on Computer Vision and Pattern Recognition, CVPR (2019)", + "33. Netzer, Y., Wang, T., Coates, A., Bissacco, A., Wu, B., Ng, A.Y.: Reading digits in natural images with unsupervised feature learning. Advances in Neural Information Processing Systems, NeurIPSW (2011)", + "34. Petsiuk, V., Das, A., Saenko, K.: RISE: randomized input sampling for explanation of black-box models. In: British Machine Vision Conference, BMVC (2018)" + ], + "bbox": [ + 215, + 147, + 785, + 840 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "P. Yang et al.", + "bbox": [ + 271, + 114, + 362, + 128 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "35. Ribeiro, M.T., Singh, S., Guestrin, C.: \" why should i trust you?\" explaining the predictions of any classifier. In: ACM SIGKDD Conference on Knowledge Discovery & Data Mining, KDD (2016)", + "36. Ross, A.S., Doshi-Velez, F.: Improving the adversarial robustness and interpretability of deep neural networks by regularizing their input gradients. In: AAAI Conference on Artificial Intelligence, AAAI (2018)", + "37. Ross, A.S., Hughes, M.C., Doshi-Velez, F.: Right for the right reasons: Training differentiable models by constraining their explanations. In: International Joint Conference on Artificial Intelligence, IJCAI (2017)", + "38. Rudin, C.: Stop explaining black box machine learning models for high stakes decisions and use interpretable models instead. Nature Machine Intelligence (2019)", + "39. Sagawa, S., Koh, P.W., Hashimoto, T.B., Liang, P.: Distributionally robust neural networks for group shifts: On the importance of regularization for worst-case generalization. In: International Conference on Learning Representations, ICLR (2019)", + "40. Salahuddin, Z., Woodruff, H.C., Chatterjee, A., Lambin, P.: Transparency of deep neural networks for medical image analysis: A review of interpretability methods. Computers in biology and medicine 140, 105111 (2022)", + "41. Samek, W., Binder, A., Montavon, G., Lapuschkin, S., Müller, K.: Evaluating the visualization of what a deep neural network has learned. TNNLS 28(11), 2660-2673 (2017)", + "42. Schramowski, P., Stammer, W., Teso, S., Brugger, A., Herbert, F., Shao, X., Luigs, H., Mahlein, A., Kersting, K.: Making deep neural networks right for the right scientific reasons by interacting with their explanations. Nature Machine Intelligence (2020)", + "43. Shafahi, A., Najibi, M., Ghiasi, M.A., Xu, Z., Dickerson, J., Studer, C., Davis, L.S., Taylor, G., Goldstein, T.: Adversarial training for free! In: Advances in Neural Information Processing Systems, NeurIPS (2019)", + "44. Shah, H., Jain, P., Netrapalli, P.: Do input gradients highlight discriminative features? In: Advances in Neural Information Processing Systems, NeurIPS (2021)", + "45. Shrikumar, A., Greenside, P., Kundaje, A.: Learning important features through propagating activation differences. In: International Conference on Machine Learning, ICML (2017)", + "46. Simon-Gabriel, C.J., Ollivier, Y., Bottou, L., Scholkopf, B., Lopez-Paz, D.: First-order adversarial vulnerability of neural networks and input dimension. In: International Conference on Machine Learning, ICML (2019)", + "47. Simonyan, K., Vedaldi, A., Zisserman, A.: Deep inside convolutional networks: Visualising image classification models and saliency maps. In: International Conference on Learning Representations, ICLR (2014)", + "48. Smilkov, D., Thorat, N., Kim, B., Viégas, F., Wattenberg, M.: Smoothgrad: removing noise by adding noise. CoRR abs/1706.03825 (2017)", + "49. Srinivas, S., Fleuret, F.: Rethinking the role of gradient-based attribution methods for model interpretability. In: International Conference on Learning Representations, ICLR (2021)", + "50. Srinivas, S., Fleuret, F.: Full-gradient representation for neural network visualization. In: Advances in Neural Information Processing Systems, NeurIPS (2019)", + "51. Sturmfels, P., Lundberg, S., Lee, S.I.: Visualizing the impact of feature attribution baselines. Distill (2020)", + "52. Sundararajan, M., Taly, A., Yan, Q.: Axiomatic attribution for deep networks. In: International Conference on Machine Learning, ICML (2017)" + ], + "bbox": [ + 212, + 146, + 787, + 840 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "Regulating Model Reliance on Non-Robust Features", + "bbox": [ + 383, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "53. Wah, C., Branson, S., Welinder, P., Perona, P., Belongie, S.: The caltech-ucsd birds-200-2011 dataset. California Institute of Technology (2011)", + "54. Yang, P., Akhtar, N., Jiang, J., Mian, A.: Backdoor-based explainable AI benchmark for high fidelity evaluation of attribution methods. arXiv preprint arXiv:2405.02344 (2024)", + "55. Yang, P., Akhtar, N., Wen, Z., Mian, A.: Local path integration for attribution. In: AAAI Conference on Artificial Intelligence, AAAI (2023)", + "56. Yang, P., Akhtar, N., Wen, Z., Shah, M., Mian, A.S.: Re-calibrating feature attributions for model interpretation. In: International Conference on Learning Representations (2023)", + "57. Yang, P., Wen, Z., Mian, A.: Multi-grained interpretable network for image recognition. In: 2022 26th International Conference on Pattern Recognition (ICPR). IEEE (2022)", + "58. Zeiler, M.D., Fergus, R.: Visualizing and understanding convolutional networks. In: European Conference on Computer Vision, ECCV (2014)", + "59. Zhang, X., Wang, N., Shen, H., Ji, S., Luo, X., Wang, T.: Interpretable deep learning under fire. In: USENIX Security (2020)", + "60. Zhou, B., Lapedriza, A., Khosla, A., Oliva, A., Torralba, A.: Places: A 10 million image database for scene recognition. TPAMI 40(6), 1452-1464 (2017)", + "61. Zintgraf, L.M., Cohen, T.S., Adel, T., Welling, M.: Visualizing deep neural network decisions: Prediction difference analysis. In: International Conference on Learning Representations, ICLR (2017)" + ], + "bbox": [ + 215, + 146, + 784, + 452 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "P. Yang et al.", + "bbox": [ + 271, + 114, + 362, + 128 + ], + "page_idx": 17 + } +] \ No newline at end of file diff --git a/2024/Regulating Model Reliance on Non-Robust Features by Smoothing Input Marginal Density/1125a5c6-c219-425e-8e55-ef1e3d508033_model.json b/2024/Regulating Model Reliance on Non-Robust Features by Smoothing Input Marginal Density/1125a5c6-c219-425e-8e55-ef1e3d508033_model.json new file mode 100644 index 0000000000000000000000000000000000000000..afe573b656fdd9ce2b81b210b647be2b257f8ce4 --- /dev/null +++ b/2024/Regulating Model Reliance on Non-Robust Features by Smoothing Input Marginal Density/1125a5c6-c219-425e-8e55-ef1e3d508033_model.json @@ -0,0 +1,2733 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.231, + 0.141, + 0.773, + 0.187 + ], + "angle": 0, + "content": "Regulating Model Reliance on Non-Robust Features by Smoothing Input Marginal Density" + }, + { + "type": "text", + "bbox": [ + 0.241, + 0.212, + 0.761, + 0.228 + ], + "angle": 0, + "content": "Peiyu Yang1, Naveed Akhtar2, Mubarak Shah3, and Ajmal Mian1" + }, + { + "type": "text", + "bbox": [ + 0.315, + 0.239, + 0.689, + 0.253 + ], + "angle": 0, + "content": "1 The University of Western Australia, Perth, Australia" + }, + { + "type": "text", + "bbox": [ + 0.321, + 0.254, + 0.68, + 0.267 + ], + "angle": 0, + "content": "2 The University of Melbourne, Melbourne, Australia" + }, + { + "type": "text", + "bbox": [ + 0.334, + 0.267, + 0.669, + 0.281 + ], + "angle": 0, + "content": "3 University of Central Florida, Orlando, America" + }, + { + "type": "text", + "bbox": [ + 0.264, + 0.283, + 0.743, + 0.309 + ], + "angle": 0, + "content": "peiyu.yang@research.uwa.edu.au, naveed.akhtar1@unimelb.edu.au, shah@crcv.ucf.edu, ajmal.mian@uwa.edu.au" + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.347, + 0.744, + 0.611 + ], + "angle": 0, + "content": "Abstract. Trustworthy machine learning necessitates meticulous regulation of model reliance on non-robust features. We propose a framework to delineate and regulate such features by attributing model predictions to the input. Within our approach, robust feature attributions exhibit a certain consistency, while non-robust feature attributions are susceptible to fluctuations. This behavior allows identification of correlation between model reliance on non-robust features and smoothness of marginal density of the input samples. Hence, we uniquely regularize the gradients of the marginal density w.r.t. the input features for robustness. We also devise an efficient implementation of our regularization4 to address the potential numerical instability of the underlying optimization process. Moreover, we analytically reveal that, as opposed to our marginal density smoothing, the prevalent input gradient regularization smoothens conditional or joint density of the input, which can cause limited robustness. Our experiments validate the effectiveness of the proposed method, providing clear evidence of its capability to address the feature leakage problem and mitigate spurious correlations. Extensive results further establish that our technique enables the model to exhibit robustness against perturbations in pixel values, input gradients, and density." + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.624, + 0.717, + 0.638 + ], + "angle": 0, + "content": "Keywords: Robust features \\(\\cdot\\) Regularization \\(\\cdot\\) Feature attributions" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.665, + 0.377, + 0.68 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.696, + 0.787, + 0.77 + ], + "angle": 0, + "content": "Research on mitigating model reliance on non-robust input features has recently gained increasing attention due to high-stake machine learning applications [12, 18,38,49]. In this paper, we advance this direction by introducing a regularization technique that promotes a smooth marginal probability density function of the input to regulate the model's reliance on non-robust features." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.772, + 0.788, + 0.817 + ], + "angle": 0, + "content": "To distinguish between robust and non-robust features, we leverage the notion of attributions [16, 52, 58]. For a model \\( f \\) parameterized by \\( \\theta \\), attributions characterize the importance of the \\( i \\)-th feature \\( x_{i} \\) of the input \\( x \\) for the model" + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.825, + 0.733, + 0.841 + ], + "angle": 0, + "content": "4 Our code is available at https://github.com/typeiyu/Input_density_reg." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.365, + 0.129 + ], + "angle": 0, + "content": "P. Yang et al." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.144, + 0.788, + 0.28 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.286, + 0.788, + 0.343 + ], + "angle": 0, + "content": "Fig. 1: Attribution maps [45] and insertion game scores [34] for samples from (a) BlockMNIST and (b) CelebA-Hair datasets. As compared to input gradient regularization, our regularization leads to lower feature leakage while also achieving higher AUC for the insertion game." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.356, + 0.788, + 0.552 + ], + "angle": 0, + "content": "prediction by quantifying the output change between \\( f(x; \\theta) \\) and \\( f(x_{[x_i = 0]}; \\theta) \\). Since robust input features contribute to model predictions equally well across slight condition variations, their attributions exhibit a certain consistency. On the other hand, non-robust feature attributions fluctuate under such variations. This potentially identifies a correlation between the model's reliance on non-robust features and the smoothness of the marginal probability density function of the input samples \\( p_\\theta(x) \\). This is an important insight for our contribution. For robustness, it offers a possibility of model regularization using the gradients of the marginal density with respect to the input \\( \\nabla_x p_\\theta(x) \\). We propose to regularize \\( \\nabla_x p_\\theta(x) \\) to encourage the model to prioritize the use of robust features and regulate its reliance on non-robust features. However, this can also lead to numerical instability during model optimization. To address that, we further introduce a stable and efficient implementation to estimate the gradient of marginal density." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.553, + 0.788, + 0.749 + ], + "angle": 0, + "content": "We also investigate input gradient norm regularization [13, 36, 37] and reveal that input gradients can be interpreted as input gradients of the log-conditional density \\(\\nabla_{x}\\log p_{\\theta}(x|y)\\) or log-joint density \\(\\nabla_{x}\\log p_{\\theta}(x,y)\\). Input gradient regularization mitigates the model's reliance on non-robust features specific to the class label \\(y = i\\), leading to unintentional blindness to class-specific non-robust features where \\(y \\neq i\\). In contrast, our regularization encourages smoothness of the marginal density \\(p_{\\theta}(x)\\) without imposing unintended constraints, providing a comprehensive regularization of the non-robust features. In Fig. 1, we employ attribution maps [45] and insertion game scores [34] to compare the robustness of vanilla models, input gradient regularized models and models trained with our regularization on BlockMNIST [44] and CelebA-Hair [30] datasets. As the representative examples show, our method suppresses both feature leakage and feature spurious correlation to improve model robustness and interpretability." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.841 + ], + "angle": 0, + "content": "The effectiveness of our approach is extensively established through a series of experiments. First, using BlockMNIST dataset [44], we demonstrate that the model trained with our regularization considerably mitigates the problem of feature leakage. This problem occurs when a model wrongly attributes importance to irrelevant but persistent features in the data to achieve better accuracy, e.g., the null block (☑) in Fig. 1. Addressing feature leakage helps suppress" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.384, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Regulating Model Reliance on Non-Robust Features" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.222 + ], + "angle": 0, + "content": "spurious correlation between the input features and model predictions [2, 39]. Moreover, we establish the robustness of the models trained with our regularization against perturbations from adversarial attacks [17, 31], input pixels [41] and input density, demonstrating the broad applicability of our approach. Our main contributions are summarized as follows." + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.237, + 0.784, + 0.28 + ], + "angle": 0, + "content": "1. We identify robust and non-robust features by leveraging attributions, and establish the correlation between model reliance on non-robust features and the smoothness of data marginal density." + }, + { + "type": "text", + "bbox": [ + 0.222, + 0.283, + 0.784, + 0.31 + ], + "angle": 0, + "content": "2. We propose an efficient technique for calculating the gradients of log density, also addressing the numerical instability of the underlying optimization." + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.313, + 0.784, + 0.356 + ], + "angle": 0, + "content": "3. Through extensive experiments, we demonstrate the effectiveness of the proposed regularization and additionally establish that our approach exhibits general robustness against perturbations." + }, + { + "type": "list", + "bbox": [ + 0.222, + 0.237, + 0.784, + 0.356 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.384, + 0.388, + 0.399 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.417, + 0.788, + 0.612 + ], + "angle": 0, + "content": "Regularization for Interpretability Robustness: Despite advancements in model transparency [26, 27, 40, 57], current deep neural networks still lack interpretability in their decision-making process, which is exacerbated by their reliance on non-robust input features. Prior studies, such as [1, 2, 44], identify that standard models are prone to relying on irrelevant or spuriously correlated features. To address that, several regularization techniques are proposed to improve model interpretability. In [14, 42], the authors incorporated prior knowledge into the model training process to regularize the model behavior. Dombrowski et al. [11, 12] found that regularizing the input Hessian using SoftPlus activations or weight decay can boost resilience against manipulated inputs. In [18], a joint energy-based model is trained as a discriminative model for improved robustness. Srinivas and Fleuret [49] enhanced the model interpretability by improving the alignment between the implicit density and the data distribution." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.614, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Regularization for Adversarial Robustness: In addition to the other sources of prediction unreliability, adversarial attacks can manipulate model outputs with imperceptible perturbations to inputs [17, 31]. To address this, adversarial training through data augmentation with adversarial samples is widely employed [9, 31, 43]. Certified adversarial robustness through regularizations [4, 36, 46] is another branch of methods to defend against adversarial perturbations. Inspired by the classic double backpropagation [13], Ross and Doshi-Velez [36] regularized the input gradient norm for adversarial robustness. Etmann [15] further explored different variants of double backpropagation regularizations for various real-world scenarios. Moosavi-Dezfooli et al. [32] also proposed regularization to encourage a low curvature for adversarial robustness. To improve adversarial robustness, Chen et al. [10] computed the norm of attributions integrated from clean samples to adversarial samples as the regularization term. Ilyas et al. [24] proposed a disentangling method to distinguish feature robustness for explaining adversarial examples. However, they do not focus on general feature robustness." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.364, + 0.129 + ], + "angle": 0, + "content": "P. Yang et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.192 + ], + "angle": 0, + "content": "From their perspective, non-robust features are highly predictive; yet adversarially brittle. In contrast, our work focuses on general robustness derived directly from natural images." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.193, + 0.788, + 0.404 + ], + "angle": 0, + "content": "Attribution Methods: Feature attribution methods are used to estimate the importance of input features for a model's prediction and can be categorized as either perturbation-based or back-propagation-based techniques. Perturbation-based methods [16,58,61] calculate the attribution scores by repeatedly perturbing the input features and analyzing the resulting effects on the model prediction. These methods are also extended for evaluating the reliability of the computed attributions [3,34,41,55]. The back-propagation-based techniques [47,52,54,56] estimate the attribution scores by computing the gradients or integrated gradients with respect to the input features in the backward propagation process. In contrast to perturbation-based techniques, back-propagation-based attribution methods offer notable advantages in terms of computational efficiency and reliability. Given the transparency of attributions, our work leverages the attribution framework to distinguish between robust and non-robust features, which enables us to systematically analyze the robustness of input features." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.427, + 0.6, + 0.444 + ], + "angle": 0, + "content": "3 Feature Robustness by Attributions" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.457, + 0.791, + 0.517 + ], + "angle": 0, + "content": "We first provide a framework for distinguishing between robust and non-robust features by analyzing their attributions. Herein, attribution inconsistencies among the features with distinct degrees of robustness identify a correlation between the model's reliance on non-robust features and the smoothness of output logits." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.518, + 0.788, + 0.623 + ], + "angle": 0, + "content": "Let us consider an input sample \\( x \\in \\mathbb{R}^n \\) with label \\( y \\in \\mathbb{R}^c \\) from a dataset \\( \\mathcal{D} \\), and a classifier \\( f: \\mathbb{R}^n \\to \\mathbb{R}^c \\) parameterized by \\( \\theta \\). We denote robust and non-robust features within the input \\( x \\) as \\( x_{rob}, x_{nrob} \\subseteq x \\). Consider an attribution method \\( \\phi: \\mathbb{R}^c \\to \\mathbb{R}^n \\) attributing model predictions to input features by estimating their importance, resulting in an attribution map \\( M = \\phi(f(x; \\theta)) \\). Inspired by the success of attributions in model explanation, we identify robust and non-robust features by leveraging their attributions." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.624, + 0.788, + 0.729 + ], + "angle": 0, + "content": "Without loss of generality, we assume that attributions \\(M\\) of the features can be estimated by calculating the change in output logits when these features are removed from the input, following perturbation-based methods [35,58]: \\(M_{x_{rob}} = f(x;\\theta) - f(x_{[x_{rob} = 0]};\\theta)\\) and \\(M_{x_{nrob}} = f(x;\\theta) - f(x_{[x_{nrob} = 0]};\\theta)\\). For ease of understanding, we alternatively use \\(f(x_{nrob};\\theta)\\) and \\(f(x_{rob};\\theta)\\) to represent attributions \\(M_{rob}\\) and \\(M_{nrob}\\) in the text to follow. We define robust features within the attribution framework as follows." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.74, + 0.788, + 0.802 + ], + "angle": 0, + "content": "Definition 1. A feature \\( x_{feat} \\) shared among different input instances under its domain \\( \\Delta_{x_{feat}} \\) is robust if, for a randomly chosen class \\( y = i \\), its attribution \\( M_{x_{feat}} \\) is bounded by a small constant \\( h \\) under a metric \\( c(\\cdot) \\), i.e., \\( c(f(x_{feat};\\theta) - f(\\tilde{x}_{feat};\\theta)) \\leq h : \\tilde{x}_{feat} \\in \\Delta_{x_{feat}} \\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Under Definition 1, robust features are expected to contribute consistently to the model's prediction across different input samples. Non-robust features, on" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.384, + 0.115, + 0.733, + 0.13 + ], + "angle": 0, + "content": "Regulating Model Reliance on Non-Robust Features" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.253 + ], + "angle": 0, + "content": "the other hand, are those that contribute to the prediction score inconsistently or only under specific conditions. Our focus here is on distinguishing between robust and non-robust features, without requiring to specify a particular metric. Definition 1 emphasizes attribution consistency for robust features rather than attribution positivity, thereby allowing for robust features that can also make a negative contribution to the model's prediction. Building further upon the above definition, we make the following remark." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.266, + 0.788, + 0.341 + ], + "angle": 0, + "content": "Remark 1. Robust features are largely condition-invariant in that they retain similar attributions despite slight changes to the input. In contrast, non-robust features are condition-specific in that their attributions either vary drastically with slightly varying input conditions, or behave robustly only under specific conditions." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.354, + 0.785, + 0.46 + ], + "angle": 0, + "content": "Robust features exhibit stable behavior across the input space, which is observable through consistent output logits \\( f(x_{rob};\\theta) \\) in classifiers regardless of the input instance or class \\( y \\). In contrast, non-robust features rely on specific conditions to exhibit a particular behavior, which is tailored to a specific class \\( y = i \\) or the input instance. Robust modeling aims for a stronger reliance on robust features for prediction. Due to the consistency of output logits \\( f(x;\\theta) \\) for robust features, a smooth \\( f(x;\\theta) \\) is a desirable property for model robustness." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.483, + 0.618, + 0.502 + ], + "angle": 0, + "content": "4 Smoothing Marginal Density of Input" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.515, + 0.784, + 0.56 + ], + "angle": 0, + "content": "Here, we establish the relation between model robustness and gradients of the input marginal density. Then, a robust regularization is proposed for regulating model reliance on non-robust features by smoothing marginal density." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.561, + 0.784, + 0.606 + ], + "angle": 0, + "content": "We commence our analytical analysis with probability density, following Bridle [7]. Given a class \\( y = i \\), a joint probability density function over the input with the output logit \\( f_{i}(x;\\theta) \\) is defined as" + }, + { + "type": "equation", + "bbox": [ + 0.411, + 0.617, + 0.785, + 0.634 + ], + "angle": 0, + "content": "\\[\np _ {\\theta} (x, y = i) = e ^ {f _ {i} (x; \\theta)} / Z _ {\\theta}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.645, + 0.785, + 0.737 + ], + "angle": 0, + "content": "where the constant \\( Z_{\\theta} = \\int e^{f_i(x;\\theta)}dx \\) is the partition function. \\( Z_{\\theta} \\) normalizes the input \\( x \\) to a probability density by integrating over all possible input points \\( x \\) in the input space via the model \\( f \\). By applying Bayes' rule, we eliminate the condition \\( y = i \\), resulting in the marginal density being defined solely on the input \\( x \\): \\( p_{\\theta}(x) = p_{\\theta}(y = i,x) / p_{\\theta}(y = i|x) \\). The conditional density function \\( p_{\\theta}(y = i|x) \\) can be further defined as" + }, + { + "type": "equation", + "bbox": [ + 0.399, + 0.748, + 0.785, + 0.766 + ], + "angle": 0, + "content": "\\[\np _ {\\theta} (y = i | x) = e ^ {f _ {i} (x; \\theta)} / Z _ {f (x; \\theta)}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.779, + 0.788, + 0.843 + ], + "angle": 0, + "content": "where \\( Z_{f(x;\\theta)} = \\sum_{i=1}^{C} e^{f_i(x;\\theta)} \\) is the partition function for the output logits \\( f_i(x;\\theta) \\) defined on all the \\( C \\) classes. To simplify the notation, we use \\( Z_{f(x)} \\) to represent \\( Z_{f(x;\\theta)} \\) in the subsequent discussion. Exploiting the symmetry property of the joint density defined in Eq. 1, i.e., \\( p_\\theta(x,y=i) = p_\\theta(y=i,x) \\), the" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.364, + 0.129 + ], + "angle": 0, + "content": "P. Yang et al." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.147, + 0.529, + 0.162 + ], + "angle": 0, + "content": "marginal density \\(p_{\\theta}(x)\\) can be expressed as" + }, + { + "type": "equation", + "bbox": [ + 0.389, + 0.175, + 0.786, + 0.211 + ], + "angle": 0, + "content": "\\[\np _ {\\theta} (x) = \\frac {e ^ {f _ {i} (x ; \\theta)} / Z _ {\\theta}}{e ^ {f _ {i} (x ; \\theta)} / Z _ {f (x)}} = \\frac {Z _ {f (x)}}{Z _ {\\theta}}. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.222, + 0.789, + 0.419 + ], + "angle": 0, + "content": "As identified in the previous section, smoothness of output logits across inputs encourages the use of robust features by the model. Hence, we consider the marginal density \\( p_{\\theta}(x) \\) defined on the output logits \\( f(x; \\theta) \\) across the input space. Promoting a smaller gradient of the marginal density with respect to the input \\( x \\), denoted as \\( \\nabla_x p_{\\theta}(x) \\), contributes to the smoothness of the output logits. Thus, a positive correlation can be established between the use of robust features and the smoothness of the probability marginal density \\( p_{\\theta}(x) \\). In particular, the smooth output logits of robust features across input samples suggest that these features will have relatively small gradients of the density \\( p_{\\theta}(x) \\) with respect to the input values. On the other hand, non-robust features with fluctuating output logits will have large gradients of the density that need to be suppressed during the training process for model robustness. Therefore, we can conclude with the following remark." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.433, + 0.79, + 0.48 + ], + "angle": 0, + "content": "Remark 2. Model reliance on non-robust features \\( x_{nrob} \\) can be regulated by regularizing the gradients of the marginal density \\( p_{\\theta}(x) \\) w.r.t. \\( x \\), and this regularization can be achieved through optimizing the model parameters." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.493, + 0.789, + 0.659 + ], + "angle": 0, + "content": "In the light of Remark 2., we propose a regularization term for minimizing the gradients of marginal density. However, computing the gradients of the marginal density \\(\\nabla_{x}p_{\\theta}(x)\\) is not feasible because the partition function defined on the entire input space is intractable. To avoid the estimation of the intractable \\(Z_{\\theta}\\), we instead compute the gradients with respect to the log density as \\(\\nabla_{x}\\log p_{\\theta}(x) = \\nabla_{x}\\log Z_{f(x)}\\). This is possible because \\(Z_{\\theta}\\) solely depends on the model parameter \\(\\theta\\) and not the input \\(x\\). Expanding the partition function \\(Z_{f(x)}\\) in \\(\\nabla_{x}\\log p_{\\theta}(x)\\), we obtain \\(\\nabla_{x}\\log p_{\\theta}(x) = \\sum_{i=1}^{C}\\nabla_{x}e^{f_{i}(x;\\theta)} / \\sum_{i=1}^{C}e^{f_{i}(x;\\theta)}\\). The \\(p\\)-norm of this gradient is computed as the regularization term. In the optimization process, the goal is to find the optimal parameter \\(\\theta^{*}\\), cf. Remark 2., by minimizing the loss \\(\\ell\\) as" + }, + { + "type": "equation", + "bbox": [ + 0.339, + 0.66, + 0.786, + 0.699 + ], + "angle": 0, + "content": "\\[\n\\theta^ {*} = \\min _ {\\theta} \\ell (f (x; \\theta), y) + \\lambda | | \\frac {\\sum_ {i = 1} ^ {C} \\nabla_ {x} e ^ {f _ {i} (x ; \\theta)}}{\\sum_ {i = 1} ^ {C} e ^ {f _ {i} (x ; \\theta)}} | | _ {p}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.704, + 0.789, + 0.734 + ], + "angle": 0, + "content": "where \\(\\lambda\\) indicates the magnitude of the coefficient for controlling the strength of the regularization." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.789, + 0.843 + ], + "angle": 0, + "content": "To regulate model reliance on non-robust features, our regularization encourages the smoothness of marginal density by regularizing its gradients. Since the output logit change in the log-marginal density \\( \\log p_{\\theta}(x) = \\log Z_{f(x)} - \\log Z_{\\theta} \\), and \\( Z_{\\theta} \\) is independent of the input \\( x \\), we can only focus on the first term \\( Z_{f(x)} = \\sum_{i=1}^{C} e^{f_i(x;\\theta)} \\). Recall, Definition 1 of robust features. Assuming the robust input feature \\( x_{rob} \\) exists in a random input \\( x \\), the corresponding output logit \\( f_i(x_{rob}) \\) will consistently attribute to the model prediction. This property" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.384, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "Regulating Model Reliance on Non-Robust Features" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.145, + 0.788, + 0.314 + ], + "angle": 0, + "content": "of \\( x_{rob} \\) leads to the smoothness of output change \\( \\sum_{i=1}^{C} e^{f_i(x_{rob};\\theta)} \\) across different input samples and class labels. In contrast, non-robust input features \\( x_{nrob} \\) show relatively high attributions for the output logit \\( f_i(x_{nrob}) \\) for a given class \\( i \\). However, they cannot maintain consistency in the attributions across different inputs or labels, leading to fluctuations in the output change \\( \\sum_{i=1}^{C} e^{f_i(x_{nrob};\\theta)} \\). In Fig. 1(a) and Fig. 13 in Supp. 17, non-robust features in the Null block exhibit fluctuating attributions across different samples on the standard trained model. It is demonstrated that the magnitude of gradients for input features in the marginal density \\( p_\\theta(x) \\) reflects the model's sensitivity to those features. We leverage this relation to mitigate model reliance on the non-robust features by smoothing the marginal density of the input samples." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.338, + 0.782, + 0.357 + ], + "angle": 0, + "content": "5 Stable and Efficient Implementation for Regularization" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.372, + 0.789, + 0.599 + ], + "angle": 0, + "content": "From the implementation perspective, the gradient computation of marginal density involves multiple exponential operations in both the numerator and denominator of Eq. 4 which can introduce numerical instability in the optimization process, leading to gradient vanishing and explosion problems. Such issues can potentially hinder the application of our regularization to large non-linear models or wide-distribution data. For instance, batch normalization (BN) layers [25] solving internal covariate shifts with learnable scaling and shifting parameters can amplify the errors caused by the exponential operations during the backpropagation. In Fig. 2(a), yellow and red curves indicate original and inverse implementations for minimizing the gradient of log density \\(\\nabla_{x}\\log Z_{f(x)}\\). It is illustrated that the Frobenius gradient norm of both implementations undergoes rapid numerical overflow as the number of training iterations increases, revealing the serious issue of exploding and vanishing gradients using the proposed regularization. Therefore, it is crucial to address these through careful implementation to ensure the feasibility of our regularization." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.599, + 0.788, + 0.675 + ], + "angle": 0, + "content": "To address this challenge, we transform the computation from the summation of exponential operations to softmax. By subtracting a constant value \\(\\eta\\) from the logits before exponentiation, softmax can prevent numerical overflow during the exponential operations, i.e., \\(e^{f_i(x;\\theta)} / \\sum_j e^{f_j(x;\\theta)} = e^{f_i(x;\\theta) - \\eta} / \\sum_j e^{f_j(x;\\theta) - \\eta}\\). Thus, we incorporate the softmax function into the computation of log density" + }, + { + "type": "image", + "bbox": [ + 0.258, + 0.683, + 0.407, + 0.779 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.277, + 0.785, + 0.396, + 0.797 + ], + "angle": 0, + "content": "(a) Numerical Stability." + }, + { + "type": "image", + "bbox": [ + 0.431, + 0.682, + 0.583, + 0.777 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.445, + 0.785, + 0.565, + 0.797 + ], + "angle": 0, + "content": "(b) Training Efficiency." + }, + { + "type": "image", + "bbox": [ + 0.596, + 0.684, + 0.745, + 0.782 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.634, + 0.785, + 0.726, + 0.797 + ], + "angle": 0, + "content": "(c) Training Time." + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.808, + 0.788, + 0.836 + ], + "angle": 0, + "content": "Fig. 2: The comparison of numerical stability, training efficiency and training time across different ResNet-34 models." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.365, + 0.129 + ], + "angle": 0, + "content": "P. Yang et al." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.147, + 0.383, + 0.164 + ], + "angle": 0, + "content": "gradient \\(\\nabla_{x}logz_{f(x)}\\) as" + }, + { + "type": "equation", + "bbox": [ + 0.32, + 0.176, + 0.684, + 0.213 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\nabla_ {x} l o g z _ {f (x)} = \\nabla_ {x} f _ {i} (x; \\theta) - \\nabla_ {x} (l o g e ^ {f _ {i} (x; \\theta)} - l o g z _ {f (x)}), \\\\ = \\nabla_ {x} f _ {i} (x; \\theta) - \\nabla_ {x} \\left(\\log \\left(f _ {i} (x; \\theta) / z _ {f (x)}\\right)\\right), \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.762, + 0.187, + 0.784, + 0.201 + ], + "angle": 0, + "content": "(5)" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.225, + 0.785, + 0.287 + ], + "angle": 0, + "content": "where \\( f_{i}(x) \\) indicates a logit of a random \\( i \\)-th class, and \\( z_{f(x)} \\) equals \\( \\sum_{j} e^{f_{j}(x; \\theta)} \\). Thus, the gradients with respect to the log-marginal density \\( p_{\\theta}(x) \\) can be replaced by computing the difference between the gradient \\( \\nabla_{x} f_{i}(x; \\theta) \\) and the gradient of a log-softmax output \\( \\nabla_{x} \\log(e^{f_{i}(x; \\theta)} / Z_{f(x)}) \\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.287, + 0.787, + 0.454 + ], + "angle": 0, + "content": "Our technique improves upon the common approach [5, 6] to achieve numerical stability in log exponential sum calculations, which typically employs the formula \\( \\log(\\sum_{i \\in \\{1, \\ldots, n\\}} e^{x_i}) = \\eta + \\log(\\sum_{i \\in \\{1, \\ldots, n\\}} e^{x_i - \\eta}) \\), with \\( \\eta \\) being the maximum value of inputs \\( \\{x_1, \\ldots, x_n\\} \\). We address the numerical instability by employing softmax, avoiding computationally expensive comparisons of the maximum gradient values. Specifically, the basic stable implementation involves finding the maximum gradient within the gradients of various classes, leading to a time complexity of \\( O(n) \\), where \\( n \\) represents the number of classes. In contrast, our method eliminates numerical instabilities associated with a randomly selected class, eliminating the need for maximum value comparisons and leading to a more favorable time complexity of \\( O(1) \\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.456, + 0.788, + 0.545 + ], + "angle": 0, + "content": "While our solution helps avoid numerical instability, it requires twice the gradient computations as compared to the sole calculation of density gradients. Hence, we propose an efficient mechanism for estimating the difference in the gradients. Specifically, we compute the gradient of the difference between two outputs to approximate the difference between the two gradients of outputs using Taylor series as" + }, + { + "type": "equation", + "bbox": [ + 0.244, + 0.557, + 0.785, + 0.575 + ], + "angle": 0, + "content": "\\[\n\\nabla_ {x} f _ {i} (x; \\theta) - \\nabla_ {x} \\log \\left(e ^ {f _ {i} (x; \\theta)} / Z _ {f (x)}\\right) \\approx \\nabla_ {x} \\left(f _ {i} (x; \\theta) - \\log \\left(e ^ {f _ {i} (x; \\theta)} / Z _ {f (x)}\\right)\\right). \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.587, + 0.788, + 0.753 + ], + "angle": 0, + "content": "The proof of Eq. 6 is provided in Supp. 9. As such, the proposed approach enables stable and efficient model optimization. The blue curve in Fig. 2(a) shows that our efficient implementation can effectively avoid numerical instability in the gradient computation. Fig. 2(b) provides a comparison of training efficiency on CIFAR-10 [28] between the basic numerically stable implementation and our two derived alternatives. More results regarding the comparison of numerical stability are provided in Supp. 15. In Fig. 2(c), we delve into a specific examination of the training times of models trained across the three datasets including BlockMNIST, CelebA and CIFAR-10. It is evident that our methods substantially improve the per-epoch training time. Moreover, our methods can yield further beneficial trade-offs as the number of classes increases." + }, + { + "type": "title", + "bbox": [ + 0.214, + 0.777, + 0.771, + 0.795 + ], + "angle": 0, + "content": "6 Limited Robustness in Input Gradient Regularization" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.842 + ], + "angle": 0, + "content": "Input gradient regularization (InputGrad Reg.) [29,36] computes the Frobenius norm of input gradients \\(||\\nabla_x f_i(x;\\theta)||_F\\) for a given class label \\(y = i\\), which is a" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.384, + 0.115, + 0.733, + 0.13 + ], + "angle": 0, + "content": "Regulating Model Reliance on Non-Robust Features" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.223 + ], + "angle": 0, + "content": "baseline robust regularization for model optimization. Existing works [29,36,37] explain InputGrad Reg. as a prediction stability technique for robustness against perturbations. Here, we reveal that the input gradient norm potentially regularizes the gradients of implicit data density. Moreover, we provide an understanding of how InputGrad Reg. encourages robustness as well as its limitations." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.223, + 0.784, + 0.253 + ], + "angle": 0, + "content": "Suppose all the classes have equal probability, i.e., \\( p_{\\theta}(y = i) = 1 / C \\). We can express the class-conditional density \\( p_{\\theta}(x|y) \\) by using Bayes's rule as" + }, + { + "type": "equation", + "bbox": [ + 0.37, + 0.266, + 0.786, + 0.299 + ], + "angle": 0, + "content": "\\[\np _ {\\theta} (x \\mid y = i) = \\frac {p _ {\\theta} (x , y = i)}{p _ {\\theta} (y = i)} = \\frac {e ^ {f _ {i} (x ; \\theta)}}{Z _ {\\theta} / C}. \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.309, + 0.784, + 0.339 + ], + "angle": 0, + "content": "Now, we compute the gradients of the log density defined in Eq. 7 with respect to the input \\( x \\) as" + }, + { + "type": "equation", + "bbox": [ + 0.314, + 0.354, + 0.785, + 0.369 + ], + "angle": 0, + "content": "\\[\n\\nabla_ {x} \\log p _ {\\theta} (x, y = i) = \\nabla_ {x} \\log p _ {\\theta} (x | y = i) = \\nabla_ {x} f _ {i} (x; \\theta). \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.381, + 0.788, + 0.592 + ], + "angle": 0, + "content": "Eq. 8 demonstrates that the input gradients can be interpreted as the gradients of either the log joint density or the log conditional density with respect to the input \\( x \\). This formulation highlights that InputGrad Reg. encourages consistent attributions of input features for the model's predictions. However, it is important to note that InputGrad Reg. is formulated under the specific condition \\( y = i \\), which limits its effectiveness in resolving inconsistencies when predicting a different class \\( y = j \\), where \\( j \\neq i \\). Although input features consistently contribute to the model's prediction under a given class, InputGrad Reg. fails to consider inconsistent attributions of these features across different classes, thereby allowing model non-robust behavior to exist. Consequently, a model trained with input gradient regularization may still exhibit spurious robustness relying on specific conditions. Fig. 1(a) illustrates an example where the model trained with InputGrad Reg. is incapable of suppressing non-robust features in Null block." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.605, + 0.788, + 0.651 + ], + "angle": 0, + "content": "Remark 3. Input gradient regularization smoothens the joint and conditional density of the input \\( x \\) under a specific label \\( y = i \\), compromising its ability to resist the class-specific non-robust features." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.663, + 0.788, + 0.755 + ], + "angle": 0, + "content": "In contrast to the existing works [36, 49] that highlight the reasons for Input-Grad Reg. efficacy, we reveal a weakness of this technique, cf. Remark 3.. Unlike regularizing gradients based on joint or conditional densities, our approach allows for more effective regularization without imposing the condition \\( y = i \\). Our method focuses on regularizing the gradients of marginal density \\( \\nabla_{x}p_{\\theta}(x) \\), thereby smoothing the output logits across the input samples." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.778, + 0.377, + 0.796 + ], + "angle": 0, + "content": "7 Experiments" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.84 + ], + "angle": 0, + "content": "In this section, we perform extensive experiments to validate the efficacy of our regularization and the newly established correlation between the smooth" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.364, + 0.129 + ], + "angle": 0, + "content": "P. Yang et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.239 + ], + "angle": 0, + "content": "marginal density and model reliance on non-robust features. Specifically, we present measurement results for addressing feature leakage and mitigating spurious correlations. In addition to specific applications, further results on the robustness against perturbations in input pixels, gradients and density are presented. Additional details about the datasets and the models used in our experiments can be found in Supp. 19." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.267, + 0.735, + 0.283 + ], + "angle": 0, + "content": "7.1 Efficacy against Feature Leakage and Adversarial Attacks" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.297, + 0.789, + 0.463 + ], + "angle": 0, + "content": "In [2, 44], it is demonstrated that deep models end up assigning importance to irrelevant input features. Shah et al. [44] used BlockMNIST in their experiments, which is a synthetic dataset extended from MNIST [29]. To each MNIST sample, BlockMNIST attaches a null block (an irrelevant pattern) randomly at the top or bottom of the image, as shown in Fig. 3(a). Shah et al. [44] observed that the explaining tool InputGrad [47] attributes importance to both the informative number block and the uninformative null block in the standard trained model. This phenomenon is termed as feature leakage by the authors. Interestingly, the issue is mitigated in adversarially trained models. Since this dataset allows for a controlled robustness assessment, we first evaluate our method on BlockMNIST to analyze the model's reliance on irrelevant features." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.465, + 0.784, + 0.48 + ], + "angle": 0, + "content": "Reproducibility and Quantitative Measurement of Feature Leakage." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.48, + 0.473, + 0.798 + ], + "angle": 0, + "content": "Owing to the unreliability of InputGrad caused by model saturation [45], we employ Integrated Gradients (IG) [52], an axiomatic explanation tool, to faithfully reinvestigate the feature leakage phenomenon. In Fig. 3(b)-(c), we show that the attributions computed by IG can reproduce different leakage phenomena from the informative block region to the null block region on standard and adversarially trained models. Feature leakage is an important phenomenon in the context of model robustness. However, there is a lack of a quantitative metric in the current literature to quantify its extent. We use integrated gradients to define the metric \\( M_{leakage} \\) to address this gap. Mathematically," + }, + { + "type": "image", + "bbox": [ + 0.487, + 0.489, + 0.78, + 0.574 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.522, + 0.575, + 0.707, + 0.587 + ], + "angle": 0, + "content": "(a) Input samples of BlockMNIST" + }, + { + "type": "image", + "bbox": [ + 0.483, + 0.591, + 0.779, + 0.647 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.483, + 0.651, + 0.78, + 0.677 + ], + "angle": 0, + "content": "(b) Integrated Gradients on (c) Integrated Gradients on standard trained model adversarially trained model" + }, + { + "type": "image_caption", + "bbox": [ + 0.48, + 0.687, + 0.789, + 0.77 + ], + "angle": 0, + "content": "Fig.3: BlockMNIST samples and feature leakage problem. (a) BlockMNIST randomly appends a null block at the top or bottom of MNIST samples. (b&c) Attribution maps are calculated by IG on the standard and adversarially trained models." + }, + { + "type": "equation", + "bbox": [ + 0.297, + 0.809, + 0.787, + 0.845 + ], + "angle": 0, + "content": "\\[\nM _ {\\text {l e a k a g e}} = \\mathbb {E} _ {x _ {n r o b} \\sim D} | | x _ {n r o b} \\times \\int_ {\\alpha = 0} ^ {1} \\frac {\\partial f (\\alpha \\cdot x _ {n r o b} ; \\theta))}{\\partial x _ {n r o b}} d \\alpha | | _ {2}, \\tag {9}\n\\]" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.384, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Regulating Model Reliance on Non-Robust Features" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.208 + ], + "angle": 0, + "content": "where \\(\\alpha\\) is the step from the absence to the presence of the features, and \\(x_{nrob}\\) specifies the non-robust features in the null block. Since the attributions of \\(x_{nrob}\\) for the model's prediction should ideally be zero, we use the \\(L_{2}\\) norm of attributions to quantify the extent of feature leakage." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.211, + 0.789, + 0.514 + ], + "angle": 0, + "content": "Robustness against Feature Leakage. Table 1 presents the experimental results on the BlockMNIST dataset. We compare our method with other robust regularizations and techniques including InputGrad [36], IG-SUM [10], SoftPlus activations [11] and Hessian [12]. InputGrad and Hessian regularize the first-order and second-order gradients w.r.t. the input. Models trained with SoftPlus activations and Hessian regularization fail to suppress the leakage problem, which indicates that feature leakage is not caused by the geometry of the model output manifold or high curvature [11, 59]. InputGrad regularization demonstrates robustness against both \\( L_{2} \\) and \\( L_{\\infty} \\) adversarial attacks, yet it still fails to address the leakage problem. The result aligns well with Remark 3, which highlights the allowance of non-robust features across different classes in InputGrad regularization. In addition, the use of IG in IG-Norm regularization that accumulates input gradients as a regularization term results in behavior similar to InputGrad regularization, leading to compromised robustness. These results further reveal that adversarial robustness is not a sufficient condition for suppressing feature leakage. Our method demonstrates a considerable improvement over other techniques for feature leakage, while also maintaining superiority in adversarial robustness. In Tab. 1, we use \\( L_{2} \\) norm for the compared regularization terms for fair benchmarking. We show in Supp. 10 that optimal norm exploration can yield an even more favorable trade-off for our technique for model robustness." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.517, + 0.79, + 0.639 + ], + "angle": 0, + "content": "Feature Leakage in Adversarily Robust Models. A FGSM adversarily trained model [17] augments the training samples by adversarial examples \\( x + \\epsilon \\cdot sign(\\nabla_x f_i(x; \\theta)) \\). Notably minimizing the loss of the perturbed input \\( x + \\epsilon \\cdot sign(\\nabla_x f_i(x; \\theta)) \\) is similar to the InputGrad regularization. Thus, training with FGSM is still limited in its ability to suppress the leakage problem. In contrast, PGD attack [31] weakens the effect of the condition \\( y = i \\) by iteratively searching for the perturbations from a random starting point. This process leads to a substantial enhancement in suppressing feature leakage, see Tab. 1." + }, + { + "type": "table_caption", + "bbox": [ + 0.215, + 0.662, + 0.79, + 0.704 + ], + "angle": 0, + "content": "Table 1: Results on BlockMNIST. Feature leakage, standard accuracy and adversarial accuracy under \\( L_{2} \\) and \\( L_{\\infty} \\) PGD-20 attacks are reported. ST & AT: Standard and Adversarial Training." + }, + { + "type": "table", + "bbox": [ + 0.216, + 0.717, + 0.794, + 0.855 + ], + "angle": 0, + "content": "
MethodFeature Leakage ↓PGD-20 (L2) ↑PGD-20 (L∞) ↑Accu. ↑
AT (FGSM)3.32487.480.0099.02
AT (PGD)2.31392.7528.0598.97
ST3.65773.570.0099.12
+ SoftPlus Acts.3.53367.950.0298.52
+ Hessian Reg.4.25880.060.0098.48
+ InputGrad Reg.3.46183.4621.1494.56
+ IG-SUM Reg.3.49782.1122.7392.87
+ Our Reg.2.25985.4129.3693.05
" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.364, + 0.129 + ], + "angle": 0, + "content": "P. Yang et al." + }, + { + "type": "image", + "bbox": [ + 0.217, + 0.146, + 0.353, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.223, + 0.227, + 0.346, + 0.238 + ], + "angle": 0, + "content": "(a) Feature Leakage." + }, + { + "type": "image", + "bbox": [ + 0.362, + 0.144, + 0.497, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.383, + 0.228, + 0.476, + 0.239 + ], + "angle": 0, + "content": "(b) \\(L_{\\infty}\\) Attack." + }, + { + "type": "image", + "bbox": [ + 0.507, + 0.144, + 0.642, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.532, + 0.228, + 0.617, + 0.24 + ], + "angle": 0, + "content": "(c) \\(L_{2}\\) Attack." + }, + { + "type": "image", + "bbox": [ + 0.653, + 0.144, + 0.787, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.679, + 0.228, + 0.76, + 0.239 + ], + "angle": 0, + "content": "(d) Accuracy." + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.251, + 0.785, + 0.294 + ], + "angle": 0, + "content": "Fig. 4: Performance comparison between our method and InputGrad regularization under varying regularization coefficient for (a) Feature leakage, (b)-(c) Adversarial Accuracy under \\( L_{\\infty} \\) and \\( L_{2} \\) PGD-20 attacks, and (d) Accuracy." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.306, + 0.785, + 0.442 + ], + "angle": 0, + "content": "Magnitude of Coefficient for Regularization. Figs. 4(a)-(d) present a comparison of results for feature leakage and adversarial accuracy under PGD attacks, as well as the standard accuracy across varying magnitudes for the regularization strength. The results affirm that our method effectively regulates feature leakage by imposing a penalty on non-robust features. Moreover, our regularization enables the model to defend against both \\( L_{2} \\) and \\( L_{\\infty} \\) attacks while maintaining high accuracy, showing an outstanding trade-off across four criteria. More adversarial robustness comparisons on CIFAR dataset [28] are reported in Supp. 11." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.466, + 0.533, + 0.481 + ], + "angle": 0, + "content": "7.2 Efficacy for Spurious Correlation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.493, + 0.487, + 0.704 + ], + "angle": 0, + "content": "Recent research has highlighted the susceptibility of neural models in learning spurious correlations that enhance performance on a given data but fail to generalize [2,8,39]. For instance, in CelebA-Hair dataset [30], which commonly consists of samples containing female celebrities with blond hair and male celebrities with dark hair, models heavily rely on the spuriously correlated gender feature to predict the target hair color [19, 39]. Consequently, accuracy tends to be lower for the samples containing" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.705, + 0.528, + 0.719 + ], + "angle": 0, + "content": "male celebrities with blond hair, see Fig. 5." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.785, + 0.84 + ], + "angle": 0, + "content": "To mitigate spurious correlations, distributional robust optimization (DRO) techniques have been proposed to re-weight the training loss of input samples from different groups [23,39]. In our regularization, an additional penalty is imposed to penalize the model's reliance on these spuriously correlated features because of their inconsistent attributions. This encourages the use of robust features while suppressing the model's reliance on spuriously correlated features. Reliable quantification of the model's robustness on natural images for spurious correlation suppression is an unresolved issue in the literature. We employ" + }, + { + "type": "image_caption", + "bbox": [ + 0.502, + 0.503, + 0.642, + 0.514 + ], + "angle": 0, + "content": "Common training samples:" + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.514, + 0.589, + 0.581 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.501, + 0.582, + 0.588, + 0.591 + ], + "angle": 0, + "content": "Label: Blond hair" + }, + { + "type": "image_caption", + "bbox": [ + 0.501, + 0.592, + 0.592, + 0.602 + ], + "angle": 0, + "content": "Spur. Feat.: Female" + }, + { + "type": "image_caption", + "bbox": [ + 0.502, + 0.603, + 0.583, + 0.612 + ], + "angle": 0, + "content": "Accuracy: \\(92.1\\%\\)" + }, + { + "type": "image", + "bbox": [ + 0.596, + 0.515, + 0.681, + 0.581 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.597, + 0.582, + 0.677, + 0.591 + ], + "angle": 0, + "content": "Label: Dark hair" + }, + { + "type": "image_caption", + "bbox": [ + 0.597, + 0.592, + 0.677, + 0.602 + ], + "angle": 0, + "content": "Spur. Feat.: Male" + }, + { + "type": "image_caption", + "bbox": [ + 0.597, + 0.603, + 0.677, + 0.612 + ], + "angle": 0, + "content": "Accuracy: \\(99.7\\%\\)" + }, + { + "type": "image_caption", + "bbox": [ + 0.699, + 0.503, + 0.765, + 0.514 + ], + "angle": 0, + "content": "Test sample:" + }, + { + "type": "image", + "bbox": [ + 0.699, + 0.515, + 0.784, + 0.58 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.7, + 0.582, + 0.784, + 0.591 + ], + "angle": 0, + "content": "Label: Blond hair" + }, + { + "type": "image_caption", + "bbox": [ + 0.7, + 0.592, + 0.78, + 0.602 + ], + "angle": 0, + "content": "Spur.Feat.:Male" + }, + { + "type": "image_caption", + "bbox": [ + 0.7, + 0.603, + 0.78, + 0.612 + ], + "angle": 0, + "content": "Accuracy: \\(23.6\\%\\)" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.384, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Regulating Model Reliance on Non-Robust Features" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.788, + 0.187 + ], + "angle": 0, + "content": "Table 2: Worst-group accuracy and overall accuracy comparisons between Vanilla ResNet-34, group DRO, Score-Matching, InputGrad and our regularization on CelebA-Hair and Waterbirds datasets." + }, + { + "type": "table", + "bbox": [ + 0.217, + 0.199, + 0.789, + 0.33 + ], + "angle": 0, + "content": "
MethodWorst-Group Accuracy (%)Overall Accuracy (%)
CelebA-HairWaterbirdsCelebA-HairWaterbirds
Vanilla Model49.90 ± 8.6962.90 ± 0.1094.90 ± 0.3987.70 ± 0.08
Group DRO59.44 ± 5.9863.60 ± 0.1794.96 ± 0.2187.60 ± 0.05
Score-Matching Reg.59.78 ± 7.5658.19 ± 1.5593.46 ± 0.7185.64 ± 0.38
InputGrad Reg.82.66 ± 3.6358.18 ± 1.2292.12 ± 2.1885.50 ± 0.27
Our Reg.85.62 ± 5.3663.78 ± 2.8392.30 ± 1.3886.48 ± 0.38
Group DRO + Ours82.98 ± 4.6973.82 ± 2.2793.62 ± 0.7490.52 ± 0.17
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.353, + 0.787, + 0.429 + ], + "angle": 0, + "content": "attribute maps [52] and insertion game [34] to demonstrate the effectiveness of our regularization in suppressing the use of the spuriously correlated gender feature and promoting the use of the robust hair feature, as shown in Fig. 1(b). This superiority leads to performance improvements on worst-case samples in the model trained with our regularization." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.43, + 0.789, + 0.673 + ], + "angle": 0, + "content": "Table 2 presents comparison of accuracy on worst-group samples and overall samples from the CelebA-Hair and Waterbirds datasets [30,39]. The Waterbirds dataset consists of synthetic bird images from CUB-200-2011 [53] and Places [60] datasets, incorporating spurious background features, such as land and water scenes, to confuse true labels of bird categories. We compare the proposed regularization method with InputGrad and Score-Matching regularizations [49], as well as Group DRO [39]. Score-Matching regularization is proposed to enhance the interpretability of the model by improving the alignment of implicit density models. Experimental results clearly demonstrate the effectiveness of our method in enhancing worst-group accuracy while maintaining overall sample accuracy. Furthermore, the performance gains can be further enhanced by incorporating our regularization technique into Group DRO. This enhancement highlights the efficacy of our regularization in terms of its practicality. More results using attribution maps and the insertion game on both CelebA-Hair and Waterbirds datasets are provided in Supp. 17. Moreover, out-of-distribution detection [21] is also performed on CIFAR [28] and SVHN [33] in Supp. 14." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.694, + 0.764, + 0.71 + ], + "angle": 0, + "content": "7.3 Efficacy against Pixels, Gradients and Density Perturbations" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.788, + 0.841 + ], + "angle": 0, + "content": "We first employ pixel perturbation [41, 55] to quantitatively compare the robustness of different models following Srinivas and Fleuret [49] who iteratively removed the most important input pixels identified by attribution maps for model robustness evaluation. Robust models are expected to exhibit increased sensitivity when removing the most important pixels and decreased sensitivity when removing the least important ones. We assess the difference in fractional output logit change between the images with the top and bottom \\( k\\% \\) most salient pixels using SmoothGrad [48] on ResNet-18 [20] trained on CIFAR-10" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.364, + 0.129 + ], + "angle": 0, + "content": "P. Yang et al." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.145, + 0.357, + 0.233 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.225, + 0.236, + 0.498, + 0.248 + ], + "angle": 0, + "content": "(a) Pixel Perturbation on CIFAR-10 and CIFAR-100." + }, + { + "type": "image", + "bbox": [ + 0.358, + 0.145, + 0.498, + 0.233 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.498, + 0.144, + 0.642, + 0.233 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.642, + 0.144, + 0.787, + 0.233 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.55, + 0.236, + 0.763, + 0.249 + ], + "angle": 0, + "content": "(b) Robustness of Gradients and Density." + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.26, + 0.785, + 0.314 + ], + "angle": 0, + "content": "Fig. 6: Robustness comparison. (a) Pixel perturbation experimental results on ResNet-18 trained on CIFAR-10 and CIFAR-100. Higher curves indicate better results. (b) Robustness of relative gradients and absolute density on CIFAR-10. Lower curves indicate better results." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.336, + 0.785, + 0.457 + ], + "angle": 0, + "content": "and CIFAR-100 [28], as depicted in Fig. 6(a). Given the recognized challenge of reference image ambiguity in IG for explaining natural images [14, 51], we opt to use SmoothGrad for explaining predictions of images in the CIFAR dataset. SmoothGrad has demonstrated to outperform IG in such conditions [22, 50]. Recognizing their respective strengths, we use different attribution methods in appropriate conditions for reliable results. It can be observed that models trained with our regularization significantly outperform different robust regularizations including Score-Matching, InputGrad and CURE regularizations [32]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.458, + 0.787, + 0.625 + ], + "angle": 0, + "content": "We also test the robustness of the relative input gradients \\(||\\nabla_{x}f(x + \\delta) - \\nabla_{x}f(x)||_{2} / ||\\nabla_{x}f(x)||_{2}\\) and the absolute density through \\(\\sum_{i = 1}^{C}e^{f_i(x + \\delta) - f_i(x)}\\) on the input with Gaussian noise \\(\\delta\\) in increasing standard deviation on CIFAR-10 dataset, as shown in Fig. 6(b). We can observe that our regularization leads to competitive robustness for the relative gradient in comparison with regularizing the Hessian norm in CURE. Moreover, our regularization naturally leads to density robustness, which is associated with a strong generative ability of the models [49]. More tests on CIFAR-100 are provided in Supp. 13. Supp. 16 provides the visualizations calculated by activation optimization. These results affirm that our regularization improves both the discriminative and generative abilities of models." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.653, + 0.36, + 0.67 + ], + "angle": 0, + "content": "8 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.788, + 0.841 + ], + "angle": 0, + "content": "In this paper, we define robust and non-robust features from a feature attribution perspective, and establish a correlation between the smoothness of input marginal density and model reliance on non-robust features. This connection motivates us to propose a regularization that targets the gradients of the marginal density, aiming to regulate the reliance on non-robust features. Extensive experiments demonstrate the effectiveness of our regularization in boosting model robustness across a wide range of applications. We emphasize that our approach does not advocate for the complete removal of model reliance on non-robust features, but instead seeks to achieve a balance between model performance and robustness through appropriate regularization strength." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.384, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Regulating Model Reliance on Non-Robust Features" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.393, + 0.163 + ], + "angle": 0, + "content": "Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.178, + 0.788, + 0.283 + ], + "angle": 0, + "content": "This research was supported by ARC Discovery Grant 190102443 and by Google Research under Google Research Scholar Program. Dr. Naveed Akhtar is a recipient of the Australian Research Council Discovery Early Career Researcher Award (project number DE230101058) funded by the Australian Government. Professor Ajmal Mian is the recipient of an Australian Research Council Future Fellowship Award (project number FT210100268) funded by the Australian Government." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.308, + 0.323, + 0.323 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.34, + 0.785, + 0.381 + ], + "angle": 0, + "content": "1. Adebayo, J., Muelly, M., Abelson, H., Kim, B.: Post hoc explanations may be ineffective for detecting unknown spurious correlation. In: International Conference on Learning Representations, ICLR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.383, + 0.785, + 0.409 + ], + "angle": 0, + "content": "2. Adebayo, J., Muelly, M., Liccardi, I., Kim, B.: Debugging tests for model explanations. In: Advances in Neural Information Processing Systems, NeurIPS (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.411, + 0.785, + 0.451 + ], + "angle": 0, + "content": "3. Ancona, M., Ceolini, E., Öztireli, C., Gross, M.: Towards better understanding of gradient-based attribution methods for deep neural networks. In: International Conference on Learning Representations, ICLR (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.452, + 0.785, + 0.478 + ], + "angle": 0, + "content": "4. Anil, C., Lucas, J., Grosse, R.: Sorting out lipschitz function approximation. In: International Conference on Machine Learning, ICML (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.48, + 0.785, + 0.52 + ], + "angle": 0, + "content": "5. Blanchard, P., Higham, D.J., Higham, N.J.: Accurately computing the log-sum-exp and softmax functions. IMA Journal of Numerical Analysis 41(4), 2311-2330 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.522, + 0.785, + 0.561 + ], + "angle": 0, + "content": "6. Bridle, J.: Training stochastic model recognition algorithms as networks can lead to maximum mutual information estimation of parameters. Advances in Neural Information Processing Systems, NeurIPS (1989)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.564, + 0.785, + 0.604 + ], + "angle": 0, + "content": "7. Bridle, J.S.: Probabilistic interpretation of feedforward classification network outputs, with relationships to statistical pattern recognition. In: Neurocomputing: Algorithms, Architectures and Applications, pp. 227-236. Springer (1990)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.606, + 0.785, + 0.645 + ], + "angle": 0, + "content": "8. Buolamwini, J., Gebru, T.: Gender shades: Intersectional accuracy disparities in commercial gender classification. In: ACM Conference on Fairness, Accountability, and Transparency, FAccT (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.647, + 0.785, + 0.673 + ], + "angle": 0, + "content": "9. Carlini, N., Wagner, D.: Towards evaluating the robustness of neural networks. In: IEEE Symposium on Security and Privacy, EuroS&P (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.221, + 0.675, + 0.785, + 0.701 + ], + "angle": 0, + "content": "10. Chen, J., Wu, X., Rastogi, V., Liang, Y., Jha, S.: Robust attribution regularization. In: Advances in Neural Information Processing Systems, NeurIPS (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.703, + 0.785, + 0.742 + ], + "angle": 0, + "content": "11. Dombrowski, A.K., Alber, M., Anders, C., Ackermann, M., Müller, K.R., Kessel, P.: Explanations can be manipulated and geometry is to blame. In: Advances in Neural Information Processing Systems, NeurIPS (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.744, + 0.785, + 0.77 + ], + "angle": 0, + "content": "12. Dombrowski, A.K., Anders, C.J., Müller, K.R., Kessel, P.: Towards robust explanations for deep neural networks. Pattern Recognition 121, 108194 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.772, + 0.785, + 0.798 + ], + "angle": 0, + "content": "13. Drucker, H., LeCun, Y.: Improving generalization performance using double backpropagation. IEEE Trans. Neural Networks (1992)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.8, + 0.785, + 0.84 + ], + "angle": 0, + "content": "14. Erion, G.G., Janizek, J.D., Sturmfels, P., Lundberg, S.M., Lee, S.: Improving performance of deep learning models with axiomatic attribution priors and expected gradients. Nature Machine Intelligence (2021)" + }, + { + "type": "list", + "bbox": [ + 0.22, + 0.34, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.364, + 0.129 + ], + "angle": 0, + "content": "P. Yang et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.148, + 0.697, + 0.162 + ], + "angle": 0, + "content": "15. Etmann, C.: A closer look at double backpropagation. CoRR (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.162, + 0.786, + 0.19 + ], + "angle": 0, + "content": "16. Fong, R.C., Vedaldi, A.: Interpretable explanations of black boxes by meaningful perturbation. In: International Conference on Computer Vision, ICCV (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.19, + 0.786, + 0.218 + ], + "angle": 0, + "content": "17. Goodfellow, I.J., Shlens, J., Szegedy, C.: Explaining and harnessing adversarial examples. In: International Conference on Learning Representations, ICLR (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.218, + 0.786, + 0.259 + ], + "angle": 0, + "content": "18. Grathwohl, W., Wang, K.C., Jacobsen, J.H., Duvenaud, D., Norouzi, M., Swersky, K.: Your classifier is secretly an energy based model and you should treat it like one. In: International Conference on Learning Representations, ICLR (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.259, + 0.786, + 0.3 + ], + "angle": 0, + "content": "19. Hashimoto, T., Srivastava, M., Namkoong, H., Liang, P.: Fairness without demographics in repeated loss minimization. In: International Conference on Machine Learning, ICML (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.3, + 0.786, + 0.342 + ], + "angle": 0, + "content": "20. He, K., Zhang, X., Ren, S., Sun, J.: Identity mappings in deep residual networks. In: European Conference on Computer Vision, European Conference on Computer Vision, ECCV (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.342, + 0.786, + 0.37 + ], + "angle": 0, + "content": "21. Hendrycks, D., Gimpel, K.: A baseline for detecting misclassified and out-of-distribution examples in neural networks. CoRR (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.37, + 0.786, + 0.411 + ], + "angle": 0, + "content": "22. Hooker, S., Erhan, D., Kindermans, P., Kim, B.: A benchmark for interpretability methods in deep neural networks. In: Advances in Neural Information Processing Systems, NeurIPS (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.411, + 0.786, + 0.453 + ], + "angle": 0, + "content": "23. Hu, W., Niu, G., Sato, I., Sugiyama, M.: Does distributionally robust supervised learning give robust classifiers? In: International Conference on Machine Learning, ICML (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.453, + 0.786, + 0.495 + ], + "angle": 0, + "content": "24. Ilyas, A., Santurkar, S., Tsipras, D., Engstrom, L., Tran, B., Madry, A.: Adversarial examples are not bugs, they are features. Advances in neural information processing systems 32 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.495, + 0.786, + 0.536 + ], + "angle": 0, + "content": "25. Ioffe, S., Szegedy, C.: Batch normalization: Accelerating deep network training by reducing internal covariate shift. In: International Conference on Machine Learning, ICML (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.536, + 0.786, + 0.577 + ], + "angle": 0, + "content": "26. Jiang, J., Wen, Z., Mansoor, A., Mian, A.: Efficient hyperparameter optimization with adaptive fidelity identification. In: IEEE Conference on Computer Vision and Pattern Recognition (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.577, + 0.786, + 0.606 + ], + "angle": 0, + "content": "27. Jiang, J., Wen, Z., Yang, P., Mansoor, A., Mian, A.: Fast-pgm: Fast probabilistic graphical model learning and inference. arXiv preprint arXiv:2405.15605 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.606, + 0.786, + 0.633 + ], + "angle": 0, + "content": "28. Krizhevsky, A., Hinton, G., et al.: Learning multiple layers of features from tiny images. Technical Report, University of Toronto (2009)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.633, + 0.786, + 0.661 + ], + "angle": 0, + "content": "29. LeCun, Y., Bottou, L., Bengio, Y., Haffner, P.: Gradient-based learning applied to document recognition. Proc. IEEE (1998)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.661, + 0.786, + 0.689 + ], + "angle": 0, + "content": "30. Liu, Z., Luo, P., Wang, X., Tang, X.: Deep learning face attributes in the wild. In: International Conference on Computer Vision, ICCV (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.689, + 0.786, + 0.73 + ], + "angle": 0, + "content": "31. Madry, A., Makelov, A., Schmidt, L., Tsipras, D., Vladu, A.: Towards deep learning models resistant to adversarial attacks. In: International Conference on Learning Representations, ICLR (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.73, + 0.786, + 0.771 + ], + "angle": 0, + "content": "32. Moosavi-Dezfooli, S.M., Fawzi, A., Uesato, J., Frossard, P.: Robustness via curvature regularization, and vice versa. In: IEEE Conference on Computer Vision and Pattern Recognition, CVPR (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.771, + 0.786, + 0.813 + ], + "angle": 0, + "content": "33. Netzer, Y., Wang, T., Coates, A., Bissacco, A., Wu, B., Ng, A.Y.: Reading digits in natural images with unsupervised feature learning. Advances in Neural Information Processing Systems, NeurIPSW (2011)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.813, + 0.786, + 0.841 + ], + "angle": 0, + "content": "34. Petsiuk, V., Das, A., Saenko, K.: RISE: randomized input sampling for explanation of black-box models. In: British Machine Vision Conference, BMVC (2018)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.148, + 0.786, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.384, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Regulating Model Reliance on Non-Robust Features" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.189 + ], + "angle": 0, + "content": "35. Ribeiro, M.T., Singh, S., Guestrin, C.: \" why should i trust you?\" explaining the predictions of any classifier. In: ACM SIGKDD Conference on Knowledge Discovery & Data Mining, KDD (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.19, + 0.788, + 0.232 + ], + "angle": 0, + "content": "36. Ross, A.S., Doshi-Velez, F.: Improving the adversarial robustness and interpretability of deep neural networks by regularizing their input gradients. In: AAAI Conference on Artificial Intelligence, AAAI (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.232, + 0.788, + 0.273 + ], + "angle": 0, + "content": "37. Ross, A.S., Hughes, M.C., Doshi-Velez, F.: Right for the right reasons: Training differentiable models by constraining their explanations. In: International Joint Conference on Artificial Intelligence, IJCAI (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.273, + 0.788, + 0.3 + ], + "angle": 0, + "content": "38. Rudin, C.: Stop explaining black box machine learning models for high stakes decisions and use interpretable models instead. Nature Machine Intelligence (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.3, + 0.788, + 0.355 + ], + "angle": 0, + "content": "39. Sagawa, S., Koh, P.W., Hashimoto, T.B., Liang, P.: Distributionally robust neural networks for group shifts: On the importance of regularization for worst-case generalization. In: International Conference on Learning Representations, ICLR (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.356, + 0.788, + 0.398 + ], + "angle": 0, + "content": "40. Salahuddin, Z., Woodruff, H.C., Chatterjee, A., Lambin, P.: Transparency of deep neural networks for medical image analysis: A review of interpretability methods. Computers in biology and medicine 140, 105111 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.398, + 0.788, + 0.439 + ], + "angle": 0, + "content": "41. Samek, W., Binder, A., Montavon, G., Lapuschkin, S., Müller, K.: Evaluating the visualization of what a deep neural network has learned. TNNLS 28(11), 2660-2673 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.439, + 0.788, + 0.495 + ], + "angle": 0, + "content": "42. Schramowski, P., Stammer, W., Teso, S., Brugger, A., Herbert, F., Shao, X., Luigs, H., Mahlein, A., Kersting, K.: Making deep neural networks right for the right scientific reasons by interacting with their explanations. Nature Machine Intelligence (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.495, + 0.788, + 0.536 + ], + "angle": 0, + "content": "43. Shafahi, A., Najibi, M., Ghiasi, M.A., Xu, Z., Dickerson, J., Studer, C., Davis, L.S., Taylor, G., Goldstein, T.: Adversarial training for free! In: Advances in Neural Information Processing Systems, NeurIPS (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.536, + 0.788, + 0.563 + ], + "angle": 0, + "content": "44. Shah, H., Jain, P., Netrapalli, P.: Do input gradients highlight discriminative features? In: Advances in Neural Information Processing Systems, NeurIPS (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.564, + 0.788, + 0.605 + ], + "angle": 0, + "content": "45. Shrikumar, A., Greenside, P., Kundaje, A.: Learning important features through propagating activation differences. In: International Conference on Machine Learning, ICML (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.605, + 0.788, + 0.647 + ], + "angle": 0, + "content": "46. Simon-Gabriel, C.J., Ollivier, Y., Bottou, L., Scholkopf, B., Lopez-Paz, D.: First-order adversarial vulnerability of neural networks and input dimension. In: International Conference on Machine Learning, ICML (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.647, + 0.788, + 0.688 + ], + "angle": 0, + "content": "47. Simonyan, K., Vedaldi, A., Zisserman, A.: Deep inside convolutional networks: Visualising image classification models and saliency maps. In: International Conference on Learning Representations, ICLR (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.688, + 0.788, + 0.716 + ], + "angle": 0, + "content": "48. Smilkov, D., Thorat, N., Kim, B., Viégas, F., Wattenberg, M.: Smoothgrad: removing noise by adding noise. CoRR abs/1706.03825 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.716, + 0.788, + 0.758 + ], + "angle": 0, + "content": "49. Srinivas, S., Fleuret, F.: Rethinking the role of gradient-based attribution methods for model interpretability. In: International Conference on Learning Representations, ICLR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.758, + 0.788, + 0.786 + ], + "angle": 0, + "content": "50. Srinivas, S., Fleuret, F.: Full-gradient representation for neural network visualization. In: Advances in Neural Information Processing Systems, NeurIPS (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.786, + 0.788, + 0.813 + ], + "angle": 0, + "content": "51. Sturmfels, P., Lundberg, S., Lee, S.I.: Visualizing the impact of feature attribution baselines. Distill (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.813, + 0.788, + 0.841 + ], + "angle": 0, + "content": "52. Sundararajan, M., Taly, A., Yan, Q.: Axiomatic attribution for deep networks. In: International Conference on Machine Learning, ICML (2017)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.364, + 0.129 + ], + "angle": 0, + "content": "P. Yang et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.175 + ], + "angle": 0, + "content": "53. Wah, C., Branson, S., Welinder, P., Perona, P., Belongie, S.: The caltech-ucsd birds-200-2011 dataset. California Institute of Technology (2011)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.175, + 0.785, + 0.217 + ], + "angle": 0, + "content": "54. Yang, P., Akhtar, N., Jiang, J., Mian, A.: Backdoor-based explainable AI benchmark for high fidelity evaluation of attribution methods. arXiv preprint arXiv:2405.02344 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.218, + 0.785, + 0.244 + ], + "angle": 0, + "content": "55. Yang, P., Akhtar, N., Wen, Z., Mian, A.: Local path integration for attribution. In: AAAI Conference on Artificial Intelligence, AAAI (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.245, + 0.785, + 0.285 + ], + "angle": 0, + "content": "56. Yang, P., Akhtar, N., Wen, Z., Shah, M., Mian, A.S.: Re-calibrating feature attributions for model interpretation. In: International Conference on Learning Representations (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.287, + 0.785, + 0.327 + ], + "angle": 0, + "content": "57. Yang, P., Wen, Z., Mian, A.: Multi-grained interpretable network for image recognition. In: 2022 26th International Conference on Pattern Recognition (ICPR). IEEE (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.328, + 0.785, + 0.355 + ], + "angle": 0, + "content": "58. Zeiler, M.D., Fergus, R.: Visualizing and understanding convolutional networks. In: European Conference on Computer Vision, ECCV (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.356, + 0.785, + 0.383 + ], + "angle": 0, + "content": "59. Zhang, X., Wang, N., Shen, H., Ji, S., Luo, X., Wang, T.: Interpretable deep learning under fire. In: USENIX Security (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.384, + 0.785, + 0.411 + ], + "angle": 0, + "content": "60. Zhou, B., Lapedriza, A., Khosla, A., Oliva, A., Torralba, A.: Places: A 10 million image database for scene recognition. TPAMI 40(6), 1452-1464 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.411, + 0.785, + 0.453 + ], + "angle": 0, + "content": "61. Zintgraf, L.M., Cohen, T.S., Adel, T., Welling, M.: Visualizing deep neural network decisions: Prediction difference analysis. In: International Conference on Learning Representations, ICLR (2017)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.453 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/Regulating Model Reliance on Non-Robust Features by Smoothing Input Marginal Density/1125a5c6-c219-425e-8e55-ef1e3d508033_origin.pdf b/2024/Regulating Model Reliance on Non-Robust Features by Smoothing Input Marginal Density/1125a5c6-c219-425e-8e55-ef1e3d508033_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..4bdc5612ca81f3167d535fad025298ac3467c096 --- /dev/null +++ b/2024/Regulating Model Reliance on Non-Robust Features by Smoothing Input Marginal Density/1125a5c6-c219-425e-8e55-ef1e3d508033_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:da0aa598d274ae563e84ac668b9abf93fa304e1383308283b510f693bbbeef67 +size 1855151 diff --git a/2024/Regulating Model Reliance on Non-Robust Features by Smoothing Input Marginal Density/full.md b/2024/Regulating Model Reliance on Non-Robust Features by Smoothing Input Marginal Density/full.md new file mode 100644 index 0000000000000000000000000000000000000000..a8379886c60d82d0c8052f586521861a85028806 --- /dev/null +++ b/2024/Regulating Model Reliance on Non-Robust Features by Smoothing Input Marginal Density/full.md @@ -0,0 +1,341 @@ +# Regulating Model Reliance on Non-Robust Features by Smoothing Input Marginal Density + +Peiyu Yang1, Naveed Akhtar2, Mubarak Shah3, and Ajmal Mian1 + +1 The University of Western Australia, Perth, Australia + +2 The University of Melbourne, Melbourne, Australia + +3 University of Central Florida, Orlando, America + +peiyu.yang@research.uwa.edu.au, naveed.akhtar1@unimelb.edu.au, shah@crcv.ucf.edu, ajmal.mian@uwa.edu.au + +Abstract. Trustworthy machine learning necessitates meticulous regulation of model reliance on non-robust features. We propose a framework to delineate and regulate such features by attributing model predictions to the input. Within our approach, robust feature attributions exhibit a certain consistency, while non-robust feature attributions are susceptible to fluctuations. This behavior allows identification of correlation between model reliance on non-robust features and smoothness of marginal density of the input samples. Hence, we uniquely regularize the gradients of the marginal density w.r.t. the input features for robustness. We also devise an efficient implementation of our regularization4 to address the potential numerical instability of the underlying optimization process. Moreover, we analytically reveal that, as opposed to our marginal density smoothing, the prevalent input gradient regularization smoothens conditional or joint density of the input, which can cause limited robustness. Our experiments validate the effectiveness of the proposed method, providing clear evidence of its capability to address the feature leakage problem and mitigate spurious correlations. Extensive results further establish that our technique enables the model to exhibit robustness against perturbations in pixel values, input gradients, and density. + +Keywords: Robust features $\cdot$ Regularization $\cdot$ Feature attributions + +# 1 Introduction + +Research on mitigating model reliance on non-robust input features has recently gained increasing attention due to high-stake machine learning applications [12, 18,38,49]. In this paper, we advance this direction by introducing a regularization technique that promotes a smooth marginal probability density function of the input to regulate the model's reliance on non-robust features. + +To distinguish between robust and non-robust features, we leverage the notion of attributions [16, 52, 58]. For a model $f$ parameterized by $\theta$ , attributions characterize the importance of the $i$ -th feature $x_{i}$ of the input $x$ for the model + +![](images/9cca57a90c893a17217a53e9d82aeb00daa20dac0f98771c14b116809af6dcb9.jpg) +Fig. 1: Attribution maps [45] and insertion game scores [34] for samples from (a) BlockMNIST and (b) CelebA-Hair datasets. As compared to input gradient regularization, our regularization leads to lower feature leakage while also achieving higher AUC for the insertion game. + +prediction by quantifying the output change between $f(x; \theta)$ and $f(x_{[x_i = 0]}; \theta)$ . Since robust input features contribute to model predictions equally well across slight condition variations, their attributions exhibit a certain consistency. On the other hand, non-robust feature attributions fluctuate under such variations. This potentially identifies a correlation between the model's reliance on non-robust features and the smoothness of the marginal probability density function of the input samples $p_\theta(x)$ . This is an important insight for our contribution. For robustness, it offers a possibility of model regularization using the gradients of the marginal density with respect to the input $\nabla_x p_\theta(x)$ . We propose to regularize $\nabla_x p_\theta(x)$ to encourage the model to prioritize the use of robust features and regulate its reliance on non-robust features. However, this can also lead to numerical instability during model optimization. To address that, we further introduce a stable and efficient implementation to estimate the gradient of marginal density. + +We also investigate input gradient norm regularization [13, 36, 37] and reveal that input gradients can be interpreted as input gradients of the log-conditional density $\nabla_{x}\log p_{\theta}(x|y)$ or log-joint density $\nabla_{x}\log p_{\theta}(x,y)$ . Input gradient regularization mitigates the model's reliance on non-robust features specific to the class label $y = i$ , leading to unintentional blindness to class-specific non-robust features where $y \neq i$ . In contrast, our regularization encourages smoothness of the marginal density $p_{\theta}(x)$ without imposing unintended constraints, providing a comprehensive regularization of the non-robust features. In Fig. 1, we employ attribution maps [45] and insertion game scores [34] to compare the robustness of vanilla models, input gradient regularized models and models trained with our regularization on BlockMNIST [44] and CelebA-Hair [30] datasets. As the representative examples show, our method suppresses both feature leakage and feature spurious correlation to improve model robustness and interpretability. + +The effectiveness of our approach is extensively established through a series of experiments. First, using BlockMNIST dataset [44], we demonstrate that the model trained with our regularization considerably mitigates the problem of feature leakage. This problem occurs when a model wrongly attributes importance to irrelevant but persistent features in the data to achieve better accuracy, e.g., the null block (☑) in Fig. 1. Addressing feature leakage helps suppress + +spurious correlation between the input features and model predictions [2, 39]. Moreover, we establish the robustness of the models trained with our regularization against perturbations from adversarial attacks [17, 31], input pixels [41] and input density, demonstrating the broad applicability of our approach. Our main contributions are summarized as follows. + +1. We identify robust and non-robust features by leveraging attributions, and establish the correlation between model reliance on non-robust features and the smoothness of data marginal density. +2. We propose an efficient technique for calculating the gradients of log density, also addressing the numerical instability of the underlying optimization. +3. Through extensive experiments, we demonstrate the effectiveness of the proposed regularization and additionally establish that our approach exhibits general robustness against perturbations. + +# 2 Related Work + +Regularization for Interpretability Robustness: Despite advancements in model transparency [26, 27, 40, 57], current deep neural networks still lack interpretability in their decision-making process, which is exacerbated by their reliance on non-robust input features. Prior studies, such as [1, 2, 44], identify that standard models are prone to relying on irrelevant or spuriously correlated features. To address that, several regularization techniques are proposed to improve model interpretability. In [14, 42], the authors incorporated prior knowledge into the model training process to regularize the model behavior. Dombrowski et al. [11, 12] found that regularizing the input Hessian using SoftPlus activations or weight decay can boost resilience against manipulated inputs. In [18], a joint energy-based model is trained as a discriminative model for improved robustness. Srinivas and Fleuret [49] enhanced the model interpretability by improving the alignment between the implicit density and the data distribution. + +Regularization for Adversarial Robustness: In addition to the other sources of prediction unreliability, adversarial attacks can manipulate model outputs with imperceptible perturbations to inputs [17, 31]. To address this, adversarial training through data augmentation with adversarial samples is widely employed [9, 31, 43]. Certified adversarial robustness through regularizations [4, 36, 46] is another branch of methods to defend against adversarial perturbations. Inspired by the classic double backpropagation [13], Ross and Doshi-Velez [36] regularized the input gradient norm for adversarial robustness. Etmann [15] further explored different variants of double backpropagation regularizations for various real-world scenarios. Moosavi-Dezfooli et al. [32] also proposed regularization to encourage a low curvature for adversarial robustness. To improve adversarial robustness, Chen et al. [10] computed the norm of attributions integrated from clean samples to adversarial samples as the regularization term. Ilyas et al. [24] proposed a disentangling method to distinguish feature robustness for explaining adversarial examples. However, they do not focus on general feature robustness. + +From their perspective, non-robust features are highly predictive; yet adversarially brittle. In contrast, our work focuses on general robustness derived directly from natural images. + +Attribution Methods: Feature attribution methods are used to estimate the importance of input features for a model's prediction and can be categorized as either perturbation-based or back-propagation-based techniques. Perturbation-based methods [16,58,61] calculate the attribution scores by repeatedly perturbing the input features and analyzing the resulting effects on the model prediction. These methods are also extended for evaluating the reliability of the computed attributions [3,34,41,55]. The back-propagation-based techniques [47,52,54,56] estimate the attribution scores by computing the gradients or integrated gradients with respect to the input features in the backward propagation process. In contrast to perturbation-based techniques, back-propagation-based attribution methods offer notable advantages in terms of computational efficiency and reliability. Given the transparency of attributions, our work leverages the attribution framework to distinguish between robust and non-robust features, which enables us to systematically analyze the robustness of input features. + +# 3 Feature Robustness by Attributions + +We first provide a framework for distinguishing between robust and non-robust features by analyzing their attributions. Herein, attribution inconsistencies among the features with distinct degrees of robustness identify a correlation between the model's reliance on non-robust features and the smoothness of output logits. + +Let us consider an input sample $x \in \mathbb{R}^n$ with label $y \in \mathbb{R}^c$ from a dataset $\mathcal{D}$ , and a classifier $f: \mathbb{R}^n \to \mathbb{R}^c$ parameterized by $\theta$ . We denote robust and non-robust features within the input $x$ as $x_{rob}, x_{nrob} \subseteq x$ . Consider an attribution method $\phi: \mathbb{R}^c \to \mathbb{R}^n$ attributing model predictions to input features by estimating their importance, resulting in an attribution map $M = \phi(f(x; \theta))$ . Inspired by the success of attributions in model explanation, we identify robust and non-robust features by leveraging their attributions. + +Without loss of generality, we assume that attributions $M$ of the features can be estimated by calculating the change in output logits when these features are removed from the input, following perturbation-based methods [35,58]: $M_{x_{rob}} = f(x;\theta) - f(x_{[x_{rob} = 0]};\theta)$ and $M_{x_{nrob}} = f(x;\theta) - f(x_{[x_{nrob} = 0]};\theta)$ . For ease of understanding, we alternatively use $f(x_{nrob};\theta)$ and $f(x_{rob};\theta)$ to represent attributions $M_{rob}$ and $M_{nrob}$ in the text to follow. We define robust features within the attribution framework as follows. + +Definition 1. A feature $x_{feat}$ shared among different input instances under its domain $\Delta_{x_{feat}}$ is robust if, for a randomly chosen class $y = i$ , its attribution $M_{x_{feat}}$ is bounded by a small constant $h$ under a metric $c(\cdot)$ , i.e., $c(f(x_{feat};\theta) - f(\tilde{x}_{feat};\theta)) \leq h : \tilde{x}_{feat} \in \Delta_{x_{feat}}$ . + +Under Definition 1, robust features are expected to contribute consistently to the model's prediction across different input samples. Non-robust features, on + +the other hand, are those that contribute to the prediction score inconsistently or only under specific conditions. Our focus here is on distinguishing between robust and non-robust features, without requiring to specify a particular metric. Definition 1 emphasizes attribution consistency for robust features rather than attribution positivity, thereby allowing for robust features that can also make a negative contribution to the model's prediction. Building further upon the above definition, we make the following remark. + +Remark 1. Robust features are largely condition-invariant in that they retain similar attributions despite slight changes to the input. In contrast, non-robust features are condition-specific in that their attributions either vary drastically with slightly varying input conditions, or behave robustly only under specific conditions. + +Robust features exhibit stable behavior across the input space, which is observable through consistent output logits $f(x_{rob};\theta)$ in classifiers regardless of the input instance or class $y$ . In contrast, non-robust features rely on specific conditions to exhibit a particular behavior, which is tailored to a specific class $y = i$ or the input instance. Robust modeling aims for a stronger reliance on robust features for prediction. Due to the consistency of output logits $f(x;\theta)$ for robust features, a smooth $f(x;\theta)$ is a desirable property for model robustness. + +# 4 Smoothing Marginal Density of Input + +Here, we establish the relation between model robustness and gradients of the input marginal density. Then, a robust regularization is proposed for regulating model reliance on non-robust features by smoothing marginal density. + +We commence our analytical analysis with probability density, following Bridle [7]. Given a class $y = i$ , a joint probability density function over the input with the output logit $f_{i}(x;\theta)$ is defined as + +$$ +p _ {\theta} (x, y = i) = e ^ {f _ {i} (x; \theta)} / Z _ {\theta}, \tag {1} +$$ + +where the constant $Z_{\theta} = \int e^{f_i(x;\theta)}dx$ is the partition function. $Z_{\theta}$ normalizes the input $x$ to a probability density by integrating over all possible input points $x$ in the input space via the model $f$ . By applying Bayes' rule, we eliminate the condition $y = i$ , resulting in the marginal density being defined solely on the input $x$ : $p_{\theta}(x) = p_{\theta}(y = i,x) / p_{\theta}(y = i|x)$ . The conditional density function $p_{\theta}(y = i|x)$ can be further defined as + +$$ +p _ {\theta} (y = i | x) = e ^ {f _ {i} (x; \theta)} / Z _ {f (x; \theta)}, \tag {2} +$$ + +where $Z_{f(x;\theta)} = \sum_{i=1}^{C} e^{f_i(x;\theta)}$ is the partition function for the output logits $f_i(x;\theta)$ defined on all the $C$ classes. To simplify the notation, we use $Z_{f(x)}$ to represent $Z_{f(x;\theta)}$ in the subsequent discussion. Exploiting the symmetry property of the joint density defined in Eq. 1, i.e., $p_\theta(x,y=i) = p_\theta(y=i,x)$ , the + +marginal density $p_{\theta}(x)$ can be expressed as + +$$ +p _ {\theta} (x) = \frac {e ^ {f _ {i} (x ; \theta)} / Z _ {\theta}}{e ^ {f _ {i} (x ; \theta)} / Z _ {f (x)}} = \frac {Z _ {f (x)}}{Z _ {\theta}}. \tag {3} +$$ + +As identified in the previous section, smoothness of output logits across inputs encourages the use of robust features by the model. Hence, we consider the marginal density $p_{\theta}(x)$ defined on the output logits $f(x; \theta)$ across the input space. Promoting a smaller gradient of the marginal density with respect to the input $x$ , denoted as $\nabla_x p_{\theta}(x)$ , contributes to the smoothness of the output logits. Thus, a positive correlation can be established between the use of robust features and the smoothness of the probability marginal density $p_{\theta}(x)$ . In particular, the smooth output logits of robust features across input samples suggest that these features will have relatively small gradients of the density $p_{\theta}(x)$ with respect to the input values. On the other hand, non-robust features with fluctuating output logits will have large gradients of the density that need to be suppressed during the training process for model robustness. Therefore, we can conclude with the following remark. + +Remark 2. Model reliance on non-robust features $x_{nrob}$ can be regulated by regularizing the gradients of the marginal density $p_{\theta}(x)$ w.r.t. $x$ , and this regularization can be achieved through optimizing the model parameters. + +In the light of Remark 2., we propose a regularization term for minimizing the gradients of marginal density. However, computing the gradients of the marginal density $\nabla_{x}p_{\theta}(x)$ is not feasible because the partition function defined on the entire input space is intractable. To avoid the estimation of the intractable $Z_{\theta}$ , we instead compute the gradients with respect to the log density as $\nabla_{x}\log p_{\theta}(x) = \nabla_{x}\log Z_{f(x)}$ . This is possible because $Z_{\theta}$ solely depends on the model parameter $\theta$ and not the input $x$ . Expanding the partition function $Z_{f(x)}$ in $\nabla_{x}\log p_{\theta}(x)$ , we obtain $\nabla_{x}\log p_{\theta}(x) = \sum_{i=1}^{C}\nabla_{x}e^{f_{i}(x;\theta)} / \sum_{i=1}^{C}e^{f_{i}(x;\theta)}$ . The $p$ -norm of this gradient is computed as the regularization term. In the optimization process, the goal is to find the optimal parameter $\theta^{*}$ , cf. Remark 2., by minimizing the loss $\ell$ as + +$$ +\theta^ {*} = \min _ {\theta} \ell (f (x; \theta), y) + \lambda | | \frac {\sum_ {i = 1} ^ {C} \nabla_ {x} e ^ {f _ {i} (x ; \theta)}}{\sum_ {i = 1} ^ {C} e ^ {f _ {i} (x ; \theta)}} | | _ {p}, \tag {4} +$$ + +where $\lambda$ indicates the magnitude of the coefficient for controlling the strength of the regularization. + +To regulate model reliance on non-robust features, our regularization encourages the smoothness of marginal density by regularizing its gradients. Since the output logit change in the log-marginal density $\log p_{\theta}(x) = \log Z_{f(x)} - \log Z_{\theta}$ , and $Z_{\theta}$ is independent of the input $x$ , we can only focus on the first term $Z_{f(x)} = \sum_{i=1}^{C} e^{f_i(x;\theta)}$ . Recall, Definition 1 of robust features. Assuming the robust input feature $x_{rob}$ exists in a random input $x$ , the corresponding output logit $f_i(x_{rob})$ will consistently attribute to the model prediction. This property + +of $x_{rob}$ leads to the smoothness of output change $\sum_{i=1}^{C} e^{f_i(x_{rob};\theta)}$ across different input samples and class labels. In contrast, non-robust input features $x_{nrob}$ show relatively high attributions for the output logit $f_i(x_{nrob})$ for a given class $i$ . However, they cannot maintain consistency in the attributions across different inputs or labels, leading to fluctuations in the output change $\sum_{i=1}^{C} e^{f_i(x_{nrob};\theta)}$ . In Fig. 1(a) and Fig. 13 in Supp. 17, non-robust features in the Null block exhibit fluctuating attributions across different samples on the standard trained model. It is demonstrated that the magnitude of gradients for input features in the marginal density $p_\theta(x)$ reflects the model's sensitivity to those features. We leverage this relation to mitigate model reliance on the non-robust features by smoothing the marginal density of the input samples. + +# 5 Stable and Efficient Implementation for Regularization + +From the implementation perspective, the gradient computation of marginal density involves multiple exponential operations in both the numerator and denominator of Eq. 4 which can introduce numerical instability in the optimization process, leading to gradient vanishing and explosion problems. Such issues can potentially hinder the application of our regularization to large non-linear models or wide-distribution data. For instance, batch normalization (BN) layers [25] solving internal covariate shifts with learnable scaling and shifting parameters can amplify the errors caused by the exponential operations during the backpropagation. In Fig. 2(a), yellow and red curves indicate original and inverse implementations for minimizing the gradient of log density $\nabla_{x}\log Z_{f(x)}$ . It is illustrated that the Frobenius gradient norm of both implementations undergoes rapid numerical overflow as the number of training iterations increases, revealing the serious issue of exploding and vanishing gradients using the proposed regularization. Therefore, it is crucial to address these through careful implementation to ensure the feasibility of our regularization. + +To address this challenge, we transform the computation from the summation of exponential operations to softmax. By subtracting a constant value $\eta$ from the logits before exponentiation, softmax can prevent numerical overflow during the exponential operations, i.e., $e^{f_i(x;\theta)} / \sum_j e^{f_j(x;\theta)} = e^{f_i(x;\theta) - \eta} / \sum_j e^{f_j(x;\theta) - \eta}$ . Thus, we incorporate the softmax function into the computation of log density + +![](images/4883dab3090e5047ad2a3d730cd81760d88a43f83de7a679acb12b363c6048f8.jpg) +(a) Numerical Stability. + +![](images/df2bee27ab0b88dec8a8b67ecd54d5d671cc737f30fec82477cb9670f6b7db52.jpg) +(b) Training Efficiency. + +![](images/421827394b8016dee901706db3a42f0c89dd6b46b3628ba42c3a297682adb77e.jpg) +(c) Training Time. +Fig. 2: The comparison of numerical stability, training efficiency and training time across different ResNet-34 models. + +gradient $\nabla_{x}logz_{f(x)}$ as + +$$ +\begin{array}{l} \nabla_ {x} l o g z _ {f (x)} = \nabla_ {x} f _ {i} (x; \theta) - \nabla_ {x} (l o g e ^ {f _ {i} (x; \theta)} - l o g z _ {f (x)}), \\ = \nabla_ {x} f _ {i} (x; \theta) - \nabla_ {x} \left(\log \left(f _ {i} (x; \theta) / z _ {f (x)}\right)\right), \\ \end{array} +$$ + +(5) + +where $f_{i}(x)$ indicates a logit of a random $i$ -th class, and $z_{f(x)}$ equals $\sum_{j} e^{f_{j}(x; \theta)}$ . Thus, the gradients with respect to the log-marginal density $p_{\theta}(x)$ can be replaced by computing the difference between the gradient $\nabla_{x} f_{i}(x; \theta)$ and the gradient of a log-softmax output $\nabla_{x} \log(e^{f_{i}(x; \theta)} / Z_{f(x)})$ . + +Our technique improves upon the common approach [5, 6] to achieve numerical stability in log exponential sum calculations, which typically employs the formula $\log(\sum_{i \in \{1, \ldots, n\}} e^{x_i}) = \eta + \log(\sum_{i \in \{1, \ldots, n\}} e^{x_i - \eta})$ , with $\eta$ being the maximum value of inputs $\{x_1, \ldots, x_n\}$ . We address the numerical instability by employing softmax, avoiding computationally expensive comparisons of the maximum gradient values. Specifically, the basic stable implementation involves finding the maximum gradient within the gradients of various classes, leading to a time complexity of $O(n)$ , where $n$ represents the number of classes. In contrast, our method eliminates numerical instabilities associated with a randomly selected class, eliminating the need for maximum value comparisons and leading to a more favorable time complexity of $O(1)$ . + +While our solution helps avoid numerical instability, it requires twice the gradient computations as compared to the sole calculation of density gradients. Hence, we propose an efficient mechanism for estimating the difference in the gradients. Specifically, we compute the gradient of the difference between two outputs to approximate the difference between the two gradients of outputs using Taylor series as + +$$ +\nabla_ {x} f _ {i} (x; \theta) - \nabla_ {x} \log \left(e ^ {f _ {i} (x; \theta)} / Z _ {f (x)}\right) \approx \nabla_ {x} \left(f _ {i} (x; \theta) - \log \left(e ^ {f _ {i} (x; \theta)} / Z _ {f (x)}\right)\right). \tag {6} +$$ + +The proof of Eq. 6 is provided in Supp. 9. As such, the proposed approach enables stable and efficient model optimization. The blue curve in Fig. 2(a) shows that our efficient implementation can effectively avoid numerical instability in the gradient computation. Fig. 2(b) provides a comparison of training efficiency on CIFAR-10 [28] between the basic numerically stable implementation and our two derived alternatives. More results regarding the comparison of numerical stability are provided in Supp. 15. In Fig. 2(c), we delve into a specific examination of the training times of models trained across the three datasets including BlockMNIST, CelebA and CIFAR-10. It is evident that our methods substantially improve the per-epoch training time. Moreover, our methods can yield further beneficial trade-offs as the number of classes increases. + +# 6 Limited Robustness in Input Gradient Regularization + +Input gradient regularization (InputGrad Reg.) [29,36] computes the Frobenius norm of input gradients $||\nabla_x f_i(x;\theta)||_F$ for a given class label $y = i$ , which is a + +baseline robust regularization for model optimization. Existing works [29,36,37] explain InputGrad Reg. as a prediction stability technique for robustness against perturbations. Here, we reveal that the input gradient norm potentially regularizes the gradients of implicit data density. Moreover, we provide an understanding of how InputGrad Reg. encourages robustness as well as its limitations. + +Suppose all the classes have equal probability, i.e., $p_{\theta}(y = i) = 1 / C$ . We can express the class-conditional density $p_{\theta}(x|y)$ by using Bayes's rule as + +$$ +p _ {\theta} (x \mid y = i) = \frac {p _ {\theta} (x , y = i)}{p _ {\theta} (y = i)} = \frac {e ^ {f _ {i} (x ; \theta)}}{Z _ {\theta} / C}. \tag {7} +$$ + +Now, we compute the gradients of the log density defined in Eq. 7 with respect to the input $x$ as + +$$ +\nabla_ {x} \log p _ {\theta} (x, y = i) = \nabla_ {x} \log p _ {\theta} (x | y = i) = \nabla_ {x} f _ {i} (x; \theta). \tag {8} +$$ + +Eq. 8 demonstrates that the input gradients can be interpreted as the gradients of either the log joint density or the log conditional density with respect to the input $x$ . This formulation highlights that InputGrad Reg. encourages consistent attributions of input features for the model's predictions. However, it is important to note that InputGrad Reg. is formulated under the specific condition $y = i$ , which limits its effectiveness in resolving inconsistencies when predicting a different class $y = j$ , where $j \neq i$ . Although input features consistently contribute to the model's prediction under a given class, InputGrad Reg. fails to consider inconsistent attributions of these features across different classes, thereby allowing model non-robust behavior to exist. Consequently, a model trained with input gradient regularization may still exhibit spurious robustness relying on specific conditions. Fig. 1(a) illustrates an example where the model trained with InputGrad Reg. is incapable of suppressing non-robust features in Null block. + +Remark 3. Input gradient regularization smoothens the joint and conditional density of the input $x$ under a specific label $y = i$ , compromising its ability to resist the class-specific non-robust features. + +In contrast to the existing works [36, 49] that highlight the reasons for Input-Grad Reg. efficacy, we reveal a weakness of this technique, cf. Remark 3.. Unlike regularizing gradients based on joint or conditional densities, our approach allows for more effective regularization without imposing the condition $y = i$ . Our method focuses on regularizing the gradients of marginal density $\nabla_{x}p_{\theta}(x)$ , thereby smoothing the output logits across the input samples. + +# 7 Experiments + +In this section, we perform extensive experiments to validate the efficacy of our regularization and the newly established correlation between the smooth + +marginal density and model reliance on non-robust features. Specifically, we present measurement results for addressing feature leakage and mitigating spurious correlations. In addition to specific applications, further results on the robustness against perturbations in input pixels, gradients and density are presented. Additional details about the datasets and the models used in our experiments can be found in Supp. 19. + +# 7.1 Efficacy against Feature Leakage and Adversarial Attacks + +In [2, 44], it is demonstrated that deep models end up assigning importance to irrelevant input features. Shah et al. [44] used BlockMNIST in their experiments, which is a synthetic dataset extended from MNIST [29]. To each MNIST sample, BlockMNIST attaches a null block (an irrelevant pattern) randomly at the top or bottom of the image, as shown in Fig. 3(a). Shah et al. [44] observed that the explaining tool InputGrad [47] attributes importance to both the informative number block and the uninformative null block in the standard trained model. This phenomenon is termed as feature leakage by the authors. Interestingly, the issue is mitigated in adversarially trained models. Since this dataset allows for a controlled robustness assessment, we first evaluate our method on BlockMNIST to analyze the model's reliance on irrelevant features. + +Reproducibility and Quantitative Measurement of Feature Leakage. + +Owing to the unreliability of InputGrad caused by model saturation [45], we employ Integrated Gradients (IG) [52], an axiomatic explanation tool, to faithfully reinvestigate the feature leakage phenomenon. In Fig. 3(b)-(c), we show that the attributions computed by IG can reproduce different leakage phenomena from the informative block region to the null block region on standard and adversarially trained models. Feature leakage is an important phenomenon in the context of model robustness. However, there is a lack of a quantitative metric in the current literature to quantify its extent. We use integrated gradients to define the metric $M_{leakage}$ to address this gap. Mathematically, + +![](images/5db74f99dcb8a59b9ca5a2787382f37ed2a0821c46b25f20de72903915974618.jpg) +(a) Input samples of BlockMNIST + +![](images/07d33db9873d2df2789915643336c04869b09dda61165952f19dc25b5460305e.jpg) +(b) Integrated Gradients on (c) Integrated Gradients on standard trained model adversarially trained model +Fig.3: BlockMNIST samples and feature leakage problem. (a) BlockMNIST randomly appends a null block at the top or bottom of MNIST samples. (b&c) Attribution maps are calculated by IG on the standard and adversarially trained models. + +$$ +M _ {\text {l e a k a g e}} = \mathbb {E} _ {x _ {n r o b} \sim D} | | x _ {n r o b} \times \int_ {\alpha = 0} ^ {1} \frac {\partial f (\alpha \cdot x _ {n r o b} ; \theta))}{\partial x _ {n r o b}} d \alpha | | _ {2}, \tag {9} +$$ + +where $\alpha$ is the step from the absence to the presence of the features, and $x_{nrob}$ specifies the non-robust features in the null block. Since the attributions of $x_{nrob}$ for the model's prediction should ideally be zero, we use the $L_{2}$ norm of attributions to quantify the extent of feature leakage. + +Robustness against Feature Leakage. Table 1 presents the experimental results on the BlockMNIST dataset. We compare our method with other robust regularizations and techniques including InputGrad [36], IG-SUM [10], SoftPlus activations [11] and Hessian [12]. InputGrad and Hessian regularize the first-order and second-order gradients w.r.t. the input. Models trained with SoftPlus activations and Hessian regularization fail to suppress the leakage problem, which indicates that feature leakage is not caused by the geometry of the model output manifold or high curvature [11, 59]. InputGrad regularization demonstrates robustness against both $L_{2}$ and $L_{\infty}$ adversarial attacks, yet it still fails to address the leakage problem. The result aligns well with Remark 3, which highlights the allowance of non-robust features across different classes in InputGrad regularization. In addition, the use of IG in IG-Norm regularization that accumulates input gradients as a regularization term results in behavior similar to InputGrad regularization, leading to compromised robustness. These results further reveal that adversarial robustness is not a sufficient condition for suppressing feature leakage. Our method demonstrates a considerable improvement over other techniques for feature leakage, while also maintaining superiority in adversarial robustness. In Tab. 1, we use $L_{2}$ norm for the compared regularization terms for fair benchmarking. We show in Supp. 10 that optimal norm exploration can yield an even more favorable trade-off for our technique for model robustness. + +Feature Leakage in Adversarily Robust Models. A FGSM adversarily trained model [17] augments the training samples by adversarial examples $x + \epsilon \cdot sign(\nabla_x f_i(x; \theta))$ . Notably minimizing the loss of the perturbed input $x + \epsilon \cdot sign(\nabla_x f_i(x; \theta))$ is similar to the InputGrad regularization. Thus, training with FGSM is still limited in its ability to suppress the leakage problem. In contrast, PGD attack [31] weakens the effect of the condition $y = i$ by iteratively searching for the perturbations from a random starting point. This process leads to a substantial enhancement in suppressing feature leakage, see Tab. 1. + +Table 1: Results on BlockMNIST. Feature leakage, standard accuracy and adversarial accuracy under $L_{2}$ and $L_{\infty}$ PGD-20 attacks are reported. ST & AT: Standard and Adversarial Training. + +
MethodFeature Leakage ↓PGD-20 (L2) ↑PGD-20 (L∞) ↑Accu. ↑
AT (FGSM)3.32487.480.0099.02
AT (PGD)2.31392.7528.0598.97
ST3.65773.570.0099.12
+ SoftPlus Acts.3.53367.950.0298.52
+ Hessian Reg.4.25880.060.0098.48
+ InputGrad Reg.3.46183.4621.1494.56
+ IG-SUM Reg.3.49782.1122.7392.87
+ Our Reg.2.25985.4129.3693.05
+ +![](images/7e1934dfae16c366802ad5283e745e806f06518ced504975b62d9a4042b22a52.jpg) +(a) Feature Leakage. + +![](images/165ac84da9f9c1381686ae165c5c1b522eb2606fba7946221a49bb74a8cefc11.jpg) +(b) $L_{\infty}$ Attack. +Fig. 4: Performance comparison between our method and InputGrad regularization under varying regularization coefficient for (a) Feature leakage, (b)-(c) Adversarial Accuracy under $L_{\infty}$ and $L_{2}$ PGD-20 attacks, and (d) Accuracy. + +![](images/b8cd40a4f5e1ade19ebe033e88077d01c8263368860ddbdbbb9cc01086363987.jpg) +(c) $L_{2}$ Attack. + +![](images/5a9d1441cb3f3bbf7e71b1dbb4d8e2057bd84eeffb98bd7115c45bdb03f51e22.jpg) +(d) Accuracy. + +Magnitude of Coefficient for Regularization. Figs. 4(a)-(d) present a comparison of results for feature leakage and adversarial accuracy under PGD attacks, as well as the standard accuracy across varying magnitudes for the regularization strength. The results affirm that our method effectively regulates feature leakage by imposing a penalty on non-robust features. Moreover, our regularization enables the model to defend against both $L_{2}$ and $L_{\infty}$ attacks while maintaining high accuracy, showing an outstanding trade-off across four criteria. More adversarial robustness comparisons on CIFAR dataset [28] are reported in Supp. 11. + +# 7.2 Efficacy for Spurious Correlation + +Recent research has highlighted the susceptibility of neural models in learning spurious correlations that enhance performance on a given data but fail to generalize [2,8,39]. For instance, in CelebA-Hair dataset [30], which commonly consists of samples containing female celebrities with blond hair and male celebrities with dark hair, models heavily rely on the spuriously correlated gender feature to predict the target hair color [19, 39]. Consequently, accuracy tends to be lower for the samples containing + +male celebrities with blond hair, see Fig. 5. + +To mitigate spurious correlations, distributional robust optimization (DRO) techniques have been proposed to re-weight the training loss of input samples from different groups [23,39]. In our regularization, an additional penalty is imposed to penalize the model's reliance on these spuriously correlated features because of their inconsistent attributions. This encourages the use of robust features while suppressing the model's reliance on spuriously correlated features. Reliable quantification of the model's robustness on natural images for spurious correlation suppression is an unresolved issue in the literature. We employ + +![](images/00aa977129fc7b52da8eb0c90bfa240334d443e45459bb1086ea76ef90c2ea3a.jpg) +Common training samples: +Label: Blond hair +Spur. Feat.: Female +Accuracy: $92.1\%$ + +![](images/449779662b8899761d692b6e86ee95d564b3e117dfd4d58035d352ff5eeba19d.jpg) +Label: Dark hair +Spur. Feat.: Male +Accuracy: $99.7\%$ + +![](images/37c14f47adcb222d4da384c9736006a1eeb053e0319fd902d3f904720f19592c.jpg) +Test sample: +Label: Blond hair +Spur.Feat.:Male +Accuracy: $23.6\%$ + +Table 2: Worst-group accuracy and overall accuracy comparisons between Vanilla ResNet-34, group DRO, Score-Matching, InputGrad and our regularization on CelebA-Hair and Waterbirds datasets. + +
MethodWorst-Group Accuracy (%)Overall Accuracy (%)
CelebA-HairWaterbirdsCelebA-HairWaterbirds
Vanilla Model49.90 ± 8.6962.90 ± 0.1094.90 ± 0.3987.70 ± 0.08
Group DRO59.44 ± 5.9863.60 ± 0.1794.96 ± 0.2187.60 ± 0.05
Score-Matching Reg.59.78 ± 7.5658.19 ± 1.5593.46 ± 0.7185.64 ± 0.38
InputGrad Reg.82.66 ± 3.6358.18 ± 1.2292.12 ± 2.1885.50 ± 0.27
Our Reg.85.62 ± 5.3663.78 ± 2.8392.30 ± 1.3886.48 ± 0.38
Group DRO + Ours82.98 ± 4.6973.82 ± 2.2793.62 ± 0.7490.52 ± 0.17
+ +attribute maps [52] and insertion game [34] to demonstrate the effectiveness of our regularization in suppressing the use of the spuriously correlated gender feature and promoting the use of the robust hair feature, as shown in Fig. 1(b). This superiority leads to performance improvements on worst-case samples in the model trained with our regularization. + +Table 2 presents comparison of accuracy on worst-group samples and overall samples from the CelebA-Hair and Waterbirds datasets [30,39]. The Waterbirds dataset consists of synthetic bird images from CUB-200-2011 [53] and Places [60] datasets, incorporating spurious background features, such as land and water scenes, to confuse true labels of bird categories. We compare the proposed regularization method with InputGrad and Score-Matching regularizations [49], as well as Group DRO [39]. Score-Matching regularization is proposed to enhance the interpretability of the model by improving the alignment of implicit density models. Experimental results clearly demonstrate the effectiveness of our method in enhancing worst-group accuracy while maintaining overall sample accuracy. Furthermore, the performance gains can be further enhanced by incorporating our regularization technique into Group DRO. This enhancement highlights the efficacy of our regularization in terms of its practicality. More results using attribution maps and the insertion game on both CelebA-Hair and Waterbirds datasets are provided in Supp. 17. Moreover, out-of-distribution detection [21] is also performed on CIFAR [28] and SVHN [33] in Supp. 14. + +# 7.3 Efficacy against Pixels, Gradients and Density Perturbations + +We first employ pixel perturbation [41, 55] to quantitatively compare the robustness of different models following Srinivas and Fleuret [49] who iteratively removed the most important input pixels identified by attribution maps for model robustness evaluation. Robust models are expected to exhibit increased sensitivity when removing the most important pixels and decreased sensitivity when removing the least important ones. We assess the difference in fractional output logit change between the images with the top and bottom $k\%$ most salient pixels using SmoothGrad [48] on ResNet-18 [20] trained on CIFAR-10 + +![](images/4a556c4dc6a686a71038b588c4cf2cb3128f7e07eb4dc4fe0adb53238a35c9db.jpg) +(a) Pixel Perturbation on CIFAR-10 and CIFAR-100. + +![](images/d063c23504d6e04835e0415096766816b2189cd8b350fa9825a055ef65593ab6.jpg) +Fig. 6: Robustness comparison. (a) Pixel perturbation experimental results on ResNet-18 trained on CIFAR-10 and CIFAR-100. Higher curves indicate better results. (b) Robustness of relative gradients and absolute density on CIFAR-10. Lower curves indicate better results. + +![](images/57ccbb079a2189b48c2834c2a0f237da9f841be08faf3239d05910a3d8139e79.jpg) +(b) Robustness of Gradients and Density. + +![](images/1c87bd4c79667abbf7b60f57532a6bfabf57e32c68bd9dd48ef483fd5d4367ae.jpg) + +and CIFAR-100 [28], as depicted in Fig. 6(a). Given the recognized challenge of reference image ambiguity in IG for explaining natural images [14, 51], we opt to use SmoothGrad for explaining predictions of images in the CIFAR dataset. SmoothGrad has demonstrated to outperform IG in such conditions [22, 50]. Recognizing their respective strengths, we use different attribution methods in appropriate conditions for reliable results. It can be observed that models trained with our regularization significantly outperform different robust regularizations including Score-Matching, InputGrad and CURE regularizations [32]. + +We also test the robustness of the relative input gradients $||\nabla_{x}f(x + \delta) - \nabla_{x}f(x)||_{2} / ||\nabla_{x}f(x)||_{2}$ and the absolute density through $\sum_{i = 1}^{C}e^{f_i(x + \delta) - f_i(x)}$ on the input with Gaussian noise $\delta$ in increasing standard deviation on CIFAR-10 dataset, as shown in Fig. 6(b). We can observe that our regularization leads to competitive robustness for the relative gradient in comparison with regularizing the Hessian norm in CURE. Moreover, our regularization naturally leads to density robustness, which is associated with a strong generative ability of the models [49]. More tests on CIFAR-100 are provided in Supp. 13. Supp. 16 provides the visualizations calculated by activation optimization. These results affirm that our regularization improves both the discriminative and generative abilities of models. + +# 8 Conclusion + +In this paper, we define robust and non-robust features from a feature attribution perspective, and establish a correlation between the smoothness of input marginal density and model reliance on non-robust features. This connection motivates us to propose a regularization that targets the gradients of the marginal density, aiming to regulate the reliance on non-robust features. Extensive experiments demonstrate the effectiveness of our regularization in boosting model robustness across a wide range of applications. We emphasize that our approach does not advocate for the complete removal of model reliance on non-robust features, but instead seeks to achieve a balance between model performance and robustness through appropriate regularization strength. + +# Acknowledgments + +This research was supported by ARC Discovery Grant 190102443 and by Google Research under Google Research Scholar Program. Dr. Naveed Akhtar is a recipient of the Australian Research Council Discovery Early Career Researcher Award (project number DE230101058) funded by the Australian Government. Professor Ajmal Mian is the recipient of an Australian Research Council Future Fellowship Award (project number FT210100268) funded by the Australian Government. + +# References + +1. Adebayo, J., Muelly, M., Abelson, H., Kim, B.: Post hoc explanations may be ineffective for detecting unknown spurious correlation. In: International Conference on Learning Representations, ICLR (2022) +2. Adebayo, J., Muelly, M., Liccardi, I., Kim, B.: Debugging tests for model explanations. In: Advances in Neural Information Processing Systems, NeurIPS (2020) +3. Ancona, M., Ceolini, E., Öztireli, C., Gross, M.: Towards better understanding of gradient-based attribution methods for deep neural networks. In: International Conference on Learning Representations, ICLR (2018) +4. Anil, C., Lucas, J., Grosse, R.: Sorting out lipschitz function approximation. In: International Conference on Machine Learning, ICML (2019) +5. Blanchard, P., Higham, D.J., Higham, N.J.: Accurately computing the log-sum-exp and softmax functions. IMA Journal of Numerical Analysis 41(4), 2311-2330 (2021) +6. Bridle, J.: Training stochastic model recognition algorithms as networks can lead to maximum mutual information estimation of parameters. Advances in Neural Information Processing Systems, NeurIPS (1989) +7. Bridle, J.S.: Probabilistic interpretation of feedforward classification network outputs, with relationships to statistical pattern recognition. In: Neurocomputing: Algorithms, Architectures and Applications, pp. 227-236. Springer (1990) +8. Buolamwini, J., Gebru, T.: Gender shades: Intersectional accuracy disparities in commercial gender classification. In: ACM Conference on Fairness, Accountability, and Transparency, FAccT (2018) +9. Carlini, N., Wagner, D.: Towards evaluating the robustness of neural networks. In: IEEE Symposium on Security and Privacy, EuroS&P (2017) +10. Chen, J., Wu, X., Rastogi, V., Liang, Y., Jha, S.: Robust attribution regularization. In: Advances in Neural Information Processing Systems, NeurIPS (2019) +11. Dombrowski, A.K., Alber, M., Anders, C., Ackermann, M., Müller, K.R., Kessel, P.: Explanations can be manipulated and geometry is to blame. In: Advances in Neural Information Processing Systems, NeurIPS (2019) +12. Dombrowski, A.K., Anders, C.J., Müller, K.R., Kessel, P.: Towards robust explanations for deep neural networks. Pattern Recognition 121, 108194 (2022) +13. Drucker, H., LeCun, Y.: Improving generalization performance using double backpropagation. IEEE Trans. Neural Networks (1992) +14. Erion, G.G., Janizek, J.D., Sturmfels, P., Lundberg, S.M., Lee, S.: Improving performance of deep learning models with axiomatic attribution priors and expected gradients. Nature Machine Intelligence (2021) + +15. Etmann, C.: A closer look at double backpropagation. CoRR (2019) +16. Fong, R.C., Vedaldi, A.: Interpretable explanations of black boxes by meaningful perturbation. In: International Conference on Computer Vision, ICCV (2017) +17. Goodfellow, I.J., Shlens, J., Szegedy, C.: Explaining and harnessing adversarial examples. In: International Conference on Learning Representations, ICLR (2015) +18. Grathwohl, W., Wang, K.C., Jacobsen, J.H., Duvenaud, D., Norouzi, M., Swersky, K.: Your classifier is secretly an energy based model and you should treat it like one. In: International Conference on Learning Representations, ICLR (2020) +19. Hashimoto, T., Srivastava, M., Namkoong, H., Liang, P.: Fairness without demographics in repeated loss minimization. In: International Conference on Machine Learning, ICML (2018) +20. He, K., Zhang, X., Ren, S., Sun, J.: Identity mappings in deep residual networks. In: European Conference on Computer Vision, European Conference on Computer Vision, ECCV (2016) +21. Hendrycks, D., Gimpel, K.: A baseline for detecting misclassified and out-of-distribution examples in neural networks. CoRR (2016) +22. Hooker, S., Erhan, D., Kindermans, P., Kim, B.: A benchmark for interpretability methods in deep neural networks. In: Advances in Neural Information Processing Systems, NeurIPS (2019) +23. Hu, W., Niu, G., Sato, I., Sugiyama, M.: Does distributionally robust supervised learning give robust classifiers? In: International Conference on Machine Learning, ICML (2018) +24. Ilyas, A., Santurkar, S., Tsipras, D., Engstrom, L., Tran, B., Madry, A.: Adversarial examples are not bugs, they are features. Advances in neural information processing systems 32 (2019) +25. Ioffe, S., Szegedy, C.: Batch normalization: Accelerating deep network training by reducing internal covariate shift. In: International Conference on Machine Learning, ICML (2015) +26. Jiang, J., Wen, Z., Mansoor, A., Mian, A.: Efficient hyperparameter optimization with adaptive fidelity identification. In: IEEE Conference on Computer Vision and Pattern Recognition (2024) +27. Jiang, J., Wen, Z., Yang, P., Mansoor, A., Mian, A.: Fast-pgm: Fast probabilistic graphical model learning and inference. arXiv preprint arXiv:2405.15605 (2024) +28. Krizhevsky, A., Hinton, G., et al.: Learning multiple layers of features from tiny images. Technical Report, University of Toronto (2009) +29. LeCun, Y., Bottou, L., Bengio, Y., Haffner, P.: Gradient-based learning applied to document recognition. Proc. IEEE (1998) +30. Liu, Z., Luo, P., Wang, X., Tang, X.: Deep learning face attributes in the wild. In: International Conference on Computer Vision, ICCV (2015) +31. Madry, A., Makelov, A., Schmidt, L., Tsipras, D., Vladu, A.: Towards deep learning models resistant to adversarial attacks. In: International Conference on Learning Representations, ICLR (2018) +32. Moosavi-Dezfooli, S.M., Fawzi, A., Uesato, J., Frossard, P.: Robustness via curvature regularization, and vice versa. In: IEEE Conference on Computer Vision and Pattern Recognition, CVPR (2019) +33. Netzer, Y., Wang, T., Coates, A., Bissacco, A., Wu, B., Ng, A.Y.: Reading digits in natural images with unsupervised feature learning. Advances in Neural Information Processing Systems, NeurIPSW (2011) +34. Petsiuk, V., Das, A., Saenko, K.: RISE: randomized input sampling for explanation of black-box models. In: British Machine Vision Conference, BMVC (2018) + +35. Ribeiro, M.T., Singh, S., Guestrin, C.: " why should i trust you?" explaining the predictions of any classifier. In: ACM SIGKDD Conference on Knowledge Discovery & Data Mining, KDD (2016) +36. Ross, A.S., Doshi-Velez, F.: Improving the adversarial robustness and interpretability of deep neural networks by regularizing their input gradients. In: AAAI Conference on Artificial Intelligence, AAAI (2018) +37. Ross, A.S., Hughes, M.C., Doshi-Velez, F.: Right for the right reasons: Training differentiable models by constraining their explanations. In: International Joint Conference on Artificial Intelligence, IJCAI (2017) +38. Rudin, C.: Stop explaining black box machine learning models for high stakes decisions and use interpretable models instead. Nature Machine Intelligence (2019) +39. Sagawa, S., Koh, P.W., Hashimoto, T.B., Liang, P.: Distributionally robust neural networks for group shifts: On the importance of regularization for worst-case generalization. In: International Conference on Learning Representations, ICLR (2019) +40. Salahuddin, Z., Woodruff, H.C., Chatterjee, A., Lambin, P.: Transparency of deep neural networks for medical image analysis: A review of interpretability methods. Computers in biology and medicine 140, 105111 (2022) +41. Samek, W., Binder, A., Montavon, G., Lapuschkin, S., Müller, K.: Evaluating the visualization of what a deep neural network has learned. TNNLS 28(11), 2660-2673 (2017) +42. Schramowski, P., Stammer, W., Teso, S., Brugger, A., Herbert, F., Shao, X., Luigs, H., Mahlein, A., Kersting, K.: Making deep neural networks right for the right scientific reasons by interacting with their explanations. Nature Machine Intelligence (2020) +43. Shafahi, A., Najibi, M., Ghiasi, M.A., Xu, Z., Dickerson, J., Studer, C., Davis, L.S., Taylor, G., Goldstein, T.: Adversarial training for free! In: Advances in Neural Information Processing Systems, NeurIPS (2019) +44. Shah, H., Jain, P., Netrapalli, P.: Do input gradients highlight discriminative features? In: Advances in Neural Information Processing Systems, NeurIPS (2021) +45. Shrikumar, A., Greenside, P., Kundaje, A.: Learning important features through propagating activation differences. In: International Conference on Machine Learning, ICML (2017) +46. Simon-Gabriel, C.J., Ollivier, Y., Bottou, L., Scholkopf, B., Lopez-Paz, D.: First-order adversarial vulnerability of neural networks and input dimension. In: International Conference on Machine Learning, ICML (2019) +47. Simonyan, K., Vedaldi, A., Zisserman, A.: Deep inside convolutional networks: Visualising image classification models and saliency maps. In: International Conference on Learning Representations, ICLR (2014) +48. Smilkov, D., Thorat, N., Kim, B., Viégas, F., Wattenberg, M.: Smoothgrad: removing noise by adding noise. CoRR abs/1706.03825 (2017) +49. Srinivas, S., Fleuret, F.: Rethinking the role of gradient-based attribution methods for model interpretability. In: International Conference on Learning Representations, ICLR (2021) +50. Srinivas, S., Fleuret, F.: Full-gradient representation for neural network visualization. In: Advances in Neural Information Processing Systems, NeurIPS (2019) +51. Sturmfels, P., Lundberg, S., Lee, S.I.: Visualizing the impact of feature attribution baselines. Distill (2020) +52. Sundararajan, M., Taly, A., Yan, Q.: Axiomatic attribution for deep networks. In: International Conference on Machine Learning, ICML (2017) + +53. Wah, C., Branson, S., Welinder, P., Perona, P., Belongie, S.: The caltech-ucsd birds-200-2011 dataset. California Institute of Technology (2011) +54. Yang, P., Akhtar, N., Jiang, J., Mian, A.: Backdoor-based explainable AI benchmark for high fidelity evaluation of attribution methods. arXiv preprint arXiv:2405.02344 (2024) +55. Yang, P., Akhtar, N., Wen, Z., Mian, A.: Local path integration for attribution. In: AAAI Conference on Artificial Intelligence, AAAI (2023) +56. Yang, P., Akhtar, N., Wen, Z., Shah, M., Mian, A.S.: Re-calibrating feature attributions for model interpretation. In: International Conference on Learning Representations (2023) +57. Yang, P., Wen, Z., Mian, A.: Multi-grained interpretable network for image recognition. In: 2022 26th International Conference on Pattern Recognition (ICPR). IEEE (2022) +58. Zeiler, M.D., Fergus, R.: Visualizing and understanding convolutional networks. In: European Conference on Computer Vision, ECCV (2014) +59. Zhang, X., Wang, N., Shen, H., Ji, S., Luo, X., Wang, T.: Interpretable deep learning under fire. In: USENIX Security (2020) +60. Zhou, B., Lapedriza, A., Khosla, A., Oliva, A., Torralba, A.: Places: A 10 million image database for scene recognition. TPAMI 40(6), 1452-1464 (2017) +61. Zintgraf, L.M., Cohen, T.S., Adel, T., Welling, M.: Visualizing deep neural network decisions: Prediction difference analysis. In: International Conference on Learning Representations, ICLR (2017) \ No newline at end of file diff --git a/2024/Regulating Model Reliance on Non-Robust Features by Smoothing Input Marginal Density/images.zip b/2024/Regulating Model Reliance on Non-Robust Features by Smoothing Input Marginal Density/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..dab3cc7d1ad8702c1ba7a5a733adb083728c4503 --- /dev/null +++ b/2024/Regulating Model Reliance on Non-Robust Features by Smoothing Input Marginal Density/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:33ace8440536d9634c9666127b289d151a6a6ccc363a737fe25544dd1ac0a135 +size 369355 diff --git a/2024/Regulating Model Reliance on Non-Robust Features by Smoothing Input Marginal Density/layout.json b/2024/Regulating Model Reliance on Non-Robust Features by Smoothing Input Marginal Density/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..40acb56691641c5d969d6b3a3a8699c542f0ecf5 --- /dev/null +++ b/2024/Regulating Model Reliance on Non-Robust Features by Smoothing Input Marginal Density/layout.json @@ -0,0 +1,11503 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 141, + 111, + 473, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 111, + 473, + 148 + ], + "spans": [ + { + "bbox": [ + 141, + 111, + 473, + 148 + ], + "type": "text", + "content": "Regulating Model Reliance on Non-Robust Features by Smoothing Input Marginal Density" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 147, + 167, + 465, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 147, + 167, + 465, + 180 + ], + "spans": [ + { + "bbox": [ + 147, + 167, + 465, + 180 + ], + "type": "text", + "content": "Peiyu Yang1, Naveed Akhtar2, Mubarak Shah3, and Ajmal Mian1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 192, + 189, + 421, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 192, + 189, + 421, + 200 + ], + "spans": [ + { + "bbox": [ + 192, + 189, + 421, + 200 + ], + "type": "text", + "content": "1 The University of Western Australia, Perth, Australia" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 196, + 201, + 416, + 211 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 196, + 201, + 416, + 211 + ], + "spans": [ + { + "bbox": [ + 196, + 201, + 416, + 211 + ], + "type": "text", + "content": "2 The University of Melbourne, Melbourne, Australia" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 204, + 211, + 409, + 222 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 204, + 211, + 409, + 222 + ], + "spans": [ + { + "bbox": [ + 204, + 211, + 409, + 222 + ], + "type": "text", + "content": "3 University of Central Florida, Orlando, America" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 161, + 224, + 454, + 244 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 161, + 224, + 454, + 244 + ], + "spans": [ + { + "bbox": [ + 161, + 224, + 454, + 244 + ], + "type": "text", + "content": "peiyu.yang@research.uwa.edu.au, naveed.akhtar1@unimelb.edu.au, shah@crcv.ucf.edu, ajmal.mian@uwa.edu.au" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 159, + 274, + 455, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 274, + 455, + 483 + ], + "spans": [ + { + "bbox": [ + 159, + 274, + 455, + 483 + ], + "type": "text", + "content": "Abstract. Trustworthy machine learning necessitates meticulous regulation of model reliance on non-robust features. We propose a framework to delineate and regulate such features by attributing model predictions to the input. Within our approach, robust feature attributions exhibit a certain consistency, while non-robust feature attributions are susceptible to fluctuations. This behavior allows identification of correlation between model reliance on non-robust features and smoothness of marginal density of the input samples. Hence, we uniquely regularize the gradients of the marginal density w.r.t. the input features for robustness. We also devise an efficient implementation of our regularization4 to address the potential numerical instability of the underlying optimization process. Moreover, we analytically reveal that, as opposed to our marginal density smoothing, the prevalent input gradient regularization smoothens conditional or joint density of the input, which can cause limited robustness. Our experiments validate the effectiveness of the proposed method, providing clear evidence of its capability to address the feature leakage problem and mitigate spurious correlations. Extensive results further establish that our technique enables the model to exhibit robustness against perturbations in pixel values, input gradients, and density." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 160, + 494, + 438, + 505 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 494, + 438, + 505 + ], + "spans": [ + { + "bbox": [ + 160, + 494, + 438, + 505 + ], + "type": "text", + "content": "Keywords: Robust features " + }, + { + "bbox": [ + 160, + 494, + 438, + 505 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 494, + 438, + 505 + ], + "type": "text", + "content": " Regularization " + }, + { + "bbox": [ + 160, + 494, + 438, + 505 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 494, + 438, + 505 + ], + "type": "text", + "content": " Feature attributions" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 526, + 230, + 538 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 526, + 230, + 538 + ], + "spans": [ + { + "bbox": [ + 132, + 526, + 230, + 538 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 551, + 481, + 609 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 551, + 481, + 609 + ], + "spans": [ + { + "bbox": [ + 130, + 551, + 481, + 609 + ], + "type": "text", + "content": "Research on mitigating model reliance on non-robust input features has recently gained increasing attention due to high-stake machine learning applications [12, 18,38,49]. In this paper, we advance this direction by introducing a regularization technique that promotes a smooth marginal probability density function of the input to regulate the model's reliance on non-robust features." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 611, + 482, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 611, + 482, + 647 + ], + "spans": [ + { + "bbox": [ + 130, + 611, + 482, + 647 + ], + "type": "text", + "content": "To distinguish between robust and non-robust features, we leverage the notion of attributions [16, 52, 58]. For a model " + }, + { + "bbox": [ + 130, + 611, + 482, + 647 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 130, + 611, + 482, + 647 + ], + "type": "text", + "content": " parameterized by " + }, + { + "bbox": [ + 130, + 611, + 482, + 647 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 130, + 611, + 482, + 647 + ], + "type": "text", + "content": ", attributions characterize the importance of the " + }, + { + "bbox": [ + 130, + 611, + 482, + 647 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 611, + 482, + 647 + ], + "type": "text", + "content": "-th feature " + }, + { + "bbox": [ + 130, + 611, + 482, + 647 + ], + "type": "inline_equation", + "content": "x_{i}" + }, + { + "bbox": [ + 130, + 611, + 482, + 647 + ], + "type": "text", + "content": " of the input " + }, + { + "bbox": [ + 130, + 611, + 482, + 647 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 130, + 611, + 482, + 647 + ], + "type": "text", + "content": " for the model" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 653, + 448, + 666 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 653, + 448, + 666 + ], + "spans": [ + { + "bbox": [ + 133, + 653, + 448, + 666 + ], + "type": "text", + "content": "4 Our code is available at https://github.com/typeiyu/Input_density_reg." + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 114, + 482, + 221 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 482, + 221 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 482, + 221 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 482, + 221 + ], + "type": "image", + "image_path": "9cca57a90c893a17217a53e9d82aeb00daa20dac0f98771c14b116809af6dcb9.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 226, + 482, + 271 + ], + "lines": [ + { + "bbox": [ + 130, + 226, + 482, + 271 + ], + "spans": [ + { + "bbox": [ + 130, + 226, + 482, + 271 + ], + "type": "text", + "content": "Fig. 1: Attribution maps [45] and insertion game scores [34] for samples from (a) BlockMNIST and (b) CelebA-Hair datasets. As compared to input gradient regularization, our regularization leads to lower feature leakage while also achieving higher AUC for the insertion game." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 281, + 482, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 281, + 482, + 437 + ], + "spans": [ + { + "bbox": [ + 130, + 281, + 482, + 437 + ], + "type": "text", + "content": "prediction by quantifying the output change between " + }, + { + "bbox": [ + 130, + 281, + 482, + 437 + ], + "type": "inline_equation", + "content": "f(x; \\theta)" + }, + { + "bbox": [ + 130, + 281, + 482, + 437 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 281, + 482, + 437 + ], + "type": "inline_equation", + "content": "f(x_{[x_i = 0]}; \\theta)" + }, + { + "bbox": [ + 130, + 281, + 482, + 437 + ], + "type": "text", + "content": ". Since robust input features contribute to model predictions equally well across slight condition variations, their attributions exhibit a certain consistency. On the other hand, non-robust feature attributions fluctuate under such variations. This potentially identifies a correlation between the model's reliance on non-robust features and the smoothness of the marginal probability density function of the input samples " + }, + { + "bbox": [ + 130, + 281, + 482, + 437 + ], + "type": "inline_equation", + "content": "p_\\theta(x)" + }, + { + "bbox": [ + 130, + 281, + 482, + 437 + ], + "type": "text", + "content": ". This is an important insight for our contribution. For robustness, it offers a possibility of model regularization using the gradients of the marginal density with respect to the input " + }, + { + "bbox": [ + 130, + 281, + 482, + 437 + ], + "type": "inline_equation", + "content": "\\nabla_x p_\\theta(x)" + }, + { + "bbox": [ + 130, + 281, + 482, + 437 + ], + "type": "text", + "content": ". We propose to regularize " + }, + { + "bbox": [ + 130, + 281, + 482, + 437 + ], + "type": "inline_equation", + "content": "\\nabla_x p_\\theta(x)" + }, + { + "bbox": [ + 130, + 281, + 482, + 437 + ], + "type": "text", + "content": " to encourage the model to prioritize the use of robust features and regulate its reliance on non-robust features. However, this can also lead to numerical instability during model optimization. To address that, we further introduce a stable and efficient implementation to estimate the gradient of marginal density." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 437, + 482, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 437, + 482, + 593 + ], + "spans": [ + { + "bbox": [ + 130, + 437, + 482, + 593 + ], + "type": "text", + "content": "We also investigate input gradient norm regularization [13, 36, 37] and reveal that input gradients can be interpreted as input gradients of the log-conditional density " + }, + { + "bbox": [ + 130, + 437, + 482, + 593 + ], + "type": "inline_equation", + "content": "\\nabla_{x}\\log p_{\\theta}(x|y)" + }, + { + "bbox": [ + 130, + 437, + 482, + 593 + ], + "type": "text", + "content": " or log-joint density " + }, + { + "bbox": [ + 130, + 437, + 482, + 593 + ], + "type": "inline_equation", + "content": "\\nabla_{x}\\log p_{\\theta}(x,y)" + }, + { + "bbox": [ + 130, + 437, + 482, + 593 + ], + "type": "text", + "content": ". Input gradient regularization mitigates the model's reliance on non-robust features specific to the class label " + }, + { + "bbox": [ + 130, + 437, + 482, + 593 + ], + "type": "inline_equation", + "content": "y = i" + }, + { + "bbox": [ + 130, + 437, + 482, + 593 + ], + "type": "text", + "content": ", leading to unintentional blindness to class-specific non-robust features where " + }, + { + "bbox": [ + 130, + 437, + 482, + 593 + ], + "type": "inline_equation", + "content": "y \\neq i" + }, + { + "bbox": [ + 130, + 437, + 482, + 593 + ], + "type": "text", + "content": ". In contrast, our regularization encourages smoothness of the marginal density " + }, + { + "bbox": [ + 130, + 437, + 482, + 593 + ], + "type": "inline_equation", + "content": "p_{\\theta}(x)" + }, + { + "bbox": [ + 130, + 437, + 482, + 593 + ], + "type": "text", + "content": " without imposing unintended constraints, providing a comprehensive regularization of the non-robust features. In Fig. 1, we employ attribution maps [45] and insertion game scores [34] to compare the robustness of vanilla models, input gradient regularized models and models trained with our regularization on BlockMNIST [44] and CelebA-Hair [30] datasets. As the representative examples show, our method suppresses both feature leakage and feature spurious correlation to improve model robustness and interpretability." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": "The effectiveness of our approach is extensively established through a series of experiments. First, using BlockMNIST dataset [44], we demonstrate that the model trained with our regularization considerably mitigates the problem of feature leakage. This problem occurs when a model wrongly attributes importance to irrelevant but persistent features in the data to achieve better accuracy, e.g., the null block (☑) in Fig. 1. Addressing feature leakage helps suppress" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "type": "text", + "content": "P. Yang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 175 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 175 + ], + "type": "text", + "content": "spurious correlation between the input features and model predictions [2, 39]. Moreover, we establish the robustness of the models trained with our regularization against perturbations from adversarial attacks [17, 31], input pixels [41] and input density, demonstrating the broad applicability of our approach. Our main contributions are summarized as follows." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 135, + 187, + 479, + 281 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 136, + 187, + 479, + 221 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 187, + 479, + 221 + ], + "spans": [ + { + "bbox": [ + 136, + 187, + 479, + 221 + ], + "type": "text", + "content": "1. We identify robust and non-robust features by leveraging attributions, and establish the correlation between model reliance on non-robust features and the smoothness of data marginal density." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 135, + 224, + 479, + 245 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 224, + 479, + 245 + ], + "spans": [ + { + "bbox": [ + 135, + 224, + 479, + 245 + ], + "type": "text", + "content": "2. We propose an efficient technique for calculating the gradients of log density, also addressing the numerical instability of the underlying optimization." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 136, + 247, + 479, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 247, + 479, + 281 + ], + "spans": [ + { + "bbox": [ + 136, + 247, + 479, + 281 + ], + "type": "text", + "content": "3. Through extensive experiments, we demonstrate the effectiveness of the proposed regularization and additionally establish that our approach exhibits general robustness against perturbations." + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 132, + 304, + 237, + 316 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 304, + 237, + 316 + ], + "spans": [ + { + "bbox": [ + 132, + 304, + 237, + 316 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 330, + 482, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 330, + 482, + 484 + ], + "spans": [ + { + "bbox": [ + 130, + 330, + 482, + 484 + ], + "type": "text", + "content": "Regularization for Interpretability Robustness: Despite advancements in model transparency [26, 27, 40, 57], current deep neural networks still lack interpretability in their decision-making process, which is exacerbated by their reliance on non-robust input features. Prior studies, such as [1, 2, 44], identify that standard models are prone to relying on irrelevant or spuriously correlated features. To address that, several regularization techniques are proposed to improve model interpretability. In [14, 42], the authors incorporated prior knowledge into the model training process to regularize the model behavior. Dombrowski et al. [11, 12] found that regularizing the input Hessian using SoftPlus activations or weight decay can boost resilience against manipulated inputs. In [18], a joint energy-based model is trained as a discriminative model for improved robustness. Srinivas and Fleuret [49] enhanced the model interpretability by improving the alignment between the implicit density and the data distribution." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 486, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 486, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 486, + 482, + 665 + ], + "type": "text", + "content": "Regularization for Adversarial Robustness: In addition to the other sources of prediction unreliability, adversarial attacks can manipulate model outputs with imperceptible perturbations to inputs [17, 31]. To address this, adversarial training through data augmentation with adversarial samples is widely employed [9, 31, 43]. Certified adversarial robustness through regularizations [4, 36, 46] is another branch of methods to defend against adversarial perturbations. Inspired by the classic double backpropagation [13], Ross and Doshi-Velez [36] regularized the input gradient norm for adversarial robustness. Etmann [15] further explored different variants of double backpropagation regularizations for various real-world scenarios. Moosavi-Dezfooli et al. [32] also proposed regularization to encourage a low curvature for adversarial robustness. To improve adversarial robustness, Chen et al. [10] computed the norm of attributions integrated from clean samples to adversarial samples as the regularization term. Ilyas et al. [24] proposed a disentangling method to distinguish feature robustness for explaining adversarial examples. However, they do not focus on general feature robustness." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 235, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 235, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 235, + 91, + 447, + 102 + ], + "type": "text", + "content": "Regulating Model Reliance on Non-Robust Features" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "content": "From their perspective, non-robust features are highly predictive; yet adversarially brittle. In contrast, our work focuses on general robustness derived directly from natural images." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 152, + 482, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 152, + 482, + 319 + ], + "spans": [ + { + "bbox": [ + 130, + 152, + 482, + 319 + ], + "type": "text", + "content": "Attribution Methods: Feature attribution methods are used to estimate the importance of input features for a model's prediction and can be categorized as either perturbation-based or back-propagation-based techniques. Perturbation-based methods [16,58,61] calculate the attribution scores by repeatedly perturbing the input features and analyzing the resulting effects on the model prediction. These methods are also extended for evaluating the reliability of the computed attributions [3,34,41,55]. The back-propagation-based techniques [47,52,54,56] estimate the attribution scores by computing the gradients or integrated gradients with respect to the input features in the backward propagation process. In contrast to perturbation-based techniques, back-propagation-based attribution methods offer notable advantages in terms of computational efficiency and reliability. Given the transparency of attributions, our work leverages the attribution framework to distinguish between robust and non-robust features, which enables us to systematically analyze the robustness of input features." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 338, + 367, + 351 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 338, + 367, + 351 + ], + "spans": [ + { + "bbox": [ + 131, + 338, + 367, + 351 + ], + "type": "text", + "content": "3 Feature Robustness by Attributions" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 361, + 484, + 409 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 361, + 484, + 409 + ], + "spans": [ + { + "bbox": [ + 130, + 361, + 484, + 409 + ], + "type": "text", + "content": "We first provide a framework for distinguishing between robust and non-robust features by analyzing their attributions. Herein, attribution inconsistencies among the features with distinct degrees of robustness identify a correlation between the model's reliance on non-robust features and the smoothness of output logits." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 410, + 482, + 493 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 410, + 482, + 493 + ], + "spans": [ + { + "bbox": [ + 130, + 410, + 482, + 493 + ], + "type": "text", + "content": "Let us consider an input sample " + }, + { + "bbox": [ + 130, + 410, + 482, + 493 + ], + "type": "inline_equation", + "content": "x \\in \\mathbb{R}^n" + }, + { + "bbox": [ + 130, + 410, + 482, + 493 + ], + "type": "text", + "content": " with label " + }, + { + "bbox": [ + 130, + 410, + 482, + 493 + ], + "type": "inline_equation", + "content": "y \\in \\mathbb{R}^c" + }, + { + "bbox": [ + 130, + 410, + 482, + 493 + ], + "type": "text", + "content": " from a dataset " + }, + { + "bbox": [ + 130, + 410, + 482, + 493 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 130, + 410, + 482, + 493 + ], + "type": "text", + "content": ", and a classifier " + }, + { + "bbox": [ + 130, + 410, + 482, + 493 + ], + "type": "inline_equation", + "content": "f: \\mathbb{R}^n \\to \\mathbb{R}^c" + }, + { + "bbox": [ + 130, + 410, + 482, + 493 + ], + "type": "text", + "content": " parameterized by " + }, + { + "bbox": [ + 130, + 410, + 482, + 493 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 130, + 410, + 482, + 493 + ], + "type": "text", + "content": ". We denote robust and non-robust features within the input " + }, + { + "bbox": [ + 130, + 410, + 482, + 493 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 130, + 410, + 482, + 493 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 130, + 410, + 482, + 493 + ], + "type": "inline_equation", + "content": "x_{rob}, x_{nrob} \\subseteq x" + }, + { + "bbox": [ + 130, + 410, + 482, + 493 + ], + "type": "text", + "content": ". Consider an attribution method " + }, + { + "bbox": [ + 130, + 410, + 482, + 493 + ], + "type": "inline_equation", + "content": "\\phi: \\mathbb{R}^c \\to \\mathbb{R}^n" + }, + { + "bbox": [ + 130, + 410, + 482, + 493 + ], + "type": "text", + "content": " attributing model predictions to input features by estimating their importance, resulting in an attribution map " + }, + { + "bbox": [ + 130, + 410, + 482, + 493 + ], + "type": "inline_equation", + "content": "M = \\phi(f(x; \\theta))" + }, + { + "bbox": [ + 130, + 410, + 482, + 493 + ], + "type": "text", + "content": ". Inspired by the success of attributions in model explanation, we identify robust and non-robust features by leveraging their attributions." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 494, + 482, + 577 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 494, + 482, + 577 + ], + "spans": [ + { + "bbox": [ + 130, + 494, + 482, + 577 + ], + "type": "text", + "content": "Without loss of generality, we assume that attributions " + }, + { + "bbox": [ + 130, + 494, + 482, + 577 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 130, + 494, + 482, + 577 + ], + "type": "text", + "content": " of the features can be estimated by calculating the change in output logits when these features are removed from the input, following perturbation-based methods [35,58]: " + }, + { + "bbox": [ + 130, + 494, + 482, + 577 + ], + "type": "inline_equation", + "content": "M_{x_{rob}} = f(x;\\theta) - f(x_{[x_{rob} = 0]};\\theta)" + }, + { + "bbox": [ + 130, + 494, + 482, + 577 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 494, + 482, + 577 + ], + "type": "inline_equation", + "content": "M_{x_{nrob}} = f(x;\\theta) - f(x_{[x_{nrob} = 0]};\\theta)" + }, + { + "bbox": [ + 130, + 494, + 482, + 577 + ], + "type": "text", + "content": ". For ease of understanding, we alternatively use " + }, + { + "bbox": [ + 130, + 494, + 482, + 577 + ], + "type": "inline_equation", + "content": "f(x_{nrob};\\theta)" + }, + { + "bbox": [ + 130, + 494, + 482, + 577 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 494, + 482, + 577 + ], + "type": "inline_equation", + "content": "f(x_{rob};\\theta)" + }, + { + "bbox": [ + 130, + 494, + 482, + 577 + ], + "type": "text", + "content": " to represent attributions " + }, + { + "bbox": [ + 130, + 494, + 482, + 577 + ], + "type": "inline_equation", + "content": "M_{rob}" + }, + { + "bbox": [ + 130, + 494, + 482, + 577 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 494, + 482, + 577 + ], + "type": "inline_equation", + "content": "M_{nrob}" + }, + { + "bbox": [ + 130, + 494, + 482, + 577 + ], + "type": "text", + "content": " in the text to follow. We define robust features within the attribution framework as follows." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 586, + 482, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 586, + 482, + 635 + ], + "spans": [ + { + "bbox": [ + 131, + 586, + 482, + 635 + ], + "type": "text", + "content": "Definition 1. A feature " + }, + { + "bbox": [ + 131, + 586, + 482, + 635 + ], + "type": "inline_equation", + "content": "x_{feat}" + }, + { + "bbox": [ + 131, + 586, + 482, + 635 + ], + "type": "text", + "content": " shared among different input instances under its domain " + }, + { + "bbox": [ + 131, + 586, + 482, + 635 + ], + "type": "inline_equation", + "content": "\\Delta_{x_{feat}}" + }, + { + "bbox": [ + 131, + 586, + 482, + 635 + ], + "type": "text", + "content": " is robust if, for a randomly chosen class " + }, + { + "bbox": [ + 131, + 586, + 482, + 635 + ], + "type": "inline_equation", + "content": "y = i" + }, + { + "bbox": [ + 131, + 586, + 482, + 635 + ], + "type": "text", + "content": ", its attribution " + }, + { + "bbox": [ + 131, + 586, + 482, + 635 + ], + "type": "inline_equation", + "content": "M_{x_{feat}}" + }, + { + "bbox": [ + 131, + 586, + 482, + 635 + ], + "type": "text", + "content": " is bounded by a small constant " + }, + { + "bbox": [ + 131, + 586, + 482, + 635 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 131, + 586, + 482, + 635 + ], + "type": "text", + "content": " under a metric " + }, + { + "bbox": [ + 131, + 586, + 482, + 635 + ], + "type": "inline_equation", + "content": "c(\\cdot)" + }, + { + "bbox": [ + 131, + 586, + 482, + 635 + ], + "type": "text", + "content": ", i.e., " + }, + { + "bbox": [ + 131, + 586, + 482, + 635 + ], + "type": "inline_equation", + "content": "c(f(x_{feat};\\theta) - f(\\tilde{x}_{feat};\\theta)) \\leq h : \\tilde{x}_{feat} \\in \\Delta_{x_{feat}}" + }, + { + "bbox": [ + 131, + 586, + 482, + 635 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "content": "Under Definition 1, robust features are expected to contribute consistently to the model's prediction across different input samples. Non-robust features, on" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 222, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 222, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 222, + 102 + ], + "type": "text", + "content": "P. Yang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 200 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 200 + ], + "type": "text", + "content": "the other hand, are those that contribute to the prediction score inconsistently or only under specific conditions. Our focus here is on distinguishing between robust and non-robust features, without requiring to specify a particular metric. Definition 1 emphasizes attribution consistency for robust features rather than attribution positivity, thereby allowing for robust features that can also make a negative contribution to the model's prediction. Building further upon the above definition, we make the following remark." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 210, + 482, + 270 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 210, + 482, + 270 + ], + "spans": [ + { + "bbox": [ + 131, + 210, + 482, + 270 + ], + "type": "text", + "content": "Remark 1. Robust features are largely condition-invariant in that they retain similar attributions despite slight changes to the input. In contrast, non-robust features are condition-specific in that their attributions either vary drastically with slightly varying input conditions, or behave robustly only under specific conditions." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 280, + 480, + 364 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 280, + 480, + 364 + ], + "spans": [ + { + "bbox": [ + 130, + 280, + 480, + 364 + ], + "type": "text", + "content": "Robust features exhibit stable behavior across the input space, which is observable through consistent output logits " + }, + { + "bbox": [ + 130, + 280, + 480, + 364 + ], + "type": "inline_equation", + "content": "f(x_{rob};\\theta)" + }, + { + "bbox": [ + 130, + 280, + 480, + 364 + ], + "type": "text", + "content": " in classifiers regardless of the input instance or class " + }, + { + "bbox": [ + 130, + 280, + 480, + 364 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 130, + 280, + 480, + 364 + ], + "type": "text", + "content": ". In contrast, non-robust features rely on specific conditions to exhibit a particular behavior, which is tailored to a specific class " + }, + { + "bbox": [ + 130, + 280, + 480, + 364 + ], + "type": "inline_equation", + "content": "y = i" + }, + { + "bbox": [ + 130, + 280, + 480, + 364 + ], + "type": "text", + "content": " or the input instance. Robust modeling aims for a stronger reliance on robust features for prediction. Due to the consistency of output logits " + }, + { + "bbox": [ + 130, + 280, + 480, + 364 + ], + "type": "inline_equation", + "content": "f(x;\\theta)" + }, + { + "bbox": [ + 130, + 280, + 480, + 364 + ], + "type": "text", + "content": " for robust features, a smooth " + }, + { + "bbox": [ + 130, + 280, + 480, + 364 + ], + "type": "inline_equation", + "content": "f(x;\\theta)" + }, + { + "bbox": [ + 130, + 280, + 480, + 364 + ], + "type": "text", + "content": " is a desirable property for model robustness." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 382, + 378, + 397 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 382, + 378, + 397 + ], + "spans": [ + { + "bbox": [ + 131, + 382, + 378, + 397 + ], + "type": "text", + "content": "4 Smoothing Marginal Density of Input" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 407, + 479, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 407, + 479, + 443 + ], + "spans": [ + { + "bbox": [ + 130, + 407, + 479, + 443 + ], + "type": "text", + "content": "Here, we establish the relation between model robustness and gradients of the input marginal density. Then, a robust regularization is proposed for regulating model reliance on non-robust features by smoothing marginal density." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 444, + 479, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 444, + 479, + 479 + ], + "spans": [ + { + "bbox": [ + 130, + 444, + 479, + 479 + ], + "type": "text", + "content": "We commence our analytical analysis with probability density, following Bridle [7]. Given a class " + }, + { + "bbox": [ + 130, + 444, + 479, + 479 + ], + "type": "inline_equation", + "content": "y = i" + }, + { + "bbox": [ + 130, + 444, + 479, + 479 + ], + "type": "text", + "content": ", a joint probability density function over the input with the output logit " + }, + { + "bbox": [ + 130, + 444, + 479, + 479 + ], + "type": "inline_equation", + "content": "f_{i}(x;\\theta)" + }, + { + "bbox": [ + 130, + 444, + 479, + 479 + ], + "type": "text", + "content": " is defined as" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 251, + 488, + 480, + 502 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 251, + 488, + 480, + 502 + ], + "spans": [ + { + "bbox": [ + 251, + 488, + 480, + 502 + ], + "type": "interline_equation", + "content": "p _ {\\theta} (x, y = i) = e ^ {f _ {i} (x; \\theta)} / Z _ {\\theta}, \\tag {1}", + "image_path": "2e5c37ca7f5709c080b25502b8710433edbae17779d6c922d813f5016cfc7305.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 510, + 480, + 583 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 480, + 583 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 480, + 583 + ], + "type": "text", + "content": "where the constant " + }, + { + "bbox": [ + 130, + 510, + 480, + 583 + ], + "type": "inline_equation", + "content": "Z_{\\theta} = \\int e^{f_i(x;\\theta)}dx" + }, + { + "bbox": [ + 130, + 510, + 480, + 583 + ], + "type": "text", + "content": " is the partition function. " + }, + { + "bbox": [ + 130, + 510, + 480, + 583 + ], + "type": "inline_equation", + "content": "Z_{\\theta}" + }, + { + "bbox": [ + 130, + 510, + 480, + 583 + ], + "type": "text", + "content": " normalizes the input " + }, + { + "bbox": [ + 130, + 510, + 480, + 583 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 130, + 510, + 480, + 583 + ], + "type": "text", + "content": " to a probability density by integrating over all possible input points " + }, + { + "bbox": [ + 130, + 510, + 480, + 583 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 130, + 510, + 480, + 583 + ], + "type": "text", + "content": " in the input space via the model " + }, + { + "bbox": [ + 130, + 510, + 480, + 583 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 130, + 510, + 480, + 583 + ], + "type": "text", + "content": ". By applying Bayes' rule, we eliminate the condition " + }, + { + "bbox": [ + 130, + 510, + 480, + 583 + ], + "type": "inline_equation", + "content": "y = i" + }, + { + "bbox": [ + 130, + 510, + 480, + 583 + ], + "type": "text", + "content": ", resulting in the marginal density being defined solely on the input " + }, + { + "bbox": [ + 130, + 510, + 480, + 583 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 130, + 510, + 480, + 583 + ], + "type": "text", + "content": ": " + }, + { + "bbox": [ + 130, + 510, + 480, + 583 + ], + "type": "inline_equation", + "content": "p_{\\theta}(x) = p_{\\theta}(y = i,x) / p_{\\theta}(y = i|x)" + }, + { + "bbox": [ + 130, + 510, + 480, + 583 + ], + "type": "text", + "content": ". The conditional density function " + }, + { + "bbox": [ + 130, + 510, + 480, + 583 + ], + "type": "inline_equation", + "content": "p_{\\theta}(y = i|x)" + }, + { + "bbox": [ + 130, + 510, + 480, + 583 + ], + "type": "text", + "content": " can be further defined as" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 244, + 592, + 480, + 606 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 244, + 592, + 480, + 606 + ], + "spans": [ + { + "bbox": [ + 244, + 592, + 480, + 606 + ], + "type": "interline_equation", + "content": "p _ {\\theta} (y = i | x) = e ^ {f _ {i} (x; \\theta)} / Z _ {f (x; \\theta)}, \\tag {2}", + "image_path": "ebd7786d165698174075bae75628b1ed9fb1d20a8091991f66218712e21a2d2e.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 616, + 482, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 616, + 482, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 616, + 482, + 667 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 616, + 482, + 667 + ], + "type": "inline_equation", + "content": "Z_{f(x;\\theta)} = \\sum_{i=1}^{C} e^{f_i(x;\\theta)}" + }, + { + "bbox": [ + 130, + 616, + 482, + 667 + ], + "type": "text", + "content": " is the partition function for the output logits " + }, + { + "bbox": [ + 130, + 616, + 482, + 667 + ], + "type": "inline_equation", + "content": "f_i(x;\\theta)" + }, + { + "bbox": [ + 130, + 616, + 482, + 667 + ], + "type": "text", + "content": " defined on all the " + }, + { + "bbox": [ + 130, + 616, + 482, + 667 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 130, + 616, + 482, + 667 + ], + "type": "text", + "content": " classes. To simplify the notation, we use " + }, + { + "bbox": [ + 130, + 616, + 482, + 667 + ], + "type": "inline_equation", + "content": "Z_{f(x)}" + }, + { + "bbox": [ + 130, + 616, + 482, + 667 + ], + "type": "text", + "content": " to represent " + }, + { + "bbox": [ + 130, + 616, + 482, + 667 + ], + "type": "inline_equation", + "content": "Z_{f(x;\\theta)}" + }, + { + "bbox": [ + 130, + 616, + 482, + 667 + ], + "type": "text", + "content": " in the subsequent discussion. Exploiting the symmetry property of the joint density defined in Eq. 1, i.e., " + }, + { + "bbox": [ + 130, + 616, + 482, + 667 + ], + "type": "inline_equation", + "content": "p_\\theta(x,y=i) = p_\\theta(y=i,x)" + }, + { + "bbox": [ + 130, + 616, + 482, + 667 + ], + "type": "text", + "content": ", the" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 235, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 235, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 235, + 91, + 448, + 102 + ], + "type": "text", + "content": "Regulating Model Reliance on Non-Robust Features" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 323, + 128 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 323, + 128 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 323, + 128 + ], + "type": "text", + "content": "marginal density " + }, + { + "bbox": [ + 132, + 116, + 323, + 128 + ], + "type": "inline_equation", + "content": "p_{\\theta}(x)" + }, + { + "bbox": [ + 132, + 116, + 323, + 128 + ], + "type": "text", + "content": " can be expressed as" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 238, + 138, + 481, + 167 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 238, + 138, + 481, + 167 + ], + "spans": [ + { + "bbox": [ + 238, + 138, + 481, + 167 + ], + "type": "interline_equation", + "content": "p _ {\\theta} (x) = \\frac {e ^ {f _ {i} (x ; \\theta)} / Z _ {\\theta}}{e ^ {f _ {i} (x ; \\theta)} / Z _ {f (x)}} = \\frac {Z _ {f (x)}}{Z _ {\\theta}}. \\tag {3}", + "image_path": "ed10d871b870ec45f6eee565a961771196d31516bf6f1c500d9a1d485040dce6.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 175, + 482, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 175, + 482, + 331 + ], + "spans": [ + { + "bbox": [ + 130, + 175, + 482, + 331 + ], + "type": "text", + "content": "As identified in the previous section, smoothness of output logits across inputs encourages the use of robust features by the model. Hence, we consider the marginal density " + }, + { + "bbox": [ + 130, + 175, + 482, + 331 + ], + "type": "inline_equation", + "content": "p_{\\theta}(x)" + }, + { + "bbox": [ + 130, + 175, + 482, + 331 + ], + "type": "text", + "content": " defined on the output logits " + }, + { + "bbox": [ + 130, + 175, + 482, + 331 + ], + "type": "inline_equation", + "content": "f(x; \\theta)" + }, + { + "bbox": [ + 130, + 175, + 482, + 331 + ], + "type": "text", + "content": " across the input space. Promoting a smaller gradient of the marginal density with respect to the input " + }, + { + "bbox": [ + 130, + 175, + 482, + 331 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 130, + 175, + 482, + 331 + ], + "type": "text", + "content": ", denoted as " + }, + { + "bbox": [ + 130, + 175, + 482, + 331 + ], + "type": "inline_equation", + "content": "\\nabla_x p_{\\theta}(x)" + }, + { + "bbox": [ + 130, + 175, + 482, + 331 + ], + "type": "text", + "content": ", contributes to the smoothness of the output logits. Thus, a positive correlation can be established between the use of robust features and the smoothness of the probability marginal density " + }, + { + "bbox": [ + 130, + 175, + 482, + 331 + ], + "type": "inline_equation", + "content": "p_{\\theta}(x)" + }, + { + "bbox": [ + 130, + 175, + 482, + 331 + ], + "type": "text", + "content": ". In particular, the smooth output logits of robust features across input samples suggest that these features will have relatively small gradients of the density " + }, + { + "bbox": [ + 130, + 175, + 482, + 331 + ], + "type": "inline_equation", + "content": "p_{\\theta}(x)" + }, + { + "bbox": [ + 130, + 175, + 482, + 331 + ], + "type": "text", + "content": " with respect to the input values. On the other hand, non-robust features with fluctuating output logits will have large gradients of the density that need to be suppressed during the training process for model robustness. Therefore, we can conclude with the following remark." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 342, + 483, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 342, + 483, + 380 + ], + "spans": [ + { + "bbox": [ + 132, + 342, + 483, + 380 + ], + "type": "text", + "content": "Remark 2. Model reliance on non-robust features " + }, + { + "bbox": [ + 132, + 342, + 483, + 380 + ], + "type": "inline_equation", + "content": "x_{nrob}" + }, + { + "bbox": [ + 132, + 342, + 483, + 380 + ], + "type": "text", + "content": " can be regulated by regularizing the gradients of the marginal density " + }, + { + "bbox": [ + 132, + 342, + 483, + 380 + ], + "type": "inline_equation", + "content": "p_{\\theta}(x)" + }, + { + "bbox": [ + 132, + 342, + 483, + 380 + ], + "type": "text", + "content": " w.r.t. " + }, + { + "bbox": [ + 132, + 342, + 483, + 380 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 132, + 342, + 483, + 380 + ], + "type": "text", + "content": ", and this regularization can be achieved through optimizing the model parameters." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 390, + 482, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 390, + 482, + 521 + ], + "spans": [ + { + "bbox": [ + 130, + 390, + 482, + 521 + ], + "type": "text", + "content": "In the light of Remark 2., we propose a regularization term for minimizing the gradients of marginal density. However, computing the gradients of the marginal density " + }, + { + "bbox": [ + 130, + 390, + 482, + 521 + ], + "type": "inline_equation", + "content": "\\nabla_{x}p_{\\theta}(x)" + }, + { + "bbox": [ + 130, + 390, + 482, + 521 + ], + "type": "text", + "content": " is not feasible because the partition function defined on the entire input space is intractable. To avoid the estimation of the intractable " + }, + { + "bbox": [ + 130, + 390, + 482, + 521 + ], + "type": "inline_equation", + "content": "Z_{\\theta}" + }, + { + "bbox": [ + 130, + 390, + 482, + 521 + ], + "type": "text", + "content": ", we instead compute the gradients with respect to the log density as " + }, + { + "bbox": [ + 130, + 390, + 482, + 521 + ], + "type": "inline_equation", + "content": "\\nabla_{x}\\log p_{\\theta}(x) = \\nabla_{x}\\log Z_{f(x)}" + }, + { + "bbox": [ + 130, + 390, + 482, + 521 + ], + "type": "text", + "content": ". This is possible because " + }, + { + "bbox": [ + 130, + 390, + 482, + 521 + ], + "type": "inline_equation", + "content": "Z_{\\theta}" + }, + { + "bbox": [ + 130, + 390, + 482, + 521 + ], + "type": "text", + "content": " solely depends on the model parameter " + }, + { + "bbox": [ + 130, + 390, + 482, + 521 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 130, + 390, + 482, + 521 + ], + "type": "text", + "content": " and not the input " + }, + { + "bbox": [ + 130, + 390, + 482, + 521 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 130, + 390, + 482, + 521 + ], + "type": "text", + "content": ". Expanding the partition function " + }, + { + "bbox": [ + 130, + 390, + 482, + 521 + ], + "type": "inline_equation", + "content": "Z_{f(x)}" + }, + { + "bbox": [ + 130, + 390, + 482, + 521 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 130, + 390, + 482, + 521 + ], + "type": "inline_equation", + "content": "\\nabla_{x}\\log p_{\\theta}(x)" + }, + { + "bbox": [ + 130, + 390, + 482, + 521 + ], + "type": "text", + "content": ", we obtain " + }, + { + "bbox": [ + 130, + 390, + 482, + 521 + ], + "type": "inline_equation", + "content": "\\nabla_{x}\\log p_{\\theta}(x) = \\sum_{i=1}^{C}\\nabla_{x}e^{f_{i}(x;\\theta)} / \\sum_{i=1}^{C}e^{f_{i}(x;\\theta)}" + }, + { + "bbox": [ + 130, + 390, + 482, + 521 + ], + "type": "text", + "content": ". The " + }, + { + "bbox": [ + 130, + 390, + 482, + 521 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 130, + 390, + 482, + 521 + ], + "type": "text", + "content": "-norm of this gradient is computed as the regularization term. In the optimization process, the goal is to find the optimal parameter " + }, + { + "bbox": [ + 130, + 390, + 482, + 521 + ], + "type": "inline_equation", + "content": "\\theta^{*}" + }, + { + "bbox": [ + 130, + 390, + 482, + 521 + ], + "type": "text", + "content": ", cf. Remark 2., by minimizing the loss " + }, + { + "bbox": [ + 130, + 390, + 482, + 521 + ], + "type": "inline_equation", + "content": "\\ell" + }, + { + "bbox": [ + 130, + 390, + 482, + 521 + ], + "type": "text", + "content": " as" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 207, + 522, + 481, + 553 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 207, + 522, + 481, + 553 + ], + "spans": [ + { + "bbox": [ + 207, + 522, + 481, + 553 + ], + "type": "interline_equation", + "content": "\\theta^ {*} = \\min _ {\\theta} \\ell (f (x; \\theta), y) + \\lambda | | \\frac {\\sum_ {i = 1} ^ {C} \\nabla_ {x} e ^ {f _ {i} (x ; \\theta)}}{\\sum_ {i = 1} ^ {C} e ^ {f _ {i} (x ; \\theta)}} | | _ {p}, \\tag {4}", + "image_path": "29efbd4a5d3a0af7deb35a7c23ab98a4a6ca2ec2b966702ee4a68e00dc1ee5d4.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 557, + 482, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 557, + 482, + 581 + ], + "spans": [ + { + "bbox": [ + 130, + 557, + 482, + 581 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 557, + 482, + 581 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 130, + 557, + 482, + 581 + ], + "type": "text", + "content": " indicates the magnitude of the coefficient for controlling the strength of the regularization." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 582, + 482, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 482, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 482, + 667 + ], + "type": "text", + "content": "To regulate model reliance on non-robust features, our regularization encourages the smoothness of marginal density by regularizing its gradients. Since the output logit change in the log-marginal density " + }, + { + "bbox": [ + 130, + 582, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\log p_{\\theta}(x) = \\log Z_{f(x)} - \\log Z_{\\theta}" + }, + { + "bbox": [ + 130, + 582, + 482, + 667 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 582, + 482, + 667 + ], + "type": "inline_equation", + "content": "Z_{\\theta}" + }, + { + "bbox": [ + 130, + 582, + 482, + 667 + ], + "type": "text", + "content": " is independent of the input " + }, + { + "bbox": [ + 130, + 582, + 482, + 667 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 130, + 582, + 482, + 667 + ], + "type": "text", + "content": ", we can only focus on the first term " + }, + { + "bbox": [ + 130, + 582, + 482, + 667 + ], + "type": "inline_equation", + "content": "Z_{f(x)} = \\sum_{i=1}^{C} e^{f_i(x;\\theta)}" + }, + { + "bbox": [ + 130, + 582, + 482, + 667 + ], + "type": "text", + "content": ". Recall, Definition 1 of robust features. Assuming the robust input feature " + }, + { + "bbox": [ + 130, + 582, + 482, + 667 + ], + "type": "inline_equation", + "content": "x_{rob}" + }, + { + "bbox": [ + 130, + 582, + 482, + 667 + ], + "type": "text", + "content": " exists in a random input " + }, + { + "bbox": [ + 130, + 582, + 482, + 667 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 130, + 582, + 482, + 667 + ], + "type": "text", + "content": ", the corresponding output logit " + }, + { + "bbox": [ + 130, + 582, + 482, + 667 + ], + "type": "inline_equation", + "content": "f_i(x_{rob})" + }, + { + "bbox": [ + 130, + 582, + 482, + 667 + ], + "type": "text", + "content": " will consistently attribute to the model prediction. This property" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 222, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 222, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 222, + 102 + ], + "type": "text", + "content": "P. Yang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 114, + 482, + 248 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 114, + 482, + 248 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 482, + 248 + ], + "type": "text", + "content": "of " + }, + { + "bbox": [ + 130, + 114, + 482, + 248 + ], + "type": "inline_equation", + "content": "x_{rob}" + }, + { + "bbox": [ + 130, + 114, + 482, + 248 + ], + "type": "text", + "content": " leads to the smoothness of output change " + }, + { + "bbox": [ + 130, + 114, + 482, + 248 + ], + "type": "inline_equation", + "content": "\\sum_{i=1}^{C} e^{f_i(x_{rob};\\theta)}" + }, + { + "bbox": [ + 130, + 114, + 482, + 248 + ], + "type": "text", + "content": " across different input samples and class labels. In contrast, non-robust input features " + }, + { + "bbox": [ + 130, + 114, + 482, + 248 + ], + "type": "inline_equation", + "content": "x_{nrob}" + }, + { + "bbox": [ + 130, + 114, + 482, + 248 + ], + "type": "text", + "content": " show relatively high attributions for the output logit " + }, + { + "bbox": [ + 130, + 114, + 482, + 248 + ], + "type": "inline_equation", + "content": "f_i(x_{nrob})" + }, + { + "bbox": [ + 130, + 114, + 482, + 248 + ], + "type": "text", + "content": " for a given class " + }, + { + "bbox": [ + 130, + 114, + 482, + 248 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 114, + 482, + 248 + ], + "type": "text", + "content": ". However, they cannot maintain consistency in the attributions across different inputs or labels, leading to fluctuations in the output change " + }, + { + "bbox": [ + 130, + 114, + 482, + 248 + ], + "type": "inline_equation", + "content": "\\sum_{i=1}^{C} e^{f_i(x_{nrob};\\theta)}" + }, + { + "bbox": [ + 130, + 114, + 482, + 248 + ], + "type": "text", + "content": ". In Fig. 1(a) and Fig. 13 in Supp. 17, non-robust features in the Null block exhibit fluctuating attributions across different samples on the standard trained model. It is demonstrated that the magnitude of gradients for input features in the marginal density " + }, + { + "bbox": [ + 130, + 114, + 482, + 248 + ], + "type": "inline_equation", + "content": "p_\\theta(x)" + }, + { + "bbox": [ + 130, + 114, + 482, + 248 + ], + "type": "text", + "content": " reflects the model's sensitivity to those features. We leverage this relation to mitigate model reliance on the non-robust features by smoothing the marginal density of the input samples." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 267, + 478, + 282 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 267, + 478, + 282 + ], + "spans": [ + { + "bbox": [ + 132, + 267, + 478, + 282 + ], + "type": "text", + "content": "5 Stable and Efficient Implementation for Regularization" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 294, + 482, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 294, + 482, + 474 + ], + "spans": [ + { + "bbox": [ + 130, + 294, + 482, + 474 + ], + "type": "text", + "content": "From the implementation perspective, the gradient computation of marginal density involves multiple exponential operations in both the numerator and denominator of Eq. 4 which can introduce numerical instability in the optimization process, leading to gradient vanishing and explosion problems. Such issues can potentially hinder the application of our regularization to large non-linear models or wide-distribution data. For instance, batch normalization (BN) layers [25] solving internal covariate shifts with learnable scaling and shifting parameters can amplify the errors caused by the exponential operations during the backpropagation. In Fig. 2(a), yellow and red curves indicate original and inverse implementations for minimizing the gradient of log density " + }, + { + "bbox": [ + 130, + 294, + 482, + 474 + ], + "type": "inline_equation", + "content": "\\nabla_{x}\\log Z_{f(x)}" + }, + { + "bbox": [ + 130, + 294, + 482, + 474 + ], + "type": "text", + "content": ". It is illustrated that the Frobenius gradient norm of both implementations undergoes rapid numerical overflow as the number of training iterations increases, revealing the serious issue of exploding and vanishing gradients using the proposed regularization. Therefore, it is crucial to address these through careful implementation to ensure the feasibility of our regularization." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 474, + 482, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 474, + 482, + 534 + ], + "spans": [ + { + "bbox": [ + 130, + 474, + 482, + 534 + ], + "type": "text", + "content": "To address this challenge, we transform the computation from the summation of exponential operations to softmax. By subtracting a constant value " + }, + { + "bbox": [ + 130, + 474, + 482, + 534 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 130, + 474, + 482, + 534 + ], + "type": "text", + "content": " from the logits before exponentiation, softmax can prevent numerical overflow during the exponential operations, i.e., " + }, + { + "bbox": [ + 130, + 474, + 482, + 534 + ], + "type": "inline_equation", + "content": "e^{f_i(x;\\theta)} / \\sum_j e^{f_j(x;\\theta)} = e^{f_i(x;\\theta) - \\eta} / \\sum_j e^{f_j(x;\\theta) - \\eta}" + }, + { + "bbox": [ + 130, + 474, + 482, + 534 + ], + "type": "text", + "content": ". Thus, we incorporate the softmax function into the computation of log density" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 157, + 540, + 249, + 616 + ], + "blocks": [ + { + "bbox": [ + 157, + 540, + 249, + 616 + ], + "lines": [ + { + "bbox": [ + 157, + 540, + 249, + 616 + ], + "spans": [ + { + "bbox": [ + 157, + 540, + 249, + 616 + ], + "type": "image", + "image_path": "4883dab3090e5047ad2a3d730cd81760d88a43f83de7a679acb12b363c6048f8.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 169, + 621, + 242, + 631 + ], + "lines": [ + { + "bbox": [ + 169, + 621, + 242, + 631 + ], + "spans": [ + { + "bbox": [ + 169, + 621, + 242, + 631 + ], + "type": "text", + "content": "(a) Numerical Stability." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 263, + 540, + 356, + 615 + ], + "blocks": [ + { + "bbox": [ + 263, + 540, + 356, + 615 + ], + "lines": [ + { + "bbox": [ + 263, + 540, + 356, + 615 + ], + "spans": [ + { + "bbox": [ + 263, + 540, + 356, + 615 + ], + "type": "image", + "image_path": "df2bee27ab0b88dec8a8b67ecd54d5d671cc737f30fec82477cb9670f6b7db52.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 272, + 621, + 345, + 631 + ], + "lines": [ + { + "bbox": [ + 272, + 621, + 345, + 631 + ], + "spans": [ + { + "bbox": [ + 272, + 621, + 345, + 631 + ], + "type": "text", + "content": "(b) Training Efficiency." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 364, + 541, + 455, + 619 + ], + "blocks": [ + { + "bbox": [ + 364, + 541, + 455, + 619 + ], + "lines": [ + { + "bbox": [ + 364, + 541, + 455, + 619 + ], + "spans": [ + { + "bbox": [ + 364, + 541, + 455, + 619 + ], + "type": "image", + "image_path": "421827394b8016dee901706db3a42f0c89dd6b46b3628ba42c3a297682adb77e.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 388, + 621, + 444, + 631 + ], + "lines": [ + { + "bbox": [ + 388, + 621, + 444, + 631 + ], + "spans": [ + { + "bbox": [ + 388, + 621, + 444, + 631 + ], + "type": "text", + "content": "(c) Training Time." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 639, + 482, + 662 + ], + "lines": [ + { + "bbox": [ + 130, + 639, + 482, + 662 + ], + "spans": [ + { + "bbox": [ + 130, + 639, + 482, + 662 + ], + "type": "text", + "content": "Fig. 2: The comparison of numerical stability, training efficiency and training time across different ResNet-34 models." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 235, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 235, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 235, + 91, + 448, + 102 + ], + "type": "text", + "content": "Regulating Model Reliance on Non-Robust Features" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 234, + 129 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 234, + 129 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 234, + 129 + ], + "type": "text", + "content": "gradient " + }, + { + "bbox": [ + 132, + 116, + 234, + 129 + ], + "type": "inline_equation", + "content": "\\nabla_{x}logz_{f(x)}" + }, + { + "bbox": [ + 132, + 116, + 234, + 129 + ], + "type": "text", + "content": " as" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 195, + 139, + 418, + 168 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 195, + 139, + 418, + 168 + ], + "spans": [ + { + "bbox": [ + 195, + 139, + 418, + 168 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\nabla_ {x} l o g z _ {f (x)} = \\nabla_ {x} f _ {i} (x; \\theta) - \\nabla_ {x} (l o g e ^ {f _ {i} (x; \\theta)} - l o g z _ {f (x)}), \\\\ = \\nabla_ {x} f _ {i} (x; \\theta) - \\nabla_ {x} \\left(\\log \\left(f _ {i} (x; \\theta) / z _ {f (x)}\\right)\\right), \\\\ \\end{array}", + "image_path": "7d2ffa71317c6c5756dd5424ec6abfaf4747f5c100ad8950546a1588c7f2b2c0.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 466, + 148, + 479, + 159 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 466, + 148, + 479, + 159 + ], + "spans": [ + { + "bbox": [ + 466, + 148, + 479, + 159 + ], + "type": "text", + "content": "(5)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 178, + 480, + 227 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 178, + 480, + 227 + ], + "spans": [ + { + "bbox": [ + 130, + 178, + 480, + 227 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 178, + 480, + 227 + ], + "type": "inline_equation", + "content": "f_{i}(x)" + }, + { + "bbox": [ + 130, + 178, + 480, + 227 + ], + "type": "text", + "content": " indicates a logit of a random " + }, + { + "bbox": [ + 130, + 178, + 480, + 227 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 178, + 480, + 227 + ], + "type": "text", + "content": "-th class, and " + }, + { + "bbox": [ + 130, + 178, + 480, + 227 + ], + "type": "inline_equation", + "content": "z_{f(x)}" + }, + { + "bbox": [ + 130, + 178, + 480, + 227 + ], + "type": "text", + "content": " equals " + }, + { + "bbox": [ + 130, + 178, + 480, + 227 + ], + "type": "inline_equation", + "content": "\\sum_{j} e^{f_{j}(x; \\theta)}" + }, + { + "bbox": [ + 130, + 178, + 480, + 227 + ], + "type": "text", + "content": ". Thus, the gradients with respect to the log-marginal density " + }, + { + "bbox": [ + 130, + 178, + 480, + 227 + ], + "type": "inline_equation", + "content": "p_{\\theta}(x)" + }, + { + "bbox": [ + 130, + 178, + 480, + 227 + ], + "type": "text", + "content": " can be replaced by computing the difference between the gradient " + }, + { + "bbox": [ + 130, + 178, + 480, + 227 + ], + "type": "inline_equation", + "content": "\\nabla_{x} f_{i}(x; \\theta)" + }, + { + "bbox": [ + 130, + 178, + 480, + 227 + ], + "type": "text", + "content": " and the gradient of a log-softmax output " + }, + { + "bbox": [ + 130, + 178, + 480, + 227 + ], + "type": "inline_equation", + "content": "\\nabla_{x} \\log(e^{f_{i}(x; \\theta)} / Z_{f(x)})" + }, + { + "bbox": [ + 130, + 178, + 480, + 227 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 227, + 481, + 359 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 227, + 481, + 359 + ], + "spans": [ + { + "bbox": [ + 130, + 227, + 481, + 359 + ], + "type": "text", + "content": "Our technique improves upon the common approach [5, 6] to achieve numerical stability in log exponential sum calculations, which typically employs the formula " + }, + { + "bbox": [ + 130, + 227, + 481, + 359 + ], + "type": "inline_equation", + "content": "\\log(\\sum_{i \\in \\{1, \\ldots, n\\}} e^{x_i}) = \\eta + \\log(\\sum_{i \\in \\{1, \\ldots, n\\}} e^{x_i - \\eta})" + }, + { + "bbox": [ + 130, + 227, + 481, + 359 + ], + "type": "text", + "content": ", with " + }, + { + "bbox": [ + 130, + 227, + 481, + 359 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 130, + 227, + 481, + 359 + ], + "type": "text", + "content": " being the maximum value of inputs " + }, + { + "bbox": [ + 130, + 227, + 481, + 359 + ], + "type": "inline_equation", + "content": "\\{x_1, \\ldots, x_n\\}" + }, + { + "bbox": [ + 130, + 227, + 481, + 359 + ], + "type": "text", + "content": ". We address the numerical instability by employing softmax, avoiding computationally expensive comparisons of the maximum gradient values. Specifically, the basic stable implementation involves finding the maximum gradient within the gradients of various classes, leading to a time complexity of " + }, + { + "bbox": [ + 130, + 227, + 481, + 359 + ], + "type": "inline_equation", + "content": "O(n)" + }, + { + "bbox": [ + 130, + 227, + 481, + 359 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 227, + 481, + 359 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 130, + 227, + 481, + 359 + ], + "type": "text", + "content": " represents the number of classes. In contrast, our method eliminates numerical instabilities associated with a randomly selected class, eliminating the need for maximum value comparisons and leading to a more favorable time complexity of " + }, + { + "bbox": [ + 130, + 227, + 481, + 359 + ], + "type": "inline_equation", + "content": "O(1)" + }, + { + "bbox": [ + 130, + 227, + 481, + 359 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 361, + 482, + 431 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 361, + 482, + 431 + ], + "spans": [ + { + "bbox": [ + 130, + 361, + 482, + 431 + ], + "type": "text", + "content": "While our solution helps avoid numerical instability, it requires twice the gradient computations as compared to the sole calculation of density gradients. Hence, we propose an efficient mechanism for estimating the difference in the gradients. Specifically, we compute the gradient of the difference between two outputs to approximate the difference between the two gradients of outputs using Taylor series as" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 149, + 441, + 480, + 455 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 441, + 480, + 455 + ], + "spans": [ + { + "bbox": [ + 149, + 441, + 480, + 455 + ], + "type": "interline_equation", + "content": "\\nabla_ {x} f _ {i} (x; \\theta) - \\nabla_ {x} \\log \\left(e ^ {f _ {i} (x; \\theta)} / Z _ {f (x)}\\right) \\approx \\nabla_ {x} \\left(f _ {i} (x; \\theta) - \\log \\left(e ^ {f _ {i} (x; \\theta)} / Z _ {f (x)}\\right)\\right). \\tag {6}", + "image_path": "fce8a2c8ab1e3e8c5b113041a4ab93a9283457cd7539d782b6f7c6ddd163072b.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 464, + 482, + 596 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 464, + 482, + 596 + ], + "spans": [ + { + "bbox": [ + 130, + 464, + 482, + 596 + ], + "type": "text", + "content": "The proof of Eq. 6 is provided in Supp. 9. As such, the proposed approach enables stable and efficient model optimization. The blue curve in Fig. 2(a) shows that our efficient implementation can effectively avoid numerical instability in the gradient computation. Fig. 2(b) provides a comparison of training efficiency on CIFAR-10 [28] between the basic numerically stable implementation and our two derived alternatives. More results regarding the comparison of numerical stability are provided in Supp. 15. In Fig. 2(c), we delve into a specific examination of the training times of models trained across the three datasets including BlockMNIST, CelebA and CIFAR-10. It is evident that our methods substantially improve the per-epoch training time. Moreover, our methods can yield further beneficial trade-offs as the number of classes increases." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 615, + 471, + 629 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 615, + 471, + 629 + ], + "spans": [ + { + "bbox": [ + 130, + 615, + 471, + 629 + ], + "type": "text", + "content": "6 Limited Robustness in Input Gradient Regularization" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "content": "Input gradient regularization (InputGrad Reg.) [29,36] computes the Frobenius norm of input gradients " + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "inline_equation", + "content": "||\\nabla_x f_i(x;\\theta)||_F" + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "content": " for a given class label " + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "inline_equation", + "content": "y = i" + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "content": ", which is a" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "type": "text", + "content": "P. Yang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 176 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 176 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 176 + ], + "type": "text", + "content": "baseline robust regularization for model optimization. Existing works [29,36,37] explain InputGrad Reg. as a prediction stability technique for robustness against perturbations. Here, we reveal that the input gradient norm potentially regularizes the gradients of implicit data density. Moreover, we provide an understanding of how InputGrad Reg. encourages robustness as well as its limitations." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 176, + 479, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 176, + 479, + 200 + ], + "spans": [ + { + "bbox": [ + 130, + 176, + 479, + 200 + ], + "type": "text", + "content": "Suppose all the classes have equal probability, i.e., " + }, + { + "bbox": [ + 130, + 176, + 479, + 200 + ], + "type": "inline_equation", + "content": "p_{\\theta}(y = i) = 1 / C" + }, + { + "bbox": [ + 130, + 176, + 479, + 200 + ], + "type": "text", + "content": ". We can express the class-conditional density " + }, + { + "bbox": [ + 130, + 176, + 479, + 200 + ], + "type": "inline_equation", + "content": "p_{\\theta}(x|y)" + }, + { + "bbox": [ + 130, + 176, + 479, + 200 + ], + "type": "text", + "content": " by using Bayes's rule as" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 226, + 210, + 481, + 236 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 226, + 210, + 481, + 236 + ], + "spans": [ + { + "bbox": [ + 226, + 210, + 481, + 236 + ], + "type": "interline_equation", + "content": "p _ {\\theta} (x \\mid y = i) = \\frac {p _ {\\theta} (x , y = i)}{p _ {\\theta} (y = i)} = \\frac {e ^ {f _ {i} (x ; \\theta)}}{Z _ {\\theta} / C}. \\tag {7}", + "image_path": "c8bd49c7a90e40b2f5763b5b69908c588f8d67a3eed422ea15b8a3018555e5e0.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 244, + 479, + 268 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 244, + 479, + 268 + ], + "spans": [ + { + "bbox": [ + 131, + 244, + 479, + 268 + ], + "type": "text", + "content": "Now, we compute the gradients of the log density defined in Eq. 7 with respect to the input " + }, + { + "bbox": [ + 131, + 244, + 479, + 268 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 131, + 244, + 479, + 268 + ], + "type": "text", + "content": " as" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 192, + 280, + 480, + 292 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 192, + 280, + 480, + 292 + ], + "spans": [ + { + "bbox": [ + 192, + 280, + 480, + 292 + ], + "type": "interline_equation", + "content": "\\nabla_ {x} \\log p _ {\\theta} (x, y = i) = \\nabla_ {x} \\log p _ {\\theta} (x | y = i) = \\nabla_ {x} f _ {i} (x; \\theta). \\tag {8}", + "image_path": "9fa968d92d68947e93a30ae978f6dc5490676f2e0f34d617ce04a6b5907e9b6f.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 301, + 482, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 301, + 482, + 468 + ], + "spans": [ + { + "bbox": [ + 130, + 301, + 482, + 468 + ], + "type": "text", + "content": "Eq. 8 demonstrates that the input gradients can be interpreted as the gradients of either the log joint density or the log conditional density with respect to the input " + }, + { + "bbox": [ + 130, + 301, + 482, + 468 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 130, + 301, + 482, + 468 + ], + "type": "text", + "content": ". This formulation highlights that InputGrad Reg. encourages consistent attributions of input features for the model's predictions. However, it is important to note that InputGrad Reg. is formulated under the specific condition " + }, + { + "bbox": [ + 130, + 301, + 482, + 468 + ], + "type": "inline_equation", + "content": "y = i" + }, + { + "bbox": [ + 130, + 301, + 482, + 468 + ], + "type": "text", + "content": ", which limits its effectiveness in resolving inconsistencies when predicting a different class " + }, + { + "bbox": [ + 130, + 301, + 482, + 468 + ], + "type": "inline_equation", + "content": "y = j" + }, + { + "bbox": [ + 130, + 301, + 482, + 468 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 301, + 482, + 468 + ], + "type": "inline_equation", + "content": "j \\neq i" + }, + { + "bbox": [ + 130, + 301, + 482, + 468 + ], + "type": "text", + "content": ". Although input features consistently contribute to the model's prediction under a given class, InputGrad Reg. fails to consider inconsistent attributions of these features across different classes, thereby allowing model non-robust behavior to exist. Consequently, a model trained with input gradient regularization may still exhibit spurious robustness relying on specific conditions. Fig. 1(a) illustrates an example where the model trained with InputGrad Reg. is incapable of suppressing non-robust features in Null block." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 479, + 482, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 479, + 482, + 515 + ], + "spans": [ + { + "bbox": [ + 131, + 479, + 482, + 515 + ], + "type": "text", + "content": "Remark 3. Input gradient regularization smoothens the joint and conditional density of the input " + }, + { + "bbox": [ + 131, + 479, + 482, + 515 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 131, + 479, + 482, + 515 + ], + "type": "text", + "content": " under a specific label " + }, + { + "bbox": [ + 131, + 479, + 482, + 515 + ], + "type": "inline_equation", + "content": "y = i" + }, + { + "bbox": [ + 131, + 479, + 482, + 515 + ], + "type": "text", + "content": ", compromising its ability to resist the class-specific non-robust features." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 525, + 482, + 597 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 525, + 482, + 597 + ], + "spans": [ + { + "bbox": [ + 130, + 525, + 482, + 597 + ], + "type": "text", + "content": "In contrast to the existing works [36, 49] that highlight the reasons for Input-Grad Reg. efficacy, we reveal a weakness of this technique, cf. Remark 3.. Unlike regularizing gradients based on joint or conditional densities, our approach allows for more effective regularization without imposing the condition " + }, + { + "bbox": [ + 130, + 525, + 482, + 597 + ], + "type": "inline_equation", + "content": "y = i" + }, + { + "bbox": [ + 130, + 525, + 482, + 597 + ], + "type": "text", + "content": ". Our method focuses on regularizing the gradients of marginal density " + }, + { + "bbox": [ + 130, + 525, + 482, + 597 + ], + "type": "inline_equation", + "content": "\\nabla_{x}p_{\\theta}(x)" + }, + { + "bbox": [ + 130, + 525, + 482, + 597 + ], + "type": "text", + "content": ", thereby smoothing the output logits across the input samples." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 616, + 230, + 630 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 616, + 230, + 630 + ], + "spans": [ + { + "bbox": [ + 132, + 616, + 230, + 630 + ], + "type": "text", + "content": "7 Experiments" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "content": "In this section, we perform extensive experiments to validate the efficacy of our regularization and the newly established correlation between the smooth" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 235, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 235, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 235, + 91, + 448, + 102 + ], + "type": "text", + "content": "Regulating Model Reliance on Non-Robust Features" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "text", + "content": "marginal density and model reliance on non-robust features. Specifically, we present measurement results for addressing feature leakage and mitigating spurious correlations. In addition to specific applications, further results on the robustness against perturbations in input pixels, gradients and density are presented. Additional details about the datasets and the models used in our experiments can be found in Supp. 19." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 211, + 449, + 224 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 211, + 449, + 224 + ], + "spans": [ + { + "bbox": [ + 131, + 211, + 449, + 224 + ], + "type": "text", + "content": "7.1 Efficacy against Feature Leakage and Adversarial Attacks" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 235, + 482, + 366 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 235, + 482, + 366 + ], + "spans": [ + { + "bbox": [ + 130, + 235, + 482, + 366 + ], + "type": "text", + "content": "In [2, 44], it is demonstrated that deep models end up assigning importance to irrelevant input features. Shah et al. [44] used BlockMNIST in their experiments, which is a synthetic dataset extended from MNIST [29]. To each MNIST sample, BlockMNIST attaches a null block (an irrelevant pattern) randomly at the top or bottom of the image, as shown in Fig. 3(a). Shah et al. [44] observed that the explaining tool InputGrad [47] attributes importance to both the informative number block and the uninformative null block in the standard trained model. This phenomenon is termed as feature leakage by the authors. Interestingly, the issue is mitigated in adversarially trained models. Since this dataset allows for a controlled robustness assessment, we first evaluate our method on BlockMNIST to analyze the model's reliance on irrelevant features." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 368, + 479, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 368, + 479, + 380 + ], + "spans": [ + { + "bbox": [ + 131, + 368, + 479, + 380 + ], + "type": "text", + "content": "Reproducibility and Quantitative Measurement of Feature Leakage." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 380, + 289, + 632 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 380, + 289, + 632 + ], + "spans": [ + { + "bbox": [ + 130, + 380, + 289, + 632 + ], + "type": "text", + "content": "Owing to the unreliability of InputGrad caused by model saturation [45], we employ Integrated Gradients (IG) [52], an axiomatic explanation tool, to faithfully reinvestigate the feature leakage phenomenon. In Fig. 3(b)-(c), we show that the attributions computed by IG can reproduce different leakage phenomena from the informative block region to the null block region on standard and adversarially trained models. Feature leakage is an important phenomenon in the context of model robustness. However, there is a lack of a quantitative metric in the current literature to quantify its extent. We use integrated gradients to define the metric " + }, + { + "bbox": [ + 130, + 380, + 289, + 632 + ], + "type": "inline_equation", + "content": "M_{leakage}" + }, + { + "bbox": [ + 130, + 380, + 289, + 632 + ], + "type": "text", + "content": " to address this gap. Mathematically," + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 298, + 387, + 477, + 454 + ], + "blocks": [ + { + "bbox": [ + 298, + 387, + 477, + 454 + ], + "lines": [ + { + "bbox": [ + 298, + 387, + 477, + 454 + ], + "spans": [ + { + "bbox": [ + 298, + 387, + 477, + 454 + ], + "type": "image", + "image_path": "5db74f99dcb8a59b9ca5a2787382f37ed2a0821c46b25f20de72903915974618.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 319, + 455, + 432, + 464 + ], + "lines": [ + { + "bbox": [ + 319, + 455, + 432, + 464 + ], + "spans": [ + { + "bbox": [ + 319, + 455, + 432, + 464 + ], + "type": "text", + "content": "(a) Input samples of BlockMNIST" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 295, + 468, + 476, + 512 + ], + "blocks": [ + { + "bbox": [ + 295, + 468, + 476, + 512 + ], + "lines": [ + { + "bbox": [ + 295, + 468, + 476, + 512 + ], + "spans": [ + { + "bbox": [ + 295, + 468, + 476, + 512 + ], + "type": "image", + "image_path": "07d33db9873d2df2789915643336c04869b09dda61165952f19dc25b5460305e.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 295, + 515, + 477, + 536 + ], + "lines": [ + { + "bbox": [ + 295, + 515, + 477, + 536 + ], + "spans": [ + { + "bbox": [ + 295, + 515, + 477, + 536 + ], + "type": "text", + "content": "(b) Integrated Gradients on (c) Integrated Gradients on standard trained model adversarially trained model" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 293, + 544, + 482, + 609 + ], + "lines": [ + { + "bbox": [ + 293, + 544, + 482, + 609 + ], + "spans": [ + { + "bbox": [ + 293, + 544, + 482, + 609 + ], + "type": "text", + "content": "Fig.3: BlockMNIST samples and feature leakage problem. (a) BlockMNIST randomly appends a null block at the top or bottom of MNIST samples. (b&c) Attribution maps are calculated by IG on the standard and adversarially trained models." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 181, + 640, + 481, + 669 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 181, + 640, + 481, + 669 + ], + "spans": [ + { + "bbox": [ + 181, + 640, + 481, + 669 + ], + "type": "interline_equation", + "content": "M _ {\\text {l e a k a g e}} = \\mathbb {E} _ {x _ {n r o b} \\sim D} | | x _ {n r o b} \\times \\int_ {\\alpha = 0} ^ {1} \\frac {\\partial f (\\alpha \\cdot x _ {n r o b} ; \\theta))}{\\partial x _ {n r o b}} d \\alpha | | _ {2}, \\tag {9}", + "image_path": "e0218e43029aa1d87d8ca4cffd67afb82ad18ed4b3b9e8fae87d77126323f144.jpg" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 222, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 222, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 222, + 102 + ], + "type": "text", + "content": "P. Yang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 164 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 164 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 164 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 116, + 482, + 164 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 130, + 116, + 482, + 164 + ], + "type": "text", + "content": " is the step from the absence to the presence of the features, and " + }, + { + "bbox": [ + 130, + 116, + 482, + 164 + ], + "type": "inline_equation", + "content": "x_{nrob}" + }, + { + "bbox": [ + 130, + 116, + 482, + 164 + ], + "type": "text", + "content": " specifies the non-robust features in the null block. Since the attributions of " + }, + { + "bbox": [ + 130, + 116, + 482, + 164 + ], + "type": "inline_equation", + "content": "x_{nrob}" + }, + { + "bbox": [ + 130, + 116, + 482, + 164 + ], + "type": "text", + "content": " for the model's prediction should ideally be zero, we use the " + }, + { + "bbox": [ + 130, + 116, + 482, + 164 + ], + "type": "inline_equation", + "content": "L_{2}" + }, + { + "bbox": [ + 130, + 116, + 482, + 164 + ], + "type": "text", + "content": " norm of attributions to quantify the extent of feature leakage." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 167, + 482, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 167, + 482, + 407 + ], + "spans": [ + { + "bbox": [ + 130, + 167, + 482, + 407 + ], + "type": "text", + "content": "Robustness against Feature Leakage. Table 1 presents the experimental results on the BlockMNIST dataset. We compare our method with other robust regularizations and techniques including InputGrad [36], IG-SUM [10], SoftPlus activations [11] and Hessian [12]. InputGrad and Hessian regularize the first-order and second-order gradients w.r.t. the input. Models trained with SoftPlus activations and Hessian regularization fail to suppress the leakage problem, which indicates that feature leakage is not caused by the geometry of the model output manifold or high curvature [11, 59]. InputGrad regularization demonstrates robustness against both " + }, + { + "bbox": [ + 130, + 167, + 482, + 407 + ], + "type": "inline_equation", + "content": "L_{2}" + }, + { + "bbox": [ + 130, + 167, + 482, + 407 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 167, + 482, + 407 + ], + "type": "inline_equation", + "content": "L_{\\infty}" + }, + { + "bbox": [ + 130, + 167, + 482, + 407 + ], + "type": "text", + "content": " adversarial attacks, yet it still fails to address the leakage problem. The result aligns well with Remark 3, which highlights the allowance of non-robust features across different classes in InputGrad regularization. In addition, the use of IG in IG-Norm regularization that accumulates input gradients as a regularization term results in behavior similar to InputGrad regularization, leading to compromised robustness. These results further reveal that adversarial robustness is not a sufficient condition for suppressing feature leakage. Our method demonstrates a considerable improvement over other techniques for feature leakage, while also maintaining superiority in adversarial robustness. In Tab. 1, we use " + }, + { + "bbox": [ + 130, + 167, + 482, + 407 + ], + "type": "inline_equation", + "content": "L_{2}" + }, + { + "bbox": [ + 130, + 167, + 482, + 407 + ], + "type": "text", + "content": " norm for the compared regularization terms for fair benchmarking. We show in Supp. 10 that optimal norm exploration can yield an even more favorable trade-off for our technique for model robustness." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 409, + 483, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 409, + 483, + 506 + ], + "spans": [ + { + "bbox": [ + 130, + 409, + 483, + 506 + ], + "type": "text", + "content": "Feature Leakage in Adversarily Robust Models. A FGSM adversarily trained model [17] augments the training samples by adversarial examples " + }, + { + "bbox": [ + 130, + 409, + 483, + 506 + ], + "type": "inline_equation", + "content": "x + \\epsilon \\cdot sign(\\nabla_x f_i(x; \\theta))" + }, + { + "bbox": [ + 130, + 409, + 483, + 506 + ], + "type": "text", + "content": ". Notably minimizing the loss of the perturbed input " + }, + { + "bbox": [ + 130, + 409, + 483, + 506 + ], + "type": "inline_equation", + "content": "x + \\epsilon \\cdot sign(\\nabla_x f_i(x; \\theta))" + }, + { + "bbox": [ + 130, + 409, + 483, + 506 + ], + "type": "text", + "content": " is similar to the InputGrad regularization. Thus, training with FGSM is still limited in its ability to suppress the leakage problem. In contrast, PGD attack [31] weakens the effect of the condition " + }, + { + "bbox": [ + 130, + 409, + 483, + 506 + ], + "type": "inline_equation", + "content": "y = i" + }, + { + "bbox": [ + 130, + 409, + 483, + 506 + ], + "type": "text", + "content": " by iteratively searching for the perturbations from a random starting point. This process leads to a substantial enhancement in suppressing feature leakage, see Tab. 1." + } + ] + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 132, + 567, + 485, + 677 + ], + "blocks": [ + { + "bbox": [ + 131, + 524, + 483, + 557 + ], + "lines": [ + { + "bbox": [ + 131, + 524, + 483, + 557 + ], + "spans": [ + { + "bbox": [ + 131, + 524, + 483, + 557 + ], + "type": "text", + "content": "Table 1: Results on BlockMNIST. Feature leakage, standard accuracy and adversarial accuracy under " + }, + { + "bbox": [ + 131, + 524, + 483, + 557 + ], + "type": "inline_equation", + "content": "L_{2}" + }, + { + "bbox": [ + 131, + 524, + 483, + 557 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 524, + 483, + 557 + ], + "type": "inline_equation", + "content": "L_{\\infty}" + }, + { + "bbox": [ + 131, + 524, + 483, + 557 + ], + "type": "text", + "content": " PGD-20 attacks are reported. ST & AT: Standard and Adversarial Training." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 132, + 567, + 485, + 677 + ], + "lines": [ + { + "bbox": [ + 132, + 567, + 485, + 677 + ], + "spans": [ + { + "bbox": [ + 132, + 567, + 485, + 677 + ], + "type": "table", + "html": "
MethodFeature Leakage ↓PGD-20 (L2) ↑PGD-20 (L∞) ↑Accu. ↑
AT (FGSM)3.32487.480.0099.02
AT (PGD)2.31392.7528.0598.97
ST3.65773.570.0099.12
+ SoftPlus Acts.3.53367.950.0298.52
+ Hessian Reg.4.25880.060.0098.48
+ InputGrad Reg.3.46183.4621.1494.56
+ IG-SUM Reg.3.49782.1122.7392.87
+ Our Reg.2.25985.4129.3693.05
", + "image_path": "4ff636f4d5661e11fb8722e8ae8cbae85fb390a6e7cbc10c46d13fb463a9dbe2.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 235, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 235, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 235, + 91, + 447, + 102 + ], + "type": "text", + "content": "Regulating Model Reliance on Non-Robust Features" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 132, + 115, + 216, + 178 + ], + "blocks": [ + { + "bbox": [ + 132, + 115, + 216, + 178 + ], + "lines": [ + { + "bbox": [ + 132, + 115, + 216, + 178 + ], + "spans": [ + { + "bbox": [ + 132, + 115, + 216, + 178 + ], + "type": "image", + "image_path": "7e1934dfae16c366802ad5283e745e806f06518ced504975b62d9a4042b22a52.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 136, + 179, + 211, + 188 + ], + "lines": [ + { + "bbox": [ + 136, + 179, + 211, + 188 + ], + "spans": [ + { + "bbox": [ + 136, + 179, + 211, + 188 + ], + "type": "text", + "content": "(a) Feature Leakage." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 221, + 114, + 304, + 178 + ], + "blocks": [ + { + "bbox": [ + 221, + 114, + 304, + 178 + ], + "lines": [ + { + "bbox": [ + 221, + 114, + 304, + 178 + ], + "spans": [ + { + "bbox": [ + 221, + 114, + 304, + 178 + ], + "type": "image", + "image_path": "165ac84da9f9c1381686ae165c5c1b522eb2606fba7946221a49bb74a8cefc11.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 234, + 180, + 291, + 189 + ], + "lines": [ + { + "bbox": [ + 234, + 180, + 291, + 189 + ], + "spans": [ + { + "bbox": [ + 234, + 180, + 291, + 189 + ], + "type": "text", + "content": "(b) " + }, + { + "bbox": [ + 234, + 180, + 291, + 189 + ], + "type": "inline_equation", + "content": "L_{\\infty}" + }, + { + "bbox": [ + 234, + 180, + 291, + 189 + ], + "type": "text", + "content": " Attack." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 131, + 198, + 480, + 232 + ], + "lines": [ + { + "bbox": [ + 131, + 198, + 480, + 232 + ], + "spans": [ + { + "bbox": [ + 131, + 198, + 480, + 232 + ], + "type": "text", + "content": "Fig. 4: Performance comparison between our method and InputGrad regularization under varying regularization coefficient for (a) Feature leakage, (b)-(c) Adversarial Accuracy under " + }, + { + "bbox": [ + 131, + 198, + 480, + 232 + ], + "type": "inline_equation", + "content": "L_{\\infty}" + }, + { + "bbox": [ + 131, + 198, + 480, + 232 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 198, + 480, + 232 + ], + "type": "inline_equation", + "content": "L_{2}" + }, + { + "bbox": [ + 131, + 198, + 480, + 232 + ], + "type": "text", + "content": " PGD-20 attacks, and (d) Accuracy." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 310, + 114, + 392, + 178 + ], + "blocks": [ + { + "bbox": [ + 310, + 114, + 392, + 178 + ], + "lines": [ + { + "bbox": [ + 310, + 114, + 392, + 178 + ], + "spans": [ + { + "bbox": [ + 310, + 114, + 392, + 178 + ], + "type": "image", + "image_path": "b8cd40a4f5e1ade19ebe033e88077d01c8263368860ddbdbbb9cc01086363987.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 325, + 180, + 377, + 190 + ], + "lines": [ + { + "bbox": [ + 325, + 180, + 377, + 190 + ], + "spans": [ + { + "bbox": [ + 325, + 180, + 377, + 190 + ], + "type": "text", + "content": "(c) " + }, + { + "bbox": [ + 325, + 180, + 377, + 190 + ], + "type": "inline_equation", + "content": "L_{2}" + }, + { + "bbox": [ + 325, + 180, + 377, + 190 + ], + "type": "text", + "content": " Attack." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 399, + 114, + 481, + 178 + ], + "blocks": [ + { + "bbox": [ + 399, + 114, + 481, + 178 + ], + "lines": [ + { + "bbox": [ + 399, + 114, + 481, + 178 + ], + "spans": [ + { + "bbox": [ + 399, + 114, + 481, + 178 + ], + "type": "image", + "image_path": "5a9d1441cb3f3bbf7e71b1dbb4d8e2057bd84eeffb98bd7115c45bdb03f51e22.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 415, + 180, + 465, + 189 + ], + "lines": [ + { + "bbox": [ + 415, + 180, + 465, + 189 + ], + "spans": [ + { + "bbox": [ + 415, + 180, + 465, + 189 + ], + "type": "text", + "content": "(d) Accuracy." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 242, + 480, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 242, + 480, + 350 + ], + "spans": [ + { + "bbox": [ + 130, + 242, + 480, + 350 + ], + "type": "text", + "content": "Magnitude of Coefficient for Regularization. Figs. 4(a)-(d) present a comparison of results for feature leakage and adversarial accuracy under PGD attacks, as well as the standard accuracy across varying magnitudes for the regularization strength. The results affirm that our method effectively regulates feature leakage by imposing a penalty on non-robust features. Moreover, our regularization enables the model to defend against both " + }, + { + "bbox": [ + 130, + 242, + 480, + 350 + ], + "type": "inline_equation", + "content": "L_{2}" + }, + { + "bbox": [ + 130, + 242, + 480, + 350 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 242, + 480, + 350 + ], + "type": "inline_equation", + "content": "L_{\\infty}" + }, + { + "bbox": [ + 130, + 242, + 480, + 350 + ], + "type": "text", + "content": " attacks while maintaining high accuracy, showing an outstanding trade-off across four criteria. More adversarial robustness comparisons on CIFAR dataset [28] are reported in Supp. 11." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 369, + 326, + 380 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 369, + 326, + 380 + ], + "spans": [ + { + "bbox": [ + 132, + 369, + 326, + 380 + ], + "type": "text", + "content": "7.2 Efficacy for Spurious Correlation" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 390, + 298, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 390, + 298, + 557 + ], + "spans": [ + { + "bbox": [ + 130, + 390, + 298, + 557 + ], + "type": "text", + "content": "Recent research has highlighted the susceptibility of neural models in learning spurious correlations that enhance performance on a given data but fail to generalize [2,8,39]. For instance, in CelebA-Hair dataset [30], which commonly consists of samples containing female celebrities with blond hair and male celebrities with dark hair, models heavily rely on the spuriously correlated gender feature to predict the target hair color [19, 39]. Consequently, accuracy tends to be lower for the samples containing" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 558, + 323, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 558, + 323, + 569 + ], + "spans": [ + { + "bbox": [ + 132, + 558, + 323, + 569 + ], + "type": "text", + "content": "male celebrities with blond hair, see Fig. 5." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 570, + 480, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 480, + 665 + ], + "type": "text", + "content": "To mitigate spurious correlations, distributional robust optimization (DRO) techniques have been proposed to re-weight the training loss of input samples from different groups [23,39]. In our regularization, an additional penalty is imposed to penalize the model's reliance on these spuriously correlated features because of their inconsistent attributions. This encourages the use of robust features while suppressing the model's reliance on spuriously correlated features. Reliable quantification of the model's robustness on natural images for spurious correlation suppression is an unresolved issue in the literature. We employ" + } + ] + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 306, + 407, + 360, + 460 + ], + "blocks": [ + { + "bbox": [ + 307, + 398, + 392, + 407 + ], + "lines": [ + { + "bbox": [ + 307, + 398, + 392, + 407 + ], + "spans": [ + { + "bbox": [ + 307, + 398, + 392, + 407 + ], + "type": "text", + "content": "Common training samples:" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 306, + 407, + 360, + 460 + ], + "lines": [ + { + "bbox": [ + 306, + 407, + 360, + 460 + ], + "spans": [ + { + "bbox": [ + 306, + 407, + 360, + 460 + ], + "type": "image", + "image_path": "00aa977129fc7b52da8eb0c90bfa240334d443e45459bb1086ea76ef90c2ea3a.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 306, + 460, + 359, + 468 + ], + "lines": [ + { + "bbox": [ + 306, + 460, + 359, + 468 + ], + "spans": [ + { + "bbox": [ + 306, + 460, + 359, + 468 + ], + "type": "text", + "content": "Label: Blond hair" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 306, + 468, + 362, + 476 + ], + "lines": [ + { + "bbox": [ + 306, + 468, + 362, + 476 + ], + "spans": [ + { + "bbox": [ + 306, + 468, + 362, + 476 + ], + "type": "text", + "content": "Spur. Feat.: Female" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 307, + 477, + 356, + 484 + ], + "lines": [ + { + "bbox": [ + 307, + 477, + 356, + 484 + ], + "spans": [ + { + "bbox": [ + 307, + 477, + 356, + 484 + ], + "type": "text", + "content": "Accuracy: " + }, + { + "bbox": [ + 307, + 477, + 356, + 484 + ], + "type": "inline_equation", + "content": "92.1\\%" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 364, + 407, + 416, + 460 + ], + "blocks": [ + { + "bbox": [ + 364, + 407, + 416, + 460 + ], + "lines": [ + { + "bbox": [ + 364, + 407, + 416, + 460 + ], + "spans": [ + { + "bbox": [ + 364, + 407, + 416, + 460 + ], + "type": "image", + "image_path": "449779662b8899761d692b6e86ee95d564b3e117dfd4d58035d352ff5eeba19d.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 365, + 460, + 414, + 468 + ], + "lines": [ + { + "bbox": [ + 365, + 460, + 414, + 468 + ], + "spans": [ + { + "bbox": [ + 365, + 460, + 414, + 468 + ], + "type": "text", + "content": "Label: Dark hair" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 365, + 468, + 414, + 476 + ], + "lines": [ + { + "bbox": [ + 365, + 468, + 414, + 476 + ], + "spans": [ + { + "bbox": [ + 365, + 468, + 414, + 476 + ], + "type": "text", + "content": "Spur. Feat.: Male" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 365, + 477, + 414, + 484 + ], + "lines": [ + { + "bbox": [ + 365, + 477, + 414, + 484 + ], + "spans": [ + { + "bbox": [ + 365, + 477, + 414, + 484 + ], + "type": "text", + "content": "Accuracy: " + }, + { + "bbox": [ + 365, + 477, + 414, + 484 + ], + "type": "inline_equation", + "content": "99.7\\%" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 427, + 407, + 479, + 459 + ], + "blocks": [ + { + "bbox": [ + 427, + 398, + 468, + 407 + ], + "lines": [ + { + "bbox": [ + 427, + 398, + 468, + 407 + ], + "spans": [ + { + "bbox": [ + 427, + 398, + 468, + 407 + ], + "type": "text", + "content": "Test sample:" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 427, + 407, + 479, + 459 + ], + "lines": [ + { + "bbox": [ + 427, + 407, + 479, + 459 + ], + "spans": [ + { + "bbox": [ + 427, + 407, + 479, + 459 + ], + "type": "image", + "image_path": "37c14f47adcb222d4da384c9736006a1eeb053e0319fd902d3f904720f19592c.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 428, + 460, + 479, + 468 + ], + "lines": [ + { + "bbox": [ + 428, + 460, + 479, + 468 + ], + "spans": [ + { + "bbox": [ + 428, + 460, + 479, + 468 + ], + "type": "text", + "content": "Label: Blond hair" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 428, + 468, + 477, + 476 + ], + "lines": [ + { + "bbox": [ + 428, + 468, + 477, + 476 + ], + "spans": [ + { + "bbox": [ + 428, + 468, + 477, + 476 + ], + "type": "text", + "content": "Spur.Feat.:Male" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 428, + 477, + 477, + 484 + ], + "lines": [ + { + "bbox": [ + 428, + 477, + 477, + 484 + ], + "spans": [ + { + "bbox": [ + 428, + 477, + 477, + 484 + ], + "type": "text", + "content": "Accuracy: " + }, + { + "bbox": [ + 428, + 477, + 477, + 484 + ], + "type": "inline_equation", + "content": "23.6\\%" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_caption" + } + ], + "index": 26 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 222, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 222, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 222, + 102 + ], + "type": "text", + "content": "P. Yang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 132, + 157, + 482, + 261 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "type": "text", + "content": "Table 2: Worst-group accuracy and overall accuracy comparisons between Vanilla ResNet-34, group DRO, Score-Matching, InputGrad and our regularization on CelebA-Hair and Waterbirds datasets." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 132, + 157, + 482, + 261 + ], + "lines": [ + { + "bbox": [ + 132, + 157, + 482, + 261 + ], + "spans": [ + { + "bbox": [ + 132, + 157, + 482, + 261 + ], + "type": "table", + "html": "
MethodWorst-Group Accuracy (%)Overall Accuracy (%)
CelebA-HairWaterbirdsCelebA-HairWaterbirds
Vanilla Model49.90 ± 8.6962.90 ± 0.1094.90 ± 0.3987.70 ± 0.08
Group DRO59.44 ± 5.9863.60 ± 0.1794.96 ± 0.2187.60 ± 0.05
Score-Matching Reg.59.78 ± 7.5658.19 ± 1.5593.46 ± 0.7185.64 ± 0.38
InputGrad Reg.82.66 ± 3.6358.18 ± 1.2292.12 ± 2.1885.50 ± 0.27
Our Reg.85.62 ± 5.3663.78 ± 2.8392.30 ± 1.3886.48 ± 0.38
Group DRO + Ours82.98 ± 4.6973.82 ± 2.2793.62 ± 0.7490.52 ± 0.17
", + "image_path": "b81f59740d7fb43ffc6e795bfa460d39853adb7e490bd80ea790ee40f0bb855c.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 279, + 481, + 339 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 279, + 481, + 339 + ], + "spans": [ + { + "bbox": [ + 130, + 279, + 481, + 339 + ], + "type": "text", + "content": "attribute maps [52] and insertion game [34] to demonstrate the effectiveness of our regularization in suppressing the use of the spuriously correlated gender feature and promoting the use of the robust hair feature, as shown in Fig. 1(b). This superiority leads to performance improvements on worst-case samples in the model trained with our regularization." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 340, + 482, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 340, + 482, + 533 + ], + "spans": [ + { + "bbox": [ + 130, + 340, + 482, + 533 + ], + "type": "text", + "content": "Table 2 presents comparison of accuracy on worst-group samples and overall samples from the CelebA-Hair and Waterbirds datasets [30,39]. The Waterbirds dataset consists of synthetic bird images from CUB-200-2011 [53] and Places [60] datasets, incorporating spurious background features, such as land and water scenes, to confuse true labels of bird categories. We compare the proposed regularization method with InputGrad and Score-Matching regularizations [49], as well as Group DRO [39]. Score-Matching regularization is proposed to enhance the interpretability of the model by improving the alignment of implicit density models. Experimental results clearly demonstrate the effectiveness of our method in enhancing worst-group accuracy while maintaining overall sample accuracy. Furthermore, the performance gains can be further enhanced by incorporating our regularization technique into Group DRO. This enhancement highlights the efficacy of our regularization in terms of its practicality. More results using attribution maps and the insertion game on both CelebA-Hair and Waterbirds datasets are provided in Supp. 17. Moreover, out-of-distribution detection [21] is also performed on CIFAR [28] and SVHN [33] in Supp. 14." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 549, + 467, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 549, + 467, + 562 + ], + "spans": [ + { + "bbox": [ + 132, + 549, + 467, + 562 + ], + "type": "text", + "content": "7.3 Efficacy against Pixels, Gradients and Density Perturbations" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": "We first employ pixel perturbation [41, 55] to quantitatively compare the robustness of different models following Srinivas and Fleuret [49] who iteratively removed the most important input pixels identified by attribution maps for model robustness evaluation. Robust models are expected to exhibit increased sensitivity when removing the most important pixels and decreased sensitivity when removing the least important ones. We assess the difference in fractional output logit change between the images with the top and bottom " + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "inline_equation", + "content": "k\\%" + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": " most salient pixels using SmoothGrad [48] on ResNet-18 [20] trained on CIFAR-10" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 235, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 235, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 235, + 91, + 447, + 102 + ], + "type": "text", + "content": "Regulating Model Reliance on Non-Robust Features" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 114, + 218, + 184 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 218, + 184 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 218, + 184 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 218, + 184 + ], + "type": "image", + "image_path": "4a556c4dc6a686a71038b588c4cf2cb3128f7e07eb4dc4fe0adb53238a35c9db.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 137, + 186, + 304, + 196 + ], + "lines": [ + { + "bbox": [ + 137, + 186, + 304, + 196 + ], + "spans": [ + { + "bbox": [ + 137, + 186, + 304, + 196 + ], + "type": "text", + "content": "(a) Pixel Perturbation on CIFAR-10 and CIFAR-100." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 219, + 114, + 304, + 184 + ], + "blocks": [ + { + "bbox": [ + 219, + 114, + 304, + 184 + ], + "lines": [ + { + "bbox": [ + 219, + 114, + 304, + 184 + ], + "spans": [ + { + "bbox": [ + 219, + 114, + 304, + 184 + ], + "type": "image", + "image_path": "d063c23504d6e04835e0415096766816b2189cd8b350fa9825a055ef65593ab6.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 205, + 480, + 248 + ], + "lines": [ + { + "bbox": [ + 132, + 205, + 480, + 248 + ], + "spans": [ + { + "bbox": [ + 132, + 205, + 480, + 248 + ], + "type": "text", + "content": "Fig. 6: Robustness comparison. (a) Pixel perturbation experimental results on ResNet-18 trained on CIFAR-10 and CIFAR-100. Higher curves indicate better results. (b) Robustness of relative gradients and absolute density on CIFAR-10. Lower curves indicate better results." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 304, + 114, + 392, + 184 + ], + "blocks": [ + { + "bbox": [ + 304, + 114, + 392, + 184 + ], + "lines": [ + { + "bbox": [ + 304, + 114, + 392, + 184 + ], + "spans": [ + { + "bbox": [ + 304, + 114, + 392, + 184 + ], + "type": "image", + "image_path": "57ccbb079a2189b48c2834c2a0f237da9f841be08faf3239d05910a3d8139e79.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 336, + 186, + 466, + 197 + ], + "lines": [ + { + "bbox": [ + 336, + 186, + 466, + 197 + ], + "spans": [ + { + "bbox": [ + 336, + 186, + 466, + 197 + ], + "type": "text", + "content": "(b) Robustness of Gradients and Density." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 392, + 114, + 481, + 184 + ], + "blocks": [ + { + "bbox": [ + 392, + 114, + 481, + 184 + ], + "lines": [ + { + "bbox": [ + 392, + 114, + 481, + 184 + ], + "spans": [ + { + "bbox": [ + 392, + 114, + 481, + 184 + ], + "type": "image", + "image_path": "1c87bd4c79667abbf7b60f57532a6bfabf57e32c68bd9dd48ef483fd5d4367ae.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 266, + 480, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 266, + 480, + 361 + ], + "spans": [ + { + "bbox": [ + 130, + 266, + 480, + 361 + ], + "type": "text", + "content": "and CIFAR-100 [28], as depicted in Fig. 6(a). Given the recognized challenge of reference image ambiguity in IG for explaining natural images [14, 51], we opt to use SmoothGrad for explaining predictions of images in the CIFAR dataset. SmoothGrad has demonstrated to outperform IG in such conditions [22, 50]. Recognizing their respective strengths, we use different attribution methods in appropriate conditions for reliable results. It can be observed that models trained with our regularization significantly outperform different robust regularizations including Score-Matching, InputGrad and CURE regularizations [32]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 362, + 481, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 362, + 481, + 495 + ], + "spans": [ + { + "bbox": [ + 130, + 362, + 481, + 495 + ], + "type": "text", + "content": "We also test the robustness of the relative input gradients " + }, + { + "bbox": [ + 130, + 362, + 481, + 495 + ], + "type": "inline_equation", + "content": "||\\nabla_{x}f(x + \\delta) - \\nabla_{x}f(x)||_{2} / ||\\nabla_{x}f(x)||_{2}" + }, + { + "bbox": [ + 130, + 362, + 481, + 495 + ], + "type": "text", + "content": " and the absolute density through " + }, + { + "bbox": [ + 130, + 362, + 481, + 495 + ], + "type": "inline_equation", + "content": "\\sum_{i = 1}^{C}e^{f_i(x + \\delta) - f_i(x)}" + }, + { + "bbox": [ + 130, + 362, + 481, + 495 + ], + "type": "text", + "content": " on the input with Gaussian noise " + }, + { + "bbox": [ + 130, + 362, + 481, + 495 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 130, + 362, + 481, + 495 + ], + "type": "text", + "content": " in increasing standard deviation on CIFAR-10 dataset, as shown in Fig. 6(b). We can observe that our regularization leads to competitive robustness for the relative gradient in comparison with regularizing the Hessian norm in CURE. Moreover, our regularization naturally leads to density robustness, which is associated with a strong generative ability of the models [49]. More tests on CIFAR-100 are provided in Supp. 13. Supp. 16 provides the visualizations calculated by activation optimization. These results affirm that our regularization improves both the discriminative and generative abilities of models." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 517, + 220, + 530 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 517, + 220, + 530 + ], + "spans": [ + { + "bbox": [ + 132, + 517, + 220, + 530 + ], + "type": "text", + "content": "8 Conclusion" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": "In this paper, we define robust and non-robust features from a feature attribution perspective, and establish a correlation between the smoothness of input marginal density and model reliance on non-robust features. This connection motivates us to propose a regularization that targets the gradients of the marginal density, aiming to regulate the reliance on non-robust features. Extensive experiments demonstrate the effectiveness of our regularization in boosting model robustness across a wide range of applications. We emphasize that our approach does not advocate for the complete removal of model reliance on non-robust features, but instead seeks to achieve a balance between model performance and robustness through appropriate regularization strength." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 222, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 222, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 222, + 102 + ], + "type": "text", + "content": "P. Yang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 240, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 240, + 129 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 240, + 129 + ], + "type": "text", + "content": "Acknowledgments" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 140, + 482, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 140, + 482, + 224 + ], + "spans": [ + { + "bbox": [ + 130, + 140, + 482, + 224 + ], + "type": "text", + "content": "This research was supported by ARC Discovery Grant 190102443 and by Google Research under Google Research Scholar Program. Dr. Naveed Akhtar is a recipient of the Australian Research Council Discovery Early Career Researcher Award (project number DE230101058) funded by the Australian Government. Professor Ajmal Mian is the recipient of an Australian Research Council Future Fellowship Award (project number FT210100268) funded by the Australian Government." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 243, + 197, + 255 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 243, + 197, + 255 + ], + "spans": [ + { + "bbox": [ + 133, + 243, + 197, + 255 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 134, + 269, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 138, + 269, + 480, + 301 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 269, + 480, + 301 + ], + "spans": [ + { + "bbox": [ + 138, + 269, + 480, + 301 + ], + "type": "text", + "content": "1. Adebayo, J., Muelly, M., Abelson, H., Kim, B.: Post hoc explanations may be ineffective for detecting unknown spurious correlation. In: International Conference on Learning Representations, ICLR (2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 303, + 480, + 323 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 303, + 480, + 323 + ], + "spans": [ + { + "bbox": [ + 138, + 303, + 480, + 323 + ], + "type": "text", + "content": "2. Adebayo, J., Muelly, M., Liccardi, I., Kim, B.: Debugging tests for model explanations. In: Advances in Neural Information Processing Systems, NeurIPS (2020)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 325, + 480, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 325, + 480, + 357 + ], + "spans": [ + { + "bbox": [ + 138, + 325, + 480, + 357 + ], + "type": "text", + "content": "3. Ancona, M., Ceolini, E., Öztireli, C., Gross, M.: Towards better understanding of gradient-based attribution methods for deep neural networks. In: International Conference on Learning Representations, ICLR (2018)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 357, + 480, + 378 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 357, + 480, + 378 + ], + "spans": [ + { + "bbox": [ + 138, + 357, + 480, + 378 + ], + "type": "text", + "content": "4. Anil, C., Lucas, J., Grosse, R.: Sorting out lipschitz function approximation. In: International Conference on Machine Learning, ICML (2019)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 380, + 480, + 411 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 380, + 480, + 411 + ], + "spans": [ + { + "bbox": [ + 138, + 380, + 480, + 411 + ], + "type": "text", + "content": "5. Blanchard, P., Higham, D.J., Higham, N.J.: Accurately computing the log-sum-exp and softmax functions. IMA Journal of Numerical Analysis 41(4), 2311-2330 (2021)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 413, + 480, + 444 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 413, + 480, + 444 + ], + "spans": [ + { + "bbox": [ + 138, + 413, + 480, + 444 + ], + "type": "text", + "content": "6. Bridle, J.: Training stochastic model recognition algorithms as networks can lead to maximum mutual information estimation of parameters. Advances in Neural Information Processing Systems, NeurIPS (1989)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 446, + 480, + 478 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 446, + 480, + 478 + ], + "spans": [ + { + "bbox": [ + 138, + 446, + 480, + 478 + ], + "type": "text", + "content": "7. Bridle, J.S.: Probabilistic interpretation of feedforward classification network outputs, with relationships to statistical pattern recognition. In: Neurocomputing: Algorithms, Architectures and Applications, pp. 227-236. Springer (1990)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 479, + 480, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 479, + 480, + 510 + ], + "spans": [ + { + "bbox": [ + 138, + 479, + 480, + 510 + ], + "type": "text", + "content": "8. Buolamwini, J., Gebru, T.: Gender shades: Intersectional accuracy disparities in commercial gender classification. In: ACM Conference on Fairness, Accountability, and Transparency, FAccT (2018)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 512, + 480, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 512, + 480, + 533 + ], + "spans": [ + { + "bbox": [ + 138, + 512, + 480, + 533 + ], + "type": "text", + "content": "9. Carlini, N., Wagner, D.: Towards evaluating the robustness of neural networks. In: IEEE Symposium on Security and Privacy, EuroS&P (2017)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 135, + 534, + 480, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 534, + 480, + 555 + ], + "spans": [ + { + "bbox": [ + 135, + 534, + 480, + 555 + ], + "type": "text", + "content": "10. Chen, J., Wu, X., Rastogi, V., Liang, Y., Jha, S.: Robust attribution regularization. In: Advances in Neural Information Processing Systems, NeurIPS (2019)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 134, + 556, + 480, + 587 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 556, + 480, + 587 + ], + "spans": [ + { + "bbox": [ + 134, + 556, + 480, + 587 + ], + "type": "text", + "content": "11. Dombrowski, A.K., Alber, M., Anders, C., Ackermann, M., Müller, K.R., Kessel, P.: Explanations can be manipulated and geometry is to blame. In: Advances in Neural Information Processing Systems, NeurIPS (2019)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 134, + 589, + 480, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 589, + 480, + 609 + ], + "spans": [ + { + "bbox": [ + 134, + 589, + 480, + 609 + ], + "type": "text", + "content": "12. Dombrowski, A.K., Anders, C.J., Müller, K.R., Kessel, P.: Towards robust explanations for deep neural networks. Pattern Recognition 121, 108194 (2022)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 134, + 611, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 611, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 134, + 611, + 480, + 632 + ], + "type": "text", + "content": "13. Drucker, H., LeCun, Y.: Improving generalization performance using double backpropagation. IEEE Trans. Neural Networks (1992)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 134, + 633, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 633, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 134, + 633, + 480, + 665 + ], + "type": "text", + "content": "14. Erion, G.G., Janizek, J.D., Sturmfels, P., Lundberg, S.M., Lee, S.: Improving performance of deep learning models with axiomatic attribution priors and expected gradients. Nature Machine Intelligence (2021)" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 235, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 235, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 235, + 91, + 447, + 102 + ], + "type": "text", + "content": "Regulating Model Reliance on Non-Robust Features" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 117, + 481, + 666 + ], + "type": "list", + "angle": 0, + "index": 22, + "blocks": [ + { + "bbox": [ + 133, + 117, + 426, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 117, + 426, + 128 + ], + "spans": [ + { + "bbox": [ + 133, + 117, + 426, + 128 + ], + "type": "text", + "content": "15. Etmann, C.: A closer look at double backpropagation. CoRR (2019)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 128, + 481, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 128, + 481, + 150 + ], + "spans": [ + { + "bbox": [ + 133, + 128, + 481, + 150 + ], + "type": "text", + "content": "16. Fong, R.C., Vedaldi, A.: Interpretable explanations of black boxes by meaningful perturbation. In: International Conference on Computer Vision, ICCV (2017)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 150, + 481, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 150, + 481, + 172 + ], + "spans": [ + { + "bbox": [ + 133, + 150, + 481, + 172 + ], + "type": "text", + "content": "17. Goodfellow, I.J., Shlens, J., Szegedy, C.: Explaining and harnessing adversarial examples. In: International Conference on Learning Representations, ICLR (2015)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 172, + 481, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 172, + 481, + 205 + ], + "spans": [ + { + "bbox": [ + 133, + 172, + 481, + 205 + ], + "type": "text", + "content": "18. Grathwohl, W., Wang, K.C., Jacobsen, J.H., Duvenaud, D., Norouzi, M., Swersky, K.: Your classifier is secretly an energy based model and you should treat it like one. In: International Conference on Learning Representations, ICLR (2020)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 133, + 205, + 481, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 205, + 481, + 237 + ], + "spans": [ + { + "bbox": [ + 133, + 205, + 481, + 237 + ], + "type": "text", + "content": "19. Hashimoto, T., Srivastava, M., Namkoong, H., Liang, P.: Fairness without demographics in repeated loss minimization. In: International Conference on Machine Learning, ICML (2018)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 237, + 481, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 237, + 481, + 270 + ], + "spans": [ + { + "bbox": [ + 132, + 237, + 481, + 270 + ], + "type": "text", + "content": "20. He, K., Zhang, X., Ren, S., Sun, J.: Identity mappings in deep residual networks. In: European Conference on Computer Vision, European Conference on Computer Vision, ECCV (2016)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 270, + 481, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 270, + 481, + 293 + ], + "spans": [ + { + "bbox": [ + 132, + 270, + 481, + 293 + ], + "type": "text", + "content": "21. Hendrycks, D., Gimpel, K.: A baseline for detecting misclassified and out-of-distribution examples in neural networks. CoRR (2016)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 293, + 481, + 325 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 293, + 481, + 325 + ], + "spans": [ + { + "bbox": [ + 132, + 293, + 481, + 325 + ], + "type": "text", + "content": "22. Hooker, S., Erhan, D., Kindermans, P., Kim, B.: A benchmark for interpretability methods in deep neural networks. In: Advances in Neural Information Processing Systems, NeurIPS (2019)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 325, + 481, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 325, + 481, + 358 + ], + "spans": [ + { + "bbox": [ + 132, + 325, + 481, + 358 + ], + "type": "text", + "content": "23. Hu, W., Niu, G., Sato, I., Sugiyama, M.: Does distributionally robust supervised learning give robust classifiers? In: International Conference on Machine Learning, ICML (2018)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 358, + 481, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 358, + 481, + 392 + ], + "spans": [ + { + "bbox": [ + 132, + 358, + 481, + 392 + ], + "type": "text", + "content": "24. Ilyas, A., Santurkar, S., Tsipras, D., Engstrom, L., Tran, B., Madry, A.: Adversarial examples are not bugs, they are features. Advances in neural information processing systems 32 (2019)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 392, + 481, + 424 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 392, + 481, + 424 + ], + "spans": [ + { + "bbox": [ + 132, + 392, + 481, + 424 + ], + "type": "text", + "content": "25. Ioffe, S., Szegedy, C.: Batch normalization: Accelerating deep network training by reducing internal covariate shift. In: International Conference on Machine Learning, ICML (2015)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 424, + 481, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 424, + 481, + 456 + ], + "spans": [ + { + "bbox": [ + 132, + 424, + 481, + 456 + ], + "type": "text", + "content": "26. Jiang, J., Wen, Z., Mansoor, A., Mian, A.: Efficient hyperparameter optimization with adaptive fidelity identification. In: IEEE Conference on Computer Vision and Pattern Recognition (2024)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 456, + 481, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 456, + 481, + 479 + ], + "spans": [ + { + "bbox": [ + 132, + 456, + 481, + 479 + ], + "type": "text", + "content": "27. Jiang, J., Wen, Z., Yang, P., Mansoor, A., Mian, A.: Fast-pgm: Fast probabilistic graphical model learning and inference. arXiv preprint arXiv:2405.15605 (2024)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 479, + 481, + 501 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 479, + 481, + 501 + ], + "spans": [ + { + "bbox": [ + 132, + 479, + 481, + 501 + ], + "type": "text", + "content": "28. Krizhevsky, A., Hinton, G., et al.: Learning multiple layers of features from tiny images. Technical Report, University of Toronto (2009)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 501, + 481, + 523 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 501, + 481, + 523 + ], + "spans": [ + { + "bbox": [ + 132, + 501, + 481, + 523 + ], + "type": "text", + "content": "29. LeCun, Y., Bottou, L., Bengio, Y., Haffner, P.: Gradient-based learning applied to document recognition. Proc. IEEE (1998)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 523, + 481, + 545 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 523, + 481, + 545 + ], + "spans": [ + { + "bbox": [ + 132, + 523, + 481, + 545 + ], + "type": "text", + "content": "30. Liu, Z., Luo, P., Wang, X., Tang, X.: Deep learning face attributes in the wild. In: International Conference on Computer Vision, ICCV (2015)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 545, + 481, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 545, + 481, + 578 + ], + "spans": [ + { + "bbox": [ + 132, + 545, + 481, + 578 + ], + "type": "text", + "content": "31. Madry, A., Makelov, A., Schmidt, L., Tsipras, D., Vladu, A.: Towards deep learning models resistant to adversarial attacks. In: International Conference on Learning Representations, ICLR (2018)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 578, + 481, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 578, + 481, + 610 + ], + "spans": [ + { + "bbox": [ + 132, + 578, + 481, + 610 + ], + "type": "text", + "content": "32. Moosavi-Dezfooli, S.M., Fawzi, A., Uesato, J., Frossard, P.: Robustness via curvature regularization, and vice versa. In: IEEE Conference on Computer Vision and Pattern Recognition, CVPR (2019)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 132, + 610, + 481, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 610, + 481, + 643 + ], + "spans": [ + { + "bbox": [ + 132, + 610, + 481, + 643 + ], + "type": "text", + "content": "33. Netzer, Y., Wang, T., Coates, A., Bissacco, A., Wu, B., Ng, A.Y.: Reading digits in natural images with unsupervised feature learning. Advances in Neural Information Processing Systems, NeurIPSW (2011)" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 132, + 643, + 481, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 643, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 643, + 481, + 666 + ], + "type": "text", + "content": "34. Petsiuk, V., Das, A., Saenko, K.: RISE: randomized input sampling for explanation of black-box models. In: British Machine Vision Conference, BMVC (2018)" + } + ] + } + ], + "index": 21 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 222, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 222, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 222, + 102 + ], + "type": "text", + "content": "P. Yang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 666 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "text", + "content": "35. Ribeiro, M.T., Singh, S., Guestrin, C.: \" why should i trust you?\" explaining the predictions of any classifier. In: ACM SIGKDD Conference on Knowledge Discovery & Data Mining, KDD (2016)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 150, + 482, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 150, + 482, + 183 + ], + "spans": [ + { + "bbox": [ + 130, + 150, + 482, + 183 + ], + "type": "text", + "content": "36. Ross, A.S., Doshi-Velez, F.: Improving the adversarial robustness and interpretability of deep neural networks by regularizing their input gradients. In: AAAI Conference on Artificial Intelligence, AAAI (2018)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 183, + 482, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 183, + 482, + 216 + ], + "spans": [ + { + "bbox": [ + 130, + 183, + 482, + 216 + ], + "type": "text", + "content": "37. Ross, A.S., Hughes, M.C., Doshi-Velez, F.: Right for the right reasons: Training differentiable models by constraining their explanations. In: International Joint Conference on Artificial Intelligence, IJCAI (2017)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 216, + 482, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 216, + 482, + 237 + ], + "spans": [ + { + "bbox": [ + 130, + 216, + 482, + 237 + ], + "type": "text", + "content": "38. Rudin, C.: Stop explaining black box machine learning models for high stakes decisions and use interpretable models instead. Nature Machine Intelligence (2019)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 237, + 482, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 237, + 482, + 281 + ], + "spans": [ + { + "bbox": [ + 130, + 237, + 482, + 281 + ], + "type": "text", + "content": "39. Sagawa, S., Koh, P.W., Hashimoto, T.B., Liang, P.: Distributionally robust neural networks for group shifts: On the importance of regularization for worst-case generalization. In: International Conference on Learning Representations, ICLR (2019)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 281, + 482, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 281, + 482, + 315 + ], + "spans": [ + { + "bbox": [ + 130, + 281, + 482, + 315 + ], + "type": "text", + "content": "40. Salahuddin, Z., Woodruff, H.C., Chatterjee, A., Lambin, P.: Transparency of deep neural networks for medical image analysis: A review of interpretability methods. Computers in biology and medicine 140, 105111 (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 315, + 482, + 347 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 315, + 482, + 347 + ], + "spans": [ + { + "bbox": [ + 130, + 315, + 482, + 347 + ], + "type": "text", + "content": "41. Samek, W., Binder, A., Montavon, G., Lapuschkin, S., Müller, K.: Evaluating the visualization of what a deep neural network has learned. TNNLS 28(11), 2660-2673 (2017)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 347, + 482, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 347, + 482, + 392 + ], + "spans": [ + { + "bbox": [ + 130, + 347, + 482, + 392 + ], + "type": "text", + "content": "42. Schramowski, P., Stammer, W., Teso, S., Brugger, A., Herbert, F., Shao, X., Luigs, H., Mahlein, A., Kersting, K.: Making deep neural networks right for the right scientific reasons by interacting with their explanations. Nature Machine Intelligence (2020)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 392, + 482, + 424 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 392, + 482, + 424 + ], + "spans": [ + { + "bbox": [ + 130, + 392, + 482, + 424 + ], + "type": "text", + "content": "43. Shafahi, A., Najibi, M., Ghiasi, M.A., Xu, Z., Dickerson, J., Studer, C., Davis, L.S., Taylor, G., Goldstein, T.: Adversarial training for free! In: Advances in Neural Information Processing Systems, NeurIPS (2019)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 424, + 482, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 424, + 482, + 445 + ], + "spans": [ + { + "bbox": [ + 130, + 424, + 482, + 445 + ], + "type": "text", + "content": "44. Shah, H., Jain, P., Netrapalli, P.: Do input gradients highlight discriminative features? In: Advances in Neural Information Processing Systems, NeurIPS (2021)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 446, + 482, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 446, + 482, + 479 + ], + "spans": [ + { + "bbox": [ + 130, + 446, + 482, + 479 + ], + "type": "text", + "content": "45. Shrikumar, A., Greenside, P., Kundaje, A.: Learning important features through propagating activation differences. In: International Conference on Machine Learning, ICML (2017)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 479, + 482, + 512 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 479, + 482, + 512 + ], + "spans": [ + { + "bbox": [ + 130, + 479, + 482, + 512 + ], + "type": "text", + "content": "46. Simon-Gabriel, C.J., Ollivier, Y., Bottou, L., Scholkopf, B., Lopez-Paz, D.: First-order adversarial vulnerability of neural networks and input dimension. In: International Conference on Machine Learning, ICML (2019)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 512, + 482, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 512, + 482, + 544 + ], + "spans": [ + { + "bbox": [ + 130, + 512, + 482, + 544 + ], + "type": "text", + "content": "47. Simonyan, K., Vedaldi, A., Zisserman, A.: Deep inside convolutional networks: Visualising image classification models and saliency maps. In: International Conference on Learning Representations, ICLR (2014)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 544, + 482, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 544, + 482, + 567 + ], + "spans": [ + { + "bbox": [ + 130, + 544, + 482, + 567 + ], + "type": "text", + "content": "48. Smilkov, D., Thorat, N., Kim, B., Viégas, F., Wattenberg, M.: Smoothgrad: removing noise by adding noise. CoRR abs/1706.03825 (2017)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 567, + 482, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 567, + 482, + 600 + ], + "spans": [ + { + "bbox": [ + 130, + 567, + 482, + 600 + ], + "type": "text", + "content": "49. Srinivas, S., Fleuret, F.: Rethinking the role of gradient-based attribution methods for model interpretability. In: International Conference on Learning Representations, ICLR (2021)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 130, + 600, + 482, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 600, + 482, + 622 + ], + "spans": [ + { + "bbox": [ + 130, + 600, + 482, + 622 + ], + "type": "text", + "content": "50. Srinivas, S., Fleuret, F.: Full-gradient representation for neural network visualization. In: Advances in Neural Information Processing Systems, NeurIPS (2019)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 130, + 622, + 482, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 622, + 482, + 643 + ], + "spans": [ + { + "bbox": [ + 130, + 622, + 482, + 643 + ], + "type": "text", + "content": "51. Sturmfels, P., Lundberg, S., Lee, S.I.: Visualizing the impact of feature attribution baselines. Distill (2020)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 130, + 643, + 482, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 643, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 643, + 482, + 666 + ], + "type": "text", + "content": "52. Sundararajan, M., Taly, A., Yan, Q.: Axiomatic attribution for deep networks. In: International Conference on Machine Learning, ICML (2017)" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 235, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 235, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 235, + 91, + 447, + 102 + ], + "type": "text", + "content": "Regulating Model Reliance on Non-Robust Features" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 358 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 480, + 138 + ], + "type": "text", + "content": "53. Wah, C., Branson, S., Welinder, P., Perona, P., Belongie, S.: The caltech-ucsd birds-200-2011 dataset. California Institute of Technology (2011)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 138, + 480, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 138, + 480, + 171 + ], + "spans": [ + { + "bbox": [ + 132, + 138, + 480, + 171 + ], + "type": "text", + "content": "54. Yang, P., Akhtar, N., Jiang, J., Mian, A.: Backdoor-based explainable AI benchmark for high fidelity evaluation of attribution methods. arXiv preprint arXiv:2405.02344 (2024)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 172, + 480, + 193 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 172, + 480, + 193 + ], + "spans": [ + { + "bbox": [ + 132, + 172, + 480, + 193 + ], + "type": "text", + "content": "55. Yang, P., Akhtar, N., Wen, Z., Mian, A.: Local path integration for attribution. In: AAAI Conference on Artificial Intelligence, AAAI (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 194, + 480, + 225 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 194, + 480, + 225 + ], + "spans": [ + { + "bbox": [ + 132, + 194, + 480, + 225 + ], + "type": "text", + "content": "56. Yang, P., Akhtar, N., Wen, Z., Shah, M., Mian, A.S.: Re-calibrating feature attributions for model interpretation. In: International Conference on Learning Representations (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 227, + 480, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 227, + 480, + 258 + ], + "spans": [ + { + "bbox": [ + 132, + 227, + 480, + 258 + ], + "type": "text", + "content": "57. Yang, P., Wen, Z., Mian, A.: Multi-grained interpretable network for image recognition. In: 2022 26th International Conference on Pattern Recognition (ICPR). IEEE (2022)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 259, + 480, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 259, + 480, + 281 + ], + "spans": [ + { + "bbox": [ + 132, + 259, + 480, + 281 + ], + "type": "text", + "content": "58. Zeiler, M.D., Fergus, R.: Visualizing and understanding convolutional networks. In: European Conference on Computer Vision, ECCV (2014)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 281, + 480, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 281, + 480, + 303 + ], + "spans": [ + { + "bbox": [ + 132, + 281, + 480, + 303 + ], + "type": "text", + "content": "59. Zhang, X., Wang, N., Shen, H., Ji, S., Luo, X., Wang, T.: Interpretable deep learning under fire. In: USENIX Security (2020)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 304, + 480, + 325 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 304, + 480, + 325 + ], + "spans": [ + { + "bbox": [ + 132, + 304, + 480, + 325 + ], + "type": "text", + "content": "60. Zhou, B., Lapedriza, A., Khosla, A., Oliva, A., Torralba, A.: Places: A 10 million image database for scene recognition. TPAMI 40(6), 1452-1464 (2017)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 325, + 480, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 325, + 480, + 358 + ], + "spans": [ + { + "bbox": [ + 132, + 325, + 480, + 358 + ], + "type": "text", + "content": "61. Zintgraf, L.M., Cohen, T.S., Adel, T., Welling, M.: Visualizing deep neural network decisions: Prediction difference analysis. In: International Conference on Learning Representations, ICLR (2017)" + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 222, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 222, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 222, + 102 + ], + "type": "text", + "content": "P. Yang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Reinforcement Learning Friendly Vision-Language Model for Minecraft/c4770d1a-703c-494e-ba79-18744626ca70_content_list.json b/2024/Reinforcement Learning Friendly Vision-Language Model for Minecraft/c4770d1a-703c-494e-ba79-18744626ca70_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..d945315eafebf331658cac8fb89066fe8c3113b4 --- /dev/null +++ b/2024/Reinforcement Learning Friendly Vision-Language Model for Minecraft/c4770d1a-703c-494e-ba79-18744626ca70_content_list.json @@ -0,0 +1,1763 @@ +[ + { + "type": "text", + "text": "Reinforcement Learning Friendly Vision-Language Model for Minecraft", + "text_level": 1, + "bbox": [ + 287, + 140, + 715, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Haobin Jiang $^{1\\star}$ , Junpeng Yue $^{1\\star}$ , Hao Luo $^{1}$ , Ziluo Ding $^{2}$ , and Zongqing Lu $^{1,2\\dagger}$", + "bbox": [ + 331, + 210, + 669, + 244 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ School of Computer Science, Peking University $^{2}$ Beijing Academy of Artificial Intelligence {haobin.jiang,zongqing.lu}@pku.edu.cn", + "bbox": [ + 336, + 253, + 665, + 297 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. One of the essential missions in the AI research community is to build an autonomous embodied agent that can achieve high-level performance across a wide spectrum of tasks. However, acquiring or manually designing rewards for all open-ended tasks is unrealistic. In this paper, we propose a novel cross-modal contrastive learning framework architecture, CLIP4MC, aiming to learn a reinforcement learning (RL) friendly vision-language model (VLM) that serves as an intrinsic reward function for open-ended tasks. Simply utilizing the similarity between the video snippet and the language prompt is not RL-friendly since standard VLMs may only capture the similarity at a coarse level. To achieve RL-friendliness, we incorporate the task completion degree into the VLM training objective, as this information can assist agents in distinguishing the importance between different states. Moreover, we provide neat YouTube datasets based on the large-scale YouTube database provided by MineDojo. Specifically, two rounds of filtering operations guarantee that the dataset covers enough essential information and that the video-text pair is highly correlated. Empirically, we demonstrate that the proposed method achieves better performance on RL tasks compared with baselines. The code and datasets are available at https://github.com/PKU-RL/CLIP4MC.", + "bbox": [ + 261, + 333, + 743, + 612 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Dataset $\\cdot$ Multimodal model $\\cdot$ Reinforcement learning", + "bbox": [ + 261, + 623, + 710, + 640 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 664, + 375, + 680 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Training reinforcement learning (RL) agents to perform complex tasks in a vision-based and open-ended world can be difficult. One main challenge is that manually specifying reward functions in all open-ended tasks is unrealistic [3,21], especially when we cannot access the internal configuration of the environment. In addition, learning a reward model from human feedback is typically expensive. To this end, MineDojo [8] has proposed an internet-scale, multi-modal knowledge base of YouTube videos to facilitate learning in an open-ended world. With the advent of a such large-scale database, agents are able to harvest practical", + "bbox": [ + 212, + 695, + 787, + 816 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Equal contribution. † Corresponding author.", + "bbox": [ + 230, + 824, + 542, + 840 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "knowledge encoded in large amounts of media like human beings. Moreover, a vision-language model (VLM), MineCLIP [8], is proposed to utilize the internet-scale domain knowledge. In more detail, the learned correlation score between the visual observation and the language prompt can be used effectively as an open-vocabulary, massively multi-task reward function for RL training. Therefore, no further task-specific reward design is needed for open-ended tasks.", + "bbox": [ + 212, + 146, + 787, + 238 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "However, the autonomous embodied agent requires the vision-language model to provide a more instructive correlation score. Given the partial observations, e.g., video snippet, and the language prompt that describes the task, the agent needs to figure out two non-trivial matters to better evaluate the current state. On the one hand, whether the target entities are present within its field of vision? MineCLIP tried to address this question. However, the alignment of texts and videos in the YouTube database is totally a catastrophe, which impedes the learning of VLM. On the other hand, what is the relationship between each video snippet and the degree of completion of the task? Normally, in an open-ended world, e.g. Minecraft, the agent explores first, then approaches and interacts with the target object. In other words, the agent requires approaching the target object before it takes trial-and-error, even for the target that needs to be kept away. Therefore, it is reasonable to make an assumption, namely, that the higher the level of completion of the task, the closer the agent is to the targets in the video snippet. We also argue it is important to incorporate the level of completion of the task into the reward function.", + "bbox": [ + 212, + 241, + 789, + 482 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we first construct neat YouTube datasets to facilitate the learning of basic game concepts, mainly the correspondence between the videos and texts. Though a large-scale database is provided by MineDojo, it contains significant noise due to its nature as an online resource. In addition, MineDojo only claims the training dataset is randomly sampled from the database, making it hard to reproduce. To overcome the catastrophic misalignment of video-text pairs in the original database, we have done four steps of dataset processing to guarantee the dataset is clean. Firstly, transcript cleaning enhances the accuracy of transcripts and ensures they are complete sentences. Secondly, keyword filtering ensures that the content of clips is relevant to the key entities in Minecraft, thereby facilitating the learning of basic game concepts. Thirdly, video partitioning and selection can handle the issue of scene transitions and thereby mitigate interference from other extraneous information. Lastly, correlation filtering can effectively address the issue of mismatch between video clips and transcripts.", + "bbox": [ + 212, + 488, + 789, + 700 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We also propose an upgraded vision-language model, CLIP4MC, to provide a more RL-friendly reward function. In RL, simply utilizing the similarity between the video snippet and the language prompt is not RL-friendly since MineCLIP tends to only capture the similarity at the entity level. In other words, VLM can hardly reflect the relationship between each video snippet and the degree of completion of the task. However, this information can better help agents distinguish the importance between similar states. To achieve this, we incorporate the degree of task completion into the VLM training objective. In more detail, CLIP has exhibited remarkable segmentation capabilities without fine-tuning. After we", + "bbox": [ + 212, + 704, + 789, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "H. Jiang et al.", + "bbox": [ + 271, + 114, + 367, + 128 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "extend pre-trained MineCLIP with modifications inspired by MaskCLIP [34], it can segment the specified object from the image and label the size of the target shown in the corresponding video. Intuitively, the closer the agent is to the target, the larger the target size becomes. During the learning procedure, we dynamically control the degree of contrasting positive and negative pairs of instances based on the target size in this positive video sample. Thus, CLIP4MC can render a more RL-friendly reward signal that instructs the agent to learn tasks faster. Our proposed method is trained on our YouTube dataset and evaluated on MineDojo Programmatic tasks, including harvest, hunt, and combat tasks. Empirically, our results show that CLIP4MC can provide a more friendly reward signal for the RL training procedure.", + "bbox": [ + 212, + 146, + 787, + 311 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To summarize, our contributions are as follows:", + "bbox": [ + 214, + 313, + 560, + 325 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Open-sourced datasets: We provide two high-quality datasets. The first one undergoes data cleaning (Sections 4.1 to 4.3) and global-level correlation filtering (Section 4.4). The VLM trained on this dataset matches the performance of the officially released MineCLIP, which lacks a publicly available training set.", + "- RL-friendlyly dataset: Our second dataset further incorporates local-level correlation filtering (Section 4.4), making it more suited for RL. The VLM trained on this dataset outperforms that on the first dataset.", + "- RL-friendly VLM: To better evaluate and leverage the advantages of our more RL-friendly dataset, we introduce CLIP4MC, a novel method to train a VLM that could improve downstream RL performance (Section 5.2)." + ], + "bbox": [ + 225, + 335, + 782, + 500 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 214, + 522, + 387, + 537 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Video-Text Retrieval. Video-text retrieval plays an essential role in multimodal research and has been widely used in many real-world web applications. Recently, the pre-trained models have dominated this line of research with noticeable results on both zero-shot and fine-tuned retrieval. Especially, BERT [5], ViT [6], and CLIP [25], are used as the backbones to extract the text or video embedding. The cross-modal embeddings are then matched with specific fusion networks to find the correct video-text pair.", + "bbox": [ + 212, + 551, + 787, + 657 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In more detail, CLIP4Clip [20] proposes three different similarity modules to calculate the correlation between video and text embeddings. HiT [19] performs hierarchical matching at two different levels, i.e. semantic level and feature level. Note that semantic level and feature level features are from the transformer network's higher and lower feature layers, respectively. Frozen [1] proposes a dual encoder architecture that utilizes the flexibility of a transformer visual encoder to train from images or video clips with text captions. Moreover, MDMMT [7] adopts several pre-training models as encoders and it shows the CLIP-based model performs the best. Therefore, our model follows this line of research by using the pre-trained model, CLIP [25], to extract the feature embeddings.", + "bbox": [ + 212, + 657, + 787, + 809 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Minecraft for AI Research. As an open-ended video game with an egocentric vision, Minecraft is a noticeable and important domain in RL due to the nature", + "bbox": [ + 214, + 809, + 785, + 839 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "CLIP4MC", + "bbox": [ + 658, + 113, + 730, + 127 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 114, + 785, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "of the sparse reward, large exploration space, and long-term episodes. Since the release of the Malmo simulator [14] and later the MineDojo simulator [8], various methods have attempted to train agents to complete tasks in Minecraft [10, 12, 18, 29, 30]. Approaches such as model-based RL, hierarchical RL, goal-based RL, and reward shaping have been adopted to alleviate the sparse reward and exploration difficulty for the agent in this environment.", + "bbox": [ + 212, + 146, + 787, + 236 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Recently, with the development of large language models (LLM) like GPT-4 [24], a series of methods leveraging LLMs for high-level planning in Minecraft have been proposed [16, 22, 31, 33, 35]. These methods have demonstrated remarkable capabilities in guiding the agent to complete multiple complicated, long-horizon tasks, such as mining diamonds. These LLMs play a crucial role in decision-making, determining the sequence of basic skills required to accomplish specific tasks. Their effectiveness is due to their extensive knowledge about Minecraft, learned from the Internet, and their ability to reflect on real-time feedback from the game environment.", + "bbox": [ + 212, + 237, + 787, + 372 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In addition to the use of LLMs, recent research attempts to incorporate Internet visual data into basic skill learning in Minecraft, beyond the traditional RL methods. MineRL [11] collected 60M player demonstrations with action labels, motivating some methods [16,28] based on behavior cloning. As well-labeled data is limited in quantity, MineCLIP [8] instead uses over 730K narrated Minecraft videos without action labels from YouTube. It aims to learn a vision-language model providing auxiliary reward signals, utilizing the vast and diverse data available on the Internet. Different from MineCLIP, VPT [2] uses action-labeled data to train an inverse dynamic model to label 70K hours of Internet videos and then conduct behavior cloning.", + "bbox": [ + 212, + 373, + 787, + 523 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Unlike the existing approaches, which incorporate the human experience and require a large number of demonstrations with action labels to train the agent, our work follows the line of MineCLIP [8] and focuses on only using the data without action labels to assist agent learning in Minecraft, which is more friendly with data collection and has the potential to scale in the future.", + "bbox": [ + 212, + 525, + 787, + 602 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3 Background", + "text_level": 1, + "bbox": [ + 215, + 625, + 370, + 643 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "MineDojo tasks. MineDojo [8] provides thousands of benchmark tasks, which can be used to develop generally capable agents in Minecraft. These tasks can be divided into two categories, Programmatic and Creative tasks. The former has ground-truth simulator states to assess whether the task has been completed. The latter, however, do not have well-defined success criteria and tend to be more open-ended, but have to be evaluated by humans.", + "bbox": [ + 212, + 657, + 787, + 748 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We mainly focus on Programmatic tasks since they can be automatically assessed. Specifically, MineDojo provides 4 categories of programmatic tasks, including Harvest, Combat, Survival, and Tech Tree, with 1581 template-generated natural language goals to evaluate the agent's different capabilities. Among these tasks, Survival and Tech Tree tasks are harder than Harvest and Combat tasks. Currently, MineCLIP [8] only expresses promising potential in some Harvest and", + "bbox": [ + 212, + 750, + 787, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 230, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "H. Jiang et al.", + "bbox": [ + 271, + 114, + 369, + 128 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/be300fae5c1064a448f8b924e71b7a754723634d5ac4539840950a0e1c1966db.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 263, + 146, + 609, + 204 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "irrelevant video content", + "text_level": 1, + "bbox": [ + 647, + 148, + 733, + 156 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "... so here I'm with nick and we've got look at my inventory and his inventory looks similar this is gonna be interesting so four...", + "bbox": [ + 614, + 160, + 730, + 191 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/b32ef89bcb2fe46bf77fcec0507609a18eaf98a96463e7ac3995570d3b8b0571.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 263, + 212, + 609, + 267 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "mismatched video content", + "text_level": 1, + "bbox": [ + 640, + 210, + 733, + 219 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "... give it I'm a little chicken just laying around by any chance a lot of arrows going on down here I'm guessing most of ...", + "bbox": [ + 614, + 224, + 730, + 255 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/da7af01c9c11db865bacd59f0bd001029674962d8210b69d0c492f200e7712dc.jpg", + "image_caption": [ + "Fig. 1: Illustration of the YouTube video database. The screenshots of video clips are on the left and key entities are circled in red. The corresponding transcript clips are on the right and key entities are marked in red. We give examples of irrelevant, mismatched, and matched video content in the YouTube video database." + ], + "image_footnote": [], + "bbox": [ + 263, + 273, + 609, + 329 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "matched video content", + "text_level": 1, + "bbox": [ + 653, + 273, + 733, + 281 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "... at some point you will lose there too or also hearts there spider was like being completely blown away! I don't assume anything at ...", + "bbox": [ + 614, + 287, + 732, + 318 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Combat tasks. Harvest means finding, obtaining, cultivating, or manufacturing hundreds of materials and objects. Combat means fighting various monsters and creatures that require fast reflexes and martial skills.", + "bbox": [ + 212, + 433, + 784, + 477 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "POMDP. We model the programmatic task as a partially observable Markov decision process (POMDP) [15]. At each timestep $t$ , the agent obtains the partial observation $o_t$ from the global state $s_t$ and a language prompt $G$ , takes action $a_t$ following its policy $\\pi(a_t|o_t)$ , and receives a reward $r_t = \\Phi(V_t, G)$ , where $V_t$ is the fixed-length sequence of observations till $t$ (thus a video snippet) and $\\Phi$ maps $V_t$ and $G$ to a scalar value. Then the environment transitions to the next state $s_{t+1}$ given the current state and action according to transition probability function $\\mathcal{T}(s_{t+1}|s_t, a_t)$ . The agent aims to maximize the expected return $R = \\mathbb{E}_{\\pi} \\sum_{t=1}^{T} \\gamma^{t-1} r_t$ , where $\\gamma$ is the discount factor and $T$ is the episode time horizon.", + "bbox": [ + 212, + 479, + 785, + 617 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4 YouTube Dataset", + "text_level": 1, + "bbox": [ + 214, + 641, + 419, + 656 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The Internet is rich in Minecraft-related data, containing a wealth of weakly labeled or even unlabeled Minecraft knowledge, including crucial entities, plausible actions, and common-sense event processes. With these multi-modal data, it is possible to create dense language-conditioned rewards, making open-ended long-term task learning feasible. MineDojo [8] collected over 730K YouTube videos and their corresponding transcripts, totaling 33 years of video content and 2.2B words in transcripts. Around 640K video clips are selected using keyword-based filtering from these videos, and these clips are used to train MineCLIP. However, MineDojo open-sourced a 13.8M dataset, and the 640K video clips were randomly sampled in their paper. However, as illustrated in Figure 1, most videos feature irrelevant game content that is not conducive to learning basic game", + "bbox": [ + 212, + 672, + 787, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "CLIP4MC", + "bbox": [ + 658, + 114, + 730, + 126 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "concepts. Meanwhile, the alignment between the transcripts and videos may not always be precise, leading to temporal or content discrepancies that could hinder the learning of retrieval and RL tasks. Given the low quality of the data, we found it necessary to provide a neat 640k dataset to the community for the training usage. To address these issues, we adopt the following few steps to acquire a clean dataset with high quality.", + "bbox": [ + 212, + 146, + 787, + 238 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1 Transcript Cleaning", + "text_level": 1, + "bbox": [ + 215, + 258, + 426, + 273 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Downloading YouTube's automatic transcripts directly can lead to several issues. Firstly, transcript blocks may overlap, resulting in overlapping timestamps in the transcript. Additionally, the caption quality is typically mediocre, with a higher occurrence of misidentifications, especially in non-English videos. Moreover, the transcript lacks punctuation, making it less friendly for understanding semantics. Based on the aforementioned issues, we implement a pipeline to construct high-quality transcripts as follows: (1) Extract audio from the videos and use Whisper [26] to obtain high-quality, temporally non-overlapping transcripts. (2) Employ FullStop [9] to generate punctuation, resulting in complete sentences of 10-35 words in length.", + "bbox": [ + 212, + 282, + 787, + 434 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2 Keyword Filtering", + "text_level": 1, + "bbox": [ + 215, + 454, + 415, + 469 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Following MineDojo [8], we also implement keyword-based filtering to ensure that the content in our dataset is pertinent to the key entities in Minecraft, thereby facilitating the learning of basic game concepts. As essential components of Minecraft, the key entities, such as stones, trees, and sheep, are common across multiple tasks in MineDojo and videos from YouTube. Specifically, we identify entity keywords in the transcripts using a keyword list from MineDojo and extract transcript clips formed into sentences to encompass as many keywords as possible. These extracted transcript clips then serve as the textual component of our dataset, determining the location of corresponding video clips.", + "bbox": [ + 212, + 477, + 787, + 613 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3 Video Partition and Selection", + "text_level": 1, + "bbox": [ + 215, + 633, + 506, + 648 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "After completing the previous two steps, we obtain transcript clips relevant to the keywords. For each transcript clip, we calculate the central timestamp that corresponds to the clip based on the transcript timestamps and then use this central timestamp to extract a video clip with a duration of $D$ seconds from the video. This process allows us to obtain temporally-aligned video clips. However, the video clips obtained in this manner still exhibit some issues, including scene transitions and discrepancies between video content and transcripts. We handle the former problem through video partition and filtering, while the latter problem is addressed in Section 4.4.", + "bbox": [ + 212, + 659, + 787, + 792 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Owing to the informal nature of YouTube content, there is often a lack of semantic congruence between the video clips and their corresponding transcriptions, as noted in VideoCLIP [32]. Moreover, video clips frequently contain a", + "bbox": [ + 212, + 795, + 787, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "H. Jiang et al.", + "bbox": [ + 271, + 114, + 367, + 128 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/a8d0bf9dbe81d2743b2257b8a10eebd1e3d489606188ad5647f2d136edf00ce6.jpg", + "image_caption": [ + "Fig. 2: Examples of how we estimate the size of key entities in video frames. Red bounding boxes are generated by our modified MineCLIP visual encoder, following the approach proposed in MaskCLIP [34]. These boxes are then used to calculate the size." + ], + "image_footnote": [], + "bbox": [ + 330, + 143, + 674, + 313 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "few different behaviors which cause scene transitions, e.g. Chopping down the tree first, then suddenly switching the inventory bar. Since some scene transitions lead to irrelevant information, we partition the video content into several semantically coherent segments based on the semantic structure of the video. Then we select the segment that aligns best with the transcript.", + "bbox": [ + 212, + 396, + 784, + 470 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To achieve semantic partition of video content, we employ the Bellman $k$ -segmentation algorithm [13]. This algorithm divides a sequence of data points into $k$ distinct and constant-line segments, providing a piece-wise constant approximation of the data. To process the video, we first use the officially released MineCLIP [8] video encoder to obtain the embedding of each frame, since MineCLIP can capture video semantics to some extent. Subsequently, we partition these embeddings into $k$ segments and select the segment with the highest similarity score, as calculated in MineCLIP.", + "bbox": [ + 212, + 472, + 787, + 592 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4 Correlation Filtering", + "text_level": 1, + "bbox": [ + 215, + 616, + 434, + 633 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We employ correlation filtering techniques to address disparities that persist between video content and transcripts. The correlation filtering is done at two levels, the global and the local levels. From the global level, we calculate the cosine similarities between video embeddings and text embeddings via the original MineCLIP and then select clips based on these similarities.", + "bbox": [ + 212, + 643, + 784, + 717 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Recent research [17, 34] has demonstrated that CLIP [25], though trained on whole images, can generate meaningful local features. Inspired by this and following MaskCLIP [34], we make modifications to the original MineCLIP visual encoder, empowering it with the ability to estimate and label the size of the key entity, which is mentioned in the corresponding transcript, in each frame of a video clip without fine-tuning. Then the correlations between video clips and transcripts are calculated as the summation of the sizes across all frames in each video clip. We consider this correlation at the local level.", + "bbox": [ + 212, + 719, + 785, + 839 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "CLIP4MC", + "bbox": [ + 658, + 113, + 730, + 126 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 114, + 785, + 126 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Based on these two criteria, we select the top $k\\%$ of clips, resulting in the final training set. We provide an elaboration on our implementation in Appendix A and illustrate some examples in Figure 2.", + "bbox": [ + 212, + 146, + 782, + 191 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The aforementioned four-step approach creates a dataset consisting of 640K video-text clip pairs, with an additional 4K pairs extracted for validation of video-text retrieval. Regarding the constants of the approach, $D$ and $k$ are set to 16 and 50, respectively. Therefore, our dataset comprises videos with a total duration of one week and approximately 0.16B words, significantly smaller in scale compared to the original low-quality 13.8M dataset. Importantly, we will open-source our entire YouTube dataset, serving as an upgraded version of the unreleased MineCLIP training data.", + "bbox": [ + 212, + 191, + 787, + 313 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5 CLIP4MC", + "text_level": 1, + "bbox": [ + 215, + 335, + 354, + 351 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Given a video snippet $V$ and a language prompt $G$ , the vision-language model outputs a correlation score, $C$ , that measures the similarity between the video snippet and the language prompt. Ideally, if the agent performs behaviors following the description of the language prompt, the vision-language model will generate a higher correlation score, leading to a higher reward. Otherwise, the agent will be given a lower reward.", + "bbox": [ + 212, + 368, + 784, + 458 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.1 Architecture", + "text_level": 1, + "bbox": [ + 215, + 481, + 366, + 494 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We follow the same architecture design as MineCLIP [8], including a video encoder, a text encoder, and a similarity calculator. All the video frames first go through the spatial transformer to obtain a sequence of frame features. The temporal transformer is then utilized to summarize the sequence of frame features into a single video embedding. An adapter further processes the video embedding for better features. Refer to Appendix B for more details about the architecture.", + "bbox": [ + 212, + 507, + 782, + 595 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Empirically, we found that the video encoder (essentially MineCLIP) can provide a bond between the entities and the language prompts and give a similar high reward as long as the target entities are present in the video frames, similar to observations in [4]. However, such a reward is not instructive enough for RL tasks since it does not reflect the behavioral trends of agents.", + "bbox": [ + 212, + 598, + 782, + 674 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.2 Contrastive Training", + "text_level": 1, + "bbox": [ + 215, + 695, + 433, + 710 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We aim to minimize the sum of the multi-modal contrastive losses [23], including video-to-text, and text-to-video:", + "bbox": [ + 212, + 720, + 782, + 750 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = - \\sum_ {\\left(z _ {v}, z _ {m}, z _ {t}\\right) \\in \\mathcal {B}} \\left(\\log \\operatorname {N C E} \\left(z _ {v}, z _ {t}\\right) + \\log \\operatorname {N C E} \\left(z _ {t}, z _ {v}\\right)\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 297, + 763, + 784, + 797 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $\\mathcal{B}$ is the batch containing sampled video-text pairs, and $\\mathrm{NCE}(\\cdot ,\\cdot)$ is the contrastive loss that calculates the similarity of two inputs. To illustrate, the", + "bbox": [ + 212, + 809, + 782, + 839 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "H. Jiang et al.", + "bbox": [ + 271, + 114, + 367, + 128 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/1e7069af898432d6a42b90f799985badac834b84318181f38c21b8036c82daf4.jpg", + "image_caption": [ + "Fig. 3: Illustration of CLIP4MC training. The upper part shows the concept of contrastive learning, while the lower part explains the swapping operation." + ], + "image_footnote": [], + "bbox": [ + 313, + 146, + 694, + 429 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "video-to-text contrastive loss is given by", + "bbox": [ + 215, + 498, + 506, + 513 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {N C E} \\left(z _ {v}, z _ {t}\\right) = \\frac {\\exp \\left(z _ {v} \\cdot z _ {t} ^ {+} / \\lambda\\right)}{\\sum_ {z \\in \\left\\{z _ {t} ^ {+} , z _ {t} ^ {-} \\right\\}} \\exp \\left(z _ {v} \\cdot z / \\lambda\\right)}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 351, + 521, + 784, + 559 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where $\\lambda$ is a temperature hyperparameter, $z_{t}^{+}$ is the positive text embedding matching with the video embedding $z_{v}$ , and $\\{z_t^-\\}$ are negative text embeddings that are implicitly formed by other text clips in the training batch. Contrastive loss of text-to-video is defined in the same way.", + "bbox": [ + 212, + 568, + 784, + 628 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "In addition, we also need to incorporate the degree of task completion into the training objective. Specifically, we hope the similarity score between video clip and text embeddings could reflect the task completion degree, i.e., higher completion leads to higher similarity. The sizes of key entities provided by the local correlation filtering serve as a surrogate for the task completion degree. Initially, we try to dynamically adjust the weight of the positive pairs based on the size of the target object. However, it did not work. Essentially, positive and negative pairs will still be separated, even with smaller sample weights.", + "bbox": [ + 212, + 628, + 782, + 750 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Therefore, we instead forcibly change the labels of the positive and negative video samples, as illustrated in Figure 3, thereby reducing the model's confidence in certain positive pairs. Specifically, during training, a positive video sample is swapped with a random negative video sample from the $\\mathcal{B}$ based on probability $P_{\\mathrm{swap}}$ . This swap only occurs when the labeled target size $H_bW_b$ in the positive video sample is below a certain threshold $\\tau$ , and the smaller the size, the greater", + "bbox": [ + 212, + 750, + 784, + 839 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "CLIP4MC", + "bbox": [ + 658, + 114, + 730, + 126 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "the probability,", + "bbox": [ + 215, + 146, + 331, + 161 + ], + "page_idx": 9 + }, + { + "type": "equation", + "text": "\n$$\nP _ {\\mathrm {s w a p}} = P _ {\\mathrm {m a x}} * \\mathrm {R e L U} \\left(1 - \\frac {H _ {b} W _ {b}}{\\tau H W}\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 370, + 161, + 785, + 193 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "where $HW$ represents the size of the image.", + "bbox": [ + 215, + 195, + 534, + 209 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Since contrastive learning brings positive samples closer and pushes negative samples away, converting a positive pair to a negative one with a certain probability, i.e., the aforementioned random swap, during training will lower the similarity score. As the decrease in similarity is directly proportional to the swapping probability, we set $P_{\\mathrm{swap}}$ to be inversely related to task completion, ensuing similarity score increases as completion increases.", + "bbox": [ + 212, + 210, + 785, + 299 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "In addition, the upper limit $P_{\\mathrm{max}}$ in Equation (3) is set to 0.5, as the swap should not disrupt severely the normal distinction between positive and negative pairs in most cases. This constraint enables CLIP4MC to preserve the understanding capability of MineCLIP for general behaviors that may lack explicit target entities in Minecraft, as discussed in Appendix F.", + "bbox": [ + 212, + 300, + 785, + 376 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "5.3 RL Training", + "text_level": 1, + "bbox": [ + 215, + 393, + 364, + 410 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "For RL training, the first step is reward generation. At timestep $t$ , we concatenate the agent's latest 16 egocentric RGB frames in a temporal window to form a video snippet, $V_{t}$ . CLIP4MC outputs the probability $P_{G,t}$ that calculates the similarity of $V_{t}$ to the task prompt, $G$ , against all other negative prompts. To compute the reward, we further process the raw probability as previous work [8] $r_t = \\max (P_{G,t} - 1 / N_t,0)$ , where $N_{t}$ is the number of prompts passed to CLIP4MC. Note that CLIP4MC can handle unseen language prompts without any further fine-tuning due to the open-vocabulary ability of CLIP [25].", + "bbox": [ + 212, + 415, + 785, + 536 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "The ultimate goal is to train a policy network that takes as input raw pixels and other structural data and outputs discrete actions to accomplish the task that is described by the language prompt, $G$ . We use PPO [27] as our RL training backbone and the policy is trained on the CLIP4MC reward together with the sparse task reward if any. The policy input contains several modality-specific components and more details can be found in Appendix D.", + "bbox": [ + 212, + 537, + 785, + 628 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "6 Experiments", + "text_level": 1, + "bbox": [ + 215, + 646, + 375, + 662 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "In this section, we comprehensively evaluate and analyze our proposed model CLIP4MC, utilizing the open-ended platform MineDojo [8], which comprises thousands of diverse, open-ended Minecraft tasks designed for embodied agents.", + "bbox": [ + 212, + 672, + 785, + 718 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "We compare CLIP4MC against two baselines: (1) MineCLIP [official], the officially released MineCLIP model [8]. (2) MineCLIP [ours], using the same architecture as MineCLIP [official] but trained on our cleaned YouTube dataset. It also serves as the ablation of CLIP4MC without the swap operation. We train CLIP4MC and MineCLIP [ours] for 20 epochs and select the models with the highest performance on RL tasks. Please refer to Appendix C for more training details. All results are presented in terms of the mean and standard error of four runs with different random seeds.", + "bbox": [ + 212, + 719, + 787, + 839 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "H. Jiang et al.", + "bbox": [ + 271, + 114, + 367, + 128 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "6.1 Environment Settings", + "text_level": 1, + "bbox": [ + 215, + 146, + 442, + 162 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "We conduct experiments on eight Programmatic tasks, comprising two harvest tasks: milk a cow and shear wool, two combat tasks: combat a spider and combat a zombie, and four hunt tasks: hunt a cow, hunt a sheep, hunt a pig, and hunt a chicken. These tasks are all built-in tasks in the MineDojo benchmark.", + "bbox": [ + 212, + 171, + 787, + 231 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Harvest. Milk a cow requires the agent to obtain milk from a cow with an empty bucket. Similarly, shear wool requires the agent to obtain wool from a sheep with shears. A harvest task is terminated and considered completed when the target item is obtained by the agent with a specified quantity. The prompts used to calculate the reward is \"obtain milk from a cow in plains with an empty bucket\"; for shear wool, it is \"shear a sheep in plains with shears\".", + "bbox": [ + 212, + 232, + 787, + 323 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Combat. In these tasks, target animals, spiders, and zombies, are hostile and will actively approach and attack the agent. The agent's goal is to fight and kill the target animal. The prompt for each combat task is \"combat a spider/zombie in plains with a diamond sword\".", + "bbox": [ + 212, + 323, + 787, + 382 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Hunt. Hunt tasks consist of hunt a cow, hunt a sheep, hunt a pig, and hunt a sheep. For each task, the agent's goal is to kill the target animal as indicated in the task name. Different from combat tasks, the target animals will flee from the agent after being attacked. Therefore, these tasks require the agent to keep chasing and attacking the target, making them challenging. As noted in [4], the original MineCLIP reward fails in these tasks since it cannot consistently increase when the agent approaches the agent. This observation aligns with our assertion that the original MineCLIP model can hardly capture the degree of task completion. The prompt for each task is \"hunt a {target} in plains with a diamond sword\" where {target} is replaced with the corresponding animal name.", + "bbox": [ + 212, + 383, + 787, + 534 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "More elaborated introduction of the Minecraft environment and settings of these tasks are available in Appendix D. To guarantee a fair comparison, we adopt the same RL hyperparameters for all models and tasks. These hyperparameters are listed in Appendix E.", + "bbox": [ + 212, + 534, + 787, + 595 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "6.2 RL Results", + "text_level": 1, + "bbox": [ + 215, + 617, + 356, + 631 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Through our evaluation of CLIP4MC, MineCLIP [ours], and MineCLIP [official] across eight Minecraft tasks, we want to answer two key questions:", + "bbox": [ + 212, + 642, + 785, + 672 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "(1) Whether the YouTube dataset we constructed enables MineCLIP, when trained on it, to provide a more effective reward signal for task learning?", + "(2) Whether our upgraded model, CLIP4MC, based on our YouTube dataset, offers a reward signal that is further friendly for the RL training procedure?" + ], + "bbox": [ + 214, + 674, + 785, + 734 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "These questions are central to verifying the effectiveness of our dataset and CLIP4MC model in Minecraft. Table 1 shows the success rates of all methods on the eight Minecraft tasks.", + "bbox": [ + 212, + 734, + 785, + 777 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "It is noticeable that, in four hunt tasks, three models demonstrate varying performance on RL. Firstly, MineCLIP [ours] consistently achieves better results compared to MineCLIP [official] across all hunt tasks. The superior performance of MineCLIP [ours] provides a positive answer to our first question, suggesting", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "CLIP4MC", + "bbox": [ + 658, + 113, + 730, + 127 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 785, + 127 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/5a34dfc5818bca4c7a693da0018064b5724a868fe3def8236b5e57804484ca7a.jpg", + "table_caption": [ + "Table 1: Success rates $(\\%)$ of RL trained with rewards provided by different models on eight Minecraft tasks. Each mean and standard error of success rates are calculated on four models after training 1e6 environment steps with different random seeds." + ], + "table_footnote": [], + "table_body": "
ModelsHarvestCombat
milk a cowshear woolcombat a spidercombat a zombie
CLIP4MC84.5±2.074.6±2.185.8±0.970.4±8.3
MineCLIP[ours]84.4±1.171.6±3.575.4±10.163.6±8.7
MineCLIP[official]84.1±0.573.2±1.882.7±2.557.4±3.7
ModelsHunt
hunt a cowhunt a sheephunt a pighunt a chicken
CLIP4MC39.8±2.545.9±7.230.6±8.426.1±3.5
MineCLIP[ours]17.3±10.633.0±18.114.1±10.615.3±10.6
MineCLIP[official]11.6±11.128.5±16.71.5±0.60.0±0.0
", + "bbox": [ + 220, + 198, + 777, + 402 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "that the YouTube dataset we construct indeed enhances the effectiveness of the reward signal in MineCLIP for task learning. Note that this is the dataset we plan to release for better training the VLM model for RL tasks on Minecraft. Secondly, CLIP4MC shows significantly higher success rates on hunt tasks compared to both MineCLIP [ours] and MineCLIP [official], meaning that the answer to the second question is also positive. As CLIP4MC provides a reward signal taking into account a surrogate for the degree of task completion, i.e., the size of the target object in our implementation, it becomes more RL-friendly in these challenging tasks.", + "bbox": [ + 212, + 438, + 787, + 575 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "In addition, we notice a practical example of misalignment in the officially released MineCLIP model. Specifically, we observe that in hunt a chicken, the agent trained with MineCLIP [official] tends to keep looking at the sky, indicating such behavior can provide a high intrinsic reward. This phenomenon suggests that the officially released MineCLIP indeed suffers from the misalignment problem in the YouTube dataset. In contrast, our MineCLIP [official] shows promising behaviors on this task, demonstrating that our dataset processing improves the alignment between the transcripts and videos.", + "bbox": [ + 212, + 579, + 787, + 700 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "In contrast to hunt tasks, these methods perform comparably on harvest tasks and combat tasks, while CLIP4MC still shows marginal advantage over the other two methods. This finding aligns with the results reported in [8], where the MineCLIP reward already achieves saturated performance on these easy tasks. Therefore, further enhancements in the reward signal, like MineCLIP [ours] and CLIP4MC, do not significantly improve performance. Unlike hunt tasks, neither harvest nor combat tasks require the agent to take multiple rounds of chasing. This result does not detract from the superiority of CLIP4MC, evidenced by its performance on hunt tasks, which are considered more difficult [2].", + "bbox": [ + 212, + 704, + 787, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "H. Jiang et al.", + "bbox": [ + 271, + 114, + 367, + 128 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/fef7ef70be28bbb120a7d8b88b5956ec245bd7ca16eb0c4529f81fdf38f51cac.jpg", + "image_caption": [ + "(a) CLIP4MC" + ], + "image_footnote": [], + "bbox": [ + 287, + 143, + 416, + 223 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/d6f8936764fc0ef98eb713c07bd06f885e82f4f3a81bd090e352396984838625.jpg", + "image_caption": [ + "(b) MineCLIP [ours]", + "Fig. 4: Scatter plots illustrating the relationship between the entity size and the intrinsic reward. The red line indicates a linear fit to the data." + ], + "image_footnote": [], + "bbox": [ + 436, + 143, + 568, + 223 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/7bc8aa2f72891617c881769faa584b7b2d959cf9b8226567d813845665890831.jpg", + "image_caption": [ + "(c) MineCLIP [official]" + ], + "image_footnote": [], + "bbox": [ + 586, + 143, + 717, + 223 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/0efcb99cd90cc4676c4ee37b9a0b0a7ac7c29307ea06a95389dcca6a65e5ec90.jpg", + "table_caption": [ + "Table 2: Success rates (%) of RL trained with rewards provided by different models on three Minecraft tasks. Each mean and standard error of success rates are calculated on four models after training 5e5 and 1e6 environment steps with different random seeds." + ], + "table_footnote": [], + "table_body": "
Modelsmilk a cowshear woolhunt a cow
5e51e65e51e65e51e6
ours73.1±0.984.4±1.147.5±6.371.6±3.515.6±8.717.3±10.6
w/o LCF71.3±1.880.7±1.153.8±3.373.6±1.50.7±0.712.1±12.0
RS65.1±3.981.7±1.021.7±6.568.5±3.61.0±0.54.0±2.6
official69.8±1.384.1±0.547.0±13.573.2±1.82.7±1.111.6±11.1
", + "bbox": [ + 217, + 348, + 785, + 467 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "6.3 Reward Analysis", + "text_level": 1, + "bbox": [ + 214, + 500, + 401, + 513 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "To quantitatively verify that CLIP4MC captures the size of the target entity specified in the language prompt, we collect 5000 steps in task hunt a cow and apply the method described in Appendix A to estimate the maximal size of the cow in consecutive 16 frames. Then we use CLIP4MC, MineCLIP [ours], and MineCLIP [official] to calculate intrinsic rewards respectively. Before calculating the correlation between the size and intrinsic rewards, we transform the size value using $f(x) = \\ln (x + e^{-2})$ , focusing on smaller values. The relationship between the transformed size and intrinsic rewards is visualized in Figure 4. The corresponding Pearson correlation coefficients from left to right are 0.81, 0.66, and 0.62, indicating that CLIP4MC reward has a higher correlation with the size, especially when it is relatively small. This is crucial in RL, as the agent needs dense and distinguishing reward signals to guide the learning process, particularly when the target is distant. Such characteristic is the essential benefit of CLIP4MC in RL.", + "bbox": [ + 212, + 527, + 787, + 739 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "6.4 Ablation on Dataset Filtering", + "text_level": 1, + "bbox": [ + 214, + 765, + 506, + 780 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "To evaluate the influence of the dataset quality on RL training and verify the effectiveness of our proposed correlation filter at the local level, we compare MineCLIP [ours] with two ablations: (1) MineCLIP [w/o LCF] is trained", + "bbox": [ + 212, + 794, + 784, + 840 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "CLIP4MC", + "bbox": [ + 658, + 113, + 730, + 126 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "on the dataset processed with only global correlation filtering, omitting the local correlation filtering. (2) MineCLIP [RS] is trained on 640K video clips randomly sampled from the MineDojo released database, consistent with their stated dataset construction method [8]. Note that methods evaluated here do not apply random swap introduced in Section 5.2 since these datasets cannot provide entity size. The results are presented in Table 2. Given the overall performance across all three presented tasks, firstly, [RS] does not perfectly reproduce the performance of [official], suggesting that misalignment in the original database hinders the reproduction of [official]. Secondly, part of our proposed data filtering makes the model, [w/o LCF], comparable to [official]. In addition, with all of our proposed data filtering techniques, [ours] outperforms [official].", + "bbox": [ + 212, + 146, + 787, + 313 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "6.5 Video-Text Retrieval Results", + "text_level": 1, + "bbox": [ + 214, + 334, + 501, + 349 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Table 3 shows the results of video-to-text retrieval and text-to-video on the test set with the same MineCLIP model for a fair comparison. We train these models for 20 epochs and select the models with the highest R@1 value on the test set, respectively. From the results, the model trained on the randomly sampled dataset gets the worst performance than those trained on our YouTube dataset.", + "bbox": [ + 212, + 359, + 460, + 541 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "The results demonstrate that our neat dataset indeed can facilitate table observation is that [ours] a lower performance on retrieval to the video-text alignment objective [w/o LCF] is trained on the data which is directly conducted at the the dataset further filtered by local signing for RL training. Note that for better training the VLM model", + "bbox": [ + 212, + 542, + 460, + 693 + ], + "page_idx": 13 + }, + { + "type": "table", + "img_path": "images/b3d43eac0b62f5f9e4505d62be3aa52d85bd7da4fc77f4307099a8ff0e3a9df1.jpg", + "table_caption": [ + "Table 3: Results of video-to-text / text-to-video retrieval on the test set. The best results are highlighted in bold." + ], + "table_footnote": [], + "table_body": "
ModelsR@1R@5R@10
ours12.4/13.127.5/27.835.3/35.9
w/o LCF12.7/14.129.4/29.737.9/37.5
RS11.1/11.624.8/25.032.4/32.2
", + "bbox": [ + 475, + 446, + 771, + 530 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "7 Conclusion", + "text_level": 1, + "bbox": [ + 214, + 715, + 359, + 732 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "We construct a neat YouTube dataset based on the large-scale YouTube database provided by MineDojo. Moreover, we introduce a novel cross-modal contrastive learning framework architecture, CLIP4MC, to learn an RL-friendly VLM that serves as an intrinsic reward function for open-ended tasks. Our findings suggest that our dataset enhances the acquisition of fundamental game concepts and CLIP4MC delivers a more effective reward signal for RL training.", + "bbox": [ + 212, + 748, + 787, + 840 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "H. Jiang et al.", + "bbox": [ + 271, + 114, + 367, + 128 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 217, + 145, + 401, + 162 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "This work was supported by NSFC under grant 62250068. The authors would like to thank the anonymous reviewers for their valuable comments.", + "bbox": [ + 215, + 178, + 784, + 208 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 234, + 321, + 250 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Bain, M., Nagrani, A., Varol, G., Zisserman, A.: Frozen in time: A joint video and image encoder for end-to-end retrieval. In: IEEE/CVF International Conference on Computer Vision (ICCV) (2021)", + "2. Baker, B., Akkaya, I., Zhokhov, P., Huizinga, J., Tang, J., Ecoffet, A., Houghton, B., Sampedro, R., Clune, J.: Video pretraining (VPT): learning to act by watching unlabeled online videos. arXiv preprint arXiv:2206.11795 (2022)", + "3. Baumli, K., Baveja, S., Behbahani, F., Chan, H., Comanici, G., Flennerhag, S., Gazeau, M., Holsheimer, K., Horgan, D., Laskin, M., et al.: Vision-language models as a source of rewards. arXiv preprint arXiv:2312.09187 (2023)", + "4. Cai, S., Wang, Z., Ma, X., Liu, A., Liang, Y.: Open-world multi-task control through goal-aware representation learning and adaptive horizon prediction. arXiv preprint arXiv:2301.10034 (2023)", + "5. Devlin, J., Chang, M.W., Lee, K., Toutanova, K.: Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805 (2018)", + "6. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., Uszkoreit, J., Houlsby, N.: An image is worth 16x16 words: Transformers for image recognition at scale. In: International Conference on Learning Representations (ICLR) (2021)", + "7. Dzabraev, M., Kalashnikov, M., Komkov, S., Petiushko, A.: MDMMT: multidomain multimodal transformer for video retrieval. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops (2021)", + "8. Fan, L., Wang, G., Jiang, Y., Mandlekar, A., Yang, Y., Zhu, H., Tang, A., Huang, D.A., Zhu, Y., Anandkumar, A.: Minedojo: Building open-ended embodied agents with internet-scale knowledge. In: Neural Information Processing Systems (NeurIPS) Datasets and Benchmarks Track (2022)", + "9. Guhr, O., Schumann, A.K., Bahrmann, F., Böhme, H.J.: Fullstop: Multilingual deep models for punctuation prediction (June 2021), http://ceur-ws.org/Vol-2957/sepp_paper4.pdf", + "10. Guss, W.H., Codel, C., Hofmann, K., Houghton, B., Kuno, N., Milani, S., Mohanty, S., Liebana, D.P., Salakhutdinov, R., Topin, N., et al.: Neurips 2019 competition: the minerl competition on sample efficient reinforcement learning using human priors. arXiv preprint arXiv:1904.10079 (2019)", + "11. Guss, W.H., Houghton, B., Topin, N., Wang, P., Codel, C., Veloso, M., Salakhutdinov, R.: Minerl: A large-scale dataset of apache demonstrations. In: International Joint Conference on Artificial Intelligence (IJCAI) (2019)", + "12. Hafner, D., Pasukonis, J., Ba, J., Lillicrap, T.P.: Mastering diverse domains through world models. arXiv preprint arXiv:2301.04104 (2023)", + "13. Haiminen, N., Gionis, A., Laasonen, K.: Algorithms for unimodal segmentation with applications to unimodality detection. Knowledge and information systems 14, 39-57 (2008)" + ], + "bbox": [ + 218, + 267, + 785, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "CLIP4MC", + "bbox": [ + 658, + 114, + 730, + 126 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 785, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "14. Johnson, M., Hofmann, K., Hutton, T., Bignell, D.: The malmo platform for artificial intelligence experimentation. In: International Joint Conference on Artificial Intelligence (IJCAI) (2016)", + "15. Kaelbling, L.P., Littman, M.L., Cassandra, A.R.: Planning and acting in partially observable stochastic domains. Artificial intelligence 101(1-2), 99-134 (1998)", + "16. Kanervisto, A., Milani, S., Ramanauskas, K., Topin, N., Lin, Z., Li, J., Shi, J., Ye, D., Fu, Q., Yang, W., Hong, W., Huang, Z., Chen, H., Zeng, G., Lin, Y., Micheli, V., Alonso, E., Fleuret, F., Nikulin, A., Belousov, Y., Svidchenko, O., Shpilman, A.: Minerl diamond 2021 competition: Overview, results, and lessons learned. arXiv preprint arXiv:2202.10583 (2022)", + "17. Li, Y., Wang, H., Duan, Y., Li, X.: Clip surgery for better explainability with enhancement in open-vocabulary tasks. arXiv preprint arXiv:2304.05653 (2023)", + "18. Lin, Z., Li, J., Shi, J., Ye, D., Fu, Q., Yang, W.: Juewu-mc: Playing minecraft with sample-efficient hierarchical reinforcement learning. arXiv preprint arXiv:2112.04907 (2021)", + "19. Liu, S., Fan, H., Qian, S., Chen, Y., Ding, W., Wang, Z.: Hit: Hierarchical transformer with momentum contrast for video-text retrieval. In: IEEE/CVF International Conference on Computer Vision (ICCV) (2021)", + "20. Luo, H., Ji, L., Zhong, M., Chen, Y., Lei, W., Duan, N., Li, T.: Clip4clip: An empirical study of CLIP for end to end video clip retrieval and captioning. Neurocomputing 508, 293-304 (2022)", + "21. Ma, Y.J., Liang, W., Wang, G., Huang, D.A., Bastani, O., Jayaraman, D., Zhu, Y., Fan, L., Anandkumar, A.: Eureka: Human-level reward design via coding large language models. arXiv preprint arXiv:2310.12931 (2023)", + "22. Nottingham, K., Ammanabrolu, P., Suhr, A., Choi, Y., Hajishirzi, H., Singh, S., Fox, R.: Do embodied agents dream of pixelated sheep?: Embodied decision making using language guided world modelling. arXiv preprint arXiv:2301.12050 (2023)", + "23. Oord, A.v.d., Li, Y., Vinyals, O.: Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748 (2018)", + "24. OpenAI: Gpt-4 technical report (2023)", + "25. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., Krueger, G., Sutskever, I.: Learning transferable visual models from natural language supervision. In: International Conference on Machine Learning, ICML (2021)", + "26. Radford, A., Kim, J.W., Xu, T., Brockman, G., McLeavey, C., Sutskever, I.: Robust speech recognition via large-scale weak supervision (2022). https://doi.org/10.48550/ARXIV.2212.04356, https://arxiv.org/abs/2212.04356", + "27. Schulman, J., Wolski, F., Dhariwal, P., Radford, A., Klimov, O.: Proximal policy optimization algorithms. arXiv preprint arXiv:1707.06347 (2017)", + "28. Shah, R., Wang, S.H., Wild, C., Milani, S., Kanervisto, A., Goecks, V.G., Waytowich, N.R., Watkins-Valls, D., Prakash, B., Mills, E., Garg, D., Fries, A., Souly, A., Chan, J.S., del Castillo, D., Lieberum, T.: Retrospective on the 2021 minerl BASALT competition on learning from human feedback. In: Neural Information Processing Systems (NeurIPS) Competitions and Demonstrations Track (2021)", + "29. Shu, T., Xiong, C., Socher, R.: Hierarchical and interpretable skill acquisition in multi-task reinforcement learning. In: International Conference on Learning Representations (ICLR) (2018)", + "30. Tessler, C., Givony, S., Zahavy, T., Mankowitz, D., Mannor, S.: A deep hierarchical approach to lifelong learning in apacheft. In: AAAI conference on artificial intelligence (AAAI) (2017)" + ], + "bbox": [ + 215, + 147, + 784, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "H. Jiang et al.", + "bbox": [ + 271, + 114, + 367, + 128 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "31. Wang, G., Xie, Y., Jiang, Y., Mandlekar, A., Xiao, C., Zhu, Y., Fan, L., Anandkumar, A.: Voyager: An open-ended embodied agent with large language models. arXiv preprint arXiv:2305.16291 (2023)", + "32. Xu, H., Ghosh, G., Huang, P.Y., Okhonko, D., Aghajanyan, A., Metze, F., Zettle-moyer, L., Feichtenhofer, C.: Videoclip: Contrastive pre-training for zero-shot video-text understanding. arXiv preprint arXiv:2109.14084 (2021)", + "33. Yuan, H., Zhang, C., Wang, H., Xie, F., Cai, P., Dong, H., Lu, Z.: Plan4mc: Skill reinforcement learning and planning for open-world minecraft tasks. arXiv preprint arXiv:2303.16563 (2023)", + "34. Zhou, C., Loy, C.C., Dai, B.: Extract free dense labels from clip. In: European Conference on Computer Vision. pp. 696-712. Springer (2022)", + "35. Zhu, X., Chen, Y., Tian, H., Tao, C., Su, W., Yang, C., Huang, G., Li, B., Lu, L., Wang, X., et al.: Ghost in the minecraft: Generally capable agents for open-world enviroments via large language models with text-based knowledge and memory. arXiv preprint arXiv:2305.17144 (2023)" + ], + "bbox": [ + 212, + 146, + 787, + 356 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "CLIP4MC", + "bbox": [ + 658, + 114, + 730, + 126 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + } +] \ No newline at end of file diff --git a/2024/Reinforcement Learning Friendly Vision-Language Model for Minecraft/c4770d1a-703c-494e-ba79-18744626ca70_model.json b/2024/Reinforcement Learning Friendly Vision-Language Model for Minecraft/c4770d1a-703c-494e-ba79-18744626ca70_model.json new file mode 100644 index 0000000000000000000000000000000000000000..0fe8eb303be3911bfc613c32f84ebd7dc77d7904 --- /dev/null +++ b/2024/Reinforcement Learning Friendly Vision-Language Model for Minecraft/c4770d1a-703c-494e-ba79-18744626ca70_model.json @@ -0,0 +1,2225 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.288, + 0.141, + 0.717, + 0.187 + ], + "angle": 0, + "content": "Reinforcement Learning Friendly Vision-Language Model for Minecraft" + }, + { + "type": "text", + "bbox": [ + 0.333, + 0.212, + 0.67, + 0.245 + ], + "angle": 0, + "content": "Haobin Jiang\\(^{1\\star}\\), Junpeng Yue\\(^{1\\star}\\), Hao Luo\\(^{1}\\), Ziluo Ding\\(^{2}\\), and Zongqing Lu\\(^{1,2\\dagger}\\)" + }, + { + "type": "text", + "bbox": [ + 0.337, + 0.255, + 0.666, + 0.298 + ], + "angle": 0, + "content": "\\(^{1}\\) School of Computer Science, Peking University \\(^{2}\\) Beijing Academy of Artificial Intelligence {haobin.jiang,zongqing.lu}@pku.edu.cn" + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.334, + 0.744, + 0.613 + ], + "angle": 0, + "content": "Abstract. One of the essential missions in the AI research community is to build an autonomous embodied agent that can achieve high-level performance across a wide spectrum of tasks. However, acquiring or manually designing rewards for all open-ended tasks is unrealistic. In this paper, we propose a novel cross-modal contrastive learning framework architecture, CLIP4MC, aiming to learn a reinforcement learning (RL) friendly vision-language model (VLM) that serves as an intrinsic reward function for open-ended tasks. Simply utilizing the similarity between the video snippet and the language prompt is not RL-friendly since standard VLMs may only capture the similarity at a coarse level. To achieve RL-friendliness, we incorporate the task completion degree into the VLM training objective, as this information can assist agents in distinguishing the importance between different states. Moreover, we provide neat YouTube datasets based on the large-scale YouTube database provided by MineDojo. Specifically, two rounds of filtering operations guarantee that the dataset covers enough essential information and that the video-text pair is highly correlated. Empirically, we demonstrate that the proposed method achieves better performance on RL tasks compared with baselines. The code and datasets are available at https://github.com/PKU-RL/CLIP4MC." + }, + { + "type": "text", + "bbox": [ + 0.263, + 0.624, + 0.711, + 0.641 + ], + "angle": 0, + "content": "Keywords: Dataset \\(\\cdot\\) Multimodal model \\(\\cdot\\) Reinforcement learning" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.665, + 0.377, + 0.681 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.696, + 0.788, + 0.817 + ], + "angle": 0, + "content": "Training reinforcement learning (RL) agents to perform complex tasks in a vision-based and open-ended world can be difficult. One main challenge is that manually specifying reward functions in all open-ended tasks is unrealistic [3,21], especially when we cannot access the internal configuration of the environment. In addition, learning a reward model from human feedback is typically expensive. To this end, MineDojo [8] has proposed an internet-scale, multi-modal knowledge base of YouTube videos to facilitate learning in an open-ended world. With the advent of a such large-scale database, agents are able to harvest practical" + }, + { + "type": "page_footnote", + "bbox": [ + 0.232, + 0.825, + 0.544, + 0.841 + ], + "angle": 0, + "content": "* Equal contribution. † Corresponding author." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.369, + 0.129 + ], + "angle": 0, + "content": "H. Jiang et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.239 + ], + "angle": 0, + "content": "knowledge encoded in large amounts of media like human beings. Moreover, a vision-language model (VLM), MineCLIP [8], is proposed to utilize the internet-scale domain knowledge. In more detail, the learned correlation score between the visual observation and the language prompt can be used effectively as an open-vocabulary, massively multi-task reward function for RL training. Therefore, no further task-specific reward design is needed for open-ended tasks." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.242, + 0.79, + 0.483 + ], + "angle": 0, + "content": "However, the autonomous embodied agent requires the vision-language model to provide a more instructive correlation score. Given the partial observations, e.g., video snippet, and the language prompt that describes the task, the agent needs to figure out two non-trivial matters to better evaluate the current state. On the one hand, whether the target entities are present within its field of vision? MineCLIP tried to address this question. However, the alignment of texts and videos in the YouTube database is totally a catastrophe, which impedes the learning of VLM. On the other hand, what is the relationship between each video snippet and the degree of completion of the task? Normally, in an open-ended world, e.g. Minecraft, the agent explores first, then approaches and interacts with the target object. In other words, the agent requires approaching the target object before it takes trial-and-error, even for the target that needs to be kept away. Therefore, it is reasonable to make an assumption, namely, that the higher the level of completion of the task, the closer the agent is to the targets in the video snippet. We also argue it is important to incorporate the level of completion of the task into the reward function." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.489, + 0.79, + 0.701 + ], + "angle": 0, + "content": "In this paper, we first construct neat YouTube datasets to facilitate the learning of basic game concepts, mainly the correspondence between the videos and texts. Though a large-scale database is provided by MineDojo, it contains significant noise due to its nature as an online resource. In addition, MineDojo only claims the training dataset is randomly sampled from the database, making it hard to reproduce. To overcome the catastrophic misalignment of video-text pairs in the original database, we have done four steps of dataset processing to guarantee the dataset is clean. Firstly, transcript cleaning enhances the accuracy of transcripts and ensures they are complete sentences. Secondly, keyword filtering ensures that the content of clips is relevant to the key entities in Minecraft, thereby facilitating the learning of basic game concepts. Thirdly, video partitioning and selection can handle the issue of scene transitions and thereby mitigate interference from other extraneous information. Lastly, correlation filtering can effectively address the issue of mismatch between video clips and transcripts." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.705, + 0.79, + 0.842 + ], + "angle": 0, + "content": "We also propose an upgraded vision-language model, CLIP4MC, to provide a more RL-friendly reward function. In RL, simply utilizing the similarity between the video snippet and the language prompt is not RL-friendly since MineCLIP tends to only capture the similarity at the entity level. In other words, VLM can hardly reflect the relationship between each video snippet and the degree of completion of the task. However, this information can better help agents distinguish the importance between similar states. To achieve this, we incorporate the degree of task completion into the VLM training objective. In more detail, CLIP has exhibited remarkable segmentation capabilities without fine-tuning. After we" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.66, + 0.114, + 0.732, + 0.128 + ], + "angle": 0, + "content": "CLIP4MC" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.787, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.313 + ], + "angle": 0, + "content": "extend pre-trained MineCLIP with modifications inspired by MaskCLIP [34], it can segment the specified object from the image and label the size of the target shown in the corresponding video. Intuitively, the closer the agent is to the target, the larger the target size becomes. During the learning procedure, we dynamically control the degree of contrasting positive and negative pairs of instances based on the target size in this positive video sample. Thus, CLIP4MC can render a more RL-friendly reward signal that instructs the agent to learn tasks faster. Our proposed method is trained on our YouTube dataset and evaluated on MineDojo Programmatic tasks, including harvest, hunt, and combat tasks. Empirically, our results show that CLIP4MC can provide a more friendly reward signal for the RL training procedure." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.314, + 0.561, + 0.327 + ], + "angle": 0, + "content": "To summarize, our contributions are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.336, + 0.784, + 0.41 + ], + "angle": 0, + "content": "- Open-sourced datasets: We provide two high-quality datasets. The first one undergoes data cleaning (Sections 4.1 to 4.3) and global-level correlation filtering (Section 4.4). The VLM trained on this dataset matches the performance of the officially released MineCLIP, which lacks a publicly available training set." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.411, + 0.784, + 0.455 + ], + "angle": 0, + "content": "- RL-friendlyly dataset: Our second dataset further incorporates local-level correlation filtering (Section 4.4), making it more suited for RL. The VLM trained on this dataset outperforms that on the first dataset." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.456, + 0.784, + 0.501 + ], + "angle": 0, + "content": "- RL-friendly VLM: To better evaluate and leverage the advantages of our more RL-friendly dataset, we introduce CLIP4MC, a novel method to train a VLM that could improve downstream RL performance (Section 5.2)." + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.336, + 0.784, + 0.501 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.523, + 0.388, + 0.539 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.553, + 0.788, + 0.658 + ], + "angle": 0, + "content": "Video-Text Retrieval. Video-text retrieval plays an essential role in multimodal research and has been widely used in many real-world web applications. Recently, the pre-trained models have dominated this line of research with noticeable results on both zero-shot and fine-tuned retrieval. Especially, BERT [5], ViT [6], and CLIP [25], are used as the backbones to extract the text or video embedding. The cross-modal embeddings are then matched with specific fusion networks to find the correct video-text pair." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.659, + 0.788, + 0.81 + ], + "angle": 0, + "content": "In more detail, CLIP4Clip [20] proposes three different similarity modules to calculate the correlation between video and text embeddings. HiT [19] performs hierarchical matching at two different levels, i.e. semantic level and feature level. Note that semantic level and feature level features are from the transformer network's higher and lower feature layers, respectively. Frozen [1] proposes a dual encoder architecture that utilizes the flexibility of a transformer visual encoder to train from images or video clips with text captions. Moreover, MDMMT [7] adopts several pre-training models as encoders and it shows the CLIP-based model performs the best. Therefore, our model follows this line of research by using the pre-trained model, CLIP [25], to extract the feature embeddings." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.81, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Minecraft for AI Research. As an open-ended video game with an egocentric vision, Minecraft is a noticeable and important domain in RL due to the nature" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.232, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.37, + 0.129 + ], + "angle": 0, + "content": "H. Jiang et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.237 + ], + "angle": 0, + "content": "of the sparse reward, large exploration space, and long-term episodes. Since the release of the Malmo simulator [14] and later the MineDojo simulator [8], various methods have attempted to train agents to complete tasks in Minecraft [10, 12, 18, 29, 30]. Approaches such as model-based RL, hierarchical RL, goal-based RL, and reward shaping have been adopted to alleviate the sparse reward and exploration difficulty for the agent in this environment." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.238, + 0.788, + 0.373 + ], + "angle": 0, + "content": "Recently, with the development of large language models (LLM) like GPT-4 [24], a series of methods leveraging LLMs for high-level planning in Minecraft have been proposed [16, 22, 31, 33, 35]. These methods have demonstrated remarkable capabilities in guiding the agent to complete multiple complicated, long-horizon tasks, such as mining diamonds. These LLMs play a crucial role in decision-making, determining the sequence of basic skills required to accomplish specific tasks. Their effectiveness is due to their extensive knowledge about Minecraft, learned from the Internet, and their ability to reflect on real-time feedback from the game environment." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.374, + 0.788, + 0.525 + ], + "angle": 0, + "content": "In addition to the use of LLMs, recent research attempts to incorporate Internet visual data into basic skill learning in Minecraft, beyond the traditional RL methods. MineRL [11] collected 60M player demonstrations with action labels, motivating some methods [16,28] based on behavior cloning. As well-labeled data is limited in quantity, MineCLIP [8] instead uses over 730K narrated Minecraft videos without action labels from YouTube. It aims to learn a vision-language model providing auxiliary reward signals, utilizing the vast and diverse data available on the Internet. Different from MineCLIP, VPT [2] uses action-labeled data to train an inverse dynamic model to label 70K hours of Internet videos and then conduct behavior cloning." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.526, + 0.788, + 0.603 + ], + "angle": 0, + "content": "Unlike the existing approaches, which incorporate the human experience and require a large number of demonstrations with action labels to train the agent, our work follows the line of MineCLIP [8] and focuses on only using the data without action labels to assist agent learning in Minecraft, which is more friendly with data collection and has the potential to scale in the future." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.626, + 0.371, + 0.645 + ], + "angle": 0, + "content": "3 Background" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.659, + 0.788, + 0.749 + ], + "angle": 0, + "content": "MineDojo tasks. MineDojo [8] provides thousands of benchmark tasks, which can be used to develop generally capable agents in Minecraft. These tasks can be divided into two categories, Programmatic and Creative tasks. The former has ground-truth simulator states to assess whether the task has been completed. The latter, however, do not have well-defined success criteria and tend to be more open-ended, but have to be evaluated by humans." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.841 + ], + "angle": 0, + "content": "We mainly focus on Programmatic tasks since they can be automatically assessed. Specifically, MineDojo provides 4 categories of programmatic tasks, including Harvest, Combat, Survival, and Tech Tree, with 1581 template-generated natural language goals to evaluate the agent's different capabilities. Among these tasks, Survival and Tech Tree tasks are harder than Harvest and Combat tasks. Currently, MineCLIP [8] only expresses promising potential in some Harvest and" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.66, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "CLIP4MC" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "image", + "bbox": [ + 0.264, + 0.147, + 0.61, + 0.205 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.648, + 0.149, + 0.735, + 0.157 + ], + "angle": 0, + "content": "irrelevant video content" + }, + { + "type": "text", + "bbox": [ + 0.615, + 0.161, + 0.732, + 0.192 + ], + "angle": 0, + "content": "... so here I'm with nick and we've got look at my inventory and his inventory looks similar this is gonna be interesting so four..." + }, + { + "type": "image", + "bbox": [ + 0.264, + 0.213, + 0.61, + 0.268 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.642, + 0.212, + 0.734, + 0.22 + ], + "angle": 0, + "content": "mismatched video content" + }, + { + "type": "text", + "bbox": [ + 0.615, + 0.225, + 0.731, + 0.256 + ], + "angle": 0, + "content": "... give it I'm a little chicken just laying around by any chance a lot of arrows going on down here I'm guessing most of ..." + }, + { + "type": "image", + "bbox": [ + 0.264, + 0.275, + 0.61, + 0.33 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.655, + 0.275, + 0.735, + 0.282 + ], + "angle": 0, + "content": "matched video content" + }, + { + "type": "text", + "bbox": [ + 0.615, + 0.288, + 0.733, + 0.319 + ], + "angle": 0, + "content": "... at some point you will lose there too or also hearts there spider was like being completely blown away! I don't assume anything at ..." + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.348, + 0.785, + 0.402 + ], + "angle": 0, + "content": "Fig. 1: Illustration of the YouTube video database. The screenshots of video clips are on the left and key entities are circled in red. The corresponding transcript clips are on the right and key entities are marked in red. We give examples of irrelevant, mismatched, and matched video content in the YouTube video database." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.434, + 0.785, + 0.478 + ], + "angle": 0, + "content": "Combat tasks. Harvest means finding, obtaining, cultivating, or manufacturing hundreds of materials and objects. Combat means fighting various monsters and creatures that require fast reflexes and martial skills." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.48, + 0.787, + 0.618 + ], + "angle": 0, + "content": "POMDP. We model the programmatic task as a partially observable Markov decision process (POMDP) [15]. At each timestep \\( t \\), the agent obtains the partial observation \\( o_t \\) from the global state \\( s_t \\) and a language prompt \\( G \\), takes action \\( a_t \\) following its policy \\( \\pi(a_t|o_t) \\), and receives a reward \\( r_t = \\Phi(V_t, G) \\), where \\( V_t \\) is the fixed-length sequence of observations till \\( t \\) (thus a video snippet) and \\( \\Phi \\) maps \\( V_t \\) and \\( G \\) to a scalar value. Then the environment transitions to the next state \\( s_{t+1} \\) given the current state and action according to transition probability function \\( \\mathcal{T}(s_{t+1}|s_t, a_t) \\). The agent aims to maximize the expected return \\( R = \\mathbb{E}_{\\pi} \\sum_{t=1}^{T} \\gamma^{t-1} r_t \\), where \\( \\gamma \\) is the discount factor and \\( T \\) is the episode time horizon." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.642, + 0.421, + 0.657 + ], + "angle": 0, + "content": "4 YouTube Dataset" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.788, + 0.841 + ], + "angle": 0, + "content": "The Internet is rich in Minecraft-related data, containing a wealth of weakly labeled or even unlabeled Minecraft knowledge, including crucial entities, plausible actions, and common-sense event processes. With these multi-modal data, it is possible to create dense language-conditioned rewards, making open-ended long-term task learning feasible. MineDojo [8] collected over 730K YouTube videos and their corresponding transcripts, totaling 33 years of video content and 2.2B words in transcripts. Around 640K video clips are selected using keyword-based filtering from these videos, and these clips are used to train MineCLIP. However, MineDojo open-sourced a 13.8M dataset, and the 640K video clips were randomly sampled in their paper. However, as illustrated in Figure 1, most videos feature irrelevant game content that is not conducive to learning basic game" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.369, + 0.129 + ], + "angle": 0, + "content": "H. Jiang et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.239 + ], + "angle": 0, + "content": "concepts. Meanwhile, the alignment between the transcripts and videos may not always be precise, leading to temporal or content discrepancies that could hinder the learning of retrieval and RL tasks. Given the low quality of the data, we found it necessary to provide a neat 640k dataset to the community for the training usage. To address these issues, we adopt the following few steps to acquire a clean dataset with high quality." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.259, + 0.428, + 0.274 + ], + "angle": 0, + "content": "4.1 Transcript Cleaning" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.283, + 0.789, + 0.435 + ], + "angle": 0, + "content": "Downloading YouTube's automatic transcripts directly can lead to several issues. Firstly, transcript blocks may overlap, resulting in overlapping timestamps in the transcript. Additionally, the caption quality is typically mediocre, with a higher occurrence of misidentifications, especially in non-English videos. Moreover, the transcript lacks punctuation, making it less friendly for understanding semantics. Based on the aforementioned issues, we implement a pipeline to construct high-quality transcripts as follows: (1) Extract audio from the videos and use Whisper [26] to obtain high-quality, temporally non-overlapping transcripts. (2) Employ FullStop [9] to generate punctuation, resulting in complete sentences of 10-35 words in length." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.455, + 0.416, + 0.47 + ], + "angle": 0, + "content": "4.2 Keyword Filtering" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.478, + 0.789, + 0.614 + ], + "angle": 0, + "content": "Following MineDojo [8], we also implement keyword-based filtering to ensure that the content in our dataset is pertinent to the key entities in Minecraft, thereby facilitating the learning of basic game concepts. As essential components of Minecraft, the key entities, such as stones, trees, and sheep, are common across multiple tasks in MineDojo and videos from YouTube. Specifically, we identify entity keywords in the transcripts using a keyword list from MineDojo and extract transcript clips formed into sentences to encompass as many keywords as possible. These extracted transcript clips then serve as the textual component of our dataset, determining the location of corresponding video clips." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.635, + 0.507, + 0.65 + ], + "angle": 0, + "content": "4.3 Video Partition and Selection" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.66, + 0.788, + 0.794 + ], + "angle": 0, + "content": "After completing the previous two steps, we obtain transcript clips relevant to the keywords. For each transcript clip, we calculate the central timestamp that corresponds to the clip based on the transcript timestamps and then use this central timestamp to extract a video clip with a duration of \\( D \\) seconds from the video. This process allows us to obtain temporally-aligned video clips. However, the video clips obtained in this manner still exhibit some issues, including scene transitions and discrepancies between video content and transcripts. We handle the former problem through video partition and filtering, while the latter problem is addressed in Section 4.4." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.796, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Owing to the informal nature of YouTube content, there is often a lack of semantic congruence between the video clips and their corresponding transcriptions, as noted in VideoCLIP [32]. Moreover, video clips frequently contain a" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.66, + 0.114, + 0.732, + 0.127 + ], + "angle": 0, + "content": "CLIP4MC" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.787, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "image", + "bbox": [ + 0.331, + 0.144, + 0.675, + 0.314 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.324, + 0.788, + 0.368 + ], + "angle": 0, + "content": "Fig. 2: Examples of how we estimate the size of key entities in video frames. Red bounding boxes are generated by our modified MineCLIP visual encoder, following the approach proposed in MaskCLIP [34]. These boxes are then used to calculate the size." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.397, + 0.785, + 0.472 + ], + "angle": 0, + "content": "few different behaviors which cause scene transitions, e.g. Chopping down the tree first, then suddenly switching the inventory bar. Since some scene transitions lead to irrelevant information, we partition the video content into several semantically coherent segments based on the semantic structure of the video. Then we select the segment that aligns best with the transcript." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.473, + 0.788, + 0.593 + ], + "angle": 0, + "content": "To achieve semantic partition of video content, we employ the Bellman \\(k\\)-segmentation algorithm [13]. This algorithm divides a sequence of data points into \\(k\\) distinct and constant-line segments, providing a piece-wise constant approximation of the data. To process the video, we first use the officially released MineCLIP [8] video encoder to obtain the embedding of each frame, since MineCLIP can capture video semantics to some extent. Subsequently, we partition these embeddings into \\(k\\) segments and select the segment with the highest similarity score, as calculated in MineCLIP." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.617, + 0.436, + 0.634 + ], + "angle": 0, + "content": "4.4 Correlation Filtering" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.644, + 0.785, + 0.718 + ], + "angle": 0, + "content": "We employ correlation filtering techniques to address disparities that persist between video content and transcripts. The correlation filtering is done at two levels, the global and the local levels. From the global level, we calculate the cosine similarities between video embeddings and text embeddings via the original MineCLIP and then select clips based on these similarities." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Recent research [17, 34] has demonstrated that CLIP [25], though trained on whole images, can generate meaningful local features. Inspired by this and following MaskCLIP [34], we make modifications to the original MineCLIP visual encoder, empowering it with the ability to estimate and label the size of the key entity, which is mentioned in the corresponding transcript, in each frame of a video clip without fine-tuning. Then the correlations between video clips and transcripts are calculated as the summation of the sizes across all frames in each video clip. We consider this correlation at the local level." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.369, + 0.129 + ], + "angle": 0, + "content": "H. Jiang et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.192 + ], + "angle": 0, + "content": "Based on these two criteria, we select the top \\( k\\% \\) of clips, resulting in the final training set. We provide an elaboration on our implementation in Appendix A and illustrate some examples in Figure 2." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.193, + 0.788, + 0.314 + ], + "angle": 0, + "content": "The aforementioned four-step approach creates a dataset consisting of 640K video-text clip pairs, with an additional 4K pairs extracted for validation of video-text retrieval. Regarding the constants of the approach, \\( D \\) and \\( k \\) are set to 16 and 50, respectively. Therefore, our dataset comprises videos with a total duration of one week and approximately 0.16B words, significantly smaller in scale compared to the original low-quality 13.8M dataset. Importantly, we will open-source our entire YouTube dataset, serving as an upgraded version of the unreleased MineCLIP training data." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.337, + 0.355, + 0.352 + ], + "angle": 0, + "content": "5 CLIP4MC" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.369, + 0.785, + 0.459 + ], + "angle": 0, + "content": "Given a video snippet \\( V \\) and a language prompt \\( G \\), the vision-language model outputs a correlation score, \\( C \\), that measures the similarity between the video snippet and the language prompt. Ideally, if the agent performs behaviors following the description of the language prompt, the vision-language model will generate a higher correlation score, leading to a higher reward. Otherwise, the agent will be given a lower reward." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.482, + 0.367, + 0.495 + ], + "angle": 0, + "content": "5.1 Architecture" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.508, + 0.784, + 0.597 + ], + "angle": 0, + "content": "We follow the same architecture design as MineCLIP [8], including a video encoder, a text encoder, and a similarity calculator. All the video frames first go through the spatial transformer to obtain a sequence of frame features. The temporal transformer is then utilized to summarize the sequence of frame features into a single video embedding. An adapter further processes the video embedding for better features. Refer to Appendix B for more details about the architecture." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.599, + 0.784, + 0.675 + ], + "angle": 0, + "content": "Empirically, we found that the video encoder (essentially MineCLIP) can provide a bond between the entities and the language prompts and give a similar high reward as long as the target entities are present in the video frames, similar to observations in [4]. However, such a reward is not instructive enough for RL tasks since it does not reflect the behavioral trends of agents." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.696, + 0.434, + 0.712 + ], + "angle": 0, + "content": "5.2 Contrastive Training" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.722, + 0.784, + 0.751 + ], + "angle": 0, + "content": "We aim to minimize the sum of the multi-modal contrastive losses [23], including video-to-text, and text-to-video:" + }, + { + "type": "equation", + "bbox": [ + 0.299, + 0.764, + 0.785, + 0.799 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = - \\sum_ {\\left(z _ {v}, z _ {m}, z _ {t}\\right) \\in \\mathcal {B}} \\left(\\log \\operatorname {N C E} \\left(z _ {v}, z _ {t}\\right) + \\log \\operatorname {N C E} \\left(z _ {t}, z _ {v}\\right)\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.784, + 0.84 + ], + "angle": 0, + "content": "where \\(\\mathcal{B}\\) is the batch containing sampled video-text pairs, and \\(\\mathrm{NCE}(\\cdot ,\\cdot)\\) is the contrastive loss that calculates the similarity of two inputs. To illustrate, the" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.66, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "CLIP4MC" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "image", + "bbox": [ + 0.315, + 0.147, + 0.695, + 0.43 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.442, + 0.785, + 0.471 + ], + "angle": 0, + "content": "Fig. 3: Illustration of CLIP4MC training. The upper part shows the concept of contrastive learning, while the lower part explains the swapping operation." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.499, + 0.507, + 0.514 + ], + "angle": 0, + "content": "video-to-text contrastive loss is given by" + }, + { + "type": "equation", + "bbox": [ + 0.352, + 0.522, + 0.785, + 0.56 + ], + "angle": 0, + "content": "\\[\n\\mathrm {N C E} \\left(z _ {v}, z _ {t}\\right) = \\frac {\\exp \\left(z _ {v} \\cdot z _ {t} ^ {+} / \\lambda\\right)}{\\sum_ {z \\in \\left\\{z _ {t} ^ {+} , z _ {t} ^ {-} \\right\\}} \\exp \\left(z _ {v} \\cdot z / \\lambda\\right)}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.569, + 0.785, + 0.629 + ], + "angle": 0, + "content": "where \\(\\lambda\\) is a temperature hyperparameter, \\(z_{t}^{+}\\) is the positive text embedding matching with the video embedding \\(z_{v}\\), and \\(\\{z_t^-\\}\\) are negative text embeddings that are implicitly formed by other text clips in the training batch. Contrastive loss of text-to-video is defined in the same way." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.63, + 0.784, + 0.75 + ], + "angle": 0, + "content": "In addition, we also need to incorporate the degree of task completion into the training objective. Specifically, we hope the similarity score between video clip and text embeddings could reflect the task completion degree, i.e., higher completion leads to higher similarity. The sizes of key entities provided by the local correlation filtering serve as a surrogate for the task completion degree. Initially, we try to dynamically adjust the weight of the positive pairs based on the size of the target object. However, it did not work. Essentially, positive and negative pairs will still be separated, even with smaller sample weights." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.785, + 0.84 + ], + "angle": 0, + "content": "Therefore, we instead forcibly change the labels of the positive and negative video samples, as illustrated in Figure 3, thereby reducing the model's confidence in certain positive pairs. Specifically, during training, a positive video sample is swapped with a random negative video sample from the \\(\\mathcal{B}\\) based on probability \\(P_{\\mathrm{swap}}\\). This swap only occurs when the labeled target size \\(H_bW_b\\) in the positive video sample is below a certain threshold \\(\\tau\\), and the smaller the size, the greater" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.369, + 0.129 + ], + "angle": 0, + "content": "H. Jiang et al." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.147, + 0.332, + 0.162 + ], + "angle": 0, + "content": "the probability," + }, + { + "type": "equation", + "bbox": [ + 0.371, + 0.162, + 0.786, + 0.194 + ], + "angle": 0, + "content": "\\[\nP _ {\\mathrm {s w a p}} = P _ {\\mathrm {m a x}} * \\mathrm {R e L U} \\left(1 - \\frac {H _ {b} W _ {b}}{\\tau H W}\\right), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.196, + 0.535, + 0.21 + ], + "angle": 0, + "content": "where \\( HW \\) represents the size of the image." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.211, + 0.787, + 0.3 + ], + "angle": 0, + "content": "Since contrastive learning brings positive samples closer and pushes negative samples away, converting a positive pair to a negative one with a certain probability, i.e., the aforementioned random swap, during training will lower the similarity score. As the decrease in similarity is directly proportional to the swapping probability, we set \\( P_{\\mathrm{swap}} \\) to be inversely related to task completion, ensuing similarity score increases as completion increases." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.301, + 0.787, + 0.377 + ], + "angle": 0, + "content": "In addition, the upper limit \\( P_{\\mathrm{max}} \\) in Equation (3) is set to 0.5, as the swap should not disrupt severely the normal distinction between positive and negative pairs in most cases. This constraint enables CLIP4MC to preserve the understanding capability of MineCLIP for general behaviors that may lack explicit target entities in Minecraft, as discussed in Appendix F." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.395, + 0.365, + 0.411 + ], + "angle": 0, + "content": "5.3 RL Training" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.416, + 0.787, + 0.537 + ], + "angle": 0, + "content": "For RL training, the first step is reward generation. At timestep \\( t \\), we concatenate the agent's latest 16 egocentric RGB frames in a temporal window to form a video snippet, \\( V_{t} \\). CLIP4MC outputs the probability \\( P_{G,t} \\) that calculates the similarity of \\( V_{t} \\) to the task prompt, \\( G \\), against all other negative prompts. To compute the reward, we further process the raw probability as previous work [8] \\( r_t = \\max (P_{G,t} - 1 / N_t,0) \\), where \\( N_{t} \\) is the number of prompts passed to CLIP4MC. Note that CLIP4MC can handle unseen language prompts without any further fine-tuning due to the open-vocabulary ability of CLIP [25]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.538, + 0.787, + 0.629 + ], + "angle": 0, + "content": "The ultimate goal is to train a policy network that takes as input raw pixels and other structural data and outputs discrete actions to accomplish the task that is described by the language prompt, \\( G \\). We use PPO [27] as our RL training backbone and the policy is trained on the CLIP4MC reward together with the sparse task reward if any. The policy input contains several modality-specific components and more details can be found in Appendix D." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.647, + 0.377, + 0.664 + ], + "angle": 0, + "content": "6 Experiments" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.787, + 0.719 + ], + "angle": 0, + "content": "In this section, we comprehensively evaluate and analyze our proposed model CLIP4MC, utilizing the open-ended platform MineDojo [8], which comprises thousands of diverse, open-ended Minecraft tasks designed for embodied agents." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.788, + 0.84 + ], + "angle": 0, + "content": "We compare CLIP4MC against two baselines: (1) MineCLIP [official], the officially released MineCLIP model [8]. (2) MineCLIP [ours], using the same architecture as MineCLIP [official] but trained on our cleaned YouTube dataset. It also serves as the ablation of CLIP4MC without the swap operation. We train CLIP4MC and MineCLIP [ours] for 20 epochs and select the models with the highest performance on RL tasks. Please refer to Appendix C for more training details. All results are presented in terms of the mean and standard error of four runs with different random seeds." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.66, + 0.114, + 0.732, + 0.128 + ], + "angle": 0, + "content": "CLIP4MC" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.128 + ], + "angle": 0, + "content": "11" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.147, + 0.444, + 0.163 + ], + "angle": 0, + "content": "6.1 Environment Settings" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.172, + 0.788, + 0.232 + ], + "angle": 0, + "content": "We conduct experiments on eight Programmatic tasks, comprising two harvest tasks: milk a cow and shear wool, two combat tasks: combat a spider and combat a zombie, and four hunt tasks: hunt a cow, hunt a sheep, hunt a pig, and hunt a chicken. These tasks are all built-in tasks in the MineDojo benchmark." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.233, + 0.788, + 0.324 + ], + "angle": 0, + "content": "Harvest. Milk a cow requires the agent to obtain milk from a cow with an empty bucket. Similarly, shear wool requires the agent to obtain wool from a sheep with shears. A harvest task is terminated and considered completed when the target item is obtained by the agent with a specified quantity. The prompts used to calculate the reward is \"obtain milk from a cow in plains with an empty bucket\"; for shear wool, it is \"shear a sheep in plains with shears\"." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.324, + 0.788, + 0.383 + ], + "angle": 0, + "content": "Combat. In these tasks, target animals, spiders, and zombies, are hostile and will actively approach and attack the agent. The agent's goal is to fight and kill the target animal. The prompt for each combat task is \"combat a spider/zombie in plains with a diamond sword\"." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.384, + 0.788, + 0.535 + ], + "angle": 0, + "content": "Hunt. Hunt tasks consist of hunt a cow, hunt a sheep, hunt a pig, and hunt a sheep. For each task, the agent's goal is to kill the target animal as indicated in the task name. Different from combat tasks, the target animals will flee from the agent after being attacked. Therefore, these tasks require the agent to keep chasing and attacking the target, making them challenging. As noted in [4], the original MineCLIP reward fails in these tasks since it cannot consistently increase when the agent approaches the agent. This observation aligns with our assertion that the original MineCLIP model can hardly capture the degree of task completion. The prompt for each task is \"hunt a {target} in plains with a diamond sword\" where {target} is replaced with the corresponding animal name." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.535, + 0.788, + 0.596 + ], + "angle": 0, + "content": "More elaborated introduction of the Minecraft environment and settings of these tasks are available in Appendix D. To guarantee a fair comparison, we adopt the same RL hyperparameters for all models and tasks. These hyperparameters are listed in Appendix E." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.618, + 0.357, + 0.632 + ], + "angle": 0, + "content": "6.2 RL Results" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.643, + 0.787, + 0.674 + ], + "angle": 0, + "content": "Through our evaluation of CLIP4MC, MineCLIP [ours], and MineCLIP [official] across eight Minecraft tasks, we want to answer two key questions:" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.675, + 0.787, + 0.703 + ], + "angle": 0, + "content": "(1) Whether the YouTube dataset we constructed enables MineCLIP, when trained on it, to provide a more effective reward signal for task learning?" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.704, + 0.786, + 0.735 + ], + "angle": 0, + "content": "(2) Whether our upgraded model, CLIP4MC, based on our YouTube dataset, offers a reward signal that is further friendly for the RL training procedure?" + }, + { + "type": "list", + "bbox": [ + 0.215, + 0.675, + 0.787, + 0.735 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.787, + 0.779 + ], + "angle": 0, + "content": "These questions are central to verifying the effectiveness of our dataset and CLIP4MC model in Minecraft. Table 1 shows the success rates of all methods on the eight Minecraft tasks." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.841 + ], + "angle": 0, + "content": "It is noticeable that, in four hunt tasks, three models demonstrate varying performance on RL. Firstly, MineCLIP [ours] consistently achieves better results compared to MineCLIP [official] across all hunt tasks. The superior performance of MineCLIP [ours] provides a positive answer to our first question, suggesting" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.369, + 0.129 + ], + "angle": 0, + "content": "H. Jiang et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.145, + 0.788, + 0.189 + ], + "angle": 0, + "content": "Table 1: Success rates \\((\\%)\\) of RL trained with rewards provided by different models on eight Minecraft tasks. Each mean and standard error of success rates are calculated on four models after training 1e6 environment steps with different random seeds." + }, + { + "type": "table", + "bbox": [ + 0.222, + 0.199, + 0.779, + 0.403 + ], + "angle": 0, + "content": "
ModelsHarvestCombat
milk a cowshear woolcombat a spidercombat a zombie
CLIP4MC84.5±2.074.6±2.185.8±0.970.4±8.3
MineCLIP[ours]84.4±1.171.6±3.575.4±10.163.6±8.7
MineCLIP[official]84.1±0.573.2±1.882.7±2.557.4±3.7
ModelsHunt
hunt a cowhunt a sheephunt a pighunt a chicken
CLIP4MC39.8±2.545.9±7.230.6±8.426.1±3.5
MineCLIP[ours]17.3±10.633.0±18.114.1±10.615.3±10.6
MineCLIP[official]11.6±11.128.5±16.71.5±0.60.0±0.0
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.439, + 0.788, + 0.576 + ], + "angle": 0, + "content": "that the YouTube dataset we construct indeed enhances the effectiveness of the reward signal in MineCLIP for task learning. Note that this is the dataset we plan to release for better training the VLM model for RL tasks on Minecraft. Secondly, CLIP4MC shows significantly higher success rates on hunt tasks compared to both MineCLIP [ours] and MineCLIP [official], meaning that the answer to the second question is also positive. As CLIP4MC provides a reward signal taking into account a surrogate for the degree of task completion, i.e., the size of the target object in our implementation, it becomes more RL-friendly in these challenging tasks." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.58, + 0.788, + 0.701 + ], + "angle": 0, + "content": "In addition, we notice a practical example of misalignment in the officially released MineCLIP model. Specifically, we observe that in hunt a chicken, the agent trained with MineCLIP [official] tends to keep looking at the sky, indicating such behavior can provide a high intrinsic reward. This phenomenon suggests that the officially released MineCLIP indeed suffers from the misalignment problem in the YouTube dataset. In contrast, our MineCLIP [official] shows promising behaviors on this task, demonstrating that our dataset processing improves the alignment between the transcripts and videos." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.705, + 0.788, + 0.842 + ], + "angle": 0, + "content": "In contrast to hunt tasks, these methods perform comparably on harvest tasks and combat tasks, while CLIP4MC still shows marginal advantage over the other two methods. This finding aligns with the results reported in [8], where the MineCLIP reward already achieves saturated performance on these easy tasks. Therefore, further enhancements in the reward signal, like MineCLIP [ours] and CLIP4MC, do not significantly improve performance. Unlike hunt tasks, neither harvest nor combat tasks require the agent to take multiple rounds of chasing. This result does not detract from the superiority of CLIP4MC, evidenced by its performance on hunt tasks, which are considered more difficult [2]." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.66, + 0.114, + 0.732, + 0.127 + ], + "angle": 0, + "content": "CLIP4MC" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "image", + "bbox": [ + 0.288, + 0.145, + 0.418, + 0.224 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.31, + 0.227, + 0.395, + 0.24 + ], + "angle": 0, + "content": "(a) CLIP4MC" + }, + { + "type": "image", + "bbox": [ + 0.437, + 0.145, + 0.569, + 0.224 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.441, + 0.227, + 0.563, + 0.24 + ], + "angle": 0, + "content": "(b) MineCLIP [ours]" + }, + { + "type": "image", + "bbox": [ + 0.587, + 0.145, + 0.718, + 0.224 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.587, + 0.227, + 0.718, + 0.24 + ], + "angle": 0, + "content": "(c) MineCLIP [official]" + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.251, + 0.784, + 0.278 + ], + "angle": 0, + "content": "Fig. 4: Scatter plots illustrating the relationship between the entity size and the intrinsic reward. The red line indicates a linear fit to the data." + }, + { + "type": "table_caption", + "bbox": [ + 0.215, + 0.295, + 0.785, + 0.337 + ], + "angle": 0, + "content": "Table 2: Success rates (%) of RL trained with rewards provided by different models on three Minecraft tasks. Each mean and standard error of success rates are calculated on four models after training 5e5 and 1e6 environment steps with different random seeds." + }, + { + "type": "table", + "bbox": [ + 0.218, + 0.349, + 0.787, + 0.468 + ], + "angle": 0, + "content": "
Modelsmilk a cowshear woolhunt a cow
5e51e65e51e65e51e6
ours73.1±0.984.4±1.147.5±6.371.6±3.515.6±8.717.3±10.6
w/o LCF71.3±1.880.7±1.153.8±3.373.6±1.50.7±0.712.1±12.0
RS65.1±3.981.7±1.021.7±6.568.5±3.61.0±0.54.0±2.6
official69.8±1.384.1±0.547.0±13.573.2±1.82.7±1.111.6±11.1
" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.5, + 0.402, + 0.515 + ], + "angle": 0, + "content": "6.3 Reward Analysis" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.529, + 0.788, + 0.74 + ], + "angle": 0, + "content": "To quantitatively verify that CLIP4MC captures the size of the target entity specified in the language prompt, we collect 5000 steps in task hunt a cow and apply the method described in Appendix A to estimate the maximal size of the cow in consecutive 16 frames. Then we use CLIP4MC, MineCLIP [ours], and MineCLIP [official] to calculate intrinsic rewards respectively. Before calculating the correlation between the size and intrinsic rewards, we transform the size value using \\( f(x) = \\ln (x + e^{-2}) \\), focusing on smaller values. The relationship between the transformed size and intrinsic rewards is visualized in Figure 4. The corresponding Pearson correlation coefficients from left to right are 0.81, 0.66, and 0.62, indicating that CLIP4MC reward has a higher correlation with the size, especially when it is relatively small. This is crucial in RL, as the agent needs dense and distinguishing reward signals to guide the learning process, particularly when the target is distant. Such characteristic is the essential benefit of CLIP4MC in RL." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.766, + 0.507, + 0.781 + ], + "angle": 0, + "content": "6.4 Ablation on Dataset Filtering" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.785, + 0.842 + ], + "angle": 0, + "content": "To evaluate the influence of the dataset quality on RL training and verify the effectiveness of our proposed correlation filter at the local level, we compare MineCLIP [ours] with two ablations: (1) MineCLIP [w/o LCF] is trained" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.369, + 0.129 + ], + "angle": 0, + "content": "H. Jiang et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.314 + ], + "angle": 0, + "content": "on the dataset processed with only global correlation filtering, omitting the local correlation filtering. (2) MineCLIP [RS] is trained on 640K video clips randomly sampled from the MineDojo released database, consistent with their stated dataset construction method [8]. Note that methods evaluated here do not apply random swap introduced in Section 5.2 since these datasets cannot provide entity size. The results are presented in Table 2. Given the overall performance across all three presented tasks, firstly, [RS] does not perfectly reproduce the performance of [official], suggesting that misalignment in the original database hinders the reproduction of [official]. Secondly, part of our proposed data filtering makes the model, [w/o LCF], comparable to [official]. In addition, with all of our proposed data filtering techniques, [ours] outperforms [official]." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.335, + 0.503, + 0.35 + ], + "angle": 0, + "content": "6.5 Video-Text Retrieval Results" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.361, + 0.461, + 0.542 + ], + "angle": 0, + "content": "Table 3 shows the results of video-to-text retrieval and text-to-video on the test set with the same MineCLIP model for a fair comparison. We train these models for 20 epochs and select the models with the highest R@1 value on the test set, respectively. From the results, the model trained on the randomly sampled dataset gets the worst performance than those trained on our YouTube dataset." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.543, + 0.461, + 0.694 + ], + "angle": 0, + "content": "The results demonstrate that our neat dataset indeed can facilitate table observation is that [ours] a lower performance on retrieval to the video-text alignment objective [w/o LCF] is trained on the data which is directly conducted at the the dataset further filtered by local signing for RL training. Note that for better training the VLM model" + }, + { + "type": "table_caption", + "bbox": [ + 0.469, + 0.406, + 0.788, + 0.447 + ], + "angle": 0, + "content": "Table 3: Results of video-to-text / text-to-video retrieval on the test set. The best results are highlighted in bold." + }, + { + "type": "table", + "bbox": [ + 0.476, + 0.448, + 0.772, + 0.531 + ], + "angle": 0, + "content": "
ModelsR@1R@5R@10
ours12.4/13.127.5/27.835.3/35.9
w/o LCF12.7/14.129.4/29.737.9/37.5
RS11.1/11.624.8/25.032.4/32.2
" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.717, + 0.36, + 0.733 + ], + "angle": 0, + "content": "7 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.749, + 0.788, + 0.841 + ], + "angle": 0, + "content": "We construct a neat YouTube dataset based on the large-scale YouTube database provided by MineDojo. Moreover, we introduce a novel cross-modal contrastive learning framework architecture, CLIP4MC, to learn an RL-friendly VLM that serves as an intrinsic reward function for open-ended tasks. Our findings suggest that our dataset enhances the acquisition of fundamental game concepts and CLIP4MC delivers a more effective reward signal for RL training." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.66, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "CLIP4MC" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.146, + 0.403, + 0.163 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.179, + 0.785, + 0.209 + ], + "angle": 0, + "content": "This work was supported by NSFC under grant 62250068. The authors would like to thank the anonymous reviewers for their valuable comments." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.235, + 0.323, + 0.25 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.268, + 0.786, + 0.31 + ], + "angle": 0, + "content": "1. Bain, M., Nagrani, A., Varol, G., Zisserman, A.: Frozen in time: A joint video and image encoder for end-to-end retrieval. In: IEEE/CVF International Conference on Computer Vision (ICCV) (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.311, + 0.787, + 0.351 + ], + "angle": 0, + "content": "2. Baker, B., Akkaya, I., Zhokhov, P., Huizinga, J., Tang, J., Ecoffet, A., Houghton, B., Sampedro, R., Clune, J.: Video pretraining (VPT): learning to act by watching unlabeled online videos. arXiv preprint arXiv:2206.11795 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.353, + 0.786, + 0.393 + ], + "angle": 0, + "content": "3. Baumli, K., Baveja, S., Behbahani, F., Chan, H., Comanici, G., Flennerhag, S., Gazeau, M., Holsheimer, K., Horgan, D., Laskin, M., et al.: Vision-language models as a source of rewards. arXiv preprint arXiv:2312.09187 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.395, + 0.786, + 0.435 + ], + "angle": 0, + "content": "4. Cai, S., Wang, Z., Ma, X., Liu, A., Liang, Y.: Open-world multi-task control through goal-aware representation learning and adaptive horizon prediction. arXiv preprint arXiv:2301.10034 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.437, + 0.786, + 0.476 + ], + "angle": 0, + "content": "5. Devlin, J., Chang, M.W., Lee, K., Toutanova, K.: Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.479, + 0.786, + 0.533 + ], + "angle": 0, + "content": "6. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., Uszkoreit, J., Houlsby, N.: An image is worth 16x16 words: Transformers for image recognition at scale. In: International Conference on Learning Representations (ICLR) (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.535, + 0.786, + 0.575 + ], + "angle": 0, + "content": "7. Dzabraev, M., Kalashnikov, M., Komkov, S., Petiushko, A.: MDMMT: multidomain multimodal transformer for video retrieval. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.577, + 0.786, + 0.631 + ], + "angle": 0, + "content": "8. Fan, L., Wang, G., Jiang, Y., Mandlekar, A., Yang, Y., Zhu, H., Tang, A., Huang, D.A., Zhu, Y., Anandkumar, A.: Minedojo: Building open-ended embodied agents with internet-scale knowledge. In: Neural Information Processing Systems (NeurIPS) Datasets and Benchmarks Track (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.633, + 0.786, + 0.673 + ], + "angle": 0, + "content": "9. Guhr, O., Schumann, A.K., Bahrmann, F., Böhme, H.J.: Fullstop: Multilingual deep models for punctuation prediction (June 2021), http://ceur-ws.org/Vol-2957/sepp_paper4.pdf" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.674, + 0.786, + 0.728 + ], + "angle": 0, + "content": "10. Guss, W.H., Codel, C., Hofmann, K., Houghton, B., Kuno, N., Milani, S., Mohanty, S., Liebana, D.P., Salakhutdinov, R., Topin, N., et al.: Neurips 2019 competition: the minerl competition on sample efficient reinforcement learning using human priors. arXiv preprint arXiv:1904.10079 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.729, + 0.786, + 0.77 + ], + "angle": 0, + "content": "11. Guss, W.H., Houghton, B., Topin, N., Wang, P., Codel, C., Veloso, M., Salakhutdinov, R.: Minerl: A large-scale dataset of apache demonstrations. In: International Joint Conference on Artificial Intelligence (IJCAI) (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.772, + 0.786, + 0.798 + ], + "angle": 0, + "content": "12. Hafner, D., Pasukonis, J., Ba, J., Lillicrap, T.P.: Mastering diverse domains through world models. arXiv preprint arXiv:2301.04104 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.8, + 0.786, + 0.84 + ], + "angle": 0, + "content": "13. Haiminen, N., Gionis, A., Laasonen, K.: Algorithms for unimodal segmentation with applications to unimodality detection. Knowledge and information systems 14, 39-57 (2008)" + }, + { + "type": "list", + "bbox": [ + 0.22, + 0.268, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.369, + 0.129 + ], + "angle": 0, + "content": "H. Jiang et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.148, + 0.785, + 0.189 + ], + "angle": 0, + "content": "14. Johnson, M., Hofmann, K., Hutton, T., Bignell, D.: The malmo platform for artificial intelligence experimentation. In: International Joint Conference on Artificial Intelligence (IJCAI) (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.19, + 0.785, + 0.218 + ], + "angle": 0, + "content": "15. Kaelbling, L.P., Littman, M.L., Cassandra, A.R.: Planning and acting in partially observable stochastic domains. Artificial intelligence 101(1-2), 99-134 (1998)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.218, + 0.785, + 0.286 + ], + "angle": 0, + "content": "16. Kanervisto, A., Milani, S., Ramanauskas, K., Topin, N., Lin, Z., Li, J., Shi, J., Ye, D., Fu, Q., Yang, W., Hong, W., Huang, Z., Chen, H., Zeng, G., Lin, Y., Micheli, V., Alonso, E., Fleuret, F., Nikulin, A., Belousov, Y., Svidchenko, O., Shpilman, A.: Minerl diamond 2021 competition: Overview, results, and lessons learned. arXiv preprint arXiv:2202.10583 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.287, + 0.785, + 0.314 + ], + "angle": 0, + "content": "17. Li, Y., Wang, H., Duan, Y., Li, X.: Clip surgery for better explainability with enhancement in open-vocabulary tasks. arXiv preprint arXiv:2304.05653 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.315, + 0.785, + 0.355 + ], + "angle": 0, + "content": "18. Lin, Z., Li, J., Shi, J., Ye, D., Fu, Q., Yang, W.: Juewu-mc: Playing minecraft with sample-efficient hierarchical reinforcement learning. arXiv preprint arXiv:2112.04907 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.356, + 0.785, + 0.398 + ], + "angle": 0, + "content": "19. Liu, S., Fan, H., Qian, S., Chen, Y., Ding, W., Wang, Z.: Hit: Hierarchical transformer with momentum contrast for video-text retrieval. In: IEEE/CVF International Conference on Computer Vision (ICCV) (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.398, + 0.785, + 0.439 + ], + "angle": 0, + "content": "20. Luo, H., Ji, L., Zhong, M., Chen, Y., Lei, W., Duan, N., Li, T.: Clip4clip: An empirical study of CLIP for end to end video clip retrieval and captioning. Neurocomputing 508, 293-304 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.439, + 0.785, + 0.481 + ], + "angle": 0, + "content": "21. Ma, Y.J., Liang, W., Wang, G., Huang, D.A., Bastani, O., Jayaraman, D., Zhu, Y., Fan, L., Anandkumar, A.: Eureka: Human-level reward design via coding large language models. arXiv preprint arXiv:2310.12931 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.481, + 0.785, + 0.522 + ], + "angle": 0, + "content": "22. Nottingham, K., Ammanabrolu, P., Suhr, A., Choi, Y., Hajishirzi, H., Singh, S., Fox, R.: Do embodied agents dream of pixelated sheep?: Embodied decision making using language guided world modelling. arXiv preprint arXiv:2301.12050 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.522, + 0.785, + 0.55 + ], + "angle": 0, + "content": "23. Oord, A.v.d., Li, Y., Vinyals, O.: Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.55, + 0.503, + 0.564 + ], + "angle": 0, + "content": "24. OpenAI: Gpt-4 technical report (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.564, + 0.785, + 0.618 + ], + "angle": 0, + "content": "25. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., Krueger, G., Sutskever, I.: Learning transferable visual models from natural language supervision. In: International Conference on Machine Learning, ICML (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.619, + 0.785, + 0.66 + ], + "angle": 0, + "content": "26. Radford, A., Kim, J.W., Xu, T., Brockman, G., McLeavey, C., Sutskever, I.: Robust speech recognition via large-scale weak supervision (2022). https://doi.org/10.48550/ARXIV.2212.04356, https://arxiv.org/abs/2212.04356" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.66, + 0.785, + 0.688 + ], + "angle": 0, + "content": "27. Schulman, J., Wolski, F., Dhariwal, P., Radford, A., Klimov, O.: Proximal policy optimization algorithms. arXiv preprint arXiv:1707.06347 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.688, + 0.785, + 0.757 + ], + "angle": 0, + "content": "28. Shah, R., Wang, S.H., Wild, C., Milani, S., Kanervisto, A., Goecks, V.G., Waytowich, N.R., Watkins-Valls, D., Prakash, B., Mills, E., Garg, D., Fries, A., Souly, A., Chan, J.S., del Castillo, D., Lieberum, T.: Retrospective on the 2021 minerl BASALT competition on learning from human feedback. In: Neural Information Processing Systems (NeurIPS) Competitions and Demonstrations Track (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.758, + 0.785, + 0.799 + ], + "angle": 0, + "content": "29. Shu, T., Xiong, C., Socher, R.: Hierarchical and interpretable skill acquisition in multi-task reinforcement learning. In: International Conference on Learning Representations (ICLR) (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.799, + 0.785, + 0.84 + ], + "angle": 0, + "content": "30. Tessler, C., Givony, S., Zahavy, T., Mankowitz, D., Mannor, S.: A deep hierarchical approach to lifelong learning in apacheft. In: AAAI conference on artificial intelligence (AAAI) (2017)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.148, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.66, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "CLIP4MC" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.189 + ], + "angle": 0, + "content": "31. Wang, G., Xie, Y., Jiang, Y., Mandlekar, A., Xiao, C., Zhu, Y., Fan, L., Anandkumar, A.: Voyager: An open-ended embodied agent with large language models. arXiv preprint arXiv:2305.16291 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.19, + 0.788, + 0.232 + ], + "angle": 0, + "content": "32. Xu, H., Ghosh, G., Huang, P.Y., Okhonko, D., Aghajanyan, A., Metze, F., Zettle-moyer, L., Feichtenhofer, C.: Videoclip: Contrastive pre-training for zero-shot video-text understanding. arXiv preprint arXiv:2109.14084 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.232, + 0.788, + 0.272 + ], + "angle": 0, + "content": "33. Yuan, H., Zhang, C., Wang, H., Xie, F., Cai, P., Dong, H., Lu, Z.: Plan4mc: Skill reinforcement learning and planning for open-world minecraft tasks. arXiv preprint arXiv:2303.16563 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.273, + 0.788, + 0.3 + ], + "angle": 0, + "content": "34. Zhou, C., Loy, C.C., Dai, B.: Extract free dense labels from clip. In: European Conference on Computer Vision. pp. 696-712. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.3, + 0.788, + 0.357 + ], + "angle": 0, + "content": "35. Zhu, X., Chen, Y., Tian, H., Tao, C., Su, W., Yang, C., Huang, G., Li, B., Lu, L., Wang, X., et al.: Ghost in the minecraft: Generally capable agents for open-world enviroments via large language models with text-based knowledge and memory. arXiv preprint arXiv:2305.17144 (2023)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.357 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/Reinforcement Learning Friendly Vision-Language Model for Minecraft/c4770d1a-703c-494e-ba79-18744626ca70_origin.pdf b/2024/Reinforcement Learning Friendly Vision-Language Model for Minecraft/c4770d1a-703c-494e-ba79-18744626ca70_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..9a9ac24ada1a3d90d4c83300c1701cc9ae7f04bb --- /dev/null +++ b/2024/Reinforcement Learning Friendly Vision-Language Model for Minecraft/c4770d1a-703c-494e-ba79-18744626ca70_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:873eca01ffe631751f25cd3004b86de81e4976317741de5867e7abd90df8082f +size 6368263 diff --git a/2024/Reinforcement Learning Friendly Vision-Language Model for Minecraft/full.md b/2024/Reinforcement Learning Friendly Vision-Language Model for Minecraft/full.md new file mode 100644 index 0000000000000000000000000000000000000000..2e3fb896622abb5c8f86dc112ee7024e7144daa2 --- /dev/null +++ b/2024/Reinforcement Learning Friendly Vision-Language Model for Minecraft/full.md @@ -0,0 +1,285 @@ +# Reinforcement Learning Friendly Vision-Language Model for Minecraft + +Haobin Jiang $^{1\star}$ , Junpeng Yue $^{1\star}$ , Hao Luo $^{1}$ , Ziluo Ding $^{2}$ , and Zongqing Lu $^{1,2\dagger}$ + +$^{1}$ School of Computer Science, Peking University $^{2}$ Beijing Academy of Artificial Intelligence {haobin.jiang,zongqing.lu}@pku.edu.cn + +Abstract. One of the essential missions in the AI research community is to build an autonomous embodied agent that can achieve high-level performance across a wide spectrum of tasks. However, acquiring or manually designing rewards for all open-ended tasks is unrealistic. In this paper, we propose a novel cross-modal contrastive learning framework architecture, CLIP4MC, aiming to learn a reinforcement learning (RL) friendly vision-language model (VLM) that serves as an intrinsic reward function for open-ended tasks. Simply utilizing the similarity between the video snippet and the language prompt is not RL-friendly since standard VLMs may only capture the similarity at a coarse level. To achieve RL-friendliness, we incorporate the task completion degree into the VLM training objective, as this information can assist agents in distinguishing the importance between different states. Moreover, we provide neat YouTube datasets based on the large-scale YouTube database provided by MineDojo. Specifically, two rounds of filtering operations guarantee that the dataset covers enough essential information and that the video-text pair is highly correlated. Empirically, we demonstrate that the proposed method achieves better performance on RL tasks compared with baselines. The code and datasets are available at https://github.com/PKU-RL/CLIP4MC. + +Keywords: Dataset $\cdot$ Multimodal model $\cdot$ Reinforcement learning + +# 1 Introduction + +Training reinforcement learning (RL) agents to perform complex tasks in a vision-based and open-ended world can be difficult. One main challenge is that manually specifying reward functions in all open-ended tasks is unrealistic [3,21], especially when we cannot access the internal configuration of the environment. In addition, learning a reward model from human feedback is typically expensive. To this end, MineDojo [8] has proposed an internet-scale, multi-modal knowledge base of YouTube videos to facilitate learning in an open-ended world. With the advent of a such large-scale database, agents are able to harvest practical + +knowledge encoded in large amounts of media like human beings. Moreover, a vision-language model (VLM), MineCLIP [8], is proposed to utilize the internet-scale domain knowledge. In more detail, the learned correlation score between the visual observation and the language prompt can be used effectively as an open-vocabulary, massively multi-task reward function for RL training. Therefore, no further task-specific reward design is needed for open-ended tasks. + +However, the autonomous embodied agent requires the vision-language model to provide a more instructive correlation score. Given the partial observations, e.g., video snippet, and the language prompt that describes the task, the agent needs to figure out two non-trivial matters to better evaluate the current state. On the one hand, whether the target entities are present within its field of vision? MineCLIP tried to address this question. However, the alignment of texts and videos in the YouTube database is totally a catastrophe, which impedes the learning of VLM. On the other hand, what is the relationship between each video snippet and the degree of completion of the task? Normally, in an open-ended world, e.g. Minecraft, the agent explores first, then approaches and interacts with the target object. In other words, the agent requires approaching the target object before it takes trial-and-error, even for the target that needs to be kept away. Therefore, it is reasonable to make an assumption, namely, that the higher the level of completion of the task, the closer the agent is to the targets in the video snippet. We also argue it is important to incorporate the level of completion of the task into the reward function. + +In this paper, we first construct neat YouTube datasets to facilitate the learning of basic game concepts, mainly the correspondence between the videos and texts. Though a large-scale database is provided by MineDojo, it contains significant noise due to its nature as an online resource. In addition, MineDojo only claims the training dataset is randomly sampled from the database, making it hard to reproduce. To overcome the catastrophic misalignment of video-text pairs in the original database, we have done four steps of dataset processing to guarantee the dataset is clean. Firstly, transcript cleaning enhances the accuracy of transcripts and ensures they are complete sentences. Secondly, keyword filtering ensures that the content of clips is relevant to the key entities in Minecraft, thereby facilitating the learning of basic game concepts. Thirdly, video partitioning and selection can handle the issue of scene transitions and thereby mitigate interference from other extraneous information. Lastly, correlation filtering can effectively address the issue of mismatch between video clips and transcripts. + +We also propose an upgraded vision-language model, CLIP4MC, to provide a more RL-friendly reward function. In RL, simply utilizing the similarity between the video snippet and the language prompt is not RL-friendly since MineCLIP tends to only capture the similarity at the entity level. In other words, VLM can hardly reflect the relationship between each video snippet and the degree of completion of the task. However, this information can better help agents distinguish the importance between similar states. To achieve this, we incorporate the degree of task completion into the VLM training objective. In more detail, CLIP has exhibited remarkable segmentation capabilities without fine-tuning. After we + +extend pre-trained MineCLIP with modifications inspired by MaskCLIP [34], it can segment the specified object from the image and label the size of the target shown in the corresponding video. Intuitively, the closer the agent is to the target, the larger the target size becomes. During the learning procedure, we dynamically control the degree of contrasting positive and negative pairs of instances based on the target size in this positive video sample. Thus, CLIP4MC can render a more RL-friendly reward signal that instructs the agent to learn tasks faster. Our proposed method is trained on our YouTube dataset and evaluated on MineDojo Programmatic tasks, including harvest, hunt, and combat tasks. Empirically, our results show that CLIP4MC can provide a more friendly reward signal for the RL training procedure. + +To summarize, our contributions are as follows: + +- Open-sourced datasets: We provide two high-quality datasets. The first one undergoes data cleaning (Sections 4.1 to 4.3) and global-level correlation filtering (Section 4.4). The VLM trained on this dataset matches the performance of the officially released MineCLIP, which lacks a publicly available training set. +- RL-friendlyly dataset: Our second dataset further incorporates local-level correlation filtering (Section 4.4), making it more suited for RL. The VLM trained on this dataset outperforms that on the first dataset. +- RL-friendly VLM: To better evaluate and leverage the advantages of our more RL-friendly dataset, we introduce CLIP4MC, a novel method to train a VLM that could improve downstream RL performance (Section 5.2). + +# 2 Related Work + +Video-Text Retrieval. Video-text retrieval plays an essential role in multimodal research and has been widely used in many real-world web applications. Recently, the pre-trained models have dominated this line of research with noticeable results on both zero-shot and fine-tuned retrieval. Especially, BERT [5], ViT [6], and CLIP [25], are used as the backbones to extract the text or video embedding. The cross-modal embeddings are then matched with specific fusion networks to find the correct video-text pair. + +In more detail, CLIP4Clip [20] proposes three different similarity modules to calculate the correlation between video and text embeddings. HiT [19] performs hierarchical matching at two different levels, i.e. semantic level and feature level. Note that semantic level and feature level features are from the transformer network's higher and lower feature layers, respectively. Frozen [1] proposes a dual encoder architecture that utilizes the flexibility of a transformer visual encoder to train from images or video clips with text captions. Moreover, MDMMT [7] adopts several pre-training models as encoders and it shows the CLIP-based model performs the best. Therefore, our model follows this line of research by using the pre-trained model, CLIP [25], to extract the feature embeddings. + +Minecraft for AI Research. As an open-ended video game with an egocentric vision, Minecraft is a noticeable and important domain in RL due to the nature + +of the sparse reward, large exploration space, and long-term episodes. Since the release of the Malmo simulator [14] and later the MineDojo simulator [8], various methods have attempted to train agents to complete tasks in Minecraft [10, 12, 18, 29, 30]. Approaches such as model-based RL, hierarchical RL, goal-based RL, and reward shaping have been adopted to alleviate the sparse reward and exploration difficulty for the agent in this environment. + +Recently, with the development of large language models (LLM) like GPT-4 [24], a series of methods leveraging LLMs for high-level planning in Minecraft have been proposed [16, 22, 31, 33, 35]. These methods have demonstrated remarkable capabilities in guiding the agent to complete multiple complicated, long-horizon tasks, such as mining diamonds. These LLMs play a crucial role in decision-making, determining the sequence of basic skills required to accomplish specific tasks. Their effectiveness is due to their extensive knowledge about Minecraft, learned from the Internet, and their ability to reflect on real-time feedback from the game environment. + +In addition to the use of LLMs, recent research attempts to incorporate Internet visual data into basic skill learning in Minecraft, beyond the traditional RL methods. MineRL [11] collected 60M player demonstrations with action labels, motivating some methods [16,28] based on behavior cloning. As well-labeled data is limited in quantity, MineCLIP [8] instead uses over 730K narrated Minecraft videos without action labels from YouTube. It aims to learn a vision-language model providing auxiliary reward signals, utilizing the vast and diverse data available on the Internet. Different from MineCLIP, VPT [2] uses action-labeled data to train an inverse dynamic model to label 70K hours of Internet videos and then conduct behavior cloning. + +Unlike the existing approaches, which incorporate the human experience and require a large number of demonstrations with action labels to train the agent, our work follows the line of MineCLIP [8] and focuses on only using the data without action labels to assist agent learning in Minecraft, which is more friendly with data collection and has the potential to scale in the future. + +# 3 Background + +MineDojo tasks. MineDojo [8] provides thousands of benchmark tasks, which can be used to develop generally capable agents in Minecraft. These tasks can be divided into two categories, Programmatic and Creative tasks. The former has ground-truth simulator states to assess whether the task has been completed. The latter, however, do not have well-defined success criteria and tend to be more open-ended, but have to be evaluated by humans. + +We mainly focus on Programmatic tasks since they can be automatically assessed. Specifically, MineDojo provides 4 categories of programmatic tasks, including Harvest, Combat, Survival, and Tech Tree, with 1581 template-generated natural language goals to evaluate the agent's different capabilities. Among these tasks, Survival and Tech Tree tasks are harder than Harvest and Combat tasks. Currently, MineCLIP [8] only expresses promising potential in some Harvest and + +![](images/be300fae5c1064a448f8b924e71b7a754723634d5ac4539840950a0e1c1966db.jpg) + +# irrelevant video content + +... so here I'm with nick and we've got look at my inventory and his inventory looks similar this is gonna be interesting so four... + +![](images/b32ef89bcb2fe46bf77fcec0507609a18eaf98a96463e7ac3995570d3b8b0571.jpg) + +# mismatched video content + +... give it I'm a little chicken just laying around by any chance a lot of arrows going on down here I'm guessing most of ... + +![](images/da7af01c9c11db865bacd59f0bd001029674962d8210b69d0c492f200e7712dc.jpg) +Fig. 1: Illustration of the YouTube video database. The screenshots of video clips are on the left and key entities are circled in red. The corresponding transcript clips are on the right and key entities are marked in red. We give examples of irrelevant, mismatched, and matched video content in the YouTube video database. + +# matched video content + +... at some point you will lose there too or also hearts there spider was like being completely blown away! I don't assume anything at ... + +Combat tasks. Harvest means finding, obtaining, cultivating, or manufacturing hundreds of materials and objects. Combat means fighting various monsters and creatures that require fast reflexes and martial skills. + +POMDP. We model the programmatic task as a partially observable Markov decision process (POMDP) [15]. At each timestep $t$ , the agent obtains the partial observation $o_t$ from the global state $s_t$ and a language prompt $G$ , takes action $a_t$ following its policy $\pi(a_t|o_t)$ , and receives a reward $r_t = \Phi(V_t, G)$ , where $V_t$ is the fixed-length sequence of observations till $t$ (thus a video snippet) and $\Phi$ maps $V_t$ and $G$ to a scalar value. Then the environment transitions to the next state $s_{t+1}$ given the current state and action according to transition probability function $\mathcal{T}(s_{t+1}|s_t, a_t)$ . The agent aims to maximize the expected return $R = \mathbb{E}_{\pi} \sum_{t=1}^{T} \gamma^{t-1} r_t$ , where $\gamma$ is the discount factor and $T$ is the episode time horizon. + +# 4 YouTube Dataset + +The Internet is rich in Minecraft-related data, containing a wealth of weakly labeled or even unlabeled Minecraft knowledge, including crucial entities, plausible actions, and common-sense event processes. With these multi-modal data, it is possible to create dense language-conditioned rewards, making open-ended long-term task learning feasible. MineDojo [8] collected over 730K YouTube videos and their corresponding transcripts, totaling 33 years of video content and 2.2B words in transcripts. Around 640K video clips are selected using keyword-based filtering from these videos, and these clips are used to train MineCLIP. However, MineDojo open-sourced a 13.8M dataset, and the 640K video clips were randomly sampled in their paper. However, as illustrated in Figure 1, most videos feature irrelevant game content that is not conducive to learning basic game + +concepts. Meanwhile, the alignment between the transcripts and videos may not always be precise, leading to temporal or content discrepancies that could hinder the learning of retrieval and RL tasks. Given the low quality of the data, we found it necessary to provide a neat 640k dataset to the community for the training usage. To address these issues, we adopt the following few steps to acquire a clean dataset with high quality. + +# 4.1 Transcript Cleaning + +Downloading YouTube's automatic transcripts directly can lead to several issues. Firstly, transcript blocks may overlap, resulting in overlapping timestamps in the transcript. Additionally, the caption quality is typically mediocre, with a higher occurrence of misidentifications, especially in non-English videos. Moreover, the transcript lacks punctuation, making it less friendly for understanding semantics. Based on the aforementioned issues, we implement a pipeline to construct high-quality transcripts as follows: (1) Extract audio from the videos and use Whisper [26] to obtain high-quality, temporally non-overlapping transcripts. (2) Employ FullStop [9] to generate punctuation, resulting in complete sentences of 10-35 words in length. + +# 4.2 Keyword Filtering + +Following MineDojo [8], we also implement keyword-based filtering to ensure that the content in our dataset is pertinent to the key entities in Minecraft, thereby facilitating the learning of basic game concepts. As essential components of Minecraft, the key entities, such as stones, trees, and sheep, are common across multiple tasks in MineDojo and videos from YouTube. Specifically, we identify entity keywords in the transcripts using a keyword list from MineDojo and extract transcript clips formed into sentences to encompass as many keywords as possible. These extracted transcript clips then serve as the textual component of our dataset, determining the location of corresponding video clips. + +# 4.3 Video Partition and Selection + +After completing the previous two steps, we obtain transcript clips relevant to the keywords. For each transcript clip, we calculate the central timestamp that corresponds to the clip based on the transcript timestamps and then use this central timestamp to extract a video clip with a duration of $D$ seconds from the video. This process allows us to obtain temporally-aligned video clips. However, the video clips obtained in this manner still exhibit some issues, including scene transitions and discrepancies between video content and transcripts. We handle the former problem through video partition and filtering, while the latter problem is addressed in Section 4.4. + +Owing to the informal nature of YouTube content, there is often a lack of semantic congruence between the video clips and their corresponding transcriptions, as noted in VideoCLIP [32]. Moreover, video clips frequently contain a + +![](images/a8d0bf9dbe81d2743b2257b8a10eebd1e3d489606188ad5647f2d136edf00ce6.jpg) +Fig. 2: Examples of how we estimate the size of key entities in video frames. Red bounding boxes are generated by our modified MineCLIP visual encoder, following the approach proposed in MaskCLIP [34]. These boxes are then used to calculate the size. + +few different behaviors which cause scene transitions, e.g. Chopping down the tree first, then suddenly switching the inventory bar. Since some scene transitions lead to irrelevant information, we partition the video content into several semantically coherent segments based on the semantic structure of the video. Then we select the segment that aligns best with the transcript. + +To achieve semantic partition of video content, we employ the Bellman $k$ -segmentation algorithm [13]. This algorithm divides a sequence of data points into $k$ distinct and constant-line segments, providing a piece-wise constant approximation of the data. To process the video, we first use the officially released MineCLIP [8] video encoder to obtain the embedding of each frame, since MineCLIP can capture video semantics to some extent. Subsequently, we partition these embeddings into $k$ segments and select the segment with the highest similarity score, as calculated in MineCLIP. + +# 4.4 Correlation Filtering + +We employ correlation filtering techniques to address disparities that persist between video content and transcripts. The correlation filtering is done at two levels, the global and the local levels. From the global level, we calculate the cosine similarities between video embeddings and text embeddings via the original MineCLIP and then select clips based on these similarities. + +Recent research [17, 34] has demonstrated that CLIP [25], though trained on whole images, can generate meaningful local features. Inspired by this and following MaskCLIP [34], we make modifications to the original MineCLIP visual encoder, empowering it with the ability to estimate and label the size of the key entity, which is mentioned in the corresponding transcript, in each frame of a video clip without fine-tuning. Then the correlations between video clips and transcripts are calculated as the summation of the sizes across all frames in each video clip. We consider this correlation at the local level. + +Based on these two criteria, we select the top $k\%$ of clips, resulting in the final training set. We provide an elaboration on our implementation in Appendix A and illustrate some examples in Figure 2. + +The aforementioned four-step approach creates a dataset consisting of 640K video-text clip pairs, with an additional 4K pairs extracted for validation of video-text retrieval. Regarding the constants of the approach, $D$ and $k$ are set to 16 and 50, respectively. Therefore, our dataset comprises videos with a total duration of one week and approximately 0.16B words, significantly smaller in scale compared to the original low-quality 13.8M dataset. Importantly, we will open-source our entire YouTube dataset, serving as an upgraded version of the unreleased MineCLIP training data. + +# 5 CLIP4MC + +Given a video snippet $V$ and a language prompt $G$ , the vision-language model outputs a correlation score, $C$ , that measures the similarity between the video snippet and the language prompt. Ideally, if the agent performs behaviors following the description of the language prompt, the vision-language model will generate a higher correlation score, leading to a higher reward. Otherwise, the agent will be given a lower reward. + +# 5.1 Architecture + +We follow the same architecture design as MineCLIP [8], including a video encoder, a text encoder, and a similarity calculator. All the video frames first go through the spatial transformer to obtain a sequence of frame features. The temporal transformer is then utilized to summarize the sequence of frame features into a single video embedding. An adapter further processes the video embedding for better features. Refer to Appendix B for more details about the architecture. + +Empirically, we found that the video encoder (essentially MineCLIP) can provide a bond between the entities and the language prompts and give a similar high reward as long as the target entities are present in the video frames, similar to observations in [4]. However, such a reward is not instructive enough for RL tasks since it does not reflect the behavioral trends of agents. + +# 5.2 Contrastive Training + +We aim to minimize the sum of the multi-modal contrastive losses [23], including video-to-text, and text-to-video: + +$$ +\mathcal {L} = - \sum_ {\left(z _ {v}, z _ {m}, z _ {t}\right) \in \mathcal {B}} \left(\log \operatorname {N C E} \left(z _ {v}, z _ {t}\right) + \log \operatorname {N C E} \left(z _ {t}, z _ {v}\right)\right), \tag {1} +$$ + +where $\mathcal{B}$ is the batch containing sampled video-text pairs, and $\mathrm{NCE}(\cdot ,\cdot)$ is the contrastive loss that calculates the similarity of two inputs. To illustrate, the + +![](images/1e7069af898432d6a42b90f799985badac834b84318181f38c21b8036c82daf4.jpg) +Fig. 3: Illustration of CLIP4MC training. The upper part shows the concept of contrastive learning, while the lower part explains the swapping operation. + +video-to-text contrastive loss is given by + +$$ +\mathrm {N C E} \left(z _ {v}, z _ {t}\right) = \frac {\exp \left(z _ {v} \cdot z _ {t} ^ {+} / \lambda\right)}{\sum_ {z \in \left\{z _ {t} ^ {+} , z _ {t} ^ {-} \right\}} \exp \left(z _ {v} \cdot z / \lambda\right)}, \tag {2} +$$ + +where $\lambda$ is a temperature hyperparameter, $z_{t}^{+}$ is the positive text embedding matching with the video embedding $z_{v}$ , and $\{z_t^-\}$ are negative text embeddings that are implicitly formed by other text clips in the training batch. Contrastive loss of text-to-video is defined in the same way. + +In addition, we also need to incorporate the degree of task completion into the training objective. Specifically, we hope the similarity score between video clip and text embeddings could reflect the task completion degree, i.e., higher completion leads to higher similarity. The sizes of key entities provided by the local correlation filtering serve as a surrogate for the task completion degree. Initially, we try to dynamically adjust the weight of the positive pairs based on the size of the target object. However, it did not work. Essentially, positive and negative pairs will still be separated, even with smaller sample weights. + +Therefore, we instead forcibly change the labels of the positive and negative video samples, as illustrated in Figure 3, thereby reducing the model's confidence in certain positive pairs. Specifically, during training, a positive video sample is swapped with a random negative video sample from the $\mathcal{B}$ based on probability $P_{\mathrm{swap}}$ . This swap only occurs when the labeled target size $H_bW_b$ in the positive video sample is below a certain threshold $\tau$ , and the smaller the size, the greater + +the probability, + +$$ +P _ {\mathrm {s w a p}} = P _ {\mathrm {m a x}} * \mathrm {R e L U} \left(1 - \frac {H _ {b} W _ {b}}{\tau H W}\right), \tag {3} +$$ + +where $HW$ represents the size of the image. + +Since contrastive learning brings positive samples closer and pushes negative samples away, converting a positive pair to a negative one with a certain probability, i.e., the aforementioned random swap, during training will lower the similarity score. As the decrease in similarity is directly proportional to the swapping probability, we set $P_{\mathrm{swap}}$ to be inversely related to task completion, ensuing similarity score increases as completion increases. + +In addition, the upper limit $P_{\mathrm{max}}$ in Equation (3) is set to 0.5, as the swap should not disrupt severely the normal distinction between positive and negative pairs in most cases. This constraint enables CLIP4MC to preserve the understanding capability of MineCLIP for general behaviors that may lack explicit target entities in Minecraft, as discussed in Appendix F. + +# 5.3 RL Training + +For RL training, the first step is reward generation. At timestep $t$ , we concatenate the agent's latest 16 egocentric RGB frames in a temporal window to form a video snippet, $V_{t}$ . CLIP4MC outputs the probability $P_{G,t}$ that calculates the similarity of $V_{t}$ to the task prompt, $G$ , against all other negative prompts. To compute the reward, we further process the raw probability as previous work [8] $r_t = \max (P_{G,t} - 1 / N_t,0)$ , where $N_{t}$ is the number of prompts passed to CLIP4MC. Note that CLIP4MC can handle unseen language prompts without any further fine-tuning due to the open-vocabulary ability of CLIP [25]. + +The ultimate goal is to train a policy network that takes as input raw pixels and other structural data and outputs discrete actions to accomplish the task that is described by the language prompt, $G$ . We use PPO [27] as our RL training backbone and the policy is trained on the CLIP4MC reward together with the sparse task reward if any. The policy input contains several modality-specific components and more details can be found in Appendix D. + +# 6 Experiments + +In this section, we comprehensively evaluate and analyze our proposed model CLIP4MC, utilizing the open-ended platform MineDojo [8], which comprises thousands of diverse, open-ended Minecraft tasks designed for embodied agents. + +We compare CLIP4MC against two baselines: (1) MineCLIP [official], the officially released MineCLIP model [8]. (2) MineCLIP [ours], using the same architecture as MineCLIP [official] but trained on our cleaned YouTube dataset. It also serves as the ablation of CLIP4MC without the swap operation. We train CLIP4MC and MineCLIP [ours] for 20 epochs and select the models with the highest performance on RL tasks. Please refer to Appendix C for more training details. All results are presented in terms of the mean and standard error of four runs with different random seeds. + +# 6.1 Environment Settings + +We conduct experiments on eight Programmatic tasks, comprising two harvest tasks: milk a cow and shear wool, two combat tasks: combat a spider and combat a zombie, and four hunt tasks: hunt a cow, hunt a sheep, hunt a pig, and hunt a chicken. These tasks are all built-in tasks in the MineDojo benchmark. + +Harvest. Milk a cow requires the agent to obtain milk from a cow with an empty bucket. Similarly, shear wool requires the agent to obtain wool from a sheep with shears. A harvest task is terminated and considered completed when the target item is obtained by the agent with a specified quantity. The prompts used to calculate the reward is "obtain milk from a cow in plains with an empty bucket"; for shear wool, it is "shear a sheep in plains with shears". + +Combat. In these tasks, target animals, spiders, and zombies, are hostile and will actively approach and attack the agent. The agent's goal is to fight and kill the target animal. The prompt for each combat task is "combat a spider/zombie in plains with a diamond sword". + +Hunt. Hunt tasks consist of hunt a cow, hunt a sheep, hunt a pig, and hunt a sheep. For each task, the agent's goal is to kill the target animal as indicated in the task name. Different from combat tasks, the target animals will flee from the agent after being attacked. Therefore, these tasks require the agent to keep chasing and attacking the target, making them challenging. As noted in [4], the original MineCLIP reward fails in these tasks since it cannot consistently increase when the agent approaches the agent. This observation aligns with our assertion that the original MineCLIP model can hardly capture the degree of task completion. The prompt for each task is "hunt a {target} in plains with a diamond sword" where {target} is replaced with the corresponding animal name. + +More elaborated introduction of the Minecraft environment and settings of these tasks are available in Appendix D. To guarantee a fair comparison, we adopt the same RL hyperparameters for all models and tasks. These hyperparameters are listed in Appendix E. + +# 6.2 RL Results + +Through our evaluation of CLIP4MC, MineCLIP [ours], and MineCLIP [official] across eight Minecraft tasks, we want to answer two key questions: + +(1) Whether the YouTube dataset we constructed enables MineCLIP, when trained on it, to provide a more effective reward signal for task learning? +(2) Whether our upgraded model, CLIP4MC, based on our YouTube dataset, offers a reward signal that is further friendly for the RL training procedure? + +These questions are central to verifying the effectiveness of our dataset and CLIP4MC model in Minecraft. Table 1 shows the success rates of all methods on the eight Minecraft tasks. + +It is noticeable that, in four hunt tasks, three models demonstrate varying performance on RL. Firstly, MineCLIP [ours] consistently achieves better results compared to MineCLIP [official] across all hunt tasks. The superior performance of MineCLIP [ours] provides a positive answer to our first question, suggesting + +Table 1: Success rates $(\%)$ of RL trained with rewards provided by different models on eight Minecraft tasks. Each mean and standard error of success rates are calculated on four models after training 1e6 environment steps with different random seeds. + +
ModelsHarvestCombat
milk a cowshear woolcombat a spidercombat a zombie
CLIP4MC84.5±2.074.6±2.185.8±0.970.4±8.3
MineCLIP[ours]84.4±1.171.6±3.575.4±10.163.6±8.7
MineCLIP[official]84.1±0.573.2±1.882.7±2.557.4±3.7
ModelsHunt
hunt a cowhunt a sheephunt a pighunt a chicken
CLIP4MC39.8±2.545.9±7.230.6±8.426.1±3.5
MineCLIP[ours]17.3±10.633.0±18.114.1±10.615.3±10.6
MineCLIP[official]11.6±11.128.5±16.71.5±0.60.0±0.0
+ +that the YouTube dataset we construct indeed enhances the effectiveness of the reward signal in MineCLIP for task learning. Note that this is the dataset we plan to release for better training the VLM model for RL tasks on Minecraft. Secondly, CLIP4MC shows significantly higher success rates on hunt tasks compared to both MineCLIP [ours] and MineCLIP [official], meaning that the answer to the second question is also positive. As CLIP4MC provides a reward signal taking into account a surrogate for the degree of task completion, i.e., the size of the target object in our implementation, it becomes more RL-friendly in these challenging tasks. + +In addition, we notice a practical example of misalignment in the officially released MineCLIP model. Specifically, we observe that in hunt a chicken, the agent trained with MineCLIP [official] tends to keep looking at the sky, indicating such behavior can provide a high intrinsic reward. This phenomenon suggests that the officially released MineCLIP indeed suffers from the misalignment problem in the YouTube dataset. In contrast, our MineCLIP [official] shows promising behaviors on this task, demonstrating that our dataset processing improves the alignment between the transcripts and videos. + +In contrast to hunt tasks, these methods perform comparably on harvest tasks and combat tasks, while CLIP4MC still shows marginal advantage over the other two methods. This finding aligns with the results reported in [8], where the MineCLIP reward already achieves saturated performance on these easy tasks. Therefore, further enhancements in the reward signal, like MineCLIP [ours] and CLIP4MC, do not significantly improve performance. Unlike hunt tasks, neither harvest nor combat tasks require the agent to take multiple rounds of chasing. This result does not detract from the superiority of CLIP4MC, evidenced by its performance on hunt tasks, which are considered more difficult [2]. + +![](images/fef7ef70be28bbb120a7d8b88b5956ec245bd7ca16eb0c4529f81fdf38f51cac.jpg) +(a) CLIP4MC + +![](images/d6f8936764fc0ef98eb713c07bd06f885e82f4f3a81bd090e352396984838625.jpg) +(b) MineCLIP [ours] +Fig. 4: Scatter plots illustrating the relationship between the entity size and the intrinsic reward. The red line indicates a linear fit to the data. + +![](images/7bc8aa2f72891617c881769faa584b7b2d959cf9b8226567d813845665890831.jpg) +(c) MineCLIP [official] + +Table 2: Success rates (%) of RL trained with rewards provided by different models on three Minecraft tasks. Each mean and standard error of success rates are calculated on four models after training 5e5 and 1e6 environment steps with different random seeds. + +
Modelsmilk a cowshear woolhunt a cow
5e51e65e51e65e51e6
ours73.1±0.984.4±1.147.5±6.371.6±3.515.6±8.717.3±10.6
w/o LCF71.3±1.880.7±1.153.8±3.373.6±1.50.7±0.712.1±12.0
RS65.1±3.981.7±1.021.7±6.568.5±3.61.0±0.54.0±2.6
official69.8±1.384.1±0.547.0±13.573.2±1.82.7±1.111.6±11.1
+ +# 6.3 Reward Analysis + +To quantitatively verify that CLIP4MC captures the size of the target entity specified in the language prompt, we collect 5000 steps in task hunt a cow and apply the method described in Appendix A to estimate the maximal size of the cow in consecutive 16 frames. Then we use CLIP4MC, MineCLIP [ours], and MineCLIP [official] to calculate intrinsic rewards respectively. Before calculating the correlation between the size and intrinsic rewards, we transform the size value using $f(x) = \ln (x + e^{-2})$ , focusing on smaller values. The relationship between the transformed size and intrinsic rewards is visualized in Figure 4. The corresponding Pearson correlation coefficients from left to right are 0.81, 0.66, and 0.62, indicating that CLIP4MC reward has a higher correlation with the size, especially when it is relatively small. This is crucial in RL, as the agent needs dense and distinguishing reward signals to guide the learning process, particularly when the target is distant. Such characteristic is the essential benefit of CLIP4MC in RL. + +# 6.4 Ablation on Dataset Filtering + +To evaluate the influence of the dataset quality on RL training and verify the effectiveness of our proposed correlation filter at the local level, we compare MineCLIP [ours] with two ablations: (1) MineCLIP [w/o LCF] is trained + +on the dataset processed with only global correlation filtering, omitting the local correlation filtering. (2) MineCLIP [RS] is trained on 640K video clips randomly sampled from the MineDojo released database, consistent with their stated dataset construction method [8]. Note that methods evaluated here do not apply random swap introduced in Section 5.2 since these datasets cannot provide entity size. The results are presented in Table 2. Given the overall performance across all three presented tasks, firstly, [RS] does not perfectly reproduce the performance of [official], suggesting that misalignment in the original database hinders the reproduction of [official]. Secondly, part of our proposed data filtering makes the model, [w/o LCF], comparable to [official]. In addition, with all of our proposed data filtering techniques, [ours] outperforms [official]. + +# 6.5 Video-Text Retrieval Results + +Table 3 shows the results of video-to-text retrieval and text-to-video on the test set with the same MineCLIP model for a fair comparison. We train these models for 20 epochs and select the models with the highest R@1 value on the test set, respectively. From the results, the model trained on the randomly sampled dataset gets the worst performance than those trained on our YouTube dataset. + +The results demonstrate that our neat dataset indeed can facilitate table observation is that [ours] a lower performance on retrieval to the video-text alignment objective [w/o LCF] is trained on the data which is directly conducted at the the dataset further filtered by local signing for RL training. Note that for better training the VLM model + +Table 3: Results of video-to-text / text-to-video retrieval on the test set. The best results are highlighted in bold. + +
ModelsR@1R@5R@10
ours12.4/13.127.5/27.835.3/35.9
w/o LCF12.7/14.129.4/29.737.9/37.5
RS11.1/11.624.8/25.032.4/32.2
+ +# 7 Conclusion + +We construct a neat YouTube dataset based on the large-scale YouTube database provided by MineDojo. Moreover, we introduce a novel cross-modal contrastive learning framework architecture, CLIP4MC, to learn an RL-friendly VLM that serves as an intrinsic reward function for open-ended tasks. Our findings suggest that our dataset enhances the acquisition of fundamental game concepts and CLIP4MC delivers a more effective reward signal for RL training. + +# Acknowledgements + +This work was supported by NSFC under grant 62250068. The authors would like to thank the anonymous reviewers for their valuable comments. + +# References + +1. Bain, M., Nagrani, A., Varol, G., Zisserman, A.: Frozen in time: A joint video and image encoder for end-to-end retrieval. In: IEEE/CVF International Conference on Computer Vision (ICCV) (2021) +2. Baker, B., Akkaya, I., Zhokhov, P., Huizinga, J., Tang, J., Ecoffet, A., Houghton, B., Sampedro, R., Clune, J.: Video pretraining (VPT): learning to act by watching unlabeled online videos. arXiv preprint arXiv:2206.11795 (2022) +3. Baumli, K., Baveja, S., Behbahani, F., Chan, H., Comanici, G., Flennerhag, S., Gazeau, M., Holsheimer, K., Horgan, D., Laskin, M., et al.: Vision-language models as a source of rewards. arXiv preprint arXiv:2312.09187 (2023) +4. Cai, S., Wang, Z., Ma, X., Liu, A., Liang, Y.: Open-world multi-task control through goal-aware representation learning and adaptive horizon prediction. arXiv preprint arXiv:2301.10034 (2023) +5. Devlin, J., Chang, M.W., Lee, K., Toutanova, K.: Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805 (2018) +6. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., Uszkoreit, J., Houlsby, N.: An image is worth 16x16 words: Transformers for image recognition at scale. In: International Conference on Learning Representations (ICLR) (2021) +7. Dzabraev, M., Kalashnikov, M., Komkov, S., Petiushko, A.: MDMMT: multidomain multimodal transformer for video retrieval. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops (2021) +8. Fan, L., Wang, G., Jiang, Y., Mandlekar, A., Yang, Y., Zhu, H., Tang, A., Huang, D.A., Zhu, Y., Anandkumar, A.: Minedojo: Building open-ended embodied agents with internet-scale knowledge. In: Neural Information Processing Systems (NeurIPS) Datasets and Benchmarks Track (2022) +9. Guhr, O., Schumann, A.K., Bahrmann, F., Böhme, H.J.: Fullstop: Multilingual deep models for punctuation prediction (June 2021), http://ceur-ws.org/Vol-2957/sepp_paper4.pdf +10. Guss, W.H., Codel, C., Hofmann, K., Houghton, B., Kuno, N., Milani, S., Mohanty, S., Liebana, D.P., Salakhutdinov, R., Topin, N., et al.: Neurips 2019 competition: the minerl competition on sample efficient reinforcement learning using human priors. arXiv preprint arXiv:1904.10079 (2019) +11. Guss, W.H., Houghton, B., Topin, N., Wang, P., Codel, C., Veloso, M., Salakhutdinov, R.: Minerl: A large-scale dataset of apache demonstrations. In: International Joint Conference on Artificial Intelligence (IJCAI) (2019) +12. Hafner, D., Pasukonis, J., Ba, J., Lillicrap, T.P.: Mastering diverse domains through world models. arXiv preprint arXiv:2301.04104 (2023) +13. Haiminen, N., Gionis, A., Laasonen, K.: Algorithms for unimodal segmentation with applications to unimodality detection. Knowledge and information systems 14, 39-57 (2008) + +14. Johnson, M., Hofmann, K., Hutton, T., Bignell, D.: The malmo platform for artificial intelligence experimentation. In: International Joint Conference on Artificial Intelligence (IJCAI) (2016) +15. Kaelbling, L.P., Littman, M.L., Cassandra, A.R.: Planning and acting in partially observable stochastic domains. Artificial intelligence 101(1-2), 99-134 (1998) +16. Kanervisto, A., Milani, S., Ramanauskas, K., Topin, N., Lin, Z., Li, J., Shi, J., Ye, D., Fu, Q., Yang, W., Hong, W., Huang, Z., Chen, H., Zeng, G., Lin, Y., Micheli, V., Alonso, E., Fleuret, F., Nikulin, A., Belousov, Y., Svidchenko, O., Shpilman, A.: Minerl diamond 2021 competition: Overview, results, and lessons learned. arXiv preprint arXiv:2202.10583 (2022) +17. Li, Y., Wang, H., Duan, Y., Li, X.: Clip surgery for better explainability with enhancement in open-vocabulary tasks. arXiv preprint arXiv:2304.05653 (2023) +18. Lin, Z., Li, J., Shi, J., Ye, D., Fu, Q., Yang, W.: Juewu-mc: Playing minecraft with sample-efficient hierarchical reinforcement learning. arXiv preprint arXiv:2112.04907 (2021) +19. Liu, S., Fan, H., Qian, S., Chen, Y., Ding, W., Wang, Z.: Hit: Hierarchical transformer with momentum contrast for video-text retrieval. In: IEEE/CVF International Conference on Computer Vision (ICCV) (2021) +20. Luo, H., Ji, L., Zhong, M., Chen, Y., Lei, W., Duan, N., Li, T.: Clip4clip: An empirical study of CLIP for end to end video clip retrieval and captioning. Neurocomputing 508, 293-304 (2022) +21. Ma, Y.J., Liang, W., Wang, G., Huang, D.A., Bastani, O., Jayaraman, D., Zhu, Y., Fan, L., Anandkumar, A.: Eureka: Human-level reward design via coding large language models. arXiv preprint arXiv:2310.12931 (2023) +22. Nottingham, K., Ammanabrolu, P., Suhr, A., Choi, Y., Hajishirzi, H., Singh, S., Fox, R.: Do embodied agents dream of pixelated sheep?: Embodied decision making using language guided world modelling. arXiv preprint arXiv:2301.12050 (2023) +23. Oord, A.v.d., Li, Y., Vinyals, O.: Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748 (2018) +24. OpenAI: Gpt-4 technical report (2023) +25. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., Krueger, G., Sutskever, I.: Learning transferable visual models from natural language supervision. In: International Conference on Machine Learning, ICML (2021) +26. Radford, A., Kim, J.W., Xu, T., Brockman, G., McLeavey, C., Sutskever, I.: Robust speech recognition via large-scale weak supervision (2022). https://doi.org/10.48550/ARXIV.2212.04356, https://arxiv.org/abs/2212.04356 +27. Schulman, J., Wolski, F., Dhariwal, P., Radford, A., Klimov, O.: Proximal policy optimization algorithms. arXiv preprint arXiv:1707.06347 (2017) +28. Shah, R., Wang, S.H., Wild, C., Milani, S., Kanervisto, A., Goecks, V.G., Waytowich, N.R., Watkins-Valls, D., Prakash, B., Mills, E., Garg, D., Fries, A., Souly, A., Chan, J.S., del Castillo, D., Lieberum, T.: Retrospective on the 2021 minerl BASALT competition on learning from human feedback. In: Neural Information Processing Systems (NeurIPS) Competitions and Demonstrations Track (2021) +29. Shu, T., Xiong, C., Socher, R.: Hierarchical and interpretable skill acquisition in multi-task reinforcement learning. In: International Conference on Learning Representations (ICLR) (2018) +30. Tessler, C., Givony, S., Zahavy, T., Mankowitz, D., Mannor, S.: A deep hierarchical approach to lifelong learning in apacheft. In: AAAI conference on artificial intelligence (AAAI) (2017) + +31. Wang, G., Xie, Y., Jiang, Y., Mandlekar, A., Xiao, C., Zhu, Y., Fan, L., Anandkumar, A.: Voyager: An open-ended embodied agent with large language models. arXiv preprint arXiv:2305.16291 (2023) +32. Xu, H., Ghosh, G., Huang, P.Y., Okhonko, D., Aghajanyan, A., Metze, F., Zettle-moyer, L., Feichtenhofer, C.: Videoclip: Contrastive pre-training for zero-shot video-text understanding. arXiv preprint arXiv:2109.14084 (2021) +33. Yuan, H., Zhang, C., Wang, H., Xie, F., Cai, P., Dong, H., Lu, Z.: Plan4mc: Skill reinforcement learning and planning for open-world minecraft tasks. arXiv preprint arXiv:2303.16563 (2023) +34. Zhou, C., Loy, C.C., Dai, B.: Extract free dense labels from clip. In: European Conference on Computer Vision. pp. 696-712. Springer (2022) +35. Zhu, X., Chen, Y., Tian, H., Tao, C., Su, W., Yang, C., Huang, G., Li, B., Lu, L., Wang, X., et al.: Ghost in the minecraft: Generally capable agents for open-world enviroments via large language models with text-based knowledge and memory. arXiv preprint arXiv:2305.17144 (2023) \ No newline at end of file diff --git a/2024/Reinforcement Learning Friendly Vision-Language Model for Minecraft/images.zip b/2024/Reinforcement Learning Friendly Vision-Language Model for Minecraft/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..9c31f12eb0c77b991d9bac134c1f9361b4e06a14 --- /dev/null +++ b/2024/Reinforcement Learning Friendly Vision-Language Model for Minecraft/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ddc4fd28a2e26ab93f322d485bd3e6c8ff63e9eb7f395eac464f6be6fc904b64 +size 320782 diff --git a/2024/Reinforcement Learning Friendly Vision-Language Model for Minecraft/layout.json b/2024/Reinforcement Learning Friendly Vision-Language Model for Minecraft/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..1e128e4ed1c86a636ff2698e84b5d7eb51a32406 --- /dev/null +++ b/2024/Reinforcement Learning Friendly Vision-Language Model for Minecraft/layout.json @@ -0,0 +1,7963 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 176, + 111, + 438, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 176, + 111, + 438, + 148 + ], + "spans": [ + { + "bbox": [ + 176, + 111, + 438, + 148 + ], + "type": "text", + "content": "Reinforcement Learning Friendly Vision-Language Model for Minecraft" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 203, + 167, + 410, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 203, + 167, + 410, + 194 + ], + "spans": [ + { + "bbox": [ + 203, + 167, + 410, + 194 + ], + "type": "text", + "content": "Haobin Jiang" + }, + { + "bbox": [ + 203, + 167, + 410, + 194 + ], + "type": "inline_equation", + "content": "^{1\\star}" + }, + { + "bbox": [ + 203, + 167, + 410, + 194 + ], + "type": "text", + "content": ", Junpeng Yue" + }, + { + "bbox": [ + 203, + 167, + 410, + 194 + ], + "type": "inline_equation", + "content": "^{1\\star}" + }, + { + "bbox": [ + 203, + 167, + 410, + 194 + ], + "type": "text", + "content": ", Hao Luo" + }, + { + "bbox": [ + 203, + 167, + 410, + 194 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 203, + 167, + 410, + 194 + ], + "type": "text", + "content": ", Ziluo Ding" + }, + { + "bbox": [ + 203, + 167, + 410, + 194 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 203, + 167, + 410, + 194 + ], + "type": "text", + "content": ", and Zongqing Lu" + }, + { + "bbox": [ + 203, + 167, + 410, + 194 + ], + "type": "inline_equation", + "content": "^{1,2\\dagger}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 206, + 201, + 407, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 206, + 201, + 407, + 236 + ], + "spans": [ + { + "bbox": [ + 206, + 201, + 407, + 236 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 206, + 201, + 407, + 236 + ], + "type": "text", + "content": " School of Computer Science, Peking University " + }, + { + "bbox": [ + 206, + 201, + 407, + 236 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 206, + 201, + 407, + 236 + ], + "type": "text", + "content": " Beijing Academy of Artificial Intelligence {haobin.jiang,zongqing.lu}@pku.edu.cn" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 160, + 264, + 455, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 264, + 455, + 485 + ], + "spans": [ + { + "bbox": [ + 160, + 264, + 455, + 485 + ], + "type": "text", + "content": "Abstract. One of the essential missions in the AI research community is to build an autonomous embodied agent that can achieve high-level performance across a wide spectrum of tasks. However, acquiring or manually designing rewards for all open-ended tasks is unrealistic. In this paper, we propose a novel cross-modal contrastive learning framework architecture, CLIP4MC, aiming to learn a reinforcement learning (RL) friendly vision-language model (VLM) that serves as an intrinsic reward function for open-ended tasks. Simply utilizing the similarity between the video snippet and the language prompt is not RL-friendly since standard VLMs may only capture the similarity at a coarse level. To achieve RL-friendliness, we incorporate the task completion degree into the VLM training objective, as this information can assist agents in distinguishing the importance between different states. Moreover, we provide neat YouTube datasets based on the large-scale YouTube database provided by MineDojo. Specifically, two rounds of filtering operations guarantee that the dataset covers enough essential information and that the video-text pair is highly correlated. Empirically, we demonstrate that the proposed method achieves better performance on RL tasks compared with baselines. The code and datasets are available at https://github.com/PKU-RL/CLIP4MC." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 160, + 494, + 435, + 507 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 494, + 435, + 507 + ], + "spans": [ + { + "bbox": [ + 160, + 494, + 435, + 507 + ], + "type": "text", + "content": "Keywords: Dataset " + }, + { + "bbox": [ + 160, + 494, + 435, + 507 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 494, + 435, + 507 + ], + "type": "text", + "content": " Multimodal model " + }, + { + "bbox": [ + 160, + 494, + 435, + 507 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 494, + 435, + 507 + ], + "type": "text", + "content": " Reinforcement learning" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 526, + 230, + 539 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 526, + 230, + 539 + ], + "spans": [ + { + "bbox": [ + 132, + 526, + 230, + 539 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 551, + 482, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 551, + 482, + 647 + ], + "spans": [ + { + "bbox": [ + 130, + 551, + 482, + 647 + ], + "type": "text", + "content": "Training reinforcement learning (RL) agents to perform complex tasks in a vision-based and open-ended world can be difficult. One main challenge is that manually specifying reward functions in all open-ended tasks is unrealistic [3,21], especially when we cannot access the internal configuration of the environment. In addition, learning a reward model from human feedback is typically expensive. To this end, MineDojo [8] has proposed an internet-scale, multi-modal knowledge base of YouTube videos to facilitate learning in an open-ended world. With the advent of a such large-scale database, agents are able to harvest practical" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 141, + 653, + 332, + 666 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 653, + 332, + 666 + ], + "spans": [ + { + "bbox": [ + 141, + 653, + 332, + 666 + ], + "type": "text", + "content": "* Equal contribution. † Corresponding author." + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "text", + "content": "knowledge encoded in large amounts of media like human beings. Moreover, a vision-language model (VLM), MineCLIP [8], is proposed to utilize the internet-scale domain knowledge. In more detail, the learned correlation score between the visual observation and the language prompt can be used effectively as an open-vocabulary, massively multi-task reward function for RL training. Therefore, no further task-specific reward design is needed for open-ended tasks." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 191, + 483, + 382 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 191, + 483, + 382 + ], + "spans": [ + { + "bbox": [ + 130, + 191, + 483, + 382 + ], + "type": "text", + "content": "However, the autonomous embodied agent requires the vision-language model to provide a more instructive correlation score. Given the partial observations, e.g., video snippet, and the language prompt that describes the task, the agent needs to figure out two non-trivial matters to better evaluate the current state. On the one hand, whether the target entities are present within its field of vision? MineCLIP tried to address this question. However, the alignment of texts and videos in the YouTube database is totally a catastrophe, which impedes the learning of VLM. On the other hand, what is the relationship between each video snippet and the degree of completion of the task? Normally, in an open-ended world, e.g. Minecraft, the agent explores first, then approaches and interacts with the target object. In other words, the agent requires approaching the target object before it takes trial-and-error, even for the target that needs to be kept away. Therefore, it is reasonable to make an assumption, namely, that the higher the level of completion of the task, the closer the agent is to the targets in the video snippet. We also argue it is important to incorporate the level of completion of the task into the reward function." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 387, + 483, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 387, + 483, + 555 + ], + "spans": [ + { + "bbox": [ + 130, + 387, + 483, + 555 + ], + "type": "text", + "content": "In this paper, we first construct neat YouTube datasets to facilitate the learning of basic game concepts, mainly the correspondence between the videos and texts. Though a large-scale database is provided by MineDojo, it contains significant noise due to its nature as an online resource. In addition, MineDojo only claims the training dataset is randomly sampled from the database, making it hard to reproduce. To overcome the catastrophic misalignment of video-text pairs in the original database, we have done four steps of dataset processing to guarantee the dataset is clean. Firstly, transcript cleaning enhances the accuracy of transcripts and ensures they are complete sentences. Secondly, keyword filtering ensures that the content of clips is relevant to the key entities in Minecraft, thereby facilitating the learning of basic game concepts. Thirdly, video partitioning and selection can handle the issue of scene transitions and thereby mitigate interference from other extraneous information. Lastly, correlation filtering can effectively address the issue of mismatch between video clips and transcripts." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 558, + 483, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 558, + 483, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 558, + 483, + 666 + ], + "type": "text", + "content": "We also propose an upgraded vision-language model, CLIP4MC, to provide a more RL-friendly reward function. In RL, simply utilizing the similarity between the video snippet and the language prompt is not RL-friendly since MineCLIP tends to only capture the similarity at the entity level. In other words, VLM can hardly reflect the relationship between each video snippet and the degree of completion of the task. However, this information can better help agents distinguish the importance between similar states. To achieve this, we incorporate the degree of task completion into the VLM training objective. In more detail, CLIP has exhibited remarkable segmentation capabilities without fine-tuning. After we" + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "type": "text", + "content": "H. Jiang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 247 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 247 + ], + "type": "text", + "content": "extend pre-trained MineCLIP with modifications inspired by MaskCLIP [34], it can segment the specified object from the image and label the size of the target shown in the corresponding video. Intuitively, the closer the agent is to the target, the larger the target size becomes. During the learning procedure, we dynamically control the degree of contrasting positive and negative pairs of instances based on the target size in this positive video sample. Thus, CLIP4MC can render a more RL-friendly reward signal that instructs the agent to learn tasks faster. Our proposed method is trained on our YouTube dataset and evaluated on MineDojo Programmatic tasks, including harvest, hunt, and combat tasks. Empirically, our results show that CLIP4MC can provide a more friendly reward signal for the RL training procedure." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 248, + 343, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 248, + 343, + 258 + ], + "spans": [ + { + "bbox": [ + 131, + 248, + 343, + 258 + ], + "type": "text", + "content": "To summarize, our contributions are as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 266, + 479, + 396 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 138, + 266, + 479, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 266, + 479, + 324 + ], + "spans": [ + { + "bbox": [ + 138, + 266, + 479, + 324 + ], + "type": "text", + "content": "- Open-sourced datasets: We provide two high-quality datasets. The first one undergoes data cleaning (Sections 4.1 to 4.3) and global-level correlation filtering (Section 4.4). The VLM trained on this dataset matches the performance of the officially released MineCLIP, which lacks a publicly available training set." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 325, + 479, + 360 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 325, + 479, + 360 + ], + "spans": [ + { + "bbox": [ + 138, + 325, + 479, + 360 + ], + "type": "text", + "content": "- RL-friendlyly dataset: Our second dataset further incorporates local-level correlation filtering (Section 4.4), making it more suited for RL. The VLM trained on this dataset outperforms that on the first dataset." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 361, + 479, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 361, + 479, + 396 + ], + "spans": [ + { + "bbox": [ + 138, + 361, + 479, + 396 + ], + "type": "text", + "content": "- RL-friendly VLM: To better evaluate and leverage the advantages of our more RL-friendly dataset, we introduce CLIP4MC, a novel method to train a VLM that could improve downstream RL performance (Section 5.2)." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 131, + 414, + 237, + 426 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 414, + 237, + 426 + ], + "spans": [ + { + "bbox": [ + 131, + 414, + 237, + 426 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 437, + 482, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 437, + 482, + 521 + ], + "spans": [ + { + "bbox": [ + 130, + 437, + 482, + 521 + ], + "type": "text", + "content": "Video-Text Retrieval. Video-text retrieval plays an essential role in multimodal research and has been widely used in many real-world web applications. Recently, the pre-trained models have dominated this line of research with noticeable results on both zero-shot and fine-tuned retrieval. Especially, BERT [5], ViT [6], and CLIP [25], are used as the backbones to extract the text or video embedding. The cross-modal embeddings are then matched with specific fusion networks to find the correct video-text pair." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 521, + 482, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 482, + 641 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 482, + 641 + ], + "type": "text", + "content": "In more detail, CLIP4Clip [20] proposes three different similarity modules to calculate the correlation between video and text embeddings. HiT [19] performs hierarchical matching at two different levels, i.e. semantic level and feature level. Note that semantic level and feature level features are from the transformer network's higher and lower feature layers, respectively. Frozen [1] proposes a dual encoder architecture that utilizes the flexibility of a transformer visual encoder to train from images or video clips with text captions. Moreover, MDMMT [7] adopts several pre-training models as encoders and it shows the CLIP-based model performs the best. Therefore, our model follows this line of research by using the pre-trained model, CLIP [25], to extract the feature embeddings." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 131, + 641, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 641, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 131, + 641, + 481, + 665 + ], + "type": "text", + "content": "Minecraft for AI Research. As an open-ended video game with an egocentric vision, Minecraft is a noticeable and important domain in RL due to the nature" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 403, + 90, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 403, + 90, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 403, + 90, + 447, + 101 + ], + "type": "text", + "content": "CLIP4MC" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 187 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 187 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 187 + ], + "type": "text", + "content": "of the sparse reward, large exploration space, and long-term episodes. Since the release of the Malmo simulator [14] and later the MineDojo simulator [8], various methods have attempted to train agents to complete tasks in Minecraft [10, 12, 18, 29, 30]. Approaches such as model-based RL, hierarchical RL, goal-based RL, and reward shaping have been adopted to alleviate the sparse reward and exploration difficulty for the agent in this environment." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 188, + 482, + 295 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 188, + 482, + 295 + ], + "spans": [ + { + "bbox": [ + 130, + 188, + 482, + 295 + ], + "type": "text", + "content": "Recently, with the development of large language models (LLM) like GPT-4 [24], a series of methods leveraging LLMs for high-level planning in Minecraft have been proposed [16, 22, 31, 33, 35]. These methods have demonstrated remarkable capabilities in guiding the agent to complete multiple complicated, long-horizon tasks, such as mining diamonds. These LLMs play a crucial role in decision-making, determining the sequence of basic skills required to accomplish specific tasks. Their effectiveness is due to their extensive knowledge about Minecraft, learned from the Internet, and their ability to reflect on real-time feedback from the game environment." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 296, + 482, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 296, + 482, + 415 + ], + "spans": [ + { + "bbox": [ + 130, + 296, + 482, + 415 + ], + "type": "text", + "content": "In addition to the use of LLMs, recent research attempts to incorporate Internet visual data into basic skill learning in Minecraft, beyond the traditional RL methods. MineRL [11] collected 60M player demonstrations with action labels, motivating some methods [16,28] based on behavior cloning. As well-labeled data is limited in quantity, MineCLIP [8] instead uses over 730K narrated Minecraft videos without action labels from YouTube. It aims to learn a vision-language model providing auxiliary reward signals, utilizing the vast and diverse data available on the Internet. Different from MineCLIP, VPT [2] uses action-labeled data to train an inverse dynamic model to label 70K hours of Internet videos and then conduct behavior cloning." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 416, + 482, + 477 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 416, + 482, + 477 + ], + "spans": [ + { + "bbox": [ + 130, + 416, + 482, + 477 + ], + "type": "text", + "content": "Unlike the existing approaches, which incorporate the human experience and require a large number of demonstrations with action labels to train the agent, our work follows the line of MineCLIP [8] and focuses on only using the data without action labels to assist agent learning in Minecraft, which is more friendly with data collection and has the potential to scale in the future." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 495, + 227, + 510 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 495, + 227, + 510 + ], + "spans": [ + { + "bbox": [ + 132, + 495, + 227, + 510 + ], + "type": "text", + "content": "3 Background" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 521, + 482, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 482, + 593 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 482, + 593 + ], + "type": "text", + "content": "MineDojo tasks. MineDojo [8] provides thousands of benchmark tasks, which can be used to develop generally capable agents in Minecraft. These tasks can be divided into two categories, Programmatic and Creative tasks. The former has ground-truth simulator states to assess whether the task has been completed. The latter, however, do not have well-defined success criteria and tend to be more open-ended, but have to be evaluated by humans." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": "We mainly focus on Programmatic tasks since they can be automatically assessed. Specifically, MineDojo provides 4 categories of programmatic tasks, including Harvest, Combat, Survival, and Tech Tree, with 1581 template-generated natural language goals to evaluate the agent's different capabilities. Among these tasks, Survival and Tech Tree tasks are harder than Harvest and Combat tasks. Currently, MineCLIP [8] only expresses promising potential in some Harvest and" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 141, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 141, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 141, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 226, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 226, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 226, + 102 + ], + "type": "text", + "content": "H. Jiang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 161, + 116, + 373, + 162 + ], + "blocks": [ + { + "bbox": [ + 161, + 116, + 373, + 162 + ], + "lines": [ + { + "bbox": [ + 161, + 116, + 373, + 162 + ], + "spans": [ + { + "bbox": [ + 161, + 116, + 373, + 162 + ], + "type": "image", + "image_path": "be300fae5c1064a448f8b924e71b7a754723634d5ac4539840950a0e1c1966db.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 396, + 118, + 449, + 124 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 396, + 118, + 449, + 124 + ], + "spans": [ + { + "bbox": [ + 396, + 118, + 449, + 124 + ], + "type": "text", + "content": "irrelevant video content" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 376, + 127, + 447, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 376, + 127, + 447, + 152 + ], + "spans": [ + { + "bbox": [ + 376, + 127, + 447, + 152 + ], + "type": "text", + "content": "... so here I'm with nick and we've got look at my inventory and his inventory looks similar this is gonna be interesting so four..." + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 161, + 168, + 373, + 212 + ], + "blocks": [ + { + "bbox": [ + 161, + 168, + 373, + 212 + ], + "lines": [ + { + "bbox": [ + 161, + 168, + 373, + 212 + ], + "spans": [ + { + "bbox": [ + 161, + 168, + 373, + 212 + ], + "type": "image", + "image_path": "b32ef89bcb2fe46bf77fcec0507609a18eaf98a96463e7ac3995570d3b8b0571.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 392, + 167, + 449, + 174 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 392, + 167, + 449, + 174 + ], + "spans": [ + { + "bbox": [ + 392, + 167, + 449, + 174 + ], + "type": "text", + "content": "mismatched video content" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 376, + 178, + 447, + 202 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 376, + 178, + 447, + 202 + ], + "spans": [ + { + "bbox": [ + 376, + 178, + 447, + 202 + ], + "type": "text", + "content": "... give it I'm a little chicken just laying around by any chance a lot of arrows going on down here I'm guessing most of ..." + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 161, + 217, + 373, + 261 + ], + "blocks": [ + { + "bbox": [ + 161, + 217, + 373, + 261 + ], + "lines": [ + { + "bbox": [ + 161, + 217, + 373, + 261 + ], + "spans": [ + { + "bbox": [ + 161, + 217, + 373, + 261 + ], + "type": "image", + "image_path": "da7af01c9c11db865bacd59f0bd001029674962d8210b69d0c492f200e7712dc.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 275, + 480, + 318 + ], + "lines": [ + { + "bbox": [ + 130, + 275, + 480, + 318 + ], + "spans": [ + { + "bbox": [ + 130, + 275, + 480, + 318 + ], + "type": "text", + "content": "Fig. 1: Illustration of the YouTube video database. The screenshots of video clips are on the left and key entities are circled in red. The corresponding transcript clips are on the right and key entities are marked in red. We give examples of irrelevant, mismatched, and matched video content in the YouTube video database." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 400, + 217, + 449, + 223 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 400, + 217, + 449, + 223 + ], + "spans": [ + { + "bbox": [ + 400, + 217, + 449, + 223 + ], + "type": "text", + "content": "matched video content" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 376, + 228, + 448, + 252 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 376, + 228, + 448, + 252 + ], + "spans": [ + { + "bbox": [ + 376, + 228, + 448, + 252 + ], + "type": "text", + "content": "... at some point you will lose there too or also hearts there spider was like being completely blown away! I don't assume anything at ..." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 343, + 480, + 378 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 343, + 480, + 378 + ], + "spans": [ + { + "bbox": [ + 130, + 343, + 480, + 378 + ], + "type": "text", + "content": "Combat tasks. Harvest means finding, obtaining, cultivating, or manufacturing hundreds of materials and objects. Combat means fighting various monsters and creatures that require fast reflexes and martial skills." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 380, + 481, + 489 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 380, + 481, + 489 + ], + "spans": [ + { + "bbox": [ + 130, + 380, + 481, + 489 + ], + "type": "text", + "content": "POMDP. We model the programmatic task as a partially observable Markov decision process (POMDP) [15]. At each timestep " + }, + { + "bbox": [ + 130, + 380, + 481, + 489 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 380, + 481, + 489 + ], + "type": "text", + "content": ", the agent obtains the partial observation " + }, + { + "bbox": [ + 130, + 380, + 481, + 489 + ], + "type": "inline_equation", + "content": "o_t" + }, + { + "bbox": [ + 130, + 380, + 481, + 489 + ], + "type": "text", + "content": " from the global state " + }, + { + "bbox": [ + 130, + 380, + 481, + 489 + ], + "type": "inline_equation", + "content": "s_t" + }, + { + "bbox": [ + 130, + 380, + 481, + 489 + ], + "type": "text", + "content": " and a language prompt " + }, + { + "bbox": [ + 130, + 380, + 481, + 489 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 130, + 380, + 481, + 489 + ], + "type": "text", + "content": ", takes action " + }, + { + "bbox": [ + 130, + 380, + 481, + 489 + ], + "type": "inline_equation", + "content": "a_t" + }, + { + "bbox": [ + 130, + 380, + 481, + 489 + ], + "type": "text", + "content": " following its policy " + }, + { + "bbox": [ + 130, + 380, + 481, + 489 + ], + "type": "inline_equation", + "content": "\\pi(a_t|o_t)" + }, + { + "bbox": [ + 130, + 380, + 481, + 489 + ], + "type": "text", + "content": ", and receives a reward " + }, + { + "bbox": [ + 130, + 380, + 481, + 489 + ], + "type": "inline_equation", + "content": "r_t = \\Phi(V_t, G)" + }, + { + "bbox": [ + 130, + 380, + 481, + 489 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 380, + 481, + 489 + ], + "type": "inline_equation", + "content": "V_t" + }, + { + "bbox": [ + 130, + 380, + 481, + 489 + ], + "type": "text", + "content": " is the fixed-length sequence of observations till " + }, + { + "bbox": [ + 130, + 380, + 481, + 489 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 380, + 481, + 489 + ], + "type": "text", + "content": " (thus a video snippet) and " + }, + { + "bbox": [ + 130, + 380, + 481, + 489 + ], + "type": "inline_equation", + "content": "\\Phi" + }, + { + "bbox": [ + 130, + 380, + 481, + 489 + ], + "type": "text", + "content": " maps " + }, + { + "bbox": [ + 130, + 380, + 481, + 489 + ], + "type": "inline_equation", + "content": "V_t" + }, + { + "bbox": [ + 130, + 380, + 481, + 489 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 380, + 481, + 489 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 130, + 380, + 481, + 489 + ], + "type": "text", + "content": " to a scalar value. Then the environment transitions to the next state " + }, + { + "bbox": [ + 130, + 380, + 481, + 489 + ], + "type": "inline_equation", + "content": "s_{t+1}" + }, + { + "bbox": [ + 130, + 380, + 481, + 489 + ], + "type": "text", + "content": " given the current state and action according to transition probability function " + }, + { + "bbox": [ + 130, + 380, + 481, + 489 + ], + "type": "inline_equation", + "content": "\\mathcal{T}(s_{t+1}|s_t, a_t)" + }, + { + "bbox": [ + 130, + 380, + 481, + 489 + ], + "type": "text", + "content": ". The agent aims to maximize the expected return " + }, + { + "bbox": [ + 130, + 380, + 481, + 489 + ], + "type": "inline_equation", + "content": "R = \\mathbb{E}_{\\pi} \\sum_{t=1}^{T} \\gamma^{t-1} r_t" + }, + { + "bbox": [ + 130, + 380, + 481, + 489 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 380, + 481, + 489 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 130, + 380, + 481, + 489 + ], + "type": "text", + "content": " is the discount factor and " + }, + { + "bbox": [ + 130, + 380, + 481, + 489 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 130, + 380, + 481, + 489 + ], + "type": "text", + "content": " is the episode time horizon." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 131, + 508, + 257, + 520 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 508, + 257, + 520 + ], + "spans": [ + { + "bbox": [ + 131, + 508, + 257, + 520 + ], + "type": "text", + "content": "4 YouTube Dataset" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": "The Internet is rich in Minecraft-related data, containing a wealth of weakly labeled or even unlabeled Minecraft knowledge, including crucial entities, plausible actions, and common-sense event processes. With these multi-modal data, it is possible to create dense language-conditioned rewards, making open-ended long-term task learning feasible. MineDojo [8] collected over 730K YouTube videos and their corresponding transcripts, totaling 33 years of video content and 2.2B words in transcripts. Around 640K video clips are selected using keyword-based filtering from these videos, and these clips are used to train MineCLIP. However, MineDojo open-sourced a 13.8M dataset, and the 640K video clips were randomly sampled in their paper. However, as illustrated in Figure 1, most videos feature irrelevant game content that is not conducive to learning basic game" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 403, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 403, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 403, + 91, + 447, + 100 + ], + "type": "text", + "content": "CLIP4MC" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "text", + "content": "concepts. Meanwhile, the alignment between the transcripts and videos may not always be precise, leading to temporal or content discrepancies that could hinder the learning of retrieval and RL tasks. Given the low quality of the data, we found it necessary to provide a neat 640k dataset to the community for the training usage. To address these issues, we adopt the following few steps to acquire a clean dataset with high quality." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 205, + 261, + 217 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 205, + 261, + 217 + ], + "spans": [ + { + "bbox": [ + 132, + 205, + 261, + 217 + ], + "type": "text", + "content": "4.1 Transcript Cleaning" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 224, + 482, + 344 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 224, + 482, + 344 + ], + "spans": [ + { + "bbox": [ + 130, + 224, + 482, + 344 + ], + "type": "text", + "content": "Downloading YouTube's automatic transcripts directly can lead to several issues. Firstly, transcript blocks may overlap, resulting in overlapping timestamps in the transcript. Additionally, the caption quality is typically mediocre, with a higher occurrence of misidentifications, especially in non-English videos. Moreover, the transcript lacks punctuation, making it less friendly for understanding semantics. Based on the aforementioned issues, we implement a pipeline to construct high-quality transcripts as follows: (1) Extract audio from the videos and use Whisper [26] to obtain high-quality, temporally non-overlapping transcripts. (2) Employ FullStop [9] to generate punctuation, resulting in complete sentences of 10-35 words in length." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 360, + 254, + 372 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 360, + 254, + 372 + ], + "spans": [ + { + "bbox": [ + 132, + 360, + 254, + 372 + ], + "type": "text", + "content": "4.2 Keyword Filtering" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 378, + 482, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 378, + 482, + 486 + ], + "spans": [ + { + "bbox": [ + 130, + 378, + 482, + 486 + ], + "type": "text", + "content": "Following MineDojo [8], we also implement keyword-based filtering to ensure that the content in our dataset is pertinent to the key entities in Minecraft, thereby facilitating the learning of basic game concepts. As essential components of Minecraft, the key entities, such as stones, trees, and sheep, are common across multiple tasks in MineDojo and videos from YouTube. Specifically, we identify entity keywords in the transcripts using a keyword list from MineDojo and extract transcript clips formed into sentences to encompass as many keywords as possible. These extracted transcript clips then serve as the textual component of our dataset, determining the location of corresponding video clips." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 502, + 310, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 502, + 310, + 514 + ], + "spans": [ + { + "bbox": [ + 132, + 502, + 310, + 514 + ], + "type": "text", + "content": "4.3 Video Partition and Selection" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 522, + 482, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 522, + 482, + 628 + ], + "spans": [ + { + "bbox": [ + 130, + 522, + 482, + 628 + ], + "type": "text", + "content": "After completing the previous two steps, we obtain transcript clips relevant to the keywords. For each transcript clip, we calculate the central timestamp that corresponds to the clip based on the transcript timestamps and then use this central timestamp to extract a video clip with a duration of " + }, + { + "bbox": [ + 130, + 522, + 482, + 628 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 130, + 522, + 482, + 628 + ], + "type": "text", + "content": " seconds from the video. This process allows us to obtain temporally-aligned video clips. However, the video clips obtained in this manner still exhibit some issues, including scene transitions and discrepancies between video content and transcripts. We handle the former problem through video partition and filtering, while the latter problem is addressed in Section 4.4." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 630, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 630, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 630, + 482, + 666 + ], + "type": "text", + "content": "Owing to the informal nature of YouTube content, there is often a lack of semantic congruence between the video clips and their corresponding transcriptions, as noted in VideoCLIP [32]. Moreover, video clips frequently contain a" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "type": "text", + "content": "H. Jiang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 202, + 114, + 413, + 248 + ], + "blocks": [ + { + "bbox": [ + 202, + 114, + 413, + 248 + ], + "lines": [ + { + "bbox": [ + 202, + 114, + 413, + 248 + ], + "spans": [ + { + "bbox": [ + 202, + 114, + 413, + 248 + ], + "type": "image", + "image_path": "a8d0bf9dbe81d2743b2257b8a10eebd1e3d489606188ad5647f2d136edf00ce6.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 256, + 482, + 291 + ], + "lines": [ + { + "bbox": [ + 131, + 256, + 482, + 291 + ], + "spans": [ + { + "bbox": [ + 131, + 256, + 482, + 291 + ], + "type": "text", + "content": "Fig. 2: Examples of how we estimate the size of key entities in video frames. Red bounding boxes are generated by our modified MineCLIP visual encoder, following the approach proposed in MaskCLIP [34]. These boxes are then used to calculate the size." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 314, + 480, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 314, + 480, + 373 + ], + "spans": [ + { + "bbox": [ + 130, + 314, + 480, + 373 + ], + "type": "text", + "content": "few different behaviors which cause scene transitions, e.g. Chopping down the tree first, then suddenly switching the inventory bar. Since some scene transitions lead to irrelevant information, we partition the video content into several semantically coherent segments based on the semantic structure of the video. Then we select the segment that aligns best with the transcript." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 374, + 482, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 374, + 482, + 469 + ], + "spans": [ + { + "bbox": [ + 130, + 374, + 482, + 469 + ], + "type": "text", + "content": "To achieve semantic partition of video content, we employ the Bellman " + }, + { + "bbox": [ + 130, + 374, + 482, + 469 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 374, + 482, + 469 + ], + "type": "text", + "content": "-segmentation algorithm [13]. This algorithm divides a sequence of data points into " + }, + { + "bbox": [ + 130, + 374, + 482, + 469 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 374, + 482, + 469 + ], + "type": "text", + "content": " distinct and constant-line segments, providing a piece-wise constant approximation of the data. To process the video, we first use the officially released MineCLIP [8] video encoder to obtain the embedding of each frame, since MineCLIP can capture video semantics to some extent. Subsequently, we partition these embeddings into " + }, + { + "bbox": [ + 130, + 374, + 482, + 469 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 374, + 482, + 469 + ], + "type": "text", + "content": " segments and select the segment with the highest similarity score, as calculated in MineCLIP." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 488, + 266, + 502 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 488, + 266, + 502 + ], + "spans": [ + { + "bbox": [ + 132, + 488, + 266, + 502 + ], + "type": "text", + "content": "4.4 Correlation Filtering" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 510, + 480, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 480, + 568 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 480, + 568 + ], + "type": "text", + "content": "We employ correlation filtering techniques to address disparities that persist between video content and transcripts. The correlation filtering is done at two levels, the global and the local levels. From the global level, we calculate the cosine similarities between video embeddings and text embeddings via the original MineCLIP and then select clips based on these similarities." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 570, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 481, + 665 + ], + "type": "text", + "content": "Recent research [17, 34] has demonstrated that CLIP [25], though trained on whole images, can generate meaningful local features. Inspired by this and following MaskCLIP [34], we make modifications to the original MineCLIP visual encoder, empowering it with the ability to estimate and label the size of the key entity, which is mentioned in the corresponding transcript, in each frame of a video clip without fine-tuning. Then the correlations between video clips and transcripts are calculated as the summation of the sizes across all frames in each video clip. We consider this correlation at the local level." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 403, + 90, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 403, + 90, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 403, + 90, + 447, + 100 + ], + "type": "text", + "content": "CLIP4MC" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "content": "Based on these two criteria, we select the top " + }, + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "inline_equation", + "content": "k\\%" + }, + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "content": " of clips, resulting in the final training set. We provide an elaboration on our implementation in Appendix A and illustrate some examples in Figure 2." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 152, + 482, + 248 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 152, + 482, + 248 + ], + "spans": [ + { + "bbox": [ + 130, + 152, + 482, + 248 + ], + "type": "text", + "content": "The aforementioned four-step approach creates a dataset consisting of 640K video-text clip pairs, with an additional 4K pairs extracted for validation of video-text retrieval. Regarding the constants of the approach, " + }, + { + "bbox": [ + 130, + 152, + 482, + 248 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 130, + 152, + 482, + 248 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 152, + 482, + 248 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 152, + 482, + 248 + ], + "type": "text", + "content": " are set to 16 and 50, respectively. Therefore, our dataset comprises videos with a total duration of one week and approximately 0.16B words, significantly smaller in scale compared to the original low-quality 13.8M dataset. Importantly, we will open-source our entire YouTube dataset, serving as an upgraded version of the unreleased MineCLIP training data." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 266, + 217, + 278 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 266, + 217, + 278 + ], + "spans": [ + { + "bbox": [ + 132, + 266, + 217, + 278 + ], + "type": "text", + "content": "5 CLIP4MC" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 292, + 480, + 363 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 292, + 480, + 363 + ], + "spans": [ + { + "bbox": [ + 130, + 292, + 480, + 363 + ], + "type": "text", + "content": "Given a video snippet " + }, + { + "bbox": [ + 130, + 292, + 480, + 363 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 130, + 292, + 480, + 363 + ], + "type": "text", + "content": " and a language prompt " + }, + { + "bbox": [ + 130, + 292, + 480, + 363 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 130, + 292, + 480, + 363 + ], + "type": "text", + "content": ", the vision-language model outputs a correlation score, " + }, + { + "bbox": [ + 130, + 292, + 480, + 363 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 130, + 292, + 480, + 363 + ], + "type": "text", + "content": ", that measures the similarity between the video snippet and the language prompt. Ideally, if the agent performs behaviors following the description of the language prompt, the vision-language model will generate a higher correlation score, leading to a higher reward. Otherwise, the agent will be given a lower reward." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 381, + 224, + 392 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 381, + 224, + 392 + ], + "spans": [ + { + "bbox": [ + 132, + 381, + 224, + 392 + ], + "type": "text", + "content": "5.1 Architecture" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 402, + 479, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 402, + 479, + 472 + ], + "spans": [ + { + "bbox": [ + 130, + 402, + 479, + 472 + ], + "type": "text", + "content": "We follow the same architecture design as MineCLIP [8], including a video encoder, a text encoder, and a similarity calculator. All the video frames first go through the spatial transformer to obtain a sequence of frame features. The temporal transformer is then utilized to summarize the sequence of frame features into a single video embedding. An adapter further processes the video embedding for better features. Refer to Appendix B for more details about the architecture." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 474, + 479, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 474, + 479, + 534 + ], + "spans": [ + { + "bbox": [ + 130, + 474, + 479, + 534 + ], + "type": "text", + "content": "Empirically, we found that the video encoder (essentially MineCLIP) can provide a bond between the entities and the language prompts and give a similar high reward as long as the target entities are present in the video frames, similar to observations in [4]. However, such a reward is not instructive enough for RL tasks since it does not reflect the behavioral trends of agents." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 551, + 265, + 563 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 551, + 265, + 563 + ], + "spans": [ + { + "bbox": [ + 132, + 551, + 265, + 563 + ], + "type": "text", + "content": "5.2 Contrastive Training" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 571, + 479, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 571, + 479, + 594 + ], + "spans": [ + { + "bbox": [ + 130, + 571, + 479, + 594 + ], + "type": "text", + "content": "We aim to minimize the sum of the multi-modal contrastive losses [23], including video-to-text, and text-to-video:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 182, + 605, + 480, + 632 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 182, + 605, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 182, + 605, + 480, + 632 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = - \\sum_ {\\left(z _ {v}, z _ {m}, z _ {t}\\right) \\in \\mathcal {B}} \\left(\\log \\operatorname {N C E} \\left(z _ {v}, z _ {t}\\right) + \\log \\operatorname {N C E} \\left(z _ {t}, z _ {v}\\right)\\right), \\tag {1}", + "image_path": "a333250d2a0ad5e03ac2af0e8ddb5e8dd460034ca713f69763d4b1b0e99d00bb.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 641, + 479, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 479, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 479, + 665 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 641, + 479, + 665 + ], + "type": "inline_equation", + "content": "\\mathcal{B}" + }, + { + "bbox": [ + 130, + 641, + 479, + 665 + ], + "type": "text", + "content": " is the batch containing sampled video-text pairs, and " + }, + { + "bbox": [ + 130, + 641, + 479, + 665 + ], + "type": "inline_equation", + "content": "\\mathrm{NCE}(\\cdot ,\\cdot)" + }, + { + "bbox": [ + 130, + 641, + 479, + 665 + ], + "type": "text", + "content": " is the contrastive loss that calculates the similarity of two inputs. To illustrate, the" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "type": "text", + "content": "H. Jiang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 192, + 116, + 425, + 340 + ], + "blocks": [ + { + "bbox": [ + 192, + 116, + 425, + 340 + ], + "lines": [ + { + "bbox": [ + 192, + 116, + 425, + 340 + ], + "spans": [ + { + "bbox": [ + 192, + 116, + 425, + 340 + ], + "type": "image", + "image_path": "1e7069af898432d6a42b90f799985badac834b84318181f38c21b8036c82daf4.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 350, + 480, + 373 + ], + "lines": [ + { + "bbox": [ + 131, + 350, + 480, + 373 + ], + "spans": [ + { + "bbox": [ + 131, + 350, + 480, + 373 + ], + "type": "text", + "content": "Fig. 3: Illustration of CLIP4MC training. The upper part shows the concept of contrastive learning, while the lower part explains the swapping operation." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 395, + 310, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 395, + 310, + 407 + ], + "spans": [ + { + "bbox": [ + 132, + 395, + 310, + 407 + ], + "type": "text", + "content": "video-to-text contrastive loss is given by" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 215, + 413, + 480, + 443 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 215, + 413, + 480, + 443 + ], + "spans": [ + { + "bbox": [ + 215, + 413, + 480, + 443 + ], + "type": "interline_equation", + "content": "\\mathrm {N C E} \\left(z _ {v}, z _ {t}\\right) = \\frac {\\exp \\left(z _ {v} \\cdot z _ {t} ^ {+} / \\lambda\\right)}{\\sum_ {z \\in \\left\\{z _ {t} ^ {+} , z _ {t} ^ {-} \\right\\}} \\exp \\left(z _ {v} \\cdot z / \\lambda\\right)}, \\tag {2}", + "image_path": "452a1b0d071b74530f176084d241af39c217cdc274856a1077aab148a1f8da75.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 450, + 480, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 450, + 480, + 498 + ], + "spans": [ + { + "bbox": [ + 130, + 450, + 480, + 498 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 450, + 480, + 498 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 130, + 450, + 480, + 498 + ], + "type": "text", + "content": " is a temperature hyperparameter, " + }, + { + "bbox": [ + 130, + 450, + 480, + 498 + ], + "type": "inline_equation", + "content": "z_{t}^{+}" + }, + { + "bbox": [ + 130, + 450, + 480, + 498 + ], + "type": "text", + "content": " is the positive text embedding matching with the video embedding " + }, + { + "bbox": [ + 130, + 450, + 480, + 498 + ], + "type": "inline_equation", + "content": "z_{v}" + }, + { + "bbox": [ + 130, + 450, + 480, + 498 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 450, + 480, + 498 + ], + "type": "inline_equation", + "content": "\\{z_t^-\\}" + }, + { + "bbox": [ + 130, + 450, + 480, + 498 + ], + "type": "text", + "content": " are negative text embeddings that are implicitly formed by other text clips in the training batch. Contrastive loss of text-to-video is defined in the same way." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 498, + 479, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 498, + 479, + 594 + ], + "spans": [ + { + "bbox": [ + 130, + 498, + 479, + 594 + ], + "type": "text", + "content": "In addition, we also need to incorporate the degree of task completion into the training objective. Specifically, we hope the similarity score between video clip and text embeddings could reflect the task completion degree, i.e., higher completion leads to higher similarity. The sizes of key entities provided by the local correlation filtering serve as a surrogate for the task completion degree. Initially, we try to dynamically adjust the weight of the positive pairs based on the size of the target object. However, it did not work. Essentially, positive and negative pairs will still be separated, even with smaller sample weights." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 594, + 480, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 480, + 665 + ], + "type": "text", + "content": "Therefore, we instead forcibly change the labels of the positive and negative video samples, as illustrated in Figure 3, thereby reducing the model's confidence in certain positive pairs. Specifically, during training, a positive video sample is swapped with a random negative video sample from the " + }, + { + "bbox": [ + 130, + 594, + 480, + 665 + ], + "type": "inline_equation", + "content": "\\mathcal{B}" + }, + { + "bbox": [ + 130, + 594, + 480, + 665 + ], + "type": "text", + "content": " based on probability " + }, + { + "bbox": [ + 130, + 594, + 480, + 665 + ], + "type": "inline_equation", + "content": "P_{\\mathrm{swap}}" + }, + { + "bbox": [ + 130, + 594, + 480, + 665 + ], + "type": "text", + "content": ". This swap only occurs when the labeled target size " + }, + { + "bbox": [ + 130, + 594, + 480, + 665 + ], + "type": "inline_equation", + "content": "H_bW_b" + }, + { + "bbox": [ + 130, + 594, + 480, + 665 + ], + "type": "text", + "content": " in the positive video sample is below a certain threshold " + }, + { + "bbox": [ + 130, + 594, + 480, + 665 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 130, + 594, + 480, + 665 + ], + "type": "text", + "content": ", and the smaller the size, the greater" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 403, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 403, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 403, + 91, + 447, + 100 + ], + "type": "text", + "content": "CLIP4MC" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 203, + 128 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 203, + 128 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 203, + 128 + ], + "type": "text", + "content": "the probability," + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 227, + 128, + 481, + 153 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 227, + 128, + 481, + 153 + ], + "spans": [ + { + "bbox": [ + 227, + 128, + 481, + 153 + ], + "type": "interline_equation", + "content": "P _ {\\mathrm {s w a p}} = P _ {\\mathrm {m a x}} * \\mathrm {R e L U} \\left(1 - \\frac {H _ {b} W _ {b}}{\\tau H W}\\right), \\tag {3}", + "image_path": "c4882dd2006608477e87a88184488fa9afedfcdf0bca0ce2b17b256ce693c352.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 155, + 327, + 166 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 155, + 327, + 166 + ], + "spans": [ + { + "bbox": [ + 132, + 155, + 327, + 166 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 132, + 155, + 327, + 166 + ], + "type": "inline_equation", + "content": "HW" + }, + { + "bbox": [ + 132, + 155, + 327, + 166 + ], + "type": "text", + "content": " represents the size of the image." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 167, + 481, + 237 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 167, + 481, + 237 + ], + "spans": [ + { + "bbox": [ + 130, + 167, + 481, + 237 + ], + "type": "text", + "content": "Since contrastive learning brings positive samples closer and pushes negative samples away, converting a positive pair to a negative one with a certain probability, i.e., the aforementioned random swap, during training will lower the similarity score. As the decrease in similarity is directly proportional to the swapping probability, we set " + }, + { + "bbox": [ + 130, + 167, + 481, + 237 + ], + "type": "inline_equation", + "content": "P_{\\mathrm{swap}}" + }, + { + "bbox": [ + 130, + 167, + 481, + 237 + ], + "type": "text", + "content": " to be inversely related to task completion, ensuing similarity score increases as completion increases." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 238, + 481, + 298 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 238, + 481, + 298 + ], + "spans": [ + { + "bbox": [ + 130, + 238, + 481, + 298 + ], + "type": "text", + "content": "In addition, the upper limit " + }, + { + "bbox": [ + 130, + 238, + 481, + 298 + ], + "type": "inline_equation", + "content": "P_{\\mathrm{max}}" + }, + { + "bbox": [ + 130, + 238, + 481, + 298 + ], + "type": "text", + "content": " in Equation (3) is set to 0.5, as the swap should not disrupt severely the normal distinction between positive and negative pairs in most cases. This constraint enables CLIP4MC to preserve the understanding capability of MineCLIP for general behaviors that may lack explicit target entities in Minecraft, as discussed in Appendix F." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 312, + 223, + 325 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 312, + 223, + 325 + ], + "spans": [ + { + "bbox": [ + 132, + 312, + 223, + 325 + ], + "type": "text", + "content": "5.3 RL Training" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 329, + 481, + 425 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 329, + 481, + 425 + ], + "spans": [ + { + "bbox": [ + 130, + 329, + 481, + 425 + ], + "type": "text", + "content": "For RL training, the first step is reward generation. At timestep " + }, + { + "bbox": [ + 130, + 329, + 481, + 425 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 329, + 481, + 425 + ], + "type": "text", + "content": ", we concatenate the agent's latest 16 egocentric RGB frames in a temporal window to form a video snippet, " + }, + { + "bbox": [ + 130, + 329, + 481, + 425 + ], + "type": "inline_equation", + "content": "V_{t}" + }, + { + "bbox": [ + 130, + 329, + 481, + 425 + ], + "type": "text", + "content": ". CLIP4MC outputs the probability " + }, + { + "bbox": [ + 130, + 329, + 481, + 425 + ], + "type": "inline_equation", + "content": "P_{G,t}" + }, + { + "bbox": [ + 130, + 329, + 481, + 425 + ], + "type": "text", + "content": " that calculates the similarity of " + }, + { + "bbox": [ + 130, + 329, + 481, + 425 + ], + "type": "inline_equation", + "content": "V_{t}" + }, + { + "bbox": [ + 130, + 329, + 481, + 425 + ], + "type": "text", + "content": " to the task prompt, " + }, + { + "bbox": [ + 130, + 329, + 481, + 425 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 130, + 329, + 481, + 425 + ], + "type": "text", + "content": ", against all other negative prompts. To compute the reward, we further process the raw probability as previous work [8] " + }, + { + "bbox": [ + 130, + 329, + 481, + 425 + ], + "type": "inline_equation", + "content": "r_t = \\max (P_{G,t} - 1 / N_t,0)" + }, + { + "bbox": [ + 130, + 329, + 481, + 425 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 329, + 481, + 425 + ], + "type": "inline_equation", + "content": "N_{t}" + }, + { + "bbox": [ + 130, + 329, + 481, + 425 + ], + "type": "text", + "content": " is the number of prompts passed to CLIP4MC. Note that CLIP4MC can handle unseen language prompts without any further fine-tuning due to the open-vocabulary ability of CLIP [25]." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 426, + 481, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 426, + 481, + 498 + ], + "spans": [ + { + "bbox": [ + 130, + 426, + 481, + 498 + ], + "type": "text", + "content": "The ultimate goal is to train a policy network that takes as input raw pixels and other structural data and outputs discrete actions to accomplish the task that is described by the language prompt, " + }, + { + "bbox": [ + 130, + 426, + 481, + 498 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 130, + 426, + 481, + 498 + ], + "type": "text", + "content": ". We use PPO [27] as our RL training backbone and the policy is trained on the CLIP4MC reward together with the sparse task reward if any. The policy input contains several modality-specific components and more details can be found in Appendix D." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 512, + 230, + 525 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 512, + 230, + 525 + ], + "spans": [ + { + "bbox": [ + 132, + 512, + 230, + 525 + ], + "type": "text", + "content": "6 Experiments" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 533, + 481, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 481, + 569 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 481, + 569 + ], + "type": "text", + "content": "In this section, we comprehensively evaluate and analyze our proposed model CLIP4MC, utilizing the open-ended platform MineDojo [8], which comprises thousands of diverse, open-ended Minecraft tasks designed for embodied agents." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "text", + "content": "We compare CLIP4MC against two baselines: (1) MineCLIP [official], the officially released MineCLIP model [8]. (2) MineCLIP [ours], using the same architecture as MineCLIP [official] but trained on our cleaned YouTube dataset. It also serves as the ablation of CLIP4MC without the swap operation. We train CLIP4MC and MineCLIP [ours] for 20 epochs and select the models with the highest performance on RL tasks. Please refer to Appendix C for more training details. All results are presented in terms of the mean and standard error of four runs with different random seeds." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "type": "text", + "content": "H. Jiang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 271, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 271, + 129 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 271, + 129 + ], + "type": "text", + "content": "6.1 Environment Settings" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 136, + 482, + 183 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 136, + 482, + 183 + ], + "spans": [ + { + "bbox": [ + 130, + 136, + 482, + 183 + ], + "type": "text", + "content": "We conduct experiments on eight Programmatic tasks, comprising two harvest tasks: milk a cow and shear wool, two combat tasks: combat a spider and combat a zombie, and four hunt tasks: hunt a cow, hunt a sheep, hunt a pig, and hunt a chicken. These tasks are all built-in tasks in the MineDojo benchmark." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 184, + 482, + 256 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 184, + 482, + 256 + ], + "spans": [ + { + "bbox": [ + 130, + 184, + 482, + 256 + ], + "type": "text", + "content": "Harvest. Milk a cow requires the agent to obtain milk from a cow with an empty bucket. Similarly, shear wool requires the agent to obtain wool from a sheep with shears. A harvest task is terminated and considered completed when the target item is obtained by the agent with a specified quantity. The prompts used to calculate the reward is \"obtain milk from a cow in plains with an empty bucket\"; for shear wool, it is \"shear a sheep in plains with shears\"." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 256, + 482, + 303 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 256, + 482, + 303 + ], + "spans": [ + { + "bbox": [ + 130, + 256, + 482, + 303 + ], + "type": "text", + "content": "Combat. In these tasks, target animals, spiders, and zombies, are hostile and will actively approach and attack the agent. The agent's goal is to fight and kill the target animal. The prompt for each combat task is \"combat a spider/zombie in plains with a diamond sword\"." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 304, + 482, + 423 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 304, + 482, + 423 + ], + "spans": [ + { + "bbox": [ + 130, + 304, + 482, + 423 + ], + "type": "text", + "content": "Hunt. Hunt tasks consist of hunt a cow, hunt a sheep, hunt a pig, and hunt a sheep. For each task, the agent's goal is to kill the target animal as indicated in the task name. Different from combat tasks, the target animals will flee from the agent after being attacked. Therefore, these tasks require the agent to keep chasing and attacking the target, making them challenging. As noted in [4], the original MineCLIP reward fails in these tasks since it cannot consistently increase when the agent approaches the agent. This observation aligns with our assertion that the original MineCLIP model can hardly capture the degree of task completion. The prompt for each task is \"hunt a {target} in plains with a diamond sword\" where {target} is replaced with the corresponding animal name." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 423, + 482, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 423, + 482, + 472 + ], + "spans": [ + { + "bbox": [ + 130, + 423, + 482, + 472 + ], + "type": "text", + "content": "More elaborated introduction of the Minecraft environment and settings of these tasks are available in Appendix D. To guarantee a fair comparison, we adopt the same RL hyperparameters for all models and tasks. These hyperparameters are listed in Appendix E." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 489, + 218, + 500 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 489, + 218, + 500 + ], + "spans": [ + { + "bbox": [ + 132, + 489, + 218, + 500 + ], + "type": "text", + "content": "6.2 RL Results" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 509, + 481, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 509, + 481, + 533 + ], + "spans": [ + { + "bbox": [ + 130, + 509, + 481, + 533 + ], + "type": "text", + "content": "Through our evaluation of CLIP4MC, MineCLIP [ours], and MineCLIP [official] across eight Minecraft tasks, we want to answer two key questions:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 534, + 481, + 582 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 131, + 534, + 481, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 534, + 481, + 556 + ], + "spans": [ + { + "bbox": [ + 131, + 534, + 481, + 556 + ], + "type": "text", + "content": "(1) Whether the YouTube dataset we constructed enables MineCLIP, when trained on it, to provide a more effective reward signal for task learning?" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 131, + 557, + 481, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 557, + 481, + 582 + ], + "spans": [ + { + "bbox": [ + 131, + 557, + 481, + 582 + ], + "type": "text", + "content": "(2) Whether our upgraded model, CLIP4MC, based on our YouTube dataset, offers a reward signal that is further friendly for the RL training procedure?" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 130, + 582, + 481, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 481, + 616 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 481, + 616 + ], + "type": "text", + "content": "These questions are central to verifying the effectiveness of our dataset and CLIP4MC model in Minecraft. Table 1 shows the success rates of all methods on the eight Minecraft tasks." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "It is noticeable that, in four hunt tasks, three models demonstrate varying performance on RL. Firstly, MineCLIP [ours] consistently achieves better results compared to MineCLIP [official] across all hunt tasks. The superior performance of MineCLIP [ours] provides a positive answer to our first question, suggesting" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 403, + 90, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 403, + 90, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 403, + 90, + 447, + 101 + ], + "type": "text", + "content": "CLIP4MC" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 101 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 101 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 101 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 135, + 157, + 476, + 319 + ], + "blocks": [ + { + "bbox": [ + 130, + 114, + 482, + 149 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 482, + 149 + ], + "type": "text", + "content": "Table 1: Success rates " + }, + { + "bbox": [ + 130, + 114, + 482, + 149 + ], + "type": "inline_equation", + "content": "(\\%)" + }, + { + "bbox": [ + 130, + 114, + 482, + 149 + ], + "type": "text", + "content": " of RL trained with rewards provided by different models on eight Minecraft tasks. Each mean and standard error of success rates are calculated on four models after training 1e6 environment steps with different random seeds." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 135, + 157, + 476, + 319 + ], + "lines": [ + { + "bbox": [ + 135, + 157, + 476, + 319 + ], + "spans": [ + { + "bbox": [ + 135, + 157, + 476, + 319 + ], + "type": "table", + "html": "
ModelsHarvestCombat
milk a cowshear woolcombat a spidercombat a zombie
CLIP4MC84.5±2.074.6±2.185.8±0.970.4±8.3
MineCLIP[ours]84.4±1.171.6±3.575.4±10.163.6±8.7
MineCLIP[official]84.1±0.573.2±1.882.7±2.557.4±3.7
ModelsHunt
hunt a cowhunt a sheephunt a pighunt a chicken
CLIP4MC39.8±2.545.9±7.230.6±8.426.1±3.5
MineCLIP[ours]17.3±10.633.0±18.114.1±10.615.3±10.6
MineCLIP[official]11.6±11.128.5±16.71.5±0.60.0±0.0
", + "image_path": "5a34dfc5818bca4c7a693da0018064b5724a868fe3def8236b5e57804484ca7a.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 347, + 482, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 347, + 482, + 456 + ], + "spans": [ + { + "bbox": [ + 130, + 347, + 482, + 456 + ], + "type": "text", + "content": "that the YouTube dataset we construct indeed enhances the effectiveness of the reward signal in MineCLIP for task learning. Note that this is the dataset we plan to release for better training the VLM model for RL tasks on Minecraft. Secondly, CLIP4MC shows significantly higher success rates on hunt tasks compared to both MineCLIP [ours] and MineCLIP [official], meaning that the answer to the second question is also positive. As CLIP4MC provides a reward signal taking into account a surrogate for the degree of task completion, i.e., the size of the target object in our implementation, it becomes more RL-friendly in these challenging tasks." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 459, + 482, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 459, + 482, + 555 + ], + "spans": [ + { + "bbox": [ + 130, + 459, + 482, + 555 + ], + "type": "text", + "content": "In addition, we notice a practical example of misalignment in the officially released MineCLIP model. Specifically, we observe that in hunt a chicken, the agent trained with MineCLIP [official] tends to keep looking at the sky, indicating such behavior can provide a high intrinsic reward. This phenomenon suggests that the officially released MineCLIP indeed suffers from the misalignment problem in the YouTube dataset. In contrast, our MineCLIP [official] shows promising behaviors on this task, demonstrating that our dataset processing improves the alignment between the transcripts and videos." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "text", + "content": "In contrast to hunt tasks, these methods perform comparably on harvest tasks and combat tasks, while CLIP4MC still shows marginal advantage over the other two methods. This finding aligns with the results reported in [8], where the MineCLIP reward already achieves saturated performance on these easy tasks. Therefore, further enhancements in the reward signal, like MineCLIP [ours] and CLIP4MC, do not significantly improve performance. Unlike hunt tasks, neither harvest nor combat tasks require the agent to take multiple rounds of chasing. This result does not detract from the superiority of CLIP4MC, evidenced by its performance on hunt tasks, which are considered more difficult [2]." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "type": "text", + "content": "H. Jiang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 176, + 114, + 255, + 177 + ], + "blocks": [ + { + "bbox": [ + 176, + 114, + 255, + 177 + ], + "lines": [ + { + "bbox": [ + 176, + 114, + 255, + 177 + ], + "spans": [ + { + "bbox": [ + 176, + 114, + 255, + 177 + ], + "type": "image", + "image_path": "fef7ef70be28bbb120a7d8b88b5956ec245bd7ca16eb0c4529f81fdf38f51cac.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 189, + 179, + 241, + 190 + ], + "lines": [ + { + "bbox": [ + 189, + 179, + 241, + 190 + ], + "spans": [ + { + "bbox": [ + 189, + 179, + 241, + 190 + ], + "type": "text", + "content": "(a) CLIP4MC" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 267, + 114, + 348, + 177 + ], + "blocks": [ + { + "bbox": [ + 267, + 114, + 348, + 177 + ], + "lines": [ + { + "bbox": [ + 267, + 114, + 348, + 177 + ], + "spans": [ + { + "bbox": [ + 267, + 114, + 348, + 177 + ], + "type": "image", + "image_path": "d6f8936764fc0ef98eb713c07bd06f885e82f4f3a81bd090e352396984838625.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 269, + 179, + 344, + 190 + ], + "lines": [ + { + "bbox": [ + 269, + 179, + 344, + 190 + ], + "spans": [ + { + "bbox": [ + 269, + 179, + 344, + 190 + ], + "type": "text", + "content": "(b) MineCLIP [ours]" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 131, + 198, + 479, + 220 + ], + "lines": [ + { + "bbox": [ + 131, + 198, + 479, + 220 + ], + "spans": [ + { + "bbox": [ + 131, + 198, + 479, + 220 + ], + "type": "text", + "content": "Fig. 4: Scatter plots illustrating the relationship between the entity size and the intrinsic reward. The red line indicates a linear fit to the data." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 359, + 114, + 439, + 177 + ], + "blocks": [ + { + "bbox": [ + 359, + 114, + 439, + 177 + ], + "lines": [ + { + "bbox": [ + 359, + 114, + 439, + 177 + ], + "spans": [ + { + "bbox": [ + 359, + 114, + 439, + 177 + ], + "type": "image", + "image_path": "7bc8aa2f72891617c881769faa584b7b2d959cf9b8226567d813845665890831.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 359, + 179, + 439, + 190 + ], + "lines": [ + { + "bbox": [ + 359, + 179, + 439, + 190 + ], + "spans": [ + { + "bbox": [ + 359, + 179, + 439, + 190 + ], + "type": "text", + "content": "(c) MineCLIP [official]" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 133, + 276, + 481, + 370 + ], + "blocks": [ + { + "bbox": [ + 131, + 233, + 480, + 266 + ], + "lines": [ + { + "bbox": [ + 131, + 233, + 480, + 266 + ], + "spans": [ + { + "bbox": [ + 131, + 233, + 480, + 266 + ], + "type": "text", + "content": "Table 2: Success rates (%) of RL trained with rewards provided by different models on three Minecraft tasks. Each mean and standard error of success rates are calculated on four models after training 5e5 and 1e6 environment steps with different random seeds." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 133, + 276, + 481, + 370 + ], + "lines": [ + { + "bbox": [ + 133, + 276, + 481, + 370 + ], + "spans": [ + { + "bbox": [ + 133, + 276, + 481, + 370 + ], + "type": "table", + "html": "
Modelsmilk a cowshear woolhunt a cow
5e51e65e51e65e51e6
ours73.1±0.984.4±1.147.5±6.371.6±3.515.6±8.717.3±10.6
w/o LCF71.3±1.880.7±1.153.8±3.373.6±1.50.7±0.712.1±12.0
RS65.1±3.981.7±1.021.7±6.568.5±3.61.0±0.54.0±2.6
official69.8±1.384.1±0.547.0±13.573.2±1.82.7±1.111.6±11.1
", + "image_path": "0efcb99cd90cc4676c4ee37b9a0b0a7ac7c29307ea06a95389dcca6a65e5ec90.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 131, + 396, + 246, + 407 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 396, + 246, + 407 + ], + "spans": [ + { + "bbox": [ + 131, + 396, + 246, + 407 + ], + "type": "text", + "content": "6.3 Reward Analysis" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 418, + 482, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 418, + 482, + 586 + ], + "spans": [ + { + "bbox": [ + 130, + 418, + 482, + 586 + ], + "type": "text", + "content": "To quantitatively verify that CLIP4MC captures the size of the target entity specified in the language prompt, we collect 5000 steps in task hunt a cow and apply the method described in Appendix A to estimate the maximal size of the cow in consecutive 16 frames. Then we use CLIP4MC, MineCLIP [ours], and MineCLIP [official] to calculate intrinsic rewards respectively. Before calculating the correlation between the size and intrinsic rewards, we transform the size value using " + }, + { + "bbox": [ + 130, + 418, + 482, + 586 + ], + "type": "inline_equation", + "content": "f(x) = \\ln (x + e^{-2})" + }, + { + "bbox": [ + 130, + 418, + 482, + 586 + ], + "type": "text", + "content": ", focusing on smaller values. The relationship between the transformed size and intrinsic rewards is visualized in Figure 4. The corresponding Pearson correlation coefficients from left to right are 0.81, 0.66, and 0.62, indicating that CLIP4MC reward has a higher correlation with the size, especially when it is relatively small. This is crucial in RL, as the agent needs dense and distinguishing reward signals to guide the learning process, particularly when the target is distant. Such characteristic is the essential benefit of CLIP4MC in RL." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 131, + 606, + 310, + 618 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 606, + 310, + 618 + ], + "spans": [ + { + "bbox": [ + 131, + 606, + 310, + 618 + ], + "type": "text", + "content": "6.4 Ablation on Dataset Filtering" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 629, + 480, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 480, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 480, + 666 + ], + "type": "text", + "content": "To evaluate the influence of the dataset quality on RL training and verify the effectiveness of our proposed correlation filter at the local level, we compare MineCLIP [ours] with two ablations: (1) MineCLIP [w/o LCF] is trained" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 403, + 90, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 403, + 90, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 403, + 90, + 447, + 100 + ], + "type": "text", + "content": "CLIP4MC" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 248 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 248 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 248 + ], + "type": "text", + "content": "on the dataset processed with only global correlation filtering, omitting the local correlation filtering. (2) MineCLIP [RS] is trained on 640K video clips randomly sampled from the MineDojo released database, consistent with their stated dataset construction method [8]. Note that methods evaluated here do not apply random swap introduced in Section 5.2 since these datasets cannot provide entity size. The results are presented in Table 2. Given the overall performance across all three presented tasks, firstly, [RS] does not perfectly reproduce the performance of [official], suggesting that misalignment in the original database hinders the reproduction of [official]. Secondly, part of our proposed data filtering makes the model, [w/o LCF], comparable to [official]. In addition, with all of our proposed data filtering techniques, [ours] outperforms [official]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 265, + 307, + 277 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 265, + 307, + 277 + ], + "spans": [ + { + "bbox": [ + 131, + 265, + 307, + 277 + ], + "type": "text", + "content": "6.5 Video-Text Retrieval Results" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 285, + 282, + 429 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 285, + 282, + 429 + ], + "spans": [ + { + "bbox": [ + 130, + 285, + 282, + 429 + ], + "type": "text", + "content": "Table 3 shows the results of video-to-text retrieval and text-to-video on the test set with the same MineCLIP model for a fair comparison. We train these models for 20 epochs and select the models with the highest R@1 value on the test set, respectively. From the results, the model trained on the randomly sampled dataset gets the worst performance than those trained on our YouTube dataset." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 430, + 282, + 549 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 430, + 282, + 549 + ], + "spans": [ + { + "bbox": [ + 130, + 430, + 282, + 549 + ], + "type": "text", + "content": "The results demonstrate that our neat dataset indeed can facilitate table observation is that [ours] a lower performance on retrieval to the video-text alignment objective [w/o LCF] is trained on the data which is directly conducted at the the dataset further filtered by local signing for RL training. Note that for better training the VLM model" + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 291, + 354, + 472, + 420 + ], + "blocks": [ + { + "bbox": [ + 287, + 321, + 482, + 354 + ], + "lines": [ + { + "bbox": [ + 287, + 321, + 482, + 354 + ], + "spans": [ + { + "bbox": [ + 287, + 321, + 482, + 354 + ], + "type": "text", + "content": "Table 3: Results of video-to-text / text-to-video retrieval on the test set. The best results are highlighted in bold." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 291, + 354, + 472, + 420 + ], + "lines": [ + { + "bbox": [ + 291, + 354, + 472, + 420 + ], + "spans": [ + { + "bbox": [ + 291, + 354, + 472, + 420 + ], + "type": "table", + "html": "
ModelsR@1R@5R@10
ours12.4/13.127.5/27.835.3/35.9
w/o LCF12.7/14.129.4/29.737.9/37.5
RS11.1/11.624.8/25.032.4/32.2
", + "image_path": "b3d43eac0b62f5f9e4505d62be3aa52d85bd7da4fc77f4307099a8ff0e3a9df1.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 567, + 220, + 580 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 567, + 220, + 580 + ], + "spans": [ + { + "bbox": [ + 131, + 567, + 220, + 580 + ], + "type": "text", + "content": "7 Conclusion" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 593, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 593, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 593, + 482, + 666 + ], + "type": "text", + "content": "We construct a neat YouTube dataset based on the large-scale YouTube database provided by MineDojo. Moreover, we introduce a novel cross-modal contrastive learning framework architecture, CLIP4MC, to learn an RL-friendly VLM that serves as an intrinsic reward function for open-ended tasks. Our findings suggest that our dataset enhances the acquisition of fundamental game concepts and CLIP4MC delivers a more effective reward signal for RL training." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "type": "text", + "content": "H. Jiang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 115, + 246, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 115, + 246, + 129 + ], + "spans": [ + { + "bbox": [ + 133, + 115, + 246, + 129 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 141, + 480, + 165 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 141, + 480, + 165 + ], + "spans": [ + { + "bbox": [ + 132, + 141, + 480, + 165 + ], + "type": "text", + "content": "This work was supported by NSFC under grant 62250068. The authors would like to thank the anonymous reviewers for their valuable comments." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 186, + 197, + 198 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 186, + 197, + 198 + ], + "spans": [ + { + "bbox": [ + 133, + 186, + 197, + 198 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 134, + 212, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 138, + 212, + 481, + 245 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 212, + 481, + 245 + ], + "spans": [ + { + "bbox": [ + 138, + 212, + 481, + 245 + ], + "type": "text", + "content": "1. Bain, M., Nagrani, A., Varol, G., Zisserman, A.: Frozen in time: A joint video and image encoder for end-to-end retrieval. In: IEEE/CVF International Conference on Computer Vision (ICCV) (2021)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 246, + 481, + 277 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 246, + 481, + 277 + ], + "spans": [ + { + "bbox": [ + 138, + 246, + 481, + 277 + ], + "type": "text", + "content": "2. Baker, B., Akkaya, I., Zhokhov, P., Huizinga, J., Tang, J., Ecoffet, A., Houghton, B., Sampedro, R., Clune, J.: Video pretraining (VPT): learning to act by watching unlabeled online videos. arXiv preprint arXiv:2206.11795 (2022)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 279, + 481, + 311 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 279, + 481, + 311 + ], + "spans": [ + { + "bbox": [ + 138, + 279, + 481, + 311 + ], + "type": "text", + "content": "3. Baumli, K., Baveja, S., Behbahani, F., Chan, H., Comanici, G., Flennerhag, S., Gazeau, M., Holsheimer, K., Horgan, D., Laskin, M., et al.: Vision-language models as a source of rewards. arXiv preprint arXiv:2312.09187 (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 312, + 481, + 344 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 312, + 481, + 344 + ], + "spans": [ + { + "bbox": [ + 138, + 312, + 481, + 344 + ], + "type": "text", + "content": "4. Cai, S., Wang, Z., Ma, X., Liu, A., Liang, Y.: Open-world multi-task control through goal-aware representation learning and adaptive horizon prediction. arXiv preprint arXiv:2301.10034 (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 346, + 481, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 346, + 481, + 376 + ], + "spans": [ + { + "bbox": [ + 138, + 346, + 481, + 376 + ], + "type": "text", + "content": "5. Devlin, J., Chang, M.W., Lee, K., Toutanova, K.: Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805 (2018)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 379, + 481, + 422 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 379, + 481, + 422 + ], + "spans": [ + { + "bbox": [ + 138, + 379, + 481, + 422 + ], + "type": "text", + "content": "6. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., Uszkoreit, J., Houlsby, N.: An image is worth 16x16 words: Transformers for image recognition at scale. In: International Conference on Learning Representations (ICLR) (2021)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 423, + 481, + 455 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 423, + 481, + 455 + ], + "spans": [ + { + "bbox": [ + 138, + 423, + 481, + 455 + ], + "type": "text", + "content": "7. Dzabraev, M., Kalashnikov, M., Komkov, S., Petiushko, A.: MDMMT: multidomain multimodal transformer for video retrieval. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops (2021)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 456, + 481, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 456, + 481, + 499 + ], + "spans": [ + { + "bbox": [ + 138, + 456, + 481, + 499 + ], + "type": "text", + "content": "8. Fan, L., Wang, G., Jiang, Y., Mandlekar, A., Yang, Y., Zhu, H., Tang, A., Huang, D.A., Zhu, Y., Anandkumar, A.: Minedojo: Building open-ended embodied agents with internet-scale knowledge. In: Neural Information Processing Systems (NeurIPS) Datasets and Benchmarks Track (2022)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 501, + 481, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 501, + 481, + 533 + ], + "spans": [ + { + "bbox": [ + 138, + 501, + 481, + 533 + ], + "type": "text", + "content": "9. Guhr, O., Schumann, A.K., Bahrmann, F., Böhme, H.J.: Fullstop: Multilingual deep models for punctuation prediction (June 2021), http://ceur-ws.org/Vol-2957/sepp_paper4.pdf" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 134, + 533, + 481, + 576 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 533, + 481, + 576 + ], + "spans": [ + { + "bbox": [ + 134, + 533, + 481, + 576 + ], + "type": "text", + "content": "10. Guss, W.H., Codel, C., Hofmann, K., Houghton, B., Kuno, N., Milani, S., Mohanty, S., Liebana, D.P., Salakhutdinov, R., Topin, N., et al.: Neurips 2019 competition: the minerl competition on sample efficient reinforcement learning using human priors. arXiv preprint arXiv:1904.10079 (2019)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 134, + 577, + 481, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 577, + 481, + 609 + ], + "spans": [ + { + "bbox": [ + 134, + 577, + 481, + 609 + ], + "type": "text", + "content": "11. Guss, W.H., Houghton, B., Topin, N., Wang, P., Codel, C., Veloso, M., Salakhutdinov, R.: Minerl: A large-scale dataset of apache demonstrations. In: International Joint Conference on Artificial Intelligence (IJCAI) (2019)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 134, + 611, + 481, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 611, + 481, + 632 + ], + "spans": [ + { + "bbox": [ + 134, + 611, + 481, + 632 + ], + "type": "text", + "content": "12. Hafner, D., Pasukonis, J., Ba, J., Lillicrap, T.P.: Mastering diverse domains through world models. arXiv preprint arXiv:2301.04104 (2023)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 134, + 633, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 633, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 134, + 633, + 481, + 665 + ], + "type": "text", + "content": "13. Haiminen, N., Gionis, A., Laasonen, K.: Algorithms for unimodal segmentation with applications to unimodality detection. Knowledge and information systems 14, 39-57 (2008)" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 403, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 403, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 403, + 91, + 447, + 100 + ], + "type": "text", + "content": "CLIP4MC" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 117, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 133, + 117, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 117, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 117, + 480, + 149 + ], + "type": "text", + "content": "14. Johnson, M., Hofmann, K., Hutton, T., Bignell, D.: The malmo platform for artificial intelligence experimentation. In: International Joint Conference on Artificial Intelligence (IJCAI) (2016)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 150, + 480, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 150, + 480, + 172 + ], + "spans": [ + { + "bbox": [ + 133, + 150, + 480, + 172 + ], + "type": "text", + "content": "15. Kaelbling, L.P., Littman, M.L., Cassandra, A.R.: Planning and acting in partially observable stochastic domains. Artificial intelligence 101(1-2), 99-134 (1998)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 172, + 480, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 172, + 480, + 226 + ], + "spans": [ + { + "bbox": [ + 133, + 172, + 480, + 226 + ], + "type": "text", + "content": "16. Kanervisto, A., Milani, S., Ramanauskas, K., Topin, N., Lin, Z., Li, J., Shi, J., Ye, D., Fu, Q., Yang, W., Hong, W., Huang, Z., Chen, H., Zeng, G., Lin, Y., Micheli, V., Alonso, E., Fleuret, F., Nikulin, A., Belousov, Y., Svidchenko, O., Shpilman, A.: Minerl diamond 2021 competition: Overview, results, and lessons learned. arXiv preprint arXiv:2202.10583 (2022)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 227, + 480, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 227, + 480, + 248 + ], + "spans": [ + { + "bbox": [ + 132, + 227, + 480, + 248 + ], + "type": "text", + "content": "17. Li, Y., Wang, H., Duan, Y., Li, X.: Clip surgery for better explainability with enhancement in open-vocabulary tasks. arXiv preprint arXiv:2304.05653 (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 249, + 480, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 249, + 480, + 281 + ], + "spans": [ + { + "bbox": [ + 132, + 249, + 480, + 281 + ], + "type": "text", + "content": "18. Lin, Z., Li, J., Shi, J., Ye, D., Fu, Q., Yang, W.: Juewu-mc: Playing minecraft with sample-efficient hierarchical reinforcement learning. arXiv preprint arXiv:2112.04907 (2021)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 281, + 480, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 281, + 480, + 315 + ], + "spans": [ + { + "bbox": [ + 132, + 281, + 480, + 315 + ], + "type": "text", + "content": "19. Liu, S., Fan, H., Qian, S., Chen, Y., Ding, W., Wang, Z.: Hit: Hierarchical transformer with momentum contrast for video-text retrieval. In: IEEE/CVF International Conference on Computer Vision (ICCV) (2021)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 315, + 480, + 347 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 315, + 480, + 347 + ], + "spans": [ + { + "bbox": [ + 132, + 315, + 480, + 347 + ], + "type": "text", + "content": "20. Luo, H., Ji, L., Zhong, M., Chen, Y., Lei, W., Duan, N., Li, T.: Clip4clip: An empirical study of CLIP for end to end video clip retrieval and captioning. Neurocomputing 508, 293-304 (2022)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 347, + 480, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 347, + 480, + 380 + ], + "spans": [ + { + "bbox": [ + 132, + 347, + 480, + 380 + ], + "type": "text", + "content": "21. Ma, Y.J., Liang, W., Wang, G., Huang, D.A., Bastani, O., Jayaraman, D., Zhu, Y., Fan, L., Anandkumar, A.: Eureka: Human-level reward design via coding large language models. arXiv preprint arXiv:2310.12931 (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 380, + 480, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 380, + 480, + 413 + ], + "spans": [ + { + "bbox": [ + 132, + 380, + 480, + 413 + ], + "type": "text", + "content": "22. Nottingham, K., Ammanabrolu, P., Suhr, A., Choi, Y., Hajishirzi, H., Singh, S., Fox, R.: Do embodied agents dream of pixelated sheep?: Embodied decision making using language guided world modelling. arXiv preprint arXiv:2301.12050 (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 413, + 480, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 413, + 480, + 435 + ], + "spans": [ + { + "bbox": [ + 132, + 413, + 480, + 435 + ], + "type": "text", + "content": "23. Oord, A.v.d., Li, Y., Vinyals, O.: Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748 (2018)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 435, + 307, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 435, + 307, + 446 + ], + "spans": [ + { + "bbox": [ + 132, + 435, + 307, + 446 + ], + "type": "text", + "content": "24. OpenAI: Gpt-4 technical report (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 446, + 480, + 489 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 446, + 480, + 489 + ], + "spans": [ + { + "bbox": [ + 132, + 446, + 480, + 489 + ], + "type": "text", + "content": "25. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., Krueger, G., Sutskever, I.: Learning transferable visual models from natural language supervision. In: International Conference on Machine Learning, ICML (2021)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 490, + 480, + 522 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 490, + 480, + 522 + ], + "spans": [ + { + "bbox": [ + 132, + 490, + 480, + 522 + ], + "type": "text", + "content": "26. Radford, A., Kim, J.W., Xu, T., Brockman, G., McLeavey, C., Sutskever, I.: Robust speech recognition via large-scale weak supervision (2022). https://doi.org/10.48550/ARXIV.2212.04356, https://arxiv.org/abs/2212.04356" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 522, + 480, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 522, + 480, + 544 + ], + "spans": [ + { + "bbox": [ + 132, + 522, + 480, + 544 + ], + "type": "text", + "content": "27. Schulman, J., Wolski, F., Dhariwal, P., Radford, A., Klimov, O.: Proximal policy optimization algorithms. arXiv preprint arXiv:1707.06347 (2017)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 544, + 480, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 544, + 480, + 599 + ], + "spans": [ + { + "bbox": [ + 132, + 544, + 480, + 599 + ], + "type": "text", + "content": "28. Shah, R., Wang, S.H., Wild, C., Milani, S., Kanervisto, A., Goecks, V.G., Waytowich, N.R., Watkins-Valls, D., Prakash, B., Mills, E., Garg, D., Fries, A., Souly, A., Chan, J.S., del Castillo, D., Lieberum, T.: Retrospective on the 2021 minerl BASALT competition on learning from human feedback. In: Neural Information Processing Systems (NeurIPS) Competitions and Demonstrations Track (2021)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 600, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 600, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 600, + 480, + 632 + ], + "type": "text", + "content": "29. Shu, T., Xiong, C., Socher, R.: Hierarchical and interpretable skill acquisition in multi-task reinforcement learning. In: International Conference on Learning Representations (ICLR) (2018)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "type": "text", + "content": "30. Tessler, C., Givony, S., Zahavy, T., Mankowitz, D., Mannor, S.: A deep hierarchical approach to lifelong learning in apacheft. In: AAAI conference on artificial intelligence (AAAI) (2017)" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "type": "text", + "content": "H. Jiang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 282 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "text", + "content": "31. Wang, G., Xie, Y., Jiang, Y., Mandlekar, A., Xiao, C., Zhu, Y., Fan, L., Anandkumar, A.: Voyager: An open-ended embodied agent with large language models. arXiv preprint arXiv:2305.16291 (2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 150, + 482, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 150, + 482, + 183 + ], + "spans": [ + { + "bbox": [ + 130, + 150, + 482, + 183 + ], + "type": "text", + "content": "32. Xu, H., Ghosh, G., Huang, P.Y., Okhonko, D., Aghajanyan, A., Metze, F., Zettle-moyer, L., Feichtenhofer, C.: Videoclip: Contrastive pre-training for zero-shot video-text understanding. arXiv preprint arXiv:2109.14084 (2021)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 183, + 482, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 183, + 482, + 215 + ], + "spans": [ + { + "bbox": [ + 130, + 183, + 482, + 215 + ], + "type": "text", + "content": "33. Yuan, H., Zhang, C., Wang, H., Xie, F., Cai, P., Dong, H., Lu, Z.: Plan4mc: Skill reinforcement learning and planning for open-world minecraft tasks. arXiv preprint arXiv:2303.16563 (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 216, + 482, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 216, + 482, + 237 + ], + "spans": [ + { + "bbox": [ + 130, + 216, + 482, + 237 + ], + "type": "text", + "content": "34. Zhou, C., Loy, C.C., Dai, B.: Extract free dense labels from clip. In: European Conference on Computer Vision. pp. 696-712. Springer (2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 237, + 482, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 237, + 482, + 282 + ], + "spans": [ + { + "bbox": [ + 130, + 237, + 482, + 282 + ], + "type": "text", + "content": "35. Zhu, X., Chen, Y., Tian, H., Tao, C., Su, W., Yang, C., Huang, G., Li, B., Lu, L., Wang, X., et al.: Ghost in the minecraft: Generally capable agents for open-world enviroments via large language models with text-based knowledge and memory. arXiv preprint arXiv:2305.17144 (2023)" + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 403, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 403, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 403, + 91, + 447, + 100 + ], + "type": "text", + "content": "CLIP4MC" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Reinforcement Learning Meets Visual Odometry/4903ec8e-7c0d-4a2f-9524-76e74784ea76_content_list.json b/2024/Reinforcement Learning Meets Visual Odometry/4903ec8e-7c0d-4a2f-9524-76e74784ea76_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..1a9c9918ab5bcab04842b03c92187712c388bf86 --- /dev/null +++ b/2024/Reinforcement Learning Meets Visual Odometry/4903ec8e-7c0d-4a2f-9524-76e74784ea76_content_list.json @@ -0,0 +1,1527 @@ +[ + { + "type": "text", + "text": "Reinforcement Learning Meets Visual Odometry", + "text_level": 1, + "bbox": [ + 223, + 141, + 779, + 165 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Nico Messikommer*, Giovanni Cioffi*, Mathias Gehrig, and Davide Scaramuzza", + "bbox": [ + 334, + 189, + 666, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Dept. of Informatics, University of Zurich {nmessi, cioffi, mgehrig, sdavide}@ifi.uzh.ch", + "bbox": [ + 336, + 232, + 666, + 262 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Visual Odometry (VO) is essential to downstream mobile robotics and augmented/virtual reality tasks. Despite recent advances, existing VO methods still rely on heuristic design choices that require several weeks of hyperparameter tuning by human experts, hindering generalizability and robustness. We address these challenges by reframing VO as a sequential decision-making task and applying Reinforcement Learning (RL) to adapt the VO process dynamically. Our approach introduces a neural network, operating as an agent within the VO pipeline, to make decisions such as keyframe and grid-size selection based on real-time conditions. Our method minimizes reliance on heuristic choices using a reward function based on pose error, runtime, and other metrics to guide the system. Our RL framework treats the VO system and the image sequence as an environment, with the agent receiving observations from keypoints, map statistics, and prior poses. Experimental results using classical VO methods and public benchmarks demonstrate improvements in accuracy and robustness, validating the generalizability of our RL-enhanced VO approach to different scenarios. We believe this paradigm shift advances VO technology by eliminating the need for time-intensive parameter tuning of heuristics.", + "bbox": [ + 261, + 289, + 743, + 554 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Visual Odometry $\\cdot$ Reinforcement Learning", + "bbox": [ + 261, + 566, + 640, + 583 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Multimedia Material The code is available at https://github.com/uzh-rpg/ml_vo and video at https://youtu.be/pt6yPTdQd6M", + "bbox": [ + 212, + 602, + 785, + 633 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 652, + 375, + 669 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The task of estimating the camera pose from a sequence of images, referred to as Visual Odometry (VO), finds broad applications in fields such as robotics and augmented/virtual reality [2, 13]. Despite considerable research and industry efforts to enhance the robustness of VO algorithms, state-of-the-art approaches grapple with generalizability [2, 25]. This limitation stems from the reliance on hyperparameters dictated by heuristic decision rules, which are highly scenario-dependent, varying with factors such as motion and lighting. Consequently, identifying the optimal hyperparameters often requires the intensive effort of domain experts, consuming several weeks.", + "bbox": [ + 212, + 681, + 787, + 818 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* equal contribution", + "bbox": [ + 217, + 825, + 357, + 840 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/b3ab776b463802e4d61e55173ddf90999eba927eb4852e6a6bc687c5f8e870ee.jpg", + "image_caption": [ + "Fig. 1: Our Framework. We propose to employ a learned agent to adaptively guide a VO method using real-time observations for enhanced robustness and accuracy. By considering the problem as a sequential decision process, we use RL to train the agent primarily based on the position error computed within a sliding window (dashed lines)." + ], + "image_footnote": [], + "bbox": [ + 305, + 147, + 697, + 330 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Existing VO methods can be positioned on a spectrum ranging from classical geometry-based approaches to end-to-end methods trained on extensive datasets. End-to-end approaches [23, 24, 26] aim to increase robustness by leveraging a high amount of image sequences with ground truth poses. However, they tend to struggle to adapt to visual scenarios and motion patterns beyond their training data distribution. On the other side of the spectrum are classical methods [3,5,6], which leverage geometry to map image points to and from the 3D world to compute the camera pose. Hybrid approaches [18, 19], positioned between these extremes, replace certain parts of the VO pipeline, usually the ones processing raw image pixels, with learned components while retaining well-defined geometric components to increase robustness.", + "bbox": [ + 212, + 429, + 787, + 594 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Despite these advancements, both hybrid and classical approaches still rely on heuristic design choices, rendering the optimal VO settings scene- and motion-dependent. Furthermore, these heuristics are defined by a fixed set of hyperparameters found through extensive offline tuning.", + "bbox": [ + 212, + 595, + 787, + 657 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "For example, one of the most impactful decisions to be taken by most of the VO systems during operation is the selection of keyframes. An adequate keyframe selection strategy should add informative keyframes to the current map to maximize the map coverage and, consequently, the accuracy of the registration of the pose of the incoming frame. Considering the limited map size due to runtime and memory usage constraints, the naive strategy of selecting each frame as a keyframe is oftentimes not the optimal one. In the current state-of-the-art VO systems [5, 6], keyframe selection strategies are based on heuristic if-else conditions, which rely on a fixed set of hyperparameters. These hyperparameters are generally tuned offline by expert users, which exploit the prior knowledge of the scene and the expected camera motion. For example, a strategy prone to often selecting a keyframe is preferred if the camera undergoes fast motion. Other", + "bbox": [ + 212, + 659, + 789, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "Messikommer et al.", + "bbox": [ + 271, + 114, + 401, + 126 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "hyperparameters such as feature extraction thresholds, brightness compensation, and weighting factors are also important, however, their impact on the overall performance depends on the specific VO system.", + "bbox": [ + 212, + 146, + 782, + 191 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To reduce the dependence on heuristic-based design, we approach the Visual Odometry problem from a sequential decision-making perspective, where current actions influence future states. Given the temporal dependencies inherent in this task, coupled with the absence of ground truth for optimal actions, we propose to train a neural network within a Reinforcement Learning (RL) framework to adapt a VO method at each timestep of the sequence. Operating as a dynamic agent, this neural network guides decisions within the VO pipeline, such as keyframe selection, or, in theory, any hyperparameter, based on real-time observations representing the current state of the VO system (for example, the number of points currently tracked and the distance from the past keyframe). Thereby minimizing the reliance on heuristic design choices while increasing the generalizability across diverse scenes and motions.", + "bbox": [ + 212, + 196, + 787, + 378 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To enforce an accurate and robust pipeline, the reward function is constructed from the pose error between predicted and ground truth poses but also includes other non-differentiable metrics such as runtime and size of the map. Importantly, our proposed framework leaves the underlying VO method unaltered while introducing a learned agent to enhance its robustness.", + "bbox": [ + 212, + 382, + 787, + 458 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In our RL approach, we consider the underlying image sequence and the VO system as the environment. From the environment, the agent receives observations in the form of tracked keypoints, map statistics, previously selected keyframes, and previously estimated poses. The agent itself consists of a multi-head attention layer to downproject a variable number of keypoints, which are then further processed along the fixed-sized inputs in a two-layer MLP. For training, we employ the on-policy RL algorithm Proximal Policy Optimization (PPO) [14] with a privileged critic network to stabilize the training. Finally, the reward includes the pose error between the predicted and ground truth pose computed in a sliding window approach while also penalizing decisions that lead to increased runtime.", + "bbox": [ + 212, + 462, + 787, + 628 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our experiments with the state-of-the-art VO methods [5, 6] on the VO benchmarks [1, 7, 16] show that our learned agent improves accuracy up to $19\\%$ , see RL DSO in Tab. 1, and increases robustness, see RL SVO in Tab. 1, where it tracks all the camera trajectories while the w/o RL baseline gets lost. Our contributions can be summarized as follows.", + "bbox": [ + 212, + 633, + 787, + 708 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. We propose to approach the Visual Odometry (VO) problem from a sequential decision perspective, where current actions influence future states.", + "2. We introduce an RL training framework designed to train an adaptive agent to replace heuristic design choices in VO methods. Our experimental results demonstrate that our approach not only boosts VO robustness and accuracy but also highlights the potential to eliminate labor-intensive fine-tuning." + ], + "bbox": [ + 222, + 744, + 784, + 839 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "Reinforcement Learning Meets Visual Odometry", + "bbox": [ + 408, + 114, + 730, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 143, + 388, + 160 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Three paradigms exist to approach Visual Odometry: classical geometric-based, end-to-end, and hybrid approaches.", + "bbox": [ + 212, + 175, + 782, + 205 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Classical approaches have been extensively studied in computer vision and robotics in the past 30 years [2, 13]. They achieve commercial-level accuracy and robustness. Indeed, classical VO algorithms are present in products like smartphones, VR-AR devices, and mobile robots. Classical VO algorithms are composed of two main modules: tracking and mapping. The tracking module uses camera images to estimate the motion. Three main approaches exist in the literature: direct methods [5], feature-based methods [11], and semi-direct methods [6]. Direct methods, such as DSO [5], work directly on the raw pixel intensities. These methods commonly extract image patches and estimate the camera trajectory by tracking the motion of such patches through consecutive images. The tracking is achieved by minimizing a photometric error defined on the raw pixel intensities. On the contrary, feature-based methods, such as ORBSLAM [11], extract points of interest, commonly known as visual features or keypoints, from the raw image pixels. The camera trajectory is estimated by tracking these points through consecutive images. Feature-based methods rely on scene texture to detect and track distinctive points, while direct methods use raw image pixels and can achieve higher reliability in low-texture environments. Semi-direct methods, such as SVO [6], combine keypoints and patches of raw pixel intensities to estimate the camera motion. SVO uses keypoints to find regions of interest in the current camera view and raw pixel intensities to track the patches centered at the detected keypoints.", + "bbox": [ + 212, + 205, + 787, + 521 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Following the deep learning revolution seen in computer vision in the past years, learning-based end-to-end VO systems [23, 24, 26] have been proposed. These methods usually rely on convolutional layers to process the input images and subsequently, recurrent and/or feedforward layers to predict the camera 6-DoF pose. They are trained with ground truth pose supervision. End-to-end VO systems show better performance than classical approaches in scenarios covered in the training data. However, they lack robustness when deployed on environments (scene and motion) outside the training data distribution.", + "bbox": [ + 212, + 523, + 787, + 643 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Hybrid methods [18, 19] combine the best of two paradigms: the geometric representation of the 3D world (classical methods), and the ability to process highly dimensional inputs (end-to-end methods). One of the state-of-the-art hybrid approaches is DROID-SLAM [18]. DROID-SLAM is a SLAM system that focuses on dense scene reconstruction, as well as, camera pose estimation. It uses a RAFT-inspired network architecture [17] followed by a differentiable dense bundle adjustment (BA) layer to iteratively update the camera poses and depth. The BA layer includes geometry constraints that improve the consistency of consecutive camera pose estimates. Using a similar differential BA layer, the hybrid method DPVO [19] tracks image patches over time using a recurrent network.", + "bbox": [ + 212, + 643, + 787, + 794 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The performance of VO algorithms highly depends on heuristics and hyperparameters. Usually, hyperparameter tuning is done manually by expert users. Due to the high dimensional space and the sequential nature of the VO prob-", + "bbox": [ + 212, + 795, + 787, + 839 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 230, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "Messikommer et al.", + "bbox": [ + 271, + 114, + 401, + 127 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "lem, automatic optimization-based hyperparameters tuning methods, such as Bayesian Optimization are difficult to use. An attempt to automatically select keyframes has been made in [15], where a deep network is used to decide if a new keyframe needs to be selected, as well as, to predict the camera pose.", + "bbox": [ + 212, + 146, + 782, + 205 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this paper, we model VO as a sequential decision-making task and apply Reinforcement Learning. Related to the problem of hyperparameter tuning with RL, Hyp-RL [10], poses the search for the best hyperparameters for training neural networks as a sequential decision process and uses RL to select the next hyperparameter set to be evaluated. In [4], RL is applied to dynamically choose the best set of hyperparameters of a deep network that performs object tracking.", + "bbox": [ + 212, + 207, + 782, + 297 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3 Method", + "text_level": 1, + "bbox": [ + 215, + 319, + 330, + 335 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/ad8f2e88031e9c9e80aa038834d2ce20128f278c540796b86fc1eb236ca40c3a.jpg", + "image_caption": [ + "Fig. 2: Agent Overview. Our agent takes as input the map statistics computed by the VO method at timestep $t_{i-1}$ and a variable number of keypoints using a multi-head attention layer. Given these inputs, a two-layer MLP computes a multi-discrete probability distribution over the binary keyframe action and the grid size action. These actions are then used inside the VO method to process the current frame at timestep $t_i$ , which will lead to the observation for the next timestep $t_{i+1}$ ." + ], + "image_footnote": [], + "bbox": [ + 233, + 375, + 767, + 518 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We consider monocular Visual Odometry (VO), whose primary objective is to estimate the camera pose for each image within an image sequence. Despite extensive research and widespread application, the robustness and generalizability of VO methods in different scenes remain an ongoing challenge. Both traditional geometry-based approaches and hybrid methods incorporating learned components often rely on heuristic decision rules for critical tasks such as keyframe selection and keypoint filtering. These heuristic rules significantly affect the performance of the VO method and are closely linked to the deployed scene. Consequently, a substantial amount of fine-tuning is required to adapt these rules effectively to each deployed scenario.", + "bbox": [ + 212, + 643, + 787, + 792 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To increase the robustness and generalizability of VO pipelines, we formulate the VO task as a sequential decision problem. Within this sequential framework, a deep agent is trained through RL. This agent is designed to make decisions", + "bbox": [ + 212, + 795, + 787, + 839 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "Reinforcement Learning Meets Visual Odometry", + "bbox": [ + 408, + 114, + 730, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "at each timestep for the VO method, taking into account the current state of the system, including map statistics, keypoints information and previous pose estimations. By training an adaptive agent, we can effectively reduce the reliance on heuristics and increase the generalizability of the underlying VO method across diverse scenes.", + "bbox": [ + 212, + 146, + 787, + 220 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The structure of this section is the following. In Sec. 3.1, we present our problem formulation. In Sec. 3.2, we introduce our proposed deep neural agent with its corresponding input observations and output actions. In Sec. 3.3, we formulate the reward function, while in Sec. 3.4, we present the RL training framework. Finally, in Sec. 3.5, we explain the specific implementation of our RL framework using SVO [6] and DSO [5] as the underlying VO methods.", + "bbox": [ + 212, + 223, + 787, + 315 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.1 Problem Formulation", + "text_level": 1, + "bbox": [ + 215, + 344, + 439, + 359 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We formulate the VO task as a sequential decision problem in which an agent interacts with its environment, see Fig. 1. Specifically, we consider the image sequence and the underlying VO pipeline to be fixed, both constituting the environment. The agent is a neural network that, at each timestep, selects actions for the VO pipeline, i.e., selecting keyframes and adapting the keypoint distribution by setting the grid size constraint (a grid is fitted to the image, and maximum one keypoint per grid cell is selected). Based on these actions, the VO method processes the next image of the sequence and computes the corresponding pose.", + "bbox": [ + 212, + 377, + 787, + 500 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "With this definition of the agent and environment, we formulate the VO task as a Markov decision process (MDP), denoted by the tuple $M = (S, A, p, r, \\gamma)$ . A state $s \\in S$ encapsulates the current image in the image sequence and the state of the VO pipeline, which includes properties such as feature tracks, previously estimated poses, the triangulate map, the optimization state, etc. To transition between VO states, actions $a \\in A$ are executed in the VO method, while the environment supplies the next frame of the sequence. The state transition probability $p$ defines the likelihood of moving from one state $s_i$ to another $s_j$ given actions $a_i$ . Finally, each state transition is associated with a reward $r(s, a)$ , dependent on the ensuing state and the chosen actions. The rewards of the future are discounted by $\\gamma$ . In the context of VO, the reward function evaluates the accuracy of the pose estimation and can also incorporate non-differentiable criteria such as runtime and map size, among others.", + "bbox": [ + 212, + 500, + 787, + 696 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In our formulation, the agent is characterized by a control policy $\\pi : S \\times A \\to \\mathbb{R}$ , where $\\pi$ specifies the actions $a \\in A$ to be taken in each state $s \\in S$ . Since the control policy is non-deterministic, it induces a probabilistic trajectory of states and actions $\\tau = \\{s_0, a_0, s_1, a_1, \\dots, s_T, a_{T-1}\\}$ , abbreviated as $\\tau \\sim \\pi$ . Following the control policy $\\pi$ , the expected sum over the discounted rewards for each timestep can be expressed with its value function", + "bbox": [ + 212, + 698, + 787, + 789 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nV ^ {\\pi} (s _ {i}) = \\mathbb {E} _ {\\tau \\sim \\pi} \\left[ \\sum_ {t = i} ^ {\\infty} \\gamma^ {t} r (s _ {t}, a _ {t}) \\right]. \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 385, + 804, + 785, + 844 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "Messikommer et al.", + "bbox": [ + 271, + 114, + 401, + 127 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The overarching goal is to find a policy $\\pi$ that maximizes the expected sum of discounted rewards.", + "bbox": [ + 212, + 146, + 782, + 175 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\pi^ {*} = \\arg \\max _ {\\pi} V ^ {\\pi} \\left(s _ {0}\\right) \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 426, + 176, + 785, + 196 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Given the reward formulation, the optimal policy provides decisions to the VO pipeline, aiming to minimize the expected position error while accounting for additional reward terms like runtime.", + "bbox": [ + 212, + 200, + 782, + 244 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.2 Deep Neural Agent", + "text_level": 1, + "bbox": [ + 214, + 263, + 421, + 279 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The central element of our proposed framework is the learned agent embodying the policy $\\pi$ , which selects optimal settings for the VO method based on the current keypoints and map statistics, i.e., relative poses to the newest and oldest keyframe. Given the inherent complexity of the underlying VO method, supplying the agent network with a comprehensive state encompassing all stored map elements, optimization states, etc., is neither feasible nor advantageous. Instead, the learned agent receives a selected subset of observations $o \\in O$ approximating the VO state within the policy $\\pi: O \\times A \\to \\mathbb{R}$ . In our system, the observations are derived from the local map and include relative poses to previous keyframes, along with the information about the currently tracked keypoints. Notably, the number of keypoints is variable and may vary depending on the characteristics of the image sequence. Therefore, the network architecture of the agent should be capable of processing inputs of variable sizes.", + "bbox": [ + 212, + 285, + 784, + 479 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To handle inputs of varying sizes, we leverage a Perceiver [8,9] architecture, illustrated in Fig. 2. Specifically, our approach involves a Variable Encoder that incorporates a multi-head attention layer [21] for processing the image positions, and the estimated depth of each tracked keypoint $k \\in \\mathbb{R}^{N \\times 3}$ . To project the information contained in the variable number of keypoints $N$ to a fixed dimension, we use $M$ number of learned tokens $z \\in \\mathbb{R}^{M \\times D}$ with dimension of $D$ as query $Q$ values. These query tokens are attended to by the projected keypoint information serving as both key $K$ and values $V$ . Finally, the fixed output of the Variable Encoder is subsequently processed along with the statistics of the local map, using a two-layer MLP with ReLu non-linearities.", + "bbox": [ + 212, + 482, + 784, + 632 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The output of the agent network is a discrete probability distribution over a set of actions expressed in the form of multiple independent categorical distributions. In our implementation, the agent predicts at each timestep a binary distribution over the decision to take the current frame as a keyframe. Additionally, in the case of SVO, the agent also predicts a separate and independent categorical distribution over a set of values $\\{20, 25, \\dots, 40\\}$ for the keypoint grid size. This grid size ensures a uniform distribution of keypoints on the image plane by enforcing that only one keypoint can reside within a given grid cell. Since DSO does not extract keypoints, there is no grid size selection necessary.", + "bbox": [ + 212, + 633, + 784, + 768 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.3 Reward", + "text_level": 1, + "bbox": [ + 214, + 787, + 326, + 801 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Another critical component within the RL framework is the design of the reward function. The feedback of the reward at each timestep guides the deep agent", + "bbox": [ + 212, + 809, + 782, + 839 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "Reinforcement Learning Meets Visual Odometry", + "bbox": [ + 408, + 114, + 732, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/482c301693015bf0c8054787fb1f741622030cf152b411d47e51d7ee866fd520.jpg", + "image_caption": [ + "Fig. 3: Position Error. To closely relate the pose prediction accuracy to the current action, we employ a sliding window of five timesteps to align the ground truth and estimated trajectory using a scale factor $s$ , a translation vector $t$ , and a rotation $R$ . The error at the current timestep $t_i$ is then used as the negative position reward." + ], + "image_footnote": [], + "bbox": [ + 305, + 146, + 699, + 239 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "toward a desired behavioral pattern. In the case of VO, naturally, the error between the estimated and the ground truth pose is the primary focus of the reward. However, since monocular VO methods cannot inherently estimate the absolute scale of the trajectory, it becomes necessary to scale the predicted poses before computing the error against the ground truth trajectory. Furthermore, another crucial aspect of ensuring stable feedback for the RL algorithm is to closely link the reward function of a timestep $t_i$ to the action taken at that specific timestep, thus mitigating the value assignment problem. Consequently, we adopt a sliding window approach to compute the error between the estimated and ground truth poses. Specifically, we employ a window of five timestamps $\\{t_{i-4}, t_{i-3}, \\dots, t_i\\}$ to compute the error at the current timestep $t_i$ , see Fig. 3. The first three poses in the window are used to align the estimated trajectory to the ground truth by minimizing the distances between the positions with the Umeyama method [20]. This alignment method accommodates missing scale information and differing coordinate origins, aligning both trajectories through an Euclidean transformation involving rotation, translation, and scaling.", + "bbox": [ + 212, + 344, + 787, + 585 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "By designing the position reward function, we can influence the policy to weigh accuracy higher at the cost of lower robustness or the other way around. This trade-off also shows the benefit of our dynamic agent over manual tuning since the decisions can be made in an online fashion, adapting to the current conditions in the image sequence. To handle this trade-off, we clip the positional error $e_{\\mathrm{tran}}$ around zero, see Eq. 3, to positively enforce accurate states while giving a penalty for states having a large positional error.", + "bbox": [ + 212, + 588, + 787, + 694 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Given that our framework leverages RL, we have the opportunity to integrate non-differentiable performance feedback into the reward system. Thus, we can enforce a faster runtime by penalizing the insertion of keyframes, i.e., directly penalizing the keyframe action $a_{\\mathrm{keyframe}}$ . Using the keyframe action instead of directly penalizing the runtime mitigates the dependence on the usage of the hardware during training. Our final reward function is formulated in Eq. 3, where we use 0.01 for $\\lambda_{1}$ and set $\\lambda_{2}$ to $5\\times 10^{-3}$ .", + "bbox": [ + 212, + 696, + 787, + 801 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\nr = \\lambda_ {1} \\max (- 1, 0. 2 - e _ {\\mathrm {t r a n}}) - \\lambda_ {2} a _ {\\text {k e y f r a m e}} \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 352, + 820, + 785, + 835 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "Messikommer et al.", + "bbox": [ + 271, + 114, + 403, + 127 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "3.4 Reinforcement Learning Framework", + "text_level": 1, + "bbox": [ + 215, + 146, + 555, + 161 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We use the on-policy algorithm Proximal Policy Optimization (PPO) [14] implemented in Stable Baselines3 [12] for training our agent. As commonly done, the RL training is conducted in two alternating stages: the rollout and the policy-update stage. During the rollout stage, the RL agent interacts with the environment, i.e., the underlying VO method and image sequences from TartanAir [25], via its actions to collect experiences comprising the actions, observations, and rewards at each timestep. While the experiences are collected, the network weights of the agent are not updated. In the second stage, the collected experiences of the rollout stage are used to update both the agent and the critic. The task of the critic is to approximate the value function, which represents the expected return of the current policy at a given state. The critic is used to stabilize the training inside the policy update steps in accordance with the PPO algorithm, which is designed to limit large changes in the policy.", + "bbox": [ + 212, + 176, + 787, + 372 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Since the performance of the VO pipelines depends not only on the decisions of the agent but also on the current image sequence, we introduce a privileged critic to stabilize the training. This privileged critic has access to both the current and future ground truth poses, enabling a better assessment of whether tracking performance was influenced by the decisions of the agent or the difficulty of the underlying image sequence. Importantly, this privileged critic with access to ground truth poses is only used during training.", + "bbox": [ + 212, + 373, + 787, + 479 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Another crucial aspect is the handling of the different stages of the VO system, which include the normal mode of camera tracking, the mode for the relocalization in the current map, and the initialization mode. Since the VO system only provides information about its full state in the tracking mode, we declare the current state a valid state if the previous state was in the tracking mode. Notably, the agent can exclusively make decisions in valid stages. To account for the valid states in the policy-update phase, we use a masked rollout buffer, which computes the return using all collected states, but samples only valid states for updating the policy network. As states in the initialization stages can not be influenced by the agent, we directly use the ground truth poses to triangulate tracked keypoints. The initialization with the ground truth poses also has the benefit of a significantly faster runtime compared to the original initialization.", + "bbox": [ + 212, + 479, + 787, + 661 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Since RL is sample inefficient, it is crucial to have a fast runtime of the environment interactions. For our RL framework, the runtime of each of the main components in the rollout phase using 100 SVO instances in parallel is reported in Tab. 6.", + "bbox": [ + 212, + 662, + 787, + 722 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "3.5 VO System", + "text_level": 1, + "bbox": [ + 215, + 750, + 356, + 763 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "In general, our proposed agent network can be combined with any VO system that relies on decision components. In this work, we use SVO [6] and DSO [5], which feature a very small runtime. Additionally, we also test on the slower SLAM pipeline ORB-SLAM3 [3]. From a practical viewpoint, the small runtime", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "Reinforcement Learning Meets Visual Odometry", + "bbox": [ + 408, + 114, + 730, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "and parallelization accelerate training and partially compensate for the inherent low sample efficiency of RL.", + "bbox": [ + 212, + 146, + 782, + 176 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "SVO is a semi-direct visual odometry composed of two main threads: motion estimation and mapping thread. The first step of the motion estimation thread is image alignment, where the pose of the new frame is estimated by aligning pixel intensity patches from the previous keyframe. The next step is feature alignment, where the positions of the keypoints on the current image are refined while the camera poses are fixed. The last step is a sparse bundle adjustment, where both the camera poses and the locations of the 3D points included in the current map are optimized. The motion thread assumes that the depth of the mapped 3D points is known. It is the objective of the mapping thread to estimate the depth. The mapping thread estimates the depth of each tracked pixel from multiple observations by means of a recursive Bayesian filter [22].", + "bbox": [ + 212, + 176, + 785, + 343 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "DSO is a direct visual odometry also composed of two threads: the front-end thread and the mapping thread. The front-end thread performs the initial frame tracking where the pose of the current frame is tracked with respect to the previous keyframe by aligning pixel intensities patches. The front-end is also in charge of selecting the active candidate points, performing outlier rejection and occlusion detection, and selecting keyframes. The mapping thread optimizes a non-linear cost function based on the photometric error. The camera poses, intrinsics, and inverse depth value of the active points in the current map are jointly optimized. This optimization can be seen as the photometric equivalent of sparse bundle adjustment.", + "bbox": [ + 212, + 343, + 787, + 492 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "We use SVO to run ablation studies on the contributions of the main components of the RL system (runtime, inputs, actions, and reward tuning).", + "bbox": [ + 212, + 494, + 785, + 526 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 215, + 549, + 375, + 566 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Dataset For training our agent, we leverage the synthetic sequences from TartanAir [25], a comprehensive dataset offering, among other measurements, ground truth camera poses. The dataset features a wide range of simulated scenes with challenging conditions such as low light and weather effects. The total number of different training sequences is 337, amounting to 279987 images or timesteps for the training of the agent. Importantly, we consider the dataset in our RL framework to be fixed without the possibility of the agent influencing the next image. Notably, as our method is trained on abstractions rather than image pixels, the risk of domain shift is mitigated.", + "bbox": [ + 212, + 582, + 787, + 717 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "For evaluation, we use three VO benchmarks: EuRoC [1], TUM-RGBD [16] and KITTI [7]. The EuRoC dataset consists of 11 sequences recorded via a drone, with ground truth pose measurements obtained from a motion capture system or a laser tracker. The TUM-RGBD dataset features 9 sequences recorded via a hand-held camera with ground truth poses provided by a motion capture system. For the results on the KITTI dataset, we refer to the supplementary.", + "bbox": [ + 212, + 718, + 785, + 809 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Evaluation Following standard practice in evaluating VO algorithms, we use the Absolute Translation Error (ATE [m]) [27] as the error metric for the EuRoC and", + "bbox": [ + 212, + 809, + 785, + 840 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "Messikommer et al.", + "bbox": [ + 271, + 114, + 401, + 127 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "TUM-RGBD datasets. Before computing the error metric, the estimated trajectories are aligned to the ground truth trajectory using the Umeyama method [20]. For the relative position error on EuRoC and TUM-RGBD, we refer to the supplementary.", + "bbox": [ + 212, + 146, + 782, + 205 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Since we want to guarantee a fair comparison between the methods with and without RL, we remove the effect of the initialization phase at the beginning of each sequence by providing the ground truth poses to the VO algorithm. Similar to the original initialization of the underlying VO, extracted keypoints are tracked from the first frame of the sequence and triangulated once a large enough disparity is detected. However, instead of estimating a relative pose, the ground truth poses are used for the triangulation. Following this procedure, all of the tested methods start with the same high-quality map, effectively eliminating the unwanted influence of the initialization. Naturally, the ground truth initialization is not used for the comparison against state-of-the-art VO methods. Instead, we take the average of the performance metrics for five runs.", + "bbox": [ + 212, + 207, + 784, + 371 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/f2c9368a7608c352a368012f47362956f219586ba42f715df1e58e761a77042f.jpg", + "image_caption": [ + "Fig. 4: Keyframe Selection. The number of selected keyframes by SVO at different translational and angular velocities for the EuRoC dataset." + ], + "image_footnote": [], + "bbox": [ + 217, + 396, + 488, + 492 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/77dcf3b29f3fdd213401076bc0920635636f2a0adb695fb07c7cb24c5b1abc25.jpg", + "image_caption": [ + "Fig.5: Top-Down View of Trajectories. Visualization of the predicted trajectory using SVO with and without RL." + ], + "image_footnote": [], + "bbox": [ + 514, + 398, + 784, + 503 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/a7ad84ea594b807a5d9dd7706b1686fbdb9d2bac76bb4c6be6eaea58d28f4119.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodMH01MH02MH03MH04MH05V101V102V103V201V202V203Avg
SVO0.1830.4900.3813.2361.2930.3200.333-0.0840.5531.431-
RL SVO0.2360.1240.2412.1491.9820.2250.3980.3940.0690.3341.5060.696
DSO0.0870.0440.3540.1840.2601.2231.6241.1250.0590.0651.5160.594
RL DSO0.0370.0330.2410.1681.1920.2070.4991.4500.0430.1641.2860.483
", + "bbox": [ + 220, + 609, + 782, + 660 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Table 1: EuRoC. We report the ATE [m] of our RL agents and the heuristic-based VO methods with parameters obtained by an extensive grid search. Our RL agents outperform the heuristic-based VO methods in terms of accuracy (lower average error) and robustness (RL SVO can complete the sequence V103 while the baseline SVO fails to track all the camera poses due to fast rotations).", + "bbox": [ + 212, + 661, + 784, + 729 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4.1 Results", + "text_level": 1, + "bbox": [ + 214, + 789, + 323, + 801 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "EuRoC We compare our RL agent with SVO and DSO against the heuristic-based VO with parameters obtained by tuning on this dataset using an extensive", + "bbox": [ + 212, + 809, + 782, + 839 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "Reinforcement Learning Meets Visual Odometry", + "bbox": [ + 408, + 114, + 730, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/10f2f41a6ea44cb0d5f253fea05e06419d20b61b65eddf074db42601e265f0e0.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Method360deskdesk2floorplantroomrpyteddyxyzAvg
SVO0.1860.6810.898-0.3200.8050.0530.7690.0570.471
RL SVO0.1890.5560.755-0.2600.8040.0560.6970.0620.422
DSO-0.2370.893-0.3200.9310.057-0.053-
RL DSO0.1870.6820.888-0.625-0.0540.7560.105-
", + "bbox": [ + 240, + 143, + 766, + 196 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/c976f4271e3a598c8c99085310176680414cf46e93a91b6d2c5bafcd01f64549.jpg", + "table_caption": [ + "Table 2: TUM-RGBD. We report the ATE [m] of our RL agents and the heuristic-base VO methods with parameters obtained by an extensive grid search. Our RL SVO outperforms the heuristic-based SVO methods in terms of accuracy (lower average error). Although this dataset is challenging for DSO, due to the use of a rolling shutter camera, RL DSO improves the robustness with two more sequences completed." + ], + "table_footnote": [], + "table_body": "
MethodMH01MH02MH03MH04MH05V101V102V103V201V202V203Avg
TartanVO [24]0.6390.3250.5501.1531.0210.4470.3890.6220.4330.7491.1520.680
DROID-VO [18]0.1630.1210.2420.3990.2700.1030.1650.1580.1020.1150.2040.186
DPVO [19]0.0870.0550.1580.1370.1140.0500.1400.0860.0570.0490.2110.105
RL SVO0.2891.5990.5522.4071.8260.2290.8130.4960.0850.9651.4340.972
RL DSO0.0320.0370.1660.1420.1100.1070.0891.3310.0410.1521.3130.320
", + "bbox": [ + 222, + 295, + 779, + 356 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Table 3: EuRoC. We report the ATE [m] of our RL agents and the state-of-the-art VO methods. Our RL DSO achieves the best accuracy in 5 sequences.", + "bbox": [ + 215, + 356, + 784, + 383 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "grid search. Qualitative estimation results of SVO are shown in the left plot in Fig. 5 from a top-down view of the trajectory estimated by RL SVO, the baseline VO, and the ground truth. Tab. 1 reports the ATE. In the case of SVO, the combination with our RL agent achieves the best ATE in most of the sequences. Remarkably, our RL SVO can estimate the entire camera trajectory in the sequence V103 while the baseline fails. This sequence is one of the most difficult in the EuRoC dataset due to fast camera rotations. The benefit of our RL agent is also evident for DSO, which achieves consistently a lower tracking error, leading to a $19\\%$ lower mean tracking error. We report in Tab. 3 the comparison of our RL agent against the state-of-the-art VO methods, which are taken from [19]. The results reported for our methods are averaged over five runs. Our RL DSO achieves the best ATE in five sequences.", + "bbox": [ + 212, + 429, + 787, + 609 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "TUM-RGBD Similar to EuRoC, we use a parameter grid search to find the best parameters for the baseline DSO and SVO methods. We report the ATE in Tab. 2, which shows that SVO RL achieves a lower ATE in most of the sequences and a lower overall mean error. Since the images of TUM-RGBD are not captured with a global shutter camera, DSO, relying on photometric alignment, struggles to finish all the sequences. Nevertheless, our RL agent improves the robustness in some challenging conditions with two more sequences successfully completed. For completeness, the top-down view of the trajectory estimated by RL SVO, the baseline, and the ground truth are visualized in the right plots in Fig. 5. We report in Tab. 4 the comparison of our RL agent against the state-of-the-art VO methods. The results reported for our methods are averaged over five runs. The results of the state-of-the-art methods are taken from [19].", + "bbox": [ + 212, + 611, + 787, + 792 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Keyframe Selection In Fig. 4, we give insights into the keyframe selection strategy employed by our RL agent. We visualize the histogram of states at which keyframes are selected and their corresponding rotational and translation", + "bbox": [ + 212, + 794, + 787, + 839 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "Messikommer et al.", + "bbox": [ + 271, + 114, + 401, + 127 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/e9fd4308e73a2dd7c537ad1a08b169f84d341608f87f84bd1b5b2354743f82f7.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Method360deskdesk2floorplantroomryteddyxyzAvg
DROID-VO [18]0.1610.0280.0990.0330.0280.3270.0280.1690.0130.098
DPVO [19]0.1350.0380.0480.0400.0360.3940.0340.0640.0120.089
RL SVO0.1830.6480.790-0.3320.7860.0520.7410.071-
RL DSO-0.4380.765-0.479---0.057-
", + "bbox": [ + 240, + 143, + 764, + 196 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/ed8f1c26f19076dfeba5ff4b2cc25db4af9fde2c0a3dacb9a0ac9c5f436309ca.jpg", + "table_caption": [ + "Table 4: TUM-RGBD. We report the ATE [m] of our RL agents and the state-of-the-art VO methods." + ], + "table_footnote": [], + "table_body": "
MethodMH01MH02MH03MH04MH05V101V102V103V201V202V203Avg
SVO0.1830.4900.3813.2361.2930.3200.333-0.0840.5531.431-
RL SVO0.2360.1240.2412.1491.9820.2250.3980.3940.0690.3341.5060.696
RL SVO (w/o keypoints)0.2550.1600.4191.9790.5880.1730.6440.8490.1821.394--
RL SVO (w/o privileged c.)0.1300.0990.8343.2830.9580.1400.2860.4590.1160.6181.3730.754
RL SVO (w/o keyframe)0.1790.3510.3023.0721.0210.325--0.1790.2581.464-
RL SVO (w/o gridsize)0.1520.1830.2932.1541.5370.1980.2690.8080.1081.6831.4050.799
", + "bbox": [ + 223, + 251, + 782, + 318 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/0849402af912da5390584c7c03c127d0b3562934a4493a34879213a88b80950d.jpg", + "table_caption": [ + "Table 5: EuRoC Ablations. Ablation study on inputs and action of the RL agent." + ], + "table_footnote": [], + "table_body": "
ComponentsTime [ms]
- 100 Parallel SVO Step73.9
- Load 100 Images18.6
- Network Forward Pass3.1
Complete Rollout Step104.1
", + "bbox": [ + 230, + 359, + 473, + 412 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/22504a8c765525510eac479e76ecbcf05889f05c28d9ace2fa22d1cbe7ffb655.jpg", + "table_caption": [ + "Table 6: Rollout Runtime. The runtime for the main components in the rollout phase averaged over 500 iterations." + ], + "table_footnote": [], + "table_body": "
MethodTime [ms]
SVO9.06
RL SVO7.40
DSO34.45
RL DSO32.29
Network inference1.75
", + "bbox": [ + 529, + 359, + 772, + 422 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Table 7: Inference Runtime. The runtime of SVO, DSO, and one network forward pass.", + "bbox": [ + 511, + 424, + 784, + 465 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "velocity. The keyframe selection using the RL agent has a similar distribution around the motion patterns, while the heuristic rules select more keyframes. Following the motion distribution of the dataset, both methods trigger more keyframes in small motions. Despite the fewer keyframes, our RL agent still achieves a more robust performance. Without the possibility to set keyframes, namely, the agent only determines the feature grid size, robustness is drastically reduced, as it can be seen in Tab. 5 by the entry RL SVO (w/o keyframe).", + "bbox": [ + 212, + 510, + 784, + 614 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Grid size Selection The positive impact of dynamically selecting the grid size for SVO can be observed by removing the grid size action, which lowers the pose estimation accuracy, see RL SVO (w/o gridsize) in Tab. 5. The supplementary video shows that the RL agent learns to reduce the grid size to the minimal", + "bbox": [ + 212, + 616, + 784, + 676 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/0b3b9bc89f1f8d5ab149d9bc58f3ccc889990390966eeb151a405d3f1f516f8d.jpg", + "table_caption": [ + "EuRoC (ATE [m])" + ], + "table_footnote": [], + "table_body": "
MethodMH01MH02MH03MH04MH05V101V102V103V201V202V203Avg
ORB-SLAM30.0310.0320.0411.2560.0660.0351.5281.4510.2010.1901.4690.573
RL ORB-SLAM30.0290.0280.0390.1520.1060.0341.4721.4410.2030.0701.6540.475
TUM-RGBD (ATE [m])
Method360deskdesk2floorplantroomrpyteddyxyzAvg
ORB-SLAM3-0.0310.7460.0130.1541.048-0.1080.0110.302
RL ORB-SLAM3-0.0310.6040.0130.0380.717-0.0910.0110.215
", + "bbox": [ + 220, + 715, + 781, + 794 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Table 8: ORB-SLAM3. We report the ATE [m] of our RL agent applied on ORBSLAM3.", + "bbox": [ + 215, + 795, + 782, + 821 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "Reinforcement Learning Meets Visual Odometry", + "bbox": [ + 408, + 114, + 730, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 784, + 126 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "size of 20 for challenging conditions, leading to more keypoints being tracked. Another behavior that can be observed is the quick increase of the grid size for one or two frames. A possible explanation for this short increase and subsequent decrease in the grid size is that the agent tries to filter out less robust keypoints. Runtime The runtime achieved by DSO and SVO given decisions from the heuristic rules and our RL agent are reported in Tab. 7. The results are obtained by averaging the runtime required for processing all sequences of EuRoC on one NVIDIA A100. Our RL agent leads to a faster average SVO runtime, which can be related to the lower rate of keyframe selection, see Fig. 4. In comparison, RL DSO does not lead to a high runtime reduction. For the performance of the agents with different penalty terms, we refer to the supplementary. Finally, the architecture of our agent is very lightweight, with a total of 296,584 parameters. SLAM Method In addition to VO methods, we also tested our method with the full SLAM pipeline ORB-SLAM3. Differently from the VO methods, which run in a sliding-window fashion (old measurements are marginalized), ORBSLAM3 uses a local bundle adjustment that has access to the entire history of keyframes. Although the local bundle adjustment could compensate for uninformative keyframes, our agent can still provide a more suitable keyframe selection strategy than the heuristics, as it can be observed in Tab. 8 by the increased accuracy on the EuRoC and TUM-RGBD datasets. To run ORB-SLAM3 with our framework, significant adjustments with respect to the synchronization between the tracking and bundle adjustment threads were required, amongst others, the removal of the loop closure algorithm, which explains the different performance compared to results reported in [3].", + "bbox": [ + 212, + 146, + 785, + 507 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Ablation To evaluate the benefits of our Variable Encoder processing the keypoints, we report in Tab. 5 the performance of an agent network without the Variable Encoder, i.e., RL SVO (w/o keypoints). As can be observed, the agent without access to the keypoints information is not capable of successfully finishing all the sequences of the EuRoC dataset. Furthermore, we also report the performance of the agent trained without privileged critic referred to as SVO RL (w/o privileged c.), leading to a larger mean ATE. For the ablations on TUM-RGBD, we refer to the supplementary.", + "bbox": [ + 212, + 508, + 787, + 630 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 655, + 359, + 672 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "By approaching the Visual Odometry (VO) problem from a sequential decision-making perspective, we employ a neural network trained within a Reinforcement Learning (RL) framework to dynamically adapt the VO method at each timestep. Instead of hand-tuned heuristics and fixed hyperparameters, the dynamic agent predicts optimal actions for the VO method to increase its accuracy and robustness. The RL agent is trained on a large-scale synthetic dataset and tested on common real-world VO benchmarks showing improved robustness and accuracy. We believe this paradigm shift advances VO technology and opens avenues for integrating RL into visual inertial odometry and simultaneous localization and mapping approaches.", + "bbox": [ + 212, + 688, + 787, + 840 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "Messikommer et al.", + "bbox": [ + 271, + 114, + 401, + 127 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgments", + "text_level": 1, + "bbox": [ + 217, + 143, + 392, + 162 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "This work was supported by the European Research Council (ERC) under grant agreement No. 864042 (AGILEFLIGHT).", + "bbox": [ + 215, + 178, + 784, + 209 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 234, + 321, + 250 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Burri, M., Nikolic, J., Gohl, P., Schneider, T., Rehder, J., Omari, S., Achtelik, M.W., Siegwart, R.: The euroc micro aerial vehicle datasets. The International Journal of Robotics Research (2016). https://doi.org/10.1177/0278364915620033, http://ijr.sagepub.com/content/early/2016/01/21/0278364915620033abstract", + "2. Cadena, C., Carlone, L., Carrillo, H., Latif, Y., Scaramuzza, D., Neira, J., Reid, I., Leonard, J.J.: Past, present, and future of simultaneous localization and mapping: Toward the robust-perception age. IEEE Transactions on robotics 32(6), 1309-1332 (2016)", + "3. Campos, C., Elvira, R., Gomez, J.J., Montiel, J.M.M., Tardos, J.D.: ORB-SLAM3: An accurate open-source library for visual, visual-inertial and multi-map SLAM. IEEE Transactions on Robotics 37(6), 1874-1890 (2021)", + "4. Dong, X., Shen, J., Wang, W., Shao, L., Ling, H., Porikli, F.: Dynamical hyperparameter optimization via deep reinforcement learning in tracking. IEEE transactions on pattern analysis and machine intelligence 43(5), 1515-1529 (2019)", + "5. Engel, J., Koltun, V., Cremers, D.: Direct Sparse Odometry. IEEE Trans. Pattern Anal. Mach. Intell. 40(3), 611-625 (Mar 2018). https://doi.org/10.1109/TPAMI.2017.2658577", + "6. Forster, C., Zhang, Z., Gassner, M., Werlberger, M., Scaramuzza, D.: SVO: Semidirect visual odometry for monocular and multicamera systems. IEEE Trans. Robot. 33(2), 249-265 (2017). https://doi.org/10.1109/TR0.2016.2623335", + "7. Geiger, A., Lenz, P., Urtasun, R.: Are we ready for autonomous driving? the kitti vision benchmark suite. In: Conference on Computer Vision and Pattern Recognition (CVPR) (2012)", + "8. Jaegle, A., Borgeaud, S., Alayrac, J.B., Doersch, C., Ionescu, C., Ding, D., Koppula, S., Zoran, D., Brock, A., Shelhamer, E., Henaff, O., Botvinick, M.M., Zisserman, A., Vinyals, O., Carreira, J.: Perceiver io: A general architecture for structured inputs & outputs. Int. Conf. Learn. Representations (ICLR) (2022)", + "9. Jaegle, A., Gimeno, F., Brock, A., Zisserman, A., Vinyals, O., Carreira, J.: Perceiver: General perception with iterative attention. Proc. Int. Conf. Mach. Learning (ICML) (2021)", + "0. Jomaa, H.S., Grabocka, J., Schmidt-Thieme, L.: Hyp-rl: Hyperparameter optimization by reinforcement learning. arXiv preprint arXiv:1906.11527 (2019)", + "1. Mur-Artal, R., Montiel, J.M.M., Tardos, J.D.: Orb-slam: a versatile and accurate monocular slam system. IEEE transactions on robotics 31(5), 1147-1163 (2015)", + "12. Raffin, A., Hill, A., Gleave, A., Kanervisto, A., Ernestus, M., Dormann, N.: Stable-baselines3: Reliable reinforcement learning implementations. Journal of Machine Learning Research 22(268), 1-8 (2021), http://jmlr.org/papers/v22/20-1364.html", + "13. Scaramuzza, D., Fraundorfer, F.: Visual odometry [tutorial]. IEEE robotics & automation magazine 18(4), 80-92 (2011)" + ], + "bbox": [ + 218, + 267, + 785, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "Reinforcement Learning Meets Visual Odometry", + "bbox": [ + 408, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "14. Schulman, J., Wolski, F., Dhariwal, P., Radford, A., Klimov, O.: Proximal policy optimization algorithms. arXiv preprint arXiv:1707.06347 (2017)", + "15. Sheng, L., Xu, D., Ouyang, W., Wang, X.: Unsupervised collaborative learning of keyframe detection and visual odometry towards monocular deep slam. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 4302-4311 (2019)", + "16. Sturm, J., Engelhard, N., Endres, F., Burgard, W., Cremers, D.: A benchmark for the evaluation of rgb-d slam systems. In: Proc. of the International Conference on Intelligent Robot Systems (IROS) (Oct 2012)", + "17. Teed, Z., Deng, J.: Raft: Recurrent all-pairs field transforms for optical flow. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part II 16. pp. 402-419. Springer (2020)", + "18. Teed, Z., Deng, J.: Droid-slam: Deep visual slam for monocular, stereo, and rgb-d cameras. Advances in neural information processing systems 34, 16558-16569 (2021)", + "19. Teed, Z., Lipson, L., Deng, J.: Deep patch visual odometry. arXiv preprint arXiv:2208.04726 (2022)", + "20. Umeyama, S.: Least-squares estimation of transformation parameters between two point patterns. IEEE Trans. Pattern Anal. Mach. Intell. 13(4) (1991)", + "21. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L.u., Polosukhin, I.: Attention is all you need. In: Guyon, I., Luxburg, U.V., Bengio, S., Wallach, H., Fergus, R., Vishwanathan, S., Garnett, R. (eds.) Advances in Neural Information Processing Systems. vol. 30. Curran Associates, Inc. (2017), https://proceedings.neurips.cc/paper/2017/file/3f5ee243547dee91fbd053c1c4a845aa-Paper.pdf", + "22. Vogiatzis, G., Hernández, C.: Video-based, real-time multi-view stereo. Image and Vision Computing 29(7), 434-441 (2011)", + "23. Wang, S., Clark, R., Wen, H., Trigoni, N.: Deepvo: Towards end-to-end visual odometry with deep recurrent convolutional neural networks. In: IEEE Int. Conf. Robot. Autom. (ICRA). pp. 2043-2050. IEEE (2017)", + "24. Wang, W., Hu, Y., Scherer, S.: Tartanvo: A generalizable learning-based vo. In: Conference on Robot Learning. pp. 1761-1772. PMLR (2021)", + "25. Wang, W., Zhu, D., Wang, X., Hu, Y., Qiu, Y., Wang, C., Hu, Y., Kapoor, A., Scherer, S.: Tartanair: A dataset to push the limits of visual slam. 2020 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) (2020)", + "26. Ye, W., Lan, X., Chen, S., Ming, Y., Yu, X., Bao, H., Cui, Z., Zhang, G.: Pvo: Panoptic visual odometry. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9579-9589 (2023)", + "27. Zhang, Z., Scaramuzza, D.: A tutorial on quantitative trajectory evaluation for visual (-inertial) odometry. In: IEEE/RSJ Int. Conf. Intell. Robot. Syst. (IROS). pp. 7244-7251. IEEE (2018)" + ], + "bbox": [ + 215, + 146, + 785, + 715 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "Messikommer et al.", + "bbox": [ + 271, + 114, + 401, + 126 + ], + "page_idx": 15 + } +] \ No newline at end of file diff --git a/2024/Reinforcement Learning Meets Visual Odometry/4903ec8e-7c0d-4a2f-9524-76e74784ea76_model.json b/2024/Reinforcement Learning Meets Visual Odometry/4903ec8e-7c0d-4a2f-9524-76e74784ea76_model.json new file mode 100644 index 0000000000000000000000000000000000000000..c3ba2e5d063d0261c05a97c580ff7cea9ca68244 --- /dev/null +++ b/2024/Reinforcement Learning Meets Visual Odometry/4903ec8e-7c0d-4a2f-9524-76e74784ea76_model.json @@ -0,0 +1,1882 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.224, + 0.142, + 0.78, + 0.166 + ], + "angle": 0, + "content": "Reinforcement Learning Meets Visual Odometry" + }, + { + "type": "text", + "bbox": [ + 0.336, + 0.19, + 0.668, + 0.222 + ], + "angle": 0, + "content": "Nico Messikommer*, Giovanni Cioffi*, Mathias Gehrig, and Davide Scaramuzza" + }, + { + "type": "text", + "bbox": [ + 0.338, + 0.233, + 0.668, + 0.263 + ], + "angle": 0, + "content": "Dept. of Informatics, University of Zurich {nmessi, cioffi, mgehrig, sdavide}@ifi.uzh.ch" + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.29, + 0.744, + 0.555 + ], + "angle": 0, + "content": "Abstract. Visual Odometry (VO) is essential to downstream mobile robotics and augmented/virtual reality tasks. Despite recent advances, existing VO methods still rely on heuristic design choices that require several weeks of hyperparameter tuning by human experts, hindering generalizability and robustness. We address these challenges by reframing VO as a sequential decision-making task and applying Reinforcement Learning (RL) to adapt the VO process dynamically. Our approach introduces a neural network, operating as an agent within the VO pipeline, to make decisions such as keyframe and grid-size selection based on real-time conditions. Our method minimizes reliance on heuristic choices using a reward function based on pose error, runtime, and other metrics to guide the system. Our RL framework treats the VO system and the image sequence as an environment, with the agent receiving observations from keypoints, map statistics, and prior poses. Experimental results using classical VO methods and public benchmarks demonstrate improvements in accuracy and robustness, validating the generalizability of our RL-enhanced VO approach to different scenarios. We believe this paradigm shift advances VO technology by eliminating the need for time-intensive parameter tuning of heuristics." + }, + { + "type": "text", + "bbox": [ + 0.263, + 0.567, + 0.642, + 0.584 + ], + "angle": 0, + "content": "Keywords: Visual Odometry \\(\\cdot\\) Reinforcement Learning" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.603, + 0.786, + 0.634 + ], + "angle": 0, + "content": "Multimedia Material The code is available at https://github.com/uzh-rpg/ml_vo and video at https://youtu.be/pt6yPTdQd6M" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.654, + 0.377, + 0.67 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.682, + 0.788, + 0.819 + ], + "angle": 0, + "content": "The task of estimating the camera pose from a sequence of images, referred to as Visual Odometry (VO), finds broad applications in fields such as robotics and augmented/virtual reality [2, 13]. Despite considerable research and industry efforts to enhance the robustness of VO algorithms, state-of-the-art approaches grapple with generalizability [2, 25]. This limitation stems from the reliance on hyperparameters dictated by heuristic decision rules, which are highly scenario-dependent, varying with factors such as motion and lighting. Consequently, identifying the optimal hyperparameters often requires the intensive effort of domain experts, consuming several weeks." + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.826, + 0.359, + 0.841 + ], + "angle": 0, + "content": "* equal contribution" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.403, + 0.127 + ], + "angle": 0, + "content": "Messikommer et al." + }, + { + "type": "image", + "bbox": [ + 0.306, + 0.148, + 0.699, + 0.331 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.341, + 0.788, + 0.398 + ], + "angle": 0, + "content": "Fig. 1: Our Framework. We propose to employ a learned agent to adaptively guide a VO method using real-time observations for enhanced robustness and accuracy. By considering the problem as a sequential decision process, we use RL to train the agent primarily based on the position error computed within a sliding window (dashed lines)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.43, + 0.789, + 0.595 + ], + "angle": 0, + "content": "Existing VO methods can be positioned on a spectrum ranging from classical geometry-based approaches to end-to-end methods trained on extensive datasets. End-to-end approaches [23, 24, 26] aim to increase robustness by leveraging a high amount of image sequences with ground truth poses. However, they tend to struggle to adapt to visual scenarios and motion patterns beyond their training data distribution. On the other side of the spectrum are classical methods [3,5,6], which leverage geometry to map image points to and from the 3D world to compute the camera pose. Hybrid approaches [18, 19], positioned between these extremes, replace certain parts of the VO pipeline, usually the ones processing raw image pixels, with learned components while retaining well-defined geometric components to increase robustness." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.597, + 0.789, + 0.658 + ], + "angle": 0, + "content": "Despite these advancements, both hybrid and classical approaches still rely on heuristic design choices, rendering the optimal VO settings scene- and motion-dependent. Furthermore, these heuristics are defined by a fixed set of hyperparameters found through extensive offline tuning." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.66, + 0.79, + 0.841 + ], + "angle": 0, + "content": "For example, one of the most impactful decisions to be taken by most of the VO systems during operation is the selection of keyframes. An adequate keyframe selection strategy should add informative keyframes to the current map to maximize the map coverage and, consequently, the accuracy of the registration of the pose of the incoming frame. Considering the limited map size due to runtime and memory usage constraints, the naive strategy of selecting each frame as a keyframe is oftentimes not the optimal one. In the current state-of-the-art VO systems [5, 6], keyframe selection strategies are based on heuristic if-else conditions, which rely on a fixed set of hyperparameters. These hyperparameters are generally tuned offline by expert users, which exploit the prior knowledge of the scene and the expected camera motion. For example, a strategy prone to often selecting a keyframe is preferred if the camera undergoes fast motion. Other" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.409, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Reinforcement Learning Meets Visual Odometry" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.193 + ], + "angle": 0, + "content": "hyperparameters such as feature extraction thresholds, brightness compensation, and weighting factors are also important, however, their impact on the overall performance depends on the specific VO system." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.197, + 0.788, + 0.379 + ], + "angle": 0, + "content": "To reduce the dependence on heuristic-based design, we approach the Visual Odometry problem from a sequential decision-making perspective, where current actions influence future states. Given the temporal dependencies inherent in this task, coupled with the absence of ground truth for optimal actions, we propose to train a neural network within a Reinforcement Learning (RL) framework to adapt a VO method at each timestep of the sequence. Operating as a dynamic agent, this neural network guides decisions within the VO pipeline, such as keyframe selection, or, in theory, any hyperparameter, based on real-time observations representing the current state of the VO system (for example, the number of points currently tracked and the distance from the past keyframe). Thereby minimizing the reliance on heuristic design choices while increasing the generalizability across diverse scenes and motions." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.383, + 0.788, + 0.459 + ], + "angle": 0, + "content": "To enforce an accurate and robust pipeline, the reward function is constructed from the pose error between predicted and ground truth poses but also includes other non-differentiable metrics such as runtime and size of the map. Importantly, our proposed framework leaves the underlying VO method unaltered while introducing a learned agent to enhance its robustness." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.463, + 0.788, + 0.629 + ], + "angle": 0, + "content": "In our RL approach, we consider the underlying image sequence and the VO system as the environment. From the environment, the agent receives observations in the form of tracked keypoints, map statistics, previously selected keyframes, and previously estimated poses. The agent itself consists of a multi-head attention layer to downproject a variable number of keypoints, which are then further processed along the fixed-sized inputs in a two-layer MLP. For training, we employ the on-policy RL algorithm Proximal Policy Optimization (PPO) [14] with a privileged critic network to stabilize the training. Finally, the reward includes the pose error between the predicted and ground truth pose computed in a sliding window approach while also penalizing decisions that lead to increased runtime." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.634, + 0.788, + 0.709 + ], + "angle": 0, + "content": "Our experiments with the state-of-the-art VO methods [5, 6] on the VO benchmarks [1, 7, 16] show that our learned agent improves accuracy up to \\(19\\%\\), see RL DSO in Tab. 1, and increases robustness, see RL SVO in Tab. 1, where it tracks all the camera trajectories while the w/o RL baseline gets lost. Our contributions can be summarized as follows." + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.745, + 0.784, + 0.775 + ], + "angle": 0, + "content": "1. We propose to approach the Visual Odometry (VO) problem from a sequential decision perspective, where current actions influence future states." + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.78, + 0.785, + 0.84 + ], + "angle": 0, + "content": "2. We introduce an RL training framework designed to train an adaptive agent to replace heuristic design choices in VO methods. Our experimental results demonstrate that our approach not only boosts VO robustness and accuracy but also highlights the potential to eliminate labor-intensive fine-tuning." + }, + { + "type": "list", + "bbox": [ + 0.223, + 0.745, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.232, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.403, + 0.128 + ], + "angle": 0, + "content": "Messikommer et al." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.145, + 0.389, + 0.161 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.176, + 0.784, + 0.206 + ], + "angle": 0, + "content": "Three paradigms exist to approach Visual Odometry: classical geometric-based, end-to-end, and hybrid approaches." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.207, + 0.788, + 0.522 + ], + "angle": 0, + "content": "Classical approaches have been extensively studied in computer vision and robotics in the past 30 years [2, 13]. They achieve commercial-level accuracy and robustness. Indeed, classical VO algorithms are present in products like smartphones, VR-AR devices, and mobile robots. Classical VO algorithms are composed of two main modules: tracking and mapping. The tracking module uses camera images to estimate the motion. Three main approaches exist in the literature: direct methods [5], feature-based methods [11], and semi-direct methods [6]. Direct methods, such as DSO [5], work directly on the raw pixel intensities. These methods commonly extract image patches and estimate the camera trajectory by tracking the motion of such patches through consecutive images. The tracking is achieved by minimizing a photometric error defined on the raw pixel intensities. On the contrary, feature-based methods, such as ORBSLAM [11], extract points of interest, commonly known as visual features or keypoints, from the raw image pixels. The camera trajectory is estimated by tracking these points through consecutive images. Feature-based methods rely on scene texture to detect and track distinctive points, while direct methods use raw image pixels and can achieve higher reliability in low-texture environments. Semi-direct methods, such as SVO [6], combine keypoints and patches of raw pixel intensities to estimate the camera motion. SVO uses keypoints to find regions of interest in the current camera view and raw pixel intensities to track the patches centered at the detected keypoints." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.524, + 0.788, + 0.644 + ], + "angle": 0, + "content": "Following the deep learning revolution seen in computer vision in the past years, learning-based end-to-end VO systems [23, 24, 26] have been proposed. These methods usually rely on convolutional layers to process the input images and subsequently, recurrent and/or feedforward layers to predict the camera 6-DoF pose. They are trained with ground truth pose supervision. End-to-end VO systems show better performance than classical approaches in scenarios covered in the training data. However, they lack robustness when deployed on environments (scene and motion) outside the training data distribution." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.645, + 0.788, + 0.795 + ], + "angle": 0, + "content": "Hybrid methods [18, 19] combine the best of two paradigms: the geometric representation of the 3D world (classical methods), and the ability to process highly dimensional inputs (end-to-end methods). One of the state-of-the-art hybrid approaches is DROID-SLAM [18]. DROID-SLAM is a SLAM system that focuses on dense scene reconstruction, as well as, camera pose estimation. It uses a RAFT-inspired network architecture [17] followed by a differentiable dense bundle adjustment (BA) layer to iteratively update the camera poses and depth. The BA layer includes geometry constraints that improve the consistency of consecutive camera pose estimates. Using a similar differential BA layer, the hybrid method DPVO [19] tracks image patches over time using a recurrent network." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.796, + 0.788, + 0.84 + ], + "angle": 0, + "content": "The performance of VO algorithms highly depends on heuristics and hyperparameters. Usually, hyperparameter tuning is done manually by expert users. Due to the high dimensional space and the sequential nature of the VO prob-" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.409, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Reinforcement Learning Meets Visual Odometry" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.783, + 0.207 + ], + "angle": 0, + "content": "lem, automatic optimization-based hyperparameters tuning methods, such as Bayesian Optimization are difficult to use. An attempt to automatically select keyframes has been made in [15], where a deep network is used to decide if a new keyframe needs to be selected, as well as, to predict the camera pose." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.208, + 0.784, + 0.298 + ], + "angle": 0, + "content": "In this paper, we model VO as a sequential decision-making task and apply Reinforcement Learning. Related to the problem of hyperparameter tuning with RL, Hyp-RL [10], poses the search for the best hyperparameters for training neural networks as a sequential decision process and uses RL to select the next hyperparameter set to be evaluated. In [4], RL is applied to dynamically choose the best set of hyperparameters of a deep network that performs object tracking." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.32, + 0.331, + 0.336 + ], + "angle": 0, + "content": "3 Method" + }, + { + "type": "image", + "bbox": [ + 0.234, + 0.375, + 0.768, + 0.519 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.528, + 0.788, + 0.613 + ], + "angle": 0, + "content": "Fig. 2: Agent Overview. Our agent takes as input the map statistics computed by the VO method at timestep \\( t_{i-1} \\) and a variable number of keypoints using a multi-head attention layer. Given these inputs, a two-layer MLP computes a multi-discrete probability distribution over the binary keyframe action and the grid size action. These actions are then used inside the VO method to process the current frame at timestep \\( t_i \\), which will lead to the observation for the next timestep \\( t_{i+1} \\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.644, + 0.788, + 0.794 + ], + "angle": 0, + "content": "We consider monocular Visual Odometry (VO), whose primary objective is to estimate the camera pose for each image within an image sequence. Despite extensive research and widespread application, the robustness and generalizability of VO methods in different scenes remain an ongoing challenge. Both traditional geometry-based approaches and hybrid methods incorporating learned components often rely on heuristic decision rules for critical tasks such as keyframe selection and keypoint filtering. These heuristic rules significantly affect the performance of the VO method and are closely linked to the deployed scene. Consequently, a substantial amount of fine-tuning is required to adapt these rules effectively to each deployed scenario." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.796, + 0.788, + 0.84 + ], + "angle": 0, + "content": "To increase the robustness and generalizability of VO pipelines, we formulate the VO task as a sequential decision problem. Within this sequential framework, a deep agent is trained through RL. This agent is designed to make decisions" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.403, + 0.128 + ], + "angle": 0, + "content": "Messikommer et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.221 + ], + "angle": 0, + "content": "at each timestep for the VO method, taking into account the current state of the system, including map statistics, keypoints information and previous pose estimations. By training an adaptive agent, we can effectively reduce the reliance on heuristics and increase the generalizability of the underlying VO method across diverse scenes." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.224, + 0.788, + 0.316 + ], + "angle": 0, + "content": "The structure of this section is the following. In Sec. 3.1, we present our problem formulation. In Sec. 3.2, we introduce our proposed deep neural agent with its corresponding input observations and output actions. In Sec. 3.3, we formulate the reward function, while in Sec. 3.4, we present the RL training framework. Finally, in Sec. 3.5, we explain the specific implementation of our RL framework using SVO [6] and DSO [5] as the underlying VO methods." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.345, + 0.441, + 0.36 + ], + "angle": 0, + "content": "3.1 Problem Formulation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.378, + 0.788, + 0.5 + ], + "angle": 0, + "content": "We formulate the VO task as a sequential decision problem in which an agent interacts with its environment, see Fig. 1. Specifically, we consider the image sequence and the underlying VO pipeline to be fixed, both constituting the environment. The agent is a neural network that, at each timestep, selects actions for the VO pipeline, i.e., selecting keyframes and adapting the keypoint distribution by setting the grid size constraint (a grid is fitted to the image, and maximum one keypoint per grid cell is selected). Based on these actions, the VO method processes the next image of the sequence and computes the corresponding pose." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.501, + 0.788, + 0.698 + ], + "angle": 0, + "content": "With this definition of the agent and environment, we formulate the VO task as a Markov decision process (MDP), denoted by the tuple \\( M = (S, A, p, r, \\gamma) \\). A state \\( s \\in S \\) encapsulates the current image in the image sequence and the state of the VO pipeline, which includes properties such as feature tracks, previously estimated poses, the triangulate map, the optimization state, etc. To transition between VO states, actions \\( a \\in A \\) are executed in the VO method, while the environment supplies the next frame of the sequence. The state transition probability \\( p \\) defines the likelihood of moving from one state \\( s_i \\) to another \\( s_j \\) given actions \\( a_i \\). Finally, each state transition is associated with a reward \\( r(s, a) \\), dependent on the ensuing state and the chosen actions. The rewards of the future are discounted by \\( \\gamma \\). In the context of VO, the reward function evaluates the accuracy of the pose estimation and can also incorporate non-differentiable criteria such as runtime and map size, among others." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.699, + 0.788, + 0.79 + ], + "angle": 0, + "content": "In our formulation, the agent is characterized by a control policy \\(\\pi : S \\times A \\to \\mathbb{R}\\), where \\(\\pi\\) specifies the actions \\(a \\in A\\) to be taken in each state \\(s \\in S\\). Since the control policy is non-deterministic, it induces a probabilistic trajectory of states and actions \\(\\tau = \\{s_0, a_0, s_1, a_1, \\dots, s_T, a_{T-1}\\}\\), abbreviated as \\(\\tau \\sim \\pi\\). Following the control policy \\(\\pi\\), the expected sum over the discounted rewards for each timestep can be expressed with its value function" + }, + { + "type": "equation", + "bbox": [ + 0.386, + 0.805, + 0.787, + 0.845 + ], + "angle": 0, + "content": "\\[\nV ^ {\\pi} (s _ {i}) = \\mathbb {E} _ {\\tau \\sim \\pi} \\left[ \\sum_ {t = i} ^ {\\infty} \\gamma^ {t} r (s _ {t}, a _ {t}) \\right]. \\tag {1}\n\\]" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.409, + 0.115, + 0.733, + 0.13 + ], + "angle": 0, + "content": "Reinforcement Learning Meets Visual Odometry" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.176 + ], + "angle": 0, + "content": "The overarching goal is to find a policy \\(\\pi\\) that maximizes the expected sum of discounted rewards." + }, + { + "type": "equation", + "bbox": [ + 0.428, + 0.177, + 0.786, + 0.197 + ], + "angle": 0, + "content": "\\[\n\\pi^ {*} = \\arg \\max _ {\\pi} V ^ {\\pi} \\left(s _ {0}\\right) \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.201, + 0.784, + 0.245 + ], + "angle": 0, + "content": "Given the reward formulation, the optimal policy provides decisions to the VO pipeline, aiming to minimize the expected position error while accounting for additional reward terms like runtime." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.265, + 0.423, + 0.28 + ], + "angle": 0, + "content": "3.2 Deep Neural Agent" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.286, + 0.785, + 0.481 + ], + "angle": 0, + "content": "The central element of our proposed framework is the learned agent embodying the policy \\(\\pi\\), which selects optimal settings for the VO method based on the current keypoints and map statistics, i.e., relative poses to the newest and oldest keyframe. Given the inherent complexity of the underlying VO method, supplying the agent network with a comprehensive state encompassing all stored map elements, optimization states, etc., is neither feasible nor advantageous. Instead, the learned agent receives a selected subset of observations \\(o \\in O\\) approximating the VO state within the policy \\(\\pi: O \\times A \\to \\mathbb{R}\\). In our system, the observations are derived from the local map and include relative poses to previous keyframes, along with the information about the currently tracked keypoints. Notably, the number of keypoints is variable and may vary depending on the characteristics of the image sequence. Therefore, the network architecture of the agent should be capable of processing inputs of variable sizes." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.483, + 0.785, + 0.633 + ], + "angle": 0, + "content": "To handle inputs of varying sizes, we leverage a Perceiver [8,9] architecture, illustrated in Fig. 2. Specifically, our approach involves a Variable Encoder that incorporates a multi-head attention layer [21] for processing the image positions, and the estimated depth of each tracked keypoint \\( k \\in \\mathbb{R}^{N \\times 3} \\). To project the information contained in the variable number of keypoints \\( N \\) to a fixed dimension, we use \\( M \\) number of learned tokens \\( z \\in \\mathbb{R}^{M \\times D} \\) with dimension of \\( D \\) as query \\( Q \\) values. These query tokens are attended to by the projected keypoint information serving as both key \\( K \\) and values \\( V \\). Finally, the fixed output of the Variable Encoder is subsequently processed along with the statistics of the local map, using a two-layer MLP with ReLu non-linearities." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.634, + 0.785, + 0.77 + ], + "angle": 0, + "content": "The output of the agent network is a discrete probability distribution over a set of actions expressed in the form of multiple independent categorical distributions. In our implementation, the agent predicts at each timestep a binary distribution over the decision to take the current frame as a keyframe. Additionally, in the case of SVO, the agent also predicts a separate and independent categorical distribution over a set of values \\(\\{20, 25, \\dots, 40\\}\\) for the keypoint grid size. This grid size ensures a uniform distribution of keypoints on the image plane by enforcing that only one keypoint can reside within a given grid cell. Since DSO does not extract keypoints, there is no grid size selection necessary." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.789, + 0.327, + 0.802 + ], + "angle": 0, + "content": "3.3 Reward" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.784, + 0.84 + ], + "angle": 0, + "content": "Another critical component within the RL framework is the design of the reward function. The feedback of the reward at each timestep guides the deep agent" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.404, + 0.128 + ], + "angle": 0, + "content": "Messikommer et al." + }, + { + "type": "image", + "bbox": [ + 0.306, + 0.147, + 0.7, + 0.24 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.254, + 0.788, + 0.31 + ], + "angle": 0, + "content": "Fig. 3: Position Error. To closely relate the pose prediction accuracy to the current action, we employ a sliding window of five timesteps to align the ground truth and estimated trajectory using a scale factor \\( s \\), a translation vector \\( t \\), and a rotation \\( R \\). The error at the current timestep \\( t_i \\) is then used as the negative position reward." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.345, + 0.789, + 0.587 + ], + "angle": 0, + "content": "toward a desired behavioral pattern. In the case of VO, naturally, the error between the estimated and the ground truth pose is the primary focus of the reward. However, since monocular VO methods cannot inherently estimate the absolute scale of the trajectory, it becomes necessary to scale the predicted poses before computing the error against the ground truth trajectory. Furthermore, another crucial aspect of ensuring stable feedback for the RL algorithm is to closely link the reward function of a timestep \\( t_i \\) to the action taken at that specific timestep, thus mitigating the value assignment problem. Consequently, we adopt a sliding window approach to compute the error between the estimated and ground truth poses. Specifically, we employ a window of five timestamps \\( \\{t_{i-4}, t_{i-3}, \\dots, t_i\\} \\) to compute the error at the current timestep \\( t_i \\), see Fig. 3. The first three poses in the window are used to align the estimated trajectory to the ground truth by minimizing the distances between the positions with the Umeyama method [20]. This alignment method accommodates missing scale information and differing coordinate origins, aligning both trajectories through an Euclidean transformation involving rotation, translation, and scaling." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.589, + 0.789, + 0.695 + ], + "angle": 0, + "content": "By designing the position reward function, we can influence the policy to weigh accuracy higher at the cost of lower robustness or the other way around. This trade-off also shows the benefit of our dynamic agent over manual tuning since the decisions can be made in an online fashion, adapting to the current conditions in the image sequence. To handle this trade-off, we clip the positional error \\( e_{\\mathrm{tran}} \\) around zero, see Eq. 3, to positively enforce accurate states while giving a penalty for states having a large positional error." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.697, + 0.789, + 0.803 + ], + "angle": 0, + "content": "Given that our framework leverages RL, we have the opportunity to integrate non-differentiable performance feedback into the reward system. Thus, we can enforce a faster runtime by penalizing the insertion of keyframes, i.e., directly penalizing the keyframe action \\(a_{\\mathrm{keyframe}}\\). Using the keyframe action instead of directly penalizing the runtime mitigates the dependence on the usage of the hardware during training. Our final reward function is formulated in Eq. 3, where we use 0.01 for \\(\\lambda_{1}\\) and set \\(\\lambda_{2}\\) to \\(5\\times 10^{-3}\\)." + }, + { + "type": "equation", + "bbox": [ + 0.354, + 0.821, + 0.787, + 0.837 + ], + "angle": 0, + "content": "\\[\nr = \\lambda_ {1} \\max (- 1, 0. 2 - e _ {\\mathrm {t r a n}}) - \\lambda_ {2} a _ {\\text {k e y f r a m e}} \\tag {3}\n\\]" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.409, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Reinforcement Learning Meets Visual Odometry" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.147, + 0.557, + 0.162 + ], + "angle": 0, + "content": "3.4 Reinforcement Learning Framework" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.177, + 0.788, + 0.373 + ], + "angle": 0, + "content": "We use the on-policy algorithm Proximal Policy Optimization (PPO) [14] implemented in Stable Baselines3 [12] for training our agent. As commonly done, the RL training is conducted in two alternating stages: the rollout and the policy-update stage. During the rollout stage, the RL agent interacts with the environment, i.e., the underlying VO method and image sequences from TartanAir [25], via its actions to collect experiences comprising the actions, observations, and rewards at each timestep. While the experiences are collected, the network weights of the agent are not updated. In the second stage, the collected experiences of the rollout stage are used to update both the agent and the critic. The task of the critic is to approximate the value function, which represents the expected return of the current policy at a given state. The critic is used to stabilize the training inside the policy update steps in accordance with the PPO algorithm, which is designed to limit large changes in the policy." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.374, + 0.788, + 0.48 + ], + "angle": 0, + "content": "Since the performance of the VO pipelines depends not only on the decisions of the agent but also on the current image sequence, we introduce a privileged critic to stabilize the training. This privileged critic has access to both the current and future ground truth poses, enabling a better assessment of whether tracking performance was influenced by the decisions of the agent or the difficulty of the underlying image sequence. Importantly, this privileged critic with access to ground truth poses is only used during training." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.481, + 0.788, + 0.662 + ], + "angle": 0, + "content": "Another crucial aspect is the handling of the different stages of the VO system, which include the normal mode of camera tracking, the mode for the relocalization in the current map, and the initialization mode. Since the VO system only provides information about its full state in the tracking mode, we declare the current state a valid state if the previous state was in the tracking mode. Notably, the agent can exclusively make decisions in valid stages. To account for the valid states in the policy-update phase, we use a masked rollout buffer, which computes the return using all collected states, but samples only valid states for updating the policy network. As states in the initialization stages can not be influenced by the agent, we directly use the ground truth poses to triangulate tracked keypoints. The initialization with the ground truth poses also has the benefit of a significantly faster runtime compared to the original initialization." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.663, + 0.788, + 0.723 + ], + "angle": 0, + "content": "Since RL is sample inefficient, it is crucial to have a fast runtime of the environment interactions. For our RL framework, the runtime of each of the main components in the rollout phase using 100 SVO instances in parallel is reported in Tab. 6." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.75, + 0.357, + 0.765 + ], + "angle": 0, + "content": "3.5 VO System" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.842 + ], + "angle": 0, + "content": "In general, our proposed agent network can be combined with any VO system that relies on decision components. In this work, we use SVO [6] and DSO [5], which feature a very small runtime. Additionally, we also test on the slower SLAM pipeline ORB-SLAM3 [3]. From a practical viewpoint, the small runtime" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.403, + 0.128 + ], + "angle": 0, + "content": "Messikommer et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.177 + ], + "angle": 0, + "content": "and parallelization accelerate training and partially compensate for the inherent low sample efficiency of RL." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.178, + 0.787, + 0.344 + ], + "angle": 0, + "content": "SVO is a semi-direct visual odometry composed of two main threads: motion estimation and mapping thread. The first step of the motion estimation thread is image alignment, where the pose of the new frame is estimated by aligning pixel intensity patches from the previous keyframe. The next step is feature alignment, where the positions of the keypoints on the current image are refined while the camera poses are fixed. The last step is a sparse bundle adjustment, where both the camera poses and the locations of the 3D points included in the current map are optimized. The motion thread assumes that the depth of the mapped 3D points is known. It is the objective of the mapping thread to estimate the depth. The mapping thread estimates the depth of each tracked pixel from multiple observations by means of a recursive Bayesian filter [22]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.344, + 0.788, + 0.493 + ], + "angle": 0, + "content": "DSO is a direct visual odometry also composed of two threads: the front-end thread and the mapping thread. The front-end thread performs the initial frame tracking where the pose of the current frame is tracked with respect to the previous keyframe by aligning pixel intensities patches. The front-end is also in charge of selecting the active candidate points, performing outlier rejection and occlusion detection, and selecting keyframes. The mapping thread optimizes a non-linear cost function based on the photometric error. The camera poses, intrinsics, and inverse depth value of the active points in the current map are jointly optimized. This optimization can be seen as the photometric equivalent of sparse bundle adjustment." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.495, + 0.787, + 0.527 + ], + "angle": 0, + "content": "We use SVO to run ablation studies on the contributions of the main components of the RL system (runtime, inputs, actions, and reward tuning)." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.55, + 0.377, + 0.568 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.583, + 0.788, + 0.718 + ], + "angle": 0, + "content": "Dataset For training our agent, we leverage the synthetic sequences from TartanAir [25], a comprehensive dataset offering, among other measurements, ground truth camera poses. The dataset features a wide range of simulated scenes with challenging conditions such as low light and weather effects. The total number of different training sequences is 337, amounting to 279987 images or timesteps for the training of the agent. Importantly, we consider the dataset in our RL framework to be fixed without the possibility of the agent influencing the next image. Notably, as our method is trained on abstractions rather than image pixels, the risk of domain shift is mitigated." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.719, + 0.787, + 0.81 + ], + "angle": 0, + "content": "For evaluation, we use three VO benchmarks: EuRoC [1], TUM-RGBD [16] and KITTI [7]. The EuRoC dataset consists of 11 sequences recorded via a drone, with ground truth pose measurements obtained from a motion capture system or a laser tracker. The TUM-RGBD dataset features 9 sequences recorded via a hand-held camera with ground truth poses provided by a motion capture system. For the results on the KITTI dataset, we refer to the supplementary." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.842 + ], + "angle": 0, + "content": "Evaluation Following standard practice in evaluating VO algorithms, we use the Absolute Translation Error (ATE [m]) [27] as the error metric for the EuRoC and" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.409, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Reinforcement Learning Meets Visual Odometry" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.207 + ], + "angle": 0, + "content": "TUM-RGBD datasets. Before computing the error metric, the estimated trajectories are aligned to the ground truth trajectory using the Umeyama method [20]. For the relative position error on EuRoC and TUM-RGBD, we refer to the supplementary." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.208, + 0.785, + 0.372 + ], + "angle": 0, + "content": "Since we want to guarantee a fair comparison between the methods with and without RL, we remove the effect of the initialization phase at the beginning of each sequence by providing the ground truth poses to the VO algorithm. Similar to the original initialization of the underlying VO, extracted keypoints are tracked from the first frame of the sequence and triangulated once a large enough disparity is detected. However, instead of estimating a relative pose, the ground truth poses are used for the triangulation. Following this procedure, all of the tested methods start with the same high-quality map, effectively eliminating the unwanted influence of the initialization. Naturally, the ground truth initialization is not used for the comparison against state-of-the-art VO methods. Instead, we take the average of the performance metrics for five runs." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.397, + 0.489, + 0.493 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.504, + 0.49, + 0.559 + ], + "angle": 0, + "content": "Fig. 4: Keyframe Selection. The number of selected keyframes by SVO at different translational and angular velocities for the EuRoC dataset." + }, + { + "type": "image", + "bbox": [ + 0.515, + 0.399, + 0.785, + 0.505 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.511, + 0.516, + 0.788, + 0.557 + ], + "angle": 0, + "content": "Fig.5: Top-Down View of Trajectories. Visualization of the predicted trajectory using SVO with and without RL." + }, + { + "type": "table", + "bbox": [ + 0.221, + 0.611, + 0.784, + 0.661 + ], + "angle": 0, + "content": "
MethodMH01MH02MH03MH04MH05V101V102V103V201V202V203Avg
SVO0.1830.4900.3813.2361.2930.3200.333-0.0840.5531.431-
RL SVO0.2360.1240.2412.1491.9820.2250.3980.3940.0690.3341.5060.696
DSO0.0870.0440.3540.1840.2601.2231.6241.1250.0590.0651.5160.594
RL DSO0.0370.0330.2410.1681.1920.2070.4991.4500.0430.1641.2860.483
" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.662, + 0.785, + 0.73 + ], + "angle": 0, + "content": "Table 1: EuRoC. We report the ATE [m] of our RL agents and the heuristic-based VO methods with parameters obtained by an extensive grid search. Our RL agents outperform the heuristic-based VO methods in terms of accuracy (lower average error) and robustness (RL SVO can complete the sequence V103 while the baseline SVO fails to track all the camera poses due to fast rotations)." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.79, + 0.325, + 0.803 + ], + "angle": 0, + "content": "4.1 Results" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.784, + 0.84 + ], + "angle": 0, + "content": "EuRoC We compare our RL agent with SVO and DSO against the heuristic-based VO with parameters obtained by tuning on this dataset using an extensive" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.403, + 0.128 + ], + "angle": 0, + "content": "Messikommer et al." + }, + { + "type": "table", + "bbox": [ + 0.241, + 0.145, + 0.767, + 0.198 + ], + "angle": 0, + "content": "
Method360deskdesk2floorplantroomrpyteddyxyzAvg
SVO0.1860.6810.898-0.3200.8050.0530.7690.0570.471
RL SVO0.1890.5560.755-0.2600.8040.0560.6970.0620.422
DSO-0.2370.893-0.3200.9310.057-0.053-
RL DSO0.1870.6820.888-0.625-0.0540.7560.105-
" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.199, + 0.787, + 0.268 + ], + "angle": 0, + "content": "Table 2: TUM-RGBD. We report the ATE [m] of our RL agents and the heuristic-base VO methods with parameters obtained by an extensive grid search. Our RL SVO outperforms the heuristic-based SVO methods in terms of accuracy (lower average error). Although this dataset is challenging for DSO, due to the use of a rolling shutter camera, RL DSO improves the robustness with two more sequences completed." + }, + { + "type": "table", + "bbox": [ + 0.223, + 0.296, + 0.78, + 0.357 + ], + "angle": 0, + "content": "
MethodMH01MH02MH03MH04MH05V101V102V103V201V202V203Avg
TartanVO [24]0.6390.3250.5501.1531.0210.4470.3890.6220.4330.7491.1520.680
DROID-VO [18]0.1630.1210.2420.3990.2700.1030.1650.1580.1020.1150.2040.186
DPVO [19]0.0870.0550.1580.1370.1140.0500.1400.0860.0570.0490.2110.105
RL SVO0.2891.5990.5522.4071.8260.2290.8130.4960.0850.9651.4340.972
RL DSO0.0320.0370.1660.1420.1100.1070.0891.3310.0410.1521.3130.320
" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.357, + 0.785, + 0.385 + ], + "angle": 0, + "content": "Table 3: EuRoC. We report the ATE [m] of our RL agents and the state-of-the-art VO methods. Our RL DSO achieves the best accuracy in 5 sequences." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.43, + 0.789, + 0.61 + ], + "angle": 0, + "content": "grid search. Qualitative estimation results of SVO are shown in the left plot in Fig. 5 from a top-down view of the trajectory estimated by RL SVO, the baseline VO, and the ground truth. Tab. 1 reports the ATE. In the case of SVO, the combination with our RL agent achieves the best ATE in most of the sequences. Remarkably, our RL SVO can estimate the entire camera trajectory in the sequence V103 while the baseline fails. This sequence is one of the most difficult in the EuRoC dataset due to fast camera rotations. The benefit of our RL agent is also evident for DSO, which achieves consistently a lower tracking error, leading to a \\(19\\%\\) lower mean tracking error. We report in Tab. 3 the comparison of our RL agent against the state-of-the-art VO methods, which are taken from [19]. The results reported for our methods are averaged over five runs. Our RL DSO achieves the best ATE in five sequences." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.612, + 0.789, + 0.794 + ], + "angle": 0, + "content": "TUM-RGBD Similar to EuRoC, we use a parameter grid search to find the best parameters for the baseline DSO and SVO methods. We report the ATE in Tab. 2, which shows that SVO RL achieves a lower ATE in most of the sequences and a lower overall mean error. Since the images of TUM-RGBD are not captured with a global shutter camera, DSO, relying on photometric alignment, struggles to finish all the sequences. Nevertheless, our RL agent improves the robustness in some challenging conditions with two more sequences successfully completed. For completeness, the top-down view of the trajectory estimated by RL SVO, the baseline, and the ground truth are visualized in the right plots in Fig. 5. We report in Tab. 4 the comparison of our RL agent against the state-of-the-art VO methods. The results reported for our methods are averaged over five runs. The results of the state-of-the-art methods are taken from [19]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.789, + 0.84 + ], + "angle": 0, + "content": "Keyframe Selection In Fig. 4, we give insights into the keyframe selection strategy employed by our RL agent. We visualize the histogram of states at which keyframes are selected and their corresponding rotational and translation" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.409, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Reinforcement Learning Meets Visual Odometry" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "table", + "bbox": [ + 0.241, + 0.144, + 0.766, + 0.197 + ], + "angle": 0, + "content": "
Method360deskdesk2floorplantroomryteddyxyzAvg
DROID-VO [18]0.1610.0280.0990.0330.0280.3270.0280.1690.0130.098
DPVO [19]0.1350.0380.0480.0400.0360.3940.0340.0640.0120.089
RL SVO0.1830.6480.790-0.3320.7860.0520.7410.071-
RL DSO-0.4380.765-0.479---0.057-
" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.198, + 0.784, + 0.224 + ], + "angle": 0, + "content": "Table 4: TUM-RGBD. We report the ATE [m] of our RL agents and the state-of-the-art VO methods." + }, + { + "type": "table", + "bbox": [ + 0.224, + 0.252, + 0.783, + 0.319 + ], + "angle": 0, + "content": "
MethodMH01MH02MH03MH04MH05V101V102V103V201V202V203Avg
SVO0.1830.4900.3813.2361.2930.3200.333-0.0840.5531.431-
RL SVO0.2360.1240.2412.1491.9820.2250.3980.3940.0690.3341.5060.696
RL SVO (w/o keypoints)0.2550.1600.4191.9790.5880.1730.6440.8490.1821.394--
RL SVO (w/o privileged c.)0.1300.0990.8343.2830.9580.1400.2860.4590.1160.6181.3730.754
RL SVO (w/o keyframe)0.1790.3510.3023.0721.0210.325--0.1790.2581.464-
RL SVO (w/o gridsize)0.1520.1830.2932.1541.5370.1980.2690.8080.1081.6831.4050.799
" + }, + { + "type": "table_caption", + "bbox": [ + 0.218, + 0.319, + 0.784, + 0.332 + ], + "angle": 0, + "content": "Table 5: EuRoC Ablations. Ablation study on inputs and action of the RL agent." + }, + { + "type": "table", + "bbox": [ + 0.232, + 0.36, + 0.475, + 0.414 + ], + "angle": 0, + "content": "
ComponentsTime [ms]
- 100 Parallel SVO Step73.9
- Load 100 Images18.6
- Network Forward Pass3.1
Complete Rollout Step104.1
" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.425, + 0.49, + 0.467 + ], + "angle": 0, + "content": "Table 6: Rollout Runtime. The runtime for the main components in the rollout phase averaged over 500 iterations." + }, + { + "type": "table", + "bbox": [ + 0.531, + 0.36, + 0.774, + 0.424 + ], + "angle": 0, + "content": "
MethodTime [ms]
SVO9.06
RL SVO7.40
DSO34.45
RL DSO32.29
Network inference1.75
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.425, + 0.785, + 0.466 + ], + "angle": 0, + "content": "Table 7: Inference Runtime. The runtime of SVO, DSO, and one network forward pass." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.511, + 0.785, + 0.616 + ], + "angle": 0, + "content": "velocity. The keyframe selection using the RL agent has a similar distribution around the motion patterns, while the heuristic rules select more keyframes. Following the motion distribution of the dataset, both methods trigger more keyframes in small motions. Despite the fewer keyframes, our RL agent still achieves a more robust performance. Without the possibility to set keyframes, namely, the agent only determines the feature grid size, robustness is drastically reduced, as it can be seen in Tab. 5 by the entry RL SVO (w/o keyframe)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.617, + 0.785, + 0.678 + ], + "angle": 0, + "content": "Grid size Selection The positive impact of dynamically selecting the grid size for SVO can be observed by removing the grid size action, which lowers the pose estimation accuracy, see RL SVO (w/o gridsize) in Tab. 5. The supplementary video shows that the RL agent learns to reduce the grid size to the minimal" + }, + { + "type": "table_caption", + "bbox": [ + 0.504, + 0.705, + 0.593, + 0.715 + ], + "angle": 0, + "content": "EuRoC (ATE [m])" + }, + { + "type": "table", + "bbox": [ + 0.222, + 0.716, + 0.782, + 0.795 + ], + "angle": 0, + "content": "
MethodMH01MH02MH03MH04MH05V101V102V103V201V202V203Avg
ORB-SLAM30.0310.0320.0411.2560.0660.0351.5281.4510.2010.1901.4690.573
RL ORB-SLAM30.0290.0280.0390.1520.1060.0341.4721.4410.2030.0701.6540.475
TUM-RGBD (ATE [m])
Method360deskdesk2floorplantroomrpyteddyxyzAvg
ORB-SLAM3-0.0310.7460.0130.1541.048-0.1080.0110.302
RL ORB-SLAM3-0.0310.6040.0130.0380.717-0.0910.0110.215
" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.796, + 0.784, + 0.822 + ], + "angle": 0, + "content": "Table 8: ORB-SLAM3. We report the ATE [m] of our RL agent applied on ORBSLAM3." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.403, + 0.128 + ], + "angle": 0, + "content": "Messikommer et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.787, + 0.508 + ], + "angle": 0, + "content": "size of 20 for challenging conditions, leading to more keypoints being tracked. Another behavior that can be observed is the quick increase of the grid size for one or two frames. A possible explanation for this short increase and subsequent decrease in the grid size is that the agent tries to filter out less robust keypoints. Runtime The runtime achieved by DSO and SVO given decisions from the heuristic rules and our RL agent are reported in Tab. 7. The results are obtained by averaging the runtime required for processing all sequences of EuRoC on one NVIDIA A100. Our RL agent leads to a faster average SVO runtime, which can be related to the lower rate of keyframe selection, see Fig. 4. In comparison, RL DSO does not lead to a high runtime reduction. For the performance of the agents with different penalty terms, we refer to the supplementary. Finally, the architecture of our agent is very lightweight, with a total of 296,584 parameters. SLAM Method In addition to VO methods, we also tested our method with the full SLAM pipeline ORB-SLAM3. Differently from the VO methods, which run in a sliding-window fashion (old measurements are marginalized), ORBSLAM3 uses a local bundle adjustment that has access to the entire history of keyframes. Although the local bundle adjustment could compensate for uninformative keyframes, our agent can still provide a more suitable keyframe selection strategy than the heuristics, as it can be observed in Tab. 8 by the increased accuracy on the EuRoC and TUM-RGBD datasets. To run ORB-SLAM3 with our framework, significant adjustments with respect to the synchronization between the tracking and bundle adjustment threads were required, amongst others, the removal of the loop closure algorithm, which explains the different performance compared to results reported in [3]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.51, + 0.788, + 0.631 + ], + "angle": 0, + "content": "Ablation To evaluate the benefits of our Variable Encoder processing the keypoints, we report in Tab. 5 the performance of an agent network without the Variable Encoder, i.e., RL SVO (w/o keypoints). As can be observed, the agent without access to the keypoints information is not capable of successfully finishing all the sequences of the EuRoC dataset. Furthermore, we also report the performance of the agent trained without privileged critic referred to as SVO RL (w/o privileged c.), leading to a larger mean ATE. For the ablations on TUM-RGBD, we refer to the supplementary." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.656, + 0.36, + 0.673 + ], + "angle": 0, + "content": "5 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.788, + 0.841 + ], + "angle": 0, + "content": "By approaching the Visual Odometry (VO) problem from a sequential decision-making perspective, we employ a neural network trained within a Reinforcement Learning (RL) framework to dynamically adapt the VO method at each timestep. Instead of hand-tuned heuristics and fixed hyperparameters, the dynamic agent predicts optimal actions for the VO method to increase its accuracy and robustness. The RL agent is trained on a large-scale synthetic dataset and tested on common real-world VO benchmarks showing improved robustness and accuracy. We believe this paradigm shift advances VO technology and opens avenues for integrating RL into visual inertial odometry and simultaneous localization and mapping approaches." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.409, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Reinforcement Learning Meets Visual Odometry" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.393, + 0.163 + ], + "angle": 0, + "content": "Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.179, + 0.785, + 0.21 + ], + "angle": 0, + "content": "This work was supported by the European Research Council (ERC) under grant agreement No. 864042 (AGILEFLIGHT)." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.235, + 0.323, + 0.25 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.268, + 0.786, + 0.336 + ], + "angle": 0, + "content": "1. Burri, M., Nikolic, J., Gohl, P., Schneider, T., Rehder, J., Omari, S., Achtelik, M.W., Siegwart, R.: The euroc micro aerial vehicle datasets. The International Journal of Robotics Research (2016). https://doi.org/10.1177/0278364915620033, http://ijr.sagepub.com/content/early/2016/01/21/0278364915620033abstract" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.338, + 0.787, + 0.393 + ], + "angle": 0, + "content": "2. Cadena, C., Carlone, L., Carrillo, H., Latif, Y., Scaramuzza, D., Neira, J., Reid, I., Leonard, J.J.: Past, present, and future of simultaneous localization and mapping: Toward the robust-perception age. IEEE Transactions on robotics 32(6), 1309-1332 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.395, + 0.786, + 0.435 + ], + "angle": 0, + "content": "3. Campos, C., Elvira, R., Gomez, J.J., Montiel, J.M.M., Tardos, J.D.: ORB-SLAM3: An accurate open-source library for visual, visual-inertial and multi-map SLAM. IEEE Transactions on Robotics 37(6), 1874-1890 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.437, + 0.786, + 0.478 + ], + "angle": 0, + "content": "4. Dong, X., Shen, J., Wang, W., Shao, L., Ling, H., Porikli, F.: Dynamical hyperparameter optimization via deep reinforcement learning in tracking. IEEE transactions on pattern analysis and machine intelligence 43(5), 1515-1529 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.479, + 0.786, + 0.518 + ], + "angle": 0, + "content": "5. Engel, J., Koltun, V., Cremers, D.: Direct Sparse Odometry. IEEE Trans. Pattern Anal. Mach. Intell. 40(3), 611-625 (Mar 2018). https://doi.org/10.1109/TPAMI.2017.2658577" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.52, + 0.786, + 0.56 + ], + "angle": 0, + "content": "6. Forster, C., Zhang, Z., Gassner, M., Werlberger, M., Scaramuzza, D.: SVO: Semidirect visual odometry for monocular and multicamera systems. IEEE Trans. Robot. 33(2), 249-265 (2017). https://doi.org/10.1109/TR0.2016.2623335" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.562, + 0.786, + 0.603 + ], + "angle": 0, + "content": "7. Geiger, A., Lenz, P., Urtasun, R.: Are we ready for autonomous driving? the kitti vision benchmark suite. In: Conference on Computer Vision and Pattern Recognition (CVPR) (2012)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.604, + 0.786, + 0.659 + ], + "angle": 0, + "content": "8. Jaegle, A., Borgeaud, S., Alayrac, J.B., Doersch, C., Ionescu, C., Ding, D., Koppula, S., Zoran, D., Brock, A., Shelhamer, E., Henaff, O., Botvinick, M.M., Zisserman, A., Vinyals, O., Carreira, J.: Perceiver io: A general architecture for structured inputs & outputs. Int. Conf. Learn. Representations (ICLR) (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.66, + 0.786, + 0.701 + ], + "angle": 0, + "content": "9. Jaegle, A., Gimeno, F., Brock, A., Zisserman, A., Vinyals, O., Carreira, J.: Perceiver: General perception with iterative attention. Proc. Int. Conf. Mach. Learning (ICML) (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.702, + 0.786, + 0.728 + ], + "angle": 0, + "content": "0. Jomaa, H.S., Grabocka, J., Schmidt-Thieme, L.: Hyp-rl: Hyperparameter optimization by reinforcement learning. arXiv preprint arXiv:1906.11527 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.729, + 0.786, + 0.757 + ], + "angle": 0, + "content": "1. Mur-Artal, R., Montiel, J.M.M., Tardos, J.D.: Orb-slam: a versatile and accurate monocular slam system. IEEE transactions on robotics 31(5), 1147-1163 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.758, + 0.786, + 0.811 + ], + "angle": 0, + "content": "12. Raffin, A., Hill, A., Gleave, A., Kanervisto, A., Ernestus, M., Dormann, N.: Stable-baselines3: Reliable reinforcement learning implementations. Journal of Machine Learning Research 22(268), 1-8 (2021), http://jmlr.org/papers/v22/20-1364.html" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.813, + 0.786, + 0.84 + ], + "angle": 0, + "content": "13. Scaramuzza, D., Fraundorfer, F.: Visual odometry [tutorial]. IEEE robotics & automation magazine 18(4), 80-92 (2011)" + }, + { + "type": "list", + "bbox": [ + 0.22, + 0.268, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.402, + 0.127 + ], + "angle": 0, + "content": "Messikommer et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.785, + 0.175 + ], + "angle": 0, + "content": "14. Schulman, J., Wolski, F., Dhariwal, P., Radford, A., Klimov, O.: Proximal policy optimization algorithms. arXiv preprint arXiv:1707.06347 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.176, + 0.787, + 0.23 + ], + "angle": 0, + "content": "15. Sheng, L., Xu, D., Ouyang, W., Wang, X.: Unsupervised collaborative learning of keyframe detection and visual odometry towards monocular deep slam. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 4302-4311 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.231, + 0.786, + 0.272 + ], + "angle": 0, + "content": "16. Sturm, J., Engelhard, N., Endres, F., Burgard, W., Cremers, D.: A benchmark for the evaluation of rgb-d slam systems. In: Proc. of the International Conference on Intelligent Robot Systems (IROS) (Oct 2012)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.273, + 0.786, + 0.314 + ], + "angle": 0, + "content": "17. Teed, Z., Deng, J.: Raft: Recurrent all-pairs field transforms for optical flow. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part II 16. pp. 402-419. Springer (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.315, + 0.786, + 0.355 + ], + "angle": 0, + "content": "18. Teed, Z., Deng, J.: Droid-slam: Deep visual slam for monocular, stereo, and rgb-d cameras. Advances in neural information processing systems 34, 16558-16569 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.356, + 0.786, + 0.384 + ], + "angle": 0, + "content": "19. Teed, Z., Lipson, L., Deng, J.: Deep patch visual odometry. arXiv preprint arXiv:2208.04726 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.384, + 0.786, + 0.411 + ], + "angle": 0, + "content": "20. Umeyama, S.: Least-squares estimation of transformation parameters between two point patterns. IEEE Trans. Pattern Anal. Mach. Intell. 13(4) (1991)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.411, + 0.786, + 0.493 + ], + "angle": 0, + "content": "21. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L.u., Polosukhin, I.: Attention is all you need. In: Guyon, I., Luxburg, U.V., Bengio, S., Wallach, H., Fergus, R., Vishwanathan, S., Garnett, R. (eds.) Advances in Neural Information Processing Systems. vol. 30. Curran Associates, Inc. (2017), https://proceedings.neurips.cc/paper/2017/file/3f5ee243547dee91fbd053c1c4a845aa-Paper.pdf" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.494, + 0.786, + 0.522 + ], + "angle": 0, + "content": "22. Vogiatzis, G., Hernández, C.: Video-based, real-time multi-view stereo. Image and Vision Computing 29(7), 434-441 (2011)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.522, + 0.786, + 0.563 + ], + "angle": 0, + "content": "23. Wang, S., Clark, R., Wen, H., Trigoni, N.: Deepvo: Towards end-to-end visual odometry with deep recurrent convolutional neural networks. In: IEEE Int. Conf. Robot. Autom. (ICRA). pp. 2043-2050. IEEE (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.564, + 0.786, + 0.591 + ], + "angle": 0, + "content": "24. Wang, W., Hu, Y., Scherer, S.: Tartanvo: A generalizable learning-based vo. In: Conference on Robot Learning. pp. 1761-1772. PMLR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.591, + 0.786, + 0.632 + ], + "angle": 0, + "content": "25. Wang, W., Zhu, D., Wang, X., Hu, Y., Qiu, Y., Wang, C., Hu, Y., Kapoor, A., Scherer, S.: Tartanair: A dataset to push the limits of visual slam. 2020 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.633, + 0.786, + 0.674 + ], + "angle": 0, + "content": "26. Ye, W., Lan, X., Chen, S., Ming, Y., Yu, X., Bao, H., Cui, Z., Zhang, G.: Pvo: Panoptic visual odometry. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9579-9589 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.675, + 0.786, + 0.716 + ], + "angle": 0, + "content": "27. Zhang, Z., Scaramuzza, D.: A tutorial on quantitative trajectory evaluation for visual (-inertial) odometry. In: IEEE/RSJ Int. Conf. Intell. Robot. Syst. (IROS). pp. 7244-7251. IEEE (2018)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.716 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/Reinforcement Learning Meets Visual Odometry/4903ec8e-7c0d-4a2f-9524-76e74784ea76_origin.pdf b/2024/Reinforcement Learning Meets Visual Odometry/4903ec8e-7c0d-4a2f-9524-76e74784ea76_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..56a1811d9fae279992e260112e45946650a482f2 --- /dev/null +++ b/2024/Reinforcement Learning Meets Visual Odometry/4903ec8e-7c0d-4a2f-9524-76e74784ea76_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7f1265c56863318b7b20845e8cfb4e9a9fa96c3cbf57f6b62c6f24624b7aa5b6 +size 987155 diff --git a/2024/Reinforcement Learning Meets Visual Odometry/full.md b/2024/Reinforcement Learning Meets Visual Odometry/full.md new file mode 100644 index 0000000000000000000000000000000000000000..2fd23b8328b0eb68c2a7e259df0e99308938e4d8 --- /dev/null +++ b/2024/Reinforcement Learning Meets Visual Odometry/full.md @@ -0,0 +1,242 @@ +# Reinforcement Learning Meets Visual Odometry + +Nico Messikommer*, Giovanni Cioffi*, Mathias Gehrig, and Davide Scaramuzza + +Dept. of Informatics, University of Zurich {nmessi, cioffi, mgehrig, sdavide}@ifi.uzh.ch + +Abstract. Visual Odometry (VO) is essential to downstream mobile robotics and augmented/virtual reality tasks. Despite recent advances, existing VO methods still rely on heuristic design choices that require several weeks of hyperparameter tuning by human experts, hindering generalizability and robustness. We address these challenges by reframing VO as a sequential decision-making task and applying Reinforcement Learning (RL) to adapt the VO process dynamically. Our approach introduces a neural network, operating as an agent within the VO pipeline, to make decisions such as keyframe and grid-size selection based on real-time conditions. Our method minimizes reliance on heuristic choices using a reward function based on pose error, runtime, and other metrics to guide the system. Our RL framework treats the VO system and the image sequence as an environment, with the agent receiving observations from keypoints, map statistics, and prior poses. Experimental results using classical VO methods and public benchmarks demonstrate improvements in accuracy and robustness, validating the generalizability of our RL-enhanced VO approach to different scenarios. We believe this paradigm shift advances VO technology by eliminating the need for time-intensive parameter tuning of heuristics. + +Keywords: Visual Odometry $\cdot$ Reinforcement Learning + +Multimedia Material The code is available at https://github.com/uzh-rpg/ml_vo and video at https://youtu.be/pt6yPTdQd6M + +# 1 Introduction + +The task of estimating the camera pose from a sequence of images, referred to as Visual Odometry (VO), finds broad applications in fields such as robotics and augmented/virtual reality [2, 13]. Despite considerable research and industry efforts to enhance the robustness of VO algorithms, state-of-the-art approaches grapple with generalizability [2, 25]. This limitation stems from the reliance on hyperparameters dictated by heuristic decision rules, which are highly scenario-dependent, varying with factors such as motion and lighting. Consequently, identifying the optimal hyperparameters often requires the intensive effort of domain experts, consuming several weeks. + +![](images/b3ab776b463802e4d61e55173ddf90999eba927eb4852e6a6bc687c5f8e870ee.jpg) +Fig. 1: Our Framework. We propose to employ a learned agent to adaptively guide a VO method using real-time observations for enhanced robustness and accuracy. By considering the problem as a sequential decision process, we use RL to train the agent primarily based on the position error computed within a sliding window (dashed lines). + +Existing VO methods can be positioned on a spectrum ranging from classical geometry-based approaches to end-to-end methods trained on extensive datasets. End-to-end approaches [23, 24, 26] aim to increase robustness by leveraging a high amount of image sequences with ground truth poses. However, they tend to struggle to adapt to visual scenarios and motion patterns beyond their training data distribution. On the other side of the spectrum are classical methods [3,5,6], which leverage geometry to map image points to and from the 3D world to compute the camera pose. Hybrid approaches [18, 19], positioned between these extremes, replace certain parts of the VO pipeline, usually the ones processing raw image pixels, with learned components while retaining well-defined geometric components to increase robustness. + +Despite these advancements, both hybrid and classical approaches still rely on heuristic design choices, rendering the optimal VO settings scene- and motion-dependent. Furthermore, these heuristics are defined by a fixed set of hyperparameters found through extensive offline tuning. + +For example, one of the most impactful decisions to be taken by most of the VO systems during operation is the selection of keyframes. An adequate keyframe selection strategy should add informative keyframes to the current map to maximize the map coverage and, consequently, the accuracy of the registration of the pose of the incoming frame. Considering the limited map size due to runtime and memory usage constraints, the naive strategy of selecting each frame as a keyframe is oftentimes not the optimal one. In the current state-of-the-art VO systems [5, 6], keyframe selection strategies are based on heuristic if-else conditions, which rely on a fixed set of hyperparameters. These hyperparameters are generally tuned offline by expert users, which exploit the prior knowledge of the scene and the expected camera motion. For example, a strategy prone to often selecting a keyframe is preferred if the camera undergoes fast motion. Other + +hyperparameters such as feature extraction thresholds, brightness compensation, and weighting factors are also important, however, their impact on the overall performance depends on the specific VO system. + +To reduce the dependence on heuristic-based design, we approach the Visual Odometry problem from a sequential decision-making perspective, where current actions influence future states. Given the temporal dependencies inherent in this task, coupled with the absence of ground truth for optimal actions, we propose to train a neural network within a Reinforcement Learning (RL) framework to adapt a VO method at each timestep of the sequence. Operating as a dynamic agent, this neural network guides decisions within the VO pipeline, such as keyframe selection, or, in theory, any hyperparameter, based on real-time observations representing the current state of the VO system (for example, the number of points currently tracked and the distance from the past keyframe). Thereby minimizing the reliance on heuristic design choices while increasing the generalizability across diverse scenes and motions. + +To enforce an accurate and robust pipeline, the reward function is constructed from the pose error between predicted and ground truth poses but also includes other non-differentiable metrics such as runtime and size of the map. Importantly, our proposed framework leaves the underlying VO method unaltered while introducing a learned agent to enhance its robustness. + +In our RL approach, we consider the underlying image sequence and the VO system as the environment. From the environment, the agent receives observations in the form of tracked keypoints, map statistics, previously selected keyframes, and previously estimated poses. The agent itself consists of a multi-head attention layer to downproject a variable number of keypoints, which are then further processed along the fixed-sized inputs in a two-layer MLP. For training, we employ the on-policy RL algorithm Proximal Policy Optimization (PPO) [14] with a privileged critic network to stabilize the training. Finally, the reward includes the pose error between the predicted and ground truth pose computed in a sliding window approach while also penalizing decisions that lead to increased runtime. + +Our experiments with the state-of-the-art VO methods [5, 6] on the VO benchmarks [1, 7, 16] show that our learned agent improves accuracy up to $19\%$ , see RL DSO in Tab. 1, and increases robustness, see RL SVO in Tab. 1, where it tracks all the camera trajectories while the w/o RL baseline gets lost. Our contributions can be summarized as follows. + +1. We propose to approach the Visual Odometry (VO) problem from a sequential decision perspective, where current actions influence future states. +2. We introduce an RL training framework designed to train an adaptive agent to replace heuristic design choices in VO methods. Our experimental results demonstrate that our approach not only boosts VO robustness and accuracy but also highlights the potential to eliminate labor-intensive fine-tuning. + +# 2 Related Work + +Three paradigms exist to approach Visual Odometry: classical geometric-based, end-to-end, and hybrid approaches. + +Classical approaches have been extensively studied in computer vision and robotics in the past 30 years [2, 13]. They achieve commercial-level accuracy and robustness. Indeed, classical VO algorithms are present in products like smartphones, VR-AR devices, and mobile robots. Classical VO algorithms are composed of two main modules: tracking and mapping. The tracking module uses camera images to estimate the motion. Three main approaches exist in the literature: direct methods [5], feature-based methods [11], and semi-direct methods [6]. Direct methods, such as DSO [5], work directly on the raw pixel intensities. These methods commonly extract image patches and estimate the camera trajectory by tracking the motion of such patches through consecutive images. The tracking is achieved by minimizing a photometric error defined on the raw pixel intensities. On the contrary, feature-based methods, such as ORBSLAM [11], extract points of interest, commonly known as visual features or keypoints, from the raw image pixels. The camera trajectory is estimated by tracking these points through consecutive images. Feature-based methods rely on scene texture to detect and track distinctive points, while direct methods use raw image pixels and can achieve higher reliability in low-texture environments. Semi-direct methods, such as SVO [6], combine keypoints and patches of raw pixel intensities to estimate the camera motion. SVO uses keypoints to find regions of interest in the current camera view and raw pixel intensities to track the patches centered at the detected keypoints. + +Following the deep learning revolution seen in computer vision in the past years, learning-based end-to-end VO systems [23, 24, 26] have been proposed. These methods usually rely on convolutional layers to process the input images and subsequently, recurrent and/or feedforward layers to predict the camera 6-DoF pose. They are trained with ground truth pose supervision. End-to-end VO systems show better performance than classical approaches in scenarios covered in the training data. However, they lack robustness when deployed on environments (scene and motion) outside the training data distribution. + +Hybrid methods [18, 19] combine the best of two paradigms: the geometric representation of the 3D world (classical methods), and the ability to process highly dimensional inputs (end-to-end methods). One of the state-of-the-art hybrid approaches is DROID-SLAM [18]. DROID-SLAM is a SLAM system that focuses on dense scene reconstruction, as well as, camera pose estimation. It uses a RAFT-inspired network architecture [17] followed by a differentiable dense bundle adjustment (BA) layer to iteratively update the camera poses and depth. The BA layer includes geometry constraints that improve the consistency of consecutive camera pose estimates. Using a similar differential BA layer, the hybrid method DPVO [19] tracks image patches over time using a recurrent network. + +The performance of VO algorithms highly depends on heuristics and hyperparameters. Usually, hyperparameter tuning is done manually by expert users. Due to the high dimensional space and the sequential nature of the VO prob- + +lem, automatic optimization-based hyperparameters tuning methods, such as Bayesian Optimization are difficult to use. An attempt to automatically select keyframes has been made in [15], where a deep network is used to decide if a new keyframe needs to be selected, as well as, to predict the camera pose. + +In this paper, we model VO as a sequential decision-making task and apply Reinforcement Learning. Related to the problem of hyperparameter tuning with RL, Hyp-RL [10], poses the search for the best hyperparameters for training neural networks as a sequential decision process and uses RL to select the next hyperparameter set to be evaluated. In [4], RL is applied to dynamically choose the best set of hyperparameters of a deep network that performs object tracking. + +# 3 Method + +![](images/ad8f2e88031e9c9e80aa038834d2ce20128f278c540796b86fc1eb236ca40c3a.jpg) +Fig. 2: Agent Overview. Our agent takes as input the map statistics computed by the VO method at timestep $t_{i-1}$ and a variable number of keypoints using a multi-head attention layer. Given these inputs, a two-layer MLP computes a multi-discrete probability distribution over the binary keyframe action and the grid size action. These actions are then used inside the VO method to process the current frame at timestep $t_i$ , which will lead to the observation for the next timestep $t_{i+1}$ . + +We consider monocular Visual Odometry (VO), whose primary objective is to estimate the camera pose for each image within an image sequence. Despite extensive research and widespread application, the robustness and generalizability of VO methods in different scenes remain an ongoing challenge. Both traditional geometry-based approaches and hybrid methods incorporating learned components often rely on heuristic decision rules for critical tasks such as keyframe selection and keypoint filtering. These heuristic rules significantly affect the performance of the VO method and are closely linked to the deployed scene. Consequently, a substantial amount of fine-tuning is required to adapt these rules effectively to each deployed scenario. + +To increase the robustness and generalizability of VO pipelines, we formulate the VO task as a sequential decision problem. Within this sequential framework, a deep agent is trained through RL. This agent is designed to make decisions + +at each timestep for the VO method, taking into account the current state of the system, including map statistics, keypoints information and previous pose estimations. By training an adaptive agent, we can effectively reduce the reliance on heuristics and increase the generalizability of the underlying VO method across diverse scenes. + +The structure of this section is the following. In Sec. 3.1, we present our problem formulation. In Sec. 3.2, we introduce our proposed deep neural agent with its corresponding input observations and output actions. In Sec. 3.3, we formulate the reward function, while in Sec. 3.4, we present the RL training framework. Finally, in Sec. 3.5, we explain the specific implementation of our RL framework using SVO [6] and DSO [5] as the underlying VO methods. + +# 3.1 Problem Formulation + +We formulate the VO task as a sequential decision problem in which an agent interacts with its environment, see Fig. 1. Specifically, we consider the image sequence and the underlying VO pipeline to be fixed, both constituting the environment. The agent is a neural network that, at each timestep, selects actions for the VO pipeline, i.e., selecting keyframes and adapting the keypoint distribution by setting the grid size constraint (a grid is fitted to the image, and maximum one keypoint per grid cell is selected). Based on these actions, the VO method processes the next image of the sequence and computes the corresponding pose. + +With this definition of the agent and environment, we formulate the VO task as a Markov decision process (MDP), denoted by the tuple $M = (S, A, p, r, \gamma)$ . A state $s \in S$ encapsulates the current image in the image sequence and the state of the VO pipeline, which includes properties such as feature tracks, previously estimated poses, the triangulate map, the optimization state, etc. To transition between VO states, actions $a \in A$ are executed in the VO method, while the environment supplies the next frame of the sequence. The state transition probability $p$ defines the likelihood of moving from one state $s_i$ to another $s_j$ given actions $a_i$ . Finally, each state transition is associated with a reward $r(s, a)$ , dependent on the ensuing state and the chosen actions. The rewards of the future are discounted by $\gamma$ . In the context of VO, the reward function evaluates the accuracy of the pose estimation and can also incorporate non-differentiable criteria such as runtime and map size, among others. + +In our formulation, the agent is characterized by a control policy $\pi : S \times A \to \mathbb{R}$ , where $\pi$ specifies the actions $a \in A$ to be taken in each state $s \in S$ . Since the control policy is non-deterministic, it induces a probabilistic trajectory of states and actions $\tau = \{s_0, a_0, s_1, a_1, \dots, s_T, a_{T-1}\}$ , abbreviated as $\tau \sim \pi$ . Following the control policy $\pi$ , the expected sum over the discounted rewards for each timestep can be expressed with its value function + +$$ +V ^ {\pi} (s _ {i}) = \mathbb {E} _ {\tau \sim \pi} \left[ \sum_ {t = i} ^ {\infty} \gamma^ {t} r (s _ {t}, a _ {t}) \right]. \tag {1} +$$ + +The overarching goal is to find a policy $\pi$ that maximizes the expected sum of discounted rewards. + +$$ +\pi^ {*} = \arg \max _ {\pi} V ^ {\pi} \left(s _ {0}\right) \tag {2} +$$ + +Given the reward formulation, the optimal policy provides decisions to the VO pipeline, aiming to minimize the expected position error while accounting for additional reward terms like runtime. + +# 3.2 Deep Neural Agent + +The central element of our proposed framework is the learned agent embodying the policy $\pi$ , which selects optimal settings for the VO method based on the current keypoints and map statistics, i.e., relative poses to the newest and oldest keyframe. Given the inherent complexity of the underlying VO method, supplying the agent network with a comprehensive state encompassing all stored map elements, optimization states, etc., is neither feasible nor advantageous. Instead, the learned agent receives a selected subset of observations $o \in O$ approximating the VO state within the policy $\pi: O \times A \to \mathbb{R}$ . In our system, the observations are derived from the local map and include relative poses to previous keyframes, along with the information about the currently tracked keypoints. Notably, the number of keypoints is variable and may vary depending on the characteristics of the image sequence. Therefore, the network architecture of the agent should be capable of processing inputs of variable sizes. + +To handle inputs of varying sizes, we leverage a Perceiver [8,9] architecture, illustrated in Fig. 2. Specifically, our approach involves a Variable Encoder that incorporates a multi-head attention layer [21] for processing the image positions, and the estimated depth of each tracked keypoint $k \in \mathbb{R}^{N \times 3}$ . To project the information contained in the variable number of keypoints $N$ to a fixed dimension, we use $M$ number of learned tokens $z \in \mathbb{R}^{M \times D}$ with dimension of $D$ as query $Q$ values. These query tokens are attended to by the projected keypoint information serving as both key $K$ and values $V$ . Finally, the fixed output of the Variable Encoder is subsequently processed along with the statistics of the local map, using a two-layer MLP with ReLu non-linearities. + +The output of the agent network is a discrete probability distribution over a set of actions expressed in the form of multiple independent categorical distributions. In our implementation, the agent predicts at each timestep a binary distribution over the decision to take the current frame as a keyframe. Additionally, in the case of SVO, the agent also predicts a separate and independent categorical distribution over a set of values $\{20, 25, \dots, 40\}$ for the keypoint grid size. This grid size ensures a uniform distribution of keypoints on the image plane by enforcing that only one keypoint can reside within a given grid cell. Since DSO does not extract keypoints, there is no grid size selection necessary. + +# 3.3 Reward + +Another critical component within the RL framework is the design of the reward function. The feedback of the reward at each timestep guides the deep agent + +![](images/482c301693015bf0c8054787fb1f741622030cf152b411d47e51d7ee866fd520.jpg) +Fig. 3: Position Error. To closely relate the pose prediction accuracy to the current action, we employ a sliding window of five timesteps to align the ground truth and estimated trajectory using a scale factor $s$ , a translation vector $t$ , and a rotation $R$ . The error at the current timestep $t_i$ is then used as the negative position reward. + +toward a desired behavioral pattern. In the case of VO, naturally, the error between the estimated and the ground truth pose is the primary focus of the reward. However, since monocular VO methods cannot inherently estimate the absolute scale of the trajectory, it becomes necessary to scale the predicted poses before computing the error against the ground truth trajectory. Furthermore, another crucial aspect of ensuring stable feedback for the RL algorithm is to closely link the reward function of a timestep $t_i$ to the action taken at that specific timestep, thus mitigating the value assignment problem. Consequently, we adopt a sliding window approach to compute the error between the estimated and ground truth poses. Specifically, we employ a window of five timestamps $\{t_{i-4}, t_{i-3}, \dots, t_i\}$ to compute the error at the current timestep $t_i$ , see Fig. 3. The first three poses in the window are used to align the estimated trajectory to the ground truth by minimizing the distances between the positions with the Umeyama method [20]. This alignment method accommodates missing scale information and differing coordinate origins, aligning both trajectories through an Euclidean transformation involving rotation, translation, and scaling. + +By designing the position reward function, we can influence the policy to weigh accuracy higher at the cost of lower robustness or the other way around. This trade-off also shows the benefit of our dynamic agent over manual tuning since the decisions can be made in an online fashion, adapting to the current conditions in the image sequence. To handle this trade-off, we clip the positional error $e_{\mathrm{tran}}$ around zero, see Eq. 3, to positively enforce accurate states while giving a penalty for states having a large positional error. + +Given that our framework leverages RL, we have the opportunity to integrate non-differentiable performance feedback into the reward system. Thus, we can enforce a faster runtime by penalizing the insertion of keyframes, i.e., directly penalizing the keyframe action $a_{\mathrm{keyframe}}$ . Using the keyframe action instead of directly penalizing the runtime mitigates the dependence on the usage of the hardware during training. Our final reward function is formulated in Eq. 3, where we use 0.01 for $\lambda_{1}$ and set $\lambda_{2}$ to $5\times 10^{-3}$ . + +$$ +r = \lambda_ {1} \max (- 1, 0. 2 - e _ {\mathrm {t r a n}}) - \lambda_ {2} a _ {\text {k e y f r a m e}} \tag {3} +$$ + +# 3.4 Reinforcement Learning Framework + +We use the on-policy algorithm Proximal Policy Optimization (PPO) [14] implemented in Stable Baselines3 [12] for training our agent. As commonly done, the RL training is conducted in two alternating stages: the rollout and the policy-update stage. During the rollout stage, the RL agent interacts with the environment, i.e., the underlying VO method and image sequences from TartanAir [25], via its actions to collect experiences comprising the actions, observations, and rewards at each timestep. While the experiences are collected, the network weights of the agent are not updated. In the second stage, the collected experiences of the rollout stage are used to update both the agent and the critic. The task of the critic is to approximate the value function, which represents the expected return of the current policy at a given state. The critic is used to stabilize the training inside the policy update steps in accordance with the PPO algorithm, which is designed to limit large changes in the policy. + +Since the performance of the VO pipelines depends not only on the decisions of the agent but also on the current image sequence, we introduce a privileged critic to stabilize the training. This privileged critic has access to both the current and future ground truth poses, enabling a better assessment of whether tracking performance was influenced by the decisions of the agent or the difficulty of the underlying image sequence. Importantly, this privileged critic with access to ground truth poses is only used during training. + +Another crucial aspect is the handling of the different stages of the VO system, which include the normal mode of camera tracking, the mode for the relocalization in the current map, and the initialization mode. Since the VO system only provides information about its full state in the tracking mode, we declare the current state a valid state if the previous state was in the tracking mode. Notably, the agent can exclusively make decisions in valid stages. To account for the valid states in the policy-update phase, we use a masked rollout buffer, which computes the return using all collected states, but samples only valid states for updating the policy network. As states in the initialization stages can not be influenced by the agent, we directly use the ground truth poses to triangulate tracked keypoints. The initialization with the ground truth poses also has the benefit of a significantly faster runtime compared to the original initialization. + +Since RL is sample inefficient, it is crucial to have a fast runtime of the environment interactions. For our RL framework, the runtime of each of the main components in the rollout phase using 100 SVO instances in parallel is reported in Tab. 6. + +# 3.5 VO System + +In general, our proposed agent network can be combined with any VO system that relies on decision components. In this work, we use SVO [6] and DSO [5], which feature a very small runtime. Additionally, we also test on the slower SLAM pipeline ORB-SLAM3 [3]. From a practical viewpoint, the small runtime + +and parallelization accelerate training and partially compensate for the inherent low sample efficiency of RL. + +SVO is a semi-direct visual odometry composed of two main threads: motion estimation and mapping thread. The first step of the motion estimation thread is image alignment, where the pose of the new frame is estimated by aligning pixel intensity patches from the previous keyframe. The next step is feature alignment, where the positions of the keypoints on the current image are refined while the camera poses are fixed. The last step is a sparse bundle adjustment, where both the camera poses and the locations of the 3D points included in the current map are optimized. The motion thread assumes that the depth of the mapped 3D points is known. It is the objective of the mapping thread to estimate the depth. The mapping thread estimates the depth of each tracked pixel from multiple observations by means of a recursive Bayesian filter [22]. + +DSO is a direct visual odometry also composed of two threads: the front-end thread and the mapping thread. The front-end thread performs the initial frame tracking where the pose of the current frame is tracked with respect to the previous keyframe by aligning pixel intensities patches. The front-end is also in charge of selecting the active candidate points, performing outlier rejection and occlusion detection, and selecting keyframes. The mapping thread optimizes a non-linear cost function based on the photometric error. The camera poses, intrinsics, and inverse depth value of the active points in the current map are jointly optimized. This optimization can be seen as the photometric equivalent of sparse bundle adjustment. + +We use SVO to run ablation studies on the contributions of the main components of the RL system (runtime, inputs, actions, and reward tuning). + +# 4 Experiments + +Dataset For training our agent, we leverage the synthetic sequences from TartanAir [25], a comprehensive dataset offering, among other measurements, ground truth camera poses. The dataset features a wide range of simulated scenes with challenging conditions such as low light and weather effects. The total number of different training sequences is 337, amounting to 279987 images or timesteps for the training of the agent. Importantly, we consider the dataset in our RL framework to be fixed without the possibility of the agent influencing the next image. Notably, as our method is trained on abstractions rather than image pixels, the risk of domain shift is mitigated. + +For evaluation, we use three VO benchmarks: EuRoC [1], TUM-RGBD [16] and KITTI [7]. The EuRoC dataset consists of 11 sequences recorded via a drone, with ground truth pose measurements obtained from a motion capture system or a laser tracker. The TUM-RGBD dataset features 9 sequences recorded via a hand-held camera with ground truth poses provided by a motion capture system. For the results on the KITTI dataset, we refer to the supplementary. + +Evaluation Following standard practice in evaluating VO algorithms, we use the Absolute Translation Error (ATE [m]) [27] as the error metric for the EuRoC and + +TUM-RGBD datasets. Before computing the error metric, the estimated trajectories are aligned to the ground truth trajectory using the Umeyama method [20]. For the relative position error on EuRoC and TUM-RGBD, we refer to the supplementary. + +Since we want to guarantee a fair comparison between the methods with and without RL, we remove the effect of the initialization phase at the beginning of each sequence by providing the ground truth poses to the VO algorithm. Similar to the original initialization of the underlying VO, extracted keypoints are tracked from the first frame of the sequence and triangulated once a large enough disparity is detected. However, instead of estimating a relative pose, the ground truth poses are used for the triangulation. Following this procedure, all of the tested methods start with the same high-quality map, effectively eliminating the unwanted influence of the initialization. Naturally, the ground truth initialization is not used for the comparison against state-of-the-art VO methods. Instead, we take the average of the performance metrics for five runs. + +![](images/f2c9368a7608c352a368012f47362956f219586ba42f715df1e58e761a77042f.jpg) +Fig. 4: Keyframe Selection. The number of selected keyframes by SVO at different translational and angular velocities for the EuRoC dataset. + +![](images/77dcf3b29f3fdd213401076bc0920635636f2a0adb695fb07c7cb24c5b1abc25.jpg) +Fig.5: Top-Down View of Trajectories. Visualization of the predicted trajectory using SVO with and without RL. + +
MethodMH01MH02MH03MH04MH05V101V102V103V201V202V203Avg
SVO0.1830.4900.3813.2361.2930.3200.333-0.0840.5531.431-
RL SVO0.2360.1240.2412.1491.9820.2250.3980.3940.0690.3341.5060.696
DSO0.0870.0440.3540.1840.2601.2231.6241.1250.0590.0651.5160.594
RL DSO0.0370.0330.2410.1681.1920.2070.4991.4500.0430.1641.2860.483
+ +Table 1: EuRoC. We report the ATE [m] of our RL agents and the heuristic-based VO methods with parameters obtained by an extensive grid search. Our RL agents outperform the heuristic-based VO methods in terms of accuracy (lower average error) and robustness (RL SVO can complete the sequence V103 while the baseline SVO fails to track all the camera poses due to fast rotations). + +# 4.1 Results + +EuRoC We compare our RL agent with SVO and DSO against the heuristic-based VO with parameters obtained by tuning on this dataset using an extensive + +
Method360deskdesk2floorplantroomrpyteddyxyzAvg
SVO0.1860.6810.898-0.3200.8050.0530.7690.0570.471
RL SVO0.1890.5560.755-0.2600.8040.0560.6970.0620.422
DSO-0.2370.893-0.3200.9310.057-0.053-
RL DSO0.1870.6820.888-0.625-0.0540.7560.105-
+ +Table 2: TUM-RGBD. We report the ATE [m] of our RL agents and the heuristic-base VO methods with parameters obtained by an extensive grid search. Our RL SVO outperforms the heuristic-based SVO methods in terms of accuracy (lower average error). Although this dataset is challenging for DSO, due to the use of a rolling shutter camera, RL DSO improves the robustness with two more sequences completed. + +
MethodMH01MH02MH03MH04MH05V101V102V103V201V202V203Avg
TartanVO [24]0.6390.3250.5501.1531.0210.4470.3890.6220.4330.7491.1520.680
DROID-VO [18]0.1630.1210.2420.3990.2700.1030.1650.1580.1020.1150.2040.186
DPVO [19]0.0870.0550.1580.1370.1140.0500.1400.0860.0570.0490.2110.105
RL SVO0.2891.5990.5522.4071.8260.2290.8130.4960.0850.9651.4340.972
RL DSO0.0320.0370.1660.1420.1100.1070.0891.3310.0410.1521.3130.320
+ +Table 3: EuRoC. We report the ATE [m] of our RL agents and the state-of-the-art VO methods. Our RL DSO achieves the best accuracy in 5 sequences. + +grid search. Qualitative estimation results of SVO are shown in the left plot in Fig. 5 from a top-down view of the trajectory estimated by RL SVO, the baseline VO, and the ground truth. Tab. 1 reports the ATE. In the case of SVO, the combination with our RL agent achieves the best ATE in most of the sequences. Remarkably, our RL SVO can estimate the entire camera trajectory in the sequence V103 while the baseline fails. This sequence is one of the most difficult in the EuRoC dataset due to fast camera rotations. The benefit of our RL agent is also evident for DSO, which achieves consistently a lower tracking error, leading to a $19\%$ lower mean tracking error. We report in Tab. 3 the comparison of our RL agent against the state-of-the-art VO methods, which are taken from [19]. The results reported for our methods are averaged over five runs. Our RL DSO achieves the best ATE in five sequences. + +TUM-RGBD Similar to EuRoC, we use a parameter grid search to find the best parameters for the baseline DSO and SVO methods. We report the ATE in Tab. 2, which shows that SVO RL achieves a lower ATE in most of the sequences and a lower overall mean error. Since the images of TUM-RGBD are not captured with a global shutter camera, DSO, relying on photometric alignment, struggles to finish all the sequences. Nevertheless, our RL agent improves the robustness in some challenging conditions with two more sequences successfully completed. For completeness, the top-down view of the trajectory estimated by RL SVO, the baseline, and the ground truth are visualized in the right plots in Fig. 5. We report in Tab. 4 the comparison of our RL agent against the state-of-the-art VO methods. The results reported for our methods are averaged over five runs. The results of the state-of-the-art methods are taken from [19]. + +Keyframe Selection In Fig. 4, we give insights into the keyframe selection strategy employed by our RL agent. We visualize the histogram of states at which keyframes are selected and their corresponding rotational and translation + +
Method360deskdesk2floorplantroomryteddyxyzAvg
DROID-VO [18]0.1610.0280.0990.0330.0280.3270.0280.1690.0130.098
DPVO [19]0.1350.0380.0480.0400.0360.3940.0340.0640.0120.089
RL SVO0.1830.6480.790-0.3320.7860.0520.7410.071-
RL DSO-0.4380.765-0.479---0.057-
+ +Table 4: TUM-RGBD. We report the ATE [m] of our RL agents and the state-of-the-art VO methods. + +
MethodMH01MH02MH03MH04MH05V101V102V103V201V202V203Avg
SVO0.1830.4900.3813.2361.2930.3200.333-0.0840.5531.431-
RL SVO0.2360.1240.2412.1491.9820.2250.3980.3940.0690.3341.5060.696
RL SVO (w/o keypoints)0.2550.1600.4191.9790.5880.1730.6440.8490.1821.394--
RL SVO (w/o privileged c.)0.1300.0990.8343.2830.9580.1400.2860.4590.1160.6181.3730.754
RL SVO (w/o keyframe)0.1790.3510.3023.0721.0210.325--0.1790.2581.464-
RL SVO (w/o gridsize)0.1520.1830.2932.1541.5370.1980.2690.8080.1081.6831.4050.799
+ +Table 5: EuRoC Ablations. Ablation study on inputs and action of the RL agent. + +
ComponentsTime [ms]
- 100 Parallel SVO Step73.9
- Load 100 Images18.6
- Network Forward Pass3.1
Complete Rollout Step104.1
+ +Table 6: Rollout Runtime. The runtime for the main components in the rollout phase averaged over 500 iterations. + +
MethodTime [ms]
SVO9.06
RL SVO7.40
DSO34.45
RL DSO32.29
Network inference1.75
+ +Table 7: Inference Runtime. The runtime of SVO, DSO, and one network forward pass. + +velocity. The keyframe selection using the RL agent has a similar distribution around the motion patterns, while the heuristic rules select more keyframes. Following the motion distribution of the dataset, both methods trigger more keyframes in small motions. Despite the fewer keyframes, our RL agent still achieves a more robust performance. Without the possibility to set keyframes, namely, the agent only determines the feature grid size, robustness is drastically reduced, as it can be seen in Tab. 5 by the entry RL SVO (w/o keyframe). + +Grid size Selection The positive impact of dynamically selecting the grid size for SVO can be observed by removing the grid size action, which lowers the pose estimation accuracy, see RL SVO (w/o gridsize) in Tab. 5. The supplementary video shows that the RL agent learns to reduce the grid size to the minimal + +EuRoC (ATE [m]) + +
MethodMH01MH02MH03MH04MH05V101V102V103V201V202V203Avg
ORB-SLAM30.0310.0320.0411.2560.0660.0351.5281.4510.2010.1901.4690.573
RL ORB-SLAM30.0290.0280.0390.1520.1060.0341.4721.4410.2030.0701.6540.475
TUM-RGBD (ATE [m])
Method360deskdesk2floorplantroomrpyteddyxyzAvg
ORB-SLAM3-0.0310.7460.0130.1541.048-0.1080.0110.302
RL ORB-SLAM3-0.0310.6040.0130.0380.717-0.0910.0110.215
+ +Table 8: ORB-SLAM3. We report the ATE [m] of our RL agent applied on ORBSLAM3. + +size of 20 for challenging conditions, leading to more keypoints being tracked. Another behavior that can be observed is the quick increase of the grid size for one or two frames. A possible explanation for this short increase and subsequent decrease in the grid size is that the agent tries to filter out less robust keypoints. Runtime The runtime achieved by DSO and SVO given decisions from the heuristic rules and our RL agent are reported in Tab. 7. The results are obtained by averaging the runtime required for processing all sequences of EuRoC on one NVIDIA A100. Our RL agent leads to a faster average SVO runtime, which can be related to the lower rate of keyframe selection, see Fig. 4. In comparison, RL DSO does not lead to a high runtime reduction. For the performance of the agents with different penalty terms, we refer to the supplementary. Finally, the architecture of our agent is very lightweight, with a total of 296,584 parameters. SLAM Method In addition to VO methods, we also tested our method with the full SLAM pipeline ORB-SLAM3. Differently from the VO methods, which run in a sliding-window fashion (old measurements are marginalized), ORBSLAM3 uses a local bundle adjustment that has access to the entire history of keyframes. Although the local bundle adjustment could compensate for uninformative keyframes, our agent can still provide a more suitable keyframe selection strategy than the heuristics, as it can be observed in Tab. 8 by the increased accuracy on the EuRoC and TUM-RGBD datasets. To run ORB-SLAM3 with our framework, significant adjustments with respect to the synchronization between the tracking and bundle adjustment threads were required, amongst others, the removal of the loop closure algorithm, which explains the different performance compared to results reported in [3]. + +Ablation To evaluate the benefits of our Variable Encoder processing the keypoints, we report in Tab. 5 the performance of an agent network without the Variable Encoder, i.e., RL SVO (w/o keypoints). As can be observed, the agent without access to the keypoints information is not capable of successfully finishing all the sequences of the EuRoC dataset. Furthermore, we also report the performance of the agent trained without privileged critic referred to as SVO RL (w/o privileged c.), leading to a larger mean ATE. For the ablations on TUM-RGBD, we refer to the supplementary. + +# 5 Conclusion + +By approaching the Visual Odometry (VO) problem from a sequential decision-making perspective, we employ a neural network trained within a Reinforcement Learning (RL) framework to dynamically adapt the VO method at each timestep. Instead of hand-tuned heuristics and fixed hyperparameters, the dynamic agent predicts optimal actions for the VO method to increase its accuracy and robustness. The RL agent is trained on a large-scale synthetic dataset and tested on common real-world VO benchmarks showing improved robustness and accuracy. We believe this paradigm shift advances VO technology and opens avenues for integrating RL into visual inertial odometry and simultaneous localization and mapping approaches. + +# Acknowledgments + +This work was supported by the European Research Council (ERC) under grant agreement No. 864042 (AGILEFLIGHT). + +# References + +1. Burri, M., Nikolic, J., Gohl, P., Schneider, T., Rehder, J., Omari, S., Achtelik, M.W., Siegwart, R.: The euroc micro aerial vehicle datasets. The International Journal of Robotics Research (2016). https://doi.org/10.1177/0278364915620033, http://ijr.sagepub.com/content/early/2016/01/21/0278364915620033abstract +2. Cadena, C., Carlone, L., Carrillo, H., Latif, Y., Scaramuzza, D., Neira, J., Reid, I., Leonard, J.J.: Past, present, and future of simultaneous localization and mapping: Toward the robust-perception age. IEEE Transactions on robotics 32(6), 1309-1332 (2016) +3. Campos, C., Elvira, R., Gomez, J.J., Montiel, J.M.M., Tardos, J.D.: ORB-SLAM3: An accurate open-source library for visual, visual-inertial and multi-map SLAM. IEEE Transactions on Robotics 37(6), 1874-1890 (2021) +4. Dong, X., Shen, J., Wang, W., Shao, L., Ling, H., Porikli, F.: Dynamical hyperparameter optimization via deep reinforcement learning in tracking. IEEE transactions on pattern analysis and machine intelligence 43(5), 1515-1529 (2019) +5. Engel, J., Koltun, V., Cremers, D.: Direct Sparse Odometry. IEEE Trans. Pattern Anal. Mach. Intell. 40(3), 611-625 (Mar 2018). https://doi.org/10.1109/TPAMI.2017.2658577 +6. Forster, C., Zhang, Z., Gassner, M., Werlberger, M., Scaramuzza, D.: SVO: Semidirect visual odometry for monocular and multicamera systems. IEEE Trans. Robot. 33(2), 249-265 (2017). https://doi.org/10.1109/TR0.2016.2623335 +7. Geiger, A., Lenz, P., Urtasun, R.: Are we ready for autonomous driving? the kitti vision benchmark suite. In: Conference on Computer Vision and Pattern Recognition (CVPR) (2012) +8. Jaegle, A., Borgeaud, S., Alayrac, J.B., Doersch, C., Ionescu, C., Ding, D., Koppula, S., Zoran, D., Brock, A., Shelhamer, E., Henaff, O., Botvinick, M.M., Zisserman, A., Vinyals, O., Carreira, J.: Perceiver io: A general architecture for structured inputs & outputs. Int. Conf. Learn. Representations (ICLR) (2022) +9. Jaegle, A., Gimeno, F., Brock, A., Zisserman, A., Vinyals, O., Carreira, J.: Perceiver: General perception with iterative attention. Proc. Int. Conf. Mach. Learning (ICML) (2021) +0. Jomaa, H.S., Grabocka, J., Schmidt-Thieme, L.: Hyp-rl: Hyperparameter optimization by reinforcement learning. arXiv preprint arXiv:1906.11527 (2019) +1. Mur-Artal, R., Montiel, J.M.M., Tardos, J.D.: Orb-slam: a versatile and accurate monocular slam system. IEEE transactions on robotics 31(5), 1147-1163 (2015) +12. Raffin, A., Hill, A., Gleave, A., Kanervisto, A., Ernestus, M., Dormann, N.: Stable-baselines3: Reliable reinforcement learning implementations. Journal of Machine Learning Research 22(268), 1-8 (2021), http://jmlr.org/papers/v22/20-1364.html +13. Scaramuzza, D., Fraundorfer, F.: Visual odometry [tutorial]. IEEE robotics & automation magazine 18(4), 80-92 (2011) + +14. Schulman, J., Wolski, F., Dhariwal, P., Radford, A., Klimov, O.: Proximal policy optimization algorithms. arXiv preprint arXiv:1707.06347 (2017) +15. Sheng, L., Xu, D., Ouyang, W., Wang, X.: Unsupervised collaborative learning of keyframe detection and visual odometry towards monocular deep slam. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 4302-4311 (2019) +16. Sturm, J., Engelhard, N., Endres, F., Burgard, W., Cremers, D.: A benchmark for the evaluation of rgb-d slam systems. In: Proc. of the International Conference on Intelligent Robot Systems (IROS) (Oct 2012) +17. Teed, Z., Deng, J.: Raft: Recurrent all-pairs field transforms for optical flow. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part II 16. pp. 402-419. Springer (2020) +18. Teed, Z., Deng, J.: Droid-slam: Deep visual slam for monocular, stereo, and rgb-d cameras. Advances in neural information processing systems 34, 16558-16569 (2021) +19. Teed, Z., Lipson, L., Deng, J.: Deep patch visual odometry. arXiv preprint arXiv:2208.04726 (2022) +20. Umeyama, S.: Least-squares estimation of transformation parameters between two point patterns. IEEE Trans. Pattern Anal. Mach. Intell. 13(4) (1991) +21. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L.u., Polosukhin, I.: Attention is all you need. In: Guyon, I., Luxburg, U.V., Bengio, S., Wallach, H., Fergus, R., Vishwanathan, S., Garnett, R. (eds.) Advances in Neural Information Processing Systems. vol. 30. Curran Associates, Inc. (2017), https://proceedings.neurips.cc/paper/2017/file/3f5ee243547dee91fbd053c1c4a845aa-Paper.pdf +22. Vogiatzis, G., Hernández, C.: Video-based, real-time multi-view stereo. Image and Vision Computing 29(7), 434-441 (2011) +23. Wang, S., Clark, R., Wen, H., Trigoni, N.: Deepvo: Towards end-to-end visual odometry with deep recurrent convolutional neural networks. In: IEEE Int. Conf. Robot. Autom. (ICRA). pp. 2043-2050. IEEE (2017) +24. Wang, W., Hu, Y., Scherer, S.: Tartanvo: A generalizable learning-based vo. In: Conference on Robot Learning. pp. 1761-1772. PMLR (2021) +25. Wang, W., Zhu, D., Wang, X., Hu, Y., Qiu, Y., Wang, C., Hu, Y., Kapoor, A., Scherer, S.: Tartanair: A dataset to push the limits of visual slam. 2020 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) (2020) +26. Ye, W., Lan, X., Chen, S., Ming, Y., Yu, X., Bao, H., Cui, Z., Zhang, G.: Pvo: Panoptic visual odometry. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9579-9589 (2023) +27. Zhang, Z., Scaramuzza, D.: A tutorial on quantitative trajectory evaluation for visual (-inertial) odometry. In: IEEE/RSJ Int. Conf. Intell. Robot. Syst. (IROS). pp. 7244-7251. IEEE (2018) \ No newline at end of file diff --git a/2024/Reinforcement Learning Meets Visual Odometry/images.zip b/2024/Reinforcement Learning Meets Visual Odometry/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..23ac49995346c518984957623fc7c5e96d1a5a64 --- /dev/null +++ b/2024/Reinforcement Learning Meets Visual Odometry/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:832c3f845733a0cf0da64a5116d586a0c4aac5f7b5ccb7711bff9a945a4c1a95 +size 326675 diff --git a/2024/Reinforcement Learning Meets Visual Odometry/layout.json b/2024/Reinforcement Learning Meets Visual Odometry/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..8068bd0f6b724bb7856cfeebe791e139415b2886 --- /dev/null +++ b/2024/Reinforcement Learning Meets Visual Odometry/layout.json @@ -0,0 +1,6776 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 137, + 112, + 477, + 131 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 112, + 477, + 131 + ], + "spans": [ + { + "bbox": [ + 137, + 112, + 477, + 131 + ], + "type": "text", + "content": "Reinforcement Learning Meets Visual Odometry" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 205, + 150, + 408, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 205, + 150, + 408, + 175 + ], + "spans": [ + { + "bbox": [ + 205, + 150, + 408, + 175 + ], + "type": "text", + "content": "Nico Messikommer*, Giovanni Cioffi*, Mathias Gehrig, and Davide Scaramuzza" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 206, + 184, + 408, + 208 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 206, + 184, + 408, + 208 + ], + "spans": [ + { + "bbox": [ + 206, + 184, + 408, + 208 + ], + "type": "text", + "content": "Dept. of Informatics, University of Zurich {nmessi, cioffi, mgehrig, sdavide}@ifi.uzh.ch" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 160, + 229, + 455, + 439 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 229, + 455, + 439 + ], + "spans": [ + { + "bbox": [ + 160, + 229, + 455, + 439 + ], + "type": "text", + "content": "Abstract. Visual Odometry (VO) is essential to downstream mobile robotics and augmented/virtual reality tasks. Despite recent advances, existing VO methods still rely on heuristic design choices that require several weeks of hyperparameter tuning by human experts, hindering generalizability and robustness. We address these challenges by reframing VO as a sequential decision-making task and applying Reinforcement Learning (RL) to adapt the VO process dynamically. Our approach introduces a neural network, operating as an agent within the VO pipeline, to make decisions such as keyframe and grid-size selection based on real-time conditions. Our method minimizes reliance on heuristic choices using a reward function based on pose error, runtime, and other metrics to guide the system. Our RL framework treats the VO system and the image sequence as an environment, with the agent receiving observations from keypoints, map statistics, and prior poses. Experimental results using classical VO methods and public benchmarks demonstrate improvements in accuracy and robustness, validating the generalizability of our RL-enhanced VO approach to different scenarios. We believe this paradigm shift advances VO technology by eliminating the need for time-intensive parameter tuning of heuristics." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 160, + 449, + 392, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 449, + 392, + 462 + ], + "spans": [ + { + "bbox": [ + 160, + 449, + 392, + 462 + ], + "type": "text", + "content": "Keywords: Visual Odometry " + }, + { + "bbox": [ + 160, + 449, + 392, + 462 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 449, + 392, + 462 + ], + "type": "text", + "content": " Reinforcement Learning" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 477, + 481, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 477, + 481, + 502 + ], + "spans": [ + { + "bbox": [ + 130, + 477, + 481, + 502 + ], + "type": "text", + "content": "Multimedia Material The code is available at https://github.com/uzh-rpg/ml_vo and video at https://youtu.be/pt6yPTdQd6M" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 517, + 230, + 530 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 517, + 230, + 530 + ], + "spans": [ + { + "bbox": [ + 132, + 517, + 230, + 530 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 540, + 482, + 648 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 540, + 482, + 648 + ], + "spans": [ + { + "bbox": [ + 130, + 540, + 482, + 648 + ], + "type": "text", + "content": "The task of estimating the camera pose from a sequence of images, referred to as Visual Odometry (VO), finds broad applications in fields such as robotics and augmented/virtual reality [2, 13]. Despite considerable research and industry efforts to enhance the robustness of VO algorithms, state-of-the-art approaches grapple with generalizability [2, 25]. This limitation stems from the reliance on hyperparameters dictated by heuristic decision rules, which are highly scenario-dependent, varying with factors such as motion and lighting. Consequently, identifying the optimal hyperparameters often requires the intensive effort of domain experts, consuming several weeks." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 654, + 219, + 666 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 654, + 219, + 666 + ], + "spans": [ + { + "bbox": [ + 133, + 654, + 219, + 666 + ], + "type": "text", + "content": "* equal contribution" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 187, + 117, + 427, + 262 + ], + "blocks": [ + { + "bbox": [ + 187, + 117, + 427, + 262 + ], + "lines": [ + { + "bbox": [ + 187, + 117, + 427, + 262 + ], + "spans": [ + { + "bbox": [ + 187, + 117, + 427, + 262 + ], + "type": "image", + "image_path": "b3ab776b463802e4d61e55173ddf90999eba927eb4852e6a6bc687c5f8e870ee.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 270, + 482, + 315 + ], + "lines": [ + { + "bbox": [ + 130, + 270, + 482, + 315 + ], + "spans": [ + { + "bbox": [ + 130, + 270, + 482, + 315 + ], + "type": "text", + "content": "Fig. 1: Our Framework. We propose to employ a learned agent to adaptively guide a VO method using real-time observations for enhanced robustness and accuracy. By considering the problem as a sequential decision process, we use RL to train the agent primarily based on the position error computed within a sliding window (dashed lines)." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 340, + 482, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 340, + 482, + 471 + ], + "spans": [ + { + "bbox": [ + 130, + 340, + 482, + 471 + ], + "type": "text", + "content": "Existing VO methods can be positioned on a spectrum ranging from classical geometry-based approaches to end-to-end methods trained on extensive datasets. End-to-end approaches [23, 24, 26] aim to increase robustness by leveraging a high amount of image sequences with ground truth poses. However, they tend to struggle to adapt to visual scenarios and motion patterns beyond their training data distribution. On the other side of the spectrum are classical methods [3,5,6], which leverage geometry to map image points to and from the 3D world to compute the camera pose. Hybrid approaches [18, 19], positioned between these extremes, replace certain parts of the VO pipeline, usually the ones processing raw image pixels, with learned components while retaining well-defined geometric components to increase robustness." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 472, + 482, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 472, + 482, + 521 + ], + "spans": [ + { + "bbox": [ + 130, + 472, + 482, + 521 + ], + "type": "text", + "content": "Despite these advancements, both hybrid and classical approaches still rely on heuristic design choices, rendering the optimal VO settings scene- and motion-dependent. Furthermore, these heuristics are defined by a fixed set of hyperparameters found through extensive offline tuning." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 522, + 483, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 522, + 483, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 522, + 483, + 666 + ], + "type": "text", + "content": "For example, one of the most impactful decisions to be taken by most of the VO systems during operation is the selection of keyframes. An adequate keyframe selection strategy should add informative keyframes to the current map to maximize the map coverage and, consequently, the accuracy of the registration of the pose of the incoming frame. Considering the limited map size due to runtime and memory usage constraints, the naive strategy of selecting each frame as a keyframe is oftentimes not the optimal one. In the current state-of-the-art VO systems [5, 6], keyframe selection strategies are based on heuristic if-else conditions, which rely on a fixed set of hyperparameters. These hyperparameters are generally tuned offline by expert users, which exploit the prior knowledge of the scene and the expected camera motion. For example, a strategy prone to often selecting a keyframe is preferred if the camera undergoes fast motion. Other" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 246, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 246, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 246, + 100 + ], + "type": "text", + "content": "Messikommer et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "content": "hyperparameters such as feature extraction thresholds, brightness compensation, and weighting factors are also important, however, their impact on the overall performance depends on the specific VO system." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 156, + 482, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 156, + 482, + 300 + ], + "spans": [ + { + "bbox": [ + 130, + 156, + 482, + 300 + ], + "type": "text", + "content": "To reduce the dependence on heuristic-based design, we approach the Visual Odometry problem from a sequential decision-making perspective, where current actions influence future states. Given the temporal dependencies inherent in this task, coupled with the absence of ground truth for optimal actions, we propose to train a neural network within a Reinforcement Learning (RL) framework to adapt a VO method at each timestep of the sequence. Operating as a dynamic agent, this neural network guides decisions within the VO pipeline, such as keyframe selection, or, in theory, any hyperparameter, based on real-time observations representing the current state of the VO system (for example, the number of points currently tracked and the distance from the past keyframe). Thereby minimizing the reliance on heuristic design choices while increasing the generalizability across diverse scenes and motions." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 303, + 482, + 363 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 303, + 482, + 363 + ], + "spans": [ + { + "bbox": [ + 130, + 303, + 482, + 363 + ], + "type": "text", + "content": "To enforce an accurate and robust pipeline, the reward function is constructed from the pose error between predicted and ground truth poses but also includes other non-differentiable metrics such as runtime and size of the map. Importantly, our proposed framework leaves the underlying VO method unaltered while introducing a learned agent to enhance its robustness." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 366, + 482, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 366, + 482, + 498 + ], + "spans": [ + { + "bbox": [ + 130, + 366, + 482, + 498 + ], + "type": "text", + "content": "In our RL approach, we consider the underlying image sequence and the VO system as the environment. From the environment, the agent receives observations in the form of tracked keypoints, map statistics, previously selected keyframes, and previously estimated poses. The agent itself consists of a multi-head attention layer to downproject a variable number of keypoints, which are then further processed along the fixed-sized inputs in a two-layer MLP. For training, we employ the on-policy RL algorithm Proximal Policy Optimization (PPO) [14] with a privileged critic network to stabilize the training. Finally, the reward includes the pose error between the predicted and ground truth pose computed in a sliding window approach while also penalizing decisions that lead to increased runtime." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 502, + 482, + 561 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 502, + 482, + 561 + ], + "spans": [ + { + "bbox": [ + 130, + 502, + 482, + 561 + ], + "type": "text", + "content": "Our experiments with the state-of-the-art VO methods [5, 6] on the VO benchmarks [1, 7, 16] show that our learned agent improves accuracy up to " + }, + { + "bbox": [ + 130, + 502, + 482, + 561 + ], + "type": "inline_equation", + "content": "19\\%" + }, + { + "bbox": [ + 130, + 502, + 482, + 561 + ], + "type": "text", + "content": ", see RL DSO in Tab. 1, and increases robustness, see RL SVO in Tab. 1, where it tracks all the camera trajectories while the w/o RL baseline gets lost. Our contributions can be summarized as follows." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 136, + 590, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 136, + 590, + 479, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 590, + 479, + 613 + ], + "spans": [ + { + "bbox": [ + 136, + 590, + 479, + 613 + ], + "type": "text", + "content": "1. We propose to approach the Visual Odometry (VO) problem from a sequential decision perspective, where current actions influence future states." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 136, + 617, + 480, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 617, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 136, + 617, + 480, + 665 + ], + "type": "text", + "content": "2. We introduce an RL training framework designed to train an adaptive agent to replace heuristic design choices in VO methods. Our experimental results demonstrate that our approach not only boosts VO robustness and accuracy but also highlights the potential to eliminate labor-intensive fine-tuning." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 250, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 250, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 250, + 91, + 447, + 102 + ], + "type": "text", + "content": "Reinforcement Learning Meets Visual Odometry" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 238, + 127 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 238, + 127 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 238, + 127 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 139, + 479, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 139, + 479, + 163 + ], + "spans": [ + { + "bbox": [ + 130, + 139, + 479, + 163 + ], + "type": "text", + "content": "Three paradigms exist to approach Visual Odometry: classical geometric-based, end-to-end, and hybrid approaches." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 163, + 482, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 163, + 482, + 413 + ], + "spans": [ + { + "bbox": [ + 130, + 163, + 482, + 413 + ], + "type": "text", + "content": "Classical approaches have been extensively studied in computer vision and robotics in the past 30 years [2, 13]. They achieve commercial-level accuracy and robustness. Indeed, classical VO algorithms are present in products like smartphones, VR-AR devices, and mobile robots. Classical VO algorithms are composed of two main modules: tracking and mapping. The tracking module uses camera images to estimate the motion. Three main approaches exist in the literature: direct methods [5], feature-based methods [11], and semi-direct methods [6]. Direct methods, such as DSO [5], work directly on the raw pixel intensities. These methods commonly extract image patches and estimate the camera trajectory by tracking the motion of such patches through consecutive images. The tracking is achieved by minimizing a photometric error defined on the raw pixel intensities. On the contrary, feature-based methods, such as ORBSLAM [11], extract points of interest, commonly known as visual features or keypoints, from the raw image pixels. The camera trajectory is estimated by tracking these points through consecutive images. Feature-based methods rely on scene texture to detect and track distinctive points, while direct methods use raw image pixels and can achieve higher reliability in low-texture environments. Semi-direct methods, such as SVO [6], combine keypoints and patches of raw pixel intensities to estimate the camera motion. SVO uses keypoints to find regions of interest in the current camera view and raw pixel intensities to track the patches centered at the detected keypoints." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 415, + 482, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 415, + 482, + 510 + ], + "spans": [ + { + "bbox": [ + 130, + 415, + 482, + 510 + ], + "type": "text", + "content": "Following the deep learning revolution seen in computer vision in the past years, learning-based end-to-end VO systems [23, 24, 26] have been proposed. These methods usually rely on convolutional layers to process the input images and subsequently, recurrent and/or feedforward layers to predict the camera 6-DoF pose. They are trained with ground truth pose supervision. End-to-end VO systems show better performance than classical approaches in scenarios covered in the training data. However, they lack robustness when deployed on environments (scene and motion) outside the training data distribution." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 510, + 482, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 482, + 629 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 482, + 629 + ], + "type": "text", + "content": "Hybrid methods [18, 19] combine the best of two paradigms: the geometric representation of the 3D world (classical methods), and the ability to process highly dimensional inputs (end-to-end methods). One of the state-of-the-art hybrid approaches is DROID-SLAM [18]. DROID-SLAM is a SLAM system that focuses on dense scene reconstruction, as well as, camera pose estimation. It uses a RAFT-inspired network architecture [17] followed by a differentiable dense bundle adjustment (BA) layer to iteratively update the camera poses and depth. The BA layer includes geometry constraints that improve the consistency of consecutive camera pose estimates. Using a similar differential BA layer, the hybrid method DPVO [19] tracks image patches over time using a recurrent network." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 630, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 630, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 630, + 482, + 665 + ], + "type": "text", + "content": "The performance of VO algorithms highly depends on heuristics and hyperparameters. Usually, hyperparameter tuning is done manually by expert users. Due to the high dimensional space and the sequential nature of the VO prob-" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 141, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 141, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 141, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 246, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 246, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 246, + 101 + ], + "type": "text", + "content": "Messikommer et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "text", + "content": "lem, automatic optimization-based hyperparameters tuning methods, such as Bayesian Optimization are difficult to use. An attempt to automatically select keyframes has been made in [15], where a deep network is used to decide if a new keyframe needs to be selected, as well as, to predict the camera pose." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 164, + 479, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 164, + 479, + 236 + ], + "spans": [ + { + "bbox": [ + 130, + 164, + 479, + 236 + ], + "type": "text", + "content": "In this paper, we model VO as a sequential decision-making task and apply Reinforcement Learning. Related to the problem of hyperparameter tuning with RL, Hyp-RL [10], poses the search for the best hyperparameters for training neural networks as a sequential decision process and uses RL to select the next hyperparameter set to be evaluated. In [4], RL is applied to dynamically choose the best set of hyperparameters of a deep network that performs object tracking." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 253, + 202, + 266 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 253, + 202, + 266 + ], + "spans": [ + { + "bbox": [ + 132, + 253, + 202, + 266 + ], + "type": "text", + "content": "3 Method" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 143, + 297, + 470, + 411 + ], + "blocks": [ + { + "bbox": [ + 143, + 297, + 470, + 411 + ], + "lines": [ + { + "bbox": [ + 143, + 297, + 470, + 411 + ], + "spans": [ + { + "bbox": [ + 143, + 297, + 470, + 411 + ], + "type": "image", + "image_path": "ad8f2e88031e9c9e80aa038834d2ce20128f278c540796b86fc1eb236ca40c3a.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 418, + 482, + 485 + ], + "lines": [ + { + "bbox": [ + 130, + 418, + 482, + 485 + ], + "spans": [ + { + "bbox": [ + 130, + 418, + 482, + 485 + ], + "type": "text", + "content": "Fig. 2: Agent Overview. Our agent takes as input the map statistics computed by the VO method at timestep " + }, + { + "bbox": [ + 130, + 418, + 482, + 485 + ], + "type": "inline_equation", + "content": "t_{i-1}" + }, + { + "bbox": [ + 130, + 418, + 482, + 485 + ], + "type": "text", + "content": " and a variable number of keypoints using a multi-head attention layer. Given these inputs, a two-layer MLP computes a multi-discrete probability distribution over the binary keyframe action and the grid size action. These actions are then used inside the VO method to process the current frame at timestep " + }, + { + "bbox": [ + 130, + 418, + 482, + 485 + ], + "type": "inline_equation", + "content": "t_i" + }, + { + "bbox": [ + 130, + 418, + 482, + 485 + ], + "type": "text", + "content": ", which will lead to the observation for the next timestep " + }, + { + "bbox": [ + 130, + 418, + 482, + 485 + ], + "type": "inline_equation", + "content": "t_{i+1}" + }, + { + "bbox": [ + 130, + 418, + 482, + 485 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 510, + 482, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 482, + 628 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 482, + 628 + ], + "type": "text", + "content": "We consider monocular Visual Odometry (VO), whose primary objective is to estimate the camera pose for each image within an image sequence. Despite extensive research and widespread application, the robustness and generalizability of VO methods in different scenes remain an ongoing challenge. Both traditional geometry-based approaches and hybrid methods incorporating learned components often rely on heuristic decision rules for critical tasks such as keyframe selection and keypoint filtering. These heuristic rules significantly affect the performance of the VO method and are closely linked to the deployed scene. Consequently, a substantial amount of fine-tuning is required to adapt these rules effectively to each deployed scenario." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 630, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 630, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 630, + 482, + 665 + ], + "type": "text", + "content": "To increase the robustness and generalizability of VO pipelines, we formulate the VO task as a sequential decision problem. Within this sequential framework, a deep agent is trained through RL. This agent is designed to make decisions" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 250, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 250, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 250, + 91, + 447, + 102 + ], + "type": "text", + "content": "Reinforcement Learning Meets Visual Odometry" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 175 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 175 + ], + "type": "text", + "content": "at each timestep for the VO method, taking into account the current state of the system, including map statistics, keypoints information and previous pose estimations. By training an adaptive agent, we can effectively reduce the reliance on heuristics and increase the generalizability of the underlying VO method across diverse scenes." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 177, + 482, + 250 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 177, + 482, + 250 + ], + "spans": [ + { + "bbox": [ + 130, + 177, + 482, + 250 + ], + "type": "text", + "content": "The structure of this section is the following. In Sec. 3.1, we present our problem formulation. In Sec. 3.2, we introduce our proposed deep neural agent with its corresponding input observations and output actions. In Sec. 3.3, we formulate the reward function, while in Sec. 3.4, we present the RL training framework. Finally, in Sec. 3.5, we explain the specific implementation of our RL framework using SVO [6] and DSO [5] as the underlying VO methods." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 273, + 269, + 285 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 273, + 269, + 285 + ], + "spans": [ + { + "bbox": [ + 132, + 273, + 269, + 285 + ], + "type": "text", + "content": "3.1 Problem Formulation" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 299, + 482, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 299, + 482, + 396 + ], + "spans": [ + { + "bbox": [ + 130, + 299, + 482, + 396 + ], + "type": "text", + "content": "We formulate the VO task as a sequential decision problem in which an agent interacts with its environment, see Fig. 1. Specifically, we consider the image sequence and the underlying VO pipeline to be fixed, both constituting the environment. The agent is a neural network that, at each timestep, selects actions for the VO pipeline, i.e., selecting keyframes and adapting the keypoint distribution by setting the grid size constraint (a grid is fitted to the image, and maximum one keypoint per grid cell is selected). Based on these actions, the VO method processes the next image of the sequence and computes the corresponding pose." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 396, + 482, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 396, + 482, + 552 + ], + "spans": [ + { + "bbox": [ + 130, + 396, + 482, + 552 + ], + "type": "text", + "content": "With this definition of the agent and environment, we formulate the VO task as a Markov decision process (MDP), denoted by the tuple " + }, + { + "bbox": [ + 130, + 396, + 482, + 552 + ], + "type": "inline_equation", + "content": "M = (S, A, p, r, \\gamma)" + }, + { + "bbox": [ + 130, + 396, + 482, + 552 + ], + "type": "text", + "content": ". A state " + }, + { + "bbox": [ + 130, + 396, + 482, + 552 + ], + "type": "inline_equation", + "content": "s \\in S" + }, + { + "bbox": [ + 130, + 396, + 482, + 552 + ], + "type": "text", + "content": " encapsulates the current image in the image sequence and the state of the VO pipeline, which includes properties such as feature tracks, previously estimated poses, the triangulate map, the optimization state, etc. To transition between VO states, actions " + }, + { + "bbox": [ + 130, + 396, + 482, + 552 + ], + "type": "inline_equation", + "content": "a \\in A" + }, + { + "bbox": [ + 130, + 396, + 482, + 552 + ], + "type": "text", + "content": " are executed in the VO method, while the environment supplies the next frame of the sequence. The state transition probability " + }, + { + "bbox": [ + 130, + 396, + 482, + 552 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 130, + 396, + 482, + 552 + ], + "type": "text", + "content": " defines the likelihood of moving from one state " + }, + { + "bbox": [ + 130, + 396, + 482, + 552 + ], + "type": "inline_equation", + "content": "s_i" + }, + { + "bbox": [ + 130, + 396, + 482, + 552 + ], + "type": "text", + "content": " to another " + }, + { + "bbox": [ + 130, + 396, + 482, + 552 + ], + "type": "inline_equation", + "content": "s_j" + }, + { + "bbox": [ + 130, + 396, + 482, + 552 + ], + "type": "text", + "content": " given actions " + }, + { + "bbox": [ + 130, + 396, + 482, + 552 + ], + "type": "inline_equation", + "content": "a_i" + }, + { + "bbox": [ + 130, + 396, + 482, + 552 + ], + "type": "text", + "content": ". Finally, each state transition is associated with a reward " + }, + { + "bbox": [ + 130, + 396, + 482, + 552 + ], + "type": "inline_equation", + "content": "r(s, a)" + }, + { + "bbox": [ + 130, + 396, + 482, + 552 + ], + "type": "text", + "content": ", dependent on the ensuing state and the chosen actions. The rewards of the future are discounted by " + }, + { + "bbox": [ + 130, + 396, + 482, + 552 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 130, + 396, + 482, + 552 + ], + "type": "text", + "content": ". In the context of VO, the reward function evaluates the accuracy of the pose estimation and can also incorporate non-differentiable criteria such as runtime and map size, among others." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 553, + 482, + 625 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 553, + 482, + 625 + ], + "spans": [ + { + "bbox": [ + 130, + 553, + 482, + 625 + ], + "type": "text", + "content": "In our formulation, the agent is characterized by a control policy " + }, + { + "bbox": [ + 130, + 553, + 482, + 625 + ], + "type": "inline_equation", + "content": "\\pi : S \\times A \\to \\mathbb{R}" + }, + { + "bbox": [ + 130, + 553, + 482, + 625 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 553, + 482, + 625 + ], + "type": "inline_equation", + "content": "\\pi" + }, + { + "bbox": [ + 130, + 553, + 482, + 625 + ], + "type": "text", + "content": " specifies the actions " + }, + { + "bbox": [ + 130, + 553, + 482, + 625 + ], + "type": "inline_equation", + "content": "a \\in A" + }, + { + "bbox": [ + 130, + 553, + 482, + 625 + ], + "type": "text", + "content": " to be taken in each state " + }, + { + "bbox": [ + 130, + 553, + 482, + 625 + ], + "type": "inline_equation", + "content": "s \\in S" + }, + { + "bbox": [ + 130, + 553, + 482, + 625 + ], + "type": "text", + "content": ". Since the control policy is non-deterministic, it induces a probabilistic trajectory of states and actions " + }, + { + "bbox": [ + 130, + 553, + 482, + 625 + ], + "type": "inline_equation", + "content": "\\tau = \\{s_0, a_0, s_1, a_1, \\dots, s_T, a_{T-1}\\}" + }, + { + "bbox": [ + 130, + 553, + 482, + 625 + ], + "type": "text", + "content": ", abbreviated as " + }, + { + "bbox": [ + 130, + 553, + 482, + 625 + ], + "type": "inline_equation", + "content": "\\tau \\sim \\pi" + }, + { + "bbox": [ + 130, + 553, + 482, + 625 + ], + "type": "text", + "content": ". Following the control policy " + }, + { + "bbox": [ + 130, + 553, + 482, + 625 + ], + "type": "inline_equation", + "content": "\\pi" + }, + { + "bbox": [ + 130, + 553, + 482, + 625 + ], + "type": "text", + "content": ", the expected sum over the discounted rewards for each timestep can be expressed with its value function" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 236, + 637, + 481, + 669 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 236, + 637, + 481, + 669 + ], + "spans": [ + { + "bbox": [ + 236, + 637, + 481, + 669 + ], + "type": "interline_equation", + "content": "V ^ {\\pi} (s _ {i}) = \\mathbb {E} _ {\\tau \\sim \\pi} \\left[ \\sum_ {t = i} ^ {\\infty} \\gamma^ {t} r (s _ {t}, a _ {t}) \\right]. \\tag {1}", + "image_path": "bcf923c2864241cad335c95bfa70daad4fd26b93bc80ae42c4caa980228e5e9a.jpg" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 246, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 246, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 246, + 101 + ], + "type": "text", + "content": "Messikommer et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 139 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 139 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 139 + ], + "type": "text", + "content": "The overarching goal is to find a policy " + }, + { + "bbox": [ + 130, + 116, + 479, + 139 + ], + "type": "inline_equation", + "content": "\\pi" + }, + { + "bbox": [ + 130, + 116, + 479, + 139 + ], + "type": "text", + "content": " that maximizes the expected sum of discounted rewards." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 261, + 140, + 481, + 156 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 261, + 140, + 481, + 156 + ], + "spans": [ + { + "bbox": [ + 261, + 140, + 481, + 156 + ], + "type": "interline_equation", + "content": "\\pi^ {*} = \\arg \\max _ {\\pi} V ^ {\\pi} \\left(s _ {0}\\right) \\tag {2}", + "image_path": "8a3bdf00e4d14d3ed17a3fbd857aa9286137fa17875763844d3b869bd8cad5f4.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 159, + 479, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 159, + 479, + 194 + ], + "spans": [ + { + "bbox": [ + 130, + 159, + 479, + 194 + ], + "type": "text", + "content": "Given the reward formulation, the optimal policy provides decisions to the VO pipeline, aiming to minimize the expected position error while accounting for additional reward terms like runtime." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 209, + 258, + 221 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 209, + 258, + 221 + ], + "spans": [ + { + "bbox": [ + 131, + 209, + 258, + 221 + ], + "type": "text", + "content": "3.2 Deep Neural Agent" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 226, + 480, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 226, + 480, + 380 + ], + "spans": [ + { + "bbox": [ + 130, + 226, + 480, + 380 + ], + "type": "text", + "content": "The central element of our proposed framework is the learned agent embodying the policy " + }, + { + "bbox": [ + 130, + 226, + 480, + 380 + ], + "type": "inline_equation", + "content": "\\pi" + }, + { + "bbox": [ + 130, + 226, + 480, + 380 + ], + "type": "text", + "content": ", which selects optimal settings for the VO method based on the current keypoints and map statistics, i.e., relative poses to the newest and oldest keyframe. Given the inherent complexity of the underlying VO method, supplying the agent network with a comprehensive state encompassing all stored map elements, optimization states, etc., is neither feasible nor advantageous. Instead, the learned agent receives a selected subset of observations " + }, + { + "bbox": [ + 130, + 226, + 480, + 380 + ], + "type": "inline_equation", + "content": "o \\in O" + }, + { + "bbox": [ + 130, + 226, + 480, + 380 + ], + "type": "text", + "content": " approximating the VO state within the policy " + }, + { + "bbox": [ + 130, + 226, + 480, + 380 + ], + "type": "inline_equation", + "content": "\\pi: O \\times A \\to \\mathbb{R}" + }, + { + "bbox": [ + 130, + 226, + 480, + 380 + ], + "type": "text", + "content": ". In our system, the observations are derived from the local map and include relative poses to previous keyframes, along with the information about the currently tracked keypoints. Notably, the number of keypoints is variable and may vary depending on the characteristics of the image sequence. Therefore, the network architecture of the agent should be capable of processing inputs of variable sizes." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 382, + 480, + 501 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 382, + 480, + 501 + ], + "spans": [ + { + "bbox": [ + 130, + 382, + 480, + 501 + ], + "type": "text", + "content": "To handle inputs of varying sizes, we leverage a Perceiver [8,9] architecture, illustrated in Fig. 2. Specifically, our approach involves a Variable Encoder that incorporates a multi-head attention layer [21] for processing the image positions, and the estimated depth of each tracked keypoint " + }, + { + "bbox": [ + 130, + 382, + 480, + 501 + ], + "type": "inline_equation", + "content": "k \\in \\mathbb{R}^{N \\times 3}" + }, + { + "bbox": [ + 130, + 382, + 480, + 501 + ], + "type": "text", + "content": ". To project the information contained in the variable number of keypoints " + }, + { + "bbox": [ + 130, + 382, + 480, + 501 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 382, + 480, + 501 + ], + "type": "text", + "content": " to a fixed dimension, we use " + }, + { + "bbox": [ + 130, + 382, + 480, + 501 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 130, + 382, + 480, + 501 + ], + "type": "text", + "content": " number of learned tokens " + }, + { + "bbox": [ + 130, + 382, + 480, + 501 + ], + "type": "inline_equation", + "content": "z \\in \\mathbb{R}^{M \\times D}" + }, + { + "bbox": [ + 130, + 382, + 480, + 501 + ], + "type": "text", + "content": " with dimension of " + }, + { + "bbox": [ + 130, + 382, + 480, + 501 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 130, + 382, + 480, + 501 + ], + "type": "text", + "content": " as query " + }, + { + "bbox": [ + 130, + 382, + 480, + 501 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 130, + 382, + 480, + 501 + ], + "type": "text", + "content": " values. These query tokens are attended to by the projected keypoint information serving as both key " + }, + { + "bbox": [ + 130, + 382, + 480, + 501 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 382, + 480, + 501 + ], + "type": "text", + "content": " and values " + }, + { + "bbox": [ + 130, + 382, + 480, + 501 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 130, + 382, + 480, + 501 + ], + "type": "text", + "content": ". Finally, the fixed output of the Variable Encoder is subsequently processed along with the statistics of the local map, using a two-layer MLP with ReLu non-linearities." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 502, + 480, + 609 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 502, + 480, + 609 + ], + "spans": [ + { + "bbox": [ + 130, + 502, + 480, + 609 + ], + "type": "text", + "content": "The output of the agent network is a discrete probability distribution over a set of actions expressed in the form of multiple independent categorical distributions. In our implementation, the agent predicts at each timestep a binary distribution over the decision to take the current frame as a keyframe. Additionally, in the case of SVO, the agent also predicts a separate and independent categorical distribution over a set of values " + }, + { + "bbox": [ + 130, + 502, + 480, + 609 + ], + "type": "inline_equation", + "content": "\\{20, 25, \\dots, 40\\}" + }, + { + "bbox": [ + 130, + 502, + 480, + 609 + ], + "type": "text", + "content": " for the keypoint grid size. This grid size ensures a uniform distribution of keypoints on the image plane by enforcing that only one keypoint can reside within a given grid cell. Since DSO does not extract keypoints, there is no grid size selection necessary." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 624, + 200, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 624, + 200, + 635 + ], + "spans": [ + { + "bbox": [ + 131, + 624, + 200, + 635 + ], + "type": "text", + "content": "3.3 Reward" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 641, + 479, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 479, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 479, + 665 + ], + "type": "text", + "content": "Another critical component within the RL framework is the design of the reward function. The feedback of the reward at each timestep guides the deep agent" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 250, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 250, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 250, + 91, + 448, + 102 + ], + "type": "text", + "content": "Reinforcement Learning Meets Visual Odometry" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 187, + 116, + 428, + 190 + ], + "blocks": [ + { + "bbox": [ + 187, + 116, + 428, + 190 + ], + "lines": [ + { + "bbox": [ + 187, + 116, + 428, + 190 + ], + "spans": [ + { + "bbox": [ + 187, + 116, + 428, + 190 + ], + "type": "image", + "image_path": "482c301693015bf0c8054787fb1f741622030cf152b411d47e51d7ee866fd520.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 201, + 482, + 245 + ], + "lines": [ + { + "bbox": [ + 131, + 201, + 482, + 245 + ], + "spans": [ + { + "bbox": [ + 131, + 201, + 482, + 245 + ], + "type": "text", + "content": "Fig. 3: Position Error. To closely relate the pose prediction accuracy to the current action, we employ a sliding window of five timesteps to align the ground truth and estimated trajectory using a scale factor " + }, + { + "bbox": [ + 131, + 201, + 482, + 245 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 131, + 201, + 482, + 245 + ], + "type": "text", + "content": ", a translation vector " + }, + { + "bbox": [ + 131, + 201, + 482, + 245 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 131, + 201, + 482, + 245 + ], + "type": "text", + "content": ", and a rotation " + }, + { + "bbox": [ + 131, + 201, + 482, + 245 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 131, + 201, + 482, + 245 + ], + "type": "text", + "content": ". The error at the current timestep " + }, + { + "bbox": [ + 131, + 201, + 482, + 245 + ], + "type": "inline_equation", + "content": "t_i" + }, + { + "bbox": [ + 131, + 201, + 482, + 245 + ], + "type": "text", + "content": " is then used as the negative position reward." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 273, + 482, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 273, + 482, + 464 + ], + "spans": [ + { + "bbox": [ + 130, + 273, + 482, + 464 + ], + "type": "text", + "content": "toward a desired behavioral pattern. In the case of VO, naturally, the error between the estimated and the ground truth pose is the primary focus of the reward. However, since monocular VO methods cannot inherently estimate the absolute scale of the trajectory, it becomes necessary to scale the predicted poses before computing the error against the ground truth trajectory. Furthermore, another crucial aspect of ensuring stable feedback for the RL algorithm is to closely link the reward function of a timestep " + }, + { + "bbox": [ + 130, + 273, + 482, + 464 + ], + "type": "inline_equation", + "content": "t_i" + }, + { + "bbox": [ + 130, + 273, + 482, + 464 + ], + "type": "text", + "content": " to the action taken at that specific timestep, thus mitigating the value assignment problem. Consequently, we adopt a sliding window approach to compute the error between the estimated and ground truth poses. Specifically, we employ a window of five timestamps " + }, + { + "bbox": [ + 130, + 273, + 482, + 464 + ], + "type": "inline_equation", + "content": "\\{t_{i-4}, t_{i-3}, \\dots, t_i\\}" + }, + { + "bbox": [ + 130, + 273, + 482, + 464 + ], + "type": "text", + "content": " to compute the error at the current timestep " + }, + { + "bbox": [ + 130, + 273, + 482, + 464 + ], + "type": "inline_equation", + "content": "t_i" + }, + { + "bbox": [ + 130, + 273, + 482, + 464 + ], + "type": "text", + "content": ", see Fig. 3. The first three poses in the window are used to align the estimated trajectory to the ground truth by minimizing the distances between the positions with the Umeyama method [20]. This alignment method accommodates missing scale information and differing coordinate origins, aligning both trajectories through an Euclidean transformation involving rotation, translation, and scaling." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 466, + 482, + 550 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 466, + 482, + 550 + ], + "spans": [ + { + "bbox": [ + 130, + 466, + 482, + 550 + ], + "type": "text", + "content": "By designing the position reward function, we can influence the policy to weigh accuracy higher at the cost of lower robustness or the other way around. This trade-off also shows the benefit of our dynamic agent over manual tuning since the decisions can be made in an online fashion, adapting to the current conditions in the image sequence. To handle this trade-off, we clip the positional error " + }, + { + "bbox": [ + 130, + 466, + 482, + 550 + ], + "type": "inline_equation", + "content": "e_{\\mathrm{tran}}" + }, + { + "bbox": [ + 130, + 466, + 482, + 550 + ], + "type": "text", + "content": " around zero, see Eq. 3, to positively enforce accurate states while giving a penalty for states having a large positional error." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 552, + 482, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 552, + 482, + 635 + ], + "spans": [ + { + "bbox": [ + 130, + 552, + 482, + 635 + ], + "type": "text", + "content": "Given that our framework leverages RL, we have the opportunity to integrate non-differentiable performance feedback into the reward system. Thus, we can enforce a faster runtime by penalizing the insertion of keyframes, i.e., directly penalizing the keyframe action " + }, + { + "bbox": [ + 130, + 552, + 482, + 635 + ], + "type": "inline_equation", + "content": "a_{\\mathrm{keyframe}}" + }, + { + "bbox": [ + 130, + 552, + 482, + 635 + ], + "type": "text", + "content": ". Using the keyframe action instead of directly penalizing the runtime mitigates the dependence on the usage of the hardware during training. Our final reward function is formulated in Eq. 3, where we use 0.01 for " + }, + { + "bbox": [ + 130, + 552, + 482, + 635 + ], + "type": "inline_equation", + "content": "\\lambda_{1}" + }, + { + "bbox": [ + 130, + 552, + 482, + 635 + ], + "type": "text", + "content": " and set " + }, + { + "bbox": [ + 130, + 552, + 482, + 635 + ], + "type": "inline_equation", + "content": "\\lambda_{2}" + }, + { + "bbox": [ + 130, + 552, + 482, + 635 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 130, + 552, + 482, + 635 + ], + "type": "inline_equation", + "content": "5\\times 10^{-3}" + }, + { + "bbox": [ + 130, + 552, + 482, + 635 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 216, + 650, + 481, + 662 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 216, + 650, + 481, + 662 + ], + "spans": [ + { + "bbox": [ + 216, + 650, + 481, + 662 + ], + "type": "interline_equation", + "content": "r = \\lambda_ {1} \\max (- 1, 0. 2 - e _ {\\mathrm {t r a n}}) - \\lambda_ {2} a _ {\\text {k e y f r a m e}} \\tag {3}", + "image_path": "9092f79e138655feeb472c267ebafd7df0cdf4678d7d6f9235ea45f89e1ad998.jpg" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 247, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 247, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 247, + 101 + ], + "type": "text", + "content": "Messikommer et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 340, + 128 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 340, + 128 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 340, + 128 + ], + "type": "text", + "content": "3.4 Reinforcement Learning Framework" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 140, + 482, + 295 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 140, + 482, + 295 + ], + "spans": [ + { + "bbox": [ + 130, + 140, + 482, + 295 + ], + "type": "text", + "content": "We use the on-policy algorithm Proximal Policy Optimization (PPO) [14] implemented in Stable Baselines3 [12] for training our agent. As commonly done, the RL training is conducted in two alternating stages: the rollout and the policy-update stage. During the rollout stage, the RL agent interacts with the environment, i.e., the underlying VO method and image sequences from TartanAir [25], via its actions to collect experiences comprising the actions, observations, and rewards at each timestep. While the experiences are collected, the network weights of the agent are not updated. In the second stage, the collected experiences of the rollout stage are used to update both the agent and the critic. The task of the critic is to approximate the value function, which represents the expected return of the current policy at a given state. The critic is used to stabilize the training inside the policy update steps in accordance with the PPO algorithm, which is designed to limit large changes in the policy." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 296, + 482, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 296, + 482, + 380 + ], + "spans": [ + { + "bbox": [ + 130, + 296, + 482, + 380 + ], + "type": "text", + "content": "Since the performance of the VO pipelines depends not only on the decisions of the agent but also on the current image sequence, we introduce a privileged critic to stabilize the training. This privileged critic has access to both the current and future ground truth poses, enabling a better assessment of whether tracking performance was influenced by the decisions of the agent or the difficulty of the underlying image sequence. Importantly, this privileged critic with access to ground truth poses is only used during training." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 380, + 482, + 524 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 380, + 482, + 524 + ], + "spans": [ + { + "bbox": [ + 130, + 380, + 482, + 524 + ], + "type": "text", + "content": "Another crucial aspect is the handling of the different stages of the VO system, which include the normal mode of camera tracking, the mode for the relocalization in the current map, and the initialization mode. Since the VO system only provides information about its full state in the tracking mode, we declare the current state a valid state if the previous state was in the tracking mode. Notably, the agent can exclusively make decisions in valid stages. To account for the valid states in the policy-update phase, we use a masked rollout buffer, which computes the return using all collected states, but samples only valid states for updating the policy network. As states in the initialization stages can not be influenced by the agent, we directly use the ground truth poses to triangulate tracked keypoints. The initialization with the ground truth poses also has the benefit of a significantly faster runtime compared to the original initialization." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 525, + 482, + 572 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 525, + 482, + 572 + ], + "spans": [ + { + "bbox": [ + 130, + 525, + 482, + 572 + ], + "type": "text", + "content": "Since RL is sample inefficient, it is crucial to have a fast runtime of the environment interactions. For our RL framework, the runtime of each of the main components in the rollout phase using 100 SVO instances in parallel is reported in Tab. 6." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 594, + 218, + 605 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 594, + 218, + 605 + ], + "spans": [ + { + "bbox": [ + 132, + 594, + 218, + 605 + ], + "type": "text", + "content": "3.5 VO System" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "In general, our proposed agent network can be combined with any VO system that relies on decision components. In this work, we use SVO [6] and DSO [5], which feature a very small runtime. Additionally, we also test on the slower SLAM pipeline ORB-SLAM3 [3]. From a practical viewpoint, the small runtime" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 250, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 250, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 250, + 91, + 447, + 102 + ], + "type": "text", + "content": "Reinforcement Learning Meets Visual Odometry" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": "and parallelization accelerate training and partially compensate for the inherent low sample efficiency of RL." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 140, + 481, + 272 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 140, + 481, + 272 + ], + "spans": [ + { + "bbox": [ + 130, + 140, + 481, + 272 + ], + "type": "text", + "content": "SVO is a semi-direct visual odometry composed of two main threads: motion estimation and mapping thread. The first step of the motion estimation thread is image alignment, where the pose of the new frame is estimated by aligning pixel intensity patches from the previous keyframe. The next step is feature alignment, where the positions of the keypoints on the current image are refined while the camera poses are fixed. The last step is a sparse bundle adjustment, where both the camera poses and the locations of the 3D points included in the current map are optimized. The motion thread assumes that the depth of the mapped 3D points is known. It is the objective of the mapping thread to estimate the depth. The mapping thread estimates the depth of each tracked pixel from multiple observations by means of a recursive Bayesian filter [22]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 272, + 482, + 390 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 272, + 482, + 390 + ], + "spans": [ + { + "bbox": [ + 130, + 272, + 482, + 390 + ], + "type": "text", + "content": "DSO is a direct visual odometry also composed of two threads: the front-end thread and the mapping thread. The front-end thread performs the initial frame tracking where the pose of the current frame is tracked with respect to the previous keyframe by aligning pixel intensities patches. The front-end is also in charge of selecting the active candidate points, performing outlier rejection and occlusion detection, and selecting keyframes. The mapping thread optimizes a non-linear cost function based on the photometric error. The camera poses, intrinsics, and inverse depth value of the active points in the current map are jointly optimized. This optimization can be seen as the photometric equivalent of sparse bundle adjustment." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 392, + 481, + 417 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 392, + 481, + 417 + ], + "spans": [ + { + "bbox": [ + 130, + 392, + 481, + 417 + ], + "type": "text", + "content": "We use SVO to run ablation studies on the contributions of the main components of the RL system (runtime, inputs, actions, and reward tuning)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 435, + 230, + 449 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 435, + 230, + 449 + ], + "spans": [ + { + "bbox": [ + 132, + 435, + 230, + 449 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 461, + 482, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 461, + 482, + 568 + ], + "spans": [ + { + "bbox": [ + 130, + 461, + 482, + 568 + ], + "type": "text", + "content": "Dataset For training our agent, we leverage the synthetic sequences from TartanAir [25], a comprehensive dataset offering, among other measurements, ground truth camera poses. The dataset features a wide range of simulated scenes with challenging conditions such as low light and weather effects. The total number of different training sequences is 337, amounting to 279987 images or timesteps for the training of the agent. Importantly, we consider the dataset in our RL framework to be fixed without the possibility of the agent influencing the next image. Notably, as our method is trained on abstractions rather than image pixels, the risk of domain shift is mitigated." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 569, + 481, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 569, + 481, + 641 + ], + "spans": [ + { + "bbox": [ + 130, + 569, + 481, + 641 + ], + "type": "text", + "content": "For evaluation, we use three VO benchmarks: EuRoC [1], TUM-RGBD [16] and KITTI [7]. The EuRoC dataset consists of 11 sequences recorded via a drone, with ground truth pose measurements obtained from a motion capture system or a laser tracker. The TUM-RGBD dataset features 9 sequences recorded via a hand-held camera with ground truth poses provided by a motion capture system. For the results on the KITTI dataset, we refer to the supplementary." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "content": "Evaluation Following standard practice in evaluating VO algorithms, we use the Absolute Translation Error (ATE [m]) [27] as the error metric for the EuRoC and" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 246, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 246, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 246, + 101 + ], + "type": "text", + "content": "Messikommer et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "text", + "content": "TUM-RGBD datasets. Before computing the error metric, the estimated trajectories are aligned to the ground truth trajectory using the Umeyama method [20]. For the relative position error on EuRoC and TUM-RGBD, we refer to the supplementary." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 164, + 480, + 294 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 164, + 480, + 294 + ], + "spans": [ + { + "bbox": [ + 130, + 164, + 480, + 294 + ], + "type": "text", + "content": "Since we want to guarantee a fair comparison between the methods with and without RL, we remove the effect of the initialization phase at the beginning of each sequence by providing the ground truth poses to the VO algorithm. Similar to the original initialization of the underlying VO, extracted keypoints are tracked from the first frame of the sequence and triangulated once a large enough disparity is detected. However, instead of estimating a relative pose, the ground truth poses are used for the triangulation. Following this procedure, all of the tested methods start with the same high-quality map, effectively eliminating the unwanted influence of the initialization. Naturally, the ground truth initialization is not used for the comparison against state-of-the-art VO methods. Instead, we take the average of the performance metrics for five runs." + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 133, + 314, + 299, + 390 + ], + "blocks": [ + { + "bbox": [ + 133, + 314, + 299, + 390 + ], + "lines": [ + { + "bbox": [ + 133, + 314, + 299, + 390 + ], + "spans": [ + { + "bbox": [ + 133, + 314, + 299, + 390 + ], + "type": "image", + "image_path": "f2c9368a7608c352a368012f47362956f219586ba42f715df1e58e761a77042f.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 399, + 299, + 442 + ], + "lines": [ + { + "bbox": [ + 131, + 399, + 299, + 442 + ], + "spans": [ + { + "bbox": [ + 131, + 399, + 299, + 442 + ], + "type": "text", + "content": "Fig. 4: Keyframe Selection. The number of selected keyframes by SVO at different translational and angular velocities for the EuRoC dataset." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 315, + 316, + 480, + 399 + ], + "blocks": [ + { + "bbox": [ + 315, + 316, + 480, + 399 + ], + "lines": [ + { + "bbox": [ + 315, + 316, + 480, + 399 + ], + "spans": [ + { + "bbox": [ + 315, + 316, + 480, + 399 + ], + "type": "image", + "image_path": "77dcf3b29f3fdd213401076bc0920635636f2a0adb695fb07c7cb24c5b1abc25.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 312, + 408, + 482, + 441 + ], + "lines": [ + { + "bbox": [ + 312, + 408, + 482, + 441 + ], + "spans": [ + { + "bbox": [ + 312, + 408, + 482, + 441 + ], + "type": "text", + "content": "Fig.5: Top-Down View of Trajectories. Visualization of the predicted trajectory using SVO with and without RL." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 135, + 483, + 479, + 523 + ], + "blocks": [ + { + "bbox": [ + 135, + 483, + 479, + 523 + ], + "lines": [ + { + "bbox": [ + 135, + 483, + 479, + 523 + ], + "spans": [ + { + "bbox": [ + 135, + 483, + 479, + 523 + ], + "type": "table", + "html": "
MethodMH01MH02MH03MH04MH05V101V102V103V201V202V203Avg
SVO0.1830.4900.3813.2361.2930.3200.333-0.0840.5531.431-
RL SVO0.2360.1240.2412.1491.9820.2250.3980.3940.0690.3341.5060.696
DSO0.0870.0440.3540.1840.2601.2231.6241.1250.0590.0651.5160.594
RL DSO0.0370.0330.2410.1681.1920.2070.4991.4500.0430.1641.2860.483
", + "image_path": "a7ad84ea594b807a5d9dd7706b1686fbdb9d2bac76bb4c6be6eaea58d28f4119.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 524, + 480, + 578 + ], + "lines": [ + { + "bbox": [ + 130, + 524, + 480, + 578 + ], + "spans": [ + { + "bbox": [ + 130, + 524, + 480, + 578 + ], + "type": "text", + "content": "Table 1: EuRoC. We report the ATE [m] of our RL agents and the heuristic-based VO methods with parameters obtained by an extensive grid search. Our RL agents outperform the heuristic-based VO methods in terms of accuracy (lower average error) and robustness (RL SVO can complete the sequence V103 while the baseline SVO fails to track all the camera poses due to fast rotations)." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 131, + 625, + 198, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 625, + 198, + 635 + ], + "spans": [ + { + "bbox": [ + 131, + 625, + 198, + 635 + ], + "type": "text", + "content": "4.1 Results" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 641, + 479, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 479, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 479, + 665 + ], + "type": "text", + "content": "EuRoC We compare our RL agent with SVO and DSO against the heuristic-based VO with parameters obtained by tuning on this dataset using an extensive" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 250, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 250, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 250, + 91, + 447, + 102 + ], + "type": "text", + "content": "Reinforcement Learning Meets Visual Odometry" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 147, + 114, + 469, + 156 + ], + "blocks": [ + { + "bbox": [ + 147, + 114, + 469, + 156 + ], + "lines": [ + { + "bbox": [ + 147, + 114, + 469, + 156 + ], + "spans": [ + { + "bbox": [ + 147, + 114, + 469, + 156 + ], + "type": "table", + "html": "
Method360deskdesk2floorplantroomrpyteddyxyzAvg
SVO0.1860.6810.898-0.3200.8050.0530.7690.0570.471
RL SVO0.1890.5560.755-0.2600.8040.0560.6970.0620.422
DSO-0.2370.893-0.3200.9310.057-0.053-
RL DSO0.1870.6820.888-0.625-0.0540.7560.105-
", + "image_path": "10f2f41a6ea44cb0d5f253fea05e06419d20b61b65eddf074db42601e265f0e0.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 136, + 234, + 477, + 282 + ], + "blocks": [ + { + "bbox": [ + 130, + 157, + 481, + 212 + ], + "lines": [ + { + "bbox": [ + 130, + 157, + 481, + 212 + ], + "spans": [ + { + "bbox": [ + 130, + 157, + 481, + 212 + ], + "type": "text", + "content": "Table 2: TUM-RGBD. We report the ATE [m] of our RL agents and the heuristic-base VO methods with parameters obtained by an extensive grid search. Our RL SVO outperforms the heuristic-based SVO methods in terms of accuracy (lower average error). Although this dataset is challenging for DSO, due to the use of a rolling shutter camera, RL DSO improves the robustness with two more sequences completed." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 136, + 234, + 477, + 282 + ], + "lines": [ + { + "bbox": [ + 136, + 234, + 477, + 282 + ], + "spans": [ + { + "bbox": [ + 136, + 234, + 477, + 282 + ], + "type": "table", + "html": "
MethodMH01MH02MH03MH04MH05V101V102V103V201V202V203Avg
TartanVO [24]0.6390.3250.5501.1531.0210.4470.3890.6220.4330.7491.1520.680
DROID-VO [18]0.1630.1210.2420.3990.2700.1030.1650.1580.1020.1150.2040.186
DPVO [19]0.0870.0550.1580.1370.1140.0500.1400.0860.0570.0490.2110.105
RL SVO0.2891.5990.5522.4071.8260.2290.8130.4960.0850.9651.4340.972
RL DSO0.0320.0370.1660.1420.1100.1070.0891.3310.0410.1521.3130.320
", + "image_path": "c976f4271e3a598c8c99085310176680414cf46e93a91b6d2c5bafcd01f64549.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 282, + 480, + 304 + ], + "lines": [ + { + "bbox": [ + 132, + 282, + 480, + 304 + ], + "spans": [ + { + "bbox": [ + 132, + 282, + 480, + 304 + ], + "type": "text", + "content": "Table 3: EuRoC. We report the ATE [m] of our RL agents and the state-of-the-art VO methods. Our RL DSO achieves the best accuracy in 5 sequences." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 130, + 340, + 482, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 340, + 482, + 483 + ], + "spans": [ + { + "bbox": [ + 130, + 340, + 482, + 483 + ], + "type": "text", + "content": "grid search. Qualitative estimation results of SVO are shown in the left plot in Fig. 5 from a top-down view of the trajectory estimated by RL SVO, the baseline VO, and the ground truth. Tab. 1 reports the ATE. In the case of SVO, the combination with our RL agent achieves the best ATE in most of the sequences. Remarkably, our RL SVO can estimate the entire camera trajectory in the sequence V103 while the baseline fails. This sequence is one of the most difficult in the EuRoC dataset due to fast camera rotations. The benefit of our RL agent is also evident for DSO, which achieves consistently a lower tracking error, leading to a " + }, + { + "bbox": [ + 130, + 340, + 482, + 483 + ], + "type": "inline_equation", + "content": "19\\%" + }, + { + "bbox": [ + 130, + 340, + 482, + 483 + ], + "type": "text", + "content": " lower mean tracking error. We report in Tab. 3 the comparison of our RL agent against the state-of-the-art VO methods, which are taken from [19]. The results reported for our methods are averaged over five runs. Our RL DSO achieves the best ATE in five sequences." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 484, + 482, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 484, + 482, + 628 + ], + "spans": [ + { + "bbox": [ + 130, + 484, + 482, + 628 + ], + "type": "text", + "content": "TUM-RGBD Similar to EuRoC, we use a parameter grid search to find the best parameters for the baseline DSO and SVO methods. We report the ATE in Tab. 2, which shows that SVO RL achieves a lower ATE in most of the sequences and a lower overall mean error. Since the images of TUM-RGBD are not captured with a global shutter camera, DSO, relying on photometric alignment, struggles to finish all the sequences. Nevertheless, our RL agent improves the robustness in some challenging conditions with two more sequences successfully completed. For completeness, the top-down view of the trajectory estimated by RL SVO, the baseline, and the ground truth are visualized in the right plots in Fig. 5. We report in Tab. 4 the comparison of our RL agent against the state-of-the-art VO methods. The results reported for our methods are averaged over five runs. The results of the state-of-the-art methods are taken from [19]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 629, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 482, + 665 + ], + "type": "text", + "content": "Keyframe Selection In Fig. 4, we give insights into the keyframe selection strategy employed by our RL agent. We visualize the histogram of states at which keyframes are selected and their corresponding rotational and translation" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 246, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 246, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 246, + 101 + ], + "type": "text", + "content": "Messikommer et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 147, + 114, + 468, + 156 + ], + "blocks": [ + { + "bbox": [ + 147, + 114, + 468, + 156 + ], + "lines": [ + { + "bbox": [ + 147, + 114, + 468, + 156 + ], + "spans": [ + { + "bbox": [ + 147, + 114, + 468, + 156 + ], + "type": "table", + "html": "
Method360deskdesk2floorplantroomryteddyxyzAvg
DROID-VO [18]0.1610.0280.0990.0330.0280.3270.0280.1690.0130.098
DPVO [19]0.1350.0380.0480.0400.0360.3940.0340.0640.0120.089
RL SVO0.1830.6480.790-0.3320.7860.0520.7410.071-
RL DSO-0.4380.765-0.479---0.057-
", + "image_path": "e9fd4308e73a2dd7c537ad1a08b169f84d341608f87f84bd1b5b2354743f82f7.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 137, + 199, + 479, + 252 + ], + "blocks": [ + { + "bbox": [ + 132, + 156, + 479, + 177 + ], + "lines": [ + { + "bbox": [ + 132, + 156, + 479, + 177 + ], + "spans": [ + { + "bbox": [ + 132, + 156, + 479, + 177 + ], + "type": "text", + "content": "Table 4: TUM-RGBD. We report the ATE [m] of our RL agents and the state-of-the-art VO methods." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 137, + 199, + 479, + 252 + ], + "lines": [ + { + "bbox": [ + 137, + 199, + 479, + 252 + ], + "spans": [ + { + "bbox": [ + 137, + 199, + 479, + 252 + ], + "type": "table", + "html": "
MethodMH01MH02MH03MH04MH05V101V102V103V201V202V203Avg
SVO0.1830.4900.3813.2361.2930.3200.333-0.0840.5531.431-
RL SVO0.2360.1240.2412.1491.9820.2250.3980.3940.0690.3341.5060.696
RL SVO (w/o keypoints)0.2550.1600.4191.9790.5880.1730.6440.8490.1821.394--
RL SVO (w/o privileged c.)0.1300.0990.8343.2830.9580.1400.2860.4590.1160.6181.3730.754
RL SVO (w/o keyframe)0.1790.3510.3023.0721.0210.325--0.1790.2581.464-
RL SVO (w/o gridsize)0.1520.1830.2932.1541.5370.1980.2690.8080.1081.6831.4050.799
", + "image_path": "ed8f1c26f19076dfeba5ff4b2cc25db4af9fde2c0a3dacb9a0ac9c5f436309ca.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 141, + 285, + 290, + 327 + ], + "blocks": [ + { + "bbox": [ + 133, + 252, + 479, + 262 + ], + "lines": [ + { + "bbox": [ + 133, + 252, + 479, + 262 + ], + "spans": [ + { + "bbox": [ + 133, + 252, + 479, + 262 + ], + "type": "text", + "content": "Table 5: EuRoC Ablations. Ablation study on inputs and action of the RL agent." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 141, + 285, + 290, + 327 + ], + "lines": [ + { + "bbox": [ + 141, + 285, + 290, + 327 + ], + "spans": [ + { + "bbox": [ + 141, + 285, + 290, + 327 + ], + "type": "table", + "html": "
ComponentsTime [ms]
- 100 Parallel SVO Step73.9
- Load 100 Images18.6
- Network Forward Pass3.1
Complete Rollout Step104.1
", + "image_path": "0849402af912da5390584c7c03c127d0b3562934a4493a34879213a88b80950d.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 324, + 285, + 473, + 335 + ], + "blocks": [ + { + "bbox": [ + 132, + 336, + 299, + 369 + ], + "lines": [ + { + "bbox": [ + 132, + 336, + 299, + 369 + ], + "spans": [ + { + "bbox": [ + 132, + 336, + 299, + 369 + ], + "type": "text", + "content": "Table 6: Rollout Runtime. The runtime for the main components in the rollout phase averaged over 500 iterations." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 324, + 285, + 473, + 335 + ], + "lines": [ + { + "bbox": [ + 324, + 285, + 473, + 335 + ], + "spans": [ + { + "bbox": [ + 324, + 285, + 473, + 335 + ], + "type": "table", + "html": "
MethodTime [ms]
SVO9.06
RL SVO7.40
DSO34.45
RL DSO32.29
Network inference1.75
", + "image_path": "22504a8c765525510eac479e76ecbcf05889f05c28d9ace2fa22d1cbe7ffb655.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 336, + 480, + 369 + ], + "lines": [ + { + "bbox": [ + 313, + 336, + 480, + 369 + ], + "spans": [ + { + "bbox": [ + 313, + 336, + 480, + 369 + ], + "type": "text", + "content": "Table 7: Inference Runtime. The runtime of SVO, DSO, and one network forward pass." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 130, + 404, + 480, + 487 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 404, + 480, + 487 + ], + "spans": [ + { + "bbox": [ + 130, + 404, + 480, + 487 + ], + "type": "text", + "content": "velocity. The keyframe selection using the RL agent has a similar distribution around the motion patterns, while the heuristic rules select more keyframes. Following the motion distribution of the dataset, both methods trigger more keyframes in small motions. Despite the fewer keyframes, our RL agent still achieves a more robust performance. Without the possibility to set keyframes, namely, the agent only determines the feature grid size, robustness is drastically reduced, as it can be seen in Tab. 5 by the entry RL SVO (w/o keyframe)." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 488, + 480, + 536 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 488, + 480, + 536 + ], + "spans": [ + { + "bbox": [ + 130, + 488, + 480, + 536 + ], + "type": "text", + "content": "Grid size Selection The positive impact of dynamically selecting the grid size for SVO can be observed by removing the grid size action, which lowers the pose estimation accuracy, see RL SVO (w/o gridsize) in Tab. 5. The supplementary video shows that the RL agent learns to reduce the grid size to the minimal" + } + ] + } + ], + "index": 11 + }, + { + "type": "table", + "bbox": [ + 135, + 567, + 478, + 629 + ], + "blocks": [ + { + "bbox": [ + 308, + 558, + 362, + 566 + ], + "lines": [ + { + "bbox": [ + 308, + 558, + 362, + 566 + ], + "spans": [ + { + "bbox": [ + 308, + 558, + 362, + 566 + ], + "type": "text", + "content": "EuRoC (ATE [m])" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 135, + 567, + 478, + 629 + ], + "lines": [ + { + "bbox": [ + 135, + 567, + 478, + 629 + ], + "spans": [ + { + "bbox": [ + 135, + 567, + 478, + 629 + ], + "type": "table", + "html": "
MethodMH01MH02MH03MH04MH05V101V102V103V201V202V203Avg
ORB-SLAM30.0310.0320.0411.2560.0660.0351.5281.4510.2010.1901.4690.573
RL ORB-SLAM30.0290.0280.0390.1520.1060.0341.4721.4410.2030.0701.6540.475
TUM-RGBD (ATE [m])
Method360deskdesk2floorplantroomrpyteddyxyzAvg
ORB-SLAM3-0.0310.7460.0130.1541.048-0.1080.0110.302
RL ORB-SLAM3-0.0310.6040.0130.0380.717-0.0910.0110.215
", + "image_path": "0b3b9bc89f1f8d5ab149d9bc58f3ccc889990390966eeb151a405d3f1f516f8d.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_body" + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 630, + 479, + 651 + ], + "lines": [ + { + "bbox": [ + 132, + 630, + 479, + 651 + ], + "spans": [ + { + "bbox": [ + 132, + 630, + 479, + 651 + ], + "type": "text", + "content": "Table 8: ORB-SLAM3. We report the ATE [m] of our RL agent applied on ORBSLAM3." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 250, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 250, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 250, + 91, + 447, + 102 + ], + "type": "text", + "content": "Reinforcement Learning Meets Visual Odometry" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 481, + 402 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 481, + 402 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 481, + 402 + ], + "type": "text", + "content": "size of 20 for challenging conditions, leading to more keypoints being tracked. Another behavior that can be observed is the quick increase of the grid size for one or two frames. A possible explanation for this short increase and subsequent decrease in the grid size is that the agent tries to filter out less robust keypoints. Runtime The runtime achieved by DSO and SVO given decisions from the heuristic rules and our RL agent are reported in Tab. 7. The results are obtained by averaging the runtime required for processing all sequences of EuRoC on one NVIDIA A100. Our RL agent leads to a faster average SVO runtime, which can be related to the lower rate of keyframe selection, see Fig. 4. In comparison, RL DSO does not lead to a high runtime reduction. For the performance of the agents with different penalty terms, we refer to the supplementary. Finally, the architecture of our agent is very lightweight, with a total of 296,584 parameters. SLAM Method In addition to VO methods, we also tested our method with the full SLAM pipeline ORB-SLAM3. Differently from the VO methods, which run in a sliding-window fashion (old measurements are marginalized), ORBSLAM3 uses a local bundle adjustment that has access to the entire history of keyframes. Although the local bundle adjustment could compensate for uninformative keyframes, our agent can still provide a more suitable keyframe selection strategy than the heuristics, as it can be observed in Tab. 8 by the increased accuracy on the EuRoC and TUM-RGBD datasets. To run ORB-SLAM3 with our framework, significant adjustments with respect to the synchronization between the tracking and bundle adjustment threads were required, amongst others, the removal of the loop closure algorithm, which explains the different performance compared to results reported in [3]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 403, + 482, + 499 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 403, + 482, + 499 + ], + "spans": [ + { + "bbox": [ + 130, + 403, + 482, + 499 + ], + "type": "text", + "content": "Ablation To evaluate the benefits of our Variable Encoder processing the keypoints, we report in Tab. 5 the performance of an agent network without the Variable Encoder, i.e., RL SVO (w/o keypoints). As can be observed, the agent without access to the keypoints information is not capable of successfully finishing all the sequences of the EuRoC dataset. Furthermore, we also report the performance of the agent trained without privileged critic referred to as SVO RL (w/o privileged c.), leading to a larger mean ATE. For the ablations on TUM-RGBD, we refer to the supplementary." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 519, + 220, + 533 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 519, + 220, + 533 + ], + "spans": [ + { + "bbox": [ + 132, + 519, + 220, + 533 + ], + "type": "text", + "content": "5 Conclusion" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": "By approaching the Visual Odometry (VO) problem from a sequential decision-making perspective, we employ a neural network trained within a Reinforcement Learning (RL) framework to dynamically adapt the VO method at each timestep. Instead of hand-tuned heuristics and fixed hyperparameters, the dynamic agent predicts optimal actions for the VO method to increase its accuracy and robustness. The RL agent is trained on a large-scale synthetic dataset and tested on common real-world VO benchmarks showing improved robustness and accuracy. We believe this paradigm shift advances VO technology and opens avenues for integrating RL into visual inertial odometry and simultaneous localization and mapping approaches." + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 246, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 246, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 246, + 101 + ], + "type": "text", + "content": "Messikommer et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 240, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 240, + 129 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 240, + 129 + ], + "type": "text", + "content": "Acknowledgments" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 141, + 480, + 166 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 141, + 480, + 166 + ], + "spans": [ + { + "bbox": [ + 132, + 141, + 480, + 166 + ], + "type": "text", + "content": "This work was supported by the European Research Council (ERC) under grant agreement No. 864042 (AGILEFLIGHT)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 186, + 197, + 198 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 186, + 197, + 198 + ], + "spans": [ + { + "bbox": [ + 133, + 186, + 197, + 198 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 134, + 212, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 138, + 212, + 481, + 266 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 212, + 481, + 266 + ], + "spans": [ + { + "bbox": [ + 138, + 212, + 481, + 266 + ], + "type": "text", + "content": "1. Burri, M., Nikolic, J., Gohl, P., Schneider, T., Rehder, J., Omari, S., Achtelik, M.W., Siegwart, R.: The euroc micro aerial vehicle datasets. The International Journal of Robotics Research (2016). https://doi.org/10.1177/0278364915620033, http://ijr.sagepub.com/content/early/2016/01/21/0278364915620033abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 267, + 481, + 311 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 267, + 481, + 311 + ], + "spans": [ + { + "bbox": [ + 138, + 267, + 481, + 311 + ], + "type": "text", + "content": "2. Cadena, C., Carlone, L., Carrillo, H., Latif, Y., Scaramuzza, D., Neira, J., Reid, I., Leonard, J.J.: Past, present, and future of simultaneous localization and mapping: Toward the robust-perception age. IEEE Transactions on robotics 32(6), 1309-1332 (2016)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 312, + 481, + 344 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 312, + 481, + 344 + ], + "spans": [ + { + "bbox": [ + 138, + 312, + 481, + 344 + ], + "type": "text", + "content": "3. Campos, C., Elvira, R., Gomez, J.J., Montiel, J.M.M., Tardos, J.D.: ORB-SLAM3: An accurate open-source library for visual, visual-inertial and multi-map SLAM. IEEE Transactions on Robotics 37(6), 1874-1890 (2021)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 346, + 481, + 378 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 346, + 481, + 378 + ], + "spans": [ + { + "bbox": [ + 138, + 346, + 481, + 378 + ], + "type": "text", + "content": "4. Dong, X., Shen, J., Wang, W., Shao, L., Ling, H., Porikli, F.: Dynamical hyperparameter optimization via deep reinforcement learning in tracking. IEEE transactions on pattern analysis and machine intelligence 43(5), 1515-1529 (2019)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 379, + 481, + 410 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 379, + 481, + 410 + ], + "spans": [ + { + "bbox": [ + 138, + 379, + 481, + 410 + ], + "type": "text", + "content": "5. Engel, J., Koltun, V., Cremers, D.: Direct Sparse Odometry. IEEE Trans. Pattern Anal. Mach. Intell. 40(3), 611-625 (Mar 2018). https://doi.org/10.1109/TPAMI.2017.2658577" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 411, + 481, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 411, + 481, + 443 + ], + "spans": [ + { + "bbox": [ + 138, + 411, + 481, + 443 + ], + "type": "text", + "content": "6. Forster, C., Zhang, Z., Gassner, M., Werlberger, M., Scaramuzza, D.: SVO: Semidirect visual odometry for monocular and multicamera systems. IEEE Trans. Robot. 33(2), 249-265 (2017). https://doi.org/10.1109/TR0.2016.2623335" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 445, + 481, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 445, + 481, + 477 + ], + "spans": [ + { + "bbox": [ + 138, + 445, + 481, + 477 + ], + "type": "text", + "content": "7. Geiger, A., Lenz, P., Urtasun, R.: Are we ready for autonomous driving? the kitti vision benchmark suite. In: Conference on Computer Vision and Pattern Recognition (CVPR) (2012)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 478, + 481, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 478, + 481, + 521 + ], + "spans": [ + { + "bbox": [ + 138, + 478, + 481, + 521 + ], + "type": "text", + "content": "8. Jaegle, A., Borgeaud, S., Alayrac, J.B., Doersch, C., Ionescu, C., Ding, D., Koppula, S., Zoran, D., Brock, A., Shelhamer, E., Henaff, O., Botvinick, M.M., Zisserman, A., Vinyals, O., Carreira, J.: Perceiver io: A general architecture for structured inputs & outputs. Int. Conf. Learn. Representations (ICLR) (2022)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 522, + 481, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 522, + 481, + 555 + ], + "spans": [ + { + "bbox": [ + 138, + 522, + 481, + 555 + ], + "type": "text", + "content": "9. Jaegle, A., Gimeno, F., Brock, A., Zisserman, A., Vinyals, O., Carreira, J.: Perceiver: General perception with iterative attention. Proc. Int. Conf. Mach. Learning (ICML) (2021)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 555, + 481, + 576 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 555, + 481, + 576 + ], + "spans": [ + { + "bbox": [ + 138, + 555, + 481, + 576 + ], + "type": "text", + "content": "0. Jomaa, H.S., Grabocka, J., Schmidt-Thieme, L.: Hyp-rl: Hyperparameter optimization by reinforcement learning. arXiv preprint arXiv:1906.11527 (2019)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 577, + 481, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 577, + 481, + 599 + ], + "spans": [ + { + "bbox": [ + 138, + 577, + 481, + 599 + ], + "type": "text", + "content": "1. Mur-Artal, R., Montiel, J.M.M., Tardos, J.D.: Orb-slam: a versatile and accurate monocular slam system. IEEE transactions on robotics 31(5), 1147-1163 (2015)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 134, + 600, + 481, + 642 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 600, + 481, + 642 + ], + "spans": [ + { + "bbox": [ + 134, + 600, + 481, + 642 + ], + "type": "text", + "content": "12. Raffin, A., Hill, A., Gleave, A., Kanervisto, A., Ernestus, M., Dormann, N.: Stable-baselines3: Reliable reinforcement learning implementations. Journal of Machine Learning Research 22(268), 1-8 (2021), http://jmlr.org/papers/v22/20-1364.html" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 134, + 643, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 643, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 134, + 643, + 481, + 665 + ], + "type": "text", + "content": "13. Scaramuzza, D., Fraundorfer, F.: Visual odometry [tutorial]. IEEE robotics & automation magazine 18(4), 80-92 (2011)" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 250, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 250, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 250, + 91, + 447, + 102 + ], + "type": "text", + "content": "Reinforcement Learning Meets Visual Odometry" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 567 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 133, + 116, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 480, + 138 + ], + "type": "text", + "content": "14. Schulman, J., Wolski, F., Dhariwal, P., Radford, A., Klimov, O.: Proximal policy optimization algorithms. arXiv preprint arXiv:1707.06347 (2017)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 139, + 481, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 139, + 481, + 182 + ], + "spans": [ + { + "bbox": [ + 133, + 139, + 481, + 182 + ], + "type": "text", + "content": "15. Sheng, L., Xu, D., Ouyang, W., Wang, X.: Unsupervised collaborative learning of keyframe detection and visual odometry towards monocular deep slam. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 4302-4311 (2019)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 182, + 481, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 182, + 481, + 215 + ], + "spans": [ + { + "bbox": [ + 132, + 182, + 481, + 215 + ], + "type": "text", + "content": "16. Sturm, J., Engelhard, N., Endres, F., Burgard, W., Cremers, D.: A benchmark for the evaluation of rgb-d slam systems. In: Proc. of the International Conference on Intelligent Robot Systems (IROS) (Oct 2012)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 216, + 481, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 216, + 481, + 248 + ], + "spans": [ + { + "bbox": [ + 132, + 216, + 481, + 248 + ], + "type": "text", + "content": "17. Teed, Z., Deng, J.: Raft: Recurrent all-pairs field transforms for optical flow. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part II 16. pp. 402-419. Springer (2020)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 249, + 481, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 249, + 481, + 281 + ], + "spans": [ + { + "bbox": [ + 132, + 249, + 481, + 281 + ], + "type": "text", + "content": "18. Teed, Z., Deng, J.: Droid-slam: Deep visual slam for monocular, stereo, and rgb-d cameras. Advances in neural information processing systems 34, 16558-16569 (2021)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 281, + 481, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 281, + 481, + 304 + ], + "spans": [ + { + "bbox": [ + 132, + 281, + 481, + 304 + ], + "type": "text", + "content": "19. Teed, Z., Lipson, L., Deng, J.: Deep patch visual odometry. arXiv preprint arXiv:2208.04726 (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 304, + 481, + 325 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 304, + 481, + 325 + ], + "spans": [ + { + "bbox": [ + 132, + 304, + 481, + 325 + ], + "type": "text", + "content": "20. Umeyama, S.: Least-squares estimation of transformation parameters between two point patterns. IEEE Trans. Pattern Anal. Mach. Intell. 13(4) (1991)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 325, + 481, + 390 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 325, + 481, + 390 + ], + "spans": [ + { + "bbox": [ + 132, + 325, + 481, + 390 + ], + "type": "text", + "content": "21. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L.u., Polosukhin, I.: Attention is all you need. In: Guyon, I., Luxburg, U.V., Bengio, S., Wallach, H., Fergus, R., Vishwanathan, S., Garnett, R. (eds.) Advances in Neural Information Processing Systems. vol. 30. Curran Associates, Inc. (2017), https://proceedings.neurips.cc/paper/2017/file/3f5ee243547dee91fbd053c1c4a845aa-Paper.pdf" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 391, + 481, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 391, + 481, + 413 + ], + "spans": [ + { + "bbox": [ + 132, + 391, + 481, + 413 + ], + "type": "text", + "content": "22. Vogiatzis, G., Hernández, C.: Video-based, real-time multi-view stereo. Image and Vision Computing 29(7), 434-441 (2011)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 413, + 481, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 413, + 481, + 445 + ], + "spans": [ + { + "bbox": [ + 132, + 413, + 481, + 445 + ], + "type": "text", + "content": "23. Wang, S., Clark, R., Wen, H., Trigoni, N.: Deepvo: Towards end-to-end visual odometry with deep recurrent convolutional neural networks. In: IEEE Int. Conf. Robot. Autom. (ICRA). pp. 2043-2050. IEEE (2017)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 446, + 481, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 446, + 481, + 468 + ], + "spans": [ + { + "bbox": [ + 132, + 446, + 481, + 468 + ], + "type": "text", + "content": "24. Wang, W., Hu, Y., Scherer, S.: Tartanvo: A generalizable learning-based vo. In: Conference on Robot Learning. pp. 1761-1772. PMLR (2021)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 468, + 481, + 500 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 468, + 481, + 500 + ], + "spans": [ + { + "bbox": [ + 132, + 468, + 481, + 500 + ], + "type": "text", + "content": "25. Wang, W., Zhu, D., Wang, X., Hu, Y., Qiu, Y., Wang, C., Hu, Y., Kapoor, A., Scherer, S.: Tartanair: A dataset to push the limits of visual slam. 2020 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) (2020)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 501, + 481, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 501, + 481, + 533 + ], + "spans": [ + { + "bbox": [ + 132, + 501, + 481, + 533 + ], + "type": "text", + "content": "26. Ye, W., Lan, X., Chen, S., Ming, Y., Yu, X., Bao, H., Cui, Z., Zhang, G.: Pvo: Panoptic visual odometry. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9579-9589 (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 534, + 481, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 534, + 481, + 567 + ], + "spans": [ + { + "bbox": [ + 132, + 534, + 481, + 567 + ], + "type": "text", + "content": "27. Zhang, Z., Scaramuzza, D.: A tutorial on quantitative trajectory evaluation for visual (-inertial) odometry. In: IEEE/RSJ Int. Conf. Intell. Robot. Syst. (IROS). pp. 7244-7251. IEEE (2018)" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 246, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 246, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 246, + 100 + ], + "type": "text", + "content": "Messikommer et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Reinforcement Learning via Auxillary Task Distillation/0d1f16d8-0d98-408f-b999-123ebb200f38_content_list.json b/2024/Reinforcement Learning via Auxillary Task Distillation/0d1f16d8-0d98-408f-b999-123ebb200f38_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..7729e4dc98e9f6910d4ce1331823174074c80b68 --- /dev/null +++ b/2024/Reinforcement Learning via Auxillary Task Distillation/0d1f16d8-0d98-408f-b999-123ebb200f38_content_list.json @@ -0,0 +1,1398 @@ +[ + { + "type": "text", + "text": "Reinforcement Learning via Auxiliary Task Distillation", + "text_level": 1, + "bbox": [ + 255, + 171, + 749, + 211 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abhinav Narayan Harish1, Larry Heck2, Josiah P. Hanna1 \nZsolt Kira2, Andrew Szot2", + "bbox": [ + 290, + 236, + 709, + 263 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 University of Wisconsin-Madison \n2 Georgia-Tech", + "bbox": [ + 381, + 274, + 616, + 300 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. We present Reinforcement Learning via Auxiliary Task Distillation (AuxDistill), a new method that enables reinforcement learning (RL) to perform long-horizon robot control problems by distilling behaviors from auxiliary RL tasks. AuxDistill achieves this by concurrently carrying out multi-task RL with auxiliary tasks, which are easier to learn and relevant to the main task. A weighted distillation loss transfers behaviors from these auxiliary tasks to solve the main task. We demonstrate that AuxDistill can learn a pixels-to-actions policy for a challenging multi-stage embodied object rearrangement task from the environment reward without demonstrations, a learning curriculum, or pre-trained skills. AuxDistill achieves $2.3 \\times$ higher success than the previous state-of-the-art baseline in the Habitat Object Rearrangement benchmark and outperforms methods that use pre-trained skills and expert demonstrations.", + "bbox": [ + 260, + 332, + 742, + 511 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 216, + 533, + 376, + 548 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "While reinforcement learning (RL) is successful in a variety of settings including games [9, 31, 41], chatbots [1, 29, 32, 40, 45], and robotics [3, 23, 35], long-horizon tasks such as embodied object rearrangement where an embodied agent must rearrange objects to target positions still remains a challenge [7]. Object rearrangement requires learning heterogeneous behaviors like picking, navigating, placing, and opening concurrently with sequencing these behaviors to solve the overall task all while operating from egocentric vision. Furthermore, the behaviors are interdependent, meaning the robot can only learn how to pick an object if it has first learned how to open the fridge containing the object. Since object rearrangement requires interacting with objects across house-scale spaces with low-level control, the episodes consist of thousands of low-level control steps, exacerbating the problem of credit assignment [19, 34]. Overall, the problems of concurrently learning low-level control, high-level decision-making, interdependent task stages, and episodes with many time steps make object rearrangement challenging even with dense rewards. While some prior methods have applied end-to-end RL to rearrangement problems, they require excessive experience in simplified simulators [8] or expert skill demonstrations [22]. Other prior work has made such complex tasks more tractable by decomposing the full task into", + "bbox": [ + 213, + 561, + 788, + 813 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "a hierarchy of separately trained skills that a high-level policy sequences together to solve the overall task. However, such hierarchical methods suffer from compounding errors between skills [17] and dynamically selecting between skills.", + "bbox": [ + 213, + 175, + 786, + 219 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We therefore present Reinforcement Learning via Auxiliary Task Distillation (AuxDistill), a method for training policies for long-horizon tasks from scratch with end-to-end RL from reward alone. AuxDistill also learns in auxiliary tasks using a multi-task RL framework and transfers the knowledge from these auxiliary tasks to help solve the desired \"main\" task of interest. Importantly, AuxDistill concurrently learns the main task and all auxiliary tasks, which are all initialized from scratch and not pre-trained. These auxiliary tasks help because they are easier to learn with RL than the main task and contain relevant behaviors for the full task. For example, in object rearrangement, the agent should also practice picking up objects in isolation from the complexities of the full task. Unlike curriculum learning in RL [4,12,13,33], which also leverages easier tasks to learn harder tasks, AuxDistill does not have distinct curriculum phases. Instead, it has a single training objective for distilling sub-behaviors into the overall task behavior. Crucially, the agent end-to-end learns which auxiliary behaviors are relevant for the main task. This relevance is encoded by a scalar weight value determined by the relevance of the robot state to the auxiliary task and it grounded in the oracle task plan. A distillation objective transfers behaviors from relevant auxiliary tasks by encouraging the policy to act consistently between the main and auxiliary tasks for relevant states in the main task. For instance, in object rearrangement, when the robot is near a drawer with the target object inside, it is supervised with the distillation loss from an \"open drawer\" auxiliary task. Despite the policy also concurrently learning the \"open drawer\" auxiliary task from scratch, AuxDistill learns the auxiliary tasks faster since they are easier. Thus, the distillation loss is a useful dense per-time step supervision signal that helps overcome the challenges of learning from the main task reward alone.", + "bbox": [ + 213, + 228, + 788, + 588 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We empirically demonstrate that AuxDistill outperforms a variety of baselines in terms of success rate on the home rearrangement benchmark in Habitat 2.0 [43]. These include baselines such as hierarchical RL, end-to-end RL with and without a curriculum, and imitation learning. On rearrangement episodes where objects start in open receptacles, AuxDistill achieves $1.4 \\times$ higher success than baselines and $2.6 \\times$ higher success in harder episodes where objects can start in closed receptacles. These results highlight the value of AuxDistill to be able to learn the entire rearrangement task end-to-end with RL. We show results beyond object rearrangement on a category-conditioned object manipulation task where AuxDistill achieves $1.75 \\times$ higher success than baselines. We also conduct extensive ablation studies where we analyze the properties of auxiliary tasks needed by AuxDistill, and find that AuxDistill is generally robust to the choice of auxiliary tasks. An analysis of the distillation objective reveals the importance of the distillation loss, with the absence of the distillation loss leading to no success on the rearrangement task. All code is available at https://github.com/absdnd/aux_distill", + "bbox": [ + 213, + 598, + 789, + 820 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 216, + 148, + 228, + 158 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "Harish et. al", + "bbox": [ + 270, + 148, + 357, + 159 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 174, + 388, + 189 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Some prior works address complex and long-horizon decision-making problems using a hierarchical breakdown of skills. One such category of approaches is option-critic methods [5, 42, 51], which seek to learn options to temporally abstract the high-level policy decision making. However, discovering and learning such options is unstable and results in a challenging credit assignment problem. Another line of work first trains or is given pre-trained skills and then learns a high-level policy to sequence these skills together to complete longer tasks [11, 18, 24, 48]. Using such a hierarchy results in compounding errors resulting from sequencing skills that were not trained to properly transition between each other [10, 17, 28, 30, 46]. Our approach does not utilize a hierarchical policy, and can solve long horizon tasks better than hierarchical methods by utilizing a single policy and thus avoiding compounding errors that emerge by sequencing independently trained policies.", + "bbox": [ + 213, + 205, + 786, + 370 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Like our method, prior works have tackled model-free end-to-end RL to solve long, complex tasks. Scaling PPO [39] training has enabled robots to learn complex locomotion behaviors [2, 15, 25, 26, 36, 38], manipulate unseen objects in a robotic hand [35], and play competitive video games [9]. Likewise, berges2023galactic showed scaling end-to-end RL can learn object rearrangement. However, their approach requires training agents for billions of environment steps with a simplified kinematic physics simulation. We show results in the full dynamic Habitat simulation, achieving better results in an order of magnitude fewer environment interactions, and show results in harder rearrangement tasks. Other works have incorporated auxiliary objectives in RL to boost performance in embodied tasks [16, 27, 49, 50]. These works add auxiliary self-supervised prediction objectives, whereas our method adds an auxiliary policy learning task via multi-task RL. jia2022improving also uses easier auxiliary RL tasks that specialist agents learn to complete and then distill back into a generalist agent. While our work also uses RL in easier tasks to boost performance, it doesn't require alternating between learning specialist and generalist policies.", + "bbox": [ + 213, + 370, + 788, + 592 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Works have also attempted to learn a curriculum to leverage easier tasks to learn harder tasks that require composing multiple behaviors. Some works learn curriculum generation policies that adjust the environment to suit the agent's capabilities [4, 12, 13]. In other works, a curriculum naturally emerges as a result of downstream training [6]. Finally, some works hand design curricula that change aspects of the environment, such as the starting and goal distributions depending on the agent performance during training [38]. Our work shares a similar spirit of leveraging easier auxiliary tasks to learn a harder task but it does not enforce an explicit curriculum and instead learns all tasks simultaneously.", + "bbox": [ + 213, + 592, + 788, + 715 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3 Method", + "text_level": 1, + "bbox": [ + 215, + 738, + 329, + 753 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We present Reinforcement Learning via Auxiliary Task Distillation (AuxDistill), a new method for training long-horizon policies from scratch using rewards alone. Learning from reward alone in complex tasks like embodied object rearrangement", + "bbox": [ + 213, + 769, + 786, + 812 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "AuxDistill", + "bbox": [ + 658, + 148, + 731, + 159 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 148, + 784, + 158 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/0bbc9144d06b05ec8141d6a4d778452164d5b7a3f68cd50eb7f82b46b6b8ae57.jpg", + "image_caption": [ + "Fig. 1: AuxDistill learns a rearrangement policy operating from egocentric depth perception and coordinate-based task specification. The full object rearrangement task decomposes into modular abilities that can be learned by auxiliary task with indicator vectors $T_{1} \\cdots T_{N}$ which are trained along with the main task using end-to-end RL. The distillation loss is computed as a weighted combination of the task relevance of $o_{t}$ in the main task $T_{0}$ under all auxiliary tasks. The task relevance function computes $w_{i}(s_{t})$ based on the relevance of the current observation and robot state to the auxiliary task $T_{i}$ . The distillation loss and RL-training loss are then used to update the policy." + ], + "image_footnote": [], + "bbox": [ + 218, + 174, + 783, + 336 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "is challenging because an agent needs to combine thousands of low-level actions controlling the arm and base, operate from egocentric visual perception, and dynamically sequence distinct behaviors such as navigating, picking, placing, and opening. AuxDistill addresses these challenges of RL in complex problems, like embodied rearrangement, by learning to leverage knowledge from easier auxiliary tasks related to the desired task we are trying to solve. Unlike a curriculum, which learns gradually harder tasks in stages, AuxDistill learns the desired task concurrently with the auxiliary task and includes a novel distillation mechanism to transfer knowledge from easier to harder tasks.", + "bbox": [ + 213, + 483, + 788, + 609 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1 Preliminaries", + "text_level": 1, + "bbox": [ + 215, + 646, + 371, + 659 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our problem is formulated as a goal-specified Partially-Observable Markov Decision Process (POMDP) defined as a tuple $\\mathcal{M} = (\\mathcal{S},\\mathcal{O},\\mathcal{A},\\mathcal{P},\\mathcal{R},\\mathcal{G},\\rho,\\gamma)$ with underlying state space $\\mathcal{S}$ , observation space $\\mathcal{O}$ , action space $\\mathcal{A}$ , transition function $\\mathcal{P}$ , reward function $\\mathcal{R}$ , goal space $\\mathcal{G}$ , initial state distribution $\\rho$ and discount factor $\\gamma$ . For a task like rearrangement, the goal space $\\mathcal{G}$ is specified using the 3D coordinate of the object's start location and the goal location where the object has to be placed. Our objective is to learn a goal-conditioned policy $\\pi(a|o,g)$ mapping an observation $o$ and goal $g$ to an action $a$ that maximizes the sum of discounted rewards $\\mathbb{E}_{s_0\\sim\\rho_0,g\\sim\\mathcal{G}}\\sum_t\\gamma^t\\mathcal{R}(s_t,g)$ .", + "bbox": [ + 213, + 686, + 788, + 813 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 216, + 149, + 228, + 159 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "Harish et. al", + "bbox": [ + 270, + 148, + 357, + 158 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2 Reinforcement Learning via Auxiliary Task Distillation", + "text_level": 1, + "bbox": [ + 213, + 175, + 715, + 191 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The core idea of AuxDistill is to learn a policy in a difficult desired \"main\" task by transferring knowledge from easier auxiliary tasks. This is done without using an explicit curriculum by learning the main and auxiliary tasks in a single training loop. We refer to the main task we wish to learn as $\\mathcal{M}_0$ with a task identifier $T_0$ . We define $N$ auxiliary tasks, defined as $\\{\\mathcal{M}_i\\}_{n=1}^N$ with task identifiers $\\{T_i\\}_{n=1}^N$ . The auxiliary tasks $\\{\\mathcal{M}_i\\}$ share the same state, observation, goal, and action space as the target task $\\mathcal{M}_0$ . Each $\\mathcal{M}_i$ has a separately defined starting state distribution and reward function. We assume these auxiliary tasks are related to the full task yet are easier to solve than the full task. Specifically, the auxiliary tasks can be easier instantiations or sub-parts of the overall task. For example, in rearrangement, we define the auxiliary tasks in terms of interactions the agent needs to complete the entire rearrangement episode, such as picking, placing, and opening. Prior works use similar task definitions to train skills for hierarchical policies in rearrangement [17], but these works suffer from a two-stage pipeline of first needing to separately train each skill and then decide how to combine them. AuxDistill directly learns the full task from scratch while also performing better.", + "bbox": [ + 213, + 200, + 786, + 420 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "AuxDistill learns a single policy with RL that concurrently learns to perform the main and auxiliary tasks. We illustrate the implementation of AuxDistill in Figure 1. This policy, parameterized by $\\theta$ , is expressed as $\\pi_{\\theta}(a_t \\mid o_t, g, T)$ for observation $o_t$ , episode goal $g$ , and per- $\\mathcal{M}$ task identifier $T$ which is encoded as a one-hot embedding in a vector which has the same size as the maximum number of auxiliary tasks. Note that since all tasks share the same observation and action space, $\\pi_{\\theta}$ can act and observe in all tasks based on the input task identifier $T$ .", + "bbox": [ + 213, + 423, + 786, + 519 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "AuxDistill updates the policy based on an average RL loss from the main task and all auxiliary tasks. Let $\\mathcal{L}_{\\mathcal{M}_i}^{\\mathrm{RL}}(\\theta)$ denote the RL loss for $\\pi_{\\theta}$ in MDP $\\mathcal{M}_i$ . These auxiliary tasks are designed to capture a subset of the main task which are easier to accomplish than the main task. To compute these losses, we assume we can collect experience in the auxiliary tasks. For example, to collect experience in a \"pick object\" auxiliary task in object rearrangement, the robot is spawned close to the object. We then update based on the average of the task and auxiliary task losses: $\\frac{1}{N}\\sum_{i=0}^{N}\\mathcal{L}_{\\mathcal{M}_i}^{\\mathrm{RL}}(\\theta)$ . Intuitively, $\\pi_{\\theta}$ will first learn to complete the easier auxiliary tasks, and this auxiliary task competency can aid in solving the main task. This shares a similar insight as curriculum learning, except all tasks are learned concurrently, and the curriculum stages are naturally induced by the policy naturally learning on easier tasks first.", + "bbox": [ + 213, + 520, + 786, + 684 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In addition to optimizing the average RL loss between the main and auxiliary tasks, AuxDistill also optimizes a distillation loss that encourages the policy to transfer relevant knowledge from the auxiliary tasks to the main task. To achieve this, for a particular observation $o_t$ , we consider the policy distribution $\\pi_\\theta(\\cdot | o_t, g, T)$ under different task identifiers $T$ . For compactness, notate $\\pi_\\theta(\\cdot | o_t, g, T_i)$ where $T_i$ indicates the task identifier for $\\mathcal{M}_i$ as $\\pi_\\theta^{T_i}(o_t)$ . We want $\\pi_\\theta^{T_0}$ (the policy in the main task) to match the behaviors from the policy in the relevant auxiliary tasks $\\pi_\\theta^{T_i}$ . We measure this relevance of an auxiliary task $\\mathcal{M}_i$ to the main task at time step $t$ via an auxiliary task relevance function $w_i(s_t)$ .", + "bbox": [ + 213, + 686, + 786, + 810 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "AuxDistill", + "bbox": [ + 658, + 148, + 731, + 159 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 148, + 784, + 158 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "This function denotes how much the knowledge from $\\mathcal{M}_i$ should apply to $\\mathcal{M}_0$ in the underlying simulator state $s_t$ . This relevance function can be grounded in the task plan generated by an oracle planner. For example, consider the state of the robot before picking up the object. In this state, the pick auxiliary task would be relevant to the main task, and the place auxiliary task would not. If an episode requires opening a fridge before accessing the object, the relevant task before the fridge is opened would be open-fridge. Note that computing the distillation weight can utilize oracle knowledge of the simulator state (for instance, whether the object is inside a fridge or a cabinet) since this information is only provided as a training signal and not used during inference. This information is utilized by all methods we compare against, including our strongest baseline [22].", + "bbox": [ + 213, + 175, + 786, + 328 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We can then distill experience from $\\mathcal{M}_i$ by supervising the policy to match the action distribution of the target task $\\mathcal{M}_0$ . The distillation loss is computed as the KL-divergence of the action distribution under $\\mathcal{M}_i$ and $\\mathcal{M}_0$ weighted by the relevance of auxiliary task $\\mathcal{M}_i$ given by $w_i(s_t)$ . The overall loss function of AuxDistill, including the distillation loss, is then:", + "bbox": [ + 213, + 329, + 786, + 398 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\frac {1}{N} \\sum_ {i = 0} ^ {N} \\mathcal {L} _ {\\mathcal {M} _ {i}} ^ {\\mathrm {R L}} (\\theta) + \\lambda \\sum_ {i = 1} ^ {N} w _ {i} \\left(s _ {t}\\right) D _ {\\mathrm {K L}} \\left(\\pi_ {\\theta} ^ {T _ {0}} \\left(o _ {t}\\right) \\parallel \\pi_ {\\theta} ^ {T _ {i}} \\left(o _ {t}\\right)\\right) \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 317, + 409, + 786, + 448 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Here $\\lambda$ represents the distillation weight relative to other PPO losses. Note that our formulation relies on the auxiliary tasks being easier to solve and relevant to the main task. Distillation using the task relevance function offers an additional supervisory signal that the policy can optimize along with its own reward. This doesn't impose the large overhead of hierarchical RL methods where each auxiliary task must be pre-trained in isolation with its own reward function and goal specification. Auxiliary tasks are flexibly incorporated to address parts of the task that are challenging for the agent to learn.", + "bbox": [ + 213, + 457, + 786, + 568 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.3 Implementation Details", + "text_level": 1, + "bbox": [ + 213, + 590, + 455, + 605 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We use PPO [39] to compute the RL loss $\\mathcal{L}_{\\mathcal{M}_i}^{\\mathrm{RL}}$ in Equation (1). We create $M$ environment instances for each of the main and $N$ auxiliary tasks and vectorize them for parallel action execution. We rollout the policy in all these $M(N + 1)$ environments and collect a batch of data. We then compute Equation (1), using the PPO loss, update the policy, and then repeat this process.", + "bbox": [ + 213, + 615, + 786, + 686 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The policy is represented as a 2-layer LSTM network with 512 hidden units per layer, similar to the architecture employed in [43]. In our experiments, $o_t$ is a depth or RGB egocentric image, which we encode with a ResNet50 [20] network. The goal and the proprioceptive state of the agent are concatenated with the visual embeddings along with the index of the POMDP $\\mathcal{M}_i$ , for which we use a one-hot embedding. This vector is then passed as an input to the LSTM. A linear projection on the output of the LSTM then produces an action to execute on the next step in the environment and a vector of value-predictions corresponding to each of the tasks $\\mathcal{M}_i$ used during training.", + "bbox": [ + 213, + 686, + 786, + 810 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 216, + 148, + 228, + 158 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "Harish et. al", + "bbox": [ + 270, + 148, + 357, + 159 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/7ebf77ca445d01bc4dcc365636589526d8c7b127b552421a7167a553470b4394.jpg", + "table_caption": [], + "table_footnote": [ + "Table 1: Success rates on the rearrangement task for our method, AuxDistill (highlighted in blue), and baselines. Displayed are the average and standard deviations for 3 seeds for M3, Monolithic RL, Skill Transformer, and M3 Oracle (numbers from [22]), and 3 seeds for the remaining methods with the highest numbers per setting bolded. Numbers in the easy and hard columns are averages over 100 episodes and 200 episodes, respectively. The All Episodes column is an average across both the Easy and Hard episodes. AuxDistill outperforms all baselines in all settings." + ], + "table_body": "
MethodTrain (Seen)Eval (Unseen)
All EpisodesEasyHardAll EpisodesEasyHard
M3 (Oracle) [17]27 ± 057 ± 212 ± 128 ± 058 ± 213 ± 2
M3 [17]25 ± 156 ± 19 ± 213 ± 153 ± 40 ± 0
Galactic [8]-37 ± 0--26 ± 0-
Monolithic RL0 ± 00 ± 00 ± 00 ± 00 ± 00 ± 0
Skill Transformer [22]25 ± 144 ± 116 ± 123 ± 137 ± 116 ± 1
AuxDistill (No Distillation)0 ± 00 ± 00 ± 00 ± 00 ± 00 ± 0
RL Curriculum0 ± 01 ± 00 ± 00 ± 01 ± 10 ± 0
AuxDistill49 ± 274 ± 236 ± 252 ± 275 ± 241 ± 2
", + "bbox": [ + 233, + 173, + 759, + 306 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Our policy is learned via a multi-task RL formulation, with the main and auxiliary tasks being concurrently learned. Each of the tasks may have different reward magnitudes and relative performance during training. To address this, we use PopArt [21], with $\\beta = 3e^{-4}$ to enable learning across different return scales. Additional AuxDistill details are in Appendix A of the supplementary.", + "bbox": [ + 213, + 436, + 786, + 506 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 213, + 530, + 374, + 546 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.1 Object-Rearrangement", + "text_level": 1, + "bbox": [ + 213, + 562, + 450, + 577 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this section, we compare AuxDistill with baselines on the Habitat 2.0 Object Rearrangement task [43]. For comparison with baselines, we run our experiments using the setup from [22]. In this task, a Fetch robot [37] must move an object from a specified start position to a desired goal position in an indoor home environment using only onboard sensing. The agent has no privileged information like existing maps of the environment, 3D object models, or exact object positions. The Fetch robot senses the world through a $256 \\times 256$ depth camera mounted on the robot's head, robot joint positions, gripper state, and base egomotion, giving the relative position of the robot from the start of the episode. The task is specified by a starting 3D object coordinate for the object to move and a 3D goal coordinate to move the object to. Both coordinates are specified relative to the robot's position at the start of the episode. Only the starting object coordinate is specified, and this information is not updated based on the current object position (e.g. if it is moved).", + "bbox": [ + 213, + 588, + 786, + 781 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The robot interacts with the world via a 7DoF arm, a suction gripper attached to the end of the arm, and a mobile base. The episode is successful if the target", + "bbox": [ + 213, + 783, + 786, + 812 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "AuxDistill", + "bbox": [ + 658, + 148, + 731, + 159 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 148, + 784, + 158 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "object is within $15\\mathrm{cm}$ of the goal position. The robot has a budget of 1,500 steps to complete the task.", + "bbox": [ + 213, + 175, + 783, + 203 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We report performance on the easy and hard evaluation episodes from [22]. In easy episodes, both the target object and goal are on an open receptacle. This means the robot does not have to first open a receptacle before accessing the target object or goal. Instead, the agent can always execute the same sequence of navigate to object, pick up object, navigate to goal, and place object at goal. In hard episodes, the object may start in a closed receptacle. The agent, therefore, needs to use its visual input to perceive if the target object is in a closed receptacle. If so, the robot must then open the receptacle before picking the object. The object may start in either the fridge or a cabinet.", + "bbox": [ + 213, + 204, + 786, + 328 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Training Setup: We train AuxDistill using 11,791 episodes with a mix of easy and hard episodes across 63 scenes using the same rearrangement training dataset as in [44]. These episodes were obtained by sampling an equal number of episodes in the easy and hard categories, with easy episodes being equally sampled across episodes with open-cabinet, open-fridge and non-articulated episodes. For hard episodes, sampling is done uniformly between closed fridge and closed cabinet episodes. We use auxiliary tasks covering abilities included as a part of the standard rearrangement benchmark used in [44]. The auxiliary tasks are:", + "bbox": [ + 213, + 346, + 786, + 456 + ], + "page_idx": 7 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Pick: The agent spawns randomly within the house and must navigate to and pick up the object.", + "- Place the agent is spawned randomly in the house with the object in its gripper. The robot must navigate to and place the object within $15\\mathrm{cm}$ of the target 3D location on the receptacle.", + "- Open Fridge: The robot spawns randomly within the house and must navigate to the fridge and use its base and arm to open the fridge door.", + "- Open Cabinet: The robot spawns randomly in the house and must navigate to the kitchen area and open the cabinet using its base and arm.", + "- Pick from Fridge: This task is similar to Pick except the agent has to pick up the object from inside an open fridge, which requires careful manipulation to minimize collisions." + ], + "bbox": [ + 213, + 456, + 784, + 630 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The auxiliary relevance function is grounded in the oracle task plan [14] of the episode. This is formulated as an indicator function with $w_{i}(s_{t}) = 1$ if the state $s_t$ is relevant to $\\mathcal{M}_i$ based on the oracle task plan of the episode and $w_{i}(s_{t}) = 0$ otherwise.", + "bbox": [ + 213, + 631, + 786, + 684 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "For example, if at step $t$ the agent is not holding an object and the object is in an open receptacle, the agent must first pick up the object so $w_{i_{\\text{pick}}}(s_t) = 1$ , and $w_i(s_t) = 0$ for all other skills. For a complete description of the auxiliary tasks, including their reward functions, along with the precise auxiliary relevance function definition in object rearrangement, see Appendix B. We train methods for 475M steps of environment interactions. We found this to be a sufficient number of environment interactions to ensure sufficient progress on auxiliary tasks for distillation to aid rearrangement performance on both the easy and hard episodes. For AuxDistill, we count the environment interactions in the auxiliary", + "bbox": [ + 213, + 686, + 786, + 810 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 216, + 148, + 228, + 158 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "Harish et. al", + "bbox": [ + 270, + 148, + 357, + 159 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/46dde73ece492b83bd763ae487f041d29624ef921fd4fc83708f857e39b9bc2b.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodTrain (Seen)Eval (Unseen)
All EpisodesEasyHardAll EpisodesEasyHard
All Skills283619303921
No Pick From Fridge2032712232
No Open Fridge414933485739
No Pick000000
", + "bbox": [ + 230, + 173, + 766, + 269 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Table 2: Robustness to auxiliary task selection. We compare the train and generalization performance of AuxDistill with 4 different auxiliary task selections. Results are averages across 100 episodes on a single seed of training. The train and evaluation episodes are in a simplified setting compared to Table 1 as described in Section 4.3", + "bbox": [ + 213, + 269, + 784, + 319 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "tasks towards this 475M step budget. We train with a learning rate of $3e^{-4}$ and linearly decay the learning rate to 0 over the course of learning.", + "bbox": [ + 213, + 369, + 783, + 397 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Baselines We compare AuxDistill to both hierarchical methods governed by a task plan as well as monolithic baselines, which directly learn a pixels-to-actions policy using sensor observations. We compare to relevant baselines from huang2023skill and two more baselines that use the auxiliary tasks.", + "bbox": [ + 213, + 422, + 784, + 476 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Monolithic RL: A monolithic neural network is trained to map sensor observations directly to actions trained with end-to-end RL. The policy showed no signs of learning the main task, so we stopped training early after $100M$ steps of training. This outcome is consistent with the monolithic RL baseline in prior works [8, 43].", + "- Galactic: [8] An end-to-end RL framework similar to monolithic RL to map sensor observations directly to actions with the policy being trained with a simplified kinematic simulation on over $>1e^{9}$ simulation steps and transferred to dynamic simulation utilized in our setting.", + "- M3: [17] Each of the skill policies navigate, pick, place, and open are separately trained. These skills are then sequenced using a task planner.", + "- M3 (Oracle): [17] This uses the same skill training as M3, but uses an oracle task planner.", + "- Skill Transformer [22]: This method uses pre-trained skills to collect successful rearrangement demonstrations and then train on these demonstrations with imitation learning.", + "- RL Curriculum This method is trained in two stages: i) First, we train the policy on the auxiliary tasks for $200M$ steps to ensure the relevant auxiliary tasks have good training performance ii) This is followed by a training on the entire rearrangement task for another $300M$ steps.", + "- AuxDistill: No-Distillation Train AuxDistill under a similar setting as described in Section 4.1 except with the distillation strength $\\lambda = 0.0$ ." + ], + "bbox": [ + 213, + 479, + 784, + 794 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "For more details on the baselines, see Appendix B.2.", + "bbox": [ + 213, + 797, + 591, + 810 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "AuxDistill", + "bbox": [ + 658, + 147, + 731, + 159 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 148, + 784, + 158 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/9eb27ef133401ffcf64610759d9858dfbd0124833233b2fe93d8aff0d916c27c.jpg", + "image_caption": [ + "Fig. 2: Comparison of skill-robustness to different choices of auxiliary task on the hard distribution. Including both the Pick and Pick from Fridge is crucial for successful rearrangement on this distribution. Not utilizing Open-Fridge leads to a boost in rearrangement success. This improvement arises because the open-fridge skill is the easiest of all auxiliary tasks and utilizing it reduces the number of samples for the main task (from the 100M step budget). We discuss the auxiliary task learning curves in Appendix B.3" + ], + "image_footnote": [], + "bbox": [ + 300, + 191, + 685, + 335 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Rearrangement Performance: In Table 1, we compare AuxDistill to baselines in the rearrangement task. On all evaluation settings, AuxDistill outperforms all baselines. AuxDistill shows an absolute improvement of $22\\%$ and $25\\%$ on the easy and hard episodes respectively over the best performing baseline. Most baselines struggle to achieve any performance on the unseen episodes. Monolithic RL [43] achieves no success in any of the evaluation settings, demonstrating the impracticality of learning directly from the dense reward in the full rearrangement task. Our method also outperforms Galactic [8] despite learning on less than half the number of training samples (475M for AuxDistill vs. $>1e^{9}$ for Galactic) on the easy split with $74\\%$ success over $37\\%$ on easy episodes during training and $75\\%$ vs. $26\\%$ on evaluation. Note that Galactic does not report performance on the hard distribution.", + "bbox": [ + 213, + 463, + 786, + 627 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Like AuxDistill, RL Curriculum also utilizes the auxiliary tasks yet finds no success. This demonstrates the value of concurrent training on the main and auxiliary tasks with the distillation loss in AuxDistill. Likewise, AuxDistill (No Distill) achieves no success, illustrating the importance of the distillation loss. The learning curve in Figure 3a as AuxDistill can gradually learn to complete the task with more environment interactions. On the other hand, RL Curriculum and AuxDistill (No-Distill) remain at no success regardless of the number of learning samples.", + "bbox": [ + 213, + 630, + 786, + 740 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "AuxDistill also outperforms the strongest baseline, Skill Transformer, by a significant margin with $75\\%$ vs. $37\\%$ on the unseen easy episodes and $41\\%$ vs. $16\\%$ on the unseen hard episodes. This demonstrates the advantages of using online, end-to-end RL, rather than solely training offline with demonstrations. As shown in Figure 3a, AuxDistill is able to improve with subsequent environment", + "bbox": [ + 213, + 741, + 786, + 810 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 216, + 148, + 235, + 158 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "Harish et. al", + "bbox": [ + 270, + 148, + 357, + 158 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/953050dab1c464a1ce6a03b8ab1372bcd711f71f1b4f12a3b958cf138f4c6b5e.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Method Train (Seen) Eval (Unseen)
AuxDistill12± 211± 1
Monolithic8± 09± 1
AuxDistill (No Distillation)3± 14± 1
RL-Curriculum6± 18± 0
", + "bbox": [ + 315, + 173, + 647, + 234 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Table 3: Comparison of our method on the Category Pick task with auxiliary tasks on three random seeds. AuxDistill outperforms all methods, including the monolithic baseline, on the unseen evaluation split. Evaluation is conducted over 1000 seen episodes sampled from the training distribution and 1000 held-out episodes from evaluation.", + "bbox": [ + 213, + 235, + 784, + 285 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "interactions, yet Skill Transformer is limited by the performance of the expert demonstrations.", + "bbox": [ + 213, + 324, + 783, + 350 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Table 1 also demonstrates that AuxDistill outperforms hierarchical baselines. M3 is able to achieve $53\\%$ success on the unseen easy episodes since it utilizes strong pretrained skills. Even in this setting, AuxDistill achieves a higher success of $75\\%$ . However, on the hard episodes, M3 cannot dynamically plan skills and achieves no success in the hard unseen episodes. AuxDistill achieves $41\\%$ on this setting because it learns a monolithic policy with RL that combines low-level and high-level decision-making. AuxDistill even outperforms an oracle version of M3 that dynamically plans the skill sequence based on oracle state information (13% vs. 41% on the unseen hard episodes).", + "bbox": [ + 213, + 351, + 784, + 475 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "We observe that AuxDistill performs $7\\%$ worse on the Habitat 2.0 rearrangement training episodes than the evaluation episodes. We find this performance gap is due to additional easier episodes in the evaluation distribution where the object is closer to the target location. In the hard split, there are 55 eval episodes and 29 train episodes where the object is picked up from the fridge and the target location is $< 0.5m$ from the object location. On removing these episodes, train performance is higher than evaluation, with $32\\%$ success on train and $31\\%$ success on evaluation.", + "bbox": [ + 213, + 476, + 786, + 586 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4.2 Category Pick", + "text_level": 1, + "bbox": [ + 213, + 607, + 379, + 621 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "The merits of AuxDistill extend to other challenging embodied tasks. In particular, consider a variant of the pick task described in Section 4.1 where a robot has to pick up an object using the object name passed as a one-hot embedding to the policy. Category pick requires the policy to discern the object to pick up by correlating the RGB observation with the object category passed as a one-hot embedding. This task is more challenging than geometric pick where the policy has access to the initial coordinate specification of the object to be picked. We include additional details about the task specification in Section B.4 of our supplementary material.", + "bbox": [ + 213, + 631, + 784, + 753 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "We leverage the easier coordinate pick task to aid the learning of the Category Pick task using AuxDistill. More specifically, we consider the task of interest $M_0$ to be the Category Pick task and $M_1$ (auxiliary task we would like to distill from) to be the coordinate pick task. Both tasks are trained jointly using AuxDistill", + "bbox": [ + 213, + 755, + 786, + 810 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "AuxDistill", + "bbox": [ + 658, + 147, + 731, + 159 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 768, + 148, + 783, + 158 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/0a95d95ec630ec17738dbe92b8640ed87facb1651b4a55af1dd8e808e9704d23.jpg", + "image_caption": [ + "(a) Rearrange learning curve." + ], + "image_footnote": [], + "bbox": [ + 245, + 186, + 475, + 310 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/97823b0f54d475f18f16446b921070e36c9ddb6bc7ed0a89da9d79b090b2be24.jpg", + "image_caption": [ + "(b) Category Pick learning.", + "Fig. 3: Left: RL training success rates on training episodes on the rearrangement task from Table 1. Note that AuxDistill (No Distill) and RL-Curriculum are displayed but achieve $0\\%$ success throughout training. Right: Learning curve on the Category Pick task of AuxDistill utilizing coordinate pick distillation vs. monolithic RL. AuxDistill outperforms baselines in both settings. Displayed are averages and standard deviation over 3 random seeds." + ], + "image_footnote": [], + "bbox": [ + 500, + 186, + 737, + 311 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "with the same reward structure with $w_{i}(s_{t}) = 1$ and use $\\lambda = 0.5$ for distillation in Equation (1). For this task, we train on the full rearrange-easy dataset from [44] with 50,000 episodes.", + "bbox": [ + 213, + 436, + 783, + 476 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "We report the training curves of the Category Pick task in comparison with the RL Curriculum and No Distillation baseline used in Table 1. In addition, we introduce a monolithic baseline for this task, which directly trains Category Pick without leveraging the coordinate pick task during training. We report the training curves in Figure 3b. We evaluate the performance of the trained policy on 1000 episodes sampled from the training and 1000 episodes from the validation split.", + "bbox": [ + 213, + 478, + 786, + 574 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "In Table 3, we show that AuxDistill shows better performance with a success rate of $11\\%$ outperforming all baselines that do not leverage distillation during RL training. The closest performing baseline to ours is the monolithic baseline, which achieves a success of $9\\%$ on the held-out distribution. AuxDistill (No-Distillation) and RL-curriculum perform worse with an evaluation success of $4\\%$ and $8\\%$ , respectively.", + "bbox": [ + 213, + 575, + 786, + 658 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "4.3 AuxDistill Analysis", + "text_level": 1, + "bbox": [ + 213, + 677, + 423, + 692 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Robustness to Auxiliary Task Selection A critical consideration for training AuxDistill is the selection of auxiliary tasks. In this section, we compare four different selections of auxiliary tasks in the object rearrangement task from Section 4.1. We conduct this ablation on a smaller distribution of episodes, which only include the Easy episodes and Hard episodes where the object starts in the fridge and train AuxDistill for 100M steps.", + "bbox": [ + 213, + 700, + 786, + 782 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "As we conduct this study only on the fridge category of articulated episodes, we modify the auxiliary task selection to be: Pick, Open Fridge, Place and Pick", + "bbox": [ + 213, + 783, + 786, + 810 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 216, + 148, + 236, + 158 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "Harish et. al", + "bbox": [ + 270, + 148, + 357, + 158 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "from Fridge. Note that the first three auxiliary tasks are the same from Section 4.1. The other auxiliary task selections are the same as this original selection, but excluding one of: Pick, Open Fridge, and Pick from Fridge.", + "bbox": [ + 213, + 175, + 784, + 218 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "In Table 2, we report the performance for each skill selection on the easy and hard episodes. We observe that among all the sub-tasks, the most important is Pick, as excluding it results in $0.0\\%$ success. Note that Pick is only relevant $(w_{i}(s_{t}) > 0)$ to easy episodes during training i.e. removing it should only impact the performance on easy distribution of our training setup. However, we find that it affects both the easy and hard episodes as the task is not successfully carried out in either case if removed. In Figure 2, we analyze the success of individual stages of rearrangement and notice that not including Pick leads to failure earlier on in the task.", + "bbox": [ + 213, + 218, + 786, + 340 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "In Table 5 we notice that introducing the Open-Fridge skill can worsen rearrangement performance (30% vs. 48%). The reason for this is that the open-fridge skill is the easiest of all auxiliary tasks (see Appendix B.3), and utilizing it reduces the number of samples allocated to the main task. However, certain auxiliary tasks can boost performance by addressing specific bottlenecks in the rearrangement. One such example is the Pick from Fridge, which boosts performance 7% to 19% by addressing the task of picking from an open fridge, which can be challenging due to the difficulty of avoiding collisions with the fridge door. Further, the 4% performance improvement from 32% to 36% on easy distribution can be attributed to the presence of easy episodes in our dataset with an open-refrigerator where Pick from Fridge helps boost performance.", + "bbox": [ + 213, + 342, + 786, + 494 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Distillation Coefficient Variation We analyze the effect of the distillation loss coefficient $\\lambda$ in AuxDistill in Table 4. The distillation loss coefficient controls how the agent balances minimizing the distillation from the auxiliary tasks versus maximizing the reward on the main task. In Table 4, we notice that using high values of distillation weighting $\\lambda = 1.0$ shifts the objective of the rearrangement task from maximizing cumulative reward to distilling from the auxiliary task leading to $1\\%$ success rate on rearrangement and similarly, using a very small value of distillation coefficient $\\lambda = 0.01$ leads to insufficient leveraging of auxiliary task information with a success rate of $7\\%$ . We find $\\lambda = 0.1, 0.05$ to be good choices for optimizing task reward and distilling behaviors from the auxiliary tasks.", + "bbox": [ + 213, + 513, + 786, + 665 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "5 Conclusion", + "text_level": 1, + "bbox": [ + 213, + 686, + 359, + 700 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "In this work, we presented AuxDistill, a new method for end-to-end RL on complex tasks by leveraging auxiliary tasks. AuxDistill learns in the auxiliary tasks concurrently with the main task through multi-task RL. A distillation loss transfers relevant behaviors from the auxiliary tasks to the main task. We show that AuxDistill outperforms a variety of baselines in Habitat object rearrangement by up to $27\\%$ . We also show another application of AuxDistill in a category-conditioned manipulation task. Finally, we analyze AuxDistill in", + "bbox": [ + 213, + 713, + 786, + 810 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "AuxDistill", + "bbox": [ + 658, + 148, + 731, + 159 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 768, + 148, + 784, + 158 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/205a116127971f7e5b3b05e8445b1a61054e8624ddbb52618700ae9e7e3c68af.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodTrain (Seen)Eval (Unseen)
All EpisodesEasyHardAll EpisodesEasyHard
λ = 0.0166510182
λ = 0.05303128363338
λ = 0.1222619303921
λ = 0.5230250
λ = 1.0010120
", + "bbox": [ + 289, + 173, + 700, + 275 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Table 4: Performance of our method for a single seed on varying distillation coefficient during training. Using a large distillation coefficient $(\\lambda = 1.0)$ makes reward optimization challenging, and too small $(\\lambda = 0.01)$ results in insufficient auxiliary distillation to succeed on the rearrangement task. The intermediate values $\\lambda = 0.05$ and $\\lambda = 0.1$ show the best performance on training and evaluation.", + "bbox": [ + 213, + 277, + 786, + 340 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "different auxiliary task selections and magnitudes of distillation strengths. Overall, AuxDistill presents a new way to tackle compound tasks with RL alone without requiring pre-trained skills or expert demonstrations.", + "bbox": [ + 213, + 379, + 784, + 420 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Limitations of AuxDistill include the need for the auxiliary tasks and the auxiliary task relevance function. The auxiliary tasks require knowing behaviors that are relevant and easier to learn than the main task. Each auxiliary task requires defining a new start state distribution and reward function. However, designing these can rely on privileged state information which is less restrictive than the requirement of pre-trained skills or expert demonstrations.", + "bbox": [ + 213, + 420, + 786, + 504 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 215, + 526, + 324, + 539 + ], + "page_idx": 13 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Achiam, J., Adler, S., Agarwal, S., Ahmad, L., Akkaya, I., Aleman, F.L., Almeida, D., Altenschmidt, J., Altman, S., Anadkat, S., et al.: Gpt-4 technical report. arXiv preprint arXiv:2303.08774 (2023)", + "2. Agarwal, A., Kumar, A., Malik, J., Pathak, D.: Legged locomotion in challenging terrains using egocentric vision. In: Conference on Robot Learning. pp. 403-415. PMLR (2023)", + "3. Akkaya, I., Andrychowicz, M., Chociej, M., Litwin, M., McGrew, B., Petron, A., Paino, A., Plappert, M., Powell, G., Ribas, R., et al.: Solving rubik's cube with a robot hand. arXiv preprint arXiv:1910.07113 (2019)", + "4. Azad, A.S., Gur, I., Emhoff, J., Alexis, N., Faust, A., Abbeel, P., Stoica, I.: Clutr: Curriculum learning via unsupervised task representation learning. In: International Conference on Machine Learning. pp. 1361-1395. PMLR (2023)", + "5. Bacon, P.L., Harb, J., Precup, D.: The option-critic architecture. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 31 (2017)", + "6. Baker, B., Kanitscheider, I., Markov, T., Wu, Y., Powell, G., McGrew, B., Mordatch, I.: Emergent tool use from multi-agent autocurricula. arXiv preprint arXiv:1909.07528 (2019)", + "7. Batra, D., Chang, A.X., Chernova, S., Davison, A.J., Deng, J., Koltun, V., Levine, S., Malik, J., Mordatch, I., Mottaghi, R., et al.: Rearrangement: A challenge for embodied ai. arXiv preprint arXiv:2011.01975 (2020)" + ], + "bbox": [ + 221, + 555, + 786, + 810 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 216, + 148, + 235, + 158 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "Harish et. al", + "bbox": [ + 270, + 148, + 357, + 158 + ], + "page_idx": 13 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "8. Berges, V.P., Szot, A., Chaplot, D.S., Gokaslan, A., Mottaghi, R., Batra, D., Undersander, E.: Galactic: Scaling end-to-end reinforcement learning for rearrangement at 100k steps-per-second. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13767-13777 (2023)", + "9. Berner, C., Brockman, G., Chan, B., Cheung, V., Debiak, P., Dennison, C., Farhi, D., Fischer, Q., Hashme, S., Hesse, C., et al.: Dota 2 with large scale deep reinforcement learning. arXiv preprint arXiv:1912.06680 (2019)", + "10. Chen, Y., Wang, C., Fei-Fei, L., Liu, C.K.: Sequential dexterity: Chaining dexterous policies for long-horizon manipulation. arXiv preprint arXiv:2309.00987 (2023)", + "11. Dalal, M., Pathak, D., Salakhutdinov, R.R.: Accelerating robotic reinforcement learning via parameterized action primitives. Advances in Neural Information Processing Systems 34, 21847-21859 (2021)", + "12. Dennis, M., Jaques, N., Vinitsky, E., Bayen, A., Russell, S., Critch, A., Levine, S.: Emergent complexity and zero-shot transfer via unsupervised environment design. Advances in neural information processing systems 33, 13049-13061 (2020)", + "13. Fang, K., Migimatsu, T., Mandlekar, A., Fei-Fei, L., Bohg, J.: Active task randomization: Learning visuomotor skills for sequential manipulation by proposing feasible and novel tasks. arXiv preprint arXiv:2211.06134 (2022)", + "14. Fikes, R.E., Nilsson, N.J.: Strips: A new approach to the application of theorem proving to problem solving. Artificial intelligence 2(3-4), 189-208 (1971)", + "15. Fu, Z., Cheng, X., Pathak, D.: Deep whole-body control: learning a unified policy for manipulation and locomotion. In: Conference on Robot Learning. pp. 138-149. PMLR (2023)", + "16. Gregor, K., Jimenez Rezende, D., Besse, F., Wu, Y., Merzic, H., van den Oord, A.: Shaping belief states with generative environment models for rl. Advances in Neural Information Processing Systems 32 (2019)", + "17. Gu, J., Chaplot, D.S., Su, H., Malik, J.: Multi-skill mobile manipulation for object rearrangement. arXiv preprint arXiv:2209.02778 (2022)", + "18. Hafner, D., Lee, K.H., Fischer, I., Abbeel, P.: Deep hierarchical planning from pixels. Advances in Neural Information Processing Systems 35, 26091-26104 (2022)", + "19. Harutyunyan, A., Dabney, W., Mesnard, T., Gheshlaghi Azar, M., Piot, B., Heess, N., van Hasselt, H.P., Wayne, G., Singh, S., Precup, D., et al.: Hindsight credit assignment. Advances in neural information processing systems 32 (2019)", + "20. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. CVPR (2016)", + "21. Hessel, M., Soyer, H., Espeholt, L., Czarnecki, W., Schmitt, S., Van Hasselt, H.: Multi-task deep reinforcement learning with popart. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 33, pp. 3796-3803 (2019)", + "22. Huang, X., Batra, D., Rai, A., Szot, A.: Skill transformer: A monolithic policy for mobile manipulation. arXiv preprint arXiv:2308.09873 (2023)", + "23. Kalashnikov, D., Irpan, A., Pastor, P., Ibarz, J., Herzog, A., Jang, E., Quillen, D., Holly, E., Kalakrishnan, M., Vanhoucke, V., Levine, S.: Scalable deep reinforcement learning for vision-based robotic manipulation. In: 2nd Annual Conference on Robot Learning, CoRL 2018, Zürich, Switzerland, 29-31 October 2018, Proceedings. Proceedings of Machine Learning Research, vol. 87, pp. 651-673. PMLR (2018), http://proceedings.mlr.press/v87 kalashnikov18a.html", + "24. Karkus, P., Mirza, M., Guez, A., Jaegle, A., Lillicrap, T., Buesing, L., Heess, N., Weber, T.: Beyond tabula-rama: a modular reinforcement learning approach for physically embedded 3d sukoban. arXiv preprint arXiv:2010.01298 (2020)", + "25. Katara, P., Xian, Z., Fragkiadaki, K.: Gen2sim: Scaling up robot learning in simulation with generative models. arXiv preprint arXiv:2310.18308 (2023)" + ], + "bbox": [ + 216, + 177, + 786, + 810 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "AuxDistill", + "bbox": [ + 658, + 147, + 731, + 159 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 768, + 148, + 784, + 158 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "26. Kumar, A., Fu, Z., Pathak, D., Malik, J.: Rapid motor adaptation for legged robots. RSS (2021)", + "27. Kuo, C.W., Ma, C.Y., Hoffman, J., Kira, Z.: Structure-encoding auxiliary tasks for improved visual representation in vision-and-language navigation. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 1104-1113 (2023)", + "28. Lee, Y., Lim, J.J., Anandkumar, A., Zhu, Y.: Adversarial skill chaining for long-horizon robot manipulation via terminal state regularization. arXiv preprint arXiv:2111.07999 (2021)", + "29. Liu, B., Tur, G., Hakkani-Tur, D., Shah, P., Heck, L.: End-to-end optimization of task-oriented dialogue model with deep reinforcement learning. arXiv preprint arXiv:1711.10712 (2017)", + "30. Mishra, U.A., Xue, S., Chen, Y., Xu, D.: Generative skill chaining: Long-horizon skill planning with diffusion models. In: Conference on Robot Learning. pp. 2905-2925. PMLR (2023)", + "31. Mnih, V., Kavukcuoglu, K., Silver, D., Graves, A., Antonoglou, I., Wierstra, D., Riedmiller, M.: Playing atari with deep reinforcement learning. arXiv preprint arXiv:1312.5602 (2013)", + "32. Nakano, R., Hilton, J., Balaji, S., Wu, J., Ouyang, L., Kim, C., Hesse, C., Jain, S., Kosaraju, V., Saunders, W., et al.: Webgpt: Browser-assisted question-answering with human feedback. arXiv preprint arXiv:2112.09332 (2021)", + "33. Narvekar, S., Peng, B., Leonetti, M., Sinapov, J., Taylor, M.E., Stone, P.: Curriculum learning for reinforcement learning domains: A framework and survey. The Journal of Machine Learning Research 21(1), 7382-7431 (2020)", + "34. Ni, T., Ma, M., Eysenbach, B., Bacon, P.L.: When do transformers shine in rl? decoupling memory from credit assignment. Advances in Neural Information Processing Systems 36 (2024)", + "35. Qi, H., Kumar, A., Calandra, R., Ma, Y., Malik, J.: In-hand object rotation via rapid motor adaptation. In: Conference on Robot Learning. pp. 1722-1732. PMLR (2023)", + "36. Radosavovic, I., Xiao, T., Zhang, B., Darrell, T., Malik, J., Sreenath, K.: Learning humanoid locomotion with transformers. arXiv preprint arXiv:2303.03381 (2023)", + "37. robotics, F.: Fetch. http://fetchrobotics.com/ (2020)", + "38. Rudin, N., Hoeller, D., Reist, P., Hutter, M.: Learning to walk in minutes using massively parallel deep reinforcement learning. In: Conference on Robot Learning pp. 91-100. PMLR (2022)", + "39. Schulman, J., Wolski, F., Dhariwal, P., Radford, A., Klimov, O.: Proximal policy optimization algorithms. arXiv preprint arXiv:1707.06347 (2017)", + "40. Shah, P., Hakkani-Tür, D., Heck, L.: Interactive reinforcement learning for task-oriented dialogue management. In: NIPS 2016 Deep Learning for Action and Interaction Workshop. vol. 11 (2016)", + "41. Silver, D., Schrittwieser, J., Simonyan, K., Antonoglou, I., Huang, A., Guez, A., Hubert, T., Baker, L., Lai, M., Bolton, A., et al.: Mastering the game of go without human knowledge. nature 550(7676), 354-359 (2017)", + "42. Sutton, R.S., Precup, D., Singh, S.: Between mdps and semi-mdps: A framework for temporal abstraction in reinforcement learning. Artificial intelligence 112(1-2), 181-211 (1999)", + "43. Szot, A., Clegg, A., Undersander, E., Wijmans, E., Zhao, Y., Turner, J., Maestre, N., Mukadam, M., Chaplot, D.S., Maksymets, O., et al.: Habitat 2.0: Training home assistants to rearrange their habitat. Advances in Neural Information Processing Systems 34 (2021)" + ], + "bbox": [ + 216, + 177, + 786, + 810 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 216, + 148, + 236, + 158 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "Harish et. al", + "bbox": [ + 270, + 148, + 357, + 158 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "44. Szot, A., Yadav, K., Clegg, A., Berges, V.P., Gokaslan, A., Chang, A., Savva, M., Kira, Z., Batra, D.: Habitat rearrangement challenge 2022. https://aihabitat.org/challenge/2022_rearrange (2022)", + "45. Touvron, H., Lavril, T., Izacard, G., Martinet, X., Lachaux, M.A., Lacroix, T., Rozière, B., Goyal, N., Hambro, E., Azhar, F., et al.: Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971 (2023)", + "46. Vezzani, G., Tirumala, D., Wulfmeier, M., Rao, D., Abdelmaleki, A., Moran, B., Haarnoja, T., Humplik, J., Hafner, R., Neunert, M., et al.: Skills: Adaptive skill sequencing for efficient temporally-extended exploration. arXiv preprint arXiv:2211.13743 (2022)", + "47. Wijmans, E., Kadian, A., Morcos, A., Lee, S., Essa, I., Parikh, D., Savva, M., Batra, D.: Dd-ppo: Learning near-perfect pointgoal navigators from 2.5 billion frames. arXiv preprint arXiv:1911.00357 (2019)", + "48. Xia, F., Li, C., Martin-Martin, R., Litany, O., Toshev, A., Savarese, S.: Relmogen: Leveraging motion generation in reinforcement learning for mobile manipulation. arXiv preprint arXiv:2008.07792 (2020)", + "49. Ye, J., Batra, D., Das, A., Wijmans, E.: Auxiliary tasks and exploration enable objectgoal navigation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 16117-16126 (2021)", + "50. Ye, J., Batra, D., Wijmans, E., Das, A.: Auxiliary tasks speed up learning point goal navigation. In: Conference on Robot Learning. pp. 498-516. PMLR (2021)", + "51. Zhang, S., Whiteson, S.: Dac: The double actor-critic architecture for learning options. Advances in Neural Information Processing Systems 32 (2019)" + ], + "bbox": [ + 213, + 175, + 786, + 468 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "AuxDistill", + "bbox": [ + 660, + 148, + 731, + 159 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 768, + 148, + 784, + 158 + ], + "page_idx": 16 + } +] \ No newline at end of file diff --git a/2024/Reinforcement Learning via Auxillary Task Distillation/0d1f16d8-0d98-408f-b999-123ebb200f38_model.json b/2024/Reinforcement Learning via Auxillary Task Distillation/0d1f16d8-0d98-408f-b999-123ebb200f38_model.json new file mode 100644 index 0000000000000000000000000000000000000000..8f2862707da06a7bd444a6452fcf7520dbe99bb6 --- /dev/null +++ b/2024/Reinforcement Learning via Auxillary Task Distillation/0d1f16d8-0d98-408f-b999-123ebb200f38_model.json @@ -0,0 +1,2071 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.257, + 0.172, + 0.75, + 0.212 + ], + "angle": 0, + "content": "Reinforcement Learning via Auxiliary Task Distillation" + }, + { + "type": "text", + "bbox": [ + 0.292, + 0.237, + 0.71, + 0.265 + ], + "angle": 0, + "content": "Abhinav Narayan Harish1, Larry Heck2, Josiah P. Hanna1 \nZsolt Kira2, Andrew Szot2" + }, + { + "type": "text", + "bbox": [ + 0.383, + 0.275, + 0.618, + 0.302 + ], + "angle": 0, + "content": "1 University of Wisconsin-Madison \n2 Georgia-Tech" + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.334, + 0.744, + 0.512 + ], + "angle": 0, + "content": "Abstract. We present Reinforcement Learning via Auxiliary Task Distillation (AuxDistill), a new method that enables reinforcement learning (RL) to perform long-horizon robot control problems by distilling behaviors from auxiliary RL tasks. AuxDistill achieves this by concurrently carrying out multi-task RL with auxiliary tasks, which are easier to learn and relevant to the main task. A weighted distillation loss transfers behaviors from these auxiliary tasks to solve the main task. We demonstrate that AuxDistill can learn a pixels-to-actions policy for a challenging multi-stage embodied object rearrangement task from the environment reward without demonstrations, a learning curriculum, or pre-trained skills. AuxDistill achieves \\(2.3 \\times\\) higher success than the previous state-of-the-art baseline in the Habitat Object Rearrangement benchmark and outperforms methods that use pre-trained skills and expert demonstrations." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.535, + 0.377, + 0.549 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.562, + 0.789, + 0.814 + ], + "angle": 0, + "content": "While reinforcement learning (RL) is successful in a variety of settings including games [9, 31, 41], chatbots [1, 29, 32, 40, 45], and robotics [3, 23, 35], long-horizon tasks such as embodied object rearrangement where an embodied agent must rearrange objects to target positions still remains a challenge [7]. Object rearrangement requires learning heterogeneous behaviors like picking, navigating, placing, and opening concurrently with sequencing these behaviors to solve the overall task all while operating from egocentric vision. Furthermore, the behaviors are interdependent, meaning the robot can only learn how to pick an object if it has first learned how to open the fridge containing the object. Since object rearrangement requires interacting with objects across house-scale spaces with low-level control, the episodes consist of thousands of low-level control steps, exacerbating the problem of credit assignment [19, 34]. Overall, the problems of concurrently learning low-level control, high-level decision-making, interdependent task stages, and episodes with many time steps make object rearrangement challenging even with dense rewards. While some prior methods have applied end-to-end RL to rearrangement problems, they require excessive experience in simplified simulators [8] or expert skill demonstrations [22]. Other prior work has made such complex tasks more tractable by decomposing the full task into" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.217, + 0.149, + 0.23, + 0.159 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.149, + 0.358, + 0.16 + ], + "angle": 0, + "content": "Harish et. al" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.177, + 0.788, + 0.22 + ], + "angle": 0, + "content": "a hierarchy of separately trained skills that a high-level policy sequences together to solve the overall task. However, such hierarchical methods suffer from compounding errors between skills [17] and dynamically selecting between skills." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.229, + 0.789, + 0.589 + ], + "angle": 0, + "content": "We therefore present Reinforcement Learning via Auxiliary Task Distillation (AuxDistill), a method for training policies for long-horizon tasks from scratch with end-to-end RL from reward alone. AuxDistill also learns in auxiliary tasks using a multi-task RL framework and transfers the knowledge from these auxiliary tasks to help solve the desired \"main\" task of interest. Importantly, AuxDistill concurrently learns the main task and all auxiliary tasks, which are all initialized from scratch and not pre-trained. These auxiliary tasks help because they are easier to learn with RL than the main task and contain relevant behaviors for the full task. For example, in object rearrangement, the agent should also practice picking up objects in isolation from the complexities of the full task. Unlike curriculum learning in RL [4,12,13,33], which also leverages easier tasks to learn harder tasks, AuxDistill does not have distinct curriculum phases. Instead, it has a single training objective for distilling sub-behaviors into the overall task behavior. Crucially, the agent end-to-end learns which auxiliary behaviors are relevant for the main task. This relevance is encoded by a scalar weight value determined by the relevance of the robot state to the auxiliary task and it grounded in the oracle task plan. A distillation objective transfers behaviors from relevant auxiliary tasks by encouraging the policy to act consistently between the main and auxiliary tasks for relevant states in the main task. For instance, in object rearrangement, when the robot is near a drawer with the target object inside, it is supervised with the distillation loss from an \"open drawer\" auxiliary task. Despite the policy also concurrently learning the \"open drawer\" auxiliary task from scratch, AuxDistill learns the auxiliary tasks faster since they are easier. Thus, the distillation loss is a useful dense per-time step supervision signal that helps overcome the challenges of learning from the main task reward alone." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.599, + 0.79, + 0.821 + ], + "angle": 0, + "content": "We empirically demonstrate that AuxDistill outperforms a variety of baselines in terms of success rate on the home rearrangement benchmark in Habitat 2.0 [43]. These include baselines such as hierarchical RL, end-to-end RL with and without a curriculum, and imitation learning. On rearrangement episodes where objects start in open receptacles, AuxDistill achieves \\(1.4 \\times\\) higher success than baselines and \\(2.6 \\times\\) higher success in harder episodes where objects can start in closed receptacles. These results highlight the value of AuxDistill to be able to learn the entire rearrangement task end-to-end with RL. We show results beyond object rearrangement on a category-conditioned object manipulation task where AuxDistill achieves \\(1.75 \\times\\) higher success than baselines. We also conduct extensive ablation studies where we analyze the properties of auxiliary tasks needed by AuxDistill, and find that AuxDistill is generally robust to the choice of auxiliary tasks. An analysis of the distillation objective reveals the importance of the distillation loss, with the absence of the distillation loss leading to no success on the rearrangement task. All code is available at https://github.com/absdnd/aux_distill" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.66, + 0.149, + 0.732, + 0.16 + ], + "angle": 0, + "content": "AuxDistill" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.149, + 0.785, + 0.159 + ], + "angle": 0, + "content": "3" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.175, + 0.389, + 0.19 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.206, + 0.788, + 0.372 + ], + "angle": 0, + "content": "Some prior works address complex and long-horizon decision-making problems using a hierarchical breakdown of skills. One such category of approaches is option-critic methods [5, 42, 51], which seek to learn options to temporally abstract the high-level policy decision making. However, discovering and learning such options is unstable and results in a challenging credit assignment problem. Another line of work first trains or is given pre-trained skills and then learns a high-level policy to sequence these skills together to complete longer tasks [11, 18, 24, 48]. Using such a hierarchy results in compounding errors resulting from sequencing skills that were not trained to properly transition between each other [10, 17, 28, 30, 46]. Our approach does not utilize a hierarchical policy, and can solve long horizon tasks better than hierarchical methods by utilizing a single policy and thus avoiding compounding errors that emerge by sequencing independently trained policies." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.372, + 0.789, + 0.593 + ], + "angle": 0, + "content": "Like our method, prior works have tackled model-free end-to-end RL to solve long, complex tasks. Scaling PPO [39] training has enabled robots to learn complex locomotion behaviors [2, 15, 25, 26, 36, 38], manipulate unseen objects in a robotic hand [35], and play competitive video games [9]. Likewise, berges2023galactic showed scaling end-to-end RL can learn object rearrangement. However, their approach requires training agents for billions of environment steps with a simplified kinematic physics simulation. We show results in the full dynamic Habitat simulation, achieving better results in an order of magnitude fewer environment interactions, and show results in harder rearrangement tasks. Other works have incorporated auxiliary objectives in RL to boost performance in embodied tasks [16, 27, 49, 50]. These works add auxiliary self-supervised prediction objectives, whereas our method adds an auxiliary policy learning task via multi-task RL. jia2022improving also uses easier auxiliary RL tasks that specialist agents learn to complete and then distill back into a generalist agent. While our work also uses RL in easier tasks to boost performance, it doesn't require alternating between learning specialist and generalist policies." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.593, + 0.789, + 0.717 + ], + "angle": 0, + "content": "Works have also attempted to learn a curriculum to leverage easier tasks to learn harder tasks that require composing multiple behaviors. Some works learn curriculum generation policies that adjust the environment to suit the agent's capabilities [4, 12, 13]. In other works, a curriculum naturally emerges as a result of downstream training [6]. Finally, some works hand design curricula that change aspects of the environment, such as the starting and goal distributions depending on the agent performance during training [38]. Our work shares a similar spirit of leveraging easier auxiliary tasks to learn a harder task but it does not enforce an explicit curriculum and instead learns all tasks simultaneously." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.739, + 0.331, + 0.754 + ], + "angle": 0, + "content": "3 Method" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.77, + 0.788, + 0.813 + ], + "angle": 0, + "content": "We present Reinforcement Learning via Auxiliary Task Distillation (AuxDistill), a new method for training long-horizon policies from scratch using rewards alone. Learning from reward alone in complex tasks like embodied object rearrangement" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.15, + 0.23, + 0.16 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.149, + 0.358, + 0.159 + ], + "angle": 0, + "content": "Harish et. al" + }, + { + "type": "image", + "bbox": [ + 0.219, + 0.175, + 0.784, + 0.337 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.349, + 0.788, + 0.452 + ], + "angle": 0, + "content": "Fig. 1: AuxDistill learns a rearrangement policy operating from egocentric depth perception and coordinate-based task specification. The full object rearrangement task decomposes into modular abilities that can be learned by auxiliary task with indicator vectors \\( T_{1} \\cdots T_{N} \\) which are trained along with the main task using end-to-end RL. The distillation loss is computed as a weighted combination of the task relevance of \\( o_{t} \\) in the main task \\( T_{0} \\) under all auxiliary tasks. The task relevance function computes \\( w_{i}(s_{t}) \\) based on the relevance of the current observation and robot state to the auxiliary task \\( T_{i} \\). The distillation loss and RL-training loss are then used to update the policy." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.485, + 0.789, + 0.61 + ], + "angle": 0, + "content": "is challenging because an agent needs to combine thousands of low-level actions controlling the arm and base, operate from egocentric visual perception, and dynamically sequence distinct behaviors such as navigating, picking, placing, and opening. AuxDistill addresses these challenges of RL in complex problems, like embodied rearrangement, by learning to leverage knowledge from easier auxiliary tasks related to the desired task we are trying to solve. Unlike a curriculum, which learns gradually harder tasks in stages, AuxDistill learns the desired task concurrently with the auxiliary task and includes a novel distillation mechanism to transfer knowledge from easier to harder tasks." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.647, + 0.373, + 0.66 + ], + "angle": 0, + "content": "3.1 Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.687, + 0.789, + 0.814 + ], + "angle": 0, + "content": "Our problem is formulated as a goal-specified Partially-Observable Markov Decision Process (POMDP) defined as a tuple \\(\\mathcal{M} = (\\mathcal{S},\\mathcal{O},\\mathcal{A},\\mathcal{P},\\mathcal{R},\\mathcal{G},\\rho,\\gamma)\\) with underlying state space \\(\\mathcal{S}\\), observation space \\(\\mathcal{O}\\), action space \\(\\mathcal{A}\\), transition function \\(\\mathcal{P}\\), reward function \\(\\mathcal{R}\\), goal space \\(\\mathcal{G}\\), initial state distribution \\(\\rho\\) and discount factor \\(\\gamma\\). For a task like rearrangement, the goal space \\(\\mathcal{G}\\) is specified using the 3D coordinate of the object's start location and the goal location where the object has to be placed. Our objective is to learn a goal-conditioned policy \\(\\pi(a|o,g)\\) mapping an observation \\(o\\) and goal \\(g\\) to an action \\(a\\) that maximizes the sum of discounted rewards \\(\\mathbb{E}_{s_0\\sim\\rho_0,g\\sim\\mathcal{G}}\\sum_t\\gamma^t\\mathcal{R}(s_t,g)\\)." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.66, + 0.149, + 0.732, + 0.16 + ], + "angle": 0, + "content": "AuxDistill" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.149, + 0.785, + 0.159 + ], + "angle": 0, + "content": "5" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.177, + 0.717, + 0.192 + ], + "angle": 0, + "content": "3.2 Reinforcement Learning via Auxiliary Task Distillation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.202, + 0.788, + 0.422 + ], + "angle": 0, + "content": "The core idea of AuxDistill is to learn a policy in a difficult desired \"main\" task by transferring knowledge from easier auxiliary tasks. This is done without using an explicit curriculum by learning the main and auxiliary tasks in a single training loop. We refer to the main task we wish to learn as \\(\\mathcal{M}_0\\) with a task identifier \\(T_0\\). We define \\(N\\) auxiliary tasks, defined as \\(\\{\\mathcal{M}_i\\}_{n=1}^N\\) with task identifiers \\(\\{T_i\\}_{n=1}^N\\). The auxiliary tasks \\(\\{\\mathcal{M}_i\\}\\) share the same state, observation, goal, and action space as the target task \\(\\mathcal{M}_0\\). Each \\(\\mathcal{M}_i\\) has a separately defined starting state distribution and reward function. We assume these auxiliary tasks are related to the full task yet are easier to solve than the full task. Specifically, the auxiliary tasks can be easier instantiations or sub-parts of the overall task. For example, in rearrangement, we define the auxiliary tasks in terms of interactions the agent needs to complete the entire rearrangement episode, such as picking, placing, and opening. Prior works use similar task definitions to train skills for hierarchical policies in rearrangement [17], but these works suffer from a two-stage pipeline of first needing to separately train each skill and then decide how to combine them. AuxDistill directly learns the full task from scratch while also performing better." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.424, + 0.788, + 0.52 + ], + "angle": 0, + "content": "AuxDistill learns a single policy with RL that concurrently learns to perform the main and auxiliary tasks. We illustrate the implementation of AuxDistill in Figure 1. This policy, parameterized by \\(\\theta\\), is expressed as \\(\\pi_{\\theta}(a_t \\mid o_t, g, T)\\) for observation \\(o_t\\), episode goal \\(g\\), and per-\\(\\mathcal{M}\\) task identifier \\(T\\) which is encoded as a one-hot embedding in a vector which has the same size as the maximum number of auxiliary tasks. Note that since all tasks share the same observation and action space, \\(\\pi_{\\theta}\\) can act and observe in all tasks based on the input task identifier \\(T\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.521, + 0.788, + 0.686 + ], + "angle": 0, + "content": "AuxDistill updates the policy based on an average RL loss from the main task and all auxiliary tasks. Let \\(\\mathcal{L}_{\\mathcal{M}_i}^{\\mathrm{RL}}(\\theta)\\) denote the RL loss for \\(\\pi_{\\theta}\\) in MDP \\(\\mathcal{M}_i\\). These auxiliary tasks are designed to capture a subset of the main task which are easier to accomplish than the main task. To compute these losses, we assume we can collect experience in the auxiliary tasks. For example, to collect experience in a \"pick object\" auxiliary task in object rearrangement, the robot is spawned close to the object. We then update based on the average of the task and auxiliary task losses: \\(\\frac{1}{N}\\sum_{i=0}^{N}\\mathcal{L}_{\\mathcal{M}_i}^{\\mathrm{RL}}(\\theta)\\). Intuitively, \\(\\pi_{\\theta}\\) will first learn to complete the easier auxiliary tasks, and this auxiliary task competency can aid in solving the main task. This shares a similar insight as curriculum learning, except all tasks are learned concurrently, and the curriculum stages are naturally induced by the policy naturally learning on easier tasks first." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.687, + 0.788, + 0.812 + ], + "angle": 0, + "content": "In addition to optimizing the average RL loss between the main and auxiliary tasks, AuxDistill also optimizes a distillation loss that encourages the policy to transfer relevant knowledge from the auxiliary tasks to the main task. To achieve this, for a particular observation \\( o_t \\), we consider the policy distribution \\( \\pi_\\theta(\\cdot | o_t, g, T) \\) under different task identifiers \\( T \\). For compactness, notate \\( \\pi_\\theta(\\cdot | o_t, g, T_i) \\) where \\( T_i \\) indicates the task identifier for \\( \\mathcal{M}_i \\) as \\( \\pi_\\theta^{T_i}(o_t) \\). We want \\( \\pi_\\theta^{T_0} \\) (the policy in the main task) to match the behaviors from the policy in the relevant auxiliary tasks \\( \\pi_\\theta^{T_i} \\). We measure this relevance of an auxiliary task \\( \\mathcal{M}_i \\) to the main task at time step \\( t \\) via an auxiliary task relevance function \\( w_i(s_t) \\)." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.217, + 0.149, + 0.23, + 0.159 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.149, + 0.358, + 0.16 + ], + "angle": 0, + "content": "Harish et. al" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.177, + 0.788, + 0.329 + ], + "angle": 0, + "content": "This function denotes how much the knowledge from \\(\\mathcal{M}_i\\) should apply to \\(\\mathcal{M}_0\\) in the underlying simulator state \\(s_t\\). This relevance function can be grounded in the task plan generated by an oracle planner. For example, consider the state of the robot before picking up the object. In this state, the pick auxiliary task would be relevant to the main task, and the place auxiliary task would not. If an episode requires opening a fridge before accessing the object, the relevant task before the fridge is opened would be open-fridge. Note that computing the distillation weight can utilize oracle knowledge of the simulator state (for instance, whether the object is inside a fridge or a cabinet) since this information is only provided as a training signal and not used during inference. This information is utilized by all methods we compare against, including our strongest baseline [22]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.33, + 0.788, + 0.399 + ], + "angle": 0, + "content": "We can then distill experience from \\(\\mathcal{M}_i\\) by supervising the policy to match the action distribution of the target task \\(\\mathcal{M}_0\\). The distillation loss is computed as the KL-divergence of the action distribution under \\(\\mathcal{M}_i\\) and \\(\\mathcal{M}_0\\) weighted by the relevance of auxiliary task \\(\\mathcal{M}_i\\) given by \\(w_i(s_t)\\). The overall loss function of AuxDistill, including the distillation loss, is then:" + }, + { + "type": "equation", + "bbox": [ + 0.318, + 0.41, + 0.788, + 0.449 + ], + "angle": 0, + "content": "\\[\n\\frac {1}{N} \\sum_ {i = 0} ^ {N} \\mathcal {L} _ {\\mathcal {M} _ {i}} ^ {\\mathrm {R L}} (\\theta) + \\lambda \\sum_ {i = 1} ^ {N} w _ {i} \\left(s _ {t}\\right) D _ {\\mathrm {K L}} \\left(\\pi_ {\\theta} ^ {T _ {0}} \\left(o _ {t}\\right) \\parallel \\pi_ {\\theta} ^ {T _ {i}} \\left(o _ {t}\\right)\\right) \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.458, + 0.788, + 0.569 + ], + "angle": 0, + "content": "Here \\(\\lambda\\) represents the distillation weight relative to other PPO losses. Note that our formulation relies on the auxiliary tasks being easier to solve and relevant to the main task. Distillation using the task relevance function offers an additional supervisory signal that the policy can optimize along with its own reward. This doesn't impose the large overhead of hierarchical RL methods where each auxiliary task must be pre-trained in isolation with its own reward function and goal specification. Auxiliary tasks are flexibly incorporated to address parts of the task that are challenging for the agent to learn." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.592, + 0.457, + 0.606 + ], + "angle": 0, + "content": "3.3 Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.617, + 0.788, + 0.687 + ], + "angle": 0, + "content": "We use PPO [39] to compute the RL loss \\(\\mathcal{L}_{\\mathcal{M}_i}^{\\mathrm{RL}}\\) in Equation (1). We create \\(M\\) environment instances for each of the main and \\(N\\) auxiliary tasks and vectorize them for parallel action execution. We rollout the policy in all these \\(M(N + 1)\\) environments and collect a batch of data. We then compute Equation (1), using the PPO loss, update the policy, and then repeat this process." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.687, + 0.788, + 0.812 + ], + "angle": 0, + "content": "The policy is represented as a 2-layer LSTM network with 512 hidden units per layer, similar to the architecture employed in [43]. In our experiments, \\(o_t\\) is a depth or RGB egocentric image, which we encode with a ResNet50 [20] network. The goal and the proprioceptive state of the agent are concatenated with the visual embeddings along with the index of the POMDP \\(\\mathcal{M}_i\\), for which we use a one-hot embedding. This vector is then passed as an input to the LSTM. A linear projection on the output of the LSTM then produces an action to execute on the next step in the environment and a vector of value-predictions corresponding to each of the tasks \\(\\mathcal{M}_i\\) used during training." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.66, + 0.149, + 0.732, + 0.16 + ], + "angle": 0, + "content": "AuxDistill" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.149, + 0.786, + 0.159 + ], + "angle": 0, + "content": "7" + }, + { + "type": "table", + "bbox": [ + 0.234, + 0.174, + 0.761, + 0.307 + ], + "angle": 0, + "content": "
MethodTrain (Seen)Eval (Unseen)
All EpisodesEasyHardAll EpisodesEasyHard
M3 (Oracle) [17]27 ± 057 ± 212 ± 128 ± 058 ± 213 ± 2
M3 [17]25 ± 156 ± 19 ± 213 ± 153 ± 40 ± 0
Galactic [8]-37 ± 0--26 ± 0-
Monolithic RL0 ± 00 ± 00 ± 00 ± 00 ± 00 ± 0
Skill Transformer [22]25 ± 144 ± 116 ± 123 ± 137 ± 116 ± 1
AuxDistill (No Distillation)0 ± 00 ± 00 ± 00 ± 00 ± 00 ± 0
RL Curriculum0 ± 01 ± 00 ± 00 ± 01 ± 10 ± 0
AuxDistill49 ± 274 ± 236 ± 252 ± 275 ± 241 ± 2
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.214, + 0.308, + 0.788, + 0.397 + ], + "angle": 0, + "content": "Table 1: Success rates on the rearrangement task for our method, AuxDistill (highlighted in blue), and baselines. Displayed are the average and standard deviations for 3 seeds for M3, Monolithic RL, Skill Transformer, and M3 Oracle (numbers from [22]), and 3 seeds for the remaining methods with the highest numbers per setting bolded. Numbers in the easy and hard columns are averages over 100 episodes and 200 episodes, respectively. The All Episodes column is an average across both the Easy and Hard episodes. AuxDistill outperforms all baselines in all settings." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.437, + 0.788, + 0.507 + ], + "angle": 0, + "content": "Our policy is learned via a multi-task RL formulation, with the main and auxiliary tasks being concurrently learned. Each of the tasks may have different reward magnitudes and relative performance during training. To address this, we use PopArt [21], with \\(\\beta = 3e^{-4}\\) to enable learning across different return scales. Additional AuxDistill details are in Appendix A of the supplementary." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.531, + 0.376, + 0.547 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.563, + 0.452, + 0.578 + ], + "angle": 0, + "content": "4.1 Object-Rearrangement" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.589, + 0.788, + 0.782 + ], + "angle": 0, + "content": "In this section, we compare AuxDistill with baselines on the Habitat 2.0 Object Rearrangement task [43]. For comparison with baselines, we run our experiments using the setup from [22]. In this task, a Fetch robot [37] must move an object from a specified start position to a desired goal position in an indoor home environment using only onboard sensing. The agent has no privileged information like existing maps of the environment, 3D object models, or exact object positions. The Fetch robot senses the world through a \\(256 \\times 256\\) depth camera mounted on the robot's head, robot joint positions, gripper state, and base egomotion, giving the relative position of the robot from the start of the episode. The task is specified by a starting 3D object coordinate for the object to move and a 3D goal coordinate to move the object to. Both coordinates are specified relative to the robot's position at the start of the episode. Only the starting object coordinate is specified, and this information is not updated based on the current object position (e.g. if it is moved)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.784, + 0.788, + 0.813 + ], + "angle": 0, + "content": "The robot interacts with the world via a 7DoF arm, a suction gripper attached to the end of the arm, and a mobile base. The episode is successful if the target" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.217, + 0.149, + 0.23, + 0.159 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.149, + 0.358, + 0.16 + ], + "angle": 0, + "content": "Harish et. al" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.177, + 0.784, + 0.204 + ], + "angle": 0, + "content": "object is within \\(15\\mathrm{cm}\\) of the goal position. The robot has a budget of 1,500 steps to complete the task." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.205, + 0.788, + 0.329 + ], + "angle": 0, + "content": "We report performance on the easy and hard evaluation episodes from [22]. In easy episodes, both the target object and goal are on an open receptacle. This means the robot does not have to first open a receptacle before accessing the target object or goal. Instead, the agent can always execute the same sequence of navigate to object, pick up object, navigate to goal, and place object at goal. In hard episodes, the object may start in a closed receptacle. The agent, therefore, needs to use its visual input to perceive if the target object is in a closed receptacle. If so, the robot must then open the receptacle before picking the object. The object may start in either the fridge or a cabinet." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.347, + 0.788, + 0.457 + ], + "angle": 0, + "content": "Training Setup: We train AuxDistill using 11,791 episodes with a mix of easy and hard episodes across 63 scenes using the same rearrangement training dataset as in [44]. These episodes were obtained by sampling an equal number of episodes in the easy and hard categories, with easy episodes being equally sampled across episodes with open-cabinet, open-fridge and non-articulated episodes. For hard episodes, sampling is done uniformly between closed fridge and closed cabinet episodes. We use auxiliary tasks covering abilities included as a part of the standard rearrangement benchmark used in [44]. The auxiliary tasks are:" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.457, + 0.785, + 0.485 + ], + "angle": 0, + "content": "- Pick: The agent spawns randomly within the house and must navigate to and pick up the object." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.487, + 0.785, + 0.528 + ], + "angle": 0, + "content": "- Place the agent is spawned randomly in the house with the object in its gripper. The robot must navigate to and place the object within \\(15\\mathrm{cm}\\) of the target 3D location on the receptacle." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.531, + 0.785, + 0.559 + ], + "angle": 0, + "content": "- Open Fridge: The robot spawns randomly within the house and must navigate to the fridge and use its base and arm to open the fridge door." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.561, + 0.785, + 0.589 + ], + "angle": 0, + "content": "- Open Cabinet: The robot spawns randomly in the house and must navigate to the kitchen area and open the cabinet using its base and arm." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.59, + 0.785, + 0.631 + ], + "angle": 0, + "content": "- Pick from Fridge: This task is similar to Pick except the agent has to pick up the object from inside an open fridge, which requires careful manipulation to minimize collisions." + }, + { + "type": "list", + "bbox": [ + 0.215, + 0.457, + 0.785, + 0.631 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.632, + 0.788, + 0.686 + ], + "angle": 0, + "content": "The auxiliary relevance function is grounded in the oracle task plan [14] of the episode. This is formulated as an indicator function with \\( w_{i}(s_{t}) = 1 \\) if the state \\( s_t \\) is relevant to \\( \\mathcal{M}_i \\) based on the oracle task plan of the episode and \\( w_{i}(s_{t}) = 0 \\) otherwise." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.687, + 0.788, + 0.812 + ], + "angle": 0, + "content": "For example, if at step \\( t \\) the agent is not holding an object and the object is in an open receptacle, the agent must first pick up the object so \\( w_{i_{\\text{pick}}}(s_t) = 1 \\), and \\( w_i(s_t) = 0 \\) for all other skills. For a complete description of the auxiliary tasks, including their reward functions, along with the precise auxiliary relevance function definition in object rearrangement, see Appendix B. We train methods for 475M steps of environment interactions. We found this to be a sufficient number of environment interactions to ensure sufficient progress on auxiliary tasks for distillation to aid rearrangement performance on both the easy and hard episodes. For AuxDistill, we count the environment interactions in the auxiliary" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.66, + 0.148, + 0.732, + 0.16 + ], + "angle": 0, + "content": "AuxDistill" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.149, + 0.785, + 0.159 + ], + "angle": 0, + "content": "9" + }, + { + "type": "table", + "bbox": [ + 0.231, + 0.174, + 0.768, + 0.27 + ], + "angle": 0, + "content": "
MethodTrain (Seen)Eval (Unseen)
All EpisodesEasyHardAll EpisodesEasyHard
All Skills283619303921
No Pick From Fridge2032712232
No Open Fridge414933485739
No Pick000000
" + }, + { + "type": "table_caption", + "bbox": [ + 0.215, + 0.271, + 0.785, + 0.32 + ], + "angle": 0, + "content": "Table 2: Robustness to auxiliary task selection. We compare the train and generalization performance of AuxDistill with 4 different auxiliary task selections. Results are averages across 100 episodes on a single seed of training. The train and evaluation episodes are in a simplified setting compared to Table 1 as described in Section 4.3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.37, + 0.784, + 0.398 + ], + "angle": 0, + "content": "tasks towards this 475M step budget. We train with a learning rate of \\(3e^{-4}\\) and linearly decay the learning rate to 0 over the course of learning." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.423, + 0.785, + 0.478 + ], + "angle": 0, + "content": "Baselines We compare AuxDistill to both hierarchical methods governed by a task plan as well as monolithic baselines, which directly learn a pixels-to-actions policy using sensor observations. We compare to relevant baselines from huang2023skill and two more baselines that use the auxiliary tasks." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.48, + 0.784, + 0.547 + ], + "angle": 0, + "content": "- Monolithic RL: A monolithic neural network is trained to map sensor observations directly to actions trained with end-to-end RL. The policy showed no signs of learning the main task, so we stopped training early after \\(100M\\) steps of training. This outcome is consistent with the monolithic RL baseline in prior works [8, 43]." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.55, + 0.784, + 0.605 + ], + "angle": 0, + "content": "- Galactic: [8] An end-to-end RL framework similar to monolithic RL to map sensor observations directly to actions with the policy being trained with a simplified kinematic simulation on over \\(>1e^{9}\\) simulation steps and transferred to dynamic simulation utilized in our setting." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.608, + 0.784, + 0.634 + ], + "angle": 0, + "content": "- M3: [17] Each of the skill policies navigate, pick, place, and open are separately trained. These skills are then sequenced using a task planner." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.638, + 0.784, + 0.664 + ], + "angle": 0, + "content": "- M3 (Oracle): [17] This uses the same skill training as M3, but uses an oracle task planner." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.668, + 0.785, + 0.708 + ], + "angle": 0, + "content": "- Skill Transformer [22]: This method uses pre-trained skills to collect successful rearrangement demonstrations and then train on these demonstrations with imitation learning." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.712, + 0.784, + 0.765 + ], + "angle": 0, + "content": "- RL Curriculum This method is trained in two stages: i) First, we train the policy on the auxiliary tasks for \\(200M\\) steps to ensure the relevant auxiliary tasks have good training performance ii) This is followed by a training on the entire rearrangement task for another \\(300M\\) steps." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.769, + 0.784, + 0.795 + ], + "angle": 0, + "content": "- AuxDistill: No-Distillation Train AuxDistill under a similar setting as described in Section 4.1 except with the distillation strength \\(\\lambda = 0.0\\)." + }, + { + "type": "list", + "bbox": [ + 0.215, + 0.48, + 0.785, + 0.795 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.798, + 0.593, + 0.811 + ], + "angle": 0, + "content": "For more details on the baselines, see Appendix B.2." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.149, + 0.236, + 0.159 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.149, + 0.358, + 0.159 + ], + "angle": 0, + "content": "Harish et. al" + }, + { + "type": "image", + "bbox": [ + 0.301, + 0.192, + 0.686, + 0.336 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.346, + 0.788, + 0.435 + ], + "angle": 0, + "content": "Fig. 2: Comparison of skill-robustness to different choices of auxiliary task on the hard distribution. Including both the Pick and Pick from Fridge is crucial for successful rearrangement on this distribution. Not utilizing Open-Fridge leads to a boost in rearrangement success. This improvement arises because the open-fridge skill is the easiest of all auxiliary tasks and utilizing it reduces the number of samples for the main task (from the 100M step budget). We discuss the auxiliary task learning curves in Appendix B.3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.464, + 0.788, + 0.629 + ], + "angle": 0, + "content": "Rearrangement Performance: In Table 1, we compare AuxDistill to baselines in the rearrangement task. On all evaluation settings, AuxDistill outperforms all baselines. AuxDistill shows an absolute improvement of \\(22\\%\\) and \\(25\\%\\) on the easy and hard episodes respectively over the best performing baseline. Most baselines struggle to achieve any performance on the unseen episodes. Monolithic RL [43] achieves no success in any of the evaluation settings, demonstrating the impracticality of learning directly from the dense reward in the full rearrangement task. Our method also outperforms Galactic [8] despite learning on less than half the number of training samples (475M for AuxDistill vs. \\(>1e^{9}\\) for Galactic) on the easy split with \\(74\\%\\) success over \\(37\\%\\) on easy episodes during training and \\(75\\%\\) vs. \\(26\\%\\) on evaluation. Note that Galactic does not report performance on the hard distribution." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.631, + 0.788, + 0.741 + ], + "angle": 0, + "content": "Like AuxDistill, RL Curriculum also utilizes the auxiliary tasks yet finds no success. This demonstrates the value of concurrent training on the main and auxiliary tasks with the distillation loss in AuxDistill. Likewise, AuxDistill (No Distill) achieves no success, illustrating the importance of the distillation loss. The learning curve in Figure 3a as AuxDistill can gradually learn to complete the task with more environment interactions. On the other hand, RL Curriculum and AuxDistill (No-Distill) remain at no success regardless of the number of learning samples." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.742, + 0.788, + 0.812 + ], + "angle": 0, + "content": "AuxDistill also outperforms the strongest baseline, Skill Transformer, by a significant margin with \\(75\\%\\) vs. \\(37\\%\\) on the unseen easy episodes and \\(41\\%\\) vs. \\(16\\%\\) on the unseen hard episodes. This demonstrates the advantages of using online, end-to-end RL, rather than solely training offline with demonstrations. As shown in Figure 3a, AuxDistill is able to improve with subsequent environment" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.66, + 0.148, + 0.732, + 0.16 + ], + "angle": 0, + "content": "AuxDistill" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.149, + 0.784, + 0.159 + ], + "angle": 0, + "content": "11" + }, + { + "type": "table", + "bbox": [ + 0.316, + 0.174, + 0.648, + 0.235 + ], + "angle": 0, + "content": "
Method Train (Seen) Eval (Unseen)
AuxDistill12± 211± 1
Monolithic8± 09± 1
AuxDistill (No Distillation)3± 14± 1
RL-Curriculum6± 18± 0
" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.236, + 0.785, + 0.286 + ], + "angle": 0, + "content": "Table 3: Comparison of our method on the Category Pick task with auxiliary tasks on three random seeds. AuxDistill outperforms all methods, including the monolithic baseline, on the unseen evaluation split. Evaluation is conducted over 1000 seen episodes sampled from the training distribution and 1000 held-out episodes from evaluation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.325, + 0.784, + 0.351 + ], + "angle": 0, + "content": "interactions, yet Skill Transformer is limited by the performance of the expert demonstrations." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.353, + 0.785, + 0.476 + ], + "angle": 0, + "content": "Table 1 also demonstrates that AuxDistill outperforms hierarchical baselines. M3 is able to achieve \\(53\\%\\) success on the unseen easy episodes since it utilizes strong pretrained skills. Even in this setting, AuxDistill achieves a higher success of \\(75\\%\\). However, on the hard episodes, M3 cannot dynamically plan skills and achieves no success in the hard unseen episodes. AuxDistill achieves \\(41\\%\\) on this setting because it learns a monolithic policy with RL that combines low-level and high-level decision-making. AuxDistill even outperforms an oracle version of M3 that dynamically plans the skill sequence based on oracle state information (13% vs. 41% on the unseen hard episodes)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.478, + 0.787, + 0.587 + ], + "angle": 0, + "content": "We observe that AuxDistill performs \\(7\\%\\) worse on the Habitat 2.0 rearrangement training episodes than the evaluation episodes. We find this performance gap is due to additional easier episodes in the evaluation distribution where the object is closer to the target location. In the hard split, there are 55 eval episodes and 29 train episodes where the object is picked up from the fridge and the target location is \\(< 0.5m\\) from the object location. On removing these episodes, train performance is higher than evaluation, with \\(32\\%\\) success on train and \\(31\\%\\) success on evaluation." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.608, + 0.381, + 0.623 + ], + "angle": 0, + "content": "4.2 Category Pick" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.632, + 0.785, + 0.755 + ], + "angle": 0, + "content": "The merits of AuxDistill extend to other challenging embodied tasks. In particular, consider a variant of the pick task described in Section 4.1 where a robot has to pick up an object using the object name passed as a one-hot embedding to the policy. Category pick requires the policy to discern the object to pick up by correlating the RGB observation with the object category passed as a one-hot embedding. This task is more challenging than geometric pick where the policy has access to the initial coordinate specification of the object to be picked. We include additional details about the task specification in Section B.4 of our supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.756, + 0.787, + 0.811 + ], + "angle": 0, + "content": "We leverage the easier coordinate pick task to aid the learning of the Category Pick task using AuxDistill. More specifically, we consider the task of interest \\( M_0 \\) to be the Category Pick task and \\( M_1 \\) (auxiliary task we would like to distill from) to be the coordinate pick task. Both tasks are trained jointly using AuxDistill" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.149, + 0.237, + 0.159 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.149, + 0.358, + 0.159 + ], + "angle": 0, + "content": "Harish et. al" + }, + { + "type": "image", + "bbox": [ + 0.247, + 0.187, + 0.476, + 0.311 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.285, + 0.313, + 0.457, + 0.324 + ], + "angle": 0, + "content": "(a) Rearrange learning curve." + }, + { + "type": "image", + "bbox": [ + 0.502, + 0.187, + 0.738, + 0.312 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.549, + 0.313, + 0.711, + 0.324 + ], + "angle": 0, + "content": "(b) Category Pick learning." + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.334, + 0.788, + 0.41 + ], + "angle": 0, + "content": "Fig. 3: Left: RL training success rates on training episodes on the rearrangement task from Table 1. Note that AuxDistill (No Distill) and RL-Curriculum are displayed but achieve \\(0\\%\\) success throughout training. Right: Learning curve on the Category Pick task of AuxDistill utilizing coordinate pick distillation vs. monolithic RL. AuxDistill outperforms baselines in both settings. Displayed are averages and standard deviation over 3 random seeds." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.437, + 0.784, + 0.478 + ], + "angle": 0, + "content": "with the same reward structure with \\( w_{i}(s_{t}) = 1 \\) and use \\( \\lambda = 0.5 \\) for distillation in Equation (1). For this task, we train on the full rearrange-easy dataset from [44] with 50,000 episodes." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.479, + 0.788, + 0.575 + ], + "angle": 0, + "content": "We report the training curves of the Category Pick task in comparison with the RL Curriculum and No Distillation baseline used in Table 1. In addition, we introduce a monolithic baseline for this task, which directly trains Category Pick without leveraging the coordinate pick task during training. We report the training curves in Figure 3b. We evaluate the performance of the trained policy on 1000 episodes sampled from the training and 1000 episodes from the validation split." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.576, + 0.788, + 0.659 + ], + "angle": 0, + "content": "In Table 3, we show that AuxDistill shows better performance with a success rate of \\(11\\%\\) outperforming all baselines that do not leverage distillation during RL training. The closest performing baseline to ours is the monolithic baseline, which achieves a success of \\(9\\%\\) on the held-out distribution. AuxDistill (No-Distillation) and RL-curriculum perform worse with an evaluation success of \\(4\\%\\) and \\(8\\%\\), respectively." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.678, + 0.424, + 0.693 + ], + "angle": 0, + "content": "4.3 AuxDistill Analysis" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.701, + 0.788, + 0.783 + ], + "angle": 0, + "content": "Robustness to Auxiliary Task Selection A critical consideration for training AuxDistill is the selection of auxiliary tasks. In this section, we compare four different selections of auxiliary tasks in the object rearrangement task from Section 4.1. We conduct this ablation on a smaller distribution of episodes, which only include the Easy episodes and Hard episodes where the object starts in the fridge and train AuxDistill for 100M steps." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.784, + 0.788, + 0.812 + ], + "angle": 0, + "content": "As we conduct this study only on the fridge category of articulated episodes, we modify the auxiliary task selection to be: Pick, Open Fridge, Place and Pick" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.66, + 0.149, + 0.732, + 0.16 + ], + "angle": 0, + "content": "AuxDistill" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.149, + 0.786, + 0.159 + ], + "angle": 0, + "content": "13" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.177, + 0.785, + 0.219 + ], + "angle": 0, + "content": "from Fridge. Note that the first three auxiliary tasks are the same from Section 4.1. The other auxiliary task selections are the same as this original selection, but excluding one of: Pick, Open Fridge, and Pick from Fridge." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.219, + 0.787, + 0.341 + ], + "angle": 0, + "content": "In Table 2, we report the performance for each skill selection on the easy and hard episodes. We observe that among all the sub-tasks, the most important is Pick, as excluding it results in \\(0.0\\%\\) success. Note that Pick is only relevant \\((w_{i}(s_{t}) > 0)\\) to easy episodes during training i.e. removing it should only impact the performance on easy distribution of our training setup. However, we find that it affects both the easy and hard episodes as the task is not successfully carried out in either case if removed. In Figure 2, we analyze the success of individual stages of rearrangement and notice that not including Pick leads to failure earlier on in the task." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.343, + 0.788, + 0.495 + ], + "angle": 0, + "content": "In Table 5 we notice that introducing the Open-Fridge skill can worsen rearrangement performance (30% vs. 48%). The reason for this is that the open-fridge skill is the easiest of all auxiliary tasks (see Appendix B.3), and utilizing it reduces the number of samples allocated to the main task. However, certain auxiliary tasks can boost performance by addressing specific bottlenecks in the rearrangement. One such example is the Pick from Fridge, which boosts performance 7% to 19% by addressing the task of picking from an open fridge, which can be challenging due to the difficulty of avoiding collisions with the fridge door. Further, the 4% performance improvement from 32% to 36% on easy distribution can be attributed to the presence of easy episodes in our dataset with an open-refrigerator where Pick from Fridge helps boost performance." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.514, + 0.788, + 0.666 + ], + "angle": 0, + "content": "Distillation Coefficient Variation We analyze the effect of the distillation loss coefficient \\(\\lambda\\) in AuxDistill in Table 4. The distillation loss coefficient controls how the agent balances minimizing the distillation from the auxiliary tasks versus maximizing the reward on the main task. In Table 4, we notice that using high values of distillation weighting \\(\\lambda = 1.0\\) shifts the objective of the rearrangement task from maximizing cumulative reward to distilling from the auxiliary task leading to \\(1\\%\\) success rate on rearrangement and similarly, using a very small value of distillation coefficient \\(\\lambda = 0.01\\) leads to insufficient leveraging of auxiliary task information with a success rate of \\(7\\%\\). We find \\(\\lambda = 0.1, 0.05\\) to be good choices for optimizing task reward and distilling behaviors from the auxiliary tasks." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.687, + 0.36, + 0.701 + ], + "angle": 0, + "content": "5 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.714, + 0.788, + 0.812 + ], + "angle": 0, + "content": "In this work, we presented AuxDistill, a new method for end-to-end RL on complex tasks by leveraging auxiliary tasks. AuxDistill learns in the auxiliary tasks concurrently with the main task through multi-task RL. A distillation loss transfers relevant behaviors from the auxiliary tasks to the main task. We show that AuxDistill outperforms a variety of baselines in Habitat object rearrangement by up to \\(27\\%\\). We also show another application of AuxDistill in a category-conditioned manipulation task. Finally, we analyze AuxDistill in" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.149, + 0.236, + 0.159 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.149, + 0.358, + 0.159 + ], + "angle": 0, + "content": "Harish et. al" + }, + { + "type": "table", + "bbox": [ + 0.29, + 0.174, + 0.702, + 0.277 + ], + "angle": 0, + "content": "
MethodTrain (Seen)Eval (Unseen)
All EpisodesEasyHardAll EpisodesEasyHard
λ = 0.0166510182
λ = 0.05303128363338
λ = 0.1222619303921
λ = 0.5230250
λ = 1.0010120
" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.278, + 0.787, + 0.341 + ], + "angle": 0, + "content": "Table 4: Performance of our method for a single seed on varying distillation coefficient during training. Using a large distillation coefficient \\((\\lambda = 1.0)\\) makes reward optimization challenging, and too small \\((\\lambda = 0.01)\\) results in insufficient auxiliary distillation to succeed on the rearrangement task. The intermediate values \\(\\lambda = 0.05\\) and \\(\\lambda = 0.1\\) show the best performance on training and evaluation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.38, + 0.786, + 0.422 + ], + "angle": 0, + "content": "different auxiliary task selections and magnitudes of distillation strengths. Overall, AuxDistill presents a new way to tackle compound tasks with RL alone without requiring pre-trained skills or expert demonstrations." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.422, + 0.787, + 0.505 + ], + "angle": 0, + "content": "Limitations of AuxDistill include the need for the auxiliary tasks and the auxiliary task relevance function. The auxiliary tasks require knowing behaviors that are relevant and easier to learn than the main task. Each auxiliary task requires defining a new start state distribution and reward function. However, designing these can rely on privileged state information which is less restrictive than the requirement of pre-trained skills or expert demonstrations." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.527, + 0.325, + 0.541 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.223, + 0.556, + 0.787, + 0.594 + ], + "angle": 0, + "content": "1. Achiam, J., Adler, S., Agarwal, S., Ahmad, L., Akkaya, I., Aleman, F.L., Almeida, D., Altenschmidt, J., Altman, S., Anadkat, S., et al.: Gpt-4 technical report. arXiv preprint arXiv:2303.08774 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.223, + 0.596, + 0.788, + 0.633 + ], + "angle": 0, + "content": "2. Agarwal, A., Kumar, A., Malik, J., Pathak, D.: Legged locomotion in challenging terrains using egocentric vision. In: Conference on Robot Learning. pp. 403-415. PMLR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.223, + 0.634, + 0.788, + 0.671 + ], + "angle": 0, + "content": "3. Akkaya, I., Andrychowicz, M., Chociej, M., Litwin, M., McGrew, B., Petron, A., Paino, A., Plappert, M., Powell, G., Ribas, R., et al.: Solving rubik's cube with a robot hand. arXiv preprint arXiv:1910.07113 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.223, + 0.672, + 0.788, + 0.71 + ], + "angle": 0, + "content": "4. Azad, A.S., Gur, I., Emhoff, J., Alexis, N., Faust, A., Abbeel, P., Stoica, I.: Clutr: Curriculum learning via unsupervised task representation learning. In: International Conference on Machine Learning. pp. 1361-1395. PMLR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.223, + 0.71, + 0.788, + 0.735 + ], + "angle": 0, + "content": "5. Bacon, P.L., Harb, J., Precup, D.: The option-critic architecture. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 31 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.223, + 0.736, + 0.788, + 0.772 + ], + "angle": 0, + "content": "6. Baker, B., Kanitscheider, I., Markov, T., Wu, Y., Powell, G., McGrew, B., Mordatch, I.: Emergent tool use from multi-agent autocurricula. arXiv preprint arXiv:1909.07528 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.223, + 0.773, + 0.788, + 0.811 + ], + "angle": 0, + "content": "7. Batra, D., Chang, A.X., Chernova, S., Davison, A.J., Deng, J., Koltun, V., Levine, S., Malik, J., Mordatch, I., Mottaghi, R., et al.: Rearrangement: A challenge for embodied ai. arXiv preprint arXiv:2011.01975 (2020)" + }, + { + "type": "list", + "bbox": [ + 0.223, + 0.556, + 0.788, + 0.811 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.66, + 0.148, + 0.732, + 0.16 + ], + "angle": 0, + "content": "AuxDistill" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.149, + 0.786, + 0.159 + ], + "angle": 0, + "content": "15" + }, + { + "type": "ref_text", + "bbox": [ + 0.223, + 0.178, + 0.786, + 0.228 + ], + "angle": 0, + "content": "8. Berges, V.P., Szot, A., Chaplot, D.S., Gokaslan, A., Mottaghi, R., Batra, D., Undersander, E.: Galactic: Scaling end-to-end reinforcement learning for rearrangement at 100k steps-per-second. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13767-13777 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.223, + 0.229, + 0.786, + 0.265 + ], + "angle": 0, + "content": "9. Berner, C., Brockman, G., Chan, B., Cheung, V., Debiak, P., Dennison, C., Farhi, D., Fischer, Q., Hashme, S., Hesse, C., et al.: Dota 2 with large scale deep reinforcement learning. arXiv preprint arXiv:1912.06680 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.221, + 0.266, + 0.785, + 0.29 + ], + "angle": 0, + "content": "10. Chen, Y., Wang, C., Fei-Fei, L., Liu, C.K.: Sequential dexterity: Chaining dexterous policies for long-horizon manipulation. arXiv preprint arXiv:2309.00987 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.291, + 0.785, + 0.327 + ], + "angle": 0, + "content": "11. Dalal, M., Pathak, D., Salakhutdinov, R.R.: Accelerating robotic reinforcement learning via parameterized action primitives. Advances in Neural Information Processing Systems 34, 21847-21859 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.328, + 0.787, + 0.365 + ], + "angle": 0, + "content": "12. Dennis, M., Jaques, N., Vinitsky, E., Bayen, A., Russell, S., Critch, A., Levine, S.: Emergent complexity and zero-shot transfer via unsupervised environment design. Advances in neural information processing systems 33, 13049-13061 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.365, + 0.787, + 0.402 + ], + "angle": 0, + "content": "13. Fang, K., Migimatsu, T., Mandlekar, A., Fei-Fei, L., Bohg, J.: Active task randomization: Learning visuomotor skills for sequential manipulation by proposing feasible and novel tasks. arXiv preprint arXiv:2211.06134 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.402, + 0.787, + 0.427 + ], + "angle": 0, + "content": "14. Fikes, R.E., Nilsson, N.J.: Strips: A new approach to the application of theorem proving to problem solving. Artificial intelligence 2(3-4), 189-208 (1971)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.427, + 0.787, + 0.464 + ], + "angle": 0, + "content": "15. Fu, Z., Cheng, X., Pathak, D.: Deep whole-body control: learning a unified policy for manipulation and locomotion. In: Conference on Robot Learning. pp. 138-149. PMLR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.464, + 0.787, + 0.501 + ], + "angle": 0, + "content": "16. Gregor, K., Jimenez Rezende, D., Besse, F., Wu, Y., Merzic, H., van den Oord, A.: Shaping belief states with generative environment models for rl. Advances in Neural Information Processing Systems 32 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.501, + 0.787, + 0.526 + ], + "angle": 0, + "content": "17. Gu, J., Chaplot, D.S., Su, H., Malik, J.: Multi-skill mobile manipulation for object rearrangement. arXiv preprint arXiv:2209.02778 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.526, + 0.787, + 0.551 + ], + "angle": 0, + "content": "18. Hafner, D., Lee, K.H., Fischer, I., Abbeel, P.: Deep hierarchical planning from pixels. Advances in Neural Information Processing Systems 35, 26091-26104 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.551, + 0.787, + 0.588 + ], + "angle": 0, + "content": "19. Harutyunyan, A., Dabney, W., Mesnard, T., Gheshlaghi Azar, M., Piot, B., Heess, N., van Hasselt, H.P., Wayne, G., Singh, S., Precup, D., et al.: Hindsight credit assignment. Advances in neural information processing systems 32 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.588, + 0.787, + 0.611 + ], + "angle": 0, + "content": "20. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. CVPR (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.612, + 0.787, + 0.649 + ], + "angle": 0, + "content": "21. Hessel, M., Soyer, H., Espeholt, L., Czarnecki, W., Schmitt, S., Van Hasselt, H.: Multi-task deep reinforcement learning with popart. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 33, pp. 3796-3803 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.649, + 0.787, + 0.674 + ], + "angle": 0, + "content": "22. Huang, X., Batra, D., Rai, A., Szot, A.: Skill transformer: A monolithic policy for mobile manipulation. arXiv preprint arXiv:2308.09873 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.674, + 0.787, + 0.749 + ], + "angle": 0, + "content": "23. Kalashnikov, D., Irpan, A., Pastor, P., Ibarz, J., Herzog, A., Jang, E., Quillen, D., Holly, E., Kalakrishnan, M., Vanhoucke, V., Levine, S.: Scalable deep reinforcement learning for vision-based robotic manipulation. In: 2nd Annual Conference on Robot Learning, CoRL 2018, Zürich, Switzerland, 29-31 October 2018, Proceedings. Proceedings of Machine Learning Research, vol. 87, pp. 651-673. PMLR (2018), http://proceedings.mlr.press/v87 kalashnikov18a.html" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.749, + 0.787, + 0.786 + ], + "angle": 0, + "content": "24. Karkus, P., Mirza, M., Guez, A., Jaegle, A., Lillicrap, T., Buesing, L., Heess, N., Weber, T.: Beyond tabula-rama: a modular reinforcement learning approach for physically embedded 3d sukoban. arXiv preprint arXiv:2010.01298 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.786, + 0.787, + 0.812 + ], + "angle": 0, + "content": "25. Katara, P., Xian, Z., Fragkiadaki, K.: Gen2sim: Scaling up robot learning in simulation with generative models. arXiv preprint arXiv:2310.18308 (2023)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.178, + 0.787, + 0.812 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.149, + 0.237, + 0.159 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.149, + 0.358, + 0.159 + ], + "angle": 0, + "content": "Harish et. al" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.178, + 0.785, + 0.203 + ], + "angle": 0, + "content": "26. Kumar, A., Fu, Z., Pathak, D., Malik, J.: Rapid motor adaptation for legged robots. RSS (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.203, + 0.787, + 0.252 + ], + "angle": 0, + "content": "27. Kuo, C.W., Ma, C.Y., Hoffman, J., Kira, Z.: Structure-encoding auxiliary tasks for improved visual representation in vision-and-language navigation. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 1104-1113 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.253, + 0.785, + 0.29 + ], + "angle": 0, + "content": "28. Lee, Y., Lim, J.J., Anandkumar, A., Zhu, Y.: Adversarial skill chaining for long-horizon robot manipulation via terminal state regularization. arXiv preprint arXiv:2111.07999 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.29, + 0.785, + 0.327 + ], + "angle": 0, + "content": "29. Liu, B., Tur, G., Hakkani-Tur, D., Shah, P., Heck, L.: End-to-end optimization of task-oriented dialogue model with deep reinforcement learning. arXiv preprint arXiv:1711.10712 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.327, + 0.785, + 0.365 + ], + "angle": 0, + "content": "30. Mishra, U.A., Xue, S., Chen, Y., Xu, D.: Generative skill chaining: Long-horizon skill planning with diffusion models. In: Conference on Robot Learning. pp. 2905-2925. PMLR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.365, + 0.785, + 0.402 + ], + "angle": 0, + "content": "31. Mnih, V., Kavukcuoglu, K., Silver, D., Graves, A., Antonoglou, I., Wierstra, D., Riedmiller, M.: Playing atari with deep reinforcement learning. arXiv preprint arXiv:1312.5602 (2013)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.402, + 0.785, + 0.439 + ], + "angle": 0, + "content": "32. Nakano, R., Hilton, J., Balaji, S., Wu, J., Ouyang, L., Kim, C., Hesse, C., Jain, S., Kosaraju, V., Saunders, W., et al.: Webgpt: Browser-assisted question-answering with human feedback. arXiv preprint arXiv:2112.09332 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.439, + 0.785, + 0.476 + ], + "angle": 0, + "content": "33. Narvekar, S., Peng, B., Leonetti, M., Sinapov, J., Taylor, M.E., Stone, P.: Curriculum learning for reinforcement learning domains: A framework and survey. The Journal of Machine Learning Research 21(1), 7382-7431 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.476, + 0.785, + 0.514 + ], + "angle": 0, + "content": "34. Ni, T., Ma, M., Eysenbach, B., Bacon, P.L.: When do transformers shine in rl? decoupling memory from credit assignment. Advances in Neural Information Processing Systems 36 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.514, + 0.785, + 0.551 + ], + "angle": 0, + "content": "35. Qi, H., Kumar, A., Calandra, R., Ma, Y., Malik, J.: In-hand object rotation via rapid motor adaptation. In: Conference on Robot Learning. pp. 1722-1732. PMLR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.551, + 0.785, + 0.576 + ], + "angle": 0, + "content": "36. Radosavovic, I., Xiao, T., Zhang, B., Darrell, T., Malik, J., Sreenath, K.: Learning humanoid locomotion with transformers. arXiv preprint arXiv:2303.03381 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.576, + 0.616, + 0.587 + ], + "angle": 0, + "content": "37. robotics, F.: Fetch. http://fetchrobotics.com/ (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.587, + 0.785, + 0.625 + ], + "angle": 0, + "content": "38. Rudin, N., Hoeller, D., Reist, P., Hutter, M.: Learning to walk in minutes using massively parallel deep reinforcement learning. In: Conference on Robot Learning pp. 91-100. PMLR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.625, + 0.785, + 0.649 + ], + "angle": 0, + "content": "39. Schulman, J., Wolski, F., Dhariwal, P., Radford, A., Klimov, O.: Proximal policy optimization algorithms. arXiv preprint arXiv:1707.06347 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.649, + 0.785, + 0.687 + ], + "angle": 0, + "content": "40. Shah, P., Hakkani-Tür, D., Heck, L.: Interactive reinforcement learning for task-oriented dialogue management. In: NIPS 2016 Deep Learning for Action and Interaction Workshop. vol. 11 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.687, + 0.785, + 0.724 + ], + "angle": 0, + "content": "41. Silver, D., Schrittwieser, J., Simonyan, K., Antonoglou, I., Huang, A., Guez, A., Hubert, T., Baker, L., Lai, M., Bolton, A., et al.: Mastering the game of go without human knowledge. nature 550(7676), 354-359 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.724, + 0.785, + 0.761 + ], + "angle": 0, + "content": "42. Sutton, R.S., Precup, D., Singh, S.: Between mdps and semi-mdps: A framework for temporal abstraction in reinforcement learning. Artificial intelligence 112(1-2), 181-211 (1999)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.761, + 0.785, + 0.811 + ], + "angle": 0, + "content": "43. Szot, A., Clegg, A., Undersander, E., Wijmans, E., Zhao, Y., Turner, J., Maestre, N., Mukadam, M., Chaplot, D.S., Maksymets, O., et al.: Habitat 2.0: Training home assistants to rearrange their habitat. Advances in Neural Information Processing Systems 34 (2021)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.178, + 0.787, + 0.811 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.661, + 0.149, + 0.732, + 0.16 + ], + "angle": 0, + "content": "AuxDistill" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.149, + 0.786, + 0.159 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.177, + 0.788, + 0.216 + ], + "angle": 0, + "content": "44. Szot, A., Yadav, K., Clegg, A., Berges, V.P., Gokaslan, A., Chang, A., Savva, M., Kira, Z., Batra, D.: Habitat rearrangement challenge 2022. https://aihabitat.org/challenge/2022_rearrange (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.217, + 0.787, + 0.254 + ], + "angle": 0, + "content": "45. Touvron, H., Lavril, T., Izacard, G., Martinet, X., Lachaux, M.A., Lacroix, T., Rozière, B., Goyal, N., Hambro, E., Azhar, F., et al.: Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.255, + 0.788, + 0.304 + ], + "angle": 0, + "content": "46. Vezzani, G., Tirumala, D., Wulfmeier, M., Rao, D., Abdelmaleki, A., Moran, B., Haarnoja, T., Humplik, J., Hafner, R., Neunert, M., et al.: Skills: Adaptive skill sequencing for efficient temporally-extended exploration. arXiv preprint arXiv:2211.13743 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.305, + 0.787, + 0.342 + ], + "angle": 0, + "content": "47. Wijmans, E., Kadian, A., Morcos, A., Lee, S., Essa, I., Parikh, D., Savva, M., Batra, D.: Dd-ppo: Learning near-perfect pointgoal navigators from 2.5 billion frames. arXiv preprint arXiv:1911.00357 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.343, + 0.787, + 0.38 + ], + "angle": 0, + "content": "48. Xia, F., Li, C., Martin-Martin, R., Litany, O., Toshev, A., Savarese, S.: Relmogen: Leveraging motion generation in reinforcement learning for mobile manipulation. arXiv preprint arXiv:2008.07792 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.381, + 0.786, + 0.419 + ], + "angle": 0, + "content": "49. Ye, J., Batra, D., Das, A., Wijmans, E.: Auxiliary tasks and exploration enable objectgoal navigation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 16117-16126 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.42, + 0.786, + 0.444 + ], + "angle": 0, + "content": "50. Ye, J., Batra, D., Wijmans, E., Das, A.: Auxiliary tasks speed up learning point goal navigation. In: Conference on Robot Learning. pp. 498-516. PMLR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.445, + 0.786, + 0.469 + ], + "angle": 0, + "content": "51. Zhang, S., Whiteson, S.: Dac: The double actor-critic architecture for learning options. Advances in Neural Information Processing Systems 32 (2019)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.177, + 0.788, + 0.469 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/Reinforcement Learning via Auxillary Task Distillation/0d1f16d8-0d98-408f-b999-123ebb200f38_origin.pdf b/2024/Reinforcement Learning via Auxillary Task Distillation/0d1f16d8-0d98-408f-b999-123ebb200f38_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..97d8ef1d1287d2cbfc9ae8d149468f2adbb7d267 --- /dev/null +++ b/2024/Reinforcement Learning via Auxillary Task Distillation/0d1f16d8-0d98-408f-b999-123ebb200f38_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3b0fda5343862a62187f709b35e5a11f3da3d4b107eca472198f458dea9f8565 +size 4577154 diff --git a/2024/Reinforcement Learning via Auxillary Task Distillation/full.md b/2024/Reinforcement Learning via Auxillary Task Distillation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..a2f67ad4af919bf6a1942380579917ed2aea027e --- /dev/null +++ b/2024/Reinforcement Learning via Auxillary Task Distillation/full.md @@ -0,0 +1,235 @@ +# Reinforcement Learning via Auxiliary Task Distillation + +Abhinav Narayan Harish1, Larry Heck2, Josiah P. Hanna1 +Zsolt Kira2, Andrew Szot2 + +1 University of Wisconsin-Madison +2 Georgia-Tech + +Abstract. We present Reinforcement Learning via Auxiliary Task Distillation (AuxDistill), a new method that enables reinforcement learning (RL) to perform long-horizon robot control problems by distilling behaviors from auxiliary RL tasks. AuxDistill achieves this by concurrently carrying out multi-task RL with auxiliary tasks, which are easier to learn and relevant to the main task. A weighted distillation loss transfers behaviors from these auxiliary tasks to solve the main task. We demonstrate that AuxDistill can learn a pixels-to-actions policy for a challenging multi-stage embodied object rearrangement task from the environment reward without demonstrations, a learning curriculum, or pre-trained skills. AuxDistill achieves $2.3 \times$ higher success than the previous state-of-the-art baseline in the Habitat Object Rearrangement benchmark and outperforms methods that use pre-trained skills and expert demonstrations. + +# 1 Introduction + +While reinforcement learning (RL) is successful in a variety of settings including games [9, 31, 41], chatbots [1, 29, 32, 40, 45], and robotics [3, 23, 35], long-horizon tasks such as embodied object rearrangement where an embodied agent must rearrange objects to target positions still remains a challenge [7]. Object rearrangement requires learning heterogeneous behaviors like picking, navigating, placing, and opening concurrently with sequencing these behaviors to solve the overall task all while operating from egocentric vision. Furthermore, the behaviors are interdependent, meaning the robot can only learn how to pick an object if it has first learned how to open the fridge containing the object. Since object rearrangement requires interacting with objects across house-scale spaces with low-level control, the episodes consist of thousands of low-level control steps, exacerbating the problem of credit assignment [19, 34]. Overall, the problems of concurrently learning low-level control, high-level decision-making, interdependent task stages, and episodes with many time steps make object rearrangement challenging even with dense rewards. While some prior methods have applied end-to-end RL to rearrangement problems, they require excessive experience in simplified simulators [8] or expert skill demonstrations [22]. Other prior work has made such complex tasks more tractable by decomposing the full task into + +a hierarchy of separately trained skills that a high-level policy sequences together to solve the overall task. However, such hierarchical methods suffer from compounding errors between skills [17] and dynamically selecting between skills. + +We therefore present Reinforcement Learning via Auxiliary Task Distillation (AuxDistill), a method for training policies for long-horizon tasks from scratch with end-to-end RL from reward alone. AuxDistill also learns in auxiliary tasks using a multi-task RL framework and transfers the knowledge from these auxiliary tasks to help solve the desired "main" task of interest. Importantly, AuxDistill concurrently learns the main task and all auxiliary tasks, which are all initialized from scratch and not pre-trained. These auxiliary tasks help because they are easier to learn with RL than the main task and contain relevant behaviors for the full task. For example, in object rearrangement, the agent should also practice picking up objects in isolation from the complexities of the full task. Unlike curriculum learning in RL [4,12,13,33], which also leverages easier tasks to learn harder tasks, AuxDistill does not have distinct curriculum phases. Instead, it has a single training objective for distilling sub-behaviors into the overall task behavior. Crucially, the agent end-to-end learns which auxiliary behaviors are relevant for the main task. This relevance is encoded by a scalar weight value determined by the relevance of the robot state to the auxiliary task and it grounded in the oracle task plan. A distillation objective transfers behaviors from relevant auxiliary tasks by encouraging the policy to act consistently between the main and auxiliary tasks for relevant states in the main task. For instance, in object rearrangement, when the robot is near a drawer with the target object inside, it is supervised with the distillation loss from an "open drawer" auxiliary task. Despite the policy also concurrently learning the "open drawer" auxiliary task from scratch, AuxDistill learns the auxiliary tasks faster since they are easier. Thus, the distillation loss is a useful dense per-time step supervision signal that helps overcome the challenges of learning from the main task reward alone. + +We empirically demonstrate that AuxDistill outperforms a variety of baselines in terms of success rate on the home rearrangement benchmark in Habitat 2.0 [43]. These include baselines such as hierarchical RL, end-to-end RL with and without a curriculum, and imitation learning. On rearrangement episodes where objects start in open receptacles, AuxDistill achieves $1.4 \times$ higher success than baselines and $2.6 \times$ higher success in harder episodes where objects can start in closed receptacles. These results highlight the value of AuxDistill to be able to learn the entire rearrangement task end-to-end with RL. We show results beyond object rearrangement on a category-conditioned object manipulation task where AuxDistill achieves $1.75 \times$ higher success than baselines. We also conduct extensive ablation studies where we analyze the properties of auxiliary tasks needed by AuxDistill, and find that AuxDistill is generally robust to the choice of auxiliary tasks. An analysis of the distillation objective reveals the importance of the distillation loss, with the absence of the distillation loss leading to no success on the rearrangement task. All code is available at https://github.com/absdnd/aux_distill + +# 2 Related Work + +Some prior works address complex and long-horizon decision-making problems using a hierarchical breakdown of skills. One such category of approaches is option-critic methods [5, 42, 51], which seek to learn options to temporally abstract the high-level policy decision making. However, discovering and learning such options is unstable and results in a challenging credit assignment problem. Another line of work first trains or is given pre-trained skills and then learns a high-level policy to sequence these skills together to complete longer tasks [11, 18, 24, 48]. Using such a hierarchy results in compounding errors resulting from sequencing skills that were not trained to properly transition between each other [10, 17, 28, 30, 46]. Our approach does not utilize a hierarchical policy, and can solve long horizon tasks better than hierarchical methods by utilizing a single policy and thus avoiding compounding errors that emerge by sequencing independently trained policies. + +Like our method, prior works have tackled model-free end-to-end RL to solve long, complex tasks. Scaling PPO [39] training has enabled robots to learn complex locomotion behaviors [2, 15, 25, 26, 36, 38], manipulate unseen objects in a robotic hand [35], and play competitive video games [9]. Likewise, berges2023galactic showed scaling end-to-end RL can learn object rearrangement. However, their approach requires training agents for billions of environment steps with a simplified kinematic physics simulation. We show results in the full dynamic Habitat simulation, achieving better results in an order of magnitude fewer environment interactions, and show results in harder rearrangement tasks. Other works have incorporated auxiliary objectives in RL to boost performance in embodied tasks [16, 27, 49, 50]. These works add auxiliary self-supervised prediction objectives, whereas our method adds an auxiliary policy learning task via multi-task RL. jia2022improving also uses easier auxiliary RL tasks that specialist agents learn to complete and then distill back into a generalist agent. While our work also uses RL in easier tasks to boost performance, it doesn't require alternating between learning specialist and generalist policies. + +Works have also attempted to learn a curriculum to leverage easier tasks to learn harder tasks that require composing multiple behaviors. Some works learn curriculum generation policies that adjust the environment to suit the agent's capabilities [4, 12, 13]. In other works, a curriculum naturally emerges as a result of downstream training [6]. Finally, some works hand design curricula that change aspects of the environment, such as the starting and goal distributions depending on the agent performance during training [38]. Our work shares a similar spirit of leveraging easier auxiliary tasks to learn a harder task but it does not enforce an explicit curriculum and instead learns all tasks simultaneously. + +# 3 Method + +We present Reinforcement Learning via Auxiliary Task Distillation (AuxDistill), a new method for training long-horizon policies from scratch using rewards alone. Learning from reward alone in complex tasks like embodied object rearrangement + +![](images/0bbc9144d06b05ec8141d6a4d778452164d5b7a3f68cd50eb7f82b46b6b8ae57.jpg) +Fig. 1: AuxDistill learns a rearrangement policy operating from egocentric depth perception and coordinate-based task specification. The full object rearrangement task decomposes into modular abilities that can be learned by auxiliary task with indicator vectors $T_{1} \cdots T_{N}$ which are trained along with the main task using end-to-end RL. The distillation loss is computed as a weighted combination of the task relevance of $o_{t}$ in the main task $T_{0}$ under all auxiliary tasks. The task relevance function computes $w_{i}(s_{t})$ based on the relevance of the current observation and robot state to the auxiliary task $T_{i}$ . The distillation loss and RL-training loss are then used to update the policy. + +is challenging because an agent needs to combine thousands of low-level actions controlling the arm and base, operate from egocentric visual perception, and dynamically sequence distinct behaviors such as navigating, picking, placing, and opening. AuxDistill addresses these challenges of RL in complex problems, like embodied rearrangement, by learning to leverage knowledge from easier auxiliary tasks related to the desired task we are trying to solve. Unlike a curriculum, which learns gradually harder tasks in stages, AuxDistill learns the desired task concurrently with the auxiliary task and includes a novel distillation mechanism to transfer knowledge from easier to harder tasks. + +# 3.1 Preliminaries + +Our problem is formulated as a goal-specified Partially-Observable Markov Decision Process (POMDP) defined as a tuple $\mathcal{M} = (\mathcal{S},\mathcal{O},\mathcal{A},\mathcal{P},\mathcal{R},\mathcal{G},\rho,\gamma)$ with underlying state space $\mathcal{S}$ , observation space $\mathcal{O}$ , action space $\mathcal{A}$ , transition function $\mathcal{P}$ , reward function $\mathcal{R}$ , goal space $\mathcal{G}$ , initial state distribution $\rho$ and discount factor $\gamma$ . For a task like rearrangement, the goal space $\mathcal{G}$ is specified using the 3D coordinate of the object's start location and the goal location where the object has to be placed. Our objective is to learn a goal-conditioned policy $\pi(a|o,g)$ mapping an observation $o$ and goal $g$ to an action $a$ that maximizes the sum of discounted rewards $\mathbb{E}_{s_0\sim\rho_0,g\sim\mathcal{G}}\sum_t\gamma^t\mathcal{R}(s_t,g)$ . + +# 3.2 Reinforcement Learning via Auxiliary Task Distillation + +The core idea of AuxDistill is to learn a policy in a difficult desired "main" task by transferring knowledge from easier auxiliary tasks. This is done without using an explicit curriculum by learning the main and auxiliary tasks in a single training loop. We refer to the main task we wish to learn as $\mathcal{M}_0$ with a task identifier $T_0$ . We define $N$ auxiliary tasks, defined as $\{\mathcal{M}_i\}_{n=1}^N$ with task identifiers $\{T_i\}_{n=1}^N$ . The auxiliary tasks $\{\mathcal{M}_i\}$ share the same state, observation, goal, and action space as the target task $\mathcal{M}_0$ . Each $\mathcal{M}_i$ has a separately defined starting state distribution and reward function. We assume these auxiliary tasks are related to the full task yet are easier to solve than the full task. Specifically, the auxiliary tasks can be easier instantiations or sub-parts of the overall task. For example, in rearrangement, we define the auxiliary tasks in terms of interactions the agent needs to complete the entire rearrangement episode, such as picking, placing, and opening. Prior works use similar task definitions to train skills for hierarchical policies in rearrangement [17], but these works suffer from a two-stage pipeline of first needing to separately train each skill and then decide how to combine them. AuxDistill directly learns the full task from scratch while also performing better. + +AuxDistill learns a single policy with RL that concurrently learns to perform the main and auxiliary tasks. We illustrate the implementation of AuxDistill in Figure 1. This policy, parameterized by $\theta$ , is expressed as $\pi_{\theta}(a_t \mid o_t, g, T)$ for observation $o_t$ , episode goal $g$ , and per- $\mathcal{M}$ task identifier $T$ which is encoded as a one-hot embedding in a vector which has the same size as the maximum number of auxiliary tasks. Note that since all tasks share the same observation and action space, $\pi_{\theta}$ can act and observe in all tasks based on the input task identifier $T$ . + +AuxDistill updates the policy based on an average RL loss from the main task and all auxiliary tasks. Let $\mathcal{L}_{\mathcal{M}_i}^{\mathrm{RL}}(\theta)$ denote the RL loss for $\pi_{\theta}$ in MDP $\mathcal{M}_i$ . These auxiliary tasks are designed to capture a subset of the main task which are easier to accomplish than the main task. To compute these losses, we assume we can collect experience in the auxiliary tasks. For example, to collect experience in a "pick object" auxiliary task in object rearrangement, the robot is spawned close to the object. We then update based on the average of the task and auxiliary task losses: $\frac{1}{N}\sum_{i=0}^{N}\mathcal{L}_{\mathcal{M}_i}^{\mathrm{RL}}(\theta)$ . Intuitively, $\pi_{\theta}$ will first learn to complete the easier auxiliary tasks, and this auxiliary task competency can aid in solving the main task. This shares a similar insight as curriculum learning, except all tasks are learned concurrently, and the curriculum stages are naturally induced by the policy naturally learning on easier tasks first. + +In addition to optimizing the average RL loss between the main and auxiliary tasks, AuxDistill also optimizes a distillation loss that encourages the policy to transfer relevant knowledge from the auxiliary tasks to the main task. To achieve this, for a particular observation $o_t$ , we consider the policy distribution $\pi_\theta(\cdot | o_t, g, T)$ under different task identifiers $T$ . For compactness, notate $\pi_\theta(\cdot | o_t, g, T_i)$ where $T_i$ indicates the task identifier for $\mathcal{M}_i$ as $\pi_\theta^{T_i}(o_t)$ . We want $\pi_\theta^{T_0}$ (the policy in the main task) to match the behaviors from the policy in the relevant auxiliary tasks $\pi_\theta^{T_i}$ . We measure this relevance of an auxiliary task $\mathcal{M}_i$ to the main task at time step $t$ via an auxiliary task relevance function $w_i(s_t)$ . + +This function denotes how much the knowledge from $\mathcal{M}_i$ should apply to $\mathcal{M}_0$ in the underlying simulator state $s_t$ . This relevance function can be grounded in the task plan generated by an oracle planner. For example, consider the state of the robot before picking up the object. In this state, the pick auxiliary task would be relevant to the main task, and the place auxiliary task would not. If an episode requires opening a fridge before accessing the object, the relevant task before the fridge is opened would be open-fridge. Note that computing the distillation weight can utilize oracle knowledge of the simulator state (for instance, whether the object is inside a fridge or a cabinet) since this information is only provided as a training signal and not used during inference. This information is utilized by all methods we compare against, including our strongest baseline [22]. + +We can then distill experience from $\mathcal{M}_i$ by supervising the policy to match the action distribution of the target task $\mathcal{M}_0$ . The distillation loss is computed as the KL-divergence of the action distribution under $\mathcal{M}_i$ and $\mathcal{M}_0$ weighted by the relevance of auxiliary task $\mathcal{M}_i$ given by $w_i(s_t)$ . The overall loss function of AuxDistill, including the distillation loss, is then: + +$$ +\frac {1}{N} \sum_ {i = 0} ^ {N} \mathcal {L} _ {\mathcal {M} _ {i}} ^ {\mathrm {R L}} (\theta) + \lambda \sum_ {i = 1} ^ {N} w _ {i} \left(s _ {t}\right) D _ {\mathrm {K L}} \left(\pi_ {\theta} ^ {T _ {0}} \left(o _ {t}\right) \parallel \pi_ {\theta} ^ {T _ {i}} \left(o _ {t}\right)\right) \tag {1} +$$ + +Here $\lambda$ represents the distillation weight relative to other PPO losses. Note that our formulation relies on the auxiliary tasks being easier to solve and relevant to the main task. Distillation using the task relevance function offers an additional supervisory signal that the policy can optimize along with its own reward. This doesn't impose the large overhead of hierarchical RL methods where each auxiliary task must be pre-trained in isolation with its own reward function and goal specification. Auxiliary tasks are flexibly incorporated to address parts of the task that are challenging for the agent to learn. + +# 3.3 Implementation Details + +We use PPO [39] to compute the RL loss $\mathcal{L}_{\mathcal{M}_i}^{\mathrm{RL}}$ in Equation (1). We create $M$ environment instances for each of the main and $N$ auxiliary tasks and vectorize them for parallel action execution. We rollout the policy in all these $M(N + 1)$ environments and collect a batch of data. We then compute Equation (1), using the PPO loss, update the policy, and then repeat this process. + +The policy is represented as a 2-layer LSTM network with 512 hidden units per layer, similar to the architecture employed in [43]. In our experiments, $o_t$ is a depth or RGB egocentric image, which we encode with a ResNet50 [20] network. The goal and the proprioceptive state of the agent are concatenated with the visual embeddings along with the index of the POMDP $\mathcal{M}_i$ , for which we use a one-hot embedding. This vector is then passed as an input to the LSTM. A linear projection on the output of the LSTM then produces an action to execute on the next step in the environment and a vector of value-predictions corresponding to each of the tasks $\mathcal{M}_i$ used during training. + +
MethodTrain (Seen)Eval (Unseen)
All EpisodesEasyHardAll EpisodesEasyHard
M3 (Oracle) [17]27 ± 057 ± 212 ± 128 ± 058 ± 213 ± 2
M3 [17]25 ± 156 ± 19 ± 213 ± 153 ± 40 ± 0
Galactic [8]-37 ± 0--26 ± 0-
Monolithic RL0 ± 00 ± 00 ± 00 ± 00 ± 00 ± 0
Skill Transformer [22]25 ± 144 ± 116 ± 123 ± 137 ± 116 ± 1
AuxDistill (No Distillation)0 ± 00 ± 00 ± 00 ± 00 ± 00 ± 0
RL Curriculum0 ± 01 ± 00 ± 00 ± 01 ± 10 ± 0
AuxDistill49 ± 274 ± 236 ± 252 ± 275 ± 241 ± 2
+ +Table 1: Success rates on the rearrangement task for our method, AuxDistill (highlighted in blue), and baselines. Displayed are the average and standard deviations for 3 seeds for M3, Monolithic RL, Skill Transformer, and M3 Oracle (numbers from [22]), and 3 seeds for the remaining methods with the highest numbers per setting bolded. Numbers in the easy and hard columns are averages over 100 episodes and 200 episodes, respectively. The All Episodes column is an average across both the Easy and Hard episodes. AuxDistill outperforms all baselines in all settings. + +Our policy is learned via a multi-task RL formulation, with the main and auxiliary tasks being concurrently learned. Each of the tasks may have different reward magnitudes and relative performance during training. To address this, we use PopArt [21], with $\beta = 3e^{-4}$ to enable learning across different return scales. Additional AuxDistill details are in Appendix A of the supplementary. + +# 4 Experiments + +# 4.1 Object-Rearrangement + +In this section, we compare AuxDistill with baselines on the Habitat 2.0 Object Rearrangement task [43]. For comparison with baselines, we run our experiments using the setup from [22]. In this task, a Fetch robot [37] must move an object from a specified start position to a desired goal position in an indoor home environment using only onboard sensing. The agent has no privileged information like existing maps of the environment, 3D object models, or exact object positions. The Fetch robot senses the world through a $256 \times 256$ depth camera mounted on the robot's head, robot joint positions, gripper state, and base egomotion, giving the relative position of the robot from the start of the episode. The task is specified by a starting 3D object coordinate for the object to move and a 3D goal coordinate to move the object to. Both coordinates are specified relative to the robot's position at the start of the episode. Only the starting object coordinate is specified, and this information is not updated based on the current object position (e.g. if it is moved). + +The robot interacts with the world via a 7DoF arm, a suction gripper attached to the end of the arm, and a mobile base. The episode is successful if the target + +object is within $15\mathrm{cm}$ of the goal position. The robot has a budget of 1,500 steps to complete the task. + +We report performance on the easy and hard evaluation episodes from [22]. In easy episodes, both the target object and goal are on an open receptacle. This means the robot does not have to first open a receptacle before accessing the target object or goal. Instead, the agent can always execute the same sequence of navigate to object, pick up object, navigate to goal, and place object at goal. In hard episodes, the object may start in a closed receptacle. The agent, therefore, needs to use its visual input to perceive if the target object is in a closed receptacle. If so, the robot must then open the receptacle before picking the object. The object may start in either the fridge or a cabinet. + +Training Setup: We train AuxDistill using 11,791 episodes with a mix of easy and hard episodes across 63 scenes using the same rearrangement training dataset as in [44]. These episodes were obtained by sampling an equal number of episodes in the easy and hard categories, with easy episodes being equally sampled across episodes with open-cabinet, open-fridge and non-articulated episodes. For hard episodes, sampling is done uniformly between closed fridge and closed cabinet episodes. We use auxiliary tasks covering abilities included as a part of the standard rearrangement benchmark used in [44]. The auxiliary tasks are: + +- Pick: The agent spawns randomly within the house and must navigate to and pick up the object. +- Place the agent is spawned randomly in the house with the object in its gripper. The robot must navigate to and place the object within $15\mathrm{cm}$ of the target 3D location on the receptacle. +- Open Fridge: The robot spawns randomly within the house and must navigate to the fridge and use its base and arm to open the fridge door. +- Open Cabinet: The robot spawns randomly in the house and must navigate to the kitchen area and open the cabinet using its base and arm. +- Pick from Fridge: This task is similar to Pick except the agent has to pick up the object from inside an open fridge, which requires careful manipulation to minimize collisions. + +The auxiliary relevance function is grounded in the oracle task plan [14] of the episode. This is formulated as an indicator function with $w_{i}(s_{t}) = 1$ if the state $s_t$ is relevant to $\mathcal{M}_i$ based on the oracle task plan of the episode and $w_{i}(s_{t}) = 0$ otherwise. + +For example, if at step $t$ the agent is not holding an object and the object is in an open receptacle, the agent must first pick up the object so $w_{i_{\text{pick}}}(s_t) = 1$ , and $w_i(s_t) = 0$ for all other skills. For a complete description of the auxiliary tasks, including their reward functions, along with the precise auxiliary relevance function definition in object rearrangement, see Appendix B. We train methods for 475M steps of environment interactions. We found this to be a sufficient number of environment interactions to ensure sufficient progress on auxiliary tasks for distillation to aid rearrangement performance on both the easy and hard episodes. For AuxDistill, we count the environment interactions in the auxiliary + +
MethodTrain (Seen)Eval (Unseen)
All EpisodesEasyHardAll EpisodesEasyHard
All Skills283619303921
No Pick From Fridge2032712232
No Open Fridge414933485739
No Pick000000
+ +Table 2: Robustness to auxiliary task selection. We compare the train and generalization performance of AuxDistill with 4 different auxiliary task selections. Results are averages across 100 episodes on a single seed of training. The train and evaluation episodes are in a simplified setting compared to Table 1 as described in Section 4.3 + +tasks towards this 475M step budget. We train with a learning rate of $3e^{-4}$ and linearly decay the learning rate to 0 over the course of learning. + +Baselines We compare AuxDistill to both hierarchical methods governed by a task plan as well as monolithic baselines, which directly learn a pixels-to-actions policy using sensor observations. We compare to relevant baselines from huang2023skill and two more baselines that use the auxiliary tasks. + +- Monolithic RL: A monolithic neural network is trained to map sensor observations directly to actions trained with end-to-end RL. The policy showed no signs of learning the main task, so we stopped training early after $100M$ steps of training. This outcome is consistent with the monolithic RL baseline in prior works [8, 43]. +- Galactic: [8] An end-to-end RL framework similar to monolithic RL to map sensor observations directly to actions with the policy being trained with a simplified kinematic simulation on over $>1e^{9}$ simulation steps and transferred to dynamic simulation utilized in our setting. +- M3: [17] Each of the skill policies navigate, pick, place, and open are separately trained. These skills are then sequenced using a task planner. +- M3 (Oracle): [17] This uses the same skill training as M3, but uses an oracle task planner. +- Skill Transformer [22]: This method uses pre-trained skills to collect successful rearrangement demonstrations and then train on these demonstrations with imitation learning. +- RL Curriculum This method is trained in two stages: i) First, we train the policy on the auxiliary tasks for $200M$ steps to ensure the relevant auxiliary tasks have good training performance ii) This is followed by a training on the entire rearrangement task for another $300M$ steps. +- AuxDistill: No-Distillation Train AuxDistill under a similar setting as described in Section 4.1 except with the distillation strength $\lambda = 0.0$ . + +For more details on the baselines, see Appendix B.2. + +![](images/9eb27ef133401ffcf64610759d9858dfbd0124833233b2fe93d8aff0d916c27c.jpg) +Fig. 2: Comparison of skill-robustness to different choices of auxiliary task on the hard distribution. Including both the Pick and Pick from Fridge is crucial for successful rearrangement on this distribution. Not utilizing Open-Fridge leads to a boost in rearrangement success. This improvement arises because the open-fridge skill is the easiest of all auxiliary tasks and utilizing it reduces the number of samples for the main task (from the 100M step budget). We discuss the auxiliary task learning curves in Appendix B.3 + +Rearrangement Performance: In Table 1, we compare AuxDistill to baselines in the rearrangement task. On all evaluation settings, AuxDistill outperforms all baselines. AuxDistill shows an absolute improvement of $22\%$ and $25\%$ on the easy and hard episodes respectively over the best performing baseline. Most baselines struggle to achieve any performance on the unseen episodes. Monolithic RL [43] achieves no success in any of the evaluation settings, demonstrating the impracticality of learning directly from the dense reward in the full rearrangement task. Our method also outperforms Galactic [8] despite learning on less than half the number of training samples (475M for AuxDistill vs. $>1e^{9}$ for Galactic) on the easy split with $74\%$ success over $37\%$ on easy episodes during training and $75\%$ vs. $26\%$ on evaluation. Note that Galactic does not report performance on the hard distribution. + +Like AuxDistill, RL Curriculum also utilizes the auxiliary tasks yet finds no success. This demonstrates the value of concurrent training on the main and auxiliary tasks with the distillation loss in AuxDistill. Likewise, AuxDistill (No Distill) achieves no success, illustrating the importance of the distillation loss. The learning curve in Figure 3a as AuxDistill can gradually learn to complete the task with more environment interactions. On the other hand, RL Curriculum and AuxDistill (No-Distill) remain at no success regardless of the number of learning samples. + +AuxDistill also outperforms the strongest baseline, Skill Transformer, by a significant margin with $75\%$ vs. $37\%$ on the unseen easy episodes and $41\%$ vs. $16\%$ on the unseen hard episodes. This demonstrates the advantages of using online, end-to-end RL, rather than solely training offline with demonstrations. As shown in Figure 3a, AuxDistill is able to improve with subsequent environment + +
Method Train (Seen) Eval (Unseen)
AuxDistill12± 211± 1
Monolithic8± 09± 1
AuxDistill (No Distillation)3± 14± 1
RL-Curriculum6± 18± 0
+ +Table 3: Comparison of our method on the Category Pick task with auxiliary tasks on three random seeds. AuxDistill outperforms all methods, including the monolithic baseline, on the unseen evaluation split. Evaluation is conducted over 1000 seen episodes sampled from the training distribution and 1000 held-out episodes from evaluation. + +interactions, yet Skill Transformer is limited by the performance of the expert demonstrations. + +Table 1 also demonstrates that AuxDistill outperforms hierarchical baselines. M3 is able to achieve $53\%$ success on the unseen easy episodes since it utilizes strong pretrained skills. Even in this setting, AuxDistill achieves a higher success of $75\%$ . However, on the hard episodes, M3 cannot dynamically plan skills and achieves no success in the hard unseen episodes. AuxDistill achieves $41\%$ on this setting because it learns a monolithic policy with RL that combines low-level and high-level decision-making. AuxDistill even outperforms an oracle version of M3 that dynamically plans the skill sequence based on oracle state information (13% vs. 41% on the unseen hard episodes). + +We observe that AuxDistill performs $7\%$ worse on the Habitat 2.0 rearrangement training episodes than the evaluation episodes. We find this performance gap is due to additional easier episodes in the evaluation distribution where the object is closer to the target location. In the hard split, there are 55 eval episodes and 29 train episodes where the object is picked up from the fridge and the target location is $< 0.5m$ from the object location. On removing these episodes, train performance is higher than evaluation, with $32\%$ success on train and $31\%$ success on evaluation. + +# 4.2 Category Pick + +The merits of AuxDistill extend to other challenging embodied tasks. In particular, consider a variant of the pick task described in Section 4.1 where a robot has to pick up an object using the object name passed as a one-hot embedding to the policy. Category pick requires the policy to discern the object to pick up by correlating the RGB observation with the object category passed as a one-hot embedding. This task is more challenging than geometric pick where the policy has access to the initial coordinate specification of the object to be picked. We include additional details about the task specification in Section B.4 of our supplementary material. + +We leverage the easier coordinate pick task to aid the learning of the Category Pick task using AuxDistill. More specifically, we consider the task of interest $M_0$ to be the Category Pick task and $M_1$ (auxiliary task we would like to distill from) to be the coordinate pick task. Both tasks are trained jointly using AuxDistill + +![](images/0a95d95ec630ec17738dbe92b8640ed87facb1651b4a55af1dd8e808e9704d23.jpg) +(a) Rearrange learning curve. + +![](images/97823b0f54d475f18f16446b921070e36c9ddb6bc7ed0a89da9d79b090b2be24.jpg) +(b) Category Pick learning. +Fig. 3: Left: RL training success rates on training episodes on the rearrangement task from Table 1. Note that AuxDistill (No Distill) and RL-Curriculum are displayed but achieve $0\%$ success throughout training. Right: Learning curve on the Category Pick task of AuxDistill utilizing coordinate pick distillation vs. monolithic RL. AuxDistill outperforms baselines in both settings. Displayed are averages and standard deviation over 3 random seeds. + +with the same reward structure with $w_{i}(s_{t}) = 1$ and use $\lambda = 0.5$ for distillation in Equation (1). For this task, we train on the full rearrange-easy dataset from [44] with 50,000 episodes. + +We report the training curves of the Category Pick task in comparison with the RL Curriculum and No Distillation baseline used in Table 1. In addition, we introduce a monolithic baseline for this task, which directly trains Category Pick without leveraging the coordinate pick task during training. We report the training curves in Figure 3b. We evaluate the performance of the trained policy on 1000 episodes sampled from the training and 1000 episodes from the validation split. + +In Table 3, we show that AuxDistill shows better performance with a success rate of $11\%$ outperforming all baselines that do not leverage distillation during RL training. The closest performing baseline to ours is the monolithic baseline, which achieves a success of $9\%$ on the held-out distribution. AuxDistill (No-Distillation) and RL-curriculum perform worse with an evaluation success of $4\%$ and $8\%$ , respectively. + +# 4.3 AuxDistill Analysis + +Robustness to Auxiliary Task Selection A critical consideration for training AuxDistill is the selection of auxiliary tasks. In this section, we compare four different selections of auxiliary tasks in the object rearrangement task from Section 4.1. We conduct this ablation on a smaller distribution of episodes, which only include the Easy episodes and Hard episodes where the object starts in the fridge and train AuxDistill for 100M steps. + +As we conduct this study only on the fridge category of articulated episodes, we modify the auxiliary task selection to be: Pick, Open Fridge, Place and Pick + +from Fridge. Note that the first three auxiliary tasks are the same from Section 4.1. The other auxiliary task selections are the same as this original selection, but excluding one of: Pick, Open Fridge, and Pick from Fridge. + +In Table 2, we report the performance for each skill selection on the easy and hard episodes. We observe that among all the sub-tasks, the most important is Pick, as excluding it results in $0.0\%$ success. Note that Pick is only relevant $(w_{i}(s_{t}) > 0)$ to easy episodes during training i.e. removing it should only impact the performance on easy distribution of our training setup. However, we find that it affects both the easy and hard episodes as the task is not successfully carried out in either case if removed. In Figure 2, we analyze the success of individual stages of rearrangement and notice that not including Pick leads to failure earlier on in the task. + +In Table 5 we notice that introducing the Open-Fridge skill can worsen rearrangement performance (30% vs. 48%). The reason for this is that the open-fridge skill is the easiest of all auxiliary tasks (see Appendix B.3), and utilizing it reduces the number of samples allocated to the main task. However, certain auxiliary tasks can boost performance by addressing specific bottlenecks in the rearrangement. One such example is the Pick from Fridge, which boosts performance 7% to 19% by addressing the task of picking from an open fridge, which can be challenging due to the difficulty of avoiding collisions with the fridge door. Further, the 4% performance improvement from 32% to 36% on easy distribution can be attributed to the presence of easy episodes in our dataset with an open-refrigerator where Pick from Fridge helps boost performance. + +Distillation Coefficient Variation We analyze the effect of the distillation loss coefficient $\lambda$ in AuxDistill in Table 4. The distillation loss coefficient controls how the agent balances minimizing the distillation from the auxiliary tasks versus maximizing the reward on the main task. In Table 4, we notice that using high values of distillation weighting $\lambda = 1.0$ shifts the objective of the rearrangement task from maximizing cumulative reward to distilling from the auxiliary task leading to $1\%$ success rate on rearrangement and similarly, using a very small value of distillation coefficient $\lambda = 0.01$ leads to insufficient leveraging of auxiliary task information with a success rate of $7\%$ . We find $\lambda = 0.1, 0.05$ to be good choices for optimizing task reward and distilling behaviors from the auxiliary tasks. + +# 5 Conclusion + +In this work, we presented AuxDistill, a new method for end-to-end RL on complex tasks by leveraging auxiliary tasks. AuxDistill learns in the auxiliary tasks concurrently with the main task through multi-task RL. A distillation loss transfers relevant behaviors from the auxiliary tasks to the main task. We show that AuxDistill outperforms a variety of baselines in Habitat object rearrangement by up to $27\%$ . We also show another application of AuxDistill in a category-conditioned manipulation task. Finally, we analyze AuxDistill in + +
MethodTrain (Seen)Eval (Unseen)
All EpisodesEasyHardAll EpisodesEasyHard
λ = 0.0166510182
λ = 0.05303128363338
λ = 0.1222619303921
λ = 0.5230250
λ = 1.0010120
+ +Table 4: Performance of our method for a single seed on varying distillation coefficient during training. Using a large distillation coefficient $(\lambda = 1.0)$ makes reward optimization challenging, and too small $(\lambda = 0.01)$ results in insufficient auxiliary distillation to succeed on the rearrangement task. The intermediate values $\lambda = 0.05$ and $\lambda = 0.1$ show the best performance on training and evaluation. + +different auxiliary task selections and magnitudes of distillation strengths. Overall, AuxDistill presents a new way to tackle compound tasks with RL alone without requiring pre-trained skills or expert demonstrations. + +Limitations of AuxDistill include the need for the auxiliary tasks and the auxiliary task relevance function. The auxiliary tasks require knowing behaviors that are relevant and easier to learn than the main task. Each auxiliary task requires defining a new start state distribution and reward function. However, designing these can rely on privileged state information which is less restrictive than the requirement of pre-trained skills or expert demonstrations. + +# References + +1. Achiam, J., Adler, S., Agarwal, S., Ahmad, L., Akkaya, I., Aleman, F.L., Almeida, D., Altenschmidt, J., Altman, S., Anadkat, S., et al.: Gpt-4 technical report. arXiv preprint arXiv:2303.08774 (2023) +2. Agarwal, A., Kumar, A., Malik, J., Pathak, D.: Legged locomotion in challenging terrains using egocentric vision. In: Conference on Robot Learning. pp. 403-415. PMLR (2023) +3. Akkaya, I., Andrychowicz, M., Chociej, M., Litwin, M., McGrew, B., Petron, A., Paino, A., Plappert, M., Powell, G., Ribas, R., et al.: Solving rubik's cube with a robot hand. arXiv preprint arXiv:1910.07113 (2019) +4. Azad, A.S., Gur, I., Emhoff, J., Alexis, N., Faust, A., Abbeel, P., Stoica, I.: Clutr: Curriculum learning via unsupervised task representation learning. In: International Conference on Machine Learning. pp. 1361-1395. PMLR (2023) +5. Bacon, P.L., Harb, J., Precup, D.: The option-critic architecture. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 31 (2017) +6. Baker, B., Kanitscheider, I., Markov, T., Wu, Y., Powell, G., McGrew, B., Mordatch, I.: Emergent tool use from multi-agent autocurricula. arXiv preprint arXiv:1909.07528 (2019) +7. Batra, D., Chang, A.X., Chernova, S., Davison, A.J., Deng, J., Koltun, V., Levine, S., Malik, J., Mordatch, I., Mottaghi, R., et al.: Rearrangement: A challenge for embodied ai. arXiv preprint arXiv:2011.01975 (2020) + +8. Berges, V.P., Szot, A., Chaplot, D.S., Gokaslan, A., Mottaghi, R., Batra, D., Undersander, E.: Galactic: Scaling end-to-end reinforcement learning for rearrangement at 100k steps-per-second. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13767-13777 (2023) +9. Berner, C., Brockman, G., Chan, B., Cheung, V., Debiak, P., Dennison, C., Farhi, D., Fischer, Q., Hashme, S., Hesse, C., et al.: Dota 2 with large scale deep reinforcement learning. arXiv preprint arXiv:1912.06680 (2019) +10. Chen, Y., Wang, C., Fei-Fei, L., Liu, C.K.: Sequential dexterity: Chaining dexterous policies for long-horizon manipulation. arXiv preprint arXiv:2309.00987 (2023) +11. Dalal, M., Pathak, D., Salakhutdinov, R.R.: Accelerating robotic reinforcement learning via parameterized action primitives. Advances in Neural Information Processing Systems 34, 21847-21859 (2021) +12. Dennis, M., Jaques, N., Vinitsky, E., Bayen, A., Russell, S., Critch, A., Levine, S.: Emergent complexity and zero-shot transfer via unsupervised environment design. Advances in neural information processing systems 33, 13049-13061 (2020) +13. Fang, K., Migimatsu, T., Mandlekar, A., Fei-Fei, L., Bohg, J.: Active task randomization: Learning visuomotor skills for sequential manipulation by proposing feasible and novel tasks. arXiv preprint arXiv:2211.06134 (2022) +14. Fikes, R.E., Nilsson, N.J.: Strips: A new approach to the application of theorem proving to problem solving. Artificial intelligence 2(3-4), 189-208 (1971) +15. Fu, Z., Cheng, X., Pathak, D.: Deep whole-body control: learning a unified policy for manipulation and locomotion. In: Conference on Robot Learning. pp. 138-149. PMLR (2023) +16. Gregor, K., Jimenez Rezende, D., Besse, F., Wu, Y., Merzic, H., van den Oord, A.: Shaping belief states with generative environment models for rl. Advances in Neural Information Processing Systems 32 (2019) +17. Gu, J., Chaplot, D.S., Su, H., Malik, J.: Multi-skill mobile manipulation for object rearrangement. arXiv preprint arXiv:2209.02778 (2022) +18. Hafner, D., Lee, K.H., Fischer, I., Abbeel, P.: Deep hierarchical planning from pixels. Advances in Neural Information Processing Systems 35, 26091-26104 (2022) +19. Harutyunyan, A., Dabney, W., Mesnard, T., Gheshlaghi Azar, M., Piot, B., Heess, N., van Hasselt, H.P., Wayne, G., Singh, S., Precup, D., et al.: Hindsight credit assignment. Advances in neural information processing systems 32 (2019) +20. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. CVPR (2016) +21. Hessel, M., Soyer, H., Espeholt, L., Czarnecki, W., Schmitt, S., Van Hasselt, H.: Multi-task deep reinforcement learning with popart. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 33, pp. 3796-3803 (2019) +22. Huang, X., Batra, D., Rai, A., Szot, A.: Skill transformer: A monolithic policy for mobile manipulation. arXiv preprint arXiv:2308.09873 (2023) +23. Kalashnikov, D., Irpan, A., Pastor, P., Ibarz, J., Herzog, A., Jang, E., Quillen, D., Holly, E., Kalakrishnan, M., Vanhoucke, V., Levine, S.: Scalable deep reinforcement learning for vision-based robotic manipulation. In: 2nd Annual Conference on Robot Learning, CoRL 2018, Zürich, Switzerland, 29-31 October 2018, Proceedings. Proceedings of Machine Learning Research, vol. 87, pp. 651-673. PMLR (2018), http://proceedings.mlr.press/v87 kalashnikov18a.html +24. Karkus, P., Mirza, M., Guez, A., Jaegle, A., Lillicrap, T., Buesing, L., Heess, N., Weber, T.: Beyond tabula-rama: a modular reinforcement learning approach for physically embedded 3d sukoban. arXiv preprint arXiv:2010.01298 (2020) +25. Katara, P., Xian, Z., Fragkiadaki, K.: Gen2sim: Scaling up robot learning in simulation with generative models. arXiv preprint arXiv:2310.18308 (2023) + +26. Kumar, A., Fu, Z., Pathak, D., Malik, J.: Rapid motor adaptation for legged robots. RSS (2021) +27. Kuo, C.W., Ma, C.Y., Hoffman, J., Kira, Z.: Structure-encoding auxiliary tasks for improved visual representation in vision-and-language navigation. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 1104-1113 (2023) +28. Lee, Y., Lim, J.J., Anandkumar, A., Zhu, Y.: Adversarial skill chaining for long-horizon robot manipulation via terminal state regularization. arXiv preprint arXiv:2111.07999 (2021) +29. Liu, B., Tur, G., Hakkani-Tur, D., Shah, P., Heck, L.: End-to-end optimization of task-oriented dialogue model with deep reinforcement learning. arXiv preprint arXiv:1711.10712 (2017) +30. Mishra, U.A., Xue, S., Chen, Y., Xu, D.: Generative skill chaining: Long-horizon skill planning with diffusion models. In: Conference on Robot Learning. pp. 2905-2925. PMLR (2023) +31. Mnih, V., Kavukcuoglu, K., Silver, D., Graves, A., Antonoglou, I., Wierstra, D., Riedmiller, M.: Playing atari with deep reinforcement learning. arXiv preprint arXiv:1312.5602 (2013) +32. Nakano, R., Hilton, J., Balaji, S., Wu, J., Ouyang, L., Kim, C., Hesse, C., Jain, S., Kosaraju, V., Saunders, W., et al.: Webgpt: Browser-assisted question-answering with human feedback. arXiv preprint arXiv:2112.09332 (2021) +33. Narvekar, S., Peng, B., Leonetti, M., Sinapov, J., Taylor, M.E., Stone, P.: Curriculum learning for reinforcement learning domains: A framework and survey. The Journal of Machine Learning Research 21(1), 7382-7431 (2020) +34. Ni, T., Ma, M., Eysenbach, B., Bacon, P.L.: When do transformers shine in rl? decoupling memory from credit assignment. Advances in Neural Information Processing Systems 36 (2024) +35. Qi, H., Kumar, A., Calandra, R., Ma, Y., Malik, J.: In-hand object rotation via rapid motor adaptation. In: Conference on Robot Learning. pp. 1722-1732. PMLR (2023) +36. Radosavovic, I., Xiao, T., Zhang, B., Darrell, T., Malik, J., Sreenath, K.: Learning humanoid locomotion with transformers. arXiv preprint arXiv:2303.03381 (2023) +37. robotics, F.: Fetch. http://fetchrobotics.com/ (2020) +38. Rudin, N., Hoeller, D., Reist, P., Hutter, M.: Learning to walk in minutes using massively parallel deep reinforcement learning. In: Conference on Robot Learning pp. 91-100. PMLR (2022) +39. Schulman, J., Wolski, F., Dhariwal, P., Radford, A., Klimov, O.: Proximal policy optimization algorithms. arXiv preprint arXiv:1707.06347 (2017) +40. Shah, P., Hakkani-Tür, D., Heck, L.: Interactive reinforcement learning for task-oriented dialogue management. In: NIPS 2016 Deep Learning for Action and Interaction Workshop. vol. 11 (2016) +41. Silver, D., Schrittwieser, J., Simonyan, K., Antonoglou, I., Huang, A., Guez, A., Hubert, T., Baker, L., Lai, M., Bolton, A., et al.: Mastering the game of go without human knowledge. nature 550(7676), 354-359 (2017) +42. Sutton, R.S., Precup, D., Singh, S.: Between mdps and semi-mdps: A framework for temporal abstraction in reinforcement learning. Artificial intelligence 112(1-2), 181-211 (1999) +43. Szot, A., Clegg, A., Undersander, E., Wijmans, E., Zhao, Y., Turner, J., Maestre, N., Mukadam, M., Chaplot, D.S., Maksymets, O., et al.: Habitat 2.0: Training home assistants to rearrange their habitat. Advances in Neural Information Processing Systems 34 (2021) + +44. Szot, A., Yadav, K., Clegg, A., Berges, V.P., Gokaslan, A., Chang, A., Savva, M., Kira, Z., Batra, D.: Habitat rearrangement challenge 2022. https://aihabitat.org/challenge/2022_rearrange (2022) +45. Touvron, H., Lavril, T., Izacard, G., Martinet, X., Lachaux, M.A., Lacroix, T., Rozière, B., Goyal, N., Hambro, E., Azhar, F., et al.: Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971 (2023) +46. Vezzani, G., Tirumala, D., Wulfmeier, M., Rao, D., Abdelmaleki, A., Moran, B., Haarnoja, T., Humplik, J., Hafner, R., Neunert, M., et al.: Skills: Adaptive skill sequencing for efficient temporally-extended exploration. arXiv preprint arXiv:2211.13743 (2022) +47. Wijmans, E., Kadian, A., Morcos, A., Lee, S., Essa, I., Parikh, D., Savva, M., Batra, D.: Dd-ppo: Learning near-perfect pointgoal navigators from 2.5 billion frames. arXiv preprint arXiv:1911.00357 (2019) +48. Xia, F., Li, C., Martin-Martin, R., Litany, O., Toshev, A., Savarese, S.: Relmogen: Leveraging motion generation in reinforcement learning for mobile manipulation. arXiv preprint arXiv:2008.07792 (2020) +49. Ye, J., Batra, D., Das, A., Wijmans, E.: Auxiliary tasks and exploration enable objectgoal navigation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 16117-16126 (2021) +50. Ye, J., Batra, D., Wijmans, E., Das, A.: Auxiliary tasks speed up learning point goal navigation. In: Conference on Robot Learning. pp. 498-516. PMLR (2021) +51. Zhang, S., Whiteson, S.: Dac: The double actor-critic architecture for learning options. Advances in Neural Information Processing Systems 32 (2019) \ No newline at end of file diff --git a/2024/Reinforcement Learning via Auxillary Task Distillation/images.zip b/2024/Reinforcement Learning via Auxillary Task Distillation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..23e5a93bc87dca5d46add9f6ae2584dbfe9fb5ca --- /dev/null +++ b/2024/Reinforcement Learning via Auxillary Task Distillation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ec364e76e063fa5ef30b099bd16476299993649fb3e1db177db67f571ddd3921 +size 223362 diff --git a/2024/Reinforcement Learning via Auxillary Task Distillation/layout.json b/2024/Reinforcement Learning via Auxillary Task Distillation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..a98af66946413f6d65994e0a592c0ac177c9a889 --- /dev/null +++ b/2024/Reinforcement Learning via Auxillary Task Distillation/layout.json @@ -0,0 +1,8904 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 152, + 144, + 446, + 178 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 144, + 446, + 178 + ], + "spans": [ + { + "bbox": [ + 152, + 144, + 446, + 178 + ], + "type": "text", + "content": "Reinforcement Learning via Auxiliary Task Distillation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 173, + 199, + 422, + 222 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 173, + 199, + 422, + 222 + ], + "spans": [ + { + "bbox": [ + 173, + 199, + 422, + 222 + ], + "type": "text", + "content": "Abhinav Narayan Harish1, Larry Heck2, Josiah P. Hanna1 \nZsolt Kira2, Andrew Szot2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 227, + 231, + 367, + 253 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 227, + 231, + 367, + 253 + ], + "spans": [ + { + "bbox": [ + 227, + 231, + 367, + 253 + ], + "type": "text", + "content": "1 University of Wisconsin-Madison \n2 Georgia-Tech" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 155, + 280, + 442, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 155, + 280, + 442, + 430 + ], + "spans": [ + { + "bbox": [ + 155, + 280, + 442, + 430 + ], + "type": "text", + "content": "Abstract. We present Reinforcement Learning via Auxiliary Task Distillation (AuxDistill), a new method that enables reinforcement learning (RL) to perform long-horizon robot control problems by distilling behaviors from auxiliary RL tasks. AuxDistill achieves this by concurrently carrying out multi-task RL with auxiliary tasks, which are easier to learn and relevant to the main task. A weighted distillation loss transfers behaviors from these auxiliary tasks to solve the main task. We demonstrate that AuxDistill can learn a pixels-to-actions policy for a challenging multi-stage embodied object rearrangement task from the environment reward without demonstrations, a learning curriculum, or pre-trained skills. AuxDistill achieves " + }, + { + "bbox": [ + 155, + 280, + 442, + 430 + ], + "type": "inline_equation", + "content": "2.3 \\times" + }, + { + "bbox": [ + 155, + 280, + 442, + 430 + ], + "type": "text", + "content": " higher success than the previous state-of-the-art baseline in the Habitat Object Rearrangement benchmark and outperforms methods that use pre-trained skills and expert demonstrations." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 129, + 449, + 224, + 461 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 129, + 449, + 224, + 461 + ], + "spans": [ + { + "bbox": [ + 129, + 449, + 224, + 461 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 127, + 472, + 469, + 684 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 127, + 472, + 469, + 684 + ], + "spans": [ + { + "bbox": [ + 127, + 472, + 469, + 684 + ], + "type": "text", + "content": "While reinforcement learning (RL) is successful in a variety of settings including games [9, 31, 41], chatbots [1, 29, 32, 40, 45], and robotics [3, 23, 35], long-horizon tasks such as embodied object rearrangement where an embodied agent must rearrange objects to target positions still remains a challenge [7]. Object rearrangement requires learning heterogeneous behaviors like picking, navigating, placing, and opening concurrently with sequencing these behaviors to solve the overall task all while operating from egocentric vision. Furthermore, the behaviors are interdependent, meaning the robot can only learn how to pick an object if it has first learned how to open the fridge containing the object. Since object rearrangement requires interacting with objects across house-scale spaces with low-level control, the episodes consist of thousands of low-level control steps, exacerbating the problem of credit assignment [19, 34]. Overall, the problems of concurrently learning low-level control, high-level decision-making, interdependent task stages, and episodes with many time steps make object rearrangement challenging even with dense rewards. While some prior methods have applied end-to-end RL to rearrangement problems, they require excessive experience in simplified simulators [8] or expert skill demonstrations [22]. Other prior work has made such complex tasks more tractable by decomposing the full task into" + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [], + "page_size": [ + 595, + 841 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 127, + 148, + 468, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 127, + 148, + 468, + 185 + ], + "spans": [ + { + "bbox": [ + 127, + 148, + 468, + 185 + ], + "type": "text", + "content": "a hierarchy of separately trained skills that a high-level policy sequences together to solve the overall task. However, such hierarchical methods suffer from compounding errors between skills [17] and dynamically selecting between skills." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 127, + 192, + 469, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 127, + 192, + 469, + 495 + ], + "spans": [ + { + "bbox": [ + 127, + 192, + 469, + 495 + ], + "type": "text", + "content": "We therefore present Reinforcement Learning via Auxiliary Task Distillation (AuxDistill), a method for training policies for long-horizon tasks from scratch with end-to-end RL from reward alone. AuxDistill also learns in auxiliary tasks using a multi-task RL framework and transfers the knowledge from these auxiliary tasks to help solve the desired \"main\" task of interest. Importantly, AuxDistill concurrently learns the main task and all auxiliary tasks, which are all initialized from scratch and not pre-trained. These auxiliary tasks help because they are easier to learn with RL than the main task and contain relevant behaviors for the full task. For example, in object rearrangement, the agent should also practice picking up objects in isolation from the complexities of the full task. Unlike curriculum learning in RL [4,12,13,33], which also leverages easier tasks to learn harder tasks, AuxDistill does not have distinct curriculum phases. Instead, it has a single training objective for distilling sub-behaviors into the overall task behavior. Crucially, the agent end-to-end learns which auxiliary behaviors are relevant for the main task. This relevance is encoded by a scalar weight value determined by the relevance of the robot state to the auxiliary task and it grounded in the oracle task plan. A distillation objective transfers behaviors from relevant auxiliary tasks by encouraging the policy to act consistently between the main and auxiliary tasks for relevant states in the main task. For instance, in object rearrangement, when the robot is near a drawer with the target object inside, it is supervised with the distillation loss from an \"open drawer\" auxiliary task. Despite the policy also concurrently learning the \"open drawer\" auxiliary task from scratch, AuxDistill learns the auxiliary tasks faster since they are easier. Thus, the distillation loss is a useful dense per-time step supervision signal that helps overcome the challenges of learning from the main task reward alone." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 127, + 503, + 470, + 690 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 127, + 503, + 470, + 690 + ], + "spans": [ + { + "bbox": [ + 127, + 503, + 470, + 690 + ], + "type": "text", + "content": "We empirically demonstrate that AuxDistill outperforms a variety of baselines in terms of success rate on the home rearrangement benchmark in Habitat 2.0 [43]. These include baselines such as hierarchical RL, end-to-end RL with and without a curriculum, and imitation learning. On rearrangement episodes where objects start in open receptacles, AuxDistill achieves " + }, + { + "bbox": [ + 127, + 503, + 470, + 690 + ], + "type": "inline_equation", + "content": "1.4 \\times" + }, + { + "bbox": [ + 127, + 503, + 470, + 690 + ], + "type": "text", + "content": " higher success than baselines and " + }, + { + "bbox": [ + 127, + 503, + 470, + 690 + ], + "type": "inline_equation", + "content": "2.6 \\times" + }, + { + "bbox": [ + 127, + 503, + 470, + 690 + ], + "type": "text", + "content": " higher success in harder episodes where objects can start in closed receptacles. These results highlight the value of AuxDistill to be able to learn the entire rearrangement task end-to-end with RL. We show results beyond object rearrangement on a category-conditioned object manipulation task where AuxDistill achieves " + }, + { + "bbox": [ + 127, + 503, + 470, + 690 + ], + "type": "inline_equation", + "content": "1.75 \\times" + }, + { + "bbox": [ + 127, + 503, + 470, + 690 + ], + "type": "text", + "content": " higher success than baselines. We also conduct extensive ablation studies where we analyze the properties of auxiliary tasks needed by AuxDistill, and find that AuxDistill is generally robust to the choice of auxiliary tasks. An analysis of the distillation objective reveals the importance of the distillation loss, with the absence of the distillation loss leading to no success on the rearrangement task. All code is available at https://github.com/absdnd/aux_distill" + } + ] + } + ], + "index": 4 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 129, + 125, + 136, + 133 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 129, + 125, + 136, + 133 + ], + "spans": [ + { + "bbox": [ + 129, + 125, + 136, + 133 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 161, + 125, + 213, + 134 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 161, + 125, + 213, + 134 + ], + "spans": [ + { + "bbox": [ + 161, + 125, + 213, + 134 + ], + "type": "text", + "content": "Harish et. al" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 128, + 147, + 231, + 159 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 128, + 147, + 231, + 159 + ], + "spans": [ + { + "bbox": [ + 128, + 147, + 231, + 159 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 127, + 173, + 468, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 127, + 173, + 468, + 312 + ], + "spans": [ + { + "bbox": [ + 127, + 173, + 468, + 312 + ], + "type": "text", + "content": "Some prior works address complex and long-horizon decision-making problems using a hierarchical breakdown of skills. One such category of approaches is option-critic methods [5, 42, 51], which seek to learn options to temporally abstract the high-level policy decision making. However, discovering and learning such options is unstable and results in a challenging credit assignment problem. Another line of work first trains or is given pre-trained skills and then learns a high-level policy to sequence these skills together to complete longer tasks [11, 18, 24, 48]. Using such a hierarchy results in compounding errors resulting from sequencing skills that were not trained to properly transition between each other [10, 17, 28, 30, 46]. Our approach does not utilize a hierarchical policy, and can solve long horizon tasks better than hierarchical methods by utilizing a single policy and thus avoiding compounding errors that emerge by sequencing independently trained policies." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 127, + 312, + 469, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 127, + 312, + 469, + 498 + ], + "spans": [ + { + "bbox": [ + 127, + 312, + 469, + 498 + ], + "type": "text", + "content": "Like our method, prior works have tackled model-free end-to-end RL to solve long, complex tasks. Scaling PPO [39] training has enabled robots to learn complex locomotion behaviors [2, 15, 25, 26, 36, 38], manipulate unseen objects in a robotic hand [35], and play competitive video games [9]. Likewise, berges2023galactic showed scaling end-to-end RL can learn object rearrangement. However, their approach requires training agents for billions of environment steps with a simplified kinematic physics simulation. We show results in the full dynamic Habitat simulation, achieving better results in an order of magnitude fewer environment interactions, and show results in harder rearrangement tasks. Other works have incorporated auxiliary objectives in RL to boost performance in embodied tasks [16, 27, 49, 50]. These works add auxiliary self-supervised prediction objectives, whereas our method adds an auxiliary policy learning task via multi-task RL. jia2022improving also uses easier auxiliary RL tasks that specialist agents learn to complete and then distill back into a generalist agent. While our work also uses RL in easier tasks to boost performance, it doesn't require alternating between learning specialist and generalist policies." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 127, + 498, + 469, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 127, + 498, + 469, + 602 + ], + "spans": [ + { + "bbox": [ + 127, + 498, + 469, + 602 + ], + "type": "text", + "content": "Works have also attempted to learn a curriculum to leverage easier tasks to learn harder tasks that require composing multiple behaviors. Some works learn curriculum generation policies that adjust the environment to suit the agent's capabilities [4, 12, 13]. In other works, a curriculum naturally emerges as a result of downstream training [6]. Finally, some works hand design curricula that change aspects of the environment, such as the starting and goal distributions depending on the agent performance during training [38]. Our work shares a similar spirit of leveraging easier auxiliary tasks to learn a harder task but it does not enforce an explicit curriculum and instead learns all tasks simultaneously." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 128, + 621, + 196, + 634 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 128, + 621, + 196, + 634 + ], + "spans": [ + { + "bbox": [ + 128, + 621, + 196, + 634 + ], + "type": "text", + "content": "3 Method" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 127, + 647, + 468, + 683 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 127, + 647, + 468, + 683 + ], + "spans": [ + { + "bbox": [ + 127, + 647, + 468, + 683 + ], + "type": "text", + "content": "We present Reinforcement Learning via Auxiliary Task Distillation (AuxDistill), a new method for training long-horizon policies from scratch using rewards alone. Learning from reward alone in complex tasks like embodied object rearrangement" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 392, + 125, + 435, + 134 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 392, + 125, + 435, + 134 + ], + "spans": [ + { + "bbox": [ + 392, + 125, + 435, + 134 + ], + "type": "text", + "content": "AuxDistill" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 461, + 125, + 467, + 133 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 461, + 125, + 467, + 133 + ], + "spans": [ + { + "bbox": [ + 461, + 125, + 467, + 133 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 130, + 147, + 466, + 283 + ], + "blocks": [ + { + "bbox": [ + 130, + 147, + 466, + 283 + ], + "lines": [ + { + "bbox": [ + 130, + 147, + 466, + 283 + ], + "spans": [ + { + "bbox": [ + 130, + 147, + 466, + 283 + ], + "type": "image", + "image_path": "0bbc9144d06b05ec8141d6a4d778452164d5b7a3f68cd50eb7f82b46b6b8ae57.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 127, + 293, + 468, + 380 + ], + "lines": [ + { + "bbox": [ + 127, + 293, + 468, + 380 + ], + "spans": [ + { + "bbox": [ + 127, + 293, + 468, + 380 + ], + "type": "text", + "content": "Fig. 1: AuxDistill learns a rearrangement policy operating from egocentric depth perception and coordinate-based task specification. The full object rearrangement task decomposes into modular abilities that can be learned by auxiliary task with indicator vectors " + }, + { + "bbox": [ + 127, + 293, + 468, + 380 + ], + "type": "inline_equation", + "content": "T_{1} \\cdots T_{N}" + }, + { + "bbox": [ + 127, + 293, + 468, + 380 + ], + "type": "text", + "content": " which are trained along with the main task using end-to-end RL. The distillation loss is computed as a weighted combination of the task relevance of " + }, + { + "bbox": [ + 127, + 293, + 468, + 380 + ], + "type": "inline_equation", + "content": "o_{t}" + }, + { + "bbox": [ + 127, + 293, + 468, + 380 + ], + "type": "text", + "content": " in the main task " + }, + { + "bbox": [ + 127, + 293, + 468, + 380 + ], + "type": "inline_equation", + "content": "T_{0}" + }, + { + "bbox": [ + 127, + 293, + 468, + 380 + ], + "type": "text", + "content": " under all auxiliary tasks. The task relevance function computes " + }, + { + "bbox": [ + 127, + 293, + 468, + 380 + ], + "type": "inline_equation", + "content": "w_{i}(s_{t})" + }, + { + "bbox": [ + 127, + 293, + 468, + 380 + ], + "type": "text", + "content": " based on the relevance of the current observation and robot state to the auxiliary task " + }, + { + "bbox": [ + 127, + 293, + 468, + 380 + ], + "type": "inline_equation", + "content": "T_{i}" + }, + { + "bbox": [ + 127, + 293, + 468, + 380 + ], + "type": "text", + "content": ". The distillation loss and RL-training loss are then used to update the policy." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 127, + 407, + 469, + 513 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 127, + 407, + 469, + 513 + ], + "spans": [ + { + "bbox": [ + 127, + 407, + 469, + 513 + ], + "type": "text", + "content": "is challenging because an agent needs to combine thousands of low-level actions controlling the arm and base, operate from egocentric visual perception, and dynamically sequence distinct behaviors such as navigating, picking, placing, and opening. AuxDistill addresses these challenges of RL in complex problems, like embodied rearrangement, by learning to leverage knowledge from easier auxiliary tasks related to the desired task we are trying to solve. Unlike a curriculum, which learns gradually harder tasks in stages, AuxDistill learns the desired task concurrently with the auxiliary task and includes a novel distillation mechanism to transfer knowledge from easier to harder tasks." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 128, + 544, + 221, + 555 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 128, + 544, + 221, + 555 + ], + "spans": [ + { + "bbox": [ + 128, + 544, + 221, + 555 + ], + "type": "text", + "content": "3.1 Preliminaries" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 127, + 577, + 469, + 684 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 127, + 577, + 469, + 684 + ], + "spans": [ + { + "bbox": [ + 127, + 577, + 469, + 684 + ], + "type": "text", + "content": "Our problem is formulated as a goal-specified Partially-Observable Markov Decision Process (POMDP) defined as a tuple " + }, + { + "bbox": [ + 127, + 577, + 469, + 684 + ], + "type": "inline_equation", + "content": "\\mathcal{M} = (\\mathcal{S},\\mathcal{O},\\mathcal{A},\\mathcal{P},\\mathcal{R},\\mathcal{G},\\rho,\\gamma)" + }, + { + "bbox": [ + 127, + 577, + 469, + 684 + ], + "type": "text", + "content": " with underlying state space " + }, + { + "bbox": [ + 127, + 577, + 469, + 684 + ], + "type": "inline_equation", + "content": "\\mathcal{S}" + }, + { + "bbox": [ + 127, + 577, + 469, + 684 + ], + "type": "text", + "content": ", observation space " + }, + { + "bbox": [ + 127, + 577, + 469, + 684 + ], + "type": "inline_equation", + "content": "\\mathcal{O}" + }, + { + "bbox": [ + 127, + 577, + 469, + 684 + ], + "type": "text", + "content": ", action space " + }, + { + "bbox": [ + 127, + 577, + 469, + 684 + ], + "type": "inline_equation", + "content": "\\mathcal{A}" + }, + { + "bbox": [ + 127, + 577, + 469, + 684 + ], + "type": "text", + "content": ", transition function " + }, + { + "bbox": [ + 127, + 577, + 469, + 684 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 127, + 577, + 469, + 684 + ], + "type": "text", + "content": ", reward function " + }, + { + "bbox": [ + 127, + 577, + 469, + 684 + ], + "type": "inline_equation", + "content": "\\mathcal{R}" + }, + { + "bbox": [ + 127, + 577, + 469, + 684 + ], + "type": "text", + "content": ", goal space " + }, + { + "bbox": [ + 127, + 577, + 469, + 684 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 127, + 577, + 469, + 684 + ], + "type": "text", + "content": ", initial state distribution " + }, + { + "bbox": [ + 127, + 577, + 469, + 684 + ], + "type": "inline_equation", + "content": "\\rho" + }, + { + "bbox": [ + 127, + 577, + 469, + 684 + ], + "type": "text", + "content": " and discount factor " + }, + { + "bbox": [ + 127, + 577, + 469, + 684 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 127, + 577, + 469, + 684 + ], + "type": "text", + "content": ". For a task like rearrangement, the goal space " + }, + { + "bbox": [ + 127, + 577, + 469, + 684 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 127, + 577, + 469, + 684 + ], + "type": "text", + "content": " is specified using the 3D coordinate of the object's start location and the goal location where the object has to be placed. Our objective is to learn a goal-conditioned policy " + }, + { + "bbox": [ + 127, + 577, + 469, + 684 + ], + "type": "inline_equation", + "content": "\\pi(a|o,g)" + }, + { + "bbox": [ + 127, + 577, + 469, + 684 + ], + "type": "text", + "content": " mapping an observation " + }, + { + "bbox": [ + 127, + 577, + 469, + 684 + ], + "type": "inline_equation", + "content": "o" + }, + { + "bbox": [ + 127, + 577, + 469, + 684 + ], + "type": "text", + "content": " and goal " + }, + { + "bbox": [ + 127, + 577, + 469, + 684 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 127, + 577, + 469, + 684 + ], + "type": "text", + "content": " to an action " + }, + { + "bbox": [ + 127, + 577, + 469, + 684 + ], + "type": "inline_equation", + "content": "a" + }, + { + "bbox": [ + 127, + 577, + 469, + 684 + ], + "type": "text", + "content": " that maximizes the sum of discounted rewards " + }, + { + "bbox": [ + 127, + 577, + 469, + 684 + ], + "type": "inline_equation", + "content": "\\mathbb{E}_{s_0\\sim\\rho_0,g\\sim\\mathcal{G}}\\sum_t\\gamma^t\\mathcal{R}(s_t,g)" + }, + { + "bbox": [ + 127, + 577, + 469, + 684 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 129, + 126, + 136, + 134 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 129, + 126, + 136, + 134 + ], + "spans": [ + { + "bbox": [ + 129, + 126, + 136, + 134 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 161, + 125, + 213, + 133 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 161, + 125, + 213, + 133 + ], + "spans": [ + { + "bbox": [ + 161, + 125, + 213, + 133 + ], + "type": "text", + "content": "Harish et. al" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 127, + 148, + 426, + 161 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 127, + 148, + 426, + 161 + ], + "spans": [ + { + "bbox": [ + 127, + 148, + 426, + 161 + ], + "type": "text", + "content": "3.2 Reinforcement Learning via Auxiliary Task Distillation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 127, + 169, + 468, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 127, + 169, + 468, + 354 + ], + "spans": [ + { + "bbox": [ + 127, + 169, + 468, + 354 + ], + "type": "text", + "content": "The core idea of AuxDistill is to learn a policy in a difficult desired \"main\" task by transferring knowledge from easier auxiliary tasks. This is done without using an explicit curriculum by learning the main and auxiliary tasks in a single training loop. We refer to the main task we wish to learn as " + }, + { + "bbox": [ + 127, + 169, + 468, + 354 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_0" + }, + { + "bbox": [ + 127, + 169, + 468, + 354 + ], + "type": "text", + "content": " with a task identifier " + }, + { + "bbox": [ + 127, + 169, + 468, + 354 + ], + "type": "inline_equation", + "content": "T_0" + }, + { + "bbox": [ + 127, + 169, + 468, + 354 + ], + "type": "text", + "content": ". We define " + }, + { + "bbox": [ + 127, + 169, + 468, + 354 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 127, + 169, + 468, + 354 + ], + "type": "text", + "content": " auxiliary tasks, defined as " + }, + { + "bbox": [ + 127, + 169, + 468, + 354 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{M}_i\\}_{n=1}^N" + }, + { + "bbox": [ + 127, + 169, + 468, + 354 + ], + "type": "text", + "content": " with task identifiers " + }, + { + "bbox": [ + 127, + 169, + 468, + 354 + ], + "type": "inline_equation", + "content": "\\{T_i\\}_{n=1}^N" + }, + { + "bbox": [ + 127, + 169, + 468, + 354 + ], + "type": "text", + "content": ". The auxiliary tasks " + }, + { + "bbox": [ + 127, + 169, + 468, + 354 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{M}_i\\}" + }, + { + "bbox": [ + 127, + 169, + 468, + 354 + ], + "type": "text", + "content": " share the same state, observation, goal, and action space as the target task " + }, + { + "bbox": [ + 127, + 169, + 468, + 354 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_0" + }, + { + "bbox": [ + 127, + 169, + 468, + 354 + ], + "type": "text", + "content": ". Each " + }, + { + "bbox": [ + 127, + 169, + 468, + 354 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_i" + }, + { + "bbox": [ + 127, + 169, + 468, + 354 + ], + "type": "text", + "content": " has a separately defined starting state distribution and reward function. We assume these auxiliary tasks are related to the full task yet are easier to solve than the full task. Specifically, the auxiliary tasks can be easier instantiations or sub-parts of the overall task. For example, in rearrangement, we define the auxiliary tasks in terms of interactions the agent needs to complete the entire rearrangement episode, such as picking, placing, and opening. Prior works use similar task definitions to train skills for hierarchical policies in rearrangement [17], but these works suffer from a two-stage pipeline of first needing to separately train each skill and then decide how to combine them. AuxDistill directly learns the full task from scratch while also performing better." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 127, + 356, + 468, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 127, + 356, + 468, + 437 + ], + "spans": [ + { + "bbox": [ + 127, + 356, + 468, + 437 + ], + "type": "text", + "content": "AuxDistill learns a single policy with RL that concurrently learns to perform the main and auxiliary tasks. We illustrate the implementation of AuxDistill in Figure 1. This policy, parameterized by " + }, + { + "bbox": [ + 127, + 356, + 468, + 437 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 127, + 356, + 468, + 437 + ], + "type": "text", + "content": ", is expressed as " + }, + { + "bbox": [ + 127, + 356, + 468, + 437 + ], + "type": "inline_equation", + "content": "\\pi_{\\theta}(a_t \\mid o_t, g, T)" + }, + { + "bbox": [ + 127, + 356, + 468, + 437 + ], + "type": "text", + "content": " for observation " + }, + { + "bbox": [ + 127, + 356, + 468, + 437 + ], + "type": "inline_equation", + "content": "o_t" + }, + { + "bbox": [ + 127, + 356, + 468, + 437 + ], + "type": "text", + "content": ", episode goal " + }, + { + "bbox": [ + 127, + 356, + 468, + 437 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 127, + 356, + 468, + 437 + ], + "type": "text", + "content": ", and per-" + }, + { + "bbox": [ + 127, + 356, + 468, + 437 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 127, + 356, + 468, + 437 + ], + "type": "text", + "content": " task identifier " + }, + { + "bbox": [ + 127, + 356, + 468, + 437 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 127, + 356, + 468, + 437 + ], + "type": "text", + "content": " which is encoded as a one-hot embedding in a vector which has the same size as the maximum number of auxiliary tasks. Note that since all tasks share the same observation and action space, " + }, + { + "bbox": [ + 127, + 356, + 468, + 437 + ], + "type": "inline_equation", + "content": "\\pi_{\\theta}" + }, + { + "bbox": [ + 127, + 356, + 468, + 437 + ], + "type": "text", + "content": " can act and observe in all tasks based on the input task identifier " + }, + { + "bbox": [ + 127, + 356, + 468, + 437 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 127, + 356, + 468, + 437 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 127, + 438, + 468, + 576 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 127, + 438, + 468, + 576 + ], + "spans": [ + { + "bbox": [ + 127, + 438, + 468, + 576 + ], + "type": "text", + "content": "AuxDistill updates the policy based on an average RL loss from the main task and all auxiliary tasks. Let " + }, + { + "bbox": [ + 127, + 438, + 468, + 576 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathcal{M}_i}^{\\mathrm{RL}}(\\theta)" + }, + { + "bbox": [ + 127, + 438, + 468, + 576 + ], + "type": "text", + "content": " denote the RL loss for " + }, + { + "bbox": [ + 127, + 438, + 468, + 576 + ], + "type": "inline_equation", + "content": "\\pi_{\\theta}" + }, + { + "bbox": [ + 127, + 438, + 468, + 576 + ], + "type": "text", + "content": " in MDP " + }, + { + "bbox": [ + 127, + 438, + 468, + 576 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_i" + }, + { + "bbox": [ + 127, + 438, + 468, + 576 + ], + "type": "text", + "content": ". These auxiliary tasks are designed to capture a subset of the main task which are easier to accomplish than the main task. To compute these losses, we assume we can collect experience in the auxiliary tasks. For example, to collect experience in a \"pick object\" auxiliary task in object rearrangement, the robot is spawned close to the object. We then update based on the average of the task and auxiliary task losses: " + }, + { + "bbox": [ + 127, + 438, + 468, + 576 + ], + "type": "inline_equation", + "content": "\\frac{1}{N}\\sum_{i=0}^{N}\\mathcal{L}_{\\mathcal{M}_i}^{\\mathrm{RL}}(\\theta)" + }, + { + "bbox": [ + 127, + 438, + 468, + 576 + ], + "type": "text", + "content": ". Intuitively, " + }, + { + "bbox": [ + 127, + 438, + 468, + 576 + ], + "type": "inline_equation", + "content": "\\pi_{\\theta}" + }, + { + "bbox": [ + 127, + 438, + 468, + 576 + ], + "type": "text", + "content": " will first learn to complete the easier auxiliary tasks, and this auxiliary task competency can aid in solving the main task. This shares a similar insight as curriculum learning, except all tasks are learned concurrently, and the curriculum stages are naturally induced by the policy naturally learning on easier tasks first." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 127, + 577, + 468, + 682 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 127, + 577, + 468, + 682 + ], + "spans": [ + { + "bbox": [ + 127, + 577, + 468, + 682 + ], + "type": "text", + "content": "In addition to optimizing the average RL loss between the main and auxiliary tasks, AuxDistill also optimizes a distillation loss that encourages the policy to transfer relevant knowledge from the auxiliary tasks to the main task. To achieve this, for a particular observation " + }, + { + "bbox": [ + 127, + 577, + 468, + 682 + ], + "type": "inline_equation", + "content": "o_t" + }, + { + "bbox": [ + 127, + 577, + 468, + 682 + ], + "type": "text", + "content": ", we consider the policy distribution " + }, + { + "bbox": [ + 127, + 577, + 468, + 682 + ], + "type": "inline_equation", + "content": "\\pi_\\theta(\\cdot | o_t, g, T)" + }, + { + "bbox": [ + 127, + 577, + 468, + 682 + ], + "type": "text", + "content": " under different task identifiers " + }, + { + "bbox": [ + 127, + 577, + 468, + 682 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 127, + 577, + 468, + 682 + ], + "type": "text", + "content": ". For compactness, notate " + }, + { + "bbox": [ + 127, + 577, + 468, + 682 + ], + "type": "inline_equation", + "content": "\\pi_\\theta(\\cdot | o_t, g, T_i)" + }, + { + "bbox": [ + 127, + 577, + 468, + 682 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 127, + 577, + 468, + 682 + ], + "type": "inline_equation", + "content": "T_i" + }, + { + "bbox": [ + 127, + 577, + 468, + 682 + ], + "type": "text", + "content": " indicates the task identifier for " + }, + { + "bbox": [ + 127, + 577, + 468, + 682 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_i" + }, + { + "bbox": [ + 127, + 577, + 468, + 682 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 127, + 577, + 468, + 682 + ], + "type": "inline_equation", + "content": "\\pi_\\theta^{T_i}(o_t)" + }, + { + "bbox": [ + 127, + 577, + 468, + 682 + ], + "type": "text", + "content": ". We want " + }, + { + "bbox": [ + 127, + 577, + 468, + 682 + ], + "type": "inline_equation", + "content": "\\pi_\\theta^{T_0}" + }, + { + "bbox": [ + 127, + 577, + 468, + 682 + ], + "type": "text", + "content": " (the policy in the main task) to match the behaviors from the policy in the relevant auxiliary tasks " + }, + { + "bbox": [ + 127, + 577, + 468, + 682 + ], + "type": "inline_equation", + "content": "\\pi_\\theta^{T_i}" + }, + { + "bbox": [ + 127, + 577, + 468, + 682 + ], + "type": "text", + "content": ". We measure this relevance of an auxiliary task " + }, + { + "bbox": [ + 127, + 577, + 468, + 682 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_i" + }, + { + "bbox": [ + 127, + 577, + 468, + 682 + ], + "type": "text", + "content": " to the main task at time step " + }, + { + "bbox": [ + 127, + 577, + 468, + 682 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 127, + 577, + 468, + 682 + ], + "type": "text", + "content": " via an auxiliary task relevance function " + }, + { + "bbox": [ + 127, + 577, + 468, + 682 + ], + "type": "inline_equation", + "content": "w_i(s_t)" + }, + { + "bbox": [ + 127, + 577, + 468, + 682 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 392, + 125, + 435, + 134 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 392, + 125, + 435, + 134 + ], + "spans": [ + { + "bbox": [ + 392, + 125, + 435, + 134 + ], + "type": "text", + "content": "AuxDistill" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 461, + 125, + 467, + 133 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 461, + 125, + 467, + 133 + ], + "spans": [ + { + "bbox": [ + 461, + 125, + 467, + 133 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 127, + 148, + 468, + 276 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 127, + 148, + 468, + 276 + ], + "spans": [ + { + "bbox": [ + 127, + 148, + 468, + 276 + ], + "type": "text", + "content": "This function denotes how much the knowledge from " + }, + { + "bbox": [ + 127, + 148, + 468, + 276 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_i" + }, + { + "bbox": [ + 127, + 148, + 468, + 276 + ], + "type": "text", + "content": " should apply to " + }, + { + "bbox": [ + 127, + 148, + 468, + 276 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_0" + }, + { + "bbox": [ + 127, + 148, + 468, + 276 + ], + "type": "text", + "content": " in the underlying simulator state " + }, + { + "bbox": [ + 127, + 148, + 468, + 276 + ], + "type": "inline_equation", + "content": "s_t" + }, + { + "bbox": [ + 127, + 148, + 468, + 276 + ], + "type": "text", + "content": ". This relevance function can be grounded in the task plan generated by an oracle planner. For example, consider the state of the robot before picking up the object. In this state, the pick auxiliary task would be relevant to the main task, and the place auxiliary task would not. If an episode requires opening a fridge before accessing the object, the relevant task before the fridge is opened would be open-fridge. Note that computing the distillation weight can utilize oracle knowledge of the simulator state (for instance, whether the object is inside a fridge or a cabinet) since this information is only provided as a training signal and not used during inference. This information is utilized by all methods we compare against, including our strongest baseline [22]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 127, + 277, + 468, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 127, + 277, + 468, + 335 + ], + "spans": [ + { + "bbox": [ + 127, + 277, + 468, + 335 + ], + "type": "text", + "content": "We can then distill experience from " + }, + { + "bbox": [ + 127, + 277, + 468, + 335 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_i" + }, + { + "bbox": [ + 127, + 277, + 468, + 335 + ], + "type": "text", + "content": " by supervising the policy to match the action distribution of the target task " + }, + { + "bbox": [ + 127, + 277, + 468, + 335 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_0" + }, + { + "bbox": [ + 127, + 277, + 468, + 335 + ], + "type": "text", + "content": ". The distillation loss is computed as the KL-divergence of the action distribution under " + }, + { + "bbox": [ + 127, + 277, + 468, + 335 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_i" + }, + { + "bbox": [ + 127, + 277, + 468, + 335 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 127, + 277, + 468, + 335 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_0" + }, + { + "bbox": [ + 127, + 277, + 468, + 335 + ], + "type": "text", + "content": " weighted by the relevance of auxiliary task " + }, + { + "bbox": [ + 127, + 277, + 468, + 335 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_i" + }, + { + "bbox": [ + 127, + 277, + 468, + 335 + ], + "type": "text", + "content": " given by " + }, + { + "bbox": [ + 127, + 277, + 468, + 335 + ], + "type": "inline_equation", + "content": "w_i(s_t)" + }, + { + "bbox": [ + 127, + 277, + 468, + 335 + ], + "type": "text", + "content": ". The overall loss function of AuxDistill, including the distillation loss, is then:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 189, + 344, + 468, + 377 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 189, + 344, + 468, + 377 + ], + "spans": [ + { + "bbox": [ + 189, + 344, + 468, + 377 + ], + "type": "interline_equation", + "content": "\\frac {1}{N} \\sum_ {i = 0} ^ {N} \\mathcal {L} _ {\\mathcal {M} _ {i}} ^ {\\mathrm {R L}} (\\theta) + \\lambda \\sum_ {i = 1} ^ {N} w _ {i} \\left(s _ {t}\\right) D _ {\\mathrm {K L}} \\left(\\pi_ {\\theta} ^ {T _ {0}} \\left(o _ {t}\\right) \\parallel \\pi_ {\\theta} ^ {T _ {i}} \\left(o _ {t}\\right)\\right) \\tag {1}", + "image_path": "dcb8944c8f739ce1ff61ac539c6d515d8ff8ba9f69f419aea6f050963dc478d0.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 127, + 385, + 468, + 478 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 127, + 385, + 468, + 478 + ], + "spans": [ + { + "bbox": [ + 127, + 385, + 468, + 478 + ], + "type": "text", + "content": "Here " + }, + { + "bbox": [ + 127, + 385, + 468, + 478 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 127, + 385, + 468, + 478 + ], + "type": "text", + "content": " represents the distillation weight relative to other PPO losses. Note that our formulation relies on the auxiliary tasks being easier to solve and relevant to the main task. Distillation using the task relevance function offers an additional supervisory signal that the policy can optimize along with its own reward. This doesn't impose the large overhead of hierarchical RL methods where each auxiliary task must be pre-trained in isolation with its own reward function and goal specification. Auxiliary tasks are flexibly incorporated to address parts of the task that are challenging for the agent to learn." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 127, + 497, + 271, + 509 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 127, + 497, + 271, + 509 + ], + "spans": [ + { + "bbox": [ + 127, + 497, + 271, + 509 + ], + "type": "text", + "content": "3.3 Implementation Details" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 127, + 518, + 468, + 577 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 127, + 518, + 468, + 577 + ], + "spans": [ + { + "bbox": [ + 127, + 518, + 468, + 577 + ], + "type": "text", + "content": "We use PPO [39] to compute the RL loss " + }, + { + "bbox": [ + 127, + 518, + 468, + 577 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathcal{M}_i}^{\\mathrm{RL}}" + }, + { + "bbox": [ + 127, + 518, + 468, + 577 + ], + "type": "text", + "content": " in Equation (1). We create " + }, + { + "bbox": [ + 127, + 518, + 468, + 577 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 127, + 518, + 468, + 577 + ], + "type": "text", + "content": " environment instances for each of the main and " + }, + { + "bbox": [ + 127, + 518, + 468, + 577 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 127, + 518, + 468, + 577 + ], + "type": "text", + "content": " auxiliary tasks and vectorize them for parallel action execution. We rollout the policy in all these " + }, + { + "bbox": [ + 127, + 518, + 468, + 577 + ], + "type": "inline_equation", + "content": "M(N + 1)" + }, + { + "bbox": [ + 127, + 518, + 468, + 577 + ], + "type": "text", + "content": " environments and collect a batch of data. We then compute Equation (1), using the PPO loss, update the policy, and then repeat this process." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 127, + 577, + 468, + 682 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 127, + 577, + 468, + 682 + ], + "spans": [ + { + "bbox": [ + 127, + 577, + 468, + 682 + ], + "type": "text", + "content": "The policy is represented as a 2-layer LSTM network with 512 hidden units per layer, similar to the architecture employed in [43]. In our experiments, " + }, + { + "bbox": [ + 127, + 577, + 468, + 682 + ], + "type": "inline_equation", + "content": "o_t" + }, + { + "bbox": [ + 127, + 577, + 468, + 682 + ], + "type": "text", + "content": " is a depth or RGB egocentric image, which we encode with a ResNet50 [20] network. The goal and the proprioceptive state of the agent are concatenated with the visual embeddings along with the index of the POMDP " + }, + { + "bbox": [ + 127, + 577, + 468, + 682 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_i" + }, + { + "bbox": [ + 127, + 577, + 468, + 682 + ], + "type": "text", + "content": ", for which we use a one-hot embedding. This vector is then passed as an input to the LSTM. A linear projection on the output of the LSTM then produces an action to execute on the next step in the environment and a vector of value-predictions corresponding to each of the tasks " + }, + { + "bbox": [ + 127, + 577, + 468, + 682 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_i" + }, + { + "bbox": [ + 127, + 577, + 468, + 682 + ], + "type": "text", + "content": " used during training." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 129, + 125, + 136, + 133 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 129, + 125, + 136, + 133 + ], + "spans": [ + { + "bbox": [ + 129, + 125, + 136, + 133 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 161, + 125, + 213, + 134 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 161, + 125, + 213, + 134 + ], + "spans": [ + { + "bbox": [ + 161, + 125, + 213, + 134 + ], + "type": "text", + "content": "Harish et. al" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 139, + 146, + 452, + 258 + ], + "blocks": [ + { + "bbox": [ + 139, + 146, + 452, + 258 + ], + "lines": [ + { + "bbox": [ + 139, + 146, + 452, + 258 + ], + "spans": [ + { + "bbox": [ + 139, + 146, + 452, + 258 + ], + "type": "table", + "html": "
MethodTrain (Seen)Eval (Unseen)
All EpisodesEasyHardAll EpisodesEasyHard
M3 (Oracle) [17]27 ± 057 ± 212 ± 128 ± 058 ± 213 ± 2
M3 [17]25 ± 156 ± 19 ± 213 ± 153 ± 40 ± 0
Galactic [8]-37 ± 0--26 ± 0-
Monolithic RL0 ± 00 ± 00 ± 00 ± 00 ± 00 ± 0
Skill Transformer [22]25 ± 144 ± 116 ± 123 ± 137 ± 116 ± 1
AuxDistill (No Distillation)0 ± 00 ± 00 ± 00 ± 00 ± 00 ± 0
RL Curriculum0 ± 01 ± 00 ± 00 ± 01 ± 10 ± 0
AuxDistill49 ± 274 ± 236 ± 252 ± 275 ± 241 ± 2
", + "image_path": "7ebf77ca445d01bc4dcc365636589526d8c7b127b552421a7167a553470b4394.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 127, + 259, + 468, + 333 + ], + "lines": [ + { + "bbox": [ + 127, + 259, + 468, + 333 + ], + "spans": [ + { + "bbox": [ + 127, + 259, + 468, + 333 + ], + "type": "text", + "content": "Table 1: Success rates on the rearrangement task for our method, AuxDistill (highlighted in blue), and baselines. Displayed are the average and standard deviations for 3 seeds for M3, Monolithic RL, Skill Transformer, and M3 Oracle (numbers from [22]), and 3 seeds for the remaining methods with the highest numbers per setting bolded. Numbers in the easy and hard columns are averages over 100 episodes and 200 episodes, respectively. The All Episodes column is an average across both the Easy and Hard episodes. AuxDistill outperforms all baselines in all settings." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 2 + }, + { + "bbox": [ + 127, + 367, + 468, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 127, + 367, + 468, + 426 + ], + "spans": [ + { + "bbox": [ + 127, + 367, + 468, + 426 + ], + "type": "text", + "content": "Our policy is learned via a multi-task RL formulation, with the main and auxiliary tasks being concurrently learned. Each of the tasks may have different reward magnitudes and relative performance during training. To address this, we use PopArt [21], with " + }, + { + "bbox": [ + 127, + 367, + 468, + 426 + ], + "type": "inline_equation", + "content": "\\beta = 3e^{-4}" + }, + { + "bbox": [ + 127, + 367, + 468, + 426 + ], + "type": "text", + "content": " to enable learning across different return scales. Additional AuxDistill details are in Appendix A of the supplementary." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 127, + 446, + 223, + 460 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 127, + 446, + 223, + 460 + ], + "spans": [ + { + "bbox": [ + 127, + 446, + 223, + 460 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 127, + 473, + 268, + 486 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 127, + 473, + 268, + 486 + ], + "spans": [ + { + "bbox": [ + 127, + 473, + 268, + 486 + ], + "type": "text", + "content": "4.1 Object-Rearrangement" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 127, + 495, + 468, + 657 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 127, + 495, + 468, + 657 + ], + "spans": [ + { + "bbox": [ + 127, + 495, + 468, + 657 + ], + "type": "text", + "content": "In this section, we compare AuxDistill with baselines on the Habitat 2.0 Object Rearrangement task [43]. For comparison with baselines, we run our experiments using the setup from [22]. In this task, a Fetch robot [37] must move an object from a specified start position to a desired goal position in an indoor home environment using only onboard sensing. The agent has no privileged information like existing maps of the environment, 3D object models, or exact object positions. The Fetch robot senses the world through a " + }, + { + "bbox": [ + 127, + 495, + 468, + 657 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 127, + 495, + 468, + 657 + ], + "type": "text", + "content": " depth camera mounted on the robot's head, robot joint positions, gripper state, and base egomotion, giving the relative position of the robot from the start of the episode. The task is specified by a starting 3D object coordinate for the object to move and a 3D goal coordinate to move the object to. Both coordinates are specified relative to the robot's position at the start of the episode. Only the starting object coordinate is specified, and this information is not updated based on the current object position (e.g. if it is moved)." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 127, + 659, + 468, + 683 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 127, + 659, + 468, + 683 + ], + "spans": [ + { + "bbox": [ + 127, + 659, + 468, + 683 + ], + "type": "text", + "content": "The robot interacts with the world via a 7DoF arm, a suction gripper attached to the end of the arm, and a mobile base. The episode is successful if the target" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 392, + 125, + 435, + 134 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 392, + 125, + 435, + 134 + ], + "spans": [ + { + "bbox": [ + 392, + 125, + 435, + 134 + ], + "type": "text", + "content": "AuxDistill" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 461, + 125, + 467, + 133 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 461, + 125, + 467, + 133 + ], + "spans": [ + { + "bbox": [ + 461, + 125, + 467, + 133 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 127, + 148, + 466, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 127, + 148, + 466, + 171 + ], + "spans": [ + { + "bbox": [ + 127, + 148, + 466, + 171 + ], + "type": "text", + "content": "object is within " + }, + { + "bbox": [ + 127, + 148, + 466, + 171 + ], + "type": "inline_equation", + "content": "15\\mathrm{cm}" + }, + { + "bbox": [ + 127, + 148, + 466, + 171 + ], + "type": "text", + "content": " of the goal position. The robot has a budget of 1,500 steps to complete the task." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 127, + 172, + 468, + 276 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 127, + 172, + 468, + 276 + ], + "spans": [ + { + "bbox": [ + 127, + 172, + 468, + 276 + ], + "type": "text", + "content": "We report performance on the easy and hard evaluation episodes from [22]. In easy episodes, both the target object and goal are on an open receptacle. This means the robot does not have to first open a receptacle before accessing the target object or goal. Instead, the agent can always execute the same sequence of navigate to object, pick up object, navigate to goal, and place object at goal. In hard episodes, the object may start in a closed receptacle. The agent, therefore, needs to use its visual input to perceive if the target object is in a closed receptacle. If so, the robot must then open the receptacle before picking the object. The object may start in either the fridge or a cabinet." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 127, + 291, + 468, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 127, + 291, + 468, + 384 + ], + "spans": [ + { + "bbox": [ + 127, + 291, + 468, + 384 + ], + "type": "text", + "content": "Training Setup: We train AuxDistill using 11,791 episodes with a mix of easy and hard episodes across 63 scenes using the same rearrangement training dataset as in [44]. These episodes were obtained by sampling an equal number of episodes in the easy and hard categories, with easy episodes being equally sampled across episodes with open-cabinet, open-fridge and non-articulated episodes. For hard episodes, sampling is done uniformly between closed fridge and closed cabinet episodes. We use auxiliary tasks covering abilities included as a part of the standard rearrangement benchmark used in [44]. The auxiliary tasks are:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 127, + 384, + 467, + 530 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 127, + 384, + 467, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 127, + 384, + 467, + 407 + ], + "spans": [ + { + "bbox": [ + 127, + 384, + 467, + 407 + ], + "type": "text", + "content": "- Pick: The agent spawns randomly within the house and must navigate to and pick up the object." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 127, + 409, + 467, + 444 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 127, + 409, + 467, + 444 + ], + "spans": [ + { + "bbox": [ + 127, + 409, + 467, + 444 + ], + "type": "text", + "content": "- Place the agent is spawned randomly in the house with the object in its gripper. The robot must navigate to and place the object within " + }, + { + "bbox": [ + 127, + 409, + 467, + 444 + ], + "type": "inline_equation", + "content": "15\\mathrm{cm}" + }, + { + "bbox": [ + 127, + 409, + 467, + 444 + ], + "type": "text", + "content": " of the target 3D location on the receptacle." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 127, + 446, + 467, + 470 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 127, + 446, + 467, + 470 + ], + "spans": [ + { + "bbox": [ + 127, + 446, + 467, + 470 + ], + "type": "text", + "content": "- Open Fridge: The robot spawns randomly within the house and must navigate to the fridge and use its base and arm to open the fridge door." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 127, + 471, + 467, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 127, + 471, + 467, + 495 + ], + "spans": [ + { + "bbox": [ + 127, + 471, + 467, + 495 + ], + "type": "text", + "content": "- Open Cabinet: The robot spawns randomly in the house and must navigate to the kitchen area and open the cabinet using its base and arm." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 127, + 496, + 467, + 530 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 127, + 496, + 467, + 530 + ], + "spans": [ + { + "bbox": [ + 127, + 496, + 467, + 530 + ], + "type": "text", + "content": "- Pick from Fridge: This task is similar to Pick except the agent has to pick up the object from inside an open fridge, which requires careful manipulation to minimize collisions." + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 127, + 531, + 468, + 576 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 127, + 531, + 468, + 576 + ], + "spans": [ + { + "bbox": [ + 127, + 531, + 468, + 576 + ], + "type": "text", + "content": "The auxiliary relevance function is grounded in the oracle task plan [14] of the episode. This is formulated as an indicator function with " + }, + { + "bbox": [ + 127, + 531, + 468, + 576 + ], + "type": "inline_equation", + "content": "w_{i}(s_{t}) = 1" + }, + { + "bbox": [ + 127, + 531, + 468, + 576 + ], + "type": "text", + "content": " if the state " + }, + { + "bbox": [ + 127, + 531, + 468, + 576 + ], + "type": "inline_equation", + "content": "s_t" + }, + { + "bbox": [ + 127, + 531, + 468, + 576 + ], + "type": "text", + "content": " is relevant to " + }, + { + "bbox": [ + 127, + 531, + 468, + 576 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_i" + }, + { + "bbox": [ + 127, + 531, + 468, + 576 + ], + "type": "text", + "content": " based on the oracle task plan of the episode and " + }, + { + "bbox": [ + 127, + 531, + 468, + 576 + ], + "type": "inline_equation", + "content": "w_{i}(s_{t}) = 0" + }, + { + "bbox": [ + 127, + 531, + 468, + 576 + ], + "type": "text", + "content": " otherwise." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 127, + 577, + 468, + 682 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 127, + 577, + 468, + 682 + ], + "spans": [ + { + "bbox": [ + 127, + 577, + 468, + 682 + ], + "type": "text", + "content": "For example, if at step " + }, + { + "bbox": [ + 127, + 577, + 468, + 682 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 127, + 577, + 468, + 682 + ], + "type": "text", + "content": " the agent is not holding an object and the object is in an open receptacle, the agent must first pick up the object so " + }, + { + "bbox": [ + 127, + 577, + 468, + 682 + ], + "type": "inline_equation", + "content": "w_{i_{\\text{pick}}}(s_t) = 1" + }, + { + "bbox": [ + 127, + 577, + 468, + 682 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 127, + 577, + 468, + 682 + ], + "type": "inline_equation", + "content": "w_i(s_t) = 0" + }, + { + "bbox": [ + 127, + 577, + 468, + 682 + ], + "type": "text", + "content": " for all other skills. For a complete description of the auxiliary tasks, including their reward functions, along with the precise auxiliary relevance function definition in object rearrangement, see Appendix B. We train methods for 475M steps of environment interactions. We found this to be a sufficient number of environment interactions to ensure sufficient progress on auxiliary tasks for distillation to aid rearrangement performance on both the easy and hard episodes. For AuxDistill, we count the environment interactions in the auxiliary" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 129, + 125, + 136, + 133 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 129, + 125, + 136, + 133 + ], + "spans": [ + { + "bbox": [ + 129, + 125, + 136, + 133 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 161, + 125, + 213, + 134 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 161, + 125, + 213, + 134 + ], + "spans": [ + { + "bbox": [ + 161, + 125, + 213, + 134 + ], + "type": "text", + "content": "Harish et. al" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 137, + 146, + 456, + 227 + ], + "blocks": [ + { + "bbox": [ + 137, + 146, + 456, + 227 + ], + "lines": [ + { + "bbox": [ + 137, + 146, + 456, + 227 + ], + "spans": [ + { + "bbox": [ + 137, + 146, + 456, + 227 + ], + "type": "table", + "html": "
MethodTrain (Seen)Eval (Unseen)
All EpisodesEasyHardAll EpisodesEasyHard
All Skills283619303921
No Pick From Fridge2032712232
No Open Fridge414933485739
No Pick000000
", + "image_path": "46dde73ece492b83bd763ae487f041d29624ef921fd4fc83708f857e39b9bc2b.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 127, + 227, + 467, + 269 + ], + "lines": [ + { + "bbox": [ + 127, + 227, + 467, + 269 + ], + "spans": [ + { + "bbox": [ + 127, + 227, + 467, + 269 + ], + "type": "text", + "content": "Table 2: Robustness to auxiliary task selection. We compare the train and generalization performance of AuxDistill with 4 different auxiliary task selections. Results are averages across 100 episodes on a single seed of training. The train and evaluation episodes are in a simplified setting compared to Table 1 as described in Section 4.3" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 127, + 311, + 466, + 334 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 127, + 311, + 466, + 334 + ], + "spans": [ + { + "bbox": [ + 127, + 311, + 466, + 334 + ], + "type": "text", + "content": "tasks towards this 475M step budget. We train with a learning rate of " + }, + { + "bbox": [ + 127, + 311, + 466, + 334 + ], + "type": "inline_equation", + "content": "3e^{-4}" + }, + { + "bbox": [ + 127, + 311, + 466, + 334 + ], + "type": "text", + "content": " and linearly decay the learning rate to 0 over the course of learning." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 127, + 355, + 467, + 401 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 127, + 355, + 467, + 401 + ], + "spans": [ + { + "bbox": [ + 127, + 355, + 467, + 401 + ], + "type": "text", + "content": "Baselines We compare AuxDistill to both hierarchical methods governed by a task plan as well as monolithic baselines, which directly learn a pixels-to-actions policy using sensor observations. We compare to relevant baselines from huang2023skill and two more baselines that use the auxiliary tasks." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 127, + 403, + 467, + 668 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 127, + 403, + 466, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 127, + 403, + 466, + 460 + ], + "spans": [ + { + "bbox": [ + 127, + 403, + 466, + 460 + ], + "type": "text", + "content": "- Monolithic RL: A monolithic neural network is trained to map sensor observations directly to actions trained with end-to-end RL. The policy showed no signs of learning the main task, so we stopped training early after " + }, + { + "bbox": [ + 127, + 403, + 466, + 460 + ], + "type": "inline_equation", + "content": "100M" + }, + { + "bbox": [ + 127, + 403, + 466, + 460 + ], + "type": "text", + "content": " steps of training. This outcome is consistent with the monolithic RL baseline in prior works [8, 43]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 127, + 462, + 466, + 508 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 127, + 462, + 466, + 508 + ], + "spans": [ + { + "bbox": [ + 127, + 462, + 466, + 508 + ], + "type": "text", + "content": "- Galactic: [8] An end-to-end RL framework similar to monolithic RL to map sensor observations directly to actions with the policy being trained with a simplified kinematic simulation on over " + }, + { + "bbox": [ + 127, + 462, + 466, + 508 + ], + "type": "inline_equation", + "content": ">1e^{9}" + }, + { + "bbox": [ + 127, + 462, + 466, + 508 + ], + "type": "text", + "content": " simulation steps and transferred to dynamic simulation utilized in our setting." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 127, + 511, + 466, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 127, + 511, + 466, + 533 + ], + "spans": [ + { + "bbox": [ + 127, + 511, + 466, + 533 + ], + "type": "text", + "content": "- M3: [17] Each of the skill policies navigate, pick, place, and open are separately trained. These skills are then sequenced using a task planner." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 127, + 536, + 466, + 558 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 127, + 536, + 466, + 558 + ], + "spans": [ + { + "bbox": [ + 127, + 536, + 466, + 558 + ], + "type": "text", + "content": "- M3 (Oracle): [17] This uses the same skill training as M3, but uses an oracle task planner." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 127, + 561, + 467, + 595 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 127, + 561, + 467, + 595 + ], + "spans": [ + { + "bbox": [ + 127, + 561, + 467, + 595 + ], + "type": "text", + "content": "- Skill Transformer [22]: This method uses pre-trained skills to collect successful rearrangement demonstrations and then train on these demonstrations with imitation learning." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 127, + 598, + 466, + 643 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 127, + 598, + 466, + 643 + ], + "spans": [ + { + "bbox": [ + 127, + 598, + 466, + 643 + ], + "type": "text", + "content": "- RL Curriculum This method is trained in two stages: i) First, we train the policy on the auxiliary tasks for " + }, + { + "bbox": [ + 127, + 598, + 466, + 643 + ], + "type": "inline_equation", + "content": "200M" + }, + { + "bbox": [ + 127, + 598, + 466, + 643 + ], + "type": "text", + "content": " steps to ensure the relevant auxiliary tasks have good training performance ii) This is followed by a training on the entire rearrangement task for another " + }, + { + "bbox": [ + 127, + 598, + 466, + 643 + ], + "type": "inline_equation", + "content": "300M" + }, + { + "bbox": [ + 127, + 598, + 466, + 643 + ], + "type": "text", + "content": " steps." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 127, + 646, + 466, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 127, + 646, + 466, + 668 + ], + "spans": [ + { + "bbox": [ + 127, + 646, + 466, + 668 + ], + "type": "text", + "content": "- AuxDistill: No-Distillation Train AuxDistill under a similar setting as described in Section 4.1 except with the distillation strength " + }, + { + "bbox": [ + 127, + 646, + 466, + 668 + ], + "type": "inline_equation", + "content": "\\lambda = 0.0" + }, + { + "bbox": [ + 127, + 646, + 466, + 668 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 127, + 671, + 352, + 682 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 127, + 671, + 352, + 682 + ], + "spans": [ + { + "bbox": [ + 127, + 671, + 352, + 682 + ], + "type": "text", + "content": "For more details on the baselines, see Appendix B.2." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 392, + 124, + 435, + 134 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 392, + 124, + 435, + 134 + ], + "spans": [ + { + "bbox": [ + 392, + 124, + 435, + 134 + ], + "type": "text", + "content": "AuxDistill" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 461, + 125, + 467, + 133 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 461, + 125, + 467, + 133 + ], + "spans": [ + { + "bbox": [ + 461, + 125, + 467, + 133 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 179, + 161, + 408, + 282 + ], + "blocks": [ + { + "bbox": [ + 179, + 161, + 408, + 282 + ], + "lines": [ + { + "bbox": [ + 179, + 161, + 408, + 282 + ], + "spans": [ + { + "bbox": [ + 179, + 161, + 408, + 282 + ], + "type": "image", + "image_path": "9eb27ef133401ffcf64610759d9858dfbd0124833233b2fe93d8aff0d916c27c.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 127, + 290, + 468, + 365 + ], + "lines": [ + { + "bbox": [ + 127, + 290, + 468, + 365 + ], + "spans": [ + { + "bbox": [ + 127, + 290, + 468, + 365 + ], + "type": "text", + "content": "Fig. 2: Comparison of skill-robustness to different choices of auxiliary task on the hard distribution. Including both the Pick and Pick from Fridge is crucial for successful rearrangement on this distribution. Not utilizing Open-Fridge leads to a boost in rearrangement success. This improvement arises because the open-fridge skill is the easiest of all auxiliary tasks and utilizing it reduces the number of samples for the main task (from the 100M step budget). We discuss the auxiliary task learning curves in Appendix B.3" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 127, + 390, + 468, + 528 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 127, + 390, + 468, + 528 + ], + "spans": [ + { + "bbox": [ + 127, + 390, + 468, + 528 + ], + "type": "text", + "content": "Rearrangement Performance: In Table 1, we compare AuxDistill to baselines in the rearrangement task. On all evaluation settings, AuxDistill outperforms all baselines. AuxDistill shows an absolute improvement of " + }, + { + "bbox": [ + 127, + 390, + 468, + 528 + ], + "type": "inline_equation", + "content": "22\\%" + }, + { + "bbox": [ + 127, + 390, + 468, + 528 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 127, + 390, + 468, + 528 + ], + "type": "inline_equation", + "content": "25\\%" + }, + { + "bbox": [ + 127, + 390, + 468, + 528 + ], + "type": "text", + "content": " on the easy and hard episodes respectively over the best performing baseline. Most baselines struggle to achieve any performance on the unseen episodes. Monolithic RL [43] achieves no success in any of the evaluation settings, demonstrating the impracticality of learning directly from the dense reward in the full rearrangement task. Our method also outperforms Galactic [8] despite learning on less than half the number of training samples (475M for AuxDistill vs. " + }, + { + "bbox": [ + 127, + 390, + 468, + 528 + ], + "type": "inline_equation", + "content": ">1e^{9}" + }, + { + "bbox": [ + 127, + 390, + 468, + 528 + ], + "type": "text", + "content": " for Galactic) on the easy split with " + }, + { + "bbox": [ + 127, + 390, + 468, + 528 + ], + "type": "inline_equation", + "content": "74\\%" + }, + { + "bbox": [ + 127, + 390, + 468, + 528 + ], + "type": "text", + "content": " success over " + }, + { + "bbox": [ + 127, + 390, + 468, + 528 + ], + "type": "inline_equation", + "content": "37\\%" + }, + { + "bbox": [ + 127, + 390, + 468, + 528 + ], + "type": "text", + "content": " on easy episodes during training and " + }, + { + "bbox": [ + 127, + 390, + 468, + 528 + ], + "type": "inline_equation", + "content": "75\\%" + }, + { + "bbox": [ + 127, + 390, + 468, + 528 + ], + "type": "text", + "content": " vs. " + }, + { + "bbox": [ + 127, + 390, + 468, + 528 + ], + "type": "inline_equation", + "content": "26\\%" + }, + { + "bbox": [ + 127, + 390, + 468, + 528 + ], + "type": "text", + "content": " on evaluation. Note that Galactic does not report performance on the hard distribution." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 127, + 530, + 468, + 623 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 127, + 530, + 468, + 623 + ], + "spans": [ + { + "bbox": [ + 127, + 530, + 468, + 623 + ], + "type": "text", + "content": "Like AuxDistill, RL Curriculum also utilizes the auxiliary tasks yet finds no success. This demonstrates the value of concurrent training on the main and auxiliary tasks with the distillation loss in AuxDistill. Likewise, AuxDistill (No Distill) achieves no success, illustrating the importance of the distillation loss. The learning curve in Figure 3a as AuxDistill can gradually learn to complete the task with more environment interactions. On the other hand, RL Curriculum and AuxDistill (No-Distill) remain at no success regardless of the number of learning samples." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 127, + 624, + 468, + 682 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 127, + 624, + 468, + 682 + ], + "spans": [ + { + "bbox": [ + 127, + 624, + 468, + 682 + ], + "type": "text", + "content": "AuxDistill also outperforms the strongest baseline, Skill Transformer, by a significant margin with " + }, + { + "bbox": [ + 127, + 624, + 468, + 682 + ], + "type": "inline_equation", + "content": "75\\%" + }, + { + "bbox": [ + 127, + 624, + 468, + 682 + ], + "type": "text", + "content": " vs. " + }, + { + "bbox": [ + 127, + 624, + 468, + 682 + ], + "type": "inline_equation", + "content": "37\\%" + }, + { + "bbox": [ + 127, + 624, + 468, + 682 + ], + "type": "text", + "content": " on the unseen easy episodes and " + }, + { + "bbox": [ + 127, + 624, + 468, + 682 + ], + "type": "inline_equation", + "content": "41\\%" + }, + { + "bbox": [ + 127, + 624, + 468, + 682 + ], + "type": "text", + "content": " vs. " + }, + { + "bbox": [ + 127, + 624, + 468, + 682 + ], + "type": "inline_equation", + "content": "16\\%" + }, + { + "bbox": [ + 127, + 624, + 468, + 682 + ], + "type": "text", + "content": " on the unseen hard episodes. This demonstrates the advantages of using online, end-to-end RL, rather than solely training offline with demonstrations. As shown in Figure 3a, AuxDistill is able to improve with subsequent environment" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 129, + 125, + 140, + 133 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 129, + 125, + 140, + 133 + ], + "spans": [ + { + "bbox": [ + 129, + 125, + 140, + 133 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 161, + 125, + 213, + 133 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 161, + 125, + 213, + 133 + ], + "spans": [ + { + "bbox": [ + 161, + 125, + 213, + 133 + ], + "type": "text", + "content": "Harish et. al" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 188, + 146, + 385, + 197 + ], + "blocks": [ + { + "bbox": [ + 188, + 146, + 385, + 197 + ], + "lines": [ + { + "bbox": [ + 188, + 146, + 385, + 197 + ], + "spans": [ + { + "bbox": [ + 188, + 146, + 385, + 197 + ], + "type": "table", + "html": "
Method Train (Seen) Eval (Unseen)
AuxDistill12± 211± 1
Monolithic8± 09± 1
AuxDistill (No Distillation)3± 14± 1
RL-Curriculum6± 18± 0
", + "image_path": "953050dab1c464a1ce6a03b8ab1372bcd711f71f1b4f12a3b958cf138f4c6b5e.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 127, + 198, + 467, + 240 + ], + "lines": [ + { + "bbox": [ + 127, + 198, + 467, + 240 + ], + "spans": [ + { + "bbox": [ + 127, + 198, + 467, + 240 + ], + "type": "text", + "content": "Table 3: Comparison of our method on the Category Pick task with auxiliary tasks on three random seeds. AuxDistill outperforms all methods, including the monolithic baseline, on the unseen evaluation split. Evaluation is conducted over 1000 seen episodes sampled from the training distribution and 1000 held-out episodes from evaluation." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 127, + 273, + 466, + 295 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 127, + 273, + 466, + 295 + ], + "spans": [ + { + "bbox": [ + 127, + 273, + 466, + 295 + ], + "type": "text", + "content": "interactions, yet Skill Transformer is limited by the performance of the expert demonstrations." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 127, + 296, + 467, + 400 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 127, + 296, + 467, + 400 + ], + "spans": [ + { + "bbox": [ + 127, + 296, + 467, + 400 + ], + "type": "text", + "content": "Table 1 also demonstrates that AuxDistill outperforms hierarchical baselines. M3 is able to achieve " + }, + { + "bbox": [ + 127, + 296, + 467, + 400 + ], + "type": "inline_equation", + "content": "53\\%" + }, + { + "bbox": [ + 127, + 296, + 467, + 400 + ], + "type": "text", + "content": " success on the unseen easy episodes since it utilizes strong pretrained skills. Even in this setting, AuxDistill achieves a higher success of " + }, + { + "bbox": [ + 127, + 296, + 467, + 400 + ], + "type": "inline_equation", + "content": "75\\%" + }, + { + "bbox": [ + 127, + 296, + 467, + 400 + ], + "type": "text", + "content": ". However, on the hard episodes, M3 cannot dynamically plan skills and achieves no success in the hard unseen episodes. AuxDistill achieves " + }, + { + "bbox": [ + 127, + 296, + 467, + 400 + ], + "type": "inline_equation", + "content": "41\\%" + }, + { + "bbox": [ + 127, + 296, + 467, + 400 + ], + "type": "text", + "content": " on this setting because it learns a monolithic policy with RL that combines low-level and high-level decision-making. AuxDistill even outperforms an oracle version of M3 that dynamically plans the skill sequence based on oracle state information (13% vs. 41% on the unseen hard episodes)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 127, + 401, + 468, + 493 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 127, + 401, + 468, + 493 + ], + "spans": [ + { + "bbox": [ + 127, + 401, + 468, + 493 + ], + "type": "text", + "content": "We observe that AuxDistill performs " + }, + { + "bbox": [ + 127, + 401, + 468, + 493 + ], + "type": "inline_equation", + "content": "7\\%" + }, + { + "bbox": [ + 127, + 401, + 468, + 493 + ], + "type": "text", + "content": " worse on the Habitat 2.0 rearrangement training episodes than the evaluation episodes. We find this performance gap is due to additional easier episodes in the evaluation distribution where the object is closer to the target location. In the hard split, there are 55 eval episodes and 29 train episodes where the object is picked up from the fridge and the target location is " + }, + { + "bbox": [ + 127, + 401, + 468, + 493 + ], + "type": "inline_equation", + "content": "< 0.5m" + }, + { + "bbox": [ + 127, + 401, + 468, + 493 + ], + "type": "text", + "content": " from the object location. On removing these episodes, train performance is higher than evaluation, with " + }, + { + "bbox": [ + 127, + 401, + 468, + 493 + ], + "type": "inline_equation", + "content": "32\\%" + }, + { + "bbox": [ + 127, + 401, + 468, + 493 + ], + "type": "text", + "content": " success on train and " + }, + { + "bbox": [ + 127, + 401, + 468, + 493 + ], + "type": "inline_equation", + "content": "31\\%" + }, + { + "bbox": [ + 127, + 401, + 468, + 493 + ], + "type": "text", + "content": " success on evaluation." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 127, + 511, + 226, + 523 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 127, + 511, + 226, + 523 + ], + "spans": [ + { + "bbox": [ + 127, + 511, + 226, + 523 + ], + "type": "text", + "content": "4.2 Category Pick" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 127, + 531, + 467, + 634 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 127, + 531, + 467, + 634 + ], + "spans": [ + { + "bbox": [ + 127, + 531, + 467, + 634 + ], + "type": "text", + "content": "The merits of AuxDistill extend to other challenging embodied tasks. In particular, consider a variant of the pick task described in Section 4.1 where a robot has to pick up an object using the object name passed as a one-hot embedding to the policy. Category pick requires the policy to discern the object to pick up by correlating the RGB observation with the object category passed as a one-hot embedding. This task is more challenging than geometric pick where the policy has access to the initial coordinate specification of the object to be picked. We include additional details about the task specification in Section B.4 of our supplementary material." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 127, + 635, + 468, + 682 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 127, + 635, + 468, + 682 + ], + "spans": [ + { + "bbox": [ + 127, + 635, + 468, + 682 + ], + "type": "text", + "content": "We leverage the easier coordinate pick task to aid the learning of the Category Pick task using AuxDistill. More specifically, we consider the task of interest " + }, + { + "bbox": [ + 127, + 635, + 468, + 682 + ], + "type": "inline_equation", + "content": "M_0" + }, + { + "bbox": [ + 127, + 635, + 468, + 682 + ], + "type": "text", + "content": " to be the Category Pick task and " + }, + { + "bbox": [ + 127, + 635, + 468, + 682 + ], + "type": "inline_equation", + "content": "M_1" + }, + { + "bbox": [ + 127, + 635, + 468, + 682 + ], + "type": "text", + "content": " (auxiliary task we would like to distill from) to be the coordinate pick task. Both tasks are trained jointly using AuxDistill" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 392, + 124, + 435, + 134 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 392, + 124, + 435, + 134 + ], + "spans": [ + { + "bbox": [ + 392, + 124, + 435, + 134 + ], + "type": "text", + "content": "AuxDistill" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 457, + 125, + 466, + 133 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 457, + 125, + 466, + 133 + ], + "spans": [ + { + "bbox": [ + 457, + 125, + 466, + 133 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 146, + 157, + 283, + 261 + ], + "blocks": [ + { + "bbox": [ + 146, + 157, + 283, + 261 + ], + "lines": [ + { + "bbox": [ + 146, + 157, + 283, + 261 + ], + "spans": [ + { + "bbox": [ + 146, + 157, + 283, + 261 + ], + "type": "image", + "image_path": "0a95d95ec630ec17738dbe92b8640ed87facb1651b4a55af1dd8e808e9704d23.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 169, + 263, + 271, + 272 + ], + "lines": [ + { + "bbox": [ + 169, + 263, + 271, + 272 + ], + "spans": [ + { + "bbox": [ + 169, + 263, + 271, + 272 + ], + "type": "text", + "content": "(a) Rearrange learning curve." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 298, + 157, + 439, + 262 + ], + "blocks": [ + { + "bbox": [ + 298, + 157, + 439, + 262 + ], + "lines": [ + { + "bbox": [ + 298, + 157, + 439, + 262 + ], + "spans": [ + { + "bbox": [ + 298, + 157, + 439, + 262 + ], + "type": "image", + "image_path": "97823b0f54d475f18f16446b921070e36c9ddb6bc7ed0a89da9d79b090b2be24.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 326, + 263, + 423, + 272 + ], + "lines": [ + { + "bbox": [ + 326, + 263, + 423, + 272 + ], + "spans": [ + { + "bbox": [ + 326, + 263, + 423, + 272 + ], + "type": "text", + "content": "(b) Category Pick learning." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 127, + 280, + 468, + 344 + ], + "lines": [ + { + "bbox": [ + 127, + 280, + 468, + 344 + ], + "spans": [ + { + "bbox": [ + 127, + 280, + 468, + 344 + ], + "type": "text", + "content": "Fig. 3: Left: RL training success rates on training episodes on the rearrangement task from Table 1. Note that AuxDistill (No Distill) and RL-Curriculum are displayed but achieve " + }, + { + "bbox": [ + 127, + 280, + 468, + 344 + ], + "type": "inline_equation", + "content": "0\\%" + }, + { + "bbox": [ + 127, + 280, + 468, + 344 + ], + "type": "text", + "content": " success throughout training. Right: Learning curve on the Category Pick task of AuxDistill utilizing coordinate pick distillation vs. monolithic RL. AuxDistill outperforms baselines in both settings. Displayed are averages and standard deviation over 3 random seeds." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 127, + 367, + 466, + 401 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 127, + 367, + 466, + 401 + ], + "spans": [ + { + "bbox": [ + 127, + 367, + 466, + 401 + ], + "type": "text", + "content": "with the same reward structure with " + }, + { + "bbox": [ + 127, + 367, + 466, + 401 + ], + "type": "inline_equation", + "content": "w_{i}(s_{t}) = 1" + }, + { + "bbox": [ + 127, + 367, + 466, + 401 + ], + "type": "text", + "content": " and use " + }, + { + "bbox": [ + 127, + 367, + 466, + 401 + ], + "type": "inline_equation", + "content": "\\lambda = 0.5" + }, + { + "bbox": [ + 127, + 367, + 466, + 401 + ], + "type": "text", + "content": " for distillation in Equation (1). For this task, we train on the full rearrange-easy dataset from [44] with 50,000 episodes." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 127, + 402, + 468, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 127, + 402, + 468, + 483 + ], + "spans": [ + { + "bbox": [ + 127, + 402, + 468, + 483 + ], + "type": "text", + "content": "We report the training curves of the Category Pick task in comparison with the RL Curriculum and No Distillation baseline used in Table 1. In addition, we introduce a monolithic baseline for this task, which directly trains Category Pick without leveraging the coordinate pick task during training. We report the training curves in Figure 3b. We evaluate the performance of the trained policy on 1000 episodes sampled from the training and 1000 episodes from the validation split." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 127, + 484, + 468, + 554 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 127, + 484, + 468, + 554 + ], + "spans": [ + { + "bbox": [ + 127, + 484, + 468, + 554 + ], + "type": "text", + "content": "In Table 3, we show that AuxDistill shows better performance with a success rate of " + }, + { + "bbox": [ + 127, + 484, + 468, + 554 + ], + "type": "inline_equation", + "content": "11\\%" + }, + { + "bbox": [ + 127, + 484, + 468, + 554 + ], + "type": "text", + "content": " outperforming all baselines that do not leverage distillation during RL training. The closest performing baseline to ours is the monolithic baseline, which achieves a success of " + }, + { + "bbox": [ + 127, + 484, + 468, + 554 + ], + "type": "inline_equation", + "content": "9\\%" + }, + { + "bbox": [ + 127, + 484, + 468, + 554 + ], + "type": "text", + "content": " on the held-out distribution. AuxDistill (No-Distillation) and RL-curriculum perform worse with an evaluation success of " + }, + { + "bbox": [ + 127, + 484, + 468, + 554 + ], + "type": "inline_equation", + "content": "4\\%" + }, + { + "bbox": [ + 127, + 484, + 468, + 554 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 127, + 484, + 468, + 554 + ], + "type": "inline_equation", + "content": "8\\%" + }, + { + "bbox": [ + 127, + 484, + 468, + 554 + ], + "type": "text", + "content": ", respectively." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 127, + 570, + 252, + 582 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 127, + 570, + 252, + 582 + ], + "spans": [ + { + "bbox": [ + 127, + 570, + 252, + 582 + ], + "type": "text", + "content": "4.3 AuxDistill Analysis" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 127, + 589, + 468, + 658 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 127, + 589, + 468, + 658 + ], + "spans": [ + { + "bbox": [ + 127, + 589, + 468, + 658 + ], + "type": "text", + "content": "Robustness to Auxiliary Task Selection A critical consideration for training AuxDistill is the selection of auxiliary tasks. In this section, we compare four different selections of auxiliary tasks in the object rearrangement task from Section 4.1. We conduct this ablation on a smaller distribution of episodes, which only include the Easy episodes and Hard episodes where the object starts in the fridge and train AuxDistill for 100M steps." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 127, + 659, + 468, + 682 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 127, + 659, + 468, + 682 + ], + "spans": [ + { + "bbox": [ + 127, + 659, + 468, + 682 + ], + "type": "text", + "content": "As we conduct this study only on the fridge category of articulated episodes, we modify the auxiliary task selection to be: Pick, Open Fridge, Place and Pick" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 129, + 125, + 141, + 133 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 129, + 125, + 141, + 133 + ], + "spans": [ + { + "bbox": [ + 129, + 125, + 141, + 133 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 161, + 125, + 213, + 133 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 161, + 125, + 213, + 133 + ], + "spans": [ + { + "bbox": [ + 161, + 125, + 213, + 133 + ], + "type": "text", + "content": "Harish et. al" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "bbox": [ + 127, + 148, + 467, + 184 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 127, + 148, + 467, + 184 + ], + "spans": [ + { + "bbox": [ + 127, + 148, + 467, + 184 + ], + "type": "text", + "content": "from Fridge. Note that the first three auxiliary tasks are the same from Section 4.1. The other auxiliary task selections are the same as this original selection, but excluding one of: Pick, Open Fridge, and Pick from Fridge." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 127, + 184, + 468, + 286 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 127, + 184, + 468, + 286 + ], + "spans": [ + { + "bbox": [ + 127, + 184, + 468, + 286 + ], + "type": "text", + "content": "In Table 2, we report the performance for each skill selection on the easy and hard episodes. We observe that among all the sub-tasks, the most important is Pick, as excluding it results in " + }, + { + "bbox": [ + 127, + 184, + 468, + 286 + ], + "type": "inline_equation", + "content": "0.0\\%" + }, + { + "bbox": [ + 127, + 184, + 468, + 286 + ], + "type": "text", + "content": " success. Note that Pick is only relevant " + }, + { + "bbox": [ + 127, + 184, + 468, + 286 + ], + "type": "inline_equation", + "content": "(w_{i}(s_{t}) > 0)" + }, + { + "bbox": [ + 127, + 184, + 468, + 286 + ], + "type": "text", + "content": " to easy episodes during training i.e. removing it should only impact the performance on easy distribution of our training setup. However, we find that it affects both the easy and hard episodes as the task is not successfully carried out in either case if removed. In Figure 2, we analyze the success of individual stages of rearrangement and notice that not including Pick leads to failure earlier on in the task." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 127, + 288, + 468, + 416 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 127, + 288, + 468, + 416 + ], + "spans": [ + { + "bbox": [ + 127, + 288, + 468, + 416 + ], + "type": "text", + "content": "In Table 5 we notice that introducing the Open-Fridge skill can worsen rearrangement performance (30% vs. 48%). The reason for this is that the open-fridge skill is the easiest of all auxiliary tasks (see Appendix B.3), and utilizing it reduces the number of samples allocated to the main task. However, certain auxiliary tasks can boost performance by addressing specific bottlenecks in the rearrangement. One such example is the Pick from Fridge, which boosts performance 7% to 19% by addressing the task of picking from an open fridge, which can be challenging due to the difficulty of avoiding collisions with the fridge door. Further, the 4% performance improvement from 32% to 36% on easy distribution can be attributed to the presence of easy episodes in our dataset with an open-refrigerator where Pick from Fridge helps boost performance." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 127, + 432, + 468, + 560 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 127, + 432, + 468, + 560 + ], + "spans": [ + { + "bbox": [ + 127, + 432, + 468, + 560 + ], + "type": "text", + "content": "Distillation Coefficient Variation We analyze the effect of the distillation loss coefficient " + }, + { + "bbox": [ + 127, + 432, + 468, + 560 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 127, + 432, + 468, + 560 + ], + "type": "text", + "content": " in AuxDistill in Table 4. The distillation loss coefficient controls how the agent balances minimizing the distillation from the auxiliary tasks versus maximizing the reward on the main task. In Table 4, we notice that using high values of distillation weighting " + }, + { + "bbox": [ + 127, + 432, + 468, + 560 + ], + "type": "inline_equation", + "content": "\\lambda = 1.0" + }, + { + "bbox": [ + 127, + 432, + 468, + 560 + ], + "type": "text", + "content": " shifts the objective of the rearrangement task from maximizing cumulative reward to distilling from the auxiliary task leading to " + }, + { + "bbox": [ + 127, + 432, + 468, + 560 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 127, + 432, + 468, + 560 + ], + "type": "text", + "content": " success rate on rearrangement and similarly, using a very small value of distillation coefficient " + }, + { + "bbox": [ + 127, + 432, + 468, + 560 + ], + "type": "inline_equation", + "content": "\\lambda = 0.01" + }, + { + "bbox": [ + 127, + 432, + 468, + 560 + ], + "type": "text", + "content": " leads to insufficient leveraging of auxiliary task information with a success rate of " + }, + { + "bbox": [ + 127, + 432, + 468, + 560 + ], + "type": "inline_equation", + "content": "7\\%" + }, + { + "bbox": [ + 127, + 432, + 468, + 560 + ], + "type": "text", + "content": ". We find " + }, + { + "bbox": [ + 127, + 432, + 468, + 560 + ], + "type": "inline_equation", + "content": "\\lambda = 0.1, 0.05" + }, + { + "bbox": [ + 127, + 432, + 468, + 560 + ], + "type": "text", + "content": " to be good choices for optimizing task reward and distilling behaviors from the auxiliary tasks." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 127, + 577, + 214, + 589 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 127, + 577, + 214, + 589 + ], + "spans": [ + { + "bbox": [ + 127, + 577, + 214, + 589 + ], + "type": "text", + "content": "5 Conclusion" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 127, + 600, + 468, + 682 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 127, + 600, + 468, + 682 + ], + "spans": [ + { + "bbox": [ + 127, + 600, + 468, + 682 + ], + "type": "text", + "content": "In this work, we presented AuxDistill, a new method for end-to-end RL on complex tasks by leveraging auxiliary tasks. AuxDistill learns in the auxiliary tasks concurrently with the main task through multi-task RL. A distillation loss transfers relevant behaviors from the auxiliary tasks to the main task. We show that AuxDistill outperforms a variety of baselines in Habitat object rearrangement by up to " + }, + { + "bbox": [ + 127, + 600, + 468, + 682 + ], + "type": "inline_equation", + "content": "27\\%" + }, + { + "bbox": [ + 127, + 600, + 468, + 682 + ], + "type": "text", + "content": ". We also show another application of AuxDistill in a category-conditioned manipulation task. Finally, we analyze AuxDistill in" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 392, + 125, + 435, + 134 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 392, + 125, + 435, + 134 + ], + "spans": [ + { + "bbox": [ + 392, + 125, + 435, + 134 + ], + "type": "text", + "content": "AuxDistill" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 457, + 125, + 467, + 133 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 457, + 125, + 467, + 133 + ], + "spans": [ + { + "bbox": [ + 457, + 125, + 467, + 133 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 172, + 146, + 417, + 232 + ], + "blocks": [ + { + "bbox": [ + 172, + 146, + 417, + 232 + ], + "lines": [ + { + "bbox": [ + 172, + 146, + 417, + 232 + ], + "spans": [ + { + "bbox": [ + 172, + 146, + 417, + 232 + ], + "type": "table", + "html": "
MethodTrain (Seen)Eval (Unseen)
All EpisodesEasyHardAll EpisodesEasyHard
λ = 0.0166510182
λ = 0.05303128363338
λ = 0.1222619303921
λ = 0.5230250
λ = 1.0010120
", + "image_path": "205a116127971f7e5b3b05e8445b1a61054e8624ddbb52618700ae9e7e3c68af.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 127, + 233, + 468, + 286 + ], + "lines": [ + { + "bbox": [ + 127, + 233, + 468, + 286 + ], + "spans": [ + { + "bbox": [ + 127, + 233, + 468, + 286 + ], + "type": "text", + "content": "Table 4: Performance of our method for a single seed on varying distillation coefficient during training. Using a large distillation coefficient " + }, + { + "bbox": [ + 127, + 233, + 468, + 286 + ], + "type": "inline_equation", + "content": "(\\lambda = 1.0)" + }, + { + "bbox": [ + 127, + 233, + 468, + 286 + ], + "type": "text", + "content": " makes reward optimization challenging, and too small " + }, + { + "bbox": [ + 127, + 233, + 468, + 286 + ], + "type": "inline_equation", + "content": "(\\lambda = 0.01)" + }, + { + "bbox": [ + 127, + 233, + 468, + 286 + ], + "type": "text", + "content": " results in insufficient auxiliary distillation to succeed on the rearrangement task. The intermediate values " + }, + { + "bbox": [ + 127, + 233, + 468, + 286 + ], + "type": "inline_equation", + "content": "\\lambda = 0.05" + }, + { + "bbox": [ + 127, + 233, + 468, + 286 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 127, + 233, + 468, + 286 + ], + "type": "inline_equation", + "content": "\\lambda = 0.1" + }, + { + "bbox": [ + 127, + 233, + 468, + 286 + ], + "type": "text", + "content": " show the best performance on training and evaluation." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 127, + 319, + 467, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 127, + 319, + 467, + 354 + ], + "spans": [ + { + "bbox": [ + 127, + 319, + 467, + 354 + ], + "type": "text", + "content": "different auxiliary task selections and magnitudes of distillation strengths. Overall, AuxDistill presents a new way to tackle compound tasks with RL alone without requiring pre-trained skills or expert demonstrations." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 127, + 354, + 468, + 424 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 127, + 354, + 468, + 424 + ], + "spans": [ + { + "bbox": [ + 127, + 354, + 468, + 424 + ], + "type": "text", + "content": "Limitations of AuxDistill include the need for the auxiliary tasks and the auxiliary task relevance function. The auxiliary tasks require knowing behaviors that are relevant and easier to learn than the main task. Each auxiliary task requires defining a new start state distribution and reward function. However, designing these can rely on privileged state information which is less restrictive than the requirement of pre-trained skills or expert demonstrations." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 128, + 443, + 193, + 454 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 128, + 443, + 193, + 454 + ], + "spans": [ + { + "bbox": [ + 128, + 443, + 193, + 454 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 467, + 468, + 682 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 132, + 467, + 468, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 467, + 468, + 499 + ], + "spans": [ + { + "bbox": [ + 132, + 467, + 468, + 499 + ], + "type": "text", + "content": "1. Achiam, J., Adler, S., Agarwal, S., Ahmad, L., Akkaya, I., Aleman, F.L., Almeida, D., Altenschmidt, J., Altman, S., Anadkat, S., et al.: Gpt-4 technical report. arXiv preprint arXiv:2303.08774 (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 501, + 468, + 532 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 501, + 468, + 532 + ], + "spans": [ + { + "bbox": [ + 132, + 501, + 468, + 532 + ], + "type": "text", + "content": "2. Agarwal, A., Kumar, A., Malik, J., Pathak, D.: Legged locomotion in challenging terrains using egocentric vision. In: Conference on Robot Learning. pp. 403-415. PMLR (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 533, + 468, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 533, + 468, + 564 + ], + "spans": [ + { + "bbox": [ + 132, + 533, + 468, + 564 + ], + "type": "text", + "content": "3. Akkaya, I., Andrychowicz, M., Chociej, M., Litwin, M., McGrew, B., Petron, A., Paino, A., Plappert, M., Powell, G., Ribas, R., et al.: Solving rubik's cube with a robot hand. arXiv preprint arXiv:1910.07113 (2019)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 565, + 468, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 565, + 468, + 597 + ], + "spans": [ + { + "bbox": [ + 132, + 565, + 468, + 597 + ], + "type": "text", + "content": "4. Azad, A.S., Gur, I., Emhoff, J., Alexis, N., Faust, A., Abbeel, P., Stoica, I.: Clutr: Curriculum learning via unsupervised task representation learning. In: International Conference on Machine Learning. pp. 1361-1395. PMLR (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 597, + 468, + 618 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 597, + 468, + 618 + ], + "spans": [ + { + "bbox": [ + 132, + 597, + 468, + 618 + ], + "type": "text", + "content": "5. Bacon, P.L., Harb, J., Precup, D.: The option-critic architecture. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 31 (2017)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 618, + 468, + 649 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 618, + 468, + 649 + ], + "spans": [ + { + "bbox": [ + 132, + 618, + 468, + 649 + ], + "type": "text", + "content": "6. Baker, B., Kanitscheider, I., Markov, T., Wu, Y., Powell, G., McGrew, B., Mordatch, I.: Emergent tool use from multi-agent autocurricula. arXiv preprint arXiv:1909.07528 (2019)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 650, + 468, + 682 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 650, + 468, + 682 + ], + "spans": [ + { + "bbox": [ + 132, + 650, + 468, + 682 + ], + "type": "text", + "content": "7. Batra, D., Chang, A.X., Chernova, S., Davison, A.J., Deng, J., Koltun, V., Levine, S., Malik, J., Mordatch, I., Mottaghi, R., et al.: Rearrangement: A challenge for embodied ai. arXiv preprint arXiv:2011.01975 (2020)" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 129, + 125, + 140, + 133 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 129, + 125, + 140, + 133 + ], + "spans": [ + { + "bbox": [ + 129, + 125, + 140, + 133 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 161, + 125, + 213, + 133 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 161, + 125, + 213, + 133 + ], + "spans": [ + { + "bbox": [ + 161, + 125, + 213, + 133 + ], + "type": "text", + "content": "Harish et. al" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 129, + 149, + 468, + 682 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 132, + 149, + 467, + 191 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 149, + 467, + 191 + ], + "spans": [ + { + "bbox": [ + 132, + 149, + 467, + 191 + ], + "type": "text", + "content": "8. Berges, V.P., Szot, A., Chaplot, D.S., Gokaslan, A., Mottaghi, R., Batra, D., Undersander, E.: Galactic: Scaling end-to-end reinforcement learning for rearrangement at 100k steps-per-second. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13767-13777 (2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 192, + 467, + 222 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 192, + 467, + 222 + ], + "spans": [ + { + "bbox": [ + 132, + 192, + 467, + 222 + ], + "type": "text", + "content": "9. Berner, C., Brockman, G., Chan, B., Cheung, V., Debiak, P., Dennison, C., Farhi, D., Fischer, Q., Hashme, S., Hesse, C., et al.: Dota 2 with large scale deep reinforcement learning. arXiv preprint arXiv:1912.06680 (2019)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 223, + 467, + 243 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 223, + 467, + 243 + ], + "spans": [ + { + "bbox": [ + 131, + 223, + 467, + 243 + ], + "type": "text", + "content": "10. Chen, Y., Wang, C., Fei-Fei, L., Liu, C.K.: Sequential dexterity: Chaining dexterous policies for long-horizon manipulation. arXiv preprint arXiv:2309.00987 (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 129, + 244, + 467, + 275 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 129, + 244, + 467, + 275 + ], + "spans": [ + { + "bbox": [ + 129, + 244, + 467, + 275 + ], + "type": "text", + "content": "11. Dalal, M., Pathak, D., Salakhutdinov, R.R.: Accelerating robotic reinforcement learning via parameterized action primitives. Advances in Neural Information Processing Systems 34, 21847-21859 (2021)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 129, + 275, + 468, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 129, + 275, + 468, + 306 + ], + "spans": [ + { + "bbox": [ + 129, + 275, + 468, + 306 + ], + "type": "text", + "content": "12. Dennis, M., Jaques, N., Vinitsky, E., Bayen, A., Russell, S., Critch, A., Levine, S.: Emergent complexity and zero-shot transfer via unsupervised environment design. Advances in neural information processing systems 33, 13049-13061 (2020)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 129, + 306, + 468, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 129, + 306, + 468, + 338 + ], + "spans": [ + { + "bbox": [ + 129, + 306, + 468, + 338 + ], + "type": "text", + "content": "13. Fang, K., Migimatsu, T., Mandlekar, A., Fei-Fei, L., Bohg, J.: Active task randomization: Learning visuomotor skills for sequential manipulation by proposing feasible and novel tasks. arXiv preprint arXiv:2211.06134 (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 129, + 338, + 468, + 359 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 129, + 338, + 468, + 359 + ], + "spans": [ + { + "bbox": [ + 129, + 338, + 468, + 359 + ], + "type": "text", + "content": "14. Fikes, R.E., Nilsson, N.J.: Strips: A new approach to the application of theorem proving to problem solving. Artificial intelligence 2(3-4), 189-208 (1971)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 129, + 359, + 468, + 390 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 129, + 359, + 468, + 390 + ], + "spans": [ + { + "bbox": [ + 129, + 359, + 468, + 390 + ], + "type": "text", + "content": "15. Fu, Z., Cheng, X., Pathak, D.: Deep whole-body control: learning a unified policy for manipulation and locomotion. In: Conference on Robot Learning. pp. 138-149. PMLR (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 129, + 390, + 468, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 129, + 390, + 468, + 421 + ], + "spans": [ + { + "bbox": [ + 129, + 390, + 468, + 421 + ], + "type": "text", + "content": "16. Gregor, K., Jimenez Rezende, D., Besse, F., Wu, Y., Merzic, H., van den Oord, A.: Shaping belief states with generative environment models for rl. Advances in Neural Information Processing Systems 32 (2019)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 129, + 421, + 468, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 129, + 421, + 468, + 442 + ], + "spans": [ + { + "bbox": [ + 129, + 421, + 468, + 442 + ], + "type": "text", + "content": "17. Gu, J., Chaplot, D.S., Su, H., Malik, J.: Multi-skill mobile manipulation for object rearrangement. arXiv preprint arXiv:2209.02778 (2022)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 129, + 442, + 468, + 463 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 129, + 442, + 468, + 463 + ], + "spans": [ + { + "bbox": [ + 129, + 442, + 468, + 463 + ], + "type": "text", + "content": "18. Hafner, D., Lee, K.H., Fischer, I., Abbeel, P.: Deep hierarchical planning from pixels. Advances in Neural Information Processing Systems 35, 26091-26104 (2022)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 129, + 463, + 468, + 494 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 129, + 463, + 468, + 494 + ], + "spans": [ + { + "bbox": [ + 129, + 463, + 468, + 494 + ], + "type": "text", + "content": "19. Harutyunyan, A., Dabney, W., Mesnard, T., Gheshlaghi Azar, M., Piot, B., Heess, N., van Hasselt, H.P., Wayne, G., Singh, S., Precup, D., et al.: Hindsight credit assignment. Advances in neural information processing systems 32 (2019)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 129, + 494, + 468, + 513 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 129, + 494, + 468, + 513 + ], + "spans": [ + { + "bbox": [ + 129, + 494, + 468, + 513 + ], + "type": "text", + "content": "20. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. CVPR (2016)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 129, + 514, + 468, + 545 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 129, + 514, + 468, + 545 + ], + "spans": [ + { + "bbox": [ + 129, + 514, + 468, + 545 + ], + "type": "text", + "content": "21. Hessel, M., Soyer, H., Espeholt, L., Czarnecki, W., Schmitt, S., Van Hasselt, H.: Multi-task deep reinforcement learning with popart. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 33, pp. 3796-3803 (2019)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 129, + 545, + 468, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 129, + 545, + 468, + 566 + ], + "spans": [ + { + "bbox": [ + 129, + 545, + 468, + 566 + ], + "type": "text", + "content": "22. Huang, X., Batra, D., Rai, A., Szot, A.: Skill transformer: A monolithic policy for mobile manipulation. arXiv preprint arXiv:2308.09873 (2023)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 129, + 566, + 468, + 629 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 129, + 566, + 468, + 629 + ], + "spans": [ + { + "bbox": [ + 129, + 566, + 468, + 629 + ], + "type": "text", + "content": "23. Kalashnikov, D., Irpan, A., Pastor, P., Ibarz, J., Herzog, A., Jang, E., Quillen, D., Holly, E., Kalakrishnan, M., Vanhoucke, V., Levine, S.: Scalable deep reinforcement learning for vision-based robotic manipulation. In: 2nd Annual Conference on Robot Learning, CoRL 2018, Zürich, Switzerland, 29-31 October 2018, Proceedings. Proceedings of Machine Learning Research, vol. 87, pp. 651-673. PMLR (2018), http://proceedings.mlr.press/v87 kalashnikov18a.html" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 129, + 629, + 468, + 661 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 129, + 629, + 468, + 661 + ], + "spans": [ + { + "bbox": [ + 129, + 629, + 468, + 661 + ], + "type": "text", + "content": "24. Karkus, P., Mirza, M., Guez, A., Jaegle, A., Lillicrap, T., Buesing, L., Heess, N., Weber, T.: Beyond tabula-rama: a modular reinforcement learning approach for physically embedded 3d sukoban. arXiv preprint arXiv:2010.01298 (2020)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 129, + 661, + 468, + 682 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 129, + 661, + 468, + 682 + ], + "spans": [ + { + "bbox": [ + 129, + 661, + 468, + 682 + ], + "type": "text", + "content": "25. Katara, P., Xian, Z., Fragkiadaki, K.: Gen2sim: Scaling up robot learning in simulation with generative models. arXiv preprint arXiv:2310.18308 (2023)" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 392, + 124, + 435, + 134 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 392, + 124, + 435, + 134 + ], + "spans": [ + { + "bbox": [ + 392, + 124, + 435, + 134 + ], + "type": "text", + "content": "AuxDistill" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 457, + 125, + 467, + 133 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 457, + 125, + 467, + 133 + ], + "spans": [ + { + "bbox": [ + 457, + 125, + 467, + 133 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 129, + 149, + 468, + 682 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 129, + 149, + 467, + 170 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 129, + 149, + 467, + 170 + ], + "spans": [ + { + "bbox": [ + 129, + 149, + 467, + 170 + ], + "type": "text", + "content": "26. Kumar, A., Fu, Z., Pathak, D., Malik, J.: Rapid motor adaptation for legged robots. RSS (2021)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 129, + 170, + 468, + 211 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 129, + 170, + 468, + 211 + ], + "spans": [ + { + "bbox": [ + 129, + 170, + 468, + 211 + ], + "type": "text", + "content": "27. Kuo, C.W., Ma, C.Y., Hoffman, J., Kira, Z.: Structure-encoding auxiliary tasks for improved visual representation in vision-and-language navigation. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 1104-1113 (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 129, + 212, + 467, + 243 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 129, + 212, + 467, + 243 + ], + "spans": [ + { + "bbox": [ + 129, + 212, + 467, + 243 + ], + "type": "text", + "content": "28. Lee, Y., Lim, J.J., Anandkumar, A., Zhu, Y.: Adversarial skill chaining for long-horizon robot manipulation via terminal state regularization. arXiv preprint arXiv:2111.07999 (2021)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 129, + 243, + 467, + 275 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 129, + 243, + 467, + 275 + ], + "spans": [ + { + "bbox": [ + 129, + 243, + 467, + 275 + ], + "type": "text", + "content": "29. Liu, B., Tur, G., Hakkani-Tur, D., Shah, P., Heck, L.: End-to-end optimization of task-oriented dialogue model with deep reinforcement learning. arXiv preprint arXiv:1711.10712 (2017)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 129, + 275, + 467, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 129, + 275, + 467, + 306 + ], + "spans": [ + { + "bbox": [ + 129, + 275, + 467, + 306 + ], + "type": "text", + "content": "30. Mishra, U.A., Xue, S., Chen, Y., Xu, D.: Generative skill chaining: Long-horizon skill planning with diffusion models. In: Conference on Robot Learning. pp. 2905-2925. PMLR (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 129, + 306, + 467, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 129, + 306, + 467, + 338 + ], + "spans": [ + { + "bbox": [ + 129, + 306, + 467, + 338 + ], + "type": "text", + "content": "31. Mnih, V., Kavukcuoglu, K., Silver, D., Graves, A., Antonoglou, I., Wierstra, D., Riedmiller, M.: Playing atari with deep reinforcement learning. arXiv preprint arXiv:1312.5602 (2013)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 129, + 338, + 467, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 129, + 338, + 467, + 369 + ], + "spans": [ + { + "bbox": [ + 129, + 338, + 467, + 369 + ], + "type": "text", + "content": "32. Nakano, R., Hilton, J., Balaji, S., Wu, J., Ouyang, L., Kim, C., Hesse, C., Jain, S., Kosaraju, V., Saunders, W., et al.: Webgpt: Browser-assisted question-answering with human feedback. arXiv preprint arXiv:2112.09332 (2021)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 129, + 369, + 467, + 400 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 129, + 369, + 467, + 400 + ], + "spans": [ + { + "bbox": [ + 129, + 369, + 467, + 400 + ], + "type": "text", + "content": "33. Narvekar, S., Peng, B., Leonetti, M., Sinapov, J., Taylor, M.E., Stone, P.: Curriculum learning for reinforcement learning domains: A framework and survey. The Journal of Machine Learning Research 21(1), 7382-7431 (2020)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 129, + 400, + 467, + 432 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 129, + 400, + 467, + 432 + ], + "spans": [ + { + "bbox": [ + 129, + 400, + 467, + 432 + ], + "type": "text", + "content": "34. Ni, T., Ma, M., Eysenbach, B., Bacon, P.L.: When do transformers shine in rl? decoupling memory from credit assignment. Advances in Neural Information Processing Systems 36 (2024)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 129, + 432, + 467, + 463 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 129, + 432, + 467, + 463 + ], + "spans": [ + { + "bbox": [ + 129, + 432, + 467, + 463 + ], + "type": "text", + "content": "35. Qi, H., Kumar, A., Calandra, R., Ma, Y., Malik, J.: In-hand object rotation via rapid motor adaptation. In: Conference on Robot Learning. pp. 1722-1732. PMLR (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 129, + 463, + 467, + 484 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 129, + 463, + 467, + 484 + ], + "spans": [ + { + "bbox": [ + 129, + 463, + 467, + 484 + ], + "type": "text", + "content": "36. Radosavovic, I., Xiao, T., Zhang, B., Darrell, T., Malik, J., Sreenath, K.: Learning humanoid locomotion with transformers. arXiv preprint arXiv:2303.03381 (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 129, + 484, + 366, + 493 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 129, + 484, + 366, + 493 + ], + "spans": [ + { + "bbox": [ + 129, + 484, + 366, + 493 + ], + "type": "text", + "content": "37. robotics, F.: Fetch. http://fetchrobotics.com/ (2020)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 129, + 493, + 467, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 129, + 493, + 467, + 525 + ], + "spans": [ + { + "bbox": [ + 129, + 493, + 467, + 525 + ], + "type": "text", + "content": "38. Rudin, N., Hoeller, D., Reist, P., Hutter, M.: Learning to walk in minutes using massively parallel deep reinforcement learning. In: Conference on Robot Learning pp. 91-100. PMLR (2022)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 129, + 525, + 467, + 545 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 129, + 525, + 467, + 545 + ], + "spans": [ + { + "bbox": [ + 129, + 525, + 467, + 545 + ], + "type": "text", + "content": "39. Schulman, J., Wolski, F., Dhariwal, P., Radford, A., Klimov, O.: Proximal policy optimization algorithms. arXiv preprint arXiv:1707.06347 (2017)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 129, + 545, + 467, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 129, + 545, + 467, + 577 + ], + "spans": [ + { + "bbox": [ + 129, + 545, + 467, + 577 + ], + "type": "text", + "content": "40. Shah, P., Hakkani-Tür, D., Heck, L.: Interactive reinforcement learning for task-oriented dialogue management. In: NIPS 2016 Deep Learning for Action and Interaction Workshop. vol. 11 (2016)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 129, + 577, + 467, + 608 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 129, + 577, + 467, + 608 + ], + "spans": [ + { + "bbox": [ + 129, + 577, + 467, + 608 + ], + "type": "text", + "content": "41. Silver, D., Schrittwieser, J., Simonyan, K., Antonoglou, I., Huang, A., Guez, A., Hubert, T., Baker, L., Lai, M., Bolton, A., et al.: Mastering the game of go without human knowledge. nature 550(7676), 354-359 (2017)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 129, + 608, + 467, + 640 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 129, + 608, + 467, + 640 + ], + "spans": [ + { + "bbox": [ + 129, + 608, + 467, + 640 + ], + "type": "text", + "content": "42. Sutton, R.S., Precup, D., Singh, S.: Between mdps and semi-mdps: A framework for temporal abstraction in reinforcement learning. Artificial intelligence 112(1-2), 181-211 (1999)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 129, + 640, + 467, + 682 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 129, + 640, + 467, + 682 + ], + "spans": [ + { + "bbox": [ + 129, + 640, + 467, + 682 + ], + "type": "text", + "content": "43. Szot, A., Clegg, A., Undersander, E., Wijmans, E., Zhao, Y., Turner, J., Maestre, N., Mukadam, M., Chaplot, D.S., Maksymets, O., et al.: Habitat 2.0: Training home assistants to rearrange their habitat. Advances in Neural Information Processing Systems 34 (2021)" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 129, + 125, + 141, + 133 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 129, + 125, + 141, + 133 + ], + "spans": [ + { + "bbox": [ + 129, + 125, + 141, + 133 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 161, + 125, + 213, + 133 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 161, + 125, + 213, + 133 + ], + "spans": [ + { + "bbox": [ + 161, + 125, + 213, + 133 + ], + "type": "text", + "content": "Harish et. al" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 127, + 148, + 468, + 394 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 127, + 148, + 468, + 181 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 127, + 148, + 468, + 181 + ], + "spans": [ + { + "bbox": [ + 127, + 148, + 468, + 181 + ], + "type": "text", + "content": "44. Szot, A., Yadav, K., Clegg, A., Berges, V.P., Gokaslan, A., Chang, A., Savva, M., Kira, Z., Batra, D.: Habitat rearrangement challenge 2022. https://aihabitat.org/challenge/2022_rearrange (2022)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 129, + 182, + 468, + 213 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 129, + 182, + 468, + 213 + ], + "spans": [ + { + "bbox": [ + 129, + 182, + 468, + 213 + ], + "type": "text", + "content": "45. Touvron, H., Lavril, T., Izacard, G., Martinet, X., Lachaux, M.A., Lacroix, T., Rozière, B., Goyal, N., Hambro, E., Azhar, F., et al.: Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971 (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 129, + 214, + 468, + 255 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 129, + 214, + 468, + 255 + ], + "spans": [ + { + "bbox": [ + 129, + 214, + 468, + 255 + ], + "type": "text", + "content": "46. Vezzani, G., Tirumala, D., Wulfmeier, M., Rao, D., Abdelmaleki, A., Moran, B., Haarnoja, T., Humplik, J., Hafner, R., Neunert, M., et al.: Skills: Adaptive skill sequencing for efficient temporally-extended exploration. arXiv preprint arXiv:2211.13743 (2022)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 129, + 256, + 468, + 287 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 129, + 256, + 468, + 287 + ], + "spans": [ + { + "bbox": [ + 129, + 256, + 468, + 287 + ], + "type": "text", + "content": "47. Wijmans, E., Kadian, A., Morcos, A., Lee, S., Essa, I., Parikh, D., Savva, M., Batra, D.: Dd-ppo: Learning near-perfect pointgoal navigators from 2.5 billion frames. arXiv preprint arXiv:1911.00357 (2019)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 129, + 288, + 468, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 129, + 288, + 468, + 319 + ], + "spans": [ + { + "bbox": [ + 129, + 288, + 468, + 319 + ], + "type": "text", + "content": "48. Xia, F., Li, C., Martin-Martin, R., Litany, O., Toshev, A., Savarese, S.: Relmogen: Leveraging motion generation in reinforcement learning for mobile manipulation. arXiv preprint arXiv:2008.07792 (2020)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 129, + 320, + 467, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 129, + 320, + 467, + 352 + ], + "spans": [ + { + "bbox": [ + 129, + 320, + 467, + 352 + ], + "type": "text", + "content": "49. Ye, J., Batra, D., Das, A., Wijmans, E.: Auxiliary tasks and exploration enable objectgoal navigation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 16117-16126 (2021)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 129, + 353, + 467, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 129, + 353, + 467, + 373 + ], + "spans": [ + { + "bbox": [ + 129, + 353, + 467, + 373 + ], + "type": "text", + "content": "50. Ye, J., Batra, D., Wijmans, E., Das, A.: Auxiliary tasks speed up learning point goal navigation. In: Conference on Robot Learning. pp. 498-516. PMLR (2021)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 129, + 374, + 467, + 394 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 129, + 374, + 467, + 394 + ], + "spans": [ + { + "bbox": [ + 129, + 374, + 467, + 394 + ], + "type": "text", + "content": "51. Zhang, S., Whiteson, S.: Dac: The double actor-critic architecture for learning options. Advances in Neural Information Processing Systems 32 (2019)" + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 393, + 125, + 435, + 134 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 393, + 125, + 435, + 134 + ], + "spans": [ + { + "bbox": [ + 393, + 125, + 435, + 134 + ], + "type": "text", + "content": "AuxDistill" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 457, + 125, + 467, + 133 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 457, + 125, + 467, + 133 + ], + "spans": [ + { + "bbox": [ + 457, + 125, + 467, + 133 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 16 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Rejection Sampling IMLE_ Designing Priors for Better Few-Shot Image Synthesis/0d5c28ac-bfc0-4636-af16-85b85c6b5148_content_list.json b/2024/Rejection Sampling IMLE_ Designing Priors for Better Few-Shot Image Synthesis/0d5c28ac-bfc0-4636-af16-85b85c6b5148_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..c896c1d8b18e33b83ca5dec3743ac29f54e881ec --- /dev/null +++ b/2024/Rejection Sampling IMLE_ Designing Priors for Better Few-Shot Image Synthesis/0d5c28ac-bfc0-4636-af16-85b85c6b5148_content_list.json @@ -0,0 +1,2065 @@ +[ + { + "type": "text", + "text": "Rejection Sampling IMLE: Designing Priors for Better Few-Shot Image Synthesis", + "text_level": 1, + "bbox": [ + 225, + 140, + 779, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Chirag Vashist, Shichong Peng, and Ke Li", + "bbox": [ + 344, + 212, + 656, + 227 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "APEX Lab", + "bbox": [ + 467, + 239, + 535, + 252 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "School of Computing Science", + "bbox": [ + 411, + 255, + 589, + 267 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Simon Fraser University", + "bbox": [ + 426, + 268, + 573, + 281 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{chirag_vashist, shichong_peng, keli}@sfu.ca", + "bbox": [ + 331, + 282, + 671, + 295 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. An emerging area of research aims to learn deep generative models with limited training data. Implicit Maximum Likelihood Estimation (IMLE), a recent technique, successfully addresses the mode collapse issue of GANs and has been adapted to the few-shot setting, achieving state-of-the-art performance. However, current IMLE-based approaches encounter challenges due to inadequate correspondence between the latent codes selected for training and those drawn during inference. This results in suboptimal test-time performance. We theoretically show a way to address this issue and propose RS-IMLE, a novel approach that changes the prior distribution used for training. This leads to substantially higher quality image generation compared to existing GAN and IMLE-based methods, as validated by comprehensive experiments conducted on nine few-shot image datasets.", + "bbox": [ + 259, + 329, + 743, + 494 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Few-shot $\\cdot$ Image Synthesis $\\cdot$ Implicit Maximum Likelihood Estimation", + "bbox": [ + 259, + 508, + 740, + 535 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 583, + 356, + 598 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recent years have witnessed significant advancements in image synthesis, driven by the emergence of deep learning and the development of a broad variety of powerful generative models. Generative Adversarial Networks (GANs) [2,7,10,11,14], Variational Autoencoders (VAEs) [3,16,29,36], Diffusion Models [4,9], Score-based Models [34,35], Normalizing Flows [5, 15, 17], and Autoregressive Models [6, 27, 28, 30] have demonstrably improved the quality of synthesized images, often achieving photorealism. Quite often, generative models often require large amounts of training data to achieve this high fidelity. In numerous situations, obtaining large volumes of training data is impractical. For instance, when we want to generate images of rare species of plants or images of anatomical structures with rare diseases, access to abundant images may be limited. In this paper, we aim to tackle the problem of high quality image synthesis using limited training data.", + "bbox": [ + 212, + 613, + 787, + 794 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In the few-shot setting, most of the leading methods to date are based on GANs [18, 23, 24, 32, 38]. Despite advances, these methods continue to be afflicted by mode collapse, a well-known issue in GANs. Mode collapse occurs when the generator network", + "bbox": [ + 212, + 795, + 787, + 839 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/be654b3d05c083e881e770900f841c83b621234de8e6a1ca159fc77b274686f6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 218, + 143, + 372, + 247 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/d0f25a73ced6cd7e287d574d3316d1d865e7807ec2dfac233f0e81e50543e202.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 388, + 157, + 658, + 247 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/a27541844bbc12b215300afd4e1935e0c03a4ddd066d3d47fb71e1181d2f6a3a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 663, + 157, + 767, + 248 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/53f47ba949c401a702e6504df83b2bae5e042cd80fdbda7903c9bd8cabe48b47.jpg", + "image_caption": [ + "Fig. 1: IMLE is an implicit generative model that maps a latent code sampled from a prior distribution to an image output. In previous IMLE-based methods, both the training and testing phases adopt a standard normal distribution as the prior distribution. However, this approach often results in poor generalization during inference. To address this limitation, we introduce RS-IMLE, which uses rejection sampling to alter the prior distribution used for training to a different distribution $\\mathcal{P}$ . This modification significantly enhances the quality of generated images during testing." + ], + "image_footnote": [], + "bbox": [ + 220, + 255, + 372, + 362 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/199fdeded655984a06003c96c65e6b5bed6adf14cc147ce66b1eb3f6bef728c7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 388, + 255, + 658, + 362 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/73a107811eae489598567b2572fdd8ea6deb35b3c3c0547008742dfb1963daec.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 663, + 255, + 766, + 363 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "fails to capture the full training data distribution and instead produces a limited subset of outputs. This phenomenon is especially problematic in scenarios where only a small number of training examples are provided.", + "bbox": [ + 212, + 489, + 787, + 536 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Implicit Maximum Likelihood Estimation (IMLE) [21] is an alternative to the GAN objective and has shown promising results in addressing mode collapse. In contrast to GANs, which aim to make each generated image resemble some training data, IMLE instead ensures that each training image has some generated sample close to it, and therefore cannot drop any of the modes present in the training data. Adaptive IMLE [1] further extends IMLE to the few-shot image synthesis setting and achieves state-of-the-art generated image quality and mode coverage.", + "bbox": [ + 212, + 537, + 787, + 643 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "However, in existing IMLE-based approaches, we observe that the latent codes used during training and those sampled during testing have different distributions, even though the same prior is used during training and inference. This phenomenon arises because of how IMLE selects latent codes during training. Some regions of the latent space are consistently rarely picked for training, despite having a high likelihood under the prior distribution (often the standard Gaussian). This is illustrated in Fig. 2a. Consequently, at test time, when latent codes drawn from the prior happen to fall in these regions, they yield low-quality samples that are far from the real data points, as illustrated in Fig. 1. We address this issue by carefully choosing a different prior so that the samples selected for training have a distribution more similar to those sampled at inference. Our method, which we call Rejection Sampling IMLE or RS-IMLE for short, demonstrably improves coverage of the latent space used during training, thereby ensuring better alignment with the prior as shown in Fig. 2b. As a result, our method yields higher quality", + "bbox": [ + 212, + 643, + 789, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "C. Vashist et al.", + "bbox": [ + 271, + 114, + 366, + 127 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/aa9c3ab7230d3eb9cf00e8be247a3420fedf3f9fa17a3cae17c24e65e4d3275b.jpg", + "image_caption": [ + "Fig. 2: Difference between the latent codes picked by IMLE and RS-IMLE over the course of training." + ], + "image_footnote": [ + "(a) Latent space of model trained by IMLE objective using standard normal prior. Dots represent the points selected by the model over the course of training, with dots of the same colour belonging to the same data-point. The contours of standard normal distribution have been shown for comparison." + ], + "bbox": [ + 218, + 143, + 480, + 337 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/12476c053cfd54019576c9267b61c255a1aac96decaeb53cf9d2ca080474c9fa.jpg", + "image_caption": [], + "image_footnote": [ + "(b) Latent space of model trained by RS-IMLE objective using prior obtained via rejection sampling. Compared to latent space of model trained by IMLE, our method over the course of training, samples latent codes that follow the distribution at test time more faithfully." + ], + "bbox": [ + 521, + 143, + 784, + 337 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "samples during testing compared to existing GAN and IMLE-based methods. We substantiate this claim through theoretical analysis and extensive experiments conducted on nine few-shot image datasets. We achieve an average of $45.9\\%$ decrease in FID [8] across datasets compared to the best baseline.", + "bbox": [ + 212, + 474, + 787, + 535 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 561, + 366, + 578 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Training deep generative models with limited data remains a significant challenge. One approach involves adapting a model pretrained on a large-scale auxiliary dataset from similar domains [22,25,26,31,37,40]. However, the availability of such large-scale auxiliary datasets across all domains is not guaranteed. Therefore, another emerging line of work focuses on training models from scratch. In this context, due to the scarcity of training data, diffusion models struggle to achieve high-quality generated images and have been demonstrated to be ineffective [1]. As a result, previous works in this area predominantly build on Generative Adversarial Networks (GANs) and design various methods to address the well-known mode collapse issue. Techniques such as ADA [13] and DiffAug [41] aim to expand training data using adaptive and differentiable augmentation strategies. FastGAN [24] introduced a skip-layer excitation module for accelerated training and used self-supervision in the discriminator to enhance feature learning, thereby improving mode coverage of the generator. FakeCLR [23] enhances image synthesis by extensive data augmentation and applies contrastive learning solely on perturbed fake samples. FreGAN [38] introduces a frequency-aware model with a self-supervised constraint to avoid generating arbitrary frequency signals. ReGAN [32] dynamically adjusts", + "bbox": [ + 212, + 598, + 787, + 843 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "Rejection Sampling IMLE", + "bbox": [ + 571, + 114, + 730, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 114, + 784, + 125 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "GANs' architecture during training to explore diverse sub-network structures at different training times. However, despite these advancements, some degree of mode collapse persists.", + "bbox": [ + 212, + 146, + 782, + 190 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In contrast, Implicit Maximum Likelihood Estimation (IMLE) [21] shows promising results in addressing mode collapse through the use of an alternative objective function compared to GANs. Building upon IMLE, Adaptive IMLE [1] adapts this approach to the few-shot image synthesis scenario by introducing individual target thresholds for each training data point. This dynamic adjustment of training progress accounts for varying difficulties across different data points, thereby effectively leveraging the limited training data. In this work, we introduce a novel algorithm, orthogonal to Adaptive IMLE, for sample selection during training.", + "bbox": [ + 212, + 191, + 787, + 313 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3 Method", + "text_level": 1, + "bbox": [ + 214, + 337, + 316, + 353 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1 Background", + "text_level": 1, + "bbox": [ + 214, + 371, + 343, + 386 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In the context of unconditional image synthesis, the primary objective is to learn the unconditional probability distribution of images $p(\\mathbf{x})$ . This distribution enables the generation of novel synthesized images through the sampling. Implicit generative models are commonly employed for this task, with the generator in GANs serving as a notable example. Generator in GANs are represented by a function $T_{\\theta}: Z \\to X$ , implemented as a neural network with parameters denoted as $\\theta$ . The function $T_{\\theta}$ learns a transformation from the latent space $Z$ to the image space $X$ by using adversarial training, which employs a discriminator that tries to distinguish between generated images $T_{\\theta}(\\mathbf{z})$ and real images $\\mathbf{x}$ , while the generator tries to produce increasingly realistic images to deceive the discriminator. However, this objective often leads to mode collapse, a well-known issue of GANs, where the generated output $T_{\\theta}(\\mathbf{z})$ only models a subset of the training examples.", + "bbox": [ + 212, + 397, + 787, + 578 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To address the issue of mode collapse, an alternative method Implicit Maximum Likelihood Estimation (IMLE) [21] has been introduced. While IMLE, like GANs, uses a generator, it differs from GANs by using an alternative objective that aims to make each generated sample similar to some training data. Instead, IMLE ensures that each training data point has similar generated samples, thereby encouraging coverage of all the modes of the training data.", + "bbox": [ + 212, + 580, + 787, + 669 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The IMLE objective is given as follows, where $d(\\cdot ,\\cdot)$ is a distance metric:", + "bbox": [ + 238, + 670, + 723, + 685 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\theta_ {\\mathrm {I M L E}} = \\arg \\min _ {\\theta} \\mathbb {E} _ {z _ {1}, \\dots , z _ {m} \\sim \\mathcal {N} (0, I)} \\left[ \\sum_ {i = 1} ^ {n} \\min _ {j \\in [ m ]} d \\left(\\mathbf {x} _ {i}, T _ {\\theta} (\\mathbf {z} _ {j})\\right) \\right] \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 294, + 696, + 785, + 739 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Here $m$ denotes the number of samples and $n$ denotes the number of data points. In simple terms, the IMLE objective first draws $m$ samples $\\mathbf{z}_j$ from the standard Gaussian distribution and transforms them into the image space using the function $T_{\\theta}$ . From these pool of samples in the image space, for each data point $x_i$ , IMLE selects a sample that is closest to the data point in some distance metric $d(\\cdot, \\cdot)$ . This operation can be done efficiently due to advances in high-dimensional nearest neighbour search [20]. Note that", + "bbox": [ + 212, + 750, + 787, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "C. Vashist et al.", + "bbox": [ + 271, + 114, + 364, + 127 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "the number of samples $m$ must at least be equal to the number of data points $n$ , since otherwise by the pigeonhole principle, some samples would be picked by multiple data points. In practice, we find that setting $m$ to be a multiplicative factor (like 10 or 20) times larger than $n$ works the best.", + "bbox": [ + 214, + 146, + 787, + 205 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2 Observation", + "text_level": 1, + "bbox": [ + 215, + 228, + 344, + 242 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In the existing IMLE-based methods, we observe that the distributions of the latent codes used for training the objective differs from the distribution of latent encountered at test time. Consider an illustrative example where the latent space is two dimensional. We train a simple generative model using IMLE on two dimensional toy dataset. The latent codes used for training over the course of training are illustrated in Figure 2a. We notice that for the latent codes belonging to the same data point (denoted by the same colour) form well-separated tight bands in the latent space. We also observe that there are large gaps between these bands, indicating that these segments of the latent space are consistently overlooked during training. Since at test time we sample from the same standard normal distribution, these unsupervised segments in the latent space have arbitrary outputs, which result in bad samples. We term this phenomenon the \"Misalignment Issue.\"", + "bbox": [ + 214, + 253, + 787, + 434 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3 Analysis of the Misalignment Issue", + "text_level": 1, + "bbox": [ + 215, + 455, + 500, + 470 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we will try to see why the phenomenon mentioned above occurs. Let us clarify the notation used in subsequent sections: $\\mathbf{x}_i$ denotes the $i^{\\mathrm{th}}$ data point, $d(\\cdot ,\\cdot)$ denotes the distance function and $T_{\\theta}(\\mathbf{z}_j)$ denote the $j^{\\mathrm{th}}$ sample where $\\mathbf{z}_j\\sim \\mathcal{N}(0,I)$ .", + "bbox": [ + 212, + 479, + 785, + 527 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Let us define a random variable, $D_{ij}$ to denote the distance of $i^{\\text{th}}$ data point to the $j^{\\text{th}}$ sample. We can also define another random variable $D_i^*$ to denote the distance of $i^{\\text{th}}$ data point to the sample closest to it. Further, let $F_{D_{ij}}$ and $F_{D_i^*}$ be the CDF of $D_{ij}$ and $D_i^*$ respectively. Let $f_{D_{ij}}$ and $f_{D_i^*}$ be the PDF of $D_{ij}$ and $D_i^*$ respectively. Now we will relate the CDF of the distances between a data point and its selected latent code, $F_{D_i^*}$ to the CDF of the distances between the same data point and a random latent code $F_{D_{ij}}$ :", + "bbox": [ + 214, + 527, + 787, + 630 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} F _ {D _ {i} ^ {*}} (t) = \\operatorname * {P r} (D _ {i} ^ {*} \\leq t) = 1 - \\operatorname * {P r} (D _ {i} ^ {*} > t) \\\\ = 1 - \\Pr (D _ {i j} > t, \\forall j \\in [ m ]) \\quad (\\text {D e f . o f} D _ {i j}) \\\\ = 1 - \\prod_ {j = 1} ^ {m} \\Pr (D _ {i j} > t) \\\\ = 1 - \\left(\\Pr \\left(D _ {i 1} > t\\right)\\right) ^ {m} (2) \\\\ = 1 - \\left(1 - F _ {D _ {i 1}} (t)\\right) ^ {m} (3) \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 292, + 652, + 785, + 784 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Note that Equation 2 is true because each $\\mathbf{z}_j$ is drawn independently from the same probability distribution which makes $D_{i1}, D_{i2} \\dots D_{im}$ identical in distribution for a particular data point $\\mathbf{x}_i$ .", + "bbox": [ + 214, + 794, + 785, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "Rejection Sampling IMLE", + "bbox": [ + 571, + 114, + 730, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 114, + 785, + 126 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/94f9f3c5429cc4a73e11fe915de277d47f4a283be323696eb592db9526c3a622.jpg", + "image_caption": [ + "(a) PDF of example distribution" + ], + "image_footnote": [], + "bbox": [ + 217, + 145, + 401, + 255 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/c196852dd0bd7dd0f1cda40ab503bbfaa75ffc7454980970d768886e4d737e68.jpg", + "image_caption": [ + "Fig. 3: Illustrative figure for demonstrating the behaviour of $F_{D_i^*}(t)$ and $F_{D_{i1}}(t)$ using Noncentral Chi-squared distribution as the example distribution." + ], + "image_footnote": [], + "bbox": [ + 411, + 145, + 591, + 255 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/434e8a8b691bba4e02e829fbdd5ee16b3b257ce536af00347c43667e993208f9.jpg", + "image_caption": [ + "(b) $F_{D_i^*}(t)$ vs $t$ for different values of $m$", + "(c) $F_{D_i^*}(t)$ vs $F_{D_{i1}}(t)$ for different values of $m$" + ], + "image_footnote": [], + "bbox": [ + 602, + 145, + 784, + 255 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Now we can try to justify our observations from Figure 2a. Equation 3 shows us how the CDF of the distance of the selected latent code (used in training) differs significantly from the CDF of distance of a random latent code (encountered at test time). This shows that the distance of the sample we choose for training is lower than the distance of the sample at testing. We can obtain a deeper understanding by analyzing the plots in Figure 3 for an example distribution. Notice that $\\forall m > 1, F_{D_i^*}(t) > F_{D_{i1}}(t)$ . Observing Equation 3 and the aforementioned plots, we can conclude the following things:", + "bbox": [ + 212, + 354, + 787, + 464 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. The $f_{D_i^*}$ is skewed towards the origin compared to $f_{D_{ij}}$ . This is intuitive because the latent codes are selected by the min operation and so their distance to their respective data point would be less than that of a random sample.", + "2. The skew towards the data point increases as $m$ increases. This piece of information will be important shortly." + ], + "bbox": [ + 222, + 474, + 785, + 555 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We can also compute the PDF $f_{D_i^*}$ in terms of $f_{D_{i1}}$ by differentiating the CDF $F_{D_i^*}(t)$ as follows:", + "bbox": [ + 212, + 566, + 787, + 597 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nf _ {D _ {i} ^ {*}} (t) = \\frac {\\mathrm {d} F _ {D _ {i} ^ {*}} (t)}{\\mathrm {d} t} = m \\left(1 - F _ {D _ {i 1}} (t)\\right) ^ {m - 1} f _ {D _ {i 1}} (t) \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 339, + 622, + 785, + 656 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.4 Solving the Misalignment Issue", + "text_level": 1, + "bbox": [ + 214, + 676, + 473, + 694 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Now that we know the reason behind the misalignment between latent codes used at training and testing, we wish to come up with a method to mitigate this phenomenon. Recall that $D_{ij}$ and $D_i^*$ are determined by the distance function $d(\\cdot, \\cdot)$ , neural network $T_\\theta$ and the prior distribution. For a given generative modelling task, it is not trivial to change the $d(\\cdot, \\cdot)$ or $T_\\theta$ . Hence, in this paper, we aim to change the prior distribution used at training such that the distribution of latent codes at training time closely matches with the distribution of latent codes (drawn from the standard normal distribution) encountered at test time. Notice that the IMLE objective (Equation 1) allows us to sample from the prior without knowing the closed form expression for its probability density", + "bbox": [ + 212, + 704, + 787, + 843 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "C. Vashist et al.", + "bbox": [ + 271, + 114, + 366, + 127 + ], + "page_idx": 5 + }, + { + "type": "code", + "sub_type": "algorithm", + "code_caption": [ + "Algorithm 1 RS-IMLE Procedure" + ], + "code_body": "Require: The set of inputs $\\{\\mathbf{x}_i\\}_{i = 1}^n$ , radius $\\epsilon$ \n1: Initialize the parameters $\\theta$ of the generator $T_{\\theta}$ \n2: for $k = 1$ to $K$ do \n3: Draw latent codes $Z\\gets \\mathbf{z}_1,\\dots,\\mathbf{z}_m$ from $\\mathcal{N}(0,\\mathbf{I})$ \n4: Compute $\\tilde{Z}\\gets \\tilde{\\mathbf{z}}_1,\\ldots ,\\tilde{\\mathbf{z}}_p$ from $Z$ such that $d\\left(\\mathbf{x}_i,T_\\theta (\\tilde{\\mathbf{z}}_j)\\right)\\geq \\epsilon ,\\quad \\forall \\tilde{\\mathbf{z}}_j\\in \\tilde{Z},i\\in [n]$ \n5: $\\sigma (i)\\gets \\arg \\min_{j\\in [m]}d(\\mathbf{x}_i,T_\\theta (\\tilde{\\mathbf{z}}_j)),\\quad \\forall i\\in [n]$ \n6: for $l = 1$ to $L$ do \n7: Pick a random batch $S\\subseteq [n]$ \n8: $\\theta \\gets \\theta -\\eta \\nabla_{\\theta}\\left(\\sum_{i\\in S}d\\left(\\mathbf{x}_i,T_\\theta \\left(\\tilde{\\mathbf{z}}_{\\sigma (i)}\\right)\\right)\\right) / |S|$ \n9: end for \n10: end for \n11: return $\\theta$", + "bbox": [ + 215, + 164, + 542, + 349 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "function. This allows us the flexibility of choosing a non-analytical prior distribution. To distinguish from $\\mathbf{z}_j\\sim \\mathcal{N}(0,I)$ , we will use $\\tilde{\\mathbf{z}}$ to denote the latent codes drawn from our desired target distribution $\\mathcal{P}$ . Identical to the previous section, let us define random variable $\\tilde{D}_{ij} = d\\left(\\mathbf{x}_i,T_\\theta (\\tilde{\\mathbf{z}}_j)\\right)$ to denote the distance between data point $\\mathbf{x}_i$ from a random sample $T_{\\theta}(\\tilde{\\mathbf{z}}_j)$ . Similarly, we define $\\tilde{D}_i^* = \\min_{j\\in [m]}\\tilde{D}_{ij}$ . Similar to the section above, let $F_{\\tilde{D}_{ij}}$ and $F_{\\tilde{D}_i^*}$ be the CDF of $\\tilde{D}_{ij}$ and $\\tilde{D}_i^*$ respectively. Then $f_{\\tilde{D}_{ij}}$ and $f_{\\tilde{D}_i^*}$ would be the PDF of $\\tilde{D}_{ij}$ and $\\tilde{D}_i^*$ respectively.", + "bbox": [ + 212, + 391, + 787, + 506 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Designing the target prior Now, we will try to think about the properties we would want to have for our target prior. Recall that the Misalignment Issue is mitigated as we decrease the number of samples. In order to differentiate between the number of samples of different priors, we use the notation $m'$ to denote the number of samples for the objective using the Gaussian prior. As hinted in the previous section, one way to avoid the Misalignment Issue is to set $m'$ to low values. However we cannot directly use a Gaussian prior with low values of $m'$ as our target prior. This is because in aforementioned case, it happens to be that many data points are forced to pick the same sample as their nearest neighbour. This in turn causes very slow or no learning. Hence when using a Gaussian prior, we need to pick $m'$ large enough to allow convergence and yet small enough such that the Misalignment Issue does not affect test time performance. In our case, we are trying to design a new distribution that solves the Misalignment Issue by having desirable properties of an ideal distribution. The ideal distribution is a Gaussian prior with $m'$ set to the lowest possible value, which is $m' = n$ .", + "bbox": [ + 212, + 553, + 787, + 765 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Now that we have found the ideal prior, we can make $\\mathcal{P}$ and the ideal prior to be equal in distribution. Similar to the analysis till Equation 4, we derive the PDF of $\\mathcal{P}$ . We get $f_{\\tilde{D}_i^*}(t) = m\\left(1 - F_{\\tilde{D}_{i1}}(t)\\right)^{m - 1}f_{\\tilde{D}_{i1}}(t)$ . Equating this PDF of $\\mathcal{P}$ to the PDF of the ideal distribution gives:", + "bbox": [ + 212, + 768, + 787, + 840 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "Rejection Sampling IMLE", + "bbox": [ + 571, + 114, + 730, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 114, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} m \\left(1 - F _ {\\tilde {D} _ {i 1}} (t)\\right) ^ {m - 1} f _ {\\tilde {D} _ {i 1}} (t) = n \\left(1 - F _ {D _ {i 1}} (t)\\right) ^ {n - 1} f _ {D _ {i 1}} (t) \\\\ \\Rightarrow f _ {\\tilde {D} _ {i 1}} (t) = \\frac {n}{m} \\frac {\\left(1 - F _ {D _ {i 1}} (t)\\right) ^ {n - 1}}{\\left(1 - F _ {\\tilde {D} _ {i 1}} (t)\\right) ^ {m - 1}} f _ {D _ {i 1}} (t) (5) \\\\ \\Longrightarrow f _ {\\tilde {D} _ {i 1}} (t) = c \\phi (t) f _ {D _ {i 1}} (t) (6) \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 316, + 170, + 785, + 282 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Here, $\\phi(t)$ is introduced to simplify notation and $c$ is a constant associated with truncating $F_{D_{i1}}(t)$ that is needed to make sure $f_{\\tilde{D}_{i1}}(t)$ is bounded. We discuss these technical details in the Appendix.", + "bbox": [ + 212, + 292, + 785, + 338 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Rejection Sampling We have expressed our target prior $\\mathcal{P}$ in terms of distribution from which we can easily sample. Now, we can use Rejection Sampling to sample from our target prior $\\mathcal{P}$ . To be concrete: $f_{\\tilde{D}_{i1}^*}(t)$ is our target distribution and $f_{D_{i1}}(t)$ acts as our proposal distribution, since we can sample from the standard Gaussian easily. We can further write it in the standard Rejection Sampling notation: $\\frac{f_{\\tilde{D}_{i1}^*}(t)}{M f_{D^*}(t)} = \\frac{c\\phi(t)}{M}$", + "bbox": [ + 212, + 359, + 782, + 450 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Here, $M$ is the scaling factor associated with Rejection Sampling. We approximate the function above using a step function. The step needs to happen at $t$ where $F_{\\tilde{D}_{i1}}(t)$ gets close to 1. Instead of trying to estimate $F_{\\tilde{D}_{i1}}(t)$ , we instead use a hyperparameter $\\epsilon$ to represent where $F_{\\tilde{D}_{i1}}(t)$ gets close to 1. We find the value of this hyperparameter $\\epsilon$ by cross-validation.", + "bbox": [ + 212, + 453, + 784, + 527 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The final procedure simplifies to this: we sample $\\mathbf{z} \\sim \\mathcal{N}(0, I)$ . If $t = d\\left(\\mathbf{x}_i, T_\\theta(\\mathbf{z})\\right) < \\epsilon$ , we reject the sample; otherwise, we accept it. Since the sampling procedure is based on Rejection Sampling, we call our method RS-IMLE. The resulting RS-IMLE procedure is included in Algorithm-1.", + "bbox": [ + 212, + 527, + 784, + 589 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "3.5 Intuitive Interpretation of the Algorithm Behaviour", + "text_level": 1, + "bbox": [ + 214, + 611, + 614, + 627 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Prior to proceeding to the Implementation details, gaining a gradient-based understanding of our new objective would provide valuable insight. Let us revisit the IMLE objective (Equation 1) again. As training progresses, we expect the loss to reduce such that: $\\forall x_{i},\\quad \\mathbb{E}_{z_{1},\\ldots ,z_{m}\\sim \\mathcal{N}(0,I)}\\min_{j\\in [m]}d\\left(\\mathbf{x}_{i},T_{\\theta}(\\mathbf{z}_{j})\\right)\\to 0.$ As the objective approaches convergence, the loss will decrease, resulting in gradients with lower magnitude. This causes smaller updates to the model parameters during training, leading to slower progress. Note that this loss is with respect to the closest sample to each data point. Thus it can be the case that even after a lot of training, although the closest sample are pretty close to their respective data point, the rest of samples are pretty far away.", + "bbox": [ + 212, + 636, + 782, + 773 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\theta_ {\\mathrm {R S - I M L E}} = \\arg \\min _ {\\theta} \\mathbb {E} _ {z _ {1}, \\dots , z _ {m} \\sim \\mathcal {P}} \\left[ \\sum_ {i = 1} ^ {n} \\min _ {j \\in [ m ]} d \\left(\\mathbf {x} _ {i}, T _ {\\theta} (\\mathbf {z} _ {j})\\right) \\right] \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 302, + 796, + 785, + 838 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 227, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "C. Vashist et al.", + "bbox": [ + 271, + 114, + 364, + 127 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/903da0e53e211d41f5c74ca952d7d4c34757878cb0b101ecf10b99f6e0909ed6.jpg", + "image_caption": [ + "(a) IMLE after epoch 100" + ], + "image_footnote": [], + "bbox": [ + 236, + 154, + 480, + 239 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/3987190656c0d473856a93f2aff5f30c29a1f264edc827aa34b32ad5a33cd6bf.jpg", + "image_caption": [ + "(b) RS-IMLE after epoch 100" + ], + "image_footnote": [], + "bbox": [ + 513, + 148, + 769, + 244 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/1d3d46be4d8958affaab93487277a459c26224141e1668a1b0b88c912695bb44.jpg", + "image_caption": [ + "(c) IMLE after epoch 500" + ], + "image_footnote": [], + "bbox": [ + 236, + 280, + 483, + 367 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/849efd1263ad76742fa1b334afcd086d74b89d0c3e2b0ba6312e80dfb278e56e.jpg", + "image_caption": [ + "(d) RS-IMLE after epoch 500" + ], + "image_footnote": [], + "bbox": [ + 514, + 275, + 779, + 369 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/02f525b8395991748243721048bbd43e663e42adc6de52da2bd099bdc5092f2a.jpg", + "image_caption": [ + "(e) IMLE after 2000 epochs" + ], + "image_footnote": [], + "bbox": [ + 236, + 406, + 478, + 489 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/756c96d4b94fd9f56b56094cf52a6c5053a063c60de10cab4e844dc7080cdea8.jpg", + "image_caption": [ + "(f) RS-IMLE after 2000 epochs", + "Fig. 4: Comparison between IMLE and RS-IMLE for 2D toy problem. Data points are denoted by $\\blacksquare$ and samples are denoted by $\\bullet$ . Samples picked as nearest neighbours are denoted by $\\star$ ." + ], + "image_footnote": [], + "bbox": [ + 514, + 401, + 769, + 498 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Consider our objective in Equation 7 and recall that we have constructed the probability distribution $\\mathcal{P}$ such that $\\forall \\mathbf{x}_i\\quad d\\left(\\mathbf{x}_i,T_\\theta (\\tilde{\\mathbf{z}})\\right)\\geq \\epsilon$ , where $\\tilde{\\mathbf{z}}\\sim \\mathcal{P}$ . In other words, all samples we obtain by using the prior $\\mathcal{P}$ are guaranteed to be $\\epsilon$ -distance away from all data points. This ensures that the loss per data point is always greater than $\\epsilon$ . The approach can be interpreted as ignoring the samples that are already close to some data point and instead training on challenging, non-trivial samples.", + "bbox": [ + 212, + 602, + 787, + 694 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "To compare the sampling behavior of the vanilla IMLE and the proposed RS-IMLE, we trained two models on a 2D toy problem as illustrated in Figure 4. The first model uses the vanilla IMLE objective (Equation 1), while the second model is trained with our proposed RS-IMLE objective (Equation 7).", + "bbox": [ + 212, + 698, + 787, + 760 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "At the initial stage of training, both the methods learn similar distributions, indicated by the straight line of orange dots ( $\\odot$ ) in Figure-4a,4b. Our method first removes all the samples that fall within an $\\epsilon$ distance from any data point before doing the nearest neighbour search. We illustrate this in Figure-4b, where samples that lie within the gray circles are not considered for the nearest neighbour search.", + "bbox": [ + 212, + 763, + 787, + 840 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "Rejection Sampling IMLE", + "bbox": [ + 571, + 114, + 730, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 114, + 785, + 126 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/1a22c15fe607f29d786b2669930a04268a35754286cd8b044c23f0129575445f.jpg", + "image_caption": [ + "Fig. 5: Qualitative comparison between our method and baselines. While analyzing the images, look for the sharpness of each image and diversity in the content of all images for a method." + ], + "image_footnote": [], + "bbox": [ + 217, + 143, + 785, + 296 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "As training progresses, we notice that for both the algorithms' samples move closer to the ground truth data points. However, for vanilla IMLE, we observe that for many data points the sample picked after the nearest neighbour search is already close to the ground truth. In this case, the loss associated with these data points would be low (indicated by the short length of arrow in the Figure-4c). As a result, the model trained by the vanilla IMLE objective does not learn anything significantly novel. In our proposed method (Figure-4d), each data point selects a sample that is at least $\\epsilon$ distance away from it. This ensures that the loss for each data point is always sufficiently high (indicated by the long arrows), resulting in meaningful updates to the model parameters.", + "bbox": [ + 212, + 364, + 787, + 501 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "3.6 Implementation Details", + "text_level": 1, + "bbox": [ + 215, + 521, + 419, + 536 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Note that computing the distance of each sample with each data point is computationally expensive. Suppose we have $n$ data points in $\\mathbb{R}^d$ and $m$ samples, calculating the distance between each pair has a time complexity of $\\mathcal{O}(mnd)$ . To get around this issue, we leverage a fast k-nearest neighbor search method, DCI [20]. This method reduces the runtime of a single query from linear to sublinear in $m$ , enabling us to efficiently filter out all the samples that are within an $\\epsilon$ distance from any data point. Subsequently, from this filtered pool of remaining samples, we select, for each data point, the sample that is closest to it.", + "bbox": [ + 212, + 545, + 787, + 665 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "In order to reduce the search time complexity further, we can project the training data to a lower dimensional subspace (which would decrease $d$ ). We use this while implementing the procedure for image synthesis task: we first flatten each image and project it to a lower dimension by using random projection. We normalize these projected vectors before using them for nearest neighbour search.", + "bbox": [ + 212, + 666, + 787, + 742 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 215, + 763, + 357, + 780 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Datasets We assess our method and the baseline approaches across a variety of datasets with $256 \\times 256$ resolution. These datasets include Animal-Face Dog [33], Animal-Face Cat [33], Obama [41], Panda [41], Grumpy-cat [41], Anime [24], Shells [24], Skulls [24]", + "bbox": [ + 212, + 794, + 785, + 840 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "C. Vashist et al.", + "bbox": [ + 271, + 114, + 364, + 126 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "and a subset of Flickr-FaceHQ (FFHQ) [12] which are standard datasets used in the few-shot learning literature.", + "bbox": [ + 215, + 146, + 784, + 176 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/b370376dac76132a7f35209bd835bac28d3057fc2de7e6c74bd9ceeb78db6275.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
DatasetFastGAN [24]FakeCLR [23]FreGAN [38]ReGAN [32]AdaIMLE [1]RS-IMLE (Ours)Imp. %
Obama41.129.933.445.725.014.044.0
Grumpy Cat26.620.624.927.319.111.539.8
Panda10.08.89.012.67.63.554.0
FFHQ-10054.262.150.587.433.212.961.1
Cat35.127.431.042.124.915.936.1
Dog50.744.447.957.243.023.146.3
Anime69.877.759.8110.865.835.845.6
Skulls109.6106.5163.3130.781.951.137.6
Shells120.9148.4169.3236.1108.555.448.9
", + "bbox": [ + 215, + 204, + 785, + 371 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Table 1: We compute FID [8] between the real data and 5000 randomly generated samples for all the methods. Lower is better.", + "bbox": [ + 215, + 372, + 785, + 398 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Baselines We compare our method to recent state-of-the-art few-shot image generation methods. These include FastGAN [24], FakeCLR [23], FreGAN [38], Re-GAN [32] and AdaIMLE [1].", + "bbox": [ + 215, + 419, + 784, + 463 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Evaluation Metrics We employ Fréchet Inception Distance (FID) [8] to assess the perceptual quality of the generated images. This involves randomly generating 5000 images and calculating the FID between these generated samples and the real images for each dataset. Additionally, we evaluate the modelling accuracy and coverage by computing precision and recall for 1000 images using the metric defined by Kynkänniemi et al. [19]. In image synthesis, precision refers to the model's capacity to generate images closely resembling the desired target or distribution. Recall measures the model's ability to encompass a wide array of diverse images within the target distribution. To ensure that the computed metrics have low variance, we generate many more samples than training images.", + "bbox": [ + 215, + 465, + 785, + 616 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Network Architecture We construct our generator network using decoder modules from VDVAE [3]. More details about the network architecture can be found in the Appendix.", + "bbox": [ + 215, + 617, + 784, + 662 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4.1 Quantitative Results", + "text_level": 1, + "bbox": [ + 215, + 688, + 398, + 704 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "In Table 1, we present the FID scores computed for all the datasets across different methods. Lower FID scores indicate that the distribution of generated images is closer to the distribution of real images. Our method performs significantly better compared to baselines.", + "bbox": [ + 215, + 717, + 784, + 777 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "In Table 2, we show the precision and recall scores. Our method has a near perfect precision (close to 1), while having a significantly higher recall compared to the baselines. In the few cases where our method is not the best, it is very close to the best metric.", + "bbox": [ + 215, + 779, + 784, + 838 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "Rejection Sampling IMLE", + "bbox": [ + 571, + 114, + 730, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 784, + 126 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/c5051ad720e597f52683eaddb7faa491c58a7db82e16ce5edebd7f67e0294f10.jpg", + "image_caption": [ + "Query", + "(a) Obama" + ], + "image_footnote": [], + "bbox": [ + 217, + 167, + 326, + 252 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/41f62913db116d5de28f06fdb521b7e5d6aff67f51d094d0b0427feb2a8fd2b1.jpg", + "image_caption": [ + "RS-IMLE" + ], + "image_footnote": [], + "bbox": [ + 328, + 167, + 439, + 252 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/5fb2b9c811adba02deb0be6ef649326622a2004b0006d35c89b87d9e8d09e4ba.jpg", + "image_caption": [ + "Ada-IMLE" + ], + "image_footnote": [], + "bbox": [ + 439, + 167, + 550, + 252 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/683bb6827b5df7121a87d723d805f452891b4d7b100c458023ee00921b181953.jpg", + "image_caption": [ + "FastGAN" + ], + "image_footnote": [], + "bbox": [ + 550, + 167, + 661, + 252 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/6b200923da9adf0ac846d23121ceb7f4296563c504eea63745369d123eaa2d46.jpg", + "image_caption": [ + "FreGAN" + ], + "image_footnote": [], + "bbox": [ + 661, + 167, + 772, + 252 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/6af369e0484a58c880c9d69c7e3729b83aa1de9fc25a6fccf1740210496bf656.jpg", + "image_caption": [ + "(b) Anime" + ], + "image_footnote": [], + "bbox": [ + 217, + 265, + 326, + 349 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/077d465dc3a703e95c0e7bd89ffaf91240843e4689860d57b7d1ca150e021853.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 328, + 266, + 437, + 349 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/ca64a04b7ca8cebcfd570d6e37e378c02c3266257779dcc8600c5e3f53f2484c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 439, + 266, + 550, + 349 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/128aa9b30dbf8a0e78a685ceaded7af59c59fc5af28ea3060b6b7911d2c05970.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 550, + 266, + 661, + 349 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/cee15c41250a9d0627202729a2fb8adebf27f42166db826bd84e8bbdd43176b4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 661, + 266, + 772, + 349 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/4d777c9d80dfb01159fcfca59e1925d711b8d11c3e889e6402a052d885e49e91.jpg", + "image_caption": [ + "(c) Dog" + ], + "image_footnote": [], + "bbox": [ + 217, + 364, + 326, + 449 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/51d2bbc677929c4ea815a88ed34099ae39df7186f2078758a21f8e99b17686bb.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 328, + 364, + 437, + 449 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/77b861d33a7080f0351f389f5c61d14af04471244e424a4c77e23cc77b89e8d0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 439, + 364, + 550, + 449 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/5beb677c92ec43c1eb5fe27dc68e09bda5977dea02cd6df50bb14e4636d1e332.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 550, + 364, + 660, + 449 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/17b6a23af1354eeeab9e9a8dbfa8ffa4c9323f61908f58ace4d9bb77628e01dc.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 661, + 364, + 772, + 449 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/966f34838f2101c1ad03df9322d86655c2c560d801012b2d2fd123aea4a05310.jpg", + "image_caption": [ + "(d) Cat" + ], + "image_footnote": [], + "bbox": [ + 217, + 462, + 326, + 546 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/fca2f453383f06982f87b9a6f1d4b0361885f3bb882b5c861b87cd3d0089b924.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 328, + 462, + 437, + 546 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/c51433fbd18b27ef6d0e8f1021f51805d2b6afb8b5bf2d18605072d5977ff4c9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 439, + 462, + 550, + 546 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/4ba24b36d82ca8c8b9067f52f5232714da42e42ce9d216b613498555f7dea1eb.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 550, + 462, + 660, + 546 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/973b997d5d0ca7a6cea7acc841c4b7a6695cb850eaa68a237e3a86f42e7fc427.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 661, + 462, + 772, + 546 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/12ddc4b196240dec46d43d2dcaebc1773ab93c166f89ecf9dfb3b728ef763069.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 217, + 560, + 326, + 643 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/a5630c9161fd5940b774d161a462fb257558ad4796aa5c744a7c660a68c3af64.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 328, + 560, + 437, + 643 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/647485f4f25d709f43e2586ab5b283eb83ec53a1d320f78d15f54b12101eddfb.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 439, + 560, + 550, + 643 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/58423e93f87222b529343d8c398d32b5d61060e924a78d12440fa5686326d004.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 550, + 560, + 660, + 643 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/026cd53b87b41f419b0b7eec78e20307404bb716cbb298d7c83773023010ab16.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 661, + 560, + 772, + 643 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/379eb13e8b7523712cf23c77e53fd5c61e1748faa79b44bb5f501b8830d4b155.jpg", + "image_caption": [ + "(e) Shells", + "(f) FFHQ-100", + "Fig. 6: Visual Recall test. The first column is the query image from the dataset. Subsequent columns are the samples produced by different methods that are closest to the query image in LPIPS feature space. The samples produced by our method are closer to the query images compared to the baselines, while remaining diverse." + ], + "image_footnote": [], + "bbox": [ + 217, + 659, + 326, + 742 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/2ec14aef5e625847b1ffcae95889948bc3adc854c24d65751048e46ef14dc396.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 328, + 657, + 437, + 742 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/8f31b4dfb65b5566bfa3c2ad96aab41c7e3e1f41019babee95e343f0bb468c92.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 439, + 657, + 550, + 742 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/57f82d76e2db2bf8b179a80f8ffc0e22f0925808bb98fc31774d70b3fb7e26d6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 550, + 657, + 660, + 742 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/68801163711ce49b64fc92ee1cb7188a38414a5ee0da640d684f6b22e2a8c303.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 661, + 657, + 772, + 742 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "C. Vashist et al.", + "bbox": [ + 271, + 114, + 364, + 126 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/6c42fc3dddcf075a3dee1cc6c0a4a3912d9ad4b6a67c9b141cb92c45757e59c8.jpg", + "table_caption": [], + "table_footnote": [ + "Table 2: We compute precision and recall [42] between the real data and 1000 randomly generated samples for all the methods. Higher values are better." + ], + "table_body": "
DatasetFastGAN [24]FakeCLR [23]FreGAN [38]ReGAN [32]AdaIMLE [1]RS-IMLE (Ours)
ObamaPrec.0.920.960.820.620.990.98
Rec.0.090.300.330.010.680.82
Grumpy CatPrec.0.910.970.900.780.970.93
Rec.0.130.390.230.040.720.95
PandaPrec.0.960.970.920.930.980.99
Rec.0.160.410.170.010.630.97
FFHQ-100Prec.0.910.710.860.390.991.00
Rec.0.130.250.210.010.770.99
CatPrec.0.970.990.950.900.980.96
Rec.0.080.550.310.150.860.98
DogPrec.0.960.950.920.840.970.98
Rec.0.190.340.200.100.610.94
AnimePrec.0.860.880.880.290.920.95
Rec.0.080.090.090.010.590.91
SkullsPrec.0.780.660.660.660.950.99
Rec.0.030.010.010.010.320.65
ShellsPrec.0.920.870.870.500.970.98
Rec.0.030.010.010.010.620.59
", + "bbox": [ + 215, + 143, + 785, + 439 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "4.2 Qualitative Results", + "text_level": 1, + "bbox": [ + 217, + 486, + 387, + 502 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Figure 5 compares the random samples of our method to that of several baselines, and we observe that our method produces overall sharper and more diverse images. In addition, we propose Visual Recall, a simple test to substantiate the qualitative superiority of our method. For each method, we first generate 1000 samples. Next, using a real image from the dataset as a query, we find the images from the pool of generated samples that are closest to the query image. We use LPIPS features [39] for the computing the distance between real images and samples. Figure 6 shows the results for the proposed test for different datasets and methods. We see that the samples produced by our method are visually similar to the query, while being sharp and diverse in attributes like hair colour, smile and jaw structure. Note that other methods do not have samples that closely resemble the query image.", + "bbox": [ + 215, + 513, + 785, + 680 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Figure 7 shows results of spherical linear interpolation between two random points in the latent space for different datasets. The images transition in a meaningful manner, indicating that our model has learnt a continuous and structured latent space representation of the image distribution.", + "bbox": [ + 215, + 681, + 785, + 742 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "4.3 Ablation Study", + "text_level": 1, + "bbox": [ + 217, + 766, + 359, + 782 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Table-3 presents the FID computed for different values of $\\epsilon$ for three datasets. We observe that our approach works best for values of $\\epsilon$ close to 0.15 and that increasing the value of $\\epsilon$ beyond a certain range degrades the performance.", + "bbox": [ + 215, + 794, + 785, + 839 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "Rejection Sampling IMLE", + "bbox": [ + 571, + 114, + 730, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 769, + 114, + 784, + 126 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/17b1ac2f0023e48788e1f9db75b7756b33d9db8e7dbba404405efade397f661b.jpg", + "image_caption": [ + "(c) Skulls", + "Fig. 7: Latent space interpolation. We observe that the output images changes smoothly in a meaningful manner." + ], + "image_footnote": [], + "bbox": [ + 215, + 143, + 785, + 446 + ], + "page_idx": 13 + }, + { + "type": "table", + "img_path": "images/08164c20d3d90fba49a992ac2f293eafa8d1538d206c8679b8910dc6d01d8c10.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Dataset0.120.150.180.22
FFHQ-10013.612.9113.0114.8
Obama14.4414.0314.6214.34
Cat17.2915.9616.1216.27
", + "bbox": [ + 300, + 512, + 702, + 583 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Table 3: FID for different values of $\\epsilon$", + "bbox": [ + 388, + 583, + 612, + 595 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Conclusion", + "text_level": 1, + "bbox": [ + 214, + 638, + 344, + 652 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In this paper, we identified a latent space misalignment between the training and testing phases of existing IMLE-based methods, resulting in poor performance in few-shot image synthesis tasks. To address this issue, we introduced a novel algorithm, RS-IMLE, which modifies the prior distribution used for training. Our experimental results demonstrate that our method significantly enhances the quality of generated images and mode coverage during inference.", + "bbox": [ + 212, + 670, + 787, + 760 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgements: This research was enabled in part by support provided by NSERC, the BC DRI Group and the Digital Research Alliance of Canada. The authors would also like to thank Tristan Engst for extensive help polishing our paper.", + "bbox": [ + 212, + 775, + 787, + 821 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "C. Vashist et al.", + "bbox": [ + 271, + 114, + 364, + 126 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 143, + 310, + 160 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Aghabozorgi, M., Peng, S., Li, K.: Adaptive IMLE for few-shot pretraining-free generative modelling. In: Krause, A., Brunskill, E., Cho, K., Engelhardt, B., Sabato, S., Scarlett, J. (eds.) Proceedings of the 40th International Conference on Machine Learning. Proceedings of Machine Learning Research, vol. 202, pp. 248-264. PMLR (23-29 Jul 2023), https://proceedings.mlr.press/v202/aghabozorgi23a.html", + "2. Brock, A., Donahue, J., Simonyan, K.: Large scale gan training for high fidelity natural image synthesis. ArXiv abs/1809.11096 (2019)", + "3. Child, R.: Very deep vaes generalize autoregressive models and can outperform them on images. ArXiv abs/2011.10650 (2021)", + "4. Dhariwal, P., Nichol, A.: Diffusion models beat gans on image synthesis. ArXiv abs/2105.05233 (2021)", + "5. Dinh, L., Sohl-Dickstein, J.N., Bengio, S.: Density estimation using real nvp. ArXiv abs/1605.08803 (2017)", + "6. Esser, P., Rombach, R., Ommer, B.: Taming transformers for high-resolution image synthesis. 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) pp. 12868-12878 (2020), https://api_semanticscholar.org/CorpusID:229297973", + "7. Goodfellow, I., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., Courville, A., Bengio, Y.: Generative adversarial nets. Advances in neural information processing systems 27 (2014)", + "8. Heusel, M., Ramsauer, H., Unterthiner, T., Nessler, B., Hochreiter, S.: Gans trained by a two time-scale update rule converge to a local nash equilibrium. In: NIPS (2017)", + "9. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. ArXiv abs/2006.11239 (2020)", + "0. Karras, T., Aittala, M., Laine, S., Harkonen, E., Hellsten, J., Lehtinen, J., Aila, T.: Alias-free generative adversarial networks. In: NeurIPS (2021)", + "1. Karras, T., Laine, S., Aila, T.: A style-based generator architecture for generative adversarial networks. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 4401-4410 (2019)", + "2. Karras, T., Laine, S., Aila, T.: A style-based generator architecture for generative adversarial networks (2019)", + "3. Karras, T., Laine, S., Aittala, M., Hellsten, J., Lehtinen, J., Aila, T.: Analyzing and improving the image quality of stylegan. 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) pp. 8107-8116 (2019), https://api_semanticscholar.org/CorpusID:209202273", + "4. Karras, T., Laine, S., Aittala, M., Hellsten, J., Lehtinen, J., Aila, T.: Analyzing and improving the image quality of stylegan. 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) pp. 8107-8116 (2020)", + "5. Kingma, D.P., Dhariwal, P.: Glow: Generative flow with invertible 1x1 convolutions. ArXiv abs/1807.03039 (2018)", + "6. Kingma, D.P., Welling, M.: Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114 (2013)", + "7. Kobyzev, I., Prince, S., Brubaker, M.A.: Normalizing flows: An introduction and review of current methods. IEEE Transactions on Pattern Analysis and Machine Intelligence 43, 3964-3979 (2021)", + "8. Kong, C., Kim, J., Han, D., Kwak, N.: Few-shot image generation with mixup-based distance learning (2022)", + "9. Kynkänniemi, T., Karras, T., Laine, S., Lehtinen, J., Aila, T.: Improved precision and recall metric for assessing generative models. Advances in Neural Information Processing Systems 32 (2019)" + ], + "bbox": [ + 225, + 172, + 785, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "Rejection Sampling IMLE", + "bbox": [ + 571, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 769, + 116, + 784, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "20. Li, K., Malik, J.: Fast k-nearest neighbour search via prioritized dci (2017)", + "21. Li, K., Malik, J.: Implicit maximum likelihood estimation. arXiv preprint arXiv:1809.09087 (2018)", + "22. Li, Y., Zhang, R., Lu, J., Shechtman, E.: Few-shot image generation with elastic weight consolidation. ArXiv abs/2012.02780 (2020)", + "23. Li, Z., Wang, C., Zheng, H., Zhang, J., Li, B.: Fakeclr: Exploring contrastive learning for solving latent discontinuity in data-efficient gans. In: ECCV (2022)", + "24. Liu, B., Zhu, Y., Song, K., Elgammal, A.: Towards faster and stabilized GAN training for high-fidelity few-shot image synthesis. CoRR abs/2101.04775 (2021), https://arxiv.org/abs/2101.04775", + "25. Mo, S., Cho, M., Shin, J.: Freeze discriminator: A simple baseline for fine-tuning gans. ArXiv abs/2002.10964 (2020)", + "26. Ojha, U., Li, Y., Lu, J., Efros, A.A., Lee, Y.J., Shechtman, E., Zhang, R.: Few-shot image generation via cross-domain correspondence. 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) pp. 10738-10747 (2021)", + "27. van den Oord, A., Kalchbrenner, N., Espeholt, L., Kavukcuoglu, K., Vinyals, O., Graves, A.: Conditional image generation with pixelCNN decoders. ArXiv abs/1606.05328 (2016)", + "28. van den Oord, A., Kalchbrenner, N., Kavukcuoglu, K.: Pixel recurrent neural networks. ArXiv abs/1601.06759 (2016)", + "29. Razavi, A., van den Oord, A., Vinyals, O.: Generating diverse high-fidelity images with vqvae-2. ArXiv abs/1906.00446 (2019)", + "30. Salimans, T., Karpathy, A., Chen, X., Kingma, D.P.: PixelCNN++: Improving the pixelCNN with discretized logistic mixture likelihood and other modifications. ArXiv abs/1701.05517 (2017)", + "31. Sauer, A., Chitta, K., Muller, J., Geiger, A.: Projected gans converge faster. In: Neural Information Processing Systems (2021), https://api-semanticscholar.org/CorpusID: 240354401", + "32. Saxena, D., Cao, J., Xu, J., Kulshrestha, T.: Re-gan: Data-efficient gans training via architectural reconfiguration. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 16230-16240 (June 2023)", + "33. Si, Z., Zhu, S.C.: Learning hybrid image templates (hit) by information projection. IEEE Transactions on Pattern Analysis and Machine Intelligence 34, 1354-1367 (2012)", + "34. Song, Y., Ermon, S.: Generative modeling by estimating gradients of the data distribution. ArXiv abs/1907.05600 (2019)", + "35. Song, Y., Sohl-Dickstein, J.N., Kingma, D.P., Kumar, A., Ermon, S., Poole, B.: Score-based generative modeling through stochastic differential equations. ArXiv abs/2011.13456 (2021)", + "36. Vahdat, A., Kautz, J.: Nvae: A deep hierarchical variational autoencoder. ArXiv abs/2007.03898 (2020)", + "37. Wang, Y., Gonzalez-Garcia, A., Berga, D., Herranz, L., Khan, F.S., van de Weijer, J.: Minegan: Effective knowledge transfer from gans to target domains with few images. 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) pp. 9329-9338 (2020)", + "38. Yang, M., Wang, Z., Chi, Z., Zhang, Y.: Fregan: Exploiting frequency components for training gans under limited data (2022)", + "39. Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual metric. In: CVPR (2018)", + "40. Zhao, M., Cong, Y., Carin, L.: On leveraging pretrained gans for generation with limited data. In: ICML (2020)", + "41. Zhao, S., Liu, Z., Lin, J., Zhu, J.Y., Han, S.: Differentiable augmentation for data-efficient gan training. ArXiv abs/2006.10738 (2020)" + ], + "bbox": [ + 215, + 147, + 785, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "C. Vashist et al.", + "bbox": [ + 271, + 114, + 364, + 127 + ], + "page_idx": 15 + }, + { + "type": "ref_text", + "text": "42. Zhu, P., Abdal, R., Qin, Y., Wonka, P.: Improved stylegan embedding: Where are the good latents? ArXiv abs/2012.09036 (2020)", + "bbox": [ + 215, + 146, + 787, + 175 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "Rejection Sampling IMLE", + "bbox": [ + 571, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + } +] \ No newline at end of file diff --git a/2024/Rejection Sampling IMLE_ Designing Priors for Better Few-Shot Image Synthesis/0d5c28ac-bfc0-4636-af16-85b85c6b5148_model.json b/2024/Rejection Sampling IMLE_ Designing Priors for Better Few-Shot Image Synthesis/0d5c28ac-bfc0-4636-af16-85b85c6b5148_model.json new file mode 100644 index 0000000000000000000000000000000000000000..f0877c2a6c9ed39cace5ff1d6f888ef6d66d9abb --- /dev/null +++ b/2024/Rejection Sampling IMLE_ Designing Priors for Better Few-Shot Image Synthesis/0d5c28ac-bfc0-4636-af16-85b85c6b5148_model.json @@ -0,0 +1,2687 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.227, + 0.141, + 0.78, + 0.187 + ], + "angle": 0, + "content": "Rejection Sampling IMLE: Designing Priors for Better Few-Shot Image Synthesis" + }, + { + "type": "text", + "bbox": [ + 0.345, + 0.213, + 0.658, + 0.228 + ], + "angle": 0, + "content": "Chirag Vashist, Shichong Peng, and Ke Li" + }, + { + "type": "text", + "bbox": [ + 0.468, + 0.241, + 0.536, + 0.253 + ], + "angle": 0, + "content": "APEX Lab" + }, + { + "type": "text", + "bbox": [ + 0.413, + 0.256, + 0.591, + 0.268 + ], + "angle": 0, + "content": "School of Computing Science" + }, + { + "type": "text", + "bbox": [ + 0.428, + 0.269, + 0.575, + 0.282 + ], + "angle": 0, + "content": "Simon Fraser University" + }, + { + "type": "text", + "bbox": [ + 0.333, + 0.283, + 0.672, + 0.296 + ], + "angle": 0, + "content": "{chirag_vashist, shichong_peng, keli}@sfu.ca" + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.33, + 0.744, + 0.496 + ], + "angle": 0, + "content": "Abstract. An emerging area of research aims to learn deep generative models with limited training data. Implicit Maximum Likelihood Estimation (IMLE), a recent technique, successfully addresses the mode collapse issue of GANs and has been adapted to the few-shot setting, achieving state-of-the-art performance. However, current IMLE-based approaches encounter challenges due to inadequate correspondence between the latent codes selected for training and those drawn during inference. This results in suboptimal test-time performance. We theoretically show a way to address this issue and propose RS-IMLE, a novel approach that changes the prior distribution used for training. This leads to substantially higher quality image generation compared to existing GAN and IMLE-based methods, as validated by comprehensive experiments conducted on nine few-shot image datasets." + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.51, + 0.741, + 0.536 + ], + "angle": 0, + "content": "Keywords: Few-shot \\(\\cdot\\) Image Synthesis \\(\\cdot\\) Implicit Maximum Likelihood Estimation" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.584, + 0.357, + 0.599 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.614, + 0.789, + 0.795 + ], + "angle": 0, + "content": "Recent years have witnessed significant advancements in image synthesis, driven by the emergence of deep learning and the development of a broad variety of powerful generative models. Generative Adversarial Networks (GANs) [2,7,10,11,14], Variational Autoencoders (VAEs) [3,16,29,36], Diffusion Models [4,9], Score-based Models [34,35], Normalizing Flows [5, 15, 17], and Autoregressive Models [6, 27, 28, 30] have demonstrably improved the quality of synthesized images, often achieving photorealism. Quite often, generative models often require large amounts of training data to achieve this high fidelity. In numerous situations, obtaining large volumes of training data is impractical. For instance, when we want to generate images of rare species of plants or images of anatomical structures with rare diseases, access to abundant images may be limited. In this paper, we aim to tackle the problem of high quality image synthesis using limited training data." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.796, + 0.789, + 0.84 + ], + "angle": 0, + "content": "In the few-shot setting, most of the leading methods to date are based on GANs [18, 23, 24, 32, 38]. Despite advances, these methods continue to be afflicted by mode collapse, a well-known issue in GANs. Mode collapse occurs when the generator network" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.367, + 0.128 + ], + "angle": 0, + "content": "C. Vashist et al." + }, + { + "type": "image", + "bbox": [ + 0.22, + 0.144, + 0.373, + 0.248 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.389, + 0.158, + 0.659, + 0.248 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.665, + 0.159, + 0.768, + 0.249 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.221, + 0.256, + 0.373, + 0.363 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.389, + 0.256, + 0.659, + 0.363 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.665, + 0.256, + 0.767, + 0.364 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.376, + 0.788, + 0.461 + ], + "angle": 0, + "content": "Fig. 1: IMLE is an implicit generative model that maps a latent code sampled from a prior distribution to an image output. In previous IMLE-based methods, both the training and testing phases adopt a standard normal distribution as the prior distribution. However, this approach often results in poor generalization during inference. To address this limitation, we introduce RS-IMLE, which uses rejection sampling to alter the prior distribution used for training to a different distribution \\(\\mathcal{P}\\). This modification significantly enhances the quality of generated images during testing." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.491, + 0.789, + 0.537 + ], + "angle": 0, + "content": "fails to capture the full training data distribution and instead produces a limited subset of outputs. This phenomenon is especially problematic in scenarios where only a small number of training examples are provided." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.538, + 0.789, + 0.644 + ], + "angle": 0, + "content": "Implicit Maximum Likelihood Estimation (IMLE) [21] is an alternative to the GAN objective and has shown promising results in addressing mode collapse. In contrast to GANs, which aim to make each generated image resemble some training data, IMLE instead ensures that each training image has some generated sample close to it, and therefore cannot drop any of the modes present in the training data. Adaptive IMLE [1] further extends IMLE to the few-shot image synthesis setting and achieves state-of-the-art generated image quality and mode coverage." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.645, + 0.79, + 0.842 + ], + "angle": 0, + "content": "However, in existing IMLE-based approaches, we observe that the latent codes used during training and those sampled during testing have different distributions, even though the same prior is used during training and inference. This phenomenon arises because of how IMLE selects latent codes during training. Some regions of the latent space are consistently rarely picked for training, despite having a high likelihood under the prior distribution (often the standard Gaussian). This is illustrated in Fig. 2a. Consequently, at test time, when latent codes drawn from the prior happen to fall in these regions, they yield low-quality samples that are far from the real data points, as illustrated in Fig. 1. We address this issue by carefully choosing a different prior so that the samples selected for training have a distribution more similar to those sampled at inference. Our method, which we call Rejection Sampling IMLE or RS-IMLE for short, demonstrably improves coverage of the latent space used during training, thereby ensuring better alignment with the prior as shown in Fig. 2b. As a result, our method yields higher quality" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.572, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Rejection Sampling IMLE" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.126 + ], + "angle": 0, + "content": "3" + }, + { + "type": "image", + "bbox": [ + 0.219, + 0.145, + 0.481, + 0.338 + ], + "angle": 0, + "content": null + }, + { + "type": "image_footnote", + "bbox": [ + 0.216, + 0.342, + 0.483, + 0.405 + ], + "angle": 0, + "content": "(a) Latent space of model trained by IMLE objective using standard normal prior. Dots represent the points selected by the model over the course of training, with dots of the same colour belonging to the same data-point. The contours of standard normal distribution have been shown for comparison." + }, + { + "type": "image", + "bbox": [ + 0.522, + 0.145, + 0.785, + 0.338 + ], + "angle": 0, + "content": null + }, + { + "type": "image_footnote", + "bbox": [ + 0.521, + 0.342, + 0.788, + 0.394 + ], + "angle": 0, + "content": "(b) Latent space of model trained by RS-IMLE objective using prior obtained via rejection sampling. Compared to latent space of model trained by IMLE, our method over the course of training, samples latent codes that follow the distribution at test time more faithfully." + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.415, + 0.789, + 0.445 + ], + "angle": 0, + "content": "Fig. 2: Difference between the latent codes picked by IMLE and RS-IMLE over the course of training." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.475, + 0.789, + 0.536 + ], + "angle": 0, + "content": "samples during testing compared to existing GAN and IMLE-based methods. We substantiate this claim through theoretical analysis and extensive experiments conducted on nine few-shot image datasets. We achieve an average of \\(45.9\\%\\) decrease in FID [8] across datasets compared to the best baseline." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.563, + 0.367, + 0.579 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.599, + 0.789, + 0.844 + ], + "angle": 0, + "content": "Training deep generative models with limited data remains a significant challenge. One approach involves adapting a model pretrained on a large-scale auxiliary dataset from similar domains [22,25,26,31,37,40]. However, the availability of such large-scale auxiliary datasets across all domains is not guaranteed. Therefore, another emerging line of work focuses on training models from scratch. In this context, due to the scarcity of training data, diffusion models struggle to achieve high-quality generated images and have been demonstrated to be ineffective [1]. As a result, previous works in this area predominantly build on Generative Adversarial Networks (GANs) and design various methods to address the well-known mode collapse issue. Techniques such as ADA [13] and DiffAug [41] aim to expand training data using adaptive and differentiable augmentation strategies. FastGAN [24] introduced a skip-layer excitation module for accelerated training and used self-supervision in the discriminator to enhance feature learning, thereby improving mode coverage of the generator. FakeCLR [23] enhances image synthesis by extensive data augmentation and applies contrastive learning solely on perturbed fake samples. FreGAN [38] introduces a frequency-aware model with a self-supervised constraint to avoid generating arbitrary frequency signals. ReGAN [32] dynamically adjusts" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.366, + 0.128 + ], + "angle": 0, + "content": "C. Vashist et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.191 + ], + "angle": 0, + "content": "GANs' architecture during training to explore diverse sub-network structures at different training times. However, despite these advancements, some degree of mode collapse persists." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.193, + 0.788, + 0.314 + ], + "angle": 0, + "content": "In contrast, Implicit Maximum Likelihood Estimation (IMLE) [21] shows promising results in addressing mode collapse through the use of an alternative objective function compared to GANs. Building upon IMLE, Adaptive IMLE [1] adapts this approach to the few-shot image synthesis scenario by introducing individual target thresholds for each training data point. This dynamic adjustment of training progress accounts for varying difficulties across different data points, thereby effectively leveraging the limited training data. In this work, we introduce a novel algorithm, orthogonal to Adaptive IMLE, for sample selection during training." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.338, + 0.317, + 0.354 + ], + "angle": 0, + "content": "3 Method" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.372, + 0.344, + 0.387 + ], + "angle": 0, + "content": "3.1 Background" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.398, + 0.788, + 0.579 + ], + "angle": 0, + "content": "In the context of unconditional image synthesis, the primary objective is to learn the unconditional probability distribution of images \\( p(\\mathbf{x}) \\). This distribution enables the generation of novel synthesized images through the sampling. Implicit generative models are commonly employed for this task, with the generator in GANs serving as a notable example. Generator in GANs are represented by a function \\( T_{\\theta}: Z \\to X \\), implemented as a neural network with parameters denoted as \\( \\theta \\). The function \\( T_{\\theta} \\) learns a transformation from the latent space \\( Z \\) to the image space \\( X \\) by using adversarial training, which employs a discriminator that tries to distinguish between generated images \\( T_{\\theta}(\\mathbf{z}) \\) and real images \\( \\mathbf{x} \\), while the generator tries to produce increasingly realistic images to deceive the discriminator. However, this objective often leads to mode collapse, a well-known issue of GANs, where the generated output \\( T_{\\theta}(\\mathbf{z}) \\) only models a subset of the training examples." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.581, + 0.788, + 0.67 + ], + "angle": 0, + "content": "To address the issue of mode collapse, an alternative method Implicit Maximum Likelihood Estimation (IMLE) [21] has been introduced. While IMLE, like GANs, uses a generator, it differs from GANs by using an alternative objective that aims to make each generated sample similar to some training data. Instead, IMLE ensures that each training data point has similar generated samples, thereby encouraging coverage of all the modes of the training data." + }, + { + "type": "text", + "bbox": [ + 0.239, + 0.671, + 0.725, + 0.686 + ], + "angle": 0, + "content": "The IMLE objective is given as follows, where \\(d(\\cdot ,\\cdot)\\) is a distance metric:" + }, + { + "type": "equation", + "bbox": [ + 0.295, + 0.697, + 0.787, + 0.74 + ], + "angle": 0, + "content": "\\[\n\\theta_ {\\mathrm {I M L E}} = \\arg \\min _ {\\theta} \\mathbb {E} _ {z _ {1}, \\dots , z _ {m} \\sim \\mathcal {N} (0, I)} \\left[ \\sum_ {i = 1} ^ {n} \\min _ {j \\in [ m ]} d \\left(\\mathbf {x} _ {i}, T _ {\\theta} (\\mathbf {z} _ {j})\\right) \\right] \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Here \\(m\\) denotes the number of samples and \\(n\\) denotes the number of data points. In simple terms, the IMLE objective first draws \\(m\\) samples \\(\\mathbf{z}_j\\) from the standard Gaussian distribution and transforms them into the image space using the function \\(T_{\\theta}\\). From these pool of samples in the image space, for each data point \\(x_i\\), IMLE selects a sample that is closest to the data point in some distance metric \\(d(\\cdot, \\cdot)\\). This operation can be done efficiently due to advances in high-dimensional nearest neighbour search [20]. Note that" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.572, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Rejection Sampling IMLE" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.147, + 0.788, + 0.207 + ], + "angle": 0, + "content": "the number of samples \\( m \\) must at least be equal to the number of data points \\( n \\), since otherwise by the pigeonhole principle, some samples would be picked by multiple data points. In practice, we find that setting \\( m \\) to be a multiplicative factor (like 10 or 20) times larger than \\( n \\) works the best." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.229, + 0.345, + 0.243 + ], + "angle": 0, + "content": "3.2 Observation" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.254, + 0.789, + 0.435 + ], + "angle": 0, + "content": "In the existing IMLE-based methods, we observe that the distributions of the latent codes used for training the objective differs from the distribution of latent encountered at test time. Consider an illustrative example where the latent space is two dimensional. We train a simple generative model using IMLE on two dimensional toy dataset. The latent codes used for training over the course of training are illustrated in Figure 2a. We notice that for the latent codes belonging to the same data point (denoted by the same colour) form well-separated tight bands in the latent space. We also observe that there are large gaps between these bands, indicating that these segments of the latent space are consistently overlooked during training. Since at test time we sample from the same standard normal distribution, these unsupervised segments in the latent space have arbitrary outputs, which result in bad samples. We term this phenomenon the \"Misalignment Issue.\"" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.456, + 0.5, + 0.472 + ], + "angle": 0, + "content": "3.3 Analysis of the Misalignment Issue" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.481, + 0.787, + 0.528 + ], + "angle": 0, + "content": "In this section, we will try to see why the phenomenon mentioned above occurs. Let us clarify the notation used in subsequent sections: \\(\\mathbf{x}_i\\) denotes the \\(i^{\\mathrm{th}}\\) data point, \\(d(\\cdot ,\\cdot)\\) denotes the distance function and \\(T_{\\theta}(\\mathbf{z}_j)\\) denote the \\(j^{\\mathrm{th}}\\) sample where \\(\\mathbf{z}_j\\sim \\mathcal{N}(0,I)\\)." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.528, + 0.788, + 0.631 + ], + "angle": 0, + "content": "Let us define a random variable, \\( D_{ij} \\) to denote the distance of \\( i^{\\text{th}} \\) data point to the \\( j^{\\text{th}} \\) sample. We can also define another random variable \\( D_i^* \\) to denote the distance of \\( i^{\\text{th}} \\) data point to the sample closest to it. Further, let \\( F_{D_{ij}} \\) and \\( F_{D_i^*} \\) be the CDF of \\( D_{ij} \\) and \\( D_i^* \\) respectively. Let \\( f_{D_{ij}} \\) and \\( f_{D_i^*} \\) be the PDF of \\( D_{ij} \\) and \\( D_i^* \\) respectively. Now we will relate the CDF of the distances between a data point and its selected latent code, \\( F_{D_i^*} \\) to the CDF of the distances between the same data point and a random latent code \\( F_{D_{ij}} \\):" + }, + { + "type": "equation", + "bbox": [ + 0.293, + 0.653, + 0.786, + 0.785 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} F _ {D _ {i} ^ {*}} (t) = \\operatorname * {P r} (D _ {i} ^ {*} \\leq t) = 1 - \\operatorname * {P r} (D _ {i} ^ {*} > t) \\\\ = 1 - \\Pr (D _ {i j} > t, \\forall j \\in [ m ]) \\quad (\\text {D e f . o f} D _ {i j}) \\\\ = 1 - \\prod_ {j = 1} ^ {m} \\Pr (D _ {i j} > t) \\\\ = 1 - \\left(\\Pr \\left(D _ {i 1} > t\\right)\\right) ^ {m} (2) \\\\ = 1 - \\left(1 - F _ {D _ {i 1}} (t)\\right) ^ {m} (3) \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.795, + 0.787, + 0.841 + ], + "angle": 0, + "content": "Note that Equation 2 is true because each \\(\\mathbf{z}_j\\) is drawn independently from the same probability distribution which makes \\(D_{i1}, D_{i2} \\dots D_{im}\\) identical in distribution for a particular data point \\(\\mathbf{x}_i\\)." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.367, + 0.128 + ], + "angle": 0, + "content": "C. Vashist et al." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.146, + 0.402, + 0.256 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.235, + 0.259, + 0.386, + 0.272 + ], + "angle": 0, + "content": "(a) PDF of example distribution" + }, + { + "type": "image", + "bbox": [ + 0.412, + 0.146, + 0.593, + 0.256 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.411, + 0.259, + 0.594, + 0.277 + ], + "angle": 0, + "content": "(b) \\(F_{D_i^*}(t)\\) vs \\(t\\) for different values of \\(m\\)" + }, + { + "type": "image", + "bbox": [ + 0.604, + 0.146, + 0.785, + 0.256 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.602, + 0.259, + 0.786, + 0.283 + ], + "angle": 0, + "content": "(c) \\(F_{D_i^*}(t)\\) vs \\(F_{D_{i1}}(t)\\) for different values of \\(m\\)" + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.296, + 0.789, + 0.325 + ], + "angle": 0, + "content": "Fig. 3: Illustrative figure for demonstrating the behaviour of \\( F_{D_i^*}(t) \\) and \\( F_{D_{i1}}(t) \\) using Noncentral Chi-squared distribution as the example distribution." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.355, + 0.788, + 0.465 + ], + "angle": 0, + "content": "Now we can try to justify our observations from Figure 2a. Equation 3 shows us how the CDF of the distance of the selected latent code (used in training) differs significantly from the CDF of distance of a random latent code (encountered at test time). This shows that the distance of the sample we choose for training is lower than the distance of the sample at testing. We can obtain a deeper understanding by analyzing the plots in Figure 3 for an example distribution. Notice that \\(\\forall m > 1, F_{D_i^*}(t) > F_{D_{i1}}(t)\\). Observing Equation 3 and the aforementioned plots, we can conclude the following things:" + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.476, + 0.787, + 0.522 + ], + "angle": 0, + "content": "1. The \\( f_{D_i^*} \\) is skewed towards the origin compared to \\( f_{D_{ij}} \\). This is intuitive because the latent codes are selected by the min operation and so their distance to their respective data point would be less than that of a random sample." + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.523, + 0.787, + 0.556 + ], + "angle": 0, + "content": "2. The skew towards the data point increases as \\( m \\) increases. This piece of information will be important shortly." + }, + { + "type": "list", + "bbox": [ + 0.223, + 0.476, + 0.787, + 0.556 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.567, + 0.788, + 0.598 + ], + "angle": 0, + "content": "We can also compute the PDF \\( f_{D_i^*} \\) in terms of \\( f_{D_{i1}} \\) by differentiating the CDF \\( F_{D_i^*}(t) \\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.341, + 0.623, + 0.787, + 0.657 + ], + "angle": 0, + "content": "\\[\nf _ {D _ {i} ^ {*}} (t) = \\frac {\\mathrm {d} F _ {D _ {i} ^ {*}} (t)}{\\mathrm {d} t} = m \\left(1 - F _ {D _ {i 1}} (t)\\right) ^ {m - 1} f _ {D _ {i 1}} (t) \\tag {4}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.678, + 0.475, + 0.695 + ], + "angle": 0, + "content": "3.4 Solving the Misalignment Issue" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.705, + 0.789, + 0.844 + ], + "angle": 0, + "content": "Now that we know the reason behind the misalignment between latent codes used at training and testing, we wish to come up with a method to mitigate this phenomenon. Recall that \\( D_{ij} \\) and \\( D_i^* \\) are determined by the distance function \\( d(\\cdot, \\cdot) \\), neural network \\( T_\\theta \\) and the prior distribution. For a given generative modelling task, it is not trivial to change the \\( d(\\cdot, \\cdot) \\) or \\( T_\\theta \\). Hence, in this paper, we aim to change the prior distribution used at training such that the distribution of latent codes at training time closely matches with the distribution of latent codes (drawn from the standard normal distribution) encountered at test time. Notice that the IMLE objective (Equation 1) allows us to sample from the prior without knowing the closed form expression for its probability density" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.572, + 0.115, + 0.732, + 0.13 + ], + "angle": 0, + "content": "Rejection Sampling IMLE" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "code_caption", + "bbox": [ + 0.217, + 0.146, + 0.445, + 0.162 + ], + "angle": 0, + "content": "Algorithm 1 RS-IMLE Procedure" + }, + { + "type": "algorithm", + "bbox": [ + 0.217, + 0.165, + 0.543, + 0.35 + ], + "angle": 0, + "content": "Require: The set of inputs \\(\\{\\mathbf{x}_i\\}_{i = 1}^n\\) , radius \\(\\epsilon\\) \n1: Initialize the parameters \\(\\theta\\) of the generator \\(T_{\\theta}\\) \n2: for \\(k = 1\\) to \\(K\\) do \n3: Draw latent codes \\(Z\\gets \\mathbf{z}_1,\\dots,\\mathbf{z}_m\\) from \\(\\mathcal{N}(0,\\mathbf{I})\\) \n4: Compute \\(\\tilde{Z}\\gets \\tilde{\\mathbf{z}}_1,\\ldots ,\\tilde{\\mathbf{z}}_p\\) from \\(Z\\) such that \\(d\\left(\\mathbf{x}_i,T_\\theta (\\tilde{\\mathbf{z}}_j)\\right)\\geq \\epsilon ,\\quad \\forall \\tilde{\\mathbf{z}}_j\\in \\tilde{Z},i\\in [n]\\) \n5: \\(\\sigma (i)\\gets \\arg \\min_{j\\in [m]}d(\\mathbf{x}_i,T_\\theta (\\tilde{\\mathbf{z}}_j)),\\quad \\forall i\\in [n]\\) \n6: for \\(l = 1\\) to \\(L\\) do \n7: Pick a random batch \\(S\\subseteq [n]\\) \n8: \\(\\theta \\gets \\theta -\\eta \\nabla_{\\theta}\\left(\\sum_{i\\in S}d\\left(\\mathbf{x}_i,T_\\theta \\left(\\tilde{\\mathbf{z}}_{\\sigma (i)}\\right)\\right)\\right) / |S|\\) \n9: end for \n10: end for \n11: return \\(\\theta\\)" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.392, + 0.788, + 0.507 + ], + "angle": 0, + "content": "function. This allows us the flexibility of choosing a non-analytical prior distribution. To distinguish from \\(\\mathbf{z}_j\\sim \\mathcal{N}(0,I)\\), we will use \\(\\tilde{\\mathbf{z}}\\) to denote the latent codes drawn from our desired target distribution \\(\\mathcal{P}\\). Identical to the previous section, let us define random variable \\(\\tilde{D}_{ij} = d\\left(\\mathbf{x}_i,T_\\theta (\\tilde{\\mathbf{z}}_j)\\right)\\) to denote the distance between data point \\(\\mathbf{x}_i\\) from a random sample \\(T_{\\theta}(\\tilde{\\mathbf{z}}_j)\\). Similarly, we define \\(\\tilde{D}_i^* = \\min_{j\\in [m]}\\tilde{D}_{ij}\\). Similar to the section above, let \\(F_{\\tilde{D}_{ij}}\\) and \\(F_{\\tilde{D}_i^*}\\) be the CDF of \\(\\tilde{D}_{ij}\\) and \\(\\tilde{D}_i^*\\) respectively. Then \\(f_{\\tilde{D}_{ij}}\\) and \\(f_{\\tilde{D}_i^*}\\) would be the PDF of \\(\\tilde{D}_{ij}\\) and \\(\\tilde{D}_i^*\\) respectively." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.554, + 0.788, + 0.766 + ], + "angle": 0, + "content": "Designing the target prior Now, we will try to think about the properties we would want to have for our target prior. Recall that the Misalignment Issue is mitigated as we decrease the number of samples. In order to differentiate between the number of samples of different priors, we use the notation \\( m' \\) to denote the number of samples for the objective using the Gaussian prior. As hinted in the previous section, one way to avoid the Misalignment Issue is to set \\( m' \\) to low values. However we cannot directly use a Gaussian prior with low values of \\( m' \\) as our target prior. This is because in aforementioned case, it happens to be that many data points are forced to pick the same sample as their nearest neighbour. This in turn causes very slow or no learning. Hence when using a Gaussian prior, we need to pick \\( m' \\) large enough to allow convergence and yet small enough such that the Misalignment Issue does not affect test time performance. In our case, we are trying to design a new distribution that solves the Misalignment Issue by having desirable properties of an ideal distribution. The ideal distribution is a Gaussian prior with \\( m' \\) set to the lowest possible value, which is \\( m' = n \\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.77, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Now that we have found the ideal prior, we can make \\(\\mathcal{P}\\) and the ideal prior to be equal in distribution. Similar to the analysis till Equation 4, we derive the PDF of \\(\\mathcal{P}\\). We get \\(f_{\\tilde{D}_i^*}(t) = m\\left(1 - F_{\\tilde{D}_{i1}}(t)\\right)^{m - 1}f_{\\tilde{D}_{i1}}(t)\\). Equating this PDF of \\(\\mathcal{P}\\) to the PDF of the ideal distribution gives:" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.228, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.366, + 0.128 + ], + "angle": 0, + "content": "C. Vashist et al." + }, + { + "type": "equation", + "bbox": [ + 0.318, + 0.171, + 0.786, + 0.284 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} m \\left(1 - F _ {\\tilde {D} _ {i 1}} (t)\\right) ^ {m - 1} f _ {\\tilde {D} _ {i 1}} (t) = n \\left(1 - F _ {D _ {i 1}} (t)\\right) ^ {n - 1} f _ {D _ {i 1}} (t) \\\\ \\Rightarrow f _ {\\tilde {D} _ {i 1}} (t) = \\frac {n}{m} \\frac {\\left(1 - F _ {D _ {i 1}} (t)\\right) ^ {n - 1}}{\\left(1 - F _ {\\tilde {D} _ {i 1}} (t)\\right) ^ {m - 1}} f _ {D _ {i 1}} (t) (5) \\\\ \\Longrightarrow f _ {\\tilde {D} _ {i 1}} (t) = c \\phi (t) f _ {D _ {i 1}} (t) (6) \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.293, + 0.786, + 0.339 + ], + "angle": 0, + "content": "Here, \\(\\phi(t)\\) is introduced to simplify notation and \\(c\\) is a constant associated with truncating \\(F_{D_{i1}}(t)\\) that is needed to make sure \\(f_{\\tilde{D}_{i1}}(t)\\) is bounded. We discuss these technical details in the Appendix." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.361, + 0.784, + 0.452 + ], + "angle": 0, + "content": "Rejection Sampling We have expressed our target prior \\(\\mathcal{P}\\) in terms of distribution from which we can easily sample. Now, we can use Rejection Sampling to sample from our target prior \\(\\mathcal{P}\\). To be concrete: \\(f_{\\tilde{D}_{i1}^*}(t)\\) is our target distribution and \\(f_{D_{i1}}(t)\\) acts as our proposal distribution, since we can sample from the standard Gaussian easily. We can further write it in the standard Rejection Sampling notation: \\(\\frac{f_{\\tilde{D}_{i1}^*}(t)}{M f_{D^*}(t)} = \\frac{c\\phi(t)}{M}\\)" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.454, + 0.785, + 0.528 + ], + "angle": 0, + "content": "Here, \\(M\\) is the scaling factor associated with Rejection Sampling. We approximate the function above using a step function. The step needs to happen at \\(t\\) where \\(F_{\\tilde{D}_{i1}}(t)\\) gets close to 1. Instead of trying to estimate \\(F_{\\tilde{D}_{i1}}(t)\\), we instead use a hyperparameter \\(\\epsilon\\) to represent where \\(F_{\\tilde{D}_{i1}}(t)\\) gets close to 1. We find the value of this hyperparameter \\(\\epsilon\\) by cross-validation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.529, + 0.785, + 0.59 + ], + "angle": 0, + "content": "The final procedure simplifies to this: we sample \\(\\mathbf{z} \\sim \\mathcal{N}(0, I)\\). If \\(t = d\\left(\\mathbf{x}_i, T_\\theta(\\mathbf{z})\\right) < \\epsilon\\), we reject the sample; otherwise, we accept it. Since the sampling procedure is based on Rejection Sampling, we call our method RS-IMLE. The resulting RS-IMLE procedure is included in Algorithm-1." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.612, + 0.616, + 0.628 + ], + "angle": 0, + "content": "3.5 Intuitive Interpretation of the Algorithm Behaviour" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.637, + 0.784, + 0.774 + ], + "angle": 0, + "content": "Prior to proceeding to the Implementation details, gaining a gradient-based understanding of our new objective would provide valuable insight. Let us revisit the IMLE objective (Equation 1) again. As training progresses, we expect the loss to reduce such that: \\(\\forall x_{i},\\quad \\mathbb{E}_{z_{1},\\ldots ,z_{m}\\sim \\mathcal{N}(0,I)}\\min_{j\\in [m]}d\\left(\\mathbf{x}_{i},T_{\\theta}(\\mathbf{z}_{j})\\right)\\to 0.\\) As the objective approaches convergence, the loss will decrease, resulting in gradients with lower magnitude. This causes smaller updates to the model parameters during training, leading to slower progress. Note that this loss is with respect to the closest sample to each data point. Thus it can be the case that even after a lot of training, although the closest sample are pretty close to their respective data point, the rest of samples are pretty far away." + }, + { + "type": "equation", + "bbox": [ + 0.303, + 0.797, + 0.786, + 0.839 + ], + "angle": 0, + "content": "\\[\n\\theta_ {\\mathrm {R S - I M L E}} = \\arg \\min _ {\\theta} \\mathbb {E} _ {z _ {1}, \\dots , z _ {m} \\sim \\mathcal {P}} \\left[ \\sum_ {i = 1} ^ {n} \\min _ {j \\in [ m ]} d \\left(\\mathbf {x} _ {i}, T _ {\\theta} (\\mathbf {z} _ {j})\\right) \\right] \\tag {7}\n\\]" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.572, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Rejection Sampling IMLE" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "image", + "bbox": [ + 0.237, + 0.155, + 0.481, + 0.24 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.299, + 0.259, + 0.422, + 0.271 + ], + "angle": 0, + "content": "(a) IMLE after epoch 100" + }, + { + "type": "image", + "bbox": [ + 0.514, + 0.149, + 0.77, + 0.245 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.576, + 0.259, + 0.715, + 0.271 + ], + "angle": 0, + "content": "(b) RS-IMLE after epoch 100" + }, + { + "type": "image", + "bbox": [ + 0.238, + 0.281, + 0.484, + 0.368 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.301, + 0.384, + 0.422, + 0.396 + ], + "angle": 0, + "content": "(c) IMLE after epoch 500" + }, + { + "type": "image", + "bbox": [ + 0.515, + 0.276, + 0.78, + 0.37 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.576, + 0.384, + 0.715, + 0.396 + ], + "angle": 0, + "content": "(d) RS-IMLE after epoch 500" + }, + { + "type": "image", + "bbox": [ + 0.238, + 0.407, + 0.479, + 0.491 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.296, + 0.51, + 0.426, + 0.522 + ], + "angle": 0, + "content": "(e) IMLE after 2000 epochs" + }, + { + "type": "image", + "bbox": [ + 0.515, + 0.402, + 0.77, + 0.499 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.571, + 0.51, + 0.72, + 0.522 + ], + "angle": 0, + "content": "(f) RS-IMLE after 2000 epochs" + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.534, + 0.788, + 0.563 + ], + "angle": 0, + "content": "Fig. 4: Comparison between IMLE and RS-IMLE for 2D toy problem. Data points are denoted by \\(\\blacksquare\\) and samples are denoted by \\(\\bullet\\). Samples picked as nearest neighbours are denoted by \\(\\star\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.603, + 0.788, + 0.695 + ], + "angle": 0, + "content": "Consider our objective in Equation 7 and recall that we have constructed the probability distribution \\(\\mathcal{P}\\) such that \\(\\forall \\mathbf{x}_i\\quad d\\left(\\mathbf{x}_i,T_\\theta (\\tilde{\\mathbf{z}})\\right)\\geq \\epsilon\\), where \\(\\tilde{\\mathbf{z}}\\sim \\mathcal{P}\\). In other words, all samples we obtain by using the prior \\(\\mathcal{P}\\) are guaranteed to be \\(\\epsilon\\)-distance away from all data points. This ensures that the loss per data point is always greater than \\(\\epsilon\\). The approach can be interpreted as ignoring the samples that are already close to some data point and instead training on challenging, non-trivial samples." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.699, + 0.788, + 0.761 + ], + "angle": 0, + "content": "To compare the sampling behavior of the vanilla IMLE and the proposed RS-IMLE, we trained two models on a 2D toy problem as illustrated in Figure 4. The first model uses the vanilla IMLE objective (Equation 1), while the second model is trained with our proposed RS-IMLE objective (Equation 7)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.788, + 0.841 + ], + "angle": 0, + "content": "At the initial stage of training, both the methods learn similar distributions, indicated by the straight line of orange dots ( \\(\\odot\\) ) in Figure-4a,4b. Our method first removes all the samples that fall within an \\(\\epsilon\\) distance from any data point before doing the nearest neighbour search. We illustrate this in Figure-4b, where samples that lie within the gray circles are not considered for the nearest neighbour search." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.366, + 0.127 + ], + "angle": 0, + "content": "C. Vashist et al." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.144, + 0.787, + 0.297 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.308, + 0.785, + 0.336 + ], + "angle": 0, + "content": "Fig. 5: Qualitative comparison between our method and baselines. While analyzing the images, look for the sharpness of each image and diversity in the content of all images for a method." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.365, + 0.788, + 0.502 + ], + "angle": 0, + "content": "As training progresses, we notice that for both the algorithms' samples move closer to the ground truth data points. However, for vanilla IMLE, we observe that for many data points the sample picked after the nearest neighbour search is already close to the ground truth. In this case, the loss associated with these data points would be low (indicated by the short length of arrow in the Figure-4c). As a result, the model trained by the vanilla IMLE objective does not learn anything significantly novel. In our proposed method (Figure-4d), each data point selects a sample that is at least \\( \\epsilon \\) distance away from it. This ensures that the loss for each data point is always sufficiently high (indicated by the long arrows), resulting in meaningful updates to the model parameters." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.522, + 0.421, + 0.537 + ], + "angle": 0, + "content": "3.6 Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.546, + 0.788, + 0.666 + ], + "angle": 0, + "content": "Note that computing the distance of each sample with each data point is computationally expensive. Suppose we have \\( n \\) data points in \\( \\mathbb{R}^d \\) and \\( m \\) samples, calculating the distance between each pair has a time complexity of \\( \\mathcal{O}(mnd) \\). To get around this issue, we leverage a fast k-nearest neighbor search method, DCI [20]. This method reduces the runtime of a single query from linear to sublinear in \\( m \\), enabling us to efficiently filter out all the samples that are within an \\( \\epsilon \\) distance from any data point. Subsequently, from this filtered pool of remaining samples, we select, for each data point, the sample that is closest to it." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.667, + 0.788, + 0.743 + ], + "angle": 0, + "content": "In order to reduce the search time complexity further, we can project the training data to a lower dimensional subspace (which would decrease \\( d \\)). We use this while implementing the procedure for image synthesis task: we first flatten each image and project it to a lower dimension by using random projection. We normalize these projected vectors before using them for nearest neighbour search." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.764, + 0.359, + 0.781 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.787, + 0.842 + ], + "angle": 0, + "content": "Datasets We assess our method and the baseline approaches across a variety of datasets with \\(256 \\times 256\\) resolution. These datasets include Animal-Face Dog [33], Animal-Face Cat [33], Obama [41], Panda [41], Grumpy-cat [41], Anime [24], Shells [24], Skulls [24]" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.572, + 0.115, + 0.732, + 0.13 + ], + "angle": 0, + "content": "Rejection Sampling IMLE" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.178 + ], + "angle": 0, + "content": "and a subset of Flickr-FaceHQ (FFHQ) [12] which are standard datasets used in the few-shot learning literature." + }, + { + "type": "table", + "bbox": [ + 0.217, + 0.205, + 0.787, + 0.372 + ], + "angle": 0, + "content": "
DatasetFastGAN [24]FakeCLR [23]FreGAN [38]ReGAN [32]AdaIMLE [1]RS-IMLE (Ours)Imp. %
Obama41.129.933.445.725.014.044.0
Grumpy Cat26.620.624.927.319.111.539.8
Panda10.08.89.012.67.63.554.0
FFHQ-10054.262.150.587.433.212.961.1
Cat35.127.431.042.124.915.936.1
Dog50.744.447.957.243.023.146.3
Anime69.877.759.8110.865.835.845.6
Skulls109.6106.5163.3130.781.951.137.6
Shells120.9148.4169.3236.1108.555.448.9
" + }, + { + "type": "table_caption", + "bbox": [ + 0.217, + 0.373, + 0.787, + 0.399 + ], + "angle": 0, + "content": "Table 1: We compute FID [8] between the real data and 5000 randomly generated samples for all the methods. Lower is better." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.42, + 0.785, + 0.464 + ], + "angle": 0, + "content": "Baselines We compare our method to recent state-of-the-art few-shot image generation methods. These include FastGAN [24], FakeCLR [23], FreGAN [38], Re-GAN [32] and AdaIMLE [1]." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.466, + 0.787, + 0.617 + ], + "angle": 0, + "content": "Evaluation Metrics We employ Fréchet Inception Distance (FID) [8] to assess the perceptual quality of the generated images. This involves randomly generating 5000 images and calculating the FID between these generated samples and the real images for each dataset. Additionally, we evaluate the modelling accuracy and coverage by computing precision and recall for 1000 images using the metric defined by Kynkänniemi et al. [19]. In image synthesis, precision refers to the model's capacity to generate images closely resembling the desired target or distribution. Recall measures the model's ability to encompass a wide array of diverse images within the target distribution. To ensure that the computed metrics have low variance, we generate many more samples than training images." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.618, + 0.785, + 0.663 + ], + "angle": 0, + "content": "Network Architecture We construct our generator network using decoder modules from VDVAE [3]. More details about the network architecture can be found in the Appendix." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.689, + 0.4, + 0.705 + ], + "angle": 0, + "content": "4.1 Quantitative Results" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.718, + 0.785, + 0.778 + ], + "angle": 0, + "content": "In Table 1, we present the FID scores computed for all the datasets across different methods. Lower FID scores indicate that the distribution of generated images is closer to the distribution of real images. Our method performs significantly better compared to baselines." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.78, + 0.785, + 0.839 + ], + "angle": 0, + "content": "In Table 2, we show the precision and recall scores. Our method has a near perfect precision (close to 1), while having a significantly higher recall compared to the baselines. In the few cases where our method is not the best, it is very close to the best metric." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.366, + 0.127 + ], + "angle": 0, + "content": "C. Vashist et al." + }, + { + "type": "image_caption", + "bbox": [ + 0.252, + 0.156, + 0.294, + 0.169 + ], + "angle": 0, + "content": "Query" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.169, + 0.328, + 0.253 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.353, + 0.156, + 0.415, + 0.168 + ], + "angle": 0, + "content": "RS-IMLE" + }, + { + "type": "image", + "bbox": [ + 0.329, + 0.169, + 0.44, + 0.253 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.461, + 0.156, + 0.53, + 0.167 + ], + "angle": 0, + "content": "Ada-IMLE" + }, + { + "type": "image", + "bbox": [ + 0.44, + 0.169, + 0.551, + 0.253 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.577, + 0.156, + 0.636, + 0.167 + ], + "angle": 0, + "content": "FastGAN" + }, + { + "type": "image", + "bbox": [ + 0.551, + 0.169, + 0.662, + 0.253 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.69, + 0.156, + 0.744, + 0.167 + ], + "angle": 0, + "content": "FreGAN" + }, + { + "type": "image", + "bbox": [ + 0.662, + 0.169, + 0.774, + 0.253 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.248, + 0.255, + 0.299, + 0.265 + ], + "angle": 0, + "content": "(a) Obama" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.266, + 0.328, + 0.351 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.33, + 0.267, + 0.439, + 0.351 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.44, + 0.267, + 0.551, + 0.351 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.552, + 0.267, + 0.662, + 0.351 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.662, + 0.267, + 0.773, + 0.351 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.248, + 0.353, + 0.298, + 0.363 + ], + "angle": 0, + "content": "(b) Anime" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.365, + 0.328, + 0.45 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.33, + 0.365, + 0.439, + 0.45 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.44, + 0.365, + 0.551, + 0.45 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.552, + 0.365, + 0.661, + 0.45 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.662, + 0.365, + 0.773, + 0.45 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.255, + 0.452, + 0.292, + 0.462 + ], + "angle": 0, + "content": "(c) Dog" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.463, + 0.328, + 0.547 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.33, + 0.463, + 0.439, + 0.547 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.44, + 0.463, + 0.551, + 0.547 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.552, + 0.463, + 0.661, + 0.547 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.662, + 0.463, + 0.773, + 0.547 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.256, + 0.55, + 0.29, + 0.559 + ], + "angle": 0, + "content": "(d) Cat" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.561, + 0.328, + 0.645 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.33, + 0.561, + 0.439, + 0.645 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.44, + 0.561, + 0.551, + 0.645 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.552, + 0.561, + 0.661, + 0.645 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.662, + 0.561, + 0.773, + 0.645 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.251, + 0.648, + 0.296, + 0.657 + ], + "angle": 0, + "content": "(e) Shells" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.66, + 0.328, + 0.743 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.33, + 0.659, + 0.439, + 0.743 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.44, + 0.659, + 0.551, + 0.743 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.552, + 0.659, + 0.661, + 0.743 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.662, + 0.659, + 0.773, + 0.743 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.239, + 0.745, + 0.307, + 0.756 + ], + "angle": 0, + "content": "(f) FFHQ-100" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.768, + 0.788, + 0.825 + ], + "angle": 0, + "content": "Fig. 6: Visual Recall test. The first column is the query image from the dataset. Subsequent columns are the samples produced by different methods that are closest to the query image in LPIPS feature space. The samples produced by our method are closer to the query images compared to the baselines, while remaining diverse." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.572, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Rejection Sampling IMLE" + }, + { + "type": "page_number", + "bbox": [ + 0.77, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "table", + "bbox": [ + 0.217, + 0.144, + 0.787, + 0.44 + ], + "angle": 0, + "content": "
DatasetFastGAN [24]FakeCLR [23]FreGAN [38]ReGAN [32]AdaIMLE [1]RS-IMLE (Ours)
ObamaPrec.0.920.960.820.620.990.98
Rec.0.090.300.330.010.680.82
Grumpy CatPrec.0.910.970.900.780.970.93
Rec.0.130.390.230.040.720.95
PandaPrec.0.960.970.920.930.980.99
Rec.0.160.410.170.010.630.97
FFHQ-100Prec.0.910.710.860.390.991.00
Rec.0.130.250.210.010.770.99
CatPrec.0.970.990.950.900.980.96
Rec.0.080.550.310.150.860.98
DogPrec.0.960.950.920.840.970.98
Rec.0.190.340.200.100.610.94
AnimePrec.0.860.880.880.290.920.95
Rec.0.080.090.090.010.590.91
SkullsPrec.0.780.660.660.660.950.99
Rec.0.030.010.010.010.320.65
ShellsPrec.0.920.870.870.500.970.98
Rec.0.030.010.010.010.620.59
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.217, + 0.441, + 0.787, + 0.468 + ], + "angle": 0, + "content": "Table 2: We compute precision and recall [42] between the real data and 1000 randomly generated samples for all the methods. Higher values are better." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.487, + 0.388, + 0.503 + ], + "angle": 0, + "content": "4.2 Qualitative Results" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.515, + 0.787, + 0.681 + ], + "angle": 0, + "content": "Figure 5 compares the random samples of our method to that of several baselines, and we observe that our method produces overall sharper and more diverse images. In addition, we propose Visual Recall, a simple test to substantiate the qualitative superiority of our method. For each method, we first generate 1000 samples. Next, using a real image from the dataset as a query, we find the images from the pool of generated samples that are closest to the query image. We use LPIPS features [39] for the computing the distance between real images and samples. Figure 6 shows the results for the proposed test for different datasets and methods. We see that the samples produced by our method are visually similar to the query, while being sharp and diverse in attributes like hair colour, smile and jaw structure. Note that other methods do not have samples that closely resemble the query image." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.682, + 0.787, + 0.743 + ], + "angle": 0, + "content": "Figure 7 shows results of spherical linear interpolation between two random points in the latent space for different datasets. The images transition in a meaningful manner, indicating that our model has learnt a continuous and structured latent space representation of the image distribution." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.767, + 0.36, + 0.783 + ], + "angle": 0, + "content": "4.3 Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.795, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Table-3 presents the FID computed for different values of \\(\\epsilon\\) for three datasets. We observe that our approach works best for values of \\(\\epsilon\\) close to 0.15 and that increasing the value of \\(\\epsilon\\) beyond a certain range degrades the performance." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.366, + 0.127 + ], + "angle": 0, + "content": "C. Vashist et al." + }, + { + "type": "image", + "bbox": [ + 0.217, + 0.144, + 0.787, + 0.447 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.478, + 0.45, + 0.526, + 0.459 + ], + "angle": 0, + "content": "(c) Skulls" + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.471, + 0.789, + 0.499 + ], + "angle": 0, + "content": "Fig. 7: Latent space interpolation. We observe that the output images changes smoothly in a meaningful manner." + }, + { + "type": "table", + "bbox": [ + 0.302, + 0.513, + 0.703, + 0.584 + ], + "angle": 0, + "content": "
Dataset0.120.150.180.22
FFHQ-10013.612.9113.0114.8
Obama14.4414.0314.6214.34
Cat17.2915.9616.1216.27
" + }, + { + "type": "table_caption", + "bbox": [ + 0.389, + 0.584, + 0.614, + 0.596 + ], + "angle": 0, + "content": "Table 3: FID for different values of \\( \\epsilon \\)" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.639, + 0.345, + 0.654 + ], + "angle": 0, + "content": "5 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.671, + 0.788, + 0.761 + ], + "angle": 0, + "content": "In this paper, we identified a latent space misalignment between the training and testing phases of existing IMLE-based methods, resulting in poor performance in few-shot image synthesis tasks. To address this issue, we introduced a novel algorithm, RS-IMLE, which modifies the prior distribution used for training. Our experimental results demonstrate that our method significantly enhances the quality of generated images and mode coverage during inference." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.776, + 0.788, + 0.823 + ], + "angle": 0, + "content": "Acknowledgements: This research was enabled in part by support provided by NSERC, the BC DRI Group and the Digital Research Alliance of Canada. The authors would also like to thank Tristan Engst for extensive help polishing our paper." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.572, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Rejection Sampling IMLE" + }, + { + "type": "page_number", + "bbox": [ + 0.77, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.312, + 0.161 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.174, + 0.787, + 0.243 + ], + "angle": 0, + "content": "1. Aghabozorgi, M., Peng, S., Li, K.: Adaptive IMLE for few-shot pretraining-free generative modelling. In: Krause, A., Brunskill, E., Cho, K., Engelhardt, B., Sabato, S., Scarlett, J. (eds.) Proceedings of the 40th International Conference on Machine Learning. Proceedings of Machine Learning Research, vol. 202, pp. 248-264. PMLR (23-29 Jul 2023), https://proceedings.mlr.press/v202/aghabozorgi23a.html" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.243, + 0.787, + 0.27 + ], + "angle": 0, + "content": "2. Brock, A., Donahue, J., Simonyan, K.: Large scale gan training for high fidelity natural image synthesis. ArXiv abs/1809.11096 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.27, + 0.787, + 0.297 + ], + "angle": 0, + "content": "3. Child, R.: Very deep vaes generalize autoregressive models and can outperform them on images. ArXiv abs/2011.10650 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.297, + 0.787, + 0.323 + ], + "angle": 0, + "content": "4. Dhariwal, P., Nichol, A.: Diffusion models beat gans on image synthesis. ArXiv abs/2105.05233 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.324, + 0.787, + 0.351 + ], + "angle": 0, + "content": "5. Dinh, L., Sohl-Dickstein, J.N., Bengio, S.: Density estimation using real nvp. ArXiv abs/1605.08803 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.351, + 0.787, + 0.392 + ], + "angle": 0, + "content": "6. Esser, P., Rombach, R., Ommer, B.: Taming transformers for high-resolution image synthesis. 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) pp. 12868-12878 (2020), https://api_semanticscholar.org/CorpusID:229297973" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.392, + 0.787, + 0.432 + ], + "angle": 0, + "content": "7. Goodfellow, I., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., Courville, A., Bengio, Y.: Generative adversarial nets. Advances in neural information processing systems 27 (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.432, + 0.787, + 0.459 + ], + "angle": 0, + "content": "8. Heusel, M., Ramsauer, H., Unterthiner, T., Nessler, B., Hochreiter, S.: Gans trained by a two time-scale update rule converge to a local nash equilibrium. In: NIPS (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.459, + 0.787, + 0.486 + ], + "angle": 0, + "content": "9. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. ArXiv abs/2006.11239 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.486, + 0.787, + 0.514 + ], + "angle": 0, + "content": "0. Karras, T., Aittala, M., Laine, S., Harkonen, E., Hellsten, J., Lehtinen, J., Aila, T.: Alias-free generative adversarial networks. In: NeurIPS (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.514, + 0.787, + 0.555 + ], + "angle": 0, + "content": "1. Karras, T., Laine, S., Aila, T.: A style-based generator architecture for generative adversarial networks. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 4401-4410 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.555, + 0.787, + 0.581 + ], + "angle": 0, + "content": "2. Karras, T., Laine, S., Aila, T.: A style-based generator architecture for generative adversarial networks (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.581, + 0.787, + 0.635 + ], + "angle": 0, + "content": "3. Karras, T., Laine, S., Aittala, M., Hellsten, J., Lehtinen, J., Aila, T.: Analyzing and improving the image quality of stylegan. 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) pp. 8107-8116 (2019), https://api_semanticscholar.org/CorpusID:209202273" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.635, + 0.787, + 0.677 + ], + "angle": 0, + "content": "4. Karras, T., Laine, S., Aittala, M., Hellsten, J., Lehtinen, J., Aila, T.: Analyzing and improving the image quality of stylegan. 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) pp. 8107-8116 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.677, + 0.787, + 0.704 + ], + "angle": 0, + "content": "5. Kingma, D.P., Dhariwal, P.: Glow: Generative flow with invertible 1x1 convolutions. ArXiv abs/1807.03039 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.704, + 0.787, + 0.731 + ], + "angle": 0, + "content": "6. Kingma, D.P., Welling, M.: Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114 (2013)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.731, + 0.787, + 0.772 + ], + "angle": 0, + "content": "7. Kobyzev, I., Prince, S., Brubaker, M.A.: Normalizing flows: An introduction and review of current methods. IEEE Transactions on Pattern Analysis and Machine Intelligence 43, 3964-3979 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.772, + 0.787, + 0.799 + ], + "angle": 0, + "content": "8. Kong, C., Kim, J., Han, D., Kwak, N.: Few-shot image generation with mixup-based distance learning (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.799, + 0.787, + 0.84 + ], + "angle": 0, + "content": "9. Kynkänniemi, T., Karras, T., Laine, S., Lehtinen, J., Aila, T.: Improved precision and recall metric for assessing generative models. Advances in Neural Information Processing Systems 32 (2019)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.174, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.366, + 0.128 + ], + "angle": 0, + "content": "C. Vashist et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.148, + 0.684, + 0.161 + ], + "angle": 0, + "content": "20. Li, K., Malik, J.: Fast k-nearest neighbour search via prioritized dci (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.162, + 0.786, + 0.188 + ], + "angle": 0, + "content": "21. Li, K., Malik, J.: Implicit maximum likelihood estimation. arXiv preprint arXiv:1809.09087 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.19, + 0.786, + 0.216 + ], + "angle": 0, + "content": "22. Li, Y., Zhang, R., Lu, J., Shechtman, E.: Few-shot image generation with elastic weight consolidation. ArXiv abs/2012.02780 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.218, + 0.786, + 0.244 + ], + "angle": 0, + "content": "23. Li, Z., Wang, C., Zheng, H., Zhang, J., Li, B.: Fakeclr: Exploring contrastive learning for solving latent discontinuity in data-efficient gans. In: ECCV (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.246, + 0.786, + 0.285 + ], + "angle": 0, + "content": "24. Liu, B., Zhu, Y., Song, K., Elgammal, A.: Towards faster and stabilized GAN training for high-fidelity few-shot image synthesis. CoRR abs/2101.04775 (2021), https://arxiv.org/abs/2101.04775" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.287, + 0.786, + 0.313 + ], + "angle": 0, + "content": "25. Mo, S., Cho, M., Shin, J.: Freeze discriminator: A simple baseline for fine-tuning gans. ArXiv abs/2002.10964 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.315, + 0.786, + 0.355 + ], + "angle": 0, + "content": "26. Ojha, U., Li, Y., Lu, J., Efros, A.A., Lee, Y.J., Shechtman, E., Zhang, R.: Few-shot image generation via cross-domain correspondence. 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) pp. 10738-10747 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.357, + 0.786, + 0.383 + ], + "angle": 0, + "content": "27. van den Oord, A., Kalchbrenner, N., Espeholt, L., Kavukcuoglu, K., Vinyals, O., Graves, A.: Conditional image generation with pixelCNN decoders. ArXiv abs/1606.05328 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.385, + 0.786, + 0.41 + ], + "angle": 0, + "content": "28. van den Oord, A., Kalchbrenner, N., Kavukcuoglu, K.: Pixel recurrent neural networks. ArXiv abs/1601.06759 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.412, + 0.786, + 0.437 + ], + "angle": 0, + "content": "29. Razavi, A., van den Oord, A., Vinyals, O.: Generating diverse high-fidelity images with vqvae-2. ArXiv abs/1906.00446 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.44, + 0.786, + 0.479 + ], + "angle": 0, + "content": "30. Salimans, T., Karpathy, A., Chen, X., Kingma, D.P.: PixelCNN++: Improving the pixelCNN with discretized logistic mixture likelihood and other modifications. ArXiv abs/1701.05517 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.481, + 0.786, + 0.52 + ], + "angle": 0, + "content": "31. Sauer, A., Chitta, K., Muller, J., Geiger, A.: Projected gans converge faster. In: Neural Information Processing Systems (2021), https://api-semanticscholar.org/CorpusID: 240354401" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.523, + 0.786, + 0.562 + ], + "angle": 0, + "content": "32. Saxena, D., Cao, J., Xu, J., Kulshrestha, T.: Re-gan: Data-efficient gans training via architectural reconfiguration. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 16230-16240 (June 2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.564, + 0.786, + 0.59 + ], + "angle": 0, + "content": "33. Si, Z., Zhu, S.C.: Learning hybrid image templates (hit) by information projection. IEEE Transactions on Pattern Analysis and Machine Intelligence 34, 1354-1367 (2012)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.592, + 0.786, + 0.617 + ], + "angle": 0, + "content": "34. Song, Y., Ermon, S.: Generative modeling by estimating gradients of the data distribution. ArXiv abs/1907.05600 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.619, + 0.786, + 0.646 + ], + "angle": 0, + "content": "35. Song, Y., Sohl-Dickstein, J.N., Kingma, D.P., Kumar, A., Ermon, S., Poole, B.: Score-based generative modeling through stochastic differential equations. ArXiv abs/2011.13456 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.648, + 0.786, + 0.673 + ], + "angle": 0, + "content": "36. Vahdat, A., Kautz, J.: Nvae: A deep hierarchical variational autoencoder. ArXiv abs/2007.03898 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.675, + 0.786, + 0.729 + ], + "angle": 0, + "content": "37. Wang, Y., Gonzalez-Garcia, A., Berga, D., Herranz, L., Khan, F.S., van de Weijer, J.: Minegan: Effective knowledge transfer from gans to target domains with few images. 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) pp. 9329-9338 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.73, + 0.786, + 0.756 + ], + "angle": 0, + "content": "38. Yang, M., Wang, Z., Chi, Z., Zhang, Y.: Fregan: Exploiting frequency components for training gans under limited data (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.758, + 0.786, + 0.784 + ], + "angle": 0, + "content": "39. Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual metric. In: CVPR (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.786, + 0.786, + 0.811 + ], + "angle": 0, + "content": "40. Zhao, M., Cong, Y., Carin, L.: On leveraging pretrained gans for generation with limited data. In: ICML (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.813, + 0.786, + 0.84 + ], + "angle": 0, + "content": "41. Zhao, S., Liu, Z., Lin, J., Zhu, J.Y., Han, S.: Differentiable augmentation for data-efficient gan training. ArXiv abs/2006.10738 (2020)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.148, + 0.786, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.572, + 0.115, + 0.732, + 0.13 + ], + "angle": 0, + "content": "Rejection Sampling IMLE" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.788, + 0.176 + ], + "angle": 0, + "content": "42. Zhu, P., Abdal, R., Qin, Y., Wonka, P.: Improved stylegan embedding: Where are the good latents? ArXiv abs/2012.09036 (2020)" + } + ] +] \ No newline at end of file diff --git a/2024/Rejection Sampling IMLE_ Designing Priors for Better Few-Shot Image Synthesis/0d5c28ac-bfc0-4636-af16-85b85c6b5148_origin.pdf b/2024/Rejection Sampling IMLE_ Designing Priors for Better Few-Shot Image Synthesis/0d5c28ac-bfc0-4636-af16-85b85c6b5148_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..5e35106a2624ccb07d3076f823b7fbcc7449f19f --- /dev/null +++ b/2024/Rejection Sampling IMLE_ Designing Priors for Better Few-Shot Image Synthesis/0d5c28ac-bfc0-4636-af16-85b85c6b5148_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6e990a7a1b8da26f52da735f57c88e5da5a3b73d67c8a8456bb40ef7af69fc34 +size 9237673 diff --git a/2024/Rejection Sampling IMLE_ Designing Priors for Better Few-Shot Image Synthesis/full.md b/2024/Rejection Sampling IMLE_ Designing Priors for Better Few-Shot Image Synthesis/full.md new file mode 100644 index 0000000000000000000000000000000000000000..9155dbf0c2d00020f244e9cf291357212fa159f2 --- /dev/null +++ b/2024/Rejection Sampling IMLE_ Designing Priors for Better Few-Shot Image Synthesis/full.md @@ -0,0 +1,361 @@ +# Rejection Sampling IMLE: Designing Priors for Better Few-Shot Image Synthesis + +Chirag Vashist, Shichong Peng, and Ke Li + +APEX Lab + +School of Computing Science + +Simon Fraser University + +{chirag_vashist, shichong_peng, keli}@sfu.ca + +Abstract. An emerging area of research aims to learn deep generative models with limited training data. Implicit Maximum Likelihood Estimation (IMLE), a recent technique, successfully addresses the mode collapse issue of GANs and has been adapted to the few-shot setting, achieving state-of-the-art performance. However, current IMLE-based approaches encounter challenges due to inadequate correspondence between the latent codes selected for training and those drawn during inference. This results in suboptimal test-time performance. We theoretically show a way to address this issue and propose RS-IMLE, a novel approach that changes the prior distribution used for training. This leads to substantially higher quality image generation compared to existing GAN and IMLE-based methods, as validated by comprehensive experiments conducted on nine few-shot image datasets. + +Keywords: Few-shot $\cdot$ Image Synthesis $\cdot$ Implicit Maximum Likelihood Estimation + +# 1 Introduction + +Recent years have witnessed significant advancements in image synthesis, driven by the emergence of deep learning and the development of a broad variety of powerful generative models. Generative Adversarial Networks (GANs) [2,7,10,11,14], Variational Autoencoders (VAEs) [3,16,29,36], Diffusion Models [4,9], Score-based Models [34,35], Normalizing Flows [5, 15, 17], and Autoregressive Models [6, 27, 28, 30] have demonstrably improved the quality of synthesized images, often achieving photorealism. Quite often, generative models often require large amounts of training data to achieve this high fidelity. In numerous situations, obtaining large volumes of training data is impractical. For instance, when we want to generate images of rare species of plants or images of anatomical structures with rare diseases, access to abundant images may be limited. In this paper, we aim to tackle the problem of high quality image synthesis using limited training data. + +In the few-shot setting, most of the leading methods to date are based on GANs [18, 23, 24, 32, 38]. Despite advances, these methods continue to be afflicted by mode collapse, a well-known issue in GANs. Mode collapse occurs when the generator network + +![](images/be654b3d05c083e881e770900f841c83b621234de8e6a1ca159fc77b274686f6.jpg) + +![](images/d0f25a73ced6cd7e287d574d3316d1d865e7807ec2dfac233f0e81e50543e202.jpg) + +![](images/a27541844bbc12b215300afd4e1935e0c03a4ddd066d3d47fb71e1181d2f6a3a.jpg) + +![](images/53f47ba949c401a702e6504df83b2bae5e042cd80fdbda7903c9bd8cabe48b47.jpg) +Fig. 1: IMLE is an implicit generative model that maps a latent code sampled from a prior distribution to an image output. In previous IMLE-based methods, both the training and testing phases adopt a standard normal distribution as the prior distribution. However, this approach often results in poor generalization during inference. To address this limitation, we introduce RS-IMLE, which uses rejection sampling to alter the prior distribution used for training to a different distribution $\mathcal{P}$ . This modification significantly enhances the quality of generated images during testing. + +![](images/199fdeded655984a06003c96c65e6b5bed6adf14cc147ce66b1eb3f6bef728c7.jpg) + +![](images/73a107811eae489598567b2572fdd8ea6deb35b3c3c0547008742dfb1963daec.jpg) + +fails to capture the full training data distribution and instead produces a limited subset of outputs. This phenomenon is especially problematic in scenarios where only a small number of training examples are provided. + +Implicit Maximum Likelihood Estimation (IMLE) [21] is an alternative to the GAN objective and has shown promising results in addressing mode collapse. In contrast to GANs, which aim to make each generated image resemble some training data, IMLE instead ensures that each training image has some generated sample close to it, and therefore cannot drop any of the modes present in the training data. Adaptive IMLE [1] further extends IMLE to the few-shot image synthesis setting and achieves state-of-the-art generated image quality and mode coverage. + +However, in existing IMLE-based approaches, we observe that the latent codes used during training and those sampled during testing have different distributions, even though the same prior is used during training and inference. This phenomenon arises because of how IMLE selects latent codes during training. Some regions of the latent space are consistently rarely picked for training, despite having a high likelihood under the prior distribution (often the standard Gaussian). This is illustrated in Fig. 2a. Consequently, at test time, when latent codes drawn from the prior happen to fall in these regions, they yield low-quality samples that are far from the real data points, as illustrated in Fig. 1. We address this issue by carefully choosing a different prior so that the samples selected for training have a distribution more similar to those sampled at inference. Our method, which we call Rejection Sampling IMLE or RS-IMLE for short, demonstrably improves coverage of the latent space used during training, thereby ensuring better alignment with the prior as shown in Fig. 2b. As a result, our method yields higher quality + +Fig. 2: Difference between the latent codes picked by IMLE and RS-IMLE over the course of training. +![](images/aa9c3ab7230d3eb9cf00e8be247a3420fedf3f9fa17a3cae17c24e65e4d3275b.jpg) +(a) Latent space of model trained by IMLE objective using standard normal prior. Dots represent the points selected by the model over the course of training, with dots of the same colour belonging to the same data-point. The contours of standard normal distribution have been shown for comparison. + +![](images/12476c053cfd54019576c9267b61c255a1aac96decaeb53cf9d2ca080474c9fa.jpg) +(b) Latent space of model trained by RS-IMLE objective using prior obtained via rejection sampling. Compared to latent space of model trained by IMLE, our method over the course of training, samples latent codes that follow the distribution at test time more faithfully. + +samples during testing compared to existing GAN and IMLE-based methods. We substantiate this claim through theoretical analysis and extensive experiments conducted on nine few-shot image datasets. We achieve an average of $45.9\%$ decrease in FID [8] across datasets compared to the best baseline. + +# 2 Related Work + +Training deep generative models with limited data remains a significant challenge. One approach involves adapting a model pretrained on a large-scale auxiliary dataset from similar domains [22,25,26,31,37,40]. However, the availability of such large-scale auxiliary datasets across all domains is not guaranteed. Therefore, another emerging line of work focuses on training models from scratch. In this context, due to the scarcity of training data, diffusion models struggle to achieve high-quality generated images and have been demonstrated to be ineffective [1]. As a result, previous works in this area predominantly build on Generative Adversarial Networks (GANs) and design various methods to address the well-known mode collapse issue. Techniques such as ADA [13] and DiffAug [41] aim to expand training data using adaptive and differentiable augmentation strategies. FastGAN [24] introduced a skip-layer excitation module for accelerated training and used self-supervision in the discriminator to enhance feature learning, thereby improving mode coverage of the generator. FakeCLR [23] enhances image synthesis by extensive data augmentation and applies contrastive learning solely on perturbed fake samples. FreGAN [38] introduces a frequency-aware model with a self-supervised constraint to avoid generating arbitrary frequency signals. ReGAN [32] dynamically adjusts + +GANs' architecture during training to explore diverse sub-network structures at different training times. However, despite these advancements, some degree of mode collapse persists. + +In contrast, Implicit Maximum Likelihood Estimation (IMLE) [21] shows promising results in addressing mode collapse through the use of an alternative objective function compared to GANs. Building upon IMLE, Adaptive IMLE [1] adapts this approach to the few-shot image synthesis scenario by introducing individual target thresholds for each training data point. This dynamic adjustment of training progress accounts for varying difficulties across different data points, thereby effectively leveraging the limited training data. In this work, we introduce a novel algorithm, orthogonal to Adaptive IMLE, for sample selection during training. + +# 3 Method + +# 3.1 Background + +In the context of unconditional image synthesis, the primary objective is to learn the unconditional probability distribution of images $p(\mathbf{x})$ . This distribution enables the generation of novel synthesized images through the sampling. Implicit generative models are commonly employed for this task, with the generator in GANs serving as a notable example. Generator in GANs are represented by a function $T_{\theta}: Z \to X$ , implemented as a neural network with parameters denoted as $\theta$ . The function $T_{\theta}$ learns a transformation from the latent space $Z$ to the image space $X$ by using adversarial training, which employs a discriminator that tries to distinguish between generated images $T_{\theta}(\mathbf{z})$ and real images $\mathbf{x}$ , while the generator tries to produce increasingly realistic images to deceive the discriminator. However, this objective often leads to mode collapse, a well-known issue of GANs, where the generated output $T_{\theta}(\mathbf{z})$ only models a subset of the training examples. + +To address the issue of mode collapse, an alternative method Implicit Maximum Likelihood Estimation (IMLE) [21] has been introduced. While IMLE, like GANs, uses a generator, it differs from GANs by using an alternative objective that aims to make each generated sample similar to some training data. Instead, IMLE ensures that each training data point has similar generated samples, thereby encouraging coverage of all the modes of the training data. + +The IMLE objective is given as follows, where $d(\cdot ,\cdot)$ is a distance metric: + +$$ +\theta_ {\mathrm {I M L E}} = \arg \min _ {\theta} \mathbb {E} _ {z _ {1}, \dots , z _ {m} \sim \mathcal {N} (0, I)} \left[ \sum_ {i = 1} ^ {n} \min _ {j \in [ m ]} d \left(\mathbf {x} _ {i}, T _ {\theta} (\mathbf {z} _ {j})\right) \right] \tag {1} +$$ + +Here $m$ denotes the number of samples and $n$ denotes the number of data points. In simple terms, the IMLE objective first draws $m$ samples $\mathbf{z}_j$ from the standard Gaussian distribution and transforms them into the image space using the function $T_{\theta}$ . From these pool of samples in the image space, for each data point $x_i$ , IMLE selects a sample that is closest to the data point in some distance metric $d(\cdot, \cdot)$ . This operation can be done efficiently due to advances in high-dimensional nearest neighbour search [20]. Note that + +the number of samples $m$ must at least be equal to the number of data points $n$ , since otherwise by the pigeonhole principle, some samples would be picked by multiple data points. In practice, we find that setting $m$ to be a multiplicative factor (like 10 or 20) times larger than $n$ works the best. + +# 3.2 Observation + +In the existing IMLE-based methods, we observe that the distributions of the latent codes used for training the objective differs from the distribution of latent encountered at test time. Consider an illustrative example where the latent space is two dimensional. We train a simple generative model using IMLE on two dimensional toy dataset. The latent codes used for training over the course of training are illustrated in Figure 2a. We notice that for the latent codes belonging to the same data point (denoted by the same colour) form well-separated tight bands in the latent space. We also observe that there are large gaps between these bands, indicating that these segments of the latent space are consistently overlooked during training. Since at test time we sample from the same standard normal distribution, these unsupervised segments in the latent space have arbitrary outputs, which result in bad samples. We term this phenomenon the "Misalignment Issue." + +# 3.3 Analysis of the Misalignment Issue + +In this section, we will try to see why the phenomenon mentioned above occurs. Let us clarify the notation used in subsequent sections: $\mathbf{x}_i$ denotes the $i^{\mathrm{th}}$ data point, $d(\cdot ,\cdot)$ denotes the distance function and $T_{\theta}(\mathbf{z}_j)$ denote the $j^{\mathrm{th}}$ sample where $\mathbf{z}_j\sim \mathcal{N}(0,I)$ . + +Let us define a random variable, $D_{ij}$ to denote the distance of $i^{\text{th}}$ data point to the $j^{\text{th}}$ sample. We can also define another random variable $D_i^*$ to denote the distance of $i^{\text{th}}$ data point to the sample closest to it. Further, let $F_{D_{ij}}$ and $F_{D_i^*}$ be the CDF of $D_{ij}$ and $D_i^*$ respectively. Let $f_{D_{ij}}$ and $f_{D_i^*}$ be the PDF of $D_{ij}$ and $D_i^*$ respectively. Now we will relate the CDF of the distances between a data point and its selected latent code, $F_{D_i^*}$ to the CDF of the distances between the same data point and a random latent code $F_{D_{ij}}$ : + +$$ +\begin{array}{l} F _ {D _ {i} ^ {*}} (t) = \operatorname * {P r} (D _ {i} ^ {*} \leq t) = 1 - \operatorname * {P r} (D _ {i} ^ {*} > t) \\ = 1 - \Pr (D _ {i j} > t, \forall j \in [ m ]) \quad (\text {D e f . o f} D _ {i j}) \\ = 1 - \prod_ {j = 1} ^ {m} \Pr (D _ {i j} > t) \\ = 1 - \left(\Pr \left(D _ {i 1} > t\right)\right) ^ {m} (2) \\ = 1 - \left(1 - F _ {D _ {i 1}} (t)\right) ^ {m} (3) \\ \end{array} +$$ + +Note that Equation 2 is true because each $\mathbf{z}_j$ is drawn independently from the same probability distribution which makes $D_{i1}, D_{i2} \dots D_{im}$ identical in distribution for a particular data point $\mathbf{x}_i$ . + +![](images/94f9f3c5429cc4a73e11fe915de277d47f4a283be323696eb592db9526c3a622.jpg) +(a) PDF of example distribution + +![](images/c196852dd0bd7dd0f1cda40ab503bbfaa75ffc7454980970d768886e4d737e68.jpg) +Fig. 3: Illustrative figure for demonstrating the behaviour of $F_{D_i^*}(t)$ and $F_{D_{i1}}(t)$ using Noncentral Chi-squared distribution as the example distribution. + +![](images/434e8a8b691bba4e02e829fbdd5ee16b3b257ce536af00347c43667e993208f9.jpg) +(b) $F_{D_i^*}(t)$ vs $t$ for different values of $m$ +(c) $F_{D_i^*}(t)$ vs $F_{D_{i1}}(t)$ for different values of $m$ + +Now we can try to justify our observations from Figure 2a. Equation 3 shows us how the CDF of the distance of the selected latent code (used in training) differs significantly from the CDF of distance of a random latent code (encountered at test time). This shows that the distance of the sample we choose for training is lower than the distance of the sample at testing. We can obtain a deeper understanding by analyzing the plots in Figure 3 for an example distribution. Notice that $\forall m > 1, F_{D_i^*}(t) > F_{D_{i1}}(t)$ . Observing Equation 3 and the aforementioned plots, we can conclude the following things: + +1. The $f_{D_i^*}$ is skewed towards the origin compared to $f_{D_{ij}}$ . This is intuitive because the latent codes are selected by the min operation and so their distance to their respective data point would be less than that of a random sample. +2. The skew towards the data point increases as $m$ increases. This piece of information will be important shortly. + +We can also compute the PDF $f_{D_i^*}$ in terms of $f_{D_{i1}}$ by differentiating the CDF $F_{D_i^*}(t)$ as follows: + +$$ +f _ {D _ {i} ^ {*}} (t) = \frac {\mathrm {d} F _ {D _ {i} ^ {*}} (t)}{\mathrm {d} t} = m \left(1 - F _ {D _ {i 1}} (t)\right) ^ {m - 1} f _ {D _ {i 1}} (t) \tag {4} +$$ + +# 3.4 Solving the Misalignment Issue + +Now that we know the reason behind the misalignment between latent codes used at training and testing, we wish to come up with a method to mitigate this phenomenon. Recall that $D_{ij}$ and $D_i^*$ are determined by the distance function $d(\cdot, \cdot)$ , neural network $T_\theta$ and the prior distribution. For a given generative modelling task, it is not trivial to change the $d(\cdot, \cdot)$ or $T_\theta$ . Hence, in this paper, we aim to change the prior distribution used at training such that the distribution of latent codes at training time closely matches with the distribution of latent codes (drawn from the standard normal distribution) encountered at test time. Notice that the IMLE objective (Equation 1) allows us to sample from the prior without knowing the closed form expression for its probability density + +Algorithm 1 RS-IMLE Procedure +Require: The set of inputs $\{\mathbf{x}_i\}_{i = 1}^n$ , radius $\epsilon$ +1: Initialize the parameters $\theta$ of the generator $T_{\theta}$ +2: for $k = 1$ to $K$ do +3: Draw latent codes $Z\gets \mathbf{z}_1,\dots,\mathbf{z}_m$ from $\mathcal{N}(0,\mathbf{I})$ +4: Compute $\tilde{Z}\gets \tilde{\mathbf{z}}_1,\ldots ,\tilde{\mathbf{z}}_p$ from $Z$ such that $d\left(\mathbf{x}_i,T_\theta (\tilde{\mathbf{z}}_j)\right)\geq \epsilon ,\quad \forall \tilde{\mathbf{z}}_j\in \tilde{Z},i\in [n]$ +5: $\sigma (i)\gets \arg \min_{j\in [m]}d(\mathbf{x}_i,T_\theta (\tilde{\mathbf{z}}_j)),\quad \forall i\in [n]$ +6: for $l = 1$ to $L$ do +7: Pick a random batch $S\subseteq [n]$ +8: $\theta \gets \theta -\eta \nabla_{\theta}\left(\sum_{i\in S}d\left(\mathbf{x}_i,T_\theta \left(\tilde{\mathbf{z}}_{\sigma (i)}\right)\right)\right) / |S|$ +9: end for +10: end for +11: return $\theta$ + +function. This allows us the flexibility of choosing a non-analytical prior distribution. To distinguish from $\mathbf{z}_j\sim \mathcal{N}(0,I)$ , we will use $\tilde{\mathbf{z}}$ to denote the latent codes drawn from our desired target distribution $\mathcal{P}$ . Identical to the previous section, let us define random variable $\tilde{D}_{ij} = d\left(\mathbf{x}_i,T_\theta (\tilde{\mathbf{z}}_j)\right)$ to denote the distance between data point $\mathbf{x}_i$ from a random sample $T_{\theta}(\tilde{\mathbf{z}}_j)$ . Similarly, we define $\tilde{D}_i^* = \min_{j\in [m]}\tilde{D}_{ij}$ . Similar to the section above, let $F_{\tilde{D}_{ij}}$ and $F_{\tilde{D}_i^*}$ be the CDF of $\tilde{D}_{ij}$ and $\tilde{D}_i^*$ respectively. Then $f_{\tilde{D}_{ij}}$ and $f_{\tilde{D}_i^*}$ would be the PDF of $\tilde{D}_{ij}$ and $\tilde{D}_i^*$ respectively. + +Designing the target prior Now, we will try to think about the properties we would want to have for our target prior. Recall that the Misalignment Issue is mitigated as we decrease the number of samples. In order to differentiate between the number of samples of different priors, we use the notation $m'$ to denote the number of samples for the objective using the Gaussian prior. As hinted in the previous section, one way to avoid the Misalignment Issue is to set $m'$ to low values. However we cannot directly use a Gaussian prior with low values of $m'$ as our target prior. This is because in aforementioned case, it happens to be that many data points are forced to pick the same sample as their nearest neighbour. This in turn causes very slow or no learning. Hence when using a Gaussian prior, we need to pick $m'$ large enough to allow convergence and yet small enough such that the Misalignment Issue does not affect test time performance. In our case, we are trying to design a new distribution that solves the Misalignment Issue by having desirable properties of an ideal distribution. The ideal distribution is a Gaussian prior with $m'$ set to the lowest possible value, which is $m' = n$ . + +Now that we have found the ideal prior, we can make $\mathcal{P}$ and the ideal prior to be equal in distribution. Similar to the analysis till Equation 4, we derive the PDF of $\mathcal{P}$ . We get $f_{\tilde{D}_i^*}(t) = m\left(1 - F_{\tilde{D}_{i1}}(t)\right)^{m - 1}f_{\tilde{D}_{i1}}(t)$ . Equating this PDF of $\mathcal{P}$ to the PDF of the ideal distribution gives: + +$$ +\begin{array}{l} m \left(1 - F _ {\tilde {D} _ {i 1}} (t)\right) ^ {m - 1} f _ {\tilde {D} _ {i 1}} (t) = n \left(1 - F _ {D _ {i 1}} (t)\right) ^ {n - 1} f _ {D _ {i 1}} (t) \\ \Rightarrow f _ {\tilde {D} _ {i 1}} (t) = \frac {n}{m} \frac {\left(1 - F _ {D _ {i 1}} (t)\right) ^ {n - 1}}{\left(1 - F _ {\tilde {D} _ {i 1}} (t)\right) ^ {m - 1}} f _ {D _ {i 1}} (t) (5) \\ \Longrightarrow f _ {\tilde {D} _ {i 1}} (t) = c \phi (t) f _ {D _ {i 1}} (t) (6) \\ \end{array} +$$ + +Here, $\phi(t)$ is introduced to simplify notation and $c$ is a constant associated with truncating $F_{D_{i1}}(t)$ that is needed to make sure $f_{\tilde{D}_{i1}}(t)$ is bounded. We discuss these technical details in the Appendix. + +Rejection Sampling We have expressed our target prior $\mathcal{P}$ in terms of distribution from which we can easily sample. Now, we can use Rejection Sampling to sample from our target prior $\mathcal{P}$ . To be concrete: $f_{\tilde{D}_{i1}^*}(t)$ is our target distribution and $f_{D_{i1}}(t)$ acts as our proposal distribution, since we can sample from the standard Gaussian easily. We can further write it in the standard Rejection Sampling notation: $\frac{f_{\tilde{D}_{i1}^*}(t)}{M f_{D^*}(t)} = \frac{c\phi(t)}{M}$ + +Here, $M$ is the scaling factor associated with Rejection Sampling. We approximate the function above using a step function. The step needs to happen at $t$ where $F_{\tilde{D}_{i1}}(t)$ gets close to 1. Instead of trying to estimate $F_{\tilde{D}_{i1}}(t)$ , we instead use a hyperparameter $\epsilon$ to represent where $F_{\tilde{D}_{i1}}(t)$ gets close to 1. We find the value of this hyperparameter $\epsilon$ by cross-validation. + +The final procedure simplifies to this: we sample $\mathbf{z} \sim \mathcal{N}(0, I)$ . If $t = d\left(\mathbf{x}_i, T_\theta(\mathbf{z})\right) < \epsilon$ , we reject the sample; otherwise, we accept it. Since the sampling procedure is based on Rejection Sampling, we call our method RS-IMLE. The resulting RS-IMLE procedure is included in Algorithm-1. + +# 3.5 Intuitive Interpretation of the Algorithm Behaviour + +Prior to proceeding to the Implementation details, gaining a gradient-based understanding of our new objective would provide valuable insight. Let us revisit the IMLE objective (Equation 1) again. As training progresses, we expect the loss to reduce such that: $\forall x_{i},\quad \mathbb{E}_{z_{1},\ldots ,z_{m}\sim \mathcal{N}(0,I)}\min_{j\in [m]}d\left(\mathbf{x}_{i},T_{\theta}(\mathbf{z}_{j})\right)\to 0.$ As the objective approaches convergence, the loss will decrease, resulting in gradients with lower magnitude. This causes smaller updates to the model parameters during training, leading to slower progress. Note that this loss is with respect to the closest sample to each data point. Thus it can be the case that even after a lot of training, although the closest sample are pretty close to their respective data point, the rest of samples are pretty far away. + +$$ +\theta_ {\mathrm {R S - I M L E}} = \arg \min _ {\theta} \mathbb {E} _ {z _ {1}, \dots , z _ {m} \sim \mathcal {P}} \left[ \sum_ {i = 1} ^ {n} \min _ {j \in [ m ]} d \left(\mathbf {x} _ {i}, T _ {\theta} (\mathbf {z} _ {j})\right) \right] \tag {7} +$$ + +![](images/903da0e53e211d41f5c74ca952d7d4c34757878cb0b101ecf10b99f6e0909ed6.jpg) +(a) IMLE after epoch 100 + +![](images/3987190656c0d473856a93f2aff5f30c29a1f264edc827aa34b32ad5a33cd6bf.jpg) +(b) RS-IMLE after epoch 100 + +![](images/1d3d46be4d8958affaab93487277a459c26224141e1668a1b0b88c912695bb44.jpg) +(c) IMLE after epoch 500 + +![](images/849efd1263ad76742fa1b334afcd086d74b89d0c3e2b0ba6312e80dfb278e56e.jpg) +(d) RS-IMLE after epoch 500 + +![](images/02f525b8395991748243721048bbd43e663e42adc6de52da2bd099bdc5092f2a.jpg) +(e) IMLE after 2000 epochs + +![](images/756c96d4b94fd9f56b56094cf52a6c5053a063c60de10cab4e844dc7080cdea8.jpg) +(f) RS-IMLE after 2000 epochs +Fig. 4: Comparison between IMLE and RS-IMLE for 2D toy problem. Data points are denoted by $\blacksquare$ and samples are denoted by $\bullet$ . Samples picked as nearest neighbours are denoted by $\star$ . + +Consider our objective in Equation 7 and recall that we have constructed the probability distribution $\mathcal{P}$ such that $\forall \mathbf{x}_i\quad d\left(\mathbf{x}_i,T_\theta (\tilde{\mathbf{z}})\right)\geq \epsilon$ , where $\tilde{\mathbf{z}}\sim \mathcal{P}$ . In other words, all samples we obtain by using the prior $\mathcal{P}$ are guaranteed to be $\epsilon$ -distance away from all data points. This ensures that the loss per data point is always greater than $\epsilon$ . The approach can be interpreted as ignoring the samples that are already close to some data point and instead training on challenging, non-trivial samples. + +To compare the sampling behavior of the vanilla IMLE and the proposed RS-IMLE, we trained two models on a 2D toy problem as illustrated in Figure 4. The first model uses the vanilla IMLE objective (Equation 1), while the second model is trained with our proposed RS-IMLE objective (Equation 7). + +At the initial stage of training, both the methods learn similar distributions, indicated by the straight line of orange dots ( $\odot$ ) in Figure-4a,4b. Our method first removes all the samples that fall within an $\epsilon$ distance from any data point before doing the nearest neighbour search. We illustrate this in Figure-4b, where samples that lie within the gray circles are not considered for the nearest neighbour search. + +![](images/1a22c15fe607f29d786b2669930a04268a35754286cd8b044c23f0129575445f.jpg) +Fig. 5: Qualitative comparison between our method and baselines. While analyzing the images, look for the sharpness of each image and diversity in the content of all images for a method. + +As training progresses, we notice that for both the algorithms' samples move closer to the ground truth data points. However, for vanilla IMLE, we observe that for many data points the sample picked after the nearest neighbour search is already close to the ground truth. In this case, the loss associated with these data points would be low (indicated by the short length of arrow in the Figure-4c). As a result, the model trained by the vanilla IMLE objective does not learn anything significantly novel. In our proposed method (Figure-4d), each data point selects a sample that is at least $\epsilon$ distance away from it. This ensures that the loss for each data point is always sufficiently high (indicated by the long arrows), resulting in meaningful updates to the model parameters. + +# 3.6 Implementation Details + +Note that computing the distance of each sample with each data point is computationally expensive. Suppose we have $n$ data points in $\mathbb{R}^d$ and $m$ samples, calculating the distance between each pair has a time complexity of $\mathcal{O}(mnd)$ . To get around this issue, we leverage a fast k-nearest neighbor search method, DCI [20]. This method reduces the runtime of a single query from linear to sublinear in $m$ , enabling us to efficiently filter out all the samples that are within an $\epsilon$ distance from any data point. Subsequently, from this filtered pool of remaining samples, we select, for each data point, the sample that is closest to it. + +In order to reduce the search time complexity further, we can project the training data to a lower dimensional subspace (which would decrease $d$ ). We use this while implementing the procedure for image synthesis task: we first flatten each image and project it to a lower dimension by using random projection. We normalize these projected vectors before using them for nearest neighbour search. + +# 4 Experiments + +Datasets We assess our method and the baseline approaches across a variety of datasets with $256 \times 256$ resolution. These datasets include Animal-Face Dog [33], Animal-Face Cat [33], Obama [41], Panda [41], Grumpy-cat [41], Anime [24], Shells [24], Skulls [24] + +and a subset of Flickr-FaceHQ (FFHQ) [12] which are standard datasets used in the few-shot learning literature. + +
DatasetFastGAN [24]FakeCLR [23]FreGAN [38]ReGAN [32]AdaIMLE [1]RS-IMLE (Ours)Imp. %
Obama41.129.933.445.725.014.044.0
Grumpy Cat26.620.624.927.319.111.539.8
Panda10.08.89.012.67.63.554.0
FFHQ-10054.262.150.587.433.212.961.1
Cat35.127.431.042.124.915.936.1
Dog50.744.447.957.243.023.146.3
Anime69.877.759.8110.865.835.845.6
Skulls109.6106.5163.3130.781.951.137.6
Shells120.9148.4169.3236.1108.555.448.9
+ +Table 1: We compute FID [8] between the real data and 5000 randomly generated samples for all the methods. Lower is better. + +Baselines We compare our method to recent state-of-the-art few-shot image generation methods. These include FastGAN [24], FakeCLR [23], FreGAN [38], Re-GAN [32] and AdaIMLE [1]. + +Evaluation Metrics We employ Fréchet Inception Distance (FID) [8] to assess the perceptual quality of the generated images. This involves randomly generating 5000 images and calculating the FID between these generated samples and the real images for each dataset. Additionally, we evaluate the modelling accuracy and coverage by computing precision and recall for 1000 images using the metric defined by Kynkänniemi et al. [19]. In image synthesis, precision refers to the model's capacity to generate images closely resembling the desired target or distribution. Recall measures the model's ability to encompass a wide array of diverse images within the target distribution. To ensure that the computed metrics have low variance, we generate many more samples than training images. + +Network Architecture We construct our generator network using decoder modules from VDVAE [3]. More details about the network architecture can be found in the Appendix. + +# 4.1 Quantitative Results + +In Table 1, we present the FID scores computed for all the datasets across different methods. Lower FID scores indicate that the distribution of generated images is closer to the distribution of real images. Our method performs significantly better compared to baselines. + +In Table 2, we show the precision and recall scores. Our method has a near perfect precision (close to 1), while having a significantly higher recall compared to the baselines. In the few cases where our method is not the best, it is very close to the best metric. + +![](images/c5051ad720e597f52683eaddb7faa491c58a7db82e16ce5edebd7f67e0294f10.jpg) +Query +(a) Obama + +![](images/41f62913db116d5de28f06fdb521b7e5d6aff67f51d094d0b0427feb2a8fd2b1.jpg) +RS-IMLE + +![](images/5fb2b9c811adba02deb0be6ef649326622a2004b0006d35c89b87d9e8d09e4ba.jpg) +Ada-IMLE + +![](images/683bb6827b5df7121a87d723d805f452891b4d7b100c458023ee00921b181953.jpg) +FastGAN + +![](images/6b200923da9adf0ac846d23121ceb7f4296563c504eea63745369d123eaa2d46.jpg) +FreGAN + +![](images/6af369e0484a58c880c9d69c7e3729b83aa1de9fc25a6fccf1740210496bf656.jpg) +(b) Anime + +![](images/077d465dc3a703e95c0e7bd89ffaf91240843e4689860d57b7d1ca150e021853.jpg) + +![](images/ca64a04b7ca8cebcfd570d6e37e378c02c3266257779dcc8600c5e3f53f2484c.jpg) + +![](images/128aa9b30dbf8a0e78a685ceaded7af59c59fc5af28ea3060b6b7911d2c05970.jpg) + +![](images/cee15c41250a9d0627202729a2fb8adebf27f42166db826bd84e8bbdd43176b4.jpg) + +![](images/4d777c9d80dfb01159fcfca59e1925d711b8d11c3e889e6402a052d885e49e91.jpg) +(c) Dog + +![](images/51d2bbc677929c4ea815a88ed34099ae39df7186f2078758a21f8e99b17686bb.jpg) + +![](images/77b861d33a7080f0351f389f5c61d14af04471244e424a4c77e23cc77b89e8d0.jpg) + +![](images/5beb677c92ec43c1eb5fe27dc68e09bda5977dea02cd6df50bb14e4636d1e332.jpg) + +![](images/17b6a23af1354eeeab9e9a8dbfa8ffa4c9323f61908f58ace4d9bb77628e01dc.jpg) + +![](images/966f34838f2101c1ad03df9322d86655c2c560d801012b2d2fd123aea4a05310.jpg) +(d) Cat + +![](images/fca2f453383f06982f87b9a6f1d4b0361885f3bb882b5c861b87cd3d0089b924.jpg) + +![](images/c51433fbd18b27ef6d0e8f1021f51805d2b6afb8b5bf2d18605072d5977ff4c9.jpg) + +![](images/4ba24b36d82ca8c8b9067f52f5232714da42e42ce9d216b613498555f7dea1eb.jpg) + +![](images/973b997d5d0ca7a6cea7acc841c4b7a6695cb850eaa68a237e3a86f42e7fc427.jpg) + +![](images/12ddc4b196240dec46d43d2dcaebc1773ab93c166f89ecf9dfb3b728ef763069.jpg) + +![](images/a5630c9161fd5940b774d161a462fb257558ad4796aa5c744a7c660a68c3af64.jpg) + +![](images/647485f4f25d709f43e2586ab5b283eb83ec53a1d320f78d15f54b12101eddfb.jpg) + +![](images/58423e93f87222b529343d8c398d32b5d61060e924a78d12440fa5686326d004.jpg) + +![](images/026cd53b87b41f419b0b7eec78e20307404bb716cbb298d7c83773023010ab16.jpg) + +![](images/379eb13e8b7523712cf23c77e53fd5c61e1748faa79b44bb5f501b8830d4b155.jpg) +(e) Shells +(f) FFHQ-100 +Fig. 6: Visual Recall test. The first column is the query image from the dataset. Subsequent columns are the samples produced by different methods that are closest to the query image in LPIPS feature space. The samples produced by our method are closer to the query images compared to the baselines, while remaining diverse. + +![](images/2ec14aef5e625847b1ffcae95889948bc3adc854c24d65751048e46ef14dc396.jpg) + +![](images/8f31b4dfb65b5566bfa3c2ad96aab41c7e3e1f41019babee95e343f0bb468c92.jpg) + +![](images/57f82d76e2db2bf8b179a80f8ffc0e22f0925808bb98fc31774d70b3fb7e26d6.jpg) + +![](images/68801163711ce49b64fc92ee1cb7188a38414a5ee0da640d684f6b22e2a8c303.jpg) + +
DatasetFastGAN [24]FakeCLR [23]FreGAN [38]ReGAN [32]AdaIMLE [1]RS-IMLE (Ours)
ObamaPrec.0.920.960.820.620.990.98
Rec.0.090.300.330.010.680.82
Grumpy CatPrec.0.910.970.900.780.970.93
Rec.0.130.390.230.040.720.95
PandaPrec.0.960.970.920.930.980.99
Rec.0.160.410.170.010.630.97
FFHQ-100Prec.0.910.710.860.390.991.00
Rec.0.130.250.210.010.770.99
CatPrec.0.970.990.950.900.980.96
Rec.0.080.550.310.150.860.98
DogPrec.0.960.950.920.840.970.98
Rec.0.190.340.200.100.610.94
AnimePrec.0.860.880.880.290.920.95
Rec.0.080.090.090.010.590.91
SkullsPrec.0.780.660.660.660.950.99
Rec.0.030.010.010.010.320.65
ShellsPrec.0.920.870.870.500.970.98
Rec.0.030.010.010.010.620.59
+ +Table 2: We compute precision and recall [42] between the real data and 1000 randomly generated samples for all the methods. Higher values are better. + +# 4.2 Qualitative Results + +Figure 5 compares the random samples of our method to that of several baselines, and we observe that our method produces overall sharper and more diverse images. In addition, we propose Visual Recall, a simple test to substantiate the qualitative superiority of our method. For each method, we first generate 1000 samples. Next, using a real image from the dataset as a query, we find the images from the pool of generated samples that are closest to the query image. We use LPIPS features [39] for the computing the distance between real images and samples. Figure 6 shows the results for the proposed test for different datasets and methods. We see that the samples produced by our method are visually similar to the query, while being sharp and diverse in attributes like hair colour, smile and jaw structure. Note that other methods do not have samples that closely resemble the query image. + +Figure 7 shows results of spherical linear interpolation between two random points in the latent space for different datasets. The images transition in a meaningful manner, indicating that our model has learnt a continuous and structured latent space representation of the image distribution. + +# 4.3 Ablation Study + +Table-3 presents the FID computed for different values of $\epsilon$ for three datasets. We observe that our approach works best for values of $\epsilon$ close to 0.15 and that increasing the value of $\epsilon$ beyond a certain range degrades the performance. + +![](images/17b1ac2f0023e48788e1f9db75b7756b33d9db8e7dbba404405efade397f661b.jpg) +(c) Skulls +Fig. 7: Latent space interpolation. We observe that the output images changes smoothly in a meaningful manner. + +
Dataset0.120.150.180.22
FFHQ-10013.612.9113.0114.8
Obama14.4414.0314.6214.34
Cat17.2915.9616.1216.27
+ +Table 3: FID for different values of $\epsilon$ + +# 5 Conclusion + +In this paper, we identified a latent space misalignment between the training and testing phases of existing IMLE-based methods, resulting in poor performance in few-shot image synthesis tasks. To address this issue, we introduced a novel algorithm, RS-IMLE, which modifies the prior distribution used for training. Our experimental results demonstrate that our method significantly enhances the quality of generated images and mode coverage during inference. + +Acknowledgements: This research was enabled in part by support provided by NSERC, the BC DRI Group and the Digital Research Alliance of Canada. The authors would also like to thank Tristan Engst for extensive help polishing our paper. + +# References + +1. Aghabozorgi, M., Peng, S., Li, K.: Adaptive IMLE for few-shot pretraining-free generative modelling. In: Krause, A., Brunskill, E., Cho, K., Engelhardt, B., Sabato, S., Scarlett, J. (eds.) Proceedings of the 40th International Conference on Machine Learning. Proceedings of Machine Learning Research, vol. 202, pp. 248-264. PMLR (23-29 Jul 2023), https://proceedings.mlr.press/v202/aghabozorgi23a.html +2. Brock, A., Donahue, J., Simonyan, K.: Large scale gan training for high fidelity natural image synthesis. ArXiv abs/1809.11096 (2019) +3. Child, R.: Very deep vaes generalize autoregressive models and can outperform them on images. ArXiv abs/2011.10650 (2021) +4. Dhariwal, P., Nichol, A.: Diffusion models beat gans on image synthesis. ArXiv abs/2105.05233 (2021) +5. Dinh, L., Sohl-Dickstein, J.N., Bengio, S.: Density estimation using real nvp. ArXiv abs/1605.08803 (2017) +6. Esser, P., Rombach, R., Ommer, B.: Taming transformers for high-resolution image synthesis. 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) pp. 12868-12878 (2020), https://api_semanticscholar.org/CorpusID:229297973 +7. Goodfellow, I., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., Courville, A., Bengio, Y.: Generative adversarial nets. Advances in neural information processing systems 27 (2014) +8. Heusel, M., Ramsauer, H., Unterthiner, T., Nessler, B., Hochreiter, S.: Gans trained by a two time-scale update rule converge to a local nash equilibrium. In: NIPS (2017) +9. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. ArXiv abs/2006.11239 (2020) +0. Karras, T., Aittala, M., Laine, S., Harkonen, E., Hellsten, J., Lehtinen, J., Aila, T.: Alias-free generative adversarial networks. In: NeurIPS (2021) +1. Karras, T., Laine, S., Aila, T.: A style-based generator architecture for generative adversarial networks. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 4401-4410 (2019) +2. Karras, T., Laine, S., Aila, T.: A style-based generator architecture for generative adversarial networks (2019) +3. Karras, T., Laine, S., Aittala, M., Hellsten, J., Lehtinen, J., Aila, T.: Analyzing and improving the image quality of stylegan. 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) pp. 8107-8116 (2019), https://api_semanticscholar.org/CorpusID:209202273 +4. Karras, T., Laine, S., Aittala, M., Hellsten, J., Lehtinen, J., Aila, T.: Analyzing and improving the image quality of stylegan. 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) pp. 8107-8116 (2020) +5. Kingma, D.P., Dhariwal, P.: Glow: Generative flow with invertible 1x1 convolutions. ArXiv abs/1807.03039 (2018) +6. Kingma, D.P., Welling, M.: Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114 (2013) +7. Kobyzev, I., Prince, S., Brubaker, M.A.: Normalizing flows: An introduction and review of current methods. IEEE Transactions on Pattern Analysis and Machine Intelligence 43, 3964-3979 (2021) +8. Kong, C., Kim, J., Han, D., Kwak, N.: Few-shot image generation with mixup-based distance learning (2022) +9. Kynkänniemi, T., Karras, T., Laine, S., Lehtinen, J., Aila, T.: Improved precision and recall metric for assessing generative models. Advances in Neural Information Processing Systems 32 (2019) + +20. Li, K., Malik, J.: Fast k-nearest neighbour search via prioritized dci (2017) +21. Li, K., Malik, J.: Implicit maximum likelihood estimation. arXiv preprint arXiv:1809.09087 (2018) +22. Li, Y., Zhang, R., Lu, J., Shechtman, E.: Few-shot image generation with elastic weight consolidation. ArXiv abs/2012.02780 (2020) +23. Li, Z., Wang, C., Zheng, H., Zhang, J., Li, B.: Fakeclr: Exploring contrastive learning for solving latent discontinuity in data-efficient gans. In: ECCV (2022) +24. Liu, B., Zhu, Y., Song, K., Elgammal, A.: Towards faster and stabilized GAN training for high-fidelity few-shot image synthesis. CoRR abs/2101.04775 (2021), https://arxiv.org/abs/2101.04775 +25. Mo, S., Cho, M., Shin, J.: Freeze discriminator: A simple baseline for fine-tuning gans. ArXiv abs/2002.10964 (2020) +26. Ojha, U., Li, Y., Lu, J., Efros, A.A., Lee, Y.J., Shechtman, E., Zhang, R.: Few-shot image generation via cross-domain correspondence. 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) pp. 10738-10747 (2021) +27. van den Oord, A., Kalchbrenner, N., Espeholt, L., Kavukcuoglu, K., Vinyals, O., Graves, A.: Conditional image generation with pixelCNN decoders. ArXiv abs/1606.05328 (2016) +28. van den Oord, A., Kalchbrenner, N., Kavukcuoglu, K.: Pixel recurrent neural networks. ArXiv abs/1601.06759 (2016) +29. Razavi, A., van den Oord, A., Vinyals, O.: Generating diverse high-fidelity images with vqvae-2. ArXiv abs/1906.00446 (2019) +30. Salimans, T., Karpathy, A., Chen, X., Kingma, D.P.: PixelCNN++: Improving the pixelCNN with discretized logistic mixture likelihood and other modifications. ArXiv abs/1701.05517 (2017) +31. Sauer, A., Chitta, K., Muller, J., Geiger, A.: Projected gans converge faster. In: Neural Information Processing Systems (2021), https://api-semanticscholar.org/CorpusID: 240354401 +32. Saxena, D., Cao, J., Xu, J., Kulshrestha, T.: Re-gan: Data-efficient gans training via architectural reconfiguration. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 16230-16240 (June 2023) +33. Si, Z., Zhu, S.C.: Learning hybrid image templates (hit) by information projection. IEEE Transactions on Pattern Analysis and Machine Intelligence 34, 1354-1367 (2012) +34. Song, Y., Ermon, S.: Generative modeling by estimating gradients of the data distribution. ArXiv abs/1907.05600 (2019) +35. Song, Y., Sohl-Dickstein, J.N., Kingma, D.P., Kumar, A., Ermon, S., Poole, B.: Score-based generative modeling through stochastic differential equations. ArXiv abs/2011.13456 (2021) +36. Vahdat, A., Kautz, J.: Nvae: A deep hierarchical variational autoencoder. ArXiv abs/2007.03898 (2020) +37. Wang, Y., Gonzalez-Garcia, A., Berga, D., Herranz, L., Khan, F.S., van de Weijer, J.: Minegan: Effective knowledge transfer from gans to target domains with few images. 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) pp. 9329-9338 (2020) +38. Yang, M., Wang, Z., Chi, Z., Zhang, Y.: Fregan: Exploiting frequency components for training gans under limited data (2022) +39. Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual metric. In: CVPR (2018) +40. Zhao, M., Cong, Y., Carin, L.: On leveraging pretrained gans for generation with limited data. In: ICML (2020) +41. Zhao, S., Liu, Z., Lin, J., Zhu, J.Y., Han, S.: Differentiable augmentation for data-efficient gan training. ArXiv abs/2006.10738 (2020) + +42. Zhu, P., Abdal, R., Qin, Y., Wonka, P.: Improved stylegan embedding: Where are the good latents? ArXiv abs/2012.09036 (2020) \ No newline at end of file diff --git a/2024/Rejection Sampling IMLE_ Designing Priors for Better Few-Shot Image Synthesis/images.zip b/2024/Rejection Sampling IMLE_ Designing Priors for Better Few-Shot Image Synthesis/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..cee39e637f3e8ba632ad0509b751d7af55954e40 --- /dev/null +++ b/2024/Rejection Sampling IMLE_ Designing Priors for Better Few-Shot Image Synthesis/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:99aeb6568d376586fdccd36260bc16412d0150c84622b0745b8d955c02f2b7de +size 852958 diff --git a/2024/Rejection Sampling IMLE_ Designing Priors for Better Few-Shot Image Synthesis/layout.json b/2024/Rejection Sampling IMLE_ Designing Priors for Better Few-Shot Image Synthesis/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..acd2adbc66b1b004bbe6a32a087b8dd12c85a313 --- /dev/null +++ b/2024/Rejection Sampling IMLE_ Designing Priors for Better Few-Shot Image Synthesis/layout.json @@ -0,0 +1,11783 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 138, + 111, + 477, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 111, + 477, + 148 + ], + "spans": [ + { + "bbox": [ + 138, + 111, + 477, + 148 + ], + "type": "text", + "content": "Rejection Sampling IMLE: Designing Priors for Better Few-Shot Image Synthesis" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 211, + 168, + 402, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 211, + 168, + 402, + 180 + ], + "spans": [ + { + "bbox": [ + 211, + 168, + 402, + 180 + ], + "type": "text", + "content": "Chirag Vashist, Shichong Peng, and Ke Li" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 286, + 190, + 328, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 286, + 190, + 328, + 200 + ], + "spans": [ + { + "bbox": [ + 286, + 190, + 328, + 200 + ], + "type": "text", + "content": "APEX Lab" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 252, + 202, + 361, + 212 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 252, + 202, + 361, + 212 + ], + "spans": [ + { + "bbox": [ + 252, + 202, + 361, + 212 + ], + "type": "text", + "content": "School of Computing Science" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 261, + 213, + 351, + 223 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 261, + 213, + 351, + 223 + ], + "spans": [ + { + "bbox": [ + 261, + 213, + 351, + 223 + ], + "type": "text", + "content": "Simon Fraser University" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 203, + 224, + 411, + 234 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 203, + 224, + 411, + 234 + ], + "spans": [ + { + "bbox": [ + 203, + 224, + 411, + 234 + ], + "type": "text", + "content": "{chirag_vashist, shichong_peng, keli}@sfu.ca" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 159, + 261, + 455, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 261, + 455, + 392 + ], + "spans": [ + { + "bbox": [ + 159, + 261, + 455, + 392 + ], + "type": "text", + "content": "Abstract. An emerging area of research aims to learn deep generative models with limited training data. Implicit Maximum Likelihood Estimation (IMLE), a recent technique, successfully addresses the mode collapse issue of GANs and has been adapted to the few-shot setting, achieving state-of-the-art performance. However, current IMLE-based approaches encounter challenges due to inadequate correspondence between the latent codes selected for training and those drawn during inference. This results in suboptimal test-time performance. We theoretically show a way to address this issue and propose RS-IMLE, a novel approach that changes the prior distribution used for training. This leads to substantially higher quality image generation compared to existing GAN and IMLE-based methods, as validated by comprehensive experiments conducted on nine few-shot image datasets." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 159, + 403, + 453, + 424 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 403, + 453, + 424 + ], + "spans": [ + { + "bbox": [ + 159, + 403, + 453, + 424 + ], + "type": "text", + "content": "Keywords: Few-shot " + }, + { + "bbox": [ + 159, + 403, + 453, + 424 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 403, + 453, + 424 + ], + "type": "text", + "content": " Image Synthesis " + }, + { + "bbox": [ + 159, + 403, + 453, + 424 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 403, + 453, + 424 + ], + "type": "text", + "content": " Implicit Maximum Likelihood Estimation" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 462, + 218, + 474 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 462, + 218, + 474 + ], + "spans": [ + { + "bbox": [ + 132, + 462, + 218, + 474 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 486, + 482, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 486, + 482, + 629 + ], + "spans": [ + { + "bbox": [ + 130, + 486, + 482, + 629 + ], + "type": "text", + "content": "Recent years have witnessed significant advancements in image synthesis, driven by the emergence of deep learning and the development of a broad variety of powerful generative models. Generative Adversarial Networks (GANs) [2,7,10,11,14], Variational Autoencoders (VAEs) [3,16,29,36], Diffusion Models [4,9], Score-based Models [34,35], Normalizing Flows [5, 15, 17], and Autoregressive Models [6, 27, 28, 30] have demonstrably improved the quality of synthesized images, often achieving photorealism. Quite often, generative models often require large amounts of training data to achieve this high fidelity. In numerous situations, obtaining large volumes of training data is impractical. For instance, when we want to generate images of rare species of plants or images of anatomical structures with rare diseases, access to abundant images may be limited. In this paper, we aim to tackle the problem of high quality image synthesis using limited training data." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 630, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 630, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 630, + 482, + 665 + ], + "type": "text", + "content": "In the few-shot setting, most of the leading methods to date are based on GANs [18, 23, 24, 32, 38]. Despite advances, these methods continue to be afflicted by mode collapse, a well-known issue in GANs. Mode collapse occurs when the generator network" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 134, + 114, + 228, + 196 + ], + "blocks": [ + { + "bbox": [ + 134, + 114, + 228, + 196 + ], + "lines": [ + { + "bbox": [ + 134, + 114, + 228, + 196 + ], + "spans": [ + { + "bbox": [ + 134, + 114, + 228, + 196 + ], + "type": "image", + "image_path": "be654b3d05c083e881e770900f841c83b621234de8e6a1ca159fc77b274686f6.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 238, + 125, + 403, + 196 + ], + "blocks": [ + { + "bbox": [ + 238, + 125, + 403, + 196 + ], + "lines": [ + { + "bbox": [ + 238, + 125, + 403, + 196 + ], + "spans": [ + { + "bbox": [ + 238, + 125, + 403, + 196 + ], + "type": "image", + "image_path": "d0f25a73ced6cd7e287d574d3316d1d865e7807ec2dfac233f0e81e50543e202.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 406, + 125, + 470, + 197 + ], + "blocks": [ + { + "bbox": [ + 406, + 125, + 470, + 197 + ], + "lines": [ + { + "bbox": [ + 406, + 125, + 470, + 197 + ], + "spans": [ + { + "bbox": [ + 406, + 125, + 470, + 197 + ], + "type": "image", + "image_path": "a27541844bbc12b215300afd4e1935e0c03a4ddd066d3d47fb71e1181d2f6a3a.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 135, + 202, + 228, + 287 + ], + "blocks": [ + { + "bbox": [ + 135, + 202, + 228, + 287 + ], + "lines": [ + { + "bbox": [ + 135, + 202, + 228, + 287 + ], + "spans": [ + { + "bbox": [ + 135, + 202, + 228, + 287 + ], + "type": "image", + "image_path": "53f47ba949c401a702e6504df83b2bae5e042cd80fdbda7903c9bd8cabe48b47.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 297, + 482, + 365 + ], + "lines": [ + { + "bbox": [ + 130, + 297, + 482, + 365 + ], + "spans": [ + { + "bbox": [ + 130, + 297, + 482, + 365 + ], + "type": "text", + "content": "Fig. 1: IMLE is an implicit generative model that maps a latent code sampled from a prior distribution to an image output. In previous IMLE-based methods, both the training and testing phases adopt a standard normal distribution as the prior distribution. However, this approach often results in poor generalization during inference. To address this limitation, we introduce RS-IMLE, which uses rejection sampling to alter the prior distribution used for training to a different distribution " + }, + { + "bbox": [ + 130, + 297, + 482, + 365 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 130, + 297, + 482, + 365 + ], + "type": "text", + "content": ". This modification significantly enhances the quality of generated images during testing." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 238, + 202, + 403, + 287 + ], + "blocks": [ + { + "bbox": [ + 238, + 202, + 403, + 287 + ], + "lines": [ + { + "bbox": [ + 238, + 202, + 403, + 287 + ], + "spans": [ + { + "bbox": [ + 238, + 202, + 403, + 287 + ], + "type": "image", + "image_path": "199fdeded655984a06003c96c65e6b5bed6adf14cc147ce66b1eb3f6bef728c7.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 406, + 202, + 469, + 288 + ], + "blocks": [ + { + "bbox": [ + 406, + 202, + 469, + 288 + ], + "lines": [ + { + "bbox": [ + 406, + 202, + 469, + 288 + ], + "spans": [ + { + "bbox": [ + 406, + 202, + 469, + 288 + ], + "type": "image", + "image_path": "73a107811eae489598567b2572fdd8ea6deb35b3c3c0547008742dfb1963daec.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 388, + 482, + 425 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 388, + 482, + 425 + ], + "spans": [ + { + "bbox": [ + 130, + 388, + 482, + 425 + ], + "type": "text", + "content": "fails to capture the full training data distribution and instead produces a limited subset of outputs. This phenomenon is especially problematic in scenarios where only a small number of training examples are provided." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 426, + 482, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 426, + 482, + 510 + ], + "spans": [ + { + "bbox": [ + 130, + 426, + 482, + 510 + ], + "type": "text", + "content": "Implicit Maximum Likelihood Estimation (IMLE) [21] is an alternative to the GAN objective and has shown promising results in addressing mode collapse. In contrast to GANs, which aim to make each generated image resemble some training data, IMLE instead ensures that each training image has some generated sample close to it, and therefore cannot drop any of the modes present in the training data. Adaptive IMLE [1] further extends IMLE to the few-shot image synthesis setting and achieves state-of-the-art generated image quality and mode coverage." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 510, + 483, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 483, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 483, + 666 + ], + "type": "text", + "content": "However, in existing IMLE-based approaches, we observe that the latent codes used during training and those sampled during testing have different distributions, even though the same prior is used during training and inference. This phenomenon arises because of how IMLE selects latent codes during training. Some regions of the latent space are consistently rarely picked for training, despite having a high likelihood under the prior distribution (often the standard Gaussian). This is illustrated in Fig. 2a. Consequently, at test time, when latent codes drawn from the prior happen to fall in these regions, they yield low-quality samples that are far from the real data points, as illustrated in Fig. 1. We address this issue by carefully choosing a different prior so that the samples selected for training have a distribution more similar to those sampled at inference. Our method, which we call Rejection Sampling IMLE or RS-IMLE for short, demonstrably improves coverage of the latent space used during training, thereby ensuring better alignment with the prior as shown in Fig. 2b. As a result, our method yields higher quality" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 224, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 224, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 224, + 101 + ], + "type": "text", + "content": "C. Vashist et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 134, + 114, + 294, + 267 + ], + "blocks": [ + { + "bbox": [ + 134, + 114, + 294, + 267 + ], + "lines": [ + { + "bbox": [ + 134, + 114, + 294, + 267 + ], + "spans": [ + { + "bbox": [ + 134, + 114, + 294, + 267 + ], + "type": "image", + "image_path": "aa9c3ab7230d3eb9cf00e8be247a3420fedf3f9fa17a3cae17c24e65e4d3275b.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 270, + 295, + 320 + ], + "lines": [ + { + "bbox": [ + 132, + 270, + 295, + 320 + ], + "spans": [ + { + "bbox": [ + 132, + 270, + 295, + 320 + ], + "type": "text", + "content": "(a) Latent space of model trained by IMLE objective using standard normal prior. Dots represent the points selected by the model over the course of training, with dots of the same colour belonging to the same data-point. The contours of standard normal distribution have been shown for comparison." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 130, + 328, + 482, + 352 + ], + "lines": [ + { + "bbox": [ + 130, + 328, + 482, + 352 + ], + "spans": [ + { + "bbox": [ + 130, + 328, + 482, + 352 + ], + "type": "text", + "content": "Fig. 2: Difference between the latent codes picked by IMLE and RS-IMLE over the course of training." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 319, + 114, + 480, + 267 + ], + "blocks": [ + { + "bbox": [ + 319, + 114, + 480, + 267 + ], + "lines": [ + { + "bbox": [ + 319, + 114, + 480, + 267 + ], + "spans": [ + { + "bbox": [ + 319, + 114, + 480, + 267 + ], + "type": "image", + "image_path": "12476c053cfd54019576c9267b61c255a1aac96decaeb53cf9d2ca080474c9fa.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 318, + 270, + 482, + 312 + ], + "lines": [ + { + "bbox": [ + 318, + 270, + 482, + 312 + ], + "spans": [ + { + "bbox": [ + 318, + 270, + 482, + 312 + ], + "type": "text", + "content": "(b) Latent space of model trained by RS-IMLE objective using prior obtained via rejection sampling. Compared to latent space of model trained by IMLE, our method over the course of training, samples latent codes that follow the distribution at test time more faithfully." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_footnote" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 376, + 482, + 424 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 376, + 482, + 424 + ], + "spans": [ + { + "bbox": [ + 130, + 376, + 482, + 424 + ], + "type": "text", + "content": "samples during testing compared to existing GAN and IMLE-based methods. We substantiate this claim through theoretical analysis and extensive experiments conducted on nine few-shot image datasets. We achieve an average of " + }, + { + "bbox": [ + 130, + 376, + 482, + 424 + ], + "type": "inline_equation", + "content": "45.9\\%" + }, + { + "bbox": [ + 130, + 376, + 482, + 424 + ], + "type": "text", + "content": " decrease in FID [8] across datasets compared to the best baseline." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 445, + 224, + 458 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 445, + 224, + 458 + ], + "spans": [ + { + "bbox": [ + 132, + 445, + 224, + 458 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 474, + 482, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 474, + 482, + 668 + ], + "spans": [ + { + "bbox": [ + 130, + 474, + 482, + 668 + ], + "type": "text", + "content": "Training deep generative models with limited data remains a significant challenge. One approach involves adapting a model pretrained on a large-scale auxiliary dataset from similar domains [22,25,26,31,37,40]. However, the availability of such large-scale auxiliary datasets across all domains is not guaranteed. Therefore, another emerging line of work focuses on training models from scratch. In this context, due to the scarcity of training data, diffusion models struggle to achieve high-quality generated images and have been demonstrated to be ineffective [1]. As a result, previous works in this area predominantly build on Generative Adversarial Networks (GANs) and design various methods to address the well-known mode collapse issue. Techniques such as ADA [13] and DiffAug [41] aim to expand training data using adaptive and differentiable augmentation strategies. FastGAN [24] introduced a skip-layer excitation module for accelerated training and used self-supervision in the discriminator to enhance feature learning, thereby improving mode coverage of the generator. FakeCLR [23] enhances image synthesis by extensive data augmentation and applies contrastive learning solely on perturbed fake samples. FreGAN [38] introduces a frequency-aware model with a self-supervised constraint to avoid generating arbitrary frequency signals. ReGAN [32] dynamically adjusts" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 350, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 350, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 350, + 91, + 447, + 102 + ], + "type": "text", + "content": "Rejection Sampling IMLE" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 99 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 99 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 99 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 151 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 151 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 151 + ], + "type": "text", + "content": "GANs' architecture during training to explore diverse sub-network structures at different training times. However, despite these advancements, some degree of mode collapse persists." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 152, + 482, + 248 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 152, + 482, + 248 + ], + "spans": [ + { + "bbox": [ + 130, + 152, + 482, + 248 + ], + "type": "text", + "content": "In contrast, Implicit Maximum Likelihood Estimation (IMLE) [21] shows promising results in addressing mode collapse through the use of an alternative objective function compared to GANs. Building upon IMLE, Adaptive IMLE [1] adapts this approach to the few-shot image synthesis scenario by introducing individual target thresholds for each training data point. This dynamic adjustment of training progress accounts for varying difficulties across different data points, thereby effectively leveraging the limited training data. In this work, we introduce a novel algorithm, orthogonal to Adaptive IMLE, for sample selection during training." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 267, + 194, + 280 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 267, + 194, + 280 + ], + "spans": [ + { + "bbox": [ + 131, + 267, + 194, + 280 + ], + "type": "text", + "content": "3 Method" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 294, + 210, + 306 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 294, + 210, + 306 + ], + "spans": [ + { + "bbox": [ + 131, + 294, + 210, + 306 + ], + "type": "text", + "content": "3.1 Background" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 315, + 482, + 458 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 315, + 482, + 458 + ], + "spans": [ + { + "bbox": [ + 130, + 315, + 482, + 458 + ], + "type": "text", + "content": "In the context of unconditional image synthesis, the primary objective is to learn the unconditional probability distribution of images " + }, + { + "bbox": [ + 130, + 315, + 482, + 458 + ], + "type": "inline_equation", + "content": "p(\\mathbf{x})" + }, + { + "bbox": [ + 130, + 315, + 482, + 458 + ], + "type": "text", + "content": ". This distribution enables the generation of novel synthesized images through the sampling. Implicit generative models are commonly employed for this task, with the generator in GANs serving as a notable example. Generator in GANs are represented by a function " + }, + { + "bbox": [ + 130, + 315, + 482, + 458 + ], + "type": "inline_equation", + "content": "T_{\\theta}: Z \\to X" + }, + { + "bbox": [ + 130, + 315, + 482, + 458 + ], + "type": "text", + "content": ", implemented as a neural network with parameters denoted as " + }, + { + "bbox": [ + 130, + 315, + 482, + 458 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 130, + 315, + 482, + 458 + ], + "type": "text", + "content": ". The function " + }, + { + "bbox": [ + 130, + 315, + 482, + 458 + ], + "type": "inline_equation", + "content": "T_{\\theta}" + }, + { + "bbox": [ + 130, + 315, + 482, + 458 + ], + "type": "text", + "content": " learns a transformation from the latent space " + }, + { + "bbox": [ + 130, + 315, + 482, + 458 + ], + "type": "inline_equation", + "content": "Z" + }, + { + "bbox": [ + 130, + 315, + 482, + 458 + ], + "type": "text", + "content": " to the image space " + }, + { + "bbox": [ + 130, + 315, + 482, + 458 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 130, + 315, + 482, + 458 + ], + "type": "text", + "content": " by using adversarial training, which employs a discriminator that tries to distinguish between generated images " + }, + { + "bbox": [ + 130, + 315, + 482, + 458 + ], + "type": "inline_equation", + "content": "T_{\\theta}(\\mathbf{z})" + }, + { + "bbox": [ + 130, + 315, + 482, + 458 + ], + "type": "text", + "content": " and real images " + }, + { + "bbox": [ + 130, + 315, + 482, + 458 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 130, + 315, + 482, + 458 + ], + "type": "text", + "content": ", while the generator tries to produce increasingly realistic images to deceive the discriminator. However, this objective often leads to mode collapse, a well-known issue of GANs, where the generated output " + }, + { + "bbox": [ + 130, + 315, + 482, + 458 + ], + "type": "inline_equation", + "content": "T_{\\theta}(\\mathbf{z})" + }, + { + "bbox": [ + 130, + 315, + 482, + 458 + ], + "type": "text", + "content": " only models a subset of the training examples." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 460, + 482, + 530 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 460, + 482, + 530 + ], + "spans": [ + { + "bbox": [ + 130, + 460, + 482, + 530 + ], + "type": "text", + "content": "To address the issue of mode collapse, an alternative method Implicit Maximum Likelihood Estimation (IMLE) [21] has been introduced. While IMLE, like GANs, uses a generator, it differs from GANs by using an alternative objective that aims to make each generated sample similar to some training data. Instead, IMLE ensures that each training data point has similar generated samples, thereby encouraging coverage of all the modes of the training data." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 146, + 531, + 443, + 543 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 531, + 443, + 543 + ], + "spans": [ + { + "bbox": [ + 146, + 531, + 443, + 543 + ], + "type": "text", + "content": "The IMLE objective is given as follows, where " + }, + { + "bbox": [ + 146, + 531, + 443, + 543 + ], + "type": "inline_equation", + "content": "d(\\cdot ,\\cdot)" + }, + { + "bbox": [ + 146, + 531, + 443, + 543 + ], + "type": "text", + "content": " is a distance metric:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 180, + 552, + 481, + 586 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 180, + 552, + 481, + 586 + ], + "spans": [ + { + "bbox": [ + 180, + 552, + 481, + 586 + ], + "type": "interline_equation", + "content": "\\theta_ {\\mathrm {I M L E}} = \\arg \\min _ {\\theta} \\mathbb {E} _ {z _ {1}, \\dots , z _ {m} \\sim \\mathcal {N} (0, I)} \\left[ \\sum_ {i = 1} ^ {n} \\min _ {j \\in [ m ]} d \\left(\\mathbf {x} _ {i}, T _ {\\theta} (\\mathbf {z} _ {j})\\right) \\right] \\tag {1}", + "image_path": "5fc59b246a0862718a9fb27b98303c786cb43147dd770b54d6cb6e7f0567f317.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": "Here " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": " denotes the number of samples and " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": " denotes the number of data points. In simple terms, the IMLE objective first draws " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": " samples " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_j" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": " from the standard Gaussian distribution and transforms them into the image space using the function " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "T_{\\theta}" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": ". From these pool of samples in the image space, for each data point " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "x_i" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": ", IMLE selects a sample that is closest to the data point in some distance metric " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "d(\\cdot, \\cdot)" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": ". This operation can be done efficiently due to advances in high-dimensional nearest neighbour search [20]. Note that" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 223, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 223, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 223, + 101 + ], + "type": "text", + "content": "C. Vashist et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 131, + 116, + 482, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 116, + 482, + 163 + ], + "spans": [ + { + "bbox": [ + 131, + 116, + 482, + 163 + ], + "type": "text", + "content": "the number of samples " + }, + { + "bbox": [ + 131, + 116, + 482, + 163 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 131, + 116, + 482, + 163 + ], + "type": "text", + "content": " must at least be equal to the number of data points " + }, + { + "bbox": [ + 131, + 116, + 482, + 163 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 131, + 116, + 482, + 163 + ], + "type": "text", + "content": ", since otherwise by the pigeonhole principle, some samples would be picked by multiple data points. In practice, we find that setting " + }, + { + "bbox": [ + 131, + 116, + 482, + 163 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 131, + 116, + 482, + 163 + ], + "type": "text", + "content": " to be a multiplicative factor (like 10 or 20) times larger than " + }, + { + "bbox": [ + 131, + 116, + 482, + 163 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 131, + 116, + 482, + 163 + ], + "type": "text", + "content": " works the best." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 181, + 211, + 192 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 181, + 211, + 192 + ], + "spans": [ + { + "bbox": [ + 132, + 181, + 211, + 192 + ], + "type": "text", + "content": "3.2 Observation" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 201, + 482, + 344 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 201, + 482, + 344 + ], + "spans": [ + { + "bbox": [ + 131, + 201, + 482, + 344 + ], + "type": "text", + "content": "In the existing IMLE-based methods, we observe that the distributions of the latent codes used for training the objective differs from the distribution of latent encountered at test time. Consider an illustrative example where the latent space is two dimensional. We train a simple generative model using IMLE on two dimensional toy dataset. The latent codes used for training over the course of training are illustrated in Figure 2a. We notice that for the latent codes belonging to the same data point (denoted by the same colour) form well-separated tight bands in the latent space. We also observe that there are large gaps between these bands, indicating that these segments of the latent space are consistently overlooked during training. Since at test time we sample from the same standard normal distribution, these unsupervised segments in the latent space have arbitrary outputs, which result in bad samples. We term this phenomenon the \"Misalignment Issue.\"" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 361, + 306, + 373 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 361, + 306, + 373 + ], + "spans": [ + { + "bbox": [ + 132, + 361, + 306, + 373 + ], + "type": "text", + "content": "3.3 Analysis of the Misalignment Issue" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 380, + 481, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 380, + 481, + 418 + ], + "spans": [ + { + "bbox": [ + 130, + 380, + 481, + 418 + ], + "type": "text", + "content": "In this section, we will try to see why the phenomenon mentioned above occurs. Let us clarify the notation used in subsequent sections: " + }, + { + "bbox": [ + 130, + 380, + 481, + 418 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_i" + }, + { + "bbox": [ + 130, + 380, + 481, + 418 + ], + "type": "text", + "content": " denotes the " + }, + { + "bbox": [ + 130, + 380, + 481, + 418 + ], + "type": "inline_equation", + "content": "i^{\\mathrm{th}}" + }, + { + "bbox": [ + 130, + 380, + 481, + 418 + ], + "type": "text", + "content": " data point, " + }, + { + "bbox": [ + 130, + 380, + 481, + 418 + ], + "type": "inline_equation", + "content": "d(\\cdot ,\\cdot)" + }, + { + "bbox": [ + 130, + 380, + 481, + 418 + ], + "type": "text", + "content": " denotes the distance function and " + }, + { + "bbox": [ + 130, + 380, + 481, + 418 + ], + "type": "inline_equation", + "content": "T_{\\theta}(\\mathbf{z}_j)" + }, + { + "bbox": [ + 130, + 380, + 481, + 418 + ], + "type": "text", + "content": " denote the " + }, + { + "bbox": [ + 130, + 380, + 481, + 418 + ], + "type": "inline_equation", + "content": "j^{\\mathrm{th}}" + }, + { + "bbox": [ + 130, + 380, + 481, + 418 + ], + "type": "text", + "content": " sample where " + }, + { + "bbox": [ + 130, + 380, + 481, + 418 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_j\\sim \\mathcal{N}(0,I)" + }, + { + "bbox": [ + 130, + 380, + 481, + 418 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 418, + 482, + 499 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 418, + 482, + 499 + ], + "spans": [ + { + "bbox": [ + 131, + 418, + 482, + 499 + ], + "type": "text", + "content": "Let us define a random variable, " + }, + { + "bbox": [ + 131, + 418, + 482, + 499 + ], + "type": "inline_equation", + "content": "D_{ij}" + }, + { + "bbox": [ + 131, + 418, + 482, + 499 + ], + "type": "text", + "content": " to denote the distance of " + }, + { + "bbox": [ + 131, + 418, + 482, + 499 + ], + "type": "inline_equation", + "content": "i^{\\text{th}}" + }, + { + "bbox": [ + 131, + 418, + 482, + 499 + ], + "type": "text", + "content": " data point to the " + }, + { + "bbox": [ + 131, + 418, + 482, + 499 + ], + "type": "inline_equation", + "content": "j^{\\text{th}}" + }, + { + "bbox": [ + 131, + 418, + 482, + 499 + ], + "type": "text", + "content": " sample. We can also define another random variable " + }, + { + "bbox": [ + 131, + 418, + 482, + 499 + ], + "type": "inline_equation", + "content": "D_i^*" + }, + { + "bbox": [ + 131, + 418, + 482, + 499 + ], + "type": "text", + "content": " to denote the distance of " + }, + { + "bbox": [ + 131, + 418, + 482, + 499 + ], + "type": "inline_equation", + "content": "i^{\\text{th}}" + }, + { + "bbox": [ + 131, + 418, + 482, + 499 + ], + "type": "text", + "content": " data point to the sample closest to it. Further, let " + }, + { + "bbox": [ + 131, + 418, + 482, + 499 + ], + "type": "inline_equation", + "content": "F_{D_{ij}}" + }, + { + "bbox": [ + 131, + 418, + 482, + 499 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 418, + 482, + 499 + ], + "type": "inline_equation", + "content": "F_{D_i^*}" + }, + { + "bbox": [ + 131, + 418, + 482, + 499 + ], + "type": "text", + "content": " be the CDF of " + }, + { + "bbox": [ + 131, + 418, + 482, + 499 + ], + "type": "inline_equation", + "content": "D_{ij}" + }, + { + "bbox": [ + 131, + 418, + 482, + 499 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 418, + 482, + 499 + ], + "type": "inline_equation", + "content": "D_i^*" + }, + { + "bbox": [ + 131, + 418, + 482, + 499 + ], + "type": "text", + "content": " respectively. Let " + }, + { + "bbox": [ + 131, + 418, + 482, + 499 + ], + "type": "inline_equation", + "content": "f_{D_{ij}}" + }, + { + "bbox": [ + 131, + 418, + 482, + 499 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 418, + 482, + 499 + ], + "type": "inline_equation", + "content": "f_{D_i^*}" + }, + { + "bbox": [ + 131, + 418, + 482, + 499 + ], + "type": "text", + "content": " be the PDF of " + }, + { + "bbox": [ + 131, + 418, + 482, + 499 + ], + "type": "inline_equation", + "content": "D_{ij}" + }, + { + "bbox": [ + 131, + 418, + 482, + 499 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 418, + 482, + 499 + ], + "type": "inline_equation", + "content": "D_i^*" + }, + { + "bbox": [ + 131, + 418, + 482, + 499 + ], + "type": "text", + "content": " respectively. Now we will relate the CDF of the distances between a data point and its selected latent code, " + }, + { + "bbox": [ + 131, + 418, + 482, + 499 + ], + "type": "inline_equation", + "content": "F_{D_i^*}" + }, + { + "bbox": [ + 131, + 418, + 482, + 499 + ], + "type": "text", + "content": " to the CDF of the distances between the same data point and a random latent code " + }, + { + "bbox": [ + 131, + 418, + 482, + 499 + ], + "type": "inline_equation", + "content": "F_{D_{ij}}" + }, + { + "bbox": [ + 131, + 418, + 482, + 499 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 179, + 517, + 481, + 621 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 179, + 517, + 481, + 621 + ], + "spans": [ + { + "bbox": [ + 179, + 517, + 481, + 621 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} F _ {D _ {i} ^ {*}} (t) = \\operatorname * {P r} (D _ {i} ^ {*} \\leq t) = 1 - \\operatorname * {P r} (D _ {i} ^ {*} > t) \\\\ = 1 - \\Pr (D _ {i j} > t, \\forall j \\in [ m ]) \\quad (\\text {D e f . o f} D _ {i j}) \\\\ = 1 - \\prod_ {j = 1} ^ {m} \\Pr (D _ {i j} > t) \\\\ = 1 - \\left(\\Pr \\left(D _ {i 1} > t\\right)\\right) ^ {m} (2) \\\\ = 1 - \\left(1 - F _ {D _ {i 1}} (t)\\right) ^ {m} (3) \\\\ \\end{array}", + "image_path": "31d6e8d36d3ec47f744575a98021c8520574fd853aae15cb37633e246758d0e0.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 629, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 629, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 131, + 629, + 481, + 666 + ], + "type": "text", + "content": "Note that Equation 2 is true because each " + }, + { + "bbox": [ + 131, + 629, + 481, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_j" + }, + { + "bbox": [ + 131, + 629, + 481, + 666 + ], + "type": "text", + "content": " is drawn independently from the same probability distribution which makes " + }, + { + "bbox": [ + 131, + 629, + 481, + 666 + ], + "type": "inline_equation", + "content": "D_{i1}, D_{i2} \\dots D_{im}" + }, + { + "bbox": [ + 131, + 629, + 481, + 666 + ], + "type": "text", + "content": " identical in distribution for a particular data point " + }, + { + "bbox": [ + 131, + 629, + 481, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_i" + }, + { + "bbox": [ + 131, + 629, + 481, + 666 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 350, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 350, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 350, + 91, + 447, + 102 + ], + "type": "text", + "content": "Rejection Sampling IMLE" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 115, + 246, + 202 + ], + "blocks": [ + { + "bbox": [ + 133, + 115, + 246, + 202 + ], + "lines": [ + { + "bbox": [ + 133, + 115, + 246, + 202 + ], + "spans": [ + { + "bbox": [ + 133, + 115, + 246, + 202 + ], + "type": "image", + "image_path": "94f9f3c5429cc4a73e11fe915de277d47f4a283be323696eb592db9526c3a622.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 143, + 205, + 236, + 215 + ], + "lines": [ + { + "bbox": [ + 143, + 205, + 236, + 215 + ], + "spans": [ + { + "bbox": [ + 143, + 205, + 236, + 215 + ], + "type": "text", + "content": "(a) PDF of example distribution" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 252, + 115, + 362, + 202 + ], + "blocks": [ + { + "bbox": [ + 252, + 115, + 362, + 202 + ], + "lines": [ + { + "bbox": [ + 252, + 115, + 362, + 202 + ], + "spans": [ + { + "bbox": [ + 252, + 115, + 362, + 202 + ], + "type": "image", + "image_path": "c196852dd0bd7dd0f1cda40ab503bbfaa75ffc7454980970d768886e4d737e68.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 234, + 482, + 257 + ], + "lines": [ + { + "bbox": [ + 131, + 234, + 482, + 257 + ], + "spans": [ + { + "bbox": [ + 131, + 234, + 482, + 257 + ], + "type": "text", + "content": "Fig. 3: Illustrative figure for demonstrating the behaviour of " + }, + { + "bbox": [ + 131, + 234, + 482, + 257 + ], + "type": "inline_equation", + "content": "F_{D_i^*}(t)" + }, + { + "bbox": [ + 131, + 234, + 482, + 257 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 234, + 482, + 257 + ], + "type": "inline_equation", + "content": "F_{D_{i1}}(t)" + }, + { + "bbox": [ + 131, + 234, + 482, + 257 + ], + "type": "text", + "content": " using Noncentral Chi-squared distribution as the example distribution." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 369, + 115, + 480, + 202 + ], + "blocks": [ + { + "bbox": [ + 251, + 205, + 363, + 219 + ], + "lines": [ + { + "bbox": [ + 251, + 205, + 363, + 219 + ], + "spans": [ + { + "bbox": [ + 251, + 205, + 363, + 219 + ], + "type": "text", + "content": "(b) " + }, + { + "bbox": [ + 251, + 205, + 363, + 219 + ], + "type": "inline_equation", + "content": "F_{D_i^*}(t)" + }, + { + "bbox": [ + 251, + 205, + 363, + 219 + ], + "type": "text", + "content": " vs " + }, + { + "bbox": [ + 251, + 205, + 363, + 219 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 251, + 205, + 363, + 219 + ], + "type": "text", + "content": " for different values of " + }, + { + "bbox": [ + 251, + 205, + 363, + 219 + ], + "type": "inline_equation", + "content": "m" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 369, + 115, + 480, + 202 + ], + "lines": [ + { + "bbox": [ + 369, + 115, + 480, + 202 + ], + "spans": [ + { + "bbox": [ + 369, + 115, + 480, + 202 + ], + "type": "image", + "image_path": "434e8a8b691bba4e02e829fbdd5ee16b3b257ce536af00347c43667e993208f9.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 368, + 205, + 481, + 224 + ], + "lines": [ + { + "bbox": [ + 368, + 205, + 481, + 224 + ], + "spans": [ + { + "bbox": [ + 368, + 205, + 481, + 224 + ], + "type": "text", + "content": "(c) " + }, + { + "bbox": [ + 368, + 205, + 481, + 224 + ], + "type": "inline_equation", + "content": "F_{D_i^*}(t)" + }, + { + "bbox": [ + 368, + 205, + 481, + 224 + ], + "type": "text", + "content": " vs " + }, + { + "bbox": [ + 368, + 205, + 481, + 224 + ], + "type": "inline_equation", + "content": "F_{D_{i1}}(t)" + }, + { + "bbox": [ + 368, + 205, + 481, + 224 + ], + "type": "text", + "content": " for different values of " + }, + { + "bbox": [ + 368, + 205, + 481, + 224 + ], + "type": "inline_equation", + "content": "m" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 281, + 482, + 368 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 281, + 482, + 368 + ], + "spans": [ + { + "bbox": [ + 130, + 281, + 482, + 368 + ], + "type": "text", + "content": "Now we can try to justify our observations from Figure 2a. Equation 3 shows us how the CDF of the distance of the selected latent code (used in training) differs significantly from the CDF of distance of a random latent code (encountered at test time). This shows that the distance of the sample we choose for training is lower than the distance of the sample at testing. We can obtain a deeper understanding by analyzing the plots in Figure 3 for an example distribution. Notice that " + }, + { + "bbox": [ + 130, + 281, + 482, + 368 + ], + "type": "inline_equation", + "content": "\\forall m > 1, F_{D_i^*}(t) > F_{D_{i1}}(t)" + }, + { + "bbox": [ + 130, + 281, + 482, + 368 + ], + "type": "text", + "content": ". Observing Equation 3 and the aforementioned plots, we can conclude the following things:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 136, + 376, + 481, + 440 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 136, + 376, + 481, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 376, + 481, + 413 + ], + "spans": [ + { + "bbox": [ + 136, + 376, + 481, + 413 + ], + "type": "text", + "content": "1. The " + }, + { + "bbox": [ + 136, + 376, + 481, + 413 + ], + "type": "inline_equation", + "content": "f_{D_i^*}" + }, + { + "bbox": [ + 136, + 376, + 481, + 413 + ], + "type": "text", + "content": " is skewed towards the origin compared to " + }, + { + "bbox": [ + 136, + 376, + 481, + 413 + ], + "type": "inline_equation", + "content": "f_{D_{ij}}" + }, + { + "bbox": [ + 136, + 376, + 481, + 413 + ], + "type": "text", + "content": ". This is intuitive because the latent codes are selected by the min operation and so their distance to their respective data point would be less than that of a random sample." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 136, + 414, + 481, + 440 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 414, + 481, + 440 + ], + "spans": [ + { + "bbox": [ + 136, + 414, + 481, + 440 + ], + "type": "text", + "content": "2. The skew towards the data point increases as " + }, + { + "bbox": [ + 136, + 414, + 481, + 440 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 136, + 414, + 481, + 440 + ], + "type": "text", + "content": " increases. This piece of information will be important shortly." + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 130, + 449, + 482, + 473 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 449, + 482, + 473 + ], + "spans": [ + { + "bbox": [ + 130, + 449, + 482, + 473 + ], + "type": "text", + "content": "We can also compute the PDF " + }, + { + "bbox": [ + 130, + 449, + 482, + 473 + ], + "type": "inline_equation", + "content": "f_{D_i^*}" + }, + { + "bbox": [ + 130, + 449, + 482, + 473 + ], + "type": "text", + "content": " in terms of " + }, + { + "bbox": [ + 130, + 449, + 482, + 473 + ], + "type": "inline_equation", + "content": "f_{D_{i1}}" + }, + { + "bbox": [ + 130, + 449, + 482, + 473 + ], + "type": "text", + "content": " by differentiating the CDF " + }, + { + "bbox": [ + 130, + 449, + 482, + 473 + ], + "type": "inline_equation", + "content": "F_{D_i^*}(t)" + }, + { + "bbox": [ + 130, + 449, + 482, + 473 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 208, + 493, + 481, + 520 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 208, + 493, + 481, + 520 + ], + "spans": [ + { + "bbox": [ + 208, + 493, + 481, + 520 + ], + "type": "interline_equation", + "content": "f _ {D _ {i} ^ {*}} (t) = \\frac {\\mathrm {d} F _ {D _ {i} ^ {*}} (t)}{\\mathrm {d} t} = m \\left(1 - F _ {D _ {i 1}} (t)\\right) ^ {m - 1} f _ {D _ {i 1}} (t) \\tag {4}", + "image_path": "b7b50b410163f77d3ea2da3d586182f8d6f6a17dc1c7568e94c0481c94795a16.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 131, + 536, + 290, + 550 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 536, + 290, + 550 + ], + "spans": [ + { + "bbox": [ + 131, + 536, + 290, + 550 + ], + "type": "text", + "content": "3.4 Solving the Misalignment Issue" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 558, + 482, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 558, + 482, + 668 + ], + "spans": [ + { + "bbox": [ + 130, + 558, + 482, + 668 + ], + "type": "text", + "content": "Now that we know the reason behind the misalignment between latent codes used at training and testing, we wish to come up with a method to mitigate this phenomenon. Recall that " + }, + { + "bbox": [ + 130, + 558, + 482, + 668 + ], + "type": "inline_equation", + "content": "D_{ij}" + }, + { + "bbox": [ + 130, + 558, + 482, + 668 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 558, + 482, + 668 + ], + "type": "inline_equation", + "content": "D_i^*" + }, + { + "bbox": [ + 130, + 558, + 482, + 668 + ], + "type": "text", + "content": " are determined by the distance function " + }, + { + "bbox": [ + 130, + 558, + 482, + 668 + ], + "type": "inline_equation", + "content": "d(\\cdot, \\cdot)" + }, + { + "bbox": [ + 130, + 558, + 482, + 668 + ], + "type": "text", + "content": ", neural network " + }, + { + "bbox": [ + 130, + 558, + 482, + 668 + ], + "type": "inline_equation", + "content": "T_\\theta" + }, + { + "bbox": [ + 130, + 558, + 482, + 668 + ], + "type": "text", + "content": " and the prior distribution. For a given generative modelling task, it is not trivial to change the " + }, + { + "bbox": [ + 130, + 558, + 482, + 668 + ], + "type": "inline_equation", + "content": "d(\\cdot, \\cdot)" + }, + { + "bbox": [ + 130, + 558, + 482, + 668 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 130, + 558, + 482, + 668 + ], + "type": "inline_equation", + "content": "T_\\theta" + }, + { + "bbox": [ + 130, + 558, + 482, + 668 + ], + "type": "text", + "content": ". Hence, in this paper, we aim to change the prior distribution used at training such that the distribution of latent codes at training time closely matches with the distribution of latent codes (drawn from the standard normal distribution) encountered at test time. Notice that the IMLE objective (Equation 1) allows us to sample from the prior without knowing the closed form expression for its probability density" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 224, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 224, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 224, + 101 + ], + "type": "text", + "content": "C. Vashist et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "code", + "bbox": [ + 132, + 130, + 332, + 277 + ], + "blocks": [ + { + "bbox": [ + 132, + 115, + 272, + 128 + ], + "lines": [ + { + "bbox": [ + 132, + 115, + 272, + 128 + ], + "spans": [ + { + "bbox": [ + 132, + 115, + 272, + 128 + ], + "type": "text", + "content": "Algorithm 1 RS-IMLE Procedure" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "code_caption" + }, + { + "bbox": [ + 132, + 130, + 332, + 277 + ], + "lines": [ + { + "bbox": [ + 132, + 130, + 332, + 277 + ], + "spans": [ + { + "bbox": [ + 132, + 130, + 332, + 277 + ], + "type": "text", + "content": "Require: The set of inputs " + }, + { + "bbox": [ + 132, + 130, + 332, + 277 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{x}_i\\}_{i = 1}^n" + }, + { + "bbox": [ + 132, + 130, + 332, + 277 + ], + "type": "text", + "content": " , radius " + }, + { + "bbox": [ + 132, + 130, + 332, + 277 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 132, + 130, + 332, + 277 + ], + "type": "text", + "content": " \n1: Initialize the parameters " + }, + { + "bbox": [ + 132, + 130, + 332, + 277 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 132, + 130, + 332, + 277 + ], + "type": "text", + "content": " of the generator " + }, + { + "bbox": [ + 132, + 130, + 332, + 277 + ], + "type": "inline_equation", + "content": "T_{\\theta}" + }, + { + "bbox": [ + 132, + 130, + 332, + 277 + ], + "type": "text", + "content": " \n2: for " + }, + { + "bbox": [ + 132, + 130, + 332, + 277 + ], + "type": "inline_equation", + "content": "k = 1" + }, + { + "bbox": [ + 132, + 130, + 332, + 277 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 132, + 130, + 332, + 277 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 132, + 130, + 332, + 277 + ], + "type": "text", + "content": " do \n3: Draw latent codes " + }, + { + "bbox": [ + 132, + 130, + 332, + 277 + ], + "type": "inline_equation", + "content": "Z\\gets \\mathbf{z}_1,\\dots,\\mathbf{z}_m" + }, + { + "bbox": [ + 132, + 130, + 332, + 277 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 132, + 130, + 332, + 277 + ], + "type": "inline_equation", + "content": "\\mathcal{N}(0,\\mathbf{I})" + }, + { + "bbox": [ + 132, + 130, + 332, + 277 + ], + "type": "text", + "content": " \n4: Compute " + }, + { + "bbox": [ + 132, + 130, + 332, + 277 + ], + "type": "inline_equation", + "content": "\\tilde{Z}\\gets \\tilde{\\mathbf{z}}_1,\\ldots ,\\tilde{\\mathbf{z}}_p" + }, + { + "bbox": [ + 132, + 130, + 332, + 277 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 132, + 130, + 332, + 277 + ], + "type": "inline_equation", + "content": "Z" + }, + { + "bbox": [ + 132, + 130, + 332, + 277 + ], + "type": "text", + "content": " such that " + }, + { + "bbox": [ + 132, + 130, + 332, + 277 + ], + "type": "inline_equation", + "content": "d\\left(\\mathbf{x}_i,T_\\theta (\\tilde{\\mathbf{z}}_j)\\right)\\geq \\epsilon ,\\quad \\forall \\tilde{\\mathbf{z}}_j\\in \\tilde{Z},i\\in [n]" + }, + { + "bbox": [ + 132, + 130, + 332, + 277 + ], + "type": "text", + "content": " \n5: " + }, + { + "bbox": [ + 132, + 130, + 332, + 277 + ], + "type": "inline_equation", + "content": "\\sigma (i)\\gets \\arg \\min_{j\\in [m]}d(\\mathbf{x}_i,T_\\theta (\\tilde{\\mathbf{z}}_j)),\\quad \\forall i\\in [n]" + }, + { + "bbox": [ + 132, + 130, + 332, + 277 + ], + "type": "text", + "content": " \n6: for " + }, + { + "bbox": [ + 132, + 130, + 332, + 277 + ], + "type": "inline_equation", + "content": "l = 1" + }, + { + "bbox": [ + 132, + 130, + 332, + 277 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 132, + 130, + 332, + 277 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 132, + 130, + 332, + 277 + ], + "type": "text", + "content": " do \n7: Pick a random batch " + }, + { + "bbox": [ + 132, + 130, + 332, + 277 + ], + "type": "inline_equation", + "content": "S\\subseteq [n]" + }, + { + "bbox": [ + 132, + 130, + 332, + 277 + ], + "type": "text", + "content": " \n8: " + }, + { + "bbox": [ + 132, + 130, + 332, + 277 + ], + "type": "inline_equation", + "content": "\\theta \\gets \\theta -\\eta \\nabla_{\\theta}\\left(\\sum_{i\\in S}d\\left(\\mathbf{x}_i,T_\\theta \\left(\\tilde{\\mathbf{z}}_{\\sigma (i)}\\right)\\right)\\right) / |S|" + }, + { + "bbox": [ + 132, + 130, + 332, + 277 + ], + "type": "text", + "content": " \n9: end for \n10: end for \n11: return " + }, + { + "bbox": [ + 132, + 130, + 332, + 277 + ], + "type": "inline_equation", + "content": "\\theta" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "code_body" + } + ], + "index": 3, + "sub_type": "algorithm" + }, + { + "bbox": [ + 130, + 310, + 482, + 401 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 310, + 482, + 401 + ], + "spans": [ + { + "bbox": [ + 130, + 310, + 482, + 401 + ], + "type": "text", + "content": "function. This allows us the flexibility of choosing a non-analytical prior distribution. To distinguish from " + }, + { + "bbox": [ + 130, + 310, + 482, + 401 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_j\\sim \\mathcal{N}(0,I)" + }, + { + "bbox": [ + 130, + 310, + 482, + 401 + ], + "type": "text", + "content": ", we will use " + }, + { + "bbox": [ + 130, + 310, + 482, + 401 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{z}}" + }, + { + "bbox": [ + 130, + 310, + 482, + 401 + ], + "type": "text", + "content": " to denote the latent codes drawn from our desired target distribution " + }, + { + "bbox": [ + 130, + 310, + 482, + 401 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 130, + 310, + 482, + 401 + ], + "type": "text", + "content": ". Identical to the previous section, let us define random variable " + }, + { + "bbox": [ + 130, + 310, + 482, + 401 + ], + "type": "inline_equation", + "content": "\\tilde{D}_{ij} = d\\left(\\mathbf{x}_i,T_\\theta (\\tilde{\\mathbf{z}}_j)\\right)" + }, + { + "bbox": [ + 130, + 310, + 482, + 401 + ], + "type": "text", + "content": " to denote the distance between data point " + }, + { + "bbox": [ + 130, + 310, + 482, + 401 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_i" + }, + { + "bbox": [ + 130, + 310, + 482, + 401 + ], + "type": "text", + "content": " from a random sample " + }, + { + "bbox": [ + 130, + 310, + 482, + 401 + ], + "type": "inline_equation", + "content": "T_{\\theta}(\\tilde{\\mathbf{z}}_j)" + }, + { + "bbox": [ + 130, + 310, + 482, + 401 + ], + "type": "text", + "content": ". Similarly, we define " + }, + { + "bbox": [ + 130, + 310, + 482, + 401 + ], + "type": "inline_equation", + "content": "\\tilde{D}_i^* = \\min_{j\\in [m]}\\tilde{D}_{ij}" + }, + { + "bbox": [ + 130, + 310, + 482, + 401 + ], + "type": "text", + "content": ". Similar to the section above, let " + }, + { + "bbox": [ + 130, + 310, + 482, + 401 + ], + "type": "inline_equation", + "content": "F_{\\tilde{D}_{ij}}" + }, + { + "bbox": [ + 130, + 310, + 482, + 401 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 310, + 482, + 401 + ], + "type": "inline_equation", + "content": "F_{\\tilde{D}_i^*}" + }, + { + "bbox": [ + 130, + 310, + 482, + 401 + ], + "type": "text", + "content": " be the CDF of " + }, + { + "bbox": [ + 130, + 310, + 482, + 401 + ], + "type": "inline_equation", + "content": "\\tilde{D}_{ij}" + }, + { + "bbox": [ + 130, + 310, + 482, + 401 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 310, + 482, + 401 + ], + "type": "inline_equation", + "content": "\\tilde{D}_i^*" + }, + { + "bbox": [ + 130, + 310, + 482, + 401 + ], + "type": "text", + "content": " respectively. Then " + }, + { + "bbox": [ + 130, + 310, + 482, + 401 + ], + "type": "inline_equation", + "content": "f_{\\tilde{D}_{ij}}" + }, + { + "bbox": [ + 130, + 310, + 482, + 401 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 310, + 482, + 401 + ], + "type": "inline_equation", + "content": "f_{\\tilde{D}_i^*}" + }, + { + "bbox": [ + 130, + 310, + 482, + 401 + ], + "type": "text", + "content": " would be the PDF of " + }, + { + "bbox": [ + 130, + 310, + 482, + 401 + ], + "type": "inline_equation", + "content": "\\tilde{D}_{ij}" + }, + { + "bbox": [ + 130, + 310, + 482, + 401 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 310, + 482, + 401 + ], + "type": "inline_equation", + "content": "\\tilde{D}_i^*" + }, + { + "bbox": [ + 130, + 310, + 482, + 401 + ], + "type": "text", + "content": " respectively." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 438, + 482, + 606 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 438, + 482, + 606 + ], + "spans": [ + { + "bbox": [ + 130, + 438, + 482, + 606 + ], + "type": "text", + "content": "Designing the target prior Now, we will try to think about the properties we would want to have for our target prior. Recall that the Misalignment Issue is mitigated as we decrease the number of samples. In order to differentiate between the number of samples of different priors, we use the notation " + }, + { + "bbox": [ + 130, + 438, + 482, + 606 + ], + "type": "inline_equation", + "content": "m'" + }, + { + "bbox": [ + 130, + 438, + 482, + 606 + ], + "type": "text", + "content": " to denote the number of samples for the objective using the Gaussian prior. As hinted in the previous section, one way to avoid the Misalignment Issue is to set " + }, + { + "bbox": [ + 130, + 438, + 482, + 606 + ], + "type": "inline_equation", + "content": "m'" + }, + { + "bbox": [ + 130, + 438, + 482, + 606 + ], + "type": "text", + "content": " to low values. However we cannot directly use a Gaussian prior with low values of " + }, + { + "bbox": [ + 130, + 438, + 482, + 606 + ], + "type": "inline_equation", + "content": "m'" + }, + { + "bbox": [ + 130, + 438, + 482, + 606 + ], + "type": "text", + "content": " as our target prior. This is because in aforementioned case, it happens to be that many data points are forced to pick the same sample as their nearest neighbour. This in turn causes very slow or no learning. Hence when using a Gaussian prior, we need to pick " + }, + { + "bbox": [ + 130, + 438, + 482, + 606 + ], + "type": "inline_equation", + "content": "m'" + }, + { + "bbox": [ + 130, + 438, + 482, + 606 + ], + "type": "text", + "content": " large enough to allow convergence and yet small enough such that the Misalignment Issue does not affect test time performance. In our case, we are trying to design a new distribution that solves the Misalignment Issue by having desirable properties of an ideal distribution. The ideal distribution is a Gaussian prior with " + }, + { + "bbox": [ + 130, + 438, + 482, + 606 + ], + "type": "inline_equation", + "content": "m'" + }, + { + "bbox": [ + 130, + 438, + 482, + 606 + ], + "type": "text", + "content": " set to the lowest possible value, which is " + }, + { + "bbox": [ + 130, + 438, + 482, + 606 + ], + "type": "inline_equation", + "content": "m' = n" + }, + { + "bbox": [ + 130, + 438, + 482, + 606 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 609, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 609, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 609, + 482, + 666 + ], + "type": "text", + "content": "Now that we have found the ideal prior, we can make " + }, + { + "bbox": [ + 130, + 609, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 130, + 609, + 482, + 666 + ], + "type": "text", + "content": " and the ideal prior to be equal in distribution. Similar to the analysis till Equation 4, we derive the PDF of " + }, + { + "bbox": [ + 130, + 609, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 130, + 609, + 482, + 666 + ], + "type": "text", + "content": ". We get " + }, + { + "bbox": [ + 130, + 609, + 482, + 666 + ], + "type": "inline_equation", + "content": "f_{\\tilde{D}_i^*}(t) = m\\left(1 - F_{\\tilde{D}_{i1}}(t)\\right)^{m - 1}f_{\\tilde{D}_{i1}}(t)" + }, + { + "bbox": [ + 130, + 609, + 482, + 666 + ], + "type": "text", + "content": ". Equating this PDF of " + }, + { + "bbox": [ + 130, + 609, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 130, + 609, + 482, + 666 + ], + "type": "text", + "content": " to the PDF of the ideal distribution gives:" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 350, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 350, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 350, + 91, + 447, + 102 + ], + "type": "text", + "content": "Rejection Sampling IMLE" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 194, + 135, + 481, + 224 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 194, + 135, + 481, + 224 + ], + "spans": [ + { + "bbox": [ + 194, + 135, + 481, + 224 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} m \\left(1 - F _ {\\tilde {D} _ {i 1}} (t)\\right) ^ {m - 1} f _ {\\tilde {D} _ {i 1}} (t) = n \\left(1 - F _ {D _ {i 1}} (t)\\right) ^ {n - 1} f _ {D _ {i 1}} (t) \\\\ \\Rightarrow f _ {\\tilde {D} _ {i 1}} (t) = \\frac {n}{m} \\frac {\\left(1 - F _ {D _ {i 1}} (t)\\right) ^ {n - 1}}{\\left(1 - F _ {\\tilde {D} _ {i 1}} (t)\\right) ^ {m - 1}} f _ {D _ {i 1}} (t) (5) \\\\ \\Longrightarrow f _ {\\tilde {D} _ {i 1}} (t) = c \\phi (t) f _ {D _ {i 1}} (t) (6) \\\\ \\end{array}", + "image_path": "b254d347840bed0fbf21e8c11fe9edf38ea14fb3da9b919f3d2dcbaf0bf4002d.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 232, + 481, + 268 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 232, + 481, + 268 + ], + "spans": [ + { + "bbox": [ + 130, + 232, + 481, + 268 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 130, + 232, + 481, + 268 + ], + "type": "inline_equation", + "content": "\\phi(t)" + }, + { + "bbox": [ + 130, + 232, + 481, + 268 + ], + "type": "text", + "content": " is introduced to simplify notation and " + }, + { + "bbox": [ + 130, + 232, + 481, + 268 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 130, + 232, + 481, + 268 + ], + "type": "text", + "content": " is a constant associated with truncating " + }, + { + "bbox": [ + 130, + 232, + 481, + 268 + ], + "type": "inline_equation", + "content": "F_{D_{i1}}(t)" + }, + { + "bbox": [ + 130, + 232, + 481, + 268 + ], + "type": "text", + "content": " that is needed to make sure " + }, + { + "bbox": [ + 130, + 232, + 481, + 268 + ], + "type": "inline_equation", + "content": "f_{\\tilde{D}_{i1}}(t)" + }, + { + "bbox": [ + 130, + 232, + 481, + 268 + ], + "type": "text", + "content": " is bounded. We discuss these technical details in the Appendix." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 285, + 479, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 285, + 479, + 357 + ], + "spans": [ + { + "bbox": [ + 130, + 285, + 479, + 357 + ], + "type": "text", + "content": "Rejection Sampling We have expressed our target prior " + }, + { + "bbox": [ + 130, + 285, + 479, + 357 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 130, + 285, + 479, + 357 + ], + "type": "text", + "content": " in terms of distribution from which we can easily sample. Now, we can use Rejection Sampling to sample from our target prior " + }, + { + "bbox": [ + 130, + 285, + 479, + 357 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 130, + 285, + 479, + 357 + ], + "type": "text", + "content": ". To be concrete: " + }, + { + "bbox": [ + 130, + 285, + 479, + 357 + ], + "type": "inline_equation", + "content": "f_{\\tilde{D}_{i1}^*}(t)" + }, + { + "bbox": [ + 130, + 285, + 479, + 357 + ], + "type": "text", + "content": " is our target distribution and " + }, + { + "bbox": [ + 130, + 285, + 479, + 357 + ], + "type": "inline_equation", + "content": "f_{D_{i1}}(t)" + }, + { + "bbox": [ + 130, + 285, + 479, + 357 + ], + "type": "text", + "content": " acts as our proposal distribution, since we can sample from the standard Gaussian easily. We can further write it in the standard Rejection Sampling notation: " + }, + { + "bbox": [ + 130, + 285, + 479, + 357 + ], + "type": "inline_equation", + "content": "\\frac{f_{\\tilde{D}_{i1}^*}(t)}{M f_{D^*}(t)} = \\frac{c\\phi(t)}{M}" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 359, + 480, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 359, + 480, + 418 + ], + "spans": [ + { + "bbox": [ + 130, + 359, + 480, + 418 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 130, + 359, + 480, + 418 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 130, + 359, + 480, + 418 + ], + "type": "text", + "content": " is the scaling factor associated with Rejection Sampling. We approximate the function above using a step function. The step needs to happen at " + }, + { + "bbox": [ + 130, + 359, + 480, + 418 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 359, + 480, + 418 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 130, + 359, + 480, + 418 + ], + "type": "inline_equation", + "content": "F_{\\tilde{D}_{i1}}(t)" + }, + { + "bbox": [ + 130, + 359, + 480, + 418 + ], + "type": "text", + "content": " gets close to 1. Instead of trying to estimate " + }, + { + "bbox": [ + 130, + 359, + 480, + 418 + ], + "type": "inline_equation", + "content": "F_{\\tilde{D}_{i1}}(t)" + }, + { + "bbox": [ + 130, + 359, + 480, + 418 + ], + "type": "text", + "content": ", we instead use a hyperparameter " + }, + { + "bbox": [ + 130, + 359, + 480, + 418 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 130, + 359, + 480, + 418 + ], + "type": "text", + "content": " to represent where " + }, + { + "bbox": [ + 130, + 359, + 480, + 418 + ], + "type": "inline_equation", + "content": "F_{\\tilde{D}_{i1}}(t)" + }, + { + "bbox": [ + 130, + 359, + 480, + 418 + ], + "type": "text", + "content": " gets close to 1. We find the value of this hyperparameter " + }, + { + "bbox": [ + 130, + 359, + 480, + 418 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 130, + 359, + 480, + 418 + ], + "type": "text", + "content": " by cross-validation." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 418, + 480, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 418, + 480, + 467 + ], + "spans": [ + { + "bbox": [ + 130, + 418, + 480, + 467 + ], + "type": "text", + "content": "The final procedure simplifies to this: we sample " + }, + { + "bbox": [ + 130, + 418, + 480, + 467 + ], + "type": "inline_equation", + "content": "\\mathbf{z} \\sim \\mathcal{N}(0, I)" + }, + { + "bbox": [ + 130, + 418, + 480, + 467 + ], + "type": "text", + "content": ". If " + }, + { + "bbox": [ + 130, + 418, + 480, + 467 + ], + "type": "inline_equation", + "content": "t = d\\left(\\mathbf{x}_i, T_\\theta(\\mathbf{z})\\right) < \\epsilon" + }, + { + "bbox": [ + 130, + 418, + 480, + 467 + ], + "type": "text", + "content": ", we reject the sample; otherwise, we accept it. Since the sampling procedure is based on Rejection Sampling, we call our method RS-IMLE. The resulting RS-IMLE procedure is included in Algorithm-1." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 484, + 376, + 497 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 484, + 376, + 497 + ], + "spans": [ + { + "bbox": [ + 131, + 484, + 376, + 497 + ], + "type": "text", + "content": "3.5 Intuitive Interpretation of the Algorithm Behaviour" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 504, + 479, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 504, + 479, + 613 + ], + "spans": [ + { + "bbox": [ + 130, + 504, + 479, + 613 + ], + "type": "text", + "content": "Prior to proceeding to the Implementation details, gaining a gradient-based understanding of our new objective would provide valuable insight. Let us revisit the IMLE objective (Equation 1) again. As training progresses, we expect the loss to reduce such that: " + }, + { + "bbox": [ + 130, + 504, + 479, + 613 + ], + "type": "inline_equation", + "content": "\\forall x_{i},\\quad \\mathbb{E}_{z_{1},\\ldots ,z_{m}\\sim \\mathcal{N}(0,I)}\\min_{j\\in [m]}d\\left(\\mathbf{x}_{i},T_{\\theta}(\\mathbf{z}_{j})\\right)\\to 0." + }, + { + "bbox": [ + 130, + 504, + 479, + 613 + ], + "type": "text", + "content": " As the objective approaches convergence, the loss will decrease, resulting in gradients with lower magnitude. This causes smaller updates to the model parameters during training, leading to slower progress. Note that this loss is with respect to the closest sample to each data point. Thus it can be the case that even after a lot of training, although the closest sample are pretty close to their respective data point, the rest of samples are pretty far away." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 185, + 631, + 481, + 664 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 185, + 631, + 481, + 664 + ], + "spans": [ + { + "bbox": [ + 185, + 631, + 481, + 664 + ], + "type": "interline_equation", + "content": "\\theta_ {\\mathrm {R S - I M L E}} = \\arg \\min _ {\\theta} \\mathbb {E} _ {z _ {1}, \\dots , z _ {m} \\sim \\mathcal {P}} \\left[ \\sum_ {i = 1} ^ {n} \\min _ {j \\in [ m ]} d \\left(\\mathbf {x} _ {i}, T _ {\\theta} (\\mathbf {z} _ {j})\\right) \\right] \\tag {7}", + "image_path": "5e8331690f9afa0e274c159600bfa161b5a933c775705872431d75250ee7547f.jpg" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 139, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 139, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 139, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 223, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 223, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 223, + 101 + ], + "type": "text", + "content": "C. Vashist et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 145, + 122, + 294, + 190 + ], + "blocks": [ + { + "bbox": [ + 145, + 122, + 294, + 190 + ], + "lines": [ + { + "bbox": [ + 145, + 122, + 294, + 190 + ], + "spans": [ + { + "bbox": [ + 145, + 122, + 294, + 190 + ], + "type": "image", + "image_path": "903da0e53e211d41f5c74ca952d7d4c34757878cb0b101ecf10b99f6e0909ed6.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 182, + 205, + 258, + 214 + ], + "lines": [ + { + "bbox": [ + 182, + 205, + 258, + 214 + ], + "spans": [ + { + "bbox": [ + 182, + 205, + 258, + 214 + ], + "type": "text", + "content": "(a) IMLE after epoch 100" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 314, + 118, + 471, + 194 + ], + "blocks": [ + { + "bbox": [ + 314, + 118, + 471, + 194 + ], + "lines": [ + { + "bbox": [ + 314, + 118, + 471, + 194 + ], + "spans": [ + { + "bbox": [ + 314, + 118, + 471, + 194 + ], + "type": "image", + "image_path": "3987190656c0d473856a93f2aff5f30c29a1f264edc827aa34b32ad5a33cd6bf.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 352, + 205, + 437, + 214 + ], + "lines": [ + { + "bbox": [ + 352, + 205, + 437, + 214 + ], + "spans": [ + { + "bbox": [ + 352, + 205, + 437, + 214 + ], + "type": "text", + "content": "(b) RS-IMLE after epoch 100" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 145, + 222, + 296, + 291 + ], + "blocks": [ + { + "bbox": [ + 145, + 222, + 296, + 291 + ], + "lines": [ + { + "bbox": [ + 145, + 222, + 296, + 291 + ], + "spans": [ + { + "bbox": [ + 145, + 222, + 296, + 291 + ], + "type": "image", + "image_path": "1d3d46be4d8958affaab93487277a459c26224141e1668a1b0b88c912695bb44.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 184, + 304, + 258, + 313 + ], + "lines": [ + { + "bbox": [ + 184, + 304, + 258, + 313 + ], + "spans": [ + { + "bbox": [ + 184, + 304, + 258, + 313 + ], + "type": "text", + "content": "(c) IMLE after epoch 500" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 315, + 218, + 477, + 293 + ], + "blocks": [ + { + "bbox": [ + 315, + 218, + 477, + 293 + ], + "lines": [ + { + "bbox": [ + 315, + 218, + 477, + 293 + ], + "spans": [ + { + "bbox": [ + 315, + 218, + 477, + 293 + ], + "type": "image", + "image_path": "849efd1263ad76742fa1b334afcd086d74b89d0c3e2b0ba6312e80dfb278e56e.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 352, + 304, + 437, + 313 + ], + "lines": [ + { + "bbox": [ + 352, + 304, + 437, + 313 + ], + "spans": [ + { + "bbox": [ + 352, + 304, + 437, + 313 + ], + "type": "text", + "content": "(d) RS-IMLE after epoch 500" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 145, + 322, + 293, + 388 + ], + "blocks": [ + { + "bbox": [ + 145, + 322, + 293, + 388 + ], + "lines": [ + { + "bbox": [ + 145, + 322, + 293, + 388 + ], + "spans": [ + { + "bbox": [ + 145, + 322, + 293, + 388 + ], + "type": "image", + "image_path": "02f525b8395991748243721048bbd43e663e42adc6de52da2bd099bdc5092f2a.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 181, + 403, + 260, + 413 + ], + "lines": [ + { + "bbox": [ + 181, + 403, + 260, + 413 + ], + "spans": [ + { + "bbox": [ + 181, + 403, + 260, + 413 + ], + "type": "text", + "content": "(e) IMLE after 2000 epochs" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 315, + 318, + 471, + 395 + ], + "blocks": [ + { + "bbox": [ + 315, + 318, + 471, + 395 + ], + "lines": [ + { + "bbox": [ + 315, + 318, + 471, + 395 + ], + "spans": [ + { + "bbox": [ + 315, + 318, + 471, + 395 + ], + "type": "image", + "image_path": "756c96d4b94fd9f56b56094cf52a6c5053a063c60de10cab4e844dc7080cdea8.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 349, + 403, + 440, + 413 + ], + "lines": [ + { + "bbox": [ + 349, + 403, + 440, + 413 + ], + "spans": [ + { + "bbox": [ + 349, + 403, + 440, + 413 + ], + "type": "text", + "content": "(f) RS-IMLE after 2000 epochs" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 131, + 422, + 482, + 445 + ], + "lines": [ + { + "bbox": [ + 131, + 422, + 482, + 445 + ], + "spans": [ + { + "bbox": [ + 131, + 422, + 482, + 445 + ], + "type": "text", + "content": "Fig. 4: Comparison between IMLE and RS-IMLE for 2D toy problem. Data points are denoted by " + }, + { + "bbox": [ + 131, + 422, + 482, + 445 + ], + "type": "inline_equation", + "content": "\\blacksquare" + }, + { + "bbox": [ + 131, + 422, + 482, + 445 + ], + "type": "text", + "content": " and samples are denoted by " + }, + { + "bbox": [ + 131, + 422, + 482, + 445 + ], + "type": "inline_equation", + "content": "\\bullet" + }, + { + "bbox": [ + 131, + 422, + 482, + 445 + ], + "type": "text", + "content": ". Samples picked as nearest neighbours are denoted by " + }, + { + "bbox": [ + 131, + 422, + 482, + 445 + ], + "type": "inline_equation", + "content": "\\star" + }, + { + "bbox": [ + 131, + 422, + 482, + 445 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 477, + 482, + 550 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 477, + 482, + 550 + ], + "spans": [ + { + "bbox": [ + 130, + 477, + 482, + 550 + ], + "type": "text", + "content": "Consider our objective in Equation 7 and recall that we have constructed the probability distribution " + }, + { + "bbox": [ + 130, + 477, + 482, + 550 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 130, + 477, + 482, + 550 + ], + "type": "text", + "content": " such that " + }, + { + "bbox": [ + 130, + 477, + 482, + 550 + ], + "type": "inline_equation", + "content": "\\forall \\mathbf{x}_i\\quad d\\left(\\mathbf{x}_i,T_\\theta (\\tilde{\\mathbf{z}})\\right)\\geq \\epsilon" + }, + { + "bbox": [ + 130, + 477, + 482, + 550 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 477, + 482, + 550 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{z}}\\sim \\mathcal{P}" + }, + { + "bbox": [ + 130, + 477, + 482, + 550 + ], + "type": "text", + "content": ". In other words, all samples we obtain by using the prior " + }, + { + "bbox": [ + 130, + 477, + 482, + 550 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 130, + 477, + 482, + 550 + ], + "type": "text", + "content": " are guaranteed to be " + }, + { + "bbox": [ + 130, + 477, + 482, + 550 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 130, + 477, + 482, + 550 + ], + "type": "text", + "content": "-distance away from all data points. This ensures that the loss per data point is always greater than " + }, + { + "bbox": [ + 130, + 477, + 482, + 550 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 130, + 477, + 482, + 550 + ], + "type": "text", + "content": ". The approach can be interpreted as ignoring the samples that are already close to some data point and instead training on challenging, non-trivial samples." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 553, + 482, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 553, + 482, + 602 + ], + "spans": [ + { + "bbox": [ + 130, + 553, + 482, + 602 + ], + "type": "text", + "content": "To compare the sampling behavior of the vanilla IMLE and the proposed RS-IMLE, we trained two models on a 2D toy problem as illustrated in Figure 4. The first model uses the vanilla IMLE objective (Equation 1), while the second model is trained with our proposed RS-IMLE objective (Equation 7)." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": "At the initial stage of training, both the methods learn similar distributions, indicated by the straight line of orange dots ( " + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\odot" + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": " ) in Figure-4a,4b. Our method first removes all the samples that fall within an " + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": " distance from any data point before doing the nearest neighbour search. We illustrate this in Figure-4b, where samples that lie within the gray circles are not considered for the nearest neighbour search." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 350, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 350, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 350, + 91, + 447, + 102 + ], + "type": "text", + "content": "Rejection Sampling IMLE" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 114, + 481, + 235 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 481, + 235 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 481, + 235 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 481, + 235 + ], + "type": "image", + "image_path": "1a22c15fe607f29d786b2669930a04268a35754286cd8b044c23f0129575445f.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 243, + 480, + 266 + ], + "lines": [ + { + "bbox": [ + 132, + 243, + 480, + 266 + ], + "spans": [ + { + "bbox": [ + 132, + 243, + 480, + 266 + ], + "type": "text", + "content": "Fig. 5: Qualitative comparison between our method and baselines. While analyzing the images, look for the sharpness of each image and diversity in the content of all images for a method." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 289, + 482, + 397 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 289, + 482, + 397 + ], + "spans": [ + { + "bbox": [ + 130, + 289, + 482, + 397 + ], + "type": "text", + "content": "As training progresses, we notice that for both the algorithms' samples move closer to the ground truth data points. However, for vanilla IMLE, we observe that for many data points the sample picked after the nearest neighbour search is already close to the ground truth. In this case, the loss associated with these data points would be low (indicated by the short length of arrow in the Figure-4c). As a result, the model trained by the vanilla IMLE objective does not learn anything significantly novel. In our proposed method (Figure-4d), each data point selects a sample that is at least " + }, + { + "bbox": [ + 130, + 289, + 482, + 397 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 130, + 289, + 482, + 397 + ], + "type": "text", + "content": " distance away from it. This ensures that the loss for each data point is always sufficiently high (indicated by the long arrows), resulting in meaningful updates to the model parameters." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 413, + 257, + 425 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 413, + 257, + 425 + ], + "spans": [ + { + "bbox": [ + 132, + 413, + 257, + 425 + ], + "type": "text", + "content": "3.6 Implementation Details" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 432, + 482, + 527 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 432, + 482, + 527 + ], + "spans": [ + { + "bbox": [ + 130, + 432, + 482, + 527 + ], + "type": "text", + "content": "Note that computing the distance of each sample with each data point is computationally expensive. Suppose we have " + }, + { + "bbox": [ + 130, + 432, + 482, + 527 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 130, + 432, + 482, + 527 + ], + "type": "text", + "content": " data points in " + }, + { + "bbox": [ + 130, + 432, + 482, + 527 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^d" + }, + { + "bbox": [ + 130, + 432, + 482, + 527 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 432, + 482, + 527 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 130, + 432, + 482, + 527 + ], + "type": "text", + "content": " samples, calculating the distance between each pair has a time complexity of " + }, + { + "bbox": [ + 130, + 432, + 482, + 527 + ], + "type": "inline_equation", + "content": "\\mathcal{O}(mnd)" + }, + { + "bbox": [ + 130, + 432, + 482, + 527 + ], + "type": "text", + "content": ". To get around this issue, we leverage a fast k-nearest neighbor search method, DCI [20]. This method reduces the runtime of a single query from linear to sublinear in " + }, + { + "bbox": [ + 130, + 432, + 482, + 527 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 130, + 432, + 482, + 527 + ], + "type": "text", + "content": ", enabling us to efficiently filter out all the samples that are within an " + }, + { + "bbox": [ + 130, + 432, + 482, + 527 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 130, + 432, + 482, + 527 + ], + "type": "text", + "content": " distance from any data point. Subsequently, from this filtered pool of remaining samples, we select, for each data point, the sample that is closest to it." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 528, + 482, + 588 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 528, + 482, + 588 + ], + "spans": [ + { + "bbox": [ + 130, + 528, + 482, + 588 + ], + "type": "text", + "content": "In order to reduce the search time complexity further, we can project the training data to a lower dimensional subspace (which would decrease " + }, + { + "bbox": [ + 130, + 528, + 482, + 588 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 130, + 528, + 482, + 588 + ], + "type": "text", + "content": "). We use this while implementing the procedure for image synthesis task: we first flatten each image and project it to a lower dimension by using random projection. We normalize these projected vectors before using them for nearest neighbour search." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 605, + 219, + 618 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 605, + 219, + 618 + ], + "spans": [ + { + "bbox": [ + 132, + 605, + 219, + 618 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "type": "text", + "content": "Datasets We assess our method and the baseline approaches across a variety of datasets with " + }, + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "type": "text", + "content": " resolution. These datasets include Animal-Face Dog [33], Animal-Face Cat [33], Obama [41], Panda [41], Grumpy-cat [41], Anime [24], Shells [24], Skulls [24]" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "type": "text", + "content": "C. Vashist et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 480, + 140 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 480, + 140 + ], + "type": "text", + "content": "and a subset of Flickr-FaceHQ (FFHQ) [12] which are standard datasets used in the few-shot learning literature." + } + ] + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 132, + 162, + 481, + 294 + ], + "blocks": [ + { + "bbox": [ + 132, + 162, + 481, + 294 + ], + "lines": [ + { + "bbox": [ + 132, + 162, + 481, + 294 + ], + "spans": [ + { + "bbox": [ + 132, + 162, + 481, + 294 + ], + "type": "table", + "html": "
DatasetFastGAN [24]FakeCLR [23]FreGAN [38]ReGAN [32]AdaIMLE [1]RS-IMLE (Ours)Imp. %
Obama41.129.933.445.725.014.044.0
Grumpy Cat26.620.624.927.319.111.539.8
Panda10.08.89.012.67.63.554.0
FFHQ-10054.262.150.587.433.212.961.1
Cat35.127.431.042.124.915.936.1
Dog50.744.447.957.243.023.146.3
Anime69.877.759.8110.865.835.845.6
Skulls109.6106.5163.3130.781.951.137.6
Shells120.9148.4169.3236.1108.555.448.9
", + "image_path": "b370376dac76132a7f35209bd835bac28d3057fc2de7e6c74bd9ceeb78db6275.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 295, + 481, + 316 + ], + "lines": [ + { + "bbox": [ + 132, + 295, + 481, + 316 + ], + "spans": [ + { + "bbox": [ + 132, + 295, + 481, + 316 + ], + "type": "text", + "content": "Table 1: We compute FID [8] between the real data and 5000 randomly generated samples for all the methods. Lower is better." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 132, + 332, + 480, + 367 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 332, + 480, + 367 + ], + "spans": [ + { + "bbox": [ + 132, + 332, + 480, + 367 + ], + "type": "text", + "content": "Baselines We compare our method to recent state-of-the-art few-shot image generation methods. These include FastGAN [24], FakeCLR [23], FreGAN [38], Re-GAN [32] and AdaIMLE [1]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 369, + 481, + 488 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 369, + 481, + 488 + ], + "spans": [ + { + "bbox": [ + 132, + 369, + 481, + 488 + ], + "type": "text", + "content": "Evaluation Metrics We employ Fréchet Inception Distance (FID) [8] to assess the perceptual quality of the generated images. This involves randomly generating 5000 images and calculating the FID between these generated samples and the real images for each dataset. Additionally, we evaluate the modelling accuracy and coverage by computing precision and recall for 1000 images using the metric defined by Kynkänniemi et al. [19]. In image synthesis, precision refers to the model's capacity to generate images closely resembling the desired target or distribution. Recall measures the model's ability to encompass a wide array of diverse images within the target distribution. To ensure that the computed metrics have low variance, we generate many more samples than training images." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 489, + 480, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 489, + 480, + 525 + ], + "spans": [ + { + "bbox": [ + 132, + 489, + 480, + 525 + ], + "type": "text", + "content": "Network Architecture We construct our generator network using decoder modules from VDVAE [3]. More details about the network architecture can be found in the Appendix." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 545, + 244, + 558 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 545, + 244, + 558 + ], + "spans": [ + { + "bbox": [ + 132, + 545, + 244, + 558 + ], + "type": "text", + "content": "4.1 Quantitative Results" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 568, + 480, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 568, + 480, + 616 + ], + "spans": [ + { + "bbox": [ + 132, + 568, + 480, + 616 + ], + "type": "text", + "content": "In Table 1, we present the FID scores computed for all the datasets across different methods. Lower FID scores indicate that the distribution of generated images is closer to the distribution of real images. Our method performs significantly better compared to baselines." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 617, + 480, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 617, + 480, + 664 + ], + "spans": [ + { + "bbox": [ + 132, + 617, + 480, + 664 + ], + "type": "text", + "content": "In Table 2, we show the precision and recall scores. Our method has a near perfect precision (close to 1), while having a significantly higher recall compared to the baselines. In the few cases where our method is not the best, it is very close to the best metric." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 350, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 350, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 350, + 91, + 447, + 102 + ], + "type": "text", + "content": "Rejection Sampling IMLE" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 133, + 200, + 200 + ], + "blocks": [ + { + "bbox": [ + 154, + 123, + 179, + 133 + ], + "lines": [ + { + "bbox": [ + 154, + 123, + 179, + 133 + ], + "spans": [ + { + "bbox": [ + 154, + 123, + 179, + 133 + ], + "type": "text", + "content": "Query" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 133, + 133, + 200, + 200 + ], + "lines": [ + { + "bbox": [ + 133, + 133, + 200, + 200 + ], + "spans": [ + { + "bbox": [ + 133, + 133, + 200, + 200 + ], + "type": "image", + "image_path": "c5051ad720e597f52683eaddb7faa491c58a7db82e16ce5edebd7f67e0294f10.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 151, + 201, + 182, + 209 + ], + "lines": [ + { + "bbox": [ + 151, + 201, + 182, + 209 + ], + "spans": [ + { + "bbox": [ + 151, + 201, + 182, + 209 + ], + "type": "text", + "content": "(a) Obama" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 201, + 133, + 269, + 200 + ], + "blocks": [ + { + "bbox": [ + 216, + 123, + 253, + 133 + ], + "lines": [ + { + "bbox": [ + 216, + 123, + 253, + 133 + ], + "spans": [ + { + "bbox": [ + 216, + 123, + 253, + 133 + ], + "type": "text", + "content": "RS-IMLE" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 201, + 133, + 269, + 200 + ], + "lines": [ + { + "bbox": [ + 201, + 133, + 269, + 200 + ], + "spans": [ + { + "bbox": [ + 201, + 133, + 269, + 200 + ], + "type": "image", + "image_path": "41f62913db116d5de28f06fdb521b7e5d6aff67f51d094d0b0427feb2a8fd2b1.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 269, + 133, + 337, + 200 + ], + "blocks": [ + { + "bbox": [ + 282, + 123, + 324, + 132 + ], + "lines": [ + { + "bbox": [ + 282, + 123, + 324, + 132 + ], + "spans": [ + { + "bbox": [ + 282, + 123, + 324, + 132 + ], + "type": "text", + "content": "Ada-IMLE" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 269, + 133, + 337, + 200 + ], + "lines": [ + { + "bbox": [ + 269, + 133, + 337, + 200 + ], + "spans": [ + { + "bbox": [ + 269, + 133, + 337, + 200 + ], + "type": "image", + "image_path": "5fb2b9c811adba02deb0be6ef649326622a2004b0006d35c89b87d9e8d09e4ba.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 337, + 133, + 405, + 200 + ], + "blocks": [ + { + "bbox": [ + 353, + 123, + 389, + 132 + ], + "lines": [ + { + "bbox": [ + 353, + 123, + 389, + 132 + ], + "spans": [ + { + "bbox": [ + 353, + 123, + 389, + 132 + ], + "type": "text", + "content": "FastGAN" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 337, + 133, + 405, + 200 + ], + "lines": [ + { + "bbox": [ + 337, + 133, + 405, + 200 + ], + "spans": [ + { + "bbox": [ + 337, + 133, + 405, + 200 + ], + "type": "image", + "image_path": "683bb6827b5df7121a87d723d805f452891b4d7b100c458023ee00921b181953.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 405, + 133, + 473, + 200 + ], + "blocks": [ + { + "bbox": [ + 422, + 123, + 455, + 132 + ], + "lines": [ + { + "bbox": [ + 422, + 123, + 455, + 132 + ], + "spans": [ + { + "bbox": [ + 422, + 123, + 455, + 132 + ], + "type": "text", + "content": "FreGAN" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 405, + 133, + 473, + 200 + ], + "lines": [ + { + "bbox": [ + 405, + 133, + 473, + 200 + ], + "spans": [ + { + "bbox": [ + 405, + 133, + 473, + 200 + ], + "type": "image", + "image_path": "6b200923da9adf0ac846d23121ceb7f4296563c504eea63745369d123eaa2d46.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 133, + 210, + 200, + 277 + ], + "blocks": [ + { + "bbox": [ + 133, + 210, + 200, + 277 + ], + "lines": [ + { + "bbox": [ + 133, + 210, + 200, + 277 + ], + "spans": [ + { + "bbox": [ + 133, + 210, + 200, + 277 + ], + "type": "image", + "image_path": "6af369e0484a58c880c9d69c7e3729b83aa1de9fc25a6fccf1740210496bf656.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 151, + 279, + 182, + 287 + ], + "lines": [ + { + "bbox": [ + 151, + 279, + 182, + 287 + ], + "spans": [ + { + "bbox": [ + 151, + 279, + 182, + 287 + ], + "type": "text", + "content": "(b) Anime" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 201, + 211, + 268, + 277 + ], + "blocks": [ + { + "bbox": [ + 201, + 211, + 268, + 277 + ], + "lines": [ + { + "bbox": [ + 201, + 211, + 268, + 277 + ], + "spans": [ + { + "bbox": [ + 201, + 211, + 268, + 277 + ], + "type": "image", + "image_path": "077d465dc3a703e95c0e7bd89ffaf91240843e4689860d57b7d1ca150e021853.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 269, + 211, + 337, + 277 + ], + "blocks": [ + { + "bbox": [ + 269, + 211, + 337, + 277 + ], + "lines": [ + { + "bbox": [ + 269, + 211, + 337, + 277 + ], + "spans": [ + { + "bbox": [ + 269, + 211, + 337, + 277 + ], + "type": "image", + "image_path": "ca64a04b7ca8cebcfd570d6e37e378c02c3266257779dcc8600c5e3f53f2484c.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 337, + 211, + 405, + 277 + ], + "blocks": [ + { + "bbox": [ + 337, + 211, + 405, + 277 + ], + "lines": [ + { + "bbox": [ + 337, + 211, + 405, + 277 + ], + "spans": [ + { + "bbox": [ + 337, + 211, + 405, + 277 + ], + "type": "image", + "image_path": "128aa9b30dbf8a0e78a685ceaded7af59c59fc5af28ea3060b6b7911d2c05970.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 405, + 211, + 473, + 277 + ], + "blocks": [ + { + "bbox": [ + 405, + 211, + 473, + 277 + ], + "lines": [ + { + "bbox": [ + 405, + 211, + 473, + 277 + ], + "spans": [ + { + "bbox": [ + 405, + 211, + 473, + 277 + ], + "type": "image", + "image_path": "cee15c41250a9d0627202729a2fb8adebf27f42166db826bd84e8bbdd43176b4.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 133, + 289, + 200, + 356 + ], + "blocks": [ + { + "bbox": [ + 133, + 289, + 200, + 356 + ], + "lines": [ + { + "bbox": [ + 133, + 289, + 200, + 356 + ], + "spans": [ + { + "bbox": [ + 133, + 289, + 200, + 356 + ], + "type": "image", + "image_path": "4d777c9d80dfb01159fcfca59e1925d711b8d11c3e889e6402a052d885e49e91.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 156, + 357, + 178, + 365 + ], + "lines": [ + { + "bbox": [ + 156, + 357, + 178, + 365 + ], + "spans": [ + { + "bbox": [ + 156, + 357, + 178, + 365 + ], + "type": "text", + "content": "(c) Dog" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 201, + 289, + 268, + 356 + ], + "blocks": [ + { + "bbox": [ + 201, + 289, + 268, + 356 + ], + "lines": [ + { + "bbox": [ + 201, + 289, + 268, + 356 + ], + "spans": [ + { + "bbox": [ + 201, + 289, + 268, + 356 + ], + "type": "image", + "image_path": "51d2bbc677929c4ea815a88ed34099ae39df7186f2078758a21f8e99b17686bb.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 269, + 289, + 337, + 356 + ], + "blocks": [ + { + "bbox": [ + 269, + 289, + 337, + 356 + ], + "lines": [ + { + "bbox": [ + 269, + 289, + 337, + 356 + ], + "spans": [ + { + "bbox": [ + 269, + 289, + 337, + 356 + ], + "type": "image", + "image_path": "77b861d33a7080f0351f389f5c61d14af04471244e424a4c77e23cc77b89e8d0.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 337, + 289, + 404, + 356 + ], + "blocks": [ + { + "bbox": [ + 337, + 289, + 404, + 356 + ], + "lines": [ + { + "bbox": [ + 337, + 289, + 404, + 356 + ], + "spans": [ + { + "bbox": [ + 337, + 289, + 404, + 356 + ], + "type": "image", + "image_path": "5beb677c92ec43c1eb5fe27dc68e09bda5977dea02cd6df50bb14e4636d1e332.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 405, + 289, + 473, + 356 + ], + "blocks": [ + { + "bbox": [ + 405, + 289, + 473, + 356 + ], + "lines": [ + { + "bbox": [ + 405, + 289, + 473, + 356 + ], + "spans": [ + { + "bbox": [ + 405, + 289, + 473, + 356 + ], + "type": "image", + "image_path": "17b6a23af1354eeeab9e9a8dbfa8ffa4c9323f61908f58ace4d9bb77628e01dc.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 133, + 366, + 200, + 433 + ], + "blocks": [ + { + "bbox": [ + 133, + 366, + 200, + 433 + ], + "lines": [ + { + "bbox": [ + 133, + 366, + 200, + 433 + ], + "spans": [ + { + "bbox": [ + 133, + 366, + 200, + 433 + ], + "type": "image", + "image_path": "966f34838f2101c1ad03df9322d86655c2c560d801012b2d2fd123aea4a05310.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 156, + 435, + 177, + 442 + ], + "lines": [ + { + "bbox": [ + 156, + 435, + 177, + 442 + ], + "spans": [ + { + "bbox": [ + 156, + 435, + 177, + 442 + ], + "type": "text", + "content": "(d) Cat" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_caption" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 201, + 366, + 268, + 433 + ], + "blocks": [ + { + "bbox": [ + 201, + 366, + 268, + 433 + ], + "lines": [ + { + "bbox": [ + 201, + 366, + 268, + 433 + ], + "spans": [ + { + "bbox": [ + 201, + 366, + 268, + 433 + ], + "type": "image", + "image_path": "fca2f453383f06982f87b9a6f1d4b0361885f3bb882b5c861b87cd3d0089b924.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 269, + 366, + 337, + 433 + ], + "blocks": [ + { + "bbox": [ + 269, + 366, + 337, + 433 + ], + "lines": [ + { + "bbox": [ + 269, + 366, + 337, + 433 + ], + "spans": [ + { + "bbox": [ + 269, + 366, + 337, + 433 + ], + "type": "image", + "image_path": "c51433fbd18b27ef6d0e8f1021f51805d2b6afb8b5bf2d18605072d5977ff4c9.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 337, + 366, + 404, + 433 + ], + "blocks": [ + { + "bbox": [ + 337, + 366, + 404, + 433 + ], + "lines": [ + { + "bbox": [ + 337, + 366, + 404, + 433 + ], + "spans": [ + { + "bbox": [ + 337, + 366, + 404, + 433 + ], + "type": "image", + "image_path": "4ba24b36d82ca8c8b9067f52f5232714da42e42ce9d216b613498555f7dea1eb.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 405, + 366, + 473, + 433 + ], + "blocks": [ + { + "bbox": [ + 405, + 366, + 473, + 433 + ], + "lines": [ + { + "bbox": [ + 405, + 366, + 473, + 433 + ], + "spans": [ + { + "bbox": [ + 405, + 366, + 473, + 433 + ], + "type": "image", + "image_path": "973b997d5d0ca7a6cea7acc841c4b7a6695cb850eaa68a237e3a86f42e7fc427.jpg" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_body" + } + ], + "index": 29 + }, + { + "type": "image", + "bbox": [ + 133, + 444, + 200, + 510 + ], + "blocks": [ + { + "bbox": [ + 133, + 444, + 200, + 510 + ], + "lines": [ + { + "bbox": [ + 133, + 444, + 200, + 510 + ], + "spans": [ + { + "bbox": [ + 133, + 444, + 200, + 510 + ], + "type": "image", + "image_path": "12ddc4b196240dec46d43d2dcaebc1773ab93c166f89ecf9dfb3b728ef763069.jpg" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_body" + } + ], + "index": 31 + }, + { + "type": "image", + "bbox": [ + 201, + 444, + 268, + 510 + ], + "blocks": [ + { + "bbox": [ + 201, + 444, + 268, + 510 + ], + "lines": [ + { + "bbox": [ + 201, + 444, + 268, + 510 + ], + "spans": [ + { + "bbox": [ + 201, + 444, + 268, + 510 + ], + "type": "image", + "image_path": "a5630c9161fd5940b774d161a462fb257558ad4796aa5c744a7c660a68c3af64.jpg" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_body" + } + ], + "index": 32 + }, + { + "type": "image", + "bbox": [ + 269, + 444, + 337, + 510 + ], + "blocks": [ + { + "bbox": [ + 269, + 444, + 337, + 510 + ], + "lines": [ + { + "bbox": [ + 269, + 444, + 337, + 510 + ], + "spans": [ + { + "bbox": [ + 269, + 444, + 337, + 510 + ], + "type": "image", + "image_path": "647485f4f25d709f43e2586ab5b283eb83ec53a1d320f78d15f54b12101eddfb.jpg" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_body" + } + ], + "index": 33 + }, + { + "type": "image", + "bbox": [ + 337, + 444, + 404, + 510 + ], + "blocks": [ + { + "bbox": [ + 337, + 444, + 404, + 510 + ], + "lines": [ + { + "bbox": [ + 337, + 444, + 404, + 510 + ], + "spans": [ + { + "bbox": [ + 337, + 444, + 404, + 510 + ], + "type": "image", + "image_path": "58423e93f87222b529343d8c398d32b5d61060e924a78d12440fa5686326d004.jpg" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_body" + } + ], + "index": 34 + }, + { + "type": "image", + "bbox": [ + 405, + 444, + 473, + 510 + ], + "blocks": [ + { + "bbox": [ + 405, + 444, + 473, + 510 + ], + "lines": [ + { + "bbox": [ + 405, + 444, + 473, + 510 + ], + "spans": [ + { + "bbox": [ + 405, + 444, + 473, + 510 + ], + "type": "image", + "image_path": "026cd53b87b41f419b0b7eec78e20307404bb716cbb298d7c83773023010ab16.jpg" + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_body" + } + ], + "index": 35 + }, + { + "type": "image", + "bbox": [ + 133, + 522, + 200, + 588 + ], + "blocks": [ + { + "bbox": [ + 153, + 513, + 181, + 520 + ], + "lines": [ + { + "bbox": [ + 153, + 513, + 181, + 520 + ], + "spans": [ + { + "bbox": [ + 153, + 513, + 181, + 520 + ], + "type": "text", + "content": "(e) Shells" + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 133, + 522, + 200, + 588 + ], + "lines": [ + { + "bbox": [ + 133, + 522, + 200, + 588 + ], + "spans": [ + { + "bbox": [ + 133, + 522, + 200, + 588 + ], + "type": "image", + "image_path": "379eb13e8b7523712cf23c77e53fd5c61e1748faa79b44bb5f501b8830d4b155.jpg" + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 146, + 590, + 187, + 598 + ], + "lines": [ + { + "bbox": [ + 146, + 590, + 187, + 598 + ], + "spans": [ + { + "bbox": [ + 146, + 590, + 187, + 598 + ], + "type": "text", + "content": "(f) FFHQ-100" + } + ] + } + ], + "index": 42, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 608, + 482, + 653 + ], + "lines": [ + { + "bbox": [ + 130, + 608, + 482, + 653 + ], + "spans": [ + { + "bbox": [ + 130, + 608, + 482, + 653 + ], + "type": "text", + "content": "Fig. 6: Visual Recall test. The first column is the query image from the dataset. Subsequent columns are the samples produced by different methods that are closest to the query image in LPIPS feature space. The samples produced by our method are closer to the query images compared to the baselines, while remaining diverse." + } + ] + } + ], + "index": 43, + "angle": 0, + "type": "image_caption" + } + ], + "index": 37 + }, + { + "type": "image", + "bbox": [ + 201, + 521, + 268, + 588 + ], + "blocks": [ + { + "bbox": [ + 201, + 521, + 268, + 588 + ], + "lines": [ + { + "bbox": [ + 201, + 521, + 268, + 588 + ], + "spans": [ + { + "bbox": [ + 201, + 521, + 268, + 588 + ], + "type": "image", + "image_path": "2ec14aef5e625847b1ffcae95889948bc3adc854c24d65751048e46ef14dc396.jpg" + } + ] + } + ], + "index": 38, + "angle": 0, + "type": "image_body" + } + ], + "index": 38 + }, + { + "type": "image", + "bbox": [ + 269, + 521, + 337, + 588 + ], + "blocks": [ + { + "bbox": [ + 269, + 521, + 337, + 588 + ], + "lines": [ + { + "bbox": [ + 269, + 521, + 337, + 588 + ], + "spans": [ + { + "bbox": [ + 269, + 521, + 337, + 588 + ], + "type": "image", + "image_path": "8f31b4dfb65b5566bfa3c2ad96aab41c7e3e1f41019babee95e343f0bb468c92.jpg" + } + ] + } + ], + "index": 39, + "angle": 0, + "type": "image_body" + } + ], + "index": 39 + }, + { + "type": "image", + "bbox": [ + 337, + 521, + 404, + 588 + ], + "blocks": [ + { + "bbox": [ + 337, + 521, + 404, + 588 + ], + "lines": [ + { + "bbox": [ + 337, + 521, + 404, + 588 + ], + "spans": [ + { + "bbox": [ + 337, + 521, + 404, + 588 + ], + "type": "image", + "image_path": "57f82d76e2db2bf8b179a80f8ffc0e22f0925808bb98fc31774d70b3fb7e26d6.jpg" + } + ] + } + ], + "index": 40, + "angle": 0, + "type": "image_body" + } + ], + "index": 40 + }, + { + "type": "image", + "bbox": [ + 405, + 521, + 473, + 588 + ], + "blocks": [ + { + "bbox": [ + 405, + 521, + 473, + 588 + ], + "lines": [ + { + "bbox": [ + 405, + 521, + 473, + 588 + ], + "spans": [ + { + "bbox": [ + 405, + 521, + 473, + 588 + ], + "type": "image", + "image_path": "68801163711ce49b64fc92ee1cb7188a38414a5ee0da640d684f6b22e2a8c303.jpg" + } + ] + } + ], + "index": 41, + "angle": 0, + "type": "image_body" + } + ], + "index": 41 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "type": "text", + "content": "C. Vashist et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 132, + 114, + 481, + 348 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 481, + 348 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 481, + 348 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 481, + 348 + ], + "type": "table", + "html": "
DatasetFastGAN [24]FakeCLR [23]FreGAN [38]ReGAN [32]AdaIMLE [1]RS-IMLE (Ours)
ObamaPrec.0.920.960.820.620.990.98
Rec.0.090.300.330.010.680.82
Grumpy CatPrec.0.910.970.900.780.970.93
Rec.0.130.390.230.040.720.95
PandaPrec.0.960.970.920.930.980.99
Rec.0.160.410.170.010.630.97
FFHQ-100Prec.0.910.710.860.390.991.00
Rec.0.130.250.210.010.770.99
CatPrec.0.970.990.950.900.980.96
Rec.0.080.550.310.150.860.98
DogPrec.0.960.950.920.840.970.98
Rec.0.190.340.200.100.610.94
AnimePrec.0.860.880.880.290.920.95
Rec.0.080.090.090.010.590.91
SkullsPrec.0.780.660.660.660.950.99
Rec.0.030.010.010.010.320.65
ShellsPrec.0.920.870.870.500.970.98
Rec.0.030.010.010.010.620.59
", + "image_path": "6c42fc3dddcf075a3dee1cc6c0a4a3912d9ad4b6a67c9b141cb92c45757e59c8.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 132, + 349, + 481, + 370 + ], + "lines": [ + { + "bbox": [ + 132, + 349, + 481, + 370 + ], + "spans": [ + { + "bbox": [ + 132, + 349, + 481, + 370 + ], + "type": "text", + "content": "Table 2: We compute precision and recall [42] between the real data and 1000 randomly generated samples for all the methods. Higher values are better." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 385, + 237, + 398 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 385, + 237, + 398 + ], + "spans": [ + { + "bbox": [ + 133, + 385, + 237, + 398 + ], + "type": "text", + "content": "4.2 Qualitative Results" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 407, + 481, + 539 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 407, + 481, + 539 + ], + "spans": [ + { + "bbox": [ + 132, + 407, + 481, + 539 + ], + "type": "text", + "content": "Figure 5 compares the random samples of our method to that of several baselines, and we observe that our method produces overall sharper and more diverse images. In addition, we propose Visual Recall, a simple test to substantiate the qualitative superiority of our method. For each method, we first generate 1000 samples. Next, using a real image from the dataset as a query, we find the images from the pool of generated samples that are closest to the query image. We use LPIPS features [39] for the computing the distance between real images and samples. Figure 6 shows the results for the proposed test for different datasets and methods. We see that the samples produced by our method are visually similar to the query, while being sharp and diverse in attributes like hair colour, smile and jaw structure. Note that other methods do not have samples that closely resemble the query image." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 540, + 481, + 588 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 540, + 481, + 588 + ], + "spans": [ + { + "bbox": [ + 132, + 540, + 481, + 588 + ], + "type": "text", + "content": "Figure 7 shows results of spherical linear interpolation between two random points in the latent space for different datasets. The images transition in a meaningful manner, indicating that our model has learnt a continuous and structured latent space representation of the image distribution." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 133, + 607, + 220, + 620 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 607, + 220, + 620 + ], + "spans": [ + { + "bbox": [ + 133, + 607, + 220, + 620 + ], + "type": "text", + "content": "4.3 Ablation Study" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 629, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 629, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 629, + 481, + 665 + ], + "type": "text", + "content": "Table-3 presents the FID computed for different values of " + }, + { + "bbox": [ + 132, + 629, + 481, + 665 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 132, + 629, + 481, + 665 + ], + "type": "text", + "content": " for three datasets. We observe that our approach works best for values of " + }, + { + "bbox": [ + 132, + 629, + 481, + 665 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 132, + 629, + 481, + 665 + ], + "type": "text", + "content": " close to 0.15 and that increasing the value of " + }, + { + "bbox": [ + 132, + 629, + 481, + 665 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 132, + 629, + 481, + 665 + ], + "type": "text", + "content": " beyond a certain range degrades the performance." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 350, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 350, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 350, + 91, + 447, + 102 + ], + "type": "text", + "content": "Rejection Sampling IMLE" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 471, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 471, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 471, + 91, + 480, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 132, + 114, + 481, + 354 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 481, + 354 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 481, + 354 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 481, + 354 + ], + "type": "image", + "image_path": "17b1ac2f0023e48788e1f9db75b7756b33d9db8e7dbba404405efade397f661b.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 292, + 356, + 321, + 363 + ], + "lines": [ + { + "bbox": [ + 292, + 356, + 321, + 363 + ], + "spans": [ + { + "bbox": [ + 292, + 356, + 321, + 363 + ], + "type": "text", + "content": "(c) Skulls" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 131, + 373, + 482, + 395 + ], + "lines": [ + { + "bbox": [ + 131, + 373, + 482, + 395 + ], + "spans": [ + { + "bbox": [ + 131, + 373, + 482, + 395 + ], + "type": "text", + "content": "Fig. 7: Latent space interpolation. We observe that the output images changes smoothly in a meaningful manner." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 184, + 406, + 430, + 462 + ], + "blocks": [ + { + "bbox": [ + 184, + 406, + 430, + 462 + ], + "lines": [ + { + "bbox": [ + 184, + 406, + 430, + 462 + ], + "spans": [ + { + "bbox": [ + 184, + 406, + 430, + 462 + ], + "type": "table", + "html": "
Dataset0.120.150.180.22
FFHQ-10013.612.9113.0114.8
Obama14.4414.0314.6214.34
Cat17.2915.9616.1216.27
", + "image_path": "08164c20d3d90fba49a992ac2f293eafa8d1538d206c8679b8910dc6d01d8c10.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 238, + 462, + 375, + 472 + ], + "lines": [ + { + "bbox": [ + 238, + 462, + 375, + 472 + ], + "spans": [ + { + "bbox": [ + 238, + 462, + 375, + 472 + ], + "type": "text", + "content": "Table 3: FID for different values of " + }, + { + "bbox": [ + 238, + 462, + 375, + 472 + ], + "type": "inline_equation", + "content": "\\epsilon" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 131, + 506, + 211, + 517 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 506, + 211, + 517 + ], + "spans": [ + { + "bbox": [ + 131, + 506, + 211, + 517 + ], + "type": "text", + "content": "5 Conclusion" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 531, + 482, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 531, + 482, + 602 + ], + "spans": [ + { + "bbox": [ + 130, + 531, + 482, + 602 + ], + "type": "text", + "content": "In this paper, we identified a latent space misalignment between the training and testing phases of existing IMLE-based methods, resulting in poor performance in few-shot image synthesis tasks. To address this issue, we introduced a novel algorithm, RS-IMLE, which modifies the prior distribution used for training. Our experimental results demonstrate that our method significantly enhances the quality of generated images and mode coverage during inference." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 614, + 482, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 614, + 482, + 651 + ], + "spans": [ + { + "bbox": [ + 130, + 614, + 482, + 651 + ], + "type": "text", + "content": "Acknowledgements: This research was enabled in part by support provided by NSERC, the BC DRI Group and the Digital Research Alliance of Canada. The authors would also like to thank Tristan Engst for extensive help polishing our paper." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 223, + 100 + ], + "type": "text", + "content": "C. Vashist et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 190, + 127 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 190, + 127 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 190, + 127 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 137, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 22, + "blocks": [ + { + "bbox": [ + 138, + 137, + 481, + 192 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 137, + 481, + 192 + ], + "spans": [ + { + "bbox": [ + 138, + 137, + 481, + 192 + ], + "type": "text", + "content": "1. Aghabozorgi, M., Peng, S., Li, K.: Adaptive IMLE for few-shot pretraining-free generative modelling. In: Krause, A., Brunskill, E., Cho, K., Engelhardt, B., Sabato, S., Scarlett, J. (eds.) Proceedings of the 40th International Conference on Machine Learning. Proceedings of Machine Learning Research, vol. 202, pp. 248-264. PMLR (23-29 Jul 2023), https://proceedings.mlr.press/v202/aghabozorgi23a.html" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 192, + 481, + 213 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 192, + 481, + 213 + ], + "spans": [ + { + "bbox": [ + 138, + 192, + 481, + 213 + ], + "type": "text", + "content": "2. Brock, A., Donahue, J., Simonyan, K.: Large scale gan training for high fidelity natural image synthesis. ArXiv abs/1809.11096 (2019)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 213, + 481, + 235 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 213, + 481, + 235 + ], + "spans": [ + { + "bbox": [ + 138, + 213, + 481, + 235 + ], + "type": "text", + "content": "3. Child, R.: Very deep vaes generalize autoregressive models and can outperform them on images. ArXiv abs/2011.10650 (2021)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 235, + 481, + 255 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 235, + 481, + 255 + ], + "spans": [ + { + "bbox": [ + 138, + 235, + 481, + 255 + ], + "type": "text", + "content": "4. Dhariwal, P., Nichol, A.: Diffusion models beat gans on image synthesis. ArXiv abs/2105.05233 (2021)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 256, + 481, + 277 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 256, + 481, + 277 + ], + "spans": [ + { + "bbox": [ + 138, + 256, + 481, + 277 + ], + "type": "text", + "content": "5. Dinh, L., Sohl-Dickstein, J.N., Bengio, S.: Density estimation using real nvp. ArXiv abs/1605.08803 (2017)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 277, + 481, + 310 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 277, + 481, + 310 + ], + "spans": [ + { + "bbox": [ + 138, + 277, + 481, + 310 + ], + "type": "text", + "content": "6. Esser, P., Rombach, R., Ommer, B.: Taming transformers for high-resolution image synthesis. 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) pp. 12868-12878 (2020), https://api_semanticscholar.org/CorpusID:229297973" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 310, + 481, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 310, + 481, + 342 + ], + "spans": [ + { + "bbox": [ + 138, + 310, + 481, + 342 + ], + "type": "text", + "content": "7. Goodfellow, I., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., Courville, A., Bengio, Y.: Generative adversarial nets. Advances in neural information processing systems 27 (2014)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 342, + 481, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 342, + 481, + 363 + ], + "spans": [ + { + "bbox": [ + 138, + 342, + 481, + 363 + ], + "type": "text", + "content": "8. Heusel, M., Ramsauer, H., Unterthiner, T., Nessler, B., Hochreiter, S.: Gans trained by a two time-scale update rule converge to a local nash equilibrium. In: NIPS (2017)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 363, + 481, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 363, + 481, + 384 + ], + "spans": [ + { + "bbox": [ + 138, + 363, + 481, + 384 + ], + "type": "text", + "content": "9. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. ArXiv abs/2006.11239 (2020)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 384, + 481, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 384, + 481, + 407 + ], + "spans": [ + { + "bbox": [ + 138, + 384, + 481, + 407 + ], + "type": "text", + "content": "0. Karras, T., Aittala, M., Laine, S., Harkonen, E., Hellsten, J., Lehtinen, J., Aila, T.: Alias-free generative adversarial networks. In: NeurIPS (2021)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 407, + 481, + 439 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 407, + 481, + 439 + ], + "spans": [ + { + "bbox": [ + 138, + 407, + 481, + 439 + ], + "type": "text", + "content": "1. Karras, T., Laine, S., Aila, T.: A style-based generator architecture for generative adversarial networks. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 4401-4410 (2019)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 439, + 481, + 460 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 439, + 481, + 460 + ], + "spans": [ + { + "bbox": [ + 138, + 439, + 481, + 460 + ], + "type": "text", + "content": "2. Karras, T., Laine, S., Aila, T.: A style-based generator architecture for generative adversarial networks (2019)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 460, + 481, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 460, + 481, + 502 + ], + "spans": [ + { + "bbox": [ + 138, + 460, + 481, + 502 + ], + "type": "text", + "content": "3. Karras, T., Laine, S., Aittala, M., Hellsten, J., Lehtinen, J., Aila, T.: Analyzing and improving the image quality of stylegan. 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) pp. 8107-8116 (2019), https://api_semanticscholar.org/CorpusID:209202273" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 502, + 481, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 502, + 481, + 536 + ], + "spans": [ + { + "bbox": [ + 138, + 502, + 481, + 536 + ], + "type": "text", + "content": "4. Karras, T., Laine, S., Aittala, M., Hellsten, J., Lehtinen, J., Aila, T.: Analyzing and improving the image quality of stylegan. 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) pp. 8107-8116 (2020)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 536, + 481, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 536, + 481, + 557 + ], + "spans": [ + { + "bbox": [ + 138, + 536, + 481, + 557 + ], + "type": "text", + "content": "5. Kingma, D.P., Dhariwal, P.: Glow: Generative flow with invertible 1x1 convolutions. ArXiv abs/1807.03039 (2018)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 557, + 481, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 557, + 481, + 578 + ], + "spans": [ + { + "bbox": [ + 138, + 557, + 481, + 578 + ], + "type": "text", + "content": "6. Kingma, D.P., Welling, M.: Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114 (2013)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 138, + 578, + 481, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 578, + 481, + 611 + ], + "spans": [ + { + "bbox": [ + 138, + 578, + 481, + 611 + ], + "type": "text", + "content": "7. Kobyzev, I., Prince, S., Brubaker, M.A.: Normalizing flows: An introduction and review of current methods. IEEE Transactions on Pattern Analysis and Machine Intelligence 43, 3964-3979 (2021)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 138, + 611, + 481, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 611, + 481, + 632 + ], + "spans": [ + { + "bbox": [ + 138, + 611, + 481, + 632 + ], + "type": "text", + "content": "8. Kong, C., Kim, J., Han, D., Kwak, N.: Few-shot image generation with mixup-based distance learning (2022)" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 138, + 632, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 632, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 632, + 481, + 665 + ], + "type": "text", + "content": "9. Kynkänniemi, T., Karras, T., Laine, S., Lehtinen, J., Aila, T.: Improved precision and recall metric for assessing generative models. Advances in Neural Information Processing Systems 32 (2019)" + } + ] + } + ], + "index": 21 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 350, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 350, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 350, + 91, + 447, + 102 + ], + "type": "text", + "content": "Rejection Sampling IMLE" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 471, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 471, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 471, + 92, + 480, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 117, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 24, + "blocks": [ + { + "bbox": [ + 132, + 117, + 418, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 117, + 418, + 127 + ], + "spans": [ + { + "bbox": [ + 132, + 117, + 418, + 127 + ], + "type": "text", + "content": "20. Li, K., Malik, J.: Fast k-nearest neighbour search via prioritized dci (2017)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 128, + 481, + 148 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 128, + 481, + 148 + ], + "spans": [ + { + "bbox": [ + 132, + 128, + 481, + 148 + ], + "type": "text", + "content": "21. Li, K., Malik, J.: Implicit maximum likelihood estimation. arXiv preprint arXiv:1809.09087 (2018)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 150, + 481, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 150, + 481, + 171 + ], + "spans": [ + { + "bbox": [ + 132, + 150, + 481, + 171 + ], + "type": "text", + "content": "22. Li, Y., Zhang, R., Lu, J., Shechtman, E.: Few-shot image generation with elastic weight consolidation. ArXiv abs/2012.02780 (2020)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 172, + 481, + 193 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 172, + 481, + 193 + ], + "spans": [ + { + "bbox": [ + 132, + 172, + 481, + 193 + ], + "type": "text", + "content": "23. Li, Z., Wang, C., Zheng, H., Zhang, J., Li, B.: Fakeclr: Exploring contrastive learning for solving latent discontinuity in data-efficient gans. In: ECCV (2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 194, + 481, + 225 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 194, + 481, + 225 + ], + "spans": [ + { + "bbox": [ + 132, + 194, + 481, + 225 + ], + "type": "text", + "content": "24. Liu, B., Zhu, Y., Song, K., Elgammal, A.: Towards faster and stabilized GAN training for high-fidelity few-shot image synthesis. CoRR abs/2101.04775 (2021), https://arxiv.org/abs/2101.04775" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 227, + 481, + 247 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 227, + 481, + 247 + ], + "spans": [ + { + "bbox": [ + 132, + 227, + 481, + 247 + ], + "type": "text", + "content": "25. Mo, S., Cho, M., Shin, J.: Freeze discriminator: A simple baseline for fine-tuning gans. ArXiv abs/2002.10964 (2020)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 249, + 481, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 249, + 481, + 281 + ], + "spans": [ + { + "bbox": [ + 132, + 249, + 481, + 281 + ], + "type": "text", + "content": "26. Ojha, U., Li, Y., Lu, J., Efros, A.A., Lee, Y.J., Shechtman, E., Zhang, R.: Few-shot image generation via cross-domain correspondence. 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) pp. 10738-10747 (2021)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 282, + 481, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 282, + 481, + 303 + ], + "spans": [ + { + "bbox": [ + 132, + 282, + 481, + 303 + ], + "type": "text", + "content": "27. van den Oord, A., Kalchbrenner, N., Espeholt, L., Kavukcuoglu, K., Vinyals, O., Graves, A.: Conditional image generation with pixelCNN decoders. ArXiv abs/1606.05328 (2016)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 304, + 481, + 324 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 304, + 481, + 324 + ], + "spans": [ + { + "bbox": [ + 132, + 304, + 481, + 324 + ], + "type": "text", + "content": "28. van den Oord, A., Kalchbrenner, N., Kavukcuoglu, K.: Pixel recurrent neural networks. ArXiv abs/1601.06759 (2016)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 326, + 481, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 326, + 481, + 346 + ], + "spans": [ + { + "bbox": [ + 132, + 326, + 481, + 346 + ], + "type": "text", + "content": "29. Razavi, A., van den Oord, A., Vinyals, O.: Generating diverse high-fidelity images with vqvae-2. ArXiv abs/1906.00446 (2019)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 348, + 481, + 379 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 348, + 481, + 379 + ], + "spans": [ + { + "bbox": [ + 132, + 348, + 481, + 379 + ], + "type": "text", + "content": "30. Salimans, T., Karpathy, A., Chen, X., Kingma, D.P.: PixelCNN++: Improving the pixelCNN with discretized logistic mixture likelihood and other modifications. ArXiv abs/1701.05517 (2017)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 380, + 481, + 411 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 380, + 481, + 411 + ], + "spans": [ + { + "bbox": [ + 132, + 380, + 481, + 411 + ], + "type": "text", + "content": "31. Sauer, A., Chitta, K., Muller, J., Geiger, A.: Projected gans converge faster. In: Neural Information Processing Systems (2021), https://api-semanticscholar.org/CorpusID: 240354401" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 414, + 481, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 414, + 481, + 445 + ], + "spans": [ + { + "bbox": [ + 132, + 414, + 481, + 445 + ], + "type": "text", + "content": "32. Saxena, D., Cao, J., Xu, J., Kulshrestha, T.: Re-gan: Data-efficient gans training via architectural reconfiguration. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 16230-16240 (June 2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 446, + 481, + 467 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 446, + 481, + 467 + ], + "spans": [ + { + "bbox": [ + 132, + 446, + 481, + 467 + ], + "type": "text", + "content": "33. Si, Z., Zhu, S.C.: Learning hybrid image templates (hit) by information projection. IEEE Transactions on Pattern Analysis and Machine Intelligence 34, 1354-1367 (2012)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 468, + 481, + 488 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 468, + 481, + 488 + ], + "spans": [ + { + "bbox": [ + 132, + 468, + 481, + 488 + ], + "type": "text", + "content": "34. Song, Y., Ermon, S.: Generative modeling by estimating gradients of the data distribution. ArXiv abs/1907.05600 (2019)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 490, + 481, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 490, + 481, + 511 + ], + "spans": [ + { + "bbox": [ + 132, + 490, + 481, + 511 + ], + "type": "text", + "content": "35. Song, Y., Sohl-Dickstein, J.N., Kingma, D.P., Kumar, A., Ermon, S., Poole, B.: Score-based generative modeling through stochastic differential equations. ArXiv abs/2011.13456 (2021)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 513, + 481, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 513, + 481, + 533 + ], + "spans": [ + { + "bbox": [ + 132, + 513, + 481, + 533 + ], + "type": "text", + "content": "36. Vahdat, A., Kautz, J.: Nvae: A deep hierarchical variational autoencoder. ArXiv abs/2007.03898 (2020)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 534, + 481, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 534, + 481, + 577 + ], + "spans": [ + { + "bbox": [ + 132, + 534, + 481, + 577 + ], + "type": "text", + "content": "37. Wang, Y., Gonzalez-Garcia, A., Berga, D., Herranz, L., Khan, F.S., van de Weijer, J.: Minegan: Effective knowledge transfer from gans to target domains with few images. 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) pp. 9329-9338 (2020)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 132, + 578, + 481, + 598 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 578, + 481, + 598 + ], + "spans": [ + { + "bbox": [ + 132, + 578, + 481, + 598 + ], + "type": "text", + "content": "38. Yang, M., Wang, Z., Chi, Z., Zhang, Y.: Fregan: Exploiting frequency components for training gans under limited data (2022)" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 132, + 600, + 481, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 600, + 481, + 620 + ], + "spans": [ + { + "bbox": [ + 132, + 600, + 481, + 620 + ], + "type": "text", + "content": "39. Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual metric. In: CVPR (2018)" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 132, + 622, + 481, + 642 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 622, + 481, + 642 + ], + "spans": [ + { + "bbox": [ + 132, + 622, + 481, + 642 + ], + "type": "text", + "content": "40. Zhao, M., Cong, Y., Carin, L.: On leveraging pretrained gans for generation with limited data. In: ICML (2020)" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 132, + 643, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 643, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 643, + 481, + 665 + ], + "type": "text", + "content": "41. Zhao, S., Liu, Z., Lin, J., Zhu, J.Y., Han, S.: Differentiable augmentation for data-efficient gan training. ArXiv abs/2006.10738 (2020)" + } + ] + } + ], + "index": 23 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 223, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 223, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 223, + 101 + ], + "type": "text", + "content": "C. Vashist et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 482, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 482, + 139 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 482, + 139 + ], + "type": "text", + "content": "42. Zhu, P., Abdal, R., Qin, Y., Wonka, P.: Improved stylegan embedding: Where are the good latents? ArXiv abs/2012.09036 (2020)" + } + ] + } + ], + "index": 2 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 350, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 350, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 350, + 91, + 447, + 102 + ], + "type": "text", + "content": "Rejection Sampling IMLE" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Relation DETR_ Exploring Explicit Position Relation Prior for Object Detection/650a9d12-9a9f-48d6-b77e-bd08f66392b3_content_list.json b/2024/Relation DETR_ Exploring Explicit Position Relation Prior for Object Detection/650a9d12-9a9f-48d6-b77e-bd08f66392b3_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..9091cb160f6042428059045f991f338b53e50dce --- /dev/null +++ b/2024/Relation DETR_ Exploring Explicit Position Relation Prior for Object Detection/650a9d12-9a9f-48d6-b77e-bd08f66392b3_content_list.json @@ -0,0 +1,1674 @@ +[ + { + "type": "text", + "text": "Relation DETR: Exploring Explicit Position Relation Prior for Object Detection", + "text_level": 1, + "bbox": [ + 250, + 140, + 751, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Xiuquan Hou $^{1}$ , Meiqin Liu $^{1,2,\\star}$ , Senlin Zhang $^{2}$ , Ping Wei $^{1}$ , Badong Chen $^{1}$ , and Xuguang Lan $^{1}$", + "bbox": [ + 238, + 215, + 761, + 247 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ National Key Laboratory of Human-Machine Hybrid Augmented Intelligence, Xi'an Jiaotong University, Xi'an 710049, China", + "bbox": [ + 215, + 257, + 785, + 287 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ College of Electrical Engineering, Zhejiang University, Hangzhou 310027, China \nxiuqhou@stu.xjtu.edu.cn, {liumeiqin, slzhang}@zju.edu.cn, {pingwei, chenbd, xglan} @mail.xjtu.edu.cn", + "bbox": [ + 228, + 286, + 772, + 329 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. This paper presents a general scheme for enhancing the convergence and performance of DETR (DEtection Transformer). We investigate the slow convergence problem in transformers from a new perspective, suggesting that it arises from the self-attention that introduces no structural bias over inputs. To address this issue, we explore incorporating position relation prior as attention bias to augment object detection, following the verification of its statistical significance using a proposed quantitative macroscopic correlation (MC) metric. Our approach, termed Relation-DETR, introduces an encoder to construct position relation embeddings for progressive attention refinement, which further extends the traditional streaming pipeline of DETR into a contrastive relation pipeline to address the conflicts between non-duplicate predictions and positive supervision. Extensive experiments on both generic and task-specific datasets demonstrate the effectiveness of our approach. Under the same configurations, Relation-DETR achieves a significant improvement $(+2.0\\%)$ AP compared to DINO), state-of-the-art performance $(51.7\\%)$ AP for $1\\times$ and $52.1\\%$ AP for $2\\times$ settings), and a remarkably faster convergence speed (over $40\\%$ AP with only 2 training epochs) than existing DETR detectors on COCO val2017. Moreover, the proposed relation encoder serves as a universal plug-in-and-play component, bringing clear improvements for theoretically any DETR-like methods. Furthermore, we introduce a class-agnostic detection dataset, SA-Det-100k. The experimental results on the dataset illustrate that the proposed explicit position relation achieves a clear improvement of $1.3\\%$ AP, highlighting its potential towards universal object detection. The code and dataset are available at https://github.com/xiuqhou/Relation-DETR.", + "bbox": [ + 261, + 397, + 738, + 758 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Detection transformer $\\cdot$ Object detection $\\cdot$ Relation network $\\cdot$ Progressive attention refinement $\\cdot$ Feature enhancement", + "bbox": [ + 261, + 773, + 740, + 803 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Corresponding author", + "bbox": [ + 220, + 824, + 380, + 839 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 143, + 374, + 160 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Object detection aims to tackle the problems of bounding box regression and object classification for each object of interest. Recently, DETection TRansformer (DETR) [4] has overcome the reliance on handcrafted designs of convolution detectors, achieving an elegant architecture in an end-to-end manner.", + "bbox": [ + 212, + 175, + 782, + 234 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Despite exhibiting impressive detection performance on large-scale datasets such as COCO [35], their performances are prone to be influenced by dataset scale and suffer from slow convergence. The root cause of the problem is the conflict between non-duplicate predictions and positive supervision [23]. During the training process, DETR employs the Hungarian algorithm to assign a single positive prediction to each ground-truth for producing unique results. However, this leads to negative predictions dominating the majority of the loss function, causing insufficient positive supervision. Therefore, more samples and iterations are required for convergence. Previous attempts have explored the issue by introducing train-only architectures (e.g. query denoising [28], multiple groups of queries [6], auxiliary queries [23], collaborative hybrid assignment training [52]) for additional supervision or by incorporating hard mining in loss functions (e.g. IA-BCE loss [3], position-supervised loss [37]). Other works have proposed specific structures for better interaction between queries and feature maps (e.g. dynamic anchor query [36], cascade window attention [46]), as well as techniques to focus on high-quality queries (e.g. hierarchical filtering [20], dense distinct process [49] and query rank layer [40]). Despite these advancements, there has been little exploration of the issue from the perspective of self-attention, which is widely used in the transformer decoders in most DETR detectors.", + "bbox": [ + 212, + 234, + 787, + 521 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The effectiveness of self-attention lies in its establishment of a high-dimensional relation representation among sequence embeddings [4, 42], which also serves as a key component for modeling relations among different detection feature representations [4]. However, such relation is an implicit representation since it assumes no structural bias over inputs, in which even position information is also needed to be learned from training data [33]. Consequently, the learning process of transformer is data-intensive and slow to converge. This analysis motivates us to introduce task-specific bias for realizing faster convergence and reducing data dependence.", + "bbox": [ + 212, + 522, + 790, + 657 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we explore enhancing DETR detectors from a novel perspective, namely explicit position relation prior. We first establish a metric for quantifying position relations in images, and analyze the distribution to verify its statistical significance. In this context, we introduce a position relation encoder to model all pairwise interactions between two bounding boxes, employing progressive attention refinement for cross-layer information interaction. To maintain the end-to-end property while providing sufficient positive supervision, we introduce a contrast relation strategy, which leverages both one-to-one and one-to-many matching while emphasizing the influence of position relation on dedduplication. The proposed method is named as Relation-DETR.", + "bbox": [ + 212, + 659, + 787, + 808 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Compared to previous works, the main feature of Relation-DETR is the integration of explicit position relation. In contrast, prior works focus on implicitly", + "bbox": [ + 212, + 809, + 785, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "X. Hou et al.", + "bbox": [ + 271, + 114, + 359, + 127 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "learned attention weights from training data, leading to slow convergence. Intuitively, our proposed position relation can be seen as a plug-in-and-play design beneficial for non-duplication predictions, since it establishes a representation of relative positions among pairs of bounding boxes (similar to IoU in NMS [16]).", + "bbox": [ + 212, + 146, + 782, + 205 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We evaluate the performance of Relation-DETR on the most popular object detection dataset, COCO 2017 [35], as well as several task-specific datasets [8,43]. The experimental results demonstrate its superior performance, surpassing previous state-of-the-art DETR detectors with clear margins. More specifically, Relation-DETR exhibits a significantly fast convergence speed. Without bells and whistles, it becomes the first DETR detector to achieve $40\\%$ AP on COCO with only 2 epochs using ResNet50 as the backbone under the $1\\times$ training configuration. In addition, the simple architecture design of our position relation encoder maintains a promising transferability. It can be easily extended to other DETR-based methods with only a few modifications to achieve consistent performance improvements. This is in contrast to some existing DETR detectors whose performance is highly dependent from complex matching strategies [22] or detection heads [52] developed by convolution-based detectors.", + "bbox": [ + 212, + 207, + 784, + 402 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 425, + 387, + 441 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Transformer for Object Detection In practice, the majority of attempts to apply transformer to object detection involve constructing a parallelizable sequence, either in the feature extractor [30] or in the detection body [4]. Specifically, transformer-based feature extractors generate token sequences based on image patches [12, 30, 39], and extract multi-scale features through aggregating local features [30, 39] or pyramid postprocess [14, 30]. DEtction TRansformer (DETR) proposed by Carion et al. [4] encodes the extracted features into object queries and decodes them into detected bounding boxes and labels. However, the self-learned attention mechanism increases the requirements for large-scale datasets and training iterations. Many works have explored the slow convergence from the perspective of structured attention (e.g. multi-scale deformable attention [51], dynamic attention [10], cascade window attention [46]), queries with explicit priors (e.g. anchor queries [44], dynamic anchor boxes queries [36], denoising queries [28], dense distinct queries [49]), and additional positive supervision (e.g. group queries [6], hybrid design [23], mixed matching [3]). However, even state-of-the-art DETR methods still utilize vanilla multi-head attention in the transformer decoder. And few works have explored the slow convergence from the perspective of implicit priors. This paper aims to address the issue with position relation.", + "bbox": [ + 212, + 455, + 787, + 743 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Relation Network Rather than processing visual features at pixel levels, patch levels or image levels, relation network captures relation features at instance levels. Existing research on relation networks involves category-based and instance-based approaches. The category-based approaches construct conceptual or statistical relations (e.g. co-occurrence probability [17, 27]) either from relation", + "bbox": [ + 212, + 763, + 785, + 840 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "Relation DETR", + "bbox": [ + 624, + 114, + 730, + 126 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 114, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "datasets like Visual Genome [24, 27, 45] or by adaptively learning from class labels [17]. Both of them, however, increase the complexity due to the assignment between instances and categories [7, 17, 45]. In contrast, instance-based approaches directly construct a fine-grained graph structure given object features as a node set and their relations as a edge set. Therefore, reasoning on the graph during the training process naturally determines the explicit relation weight [21]. Typically, the weight denotes the parametric distances between each paired object instances in high-dimensional space, such as appearance similarity [31], proposal distance [32] or even self-attention weight [42]. Since learning self-attention weight solely from training data without structural bias increases the requirement for dataset scale and iterations, we explore explicit position relation as a prior to reduce the requirement.", + "bbox": [ + 212, + 146, + 787, + 328 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Classification loss for hard mining. During object detection training, positive predictions assigned to ground truth are much fewer than negative predictions, often resulting in imbalanced supervision and slow convergence. For classification tasks, Focal Loss [34] proposes introducing a weight parameter to focus on hard samples, which is further extended to many variants like generalized focal loss (GFL) [29], vari focal loss (VFL) [29]. Moreover, for object detection tasks, using loss with modulation terms based on regression metrics (e.g. TOOD [15], IA-BCE [3], position-supervised loss [37]) further achieves high-quality alignment between classification and regression tasks.", + "bbox": [ + 212, + 349, + 787, + 488 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3 Statistical significance of object position relation", + "text_level": 1, + "bbox": [ + 215, + 510, + 723, + 529 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Are objects really correlated in object detection tasks? To answer the question, we propose a quantitative macroscopic correlation (MC) metric based on the Pearson Correlation Coefficient (PCC) to measure the position correlation among objects in a single image. Assume the objects in an image form a node set, and the PCC between each pair of bounding box annotations serves as their corresponding edge weight. We can construct an undirected graph with continuous values. In this end, the macroscopic correlation for each image can be calculated using the graph intensity, formulated as:", + "bbox": [ + 212, + 542, + 787, + 662 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nM C = \\frac {\\sum_ {i} \\sum_ {j : j \\neq i} | \\operatorname {P e a r s o n} \\left(\\boldsymbol {b} _ {i} , \\boldsymbol {b} _ {j}\\right) |}{N (N - 1)} \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 372, + 674, + 785, + 709 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $N$ denotes the number of objects, i.e. the number of nodes, $\\pmb{b} = [x,y,w,h]$ denotes the position annotation of bounding boxes in datasets. $MC = 1$ only when all objects are fully linearly correlated, whereas $MC = 0$ if there is no position correlation between any pair of objects.", + "bbox": [ + 212, + 718, + 787, + 779 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We visualize the statistical distribution of MC for datasets across various scenarios, including industrial (ESD [19], CSD [43], MSSD [8]), domestic (AI2thor [26]), urban (Cityscapes [9]) and generic settings (PascalVOC [13], COCO [35], Object365 [41], SA-1B [25]). The datasets cover a wide range of scales, from", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "X. Hou et al.", + "bbox": [ + 271, + 114, + 359, + 127 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/db1a7ca603db4b177adec9136dc6a526425d17c13b7216f61756190c50f7fc8e.jpg", + "image_caption": [ + "Fig. 1: Statistical distribution of macroscopic correlation (MC) on various datasets (normalized for better visualization), and the values in brackets indicate the number of dataset samples." + ], + "image_footnote": [], + "bbox": [ + 328, + 146, + 674, + 353 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "0.3K to 11M images. As shown in Fig. 1, all of these datasets indicate that the distribution of MC is concentrated around high numerical values, with the distribution centers closer to the upper bound. This demonstrates the presence and the statistical significance of object position relation. Specifically, task-specific datasets display more prior knowledge and clearer clustering patterns in the high-dimensional feature space, which thus results in higher MC values than generic datasets like COCO.", + "bbox": [ + 212, + 439, + 787, + 544 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4 Relation-DETR", + "text_level": 1, + "bbox": [ + 215, + 566, + 406, + 583 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Given the statistical significance of position relation, we propose a state-of-the-art detector, named Relation-DETR, that explores explicit position relation prior to enhance object detection. To address the issue of slow convergence, we present a position relation encoder (Sec. 4.1) for progressive attention refinement (Sec. 4.2). Further in Sec. 4.3, it extends the streaming pipeline of DETR into a contrast pipeline to emphasize the influence of position relation on removing duplication while maintaining sufficient positive supervision for faster convergence.", + "bbox": [ + 212, + 598, + 787, + 719 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1 Position relation encoder", + "text_level": 1, + "bbox": [ + 215, + 739, + 470, + 753 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Previous research has demonstrated the relation effectiveness for convolution detectors [21,32]. Recently, some DETR methods attempt to construct instance-level relation by indexing from category-level relation using class indices [17]. In contrast to these approaches, we directly construct instance-level relation through a simple position encoder, maintaining an end-to-end design for DETR.", + "bbox": [ + 212, + 763, + 787, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "Relation DETR", + "bbox": [ + 624, + 114, + 730, + 126 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 114, + 782, + 125 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We first review the basic pipeline in DETR detectors. Given image features extracted by the backbone, the transformer encoder produces an augmented memory $\\mathbf{Z} \\in \\mathbb{R}^{d \\times H \\times W}$ for further decoding into bounding boxes $\\pmb{b}_i = [x, y, w, h], i = 1, \\dots, N$ and class labels $\\pmb{c}_i$ as predictions. Each decoder layer refines the bounding box coordinates iteratively by predicting $\\Delta$ w.r.t. the coordinate from the last decoder layer, known as iterative bounding box refinement [51]. In addition, predictions from all decoder layers equally participate in loss calculation to compute auxiliary decoding losses [4].", + "bbox": [ + 212, + 146, + 787, + 267 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Under the aforementioned detection framework, our position relation encoder represents the high-dimensional relation embedding as an explicit prior for self-attention in the transformer. This embedding is calculated based on the predicted bounding boxes (denoted as $\\mathbf{b} = [x,y,w,h]$ ) from each decoder layer. To ensure that the relation is invariant to translation and scale transformations, we encode it based on normalized relative geometry features:", + "bbox": [ + 212, + 267, + 787, + 358 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {e} (\\boldsymbol {b} _ {i}, \\boldsymbol {b} _ {j}) = \\left[ \\log \\left(\\frac {| x _ {i} - x _ {j} |}{w _ {i}} + 1\\right), \\log \\left(\\frac {| y _ {i} - y _ {j} |}{h _ {i}} + 1\\right), \\log \\left(\\frac {w _ {i}}{w _ {j}}\\right), \\log \\left(\\frac {h _ {i}}{h _ {j}}\\right) \\right] (2)\n$$\n", + "text_format": "latex", + "bbox": [ + 228, + 362, + 785, + 397 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Unlike [21], our position relation is unbiased, as $\\pmb{e}(\\pmb{b}_i, \\pmb{b}_j) = 0$ when $i = j$ . The relation matrix $\\mathbf{E} \\in \\mathbb{R}^{N \\times N \\times 4}$ (with $\\mathbf{E}(i,j) = \\pmb{e}(\\pmb{b}_i, \\pmb{b}_j)$ ) is further transformed into high-dimensional embeddings through sine-cosine encoding [42].", + "bbox": [ + 214, + 401, + 787, + 446 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {E m b e d} (\\mathbf {E}, 2 k) = \\sin \\left(s \\mathbf {E} / T ^ {2 k / d _ {\\mathrm {r e}}}\\right) \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 395, + 453, + 784, + 477 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {E m b e d} (\\mathbf {E}, 2 k + 1) = \\cos (s \\mathbf {E} / T ^ {2 k / d _ {\\mathrm {r e}}}) \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 366, + 482, + 784, + 506 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where the shape of relation embedding is $N \\times N \\times 4d_{\\mathrm{re}}$ , and $T$ , $d_{\\mathrm{re}}$ , $s$ are encoding parameters. Finally, the embedding undergoes a linear transformation to obtain $M$ scalar weights, where $M$ denotes the number of attention heads.", + "bbox": [ + 214, + 510, + 787, + 555 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {R e l} (\\boldsymbol {b}, \\boldsymbol {b}) = \\max \\left(\\epsilon , \\mathbf {W E m b e d} (\\boldsymbol {b}, \\boldsymbol {b}) + \\mathbf {B}\\right) \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 357, + 561, + 784, + 578 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\epsilon$ makes sure a positive value for relation to avoid gradient vanishing after $\\exp$ when integrated into self-attention, and $\\mathsf{Rel}(\\pmb {b},\\pmb {b})\\in \\mathbb{R}^{N\\times N\\times M}$ .", + "bbox": [ + 214, + 584, + 787, + 614 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2 Progressive attention refinement with position relation", + "text_level": 1, + "bbox": [ + 214, + 632, + 714, + 648 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The iterative box refinement proposed by Deformable-DETR [51] has shown the effectiveness for high-quality bounding box regression. Following the motivation, we propose a progressive attention refinement method to introduce the position relation into the streaming pipeline of DETR. Specifically, the relation of layer- $i$ is determined by bounding boxes of both layer- $i - 1$ and layer- $i$ , which is further integrated into self-attention for producing the bounding boxes in layer- $(i + 1)$ .", + "bbox": [ + 212, + 652, + 787, + 744 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {A t t n} _ {\\text {S e l f}} \\left(\\mathbf {Q} ^ {l}\\right) = \\operatorname {S o f t m a x} \\left(\\operatorname {R e l} \\left(\\boldsymbol {b} ^ {l - 1}, \\boldsymbol {b} ^ {l}\\right) + \\frac {\\operatorname {Q u e} \\left(\\mathbf {Q} ^ {l}\\right) \\operatorname {K e y} \\left(\\mathbf {Q} ^ {l}\\right) ^ {\\top}}{\\sqrt {d _ {\\text {m o d e l}}}}\\right) \\operatorname {V a l} \\left(\\mathbf {Q} ^ {l}\\right) \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 228, + 761, + 784, + 795 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {Q} ^ {l + 1} = \\text {F F N} \\left(\\mathbf {Q} ^ {l} + \\operatorname {A t t n} _ {\\text {c r o s s}} \\left(\\operatorname {A t t n} _ {\\text {s e l f}} (\\mathbf {Q} ^ {l}), \\operatorname {K e y} (\\mathbf {Z}), \\operatorname {V a l} (\\mathbf {Z})\\right)\\right) \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 292, + 797, + 784, + 815 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {b} ^ {l + 1} = \\operatorname {M L P} \\left(\\mathbf {Q} ^ {l + 1}\\right), \\boldsymbol {c} ^ {l + 1} = \\operatorname {L i n e a r} \\left(\\mathbf {Q} ^ {l + 1}\\right) \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 295, + 818, + 784, + 837 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 227, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "X. Hou et al.", + "bbox": [ + 271, + 114, + 359, + 127 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\mathbf{Q}^l$ denotes the queries in the $l$ -th decoder layer in the DETR transformer, $\\mathbf{Z}$ is the memory, i.e., the enhanced image features from the transformer encoder.", + "bbox": [ + 214, + 145, + 782, + 176 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b279f4e719153e33370759b926e9c57cda4e1343abbfd03d603bdc9eb04ad97d.jpg", + "image_caption": [ + "Fig. 2: Comparison of transformer decoder in Deformable-DETR(left) and Relation-DETR(right)." + ], + "image_footnote": [], + "bbox": [ + 225, + 204, + 498, + 306 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a1b9f1abfc871fc81b2ed4376a5577eea72c11ba94458b93f22e7b8ce7668231.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 504, + 202, + 779, + 305 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The only difference between our method and the existing DETR decoder is marked in red. As depicted in Fig. 2, the requisite addition involves the introduction of a lateral branch for the computation of position relation. Therefore, our position relation and the progressive attention refinement are straightforward, allowing for plug-in-and-play integration with self attention in existing DETR detectors to achieve consistent performance improvements (see Tab. 6).", + "bbox": [ + 212, + 378, + 784, + 470 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3 Contrast relation pipeline", + "text_level": 1, + "bbox": [ + 214, + 496, + 473, + 512 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Rethinking the mechanism of existing duplication removal methods (including NMS [16], Soft-NMS [1], fast-NMS [2], Adaptive-NMS [38]), these processes heavily rely on IoU (intersection over Union), which, to some extent, signifies the position relation between bounding boxes. Therefore, we may hypothesize that integrating the position relation among queries in self-attention contributes to non-duplicated predictions in object detection, akin to [22].", + "bbox": [ + 212, + 525, + 784, + 616 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The conflicts between non-duplicate predictions and sufficient positive supervision arise from the streaming pipeline of DETR, which must navigate between one-to-one matching and one-to-many matching. To overcome this limitation, we extend it to a contrast pipeline based on the proposed position relation. Specifically, we construct two parallel sets of queries, i.e. matching queries $\\mathbf{Q}_m$ and hybrid queries $\\mathbf{Q}_h$ . Both are input into the transformer decoder but undergo distinct processing. The matching queries are processed with self-attention incorporating position relation to produce non-duplicated predictions:", + "bbox": [ + 212, + 617, + 784, + 738 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {A t t n} _ {\\text {S e l f}} \\left(\\mathbf {Q} _ {\\mathrm {m}} ^ {l}\\right) = \\operatorname {S o f t m a x} \\left(\\operatorname {R e l} \\left(\\boldsymbol {b} ^ {l - 1}, \\boldsymbol {b} ^ {l}\\right) + \\frac {\\operatorname {Q u e} \\left(\\mathbf {Q} _ {\\mathrm {m}}\\right) \\operatorname {K e y} \\left(\\mathbf {Q} _ {\\mathrm {m}}\\right) ^ {\\top}}{\\sqrt {d _ {\\text {m o d e l}}}}\\right) \\operatorname {V a l} \\left(\\mathbf {Q} _ {\\mathrm {m}}\\right) \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 217, + 750, + 784, + 800 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {A t t n} _ {\\text {S e l f}} \\left(\\mathbf {Q} _ {\\mathrm {h}} ^ {l}\\right) = \\operatorname {S o f t m a x} \\left(\\frac {\\operatorname {Q u e} \\left(\\mathbf {Q} _ {\\mathrm {h}}\\right) \\operatorname {K e y} \\left(\\mathbf {Q} _ {\\mathrm {h}}\\right) ^ {\\top}}{\\sqrt {d _ {\\text {m o d e l}}}}\\right) \\operatorname {V a l} \\left(\\mathbf {Q} _ {\\mathrm {h}}\\right) \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 222, + 801, + 784, + 838 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "Relation DETR", + "bbox": [ + 622, + 114, + 730, + 126 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 114, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "while the hybrid queries are decoded by the same decoder but skip the calculation of position relation to explore more potential candidates. Their corresponding predictions are denoted as $\\pmb{p}_m^l = (\\pmb{b}_m^l, \\pmb{c}_m^l)$ and $\\pmb{p}_h^l = (\\pmb{b}_h^l, \\pmb{c}_h^l)$ , respectively. Details of the contrast relation pipeline are illustrated in Fig. 3.", + "bbox": [ + 212, + 146, + 787, + 208 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/734495d4255b2af0c1253dff6b4627e00d65b7adb28330f26d43e0df3546b13c.jpg", + "image_caption": [ + "Fig. 3: Detailed illustration of the proposed contrast relation pipeline." + ], + "image_footnote": [], + "bbox": [ + 313, + 234, + 571, + 406 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/a28699bd84ffaf677331794c6cd14f2ea24e6f3e54c3b984eef3b7be26635ca4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 581, + 234, + 709, + 406 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Assuming $\\pmb{g}$ denotes the ground truth annotations, for $\\pmb{p}_m$ , we employ a one-to-one matching scheme to emphasize the non-duplicate property, with loss calculation similar to the original DETR approach [4]:", + "bbox": [ + 212, + 465, + 787, + 513 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {m} \\left(\\boldsymbol {p} _ {m}, \\boldsymbol {g}\\right) = \\sum_ {l = 1} ^ {L} \\mathcal {L} _ {\\text {H u n g a r i a n}} \\left(\\boldsymbol {p} _ {m} ^ {l}, \\boldsymbol {g}\\right) \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 374, + 523, + 784, + 564 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "While for $\\pmb{p}_h$ , a one-to-many matching scheme is employed to form more potential positive candidates. We simply follow H-DETR [23] and repeat the ground truth $K$ times, denoted as $\\tilde{\\pmb{g}} = \\{\\pmb{g}^1, \\pmb{g}^2, \\dots, \\pmb{g}^K\\}$ , for loss calculation:", + "bbox": [ + 214, + 574, + 787, + 622 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {h} \\left(\\boldsymbol {p} _ {h}, \\boldsymbol {g}\\right) = \\sum_ {l = 1} ^ {L} \\mathcal {L} _ {\\text {H u n g a r i a n}} \\left(\\boldsymbol {p} _ {h} ^ {l}, \\tilde {\\boldsymbol {g}}\\right) \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 380, + 632, + 784, + 674 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $\\mathcal{L}_{\\mathrm{Hungarian}}$ denotes the Hungarian loss, and $L$ denotes the number of decoder layers. It is worth noting that the hybrid queries are only involved during training, thus incurring no extra computational burden for inference.", + "bbox": [ + 212, + 684, + 787, + 731 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5 Experimental Results and Discussion", + "text_level": 1, + "bbox": [ + 214, + 752, + 609, + 771 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.1 Setup", + "text_level": 1, + "bbox": [ + 214, + 784, + 310, + 800 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "For a comprehensive evaluation, we conduct experiments on both an object detection benchmark, COCO 2017 [35], and two task-specific datasets, CSD [43]", + "bbox": [ + 212, + 809, + 785, + 842 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 227, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "X. Hou et al.", + "bbox": [ + 271, + 114, + 359, + 127 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "and MSSD [8]. Detection performance is measured using the standard Average Precision (AP) [35]. We train our model on NVIDIA A800 GPU (80GB) and RTX 3090 GPU (24GB) using the AdamW optimizer with an initial learning rate of $1 \\times 10^{-4}$ and weight decay of $1 \\times 10^{-4}$ . The Relation-DETR is implemented based on the backbone ResNet-50 [18] and Swin-L [39] pretrained from ImageNet [11], which are finetuned with a learning rate $1 \\times 10^{-5}$ during training. The learning rate is reduced by a factor of 0.1 at later stages. The parameters of the position relation encoder, $T$ , $d_{\\mathrm{re}}$ , $s$ are empirically chosen as 10000, 16, 100, respectively. The hybrid training configurations follow DINO [47] and H-DETR [23], i.e., $N_{m} = 900$ , $N_{h} = 1500$ , $k = 6$ . We adopt the VariFocal loss [48] for training Relation-DETR. The training batch size is 10 for COCO 2017 and 2 for task-specific datasets. Before fed into detectors, images undergo the same augmentations (random resize, crop and flip) as other DETR detectors.", + "bbox": [ + 212, + 145, + 787, + 344 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "5.2 Comparison with state-of-the-art methods", + "text_level": 1, + "bbox": [ + 214, + 377, + 611, + 393 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Comparison on COCO 2017. Tab. 1 presents the detection performance on COCO val 2017. Compared to other state-of-the-art DETR methods, our approach converges much faster and demonstrates significant improvements of $1.0\\%$ AP, $1.0\\% \\mathrm{AP}_{50}$ and $0.6\\% \\mathrm{AP}_{75}$ , respectively, suppressing the second best DDQ-DETR [49] with clear margins. Specifically, Relation-DETR achieves $51.7\\%$ AP using only 12 epochs with the ResNet-50 backbone and even outperforms DINO [47] (51.2% AP) with 36 epochs ( $3\\times$ faster). More importantly, in contrast to DDQ-DETR [49] and Co-DETR [52] that leverages NMS in the decoder or postprocess for improving precision, our Relation-DETR maintains an end-to-end pipeline, ensuring promising extensibility. When integrated with Swin-L backbone, Relation-DETR outperforms all counterparts, achieving the best $57.8\\%$ AP with a $0.5\\%$ AP improvement, showcasing its excellent scalability for larger model capacity.", + "bbox": [ + 212, + 412, + 787, + 609 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Comparison on task-specific datasets. Different from generic object detection benchmarks, datasets in task-specific scenarios lack sufficient samples to provide semantic information. To reveal the generalizability of Relation-DETR, we conduct a performance comparison on two defect detection datasets, i.e. CSD [43] and MSSD [8]. The results in Tab. 3 show that, Relation-DETR improves the baseline DINO by $1.4\\%$ AP on CSD, achieving the highest $54.4\\%$ AP. Tab. 4 demonstrates that Relation-DETR further increases the margin to $6.4\\%$ AP on MSSD and surpasses other counterparts. It is noteworthy that CSD [43] and MSSD [8] contain more small-sized objects than COCO 2017, which thus confirms the effectiveness of Relation-DETR to small-sized detection. Moreover, considering a stricter IoU threshold, Relation-DETR outperforms the second best method DINO by a significant margin of $11.1\\%$ AP@75, highlighting the beneficial impact of explicit position relation on high-quality predictions.", + "bbox": [ + 212, + 643, + 787, + 840 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "Relation DETR", + "bbox": [ + 622, + 113, + 730, + 127 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 114, + 785, + 127 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/f8170a23173dc2b54096e83c25ce82a3e07cf74c6402036e09f38cc1cf36ec53.jpg", + "table_caption": [ + "Table 1: Comparison with state-of-the-art methods on COCO val2017 using ResNet-50(IN-1K) backbone. The * means that we re-implement the methods and report the corresponding results." + ], + "table_footnote": [], + "table_body": "
MethodBackboneEpochsAP@50:95AP@50AP@75\\( AP_S \\)\\( AP_M \\)\\( AP_L \\)
Anchor-DETR [44]ResNet-505042.163.144.922.346.260.0
Def-DETR* [51]ResNet-501245.465.049.127.249.661.0
DAB-Def-DETR* [36]ResNet-501246.365.950.429.350.561.9
DN-Def-DETR* [28]ResNet-501247.265.351.330.250.662.4
DINO* [47]ResNet-501249.967.454.533.953.563.8
Group-DETR [6]ResNet-501249.832.453.064.2
H-Def-DETR [23]ResNet-501248.766.452.931.251.563.5
Cascade-DETR [46]ResNet-501249.767.154.132.453.565.1
Co-Def-DETR [52]ResNet-501249.567.654.332.452.763.7
Align-DETR [3]ResNet-501250.267.854.432.953.365.0
Stable-DINO [37]ResNet-501250.467.455.032.954.065.5
DAC-DETR [22]ResNet-501250.067.654.7
Salience-DETR* [20]ResNet-501250.067.754.233.354.464.4
Rank-DETR [40]ResNet-501250.467.955.233.653.864.2
MS-DETR [50]ResNet-501250.367.455.132.754.064.6
DDQ-DETR [49]ResNet-501250.768.155.7
Relation-DETRResNet-501251.769.156.336.155.666.1
DINO [47]ResNet-502450.468.354.833.353.764.8
DINO [47]ResNet-503651.269.055.835.054.365.3
H-Def-DETR [23]ResNet-503650.068.354.432.952.765.3
Salience-DETR* [20]ResNet-502451.268.955.733.955.565.6
DDQ-DETR [49]ResNet-502452.069.557.235.254.965.9
Relation-DETRResNet-502452.169.756.636.156.066.5
", + "bbox": [ + 215, + 199, + 782, + 554 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/f36de4ba72d5cbcfc49048f63c765dc650e27f9d3a40f7dc30fc9160efc394f1.jpg", + "table_caption": [ + "Table 2: Comparison with state-of-the-art methods on COCO val2017 using SwinL(IN-22K) as the backbone." + ], + "table_footnote": [], + "table_body": "
MethodBackboneEpochsAP@50:95AP@50AP@75\\( AP_S \\)\\( AP_M \\)\\( AP_L \\)
DINO [47]Swin-L1256.875.462.341.160.673.5
\\( \\mathcal{H} \\text{-Def-DETR} \\) [23]Swin-L1255.975.261.039.159.972.2
Salience-DETR* [20]Swin-L1256.575.061.540.261.272.8
Rank-DETR [40]Swin-L1257.375.962.940.861.373.2
Rank-DINO [40]Swin-L1257.676.063.441.661.473.8
Relation-DETRSwin-L1257.876.162.941.262.174.4
", + "bbox": [ + 215, + 609, + 790, + 723 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "5.3 Ablation study", + "text_level": 1, + "bbox": [ + 215, + 752, + 388, + 767 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "This part conducts an ablation study to explore how the proposed components influence the final detection performance on COCO. The results in Tab. 5 show that, each key component of Relation-DETR consistently contributes to improving AP. Even on a highly-optimized baseline with VariFocal loss [48], our posi", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "X. Hou et al.", + "bbox": [ + 271, + 114, + 359, + 126 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/ac08523168598cc8b88173451226749d79fbf128876968a0b59dd016beae9ad3.jpg", + "table_caption": [ + "Table 3: Quantitative comparison on CSD [43]." + ], + "table_footnote": [], + "table_body": "
MethodBackboneEpochsAP@50:95AP@50AP@75\\( AP_S \\)\\( AP_M \\)\\( AP_L \\)
Def-DETR [51]ResNet-5030043.786.236.640.534.910.0
DAB-Def-DETR [36]ResNet-509052.991.255.050.339.40.0
DN-Def-DETR [28]ResNet-506049.988.051.247.637.70.0
DINO [47]ResNet-506053.090.855.550.939.60.0
H-Def-DETR [23]ResNet-506053.090.655.751.239.26.7
Salience-DETR [20]ResNet-506053.292.555.151.040.90.0
Relation-DETRResNet-506054.492.956.053.640.80.0
", + "bbox": [ + 222, + 170, + 779, + 299 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/c4d7acb802e46d98cf849eee8b8b595f1d3add52e351b0e945597fcce8942b9d.jpg", + "table_caption": [ + "Table 4: Quantitative comparison on MSSD [8]." + ], + "table_footnote": [], + "table_body": "
MethodBackboneEpochsAP@50:95AP@50AP@75APSAPMAPL
Def-DETR [51]ResNet-5030033.054.332.09.811.133.3
DAB-Def-DETR [36]ResNet-5012033.760.031.015.926.729.0
DN-Def-DETR [28]ResNet-5012045.674.144.918.631.941.0
H-Def-DETR [23]ResNet-5012046.976.847.120.345.340.7
DINO [47]ResNet-5012051.080.052.520.047.444.8
Salience-DETR [20]ResNet-5012055.478.261.928.747.544.5
Relation-DETRResNet-5012057.479.263.631.453.545.5
", + "bbox": [ + 222, + 354, + 779, + 483 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "tion relation encoder and contrast pipeline bring clear improvements of $+0.3\\%$ , $+0.5\\%$ AP, respectively. Built upon normalized relative geometry features, the position relation overcomes scale bias effectively, thus benefiting consistent performance improvements for different sized objects. For instance, Tab. 5 shows that introducing relation into the baseline with VFL achieves $+1.2\\%$ $\\mathrm{AP}_S$ , $+1.0\\%$ $\\mathrm{AP}_M$ , and $+1.3\\%$ $\\mathrm{AP}_L$ .", + "bbox": [ + 212, + 526, + 787, + 619 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/18c19cd800f55af37a5bef721b5da00ab334d28f74bfaf5ebdaf7a4c915e95fd.jpg", + "table_caption": [ + "Table 5: Ablation study on key components of Relation-DETR (ResNet-50, 1x). We study both baseline and an optimized version with VariFocal Loss [48]." + ], + "table_footnote": [], + "table_body": "
VFLRel.contrastAP@50:95AP@50AP@75APSAPMAPL
49.967.454.533.953.563.8
✓50.3(↑0.4)68.1(↑0.7)55.0(↑0.5)34.7(↑0.8)53.4(↓0.1)64.7(↑0.9)
✓50.9(↑0.6)68.2(↑0.1)55.3(↑0.3)34.9(↑0.2)54.6(↑1.2)64.8(↑0.1)
✓✓51.2(↑0.3)68.5(↑0.7)55.7(↑0.4)35.3(↑0.4)54.9(↑0.3)65.5(↑0.7)
✓✓✓51.7(↑0.5)69.1(↑0.6)56.3(↑0.6)36.1(↑0.8)55.6(↑0.7)66.1(↑0.6)
", + "bbox": [ + 215, + 700, + 787, + 801 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "Relation DETR", + "bbox": [ + 622, + 114, + 730, + 126 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "5.4 Transferability of position relation", + "text_level": 1, + "bbox": [ + 215, + 146, + 542, + 161 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Our position relation encoder adopts an elegant architectural design, ensuring a promising transferability to existing DETR detectors with minimal modifications. The experimental results in Tab. 6 show that, integrating position relation encoders without any further modifications enhances detection performance with clear margins of $1.6\\%$ , $2.0\\%$ , $0.1\\%$ and $0.2\\%$ for Deformable-DETR [51], DAB-Deformable-DETR [36], DN-Deformable-DETR [28] and DINO [47], respectively. Interestingly, in comparison to $\\mathrm{AP}_M$ and $\\mathrm{AP}_L$ , the position relation has a more substantial impact on improving $\\mathrm{AP}_S$ for Deformable-DETR [51] and DAB-Deformable-DETR [36]. We attribute this to the fact that these early proposed baselines introduce relatively less structural bias and thus benefit more from our explicit position relation prior.", + "bbox": [ + 212, + 174, + 787, + 340 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/b0771129a444c8f39f822ea8f9497ba25398f0e7507322598cb3dd92377e566d.jpg", + "table_caption": [ + "Table 6: Transfer experiments for the position relation encoder (ResNet-50, $1 \\times$ ). \"+RelEnc\" denotes the version that integrates our position relation encoder." + ], + "table_footnote": [], + "table_body": "
MethodAP@50:95AP@50AP@75APSAPMAPL
Def-DETR45.465.049.127.249.661.0
+RelEnc47.0(↑1.6)65.6(↑0.6)51.2(↑2.1)29.3(↑2.1)51.0(↑1.4)62.2(↑1.2)
DAB-Def-DETR46.365.950.428.350.561.9
+RelEnc48.3(↑2.0)66.5(↑0.6)52.9(↑2.5)32.4(↑4.1)52.0(↑1.5)62.0(↑0.1)
DN-Def-DETR47.265.351.330.250.662.4
+RelEnc47.3(↑0.1)65.6(↑0.3)51.4(↑0.1)29.9(↓0.3)50.8(↑0.2)62.1(↓0.3)
", + "bbox": [ + 215, + 407, + 779, + 536 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Moreover, the proposed contrast pipeline can be seen as an extension of hybrid matching [23], utilizing the proposed position relation encoder. Table 7 compares their transferability when integrated into DINO. The results indicate that directly applying hybrid matching [23] to DINO [47] results in a decrease in performance, from $49.9\\%$ AP to $49.5\\%$ AP. In contrast, the introduction of both the proposed relation encoder and the extended contrast pipeline consistently increases the performance. This demonstrates the effectiveness of the proposed position relation prior in improve detection performance, overcoming the weak generalizability inherent in hybrid matching.", + "bbox": [ + 212, + 566, + 787, + 703 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/4c0ba0484f07d052f7280e31f669de40c39126035cfa95ec20f186631cfa4b5d.jpg", + "table_caption": [ + "Table 7: Transfer experiments compared with hybrid matching [23] on DINO." + ], + "table_footnote": [ + "† We found that hybrid matching decreases the performance of DINO. The conclusion is consistent with the results of HDINO (see here) reported by MMDetection [5]." + ], + "table_body": "
DINO+Hybrid [23]+RelEnc+RelEnc+Contrast
APAP@50AP@75APAP@50AP@75APAP@50AP@75APAP@50AP@75
49.967.454.549.5†66.654.050.368.155.051.268.755.6
", + "bbox": [ + 215, + 755, + 792, + 813 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "X. Hou et al.", + "bbox": [ + 271, + 114, + 359, + 126 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "5.5 Intuitive performance comparison", + "text_level": 1, + "bbox": [ + 215, + 146, + 540, + 161 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "To facilitate an intuitive performance comparison, Fig. 4 plots the convergence curve and the precision-recall curve. Because the position relation prior reduces the requirement for learning structural bias from data [33], Relation-DETR illustrates a faster convergence speed. When training from scratch, it achieves a higher AP than other counterparts with fewer iterations. Specifically, Relation-DETR can achieve over $40\\%$ AP with only 2 epochs, surpassing existing DETR detectors. In addition to convergence speed, the PR curves under different IoU thresholds also validate the performance gain of our Relation-DETR.", + "bbox": [ + 212, + 167, + 787, + 289 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/237f844b9eacd221aacb342100b5e9d2a9a39929a679215ff5dc8119708deed5.jpg", + "image_caption": [ + "Fig. 4: Convergence curve(left) and Precision-recall curve for $\\mathrm{IoU} = 50\\% \\sim 95\\%$ (right). All models are trained with ResNet-50 backbone under the same $1\\times$ training configuration on COCO 2017." + ], + "image_footnote": [], + "bbox": [ + 228, + 314, + 498, + 476 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/c90bae7c9bf3079eb86772dc3acbcda46503e06a5baa859fd2e881fa1d160ffd.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 504, + 315, + 774, + 474 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "5.6 Visualization", + "text_level": 1, + "bbox": [ + 215, + 580, + 370, + 593 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "For a more intuitive grasp of the relation mechanism, Fig. 5 illustrates representative objects with high relation weights when given a query object. The visualization shows that for both generic and task-specific datasets, the relation contributes to identifying other detection candidates based on the given object query. Furthermore, small-sized objects tend to establish more relation connections with other objects due to the lack of their own semantic information. Therefore, constructing relations is crucial for small-sized object detection. Fig. 6 further visualizes some failure cases of Relation-DETR, indicating that the presented model may benefit from occluded objects and dense objects with misleading semantic differences by considering more complex relations, such as occlusion and semantic relations.", + "bbox": [ + 212, + 602, + 787, + 768 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "5.7 Towards universal object detection", + "text_level": 1, + "bbox": [ + 215, + 787, + 549, + 803 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Will the position relation prior still be effective for datasets that cover a broader range of scenarios and objects? As a piece of general prior knowledge, we an-", + "bbox": [ + 212, + 809, + 785, + 840 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "Relation DETR", + "bbox": [ + 622, + 114, + 730, + 127 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 784, + 126 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/b88e6cea89b1cb55de875bdcd9cd43af9d578a420d2331248e09fe803cf1c57b.jpg", + "image_caption": [ + "Fig. 5: Representative objects(red) related to the Fig. 6: Predictions (left) and GTs given object(blue) (right) of failure cases" + ], + "image_footnote": [], + "bbox": [ + 218, + 143, + 333, + 277 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/910b96b06478edf672035823b8bebc20795af2d27ea30a702cfd0e8fdd2a862e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 333, + 143, + 462, + 277 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/4f4eada323159925d9da589f54b79022b2ec0cd7ccc3ae9d98bd3ccbfe77a7f3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 462, + 143, + 550, + 277 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/6013a8ef88a60636fba0ab2dd2cafa920d9dc5cf53c46ec8c5c6278c2dff689e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 550, + 143, + 666, + 277 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/229d92370487151532f69edba696994cf966f2f18b007b085ce8fd8b40ac4ddd.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 666, + 143, + 782, + 277 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "ticipate that the explicit position relation prior can benefit universal object detection tasks. To explore this, we have constructed a large-scale class-agnostic detection dataset with about 100,000 images, termed SA-Det-100k, by sampling a subset from SA-1B, which is one of the largest scale segmentation datasets proposed in Segment Anything [25]. We then compared the performance of our Relation-DETR with the baseline DINO [47] using VFL [48] on this dataset. The results in Tab. 8 show that Relation-DETR achieves a clear improvement of $1.3\\%$ AP, demonstrating the scalability of the proposed position relation prior.", + "bbox": [ + 212, + 345, + 787, + 467 + ], + "page_idx": 13 + }, + { + "type": "table", + "img_path": "images/9a1ebd98438d6ced4efc893f1aa48dd08aa958499bc86bb91e30653ffac0bd5c.jpg", + "table_caption": [ + "Table 8: Quantitative comparison on SA-Det-100k (ResNet-50, $1 \\times$" + ], + "table_footnote": [], + "table_body": "
MethodsAP@50:95AP@50AP@75APSAPMAPL
DINO with VFL43.752.047.75.843.061.5
Relation-DETR45.0(↑1.3)53.1(↑1.1)48.9(↑1.2)6.0(↑0.2)44.4(↑1.4)62.9(↑1.4)
", + "bbox": [ + 223, + 518, + 779, + 579 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "6 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 628, + 357, + 643 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "This paper explores explicit position relation prior for enhancing performance and convergence of DETR detectors. Built upon normalized relative geometry features, we propose a novel position relation that overcomes scale bias for progressive attention refinement. To address the conflicts between non-duplicate predictions and sufficient positive supervision in DETR frameworks, we extend the streaming pipeline to a contrast pipeline based on the proposed position relation. Combining these components produces a state-of-the-art detector, named Relation-DETR. Massive ablation studies and experimental results demonstrate the superior performance, faster convergence and promising transferability of the proposed detector. Moreover, Relation-DETR exhibits remarkable generalizability for both generic and task-specific detection tasks. We believe that this work will inspire future research on relation and structural bias for DETR detectors.", + "bbox": [ + 212, + 657, + 787, + 839 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "X. Hou et al.", + "bbox": [ + 271, + 114, + 359, + 126 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 215, + 143, + 401, + 162 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "This work was supported by the National Natural Science Foundation of China under Grant 62327808, and the Fundamental Research Funds for Xi'an Jiaotong University under Grants xtr072022001 and xzy022024009.", + "bbox": [ + 212, + 175, + 787, + 220 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 215, + 243, + 321, + 258 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Bodla, N., Singh, B., Chellappa, R., Davis, L.S.: Soft-NMS-improving object detection with one line of code. In: Int. Conf. Comput. Vis. pp. 5561-5569 (2017)", + "2. Bolya, D., Zhou, C., Xiao, F., Lee, Y.J.: YOLACT: Real-time instance segmentation. In: Int. Conf. Comput. Vis. pp. 9157-9166 (2019)", + "3. Cai, Z., Liu, S., Wang, G., Ge, Z., Zhang, X., Huang, D.: Align-DETR: Improving DETR with simple IoU-aware BCE loss. arXiv preprint arXiv:2304.07527 (2023)", + "4. Carion, N., Massa, F., Synnaeve, G., Usunier, N., Kirillov, A., Zagoruyko, S.: End-to-end object detection with transformers. In: Eur. Conf. Comput. Vis. pp. 213-229. Springer (2020)", + "5. Chen, K., Wang, J., Pang, J., Cao, Y., Xiong, Y., Li, X., Sun, S., Feng, W., Liu, Z., Xu, J., et al.: Mmdetection: Open mmlab detection toolbox and benchmark. arXiv preprint arXiv:1906.07155 (2019)", + "6. Chen, Q., Chen, X., Wang, J., Zhang, S., Yao, K., Feng, H., Han, J., Ding, E., Zeng, G., Wang, J.: Group DETR: Fast DETR training with group-wise one-to-many assignment. In: Int. Conf. Comput. Vis. pp. 6633-6642 (2023)", + "7. Chen, X., Li, L.J., Fei-Fei, L., Gupta, A.: Iterative visual reasoning beyond convolutions. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 7239-7248 (2018)", + "8. Chen, Y., Pan, J., Lei, J., Zeng, D., Wu, Z., Chen, C.: EEE-Net: Efficient edge enhanced network for surface defect detection of glass. IEEE Transactions on Instrumentation and Measurement (2023)", + "9. Cordts, M., Omran, M., Ramos, S., Rehfeld, T., Enzweiler, M., Benenson, R., Franke, U., Roth, S., Schiele, B.: The cityscapes dataset for semantic urban scene understanding. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 3213-3223 (2016)", + "10. Dai, X., Chen, Y., Yang, J., Zhang, P., Yuan, L., Zhang, L.: Dynamic DETR: End-to-end object detection with dynamic attention. In: Int. Conf. Comput. Vis. pp. 2988-2997 (2021)", + "11. Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: ImageNet: A large-scale hierarchical image database. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 248-255. IEEE (2009)", + "12. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. In: Int. Conf. Learn. Represent. (2020)", + "13. Everingham, M., Van Gool, L., Williams, C.K., Winn, J., Zisserman, A.: The pascal visual object classes (voc) challenge. Int. J. Comput. Vis. 88, 303-338 (2010)", + "14. Fang, Y., Sun, Q., Wang, X., Huang, T., Wang, X., Cao, Y.: EVA-02: A visual representation for neon genesis. arXiv preprint arXiv:2303.11331 (2023)", + "15. Feng, C., Zhong, Y., Gao, Y., Scott, M.R., Huang, W.: TOOD: Task-aligned one-stage object detection. In: Int. Conf. Comput. Vis. pp. 3490-3499. IEEE Computer Society (2021)", + "16. Girshick, R.: Fast R-CNN. In: Int. Conf. Comput. Vis. pp. 1440-1448 (2015)" + ], + "bbox": [ + 217, + 273, + 784, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "Relation DETR", + "bbox": [ + 622, + 114, + 730, + 126 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "17. Hao, X., Huang, D., Lin, J., Lin, C.Y.: Relation-enhanced DETR for component detection in graphic design reverse engineering. In: IJCAI. pp. 4785-4793 (2023)", + "18. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 770-778 (2016)", + "19. Hou, X., Liu, M., Zhang, S., Wei, P., Chen, B.: CANet: Contextual information and spatial attention based network for detecting small defects in manufacturing industry. Pattern Recognition 140, 109558 (2023)", + "20. Hou, X., Liu, M., Zhang, S., Wei, P., Chen, B.: Salience detr: Enhancing detection transformer with hierarchical salience filtering refinement. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 17574-17583 (2024)", + "21. Hu, H., Gu, J., Zhang, Z., Dai, J., Wei, Y.: Relation networks for object detection. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 3588-3597 (2018)", + "22. Hu, Z., Sun, Y., Wang, J., Yang, Y.: DAC-DETR: Divide the attention layers and conquer. Adv. Neural Inform. Process. Syst. 36 (2024)", + "23. Jia, D., Yuan, Y., He, H., Wu, X., Yu, H., Lin, W., Sun, L., Zhang, C., Hu, H.: DETRs with hybrid matching. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 19702-19712 (2023)", + "24. Jiang, C., Xu, H., Liang, X., Lin, L.: Hybrid knowledge routed modules for large-scale object detection. Adv. Neural Inform. Process. Syst. 31 (2018)", + "25. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., et al.: Segment anything. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 4015-4026 (2023)", + "26. Kolve, E., Mottaghi, R., Han, W., VanderBilt, E., Weihs, L., Herrasti, A., Deitke, M., Ehsani, K., Gordon, D., Zhu, Y., et al.: Ai2-thor: An interactive 3d environment for visual ai. arXiv preprint arXiv:1712.05474 (2017)", + "27. Krishna, R., Zhu, Y., Groth, O., Johnson, J., Hata, K., Kravitz, J., Chen, S., Kalantidis, Y., Li, L.J., Shamma, D.A., et al.: Visual genome: Connecting language and vision using crowdsourced dense image annotations. Int. J. Comput. Vis. 123, 32-73 (2017)", + "28. Li, F., Zhang, H., Liu, S., Guo, J., Ni, L.M., Zhang, L.: DN-DETR: Accelerate DETR training by introducing query denoising. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 13619-13627 (2022)", + "29. Li, X., Wang, W., Wu, L., Chen, S., Hu, X., Li, J., Tang, J., Yang, J.: Generalized focal loss: Learning qualified and distributed bounding boxes for dense object detection. Adv. Neural Inform. Process. Syst. 33, 21002-21012 (2020)", + "30. Li, Y., Mao, H., Girshick, R., He, K.: Exploring plain vision transformer backbones for object detection. In: Eur. Conf. Comput. Vis. pp. 280-296. Springer (2022)", + "31. Li, Z., Du, X., Cao, Y.: Gar: Graph assisted reasoning for object detection. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 1295-1304 (2020)", + "32. Lin, J., Pan, Y., Lai, R., Yang, X., Chao, H., Yao, T.: Core-Text: Improving scene text detection with contrastive relational reasoning. In: Int. Conf. Multimedia and Expo. pp. 1-6. IEEE (2021)", + "33. Lin, T., Wang, Y., Liu, X., Qiu, X.: A survey of transformers. AI Open (2022)", + "34. Lin, T.Y., Goyal, P., Girshick, R., He, K., Dollar, P.: Focal loss for dense object detection. In: Int. Conf. Comput. Vis. pp. 2980-2988 (2017)", + "35. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollar, P., Zitnick, C.L.: Microsoft COCO: Common objects in context. In: Eur. Conf. Comput. Vis. pp. 740-755. Springer (2014)" + ], + "bbox": [ + 215, + 146, + 784, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "X. Hou et al.", + "bbox": [ + 271, + 114, + 359, + 126 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "36. Liu, S., Li, F., Zhang, H., Yang, X., Qi, X., Su, H., Zhu, J., Zhang, L.: DABDETR: Dynamic anchor boxes are better queries for DETR. In: Int. Conf. Learn. Represent. (2021)", + "37. Liu, S., Ren, T., Chen, J., Zeng, Z., Zhang, H., Li, F., Li, H., Huang, J., Su, H., Zhu, J., et al.: Detection transformer with stable matching. arXiv preprint arXiv:2304.04742 (2023)", + "38. Liu, S., Huang, D., Wang, Y.: Adaptive NNS: Refining pedestrian detection in a crowd. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 6459-6468 (2019)", + "39. Liu, Z., Lin, Y., Cao, Y., Hu, H., Wei, Y., Zhang, Z., Lin, S., Guo, B.: Swin transformer: Hierarchical vision transformer using shifted windows. In: Int. Conf. Comput. Vis. pp. 10012-10022 (2021)", + "40. Pu, Y., Liang, W., Hao, Y., Yuan, Y., Yang, Y., Zhang, C., Hu, H., Huang, G.: Rank-DETR for high quality object detection. Adv. Neural Inform. Process. Syst. 36 (2024)", + "41. Shao, S., Li, Z., Zhang, T., Peng, C., Yu, G., Zhang, X., Li, J., Sun, J.: Objects365: A large-scale, high-quality dataset for object detection. In: Int. Conf. Comput. Vis. pp. 8430-8439 (2019)", + "42. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. Adv. Neural Inform. Process. Syst. 30 (2017)", + "43. Wang, Q., Gao, S., Xiong, L., Liang, A., Jiang, K., Zhang, W.: A casting surface dataset and benchmark for subtle and confusable defect detection in complex contexts. IEEE Sensors Journal pp. 1-1 (2024)", + "44. Wang, Y., Zhang, X., Yang, T., Sun, J.: Anchor DETR: Query design for transformer-based detector. In: AAAI. vol. 36, pp. 2567-2575 (2022)", + "45. Xu, H., Jiang, C., Liang, X., Lin, L., Li, Z.: Reasoning-RCNN: Unifying adaptive global reasoning into large-scale object detection. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 6419-6428 (2019)", + "46. Ye, M., Ke, L., Li, S., Tai, Y.W., Tang, C.K., Danelljan, M., Yu, F.: Cascade-DETR: Delving into high-quality universal object detection. In: Int. Conf. Comput. Vis. pp. 6704-6714 (2023)", + "47. Zhang, H., Li, F., Liu, S., Zhang, L., Su, H., Zhu, J., Ni, L., Shum, H.Y.: DINO: DETR with improved denoising anchor boxes for end-to-end object detection. In: Int. Conf. Learn. Represent. (2022)", + "48. Zhang, H., Wang, Y., Dayoub, F., Sunderhauf, N.: Varifocalnet: An iou-aware dense object detector. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 8514-8523 (2021)", + "49. Zhang, S., Wang, X., Wang, J., Pang, J., Lyu, C., Zhang, W., Luo, P., Chen, K.: Dense distinct query for end-to-end object detection. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 7329-7338 (2023)", + "50. Zhao, C., Sun, Y., Wang, W., Chen, Q., Ding, E., Yang, Y., Wang, J.: Ms-detr: Efficient detr training with mixed supervision. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 17027-17036 (2024)", + "51. Zhu, X., Su, W., Lu, L., Li, B., Wang, X., Dai, J.: Deformable DETR: Deformable transformers for end-to-end object detection. In: Int. Conf. Learn. Represent. (2020)", + "52. Zong, Z., Song, G., Liu, Y.: DETRs with collaborative hybrid assignments training. In: Int. Conf. Comput. Vis. pp. 6748-6758 (2023)" + ], + "bbox": [ + 212, + 146, + 787, + 811 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "Relation DETR", + "bbox": [ + 622, + 114, + 730, + 126 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 784, + 126 + ], + "page_idx": 16 + } +] \ No newline at end of file diff --git a/2024/Relation DETR_ Exploring Explicit Position Relation Prior for Object Detection/650a9d12-9a9f-48d6-b77e-bd08f66392b3_model.json b/2024/Relation DETR_ Exploring Explicit Position Relation Prior for Object Detection/650a9d12-9a9f-48d6-b77e-bd08f66392b3_model.json new file mode 100644 index 0000000000000000000000000000000000000000..d3b1a3f63e4d1d45fc024569421b2fe5189f3460 --- /dev/null +++ b/2024/Relation DETR_ Exploring Explicit Position Relation Prior for Object Detection/650a9d12-9a9f-48d6-b77e-bd08f66392b3_model.json @@ -0,0 +1,2269 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.25, + 0.141, + 0.753, + 0.187 + ], + "angle": 0, + "content": "Relation DETR: Exploring Explicit Position Relation Prior for Object Detection" + }, + { + "type": "text", + "bbox": [ + 0.24, + 0.216, + 0.763, + 0.248 + ], + "angle": 0, + "content": "Xiuquan Hou\\(^{1}\\), Meiqin Liu\\(^{1,2,\\star}\\), Senlin Zhang\\(^{2}\\), Ping Wei\\(^{1}\\), Badong Chen\\(^{1}\\), and Xuguang Lan\\(^{1}\\)" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.258, + 0.786, + 0.288 + ], + "angle": 0, + "content": "\\(^{1}\\) National Key Laboratory of Human-Machine Hybrid Augmented Intelligence, Xi'an Jiaotong University, Xi'an 710049, China" + }, + { + "type": "text", + "bbox": [ + 0.229, + 0.287, + 0.774, + 0.33 + ], + "angle": 0, + "content": "\\(^{2}\\) College of Electrical Engineering, Zhejiang University, Hangzhou 310027, China \nxiuqhou@stu.xjtu.edu.cn, {liumeiqin, slzhang}@zju.edu.cn, {pingwei, chenbd, xglan} @mail.xjtu.edu.cn" + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.398, + 0.74, + 0.759 + ], + "angle": 0, + "content": "Abstract. This paper presents a general scheme for enhancing the convergence and performance of DETR (DEtection Transformer). We investigate the slow convergence problem in transformers from a new perspective, suggesting that it arises from the self-attention that introduces no structural bias over inputs. To address this issue, we explore incorporating position relation prior as attention bias to augment object detection, following the verification of its statistical significance using a proposed quantitative macroscopic correlation (MC) metric. Our approach, termed Relation-DETR, introduces an encoder to construct position relation embeddings for progressive attention refinement, which further extends the traditional streaming pipeline of DETR into a contrastive relation pipeline to address the conflicts between non-duplicate predictions and positive supervision. Extensive experiments on both generic and task-specific datasets demonstrate the effectiveness of our approach. Under the same configurations, Relation-DETR achieves a significant improvement \\((+2.0\\%)\\) AP compared to DINO), state-of-the-art performance \\((51.7\\%)\\) AP for \\(1\\times\\) and \\(52.1\\%\\) AP for \\(2\\times\\) settings), and a remarkably faster convergence speed (over \\(40\\%\\) AP with only 2 training epochs) than existing DETR detectors on COCO val2017. Moreover, the proposed relation encoder serves as a universal plug-in-and-play component, bringing clear improvements for theoretically any DETR-like methods. Furthermore, we introduce a class-agnostic detection dataset, SA-Det-100k. The experimental results on the dataset illustrate that the proposed explicit position relation achieves a clear improvement of \\(1.3\\%\\) AP, highlighting its potential towards universal object detection. The code and dataset are available at https://github.com/xiuqhou/Relation-DETR." + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.775, + 0.741, + 0.804 + ], + "angle": 0, + "content": "Keywords: Detection transformer \\(\\cdot\\) Object detection \\(\\cdot\\) Relation network \\(\\cdot\\) Progressive attention refinement \\(\\cdot\\) Feature enhancement" + }, + { + "type": "page_footnote", + "bbox": [ + 0.221, + 0.825, + 0.382, + 0.84 + ], + "angle": 0, + "content": "* Corresponding author" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.128 + ], + "angle": 0, + "content": "X. Hou et al." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.145, + 0.375, + 0.161 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.176, + 0.784, + 0.235 + ], + "angle": 0, + "content": "Object detection aims to tackle the problems of bounding box regression and object classification for each object of interest. Recently, DETection TRansformer (DETR) [4] has overcome the reliance on handcrafted designs of convolution detectors, achieving an elegant architecture in an end-to-end manner." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.236, + 0.788, + 0.522 + ], + "angle": 0, + "content": "Despite exhibiting impressive detection performance on large-scale datasets such as COCO [35], their performances are prone to be influenced by dataset scale and suffer from slow convergence. The root cause of the problem is the conflict between non-duplicate predictions and positive supervision [23]. During the training process, DETR employs the Hungarian algorithm to assign a single positive prediction to each ground-truth for producing unique results. However, this leads to negative predictions dominating the majority of the loss function, causing insufficient positive supervision. Therefore, more samples and iterations are required for convergence. Previous attempts have explored the issue by introducing train-only architectures (e.g. query denoising [28], multiple groups of queries [6], auxiliary queries [23], collaborative hybrid assignment training [52]) for additional supervision or by incorporating hard mining in loss functions (e.g. IA-BCE loss [3], position-supervised loss [37]). Other works have proposed specific structures for better interaction between queries and feature maps (e.g. dynamic anchor query [36], cascade window attention [46]), as well as techniques to focus on high-quality queries (e.g. hierarchical filtering [20], dense distinct process [49] and query rank layer [40]). Despite these advancements, there has been little exploration of the issue from the perspective of self-attention, which is widely used in the transformer decoders in most DETR detectors." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.523, + 0.792, + 0.658 + ], + "angle": 0, + "content": "The effectiveness of self-attention lies in its establishment of a high-dimensional relation representation among sequence embeddings [4, 42], which also serves as a key component for modeling relations among different detection feature representations [4]. However, such relation is an implicit representation since it assumes no structural bias over inputs, in which even position information is also needed to be learned from training data [33]. Consequently, the learning process of transformer is data-intensive and slow to converge. This analysis motivates us to introduce task-specific bias for realizing faster convergence and reducing data dependence." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.66, + 0.788, + 0.809 + ], + "angle": 0, + "content": "In this paper, we explore enhancing DETR detectors from a novel perspective, namely explicit position relation prior. We first establish a metric for quantifying position relations in images, and analyze the distribution to verify its statistical significance. In this context, we introduce a position relation encoder to model all pairwise interactions between two bounding boxes, employing progressive attention refinement for cross-layer information interaction. To maintain the end-to-end property while providing sufficient positive supervision, we introduce a contrast relation strategy, which leverages both one-to-one and one-to-many matching while emphasizing the influence of position relation on dedduplication. The proposed method is named as Relation-DETR." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.841 + ], + "angle": 0, + "content": "Compared to previous works, the main feature of Relation-DETR is the integration of explicit position relation. In contrast, prior works focus on implicitly" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.625, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "Relation DETR" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.207 + ], + "angle": 0, + "content": "learned attention weights from training data, leading to slow convergence. Intuitively, our proposed position relation can be seen as a plug-in-and-play design beneficial for non-duplication predictions, since it establishes a representation of relative positions among pairs of bounding boxes (similar to IoU in NMS [16])." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.208, + 0.785, + 0.404 + ], + "angle": 0, + "content": "We evaluate the performance of Relation-DETR on the most popular object detection dataset, COCO 2017 [35], as well as several task-specific datasets [8,43]. The experimental results demonstrate its superior performance, surpassing previous state-of-the-art DETR detectors with clear margins. More specifically, Relation-DETR exhibits a significantly fast convergence speed. Without bells and whistles, it becomes the first DETR detector to achieve \\(40\\%\\) AP on COCO with only 2 epochs using ResNet50 as the backbone under the \\(1\\times\\) training configuration. In addition, the simple architecture design of our position relation encoder maintains a promising transferability. It can be easily extended to other DETR-based methods with only a few modifications to achieve consistent performance improvements. This is in contrast to some existing DETR detectors whose performance is highly dependent from complex matching strategies [22] or detection heads [52] developed by convolution-based detectors." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.426, + 0.388, + 0.442 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.456, + 0.788, + 0.744 + ], + "angle": 0, + "content": "Transformer for Object Detection In practice, the majority of attempts to apply transformer to object detection involve constructing a parallelizable sequence, either in the feature extractor [30] or in the detection body [4]. Specifically, transformer-based feature extractors generate token sequences based on image patches [12, 30, 39], and extract multi-scale features through aggregating local features [30, 39] or pyramid postprocess [14, 30]. DEtction TRansformer (DETR) proposed by Carion et al. [4] encodes the extracted features into object queries and decodes them into detected bounding boxes and labels. However, the self-learned attention mechanism increases the requirements for large-scale datasets and training iterations. Many works have explored the slow convergence from the perspective of structured attention (e.g. multi-scale deformable attention [51], dynamic attention [10], cascade window attention [46]), queries with explicit priors (e.g. anchor queries [44], dynamic anchor boxes queries [36], denoising queries [28], dense distinct queries [49]), and additional positive supervision (e.g. group queries [6], hybrid design [23], mixed matching [3]). However, even state-of-the-art DETR methods still utilize vanilla multi-head attention in the transformer decoder. And few works have explored the slow convergence from the perspective of implicit priors. This paper aims to address the issue with position relation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.787, + 0.842 + ], + "angle": 0, + "content": "Relation Network Rather than processing visual features at pixel levels, patch levels or image levels, relation network captures relation features at instance levels. Existing research on relation networks involves category-based and instance-based approaches. The category-based approaches construct conceptual or statistical relations (e.g. co-occurrence probability [17, 27]) either from relation" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.128 + ], + "angle": 0, + "content": "X. Hou et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.329 + ], + "angle": 0, + "content": "datasets like Visual Genome [24, 27, 45] or by adaptively learning from class labels [17]. Both of them, however, increase the complexity due to the assignment between instances and categories [7, 17, 45]. In contrast, instance-based approaches directly construct a fine-grained graph structure given object features as a node set and their relations as a edge set. Therefore, reasoning on the graph during the training process naturally determines the explicit relation weight [21]. Typically, the weight denotes the parametric distances between each paired object instances in high-dimensional space, such as appearance similarity [31], proposal distance [32] or even self-attention weight [42]. Since learning self-attention weight solely from training data without structural bias increases the requirement for dataset scale and iterations, we explore explicit position relation as a prior to reduce the requirement." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.351, + 0.788, + 0.489 + ], + "angle": 0, + "content": "Classification loss for hard mining. During object detection training, positive predictions assigned to ground truth are much fewer than negative predictions, often resulting in imbalanced supervision and slow convergence. For classification tasks, Focal Loss [34] proposes introducing a weight parameter to focus on hard samples, which is further extended to many variants like generalized focal loss (GFL) [29], vari focal loss (VFL) [29]. Moreover, for object detection tasks, using loss with modulation terms based on regression metrics (e.g. TOOD [15], IA-BCE [3], position-supervised loss [37]) further achieves high-quality alignment between classification and regression tasks." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.511, + 0.724, + 0.53 + ], + "angle": 0, + "content": "3 Statistical significance of object position relation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.543, + 0.788, + 0.664 + ], + "angle": 0, + "content": "Are objects really correlated in object detection tasks? To answer the question, we propose a quantitative macroscopic correlation (MC) metric based on the Pearson Correlation Coefficient (PCC) to measure the position correlation among objects in a single image. Assume the objects in an image form a node set, and the PCC between each pair of bounding box annotations serves as their corresponding edge weight. We can construct an undirected graph with continuous values. In this end, the macroscopic correlation for each image can be calculated using the graph intensity, formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.373, + 0.675, + 0.786, + 0.71 + ], + "angle": 0, + "content": "\\[\nM C = \\frac {\\sum_ {i} \\sum_ {j : j \\neq i} | \\operatorname {P e a r s o n} \\left(\\boldsymbol {b} _ {i} , \\boldsymbol {b} _ {j}\\right) |}{N (N - 1)} \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.719, + 0.788, + 0.78 + ], + "angle": 0, + "content": "where \\(N\\) denotes the number of objects, i.e. the number of nodes, \\(\\pmb{b} = [x,y,w,h]\\) denotes the position annotation of bounding boxes in datasets. \\(MC = 1\\) only when all objects are fully linearly correlated, whereas \\(MC = 0\\) if there is no position correlation between any pair of objects." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.842 + ], + "angle": 0, + "content": "We visualize the statistical distribution of MC for datasets across various scenarios, including industrial (ESD [19], CSD [43], MSSD [8]), domestic (AI2thor [26]), urban (Cityscapes [9]) and generic settings (PascalVOC [13], COCO [35], Object365 [41], SA-1B [25]). The datasets cover a wide range of scales, from" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.625, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "Relation DETR" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.784, + 0.126 + ], + "angle": 0, + "content": "5" + }, + { + "type": "image", + "bbox": [ + 0.33, + 0.147, + 0.675, + 0.354 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.369, + 0.788, + 0.412 + ], + "angle": 0, + "content": "Fig. 1: Statistical distribution of macroscopic correlation (MC) on various datasets (normalized for better visualization), and the values in brackets indicate the number of dataset samples." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.44, + 0.788, + 0.545 + ], + "angle": 0, + "content": "0.3K to 11M images. As shown in Fig. 1, all of these datasets indicate that the distribution of MC is concentrated around high numerical values, with the distribution centers closer to the upper bound. This demonstrates the presence and the statistical significance of object position relation. Specifically, task-specific datasets display more prior knowledge and clearer clustering patterns in the high-dimensional feature space, which thus results in higher MC values than generic datasets like COCO." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.568, + 0.407, + 0.584 + ], + "angle": 0, + "content": "4 Relation-DETR" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.599, + 0.788, + 0.72 + ], + "angle": 0, + "content": "Given the statistical significance of position relation, we propose a state-of-the-art detector, named Relation-DETR, that explores explicit position relation prior to enhance object detection. To address the issue of slow convergence, we present a position relation encoder (Sec. 4.1) for progressive attention refinement (Sec. 4.2). Further in Sec. 4.3, it extends the streaming pipeline of DETR into a contrast pipeline to emphasize the influence of position relation on removing duplication while maintaining sufficient positive supervision for faster convergence." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.741, + 0.471, + 0.755 + ], + "angle": 0, + "content": "4.1 Position relation encoder" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Previous research has demonstrated the relation effectiveness for convolution detectors [21,32]. Recently, some DETR methods attempt to construct instance-level relation by indexing from category-level relation using class indices [17]. In contrast to these approaches, we directly construct instance-level relation through a simple position encoder, maintaining an end-to-end design for DETR." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.228, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.128 + ], + "angle": 0, + "content": "X. Hou et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.268 + ], + "angle": 0, + "content": "We first review the basic pipeline in DETR detectors. Given image features extracted by the backbone, the transformer encoder produces an augmented memory \\(\\mathbf{Z} \\in \\mathbb{R}^{d \\times H \\times W}\\) for further decoding into bounding boxes \\(\\pmb{b}_i = [x, y, w, h], i = 1, \\dots, N\\) and class labels \\(\\pmb{c}_i\\) as predictions. Each decoder layer refines the bounding box coordinates iteratively by predicting \\(\\Delta\\) w.r.t. the coordinate from the last decoder layer, known as iterative bounding box refinement [51]. In addition, predictions from all decoder layers equally participate in loss calculation to compute auxiliary decoding losses [4]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.268, + 0.788, + 0.359 + ], + "angle": 0, + "content": "Under the aforementioned detection framework, our position relation encoder represents the high-dimensional relation embedding as an explicit prior for self-attention in the transformer. This embedding is calculated based on the predicted bounding boxes (denoted as \\(\\mathbf{b} = [x,y,w,h]\\)) from each decoder layer. To ensure that the relation is invariant to translation and scale transformations, we encode it based on normalized relative geometry features:" + }, + { + "type": "equation", + "bbox": [ + 0.229, + 0.363, + 0.786, + 0.398 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {e} (\\boldsymbol {b} _ {i}, \\boldsymbol {b} _ {j}) = \\left[ \\log \\left(\\frac {| x _ {i} - x _ {j} |}{w _ {i}} + 1\\right), \\log \\left(\\frac {| y _ {i} - y _ {j} |}{h _ {i}} + 1\\right), \\log \\left(\\frac {w _ {i}}{w _ {j}}\\right), \\log \\left(\\frac {h _ {i}}{h _ {j}}\\right) \\right] (2)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.402, + 0.788, + 0.448 + ], + "angle": 0, + "content": "Unlike [21], our position relation is unbiased, as \\( \\pmb{e}(\\pmb{b}_i, \\pmb{b}_j) = 0 \\) when \\( i = j \\). The relation matrix \\( \\mathbf{E} \\in \\mathbb{R}^{N \\times N \\times 4} \\) (with \\( \\mathbf{E}(i,j) = \\pmb{e}(\\pmb{b}_i, \\pmb{b}_j) \\)) is further transformed into high-dimensional embeddings through sine-cosine encoding [42]." + }, + { + "type": "equation", + "bbox": [ + 0.397, + 0.454, + 0.785, + 0.478 + ], + "angle": 0, + "content": "\\[\n\\operatorname {E m b e d} (\\mathbf {E}, 2 k) = \\sin \\left(s \\mathbf {E} / T ^ {2 k / d _ {\\mathrm {r e}}}\\right) \\tag {3}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.367, + 0.483, + 0.785, + 0.507 + ], + "angle": 0, + "content": "\\[\n\\operatorname {E m b e d} (\\mathbf {E}, 2 k + 1) = \\cos (s \\mathbf {E} / T ^ {2 k / d _ {\\mathrm {r e}}}) \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.511, + 0.788, + 0.556 + ], + "angle": 0, + "content": "where the shape of relation embedding is \\( N \\times N \\times 4d_{\\mathrm{re}} \\), and \\( T \\), \\( d_{\\mathrm{re}} \\), \\( s \\) are encoding parameters. Finally, the embedding undergoes a linear transformation to obtain \\( M \\) scalar weights, where \\( M \\) denotes the number of attention heads." + }, + { + "type": "equation", + "bbox": [ + 0.358, + 0.563, + 0.785, + 0.579 + ], + "angle": 0, + "content": "\\[\n\\operatorname {R e l} (\\boldsymbol {b}, \\boldsymbol {b}) = \\max \\left(\\epsilon , \\mathbf {W E m b e d} (\\boldsymbol {b}, \\boldsymbol {b}) + \\mathbf {B}\\right) \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.585, + 0.788, + 0.615 + ], + "angle": 0, + "content": "where \\(\\epsilon\\) makes sure a positive value for relation to avoid gradient vanishing after \\(\\exp\\) when integrated into self-attention, and \\(\\mathsf{Rel}(\\pmb {b},\\pmb {b})\\in \\mathbb{R}^{N\\times N\\times M}\\)." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.633, + 0.715, + 0.649 + ], + "angle": 0, + "content": "4.2 Progressive attention refinement with position relation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.653, + 0.788, + 0.746 + ], + "angle": 0, + "content": "The iterative box refinement proposed by Deformable-DETR [51] has shown the effectiveness for high-quality bounding box regression. Following the motivation, we propose a progressive attention refinement method to introduce the position relation into the streaming pipeline of DETR. Specifically, the relation of layer-\\(i\\) is determined by bounding boxes of both layer-\\(i - 1\\) and layer-\\(i\\), which is further integrated into self-attention for producing the bounding boxes in layer-\\((i + 1)\\)." + }, + { + "type": "equation", + "bbox": [ + 0.229, + 0.762, + 0.785, + 0.796 + ], + "angle": 0, + "content": "\\[\n\\operatorname {A t t n} _ {\\text {S e l f}} \\left(\\mathbf {Q} ^ {l}\\right) = \\operatorname {S o f t m a x} \\left(\\operatorname {R e l} \\left(\\boldsymbol {b} ^ {l - 1}, \\boldsymbol {b} ^ {l}\\right) + \\frac {\\operatorname {Q u e} \\left(\\mathbf {Q} ^ {l}\\right) \\operatorname {K e y} \\left(\\mathbf {Q} ^ {l}\\right) ^ {\\top}}{\\sqrt {d _ {\\text {m o d e l}}}}\\right) \\operatorname {V a l} \\left(\\mathbf {Q} ^ {l}\\right) \\tag {6}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.293, + 0.798, + 0.785, + 0.816 + ], + "angle": 0, + "content": "\\[\n\\mathbf {Q} ^ {l + 1} = \\text {F F N} \\left(\\mathbf {Q} ^ {l} + \\operatorname {A t t n} _ {\\text {c r o s s}} \\left(\\operatorname {A t t n} _ {\\text {s e l f}} (\\mathbf {Q} ^ {l}), \\operatorname {K e y} (\\mathbf {Z}), \\operatorname {V a l} (\\mathbf {Z})\\right)\\right) \\tag {7}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.297, + 0.819, + 0.785, + 0.838 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {b} ^ {l + 1} = \\operatorname {M L P} \\left(\\mathbf {Q} ^ {l + 1}\\right), \\boldsymbol {c} ^ {l + 1} = \\operatorname {L i n e a r} \\left(\\mathbf {Q} ^ {l + 1}\\right) \\tag {8}\n\\]" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.624, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "Relation DETR" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.146, + 0.784, + 0.178 + ], + "angle": 0, + "content": "where \\(\\mathbf{Q}^l\\) denotes the queries in the \\(l\\)-th decoder layer in the DETR transformer, \\(\\mathbf{Z}\\) is the memory, i.e., the enhanced image features from the transformer encoder." + }, + { + "type": "image", + "bbox": [ + 0.226, + 0.205, + 0.499, + 0.307 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.506, + 0.203, + 0.78, + 0.306 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.32, + 0.784, + 0.349 + ], + "angle": 0, + "content": "Fig. 2: Comparison of transformer decoder in Deformable-DETR(left) and Relation-DETR(right)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.38, + 0.785, + 0.471 + ], + "angle": 0, + "content": "The only difference between our method and the existing DETR decoder is marked in red. As depicted in Fig. 2, the requisite addition involves the introduction of a lateral branch for the computation of position relation. Therefore, our position relation and the progressive attention refinement are straightforward, allowing for plug-in-and-play integration with self attention in existing DETR detectors to achieve consistent performance improvements (see Tab. 6)." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.497, + 0.475, + 0.513 + ], + "angle": 0, + "content": "4.3 Contrast relation pipeline" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.526, + 0.785, + 0.617 + ], + "angle": 0, + "content": "Rethinking the mechanism of existing duplication removal methods (including NMS [16], Soft-NMS [1], fast-NMS [2], Adaptive-NMS [38]), these processes heavily rely on IoU (intersection over Union), which, to some extent, signifies the position relation between bounding boxes. Therefore, we may hypothesize that integrating the position relation among queries in self-attention contributes to non-duplicated predictions in object detection, akin to [22]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.618, + 0.785, + 0.739 + ], + "angle": 0, + "content": "The conflicts between non-duplicate predictions and sufficient positive supervision arise from the streaming pipeline of DETR, which must navigate between one-to-one matching and one-to-many matching. To overcome this limitation, we extend it to a contrast pipeline based on the proposed position relation. Specifically, we construct two parallel sets of queries, i.e. matching queries \\(\\mathbf{Q}_m\\) and hybrid queries \\(\\mathbf{Q}_h\\). Both are input into the transformer decoder but undergo distinct processing. The matching queries are processed with self-attention incorporating position relation to produce non-duplicated predictions:" + }, + { + "type": "equation", + "bbox": [ + 0.218, + 0.75, + 0.785, + 0.801 + ], + "angle": 0, + "content": "\\[\n\\operatorname {A t t n} _ {\\text {S e l f}} \\left(\\mathbf {Q} _ {\\mathrm {m}} ^ {l}\\right) = \\operatorname {S o f t m a x} \\left(\\operatorname {R e l} \\left(\\boldsymbol {b} ^ {l - 1}, \\boldsymbol {b} ^ {l}\\right) + \\frac {\\operatorname {Q u e} \\left(\\mathbf {Q} _ {\\mathrm {m}}\\right) \\operatorname {K e y} \\left(\\mathbf {Q} _ {\\mathrm {m}}\\right) ^ {\\top}}{\\sqrt {d _ {\\text {m o d e l}}}}\\right) \\operatorname {V a l} \\left(\\mathbf {Q} _ {\\mathrm {m}}\\right) \\tag {9}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.223, + 0.803, + 0.785, + 0.839 + ], + "angle": 0, + "content": "\\[\n\\operatorname {A t t n} _ {\\text {S e l f}} \\left(\\mathbf {Q} _ {\\mathrm {h}} ^ {l}\\right) = \\operatorname {S o f t m a x} \\left(\\frac {\\operatorname {Q u e} \\left(\\mathbf {Q} _ {\\mathrm {h}}\\right) \\operatorname {K e y} \\left(\\mathbf {Q} _ {\\mathrm {h}}\\right) ^ {\\top}}{\\sqrt {d _ {\\text {m o d e l}}}}\\right) \\operatorname {V a l} \\left(\\mathbf {Q} _ {\\mathrm {h}}\\right) \\tag {10}\n\\]" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.228, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.128 + ], + "angle": 0, + "content": "X. Hou et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.209 + ], + "angle": 0, + "content": "while the hybrid queries are decoded by the same decoder but skip the calculation of position relation to explore more potential candidates. Their corresponding predictions are denoted as \\(\\pmb{p}_m^l = (\\pmb{b}_m^l, \\pmb{c}_m^l)\\) and \\(\\pmb{p}_h^l = (\\pmb{b}_h^l, \\pmb{c}_h^l)\\), respectively. Details of the contrast relation pipeline are illustrated in Fig. 3." + }, + { + "type": "image", + "bbox": [ + 0.315, + 0.235, + 0.573, + 0.407 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.582, + 0.236, + 0.71, + 0.407 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.266, + 0.422, + 0.735, + 0.437 + ], + "angle": 0, + "content": "Fig. 3: Detailed illustration of the proposed contrast relation pipeline." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.467, + 0.789, + 0.514 + ], + "angle": 0, + "content": "Assuming \\(\\pmb{g}\\) denotes the ground truth annotations, for \\(\\pmb{p}_m\\), we employ a one-to-one matching scheme to emphasize the non-duplicate property, with loss calculation similar to the original DETR approach [4]:" + }, + { + "type": "equation", + "bbox": [ + 0.375, + 0.524, + 0.785, + 0.565 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {m} \\left(\\boldsymbol {p} _ {m}, \\boldsymbol {g}\\right) = \\sum_ {l = 1} ^ {L} \\mathcal {L} _ {\\text {H u n g a r i a n}} \\left(\\boldsymbol {p} _ {m} ^ {l}, \\boldsymbol {g}\\right) \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.575, + 0.788, + 0.623 + ], + "angle": 0, + "content": "While for \\( \\pmb{p}_h \\), a one-to-many matching scheme is employed to form more potential positive candidates. We simply follow H-DETR [23] and repeat the ground truth \\( K \\) times, denoted as \\( \\tilde{\\pmb{g}} = \\{\\pmb{g}^1, \\pmb{g}^2, \\dots, \\pmb{g}^K\\} \\), for loss calculation:" + }, + { + "type": "equation", + "bbox": [ + 0.381, + 0.633, + 0.785, + 0.675 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {h} \\left(\\boldsymbol {p} _ {h}, \\boldsymbol {g}\\right) = \\sum_ {l = 1} ^ {L} \\mathcal {L} _ {\\text {H u n g a r i a n}} \\left(\\boldsymbol {p} _ {h} ^ {l}, \\tilde {\\boldsymbol {g}}\\right) \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.685, + 0.788, + 0.732 + ], + "angle": 0, + "content": "where \\(\\mathcal{L}_{\\mathrm{Hungarian}}\\) denotes the Hungarian loss, and \\(L\\) denotes the number of decoder layers. It is worth noting that the hybrid queries are only involved during training, thus incurring no extra computational burden for inference." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.753, + 0.611, + 0.772 + ], + "angle": 0, + "content": "5 Experimental Results and Discussion" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.785, + 0.312, + 0.801 + ], + "angle": 0, + "content": "5.1 Setup" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.843 + ], + "angle": 0, + "content": "For a comprehensive evaluation, we conduct experiments on both an object detection benchmark, COCO 2017 [35], and two task-specific datasets, CSD [43]" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.624, + 0.114, + 0.732, + 0.128 + ], + "angle": 0, + "content": "Relation DETR" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.787, + 0.128 + ], + "angle": 0, + "content": "9" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.146, + 0.788, + 0.345 + ], + "angle": 0, + "content": "and MSSD [8]. Detection performance is measured using the standard Average Precision (AP) [35]. We train our model on NVIDIA A800 GPU (80GB) and RTX 3090 GPU (24GB) using the AdamW optimizer with an initial learning rate of \\(1 \\times 10^{-4}\\) and weight decay of \\(1 \\times 10^{-4}\\). The Relation-DETR is implemented based on the backbone ResNet-50 [18] and Swin-L [39] pretrained from ImageNet [11], which are finetuned with a learning rate \\(1 \\times 10^{-5}\\) during training. The learning rate is reduced by a factor of 0.1 at later stages. The parameters of the position relation encoder, \\(T\\), \\(d_{\\mathrm{re}}\\), \\(s\\) are empirically chosen as 10000, 16, 100, respectively. The hybrid training configurations follow DINO [47] and H-DETR [23], i.e., \\(N_{m} = 900\\), \\(N_{h} = 1500\\), \\(k = 6\\). We adopt the VariFocal loss [48] for training Relation-DETR. The training batch size is 10 for COCO 2017 and 2 for task-specific datasets. Before fed into detectors, images undergo the same augmentations (random resize, crop and flip) as other DETR detectors." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.378, + 0.612, + 0.394 + ], + "angle": 0, + "content": "5.2 Comparison with state-of-the-art methods" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.413, + 0.789, + 0.611 + ], + "angle": 0, + "content": "Comparison on COCO 2017. Tab. 1 presents the detection performance on COCO val 2017. Compared to other state-of-the-art DETR methods, our approach converges much faster and demonstrates significant improvements of \\(1.0\\%\\) AP, \\(1.0\\% \\mathrm{AP}_{50}\\) and \\(0.6\\% \\mathrm{AP}_{75}\\), respectively, suppressing the second best DDQ-DETR [49] with clear margins. Specifically, Relation-DETR achieves \\(51.7\\%\\) AP using only 12 epochs with the ResNet-50 backbone and even outperforms DINO [47] (51.2% AP) with 36 epochs (\\(3\\times\\) faster). More importantly, in contrast to DDQ-DETR [49] and Co-DETR [52] that leverages NMS in the decoder or postprocess for improving precision, our Relation-DETR maintains an end-to-end pipeline, ensuring promising extensibility. When integrated with Swin-L backbone, Relation-DETR outperforms all counterparts, achieving the best \\(57.8\\%\\) AP with a \\(0.5\\%\\) AP improvement, showcasing its excellent scalability for larger model capacity." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.644, + 0.789, + 0.842 + ], + "angle": 0, + "content": "Comparison on task-specific datasets. Different from generic object detection benchmarks, datasets in task-specific scenarios lack sufficient samples to provide semantic information. To reveal the generalizability of Relation-DETR, we conduct a performance comparison on two defect detection datasets, i.e. CSD [43] and MSSD [8]. The results in Tab. 3 show that, Relation-DETR improves the baseline DINO by \\(1.4\\%\\) AP on CSD, achieving the highest \\(54.4\\%\\) AP. Tab. 4 demonstrates that Relation-DETR further increases the margin to \\(6.4\\%\\) AP on MSSD and surpasses other counterparts. It is noteworthy that CSD [43] and MSSD [8] contain more small-sized objects than COCO 2017, which thus confirms the effectiveness of Relation-DETR to small-sized detection. Moreover, considering a stricter IoU threshold, Relation-DETR outperforms the second best method DINO by a significant margin of \\(11.1\\%\\) AP@75, highlighting the beneficial impact of explicit position relation on high-quality predictions." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.127 + ], + "angle": 0, + "content": "X. Hou et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.145, + 0.788, + 0.188 + ], + "angle": 0, + "content": "Table 1: Comparison with state-of-the-art methods on COCO val2017 using ResNet-50(IN-1K) backbone. The * means that we re-implement the methods and report the corresponding results." + }, + { + "type": "table", + "bbox": [ + 0.216, + 0.2, + 0.784, + 0.555 + ], + "angle": 0, + "content": "
MethodBackboneEpochsAP@50:95AP@50AP@75\\( AP_S \\)\\( AP_M \\)\\( AP_L \\)
Anchor-DETR [44]ResNet-505042.163.144.922.346.260.0
Def-DETR* [51]ResNet-501245.465.049.127.249.661.0
DAB-Def-DETR* [36]ResNet-501246.365.950.429.350.561.9
DN-Def-DETR* [28]ResNet-501247.265.351.330.250.662.4
DINO* [47]ResNet-501249.967.454.533.953.563.8
Group-DETR [6]ResNet-501249.832.453.064.2
H-Def-DETR [23]ResNet-501248.766.452.931.251.563.5
Cascade-DETR [46]ResNet-501249.767.154.132.453.565.1
Co-Def-DETR [52]ResNet-501249.567.654.332.452.763.7
Align-DETR [3]ResNet-501250.267.854.432.953.365.0
Stable-DINO [37]ResNet-501250.467.455.032.954.065.5
DAC-DETR [22]ResNet-501250.067.654.7
Salience-DETR* [20]ResNet-501250.067.754.233.354.464.4
Rank-DETR [40]ResNet-501250.467.955.233.653.864.2
MS-DETR [50]ResNet-501250.367.455.132.754.064.6
DDQ-DETR [49]ResNet-501250.768.155.7
Relation-DETRResNet-501251.769.156.336.155.666.1
DINO [47]ResNet-502450.468.354.833.353.764.8
DINO [47]ResNet-503651.269.055.835.054.365.3
H-Def-DETR [23]ResNet-503650.068.354.432.952.765.3
Salience-DETR* [20]ResNet-502451.268.955.733.955.565.6
DDQ-DETR [49]ResNet-502452.069.557.235.254.965.9
Relation-DETRResNet-502452.169.756.636.156.066.5
" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.568, + 0.784, + 0.597 + ], + "angle": 0, + "content": "Table 2: Comparison with state-of-the-art methods on COCO val2017 using SwinL(IN-22K) as the backbone." + }, + { + "type": "table", + "bbox": [ + 0.216, + 0.61, + 0.791, + 0.724 + ], + "angle": 0, + "content": "
MethodBackboneEpochsAP@50:95AP@50AP@75\\( AP_S \\)\\( AP_M \\)\\( AP_L \\)
DINO [47]Swin-L1256.875.462.341.160.673.5
\\( \\mathcal{H} \\text{-Def-DETR} \\) [23]Swin-L1255.975.261.039.159.972.2
Salience-DETR* [20]Swin-L1256.575.061.540.261.272.8
Rank-DETR [40]Swin-L1257.375.962.940.861.373.2
Rank-DINO [40]Swin-L1257.676.063.441.661.473.8
Relation-DETRSwin-L1257.876.162.941.262.174.4
" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.753, + 0.389, + 0.768 + ], + "angle": 0, + "content": "5.3 Ablation study" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.842 + ], + "angle": 0, + "content": "This part conducts an ablation study to explore how the proposed components influence the final detection performance on COCO. The results in Tab. 5 show that, each key component of Relation-DETR consistently contributes to improving AP. Even on a highly-optimized baseline with VariFocal loss [48], our posi" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.624, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "Relation DETR" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "table_caption", + "bbox": [ + 0.34, + 0.145, + 0.663, + 0.16 + ], + "angle": 0, + "content": "Table 3: Quantitative comparison on CSD [43]." + }, + { + "type": "table", + "bbox": [ + 0.223, + 0.171, + 0.781, + 0.3 + ], + "angle": 0, + "content": "
MethodBackboneEpochsAP@50:95AP@50AP@75\\( AP_S \\)\\( AP_M \\)\\( AP_L \\)
Def-DETR [51]ResNet-5030043.786.236.640.534.910.0
DAB-Def-DETR [36]ResNet-509052.991.255.050.339.40.0
DN-Def-DETR [28]ResNet-506049.988.051.247.637.70.0
DINO [47]ResNet-506053.090.855.550.939.60.0
H-Def-DETR [23]ResNet-506053.090.655.751.239.26.7
Salience-DETR [20]ResNet-506053.292.555.151.040.90.0
Relation-DETRResNet-506054.492.956.053.640.80.0
" + }, + { + "type": "table_caption", + "bbox": [ + 0.338, + 0.329, + 0.665, + 0.344 + ], + "angle": 0, + "content": "Table 4: Quantitative comparison on MSSD [8]." + }, + { + "type": "table", + "bbox": [ + 0.223, + 0.355, + 0.781, + 0.484 + ], + "angle": 0, + "content": "
MethodBackboneEpochsAP@50:95AP@50AP@75APSAPMAPL
Def-DETR [51]ResNet-5030033.054.332.09.811.133.3
DAB-Def-DETR [36]ResNet-5012033.760.031.015.926.729.0
DN-Def-DETR [28]ResNet-5012045.674.144.918.631.941.0
H-Def-DETR [23]ResNet-5012046.976.847.120.345.340.7
DINO [47]ResNet-5012051.080.052.520.047.444.8
Salience-DETR [20]ResNet-5012055.478.261.928.747.544.5
Relation-DETRResNet-5012057.479.263.631.453.545.5
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.527, + 0.788, + 0.62 + ], + "angle": 0, + "content": "tion relation encoder and contrast pipeline bring clear improvements of \\(+0.3\\%\\), \\(+0.5\\%\\) AP, respectively. Built upon normalized relative geometry features, the position relation overcomes scale bias effectively, thus benefiting consistent performance improvements for different sized objects. For instance, Tab. 5 shows that introducing relation into the baseline with VFL achieves \\(+1.2\\%\\) \\(\\mathrm{AP}_S\\), \\(+1.0\\%\\) \\(\\mathrm{AP}_M\\), and \\(+1.3\\%\\) \\(\\mathrm{AP}_L\\)." + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.66, + 0.788, + 0.69 + ], + "angle": 0, + "content": "Table 5: Ablation study on key components of Relation-DETR (ResNet-50, 1x). We study both baseline and an optimized version with VariFocal Loss [48]." + }, + { + "type": "table", + "bbox": [ + 0.216, + 0.701, + 0.789, + 0.802 + ], + "angle": 0, + "content": "
VFLRel.contrastAP@50:95AP@50AP@75APSAPMAPL
49.967.454.533.953.563.8
✓50.3(↑0.4)68.1(↑0.7)55.0(↑0.5)34.7(↑0.8)53.4(↓0.1)64.7(↑0.9)
✓50.9(↑0.6)68.2(↑0.1)55.3(↑0.3)34.9(↑0.2)54.6(↑1.2)64.8(↑0.1)
✓✓51.2(↑0.3)68.5(↑0.7)55.7(↑0.4)35.3(↑0.4)54.9(↑0.3)65.5(↑0.7)
✓✓✓51.7(↑0.5)69.1(↑0.6)56.3(↑0.6)36.1(↑0.8)55.6(↑0.7)66.1(↑0.6)
" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.127 + ], + "angle": 0, + "content": "X. Hou et al." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.147, + 0.544, + 0.162 + ], + "angle": 0, + "content": "5.4 Transferability of position relation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.175, + 0.788, + 0.342 + ], + "angle": 0, + "content": "Our position relation encoder adopts an elegant architectural design, ensuring a promising transferability to existing DETR detectors with minimal modifications. The experimental results in Tab. 6 show that, integrating position relation encoders without any further modifications enhances detection performance with clear margins of \\(1.6\\%\\), \\(2.0\\%\\), \\(0.1\\%\\) and \\(0.2\\%\\) for Deformable-DETR [51], DAB-Deformable-DETR [36], DN-Deformable-DETR [28] and DINO [47], respectively. Interestingly, in comparison to \\(\\mathrm{AP}_M\\) and \\(\\mathrm{AP}_L\\), the position relation has a more substantial impact on improving \\(\\mathrm{AP}_S\\) for Deformable-DETR [51] and DAB-Deformable-DETR [36]. We attribute this to the fact that these early proposed baselines introduce relatively less structural bias and thus benefit more from our explicit position relation prior." + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.369, + 0.784, + 0.397 + ], + "angle": 0, + "content": "Table 6: Transfer experiments for the position relation encoder (ResNet-50, \\(1 \\times\\)). \"+RelEnc\" denotes the version that integrates our position relation encoder." + }, + { + "type": "table", + "bbox": [ + 0.216, + 0.409, + 0.78, + 0.537 + ], + "angle": 0, + "content": "
MethodAP@50:95AP@50AP@75APSAPMAPL
Def-DETR45.465.049.127.249.661.0
+RelEnc47.0(↑1.6)65.6(↑0.6)51.2(↑2.1)29.3(↑2.1)51.0(↑1.4)62.2(↑1.2)
DAB-Def-DETR46.365.950.428.350.561.9
+RelEnc48.3(↑2.0)66.5(↑0.6)52.9(↑2.5)32.4(↑4.1)52.0(↑1.5)62.0(↑0.1)
DN-Def-DETR47.265.351.330.250.662.4
+RelEnc47.3(↑0.1)65.6(↑0.3)51.4(↑0.1)29.9(↓0.3)50.8(↑0.2)62.1(↓0.3)
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.567, + 0.788, + 0.704 + ], + "angle": 0, + "content": "Moreover, the proposed contrast pipeline can be seen as an extension of hybrid matching [23], utilizing the proposed position relation encoder. Table 7 compares their transferability when integrated into DINO. The results indicate that directly applying hybrid matching [23] to DINO [47] results in a decrease in performance, from \\(49.9\\%\\) AP to \\(49.5\\%\\) AP. In contrast, the introduction of both the proposed relation encoder and the extended contrast pipeline consistently increases the performance. This demonstrates the effectiveness of the proposed position relation prior in improve detection performance, overcoming the weak generalizability inherent in hybrid matching." + }, + { + "type": "table_caption", + "bbox": [ + 0.238, + 0.73, + 0.763, + 0.745 + ], + "angle": 0, + "content": "Table 7: Transfer experiments compared with hybrid matching [23] on DINO." + }, + { + "type": "table", + "bbox": [ + 0.216, + 0.756, + 0.794, + 0.814 + ], + "angle": 0, + "content": "
DINO+Hybrid [23]+RelEnc+RelEnc+Contrast
APAP@50AP@75APAP@50AP@75APAP@50AP@75APAP@50AP@75
49.967.454.549.5†66.654.050.368.155.051.268.755.6
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.214, + 0.815, + 0.788, + 0.838 + ], + "angle": 0, + "content": "† We found that hybrid matching decreases the performance of DINO. The conclusion is consistent with the results of HDINO (see here) reported by MMDetection [5]." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.624, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "Relation DETR" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.147, + 0.542, + 0.162 + ], + "angle": 0, + "content": "5.5 Intuitive performance comparison" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.169, + 0.788, + 0.29 + ], + "angle": 0, + "content": "To facilitate an intuitive performance comparison, Fig. 4 plots the convergence curve and the precision-recall curve. Because the position relation prior reduces the requirement for learning structural bias from data [33], Relation-DETR illustrates a faster convergence speed. When training from scratch, it achieves a higher AP than other counterparts with fewer iterations. Specifically, Relation-DETR can achieve over \\(40\\%\\) AP with only 2 epochs, surpassing existing DETR detectors. In addition to convergence speed, the PR curves under different IoU thresholds also validate the performance gain of our Relation-DETR." + }, + { + "type": "image", + "bbox": [ + 0.229, + 0.315, + 0.499, + 0.477 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.506, + 0.316, + 0.776, + 0.476 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.491, + 0.785, + 0.533 + ], + "angle": 0, + "content": "Fig. 4: Convergence curve(left) and Precision-recall curve for \\(\\mathrm{IoU} = 50\\% \\sim 95\\%\\) (right). All models are trained with ResNet-50 backbone under the same \\(1\\times\\) training configuration on COCO 2017." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.581, + 0.371, + 0.594 + ], + "angle": 0, + "content": "5.6 Visualization" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.603, + 0.788, + 0.769 + ], + "angle": 0, + "content": "For a more intuitive grasp of the relation mechanism, Fig. 5 illustrates representative objects with high relation weights when given a query object. The visualization shows that for both generic and task-specific datasets, the relation contributes to identifying other detection candidates based on the given object query. Furthermore, small-sized objects tend to establish more relation connections with other objects due to the lack of their own semantic information. Therefore, constructing relations is crucial for small-sized object detection. Fig. 6 further visualizes some failure cases of Relation-DETR, indicating that the presented model may benefit from occluded objects and dense objects with misleading semantic differences by considering more complex relations, such as occlusion and semantic relations." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.788, + 0.55, + 0.804 + ], + "angle": 0, + "content": "5.7 Towards universal object detection" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.842 + ], + "angle": 0, + "content": "Will the position relation prior still be effective for datasets that cover a broader range of scenarios and objects? As a piece of general prior knowledge, we an-" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.127 + ], + "angle": 0, + "content": "X. Hou et al." + }, + { + "type": "image", + "bbox": [ + 0.22, + 0.145, + 0.334, + 0.279 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.334, + 0.145, + 0.464, + 0.279 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.464, + 0.145, + 0.551, + 0.279 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.552, + 0.145, + 0.667, + 0.279 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.668, + 0.145, + 0.784, + 0.279 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.217, + 0.289, + 0.785, + 0.319 + ], + "angle": 0, + "content": "Fig. 5: Representative objects(red) related to the Fig. 6: Predictions (left) and GTs given object(blue) (right) of failure cases" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.346, + 0.788, + 0.468 + ], + "angle": 0, + "content": "ticipate that the explicit position relation prior can benefit universal object detection tasks. To explore this, we have constructed a large-scale class-agnostic detection dataset with about 100,000 images, termed SA-Det-100k, by sampling a subset from SA-1B, which is one of the largest scale segmentation datasets proposed in Segment Anything [25]. We then compared the performance of our Relation-DETR with the baseline DINO [47] using VFL [48] on this dataset. The results in Tab. 8 show that Relation-DETR achieves a clear improvement of \\(1.3\\%\\) AP, demonstrating the scalability of the proposed position relation prior." + }, + { + "type": "table_caption", + "bbox": [ + 0.27, + 0.493, + 0.733, + 0.509 + ], + "angle": 0, + "content": "Table 8: Quantitative comparison on SA-Det-100k (ResNet-50, \\(1 \\times\\)" + }, + { + "type": "table", + "bbox": [ + 0.224, + 0.52, + 0.78, + 0.58 + ], + "angle": 0, + "content": "
MethodsAP@50:95AP@50AP@75APSAPMAPL
DINO with VFL43.752.047.75.843.061.5
Relation-DETR45.0(↑1.3)53.1(↑1.1)48.9(↑1.2)6.0(↑0.2)44.4(↑1.4)62.9(↑1.4)
" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.629, + 0.359, + 0.645 + ], + "angle": 0, + "content": "6 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.659, + 0.788, + 0.84 + ], + "angle": 0, + "content": "This paper explores explicit position relation prior for enhancing performance and convergence of DETR detectors. Built upon normalized relative geometry features, we propose a novel position relation that overcomes scale bias for progressive attention refinement. To address the conflicts between non-duplicate predictions and sufficient positive supervision in DETR frameworks, we extend the streaming pipeline to a contrast pipeline based on the proposed position relation. Combining these components produces a state-of-the-art detector, named Relation-DETR. Massive ablation studies and experimental results demonstrate the superior performance, faster convergence and promising transferability of the proposed detector. Moreover, Relation-DETR exhibits remarkable generalizability for both generic and task-specific detection tasks. We believe that this work will inspire future research on relation and structural bias for DETR detectors." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.624, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "Relation DETR" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.145, + 0.403, + 0.163 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.176, + 0.788, + 0.222 + ], + "angle": 0, + "content": "This work was supported by the National Natural Science Foundation of China under Grant 62327808, and the Fundamental Research Funds for Xi'an Jiaotong University under Grants xtr072022001 and xzy022024009." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.244, + 0.323, + 0.26 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.274, + 0.785, + 0.303 + ], + "angle": 0, + "content": "1. Bodla, N., Singh, B., Chellappa, R., Davis, L.S.: Soft-NMS-improving object detection with one line of code. In: Int. Conf. Comput. Vis. pp. 5561-5569 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.304, + 0.785, + 0.33 + ], + "angle": 0, + "content": "2. Bolya, D., Zhou, C., Xiao, F., Lee, Y.J.: YOLACT: Real-time instance segmentation. In: Int. Conf. Comput. Vis. pp. 9157-9166 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.331, + 0.785, + 0.358 + ], + "angle": 0, + "content": "3. Cai, Z., Liu, S., Wang, G., Ge, Z., Zhang, X., Huang, D.: Align-DETR: Improving DETR with simple IoU-aware BCE loss. arXiv preprint arXiv:2304.07527 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.359, + 0.785, + 0.399 + ], + "angle": 0, + "content": "4. Carion, N., Massa, F., Synnaeve, G., Usunier, N., Kirillov, A., Zagoruyko, S.: End-to-end object detection with transformers. In: Eur. Conf. Comput. Vis. pp. 213-229. Springer (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.4, + 0.785, + 0.44 + ], + "angle": 0, + "content": "5. Chen, K., Wang, J., Pang, J., Cao, Y., Xiong, Y., Li, X., Sun, S., Feng, W., Liu, Z., Xu, J., et al.: Mmdetection: Open mmlab detection toolbox and benchmark. arXiv preprint arXiv:1906.07155 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.441, + 0.785, + 0.482 + ], + "angle": 0, + "content": "6. Chen, Q., Chen, X., Wang, J., Zhang, S., Yao, K., Feng, H., Han, J., Ding, E., Zeng, G., Wang, J.: Group DETR: Fast DETR training with group-wise one-to-many assignment. In: Int. Conf. Comput. Vis. pp. 6633-6642 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.483, + 0.785, + 0.509 + ], + "angle": 0, + "content": "7. Chen, X., Li, L.J., Fei-Fei, L., Gupta, A.: Iterative visual reasoning beyond convolutions. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 7239-7248 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.51, + 0.785, + 0.55 + ], + "angle": 0, + "content": "8. Chen, Y., Pan, J., Lei, J., Zeng, D., Wu, Z., Chen, C.: EEE-Net: Efficient edge enhanced network for surface defect detection of glass. IEEE Transactions on Instrumentation and Measurement (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.551, + 0.785, + 0.592 + ], + "angle": 0, + "content": "9. Cordts, M., Omran, M., Ramos, S., Rehfeld, T., Enzweiler, M., Benenson, R., Franke, U., Roth, S., Schiele, B.: The cityscapes dataset for semantic urban scene understanding. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 3213-3223 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.593, + 0.785, + 0.634 + ], + "angle": 0, + "content": "10. Dai, X., Chen, Y., Yang, J., Zhang, P., Yuan, L., Zhang, L.: Dynamic DETR: End-to-end object detection with dynamic attention. In: Int. Conf. Comput. Vis. pp. 2988-2997 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.634, + 0.785, + 0.675 + ], + "angle": 0, + "content": "11. Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: ImageNet: A large-scale hierarchical image database. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 248-255. IEEE (2009)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.676, + 0.785, + 0.73 + ], + "angle": 0, + "content": "12. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. In: Int. Conf. Learn. Represent. (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.731, + 0.785, + 0.757 + ], + "angle": 0, + "content": "13. Everingham, M., Van Gool, L., Williams, C.K., Winn, J., Zisserman, A.: The pascal visual object classes (voc) challenge. Int. J. Comput. Vis. 88, 303-338 (2010)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.758, + 0.785, + 0.785 + ], + "angle": 0, + "content": "14. Fang, Y., Sun, Q., Wang, X., Huang, T., Wang, X., Cao, Y.: EVA-02: A visual representation for neon genesis. arXiv preprint arXiv:2303.11331 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.786, + 0.785, + 0.826 + ], + "angle": 0, + "content": "15. Feng, C., Zhong, Y., Gao, Y., Scott, M.R., Huang, W.: TOOD: Task-aligned one-stage object detection. In: Int. Conf. Comput. Vis. pp. 3490-3499. IEEE Computer Society (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.827, + 0.753, + 0.84 + ], + "angle": 0, + "content": "16. Girshick, R.: Fast R-CNN. In: Int. Conf. Comput. Vis. pp. 1440-1448 (2015)" + }, + { + "type": "list", + "bbox": [ + 0.218, + 0.274, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.127 + ], + "angle": 0, + "content": "X. Hou et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.175 + ], + "angle": 0, + "content": "17. Hao, X., Huang, D., Lin, J., Lin, C.Y.: Relation-enhanced DETR for component detection in graphic design reverse engineering. In: IJCAI. pp. 4785-4793 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.177, + 0.785, + 0.204 + ], + "angle": 0, + "content": "18. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 770-778 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.205, + 0.785, + 0.247 + ], + "angle": 0, + "content": "19. Hou, X., Liu, M., Zhang, S., Wei, P., Chen, B.: CANet: Contextual information and spatial attention based network for detecting small defects in manufacturing industry. Pattern Recognition 140, 109558 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.248, + 0.785, + 0.289 + ], + "angle": 0, + "content": "20. Hou, X., Liu, M., Zhang, S., Wei, P., Chen, B.: Salience detr: Enhancing detection transformer with hierarchical salience filtering refinement. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 17574-17583 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.29, + 0.785, + 0.317 + ], + "angle": 0, + "content": "21. Hu, H., Gu, J., Zhang, Z., Dai, J., Wei, Y.: Relation networks for object detection. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 3588-3597 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.318, + 0.785, + 0.346 + ], + "angle": 0, + "content": "22. Hu, Z., Sun, Y., Wang, J., Yang, Y.: DAC-DETR: Divide the attention layers and conquer. Adv. Neural Inform. Process. Syst. 36 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.347, + 0.785, + 0.388 + ], + "angle": 0, + "content": "23. Jia, D., Yuan, Y., He, H., Wu, X., Yu, H., Lin, W., Sun, L., Zhang, C., Hu, H.: DETRs with hybrid matching. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 19702-19712 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.389, + 0.785, + 0.416 + ], + "angle": 0, + "content": "24. Jiang, C., Xu, H., Liang, X., Lin, L.: Hybrid knowledge routed modules for large-scale object detection. Adv. Neural Inform. Process. Syst. 31 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.418, + 0.785, + 0.459 + ], + "angle": 0, + "content": "25. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., et al.: Segment anything. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 4015-4026 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.46, + 0.785, + 0.501 + ], + "angle": 0, + "content": "26. Kolve, E., Mottaghi, R., Han, W., VanderBilt, E., Weihs, L., Herrasti, A., Deitke, M., Ehsani, K., Gordon, D., Zhu, Y., et al.: Ai2-thor: An interactive 3d environment for visual ai. arXiv preprint arXiv:1712.05474 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.502, + 0.785, + 0.557 + ], + "angle": 0, + "content": "27. Krishna, R., Zhu, Y., Groth, O., Johnson, J., Hata, K., Kravitz, J., Chen, S., Kalantidis, Y., Li, L.J., Shamma, D.A., et al.: Visual genome: Connecting language and vision using crowdsourced dense image annotations. Int. J. Comput. Vis. 123, 32-73 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.558, + 0.785, + 0.6 + ], + "angle": 0, + "content": "28. Li, F., Zhang, H., Liu, S., Guo, J., Ni, L.M., Zhang, L.: DN-DETR: Accelerate DETR training by introducing query denoising. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 13619-13627 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.601, + 0.785, + 0.642 + ], + "angle": 0, + "content": "29. Li, X., Wang, W., Wu, L., Chen, S., Hu, X., Li, J., Tang, J., Yang, J.: Generalized focal loss: Learning qualified and distributed bounding boxes for dense object detection. Adv. Neural Inform. Process. Syst. 33, 21002-21012 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.643, + 0.785, + 0.671 + ], + "angle": 0, + "content": "30. Li, Y., Mao, H., Girshick, R., He, K.: Exploring plain vision transformer backbones for object detection. In: Eur. Conf. Comput. Vis. pp. 280-296. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.672, + 0.785, + 0.712 + ], + "angle": 0, + "content": "31. Li, Z., Du, X., Cao, Y.: Gar: Graph assisted reasoning for object detection. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 1295-1304 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.713, + 0.785, + 0.755 + ], + "angle": 0, + "content": "32. Lin, J., Pan, Y., Lai, R., Yang, X., Chao, H., Yao, T.: Core-Text: Improving scene text detection with contrastive relational reasoning. In: Int. Conf. Multimedia and Expo. pp. 1-6. IEEE (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.756, + 0.765, + 0.77 + ], + "angle": 0, + "content": "33. Lin, T., Wang, Y., Liu, X., Qiu, X.: A survey of transformers. AI Open (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.771, + 0.785, + 0.798 + ], + "angle": 0, + "content": "34. Lin, T.Y., Goyal, P., Girshick, R., He, K., Dollar, P.: Focal loss for dense object detection. In: Int. Conf. Comput. Vis. pp. 2980-2988 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.799, + 0.785, + 0.84 + ], + "angle": 0, + "content": "35. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollar, P., Zitnick, C.L.: Microsoft COCO: Common objects in context. In: Eur. Conf. Comput. Vis. pp. 740-755. Springer (2014)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.624, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "Relation DETR" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.189 + ], + "angle": 0, + "content": "36. Liu, S., Li, F., Zhang, H., Yang, X., Qi, X., Su, H., Zhu, J., Zhang, L.: DABDETR: Dynamic anchor boxes are better queries for DETR. In: Int. Conf. Learn. Represent. (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.19, + 0.787, + 0.232 + ], + "angle": 0, + "content": "37. Liu, S., Ren, T., Chen, J., Zeng, Z., Zhang, H., Li, F., Li, H., Huang, J., Su, H., Zhu, J., et al.: Detection transformer with stable matching. arXiv preprint arXiv:2304.04742 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.232, + 0.786, + 0.259 + ], + "angle": 0, + "content": "38. Liu, S., Huang, D., Wang, Y.: Adaptive NNS: Refining pedestrian detection in a crowd. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 6459-6468 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.26, + 0.787, + 0.3 + ], + "angle": 0, + "content": "39. Liu, Z., Lin, Y., Cao, Y., Hu, H., Wei, Y., Zhang, Z., Lin, S., Guo, B.: Swin transformer: Hierarchical vision transformer using shifted windows. In: Int. Conf. Comput. Vis. pp. 10012-10022 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.301, + 0.787, + 0.342 + ], + "angle": 0, + "content": "40. Pu, Y., Liang, W., Hao, Y., Yuan, Y., Yang, Y., Zhang, C., Hu, H., Huang, G.: Rank-DETR for high quality object detection. Adv. Neural Inform. Process. Syst. 36 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.343, + 0.786, + 0.384 + ], + "angle": 0, + "content": "41. Shao, S., Li, Z., Zhang, T., Peng, C., Yu, G., Zhang, X., Li, J., Sun, J.: Objects365: A large-scale, high-quality dataset for object detection. In: Int. Conf. Comput. Vis. pp. 8430-8439 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.384, + 0.787, + 0.424 + ], + "angle": 0, + "content": "42. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. Adv. Neural Inform. Process. Syst. 30 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.425, + 0.787, + 0.467 + ], + "angle": 0, + "content": "43. Wang, Q., Gao, S., Xiong, L., Liang, A., Jiang, K., Zhang, W.: A casting surface dataset and benchmark for subtle and confusable defect detection in complex contexts. IEEE Sensors Journal pp. 1-1 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.468, + 0.787, + 0.494 + ], + "angle": 0, + "content": "44. Wang, Y., Zhang, X., Yang, T., Sun, J.: Anchor DETR: Query design for transformer-based detector. In: AAAI. vol. 36, pp. 2567-2575 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.495, + 0.787, + 0.535 + ], + "angle": 0, + "content": "45. Xu, H., Jiang, C., Liang, X., Lin, L., Li, Z.: Reasoning-RCNN: Unifying adaptive global reasoning into large-scale object detection. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 6419-6428 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.536, + 0.787, + 0.577 + ], + "angle": 0, + "content": "46. Ye, M., Ke, L., Li, S., Tai, Y.W., Tang, C.K., Danelljan, M., Yu, F.: Cascade-DETR: Delving into high-quality universal object detection. In: Int. Conf. Comput. Vis. pp. 6704-6714 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.578, + 0.787, + 0.618 + ], + "angle": 0, + "content": "47. Zhang, H., Li, F., Liu, S., Zhang, L., Su, H., Zhu, J., Ni, L., Shum, H.Y.: DINO: DETR with improved denoising anchor boxes for end-to-end object detection. In: Int. Conf. Learn. Represent. (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.619, + 0.787, + 0.659 + ], + "angle": 0, + "content": "48. Zhang, H., Wang, Y., Dayoub, F., Sunderhauf, N.: Varifocalnet: An iou-aware dense object detector. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 8514-8523 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.66, + 0.787, + 0.702 + ], + "angle": 0, + "content": "49. Zhang, S., Wang, X., Wang, J., Pang, J., Lyu, C., Zhang, W., Luo, P., Chen, K.: Dense distinct query for end-to-end object detection. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 7329-7338 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.702, + 0.787, + 0.743 + ], + "angle": 0, + "content": "50. Zhao, C., Sun, Y., Wang, W., Chen, Q., Ding, E., Yang, Y., Wang, J.: Ms-detr: Efficient detr training with mixed supervision. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 17027-17036 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.744, + 0.787, + 0.785 + ], + "angle": 0, + "content": "51. Zhu, X., Su, W., Lu, L., Li, B., Wang, X., Dai, J.: Deformable DETR: Deformable transformers for end-to-end object detection. In: Int. Conf. Learn. Represent. (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.786, + 0.787, + 0.813 + ], + "angle": 0, + "content": "52. Zong, Z., Song, G., Liu, Y.: DETRs with collaborative hybrid assignments training. In: Int. Conf. Comput. Vis. pp. 6748-6758 (2023)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.813 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/Relation DETR_ Exploring Explicit Position Relation Prior for Object Detection/650a9d12-9a9f-48d6-b77e-bd08f66392b3_origin.pdf b/2024/Relation DETR_ Exploring Explicit Position Relation Prior for Object Detection/650a9d12-9a9f-48d6-b77e-bd08f66392b3_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..df04d91dea5aed3c8b40f2e72a3ed409517665dd --- /dev/null +++ b/2024/Relation DETR_ Exploring Explicit Position Relation Prior for Object Detection/650a9d12-9a9f-48d6-b77e-bd08f66392b3_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e59a0d52c7af955e7424725e29e46b14637ec4dc6737419087ceea407a07a20e +size 1351506 diff --git a/2024/Relation DETR_ Exploring Explicit Position Relation Prior for Object Detection/full.md b/2024/Relation DETR_ Exploring Explicit Position Relation Prior for Object Detection/full.md new file mode 100644 index 0000000000000000000000000000000000000000..fb35d9fa2b1c5425c8f885810cc34bf0f77cf4d4 --- /dev/null +++ b/2024/Relation DETR_ Exploring Explicit Position Relation Prior for Object Detection/full.md @@ -0,0 +1,304 @@ +# Relation DETR: Exploring Explicit Position Relation Prior for Object Detection + +Xiuquan Hou $^{1}$ , Meiqin Liu $^{1,2,\star}$ , Senlin Zhang $^{2}$ , Ping Wei $^{1}$ , Badong Chen $^{1}$ , and Xuguang Lan $^{1}$ + +$^{1}$ National Key Laboratory of Human-Machine Hybrid Augmented Intelligence, Xi'an Jiaotong University, Xi'an 710049, China + +$^{2}$ College of Electrical Engineering, Zhejiang University, Hangzhou 310027, China +xiuqhou@stu.xjtu.edu.cn, {liumeiqin, slzhang}@zju.edu.cn, {pingwei, chenbd, xglan} @mail.xjtu.edu.cn + +Abstract. This paper presents a general scheme for enhancing the convergence and performance of DETR (DEtection Transformer). We investigate the slow convergence problem in transformers from a new perspective, suggesting that it arises from the self-attention that introduces no structural bias over inputs. To address this issue, we explore incorporating position relation prior as attention bias to augment object detection, following the verification of its statistical significance using a proposed quantitative macroscopic correlation (MC) metric. Our approach, termed Relation-DETR, introduces an encoder to construct position relation embeddings for progressive attention refinement, which further extends the traditional streaming pipeline of DETR into a contrastive relation pipeline to address the conflicts between non-duplicate predictions and positive supervision. Extensive experiments on both generic and task-specific datasets demonstrate the effectiveness of our approach. Under the same configurations, Relation-DETR achieves a significant improvement $(+2.0\%)$ AP compared to DINO), state-of-the-art performance $(51.7\%)$ AP for $1\times$ and $52.1\%$ AP for $2\times$ settings), and a remarkably faster convergence speed (over $40\%$ AP with only 2 training epochs) than existing DETR detectors on COCO val2017. Moreover, the proposed relation encoder serves as a universal plug-in-and-play component, bringing clear improvements for theoretically any DETR-like methods. Furthermore, we introduce a class-agnostic detection dataset, SA-Det-100k. The experimental results on the dataset illustrate that the proposed explicit position relation achieves a clear improvement of $1.3\%$ AP, highlighting its potential towards universal object detection. The code and dataset are available at https://github.com/xiuqhou/Relation-DETR. + +Keywords: Detection transformer $\cdot$ Object detection $\cdot$ Relation network $\cdot$ Progressive attention refinement $\cdot$ Feature enhancement + +# 1 Introduction + +Object detection aims to tackle the problems of bounding box regression and object classification for each object of interest. Recently, DETection TRansformer (DETR) [4] has overcome the reliance on handcrafted designs of convolution detectors, achieving an elegant architecture in an end-to-end manner. + +Despite exhibiting impressive detection performance on large-scale datasets such as COCO [35], their performances are prone to be influenced by dataset scale and suffer from slow convergence. The root cause of the problem is the conflict between non-duplicate predictions and positive supervision [23]. During the training process, DETR employs the Hungarian algorithm to assign a single positive prediction to each ground-truth for producing unique results. However, this leads to negative predictions dominating the majority of the loss function, causing insufficient positive supervision. Therefore, more samples and iterations are required for convergence. Previous attempts have explored the issue by introducing train-only architectures (e.g. query denoising [28], multiple groups of queries [6], auxiliary queries [23], collaborative hybrid assignment training [52]) for additional supervision or by incorporating hard mining in loss functions (e.g. IA-BCE loss [3], position-supervised loss [37]). Other works have proposed specific structures for better interaction between queries and feature maps (e.g. dynamic anchor query [36], cascade window attention [46]), as well as techniques to focus on high-quality queries (e.g. hierarchical filtering [20], dense distinct process [49] and query rank layer [40]). Despite these advancements, there has been little exploration of the issue from the perspective of self-attention, which is widely used in the transformer decoders in most DETR detectors. + +The effectiveness of self-attention lies in its establishment of a high-dimensional relation representation among sequence embeddings [4, 42], which also serves as a key component for modeling relations among different detection feature representations [4]. However, such relation is an implicit representation since it assumes no structural bias over inputs, in which even position information is also needed to be learned from training data [33]. Consequently, the learning process of transformer is data-intensive and slow to converge. This analysis motivates us to introduce task-specific bias for realizing faster convergence and reducing data dependence. + +In this paper, we explore enhancing DETR detectors from a novel perspective, namely explicit position relation prior. We first establish a metric for quantifying position relations in images, and analyze the distribution to verify its statistical significance. In this context, we introduce a position relation encoder to model all pairwise interactions between two bounding boxes, employing progressive attention refinement for cross-layer information interaction. To maintain the end-to-end property while providing sufficient positive supervision, we introduce a contrast relation strategy, which leverages both one-to-one and one-to-many matching while emphasizing the influence of position relation on dedduplication. The proposed method is named as Relation-DETR. + +Compared to previous works, the main feature of Relation-DETR is the integration of explicit position relation. In contrast, prior works focus on implicitly + +learned attention weights from training data, leading to slow convergence. Intuitively, our proposed position relation can be seen as a plug-in-and-play design beneficial for non-duplication predictions, since it establishes a representation of relative positions among pairs of bounding boxes (similar to IoU in NMS [16]). + +We evaluate the performance of Relation-DETR on the most popular object detection dataset, COCO 2017 [35], as well as several task-specific datasets [8,43]. The experimental results demonstrate its superior performance, surpassing previous state-of-the-art DETR detectors with clear margins. More specifically, Relation-DETR exhibits a significantly fast convergence speed. Without bells and whistles, it becomes the first DETR detector to achieve $40\%$ AP on COCO with only 2 epochs using ResNet50 as the backbone under the $1\times$ training configuration. In addition, the simple architecture design of our position relation encoder maintains a promising transferability. It can be easily extended to other DETR-based methods with only a few modifications to achieve consistent performance improvements. This is in contrast to some existing DETR detectors whose performance is highly dependent from complex matching strategies [22] or detection heads [52] developed by convolution-based detectors. + +# 2 Related Work + +Transformer for Object Detection In practice, the majority of attempts to apply transformer to object detection involve constructing a parallelizable sequence, either in the feature extractor [30] or in the detection body [4]. Specifically, transformer-based feature extractors generate token sequences based on image patches [12, 30, 39], and extract multi-scale features through aggregating local features [30, 39] or pyramid postprocess [14, 30]. DEtction TRansformer (DETR) proposed by Carion et al. [4] encodes the extracted features into object queries and decodes them into detected bounding boxes and labels. However, the self-learned attention mechanism increases the requirements for large-scale datasets and training iterations. Many works have explored the slow convergence from the perspective of structured attention (e.g. multi-scale deformable attention [51], dynamic attention [10], cascade window attention [46]), queries with explicit priors (e.g. anchor queries [44], dynamic anchor boxes queries [36], denoising queries [28], dense distinct queries [49]), and additional positive supervision (e.g. group queries [6], hybrid design [23], mixed matching [3]). However, even state-of-the-art DETR methods still utilize vanilla multi-head attention in the transformer decoder. And few works have explored the slow convergence from the perspective of implicit priors. This paper aims to address the issue with position relation. + +Relation Network Rather than processing visual features at pixel levels, patch levels or image levels, relation network captures relation features at instance levels. Existing research on relation networks involves category-based and instance-based approaches. The category-based approaches construct conceptual or statistical relations (e.g. co-occurrence probability [17, 27]) either from relation + +datasets like Visual Genome [24, 27, 45] or by adaptively learning from class labels [17]. Both of them, however, increase the complexity due to the assignment between instances and categories [7, 17, 45]. In contrast, instance-based approaches directly construct a fine-grained graph structure given object features as a node set and their relations as a edge set. Therefore, reasoning on the graph during the training process naturally determines the explicit relation weight [21]. Typically, the weight denotes the parametric distances between each paired object instances in high-dimensional space, such as appearance similarity [31], proposal distance [32] or even self-attention weight [42]. Since learning self-attention weight solely from training data without structural bias increases the requirement for dataset scale and iterations, we explore explicit position relation as a prior to reduce the requirement. + +Classification loss for hard mining. During object detection training, positive predictions assigned to ground truth are much fewer than negative predictions, often resulting in imbalanced supervision and slow convergence. For classification tasks, Focal Loss [34] proposes introducing a weight parameter to focus on hard samples, which is further extended to many variants like generalized focal loss (GFL) [29], vari focal loss (VFL) [29]. Moreover, for object detection tasks, using loss with modulation terms based on regression metrics (e.g. TOOD [15], IA-BCE [3], position-supervised loss [37]) further achieves high-quality alignment between classification and regression tasks. + +# 3 Statistical significance of object position relation + +Are objects really correlated in object detection tasks? To answer the question, we propose a quantitative macroscopic correlation (MC) metric based on the Pearson Correlation Coefficient (PCC) to measure the position correlation among objects in a single image. Assume the objects in an image form a node set, and the PCC between each pair of bounding box annotations serves as their corresponding edge weight. We can construct an undirected graph with continuous values. In this end, the macroscopic correlation for each image can be calculated using the graph intensity, formulated as: + +$$ +M C = \frac {\sum_ {i} \sum_ {j : j \neq i} | \operatorname {P e a r s o n} \left(\boldsymbol {b} _ {i} , \boldsymbol {b} _ {j}\right) |}{N (N - 1)} \tag {1} +$$ + +where $N$ denotes the number of objects, i.e. the number of nodes, $\pmb{b} = [x,y,w,h]$ denotes the position annotation of bounding boxes in datasets. $MC = 1$ only when all objects are fully linearly correlated, whereas $MC = 0$ if there is no position correlation between any pair of objects. + +We visualize the statistical distribution of MC for datasets across various scenarios, including industrial (ESD [19], CSD [43], MSSD [8]), domestic (AI2thor [26]), urban (Cityscapes [9]) and generic settings (PascalVOC [13], COCO [35], Object365 [41], SA-1B [25]). The datasets cover a wide range of scales, from + +![](images/db1a7ca603db4b177adec9136dc6a526425d17c13b7216f61756190c50f7fc8e.jpg) +Fig. 1: Statistical distribution of macroscopic correlation (MC) on various datasets (normalized for better visualization), and the values in brackets indicate the number of dataset samples. + +0.3K to 11M images. As shown in Fig. 1, all of these datasets indicate that the distribution of MC is concentrated around high numerical values, with the distribution centers closer to the upper bound. This demonstrates the presence and the statistical significance of object position relation. Specifically, task-specific datasets display more prior knowledge and clearer clustering patterns in the high-dimensional feature space, which thus results in higher MC values than generic datasets like COCO. + +# 4 Relation-DETR + +Given the statistical significance of position relation, we propose a state-of-the-art detector, named Relation-DETR, that explores explicit position relation prior to enhance object detection. To address the issue of slow convergence, we present a position relation encoder (Sec. 4.1) for progressive attention refinement (Sec. 4.2). Further in Sec. 4.3, it extends the streaming pipeline of DETR into a contrast pipeline to emphasize the influence of position relation on removing duplication while maintaining sufficient positive supervision for faster convergence. + +# 4.1 Position relation encoder + +Previous research has demonstrated the relation effectiveness for convolution detectors [21,32]. Recently, some DETR methods attempt to construct instance-level relation by indexing from category-level relation using class indices [17]. In contrast to these approaches, we directly construct instance-level relation through a simple position encoder, maintaining an end-to-end design for DETR. + +We first review the basic pipeline in DETR detectors. Given image features extracted by the backbone, the transformer encoder produces an augmented memory $\mathbf{Z} \in \mathbb{R}^{d \times H \times W}$ for further decoding into bounding boxes $\pmb{b}_i = [x, y, w, h], i = 1, \dots, N$ and class labels $\pmb{c}_i$ as predictions. Each decoder layer refines the bounding box coordinates iteratively by predicting $\Delta$ w.r.t. the coordinate from the last decoder layer, known as iterative bounding box refinement [51]. In addition, predictions from all decoder layers equally participate in loss calculation to compute auxiliary decoding losses [4]. + +Under the aforementioned detection framework, our position relation encoder represents the high-dimensional relation embedding as an explicit prior for self-attention in the transformer. This embedding is calculated based on the predicted bounding boxes (denoted as $\mathbf{b} = [x,y,w,h]$ ) from each decoder layer. To ensure that the relation is invariant to translation and scale transformations, we encode it based on normalized relative geometry features: + +$$ +\boldsymbol {e} (\boldsymbol {b} _ {i}, \boldsymbol {b} _ {j}) = \left[ \log \left(\frac {| x _ {i} - x _ {j} |}{w _ {i}} + 1\right), \log \left(\frac {| y _ {i} - y _ {j} |}{h _ {i}} + 1\right), \log \left(\frac {w _ {i}}{w _ {j}}\right), \log \left(\frac {h _ {i}}{h _ {j}}\right) \right] (2) +$$ + +Unlike [21], our position relation is unbiased, as $\pmb{e}(\pmb{b}_i, \pmb{b}_j) = 0$ when $i = j$ . The relation matrix $\mathbf{E} \in \mathbb{R}^{N \times N \times 4}$ (with $\mathbf{E}(i,j) = \pmb{e}(\pmb{b}_i, \pmb{b}_j)$ ) is further transformed into high-dimensional embeddings through sine-cosine encoding [42]. + +$$ +\operatorname {E m b e d} (\mathbf {E}, 2 k) = \sin \left(s \mathbf {E} / T ^ {2 k / d _ {\mathrm {r e}}}\right) \tag {3} +$$ + +$$ +\operatorname {E m b e d} (\mathbf {E}, 2 k + 1) = \cos (s \mathbf {E} / T ^ {2 k / d _ {\mathrm {r e}}}) \tag {4} +$$ + +where the shape of relation embedding is $N \times N \times 4d_{\mathrm{re}}$ , and $T$ , $d_{\mathrm{re}}$ , $s$ are encoding parameters. Finally, the embedding undergoes a linear transformation to obtain $M$ scalar weights, where $M$ denotes the number of attention heads. + +$$ +\operatorname {R e l} (\boldsymbol {b}, \boldsymbol {b}) = \max \left(\epsilon , \mathbf {W E m b e d} (\boldsymbol {b}, \boldsymbol {b}) + \mathbf {B}\right) \tag {5} +$$ + +where $\epsilon$ makes sure a positive value for relation to avoid gradient vanishing after $\exp$ when integrated into self-attention, and $\mathsf{Rel}(\pmb {b},\pmb {b})\in \mathbb{R}^{N\times N\times M}$ . + +# 4.2 Progressive attention refinement with position relation + +The iterative box refinement proposed by Deformable-DETR [51] has shown the effectiveness for high-quality bounding box regression. Following the motivation, we propose a progressive attention refinement method to introduce the position relation into the streaming pipeline of DETR. Specifically, the relation of layer- $i$ is determined by bounding boxes of both layer- $i - 1$ and layer- $i$ , which is further integrated into self-attention for producing the bounding boxes in layer- $(i + 1)$ . + +$$ +\operatorname {A t t n} _ {\text {S e l f}} \left(\mathbf {Q} ^ {l}\right) = \operatorname {S o f t m a x} \left(\operatorname {R e l} \left(\boldsymbol {b} ^ {l - 1}, \boldsymbol {b} ^ {l}\right) + \frac {\operatorname {Q u e} \left(\mathbf {Q} ^ {l}\right) \operatorname {K e y} \left(\mathbf {Q} ^ {l}\right) ^ {\top}}{\sqrt {d _ {\text {m o d e l}}}}\right) \operatorname {V a l} \left(\mathbf {Q} ^ {l}\right) \tag {6} +$$ + +$$ +\mathbf {Q} ^ {l + 1} = \text {F F N} \left(\mathbf {Q} ^ {l} + \operatorname {A t t n} _ {\text {c r o s s}} \left(\operatorname {A t t n} _ {\text {s e l f}} (\mathbf {Q} ^ {l}), \operatorname {K e y} (\mathbf {Z}), \operatorname {V a l} (\mathbf {Z})\right)\right) \tag {7} +$$ + +$$ +\boldsymbol {b} ^ {l + 1} = \operatorname {M L P} \left(\mathbf {Q} ^ {l + 1}\right), \boldsymbol {c} ^ {l + 1} = \operatorname {L i n e a r} \left(\mathbf {Q} ^ {l + 1}\right) \tag {8} +$$ + +where $\mathbf{Q}^l$ denotes the queries in the $l$ -th decoder layer in the DETR transformer, $\mathbf{Z}$ is the memory, i.e., the enhanced image features from the transformer encoder. + +![](images/b279f4e719153e33370759b926e9c57cda4e1343abbfd03d603bdc9eb04ad97d.jpg) +Fig. 2: Comparison of transformer decoder in Deformable-DETR(left) and Relation-DETR(right). + +![](images/a1b9f1abfc871fc81b2ed4376a5577eea72c11ba94458b93f22e7b8ce7668231.jpg) + +The only difference between our method and the existing DETR decoder is marked in red. As depicted in Fig. 2, the requisite addition involves the introduction of a lateral branch for the computation of position relation. Therefore, our position relation and the progressive attention refinement are straightforward, allowing for plug-in-and-play integration with self attention in existing DETR detectors to achieve consistent performance improvements (see Tab. 6). + +# 4.3 Contrast relation pipeline + +Rethinking the mechanism of existing duplication removal methods (including NMS [16], Soft-NMS [1], fast-NMS [2], Adaptive-NMS [38]), these processes heavily rely on IoU (intersection over Union), which, to some extent, signifies the position relation between bounding boxes. Therefore, we may hypothesize that integrating the position relation among queries in self-attention contributes to non-duplicated predictions in object detection, akin to [22]. + +The conflicts between non-duplicate predictions and sufficient positive supervision arise from the streaming pipeline of DETR, which must navigate between one-to-one matching and one-to-many matching. To overcome this limitation, we extend it to a contrast pipeline based on the proposed position relation. Specifically, we construct two parallel sets of queries, i.e. matching queries $\mathbf{Q}_m$ and hybrid queries $\mathbf{Q}_h$ . Both are input into the transformer decoder but undergo distinct processing. The matching queries are processed with self-attention incorporating position relation to produce non-duplicated predictions: + +$$ +\operatorname {A t t n} _ {\text {S e l f}} \left(\mathbf {Q} _ {\mathrm {m}} ^ {l}\right) = \operatorname {S o f t m a x} \left(\operatorname {R e l} \left(\boldsymbol {b} ^ {l - 1}, \boldsymbol {b} ^ {l}\right) + \frac {\operatorname {Q u e} \left(\mathbf {Q} _ {\mathrm {m}}\right) \operatorname {K e y} \left(\mathbf {Q} _ {\mathrm {m}}\right) ^ {\top}}{\sqrt {d _ {\text {m o d e l}}}}\right) \operatorname {V a l} \left(\mathbf {Q} _ {\mathrm {m}}\right) \tag {9} +$$ + +$$ +\operatorname {A t t n} _ {\text {S e l f}} \left(\mathbf {Q} _ {\mathrm {h}} ^ {l}\right) = \operatorname {S o f t m a x} \left(\frac {\operatorname {Q u e} \left(\mathbf {Q} _ {\mathrm {h}}\right) \operatorname {K e y} \left(\mathbf {Q} _ {\mathrm {h}}\right) ^ {\top}}{\sqrt {d _ {\text {m o d e l}}}}\right) \operatorname {V a l} \left(\mathbf {Q} _ {\mathrm {h}}\right) \tag {10} +$$ + +while the hybrid queries are decoded by the same decoder but skip the calculation of position relation to explore more potential candidates. Their corresponding predictions are denoted as $\pmb{p}_m^l = (\pmb{b}_m^l, \pmb{c}_m^l)$ and $\pmb{p}_h^l = (\pmb{b}_h^l, \pmb{c}_h^l)$ , respectively. Details of the contrast relation pipeline are illustrated in Fig. 3. + +![](images/734495d4255b2af0c1253dff6b4627e00d65b7adb28330f26d43e0df3546b13c.jpg) +Fig. 3: Detailed illustration of the proposed contrast relation pipeline. + +![](images/a28699bd84ffaf677331794c6cd14f2ea24e6f3e54c3b984eef3b7be26635ca4.jpg) + +Assuming $\pmb{g}$ denotes the ground truth annotations, for $\pmb{p}_m$ , we employ a one-to-one matching scheme to emphasize the non-duplicate property, with loss calculation similar to the original DETR approach [4]: + +$$ +\mathcal {L} _ {m} \left(\boldsymbol {p} _ {m}, \boldsymbol {g}\right) = \sum_ {l = 1} ^ {L} \mathcal {L} _ {\text {H u n g a r i a n}} \left(\boldsymbol {p} _ {m} ^ {l}, \boldsymbol {g}\right) \tag {11} +$$ + +While for $\pmb{p}_h$ , a one-to-many matching scheme is employed to form more potential positive candidates. We simply follow H-DETR [23] and repeat the ground truth $K$ times, denoted as $\tilde{\pmb{g}} = \{\pmb{g}^1, \pmb{g}^2, \dots, \pmb{g}^K\}$ , for loss calculation: + +$$ +\mathcal {L} _ {h} \left(\boldsymbol {p} _ {h}, \boldsymbol {g}\right) = \sum_ {l = 1} ^ {L} \mathcal {L} _ {\text {H u n g a r i a n}} \left(\boldsymbol {p} _ {h} ^ {l}, \tilde {\boldsymbol {g}}\right) \tag {12} +$$ + +where $\mathcal{L}_{\mathrm{Hungarian}}$ denotes the Hungarian loss, and $L$ denotes the number of decoder layers. It is worth noting that the hybrid queries are only involved during training, thus incurring no extra computational burden for inference. + +# 5 Experimental Results and Discussion + +# 5.1 Setup + +For a comprehensive evaluation, we conduct experiments on both an object detection benchmark, COCO 2017 [35], and two task-specific datasets, CSD [43] + +and MSSD [8]. Detection performance is measured using the standard Average Precision (AP) [35]. We train our model on NVIDIA A800 GPU (80GB) and RTX 3090 GPU (24GB) using the AdamW optimizer with an initial learning rate of $1 \times 10^{-4}$ and weight decay of $1 \times 10^{-4}$ . The Relation-DETR is implemented based on the backbone ResNet-50 [18] and Swin-L [39] pretrained from ImageNet [11], which are finetuned with a learning rate $1 \times 10^{-5}$ during training. The learning rate is reduced by a factor of 0.1 at later stages. The parameters of the position relation encoder, $T$ , $d_{\mathrm{re}}$ , $s$ are empirically chosen as 10000, 16, 100, respectively. The hybrid training configurations follow DINO [47] and H-DETR [23], i.e., $N_{m} = 900$ , $N_{h} = 1500$ , $k = 6$ . We adopt the VariFocal loss [48] for training Relation-DETR. The training batch size is 10 for COCO 2017 and 2 for task-specific datasets. Before fed into detectors, images undergo the same augmentations (random resize, crop and flip) as other DETR detectors. + +# 5.2 Comparison with state-of-the-art methods + +Comparison on COCO 2017. Tab. 1 presents the detection performance on COCO val 2017. Compared to other state-of-the-art DETR methods, our approach converges much faster and demonstrates significant improvements of $1.0\%$ AP, $1.0\% \mathrm{AP}_{50}$ and $0.6\% \mathrm{AP}_{75}$ , respectively, suppressing the second best DDQ-DETR [49] with clear margins. Specifically, Relation-DETR achieves $51.7\%$ AP using only 12 epochs with the ResNet-50 backbone and even outperforms DINO [47] (51.2% AP) with 36 epochs ( $3\times$ faster). More importantly, in contrast to DDQ-DETR [49] and Co-DETR [52] that leverages NMS in the decoder or postprocess for improving precision, our Relation-DETR maintains an end-to-end pipeline, ensuring promising extensibility. When integrated with Swin-L backbone, Relation-DETR outperforms all counterparts, achieving the best $57.8\%$ AP with a $0.5\%$ AP improvement, showcasing its excellent scalability for larger model capacity. + +Comparison on task-specific datasets. Different from generic object detection benchmarks, datasets in task-specific scenarios lack sufficient samples to provide semantic information. To reveal the generalizability of Relation-DETR, we conduct a performance comparison on two defect detection datasets, i.e. CSD [43] and MSSD [8]. The results in Tab. 3 show that, Relation-DETR improves the baseline DINO by $1.4\%$ AP on CSD, achieving the highest $54.4\%$ AP. Tab. 4 demonstrates that Relation-DETR further increases the margin to $6.4\%$ AP on MSSD and surpasses other counterparts. It is noteworthy that CSD [43] and MSSD [8] contain more small-sized objects than COCO 2017, which thus confirms the effectiveness of Relation-DETR to small-sized detection. Moreover, considering a stricter IoU threshold, Relation-DETR outperforms the second best method DINO by a significant margin of $11.1\%$ AP@75, highlighting the beneficial impact of explicit position relation on high-quality predictions. + +Table 1: Comparison with state-of-the-art methods on COCO val2017 using ResNet-50(IN-1K) backbone. The * means that we re-implement the methods and report the corresponding results. + +
MethodBackboneEpochsAP@50:95AP@50AP@75\( AP_S \)\( AP_M \)\( AP_L \)
Anchor-DETR [44]ResNet-505042.163.144.922.346.260.0
Def-DETR* [51]ResNet-501245.465.049.127.249.661.0
DAB-Def-DETR* [36]ResNet-501246.365.950.429.350.561.9
DN-Def-DETR* [28]ResNet-501247.265.351.330.250.662.4
DINO* [47]ResNet-501249.967.454.533.953.563.8
Group-DETR [6]ResNet-501249.832.453.064.2
H-Def-DETR [23]ResNet-501248.766.452.931.251.563.5
Cascade-DETR [46]ResNet-501249.767.154.132.453.565.1
Co-Def-DETR [52]ResNet-501249.567.654.332.452.763.7
Align-DETR [3]ResNet-501250.267.854.432.953.365.0
Stable-DINO [37]ResNet-501250.467.455.032.954.065.5
DAC-DETR [22]ResNet-501250.067.654.7
Salience-DETR* [20]ResNet-501250.067.754.233.354.464.4
Rank-DETR [40]ResNet-501250.467.955.233.653.864.2
MS-DETR [50]ResNet-501250.367.455.132.754.064.6
DDQ-DETR [49]ResNet-501250.768.155.7
Relation-DETRResNet-501251.769.156.336.155.666.1
DINO [47]ResNet-502450.468.354.833.353.764.8
DINO [47]ResNet-503651.269.055.835.054.365.3
H-Def-DETR [23]ResNet-503650.068.354.432.952.765.3
Salience-DETR* [20]ResNet-502451.268.955.733.955.565.6
DDQ-DETR [49]ResNet-502452.069.557.235.254.965.9
Relation-DETRResNet-502452.169.756.636.156.066.5
+ +Table 2: Comparison with state-of-the-art methods on COCO val2017 using SwinL(IN-22K) as the backbone. + +
MethodBackboneEpochsAP@50:95AP@50AP@75\( AP_S \)\( AP_M \)\( AP_L \)
DINO [47]Swin-L1256.875.462.341.160.673.5
\( \mathcal{H} \text{-Def-DETR} \) [23]Swin-L1255.975.261.039.159.972.2
Salience-DETR* [20]Swin-L1256.575.061.540.261.272.8
Rank-DETR [40]Swin-L1257.375.962.940.861.373.2
Rank-DINO [40]Swin-L1257.676.063.441.661.473.8
Relation-DETRSwin-L1257.876.162.941.262.174.4
+ +# 5.3 Ablation study + +This part conducts an ablation study to explore how the proposed components influence the final detection performance on COCO. The results in Tab. 5 show that, each key component of Relation-DETR consistently contributes to improving AP. Even on a highly-optimized baseline with VariFocal loss [48], our posi + +Table 3: Quantitative comparison on CSD [43]. + +
MethodBackboneEpochsAP@50:95AP@50AP@75\( AP_S \)\( AP_M \)\( AP_L \)
Def-DETR [51]ResNet-5030043.786.236.640.534.910.0
DAB-Def-DETR [36]ResNet-509052.991.255.050.339.40.0
DN-Def-DETR [28]ResNet-506049.988.051.247.637.70.0
DINO [47]ResNet-506053.090.855.550.939.60.0
H-Def-DETR [23]ResNet-506053.090.655.751.239.26.7
Salience-DETR [20]ResNet-506053.292.555.151.040.90.0
Relation-DETRResNet-506054.492.956.053.640.80.0
+ +Table 4: Quantitative comparison on MSSD [8]. + +
MethodBackboneEpochsAP@50:95AP@50AP@75APSAPMAPL
Def-DETR [51]ResNet-5030033.054.332.09.811.133.3
DAB-Def-DETR [36]ResNet-5012033.760.031.015.926.729.0
DN-Def-DETR [28]ResNet-5012045.674.144.918.631.941.0
H-Def-DETR [23]ResNet-5012046.976.847.120.345.340.7
DINO [47]ResNet-5012051.080.052.520.047.444.8
Salience-DETR [20]ResNet-5012055.478.261.928.747.544.5
Relation-DETRResNet-5012057.479.263.631.453.545.5
+ +tion relation encoder and contrast pipeline bring clear improvements of $+0.3\%$ , $+0.5\%$ AP, respectively. Built upon normalized relative geometry features, the position relation overcomes scale bias effectively, thus benefiting consistent performance improvements for different sized objects. For instance, Tab. 5 shows that introducing relation into the baseline with VFL achieves $+1.2\%$ $\mathrm{AP}_S$ , $+1.0\%$ $\mathrm{AP}_M$ , and $+1.3\%$ $\mathrm{AP}_L$ . + +Table 5: Ablation study on key components of Relation-DETR (ResNet-50, 1x). We study both baseline and an optimized version with VariFocal Loss [48]. + +
VFLRel.contrastAP@50:95AP@50AP@75APSAPMAPL
49.967.454.533.953.563.8
✓50.3(↑0.4)68.1(↑0.7)55.0(↑0.5)34.7(↑0.8)53.4(↓0.1)64.7(↑0.9)
✓50.9(↑0.6)68.2(↑0.1)55.3(↑0.3)34.9(↑0.2)54.6(↑1.2)64.8(↑0.1)
✓✓51.2(↑0.3)68.5(↑0.7)55.7(↑0.4)35.3(↑0.4)54.9(↑0.3)65.5(↑0.7)
✓✓✓51.7(↑0.5)69.1(↑0.6)56.3(↑0.6)36.1(↑0.8)55.6(↑0.7)66.1(↑0.6)
+ +# 5.4 Transferability of position relation + +Our position relation encoder adopts an elegant architectural design, ensuring a promising transferability to existing DETR detectors with minimal modifications. The experimental results in Tab. 6 show that, integrating position relation encoders without any further modifications enhances detection performance with clear margins of $1.6\%$ , $2.0\%$ , $0.1\%$ and $0.2\%$ for Deformable-DETR [51], DAB-Deformable-DETR [36], DN-Deformable-DETR [28] and DINO [47], respectively. Interestingly, in comparison to $\mathrm{AP}_M$ and $\mathrm{AP}_L$ , the position relation has a more substantial impact on improving $\mathrm{AP}_S$ for Deformable-DETR [51] and DAB-Deformable-DETR [36]. We attribute this to the fact that these early proposed baselines introduce relatively less structural bias and thus benefit more from our explicit position relation prior. + +Table 6: Transfer experiments for the position relation encoder (ResNet-50, $1 \times$ ). "+RelEnc" denotes the version that integrates our position relation encoder. + +
MethodAP@50:95AP@50AP@75APSAPMAPL
Def-DETR45.465.049.127.249.661.0
+RelEnc47.0(↑1.6)65.6(↑0.6)51.2(↑2.1)29.3(↑2.1)51.0(↑1.4)62.2(↑1.2)
DAB-Def-DETR46.365.950.428.350.561.9
+RelEnc48.3(↑2.0)66.5(↑0.6)52.9(↑2.5)32.4(↑4.1)52.0(↑1.5)62.0(↑0.1)
DN-Def-DETR47.265.351.330.250.662.4
+RelEnc47.3(↑0.1)65.6(↑0.3)51.4(↑0.1)29.9(↓0.3)50.8(↑0.2)62.1(↓0.3)
+ +Moreover, the proposed contrast pipeline can be seen as an extension of hybrid matching [23], utilizing the proposed position relation encoder. Table 7 compares their transferability when integrated into DINO. The results indicate that directly applying hybrid matching [23] to DINO [47] results in a decrease in performance, from $49.9\%$ AP to $49.5\%$ AP. In contrast, the introduction of both the proposed relation encoder and the extended contrast pipeline consistently increases the performance. This demonstrates the effectiveness of the proposed position relation prior in improve detection performance, overcoming the weak generalizability inherent in hybrid matching. + +Table 7: Transfer experiments compared with hybrid matching [23] on DINO. + +
DINO+Hybrid [23]+RelEnc+RelEnc+Contrast
APAP@50AP@75APAP@50AP@75APAP@50AP@75APAP@50AP@75
49.967.454.549.5†66.654.050.368.155.051.268.755.6
+ +† We found that hybrid matching decreases the performance of DINO. The conclusion is consistent with the results of HDINO (see here) reported by MMDetection [5]. + +# 5.5 Intuitive performance comparison + +To facilitate an intuitive performance comparison, Fig. 4 plots the convergence curve and the precision-recall curve. Because the position relation prior reduces the requirement for learning structural bias from data [33], Relation-DETR illustrates a faster convergence speed. When training from scratch, it achieves a higher AP than other counterparts with fewer iterations. Specifically, Relation-DETR can achieve over $40\%$ AP with only 2 epochs, surpassing existing DETR detectors. In addition to convergence speed, the PR curves under different IoU thresholds also validate the performance gain of our Relation-DETR. + +![](images/237f844b9eacd221aacb342100b5e9d2a9a39929a679215ff5dc8119708deed5.jpg) +Fig. 4: Convergence curve(left) and Precision-recall curve for $\mathrm{IoU} = 50\% \sim 95\%$ (right). All models are trained with ResNet-50 backbone under the same $1\times$ training configuration on COCO 2017. + +![](images/c90bae7c9bf3079eb86772dc3acbcda46503e06a5baa859fd2e881fa1d160ffd.jpg) + +# 5.6 Visualization + +For a more intuitive grasp of the relation mechanism, Fig. 5 illustrates representative objects with high relation weights when given a query object. The visualization shows that for both generic and task-specific datasets, the relation contributes to identifying other detection candidates based on the given object query. Furthermore, small-sized objects tend to establish more relation connections with other objects due to the lack of their own semantic information. Therefore, constructing relations is crucial for small-sized object detection. Fig. 6 further visualizes some failure cases of Relation-DETR, indicating that the presented model may benefit from occluded objects and dense objects with misleading semantic differences by considering more complex relations, such as occlusion and semantic relations. + +# 5.7 Towards universal object detection + +Will the position relation prior still be effective for datasets that cover a broader range of scenarios and objects? As a piece of general prior knowledge, we an- + +![](images/b88e6cea89b1cb55de875bdcd9cd43af9d578a420d2331248e09fe803cf1c57b.jpg) +Fig. 5: Representative objects(red) related to the Fig. 6: Predictions (left) and GTs given object(blue) (right) of failure cases + +![](images/910b96b06478edf672035823b8bebc20795af2d27ea30a702cfd0e8fdd2a862e.jpg) + +![](images/4f4eada323159925d9da589f54b79022b2ec0cd7ccc3ae9d98bd3ccbfe77a7f3.jpg) + +![](images/6013a8ef88a60636fba0ab2dd2cafa920d9dc5cf53c46ec8c5c6278c2dff689e.jpg) + +![](images/229d92370487151532f69edba696994cf966f2f18b007b085ce8fd8b40ac4ddd.jpg) + +ticipate that the explicit position relation prior can benefit universal object detection tasks. To explore this, we have constructed a large-scale class-agnostic detection dataset with about 100,000 images, termed SA-Det-100k, by sampling a subset from SA-1B, which is one of the largest scale segmentation datasets proposed in Segment Anything [25]. We then compared the performance of our Relation-DETR with the baseline DINO [47] using VFL [48] on this dataset. The results in Tab. 8 show that Relation-DETR achieves a clear improvement of $1.3\%$ AP, demonstrating the scalability of the proposed position relation prior. + +Table 8: Quantitative comparison on SA-Det-100k (ResNet-50, $1 \times$ + +
MethodsAP@50:95AP@50AP@75APSAPMAPL
DINO with VFL43.752.047.75.843.061.5
Relation-DETR45.0(↑1.3)53.1(↑1.1)48.9(↑1.2)6.0(↑0.2)44.4(↑1.4)62.9(↑1.4)
+ +# 6 Conclusion + +This paper explores explicit position relation prior for enhancing performance and convergence of DETR detectors. Built upon normalized relative geometry features, we propose a novel position relation that overcomes scale bias for progressive attention refinement. To address the conflicts between non-duplicate predictions and sufficient positive supervision in DETR frameworks, we extend the streaming pipeline to a contrast pipeline based on the proposed position relation. Combining these components produces a state-of-the-art detector, named Relation-DETR. Massive ablation studies and experimental results demonstrate the superior performance, faster convergence and promising transferability of the proposed detector. Moreover, Relation-DETR exhibits remarkable generalizability for both generic and task-specific detection tasks. We believe that this work will inspire future research on relation and structural bias for DETR detectors. + +# Acknowledgements + +This work was supported by the National Natural Science Foundation of China under Grant 62327808, and the Fundamental Research Funds for Xi'an Jiaotong University under Grants xtr072022001 and xzy022024009. + +# References + +1. Bodla, N., Singh, B., Chellappa, R., Davis, L.S.: Soft-NMS-improving object detection with one line of code. In: Int. Conf. Comput. Vis. pp. 5561-5569 (2017) +2. Bolya, D., Zhou, C., Xiao, F., Lee, Y.J.: YOLACT: Real-time instance segmentation. In: Int. Conf. Comput. Vis. pp. 9157-9166 (2019) +3. Cai, Z., Liu, S., Wang, G., Ge, Z., Zhang, X., Huang, D.: Align-DETR: Improving DETR with simple IoU-aware BCE loss. arXiv preprint arXiv:2304.07527 (2023) +4. Carion, N., Massa, F., Synnaeve, G., Usunier, N., Kirillov, A., Zagoruyko, S.: End-to-end object detection with transformers. In: Eur. Conf. Comput. Vis. pp. 213-229. Springer (2020) +5. Chen, K., Wang, J., Pang, J., Cao, Y., Xiong, Y., Li, X., Sun, S., Feng, W., Liu, Z., Xu, J., et al.: Mmdetection: Open mmlab detection toolbox and benchmark. arXiv preprint arXiv:1906.07155 (2019) +6. Chen, Q., Chen, X., Wang, J., Zhang, S., Yao, K., Feng, H., Han, J., Ding, E., Zeng, G., Wang, J.: Group DETR: Fast DETR training with group-wise one-to-many assignment. In: Int. Conf. Comput. Vis. pp. 6633-6642 (2023) +7. Chen, X., Li, L.J., Fei-Fei, L., Gupta, A.: Iterative visual reasoning beyond convolutions. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 7239-7248 (2018) +8. Chen, Y., Pan, J., Lei, J., Zeng, D., Wu, Z., Chen, C.: EEE-Net: Efficient edge enhanced network for surface defect detection of glass. IEEE Transactions on Instrumentation and Measurement (2023) +9. Cordts, M., Omran, M., Ramos, S., Rehfeld, T., Enzweiler, M., Benenson, R., Franke, U., Roth, S., Schiele, B.: The cityscapes dataset for semantic urban scene understanding. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 3213-3223 (2016) +10. Dai, X., Chen, Y., Yang, J., Zhang, P., Yuan, L., Zhang, L.: Dynamic DETR: End-to-end object detection with dynamic attention. In: Int. Conf. Comput. Vis. pp. 2988-2997 (2021) +11. Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: ImageNet: A large-scale hierarchical image database. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 248-255. IEEE (2009) +12. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. In: Int. Conf. Learn. Represent. (2020) +13. Everingham, M., Van Gool, L., Williams, C.K., Winn, J., Zisserman, A.: The pascal visual object classes (voc) challenge. Int. J. Comput. Vis. 88, 303-338 (2010) +14. Fang, Y., Sun, Q., Wang, X., Huang, T., Wang, X., Cao, Y.: EVA-02: A visual representation for neon genesis. arXiv preprint arXiv:2303.11331 (2023) +15. Feng, C., Zhong, Y., Gao, Y., Scott, M.R., Huang, W.: TOOD: Task-aligned one-stage object detection. In: Int. Conf. Comput. Vis. pp. 3490-3499. IEEE Computer Society (2021) +16. Girshick, R.: Fast R-CNN. In: Int. Conf. Comput. Vis. pp. 1440-1448 (2015) + +17. Hao, X., Huang, D., Lin, J., Lin, C.Y.: Relation-enhanced DETR for component detection in graphic design reverse engineering. In: IJCAI. pp. 4785-4793 (2023) +18. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 770-778 (2016) +19. Hou, X., Liu, M., Zhang, S., Wei, P., Chen, B.: CANet: Contextual information and spatial attention based network for detecting small defects in manufacturing industry. Pattern Recognition 140, 109558 (2023) +20. Hou, X., Liu, M., Zhang, S., Wei, P., Chen, B.: Salience detr: Enhancing detection transformer with hierarchical salience filtering refinement. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 17574-17583 (2024) +21. Hu, H., Gu, J., Zhang, Z., Dai, J., Wei, Y.: Relation networks for object detection. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 3588-3597 (2018) +22. Hu, Z., Sun, Y., Wang, J., Yang, Y.: DAC-DETR: Divide the attention layers and conquer. Adv. Neural Inform. Process. Syst. 36 (2024) +23. Jia, D., Yuan, Y., He, H., Wu, X., Yu, H., Lin, W., Sun, L., Zhang, C., Hu, H.: DETRs with hybrid matching. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 19702-19712 (2023) +24. Jiang, C., Xu, H., Liang, X., Lin, L.: Hybrid knowledge routed modules for large-scale object detection. Adv. Neural Inform. Process. Syst. 31 (2018) +25. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., et al.: Segment anything. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 4015-4026 (2023) +26. Kolve, E., Mottaghi, R., Han, W., VanderBilt, E., Weihs, L., Herrasti, A., Deitke, M., Ehsani, K., Gordon, D., Zhu, Y., et al.: Ai2-thor: An interactive 3d environment for visual ai. arXiv preprint arXiv:1712.05474 (2017) +27. Krishna, R., Zhu, Y., Groth, O., Johnson, J., Hata, K., Kravitz, J., Chen, S., Kalantidis, Y., Li, L.J., Shamma, D.A., et al.: Visual genome: Connecting language and vision using crowdsourced dense image annotations. Int. J. Comput. Vis. 123, 32-73 (2017) +28. Li, F., Zhang, H., Liu, S., Guo, J., Ni, L.M., Zhang, L.: DN-DETR: Accelerate DETR training by introducing query denoising. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 13619-13627 (2022) +29. Li, X., Wang, W., Wu, L., Chen, S., Hu, X., Li, J., Tang, J., Yang, J.: Generalized focal loss: Learning qualified and distributed bounding boxes for dense object detection. Adv. Neural Inform. Process. Syst. 33, 21002-21012 (2020) +30. Li, Y., Mao, H., Girshick, R., He, K.: Exploring plain vision transformer backbones for object detection. In: Eur. Conf. Comput. Vis. pp. 280-296. Springer (2022) +31. Li, Z., Du, X., Cao, Y.: Gar: Graph assisted reasoning for object detection. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 1295-1304 (2020) +32. Lin, J., Pan, Y., Lai, R., Yang, X., Chao, H., Yao, T.: Core-Text: Improving scene text detection with contrastive relational reasoning. In: Int. Conf. Multimedia and Expo. pp. 1-6. IEEE (2021) +33. Lin, T., Wang, Y., Liu, X., Qiu, X.: A survey of transformers. AI Open (2022) +34. Lin, T.Y., Goyal, P., Girshick, R., He, K., Dollar, P.: Focal loss for dense object detection. In: Int. Conf. Comput. Vis. pp. 2980-2988 (2017) +35. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollar, P., Zitnick, C.L.: Microsoft COCO: Common objects in context. In: Eur. Conf. Comput. Vis. pp. 740-755. Springer (2014) + +36. Liu, S., Li, F., Zhang, H., Yang, X., Qi, X., Su, H., Zhu, J., Zhang, L.: DABDETR: Dynamic anchor boxes are better queries for DETR. In: Int. Conf. Learn. Represent. (2021) +37. Liu, S., Ren, T., Chen, J., Zeng, Z., Zhang, H., Li, F., Li, H., Huang, J., Su, H., Zhu, J., et al.: Detection transformer with stable matching. arXiv preprint arXiv:2304.04742 (2023) +38. Liu, S., Huang, D., Wang, Y.: Adaptive NNS: Refining pedestrian detection in a crowd. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 6459-6468 (2019) +39. Liu, Z., Lin, Y., Cao, Y., Hu, H., Wei, Y., Zhang, Z., Lin, S., Guo, B.: Swin transformer: Hierarchical vision transformer using shifted windows. In: Int. Conf. Comput. Vis. pp. 10012-10022 (2021) +40. Pu, Y., Liang, W., Hao, Y., Yuan, Y., Yang, Y., Zhang, C., Hu, H., Huang, G.: Rank-DETR for high quality object detection. Adv. Neural Inform. Process. Syst. 36 (2024) +41. Shao, S., Li, Z., Zhang, T., Peng, C., Yu, G., Zhang, X., Li, J., Sun, J.: Objects365: A large-scale, high-quality dataset for object detection. In: Int. Conf. Comput. Vis. pp. 8430-8439 (2019) +42. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. Adv. Neural Inform. Process. Syst. 30 (2017) +43. Wang, Q., Gao, S., Xiong, L., Liang, A., Jiang, K., Zhang, W.: A casting surface dataset and benchmark for subtle and confusable defect detection in complex contexts. IEEE Sensors Journal pp. 1-1 (2024) +44. Wang, Y., Zhang, X., Yang, T., Sun, J.: Anchor DETR: Query design for transformer-based detector. In: AAAI. vol. 36, pp. 2567-2575 (2022) +45. Xu, H., Jiang, C., Liang, X., Lin, L., Li, Z.: Reasoning-RCNN: Unifying adaptive global reasoning into large-scale object detection. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 6419-6428 (2019) +46. Ye, M., Ke, L., Li, S., Tai, Y.W., Tang, C.K., Danelljan, M., Yu, F.: Cascade-DETR: Delving into high-quality universal object detection. In: Int. Conf. Comput. Vis. pp. 6704-6714 (2023) +47. Zhang, H., Li, F., Liu, S., Zhang, L., Su, H., Zhu, J., Ni, L., Shum, H.Y.: DINO: DETR with improved denoising anchor boxes for end-to-end object detection. In: Int. Conf. Learn. Represent. (2022) +48. Zhang, H., Wang, Y., Dayoub, F., Sunderhauf, N.: Varifocalnet: An iou-aware dense object detector. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 8514-8523 (2021) +49. Zhang, S., Wang, X., Wang, J., Pang, J., Lyu, C., Zhang, W., Luo, P., Chen, K.: Dense distinct query for end-to-end object detection. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 7329-7338 (2023) +50. Zhao, C., Sun, Y., Wang, W., Chen, Q., Ding, E., Yang, Y., Wang, J.: Ms-detr: Efficient detr training with mixed supervision. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 17027-17036 (2024) +51. Zhu, X., Su, W., Lu, L., Li, B., Wang, X., Dai, J.: Deformable DETR: Deformable transformers for end-to-end object detection. In: Int. Conf. Learn. Represent. (2020) +52. Zong, Z., Song, G., Liu, Y.: DETRs with collaborative hybrid assignments training. In: Int. Conf. Comput. Vis. pp. 6748-6758 (2023) \ No newline at end of file diff --git a/2024/Relation DETR_ Exploring Explicit Position Relation Prior for Object Detection/images.zip b/2024/Relation DETR_ Exploring Explicit Position Relation Prior for Object Detection/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..62a4c39ec1e5fe94e66eea8ff4d0bb0e65277d3b --- /dev/null +++ b/2024/Relation DETR_ Exploring Explicit Position Relation Prior for Object Detection/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:631622dbaa5c0f7cac142528bea6f15917e1f0167a5c3c6ae8ee012557decc7d +size 799075 diff --git a/2024/Relation DETR_ Exploring Explicit Position Relation Prior for Object Detection/layout.json b/2024/Relation DETR_ Exploring Explicit Position Relation Prior for Object Detection/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..c096a61a4565a2fb7d7a33bf07873da7040e270c --- /dev/null +++ b/2024/Relation DETR_ Exploring Explicit Position Relation Prior for Object Detection/layout.json @@ -0,0 +1,9105 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 153, + 111, + 460, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 153, + 111, + 460, + 148 + ], + "spans": [ + { + "bbox": [ + 153, + 111, + 460, + 148 + ], + "type": "text", + "content": "Relation DETR: Exploring Explicit Position Relation Prior for Object Detection" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 171, + 466, + 196 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 171, + 466, + 196 + ], + "spans": [ + { + "bbox": [ + 146, + 171, + 466, + 196 + ], + "type": "text", + "content": "Xiuquan Hou" + }, + { + "bbox": [ + 146, + 171, + 466, + 196 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 146, + 171, + 466, + 196 + ], + "type": "text", + "content": ", Meiqin Liu" + }, + { + "bbox": [ + 146, + 171, + 466, + 196 + ], + "type": "inline_equation", + "content": "^{1,2,\\star}" + }, + { + "bbox": [ + 146, + 171, + 466, + 196 + ], + "type": "text", + "content": ", Senlin Zhang" + }, + { + "bbox": [ + 146, + 171, + 466, + 196 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 146, + 171, + 466, + 196 + ], + "type": "text", + "content": ", Ping Wei" + }, + { + "bbox": [ + 146, + 171, + 466, + 196 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 146, + 171, + 466, + 196 + ], + "type": "text", + "content": ", Badong Chen" + }, + { + "bbox": [ + 146, + 171, + 466, + 196 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 146, + 171, + 466, + 196 + ], + "type": "text", + "content": ", and Xuguang Lan" + }, + { + "bbox": [ + 146, + 171, + 466, + 196 + ], + "type": "inline_equation", + "content": "^{1}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 132, + 204, + 481, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 204, + 481, + 228 + ], + "spans": [ + { + "bbox": [ + 132, + 204, + 481, + 228 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 132, + 204, + 481, + 228 + ], + "type": "text", + "content": " National Key Laboratory of Human-Machine Hybrid Augmented Intelligence, Xi'an Jiaotong University, Xi'an 710049, China" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 140, + 227, + 473, + 261 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 140, + 227, + 473, + 261 + ], + "spans": [ + { + "bbox": [ + 140, + 227, + 473, + 261 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 140, + 227, + 473, + 261 + ], + "type": "text", + "content": " College of Electrical Engineering, Zhejiang University, Hangzhou 310027, China \nxiuqhou@stu.xjtu.edu.cn, {liumeiqin, slzhang}@zju.edu.cn, {pingwei, chenbd, xglan} @mail.xjtu.edu.cn" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 160, + 315, + 452, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 315, + 452, + 601 + ], + "spans": [ + { + "bbox": [ + 160, + 315, + 452, + 601 + ], + "type": "text", + "content": "Abstract. This paper presents a general scheme for enhancing the convergence and performance of DETR (DEtection Transformer). We investigate the slow convergence problem in transformers from a new perspective, suggesting that it arises from the self-attention that introduces no structural bias over inputs. To address this issue, we explore incorporating position relation prior as attention bias to augment object detection, following the verification of its statistical significance using a proposed quantitative macroscopic correlation (MC) metric. Our approach, termed Relation-DETR, introduces an encoder to construct position relation embeddings for progressive attention refinement, which further extends the traditional streaming pipeline of DETR into a contrastive relation pipeline to address the conflicts between non-duplicate predictions and positive supervision. Extensive experiments on both generic and task-specific datasets demonstrate the effectiveness of our approach. Under the same configurations, Relation-DETR achieves a significant improvement " + }, + { + "bbox": [ + 160, + 315, + 452, + 601 + ], + "type": "inline_equation", + "content": "(+2.0\\%)" + }, + { + "bbox": [ + 160, + 315, + 452, + 601 + ], + "type": "text", + "content": " AP compared to DINO), state-of-the-art performance " + }, + { + "bbox": [ + 160, + 315, + 452, + 601 + ], + "type": "inline_equation", + "content": "(51.7\\%)" + }, + { + "bbox": [ + 160, + 315, + 452, + 601 + ], + "type": "text", + "content": " AP for " + }, + { + "bbox": [ + 160, + 315, + 452, + 601 + ], + "type": "inline_equation", + "content": "1\\times" + }, + { + "bbox": [ + 160, + 315, + 452, + 601 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 160, + 315, + 452, + 601 + ], + "type": "inline_equation", + "content": "52.1\\%" + }, + { + "bbox": [ + 160, + 315, + 452, + 601 + ], + "type": "text", + "content": " AP for " + }, + { + "bbox": [ + 160, + 315, + 452, + 601 + ], + "type": "inline_equation", + "content": "2\\times" + }, + { + "bbox": [ + 160, + 315, + 452, + 601 + ], + "type": "text", + "content": " settings), and a remarkably faster convergence speed (over " + }, + { + "bbox": [ + 160, + 315, + 452, + 601 + ], + "type": "inline_equation", + "content": "40\\%" + }, + { + "bbox": [ + 160, + 315, + 452, + 601 + ], + "type": "text", + "content": " AP with only 2 training epochs) than existing DETR detectors on COCO val2017. Moreover, the proposed relation encoder serves as a universal plug-in-and-play component, bringing clear improvements for theoretically any DETR-like methods. Furthermore, we introduce a class-agnostic detection dataset, SA-Det-100k. The experimental results on the dataset illustrate that the proposed explicit position relation achieves a clear improvement of " + }, + { + "bbox": [ + 160, + 315, + 452, + 601 + ], + "type": "inline_equation", + "content": "1.3\\%" + }, + { + "bbox": [ + 160, + 315, + 452, + 601 + ], + "type": "text", + "content": " AP, highlighting its potential towards universal object detection. The code and dataset are available at https://github.com/xiuqhou/Relation-DETR." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 160, + 613, + 453, + 636 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 613, + 453, + 636 + ], + "spans": [ + { + "bbox": [ + 160, + 613, + 453, + 636 + ], + "type": "text", + "content": "Keywords: Detection transformer " + }, + { + "bbox": [ + 160, + 613, + 453, + 636 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 613, + 453, + 636 + ], + "type": "text", + "content": " Object detection " + }, + { + "bbox": [ + 160, + 613, + 453, + 636 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 613, + 453, + 636 + ], + "type": "text", + "content": " Relation network " + }, + { + "bbox": [ + 160, + 613, + 453, + 636 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 613, + 453, + 636 + ], + "type": "text", + "content": " Progressive attention refinement " + }, + { + "bbox": [ + 160, + 613, + 453, + 636 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 613, + 453, + 636 + ], + "type": "text", + "content": " Feature enhancement" + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 135, + 653, + 233, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 653, + 233, + 665 + ], + "spans": [ + { + "bbox": [ + 135, + 653, + 233, + 665 + ], + "type": "text", + "content": "* Corresponding author" + } + ] + } + ], + "index": 6 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 229, + 127 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 229, + 127 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 229, + 127 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 139, + 479, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 139, + 479, + 186 + ], + "spans": [ + { + "bbox": [ + 130, + 139, + 479, + 186 + ], + "type": "text", + "content": "Object detection aims to tackle the problems of bounding box regression and object classification for each object of interest. Recently, DETection TRansformer (DETR) [4] has overcome the reliance on handcrafted designs of convolution detectors, achieving an elegant architecture in an end-to-end manner." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 186, + 482, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 186, + 482, + 413 + ], + "spans": [ + { + "bbox": [ + 130, + 186, + 482, + 413 + ], + "type": "text", + "content": "Despite exhibiting impressive detection performance on large-scale datasets such as COCO [35], their performances are prone to be influenced by dataset scale and suffer from slow convergence. The root cause of the problem is the conflict between non-duplicate predictions and positive supervision [23]. During the training process, DETR employs the Hungarian algorithm to assign a single positive prediction to each ground-truth for producing unique results. However, this leads to negative predictions dominating the majority of the loss function, causing insufficient positive supervision. Therefore, more samples and iterations are required for convergence. Previous attempts have explored the issue by introducing train-only architectures (e.g. query denoising [28], multiple groups of queries [6], auxiliary queries [23], collaborative hybrid assignment training [52]) for additional supervision or by incorporating hard mining in loss functions (e.g. IA-BCE loss [3], position-supervised loss [37]). Other works have proposed specific structures for better interaction between queries and feature maps (e.g. dynamic anchor query [36], cascade window attention [46]), as well as techniques to focus on high-quality queries (e.g. hierarchical filtering [20], dense distinct process [49] and query rank layer [40]). Despite these advancements, there has been little exploration of the issue from the perspective of self-attention, which is widely used in the transformer decoders in most DETR detectors." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 414, + 484, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 414, + 484, + 521 + ], + "spans": [ + { + "bbox": [ + 130, + 414, + 484, + 521 + ], + "type": "text", + "content": "The effectiveness of self-attention lies in its establishment of a high-dimensional relation representation among sequence embeddings [4, 42], which also serves as a key component for modeling relations among different detection feature representations [4]. However, such relation is an implicit representation since it assumes no structural bias over inputs, in which even position information is also needed to be learned from training data [33]. Consequently, the learning process of transformer is data-intensive and slow to converge. This analysis motivates us to introduce task-specific bias for realizing faster convergence and reducing data dependence." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 522, + 482, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 522, + 482, + 640 + ], + "spans": [ + { + "bbox": [ + 130, + 522, + 482, + 640 + ], + "type": "text", + "content": "In this paper, we explore enhancing DETR detectors from a novel perspective, namely explicit position relation prior. We first establish a metric for quantifying position relations in images, and analyze the distribution to verify its statistical significance. In this context, we introduce a position relation encoder to model all pairwise interactions between two bounding boxes, employing progressive attention refinement for cross-layer information interaction. To maintain the end-to-end property while providing sufficient positive supervision, we introduce a contrast relation strategy, which leverages both one-to-one and one-to-many matching while emphasizing the influence of position relation on dedduplication. The proposed method is named as Relation-DETR." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "content": "Compared to previous works, the main feature of Relation-DETR is the integration of explicit position relation. In contrast, prior works focus on implicitly" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "type": "text", + "content": "X. Hou et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "text", + "content": "learned attention weights from training data, leading to slow convergence. Intuitively, our proposed position relation can be seen as a plug-in-and-play design beneficial for non-duplication predictions, since it establishes a representation of relative positions among pairs of bounding boxes (similar to IoU in NMS [16])." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 164, + 480, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 164, + 480, + 319 + ], + "spans": [ + { + "bbox": [ + 130, + 164, + 480, + 319 + ], + "type": "text", + "content": "We evaluate the performance of Relation-DETR on the most popular object detection dataset, COCO 2017 [35], as well as several task-specific datasets [8,43]. The experimental results demonstrate its superior performance, surpassing previous state-of-the-art DETR detectors with clear margins. More specifically, Relation-DETR exhibits a significantly fast convergence speed. Without bells and whistles, it becomes the first DETR detector to achieve " + }, + { + "bbox": [ + 130, + 164, + 480, + 319 + ], + "type": "inline_equation", + "content": "40\\%" + }, + { + "bbox": [ + 130, + 164, + 480, + 319 + ], + "type": "text", + "content": " AP on COCO with only 2 epochs using ResNet50 as the backbone under the " + }, + { + "bbox": [ + 130, + 164, + 480, + 319 + ], + "type": "inline_equation", + "content": "1\\times" + }, + { + "bbox": [ + 130, + 164, + 480, + 319 + ], + "type": "text", + "content": " training configuration. In addition, the simple architecture design of our position relation encoder maintains a promising transferability. It can be easily extended to other DETR-based methods with only a few modifications to achieve consistent performance improvements. This is in contrast to some existing DETR detectors whose performance is highly dependent from complex matching strategies [22] or detection heads [52] developed by convolution-based detectors." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 337, + 237, + 350 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 337, + 237, + 350 + ], + "spans": [ + { + "bbox": [ + 132, + 337, + 237, + 350 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 361, + 482, + 589 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 361, + 482, + 589 + ], + "spans": [ + { + "bbox": [ + 130, + 361, + 482, + 589 + ], + "type": "text", + "content": "Transformer for Object Detection In practice, the majority of attempts to apply transformer to object detection involve constructing a parallelizable sequence, either in the feature extractor [30] or in the detection body [4]. Specifically, transformer-based feature extractors generate token sequences based on image patches [12, 30, 39], and extract multi-scale features through aggregating local features [30, 39] or pyramid postprocess [14, 30]. DEtction TRansformer (DETR) proposed by Carion et al. [4] encodes the extracted features into object queries and decodes them into detected bounding boxes and labels. However, the self-learned attention mechanism increases the requirements for large-scale datasets and training iterations. Many works have explored the slow convergence from the perspective of structured attention (e.g. multi-scale deformable attention [51], dynamic attention [10], cascade window attention [46]), queries with explicit priors (e.g. anchor queries [44], dynamic anchor boxes queries [36], denoising queries [28], dense distinct queries [49]), and additional positive supervision (e.g. group queries [6], hybrid design [23], mixed matching [3]). However, even state-of-the-art DETR methods still utilize vanilla multi-head attention in the transformer decoder. And few works have explored the slow convergence from the perspective of implicit priors. This paper aims to address the issue with position relation." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 605, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 481, + 666 + ], + "type": "text", + "content": "Relation Network Rather than processing visual features at pixel levels, patch levels or image levels, relation network captures relation features at instance levels. Existing research on relation networks involves category-based and instance-based approaches. The category-based approaches construct conceptual or statistical relations (e.g. co-occurrence probability [17, 27]) either from relation" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 382, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 382, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 382, + 91, + 447, + 100 + ], + "type": "text", + "content": "Relation DETR" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 260 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 260 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 260 + ], + "type": "text", + "content": "datasets like Visual Genome [24, 27, 45] or by adaptively learning from class labels [17]. Both of them, however, increase the complexity due to the assignment between instances and categories [7, 17, 45]. In contrast, instance-based approaches directly construct a fine-grained graph structure given object features as a node set and their relations as a edge set. Therefore, reasoning on the graph during the training process naturally determines the explicit relation weight [21]. Typically, the weight denotes the parametric distances between each paired object instances in high-dimensional space, such as appearance similarity [31], proposal distance [32] or even self-attention weight [42]. Since learning self-attention weight solely from training data without structural bias increases the requirement for dataset scale and iterations, we explore explicit position relation as a prior to reduce the requirement." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 277, + 482, + 387 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 277, + 482, + 387 + ], + "spans": [ + { + "bbox": [ + 130, + 277, + 482, + 387 + ], + "type": "text", + "content": "Classification loss for hard mining. During object detection training, positive predictions assigned to ground truth are much fewer than negative predictions, often resulting in imbalanced supervision and slow convergence. For classification tasks, Focal Loss [34] proposes introducing a weight parameter to focus on hard samples, which is further extended to many variants like generalized focal loss (GFL) [29], vari focal loss (VFL) [29]. Moreover, for object detection tasks, using loss with modulation terms based on regression metrics (e.g. TOOD [15], IA-BCE [3], position-supervised loss [37]) further achieves high-quality alignment between classification and regression tasks." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 404, + 443, + 419 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 404, + 443, + 419 + ], + "spans": [ + { + "bbox": [ + 132, + 404, + 443, + 419 + ], + "type": "text", + "content": "3 Statistical significance of object position relation" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 430, + 482, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 430, + 482, + 525 + ], + "spans": [ + { + "bbox": [ + 130, + 430, + 482, + 525 + ], + "type": "text", + "content": "Are objects really correlated in object detection tasks? To answer the question, we propose a quantitative macroscopic correlation (MC) metric based on the Pearson Correlation Coefficient (PCC) to measure the position correlation among objects in a single image. Assume the objects in an image form a node set, and the PCC between each pair of bounding box annotations serves as their corresponding edge weight. We can construct an undirected graph with continuous values. In this end, the macroscopic correlation for each image can be calculated using the graph intensity, formulated as:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 228, + 534, + 481, + 562 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 228, + 534, + 481, + 562 + ], + "spans": [ + { + "bbox": [ + 228, + 534, + 481, + 562 + ], + "type": "interline_equation", + "content": "M C = \\frac {\\sum_ {i} \\sum_ {j : j \\neq i} | \\operatorname {P e a r s o n} \\left(\\boldsymbol {b} _ {i} , \\boldsymbol {b} _ {j}\\right) |}{N (N - 1)} \\tag {1}", + "image_path": "7abde49d61cbc2dd972fea0c00548f750f98ded4475160b0671414331ee4e149.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 569, + 482, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 569, + 482, + 617 + ], + "spans": [ + { + "bbox": [ + 130, + 569, + 482, + 617 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 569, + 482, + 617 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 569, + 482, + 617 + ], + "type": "text", + "content": " denotes the number of objects, i.e. the number of nodes, " + }, + { + "bbox": [ + 130, + 569, + 482, + 617 + ], + "type": "inline_equation", + "content": "\\pmb{b} = [x,y,w,h]" + }, + { + "bbox": [ + 130, + 569, + 482, + 617 + ], + "type": "text", + "content": " denotes the position annotation of bounding boxes in datasets. " + }, + { + "bbox": [ + 130, + 569, + 482, + 617 + ], + "type": "inline_equation", + "content": "MC = 1" + }, + { + "bbox": [ + 130, + 569, + 482, + 617 + ], + "type": "text", + "content": " only when all objects are fully linearly correlated, whereas " + }, + { + "bbox": [ + 130, + 569, + 482, + 617 + ], + "type": "inline_equation", + "content": "MC = 0" + }, + { + "bbox": [ + 130, + 569, + 482, + 617 + ], + "type": "text", + "content": " if there is no position correlation between any pair of objects." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "We visualize the statistical distribution of MC for datasets across various scenarios, including industrial (ESD [19], CSD [43], MSSD [8]), domestic (AI2thor [26]), urban (Cityscapes [9]) and generic settings (PascalVOC [13], COCO [35], Object365 [41], SA-1B [25]). The datasets cover a wide range of scales, from" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "type": "text", + "content": "X. Hou et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 201, + 116, + 413, + 280 + ], + "blocks": [ + { + "bbox": [ + 201, + 116, + 413, + 280 + ], + "lines": [ + { + "bbox": [ + 201, + 116, + 413, + 280 + ], + "spans": [ + { + "bbox": [ + 201, + 116, + 413, + 280 + ], + "type": "image", + "image_path": "db1a7ca603db4b177adec9136dc6a526425d17c13b7216f61756190c50f7fc8e.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 292, + 482, + 326 + ], + "lines": [ + { + "bbox": [ + 130, + 292, + 482, + 326 + ], + "spans": [ + { + "bbox": [ + 130, + 292, + 482, + 326 + ], + "type": "text", + "content": "Fig. 1: Statistical distribution of macroscopic correlation (MC) on various datasets (normalized for better visualization), and the values in brackets indicate the number of dataset samples." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 348, + 482, + 431 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 348, + 482, + 431 + ], + "spans": [ + { + "bbox": [ + 130, + 348, + 482, + 431 + ], + "type": "text", + "content": "0.3K to 11M images. As shown in Fig. 1, all of these datasets indicate that the distribution of MC is concentrated around high numerical values, with the distribution centers closer to the upper bound. This demonstrates the presence and the statistical significance of object position relation. Specifically, task-specific datasets display more prior knowledge and clearer clustering patterns in the high-dimensional feature space, which thus results in higher MC values than generic datasets like COCO." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 449, + 249, + 462 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 449, + 249, + 462 + ], + "spans": [ + { + "bbox": [ + 132, + 449, + 249, + 462 + ], + "type": "text", + "content": "4 Relation-DETR" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 474, + 482, + 570 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 474, + 482, + 570 + ], + "spans": [ + { + "bbox": [ + 130, + 474, + 482, + 570 + ], + "type": "text", + "content": "Given the statistical significance of position relation, we propose a state-of-the-art detector, named Relation-DETR, that explores explicit position relation prior to enhance object detection. To address the issue of slow convergence, we present a position relation encoder (Sec. 4.1) for progressive attention refinement (Sec. 4.2). Further in Sec. 4.3, it extends the streaming pipeline of DETR into a contrast pipeline to emphasize the influence of position relation on removing duplication while maintaining sufficient positive supervision for faster convergence." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 586, + 288, + 597 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 586, + 288, + 597 + ], + "spans": [ + { + "bbox": [ + 132, + 586, + 288, + 597 + ], + "type": "text", + "content": "4.1 Position relation encoder" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": "Previous research has demonstrated the relation effectiveness for convolution detectors [21,32]. Recently, some DETR methods attempt to construct instance-level relation by indexing from category-level relation using class indices [17]. In contrast to these approaches, we directly construct instance-level relation through a simple position encoder, maintaining an end-to-end design for DETR." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 382, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 382, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 382, + 91, + 447, + 100 + ], + "type": "text", + "content": "Relation DETR" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 479, + 99 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 479, + 99 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 479, + 99 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 212 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 212 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 212 + ], + "type": "text", + "content": "We first review the basic pipeline in DETR detectors. Given image features extracted by the backbone, the transformer encoder produces an augmented memory " + }, + { + "bbox": [ + 130, + 116, + 482, + 212 + ], + "type": "inline_equation", + "content": "\\mathbf{Z} \\in \\mathbb{R}^{d \\times H \\times W}" + }, + { + "bbox": [ + 130, + 116, + 482, + 212 + ], + "type": "text", + "content": " for further decoding into bounding boxes " + }, + { + "bbox": [ + 130, + 116, + 482, + 212 + ], + "type": "inline_equation", + "content": "\\pmb{b}_i = [x, y, w, h], i = 1, \\dots, N" + }, + { + "bbox": [ + 130, + 116, + 482, + 212 + ], + "type": "text", + "content": " and class labels " + }, + { + "bbox": [ + 130, + 116, + 482, + 212 + ], + "type": "inline_equation", + "content": "\\pmb{c}_i" + }, + { + "bbox": [ + 130, + 116, + 482, + 212 + ], + "type": "text", + "content": " as predictions. Each decoder layer refines the bounding box coordinates iteratively by predicting " + }, + { + "bbox": [ + 130, + 116, + 482, + 212 + ], + "type": "inline_equation", + "content": "\\Delta" + }, + { + "bbox": [ + 130, + 116, + 482, + 212 + ], + "type": "text", + "content": " w.r.t. the coordinate from the last decoder layer, known as iterative bounding box refinement [51]. In addition, predictions from all decoder layers equally participate in loss calculation to compute auxiliary decoding losses [4]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 212, + 482, + 284 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 212, + 482, + 284 + ], + "spans": [ + { + "bbox": [ + 130, + 212, + 482, + 284 + ], + "type": "text", + "content": "Under the aforementioned detection framework, our position relation encoder represents the high-dimensional relation embedding as an explicit prior for self-attention in the transformer. This embedding is calculated based on the predicted bounding boxes (denoted as " + }, + { + "bbox": [ + 130, + 212, + 482, + 284 + ], + "type": "inline_equation", + "content": "\\mathbf{b} = [x,y,w,h]" + }, + { + "bbox": [ + 130, + 212, + 482, + 284 + ], + "type": "text", + "content": ") from each decoder layer. To ensure that the relation is invariant to translation and scale transformations, we encode it based on normalized relative geometry features:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 140, + 287, + 481, + 315 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 140, + 287, + 481, + 315 + ], + "spans": [ + { + "bbox": [ + 140, + 287, + 481, + 315 + ], + "type": "interline_equation", + "content": "\\boldsymbol {e} (\\boldsymbol {b} _ {i}, \\boldsymbol {b} _ {j}) = \\left[ \\log \\left(\\frac {| x _ {i} - x _ {j} |}{w _ {i}} + 1\\right), \\log \\left(\\frac {| y _ {i} - y _ {j} |}{h _ {i}} + 1\\right), \\log \\left(\\frac {w _ {i}}{w _ {j}}\\right), \\log \\left(\\frac {h _ {i}}{h _ {j}}\\right) \\right] (2)", + "image_path": "3057ee676aa800b6908c55aa5cb6e88121fc950c2ff1d90f6b65f2a06dcc6124.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 318, + 482, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 318, + 482, + 354 + ], + "spans": [ + { + "bbox": [ + 131, + 318, + 482, + 354 + ], + "type": "text", + "content": "Unlike [21], our position relation is unbiased, as " + }, + { + "bbox": [ + 131, + 318, + 482, + 354 + ], + "type": "inline_equation", + "content": "\\pmb{e}(\\pmb{b}_i, \\pmb{b}_j) = 0" + }, + { + "bbox": [ + 131, + 318, + 482, + 354 + ], + "type": "text", + "content": " when " + }, + { + "bbox": [ + 131, + 318, + 482, + 354 + ], + "type": "inline_equation", + "content": "i = j" + }, + { + "bbox": [ + 131, + 318, + 482, + 354 + ], + "type": "text", + "content": ". The relation matrix " + }, + { + "bbox": [ + 131, + 318, + 482, + 354 + ], + "type": "inline_equation", + "content": "\\mathbf{E} \\in \\mathbb{R}^{N \\times N \\times 4}" + }, + { + "bbox": [ + 131, + 318, + 482, + 354 + ], + "type": "text", + "content": " (with " + }, + { + "bbox": [ + 131, + 318, + 482, + 354 + ], + "type": "inline_equation", + "content": "\\mathbf{E}(i,j) = \\pmb{e}(\\pmb{b}_i, \\pmb{b}_j)" + }, + { + "bbox": [ + 131, + 318, + 482, + 354 + ], + "type": "text", + "content": ") is further transformed into high-dimensional embeddings through sine-cosine encoding [42]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 242, + 359, + 480, + 378 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 242, + 359, + 480, + 378 + ], + "spans": [ + { + "bbox": [ + 242, + 359, + 480, + 378 + ], + "type": "interline_equation", + "content": "\\operatorname {E m b e d} (\\mathbf {E}, 2 k) = \\sin \\left(s \\mathbf {E} / T ^ {2 k / d _ {\\mathrm {r e}}}\\right) \\tag {3}", + "image_path": "88ef2595ba8a670fb09fd622a32d8fe8a4b4dfc69d2653dee38a2df41a846ad2.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 224, + 382, + 480, + 401 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 224, + 382, + 480, + 401 + ], + "spans": [ + { + "bbox": [ + 224, + 382, + 480, + 401 + ], + "type": "interline_equation", + "content": "\\operatorname {E m b e d} (\\mathbf {E}, 2 k + 1) = \\cos (s \\mathbf {E} / T ^ {2 k / d _ {\\mathrm {r e}}}) \\tag {4}", + "image_path": "4ae474d9c80cae7d83e82514ce14ee326e630e43b69ff917bd915f9a7517a972.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 404, + 482, + 440 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 404, + 482, + 440 + ], + "spans": [ + { + "bbox": [ + 131, + 404, + 482, + 440 + ], + "type": "text", + "content": "where the shape of relation embedding is " + }, + { + "bbox": [ + 131, + 404, + 482, + 440 + ], + "type": "inline_equation", + "content": "N \\times N \\times 4d_{\\mathrm{re}}" + }, + { + "bbox": [ + 131, + 404, + 482, + 440 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 131, + 404, + 482, + 440 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 131, + 404, + 482, + 440 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 131, + 404, + 482, + 440 + ], + "type": "inline_equation", + "content": "d_{\\mathrm{re}}" + }, + { + "bbox": [ + 131, + 404, + 482, + 440 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 131, + 404, + 482, + 440 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 131, + 404, + 482, + 440 + ], + "type": "text", + "content": " are encoding parameters. Finally, the embedding undergoes a linear transformation to obtain " + }, + { + "bbox": [ + 131, + 404, + 482, + 440 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 131, + 404, + 482, + 440 + ], + "type": "text", + "content": " scalar weights, where " + }, + { + "bbox": [ + 131, + 404, + 482, + 440 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 131, + 404, + 482, + 440 + ], + "type": "text", + "content": " denotes the number of attention heads." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 219, + 445, + 480, + 458 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 219, + 445, + 480, + 458 + ], + "spans": [ + { + "bbox": [ + 219, + 445, + 480, + 458 + ], + "type": "interline_equation", + "content": "\\operatorname {R e l} (\\boldsymbol {b}, \\boldsymbol {b}) = \\max \\left(\\epsilon , \\mathbf {W E m b e d} (\\boldsymbol {b}, \\boldsymbol {b}) + \\mathbf {B}\\right) \\tag {5}", + "image_path": "df267c425b4e211b87f8caeadae039464f3a8f6abca32a5e4e20a4a699cdb844.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 463, + 482, + 487 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 463, + 482, + 487 + ], + "spans": [ + { + "bbox": [ + 131, + 463, + 482, + 487 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 463, + 482, + 487 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 131, + 463, + 482, + 487 + ], + "type": "text", + "content": " makes sure a positive value for relation to avoid gradient vanishing after " + }, + { + "bbox": [ + 131, + 463, + 482, + 487 + ], + "type": "inline_equation", + "content": "\\exp" + }, + { + "bbox": [ + 131, + 463, + 482, + 487 + ], + "type": "text", + "content": " when integrated into self-attention, and " + }, + { + "bbox": [ + 131, + 463, + 482, + 487 + ], + "type": "inline_equation", + "content": "\\mathsf{Rel}(\\pmb {b},\\pmb {b})\\in \\mathbb{R}^{N\\times N\\times M}" + }, + { + "bbox": [ + 131, + 463, + 482, + 487 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 131, + 501, + 437, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 501, + 437, + 514 + ], + "spans": [ + { + "bbox": [ + 131, + 501, + 437, + 514 + ], + "type": "text", + "content": "4.2 Progressive attention refinement with position relation" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 517, + 482, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 517, + 482, + 590 + ], + "spans": [ + { + "bbox": [ + 130, + 517, + 482, + 590 + ], + "type": "text", + "content": "The iterative box refinement proposed by Deformable-DETR [51] has shown the effectiveness for high-quality bounding box regression. Following the motivation, we propose a progressive attention refinement method to introduce the position relation into the streaming pipeline of DETR. Specifically, the relation of layer-" + }, + { + "bbox": [ + 130, + 517, + 482, + 590 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 517, + 482, + 590 + ], + "type": "text", + "content": " is determined by bounding boxes of both layer-" + }, + { + "bbox": [ + 130, + 517, + 482, + 590 + ], + "type": "inline_equation", + "content": "i - 1" + }, + { + "bbox": [ + 130, + 517, + 482, + 590 + ], + "type": "text", + "content": " and layer-" + }, + { + "bbox": [ + 130, + 517, + 482, + 590 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 517, + 482, + 590 + ], + "type": "text", + "content": ", which is further integrated into self-attention for producing the bounding boxes in layer-" + }, + { + "bbox": [ + 130, + 517, + 482, + 590 + ], + "type": "inline_equation", + "content": "(i + 1)" + }, + { + "bbox": [ + 130, + 517, + 482, + 590 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 140, + 603, + 480, + 630 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 140, + 603, + 480, + 630 + ], + "spans": [ + { + "bbox": [ + 140, + 603, + 480, + 630 + ], + "type": "interline_equation", + "content": "\\operatorname {A t t n} _ {\\text {S e l f}} \\left(\\mathbf {Q} ^ {l}\\right) = \\operatorname {S o f t m a x} \\left(\\operatorname {R e l} \\left(\\boldsymbol {b} ^ {l - 1}, \\boldsymbol {b} ^ {l}\\right) + \\frac {\\operatorname {Q u e} \\left(\\mathbf {Q} ^ {l}\\right) \\operatorname {K e y} \\left(\\mathbf {Q} ^ {l}\\right) ^ {\\top}}{\\sqrt {d _ {\\text {m o d e l}}}}\\right) \\operatorname {V a l} \\left(\\mathbf {Q} ^ {l}\\right) \\tag {6}", + "image_path": "cb282a31c202c5ccc67d2755570c2666268a661ffba40913117cac4b347ba1cd.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 179, + 632, + 480, + 646 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 179, + 632, + 480, + 646 + ], + "spans": [ + { + "bbox": [ + 179, + 632, + 480, + 646 + ], + "type": "interline_equation", + "content": "\\mathbf {Q} ^ {l + 1} = \\text {F F N} \\left(\\mathbf {Q} ^ {l} + \\operatorname {A t t n} _ {\\text {c r o s s}} \\left(\\operatorname {A t t n} _ {\\text {s e l f}} (\\mathbf {Q} ^ {l}), \\operatorname {K e y} (\\mathbf {Z}), \\operatorname {V a l} (\\mathbf {Z})\\right)\\right) \\tag {7}", + "image_path": "b6a8e06c41b2f0a0896defb096d6a0176eff5e5974126d2daf8f0047489865b2.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 181, + 648, + 480, + 663 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 181, + 648, + 480, + 663 + ], + "spans": [ + { + "bbox": [ + 181, + 648, + 480, + 663 + ], + "type": "interline_equation", + "content": "\\boldsymbol {b} ^ {l + 1} = \\operatorname {M L P} \\left(\\mathbf {Q} ^ {l + 1}\\right), \\boldsymbol {c} ^ {l + 1} = \\operatorname {L i n e a r} \\left(\\mathbf {Q} ^ {l + 1}\\right) \\tag {8}", + "image_path": "94bc3713388dbf0612c8c41c1b88e66eeb7dd18644693358f322f8d21eb82c53.jpg" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 139, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 139, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 139, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "type": "text", + "content": "X. Hou et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 131, + 115, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 115, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 131, + 115, + 479, + 140 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 115, + 479, + 140 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}^l" + }, + { + "bbox": [ + 131, + 115, + 479, + 140 + ], + "type": "text", + "content": " denotes the queries in the " + }, + { + "bbox": [ + 131, + 115, + 479, + 140 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 131, + 115, + 479, + 140 + ], + "type": "text", + "content": "-th decoder layer in the DETR transformer, " + }, + { + "bbox": [ + 131, + 115, + 479, + 140 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}" + }, + { + "bbox": [ + 131, + 115, + 479, + 140 + ], + "type": "text", + "content": " is the memory, i.e., the enhanced image features from the transformer encoder." + } + ] + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 138, + 162, + 305, + 243 + ], + "blocks": [ + { + "bbox": [ + 138, + 162, + 305, + 243 + ], + "lines": [ + { + "bbox": [ + 138, + 162, + 305, + 243 + ], + "spans": [ + { + "bbox": [ + 138, + 162, + 305, + 243 + ], + "type": "image", + "image_path": "b279f4e719153e33370759b926e9c57cda4e1343abbfd03d603bdc9eb04ad97d.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 253, + 479, + 276 + ], + "lines": [ + { + "bbox": [ + 131, + 253, + 479, + 276 + ], + "spans": [ + { + "bbox": [ + 131, + 253, + 479, + 276 + ], + "type": "text", + "content": "Fig. 2: Comparison of transformer decoder in Deformable-DETR(left) and Relation-DETR(right)." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 309, + 160, + 477, + 242 + ], + "blocks": [ + { + "bbox": [ + 309, + 160, + 477, + 242 + ], + "lines": [ + { + "bbox": [ + 309, + 160, + 477, + 242 + ], + "spans": [ + { + "bbox": [ + 309, + 160, + 477, + 242 + ], + "type": "image", + "image_path": "a1b9f1abfc871fc81b2ed4376a5577eea72c11ba94458b93f22e7b8ce7668231.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 300, + 480, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 300, + 480, + 373 + ], + "spans": [ + { + "bbox": [ + 130, + 300, + 480, + 373 + ], + "type": "text", + "content": "The only difference between our method and the existing DETR decoder is marked in red. As depicted in Fig. 2, the requisite addition involves the introduction of a lateral branch for the computation of position relation. Therefore, our position relation and the progressive attention refinement are straightforward, allowing for plug-in-and-play integration with self attention in existing DETR detectors to achieve consistent performance improvements (see Tab. 6)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 393, + 290, + 406 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 393, + 290, + 406 + ], + "spans": [ + { + "bbox": [ + 131, + 393, + 290, + 406 + ], + "type": "text", + "content": "4.3 Contrast relation pipeline" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 416, + 480, + 488 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 416, + 480, + 488 + ], + "spans": [ + { + "bbox": [ + 130, + 416, + 480, + 488 + ], + "type": "text", + "content": "Rethinking the mechanism of existing duplication removal methods (including NMS [16], Soft-NMS [1], fast-NMS [2], Adaptive-NMS [38]), these processes heavily rely on IoU (intersection over Union), which, to some extent, signifies the position relation between bounding boxes. Therefore, we may hypothesize that integrating the position relation among queries in self-attention contributes to non-duplicated predictions in object detection, akin to [22]." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 489, + 480, + 585 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 489, + 480, + 585 + ], + "spans": [ + { + "bbox": [ + 130, + 489, + 480, + 585 + ], + "type": "text", + "content": "The conflicts between non-duplicate predictions and sufficient positive supervision arise from the streaming pipeline of DETR, which must navigate between one-to-one matching and one-to-many matching. To overcome this limitation, we extend it to a contrast pipeline based on the proposed position relation. Specifically, we construct two parallel sets of queries, i.e. matching queries " + }, + { + "bbox": [ + 130, + 489, + 480, + 585 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}_m" + }, + { + "bbox": [ + 130, + 489, + 480, + 585 + ], + "type": "text", + "content": " and hybrid queries " + }, + { + "bbox": [ + 130, + 489, + 480, + 585 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}_h" + }, + { + "bbox": [ + 130, + 489, + 480, + 585 + ], + "type": "text", + "content": ". Both are input into the transformer decoder but undergo distinct processing. The matching queries are processed with self-attention incorporating position relation to produce non-duplicated predictions:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 133, + 594, + 480, + 634 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 594, + 480, + 634 + ], + "spans": [ + { + "bbox": [ + 133, + 594, + 480, + 634 + ], + "type": "interline_equation", + "content": "\\operatorname {A t t n} _ {\\text {S e l f}} \\left(\\mathbf {Q} _ {\\mathrm {m}} ^ {l}\\right) = \\operatorname {S o f t m a x} \\left(\\operatorname {R e l} \\left(\\boldsymbol {b} ^ {l - 1}, \\boldsymbol {b} ^ {l}\\right) + \\frac {\\operatorname {Q u e} \\left(\\mathbf {Q} _ {\\mathrm {m}}\\right) \\operatorname {K e y} \\left(\\mathbf {Q} _ {\\mathrm {m}}\\right) ^ {\\top}}{\\sqrt {d _ {\\text {m o d e l}}}}\\right) \\operatorname {V a l} \\left(\\mathbf {Q} _ {\\mathrm {m}}\\right) \\tag {9}", + "image_path": "182f4be32f1099d1a2914b1b06b33823109b928bdda1a294031abb3ec62bc57f.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 136, + 635, + 480, + 664 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 635, + 480, + 664 + ], + "spans": [ + { + "bbox": [ + 136, + 635, + 480, + 664 + ], + "type": "interline_equation", + "content": "\\operatorname {A t t n} _ {\\text {S e l f}} \\left(\\mathbf {Q} _ {\\mathrm {h}} ^ {l}\\right) = \\operatorname {S o f t m a x} \\left(\\frac {\\operatorname {Q u e} \\left(\\mathbf {Q} _ {\\mathrm {h}}\\right) \\operatorname {K e y} \\left(\\mathbf {Q} _ {\\mathrm {h}}\\right) ^ {\\top}}{\\sqrt {d _ {\\text {m o d e l}}}}\\right) \\operatorname {V a l} \\left(\\mathbf {Q} _ {\\mathrm {h}}\\right) \\tag {10}", + "image_path": "3ad96d63aa8336c31955869b67a03b284a12a3f606db53e4ede9b3c5224cfdf4.jpg" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 381, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 381, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 381, + 91, + 447, + 100 + ], + "type": "text", + "content": "Relation DETR" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 165 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 165 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 165 + ], + "type": "text", + "content": "while the hybrid queries are decoded by the same decoder but skip the calculation of position relation to explore more potential candidates. Their corresponding predictions are denoted as " + }, + { + "bbox": [ + 130, + 116, + 482, + 165 + ], + "type": "inline_equation", + "content": "\\pmb{p}_m^l = (\\pmb{b}_m^l, \\pmb{c}_m^l)" + }, + { + "bbox": [ + 130, + 116, + 482, + 165 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 116, + 482, + 165 + ], + "type": "inline_equation", + "content": "\\pmb{p}_h^l = (\\pmb{b}_h^l, \\pmb{c}_h^l)" + }, + { + "bbox": [ + 130, + 116, + 482, + 165 + ], + "type": "text", + "content": ", respectively. Details of the contrast relation pipeline are illustrated in Fig. 3." + } + ] + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 192, + 186, + 350, + 322 + ], + "blocks": [ + { + "bbox": [ + 192, + 186, + 350, + 322 + ], + "lines": [ + { + "bbox": [ + 192, + 186, + 350, + 322 + ], + "spans": [ + { + "bbox": [ + 192, + 186, + 350, + 322 + ], + "type": "image", + "image_path": "734495d4255b2af0c1253dff6b4627e00d65b7adb28330f26d43e0df3546b13c.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 162, + 334, + 449, + 346 + ], + "lines": [ + { + "bbox": [ + 162, + 334, + 449, + 346 + ], + "spans": [ + { + "bbox": [ + 162, + 334, + 449, + 346 + ], + "type": "text", + "content": "Fig. 3: Detailed illustration of the proposed contrast relation pipeline." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 356, + 186, + 434, + 322 + ], + "blocks": [ + { + "bbox": [ + 356, + 186, + 434, + 322 + ], + "lines": [ + { + "bbox": [ + 356, + 186, + 434, + 322 + ], + "spans": [ + { + "bbox": [ + 356, + 186, + 434, + 322 + ], + "type": "image", + "image_path": "a28699bd84ffaf677331794c6cd14f2ea24e6f3e54c3b984eef3b7be26635ca4.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 369, + 482, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 369, + 482, + 407 + ], + "spans": [ + { + "bbox": [ + 130, + 369, + 482, + 407 + ], + "type": "text", + "content": "Assuming " + }, + { + "bbox": [ + 130, + 369, + 482, + 407 + ], + "type": "inline_equation", + "content": "\\pmb{g}" + }, + { + "bbox": [ + 130, + 369, + 482, + 407 + ], + "type": "text", + "content": " denotes the ground truth annotations, for " + }, + { + "bbox": [ + 130, + 369, + 482, + 407 + ], + "type": "inline_equation", + "content": "\\pmb{p}_m" + }, + { + "bbox": [ + 130, + 369, + 482, + 407 + ], + "type": "text", + "content": ", we employ a one-to-one matching scheme to emphasize the non-duplicate property, with loss calculation similar to the original DETR approach [4]:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 229, + 415, + 480, + 447 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 229, + 415, + 480, + 447 + ], + "spans": [ + { + "bbox": [ + 229, + 415, + 480, + 447 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {m} \\left(\\boldsymbol {p} _ {m}, \\boldsymbol {g}\\right) = \\sum_ {l = 1} ^ {L} \\mathcal {L} _ {\\text {H u n g a r i a n}} \\left(\\boldsymbol {p} _ {m} ^ {l}, \\boldsymbol {g}\\right) \\tag {11}", + "image_path": "96fd1b361956043485c56a496c74811ae4ee0916c4f1cbac057b3af2f5351b94.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 455, + 482, + 493 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 455, + 482, + 493 + ], + "spans": [ + { + "bbox": [ + 131, + 455, + 482, + 493 + ], + "type": "text", + "content": "While for " + }, + { + "bbox": [ + 131, + 455, + 482, + 493 + ], + "type": "inline_equation", + "content": "\\pmb{p}_h" + }, + { + "bbox": [ + 131, + 455, + 482, + 493 + ], + "type": "text", + "content": ", a one-to-many matching scheme is employed to form more potential positive candidates. We simply follow H-DETR [23] and repeat the ground truth " + }, + { + "bbox": [ + 131, + 455, + 482, + 493 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 131, + 455, + 482, + 493 + ], + "type": "text", + "content": " times, denoted as " + }, + { + "bbox": [ + 131, + 455, + 482, + 493 + ], + "type": "inline_equation", + "content": "\\tilde{\\pmb{g}} = \\{\\pmb{g}^1, \\pmb{g}^2, \\dots, \\pmb{g}^K\\}" + }, + { + "bbox": [ + 131, + 455, + 482, + 493 + ], + "type": "text", + "content": ", for loss calculation:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 233, + 501, + 480, + 534 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 233, + 501, + 480, + 534 + ], + "spans": [ + { + "bbox": [ + 233, + 501, + 480, + 534 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {h} \\left(\\boldsymbol {p} _ {h}, \\boldsymbol {g}\\right) = \\sum_ {l = 1} ^ {L} \\mathcal {L} _ {\\text {H u n g a r i a n}} \\left(\\boldsymbol {p} _ {h} ^ {l}, \\tilde {\\boldsymbol {g}}\\right) \\tag {12}", + "image_path": "b1e9bf8198c3a12c7744b49a23c00d1871246f52b5c5ea481c18659cab595f67.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 542, + 482, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 542, + 482, + 579 + ], + "spans": [ + { + "bbox": [ + 130, + 542, + 482, + 579 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 542, + 482, + 579 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{Hungarian}}" + }, + { + "bbox": [ + 130, + 542, + 482, + 579 + ], + "type": "text", + "content": " denotes the Hungarian loss, and " + }, + { + "bbox": [ + 130, + 542, + 482, + 579 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 130, + 542, + 482, + 579 + ], + "type": "text", + "content": " denotes the number of decoder layers. It is worth noting that the hybrid queries are only involved during training, thus incurring no extra computational burden for inference." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 131, + 596, + 373, + 611 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 596, + 373, + 611 + ], + "spans": [ + { + "bbox": [ + 131, + 596, + 373, + 611 + ], + "type": "text", + "content": "5 Experimental Results and Discussion" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 131, + 621, + 190, + 634 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 621, + 190, + 634 + ], + "spans": [ + { + "bbox": [ + 131, + 621, + 190, + 634 + ], + "type": "text", + "content": "5.1 Setup" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 641, + 481, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 667 + ], + "type": "text", + "content": "For a comprehensive evaluation, we conduct experiments on both an object detection benchmark, COCO 2017 [35], and two task-specific datasets, CSD [43]" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 139, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 139, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 139, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "type": "text", + "content": "X. Hou et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 115, + 482, + 273 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 115, + 482, + 273 + ], + "spans": [ + { + "bbox": [ + 130, + 115, + 482, + 273 + ], + "type": "text", + "content": "and MSSD [8]. Detection performance is measured using the standard Average Precision (AP) [35]. We train our model on NVIDIA A800 GPU (80GB) and RTX 3090 GPU (24GB) using the AdamW optimizer with an initial learning rate of " + }, + { + "bbox": [ + 130, + 115, + 482, + 273 + ], + "type": "inline_equation", + "content": "1 \\times 10^{-4}" + }, + { + "bbox": [ + 130, + 115, + 482, + 273 + ], + "type": "text", + "content": " and weight decay of " + }, + { + "bbox": [ + 130, + 115, + 482, + 273 + ], + "type": "inline_equation", + "content": "1 \\times 10^{-4}" + }, + { + "bbox": [ + 130, + 115, + 482, + 273 + ], + "type": "text", + "content": ". The Relation-DETR is implemented based on the backbone ResNet-50 [18] and Swin-L [39] pretrained from ImageNet [11], which are finetuned with a learning rate " + }, + { + "bbox": [ + 130, + 115, + 482, + 273 + ], + "type": "inline_equation", + "content": "1 \\times 10^{-5}" + }, + { + "bbox": [ + 130, + 115, + 482, + 273 + ], + "type": "text", + "content": " during training. The learning rate is reduced by a factor of 0.1 at later stages. The parameters of the position relation encoder, " + }, + { + "bbox": [ + 130, + 115, + 482, + 273 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 130, + 115, + 482, + 273 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 115, + 482, + 273 + ], + "type": "inline_equation", + "content": "d_{\\mathrm{re}}" + }, + { + "bbox": [ + 130, + 115, + 482, + 273 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 115, + 482, + 273 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 130, + 115, + 482, + 273 + ], + "type": "text", + "content": " are empirically chosen as 10000, 16, 100, respectively. The hybrid training configurations follow DINO [47] and H-DETR [23], i.e., " + }, + { + "bbox": [ + 130, + 115, + 482, + 273 + ], + "type": "inline_equation", + "content": "N_{m} = 900" + }, + { + "bbox": [ + 130, + 115, + 482, + 273 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 115, + 482, + 273 + ], + "type": "inline_equation", + "content": "N_{h} = 1500" + }, + { + "bbox": [ + 130, + 115, + 482, + 273 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 115, + 482, + 273 + ], + "type": "inline_equation", + "content": "k = 6" + }, + { + "bbox": [ + 130, + 115, + 482, + 273 + ], + "type": "text", + "content": ". We adopt the VariFocal loss [48] for training Relation-DETR. The training batch size is 10 for COCO 2017 and 2 for task-specific datasets. Before fed into detectors, images undergo the same augmentations (random resize, crop and flip) as other DETR detectors." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 299, + 374, + 312 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 299, + 374, + 312 + ], + "spans": [ + { + "bbox": [ + 131, + 299, + 374, + 312 + ], + "type": "text", + "content": "5.2 Comparison with state-of-the-art methods" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 327, + 482, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 327, + 482, + 483 + ], + "spans": [ + { + "bbox": [ + 130, + 327, + 482, + 483 + ], + "type": "text", + "content": "Comparison on COCO 2017. Tab. 1 presents the detection performance on COCO val 2017. Compared to other state-of-the-art DETR methods, our approach converges much faster and demonstrates significant improvements of " + }, + { + "bbox": [ + 130, + 327, + 482, + 483 + ], + "type": "inline_equation", + "content": "1.0\\%" + }, + { + "bbox": [ + 130, + 327, + 482, + 483 + ], + "type": "text", + "content": " AP, " + }, + { + "bbox": [ + 130, + 327, + 482, + 483 + ], + "type": "inline_equation", + "content": "1.0\\% \\mathrm{AP}_{50}" + }, + { + "bbox": [ + 130, + 327, + 482, + 483 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 327, + 482, + 483 + ], + "type": "inline_equation", + "content": "0.6\\% \\mathrm{AP}_{75}" + }, + { + "bbox": [ + 130, + 327, + 482, + 483 + ], + "type": "text", + "content": ", respectively, suppressing the second best DDQ-DETR [49] with clear margins. Specifically, Relation-DETR achieves " + }, + { + "bbox": [ + 130, + 327, + 482, + 483 + ], + "type": "inline_equation", + "content": "51.7\\%" + }, + { + "bbox": [ + 130, + 327, + 482, + 483 + ], + "type": "text", + "content": " AP using only 12 epochs with the ResNet-50 backbone and even outperforms DINO [47] (51.2% AP) with 36 epochs (" + }, + { + "bbox": [ + 130, + 327, + 482, + 483 + ], + "type": "inline_equation", + "content": "3\\times" + }, + { + "bbox": [ + 130, + 327, + 482, + 483 + ], + "type": "text", + "content": " faster). More importantly, in contrast to DDQ-DETR [49] and Co-DETR [52] that leverages NMS in the decoder or postprocess for improving precision, our Relation-DETR maintains an end-to-end pipeline, ensuring promising extensibility. When integrated with Swin-L backbone, Relation-DETR outperforms all counterparts, achieving the best " + }, + { + "bbox": [ + 130, + 327, + 482, + 483 + ], + "type": "inline_equation", + "content": "57.8\\%" + }, + { + "bbox": [ + 130, + 327, + 482, + 483 + ], + "type": "text", + "content": " AP with a " + }, + { + "bbox": [ + 130, + 327, + 482, + 483 + ], + "type": "inline_equation", + "content": "0.5\\%" + }, + { + "bbox": [ + 130, + 327, + 482, + 483 + ], + "type": "text", + "content": " AP improvement, showcasing its excellent scalability for larger model capacity." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "content": "Comparison on task-specific datasets. Different from generic object detection benchmarks, datasets in task-specific scenarios lack sufficient samples to provide semantic information. To reveal the generalizability of Relation-DETR, we conduct a performance comparison on two defect detection datasets, i.e. CSD [43] and MSSD [8]. The results in Tab. 3 show that, Relation-DETR improves the baseline DINO by " + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "inline_equation", + "content": "1.4\\%" + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "content": " AP on CSD, achieving the highest " + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "inline_equation", + "content": "54.4\\%" + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "content": " AP. Tab. 4 demonstrates that Relation-DETR further increases the margin to " + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "inline_equation", + "content": "6.4\\%" + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "content": " AP on MSSD and surpasses other counterparts. It is noteworthy that CSD [43] and MSSD [8] contain more small-sized objects than COCO 2017, which thus confirms the effectiveness of Relation-DETR to small-sized detection. Moreover, considering a stricter IoU threshold, Relation-DETR outperforms the second best method DINO by a significant margin of " + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "inline_equation", + "content": "11.1\\%" + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "content": " AP@75, highlighting the beneficial impact of explicit position relation on high-quality predictions." + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 381, + 90, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 381, + 90, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 381, + 90, + 447, + 101 + ], + "type": "text", + "content": "Relation DETR" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 101 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 101 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 101 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 132, + 158, + 479, + 439 + ], + "blocks": [ + { + "bbox": [ + 130, + 114, + 482, + 148 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 482, + 148 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 482, + 148 + ], + "type": "text", + "content": "Table 1: Comparison with state-of-the-art methods on COCO val2017 using ResNet-50(IN-1K) backbone. The * means that we re-implement the methods and report the corresponding results." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 132, + 158, + 479, + 439 + ], + "lines": [ + { + "bbox": [ + 132, + 158, + 479, + 439 + ], + "spans": [ + { + "bbox": [ + 132, + 158, + 479, + 439 + ], + "type": "table", + "html": "
MethodBackboneEpochsAP@50:95AP@50AP@75\\( AP_S \\)\\( AP_M \\)\\( AP_L \\)
Anchor-DETR [44]ResNet-505042.163.144.922.346.260.0
Def-DETR* [51]ResNet-501245.465.049.127.249.661.0
DAB-Def-DETR* [36]ResNet-501246.365.950.429.350.561.9
DN-Def-DETR* [28]ResNet-501247.265.351.330.250.662.4
DINO* [47]ResNet-501249.967.454.533.953.563.8
Group-DETR [6]ResNet-501249.832.453.064.2
H-Def-DETR [23]ResNet-501248.766.452.931.251.563.5
Cascade-DETR [46]ResNet-501249.767.154.132.453.565.1
Co-Def-DETR [52]ResNet-501249.567.654.332.452.763.7
Align-DETR [3]ResNet-501250.267.854.432.953.365.0
Stable-DINO [37]ResNet-501250.467.455.032.954.065.5
DAC-DETR [22]ResNet-501250.067.654.7
Salience-DETR* [20]ResNet-501250.067.754.233.354.464.4
Rank-DETR [40]ResNet-501250.467.955.233.653.864.2
MS-DETR [50]ResNet-501250.367.455.132.754.064.6
DDQ-DETR [49]ResNet-501250.768.155.7
Relation-DETRResNet-501251.769.156.336.155.666.1
DINO [47]ResNet-502450.468.354.833.353.764.8
DINO [47]ResNet-503651.269.055.835.054.365.3
H-Def-DETR [23]ResNet-503650.068.354.432.952.765.3
Salience-DETR* [20]ResNet-502451.268.955.733.955.565.6
DDQ-DETR [49]ResNet-502452.069.557.235.254.965.9
Relation-DETRResNet-502452.169.756.636.156.066.5
", + "image_path": "f8170a23173dc2b54096e83c25ce82a3e07cf74c6402036e09f38cc1cf36ec53.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 132, + 483, + 484, + 573 + ], + "blocks": [ + { + "bbox": [ + 130, + 449, + 479, + 472 + ], + "lines": [ + { + "bbox": [ + 130, + 449, + 479, + 472 + ], + "spans": [ + { + "bbox": [ + 130, + 449, + 479, + 472 + ], + "type": "text", + "content": "Table 2: Comparison with state-of-the-art methods on COCO val2017 using SwinL(IN-22K) as the backbone." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 132, + 483, + 484, + 573 + ], + "lines": [ + { + "bbox": [ + 132, + 483, + 484, + 573 + ], + "spans": [ + { + "bbox": [ + 132, + 483, + 484, + 573 + ], + "type": "table", + "html": "
MethodBackboneEpochsAP@50:95AP@50AP@75\\( AP_S \\)\\( AP_M \\)\\( AP_L \\)
DINO [47]Swin-L1256.875.462.341.160.673.5
\\( \\mathcal{H} \\text{-Def-DETR} \\) [23]Swin-L1255.975.261.039.159.972.2
Salience-DETR* [20]Swin-L1256.575.061.540.261.272.8
Rank-DETR [40]Swin-L1257.375.962.940.861.373.2
Rank-DINO [40]Swin-L1257.676.063.441.661.473.8
Relation-DETRSwin-L1257.876.162.941.262.174.4
", + "image_path": "f36de4ba72d5cbcfc49048f63c765dc650e27f9d3a40f7dc30fc9160efc394f1.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 596, + 238, + 608 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 596, + 238, + 608 + ], + "spans": [ + { + "bbox": [ + 132, + 596, + 238, + 608 + ], + "type": "text", + "content": "5.3 Ablation study" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "This part conducts an ablation study to explore how the proposed components influence the final detection performance on COCO. The results in Tab. 5 show that, each key component of Relation-DETR consistently contributes to improving AP. Even on a highly-optimized baseline with VariFocal loss [48], our posi" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "text", + "content": "X. Hou et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 136, + 135, + 477, + 237 + ], + "blocks": [ + { + "bbox": [ + 208, + 114, + 405, + 126 + ], + "lines": [ + { + "bbox": [ + 208, + 114, + 405, + 126 + ], + "spans": [ + { + "bbox": [ + 208, + 114, + 405, + 126 + ], + "type": "text", + "content": "Table 3: Quantitative comparison on CSD [43]." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 136, + 135, + 477, + 237 + ], + "lines": [ + { + "bbox": [ + 136, + 135, + 477, + 237 + ], + "spans": [ + { + "bbox": [ + 136, + 135, + 477, + 237 + ], + "type": "table", + "html": "
MethodBackboneEpochsAP@50:95AP@50AP@75\\( AP_S \\)\\( AP_M \\)\\( AP_L \\)
Def-DETR [51]ResNet-5030043.786.236.640.534.910.0
DAB-Def-DETR [36]ResNet-509052.991.255.050.339.40.0
DN-Def-DETR [28]ResNet-506049.988.051.247.637.70.0
DINO [47]ResNet-506053.090.855.550.939.60.0
H-Def-DETR [23]ResNet-506053.090.655.751.239.26.7
Salience-DETR [20]ResNet-506053.292.555.151.040.90.0
Relation-DETRResNet-506054.492.956.053.640.80.0
", + "image_path": "ac08523168598cc8b88173451226749d79fbf128876968a0b59dd016beae9ad3.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 136, + 281, + 477, + 383 + ], + "blocks": [ + { + "bbox": [ + 206, + 260, + 406, + 272 + ], + "lines": [ + { + "bbox": [ + 206, + 260, + 406, + 272 + ], + "spans": [ + { + "bbox": [ + 206, + 260, + 406, + 272 + ], + "type": "text", + "content": "Table 4: Quantitative comparison on MSSD [8]." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 136, + 281, + 477, + 383 + ], + "lines": [ + { + "bbox": [ + 136, + 281, + 477, + 383 + ], + "spans": [ + { + "bbox": [ + 136, + 281, + 477, + 383 + ], + "type": "table", + "html": "
MethodBackboneEpochsAP@50:95AP@50AP@75APSAPMAPL
Def-DETR [51]ResNet-5030033.054.332.09.811.133.3
DAB-Def-DETR [36]ResNet-5012033.760.031.015.926.729.0
DN-Def-DETR [28]ResNet-5012045.674.144.918.631.941.0
H-Def-DETR [23]ResNet-5012046.976.847.120.345.340.7
DINO [47]ResNet-5012051.080.052.520.047.444.8
Salience-DETR [20]ResNet-5012055.478.261.928.747.544.5
Relation-DETRResNet-5012057.479.263.631.453.545.5
", + "image_path": "c4d7acb802e46d98cf849eee8b8b595f1d3add52e351b0e945597fcce8942b9d.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 417, + 482, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 417, + 482, + 491 + ], + "spans": [ + { + "bbox": [ + 130, + 417, + 482, + 491 + ], + "type": "text", + "content": "tion relation encoder and contrast pipeline bring clear improvements of " + }, + { + "bbox": [ + 130, + 417, + 482, + 491 + ], + "type": "inline_equation", + "content": "+0.3\\%" + }, + { + "bbox": [ + 130, + 417, + 482, + 491 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 417, + 482, + 491 + ], + "type": "inline_equation", + "content": "+0.5\\%" + }, + { + "bbox": [ + 130, + 417, + 482, + 491 + ], + "type": "text", + "content": " AP, respectively. Built upon normalized relative geometry features, the position relation overcomes scale bias effectively, thus benefiting consistent performance improvements for different sized objects. For instance, Tab. 5 shows that introducing relation into the baseline with VFL achieves " + }, + { + "bbox": [ + 130, + 417, + 482, + 491 + ], + "type": "inline_equation", + "content": "+1.2\\%" + }, + { + "bbox": [ + 130, + 417, + 482, + 491 + ], + "type": "inline_equation", + "content": "\\mathrm{AP}_S" + }, + { + "bbox": [ + 130, + 417, + 482, + 491 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 417, + 482, + 491 + ], + "type": "inline_equation", + "content": "+1.0\\%" + }, + { + "bbox": [ + 130, + 417, + 482, + 491 + ], + "type": "inline_equation", + "content": "\\mathrm{AP}_M" + }, + { + "bbox": [ + 130, + 417, + 482, + 491 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 417, + 482, + 491 + ], + "type": "inline_equation", + "content": "+1.3\\%" + }, + { + "bbox": [ + 130, + 417, + 482, + 491 + ], + "type": "inline_equation", + "content": "\\mathrm{AP}_L" + }, + { + "bbox": [ + 130, + 417, + 482, + 491 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 132, + 555, + 482, + 635 + ], + "blocks": [ + { + "bbox": [ + 130, + 522, + 482, + 546 + ], + "lines": [ + { + "bbox": [ + 130, + 522, + 482, + 546 + ], + "spans": [ + { + "bbox": [ + 130, + 522, + 482, + 546 + ], + "type": "text", + "content": "Table 5: Ablation study on key components of Relation-DETR (ResNet-50, 1x). We study both baseline and an optimized version with VariFocal Loss [48]." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 132, + 555, + 482, + 635 + ], + "lines": [ + { + "bbox": [ + 132, + 555, + 482, + 635 + ], + "spans": [ + { + "bbox": [ + 132, + 555, + 482, + 635 + ], + "type": "table", + "html": "
VFLRel.contrastAP@50:95AP@50AP@75APSAPMAPL
49.967.454.533.953.563.8
✓50.3(↑0.4)68.1(↑0.7)55.0(↑0.5)34.7(↑0.8)53.4(↓0.1)64.7(↑0.9)
✓50.9(↑0.6)68.2(↑0.1)55.3(↑0.3)34.9(↑0.2)54.6(↑1.2)64.8(↑0.1)
✓✓51.2(↑0.3)68.5(↑0.7)55.7(↑0.4)35.3(↑0.4)54.9(↑0.3)65.5(↑0.7)
✓✓✓51.7(↑0.5)69.1(↑0.6)56.3(↑0.6)36.1(↑0.8)55.6(↑0.7)66.1(↑0.6)
", + "image_path": "18c19cd800f55af37a5bef721b5da00ab334d28f74bfaf5ebdaf7a4c915e95fd.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 381, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 381, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 381, + 91, + 447, + 100 + ], + "type": "text", + "content": "Relation DETR" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 332, + 128 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 332, + 128 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 332, + 128 + ], + "type": "text", + "content": "5.4 Transferability of position relation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 138, + 482, + 270 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 138, + 482, + 270 + ], + "spans": [ + { + "bbox": [ + 130, + 138, + 482, + 270 + ], + "type": "text", + "content": "Our position relation encoder adopts an elegant architectural design, ensuring a promising transferability to existing DETR detectors with minimal modifications. The experimental results in Tab. 6 show that, integrating position relation encoders without any further modifications enhances detection performance with clear margins of " + }, + { + "bbox": [ + 130, + 138, + 482, + 270 + ], + "type": "inline_equation", + "content": "1.6\\%" + }, + { + "bbox": [ + 130, + 138, + 482, + 270 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 138, + 482, + 270 + ], + "type": "inline_equation", + "content": "2.0\\%" + }, + { + "bbox": [ + 130, + 138, + 482, + 270 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 138, + 482, + 270 + ], + "type": "inline_equation", + "content": "0.1\\%" + }, + { + "bbox": [ + 130, + 138, + 482, + 270 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 138, + 482, + 270 + ], + "type": "inline_equation", + "content": "0.2\\%" + }, + { + "bbox": [ + 130, + 138, + 482, + 270 + ], + "type": "text", + "content": " for Deformable-DETR [51], DAB-Deformable-DETR [36], DN-Deformable-DETR [28] and DINO [47], respectively. Interestingly, in comparison to " + }, + { + "bbox": [ + 130, + 138, + 482, + 270 + ], + "type": "inline_equation", + "content": "\\mathrm{AP}_M" + }, + { + "bbox": [ + 130, + 138, + 482, + 270 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 138, + 482, + 270 + ], + "type": "inline_equation", + "content": "\\mathrm{AP}_L" + }, + { + "bbox": [ + 130, + 138, + 482, + 270 + ], + "type": "text", + "content": ", the position relation has a more substantial impact on improving " + }, + { + "bbox": [ + 130, + 138, + 482, + 270 + ], + "type": "inline_equation", + "content": "\\mathrm{AP}_S" + }, + { + "bbox": [ + 130, + 138, + 482, + 270 + ], + "type": "text", + "content": " for Deformable-DETR [51] and DAB-Deformable-DETR [36]. We attribute this to the fact that these early proposed baselines introduce relatively less structural bias and thus benefit more from our explicit position relation prior." + } + ] + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 132, + 323, + 477, + 425 + ], + "blocks": [ + { + "bbox": [ + 130, + 292, + 479, + 314 + ], + "lines": [ + { + "bbox": [ + 130, + 292, + 479, + 314 + ], + "spans": [ + { + "bbox": [ + 130, + 292, + 479, + 314 + ], + "type": "text", + "content": "Table 6: Transfer experiments for the position relation encoder (ResNet-50, " + }, + { + "bbox": [ + 130, + 292, + 479, + 314 + ], + "type": "inline_equation", + "content": "1 \\times" + }, + { + "bbox": [ + 130, + 292, + 479, + 314 + ], + "type": "text", + "content": "). \"+RelEnc\" denotes the version that integrates our position relation encoder." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 132, + 323, + 477, + 425 + ], + "lines": [ + { + "bbox": [ + 132, + 323, + 477, + 425 + ], + "spans": [ + { + "bbox": [ + 132, + 323, + 477, + 425 + ], + "type": "table", + "html": "
MethodAP@50:95AP@50AP@75APSAPMAPL
Def-DETR45.465.049.127.249.661.0
+RelEnc47.0(↑1.6)65.6(↑0.6)51.2(↑2.1)29.3(↑2.1)51.0(↑1.4)62.2(↑1.2)
DAB-Def-DETR46.365.950.428.350.561.9
+RelEnc48.3(↑2.0)66.5(↑0.6)52.9(↑2.5)32.4(↑4.1)52.0(↑1.5)62.0(↑0.1)
DN-Def-DETR47.265.351.330.250.662.4
+RelEnc47.3(↑0.1)65.6(↑0.3)51.4(↑0.1)29.9(↓0.3)50.8(↑0.2)62.1(↓0.3)
", + "image_path": "b0771129a444c8f39f822ea8f9497ba25398f0e7507322598cb3dd92377e566d.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 449, + 482, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 449, + 482, + 557 + ], + "spans": [ + { + "bbox": [ + 130, + 449, + 482, + 557 + ], + "type": "text", + "content": "Moreover, the proposed contrast pipeline can be seen as an extension of hybrid matching [23], utilizing the proposed position relation encoder. Table 7 compares their transferability when integrated into DINO. The results indicate that directly applying hybrid matching [23] to DINO [47] results in a decrease in performance, from " + }, + { + "bbox": [ + 130, + 449, + 482, + 557 + ], + "type": "inline_equation", + "content": "49.9\\%" + }, + { + "bbox": [ + 130, + 449, + 482, + 557 + ], + "type": "text", + "content": " AP to " + }, + { + "bbox": [ + 130, + 449, + 482, + 557 + ], + "type": "inline_equation", + "content": "49.5\\%" + }, + { + "bbox": [ + 130, + 449, + 482, + 557 + ], + "type": "text", + "content": " AP. In contrast, the introduction of both the proposed relation encoder and the extended contrast pipeline consistently increases the performance. This demonstrates the effectiveness of the proposed position relation prior in improve detection performance, overcoming the weak generalizability inherent in hybrid matching." + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 132, + 598, + 485, + 644 + ], + "blocks": [ + { + "bbox": [ + 145, + 578, + 466, + 590 + ], + "lines": [ + { + "bbox": [ + 145, + 578, + 466, + 590 + ], + "spans": [ + { + "bbox": [ + 145, + 578, + 466, + 590 + ], + "type": "text", + "content": "Table 7: Transfer experiments compared with hybrid matching [23] on DINO." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 132, + 598, + 485, + 644 + ], + "lines": [ + { + "bbox": [ + 132, + 598, + 485, + 644 + ], + "spans": [ + { + "bbox": [ + 132, + 598, + 485, + 644 + ], + "type": "table", + "html": "
DINO+Hybrid [23]+RelEnc+RelEnc+Contrast
APAP@50AP@75APAP@50AP@75APAP@50AP@75APAP@50AP@75
49.967.454.549.5†66.654.050.368.155.051.268.755.6
", + "image_path": "4c0ba0484f07d052f7280e31f669de40c39126035cfa95ec20f186631cfa4b5d.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 130, + 645, + 482, + 663 + ], + "lines": [ + { + "bbox": [ + 130, + 645, + 482, + 663 + ], + "spans": [ + { + "bbox": [ + 130, + 645, + 482, + 663 + ], + "type": "text", + "content": "† We found that hybrid matching decreases the performance of DINO. The conclusion is consistent with the results of HDINO (see here) reported by MMDetection [5]." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "text", + "content": "X. Hou et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 331, + 128 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 331, + 128 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 331, + 128 + ], + "type": "text", + "content": "5.5 Intuitive performance comparison" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 133, + 482, + 229 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 133, + 482, + 229 + ], + "spans": [ + { + "bbox": [ + 130, + 133, + 482, + 229 + ], + "type": "text", + "content": "To facilitate an intuitive performance comparison, Fig. 4 plots the convergence curve and the precision-recall curve. Because the position relation prior reduces the requirement for learning structural bias from data [33], Relation-DETR illustrates a faster convergence speed. When training from scratch, it achieves a higher AP than other counterparts with fewer iterations. Specifically, Relation-DETR can achieve over " + }, + { + "bbox": [ + 130, + 133, + 482, + 229 + ], + "type": "inline_equation", + "content": "40\\%" + }, + { + "bbox": [ + 130, + 133, + 482, + 229 + ], + "type": "text", + "content": " AP with only 2 epochs, surpassing existing DETR detectors. In addition to convergence speed, the PR curves under different IoU thresholds also validate the performance gain of our Relation-DETR." + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 140, + 249, + 305, + 377 + ], + "blocks": [ + { + "bbox": [ + 140, + 249, + 305, + 377 + ], + "lines": [ + { + "bbox": [ + 140, + 249, + 305, + 377 + ], + "spans": [ + { + "bbox": [ + 140, + 249, + 305, + 377 + ], + "type": "image", + "image_path": "237f844b9eacd221aacb342100b5e9d2a9a39929a679215ff5dc8119708deed5.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 388, + 480, + 422 + ], + "lines": [ + { + "bbox": [ + 130, + 388, + 480, + 422 + ], + "spans": [ + { + "bbox": [ + 130, + 388, + 480, + 422 + ], + "type": "text", + "content": "Fig. 4: Convergence curve(left) and Precision-recall curve for " + }, + { + "bbox": [ + 130, + 388, + 480, + 422 + ], + "type": "inline_equation", + "content": "\\mathrm{IoU} = 50\\% \\sim 95\\%" + }, + { + "bbox": [ + 130, + 388, + 480, + 422 + ], + "type": "text", + "content": " (right). All models are trained with ResNet-50 backbone under the same " + }, + { + "bbox": [ + 130, + 388, + 480, + 422 + ], + "type": "inline_equation", + "content": "1\\times" + }, + { + "bbox": [ + 130, + 388, + 480, + 422 + ], + "type": "text", + "content": " training configuration on COCO 2017." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 309, + 250, + 474, + 376 + ], + "blocks": [ + { + "bbox": [ + 309, + 250, + 474, + 376 + ], + "lines": [ + { + "bbox": [ + 309, + 250, + 474, + 376 + ], + "spans": [ + { + "bbox": [ + 309, + 250, + 474, + 376 + ], + "type": "image", + "image_path": "c90bae7c9bf3079eb86772dc3acbcda46503e06a5baa859fd2e881fa1d160ffd.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 460, + 227, + 470 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 460, + 227, + 470 + ], + "spans": [ + { + "bbox": [ + 132, + 460, + 227, + 470 + ], + "type": "text", + "content": "5.6 Visualization" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 477, + 482, + 609 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 477, + 482, + 609 + ], + "spans": [ + { + "bbox": [ + 130, + 477, + 482, + 609 + ], + "type": "text", + "content": "For a more intuitive grasp of the relation mechanism, Fig. 5 illustrates representative objects with high relation weights when given a query object. The visualization shows that for both generic and task-specific datasets, the relation contributes to identifying other detection candidates based on the given object query. Furthermore, small-sized objects tend to establish more relation connections with other objects due to the lack of their own semantic information. Therefore, constructing relations is crucial for small-sized object detection. Fig. 6 further visualizes some failure cases of Relation-DETR, indicating that the presented model may benefit from occluded objects and dense objects with misleading semantic differences by considering more complex relations, such as occlusion and semantic relations." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 624, + 336, + 636 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 624, + 336, + 636 + ], + "spans": [ + { + "bbox": [ + 132, + 624, + 336, + 636 + ], + "type": "text", + "content": "5.7 Towards universal object detection" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "content": "Will the position relation prior still be effective for datasets that cover a broader range of scenarios and objects? As a piece of general prior knowledge, we an-" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 381, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 381, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 381, + 91, + 447, + 101 + ], + "type": "text", + "content": "Relation DETR" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 134, + 114, + 204, + 220 + ], + "blocks": [ + { + "bbox": [ + 134, + 114, + 204, + 220 + ], + "lines": [ + { + "bbox": [ + 134, + 114, + 204, + 220 + ], + "spans": [ + { + "bbox": [ + 134, + 114, + 204, + 220 + ], + "type": "image", + "image_path": "b88e6cea89b1cb55de875bdcd9cd43af9d578a420d2331248e09fe803cf1c57b.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 228, + 480, + 252 + ], + "lines": [ + { + "bbox": [ + 132, + 228, + 480, + 252 + ], + "spans": [ + { + "bbox": [ + 132, + 228, + 480, + 252 + ], + "type": "text", + "content": "Fig. 5: Representative objects(red) related to the Fig. 6: Predictions (left) and GTs given object(blue) (right) of failure cases" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 204, + 114, + 283, + 220 + ], + "blocks": [ + { + "bbox": [ + 204, + 114, + 283, + 220 + ], + "lines": [ + { + "bbox": [ + 204, + 114, + 283, + 220 + ], + "spans": [ + { + "bbox": [ + 204, + 114, + 283, + 220 + ], + "type": "image", + "image_path": "910b96b06478edf672035823b8bebc20795af2d27ea30a702cfd0e8fdd2a862e.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 283, + 114, + 337, + 220 + ], + "blocks": [ + { + "bbox": [ + 283, + 114, + 337, + 220 + ], + "lines": [ + { + "bbox": [ + 283, + 114, + 337, + 220 + ], + "spans": [ + { + "bbox": [ + 283, + 114, + 337, + 220 + ], + "type": "image", + "image_path": "4f4eada323159925d9da589f54b79022b2ec0cd7ccc3ae9d98bd3ccbfe77a7f3.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 337, + 114, + 408, + 220 + ], + "blocks": [ + { + "bbox": [ + 337, + 114, + 408, + 220 + ], + "lines": [ + { + "bbox": [ + 337, + 114, + 408, + 220 + ], + "spans": [ + { + "bbox": [ + 337, + 114, + 408, + 220 + ], + "type": "image", + "image_path": "6013a8ef88a60636fba0ab2dd2cafa920d9dc5cf53c46ec8c5c6278c2dff689e.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 408, + 114, + 479, + 220 + ], + "blocks": [ + { + "bbox": [ + 408, + 114, + 479, + 220 + ], + "lines": [ + { + "bbox": [ + 408, + 114, + 479, + 220 + ], + "spans": [ + { + "bbox": [ + 408, + 114, + 479, + 220 + ], + "type": "image", + "image_path": "229d92370487151532f69edba696994cf966f2f18b007b085ce8fd8b40ac4ddd.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 274, + 482, + 370 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 274, + 482, + 370 + ], + "spans": [ + { + "bbox": [ + 130, + 274, + 482, + 370 + ], + "type": "text", + "content": "ticipate that the explicit position relation prior can benefit universal object detection tasks. To explore this, we have constructed a large-scale class-agnostic detection dataset with about 100,000 images, termed SA-Det-100k, by sampling a subset from SA-1B, which is one of the largest scale segmentation datasets proposed in Segment Anything [25]. We then compared the performance of our Relation-DETR with the baseline DINO [47] using VFL [48] on this dataset. The results in Tab. 8 show that Relation-DETR achieves a clear improvement of " + }, + { + "bbox": [ + 130, + 274, + 482, + 370 + ], + "type": "inline_equation", + "content": "1.3\\%" + }, + { + "bbox": [ + 130, + 274, + 482, + 370 + ], + "type": "text", + "content": " AP, demonstrating the scalability of the proposed position relation prior." + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 137, + 411, + 477, + 459 + ], + "blocks": [ + { + "bbox": [ + 165, + 390, + 448, + 403 + ], + "lines": [ + { + "bbox": [ + 165, + 390, + 448, + 403 + ], + "spans": [ + { + "bbox": [ + 165, + 390, + 448, + 403 + ], + "type": "text", + "content": "Table 8: Quantitative comparison on SA-Det-100k (ResNet-50, " + }, + { + "bbox": [ + 165, + 390, + 448, + 403 + ], + "type": "inline_equation", + "content": "1 \\times" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 137, + 411, + 477, + 459 + ], + "lines": [ + { + "bbox": [ + 137, + 411, + 477, + 459 + ], + "spans": [ + { + "bbox": [ + 137, + 411, + 477, + 459 + ], + "type": "table", + "html": "
MethodsAP@50:95AP@50AP@75APSAPMAPL
DINO with VFL43.752.047.75.843.061.5
Relation-DETR45.0(↑1.3)53.1(↑1.1)48.9(↑1.2)6.0(↑0.2)44.4(↑1.4)62.9(↑1.4)
", + "image_path": "9a1ebd98438d6ced4efc893f1aa48dd08aa958499bc86bb91e30653ffac0bd5c.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 498, + 219, + 510 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 498, + 219, + 510 + ], + "spans": [ + { + "bbox": [ + 132, + 498, + 219, + 510 + ], + "type": "text", + "content": "6 Conclusion" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 521, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 482, + 665 + ], + "type": "text", + "content": "This paper explores explicit position relation prior for enhancing performance and convergence of DETR detectors. Built upon normalized relative geometry features, we propose a novel position relation that overcomes scale bias for progressive attention refinement. To address the conflicts between non-duplicate predictions and sufficient positive supervision in DETR frameworks, we extend the streaming pipeline to a contrast pipeline based on the proposed position relation. Combining these components produces a state-of-the-art detector, named Relation-DETR. Massive ablation studies and experimental results demonstrate the superior performance, faster convergence and promising transferability of the proposed detector. Moreover, Relation-DETR exhibits remarkable generalizability for both generic and task-specific detection tasks. We believe that this work will inspire future research on relation and structural bias for DETR detectors." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "text", + "content": "X. Hou et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 246, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 246, + 129 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 246, + 129 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 139, + 482, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 139, + 482, + 175 + ], + "spans": [ + { + "bbox": [ + 130, + 139, + 482, + 175 + ], + "type": "text", + "content": "This work was supported by the National Natural Science Foundation of China under Grant 62327808, and the Fundamental Research Funds for Xi'an Jiaotong University under Grants xtr072022001 and xzy022024009." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 193, + 197, + 205 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 193, + 197, + 205 + ], + "spans": [ + { + "bbox": [ + 132, + 193, + 197, + 205 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 217, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 138, + 217, + 480, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 217, + 480, + 239 + ], + "spans": [ + { + "bbox": [ + 138, + 217, + 480, + 239 + ], + "type": "text", + "content": "1. Bodla, N., Singh, B., Chellappa, R., Davis, L.S.: Soft-NMS-improving object detection with one line of code. In: Int. Conf. Comput. Vis. pp. 5561-5569 (2017)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 240, + 480, + 261 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 240, + 480, + 261 + ], + "spans": [ + { + "bbox": [ + 138, + 240, + 480, + 261 + ], + "type": "text", + "content": "2. Bolya, D., Zhou, C., Xiao, F., Lee, Y.J.: YOLACT: Real-time instance segmentation. In: Int. Conf. Comput. Vis. pp. 9157-9166 (2019)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 262, + 480, + 283 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 262, + 480, + 283 + ], + "spans": [ + { + "bbox": [ + 138, + 262, + 480, + 283 + ], + "type": "text", + "content": "3. Cai, Z., Liu, S., Wang, G., Ge, Z., Zhang, X., Huang, D.: Align-DETR: Improving DETR with simple IoU-aware BCE loss. arXiv preprint arXiv:2304.07527 (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 284, + 480, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 284, + 480, + 316 + ], + "spans": [ + { + "bbox": [ + 138, + 284, + 480, + 316 + ], + "type": "text", + "content": "4. Carion, N., Massa, F., Synnaeve, G., Usunier, N., Kirillov, A., Zagoruyko, S.: End-to-end object detection with transformers. In: Eur. Conf. Comput. Vis. pp. 213-229. Springer (2020)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 316, + 480, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 316, + 480, + 348 + ], + "spans": [ + { + "bbox": [ + 138, + 316, + 480, + 348 + ], + "type": "text", + "content": "5. Chen, K., Wang, J., Pang, J., Cao, Y., Xiong, Y., Li, X., Sun, S., Feng, W., Liu, Z., Xu, J., et al.: Mmdetection: Open mmlab detection toolbox and benchmark. arXiv preprint arXiv:1906.07155 (2019)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 349, + 480, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 349, + 480, + 381 + ], + "spans": [ + { + "bbox": [ + 138, + 349, + 480, + 381 + ], + "type": "text", + "content": "6. Chen, Q., Chen, X., Wang, J., Zhang, S., Yao, K., Feng, H., Han, J., Ding, E., Zeng, G., Wang, J.: Group DETR: Fast DETR training with group-wise one-to-many assignment. In: Int. Conf. Comput. Vis. pp. 6633-6642 (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 382, + 480, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 382, + 480, + 403 + ], + "spans": [ + { + "bbox": [ + 138, + 382, + 480, + 403 + ], + "type": "text", + "content": "7. Chen, X., Li, L.J., Fei-Fei, L., Gupta, A.: Iterative visual reasoning beyond convolutions. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 7239-7248 (2018)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 403, + 480, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 403, + 480, + 435 + ], + "spans": [ + { + "bbox": [ + 138, + 403, + 480, + 435 + ], + "type": "text", + "content": "8. Chen, Y., Pan, J., Lei, J., Zeng, D., Wu, Z., Chen, C.: EEE-Net: Efficient edge enhanced network for surface defect detection of glass. IEEE Transactions on Instrumentation and Measurement (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 436, + 480, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 436, + 480, + 468 + ], + "spans": [ + { + "bbox": [ + 138, + 436, + 480, + 468 + ], + "type": "text", + "content": "9. Cordts, M., Omran, M., Ramos, S., Rehfeld, T., Enzweiler, M., Benenson, R., Franke, U., Roth, S., Schiele, B.: The cityscapes dataset for semantic urban scene understanding. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 3213-3223 (2016)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 133, + 469, + 480, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 469, + 480, + 502 + ], + "spans": [ + { + "bbox": [ + 133, + 469, + 480, + 502 + ], + "type": "text", + "content": "10. Dai, X., Chen, Y., Yang, J., Zhang, P., Yuan, L., Zhang, L.: Dynamic DETR: End-to-end object detection with dynamic attention. In: Int. Conf. Comput. Vis. pp. 2988-2997 (2021)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 133, + 502, + 480, + 534 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 502, + 480, + 534 + ], + "spans": [ + { + "bbox": [ + 133, + 502, + 480, + 534 + ], + "type": "text", + "content": "11. Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: ImageNet: A large-scale hierarchical image database. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 248-255. IEEE (2009)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 133, + 535, + 480, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 535, + 480, + 578 + ], + "spans": [ + { + "bbox": [ + 133, + 535, + 480, + 578 + ], + "type": "text", + "content": "12. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. In: Int. Conf. Learn. Represent. (2020)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 133, + 578, + 480, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 578, + 480, + 599 + ], + "spans": [ + { + "bbox": [ + 133, + 578, + 480, + 599 + ], + "type": "text", + "content": "13. Everingham, M., Van Gool, L., Williams, C.K., Winn, J., Zisserman, A.: The pascal visual object classes (voc) challenge. Int. J. Comput. Vis. 88, 303-338 (2010)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 133, + 600, + 480, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 600, + 480, + 621 + ], + "spans": [ + { + "bbox": [ + 133, + 600, + 480, + 621 + ], + "type": "text", + "content": "14. Fang, Y., Sun, Q., Wang, X., Huang, T., Wang, X., Cao, Y.: EVA-02: A visual representation for neon genesis. arXiv preprint arXiv:2303.11331 (2023)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 133, + 622, + 480, + 654 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 622, + 480, + 654 + ], + "spans": [ + { + "bbox": [ + 133, + 622, + 480, + 654 + ], + "type": "text", + "content": "15. Feng, C., Zhong, Y., Gao, Y., Scott, M.R., Huang, W.: TOOD: Task-aligned one-stage object detection. In: Int. Conf. Comput. Vis. pp. 3490-3499. IEEE Computer Society (2021)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 133, + 654, + 460, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 654, + 460, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 654, + 460, + 665 + ], + "type": "text", + "content": "16. Girshick, R.: Fast R-CNN. In: Int. Conf. Comput. Vis. pp. 1440-1448 (2015)" + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 381, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 381, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 381, + 91, + 447, + 100 + ], + "type": "text", + "content": "Relation DETR" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 480, + 138 + ], + "type": "text", + "content": "17. Hao, X., Huang, D., Lin, J., Lin, C.Y.: Relation-enhanced DETR for component detection in graphic design reverse engineering. In: IJCAI. pp. 4785-4793 (2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 140, + 480, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 140, + 480, + 161 + ], + "spans": [ + { + "bbox": [ + 132, + 140, + 480, + 161 + ], + "type": "text", + "content": "18. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 770-778 (2016)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 162, + 480, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 162, + 480, + 195 + ], + "spans": [ + { + "bbox": [ + 132, + 162, + 480, + 195 + ], + "type": "text", + "content": "19. Hou, X., Liu, M., Zhang, S., Wei, P., Chen, B.: CANet: Contextual information and spatial attention based network for detecting small defects in manufacturing industry. Pattern Recognition 140, 109558 (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 196, + 480, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 196, + 480, + 228 + ], + "spans": [ + { + "bbox": [ + 132, + 196, + 480, + 228 + ], + "type": "text", + "content": "20. Hou, X., Liu, M., Zhang, S., Wei, P., Chen, B.: Salience detr: Enhancing detection transformer with hierarchical salience filtering refinement. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 17574-17583 (2024)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 229, + 480, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 229, + 480, + 251 + ], + "spans": [ + { + "bbox": [ + 132, + 229, + 480, + 251 + ], + "type": "text", + "content": "21. Hu, H., Gu, J., Zhang, Z., Dai, J., Wei, Y.: Relation networks for object detection. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 3588-3597 (2018)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 251, + 480, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 251, + 480, + 274 + ], + "spans": [ + { + "bbox": [ + 132, + 251, + 480, + 274 + ], + "type": "text", + "content": "22. Hu, Z., Sun, Y., Wang, J., Yang, Y.: DAC-DETR: Divide the attention layers and conquer. Adv. Neural Inform. Process. Syst. 36 (2024)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 274, + 480, + 307 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 274, + 480, + 307 + ], + "spans": [ + { + "bbox": [ + 132, + 274, + 480, + 307 + ], + "type": "text", + "content": "23. Jia, D., Yuan, Y., He, H., Wu, X., Yu, H., Lin, W., Sun, L., Zhang, C., Hu, H.: DETRs with hybrid matching. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 19702-19712 (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 308, + 480, + 329 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 308, + 480, + 329 + ], + "spans": [ + { + "bbox": [ + 132, + 308, + 480, + 329 + ], + "type": "text", + "content": "24. Jiang, C., Xu, H., Liang, X., Lin, L.: Hybrid knowledge routed modules for large-scale object detection. Adv. Neural Inform. Process. Syst. 31 (2018)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 331, + 480, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 331, + 480, + 363 + ], + "spans": [ + { + "bbox": [ + 132, + 331, + 480, + 363 + ], + "type": "text", + "content": "25. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., et al.: Segment anything. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 4015-4026 (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 364, + 480, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 364, + 480, + 396 + ], + "spans": [ + { + "bbox": [ + 132, + 364, + 480, + 396 + ], + "type": "text", + "content": "26. Kolve, E., Mottaghi, R., Han, W., VanderBilt, E., Weihs, L., Herrasti, A., Deitke, M., Ehsani, K., Gordon, D., Zhu, Y., et al.: Ai2-thor: An interactive 3d environment for visual ai. arXiv preprint arXiv:1712.05474 (2017)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 397, + 480, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 397, + 480, + 441 + ], + "spans": [ + { + "bbox": [ + 132, + 397, + 480, + 441 + ], + "type": "text", + "content": "27. Krishna, R., Zhu, Y., Groth, O., Johnson, J., Hata, K., Kravitz, J., Chen, S., Kalantidis, Y., Li, L.J., Shamma, D.A., et al.: Visual genome: Connecting language and vision using crowdsourced dense image annotations. Int. J. Comput. Vis. 123, 32-73 (2017)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 441, + 480, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 441, + 480, + 475 + ], + "spans": [ + { + "bbox": [ + 132, + 441, + 480, + 475 + ], + "type": "text", + "content": "28. Li, F., Zhang, H., Liu, S., Guo, J., Ni, L.M., Zhang, L.: DN-DETR: Accelerate DETR training by introducing query denoising. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 13619-13627 (2022)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 475, + 480, + 508 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 475, + 480, + 508 + ], + "spans": [ + { + "bbox": [ + 132, + 475, + 480, + 508 + ], + "type": "text", + "content": "29. Li, X., Wang, W., Wu, L., Chen, S., Hu, X., Li, J., Tang, J., Yang, J.: Generalized focal loss: Learning qualified and distributed bounding boxes for dense object detection. Adv. Neural Inform. Process. Syst. 33, 21002-21012 (2020)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 509, + 480, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 509, + 480, + 531 + ], + "spans": [ + { + "bbox": [ + 132, + 509, + 480, + 531 + ], + "type": "text", + "content": "30. Li, Y., Mao, H., Girshick, R., He, K.: Exploring plain vision transformer backbones for object detection. In: Eur. Conf. Comput. Vis. pp. 280-296. Springer (2022)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 532, + 480, + 563 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 532, + 480, + 563 + ], + "spans": [ + { + "bbox": [ + 132, + 532, + 480, + 563 + ], + "type": "text", + "content": "31. Li, Z., Du, X., Cao, Y.: Gar: Graph assisted reasoning for object detection. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 1295-1304 (2020)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 564, + 480, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 564, + 480, + 597 + ], + "spans": [ + { + "bbox": [ + 132, + 564, + 480, + 597 + ], + "type": "text", + "content": "32. Lin, J., Pan, Y., Lai, R., Yang, X., Chao, H., Yao, T.: Core-Text: Improving scene text detection with contrastive relational reasoning. In: Int. Conf. Multimedia and Expo. pp. 1-6. IEEE (2021)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 598, + 468, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 598, + 468, + 609 + ], + "spans": [ + { + "bbox": [ + 132, + 598, + 468, + 609 + ], + "type": "text", + "content": "33. Lin, T., Wang, Y., Liu, X., Qiu, X.: A survey of transformers. AI Open (2022)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 610, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 610, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 610, + 480, + 632 + ], + "type": "text", + "content": "34. Lin, T.Y., Goyal, P., Girshick, R., He, K., Dollar, P.: Focal loss for dense object detection. In: Int. Conf. Comput. Vis. pp. 2980-2988 (2017)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "type": "text", + "content": "35. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollar, P., Zitnick, C.L.: Microsoft COCO: Common objects in context. In: Eur. Conf. Comput. Vis. pp. 740-755. Springer (2014)" + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "text", + "content": "X. Hou et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 643 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "text", + "content": "36. Liu, S., Li, F., Zhang, H., Yang, X., Qi, X., Su, H., Zhu, J., Zhang, L.: DABDETR: Dynamic anchor boxes are better queries for DETR. In: Int. Conf. Learn. Represent. (2021)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 150, + 481, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 150, + 481, + 183 + ], + "spans": [ + { + "bbox": [ + 130, + 150, + 481, + 183 + ], + "type": "text", + "content": "37. Liu, S., Ren, T., Chen, J., Zeng, Z., Zhang, H., Li, F., Li, H., Huang, J., Su, H., Zhu, J., et al.: Detection transformer with stable matching. arXiv preprint arXiv:2304.04742 (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 183, + 481, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 183, + 481, + 205 + ], + "spans": [ + { + "bbox": [ + 132, + 183, + 481, + 205 + ], + "type": "text", + "content": "38. Liu, S., Huang, D., Wang, Y.: Adaptive NNS: Refining pedestrian detection in a crowd. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 6459-6468 (2019)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 205, + 481, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 205, + 481, + 237 + ], + "spans": [ + { + "bbox": [ + 132, + 205, + 481, + 237 + ], + "type": "text", + "content": "39. Liu, Z., Lin, Y., Cao, Y., Hu, H., Wei, Y., Zhang, Z., Lin, S., Guo, B.: Swin transformer: Hierarchical vision transformer using shifted windows. In: Int. Conf. Comput. Vis. pp. 10012-10022 (2021)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 238, + 481, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 238, + 481, + 270 + ], + "spans": [ + { + "bbox": [ + 132, + 238, + 481, + 270 + ], + "type": "text", + "content": "40. Pu, Y., Liang, W., Hao, Y., Yuan, Y., Yang, Y., Zhang, C., Hu, H., Huang, G.: Rank-DETR for high quality object detection. Adv. Neural Inform. Process. Syst. 36 (2024)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 271, + 481, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 271, + 481, + 304 + ], + "spans": [ + { + "bbox": [ + 132, + 271, + 481, + 304 + ], + "type": "text", + "content": "41. Shao, S., Li, Z., Zhang, T., Peng, C., Yu, G., Zhang, X., Li, J., Sun, J.: Objects365: A large-scale, high-quality dataset for object detection. In: Int. Conf. Comput. Vis. pp. 8430-8439 (2019)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 304, + 481, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 304, + 481, + 335 + ], + "spans": [ + { + "bbox": [ + 132, + 304, + 481, + 335 + ], + "type": "text", + "content": "42. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. Adv. Neural Inform. Process. Syst. 30 (2017)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 336, + 481, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 336, + 481, + 369 + ], + "spans": [ + { + "bbox": [ + 132, + 336, + 481, + 369 + ], + "type": "text", + "content": "43. Wang, Q., Gao, S., Xiong, L., Liang, A., Jiang, K., Zhang, W.: A casting surface dataset and benchmark for subtle and confusable defect detection in complex contexts. IEEE Sensors Journal pp. 1-1 (2024)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 370, + 481, + 391 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 370, + 481, + 391 + ], + "spans": [ + { + "bbox": [ + 132, + 370, + 481, + 391 + ], + "type": "text", + "content": "44. Wang, Y., Zhang, X., Yang, T., Sun, J.: Anchor DETR: Query design for transformer-based detector. In: AAAI. vol. 36, pp. 2567-2575 (2022)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 392, + 481, + 423 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 392, + 481, + 423 + ], + "spans": [ + { + "bbox": [ + 132, + 392, + 481, + 423 + ], + "type": "text", + "content": "45. Xu, H., Jiang, C., Liang, X., Lin, L., Li, Z.: Reasoning-RCNN: Unifying adaptive global reasoning into large-scale object detection. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 6419-6428 (2019)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 424, + 481, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 424, + 481, + 456 + ], + "spans": [ + { + "bbox": [ + 132, + 424, + 481, + 456 + ], + "type": "text", + "content": "46. Ye, M., Ke, L., Li, S., Tai, Y.W., Tang, C.K., Danelljan, M., Yu, F.: Cascade-DETR: Delving into high-quality universal object detection. In: Int. Conf. Comput. Vis. pp. 6704-6714 (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 457, + 481, + 489 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 457, + 481, + 489 + ], + "spans": [ + { + "bbox": [ + 132, + 457, + 481, + 489 + ], + "type": "text", + "content": "47. Zhang, H., Li, F., Liu, S., Zhang, L., Su, H., Zhu, J., Ni, L., Shum, H.Y.: DINO: DETR with improved denoising anchor boxes for end-to-end object detection. In: Int. Conf. Learn. Represent. (2022)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 490, + 481, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 490, + 481, + 521 + ], + "spans": [ + { + "bbox": [ + 132, + 490, + 481, + 521 + ], + "type": "text", + "content": "48. Zhang, H., Wang, Y., Dayoub, F., Sunderhauf, N.: Varifocalnet: An iou-aware dense object detector. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 8514-8523 (2021)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 522, + 481, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 522, + 481, + 555 + ], + "spans": [ + { + "bbox": [ + 132, + 522, + 481, + 555 + ], + "type": "text", + "content": "49. Zhang, S., Wang, X., Wang, J., Pang, J., Lyu, C., Zhang, W., Luo, P., Chen, K.: Dense distinct query for end-to-end object detection. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 7329-7338 (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 555, + 481, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 555, + 481, + 588 + ], + "spans": [ + { + "bbox": [ + 132, + 555, + 481, + 588 + ], + "type": "text", + "content": "50. Zhao, C., Sun, Y., Wang, W., Chen, Q., Ding, E., Yang, Y., Wang, J.: Ms-detr: Efficient detr training with mixed supervision. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 17027-17036 (2024)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 589, + 481, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 589, + 481, + 621 + ], + "spans": [ + { + "bbox": [ + 132, + 589, + 481, + 621 + ], + "type": "text", + "content": "51. Zhu, X., Su, W., Lu, L., Li, B., Wang, X., Dai, J.: Deformable DETR: Deformable transformers for end-to-end object detection. In: Int. Conf. Learn. Represent. (2020)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 622, + 481, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 622, + 481, + 643 + ], + "spans": [ + { + "bbox": [ + 132, + 622, + 481, + 643 + ], + "type": "text", + "content": "52. Zong, Z., Song, G., Liu, Y.: DETRs with collaborative hybrid assignments training. In: Int. Conf. Comput. Vis. pp. 6748-6758 (2023)" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 381, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 381, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 381, + 91, + 447, + 100 + ], + "type": "text", + "content": "Relation DETR" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Reliability in Semantic Segmentation_ Can We Use Synthetic Data_/3d29ee45-5380-4719-adfd-255f20df0905_content_list.json b/2024/Reliability in Semantic Segmentation_ Can We Use Synthetic Data_/3d29ee45-5380-4719-adfd-255f20df0905_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..e4cd8acc47d78c985d5fc8d7fb2557a3496a00ce --- /dev/null +++ b/2024/Reliability in Semantic Segmentation_ Can We Use Synthetic Data_/3d29ee45-5380-4719-adfd-255f20df0905_content_list.json @@ -0,0 +1,1499 @@ +[ + { + "type": "text", + "text": "Reliability in Semantic Segmentation: Can We Use Synthetic Data?", + "text_level": 1, + "bbox": [ + 285, + 140, + 715, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Thibaut Loiseau $^{2\\dagger}$ , Tuan-Hung Vu $^{1}$ , Mickael Chen $^{1}$ , Patrick Pérez $^{3}$ , and Matthieu Cord $^{1,4}$", + "bbox": [ + 238, + 210, + 764, + 244 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ valeo.ai, Paris, France \n $^{2}$ LIGM, Ecole des Ponts, Univ Gustave Eiffel, CNRS, Marne-la-Vallee, France", + "bbox": [ + 236, + 253, + 764, + 281 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3 kyutai, Paris, France", + "bbox": [ + 423, + 282, + 580, + 296 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "4 Sorbonne Université, Paris, France", + "bbox": [ + 377, + 296, + 624, + 309 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Assessing the robustness of perception models to covariate shifts and their ability to detect out-of-distribution (OOD) inputs is crucial for safety-critical applications such as autonomous vehicles. By nature of such applications, however, the relevant data is difficult to collect and annotate. In this paper, we show for the first time how synthetic data can be specifically generated to assess comprehensively the real-world reliability of semantic segmentation models. By fine-tuning Stable Diffusion [31] with only in-domain data, we perform zero-shot generation of visual scenes in OOD domains or inpainted with OOD objects. This synthetic data is employed to evaluate the robustness of pretrained segmenters, thereby offering insights into their performance when confronted with real edge cases. Through extensive experiments, we demonstrate a high correlation between the performance of models when evaluated on our synthetic OOD data and when evaluated on real OOD inputs, showing the relevance of such virtual testing. Furthermore, we demonstrate how our approach can be utilized to enhance the calibration and OOD detection capabilities of segmenters. Code and data are made public.", + "bbox": [ + 259, + 343, + 743, + 580 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 625, + 375, + 641 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Despite the rapid adoption of deep networks in safety-critical applications, reliability [26,36,38] has been an overlooked factor when designing and training these models. Recent efforts are geared toward enhancing model robustness under covariate shifts in data distributions [26, 30] and improving the model's ability to detect the unknown [10, 12, 41]. For open-world validation, in-domain data is no longer sufficient [16]; reliable and trustworthy systems demand more rigorous testing on diverse distributions, potentially exhibiting unknown objects. However, data collection campaigns can be quickly overwhelmed by the growing number of out-of-distribution (OOD) objects and conditions, sometimes extreme.", + "bbox": [ + 212, + 654, + 787, + 790 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this paper, we propose to leverage pretrained generative models, $e.g.$ Stable Diffusion 1.5 (SD), to alleviate the need for real OOD data. We use SD as a", + "bbox": [ + 212, + 791, + 787, + 821 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Work done during an internship at valeo.ai", + "bbox": [ + 230, + 825, + 532, + 840 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/7d42448a0ecaa80a6205718151f8dd2cc847c2bd0d767f1bd6866fa951c07912.jpg", + "image_caption": [ + "Fig. 1: Assessing 40 pretrained segmenters under covariate shifts. Segmentation models under scrutiny were trained on Cityscapes train set only (in-domain data). They are evaluated on (i) Cityscapes validation set, (ii) real OOD data, and (iii) proposed synthetic data. We observe a strong correlation between results on (ii) and (iii)." + ], + "image_footnote": [], + "bbox": [ + 241, + 145, + 763, + 300 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "general validator, targeting multiple faces of reliability:", + "bbox": [ + 212, + 407, + 612, + 421 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "- First, we fine-tune SD on in-domain data to enable mask conditioning; in action, zero-shot prompting generates covariate shift images for testing. Fig. 1 illustrates the generation of synthetic data in OOD domains and its use for model evaluation. By assessing 40 pretrained segmenters under covariate shifts, our first contribution demonstrates how our generated synthetic OOD benchmark can act as a powerful proxy for the real OOD benchmark ACDC [33].", + "bbox": [ + 212, + 422, + 785, + 513 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "- Second, we explore a similar strategy for OOD object detection assessment: we inpaint objects of unknown classes into in-domain data by injecting appropriate prompts during the zero-shot inpainting process. Inpainted images and OOD masks are used to benchmark the 40 models to see how well they can recognize OOD objects. Our results are strongly correlated to the real OOD benchmark [4]. Our high-quality synthetic data is featured in the official BRAVO benchmark.", + "bbox": [ + 212, + 513, + 787, + 602 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "- Third, we demonstrate the usefulness of our two synthetic OOD benchmarks for hyperparameter tuning (here model calibration) and training. We calibrate the pretrained models to targeted OOD domains using our synthetic data and validate our strategy by comparing results to using real OOD data. We also train segmenters on inpainted data for OOD detection, obtaining competitive results.", + "bbox": [ + 212, + 603, + 785, + 679 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 702, + 388, + 718 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Covariate Shifts. Modern machine learning models, notably deep networks, fall short in preserving their robustness and estimating their prediction confidence in the presence of covariate shifts [26,30]. Various benchmarks [5,11,16,35,36] have emerged to address the need for assessing models' reliability under different distributional shifts. Hendrycks et al. [11] propose a pioneering benchmark featuring data corrupted by various synthetic perturbations such as noise, blur, and brightness. Taori et al. [36] emphasize the importance of realistic shifts in", + "bbox": [ + 212, + 733, + 787, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "T.Loiseau et al.", + "bbox": [ + 271, + 114, + 377, + 126 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "reliability assessment; they highlight the disparity between natural and synthetic shifts, asserting that there is minimal to no robustness transfer from synthetic to natural distribution shifts. Sign et al. [35] study low-shot robustness to natural distribution shifts, highlighting the robustness properties of advanced architecture and pretraining strategies. In [14], de Jorge et al. extend beyond the standard classification setup to address the reliability problem in semantic segmentation, with findings largely aligned with previous works on classification; interestingly, they point to the disconnection between robustness and confidence calibration, urging more attention to calibration during segmenter design and training. On a related line, prior works study the connections between in-domain and out-of-domain robustness [14, 23, 37]; they suggest either positive, none, or even negative correlations between ID and OOD robustness, which indeed largely depends on the type of shifts. In line with [14], we focus on segmentation as the visual perception task of interest. Different from previous works, we advocate for the use of advanced generative models to generate realistic synthetic data for testing segmenters under arbitrary covariate shifts. Taori et al. [36] criticize the synthetic robustness benchmarks and advocate for using real-shift data; while we agree on the importance of realism in testing, we demonstrate that the rapid advancement of generative models now permits very meaningful virtual assessment. Our goal is to study whether synthetic data can be a superior choice compared to ID data in correlation studies against OOD robustness – referred to as real-shift robustness here to avoid confusion with OOD detection. Inspired by [14], we also calibrate models but using synthetic OOD data instead.", + "bbox": [ + 217, + 146, + 785, + 492 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "OOD Object Detection. In addition to robustness and calibration, the ability to detect the \"unknown\" is equally important to assess for trustworthy systems [10,12,41]. In semantic segmentation, several datasets are available for evaluating OOD detection on the road, namely LostAndFound [27], StreetHazards [9] (synthetic), BDD-Anomaly [9], Fishyscapes [3] (synthetic), and SegmentMeIfYouCan [4]. Encountering and capturing images of OOD objects in real-world scenarios, without deliberately placing them on the road, is quite uncommon. Effectively, existing datasets are limited in scale, both in terms of the number of images and the variety of object classes. In this work, we propose leveraging advanced zero-shot inpainting techniques to augment an existing segmentation dataset with the insertion of OOD objects; this enables the generation of highly realistic synthetic data for testing and training OOD detection.", + "bbox": [ + 217, + 507, + 785, + 688 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Synthetic Data for Testing. Generative models have been exploited to create training data for image classification [2,8,34], object detection [22], or semantic segmentation [7,17,18,40,43]. Only recently, a few works have delved into the topic of generative testing data. Li et al. [19] exploit diffusion models to realistically edit images, controlling over various object attributes. This approach enables stress-testing models and understanding their sensitivity to different attributes. Using SD, LANCE [29] generates counterfactual images capable of challenging any given perception model. To the extent of our knowledge, no existing works have proposed to generate testing data for segmentation reliability.", + "bbox": [ + 217, + 703, + 785, + 839 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "Reliability in Semantic Segmentation: Can We Use Synthetic Data?", + "bbox": [ + 281, + 114, + 730, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3 Reliability Under Covariate Shifts", + "text_level": 1, + "bbox": [ + 215, + 143, + 581, + 162 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this section, we explore whether synthetic data can be used to assess the robustness of pretrained segmenters in the presence of covariate shifts to unseen OOD domains. We describe the data generation process and present the benchmarking results for a wide range of segmenters on our synthetic data. We demonstrate the validity of the approach using domains for which a real OOD dataset exists so that we can have access to a gold standard. We stress that we do not use the OOD data at any point in the method itself. This aspect is critical for the method to be applicable to benchmark robustness in the presence of extreme or hazardous conditions.", + "bbox": [ + 212, + 176, + 787, + 311 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1 Generating images in arbitrary domains", + "text_level": 1, + "bbox": [ + 215, + 335, + 591, + 351 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/ca8c2188a25cf14eb2482daa909b816ba33e7b6d7aaf38bef4ea68a02a703ebf.jpg", + "image_caption": [ + "Fig. 2: Generating data with covariate shifts. Training (left) and Sampling (right) processes for producing the synthetic data with shifts. For training, only in-domain images and masks are used. For inference, we use the in-domain masks to generate OOD images. No real OOD data is required in the framework." + ], + "image_footnote": [], + "bbox": [ + 218, + 387, + 488, + 483 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/48399ad6288318a3e5ae8042f87e2998c51f7f1f113d94ac96dcec34c4d676df.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 509, + 388, + 782, + 484 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our goal here is to obtain pairs of images and semantic masks, with the images belonging to visual domains for which we lack data. To this end, we leverage a pretrained text-to-image Stable Diffusion 1.5 (SD) model [31], repurposed as a semantic-conditioned model called ControlNet [42]. ControlNet is trained solely on images and segmentation ground truths from the Cityscapes dataset; at train time, the text prompts are captions automatically extracted using CLIP-interrogator [1]. As a result, the model is able to perform mask-to-image generation of driving scenes while retaining the ability to steer the generation through text prompting of Stable Diffusion.", + "bbox": [ + 212, + 583, + 787, + 718 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To generate synthetic data, we prompt a trained ControlNet by the concatenation of OOD domain descriptions and CLIP-interrogator captions obtained from Cityscapes validation images. Also, segmentation masks in Cityscapes validation set are used to condition the generative process. Thanks to zero-shot prompting, the synthetic images are aligned with the semantic condition while displaying the visual properties of arbitrary OOD domains. Fig. 2 illustrates the training and generation steps. Detailed technical descriptions and more visualizations are in the Supplementary Material.", + "bbox": [ + 212, + 719, + 787, + 839 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "T.Loiseau et al.", + "bbox": [ + 271, + 114, + 377, + 127 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/12e108e582f8a05ae89529ef01ed428fe795cc5eadf7c01591aee891e20372cb.jpg", + "image_caption": [ + "Fig. 3: Robustness correlation between real and synthetic covariate shifts across 40 pretrained segmenters. The tested models, see families in bottom legend, cover different architectures and sizes. (top) Pearson Correlation Coefficients of mIoUs between Cityscapes and real-shifts ('PCC_CS'), and between synthetic shifts and real ones ('PCC_Syn'). (bottom) Scatter plots of synthetic vs. real mIoUs along with the linear regression line accompanied by $95\\%$ confidence intervals ('CI'). $(a - e)$ Five types of domain shifts from Cityscapes in-domain distribution, sorted by increasing gap as assessed by decreasing PCC_CS. The robustness results on synthetic data exhibit a strong correlation with those on real data, particularly in the case of the most distant shifts like 'snow' and 'night'. More details are provided as Supplementary Materials." + ], + "image_footnote": [], + "bbox": [ + 217, + 145, + 782, + 276 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2 Robustness Assessment with Synthetic Data", + "text_level": 1, + "bbox": [ + 215, + 455, + 627, + 469 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "With the pipeline outlined in Sec. 3.1, one can generate synthetic data to assess the robustness of pretrained segmenters in any unseen OOD domains through zero-shot prompting. To quantify robustness, we employ the traditional mean Intersection-over-Union (mIoU) score, measuring the correct overlap between semantic predictions and ground-truth masks. Given that our synthetic dataset comprises pairs of segmentation masks and synthetic images, one can straightforwardly derive synthetic scores for any pretrained segmenters. We here investigate whether synthetic performance can faithfully reflect the performance on real data in OOD domains under covariate shifts.", + "bbox": [ + 212, + 482, + 787, + 618 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Experimental Setups. We address weather shifts and geographical shifts, which are often encountered in autonomous driving. These covariate shifts are exhibited in two existing real datasets: the Adverse Conditions Dataset (ACDC) [33] and the Indian Driving Dataset (IDD) [39]. We utilize those real data to quantify the quality of our synthetic data and to validate its usefulness.", + "bbox": [ + 212, + 623, + 787, + 698 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Synthetic data are generated by conditioning on semantic masks from the Cityscapes validation set. To prompt ControlNet, we concatenate CLIP-interrogator's caption with a domain description following a simple template $[<\\text{caption}>$ , in $<\\text{domain}>$ ] where domain is either 'india', 'fog', 'rain', 'snow' or 'night'.", + "bbox": [ + 212, + 699, + 816, + 760 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For testing, we gather a collection of 40 publicly available segmenters trained only on Cityscapes, representative of different backbones, segmentation architectures, and sizes. The full list of models is in Supplementary Material.", + "bbox": [ + 212, + 760, + 787, + 805 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Results. In Fig. 3, we present our main results. Our primary metric is the Pearson Correlation Coefficient (PCC) between the mIoUs on testing data and on", + "bbox": [ + 212, + 809, + 787, + 839 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "Reliability in Semantic Segmentation: Can We Use Synthetic Data?", + "bbox": [ + 279, + 114, + 732, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "real-shift data from ACDC's splits or from IDD. The testing data can be either the Cityscapes validation set (CS) or our synthetic data (syn); the idea is to see which testing data –whether real CS or our synthetic one–correlates more with the real-shift data. Note that in the absence of OOD data, the Cityscapes validation set acts as the closest easily available proxy and is a reasonable predictor for OOD performance as pointed out by Jorge et al. [14].", + "bbox": [ + 212, + 146, + 787, + 238 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We organize our results based on increasing domain gaps relative to the Cityscapes domain. The domain gaps are quantified by the Pearson correlation between Cityscapes mIoUs and real-shift mIoUs, annotated as PCC_CS and visualized as red bars in the subplots of Fig. 3. Moving from left to right, i.e. with growing domain gaps, we observe a widening discrepancy between PCC_CS and PCC_Syn. Here, PCC_Syn represents the Pearson correlation between synthetic mIoUs and real-shift mIoUs. In domains with small gaps, PCC_CS and PCC_Syn are relatively comparable. However, in domains with more adverse shifts, such as 'snow' and 'night', PCC_Syn outperforms PCC_CS significantly, exceeding PCC_CS in 'night' by more than double.", + "bbox": [ + 212, + 239, + 789, + 391 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In Fig. 4, we analyze the results for the 'night' condition using the most robust models across different architectures, ranging from ConvNets to recent transformer networks. We use the Semantic-FPN score as the reference to normalize the scores of other models. This normalization aims to illustrate the relative improvement in robustness in terms of architecture. We rank the models from left to right based on their performance on real night data from the ACDC-night split. The consistently increasing trend of synthetic scores ( $\\rightarrow$ ) from left to right demonstrates a strong correlation with the ranking based on real scores. In contrast, Cityscapes scores ( $\\rightarrow$ ) are not indicative of night performance: a higher mIoU obtained on Cityscapes does not immediately translate into a higher mIoU at night.", + "bbox": [ + 212, + 395, + 789, + 561 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Since synthetic data can be generated in any desired quantity, a natural question arises: how many images are sufficient? In addressing this question, we conducted experiments and presented the results in Fig. 5. Our empirical finding suggests that $\\sim 500$ synthetic images are adequate for a stable and reliable assessment of robustness.", + "bbox": [ + 212, + 564, + 787, + 638 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Discussion. In their recent work on reliability in semantic segmentation, Jorge et al. [14] systematically quantified the robustness of segmenters on real-shift data; similarly to ours, they draw comparisons from ACDC and IDD datasets. One intriguing finding in this paper is that “[...] the larger the domain shift, the larger the improvement brought by more recent segmentation models”, hinting at a correlation between model robustness on in-domain data and covariate-shift data; that corresponds to the CS baseline we consider here. In our study, we delve deeper into this correlation, choosing to separately address different weather types instead of grouping them all together as done in [14]. In domains exhibiting small gaps to Cityscapes, such as IDD or ACDC-Fog, our conclusion aligns with [14]. However, as domain gaps increase, the discrepancy between Cityscapes mIoUs and real-shift mIoUs becomes more pronounced, resulting in poor PCC_CS scores. On the contrary, synthetic mIoUs and real-shift mIoUs", + "bbox": [ + 212, + 643, + 787, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "T.Loiseau et al.", + "bbox": [ + 271, + 114, + 377, + 127 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/b97ba5774136dbc05db5cee6f177ffa0f880fad6cc7df356d3e2cfa277f5d7fa.jpg", + "image_caption": [ + "Fig. 4: Day-night shift. Models are ranked from left to right by their robustness on real night data - ACDC-Night mIoUs are shown on top of model names. For each presented architecture, the most robust model on Cityscapes is tested; the SemanticFPN, DeeplabV3+, and PSPNet models have ResNet-101 as backbone. The SemanticFPN model (lowest mIoU on ACDC-Night) serves as the reference for computing the relative mIoUs. Blue bars or orange bars show the relative mIoUs when testing on our synthetic data (□) or testing on Cityscapes validation data (■). Cityscapes scores are not reliable for ranking models in the night domain. Synthetic scores exhibit a stronger correlation with real night scores, as evidenced by the more consistently increasing trend in the blue bars from left to right." + ], + "image_footnote": [], + "bbox": [ + 277, + 146, + 725, + 319 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1203cbdf388637c7316ca97cc2c62e0bf2c429d9976ef85d79920e6002ecf354.jpg", + "image_caption": [ + "Fig. 5: Pearson Correlation vs. # Synthetic Samples. Using more synthetic samples contributes to increased stability in the results. Empirical plots demonstrate that approximately 500 samples are sufficient for a stable correlation assessment." + ], + "image_footnote": [], + "bbox": [ + 209, + 489, + 493, + 623 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "exhibit a strong correlation across shifts. Our empirical study has validated that synthetic performance is a reliable indicator of model robustness in the presence of covariate shifts.", + "bbox": [ + 212, + 656, + 785, + 700 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "With synthetic data, we are able to confirm observations from the literature on real data, such as (i) the robustness of transformer and ConvNext backbones and that (ii) within an architecture family, robustness correlates with the number of parameters and the robustness of the backbone.", + "bbox": [ + 212, + 702, + 785, + 762 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Different generative models? The important ingredient in our pipeline is the generative model. In Tab. 1, we ablate by replacing the default SD1.5 model with other available ones: image-2-image GAN called TSIT [13], physics-based fog simulator [32], and a bigger SD variant called SDXL [28]. On 'night', 'rain', and 'snow', SD variants perform better than TSIT, while of note, TSIT was", + "bbox": [ + 212, + 763, + 785, + 839 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "Reliability in Semantic Segmentation: Can We Use Synthetic Data?", + "bbox": [ + 279, + 114, + 730, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/6a47315201b4cf839c9f22ff61a5f3dc94c169897f5ae4459a9cfdf326fb8418.jpg", + "table_caption": [], + "table_footnote": [ + "Table 1: Different generative models. Our pipeline achieves better performance (PCC with real-shifts) while does not require any OOD knowledge. Of note, it's very easy to apply to new unseen OOD domains using our pipeline, which remains challenging for GAN-based and physics-based models." + ], + "table_body": "
OOD expertise?OOD data?Night Rain SnowFogIndia
GAN-based TSIT [13]norequired0.830.840.81--
Physics-based Fog Sim. [32]requiredno---0.82-
Ours w/ SD1.5 (default)nono0.850.860.850.770.71
Ours w/ SDXL [28]nono0.840.900.820.890.93
", + "bbox": [ + 215, + 143, + 789, + 227 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "trained on real OOD data. The dedicated fog simulator performs very well, on par with our results using the larger model SDXL. Physics-based simulators are definitely valuable and this direction should be investigated further; however, such simulators require in-depth knowledge of OOD domains and hence are very difficult to design. Both GAN-based and physics-based approaches are limited in their scalability to many more OOD domains. Comparing SD1.5 vs. SDXL, we observe comparable results on challenging OOD domains while SDXL performs much better on 'fog' and 'india'; such results hint at the future potential of stronger and better generative models in further advancing virtual testing. Unfortunately, as SDXL is much more memory-demanding, we only limit our experiments with SDXL in this particular study.", + "bbox": [ + 212, + 325, + 784, + 489 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In the Supplementary Material, we report the FID scores, which measure the direct distance between synthetic and real distributions; we also extend a discussion on some limitations of our framework.", + "bbox": [ + 212, + 491, + 784, + 536 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "3.3 Confidence Calibration with Synthetic Data", + "text_level": 1, + "bbox": [ + 215, + 558, + 624, + 573 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Confidence calibration is a crucial aspect of deep networks, particularly when employed in safety-critical applications such as autonomous driving. Jorge et al. [14] highlighted a disconnection between model robustness and calibration, asserting that \"... despite the remarkable improvements in terms of robustness, recent models are not significantly better calibrated\". Therefore, it is essential to devise techniques and protocols for recalibrating data, particularly in domains exhibiting covariate shifts. Drawing inspiration from this, we explore the feasibility of using synthetic data to recalibrate pretrained segmenters.", + "bbox": [ + 212, + 583, + 784, + 703 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We perform temperature scaling using our synthetic data. Temperature scaling [6] is a well-established technique for calibrating pretrained models, typically conducted on a small validation set within the OOD domain. In our study, for each segmenter, we utilized the same sets of synthetic data generated in Sec. 3.2 to optimize temperature scaling factors, with one adjustment made for each covariate shift. For comparison, we replicate the process using real-shift data from ACDC and IDD.", + "bbox": [ + 212, + 704, + 784, + 808 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Fig. 6 reports the calibration improvement for the 40 pretrained segmenters using either real-shift data (■) or synthetic data (■). The subplots are arranged", + "bbox": [ + 215, + 809, + 785, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "T.Loiseau et al.", + "bbox": [ + 271, + 114, + 377, + 127 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c8db3259078009d1748afed9294c1244a65f5834f47cd41326d10c59c1b4fe3b.jpg", + "image_caption": [ + "Fig. 6: ECE improvement using synthetic data. The 40 segmenters are calibrated using either real-shift data or synthetic data. For each model, the relative ECE improvement $(\\%)$ over its non-calibrated version is computed, visualized by (synthetic shift) and (real shift). The subplots are ranked by model robustness on real-shift data (left-to-right) and by the increased domain shift (top-to-bottom). The titles of the subplots indicate the percentage of models that showed improvement with synthetic data." + ], + "image_footnote": [], + "bbox": [ + 235, + 143, + 485, + 342 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "in increasing domain gap order from top to bottom, with the segmenters ranked from left to right based on increasing robustness on real-shift data. The Expected Calibration Error (ECE) [24] quantifies the calibration results, with a lower ECE indicating a better-calibrated model. For better interpretation, we present the relative ECE improvement, computed as the percentage decrease in ECE after calibration compared to the original ECE without calibration. For example, a model with an ECE of 0.4 before calibration and an ECE of 0.2 after calibration will achieve a $(0.4 - 0.2) / 0.4 = 50\\%$ relative improvement.", + "bbox": [ + 212, + 369, + 787, + 491 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We observe promising calibration results when employing our synthetic data. While not as good as real-shift data, synthetic data achieves a promising success rate of $72.5\\%$ on IDD and exceeds $90\\%$ on the four ACDC shifts. Interestingly, in weather shifts, we empirically observe that more robust models derive greater benefits when calibrated using our synthetic data; the reverse is observed for 'europe-india' geographical shift. While with real-shift data, robustness and calibration are not well correlated [14], our results suggest that a potential correlation might exist between the two factors when using synthetic data. We note that calibration with temperature scaling does not always guarantee ECE improvement. Such phenomenon may happen even using real data, especially under domain shifts as explored in prior work [26]. In the Supplementary Material, we provide more technical details and results.", + "bbox": [ + 212, + 492, + 787, + 674 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "3.4 On Practical Applicability", + "text_level": 1, + "bbox": [ + 214, + 694, + 478, + 709 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "One significant advantage of our framework lies in its potential to address rare conditions simply through prompting. The practical applicability of generative testing is tremendous. Our results demonstrate promising signals; practitioners can begin assessing and ranking their pretrained models for a new, unseen OOD domain of interest without the need for real data collection. In practice, our proposed generative benchmarking can serve as the initial step in a full validation pipeline, helping filter out non-robust prototypes and thereby saving on total operational costs. Starting from complementing real-data validation, one", + "bbox": [ + 212, + 718, + 787, + 840 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "Reliability in Semantic Segmentation: Can We Use Synthetic Data?", + "bbox": [ + 279, + 113, + 730, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 114, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/81c42b9df908182830ed82081068e30e7d063f92d6bdeb6cfa9dd1a600cef908.jpg", + "image_caption": [ + "Fig. 7: Qualitative results. Examples of rare conditions generated for testing and predictions from different models. Results of the strong model like SegFormer-B5 are visibly better than the Semantic-FPN and MobileNetV3." + ], + "image_footnote": [], + "bbox": [ + 230, + 142, + 496, + 303 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "can envision a future where generative techniques mature to the point of fully replacing real-data validation. In Fig. 7, we visualize some synthetic images and model predictions under rare conditions, such as being flooded with water, having autumn leaves scattered across the road, or having a building on fire. We observe clear visual distinctions between weaker (Semantic-FPN and MobileNetV3) and stronger (SegF-B5) models, knowing that their Cityscapes scores do not differ significantly. More examples are provided in the Supplementary Material.", + "bbox": [ + 212, + 330, + 787, + 439 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4 Reliability Against OOD Objects", + "text_level": 1, + "bbox": [ + 215, + 459, + 571, + 478 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "We now address the reliability of segmentation models in the presence of Out-of-Distribution (OOD) objects. To begin, we explain our pipeline for inpainting random OOD objects into existing Cityscapes images. Following that, we demonstrate how one can utilize inpainted images for OOD detection assessment and for enhancing OOD detection.", + "bbox": [ + 212, + 491, + 787, + 566 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.1 Inpainting Anomaly Objects", + "text_level": 1, + "bbox": [ + 215, + 588, + 496, + 604 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "We inpaint random objects into Cityscapes images. To this end, we initially sample a location — a square box to which we inpaint the new object. We crop the box, upsample its content to match the preferred output size of the generative model, and inpaint an object guided by a text prompt. In this step, we leverage Stable Diffusion inpainting capabilities, obtaining high-definition square images of the desired object. This image is then resized and pasted back into the original image, creating a final high-definition synthetic image. To ensure compositional consistency, we employ two techniques: Firstly, we divide the cropped box into two regions by center cropping it again. We inpaint only the inner region, leaving the outer region untouched, similar to the approach in RePaint [21]. Secondly, after composing the final image, we address any remaining inconsistencies by applying a light noise over the entire picture and performing reverse diffusion again. Details and visualizations are provided in the Supplementary Material.", + "bbox": [ + 212, + 613, + 787, + 809 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "After inpainting, it is necessary to extract the mask corresponding to the new object. To achieve this, we begin with a high-definition square image and", + "bbox": [ + 214, + 810, + 785, + 840 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "T.Loiseau et al.", + "bbox": [ + 271, + 114, + 377, + 126 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/5fc2d2ed8532433e179ae126cbd58922d8f740ca92b92d23527f455654d6d1a9.jpg", + "image_caption": [ + "Fig. 8: Entropy Correlation. The top-left inset reports the Pearson correlations between real-OOD entropy vs. synthetic-OOD entropy, computed either on 'curated' (■) or all synthetic inpainted images (□). Evaluations are performed on the same model set used in Fig. 3, with similar markers." + ], + "image_footnote": [], + "bbox": [ + 233, + 142, + 495, + 280 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "apply the Grounded Segment Anything [15,20], prompted with the name of the object. This process yields a mask within the square image, which can then be repositioned in the full image.", + "bbox": [ + 212, + 308, + 784, + 354 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/a4cce56da056ea51dd830addc8064a23c0488559d3d89fbf17558e5055ca42fc.jpg", + "image_caption": [ + "Fig. 9: Correlation in OOD Detection. Each subplot scatters computed anomaly scores of segmenters on real OODs (y-axis) and on synthetic OODs (x-axis). The top row shows the four anomaly metrics utilized: $\\mathrm{FPR}_{95}$ , AUROC, $\\mathrm{AUPR}_{IN}$ , and $\\mathrm{AUPR}_{OUT}$ . The results are organized into two rows corresponding to two different confidence measures (i) Entropy and (ii) MaxLogit. In the top-left corner of each subplot, an inset plots Pearson correlations to real OOD for 'curated' (■) and 'all' (■) synthetic sets. Evaluations are performed on the same model set used in Fig. 3, with similar markers." + ], + "image_footnote": [], + "bbox": [ + 212, + 378, + 787, + 566 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Our end-to-end generation pipeline is fully automatic. Through qualitative assessment, we achieve a satisfactory success rate in terms of generation realism; some inpainted images are illustrated in Fig. 10 and much more in the Supplementary Material. However, this still leaves a few generations with artifacts, characterized by either unusual compositions or unrealistic details. We here question the criticality of realism in assessing OOD detection and, furthermore, in improving OOD detection. To this end, we construct two different sets: (i) all 23,040 images generated automatically and (ii) 656 curated images where", + "bbox": [ + 212, + 719, + 787, + 840 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "Reliability in Semantic Segmentation: Can We Use Synthetic Data?", + "bbox": [ + 279, + 113, + 730, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "we manually select the best images in terms of visual quality and realism; specifically, curators filter out strong color saturation differences or partial objects, e.g. animal heads. We note that the manual selection process for the curated set is not exhaustive and is constrained by our allocated resources; there are many more high-quality images in the 'all' set. In what follows, we present results using both curated and uncurated sets.", + "bbox": [ + 212, + 146, + 787, + 238 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "4.2 Assess OOD Detection", + "text_level": 1, + "bbox": [ + 215, + 258, + 452, + 272 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Experimental setup. To measure how the segmenters react to unseen OOD objects, we use standard anomaly detection metrics [41], which are False Positive Rate at $95\\%$ true positives $\\mathrm{(FPR_{95})}$ , Area Under ROC curve (AUROC), and Area Under Precision-Recall curve (AUPR). AUPR are declined into $\\mathrm{AUPR_{IN}}$ and $\\mathrm{AUPR_{OUT}}$ , which consider the in-distribution regions, respectively the out-of-distribution regions (the inpainted object), as positive regions to compute the Precision-Recall curves.", + "bbox": [ + 212, + 281, + 787, + 385 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "All segmenters in our study are not designed to produce confidence scores. We thus seek various techniques to derive confidence scores from pretrained models [9] and eventually narrow down the options to two measures: (i) Entropy of soft-probability predictions, and (ii) MaxLogit as the maximum logit value (before softmax) among the classes. While Entropy is the traditional measure of uncertainty, MaxLogit is a recent and surprising finding that has been proven to be more effective in estimating OOD confidence [9].", + "bbox": [ + 212, + 387, + 787, + 492 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Results. For quantitative comparison, we leverage the SegmentMeIfYouCan (SMIYC) dataset [4], a recent dataset for OOD detection. We resort to the RoadAnomaly21 split in SMIYC, due to similarity in object scales to our synthetic data. We analyze the correlation between the OOD scores obtained on RoadAnomaly21 and one using our synthetic inpainted data. Fig. 8 reports our first analysis on the entropy in the OOD areas, either real or generated. For each model, we compute the Pearson Correlation (PCC) between real-OD entropy and synthetic-OD entropy; the computation is done on both 'curated' and 'all' sets. We observe a very high entropy correlation between real- and synthetic-OD, reaching 0.94 PCC using both 'curated' and 'all' sets. In Fig. 11, we show a control experiment in which we inpaint in-domain object class 'car' into Cityscapes scenes, and we analyze models' responses to synthetic cars, real cars, and OOD objects.", + "bbox": [ + 212, + 493, + 787, + 688 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "We then analyze the correlation between real and synthetic anomaly metrics. Fig. 9 presents our primary findings and Fig. 10 illustrates some qualitative results. We observe a strong correlation with real scores when utilizing the 'curated' set for computing synthetic scores; the curated PCCs (■) are consistently around 0.8 across multiple metrics, irrespective of the two confidence measures. Although the correlations are somewhat weaker when using all uncurated synthetic data (■), such results remain acceptable, particularly when no effort is dedicated to curation.", + "bbox": [ + 212, + 689, + 787, + 808 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Our results validate the potential of utilizing realistic synthetic data, inpainted with anomaly objects, for assessing OOD detection. In OOD testing, it is quite", + "bbox": [ + 215, + 809, + 787, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "T.Loiseau et al.", + "bbox": [ + 271, + 114, + 377, + 127 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/ec7b8848aed786514c1e23c0dd46d52431ba1d74d59c85d2070199b27fa49b3d.jpg", + "image_caption": [ + "Fig. 10: Qualitative results. Confidence maps are visualized for the four exemplified models on real data (first row) and synthetic inpainted data (second and third rows). Hotter colors correspond to higher OOD likelihood. Ideally, results should exhibit hot colors in OOD areas and cold colors everywhere else. We observe a strong correlation in model reactions to real and synthetic OOD regions, particularly for more recent and robust models. We note that the real data in SMIYC also exhibit distributional shifts against Cityscapes; which already causes confusion to weak models in the background." + ], + "image_footnote": [], + "bbox": [ + 228, + 143, + 769, + 306 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/f53f0840b5bea16eba753eab91b21444b315b7763ef8fdb23367866aeeb73b43.jpg", + "image_caption": [ + "Fig. 11: In-domain vs OOD inpainted objects. Models' responses to synthetic cars are close to real cars, and far from synthetic OOD objects" + ], + "image_footnote": [], + "bbox": [ + 209, + 436, + 493, + 489 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "important to use high-quality synthetic inpainted data. Nonetheless, even non-curated synthetic data can offer an acceptable estimation of real performance with minimal curation costs.", + "bbox": [ + 215, + 517, + 787, + 561 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/10a4b2e23f3fdc81589460076374164ca045e6357cecd6c24271d504ce897054.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodAUROC (↑)AUPR (↑)FPR95 (↓)
RbA [25] Swin-B95.678.411.8
+ COCO [25]97.885.38.5
+ Ours (curated)97.284.98.1
+ Ours (all)97.384.88.2
RbA [25] Swin-L96.479.615.0
+ COCO [25]98.288.78.2
+ Ours (curated)97.288.07.9
+ Ours (all)98.188.68.3
", + "bbox": [ + 215, + 585, + 599, + 734 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Table 2: Improving OOD detection on real SMIYC benchmarking using our synthetic data. All results are obtained using the published code and default parameters.", + "bbox": [ + 611, + 585, + 787, + 683 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "4.3 Improve OOD Detection", + "text_level": 1, + "bbox": [ + 215, + 785, + 465, + 800 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "In this experiment, we investigate if synthetic inpainted data can be used to enhance a deep network's ability to detect OOD objects. To this end, we adopt", + "bbox": [ + 212, + 809, + 785, + 839 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "Reliability in Semantic Segmentation: Can We Use Synthetic Data?", + "bbox": [ + 279, + 114, + 730, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "the state-of-the-art RbA [25] approach for OOD detection and train RbA models on our synthetic data.", + "bbox": [ + 212, + 146, + 782, + 175 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "OOD metrics are computed on RoadAnomaly21 and reported in Tab. 2. The RbA models trained on our data significantly outperform the vanilla RbA model. We reach comparable performance to the RbA variants that leverage the external COCO dataset for augmentation. Notably, there are no clear differences between using 'curated' or 'all' sets. We conjecture that, unlike benchmarking, training for OOD detection does not demand a high degree of realism from synthetic data. This explains why the simple strategy of copy-pasting COCO objects [25] already proves effective. All results are consistent across the two addressed backbones. Fig. 10 illustrates a few qualitative results.", + "bbox": [ + 212, + 176, + 785, + 311 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Discussion. In our work, we take all available levers and study the extent to which synthetic data can be used for evaluation and specific training purposes. We acknowledge the fact that the models used in this work were trained on substantial amounts of data. Of note, we do not claim the efficacy of synthetic data in all aspects, particularly regarding the total amount of training data required. Our findings are limited to highlighting the significant potential of published generative models in the task of reliability assessment. Advances in efficient training of generative models may address concerns regarding data quantity but are beyond the scope of this work.", + "bbox": [ + 212, + 311, + 787, + 448 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Takeaways", + "text_level": 1, + "bbox": [ + 215, + 470, + 354, + 488 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In this work, we explore the potential of synthetic data in reliability assessment for semantic segmentation networks. We introduce two automatic zero-shot pipelines to generate data in OOD domains and to inpaint OOD objects for virtual reliability assessment. Our promising results encourage further collective investigations into this research problem, paving the way for synthetic system validation, especially in safety-critical applications. We summarize here our findings:", + "bbox": [ + 212, + 502, + 787, + 607 + ], + "page_idx": 13 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "$\\triangleright$ Reliability Under Covariate Shifts: synthetic data can help assess the relative robustness of models in real-life covariate shifts, especially when shifts to the training condition are significant. Synthetic data can well complement real data in system validation, helping reduce the total operational cost. Pretrained models can be calibrated using synthetic data to better estimate prediction confidence in any shifted domains.", + "$\\triangleright$ Reliability Against OOD Objects: synthetic data is useful in both OOD testing and OOD training; however, the demands on synthetic data quality differ in these two cases. In OOD testing, the best result estimations are obtained with the most realistically inpainted data, which may require a certain amount of curation time for qualitative assessment. The curation task is not time-demanding and can be done quickly with a reasonable budget. On the other hand, for OOD training, no curation is actually needed to achieve improvements." + ], + "bbox": [ + 225, + 616, + 782, + 825 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "T.Loiseau et al.", + "bbox": [ + 271, + 114, + 377, + 127 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 217, + 143, + 401, + 162 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "This work is supported by ELSA - European Lighthouse on Secure and Safe AI funded by the European Union under grant agreement No. 101070617. We thank the authors of RELIS [14] for providing us their pretrained checkpoints.", + "bbox": [ + 212, + 175, + 787, + 222 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 243, + 321, + 258 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. clip-interrogator. https://github.com/pharmapsychotic/clip-interrogator (2023)", + "2. Besnier, V., Jain, H., Bursuc, A., Cord, M., Pérez, P.: This dataset does not exist: training models from generated images. In: ICASSP (2020)", + "3. Blum, H., Sarlin, P.E., Nieto, J., Siegwart, R., Cadena, C.: The fishyscapes benchmark: Measuring blind spots in semantic segmentation (2021)", + "4. Chan, R., Lis, K., Uhlemeyer, S., Blum, H., Honari, S., Siegwart, R., Fua, P., Salzmann, M., Rottmann, M.: Segmentmeifyoucan: A benchmark for anomaly segmentation. In: NeurIPS (2021)", + "5. Geirhos, R., Narayanappa, K., Mitzkus, B., Thieringer, T., Bethge, M., Wichmann, F.A., Brendel, W.: Partial success in closing the gap between human and machine vision. NeurIPS (2021)", + "6. Guo, C., Pleiss, G., Sun, Y., Weinberger, K.Q.: On calibration of modern neural networks. In: ICML (2017)", + "7. Hariat, M., Laurent, O., Kazmierczak, R., Zhang, S., Bursuc, A., Yao, A., Franchi, G.: Learning to generate training datasets for robust semantic segmentation. In: WACV (2024)", + "8. He, R., Sun, S., Yu, X., Xue, C., Zhang, W., Torr, P., Bai, S., Qi, X.: Is synthetic data from generative models ready for image recognition? (2023)", + "9. Hendrycks, D., Basart, S., Mazeika, M., Zou, A., Kwon, J., Mostajabi, M., Steinhardt, J., Song, D.: Scaling out-of-distribution detection for real-world settings. In: ICML (2022)", + "0. Hendrycks, D., Basart, S., Mu, N., Kadavath, S., Wang, F., Dorundo, E., Desai, R., Zhu, T., Parajuli, S., Guo, M., et al.: The many faces of robustness: A critical analysis of out-of-distribution generalization. In: ICCV (2021)", + "1. Hendrycks, D., Dietterich, T.: Benchmarking neural network robustness to common corruptions and perturbations. In: ICLR (2019)", + "2. Hendrycks, D., Gimpel, K.: A baseline for detecting misclassified and out-of-distribution examples in neural networks. In: ICLR (2017)", + "3. Jiang, L., Zhang, C., Huang, M., Liu, C., Shi, J., Loy, C.C.: Tsit: A simple and versatile framework for image-to-image translation. In: ECCV (2020)", + "4. de Jorge, P., Volpi, R., Torr, P.H., Rogez, G.: Reliability in semantic segmentation: Are we on the right track? In: CVPR (2023)", + "5. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., Dollar, P., Girshick, R.: Segment anything. arXiv:2304.02643 (2023)", + "6. Koh, P.W., Sagawa, S., Marklund, H., Xie, S.M., Zhang, M., Balsubramani, A., Hu, W., Yasunaga, M., Phillips, R.L., Gao, I., et al.: Wilds: A benchmark of in-the-wild distribution shifts. In: ICLR (2021)", + "7. Le Moing, G., Vu, T.H., Jain, H., Pérez, P., Cord, M.: Semantic palette: Guiding scene generation with class proportions. In: CVPR (2021)" + ], + "bbox": [ + 225, + 273, + 785, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "Reliability in Semantic Segmentation: Can We Use Synthetic Data?", + "bbox": [ + 279, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 785, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "18. Li, D., Ling, H., Kim, S.W., Kreis, K., Fidler, S., Torralba, A.: Bigdatasetgan: Synthesizing imagenet with pixel-wise annotations. In: CVPR (2022)", + "19. Li, X., Chen, Y., Zhu, Y., Wang, S., Zhang, R., Xue, H.: Imagenet-e: Benchmarking neural network robustness via attribute editing. In: CVPR (2023)", + "20. Liu, S., Zeng, Z., Ren, T., Li, F., Zhang, H., Yang, J., Li, C., Yang, J., Su, H., Zhu, J., et al.: Grounding dino: Marrying dino with grounded pre-training for open-set object detection. arXiv preprint arXiv:2303.05499 (2023)", + "21. Lugmayr, A., Danelljan, M., Romero, A., Yu, F., Timofte, R., Gool, L.V.: Repaint: Inpainting using denoising diffusion probabilistic models. In: CVPR (2022)", + "22. Marathe, A., Ramanan, D., Walambe, R., Kotecha, K.: Wedge: A multi-weather autonomous driving dataset built from generative vision-language models. In: CVPRW (2023)", + "23. Miller, J.P., Taori, R., Raghunathan, A., Sagawa, S., Koh, P.W., Shankar, V., Liang, P., Carmon, Y., Schmidt, L.: Accuracy on the line: on the strong correlation between out-of-distribution and in-distribution generalization. In: ICLR (2021)", + "24. Naeini, M.P., Cooper, G., Hauskrecht, M.: Obtaining well calibrated probabilities using bayesian binning. In: AAAI (2015)", + "25. Nayal, N., Yavuz, M., Henriques, J.F., Güney, F.: Rba: Segmenting unknown regions rejected by all. In: ICCV (2023)", + "26. Ovadia, Y., Fertig, E., Ren, J., Nado, Z., Sculley, D., Nowozin, S., Dillon, J., Lakshminarayanan, B., Snoek, J.: Can you trust your model's uncertainty? evaluating predictive uncertainty under dataset shift. NeurIPS (2019)", + "27. Pinggera, P., Ramos, S., Gehrig, S., Franke, U., Rother, C., Mester, R.: Lost and found: detecting small road hazards for self-driving vehicles. In: IROS (2016)", + "28. Podell, D., English, Z., Lacey, K., Blattmann, A., Dockhorn, T., Müller, J., Penna, J., Rombach, R.: Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952 (2023)", + "29. Prabhu, V., Yenamandra, S., Chattopadhyay, P., Hoffman, J.: Lance: Stress-testing visual models by generating language-guided counterfactual images. In: NeurIPS (2023)", + "30. Recht, B., Roelofs, R., Schmidt, L., Shankar, V.: DoImagenet classifiers generalize toImagenet? In: ICLR (2019)", + "31. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: CVPR (2022)", + "32. Sakaridis, C., Dai, D., Van Gool, L.: Semantic foggy scene understanding with synthetic data. IJCV (2018)", + "33. Sakaridis, C., Dai, D., Van Gool, L.: Acdc: The adverse conditions dataset with correspondences for semantic driving scene understanding. In: ICCV (2021)", + "34. Sariyildiz, M.B., Alahari, K., Larlus, D., Kalantidis, Y.: Fake it till you make it: Learning transferable representations from synthetic imagenet clones. In: CVPR (2023)", + "35. Singh, A., Sarangmath, K., Chattopadhyay, P., Hoffman, J.: Benchmarking low-shot robustness to natural distribution shifts. In: ICCV (2023)", + "36. Taori, R., Dave, A., Shankar, V., Carlini, N., Recht, B., Schmidt, L.: Measuring robustness to natural distribution shifts in image classification. NeurIPS (2020)", + "37. Teney, D., Lin, Y., Oh, S.J., Abbasnejad, E.: Id and ood performance are sometimes inversely correlated on real-world datasets. arXiv (2022)", + "38. Tran, D., Liu, J., Dusenberry, M.W., Phan, D., Collier, M., Ren, J., Han, K., Wang, Z., Mariet, Z., Hu, H., et al.: Plex: Towards reliability using pretrained large model extensions. arXiv (2022)" + ], + "bbox": [ + 215, + 147, + 784, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "T.Loiseau et al.", + "bbox": [ + 271, + 114, + 377, + 127 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "39. Varma, G., Subramanian, A., Namboodiri, A., Chandraker, M., Jawahar, C.: Idd: A dataset for exploring problems of autonomous navigation in unconstrained environments. In: WACV (2019)", + "40. Wu, W., Zhao, Y., Shou, M.Z., Zhou, H., Shen, C.: Diffumask: Synthesizing images with pixel-level annotations for semantic segmentation using diffusion models. In: ICCV (2023)", + "41. Yang, J., Wang, P., Zou, D., Zhou, Z., Ding, K., Peng, W., Wang, H., Chen, G., Li, B., Sun, Y., et al.: Openood: Benchmarking generalized out-of-distribution detection. NeurIPS (2022)", + "42. Zhang, L., Rao, A., Agrawala, M.: Adding conditional control to text-to-image diffusion models. In: ICCV (2023)", + "43. Zhang, Y., Ling, H., Gao, J., Yin, K., Lafleche, J.F., Barriuso, A., Torralba, A., Fidler, S.: Datasetgan: Efficient labeled data factory with minimal human effort. In: CVPR (2021)" + ], + "bbox": [ + 215, + 146, + 785, + 340 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "Reliability in Semantic Segmentation: Can We Use Synthetic Data?", + "bbox": [ + 279, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 16 + } +] \ No newline at end of file diff --git a/2024/Reliability in Semantic Segmentation_ Can We Use Synthetic Data_/3d29ee45-5380-4719-adfd-255f20df0905_model.json b/2024/Reliability in Semantic Segmentation_ Can We Use Synthetic Data_/3d29ee45-5380-4719-adfd-255f20df0905_model.json new file mode 100644 index 0000000000000000000000000000000000000000..7f2a0fedd5ecb4dc67922d20cfd31dd82dc526d8 --- /dev/null +++ b/2024/Reliability in Semantic Segmentation_ Can We Use Synthetic Data_/3d29ee45-5380-4719-adfd-255f20df0905_model.json @@ -0,0 +1,2038 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.286, + 0.141, + 0.717, + 0.187 + ], + "angle": 0, + "content": "Reliability in Semantic Segmentation: Can We Use Synthetic Data?" + }, + { + "type": "text", + "bbox": [ + 0.24, + 0.212, + 0.765, + 0.245 + ], + "angle": 0, + "content": "Thibaut Loiseau\\(^{2\\dagger}\\), Tuan-Hung Vu\\(^{1}\\), Mickael Chen\\(^{1}\\), Patrick Pérez\\(^{3}\\), and Matthieu Cord\\(^{1,4}\\)" + }, + { + "type": "text", + "bbox": [ + 0.237, + 0.254, + 0.765, + 0.282 + ], + "angle": 0, + "content": "\\(^{1}\\) valeo.ai, Paris, France \n\\(^{2}\\) LIGM, Ecole des Ponts, Univ Gustave Eiffel, CNRS, Marne-la-Vallee, France" + }, + { + "type": "text", + "bbox": [ + 0.424, + 0.283, + 0.581, + 0.297 + ], + "angle": 0, + "content": "3 kyutai, Paris, France" + }, + { + "type": "text", + "bbox": [ + 0.379, + 0.297, + 0.625, + 0.31 + ], + "angle": 0, + "content": "4 Sorbonne Université, Paris, France" + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.344, + 0.744, + 0.581 + ], + "angle": 0, + "content": "Abstract. Assessing the robustness of perception models to covariate shifts and their ability to detect out-of-distribution (OOD) inputs is crucial for safety-critical applications such as autonomous vehicles. By nature of such applications, however, the relevant data is difficult to collect and annotate. In this paper, we show for the first time how synthetic data can be specifically generated to assess comprehensively the real-world reliability of semantic segmentation models. By fine-tuning Stable Diffusion [31] with only in-domain data, we perform zero-shot generation of visual scenes in OOD domains or inpainted with OOD objects. This synthetic data is employed to evaluate the robustness of pretrained segmenters, thereby offering insights into their performance when confronted with real edge cases. Through extensive experiments, we demonstrate a high correlation between the performance of models when evaluated on our synthetic OOD data and when evaluated on real OOD inputs, showing the relevance of such virtual testing. Furthermore, we demonstrate how our approach can be utilized to enhance the calibration and OOD detection capabilities of segmenters. Code and data are made public." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.625, + 0.377, + 0.642 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.655, + 0.788, + 0.791 + ], + "angle": 0, + "content": "Despite the rapid adoption of deep networks in safety-critical applications, reliability [26,36,38] has been an overlooked factor when designing and training these models. Recent efforts are geared toward enhancing model robustness under covariate shifts in data distributions [26, 30] and improving the model's ability to detect the unknown [10, 12, 41]. For open-world validation, in-domain data is no longer sufficient [16]; reliable and trustworthy systems demand more rigorous testing on diverse distributions, potentially exhibiting unknown objects. However, data collection campaigns can be quickly overwhelmed by the growing number of out-of-distribution (OOD) objects and conditions, sometimes extreme." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.792, + 0.788, + 0.822 + ], + "angle": 0, + "content": "In this paper, we propose to leverage pretrained generative models, \\( e.g. \\) Stable Diffusion 1.5 (SD), to alleviate the need for real OOD data. We use SD as a" + }, + { + "type": "page_footnote", + "bbox": [ + 0.232, + 0.826, + 0.533, + 0.841 + ], + "angle": 0, + "content": "† Work done during an internship at valeo.ai" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.378, + 0.127 + ], + "angle": 0, + "content": "T.Loiseau et al." + }, + { + "type": "image", + "bbox": [ + 0.242, + 0.146, + 0.764, + 0.301 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.322, + 0.788, + 0.379 + ], + "angle": 0, + "content": "Fig. 1: Assessing 40 pretrained segmenters under covariate shifts. Segmentation models under scrutiny were trained on Cityscapes train set only (in-domain data). They are evaluated on (i) Cityscapes validation set, (ii) real OOD data, and (iii) proposed synthetic data. We observe a strong correlation between results on (ii) and (iii)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.408, + 0.614, + 0.422 + ], + "angle": 0, + "content": "general validator, targeting multiple faces of reliability:" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.423, + 0.787, + 0.514 + ], + "angle": 0, + "content": "- First, we fine-tune SD on in-domain data to enable mask conditioning; in action, zero-shot prompting generates covariate shift images for testing. Fig. 1 illustrates the generation of synthetic data in OOD domains and its use for model evaluation. By assessing 40 pretrained segmenters under covariate shifts, our first contribution demonstrates how our generated synthetic OOD benchmark can act as a powerful proxy for the real OOD benchmark ACDC [33]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.514, + 0.788, + 0.603 + ], + "angle": 0, + "content": "- Second, we explore a similar strategy for OOD object detection assessment: we inpaint objects of unknown classes into in-domain data by injecting appropriate prompts during the zero-shot inpainting process. Inpainted images and OOD masks are used to benchmark the 40 models to see how well they can recognize OOD objects. Our results are strongly correlated to the real OOD benchmark [4]. Our high-quality synthetic data is featured in the official BRAVO benchmark." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.604, + 0.787, + 0.68 + ], + "angle": 0, + "content": "- Third, we demonstrate the usefulness of our two synthetic OOD benchmarks for hyperparameter tuning (here model calibration) and training. We calibrate the pretrained models to targeted OOD domains using our synthetic data and validate our strategy by comparing results to using real OOD data. We also train segmenters on inpainted data for OOD detection, obtaining competitive results." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.703, + 0.389, + 0.719 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.734, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Covariate Shifts. Modern machine learning models, notably deep networks, fall short in preserving their robustness and estimating their prediction confidence in the presence of covariate shifts [26,30]. Various benchmarks [5,11,16,35,36] have emerged to address the need for assessing models' reliability under different distributional shifts. Hendrycks et al. [11] propose a pioneering benchmark featuring data corrupted by various synthetic perturbations such as noise, blur, and brightness. Taori et al. [36] emphasize the importance of realistic shifts in" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.282, + 0.115, + 0.731, + 0.129 + ], + "angle": 0, + "content": "Reliability in Semantic Segmentation: Can We Use Synthetic Data?" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.147, + 0.786, + 0.493 + ], + "angle": 0, + "content": "reliability assessment; they highlight the disparity between natural and synthetic shifts, asserting that there is minimal to no robustness transfer from synthetic to natural distribution shifts. Sign et al. [35] study low-shot robustness to natural distribution shifts, highlighting the robustness properties of advanced architecture and pretraining strategies. In [14], de Jorge et al. extend beyond the standard classification setup to address the reliability problem in semantic segmentation, with findings largely aligned with previous works on classification; interestingly, they point to the disconnection between robustness and confidence calibration, urging more attention to calibration during segmenter design and training. On a related line, prior works study the connections between in-domain and out-of-domain robustness [14, 23, 37]; they suggest either positive, none, or even negative correlations between ID and OOD robustness, which indeed largely depends on the type of shifts. In line with [14], we focus on segmentation as the visual perception task of interest. Different from previous works, we advocate for the use of advanced generative models to generate realistic synthetic data for testing segmenters under arbitrary covariate shifts. Taori et al. [36] criticize the synthetic robustness benchmarks and advocate for using real-shift data; while we agree on the importance of realism in testing, we demonstrate that the rapid advancement of generative models now permits very meaningful virtual assessment. Our goal is to study whether synthetic data can be a superior choice compared to ID data in correlation studies against OOD robustness – referred to as real-shift robustness here to avoid confusion with OOD detection. Inspired by [14], we also calibrate models but using synthetic OOD data instead." + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.508, + 0.786, + 0.689 + ], + "angle": 0, + "content": "OOD Object Detection. In addition to robustness and calibration, the ability to detect the \"unknown\" is equally important to assess for trustworthy systems [10,12,41]. In semantic segmentation, several datasets are available for evaluating OOD detection on the road, namely LostAndFound [27], StreetHazards [9] (synthetic), BDD-Anomaly [9], Fishyscapes [3] (synthetic), and SegmentMeIfYouCan [4]. Encountering and capturing images of OOD objects in real-world scenarios, without deliberately placing them on the road, is quite uncommon. Effectively, existing datasets are limited in scale, both in terms of the number of images and the variety of object classes. In this work, we propose leveraging advanced zero-shot inpainting techniques to augment an existing segmentation dataset with the insertion of OOD objects; this enables the generation of highly realistic synthetic data for testing and training OOD detection." + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.704, + 0.786, + 0.84 + ], + "angle": 0, + "content": "Synthetic Data for Testing. Generative models have been exploited to create training data for image classification [2,8,34], object detection [22], or semantic segmentation [7,17,18,40,43]. Only recently, a few works have delved into the topic of generative testing data. Li et al. [19] exploit diffusion models to realistically edit images, controlling over various object attributes. This approach enables stress-testing models and understanding their sensitivity to different attributes. Using SD, LANCE [29] generates counterfactual images capable of challenging any given perception model. To the extent of our knowledge, no existing works have proposed to generate testing data for segmentation reliability." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.378, + 0.128 + ], + "angle": 0, + "content": "T.Loiseau et al." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.145, + 0.583, + 0.163 + ], + "angle": 0, + "content": "3 Reliability Under Covariate Shifts" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.178, + 0.788, + 0.313 + ], + "angle": 0, + "content": "In this section, we explore whether synthetic data can be used to assess the robustness of pretrained segmenters in the presence of covariate shifts to unseen OOD domains. We describe the data generation process and present the benchmarking results for a wide range of segmenters on our synthetic data. We demonstrate the validity of the approach using domains for which a real OOD dataset exists so that we can have access to a gold standard. We stress that we do not use the OOD data at any point in the method itself. This aspect is critical for the method to be applicable to benchmark robustness in the presence of extreme or hazardous conditions." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.337, + 0.593, + 0.352 + ], + "angle": 0, + "content": "3.1 Generating images in arbitrary domains" + }, + { + "type": "image", + "bbox": [ + 0.219, + 0.388, + 0.49, + 0.484 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.511, + 0.389, + 0.784, + 0.485 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.497, + 0.788, + 0.553 + ], + "angle": 0, + "content": "Fig. 2: Generating data with covariate shifts. Training (left) and Sampling (right) processes for producing the synthetic data with shifts. For training, only in-domain images and masks are used. For inference, we use the in-domain masks to generate OOD images. No real OOD data is required in the framework." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.584, + 0.788, + 0.719 + ], + "angle": 0, + "content": "Our goal here is to obtain pairs of images and semantic masks, with the images belonging to visual domains for which we lack data. To this end, we leverage a pretrained text-to-image Stable Diffusion 1.5 (SD) model [31], repurposed as a semantic-conditioned model called ControlNet [42]. ControlNet is trained solely on images and segmentation ground truths from the Cityscapes dataset; at train time, the text prompts are captions automatically extracted using CLIP-interrogator [1]. As a result, the model is able to perform mask-to-image generation of driving scenes while retaining the ability to steer the generation through text prompting of Stable Diffusion." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.788, + 0.84 + ], + "angle": 0, + "content": "To generate synthetic data, we prompt a trained ControlNet by the concatenation of OOD domain descriptions and CLIP-interrogator captions obtained from Cityscapes validation images. Also, segmentation masks in Cityscapes validation set are used to condition the generative process. Thanks to zero-shot prompting, the synthetic images are aligned with the semantic condition while displaying the visual properties of arbitrary OOD domains. Fig. 2 illustrates the training and generation steps. Detailed technical descriptions and more visualizations are in the Supplementary Material." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.281, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "Reliability in Semantic Segmentation: Can We Use Synthetic Data?" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.146, + 0.784, + 0.277 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.286, + 0.788, + 0.427 + ], + "angle": 0, + "content": "Fig. 3: Robustness correlation between real and synthetic covariate shifts across 40 pretrained segmenters. The tested models, see families in bottom legend, cover different architectures and sizes. (top) Pearson Correlation Coefficients of mIoUs between Cityscapes and real-shifts ('PCC_CS'), and between synthetic shifts and real ones ('PCC_Syn'). (bottom) Scatter plots of synthetic vs. real mIoUs along with the linear regression line accompanied by \\(95\\%\\) confidence intervals ('CI'). \\((a - e)\\) Five types of domain shifts from Cityscapes in-domain distribution, sorted by increasing gap as assessed by decreasing PCC_CS. The robustness results on synthetic data exhibit a strong correlation with those on real data, particularly in the case of the most distant shifts like 'snow' and 'night'. More details are provided as Supplementary Materials." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.456, + 0.628, + 0.47 + ], + "angle": 0, + "content": "3.2 Robustness Assessment with Synthetic Data" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.483, + 0.789, + 0.619 + ], + "angle": 0, + "content": "With the pipeline outlined in Sec. 3.1, one can generate synthetic data to assess the robustness of pretrained segmenters in any unseen OOD domains through zero-shot prompting. To quantify robustness, we employ the traditional mean Intersection-over-Union (mIoU) score, measuring the correct overlap between semantic predictions and ground-truth masks. Given that our synthetic dataset comprises pairs of segmentation masks and synthetic images, one can straightforwardly derive synthetic scores for any pretrained segmenters. We here investigate whether synthetic performance can faithfully reflect the performance on real data in OOD domains under covariate shifts." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.624, + 0.789, + 0.699 + ], + "angle": 0, + "content": "Experimental Setups. We address weather shifts and geographical shifts, which are often encountered in autonomous driving. These covariate shifts are exhibited in two existing real datasets: the Adverse Conditions Dataset (ACDC) [33] and the Indian Driving Dataset (IDD) [39]. We utilize those real data to quantify the quality of our synthetic data and to validate its usefulness." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.7, + 0.817, + 0.761 + ], + "angle": 0, + "content": "Synthetic data are generated by conditioning on semantic masks from the Cityscapes validation set. To prompt ControlNet, we concatenate CLIP-interrogator's caption with a domain description following a simple template \\([<\\text{caption}>\\), in \\(<\\text{domain}>\\)] where domain is either 'india', 'fog', 'rain', 'snow' or 'night'." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.761, + 0.788, + 0.806 + ], + "angle": 0, + "content": "For testing, we gather a collection of 40 publicly available segmenters trained only on Cityscapes, representative of different backbones, segmentation architectures, and sizes. The full list of models is in Supplementary Material." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Results. In Fig. 3, we present our main results. Our primary metric is the Pearson Correlation Coefficient (PCC) between the mIoUs on testing data and on" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.378, + 0.128 + ], + "angle": 0, + "content": "T.Loiseau et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.239 + ], + "angle": 0, + "content": "real-shift data from ACDC's splits or from IDD. The testing data can be either the Cityscapes validation set (CS) or our synthetic data (syn); the idea is to see which testing data –whether real CS or our synthetic one–correlates more with the real-shift data. Note that in the absence of OOD data, the Cityscapes validation set acts as the closest easily available proxy and is a reasonable predictor for OOD performance as pointed out by Jorge et al. [14]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.241, + 0.79, + 0.392 + ], + "angle": 0, + "content": "We organize our results based on increasing domain gaps relative to the Cityscapes domain. The domain gaps are quantified by the Pearson correlation between Cityscapes mIoUs and real-shift mIoUs, annotated as PCC_CS and visualized as red bars in the subplots of Fig. 3. Moving from left to right, i.e. with growing domain gaps, we observe a widening discrepancy between PCC_CS and PCC_Syn. Here, PCC_Syn represents the Pearson correlation between synthetic mIoUs and real-shift mIoUs. In domains with small gaps, PCC_CS and PCC_Syn are relatively comparable. However, in domains with more adverse shifts, such as 'snow' and 'night', PCC_Syn outperforms PCC_CS significantly, exceeding PCC_CS in 'night' by more than double." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.396, + 0.79, + 0.562 + ], + "angle": 0, + "content": "In Fig. 4, we analyze the results for the 'night' condition using the most robust models across different architectures, ranging from ConvNets to recent transformer networks. We use the Semantic-FPN score as the reference to normalize the scores of other models. This normalization aims to illustrate the relative improvement in robustness in terms of architecture. We rank the models from left to right based on their performance on real night data from the ACDC-night split. The consistently increasing trend of synthetic scores ( \\(\\rightarrow\\) ) from left to right demonstrates a strong correlation with the ranking based on real scores. In contrast, Cityscapes scores ( \\(\\rightarrow\\) ) are not indicative of night performance: a higher mIoU obtained on Cityscapes does not immediately translate into a higher mIoU at night." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.565, + 0.788, + 0.64 + ], + "angle": 0, + "content": "Since synthetic data can be generated in any desired quantity, a natural question arises: how many images are sufficient? In addressing this question, we conducted experiments and presented the results in Fig. 5. Our empirical finding suggests that \\(\\sim 500\\) synthetic images are adequate for a stable and reliable assessment of robustness." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.644, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Discussion. In their recent work on reliability in semantic segmentation, Jorge et al. [14] systematically quantified the robustness of segmenters on real-shift data; similarly to ours, they draw comparisons from ACDC and IDD datasets. One intriguing finding in this paper is that “[...] the larger the domain shift, the larger the improvement brought by more recent segmentation models”, hinting at a correlation between model robustness on in-domain data and covariate-shift data; that corresponds to the CS baseline we consider here. In our study, we delve deeper into this correlation, choosing to separately address different weather types instead of grouping them all together as done in [14]. In domains exhibiting small gaps to Cityscapes, such as IDD or ACDC-Fog, our conclusion aligns with [14]. However, as domain gaps increase, the discrepancy between Cityscapes mIoUs and real-shift mIoUs becomes more pronounced, resulting in poor PCC_CS scores. On the contrary, synthetic mIoUs and real-shift mIoUs" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.281, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Reliability in Semantic Segmentation: Can We Use Synthetic Data?" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "image", + "bbox": [ + 0.278, + 0.147, + 0.726, + 0.32 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.333, + 0.788, + 0.473 + ], + "angle": 0, + "content": "Fig. 4: Day-night shift. Models are ranked from left to right by their robustness on real night data - ACDC-Night mIoUs are shown on top of model names. For each presented architecture, the most robust model on Cityscapes is tested; the SemanticFPN, DeeplabV3+, and PSPNet models have ResNet-101 as backbone. The SemanticFPN model (lowest mIoU on ACDC-Night) serves as the reference for computing the relative mIoUs. Blue bars or orange bars show the relative mIoUs when testing on our synthetic data (□) or testing on Cityscapes validation data (■). Cityscapes scores are not reliable for ranking models in the night domain. Synthetic scores exhibit a stronger correlation with real night scores, as evidenced by the more consistently increasing trend in the blue bars from left to right." + }, + { + "type": "image", + "bbox": [ + 0.21, + 0.491, + 0.495, + 0.624 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.506, + 0.54, + 0.795, + 0.622 + ], + "angle": 0, + "content": "Fig. 5: Pearson Correlation vs. # Synthetic Samples. Using more synthetic samples contributes to increased stability in the results. Empirical plots demonstrate that approximately 500 samples are sufficient for a stable correlation assessment." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.657, + 0.787, + 0.701 + ], + "angle": 0, + "content": "exhibit a strong correlation across shifts. Our empirical study has validated that synthetic performance is a reliable indicator of model robustness in the presence of covariate shifts." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.703, + 0.787, + 0.763 + ], + "angle": 0, + "content": "With synthetic data, we are able to confirm observations from the literature on real data, such as (i) the robustness of transformer and ConvNext backbones and that (ii) within an architecture family, robustness correlates with the number of parameters and the robustness of the backbone." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Different generative models? The important ingredient in our pipeline is the generative model. In Tab. 1, we ablate by replacing the default SD1.5 model with other available ones: image-2-image GAN called TSIT [13], physics-based fog simulator [32], and a bigger SD variant called SDXL [28]. On 'night', 'rain', and 'snow', SD variants perform better than TSIT, while of note, TSIT was" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.378, + 0.128 + ], + "angle": 0, + "content": "T.Loiseau et al." + }, + { + "type": "table", + "bbox": [ + 0.217, + 0.145, + 0.79, + 0.228 + ], + "angle": 0, + "content": "
OOD expertise?OOD data?Night Rain SnowFogIndia
GAN-based TSIT [13]norequired0.830.840.81--
Physics-based Fog Sim. [32]requiredno---0.82-
Ours w/ SD1.5 (default)nono0.850.860.850.770.71
Ours w/ SDXL [28]nono0.840.900.820.890.93
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.216, + 0.229, + 0.787, + 0.283 + ], + "angle": 0, + "content": "Table 1: Different generative models. Our pipeline achieves better performance (PCC with real-shifts) while does not require any OOD knowledge. Of note, it's very easy to apply to new unseen OOD domains using our pipeline, which remains challenging for GAN-based and physics-based models." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.326, + 0.785, + 0.491 + ], + "angle": 0, + "content": "trained on real OOD data. The dedicated fog simulator performs very well, on par with our results using the larger model SDXL. Physics-based simulators are definitely valuable and this direction should be investigated further; however, such simulators require in-depth knowledge of OOD domains and hence are very difficult to design. Both GAN-based and physics-based approaches are limited in their scalability to many more OOD domains. Comparing SD1.5 vs. SDXL, we observe comparable results on challenging OOD domains while SDXL performs much better on 'fog' and 'india'; such results hint at the future potential of stronger and better generative models in further advancing virtual testing. Unfortunately, as SDXL is much more memory-demanding, we only limit our experiments with SDXL in this particular study." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.492, + 0.785, + 0.537 + ], + "angle": 0, + "content": "In the Supplementary Material, we report the FID scores, which measure the direct distance between synthetic and real distributions; we also extend a discussion on some limitations of our framework." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.559, + 0.625, + 0.574 + ], + "angle": 0, + "content": "3.3 Confidence Calibration with Synthetic Data" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.584, + 0.785, + 0.704 + ], + "angle": 0, + "content": "Confidence calibration is a crucial aspect of deep networks, particularly when employed in safety-critical applications such as autonomous driving. Jorge et al. [14] highlighted a disconnection between model robustness and calibration, asserting that \"... despite the remarkable improvements in terms of robustness, recent models are not significantly better calibrated\". Therefore, it is essential to devise techniques and protocols for recalibrating data, particularly in domains exhibiting covariate shifts. Drawing inspiration from this, we explore the feasibility of using synthetic data to recalibrate pretrained segmenters." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.705, + 0.785, + 0.809 + ], + "angle": 0, + "content": "We perform temperature scaling using our synthetic data. Temperature scaling [6] is a well-established technique for calibrating pretrained models, typically conducted on a small validation set within the OOD domain. In our study, for each segmenter, we utilized the same sets of synthetic data generated in Sec. 3.2 to optimize temperature scaling factors, with one adjustment made for each covariate shift. For comparison, we replicate the process using real-shift data from ACDC and IDD." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.81, + 0.787, + 0.841 + ], + "angle": 0, + "content": "Fig. 6 reports the calibration improvement for the 40 pretrained segmenters using either real-shift data (■) or synthetic data (■). The subplots are arranged" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.281, + 0.114, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Reliability in Semantic Segmentation: Can We Use Synthetic Data?" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "image", + "bbox": [ + 0.236, + 0.144, + 0.486, + 0.343 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.507, + 0.147, + 0.768, + 0.342 + ], + "angle": 0, + "content": "Fig. 6: ECE improvement using synthetic data. The 40 segmenters are calibrated using either real-shift data or synthetic data. For each model, the relative ECE improvement \\((\\%)\\) over its non-calibrated version is computed, visualized by (synthetic shift) and (real shift). The subplots are ranked by model robustness on real-shift data (left-to-right) and by the increased domain shift (top-to-bottom). The titles of the subplots indicate the percentage of models that showed improvement with synthetic data." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.371, + 0.788, + 0.492 + ], + "angle": 0, + "content": "in increasing domain gap order from top to bottom, with the segmenters ranked from left to right based on increasing robustness on real-shift data. The Expected Calibration Error (ECE) [24] quantifies the calibration results, with a lower ECE indicating a better-calibrated model. For better interpretation, we present the relative ECE improvement, computed as the percentage decrease in ECE after calibration compared to the original ECE without calibration. For example, a model with an ECE of 0.4 before calibration and an ECE of 0.2 after calibration will achieve a \\((0.4 - 0.2) / 0.4 = 50\\%\\) relative improvement." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.493, + 0.788, + 0.675 + ], + "angle": 0, + "content": "We observe promising calibration results when employing our synthetic data. While not as good as real-shift data, synthetic data achieves a promising success rate of \\(72.5\\%\\) on IDD and exceeds \\(90\\%\\) on the four ACDC shifts. Interestingly, in weather shifts, we empirically observe that more robust models derive greater benefits when calibrated using our synthetic data; the reverse is observed for 'europe-india' geographical shift. While with real-shift data, robustness and calibration are not well correlated [14], our results suggest that a potential correlation might exist between the two factors when using synthetic data. We note that calibration with temperature scaling does not always guarantee ECE improvement. Such phenomenon may happen even using real data, especially under domain shifts as explored in prior work [26]. In the Supplementary Material, we provide more technical details and results." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.695, + 0.48, + 0.71 + ], + "angle": 0, + "content": "3.4 On Practical Applicability" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.719, + 0.788, + 0.841 + ], + "angle": 0, + "content": "One significant advantage of our framework lies in its potential to address rare conditions simply through prompting. The practical applicability of generative testing is tremendous. Our results demonstrate promising signals; practitioners can begin assessing and ranking their pretrained models for a new, unseen OOD domain of interest without the need for real data collection. In practice, our proposed generative benchmarking can serve as the initial step in a full validation pipeline, helping filter out non-robust prototypes and thereby saving on total operational costs. Starting from complementing real-data validation, one" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.378, + 0.127 + ], + "angle": 0, + "content": "T.Loiseau et al." + }, + { + "type": "image", + "bbox": [ + 0.232, + 0.143, + 0.498, + 0.304 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.515, + 0.166, + 0.788, + 0.249 + ], + "angle": 0, + "content": "Fig. 7: Qualitative results. Examples of rare conditions generated for testing and predictions from different models. Results of the strong model like SegFormer-B5 are visibly better than the Semantic-FPN and MobileNetV3." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.332, + 0.788, + 0.44 + ], + "angle": 0, + "content": "can envision a future where generative techniques mature to the point of fully replacing real-data validation. In Fig. 7, we visualize some synthetic images and model predictions under rare conditions, such as being flooded with water, having autumn leaves scattered across the road, or having a building on fire. We observe clear visual distinctions between weaker (Semantic-FPN and MobileNetV3) and stronger (SegF-B5) models, knowing that their Cityscapes scores do not differ significantly. More examples are provided in the Supplementary Material." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.46, + 0.573, + 0.479 + ], + "angle": 0, + "content": "4 Reliability Against OOD Objects" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.492, + 0.788, + 0.568 + ], + "angle": 0, + "content": "We now address the reliability of segmentation models in the presence of Out-of-Distribution (OOD) objects. To begin, we explain our pipeline for inpainting random OOD objects into existing Cityscapes images. Following that, we demonstrate how one can utilize inpainted images for OOD detection assessment and for enhancing OOD detection." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.589, + 0.498, + 0.605 + ], + "angle": 0, + "content": "4.1 Inpainting Anomaly Objects" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.614, + 0.788, + 0.81 + ], + "angle": 0, + "content": "We inpaint random objects into Cityscapes images. To this end, we initially sample a location — a square box to which we inpaint the new object. We crop the box, upsample its content to match the preferred output size of the generative model, and inpaint an object guided by a text prompt. In this step, we leverage Stable Diffusion inpainting capabilities, obtaining high-definition square images of the desired object. This image is then resized and pasted back into the original image, creating a final high-definition synthetic image. To ensure compositional consistency, we employ two techniques: Firstly, we divide the cropped box into two regions by center cropping it again. We inpaint only the inner region, leaving the outer region untouched, similar to the approach in RePaint [21]. Secondly, after composing the final image, we address any remaining inconsistencies by applying a light noise over the entire picture and performing reverse diffusion again. Details and visualizations are provided in the Supplementary Material." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.811, + 0.787, + 0.841 + ], + "angle": 0, + "content": "After inpainting, it is necessary to extract the mask corresponding to the new object. To achieve this, we begin with a high-definition square image and" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.281, + 0.114, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Reliability in Semantic Segmentation: Can We Use Synthetic Data?" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "image", + "bbox": [ + 0.234, + 0.143, + 0.496, + 0.281 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.507, + 0.151, + 0.768, + 0.262 + ], + "angle": 0, + "content": "Fig. 8: Entropy Correlation. The top-left inset reports the Pearson correlations between real-OOD entropy vs. synthetic-OOD entropy, computed either on 'curated' (■) or all synthetic inpainted images (□). Evaluations are performed on the same model set used in Fig. 3, with similar markers." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.309, + 0.785, + 0.356 + ], + "angle": 0, + "content": "apply the Grounded Segment Anything [15,20], prompted with the name of the object. This process yields a mask within the square image, which can then be repositioned in the full image." + }, + { + "type": "image", + "bbox": [ + 0.214, + 0.379, + 0.788, + 0.567 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.576, + 0.788, + 0.687 + ], + "angle": 0, + "content": "Fig. 9: Correlation in OOD Detection. Each subplot scatters computed anomaly scores of segmenters on real OODs (y-axis) and on synthetic OODs (x-axis). The top row shows the four anomaly metrics utilized: \\(\\mathrm{FPR}_{95}\\), AUROC, \\(\\mathrm{AUPR}_{IN}\\), and \\(\\mathrm{AUPR}_{OUT}\\). The results are organized into two rows corresponding to two different confidence measures (i) Entropy and (ii) MaxLogit. In the top-left corner of each subplot, an inset plots Pearson correlations to real OOD for 'curated' (■) and 'all' (■) synthetic sets. Evaluations are performed on the same model set used in Fig. 3, with similar markers." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Our end-to-end generation pipeline is fully automatic. Through qualitative assessment, we achieve a satisfactory success rate in terms of generation realism; some inpainted images are illustrated in Fig. 10 and much more in the Supplementary Material. However, this still leaves a few generations with artifacts, characterized by either unusual compositions or unrealistic details. We here question the criticality of realism in assessing OOD detection and, furthermore, in improving OOD detection. To this end, we construct two different sets: (i) all 23,040 images generated automatically and (ii) 656 curated images where" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.378, + 0.128 + ], + "angle": 0, + "content": "T.Loiseau et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.239 + ], + "angle": 0, + "content": "we manually select the best images in terms of visual quality and realism; specifically, curators filter out strong color saturation differences or partial objects, e.g. animal heads. We note that the manual selection process for the curated set is not exhaustive and is constrained by our allocated resources; there are many more high-quality images in the 'all' set. In what follows, we present results using both curated and uncurated sets." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.259, + 0.453, + 0.273 + ], + "angle": 0, + "content": "4.2 Assess OOD Detection" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.282, + 0.788, + 0.386 + ], + "angle": 0, + "content": "Experimental setup. To measure how the segmenters react to unseen OOD objects, we use standard anomaly detection metrics [41], which are False Positive Rate at \\(95\\%\\) true positives \\(\\mathrm{(FPR_{95})}\\), Area Under ROC curve (AUROC), and Area Under Precision-Recall curve (AUPR). AUPR are declined into \\(\\mathrm{AUPR_{IN}}\\) and \\(\\mathrm{AUPR_{OUT}}\\), which consider the in-distribution regions, respectively the out-of-distribution regions (the inpainted object), as positive regions to compute the Precision-Recall curves." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.388, + 0.788, + 0.493 + ], + "angle": 0, + "content": "All segmenters in our study are not designed to produce confidence scores. We thus seek various techniques to derive confidence scores from pretrained models [9] and eventually narrow down the options to two measures: (i) Entropy of soft-probability predictions, and (ii) MaxLogit as the maximum logit value (before softmax) among the classes. While Entropy is the traditional measure of uncertainty, MaxLogit is a recent and surprising finding that has been proven to be more effective in estimating OOD confidence [9]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.494, + 0.788, + 0.689 + ], + "angle": 0, + "content": "Results. For quantitative comparison, we leverage the SegmentMeIfYouCan (SMIYC) dataset [4], a recent dataset for OOD detection. We resort to the RoadAnomaly21 split in SMIYC, due to similarity in object scales to our synthetic data. We analyze the correlation between the OOD scores obtained on RoadAnomaly21 and one using our synthetic inpainted data. Fig. 8 reports our first analysis on the entropy in the OOD areas, either real or generated. For each model, we compute the Pearson Correlation (PCC) between real-OD entropy and synthetic-OD entropy; the computation is done on both 'curated' and 'all' sets. We observe a very high entropy correlation between real- and synthetic-OD, reaching 0.94 PCC using both 'curated' and 'all' sets. In Fig. 11, we show a control experiment in which we inpaint in-domain object class 'car' into Cityscapes scenes, and we analyze models' responses to synthetic cars, real cars, and OOD objects." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.69, + 0.788, + 0.809 + ], + "angle": 0, + "content": "We then analyze the correlation between real and synthetic anomaly metrics. Fig. 9 presents our primary findings and Fig. 10 illustrates some qualitative results. We observe a strong correlation with real scores when utilizing the 'curated' set for computing synthetic scores; the curated PCCs (■) are consistently around 0.8 across multiple metrics, irrespective of the two confidence measures. Although the correlations are somewhat weaker when using all uncurated synthetic data (■), such results remain acceptable, particularly when no effort is dedicated to curation." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.81, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Our results validate the potential of utilizing realistic synthetic data, inpainted with anomaly objects, for assessing OOD detection. In OOD testing, it is quite" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.281, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Reliability in Semantic Segmentation: Can We Use Synthetic Data?" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "image", + "bbox": [ + 0.23, + 0.145, + 0.77, + 0.307 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.321, + 0.788, + 0.419 + ], + "angle": 0, + "content": "Fig. 10: Qualitative results. Confidence maps are visualized for the four exemplified models on real data (first row) and synthetic inpainted data (second and third rows). Hotter colors correspond to higher OOD likelihood. Ideally, results should exhibit hot colors in OOD areas and cold colors everywhere else. We observe a strong correlation in model reactions to real and synthetic OOD regions, particularly for more recent and robust models. We note that the real data in SMIYC also exhibit distributional shifts against Cityscapes; which already causes confusion to weak models in the background." + }, + { + "type": "image", + "bbox": [ + 0.21, + 0.437, + 0.495, + 0.49 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.506, + 0.433, + 0.795, + 0.489 + ], + "angle": 0, + "content": "Fig. 11: In-domain vs OOD inpainted objects. Models' responses to synthetic cars are close to real cars, and far from synthetic OOD objects" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.518, + 0.788, + 0.562 + ], + "angle": 0, + "content": "important to use high-quality synthetic inpainted data. Nonetheless, even non-curated synthetic data can offer an acceptable estimation of real performance with minimal curation costs." + }, + { + "type": "table", + "bbox": [ + 0.216, + 0.586, + 0.601, + 0.735 + ], + "angle": 0, + "content": "
MethodAUROC (↑)AUPR (↑)FPR95 (↓)
RbA [25] Swin-B95.678.411.8
+ COCO [25]97.885.38.5
+ Ours (curated)97.284.98.1
+ Ours (all)97.384.88.2
RbA [25] Swin-L96.479.615.0
+ COCO [25]98.288.78.2
+ Ours (curated)97.288.07.9
+ Ours (all)98.188.68.3
" + }, + { + "type": "text", + "bbox": [ + 0.612, + 0.587, + 0.788, + 0.684 + ], + "angle": 0, + "content": "Table 2: Improving OOD detection on real SMIYC benchmarking using our synthetic data. All results are obtained using the published code and default parameters." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.786, + 0.466, + 0.801 + ], + "angle": 0, + "content": "4.3 Improve OOD Detection" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.84 + ], + "angle": 0, + "content": "In this experiment, we investigate if synthetic inpainted data can be used to enhance a deep network's ability to detect OOD objects. To this end, we adopt" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.378, + 0.128 + ], + "angle": 0, + "content": "T.Loiseau et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.176 + ], + "angle": 0, + "content": "the state-of-the-art RbA [25] approach for OOD detection and train RbA models on our synthetic data." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.177, + 0.787, + 0.312 + ], + "angle": 0, + "content": "OOD metrics are computed on RoadAnomaly21 and reported in Tab. 2. The RbA models trained on our data significantly outperform the vanilla RbA model. We reach comparable performance to the RbA variants that leverage the external COCO dataset for augmentation. Notably, there are no clear differences between using 'curated' or 'all' sets. We conjecture that, unlike benchmarking, training for OOD detection does not demand a high degree of realism from synthetic data. This explains why the simple strategy of copy-pasting COCO objects [25] already proves effective. All results are consistent across the two addressed backbones. Fig. 10 illustrates a few qualitative results." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.313, + 0.788, + 0.449 + ], + "angle": 0, + "content": "Discussion. In our work, we take all available levers and study the extent to which synthetic data can be used for evaluation and specific training purposes. We acknowledge the fact that the models used in this work were trained on substantial amounts of data. Of note, we do not claim the efficacy of synthetic data in all aspects, particularly regarding the total amount of training data required. Our findings are limited to highlighting the significant potential of published generative models in the task of reliability assessment. Advances in efficient training of generative models may address concerns regarding data quantity but are beyond the scope of this work." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.471, + 0.356, + 0.489 + ], + "angle": 0, + "content": "5 Takeaways" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.503, + 0.788, + 0.608 + ], + "angle": 0, + "content": "In this work, we explore the potential of synthetic data in reliability assessment for semantic segmentation networks. We introduce two automatic zero-shot pipelines to generate data in OOD domains and to inpaint OOD objects for virtual reliability assessment. Our promising results encourage further collective investigations into this research problem, paving the way for synthetic system validation, especially in safety-critical applications. We summarize here our findings:" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.617, + 0.784, + 0.705 + ], + "angle": 0, + "content": "\\(\\triangleright\\) Reliability Under Covariate Shifts: synthetic data can help assess the relative robustness of models in real-life covariate shifts, especially when shifts to the training condition are significant. Synthetic data can well complement real data in system validation, helping reduce the total operational cost. Pretrained models can be calibrated using synthetic data to better estimate prediction confidence in any shifted domains." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.708, + 0.784, + 0.826 + ], + "angle": 0, + "content": "\\(\\triangleright\\) Reliability Against OOD Objects: synthetic data is useful in both OOD testing and OOD training; however, the demands on synthetic data quality differ in these two cases. In OOD testing, the best result estimations are obtained with the most realistically inpainted data, which may require a certain amount of curation time for qualitative assessment. The curation task is not time-demanding and can be done quickly with a reasonable budget. On the other hand, for OOD training, no curation is actually needed to achieve improvements." + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.617, + 0.784, + 0.826 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.281, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Reliability in Semantic Segmentation: Can We Use Synthetic Data?" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.403, + 0.163 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.176, + 0.788, + 0.223 + ], + "angle": 0, + "content": "This work is supported by ELSA - European Lighthouse on Secure and Safe AI funded by the European Union under grant agreement No. 101070617. We thank the authors of RELIS [14] for providing us their pretrained checkpoints." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.244, + 0.323, + 0.26 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.274, + 0.787, + 0.302 + ], + "angle": 0, + "content": "1. clip-interrogator. https://github.com/pharmapsychotic/clip-interrogator (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.304, + 0.787, + 0.331 + ], + "angle": 0, + "content": "2. Besnier, V., Jain, H., Bursuc, A., Cord, M., Pérez, P.: This dataset does not exist: training models from generated images. In: ICASSP (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.332, + 0.786, + 0.358 + ], + "angle": 0, + "content": "3. Blum, H., Sarlin, P.E., Nieto, J., Siegwart, R., Cadena, C.: The fishyscapes benchmark: Measuring blind spots in semantic segmentation (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.359, + 0.786, + 0.399 + ], + "angle": 0, + "content": "4. Chan, R., Lis, K., Uhlemeyer, S., Blum, H., Honari, S., Siegwart, R., Fua, P., Salzmann, M., Rottmann, M.: Segmentmeifyoucan: A benchmark for anomaly segmentation. In: NeurIPS (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.4, + 0.787, + 0.44 + ], + "angle": 0, + "content": "5. Geirhos, R., Narayanappa, K., Mitzkus, B., Thieringer, T., Bethge, M., Wichmann, F.A., Brendel, W.: Partial success in closing the gap between human and machine vision. NeurIPS (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.441, + 0.786, + 0.469 + ], + "angle": 0, + "content": "6. Guo, C., Pleiss, G., Sun, Y., Weinberger, K.Q.: On calibration of modern neural networks. In: ICML (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.47, + 0.787, + 0.509 + ], + "angle": 0, + "content": "7. Hariat, M., Laurent, O., Kazmierczak, R., Zhang, S., Bursuc, A., Yao, A., Franchi, G.: Learning to generate training datasets for robust semantic segmentation. In: WACV (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.511, + 0.787, + 0.537 + ], + "angle": 0, + "content": "8. He, R., Sun, S., Yu, X., Xue, C., Zhang, W., Torr, P., Bai, S., Qi, X.: Is synthetic data from generative models ready for image recognition? (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.538, + 0.787, + 0.578 + ], + "angle": 0, + "content": "9. Hendrycks, D., Basart, S., Mazeika, M., Zou, A., Kwon, J., Mostajabi, M., Steinhardt, J., Song, D.: Scaling out-of-distribution detection for real-world settings. In: ICML (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.58, + 0.787, + 0.62 + ], + "angle": 0, + "content": "0. Hendrycks, D., Basart, S., Mu, N., Kadavath, S., Wang, F., Dorundo, E., Desai, R., Zhu, T., Parajuli, S., Guo, M., et al.: The many faces of robustness: A critical analysis of out-of-distribution generalization. In: ICCV (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.621, + 0.787, + 0.647 + ], + "angle": 0, + "content": "1. Hendrycks, D., Dietterich, T.: Benchmarking neural network robustness to common corruptions and perturbations. In: ICLR (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.649, + 0.787, + 0.674 + ], + "angle": 0, + "content": "2. Hendrycks, D., Gimpel, K.: A baseline for detecting misclassified and out-of-distribution examples in neural networks. In: ICLR (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.676, + 0.787, + 0.703 + ], + "angle": 0, + "content": "3. Jiang, L., Zhang, C., Huang, M., Liu, C., Shi, J., Loy, C.C.: Tsit: A simple and versatile framework for image-to-image translation. In: ECCV (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.704, + 0.787, + 0.73 + ], + "angle": 0, + "content": "4. de Jorge, P., Volpi, R., Torr, P.H., Rogez, G.: Reliability in semantic segmentation: Are we on the right track? In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.731, + 0.787, + 0.771 + ], + "angle": 0, + "content": "5. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., Dollar, P., Girshick, R.: Segment anything. arXiv:2304.02643 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.772, + 0.787, + 0.812 + ], + "angle": 0, + "content": "6. Koh, P.W., Sagawa, S., Marklund, H., Xie, S.M., Zhang, M., Balsubramani, A., Hu, W., Yasunaga, M., Phillips, R.L., Gao, I., et al.: Wilds: A benchmark of in-the-wild distribution shifts. In: ICLR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.813, + 0.787, + 0.84 + ], + "angle": 0, + "content": "7. Le Moing, G., Vu, T.H., Jain, H., Pérez, P., Cord, M.: Semantic palette: Guiding scene generation with class proportions. In: CVPR (2021)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.274, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.378, + 0.128 + ], + "angle": 0, + "content": "T.Loiseau et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.148, + 0.785, + 0.175 + ], + "angle": 0, + "content": "18. Li, D., Ling, H., Kim, S.W., Kreis, K., Fidler, S., Torralba, A.: Bigdatasetgan: Synthesizing imagenet with pixel-wise annotations. In: CVPR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.177, + 0.785, + 0.203 + ], + "angle": 0, + "content": "19. Li, X., Chen, Y., Zhu, Y., Wang, S., Zhang, R., Xue, H.: Imagenet-e: Benchmarking neural network robustness via attribute editing. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.204, + 0.785, + 0.244 + ], + "angle": 0, + "content": "20. Liu, S., Zeng, Z., Ren, T., Li, F., Zhang, H., Yang, J., Li, C., Yang, J., Su, H., Zhu, J., et al.: Grounding dino: Marrying dino with grounded pre-training for open-set object detection. arXiv preprint arXiv:2303.05499 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.245, + 0.785, + 0.272 + ], + "angle": 0, + "content": "21. Lugmayr, A., Danelljan, M., Romero, A., Yu, F., Timofte, R., Gool, L.V.: Repaint: Inpainting using denoising diffusion probabilistic models. In: CVPR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.273, + 0.785, + 0.313 + ], + "angle": 0, + "content": "22. Marathe, A., Ramanan, D., Walambe, R., Kotecha, K.: Wedge: A multi-weather autonomous driving dataset built from generative vision-language models. In: CVPRW (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.315, + 0.785, + 0.355 + ], + "angle": 0, + "content": "23. Miller, J.P., Taori, R., Raghunathan, A., Sagawa, S., Koh, P.W., Shankar, V., Liang, P., Carmon, Y., Schmidt, L.: Accuracy on the line: on the strong correlation between out-of-distribution and in-distribution generalization. In: ICLR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.357, + 0.785, + 0.383 + ], + "angle": 0, + "content": "24. Naeini, M.P., Cooper, G., Hauskrecht, M.: Obtaining well calibrated probabilities using bayesian binning. In: AAAI (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.384, + 0.785, + 0.411 + ], + "angle": 0, + "content": "25. Nayal, N., Yavuz, M., Henriques, J.F., Güney, F.: Rba: Segmenting unknown regions rejected by all. In: ICCV (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.412, + 0.785, + 0.452 + ], + "angle": 0, + "content": "26. Ovadia, Y., Fertig, E., Ren, J., Nado, Z., Sculley, D., Nowozin, S., Dillon, J., Lakshminarayanan, B., Snoek, J.: Can you trust your model's uncertainty? evaluating predictive uncertainty under dataset shift. NeurIPS (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.453, + 0.785, + 0.479 + ], + "angle": 0, + "content": "27. Pinggera, P., Ramos, S., Gehrig, S., Franke, U., Rother, C., Mester, R.: Lost and found: detecting small road hazards for self-driving vehicles. In: IROS (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.481, + 0.785, + 0.521 + ], + "angle": 0, + "content": "28. Podell, D., English, Z., Lacey, K., Blattmann, A., Dockhorn, T., Müller, J., Penna, J., Rombach, R.: Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.522, + 0.785, + 0.563 + ], + "angle": 0, + "content": "29. Prabhu, V., Yenamandra, S., Chattopadhyay, P., Hoffman, J.: Lance: Stress-testing visual models by generating language-guided counterfactual images. In: NeurIPS (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.564, + 0.785, + 0.59 + ], + "angle": 0, + "content": "30. Recht, B., Roelofs, R., Schmidt, L., Shankar, V.: DoImagenet classifiers generalize toImagenet? In: ICLR (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.592, + 0.785, + 0.618 + ], + "angle": 0, + "content": "31. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: CVPR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.619, + 0.785, + 0.646 + ], + "angle": 0, + "content": "32. Sakaridis, C., Dai, D., Van Gool, L.: Semantic foggy scene understanding with synthetic data. IJCV (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.647, + 0.785, + 0.673 + ], + "angle": 0, + "content": "33. Sakaridis, C., Dai, D., Van Gool, L.: Acdc: The adverse conditions dataset with correspondences for semantic driving scene understanding. In: ICCV (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.675, + 0.785, + 0.715 + ], + "angle": 0, + "content": "34. Sariyildiz, M.B., Alahari, K., Larlus, D., Kalantidis, Y.: Fake it till you make it: Learning transferable representations from synthetic imagenet clones. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.716, + 0.785, + 0.743 + ], + "angle": 0, + "content": "35. Singh, A., Sarangmath, K., Chattopadhyay, P., Hoffman, J.: Benchmarking low-shot robustness to natural distribution shifts. In: ICCV (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.744, + 0.785, + 0.771 + ], + "angle": 0, + "content": "36. Taori, R., Dave, A., Shankar, V., Carlini, N., Recht, B., Schmidt, L.: Measuring robustness to natural distribution shifts in image classification. NeurIPS (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.772, + 0.785, + 0.798 + ], + "angle": 0, + "content": "37. Teney, D., Lin, Y., Oh, S.J., Abbasnejad, E.: Id and ood performance are sometimes inversely correlated on real-world datasets. arXiv (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.799, + 0.785, + 0.84 + ], + "angle": 0, + "content": "38. Tran, D., Liu, J., Dusenberry, M.W., Phan, D., Collier, M., Ren, J., Han, K., Wang, Z., Mariet, Z., Hu, H., et al.: Plex: Towards reliability using pretrained large model extensions. arXiv (2022)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.148, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.281, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Reliability in Semantic Segmentation: Can We Use Synthetic Data?" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.189 + ], + "angle": 0, + "content": "39. Varma, G., Subramanian, A., Namboodiri, A., Chandraker, M., Jawahar, C.: Idd: A dataset for exploring problems of autonomous navigation in unconstrained environments. In: WACV (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.19, + 0.787, + 0.231 + ], + "angle": 0, + "content": "40. Wu, W., Zhao, Y., Shou, M.Z., Zhou, H., Shen, C.: Diffumask: Synthesizing images with pixel-level annotations for semantic segmentation using diffusion models. In: ICCV (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.232, + 0.785, + 0.272 + ], + "angle": 0, + "content": "41. Yang, J., Wang, P., Zou, D., Zhou, Z., Ding, K., Peng, W., Wang, H., Chen, G., Li, B., Sun, Y., et al.: Openood: Benchmarking generalized out-of-distribution detection. NeurIPS (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.273, + 0.785, + 0.3 + ], + "angle": 0, + "content": "42. Zhang, L., Rao, A., Agrawala, M.: Adding conditional control to text-to-image diffusion models. In: ICCV (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.3, + 0.785, + 0.342 + ], + "angle": 0, + "content": "43. Zhang, Y., Ling, H., Gao, J., Yin, K., Lafleche, J.F., Barriuso, A., Torralba, A., Fidler, S.: Datasetgan: Efficient labeled data factory with minimal human effort. In: CVPR (2021)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.342 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/Reliability in Semantic Segmentation_ Can We Use Synthetic Data_/3d29ee45-5380-4719-adfd-255f20df0905_origin.pdf b/2024/Reliability in Semantic Segmentation_ Can We Use Synthetic Data_/3d29ee45-5380-4719-adfd-255f20df0905_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..475747b655f654fcbdd0a14c06958381d3c096fb --- /dev/null +++ b/2024/Reliability in Semantic Segmentation_ Can We Use Synthetic Data_/3d29ee45-5380-4719-adfd-255f20df0905_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:091a09c0311206f8e071ca1ca27c1a0578698a9abf1f213c9cdb59159e85103a +size 4556063 diff --git a/2024/Reliability in Semantic Segmentation_ Can We Use Synthetic Data_/full.md b/2024/Reliability in Semantic Segmentation_ Can We Use Synthetic Data_/full.md new file mode 100644 index 0000000000000000000000000000000000000000..a596d657fc8660ece87d14f64164eb60e4d5b30b --- /dev/null +++ b/2024/Reliability in Semantic Segmentation_ Can We Use Synthetic Data_/full.md @@ -0,0 +1,238 @@ +# Reliability in Semantic Segmentation: Can We Use Synthetic Data? + +Thibaut Loiseau $^{2\dagger}$ , Tuan-Hung Vu $^{1}$ , Mickael Chen $^{1}$ , Patrick Pérez $^{3}$ , and Matthieu Cord $^{1,4}$ + +$^{1}$ valeo.ai, Paris, France + $^{2}$ LIGM, Ecole des Ponts, Univ Gustave Eiffel, CNRS, Marne-la-Vallee, France + +3 kyutai, Paris, France + +4 Sorbonne Université, Paris, France + +Abstract. Assessing the robustness of perception models to covariate shifts and their ability to detect out-of-distribution (OOD) inputs is crucial for safety-critical applications such as autonomous vehicles. By nature of such applications, however, the relevant data is difficult to collect and annotate. In this paper, we show for the first time how synthetic data can be specifically generated to assess comprehensively the real-world reliability of semantic segmentation models. By fine-tuning Stable Diffusion [31] with only in-domain data, we perform zero-shot generation of visual scenes in OOD domains or inpainted with OOD objects. This synthetic data is employed to evaluate the robustness of pretrained segmenters, thereby offering insights into their performance when confronted with real edge cases. Through extensive experiments, we demonstrate a high correlation between the performance of models when evaluated on our synthetic OOD data and when evaluated on real OOD inputs, showing the relevance of such virtual testing. Furthermore, we demonstrate how our approach can be utilized to enhance the calibration and OOD detection capabilities of segmenters. Code and data are made public. + +# 1 Introduction + +Despite the rapid adoption of deep networks in safety-critical applications, reliability [26,36,38] has been an overlooked factor when designing and training these models. Recent efforts are geared toward enhancing model robustness under covariate shifts in data distributions [26, 30] and improving the model's ability to detect the unknown [10, 12, 41]. For open-world validation, in-domain data is no longer sufficient [16]; reliable and trustworthy systems demand more rigorous testing on diverse distributions, potentially exhibiting unknown objects. However, data collection campaigns can be quickly overwhelmed by the growing number of out-of-distribution (OOD) objects and conditions, sometimes extreme. + +In this paper, we propose to leverage pretrained generative models, $e.g.$ Stable Diffusion 1.5 (SD), to alleviate the need for real OOD data. We use SD as a + +![](images/7d42448a0ecaa80a6205718151f8dd2cc847c2bd0d767f1bd6866fa951c07912.jpg) +Fig. 1: Assessing 40 pretrained segmenters under covariate shifts. Segmentation models under scrutiny were trained on Cityscapes train set only (in-domain data). They are evaluated on (i) Cityscapes validation set, (ii) real OOD data, and (iii) proposed synthetic data. We observe a strong correlation between results on (ii) and (iii). + +general validator, targeting multiple faces of reliability: + +- First, we fine-tune SD on in-domain data to enable mask conditioning; in action, zero-shot prompting generates covariate shift images for testing. Fig. 1 illustrates the generation of synthetic data in OOD domains and its use for model evaluation. By assessing 40 pretrained segmenters under covariate shifts, our first contribution demonstrates how our generated synthetic OOD benchmark can act as a powerful proxy for the real OOD benchmark ACDC [33]. + +- Second, we explore a similar strategy for OOD object detection assessment: we inpaint objects of unknown classes into in-domain data by injecting appropriate prompts during the zero-shot inpainting process. Inpainted images and OOD masks are used to benchmark the 40 models to see how well they can recognize OOD objects. Our results are strongly correlated to the real OOD benchmark [4]. Our high-quality synthetic data is featured in the official BRAVO benchmark. + +- Third, we demonstrate the usefulness of our two synthetic OOD benchmarks for hyperparameter tuning (here model calibration) and training. We calibrate the pretrained models to targeted OOD domains using our synthetic data and validate our strategy by comparing results to using real OOD data. We also train segmenters on inpainted data for OOD detection, obtaining competitive results. + +# 2 Related Work + +Covariate Shifts. Modern machine learning models, notably deep networks, fall short in preserving their robustness and estimating their prediction confidence in the presence of covariate shifts [26,30]. Various benchmarks [5,11,16,35,36] have emerged to address the need for assessing models' reliability under different distributional shifts. Hendrycks et al. [11] propose a pioneering benchmark featuring data corrupted by various synthetic perturbations such as noise, blur, and brightness. Taori et al. [36] emphasize the importance of realistic shifts in + +reliability assessment; they highlight the disparity between natural and synthetic shifts, asserting that there is minimal to no robustness transfer from synthetic to natural distribution shifts. Sign et al. [35] study low-shot robustness to natural distribution shifts, highlighting the robustness properties of advanced architecture and pretraining strategies. In [14], de Jorge et al. extend beyond the standard classification setup to address the reliability problem in semantic segmentation, with findings largely aligned with previous works on classification; interestingly, they point to the disconnection between robustness and confidence calibration, urging more attention to calibration during segmenter design and training. On a related line, prior works study the connections between in-domain and out-of-domain robustness [14, 23, 37]; they suggest either positive, none, or even negative correlations between ID and OOD robustness, which indeed largely depends on the type of shifts. In line with [14], we focus on segmentation as the visual perception task of interest. Different from previous works, we advocate for the use of advanced generative models to generate realistic synthetic data for testing segmenters under arbitrary covariate shifts. Taori et al. [36] criticize the synthetic robustness benchmarks and advocate for using real-shift data; while we agree on the importance of realism in testing, we demonstrate that the rapid advancement of generative models now permits very meaningful virtual assessment. Our goal is to study whether synthetic data can be a superior choice compared to ID data in correlation studies against OOD robustness – referred to as real-shift robustness here to avoid confusion with OOD detection. Inspired by [14], we also calibrate models but using synthetic OOD data instead. + +OOD Object Detection. In addition to robustness and calibration, the ability to detect the "unknown" is equally important to assess for trustworthy systems [10,12,41]. In semantic segmentation, several datasets are available for evaluating OOD detection on the road, namely LostAndFound [27], StreetHazards [9] (synthetic), BDD-Anomaly [9], Fishyscapes [3] (synthetic), and SegmentMeIfYouCan [4]. Encountering and capturing images of OOD objects in real-world scenarios, without deliberately placing them on the road, is quite uncommon. Effectively, existing datasets are limited in scale, both in terms of the number of images and the variety of object classes. In this work, we propose leveraging advanced zero-shot inpainting techniques to augment an existing segmentation dataset with the insertion of OOD objects; this enables the generation of highly realistic synthetic data for testing and training OOD detection. + +Synthetic Data for Testing. Generative models have been exploited to create training data for image classification [2,8,34], object detection [22], or semantic segmentation [7,17,18,40,43]. Only recently, a few works have delved into the topic of generative testing data. Li et al. [19] exploit diffusion models to realistically edit images, controlling over various object attributes. This approach enables stress-testing models and understanding their sensitivity to different attributes. Using SD, LANCE [29] generates counterfactual images capable of challenging any given perception model. To the extent of our knowledge, no existing works have proposed to generate testing data for segmentation reliability. + +# 3 Reliability Under Covariate Shifts + +In this section, we explore whether synthetic data can be used to assess the robustness of pretrained segmenters in the presence of covariate shifts to unseen OOD domains. We describe the data generation process and present the benchmarking results for a wide range of segmenters on our synthetic data. We demonstrate the validity of the approach using domains for which a real OOD dataset exists so that we can have access to a gold standard. We stress that we do not use the OOD data at any point in the method itself. This aspect is critical for the method to be applicable to benchmark robustness in the presence of extreme or hazardous conditions. + +# 3.1 Generating images in arbitrary domains + +![](images/ca8c2188a25cf14eb2482daa909b816ba33e7b6d7aaf38bef4ea68a02a703ebf.jpg) +Fig. 2: Generating data with covariate shifts. Training (left) and Sampling (right) processes for producing the synthetic data with shifts. For training, only in-domain images and masks are used. For inference, we use the in-domain masks to generate OOD images. No real OOD data is required in the framework. + +![](images/48399ad6288318a3e5ae8042f87e2998c51f7f1f113d94ac96dcec34c4d676df.jpg) + +Our goal here is to obtain pairs of images and semantic masks, with the images belonging to visual domains for which we lack data. To this end, we leverage a pretrained text-to-image Stable Diffusion 1.5 (SD) model [31], repurposed as a semantic-conditioned model called ControlNet [42]. ControlNet is trained solely on images and segmentation ground truths from the Cityscapes dataset; at train time, the text prompts are captions automatically extracted using CLIP-interrogator [1]. As a result, the model is able to perform mask-to-image generation of driving scenes while retaining the ability to steer the generation through text prompting of Stable Diffusion. + +To generate synthetic data, we prompt a trained ControlNet by the concatenation of OOD domain descriptions and CLIP-interrogator captions obtained from Cityscapes validation images. Also, segmentation masks in Cityscapes validation set are used to condition the generative process. Thanks to zero-shot prompting, the synthetic images are aligned with the semantic condition while displaying the visual properties of arbitrary OOD domains. Fig. 2 illustrates the training and generation steps. Detailed technical descriptions and more visualizations are in the Supplementary Material. + +![](images/12e108e582f8a05ae89529ef01ed428fe795cc5eadf7c01591aee891e20372cb.jpg) +Fig. 3: Robustness correlation between real and synthetic covariate shifts across 40 pretrained segmenters. The tested models, see families in bottom legend, cover different architectures and sizes. (top) Pearson Correlation Coefficients of mIoUs between Cityscapes and real-shifts ('PCC_CS'), and between synthetic shifts and real ones ('PCC_Syn'). (bottom) Scatter plots of synthetic vs. real mIoUs along with the linear regression line accompanied by $95\%$ confidence intervals ('CI'). $(a - e)$ Five types of domain shifts from Cityscapes in-domain distribution, sorted by increasing gap as assessed by decreasing PCC_CS. The robustness results on synthetic data exhibit a strong correlation with those on real data, particularly in the case of the most distant shifts like 'snow' and 'night'. More details are provided as Supplementary Materials. + +# 3.2 Robustness Assessment with Synthetic Data + +With the pipeline outlined in Sec. 3.1, one can generate synthetic data to assess the robustness of pretrained segmenters in any unseen OOD domains through zero-shot prompting. To quantify robustness, we employ the traditional mean Intersection-over-Union (mIoU) score, measuring the correct overlap between semantic predictions and ground-truth masks. Given that our synthetic dataset comprises pairs of segmentation masks and synthetic images, one can straightforwardly derive synthetic scores for any pretrained segmenters. We here investigate whether synthetic performance can faithfully reflect the performance on real data in OOD domains under covariate shifts. + +Experimental Setups. We address weather shifts and geographical shifts, which are often encountered in autonomous driving. These covariate shifts are exhibited in two existing real datasets: the Adverse Conditions Dataset (ACDC) [33] and the Indian Driving Dataset (IDD) [39]. We utilize those real data to quantify the quality of our synthetic data and to validate its usefulness. + +Synthetic data are generated by conditioning on semantic masks from the Cityscapes validation set. To prompt ControlNet, we concatenate CLIP-interrogator's caption with a domain description following a simple template $[<\text{caption}>$ , in $<\text{domain}>$ ] where domain is either 'india', 'fog', 'rain', 'snow' or 'night'. + +For testing, we gather a collection of 40 publicly available segmenters trained only on Cityscapes, representative of different backbones, segmentation architectures, and sizes. The full list of models is in Supplementary Material. + +Results. In Fig. 3, we present our main results. Our primary metric is the Pearson Correlation Coefficient (PCC) between the mIoUs on testing data and on + +real-shift data from ACDC's splits or from IDD. The testing data can be either the Cityscapes validation set (CS) or our synthetic data (syn); the idea is to see which testing data –whether real CS or our synthetic one–correlates more with the real-shift data. Note that in the absence of OOD data, the Cityscapes validation set acts as the closest easily available proxy and is a reasonable predictor for OOD performance as pointed out by Jorge et al. [14]. + +We organize our results based on increasing domain gaps relative to the Cityscapes domain. The domain gaps are quantified by the Pearson correlation between Cityscapes mIoUs and real-shift mIoUs, annotated as PCC_CS and visualized as red bars in the subplots of Fig. 3. Moving from left to right, i.e. with growing domain gaps, we observe a widening discrepancy between PCC_CS and PCC_Syn. Here, PCC_Syn represents the Pearson correlation between synthetic mIoUs and real-shift mIoUs. In domains with small gaps, PCC_CS and PCC_Syn are relatively comparable. However, in domains with more adverse shifts, such as 'snow' and 'night', PCC_Syn outperforms PCC_CS significantly, exceeding PCC_CS in 'night' by more than double. + +In Fig. 4, we analyze the results for the 'night' condition using the most robust models across different architectures, ranging from ConvNets to recent transformer networks. We use the Semantic-FPN score as the reference to normalize the scores of other models. This normalization aims to illustrate the relative improvement in robustness in terms of architecture. We rank the models from left to right based on their performance on real night data from the ACDC-night split. The consistently increasing trend of synthetic scores ( $\rightarrow$ ) from left to right demonstrates a strong correlation with the ranking based on real scores. In contrast, Cityscapes scores ( $\rightarrow$ ) are not indicative of night performance: a higher mIoU obtained on Cityscapes does not immediately translate into a higher mIoU at night. + +Since synthetic data can be generated in any desired quantity, a natural question arises: how many images are sufficient? In addressing this question, we conducted experiments and presented the results in Fig. 5. Our empirical finding suggests that $\sim 500$ synthetic images are adequate for a stable and reliable assessment of robustness. + +Discussion. In their recent work on reliability in semantic segmentation, Jorge et al. [14] systematically quantified the robustness of segmenters on real-shift data; similarly to ours, they draw comparisons from ACDC and IDD datasets. One intriguing finding in this paper is that “[...] the larger the domain shift, the larger the improvement brought by more recent segmentation models”, hinting at a correlation between model robustness on in-domain data and covariate-shift data; that corresponds to the CS baseline we consider here. In our study, we delve deeper into this correlation, choosing to separately address different weather types instead of grouping them all together as done in [14]. In domains exhibiting small gaps to Cityscapes, such as IDD or ACDC-Fog, our conclusion aligns with [14]. However, as domain gaps increase, the discrepancy between Cityscapes mIoUs and real-shift mIoUs becomes more pronounced, resulting in poor PCC_CS scores. On the contrary, synthetic mIoUs and real-shift mIoUs + +![](images/b97ba5774136dbc05db5cee6f177ffa0f880fad6cc7df356d3e2cfa277f5d7fa.jpg) +Fig. 4: Day-night shift. Models are ranked from left to right by their robustness on real night data - ACDC-Night mIoUs are shown on top of model names. For each presented architecture, the most robust model on Cityscapes is tested; the SemanticFPN, DeeplabV3+, and PSPNet models have ResNet-101 as backbone. The SemanticFPN model (lowest mIoU on ACDC-Night) serves as the reference for computing the relative mIoUs. Blue bars or orange bars show the relative mIoUs when testing on our synthetic data (□) or testing on Cityscapes validation data (■). Cityscapes scores are not reliable for ranking models in the night domain. Synthetic scores exhibit a stronger correlation with real night scores, as evidenced by the more consistently increasing trend in the blue bars from left to right. + +![](images/1203cbdf388637c7316ca97cc2c62e0bf2c429d9976ef85d79920e6002ecf354.jpg) +Fig. 5: Pearson Correlation vs. # Synthetic Samples. Using more synthetic samples contributes to increased stability in the results. Empirical plots demonstrate that approximately 500 samples are sufficient for a stable correlation assessment. + +exhibit a strong correlation across shifts. Our empirical study has validated that synthetic performance is a reliable indicator of model robustness in the presence of covariate shifts. + +With synthetic data, we are able to confirm observations from the literature on real data, such as (i) the robustness of transformer and ConvNext backbones and that (ii) within an architecture family, robustness correlates with the number of parameters and the robustness of the backbone. + +Different generative models? The important ingredient in our pipeline is the generative model. In Tab. 1, we ablate by replacing the default SD1.5 model with other available ones: image-2-image GAN called TSIT [13], physics-based fog simulator [32], and a bigger SD variant called SDXL [28]. On 'night', 'rain', and 'snow', SD variants perform better than TSIT, while of note, TSIT was + +
OOD expertise?OOD data?Night Rain SnowFogIndia
GAN-based TSIT [13]norequired0.830.840.81--
Physics-based Fog Sim. [32]requiredno---0.82-
Ours w/ SD1.5 (default)nono0.850.860.850.770.71
Ours w/ SDXL [28]nono0.840.900.820.890.93
+ +Table 1: Different generative models. Our pipeline achieves better performance (PCC with real-shifts) while does not require any OOD knowledge. Of note, it's very easy to apply to new unseen OOD domains using our pipeline, which remains challenging for GAN-based and physics-based models. + +trained on real OOD data. The dedicated fog simulator performs very well, on par with our results using the larger model SDXL. Physics-based simulators are definitely valuable and this direction should be investigated further; however, such simulators require in-depth knowledge of OOD domains and hence are very difficult to design. Both GAN-based and physics-based approaches are limited in their scalability to many more OOD domains. Comparing SD1.5 vs. SDXL, we observe comparable results on challenging OOD domains while SDXL performs much better on 'fog' and 'india'; such results hint at the future potential of stronger and better generative models in further advancing virtual testing. Unfortunately, as SDXL is much more memory-demanding, we only limit our experiments with SDXL in this particular study. + +In the Supplementary Material, we report the FID scores, which measure the direct distance between synthetic and real distributions; we also extend a discussion on some limitations of our framework. + +# 3.3 Confidence Calibration with Synthetic Data + +Confidence calibration is a crucial aspect of deep networks, particularly when employed in safety-critical applications such as autonomous driving. Jorge et al. [14] highlighted a disconnection between model robustness and calibration, asserting that "... despite the remarkable improvements in terms of robustness, recent models are not significantly better calibrated". Therefore, it is essential to devise techniques and protocols for recalibrating data, particularly in domains exhibiting covariate shifts. Drawing inspiration from this, we explore the feasibility of using synthetic data to recalibrate pretrained segmenters. + +We perform temperature scaling using our synthetic data. Temperature scaling [6] is a well-established technique for calibrating pretrained models, typically conducted on a small validation set within the OOD domain. In our study, for each segmenter, we utilized the same sets of synthetic data generated in Sec. 3.2 to optimize temperature scaling factors, with one adjustment made for each covariate shift. For comparison, we replicate the process using real-shift data from ACDC and IDD. + +Fig. 6 reports the calibration improvement for the 40 pretrained segmenters using either real-shift data (■) or synthetic data (■). The subplots are arranged + +![](images/c8db3259078009d1748afed9294c1244a65f5834f47cd41326d10c59c1b4fe3b.jpg) +Fig. 6: ECE improvement using synthetic data. The 40 segmenters are calibrated using either real-shift data or synthetic data. For each model, the relative ECE improvement $(\%)$ over its non-calibrated version is computed, visualized by (synthetic shift) and (real shift). The subplots are ranked by model robustness on real-shift data (left-to-right) and by the increased domain shift (top-to-bottom). The titles of the subplots indicate the percentage of models that showed improvement with synthetic data. + +in increasing domain gap order from top to bottom, with the segmenters ranked from left to right based on increasing robustness on real-shift data. The Expected Calibration Error (ECE) [24] quantifies the calibration results, with a lower ECE indicating a better-calibrated model. For better interpretation, we present the relative ECE improvement, computed as the percentage decrease in ECE after calibration compared to the original ECE without calibration. For example, a model with an ECE of 0.4 before calibration and an ECE of 0.2 after calibration will achieve a $(0.4 - 0.2) / 0.4 = 50\%$ relative improvement. + +We observe promising calibration results when employing our synthetic data. While not as good as real-shift data, synthetic data achieves a promising success rate of $72.5\%$ on IDD and exceeds $90\%$ on the four ACDC shifts. Interestingly, in weather shifts, we empirically observe that more robust models derive greater benefits when calibrated using our synthetic data; the reverse is observed for 'europe-india' geographical shift. While with real-shift data, robustness and calibration are not well correlated [14], our results suggest that a potential correlation might exist between the two factors when using synthetic data. We note that calibration with temperature scaling does not always guarantee ECE improvement. Such phenomenon may happen even using real data, especially under domain shifts as explored in prior work [26]. In the Supplementary Material, we provide more technical details and results. + +# 3.4 On Practical Applicability + +One significant advantage of our framework lies in its potential to address rare conditions simply through prompting. The practical applicability of generative testing is tremendous. Our results demonstrate promising signals; practitioners can begin assessing and ranking their pretrained models for a new, unseen OOD domain of interest without the need for real data collection. In practice, our proposed generative benchmarking can serve as the initial step in a full validation pipeline, helping filter out non-robust prototypes and thereby saving on total operational costs. Starting from complementing real-data validation, one + +![](images/81c42b9df908182830ed82081068e30e7d063f92d6bdeb6cfa9dd1a600cef908.jpg) +Fig. 7: Qualitative results. Examples of rare conditions generated for testing and predictions from different models. Results of the strong model like SegFormer-B5 are visibly better than the Semantic-FPN and MobileNetV3. + +can envision a future where generative techniques mature to the point of fully replacing real-data validation. In Fig. 7, we visualize some synthetic images and model predictions under rare conditions, such as being flooded with water, having autumn leaves scattered across the road, or having a building on fire. We observe clear visual distinctions between weaker (Semantic-FPN and MobileNetV3) and stronger (SegF-B5) models, knowing that their Cityscapes scores do not differ significantly. More examples are provided in the Supplementary Material. + +# 4 Reliability Against OOD Objects + +We now address the reliability of segmentation models in the presence of Out-of-Distribution (OOD) objects. To begin, we explain our pipeline for inpainting random OOD objects into existing Cityscapes images. Following that, we demonstrate how one can utilize inpainted images for OOD detection assessment and for enhancing OOD detection. + +# 4.1 Inpainting Anomaly Objects + +We inpaint random objects into Cityscapes images. To this end, we initially sample a location — a square box to which we inpaint the new object. We crop the box, upsample its content to match the preferred output size of the generative model, and inpaint an object guided by a text prompt. In this step, we leverage Stable Diffusion inpainting capabilities, obtaining high-definition square images of the desired object. This image is then resized and pasted back into the original image, creating a final high-definition synthetic image. To ensure compositional consistency, we employ two techniques: Firstly, we divide the cropped box into two regions by center cropping it again. We inpaint only the inner region, leaving the outer region untouched, similar to the approach in RePaint [21]. Secondly, after composing the final image, we address any remaining inconsistencies by applying a light noise over the entire picture and performing reverse diffusion again. Details and visualizations are provided in the Supplementary Material. + +After inpainting, it is necessary to extract the mask corresponding to the new object. To achieve this, we begin with a high-definition square image and + +![](images/5fc2d2ed8532433e179ae126cbd58922d8f740ca92b92d23527f455654d6d1a9.jpg) +Fig. 8: Entropy Correlation. The top-left inset reports the Pearson correlations between real-OOD entropy vs. synthetic-OOD entropy, computed either on 'curated' (■) or all synthetic inpainted images (□). Evaluations are performed on the same model set used in Fig. 3, with similar markers. + +apply the Grounded Segment Anything [15,20], prompted with the name of the object. This process yields a mask within the square image, which can then be repositioned in the full image. + +![](images/a4cce56da056ea51dd830addc8064a23c0488559d3d89fbf17558e5055ca42fc.jpg) +Fig. 9: Correlation in OOD Detection. Each subplot scatters computed anomaly scores of segmenters on real OODs (y-axis) and on synthetic OODs (x-axis). The top row shows the four anomaly metrics utilized: $\mathrm{FPR}_{95}$ , AUROC, $\mathrm{AUPR}_{IN}$ , and $\mathrm{AUPR}_{OUT}$ . The results are organized into two rows corresponding to two different confidence measures (i) Entropy and (ii) MaxLogit. In the top-left corner of each subplot, an inset plots Pearson correlations to real OOD for 'curated' (■) and 'all' (■) synthetic sets. Evaluations are performed on the same model set used in Fig. 3, with similar markers. + +Our end-to-end generation pipeline is fully automatic. Through qualitative assessment, we achieve a satisfactory success rate in terms of generation realism; some inpainted images are illustrated in Fig. 10 and much more in the Supplementary Material. However, this still leaves a few generations with artifacts, characterized by either unusual compositions or unrealistic details. We here question the criticality of realism in assessing OOD detection and, furthermore, in improving OOD detection. To this end, we construct two different sets: (i) all 23,040 images generated automatically and (ii) 656 curated images where + +we manually select the best images in terms of visual quality and realism; specifically, curators filter out strong color saturation differences or partial objects, e.g. animal heads. We note that the manual selection process for the curated set is not exhaustive and is constrained by our allocated resources; there are many more high-quality images in the 'all' set. In what follows, we present results using both curated and uncurated sets. + +# 4.2 Assess OOD Detection + +Experimental setup. To measure how the segmenters react to unseen OOD objects, we use standard anomaly detection metrics [41], which are False Positive Rate at $95\%$ true positives $\mathrm{(FPR_{95})}$ , Area Under ROC curve (AUROC), and Area Under Precision-Recall curve (AUPR). AUPR are declined into $\mathrm{AUPR_{IN}}$ and $\mathrm{AUPR_{OUT}}$ , which consider the in-distribution regions, respectively the out-of-distribution regions (the inpainted object), as positive regions to compute the Precision-Recall curves. + +All segmenters in our study are not designed to produce confidence scores. We thus seek various techniques to derive confidence scores from pretrained models [9] and eventually narrow down the options to two measures: (i) Entropy of soft-probability predictions, and (ii) MaxLogit as the maximum logit value (before softmax) among the classes. While Entropy is the traditional measure of uncertainty, MaxLogit is a recent and surprising finding that has been proven to be more effective in estimating OOD confidence [9]. + +Results. For quantitative comparison, we leverage the SegmentMeIfYouCan (SMIYC) dataset [4], a recent dataset for OOD detection. We resort to the RoadAnomaly21 split in SMIYC, due to similarity in object scales to our synthetic data. We analyze the correlation between the OOD scores obtained on RoadAnomaly21 and one using our synthetic inpainted data. Fig. 8 reports our first analysis on the entropy in the OOD areas, either real or generated. For each model, we compute the Pearson Correlation (PCC) between real-OD entropy and synthetic-OD entropy; the computation is done on both 'curated' and 'all' sets. We observe a very high entropy correlation between real- and synthetic-OD, reaching 0.94 PCC using both 'curated' and 'all' sets. In Fig. 11, we show a control experiment in which we inpaint in-domain object class 'car' into Cityscapes scenes, and we analyze models' responses to synthetic cars, real cars, and OOD objects. + +We then analyze the correlation between real and synthetic anomaly metrics. Fig. 9 presents our primary findings and Fig. 10 illustrates some qualitative results. We observe a strong correlation with real scores when utilizing the 'curated' set for computing synthetic scores; the curated PCCs (■) are consistently around 0.8 across multiple metrics, irrespective of the two confidence measures. Although the correlations are somewhat weaker when using all uncurated synthetic data (■), such results remain acceptable, particularly when no effort is dedicated to curation. + +Our results validate the potential of utilizing realistic synthetic data, inpainted with anomaly objects, for assessing OOD detection. In OOD testing, it is quite + +![](images/ec7b8848aed786514c1e23c0dd46d52431ba1d74d59c85d2070199b27fa49b3d.jpg) +Fig. 10: Qualitative results. Confidence maps are visualized for the four exemplified models on real data (first row) and synthetic inpainted data (second and third rows). Hotter colors correspond to higher OOD likelihood. Ideally, results should exhibit hot colors in OOD areas and cold colors everywhere else. We observe a strong correlation in model reactions to real and synthetic OOD regions, particularly for more recent and robust models. We note that the real data in SMIYC also exhibit distributional shifts against Cityscapes; which already causes confusion to weak models in the background. + +![](images/f53f0840b5bea16eba753eab91b21444b315b7763ef8fdb23367866aeeb73b43.jpg) +Fig. 11: In-domain vs OOD inpainted objects. Models' responses to synthetic cars are close to real cars, and far from synthetic OOD objects + +important to use high-quality synthetic inpainted data. Nonetheless, even non-curated synthetic data can offer an acceptable estimation of real performance with minimal curation costs. + +
MethodAUROC (↑)AUPR (↑)FPR95 (↓)
RbA [25] Swin-B95.678.411.8
+ COCO [25]97.885.38.5
+ Ours (curated)97.284.98.1
+ Ours (all)97.384.88.2
RbA [25] Swin-L96.479.615.0
+ COCO [25]98.288.78.2
+ Ours (curated)97.288.07.9
+ Ours (all)98.188.68.3
+ +Table 2: Improving OOD detection on real SMIYC benchmarking using our synthetic data. All results are obtained using the published code and default parameters. + +# 4.3 Improve OOD Detection + +In this experiment, we investigate if synthetic inpainted data can be used to enhance a deep network's ability to detect OOD objects. To this end, we adopt + +the state-of-the-art RbA [25] approach for OOD detection and train RbA models on our synthetic data. + +OOD metrics are computed on RoadAnomaly21 and reported in Tab. 2. The RbA models trained on our data significantly outperform the vanilla RbA model. We reach comparable performance to the RbA variants that leverage the external COCO dataset for augmentation. Notably, there are no clear differences between using 'curated' or 'all' sets. We conjecture that, unlike benchmarking, training for OOD detection does not demand a high degree of realism from synthetic data. This explains why the simple strategy of copy-pasting COCO objects [25] already proves effective. All results are consistent across the two addressed backbones. Fig. 10 illustrates a few qualitative results. + +Discussion. In our work, we take all available levers and study the extent to which synthetic data can be used for evaluation and specific training purposes. We acknowledge the fact that the models used in this work were trained on substantial amounts of data. Of note, we do not claim the efficacy of synthetic data in all aspects, particularly regarding the total amount of training data required. Our findings are limited to highlighting the significant potential of published generative models in the task of reliability assessment. Advances in efficient training of generative models may address concerns regarding data quantity but are beyond the scope of this work. + +# 5 Takeaways + +In this work, we explore the potential of synthetic data in reliability assessment for semantic segmentation networks. We introduce two automatic zero-shot pipelines to generate data in OOD domains and to inpaint OOD objects for virtual reliability assessment. Our promising results encourage further collective investigations into this research problem, paving the way for synthetic system validation, especially in safety-critical applications. We summarize here our findings: + +$\triangleright$ Reliability Under Covariate Shifts: synthetic data can help assess the relative robustness of models in real-life covariate shifts, especially when shifts to the training condition are significant. Synthetic data can well complement real data in system validation, helping reduce the total operational cost. Pretrained models can be calibrated using synthetic data to better estimate prediction confidence in any shifted domains. +$\triangleright$ Reliability Against OOD Objects: synthetic data is useful in both OOD testing and OOD training; however, the demands on synthetic data quality differ in these two cases. In OOD testing, the best result estimations are obtained with the most realistically inpainted data, which may require a certain amount of curation time for qualitative assessment. The curation task is not time-demanding and can be done quickly with a reasonable budget. On the other hand, for OOD training, no curation is actually needed to achieve improvements. + +# Acknowledgements + +This work is supported by ELSA - European Lighthouse on Secure and Safe AI funded by the European Union under grant agreement No. 101070617. We thank the authors of RELIS [14] for providing us their pretrained checkpoints. + +# References + +1. clip-interrogator. https://github.com/pharmapsychotic/clip-interrogator (2023) +2. Besnier, V., Jain, H., Bursuc, A., Cord, M., Pérez, P.: This dataset does not exist: training models from generated images. In: ICASSP (2020) +3. Blum, H., Sarlin, P.E., Nieto, J., Siegwart, R., Cadena, C.: The fishyscapes benchmark: Measuring blind spots in semantic segmentation (2021) +4. Chan, R., Lis, K., Uhlemeyer, S., Blum, H., Honari, S., Siegwart, R., Fua, P., Salzmann, M., Rottmann, M.: Segmentmeifyoucan: A benchmark for anomaly segmentation. In: NeurIPS (2021) +5. Geirhos, R., Narayanappa, K., Mitzkus, B., Thieringer, T., Bethge, M., Wichmann, F.A., Brendel, W.: Partial success in closing the gap between human and machine vision. NeurIPS (2021) +6. Guo, C., Pleiss, G., Sun, Y., Weinberger, K.Q.: On calibration of modern neural networks. In: ICML (2017) +7. Hariat, M., Laurent, O., Kazmierczak, R., Zhang, S., Bursuc, A., Yao, A., Franchi, G.: Learning to generate training datasets for robust semantic segmentation. In: WACV (2024) +8. He, R., Sun, S., Yu, X., Xue, C., Zhang, W., Torr, P., Bai, S., Qi, X.: Is synthetic data from generative models ready for image recognition? (2023) +9. Hendrycks, D., Basart, S., Mazeika, M., Zou, A., Kwon, J., Mostajabi, M., Steinhardt, J., Song, D.: Scaling out-of-distribution detection for real-world settings. In: ICML (2022) +0. Hendrycks, D., Basart, S., Mu, N., Kadavath, S., Wang, F., Dorundo, E., Desai, R., Zhu, T., Parajuli, S., Guo, M., et al.: The many faces of robustness: A critical analysis of out-of-distribution generalization. In: ICCV (2021) +1. Hendrycks, D., Dietterich, T.: Benchmarking neural network robustness to common corruptions and perturbations. In: ICLR (2019) +2. Hendrycks, D., Gimpel, K.: A baseline for detecting misclassified and out-of-distribution examples in neural networks. In: ICLR (2017) +3. Jiang, L., Zhang, C., Huang, M., Liu, C., Shi, J., Loy, C.C.: Tsit: A simple and versatile framework for image-to-image translation. In: ECCV (2020) +4. de Jorge, P., Volpi, R., Torr, P.H., Rogez, G.: Reliability in semantic segmentation: Are we on the right track? In: CVPR (2023) +5. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., Dollar, P., Girshick, R.: Segment anything. arXiv:2304.02643 (2023) +6. Koh, P.W., Sagawa, S., Marklund, H., Xie, S.M., Zhang, M., Balsubramani, A., Hu, W., Yasunaga, M., Phillips, R.L., Gao, I., et al.: Wilds: A benchmark of in-the-wild distribution shifts. In: ICLR (2021) +7. Le Moing, G., Vu, T.H., Jain, H., Pérez, P., Cord, M.: Semantic palette: Guiding scene generation with class proportions. In: CVPR (2021) + +18. Li, D., Ling, H., Kim, S.W., Kreis, K., Fidler, S., Torralba, A.: Bigdatasetgan: Synthesizing imagenet with pixel-wise annotations. In: CVPR (2022) +19. Li, X., Chen, Y., Zhu, Y., Wang, S., Zhang, R., Xue, H.: Imagenet-e: Benchmarking neural network robustness via attribute editing. In: CVPR (2023) +20. Liu, S., Zeng, Z., Ren, T., Li, F., Zhang, H., Yang, J., Li, C., Yang, J., Su, H., Zhu, J., et al.: Grounding dino: Marrying dino with grounded pre-training for open-set object detection. arXiv preprint arXiv:2303.05499 (2023) +21. Lugmayr, A., Danelljan, M., Romero, A., Yu, F., Timofte, R., Gool, L.V.: Repaint: Inpainting using denoising diffusion probabilistic models. In: CVPR (2022) +22. Marathe, A., Ramanan, D., Walambe, R., Kotecha, K.: Wedge: A multi-weather autonomous driving dataset built from generative vision-language models. In: CVPRW (2023) +23. Miller, J.P., Taori, R., Raghunathan, A., Sagawa, S., Koh, P.W., Shankar, V., Liang, P., Carmon, Y., Schmidt, L.: Accuracy on the line: on the strong correlation between out-of-distribution and in-distribution generalization. In: ICLR (2021) +24. Naeini, M.P., Cooper, G., Hauskrecht, M.: Obtaining well calibrated probabilities using bayesian binning. In: AAAI (2015) +25. Nayal, N., Yavuz, M., Henriques, J.F., Güney, F.: Rba: Segmenting unknown regions rejected by all. In: ICCV (2023) +26. Ovadia, Y., Fertig, E., Ren, J., Nado, Z., Sculley, D., Nowozin, S., Dillon, J., Lakshminarayanan, B., Snoek, J.: Can you trust your model's uncertainty? evaluating predictive uncertainty under dataset shift. NeurIPS (2019) +27. Pinggera, P., Ramos, S., Gehrig, S., Franke, U., Rother, C., Mester, R.: Lost and found: detecting small road hazards for self-driving vehicles. In: IROS (2016) +28. Podell, D., English, Z., Lacey, K., Blattmann, A., Dockhorn, T., Müller, J., Penna, J., Rombach, R.: Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952 (2023) +29. Prabhu, V., Yenamandra, S., Chattopadhyay, P., Hoffman, J.: Lance: Stress-testing visual models by generating language-guided counterfactual images. In: NeurIPS (2023) +30. Recht, B., Roelofs, R., Schmidt, L., Shankar, V.: DoImagenet classifiers generalize toImagenet? In: ICLR (2019) +31. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: CVPR (2022) +32. Sakaridis, C., Dai, D., Van Gool, L.: Semantic foggy scene understanding with synthetic data. IJCV (2018) +33. Sakaridis, C., Dai, D., Van Gool, L.: Acdc: The adverse conditions dataset with correspondences for semantic driving scene understanding. In: ICCV (2021) +34. Sariyildiz, M.B., Alahari, K., Larlus, D., Kalantidis, Y.: Fake it till you make it: Learning transferable representations from synthetic imagenet clones. In: CVPR (2023) +35. Singh, A., Sarangmath, K., Chattopadhyay, P., Hoffman, J.: Benchmarking low-shot robustness to natural distribution shifts. In: ICCV (2023) +36. Taori, R., Dave, A., Shankar, V., Carlini, N., Recht, B., Schmidt, L.: Measuring robustness to natural distribution shifts in image classification. NeurIPS (2020) +37. Teney, D., Lin, Y., Oh, S.J., Abbasnejad, E.: Id and ood performance are sometimes inversely correlated on real-world datasets. arXiv (2022) +38. Tran, D., Liu, J., Dusenberry, M.W., Phan, D., Collier, M., Ren, J., Han, K., Wang, Z., Mariet, Z., Hu, H., et al.: Plex: Towards reliability using pretrained large model extensions. arXiv (2022) + +39. Varma, G., Subramanian, A., Namboodiri, A., Chandraker, M., Jawahar, C.: Idd: A dataset for exploring problems of autonomous navigation in unconstrained environments. In: WACV (2019) +40. Wu, W., Zhao, Y., Shou, M.Z., Zhou, H., Shen, C.: Diffumask: Synthesizing images with pixel-level annotations for semantic segmentation using diffusion models. In: ICCV (2023) +41. Yang, J., Wang, P., Zou, D., Zhou, Z., Ding, K., Peng, W., Wang, H., Chen, G., Li, B., Sun, Y., et al.: Openood: Benchmarking generalized out-of-distribution detection. NeurIPS (2022) +42. Zhang, L., Rao, A., Agrawala, M.: Adding conditional control to text-to-image diffusion models. In: ICCV (2023) +43. Zhang, Y., Ling, H., Gao, J., Yin, K., Lafleche, J.F., Barriuso, A., Torralba, A., Fidler, S.: Datasetgan: Efficient labeled data factory with minimal human effort. In: CVPR (2021) \ No newline at end of file diff --git a/2024/Reliability in Semantic Segmentation_ Can We Use Synthetic Data_/images.zip b/2024/Reliability in Semantic Segmentation_ Can We Use Synthetic Data_/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..8db491fc999efa70a9ee49d316ce7f0f650e51a1 --- /dev/null +++ b/2024/Reliability in Semantic Segmentation_ Can We Use Synthetic Data_/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2f36315fc5f0ef023739c7b8c3f40e5d10825c5ff4dec7bfa6910b449295449d +size 455659 diff --git a/2024/Reliability in Semantic Segmentation_ Can We Use Synthetic Data_/layout.json b/2024/Reliability in Semantic Segmentation_ Can We Use Synthetic Data_/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..b5a85152cbaae911d2de2c69dcae5cbf1791e620 --- /dev/null +++ b/2024/Reliability in Semantic Segmentation_ Can We Use Synthetic Data_/layout.json @@ -0,0 +1,6813 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 175, + 111, + 438, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 175, + 111, + 438, + 148 + ], + "spans": [ + { + "bbox": [ + 175, + 111, + 438, + 148 + ], + "type": "text", + "content": "Reliability in Semantic Segmentation: Can We Use Synthetic Data?" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 167, + 468, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 167, + 468, + 194 + ], + "spans": [ + { + "bbox": [ + 146, + 167, + 468, + 194 + ], + "type": "text", + "content": "Thibaut Loiseau" + }, + { + "bbox": [ + 146, + 167, + 468, + 194 + ], + "type": "inline_equation", + "content": "^{2\\dagger}" + }, + { + "bbox": [ + 146, + 167, + 468, + 194 + ], + "type": "text", + "content": ", Tuan-Hung Vu" + }, + { + "bbox": [ + 146, + 167, + 468, + 194 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 146, + 167, + 468, + 194 + ], + "type": "text", + "content": ", Mickael Chen" + }, + { + "bbox": [ + 146, + 167, + 468, + 194 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 146, + 167, + 468, + 194 + ], + "type": "text", + "content": ", Patrick Pérez" + }, + { + "bbox": [ + 146, + 167, + 468, + 194 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 146, + 167, + 468, + 194 + ], + "type": "text", + "content": ", and Matthieu Cord" + }, + { + "bbox": [ + 146, + 167, + 468, + 194 + ], + "type": "inline_equation", + "content": "^{1,4}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 145, + 201, + 468, + 223 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 201, + 468, + 223 + ], + "spans": [ + { + "bbox": [ + 145, + 201, + 468, + 223 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 145, + 201, + 468, + 223 + ], + "type": "text", + "content": " valeo.ai, Paris, France \n" + }, + { + "bbox": [ + 145, + 201, + 468, + 223 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 145, + 201, + 468, + 223 + ], + "type": "text", + "content": " LIGM, Ecole des Ponts, Univ Gustave Eiffel, CNRS, Marne-la-Vallee, France" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 259, + 224, + 355, + 235 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 259, + 224, + 355, + 235 + ], + "spans": [ + { + "bbox": [ + 259, + 224, + 355, + 235 + ], + "type": "text", + "content": "3 kyutai, Paris, France" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 231, + 235, + 382, + 245 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 231, + 235, + 382, + 245 + ], + "spans": [ + { + "bbox": [ + 231, + 235, + 382, + 245 + ], + "type": "text", + "content": "4 Sorbonne Université, Paris, France" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 159, + 272, + 455, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 272, + 455, + 460 + ], + "spans": [ + { + "bbox": [ + 159, + 272, + 455, + 460 + ], + "type": "text", + "content": "Abstract. Assessing the robustness of perception models to covariate shifts and their ability to detect out-of-distribution (OOD) inputs is crucial for safety-critical applications such as autonomous vehicles. By nature of such applications, however, the relevant data is difficult to collect and annotate. In this paper, we show for the first time how synthetic data can be specifically generated to assess comprehensively the real-world reliability of semantic segmentation models. By fine-tuning Stable Diffusion [31] with only in-domain data, we perform zero-shot generation of visual scenes in OOD domains or inpainted with OOD objects. This synthetic data is employed to evaluate the robustness of pretrained segmenters, thereby offering insights into their performance when confronted with real edge cases. Through extensive experiments, we demonstrate a high correlation between the performance of models when evaluated on our synthetic OOD data and when evaluated on real OOD inputs, showing the relevance of such virtual testing. Furthermore, we demonstrate how our approach can be utilized to enhance the calibration and OOD detection capabilities of segmenters. Code and data are made public." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 495, + 230, + 508 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 495, + 230, + 508 + ], + "spans": [ + { + "bbox": [ + 132, + 495, + 230, + 508 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 518, + 482, + 626 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 518, + 482, + 626 + ], + "spans": [ + { + "bbox": [ + 130, + 518, + 482, + 626 + ], + "type": "text", + "content": "Despite the rapid adoption of deep networks in safety-critical applications, reliability [26,36,38] has been an overlooked factor when designing and training these models. Recent efforts are geared toward enhancing model robustness under covariate shifts in data distributions [26, 30] and improving the model's ability to detect the unknown [10, 12, 41]. For open-world validation, in-domain data is no longer sufficient [16]; reliable and trustworthy systems demand more rigorous testing on diverse distributions, potentially exhibiting unknown objects. However, data collection campaigns can be quickly overwhelmed by the growing number of out-of-distribution (OOD) objects and conditions, sometimes extreme." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 627, + 482, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 627, + 482, + 651 + ], + "spans": [ + { + "bbox": [ + 130, + 627, + 482, + 651 + ], + "type": "text", + "content": "In this paper, we propose to leverage pretrained generative models, " + }, + { + "bbox": [ + 130, + 627, + 482, + 651 + ], + "type": "inline_equation", + "content": "e.g." + }, + { + "bbox": [ + 130, + 627, + 482, + 651 + ], + "type": "text", + "content": " Stable Diffusion 1.5 (SD), to alleviate the need for real OOD data. We use SD as a" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 141, + 654, + 326, + 666 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 654, + 326, + 666 + ], + "spans": [ + { + "bbox": [ + 141, + 654, + 326, + 666 + ], + "type": "text", + "content": "† Work done during an internship at valeo.ai" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 148, + 115, + 467, + 238 + ], + "blocks": [ + { + "bbox": [ + 148, + 115, + 467, + 238 + ], + "lines": [ + { + "bbox": [ + 148, + 115, + 467, + 238 + ], + "spans": [ + { + "bbox": [ + 148, + 115, + 467, + 238 + ], + "type": "image", + "image_path": "7d42448a0ecaa80a6205718151f8dd2cc847c2bd0d767f1bd6866fa951c07912.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 255, + 482, + 300 + ], + "lines": [ + { + "bbox": [ + 130, + 255, + 482, + 300 + ], + "spans": [ + { + "bbox": [ + 130, + 255, + 482, + 300 + ], + "type": "text", + "content": "Fig. 1: Assessing 40 pretrained segmenters under covariate shifts. Segmentation models under scrutiny were trained on Cityscapes train set only (in-domain data). They are evaluated on (i) Cityscapes validation set, (ii) real OOD data, and (iii) proposed synthetic data. We observe a strong correlation between results on (ii) and (iii)." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 323, + 375, + 334 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 323, + 375, + 334 + ], + "spans": [ + { + "bbox": [ + 130, + 323, + 375, + 334 + ], + "type": "text", + "content": "general validator, targeting multiple faces of reliability:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 335, + 481, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 335, + 481, + 407 + ], + "spans": [ + { + "bbox": [ + 130, + 335, + 481, + 407 + ], + "type": "text", + "content": "- First, we fine-tune SD on in-domain data to enable mask conditioning; in action, zero-shot prompting generates covariate shift images for testing. Fig. 1 illustrates the generation of synthetic data in OOD domains and its use for model evaluation. By assessing 40 pretrained segmenters under covariate shifts, our first contribution demonstrates how our generated synthetic OOD benchmark can act as a powerful proxy for the real OOD benchmark ACDC [33]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 407, + 482, + 477 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 407, + 482, + 477 + ], + "spans": [ + { + "bbox": [ + 130, + 407, + 482, + 477 + ], + "type": "text", + "content": "- Second, we explore a similar strategy for OOD object detection assessment: we inpaint objects of unknown classes into in-domain data by injecting appropriate prompts during the zero-shot inpainting process. Inpainted images and OOD masks are used to benchmark the 40 models to see how well they can recognize OOD objects. Our results are strongly correlated to the real OOD benchmark [4]. Our high-quality synthetic data is featured in the official BRAVO benchmark." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 478, + 481, + 538 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 478, + 481, + 538 + ], + "spans": [ + { + "bbox": [ + 130, + 478, + 481, + 538 + ], + "type": "text", + "content": "- Third, we demonstrate the usefulness of our two synthetic OOD benchmarks for hyperparameter tuning (here model calibration) and training. We calibrate the pretrained models to targeted OOD domains using our synthetic data and validate our strategy by comparing results to using real OOD data. We also train segmenters on inpainted data for OOD detection, obtaining competitive results." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 556, + 238, + 569 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 556, + 238, + 569 + ], + "spans": [ + { + "bbox": [ + 132, + 556, + 238, + 569 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "content": "Covariate Shifts. Modern machine learning models, notably deep networks, fall short in preserving their robustness and estimating their prediction confidence in the presence of covariate shifts [26,30]. Various benchmarks [5,11,16,35,36] have emerged to address the need for assessing models' reliability under different distributional shifts. Hendrycks et al. [11] propose a pioneering benchmark featuring data corrupted by various synthetic perturbations such as noise, blur, and brightness. Taori et al. [36] emphasize the importance of realistic shifts in" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 231, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 231, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 231, + 100 + ], + "type": "text", + "content": "T.Loiseau et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 390 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 481, + 390 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 481, + 390 + ], + "type": "text", + "content": "reliability assessment; they highlight the disparity between natural and synthetic shifts, asserting that there is minimal to no robustness transfer from synthetic to natural distribution shifts. Sign et al. [35] study low-shot robustness to natural distribution shifts, highlighting the robustness properties of advanced architecture and pretraining strategies. In [14], de Jorge et al. extend beyond the standard classification setup to address the reliability problem in semantic segmentation, with findings largely aligned with previous works on classification; interestingly, they point to the disconnection between robustness and confidence calibration, urging more attention to calibration during segmenter design and training. On a related line, prior works study the connections between in-domain and out-of-domain robustness [14, 23, 37]; they suggest either positive, none, or even negative correlations between ID and OOD robustness, which indeed largely depends on the type of shifts. In line with [14], we focus on segmentation as the visual perception task of interest. Different from previous works, we advocate for the use of advanced generative models to generate realistic synthetic data for testing segmenters under arbitrary covariate shifts. Taori et al. [36] criticize the synthetic robustness benchmarks and advocate for using real-shift data; while we agree on the importance of realism in testing, we demonstrate that the rapid advancement of generative models now permits very meaningful virtual assessment. Our goal is to study whether synthetic data can be a superior choice compared to ID data in correlation studies against OOD robustness – referred to as real-shift robustness here to avoid confusion with OOD detection. Inspired by [14], we also calibrate models but using synthetic OOD data instead." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 402, + 481, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 402, + 481, + 545 + ], + "spans": [ + { + "bbox": [ + 133, + 402, + 481, + 545 + ], + "type": "text", + "content": "OOD Object Detection. In addition to robustness and calibration, the ability to detect the \"unknown\" is equally important to assess for trustworthy systems [10,12,41]. In semantic segmentation, several datasets are available for evaluating OOD detection on the road, namely LostAndFound [27], StreetHazards [9] (synthetic), BDD-Anomaly [9], Fishyscapes [3] (synthetic), and SegmentMeIfYouCan [4]. Encountering and capturing images of OOD objects in real-world scenarios, without deliberately placing them on the road, is quite uncommon. Effectively, existing datasets are limited in scale, both in terms of the number of images and the variety of object classes. In this work, we propose leveraging advanced zero-shot inpainting techniques to augment an existing segmentation dataset with the insertion of OOD objects; this enables the generation of highly realistic synthetic data for testing and training OOD detection." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 557, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 557, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 557, + 481, + 665 + ], + "type": "text", + "content": "Synthetic Data for Testing. Generative models have been exploited to create training data for image classification [2,8,34], object detection [22], or semantic segmentation [7,17,18,40,43]. Only recently, a few works have delved into the topic of generative testing data. Li et al. [19] exploit diffusion models to realistically edit images, controlling over various object attributes. This approach enables stress-testing models and understanding their sensitivity to different attributes. Using SD, LANCE [29] generates counterfactual images capable of challenging any given perception model. To the extent of our knowledge, no existing works have proposed to generate testing data for segmentation reliability." + } + ] + } + ], + "index": 4 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 172, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 172, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 172, + 91, + 447, + 102 + ], + "type": "text", + "content": "Reliability in Semantic Segmentation: Can We Use Synthetic Data?" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 356, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 356, + 129 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 356, + 129 + ], + "type": "text", + "content": "3 Reliability Under Covariate Shifts" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 140, + 482, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 140, + 482, + 247 + ], + "spans": [ + { + "bbox": [ + 130, + 140, + 482, + 247 + ], + "type": "text", + "content": "In this section, we explore whether synthetic data can be used to assess the robustness of pretrained segmenters in the presence of covariate shifts to unseen OOD domains. We describe the data generation process and present the benchmarking results for a wide range of segmenters on our synthetic data. We demonstrate the validity of the approach using domains for which a real OOD dataset exists so that we can have access to a gold standard. We stress that we do not use the OOD data at any point in the method itself. This aspect is critical for the method to be applicable to benchmark robustness in the presence of extreme or hazardous conditions." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 266, + 362, + 278 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 266, + 362, + 278 + ], + "spans": [ + { + "bbox": [ + 132, + 266, + 362, + 278 + ], + "type": "text", + "content": "3.1 Generating images in arbitrary domains" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 134, + 307, + 299, + 383 + ], + "blocks": [ + { + "bbox": [ + 134, + 307, + 299, + 383 + ], + "lines": [ + { + "bbox": [ + 134, + 307, + 299, + 383 + ], + "spans": [ + { + "bbox": [ + 134, + 307, + 299, + 383 + ], + "type": "image", + "image_path": "ca8c2188a25cf14eb2482daa909b816ba33e7b6d7aaf38bef4ea68a02a703ebf.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 393, + 482, + 437 + ], + "lines": [ + { + "bbox": [ + 130, + 393, + 482, + 437 + ], + "spans": [ + { + "bbox": [ + 130, + 393, + 482, + 437 + ], + "type": "text", + "content": "Fig. 2: Generating data with covariate shifts. Training (left) and Sampling (right) processes for producing the synthetic data with shifts. For training, only in-domain images and masks are used. For inference, we use the in-domain masks to generate OOD images. No real OOD data is required in the framework." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 312, + 308, + 479, + 384 + ], + "blocks": [ + { + "bbox": [ + 312, + 308, + 479, + 384 + ], + "lines": [ + { + "bbox": [ + 312, + 308, + 479, + 384 + ], + "spans": [ + { + "bbox": [ + 312, + 308, + 479, + 384 + ], + "type": "image", + "image_path": "48399ad6288318a3e5ae8042f87e2998c51f7f1f113d94ac96dcec34c4d676df.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 462, + 482, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 462, + 482, + 569 + ], + "spans": [ + { + "bbox": [ + 130, + 462, + 482, + 569 + ], + "type": "text", + "content": "Our goal here is to obtain pairs of images and semantic masks, with the images belonging to visual domains for which we lack data. To this end, we leverage a pretrained text-to-image Stable Diffusion 1.5 (SD) model [31], repurposed as a semantic-conditioned model called ControlNet [42]. ControlNet is trained solely on images and segmentation ground truths from the Cityscapes dataset; at train time, the text prompts are captions automatically extracted using CLIP-interrogator [1]. As a result, the model is able to perform mask-to-image generation of driving scenes while retaining the ability to steer the generation through text prompting of Stable Diffusion." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "text", + "content": "To generate synthetic data, we prompt a trained ControlNet by the concatenation of OOD domain descriptions and CLIP-interrogator captions obtained from Cityscapes validation images. Also, segmentation masks in Cityscapes validation set are used to condition the generative process. Thanks to zero-shot prompting, the synthetic images are aligned with the semantic condition while displaying the visual properties of arbitrary OOD domains. Fig. 2 illustrates the training and generation steps. Detailed technical descriptions and more visualizations are in the Supplementary Material." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 231, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 231, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 231, + 101 + ], + "type": "text", + "content": "T.Loiseau et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 115, + 479, + 219 + ], + "blocks": [ + { + "bbox": [ + 133, + 115, + 479, + 219 + ], + "lines": [ + { + "bbox": [ + 133, + 115, + 479, + 219 + ], + "spans": [ + { + "bbox": [ + 133, + 115, + 479, + 219 + ], + "type": "image", + "image_path": "12e108e582f8a05ae89529ef01ed428fe795cc5eadf7c01591aee891e20372cb.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 226, + 482, + 338 + ], + "lines": [ + { + "bbox": [ + 130, + 226, + 482, + 338 + ], + "spans": [ + { + "bbox": [ + 130, + 226, + 482, + 338 + ], + "type": "text", + "content": "Fig. 3: Robustness correlation between real and synthetic covariate shifts across 40 pretrained segmenters. The tested models, see families in bottom legend, cover different architectures and sizes. (top) Pearson Correlation Coefficients of mIoUs between Cityscapes and real-shifts ('PCC_CS'), and between synthetic shifts and real ones ('PCC_Syn'). (bottom) Scatter plots of synthetic vs. real mIoUs along with the linear regression line accompanied by " + }, + { + "bbox": [ + 130, + 226, + 482, + 338 + ], + "type": "inline_equation", + "content": "95\\%" + }, + { + "bbox": [ + 130, + 226, + 482, + 338 + ], + "type": "text", + "content": " confidence intervals ('CI'). " + }, + { + "bbox": [ + 130, + 226, + 482, + 338 + ], + "type": "inline_equation", + "content": "(a - e)" + }, + { + "bbox": [ + 130, + 226, + 482, + 338 + ], + "type": "text", + "content": " Five types of domain shifts from Cityscapes in-domain distribution, sorted by increasing gap as assessed by decreasing PCC_CS. The robustness results on synthetic data exhibit a strong correlation with those on real data, particularly in the case of the most distant shifts like 'snow' and 'night'. More details are provided as Supplementary Materials." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 361, + 384, + 372 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 361, + 384, + 372 + ], + "spans": [ + { + "bbox": [ + 132, + 361, + 384, + 372 + ], + "type": "text", + "content": "3.2 Robustness Assessment with Synthetic Data" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 382, + 482, + 490 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 382, + 482, + 490 + ], + "spans": [ + { + "bbox": [ + 130, + 382, + 482, + 490 + ], + "type": "text", + "content": "With the pipeline outlined in Sec. 3.1, one can generate synthetic data to assess the robustness of pretrained segmenters in any unseen OOD domains through zero-shot prompting. To quantify robustness, we employ the traditional mean Intersection-over-Union (mIoU) score, measuring the correct overlap between semantic predictions and ground-truth masks. Given that our synthetic dataset comprises pairs of segmentation masks and synthetic images, one can straightforwardly derive synthetic scores for any pretrained segmenters. We here investigate whether synthetic performance can faithfully reflect the performance on real data in OOD domains under covariate shifts." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 494, + 482, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 494, + 482, + 553 + ], + "spans": [ + { + "bbox": [ + 130, + 494, + 482, + 553 + ], + "type": "text", + "content": "Experimental Setups. We address weather shifts and geographical shifts, which are often encountered in autonomous driving. These covariate shifts are exhibited in two existing real datasets: the Adverse Conditions Dataset (ACDC) [33] and the Indian Driving Dataset (IDD) [39]. We utilize those real data to quantify the quality of our synthetic data and to validate its usefulness." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 554, + 500, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 554, + 500, + 602 + ], + "spans": [ + { + "bbox": [ + 130, + 554, + 500, + 602 + ], + "type": "text", + "content": "Synthetic data are generated by conditioning on semantic masks from the Cityscapes validation set. To prompt ControlNet, we concatenate CLIP-interrogator's caption with a domain description following a simple template " + }, + { + "bbox": [ + 130, + 554, + 500, + 602 + ], + "type": "inline_equation", + "content": "[<\\text{caption}>" + }, + { + "bbox": [ + 130, + 554, + 500, + 602 + ], + "type": "text", + "content": ", in " + }, + { + "bbox": [ + 130, + 554, + 500, + 602 + ], + "type": "inline_equation", + "content": "<\\text{domain}>" + }, + { + "bbox": [ + 130, + 554, + 500, + 602 + ], + "type": "text", + "content": "] where domain is either 'india', 'fog', 'rain', 'snow' or 'night'." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 602, + 482, + 638 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 602, + 482, + 638 + ], + "spans": [ + { + "bbox": [ + 130, + 602, + 482, + 638 + ], + "type": "text", + "content": "For testing, we gather a collection of 40 publicly available segmenters trained only on Cityscapes, representative of different backbones, segmentation architectures, and sizes. The full list of models is in Supplementary Material." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 641, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 482, + 665 + ], + "type": "text", + "content": "Results. In Fig. 3, we present our main results. Our primary metric is the Pearson Correlation Coefficient (PCC) between the mIoUs on testing data and on" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 171, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 171, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 171, + 91, + 448, + 102 + ], + "type": "text", + "content": "Reliability in Semantic Segmentation: Can We Use Synthetic Data?" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "text", + "content": "real-shift data from ACDC's splits or from IDD. The testing data can be either the Cityscapes validation set (CS) or our synthetic data (syn); the idea is to see which testing data –whether real CS or our synthetic one–correlates more with the real-shift data. Note that in the absence of OOD data, the Cityscapes validation set acts as the closest easily available proxy and is a reasonable predictor for OOD performance as pointed out by Jorge et al. [14]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 190, + 483, + 310 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 190, + 483, + 310 + ], + "spans": [ + { + "bbox": [ + 130, + 190, + 483, + 310 + ], + "type": "text", + "content": "We organize our results based on increasing domain gaps relative to the Cityscapes domain. The domain gaps are quantified by the Pearson correlation between Cityscapes mIoUs and real-shift mIoUs, annotated as PCC_CS and visualized as red bars in the subplots of Fig. 3. Moving from left to right, i.e. with growing domain gaps, we observe a widening discrepancy between PCC_CS and PCC_Syn. Here, PCC_Syn represents the Pearson correlation between synthetic mIoUs and real-shift mIoUs. In domains with small gaps, PCC_CS and PCC_Syn are relatively comparable. However, in domains with more adverse shifts, such as 'snow' and 'night', PCC_Syn outperforms PCC_CS significantly, exceeding PCC_CS in 'night' by more than double." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 313, + 483, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 313, + 483, + 445 + ], + "spans": [ + { + "bbox": [ + 130, + 313, + 483, + 445 + ], + "type": "text", + "content": "In Fig. 4, we analyze the results for the 'night' condition using the most robust models across different architectures, ranging from ConvNets to recent transformer networks. We use the Semantic-FPN score as the reference to normalize the scores of other models. This normalization aims to illustrate the relative improvement in robustness in terms of architecture. We rank the models from left to right based on their performance on real night data from the ACDC-night split. The consistently increasing trend of synthetic scores ( " + }, + { + "bbox": [ + 130, + 313, + 483, + 445 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 130, + 313, + 483, + 445 + ], + "type": "text", + "content": " ) from left to right demonstrates a strong correlation with the ranking based on real scores. In contrast, Cityscapes scores ( " + }, + { + "bbox": [ + 130, + 313, + 483, + 445 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 130, + 313, + 483, + 445 + ], + "type": "text", + "content": " ) are not indicative of night performance: a higher mIoU obtained on Cityscapes does not immediately translate into a higher mIoU at night." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 447, + 482, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 447, + 482, + 506 + ], + "spans": [ + { + "bbox": [ + 130, + 447, + 482, + 506 + ], + "type": "text", + "content": "Since synthetic data can be generated in any desired quantity, a natural question arises: how many images are sufficient? In addressing this question, we conducted experiments and presented the results in Fig. 5. Our empirical finding suggests that " + }, + { + "bbox": [ + 130, + 447, + 482, + 506 + ], + "type": "inline_equation", + "content": "\\sim 500" + }, + { + "bbox": [ + 130, + 447, + 482, + 506 + ], + "type": "text", + "content": " synthetic images are adequate for a stable and reliable assessment of robustness." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "content": "Discussion. In their recent work on reliability in semantic segmentation, Jorge et al. [14] systematically quantified the robustness of segmenters on real-shift data; similarly to ours, they draw comparisons from ACDC and IDD datasets. One intriguing finding in this paper is that “[...] the larger the domain shift, the larger the improvement brought by more recent segmentation models”, hinting at a correlation between model robustness on in-domain data and covariate-shift data; that corresponds to the CS baseline we consider here. In our study, we delve deeper into this correlation, choosing to separately address different weather types instead of grouping them all together as done in [14]. In domains exhibiting small gaps to Cityscapes, such as IDD or ACDC-Fog, our conclusion aligns with [14]. However, as domain gaps increase, the discrepancy between Cityscapes mIoUs and real-shift mIoUs becomes more pronounced, resulting in poor PCC_CS scores. On the contrary, synthetic mIoUs and real-shift mIoUs" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 231, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 231, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 231, + 101 + ], + "type": "text", + "content": "T.Loiseau et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 170, + 116, + 444, + 253 + ], + "blocks": [ + { + "bbox": [ + 170, + 116, + 444, + 253 + ], + "lines": [ + { + "bbox": [ + 170, + 116, + 444, + 253 + ], + "spans": [ + { + "bbox": [ + 170, + 116, + 444, + 253 + ], + "type": "image", + "image_path": "b97ba5774136dbc05db5cee6f177ffa0f880fad6cc7df356d3e2cfa277f5d7fa.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 263, + 482, + 374 + ], + "lines": [ + { + "bbox": [ + 130, + 263, + 482, + 374 + ], + "spans": [ + { + "bbox": [ + 130, + 263, + 482, + 374 + ], + "type": "text", + "content": "Fig. 4: Day-night shift. Models are ranked from left to right by their robustness on real night data - ACDC-Night mIoUs are shown on top of model names. For each presented architecture, the most robust model on Cityscapes is tested; the SemanticFPN, DeeplabV3+, and PSPNet models have ResNet-101 as backbone. The SemanticFPN model (lowest mIoU on ACDC-Night) serves as the reference for computing the relative mIoUs. Blue bars or orange bars show the relative mIoUs when testing on our synthetic data (□) or testing on Cityscapes validation data (■). Cityscapes scores are not reliable for ranking models in the night domain. Synthetic scores exhibit a stronger correlation with real night scores, as evidenced by the more consistently increasing trend in the blue bars from left to right." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 128, + 388, + 302, + 494 + ], + "blocks": [ + { + "bbox": [ + 128, + 388, + 302, + 494 + ], + "lines": [ + { + "bbox": [ + 128, + 388, + 302, + 494 + ], + "spans": [ + { + "bbox": [ + 128, + 388, + 302, + 494 + ], + "type": "image", + "image_path": "1203cbdf388637c7316ca97cc2c62e0bf2c429d9976ef85d79920e6002ecf354.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 309, + 427, + 486, + 492 + ], + "lines": [ + { + "bbox": [ + 309, + 427, + 486, + 492 + ], + "spans": [ + { + "bbox": [ + 309, + 427, + 486, + 492 + ], + "type": "text", + "content": "Fig. 5: Pearson Correlation vs. # Synthetic Samples. Using more synthetic samples contributes to increased stability in the results. Empirical plots demonstrate that approximately 500 samples are sufficient for a stable correlation assessment." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 520, + 481, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 520, + 481, + 555 + ], + "spans": [ + { + "bbox": [ + 130, + 520, + 481, + 555 + ], + "type": "text", + "content": "exhibit a strong correlation across shifts. Our empirical study has validated that synthetic performance is a reliable indicator of model robustness in the presence of covariate shifts." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 556, + 481, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 556, + 481, + 604 + ], + "spans": [ + { + "bbox": [ + 130, + 556, + 481, + 604 + ], + "type": "text", + "content": "With synthetic data, we are able to confirm observations from the literature on real data, such as (i) the robustness of transformer and ConvNext backbones and that (ii) within an architecture family, robustness correlates with the number of parameters and the robustness of the backbone." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 605, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 481, + 665 + ], + "type": "text", + "content": "Different generative models? The important ingredient in our pipeline is the generative model. In Tab. 1, we ablate by replacing the default SD1.5 model with other available ones: image-2-image GAN called TSIT [13], physics-based fog simulator [32], and a bigger SD variant called SDXL [28]. On 'night', 'rain', and 'snow', SD variants perform better than TSIT, while of note, TSIT was" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 171, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 171, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 171, + 91, + 447, + 102 + ], + "type": "text", + "content": "Reliability in Semantic Segmentation: Can We Use Synthetic Data?" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 132, + 114, + 483, + 180 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 483, + 180 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 483, + 180 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 483, + 180 + ], + "type": "table", + "html": "
OOD expertise?OOD data?Night Rain SnowFogIndia
GAN-based TSIT [13]norequired0.830.840.81--
Physics-based Fog Sim. [32]requiredno---0.82-
Ours w/ SD1.5 (default)nono0.850.860.850.770.71
Ours w/ SDXL [28]nono0.840.900.820.890.93
", + "image_path": "6a47315201b4cf839c9f22ff61a5f3dc94c169897f5ae4459a9cfdf326fb8418.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 132, + 181, + 481, + 224 + ], + "lines": [ + { + "bbox": [ + 132, + 181, + 481, + 224 + ], + "spans": [ + { + "bbox": [ + 132, + 181, + 481, + 224 + ], + "type": "text", + "content": "Table 1: Different generative models. Our pipeline achieves better performance (PCC with real-shifts) while does not require any OOD knowledge. Of note, it's very easy to apply to new unseen OOD domains using our pipeline, which remains challenging for GAN-based and physics-based models." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 258, + 480, + 388 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 258, + 480, + 388 + ], + "spans": [ + { + "bbox": [ + 130, + 258, + 480, + 388 + ], + "type": "text", + "content": "trained on real OOD data. The dedicated fog simulator performs very well, on par with our results using the larger model SDXL. Physics-based simulators are definitely valuable and this direction should be investigated further; however, such simulators require in-depth knowledge of OOD domains and hence are very difficult to design. Both GAN-based and physics-based approaches are limited in their scalability to many more OOD domains. Comparing SD1.5 vs. SDXL, we observe comparable results on challenging OOD domains while SDXL performs much better on 'fog' and 'india'; such results hint at the future potential of stronger and better generative models in further advancing virtual testing. Unfortunately, as SDXL is much more memory-demanding, we only limit our experiments with SDXL in this particular study." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 389, + 480, + 425 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 389, + 480, + 425 + ], + "spans": [ + { + "bbox": [ + 130, + 389, + 480, + 425 + ], + "type": "text", + "content": "In the Supplementary Material, we report the FID scores, which measure the direct distance between synthetic and real distributions; we also extend a discussion on some limitations of our framework." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 442, + 382, + 454 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 442, + 382, + 454 + ], + "spans": [ + { + "bbox": [ + 132, + 442, + 382, + 454 + ], + "type": "text", + "content": "3.3 Confidence Calibration with Synthetic Data" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 462, + 480, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 462, + 480, + 557 + ], + "spans": [ + { + "bbox": [ + 130, + 462, + 480, + 557 + ], + "type": "text", + "content": "Confidence calibration is a crucial aspect of deep networks, particularly when employed in safety-critical applications such as autonomous driving. Jorge et al. [14] highlighted a disconnection between model robustness and calibration, asserting that \"... despite the remarkable improvements in terms of robustness, recent models are not significantly better calibrated\". Therefore, it is essential to devise techniques and protocols for recalibrating data, particularly in domains exhibiting covariate shifts. Drawing inspiration from this, we explore the feasibility of using synthetic data to recalibrate pretrained segmenters." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 558, + 480, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 558, + 480, + 640 + ], + "spans": [ + { + "bbox": [ + 130, + 558, + 480, + 640 + ], + "type": "text", + "content": "We perform temperature scaling using our synthetic data. Temperature scaling [6] is a well-established technique for calibrating pretrained models, typically conducted on a small validation set within the OOD domain. In our study, for each segmenter, we utilized the same sets of synthetic data generated in Sec. 3.2 to optimize temperature scaling factors, with one adjustment made for each covariate shift. For comparison, we replicate the process using real-shift data from ACDC and IDD." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 641, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 641, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 641, + 481, + 666 + ], + "type": "text", + "content": "Fig. 6 reports the calibration improvement for the 40 pretrained segmenters using either real-shift data (■) or synthetic data (■). The subplots are arranged" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 231, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 231, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 231, + 101 + ], + "type": "text", + "content": "T.Loiseau et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 144, + 114, + 297, + 271 + ], + "blocks": [ + { + "bbox": [ + 144, + 114, + 297, + 271 + ], + "lines": [ + { + "bbox": [ + 144, + 114, + 297, + 271 + ], + "spans": [ + { + "bbox": [ + 144, + 114, + 297, + 271 + ], + "type": "image", + "image_path": "c8db3259078009d1748afed9294c1244a65f5834f47cd41326d10c59c1b4fe3b.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 310, + 116, + 470, + 270 + ], + "lines": [ + { + "bbox": [ + 310, + 116, + 470, + 270 + ], + "spans": [ + { + "bbox": [ + 310, + 116, + 470, + 270 + ], + "type": "text", + "content": "Fig. 6: ECE improvement using synthetic data. The 40 segmenters are calibrated using either real-shift data or synthetic data. For each model, the relative ECE improvement " + }, + { + "bbox": [ + 310, + 116, + 470, + 270 + ], + "type": "inline_equation", + "content": "(\\%)" + }, + { + "bbox": [ + 310, + 116, + 470, + 270 + ], + "type": "text", + "content": " over its non-calibrated version is computed, visualized by (synthetic shift) and (real shift). The subplots are ranked by model robustness on real-shift data (left-to-right) and by the increased domain shift (top-to-bottom). The titles of the subplots indicate the percentage of models that showed improvement with synthetic data." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 293, + 482, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 293, + 482, + 389 + ], + "spans": [ + { + "bbox": [ + 130, + 293, + 482, + 389 + ], + "type": "text", + "content": "in increasing domain gap order from top to bottom, with the segmenters ranked from left to right based on increasing robustness on real-shift data. The Expected Calibration Error (ECE) [24] quantifies the calibration results, with a lower ECE indicating a better-calibrated model. For better interpretation, we present the relative ECE improvement, computed as the percentage decrease in ECE after calibration compared to the original ECE without calibration. For example, a model with an ECE of 0.4 before calibration and an ECE of 0.2 after calibration will achieve a " + }, + { + "bbox": [ + 130, + 293, + 482, + 389 + ], + "type": "inline_equation", + "content": "(0.4 - 0.2) / 0.4 = 50\\%" + }, + { + "bbox": [ + 130, + 293, + 482, + 389 + ], + "type": "text", + "content": " relative improvement." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 390, + 482, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 390, + 482, + 534 + ], + "spans": [ + { + "bbox": [ + 130, + 390, + 482, + 534 + ], + "type": "text", + "content": "We observe promising calibration results when employing our synthetic data. While not as good as real-shift data, synthetic data achieves a promising success rate of " + }, + { + "bbox": [ + 130, + 390, + 482, + 534 + ], + "type": "inline_equation", + "content": "72.5\\%" + }, + { + "bbox": [ + 130, + 390, + 482, + 534 + ], + "type": "text", + "content": " on IDD and exceeds " + }, + { + "bbox": [ + 130, + 390, + 482, + 534 + ], + "type": "inline_equation", + "content": "90\\%" + }, + { + "bbox": [ + 130, + 390, + 482, + 534 + ], + "type": "text", + "content": " on the four ACDC shifts. Interestingly, in weather shifts, we empirically observe that more robust models derive greater benefits when calibrated using our synthetic data; the reverse is observed for 'europe-india' geographical shift. While with real-shift data, robustness and calibration are not well correlated [14], our results suggest that a potential correlation might exist between the two factors when using synthetic data. We note that calibration with temperature scaling does not always guarantee ECE improvement. Such phenomenon may happen even using real data, especially under domain shifts as explored in prior work [26]. In the Supplementary Material, we provide more technical details and results." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 550, + 293, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 550, + 293, + 562 + ], + "spans": [ + { + "bbox": [ + 131, + 550, + 293, + 562 + ], + "type": "text", + "content": "3.4 On Practical Applicability" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "type": "text", + "content": "One significant advantage of our framework lies in its potential to address rare conditions simply through prompting. The practical applicability of generative testing is tremendous. Our results demonstrate promising signals; practitioners can begin assessing and ranking their pretrained models for a new, unseen OOD domain of interest without the need for real data collection. In practice, our proposed generative benchmarking can serve as the initial step in a full validation pipeline, helping filter out non-robust prototypes and thereby saving on total operational costs. Starting from complementing real-data validation, one" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 171, + 90, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 171, + 90, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 171, + 90, + 447, + 102 + ], + "type": "text", + "content": "Reliability in Semantic Segmentation: Can We Use Synthetic Data?" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 141, + 113, + 304, + 240 + ], + "blocks": [ + { + "bbox": [ + 141, + 113, + 304, + 240 + ], + "lines": [ + { + "bbox": [ + 141, + 113, + 304, + 240 + ], + "spans": [ + { + "bbox": [ + 141, + 113, + 304, + 240 + ], + "type": "image", + "image_path": "81c42b9df908182830ed82081068e30e7d063f92d6bdeb6cfa9dd1a600cef908.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 315, + 131, + 482, + 197 + ], + "lines": [ + { + "bbox": [ + 315, + 131, + 482, + 197 + ], + "spans": [ + { + "bbox": [ + 315, + 131, + 482, + 197 + ], + "type": "text", + "content": "Fig. 7: Qualitative results. Examples of rare conditions generated for testing and predictions from different models. Results of the strong model like SegFormer-B5 are visibly better than the Semantic-FPN and MobileNetV3." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 262, + 482, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 262, + 482, + 348 + ], + "spans": [ + { + "bbox": [ + 130, + 262, + 482, + 348 + ], + "type": "text", + "content": "can envision a future where generative techniques mature to the point of fully replacing real-data validation. In Fig. 7, we visualize some synthetic images and model predictions under rare conditions, such as being flooded with water, having autumn leaves scattered across the road, or having a building on fire. We observe clear visual distinctions between weaker (Semantic-FPN and MobileNetV3) and stronger (SegF-B5) models, knowing that their Cityscapes scores do not differ significantly. More examples are provided in the Supplementary Material." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 364, + 350, + 379 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 364, + 350, + 379 + ], + "spans": [ + { + "bbox": [ + 132, + 364, + 350, + 379 + ], + "type": "text", + "content": "4 Reliability Against OOD Objects" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 389, + 482, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 389, + 482, + 449 + ], + "spans": [ + { + "bbox": [ + 130, + 389, + 482, + 449 + ], + "type": "text", + "content": "We now address the reliability of segmentation models in the presence of Out-of-Distribution (OOD) objects. To begin, we explain our pipeline for inpainting random OOD objects into existing Cityscapes images. Following that, we demonstrate how one can utilize inpainted images for OOD detection assessment and for enhancing OOD detection." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 466, + 304, + 479 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 466, + 304, + 479 + ], + "spans": [ + { + "bbox": [ + 132, + 466, + 304, + 479 + ], + "type": "text", + "content": "4.1 Inpainting Anomaly Objects" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 486, + 482, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 486, + 482, + 641 + ], + "spans": [ + { + "bbox": [ + 130, + 486, + 482, + 641 + ], + "type": "text", + "content": "We inpaint random objects into Cityscapes images. To this end, we initially sample a location — a square box to which we inpaint the new object. We crop the box, upsample its content to match the preferred output size of the generative model, and inpaint an object guided by a text prompt. In this step, we leverage Stable Diffusion inpainting capabilities, obtaining high-definition square images of the desired object. This image is then resized and pasted back into the original image, creating a final high-definition synthetic image. To ensure compositional consistency, we employ two techniques: Firstly, we divide the cropped box into two regions by center cropping it again. We inpaint only the inner region, leaving the outer region untouched, similar to the approach in RePaint [21]. Secondly, after composing the final image, we address any remaining inconsistencies by applying a light noise over the entire picture and performing reverse diffusion again. Details and visualizations are provided in the Supplementary Material." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 642, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 642, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 131, + 642, + 481, + 666 + ], + "type": "text", + "content": "After inpainting, it is necessary to extract the mask corresponding to the new object. To achieve this, we begin with a high-definition square image and" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 231, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 231, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 231, + 100 + ], + "type": "text", + "content": "T.Loiseau et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 143, + 113, + 303, + 222 + ], + "blocks": [ + { + "bbox": [ + 143, + 113, + 303, + 222 + ], + "lines": [ + { + "bbox": [ + 143, + 113, + 303, + 222 + ], + "spans": [ + { + "bbox": [ + 143, + 113, + 303, + 222 + ], + "type": "image", + "image_path": "5fc2d2ed8532433e179ae126cbd58922d8f740ca92b92d23527f455654d6d1a9.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 310, + 119, + 470, + 207 + ], + "lines": [ + { + "bbox": [ + 310, + 119, + 470, + 207 + ], + "spans": [ + { + "bbox": [ + 310, + 119, + 470, + 207 + ], + "type": "text", + "content": "Fig. 8: Entropy Correlation. The top-left inset reports the Pearson correlations between real-OOD entropy vs. synthetic-OOD entropy, computed either on 'curated' (■) or all synthetic inpainted images (□). Evaluations are performed on the same model set used in Fig. 3, with similar markers." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 244, + 480, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 244, + 480, + 281 + ], + "spans": [ + { + "bbox": [ + 130, + 244, + 480, + 281 + ], + "type": "text", + "content": "apply the Grounded Segment Anything [15,20], prompted with the name of the object. This process yields a mask within the square image, which can then be repositioned in the full image." + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 130, + 300, + 482, + 449 + ], + "blocks": [ + { + "bbox": [ + 130, + 300, + 482, + 449 + ], + "lines": [ + { + "bbox": [ + 130, + 300, + 482, + 449 + ], + "spans": [ + { + "bbox": [ + 130, + 300, + 482, + 449 + ], + "type": "image", + "image_path": "a4cce56da056ea51dd830addc8064a23c0488559d3d89fbf17558e5055ca42fc.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 456, + 482, + 544 + ], + "lines": [ + { + "bbox": [ + 130, + 456, + 482, + 544 + ], + "spans": [ + { + "bbox": [ + 130, + 456, + 482, + 544 + ], + "type": "text", + "content": "Fig. 9: Correlation in OOD Detection. Each subplot scatters computed anomaly scores of segmenters on real OODs (y-axis) and on synthetic OODs (x-axis). The top row shows the four anomaly metrics utilized: " + }, + { + "bbox": [ + 130, + 456, + 482, + 544 + ], + "type": "inline_equation", + "content": "\\mathrm{FPR}_{95}" + }, + { + "bbox": [ + 130, + 456, + 482, + 544 + ], + "type": "text", + "content": ", AUROC, " + }, + { + "bbox": [ + 130, + 456, + 482, + 544 + ], + "type": "inline_equation", + "content": "\\mathrm{AUPR}_{IN}" + }, + { + "bbox": [ + 130, + 456, + 482, + 544 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 456, + 482, + 544 + ], + "type": "inline_equation", + "content": "\\mathrm{AUPR}_{OUT}" + }, + { + "bbox": [ + 130, + 456, + 482, + 544 + ], + "type": "text", + "content": ". The results are organized into two rows corresponding to two different confidence measures (i) Entropy and (ii) MaxLogit. In the top-left corner of each subplot, an inset plots Pearson correlations to real OOD for 'curated' (■) and 'all' (■) synthetic sets. Evaluations are performed on the same model set used in Fig. 3, with similar markers." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": "Our end-to-end generation pipeline is fully automatic. Through qualitative assessment, we achieve a satisfactory success rate in terms of generation realism; some inpainted images are illustrated in Fig. 10 and much more in the Supplementary Material. However, this still leaves a few generations with artifacts, characterized by either unusual compositions or unrealistic details. We here question the criticality of realism in assessing OOD detection and, furthermore, in improving OOD detection. To this end, we construct two different sets: (i) all 23,040 images generated automatically and (ii) 656 curated images where" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 171, + 90, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 171, + 90, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 171, + 90, + 447, + 102 + ], + "type": "text", + "content": "Reliability in Semantic Segmentation: Can We Use Synthetic Data?" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "text", + "content": "we manually select the best images in terms of visual quality and realism; specifically, curators filter out strong color saturation differences or partial objects, e.g. animal heads. We note that the manual selection process for the curated set is not exhaustive and is constrained by our allocated resources; there are many more high-quality images in the 'all' set. In what follows, we present results using both curated and uncurated sets." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 205, + 277, + 216 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 205, + 277, + 216 + ], + "spans": [ + { + "bbox": [ + 132, + 205, + 277, + 216 + ], + "type": "text", + "content": "4.2 Assess OOD Detection" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 223, + 482, + 305 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 223, + 482, + 305 + ], + "spans": [ + { + "bbox": [ + 130, + 223, + 482, + 305 + ], + "type": "text", + "content": "Experimental setup. To measure how the segmenters react to unseen OOD objects, we use standard anomaly detection metrics [41], which are False Positive Rate at " + }, + { + "bbox": [ + 130, + 223, + 482, + 305 + ], + "type": "inline_equation", + "content": "95\\%" + }, + { + "bbox": [ + 130, + 223, + 482, + 305 + ], + "type": "text", + "content": " true positives " + }, + { + "bbox": [ + 130, + 223, + 482, + 305 + ], + "type": "inline_equation", + "content": "\\mathrm{(FPR_{95})}" + }, + { + "bbox": [ + 130, + 223, + 482, + 305 + ], + "type": "text", + "content": ", Area Under ROC curve (AUROC), and Area Under Precision-Recall curve (AUPR). AUPR are declined into " + }, + { + "bbox": [ + 130, + 223, + 482, + 305 + ], + "type": "inline_equation", + "content": "\\mathrm{AUPR_{IN}}" + }, + { + "bbox": [ + 130, + 223, + 482, + 305 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 223, + 482, + 305 + ], + "type": "inline_equation", + "content": "\\mathrm{AUPR_{OUT}}" + }, + { + "bbox": [ + 130, + 223, + 482, + 305 + ], + "type": "text", + "content": ", which consider the in-distribution regions, respectively the out-of-distribution regions (the inpainted object), as positive regions to compute the Precision-Recall curves." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 307, + 482, + 390 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 307, + 482, + 390 + ], + "spans": [ + { + "bbox": [ + 130, + 307, + 482, + 390 + ], + "type": "text", + "content": "All segmenters in our study are not designed to produce confidence scores. We thus seek various techniques to derive confidence scores from pretrained models [9] and eventually narrow down the options to two measures: (i) Entropy of soft-probability predictions, and (ii) MaxLogit as the maximum logit value (before softmax) among the classes. While Entropy is the traditional measure of uncertainty, MaxLogit is a recent and surprising finding that has been proven to be more effective in estimating OOD confidence [9]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 391, + 482, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 391, + 482, + 545 + ], + "spans": [ + { + "bbox": [ + 130, + 391, + 482, + 545 + ], + "type": "text", + "content": "Results. For quantitative comparison, we leverage the SegmentMeIfYouCan (SMIYC) dataset [4], a recent dataset for OOD detection. We resort to the RoadAnomaly21 split in SMIYC, due to similarity in object scales to our synthetic data. We analyze the correlation between the OOD scores obtained on RoadAnomaly21 and one using our synthetic inpainted data. Fig. 8 reports our first analysis on the entropy in the OOD areas, either real or generated. For each model, we compute the Pearson Correlation (PCC) between real-OD entropy and synthetic-OD entropy; the computation is done on both 'curated' and 'all' sets. We observe a very high entropy correlation between real- and synthetic-OD, reaching 0.94 PCC using both 'curated' and 'all' sets. In Fig. 11, we show a control experiment in which we inpaint in-domain object class 'car' into Cityscapes scenes, and we analyze models' responses to synthetic cars, real cars, and OOD objects." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 546, + 482, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 546, + 482, + 640 + ], + "spans": [ + { + "bbox": [ + 130, + 546, + 482, + 640 + ], + "type": "text", + "content": "We then analyze the correlation between real and synthetic anomaly metrics. Fig. 9 presents our primary findings and Fig. 10 illustrates some qualitative results. We observe a strong correlation with real scores when utilizing the 'curated' set for computing synthetic scores; the curated PCCs (■) are consistently around 0.8 across multiple metrics, irrespective of the two confidence measures. Although the correlations are somewhat weaker when using all uncurated synthetic data (■), such results remain acceptable, particularly when no effort is dedicated to curation." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 641, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 641, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 641, + 482, + 666 + ], + "type": "text", + "content": "Our results validate the potential of utilizing realistic synthetic data, inpainted with anomaly objects, for assessing OOD detection. In OOD testing, it is quite" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 231, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 231, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 231, + 101 + ], + "type": "text", + "content": "T.Loiseau et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 140, + 114, + 471, + 243 + ], + "blocks": [ + { + "bbox": [ + 140, + 114, + 471, + 243 + ], + "lines": [ + { + "bbox": [ + 140, + 114, + 471, + 243 + ], + "spans": [ + { + "bbox": [ + 140, + 114, + 471, + 243 + ], + "type": "image", + "image_path": "ec7b8848aed786514c1e23c0dd46d52431ba1d74d59c85d2070199b27fa49b3d.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 254, + 482, + 331 + ], + "lines": [ + { + "bbox": [ + 130, + 254, + 482, + 331 + ], + "spans": [ + { + "bbox": [ + 130, + 254, + 482, + 331 + ], + "type": "text", + "content": "Fig. 10: Qualitative results. Confidence maps are visualized for the four exemplified models on real data (first row) and synthetic inpainted data (second and third rows). Hotter colors correspond to higher OOD likelihood. Ideally, results should exhibit hot colors in OOD areas and cold colors everywhere else. We observe a strong correlation in model reactions to real and synthetic OOD regions, particularly for more recent and robust models. We note that the real data in SMIYC also exhibit distributional shifts against Cityscapes; which already causes confusion to weak models in the background." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 128, + 346, + 302, + 388 + ], + "blocks": [ + { + "bbox": [ + 128, + 346, + 302, + 388 + ], + "lines": [ + { + "bbox": [ + 128, + 346, + 302, + 388 + ], + "spans": [ + { + "bbox": [ + 128, + 346, + 302, + 388 + ], + "type": "image", + "image_path": "f53f0840b5bea16eba753eab91b21444b315b7763ef8fdb23367866aeeb73b43.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 309, + 342, + 486, + 387 + ], + "lines": [ + { + "bbox": [ + 309, + 342, + 486, + 387 + ], + "spans": [ + { + "bbox": [ + 309, + 342, + 486, + 387 + ], + "type": "text", + "content": "Fig. 11: In-domain vs OOD inpainted objects. Models' responses to synthetic cars are close to real cars, and far from synthetic OOD objects" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 410, + 482, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 410, + 482, + 445 + ], + "spans": [ + { + "bbox": [ + 132, + 410, + 482, + 445 + ], + "type": "text", + "content": "important to use high-quality synthetic inpainted data. Nonetheless, even non-curated synthetic data can offer an acceptable estimation of real performance with minimal curation costs." + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 132, + 464, + 367, + 582 + ], + "blocks": [ + { + "bbox": [ + 132, + 464, + 367, + 582 + ], + "lines": [ + { + "bbox": [ + 132, + 464, + 367, + 582 + ], + "spans": [ + { + "bbox": [ + 132, + 464, + 367, + 582 + ], + "type": "table", + "html": "
MethodAUROC (↑)AUPR (↑)FPR95 (↓)
RbA [25] Swin-B95.678.411.8
+ COCO [25]97.885.38.5
+ Ours (curated)97.284.98.1
+ Ours (all)97.384.88.2
RbA [25] Swin-L96.479.615.0
+ COCO [25]98.288.78.2
+ Ours (curated)97.288.07.9
+ Ours (all)98.188.68.3
", + "image_path": "10a4b2e23f3fdc81589460076374164ca045e6357cecd6c24271d504ce897054.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 374, + 464, + 482, + 541 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 374, + 464, + 482, + 541 + ], + "spans": [ + { + "bbox": [ + 374, + 464, + 482, + 541 + ], + "type": "text", + "content": "Table 2: Improving OOD detection on real SMIYC benchmarking using our synthetic data. All results are obtained using the published code and default parameters." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 622, + 285, + 634 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 622, + 285, + 634 + ], + "spans": [ + { + "bbox": [ + 132, + 622, + 285, + 634 + ], + "type": "text", + "content": "4.3 Improve OOD Detection" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "content": "In this experiment, we investigate if synthetic inpainted data can be used to enhance a deep network's ability to detect OOD objects. To this end, we adopt" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 171, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 171, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 171, + 91, + 447, + 102 + ], + "type": "text", + "content": "Reliability in Semantic Segmentation: Can We Use Synthetic Data?" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 139 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 139 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 139 + ], + "type": "text", + "content": "the state-of-the-art RbA [25] approach for OOD detection and train RbA models on our synthetic data." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 140, + 481, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 140, + 481, + 247 + ], + "spans": [ + { + "bbox": [ + 130, + 140, + 481, + 247 + ], + "type": "text", + "content": "OOD metrics are computed on RoadAnomaly21 and reported in Tab. 2. The RbA models trained on our data significantly outperform the vanilla RbA model. We reach comparable performance to the RbA variants that leverage the external COCO dataset for augmentation. Notably, there are no clear differences between using 'curated' or 'all' sets. We conjecture that, unlike benchmarking, training for OOD detection does not demand a high degree of realism from synthetic data. This explains why the simple strategy of copy-pasting COCO objects [25] already proves effective. All results are consistent across the two addressed backbones. Fig. 10 illustrates a few qualitative results." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 247, + 482, + 355 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 247, + 482, + 355 + ], + "spans": [ + { + "bbox": [ + 130, + 247, + 482, + 355 + ], + "type": "text", + "content": "Discussion. In our work, we take all available levers and study the extent to which synthetic data can be used for evaluation and specific training purposes. We acknowledge the fact that the models used in this work were trained on substantial amounts of data. Of note, we do not claim the efficacy of synthetic data in all aspects, particularly regarding the total amount of training data required. Our findings are limited to highlighting the significant potential of published generative models in the task of reliability assessment. Advances in efficient training of generative models may address concerns regarding data quantity but are beyond the scope of this work." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 373, + 217, + 387 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 373, + 217, + 387 + ], + "spans": [ + { + "bbox": [ + 132, + 373, + 217, + 387 + ], + "type": "text", + "content": "5 Takeaways" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 398, + 482, + 481 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 398, + 482, + 481 + ], + "spans": [ + { + "bbox": [ + 130, + 398, + 482, + 481 + ], + "type": "text", + "content": "In this work, we explore the potential of synthetic data in reliability assessment for semantic segmentation networks. We introduce two automatic zero-shot pipelines to generate data in OOD domains and to inpaint OOD objects for virtual reliability assessment. Our promising results encourage further collective investigations into this research problem, paving the way for synthetic system validation, especially in safety-critical applications. We summarize here our findings:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 488, + 479, + 654 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 138, + 488, + 479, + 558 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 488, + 479, + 558 + ], + "spans": [ + { + "bbox": [ + 138, + 488, + 479, + 558 + ], + "type": "inline_equation", + "content": "\\triangleright" + }, + { + "bbox": [ + 138, + 488, + 479, + 558 + ], + "type": "text", + "content": " Reliability Under Covariate Shifts: synthetic data can help assess the relative robustness of models in real-life covariate shifts, especially when shifts to the training condition are significant. Synthetic data can well complement real data in system validation, helping reduce the total operational cost. Pretrained models can be calibrated using synthetic data to better estimate prediction confidence in any shifted domains." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 560, + 479, + 654 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 560, + 479, + 654 + ], + "spans": [ + { + "bbox": [ + 138, + 560, + 479, + 654 + ], + "type": "inline_equation", + "content": "\\triangleright" + }, + { + "bbox": [ + 138, + 560, + 479, + 654 + ], + "type": "text", + "content": " Reliability Against OOD Objects: synthetic data is useful in both OOD testing and OOD training; however, the demands on synthetic data quality differ in these two cases. In OOD testing, the best result estimations are obtained with the most realistically inpainted data, which may require a certain amount of curation time for qualitative assessment. The curation task is not time-demanding and can be done quickly with a reasonable budget. On the other hand, for OOD training, no curation is actually needed to achieve improvements." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 231, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 231, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 231, + 101 + ], + "type": "text", + "content": "T.Loiseau et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 139, + 482, + 176 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 139, + 482, + 176 + ], + "spans": [ + { + "bbox": [ + 130, + 139, + 482, + 176 + ], + "type": "text", + "content": "This work is supported by ELSA - European Lighthouse on Secure and Safe AI funded by the European Union under grant agreement No. 101070617. We thank the authors of RELIS [14] for providing us their pretrained checkpoints." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 193, + 197, + 205 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 193, + 197, + 205 + ], + "spans": [ + { + "bbox": [ + 133, + 193, + 197, + 205 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 217, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 22, + "blocks": [ + { + "bbox": [ + 138, + 217, + 481, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 217, + 481, + 239 + ], + "spans": [ + { + "bbox": [ + 138, + 217, + 481, + 239 + ], + "type": "text", + "content": "1. clip-interrogator. https://github.com/pharmapsychotic/clip-interrogator (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 240, + 481, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 240, + 481, + 262 + ], + "spans": [ + { + "bbox": [ + 138, + 240, + 481, + 262 + ], + "type": "text", + "content": "2. Besnier, V., Jain, H., Bursuc, A., Cord, M., Pérez, P.: This dataset does not exist: training models from generated images. In: ICASSP (2020)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 262, + 481, + 283 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 262, + 481, + 283 + ], + "spans": [ + { + "bbox": [ + 138, + 262, + 481, + 283 + ], + "type": "text", + "content": "3. Blum, H., Sarlin, P.E., Nieto, J., Siegwart, R., Cadena, C.: The fishyscapes benchmark: Measuring blind spots in semantic segmentation (2021)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 284, + 481, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 284, + 481, + 316 + ], + "spans": [ + { + "bbox": [ + 138, + 284, + 481, + 316 + ], + "type": "text", + "content": "4. Chan, R., Lis, K., Uhlemeyer, S., Blum, H., Honari, S., Siegwart, R., Fua, P., Salzmann, M., Rottmann, M.: Segmentmeifyoucan: A benchmark for anomaly segmentation. In: NeurIPS (2021)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 316, + 481, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 316, + 481, + 348 + ], + "spans": [ + { + "bbox": [ + 138, + 316, + 481, + 348 + ], + "type": "text", + "content": "5. Geirhos, R., Narayanappa, K., Mitzkus, B., Thieringer, T., Bethge, M., Wichmann, F.A., Brendel, W.: Partial success in closing the gap between human and machine vision. NeurIPS (2021)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 349, + 481, + 371 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 349, + 481, + 371 + ], + "spans": [ + { + "bbox": [ + 138, + 349, + 481, + 371 + ], + "type": "text", + "content": "6. Guo, C., Pleiss, G., Sun, Y., Weinberger, K.Q.: On calibration of modern neural networks. In: ICML (2017)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 372, + 481, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 372, + 481, + 403 + ], + "spans": [ + { + "bbox": [ + 138, + 372, + 481, + 403 + ], + "type": "text", + "content": "7. Hariat, M., Laurent, O., Kazmierczak, R., Zhang, S., Bursuc, A., Yao, A., Franchi, G.: Learning to generate training datasets for robust semantic segmentation. In: WACV (2024)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 404, + 481, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 404, + 481, + 425 + ], + "spans": [ + { + "bbox": [ + 138, + 404, + 481, + 425 + ], + "type": "text", + "content": "8. He, R., Sun, S., Yu, X., Xue, C., Zhang, W., Torr, P., Bai, S., Qi, X.: Is synthetic data from generative models ready for image recognition? (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 426, + 481, + 457 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 426, + 481, + 457 + ], + "spans": [ + { + "bbox": [ + 138, + 426, + 481, + 457 + ], + "type": "text", + "content": "9. Hendrycks, D., Basart, S., Mazeika, M., Zou, A., Kwon, J., Mostajabi, M., Steinhardt, J., Song, D.: Scaling out-of-distribution detection for real-world settings. In: ICML (2022)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 459, + 481, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 459, + 481, + 491 + ], + "spans": [ + { + "bbox": [ + 138, + 459, + 481, + 491 + ], + "type": "text", + "content": "0. Hendrycks, D., Basart, S., Mu, N., Kadavath, S., Wang, F., Dorundo, E., Desai, R., Zhu, T., Parajuli, S., Guo, M., et al.: The many faces of robustness: A critical analysis of out-of-distribution generalization. In: ICCV (2021)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 491, + 481, + 512 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 491, + 481, + 512 + ], + "spans": [ + { + "bbox": [ + 138, + 491, + 481, + 512 + ], + "type": "text", + "content": "1. Hendrycks, D., Dietterich, T.: Benchmarking neural network robustness to common corruptions and perturbations. In: ICLR (2019)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 514, + 481, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 514, + 481, + 533 + ], + "spans": [ + { + "bbox": [ + 138, + 514, + 481, + 533 + ], + "type": "text", + "content": "2. Hendrycks, D., Gimpel, K.: A baseline for detecting misclassified and out-of-distribution examples in neural networks. In: ICLR (2017)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 535, + 481, + 556 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 535, + 481, + 556 + ], + "spans": [ + { + "bbox": [ + 138, + 535, + 481, + 556 + ], + "type": "text", + "content": "3. Jiang, L., Zhang, C., Huang, M., Liu, C., Shi, J., Loy, C.C.: Tsit: A simple and versatile framework for image-to-image translation. In: ECCV (2020)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 557, + 481, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 557, + 481, + 578 + ], + "spans": [ + { + "bbox": [ + 138, + 557, + 481, + 578 + ], + "type": "text", + "content": "4. de Jorge, P., Volpi, R., Torr, P.H., Rogez, G.: Reliability in semantic segmentation: Are we on the right track? In: CVPR (2023)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 138, + 578, + 481, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 578, + 481, + 610 + ], + "spans": [ + { + "bbox": [ + 138, + 578, + 481, + 610 + ], + "type": "text", + "content": "5. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., Dollar, P., Girshick, R.: Segment anything. arXiv:2304.02643 (2023)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 138, + 611, + 481, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 611, + 481, + 643 + ], + "spans": [ + { + "bbox": [ + 138, + 611, + 481, + 643 + ], + "type": "text", + "content": "6. Koh, P.W., Sagawa, S., Marklund, H., Xie, S.M., Zhang, M., Balsubramani, A., Hu, W., Yasunaga, M., Phillips, R.L., Gao, I., et al.: Wilds: A benchmark of in-the-wild distribution shifts. In: ICLR (2021)" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 138, + 643, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 643, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 643, + 481, + 665 + ], + "type": "text", + "content": "7. Le Moing, G., Vu, T.H., Jain, H., Pérez, P., Cord, M.: Semantic palette: Guiding scene generation with class proportions. In: CVPR (2021)" + } + ] + } + ], + "index": 21 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 171, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 171, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 171, + 91, + 447, + 102 + ], + "type": "text", + "content": "Reliability in Semantic Segmentation: Can We Use Synthetic Data?" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 117, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 23, + "blocks": [ + { + "bbox": [ + 133, + 117, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 117, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 133, + 117, + 480, + 138 + ], + "type": "text", + "content": "18. Li, D., Ling, H., Kim, S.W., Kreis, K., Fidler, S., Torralba, A.: Bigdatasetgan: Synthesizing imagenet with pixel-wise annotations. In: CVPR (2022)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 140, + 480, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 140, + 480, + 160 + ], + "spans": [ + { + "bbox": [ + 133, + 140, + 480, + 160 + ], + "type": "text", + "content": "19. Li, X., Chen, Y., Zhu, Y., Wang, S., Zhang, R., Xue, H.: Imagenet-e: Benchmarking neural network robustness via attribute editing. In: CVPR (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 161, + 480, + 193 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 161, + 480, + 193 + ], + "spans": [ + { + "bbox": [ + 133, + 161, + 480, + 193 + ], + "type": "text", + "content": "20. Liu, S., Zeng, Z., Ren, T., Li, F., Zhang, H., Yang, J., Li, C., Yang, J., Su, H., Zhu, J., et al.: Grounding dino: Marrying dino with grounded pre-training for open-set object detection. arXiv preprint arXiv:2303.05499 (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 194, + 480, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 194, + 480, + 215 + ], + "spans": [ + { + "bbox": [ + 132, + 194, + 480, + 215 + ], + "type": "text", + "content": "21. Lugmayr, A., Danelljan, M., Romero, A., Yu, F., Timofte, R., Gool, L.V.: Repaint: Inpainting using denoising diffusion probabilistic models. In: CVPR (2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 216, + 480, + 247 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 216, + 480, + 247 + ], + "spans": [ + { + "bbox": [ + 132, + 216, + 480, + 247 + ], + "type": "text", + "content": "22. Marathe, A., Ramanan, D., Walambe, R., Kotecha, K.: Wedge: A multi-weather autonomous driving dataset built from generative vision-language models. In: CVPRW (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 249, + 480, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 249, + 480, + 281 + ], + "spans": [ + { + "bbox": [ + 132, + 249, + 480, + 281 + ], + "type": "text", + "content": "23. Miller, J.P., Taori, R., Raghunathan, A., Sagawa, S., Koh, P.W., Shankar, V., Liang, P., Carmon, Y., Schmidt, L.: Accuracy on the line: on the strong correlation between out-of-distribution and in-distribution generalization. In: ICLR (2021)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 282, + 480, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 282, + 480, + 303 + ], + "spans": [ + { + "bbox": [ + 132, + 282, + 480, + 303 + ], + "type": "text", + "content": "24. Naeini, M.P., Cooper, G., Hauskrecht, M.: Obtaining well calibrated probabilities using bayesian binning. In: AAAI (2015)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 304, + 480, + 325 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 304, + 480, + 325 + ], + "spans": [ + { + "bbox": [ + 132, + 304, + 480, + 325 + ], + "type": "text", + "content": "25. Nayal, N., Yavuz, M., Henriques, J.F., Güney, F.: Rba: Segmenting unknown regions rejected by all. In: ICCV (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 326, + 480, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 326, + 480, + 357 + ], + "spans": [ + { + "bbox": [ + 132, + 326, + 480, + 357 + ], + "type": "text", + "content": "26. Ovadia, Y., Fertig, E., Ren, J., Nado, Z., Sculley, D., Nowozin, S., Dillon, J., Lakshminarayanan, B., Snoek, J.: Can you trust your model's uncertainty? evaluating predictive uncertainty under dataset shift. NeurIPS (2019)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 358, + 480, + 379 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 358, + 480, + 379 + ], + "spans": [ + { + "bbox": [ + 132, + 358, + 480, + 379 + ], + "type": "text", + "content": "27. Pinggera, P., Ramos, S., Gehrig, S., Franke, U., Rother, C., Mester, R.: Lost and found: detecting small road hazards for self-driving vehicles. In: IROS (2016)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 380, + 480, + 412 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 380, + 480, + 412 + ], + "spans": [ + { + "bbox": [ + 132, + 380, + 480, + 412 + ], + "type": "text", + "content": "28. Podell, D., English, Z., Lacey, K., Blattmann, A., Dockhorn, T., Müller, J., Penna, J., Rombach, R.: Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952 (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 413, + 480, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 413, + 480, + 445 + ], + "spans": [ + { + "bbox": [ + 132, + 413, + 480, + 445 + ], + "type": "text", + "content": "29. Prabhu, V., Yenamandra, S., Chattopadhyay, P., Hoffman, J.: Lance: Stress-testing visual models by generating language-guided counterfactual images. In: NeurIPS (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 446, + 480, + 467 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 446, + 480, + 467 + ], + "spans": [ + { + "bbox": [ + 132, + 446, + 480, + 467 + ], + "type": "text", + "content": "30. Recht, B., Roelofs, R., Schmidt, L., Shankar, V.: DoImagenet classifiers generalize toImagenet? In: ICLR (2019)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 468, + 480, + 489 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 468, + 480, + 489 + ], + "spans": [ + { + "bbox": [ + 132, + 468, + 480, + 489 + ], + "type": "text", + "content": "31. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: CVPR (2022)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 490, + 480, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 490, + 480, + 511 + ], + "spans": [ + { + "bbox": [ + 132, + 490, + 480, + 511 + ], + "type": "text", + "content": "32. Sakaridis, C., Dai, D., Van Gool, L.: Semantic foggy scene understanding with synthetic data. IJCV (2018)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 512, + 480, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 512, + 480, + 533 + ], + "spans": [ + { + "bbox": [ + 132, + 512, + 480, + 533 + ], + "type": "text", + "content": "33. Sakaridis, C., Dai, D., Van Gool, L.: Acdc: The adverse conditions dataset with correspondences for semantic driving scene understanding. In: ICCV (2021)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 534, + 480, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 534, + 480, + 566 + ], + "spans": [ + { + "bbox": [ + 132, + 534, + 480, + 566 + ], + "type": "text", + "content": "34. Sariyildiz, M.B., Alahari, K., Larlus, D., Kalantidis, Y.: Fake it till you make it: Learning transferable representations from synthetic imagenet clones. In: CVPR (2023)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 567, + 480, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 567, + 480, + 588 + ], + "spans": [ + { + "bbox": [ + 132, + 567, + 480, + 588 + ], + "type": "text", + "content": "35. Singh, A., Sarangmath, K., Chattopadhyay, P., Hoffman, J.: Benchmarking low-shot robustness to natural distribution shifts. In: ICCV (2023)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 132, + 589, + 480, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 589, + 480, + 610 + ], + "spans": [ + { + "bbox": [ + 132, + 589, + 480, + 610 + ], + "type": "text", + "content": "36. Taori, R., Dave, A., Shankar, V., Carlini, N., Recht, B., Schmidt, L.: Measuring robustness to natural distribution shifts in image classification. NeurIPS (2020)" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 132, + 611, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 611, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 611, + 480, + 632 + ], + "type": "text", + "content": "37. Teney, D., Lin, Y., Oh, S.J., Abbasnejad, E.: Id and ood performance are sometimes inversely correlated on real-world datasets. arXiv (2022)" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "type": "text", + "content": "38. Tran, D., Liu, J., Dusenberry, M.W., Phan, D., Collier, M., Ren, J., Han, K., Wang, Z., Mariet, Z., Hu, H., et al.: Plex: Towards reliability using pretrained large model extensions. arXiv (2022)" + } + ] + } + ], + "index": 22 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 231, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 231, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 231, + 101 + ], + "type": "text", + "content": "T.Loiseau et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 270 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 480, + 149 + ], + "type": "text", + "content": "39. Varma, G., Subramanian, A., Namboodiri, A., Chandraker, M., Jawahar, C.: Idd: A dataset for exploring problems of autonomous navigation in unconstrained environments. In: WACV (2019)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 150, + 481, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 150, + 481, + 182 + ], + "spans": [ + { + "bbox": [ + 132, + 150, + 481, + 182 + ], + "type": "text", + "content": "40. Wu, W., Zhao, Y., Shou, M.Z., Zhou, H., Shen, C.: Diffumask: Synthesizing images with pixel-level annotations for semantic segmentation using diffusion models. In: ICCV (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 183, + 480, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 183, + 480, + 215 + ], + "spans": [ + { + "bbox": [ + 132, + 183, + 480, + 215 + ], + "type": "text", + "content": "41. Yang, J., Wang, P., Zou, D., Zhou, Z., Ding, K., Peng, W., Wang, H., Chen, G., Li, B., Sun, Y., et al.: Openood: Benchmarking generalized out-of-distribution detection. NeurIPS (2022)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 216, + 480, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 216, + 480, + 237 + ], + "spans": [ + { + "bbox": [ + 132, + 216, + 480, + 237 + ], + "type": "text", + "content": "42. Zhang, L., Rao, A., Agrawala, M.: Adding conditional control to text-to-image diffusion models. In: ICCV (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 237, + 480, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 237, + 480, + 270 + ], + "spans": [ + { + "bbox": [ + 132, + 237, + 480, + 270 + ], + "type": "text", + "content": "43. Zhang, Y., Ling, H., Gao, J., Yin, K., Lafleche, J.F., Barriuso, A., Torralba, A., Fidler, S.: Datasetgan: Efficient labeled data factory with minimal human effort. In: CVPR (2021)" + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 171, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 171, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 171, + 91, + 447, + 102 + ], + "type": "text", + "content": "Reliability in Semantic Segmentation: Can We Use Synthetic Data?" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Reliable Spatial-Temporal Voxels For Multi-Modal Test-Time Adaptation/c906fc92-ebc1-45f1-bc7d-fa0568ba9181_content_list.json b/2024/Reliable Spatial-Temporal Voxels For Multi-Modal Test-Time Adaptation/c906fc92-ebc1-45f1-bc7d-fa0568ba9181_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..bcedbbbf73d98ec5d06deb0502e4eae33799c190 --- /dev/null +++ b/2024/Reliable Spatial-Temporal Voxels For Multi-Modal Test-Time Adaptation/c906fc92-ebc1-45f1-bc7d-fa0568ba9181_content_list.json @@ -0,0 +1,1657 @@ +[ + { + "type": "text", + "text": "Reliable Spatial-Temporal Voxels For Multi-Modal Test-Time Adaptation", + "text_level": 1, + "bbox": [ + 289, + 140, + 712, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Haozhi Cao $^{1}$ , Yuecong Xu $^{1,2}$ , Jianfei Yang $^{1,3}$ , Pengyu Yin $^{1}$ , Xingyu Ji $^{1}$ , Shenghai Yuan $^{1}$ , and Lihua Xie $^{1}$", + "bbox": [ + 223, + 210, + 779, + 243 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 School of Electrical and Electronic Engineering, Nanyang Technological University {haozhi002, jianfei.yang, pengyu001, xinyu001, shyuan, elhxie}@ntu.edu.sg", + "bbox": [ + 217, + 253, + 781, + 282 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ Department of Electrical and Computer Engineering, National University of Singapore yc.xu@nus.edu.sg", + "bbox": [ + 315, + 282, + 684, + 324 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3 School of Mechanical and Aerospace Engineering, Nanyang Technological University", + "bbox": [ + 215, + 324, + 784, + 339 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Multi-modal test-time adaptation (MM-TTA) is proposed to adapt models to an unlabeled target domain by leveraging the complementary multi-modal inputs in an online manner. Previous MM-TTA methods for 3D segmentation rely on predictions of cross-modal information in each input frame, while they ignore the fact that predictions of geometric neighborhoods within consecutive frames are highly correlated, leading to unstable predictions across time. To fulfill this gap, we propose ReLiable Spatial-temporal Voxels (Latte), an MM-TTA method that leverages reliable cross-modal spatial-temporal correspondences for multi-modal 3D segmentation. Motivated by the fact that reliable predictions should be consistent with their spatial-temporal correspondences, Latte aggregates consecutive frames in a slide window manner and constructs Spatial-Temopral (ST) voxels to capture temporally local prediction consistency for each modality. After filtering out ST voxels with high ST entropy, Latte conducts cross-modal learning for each point and pixel by attending to those with reliable and consistent predictions among both spatial and temporal neighborhoods. Experimental results show that Latte achieves state-of-the-art performance on three different MM-TTA benchmarks compared to previous MM-TTA or TTA methods. Visit our project site https://sites.google.com/view/eccv24-latte.", + "bbox": [ + 261, + 375, + 738, + 652 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Test-time adaptation $\\cdot$ Multi-modal learning $\\cdot$ 3D semantic segmentation", + "bbox": [ + 259, + 664, + 740, + 691 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 718, + 375, + 734 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3D semantic segmentation is a fundamental task to achieve various fully autonomous applications such as autonomous driving and robot navigation [9,14]. With the increasing urge for robust sensing, multi-modal sensors (e.g., cameras and LiDARs) are widely adopted in autonomous systems containing rich complementary information from different sensors. However, conventional deep-learning-based solutions lean on expensive point-wise annotations and perform", + "bbox": [ + 212, + 750, + 787, + 840 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/b8e4c8e876345fb3c69ae9096facc9e311366ae042e737c0a55381edfca585f2.jpg", + "image_caption": [ + "Fig. 1: Illustration of previous MM-TTA methods and our Latte. (a) visualizes the predictions from the previous state-of-the-art MM-TTA method [29] in consecutive frames on SemanticKITTI [1], where regions in white boxes exhibit noisy predictions when considering single-frame input. (b) presents the overall framework of Latte, which firstly obtains ST voxels for each modality given the merged input of consecutive frames within each slide window and then computes the modality-specific ST entropy $E^{2\\mathrm{D}}$ , $E^{3\\mathrm{D}}$ . ST entropy indicates the relative reliability of each voxel in each modality and therefore is utilized in the following adaptive cross-modal attending." + ], + "image_footnote": [], + "bbox": [ + 276, + 143, + 454, + 248 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/a500e2fb9aa47cb9e5e77e60915ad7cb7531a73e950dab13d544e87e16d75802.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 464, + 142, + 725, + 250 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "poorly on data from different domains due to the domain shift [6, 26, 33]. To resolve this problem, previous works have proposed Multi-Modal Unsupervised Domain Adaptation (MM-UDA) approaches [16, 18, 24, 44] that employ multimodal information to transfer knowledge learned from the labeled source domain to the unlabeled target domain without expensive human annotations.", + "bbox": [ + 212, + 401, + 784, + 476 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Nevertheless, the adaptation process of MM-UDA is completely offline, which requires full access to the source domain dataset and multiple training epochs. Despite its effectiveness, offline training is usually infeasible when encountering distribution shifts during the inference stage that require adapting models in real time. Inspired by Test-Time Adaptation (TTA) [22, 23], a recent work [29] proposes the first Multi-Modal Test-Time Adaptation (MM-TTA) method for 3D segmentation. Similar to the settings of TTA, MM-TTA prohibits access to any raw sample from the source domain dataset and adapts the model during the testing stage. This results in a quick adaptation scenario (i.e., one epoch for training only as in [29, 32]) that requires stable optimization in an online manner. To this end, MM-TTA methods must seek a reliable source of supervision signals.", + "bbox": [ + 212, + 477, + 787, + 643 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To fulfill the efficiency and stability of adaptation, the current state-of-the-art MM-TTA method [29] evaluates modal reliability by the prediction consistency between different experts in a frame-wise manner. However, due to the domain shift between the online frame and the pre-trained source dataset, the single-frame prediction is usually unstable. Existing approaches rely on pseudo labels to refine the single-frame prediction but still fail in some cases. As shown in Fig. 1a, the single-frame refinement strategy [29] suffers from noisy single-frame predictions on the car in the white rectangle across time. Such inconsistent predictions across time degenerate the segmentation performance and aggravate denoising stress in downstream tasks (e.g., semantic-based retrivial [46] and obstacle recognition). On one hand, these temporally unstable single-frame predictions could be wrongly regarded as reliable ones, propagating prediction noise to the other modality and causing error accumulation or even catastrophic forgetting [22,23].", + "bbox": [ + 212, + 643, + 787, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "H. Cao et al.", + "bbox": [ + 271, + 114, + 359, + 126 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "On the other hand, while previous methods [4, 41] alleviate this instability by utilizing the average predictions of multiple augmented frames instead, they are computationally expensive for online adaptation since the inference time grows linearly with the increasing number of augmented frames.", + "bbox": [ + 212, + 146, + 784, + 205 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "This work aims to efficiently suppress the single-frame instability for MM-TTA by exploiting the correlation between multi-frame correspondences. Specifically, we propose that 3D space can be efficiently divided into multiple voxels [33], each of which contains points and point-corresponding pixels located at the same geometric region. While captured at different timestamps, it can be observed that points or pixels located at the same voxel can be viewed as different observations on the same semantic objects [28]. Reliable predictions should therefore be certain and consistent within their corresponding voxels across time, whether we evaluate them from a temporally global perspective (i.e., all input frames) or a temporally local perspective (i.e., certain consecutive frames). By aggregating the predictions in a voxel-wise manner and re-evaluating their reliability, the single-frame prediction noise as in Fig. 1a can be effectively alleviated.", + "bbox": [ + 212, + 207, + 785, + 387 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Motivated by the above observation, we propose a novel MM-TTA method called ReLiable Spatial-temporal Voxels (Latte), which leverages the spatial-temporal correspondences within consecutive frames in a cross-modal manner as in Fig. 1b. Specifically, given frames of point clouds with their estimated poses, we regard points in the same voxel extracted from the merged point cloud frames as spatial-temporal correspondences. Different from previous works that merge all input frames [5,7] or regularize frame-to-frame consistency [28], Latte aggregates consecutive frames in a sliding window manner to estimate the temporally local prediction consistency through the proposed Spatial-Temporal voxels (ST voxels) and entropy (ST entropy). Based on the reliability estimated by ST entropy, we conduct cross-modal learning in an adaptive attending manner to reduce the contributions of predictions from the noisy modality. To better facilitate the progress of MM-TTA, we evaluate our Latte with more recent backbones (SegFormer [43] and SPVCNN [33]) on three different benchmarks, where Latte consistently outperforms previous TTA or MM-TTA methods.", + "bbox": [ + 212, + 388, + 785, + 614 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In summary, our contributions are three-fold. Firstly, we propose Latte, a novel MM-TTA method for multi-modal 3D segmentation, which is the first work that incorporates spatial-temporal correlations for MM-TTA segmentation to our best knowledge. Secondly, Latte efficiently extracts spatial-temporal correspondence through ST voxels and estimates their reliability by ST entropy, which is further incorporated to enhance adaptive cross-modal attending. Thirdly, our experimental results show that Latte outperforms previous state-of-the-art TTA and MM-TTA methods on three different benchmarks.", + "bbox": [ + 212, + 616, + 785, + 734 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Works", + "text_level": 1, + "bbox": [ + 215, + 760, + 395, + 777 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Test-Time Adaptation (TTA). TTA is proposed to mitigate the domain shift between the training data and the testing data in an online manner, which is firstly proposed by TENT [39]. Due to its practical and challenging setting,", + "bbox": [ + 212, + 794, + 785, + 840 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "Reliable Spatial-Temporal Voxels For Multi-Modal Test-Time Adaptation", + "bbox": [ + 241, + 114, + 730, + 130 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "TTA is attracting more and more attention. Previously proposed TTA methods attempt to address this task from different perspectives, such as entropy minimization [22, 23, 39], self-training with pseudo-labels [12, 40], and augmentation invariance [41, 48]. In addition to regularizing predictions, some methods propose to adapt networks from the feature level instead [21, 32], while more recent methods begin to consider different variants of TTA scenarios, such as continual TTA [4, 22, 41], non-i.i.d TTA [11], or the mix of aforementioned cases [47]. Different from source-free adaptation [19, 30] which also discards the access to the source domain yet requires multiple training epochs, the adaptation process of TTA is completely online, following the \"one-pass\" protocol as in [32]. While most existing TTA methods are designed for image classification or image segmentation without explicit temporal correlation, we argue temporal information in 3D segmentation can be effectively leveraged for TTA as Latte in this work.", + "bbox": [ + 212, + 146, + 787, + 343 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Multi-modal domain adaptation for 3D segmentation. To avoid expensive annotation costs and overcome the poor generalizability of fully supervised solutions, various multi-modal domain adaptation methods have been investigated. Specifically for MM-UDA, xMUDA [16] is the primary works that explore incorporating cross-modal learning with MM-UDA. It regards cross-modal prediction consistency and pseudo-labels as its supervision signals in the unlabeled target domain. Most subsequent works propose solutions to address its limitations from different perspectives, such as more diverse point-pixel correspondence [24, 44], procedures to mitigate domain gaps [18, 20], and alleviating class-imbalanced problem [3]. Since MM-UDA is infeasible when facing online domain shifts, some previous works [4, 29] develop different methods to address the domain gap in an online manner, including cross-modal pseudo-label generation [29] and adaptive modality attending [4]. A recent work [30] proposes a source-free domain adaptation method by estimating cross-modal prediction agreement, yet it requires multiple offline training epochs. Different from previous multi-modal domain adaptation methods that ignore temporal relationships, Latte couples spatial-temporal correspondence to mitigate the single-frame instability and achieve better online adaptation performance.", + "bbox": [ + 212, + 349, + 787, + 622 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Temporal processing of point clouds. Since point clouds are naturally in the form of temporally consecutive frames, previous methods have widely explored how to leverage such temporal information for different tasks. Specifically, scene flow [15,35,37,38] contains the movement of every point in the 3D world. Considering such point-wise computation expensive, previous works on different tasks (e.g., segmentation [5,6,28] and object detection [25]) usually interpret the temporal information from a simpler perspective. In terms of fully supervised learning, Choy et al. [6] propose the first 4D segmentation method by generalized sparse convolution kernels [13] and Fan et al. [8] turn to the fully connected learning scheme as in [26], while Xu et al. [45] extract long-term temporal information by an efficient memory bank. On the other hand, considering scenarios with insufficient or zero annotations, some works propose different strategies to couple spatial-temporal information without ground-truth labels. For instance, Chen et al. [5] proposes to guide voxelized spatial-temporal representation with", + "bbox": [ + 212, + 628, + 787, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "H. Cao et al.", + "bbox": [ + 271, + 114, + 359, + 127 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/84f048bd8d47220c0c1f832940d02649d5471164a1790f6cdf6db3e16604bd55.jpg", + "image_caption": [ + "Fig. 2: Overall structure of Latte. Taking a student prediction frame of one modality as the query input, our slide-window extraction searches its spatial-temporal correspondences through voxelization within a time window to establish the temporally local prediction consistency. ST voxels are then generated, where those with high ST entropy (larger than $\\alpha$ -quantile) are discarded as unreliable correspondences, while the others are leveraged for adaptive cross-modal learning by attending to the modality with lower ST entropy in a voxel-wise manner." + ], + "image_footnote": [], + "bbox": [ + 230, + 143, + 772, + 281 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "strong text embedding from CLIP [27] instead of labels in a multi-modal manner, while Saitor et al. [28] interpret such representation by nearest neighbor search in a frame-to-frame manner. In this work, we design a slide window aggregation and voxelization rather than aggregating all frames as in [5], which yields a better indication of the prediction reliability for each point.", + "bbox": [ + 212, + 422, + 784, + 500 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3 Methodology", + "text_level": 1, + "bbox": [ + 215, + 521, + 379, + 539 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Problem definition. During MM-TTA for 3D segmentation, both 2D RGB images and 3D point clouds are captured consecutively in the target domain $\\mathcal{T}$ , denoted as $\\mathbf{x}_{\\mathcal{T},t}^{2\\mathrm{D}} \\in \\mathbb{R}^{3 \\times H \\times W}$ and $\\mathbf{x}_{\\mathcal{T},t}^{3\\mathrm{D}} \\in \\mathbb{R}^{N \\times 4}$ , respectively, where $t$ represents the frame order and $H, W$ are the height and width of input images. The point-wise input contains the 3D coordinate $\\{x, y, z\\}$ and the point feature (e.g., intensity, reflectivity, etc.). Modality-specific networks are pre-trained on the labeled source domain before adaptation, denoted as $\\phi_{\\mathcal{T},t}^{\\mathrm{m}}(\\cdot)$ with $\\mathrm{m} \\in \\{2\\mathrm{D}, 3\\mathrm{D}\\}$ . During MM-TTA, both networks are initialized from the parameters pre-trained on the source domain. Following previous works [16,29], 3D points are projected to the 2D image based on the relative projection matrix to obtain cross-modal correspondences within each frame, resulting in the modality-specific predictions $\\mathbf{p}_{\\mathcal{T},t}^{\\mathrm{m}} = \\phi_{\\mathcal{T},t}^{\\mathrm{m}}(\\mathbf{x}_{\\mathcal{T},t}^{\\mathrm{m}})$ , $\\mathbf{p}_{\\mathcal{T},t}^{\\mathrm{m}} \\in \\mathbb{R}^{N \\times K}$ where $K$ denotes the number of semantic classes. Data from the labeled source domain are strictly inaccessible and we omit the target domain subscript $\\mathcal{T}$ by default.", + "bbox": [ + 212, + 551, + 787, + 763 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The multi-modal input of online 3D segmentation is temporally consecutive, which exhibits strong temporal relationships between frames. The geometric neighborhoods in consecutive frames can be viewed as the naturally various observations on the same semantic object that can be leveraged to stabilize noisy single-frame predictions. Motivated by this observation, we develop Latte,", + "bbox": [ + 212, + 765, + 787, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "Reliable Spatial-Temporal Voxels For Multi-Modal Test-Time Adaptation", + "bbox": [ + 241, + 114, + 730, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "which discovers reliable spatial-temporal (ST) correspondences to improve both optimization stability and prediction consistency during MM-TTA. As shown in Fig. 2, given consecutive frames of 2D images and 3D point clouds as input, Latte firstly extracts their frame-wise predictions from both student and teacher networks (Sec. 3.1). These frame-wise predictions are then passed to the slide-window extraction and voxelization (Sec. 3.2). ST voxels and ST entropy are subsequently computed indicating the prediction consistency and certainty (Sec. 3.3). The final ST entropy is leveraged for cross-modal attending to alleviate the noise from unreliable modality-specific predictions (Sec. 3.4).", + "bbox": [ + 212, + 146, + 787, + 282 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.1 Frame-wise Predictions from Students and Teachers", + "text_level": 1, + "bbox": [ + 214, + 301, + 689, + 315 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Before exploiting spatial-temporal correspondence, we first generate the framewise predictions from each modality. Motivated by the fact that moving average models can provide more stable online predictions [4,34,41], Latte utilizes weight-averaged teacher models and fast student models for each modality as in previous works [29, 41], denoted as $\\tilde{\\phi}_t^{\\mathrm{m}}(\\cdot)$ and $\\phi_t^{\\mathrm{m}}(\\cdot)$ , respectively. Given each input frame $\\mathbf{x}_t^{\\mathrm{m}}$ , the student predictions $\\mathbf{p}_t^{\\mathrm{m}}$ and teacher predictions $\\tilde{\\mathbf{p}}_t^{\\mathrm{m}}$ are computed as:", + "bbox": [ + 212, + 321, + 782, + 412 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {\\mathbf {p}} _ {t} ^ {\\mathrm {m}} = \\tilde {\\phi} _ {t} ^ {\\mathrm {m}} \\left(\\mathbf {x} _ {t} ^ {\\mathrm {m}}\\right), \\mathbf {p} _ {t} ^ {\\mathrm {m}} = \\phi_ {t} ^ {\\mathrm {m}} \\left(\\mathbf {x} _ {t} ^ {\\mathrm {m}}\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 395, + 420, + 784, + 436 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where the teacher model $\\tilde{\\phi}_t^{\\mathrm{m}}$ is initialized from the source domain pre-trained models. In the following process, the teacher's predictions of one modality are regarded as the cross-modal guidance of student predictions from the other modality. In practice, the gradient propagation of teacher model predictions is discarded to prevent directly updating teacher models.", + "bbox": [ + 212, + 443, + 787, + 518 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.2 Slide Window Frame Aggregation and Voxelization", + "text_level": 1, + "bbox": [ + 214, + 537, + 683, + 553 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Our goal is to efficiently leverage spatial-temporal correspondences between predictions and seek reliability estimation through prediction consistency within correspondences. Establishing spatial-temporal correspondences in 3D space has been explored in other deep learning tasks, where they mainly consider temporally global relationships of all input frames [5,28] or frame-to-frame correspondences [28]. However, they both have some limitations in terms of consistency and certainty evaluation: the former can not highlight the risk of some inconsistent predictions within a short time window (i.e., temporally local inconsistency), while the latter suffers from an insufficient number of correspondences.", + "bbox": [ + 212, + 558, + 787, + 694 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To effectively empower spatial-temporal information with Latte, we propose a slide window frame aggregation that focuses on temporally local correspondences that lie within a time window to evaluate temporally local consistency. Specifically, given a student prediction frame $\\mathbf{p}_i^{\\mathrm{m}}$ with the temporal index $i$ as the query, the correspondence search is conducted within a time window of consecutive frames denoted as $\\{j||j - i|\\leq w_t\\}$ , where $w_{t}$ is the pre-defined time window size. The merged point cloud $\\hat{\\mathbf{x}}_i^{3\\mathrm{D}}\\in \\mathbb{R}^{\\hat{N}\\times 3}$ (point features are omitted for simplicity) for correspondence search is then formulated as:", + "bbox": [ + 212, + 694, + 787, + 816 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathbf {x}} _ {i} ^ {\\mathrm {3 D}} = \\operatorname {c a t} \\left( \\right.\\left\\{ \\right.\\mathbf {T} _ {j \\rightarrow i} * \\mathbf {x} _ {j} ^ {\\mathrm {3 D}} \\left. \\right| | j - i | \\leq w _ {t} \\}\\left. \\right), \\quad \\mathbf {T} _ {j \\rightarrow i} = \\mathbf {T} _ {i} ^ {- 1} \\mathbf {T} _ {j}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 302, + 823, + 784, + 842 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "H. Cao et al.", + "bbox": [ + 271, + 114, + 359, + 127 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\mathbf{T}_i$ is the estimated pose at frame $i$ and so as $\\mathbf{T}_j$ , which can be easily obtained online through off-the-shelf SLAM algorithms [17,36]. $\\operatorname{cat}(\\cdot)$ is the concatenation operation along the point number dimension while $*$ denotes the pose transformation. The voxelization is then performed on $\\hat{\\mathbf{x}}_i^{3D}$ , formulated as:", + "bbox": [ + 212, + 146, + 787, + 208 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {v} _ {i} ^ {\\mathrm {3 D}} = \\mathcal {V} _ {\\mathbf {s}} \\left(\\hat {\\mathbf {x}} _ {i} ^ {\\mathrm {3 D}}\\right), \\quad \\mathbf {v} _ {i} ^ {\\mathrm {3 D}} \\in \\mathbb {R} ^ {N _ {\\mathrm {v}} \\times 3}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 318, + 215, + 784, + 234 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {M} _ {i} ^ {\\mathrm {3 D}} = \\left\\{k | \\forall g \\in [ 1, \\hat {N} ], \\left\\lfloor \\hat {\\mathbf {x}} _ {i, g} ^ {\\mathrm {3 D}} / \\mathbf {s} \\right\\rfloor = \\mathbf {v} _ {i, k} ^ {\\mathrm {3 D}} \\right\\}, \\quad \\mathbf {M} _ {i} ^ {\\mathrm {3 D}} \\in \\mathbb {R} ^ {\\hat {N}}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 310, + 238, + 784, + 258 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $\\mathbf{v}_i^{3\\mathrm{D}}$ is the extracted voxels and $\\nu_{s}$ is the voxelization operation with a voxel size s. After voxelization, points located in the same voxel can be regarded as correspondences since they are geometric neighborhoods, where the hash table $\\mathbf{M}_i^{3\\mathrm{D}}$ that maps each voxel back to its containing points (e.g., $\\mathbf{v}_{i,k}^{3\\mathrm{D}}$ to $\\hat{\\mathbf{x}}_{i,g}^{3\\mathrm{D}}$ as in Eq. (4)) is preserved for the subsequent correspondence mapping. In this way, it provides more diverse neighborhood selection strategies with temporally local correspondences for consistency and certainty evaluation compared to globally merging all frames in one single step. In practice, the slide window aggregation and voxelization are conducted iteratively for each frame in the input batch.", + "bbox": [ + 214, + 268, + 787, + 405 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.3 Spatial-Temporal Voxels and Entropy", + "text_level": 1, + "bbox": [ + 214, + 426, + 571, + 441 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Given the geometric correspondences captured by voxelization, our goal is to emphasize the predictions with correspondence consistency within each voxel. After slide window aggregation and voxelization, there inevitably exist some unreliable correspondences of two different types: (i) voxels that contain different types of semantic objects (e.g., those located on the contact surface between cars and roads) and (ii) voxels with highly inconsistent predictions due to uncertain predictions across time. These unreliable correspondences contain high uncertainty and inconsistency, which could harm the online optimization of models if they as references. To alleviate the side-effect introduced by such unreliable correspondences, we propose ST voxels that leverage ST entropy for voxel-wise reliability evaluation and cross-modal attending.", + "bbox": [ + 212, + 449, + 787, + 614 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Specifically, an ST voxel contains two components, including the query and the reference, where the query is encouraged to be consistent with the reference. Considering that the moving average teacher can usually generate more stable predictions [4, 41], we regard the single-frame student predictions as the query while multi-frame teacher predictions as our reference, forming ST voxels denoted as $\\mathbf{v}_i^{\\mathrm{ST}}$ . Without losing generality, here we take a single ST voxel indexed by $k$ in frame $i$ as an example, denoted as $\\mathbf{v}_{i,k}^{\\mathrm{ST}} = \\{\\mathbf{p}_{\\mathrm{q}}^{\\mathrm{m}},\\mathbf{p}_{\\mathrm{r}}^{\\mathrm{m}}\\}$ , where $\\mathbf{p}_{\\mathrm{q}}^{\\mathrm{m}}$ and $\\mathbf{p}_{\\mathrm{r}}^{\\mathrm{m}}$ are the point-wise student predictions at the query frame $i$ and teacher predictions in the frame searching range $j$ , respectively, formulated as:", + "bbox": [ + 212, + 616, + 787, + 752 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {p} _ {\\mathrm {q}} ^ {\\mathrm {m}} = \\left\\{\\mathbf {p} _ {i, g} ^ {\\mathrm {m}} \\mid \\left\\lfloor \\mathbf {x} _ {i, g} ^ {\\mathrm {3 D}} / \\mathbf {s} \\right\\rfloor = \\mathbf {v} _ {i, k} ^ {\\mathrm {3 D}} \\right\\}, \\quad \\mathbf {p} _ {\\mathrm {q}} ^ {\\mathrm {m}} \\in \\mathbb {R} ^ {N _ {\\mathrm {q}} \\times K} \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 303, + 761, + 784, + 780 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {p} _ {\\mathrm {r}} ^ {\\mathrm {m}} = \\left\\{\\tilde {\\mathbf {p}} _ {j, g} ^ {\\mathrm {m}} \\mid \\left| \\mathbf {x} _ {j, g} ^ {3 \\mathrm {D}} / \\mathbf {s} \\right| = \\mathbf {v} _ {i, k} ^ {3 \\mathrm {D}}, | j - i | \\leq w _ {t} \\right\\}, \\quad \\mathbf {p} _ {\\mathrm {r}} ^ {\\mathrm {m}} \\in \\mathbb {R} ^ {N _ {\\mathrm {r}} \\times K}. \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 303, + 782, + 784, + 801 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The reliability of each voxel is then evaluated as the Shannon's entropy [42] of the average teacher predictions, where the unreliable ST voxels are then filtered", + "bbox": [ + 214, + 809, + 787, + 839 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "Reliable Spatial-Temporal Voxels For Multi-Modal Test-Time Adaptation", + "bbox": [ + 241, + 114, + 732, + 130 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "out given a pre-defined quantile $\\alpha$ as in Fig. 2, denoted as:", + "bbox": [ + 212, + 146, + 640, + 162 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\nE _ {i, k} ^ {\\mathrm {m}} = - \\sum_ {c} ^ {K} \\bar {\\mathbf {p}} _ {\\mathrm {r}, c} ^ {\\mathrm {m}} \\log \\bar {\\mathbf {p}} _ {\\mathrm {r}, c} ^ {\\mathrm {m}}, \\quad \\bar {\\mathbf {p}} _ {\\mathrm {r}} ^ {\\mathrm {m}} = \\sum_ {n = 1} ^ {N _ {\\mathrm {r}}} \\psi \\left(\\mathbf {p} _ {\\mathrm {r}, n} ^ {\\mathrm {m}}\\right) / N _ {\\mathrm {r}}, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 330, + 171, + 785, + 210 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\nh _ {i, k} = \\left\\{ \\begin{array}{l l} 0, & \\text {i f} E _ {i, k} ^ {m} > Q ^ {m} (\\alpha) \\\\ 1, & \\text {i f} E _ {i, k} ^ {m} \\leq Q ^ {m} (\\alpha) \\end{array} , \\right. \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 331, + 214, + 785, + 253 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $E_{i,k}^{\\mathrm{m}}$ is regarded as the ST entropy of voxel $k$ for modality $\\mathrm{m}$ , indicating the modal-specific prediction reliability within this voxel. $Q^{m}(\\alpha)$ is the $\\alpha$ -quantile of all ST entropy in modality $\\mathrm{m}$ in the same batch and $\\psi(\\cdot)$ denotes the Softmax function along the class dimension.", + "bbox": [ + 214, + 260, + 787, + 321 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "3.4 ST Voxel Aided Cross-Modal Learning", + "text_level": 1, + "bbox": [ + 214, + 340, + 583, + 358 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Each modality has its pros and cons under different conditions and cross-modal learning for online scenarios like MM-TTA should therefore possess an adaptive mechanism that can attend to \"pros\" and meanwhile suppress the \"cons\". To achieve that, Latte performs cross-modal learning by attending to the modality with more consistent and certain predictions from a spatial-temporal perspective. Specifically, taking ST voxel $\\mathbf{v}_{i,k}^{\\mathrm{ST}}$ and its corresponding ST entropy $E_{i,k}^{\\mathrm{m}}$ as an example, the voxel-wise cross-modal attending weights and the weighted cross-modal consistency loss $\\mathcal{L}_{i,k}^{\\mathrm{xM}}$ are formulated as:", + "bbox": [ + 212, + 364, + 787, + 487 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\nw _ {\\mathrm {v}} ^ {\\mathrm {2 D}} = \\frac {\\exp \\left(E _ {i , k} ^ {\\mathrm {2 D}}\\right)}{\\exp \\left(E _ {i , k} ^ {\\mathrm {2 D}}\\right) + \\exp \\left(E _ {i , k} ^ {\\mathrm {3 D}}\\right)}, w _ {\\mathrm {v}} ^ {\\mathrm {3 D}} = 1 - w _ {\\mathrm {v}} ^ {\\mathrm {2 D}}, \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 326, + 496, + 785, + 532 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {i, k} ^ {\\mathrm {x M}} = w _ {\\mathrm {v}} ^ {\\mathrm {2 D}} D _ {\\mathrm {K L}} \\left(\\bar {\\mathbf {p}} _ {\\mathrm {q}} ^ {\\mathrm {3 D}} \\| \\bar {\\mathbf {p}} _ {\\mathrm {r}} ^ {\\mathrm {2 D}}\\right) + w _ {\\mathrm {v}} ^ {\\mathrm {3 D}} D _ {\\mathrm {K L}} \\left(\\bar {\\mathbf {p}} _ {\\mathrm {q}} ^ {\\mathrm {2 D}} \\| \\bar {\\mathbf {p}} _ {\\mathrm {r}} ^ {\\mathrm {3 D}}\\right), \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 326, + 535, + 784, + 554 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $D_{\\mathrm{KL}}(\\cdot)$ represents the KL-divergence between two probability. $\\bar{\\mathbf{p}}_{\\mathrm{q}}^{3\\mathrm{D}}$ is the average query predictions in the ST voxel, similarly computed as in Eq. (7).", + "bbox": [ + 212, + 561, + 787, + 593 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The cross-modal attending is further extended from the voxel level to the point level for better online predictions and cross-modal pseudo-label generation. Specifically, the ST entropy is propagated from the ST voxel to its reference point-wise prediction, indicating its confidence and consistency within its spatial-temporal neighborhoods. If one's ST entropy has been filtered out as in Eq. (8), its ST entropy would fall back to the point-level entropy. For arbitrary pointwise predictions $\\mathbf{p}^{\\mathrm{m}}\\in \\mathbf{p}_{\\mathrm{r}}^{\\mathrm{m}}$ , the cross-modal predictions $y^{\\mathrm{xM}}$ are formulated as:", + "bbox": [ + 212, + 594, + 787, + 700 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {E} _ {\\mathrm {r}} ^ {\\mathrm {m}} = h _ {i, k} E _ {i, k} ^ {\\mathrm {m}} - \\left(1 - h _ {i, k}\\right) \\sum_ {c} ^ {K} \\mathbf {p} _ {c} ^ {\\mathrm {m}} \\log \\mathbf {p} _ {c} ^ {\\mathrm {m}}, \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 334, + 720, + 785, + 760 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\nw _ {\\mathrm {p}} ^ {\\mathrm {2 D}} = \\frac {\\exp \\left(\\hat {E} ^ {\\mathrm {2 D}}\\right)}{\\exp \\left(\\hat {E} ^ {\\mathrm {2 D}}\\right) + \\exp \\left(\\hat {E} ^ {\\mathrm {3 D}}\\right)}, w _ {\\mathrm {p}} ^ {\\mathrm {3 D}} = 1 - w _ {\\mathrm {p}} ^ {\\mathrm {2 D}}, \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 328, + 763, + 784, + 800 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {p} ^ {\\mathrm {x M}} = w _ {\\mathrm {p}} ^ {\\mathrm {2 D}} \\mathbf {p} ^ {\\mathrm {2 D}} + w _ {\\mathrm {p}} ^ {\\mathrm {3 D}} \\mathbf {p} ^ {\\mathrm {3 D}}, y ^ {\\mathrm {x M}} = \\underset {c} {\\arg \\max } \\mathbf {p} ^ {\\mathrm {x M}}. \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 328, + 801, + 784, + 827 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "H. Cao et al.", + "bbox": [ + 271, + 114, + 359, + 127 + ], + "page_idx": 7 + }, + { + "type": "code", + "sub_type": "algorithm", + "code_caption": [ + "Algorithm 1: Adaptation and Online Prediction Process of Latte" + ], + "code_body": "Target Domain: Multi-modal input $\\mathcal{X}_{\\mathcal{T}} = \\{(\\mathbf{x}_{t}^{2\\mathrm{D}},\\mathbf{x}_{t}^{3\\mathrm{D}})\\}$ \nInit Model: Teacher $\\tilde{\\phi}_t^{\\mathrm{m}}$ , student $\\phi_t^{\\mathrm{m}}$ $\\mathfrak{m}\\in \\{2\\mathrm{D},3\\mathrm{D}\\}$ , Slide window size $w_{t}$ \nfor $\\pmb {x} = \\{(x_{t}^{2D},x_{t}^{3D})\\}_{t = t_0 + 1}^B\\in \\mathcal{X}_T$ do 1. Get student and teacher predictions $\\tilde{\\mathbf{p}}_t^{\\mathrm{m}},\\mathbf{p}_t^{\\mathrm{m}}$ by Eq. (1) for $i\\in [t_0 + 1,t_0 + B]$ do 2. Get $\\hat{\\mathbf{x}}_i^{3\\mathrm{D}},\\mathbf{v}_i^{3\\mathrm{D}},\\mathbf{M}_i^{3\\mathrm{D}}$ within the time window of size $w_{t}$ by Eq. (2)-(4) 3. Compute ST voxels $\\mathbf{v}_i^{\\mathrm{ST}}$ and ST entropy $E_i^{\\mathrm{m}}$ by Eq. (5)-(8) 4. Compute $\\mathcal{L}_i^{\\mathrm{xM}}$ by Eq. (9)-(10) end 5. Compute $\\hat{E}_{\\mathrm{r}}^{\\mathrm{m}}$ and get $\\mathbf{y}_t^{\\mathrm{xM}}$ as predictions by Eq. (11)-(13) 6. Compute overall loss by Eq. (14) and update models by Eq. (15) \nend", + "bbox": [ + 212, + 166, + 743, + 344 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "3.5 Online Predictions and Optimization", + "text_level": 1, + "bbox": [ + 215, + 358, + 565, + 373 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "As shown in Fig. 2, given a batch of $B$ consecutive frames $\\{\\mathbf{x}_t^{\\mathrm{m}}\\}_{t = t_0 + 1}^{t_0 + B}$ as input, we first extract the student and teacher predictions through Eq. (1). Subsequently, frames are aggregated and voxelized as in Eq. (2), resulting in a batch of merged point cloud $\\{\\hat{\\mathbf{x}}_t^{3\\mathrm{D}}\\}_{t = t_0 + 1}^{t_0 + B}$ and voxels $\\{\\mathbf{v}_t^{3\\mathrm{D}}\\}_{t = t_0 + 1}^{t_0 + B}$ . The ST voxels and ST entropy computation process is then applied to each frame of the merged point cloud and voxels. Due to the overlap of the slide window, multiple ST entropy values could be propagated to the same point prediction, where we empirically take the average value of all ST entropy received to proceed Eq. (11). The overall loss function and the updating scheme can be formulated as:", + "bbox": [ + 212, + 391, + 787, + 529 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\sum_ {t} \\mathcal {F} \\left(\\mathbf {p} _ {t} ^ {\\mathrm {m}}, \\mathbf {y} _ {t} ^ {\\mathrm {x M}}\\right) + \\frac {\\lambda_ {\\mathrm {x M}}}{B} \\sum_ {t} \\sum_ {k} \\mathcal {L} _ {t, k} ^ {\\mathrm {x M}}, \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 357, + 542, + 785, + 578 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {\\theta} _ {t} ^ {\\mathrm {m}} = \\lambda_ {s} \\tilde {\\theta} _ {t - 1} ^ {\\mathrm {m}} + (1 - \\lambda_ {s}) \\theta_ {t} ^ {\\mathrm {m}}, \\tag {15}\n$$\n", + "text_format": "latex", + "bbox": [ + 349, + 580, + 784, + 599 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where $\\mathcal{F}(\\cdot)$ is the cross-entropy function and $\\mathbf{y}_t^{\\mathrm{xM}}$ is the point-wise cross-modal pseudo-label frame from Eq. (13), which is also regarded as our cross-modal prediction for evaluation. $\\tilde{\\theta}_t^{\\mathrm{m}}$ and $\\theta_t^{\\mathrm{m}}$ are the model parameters of teacher and student models, respectively, where $\\lambda_s$ is the momentum update coefficient. $\\lambda_{\\mathrm{xM}}$ is pre-defined coefficient for the cross-modal consistency loss. Our adaptation and inference process is summarized in Algo. 1", + "bbox": [ + 212, + 614, + 787, + 708 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4 Experimental Results", + "text_level": 1, + "bbox": [ + 215, + 738, + 460, + 756 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "In this section, we present our thorough experimental results on three different benchmarks. Details of benchmarks, backbones, and settings are first present in Sec. 4.1. The main results are then illustrated in Sec. 4.2, followed by detailed ablation studies and qualitative results in Sec. 4.3.", + "bbox": [ + 212, + 779, + 787, + 839 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "Reliable Spatial-Temporal Voxels For Multi-Modal Test-Time Adaptation", + "bbox": [ + 241, + 113, + 730, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 114, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.1 Benchmarks and settings", + "text_level": 1, + "bbox": [ + 215, + 146, + 470, + 162 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Benchmark Details. To thoroughly investigate the effectiveness of Latte, we conduct our experiments on three different MM-TTA benchmarks: (i) USA-to-Singapore (U-to-S), (ii) A2D2-to-SemanticKITTI (A-to-S), and (iii) Synthia-to-SemanticKITTI (S-to-S). Specifically, U-to-S is developed from NuScenes-LiDARSeg [2] dataset, where the domain gap is mainly credited to the infrastructure difference between countries. Same as [16], the source domain and the target domain data are selected by filtering country keywords based on the data recording description in NuScenes. Different from previous UDA methods [16,24] utilizing self-defined class mappings, we follow the official mapping in NuScenes-LiDARSeg, forming a 16-class segmentation benchmark. The remaining two benchmarks follow a similar setting as in [29], where the domain gap of A-to-S [1,10] lies in different LiDAR mounting positions and image resolutions while the one of S-to-S [1] lies in different patterns between synthetic and real data for both modalities. For A-to-S, we adopt the same class-mapping as in [16,29] which leads to a 10-class benchmark, while we design our class-mapping for S-to-S since its original class map has not been revealed previously [29], forming a 9-class benchmark. More details are presented in our appendix.", + "bbox": [ + 212, + 174, + 787, + 430 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Baseline methods. Previous TTA and MM-TTA methods are included in comparison with Latte. For TTA methods, we compare Latte with point-wise pseudolabels with filtering (PsLabel), TENT [39], ETA [22], SAR [23] with and without restoring (denoted as SAR-rs and SAR, respectively). MMTTA [29] as well as its online version of xMUDA [16] (including pure xMUDA and pseudo-label version xMUDA+PL) are additionally included as the previous SOTA methods for comparison. Considering that EATA [22] with Fisher regularizer requires pre-access to the target domain samples, we compare Latte with ETA, an official variant of EATA without this regularizer for a fair comparison. We further present a performance lower bound by directly testing with source pre-trained models (i.e., Source only) and an upper bound by online adapting networks with ground-truth labels from the target domain for one epoch (i.e., Oracle TTA).", + "bbox": [ + 212, + 431, + 787, + 611 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Implementation details. With the rapid progress in 2D and 3D segmentation, we employ all baseline methods and Latte with more recent SegFormer [43] (SegFormer-B1) and SPVCNN [33] (SPVCNN-cr1) as their 2D and 3D backbones, respectively, to investigate the improvement of existing TTA methods when combined with advanced networks. The pre-training procedures on the source domain follow a similar setting as in [16], except for 2D SegFormer which is trained with a base learning rate of 6E-5 with an AdamW optimizer and Poly scheduler as in [43]. For Latte, we maintain the same learning rate and optimizer as in the pre-training stage, while the scheduler is disabled. For our voxelization, we leverage the off-the-shelf SLAM algorithm KISS-ICP [36] to generate poses and utilize a window size $w_{t}$ of 3 with a voxel size of $0.2\\mathrm{m}$ , while the quantile $\\alpha$ and coefficient $\\lambda_{\\mathrm{xM}}$ are set to 0.9 and 0.3, respectively. $\\lambda_{s}$ is empirically set to 0.99 as in [4,31]. For all methods, we strictly follow the one-pass protocol to first evaluate the predictions of the input batch and then update the networks, where we only update the trainable parameters in normalization layers. For all base", + "bbox": [ + 212, + 613, + 787, + 840 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "H. Cao et al.", + "bbox": [ + 271, + 114, + 359, + 127 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/2da61598510481053582ff88a330dba2fe43ac7a14fc57167eccd5491918fd25.jpg", + "table_caption": [ + "Table 1: Performance (mIoU) comparison of Latte. Latte outperforms all previous SOTA methods on the cross-modal metric (xM) across three benchmarks. Here \"MM\" denotes whether the method contains multi-modal interaction or not. Cross-modal predictions are computed as the Softmax average of modal outputs except for Latte. \"Avg\" present the average xM performance across three benchmarks." + ], + "table_footnote": [], + "table_body": "
MethodPublicationMMU-to-SxMA-to-SxMS-to-SAvg
2D3D2D3D2D3D
Source only-X31.441.143.947.417.944.323.436.448.2
Oracle TTA-X38.745.550.349.161.362.536.054.554.6
TENT [39]ICML-21X36.836.141.143.344.449.122.435.537.5
ETA [22]ICML-22X36.734.743.743.242.549.722.430.632.9
SAR [23]ICLR-23X36.637.243.943.042.750.320.731.332.9
SAR-rs [23]ICLR-23X36.635.543.943.143.850.124.225.128.8
xMUDA [16]PAMI-22✓19.422.924.213.133.032.212.214.814.1
xMUDA+PL [16]PAMI-22✓36.238.243.042.848.350.930.934.136.3
MMTTA [29]CVPR-22✓37.241.545.444.551.753.727.535.135.5
PsLabel-✓37.735.841.846.744.350.029.233.335.3
Latte (Ours)-✓37.441.046.046.152.654.333.239.341.6
", + "bbox": [ + 220, + 227, + 781, + 364 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "lines, we conduct a parameter search and report their best results. We utilize the mean intersection over union (mIoU) as our evaluation metric. All experiments are conducted with PyTorch on a single RTX 3090. More details are presented in our appendix.", + "bbox": [ + 212, + 373, + 789, + 434 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4.2 Overall Results", + "text_level": 1, + "bbox": [ + 215, + 460, + 388, + 474 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Tab. 1 presents the overall results of our Latte on three benchmarks compared with previous SOTA methods. In terms of average cross-modal predictions (Avg), Latte surpasses all previous TTA or MM-TTA methods, with a significant relative improvement of more than $5.3\\%$ . For the less challenging U-to-S and A-to-S (i.e., gap between Oracle TTA and best TTA less than $10\\%$ ), Latte achieves consistent relative improvements of $1.3\\%$ on U-to-S and $1.1\\%$ on A-to-S compared to MMTTA, respectively. Most existing methods except for xMUDA [16] can achieve superior performance compared to source-only results when one modality is reliable as in U-to-S (3D prevails) and A-to-S (2D prevails).", + "bbox": [ + 212, + 491, + 787, + 628 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "In terms of the most challenging S-to-S, Latte exhibits significant performance improvement compared to previous SOTA MM-TTA methods such as MMTTA [29] and xMUDA+PL [16], with a relative gap of $14.6\\%$ . In fact, all previous methods except for our Latte perform inferiorly compared to source pre-trained models (Source only), due to the significant domain gaps caused by the non-trivial pattern difference in both modalities between synthetic and real data. In this case, the instability of single-frame predictions becomes exacerbated, leading to performance degeneration in MMTTA which mainly relies on single-frame refinement. This justifies the effectiveness of Latte under either challenging or less challenging TTA scenarios. Another interesting observation is that almost all cross-modal methods outperform other single-modal TTA methods, e.g., even the simplest PsLabel can outperform recent TTA methods like SAR [23]. This reveals the non-trivial benefits of cross-modal learning for TTA with segmentation. More comparisons about efficiency are included in appendix.", + "bbox": [ + 212, + 628, + 787, + 840 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "Reliable Spatial-Temporal Voxels For Multi-Modal Test-Time Adaptation", + "bbox": [ + 241, + 113, + 730, + 130 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/124df0ad9b6644ef2b50a453f9c3e161ebbea36301c3cd451f48609daf2f03fd.jpg", + "table_caption": [ + "Table 2: Ablation studies about the effectiveness of cross-modal learning. When disabling $\\mathbf{y}^{\\mathrm{xM}}$ , we utilize the average Softmax of all modalities as the pseudo-labels as a replacement. In experiment No. 8, we replace the ST entropy (ST ety) with point-wise entropy in Eq. (13) and set $w_{\\mathrm{v}}^{\\mathrm{m}}$ in Eq. (10) to 0.5. Here $\\mathcal{L}_{32}^{\\mathrm{xM}}$ and $\\mathcal{L}_{23}^{\\mathrm{xM}}$ refer to the former half and the latter half in Eq. 10, respectively." + ], + "table_footnote": [], + "table_body": "
No.Description\\( y^{xM} \\)\\( \\mathcal{L}_{23}^{xM} \\)\\( \\mathcal{L}_{32}^{xM} \\)U-to-SA-to-SS-to-S
2D3DxM2D3DxM2D3DxM
0Source only---31.441.143.947.417.944.323.436.438.2
1only ST ety label✓36.834.543.743.342.249.922.430.633.3
2only ST 2D→3D✓32.534.541.041.742.248.227.730.632.9
3only ST 3D→2D✓36.839.642.643.350.248.422.426.725.6
4ST ety label + ST 2D→3D✓✓37.535.841.043.647.549.831.036.038.4
5ST ety label + ST 3D→2D✓✓37.440.945.946.252.153.432.538.739.7
6ST 2D↔3D✓✓36.840.343.844.047.051.230.435.637.7
7w/o percentile in Eq. (8)✓✓✓37.440.345.445.751.453.232.939.041.7
8Point ety vs. ST ety✓✓✓37.040.645.145.151.452.731.636.738.2
9Latte✓✓✓37.441.046.046.152.654.333.239.341.6
", + "bbox": [ + 230, + 227, + 767, + 349 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/2e91f6bc715966bf0af65128b1ec32013bf4e3b4520d2bf8a2f5912c3115700c.jpg", + "image_caption": [ + "Fig. 3: Ablation studies of different frame aggregation mechanisms. Besides (a) Latte's slide window aggregation, we test two different aggregation methods, including (b) replacing single frame student predictions with multiple frames in Eq. (5) and (c) adding non-overlapping aggregation to (b). Results show that our slide window aggregation can better evaluate the local consistency and results in consistent improvement." + ], + "image_footnote": [], + "bbox": [ + 274, + 359, + 725, + 450 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "4.3 Ablation Studies and Visualization", + "text_level": 1, + "bbox": [ + 214, + 539, + 549, + 554 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "In this section, we investigate the effectiveness of each component in Latte by thorough ablation studies and qualitative results.", + "bbox": [ + 212, + 561, + 785, + 593 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Effectiveness of cross-modal learning. The main optimization objectives of Latte contain three parts, including ST-entropy-based cross-modal consistency (both 2D-to-3D and 3D-to-2D) and cross-entropy loss between predictions and cross-modal pseudo-labels. We justify the effectiveness by using different combinations of these components as in Tab. 2. As shown from experiments No. 1 to No. 6, disabling any part of Latte causes performance decreases of more than $1.6\\%$ relatively on the cross-modal prediction (xM), which proves the effectiveness of each component we included. A special case is No. 5 on U-to-S, where disabling ST $\\mathcal{L}_{23}^{\\mathrm{xM}}$ only leads to a minor $0.1\\%$ drop in accuracy. This is mainly because the 2D performance is already close to saturation (37.4 vs. 38.7 of Oracle TTA) without the guidance of 3D predictions. In this case, introducing 3D information to 2D is less effective compared to the other benchmarks.", + "bbox": [ + 212, + 613, + 787, + 792 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Besides optimization objects, we further testify two important components of Latte, including ST-entropy-based weighting in Eq. (10) and Eq. (13) as well as percentile filtering in Eq. (8). Specifically, disabling percentile filtering as in No.", + "bbox": [ + 212, + 795, + 787, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "H. Cao et al.", + "bbox": [ + 271, + 114, + 359, + 127 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/6b7836395acd46ec5d09642e0d8b6e842ddff0d8b006624ac84dfe69592869ed.jpg", + "image_caption": [ + "(a) Ablation studies of various ST voxel sizes." + ], + "image_footnote": [], + "bbox": [ + 259, + 143, + 413, + 222 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/2755e97e95cc69d4cd070c321af5a2940b87b922677f4bcd78ebc54a8359a136.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 423, + 143, + 576, + 222 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/7c9ec6155f3a12c5f3c433b3d33768baae76b4c4c4c957fe958691ec0172403d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 588, + 143, + 740, + 222 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/3e4fc3aefee68e3c50682ba0247c3960437614d62b6cb5aecdffd28f29d298db.jpg", + "image_caption": [ + "(b) Ablation studies of filtering percentile $\\alpha$ in Eq. (8)." + ], + "image_footnote": [], + "bbox": [ + 259, + 236, + 411, + 313 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/8216ef83bcc00af7ce9fc4349c5ebfc892298cd3257daca86fce94f832a9c0cf.jpg", + "image_caption": [ + "Fig. 4: Sensitivity analysis about the ST voxel size and confident percentile $\\alpha$ ." + ], + "image_footnote": [], + "bbox": [ + 424, + 236, + 573, + 313 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/b81f9446c37d96988e78a39195bf51e0b51e01109ba2865ae713e45acb63b0f3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 588, + 236, + 738, + 313 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "7 leads to a relative accuracy drop of more than $0.5\\%$ across all benchmarks while utilizing point-wise entropy instead of our ST entropy as in No. 8 degenerates the performance by more than $2.0\\%$ relatively, which justifies that ST entropy can better estimate the prediction reliability compared to the point-wise entropy.", + "bbox": [ + 212, + 359, + 787, + 421 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Different frame aggregation mechanisms. One of the core designs of Latte is the slide window frame aggregation, which is designed to capture temporally local correspondences and estimate their prediction consistency. To justify the aforementioned claim, we compare our slide window aggregation with two aggregation variants as shown in Fig. 3. Specifically, variant (b) regards all student prediction frames rather a single frame within the time window as the query in Eq. (5), while variant (c) further utilizes a non-overlapping sliding strategy. According to Fig. 3, our slide window aggregation surpasses the other two aggregation variants under all sizes of time windows across three benchmarks, which justifies the effectiveness of our method. The inferior performance of variants (b) and (c) could be due to two factors. Firstly, aggregating student predictions could be an inferior option since student predictions tend to be less stable, and aggregating multiple frames can therefore lead to noisy queries for cross-modal learning. Secondly, non-overlapping aggregation can result in a miss-checking area on the boundary of each time window, which leads to unrepresentative entropy estimation for some ST voxels. Another observation is that utilizing a smaller slide window can usually yield better performance since temporally local consistency can be captured. This justifies the effectiveness of estimating prediction reliability through temporally local consistency.", + "bbox": [ + 212, + 441, + 787, + 729 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Parameter sensitivity analysis. To testify whether Latte is sensitive towards hyper-parameter settings, we conduct sensitivity analysis on Latte across three benchmarks. Here we illustrate the sensitivity analysis of two hyper-parameters of Latte, including ST voxel size and filtering percentile $\\alpha$ as shown in Fig. 4. In terms of ST voxel sizes, one can tell from Fig. 4a that all benchmarks share the same optimal voxel size as 0.2, where performance drops as the voxel size", + "bbox": [ + 212, + 750, + 789, + 840 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "Reliable Spatial-Temporal Voxels For Multi-Modal Test-Time Adaptation", + "bbox": [ + 241, + 114, + 730, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/e7f6804bd9099bb2c160d639ef555a80d2aa6b99dcd01d54f1c9587948f9e2da.jpg", + "image_caption": [ + "Fig. 5: Qualitative results of Latte on S-to-S. Here we visualize modality-specific predictions and cross-modal predictions from Latte during two sets of consecutive frames." + ], + "image_footnote": [], + "bbox": [ + 243, + 143, + 756, + 305 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "becomes too large or too small. This is mainly because large voxels could cause boundary ambiguity while small voxels contain insufficient representative points for evaluation. As for filtering percentile $\\alpha$ , non-trivial improvement can be observed on U-to-S and A-to-S when $\\alpha \\geq 0.8$ , while a lower $\\alpha$ begins to discard more confident ST voxels leading to a slight drop in accuracy. On S-to-S, Latte without filtering (i.e., No. 7 in Tab. 2) performs slightly better on cross-modal predictions with a minor gap of $0.1\\%$ , while it can also be observed that setting $\\alpha$ as 0.9 leads to an improvement of $0.3\\%$ on both modality-specific predictions, which justifies the overall effectiveness of our filtering procedures.", + "bbox": [ + 212, + 349, + 787, + 486 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Qualitative results. Our main motivation is utilizing prediction consistency of spatial-temporal correspondences to estimate modality reliability and achieve better cross-modal attending. As shown in Fig. 5, Latte can effectively attend the modality with more consistent predictions in spatial-temporal correspondences and therefore improve the cross-modal prediction consistency across time. For instance, Latte successfully attends to the stable 2D predictions pedestrian in the upper set of consecutive frames, suppressing the spatial-temporal inconsistency from 3D predictions. Similar observations can also be found in the lower sample, where the more consistent 2D predictions of buildings in the red rectangle and 3D predictions of roads in the blue rectangle prevail. More qualitative comparisons with previous SOTA methods are presented in our appendix.", + "bbox": [ + 212, + 505, + 787, + 672 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 690, + 359, + 707 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In this paper, we propose a novel MM-TTA method called Latte, which leverages prediction consistency of spatial-temporal correspondences in MM-TTA for 3D segmentation. Latte utilizes a slide window frame aggregation to extract ST voxels and estimates temporally local consistency by the ST entropy of each modality. By attending to the modality with more consistency, Latte can achieve stable improvement on challenging MM-TTA scenarios. Despite its effectiveness, Latte struggles to rectify consistently wrong predictions across time, which is challenging and worth further investigation in the future.", + "bbox": [ + 212, + 719, + 787, + 839 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "H. Cao et al.", + "bbox": [ + 271, + 114, + 359, + 126 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgment", + "text_level": 1, + "bbox": [ + 217, + 145, + 382, + 162 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "This research is supported by the National Research Foundation, Singapore, under the NRF Medium Sized Centre for Advanced Robotics Technology Innovation (CARTIN). Any opinions, findings and conclusions or recommendations expressed in this material are those of the author(s) and do not reflect the views of National Research Foundation, Singapore.", + "bbox": [ + 212, + 175, + 787, + 251 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 273, + 321, + 287 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Behley, J., Garbade, M., Milioto, A., Quenzel, J., Behnke, S., Stachniss, C., Gall, J.: Semantickitti: A dataset for semantic scene understanding of lidar sequences. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 9297-9307 (2019)", + "2. Caesar, H., Bankiti, V., Lang, A.H., Vora, S., Liong, V.E., Xu, Q., Krishnan, A., Pan, Y., Baldan, G., Beijbom, O.: nuscenes: A multimodal dataset for autonomous driving. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11621-11631 (2020)", + "3. Cao, H., Xu, Y., Yang, J., Yin, P., Yuan, S., Xie, L.: Mopa: Multi-modal prior aided domain adaptation for 3d semantic segmentation. arXiv preprint arXiv:2309.11839 (2023)", + "4. Cao, H., Xu, Y., Yang, J., Yin, P., Yuan, S., Xie, L.: Multi-modal continual test-time adaptation for 3d semantic segmentation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 18809-18819 (October 2023)", + "5. Chen, R., Liu, Y., Kong, L., Zhu, X., Ma, Y., Li, Y., Hou, Y., Qiao, Y., Wang, W.: Clip2scene: Towards label-efficient 3d scene understanding by clip. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7020-7030 (2023)", + "6. Choy, C., Gwak, J., Savarese, S.: 4d spatio-temporal convnets: Minkowski convolutional neural networks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3075-3084 (2019)", + "7. Fan, H., Yang, Y., Kankanhalli, M.: Point 4d transformer networks for spatiotemporal modeling in point cloud videos. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14204-14213 (2021)", + "8. Fan, H., Yu, X., Ding, Y., Yang, Y., Kankanhalli, M.: Pistnet: Point spatio-temporal convolution on point cloud sequences. arXiv preprint arXiv:2205.13713 (2022)", + "9. Feng, D., Haase-Schütz, C., Rosenbaum, L., Hertlein, H., Glaeser, C., Timm, F., Wiesbeck, W., Dietmayer, K.: Deep multi-modal object detection and semantic segmentation for autonomous driving: Datasets, methods, and challenges. IEEE Transactions on Intelligent Transportation Systems 22(3), 1341-1360 (2020)", + "10. Geyer, J., Kassahun, Y., Mahmudi, M., Ricou, X., Durgesh, R., Chung, A.S., Hauswald, L., Pham, V.H., Mühlegg, M., Dorn, S., et al.: A2d2: Audi autonomous driving dataset. arXiv preprint arXiv:2004.06320 (2020)", + "11. Gong, T., Jeong, J., Kim, T., Kim, Y., Shin, J., Lee, S.J.: Robust continual test-time adaptation: Instance-aware bn and prediction-balanced memory. arXiv preprint arXiv:2208.05117 (2022)", + "12. Goyal, S., Sun, M., Raghunathan, A., Kolter, J.Z.: Test time adaptation via conjugate pseudo-labels. In: Advances in Neural Information Processing Systems (2022)" + ], + "bbox": [ + 218, + 301, + 784, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "Reliable Spatial-Temporal Voxels For Multi-Modal Test-Time Adaptation", + "bbox": [ + 241, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "13. Graham, B.: Sparse 3d convolutional neural networks. arXiv preprint arXiv:1505.02890 (2015)", + "14. Guo, Y., Wang, H., Hu, Q., Liu, H., Liu, L., Bennamoun, M.: Deep learning for 3d point clouds: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence 43(12), 4338-4364 (2020)", + "15. Huang, S., Gojcic, Z., Huang, J., Wieser, A., Schindler, K.: Dynamic 3d scene analysis by point cloud accumulation. In: European Conference on Computer Vision. pp. 674-690. Springer (2022)", + "16. Jaritz, M., Vu, T.H., De Charette, R., Wirbel, E., Pérez, P.: Cross-modal learning for domain adaptation in 3d semantic segmentation. IEEE Transactions on Pattern Analysis and Machine Intelligence 45(2), 1533-1544 (2022)", + "17. Ji, X., Yuan, S., Yin, P., Xie, L.: Lio-gvm: an accurate, tightly-coupled lidar-inertial odometry with gaussian voxel map. IEEE Robotics and Automation Letters (2024)", + "18. Li, M., Zhang, Y., Xie, Y., Gao, Z., Li, C., Zhang, Z., Qu, Y.: Cross-domain and cross-modal knowledge distillation in domain adaptation for 3d semantic segmentation. In: Proceedings of the 30th ACM International Conference on Multimedia. pp. 3829-3837 (2022)", + "19. Liang, J., Hu, D., Feng, J.: Do we really need to access the source data? source hypothesis transfer for unsupervised domain adaptation. In: International Conference on Machine Learning. pp. 6028-6039. PMLR (2020)", + "20. Liu, W., Luo, Z., Cai, Y., Yu, Y., Ke, Y., Junior, J.M., Gonçalves, W.N., Li, J.: Adversarial unsupervised domain adaptation for 3d semantic segmentation with multi-modal learning. ISPRS Journal of Photogrammetry and Remote Sensing 176, 211-221 (2021)", + "21. Liu, Y., Kothari, P., Van Delft, B., Bellot-Gurlet, B., Mordan, T., Alahi, A.: Ttt++: When does self-supervised test-time training fail or thrive? Advances in Neural Information Processing Systems 34, 21808-21820 (2021)", + "22. Niu, S., Wu, J., Zhang, Y., Chen, Y., Zheng, S., Zhao, P., Tan, M.: Efficient test-time model adaptation without forgetting. In: International Conference on Machine Learning. pp. 16888-16905. PMLR (2022)", + "23. Niu, S., Wu, J., Zhang, Y., Wen, Z., Chen, Y., Zhao, P., Tan, M.: Towards stable test-time adaptation in dynamic wild world. arXiv preprint arXiv:2302.12400 (2023)", + "24. Peng, D., Lei, Y., Li, W., Zhang, P., Guo, Y.: Sparse-to-dense feature matching: Intra and inter domain cross-modal learning in domain adaptation for 3d semantic segmentation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 7108-7117 (2021)", + "25. Piergiovanni, A., Casser, V., Ryoo, M.S., Angelova, A.: 4d-net for learned multimodal alignment. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 15435-15445 (2021)", + "26. Qi, C.R., Su, H., Mo, K., Guibas, L.J.: Pointnet: Deep learning on point sets for 3d classification and segmentation. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 652-660 (2017)", + "27. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International Conference on Machine Learning. pp. 8748-8763. PMLR (2021)", + "28. Saltori, C., Krivosheev, E., Lathuilière, S., Sebe, N., Galasso, F., Fiameni, G., Ricci, E., Poiesi, F.: Gipso: Geometrically informed propagation for online adaptation in 3d lidar segmentation. In: European Conference on Computer Vision. pp. 567-585. Springer (2022)" + ], + "bbox": [ + 217, + 147, + 784, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "H. Cao et al.", + "bbox": [ + 271, + 114, + 359, + 126 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "29. Shin, I., Tsai, Y.H., Zhuang, B., Schulter, S., Liu, B., Garg, S., Kweon, I.S., Yoon, K.J.: Mm-tta: multi-modal test-time adaptation for 3d semantic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16928-16937 (2022)", + "30. Simons, C., Raychaudhuri, D.S., Ahmed, S.M., You, S., Karydis, K., Roy-Chowdhury, A.K.: Summit: Source-free adaptation of uni-modal models to multimodal targets. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 1239-1249 (October 2023)", + "31. Song, J., Lee, J., Kweon, I.S., Choi, S.: Ecotta: Memory-efficient continual test-time adaptation via self-distilled regularization. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11920-11929 (2023)", + "32. Su, Y., Xu, X., Li, T., Jia, K.: Revisiting realistic test-time training: Sequential inference and adaptation by anchored clustering regularized self-training. arXiv preprint arXiv:2303.10856 (2023)", + "33. Tang, H., Liu, Z., Zhao, S., Lin, Y., Lin, J., Wang, H., Han, S.: Searching efficient 3d architectures with sparse point-voxel convolution. In: European Conference on Computer Vision. pp. 685–702. Springer (2020)", + "34. Tarvainen, A., Valpola, H.: Mean teachers are better role models: Weight-averaged consistency targets improve semi-supervised deep learning results. Advances in Neural Information Processing Systems 30 (2017)", + "35. Vedula, S., Baker, S., Rander, P., Collins, R., Kanade, T.: Three-dimensional scene flow. In: Proceedings of the Seventh IEEE International Conference on Computer Vision. vol. 2, pp. 722-729. IEEE (1999)", + "36. Vizzo, I., Guadagnino, T., Mersch, B., Wiesmann, L., Behley, J., Stachniss, C.: Kiss-icp: In defense of point-to-point icp-simple, accurate, and robust registration if done the right way. IEEE Robotics and Automation Letters 8(2), 1029-1036 (2023)", + "37. Vogel, C., Schindler, K., Roth, S.: 3d scene flow estimation with a rigid motion prior. In: 2011 International Conference on Computer Vision. pp. 1291-1298. IEEE (2011)", + "38. Vogel, C., Schindler, K., Roth, S.: Piecewise rigid scene flow. In: Proceedings of the IEEE International Conference on Computer Vision. pp. 1377-1384 (2013)", + "39. Wang, D., Shelhamer, E., Liu, S., Olshausen, B., Darrell, T.: Tent: Fully test-time adaptation by entropy minimization. In: International Conference on Learning Representations (2021), https://openreview.net/forum?id=uX13bZLkr3c", + "40. Wang, J.K., Wibisono, A.: Towards understanding gd with hard and conjugate pseudo-labels for test-time adaptation. arXiv preprint arXiv:2210.10019 (2022)", + "41. Wang, Q., Fink, O., Van Gool, L., Dai, D.: Continual test-time domain adaptation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7201-7211 (2022)", + "42. Wyner, A.: Recent results in the shannon theory. IEEE Transactions on information Theory 20(1), 2-10 (1974)", + "43. Xie, E., Wang, W., Yu, Z., Anandkumar, A., Alvarez, J.M., Luo, P.: Segformer: Simple and efficient design for semantic segmentation with transformers. Advances in Neural Information Processing Systems 34, 12077-12090 (2021)", + "44. Xing, B., Ying, X., Wang, R., Yang, J., Chen, T.: Cross-modal contrastive learning for domain adaptation in 3d semantic segmentation. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 37, pp. 2974-2982 (2023)", + "45. Xu, J., Miao, Z., Zhang, D., Pan, H., Liu, K., Hao, P., Zhu, J., Sun, Z., Li, H., Zhan, X.: Int: Towards infinite-frames 3d detection with an efficient framework. In: European Conference on Computer Vision. pp. 193–209. Springer (2022)" + ], + "bbox": [ + 215, + 146, + 784, + 840 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "Reliable Spatial-Temporal Voxels For Multi-Modal Test-Time Adaptation", + "bbox": [ + 241, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "46. Yin, P., Cao, H., Nguyen, T.M., Yuan, S., Zhang, S., Liu, K., Xie, L.: Outram: One-shot global localization via triangulated scene graph and global outlier pruning. arXiv preprint arXiv:2309.08914 (2023)", + "47. Yuan, L., Xie, B., Li, S.: Robust test-time adaptation in dynamic scenarios. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15922-15932 (2023)", + "48. Zhang, M., Levine, S., Finn, C.: Memo: Test time robustness via adaptation and augmentation. Advances in Neural Information Processing Systems 35, 38629-38642 (2022)" + ], + "bbox": [ + 215, + 146, + 784, + 272 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "H. Cao et al.", + "bbox": [ + 271, + 114, + 359, + 126 + ], + "page_idx": 17 + } +] \ No newline at end of file diff --git a/2024/Reliable Spatial-Temporal Voxels For Multi-Modal Test-Time Adaptation/c906fc92-ebc1-45f1-bc7d-fa0568ba9181_model.json b/2024/Reliable Spatial-Temporal Voxels For Multi-Modal Test-Time Adaptation/c906fc92-ebc1-45f1-bc7d-fa0568ba9181_model.json new file mode 100644 index 0000000000000000000000000000000000000000..274d64af809d4f86afbfd220117998f2448ed792 --- /dev/null +++ b/2024/Reliable Spatial-Temporal Voxels For Multi-Modal Test-Time Adaptation/c906fc92-ebc1-45f1-bc7d-fa0568ba9181_model.json @@ -0,0 +1,2194 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.29, + 0.141, + 0.713, + 0.187 + ], + "angle": 0, + "content": "Reliable Spatial-Temporal Voxels For Multi-Modal Test-Time Adaptation" + }, + { + "type": "text", + "bbox": [ + 0.224, + 0.212, + 0.78, + 0.244 + ], + "angle": 0, + "content": "Haozhi Cao\\(^{1}\\), Yuecong Xu\\(^{1,2}\\), Jianfei Yang\\(^{1,3}\\), Pengyu Yin\\(^{1}\\), Xingyu Ji\\(^{1}\\), Shenghai Yuan\\(^{1}\\), and Lihua Xie\\(^{1}\\)" + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.255, + 0.782, + 0.283 + ], + "angle": 0, + "content": "1 School of Electrical and Electronic Engineering, Nanyang Technological University {haozhi002, jianfei.yang, pengyu001, xinyu001, shyuan, elhxie}@ntu.edu.sg" + }, + { + "type": "text", + "bbox": [ + 0.316, + 0.283, + 0.686, + 0.325 + ], + "angle": 0, + "content": "\\(^{2}\\) Department of Electrical and Computer Engineering, National University of Singapore yc.xu@nus.edu.sg" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.325, + 0.785, + 0.34 + ], + "angle": 0, + "content": "3 School of Mechanical and Aerospace Engineering, Nanyang Technological University" + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.375, + 0.74, + 0.653 + ], + "angle": 0, + "content": "Abstract. Multi-modal test-time adaptation (MM-TTA) is proposed to adapt models to an unlabeled target domain by leveraging the complementary multi-modal inputs in an online manner. Previous MM-TTA methods for 3D segmentation rely on predictions of cross-modal information in each input frame, while they ignore the fact that predictions of geometric neighborhoods within consecutive frames are highly correlated, leading to unstable predictions across time. To fulfill this gap, we propose ReLiable Spatial-temporal Voxels (Latte), an MM-TTA method that leverages reliable cross-modal spatial-temporal correspondences for multi-modal 3D segmentation. Motivated by the fact that reliable predictions should be consistent with their spatial-temporal correspondences, Latte aggregates consecutive frames in a slide window manner and constructs Spatial-Temopral (ST) voxels to capture temporally local prediction consistency for each modality. After filtering out ST voxels with high ST entropy, Latte conducts cross-modal learning for each point and pixel by attending to those with reliable and consistent predictions among both spatial and temporal neighborhoods. Experimental results show that Latte achieves state-of-the-art performance on three different MM-TTA benchmarks compared to previous MM-TTA or TTA methods. Visit our project site https://sites.google.com/view/eccv24-latte." + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.665, + 0.741, + 0.693 + ], + "angle": 0, + "content": "Keywords: Test-time adaptation \\(\\cdot\\) Multi-modal learning \\(\\cdot\\) 3D semantic segmentation" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.719, + 0.377, + 0.735 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.841 + ], + "angle": 0, + "content": "3D semantic segmentation is a fundamental task to achieve various fully autonomous applications such as autonomous driving and robot navigation [9,14]. With the increasing urge for robust sensing, multi-modal sensors (e.g., cameras and LiDARs) are widely adopted in autonomous systems containing rich complementary information from different sensors. However, conventional deep-learning-based solutions lean on expensive point-wise annotations and perform" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.127 + ], + "angle": 0, + "content": "H. Cao et al." + }, + { + "type": "image", + "bbox": [ + 0.277, + 0.145, + 0.455, + 0.249 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.465, + 0.143, + 0.727, + 0.251 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.261, + 0.788, + 0.372 + ], + "angle": 0, + "content": "Fig. 1: Illustration of previous MM-TTA methods and our Latte. (a) visualizes the predictions from the previous state-of-the-art MM-TTA method [29] in consecutive frames on SemanticKITTI [1], where regions in white boxes exhibit noisy predictions when considering single-frame input. (b) presents the overall framework of Latte, which firstly obtains ST voxels for each modality given the merged input of consecutive frames within each slide window and then computes the modality-specific ST entropy \\( E^{2\\mathrm{D}} \\), \\( E^{3\\mathrm{D}} \\). ST entropy indicates the relative reliability of each voxel in each modality and therefore is utilized in the following adaptive cross-modal attending." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.402, + 0.785, + 0.477 + ], + "angle": 0, + "content": "poorly on data from different domains due to the domain shift [6, 26, 33]. To resolve this problem, previous works have proposed Multi-Modal Unsupervised Domain Adaptation (MM-UDA) approaches [16, 18, 24, 44] that employ multimodal information to transfer knowledge learned from the labeled source domain to the unlabeled target domain without expensive human annotations." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.478, + 0.788, + 0.644 + ], + "angle": 0, + "content": "Nevertheless, the adaptation process of MM-UDA is completely offline, which requires full access to the source domain dataset and multiple training epochs. Despite its effectiveness, offline training is usually infeasible when encountering distribution shifts during the inference stage that require adapting models in real time. Inspired by Test-Time Adaptation (TTA) [22, 23], a recent work [29] proposes the first Multi-Modal Test-Time Adaptation (MM-TTA) method for 3D segmentation. Similar to the settings of TTA, MM-TTA prohibits access to any raw sample from the source domain dataset and adapts the model during the testing stage. This results in a quick adaptation scenario (i.e., one epoch for training only as in [29, 32]) that requires stable optimization in an online manner. To this end, MM-TTA methods must seek a reliable source of supervision signals." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.644, + 0.789, + 0.841 + ], + "angle": 0, + "content": "To fulfill the efficiency and stability of adaptation, the current state-of-the-art MM-TTA method [29] evaluates modal reliability by the prediction consistency between different experts in a frame-wise manner. However, due to the domain shift between the online frame and the pre-trained source dataset, the single-frame prediction is usually unstable. Existing approaches rely on pseudo labels to refine the single-frame prediction but still fail in some cases. As shown in Fig. 1a, the single-frame refinement strategy [29] suffers from noisy single-frame predictions on the car in the white rectangle across time. Such inconsistent predictions across time degenerate the segmentation performance and aggravate denoising stress in downstream tasks (e.g., semantic-based retrivial [46] and obstacle recognition). On one hand, these temporally unstable single-frame predictions could be wrongly regarded as reliable ones, propagating prediction noise to the other modality and causing error accumulation or even catastrophic forgetting [22,23]." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.242, + 0.115, + 0.732, + 0.131 + ], + "angle": 0, + "content": "Reliable Spatial-Temporal Voxels For Multi-Modal Test-Time Adaptation" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.785, + 0.207 + ], + "angle": 0, + "content": "On the other hand, while previous methods [4, 41] alleviate this instability by utilizing the average predictions of multiple augmented frames instead, they are computationally expensive for online adaptation since the inference time grows linearly with the increasing number of augmented frames." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.208, + 0.787, + 0.388 + ], + "angle": 0, + "content": "This work aims to efficiently suppress the single-frame instability for MM-TTA by exploiting the correlation between multi-frame correspondences. Specifically, we propose that 3D space can be efficiently divided into multiple voxels [33], each of which contains points and point-corresponding pixels located at the same geometric region. While captured at different timestamps, it can be observed that points or pixels located at the same voxel can be viewed as different observations on the same semantic objects [28]. Reliable predictions should therefore be certain and consistent within their corresponding voxels across time, whether we evaluate them from a temporally global perspective (i.e., all input frames) or a temporally local perspective (i.e., certain consecutive frames). By aggregating the predictions in a voxel-wise manner and re-evaluating their reliability, the single-frame prediction noise as in Fig. 1a can be effectively alleviated." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.389, + 0.787, + 0.616 + ], + "angle": 0, + "content": "Motivated by the above observation, we propose a novel MM-TTA method called ReLiable Spatial-temporal Voxels (Latte), which leverages the spatial-temporal correspondences within consecutive frames in a cross-modal manner as in Fig. 1b. Specifically, given frames of point clouds with their estimated poses, we regard points in the same voxel extracted from the merged point cloud frames as spatial-temporal correspondences. Different from previous works that merge all input frames [5,7] or regularize frame-to-frame consistency [28], Latte aggregates consecutive frames in a sliding window manner to estimate the temporally local prediction consistency through the proposed Spatial-Temporal voxels (ST voxels) and entropy (ST entropy). Based on the reliability estimated by ST entropy, we conduct cross-modal learning in an adaptive attending manner to reduce the contributions of predictions from the noisy modality. To better facilitate the progress of MM-TTA, we evaluate our Latte with more recent backbones (SegFormer [43] and SPVCNN [33]) on three different benchmarks, where Latte consistently outperforms previous TTA or MM-TTA methods." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.617, + 0.787, + 0.735 + ], + "angle": 0, + "content": "In summary, our contributions are three-fold. Firstly, we propose Latte, a novel MM-TTA method for multi-modal 3D segmentation, which is the first work that incorporates spatial-temporal correlations for MM-TTA segmentation to our best knowledge. Secondly, Latte efficiently extracts spatial-temporal correspondence through ST voxels and estimates their reliability by ST entropy, which is further incorporated to enhance adaptive cross-modal attending. Thirdly, our experimental results show that Latte outperforms previous state-of-the-art TTA and MM-TTA methods on three different benchmarks." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.761, + 0.396, + 0.779 + ], + "angle": 0, + "content": "2 Related Works" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.787, + 0.842 + ], + "angle": 0, + "content": "Test-Time Adaptation (TTA). TTA is proposed to mitigate the domain shift between the training data and the testing data in an online manner, which is firstly proposed by TENT [39]. Due to its practical and challenging setting," + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.128 + ], + "angle": 0, + "content": "H. Cao et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.344 + ], + "angle": 0, + "content": "TTA is attracting more and more attention. Previously proposed TTA methods attempt to address this task from different perspectives, such as entropy minimization [22, 23, 39], self-training with pseudo-labels [12, 40], and augmentation invariance [41, 48]. In addition to regularizing predictions, some methods propose to adapt networks from the feature level instead [21, 32], while more recent methods begin to consider different variants of TTA scenarios, such as continual TTA [4, 22, 41], non-i.i.d TTA [11], or the mix of aforementioned cases [47]. Different from source-free adaptation [19, 30] which also discards the access to the source domain yet requires multiple training epochs, the adaptation process of TTA is completely online, following the \"one-pass\" protocol as in [32]. While most existing TTA methods are designed for image classification or image segmentation without explicit temporal correlation, we argue temporal information in 3D segmentation can be effectively leveraged for TTA as Latte in this work." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.35, + 0.788, + 0.623 + ], + "angle": 0, + "content": "Multi-modal domain adaptation for 3D segmentation. To avoid expensive annotation costs and overcome the poor generalizability of fully supervised solutions, various multi-modal domain adaptation methods have been investigated. Specifically for MM-UDA, xMUDA [16] is the primary works that explore incorporating cross-modal learning with MM-UDA. It regards cross-modal prediction consistency and pseudo-labels as its supervision signals in the unlabeled target domain. Most subsequent works propose solutions to address its limitations from different perspectives, such as more diverse point-pixel correspondence [24, 44], procedures to mitigate domain gaps [18, 20], and alleviating class-imbalanced problem [3]. Since MM-UDA is infeasible when facing online domain shifts, some previous works [4, 29] develop different methods to address the domain gap in an online manner, including cross-modal pseudo-label generation [29] and adaptive modality attending [4]. A recent work [30] proposes a source-free domain adaptation method by estimating cross-modal prediction agreement, yet it requires multiple offline training epochs. Different from previous multi-modal domain adaptation methods that ignore temporal relationships, Latte couples spatial-temporal correspondence to mitigate the single-frame instability and achieve better online adaptation performance." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.629, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Temporal processing of point clouds. Since point clouds are naturally in the form of temporally consecutive frames, previous methods have widely explored how to leverage such temporal information for different tasks. Specifically, scene flow [15,35,37,38] contains the movement of every point in the 3D world. Considering such point-wise computation expensive, previous works on different tasks (e.g., segmentation [5,6,28] and object detection [25]) usually interpret the temporal information from a simpler perspective. In terms of fully supervised learning, Choy et al. [6] propose the first 4D segmentation method by generalized sparse convolution kernels [13] and Fan et al. [8] turn to the fully connected learning scheme as in [26], while Xu et al. [45] extract long-term temporal information by an efficient memory bank. On the other hand, considering scenarios with insufficient or zero annotations, some works propose different strategies to couple spatial-temporal information without ground-truth labels. For instance, Chen et al. [5] proposes to guide voxelized spatial-temporal representation with" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.242, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Reliable Spatial-Temporal Voxels For Multi-Modal Test-Time Adaptation" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "image", + "bbox": [ + 0.232, + 0.145, + 0.773, + 0.282 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.296, + 0.788, + 0.394 + ], + "angle": 0, + "content": "Fig. 2: Overall structure of Latte. Taking a student prediction frame of one modality as the query input, our slide-window extraction searches its spatial-temporal correspondences through voxelization within a time window to establish the temporally local prediction consistency. ST voxels are then generated, where those with high ST entropy (larger than \\(\\alpha\\)-quantile) are discarded as unreliable correspondences, while the others are leveraged for adaptive cross-modal learning by attending to the modality with lower ST entropy in a voxel-wise manner." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.423, + 0.785, + 0.5 + ], + "angle": 0, + "content": "strong text embedding from CLIP [27] instead of labels in a multi-modal manner, while Saitor et al. [28] interpret such representation by nearest neighbor search in a frame-to-frame manner. In this work, we design a slide window aggregation and voxelization rather than aggregating all frames as in [5], which yields a better indication of the prediction reliability for each point." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.522, + 0.38, + 0.54 + ], + "angle": 0, + "content": "3 Methodology" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.553, + 0.788, + 0.765 + ], + "angle": 0, + "content": "Problem definition. During MM-TTA for 3D segmentation, both 2D RGB images and 3D point clouds are captured consecutively in the target domain \\(\\mathcal{T}\\), denoted as \\(\\mathbf{x}_{\\mathcal{T},t}^{2\\mathrm{D}} \\in \\mathbb{R}^{3 \\times H \\times W}\\) and \\(\\mathbf{x}_{\\mathcal{T},t}^{3\\mathrm{D}} \\in \\mathbb{R}^{N \\times 4}\\), respectively, where \\(t\\) represents the frame order and \\(H, W\\) are the height and width of input images. The point-wise input contains the 3D coordinate \\(\\{x, y, z\\}\\) and the point feature (e.g., intensity, reflectivity, etc.). Modality-specific networks are pre-trained on the labeled source domain before adaptation, denoted as \\(\\phi_{\\mathcal{T},t}^{\\mathrm{m}}(\\cdot)\\) with \\(\\mathrm{m} \\in \\{2\\mathrm{D}, 3\\mathrm{D}\\}\\). During MM-TTA, both networks are initialized from the parameters pre-trained on the source domain. Following previous works [16,29], 3D points are projected to the 2D image based on the relative projection matrix to obtain cross-modal correspondences within each frame, resulting in the modality-specific predictions \\(\\mathbf{p}_{\\mathcal{T},t}^{\\mathrm{m}} = \\phi_{\\mathcal{T},t}^{\\mathrm{m}}(\\mathbf{x}_{\\mathcal{T},t}^{\\mathrm{m}})\\), \\(\\mathbf{p}_{\\mathcal{T},t}^{\\mathrm{m}} \\in \\mathbb{R}^{N \\times K}\\) where \\(K\\) denotes the number of semantic classes. Data from the labeled source domain are strictly inaccessible and we omit the target domain subscript \\(\\mathcal{T}\\) by default." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.766, + 0.788, + 0.842 + ], + "angle": 0, + "content": "The multi-modal input of online 3D segmentation is temporally consecutive, which exhibits strong temporal relationships between frames. The geometric neighborhoods in consecutive frames can be viewed as the naturally various observations on the same semantic object that can be leveraged to stabilize noisy single-frame predictions. Motivated by this observation, we develop Latte," + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.128 + ], + "angle": 0, + "content": "H. Cao et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.284 + ], + "angle": 0, + "content": "which discovers reliable spatial-temporal (ST) correspondences to improve both optimization stability and prediction consistency during MM-TTA. As shown in Fig. 2, given consecutive frames of 2D images and 3D point clouds as input, Latte firstly extracts their frame-wise predictions from both student and teacher networks (Sec. 3.1). These frame-wise predictions are then passed to the slide-window extraction and voxelization (Sec. 3.2). ST voxels and ST entropy are subsequently computed indicating the prediction consistency and certainty (Sec. 3.3). The final ST entropy is leveraged for cross-modal attending to alleviate the noise from unreliable modality-specific predictions (Sec. 3.4)." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.302, + 0.691, + 0.316 + ], + "angle": 0, + "content": "3.1 Frame-wise Predictions from Students and Teachers" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.323, + 0.784, + 0.414 + ], + "angle": 0, + "content": "Before exploiting spatial-temporal correspondence, we first generate the framewise predictions from each modality. Motivated by the fact that moving average models can provide more stable online predictions [4,34,41], Latte utilizes weight-averaged teacher models and fast student models for each modality as in previous works [29, 41], denoted as \\(\\tilde{\\phi}_t^{\\mathrm{m}}(\\cdot)\\) and \\(\\phi_t^{\\mathrm{m}}(\\cdot)\\), respectively. Given each input frame \\(\\mathbf{x}_t^{\\mathrm{m}}\\), the student predictions \\(\\mathbf{p}_t^{\\mathrm{m}}\\) and teacher predictions \\(\\tilde{\\mathbf{p}}_t^{\\mathrm{m}}\\) are computed as:" + }, + { + "type": "equation", + "bbox": [ + 0.397, + 0.421, + 0.785, + 0.438 + ], + "angle": 0, + "content": "\\[\n\\tilde {\\mathbf {p}} _ {t} ^ {\\mathrm {m}} = \\tilde {\\phi} _ {t} ^ {\\mathrm {m}} \\left(\\mathbf {x} _ {t} ^ {\\mathrm {m}}\\right), \\mathbf {p} _ {t} ^ {\\mathrm {m}} = \\phi_ {t} ^ {\\mathrm {m}} \\left(\\mathbf {x} _ {t} ^ {\\mathrm {m}}\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.444, + 0.788, + 0.52 + ], + "angle": 0, + "content": "where the teacher model \\(\\tilde{\\phi}_t^{\\mathrm{m}}\\) is initialized from the source domain pre-trained models. In the following process, the teacher's predictions of one modality are regarded as the cross-modal guidance of student predictions from the other modality. In practice, the gradient propagation of teacher model predictions is discarded to prevent directly updating teacher models." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.538, + 0.684, + 0.554 + ], + "angle": 0, + "content": "3.2 Slide Window Frame Aggregation and Voxelization" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.559, + 0.788, + 0.695 + ], + "angle": 0, + "content": "Our goal is to efficiently leverage spatial-temporal correspondences between predictions and seek reliability estimation through prediction consistency within correspondences. Establishing spatial-temporal correspondences in 3D space has been explored in other deep learning tasks, where they mainly consider temporally global relationships of all input frames [5,28] or frame-to-frame correspondences [28]. However, they both have some limitations in terms of consistency and certainty evaluation: the former can not highlight the risk of some inconsistent predictions within a short time window (i.e., temporally local inconsistency), while the latter suffers from an insufficient number of correspondences." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.695, + 0.788, + 0.818 + ], + "angle": 0, + "content": "To effectively empower spatial-temporal information with Latte, we propose a slide window frame aggregation that focuses on temporally local correspondences that lie within a time window to evaluate temporally local consistency. Specifically, given a student prediction frame \\(\\mathbf{p}_i^{\\mathrm{m}}\\) with the temporal index \\(i\\) as the query, the correspondence search is conducted within a time window of consecutive frames denoted as \\(\\{j||j - i|\\leq w_t\\}\\), where \\(w_{t}\\) is the pre-defined time window size. The merged point cloud \\(\\hat{\\mathbf{x}}_i^{3\\mathrm{D}}\\in \\mathbb{R}^{\\hat{N}\\times 3}\\) (point features are omitted for simplicity) for correspondence search is then formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.303, + 0.824, + 0.785, + 0.843 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathbf {x}} _ {i} ^ {\\mathrm {3 D}} = \\operatorname {c a t} \\left( \\right.\\left\\{ \\right.\\mathbf {T} _ {j \\rightarrow i} * \\mathbf {x} _ {j} ^ {\\mathrm {3 D}} \\left. \\right| | j - i | \\leq w _ {t} \\}\\left. \\right), \\quad \\mathbf {T} _ {j \\rightarrow i} = \\mathbf {T} _ {i} ^ {- 1} \\mathbf {T} _ {j}, \\tag {2}\n\\]" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.242, + 0.115, + 0.733, + 0.131 + ], + "angle": 0, + "content": "Reliable Spatial-Temporal Voxels For Multi-Modal Test-Time Adaptation" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.209 + ], + "angle": 0, + "content": "where \\(\\mathbf{T}_i\\) is the estimated pose at frame \\(i\\) and so as \\(\\mathbf{T}_j\\), which can be easily obtained online through off-the-shelf SLAM algorithms [17,36]. \\(\\operatorname{cat}(\\cdot)\\) is the concatenation operation along the point number dimension while \\(*\\) denotes the pose transformation. The voxelization is then performed on \\(\\hat{\\mathbf{x}}_i^{3D}\\), formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.32, + 0.217, + 0.785, + 0.235 + ], + "angle": 0, + "content": "\\[\n\\mathbf {v} _ {i} ^ {\\mathrm {3 D}} = \\mathcal {V} _ {\\mathbf {s}} \\left(\\hat {\\mathbf {x}} _ {i} ^ {\\mathrm {3 D}}\\right), \\quad \\mathbf {v} _ {i} ^ {\\mathrm {3 D}} \\in \\mathbb {R} ^ {N _ {\\mathrm {v}} \\times 3}, \\tag {3}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.312, + 0.239, + 0.785, + 0.26 + ], + "angle": 0, + "content": "\\[\n\\mathbf {M} _ {i} ^ {\\mathrm {3 D}} = \\left\\{k | \\forall g \\in [ 1, \\hat {N} ], \\left\\lfloor \\hat {\\mathbf {x}} _ {i, g} ^ {\\mathrm {3 D}} / \\mathbf {s} \\right\\rfloor = \\mathbf {v} _ {i, k} ^ {\\mathrm {3 D}} \\right\\}, \\quad \\mathbf {M} _ {i} ^ {\\mathrm {3 D}} \\in \\mathbb {R} ^ {\\hat {N}}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.269, + 0.789, + 0.406 + ], + "angle": 0, + "content": "where \\(\\mathbf{v}_i^{3\\mathrm{D}}\\) is the extracted voxels and \\(\\nu_{s}\\) is the voxelization operation with a voxel size s. After voxelization, points located in the same voxel can be regarded as correspondences since they are geometric neighborhoods, where the hash table \\(\\mathbf{M}_i^{3\\mathrm{D}}\\) that maps each voxel back to its containing points (e.g., \\(\\mathbf{v}_{i,k}^{3\\mathrm{D}}\\) to \\(\\hat{\\mathbf{x}}_{i,g}^{3\\mathrm{D}}\\) as in Eq. (4)) is preserved for the subsequent correspondence mapping. In this way, it provides more diverse neighborhood selection strategies with temporally local correspondences for consistency and certainty evaluation compared to globally merging all frames in one single step. In practice, the slide window aggregation and voxelization are conducted iteratively for each frame in the input batch." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.427, + 0.573, + 0.442 + ], + "angle": 0, + "content": "3.3 Spatial-Temporal Voxels and Entropy" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.45, + 0.789, + 0.616 + ], + "angle": 0, + "content": "Given the geometric correspondences captured by voxelization, our goal is to emphasize the predictions with correspondence consistency within each voxel. After slide window aggregation and voxelization, there inevitably exist some unreliable correspondences of two different types: (i) voxels that contain different types of semantic objects (e.g., those located on the contact surface between cars and roads) and (ii) voxels with highly inconsistent predictions due to uncertain predictions across time. These unreliable correspondences contain high uncertainty and inconsistency, which could harm the online optimization of models if they as references. To alleviate the side-effect introduced by such unreliable correspondences, we propose ST voxels that leverage ST entropy for voxel-wise reliability evaluation and cross-modal attending." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.617, + 0.789, + 0.753 + ], + "angle": 0, + "content": "Specifically, an ST voxel contains two components, including the query and the reference, where the query is encouraged to be consistent with the reference. Considering that the moving average teacher can usually generate more stable predictions [4, 41], we regard the single-frame student predictions as the query while multi-frame teacher predictions as our reference, forming ST voxels denoted as \\(\\mathbf{v}_i^{\\mathrm{ST}}\\). Without losing generality, here we take a single ST voxel indexed by \\(k\\) in frame \\(i\\) as an example, denoted as \\(\\mathbf{v}_{i,k}^{\\mathrm{ST}} = \\{\\mathbf{p}_{\\mathrm{q}}^{\\mathrm{m}},\\mathbf{p}_{\\mathrm{r}}^{\\mathrm{m}}\\}\\), where \\(\\mathbf{p}_{\\mathrm{q}}^{\\mathrm{m}}\\) and \\(\\mathbf{p}_{\\mathrm{r}}^{\\mathrm{m}}\\) are the point-wise student predictions at the query frame \\(i\\) and teacher predictions in the frame searching range \\(j\\), respectively, formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.304, + 0.762, + 0.785, + 0.781 + ], + "angle": 0, + "content": "\\[\n\\mathbf {p} _ {\\mathrm {q}} ^ {\\mathrm {m}} = \\left\\{\\mathbf {p} _ {i, g} ^ {\\mathrm {m}} \\mid \\left\\lfloor \\mathbf {x} _ {i, g} ^ {\\mathrm {3 D}} / \\mathbf {s} \\right\\rfloor = \\mathbf {v} _ {i, k} ^ {\\mathrm {3 D}} \\right\\}, \\quad \\mathbf {p} _ {\\mathrm {q}} ^ {\\mathrm {m}} \\in \\mathbb {R} ^ {N _ {\\mathrm {q}} \\times K} \\tag {5}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.304, + 0.783, + 0.785, + 0.802 + ], + "angle": 0, + "content": "\\[\n\\mathbf {p} _ {\\mathrm {r}} ^ {\\mathrm {m}} = \\left\\{\\tilde {\\mathbf {p}} _ {j, g} ^ {\\mathrm {m}} \\mid \\left| \\mathbf {x} _ {j, g} ^ {3 \\mathrm {D}} / \\mathbf {s} \\right| = \\mathbf {v} _ {i, k} ^ {3 \\mathrm {D}}, | j - i | \\leq w _ {t} \\right\\}, \\quad \\mathbf {p} _ {\\mathrm {r}} ^ {\\mathrm {m}} \\in \\mathbb {R} ^ {N _ {\\mathrm {r}} \\times K}. \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.81, + 0.789, + 0.84 + ], + "angle": 0, + "content": "The reliability of each voxel is then evaluated as the Shannon's entropy [42] of the average teacher predictions, where the unreliable ST voxels are then filtered" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.128 + ], + "angle": 0, + "content": "H. Cao et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.641, + 0.163 + ], + "angle": 0, + "content": "out given a pre-defined quantile \\(\\alpha\\) as in Fig. 2, denoted as:" + }, + { + "type": "equation", + "bbox": [ + 0.331, + 0.172, + 0.786, + 0.211 + ], + "angle": 0, + "content": "\\[\nE _ {i, k} ^ {\\mathrm {m}} = - \\sum_ {c} ^ {K} \\bar {\\mathbf {p}} _ {\\mathrm {r}, c} ^ {\\mathrm {m}} \\log \\bar {\\mathbf {p}} _ {\\mathrm {r}, c} ^ {\\mathrm {m}}, \\quad \\bar {\\mathbf {p}} _ {\\mathrm {r}} ^ {\\mathrm {m}} = \\sum_ {n = 1} ^ {N _ {\\mathrm {r}}} \\psi \\left(\\mathbf {p} _ {\\mathrm {r}, n} ^ {\\mathrm {m}}\\right) / N _ {\\mathrm {r}}, \\tag {7}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.333, + 0.215, + 0.786, + 0.254 + ], + "angle": 0, + "content": "\\[\nh _ {i, k} = \\left\\{ \\begin{array}{l l} 0, & \\text {i f} E _ {i, k} ^ {m} > Q ^ {m} (\\alpha) \\\\ 1, & \\text {i f} E _ {i, k} ^ {m} \\leq Q ^ {m} (\\alpha) \\end{array} , \\right. \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.261, + 0.789, + 0.323 + ], + "angle": 0, + "content": "where \\( E_{i,k}^{\\mathrm{m}} \\) is regarded as the ST entropy of voxel \\( k \\) for modality \\( \\mathrm{m} \\), indicating the modal-specific prediction reliability within this voxel. \\( Q^{m}(\\alpha) \\) is the \\( \\alpha \\)-quantile of all ST entropy in modality \\( \\mathrm{m} \\) in the same batch and \\( \\psi(\\cdot) \\) denotes the Softmax function along the class dimension." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.342, + 0.584, + 0.359 + ], + "angle": 0, + "content": "3.4 ST Voxel Aided Cross-Modal Learning" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.365, + 0.789, + 0.488 + ], + "angle": 0, + "content": "Each modality has its pros and cons under different conditions and cross-modal learning for online scenarios like MM-TTA should therefore possess an adaptive mechanism that can attend to \"pros\" and meanwhile suppress the \"cons\". To achieve that, Latte performs cross-modal learning by attending to the modality with more consistent and certain predictions from a spatial-temporal perspective. Specifically, taking ST voxel \\(\\mathbf{v}_{i,k}^{\\mathrm{ST}}\\) and its corresponding ST entropy \\(E_{i,k}^{\\mathrm{m}}\\) as an example, the voxel-wise cross-modal attending weights and the weighted cross-modal consistency loss \\(\\mathcal{L}_{i,k}^{\\mathrm{xM}}\\) are formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.328, + 0.497, + 0.786, + 0.533 + ], + "angle": 0, + "content": "\\[\nw _ {\\mathrm {v}} ^ {\\mathrm {2 D}} = \\frac {\\exp \\left(E _ {i , k} ^ {\\mathrm {2 D}}\\right)}{\\exp \\left(E _ {i , k} ^ {\\mathrm {2 D}}\\right) + \\exp \\left(E _ {i , k} ^ {\\mathrm {3 D}}\\right)}, w _ {\\mathrm {v}} ^ {\\mathrm {3 D}} = 1 - w _ {\\mathrm {v}} ^ {\\mathrm {2 D}}, \\tag {9}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.327, + 0.536, + 0.785, + 0.555 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {i, k} ^ {\\mathrm {x M}} = w _ {\\mathrm {v}} ^ {\\mathrm {2 D}} D _ {\\mathrm {K L}} \\left(\\bar {\\mathbf {p}} _ {\\mathrm {q}} ^ {\\mathrm {3 D}} \\| \\bar {\\mathbf {p}} _ {\\mathrm {r}} ^ {\\mathrm {2 D}}\\right) + w _ {\\mathrm {v}} ^ {\\mathrm {3 D}} D _ {\\mathrm {K L}} \\left(\\bar {\\mathbf {p}} _ {\\mathrm {q}} ^ {\\mathrm {2 D}} \\| \\bar {\\mathbf {p}} _ {\\mathrm {r}} ^ {\\mathrm {3 D}}\\right), \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.563, + 0.788, + 0.594 + ], + "angle": 0, + "content": "where \\(D_{\\mathrm{KL}}(\\cdot)\\) represents the KL-divergence between two probability. \\(\\bar{\\mathbf{p}}_{\\mathrm{q}}^{3\\mathrm{D}}\\) is the average query predictions in the ST voxel, similarly computed as in Eq. (7)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.595, + 0.789, + 0.702 + ], + "angle": 0, + "content": "The cross-modal attending is further extended from the voxel level to the point level for better online predictions and cross-modal pseudo-label generation. Specifically, the ST entropy is propagated from the ST voxel to its reference point-wise prediction, indicating its confidence and consistency within its spatial-temporal neighborhoods. If one's ST entropy has been filtered out as in Eq. (8), its ST entropy would fall back to the point-level entropy. For arbitrary pointwise predictions \\(\\mathbf{p}^{\\mathrm{m}}\\in \\mathbf{p}_{\\mathrm{r}}^{\\mathrm{m}}\\), the cross-modal predictions \\(y^{\\mathrm{xM}}\\) are formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.335, + 0.722, + 0.786, + 0.761 + ], + "angle": 0, + "content": "\\[\n\\hat {E} _ {\\mathrm {r}} ^ {\\mathrm {m}} = h _ {i, k} E _ {i, k} ^ {\\mathrm {m}} - \\left(1 - h _ {i, k}\\right) \\sum_ {c} ^ {K} \\mathbf {p} _ {c} ^ {\\mathrm {m}} \\log \\mathbf {p} _ {c} ^ {\\mathrm {m}}, \\tag {11}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.33, + 0.765, + 0.785, + 0.801 + ], + "angle": 0, + "content": "\\[\nw _ {\\mathrm {p}} ^ {\\mathrm {2 D}} = \\frac {\\exp \\left(\\hat {E} ^ {\\mathrm {2 D}}\\right)}{\\exp \\left(\\hat {E} ^ {\\mathrm {2 D}}\\right) + \\exp \\left(\\hat {E} ^ {\\mathrm {3 D}}\\right)}, w _ {\\mathrm {p}} ^ {\\mathrm {3 D}} = 1 - w _ {\\mathrm {p}} ^ {\\mathrm {2 D}}, \\tag {12}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.329, + 0.803, + 0.785, + 0.828 + ], + "angle": 0, + "content": "\\[\n\\mathbf {p} ^ {\\mathrm {x M}} = w _ {\\mathrm {p}} ^ {\\mathrm {2 D}} \\mathbf {p} ^ {\\mathrm {2 D}} + w _ {\\mathrm {p}} ^ {\\mathrm {3 D}} \\mathbf {p} ^ {\\mathrm {3 D}}, y ^ {\\mathrm {x M}} = \\underset {c} {\\arg \\max } \\mathbf {p} ^ {\\mathrm {x M}}. \\tag {13}\n\\]" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.242, + 0.114, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Reliable Spatial-Temporal Voxels For Multi-Modal Test-Time Adaptation" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "code_caption", + "bbox": [ + 0.228, + 0.15, + 0.709, + 0.165 + ], + "angle": 0, + "content": "Algorithm 1: Adaptation and Online Prediction Process of Latte" + }, + { + "type": "algorithm", + "bbox": [ + 0.214, + 0.167, + 0.745, + 0.345 + ], + "angle": 0, + "content": "Target Domain: Multi-modal input \\(\\mathcal{X}_{\\mathcal{T}} = \\{(\\mathbf{x}_{t}^{2\\mathrm{D}},\\mathbf{x}_{t}^{3\\mathrm{D}})\\}\\) \nInit Model: Teacher \\(\\tilde{\\phi}_t^{\\mathrm{m}}\\) , student \\(\\phi_t^{\\mathrm{m}}\\) \\(\\mathfrak{m}\\in \\{2\\mathrm{D},3\\mathrm{D}\\}\\) , Slide window size \\(w_{t}\\) \nfor \\(\\pmb {x} = \\{(x_{t}^{2D},x_{t}^{3D})\\}_{t = t_0 + 1}^B\\in \\mathcal{X}_T\\) do 1. Get student and teacher predictions \\(\\tilde{\\mathbf{p}}_t^{\\mathrm{m}},\\mathbf{p}_t^{\\mathrm{m}}\\) by Eq. (1) for \\(i\\in [t_0 + 1,t_0 + B]\\) do 2. Get \\(\\hat{\\mathbf{x}}_i^{3\\mathrm{D}},\\mathbf{v}_i^{3\\mathrm{D}},\\mathbf{M}_i^{3\\mathrm{D}}\\) within the time window of size \\(w_{t}\\) by Eq. (2)-(4) 3. Compute ST voxels \\(\\mathbf{v}_i^{\\mathrm{ST}}\\) and ST entropy \\(E_i^{\\mathrm{m}}\\) by Eq. (5)-(8) 4. Compute \\(\\mathcal{L}_i^{\\mathrm{xM}}\\) by Eq. (9)-(10) end 5. Compute \\(\\hat{E}_{\\mathrm{r}}^{\\mathrm{m}}\\) and get \\(\\mathbf{y}_t^{\\mathrm{xM}}\\) as predictions by Eq. (11)-(13) 6. Compute overall loss by Eq. (14) and update models by Eq. (15) \nend" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.359, + 0.566, + 0.374 + ], + "angle": 0, + "content": "3.5 Online Predictions and Optimization" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.392, + 0.788, + 0.53 + ], + "angle": 0, + "content": "As shown in Fig. 2, given a batch of \\( B \\) consecutive frames \\( \\{\\mathbf{x}_t^{\\mathrm{m}}\\}_{t = t_0 + 1}^{t_0 + B} \\) as input, we first extract the student and teacher predictions through Eq. (1). Subsequently, frames are aggregated and voxelized as in Eq. (2), resulting in a batch of merged point cloud \\( \\{\\hat{\\mathbf{x}}_t^{3\\mathrm{D}}\\}_{t = t_0 + 1}^{t_0 + B} \\) and voxels \\( \\{\\mathbf{v}_t^{3\\mathrm{D}}\\}_{t = t_0 + 1}^{t_0 + B} \\). The ST voxels and ST entropy computation process is then applied to each frame of the merged point cloud and voxels. Due to the overlap of the slide window, multiple ST entropy values could be propagated to the same point prediction, where we empirically take the average value of all ST entropy received to proceed Eq. (11). The overall loss function and the updating scheme can be formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.358, + 0.544, + 0.786, + 0.579 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\sum_ {t} \\mathcal {F} \\left(\\mathbf {p} _ {t} ^ {\\mathrm {m}}, \\mathbf {y} _ {t} ^ {\\mathrm {x M}}\\right) + \\frac {\\lambda_ {\\mathrm {x M}}}{B} \\sum_ {t} \\sum_ {k} \\mathcal {L} _ {t, k} ^ {\\mathrm {x M}}, \\tag {14}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.351, + 0.582, + 0.785, + 0.601 + ], + "angle": 0, + "content": "\\[\n\\tilde {\\theta} _ {t} ^ {\\mathrm {m}} = \\lambda_ {s} \\tilde {\\theta} _ {t - 1} ^ {\\mathrm {m}} + (1 - \\lambda_ {s}) \\theta_ {t} ^ {\\mathrm {m}}, \\tag {15}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.616, + 0.788, + 0.709 + ], + "angle": 0, + "content": "where \\(\\mathcal{F}(\\cdot)\\) is the cross-entropy function and \\(\\mathbf{y}_t^{\\mathrm{xM}}\\) is the point-wise cross-modal pseudo-label frame from Eq. (13), which is also regarded as our cross-modal prediction for evaluation. \\(\\tilde{\\theta}_t^{\\mathrm{m}}\\) and \\(\\theta_t^{\\mathrm{m}}\\) are the model parameters of teacher and student models, respectively, where \\(\\lambda_s\\) is the momentum update coefficient. \\(\\lambda_{\\mathrm{xM}}\\) is pre-defined coefficient for the cross-modal consistency loss. Our adaptation and inference process is summarized in Algo. 1" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.739, + 0.462, + 0.757 + ], + "angle": 0, + "content": "4 Experimental Results" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.84 + ], + "angle": 0, + "content": "In this section, we present our thorough experimental results on three different benchmarks. Details of benchmarks, backbones, and settings are first present in Sec. 4.1. The main results are then illustrated in Sec. 4.2, followed by detailed ablation studies and qualitative results in Sec. 4.3." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.128 + ], + "angle": 0, + "content": "H. Cao et al." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.147, + 0.471, + 0.163 + ], + "angle": 0, + "content": "4.1 Benchmarks and settings" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.175, + 0.788, + 0.431 + ], + "angle": 0, + "content": "Benchmark Details. To thoroughly investigate the effectiveness of Latte, we conduct our experiments on three different MM-TTA benchmarks: (i) USA-to-Singapore (U-to-S), (ii) A2D2-to-SemanticKITTI (A-to-S), and (iii) Synthia-to-SemanticKITTI (S-to-S). Specifically, U-to-S is developed from NuScenes-LiDARSeg [2] dataset, where the domain gap is mainly credited to the infrastructure difference between countries. Same as [16], the source domain and the target domain data are selected by filtering country keywords based on the data recording description in NuScenes. Different from previous UDA methods [16,24] utilizing self-defined class mappings, we follow the official mapping in NuScenes-LiDARSeg, forming a 16-class segmentation benchmark. The remaining two benchmarks follow a similar setting as in [29], where the domain gap of A-to-S [1,10] lies in different LiDAR mounting positions and image resolutions while the one of S-to-S [1] lies in different patterns between synthetic and real data for both modalities. For A-to-S, we adopt the same class-mapping as in [16,29] which leads to a 10-class benchmark, while we design our class-mapping for S-to-S since its original class map has not been revealed previously [29], forming a 9-class benchmark. More details are presented in our appendix." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.432, + 0.788, + 0.612 + ], + "angle": 0, + "content": "Baseline methods. Previous TTA and MM-TTA methods are included in comparison with Latte. For TTA methods, we compare Latte with point-wise pseudolabels with filtering (PsLabel), TENT [39], ETA [22], SAR [23] with and without restoring (denoted as SAR-rs and SAR, respectively). MMTTA [29] as well as its online version of xMUDA [16] (including pure xMUDA and pseudo-label version xMUDA+PL) are additionally included as the previous SOTA methods for comparison. Considering that EATA [22] with Fisher regularizer requires pre-access to the target domain samples, we compare Latte with ETA, an official variant of EATA without this regularizer for a fair comparison. We further present a performance lower bound by directly testing with source pre-trained models (i.e., Source only) and an upper bound by online adapting networks with ground-truth labels from the target domain for one epoch (i.e., Oracle TTA)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.614, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Implementation details. With the rapid progress in 2D and 3D segmentation, we employ all baseline methods and Latte with more recent SegFormer [43] (SegFormer-B1) and SPVCNN [33] (SPVCNN-cr1) as their 2D and 3D backbones, respectively, to investigate the improvement of existing TTA methods when combined with advanced networks. The pre-training procedures on the source domain follow a similar setting as in [16], except for 2D SegFormer which is trained with a base learning rate of 6E-5 with an AdamW optimizer and Poly scheduler as in [43]. For Latte, we maintain the same learning rate and optimizer as in the pre-training stage, while the scheduler is disabled. For our voxelization, we leverage the off-the-shelf SLAM algorithm KISS-ICP [36] to generate poses and utilize a window size \\( w_{t} \\) of 3 with a voxel size of \\( 0.2\\mathrm{m} \\), while the quantile \\( \\alpha \\) and coefficient \\( \\lambda_{\\mathrm{xM}} \\) are set to 0.9 and 0.3, respectively. \\( \\lambda_{s} \\) is empirically set to 0.99 as in [4,31]. For all methods, we strictly follow the one-pass protocol to first evaluate the predictions of the input batch and then update the networks, where we only update the trainable parameters in normalization layers. For all base" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.242, + 0.114, + 0.732, + 0.131 + ], + "angle": 0, + "content": "Reliable Spatial-Temporal Voxels For Multi-Modal Test-Time Adaptation" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.145, + 0.788, + 0.215 + ], + "angle": 0, + "content": "Table 1: Performance (mIoU) comparison of Latte. Latte outperforms all previous SOTA methods on the cross-modal metric (xM) across three benchmarks. Here \"MM\" denotes whether the method contains multi-modal interaction or not. Cross-modal predictions are computed as the Softmax average of modal outputs except for Latte. \"Avg\" present the average xM performance across three benchmarks." + }, + { + "type": "table", + "bbox": [ + 0.221, + 0.228, + 0.782, + 0.366 + ], + "angle": 0, + "content": "
MethodPublicationMMU-to-SxMA-to-SxMS-to-SAvg
2D3D2D3D2D3D
Source only-X31.441.143.947.417.944.323.436.448.2
Oracle TTA-X38.745.550.349.161.362.536.054.554.6
TENT [39]ICML-21X36.836.141.143.344.449.122.435.537.5
ETA [22]ICML-22X36.734.743.743.242.549.722.430.632.9
SAR [23]ICLR-23X36.637.243.943.042.750.320.731.332.9
SAR-rs [23]ICLR-23X36.635.543.943.143.850.124.225.128.8
xMUDA [16]PAMI-22✓19.422.924.213.133.032.212.214.814.1
xMUDA+PL [16]PAMI-22✓36.238.243.042.848.350.930.934.136.3
MMTTA [29]CVPR-22✓37.241.545.444.551.753.727.535.135.5
PsLabel-✓37.735.841.846.744.350.029.233.335.3
Latte (Ours)-✓37.441.046.046.152.654.333.239.341.6
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.374, + 0.79, + 0.435 + ], + "angle": 0, + "content": "lines, we conduct a parameter search and report their best results. We utilize the mean intersection over union (mIoU) as our evaluation metric. All experiments are conducted with PyTorch on a single RTX 3090. More details are presented in our appendix." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.462, + 0.39, + 0.476 + ], + "angle": 0, + "content": "4.2 Overall Results" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.492, + 0.788, + 0.629 + ], + "angle": 0, + "content": "Tab. 1 presents the overall results of our Latte on three benchmarks compared with previous SOTA methods. In terms of average cross-modal predictions (Avg), Latte surpasses all previous TTA or MM-TTA methods, with a significant relative improvement of more than \\(5.3\\%\\). For the less challenging U-to-S and A-to-S (i.e., gap between Oracle TTA and best TTA less than \\(10\\%\\)), Latte achieves consistent relative improvements of \\(1.3\\%\\) on U-to-S and \\(1.1\\%\\) on A-to-S compared to MMTTA, respectively. Most existing methods except for xMUDA [16] can achieve superior performance compared to source-only results when one modality is reliable as in U-to-S (3D prevails) and A-to-S (2D prevails)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.63, + 0.788, + 0.842 + ], + "angle": 0, + "content": "In terms of the most challenging S-to-S, Latte exhibits significant performance improvement compared to previous SOTA MM-TTA methods such as MMTTA [29] and xMUDA+PL [16], with a relative gap of \\(14.6\\%\\). In fact, all previous methods except for our Latte perform inferiorly compared to source pre-trained models (Source only), due to the significant domain gaps caused by the non-trivial pattern difference in both modalities between synthetic and real data. In this case, the instability of single-frame predictions becomes exacerbated, leading to performance degeneration in MMTTA which mainly relies on single-frame refinement. This justifies the effectiveness of Latte under either challenging or less challenging TTA scenarios. Another interesting observation is that almost all cross-modal methods outperform other single-modal TTA methods, e.g., even the simplest PsLabel can outperform recent TTA methods like SAR [23]. This reveals the non-trivial benefits of cross-modal learning for TTA with segmentation. More comparisons about efficiency are included in appendix." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.128 + ], + "angle": 0, + "content": "H. Cao et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.145, + 0.788, + 0.215 + ], + "angle": 0, + "content": "Table 2: Ablation studies about the effectiveness of cross-modal learning. When disabling \\(\\mathbf{y}^{\\mathrm{xM}}\\), we utilize the average Softmax of all modalities as the pseudo-labels as a replacement. In experiment No. 8, we replace the ST entropy (ST ety) with point-wise entropy in Eq. (13) and set \\(w_{\\mathrm{v}}^{\\mathrm{m}}\\) in Eq. (10) to 0.5. Here \\(\\mathcal{L}_{32}^{\\mathrm{xM}}\\) and \\(\\mathcal{L}_{23}^{\\mathrm{xM}}\\) refer to the former half and the latter half in Eq. 10, respectively." + }, + { + "type": "table", + "bbox": [ + 0.232, + 0.228, + 0.769, + 0.351 + ], + "angle": 0, + "content": "
No.Description\\( y^{xM} \\)\\( \\mathcal{L}_{23}^{xM} \\)\\( \\mathcal{L}_{32}^{xM} \\)U-to-SA-to-SS-to-S
2D3DxM2D3DxM2D3DxM
0Source only---31.441.143.947.417.944.323.436.438.2
1only ST ety label✓36.834.543.743.342.249.922.430.633.3
2only ST 2D→3D✓32.534.541.041.742.248.227.730.632.9
3only ST 3D→2D✓36.839.642.643.350.248.422.426.725.6
4ST ety label + ST 2D→3D✓✓37.535.841.043.647.549.831.036.038.4
5ST ety label + ST 3D→2D✓✓37.440.945.946.252.153.432.538.739.7
6ST 2D↔3D✓✓36.840.343.844.047.051.230.435.637.7
7w/o percentile in Eq. (8)✓✓✓37.440.345.445.751.453.232.939.041.7
8Point ety vs. ST ety✓✓✓37.040.645.145.151.452.731.636.738.2
9Latte✓✓✓37.441.046.046.152.654.333.239.341.6
" + }, + { + "type": "image", + "bbox": [ + 0.276, + 0.361, + 0.727, + 0.452 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.462, + 0.788, + 0.534 + ], + "angle": 0, + "content": "Fig. 3: Ablation studies of different frame aggregation mechanisms. Besides (a) Latte's slide window aggregation, we test two different aggregation methods, including (b) replacing single frame student predictions with multiple frames in Eq. (5) and (c) adding non-overlapping aggregation to (b). Results show that our slide window aggregation can better evaluate the local consistency and results in consistent improvement." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.54, + 0.55, + 0.555 + ], + "angle": 0, + "content": "4.3 Ablation Studies and Visualization" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.563, + 0.787, + 0.594 + ], + "angle": 0, + "content": "In this section, we investigate the effectiveness of each component in Latte by thorough ablation studies and qualitative results." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.614, + 0.788, + 0.794 + ], + "angle": 0, + "content": "Effectiveness of cross-modal learning. The main optimization objectives of Latte contain three parts, including ST-entropy-based cross-modal consistency (both 2D-to-3D and 3D-to-2D) and cross-entropy loss between predictions and cross-modal pseudo-labels. We justify the effectiveness by using different combinations of these components as in Tab. 2. As shown from experiments No. 1 to No. 6, disabling any part of Latte causes performance decreases of more than \\(1.6\\%\\) relatively on the cross-modal prediction (xM), which proves the effectiveness of each component we included. A special case is No. 5 on U-to-S, where disabling ST \\(\\mathcal{L}_{23}^{\\mathrm{xM}}\\) only leads to a minor \\(0.1\\%\\) drop in accuracy. This is mainly because the 2D performance is already close to saturation (37.4 vs. 38.7 of Oracle TTA) without the guidance of 3D predictions. In this case, introducing 3D information to 2D is less effective compared to the other benchmarks." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.796, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Besides optimization objects, we further testify two important components of Latte, including ST-entropy-based weighting in Eq. (10) and Eq. (13) as well as percentile filtering in Eq. (8). Specifically, disabling percentile filtering as in No." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.242, + 0.115, + 0.732, + 0.13 + ], + "angle": 0, + "content": "Reliable Spatial-Temporal Voxels For Multi-Modal Test-Time Adaptation" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "image", + "bbox": [ + 0.26, + 0.145, + 0.414, + 0.223 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.424, + 0.145, + 0.577, + 0.223 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.589, + 0.145, + 0.741, + 0.223 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.368, + 0.224, + 0.636, + 0.236 + ], + "angle": 0, + "content": "(a) Ablation studies of various ST voxel sizes." + }, + { + "type": "image", + "bbox": [ + 0.261, + 0.237, + 0.412, + 0.314 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.425, + 0.237, + 0.574, + 0.314 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.589, + 0.237, + 0.74, + 0.314 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.341, + 0.316, + 0.662, + 0.328 + ], + "angle": 0, + "content": "(b) Ablation studies of filtering percentile \\(\\alpha\\) in Eq. (8)." + }, + { + "type": "image_caption", + "bbox": [ + 0.239, + 0.339, + 0.763, + 0.354 + ], + "angle": 0, + "content": "Fig. 4: Sensitivity analysis about the ST voxel size and confident percentile \\(\\alpha\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.361, + 0.788, + 0.422 + ], + "angle": 0, + "content": "7 leads to a relative accuracy drop of more than \\(0.5\\%\\) across all benchmarks while utilizing point-wise entropy instead of our ST entropy as in No. 8 degenerates the performance by more than \\(2.0\\%\\) relatively, which justifies that ST entropy can better estimate the prediction reliability compared to the point-wise entropy." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.442, + 0.789, + 0.73 + ], + "angle": 0, + "content": "Different frame aggregation mechanisms. One of the core designs of Latte is the slide window frame aggregation, which is designed to capture temporally local correspondences and estimate their prediction consistency. To justify the aforementioned claim, we compare our slide window aggregation with two aggregation variants as shown in Fig. 3. Specifically, variant (b) regards all student prediction frames rather a single frame within the time window as the query in Eq. (5), while variant (c) further utilizes a non-overlapping sliding strategy. According to Fig. 3, our slide window aggregation surpasses the other two aggregation variants under all sizes of time windows across three benchmarks, which justifies the effectiveness of our method. The inferior performance of variants (b) and (c) could be due to two factors. Firstly, aggregating student predictions could be an inferior option since student predictions tend to be less stable, and aggregating multiple frames can therefore lead to noisy queries for cross-modal learning. Secondly, non-overlapping aggregation can result in a miss-checking area on the boundary of each time window, which leads to unrepresentative entropy estimation for some ST voxels. Another observation is that utilizing a smaller slide window can usually yield better performance since temporally local consistency can be captured. This justifies the effectiveness of estimating prediction reliability through temporally local consistency." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.79, + 0.841 + ], + "angle": 0, + "content": "Parameter sensitivity analysis. To testify whether Latte is sensitive towards hyper-parameter settings, we conduct sensitivity analysis on Latte across three benchmarks. Here we illustrate the sensitivity analysis of two hyper-parameters of Latte, including ST voxel size and filtering percentile \\(\\alpha\\) as shown in Fig. 4. In terms of ST voxel sizes, one can tell from Fig. 4a that all benchmarks share the same optimal voxel size as 0.2, where performance drops as the voxel size" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.127 + ], + "angle": 0, + "content": "H. Cao et al." + }, + { + "type": "image", + "bbox": [ + 0.245, + 0.145, + 0.758, + 0.306 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.315, + 0.785, + 0.343 + ], + "angle": 0, + "content": "Fig. 5: Qualitative results of Latte on S-to-S. Here we visualize modality-specific predictions and cross-modal predictions from Latte during two sets of consecutive frames." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.351, + 0.788, + 0.487 + ], + "angle": 0, + "content": "becomes too large or too small. This is mainly because large voxels could cause boundary ambiguity while small voxels contain insufficient representative points for evaluation. As for filtering percentile \\(\\alpha\\), non-trivial improvement can be observed on U-to-S and A-to-S when \\(\\alpha \\geq 0.8\\), while a lower \\(\\alpha\\) begins to discard more confident ST voxels leading to a slight drop in accuracy. On S-to-S, Latte without filtering (i.e., No. 7 in Tab. 2) performs slightly better on cross-modal predictions with a minor gap of \\(0.1\\%\\), while it can also be observed that setting \\(\\alpha\\) as 0.9 leads to an improvement of \\(0.3\\%\\) on both modality-specific predictions, which justifies the overall effectiveness of our filtering procedures." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.506, + 0.788, + 0.673 + ], + "angle": 0, + "content": "Qualitative results. Our main motivation is utilizing prediction consistency of spatial-temporal correspondences to estimate modality reliability and achieve better cross-modal attending. As shown in Fig. 5, Latte can effectively attend the modality with more consistent predictions in spatial-temporal correspondences and therefore improve the cross-modal prediction consistency across time. For instance, Latte successfully attends to the stable 2D predictions pedestrian in the upper set of consecutive frames, suppressing the spatial-temporal inconsistency from 3D predictions. Similar observations can also be found in the lower sample, where the more consistent 2D predictions of buildings in the red rectangle and 3D predictions of roads in the blue rectangle prevail. More qualitative comparisons with previous SOTA methods are presented in our appendix." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.691, + 0.36, + 0.708 + ], + "angle": 0, + "content": "5 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.788, + 0.84 + ], + "angle": 0, + "content": "In this paper, we propose a novel MM-TTA method called Latte, which leverages prediction consistency of spatial-temporal correspondences in MM-TTA for 3D segmentation. Latte utilizes a slide window frame aggregation to extract ST voxels and estimates temporally local consistency by the ST entropy of each modality. By attending to the modality with more consistency, Latte can achieve stable improvement on challenging MM-TTA scenarios. Despite its effectiveness, Latte struggles to rectify consistently wrong predictions across time, which is challenging and worth further investigation in the future." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.242, + 0.115, + 0.732, + 0.13 + ], + "angle": 0, + "content": "Reliable Spatial-Temporal Voxels For Multi-Modal Test-Time Adaptation" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.146, + 0.383, + 0.163 + ], + "angle": 0, + "content": "Acknowledgment" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.176, + 0.788, + 0.252 + ], + "angle": 0, + "content": "This research is supported by the National Research Foundation, Singapore, under the NRF Medium Sized Centre for Advanced Robotics Technology Innovation (CARTIN). Any opinions, findings and conclusions or recommendations expressed in this material are those of the author(s) and do not reflect the views of National Research Foundation, Singapore." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.274, + 0.323, + 0.289 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.303, + 0.785, + 0.359 + ], + "angle": 0, + "content": "1. Behley, J., Garbade, M., Milioto, A., Quenzel, J., Behnke, S., Stachniss, C., Gall, J.: Semantickitti: A dataset for semantic scene understanding of lidar sequences. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 9297-9307 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.36, + 0.785, + 0.414 + ], + "angle": 0, + "content": "2. Caesar, H., Bankiti, V., Lang, A.H., Vora, S., Liong, V.E., Xu, Q., Krishnan, A., Pan, Y., Baldan, G., Beijbom, O.: nuscenes: A multimodal dataset for autonomous driving. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11621-11631 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.415, + 0.785, + 0.455 + ], + "angle": 0, + "content": "3. Cao, H., Xu, Y., Yang, J., Yin, P., Yuan, S., Xie, L.: Mopa: Multi-modal prior aided domain adaptation for 3d semantic segmentation. arXiv preprint arXiv:2309.11839 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.456, + 0.785, + 0.509 + ], + "angle": 0, + "content": "4. Cao, H., Xu, Y., Yang, J., Yin, P., Yuan, S., Xie, L.: Multi-modal continual test-time adaptation for 3d semantic segmentation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 18809-18819 (October 2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.511, + 0.785, + 0.564 + ], + "angle": 0, + "content": "5. Chen, R., Liu, Y., Kong, L., Zhu, X., Ma, Y., Li, Y., Hou, Y., Qiao, Y., Wang, W.: Clip2scene: Towards label-efficient 3d scene understanding by clip. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7020-7030 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.566, + 0.785, + 0.607 + ], + "angle": 0, + "content": "6. Choy, C., Gwak, J., Savarese, S.: 4d spatio-temporal convnets: Minkowski convolutional neural networks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3075-3084 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.608, + 0.785, + 0.648 + ], + "angle": 0, + "content": "7. Fan, H., Yang, Y., Kankanhalli, M.: Point 4d transformer networks for spatiotemporal modeling in point cloud videos. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14204-14213 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.649, + 0.785, + 0.675 + ], + "angle": 0, + "content": "8. Fan, H., Yu, X., Ding, Y., Yang, Y., Kankanhalli, M.: Pistnet: Point spatio-temporal convolution on point cloud sequences. arXiv preprint arXiv:2205.13713 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.676, + 0.785, + 0.731 + ], + "angle": 0, + "content": "9. Feng, D., Haase-Schütz, C., Rosenbaum, L., Hertlein, H., Glaeser, C., Timm, F., Wiesbeck, W., Dietmayer, K.: Deep multi-modal object detection and semantic segmentation for autonomous driving: Datasets, methods, and challenges. IEEE Transactions on Intelligent Transportation Systems 22(3), 1341-1360 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.732, + 0.785, + 0.772 + ], + "angle": 0, + "content": "10. Geyer, J., Kassahun, Y., Mahmudi, M., Ricou, X., Durgesh, R., Chung, A.S., Hauswald, L., Pham, V.H., Mühlegg, M., Dorn, S., et al.: A2d2: Audi autonomous driving dataset. arXiv preprint arXiv:2004.06320 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.773, + 0.785, + 0.813 + ], + "angle": 0, + "content": "11. Gong, T., Jeong, J., Kim, T., Kim, Y., Shin, J., Lee, S.J.: Robust continual test-time adaptation: Instance-aware bn and prediction-balanced memory. arXiv preprint arXiv:2208.05117 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.814, + 0.785, + 0.84 + ], + "angle": 0, + "content": "12. Goyal, S., Sun, M., Raghunathan, A., Kolter, J.Z.: Test time adaptation via conjugate pseudo-labels. In: Advances in Neural Information Processing Systems (2022)" + }, + { + "type": "list", + "bbox": [ + 0.22, + 0.303, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.36, + 0.127 + ], + "angle": 0, + "content": "H. Cao et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.148, + 0.785, + 0.175 + ], + "angle": 0, + "content": "13. Graham, B.: Sparse 3d convolutional neural networks. arXiv preprint arXiv:1505.02890 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.176, + 0.785, + 0.216 + ], + "angle": 0, + "content": "14. Guo, Y., Wang, H., Hu, Q., Liu, H., Liu, L., Bennamoun, M.: Deep learning for 3d point clouds: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence 43(12), 4338-4364 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.216, + 0.785, + 0.257 + ], + "angle": 0, + "content": "15. Huang, S., Gojcic, Z., Huang, J., Wieser, A., Schindler, K.: Dynamic 3d scene analysis by point cloud accumulation. In: European Conference on Computer Vision. pp. 674-690. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.257, + 0.785, + 0.298 + ], + "angle": 0, + "content": "16. Jaritz, M., Vu, T.H., De Charette, R., Wirbel, E., Pérez, P.: Cross-modal learning for domain adaptation in 3d semantic segmentation. IEEE Transactions on Pattern Analysis and Machine Intelligence 45(2), 1533-1544 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.298, + 0.785, + 0.324 + ], + "angle": 0, + "content": "17. Ji, X., Yuan, S., Yin, P., Xie, L.: Lio-gvm: an accurate, tightly-coupled lidar-inertial odometry with gaussian voxel map. IEEE Robotics and Automation Letters (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.324, + 0.785, + 0.378 + ], + "angle": 0, + "content": "18. Li, M., Zhang, Y., Xie, Y., Gao, Z., Li, C., Zhang, Z., Qu, Y.: Cross-domain and cross-modal knowledge distillation in domain adaptation for 3d semantic segmentation. In: Proceedings of the 30th ACM International Conference on Multimedia. pp. 3829-3837 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.379, + 0.785, + 0.419 + ], + "angle": 0, + "content": "19. Liang, J., Hu, D., Feng, J.: Do we really need to access the source data? source hypothesis transfer for unsupervised domain adaptation. In: International Conference on Machine Learning. pp. 6028-6039. PMLR (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.42, + 0.785, + 0.473 + ], + "angle": 0, + "content": "20. Liu, W., Luo, Z., Cai, Y., Yu, Y., Ke, Y., Junior, J.M., Gonçalves, W.N., Li, J.: Adversarial unsupervised domain adaptation for 3d semantic segmentation with multi-modal learning. ISPRS Journal of Photogrammetry and Remote Sensing 176, 211-221 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.473, + 0.785, + 0.514 + ], + "angle": 0, + "content": "21. Liu, Y., Kothari, P., Van Delft, B., Bellot-Gurlet, B., Mordan, T., Alahi, A.: Ttt++: When does self-supervised test-time training fail or thrive? Advances in Neural Information Processing Systems 34, 21808-21820 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.514, + 0.785, + 0.554 + ], + "angle": 0, + "content": "22. Niu, S., Wu, J., Zhang, Y., Chen, Y., Zheng, S., Zhao, P., Tan, M.: Efficient test-time model adaptation without forgetting. In: International Conference on Machine Learning. pp. 16888-16905. PMLR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.554, + 0.785, + 0.595 + ], + "angle": 0, + "content": "23. Niu, S., Wu, J., Zhang, Y., Wen, Z., Chen, Y., Zhao, P., Tan, M.: Towards stable test-time adaptation in dynamic wild world. arXiv preprint arXiv:2302.12400 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.595, + 0.785, + 0.65 + ], + "angle": 0, + "content": "24. Peng, D., Lei, Y., Li, W., Zhang, P., Guo, Y.: Sparse-to-dense feature matching: Intra and inter domain cross-modal learning in domain adaptation for 3d semantic segmentation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 7108-7117 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.65, + 0.785, + 0.69 + ], + "angle": 0, + "content": "25. Piergiovanni, A., Casser, V., Ryoo, M.S., Angelova, A.: 4d-net for learned multimodal alignment. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 15435-15445 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.691, + 0.785, + 0.731 + ], + "angle": 0, + "content": "26. Qi, C.R., Su, H., Mo, K., Guibas, L.J.: Pointnet: Deep learning on point sets for 3d classification and segmentation. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 652-660 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.731, + 0.785, + 0.786 + ], + "angle": 0, + "content": "27. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International Conference on Machine Learning. pp. 8748-8763. PMLR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.786, + 0.785, + 0.84 + ], + "angle": 0, + "content": "28. Saltori, C., Krivosheev, E., Lathuilière, S., Sebe, N., Galasso, F., Fiameni, G., Ricci, E., Poiesi, F.: Gipso: Geometrically informed propagation for online adaptation in 3d lidar segmentation. In: European Conference on Computer Vision. pp. 567-585. Springer (2022)" + }, + { + "type": "list", + "bbox": [ + 0.218, + 0.148, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.242, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Reliable Spatial-Temporal Voxels For Multi-Modal Test-Time Adaptation" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.203 + ], + "angle": 0, + "content": "29. Shin, I., Tsai, Y.H., Zhuang, B., Schulter, S., Liu, B., Garg, S., Kweon, I.S., Yoon, K.J.: Mm-tta: multi-modal test-time adaptation for 3d semantic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16928-16937 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.203, + 0.785, + 0.258 + ], + "angle": 0, + "content": "30. Simons, C., Raychaudhuri, D.S., Ahmed, S.M., You, S., Karydis, K., Roy-Chowdhury, A.K.: Summit: Source-free adaptation of uni-modal models to multimodal targets. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 1239-1249 (October 2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.258, + 0.785, + 0.299 + ], + "angle": 0, + "content": "31. Song, J., Lee, J., Kweon, I.S., Choi, S.: Ecotta: Memory-efficient continual test-time adaptation via self-distilled regularization. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11920-11929 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.299, + 0.785, + 0.339 + ], + "angle": 0, + "content": "32. Su, Y., Xu, X., Li, T., Jia, K.: Revisiting realistic test-time training: Sequential inference and adaptation by anchored clustering regularized self-training. arXiv preprint arXiv:2303.10856 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.339, + 0.785, + 0.38 + ], + "angle": 0, + "content": "33. Tang, H., Liu, Z., Zhao, S., Lin, Y., Lin, J., Wang, H., Han, S.: Searching efficient 3d architectures with sparse point-voxel convolution. In: European Conference on Computer Vision. pp. 685–702. Springer (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.38, + 0.785, + 0.42 + ], + "angle": 0, + "content": "34. Tarvainen, A., Valpola, H.: Mean teachers are better role models: Weight-averaged consistency targets improve semi-supervised deep learning results. Advances in Neural Information Processing Systems 30 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.421, + 0.785, + 0.461 + ], + "angle": 0, + "content": "35. Vedula, S., Baker, S., Rander, P., Collins, R., Kanade, T.: Three-dimensional scene flow. In: Proceedings of the Seventh IEEE International Conference on Computer Vision. vol. 2, pp. 722-729. IEEE (1999)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.461, + 0.785, + 0.515 + ], + "angle": 0, + "content": "36. Vizzo, I., Guadagnino, T., Mersch, B., Wiesmann, L., Behley, J., Stachniss, C.: Kiss-icp: In defense of point-to-point icp-simple, accurate, and robust registration if done the right way. IEEE Robotics and Automation Letters 8(2), 1029-1036 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.515, + 0.785, + 0.556 + ], + "angle": 0, + "content": "37. Vogel, C., Schindler, K., Roth, S.: 3d scene flow estimation with a rigid motion prior. In: 2011 International Conference on Computer Vision. pp. 1291-1298. IEEE (2011)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.556, + 0.785, + 0.583 + ], + "angle": 0, + "content": "38. Vogel, C., Schindler, K., Roth, S.: Piecewise rigid scene flow. In: Proceedings of the IEEE International Conference on Computer Vision. pp. 1377-1384 (2013)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.583, + 0.785, + 0.624 + ], + "angle": 0, + "content": "39. Wang, D., Shelhamer, E., Liu, S., Olshausen, B., Darrell, T.: Tent: Fully test-time adaptation by entropy minimization. In: International Conference on Learning Representations (2021), https://openreview.net/forum?id=uX13bZLkr3c" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.624, + 0.785, + 0.651 + ], + "angle": 0, + "content": "40. Wang, J.K., Wibisono, A.: Towards understanding gd with hard and conjugate pseudo-labels for test-time adaptation. arXiv preprint arXiv:2210.10019 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.651, + 0.785, + 0.691 + ], + "angle": 0, + "content": "41. Wang, Q., Fink, O., Van Gool, L., Dai, D.: Continual test-time domain adaptation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7201-7211 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.691, + 0.785, + 0.718 + ], + "angle": 0, + "content": "42. Wyner, A.: Recent results in the shannon theory. IEEE Transactions on information Theory 20(1), 2-10 (1974)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.718, + 0.785, + 0.759 + ], + "angle": 0, + "content": "43. Xie, E., Wang, W., Yu, Z., Anandkumar, A., Alvarez, J.M., Luo, P.: Segformer: Simple and efficient design for semantic segmentation with transformers. Advances in Neural Information Processing Systems 34, 12077-12090 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.759, + 0.785, + 0.799 + ], + "angle": 0, + "content": "44. Xing, B., Ying, X., Wang, R., Yang, J., Chen, T.: Cross-modal contrastive learning for domain adaptation in 3d semantic segmentation. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 37, pp. 2974-2982 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.799, + 0.785, + 0.841 + ], + "angle": 0, + "content": "45. Xu, J., Miao, Z., Zhang, D., Pan, H., Liu, K., Hao, P., Zhu, J., Sun, Z., Li, H., Zhan, X.: Int: Towards infinite-frames 3d detection with an efficient framework. In: European Conference on Computer Vision. pp. 193–209. Springer (2022)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.127 + ], + "angle": 0, + "content": "H. Cao et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.189 + ], + "angle": 0, + "content": "46. Yin, P., Cao, H., Nguyen, T.M., Yuan, S., Zhang, S., Liu, K., Xie, L.: Outram: One-shot global localization via triangulated scene graph and global outlier pruning. arXiv preprint arXiv:2309.08914 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.19, + 0.785, + 0.231 + ], + "angle": 0, + "content": "47. Yuan, L., Xie, B., Li, S.: Robust test-time adaptation in dynamic scenarios. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15922-15932 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.232, + 0.785, + 0.273 + ], + "angle": 0, + "content": "48. Zhang, M., Levine, S., Finn, C.: Memo: Test time robustness via adaptation and augmentation. Advances in Neural Information Processing Systems 35, 38629-38642 (2022)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.273 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/Reliable Spatial-Temporal Voxels For Multi-Modal Test-Time Adaptation/c906fc92-ebc1-45f1-bc7d-fa0568ba9181_origin.pdf b/2024/Reliable Spatial-Temporal Voxels For Multi-Modal Test-Time Adaptation/c906fc92-ebc1-45f1-bc7d-fa0568ba9181_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..7aeac9a9787d846b107dbf0aae199838eb92da7e --- /dev/null +++ b/2024/Reliable Spatial-Temporal Voxels For Multi-Modal Test-Time Adaptation/c906fc92-ebc1-45f1-bc7d-fa0568ba9181_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e6fe429ba88442a544f672be8a6d056a876f243c6c978c16e19cad19a69ad312 +size 1216165 diff --git a/2024/Reliable Spatial-Temporal Voxels For Multi-Modal Test-Time Adaptation/full.md b/2024/Reliable Spatial-Temporal Voxels For Multi-Modal Test-Time Adaptation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..bc889726f4349996c09c5eee7c0a5276c449ae56 --- /dev/null +++ b/2024/Reliable Spatial-Temporal Voxels For Multi-Modal Test-Time Adaptation/full.md @@ -0,0 +1,296 @@ +# Reliable Spatial-Temporal Voxels For Multi-Modal Test-Time Adaptation + +Haozhi Cao $^{1}$ , Yuecong Xu $^{1,2}$ , Jianfei Yang $^{1,3}$ , Pengyu Yin $^{1}$ , Xingyu Ji $^{1}$ , Shenghai Yuan $^{1}$ , and Lihua Xie $^{1}$ + +1 School of Electrical and Electronic Engineering, Nanyang Technological University {haozhi002, jianfei.yang, pengyu001, xinyu001, shyuan, elhxie}@ntu.edu.sg + +$^{2}$ Department of Electrical and Computer Engineering, National University of Singapore yc.xu@nus.edu.sg + +3 School of Mechanical and Aerospace Engineering, Nanyang Technological University + +Abstract. Multi-modal test-time adaptation (MM-TTA) is proposed to adapt models to an unlabeled target domain by leveraging the complementary multi-modal inputs in an online manner. Previous MM-TTA methods for 3D segmentation rely on predictions of cross-modal information in each input frame, while they ignore the fact that predictions of geometric neighborhoods within consecutive frames are highly correlated, leading to unstable predictions across time. To fulfill this gap, we propose ReLiable Spatial-temporal Voxels (Latte), an MM-TTA method that leverages reliable cross-modal spatial-temporal correspondences for multi-modal 3D segmentation. Motivated by the fact that reliable predictions should be consistent with their spatial-temporal correspondences, Latte aggregates consecutive frames in a slide window manner and constructs Spatial-Temopral (ST) voxels to capture temporally local prediction consistency for each modality. After filtering out ST voxels with high ST entropy, Latte conducts cross-modal learning for each point and pixel by attending to those with reliable and consistent predictions among both spatial and temporal neighborhoods. Experimental results show that Latte achieves state-of-the-art performance on three different MM-TTA benchmarks compared to previous MM-TTA or TTA methods. Visit our project site https://sites.google.com/view/eccv24-latte. + +Keywords: Test-time adaptation $\cdot$ Multi-modal learning $\cdot$ 3D semantic segmentation + +# 1 Introduction + +3D semantic segmentation is a fundamental task to achieve various fully autonomous applications such as autonomous driving and robot navigation [9,14]. With the increasing urge for robust sensing, multi-modal sensors (e.g., cameras and LiDARs) are widely adopted in autonomous systems containing rich complementary information from different sensors. However, conventional deep-learning-based solutions lean on expensive point-wise annotations and perform + +![](images/b8e4c8e876345fb3c69ae9096facc9e311366ae042e737c0a55381edfca585f2.jpg) +Fig. 1: Illustration of previous MM-TTA methods and our Latte. (a) visualizes the predictions from the previous state-of-the-art MM-TTA method [29] in consecutive frames on SemanticKITTI [1], where regions in white boxes exhibit noisy predictions when considering single-frame input. (b) presents the overall framework of Latte, which firstly obtains ST voxels for each modality given the merged input of consecutive frames within each slide window and then computes the modality-specific ST entropy $E^{2\mathrm{D}}$ , $E^{3\mathrm{D}}$ . ST entropy indicates the relative reliability of each voxel in each modality and therefore is utilized in the following adaptive cross-modal attending. + +![](images/a500e2fb9aa47cb9e5e77e60915ad7cb7531a73e950dab13d544e87e16d75802.jpg) + +poorly on data from different domains due to the domain shift [6, 26, 33]. To resolve this problem, previous works have proposed Multi-Modal Unsupervised Domain Adaptation (MM-UDA) approaches [16, 18, 24, 44] that employ multimodal information to transfer knowledge learned from the labeled source domain to the unlabeled target domain without expensive human annotations. + +Nevertheless, the adaptation process of MM-UDA is completely offline, which requires full access to the source domain dataset and multiple training epochs. Despite its effectiveness, offline training is usually infeasible when encountering distribution shifts during the inference stage that require adapting models in real time. Inspired by Test-Time Adaptation (TTA) [22, 23], a recent work [29] proposes the first Multi-Modal Test-Time Adaptation (MM-TTA) method for 3D segmentation. Similar to the settings of TTA, MM-TTA prohibits access to any raw sample from the source domain dataset and adapts the model during the testing stage. This results in a quick adaptation scenario (i.e., one epoch for training only as in [29, 32]) that requires stable optimization in an online manner. To this end, MM-TTA methods must seek a reliable source of supervision signals. + +To fulfill the efficiency and stability of adaptation, the current state-of-the-art MM-TTA method [29] evaluates modal reliability by the prediction consistency between different experts in a frame-wise manner. However, due to the domain shift between the online frame and the pre-trained source dataset, the single-frame prediction is usually unstable. Existing approaches rely on pseudo labels to refine the single-frame prediction but still fail in some cases. As shown in Fig. 1a, the single-frame refinement strategy [29] suffers from noisy single-frame predictions on the car in the white rectangle across time. Such inconsistent predictions across time degenerate the segmentation performance and aggravate denoising stress in downstream tasks (e.g., semantic-based retrivial [46] and obstacle recognition). On one hand, these temporally unstable single-frame predictions could be wrongly regarded as reliable ones, propagating prediction noise to the other modality and causing error accumulation or even catastrophic forgetting [22,23]. + +On the other hand, while previous methods [4, 41] alleviate this instability by utilizing the average predictions of multiple augmented frames instead, they are computationally expensive for online adaptation since the inference time grows linearly with the increasing number of augmented frames. + +This work aims to efficiently suppress the single-frame instability for MM-TTA by exploiting the correlation between multi-frame correspondences. Specifically, we propose that 3D space can be efficiently divided into multiple voxels [33], each of which contains points and point-corresponding pixels located at the same geometric region. While captured at different timestamps, it can be observed that points or pixels located at the same voxel can be viewed as different observations on the same semantic objects [28]. Reliable predictions should therefore be certain and consistent within their corresponding voxels across time, whether we evaluate them from a temporally global perspective (i.e., all input frames) or a temporally local perspective (i.e., certain consecutive frames). By aggregating the predictions in a voxel-wise manner and re-evaluating their reliability, the single-frame prediction noise as in Fig. 1a can be effectively alleviated. + +Motivated by the above observation, we propose a novel MM-TTA method called ReLiable Spatial-temporal Voxels (Latte), which leverages the spatial-temporal correspondences within consecutive frames in a cross-modal manner as in Fig. 1b. Specifically, given frames of point clouds with their estimated poses, we regard points in the same voxel extracted from the merged point cloud frames as spatial-temporal correspondences. Different from previous works that merge all input frames [5,7] or regularize frame-to-frame consistency [28], Latte aggregates consecutive frames in a sliding window manner to estimate the temporally local prediction consistency through the proposed Spatial-Temporal voxels (ST voxels) and entropy (ST entropy). Based on the reliability estimated by ST entropy, we conduct cross-modal learning in an adaptive attending manner to reduce the contributions of predictions from the noisy modality. To better facilitate the progress of MM-TTA, we evaluate our Latte with more recent backbones (SegFormer [43] and SPVCNN [33]) on three different benchmarks, where Latte consistently outperforms previous TTA or MM-TTA methods. + +In summary, our contributions are three-fold. Firstly, we propose Latte, a novel MM-TTA method for multi-modal 3D segmentation, which is the first work that incorporates spatial-temporal correlations for MM-TTA segmentation to our best knowledge. Secondly, Latte efficiently extracts spatial-temporal correspondence through ST voxels and estimates their reliability by ST entropy, which is further incorporated to enhance adaptive cross-modal attending. Thirdly, our experimental results show that Latte outperforms previous state-of-the-art TTA and MM-TTA methods on three different benchmarks. + +# 2 Related Works + +Test-Time Adaptation (TTA). TTA is proposed to mitigate the domain shift between the training data and the testing data in an online manner, which is firstly proposed by TENT [39]. Due to its practical and challenging setting, + +TTA is attracting more and more attention. Previously proposed TTA methods attempt to address this task from different perspectives, such as entropy minimization [22, 23, 39], self-training with pseudo-labels [12, 40], and augmentation invariance [41, 48]. In addition to regularizing predictions, some methods propose to adapt networks from the feature level instead [21, 32], while more recent methods begin to consider different variants of TTA scenarios, such as continual TTA [4, 22, 41], non-i.i.d TTA [11], or the mix of aforementioned cases [47]. Different from source-free adaptation [19, 30] which also discards the access to the source domain yet requires multiple training epochs, the adaptation process of TTA is completely online, following the "one-pass" protocol as in [32]. While most existing TTA methods are designed for image classification or image segmentation without explicit temporal correlation, we argue temporal information in 3D segmentation can be effectively leveraged for TTA as Latte in this work. + +Multi-modal domain adaptation for 3D segmentation. To avoid expensive annotation costs and overcome the poor generalizability of fully supervised solutions, various multi-modal domain adaptation methods have been investigated. Specifically for MM-UDA, xMUDA [16] is the primary works that explore incorporating cross-modal learning with MM-UDA. It regards cross-modal prediction consistency and pseudo-labels as its supervision signals in the unlabeled target domain. Most subsequent works propose solutions to address its limitations from different perspectives, such as more diverse point-pixel correspondence [24, 44], procedures to mitigate domain gaps [18, 20], and alleviating class-imbalanced problem [3]. Since MM-UDA is infeasible when facing online domain shifts, some previous works [4, 29] develop different methods to address the domain gap in an online manner, including cross-modal pseudo-label generation [29] and adaptive modality attending [4]. A recent work [30] proposes a source-free domain adaptation method by estimating cross-modal prediction agreement, yet it requires multiple offline training epochs. Different from previous multi-modal domain adaptation methods that ignore temporal relationships, Latte couples spatial-temporal correspondence to mitigate the single-frame instability and achieve better online adaptation performance. + +Temporal processing of point clouds. Since point clouds are naturally in the form of temporally consecutive frames, previous methods have widely explored how to leverage such temporal information for different tasks. Specifically, scene flow [15,35,37,38] contains the movement of every point in the 3D world. Considering such point-wise computation expensive, previous works on different tasks (e.g., segmentation [5,6,28] and object detection [25]) usually interpret the temporal information from a simpler perspective. In terms of fully supervised learning, Choy et al. [6] propose the first 4D segmentation method by generalized sparse convolution kernels [13] and Fan et al. [8] turn to the fully connected learning scheme as in [26], while Xu et al. [45] extract long-term temporal information by an efficient memory bank. On the other hand, considering scenarios with insufficient or zero annotations, some works propose different strategies to couple spatial-temporal information without ground-truth labels. For instance, Chen et al. [5] proposes to guide voxelized spatial-temporal representation with + +![](images/84f048bd8d47220c0c1f832940d02649d5471164a1790f6cdf6db3e16604bd55.jpg) +Fig. 2: Overall structure of Latte. Taking a student prediction frame of one modality as the query input, our slide-window extraction searches its spatial-temporal correspondences through voxelization within a time window to establish the temporally local prediction consistency. ST voxels are then generated, where those with high ST entropy (larger than $\alpha$ -quantile) are discarded as unreliable correspondences, while the others are leveraged for adaptive cross-modal learning by attending to the modality with lower ST entropy in a voxel-wise manner. + +strong text embedding from CLIP [27] instead of labels in a multi-modal manner, while Saitor et al. [28] interpret such representation by nearest neighbor search in a frame-to-frame manner. In this work, we design a slide window aggregation and voxelization rather than aggregating all frames as in [5], which yields a better indication of the prediction reliability for each point. + +# 3 Methodology + +Problem definition. During MM-TTA for 3D segmentation, both 2D RGB images and 3D point clouds are captured consecutively in the target domain $\mathcal{T}$ , denoted as $\mathbf{x}_{\mathcal{T},t}^{2\mathrm{D}} \in \mathbb{R}^{3 \times H \times W}$ and $\mathbf{x}_{\mathcal{T},t}^{3\mathrm{D}} \in \mathbb{R}^{N \times 4}$ , respectively, where $t$ represents the frame order and $H, W$ are the height and width of input images. The point-wise input contains the 3D coordinate $\{x, y, z\}$ and the point feature (e.g., intensity, reflectivity, etc.). Modality-specific networks are pre-trained on the labeled source domain before adaptation, denoted as $\phi_{\mathcal{T},t}^{\mathrm{m}}(\cdot)$ with $\mathrm{m} \in \{2\mathrm{D}, 3\mathrm{D}\}$ . During MM-TTA, both networks are initialized from the parameters pre-trained on the source domain. Following previous works [16,29], 3D points are projected to the 2D image based on the relative projection matrix to obtain cross-modal correspondences within each frame, resulting in the modality-specific predictions $\mathbf{p}_{\mathcal{T},t}^{\mathrm{m}} = \phi_{\mathcal{T},t}^{\mathrm{m}}(\mathbf{x}_{\mathcal{T},t}^{\mathrm{m}})$ , $\mathbf{p}_{\mathcal{T},t}^{\mathrm{m}} \in \mathbb{R}^{N \times K}$ where $K$ denotes the number of semantic classes. Data from the labeled source domain are strictly inaccessible and we omit the target domain subscript $\mathcal{T}$ by default. + +The multi-modal input of online 3D segmentation is temporally consecutive, which exhibits strong temporal relationships between frames. The geometric neighborhoods in consecutive frames can be viewed as the naturally various observations on the same semantic object that can be leveraged to stabilize noisy single-frame predictions. Motivated by this observation, we develop Latte, + +which discovers reliable spatial-temporal (ST) correspondences to improve both optimization stability and prediction consistency during MM-TTA. As shown in Fig. 2, given consecutive frames of 2D images and 3D point clouds as input, Latte firstly extracts their frame-wise predictions from both student and teacher networks (Sec. 3.1). These frame-wise predictions are then passed to the slide-window extraction and voxelization (Sec. 3.2). ST voxels and ST entropy are subsequently computed indicating the prediction consistency and certainty (Sec. 3.3). The final ST entropy is leveraged for cross-modal attending to alleviate the noise from unreliable modality-specific predictions (Sec. 3.4). + +# 3.1 Frame-wise Predictions from Students and Teachers + +Before exploiting spatial-temporal correspondence, we first generate the framewise predictions from each modality. Motivated by the fact that moving average models can provide more stable online predictions [4,34,41], Latte utilizes weight-averaged teacher models and fast student models for each modality as in previous works [29, 41], denoted as $\tilde{\phi}_t^{\mathrm{m}}(\cdot)$ and $\phi_t^{\mathrm{m}}(\cdot)$ , respectively. Given each input frame $\mathbf{x}_t^{\mathrm{m}}$ , the student predictions $\mathbf{p}_t^{\mathrm{m}}$ and teacher predictions $\tilde{\mathbf{p}}_t^{\mathrm{m}}$ are computed as: + +$$ +\tilde {\mathbf {p}} _ {t} ^ {\mathrm {m}} = \tilde {\phi} _ {t} ^ {\mathrm {m}} \left(\mathbf {x} _ {t} ^ {\mathrm {m}}\right), \mathbf {p} _ {t} ^ {\mathrm {m}} = \phi_ {t} ^ {\mathrm {m}} \left(\mathbf {x} _ {t} ^ {\mathrm {m}}\right), \tag {1} +$$ + +where the teacher model $\tilde{\phi}_t^{\mathrm{m}}$ is initialized from the source domain pre-trained models. In the following process, the teacher's predictions of one modality are regarded as the cross-modal guidance of student predictions from the other modality. In practice, the gradient propagation of teacher model predictions is discarded to prevent directly updating teacher models. + +# 3.2 Slide Window Frame Aggregation and Voxelization + +Our goal is to efficiently leverage spatial-temporal correspondences between predictions and seek reliability estimation through prediction consistency within correspondences. Establishing spatial-temporal correspondences in 3D space has been explored in other deep learning tasks, where they mainly consider temporally global relationships of all input frames [5,28] or frame-to-frame correspondences [28]. However, they both have some limitations in terms of consistency and certainty evaluation: the former can not highlight the risk of some inconsistent predictions within a short time window (i.e., temporally local inconsistency), while the latter suffers from an insufficient number of correspondences. + +To effectively empower spatial-temporal information with Latte, we propose a slide window frame aggregation that focuses on temporally local correspondences that lie within a time window to evaluate temporally local consistency. Specifically, given a student prediction frame $\mathbf{p}_i^{\mathrm{m}}$ with the temporal index $i$ as the query, the correspondence search is conducted within a time window of consecutive frames denoted as $\{j||j - i|\leq w_t\}$ , where $w_{t}$ is the pre-defined time window size. The merged point cloud $\hat{\mathbf{x}}_i^{3\mathrm{D}}\in \mathbb{R}^{\hat{N}\times 3}$ (point features are omitted for simplicity) for correspondence search is then formulated as: + +$$ +\hat {\mathbf {x}} _ {i} ^ {\mathrm {3 D}} = \operatorname {c a t} \left( \right.\left\{ \right.\mathbf {T} _ {j \rightarrow i} * \mathbf {x} _ {j} ^ {\mathrm {3 D}} \left. \right| | j - i | \leq w _ {t} \}\left. \right), \quad \mathbf {T} _ {j \rightarrow i} = \mathbf {T} _ {i} ^ {- 1} \mathbf {T} _ {j}, \tag {2} +$$ + +where $\mathbf{T}_i$ is the estimated pose at frame $i$ and so as $\mathbf{T}_j$ , which can be easily obtained online through off-the-shelf SLAM algorithms [17,36]. $\operatorname{cat}(\cdot)$ is the concatenation operation along the point number dimension while $*$ denotes the pose transformation. The voxelization is then performed on $\hat{\mathbf{x}}_i^{3D}$ , formulated as: + +$$ +\mathbf {v} _ {i} ^ {\mathrm {3 D}} = \mathcal {V} _ {\mathbf {s}} \left(\hat {\mathbf {x}} _ {i} ^ {\mathrm {3 D}}\right), \quad \mathbf {v} _ {i} ^ {\mathrm {3 D}} \in \mathbb {R} ^ {N _ {\mathrm {v}} \times 3}, \tag {3} +$$ + +$$ +\mathbf {M} _ {i} ^ {\mathrm {3 D}} = \left\{k | \forall g \in [ 1, \hat {N} ], \left\lfloor \hat {\mathbf {x}} _ {i, g} ^ {\mathrm {3 D}} / \mathbf {s} \right\rfloor = \mathbf {v} _ {i, k} ^ {\mathrm {3 D}} \right\}, \quad \mathbf {M} _ {i} ^ {\mathrm {3 D}} \in \mathbb {R} ^ {\hat {N}}, \tag {4} +$$ + +where $\mathbf{v}_i^{3\mathrm{D}}$ is the extracted voxels and $\nu_{s}$ is the voxelization operation with a voxel size s. After voxelization, points located in the same voxel can be regarded as correspondences since they are geometric neighborhoods, where the hash table $\mathbf{M}_i^{3\mathrm{D}}$ that maps each voxel back to its containing points (e.g., $\mathbf{v}_{i,k}^{3\mathrm{D}}$ to $\hat{\mathbf{x}}_{i,g}^{3\mathrm{D}}$ as in Eq. (4)) is preserved for the subsequent correspondence mapping. In this way, it provides more diverse neighborhood selection strategies with temporally local correspondences for consistency and certainty evaluation compared to globally merging all frames in one single step. In practice, the slide window aggregation and voxelization are conducted iteratively for each frame in the input batch. + +# 3.3 Spatial-Temporal Voxels and Entropy + +Given the geometric correspondences captured by voxelization, our goal is to emphasize the predictions with correspondence consistency within each voxel. After slide window aggregation and voxelization, there inevitably exist some unreliable correspondences of two different types: (i) voxels that contain different types of semantic objects (e.g., those located on the contact surface between cars and roads) and (ii) voxels with highly inconsistent predictions due to uncertain predictions across time. These unreliable correspondences contain high uncertainty and inconsistency, which could harm the online optimization of models if they as references. To alleviate the side-effect introduced by such unreliable correspondences, we propose ST voxels that leverage ST entropy for voxel-wise reliability evaluation and cross-modal attending. + +Specifically, an ST voxel contains two components, including the query and the reference, where the query is encouraged to be consistent with the reference. Considering that the moving average teacher can usually generate more stable predictions [4, 41], we regard the single-frame student predictions as the query while multi-frame teacher predictions as our reference, forming ST voxels denoted as $\mathbf{v}_i^{\mathrm{ST}}$ . Without losing generality, here we take a single ST voxel indexed by $k$ in frame $i$ as an example, denoted as $\mathbf{v}_{i,k}^{\mathrm{ST}} = \{\mathbf{p}_{\mathrm{q}}^{\mathrm{m}},\mathbf{p}_{\mathrm{r}}^{\mathrm{m}}\}$ , where $\mathbf{p}_{\mathrm{q}}^{\mathrm{m}}$ and $\mathbf{p}_{\mathrm{r}}^{\mathrm{m}}$ are the point-wise student predictions at the query frame $i$ and teacher predictions in the frame searching range $j$ , respectively, formulated as: + +$$ +\mathbf {p} _ {\mathrm {q}} ^ {\mathrm {m}} = \left\{\mathbf {p} _ {i, g} ^ {\mathrm {m}} \mid \left\lfloor \mathbf {x} _ {i, g} ^ {\mathrm {3 D}} / \mathbf {s} \right\rfloor = \mathbf {v} _ {i, k} ^ {\mathrm {3 D}} \right\}, \quad \mathbf {p} _ {\mathrm {q}} ^ {\mathrm {m}} \in \mathbb {R} ^ {N _ {\mathrm {q}} \times K} \tag {5} +$$ + +$$ +\mathbf {p} _ {\mathrm {r}} ^ {\mathrm {m}} = \left\{\tilde {\mathbf {p}} _ {j, g} ^ {\mathrm {m}} \mid \left| \mathbf {x} _ {j, g} ^ {3 \mathrm {D}} / \mathbf {s} \right| = \mathbf {v} _ {i, k} ^ {3 \mathrm {D}}, | j - i | \leq w _ {t} \right\}, \quad \mathbf {p} _ {\mathrm {r}} ^ {\mathrm {m}} \in \mathbb {R} ^ {N _ {\mathrm {r}} \times K}. \tag {6} +$$ + +The reliability of each voxel is then evaluated as the Shannon's entropy [42] of the average teacher predictions, where the unreliable ST voxels are then filtered + +out given a pre-defined quantile $\alpha$ as in Fig. 2, denoted as: + +$$ +E _ {i, k} ^ {\mathrm {m}} = - \sum_ {c} ^ {K} \bar {\mathbf {p}} _ {\mathrm {r}, c} ^ {\mathrm {m}} \log \bar {\mathbf {p}} _ {\mathrm {r}, c} ^ {\mathrm {m}}, \quad \bar {\mathbf {p}} _ {\mathrm {r}} ^ {\mathrm {m}} = \sum_ {n = 1} ^ {N _ {\mathrm {r}}} \psi \left(\mathbf {p} _ {\mathrm {r}, n} ^ {\mathrm {m}}\right) / N _ {\mathrm {r}}, \tag {7} +$$ + +$$ +h _ {i, k} = \left\{ \begin{array}{l l} 0, & \text {i f} E _ {i, k} ^ {m} > Q ^ {m} (\alpha) \\ 1, & \text {i f} E _ {i, k} ^ {m} \leq Q ^ {m} (\alpha) \end{array} , \right. \tag {8} +$$ + +where $E_{i,k}^{\mathrm{m}}$ is regarded as the ST entropy of voxel $k$ for modality $\mathrm{m}$ , indicating the modal-specific prediction reliability within this voxel. $Q^{m}(\alpha)$ is the $\alpha$ -quantile of all ST entropy in modality $\mathrm{m}$ in the same batch and $\psi(\cdot)$ denotes the Softmax function along the class dimension. + +# 3.4 ST Voxel Aided Cross-Modal Learning + +Each modality has its pros and cons under different conditions and cross-modal learning for online scenarios like MM-TTA should therefore possess an adaptive mechanism that can attend to "pros" and meanwhile suppress the "cons". To achieve that, Latte performs cross-modal learning by attending to the modality with more consistent and certain predictions from a spatial-temporal perspective. Specifically, taking ST voxel $\mathbf{v}_{i,k}^{\mathrm{ST}}$ and its corresponding ST entropy $E_{i,k}^{\mathrm{m}}$ as an example, the voxel-wise cross-modal attending weights and the weighted cross-modal consistency loss $\mathcal{L}_{i,k}^{\mathrm{xM}}$ are formulated as: + +$$ +w _ {\mathrm {v}} ^ {\mathrm {2 D}} = \frac {\exp \left(E _ {i , k} ^ {\mathrm {2 D}}\right)}{\exp \left(E _ {i , k} ^ {\mathrm {2 D}}\right) + \exp \left(E _ {i , k} ^ {\mathrm {3 D}}\right)}, w _ {\mathrm {v}} ^ {\mathrm {3 D}} = 1 - w _ {\mathrm {v}} ^ {\mathrm {2 D}}, \tag {9} +$$ + +$$ +\mathcal {L} _ {i, k} ^ {\mathrm {x M}} = w _ {\mathrm {v}} ^ {\mathrm {2 D}} D _ {\mathrm {K L}} \left(\bar {\mathbf {p}} _ {\mathrm {q}} ^ {\mathrm {3 D}} \| \bar {\mathbf {p}} _ {\mathrm {r}} ^ {\mathrm {2 D}}\right) + w _ {\mathrm {v}} ^ {\mathrm {3 D}} D _ {\mathrm {K L}} \left(\bar {\mathbf {p}} _ {\mathrm {q}} ^ {\mathrm {2 D}} \| \bar {\mathbf {p}} _ {\mathrm {r}} ^ {\mathrm {3 D}}\right), \tag {10} +$$ + +where $D_{\mathrm{KL}}(\cdot)$ represents the KL-divergence between two probability. $\bar{\mathbf{p}}_{\mathrm{q}}^{3\mathrm{D}}$ is the average query predictions in the ST voxel, similarly computed as in Eq. (7). + +The cross-modal attending is further extended from the voxel level to the point level for better online predictions and cross-modal pseudo-label generation. Specifically, the ST entropy is propagated from the ST voxel to its reference point-wise prediction, indicating its confidence and consistency within its spatial-temporal neighborhoods. If one's ST entropy has been filtered out as in Eq. (8), its ST entropy would fall back to the point-level entropy. For arbitrary pointwise predictions $\mathbf{p}^{\mathrm{m}}\in \mathbf{p}_{\mathrm{r}}^{\mathrm{m}}$ , the cross-modal predictions $y^{\mathrm{xM}}$ are formulated as: + +$$ +\hat {E} _ {\mathrm {r}} ^ {\mathrm {m}} = h _ {i, k} E _ {i, k} ^ {\mathrm {m}} - \left(1 - h _ {i, k}\right) \sum_ {c} ^ {K} \mathbf {p} _ {c} ^ {\mathrm {m}} \log \mathbf {p} _ {c} ^ {\mathrm {m}}, \tag {11} +$$ + +$$ +w _ {\mathrm {p}} ^ {\mathrm {2 D}} = \frac {\exp \left(\hat {E} ^ {\mathrm {2 D}}\right)}{\exp \left(\hat {E} ^ {\mathrm {2 D}}\right) + \exp \left(\hat {E} ^ {\mathrm {3 D}}\right)}, w _ {\mathrm {p}} ^ {\mathrm {3 D}} = 1 - w _ {\mathrm {p}} ^ {\mathrm {2 D}}, \tag {12} +$$ + +$$ +\mathbf {p} ^ {\mathrm {x M}} = w _ {\mathrm {p}} ^ {\mathrm {2 D}} \mathbf {p} ^ {\mathrm {2 D}} + w _ {\mathrm {p}} ^ {\mathrm {3 D}} \mathbf {p} ^ {\mathrm {3 D}}, y ^ {\mathrm {x M}} = \underset {c} {\arg \max } \mathbf {p} ^ {\mathrm {x M}}. \tag {13} +$$ + +Algorithm 1: Adaptation and Online Prediction Process of Latte +Target Domain: Multi-modal input $\mathcal{X}_{\mathcal{T}} = \{(\mathbf{x}_{t}^{2\mathrm{D}},\mathbf{x}_{t}^{3\mathrm{D}})\}$ +Init Model: Teacher $\tilde{\phi}_t^{\mathrm{m}}$ , student $\phi_t^{\mathrm{m}}$ $\mathfrak{m}\in \{2\mathrm{D},3\mathrm{D}\}$ , Slide window size $w_{t}$ +for $\pmb {x} = \{(x_{t}^{2D},x_{t}^{3D})\}_{t = t_0 + 1}^B\in \mathcal{X}_T$ do 1. Get student and teacher predictions $\tilde{\mathbf{p}}_t^{\mathrm{m}},\mathbf{p}_t^{\mathrm{m}}$ by Eq. (1) for $i\in [t_0 + 1,t_0 + B]$ do 2. Get $\hat{\mathbf{x}}_i^{3\mathrm{D}},\mathbf{v}_i^{3\mathrm{D}},\mathbf{M}_i^{3\mathrm{D}}$ within the time window of size $w_{t}$ by Eq. (2)-(4) 3. Compute ST voxels $\mathbf{v}_i^{\mathrm{ST}}$ and ST entropy $E_i^{\mathrm{m}}$ by Eq. (5)-(8) 4. Compute $\mathcal{L}_i^{\mathrm{xM}}$ by Eq. (9)-(10) end 5. Compute $\hat{E}_{\mathrm{r}}^{\mathrm{m}}$ and get $\mathbf{y}_t^{\mathrm{xM}}$ as predictions by Eq. (11)-(13) 6. Compute overall loss by Eq. (14) and update models by Eq. (15) +end + +# 3.5 Online Predictions and Optimization + +As shown in Fig. 2, given a batch of $B$ consecutive frames $\{\mathbf{x}_t^{\mathrm{m}}\}_{t = t_0 + 1}^{t_0 + B}$ as input, we first extract the student and teacher predictions through Eq. (1). Subsequently, frames are aggregated and voxelized as in Eq. (2), resulting in a batch of merged point cloud $\{\hat{\mathbf{x}}_t^{3\mathrm{D}}\}_{t = t_0 + 1}^{t_0 + B}$ and voxels $\{\mathbf{v}_t^{3\mathrm{D}}\}_{t = t_0 + 1}^{t_0 + B}$ . The ST voxels and ST entropy computation process is then applied to each frame of the merged point cloud and voxels. Due to the overlap of the slide window, multiple ST entropy values could be propagated to the same point prediction, where we empirically take the average value of all ST entropy received to proceed Eq. (11). The overall loss function and the updating scheme can be formulated as: + +$$ +\mathcal {L} = \sum_ {t} \mathcal {F} \left(\mathbf {p} _ {t} ^ {\mathrm {m}}, \mathbf {y} _ {t} ^ {\mathrm {x M}}\right) + \frac {\lambda_ {\mathrm {x M}}}{B} \sum_ {t} \sum_ {k} \mathcal {L} _ {t, k} ^ {\mathrm {x M}}, \tag {14} +$$ + +$$ +\tilde {\theta} _ {t} ^ {\mathrm {m}} = \lambda_ {s} \tilde {\theta} _ {t - 1} ^ {\mathrm {m}} + (1 - \lambda_ {s}) \theta_ {t} ^ {\mathrm {m}}, \tag {15} +$$ + +where $\mathcal{F}(\cdot)$ is the cross-entropy function and $\mathbf{y}_t^{\mathrm{xM}}$ is the point-wise cross-modal pseudo-label frame from Eq. (13), which is also regarded as our cross-modal prediction for evaluation. $\tilde{\theta}_t^{\mathrm{m}}$ and $\theta_t^{\mathrm{m}}$ are the model parameters of teacher and student models, respectively, where $\lambda_s$ is the momentum update coefficient. $\lambda_{\mathrm{xM}}$ is pre-defined coefficient for the cross-modal consistency loss. Our adaptation and inference process is summarized in Algo. 1 + +# 4 Experimental Results + +In this section, we present our thorough experimental results on three different benchmarks. Details of benchmarks, backbones, and settings are first present in Sec. 4.1. The main results are then illustrated in Sec. 4.2, followed by detailed ablation studies and qualitative results in Sec. 4.3. + +# 4.1 Benchmarks and settings + +Benchmark Details. To thoroughly investigate the effectiveness of Latte, we conduct our experiments on three different MM-TTA benchmarks: (i) USA-to-Singapore (U-to-S), (ii) A2D2-to-SemanticKITTI (A-to-S), and (iii) Synthia-to-SemanticKITTI (S-to-S). Specifically, U-to-S is developed from NuScenes-LiDARSeg [2] dataset, where the domain gap is mainly credited to the infrastructure difference between countries. Same as [16], the source domain and the target domain data are selected by filtering country keywords based on the data recording description in NuScenes. Different from previous UDA methods [16,24] utilizing self-defined class mappings, we follow the official mapping in NuScenes-LiDARSeg, forming a 16-class segmentation benchmark. The remaining two benchmarks follow a similar setting as in [29], where the domain gap of A-to-S [1,10] lies in different LiDAR mounting positions and image resolutions while the one of S-to-S [1] lies in different patterns between synthetic and real data for both modalities. For A-to-S, we adopt the same class-mapping as in [16,29] which leads to a 10-class benchmark, while we design our class-mapping for S-to-S since its original class map has not been revealed previously [29], forming a 9-class benchmark. More details are presented in our appendix. + +Baseline methods. Previous TTA and MM-TTA methods are included in comparison with Latte. For TTA methods, we compare Latte with point-wise pseudolabels with filtering (PsLabel), TENT [39], ETA [22], SAR [23] with and without restoring (denoted as SAR-rs and SAR, respectively). MMTTA [29] as well as its online version of xMUDA [16] (including pure xMUDA and pseudo-label version xMUDA+PL) are additionally included as the previous SOTA methods for comparison. Considering that EATA [22] with Fisher regularizer requires pre-access to the target domain samples, we compare Latte with ETA, an official variant of EATA without this regularizer for a fair comparison. We further present a performance lower bound by directly testing with source pre-trained models (i.e., Source only) and an upper bound by online adapting networks with ground-truth labels from the target domain for one epoch (i.e., Oracle TTA). + +Implementation details. With the rapid progress in 2D and 3D segmentation, we employ all baseline methods and Latte with more recent SegFormer [43] (SegFormer-B1) and SPVCNN [33] (SPVCNN-cr1) as their 2D and 3D backbones, respectively, to investigate the improvement of existing TTA methods when combined with advanced networks. The pre-training procedures on the source domain follow a similar setting as in [16], except for 2D SegFormer which is trained with a base learning rate of 6E-5 with an AdamW optimizer and Poly scheduler as in [43]. For Latte, we maintain the same learning rate and optimizer as in the pre-training stage, while the scheduler is disabled. For our voxelization, we leverage the off-the-shelf SLAM algorithm KISS-ICP [36] to generate poses and utilize a window size $w_{t}$ of 3 with a voxel size of $0.2\mathrm{m}$ , while the quantile $\alpha$ and coefficient $\lambda_{\mathrm{xM}}$ are set to 0.9 and 0.3, respectively. $\lambda_{s}$ is empirically set to 0.99 as in [4,31]. For all methods, we strictly follow the one-pass protocol to first evaluate the predictions of the input batch and then update the networks, where we only update the trainable parameters in normalization layers. For all base + +Table 1: Performance (mIoU) comparison of Latte. Latte outperforms all previous SOTA methods on the cross-modal metric (xM) across three benchmarks. Here "MM" denotes whether the method contains multi-modal interaction or not. Cross-modal predictions are computed as the Softmax average of modal outputs except for Latte. "Avg" present the average xM performance across three benchmarks. + +
MethodPublicationMMU-to-SxMA-to-SxMS-to-SAvg
2D3D2D3D2D3D
Source only-X31.441.143.947.417.944.323.436.448.2
Oracle TTA-X38.745.550.349.161.362.536.054.554.6
TENT [39]ICML-21X36.836.141.143.344.449.122.435.537.5
ETA [22]ICML-22X36.734.743.743.242.549.722.430.632.9
SAR [23]ICLR-23X36.637.243.943.042.750.320.731.332.9
SAR-rs [23]ICLR-23X36.635.543.943.143.850.124.225.128.8
xMUDA [16]PAMI-22✓19.422.924.213.133.032.212.214.814.1
xMUDA+PL [16]PAMI-22✓36.238.243.042.848.350.930.934.136.3
MMTTA [29]CVPR-22✓37.241.545.444.551.753.727.535.135.5
PsLabel-✓37.735.841.846.744.350.029.233.335.3
Latte (Ours)-✓37.441.046.046.152.654.333.239.341.6
+ +lines, we conduct a parameter search and report their best results. We utilize the mean intersection over union (mIoU) as our evaluation metric. All experiments are conducted with PyTorch on a single RTX 3090. More details are presented in our appendix. + +# 4.2 Overall Results + +Tab. 1 presents the overall results of our Latte on three benchmarks compared with previous SOTA methods. In terms of average cross-modal predictions (Avg), Latte surpasses all previous TTA or MM-TTA methods, with a significant relative improvement of more than $5.3\%$ . For the less challenging U-to-S and A-to-S (i.e., gap between Oracle TTA and best TTA less than $10\%$ ), Latte achieves consistent relative improvements of $1.3\%$ on U-to-S and $1.1\%$ on A-to-S compared to MMTTA, respectively. Most existing methods except for xMUDA [16] can achieve superior performance compared to source-only results when one modality is reliable as in U-to-S (3D prevails) and A-to-S (2D prevails). + +In terms of the most challenging S-to-S, Latte exhibits significant performance improvement compared to previous SOTA MM-TTA methods such as MMTTA [29] and xMUDA+PL [16], with a relative gap of $14.6\%$ . In fact, all previous methods except for our Latte perform inferiorly compared to source pre-trained models (Source only), due to the significant domain gaps caused by the non-trivial pattern difference in both modalities between synthetic and real data. In this case, the instability of single-frame predictions becomes exacerbated, leading to performance degeneration in MMTTA which mainly relies on single-frame refinement. This justifies the effectiveness of Latte under either challenging or less challenging TTA scenarios. Another interesting observation is that almost all cross-modal methods outperform other single-modal TTA methods, e.g., even the simplest PsLabel can outperform recent TTA methods like SAR [23]. This reveals the non-trivial benefits of cross-modal learning for TTA with segmentation. More comparisons about efficiency are included in appendix. + +Table 2: Ablation studies about the effectiveness of cross-modal learning. When disabling $\mathbf{y}^{\mathrm{xM}}$ , we utilize the average Softmax of all modalities as the pseudo-labels as a replacement. In experiment No. 8, we replace the ST entropy (ST ety) with point-wise entropy in Eq. (13) and set $w_{\mathrm{v}}^{\mathrm{m}}$ in Eq. (10) to 0.5. Here $\mathcal{L}_{32}^{\mathrm{xM}}$ and $\mathcal{L}_{23}^{\mathrm{xM}}$ refer to the former half and the latter half in Eq. 10, respectively. + +
No.Description\( y^{xM} \)\( \mathcal{L}_{23}^{xM} \)\( \mathcal{L}_{32}^{xM} \)U-to-SA-to-SS-to-S
2D3DxM2D3DxM2D3DxM
0Source only---31.441.143.947.417.944.323.436.438.2
1only ST ety label✓36.834.543.743.342.249.922.430.633.3
2only ST 2D→3D✓32.534.541.041.742.248.227.730.632.9
3only ST 3D→2D✓36.839.642.643.350.248.422.426.725.6
4ST ety label + ST 2D→3D✓✓37.535.841.043.647.549.831.036.038.4
5ST ety label + ST 3D→2D✓✓37.440.945.946.252.153.432.538.739.7
6ST 2D↔3D✓✓36.840.343.844.047.051.230.435.637.7
7w/o percentile in Eq. (8)✓✓✓37.440.345.445.751.453.232.939.041.7
8Point ety vs. ST ety✓✓✓37.040.645.145.151.452.731.636.738.2
9Latte✓✓✓37.441.046.046.152.654.333.239.341.6
+ +![](images/2e91f6bc715966bf0af65128b1ec32013bf4e3b4520d2bf8a2f5912c3115700c.jpg) +Fig. 3: Ablation studies of different frame aggregation mechanisms. Besides (a) Latte's slide window aggregation, we test two different aggregation methods, including (b) replacing single frame student predictions with multiple frames in Eq. (5) and (c) adding non-overlapping aggregation to (b). Results show that our slide window aggregation can better evaluate the local consistency and results in consistent improvement. + +# 4.3 Ablation Studies and Visualization + +In this section, we investigate the effectiveness of each component in Latte by thorough ablation studies and qualitative results. + +Effectiveness of cross-modal learning. The main optimization objectives of Latte contain three parts, including ST-entropy-based cross-modal consistency (both 2D-to-3D and 3D-to-2D) and cross-entropy loss between predictions and cross-modal pseudo-labels. We justify the effectiveness by using different combinations of these components as in Tab. 2. As shown from experiments No. 1 to No. 6, disabling any part of Latte causes performance decreases of more than $1.6\%$ relatively on the cross-modal prediction (xM), which proves the effectiveness of each component we included. A special case is No. 5 on U-to-S, where disabling ST $\mathcal{L}_{23}^{\mathrm{xM}}$ only leads to a minor $0.1\%$ drop in accuracy. This is mainly because the 2D performance is already close to saturation (37.4 vs. 38.7 of Oracle TTA) without the guidance of 3D predictions. In this case, introducing 3D information to 2D is less effective compared to the other benchmarks. + +Besides optimization objects, we further testify two important components of Latte, including ST-entropy-based weighting in Eq. (10) and Eq. (13) as well as percentile filtering in Eq. (8). Specifically, disabling percentile filtering as in No. + +![](images/6b7836395acd46ec5d09642e0d8b6e842ddff0d8b006624ac84dfe69592869ed.jpg) +(a) Ablation studies of various ST voxel sizes. + +![](images/2755e97e95cc69d4cd070c321af5a2940b87b922677f4bcd78ebc54a8359a136.jpg) + +![](images/7c9ec6155f3a12c5f3c433b3d33768baae76b4c4c4c957fe958691ec0172403d.jpg) + +![](images/3e4fc3aefee68e3c50682ba0247c3960437614d62b6cb5aecdffd28f29d298db.jpg) +(b) Ablation studies of filtering percentile $\alpha$ in Eq. (8). + +![](images/8216ef83bcc00af7ce9fc4349c5ebfc892298cd3257daca86fce94f832a9c0cf.jpg) +Fig. 4: Sensitivity analysis about the ST voxel size and confident percentile $\alpha$ . + +![](images/b81f9446c37d96988e78a39195bf51e0b51e01109ba2865ae713e45acb63b0f3.jpg) + +7 leads to a relative accuracy drop of more than $0.5\%$ across all benchmarks while utilizing point-wise entropy instead of our ST entropy as in No. 8 degenerates the performance by more than $2.0\%$ relatively, which justifies that ST entropy can better estimate the prediction reliability compared to the point-wise entropy. + +Different frame aggregation mechanisms. One of the core designs of Latte is the slide window frame aggregation, which is designed to capture temporally local correspondences and estimate their prediction consistency. To justify the aforementioned claim, we compare our slide window aggregation with two aggregation variants as shown in Fig. 3. Specifically, variant (b) regards all student prediction frames rather a single frame within the time window as the query in Eq. (5), while variant (c) further utilizes a non-overlapping sliding strategy. According to Fig. 3, our slide window aggregation surpasses the other two aggregation variants under all sizes of time windows across three benchmarks, which justifies the effectiveness of our method. The inferior performance of variants (b) and (c) could be due to two factors. Firstly, aggregating student predictions could be an inferior option since student predictions tend to be less stable, and aggregating multiple frames can therefore lead to noisy queries for cross-modal learning. Secondly, non-overlapping aggregation can result in a miss-checking area on the boundary of each time window, which leads to unrepresentative entropy estimation for some ST voxels. Another observation is that utilizing a smaller slide window can usually yield better performance since temporally local consistency can be captured. This justifies the effectiveness of estimating prediction reliability through temporally local consistency. + +Parameter sensitivity analysis. To testify whether Latte is sensitive towards hyper-parameter settings, we conduct sensitivity analysis on Latte across three benchmarks. Here we illustrate the sensitivity analysis of two hyper-parameters of Latte, including ST voxel size and filtering percentile $\alpha$ as shown in Fig. 4. In terms of ST voxel sizes, one can tell from Fig. 4a that all benchmarks share the same optimal voxel size as 0.2, where performance drops as the voxel size + +![](images/e7f6804bd9099bb2c160d639ef555a80d2aa6b99dcd01d54f1c9587948f9e2da.jpg) +Fig. 5: Qualitative results of Latte on S-to-S. Here we visualize modality-specific predictions and cross-modal predictions from Latte during two sets of consecutive frames. + +becomes too large or too small. This is mainly because large voxels could cause boundary ambiguity while small voxels contain insufficient representative points for evaluation. As for filtering percentile $\alpha$ , non-trivial improvement can be observed on U-to-S and A-to-S when $\alpha \geq 0.8$ , while a lower $\alpha$ begins to discard more confident ST voxels leading to a slight drop in accuracy. On S-to-S, Latte without filtering (i.e., No. 7 in Tab. 2) performs slightly better on cross-modal predictions with a minor gap of $0.1\%$ , while it can also be observed that setting $\alpha$ as 0.9 leads to an improvement of $0.3\%$ on both modality-specific predictions, which justifies the overall effectiveness of our filtering procedures. + +Qualitative results. Our main motivation is utilizing prediction consistency of spatial-temporal correspondences to estimate modality reliability and achieve better cross-modal attending. As shown in Fig. 5, Latte can effectively attend the modality with more consistent predictions in spatial-temporal correspondences and therefore improve the cross-modal prediction consistency across time. For instance, Latte successfully attends to the stable 2D predictions pedestrian in the upper set of consecutive frames, suppressing the spatial-temporal inconsistency from 3D predictions. Similar observations can also be found in the lower sample, where the more consistent 2D predictions of buildings in the red rectangle and 3D predictions of roads in the blue rectangle prevail. More qualitative comparisons with previous SOTA methods are presented in our appendix. + +# 5 Conclusion + +In this paper, we propose a novel MM-TTA method called Latte, which leverages prediction consistency of spatial-temporal correspondences in MM-TTA for 3D segmentation. Latte utilizes a slide window frame aggregation to extract ST voxels and estimates temporally local consistency by the ST entropy of each modality. By attending to the modality with more consistency, Latte can achieve stable improvement on challenging MM-TTA scenarios. Despite its effectiveness, Latte struggles to rectify consistently wrong predictions across time, which is challenging and worth further investigation in the future. + +# Acknowledgment + +This research is supported by the National Research Foundation, Singapore, under the NRF Medium Sized Centre for Advanced Robotics Technology Innovation (CARTIN). Any opinions, findings and conclusions or recommendations expressed in this material are those of the author(s) and do not reflect the views of National Research Foundation, Singapore. + +# References + +1. Behley, J., Garbade, M., Milioto, A., Quenzel, J., Behnke, S., Stachniss, C., Gall, J.: Semantickitti: A dataset for semantic scene understanding of lidar sequences. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 9297-9307 (2019) +2. Caesar, H., Bankiti, V., Lang, A.H., Vora, S., Liong, V.E., Xu, Q., Krishnan, A., Pan, Y., Baldan, G., Beijbom, O.: nuscenes: A multimodal dataset for autonomous driving. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11621-11631 (2020) +3. Cao, H., Xu, Y., Yang, J., Yin, P., Yuan, S., Xie, L.: Mopa: Multi-modal prior aided domain adaptation for 3d semantic segmentation. arXiv preprint arXiv:2309.11839 (2023) +4. Cao, H., Xu, Y., Yang, J., Yin, P., Yuan, S., Xie, L.: Multi-modal continual test-time adaptation for 3d semantic segmentation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 18809-18819 (October 2023) +5. Chen, R., Liu, Y., Kong, L., Zhu, X., Ma, Y., Li, Y., Hou, Y., Qiao, Y., Wang, W.: Clip2scene: Towards label-efficient 3d scene understanding by clip. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7020-7030 (2023) +6. Choy, C., Gwak, J., Savarese, S.: 4d spatio-temporal convnets: Minkowski convolutional neural networks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3075-3084 (2019) +7. Fan, H., Yang, Y., Kankanhalli, M.: Point 4d transformer networks for spatiotemporal modeling in point cloud videos. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14204-14213 (2021) +8. Fan, H., Yu, X., Ding, Y., Yang, Y., Kankanhalli, M.: Pistnet: Point spatio-temporal convolution on point cloud sequences. arXiv preprint arXiv:2205.13713 (2022) +9. Feng, D., Haase-Schütz, C., Rosenbaum, L., Hertlein, H., Glaeser, C., Timm, F., Wiesbeck, W., Dietmayer, K.: Deep multi-modal object detection and semantic segmentation for autonomous driving: Datasets, methods, and challenges. IEEE Transactions on Intelligent Transportation Systems 22(3), 1341-1360 (2020) +10. Geyer, J., Kassahun, Y., Mahmudi, M., Ricou, X., Durgesh, R., Chung, A.S., Hauswald, L., Pham, V.H., Mühlegg, M., Dorn, S., et al.: A2d2: Audi autonomous driving dataset. arXiv preprint arXiv:2004.06320 (2020) +11. Gong, T., Jeong, J., Kim, T., Kim, Y., Shin, J., Lee, S.J.: Robust continual test-time adaptation: Instance-aware bn and prediction-balanced memory. arXiv preprint arXiv:2208.05117 (2022) +12. Goyal, S., Sun, M., Raghunathan, A., Kolter, J.Z.: Test time adaptation via conjugate pseudo-labels. In: Advances in Neural Information Processing Systems (2022) + +13. Graham, B.: Sparse 3d convolutional neural networks. arXiv preprint arXiv:1505.02890 (2015) +14. Guo, Y., Wang, H., Hu, Q., Liu, H., Liu, L., Bennamoun, M.: Deep learning for 3d point clouds: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence 43(12), 4338-4364 (2020) +15. Huang, S., Gojcic, Z., Huang, J., Wieser, A., Schindler, K.: Dynamic 3d scene analysis by point cloud accumulation. In: European Conference on Computer Vision. pp. 674-690. Springer (2022) +16. Jaritz, M., Vu, T.H., De Charette, R., Wirbel, E., Pérez, P.: Cross-modal learning for domain adaptation in 3d semantic segmentation. IEEE Transactions on Pattern Analysis and Machine Intelligence 45(2), 1533-1544 (2022) +17. Ji, X., Yuan, S., Yin, P., Xie, L.: Lio-gvm: an accurate, tightly-coupled lidar-inertial odometry with gaussian voxel map. IEEE Robotics and Automation Letters (2024) +18. Li, M., Zhang, Y., Xie, Y., Gao, Z., Li, C., Zhang, Z., Qu, Y.: Cross-domain and cross-modal knowledge distillation in domain adaptation for 3d semantic segmentation. In: Proceedings of the 30th ACM International Conference on Multimedia. pp. 3829-3837 (2022) +19. Liang, J., Hu, D., Feng, J.: Do we really need to access the source data? source hypothesis transfer for unsupervised domain adaptation. In: International Conference on Machine Learning. pp. 6028-6039. PMLR (2020) +20. Liu, W., Luo, Z., Cai, Y., Yu, Y., Ke, Y., Junior, J.M., Gonçalves, W.N., Li, J.: Adversarial unsupervised domain adaptation for 3d semantic segmentation with multi-modal learning. ISPRS Journal of Photogrammetry and Remote Sensing 176, 211-221 (2021) +21. Liu, Y., Kothari, P., Van Delft, B., Bellot-Gurlet, B., Mordan, T., Alahi, A.: Ttt++: When does self-supervised test-time training fail or thrive? Advances in Neural Information Processing Systems 34, 21808-21820 (2021) +22. Niu, S., Wu, J., Zhang, Y., Chen, Y., Zheng, S., Zhao, P., Tan, M.: Efficient test-time model adaptation without forgetting. In: International Conference on Machine Learning. pp. 16888-16905. PMLR (2022) +23. Niu, S., Wu, J., Zhang, Y., Wen, Z., Chen, Y., Zhao, P., Tan, M.: Towards stable test-time adaptation in dynamic wild world. arXiv preprint arXiv:2302.12400 (2023) +24. Peng, D., Lei, Y., Li, W., Zhang, P., Guo, Y.: Sparse-to-dense feature matching: Intra and inter domain cross-modal learning in domain adaptation for 3d semantic segmentation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 7108-7117 (2021) +25. Piergiovanni, A., Casser, V., Ryoo, M.S., Angelova, A.: 4d-net for learned multimodal alignment. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 15435-15445 (2021) +26. Qi, C.R., Su, H., Mo, K., Guibas, L.J.: Pointnet: Deep learning on point sets for 3d classification and segmentation. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 652-660 (2017) +27. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International Conference on Machine Learning. pp. 8748-8763. PMLR (2021) +28. Saltori, C., Krivosheev, E., Lathuilière, S., Sebe, N., Galasso, F., Fiameni, G., Ricci, E., Poiesi, F.: Gipso: Geometrically informed propagation for online adaptation in 3d lidar segmentation. In: European Conference on Computer Vision. pp. 567-585. Springer (2022) + +29. Shin, I., Tsai, Y.H., Zhuang, B., Schulter, S., Liu, B., Garg, S., Kweon, I.S., Yoon, K.J.: Mm-tta: multi-modal test-time adaptation for 3d semantic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16928-16937 (2022) +30. Simons, C., Raychaudhuri, D.S., Ahmed, S.M., You, S., Karydis, K., Roy-Chowdhury, A.K.: Summit: Source-free adaptation of uni-modal models to multimodal targets. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 1239-1249 (October 2023) +31. Song, J., Lee, J., Kweon, I.S., Choi, S.: Ecotta: Memory-efficient continual test-time adaptation via self-distilled regularization. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11920-11929 (2023) +32. Su, Y., Xu, X., Li, T., Jia, K.: Revisiting realistic test-time training: Sequential inference and adaptation by anchored clustering regularized self-training. arXiv preprint arXiv:2303.10856 (2023) +33. Tang, H., Liu, Z., Zhao, S., Lin, Y., Lin, J., Wang, H., Han, S.: Searching efficient 3d architectures with sparse point-voxel convolution. In: European Conference on Computer Vision. pp. 685–702. Springer (2020) +34. Tarvainen, A., Valpola, H.: Mean teachers are better role models: Weight-averaged consistency targets improve semi-supervised deep learning results. Advances in Neural Information Processing Systems 30 (2017) +35. Vedula, S., Baker, S., Rander, P., Collins, R., Kanade, T.: Three-dimensional scene flow. In: Proceedings of the Seventh IEEE International Conference on Computer Vision. vol. 2, pp. 722-729. IEEE (1999) +36. Vizzo, I., Guadagnino, T., Mersch, B., Wiesmann, L., Behley, J., Stachniss, C.: Kiss-icp: In defense of point-to-point icp-simple, accurate, and robust registration if done the right way. IEEE Robotics and Automation Letters 8(2), 1029-1036 (2023) +37. Vogel, C., Schindler, K., Roth, S.: 3d scene flow estimation with a rigid motion prior. In: 2011 International Conference on Computer Vision. pp. 1291-1298. IEEE (2011) +38. Vogel, C., Schindler, K., Roth, S.: Piecewise rigid scene flow. In: Proceedings of the IEEE International Conference on Computer Vision. pp. 1377-1384 (2013) +39. Wang, D., Shelhamer, E., Liu, S., Olshausen, B., Darrell, T.: Tent: Fully test-time adaptation by entropy minimization. In: International Conference on Learning Representations (2021), https://openreview.net/forum?id=uX13bZLkr3c +40. Wang, J.K., Wibisono, A.: Towards understanding gd with hard and conjugate pseudo-labels for test-time adaptation. arXiv preprint arXiv:2210.10019 (2022) +41. Wang, Q., Fink, O., Van Gool, L., Dai, D.: Continual test-time domain adaptation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7201-7211 (2022) +42. Wyner, A.: Recent results in the shannon theory. IEEE Transactions on information Theory 20(1), 2-10 (1974) +43. Xie, E., Wang, W., Yu, Z., Anandkumar, A., Alvarez, J.M., Luo, P.: Segformer: Simple and efficient design for semantic segmentation with transformers. Advances in Neural Information Processing Systems 34, 12077-12090 (2021) +44. Xing, B., Ying, X., Wang, R., Yang, J., Chen, T.: Cross-modal contrastive learning for domain adaptation in 3d semantic segmentation. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 37, pp. 2974-2982 (2023) +45. Xu, J., Miao, Z., Zhang, D., Pan, H., Liu, K., Hao, P., Zhu, J., Sun, Z., Li, H., Zhan, X.: Int: Towards infinite-frames 3d detection with an efficient framework. In: European Conference on Computer Vision. pp. 193–209. Springer (2022) + +46. Yin, P., Cao, H., Nguyen, T.M., Yuan, S., Zhang, S., Liu, K., Xie, L.: Outram: One-shot global localization via triangulated scene graph and global outlier pruning. arXiv preprint arXiv:2309.08914 (2023) +47. Yuan, L., Xie, B., Li, S.: Robust test-time adaptation in dynamic scenarios. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15922-15932 (2023) +48. Zhang, M., Levine, S., Finn, C.: Memo: Test time robustness via adaptation and augmentation. Advances in Neural Information Processing Systems 35, 38629-38642 (2022) \ No newline at end of file diff --git a/2024/Reliable Spatial-Temporal Voxels For Multi-Modal Test-Time Adaptation/images.zip b/2024/Reliable Spatial-Temporal Voxels For Multi-Modal Test-Time Adaptation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..168d88416ddeabd44c4d0777f408a1e2f6f79c69 --- /dev/null +++ b/2024/Reliable Spatial-Temporal Voxels For Multi-Modal Test-Time Adaptation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9656eefeffe1d6556deac6f199339733f7dee3d74b8b1e324f60ec70063860eb +size 501111 diff --git a/2024/Reliable Spatial-Temporal Voxels For Multi-Modal Test-Time Adaptation/layout.json b/2024/Reliable Spatial-Temporal Voxels For Multi-Modal Test-Time Adaptation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..475e58e8187fa8eaf9f441306814051e2cd9a8a7 --- /dev/null +++ b/2024/Reliable Spatial-Temporal Voxels For Multi-Modal Test-Time Adaptation/layout.json @@ -0,0 +1,9211 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 177, + 111, + 436, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 177, + 111, + 436, + 148 + ], + "spans": [ + { + "bbox": [ + 177, + 111, + 436, + 148 + ], + "type": "text", + "content": "Reliable Spatial-Temporal Voxels For Multi-Modal Test-Time Adaptation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 137, + 167, + 477, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 167, + 477, + 193 + ], + "spans": [ + { + "bbox": [ + 137, + 167, + 477, + 193 + ], + "type": "text", + "content": "Haozhi Cao" + }, + { + "bbox": [ + 137, + 167, + 477, + 193 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 137, + 167, + 477, + 193 + ], + "type": "text", + "content": ", Yuecong Xu" + }, + { + "bbox": [ + 137, + 167, + 477, + 193 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 137, + 167, + 477, + 193 + ], + "type": "text", + "content": ", Jianfei Yang" + }, + { + "bbox": [ + 137, + 167, + 477, + 193 + ], + "type": "inline_equation", + "content": "^{1,3}" + }, + { + "bbox": [ + 137, + 167, + 477, + 193 + ], + "type": "text", + "content": ", Pengyu Yin" + }, + { + "bbox": [ + 137, + 167, + 477, + 193 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 137, + 167, + 477, + 193 + ], + "type": "text", + "content": ", Xingyu Ji" + }, + { + "bbox": [ + 137, + 167, + 477, + 193 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 137, + 167, + 477, + 193 + ], + "type": "text", + "content": ", Shenghai Yuan" + }, + { + "bbox": [ + 137, + 167, + 477, + 193 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 137, + 167, + 477, + 193 + ], + "type": "text", + "content": ", and Lihua Xie" + }, + { + "bbox": [ + 137, + 167, + 477, + 193 + ], + "type": "inline_equation", + "content": "^{1}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 133, + 201, + 478, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 201, + 478, + 224 + ], + "spans": [ + { + "bbox": [ + 133, + 201, + 478, + 224 + ], + "type": "text", + "content": "1 School of Electrical and Electronic Engineering, Nanyang Technological University {haozhi002, jianfei.yang, pengyu001, xinyu001, shyuan, elhxie}@ntu.edu.sg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 193, + 224, + 419, + 257 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 193, + 224, + 419, + 257 + ], + "spans": [ + { + "bbox": [ + 193, + 224, + 419, + 257 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 193, + 224, + 419, + 257 + ], + "type": "text", + "content": " Department of Electrical and Computer Engineering, National University of Singapore yc.xu@nus.edu.sg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 257, + 480, + 269 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 257, + 480, + 269 + ], + "spans": [ + { + "bbox": [ + 132, + 257, + 480, + 269 + ], + "type": "text", + "content": "3 School of Mechanical and Aerospace Engineering, Nanyang Technological University" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 160, + 297, + 452, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 297, + 452, + 517 + ], + "spans": [ + { + "bbox": [ + 160, + 297, + 452, + 517 + ], + "type": "text", + "content": "Abstract. Multi-modal test-time adaptation (MM-TTA) is proposed to adapt models to an unlabeled target domain by leveraging the complementary multi-modal inputs in an online manner. Previous MM-TTA methods for 3D segmentation rely on predictions of cross-modal information in each input frame, while they ignore the fact that predictions of geometric neighborhoods within consecutive frames are highly correlated, leading to unstable predictions across time. To fulfill this gap, we propose ReLiable Spatial-temporal Voxels (Latte), an MM-TTA method that leverages reliable cross-modal spatial-temporal correspondences for multi-modal 3D segmentation. Motivated by the fact that reliable predictions should be consistent with their spatial-temporal correspondences, Latte aggregates consecutive frames in a slide window manner and constructs Spatial-Temopral (ST) voxels to capture temporally local prediction consistency for each modality. After filtering out ST voxels with high ST entropy, Latte conducts cross-modal learning for each point and pixel by attending to those with reliable and consistent predictions among both spatial and temporal neighborhoods. Experimental results show that Latte achieves state-of-the-art performance on three different MM-TTA benchmarks compared to previous MM-TTA or TTA methods. Visit our project site https://sites.google.com/view/eccv24-latte." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 159, + 526, + 453, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 526, + 453, + 548 + ], + "spans": [ + { + "bbox": [ + 159, + 526, + 453, + 548 + ], + "type": "text", + "content": "Keywords: Test-time adaptation " + }, + { + "bbox": [ + 159, + 526, + 453, + 548 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 526, + 453, + 548 + ], + "type": "text", + "content": " Multi-modal learning " + }, + { + "bbox": [ + 159, + 526, + 453, + 548 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 526, + 453, + 548 + ], + "type": "text", + "content": " 3D semantic segmentation" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 569, + 230, + 582 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 569, + 230, + 582 + ], + "spans": [ + { + "bbox": [ + 132, + 569, + 230, + 582 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": "3D semantic segmentation is a fundamental task to achieve various fully autonomous applications such as autonomous driving and robot navigation [9,14]. With the increasing urge for robust sensing, multi-modal sensors (e.g., cameras and LiDARs) are widely adopted in autonomous systems containing rich complementary information from different sensors. However, conventional deep-learning-based solutions lean on expensive point-wise annotations and perform" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 169, + 114, + 278, + 197 + ], + "blocks": [ + { + "bbox": [ + 169, + 114, + 278, + 197 + ], + "lines": [ + { + "bbox": [ + 169, + 114, + 278, + 197 + ], + "spans": [ + { + "bbox": [ + 169, + 114, + 278, + 197 + ], + "type": "image", + "image_path": "b8e4c8e876345fb3c69ae9096facc9e311366ae042e737c0a55381edfca585f2.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 206, + 482, + 294 + ], + "lines": [ + { + "bbox": [ + 130, + 206, + 482, + 294 + ], + "spans": [ + { + "bbox": [ + 130, + 206, + 482, + 294 + ], + "type": "text", + "content": "Fig. 1: Illustration of previous MM-TTA methods and our Latte. (a) visualizes the predictions from the previous state-of-the-art MM-TTA method [29] in consecutive frames on SemanticKITTI [1], where regions in white boxes exhibit noisy predictions when considering single-frame input. (b) presents the overall framework of Latte, which firstly obtains ST voxels for each modality given the merged input of consecutive frames within each slide window and then computes the modality-specific ST entropy " + }, + { + "bbox": [ + 130, + 206, + 482, + 294 + ], + "type": "inline_equation", + "content": "E^{2\\mathrm{D}}" + }, + { + "bbox": [ + 130, + 206, + 482, + 294 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 206, + 482, + 294 + ], + "type": "inline_equation", + "content": "E^{3\\mathrm{D}}" + }, + { + "bbox": [ + 130, + 206, + 482, + 294 + ], + "type": "text", + "content": ". ST entropy indicates the relative reliability of each voxel in each modality and therefore is utilized in the following adaptive cross-modal attending." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 284, + 113, + 444, + 198 + ], + "blocks": [ + { + "bbox": [ + 284, + 113, + 444, + 198 + ], + "lines": [ + { + "bbox": [ + 284, + 113, + 444, + 198 + ], + "spans": [ + { + "bbox": [ + 284, + 113, + 444, + 198 + ], + "type": "image", + "image_path": "a500e2fb9aa47cb9e5e77e60915ad7cb7531a73e950dab13d544e87e16d75802.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 318, + 480, + 377 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 318, + 480, + 377 + ], + "spans": [ + { + "bbox": [ + 130, + 318, + 480, + 377 + ], + "type": "text", + "content": "poorly on data from different domains due to the domain shift [6, 26, 33]. To resolve this problem, previous works have proposed Multi-Modal Unsupervised Domain Adaptation (MM-UDA) approaches [16, 18, 24, 44] that employ multimodal information to transfer knowledge learned from the labeled source domain to the unlabeled target domain without expensive human annotations." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 378, + 482, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 378, + 482, + 510 + ], + "spans": [ + { + "bbox": [ + 130, + 378, + 482, + 510 + ], + "type": "text", + "content": "Nevertheless, the adaptation process of MM-UDA is completely offline, which requires full access to the source domain dataset and multiple training epochs. Despite its effectiveness, offline training is usually infeasible when encountering distribution shifts during the inference stage that require adapting models in real time. Inspired by Test-Time Adaptation (TTA) [22, 23], a recent work [29] proposes the first Multi-Modal Test-Time Adaptation (MM-TTA) method for 3D segmentation. Similar to the settings of TTA, MM-TTA prohibits access to any raw sample from the source domain dataset and adapts the model during the testing stage. This results in a quick adaptation scenario (i.e., one epoch for training only as in [29, 32]) that requires stable optimization in an online manner. To this end, MM-TTA methods must seek a reliable source of supervision signals." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "content": "To fulfill the efficiency and stability of adaptation, the current state-of-the-art MM-TTA method [29] evaluates modal reliability by the prediction consistency between different experts in a frame-wise manner. However, due to the domain shift between the online frame and the pre-trained source dataset, the single-frame prediction is usually unstable. Existing approaches rely on pseudo labels to refine the single-frame prediction but still fail in some cases. As shown in Fig. 1a, the single-frame refinement strategy [29] suffers from noisy single-frame predictions on the car in the white rectangle across time. Such inconsistent predictions across time degenerate the segmentation performance and aggravate denoising stress in downstream tasks (e.g., semantic-based retrivial [46] and obstacle recognition). On one hand, these temporally unstable single-frame predictions could be wrongly regarded as reliable ones, propagating prediction noise to the other modality and causing error accumulation or even catastrophic forgetting [22,23]." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "text", + "content": "H. Cao et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 480, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 480, + 163 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 480, + 163 + ], + "type": "text", + "content": "On the other hand, while previous methods [4, 41] alleviate this instability by utilizing the average predictions of multiple augmented frames instead, they are computationally expensive for online adaptation since the inference time grows linearly with the increasing number of augmented frames." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 164, + 481, + 307 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 164, + 481, + 307 + ], + "spans": [ + { + "bbox": [ + 130, + 164, + 481, + 307 + ], + "type": "text", + "content": "This work aims to efficiently suppress the single-frame instability for MM-TTA by exploiting the correlation between multi-frame correspondences. Specifically, we propose that 3D space can be efficiently divided into multiple voxels [33], each of which contains points and point-corresponding pixels located at the same geometric region. While captured at different timestamps, it can be observed that points or pixels located at the same voxel can be viewed as different observations on the same semantic objects [28]. Reliable predictions should therefore be certain and consistent within their corresponding voxels across time, whether we evaluate them from a temporally global perspective (i.e., all input frames) or a temporally local perspective (i.e., certain consecutive frames). By aggregating the predictions in a voxel-wise manner and re-evaluating their reliability, the single-frame prediction noise as in Fig. 1a can be effectively alleviated." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 308, + 481, + 487 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 308, + 481, + 487 + ], + "spans": [ + { + "bbox": [ + 130, + 308, + 481, + 487 + ], + "type": "text", + "content": "Motivated by the above observation, we propose a novel MM-TTA method called ReLiable Spatial-temporal Voxels (Latte), which leverages the spatial-temporal correspondences within consecutive frames in a cross-modal manner as in Fig. 1b. Specifically, given frames of point clouds with their estimated poses, we regard points in the same voxel extracted from the merged point cloud frames as spatial-temporal correspondences. Different from previous works that merge all input frames [5,7] or regularize frame-to-frame consistency [28], Latte aggregates consecutive frames in a sliding window manner to estimate the temporally local prediction consistency through the proposed Spatial-Temporal voxels (ST voxels) and entropy (ST entropy). Based on the reliability estimated by ST entropy, we conduct cross-modal learning in an adaptive attending manner to reduce the contributions of predictions from the noisy modality. To better facilitate the progress of MM-TTA, we evaluate our Latte with more recent backbones (SegFormer [43] and SPVCNN [33]) on three different benchmarks, where Latte consistently outperforms previous TTA or MM-TTA methods." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 488, + 481, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 488, + 481, + 582 + ], + "spans": [ + { + "bbox": [ + 130, + 488, + 481, + 582 + ], + "type": "text", + "content": "In summary, our contributions are three-fold. Firstly, we propose Latte, a novel MM-TTA method for multi-modal 3D segmentation, which is the first work that incorporates spatial-temporal correlations for MM-TTA segmentation to our best knowledge. Secondly, Latte efficiently extracts spatial-temporal correspondence through ST voxels and estimates their reliability by ST entropy, which is further incorporated to enhance adaptive cross-modal attending. Thirdly, our experimental results show that Latte outperforms previous state-of-the-art TTA and MM-TTA methods on three different benchmarks." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 602, + 242, + 616 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 602, + 242, + 616 + ], + "spans": [ + { + "bbox": [ + 132, + 602, + 242, + 616 + ], + "type": "text", + "content": "2 Related Works" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "type": "text", + "content": "Test-Time Adaptation (TTA). TTA is proposed to mitigate the domain shift between the training data and the testing data in an online manner, which is firstly proposed by TENT [39]. Due to its practical and challenging setting," + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 148, + 91, + 447, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 148, + 91, + 447, + 103 + ], + "spans": [ + { + "bbox": [ + 148, + 91, + 447, + 103 + ], + "type": "text", + "content": "Reliable Spatial-Temporal Voxels For Multi-Modal Test-Time Adaptation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 272 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 272 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 272 + ], + "type": "text", + "content": "TTA is attracting more and more attention. Previously proposed TTA methods attempt to address this task from different perspectives, such as entropy minimization [22, 23, 39], self-training with pseudo-labels [12, 40], and augmentation invariance [41, 48]. In addition to regularizing predictions, some methods propose to adapt networks from the feature level instead [21, 32], while more recent methods begin to consider different variants of TTA scenarios, such as continual TTA [4, 22, 41], non-i.i.d TTA [11], or the mix of aforementioned cases [47]. Different from source-free adaptation [19, 30] which also discards the access to the source domain yet requires multiple training epochs, the adaptation process of TTA is completely online, following the \"one-pass\" protocol as in [32]. While most existing TTA methods are designed for image classification or image segmentation without explicit temporal correlation, we argue temporal information in 3D segmentation can be effectively leveraged for TTA as Latte in this work." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 277, + 482, + 493 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 277, + 482, + 493 + ], + "spans": [ + { + "bbox": [ + 130, + 277, + 482, + 493 + ], + "type": "text", + "content": "Multi-modal domain adaptation for 3D segmentation. To avoid expensive annotation costs and overcome the poor generalizability of fully supervised solutions, various multi-modal domain adaptation methods have been investigated. Specifically for MM-UDA, xMUDA [16] is the primary works that explore incorporating cross-modal learning with MM-UDA. It regards cross-modal prediction consistency and pseudo-labels as its supervision signals in the unlabeled target domain. Most subsequent works propose solutions to address its limitations from different perspectives, such as more diverse point-pixel correspondence [24, 44], procedures to mitigate domain gaps [18, 20], and alleviating class-imbalanced problem [3]. Since MM-UDA is infeasible when facing online domain shifts, some previous works [4, 29] develop different methods to address the domain gap in an online manner, including cross-modal pseudo-label generation [29] and adaptive modality attending [4]. A recent work [30] proposes a source-free domain adaptation method by estimating cross-modal prediction agreement, yet it requires multiple offline training epochs. Different from previous multi-modal domain adaptation methods that ignore temporal relationships, Latte couples spatial-temporal correspondence to mitigate the single-frame instability and achieve better online adaptation performance." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "text", + "content": "Temporal processing of point clouds. Since point clouds are naturally in the form of temporally consecutive frames, previous methods have widely explored how to leverage such temporal information for different tasks. Specifically, scene flow [15,35,37,38] contains the movement of every point in the 3D world. Considering such point-wise computation expensive, previous works on different tasks (e.g., segmentation [5,6,28] and object detection [25]) usually interpret the temporal information from a simpler perspective. In terms of fully supervised learning, Choy et al. [6] propose the first 4D segmentation method by generalized sparse convolution kernels [13] and Fan et al. [8] turn to the fully connected learning scheme as in [26], while Xu et al. [45] extract long-term temporal information by an efficient memory bank. On the other hand, considering scenarios with insufficient or zero annotations, some works propose different strategies to couple spatial-temporal information without ground-truth labels. For instance, Chen et al. [5] proposes to guide voxelized spatial-temporal representation with" + } + ] + } + ], + "index": 4 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "type": "text", + "content": "H. Cao et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 141, + 114, + 473, + 223 + ], + "blocks": [ + { + "bbox": [ + 141, + 114, + 473, + 223 + ], + "lines": [ + { + "bbox": [ + 141, + 114, + 473, + 223 + ], + "spans": [ + { + "bbox": [ + 141, + 114, + 473, + 223 + ], + "type": "image", + "image_path": "84f048bd8d47220c0c1f832940d02649d5471164a1790f6cdf6db3e16604bd55.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 234, + 482, + 312 + ], + "lines": [ + { + "bbox": [ + 130, + 234, + 482, + 312 + ], + "spans": [ + { + "bbox": [ + 130, + 234, + 482, + 312 + ], + "type": "text", + "content": "Fig. 2: Overall structure of Latte. Taking a student prediction frame of one modality as the query input, our slide-window extraction searches its spatial-temporal correspondences through voxelization within a time window to establish the temporally local prediction consistency. ST voxels are then generated, where those with high ST entropy (larger than " + }, + { + "bbox": [ + 130, + 234, + 482, + 312 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 130, + 234, + 482, + 312 + ], + "type": "text", + "content": "-quantile) are discarded as unreliable correspondences, while the others are leveraged for adaptive cross-modal learning by attending to the modality with lower ST entropy in a voxel-wise manner." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 335, + 480, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 335, + 480, + 396 + ], + "spans": [ + { + "bbox": [ + 130, + 335, + 480, + 396 + ], + "type": "text", + "content": "strong text embedding from CLIP [27] instead of labels in a multi-modal manner, while Saitor et al. [28] interpret such representation by nearest neighbor search in a frame-to-frame manner. In this work, we design a slide window aggregation and voxelization rather than aggregating all frames as in [5], which yields a better indication of the prediction reliability for each point." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 413, + 232, + 427 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 413, + 232, + 427 + ], + "spans": [ + { + "bbox": [ + 132, + 413, + 232, + 427 + ], + "type": "text", + "content": "3 Methodology" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 437, + 482, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 437, + 482, + 605 + ], + "spans": [ + { + "bbox": [ + 130, + 437, + 482, + 605 + ], + "type": "text", + "content": "Problem definition. During MM-TTA for 3D segmentation, both 2D RGB images and 3D point clouds are captured consecutively in the target domain " + }, + { + "bbox": [ + 130, + 437, + 482, + 605 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 130, + 437, + 482, + 605 + ], + "type": "text", + "content": ", denoted as " + }, + { + "bbox": [ + 130, + 437, + 482, + 605 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{\\mathcal{T},t}^{2\\mathrm{D}} \\in \\mathbb{R}^{3 \\times H \\times W}" + }, + { + "bbox": [ + 130, + 437, + 482, + 605 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 437, + 482, + 605 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{\\mathcal{T},t}^{3\\mathrm{D}} \\in \\mathbb{R}^{N \\times 4}" + }, + { + "bbox": [ + 130, + 437, + 482, + 605 + ], + "type": "text", + "content": ", respectively, where " + }, + { + "bbox": [ + 130, + 437, + 482, + 605 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 437, + 482, + 605 + ], + "type": "text", + "content": " represents the frame order and " + }, + { + "bbox": [ + 130, + 437, + 482, + 605 + ], + "type": "inline_equation", + "content": "H, W" + }, + { + "bbox": [ + 130, + 437, + 482, + 605 + ], + "type": "text", + "content": " are the height and width of input images. The point-wise input contains the 3D coordinate " + }, + { + "bbox": [ + 130, + 437, + 482, + 605 + ], + "type": "inline_equation", + "content": "\\{x, y, z\\}" + }, + { + "bbox": [ + 130, + 437, + 482, + 605 + ], + "type": "text", + "content": " and the point feature (e.g., intensity, reflectivity, etc.). Modality-specific networks are pre-trained on the labeled source domain before adaptation, denoted as " + }, + { + "bbox": [ + 130, + 437, + 482, + 605 + ], + "type": "inline_equation", + "content": "\\phi_{\\mathcal{T},t}^{\\mathrm{m}}(\\cdot)" + }, + { + "bbox": [ + 130, + 437, + 482, + 605 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 130, + 437, + 482, + 605 + ], + "type": "inline_equation", + "content": "\\mathrm{m} \\in \\{2\\mathrm{D}, 3\\mathrm{D}\\}" + }, + { + "bbox": [ + 130, + 437, + 482, + 605 + ], + "type": "text", + "content": ". During MM-TTA, both networks are initialized from the parameters pre-trained on the source domain. Following previous works [16,29], 3D points are projected to the 2D image based on the relative projection matrix to obtain cross-modal correspondences within each frame, resulting in the modality-specific predictions " + }, + { + "bbox": [ + 130, + 437, + 482, + 605 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_{\\mathcal{T},t}^{\\mathrm{m}} = \\phi_{\\mathcal{T},t}^{\\mathrm{m}}(\\mathbf{x}_{\\mathcal{T},t}^{\\mathrm{m}})" + }, + { + "bbox": [ + 130, + 437, + 482, + 605 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 437, + 482, + 605 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_{\\mathcal{T},t}^{\\mathrm{m}} \\in \\mathbb{R}^{N \\times K}" + }, + { + "bbox": [ + 130, + 437, + 482, + 605 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 130, + 437, + 482, + 605 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 437, + 482, + 605 + ], + "type": "text", + "content": " denotes the number of semantic classes. Data from the labeled source domain are strictly inaccessible and we omit the target domain subscript " + }, + { + "bbox": [ + 130, + 437, + 482, + 605 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 130, + 437, + 482, + 605 + ], + "type": "text", + "content": " by default." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 606, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 606, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 606, + 482, + 666 + ], + "type": "text", + "content": "The multi-modal input of online 3D segmentation is temporally consecutive, which exhibits strong temporal relationships between frames. The geometric neighborhoods in consecutive frames can be viewed as the naturally various observations on the same semantic object that can be leveraged to stabilize noisy single-frame predictions. Motivated by this observation, we develop Latte," + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 148, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 148, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 148, + 91, + 447, + 102 + ], + "type": "text", + "content": "Reliable Spatial-Temporal Voxels For Multi-Modal Test-Time Adaptation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "content": "which discovers reliable spatial-temporal (ST) correspondences to improve both optimization stability and prediction consistency during MM-TTA. As shown in Fig. 2, given consecutive frames of 2D images and 3D point clouds as input, Latte firstly extracts their frame-wise predictions from both student and teacher networks (Sec. 3.1). These frame-wise predictions are then passed to the slide-window extraction and voxelization (Sec. 3.2). ST voxels and ST entropy are subsequently computed indicating the prediction consistency and certainty (Sec. 3.3). The final ST entropy is leveraged for cross-modal attending to alleviate the noise from unreliable modality-specific predictions (Sec. 3.4)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 239, + 422, + 250 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 239, + 422, + 250 + ], + "spans": [ + { + "bbox": [ + 131, + 239, + 422, + 250 + ], + "type": "text", + "content": "3.1 Frame-wise Predictions from Students and Teachers" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 255, + 479, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 255, + 479, + 327 + ], + "spans": [ + { + "bbox": [ + 130, + 255, + 479, + 327 + ], + "type": "text", + "content": "Before exploiting spatial-temporal correspondence, we first generate the framewise predictions from each modality. Motivated by the fact that moving average models can provide more stable online predictions [4,34,41], Latte utilizes weight-averaged teacher models and fast student models for each modality as in previous works [29, 41], denoted as " + }, + { + "bbox": [ + 130, + 255, + 479, + 327 + ], + "type": "inline_equation", + "content": "\\tilde{\\phi}_t^{\\mathrm{m}}(\\cdot)" + }, + { + "bbox": [ + 130, + 255, + 479, + 327 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 255, + 479, + 327 + ], + "type": "inline_equation", + "content": "\\phi_t^{\\mathrm{m}}(\\cdot)" + }, + { + "bbox": [ + 130, + 255, + 479, + 327 + ], + "type": "text", + "content": ", respectively. Given each input frame " + }, + { + "bbox": [ + 130, + 255, + 479, + 327 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_t^{\\mathrm{m}}" + }, + { + "bbox": [ + 130, + 255, + 479, + 327 + ], + "type": "text", + "content": ", the student predictions " + }, + { + "bbox": [ + 130, + 255, + 479, + 327 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_t^{\\mathrm{m}}" + }, + { + "bbox": [ + 130, + 255, + 479, + 327 + ], + "type": "text", + "content": " and teacher predictions " + }, + { + "bbox": [ + 130, + 255, + 479, + 327 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{p}}_t^{\\mathrm{m}}" + }, + { + "bbox": [ + 130, + 255, + 479, + 327 + ], + "type": "text", + "content": " are computed as:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 242, + 333, + 480, + 346 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 242, + 333, + 480, + 346 + ], + "spans": [ + { + "bbox": [ + 242, + 333, + 480, + 346 + ], + "type": "interline_equation", + "content": "\\tilde {\\mathbf {p}} _ {t} ^ {\\mathrm {m}} = \\tilde {\\phi} _ {t} ^ {\\mathrm {m}} \\left(\\mathbf {x} _ {t} ^ {\\mathrm {m}}\\right), \\mathbf {p} _ {t} ^ {\\mathrm {m}} = \\phi_ {t} ^ {\\mathrm {m}} \\left(\\mathbf {x} _ {t} ^ {\\mathrm {m}}\\right), \\tag {1}", + "image_path": "58be026bdc2e651d476bf514f4d9b7a0e08eb45feb80bd3d8f6feb6b1319952b.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 351, + 482, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 351, + 482, + 411 + ], + "spans": [ + { + "bbox": [ + 130, + 351, + 482, + 411 + ], + "type": "text", + "content": "where the teacher model " + }, + { + "bbox": [ + 130, + 351, + 482, + 411 + ], + "type": "inline_equation", + "content": "\\tilde{\\phi}_t^{\\mathrm{m}}" + }, + { + "bbox": [ + 130, + 351, + 482, + 411 + ], + "type": "text", + "content": " is initialized from the source domain pre-trained models. In the following process, the teacher's predictions of one modality are regarded as the cross-modal guidance of student predictions from the other modality. In practice, the gradient propagation of teacher model predictions is discarded to prevent directly updating teacher models." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 426, + 418, + 438 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 426, + 418, + 438 + ], + "spans": [ + { + "bbox": [ + 131, + 426, + 418, + 438 + ], + "type": "text", + "content": "3.2 Slide Window Frame Aggregation and Voxelization" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 442, + 482, + 550 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 442, + 482, + 550 + ], + "spans": [ + { + "bbox": [ + 130, + 442, + 482, + 550 + ], + "type": "text", + "content": "Our goal is to efficiently leverage spatial-temporal correspondences between predictions and seek reliability estimation through prediction consistency within correspondences. Establishing spatial-temporal correspondences in 3D space has been explored in other deep learning tasks, where they mainly consider temporally global relationships of all input frames [5,28] or frame-to-frame correspondences [28]. However, they both have some limitations in terms of consistency and certainty evaluation: the former can not highlight the risk of some inconsistent predictions within a short time window (i.e., temporally local inconsistency), while the latter suffers from an insufficient number of correspondences." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 550, + 482, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 550, + 482, + 647 + ], + "spans": [ + { + "bbox": [ + 130, + 550, + 482, + 647 + ], + "type": "text", + "content": "To effectively empower spatial-temporal information with Latte, we propose a slide window frame aggregation that focuses on temporally local correspondences that lie within a time window to evaluate temporally local consistency. Specifically, given a student prediction frame " + }, + { + "bbox": [ + 130, + 550, + 482, + 647 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_i^{\\mathrm{m}}" + }, + { + "bbox": [ + 130, + 550, + 482, + 647 + ], + "type": "text", + "content": " with the temporal index " + }, + { + "bbox": [ + 130, + 550, + 482, + 647 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 550, + 482, + 647 + ], + "type": "text", + "content": " as the query, the correspondence search is conducted within a time window of consecutive frames denoted as " + }, + { + "bbox": [ + 130, + 550, + 482, + 647 + ], + "type": "inline_equation", + "content": "\\{j||j - i|\\leq w_t\\}" + }, + { + "bbox": [ + 130, + 550, + 482, + 647 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 550, + 482, + 647 + ], + "type": "inline_equation", + "content": "w_{t}" + }, + { + "bbox": [ + 130, + 550, + 482, + 647 + ], + "type": "text", + "content": " is the pre-defined time window size. The merged point cloud " + }, + { + "bbox": [ + 130, + 550, + 482, + 647 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{x}}_i^{3\\mathrm{D}}\\in \\mathbb{R}^{\\hat{N}\\times 3}" + }, + { + "bbox": [ + 130, + 550, + 482, + 647 + ], + "type": "text", + "content": " (point features are omitted for simplicity) for correspondence search is then formulated as:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 185, + 652, + 480, + 667 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 185, + 652, + 480, + 667 + ], + "spans": [ + { + "bbox": [ + 185, + 652, + 480, + 667 + ], + "type": "interline_equation", + "content": "\\hat {\\mathbf {x}} _ {i} ^ {\\mathrm {3 D}} = \\operatorname {c a t} \\left( \\right.\\left\\{ \\right.\\mathbf {T} _ {j \\rightarrow i} * \\mathbf {x} _ {j} ^ {\\mathrm {3 D}} \\left. \\right| | j - i | \\leq w _ {t} \\}\\left. \\right), \\quad \\mathbf {T} _ {j \\rightarrow i} = \\mathbf {T} _ {i} ^ {- 1} \\mathbf {T} _ {j}, \\tag {2}", + "image_path": "e3b81e668de5d8129a2c8d4066ea00b82f4a90b013a7f3f4b27c9321dc6e6c46.jpg" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "type": "text", + "content": "H. Cao et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 165 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 165 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 165 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 116, + 482, + 165 + ], + "type": "inline_equation", + "content": "\\mathbf{T}_i" + }, + { + "bbox": [ + 130, + 116, + 482, + 165 + ], + "type": "text", + "content": " is the estimated pose at frame " + }, + { + "bbox": [ + 130, + 116, + 482, + 165 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 116, + 482, + 165 + ], + "type": "text", + "content": " and so as " + }, + { + "bbox": [ + 130, + 116, + 482, + 165 + ], + "type": "inline_equation", + "content": "\\mathbf{T}_j" + }, + { + "bbox": [ + 130, + 116, + 482, + 165 + ], + "type": "text", + "content": ", which can be easily obtained online through off-the-shelf SLAM algorithms [17,36]. " + }, + { + "bbox": [ + 130, + 116, + 482, + 165 + ], + "type": "inline_equation", + "content": "\\operatorname{cat}(\\cdot)" + }, + { + "bbox": [ + 130, + 116, + 482, + 165 + ], + "type": "text", + "content": " is the concatenation operation along the point number dimension while " + }, + { + "bbox": [ + 130, + 116, + 482, + 165 + ], + "type": "inline_equation", + "content": "*" + }, + { + "bbox": [ + 130, + 116, + 482, + 165 + ], + "type": "text", + "content": " denotes the pose transformation. The voxelization is then performed on " + }, + { + "bbox": [ + 130, + 116, + 482, + 165 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{x}}_i^{3D}" + }, + { + "bbox": [ + 130, + 116, + 482, + 165 + ], + "type": "text", + "content": ", formulated as:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 195, + 171, + 480, + 186 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 195, + 171, + 480, + 186 + ], + "spans": [ + { + "bbox": [ + 195, + 171, + 480, + 186 + ], + "type": "interline_equation", + "content": "\\mathbf {v} _ {i} ^ {\\mathrm {3 D}} = \\mathcal {V} _ {\\mathbf {s}} \\left(\\hat {\\mathbf {x}} _ {i} ^ {\\mathrm {3 D}}\\right), \\quad \\mathbf {v} _ {i} ^ {\\mathrm {3 D}} \\in \\mathbb {R} ^ {N _ {\\mathrm {v}} \\times 3}, \\tag {3}", + "image_path": "cbe99ef0870af85f3e2664175baf856c43f155607c7dc78adcd283284180b551.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 190, + 189, + 480, + 205 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 190, + 189, + 480, + 205 + ], + "spans": [ + { + "bbox": [ + 190, + 189, + 480, + 205 + ], + "type": "interline_equation", + "content": "\\mathbf {M} _ {i} ^ {\\mathrm {3 D}} = \\left\\{k | \\forall g \\in [ 1, \\hat {N} ], \\left\\lfloor \\hat {\\mathbf {x}} _ {i, g} ^ {\\mathrm {3 D}} / \\mathbf {s} \\right\\rfloor = \\mathbf {v} _ {i, k} ^ {\\mathrm {3 D}} \\right\\}, \\quad \\mathbf {M} _ {i} ^ {\\mathrm {3 D}} \\in \\mathbb {R} ^ {\\hat {N}}, \\tag {4}", + "image_path": "f57093ab6efc273b1c81292327ee6c63a153dfd69f5345649da7cc694ae294ef.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 213, + 482, + 321 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 213, + 482, + 321 + ], + "spans": [ + { + "bbox": [ + 131, + 213, + 482, + 321 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 213, + 482, + 321 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_i^{3\\mathrm{D}}" + }, + { + "bbox": [ + 131, + 213, + 482, + 321 + ], + "type": "text", + "content": " is the extracted voxels and " + }, + { + "bbox": [ + 131, + 213, + 482, + 321 + ], + "type": "inline_equation", + "content": "\\nu_{s}" + }, + { + "bbox": [ + 131, + 213, + 482, + 321 + ], + "type": "text", + "content": " is the voxelization operation with a voxel size s. After voxelization, points located in the same voxel can be regarded as correspondences since they are geometric neighborhoods, where the hash table " + }, + { + "bbox": [ + 131, + 213, + 482, + 321 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_i^{3\\mathrm{D}}" + }, + { + "bbox": [ + 131, + 213, + 482, + 321 + ], + "type": "text", + "content": " that maps each voxel back to its containing points (e.g., " + }, + { + "bbox": [ + 131, + 213, + 482, + 321 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_{i,k}^{3\\mathrm{D}}" + }, + { + "bbox": [ + 131, + 213, + 482, + 321 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 131, + 213, + 482, + 321 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{x}}_{i,g}^{3\\mathrm{D}}" + }, + { + "bbox": [ + 131, + 213, + 482, + 321 + ], + "type": "text", + "content": " as in Eq. (4)) is preserved for the subsequent correspondence mapping. In this way, it provides more diverse neighborhood selection strategies with temporally local correspondences for consistency and certainty evaluation compared to globally merging all frames in one single step. In practice, the slide window aggregation and voxelization are conducted iteratively for each frame in the input batch." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 338, + 350, + 350 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 338, + 350, + 350 + ], + "spans": [ + { + "bbox": [ + 131, + 338, + 350, + 350 + ], + "type": "text", + "content": "3.3 Spatial-Temporal Voxels and Entropy" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 356, + 482, + 487 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 356, + 482, + 487 + ], + "spans": [ + { + "bbox": [ + 130, + 356, + 482, + 487 + ], + "type": "text", + "content": "Given the geometric correspondences captured by voxelization, our goal is to emphasize the predictions with correspondence consistency within each voxel. After slide window aggregation and voxelization, there inevitably exist some unreliable correspondences of two different types: (i) voxels that contain different types of semantic objects (e.g., those located on the contact surface between cars and roads) and (ii) voxels with highly inconsistent predictions due to uncertain predictions across time. These unreliable correspondences contain high uncertainty and inconsistency, which could harm the online optimization of models if they as references. To alleviate the side-effect introduced by such unreliable correspondences, we propose ST voxels that leverage ST entropy for voxel-wise reliability evaluation and cross-modal attending." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 488, + 482, + 596 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 488, + 482, + 596 + ], + "spans": [ + { + "bbox": [ + 130, + 488, + 482, + 596 + ], + "type": "text", + "content": "Specifically, an ST voxel contains two components, including the query and the reference, where the query is encouraged to be consistent with the reference. Considering that the moving average teacher can usually generate more stable predictions [4, 41], we regard the single-frame student predictions as the query while multi-frame teacher predictions as our reference, forming ST voxels denoted as " + }, + { + "bbox": [ + 130, + 488, + 482, + 596 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_i^{\\mathrm{ST}}" + }, + { + "bbox": [ + 130, + 488, + 482, + 596 + ], + "type": "text", + "content": ". Without losing generality, here we take a single ST voxel indexed by " + }, + { + "bbox": [ + 130, + 488, + 482, + 596 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 488, + 482, + 596 + ], + "type": "text", + "content": " in frame " + }, + { + "bbox": [ + 130, + 488, + 482, + 596 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 488, + 482, + 596 + ], + "type": "text", + "content": " as an example, denoted as " + }, + { + "bbox": [ + 130, + 488, + 482, + 596 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_{i,k}^{\\mathrm{ST}} = \\{\\mathbf{p}_{\\mathrm{q}}^{\\mathrm{m}},\\mathbf{p}_{\\mathrm{r}}^{\\mathrm{m}}\\}" + }, + { + "bbox": [ + 130, + 488, + 482, + 596 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 488, + 482, + 596 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_{\\mathrm{q}}^{\\mathrm{m}}" + }, + { + "bbox": [ + 130, + 488, + 482, + 596 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 488, + 482, + 596 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_{\\mathrm{r}}^{\\mathrm{m}}" + }, + { + "bbox": [ + 130, + 488, + 482, + 596 + ], + "type": "text", + "content": " are the point-wise student predictions at the query frame " + }, + { + "bbox": [ + 130, + 488, + 482, + 596 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 488, + 482, + 596 + ], + "type": "text", + "content": " and teacher predictions in the frame searching range " + }, + { + "bbox": [ + 130, + 488, + 482, + 596 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 130, + 488, + 482, + 596 + ], + "type": "text", + "content": ", respectively, formulated as:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 186, + 603, + 480, + 618 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 186, + 603, + 480, + 618 + ], + "spans": [ + { + "bbox": [ + 186, + 603, + 480, + 618 + ], + "type": "interline_equation", + "content": "\\mathbf {p} _ {\\mathrm {q}} ^ {\\mathrm {m}} = \\left\\{\\mathbf {p} _ {i, g} ^ {\\mathrm {m}} \\mid \\left\\lfloor \\mathbf {x} _ {i, g} ^ {\\mathrm {3 D}} / \\mathbf {s} \\right\\rfloor = \\mathbf {v} _ {i, k} ^ {\\mathrm {3 D}} \\right\\}, \\quad \\mathbf {p} _ {\\mathrm {q}} ^ {\\mathrm {m}} \\in \\mathbb {R} ^ {N _ {\\mathrm {q}} \\times K} \\tag {5}", + "image_path": "455d6da4c0a7dcd23822338e1dba625b67dc2a36bbbbfce56483a6905b52953f.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 186, + 620, + 480, + 635 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 186, + 620, + 480, + 635 + ], + "spans": [ + { + "bbox": [ + 186, + 620, + 480, + 635 + ], + "type": "interline_equation", + "content": "\\mathbf {p} _ {\\mathrm {r}} ^ {\\mathrm {m}} = \\left\\{\\tilde {\\mathbf {p}} _ {j, g} ^ {\\mathrm {m}} \\mid \\left| \\mathbf {x} _ {j, g} ^ {3 \\mathrm {D}} / \\mathbf {s} \\right| = \\mathbf {v} _ {i, k} ^ {3 \\mathrm {D}}, | j - i | \\leq w _ {t} \\right\\}, \\quad \\mathbf {p} _ {\\mathrm {r}} ^ {\\mathrm {m}} \\in \\mathbb {R} ^ {N _ {\\mathrm {r}} \\times K}. \\tag {6}", + "image_path": "d920e3b4ffeb9305ec97dfa038d192994a24b52d97ba683035cc880364a9ed16.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 131, + 641, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 641, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 131, + 641, + 482, + 665 + ], + "type": "text", + "content": "The reliability of each voxel is then evaluated as the Shannon's entropy [42] of the average teacher predictions, where the unreliable ST voxels are then filtered" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 148, + 91, + 448, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 148, + 91, + 448, + 103 + ], + "spans": [ + { + "bbox": [ + 148, + 91, + 448, + 103 + ], + "type": "text", + "content": "Reliable Spatial-Temporal Voxels For Multi-Modal Test-Time Adaptation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 392, + 129 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 392, + 129 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 392, + 129 + ], + "type": "text", + "content": "out given a pre-defined quantile " + }, + { + "bbox": [ + 130, + 116, + 392, + 129 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 130, + 116, + 392, + 129 + ], + "type": "text", + "content": " as in Fig. 2, denoted as:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 202, + 136, + 481, + 167 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 202, + 136, + 481, + 167 + ], + "spans": [ + { + "bbox": [ + 202, + 136, + 481, + 167 + ], + "type": "interline_equation", + "content": "E _ {i, k} ^ {\\mathrm {m}} = - \\sum_ {c} ^ {K} \\bar {\\mathbf {p}} _ {\\mathrm {r}, c} ^ {\\mathrm {m}} \\log \\bar {\\mathbf {p}} _ {\\mathrm {r}, c} ^ {\\mathrm {m}}, \\quad \\bar {\\mathbf {p}} _ {\\mathrm {r}} ^ {\\mathrm {m}} = \\sum_ {n = 1} ^ {N _ {\\mathrm {r}}} \\psi \\left(\\mathbf {p} _ {\\mathrm {r}, n} ^ {\\mathrm {m}}\\right) / N _ {\\mathrm {r}}, \\tag {7}", + "image_path": "3140b0852d5be0741612e0c57f16b383c1e82405547bdcfd2fc87a00cabb1c03.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 203, + 170, + 481, + 201 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 203, + 170, + 481, + 201 + ], + "spans": [ + { + "bbox": [ + 203, + 170, + 481, + 201 + ], + "type": "interline_equation", + "content": "h _ {i, k} = \\left\\{ \\begin{array}{l l} 0, & \\text {i f} E _ {i, k} ^ {m} > Q ^ {m} (\\alpha) \\\\ 1, & \\text {i f} E _ {i, k} ^ {m} \\leq Q ^ {m} (\\alpha) \\end{array} , \\right. \\tag {8}", + "image_path": "172edcf5a8e47efbe1c41152b3787893182743a3b1b1d00243d5e84808885584.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 206, + 482, + 255 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 206, + 482, + 255 + ], + "spans": [ + { + "bbox": [ + 131, + 206, + 482, + 255 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 206, + 482, + 255 + ], + "type": "inline_equation", + "content": "E_{i,k}^{\\mathrm{m}}" + }, + { + "bbox": [ + 131, + 206, + 482, + 255 + ], + "type": "text", + "content": " is regarded as the ST entropy of voxel " + }, + { + "bbox": [ + 131, + 206, + 482, + 255 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 131, + 206, + 482, + 255 + ], + "type": "text", + "content": " for modality " + }, + { + "bbox": [ + 131, + 206, + 482, + 255 + ], + "type": "inline_equation", + "content": "\\mathrm{m}" + }, + { + "bbox": [ + 131, + 206, + 482, + 255 + ], + "type": "text", + "content": ", indicating the modal-specific prediction reliability within this voxel. " + }, + { + "bbox": [ + 131, + 206, + 482, + 255 + ], + "type": "inline_equation", + "content": "Q^{m}(\\alpha)" + }, + { + "bbox": [ + 131, + 206, + 482, + 255 + ], + "type": "text", + "content": " is the " + }, + { + "bbox": [ + 131, + 206, + 482, + 255 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 131, + 206, + 482, + 255 + ], + "type": "text", + "content": "-quantile of all ST entropy in modality " + }, + { + "bbox": [ + 131, + 206, + 482, + 255 + ], + "type": "inline_equation", + "content": "\\mathrm{m}" + }, + { + "bbox": [ + 131, + 206, + 482, + 255 + ], + "type": "text", + "content": " in the same batch and " + }, + { + "bbox": [ + 131, + 206, + 482, + 255 + ], + "type": "inline_equation", + "content": "\\psi(\\cdot)" + }, + { + "bbox": [ + 131, + 206, + 482, + 255 + ], + "type": "text", + "content": " denotes the Softmax function along the class dimension." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 270, + 357, + 284 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 270, + 357, + 284 + ], + "spans": [ + { + "bbox": [ + 131, + 270, + 357, + 284 + ], + "type": "text", + "content": "3.4 ST Voxel Aided Cross-Modal Learning" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 289, + 482, + 386 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 289, + 482, + 386 + ], + "spans": [ + { + "bbox": [ + 130, + 289, + 482, + 386 + ], + "type": "text", + "content": "Each modality has its pros and cons under different conditions and cross-modal learning for online scenarios like MM-TTA should therefore possess an adaptive mechanism that can attend to \"pros\" and meanwhile suppress the \"cons\". To achieve that, Latte performs cross-modal learning by attending to the modality with more consistent and certain predictions from a spatial-temporal perspective. Specifically, taking ST voxel " + }, + { + "bbox": [ + 130, + 289, + 482, + 386 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_{i,k}^{\\mathrm{ST}}" + }, + { + "bbox": [ + 130, + 289, + 482, + 386 + ], + "type": "text", + "content": " and its corresponding ST entropy " + }, + { + "bbox": [ + 130, + 289, + 482, + 386 + ], + "type": "inline_equation", + "content": "E_{i,k}^{\\mathrm{m}}" + }, + { + "bbox": [ + 130, + 289, + 482, + 386 + ], + "type": "text", + "content": " as an example, the voxel-wise cross-modal attending weights and the weighted cross-modal consistency loss " + }, + { + "bbox": [ + 130, + 289, + 482, + 386 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{i,k}^{\\mathrm{xM}}" + }, + { + "bbox": [ + 130, + 289, + 482, + 386 + ], + "type": "text", + "content": " are formulated as:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 200, + 393, + 481, + 422 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 200, + 393, + 481, + 422 + ], + "spans": [ + { + "bbox": [ + 200, + 393, + 481, + 422 + ], + "type": "interline_equation", + "content": "w _ {\\mathrm {v}} ^ {\\mathrm {2 D}} = \\frac {\\exp \\left(E _ {i , k} ^ {\\mathrm {2 D}}\\right)}{\\exp \\left(E _ {i , k} ^ {\\mathrm {2 D}}\\right) + \\exp \\left(E _ {i , k} ^ {\\mathrm {3 D}}\\right)}, w _ {\\mathrm {v}} ^ {\\mathrm {3 D}} = 1 - w _ {\\mathrm {v}} ^ {\\mathrm {2 D}}, \\tag {9}", + "image_path": "d9f025e4d78aeb822fa73a3e5813584d1fe4b2719ac4ea9fabe01798582a9add.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 200, + 424, + 480, + 439 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 200, + 424, + 480, + 439 + ], + "spans": [ + { + "bbox": [ + 200, + 424, + 480, + 439 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {i, k} ^ {\\mathrm {x M}} = w _ {\\mathrm {v}} ^ {\\mathrm {2 D}} D _ {\\mathrm {K L}} \\left(\\bar {\\mathbf {p}} _ {\\mathrm {q}} ^ {\\mathrm {3 D}} \\| \\bar {\\mathbf {p}} _ {\\mathrm {r}} ^ {\\mathrm {2 D}}\\right) + w _ {\\mathrm {v}} ^ {\\mathrm {3 D}} D _ {\\mathrm {K L}} \\left(\\bar {\\mathbf {p}} _ {\\mathrm {q}} ^ {\\mathrm {2 D}} \\| \\bar {\\mathbf {p}} _ {\\mathrm {r}} ^ {\\mathrm {3 D}}\\right), \\tag {10}", + "image_path": "01f130fd3ceed60a7086abb55ac385c010a82d1e3239bb0f990e82c47b07b277.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 445, + 482, + 470 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 445, + 482, + 470 + ], + "spans": [ + { + "bbox": [ + 130, + 445, + 482, + 470 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 445, + 482, + 470 + ], + "type": "inline_equation", + "content": "D_{\\mathrm{KL}}(\\cdot)" + }, + { + "bbox": [ + 130, + 445, + 482, + 470 + ], + "type": "text", + "content": " represents the KL-divergence between two probability. " + }, + { + "bbox": [ + 130, + 445, + 482, + 470 + ], + "type": "inline_equation", + "content": "\\bar{\\mathbf{p}}_{\\mathrm{q}}^{3\\mathrm{D}}" + }, + { + "bbox": [ + 130, + 445, + 482, + 470 + ], + "type": "text", + "content": " is the average query predictions in the ST voxel, similarly computed as in Eq. (7)." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 471, + 482, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 471, + 482, + 555 + ], + "spans": [ + { + "bbox": [ + 130, + 471, + 482, + 555 + ], + "type": "text", + "content": "The cross-modal attending is further extended from the voxel level to the point level for better online predictions and cross-modal pseudo-label generation. Specifically, the ST entropy is propagated from the ST voxel to its reference point-wise prediction, indicating its confidence and consistency within its spatial-temporal neighborhoods. If one's ST entropy has been filtered out as in Eq. (8), its ST entropy would fall back to the point-level entropy. For arbitrary pointwise predictions " + }, + { + "bbox": [ + 130, + 471, + 482, + 555 + ], + "type": "inline_equation", + "content": "\\mathbf{p}^{\\mathrm{m}}\\in \\mathbf{p}_{\\mathrm{r}}^{\\mathrm{m}}" + }, + { + "bbox": [ + 130, + 471, + 482, + 555 + ], + "type": "text", + "content": ", the cross-modal predictions " + }, + { + "bbox": [ + 130, + 471, + 482, + 555 + ], + "type": "inline_equation", + "content": "y^{\\mathrm{xM}}" + }, + { + "bbox": [ + 130, + 471, + 482, + 555 + ], + "type": "text", + "content": " are formulated as:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 205, + 571, + 481, + 602 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 205, + 571, + 481, + 602 + ], + "spans": [ + { + "bbox": [ + 205, + 571, + 481, + 602 + ], + "type": "interline_equation", + "content": "\\hat {E} _ {\\mathrm {r}} ^ {\\mathrm {m}} = h _ {i, k} E _ {i, k} ^ {\\mathrm {m}} - \\left(1 - h _ {i, k}\\right) \\sum_ {c} ^ {K} \\mathbf {p} _ {c} ^ {\\mathrm {m}} \\log \\mathbf {p} _ {c} ^ {\\mathrm {m}}, \\tag {11}", + "image_path": "4619a67bd474df73380f2463d4593d200b039018addbadc29966f7d0203be843.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 201, + 605, + 480, + 634 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 201, + 605, + 480, + 634 + ], + "spans": [ + { + "bbox": [ + 201, + 605, + 480, + 634 + ], + "type": "interline_equation", + "content": "w _ {\\mathrm {p}} ^ {\\mathrm {2 D}} = \\frac {\\exp \\left(\\hat {E} ^ {\\mathrm {2 D}}\\right)}{\\exp \\left(\\hat {E} ^ {\\mathrm {2 D}}\\right) + \\exp \\left(\\hat {E} ^ {\\mathrm {3 D}}\\right)}, w _ {\\mathrm {p}} ^ {\\mathrm {3 D}} = 1 - w _ {\\mathrm {p}} ^ {\\mathrm {2 D}}, \\tag {12}", + "image_path": "c854396e6e7415469d2f9c7ee0e96b8e0f79e20314ad607024d3d7dc869049ad.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 201, + 635, + 480, + 655 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 201, + 635, + 480, + 655 + ], + "spans": [ + { + "bbox": [ + 201, + 635, + 480, + 655 + ], + "type": "interline_equation", + "content": "\\mathbf {p} ^ {\\mathrm {x M}} = w _ {\\mathrm {p}} ^ {\\mathrm {2 D}} \\mathbf {p} ^ {\\mathrm {2 D}} + w _ {\\mathrm {p}} ^ {\\mathrm {3 D}} \\mathbf {p} ^ {\\mathrm {3 D}}, y ^ {\\mathrm {x M}} = \\underset {c} {\\arg \\max } \\mathbf {p} ^ {\\mathrm {x M}}. \\tag {13}", + "image_path": "9de4a78af726f5742e4383b32bdb2acbce714ee814e17fd861c297781dc14976.jpg" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "type": "text", + "content": "H. Cao et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "code", + "bbox": [ + 130, + 132, + 455, + 273 + ], + "blocks": [ + { + "bbox": [ + 139, + 118, + 433, + 130 + ], + "lines": [ + { + "bbox": [ + 139, + 118, + 433, + 130 + ], + "spans": [ + { + "bbox": [ + 139, + 118, + 433, + 130 + ], + "type": "text", + "content": "Algorithm 1: Adaptation and Online Prediction Process of Latte" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "code_caption" + }, + { + "bbox": [ + 130, + 132, + 455, + 273 + ], + "lines": [ + { + "bbox": [ + 130, + 132, + 455, + 273 + ], + "spans": [ + { + "bbox": [ + 130, + 132, + 455, + 273 + ], + "type": "text", + "content": "Target Domain: Multi-modal input " + }, + { + "bbox": [ + 130, + 132, + 455, + 273 + ], + "type": "inline_equation", + "content": "\\mathcal{X}_{\\mathcal{T}} = \\{(\\mathbf{x}_{t}^{2\\mathrm{D}},\\mathbf{x}_{t}^{3\\mathrm{D}})\\}" + }, + { + "bbox": [ + 130, + 132, + 455, + 273 + ], + "type": "text", + "content": " \nInit Model: Teacher " + }, + { + "bbox": [ + 130, + 132, + 455, + 273 + ], + "type": "inline_equation", + "content": "\\tilde{\\phi}_t^{\\mathrm{m}}" + }, + { + "bbox": [ + 130, + 132, + 455, + 273 + ], + "type": "text", + "content": " , student " + }, + { + "bbox": [ + 130, + 132, + 455, + 273 + ], + "type": "inline_equation", + "content": "\\phi_t^{\\mathrm{m}}" + }, + { + "bbox": [ + 130, + 132, + 455, + 273 + ], + "type": "inline_equation", + "content": "\\mathfrak{m}\\in \\{2\\mathrm{D},3\\mathrm{D}\\}" + }, + { + "bbox": [ + 130, + 132, + 455, + 273 + ], + "type": "text", + "content": " , Slide window size " + }, + { + "bbox": [ + 130, + 132, + 455, + 273 + ], + "type": "inline_equation", + "content": "w_{t}" + }, + { + "bbox": [ + 130, + 132, + 455, + 273 + ], + "type": "text", + "content": " \nfor " + }, + { + "bbox": [ + 130, + 132, + 455, + 273 + ], + "type": "inline_equation", + "content": "\\pmb {x} = \\{(x_{t}^{2D},x_{t}^{3D})\\}_{t = t_0 + 1}^B\\in \\mathcal{X}_T" + }, + { + "bbox": [ + 130, + 132, + 455, + 273 + ], + "type": "text", + "content": " do 1. Get student and teacher predictions " + }, + { + "bbox": [ + 130, + 132, + 455, + 273 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{p}}_t^{\\mathrm{m}},\\mathbf{p}_t^{\\mathrm{m}}" + }, + { + "bbox": [ + 130, + 132, + 455, + 273 + ], + "type": "text", + "content": " by Eq. (1) for " + }, + { + "bbox": [ + 130, + 132, + 455, + 273 + ], + "type": "inline_equation", + "content": "i\\in [t_0 + 1,t_0 + B]" + }, + { + "bbox": [ + 130, + 132, + 455, + 273 + ], + "type": "text", + "content": " do 2. Get " + }, + { + "bbox": [ + 130, + 132, + 455, + 273 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{x}}_i^{3\\mathrm{D}},\\mathbf{v}_i^{3\\mathrm{D}},\\mathbf{M}_i^{3\\mathrm{D}}" + }, + { + "bbox": [ + 130, + 132, + 455, + 273 + ], + "type": "text", + "content": " within the time window of size " + }, + { + "bbox": [ + 130, + 132, + 455, + 273 + ], + "type": "inline_equation", + "content": "w_{t}" + }, + { + "bbox": [ + 130, + 132, + 455, + 273 + ], + "type": "text", + "content": " by Eq. (2)-(4) 3. Compute ST voxels " + }, + { + "bbox": [ + 130, + 132, + 455, + 273 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_i^{\\mathrm{ST}}" + }, + { + "bbox": [ + 130, + 132, + 455, + 273 + ], + "type": "text", + "content": " and ST entropy " + }, + { + "bbox": [ + 130, + 132, + 455, + 273 + ], + "type": "inline_equation", + "content": "E_i^{\\mathrm{m}}" + }, + { + "bbox": [ + 130, + 132, + 455, + 273 + ], + "type": "text", + "content": " by Eq. (5)-(8) 4. Compute " + }, + { + "bbox": [ + 130, + 132, + 455, + 273 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_i^{\\mathrm{xM}}" + }, + { + "bbox": [ + 130, + 132, + 455, + 273 + ], + "type": "text", + "content": " by Eq. (9)-(10) end 5. Compute " + }, + { + "bbox": [ + 130, + 132, + 455, + 273 + ], + "type": "inline_equation", + "content": "\\hat{E}_{\\mathrm{r}}^{\\mathrm{m}}" + }, + { + "bbox": [ + 130, + 132, + 455, + 273 + ], + "type": "text", + "content": " and get " + }, + { + "bbox": [ + 130, + 132, + 455, + 273 + ], + "type": "inline_equation", + "content": "\\mathbf{y}_t^{\\mathrm{xM}}" + }, + { + "bbox": [ + 130, + 132, + 455, + 273 + ], + "type": "text", + "content": " as predictions by Eq. (11)-(13) 6. Compute overall loss by Eq. (14) and update models by Eq. (15) \nend" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "code_body" + } + ], + "index": 3, + "sub_type": "algorithm" + }, + { + "bbox": [ + 132, + 284, + 346, + 296 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 284, + 346, + 296 + ], + "spans": [ + { + "bbox": [ + 132, + 284, + 346, + 296 + ], + "type": "text", + "content": "3.5 Online Predictions and Optimization" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 310, + 482, + 419 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 310, + 482, + 419 + ], + "spans": [ + { + "bbox": [ + 130, + 310, + 482, + 419 + ], + "type": "text", + "content": "As shown in Fig. 2, given a batch of " + }, + { + "bbox": [ + 130, + 310, + 482, + 419 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 130, + 310, + 482, + 419 + ], + "type": "text", + "content": " consecutive frames " + }, + { + "bbox": [ + 130, + 310, + 482, + 419 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{x}_t^{\\mathrm{m}}\\}_{t = t_0 + 1}^{t_0 + B}" + }, + { + "bbox": [ + 130, + 310, + 482, + 419 + ], + "type": "text", + "content": " as input, we first extract the student and teacher predictions through Eq. (1). Subsequently, frames are aggregated and voxelized as in Eq. (2), resulting in a batch of merged point cloud " + }, + { + "bbox": [ + 130, + 310, + 482, + 419 + ], + "type": "inline_equation", + "content": "\\{\\hat{\\mathbf{x}}_t^{3\\mathrm{D}}\\}_{t = t_0 + 1}^{t_0 + B}" + }, + { + "bbox": [ + 130, + 310, + 482, + 419 + ], + "type": "text", + "content": " and voxels " + }, + { + "bbox": [ + 130, + 310, + 482, + 419 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{v}_t^{3\\mathrm{D}}\\}_{t = t_0 + 1}^{t_0 + B}" + }, + { + "bbox": [ + 130, + 310, + 482, + 419 + ], + "type": "text", + "content": ". The ST voxels and ST entropy computation process is then applied to each frame of the merged point cloud and voxels. Due to the overlap of the slide window, multiple ST entropy values could be propagated to the same point prediction, where we empirically take the average value of all ST entropy received to proceed Eq. (11). The overall loss function and the updating scheme can be formulated as:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 219, + 430, + 481, + 458 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 219, + 430, + 481, + 458 + ], + "spans": [ + { + "bbox": [ + 219, + 430, + 481, + 458 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\sum_ {t} \\mathcal {F} \\left(\\mathbf {p} _ {t} ^ {\\mathrm {m}}, \\mathbf {y} _ {t} ^ {\\mathrm {x M}}\\right) + \\frac {\\lambda_ {\\mathrm {x M}}}{B} \\sum_ {t} \\sum_ {k} \\mathcal {L} _ {t, k} ^ {\\mathrm {x M}}, \\tag {14}", + "image_path": "777ec1e3630df748881c0299b97057d9f32a18331780afb9a9fdfbf7d12f19f4.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 214, + 460, + 480, + 475 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 214, + 460, + 480, + 475 + ], + "spans": [ + { + "bbox": [ + 214, + 460, + 480, + 475 + ], + "type": "interline_equation", + "content": "\\tilde {\\theta} _ {t} ^ {\\mathrm {m}} = \\lambda_ {s} \\tilde {\\theta} _ {t - 1} ^ {\\mathrm {m}} + (1 - \\lambda_ {s}) \\theta_ {t} ^ {\\mathrm {m}}, \\tag {15}", + "image_path": "24e73152bd6aee08043ce9ff2878e599b0bee97f5c7996c3c27d1c64be62e0aa.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 487, + 482, + 561 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 487, + 482, + 561 + ], + "spans": [ + { + "bbox": [ + 130, + 487, + 482, + 561 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 487, + 482, + 561 + ], + "type": "inline_equation", + "content": "\\mathcal{F}(\\cdot)" + }, + { + "bbox": [ + 130, + 487, + 482, + 561 + ], + "type": "text", + "content": " is the cross-entropy function and " + }, + { + "bbox": [ + 130, + 487, + 482, + 561 + ], + "type": "inline_equation", + "content": "\\mathbf{y}_t^{\\mathrm{xM}}" + }, + { + "bbox": [ + 130, + 487, + 482, + 561 + ], + "type": "text", + "content": " is the point-wise cross-modal pseudo-label frame from Eq. (13), which is also regarded as our cross-modal prediction for evaluation. " + }, + { + "bbox": [ + 130, + 487, + 482, + 561 + ], + "type": "inline_equation", + "content": "\\tilde{\\theta}_t^{\\mathrm{m}}" + }, + { + "bbox": [ + 130, + 487, + 482, + 561 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 487, + 482, + 561 + ], + "type": "inline_equation", + "content": "\\theta_t^{\\mathrm{m}}" + }, + { + "bbox": [ + 130, + 487, + 482, + 561 + ], + "type": "text", + "content": " are the model parameters of teacher and student models, respectively, where " + }, + { + "bbox": [ + 130, + 487, + 482, + 561 + ], + "type": "inline_equation", + "content": "\\lambda_s" + }, + { + "bbox": [ + 130, + 487, + 482, + 561 + ], + "type": "text", + "content": " is the momentum update coefficient. " + }, + { + "bbox": [ + 130, + 487, + 482, + 561 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{xM}}" + }, + { + "bbox": [ + 130, + 487, + 482, + 561 + ], + "type": "text", + "content": " is pre-defined coefficient for the cross-modal consistency loss. Our adaptation and inference process is summarized in Algo. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 585, + 282, + 599 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 585, + 282, + 599 + ], + "spans": [ + { + "bbox": [ + 132, + 585, + 282, + 599 + ], + "type": "text", + "content": "4 Experimental Results" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 617, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 665 + ], + "type": "text", + "content": "In this section, we present our thorough experimental results on three different benchmarks. Details of benchmarks, backbones, and settings are first present in Sec. 4.1. The main results are then illustrated in Sec. 4.2, followed by detailed ablation studies and qualitative results in Sec. 4.3." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 148, + 90, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 148, + 90, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 148, + 90, + 447, + 102 + ], + "type": "text", + "content": "Reliable Spatial-Temporal Voxels For Multi-Modal Test-Time Adaptation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 288, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 288, + 129 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 288, + 129 + ], + "type": "text", + "content": "4.1 Benchmarks and settings" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 138, + 482, + 341 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 138, + 482, + 341 + ], + "spans": [ + { + "bbox": [ + 130, + 138, + 482, + 341 + ], + "type": "text", + "content": "Benchmark Details. To thoroughly investigate the effectiveness of Latte, we conduct our experiments on three different MM-TTA benchmarks: (i) USA-to-Singapore (U-to-S), (ii) A2D2-to-SemanticKITTI (A-to-S), and (iii) Synthia-to-SemanticKITTI (S-to-S). Specifically, U-to-S is developed from NuScenes-LiDARSeg [2] dataset, where the domain gap is mainly credited to the infrastructure difference between countries. Same as [16], the source domain and the target domain data are selected by filtering country keywords based on the data recording description in NuScenes. Different from previous UDA methods [16,24] utilizing self-defined class mappings, we follow the official mapping in NuScenes-LiDARSeg, forming a 16-class segmentation benchmark. The remaining two benchmarks follow a similar setting as in [29], where the domain gap of A-to-S [1,10] lies in different LiDAR mounting positions and image resolutions while the one of S-to-S [1] lies in different patterns between synthetic and real data for both modalities. For A-to-S, we adopt the same class-mapping as in [16,29] which leads to a 10-class benchmark, while we design our class-mapping for S-to-S since its original class map has not been revealed previously [29], forming a 9-class benchmark. More details are presented in our appendix." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 342, + 482, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 342, + 482, + 484 + ], + "spans": [ + { + "bbox": [ + 130, + 342, + 482, + 484 + ], + "type": "text", + "content": "Baseline methods. Previous TTA and MM-TTA methods are included in comparison with Latte. For TTA methods, we compare Latte with point-wise pseudolabels with filtering (PsLabel), TENT [39], ETA [22], SAR [23] with and without restoring (denoted as SAR-rs and SAR, respectively). MMTTA [29] as well as its online version of xMUDA [16] (including pure xMUDA and pseudo-label version xMUDA+PL) are additionally included as the previous SOTA methods for comparison. Considering that EATA [22] with Fisher regularizer requires pre-access to the target domain samples, we compare Latte with ETA, an official variant of EATA without this regularizer for a fair comparison. We further present a performance lower bound by directly testing with source pre-trained models (i.e., Source only) and an upper bound by online adapting networks with ground-truth labels from the target domain for one epoch (i.e., Oracle TTA)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "type": "text", + "content": "Implementation details. With the rapid progress in 2D and 3D segmentation, we employ all baseline methods and Latte with more recent SegFormer [43] (SegFormer-B1) and SPVCNN [33] (SPVCNN-cr1) as their 2D and 3D backbones, respectively, to investigate the improvement of existing TTA methods when combined with advanced networks. The pre-training procedures on the source domain follow a similar setting as in [16], except for 2D SegFormer which is trained with a base learning rate of 6E-5 with an AdamW optimizer and Poly scheduler as in [43]. For Latte, we maintain the same learning rate and optimizer as in the pre-training stage, while the scheduler is disabled. For our voxelization, we leverage the off-the-shelf SLAM algorithm KISS-ICP [36] to generate poses and utilize a window size " + }, + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "type": "inline_equation", + "content": "w_{t}" + }, + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "type": "text", + "content": " of 3 with a voxel size of " + }, + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "type": "inline_equation", + "content": "0.2\\mathrm{m}" + }, + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "type": "text", + "content": ", while the quantile " + }, + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "type": "text", + "content": " and coefficient " + }, + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{xM}}" + }, + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "type": "text", + "content": " are set to 0.9 and 0.3, respectively. " + }, + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\lambda_{s}" + }, + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "type": "text", + "content": " is empirically set to 0.99 as in [4,31]. For all methods, we strictly follow the one-pass protocol to first evaluate the predictions of the input batch and then update the networks, where we only update the trainable parameters in normalization layers. For all base" + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "type": "text", + "content": "H. Cao et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 135, + 180, + 478, + 289 + ], + "blocks": [ + { + "bbox": [ + 130, + 114, + 482, + 170 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 482, + 170 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 482, + 170 + ], + "type": "text", + "content": "Table 1: Performance (mIoU) comparison of Latte. Latte outperforms all previous SOTA methods on the cross-modal metric (xM) across three benchmarks. Here \"MM\" denotes whether the method contains multi-modal interaction or not. Cross-modal predictions are computed as the Softmax average of modal outputs except for Latte. \"Avg\" present the average xM performance across three benchmarks." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 135, + 180, + 478, + 289 + ], + "lines": [ + { + "bbox": [ + 135, + 180, + 478, + 289 + ], + "spans": [ + { + "bbox": [ + 135, + 180, + 478, + 289 + ], + "type": "table", + "html": "
MethodPublicationMMU-to-SxMA-to-SxMS-to-SAvg
2D3D2D3D2D3D
Source only-X31.441.143.947.417.944.323.436.448.2
Oracle TTA-X38.745.550.349.161.362.536.054.554.6
TENT [39]ICML-21X36.836.141.143.344.449.122.435.537.5
ETA [22]ICML-22X36.734.743.743.242.549.722.430.632.9
SAR [23]ICLR-23X36.637.243.943.042.750.320.731.332.9
SAR-rs [23]ICLR-23X36.635.543.943.143.850.124.225.128.8
xMUDA [16]PAMI-22✓19.422.924.213.133.032.212.214.814.1
xMUDA+PL [16]PAMI-22✓36.238.243.042.848.350.930.934.136.3
MMTTA [29]CVPR-22✓37.241.545.444.551.753.727.535.135.5
PsLabel-✓37.735.841.846.744.350.029.233.335.3
Latte (Ours)-✓37.441.046.046.152.654.333.239.341.6
", + "image_path": "2da61598510481053582ff88a330dba2fe43ac7a14fc57167eccd5491918fd25.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 296, + 483, + 344 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 296, + 483, + 344 + ], + "spans": [ + { + "bbox": [ + 130, + 296, + 483, + 344 + ], + "type": "text", + "content": "lines, we conduct a parameter search and report their best results. We utilize the mean intersection over union (mIoU) as our evaluation metric. All experiments are conducted with PyTorch on a single RTX 3090. More details are presented in our appendix." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 365, + 238, + 376 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 365, + 238, + 376 + ], + "spans": [ + { + "bbox": [ + 132, + 365, + 238, + 376 + ], + "type": "text", + "content": "4.2 Overall Results" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 389, + 482, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 389, + 482, + 498 + ], + "spans": [ + { + "bbox": [ + 130, + 389, + 482, + 498 + ], + "type": "text", + "content": "Tab. 1 presents the overall results of our Latte on three benchmarks compared with previous SOTA methods. In terms of average cross-modal predictions (Avg), Latte surpasses all previous TTA or MM-TTA methods, with a significant relative improvement of more than " + }, + { + "bbox": [ + 130, + 389, + 482, + 498 + ], + "type": "inline_equation", + "content": "5.3\\%" + }, + { + "bbox": [ + 130, + 389, + 482, + 498 + ], + "type": "text", + "content": ". For the less challenging U-to-S and A-to-S (i.e., gap between Oracle TTA and best TTA less than " + }, + { + "bbox": [ + 130, + 389, + 482, + 498 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 130, + 389, + 482, + 498 + ], + "type": "text", + "content": "), Latte achieves consistent relative improvements of " + }, + { + "bbox": [ + 130, + 389, + 482, + 498 + ], + "type": "inline_equation", + "content": "1.3\\%" + }, + { + "bbox": [ + 130, + 389, + 482, + 498 + ], + "type": "text", + "content": " on U-to-S and " + }, + { + "bbox": [ + 130, + 389, + 482, + 498 + ], + "type": "inline_equation", + "content": "1.1\\%" + }, + { + "bbox": [ + 130, + 389, + 482, + 498 + ], + "type": "text", + "content": " on A-to-S compared to MMTTA, respectively. Most existing methods except for xMUDA [16] can achieve superior performance compared to source-only results when one modality is reliable as in U-to-S (3D prevails) and A-to-S (2D prevails)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "text", + "content": "In terms of the most challenging S-to-S, Latte exhibits significant performance improvement compared to previous SOTA MM-TTA methods such as MMTTA [29] and xMUDA+PL [16], with a relative gap of " + }, + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "inline_equation", + "content": "14.6\\%" + }, + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "text", + "content": ". In fact, all previous methods except for our Latte perform inferiorly compared to source pre-trained models (Source only), due to the significant domain gaps caused by the non-trivial pattern difference in both modalities between synthetic and real data. In this case, the instability of single-frame predictions becomes exacerbated, leading to performance degeneration in MMTTA which mainly relies on single-frame refinement. This justifies the effectiveness of Latte under either challenging or less challenging TTA scenarios. Another interesting observation is that almost all cross-modal methods outperform other single-modal TTA methods, e.g., even the simplest PsLabel can outperform recent TTA methods like SAR [23]. This reveals the non-trivial benefits of cross-modal learning for TTA with segmentation. More comparisons about efficiency are included in appendix." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 148, + 90, + 447, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 148, + 90, + 447, + 103 + ], + "spans": [ + { + "bbox": [ + 148, + 90, + 447, + 103 + ], + "type": "text", + "content": "Reliable Spatial-Temporal Voxels For Multi-Modal Test-Time Adaptation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 141, + 180, + 470, + 277 + ], + "blocks": [ + { + "bbox": [ + 130, + 114, + 482, + 170 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 482, + 170 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 482, + 170 + ], + "type": "text", + "content": "Table 2: Ablation studies about the effectiveness of cross-modal learning. When disabling " + }, + { + "bbox": [ + 130, + 114, + 482, + 170 + ], + "type": "inline_equation", + "content": "\\mathbf{y}^{\\mathrm{xM}}" + }, + { + "bbox": [ + 130, + 114, + 482, + 170 + ], + "type": "text", + "content": ", we utilize the average Softmax of all modalities as the pseudo-labels as a replacement. In experiment No. 8, we replace the ST entropy (ST ety) with point-wise entropy in Eq. (13) and set " + }, + { + "bbox": [ + 130, + 114, + 482, + 170 + ], + "type": "inline_equation", + "content": "w_{\\mathrm{v}}^{\\mathrm{m}}" + }, + { + "bbox": [ + 130, + 114, + 482, + 170 + ], + "type": "text", + "content": " in Eq. (10) to 0.5. Here " + }, + { + "bbox": [ + 130, + 114, + 482, + 170 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{32}^{\\mathrm{xM}}" + }, + { + "bbox": [ + 130, + 114, + 482, + 170 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 114, + 482, + 170 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{23}^{\\mathrm{xM}}" + }, + { + "bbox": [ + 130, + 114, + 482, + 170 + ], + "type": "text", + "content": " refer to the former half and the latter half in Eq. 10, respectively." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 141, + 180, + 470, + 277 + ], + "lines": [ + { + "bbox": [ + 141, + 180, + 470, + 277 + ], + "spans": [ + { + "bbox": [ + 141, + 180, + 470, + 277 + ], + "type": "table", + "html": "
No.Description\\( y^{xM} \\)\\( \\mathcal{L}_{23}^{xM} \\)\\( \\mathcal{L}_{32}^{xM} \\)U-to-SA-to-SS-to-S
2D3DxM2D3DxM2D3DxM
0Source only---31.441.143.947.417.944.323.436.438.2
1only ST ety label✓36.834.543.743.342.249.922.430.633.3
2only ST 2D→3D✓32.534.541.041.742.248.227.730.632.9
3only ST 3D→2D✓36.839.642.643.350.248.422.426.725.6
4ST ety label + ST 2D→3D✓✓37.535.841.043.647.549.831.036.038.4
5ST ety label + ST 3D→2D✓✓37.440.945.946.252.153.432.538.739.7
6ST 2D↔3D✓✓36.840.343.844.047.051.230.435.637.7
7w/o percentile in Eq. (8)✓✓✓37.440.345.445.751.453.232.939.041.7
8Point ety vs. ST ety✓✓✓37.040.645.145.151.452.731.636.738.2
9Latte✓✓✓37.441.046.046.152.654.333.239.341.6
", + "image_path": "124df0ad9b6644ef2b50a453f9c3e161ebbea36301c3cd451f48609daf2f03fd.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 168, + 285, + 444, + 357 + ], + "blocks": [ + { + "bbox": [ + 168, + 285, + 444, + 357 + ], + "lines": [ + { + "bbox": [ + 168, + 285, + 444, + 357 + ], + "spans": [ + { + "bbox": [ + 168, + 285, + 444, + 357 + ], + "type": "image", + "image_path": "2e91f6bc715966bf0af65128b1ec32013bf4e3b4520d2bf8a2f5912c3115700c.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 365, + 482, + 422 + ], + "lines": [ + { + "bbox": [ + 130, + 365, + 482, + 422 + ], + "spans": [ + { + "bbox": [ + 130, + 365, + 482, + 422 + ], + "type": "text", + "content": "Fig. 3: Ablation studies of different frame aggregation mechanisms. Besides (a) Latte's slide window aggregation, we test two different aggregation methods, including (b) replacing single frame student predictions with multiple frames in Eq. (5) and (c) adding non-overlapping aggregation to (b). Results show that our slide window aggregation can better evaluate the local consistency and results in consistent improvement." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 427, + 336, + 439 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 427, + 336, + 439 + ], + "spans": [ + { + "bbox": [ + 131, + 427, + 336, + 439 + ], + "type": "text", + "content": "4.3 Ablation Studies and Visualization" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 445, + 481, + 470 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 445, + 481, + 470 + ], + "spans": [ + { + "bbox": [ + 130, + 445, + 481, + 470 + ], + "type": "text", + "content": "In this section, we investigate the effectiveness of each component in Latte by thorough ablation studies and qualitative results." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 486, + 482, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 486, + 482, + 628 + ], + "spans": [ + { + "bbox": [ + 130, + 486, + 482, + 628 + ], + "type": "text", + "content": "Effectiveness of cross-modal learning. The main optimization objectives of Latte contain three parts, including ST-entropy-based cross-modal consistency (both 2D-to-3D and 3D-to-2D) and cross-entropy loss between predictions and cross-modal pseudo-labels. We justify the effectiveness by using different combinations of these components as in Tab. 2. As shown from experiments No. 1 to No. 6, disabling any part of Latte causes performance decreases of more than " + }, + { + "bbox": [ + 130, + 486, + 482, + 628 + ], + "type": "inline_equation", + "content": "1.6\\%" + }, + { + "bbox": [ + 130, + 486, + 482, + 628 + ], + "type": "text", + "content": " relatively on the cross-modal prediction (xM), which proves the effectiveness of each component we included. A special case is No. 5 on U-to-S, where disabling ST " + }, + { + "bbox": [ + 130, + 486, + 482, + 628 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{23}^{\\mathrm{xM}}" + }, + { + "bbox": [ + 130, + 486, + 482, + 628 + ], + "type": "text", + "content": " only leads to a minor " + }, + { + "bbox": [ + 130, + 486, + 482, + 628 + ], + "type": "inline_equation", + "content": "0.1\\%" + }, + { + "bbox": [ + 130, + 486, + 482, + 628 + ], + "type": "text", + "content": " drop in accuracy. This is mainly because the 2D performance is already close to saturation (37.4 vs. 38.7 of Oracle TTA) without the guidance of 3D predictions. In this case, introducing 3D information to 2D is less effective compared to the other benchmarks." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 630, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 630, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 630, + 482, + 666 + ], + "type": "text", + "content": "Besides optimization objects, we further testify two important components of Latte, including ST-entropy-based weighting in Eq. (10) and Eq. (13) as well as percentile filtering in Eq. (8). Specifically, disabling percentile filtering as in No." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "type": "text", + "content": "H. Cao et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 159, + 114, + 253, + 176 + ], + "blocks": [ + { + "bbox": [ + 159, + 114, + 253, + 176 + ], + "lines": [ + { + "bbox": [ + 159, + 114, + 253, + 176 + ], + "spans": [ + { + "bbox": [ + 159, + 114, + 253, + 176 + ], + "type": "image", + "image_path": "6b7836395acd46ec5d09642e0d8b6e842ddff0d8b006624ac84dfe69592869ed.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 225, + 177, + 389, + 186 + ], + "lines": [ + { + "bbox": [ + 225, + 177, + 389, + 186 + ], + "spans": [ + { + "bbox": [ + 225, + 177, + 389, + 186 + ], + "type": "text", + "content": "(a) Ablation studies of various ST voxel sizes." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 259, + 114, + 353, + 176 + ], + "blocks": [ + { + "bbox": [ + 259, + 114, + 353, + 176 + ], + "lines": [ + { + "bbox": [ + 259, + 114, + 353, + 176 + ], + "spans": [ + { + "bbox": [ + 259, + 114, + 353, + 176 + ], + "type": "image", + "image_path": "2755e97e95cc69d4cd070c321af5a2940b87b922677f4bcd78ebc54a8359a136.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 360, + 114, + 453, + 176 + ], + "blocks": [ + { + "bbox": [ + 360, + 114, + 453, + 176 + ], + "lines": [ + { + "bbox": [ + 360, + 114, + 453, + 176 + ], + "spans": [ + { + "bbox": [ + 360, + 114, + 453, + 176 + ], + "type": "image", + "image_path": "7c9ec6155f3a12c5f3c433b3d33768baae76b4c4c4c957fe958691ec0172403d.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 159, + 187, + 252, + 248 + ], + "blocks": [ + { + "bbox": [ + 159, + 187, + 252, + 248 + ], + "lines": [ + { + "bbox": [ + 159, + 187, + 252, + 248 + ], + "spans": [ + { + "bbox": [ + 159, + 187, + 252, + 248 + ], + "type": "image", + "image_path": "3e4fc3aefee68e3c50682ba0247c3960437614d62b6cb5aecdffd28f29d298db.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 208, + 250, + 405, + 259 + ], + "lines": [ + { + "bbox": [ + 208, + 250, + 405, + 259 + ], + "spans": [ + { + "bbox": [ + 208, + 250, + 405, + 259 + ], + "type": "text", + "content": "(b) Ablation studies of filtering percentile " + }, + { + "bbox": [ + 208, + 250, + 405, + 259 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 208, + 250, + 405, + 259 + ], + "type": "text", + "content": " in Eq. (8)." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 260, + 187, + 351, + 248 + ], + "blocks": [ + { + "bbox": [ + 260, + 187, + 351, + 248 + ], + "lines": [ + { + "bbox": [ + 260, + 187, + 351, + 248 + ], + "spans": [ + { + "bbox": [ + 260, + 187, + 351, + 248 + ], + "type": "image", + "image_path": "8216ef83bcc00af7ce9fc4349c5ebfc892298cd3257daca86fce94f832a9c0cf.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 146, + 268, + 466, + 280 + ], + "lines": [ + { + "bbox": [ + 146, + 268, + 466, + 280 + ], + "spans": [ + { + "bbox": [ + 146, + 268, + 466, + 280 + ], + "type": "text", + "content": "Fig. 4: Sensitivity analysis about the ST voxel size and confident percentile " + }, + { + "bbox": [ + 146, + 268, + 466, + 280 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 146, + 268, + 466, + 280 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 360, + 187, + 452, + 248 + ], + "blocks": [ + { + "bbox": [ + 360, + 187, + 452, + 248 + ], + "lines": [ + { + "bbox": [ + 360, + 187, + 452, + 248 + ], + "spans": [ + { + "bbox": [ + 360, + 187, + 452, + 248 + ], + "type": "image", + "image_path": "b81f9446c37d96988e78a39195bf51e0b51e01109ba2865ae713e45acb63b0f3.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 285, + 482, + 334 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 285, + 482, + 334 + ], + "spans": [ + { + "bbox": [ + 130, + 285, + 482, + 334 + ], + "type": "text", + "content": "7 leads to a relative accuracy drop of more than " + }, + { + "bbox": [ + 130, + 285, + 482, + 334 + ], + "type": "inline_equation", + "content": "0.5\\%" + }, + { + "bbox": [ + 130, + 285, + 482, + 334 + ], + "type": "text", + "content": " across all benchmarks while utilizing point-wise entropy instead of our ST entropy as in No. 8 degenerates the performance by more than " + }, + { + "bbox": [ + 130, + 285, + 482, + 334 + ], + "type": "inline_equation", + "content": "2.0\\%" + }, + { + "bbox": [ + 130, + 285, + 482, + 334 + ], + "type": "text", + "content": " relatively, which justifies that ST entropy can better estimate the prediction reliability compared to the point-wise entropy." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 350, + 482, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 350, + 482, + 578 + ], + "spans": [ + { + "bbox": [ + 130, + 350, + 482, + 578 + ], + "type": "text", + "content": "Different frame aggregation mechanisms. One of the core designs of Latte is the slide window frame aggregation, which is designed to capture temporally local correspondences and estimate their prediction consistency. To justify the aforementioned claim, we compare our slide window aggregation with two aggregation variants as shown in Fig. 3. Specifically, variant (b) regards all student prediction frames rather a single frame within the time window as the query in Eq. (5), while variant (c) further utilizes a non-overlapping sliding strategy. According to Fig. 3, our slide window aggregation surpasses the other two aggregation variants under all sizes of time windows across three benchmarks, which justifies the effectiveness of our method. The inferior performance of variants (b) and (c) could be due to two factors. Firstly, aggregating student predictions could be an inferior option since student predictions tend to be less stable, and aggregating multiple frames can therefore lead to noisy queries for cross-modal learning. Secondly, non-overlapping aggregation can result in a miss-checking area on the boundary of each time window, which leads to unrepresentative entropy estimation for some ST voxels. Another observation is that utilizing a smaller slide window can usually yield better performance since temporally local consistency can be captured. This justifies the effectiveness of estimating prediction reliability through temporally local consistency." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 594, + 483, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 483, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 483, + 666 + ], + "type": "text", + "content": "Parameter sensitivity analysis. To testify whether Latte is sensitive towards hyper-parameter settings, we conduct sensitivity analysis on Latte across three benchmarks. Here we illustrate the sensitivity analysis of two hyper-parameters of Latte, including ST voxel size and filtering percentile " + }, + { + "bbox": [ + 130, + 594, + 483, + 666 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 130, + 594, + 483, + 666 + ], + "type": "text", + "content": " as shown in Fig. 4. In terms of ST voxel sizes, one can tell from Fig. 4a that all benchmarks share the same optimal voxel size as 0.2, where performance drops as the voxel size" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 148, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 148, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 148, + 91, + 447, + 102 + ], + "type": "text", + "content": "Reliable Spatial-Temporal Voxels For Multi-Modal Test-Time Adaptation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 149, + 114, + 463, + 242 + ], + "blocks": [ + { + "bbox": [ + 149, + 114, + 463, + 242 + ], + "lines": [ + { + "bbox": [ + 149, + 114, + 463, + 242 + ], + "spans": [ + { + "bbox": [ + 149, + 114, + 463, + 242 + ], + "type": "image", + "image_path": "e7f6804bd9099bb2c160d639ef555a80d2aa6b99dcd01d54f1c9587948f9e2da.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 249, + 480, + 271 + ], + "lines": [ + { + "bbox": [ + 130, + 249, + 480, + 271 + ], + "spans": [ + { + "bbox": [ + 130, + 249, + 480, + 271 + ], + "type": "text", + "content": "Fig. 5: Qualitative results of Latte on S-to-S. Here we visualize modality-specific predictions and cross-modal predictions from Latte during two sets of consecutive frames." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 277, + 482, + 385 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 277, + 482, + 385 + ], + "spans": [ + { + "bbox": [ + 130, + 277, + 482, + 385 + ], + "type": "text", + "content": "becomes too large or too small. This is mainly because large voxels could cause boundary ambiguity while small voxels contain insufficient representative points for evaluation. As for filtering percentile " + }, + { + "bbox": [ + 130, + 277, + 482, + 385 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 130, + 277, + 482, + 385 + ], + "type": "text", + "content": ", non-trivial improvement can be observed on U-to-S and A-to-S when " + }, + { + "bbox": [ + 130, + 277, + 482, + 385 + ], + "type": "inline_equation", + "content": "\\alpha \\geq 0.8" + }, + { + "bbox": [ + 130, + 277, + 482, + 385 + ], + "type": "text", + "content": ", while a lower " + }, + { + "bbox": [ + 130, + 277, + 482, + 385 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 130, + 277, + 482, + 385 + ], + "type": "text", + "content": " begins to discard more confident ST voxels leading to a slight drop in accuracy. On S-to-S, Latte without filtering (i.e., No. 7 in Tab. 2) performs slightly better on cross-modal predictions with a minor gap of " + }, + { + "bbox": [ + 130, + 277, + 482, + 385 + ], + "type": "inline_equation", + "content": "0.1\\%" + }, + { + "bbox": [ + 130, + 277, + 482, + 385 + ], + "type": "text", + "content": ", while it can also be observed that setting " + }, + { + "bbox": [ + 130, + 277, + 482, + 385 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 130, + 277, + 482, + 385 + ], + "type": "text", + "content": " as 0.9 leads to an improvement of " + }, + { + "bbox": [ + 130, + 277, + 482, + 385 + ], + "type": "inline_equation", + "content": "0.3\\%" + }, + { + "bbox": [ + 130, + 277, + 482, + 385 + ], + "type": "text", + "content": " on both modality-specific predictions, which justifies the overall effectiveness of our filtering procedures." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 400, + 482, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 400, + 482, + 533 + ], + "spans": [ + { + "bbox": [ + 130, + 400, + 482, + 533 + ], + "type": "text", + "content": "Qualitative results. Our main motivation is utilizing prediction consistency of spatial-temporal correspondences to estimate modality reliability and achieve better cross-modal attending. As shown in Fig. 5, Latte can effectively attend the modality with more consistent predictions in spatial-temporal correspondences and therefore improve the cross-modal prediction consistency across time. For instance, Latte successfully attends to the stable 2D predictions pedestrian in the upper set of consecutive frames, suppressing the spatial-temporal inconsistency from 3D predictions. Similar observations can also be found in the lower sample, where the more consistent 2D predictions of buildings in the red rectangle and 3D predictions of roads in the blue rectangle prevail. More qualitative comparisons with previous SOTA methods are presented in our appendix." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 547, + 220, + 560 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 547, + 220, + 560 + ], + "spans": [ + { + "bbox": [ + 132, + 547, + 220, + 560 + ], + "type": "text", + "content": "5 Conclusion" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "text", + "content": "In this paper, we propose a novel MM-TTA method called Latte, which leverages prediction consistency of spatial-temporal correspondences in MM-TTA for 3D segmentation. Latte utilizes a slide window frame aggregation to extract ST voxels and estimates temporally local consistency by the ST entropy of each modality. By attending to the modality with more consistency, Latte can achieve stable improvement on challenging MM-TTA scenarios. Despite its effectiveness, Latte struggles to rectify consistently wrong predictions across time, which is challenging and worth further investigation in the future." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "text", + "content": "H. Cao et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 115, + 234, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 115, + 234, + 129 + ], + "spans": [ + { + "bbox": [ + 133, + 115, + 234, + 129 + ], + "type": "text", + "content": "Acknowledgment" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 139, + 482, + 199 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 139, + 482, + 199 + ], + "spans": [ + { + "bbox": [ + 130, + 139, + 482, + 199 + ], + "type": "text", + "content": "This research is supported by the National Research Foundation, Singapore, under the NRF Medium Sized Centre for Advanced Robotics Technology Innovation (CARTIN). Any opinions, findings and conclusions or recommendations expressed in this material are those of the author(s) and do not reflect the views of National Research Foundation, Singapore." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 217, + 197, + 228 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 217, + 197, + 228 + ], + "spans": [ + { + "bbox": [ + 133, + 217, + 197, + 228 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 134, + 239, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 138, + 239, + 480, + 284 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 239, + 480, + 284 + ], + "spans": [ + { + "bbox": [ + 138, + 239, + 480, + 284 + ], + "type": "text", + "content": "1. Behley, J., Garbade, M., Milioto, A., Quenzel, J., Behnke, S., Stachniss, C., Gall, J.: Semantickitti: A dataset for semantic scene understanding of lidar sequences. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 9297-9307 (2019)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 285, + 480, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 285, + 480, + 327 + ], + "spans": [ + { + "bbox": [ + 138, + 285, + 480, + 327 + ], + "type": "text", + "content": "2. Caesar, H., Bankiti, V., Lang, A.H., Vora, S., Liong, V.E., Xu, Q., Krishnan, A., Pan, Y., Baldan, G., Beijbom, O.: nuscenes: A multimodal dataset for autonomous driving. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11621-11631 (2020)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 328, + 480, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 328, + 480, + 360 + ], + "spans": [ + { + "bbox": [ + 138, + 328, + 480, + 360 + ], + "type": "text", + "content": "3. Cao, H., Xu, Y., Yang, J., Yin, P., Yuan, S., Xie, L.: Mopa: Multi-modal prior aided domain adaptation for 3d semantic segmentation. arXiv preprint arXiv:2309.11839 (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 361, + 480, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 361, + 480, + 403 + ], + "spans": [ + { + "bbox": [ + 138, + 361, + 480, + 403 + ], + "type": "text", + "content": "4. Cao, H., Xu, Y., Yang, J., Yin, P., Yuan, S., Xie, L.: Multi-modal continual test-time adaptation for 3d semantic segmentation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 18809-18819 (October 2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 404, + 480, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 404, + 480, + 446 + ], + "spans": [ + { + "bbox": [ + 138, + 404, + 480, + 446 + ], + "type": "text", + "content": "5. Chen, R., Liu, Y., Kong, L., Zhu, X., Ma, Y., Li, Y., Hou, Y., Qiao, Y., Wang, W.: Clip2scene: Towards label-efficient 3d scene understanding by clip. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7020-7030 (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 448, + 480, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 448, + 480, + 480 + ], + "spans": [ + { + "bbox": [ + 138, + 448, + 480, + 480 + ], + "type": "text", + "content": "6. Choy, C., Gwak, J., Savarese, S.: 4d spatio-temporal convnets: Minkowski convolutional neural networks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3075-3084 (2019)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 481, + 480, + 513 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 481, + 480, + 513 + ], + "spans": [ + { + "bbox": [ + 138, + 481, + 480, + 513 + ], + "type": "text", + "content": "7. Fan, H., Yang, Y., Kankanhalli, M.: Point 4d transformer networks for spatiotemporal modeling in point cloud videos. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14204-14213 (2021)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 514, + 480, + 534 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 514, + 480, + 534 + ], + "spans": [ + { + "bbox": [ + 138, + 514, + 480, + 534 + ], + "type": "text", + "content": "8. Fan, H., Yu, X., Ding, Y., Yang, Y., Kankanhalli, M.: Pistnet: Point spatio-temporal convolution on point cloud sequences. arXiv preprint arXiv:2205.13713 (2022)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 535, + 480, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 535, + 480, + 578 + ], + "spans": [ + { + "bbox": [ + 138, + 535, + 480, + 578 + ], + "type": "text", + "content": "9. Feng, D., Haase-Schütz, C., Rosenbaum, L., Hertlein, H., Glaeser, C., Timm, F., Wiesbeck, W., Dietmayer, K.: Deep multi-modal object detection and semantic segmentation for autonomous driving: Datasets, methods, and challenges. IEEE Transactions on Intelligent Transportation Systems 22(3), 1341-1360 (2020)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 134, + 579, + 480, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 579, + 480, + 611 + ], + "spans": [ + { + "bbox": [ + 134, + 579, + 480, + 611 + ], + "type": "text", + "content": "10. Geyer, J., Kassahun, Y., Mahmudi, M., Ricou, X., Durgesh, R., Chung, A.S., Hauswald, L., Pham, V.H., Mühlegg, M., Dorn, S., et al.: A2d2: Audi autonomous driving dataset. arXiv preprint arXiv:2004.06320 (2020)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 134, + 612, + 480, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 612, + 480, + 643 + ], + "spans": [ + { + "bbox": [ + 134, + 612, + 480, + 643 + ], + "type": "text", + "content": "11. Gong, T., Jeong, J., Kim, T., Kim, Y., Shin, J., Lee, S.J.: Robust continual test-time adaptation: Instance-aware bn and prediction-balanced memory. arXiv preprint arXiv:2208.05117 (2022)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 134, + 644, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 644, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 134, + 644, + 480, + 665 + ], + "type": "text", + "content": "12. Goyal, S., Sun, M., Raghunathan, A., Kolter, J.Z.: Test time adaptation via conjugate pseudo-labels. In: Advances in Neural Information Processing Systems (2022)" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 148, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 148, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 148, + 91, + 447, + 102 + ], + "type": "text", + "content": "Reliable Spatial-Temporal Voxels For Multi-Modal Test-Time Adaptation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 117, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 133, + 117, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 117, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 133, + 117, + 480, + 138 + ], + "type": "text", + "content": "13. Graham, B.: Sparse 3d convolutional neural networks. arXiv preprint arXiv:1505.02890 (2015)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 139, + 480, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 139, + 480, + 171 + ], + "spans": [ + { + "bbox": [ + 133, + 139, + 480, + 171 + ], + "type": "text", + "content": "14. Guo, Y., Wang, H., Hu, Q., Liu, H., Liu, L., Bennamoun, M.: Deep learning for 3d point clouds: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence 43(12), 4338-4364 (2020)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 171, + 480, + 203 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 171, + 480, + 203 + ], + "spans": [ + { + "bbox": [ + 133, + 171, + 480, + 203 + ], + "type": "text", + "content": "15. Huang, S., Gojcic, Z., Huang, J., Wieser, A., Schindler, K.: Dynamic 3d scene analysis by point cloud accumulation. In: European Conference on Computer Vision. pp. 674-690. Springer (2022)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 203, + 480, + 236 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 203, + 480, + 236 + ], + "spans": [ + { + "bbox": [ + 133, + 203, + 480, + 236 + ], + "type": "text", + "content": "16. Jaritz, M., Vu, T.H., De Charette, R., Wirbel, E., Pérez, P.: Cross-modal learning for domain adaptation in 3d semantic segmentation. IEEE Transactions on Pattern Analysis and Machine Intelligence 45(2), 1533-1544 (2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 133, + 236, + 480, + 256 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 236, + 480, + 256 + ], + "spans": [ + { + "bbox": [ + 133, + 236, + 480, + 256 + ], + "type": "text", + "content": "17. Ji, X., Yuan, S., Yin, P., Xie, L.: Lio-gvm: an accurate, tightly-coupled lidar-inertial odometry with gaussian voxel map. IEEE Robotics and Automation Letters (2024)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 133, + 256, + 480, + 299 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 256, + 480, + 299 + ], + "spans": [ + { + "bbox": [ + 133, + 256, + 480, + 299 + ], + "type": "text", + "content": "18. Li, M., Zhang, Y., Xie, Y., Gao, Z., Li, C., Zhang, Z., Qu, Y.: Cross-domain and cross-modal knowledge distillation in domain adaptation for 3d semantic segmentation. In: Proceedings of the 30th ACM International Conference on Multimedia. pp. 3829-3837 (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 133, + 300, + 480, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 300, + 480, + 331 + ], + "spans": [ + { + "bbox": [ + 133, + 300, + 480, + 331 + ], + "type": "text", + "content": "19. Liang, J., Hu, D., Feng, J.: Do we really need to access the source data? source hypothesis transfer for unsupervised domain adaptation. In: International Conference on Machine Learning. pp. 6028-6039. PMLR (2020)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 133, + 332, + 480, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 332, + 480, + 374 + ], + "spans": [ + { + "bbox": [ + 133, + 332, + 480, + 374 + ], + "type": "text", + "content": "20. Liu, W., Luo, Z., Cai, Y., Yu, Y., Ke, Y., Junior, J.M., Gonçalves, W.N., Li, J.: Adversarial unsupervised domain adaptation for 3d semantic segmentation with multi-modal learning. ISPRS Journal of Photogrammetry and Remote Sensing 176, 211-221 (2021)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 133, + 374, + 480, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 374, + 480, + 407 + ], + "spans": [ + { + "bbox": [ + 133, + 374, + 480, + 407 + ], + "type": "text", + "content": "21. Liu, Y., Kothari, P., Van Delft, B., Bellot-Gurlet, B., Mordan, T., Alahi, A.: Ttt++: When does self-supervised test-time training fail or thrive? Advances in Neural Information Processing Systems 34, 21808-21820 (2021)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 133, + 407, + 480, + 438 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 407, + 480, + 438 + ], + "spans": [ + { + "bbox": [ + 133, + 407, + 480, + 438 + ], + "type": "text", + "content": "22. Niu, S., Wu, J., Zhang, Y., Chen, Y., Zheng, S., Zhao, P., Tan, M.: Efficient test-time model adaptation without forgetting. In: International Conference on Machine Learning. pp. 16888-16905. PMLR (2022)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 133, + 438, + 480, + 471 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 438, + 480, + 471 + ], + "spans": [ + { + "bbox": [ + 133, + 438, + 480, + 471 + ], + "type": "text", + "content": "23. Niu, S., Wu, J., Zhang, Y., Wen, Z., Chen, Y., Zhao, P., Tan, M.: Towards stable test-time adaptation in dynamic wild world. arXiv preprint arXiv:2302.12400 (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 133, + 471, + 480, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 471, + 480, + 514 + ], + "spans": [ + { + "bbox": [ + 133, + 471, + 480, + 514 + ], + "type": "text", + "content": "24. Peng, D., Lei, Y., Li, W., Zhang, P., Guo, Y.: Sparse-to-dense feature matching: Intra and inter domain cross-modal learning in domain adaptation for 3d semantic segmentation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 7108-7117 (2021)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 133, + 514, + 480, + 546 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 514, + 480, + 546 + ], + "spans": [ + { + "bbox": [ + 133, + 514, + 480, + 546 + ], + "type": "text", + "content": "25. Piergiovanni, A., Casser, V., Ryoo, M.S., Angelova, A.: 4d-net for learned multimodal alignment. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 15435-15445 (2021)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 133, + 547, + 480, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 547, + 480, + 578 + ], + "spans": [ + { + "bbox": [ + 133, + 547, + 480, + 578 + ], + "type": "text", + "content": "26. Qi, C.R., Su, H., Mo, K., Guibas, L.J.: Pointnet: Deep learning on point sets for 3d classification and segmentation. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 652-660 (2017)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 133, + 578, + 480, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 578, + 480, + 622 + ], + "spans": [ + { + "bbox": [ + 133, + 578, + 480, + 622 + ], + "type": "text", + "content": "27. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International Conference on Machine Learning. pp. 8748-8763. PMLR (2021)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 133, + 622, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 622, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 622, + 480, + 665 + ], + "type": "text", + "content": "28. Saltori, C., Krivosheev, E., Lathuilière, S., Sebe, N., Galasso, F., Fiameni, G., Ricci, E., Poiesi, F.: Gipso: Geometrically informed propagation for online adaptation in 3d lidar segmentation. In: European Conference on Computer Vision. pp. 567-585. Springer (2022)" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "text", + "content": "H. Cao et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 666 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 480, + 160 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 480, + 160 + ], + "type": "text", + "content": "29. Shin, I., Tsai, Y.H., Zhuang, B., Schulter, S., Liu, B., Garg, S., Kweon, I.S., Yoon, K.J.: Mm-tta: multi-modal test-time adaptation for 3d semantic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16928-16937 (2022)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 160, + 480, + 204 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 160, + 480, + 204 + ], + "spans": [ + { + "bbox": [ + 132, + 160, + 480, + 204 + ], + "type": "text", + "content": "30. Simons, C., Raychaudhuri, D.S., Ahmed, S.M., You, S., Karydis, K., Roy-Chowdhury, A.K.: Summit: Source-free adaptation of uni-modal models to multimodal targets. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 1239-1249 (October 2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 204, + 480, + 236 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 204, + 480, + 236 + ], + "spans": [ + { + "bbox": [ + 132, + 204, + 480, + 236 + ], + "type": "text", + "content": "31. Song, J., Lee, J., Kweon, I.S., Choi, S.: Ecotta: Memory-efficient continual test-time adaptation via self-distilled regularization. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11920-11929 (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 236, + 480, + 268 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 236, + 480, + 268 + ], + "spans": [ + { + "bbox": [ + 132, + 236, + 480, + 268 + ], + "type": "text", + "content": "32. Su, Y., Xu, X., Li, T., Jia, K.: Revisiting realistic test-time training: Sequential inference and adaptation by anchored clustering regularized self-training. arXiv preprint arXiv:2303.10856 (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 268, + 480, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 268, + 480, + 300 + ], + "spans": [ + { + "bbox": [ + 132, + 268, + 480, + 300 + ], + "type": "text", + "content": "33. Tang, H., Liu, Z., Zhao, S., Lin, Y., Lin, J., Wang, H., Han, S.: Searching efficient 3d architectures with sparse point-voxel convolution. In: European Conference on Computer Vision. pp. 685–702. Springer (2020)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 300, + 480, + 332 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 300, + 480, + 332 + ], + "spans": [ + { + "bbox": [ + 132, + 300, + 480, + 332 + ], + "type": "text", + "content": "34. Tarvainen, A., Valpola, H.: Mean teachers are better role models: Weight-averaged consistency targets improve semi-supervised deep learning results. Advances in Neural Information Processing Systems 30 (2017)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 333, + 480, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 333, + 480, + 365 + ], + "spans": [ + { + "bbox": [ + 132, + 333, + 480, + 365 + ], + "type": "text", + "content": "35. Vedula, S., Baker, S., Rander, P., Collins, R., Kanade, T.: Three-dimensional scene flow. In: Proceedings of the Seventh IEEE International Conference on Computer Vision. vol. 2, pp. 722-729. IEEE (1999)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 365, + 480, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 365, + 480, + 407 + ], + "spans": [ + { + "bbox": [ + 132, + 365, + 480, + 407 + ], + "type": "text", + "content": "36. Vizzo, I., Guadagnino, T., Mersch, B., Wiesmann, L., Behley, J., Stachniss, C.: Kiss-icp: In defense of point-to-point icp-simple, accurate, and robust registration if done the right way. IEEE Robotics and Automation Letters 8(2), 1029-1036 (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 407, + 480, + 440 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 407, + 480, + 440 + ], + "spans": [ + { + "bbox": [ + 132, + 407, + 480, + 440 + ], + "type": "text", + "content": "37. Vogel, C., Schindler, K., Roth, S.: 3d scene flow estimation with a rigid motion prior. In: 2011 International Conference on Computer Vision. pp. 1291-1298. IEEE (2011)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 440, + 480, + 461 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 440, + 480, + 461 + ], + "spans": [ + { + "bbox": [ + 132, + 440, + 480, + 461 + ], + "type": "text", + "content": "38. Vogel, C., Schindler, K., Roth, S.: Piecewise rigid scene flow. In: Proceedings of the IEEE International Conference on Computer Vision. pp. 1377-1384 (2013)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 461, + 480, + 494 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 461, + 480, + 494 + ], + "spans": [ + { + "bbox": [ + 132, + 461, + 480, + 494 + ], + "type": "text", + "content": "39. Wang, D., Shelhamer, E., Liu, S., Olshausen, B., Darrell, T.: Tent: Fully test-time adaptation by entropy minimization. In: International Conference on Learning Representations (2021), https://openreview.net/forum?id=uX13bZLkr3c" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 494, + 480, + 515 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 494, + 480, + 515 + ], + "spans": [ + { + "bbox": [ + 132, + 494, + 480, + 515 + ], + "type": "text", + "content": "40. Wang, J.K., Wibisono, A.: Towards understanding gd with hard and conjugate pseudo-labels for test-time adaptation. arXiv preprint arXiv:2210.10019 (2022)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 515, + 480, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 515, + 480, + 547 + ], + "spans": [ + { + "bbox": [ + 132, + 515, + 480, + 547 + ], + "type": "text", + "content": "41. Wang, Q., Fink, O., Van Gool, L., Dai, D.: Continual test-time domain adaptation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7201-7211 (2022)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 547, + 480, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 547, + 480, + 568 + ], + "spans": [ + { + "bbox": [ + 132, + 547, + 480, + 568 + ], + "type": "text", + "content": "42. Wyner, A.: Recent results in the shannon theory. IEEE Transactions on information Theory 20(1), 2-10 (1974)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 568, + 480, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 568, + 480, + 601 + ], + "spans": [ + { + "bbox": [ + 132, + 568, + 480, + 601 + ], + "type": "text", + "content": "43. Xie, E., Wang, W., Yu, Z., Anandkumar, A., Alvarez, J.M., Luo, P.: Segformer: Simple and efficient design for semantic segmentation with transformers. Advances in Neural Information Processing Systems 34, 12077-12090 (2021)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 601, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 601, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 601, + 480, + 632 + ], + "type": "text", + "content": "44. Xing, B., Ying, X., Wang, R., Yang, J., Chen, T.: Cross-modal contrastive learning for domain adaptation in 3d semantic segmentation. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 37, pp. 2974-2982 (2023)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 632, + 480, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 480, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 480, + 666 + ], + "type": "text", + "content": "45. Xu, J., Miao, Z., Zhang, D., Pan, H., Liu, K., Hao, P., Zhu, J., Sun, Z., Li, H., Zhan, X.: Int: Towards infinite-frames 3d detection with an efficient framework. In: European Conference on Computer Vision. pp. 193–209. Springer (2022)" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 148, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 148, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 148, + 91, + 447, + 102 + ], + "type": "text", + "content": "Reliable Spatial-Temporal Voxels For Multi-Modal Test-Time Adaptation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 216 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 480, + 149 + ], + "type": "text", + "content": "46. Yin, P., Cao, H., Nguyen, T.M., Yuan, S., Zhang, S., Liu, K., Xie, L.: Outram: One-shot global localization via triangulated scene graph and global outlier pruning. arXiv preprint arXiv:2309.08914 (2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 150, + 480, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 150, + 480, + 182 + ], + "spans": [ + { + "bbox": [ + 132, + 150, + 480, + 182 + ], + "type": "text", + "content": "47. Yuan, L., Xie, B., Li, S.: Robust test-time adaptation in dynamic scenarios. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15922-15932 (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 183, + 480, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 183, + 480, + 216 + ], + "spans": [ + { + "bbox": [ + 132, + 183, + 480, + 216 + ], + "type": "text", + "content": "48. Zhang, M., Levine, S., Finn, C.: Memo: Test time robustness via adaptation and augmentation. Advances in Neural Information Processing Systems 35, 38629-38642 (2022)" + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "text", + "content": "H. Cao et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Reliable and Efficient Concept Erasure of Text-to-Image Diffusion Models/d5988da3-2c65-4cde-838f-fe6bf86e8cd3_content_list.json b/2024/Reliable and Efficient Concept Erasure of Text-to-Image Diffusion Models/d5988da3-2c65-4cde-838f-fe6bf86e8cd3_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..806725aeffbaebb51b25e5a0236dcbad1345b517 --- /dev/null +++ b/2024/Reliable and Efficient Concept Erasure of Text-to-Image Diffusion Models/d5988da3-2c65-4cde-838f-fe6bf86e8cd3_content_list.json @@ -0,0 +1,1840 @@ +[ + { + "type": "text", + "text": "Reliable and Efficient Concept Erasure of Text-to-Image Diffusion Models", + "text_level": 1, + "bbox": [ + 264, + 140, + 738, + 185 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Chao Gong $^{1,2\\star}$ , Kai Chen $^{1,2\\star}$ , Zhipeng Wei $^{1,2}$ , Jingjing Chen $^{1,2\\dagger}$ , and Yu-Gang Jiang $^{1,2}$", + "bbox": [ + 215, + 210, + 787, + 244 + ], + "page_idx": 0 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "$^{1}$ Shanghai Key Lab of Intell. Info. Processing, School of Computer Science, Fudan University", + "$^{2}$ Shanghai Collaborative Innovation Center on Intelligent Visual Computing {c Gong20, chenjingjing, ygj} @fudan.edu.cn, {kchen22, zpwei21} @m.fudan.edu.cn" + ], + "bbox": [ + 220, + 253, + 781, + 310 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Text-to-image models encounter safety issues, including concerns related to copyright and Not-Safe-For-Work (NSFW) content. Despite several methods have been proposed for erasing inappropriate concepts from diffusion models, they often exhibit incomplete erasure, consume a lot of computing resources, and inadvertently damage generation ability. In this work, we introduce Reliable and Efficient Concept Erasure (RECE), a novel approach that modifies the model in 3 seconds without necessitating additional fine-tuning. Specifically, RECE efficiently leverages a closed-form solution to derive new target embeddings, which are capable of regenerating erased concepts within the unlearned model. To mitigate inappropriate content potentially represented by derived embeddings, RECE further aligns them with harmless concepts in crossattention layers. The derivation and erasure of new representation embeddings are conducted iteratively to achieve a thorough erasure of inappropriate concepts. Besides, to preserve the model's generation ability, RECE introduces an additional regularization term during the derivation process, resulting in minimizing the impact on unrelated concepts during the erasure process. All the processes above are in closed-form, guaranteeing extremely efficient erasure in only 3 seconds. Benchmarking against previous approaches, our method achieves more efficient and thorough erasure with minor damage to original generation ability and demonstrates enhanced robustness against red-teaming tools. Code is available at https://github.com/CharlesGong12/RECE.", + "bbox": [ + 259, + 349, + 738, + 667 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "WARNING: This paper contains model outputs that may be offensive.", + "bbox": [ + 261, + 669, + 732, + 681 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Text-to-Image $\\cdot$ Concept Erasing $\\cdot$ Machine Unlearning", + "bbox": [ + 261, + 696, + 718, + 710 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 737, + 375, + 752 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In recent years, large-scale text-to-image (T2I) diffusion models have exhibited remarkable capability in synthesizing photo-realistic images from text prompts [15,", + "bbox": [ + 212, + 770, + 795, + 801 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Equal contributions.", + "bbox": [ + 232, + 811, + 383, + 824 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Corresponding author.", + "bbox": [ + 232, + 824, + 395, + 839 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "[20, 24, 28]. The exceptional performance of T2I diffusion models is largely due to the vast amount of training data collected from the Internet, which enables the models to imitate a wide variety of concepts. Unfortunately, such powerful models can also be misused to generate copyright infringement and Not-Safe-For-Work (NSFW) image content when conditioned on inappropriate text prompts [11, 30]. Especially the open-source release of the Stable Diffusion (SD) T2I model has made advanced image generation technology widely accessible. To alleviate this safety concern, several recent research efforts have incorporated safety mechanisms into T2I diffusion models, e.g. filtering out inappropriate training data and retraining model [22], censoring model outputs with an NSFW safety checker [21], and applying classifier-free guidance to steer the generation away from inappropriate concepts [29]. However, these safety mechanisms either demand expensive computational resources and time [25] or can be easily circumvented by malicious users due to the public availability of code and model parameters in open-source scenario [23].", + "bbox": [ + 212, + 146, + 787, + 375 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In response to the drawbacks mentioned above, an alternative is to erase inappropriate concepts from the T2I diffusion model [6-8, 13]. Specifically, given an inappropriate concept described in the text prompt, the pre-trained T2I diffusion model's parameters are fine-tuned to unlearn that concept so that the associated image content cannot be generated. Compared with previous security mechanisms, concept erasure neither requires training the entire model from scratch nor can be easily circumvented even in the case of open-source code. Despite promising progress in concept erasure, there exist several issues. On the one hand, most erasure methods require a high number of iterations to finetune considerable amounts of parameters [6, 8, 13], which inevitably degrades the generation capability and consumes a lot of computing resources. Only a recent work called UCE [7] modifies model parameters without fine-tuning using a closed-form solution, ensuring the model maintains original generation capability when erasing concepts. On the other hand, almost all methods fail to sufficiently erase inappropriate concepts, leaving them vulnerable to problematic prompts found by the red-teaming of T2I diffusion models [3, 32, 38]. This results in the unlearned model being compelled to regenerate inappropriate images.", + "bbox": [ + 212, + 380, + 787, + 636 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Inspired by the idea of adversarial fine-tuning, we propose a Reliable and Efficient Concept Erasure (RECE) method to address the aforementioned challenge, which continually finds new embeddings of the erased concepts during fine-tuning and then enables the unlearned model to erase these new concept embeddings. To speed up the unlearning process, the RECE method builds upon the previous fast and efficient concept erasure method UCE [7], which employs a closed-form editing to only modify the key and value projection matrices in cross-attention layers [33]. Similarly, the RECE method derives new embeddings that most effectively prompt the model to regenerate images of erased concepts, with a closed-form solution based on cross-attention output. Furthermore, a regularization term is introduced to preserve the image generation capability of the model by restricting the deviation of model parameters before and after modification. By editing the model and deriving embedding for multiple epochs,", + "bbox": [ + 212, + 643, + 787, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "F. Chao Gong et al.", + "bbox": [ + 271, + 114, + 405, + 128 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "RECE enables the unlearned model to preserve the image generation ability of unerased concepts and robustly refrains the model from generating images with erased concept content. All the processes above are in closed-form, guaranteeing extremely efficient erasure in 3 seconds. Our major contributions are summarized as follows:", + "bbox": [ + 212, + 146, + 787, + 220 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We present a novel concept erasure method - RECE that uses closed-form parameter editing and adversarial learning schemes for reliable and efficient concept erasing in only 3 seconds.", + "- RECE sufficiently erases concepts by deriving new embeddings that enable the unlearned model to regenerate erased concepts. In addition, a regularization term is introduced to minimize the impact on the model's capabilities.", + "- We conduct extensive experiments to validate the effectiveness of RECE for erasing unsafe contents, artistic styles and object classes. Additionally, we assess the robustness of RECE against three red-teaming tools and record fine-tuning durations to highlight the efficiency of RECE." + ], + "bbox": [ + 225, + 232, + 782, + 380 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 407, + 387, + 422 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "T2I Diffusion Models with Safety Mechanisms. In response to the issue of generating inappropriate images in T2I diffusion models, several research has explored solutions to address this concern. Briefly, existing research primarily falls into the following three distinct strategies: The first is filtering the training data and retraining the model [22]. However, retraining on curated datasets not only requires substantial computational resources investment but also results in the generation of inappropriate content [6] and performance degradation [29]. The second is censoring model output through safety checkers [21], or exploiting classifier-free guidance to steer the latent codes away from inappropriate concepts during inference [29]. However, in the case of open-source code, pretrained T2I diffusion model architectures and parameters are publicly available, so such post-hoc intervention strategies can be easily circumvented by malicious users [23]. The third is fine-tuning the partial parameters of the pre-trained T2I diffusion models to erase the model's representation capability of inappropriate concepts [6-8, 13]. While fine-tuning has been considered an effective strategy to prevent the generation of inappropriate content, current methods consume a lot of computing time and can be easily bypassed by red-teaming tools for T2I diffusion models.", + "bbox": [ + 212, + 439, + 787, + 710 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Red-Teaming Tools for T2I Diffusion Models. With the recent popularity of AI, red-teaming has been applied to AI models to enhance model stability by probing functional vulnerabilities [1, 2, 34, 35]. Recent works have also developed red-teaming tools for T2I diffusion models, which is a rarely explored field in AI red-teaming. For instance, Prompting4Debugging (P4D) [3] automatically finds the problematic prompts that would lead to inappropriate content via utilizing prompt engineering techniques and an auxiliary diffusion model", + "bbox": [ + 212, + 733, + 787, + 840 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "Reliable and Efficient Concept Erasure of Text-to-Image Diffusion Models", + "bbox": [ + 238, + 114, + 730, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "without any safety mechanisms to assess the reliability of deployed safety mechanisms. Conversely, UnlearnDiff [38] does not depend on an auxiliary diffusion model, which leverages the inherent classification capabilities of diffusion models, thereby providing computational efficiency without sacrificing effectiveness. Both works have the main weakness in assuming white-box access to the target model. In response to this issue, Ring-A-Bell [32], a model-agnostic framework capable of constructing attacks without prior knowledge of the target model. Specifically, Ring-A-Bell first performs concept extraction to obtain a holistic representation of inappropriate concepts. Subsequently, Ring-A-Bell automatically produces problematic prompts by leveraging the extracted concepts.", + "bbox": [ + 212, + 146, + 787, + 297 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3 Method", + "text_level": 1, + "bbox": [ + 215, + 321, + 330, + 338 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1 Preliminaries", + "text_level": 1, + "bbox": [ + 215, + 354, + 372, + 369 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Text-to-Image (T2I) Diffusion Models In contemporary Text-to-Image (T2I) applications, diffusion models have become the preferred choice [36] since the progressive denoising process [10] empowers them with superior image synthesis ability [4]. To reduce computational complexity, T2I often adopts latent diffusion models [12, 24, 27, 28], which operates on the low-dimensional latent space of a pre-trained variational autoencoder (VAE) [5] and employs a U-Net generative network as the denoising architecture [26]. To incorporate text conditioning into the image generation process, T2I encodes text by language models like CLIP [19, 24] and integrates text embeddings into U-Net through cross-attention layers. Specifically, these layers employ Query-Key-Value (QKV) structure [33] to represent the interactions between text and vision. For a given text embedding $c_{i}$ , keys and values are generated as $k_{i} = W_{k}c_{i}$ and $v_{i} = W_{v}c_{i}$ . These keys compute an attention map by multiplying with the query $q_{i}$ representing visual features, and then the cross-attention output is computed by attending over values $v_{i}$ :", + "bbox": [ + 212, + 380, + 787, + 607 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {O} \\propto \\operatorname {s o f t m a x} (q _ {i} k _ {i} ^ {T}) v _ {i}. \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 424, + 619, + 785, + 638 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Concept Erasing with Closed-form Solution There are existing erasure methods requiring fine-tuning, such as ESD, CA and SA [6,8,13]. However, such methods are relatively inefficient as they require thousands of fine-tuning steps. In contrast, UCE [7] is an efficient method which modifies the attention weights through a closed-form edit. UCE requires a \"source\" concept (e.g., \"nudity\") and a \"destination\" concept (e.g., empty text \"\"). Let $c_{i}$ represent the source embedding, $c_{i}^{*}$ denote the corresponding destination embedding, set $E$ denote concepts to erase, and set $P$ denote concepts to preserve. Given a K/V projection matrix $W^{\\mathrm{old}}$ (a concise notation for $W_{k}^{\\mathrm{old}}$ and $W_{v}^{\\mathrm{old}}$ ), UCE seeks new weights $W$ by editing concepts in $E$ while preserving concepts in $P$ . Specifically, the objective is to find weights such that the output $Wc_{i}$ for $c_{i} \\in E$ aligns with target values $W^{\\mathrm{old}}c_{i}^{*}$ instead of the original $W^{\\mathrm{old}}c_{i}$ . Meanwhile, to control parameter", + "bbox": [ + 212, + 657, + 787, + 843 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "F. Chao Gong et al.", + "bbox": [ + 271, + 114, + 405, + 128 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/c86ae884f2603a05e25af154bd26d60878181de1845b0ae7ce4bd072161e0486.jpg", + "image_caption": [ + "Model Editing" + ], + "image_footnote": [], + "bbox": [ + 223, + 142, + 493, + 280 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/9add03c9bdefba80762a644020336df9bd605f4c6efa04864ec39d86dfb327b6.jpg", + "image_caption": [ + "Embedding Derivation", + "Fig. 1: Overview of the proposed RECE. RECE consists of two main components: model editing and embedding derivation. First, erasing concepts by editing the model with a closed-form solution, and obtaining the edited cross-attention $W^{\\mathrm{new}}$ . Then, new embedding $c_{i}^{\\prime}$ can be derived by Eq. (7) given the original cross attention $W^{\\mathrm{old}}$ and the edited $W^{\\mathrm{new}}$ . In subsequent epochs, model editing and embedding derivation are looped." + ], + "image_footnote": [], + "bbox": [ + 495, + 142, + 781, + 280 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "changes, outputs for $c_{j} \\in P$ are preserved as $W^{\\mathrm{old}}c_{j}$ and a L2 regularization term is introduced:", + "bbox": [ + 212, + 424, + 784, + 454 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {W} \\sum_ {c _ {i} \\in E} | | W c _ {i} - W ^ {\\text {o l d}} c _ {i} ^ {*} | | _ {2} ^ {2} + \\lambda_ {1} \\sum_ {c _ {j} \\in P} | | W c _ {j} - W ^ {\\text {o l d}} c _ {j} | | _ {2} ^ {2} + \\lambda_ {2} | | W - W ^ {\\text {o l d}} | | _ {F} ^ {2}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 222, + 465, + 784, + 500 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\lambda_{1}$ and $\\lambda_{2}$ are scaling factors preserving the existing concepts. UCE [7] prove that this formula has a closed-form solution:", + "bbox": [ + 212, + 512, + 784, + 544 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nW = W ^ {\\text {o l d}} \\left(\\sum_ {c _ {i} \\in E} c _ {i} ^ {*} c _ {i} ^ {T} + \\lambda_ {1} \\sum_ {c _ {j} \\in P} c _ {j} c _ {j} ^ {T} + \\lambda_ {2} I\\right) \\left(\\sum_ {c _ {i} \\in E} c _ {i} c _ {i} ^ {T} + \\lambda_ {1} \\sum_ {c _ {j} \\in P} c _ {j} c _ {j} ^ {T} + \\lambda_ {2} I\\right) ^ {- 1}. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 222, + 555, + 784, + 590 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "UCE directly assigns cross-attention KV matrices using the closed-form solution, eliminating the need for fine-tuning. This makes UCE significantly faster, hence we use UCE in our method.", + "bbox": [ + 212, + 603, + 784, + 648 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2 Reliable and Efficient Concept Erasure (RECE)", + "text_level": 1, + "bbox": [ + 214, + 675, + 656, + 693 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "While UCE [7] offers a fast solution for removing undesired concepts from T2I diffusion models, it can still produce undesired content, as illustrated in Tab. 1. This suggests an incomplete erasure of these concepts. To effectively eliminate such undesired concepts, we efficiently erase closed-form embeddings capable of regenerating erased concepts within the unlearned model. The derivation of embeddings and erasure is conducted iteratively to achieve a thorough erasure of inappropriate concepts, as shown in Fig. 1.", + "bbox": [ + 212, + 705, + 787, + 811 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "Reliable and Efficient Concept Erasure of Text-to-Image Diffusion Models", + "bbox": [ + 238, + 114, + 732, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Finding Target Contents Let us take \"nudity\" for example. As depicted in the second column of Fig. 2, when directly providing the input prompt \"nudity\" to UCE models, only landscape or unrelated images are generated. This is because the word \"nudity\" has been aligned with the empty text\". However, the erasure of UCE is incomplete. We can generate an adversarial prompt that enables UCE's model to produce images containing nudity content again, similar to those generated by SD when provided with the prompt \"nudity\".", + "bbox": [ + 212, + 146, + 486, + 372 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this section, we introduce our method for deriving the new embedding in UCE's model, which guides UCE to generate nude images. As", + "bbox": [ + 212, + 373, + 488, + 434 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "elaborated in Sec. 3.1, T2I introduces text embeddings into image generation through cross-attention layers, where the projection matrices $W_{k}$ and $W_{v}$ are used to transform text embeddings. Let $W^{\\mathrm{old}}$ denote the projection matrices of the original U-Net before UCE's editing, $W^{\\mathrm{new}}$ represent the projection matrices after UCE's editing, $c$ denote the embedding of \"nudity\", and $c'$ signify our derived embedding. If we can find a $c'$ such that $W^{\\mathrm{new}}c'$ closely resembles $W^{\\mathrm{old}}c$ , then $c'$ can guide the edited model to generate nude images like how $c$ guides the original model. More precisely, the objective function is formulated as follows:", + "bbox": [ + 212, + 434, + 787, + 555 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {c ^ {\\prime}} \\sum_ {i} \\| W _ {i} ^ {\\mathrm {n e w}} c ^ {\\prime} - W _ {i} ^ {\\mathrm {o l d}} c \\| _ {2} ^ {2}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 397, + 566, + 785, + 597 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $W_{i}$ denotes K/V cross-attention matrices of U-Net. The solution $c'$ derived from Eq. (4) can be viewed as the actual representation of $c$ within the edited model. Evidently, Eq. (4) represents a convex function with respect to $c'$ , which possesses a unique global minimum. As derived in Appendix A, Eq. (4) admits a closed-form solution:", + "bbox": [ + 212, + 609, + 787, + 686 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nc ^ {\\prime} = \\left(\\sum_ {i} W _ {i} ^ {\\mathrm {n e w} ^ {T}} W _ {i} ^ {\\mathrm {n e w}}\\right) ^ {- 1} \\left(\\sum_ {i} W _ {i} ^ {\\mathrm {n e w} ^ {T}} W _ {i} ^ {\\mathrm {o l d}}\\right) c. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 321, + 698, + 785, + 743 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Given that text conditioning works in the form of embedding in Stable Diffusion (SD), we can use our derived embedding as text conditioning. As illustrated in the third column of Fig. 2, this derived embedding effectively guides UCE's edited model to once again generate nude images, indicating its capacity to represent the concept of \"nudity\" within edited model. Thus it demonstrates that the erasure process of UCE remains incomplete. To address this issue, we further", + "bbox": [ + 212, + 750, + 787, + 840 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/baf14b026fad6d13d18d70d1506374992031d69b73d505a07c188881514cb37c.jpg", + "image_caption": [ + "Fig.2: When given the input prompt \"nudity\", SD generates images containing nudity content, while UCE generates unrelated images. When given our derived embedding and regularized embedding, UCE generates nude images again. We use for publication purposes." + ], + "image_footnote": [], + "bbox": [ + 498, + 152, + 787, + 292 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "F. Chao Gong et al.", + "bbox": [ + 271, + 114, + 406, + 128 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "remove our derived embeddings $c'$ from UCE's model with Eq. (3) to prevent the generation of nude images.", + "bbox": [ + 214, + 146, + 782, + 176 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Regularization Term We can further erase the concept of \"nudity\" by substituting $c$ in Eq. (2) with our derived embedding $c'$ . However, upon directly erasing $c'$ , we observe a significant decline in the model's performance: it struggles to generate high-quality images for unrelated concepts as shown in Fig. 6. Hence, it becomes imperative to devise a method that retains the model's performance while erasing concepts.", + "bbox": [ + 212, + 202, + 782, + 292 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Let $W^{\\mathrm{new1}}$ denote the projection matrices after the last epoch's modification, and $W^{\\mathrm{new2}}$ denote the projection matrices after the current epoch. Partly, preserving the model's performance entails minimizing the impact on unrelated concepts. Consequently, we define our objective function as follows:", + "bbox": [ + 212, + 292, + 782, + 354 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {W} \\| W ^ {\\mathrm {n e w} 2} d - W ^ {\\mathrm {n e w} 1} d \\| _ {2} ^ {2}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 406, + 367, + 785, + 388 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $d$ represents an unrelated concept's embedding. Note that $W^{\\mathrm{new2}}$ is directly influenced by $c'$ , as it is derived after erasing $c'$ from $W^{\\mathrm{new1}}$ . We can obtain a theorem about our objective and its proof is provided in Appendix B:", + "bbox": [ + 214, + 401, + 782, + 448 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Theorem 1. If $c'$ is set to 0, Eq. (6) achieves its global minimum of 0.", + "bbox": [ + 215, + 460, + 733, + 478 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Intuitively, if $c_{i}$ in Eq. (3) is set to zero, the coefficient matrices on the right side will become an identity matrix when multiplied. As a result, $W$ will revert to being equivalent to $W^{\\mathrm{old}}$ , exerting minimal influence on unrelated concepts due to the absence of further modifications to $W$ . Therefore, we need to introduce a regularization term to the original objective Eq. (4) to ensure that the obtained $c'$ is close to zero, thereby minimizing its influence on the model's performance:", + "bbox": [ + 214, + 491, + 782, + 580 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {c ^ {\\prime}} \\sum_ {i} \\| W _ {i} ^ {\\text {n e w}} c ^ {\\prime} - W _ {i} ^ {\\text {o l d}} c \\| _ {2} ^ {2} + \\lambda \\| c ^ {\\prime} \\| _ {2} ^ {2}. \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 367, + 595, + 784, + 625 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "As derived in Appendix A, this final objective function also possesses a unique global minimum solution:", + "bbox": [ + 214, + 636, + 782, + 667 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\nc ^ {\\prime} = \\left(\\lambda I + \\sum_ {i} W _ {i} ^ {\\mathrm {n e w} ^ {T}} W _ {i} ^ {\\mathrm {n e w}}\\right) ^ {- 1} \\left(\\sum_ {i} W _ {i} ^ {\\mathrm {n e w} ^ {T}} W _ {i} ^ {\\mathrm {o l d}}\\right) c \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 308, + 679, + 784, + 722 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "As illustrated in the fourth column of Fig. 2, this regularized embedding can also guide UCE's model to once again generate nude images, indicating its ability to represent the concept \"nudity\" within UCE's model. With the incorporation of our regularization term, we iteratively apply the erasure process to the refined embedding $c'$ using Eq. (3) over multiple epochs. This ensures thorough concept erasure while safeguarding the overall performance of the model. The algorithm details are elaborated in Algorithm 1.", + "bbox": [ + 212, + 734, + 782, + 839 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "Reliable and Efficient Concept Erasure of Text-to-Image Diffusion Models", + "bbox": [ + 238, + 114, + 732, + 130 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "code", + "sub_type": "algorithm", + "code_caption": [ + "Algorithm 1: Erase Concepts with RECE" + ], + "code_body": "Input: Diffusion U-Net $\\theta$ , concepts set $E$ to erase, $P$ to preserve, epochs $T$ Output: Diffusion U-Net $\\theta^{\\prime}$ with concepts $E$ erased. \n1 /\\* Initialize\\*/ \n2 $\\theta^{\\prime}\\gets \\theta$ . \n3 Initialize text embeddings $c_{i}$ and $c_{j}$ from $E$ and $P$ . \n4 Extract K&V matrices $W^{\\mathrm{old}}$ from the cross attention of $\\theta$ . \n5 /\\* Preliminary erasing with UCE\\*/ \n6 Obtain $W^{\\mathrm{new}}$ by erasing concepts in $E$ with Eq. (3); \n7 for $t = 1,\\ldots ,T$ do \n8 /\\* Derive new embeddings\\*/ \n9 $E^{\\prime}\\gets \\{\\}$ . \nfor $c_{i}\\in E$ do \n11 Derive new embedding $c_i^\\prime$ using $W^{\\mathrm{new}}$ with Eq. (8); \n12 $\\begin{array}{r}E^{\\prime}\\leftarrow E^{\\prime}\\cup \\{c_{i}^{\\prime}\\} \\end{array}$ \n13 /\\* Erasing derived embeddings\\*/ \n14 Obtain $W^{\\mathrm{new}}$ by erasing $E^{\\prime}$ with Eq. (3); \n15 $W^{\\mathrm{new}}\\gets W^{\\mathrm{new}};$ \n16 Update $\\theta^{\\prime}$ by replacing K&V matrices with $W^{\\mathrm{new}}$ . \n17 return $\\theta^{\\prime}$", + "bbox": [ + 218, + 166, + 787, + 444 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 215, + 478, + 375, + 494 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this section, we present the results of our method for erasing inappropriate concepts and artistic styles. We also include the results of object removal in Appendix. We start with SD V1.4 as our base model. Following the implementation in [16], we set $\\lambda_{1}$ and $\\lambda_{2}$ in Eq. (2) to 0.1. For inappropriate concepts, we perform iterative erasure for 5 epochs and set $\\lambda$ in Eq. (7) to $1e - 1$ . For artistic style, we conduct erasure for 10 epochs and set $\\lambda$ to $1e - 3$ . The baselines we will compare with are: SD V1.4 [25], SD V2.1 [31](Stable Diffusion pretrained on an NSFW filtered dataset), SLD [29], ESD [6], CA [13], SA [8], UCE [7]. As for SLD, ESD, SA and UCE, we adhere to the recommended configuration in their papers [6-8,29]. For CA [13], we fine-tune the full weights of U-Net to erase unsafe contents and the cross-attention module to erase artistic styles, according to its documentation3.", + "bbox": [ + 212, + 510, + 787, + 691 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.1 Unsafe Content Removal", + "text_level": 1, + "bbox": [ + 215, + 714, + 468, + 728 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Experimental Setup In this section, we assess the effectiveness of erasing unsafe concepts. We conduct experiments on the Inappropriate Image Prompts (I2P) dataset [29]. The I2P dataset includes various inappropriate prompts, such as violence, self-harm, sexual content, and shocking content. These prompts are collected from real-world, user-generated images based on the official SD. Our", + "bbox": [ + 212, + 739, + 787, + 816 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "F. Chao Gong et al.", + "bbox": [ + 271, + 114, + 406, + 128 + ], + "page_idx": 7 + }, + { + "type": "page_footnote", + "text": "3 https://github.com/nupurkmr9/concept-ablation", + "bbox": [ + 217, + 824, + 584, + 840 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/46d37a4fa79add357abb9f51c32db6b6803b6d5aea56801b2c80a1439f0ad4d5.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodNudity DetectionCOCO-30k
Breast(F)Genitalia(F)Breast(M)Genitalia(M)ButtocksFeetBellyArmpitsTotal↓CLIP↑FID↓
SD v1.4183214610444217112964631.33-
SD v2.1121134031439146109485--
ESD-u14185524313312130.453.73
UCE3161981120553618631.261.82
SLD-Med72534318191049935430.952.60
SA399401532491516330.5717.34
CA6191041428239531.167.87
Ours80640823176630.952.82
", + "bbox": [ + 218, + 143, + 782, + 268 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Table 1: Comparison of performance metrics for content removal methods. (Left) Number of nude body parts detected by Nudenet on I2P dataset with threshold 0.6. (Right) CLIP-score and FID against original SD. F: Female. M: Male. Bold: best. Underline: second-best.", + "bbox": [ + 215, + 268, + 784, + 324 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "evaluation focuses on the erasure of nudity since it is a classical unsafe concept. For each model, we generate one image per prompt in the I2P dataset, resulting in a total of 4703 images. Nude body parts are detected using the Nudenet detector [18], with the threshold set to 0.6. This threshold follows the default settings in $\\mathrm{I2P^4}$ .", + "bbox": [ + 212, + 369, + 782, + 444 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "To verify that the unlearned models can still generate normal images, we use COCO-30k [14] with its captions as prompts. COCO-30k is a dataset devoid of unsafe concepts, making it suitable for evaluating edited models' generation capabilities. We evaluate the models' image-text consistency based on CLIP-score [9], and visual similarity against SD-generated images based on FID [17].", + "bbox": [ + 212, + 446, + 784, + 523 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Removal Results As depicted in Tab. 1, our method yields the lowest number of nude body parts, while demonstrating impressive specificity in preserving normal content of COCO-30k. CA generates the second-fewest nude body parts but it exhibits poorer performance in terms of FID, indicating a poorer trade-off between generation ability and removal effectiveness. On the other hand, CA fine-tunes full weights and our RECE only fine-tunes cross-attention modules, which will be discussed in detail in Sec. 4.4.", + "bbox": [ + 212, + 551, + 784, + 655 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Notably, our method achieves a FID score closely comparable to the top-performing UCE and the second-best SLD, both of which generates a considerably higher number of nude body parts. This suggests that our method minimally impacts the generation of normal content while striving for better removal effectiveness. Additionally, most methods exhibits favorable CLIP-score results thus we consider the CLIP-score performance acceptable as long as it remains within a reasonable range.", + "bbox": [ + 212, + 657, + 784, + 763 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "In open-sourced conditions, the inference guidance mechanism such as SLD [29] can be easily bypassed by deleting the corresponding code [24]. Large-scale model retraining on NSFW-filtered dataset demands considerable computational re", + "bbox": [ + 212, + 765, + 792, + 811 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "Reliable and Efficient Concept Erasure of Text-to-Image Diffusion Models", + "bbox": [ + 240, + 114, + 730, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "page_footnote", + "text": "4 https://github.com/ml-research/i2p", + "bbox": [ + 217, + 824, + 500, + 840 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/a8f83f03d442eab7b7a57107a1913358fbae6b46c5ab152a0b9cd23ee6385333.jpg", + "image_caption": [ + "SD v1.4" + ], + "image_footnote": [], + "bbox": [ + 222, + 143, + 289, + 244 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/6275885888c8026fe5df7b6b96838c6ae6503bb8bfd86262ccbef83e1bd21a3d.jpg", + "image_caption": [ + "SD v2.1" + ], + "image_footnote": [], + "bbox": [ + 292, + 143, + 359, + 244 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/aab5cba63815acac99ab09c739109b4d533a5cc480081220d4817399f1e3a8d0.jpg", + "image_caption": [ + "ESD" + ], + "image_footnote": [], + "bbox": [ + 362, + 143, + 429, + 244 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/7a07d949f919d07dedb8c0220f1b59c63a6bd00d8dd0ec66c1d2fcfe32796fe1.jpg", + "image_caption": [ + "UCE", + "Fig. 3: Qualitative results of different nudity removing methods. Images for each row are generated using the same prompt. All prompts originate from I2P benchmark, and we don't intentionally generate inappropriate content of any individual." + ], + "image_footnote": [], + "bbox": [ + 433, + 143, + 500, + 244 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/a6f700265c773b2839d4c4b7851947f130832029ca90abf94f07f7bd312802fd.jpg", + "image_caption": [ + "SLD-Med" + ], + "image_footnote": [], + "bbox": [ + 501, + 143, + 570, + 244 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/c6e74b0bff7785af9f1dcd9330dfc281a6b443797041c536576602838384fe65.jpg", + "image_caption": [ + "SA" + ], + "image_footnote": [], + "bbox": [ + 573, + 143, + 640, + 244 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/beac8009cce77f2f1f1d6ebf31b54ab49ffa7ce51fd6d950f27bcfa24577e05a.jpg", + "image_caption": [ + "CA" + ], + "image_footnote": [], + "bbox": [ + 642, + 143, + 710, + 244 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/9a531224f258d0878a7814fa61bf3b215c0927b55a26d99e254ff4c5820a0fba.jpg", + "image_caption": [ + "Ours" + ], + "image_footnote": [], + "bbox": [ + 712, + 143, + 779, + 244 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "sources [22], and even then, the model SD v2.1 may still generate nude images, as illustrated in Tab. 1.", + "bbox": [ + 212, + 357, + 782, + 386 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Qualitative results are illustrated in Fig. 3. In the first row, all erasing methods successfully generate non-nude images, but the results of CA and SA differ significantly from the original SD's. In the second row, all methods except for SD v2.1 avoid generating nude images. Among these, only UCE, SLD, and our method effectively capture the facial features. These findings demonstrate that our method effectively maintains unrelated concepts.", + "bbox": [ + 212, + 387, + 784, + 479 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Nudity Bias While our RECE demonstrates remarkable effectiveness in minimizing the generation of nude content, it exhibits limitations in erasing male nudity and similar limitations are also observed in other methods. We count the number of female and male nude body parts in the 4703 I2P images. As illustrated in Tab. 2, the nudity ratios between women and men almost decrease in every method, indicating an erasure bias on female information (excluding SA, which erases more sex-related concepts besides nudity [8]). We attribute this limitation to the inherent bias in the target concept \"nudity\" within SD, which tends to generate more female-oriented content. To investigate this bias, we randomly selected 20 seeds and employed SD V1.4 to generate 3 images per seed with the prompt \"nudity\", resulting in a total of 60 images. Surprisingly, almost all of these generated images depict female body part as presented in the last column of Tab. 2. Thus our derived embedding is also biased. Further improvements require an awareness of the biases inherent in the model while performing erasure.", + "bbox": [ + 212, + 502, + 787, + 729 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.2 Artistic Style Removal", + "text_level": 1, + "bbox": [ + 215, + 752, + 450, + 767 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Experimental Setup We conduct an evaluation to assess the efficacy of removing artistic styles to address copyright concerns. Following the datasets in ESD [6], we use 20 prompts for each of 5 famous artists—Van Gogh, Pablo Picasso, Rembrandt, Andy Warhol and Caravaggio—and 5 modern artists—Kelly", + "bbox": [ + 212, + 779, + 785, + 840 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "F. Chao Gong et al.", + "bbox": [ + 271, + 114, + 406, + 128 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/88ab8b50b76590323b03c6ff579cb88b872e2c2c19cb78d3b44f64e7d22ef99a.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MetricsSDESDUCESLD-MedSACAOursSD- "nudity"
# Female Nudity204153777487889
# Male Nudity56132737419101
Female/Male3.641.151.372.08120.370.889
", + "bbox": [ + 263, + 143, + 738, + 214 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/46fc7bfb5a0d212e38360d642b63b68e5cf89e9e4857d863ccd8078e3029aa3f.jpg", + "table_caption": [ + "Table 2: In the first 7 columns, we counted the numbers of female nudity and male nudity body parts for I2P dataset. The last column is the result of SD v1.4 conditioned by the prompt \"nudity\". This table highlights the inherent bias within SD. And all each erasure methods can not remove the concept of male nudity very well." + ], + "table_footnote": [], + "table_body": "
Removal MethodErase "Van Gogh"Erase "Kelly McKernan"
LPIPSe↑LPIPSu↓LPIPSd↑LPIPSe↑LPIPSu↓LPIPSd↑
ESD-x0.400.260.140.370.210.16
UCE0.250.050.200.250.030.22
SLD-Med0.210.100.110.220.180.04
CA0.300.130.170.220.170.05
Ours0.310.080.230.290.040.25
", + "bbox": [ + 263, + 294, + 741, + 402 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Table 3: Comparison of LPIPS scores for artistic removal methods. Bold: best. Underline: second-best. $\\mathrm{LPIPS_d}$ indicates overall erasure performance.", + "bbox": [ + 212, + 402, + 784, + 430 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "McKernan, Thomas Kinkade, Tyler Edlin, Kilian Eng and the series \"Ajin: Demi-Human\", which have been reported to be imitated by SD [30]. To evaluate our RECE and all the aforementioned baselines, we erase the style of two artists: Van Gogh and Kelly McKernan.", + "bbox": [ + 212, + 470, + 784, + 532 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Removal Results We conducted an evaluation based on LPIPS scores [37] compared to the original SD, as detailed in Tab. 3. LPIPS evaluates the perceptual distance between image patches, where higher values indicate greater differences and lower values indicate more similarity. The $\\mathrm{LPIPS_e}$ is calculated on the erased artist. A higher $\\mathrm{LPIPS_e}$ value suggests a more effective style removal, and both ESD and our method demonstrate successful erasure of the target style. $\\mathrm{LPIPS_u}$ is calculated on unerased artists. A lower $\\mathrm{LPIPS_u}$ indicates a lesser impact on unrelated artists, where our method and UCE effectively maintains unrelated concepts. We also calculate the overall effectiveness by $\\mathrm{LPIPS_d} = \\mathrm{LPIPS_e} - \\mathrm{LPIPS_u}$ , which is the difference between erased and unerased artists. Our method performs best in this regard. Qualitative results can be found in Appendix.", + "bbox": [ + 212, + 553, + 787, + 734 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4.3 Robustness Against Red-teaming Tools", + "text_level": 1, + "bbox": [ + 214, + 756, + 584, + 771 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Experimental Setup To demonstrate the robustness of our RECE in safeguarding against various attack methods, we employ different red-teaming tools, including white-box methods such as P4D [3] and UnlearnDiff [38], and the black-box method Ring-A-Bell [32]. For nudity, as provided by UnlearnDiff [38],", + "bbox": [ + 212, + 779, + 785, + 840 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "Reliable and Efficient Concept Erasure of Text-to-Image Diffusion Models", + "bbox": [ + 238, + 113, + 730, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/99c545c1dc29393fd647580017cc266a66ad743546f11436cdfcce1d4821b74f.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Red-teamingSD v1.4SD v2.1ESD-uUCESLD-MaxSACAOurs
UnlearnDiff--66.2079.5882.3977.4665.4965.49
P4D--63.3880.2877.4678.8760.5664.79
Ring-A-Bell83.1072.5469.7233.1066.2022.5425.3513.38
Average--66.4364.3275.3559.6250.4747.89
", + "bbox": [ + 263, + 143, + 738, + 231 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Table 4: Robustness of different methods against red-teaming tools, measured by attack success rate(%) . Bold: best. Underline: second-best.", + "bbox": [ + 215, + 232, + 782, + 258 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/b4f3e977ed9cae46da51871eed38732622a37cb76ef6a80a31a06e14db7f2c19.jpg", + "image_caption": [ + "Original SD", + "Fig. 4: Qualitative results of different artistic style removal methods for eliminating the style of Van Gogh under Ring-A-Bell's attack. The prompt is \"A depiction of a starry night over a quiet town, reminiscent of Van Gogh's famous painting\"." + ], + "image_footnote": [], + "bbox": [ + 282, + 284, + 356, + 339 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/3c1c6670da3aef5f747eebd760738458c054f438a4104715ab56c75191d6e383.jpg", + "image_caption": [ + "ESD" + ], + "image_footnote": [], + "bbox": [ + 357, + 285, + 426, + 339 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/f0921bb0eb7a2e44276d351feb20046c2d7bccf655d6781c8f472aefe08287b5.jpg", + "image_caption": [ + "UCE" + ], + "image_footnote": [], + "bbox": [ + 429, + 285, + 500, + 338 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/54906c9f9f6e704295990d1c6fcf3f84456d64e8e201f51a0003e3b04c748307.jpg", + "image_caption": [ + "SLD-Med" + ], + "image_footnote": [], + "bbox": [ + 503, + 285, + 573, + 338 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/6c108cc06dcd28666d0902f3554b48090f60776ed47d1e4779b7c7cb629c7349.jpg", + "image_caption": [ + "CA" + ], + "image_footnote": [], + "bbox": [ + 575, + 285, + 645, + 338 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/9cf2bdccd184a9659d37be6304ae5d6463b2cf4c9d41dd534f318737f24a6205.jpg", + "image_caption": [ + "Ours" + ], + "image_footnote": [], + "bbox": [ + 648, + 285, + 718, + 338 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "we use a set of 143 prompts selected from I2P, each with a nudity score (as determined by NudeNet) above 0.75, and employ the Nudenet detector with a threshold set to 0.45 for detecting inappropriate content.", + "bbox": [ + 214, + 431, + 782, + 477 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Results The attack success rates (ASR, %) are summarized in Tab. 4. Our method achieves the best robustness in average. In the case of the black-box attack Ring-A-Bell, our method achieves the lowest ASR at $13.38\\%$ , significantly outperforming other methods. In the case of the white-box attack, CA achieves the best performance, while our method performs either the best or very closely to the second-best. Although SA achieves a decent black-box result, it consumes substantial computation resources for generating 5000 prepared images [8]. CA modifies $100\\%$ of the U-Net parameters while our method only modifies K&V projection matrices, constituting a mere $2.23\\%$ of the U-Net. While UCE also modifies only $2.23\\%$ parameters, our method significantly outperforms UCE. This is attributed to our derived embeddings in Eq. (8).", + "bbox": [ + 212, + 496, + 784, + 662 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "For artistic style removal, we provide qualitative examples in Fig. 4. The first image is generated by the original SD v1.4 without any attack, and the following images are under Ring-A-Bell's attack [32]. Recall that UCE is the second-best artistic style removal method as shown in Tab. 3 but it falls short in robustness against red-teaming tools. Although our method employs a closed-form solution like UCE, it outperforms UCE in robustness. ESD, CA and our method perform similarly well. More results can be found in the appendix.", + "bbox": [ + 212, + 662, + 784, + 768 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "4.4 Model Editing Duration", + "text_level": 1, + "bbox": [ + 214, + 787, + 460, + 801 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "To demonstrate the efficiency of different methods, we measured the percentage of parameter modification and editing duration on an RTX 3090 for each", + "bbox": [ + 214, + 809, + 782, + 839 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "F. Chao Gong et al.", + "bbox": [ + 271, + 114, + 406, + 128 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/9e64228585d90f0bf25041bb9f342286090fb94826a285d84c8cb8d1b1a0169d.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ESDUCESACAOurs
Modification (%)94.652.2394.651002.23
Duration (s)37201.2-14003.4
", + "bbox": [ + 348, + 143, + 655, + 199 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Table 5: Percentage of parameter modification and model editing duration for different methods. Our method and UCE are significantly ahead of other methods.", + "bbox": [ + 215, + 199, + 782, + 227 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "method, as shown in Tab. 5. We excluded SLD from the analysis since it operates at inference time rather than modifying the model's weights which can be easily bypassed under open-source conditions. Additionally, we don't include the duration of SA, as it involves generating 5000 images, calculating the Fisher Information Matrix and fine-tuning, which makes it exceptionally slow.", + "bbox": [ + 212, + 270, + 784, + 345 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Based on Tab. 5 and Tab. 1, our method achieves the best concept erasure effect in an extremely short time of only three seconds. Our method (5 epochs) and UCE (1 epoch) modify the lowest percentage of parameters with a closed-form solution, resulting in the shortest editing durations. Despite similar durations, our method significantly outperforms UCE in removal effectiveness. Conversely, CA, ESD and SA modify a high percentage of parameters with more time but achieve less impressive removal results.", + "bbox": [ + 212, + 345, + 785, + 452 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "4.5 Ablation Study", + "text_level": 1, + "bbox": [ + 215, + 474, + 387, + 489 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "We conduct experiments to elucidate the impact of our derived embedding among different epochs and the effectiveness of the regularization term.", + "bbox": [ + 214, + 500, + 784, + 530 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Effect of Derived Embeddings among Epochs", + "text_level": 1, + "bbox": [ + 215, + 551, + 413, + 583 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "We conduct an experiment to expound the impact of our derived embedding. We perform \"nudity\" erasure for 5 epochs using Algorithm 1. In each epoch, we derive a distinct embedding that represents \"nudity\". Before the erasure of each", + "bbox": [ + 212, + 583, + 403, + 732 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "epoch, we generate an image using the embedding to test its degree of nudity information, as presented in Fig. 5. Images from epoch 0 to epoch 2 contain nude body parts, indicating that our derived embeddings successfully reveal potential nudity information in the model. Specifically, we opt for the checkpoint after epoch 2, as images from epoch 3 to 4 lack nudity information. Actually, erasing such \"not so nude\" embeddings in epoch 3-4 would impair the model's normal generation ability, which is an unworthy trade-off.", + "bbox": [ + 212, + 734, + 785, + 839 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/a6cd6fddd50fd3f3a1017cdb441762c6eabcc41be1c916ab35ab5ac81f034c98.jpg", + "image_caption": [ + "Fig. 5: There is no nudity information from epoch 3 onward; hence, we select the checkpoint after epoch 2 to avoid damaging the model's ability." + ], + "image_footnote": [], + "bbox": [ + 424, + 584, + 777, + 654 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "Reliable and Efficient Concept Erasure of Text-to-Image Diffusion Models", + "bbox": [ + 238, + 113, + 730, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/09f9728f5f221f3a57c35ee85a361d4e48298663c7ed1c8bb330ce059fead620.jpg", + "image_caption": [ + "Fig. 6: Ablation study on the impact of the regularization term on model performance." + ], + "image_footnote": [], + "bbox": [ + 261, + 145, + 743, + 234 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Effect of Regularization Coefficient We conduct an experiment to assess the influence of our regularization term. Specifically, we select different regularization coefficients $\\lambda$ in Eq. (4), which divide the interval [0, 1] into five parts. The results, presented in Fig. 6, include CLIP-score and LPIPS against the original SD on the COCO-30k validation subset. As $\\lambda$ increases, CLIP score shows an upward trend while LPIPS and the difference between new and old parameters show a downward trend. This indicates the role of the regularization term in preserving the model's ability for unerased content. Furthermore, we recorded the number of nude parts on the I2P benchmark, presented in the third column of Fig. 6. However, the number of nude parts doesn't strictly increase as $\\lambda$ increases, which is counterintuitive. Although the purpose of the regularization term in Eq. (6) is to preserve the model's generation capability, maintaining this capability does not always affect the erasure effect.", + "bbox": [ + 212, + 301, + 787, + 498 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 544, + 359, + 559 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In this paper, we propose a novel approach for reliably and efficiently erasing specific concepts from Text-to-Image (T2I) diffusion models. Our approach only modifies the cross-attention K&V matrices of U-Net, constituting a mere $2.23\\%$ of parameters. While previous methods also edited cross-attention modules, they still exhibited the ability to generate inappropriate images. To tackle this challenge, we derive and erase new embeddings that can represent target concepts within unlearned models. To mitigate the impact on unrelated concepts, a regularization term is introduced during the erasure process. All the above techniques are formulated in closed-form, facilitating rapid editing. This enables the execution of \"derive-erase\" across multiple epochs, ensuring thorough and robust erasure. Extensive experiments were conducted to validate the effectiveness of our approach in erasing artistic styles, unsafe contents and object classes. Furthermore, we recorded editing durations to underscore the efficiency of our method and evaluated the robustness against red-teaming tools. We believe our RECE has the potential to empower T2I providers in effectively removing undesired concepts, thereby fostering the development of a safer AI community.", + "bbox": [ + 212, + 598, + 787, + 843 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "F. Chao Gong et al.", + "bbox": [ + 271, + 114, + 406, + 128 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 217, + 145, + 401, + 162 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "This project was supported by National Key R&D Program of China (No. 2021ZD0112804).", + "bbox": [ + 212, + 178, + 782, + 209 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 215, + 233, + 323, + 250 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Chen, K., Wei, Z., Chen, J., Wu, Z., Jiang, Y.G.: Attacking video recognition models with bullet-screen comments. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 36, pp. 312-320 (2022)", + "2. Chen, K., Wei, Z., Chen, J., Wu, Z., Jiang, Y.G.: Gcma: Generative cross-modal transferable adversarial attacks from images to videos. In: Proceedings of the 31st ACM International Conference on Multimedia. pp. 698-708 (2023)", + "3. Chin, Z.Y., Jiang, C.M., Huang, C.C., Chen, P.Y., Chiu, W.C.: Prompting4debugging: Red-teaming text-to-image diffusion models by finding problematic prompts. arXiv preprint arXiv:2309.06135 (2023)", + "4. Dhariwal, P., Nichol, A.: Diffusion models beat gans on image synthesis. Advances in neural information processing systems 34, 8780-8794 (2021)", + "5. Doersch, C.: Tutorial on variational autoencoders. arXiv preprint arXiv:1606.05908 (2016)", + "6. Gandikota, R., Materzynska, J., Fiotto-Kaufman, J., Bau, D.: Erasing concepts from diffusion models. arXiv preprint arXiv:2303.07345 (2023)", + "7. Gandikota, R., Orgad, H., Belinkov, Y., Materzyńska, J., Bau, D.: Unified concept editing in diffusion models. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 5111-5120 (2024)", + "8. Heng, A., Soh, H.: Selective amnesia: A continual learning approach to forgetting in deep generative models. Advances in Neural Information Processing Systems 36 (2024)", + "9. Hessel, J., Holtzman, A., Forbes, M., Bras, R.L., Choi, Y.: Clipscore: A reference-free evaluation metric for image captioning. arXiv preprint arXiv:2104.08718 (2021)", + "0. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. Advances in neural information processing systems 33, 6840-6851 (2020)", + "1. Hunter, T.: Ai porn is easy to make now. for women, that's a nightmare. (2 2023)", + "2. Kawar, B., Zada, S., Lang, O., Tov, O., Chang, H., Dekel, T., Mosseri, I., Irani, M.: Imagic: Text-based real image editing with diffusion models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6007-6017 (2023)", + "3. Kumari, N., Zhang, B., Wang, S.Y., Shechtman, E., Zhang, R., Zhu, J.Y.: Ablating concepts in text-to-image diffusion models. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 22691-22702 (2023)", + "4. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollar, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13. pp. 740-755. Springer (2014)", + "5. Nichol, A.Q., Dhariwal, P., Ramesh, A., Shyam, P., Mishkin, P., Mcgrew, B., Sutskever, I., Chen, M.: Glide: Towards photorealistic image generation and editing with text-guided diffusion models. In: International Conference on Machine Learning. pp. 16784-16804. PMLR (2022)" + ], + "bbox": [ + 225, + 266, + 785, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "Reliable and Efficient Concept Erasure of Text-to-Image Diffusion Models", + "bbox": [ + 238, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 785, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "16. Orgad, H., Kawar, B., Belinkov, Y.: Editing implicit assumptions in text-to-image diffusion models. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 7053-7061 (2023)", + "17. Parmar, G., Zhang, R., Zhu, J.Y.: On aliased resizing and surprising subtleties in gan evaluation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11410-11420 (2022)", + "18. Praneeth, B.: Nudenet: Neural nets for nudity classification, detection and selective censoring (2019)", + "19. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021)", + "20. Ramesh, A., Dhariwal, P., Nichol, A., Chu, C., Chen, M.: Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125 1(2), 3 (2022)", + "21. Rando, J., Paleka, D., Lindner, D., Heim, L., Tramér, F.: Red-teaming the stable diffusion safety filter. arXiv preprint arXiv:2210.04610 (2022)", + "22. Rombach, R.: Stable diffusion 2.0 release (November 2022)", + "23. Rombach, R.: Tutorial: How to remove the safety filter in 5 seconds (August 2022)", + "24. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10684-10695 (2022)", + "25. Rombach, R., Esser, P.: Stable diffusion v1-4 model card. Model Card (2022)", + "26. Ronneberger, O., Fischer, P., Brox, T.: U-net: Convolutional networks for biomedical image segmentation. In: Medical image computing and computer-assisted intervention-MICCAI 2015: 18th international conference, Munich, Germany, October 5-9, 2015, proceedings, part III 18. pp. 234-241. Springer (2015)", + "27. Ruiz, N., Li, Y., Jampani, V., Pritch, Y., Rubinstein, M., Aberman, K.: Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 22500-22510 (2023)", + "28. Saharia, C., Chan, W., Saxena, S., Li, L., Whang, J., Denton, E.L., Ghasemipour, K., Gontijo Lopes, R., Karagol Ayan, B., Salimans, T., et al.: Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems 35, 36479-36494 (2022)", + "29. Schramowski, P., Brack, M., Deiseroth, B., Kersting, K.: Safe latent diffusion: Mitigating inappropriate degeneration in diffusion models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 22522-22531 (2023)", + "30. Setty, R.: Ai art generators hit with copyright suit over artists' images (1 2023)", + "31. StabilityAI: Stable diffusion 2.1 model card. Model Card (2022)", + "32. Tsai, Y.L., Hsu, C.Y., Xie, C., Lin, C.H., Chen, J.Y., Li, B., Chen, P.Y., Yu, C.M., Huang, C.Y.: Ring-a-bell! how reliable are concept removal methods for diffusion models? arXiv preprint arXiv:2310.10012 (2023)", + "33. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. Advances in neural information processing systems 30 (2017)", + "34. Wei, Z., Chen, J., Goldblum, M., Wu, Z., Goldstein, T., Jiang, Y.G., Davis, L.S.: Towards transferable adversarial attacks on image and video transformers. IEEE Transactions on Image Processing 32, 6346-6358 (2023)" + ], + "bbox": [ + 215, + 147, + 784, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "F. Chao Gong et al.", + "bbox": [ + 271, + 114, + 405, + 127 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "35. Wei, Z., Chen, J., Wu, Z., Jiang, Y.G.: Adaptive cross-modal transferable adversarial attacks from images to videos. IEEE Transactions on Pattern Analysis and Machine Intelligence (2023)", + "36. Yang, L., Zhang, Z., Song, Y., Hong, S., Xu, R., Zhao, Y., Zhang, W., Cui, B., Yang, M.H.: Diffusion models: A comprehensive survey of methods and applications. ACM Computing Surveys 56(4), 1-39 (2023)", + "37. Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual metric. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 586-595 (2018)", + "38. Zhang, Y., Jia, J., Chen, X., Chen, A., Zhang, Y., Liu, J., Ding, K., Liu, S.: To generate or not? safety-driven unlearned diffusion models are still easy to generate unsafe images... for now. arXiv preprint arXiv:2310.11868 (2023)" + ], + "bbox": [ + 215, + 146, + 787, + 314 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "Reliable and Efficient Concept Erasure of Text-to-Image Diffusion Models", + "bbox": [ + 238, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 16 + } +] \ No newline at end of file diff --git a/2024/Reliable and Efficient Concept Erasure of Text-to-Image Diffusion Models/d5988da3-2c65-4cde-838f-fe6bf86e8cd3_model.json b/2024/Reliable and Efficient Concept Erasure of Text-to-Image Diffusion Models/d5988da3-2c65-4cde-838f-fe6bf86e8cd3_model.json new file mode 100644 index 0000000000000000000000000000000000000000..a44bc78ab45f937d1fac195df91546916a3d7ac8 --- /dev/null +++ b/2024/Reliable and Efficient Concept Erasure of Text-to-Image Diffusion Models/d5988da3-2c65-4cde-838f-fe6bf86e8cd3_model.json @@ -0,0 +1,2445 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.266, + 0.141, + 0.74, + 0.186 + ], + "angle": 0, + "content": "Reliable and Efficient Concept Erasure of Text-to-Image Diffusion Models" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.212, + 0.788, + 0.245 + ], + "angle": 0, + "content": "Chao Gong\\(^{1,2\\star}\\), Kai Chen\\(^{1,2\\star}\\), Zhipeng Wei\\(^{1,2}\\), Jingjing Chen\\(^{1,2\\dagger}\\), and Yu-Gang Jiang\\(^{1,2}\\)" + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.255, + 0.778, + 0.283 + ], + "angle": 0, + "content": "\\(^{1}\\) Shanghai Key Lab of Intell. Info. Processing, School of Computer Science, Fudan University" + }, + { + "type": "text", + "bbox": [ + 0.222, + 0.283, + 0.782, + 0.311 + ], + "angle": 0, + "content": "\\(^{2}\\) Shanghai Collaborative Innovation Center on Intelligent Visual Computing {c Gong20, chenjingjing, ygj} @fudan.edu.cn, {kchen22, zpwei21} @m.fudan.edu.cn" + }, + { + "type": "list", + "bbox": [ + 0.222, + 0.255, + 0.782, + 0.311 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.351, + 0.74, + 0.669 + ], + "angle": 0, + "content": "Abstract. Text-to-image models encounter safety issues, including concerns related to copyright and Not-Safe-For-Work (NSFW) content. Despite several methods have been proposed for erasing inappropriate concepts from diffusion models, they often exhibit incomplete erasure, consume a lot of computing resources, and inadvertently damage generation ability. In this work, we introduce Reliable and Efficient Concept Erasure (RECE), a novel approach that modifies the model in 3 seconds without necessitating additional fine-tuning. Specifically, RECE efficiently leverages a closed-form solution to derive new target embeddings, which are capable of regenerating erased concepts within the unlearned model. To mitigate inappropriate content potentially represented by derived embeddings, RECE further aligns them with harmless concepts in crossattention layers. The derivation and erasure of new representation embeddings are conducted iteratively to achieve a thorough erasure of inappropriate concepts. Besides, to preserve the model's generation ability, RECE introduces an additional regularization term during the derivation process, resulting in minimizing the impact on unrelated concepts during the erasure process. All the processes above are in closed-form, guaranteeing extremely efficient erasure in only 3 seconds. Benchmarking against previous approaches, our method achieves more efficient and thorough erasure with minor damage to original generation ability and demonstrates enhanced robustness against red-teaming tools. Code is available at https://github.com/CharlesGong12/RECE." + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.67, + 0.733, + 0.683 + ], + "angle": 0, + "content": "WARNING: This paper contains model outputs that may be offensive." + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.697, + 0.719, + 0.711 + ], + "angle": 0, + "content": "Keywords: Text-to-Image \\(\\cdot\\) Concept Erasing \\(\\cdot\\) Machine Unlearning" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.738, + 0.376, + 0.753 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.771, + 0.797, + 0.802 + ], + "angle": 0, + "content": "In recent years, large-scale text-to-image (T2I) diffusion models have exhibited remarkable capability in synthesizing photo-realistic images from text prompts [15," + }, + { + "type": "page_footnote", + "bbox": [ + 0.233, + 0.812, + 0.384, + 0.825 + ], + "angle": 0, + "content": "* Equal contributions." + }, + { + "type": "page_footnote", + "bbox": [ + 0.233, + 0.825, + 0.396, + 0.84 + ], + "angle": 0, + "content": "† Corresponding author." + }, + { + "type": "list", + "bbox": [ + 0.233, + 0.812, + 0.396, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.406, + 0.129 + ], + "angle": 0, + "content": "F. Chao Gong et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.375 + ], + "angle": 0, + "content": "[20, 24, 28]. The exceptional performance of T2I diffusion models is largely due to the vast amount of training data collected from the Internet, which enables the models to imitate a wide variety of concepts. Unfortunately, such powerful models can also be misused to generate copyright infringement and Not-Safe-For-Work (NSFW) image content when conditioned on inappropriate text prompts [11, 30]. Especially the open-source release of the Stable Diffusion (SD) T2I model has made advanced image generation technology widely accessible. To alleviate this safety concern, several recent research efforts have incorporated safety mechanisms into T2I diffusion models, e.g. filtering out inappropriate training data and retraining model [22], censoring model outputs with an NSFW safety checker [21], and applying classifier-free guidance to steer the generation away from inappropriate concepts [29]. However, these safety mechanisms either demand expensive computational resources and time [25] or can be easily circumvented by malicious users due to the public availability of code and model parameters in open-source scenario [23]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.381, + 0.788, + 0.637 + ], + "angle": 0, + "content": "In response to the drawbacks mentioned above, an alternative is to erase inappropriate concepts from the T2I diffusion model [6-8, 13]. Specifically, given an inappropriate concept described in the text prompt, the pre-trained T2I diffusion model's parameters are fine-tuned to unlearn that concept so that the associated image content cannot be generated. Compared with previous security mechanisms, concept erasure neither requires training the entire model from scratch nor can be easily circumvented even in the case of open-source code. Despite promising progress in concept erasure, there exist several issues. On the one hand, most erasure methods require a high number of iterations to finetune considerable amounts of parameters [6, 8, 13], which inevitably degrades the generation capability and consumes a lot of computing resources. Only a recent work called UCE [7] modifies model parameters without fine-tuning using a closed-form solution, ensuring the model maintains original generation capability when erasing concepts. On the other hand, almost all methods fail to sufficiently erase inappropriate concepts, leaving them vulnerable to problematic prompts found by the red-teaming of T2I diffusion models [3, 32, 38]. This results in the unlearned model being compelled to regenerate inappropriate images." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.644, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Inspired by the idea of adversarial fine-tuning, we propose a Reliable and Efficient Concept Erasure (RECE) method to address the aforementioned challenge, which continually finds new embeddings of the erased concepts during fine-tuning and then enables the unlearned model to erase these new concept embeddings. To speed up the unlearning process, the RECE method builds upon the previous fast and efficient concept erasure method UCE [7], which employs a closed-form editing to only modify the key and value projection matrices in cross-attention layers [33]. Similarly, the RECE method derives new embeddings that most effectively prompt the model to regenerate images of erased concepts, with a closed-form solution based on cross-attention output. Furthermore, a regularization term is introduced to preserve the image generation capability of the model by restricting the deviation of model parameters before and after modification. By editing the model and deriving embedding for multiple epochs," + } + ], + [ + { + "type": "header", + "bbox": [ + 0.24, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Reliable and Efficient Concept Erasure of Text-to-Image Diffusion Models" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.222 + ], + "angle": 0, + "content": "RECE enables the unlearned model to preserve the image generation ability of unerased concepts and robustly refrains the model from generating images with erased concept content. All the processes above are in closed-form, guaranteeing extremely efficient erasure in 3 seconds. Our major contributions are summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.227, + 0.233, + 0.784, + 0.277 + ], + "angle": 0, + "content": "- We present a novel concept erasure method - RECE that uses closed-form parameter editing and adversarial learning schemes for reliable and efficient concept erasing in only 3 seconds." + }, + { + "type": "text", + "bbox": [ + 0.227, + 0.279, + 0.784, + 0.322 + ], + "angle": 0, + "content": "- RECE sufficiently erases concepts by deriving new embeddings that enable the unlearned model to regenerate erased concepts. In addition, a regularization term is introduced to minimize the impact on the model's capabilities." + }, + { + "type": "text", + "bbox": [ + 0.228, + 0.324, + 0.784, + 0.381 + ], + "angle": 0, + "content": "- We conduct extensive experiments to validate the effectiveness of RECE for erasing unsafe contents, artistic styles and object classes. Additionally, we assess the robustness of RECE against three red-teaming tools and record fine-tuning durations to highlight the efficiency of RECE." + }, + { + "type": "list", + "bbox": [ + 0.227, + 0.233, + 0.784, + 0.381 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.408, + 0.388, + 0.423 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.44, + 0.789, + 0.712 + ], + "angle": 0, + "content": "T2I Diffusion Models with Safety Mechanisms. In response to the issue of generating inappropriate images in T2I diffusion models, several research has explored solutions to address this concern. Briefly, existing research primarily falls into the following three distinct strategies: The first is filtering the training data and retraining the model [22]. However, retraining on curated datasets not only requires substantial computational resources investment but also results in the generation of inappropriate content [6] and performance degradation [29]. The second is censoring model output through safety checkers [21], or exploiting classifier-free guidance to steer the latent codes away from inappropriate concepts during inference [29]. However, in the case of open-source code, pretrained T2I diffusion model architectures and parameters are publicly available, so such post-hoc intervention strategies can be easily circumvented by malicious users [23]. The third is fine-tuning the partial parameters of the pre-trained T2I diffusion models to erase the model's representation capability of inappropriate concepts [6-8, 13]. While fine-tuning has been considered an effective strategy to prevent the generation of inappropriate content, current methods consume a lot of computing time and can be easily bypassed by red-teaming tools for T2I diffusion models." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.734, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Red-Teaming Tools for T2I Diffusion Models. With the recent popularity of AI, red-teaming has been applied to AI models to enhance model stability by probing functional vulnerabilities [1, 2, 34, 35]. Recent works have also developed red-teaming tools for T2I diffusion models, which is a rarely explored field in AI red-teaming. For instance, Prompting4Debugging (P4D) [3] automatically finds the problematic prompts that would lead to inappropriate content via utilizing prompt engineering techniques and an auxiliary diffusion model" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.406, + 0.129 + ], + "angle": 0, + "content": "F. Chao Gong et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.299 + ], + "angle": 0, + "content": "without any safety mechanisms to assess the reliability of deployed safety mechanisms. Conversely, UnlearnDiff [38] does not depend on an auxiliary diffusion model, which leverages the inherent classification capabilities of diffusion models, thereby providing computational efficiency without sacrificing effectiveness. Both works have the main weakness in assuming white-box access to the target model. In response to this issue, Ring-A-Bell [32], a model-agnostic framework capable of constructing attacks without prior knowledge of the target model. Specifically, Ring-A-Bell first performs concept extraction to obtain a holistic representation of inappropriate concepts. Subsequently, Ring-A-Bell automatically produces problematic prompts by leveraging the extracted concepts." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.322, + 0.331, + 0.339 + ], + "angle": 0, + "content": "3 Method" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.355, + 0.373, + 0.37 + ], + "angle": 0, + "content": "3.1 Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.381, + 0.789, + 0.608 + ], + "angle": 0, + "content": "Text-to-Image (T2I) Diffusion Models In contemporary Text-to-Image (T2I) applications, diffusion models have become the preferred choice [36] since the progressive denoising process [10] empowers them with superior image synthesis ability [4]. To reduce computational complexity, T2I often adopts latent diffusion models [12, 24, 27, 28], which operates on the low-dimensional latent space of a pre-trained variational autoencoder (VAE) [5] and employs a U-Net generative network as the denoising architecture [26]. To incorporate text conditioning into the image generation process, T2I encodes text by language models like CLIP [19, 24] and integrates text embeddings into U-Net through cross-attention layers. Specifically, these layers employ Query-Key-Value (QKV) structure [33] to represent the interactions between text and vision. For a given text embedding \\( c_{i} \\), keys and values are generated as \\( k_{i} = W_{k}c_{i} \\) and \\( v_{i} = W_{v}c_{i} \\). These keys compute an attention map by multiplying with the query \\( q_{i} \\) representing visual features, and then the cross-attention output is computed by attending over values \\( v_{i} \\):" + }, + { + "type": "equation", + "bbox": [ + 0.425, + 0.62, + 0.786, + 0.639 + ], + "angle": 0, + "content": "\\[\n\\mathcal {O} \\propto \\operatorname {s o f t m a x} (q _ {i} k _ {i} ^ {T}) v _ {i}. \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.659, + 0.789, + 0.844 + ], + "angle": 0, + "content": "Concept Erasing with Closed-form Solution There are existing erasure methods requiring fine-tuning, such as ESD, CA and SA [6,8,13]. However, such methods are relatively inefficient as they require thousands of fine-tuning steps. In contrast, UCE [7] is an efficient method which modifies the attention weights through a closed-form edit. UCE requires a \"source\" concept (e.g., \"nudity\") and a \"destination\" concept (e.g., empty text \"\"). Let \\( c_{i} \\) represent the source embedding, \\( c_{i}^{*} \\) denote the corresponding destination embedding, set \\( E \\) denote concepts to erase, and set \\( P \\) denote concepts to preserve. Given a K/V projection matrix \\( W^{\\mathrm{old}} \\) (a concise notation for \\( W_{k}^{\\mathrm{old}} \\) and \\( W_{v}^{\\mathrm{old}} \\)), UCE seeks new weights \\( W \\) by editing concepts in \\( E \\) while preserving concepts in \\( P \\). Specifically, the objective is to find weights such that the output \\( Wc_{i} \\) for \\( c_{i} \\in E \\) aligns with target values \\( W^{\\mathrm{old}}c_{i}^{*} \\) instead of the original \\( W^{\\mathrm{old}}c_{i} \\). Meanwhile, to control parameter" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.24, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "Reliable and Efficient Concept Erasure of Text-to-Image Diffusion Models" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "image", + "bbox": [ + 0.225, + 0.143, + 0.495, + 0.281 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.325, + 0.283, + 0.388, + 0.295 + ], + "angle": 0, + "content": "Model Editing" + }, + { + "type": "image", + "bbox": [ + 0.496, + 0.143, + 0.782, + 0.281 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.581, + 0.283, + 0.678, + 0.294 + ], + "angle": 0, + "content": "Embedding Derivation" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.309, + 0.788, + 0.395 + ], + "angle": 0, + "content": "Fig. 1: Overview of the proposed RECE. RECE consists of two main components: model editing and embedding derivation. First, erasing concepts by editing the model with a closed-form solution, and obtaining the edited cross-attention \\( W^{\\mathrm{new}} \\). Then, new embedding \\( c_{i}^{\\prime} \\) can be derived by Eq. (7) given the original cross attention \\( W^{\\mathrm{old}} \\) and the edited \\( W^{\\mathrm{new}} \\). In subsequent epochs, model editing and embedding derivation are looped." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.425, + 0.785, + 0.455 + ], + "angle": 0, + "content": "changes, outputs for \\( c_{j} \\in P \\) are preserved as \\( W^{\\mathrm{old}}c_{j} \\) and a L2 regularization term is introduced:" + }, + { + "type": "equation", + "bbox": [ + 0.223, + 0.466, + 0.785, + 0.501 + ], + "angle": 0, + "content": "\\[\n\\min _ {W} \\sum_ {c _ {i} \\in E} | | W c _ {i} - W ^ {\\text {o l d}} c _ {i} ^ {*} | | _ {2} ^ {2} + \\lambda_ {1} \\sum_ {c _ {j} \\in P} | | W c _ {j} - W ^ {\\text {o l d}} c _ {j} | | _ {2} ^ {2} + \\lambda_ {2} | | W - W ^ {\\text {o l d}} | | _ {F} ^ {2}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.513, + 0.785, + 0.545 + ], + "angle": 0, + "content": "where \\(\\lambda_{1}\\) and \\(\\lambda_{2}\\) are scaling factors preserving the existing concepts. UCE [7] prove that this formula has a closed-form solution:" + }, + { + "type": "equation", + "bbox": [ + 0.223, + 0.556, + 0.785, + 0.591 + ], + "angle": 0, + "content": "\\[\nW = W ^ {\\text {o l d}} \\left(\\sum_ {c _ {i} \\in E} c _ {i} ^ {*} c _ {i} ^ {T} + \\lambda_ {1} \\sum_ {c _ {j} \\in P} c _ {j} c _ {j} ^ {T} + \\lambda_ {2} I\\right) \\left(\\sum_ {c _ {i} \\in E} c _ {i} c _ {i} ^ {T} + \\lambda_ {1} \\sum_ {c _ {j} \\in P} c _ {j} c _ {j} ^ {T} + \\lambda_ {2} I\\right) ^ {- 1}. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.604, + 0.785, + 0.649 + ], + "angle": 0, + "content": "UCE directly assigns cross-attention KV matrices using the closed-form solution, eliminating the need for fine-tuning. This makes UCE significantly faster, hence we use UCE in our method." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.676, + 0.657, + 0.694 + ], + "angle": 0, + "content": "3.2 Reliable and Efficient Concept Erasure (RECE)" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.706, + 0.788, + 0.812 + ], + "angle": 0, + "content": "While UCE [7] offers a fast solution for removing undesired concepts from T2I diffusion models, it can still produce undesired content, as illustrated in Tab. 1. This suggests an incomplete erasure of these concepts. To effectively eliminate such undesired concepts, we efficiently erase closed-form embeddings capable of regenerating erased concepts within the unlearned model. The derivation of embeddings and erasure is conducted iteratively to achieve a thorough erasure of inappropriate concepts, as shown in Fig. 1." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.407, + 0.129 + ], + "angle": 0, + "content": "F. Chao Gong et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.488, + 0.373 + ], + "angle": 0, + "content": "Finding Target Contents Let us take \"nudity\" for example. As depicted in the second column of Fig. 2, when directly providing the input prompt \"nudity\" to UCE models, only landscape or unrelated images are generated. This is because the word \"nudity\" has been aligned with the empty text\". However, the erasure of UCE is incomplete. We can generate an adversarial prompt that enables UCE's model to produce images containing nudity content again, similar to those generated by SD when provided with the prompt \"nudity\"." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.374, + 0.489, + 0.435 + ], + "angle": 0, + "content": "In this section, we introduce our method for deriving the new embedding in UCE's model, which guides UCE to generate nude images. As" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.435, + 0.788, + 0.556 + ], + "angle": 0, + "content": "elaborated in Sec. 3.1, T2I introduces text embeddings into image generation through cross-attention layers, where the projection matrices \\( W_{k} \\) and \\( W_{v} \\) are used to transform text embeddings. Let \\( W^{\\mathrm{old}} \\) denote the projection matrices of the original U-Net before UCE's editing, \\( W^{\\mathrm{new}} \\) represent the projection matrices after UCE's editing, \\( c \\) denote the embedding of \"nudity\", and \\( c' \\) signify our derived embedding. If we can find a \\( c' \\) such that \\( W^{\\mathrm{new}}c' \\) closely resembles \\( W^{\\mathrm{old}}c \\), then \\( c' \\) can guide the edited model to generate nude images like how \\( c \\) guides the original model. More precisely, the objective function is formulated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.398, + 0.567, + 0.787, + 0.598 + ], + "angle": 0, + "content": "\\[\n\\min _ {c ^ {\\prime}} \\sum_ {i} \\| W _ {i} ^ {\\mathrm {n e w}} c ^ {\\prime} - W _ {i} ^ {\\mathrm {o l d}} c \\| _ {2} ^ {2}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.61, + 0.788, + 0.687 + ], + "angle": 0, + "content": "where \\( W_{i} \\) denotes K/V cross-attention matrices of U-Net. The solution \\( c' \\) derived from Eq. (4) can be viewed as the actual representation of \\( c \\) within the edited model. Evidently, Eq. (4) represents a convex function with respect to \\( c' \\), which possesses a unique global minimum. As derived in Appendix A, Eq. (4) admits a closed-form solution:" + }, + { + "type": "equation", + "bbox": [ + 0.323, + 0.699, + 0.787, + 0.744 + ], + "angle": 0, + "content": "\\[\nc ^ {\\prime} = \\left(\\sum_ {i} W _ {i} ^ {\\mathrm {n e w} ^ {T}} W _ {i} ^ {\\mathrm {n e w}}\\right) ^ {- 1} \\left(\\sum_ {i} W _ {i} ^ {\\mathrm {n e w} ^ {T}} W _ {i} ^ {\\mathrm {o l d}}\\right) c. \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Given that text conditioning works in the form of embedding in Stable Diffusion (SD), we can use our derived embedding as text conditioning. As illustrated in the third column of Fig. 2, this derived embedding effectively guides UCE's edited model to once again generate nude images, indicating its capacity to represent the concept of \"nudity\" within edited model. Thus it demonstrates that the erasure process of UCE remains incomplete. To address this issue, we further" + }, + { + "type": "image", + "bbox": [ + 0.499, + 0.153, + 0.788, + 0.294 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.497, + 0.306, + 0.789, + 0.404 + ], + "angle": 0, + "content": "Fig.2: When given the input prompt \"nudity\", SD generates images containing nudity content, while UCE generates unrelated images. When given our derived embedding and regularized embedding, UCE generates nude images again. We use for publication purposes." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.24, + 0.115, + 0.733, + 0.131 + ], + "angle": 0, + "content": "Reliable and Efficient Concept Erasure of Text-to-Image Diffusion Models" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.147, + 0.784, + 0.178 + ], + "angle": 0, + "content": "remove our derived embeddings \\( c' \\) from UCE's model with Eq. (3) to prevent the generation of nude images." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.203, + 0.784, + 0.293 + ], + "angle": 0, + "content": "Regularization Term We can further erase the concept of \"nudity\" by substituting \\( c \\) in Eq. (2) with our derived embedding \\( c' \\). However, upon directly erasing \\( c' \\), we observe a significant decline in the model's performance: it struggles to generate high-quality images for unrelated concepts as shown in Fig. 6. Hence, it becomes imperative to devise a method that retains the model's performance while erasing concepts." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.294, + 0.784, + 0.355 + ], + "angle": 0, + "content": "Let \\( W^{\\mathrm{new1}} \\) denote the projection matrices after the last epoch's modification, and \\( W^{\\mathrm{new2}} \\) denote the projection matrices after the current epoch. Partly, preserving the model's performance entails minimizing the impact on unrelated concepts. Consequently, we define our objective function as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.407, + 0.368, + 0.786, + 0.39 + ], + "angle": 0, + "content": "\\[\n\\min _ {W} \\| W ^ {\\mathrm {n e w} 2} d - W ^ {\\mathrm {n e w} 1} d \\| _ {2} ^ {2}, \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.402, + 0.784, + 0.449 + ], + "angle": 0, + "content": "where \\( d \\) represents an unrelated concept's embedding. Note that \\( W^{\\mathrm{new2}} \\) is directly influenced by \\( c' \\), as it is derived after erasing \\( c' \\) from \\( W^{\\mathrm{new1}} \\). We can obtain a theorem about our objective and its proof is provided in Appendix B:" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.462, + 0.734, + 0.479 + ], + "angle": 0, + "content": "Theorem 1. If \\( c' \\) is set to 0, Eq. (6) achieves its global minimum of 0." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.492, + 0.784, + 0.582 + ], + "angle": 0, + "content": "Intuitively, if \\( c_{i} \\) in Eq. (3) is set to zero, the coefficient matrices on the right side will become an identity matrix when multiplied. As a result, \\( W \\) will revert to being equivalent to \\( W^{\\mathrm{old}} \\), exerting minimal influence on unrelated concepts due to the absence of further modifications to \\( W \\). Therefore, we need to introduce a regularization term to the original objective Eq. (4) to ensure that the obtained \\( c' \\) is close to zero, thereby minimizing its influence on the model's performance:" + }, + { + "type": "equation", + "bbox": [ + 0.369, + 0.596, + 0.785, + 0.625 + ], + "angle": 0, + "content": "\\[\n\\min _ {c ^ {\\prime}} \\sum_ {i} \\| W _ {i} ^ {\\text {n e w}} c ^ {\\prime} - W _ {i} ^ {\\text {o l d}} c \\| _ {2} ^ {2} + \\lambda \\| c ^ {\\prime} \\| _ {2} ^ {2}. \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.637, + 0.784, + 0.668 + ], + "angle": 0, + "content": "As derived in Appendix A, this final objective function also possesses a unique global minimum solution:" + }, + { + "type": "equation", + "bbox": [ + 0.31, + 0.68, + 0.785, + 0.723 + ], + "angle": 0, + "content": "\\[\nc ^ {\\prime} = \\left(\\lambda I + \\sum_ {i} W _ {i} ^ {\\mathrm {n e w} ^ {T}} W _ {i} ^ {\\mathrm {n e w}}\\right) ^ {- 1} \\left(\\sum_ {i} W _ {i} ^ {\\mathrm {n e w} ^ {T}} W _ {i} ^ {\\mathrm {o l d}}\\right) c \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.784, + 0.84 + ], + "angle": 0, + "content": "As illustrated in the fourth column of Fig. 2, this regularized embedding can also guide UCE's model to once again generate nude images, indicating its ability to represent the concept \"nudity\" within UCE's model. With the incorporation of our regularization term, we iteratively apply the erasure process to the refined embedding \\( c' \\) using Eq. (3) over multiple epochs. This ensures thorough concept erasure while safeguarding the overall performance of the model. The algorithm details are elaborated in Algorithm 1." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.407, + 0.129 + ], + "angle": 0, + "content": "F. Chao Gong et al." + }, + { + "type": "code_caption", + "bbox": [ + 0.229, + 0.15, + 0.544, + 0.165 + ], + "angle": 0, + "content": "Algorithm 1: Erase Concepts with RECE" + }, + { + "type": "algorithm", + "bbox": [ + 0.219, + 0.167, + 0.788, + 0.445 + ], + "angle": 0, + "content": "Input: Diffusion U-Net \\(\\theta\\) , concepts set \\(E\\) to erase, \\(P\\) to preserve, epochs \\(T\\) Output: Diffusion U-Net \\(\\theta^{\\prime}\\) with concepts \\(E\\) erased. \n1 /\\* Initialize\\*/ \n2 \\(\\theta^{\\prime}\\gets \\theta\\) . \n3 Initialize text embeddings \\(c_{i}\\) and \\(c_{j}\\) from \\(E\\) and \\(P\\) . \n4 Extract K&V matrices \\(W^{\\mathrm{old}}\\) from the cross attention of \\(\\theta\\) . \n5 /\\* Preliminary erasing with UCE\\*/ \n6 Obtain \\(W^{\\mathrm{new}}\\) by erasing concepts in \\(E\\) with Eq. (3); \n7 for \\(t = 1,\\ldots ,T\\) do \n8 /\\* Derive new embeddings\\*/ \n9 \\(E^{\\prime}\\gets \\{\\}\\) . \nfor \\(c_{i}\\in E\\) do \n11 Derive new embedding \\(c_i^\\prime\\) using \\(W^{\\mathrm{new}}\\) with Eq. (8); \n12 \\(\\begin{array}{r}E^{\\prime}\\leftarrow E^{\\prime}\\cup \\{c_{i}^{\\prime}\\} \\end{array}\\) \n13 /\\* Erasing derived embeddings\\*/ \n14 Obtain \\(W^{\\mathrm{new}}\\) by erasing \\(E^{\\prime}\\) with Eq. (3); \n15 \\(W^{\\mathrm{new}}\\gets W^{\\mathrm{new}};\\) \n16 Update \\(\\theta^{\\prime}\\) by replacing K&V matrices with \\(W^{\\mathrm{new}}\\) . \n17 return \\(\\theta^{\\prime}\\)" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.479, + 0.376, + 0.496 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.511, + 0.789, + 0.692 + ], + "angle": 0, + "content": "In this section, we present the results of our method for erasing inappropriate concepts and artistic styles. We also include the results of object removal in Appendix. We start with SD V1.4 as our base model. Following the implementation in [16], we set \\(\\lambda_{1}\\) and \\(\\lambda_{2}\\) in Eq. (2) to 0.1. For inappropriate concepts, we perform iterative erasure for 5 epochs and set \\(\\lambda\\) in Eq. (7) to \\(1e - 1\\). For artistic style, we conduct erasure for 10 epochs and set \\(\\lambda\\) to \\(1e - 3\\). The baselines we will compare with are: SD V1.4 [25], SD V2.1 [31](Stable Diffusion pretrained on an NSFW filtered dataset), SLD [29], ESD [6], CA [13], SA [8], UCE [7]. As for SLD, ESD, SA and UCE, we adhere to the recommended configuration in their papers [6-8,29]. For CA [13], we fine-tune the full weights of U-Net to erase unsafe contents and the cross-attention module to erase artistic styles, according to its documentation3." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.715, + 0.469, + 0.729 + ], + "angle": 0, + "content": "4.1 Unsafe Content Removal" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.741, + 0.788, + 0.817 + ], + "angle": 0, + "content": "Experimental Setup In this section, we assess the effectiveness of erasing unsafe concepts. We conduct experiments on the Inappropriate Image Prompts (I2P) dataset [29]. The I2P dataset includes various inappropriate prompts, such as violence, self-harm, sexual content, and shocking content. These prompts are collected from real-world, user-generated images based on the official SD. Our" + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.825, + 0.585, + 0.841 + ], + "angle": 0, + "content": "3 https://github.com/nupurkmr9/concept-ablation" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.241, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Reliable and Efficient Concept Erasure of Text-to-Image Diffusion Models" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "table", + "bbox": [ + 0.219, + 0.144, + 0.783, + 0.269 + ], + "angle": 0, + "content": "
MethodNudity DetectionCOCO-30k
Breast(F)Genitalia(F)Breast(M)Genitalia(M)ButtocksFeetBellyArmpitsTotal↓CLIP↑FID↓
SD v1.4183214610444217112964631.33-
SD v2.1121134031439146109485--
ESD-u14185524313312130.453.73
UCE3161981120553618631.261.82
SLD-Med72534318191049935430.952.60
SA399401532491516330.5717.34
CA6191041428239531.167.87
Ours80640823176630.952.82
" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.27, + 0.785, + 0.325 + ], + "angle": 0, + "content": "Table 1: Comparison of performance metrics for content removal methods. (Left) Number of nude body parts detected by Nudenet on I2P dataset with threshold 0.6. (Right) CLIP-score and FID against original SD. F: Female. M: Male. Bold: best. Underline: second-best." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.37, + 0.783, + 0.445 + ], + "angle": 0, + "content": "evaluation focuses on the erasure of nudity since it is a classical unsafe concept. For each model, we generate one image per prompt in the I2P dataset, resulting in a total of 4703 images. Nude body parts are detected using the Nudenet detector [18], with the threshold set to 0.6. This threshold follows the default settings in \\(\\mathrm{I2P^4}\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.447, + 0.785, + 0.524 + ], + "angle": 0, + "content": "To verify that the unlearned models can still generate normal images, we use COCO-30k [14] with its captions as prompts. COCO-30k is a dataset devoid of unsafe concepts, making it suitable for evaluating edited models' generation capabilities. We evaluate the models' image-text consistency based on CLIP-score [9], and visual similarity against SD-generated images based on FID [17]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.552, + 0.785, + 0.656 + ], + "angle": 0, + "content": "Removal Results As depicted in Tab. 1, our method yields the lowest number of nude body parts, while demonstrating impressive specificity in preserving normal content of COCO-30k. CA generates the second-fewest nude body parts but it exhibits poorer performance in terms of FID, indicating a poorer trade-off between generation ability and removal effectiveness. On the other hand, CA fine-tunes full weights and our RECE only fine-tunes cross-attention modules, which will be discussed in detail in Sec. 4.4." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.659, + 0.785, + 0.764 + ], + "angle": 0, + "content": "Notably, our method achieves a FID score closely comparable to the top-performing UCE and the second-best SLD, both of which generates a considerably higher number of nude body parts. This suggests that our method minimally impacts the generation of normal content while striving for better removal effectiveness. Additionally, most methods exhibits favorable CLIP-score results thus we consider the CLIP-score performance acceptable as long as it remains within a reasonable range." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.766, + 0.793, + 0.813 + ], + "angle": 0, + "content": "In open-sourced conditions, the inference guidance mechanism such as SLD [29] can be easily bypassed by deleting the corresponding code [24]. Large-scale model retraining on NSFW-filtered dataset demands considerable computational re" + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.825, + 0.5, + 0.841 + ], + "angle": 0, + "content": "4 https://github.com/ml-research/i2p" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.407, + 0.129 + ], + "angle": 0, + "content": "F. Chao Gong et al." + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.144, + 0.29, + 0.246 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.233, + 0.248, + 0.281, + 0.258 + ], + "angle": 0, + "content": "SD v1.4" + }, + { + "type": "image", + "bbox": [ + 0.293, + 0.144, + 0.361, + 0.246 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.304, + 0.248, + 0.351, + 0.258 + ], + "angle": 0, + "content": "SD v2.1" + }, + { + "type": "image", + "bbox": [ + 0.364, + 0.144, + 0.431, + 0.246 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.383, + 0.248, + 0.411, + 0.259 + ], + "angle": 0, + "content": "ESD" + }, + { + "type": "image", + "bbox": [ + 0.434, + 0.144, + 0.5, + 0.246 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.453, + 0.248, + 0.482, + 0.259 + ], + "angle": 0, + "content": "UCE" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.144, + 0.571, + 0.246 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.509, + 0.248, + 0.565, + 0.259 + ], + "angle": 0, + "content": "SLD-Med" + }, + { + "type": "image", + "bbox": [ + 0.574, + 0.144, + 0.641, + 0.246 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.596, + 0.248, + 0.616, + 0.259 + ], + "angle": 0, + "content": "SA" + }, + { + "type": "image", + "bbox": [ + 0.643, + 0.144, + 0.711, + 0.246 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.667, + 0.248, + 0.688, + 0.258 + ], + "angle": 0, + "content": "CA" + }, + { + "type": "image", + "bbox": [ + 0.714, + 0.144, + 0.781, + 0.246 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.732, + 0.248, + 0.762, + 0.258 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.271, + 0.787, + 0.327 + ], + "angle": 0, + "content": "Fig. 3: Qualitative results of different nudity removing methods. Images for each row are generated using the same prompt. All prompts originate from I2P benchmark, and we don't intentionally generate inappropriate content of any individual." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.358, + 0.784, + 0.387 + ], + "angle": 0, + "content": "sources [22], and even then, the model SD v2.1 may still generate nude images, as illustrated in Tab. 1." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.388, + 0.785, + 0.48 + ], + "angle": 0, + "content": "Qualitative results are illustrated in Fig. 3. In the first row, all erasing methods successfully generate non-nude images, but the results of CA and SA differ significantly from the original SD's. In the second row, all methods except for SD v2.1 avoid generating nude images. Among these, only UCE, SLD, and our method effectively capture the facial features. These findings demonstrate that our method effectively maintains unrelated concepts." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.503, + 0.788, + 0.73 + ], + "angle": 0, + "content": "Nudity Bias While our RECE demonstrates remarkable effectiveness in minimizing the generation of nude content, it exhibits limitations in erasing male nudity and similar limitations are also observed in other methods. We count the number of female and male nude body parts in the 4703 I2P images. As illustrated in Tab. 2, the nudity ratios between women and men almost decrease in every method, indicating an erasure bias on female information (excluding SA, which erases more sex-related concepts besides nudity [8]). We attribute this limitation to the inherent bias in the target concept \"nudity\" within SD, which tends to generate more female-oriented content. To investigate this bias, we randomly selected 20 seeds and employed SD V1.4 to generate 3 images per seed with the prompt \"nudity\", resulting in a total of 60 images. Surprisingly, almost all of these generated images depict female body part as presented in the last column of Tab. 2. Thus our derived embedding is also biased. Further improvements require an awareness of the biases inherent in the model while performing erasure." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.753, + 0.452, + 0.768 + ], + "angle": 0, + "content": "4.2 Artistic Style Removal" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.787, + 0.842 + ], + "angle": 0, + "content": "Experimental Setup We conduct an evaluation to assess the efficacy of removing artistic styles to address copyright concerns. Following the datasets in ESD [6], we use 20 prompts for each of 5 famous artists—Van Gogh, Pablo Picasso, Rembrandt, Andy Warhol and Caravaggio—and 5 modern artists—Kelly" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.24, + 0.114, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Reliable and Efficient Concept Erasure of Text-to-Image Diffusion Models" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "table", + "bbox": [ + 0.264, + 0.144, + 0.74, + 0.215 + ], + "angle": 0, + "content": "
MetricsSDESDUCESLD-MedSACAOursSD- "nudity"
# Female Nudity204153777487889
# Male Nudity56132737419101
Female/Male3.641.151.372.08120.370.889
" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.215, + 0.787, + 0.27 + ], + "angle": 0, + "content": "Table 2: In the first 7 columns, we counted the numbers of female nudity and male nudity body parts for I2P dataset. The last column is the result of SD v1.4 conditioned by the prompt \"nudity\". This table highlights the inherent bias within SD. And all each erasure methods can not remove the concept of male nudity very well." + }, + { + "type": "table", + "bbox": [ + 0.264, + 0.295, + 0.742, + 0.403 + ], + "angle": 0, + "content": "
Removal MethodErase "Van Gogh"Erase "Kelly McKernan"
LPIPSe↑LPIPSu↓LPIPSd↑LPIPSe↑LPIPSu↓LPIPSd↑
ESD-x0.400.260.140.370.210.16
UCE0.250.050.200.250.030.22
SLD-Med0.210.100.110.220.180.04
CA0.300.130.170.220.170.05
Ours0.310.080.230.290.040.25
" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.404, + 0.785, + 0.431 + ], + "angle": 0, + "content": "Table 3: Comparison of LPIPS scores for artistic removal methods. Bold: best. Underline: second-best. \\(\\mathrm{LPIPS_d}\\) indicates overall erasure performance." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.472, + 0.785, + 0.534 + ], + "angle": 0, + "content": "McKernan, Thomas Kinkade, Tyler Edlin, Kilian Eng and the series \"Ajin: Demi-Human\", which have been reported to be imitated by SD [30]. To evaluate our RECE and all the aforementioned baselines, we erase the style of two artists: Van Gogh and Kelly McKernan." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.554, + 0.788, + 0.736 + ], + "angle": 0, + "content": "Removal Results We conducted an evaluation based on LPIPS scores [37] compared to the original SD, as detailed in Tab. 3. LPIPS evaluates the perceptual distance between image patches, where higher values indicate greater differences and lower values indicate more similarity. The \\(\\mathrm{LPIPS_e}\\) is calculated on the erased artist. A higher \\(\\mathrm{LPIPS_e}\\) value suggests a more effective style removal, and both ESD and our method demonstrate successful erasure of the target style. \\(\\mathrm{LPIPS_u}\\) is calculated on unerased artists. A lower \\(\\mathrm{LPIPS_u}\\) indicates a lesser impact on unrelated artists, where our method and UCE effectively maintains unrelated concepts. We also calculate the overall effectiveness by \\(\\mathrm{LPIPS_d} = \\mathrm{LPIPS_e} - \\mathrm{LPIPS_u}\\), which is the difference between erased and unerased artists. Our method performs best in this regard. Qualitative results can be found in Appendix." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.757, + 0.586, + 0.772 + ], + "angle": 0, + "content": "4.3 Robustness Against Red-teaming Tools" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.787, + 0.842 + ], + "angle": 0, + "content": "Experimental Setup To demonstrate the robustness of our RECE in safeguarding against various attack methods, we employ different red-teaming tools, including white-box methods such as P4D [3] and UnlearnDiff [38], and the black-box method Ring-A-Bell [32]. For nudity, as provided by UnlearnDiff [38]," + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.407, + 0.129 + ], + "angle": 0, + "content": "F. Chao Gong et al." + }, + { + "type": "table", + "bbox": [ + 0.264, + 0.145, + 0.74, + 0.232 + ], + "angle": 0, + "content": "
Red-teamingSD v1.4SD v2.1ESD-uUCESLD-MaxSACAOurs
UnlearnDiff--66.2079.5882.3977.4665.4965.49
P4D--63.3880.2877.4678.8760.5664.79
Ring-A-Bell83.1072.5469.7233.1066.2022.5425.3513.38
Average--66.4364.3275.3559.6250.4747.89
" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.233, + 0.784, + 0.26 + ], + "angle": 0, + "content": "Table 4: Robustness of different methods against red-teaming tools, measured by attack success rate(%) . Bold: best. Underline: second-best." + }, + { + "type": "image", + "bbox": [ + 0.284, + 0.285, + 0.357, + 0.34 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.285, + 0.34, + 0.355, + 0.351 + ], + "angle": 0, + "content": "Original SD" + }, + { + "type": "image", + "bbox": [ + 0.358, + 0.286, + 0.428, + 0.34 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.379, + 0.341, + 0.406, + 0.351 + ], + "angle": 0, + "content": "ESD" + }, + { + "type": "image", + "bbox": [ + 0.43, + 0.286, + 0.501, + 0.339 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.451, + 0.341, + 0.481, + 0.351 + ], + "angle": 0, + "content": "UCE" + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.286, + 0.574, + 0.339 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.51, + 0.341, + 0.567, + 0.351 + ], + "angle": 0, + "content": "SLD-Med" + }, + { + "type": "image", + "bbox": [ + 0.576, + 0.286, + 0.646, + 0.339 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.601, + 0.341, + 0.621, + 0.35 + ], + "angle": 0, + "content": "CA" + }, + { + "type": "image", + "bbox": [ + 0.649, + 0.286, + 0.72, + 0.339 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.67, + 0.341, + 0.699, + 0.35 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.364, + 0.785, + 0.405 + ], + "angle": 0, + "content": "Fig. 4: Qualitative results of different artistic style removal methods for eliminating the style of Van Gogh under Ring-A-Bell's attack. The prompt is \"A depiction of a starry night over a quiet town, reminiscent of Van Gogh's famous painting\"." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.432, + 0.784, + 0.478 + ], + "angle": 0, + "content": "we use a set of 143 prompts selected from I2P, each with a nudity score (as determined by NudeNet) above 0.75, and employ the Nudenet detector with a threshold set to 0.45 for detecting inappropriate content." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.497, + 0.785, + 0.663 + ], + "angle": 0, + "content": "Results The attack success rates (ASR, %) are summarized in Tab. 4. Our method achieves the best robustness in average. In the case of the black-box attack Ring-A-Bell, our method achieves the lowest ASR at \\(13.38\\%\\), significantly outperforming other methods. In the case of the white-box attack, CA achieves the best performance, while our method performs either the best or very closely to the second-best. Although SA achieves a decent black-box result, it consumes substantial computation resources for generating 5000 prepared images [8]. CA modifies \\(100\\%\\) of the U-Net parameters while our method only modifies K&V projection matrices, constituting a mere \\(2.23\\%\\) of the U-Net. While UCE also modifies only \\(2.23\\%\\) parameters, our method significantly outperforms UCE. This is attributed to our derived embeddings in Eq. (8)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.664, + 0.785, + 0.769 + ], + "angle": 0, + "content": "For artistic style removal, we provide qualitative examples in Fig. 4. The first image is generated by the original SD v1.4 without any attack, and the following images are under Ring-A-Bell's attack [32]. Recall that UCE is the second-best artistic style removal method as shown in Tab. 3 but it falls short in robustness against red-teaming tools. Although our method employs a closed-form solution like UCE, it outperforms UCE in robustness. ESD, CA and our method perform similarly well. More results can be found in the appendix." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.788, + 0.462, + 0.803 + ], + "angle": 0, + "content": "4.4 Model Editing Duration" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.81, + 0.784, + 0.84 + ], + "angle": 0, + "content": "To demonstrate the efficiency of different methods, we measured the percentage of parameter modification and editing duration on an RTX 3090 for each" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.24, + 0.114, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Reliable and Efficient Concept Erasure of Text-to-Image Diffusion Models" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "table", + "bbox": [ + 0.349, + 0.144, + 0.656, + 0.2 + ], + "angle": 0, + "content": "
ESDUCESACAOurs
Modification (%)94.652.2394.651002.23
Duration (s)37201.2-14003.4
" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.2, + 0.784, + 0.228 + ], + "angle": 0, + "content": "Table 5: Percentage of parameter modification and model editing duration for different methods. Our method and UCE are significantly ahead of other methods." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.271, + 0.785, + 0.346 + ], + "angle": 0, + "content": "method, as shown in Tab. 5. We excluded SLD from the analysis since it operates at inference time rather than modifying the model's weights which can be easily bypassed under open-source conditions. Additionally, we don't include the duration of SA, as it involves generating 5000 images, calculating the Fisher Information Matrix and fine-tuning, which makes it exceptionally slow." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.347, + 0.787, + 0.453 + ], + "angle": 0, + "content": "Based on Tab. 5 and Tab. 1, our method achieves the best concept erasure effect in an extremely short time of only three seconds. Our method (5 epochs) and UCE (1 epoch) modify the lowest percentage of parameters with a closed-form solution, resulting in the shortest editing durations. Despite similar durations, our method significantly outperforms UCE in removal effectiveness. Conversely, CA, ESD and SA modify a high percentage of parameters with more time but achieve less impressive removal results." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.475, + 0.388, + 0.49 + ], + "angle": 0, + "content": "4.5 Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.501, + 0.785, + 0.531 + ], + "angle": 0, + "content": "We conduct experiments to elucidate the impact of our derived embedding among different epochs and the effectiveness of the regularization term." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.553, + 0.414, + 0.584 + ], + "angle": 0, + "content": "Effect of Derived Embeddings among Epochs" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.584, + 0.404, + 0.733 + ], + "angle": 0, + "content": "We conduct an experiment to expound the impact of our derived embedding. We perform \"nudity\" erasure for 5 epochs using Algorithm 1. In each epoch, we derive a distinct embedding that represents \"nudity\". Before the erasure of each" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.787, + 0.84 + ], + "angle": 0, + "content": "epoch, we generate an image using the embedding to test its degree of nudity information, as presented in Fig. 5. Images from epoch 0 to epoch 2 contain nude body parts, indicating that our derived embeddings successfully reveal potential nudity information in the model. Specifically, we opt for the checkpoint after epoch 2, as images from epoch 3 to 4 lack nudity information. Actually, erasing such \"not so nude\" embeddings in epoch 3-4 would impair the model's normal generation ability, which is an unworthy trade-off." + }, + { + "type": "image", + "bbox": [ + 0.425, + 0.585, + 0.779, + 0.655 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.413, + 0.665, + 0.788, + 0.708 + ], + "angle": 0, + "content": "Fig. 5: There is no nudity information from epoch 3 onward; hence, we select the checkpoint after epoch 2 to avoid damaging the model's ability." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.407, + 0.129 + ], + "angle": 0, + "content": "F. Chao Gong et al." + }, + { + "type": "image", + "bbox": [ + 0.263, + 0.146, + 0.744, + 0.236 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.249, + 0.785, + 0.264 + ], + "angle": 0, + "content": "Fig. 6: Ablation study on the impact of the regularization term on model performance." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.302, + 0.789, + 0.499 + ], + "angle": 0, + "content": "Effect of Regularization Coefficient We conduct an experiment to assess the influence of our regularization term. Specifically, we select different regularization coefficients \\(\\lambda\\) in Eq. (4), which divide the interval [0, 1] into five parts. The results, presented in Fig. 6, include CLIP-score and LPIPS against the original SD on the COCO-30k validation subset. As \\(\\lambda\\) increases, CLIP score shows an upward trend while LPIPS and the difference between new and old parameters show a downward trend. This indicates the role of the regularization term in preserving the model's ability for unerased content. Furthermore, we recorded the number of nude parts on the I2P benchmark, presented in the third column of Fig. 6. However, the number of nude parts doesn't strictly increase as \\(\\lambda\\) increases, which is counterintuitive. Although the purpose of the regularization term in Eq. (6) is to preserve the model's generation capability, maintaining this capability does not always affect the erasure effect." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.545, + 0.36, + 0.56 + ], + "angle": 0, + "content": "5 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.599, + 0.789, + 0.844 + ], + "angle": 0, + "content": "In this paper, we propose a novel approach for reliably and efficiently erasing specific concepts from Text-to-Image (T2I) diffusion models. Our approach only modifies the cross-attention K&V matrices of U-Net, constituting a mere \\(2.23\\%\\) of parameters. While previous methods also edited cross-attention modules, they still exhibited the ability to generate inappropriate images. To tackle this challenge, we derive and erase new embeddings that can represent target concepts within unlearned models. To mitigate the impact on unrelated concepts, a regularization term is introduced during the erasure process. All the above techniques are formulated in closed-form, facilitating rapid editing. This enables the execution of \"derive-erase\" across multiple epochs, ensuring thorough and robust erasure. Extensive experiments were conducted to validate the effectiveness of our approach in erasing artistic styles, unsafe contents and object classes. Furthermore, we recorded editing durations to underscore the efficiency of our method and evaluated the robustness against red-teaming tools. We believe our RECE has the potential to empower T2I providers in effectively removing undesired concepts, thereby fostering the development of a safer AI community." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.24, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Reliable and Efficient Concept Erasure of Text-to-Image Diffusion Models" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.146, + 0.403, + 0.163 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.179, + 0.784, + 0.21 + ], + "angle": 0, + "content": "This project was supported by National Key R&D Program of China (No. 2021ZD0112804)." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.234, + 0.324, + 0.25 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.267, + 0.786, + 0.309 + ], + "angle": 0, + "content": "1. Chen, K., Wei, Z., Chen, J., Wu, Z., Jiang, Y.G.: Attacking video recognition models with bullet-screen comments. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 36, pp. 312-320 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.311, + 0.787, + 0.351 + ], + "angle": 0, + "content": "2. Chen, K., Wei, Z., Chen, J., Wu, Z., Jiang, Y.G.: Gcma: Generative cross-modal transferable adversarial attacks from images to videos. In: Proceedings of the 31st ACM International Conference on Multimedia. pp. 698-708 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.353, + 0.787, + 0.393 + ], + "angle": 0, + "content": "3. Chin, Z.Y., Jiang, C.M., Huang, C.C., Chen, P.Y., Chiu, W.C.: Prompting4debugging: Red-teaming text-to-image diffusion models by finding problematic prompts. arXiv preprint arXiv:2309.06135 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.395, + 0.787, + 0.421 + ], + "angle": 0, + "content": "4. Dhariwal, P., Nichol, A.: Diffusion models beat gans on image synthesis. Advances in neural information processing systems 34, 8780-8794 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.423, + 0.787, + 0.449 + ], + "angle": 0, + "content": "5. Doersch, C.: Tutorial on variational autoencoders. arXiv preprint arXiv:1606.05908 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.451, + 0.787, + 0.477 + ], + "angle": 0, + "content": "6. Gandikota, R., Materzynska, J., Fiotto-Kaufman, J., Bau, D.: Erasing concepts from diffusion models. arXiv preprint arXiv:2303.07345 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.479, + 0.787, + 0.519 + ], + "angle": 0, + "content": "7. Gandikota, R., Orgad, H., Belinkov, Y., Materzyńska, J., Bau, D.: Unified concept editing in diffusion models. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 5111-5120 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.521, + 0.787, + 0.56 + ], + "angle": 0, + "content": "8. Heng, A., Soh, H.: Selective amnesia: A continual learning approach to forgetting in deep generative models. Advances in Neural Information Processing Systems 36 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.563, + 0.787, + 0.589 + ], + "angle": 0, + "content": "9. Hessel, J., Holtzman, A., Forbes, M., Bras, R.L., Choi, Y.: Clipscore: A reference-free evaluation metric for image captioning. arXiv preprint arXiv:2104.08718 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.591, + 0.787, + 0.617 + ], + "angle": 0, + "content": "0. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. Advances in neural information processing systems 33, 6840-6851 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.619, + 0.787, + 0.632 + ], + "angle": 0, + "content": "1. Hunter, T.: Ai porn is easy to make now. for women, that's a nightmare. (2 2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.634, + 0.787, + 0.687 + ], + "angle": 0, + "content": "2. Kawar, B., Zada, S., Lang, O., Tov, O., Chang, H., Dekel, T., Mosseri, I., Irani, M.: Imagic: Text-based real image editing with diffusion models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6007-6017 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.689, + 0.787, + 0.729 + ], + "angle": 0, + "content": "3. Kumari, N., Zhang, B., Wang, S.Y., Shechtman, E., Zhang, R., Zhu, J.Y.: Ablating concepts in text-to-image diffusion models. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 22691-22702 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.731, + 0.787, + 0.785 + ], + "angle": 0, + "content": "4. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollar, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13. pp. 740-755. Springer (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.786, + 0.787, + 0.84 + ], + "angle": 0, + "content": "5. Nichol, A.Q., Dhariwal, P., Ramesh, A., Shyam, P., Mishkin, P., Mcgrew, B., Sutskever, I., Chen, M.: Glide: Towards photorealistic image generation and editing with text-guided diffusion models. In: International Conference on Machine Learning. pp. 16784-16804. PMLR (2022)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.267, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.406, + 0.128 + ], + "angle": 0, + "content": "F. Chao Gong et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.148, + 0.785, + 0.189 + ], + "angle": 0, + "content": "16. Orgad, H., Kawar, B., Belinkov, Y.: Editing implicit assumptions in text-to-image diffusion models. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 7053-7061 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.19, + 0.785, + 0.232 + ], + "angle": 0, + "content": "17. Parmar, G., Zhang, R., Zhu, J.Y.: On aliased resizing and surprising subtleties in gan evaluation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11410-11420 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.232, + 0.785, + 0.259 + ], + "angle": 0, + "content": "18. Praneeth, B.: Nudenet: Neural nets for nudity classification, detection and selective censoring (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.259, + 0.785, + 0.313 + ], + "angle": 0, + "content": "19. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.315, + 0.785, + 0.355 + ], + "angle": 0, + "content": "20. Ramesh, A., Dhariwal, P., Nichol, A., Chu, C., Chen, M.: Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125 1(2), 3 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.356, + 0.785, + 0.383 + ], + "angle": 0, + "content": "21. Rando, J., Paleka, D., Lindner, D., Heim, L., Tramér, F.: Red-teaming the stable diffusion safety filter. arXiv preprint arXiv:2210.04610 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.383, + 0.635, + 0.397 + ], + "angle": 0, + "content": "22. Rombach, R.: Stable diffusion 2.0 release (November 2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.398, + 0.785, + 0.412 + ], + "angle": 0, + "content": "23. Rombach, R.: Tutorial: How to remove the safety filter in 5 seconds (August 2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.412, + 0.785, + 0.453 + ], + "angle": 0, + "content": "24. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10684-10695 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.453, + 0.756, + 0.466 + ], + "angle": 0, + "content": "25. Rombach, R., Esser, P.: Stable diffusion v1-4 model card. Model Card (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.467, + 0.785, + 0.522 + ], + "angle": 0, + "content": "26. Ronneberger, O., Fischer, P., Brox, T.: U-net: Convolutional networks for biomedical image segmentation. In: Medical image computing and computer-assisted intervention-MICCAI 2015: 18th international conference, Munich, Germany, October 5-9, 2015, proceedings, part III 18. pp. 234-241. Springer (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.522, + 0.785, + 0.577 + ], + "angle": 0, + "content": "27. Ruiz, N., Li, Y., Jampani, V., Pritch, Y., Rubinstein, M., Aberman, K.: Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 22500-22510 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.578, + 0.785, + 0.632 + ], + "angle": 0, + "content": "28. Saharia, C., Chan, W., Saxena, S., Li, L., Whang, J., Denton, E.L., Ghasemipour, K., Gontijo Lopes, R., Karagol Ayan, B., Salimans, T., et al.: Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems 35, 36479-36494 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.633, + 0.785, + 0.688 + ], + "angle": 0, + "content": "29. Schramowski, P., Brack, M., Deiseroth, B., Kersting, K.: Safe latent diffusion: Mitigating inappropriate degeneration in diffusion models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 22522-22531 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.689, + 0.772, + 0.702 + ], + "angle": 0, + "content": "30. Setty, R.: Ai art generators hit with copyright suit over artists' images (1 2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.702, + 0.668, + 0.715 + ], + "angle": 0, + "content": "31. StabilityAI: Stable diffusion 2.1 model card. Model Card (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.716, + 0.785, + 0.757 + ], + "angle": 0, + "content": "32. Tsai, Y.L., Hsu, C.Y., Xie, C., Lin, C.H., Chen, J.Y., Li, B., Chen, P.Y., Yu, C.M., Huang, C.Y.: Ring-a-bell! how reliable are concept removal methods for diffusion models? arXiv preprint arXiv:2310.10012 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.758, + 0.785, + 0.799 + ], + "angle": 0, + "content": "33. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. Advances in neural information processing systems 30 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.799, + 0.785, + 0.84 + ], + "angle": 0, + "content": "34. Wei, Z., Chen, J., Goldblum, M., Wu, Z., Goldstein, T., Jiang, Y.G., Davis, L.S.: Towards transferable adversarial attacks on image and video transformers. IEEE Transactions on Image Processing 32, 6346-6358 (2023)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.148, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.24, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Reliable and Efficient Concept Erasure of Text-to-Image Diffusion Models" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.788, + 0.189 + ], + "angle": 0, + "content": "35. Wei, Z., Chen, J., Wu, Z., Jiang, Y.G.: Adaptive cross-modal transferable adversarial attacks from images to videos. IEEE Transactions on Pattern Analysis and Machine Intelligence (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.19, + 0.788, + 0.232 + ], + "angle": 0, + "content": "36. Yang, L., Zhang, Z., Song, Y., Hong, S., Xu, R., Zhao, Y., Zhang, W., Cui, B., Yang, M.H.: Diffusion models: A comprehensive survey of methods and applications. ACM Computing Surveys 56(4), 1-39 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.232, + 0.788, + 0.273 + ], + "angle": 0, + "content": "37. Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual metric. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 586-595 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.273, + 0.788, + 0.315 + ], + "angle": 0, + "content": "38. Zhang, Y., Jia, J., Chen, X., Chen, A., Zhang, Y., Liu, J., Ding, K., Liu, S.: To generate or not? safety-driven unlearned diffusion models are still easy to generate unsafe images... for now. arXiv preprint arXiv:2310.11868 (2023)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.788, + 0.315 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/Reliable and Efficient Concept Erasure of Text-to-Image Diffusion Models/d5988da3-2c65-4cde-838f-fe6bf86e8cd3_origin.pdf b/2024/Reliable and Efficient Concept Erasure of Text-to-Image Diffusion Models/d5988da3-2c65-4cde-838f-fe6bf86e8cd3_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..76aa162a79c7fcdc57d52edfe2a2723d7cacb21e --- /dev/null +++ b/2024/Reliable and Efficient Concept Erasure of Text-to-Image Diffusion Models/d5988da3-2c65-4cde-838f-fe6bf86e8cd3_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:fb126e8856f71d10336c699427e53a32b0b900afc50a1610f1ed7271d7d2fbf7 +size 5226051 diff --git a/2024/Reliable and Efficient Concept Erasure of Text-to-Image Diffusion Models/full.md b/2024/Reliable and Efficient Concept Erasure of Text-to-Image Diffusion Models/full.md new file mode 100644 index 0000000000000000000000000000000000000000..ca186588154215f8dcd8039f4b7eac736e220162 --- /dev/null +++ b/2024/Reliable and Efficient Concept Erasure of Text-to-Image Diffusion Models/full.md @@ -0,0 +1,325 @@ +# Reliable and Efficient Concept Erasure of Text-to-Image Diffusion Models + +Chao Gong $^{1,2\star}$ , Kai Chen $^{1,2\star}$ , Zhipeng Wei $^{1,2}$ , Jingjing Chen $^{1,2\dagger}$ , and Yu-Gang Jiang $^{1,2}$ + +$^{1}$ Shanghai Key Lab of Intell. Info. Processing, School of Computer Science, Fudan University +$^{2}$ Shanghai Collaborative Innovation Center on Intelligent Visual Computing {c Gong20, chenjingjing, ygj} @fudan.edu.cn, {kchen22, zpwei21} @m.fudan.edu.cn + +Abstract. Text-to-image models encounter safety issues, including concerns related to copyright and Not-Safe-For-Work (NSFW) content. Despite several methods have been proposed for erasing inappropriate concepts from diffusion models, they often exhibit incomplete erasure, consume a lot of computing resources, and inadvertently damage generation ability. In this work, we introduce Reliable and Efficient Concept Erasure (RECE), a novel approach that modifies the model in 3 seconds without necessitating additional fine-tuning. Specifically, RECE efficiently leverages a closed-form solution to derive new target embeddings, which are capable of regenerating erased concepts within the unlearned model. To mitigate inappropriate content potentially represented by derived embeddings, RECE further aligns them with harmless concepts in crossattention layers. The derivation and erasure of new representation embeddings are conducted iteratively to achieve a thorough erasure of inappropriate concepts. Besides, to preserve the model's generation ability, RECE introduces an additional regularization term during the derivation process, resulting in minimizing the impact on unrelated concepts during the erasure process. All the processes above are in closed-form, guaranteeing extremely efficient erasure in only 3 seconds. Benchmarking against previous approaches, our method achieves more efficient and thorough erasure with minor damage to original generation ability and demonstrates enhanced robustness against red-teaming tools. Code is available at https://github.com/CharlesGong12/RECE. + +WARNING: This paper contains model outputs that may be offensive. + +Keywords: Text-to-Image $\cdot$ Concept Erasing $\cdot$ Machine Unlearning + +# 1 Introduction + +In recent years, large-scale text-to-image (T2I) diffusion models have exhibited remarkable capability in synthesizing photo-realistic images from text prompts [15, + +[20, 24, 28]. The exceptional performance of T2I diffusion models is largely due to the vast amount of training data collected from the Internet, which enables the models to imitate a wide variety of concepts. Unfortunately, such powerful models can also be misused to generate copyright infringement and Not-Safe-For-Work (NSFW) image content when conditioned on inappropriate text prompts [11, 30]. Especially the open-source release of the Stable Diffusion (SD) T2I model has made advanced image generation technology widely accessible. To alleviate this safety concern, several recent research efforts have incorporated safety mechanisms into T2I diffusion models, e.g. filtering out inappropriate training data and retraining model [22], censoring model outputs with an NSFW safety checker [21], and applying classifier-free guidance to steer the generation away from inappropriate concepts [29]. However, these safety mechanisms either demand expensive computational resources and time [25] or can be easily circumvented by malicious users due to the public availability of code and model parameters in open-source scenario [23]. + +In response to the drawbacks mentioned above, an alternative is to erase inappropriate concepts from the T2I diffusion model [6-8, 13]. Specifically, given an inappropriate concept described in the text prompt, the pre-trained T2I diffusion model's parameters are fine-tuned to unlearn that concept so that the associated image content cannot be generated. Compared with previous security mechanisms, concept erasure neither requires training the entire model from scratch nor can be easily circumvented even in the case of open-source code. Despite promising progress in concept erasure, there exist several issues. On the one hand, most erasure methods require a high number of iterations to finetune considerable amounts of parameters [6, 8, 13], which inevitably degrades the generation capability and consumes a lot of computing resources. Only a recent work called UCE [7] modifies model parameters without fine-tuning using a closed-form solution, ensuring the model maintains original generation capability when erasing concepts. On the other hand, almost all methods fail to sufficiently erase inappropriate concepts, leaving them vulnerable to problematic prompts found by the red-teaming of T2I diffusion models [3, 32, 38]. This results in the unlearned model being compelled to regenerate inappropriate images. + +Inspired by the idea of adversarial fine-tuning, we propose a Reliable and Efficient Concept Erasure (RECE) method to address the aforementioned challenge, which continually finds new embeddings of the erased concepts during fine-tuning and then enables the unlearned model to erase these new concept embeddings. To speed up the unlearning process, the RECE method builds upon the previous fast and efficient concept erasure method UCE [7], which employs a closed-form editing to only modify the key and value projection matrices in cross-attention layers [33]. Similarly, the RECE method derives new embeddings that most effectively prompt the model to regenerate images of erased concepts, with a closed-form solution based on cross-attention output. Furthermore, a regularization term is introduced to preserve the image generation capability of the model by restricting the deviation of model parameters before and after modification. By editing the model and deriving embedding for multiple epochs, + +RECE enables the unlearned model to preserve the image generation ability of unerased concepts and robustly refrains the model from generating images with erased concept content. All the processes above are in closed-form, guaranteeing extremely efficient erasure in 3 seconds. Our major contributions are summarized as follows: + +- We present a novel concept erasure method - RECE that uses closed-form parameter editing and adversarial learning schemes for reliable and efficient concept erasing in only 3 seconds. +- RECE sufficiently erases concepts by deriving new embeddings that enable the unlearned model to regenerate erased concepts. In addition, a regularization term is introduced to minimize the impact on the model's capabilities. +- We conduct extensive experiments to validate the effectiveness of RECE for erasing unsafe contents, artistic styles and object classes. Additionally, we assess the robustness of RECE against three red-teaming tools and record fine-tuning durations to highlight the efficiency of RECE. + +# 2 Related Work + +T2I Diffusion Models with Safety Mechanisms. In response to the issue of generating inappropriate images in T2I diffusion models, several research has explored solutions to address this concern. Briefly, existing research primarily falls into the following three distinct strategies: The first is filtering the training data and retraining the model [22]. However, retraining on curated datasets not only requires substantial computational resources investment but also results in the generation of inappropriate content [6] and performance degradation [29]. The second is censoring model output through safety checkers [21], or exploiting classifier-free guidance to steer the latent codes away from inappropriate concepts during inference [29]. However, in the case of open-source code, pretrained T2I diffusion model architectures and parameters are publicly available, so such post-hoc intervention strategies can be easily circumvented by malicious users [23]. The third is fine-tuning the partial parameters of the pre-trained T2I diffusion models to erase the model's representation capability of inappropriate concepts [6-8, 13]. While fine-tuning has been considered an effective strategy to prevent the generation of inappropriate content, current methods consume a lot of computing time and can be easily bypassed by red-teaming tools for T2I diffusion models. + +Red-Teaming Tools for T2I Diffusion Models. With the recent popularity of AI, red-teaming has been applied to AI models to enhance model stability by probing functional vulnerabilities [1, 2, 34, 35]. Recent works have also developed red-teaming tools for T2I diffusion models, which is a rarely explored field in AI red-teaming. For instance, Prompting4Debugging (P4D) [3] automatically finds the problematic prompts that would lead to inappropriate content via utilizing prompt engineering techniques and an auxiliary diffusion model + +without any safety mechanisms to assess the reliability of deployed safety mechanisms. Conversely, UnlearnDiff [38] does not depend on an auxiliary diffusion model, which leverages the inherent classification capabilities of diffusion models, thereby providing computational efficiency without sacrificing effectiveness. Both works have the main weakness in assuming white-box access to the target model. In response to this issue, Ring-A-Bell [32], a model-agnostic framework capable of constructing attacks without prior knowledge of the target model. Specifically, Ring-A-Bell first performs concept extraction to obtain a holistic representation of inappropriate concepts. Subsequently, Ring-A-Bell automatically produces problematic prompts by leveraging the extracted concepts. + +# 3 Method + +# 3.1 Preliminaries + +Text-to-Image (T2I) Diffusion Models In contemporary Text-to-Image (T2I) applications, diffusion models have become the preferred choice [36] since the progressive denoising process [10] empowers them with superior image synthesis ability [4]. To reduce computational complexity, T2I often adopts latent diffusion models [12, 24, 27, 28], which operates on the low-dimensional latent space of a pre-trained variational autoencoder (VAE) [5] and employs a U-Net generative network as the denoising architecture [26]. To incorporate text conditioning into the image generation process, T2I encodes text by language models like CLIP [19, 24] and integrates text embeddings into U-Net through cross-attention layers. Specifically, these layers employ Query-Key-Value (QKV) structure [33] to represent the interactions between text and vision. For a given text embedding $c_{i}$ , keys and values are generated as $k_{i} = W_{k}c_{i}$ and $v_{i} = W_{v}c_{i}$ . These keys compute an attention map by multiplying with the query $q_{i}$ representing visual features, and then the cross-attention output is computed by attending over values $v_{i}$ : + +$$ +\mathcal {O} \propto \operatorname {s o f t m a x} (q _ {i} k _ {i} ^ {T}) v _ {i}. \tag {1} +$$ + +Concept Erasing with Closed-form Solution There are existing erasure methods requiring fine-tuning, such as ESD, CA and SA [6,8,13]. However, such methods are relatively inefficient as they require thousands of fine-tuning steps. In contrast, UCE [7] is an efficient method which modifies the attention weights through a closed-form edit. UCE requires a "source" concept (e.g., "nudity") and a "destination" concept (e.g., empty text ""). Let $c_{i}$ represent the source embedding, $c_{i}^{*}$ denote the corresponding destination embedding, set $E$ denote concepts to erase, and set $P$ denote concepts to preserve. Given a K/V projection matrix $W^{\mathrm{old}}$ (a concise notation for $W_{k}^{\mathrm{old}}$ and $W_{v}^{\mathrm{old}}$ ), UCE seeks new weights $W$ by editing concepts in $E$ while preserving concepts in $P$ . Specifically, the objective is to find weights such that the output $Wc_{i}$ for $c_{i} \in E$ aligns with target values $W^{\mathrm{old}}c_{i}^{*}$ instead of the original $W^{\mathrm{old}}c_{i}$ . Meanwhile, to control parameter + +![](images/c86ae884f2603a05e25af154bd26d60878181de1845b0ae7ce4bd072161e0486.jpg) +Model Editing + +![](images/9add03c9bdefba80762a644020336df9bd605f4c6efa04864ec39d86dfb327b6.jpg) +Embedding Derivation +Fig. 1: Overview of the proposed RECE. RECE consists of two main components: model editing and embedding derivation. First, erasing concepts by editing the model with a closed-form solution, and obtaining the edited cross-attention $W^{\mathrm{new}}$ . Then, new embedding $c_{i}^{\prime}$ can be derived by Eq. (7) given the original cross attention $W^{\mathrm{old}}$ and the edited $W^{\mathrm{new}}$ . In subsequent epochs, model editing and embedding derivation are looped. + +changes, outputs for $c_{j} \in P$ are preserved as $W^{\mathrm{old}}c_{j}$ and a L2 regularization term is introduced: + +$$ +\min _ {W} \sum_ {c _ {i} \in E} | | W c _ {i} - W ^ {\text {o l d}} c _ {i} ^ {*} | | _ {2} ^ {2} + \lambda_ {1} \sum_ {c _ {j} \in P} | | W c _ {j} - W ^ {\text {o l d}} c _ {j} | | _ {2} ^ {2} + \lambda_ {2} | | W - W ^ {\text {o l d}} | | _ {F} ^ {2}, \tag {2} +$$ + +where $\lambda_{1}$ and $\lambda_{2}$ are scaling factors preserving the existing concepts. UCE [7] prove that this formula has a closed-form solution: + +$$ +W = W ^ {\text {o l d}} \left(\sum_ {c _ {i} \in E} c _ {i} ^ {*} c _ {i} ^ {T} + \lambda_ {1} \sum_ {c _ {j} \in P} c _ {j} c _ {j} ^ {T} + \lambda_ {2} I\right) \left(\sum_ {c _ {i} \in E} c _ {i} c _ {i} ^ {T} + \lambda_ {1} \sum_ {c _ {j} \in P} c _ {j} c _ {j} ^ {T} + \lambda_ {2} I\right) ^ {- 1}. \tag {3} +$$ + +UCE directly assigns cross-attention KV matrices using the closed-form solution, eliminating the need for fine-tuning. This makes UCE significantly faster, hence we use UCE in our method. + +# 3.2 Reliable and Efficient Concept Erasure (RECE) + +While UCE [7] offers a fast solution for removing undesired concepts from T2I diffusion models, it can still produce undesired content, as illustrated in Tab. 1. This suggests an incomplete erasure of these concepts. To effectively eliminate such undesired concepts, we efficiently erase closed-form embeddings capable of regenerating erased concepts within the unlearned model. The derivation of embeddings and erasure is conducted iteratively to achieve a thorough erasure of inappropriate concepts, as shown in Fig. 1. + +Finding Target Contents Let us take "nudity" for example. As depicted in the second column of Fig. 2, when directly providing the input prompt "nudity" to UCE models, only landscape or unrelated images are generated. This is because the word "nudity" has been aligned with the empty text". However, the erasure of UCE is incomplete. We can generate an adversarial prompt that enables UCE's model to produce images containing nudity content again, similar to those generated by SD when provided with the prompt "nudity". + +In this section, we introduce our method for deriving the new embedding in UCE's model, which guides UCE to generate nude images. As + +elaborated in Sec. 3.1, T2I introduces text embeddings into image generation through cross-attention layers, where the projection matrices $W_{k}$ and $W_{v}$ are used to transform text embeddings. Let $W^{\mathrm{old}}$ denote the projection matrices of the original U-Net before UCE's editing, $W^{\mathrm{new}}$ represent the projection matrices after UCE's editing, $c$ denote the embedding of "nudity", and $c'$ signify our derived embedding. If we can find a $c'$ such that $W^{\mathrm{new}}c'$ closely resembles $W^{\mathrm{old}}c$ , then $c'$ can guide the edited model to generate nude images like how $c$ guides the original model. More precisely, the objective function is formulated as follows: + +$$ +\min _ {c ^ {\prime}} \sum_ {i} \| W _ {i} ^ {\mathrm {n e w}} c ^ {\prime} - W _ {i} ^ {\mathrm {o l d}} c \| _ {2} ^ {2}, \tag {4} +$$ + +where $W_{i}$ denotes K/V cross-attention matrices of U-Net. The solution $c'$ derived from Eq. (4) can be viewed as the actual representation of $c$ within the edited model. Evidently, Eq. (4) represents a convex function with respect to $c'$ , which possesses a unique global minimum. As derived in Appendix A, Eq. (4) admits a closed-form solution: + +$$ +c ^ {\prime} = \left(\sum_ {i} W _ {i} ^ {\mathrm {n e w} ^ {T}} W _ {i} ^ {\mathrm {n e w}}\right) ^ {- 1} \left(\sum_ {i} W _ {i} ^ {\mathrm {n e w} ^ {T}} W _ {i} ^ {\mathrm {o l d}}\right) c. \tag {5} +$$ + +Given that text conditioning works in the form of embedding in Stable Diffusion (SD), we can use our derived embedding as text conditioning. As illustrated in the third column of Fig. 2, this derived embedding effectively guides UCE's edited model to once again generate nude images, indicating its capacity to represent the concept of "nudity" within edited model. Thus it demonstrates that the erasure process of UCE remains incomplete. To address this issue, we further + +![](images/baf14b026fad6d13d18d70d1506374992031d69b73d505a07c188881514cb37c.jpg) +Fig.2: When given the input prompt "nudity", SD generates images containing nudity content, while UCE generates unrelated images. When given our derived embedding and regularized embedding, UCE generates nude images again. We use for publication purposes. + +remove our derived embeddings $c'$ from UCE's model with Eq. (3) to prevent the generation of nude images. + +Regularization Term We can further erase the concept of "nudity" by substituting $c$ in Eq. (2) with our derived embedding $c'$ . However, upon directly erasing $c'$ , we observe a significant decline in the model's performance: it struggles to generate high-quality images for unrelated concepts as shown in Fig. 6. Hence, it becomes imperative to devise a method that retains the model's performance while erasing concepts. + +Let $W^{\mathrm{new1}}$ denote the projection matrices after the last epoch's modification, and $W^{\mathrm{new2}}$ denote the projection matrices after the current epoch. Partly, preserving the model's performance entails minimizing the impact on unrelated concepts. Consequently, we define our objective function as follows: + +$$ +\min _ {W} \| W ^ {\mathrm {n e w} 2} d - W ^ {\mathrm {n e w} 1} d \| _ {2} ^ {2}, \tag {6} +$$ + +where $d$ represents an unrelated concept's embedding. Note that $W^{\mathrm{new2}}$ is directly influenced by $c'$ , as it is derived after erasing $c'$ from $W^{\mathrm{new1}}$ . We can obtain a theorem about our objective and its proof is provided in Appendix B: + +Theorem 1. If $c'$ is set to 0, Eq. (6) achieves its global minimum of 0. + +Intuitively, if $c_{i}$ in Eq. (3) is set to zero, the coefficient matrices on the right side will become an identity matrix when multiplied. As a result, $W$ will revert to being equivalent to $W^{\mathrm{old}}$ , exerting minimal influence on unrelated concepts due to the absence of further modifications to $W$ . Therefore, we need to introduce a regularization term to the original objective Eq. (4) to ensure that the obtained $c'$ is close to zero, thereby minimizing its influence on the model's performance: + +$$ +\min _ {c ^ {\prime}} \sum_ {i} \| W _ {i} ^ {\text {n e w}} c ^ {\prime} - W _ {i} ^ {\text {o l d}} c \| _ {2} ^ {2} + \lambda \| c ^ {\prime} \| _ {2} ^ {2}. \tag {7} +$$ + +As derived in Appendix A, this final objective function also possesses a unique global minimum solution: + +$$ +c ^ {\prime} = \left(\lambda I + \sum_ {i} W _ {i} ^ {\mathrm {n e w} ^ {T}} W _ {i} ^ {\mathrm {n e w}}\right) ^ {- 1} \left(\sum_ {i} W _ {i} ^ {\mathrm {n e w} ^ {T}} W _ {i} ^ {\mathrm {o l d}}\right) c \tag {8} +$$ + +As illustrated in the fourth column of Fig. 2, this regularized embedding can also guide UCE's model to once again generate nude images, indicating its ability to represent the concept "nudity" within UCE's model. With the incorporation of our regularization term, we iteratively apply the erasure process to the refined embedding $c'$ using Eq. (3) over multiple epochs. This ensures thorough concept erasure while safeguarding the overall performance of the model. The algorithm details are elaborated in Algorithm 1. + +Algorithm 1: Erase Concepts with RECE +Input: Diffusion U-Net $\theta$ , concepts set $E$ to erase, $P$ to preserve, epochs $T$ Output: Diffusion U-Net $\theta^{\prime}$ with concepts $E$ erased. +1 /\* Initialize\*/ +2 $\theta^{\prime}\gets \theta$ . +3 Initialize text embeddings $c_{i}$ and $c_{j}$ from $E$ and $P$ . +4 Extract K&V matrices $W^{\mathrm{old}}$ from the cross attention of $\theta$ . +5 /\* Preliminary erasing with UCE\*/ +6 Obtain $W^{\mathrm{new}}$ by erasing concepts in $E$ with Eq. (3); +7 for $t = 1,\ldots ,T$ do +8 /\* Derive new embeddings\*/ +9 $E^{\prime}\gets \{\}$ . +for $c_{i}\in E$ do +11 Derive new embedding $c_i^\prime$ using $W^{\mathrm{new}}$ with Eq. (8); +12 $\begin{array}{r}E^{\prime}\leftarrow E^{\prime}\cup \{c_{i}^{\prime}\} \end{array}$ +13 /\* Erasing derived embeddings\*/ +14 Obtain $W^{\mathrm{new}}$ by erasing $E^{\prime}$ with Eq. (3); +15 $W^{\mathrm{new}}\gets W^{\mathrm{new}};$ +16 Update $\theta^{\prime}$ by replacing K&V matrices with $W^{\mathrm{new}}$ . +17 return $\theta^{\prime}$ + +# 4 Experiments + +In this section, we present the results of our method for erasing inappropriate concepts and artistic styles. We also include the results of object removal in Appendix. We start with SD V1.4 as our base model. Following the implementation in [16], we set $\lambda_{1}$ and $\lambda_{2}$ in Eq. (2) to 0.1. For inappropriate concepts, we perform iterative erasure for 5 epochs and set $\lambda$ in Eq. (7) to $1e - 1$ . For artistic style, we conduct erasure for 10 epochs and set $\lambda$ to $1e - 3$ . The baselines we will compare with are: SD V1.4 [25], SD V2.1 [31](Stable Diffusion pretrained on an NSFW filtered dataset), SLD [29], ESD [6], CA [13], SA [8], UCE [7]. As for SLD, ESD, SA and UCE, we adhere to the recommended configuration in their papers [6-8,29]. For CA [13], we fine-tune the full weights of U-Net to erase unsafe contents and the cross-attention module to erase artistic styles, according to its documentation3. + +# 4.1 Unsafe Content Removal + +Experimental Setup In this section, we assess the effectiveness of erasing unsafe concepts. We conduct experiments on the Inappropriate Image Prompts (I2P) dataset [29]. The I2P dataset includes various inappropriate prompts, such as violence, self-harm, sexual content, and shocking content. These prompts are collected from real-world, user-generated images based on the official SD. Our + +
MethodNudity DetectionCOCO-30k
Breast(F)Genitalia(F)Breast(M)Genitalia(M)ButtocksFeetBellyArmpitsTotal↓CLIP↑FID↓
SD v1.4183214610444217112964631.33-
SD v2.1121134031439146109485--
ESD-u14185524313312130.453.73
UCE3161981120553618631.261.82
SLD-Med72534318191049935430.952.60
SA399401532491516330.5717.34
CA6191041428239531.167.87
Ours80640823176630.952.82
+ +Table 1: Comparison of performance metrics for content removal methods. (Left) Number of nude body parts detected by Nudenet on I2P dataset with threshold 0.6. (Right) CLIP-score and FID against original SD. F: Female. M: Male. Bold: best. Underline: second-best. + +evaluation focuses on the erasure of nudity since it is a classical unsafe concept. For each model, we generate one image per prompt in the I2P dataset, resulting in a total of 4703 images. Nude body parts are detected using the Nudenet detector [18], with the threshold set to 0.6. This threshold follows the default settings in $\mathrm{I2P^4}$ . + +To verify that the unlearned models can still generate normal images, we use COCO-30k [14] with its captions as prompts. COCO-30k is a dataset devoid of unsafe concepts, making it suitable for evaluating edited models' generation capabilities. We evaluate the models' image-text consistency based on CLIP-score [9], and visual similarity against SD-generated images based on FID [17]. + +Removal Results As depicted in Tab. 1, our method yields the lowest number of nude body parts, while demonstrating impressive specificity in preserving normal content of COCO-30k. CA generates the second-fewest nude body parts but it exhibits poorer performance in terms of FID, indicating a poorer trade-off between generation ability and removal effectiveness. On the other hand, CA fine-tunes full weights and our RECE only fine-tunes cross-attention modules, which will be discussed in detail in Sec. 4.4. + +Notably, our method achieves a FID score closely comparable to the top-performing UCE and the second-best SLD, both of which generates a considerably higher number of nude body parts. This suggests that our method minimally impacts the generation of normal content while striving for better removal effectiveness. Additionally, most methods exhibits favorable CLIP-score results thus we consider the CLIP-score performance acceptable as long as it remains within a reasonable range. + +In open-sourced conditions, the inference guidance mechanism such as SLD [29] can be easily bypassed by deleting the corresponding code [24]. Large-scale model retraining on NSFW-filtered dataset demands considerable computational re + +![](images/a8f83f03d442eab7b7a57107a1913358fbae6b46c5ab152a0b9cd23ee6385333.jpg) +SD v1.4 + +![](images/6275885888c8026fe5df7b6b96838c6ae6503bb8bfd86262ccbef83e1bd21a3d.jpg) +SD v2.1 + +![](images/aab5cba63815acac99ab09c739109b4d533a5cc480081220d4817399f1e3a8d0.jpg) +ESD + +![](images/7a07d949f919d07dedb8c0220f1b59c63a6bd00d8dd0ec66c1d2fcfe32796fe1.jpg) +UCE +Fig. 3: Qualitative results of different nudity removing methods. Images for each row are generated using the same prompt. All prompts originate from I2P benchmark, and we don't intentionally generate inappropriate content of any individual. + +![](images/a6f700265c773b2839d4c4b7851947f130832029ca90abf94f07f7bd312802fd.jpg) +SLD-Med + +![](images/c6e74b0bff7785af9f1dcd9330dfc281a6b443797041c536576602838384fe65.jpg) +SA + +![](images/beac8009cce77f2f1f1d6ebf31b54ab49ffa7ce51fd6d950f27bcfa24577e05a.jpg) +CA + +![](images/9a531224f258d0878a7814fa61bf3b215c0927b55a26d99e254ff4c5820a0fba.jpg) +Ours + +sources [22], and even then, the model SD v2.1 may still generate nude images, as illustrated in Tab. 1. + +Qualitative results are illustrated in Fig. 3. In the first row, all erasing methods successfully generate non-nude images, but the results of CA and SA differ significantly from the original SD's. In the second row, all methods except for SD v2.1 avoid generating nude images. Among these, only UCE, SLD, and our method effectively capture the facial features. These findings demonstrate that our method effectively maintains unrelated concepts. + +Nudity Bias While our RECE demonstrates remarkable effectiveness in minimizing the generation of nude content, it exhibits limitations in erasing male nudity and similar limitations are also observed in other methods. We count the number of female and male nude body parts in the 4703 I2P images. As illustrated in Tab. 2, the nudity ratios between women and men almost decrease in every method, indicating an erasure bias on female information (excluding SA, which erases more sex-related concepts besides nudity [8]). We attribute this limitation to the inherent bias in the target concept "nudity" within SD, which tends to generate more female-oriented content. To investigate this bias, we randomly selected 20 seeds and employed SD V1.4 to generate 3 images per seed with the prompt "nudity", resulting in a total of 60 images. Surprisingly, almost all of these generated images depict female body part as presented in the last column of Tab. 2. Thus our derived embedding is also biased. Further improvements require an awareness of the biases inherent in the model while performing erasure. + +# 4.2 Artistic Style Removal + +Experimental Setup We conduct an evaluation to assess the efficacy of removing artistic styles to address copyright concerns. Following the datasets in ESD [6], we use 20 prompts for each of 5 famous artists—Van Gogh, Pablo Picasso, Rembrandt, Andy Warhol and Caravaggio—and 5 modern artists—Kelly + +
MetricsSDESDUCESLD-MedSACAOursSD- "nudity"
# Female Nudity204153777487889
# Male Nudity56132737419101
Female/Male3.641.151.372.08120.370.889
+ +Table 2: In the first 7 columns, we counted the numbers of female nudity and male nudity body parts for I2P dataset. The last column is the result of SD v1.4 conditioned by the prompt "nudity". This table highlights the inherent bias within SD. And all each erasure methods can not remove the concept of male nudity very well. + +
Removal MethodErase "Van Gogh"Erase "Kelly McKernan"
LPIPSe↑LPIPSu↓LPIPSd↑LPIPSe↑LPIPSu↓LPIPSd↑
ESD-x0.400.260.140.370.210.16
UCE0.250.050.200.250.030.22
SLD-Med0.210.100.110.220.180.04
CA0.300.130.170.220.170.05
Ours0.310.080.230.290.040.25
+ +Table 3: Comparison of LPIPS scores for artistic removal methods. Bold: best. Underline: second-best. $\mathrm{LPIPS_d}$ indicates overall erasure performance. + +McKernan, Thomas Kinkade, Tyler Edlin, Kilian Eng and the series "Ajin: Demi-Human", which have been reported to be imitated by SD [30]. To evaluate our RECE and all the aforementioned baselines, we erase the style of two artists: Van Gogh and Kelly McKernan. + +Removal Results We conducted an evaluation based on LPIPS scores [37] compared to the original SD, as detailed in Tab. 3. LPIPS evaluates the perceptual distance between image patches, where higher values indicate greater differences and lower values indicate more similarity. The $\mathrm{LPIPS_e}$ is calculated on the erased artist. A higher $\mathrm{LPIPS_e}$ value suggests a more effective style removal, and both ESD and our method demonstrate successful erasure of the target style. $\mathrm{LPIPS_u}$ is calculated on unerased artists. A lower $\mathrm{LPIPS_u}$ indicates a lesser impact on unrelated artists, where our method and UCE effectively maintains unrelated concepts. We also calculate the overall effectiveness by $\mathrm{LPIPS_d} = \mathrm{LPIPS_e} - \mathrm{LPIPS_u}$ , which is the difference between erased and unerased artists. Our method performs best in this regard. Qualitative results can be found in Appendix. + +# 4.3 Robustness Against Red-teaming Tools + +Experimental Setup To demonstrate the robustness of our RECE in safeguarding against various attack methods, we employ different red-teaming tools, including white-box methods such as P4D [3] and UnlearnDiff [38], and the black-box method Ring-A-Bell [32]. For nudity, as provided by UnlearnDiff [38], + +
Red-teamingSD v1.4SD v2.1ESD-uUCESLD-MaxSACAOurs
UnlearnDiff--66.2079.5882.3977.4665.4965.49
P4D--63.3880.2877.4678.8760.5664.79
Ring-A-Bell83.1072.5469.7233.1066.2022.5425.3513.38
Average--66.4364.3275.3559.6250.4747.89
+ +Table 4: Robustness of different methods against red-teaming tools, measured by attack success rate(%) . Bold: best. Underline: second-best. + +![](images/b4f3e977ed9cae46da51871eed38732622a37cb76ef6a80a31a06e14db7f2c19.jpg) +Original SD +Fig. 4: Qualitative results of different artistic style removal methods for eliminating the style of Van Gogh under Ring-A-Bell's attack. The prompt is "A depiction of a starry night over a quiet town, reminiscent of Van Gogh's famous painting". + +![](images/3c1c6670da3aef5f747eebd760738458c054f438a4104715ab56c75191d6e383.jpg) +ESD + +![](images/f0921bb0eb7a2e44276d351feb20046c2d7bccf655d6781c8f472aefe08287b5.jpg) +UCE + +![](images/54906c9f9f6e704295990d1c6fcf3f84456d64e8e201f51a0003e3b04c748307.jpg) +SLD-Med + +![](images/6c108cc06dcd28666d0902f3554b48090f60776ed47d1e4779b7c7cb629c7349.jpg) +CA + +![](images/9cf2bdccd184a9659d37be6304ae5d6463b2cf4c9d41dd534f318737f24a6205.jpg) +Ours + +we use a set of 143 prompts selected from I2P, each with a nudity score (as determined by NudeNet) above 0.75, and employ the Nudenet detector with a threshold set to 0.45 for detecting inappropriate content. + +Results The attack success rates (ASR, %) are summarized in Tab. 4. Our method achieves the best robustness in average. In the case of the black-box attack Ring-A-Bell, our method achieves the lowest ASR at $13.38\%$ , significantly outperforming other methods. In the case of the white-box attack, CA achieves the best performance, while our method performs either the best or very closely to the second-best. Although SA achieves a decent black-box result, it consumes substantial computation resources for generating 5000 prepared images [8]. CA modifies $100\%$ of the U-Net parameters while our method only modifies K&V projection matrices, constituting a mere $2.23\%$ of the U-Net. While UCE also modifies only $2.23\%$ parameters, our method significantly outperforms UCE. This is attributed to our derived embeddings in Eq. (8). + +For artistic style removal, we provide qualitative examples in Fig. 4. The first image is generated by the original SD v1.4 without any attack, and the following images are under Ring-A-Bell's attack [32]. Recall that UCE is the second-best artistic style removal method as shown in Tab. 3 but it falls short in robustness against red-teaming tools. Although our method employs a closed-form solution like UCE, it outperforms UCE in robustness. ESD, CA and our method perform similarly well. More results can be found in the appendix. + +# 4.4 Model Editing Duration + +To demonstrate the efficiency of different methods, we measured the percentage of parameter modification and editing duration on an RTX 3090 for each + +
ESDUCESACAOurs
Modification (%)94.652.2394.651002.23
Duration (s)37201.2-14003.4
+ +Table 5: Percentage of parameter modification and model editing duration for different methods. Our method and UCE are significantly ahead of other methods. + +method, as shown in Tab. 5. We excluded SLD from the analysis since it operates at inference time rather than modifying the model's weights which can be easily bypassed under open-source conditions. Additionally, we don't include the duration of SA, as it involves generating 5000 images, calculating the Fisher Information Matrix and fine-tuning, which makes it exceptionally slow. + +Based on Tab. 5 and Tab. 1, our method achieves the best concept erasure effect in an extremely short time of only three seconds. Our method (5 epochs) and UCE (1 epoch) modify the lowest percentage of parameters with a closed-form solution, resulting in the shortest editing durations. Despite similar durations, our method significantly outperforms UCE in removal effectiveness. Conversely, CA, ESD and SA modify a high percentage of parameters with more time but achieve less impressive removal results. + +# 4.5 Ablation Study + +We conduct experiments to elucidate the impact of our derived embedding among different epochs and the effectiveness of the regularization term. + +# Effect of Derived Embeddings among Epochs + +We conduct an experiment to expound the impact of our derived embedding. We perform "nudity" erasure for 5 epochs using Algorithm 1. In each epoch, we derive a distinct embedding that represents "nudity". Before the erasure of each + +epoch, we generate an image using the embedding to test its degree of nudity information, as presented in Fig. 5. Images from epoch 0 to epoch 2 contain nude body parts, indicating that our derived embeddings successfully reveal potential nudity information in the model. Specifically, we opt for the checkpoint after epoch 2, as images from epoch 3 to 4 lack nudity information. Actually, erasing such "not so nude" embeddings in epoch 3-4 would impair the model's normal generation ability, which is an unworthy trade-off. + +![](images/a6cd6fddd50fd3f3a1017cdb441762c6eabcc41be1c916ab35ab5ac81f034c98.jpg) +Fig. 5: There is no nudity information from epoch 3 onward; hence, we select the checkpoint after epoch 2 to avoid damaging the model's ability. + +![](images/09f9728f5f221f3a57c35ee85a361d4e48298663c7ed1c8bb330ce059fead620.jpg) +Fig. 6: Ablation study on the impact of the regularization term on model performance. + +Effect of Regularization Coefficient We conduct an experiment to assess the influence of our regularization term. Specifically, we select different regularization coefficients $\lambda$ in Eq. (4), which divide the interval [0, 1] into five parts. The results, presented in Fig. 6, include CLIP-score and LPIPS against the original SD on the COCO-30k validation subset. As $\lambda$ increases, CLIP score shows an upward trend while LPIPS and the difference between new and old parameters show a downward trend. This indicates the role of the regularization term in preserving the model's ability for unerased content. Furthermore, we recorded the number of nude parts on the I2P benchmark, presented in the third column of Fig. 6. However, the number of nude parts doesn't strictly increase as $\lambda$ increases, which is counterintuitive. Although the purpose of the regularization term in Eq. (6) is to preserve the model's generation capability, maintaining this capability does not always affect the erasure effect. + +# 5 Conclusion + +In this paper, we propose a novel approach for reliably and efficiently erasing specific concepts from Text-to-Image (T2I) diffusion models. Our approach only modifies the cross-attention K&V matrices of U-Net, constituting a mere $2.23\%$ of parameters. While previous methods also edited cross-attention modules, they still exhibited the ability to generate inappropriate images. To tackle this challenge, we derive and erase new embeddings that can represent target concepts within unlearned models. To mitigate the impact on unrelated concepts, a regularization term is introduced during the erasure process. All the above techniques are formulated in closed-form, facilitating rapid editing. This enables the execution of "derive-erase" across multiple epochs, ensuring thorough and robust erasure. Extensive experiments were conducted to validate the effectiveness of our approach in erasing artistic styles, unsafe contents and object classes. Furthermore, we recorded editing durations to underscore the efficiency of our method and evaluated the robustness against red-teaming tools. We believe our RECE has the potential to empower T2I providers in effectively removing undesired concepts, thereby fostering the development of a safer AI community. + +# Acknowledgements + +This project was supported by National Key R&D Program of China (No. 2021ZD0112804). + +# References + +1. Chen, K., Wei, Z., Chen, J., Wu, Z., Jiang, Y.G.: Attacking video recognition models with bullet-screen comments. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 36, pp. 312-320 (2022) +2. Chen, K., Wei, Z., Chen, J., Wu, Z., Jiang, Y.G.: Gcma: Generative cross-modal transferable adversarial attacks from images to videos. In: Proceedings of the 31st ACM International Conference on Multimedia. pp. 698-708 (2023) +3. Chin, Z.Y., Jiang, C.M., Huang, C.C., Chen, P.Y., Chiu, W.C.: Prompting4debugging: Red-teaming text-to-image diffusion models by finding problematic prompts. arXiv preprint arXiv:2309.06135 (2023) +4. Dhariwal, P., Nichol, A.: Diffusion models beat gans on image synthesis. Advances in neural information processing systems 34, 8780-8794 (2021) +5. Doersch, C.: Tutorial on variational autoencoders. arXiv preprint arXiv:1606.05908 (2016) +6. Gandikota, R., Materzynska, J., Fiotto-Kaufman, J., Bau, D.: Erasing concepts from diffusion models. arXiv preprint arXiv:2303.07345 (2023) +7. Gandikota, R., Orgad, H., Belinkov, Y., Materzyńska, J., Bau, D.: Unified concept editing in diffusion models. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 5111-5120 (2024) +8. Heng, A., Soh, H.: Selective amnesia: A continual learning approach to forgetting in deep generative models. Advances in Neural Information Processing Systems 36 (2024) +9. Hessel, J., Holtzman, A., Forbes, M., Bras, R.L., Choi, Y.: Clipscore: A reference-free evaluation metric for image captioning. arXiv preprint arXiv:2104.08718 (2021) +0. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. Advances in neural information processing systems 33, 6840-6851 (2020) +1. Hunter, T.: Ai porn is easy to make now. for women, that's a nightmare. (2 2023) +2. Kawar, B., Zada, S., Lang, O., Tov, O., Chang, H., Dekel, T., Mosseri, I., Irani, M.: Imagic: Text-based real image editing with diffusion models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6007-6017 (2023) +3. Kumari, N., Zhang, B., Wang, S.Y., Shechtman, E., Zhang, R., Zhu, J.Y.: Ablating concepts in text-to-image diffusion models. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 22691-22702 (2023) +4. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollar, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13. pp. 740-755. Springer (2014) +5. Nichol, A.Q., Dhariwal, P., Ramesh, A., Shyam, P., Mishkin, P., Mcgrew, B., Sutskever, I., Chen, M.: Glide: Towards photorealistic image generation and editing with text-guided diffusion models. In: International Conference on Machine Learning. pp. 16784-16804. PMLR (2022) + +16. Orgad, H., Kawar, B., Belinkov, Y.: Editing implicit assumptions in text-to-image diffusion models. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 7053-7061 (2023) +17. Parmar, G., Zhang, R., Zhu, J.Y.: On aliased resizing and surprising subtleties in gan evaluation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11410-11420 (2022) +18. Praneeth, B.: Nudenet: Neural nets for nudity classification, detection and selective censoring (2019) +19. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021) +20. Ramesh, A., Dhariwal, P., Nichol, A., Chu, C., Chen, M.: Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125 1(2), 3 (2022) +21. Rando, J., Paleka, D., Lindner, D., Heim, L., Tramér, F.: Red-teaming the stable diffusion safety filter. arXiv preprint arXiv:2210.04610 (2022) +22. Rombach, R.: Stable diffusion 2.0 release (November 2022) +23. Rombach, R.: Tutorial: How to remove the safety filter in 5 seconds (August 2022) +24. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10684-10695 (2022) +25. Rombach, R., Esser, P.: Stable diffusion v1-4 model card. Model Card (2022) +26. Ronneberger, O., Fischer, P., Brox, T.: U-net: Convolutional networks for biomedical image segmentation. In: Medical image computing and computer-assisted intervention-MICCAI 2015: 18th international conference, Munich, Germany, October 5-9, 2015, proceedings, part III 18. pp. 234-241. Springer (2015) +27. Ruiz, N., Li, Y., Jampani, V., Pritch, Y., Rubinstein, M., Aberman, K.: Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 22500-22510 (2023) +28. Saharia, C., Chan, W., Saxena, S., Li, L., Whang, J., Denton, E.L., Ghasemipour, K., Gontijo Lopes, R., Karagol Ayan, B., Salimans, T., et al.: Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems 35, 36479-36494 (2022) +29. Schramowski, P., Brack, M., Deiseroth, B., Kersting, K.: Safe latent diffusion: Mitigating inappropriate degeneration in diffusion models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 22522-22531 (2023) +30. Setty, R.: Ai art generators hit with copyright suit over artists' images (1 2023) +31. StabilityAI: Stable diffusion 2.1 model card. Model Card (2022) +32. Tsai, Y.L., Hsu, C.Y., Xie, C., Lin, C.H., Chen, J.Y., Li, B., Chen, P.Y., Yu, C.M., Huang, C.Y.: Ring-a-bell! how reliable are concept removal methods for diffusion models? arXiv preprint arXiv:2310.10012 (2023) +33. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. Advances in neural information processing systems 30 (2017) +34. Wei, Z., Chen, J., Goldblum, M., Wu, Z., Goldstein, T., Jiang, Y.G., Davis, L.S.: Towards transferable adversarial attacks on image and video transformers. IEEE Transactions on Image Processing 32, 6346-6358 (2023) + +35. Wei, Z., Chen, J., Wu, Z., Jiang, Y.G.: Adaptive cross-modal transferable adversarial attacks from images to videos. IEEE Transactions on Pattern Analysis and Machine Intelligence (2023) +36. Yang, L., Zhang, Z., Song, Y., Hong, S., Xu, R., Zhao, Y., Zhang, W., Cui, B., Yang, M.H.: Diffusion models: A comprehensive survey of methods and applications. ACM Computing Surveys 56(4), 1-39 (2023) +37. Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual metric. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 586-595 (2018) +38. Zhang, Y., Jia, J., Chen, X., Chen, A., Zhang, Y., Liu, J., Ding, K., Liu, S.: To generate or not? safety-driven unlearned diffusion models are still easy to generate unsafe images... for now. arXiv preprint arXiv:2310.11868 (2023) \ No newline at end of file diff --git a/2024/Reliable and Efficient Concept Erasure of Text-to-Image Diffusion Models/images.zip b/2024/Reliable and Efficient Concept Erasure of Text-to-Image Diffusion Models/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..8ad95735c334a81e996d8498c759fbec0f882f48 --- /dev/null +++ b/2024/Reliable and Efficient Concept Erasure of Text-to-Image Diffusion Models/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1263490db9b75f9f7d9f7a2db88ae41d10e2e652ff8ed415aa6f2ea2776b0227 +size 404297 diff --git a/2024/Reliable and Efficient Concept Erasure of Text-to-Image Diffusion Models/layout.json b/2024/Reliable and Efficient Concept Erasure of Text-to-Image Diffusion Models/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..ab4d9711f461766d77d83b0ad1df64967bc38fbb --- /dev/null +++ b/2024/Reliable and Efficient Concept Erasure of Text-to-Image Diffusion Models/layout.json @@ -0,0 +1,9906 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 162, + 111, + 452, + 147 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 162, + 111, + 452, + 147 + ], + "spans": [ + { + "bbox": [ + 162, + 111, + 452, + 147 + ], + "type": "text", + "content": "Reliable and Efficient Concept Erasure of Text-to-Image Diffusion Models" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 132, + 167, + 482, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 167, + 482, + 194 + ], + "spans": [ + { + "bbox": [ + 132, + 167, + 482, + 194 + ], + "type": "text", + "content": "Chao Gong" + }, + { + "bbox": [ + 132, + 167, + 482, + 194 + ], + "type": "inline_equation", + "content": "^{1,2\\star}" + }, + { + "bbox": [ + 132, + 167, + 482, + 194 + ], + "type": "text", + "content": ", Kai Chen" + }, + { + "bbox": [ + 132, + 167, + 482, + 194 + ], + "type": "inline_equation", + "content": "^{1,2\\star}" + }, + { + "bbox": [ + 132, + 167, + 482, + 194 + ], + "type": "text", + "content": ", Zhipeng Wei" + }, + { + "bbox": [ + 132, + 167, + 482, + 194 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 132, + 167, + 482, + 194 + ], + "type": "text", + "content": ", Jingjing Chen" + }, + { + "bbox": [ + 132, + 167, + 482, + 194 + ], + "type": "inline_equation", + "content": "^{1,2\\dagger}" + }, + { + "bbox": [ + 132, + 167, + 482, + 194 + ], + "type": "text", + "content": ", and Yu-Gang Jiang" + }, + { + "bbox": [ + 132, + 167, + 482, + 194 + ], + "type": "inline_equation", + "content": "^{1,2}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 135, + 201, + 478, + 246 + ], + "type": "list", + "angle": 0, + "index": 4, + "blocks": [ + { + "bbox": [ + 136, + 201, + 476, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 201, + 476, + 224 + ], + "spans": [ + { + "bbox": [ + 136, + 201, + 476, + 224 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 136, + 201, + 476, + 224 + ], + "type": "text", + "content": " Shanghai Key Lab of Intell. Info. Processing, School of Computer Science, Fudan University" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 135, + 224, + 478, + 246 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 224, + 478, + 246 + ], + "spans": [ + { + "bbox": [ + 135, + 224, + 478, + 246 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 135, + 224, + 478, + 246 + ], + "type": "text", + "content": " Shanghai Collaborative Innovation Center on Intelligent Visual Computing {c Gong20, chenjingjing, ygj} @fudan.edu.cn, {kchen22, zpwei21} @m.fudan.edu.cn" + } + ] + } + ], + "index": 3 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 159, + 277, + 452, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 277, + 452, + 529 + ], + "spans": [ + { + "bbox": [ + 159, + 277, + 452, + 529 + ], + "type": "text", + "content": "Abstract. Text-to-image models encounter safety issues, including concerns related to copyright and Not-Safe-For-Work (NSFW) content. Despite several methods have been proposed for erasing inappropriate concepts from diffusion models, they often exhibit incomplete erasure, consume a lot of computing resources, and inadvertently damage generation ability. In this work, we introduce Reliable and Efficient Concept Erasure (RECE), a novel approach that modifies the model in 3 seconds without necessitating additional fine-tuning. Specifically, RECE efficiently leverages a closed-form solution to derive new target embeddings, which are capable of regenerating erased concepts within the unlearned model. To mitigate inappropriate content potentially represented by derived embeddings, RECE further aligns them with harmless concepts in crossattention layers. The derivation and erasure of new representation embeddings are conducted iteratively to achieve a thorough erasure of inappropriate concepts. Besides, to preserve the model's generation ability, RECE introduces an additional regularization term during the derivation process, resulting in minimizing the impact on unrelated concepts during the erasure process. All the processes above are in closed-form, guaranteeing extremely efficient erasure in only 3 seconds. Benchmarking against previous approaches, our method achieves more efficient and thorough erasure with minor damage to original generation ability and demonstrates enhanced robustness against red-teaming tools. Code is available at https://github.com/CharlesGong12/RECE." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 160, + 530, + 448, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 530, + 448, + 540 + ], + "spans": [ + { + "bbox": [ + 160, + 530, + 448, + 540 + ], + "type": "text", + "content": "WARNING: This paper contains model outputs that may be offensive." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 160, + 552, + 440, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 552, + 440, + 563 + ], + "spans": [ + { + "bbox": [ + 160, + 552, + 440, + 563 + ], + "type": "text", + "content": "Keywords: Text-to-Image " + }, + { + "bbox": [ + 160, + 552, + 440, + 563 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 552, + 440, + 563 + ], + "type": "text", + "content": " Concept Erasing " + }, + { + "bbox": [ + 160, + 552, + 440, + 563 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 552, + 440, + 563 + ], + "type": "text", + "content": " Machine Unlearning" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 584, + 230, + 596 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 584, + 230, + 596 + ], + "spans": [ + { + "bbox": [ + 132, + 584, + 230, + 596 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 610, + 487, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 610, + 487, + 635 + ], + "spans": [ + { + "bbox": [ + 130, + 610, + 487, + 635 + ], + "type": "text", + "content": "In recent years, large-scale text-to-image (T2I) diffusion models have exhibited remarkable capability in synthesizing photo-realistic images from text prompts [15," + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 142, + 643, + 235, + 653 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 142, + 643, + 235, + 653 + ], + "spans": [ + { + "bbox": [ + 142, + 643, + 235, + 653 + ], + "type": "text", + "content": "* Equal contributions." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 142, + 653, + 242, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 142, + 653, + 242, + 665 + ], + "spans": [ + { + "bbox": [ + 142, + 653, + 242, + 665 + ], + "type": "text", + "content": "† Corresponding author." + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 297 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 297 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 297 + ], + "type": "text", + "content": "[20, 24, 28]. The exceptional performance of T2I diffusion models is largely due to the vast amount of training data collected from the Internet, which enables the models to imitate a wide variety of concepts. Unfortunately, such powerful models can also be misused to generate copyright infringement and Not-Safe-For-Work (NSFW) image content when conditioned on inappropriate text prompts [11, 30]. Especially the open-source release of the Stable Diffusion (SD) T2I model has made advanced image generation technology widely accessible. To alleviate this safety concern, several recent research efforts have incorporated safety mechanisms into T2I diffusion models, e.g. filtering out inappropriate training data and retraining model [22], censoring model outputs with an NSFW safety checker [21], and applying classifier-free guidance to steer the generation away from inappropriate concepts [29]. However, these safety mechanisms either demand expensive computational resources and time [25] or can be easily circumvented by malicious users due to the public availability of code and model parameters in open-source scenario [23]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 301, + 482, + 504 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 301, + 482, + 504 + ], + "spans": [ + { + "bbox": [ + 130, + 301, + 482, + 504 + ], + "type": "text", + "content": "In response to the drawbacks mentioned above, an alternative is to erase inappropriate concepts from the T2I diffusion model [6-8, 13]. Specifically, given an inappropriate concept described in the text prompt, the pre-trained T2I diffusion model's parameters are fine-tuned to unlearn that concept so that the associated image content cannot be generated. Compared with previous security mechanisms, concept erasure neither requires training the entire model from scratch nor can be easily circumvented even in the case of open-source code. Despite promising progress in concept erasure, there exist several issues. On the one hand, most erasure methods require a high number of iterations to finetune considerable amounts of parameters [6, 8, 13], which inevitably degrades the generation capability and consumes a lot of computing resources. Only a recent work called UCE [7] modifies model parameters without fine-tuning using a closed-form solution, ensuring the model maintains original generation capability when erasing concepts. On the other hand, almost all methods fail to sufficiently erase inappropriate concepts, leaving them vulnerable to problematic prompts found by the red-teaming of T2I diffusion models [3, 32, 38]. This results in the unlearned model being compelled to regenerate inappropriate images." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "content": "Inspired by the idea of adversarial fine-tuning, we propose a Reliable and Efficient Concept Erasure (RECE) method to address the aforementioned challenge, which continually finds new embeddings of the erased concepts during fine-tuning and then enables the unlearned model to erase these new concept embeddings. To speed up the unlearning process, the RECE method builds upon the previous fast and efficient concept erasure method UCE [7], which employs a closed-form editing to only modify the key and value projection matrices in cross-attention layers [33]. Similarly, the RECE method derives new embeddings that most effectively prompt the model to regenerate images of erased concepts, with a closed-form solution based on cross-attention output. Furthermore, a regularization term is introduced to preserve the image generation capability of the model by restricting the deviation of model parameters before and after modification. By editing the model and deriving embedding for multiple epochs," + } + ] + } + ], + "index": 4 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 248, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 248, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 248, + 102 + ], + "type": "text", + "content": "F. Chao Gong et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 175 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 175 + ], + "type": "text", + "content": "RECE enables the unlearned model to preserve the image generation ability of unerased concepts and robustly refrains the model from generating images with erased concept content. All the processes above are in closed-form, guaranteeing extremely efficient erasure in 3 seconds. Our major contributions are summarized as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 184, + 479, + 301 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 138, + 184, + 479, + 219 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 184, + 479, + 219 + ], + "spans": [ + { + "bbox": [ + 138, + 184, + 479, + 219 + ], + "type": "text", + "content": "- We present a novel concept erasure method - RECE that uses closed-form parameter editing and adversarial learning schemes for reliable and efficient concept erasing in only 3 seconds." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 220, + 479, + 255 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 220, + 479, + 255 + ], + "spans": [ + { + "bbox": [ + 138, + 220, + 479, + 255 + ], + "type": "text", + "content": "- RECE sufficiently erases concepts by deriving new embeddings that enable the unlearned model to regenerate erased concepts. In addition, a regularization term is introduced to minimize the impact on the model's capabilities." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 139, + 256, + 479, + 301 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 256, + 479, + 301 + ], + "spans": [ + { + "bbox": [ + 139, + 256, + 479, + 301 + ], + "type": "text", + "content": "- We conduct extensive experiments to validate the effectiveness of RECE for erasing unsafe contents, artistic styles and object classes. Additionally, we assess the robustness of RECE against three red-teaming tools and record fine-tuning durations to highlight the efficiency of RECE." + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 132, + 323, + 237, + 335 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 323, + 237, + 335 + ], + "spans": [ + { + "bbox": [ + 132, + 323, + 237, + 335 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 348, + 482, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 348, + 482, + 563 + ], + "spans": [ + { + "bbox": [ + 130, + 348, + 482, + 563 + ], + "type": "text", + "content": "T2I Diffusion Models with Safety Mechanisms. In response to the issue of generating inappropriate images in T2I diffusion models, several research has explored solutions to address this concern. Briefly, existing research primarily falls into the following three distinct strategies: The first is filtering the training data and retraining the model [22]. However, retraining on curated datasets not only requires substantial computational resources investment but also results in the generation of inappropriate content [6] and performance degradation [29]. The second is censoring model output through safety checkers [21], or exploiting classifier-free guidance to steer the latent codes away from inappropriate concepts during inference [29]. However, in the case of open-source code, pretrained T2I diffusion model architectures and parameters are publicly available, so such post-hoc intervention strategies can be easily circumvented by malicious users [23]. The third is fine-tuning the partial parameters of the pre-trained T2I diffusion models to erase the model's representation capability of inappropriate concepts [6-8, 13]. While fine-tuning has been considered an effective strategy to prevent the generation of inappropriate content, current methods consume a lot of computing time and can be easily bypassed by red-teaming tools for T2I diffusion models." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "content": "Red-Teaming Tools for T2I Diffusion Models. With the recent popularity of AI, red-teaming has been applied to AI models to enhance model stability by probing functional vulnerabilities [1, 2, 34, 35]. Recent works have also developed red-teaming tools for T2I diffusion models, which is a rarely explored field in AI red-teaming. For instance, Prompting4Debugging (P4D) [3] automatically finds the problematic prompts that would lead to inappropriate content via utilizing prompt engineering techniques and an auxiliary diffusion model" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 146, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 146, + 91, + 447, + 102 + ], + "type": "text", + "content": "Reliable and Efficient Concept Erasure of Text-to-Image Diffusion Models" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "text", + "content": "without any safety mechanisms to assess the reliability of deployed safety mechanisms. Conversely, UnlearnDiff [38] does not depend on an auxiliary diffusion model, which leverages the inherent classification capabilities of diffusion models, thereby providing computational efficiency without sacrificing effectiveness. Both works have the main weakness in assuming white-box access to the target model. In response to this issue, Ring-A-Bell [32], a model-agnostic framework capable of constructing attacks without prior knowledge of the target model. Specifically, Ring-A-Bell first performs concept extraction to obtain a holistic representation of inappropriate concepts. Subsequently, Ring-A-Bell automatically produces problematic prompts by leveraging the extracted concepts." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 255, + 202, + 268 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 255, + 202, + 268 + ], + "spans": [ + { + "bbox": [ + 132, + 255, + 202, + 268 + ], + "type": "text", + "content": "3 Method" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 281, + 228, + 293 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 281, + 228, + 293 + ], + "spans": [ + { + "bbox": [ + 132, + 281, + 228, + 293 + ], + "type": "text", + "content": "3.1 Preliminaries" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 301, + 482, + 481 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 301, + 482, + 481 + ], + "spans": [ + { + "bbox": [ + 130, + 301, + 482, + 481 + ], + "type": "text", + "content": "Text-to-Image (T2I) Diffusion Models In contemporary Text-to-Image (T2I) applications, diffusion models have become the preferred choice [36] since the progressive denoising process [10] empowers them with superior image synthesis ability [4]. To reduce computational complexity, T2I often adopts latent diffusion models [12, 24, 27, 28], which operates on the low-dimensional latent space of a pre-trained variational autoencoder (VAE) [5] and employs a U-Net generative network as the denoising architecture [26]. To incorporate text conditioning into the image generation process, T2I encodes text by language models like CLIP [19, 24] and integrates text embeddings into U-Net through cross-attention layers. Specifically, these layers employ Query-Key-Value (QKV) structure [33] to represent the interactions between text and vision. For a given text embedding " + }, + { + "bbox": [ + 130, + 301, + 482, + 481 + ], + "type": "inline_equation", + "content": "c_{i}" + }, + { + "bbox": [ + 130, + 301, + 482, + 481 + ], + "type": "text", + "content": ", keys and values are generated as " + }, + { + "bbox": [ + 130, + 301, + 482, + 481 + ], + "type": "inline_equation", + "content": "k_{i} = W_{k}c_{i}" + }, + { + "bbox": [ + 130, + 301, + 482, + 481 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 301, + 482, + 481 + ], + "type": "inline_equation", + "content": "v_{i} = W_{v}c_{i}" + }, + { + "bbox": [ + 130, + 301, + 482, + 481 + ], + "type": "text", + "content": ". These keys compute an attention map by multiplying with the query " + }, + { + "bbox": [ + 130, + 301, + 482, + 481 + ], + "type": "inline_equation", + "content": "q_{i}" + }, + { + "bbox": [ + 130, + 301, + 482, + 481 + ], + "type": "text", + "content": " representing visual features, and then the cross-attention output is computed by attending over values " + }, + { + "bbox": [ + 130, + 301, + 482, + 481 + ], + "type": "inline_equation", + "content": "v_{i}" + }, + { + "bbox": [ + 130, + 301, + 482, + 481 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 260, + 491, + 481, + 506 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 260, + 491, + 481, + 506 + ], + "spans": [ + { + "bbox": [ + 260, + 491, + 481, + 506 + ], + "type": "interline_equation", + "content": "\\mathcal {O} \\propto \\operatorname {s o f t m a x} (q _ {i} k _ {i} ^ {T}) v _ {i}. \\tag {1}", + "image_path": "1ad52da22332a1cc08cc4054b704abd151523459d34623cb23c142d2176b7b14.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 521, + 482, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 482, + 668 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 482, + 668 + ], + "type": "text", + "content": "Concept Erasing with Closed-form Solution There are existing erasure methods requiring fine-tuning, such as ESD, CA and SA [6,8,13]. However, such methods are relatively inefficient as they require thousands of fine-tuning steps. In contrast, UCE [7] is an efficient method which modifies the attention weights through a closed-form edit. UCE requires a \"source\" concept (e.g., \"nudity\") and a \"destination\" concept (e.g., empty text \"\"). Let " + }, + { + "bbox": [ + 130, + 521, + 482, + 668 + ], + "type": "inline_equation", + "content": "c_{i}" + }, + { + "bbox": [ + 130, + 521, + 482, + 668 + ], + "type": "text", + "content": " represent the source embedding, " + }, + { + "bbox": [ + 130, + 521, + 482, + 668 + ], + "type": "inline_equation", + "content": "c_{i}^{*}" + }, + { + "bbox": [ + 130, + 521, + 482, + 668 + ], + "type": "text", + "content": " denote the corresponding destination embedding, set " + }, + { + "bbox": [ + 130, + 521, + 482, + 668 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 130, + 521, + 482, + 668 + ], + "type": "text", + "content": " denote concepts to erase, and set " + }, + { + "bbox": [ + 130, + 521, + 482, + 668 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 130, + 521, + 482, + 668 + ], + "type": "text", + "content": " denote concepts to preserve. Given a K/V projection matrix " + }, + { + "bbox": [ + 130, + 521, + 482, + 668 + ], + "type": "inline_equation", + "content": "W^{\\mathrm{old}}" + }, + { + "bbox": [ + 130, + 521, + 482, + 668 + ], + "type": "text", + "content": " (a concise notation for " + }, + { + "bbox": [ + 130, + 521, + 482, + 668 + ], + "type": "inline_equation", + "content": "W_{k}^{\\mathrm{old}}" + }, + { + "bbox": [ + 130, + 521, + 482, + 668 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 521, + 482, + 668 + ], + "type": "inline_equation", + "content": "W_{v}^{\\mathrm{old}}" + }, + { + "bbox": [ + 130, + 521, + 482, + 668 + ], + "type": "text", + "content": "), UCE seeks new weights " + }, + { + "bbox": [ + 130, + 521, + 482, + 668 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 130, + 521, + 482, + 668 + ], + "type": "text", + "content": " by editing concepts in " + }, + { + "bbox": [ + 130, + 521, + 482, + 668 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 130, + 521, + 482, + 668 + ], + "type": "text", + "content": " while preserving concepts in " + }, + { + "bbox": [ + 130, + 521, + 482, + 668 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 130, + 521, + 482, + 668 + ], + "type": "text", + "content": ". Specifically, the objective is to find weights such that the output " + }, + { + "bbox": [ + 130, + 521, + 482, + 668 + ], + "type": "inline_equation", + "content": "Wc_{i}" + }, + { + "bbox": [ + 130, + 521, + 482, + 668 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 130, + 521, + 482, + 668 + ], + "type": "inline_equation", + "content": "c_{i} \\in E" + }, + { + "bbox": [ + 130, + 521, + 482, + 668 + ], + "type": "text", + "content": " aligns with target values " + }, + { + "bbox": [ + 130, + 521, + 482, + 668 + ], + "type": "inline_equation", + "content": "W^{\\mathrm{old}}c_{i}^{*}" + }, + { + "bbox": [ + 130, + 521, + 482, + 668 + ], + "type": "text", + "content": " instead of the original " + }, + { + "bbox": [ + 130, + 521, + 482, + 668 + ], + "type": "inline_equation", + "content": "W^{\\mathrm{old}}c_{i}" + }, + { + "bbox": [ + 130, + 521, + 482, + 668 + ], + "type": "text", + "content": ". Meanwhile, to control parameter" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 248, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 248, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 248, + 102 + ], + "type": "text", + "content": "F. Chao Gong et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 137, + 113, + 302, + 222 + ], + "blocks": [ + { + "bbox": [ + 137, + 113, + 302, + 222 + ], + "lines": [ + { + "bbox": [ + 137, + 113, + 302, + 222 + ], + "spans": [ + { + "bbox": [ + 137, + 113, + 302, + 222 + ], + "type": "image", + "image_path": "c86ae884f2603a05e25af154bd26d60878181de1845b0ae7ce4bd072161e0486.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 198, + 224, + 237, + 233 + ], + "lines": [ + { + "bbox": [ + 198, + 224, + 237, + 233 + ], + "spans": [ + { + "bbox": [ + 198, + 224, + 237, + 233 + ], + "type": "text", + "content": "Model Editing" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 303, + 113, + 478, + 222 + ], + "blocks": [ + { + "bbox": [ + 303, + 113, + 478, + 222 + ], + "lines": [ + { + "bbox": [ + 303, + 113, + 478, + 222 + ], + "spans": [ + { + "bbox": [ + 303, + 113, + 478, + 222 + ], + "type": "image", + "image_path": "9add03c9bdefba80762a644020336df9bd605f4c6efa04864ec39d86dfb327b6.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 355, + 224, + 414, + 232 + ], + "lines": [ + { + "bbox": [ + 355, + 224, + 414, + 232 + ], + "spans": [ + { + "bbox": [ + 355, + 224, + 414, + 232 + ], + "type": "text", + "content": "Embedding Derivation" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 244, + 482, + 312 + ], + "lines": [ + { + "bbox": [ + 130, + 244, + 482, + 312 + ], + "spans": [ + { + "bbox": [ + 130, + 244, + 482, + 312 + ], + "type": "text", + "content": "Fig. 1: Overview of the proposed RECE. RECE consists of two main components: model editing and embedding derivation. First, erasing concepts by editing the model with a closed-form solution, and obtaining the edited cross-attention " + }, + { + "bbox": [ + 130, + 244, + 482, + 312 + ], + "type": "inline_equation", + "content": "W^{\\mathrm{new}}" + }, + { + "bbox": [ + 130, + 244, + 482, + 312 + ], + "type": "text", + "content": ". Then, new embedding " + }, + { + "bbox": [ + 130, + 244, + 482, + 312 + ], + "type": "inline_equation", + "content": "c_{i}^{\\prime}" + }, + { + "bbox": [ + 130, + 244, + 482, + 312 + ], + "type": "text", + "content": " can be derived by Eq. (7) given the original cross attention " + }, + { + "bbox": [ + 130, + 244, + 482, + 312 + ], + "type": "inline_equation", + "content": "W^{\\mathrm{old}}" + }, + { + "bbox": [ + 130, + 244, + 482, + 312 + ], + "type": "text", + "content": " and the edited " + }, + { + "bbox": [ + 130, + 244, + 482, + 312 + ], + "type": "inline_equation", + "content": "W^{\\mathrm{new}}" + }, + { + "bbox": [ + 130, + 244, + 482, + 312 + ], + "type": "text", + "content": ". In subsequent epochs, model editing and embedding derivation are looped." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 336, + 480, + 360 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 336, + 480, + 360 + ], + "spans": [ + { + "bbox": [ + 130, + 336, + 480, + 360 + ], + "type": "text", + "content": "changes, outputs for " + }, + { + "bbox": [ + 130, + 336, + 480, + 360 + ], + "type": "inline_equation", + "content": "c_{j} \\in P" + }, + { + "bbox": [ + 130, + 336, + 480, + 360 + ], + "type": "text", + "content": " are preserved as " + }, + { + "bbox": [ + 130, + 336, + 480, + 360 + ], + "type": "inline_equation", + "content": "W^{\\mathrm{old}}c_{j}" + }, + { + "bbox": [ + 130, + 336, + 480, + 360 + ], + "type": "text", + "content": " and a L2 regularization term is introduced:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 136, + 369, + 480, + 396 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 369, + 480, + 396 + ], + "spans": [ + { + "bbox": [ + 136, + 369, + 480, + 396 + ], + "type": "interline_equation", + "content": "\\min _ {W} \\sum_ {c _ {i} \\in E} | | W c _ {i} - W ^ {\\text {o l d}} c _ {i} ^ {*} | | _ {2} ^ {2} + \\lambda_ {1} \\sum_ {c _ {j} \\in P} | | W c _ {j} - W ^ {\\text {o l d}} c _ {j} | | _ {2} ^ {2} + \\lambda_ {2} | | W - W ^ {\\text {o l d}} | | _ {F} ^ {2}, \\tag {2}", + "image_path": "8e2ef690939f6fb894cd6c9deed25fa125288cef2866750f5904d2c55875d844.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 406, + 480, + 431 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 406, + 480, + 431 + ], + "spans": [ + { + "bbox": [ + 130, + 406, + 480, + 431 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 406, + 480, + 431 + ], + "type": "inline_equation", + "content": "\\lambda_{1}" + }, + { + "bbox": [ + 130, + 406, + 480, + 431 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 406, + 480, + 431 + ], + "type": "inline_equation", + "content": "\\lambda_{2}" + }, + { + "bbox": [ + 130, + 406, + 480, + 431 + ], + "type": "text", + "content": " are scaling factors preserving the existing concepts. UCE [7] prove that this formula has a closed-form solution:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 136, + 440, + 480, + 468 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 440, + 480, + 468 + ], + "spans": [ + { + "bbox": [ + 136, + 440, + 480, + 468 + ], + "type": "interline_equation", + "content": "W = W ^ {\\text {o l d}} \\left(\\sum_ {c _ {i} \\in E} c _ {i} ^ {*} c _ {i} ^ {T} + \\lambda_ {1} \\sum_ {c _ {j} \\in P} c _ {j} c _ {j} ^ {T} + \\lambda_ {2} I\\right) \\left(\\sum_ {c _ {i} \\in E} c _ {i} c _ {i} ^ {T} + \\lambda_ {1} \\sum_ {c _ {j} \\in P} c _ {j} c _ {j} ^ {T} + \\lambda_ {2} I\\right) ^ {- 1}. \\tag {3}", + "image_path": "1cb9465c125858943332f4ca30b72875fb30781481918faba881ac976e2180f6.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 478, + 480, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 478, + 480, + 514 + ], + "spans": [ + { + "bbox": [ + 130, + 478, + 480, + 514 + ], + "type": "text", + "content": "UCE directly assigns cross-attention KV matrices using the closed-form solution, eliminating the need for fine-tuning. This makes UCE significantly faster, hence we use UCE in our method." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 131, + 535, + 402, + 549 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 535, + 402, + 549 + ], + "spans": [ + { + "bbox": [ + 131, + 535, + 402, + 549 + ], + "type": "text", + "content": "3.2 Reliable and Efficient Concept Erasure (RECE)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 559, + 482, + 643 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 559, + 482, + 643 + ], + "spans": [ + { + "bbox": [ + 130, + 559, + 482, + 643 + ], + "type": "text", + "content": "While UCE [7] offers a fast solution for removing undesired concepts from T2I diffusion models, it can still produce undesired content, as illustrated in Tab. 1. This suggests an incomplete erasure of these concepts. To effectively eliminate such undesired concepts, we efficiently erase closed-form embeddings capable of regenerating erased concepts within the unlearned model. The derivation of embeddings and erasure is conducted iteratively to achieve a thorough erasure of inappropriate concepts, as shown in Fig. 1." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 146, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 146, + 91, + 448, + 102 + ], + "type": "text", + "content": "Reliable and Efficient Concept Erasure of Text-to-Image Diffusion Models" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 298, + 295 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 298, + 295 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 298, + 295 + ], + "type": "text", + "content": "Finding Target Contents Let us take \"nudity\" for example. As depicted in the second column of Fig. 2, when directly providing the input prompt \"nudity\" to UCE models, only landscape or unrelated images are generated. This is because the word \"nudity\" has been aligned with the empty text\". However, the erasure of UCE is incomplete. We can generate an adversarial prompt that enables UCE's model to produce images containing nudity content again, similar to those generated by SD when provided with the prompt \"nudity\"." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 296, + 299, + 344 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 296, + 299, + 344 + ], + "spans": [ + { + "bbox": [ + 130, + 296, + 299, + 344 + ], + "type": "text", + "content": "In this section, we introduce our method for deriving the new embedding in UCE's model, which guides UCE to generate nude images. As" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 344, + 482, + 440 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 344, + 482, + 440 + ], + "spans": [ + { + "bbox": [ + 130, + 344, + 482, + 440 + ], + "type": "text", + "content": "elaborated in Sec. 3.1, T2I introduces text embeddings into image generation through cross-attention layers, where the projection matrices " + }, + { + "bbox": [ + 130, + 344, + 482, + 440 + ], + "type": "inline_equation", + "content": "W_{k}" + }, + { + "bbox": [ + 130, + 344, + 482, + 440 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 344, + 482, + 440 + ], + "type": "inline_equation", + "content": "W_{v}" + }, + { + "bbox": [ + 130, + 344, + 482, + 440 + ], + "type": "text", + "content": " are used to transform text embeddings. Let " + }, + { + "bbox": [ + 130, + 344, + 482, + 440 + ], + "type": "inline_equation", + "content": "W^{\\mathrm{old}}" + }, + { + "bbox": [ + 130, + 344, + 482, + 440 + ], + "type": "text", + "content": " denote the projection matrices of the original U-Net before UCE's editing, " + }, + { + "bbox": [ + 130, + 344, + 482, + 440 + ], + "type": "inline_equation", + "content": "W^{\\mathrm{new}}" + }, + { + "bbox": [ + 130, + 344, + 482, + 440 + ], + "type": "text", + "content": " represent the projection matrices after UCE's editing, " + }, + { + "bbox": [ + 130, + 344, + 482, + 440 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 130, + 344, + 482, + 440 + ], + "type": "text", + "content": " denote the embedding of \"nudity\", and " + }, + { + "bbox": [ + 130, + 344, + 482, + 440 + ], + "type": "inline_equation", + "content": "c'" + }, + { + "bbox": [ + 130, + 344, + 482, + 440 + ], + "type": "text", + "content": " signify our derived embedding. If we can find a " + }, + { + "bbox": [ + 130, + 344, + 482, + 440 + ], + "type": "inline_equation", + "content": "c'" + }, + { + "bbox": [ + 130, + 344, + 482, + 440 + ], + "type": "text", + "content": " such that " + }, + { + "bbox": [ + 130, + 344, + 482, + 440 + ], + "type": "inline_equation", + "content": "W^{\\mathrm{new}}c'" + }, + { + "bbox": [ + 130, + 344, + 482, + 440 + ], + "type": "text", + "content": " closely resembles " + }, + { + "bbox": [ + 130, + 344, + 482, + 440 + ], + "type": "inline_equation", + "content": "W^{\\mathrm{old}}c" + }, + { + "bbox": [ + 130, + 344, + 482, + 440 + ], + "type": "text", + "content": ", then " + }, + { + "bbox": [ + 130, + 344, + 482, + 440 + ], + "type": "inline_equation", + "content": "c'" + }, + { + "bbox": [ + 130, + 344, + 482, + 440 + ], + "type": "text", + "content": " can guide the edited model to generate nude images like how " + }, + { + "bbox": [ + 130, + 344, + 482, + 440 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 130, + 344, + 482, + 440 + ], + "type": "text", + "content": " guides the original model. More precisely, the objective function is formulated as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 243, + 449, + 481, + 473 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 243, + 449, + 481, + 473 + ], + "spans": [ + { + "bbox": [ + 243, + 449, + 481, + 473 + ], + "type": "interline_equation", + "content": "\\min _ {c ^ {\\prime}} \\sum_ {i} \\| W _ {i} ^ {\\mathrm {n e w}} c ^ {\\prime} - W _ {i} ^ {\\mathrm {o l d}} c \\| _ {2} ^ {2}, \\tag {4}", + "image_path": "225aba6b7d2700b2bddf59b3ef4013ef4a2886f3d9e238afb93ed9e52bff8fac.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 483, + 482, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 483, + 482, + 544 + ], + "spans": [ + { + "bbox": [ + 130, + 483, + 482, + 544 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 483, + 482, + 544 + ], + "type": "inline_equation", + "content": "W_{i}" + }, + { + "bbox": [ + 130, + 483, + 482, + 544 + ], + "type": "text", + "content": " denotes K/V cross-attention matrices of U-Net. The solution " + }, + { + "bbox": [ + 130, + 483, + 482, + 544 + ], + "type": "inline_equation", + "content": "c'" + }, + { + "bbox": [ + 130, + 483, + 482, + 544 + ], + "type": "text", + "content": " derived from Eq. (4) can be viewed as the actual representation of " + }, + { + "bbox": [ + 130, + 483, + 482, + 544 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 130, + 483, + 482, + 544 + ], + "type": "text", + "content": " within the edited model. Evidently, Eq. (4) represents a convex function with respect to " + }, + { + "bbox": [ + 130, + 483, + 482, + 544 + ], + "type": "inline_equation", + "content": "c'" + }, + { + "bbox": [ + 130, + 483, + 482, + 544 + ], + "type": "text", + "content": ", which possesses a unique global minimum. As derived in Appendix A, Eq. (4) admits a closed-form solution:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 197, + 553, + 481, + 589 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 197, + 553, + 481, + 589 + ], + "spans": [ + { + "bbox": [ + 197, + 553, + 481, + 589 + ], + "type": "interline_equation", + "content": "c ^ {\\prime} = \\left(\\sum_ {i} W _ {i} ^ {\\mathrm {n e w} ^ {T}} W _ {i} ^ {\\mathrm {n e w}}\\right) ^ {- 1} \\left(\\sum_ {i} W _ {i} ^ {\\mathrm {n e w} ^ {T}} W _ {i} ^ {\\mathrm {o l d}}\\right) c. \\tag {5}", + "image_path": "8eae5ddc57e6d883a8e67660dffbe572e90bc8868ff3a7ac23405c4d4de1d37c.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": "Given that text conditioning works in the form of embedding in Stable Diffusion (SD), we can use our derived embedding as text conditioning. As illustrated in the third column of Fig. 2, this derived embedding effectively guides UCE's edited model to once again generate nude images, indicating its capacity to represent the concept of \"nudity\" within edited model. Thus it demonstrates that the erasure process of UCE remains incomplete. To address this issue, we further" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 305, + 121, + 482, + 232 + ], + "blocks": [ + { + "bbox": [ + 305, + 121, + 482, + 232 + ], + "lines": [ + { + "bbox": [ + 305, + 121, + 482, + 232 + ], + "spans": [ + { + "bbox": [ + 305, + 121, + 482, + 232 + ], + "type": "image", + "image_path": "baf14b026fad6d13d18d70d1506374992031d69b73d505a07c188881514cb37c.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 242, + 482, + 319 + ], + "lines": [ + { + "bbox": [ + 304, + 242, + 482, + 319 + ], + "spans": [ + { + "bbox": [ + 304, + 242, + 482, + 319 + ], + "type": "text", + "content": "Fig.2: When given the input prompt \"nudity\", SD generates images containing nudity content, while UCE generates unrelated images. When given our derived embedding and regularized embedding, UCE generates nude images again. We use for publication purposes." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 249, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 249, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 249, + 102 + ], + "type": "text", + "content": "F. Chao Gong et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 131, + 116, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 116, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 131, + 116, + 479, + 140 + ], + "type": "text", + "content": "remove our derived embeddings " + }, + { + "bbox": [ + 131, + 116, + 479, + 140 + ], + "type": "inline_equation", + "content": "c'" + }, + { + "bbox": [ + 131, + 116, + 479, + 140 + ], + "type": "text", + "content": " from UCE's model with Eq. (3) to prevent the generation of nude images." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 160, + 479, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 160, + 479, + 232 + ], + "spans": [ + { + "bbox": [ + 130, + 160, + 479, + 232 + ], + "type": "text", + "content": "Regularization Term We can further erase the concept of \"nudity\" by substituting " + }, + { + "bbox": [ + 130, + 160, + 479, + 232 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 130, + 160, + 479, + 232 + ], + "type": "text", + "content": " in Eq. (2) with our derived embedding " + }, + { + "bbox": [ + 130, + 160, + 479, + 232 + ], + "type": "inline_equation", + "content": "c'" + }, + { + "bbox": [ + 130, + 160, + 479, + 232 + ], + "type": "text", + "content": ". However, upon directly erasing " + }, + { + "bbox": [ + 130, + 160, + 479, + 232 + ], + "type": "inline_equation", + "content": "c'" + }, + { + "bbox": [ + 130, + 160, + 479, + 232 + ], + "type": "text", + "content": ", we observe a significant decline in the model's performance: it struggles to generate high-quality images for unrelated concepts as shown in Fig. 6. Hence, it becomes imperative to devise a method that retains the model's performance while erasing concepts." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 232, + 479, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 232, + 479, + 281 + ], + "spans": [ + { + "bbox": [ + 130, + 232, + 479, + 281 + ], + "type": "text", + "content": "Let " + }, + { + "bbox": [ + 130, + 232, + 479, + 281 + ], + "type": "inline_equation", + "content": "W^{\\mathrm{new1}}" + }, + { + "bbox": [ + 130, + 232, + 479, + 281 + ], + "type": "text", + "content": " denote the projection matrices after the last epoch's modification, and " + }, + { + "bbox": [ + 130, + 232, + 479, + 281 + ], + "type": "inline_equation", + "content": "W^{\\mathrm{new2}}" + }, + { + "bbox": [ + 130, + 232, + 479, + 281 + ], + "type": "text", + "content": " denote the projection matrices after the current epoch. Partly, preserving the model's performance entails minimizing the impact on unrelated concepts. Consequently, we define our objective function as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 249, + 291, + 481, + 308 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 249, + 291, + 481, + 308 + ], + "spans": [ + { + "bbox": [ + 249, + 291, + 481, + 308 + ], + "type": "interline_equation", + "content": "\\min _ {W} \\| W ^ {\\mathrm {n e w} 2} d - W ^ {\\mathrm {n e w} 1} d \\| _ {2} ^ {2}, \\tag {6}", + "image_path": "90d6464db0a519ea48cdb4fe60855b376427ab0d4ded061fe1f85b191230c04d.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 318, + 479, + 355 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 318, + 479, + 355 + ], + "spans": [ + { + "bbox": [ + 131, + 318, + 479, + 355 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 318, + 479, + 355 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 131, + 318, + 479, + 355 + ], + "type": "text", + "content": " represents an unrelated concept's embedding. Note that " + }, + { + "bbox": [ + 131, + 318, + 479, + 355 + ], + "type": "inline_equation", + "content": "W^{\\mathrm{new2}}" + }, + { + "bbox": [ + 131, + 318, + 479, + 355 + ], + "type": "text", + "content": " is directly influenced by " + }, + { + "bbox": [ + 131, + 318, + 479, + 355 + ], + "type": "inline_equation", + "content": "c'" + }, + { + "bbox": [ + 131, + 318, + 479, + 355 + ], + "type": "text", + "content": ", as it is derived after erasing " + }, + { + "bbox": [ + 131, + 318, + 479, + 355 + ], + "type": "inline_equation", + "content": "c'" + }, + { + "bbox": [ + 131, + 318, + 479, + 355 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 131, + 318, + 479, + 355 + ], + "type": "inline_equation", + "content": "W^{\\mathrm{new1}}" + }, + { + "bbox": [ + 131, + 318, + 479, + 355 + ], + "type": "text", + "content": ". We can obtain a theorem about our objective and its proof is provided in Appendix B:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 365, + 449, + 379 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 365, + 449, + 379 + ], + "spans": [ + { + "bbox": [ + 132, + 365, + 449, + 379 + ], + "type": "text", + "content": "Theorem 1. If " + }, + { + "bbox": [ + 132, + 365, + 449, + 379 + ], + "type": "inline_equation", + "content": "c'" + }, + { + "bbox": [ + 132, + 365, + 449, + 379 + ], + "type": "text", + "content": " is set to 0, Eq. (6) achieves its global minimum of 0." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 389, + 479, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 389, + 479, + 460 + ], + "spans": [ + { + "bbox": [ + 131, + 389, + 479, + 460 + ], + "type": "text", + "content": "Intuitively, if " + }, + { + "bbox": [ + 131, + 389, + 479, + 460 + ], + "type": "inline_equation", + "content": "c_{i}" + }, + { + "bbox": [ + 131, + 389, + 479, + 460 + ], + "type": "text", + "content": " in Eq. (3) is set to zero, the coefficient matrices on the right side will become an identity matrix when multiplied. As a result, " + }, + { + "bbox": [ + 131, + 389, + 479, + 460 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 131, + 389, + 479, + 460 + ], + "type": "text", + "content": " will revert to being equivalent to " + }, + { + "bbox": [ + 131, + 389, + 479, + 460 + ], + "type": "inline_equation", + "content": "W^{\\mathrm{old}}" + }, + { + "bbox": [ + 131, + 389, + 479, + 460 + ], + "type": "text", + "content": ", exerting minimal influence on unrelated concepts due to the absence of further modifications to " + }, + { + "bbox": [ + 131, + 389, + 479, + 460 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 131, + 389, + 479, + 460 + ], + "type": "text", + "content": ". Therefore, we need to introduce a regularization term to the original objective Eq. (4) to ensure that the obtained " + }, + { + "bbox": [ + 131, + 389, + 479, + 460 + ], + "type": "inline_equation", + "content": "c'" + }, + { + "bbox": [ + 131, + 389, + 479, + 460 + ], + "type": "text", + "content": " is close to zero, thereby minimizing its influence on the model's performance:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 225, + 472, + 480, + 495 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 225, + 472, + 480, + 495 + ], + "spans": [ + { + "bbox": [ + 225, + 472, + 480, + 495 + ], + "type": "interline_equation", + "content": "\\min _ {c ^ {\\prime}} \\sum_ {i} \\| W _ {i} ^ {\\text {n e w}} c ^ {\\prime} - W _ {i} ^ {\\text {o l d}} c \\| _ {2} ^ {2} + \\lambda \\| c ^ {\\prime} \\| _ {2} ^ {2}. \\tag {7}", + "image_path": "4c7afef5f4d10febb1b6ea48433cf7671afdfacc48d6e4afd9da4163699c0834.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 504, + 479, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 504, + 479, + 529 + ], + "spans": [ + { + "bbox": [ + 131, + 504, + 479, + 529 + ], + "type": "text", + "content": "As derived in Appendix A, this final objective function also possesses a unique global minimum solution:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 189, + 538, + 480, + 572 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 189, + 538, + 480, + 572 + ], + "spans": [ + { + "bbox": [ + 189, + 538, + 480, + 572 + ], + "type": "interline_equation", + "content": "c ^ {\\prime} = \\left(\\lambda I + \\sum_ {i} W _ {i} ^ {\\mathrm {n e w} ^ {T}} W _ {i} ^ {\\mathrm {n e w}}\\right) ^ {- 1} \\left(\\sum_ {i} W _ {i} ^ {\\mathrm {n e w} ^ {T}} W _ {i} ^ {\\mathrm {o l d}}\\right) c \\tag {8}", + "image_path": "8caf0fe336a88da4ff0584d216aa542524708126a8b42c7720f8f89aef0e42a1.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 582, + 479, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 479, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 479, + 665 + ], + "type": "text", + "content": "As illustrated in the fourth column of Fig. 2, this regularized embedding can also guide UCE's model to once again generate nude images, indicating its ability to represent the concept \"nudity\" within UCE's model. With the incorporation of our regularization term, we iteratively apply the erasure process to the refined embedding " + }, + { + "bbox": [ + 130, + 582, + 479, + 665 + ], + "type": "inline_equation", + "content": "c'" + }, + { + "bbox": [ + 130, + 582, + 479, + 665 + ], + "type": "text", + "content": " using Eq. (3) over multiple epochs. This ensures thorough concept erasure while safeguarding the overall performance of the model. The algorithm details are elaborated in Algorithm 1." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 146, + 91, + 448, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 91, + 448, + 103 + ], + "spans": [ + { + "bbox": [ + 146, + 91, + 448, + 103 + ], + "type": "text", + "content": "Reliable and Efficient Concept Erasure of Text-to-Image Diffusion Models" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "code", + "bbox": [ + 134, + 132, + 482, + 352 + ], + "blocks": [ + { + "bbox": [ + 140, + 118, + 332, + 130 + ], + "lines": [ + { + "bbox": [ + 140, + 118, + 332, + 130 + ], + "spans": [ + { + "bbox": [ + 140, + 118, + 332, + 130 + ], + "type": "text", + "content": "Algorithm 1: Erase Concepts with RECE" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "code_caption" + }, + { + "bbox": [ + 134, + 132, + 482, + 352 + ], + "lines": [ + { + "bbox": [ + 134, + 132, + 482, + 352 + ], + "spans": [ + { + "bbox": [ + 134, + 132, + 482, + 352 + ], + "type": "text", + "content": "Input: Diffusion U-Net " + }, + { + "bbox": [ + 134, + 132, + 482, + 352 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 134, + 132, + 482, + 352 + ], + "type": "text", + "content": " , concepts set " + }, + { + "bbox": [ + 134, + 132, + 482, + 352 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 134, + 132, + 482, + 352 + ], + "type": "text", + "content": " to erase, " + }, + { + "bbox": [ + 134, + 132, + 482, + 352 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 134, + 132, + 482, + 352 + ], + "type": "text", + "content": " to preserve, epochs " + }, + { + "bbox": [ + 134, + 132, + 482, + 352 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 134, + 132, + 482, + 352 + ], + "type": "text", + "content": " Output: Diffusion U-Net " + }, + { + "bbox": [ + 134, + 132, + 482, + 352 + ], + "type": "inline_equation", + "content": "\\theta^{\\prime}" + }, + { + "bbox": [ + 134, + 132, + 482, + 352 + ], + "type": "text", + "content": " with concepts " + }, + { + "bbox": [ + 134, + 132, + 482, + 352 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 134, + 132, + 482, + 352 + ], + "type": "text", + "content": " erased. \n1 /\\* Initialize\\*/ \n2 " + }, + { + "bbox": [ + 134, + 132, + 482, + 352 + ], + "type": "inline_equation", + "content": "\\theta^{\\prime}\\gets \\theta" + }, + { + "bbox": [ + 134, + 132, + 482, + 352 + ], + "type": "text", + "content": " . \n3 Initialize text embeddings " + }, + { + "bbox": [ + 134, + 132, + 482, + 352 + ], + "type": "inline_equation", + "content": "c_{i}" + }, + { + "bbox": [ + 134, + 132, + 482, + 352 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 134, + 132, + 482, + 352 + ], + "type": "inline_equation", + "content": "c_{j}" + }, + { + "bbox": [ + 134, + 132, + 482, + 352 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 134, + 132, + 482, + 352 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 134, + 132, + 482, + 352 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 134, + 132, + 482, + 352 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 134, + 132, + 482, + 352 + ], + "type": "text", + "content": " . \n4 Extract K&V matrices " + }, + { + "bbox": [ + 134, + 132, + 482, + 352 + ], + "type": "inline_equation", + "content": "W^{\\mathrm{old}}" + }, + { + "bbox": [ + 134, + 132, + 482, + 352 + ], + "type": "text", + "content": " from the cross attention of " + }, + { + "bbox": [ + 134, + 132, + 482, + 352 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 134, + 132, + 482, + 352 + ], + "type": "text", + "content": " . \n5 /\\* Preliminary erasing with UCE\\*/ \n6 Obtain " + }, + { + "bbox": [ + 134, + 132, + 482, + 352 + ], + "type": "inline_equation", + "content": "W^{\\mathrm{new}}" + }, + { + "bbox": [ + 134, + 132, + 482, + 352 + ], + "type": "text", + "content": " by erasing concepts in " + }, + { + "bbox": [ + 134, + 132, + 482, + 352 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 134, + 132, + 482, + 352 + ], + "type": "text", + "content": " with Eq. (3); \n7 for " + }, + { + "bbox": [ + 134, + 132, + 482, + 352 + ], + "type": "inline_equation", + "content": "t = 1,\\ldots ,T" + }, + { + "bbox": [ + 134, + 132, + 482, + 352 + ], + "type": "text", + "content": " do \n8 /\\* Derive new embeddings\\*/ \n9 " + }, + { + "bbox": [ + 134, + 132, + 482, + 352 + ], + "type": "inline_equation", + "content": "E^{\\prime}\\gets \\{\\}" + }, + { + "bbox": [ + 134, + 132, + 482, + 352 + ], + "type": "text", + "content": " . \nfor " + }, + { + "bbox": [ + 134, + 132, + 482, + 352 + ], + "type": "inline_equation", + "content": "c_{i}\\in E" + }, + { + "bbox": [ + 134, + 132, + 482, + 352 + ], + "type": "text", + "content": " do \n11 Derive new embedding " + }, + { + "bbox": [ + 134, + 132, + 482, + 352 + ], + "type": "inline_equation", + "content": "c_i^\\prime" + }, + { + "bbox": [ + 134, + 132, + 482, + 352 + ], + "type": "text", + "content": " using " + }, + { + "bbox": [ + 134, + 132, + 482, + 352 + ], + "type": "inline_equation", + "content": "W^{\\mathrm{new}}" + }, + { + "bbox": [ + 134, + 132, + 482, + 352 + ], + "type": "text", + "content": " with Eq. (8); \n12 " + }, + { + "bbox": [ + 134, + 132, + 482, + 352 + ], + "type": "inline_equation", + "content": "\\begin{array}{r}E^{\\prime}\\leftarrow E^{\\prime}\\cup \\{c_{i}^{\\prime}\\} \\end{array}" + }, + { + "bbox": [ + 134, + 132, + 482, + 352 + ], + "type": "text", + "content": " \n13 /\\* Erasing derived embeddings\\*/ \n14 Obtain " + }, + { + "bbox": [ + 134, + 132, + 482, + 352 + ], + "type": "inline_equation", + "content": "W^{\\mathrm{new}}" + }, + { + "bbox": [ + 134, + 132, + 482, + 352 + ], + "type": "text", + "content": " by erasing " + }, + { + "bbox": [ + 134, + 132, + 482, + 352 + ], + "type": "inline_equation", + "content": "E^{\\prime}" + }, + { + "bbox": [ + 134, + 132, + 482, + 352 + ], + "type": "text", + "content": " with Eq. (3); \n15 " + }, + { + "bbox": [ + 134, + 132, + 482, + 352 + ], + "type": "inline_equation", + "content": "W^{\\mathrm{new}}\\gets W^{\\mathrm{new}};" + }, + { + "bbox": [ + 134, + 132, + 482, + 352 + ], + "type": "text", + "content": " \n16 Update " + }, + { + "bbox": [ + 134, + 132, + 482, + 352 + ], + "type": "inline_equation", + "content": "\\theta^{\\prime}" + }, + { + "bbox": [ + 134, + 132, + 482, + 352 + ], + "type": "text", + "content": " by replacing K&V matrices with " + }, + { + "bbox": [ + 134, + 132, + 482, + 352 + ], + "type": "inline_equation", + "content": "W^{\\mathrm{new}}" + }, + { + "bbox": [ + 134, + 132, + 482, + 352 + ], + "type": "text", + "content": " . \n17 return " + }, + { + "bbox": [ + 134, + 132, + 482, + 352 + ], + "type": "inline_equation", + "content": "\\theta^{\\prime}" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "code_body" + } + ], + "index": 3, + "sub_type": "algorithm" + }, + { + "bbox": [ + 132, + 379, + 230, + 392 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 379, + 230, + 392 + ], + "spans": [ + { + "bbox": [ + 132, + 379, + 230, + 392 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 404, + 482, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 404, + 482, + 548 + ], + "spans": [ + { + "bbox": [ + 130, + 404, + 482, + 548 + ], + "type": "text", + "content": "In this section, we present the results of our method for erasing inappropriate concepts and artistic styles. We also include the results of object removal in Appendix. We start with SD V1.4 as our base model. Following the implementation in [16], we set " + }, + { + "bbox": [ + 130, + 404, + 482, + 548 + ], + "type": "inline_equation", + "content": "\\lambda_{1}" + }, + { + "bbox": [ + 130, + 404, + 482, + 548 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 404, + 482, + 548 + ], + "type": "inline_equation", + "content": "\\lambda_{2}" + }, + { + "bbox": [ + 130, + 404, + 482, + 548 + ], + "type": "text", + "content": " in Eq. (2) to 0.1. For inappropriate concepts, we perform iterative erasure for 5 epochs and set " + }, + { + "bbox": [ + 130, + 404, + 482, + 548 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 130, + 404, + 482, + 548 + ], + "type": "text", + "content": " in Eq. (7) to " + }, + { + "bbox": [ + 130, + 404, + 482, + 548 + ], + "type": "inline_equation", + "content": "1e - 1" + }, + { + "bbox": [ + 130, + 404, + 482, + 548 + ], + "type": "text", + "content": ". For artistic style, we conduct erasure for 10 epochs and set " + }, + { + "bbox": [ + 130, + 404, + 482, + 548 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 130, + 404, + 482, + 548 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 130, + 404, + 482, + 548 + ], + "type": "inline_equation", + "content": "1e - 3" + }, + { + "bbox": [ + 130, + 404, + 482, + 548 + ], + "type": "text", + "content": ". The baselines we will compare with are: SD V1.4 [25], SD V2.1 [31](Stable Diffusion pretrained on an NSFW filtered dataset), SLD [29], ESD [6], CA [13], SA [8], UCE [7]. As for SLD, ESD, SA and UCE, we adhere to the recommended configuration in their papers [6-8,29]. For CA [13], we fine-tune the full weights of U-Net to erase unsafe contents and the cross-attention module to erase artistic styles, according to its documentation3." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 566, + 287, + 577 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 566, + 287, + 577 + ], + "spans": [ + { + "bbox": [ + 132, + 566, + 287, + 577 + ], + "type": "text", + "content": "4.1 Unsafe Content Removal" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 586, + 482, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 586, + 482, + 647 + ], + "spans": [ + { + "bbox": [ + 130, + 586, + 482, + 647 + ], + "type": "text", + "content": "Experimental Setup In this section, we assess the effectiveness of erasing unsafe concepts. We conduct experiments on the Inappropriate Image Prompts (I2P) dataset [29]. The I2P dataset includes various inappropriate prompts, such as violence, self-harm, sexual content, and shocking content. These prompts are collected from real-world, user-generated images based on the official SD. Our" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 249, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 249, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 249, + 102 + ], + "type": "text", + "content": "F. Chao Gong et al." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 133, + 653, + 358, + 666 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 653, + 358, + 666 + ], + "spans": [ + { + "bbox": [ + 133, + 653, + 358, + 666 + ], + "type": "text", + "content": "3 https://github.com/nupurkmr9/concept-ablation" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 134, + 114, + 479, + 213 + ], + "blocks": [ + { + "bbox": [ + 134, + 114, + 479, + 213 + ], + "lines": [ + { + "bbox": [ + 134, + 114, + 479, + 213 + ], + "spans": [ + { + "bbox": [ + 134, + 114, + 479, + 213 + ], + "type": "table", + "html": "
MethodNudity DetectionCOCO-30k
Breast(F)Genitalia(F)Breast(M)Genitalia(M)ButtocksFeetBellyArmpitsTotal↓CLIP↑FID↓
SD v1.4183214610444217112964631.33-
SD v2.1121134031439146109485--
ESD-u14185524313312130.453.73
UCE3161981120553618631.261.82
SLD-Med72534318191049935430.952.60
SA399401532491516330.5717.34
CA6191041428239531.167.87
Ours80640823176630.952.82
", + "image_path": "46d37a4fa79add357abb9f51c32db6b6803b6d5aea56801b2c80a1439f0ad4d5.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 213, + 480, + 257 + ], + "lines": [ + { + "bbox": [ + 132, + 213, + 480, + 257 + ], + "spans": [ + { + "bbox": [ + 132, + 213, + 480, + 257 + ], + "type": "text", + "content": "Table 1: Comparison of performance metrics for content removal methods. (Left) Number of nude body parts detected by Nudenet on I2P dataset with threshold 0.6. (Right) CLIP-score and FID against original SD. F: Female. M: Male. Bold: best. Underline: second-best." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 130, + 293, + 479, + 352 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 293, + 479, + 352 + ], + "spans": [ + { + "bbox": [ + 130, + 293, + 479, + 352 + ], + "type": "text", + "content": "evaluation focuses on the erasure of nudity since it is a classical unsafe concept. For each model, we generate one image per prompt in the I2P dataset, resulting in a total of 4703 images. Nude body parts are detected using the Nudenet detector [18], with the threshold set to 0.6. This threshold follows the default settings in " + }, + { + "bbox": [ + 130, + 293, + 479, + 352 + ], + "type": "inline_equation", + "content": "\\mathrm{I2P^4}" + }, + { + "bbox": [ + 130, + 293, + 479, + 352 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 354, + 480, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 354, + 480, + 415 + ], + "spans": [ + { + "bbox": [ + 130, + 354, + 480, + 415 + ], + "type": "text", + "content": "To verify that the unlearned models can still generate normal images, we use COCO-30k [14] with its captions as prompts. COCO-30k is a dataset devoid of unsafe concepts, making it suitable for evaluating edited models' generation capabilities. We evaluate the models' image-text consistency based on CLIP-score [9], and visual similarity against SD-generated images based on FID [17]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 437, + 480, + 519 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 437, + 480, + 519 + ], + "spans": [ + { + "bbox": [ + 130, + 437, + 480, + 519 + ], + "type": "text", + "content": "Removal Results As depicted in Tab. 1, our method yields the lowest number of nude body parts, while demonstrating impressive specificity in preserving normal content of COCO-30k. CA generates the second-fewest nude body parts but it exhibits poorer performance in terms of FID, indicating a poorer trade-off between generation ability and removal effectiveness. On the other hand, CA fine-tunes full weights and our RECE only fine-tunes cross-attention modules, which will be discussed in detail in Sec. 4.4." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 521, + 480, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 480, + 605 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 480, + 605 + ], + "type": "text", + "content": "Notably, our method achieves a FID score closely comparable to the top-performing UCE and the second-best SLD, both of which generates a considerably higher number of nude body parts. This suggests that our method minimally impacts the generation of normal content while striving for better removal effectiveness. Additionally, most methods exhibits favorable CLIP-score results thus we consider the CLIP-score performance acceptable as long as it remains within a reasonable range." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 606, + 485, + 643 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 606, + 485, + 643 + ], + "spans": [ + { + "bbox": [ + 130, + 606, + 485, + 643 + ], + "type": "text", + "content": "In open-sourced conditions, the inference guidance mechanism such as SLD [29] can be easily bypassed by deleting the corresponding code [24]. Large-scale model retraining on NSFW-filtered dataset demands considerable computational re" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 147, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 147, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 147, + 91, + 447, + 102 + ], + "type": "text", + "content": "Reliable and Efficient Concept Erasure of Text-to-Image Diffusion Models" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 133, + 653, + 306, + 666 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 653, + 306, + 666 + ], + "spans": [ + { + "bbox": [ + 133, + 653, + 306, + 666 + ], + "type": "text", + "content": "4 https://github.com/ml-research/i2p" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 136, + 114, + 177, + 194 + ], + "blocks": [ + { + "bbox": [ + 136, + 114, + 177, + 194 + ], + "lines": [ + { + "bbox": [ + 136, + 114, + 177, + 194 + ], + "spans": [ + { + "bbox": [ + 136, + 114, + 177, + 194 + ], + "type": "image", + "image_path": "a8f83f03d442eab7b7a57107a1913358fbae6b46c5ab152a0b9cd23ee6385333.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 142, + 196, + 171, + 204 + ], + "lines": [ + { + "bbox": [ + 142, + 196, + 171, + 204 + ], + "spans": [ + { + "bbox": [ + 142, + 196, + 171, + 204 + ], + "type": "text", + "content": "SD v1.4" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 179, + 114, + 220, + 194 + ], + "blocks": [ + { + "bbox": [ + 179, + 114, + 220, + 194 + ], + "lines": [ + { + "bbox": [ + 179, + 114, + 220, + 194 + ], + "spans": [ + { + "bbox": [ + 179, + 114, + 220, + 194 + ], + "type": "image", + "image_path": "6275885888c8026fe5df7b6b96838c6ae6503bb8bfd86262ccbef83e1bd21a3d.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 186, + 196, + 214, + 204 + ], + "lines": [ + { + "bbox": [ + 186, + 196, + 214, + 204 + ], + "spans": [ + { + "bbox": [ + 186, + 196, + 214, + 204 + ], + "type": "text", + "content": "SD v2.1" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 222, + 114, + 263, + 194 + ], + "blocks": [ + { + "bbox": [ + 222, + 114, + 263, + 194 + ], + "lines": [ + { + "bbox": [ + 222, + 114, + 263, + 194 + ], + "spans": [ + { + "bbox": [ + 222, + 114, + 263, + 194 + ], + "type": "image", + "image_path": "aab5cba63815acac99ab09c739109b4d533a5cc480081220d4817399f1e3a8d0.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 234, + 196, + 251, + 205 + ], + "lines": [ + { + "bbox": [ + 234, + 196, + 251, + 205 + ], + "spans": [ + { + "bbox": [ + 234, + 196, + 251, + 205 + ], + "type": "text", + "content": "ESD" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 265, + 114, + 306, + 194 + ], + "blocks": [ + { + "bbox": [ + 265, + 114, + 306, + 194 + ], + "lines": [ + { + "bbox": [ + 265, + 114, + 306, + 194 + ], + "spans": [ + { + "bbox": [ + 265, + 114, + 306, + 194 + ], + "type": "image", + "image_path": "7a07d949f919d07dedb8c0220f1b59c63a6bd00d8dd0ec66c1d2fcfe32796fe1.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 277, + 196, + 294, + 205 + ], + "lines": [ + { + "bbox": [ + 277, + 196, + 294, + 205 + ], + "spans": [ + { + "bbox": [ + 277, + 196, + 294, + 205 + ], + "type": "text", + "content": "UCE" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 214, + 481, + 258 + ], + "lines": [ + { + "bbox": [ + 130, + 214, + 481, + 258 + ], + "spans": [ + { + "bbox": [ + 130, + 214, + 481, + 258 + ], + "type": "text", + "content": "Fig. 3: Qualitative results of different nudity removing methods. Images for each row are generated using the same prompt. All prompts originate from I2P benchmark, and we don't intentionally generate inappropriate content of any individual." + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 307, + 114, + 349, + 194 + ], + "blocks": [ + { + "bbox": [ + 307, + 114, + 349, + 194 + ], + "lines": [ + { + "bbox": [ + 307, + 114, + 349, + 194 + ], + "spans": [ + { + "bbox": [ + 307, + 114, + 349, + 194 + ], + "type": "image", + "image_path": "a6f700265c773b2839d4c4b7851947f130832029ca90abf94f07f7bd312802fd.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 311, + 196, + 345, + 205 + ], + "lines": [ + { + "bbox": [ + 311, + 196, + 345, + 205 + ], + "spans": [ + { + "bbox": [ + 311, + 196, + 345, + 205 + ], + "type": "text", + "content": "SLD-Med" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 351, + 114, + 392, + 194 + ], + "blocks": [ + { + "bbox": [ + 351, + 114, + 392, + 194 + ], + "lines": [ + { + "bbox": [ + 351, + 114, + 392, + 194 + ], + "spans": [ + { + "bbox": [ + 351, + 114, + 392, + 194 + ], + "type": "image", + "image_path": "c6e74b0bff7785af9f1dcd9330dfc281a6b443797041c536576602838384fe65.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 364, + 196, + 376, + 205 + ], + "lines": [ + { + "bbox": [ + 364, + 196, + 376, + 205 + ], + "spans": [ + { + "bbox": [ + 364, + 196, + 376, + 205 + ], + "type": "text", + "content": "SA" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 393, + 114, + 435, + 194 + ], + "blocks": [ + { + "bbox": [ + 393, + 114, + 435, + 194 + ], + "lines": [ + { + "bbox": [ + 393, + 114, + 435, + 194 + ], + "spans": [ + { + "bbox": [ + 393, + 114, + 435, + 194 + ], + "type": "image", + "image_path": "beac8009cce77f2f1f1d6ebf31b54ab49ffa7ce51fd6d950f27bcfa24577e05a.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 408, + 196, + 421, + 204 + ], + "lines": [ + { + "bbox": [ + 408, + 196, + 421, + 204 + ], + "spans": [ + { + "bbox": [ + 408, + 196, + 421, + 204 + ], + "type": "text", + "content": "CA" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 436, + 114, + 477, + 194 + ], + "blocks": [ + { + "bbox": [ + 436, + 114, + 477, + 194 + ], + "lines": [ + { + "bbox": [ + 436, + 114, + 477, + 194 + ], + "spans": [ + { + "bbox": [ + 436, + 114, + 477, + 194 + ], + "type": "image", + "image_path": "9a531224f258d0878a7814fa61bf3b215c0927b55a26d99e254ff4c5820a0fba.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 447, + 196, + 466, + 204 + ], + "lines": [ + { + "bbox": [ + 447, + 196, + 466, + 204 + ], + "spans": [ + { + "bbox": [ + 447, + 196, + 466, + 204 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "bbox": [ + 130, + 283, + 479, + 306 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 283, + 479, + 306 + ], + "spans": [ + { + "bbox": [ + 130, + 283, + 479, + 306 + ], + "type": "text", + "content": "sources [22], and even then, the model SD v2.1 may still generate nude images, as illustrated in Tab. 1." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 130, + 307, + 480, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 307, + 480, + 380 + ], + "spans": [ + { + "bbox": [ + 130, + 307, + 480, + 380 + ], + "type": "text", + "content": "Qualitative results are illustrated in Fig. 3. In the first row, all erasing methods successfully generate non-nude images, but the results of CA and SA differ significantly from the original SD's. In the second row, all methods except for SD v2.1 avoid generating nude images. Among these, only UCE, SLD, and our method effectively capture the facial features. These findings demonstrate that our method effectively maintains unrelated concepts." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 130, + 398, + 482, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 398, + 482, + 578 + ], + "spans": [ + { + "bbox": [ + 130, + 398, + 482, + 578 + ], + "type": "text", + "content": "Nudity Bias While our RECE demonstrates remarkable effectiveness in minimizing the generation of nude content, it exhibits limitations in erasing male nudity and similar limitations are also observed in other methods. We count the number of female and male nude body parts in the 4703 I2P images. As illustrated in Tab. 2, the nudity ratios between women and men almost decrease in every method, indicating an erasure bias on female information (excluding SA, which erases more sex-related concepts besides nudity [8]). We attribute this limitation to the inherent bias in the target concept \"nudity\" within SD, which tends to generate more female-oriented content. To investigate this bias, we randomly selected 20 seeds and employed SD V1.4 to generate 3 images per seed with the prompt \"nudity\", resulting in a total of 60 images. Surprisingly, almost all of these generated images depict female body part as presented in the last column of Tab. 2. Thus our derived embedding is also biased. Further improvements require an awareness of the biases inherent in the model while performing erasure." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 132, + 596, + 276, + 608 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 596, + 276, + 608 + ], + "spans": [ + { + "bbox": [ + 132, + 596, + 276, + 608 + ], + "type": "text", + "content": "4.2 Artistic Style Removal" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "type": "text", + "content": "Experimental Setup We conduct an evaluation to assess the efficacy of removing artistic styles to address copyright concerns. Following the datasets in ESD [6], we use 20 prompts for each of 5 famous artists—Van Gogh, Pablo Picasso, Rembrandt, Andy Warhol and Caravaggio—and 5 modern artists—Kelly" + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 249, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 249, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 249, + 102 + ], + "type": "text", + "content": "F. Chao Gong et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 161, + 114, + 452, + 170 + ], + "blocks": [ + { + "bbox": [ + 161, + 114, + 452, + 170 + ], + "lines": [ + { + "bbox": [ + 161, + 114, + 452, + 170 + ], + "spans": [ + { + "bbox": [ + 161, + 114, + 452, + 170 + ], + "type": "table", + "html": "
MetricsSDESDUCESLD-MedSACAOursSD- "nudity"
# Female Nudity204153777487889
# Male Nudity56132737419101
Female/Male3.641.151.372.08120.370.889
", + "image_path": "88ab8b50b76590323b03c6ff579cb88b872e2c2c19cb78d3b44f64e7d22ef99a.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 161, + 233, + 454, + 319 + ], + "blocks": [ + { + "bbox": [ + 130, + 170, + 481, + 213 + ], + "lines": [ + { + "bbox": [ + 130, + 170, + 481, + 213 + ], + "spans": [ + { + "bbox": [ + 130, + 170, + 481, + 213 + ], + "type": "text", + "content": "Table 2: In the first 7 columns, we counted the numbers of female nudity and male nudity body parts for I2P dataset. The last column is the result of SD v1.4 conditioned by the prompt \"nudity\". This table highlights the inherent bias within SD. And all each erasure methods can not remove the concept of male nudity very well." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 161, + 233, + 454, + 319 + ], + "lines": [ + { + "bbox": [ + 161, + 233, + 454, + 319 + ], + "spans": [ + { + "bbox": [ + 161, + 233, + 454, + 319 + ], + "type": "table", + "html": "
Removal MethodErase "Van Gogh"Erase "Kelly McKernan"
LPIPSe↑LPIPSu↓LPIPSd↑LPIPSe↑LPIPSu↓LPIPSd↑
ESD-x0.400.260.140.370.210.16
UCE0.250.050.200.250.030.22
SLD-Med0.210.100.110.220.180.04
CA0.300.130.170.220.170.05
Ours0.310.080.230.290.040.25
", + "image_path": "46fc7bfb5a0d212e38360d642b63b68e5cf89e9e4857d863ccd8078e3029aa3f.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 319, + 480, + 341 + ], + "lines": [ + { + "bbox": [ + 130, + 319, + 480, + 341 + ], + "spans": [ + { + "bbox": [ + 130, + 319, + 480, + 341 + ], + "type": "text", + "content": "Table 3: Comparison of LPIPS scores for artistic removal methods. Bold: best. Underline: second-best. " + }, + { + "bbox": [ + 130, + 319, + 480, + 341 + ], + "type": "inline_equation", + "content": "\\mathrm{LPIPS_d}" + }, + { + "bbox": [ + 130, + 319, + 480, + 341 + ], + "type": "text", + "content": " indicates overall erasure performance." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 130, + 373, + 480, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 373, + 480, + 422 + ], + "spans": [ + { + "bbox": [ + 130, + 373, + 480, + 422 + ], + "type": "text", + "content": "McKernan, Thomas Kinkade, Tyler Edlin, Kilian Eng and the series \"Ajin: Demi-Human\", which have been reported to be imitated by SD [30]. To evaluate our RECE and all the aforementioned baselines, we erase the style of two artists: Van Gogh and Kelly McKernan." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 438, + 482, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 438, + 482, + 582 + ], + "spans": [ + { + "bbox": [ + 130, + 438, + 482, + 582 + ], + "type": "text", + "content": "Removal Results We conducted an evaluation based on LPIPS scores [37] compared to the original SD, as detailed in Tab. 3. LPIPS evaluates the perceptual distance between image patches, where higher values indicate greater differences and lower values indicate more similarity. The " + }, + { + "bbox": [ + 130, + 438, + 482, + 582 + ], + "type": "inline_equation", + "content": "\\mathrm{LPIPS_e}" + }, + { + "bbox": [ + 130, + 438, + 482, + 582 + ], + "type": "text", + "content": " is calculated on the erased artist. A higher " + }, + { + "bbox": [ + 130, + 438, + 482, + 582 + ], + "type": "inline_equation", + "content": "\\mathrm{LPIPS_e}" + }, + { + "bbox": [ + 130, + 438, + 482, + 582 + ], + "type": "text", + "content": " value suggests a more effective style removal, and both ESD and our method demonstrate successful erasure of the target style. " + }, + { + "bbox": [ + 130, + 438, + 482, + 582 + ], + "type": "inline_equation", + "content": "\\mathrm{LPIPS_u}" + }, + { + "bbox": [ + 130, + 438, + 482, + 582 + ], + "type": "text", + "content": " is calculated on unerased artists. A lower " + }, + { + "bbox": [ + 130, + 438, + 482, + 582 + ], + "type": "inline_equation", + "content": "\\mathrm{LPIPS_u}" + }, + { + "bbox": [ + 130, + 438, + 482, + 582 + ], + "type": "text", + "content": " indicates a lesser impact on unrelated artists, where our method and UCE effectively maintains unrelated concepts. We also calculate the overall effectiveness by " + }, + { + "bbox": [ + 130, + 438, + 482, + 582 + ], + "type": "inline_equation", + "content": "\\mathrm{LPIPS_d} = \\mathrm{LPIPS_e} - \\mathrm{LPIPS_u}" + }, + { + "bbox": [ + 130, + 438, + 482, + 582 + ], + "type": "text", + "content": ", which is the difference between erased and unerased artists. Our method performs best in this regard. Qualitative results can be found in Appendix." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 599, + 358, + 611 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 599, + 358, + 611 + ], + "spans": [ + { + "bbox": [ + 131, + 599, + 358, + 611 + ], + "type": "text", + "content": "4.3 Robustness Against Red-teaming Tools" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "type": "text", + "content": "Experimental Setup To demonstrate the robustness of our RECE in safeguarding against various attack methods, we employ different red-teaming tools, including white-box methods such as P4D [3] and UnlearnDiff [38], and the black-box method Ring-A-Bell [32]. For nudity, as provided by UnlearnDiff [38]," + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 146, + 90, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 90, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 146, + 90, + 447, + 102 + ], + "type": "text", + "content": "Reliable and Efficient Concept Erasure of Text-to-Image Diffusion Models" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 161, + 114, + 452, + 183 + ], + "blocks": [ + { + "bbox": [ + 161, + 114, + 452, + 183 + ], + "lines": [ + { + "bbox": [ + 161, + 114, + 452, + 183 + ], + "spans": [ + { + "bbox": [ + 161, + 114, + 452, + 183 + ], + "type": "table", + "html": "
Red-teamingSD v1.4SD v2.1ESD-uUCESLD-MaxSACAOurs
UnlearnDiff--66.2079.5882.3977.4665.4965.49
P4D--63.3880.2877.4678.8760.5664.79
Ring-A-Bell83.1072.5469.7233.1066.2022.5425.3513.38
Average--66.4364.3275.3559.6250.4747.89
", + "image_path": "99c545c1dc29393fd647580017cc266a66ad743546f11436cdfcce1d4821b74f.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 184, + 479, + 205 + ], + "lines": [ + { + "bbox": [ + 132, + 184, + 479, + 205 + ], + "spans": [ + { + "bbox": [ + 132, + 184, + 479, + 205 + ], + "type": "text", + "content": "Table 4: Robustness of different methods against red-teaming tools, measured by attack success rate(%) . Bold: best. Underline: second-best." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 173, + 225, + 218, + 269 + ], + "blocks": [ + { + "bbox": [ + 173, + 225, + 218, + 269 + ], + "lines": [ + { + "bbox": [ + 173, + 225, + 218, + 269 + ], + "spans": [ + { + "bbox": [ + 173, + 225, + 218, + 269 + ], + "type": "image", + "image_path": "b4f3e977ed9cae46da51871eed38732622a37cb76ef6a80a31a06e14db7f2c19.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 174, + 269, + 217, + 277 + ], + "lines": [ + { + "bbox": [ + 174, + 269, + 217, + 277 + ], + "spans": [ + { + "bbox": [ + 174, + 269, + 217, + 277 + ], + "type": "text", + "content": "Original SD" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 131, + 288, + 480, + 320 + ], + "lines": [ + { + "bbox": [ + 131, + 288, + 480, + 320 + ], + "spans": [ + { + "bbox": [ + 131, + 288, + 480, + 320 + ], + "type": "text", + "content": "Fig. 4: Qualitative results of different artistic style removal methods for eliminating the style of Van Gogh under Ring-A-Bell's attack. The prompt is \"A depiction of a starry night over a quiet town, reminiscent of Van Gogh's famous painting\"." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 219, + 226, + 261, + 269 + ], + "blocks": [ + { + "bbox": [ + 219, + 226, + 261, + 269 + ], + "lines": [ + { + "bbox": [ + 219, + 226, + 261, + 269 + ], + "spans": [ + { + "bbox": [ + 219, + 226, + 261, + 269 + ], + "type": "image", + "image_path": "3c1c6670da3aef5f747eebd760738458c054f438a4104715ab56c75191d6e383.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 231, + 270, + 248, + 277 + ], + "lines": [ + { + "bbox": [ + 231, + 270, + 248, + 277 + ], + "spans": [ + { + "bbox": [ + 231, + 270, + 248, + 277 + ], + "type": "text", + "content": "ESD" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 263, + 226, + 306, + 268 + ], + "blocks": [ + { + "bbox": [ + 263, + 226, + 306, + 268 + ], + "lines": [ + { + "bbox": [ + 263, + 226, + 306, + 268 + ], + "spans": [ + { + "bbox": [ + 263, + 226, + 306, + 268 + ], + "type": "image", + "image_path": "f0921bb0eb7a2e44276d351feb20046c2d7bccf655d6781c8f472aefe08287b5.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 276, + 270, + 294, + 277 + ], + "lines": [ + { + "bbox": [ + 276, + 270, + 294, + 277 + ], + "spans": [ + { + "bbox": [ + 276, + 270, + 294, + 277 + ], + "type": "text", + "content": "UCE" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 308, + 226, + 351, + 268 + ], + "blocks": [ + { + "bbox": [ + 308, + 226, + 351, + 268 + ], + "lines": [ + { + "bbox": [ + 308, + 226, + 351, + 268 + ], + "spans": [ + { + "bbox": [ + 308, + 226, + 351, + 268 + ], + "type": "image", + "image_path": "54906c9f9f6e704295990d1c6fcf3f84456d64e8e201f51a0003e3b04c748307.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 312, + 270, + 347, + 277 + ], + "lines": [ + { + "bbox": [ + 312, + 270, + 347, + 277 + ], + "spans": [ + { + "bbox": [ + 312, + 270, + 347, + 277 + ], + "type": "text", + "content": "SLD-Med" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 352, + 226, + 395, + 268 + ], + "blocks": [ + { + "bbox": [ + 352, + 226, + 395, + 268 + ], + "lines": [ + { + "bbox": [ + 352, + 226, + 395, + 268 + ], + "spans": [ + { + "bbox": [ + 352, + 226, + 395, + 268 + ], + "type": "image", + "image_path": "6c108cc06dcd28666d0902f3554b48090f60776ed47d1e4779b7c7cb629c7349.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 367, + 270, + 380, + 277 + ], + "lines": [ + { + "bbox": [ + 367, + 270, + 380, + 277 + ], + "spans": [ + { + "bbox": [ + 367, + 270, + 380, + 277 + ], + "type": "text", + "content": "CA" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 397, + 226, + 440, + 268 + ], + "blocks": [ + { + "bbox": [ + 397, + 226, + 440, + 268 + ], + "lines": [ + { + "bbox": [ + 397, + 226, + 440, + 268 + ], + "spans": [ + { + "bbox": [ + 397, + 226, + 440, + 268 + ], + "type": "image", + "image_path": "9cf2bdccd184a9659d37be6304ae5d6463b2cf4c9d41dd534f318737f24a6205.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 410, + 270, + 427, + 277 + ], + "lines": [ + { + "bbox": [ + 410, + 270, + 427, + 277 + ], + "spans": [ + { + "bbox": [ + 410, + 270, + 427, + 277 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "bbox": [ + 131, + 342, + 479, + 378 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 342, + 479, + 378 + ], + "spans": [ + { + "bbox": [ + 131, + 342, + 479, + 378 + ], + "type": "text", + "content": "we use a set of 143 prompts selected from I2P, each with a nudity score (as determined by NudeNet) above 0.75, and employ the Nudenet detector with a threshold set to 0.45 for detecting inappropriate content." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 130, + 393, + 480, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 393, + 480, + 525 + ], + "spans": [ + { + "bbox": [ + 130, + 393, + 480, + 525 + ], + "type": "text", + "content": "Results The attack success rates (ASR, %) are summarized in Tab. 4. Our method achieves the best robustness in average. In the case of the black-box attack Ring-A-Bell, our method achieves the lowest ASR at " + }, + { + "bbox": [ + 130, + 393, + 480, + 525 + ], + "type": "inline_equation", + "content": "13.38\\%" + }, + { + "bbox": [ + 130, + 393, + 480, + 525 + ], + "type": "text", + "content": ", significantly outperforming other methods. In the case of the white-box attack, CA achieves the best performance, while our method performs either the best or very closely to the second-best. Although SA achieves a decent black-box result, it consumes substantial computation resources for generating 5000 prepared images [8]. CA modifies " + }, + { + "bbox": [ + 130, + 393, + 480, + 525 + ], + "type": "inline_equation", + "content": "100\\%" + }, + { + "bbox": [ + 130, + 393, + 480, + 525 + ], + "type": "text", + "content": " of the U-Net parameters while our method only modifies K&V projection matrices, constituting a mere " + }, + { + "bbox": [ + 130, + 393, + 480, + 525 + ], + "type": "inline_equation", + "content": "2.23\\%" + }, + { + "bbox": [ + 130, + 393, + 480, + 525 + ], + "type": "text", + "content": " of the U-Net. While UCE also modifies only " + }, + { + "bbox": [ + 130, + 393, + 480, + 525 + ], + "type": "inline_equation", + "content": "2.23\\%" + }, + { + "bbox": [ + 130, + 393, + 480, + 525 + ], + "type": "text", + "content": " parameters, our method significantly outperforms UCE. This is attributed to our derived embeddings in Eq. (8)." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 130, + 525, + 480, + 609 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 525, + 480, + 609 + ], + "spans": [ + { + "bbox": [ + 130, + 525, + 480, + 609 + ], + "type": "text", + "content": "For artistic style removal, we provide qualitative examples in Fig. 4. The first image is generated by the original SD v1.4 without any attack, and the following images are under Ring-A-Bell's attack [32]. Recall that UCE is the second-best artistic style removal method as shown in Tab. 3 but it falls short in robustness against red-teaming tools. Although our method employs a closed-form solution like UCE, it outperforms UCE in robustness. ESD, CA and our method perform similarly well. More results can be found in the appendix." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 131, + 624, + 282, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 624, + 282, + 635 + ], + "spans": [ + { + "bbox": [ + 131, + 624, + 282, + 635 + ], + "type": "text", + "content": "4.4 Model Editing Duration" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 131, + 641, + 479, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 641, + 479, + 665 + ], + "spans": [ + { + "bbox": [ + 131, + 641, + 479, + 665 + ], + "type": "text", + "content": "To demonstrate the efficiency of different methods, we measured the percentage of parameter modification and editing duration on an RTX 3090 for each" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 249, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 249, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 249, + 102 + ], + "type": "text", + "content": "F. Chao Gong et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 213, + 114, + 401, + 158 + ], + "blocks": [ + { + "bbox": [ + 213, + 114, + 401, + 158 + ], + "lines": [ + { + "bbox": [ + 213, + 114, + 401, + 158 + ], + "spans": [ + { + "bbox": [ + 213, + 114, + 401, + 158 + ], + "type": "table", + "html": "
ESDUCESACAOurs
Modification (%)94.652.2394.651002.23
Duration (s)37201.2-14003.4
", + "image_path": "9e64228585d90f0bf25041bb9f342286090fb94826a285d84c8cb8d1b1a0169d.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 158, + 479, + 180 + ], + "lines": [ + { + "bbox": [ + 132, + 158, + 479, + 180 + ], + "spans": [ + { + "bbox": [ + 132, + 158, + 479, + 180 + ], + "type": "text", + "content": "Table 5: Percentage of parameter modification and model editing duration for different methods. Our method and UCE are significantly ahead of other methods." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 130, + 214, + 480, + 274 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 214, + 480, + 274 + ], + "spans": [ + { + "bbox": [ + 130, + 214, + 480, + 274 + ], + "type": "text", + "content": "method, as shown in Tab. 5. We excluded SLD from the analysis since it operates at inference time rather than modifying the model's weights which can be easily bypassed under open-source conditions. Additionally, we don't include the duration of SA, as it involves generating 5000 images, calculating the Fisher Information Matrix and fine-tuning, which makes it exceptionally slow." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 274, + 481, + 358 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 274, + 481, + 358 + ], + "spans": [ + { + "bbox": [ + 130, + 274, + 481, + 358 + ], + "type": "text", + "content": "Based on Tab. 5 and Tab. 1, our method achieves the best concept erasure effect in an extremely short time of only three seconds. Our method (5 epochs) and UCE (1 epoch) modify the lowest percentage of parameters with a closed-form solution, resulting in the shortest editing durations. Despite similar durations, our method significantly outperforms UCE in removal effectiveness. Conversely, CA, ESD and SA modify a high percentage of parameters with more time but achieve less impressive removal results." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 376, + 237, + 388 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 376, + 237, + 388 + ], + "spans": [ + { + "bbox": [ + 132, + 376, + 237, + 388 + ], + "type": "text", + "content": "4.5 Ablation Study" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 396, + 480, + 420 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 396, + 480, + 420 + ], + "spans": [ + { + "bbox": [ + 131, + 396, + 480, + 420 + ], + "type": "text", + "content": "We conduct experiments to elucidate the impact of our derived embedding among different epochs and the effectiveness of the regularization term." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 437, + 253, + 462 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 437, + 253, + 462 + ], + "spans": [ + { + "bbox": [ + 132, + 437, + 253, + 462 + ], + "type": "text", + "content": "Effect of Derived Embeddings among Epochs" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 462, + 247, + 580 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 462, + 247, + 580 + ], + "spans": [ + { + "bbox": [ + 130, + 462, + 247, + 580 + ], + "type": "text", + "content": "We conduct an experiment to expound the impact of our derived embedding. We perform \"nudity\" erasure for 5 epochs using Algorithm 1. In each epoch, we derive a distinct embedding that represents \"nudity\". Before the erasure of each" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 582, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 481, + 665 + ], + "type": "text", + "content": "epoch, we generate an image using the embedding to test its degree of nudity information, as presented in Fig. 5. Images from epoch 0 to epoch 2 contain nude body parts, indicating that our derived embeddings successfully reveal potential nudity information in the model. Specifically, we opt for the checkpoint after epoch 2, as images from epoch 3 to 4 lack nudity information. Actually, erasing such \"not so nude\" embeddings in epoch 3-4 would impair the model's normal generation ability, which is an unworthy trade-off." + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 260, + 463, + 476, + 518 + ], + "blocks": [ + { + "bbox": [ + 260, + 463, + 476, + 518 + ], + "lines": [ + { + "bbox": [ + 260, + 463, + 476, + 518 + ], + "spans": [ + { + "bbox": [ + 260, + 463, + 476, + 518 + ], + "type": "image", + "image_path": "a6cd6fddd50fd3f3a1017cdb441762c6eabcc41be1c916ab35ab5ac81f034c98.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 252, + 526, + 482, + 560 + ], + "lines": [ + { + "bbox": [ + 252, + 526, + 482, + 560 + ], + "spans": [ + { + "bbox": [ + 252, + 526, + 482, + 560 + ], + "type": "text", + "content": "Fig. 5: There is no nudity information from epoch 3 onward; hence, we select the checkpoint after epoch 2 to avoid damaging the model's ability." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 146, + 90, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 90, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 146, + 90, + 447, + 102 + ], + "type": "text", + "content": "Reliable and Efficient Concept Erasure of Text-to-Image Diffusion Models" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 160, + 115, + 455, + 186 + ], + "blocks": [ + { + "bbox": [ + 160, + 115, + 455, + 186 + ], + "lines": [ + { + "bbox": [ + 160, + 115, + 455, + 186 + ], + "spans": [ + { + "bbox": [ + 160, + 115, + 455, + 186 + ], + "type": "image", + "image_path": "09f9728f5f221f3a57c35ee85a361d4e48298663c7ed1c8bb330ce059fead620.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 197, + 480, + 209 + ], + "lines": [ + { + "bbox": [ + 132, + 197, + 480, + 209 + ], + "spans": [ + { + "bbox": [ + 132, + 197, + 480, + 209 + ], + "type": "text", + "content": "Fig. 6: Ablation study on the impact of the regularization term on model performance." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 239, + 482, + 395 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 239, + 482, + 395 + ], + "spans": [ + { + "bbox": [ + 130, + 239, + 482, + 395 + ], + "type": "text", + "content": "Effect of Regularization Coefficient We conduct an experiment to assess the influence of our regularization term. Specifically, we select different regularization coefficients " + }, + { + "bbox": [ + 130, + 239, + 482, + 395 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 130, + 239, + 482, + 395 + ], + "type": "text", + "content": " in Eq. (4), which divide the interval [0, 1] into five parts. The results, presented in Fig. 6, include CLIP-score and LPIPS against the original SD on the COCO-30k validation subset. As " + }, + { + "bbox": [ + 130, + 239, + 482, + 395 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 130, + 239, + 482, + 395 + ], + "type": "text", + "content": " increases, CLIP score shows an upward trend while LPIPS and the difference between new and old parameters show a downward trend. This indicates the role of the regularization term in preserving the model's ability for unerased content. Furthermore, we recorded the number of nude parts on the I2P benchmark, presented in the third column of Fig. 6. However, the number of nude parts doesn't strictly increase as " + }, + { + "bbox": [ + 130, + 239, + 482, + 395 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 130, + 239, + 482, + 395 + ], + "type": "text", + "content": " increases, which is counterintuitive. Although the purpose of the regularization term in Eq. (6) is to preserve the model's generation capability, maintaining this capability does not always affect the erasure effect." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 431, + 220, + 443 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 431, + 220, + 443 + ], + "spans": [ + { + "bbox": [ + 132, + 431, + 220, + 443 + ], + "type": "text", + "content": "5 Conclusion" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 474, + 482, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 474, + 482, + 668 + ], + "spans": [ + { + "bbox": [ + 130, + 474, + 482, + 668 + ], + "type": "text", + "content": "In this paper, we propose a novel approach for reliably and efficiently erasing specific concepts from Text-to-Image (T2I) diffusion models. Our approach only modifies the cross-attention K&V matrices of U-Net, constituting a mere " + }, + { + "bbox": [ + 130, + 474, + 482, + 668 + ], + "type": "inline_equation", + "content": "2.23\\%" + }, + { + "bbox": [ + 130, + 474, + 482, + 668 + ], + "type": "text", + "content": " of parameters. While previous methods also edited cross-attention modules, they still exhibited the ability to generate inappropriate images. To tackle this challenge, we derive and erase new embeddings that can represent target concepts within unlearned models. To mitigate the impact on unrelated concepts, a regularization term is introduced during the erasure process. All the above techniques are formulated in closed-form, facilitating rapid editing. This enables the execution of \"derive-erase\" across multiple epochs, ensuring thorough and robust erasure. Extensive experiments were conducted to validate the effectiveness of our approach in erasing artistic styles, unsafe contents and object classes. Furthermore, we recorded editing durations to underscore the efficiency of our method and evaluated the robustness against red-teaming tools. We believe our RECE has the potential to empower T2I providers in effectively removing undesired concepts, thereby fostering the development of a safer AI community." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 249, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 249, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 249, + 102 + ], + "type": "text", + "content": "F. Chao Gong et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 115, + 246, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 115, + 246, + 129 + ], + "spans": [ + { + "bbox": [ + 133, + 115, + 246, + 129 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 141, + 479, + 166 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 141, + 479, + 166 + ], + "spans": [ + { + "bbox": [ + 130, + 141, + 479, + 166 + ], + "type": "text", + "content": "This project was supported by National Key R&D Program of China (No. 2021ZD0112804)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 185, + 198, + 198 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 185, + 198, + 198 + ], + "spans": [ + { + "bbox": [ + 132, + 185, + 198, + 198 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 211, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 138, + 211, + 481, + 244 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 211, + 481, + 244 + ], + "spans": [ + { + "bbox": [ + 138, + 211, + 481, + 244 + ], + "type": "text", + "content": "1. Chen, K., Wei, Z., Chen, J., Wu, Z., Jiang, Y.G.: Attacking video recognition models with bullet-screen comments. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 36, pp. 312-320 (2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 246, + 481, + 277 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 246, + 481, + 277 + ], + "spans": [ + { + "bbox": [ + 138, + 246, + 481, + 277 + ], + "type": "text", + "content": "2. Chen, K., Wei, Z., Chen, J., Wu, Z., Jiang, Y.G.: Gcma: Generative cross-modal transferable adversarial attacks from images to videos. In: Proceedings of the 31st ACM International Conference on Multimedia. pp. 698-708 (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 279, + 481, + 311 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 279, + 481, + 311 + ], + "spans": [ + { + "bbox": [ + 138, + 279, + 481, + 311 + ], + "type": "text", + "content": "3. Chin, Z.Y., Jiang, C.M., Huang, C.C., Chen, P.Y., Chiu, W.C.: Prompting4debugging: Red-teaming text-to-image diffusion models by finding problematic prompts. arXiv preprint arXiv:2309.06135 (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 312, + 481, + 333 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 312, + 481, + 333 + ], + "spans": [ + { + "bbox": [ + 138, + 312, + 481, + 333 + ], + "type": "text", + "content": "4. Dhariwal, P., Nichol, A.: Diffusion models beat gans on image synthesis. Advances in neural information processing systems 34, 8780-8794 (2021)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 335, + 481, + 355 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 335, + 481, + 355 + ], + "spans": [ + { + "bbox": [ + 138, + 335, + 481, + 355 + ], + "type": "text", + "content": "5. Doersch, C.: Tutorial on variational autoencoders. arXiv preprint arXiv:1606.05908 (2016)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 357, + 481, + 377 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 357, + 481, + 377 + ], + "spans": [ + { + "bbox": [ + 138, + 357, + 481, + 377 + ], + "type": "text", + "content": "6. Gandikota, R., Materzynska, J., Fiotto-Kaufman, J., Bau, D.: Erasing concepts from diffusion models. arXiv preprint arXiv:2303.07345 (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 379, + 481, + 411 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 379, + 481, + 411 + ], + "spans": [ + { + "bbox": [ + 138, + 379, + 481, + 411 + ], + "type": "text", + "content": "7. Gandikota, R., Orgad, H., Belinkov, Y., Materzyńska, J., Bau, D.: Unified concept editing in diffusion models. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 5111-5120 (2024)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 412, + 481, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 412, + 481, + 443 + ], + "spans": [ + { + "bbox": [ + 138, + 412, + 481, + 443 + ], + "type": "text", + "content": "8. Heng, A., Soh, H.: Selective amnesia: A continual learning approach to forgetting in deep generative models. Advances in Neural Information Processing Systems 36 (2024)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 445, + 481, + 466 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 445, + 481, + 466 + ], + "spans": [ + { + "bbox": [ + 138, + 445, + 481, + 466 + ], + "type": "text", + "content": "9. Hessel, J., Holtzman, A., Forbes, M., Bras, R.L., Choi, Y.: Clipscore: A reference-free evaluation metric for image captioning. arXiv preprint arXiv:2104.08718 (2021)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 468, + 481, + 488 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 468, + 481, + 488 + ], + "spans": [ + { + "bbox": [ + 138, + 468, + 481, + 488 + ], + "type": "text", + "content": "0. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. Advances in neural information processing systems 33, 6840-6851 (2020)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 490, + 481, + 500 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 490, + 481, + 500 + ], + "spans": [ + { + "bbox": [ + 138, + 490, + 481, + 500 + ], + "type": "text", + "content": "1. Hunter, T.: Ai porn is easy to make now. for women, that's a nightmare. (2 2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 502, + 481, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 502, + 481, + 544 + ], + "spans": [ + { + "bbox": [ + 138, + 502, + 481, + 544 + ], + "type": "text", + "content": "2. Kawar, B., Zada, S., Lang, O., Tov, O., Chang, H., Dekel, T., Mosseri, I., Irani, M.: Imagic: Text-based real image editing with diffusion models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6007-6017 (2023)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 545, + 481, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 545, + 481, + 577 + ], + "spans": [ + { + "bbox": [ + 138, + 545, + 481, + 577 + ], + "type": "text", + "content": "3. Kumari, N., Zhang, B., Wang, S.Y., Shechtman, E., Zhang, R., Zhu, J.Y.: Ablating concepts in text-to-image diffusion models. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 22691-22702 (2023)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 578, + 481, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 578, + 481, + 621 + ], + "spans": [ + { + "bbox": [ + 138, + 578, + 481, + 621 + ], + "type": "text", + "content": "4. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollar, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13. pp. 740-755. Springer (2014)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 138, + 622, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 622, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 622, + 481, + 665 + ], + "type": "text", + "content": "5. Nichol, A.Q., Dhariwal, P., Ramesh, A., Shyam, P., Mishkin, P., Mcgrew, B., Sutskever, I., Chen, M.: Glide: Towards photorealistic image generation and editing with text-guided diffusion models. In: International Conference on Machine Learning. pp. 16784-16804. PMLR (2022)" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 146, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 146, + 91, + 447, + 102 + ], + "type": "text", + "content": "Reliable and Efficient Concept Erasure of Text-to-Image Diffusion Models" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 117, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 133, + 117, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 117, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 117, + 480, + 149 + ], + "type": "text", + "content": "16. Orgad, H., Kawar, B., Belinkov, Y.: Editing implicit assumptions in text-to-image diffusion models. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 7053-7061 (2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 150, + 480, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 150, + 480, + 183 + ], + "spans": [ + { + "bbox": [ + 133, + 150, + 480, + 183 + ], + "type": "text", + "content": "17. Parmar, G., Zhang, R., Zhu, J.Y.: On aliased resizing and surprising subtleties in gan evaluation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11410-11420 (2022)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 183, + 480, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 183, + 480, + 205 + ], + "spans": [ + { + "bbox": [ + 133, + 183, + 480, + 205 + ], + "type": "text", + "content": "18. Praneeth, B.: Nudenet: Neural nets for nudity classification, detection and selective censoring (2019)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 134, + 205, + 480, + 247 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 205, + 480, + 247 + ], + "spans": [ + { + "bbox": [ + 134, + 205, + 480, + 247 + ], + "type": "text", + "content": "19. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 133, + 249, + 480, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 249, + 480, + 281 + ], + "spans": [ + { + "bbox": [ + 133, + 249, + 480, + 281 + ], + "type": "text", + "content": "20. Ramesh, A., Dhariwal, P., Nichol, A., Chu, C., Chen, M.: Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125 1(2), 3 (2022)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 281, + 480, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 281, + 480, + 303 + ], + "spans": [ + { + "bbox": [ + 132, + 281, + 480, + 303 + ], + "type": "text", + "content": "21. Rando, J., Paleka, D., Lindner, D., Heim, L., Tramér, F.: Red-teaming the stable diffusion safety filter. arXiv preprint arXiv:2210.04610 (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 303, + 388, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 303, + 388, + 314 + ], + "spans": [ + { + "bbox": [ + 132, + 303, + 388, + 314 + ], + "type": "text", + "content": "22. Rombach, R.: Stable diffusion 2.0 release (November 2022)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 315, + 480, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 315, + 480, + 326 + ], + "spans": [ + { + "bbox": [ + 132, + 315, + 480, + 326 + ], + "type": "text", + "content": "23. Rombach, R.: Tutorial: How to remove the safety filter in 5 seconds (August 2022)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 326, + 480, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 326, + 480, + 358 + ], + "spans": [ + { + "bbox": [ + 132, + 326, + 480, + 358 + ], + "type": "text", + "content": "24. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10684-10695 (2022)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 358, + 462, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 358, + 462, + 369 + ], + "spans": [ + { + "bbox": [ + 132, + 358, + 462, + 369 + ], + "type": "text", + "content": "25. Rombach, R., Esser, P.: Stable diffusion v1-4 model card. Model Card (2022)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 133, + 369, + 480, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 369, + 480, + 413 + ], + "spans": [ + { + "bbox": [ + 133, + 369, + 480, + 413 + ], + "type": "text", + "content": "26. Ronneberger, O., Fischer, P., Brox, T.: U-net: Convolutional networks for biomedical image segmentation. In: Medical image computing and computer-assisted intervention-MICCAI 2015: 18th international conference, Munich, Germany, October 5-9, 2015, proceedings, part III 18. pp. 234-241. Springer (2015)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 133, + 413, + 480, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 413, + 480, + 456 + ], + "spans": [ + { + "bbox": [ + 133, + 413, + 480, + 456 + ], + "type": "text", + "content": "27. Ruiz, N., Li, Y., Jampani, V., Pritch, Y., Rubinstein, M., Aberman, K.: Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 22500-22510 (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 133, + 457, + 480, + 500 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 457, + 480, + 500 + ], + "spans": [ + { + "bbox": [ + 133, + 457, + 480, + 500 + ], + "type": "text", + "content": "28. Saharia, C., Chan, W., Saxena, S., Li, L., Whang, J., Denton, E.L., Ghasemipour, K., Gontijo Lopes, R., Karagol Ayan, B., Salimans, T., et al.: Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems 35, 36479-36494 (2022)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 133, + 501, + 480, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 501, + 480, + 544 + ], + "spans": [ + { + "bbox": [ + 133, + 501, + 480, + 544 + ], + "type": "text", + "content": "29. Schramowski, P., Brack, M., Deiseroth, B., Kersting, K.: Safe latent diffusion: Mitigating inappropriate degeneration in diffusion models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 22522-22531 (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 545, + 472, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 545, + 472, + 555 + ], + "spans": [ + { + "bbox": [ + 132, + 545, + 472, + 555 + ], + "type": "text", + "content": "30. Setty, R.: Ai art generators hit with copyright suit over artists' images (1 2023)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 555, + 408, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 555, + 408, + 566 + ], + "spans": [ + { + "bbox": [ + 132, + 555, + 408, + 566 + ], + "type": "text", + "content": "31. StabilityAI: Stable diffusion 2.1 model card. Model Card (2022)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 133, + 567, + 480, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 567, + 480, + 599 + ], + "spans": [ + { + "bbox": [ + 133, + 567, + 480, + 599 + ], + "type": "text", + "content": "32. Tsai, Y.L., Hsu, C.Y., Xie, C., Lin, C.H., Chen, J.Y., Li, B., Chen, P.Y., Yu, C.M., Huang, C.Y.: Ring-a-bell! how reliable are concept removal methods for diffusion models? arXiv preprint arXiv:2310.10012 (2023)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 133, + 600, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 600, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 133, + 600, + 480, + 632 + ], + "type": "text", + "content": "33. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. Advances in neural information processing systems 30 (2017)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 133, + 632, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 632, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 632, + 480, + 665 + ], + "type": "text", + "content": "34. Wei, Z., Chen, J., Goldblum, M., Wu, Z., Goldstein, T., Jiang, Y.G., Davis, L.S.: Towards transferable adversarial attacks on image and video transformers. IEEE Transactions on Image Processing 32, 6346-6358 (2023)" + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 248, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 248, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 248, + 101 + ], + "type": "text", + "content": "F. Chao Gong et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 482, + 249 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 132, + 116, + 482, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 482, + 149 + ], + "type": "text", + "content": "35. Wei, Z., Chen, J., Wu, Z., Jiang, Y.G.: Adaptive cross-modal transferable adversarial attacks from images to videos. IEEE Transactions on Pattern Analysis and Machine Intelligence (2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 150, + 482, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 150, + 482, + 183 + ], + "spans": [ + { + "bbox": [ + 132, + 150, + 482, + 183 + ], + "type": "text", + "content": "36. Yang, L., Zhang, Z., Song, Y., Hong, S., Xu, R., Zhao, Y., Zhang, W., Cui, B., Yang, M.H.: Diffusion models: A comprehensive survey of methods and applications. ACM Computing Surveys 56(4), 1-39 (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 183, + 482, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 183, + 482, + 216 + ], + "spans": [ + { + "bbox": [ + 132, + 183, + 482, + 216 + ], + "type": "text", + "content": "37. Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual metric. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 586-595 (2018)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 216, + 482, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 216, + 482, + 249 + ], + "spans": [ + { + "bbox": [ + 132, + 216, + 482, + 249 + ], + "type": "text", + "content": "38. Zhang, Y., Jia, J., Chen, X., Chen, A., Zhang, Y., Liu, J., Ding, K., Liu, S.: To generate or not? safety-driven unlearned diffusion models are still easy to generate unsafe images... for now. arXiv preprint arXiv:2310.11868 (2023)" + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 146, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 146, + 91, + 447, + 102 + ], + "type": "text", + "content": "Reliable and Efficient Concept Erasure of Text-to-Image Diffusion Models" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Relightable 3D Gaussians_ Realistic Point Cloud Relighting with BRDF Decomposition and Ray Tracing/8f8cf818-147d-408c-857b-28f5e4e05e99_content_list.json b/2024/Relightable 3D Gaussians_ Realistic Point Cloud Relighting with BRDF Decomposition and Ray Tracing/8f8cf818-147d-408c-857b-28f5e4e05e99_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..1739acbba47004223b2990388b734202a6505835 --- /dev/null +++ b/2024/Relightable 3D Gaussians_ Realistic Point Cloud Relighting with BRDF Decomposition and Ray Tracing/8f8cf818-147d-408c-857b-28f5e4e05e99_content_list.json @@ -0,0 +1,2106 @@ +[ + { + "type": "text", + "text": "Relightable 3D Gaussians: Realistic Point Cloud Relighting with BRDF Decomposition and Ray Tracing", + "text_level": 1, + "bbox": [ + 225, + 140, + 779, + 209 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jian $\\mathrm{Gao}^{1*}$ , Chun $\\mathrm{Gu}^{2*}$ , Youtian Lin $^{1}$ , Zhihao Li $^{3}$ , Hao Zhu $^{1}$ , Xun $\\mathrm{Cao}^{1}$ , Li Zhang $^{2\\boxtimes}$ , and Yao $\\mathrm{Yao}^{1\\boxtimes}$", + "bbox": [ + 236, + 233, + 764, + 265 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Nanjing University", + "bbox": [ + 431, + 276, + 571, + 290 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ Fudan University", + "bbox": [ + 433, + 291, + 566, + 304 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3 Huawei Noah's Ark Lab", + "bbox": [ + 413, + 305, + 589, + 318 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. In this paper, we present a novel differentiable point-based rendering framework to achieve photo-realistic relighting. To make the reconstructed scene relightable, we enhance vanilla 3D Gaussians by associating extra properties, including normal vectors, BRDF parameters, and incident lighting from various directions. From a collection of multiview images, the 3D scene is optimized through 3D Gaussian Splatting while BRDF and lighting are decomposed by physically based differentiable rendering. To produce plausible shadow effects in photo-realistic relighting, we introduce an innovative point-based ray tracing with the bounding volume hierarchies for efficient visibility pre-computation. Extensive experiments demonstrate our improved BRDF estimation, novel view synthesis and relighting results compared to state-of-the-art approaches. The proposed framework showcases the potential to revolutionize the mesh-based graphics pipeline with a point-based pipeline enabling editing, tracing, and relighting.", + "bbox": [ + 259, + 357, + 743, + 566 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: 3D Gaussian Splatting, Relighting, Point based rendering", + "bbox": [ + 261, + 578, + 733, + 593 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 618, + 375, + 633 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Reconstructing 3D scenes from multi-view images for photo-realistic rendering is a fundamental problem at the intersection of computer vision and graphics. Recently, 3D Gaussian Splatting (3DGS) [26] has been proposed and has gained significant attention from the community. The method employs a set of 3D Gaussian points to represent a 3D scene and projects these points onto a designated view through a tile-based rasterization. Attributes of each 3D Gaussian point are then optimized through the point-based differentiable rendering. Notably, 3DGS achieves real-time rendering with quality comparable to or even surpassing previous state-of-the-art methods (e.g., Mip-NeRF [3]), with a training speed on par with the most efficient Instant-NGP [32]. However, the current 3DGS is unable to reconstruct a scene that can be relighted under different lighting conditions,", + "bbox": [ + 212, + 650, + 787, + 816 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equally contributed.", + "bbox": [ + 232, + 825, + 380, + 840 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/db90ca5bf82a52b5d6d04ed8397b42380b6f5d0b08db17516332ec2f153721dc.jpg", + "image_caption": [ + "(a) Point Cloud" + ], + "image_footnote": [], + "bbox": [ + 220, + 143, + 500, + 229 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/77fad46f9adc611e9ba0cf90993ce1f1e0aa703b8fead9cdd503726d1c334d57.jpg", + "image_caption": [ + "(b) Normal" + ], + "image_footnote": [], + "bbox": [ + 501, + 143, + 782, + 229 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/4764d770bb5551e8d11a38b78daafcf236583b1ee34141e43ba563b3f00de2e1.jpg", + "image_caption": [ + "(c) Ambient Occlusion" + ], + "image_footnote": [], + "bbox": [ + 220, + 243, + 500, + 325 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/0b357a25004a1d765e431fb52522809bbd1a4e14a34ff10c8b10278a46435edf.jpg", + "image_caption": [ + "(d) Physically Based Rendering", + "Fig. 1: Visual results of our pipeline on a multi-object composition scene. In our pipeline, we represent a scene as Relightable 3D Gaussians. From multi-view images, we recover the geometry and materials of individual objects with inverse rendering techniques (see Sec. 3). Then, objects are easily composited into a new scene, thanks to our explicit representation. After that, we solve the complex occlusions though point based ray tracing (see Sec. 4) and re-light the new scene. Ultimately, we achieve high-fidelity relighting with remarkably realistic shadow." + ], + "image_footnote": [], + "bbox": [ + 501, + 243, + 782, + 327 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "making the method only applicable to the task of novel view synthesis. In addition, ray tracing, a crucial component in achieving realistic rendering, remains an unresolved challenge in point-based representation, limiting 3DGS to more dedicated rendering effects such as shadowing and light reflectance.", + "bbox": [ + 212, + 498, + 784, + 559 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we aim to reconstruct a re烈table 3D Gaussian point cloud from multi-view images with a differentiable rendering framework and achieve realistic relighting. We make 3D Gaussian relightable by assigning it with additional normal, BRDF properties, and incident light information to model perpoint light reflectance. In contrast to the plain alpha blending in the original 3DGS [26], we apply physically based rendering (PBR) to get a PBR color for each 3D Gaussian point, which is then alpha-composed to obtain a rendered color for the corresponding image pixel. For robust material and lighting estimation, we split the incident light into a global environment map and an indirect incident light field. To capture accurate visibility for each 3D Gaussian, we propose a novel ray tracing method based on the bounding volume hierarchy (BVH), which enables efficient visibility pre-computing for real-time rendering. Additionally, the proposed ray tracing method can handle complex occlusion relationships in a novel multi-object composition scene, thus realizing realistic shadow effects. Moreover, proper regularizations are introduced to enhance the geometry and mitigate the material-lighting ambiguity during the optimization, including constraints on depth distribution, smoothness priors, and a lighting regularization.", + "bbox": [ + 212, + 568, + 787, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "J. Gao et al.", + "bbox": [ + 271, + 114, + 357, + 126 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Extensive experiments conducted across diverse datasets demonstrate the improved material and lighting estimation results and novel view rendering quality. Additionally, we illustrate the relightable and editable capabilities of our system through multi-object composition in a novel lighting environment. To summarize, major contributions of the paper include:", + "bbox": [ + 212, + 146, + 782, + 220 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose a material and lighting decomposition scheme for 3D Gaussian Splatting, where a normal vector, BRDF values, and incident lights are assigned and optimized for each 3D Gaussian point.", + "- We introduce a novel point-based ray tracing approach based on bounding volume hierarchy, enabling efficient visibility pre-computing for each 3D Gaussian point and rendering of a 3D scene with realistic shadow effect.", + "- We demonstrate a comprehensive graphics pipeline solely based on a discretized point representation, supporting relighting, editing, and ray tracing of a reconstructed 3D point cloud." + ], + "bbox": [ + 225, + 229, + 782, + 363 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Works", + "text_level": 1, + "bbox": [ + 215, + 386, + 395, + 402 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Neural Radiance Field. Differentiable rendering techniques, exemplified by Neural Radiance Field (NeRF) [31], have garnered significant attention [3,12,32]. NeRF utilizes an implicit Multi-Layer Perceptron (MLP) that takes 3D positions and viewing directions as inputs to generate density and view-dependent colors for differentiable volume rendering. Despite its powerful neural implicit representation, both speed and quality can be improved. Efficiency improvements mainly focus on optimizing queries on neural fields and minimizing queries [12, 13, 15, 16, 21, 22, 32, 39]. Quality enhancements involve anti-aliasing [3-5] or utilizing exterior supervision [14, 48], etc.", + "bbox": [ + 212, + 416, + 784, + 551 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Differentiable Point Based Rendering. Using points as rendering primitives was first proposed in [29]. To address the issue of resulting holey images when rasterizing discrete points directly, solutions fall into two categories. One approach directly employs points as rendering primitives and encodes the geometric and photometric features near the point using SIFT descriptors [37] or neural descriptors [1,38]. This leads to a feature image with gaps, which is then decoded to recover a hole-free RGB image. The other category models a point as a primitive occupying a specific space, like a surfel [36,47] or a 3D Gaussian [26]. The rendered image is then generated using a splatting technique. This technique saw significant development in the 2000s [36,57,58], and in the era of differentiable rendering, PointRF [52], DSS [47] and 3DGS [26] serve as notable examples. We assert that the point serves as a promising rendering primitive owing to its inherent ease of editing, as also corroborated in [54].", + "bbox": [ + 212, + 553, + 784, + 748 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Material and Lighting Estimation. Decomposing the materials and illumination of a scene from multi-view images presents a formidable challenge owing to its intrinsic high-dimensional complexity. Some methods simplified the problem under the assumption of a controllable light [2,6-8,18,34,35,40]. Subsequent research explores more complex lighting models to cope with realistic scenarios. NeRV [41] and PhySG [51] leverage an environmental map to manage arbitrary", + "bbox": [ + 212, + 750, + 782, + 839 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "Relightable 3D Gaussians", + "bbox": [ + 558, + 114, + 730, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "lighting conditions. NeRD [9] addresses the challenge posed by images captured under varying illumination by attributing Spherical Gaussians to each image. IRON [50] introduces an innovative edge sampling algorithm tailored for neural SDFs. NeRFactor [53] exploits light visibility to achieve superior material and lighting decomposition. Further studies [19, 52] address the consideration of indirect lighting, leading to enhanced BRDF estimation quality. [20, 33] utilize differentiable marching tetrahedrons for direct optimization on mesh surfaces. [10, 28] deal with varying cameras, illumination, and backgrounds. Ref-NeRF [42] suggests the employment of integrated direction encoding to ameliorate the reconstruction fidelity of reflective objects. [30, 33] use split-sum approximation to approximate the shading effects. NeMF [55] represents the scene as a microflake volume. NeFII [44] integrates lights through path tracing with Monte Carlo sampling. InvRender [56] computes indirect illumination by directly leveraging the neural radiance field, rather than concurrently estimating it alongside the decomposition of direct lighting and materials. TensoIR [23] performs inverse rendering based on tensor factorization and neural fields. NeILF [45] and NeILF++ [49] expresses the incident lights as a neural incident light field, while NeILF++ integrates VolSDF [46] with NeILF through inter-reflection. However, existing schemes rarely delve into BRDF estimation for point clouds, and the estimation and rendering quality remain to be improved.", + "bbox": [ + 212, + 145, + 787, + 450 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3 Relightable 3D Gaussians", + "text_level": 1, + "bbox": [ + 215, + 465, + 501, + 484 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this section, we introduce a novel pipeline to decompose materials and lighting from multi-view images based on 3D Gaussian Splatting (3DGS) [26]. An overview of our pipeline is shown in Fig. 2.", + "bbox": [ + 212, + 493, + 785, + 541 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1 Preliminary", + "text_level": 1, + "bbox": [ + 215, + 556, + 361, + 573 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Distinct from the widely adopted Neural Radiance Field, 3DGS [26] employs explicit 3D Gaussian points as its rendering primitives. A 3D Gaussian point is mathematically defined as:", + "bbox": [ + 212, + 575, + 785, + 622 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nG (\\boldsymbol {x}) = \\exp \\left(- \\frac {1}{2} (\\boldsymbol {x} - \\boldsymbol {\\mu}) ^ {\\top} \\Sigma^ {- 1} (\\boldsymbol {x} - \\boldsymbol {\\mu})\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 362, + 627, + 785, + 656 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\pmb{\\mu}$ and $\\Sigma$ denote the 3D spatial mean and covariance matrix, respectively. Each Gaussian is also equipped with an opacity $o$ and a view-dependent color $\\pmb{c}$ .", + "bbox": [ + 212, + 659, + 785, + 689 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The rendering process in 3DGS is divided into two main steps. Firstly, 3D Gaussians are projected to 2D Gaussians on the image plane. The 2D means are determined by accurate projection of 3D means, while the 2D covariance matrices are approximated by: $\\Sigma' = J\\boldsymbol{W} \\Sigma \\boldsymbol{W}^\\top \\boldsymbol{J}^\\top$ , where $\\boldsymbol{W}$ and $\\boldsymbol{J}$ denote the viewing transformation and the Jacobian of the affine approximation of perspective projection transformation [57]. Subsequently, the pixel color is derived by alpha blending $N$ ordered 2D Gaussians from front to back:", + "bbox": [ + 212, + 689, + 785, + 796 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {C} = \\sum_ {i \\in N} T _ {i} \\alpha_ {i} \\boldsymbol {c} _ {i}, \\quad T _ {i} = \\prod_ {j = 1} ^ {i - 1} (1 - \\alpha_ {j}). \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 380, + 801, + 785, + 844 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "J. Gao et al.", + "bbox": [ + 271, + 114, + 357, + 127 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/f0101fd25b550b2fd16a3b6c60da2720ceae47ec9045536a575998c4f6bd5039.jpg", + "image_caption": [ + "Fig. 2: The proposed differentiable rendering pipeline. Starting with a collection of 3D Gaussians that embody geometry, materials, and lighting attributes, we first execute the physically based rendering equation for each 3D Gaussian to determine its outgoing radiance, denoted as $c'$ . Following this, we perform rasterization and alpha blending to obtain vanilla color map $\\mathcal{C}$ , PBR color map $\\mathcal{C}'$ , depth map $\\mathcal{D}$ , normal map $\\mathcal{N}$ , etc. To optimize relightable 3D Gaussians, we utilize the ground truth image $\\mathcal{C}_{gt}$ and the pseudo normal map derived from $\\mathcal{D}$ for supervision." + ], + "image_footnote": [], + "bbox": [ + 220, + 146, + 781, + 297 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Here, $\\alpha$ is obtained by multiplying the opacity $o$ with the 2D covariance's contribution computed from $\\Sigma^{\\prime}$ and pixel coordinate in image space [26]. In implementation details, the covariance matrix $\\Sigma$ is parameterized as a unit quaternion $\\pmb{q}$ and a 3D scaling vector $\\pmb{s}$ to maintain its meaningful interpretation throughout optimization. Additionally, view-dependent color $\\pmb{c}_i$ is represented through a set of Spherical Harmonics (SH).", + "bbox": [ + 212, + 441, + 787, + 531 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In summation, a 3D scene is represented by a collection of 3D Gaussians, with the $i^{th}$ Gaussian $\\mathcal{P}_i$ parameterized as $\\{\\pmb{\\mu}_i, \\pmb{q}_i, \\pmb{s}_i, o_i, \\pmb{c}_i\\}$ .", + "bbox": [ + 212, + 534, + 785, + 566 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2 Geometry Enhancement", + "text_level": 1, + "bbox": [ + 214, + 599, + 462, + 616 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Scene geometry, specifically surface normal, is essential for realistic physically based rendering, which will be discussed in Sec. 3.3.", + "bbox": [ + 212, + 638, + 785, + 667 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Normal Estimation. We incorporate a normal attribute $\\pmb{n}$ for each 3D Gaussian and try to solve it. A potential methodology treats the spatial mean of a 3D Gaussian as a conventional point, and executes normal estimation based on the local planar assumption. However, this approach cannot provide accurate normal estimation. The reasons are two fold: First, the reconstructed Gaussian point cloud is often sparse. More critically, the Gaussian points are naturally soft, which means these points are not precisely aligned with the object surface.", + "bbox": [ + 212, + 671, + 787, + 777 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To address these limitations, we propose to optimize $\\pmb{n}$ from an initial random vector via back-propagation for each 3D Gaussian. We perceive the depth of all Gaussians along a ray as a distribution, and estimate the pixel depth as the expectation of this distribution. Similarly, we determine the normal for each", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "Relightable 3D Gaussians", + "bbox": [ + 558, + 114, + 732, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "pixel. This process is described by:", + "bbox": [ + 215, + 146, + 468, + 162 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\{\\mathcal {D}, \\mathcal {N} \\} = \\sum_ {i \\in N} w _ {i} \\left\\{d _ {i}, \\boldsymbol {n} _ {i} \\right\\}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 413, + 172, + 785, + 204 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $d_{i}$ , $\\pmb{n}_i$ and $w_{i} = T_{i}\\alpha_{i} / \\sum_{i\\in N}T_{i}\\alpha_{i}$ denote the depth, normal and weight of the point. We then encourage the consistency between the rendered normal $\\mathcal{N}$ and the pseudo normal $\\tilde{\\mathcal{N}}$ , which is computed from the rendered depth $\\mathcal{D}$ under the local planarity assumption. The normal consistency is quantified as follows:", + "bbox": [ + 215, + 215, + 787, + 277 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {n} = \\| \\mathcal {N} - \\tilde {\\mathcal {N}} \\| _ {2}. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 439, + 287, + 785, + 305 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Normal Gradient Based Densification. To achieve superior rendering fidelity in detail regions, vanilla 3DGS [26] densifies 3D Gaussians through the gradient of view space points. Drawing upon this, to improve normal recovery in thin regions, we introduce an additional densification criterion on the gradient of normals. Specifically, we densify Gaussians whose normal gradient exceeds a threshold $T_{n}$ .", + "bbox": [ + 215, + 316, + 782, + 405 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Constraint on Depth Distribution. Given our assumption of the object possessing an accurate surface, we enforce a constraint on the depth distribution by minimizing the uncertainty. The uncertainty is defined as:", + "bbox": [ + 215, + 407, + 782, + 452 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {u} = \\mathcal {D} _ {s q} - \\mathcal {D} ^ {2}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 446, + 467, + 784, + 483 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\mathcal{D}_{sq} = \\sum_{i\\in N}w_id_i^2$ , and $\\mathcal{D}$ is the depth estimation defined in Eq. 3. This constraint on the depth distribution drives Gaussian points to the object surface, thereby improving geometric reconstruction.", + "bbox": [ + 215, + 489, + 782, + 534 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Object Mask Constraint. If there is a mask indicating the object, we can constrain the optimization by a binary cross entropy loss [43]:", + "bbox": [ + 215, + 535, + 782, + 566 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {O} = - M \\log O - (1 - M) \\log (1 - O), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 362, + 580, + 784, + 595 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $M$ is the object mask and $O = \\sum_{i\\in N}T_i\\alpha_i$ . With this constraint, $O$ is forced to be aligned with the distribution of $M$ , and so we get opaque surface.", + "bbox": [ + 215, + 608, + 782, + 638 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.3 BRDF and Light Modeling", + "text_level": 1, + "bbox": [ + 215, + 661, + 485, + 678 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Rendering Equation. We use the rendering equation [24] to model light interaction with surfaces, accounting for their material properties and geometry. It is given by:", + "bbox": [ + 215, + 686, + 782, + 733 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nL _ {o} \\left(\\boldsymbol {\\omega} _ {o}, \\boldsymbol {x}\\right) = \\int_ {\\Omega} f \\left(\\boldsymbol {\\omega} _ {o}, \\boldsymbol {\\omega} _ {i}, \\boldsymbol {x}\\right) L _ {i} \\left(\\boldsymbol {\\omega} _ {i}, \\boldsymbol {x}\\right) \\left(\\boldsymbol {\\omega} _ {i} \\cdot \\boldsymbol {n}\\right) d \\boldsymbol {\\omega} _ {i}, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 326, + 743, + 784, + 773 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\pmb{x}$ and $\\pmb{n}$ are the surface point and its normal vector, $f$ is the Bidirectional Reflectance Distribution Function (BRDF), and $L_{i}$ and $L_{o}$ denote the incoming and outgoing radiance in directions $\\omega_{i}$ and $\\omega_{o}$ . $\\Omega$ signifies the hemispherical domain above the surface.", + "bbox": [ + 215, + 779, + 785, + 839 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "J. Gao et al.", + "bbox": [ + 271, + 114, + 356, + 127 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Prior methods [49, 51] typically begin with the acquisition of intersection points between rays and surface through differentiable rendering. Then they apply the rendering equation at these points to facilitate Physical Based Rendering (PBR). However, these approach present significant challenges in the point-based framework. Certainly, it is feasible to extract surfaces in 3DGS [17]; but, it necessitates a coordinate-based incident light field for inverse PBR. The querying of this field at millions of points during every iteration imposes a substantial computational burden. Moreover, it is pertinent to acknowledge that accurately extracting geometric surfaces presents considerable challenges in 3DGS system.", + "bbox": [ + 212, + 146, + 782, + 281 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To tackle this issue, we propose to compute PBR color $\\{c_i^{\\prime}\\}_{i = 0}^{N}$ for each 3D Gaussian, and then obtain the PBR image $\\mathcal{C}'$ through alpha-blending, as shown in Fig. 2. This method is more efficient for two main reasons. First, PBR is performed on fewer 3D Gaussians rather than image pixels in all input views, as each Gaussian affects multiple pixels based on its scale. Second, by allocating discrete attributes for each Gaussian, we avoid the need for global neural fields.", + "bbox": [ + 212, + 282, + 784, + 372 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "BRDF Parameterization. To make 3D Gaussians relightable, we assign BRDF properties to each Gaussian and adopt a simplified Disney BRDF model [11]. The BRDF properties include an albedo $\\pmb{b} \\in [0,1]^3$ and a roughness $r \\in [0,1]$ , and the BRDF $f$ in Eq. 7 is divided into a diffuse term $f_{d} = \\frac{b}{\\pi}$ and a specular term:", + "bbox": [ + 212, + 373, + 785, + 445 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\nf _ {s} \\left(\\boldsymbol {\\omega} _ {o}, \\boldsymbol {\\omega} _ {i}\\right) = \\frac {D (\\boldsymbol {h} ; r) \\cdot F \\left(\\boldsymbol {\\omega} _ {o} , \\boldsymbol {h}\\right) \\cdot G \\left(\\boldsymbol {\\omega} _ {i} , \\boldsymbol {\\omega} _ {o} , h ; r\\right)}{\\left(\\boldsymbol {n} \\cdot \\boldsymbol {\\omega} _ {i}\\right) \\cdot \\left(\\boldsymbol {n} \\cdot \\boldsymbol {\\omega} _ {o}\\right)}, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 330, + 445, + 785, + 477 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $\\pmb{h} = (\\pmb{\\omega}_i + \\pmb{\\omega}_o) / 2$ is the half vector, $D, F$ and $G$ denote the normal distribution function, Fresnel term and geometry term.", + "bbox": [ + 212, + 481, + 782, + 508 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Incident Light Modeling. Optimizing a NeILF for each Gaussian can be excessively unconstrained, leading to challenges in accurately decomposing incident lights from its appearance. To address this issue, we apply a prior by partitioning the incident light into globally shared direct components and individual per-Gaussian indirect components. The sampled incident light at a Gaussian from direction $\\omega_{i}$ is represented as:", + "bbox": [ + 212, + 512, + 782, + 602 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\nL _ {i} (\\boldsymbol {\\omega} _ {i}) = V (\\boldsymbol {\\omega} _ {i}) \\cdot L _ {\\text {d i r e c t}} (\\boldsymbol {\\omega} _ {i}) + L _ {\\text {i n d i r e c t}} (\\boldsymbol {\\omega} _ {i}), \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 344, + 614, + 784, + 630 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $V(\\omega_{i})$ represents the visibility term which will be further discussed in Sec. 4, the indirect light term $L_{\\text{indirect}}$ is parameterized by 3-level SH, denoted as $l$ , while the direct light term $L_{\\text{direct}}$ is parameterized as a globally shared 16x32 environment map, denoted as $l^{env}$ . Despite the utilization of relatively low-level Spherical Harmonics (SH), we can still capture some inter-reflection effects, as the rendering color of a given pixel arises from the collaboration of multiple relightable 3D Gaussians.", + "bbox": [ + 212, + 640, + 784, + 744 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "For each 3D Gaussian, we sample $N_{s}$ incident directions over the hemisphere space through Fibonacci sampling [45] to provide numerical integration for Eq. 7. The PBR color of each Gaussian is then given by:", + "bbox": [ + 212, + 746, + 782, + 792 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {c} ^ {\\prime} \\left(\\boldsymbol {\\omega} _ {o}\\right) = \\sum_ {i = 0} ^ {N _ {s}} \\left(f _ {d} + f _ {s} \\left(\\boldsymbol {\\omega} _ {o}, \\boldsymbol {\\omega} _ {i}\\right)\\right) L _ {i} \\left(\\boldsymbol {\\omega} _ {i}\\right) \\left(\\boldsymbol {\\omega} _ {i} \\cdot \\boldsymbol {n}\\right) \\Delta \\boldsymbol {\\omega} _ {i}, \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 326, + 801, + 785, + 842 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "Relightable 3D Gaussians", + "bbox": [ + 558, + 114, + 732, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 114, + 785, + 126 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $\\omega_{i}$ is the solid angle.", + "bbox": [ + 215, + 146, + 413, + 161 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To summarize, our method represents a 3D scene as a set of relightable 3D Gaussians and a global environment light $l^{env}$ , where the $i^{th}$ Gaussian $\\mathcal{P}_i$ is parameterized as $\\{\\pmb{\\mu}_i, \\pmb{q}_i, \\pmb{s}_i, o_i, \\pmb{c}_i, \\pmb{n}_i, \\pmb{b}_i, r_i, l_i\\}$ .", + "bbox": [ + 212, + 162, + 787, + 208 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "3.4 Regularizations", + "text_level": 1, + "bbox": [ + 215, + 229, + 390, + 244 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To mitigate the materials-lighting ambiguity [45], proper regularizations are utilized to facilitate their plausible decomposition.", + "bbox": [ + 212, + 255, + 782, + 285 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Light Regularization. We apply a light regularization assuming a near-natural white incident light [30]:", + "bbox": [ + 212, + 285, + 784, + 316 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {l i g h t}} = \\sum_ {c} \\left(L _ {c} - \\frac {1}{3} \\sum_ {c} L _ {c}\\right), c \\in \\{R, G, B \\}. \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 343, + 327, + 785, + 357 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Smoothness Priors. We expect that the BRDF properties do not change drastically in homogeneous areas [45]. We define a smooth constraint on roughness as:", + "bbox": [ + 214, + 364, + 782, + 407 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {s, r} = \\| \\nabla R \\| \\exp (- \\| \\nabla C _ {g t} \\|), \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 398, + 411, + 784, + 428 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $R = \\sum_{i\\in N}w_i r_i$ is the rendered roughness map. Similarly, we also define smoothness constraints $\\mathcal{L}_{s,n}$ and $\\mathcal{L}_{s,b}$ on normal and albedo.", + "bbox": [ + 212, + 433, + 782, + 465 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4 Point-based Ray Tracing", + "text_level": 1, + "bbox": [ + 215, + 487, + 491, + 506 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "For real-time realistic rendering with plausible shadow effects on relightable 3D Gaussians, we introduce a novel point-based ray tracing approach in this section.", + "bbox": [ + 214, + 520, + 782, + 550 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.1 Ray Tracing on 3D Gaussians", + "text_level": 1, + "bbox": [ + 215, + 571, + 506, + 588 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Our proposed ray tracing technique on 3D Gaussians is built upon the Bounding Volume Hierarchy (BVH), enabling efficient visibility querying along a ray. Our method adopts the idea from [25], an in-place algorithm for constructing a binary radix tree that maximizes parallelism and facilitates real-time BVH construction. Specifically, we construct a binary radix tree from a given set of 3D Gaussians, where each leaf node represents the tight bounding box of a Gaussian, and each internal node denotes the bounding box of its two children.", + "bbox": [ + 212, + 598, + 784, + 703 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Unlike ray tracing with opaque polygonal meshes, where only the closest intersection point is required, ray tracing with semi-transparent Gaussians necessitates accounting for all Gaussians potentially influencing the ray's transmittance. The process of ray tracing on 3D Gaussian points can be described as follows: the ray travels from the camera center and accumulates transmittance as it passes through 3D Gaussian points until the transmittance is zero. The first key issue to be addressed is how to define the contribution of a single 3D Gaussian to the transmittance. As discussed in Sec.3.1, transforming a 3D Gaussian to a 2D Gaussian involves an approximation, which complicating the", + "bbox": [ + 212, + 704, + 785, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "J. Gao et al.", + "bbox": [ + 271, + 114, + 357, + 127 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/7dd81d7ea48234217c81c0676ca8bebe9540cb52587a70d919390505ba9f1bda.jpg", + "image_caption": [ + "(a) Root Node" + ], + "image_footnote": [], + "bbox": [ + 215, + 148, + 352, + 292 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/1ea1a93651801d076c5cd91dc0dd289887767b15d82ac2bfaec1c3d2d36157c0.jpg", + "image_caption": [ + "(b) Child Node" + ], + "image_footnote": [], + "bbox": [ + 359, + 148, + 496, + 292 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/71c5c1a930d868cf6692fce5aa54ed6ad19c3852eea3f3ba3c145f0051dbcc82.jpg", + "image_caption": [ + "(c) Gaussian", + "Fig. 3: Intersection tests in point-based ray tracing. Intersection point between ray and Gaussian is obtained by three steps: (a) intersect the BVH root node; (b) dive into the intersected child nodes recursively until the leaf node; (c) perform Eq. 13 to get the equivalent intersection point. (d) shows that a 3D Gaussian actually has non-negligible effect on a segment of a ray." + ], + "image_footnote": [], + "bbox": [ + 506, + 148, + 640, + 292 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/f366c2a3d2b2fb588bf8a5819989d4f17b66aaf33f9e9f32e9464b3a32440977.jpg", + "image_caption": [ + "(d) Gaussian" + ], + "image_footnote": [], + "bbox": [ + 651, + 148, + 787, + 292 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "identification of accurate intersection between the Gaussian and a ray. As shown in Fig. 3d, a 3D Gaussian actually has non-negligible effect on a segment of a ray. Thus, drawing inspiration from [27], we approximate the intersection of ray with 3D Gaussian as a point where the 3D Gaussian's contribution peaks, as showed in Fig. 3(3). The equivalent intersection point is defined as:", + "bbox": [ + 212, + 417, + 782, + 494 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {r} _ {\\boldsymbol {x}} = \\boldsymbol {r} _ {\\boldsymbol {o}} + t _ {j} \\boldsymbol {r} _ {\\boldsymbol {d}}, \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 441, + 508, + 784, + 525 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where $\\pmb{r_o}$ denotes the origin, $\\pmb{r_d}$ is direction of the ray, corresponding to the previously mentioned $\\omega_{i}, t_{j}$ is defined as:", + "bbox": [ + 212, + 531, + 782, + 563 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\nt _ {j} = \\frac {\\left(\\boldsymbol {\\mu} - \\boldsymbol {r} _ {o}\\right) ^ {T} \\boldsymbol {\\Sigma} \\boldsymbol {r} _ {d}}{\\boldsymbol {r} _ {d} ^ {T} \\boldsymbol {\\Sigma} \\boldsymbol {r} _ {d}}. \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 423, + 573, + 784, + 608 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Subsequently, we approximate the contribution of this 3D Gaussian to the ray as $\\alpha_{j}$ at the equivalent intersection point $\\pmb{r}_{\\pmb{x}}$", + "bbox": [ + 212, + 616, + 782, + 646 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Considering the transmittance along a ray: $T_{i} = \\prod_{j=1}^{i-1}(1 - \\alpha_{j})$ , it is evident that the order of $\\alpha_{j}$ does not affect $T_{i}$ . In other words, the order in which Gaussians are encountered along a ray does not impact the overall transmittance. As illustrated in Fig. 3, starting from the root node of the binary radix tree, intersection tests are recursively performed between the ray and the bounding volumes of each node's children. Upon reaching a leaf node, the associated Gaussian is identified. Through this traversal, the transmittance $T$ is progressively attenuated:", + "bbox": [ + 212, + 648, + 787, + 768 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\nT _ {i} = \\left(1 - \\alpha_ {i - 1}\\right) T _ {i - 1}, \\quad \\text {f o r} i = 1, \\dots , j - 1 \\text {w i t h} T _ {1} = 1. \\tag {15}\n$$\n", + "text_format": "latex", + "bbox": [ + 300, + 781, + 784, + 797 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "To speed up ray tracing, the process is early terminated if a ray's transmittance drops below a certain threshold $T_{\\min}$ .", + "bbox": [ + 212, + 809, + 782, + 840 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "Relightable 3D Gaussians", + "bbox": [ + 558, + 114, + 732, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 114, + 785, + 126 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.2 Visibility Pre-computation", + "text_level": 1, + "bbox": [ + 215, + 146, + 480, + 161 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "In Sec. 3.3, a potential ambiguity in the decomposition of indirect and direct lights has been discerned. While the essential visibility term can be derived via the proposed ray tracing on 3D Gaussians (Sec. 4.1), querying the visibility term through ray tracing online proves challenging due to the computational complexity. Nevertheless, given our exclusive focus on static scenes, we can pre-compute the visibility term $V(\\omega_{i})$ across the hemispherical domain determined by $\\pmb{n}$ for each Gaussian, and subsequently integrate it into the rendering equation.", + "bbox": [ + 212, + 172, + 785, + 279 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Hence, we divide the optimization process into two stages. In the first stage, we optimize a vanilla 3D Gaussian point cloud, augmented with an additional normal vector $\\pmb{n}$ for each Gaussian. Immediately afterwards, we pre-compute the per-Gaussian visibility through the proposed ray tracing. In the second stage, we lock the geometry of 3D Gaussians and focus solely on optimizing the material and lighting parameters using the pipeline described in Fig. 2.", + "bbox": [ + 212, + 279, + 785, + 369 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.3 Point Based Relighting Pipeline", + "text_level": 1, + "bbox": [ + 215, + 393, + 526, + 410 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Based on our relightable 3D Gaussians, we devise a point-based graphics pipeline that seamlessly integrates effortless scene editing and realistic relighting. To our knowledge, there is currently no point-based rendering pipeline that effectively accomplishes both tasks. Although explicit point representations make scene editing easy, achieving photo-realistic relighting proved nearly insurmountable in existing point-based rendering pipeline. Conversely, in inverse rendering approaches based on implicit representations, while highly realistic relighting is feasible, scene editing presents a challenging endeavor.", + "bbox": [ + 212, + 421, + 785, + 542 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "As an illustration, we concentrate on relighting in a multi-object composition scene. Initially, our pipeline computes the visibility $V(\\omega_{i})$ for each Gaussian point via ray tracing (Sec. 4.1). Despite the challenging inter-object occlusions introduced by composition, our proposed point-based ray tracing method adeptly manages these complexities and ensures accurate occlusion updates within the novel scene. Subsequently, the rendering process unfolds (Sec. 3.3), starting with Gaussian-level PBR and ending with alpha blending. Throughout this process, the original indirect lighting of each object is discarded. Consequently, we achieve relighting with remarkably vivid shadow effects solely based on a discrete set of points, as illustrated in Fig. 1.", + "bbox": [ + 212, + 542, + 787, + 694 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "5 Experiments", + "text_level": 1, + "bbox": [ + 215, + 719, + 375, + 734 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "5.1 Training Details", + "text_level": 1, + "bbox": [ + 215, + 752, + 397, + 767 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "As pointed out in Sec. 4.2, the training procedure is divided into two stages. The first stage follows the setting of 3DGS [26], along with the proposed normal gradient-based densification (Sec. 3.2) where $T_{\\mathbf{n}}$ is set as $2 \\times 10^{-9}$ . In the second stage, we sample $N_{s} = 64$ incident rays per Gaussian point for PBR. We train", + "bbox": [ + 212, + 779, + 785, + 840 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "J. Gao et al.", + "bbox": [ + 271, + 114, + 354, + 126 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/dfc6f587526f1d16cd3d279c42c0818c98e03b01f67b16d14002c5e0fff82942.jpg", + "table_caption": [ + "Table 1: Quantitative results for novel view synthesis on NeRF synthetic dataset" + ], + "table_footnote": [], + "table_body": "
GeometryRelightablePSNR↑ SSIM↑ LPIPS
NPBG [1]point×28.100.9230.077
NPBG++ [38]point×28.120.9280.076
FreqPCR [54]point×31.240.9500.049
3DGS [26]point×33.880.9700.031
PhySG [51]neural✓18.910.8470.182
NeILF++ [49]neural✓26.370.9110.091
Nvdiffrec [33]mesh✓29.050.9390.081
R3DG(Ours)point✓31.220.9590.039
", + "bbox": [ + 331, + 171, + 669, + 292 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/691c40cc11a766ec29a9963794e6671aa4e0a4ef875e26e1e366231387f26dff.jpg", + "image_caption": [ + "(a) Lego" + ], + "image_footnote": [], + "bbox": [ + 228, + 303, + 395, + 446 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/d512835caae51e69be0bd4764e61a50ef532b082695459b4484b1db87d5a5b21.jpg", + "image_caption": [ + "(b) Chair" + ], + "image_footnote": [], + "bbox": [ + 423, + 303, + 573, + 446 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/5172a9fff6cd0cdfd0864a9e17ed364e479df8dd62e942fb360bd3a5a15597cb.jpg", + "image_caption": [ + "(c) Hotdog", + "Fig. 4: Visualizations on NeRF synthetic dataset [31]. Each scene is displayed in an order from left to right and from top to bottom: Ground Truth, PBR Image, Normal Map and Ambient Occlusion Map." + ], + "image_footnote": [], + "bbox": [ + 593, + 313, + 781, + 439 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "our model for 30,000 iterations in the initial stage and 10,000 iterations in the second stage. The loss in the first stage is calculated by:", + "bbox": [ + 212, + 544, + 784, + 575 + ], + "page_idx": 10 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\lambda_ {1} \\mathcal {L} _ {1} + \\lambda_ {s s i m} \\mathcal {L} _ {s s i m} + \\lambda_ {n} \\mathcal {L} _ {n} + \\lambda_ {s, n} \\mathcal {L} _ {s, n} + \\lambda_ {O} \\mathcal {L} _ {O} + \\lambda_ {u} \\mathcal {L} _ {u}, \\tag {16}\n$$\n", + "text_format": "latex", + "bbox": [ + 282, + 585, + 784, + 603 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "where $\\{\\lambda_1, \\lambda_{ssim}, \\lambda_n, \\lambda_{s,n}, \\lambda_O, \\lambda_u\\}$ are set to $\\{0.8, 0.2, 0.01, 0.01, 0.01, 0.01\\}$ , respectively. The loss in the second stage is given by:", + "bbox": [ + 212, + 614, + 785, + 646 + ], + "page_idx": 10 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\lambda_ {1} \\mathcal {L} _ {1} + \\lambda_ {s s i m} \\mathcal {L} _ {s s i m} + \\lambda_ {l} \\mathcal {L} _ {l} + \\lambda_ {s, b} \\mathcal {L} _ {s, b} + \\lambda_ {s, r} \\mathcal {L} _ {s, r}, \\tag {17}\n$$\n", + "text_format": "latex", + "bbox": [ + 310, + 657, + 784, + 674 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "where $\\{\\lambda_1,\\lambda_{ssim},\\lambda_l,\\lambda_{s,b},\\lambda_{s,r}\\}$ are set to $\\{0.8,0.2,0.0001,0.01,0.01\\}$ , respectively.", + "bbox": [ + 212, + 685, + 787, + 702 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "5.2 Performance on Novel View Synthesis", + "text_level": 1, + "bbox": [ + 214, + 723, + 576, + 739 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "We first evaluate the novel view synthesis (NVS) of the physically based rendering (PBR) on NeRF synthetic dataset [31]. We compare both point-based rendering methods and relightable rendering approaches. We report average metrics across all scenes in Tab. 1, including peak signal-to-noise ratio (PSNR), structural similarity index measure (SSIM), and learned perceptual image patch similarity (LPIPS). Compared with existing point-based rendering methods, our", + "bbox": [ + 212, + 750, + 787, + 840 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "Relightable 3D Gaussians", + "bbox": [ + 558, + 114, + 732, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/259d2722551fd50cceb1ade01cfc6d371ded05171ee9760515202be7a47bd4d7.jpg", + "table_caption": [ + "Table 2: Quantitative results on Synthetic4Relight dataset [56]" + ], + "table_footnote": [], + "table_body": "
Novel View SynthesisRelightingAlbedoRoughnessTime
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓MSE↓(hours)
NerFactor [53]22.800.9160.15021.540.8750.17119.490.8640.206N/A>48
Nvdifferc-MC [19]34.290.9670.06824.220.9430.07829.610.9450.0750.0094.17
InvRender [56]30.740.9530.08628.670.9500.09128.280.9350.0720.00814.3
TensorIR [23]35.800.9780.04929.690.9510.07930.580.9460.0650.0153.24
R3DG (Ours)36.800.9820.02831.000.9640.05028.310.9510.0580.0130.90
", + "bbox": [ + 218, + 171, + 784, + 256 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/eda3e19a9df0d153ec03f03ce255213cb4e3de1751b555409800b287014933c5.jpg", + "image_caption": [ + "Fig. 5: Qualitative results on Synthetic4Relight dataset [56]." + ], + "image_footnote": [], + "bbox": [ + 218, + 262, + 781, + 561 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "R3DG demonstrates superiority over most of them [1,38,54]. While our performance slightly marginally trails that of vanilla 3DGS [26], it is noteworthy that our approach excels in relighting capability, presenting a significant advantage. Furthermore, in comparison to other relightable methods [33,49,51], our method showcases significantly better NVS quality.", + "bbox": [ + 212, + 614, + 787, + 690 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "In Fig. 4, we visualize the our reconstruction results in NeRF synthetic dataset, including the PBR image, the normal map, and the pre-computed visibility illustrated in the form of the Ambient Occlusion (AO) map. As depicted in Fig. 4, our approach successfully achieves realistic PBR rendering, smooth normal estimation, and accurate visibility solving on discrete point clouds.", + "bbox": [ + 212, + 691, + 787, + 767 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "5.3 Performance on Relighting", + "text_level": 1, + "bbox": [ + 215, + 786, + 483, + 801 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "To comprehensively assess the relighting capabilities of our pipeline, we further undertook experiments using the Synthetic4Relight dataset [56]. Recognizing", + "bbox": [ + 212, + 809, + 785, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "J. Gao et al.", + "bbox": [ + 271, + 114, + 357, + 126 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/fec47af15235fbfd3b6c09aae284aa837196eee7701c35dff80bb8a743ccc7dd.jpg", + "image_caption": [ + "Fig. 6: Qualitative results of relighting on real-world scenes." + ], + "image_footnote": [], + "bbox": [ + 222, + 143, + 781, + 372 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "the inherent scale ambiguity between the estimated albedo and lighting, we standardized the scale of the estimated albedo against the ground truth to relighting, consistent with previous studies [51, 56].", + "bbox": [ + 212, + 429, + 782, + 474 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Our evaluation encompasses the analysis of decomposed materials, the synthesis of novel views, and the relighting outcomes. The quantitative analysis is presented in Tab. 2. Our method outperforms existing approaches in terms of Novel View Synthesis (NVS) and relighting precision. Regarding material estimation, our method exhibits superior albedo accuracy in SSIM and LPIPS. Qualitatively, our method achieves visually pleasing material decomposition, facilitating a realistic relighting effect, which is shown in 5.", + "bbox": [ + 212, + 474, + 784, + 580 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Additionally, we conduct relighting on real-world scenes in the Mip-NeRF 360 dataset [4], as shown in Fig. 6. It is worth noting that our relighting in the kitchen with the second environment map, exhibits pleasing shadow effects.", + "bbox": [ + 212, + 580, + 784, + 626 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "5.4 Ablation Study", + "text_level": 1, + "bbox": [ + 215, + 648, + 387, + 664 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "We conduct ablation studies on three principal components of our method that significantly influence the quality of inverse rendering. The results are illustrated in Fig. 7. Initially, we investigate the impact of normal gradient densification (NGD). Fig. 7 (a, b) depict the rendering normal maps with and without the densification. It shows that the proposed densification markedly enhances the details of the normal map for *Lego*. In Fig. 7 (c, d, e), we examine the effects of our full light modeling. We simplify the full model to a global environment map in the optimization. The result demonstrates that the simplified model fails to generate shadow effects when relighting. Furthermore, we evaluate the significance of the proposed constraint on depth distribution, denoted as $\\mathcal{L}_u$ . The depth uncertainty is visualized in Fig. 7 (h, i), which illustrates the distribution", + "bbox": [ + 212, + 672, + 787, + 840 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "Relightable 3D Gaussians", + "bbox": [ + 558, + 114, + 730, + 127 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 784, + 126 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/5b3b21ad98f3987eb60bbdce21c626f4daf0bf4745a4fabeed8200109a7888a4.jpg", + "image_caption": [ + "(a) w NGD." + ], + "image_footnote": [], + "bbox": [ + 217, + 148, + 320, + 228 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/6523353d9050e62fb8dc52cc480729a976107cbc8ccc0568fd9988778d08578a.jpg", + "image_caption": [ + "(b) w/o NGD." + ], + "image_footnote": [], + "bbox": [ + 330, + 148, + 431, + 228 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/5758f5d0b48f1ef7bf0d2582aadb03b628d0e685558ec0a66b0068add60c40f0.jpg", + "image_caption": [ + "(c) Relit. GT." + ], + "image_footnote": [], + "bbox": [ + 442, + 143, + 552, + 229 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/96deec950c0a1ce91651e02ca8ca1e20ae6b11f6c7e4cc0373e465b782a38585.jpg", + "image_caption": [ + "(d) Full light." + ], + "image_footnote": [], + "bbox": [ + 553, + 143, + 663, + 229 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/ff8d7021dea44ee962f1e8c41b78f3df0d10a159cf53e5904436ca8b072ecbd7.jpg", + "image_caption": [ + "(e) Simplified." + ], + "image_footnote": [], + "bbox": [ + 666, + 143, + 777, + 229 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/690f3d3716c469c06a0ee3cde99eb678c9216a4f1d55251ed51d3d4f0e96bec3.jpg", + "image_caption": [ + "(f) AO w $\\mathcal{L}_u$" + ], + "image_footnote": [], + "bbox": [ + 222, + 244, + 349, + 286 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/71cbe71748db2e43c57dbc9a9120d19c0f11c1cbff4d6c0a56c63cf57c92e180.jpg", + "image_caption": [ + "(g) AO w/o $\\mathcal{L}_u$" + ], + "image_footnote": [], + "bbox": [ + 361, + 244, + 491, + 287 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/2b3b550b97d69ba29525029c7a3c2296b1502eab4d62adcb05b70376006e012f.jpg", + "image_caption": [ + "(h) Uncertainty w $\\mathcal{L}_u$", + "Fig. 7: Ablation studies on main components of our method. (a) and (b) show normal maps with and without the proposed normal gradient densification (NGD). (d) and (e) display re-lighting results using the proposed full lighting model and only a simplified light modeling, i.e. a global environment map. (f) and (g) visualize ambient occlusion (AO) maps with and without the proposed constraint on depth distribution $\\mathcal{L}_u$ . (h) and (i) illustrate the depth uncertainty with and without $\\mathcal{L}_u$ ." + ], + "image_footnote": [], + "bbox": [ + 498, + 242, + 635, + 290 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/fa38bf3880ce90a9172ea6a6553efbb351ceff1352f95bee43ca8e6e386d8890.jpg", + "image_caption": [ + "(i) Uncertainty w/o $\\mathcal{L}_u$" + ], + "image_footnote": [], + "bbox": [ + 638, + 242, + 777, + 290 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "constraint contribute significantly to reducing the depth uncertainty. Additionally, Fig. 7 (f, g) reveals that incorporating $\\mathcal{L}_u$ leads to more plausible ambient occlusion (averaged visibility) which plays a crucial role in stage 2.", + "bbox": [ + 214, + 434, + 784, + 481 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "6 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 518, + 357, + 534 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In this paper, we have introduced a differentiable point-based rendering pipeline enabling effort less editing, accuracy ray tracing, and photo-realistic relighting. We present the scene as Relightable 3D Gaussians, an extension of traditional 3D Gaussians enriched with supplementary normals, BRDFs and indirect lighting. To reconstruct a relightable scene from multi-view images, we build a novel differentiable rendering pipeline based on the inverse rendering techniques and 3D Gaussian Splatting. Additionally, we introduce an innovative ray tracing scheme specifically designed for scenes represented by discrete points, providing precise visibility for realistic relighting. Quantitative and qualitative experiments confirm that our pipeline successfully reconstructs reasonable normals, materials for discrete point clouds, and achieves commendable accuracy in both novel view synthesis and scene relighting tasks.", + "bbox": [ + 212, + 564, + 787, + 746 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Limitations and Future work. Our present pipeline targets the reconstruction of static objects. Certain design choices pose challenges in maintaining its performance with large scale scenes. The high density of points in large scene can slow down optimization as we sample rays and perform PBR at each point. This can be mitigated through deferred rendering technique. Regarding geometry, integrating MVS shows promise for achieving more accurate representations.", + "bbox": [ + 212, + 750, + 787, + 840 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "J. Gao et al.", + "bbox": [ + 271, + 114, + 357, + 126 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 217, + 143, + 401, + 162 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "This work was supported by National Key R&D Program of China (2023YFB3209702), and NSFC (62441204).", + "bbox": [ + 215, + 172, + 825, + 204 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 224, + 321, + 239 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Aliev, K.A., Sevastopolsky, A., Kolos, M., Ulyanov, D., Lempitsky, V.: Neural point-based graphics. In: ECCV (2020)", + "2. Azinovic, D., Li, T.M., Kaplanyan, A., Nießner, M.: Inverse path tracing for joint material and lighting estimation. In: CVPR (2019)", + "3. Barron, J.T., Mildenhall, B., Tancik, M., Hedman, P., Martin-Brualla, R., Srinivasan, P.P.: Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In: ICCV (2021)", + "4. Barron, J.T., Mildenhall, B., Verbin, D., Srinivasan, P.P., Hedman, P.: Mip-nerf 360: Unbounded anti-aliased neural radiance fields. In: CVPR (2022)", + "5. Barron, J.T., Mildenhall, B., Verbin, D., Srinivasan, P.P., Hedman, P.: Zip-nerf: Anti-aliased grid-based neural radiance fields. arXiv preprint (2023)", + "6. Bi, S., Xu, Z., Srinivasan, P., Mildenhall, B., Sunkavalli, K., Hašan, M., Hold-Geoffroy, Y., Kriegman, D., Ramamoorthi, R.: Neural reflectance fields for appearance acquisition. arXiv preprint (2020)", + "7. Bi, S., Xu, Z., Sunkavalli, K., Hašan, M., Hold-Geoffroy, Y., Kriegman, D., Ramamoorthi, R.: Deep reflectance volumes: Relightable reconstructions from multiview photometric images. In: ECCV (2020)", + "8. Bi, S., Xu, Z., Sunkavalli, K., Kriegman, D., Ramamoorthi, R.: Deep 3d capture: Geometry and reflectance from sparse multi-view images. In: CVPR (2020)", + "9. Boss, M., Braun, R., Jampani, V., Barron, J.T., Liu, C., Lensch, H.: Nerd: Neural reflectance decomposition from image collections. In: ICCV (2021)", + "0. Boss, M., Engelhardt, A., Kar, A., Li, Y., Sun, D., Barron, J., Lensch, H., Jampani, V.: Samurai: Shape and material from unconstrained real-world arbitrary image collections. In: NeurIPS (2022)", + "1. Burley, B., Studios, W.D.A.: Physically-based shading at disney. In: SIGGRAPH (2012)", + "2. Chen, A., Xu, Z., Geiger, A., Yu, J., Su, H.: Tensorf: Tensorial radiance fields. In: ECCV (2022)", + "3. Chen, Z., Funkhouser, T., Hedman, P., Tagliasacchi, A.: Mobilenerf: Exploiting the polygon rasterization pipeline for efficient neural field rendering on mobile architectures. In: CVPR (2023)", + "4. Deng, K., Liu, A., Zhu, J.Y., Ramanan, D.: Depth-supervised nef: Fewer views and faster training for free. In: CVPR (2022)", + "5. Fridovich-Keil, S., Yu, A., Tancik, M., Chen, Q., Recht, B., Kanazawa, A.: Plenoxels: Radiance fields without neural networks. In: CVPR (2022)", + "6. Garbin, S.J., Kowalski, M., Johnson, M., Shotton, J., Valentin, J.: Fastnerf: High-fidelity neural rendering at 200fps. In: ICCV (2021)", + "7. Guédon, A., Lepetit, V.: Sugar: Surface-aligned gaussian splatting for efficient 3d mesh reconstruction and high-quality mesh rendering. arXiv preprint arXiv:2311.12775 (2023)", + "8. Guo, K., Lincoln, P., Davidson, P., Busch, J., Yu, X., Whalen, M., Harvey, G., Orts-Escalano, S., Pandey, R., Dourgarian, J., et al.: The relighttables: Volumetric performance capture of humans with realistic relighting. ACM TOG (2019)" + ], + "bbox": [ + 225, + 253, + 785, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "Relightable 3D Gaussians", + "bbox": [ + 558, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 785, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "19. Hasselgren, J., Hofmann, N., Munkberg, J.: Shape, light, and material decomposition from images using monte carlo rendering and denoising. In: NeurIPS (2022)", + "20. Hasselgren, J., Hofmann, N., Munkberg, J.: Shape, Light, and Material Decomposition from Images using Monte Carlo Rendering and Denoising. arXiv:2206.03380 (2022)", + "21. Hedman, P., Srinivasan, P.P., Mildenhall, B., Barron, J.T., Debevec, P.: Baking neural radiance fields for real-time view synthesis. In: ICCV (2021)", + "22. Hu, T., Liu, S., Chen, Y., Shen, T., Jia, J.: Efficientnerf efficient neural radiance fields. In: CVPR (2022)", + "23. Jin, H., Liu, I., Xu, P., Zhang, X., Han, S., Bi, S., Zhou, X., Xu, Z., Su, H.: Tensoir: Tensorial inverse rendering. In: CVPR (2023)", + "24. Kajiya, J.T.: The rendering equation. In: Proceedings of the 13th annual conference on Computer graphics and interactive techniques (1986)", + "25. Karras, T.: Maximizing parallelism in the construction of bvhs, octrees, and k-d trees. In: Proceedings of the Fourth ACM SIGGRAPH/Eurographics conference on High-Performance Graphics (2012)", + "26. Kerbl, B., Kopanas, G., Leimkuhler, T., Drettakis, G.: 3d gaussian splatting for real-time radiance field rendering. ACM TOG (2023)", + "27. Keselman, L., Hebert, M.: Flexible techniques for differentiable rendering with 3d gaussians. arXiv preprint (2023)", + "28. Kuang, Z., Olszewski, K., Chai, M., Huang, Z., Achlioptas, P., Tulyakov, S.: Neroic: Neural rendering of objects from online image collections. ACM TOG (2022)", + "29. Levoy, M., Whitted, T.: The use of points as a display primitive (1985)", + "30. Liu, Y., Wang, P., Lin, C., Long, X., Wang, J., Liu, L., Komura, T., Wang, W.: Nano: Neural geometry and brdf reconstruction of reflective objects from multiview images. In: SIGGRAPH (2023)", + "31. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. In: ECCV (2020)", + "32. Müller, T., Evans, A., Schied, C., Keller, A.: Instant neural graphics primitives with a multiresolution hash encoding. ACM TOG (2022)", + "33. Munkberg, J., Hasselgren, J., Shen, T., Gao, J., Chen, W., Evans, A., Müller, T., Fidler, S.: Extracting triangular 3d models, materials, and lighting from images. In: CVPR (2022)", + "34. Nam, G., Lee, J.H., Gutierrez, D., Kim, M.H.: Practical svbrdf acquisition of 3d objects with unstructured flash photography. ACM TOG (2018)", + "35. Park, J.J., Holynski, A., Seitz, S.M.: Seeing the world in a bag of chips. In: CVPR (2020)", + "36. Pfister, H., Zwicker, M., Van Baar, J., Gross, M.: Surfels: Surface elements as rendering primitives. In: Proceedings of the 27th annual conference on Computer graphics and interactive techniques (2000)", + "37. Pittaluga, F., Koppal, S.J., Kang, S.B., Sinha, S.N.: Revealing scenes by inverting structure from motion reconstructions. In: CVPR (2019)", + "38. Rakhimov, R., Ardelean, A.T., Lempitsky, V., Burnaev, E.: Npbg++: Accelerating neural point-based graphics. In: CVPR (2022)", + "39. Reiser, C., Peng, S., Liao, Y., Geiger, A.: Kilonerf: Speeding up neural radiance fields with thousands of tiny mlps. In: ICCV (2021)", + "40. Schmitt, C., Donne, S., Riegler, G., Koltun, V., Geiger, A.: On joint estimation of pose, geometry and svbrdf from a handheld scanner. In: CVPR (2020)" + ], + "bbox": [ + 215, + 147, + 784, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "J. Gao et al.", + "bbox": [ + 271, + 114, + 354, + 126 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "41. Srinivasan, P.P., Deng, B., Zhang, X., Tancik, M., Mildenhall, B., Barron, J.T.: Nerv: Neural reflectance and visibility fields for relighting and view synthesis. In: CVPR (2021)", + "42. Verbin, D., Hedman, P., Mildenhall, B., Zickler, T., Barron, J.T., Srinivasan, P.P.: Ref-nerf: Structured view-dependent appearance for neural radiance fields. In: CVPR (2022)", + "43. Wang, P., Liu, L., Liu, Y., Theobalt, C., Komura, T., Wang, W.: Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. arXiv preprint (2021)", + "44. Wu, H., Hu, Z., Li, L., Zhang, Y., Fan, C., Yu, X.: Nefii: Inverse rendering for reflectance decomposition with near-field indirect illumination. In: CVPR (2023)", + "45. Yao, Y., Zhang, J., Liu, J., Qu, Y., Fang, T., McKinnon, D., Tsin, Y., Quan, L.: Neilf: Neural incident light field for physically-based material estimation. In: ECCV (2022)", + "46. Yariv, L., Gu, J., Kasten, Y., Lipman, Y.: Volume rendering of neural implicit surfaces. In: NeurIPS (2021)", + "47. Yifan, W., Serena, F., Wu, S., Öztireli, C., Sorkine-Hornung, O.: Differentiable surface splatting for point-based geometry processing. ACM TOG (2019)", + "48. Yu, Z., Peng, S., Niemeyer, M., Sattler, T., Geiger, A.: Monosdf: Exploring monocular geometric cues for neural implicit surface reconstruction. In: NeurIPS (2022)", + "49. Zhang, J., Yao, Y., Li, S., Liu, J., Fang, T., McKinnon, D., Tsin, Y., Quan, L.: Neil++: Inter-reflectable light fields for geometry and material estimation. In: ICCV (2023)", + "50. Zhang, K., Luan, F., Li, Z., Snavely, N.: Iron: Inverse rendering by optimizing neural sdfs and materials from photometric images. In: CVPR (2022)", + "51. Zhang, K., Luan, F., Wang, Q., Bala, K., Snavely, N.: Physg: Inverse rendering with spherical gaussians for physics-based material editing and relighting. In: CVPR (2021)", + "52. Zhang, Q., Baek, S.H., Rusinkiewicz, S., Heide, F.: Differentiable point-based radiance fields for efficient view synthesis. In: SIGGRAPH Asia (2022)", + "53. Zhang, X., Srinivasan, P.P., Deng, B., Debevec, P., Freeman, W.T., Barron, J.T.: Nerfactor: Neural factorization of shape and reflectance under an unknown illumination. ACM TOG (2021)", + "54. Zhang, Y., Huang, X., Ni, B., Li, T., Zhang, W.: Frequency-modulated point cloud rendering with easy editing. In: CVPR (2023)", + "55. Zhang, Y., Xu, T., Yu, J., Ye, Y., Jing, Y., Wang, J., Yu, J., Yang, W.: Nemf: Inverse volume rendering with neural microflake field. In: ICCV (2023)", + "56. Zhang, Y., Sun, J., He, X., Fu, H., Jia, R., Zhou, X.: Modeling indirect illumination for inverse rendering. In: CVPR (2022)", + "57. Zwicker, M., Pfister, H., Van Baar, J., Gross, M.: Surface splatting. In: Proceedings of the 28th annual conference on computer graphics and interactive techniques (2001)", + "58. Zwicker, M., Pfister, H., Van Baar, J., Gross, M.: Ewa splatting. IEEE TVCG (2002)" + ], + "bbox": [ + 212, + 146, + 787, + 756 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "Relightable 3D Gaussians", + "bbox": [ + 558, + 114, + 730, + 127 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + } +] \ No newline at end of file diff --git a/2024/Relightable 3D Gaussians_ Realistic Point Cloud Relighting with BRDF Decomposition and Ray Tracing/8f8cf818-147d-408c-857b-28f5e4e05e99_model.json b/2024/Relightable 3D Gaussians_ Realistic Point Cloud Relighting with BRDF Decomposition and Ray Tracing/8f8cf818-147d-408c-857b-28f5e4e05e99_model.json new file mode 100644 index 0000000000000000000000000000000000000000..80e62181a0abb6cf12ff1c23fb5b62e4e4d6abd0 --- /dev/null +++ b/2024/Relightable 3D Gaussians_ Realistic Point Cloud Relighting with BRDF Decomposition and Ray Tracing/8f8cf818-147d-408c-857b-28f5e4e05e99_model.json @@ -0,0 +1,2918 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.227, + 0.141, + 0.78, + 0.21 + ], + "angle": 0, + "content": "Relightable 3D Gaussians: Realistic Point Cloud Relighting with BRDF Decomposition and Ray Tracing" + }, + { + "type": "text", + "bbox": [ + 0.238, + 0.234, + 0.765, + 0.266 + ], + "angle": 0, + "content": "Jian \\(\\mathrm{Gao}^{1*}\\), Chun \\(\\mathrm{Gu}^{2*}\\), Youtian Lin\\(^{1}\\), Zhihao Li\\(^{3}\\), Hao Zhu\\(^{1}\\), Xun \\(\\mathrm{Cao}^{1}\\), Li Zhang\\(^{2\\boxtimes}\\), and Yao \\(\\mathrm{Yao}^{1\\boxtimes}\\)" + }, + { + "type": "text", + "bbox": [ + 0.432, + 0.277, + 0.573, + 0.291 + ], + "angle": 0, + "content": "\\(^{1}\\) Nanjing University" + }, + { + "type": "text", + "bbox": [ + 0.434, + 0.292, + 0.568, + 0.305 + ], + "angle": 0, + "content": "\\(^{2}\\) Fudan University" + }, + { + "type": "text", + "bbox": [ + 0.415, + 0.306, + 0.59, + 0.319 + ], + "angle": 0, + "content": "3 Huawei Noah's Ark Lab" + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.358, + 0.744, + 0.567 + ], + "angle": 0, + "content": "Abstract. In this paper, we present a novel differentiable point-based rendering framework to achieve photo-realistic relighting. To make the reconstructed scene relightable, we enhance vanilla 3D Gaussians by associating extra properties, including normal vectors, BRDF parameters, and incident lighting from various directions. From a collection of multiview images, the 3D scene is optimized through 3D Gaussian Splatting while BRDF and lighting are decomposed by physically based differentiable rendering. To produce plausible shadow effects in photo-realistic relighting, we introduce an innovative point-based ray tracing with the bounding volume hierarchies for efficient visibility pre-computation. Extensive experiments demonstrate our improved BRDF estimation, novel view synthesis and relighting results compared to state-of-the-art approaches. The proposed framework showcases the potential to revolutionize the mesh-based graphics pipeline with a point-based pipeline enabling editing, tracing, and relighting." + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.579, + 0.735, + 0.594 + ], + "angle": 0, + "content": "Keywords: 3D Gaussian Splatting, Relighting, Point based rendering" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.619, + 0.377, + 0.635 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.651, + 0.788, + 0.818 + ], + "angle": 0, + "content": "Reconstructing 3D scenes from multi-view images for photo-realistic rendering is a fundamental problem at the intersection of computer vision and graphics. Recently, 3D Gaussian Splatting (3DGS) [26] has been proposed and has gained significant attention from the community. The method employs a set of 3D Gaussian points to represent a 3D scene and projects these points onto a designated view through a tile-based rasterization. Attributes of each 3D Gaussian point are then optimized through the point-based differentiable rendering. Notably, 3DGS achieves real-time rendering with quality comparable to or even surpassing previous state-of-the-art methods (e.g., Mip-NeRF [3]), with a training speed on par with the most efficient Instant-NGP [32]. However, the current 3DGS is unable to reconstruct a scene that can be relighted under different lighting conditions," + }, + { + "type": "page_footnote", + "bbox": [ + 0.233, + 0.826, + 0.381, + 0.841 + ], + "angle": 0, + "content": "*Equally contributed." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.358, + 0.127 + ], + "angle": 0, + "content": "J. Gao et al." + }, + { + "type": "image", + "bbox": [ + 0.222, + 0.145, + 0.501, + 0.23 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.315, + 0.232, + 0.409, + 0.242 + ], + "angle": 0, + "content": "(a) Point Cloud" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.144, + 0.783, + 0.23 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.609, + 0.232, + 0.677, + 0.242 + ], + "angle": 0, + "content": "(b) Normal" + }, + { + "type": "image", + "bbox": [ + 0.222, + 0.244, + 0.501, + 0.327 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.295, + 0.33, + 0.428, + 0.342 + ], + "angle": 0, + "content": "(c) Ambient Occlusion" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.244, + 0.783, + 0.328 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.55, + 0.33, + 0.736, + 0.342 + ], + "angle": 0, + "content": "(d) Physically Based Rendering" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.353, + 0.788, + 0.451 + ], + "angle": 0, + "content": "Fig. 1: Visual results of our pipeline on a multi-object composition scene. In our pipeline, we represent a scene as Relightable 3D Gaussians. From multi-view images, we recover the geometry and materials of individual objects with inverse rendering techniques (see Sec. 3). Then, objects are easily composited into a new scene, thanks to our explicit representation. After that, we solve the complex occlusions though point based ray tracing (see Sec. 4) and re-light the new scene. Ultimately, we achieve high-fidelity relighting with remarkably realistic shadow." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.499, + 0.785, + 0.56 + ], + "angle": 0, + "content": "making the method only applicable to the task of novel view synthesis. In addition, ray tracing, a crucial component in achieving realistic rendering, remains an unresolved challenge in point-based representation, limiting 3DGS to more dedicated rendering effects such as shadowing and light reflectance." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.569, + 0.788, + 0.841 + ], + "angle": 0, + "content": "In this paper, we aim to reconstruct a re烈table 3D Gaussian point cloud from multi-view images with a differentiable rendering framework and achieve realistic relighting. We make 3D Gaussian relightable by assigning it with additional normal, BRDF properties, and incident light information to model perpoint light reflectance. In contrast to the plain alpha blending in the original 3DGS [26], we apply physically based rendering (PBR) to get a PBR color for each 3D Gaussian point, which is then alpha-composed to obtain a rendered color for the corresponding image pixel. For robust material and lighting estimation, we split the incident light into a global environment map and an indirect incident light field. To capture accurate visibility for each 3D Gaussian, we propose a novel ray tracing method based on the bounding volume hierarchy (BVH), which enables efficient visibility pre-computing for real-time rendering. Additionally, the proposed ray tracing method can handle complex occlusion relationships in a novel multi-object composition scene, thus realizing realistic shadow effects. Moreover, proper regularizations are introduced to enhance the geometry and mitigate the material-lighting ambiguity during the optimization, including constraints on depth distribution, smoothness priors, and a lighting regularization." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.56, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Relightable 3D Gaussians" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.222 + ], + "angle": 0, + "content": "Extensive experiments conducted across diverse datasets demonstrate the improved material and lighting estimation results and novel view rendering quality. Additionally, we illustrate the relightable and editable capabilities of our system through multi-object composition in a novel lighting environment. To summarize, major contributions of the paper include:" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.231, + 0.784, + 0.275 + ], + "angle": 0, + "content": "- We propose a material and lighting decomposition scheme for 3D Gaussian Splatting, where a normal vector, BRDF values, and incident lights are assigned and optimized for each 3D Gaussian point." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.276, + 0.784, + 0.319 + ], + "angle": 0, + "content": "- We introduce a novel point-based ray tracing approach based on bounding volume hierarchy, enabling efficient visibility pre-computing for each 3D Gaussian point and rendering of a 3D scene with realistic shadow effect." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.321, + 0.784, + 0.364 + ], + "angle": 0, + "content": "- We demonstrate a comprehensive graphics pipeline solely based on a discretized point representation, supporting relighting, editing, and ray tracing of a reconstructed 3D point cloud." + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.231, + 0.784, + 0.364 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.387, + 0.396, + 0.403 + ], + "angle": 0, + "content": "2 Related Works" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.417, + 0.785, + 0.553 + ], + "angle": 0, + "content": "Neural Radiance Field. Differentiable rendering techniques, exemplified by Neural Radiance Field (NeRF) [31], have garnered significant attention [3,12,32]. NeRF utilizes an implicit Multi-Layer Perceptron (MLP) that takes 3D positions and viewing directions as inputs to generate density and view-dependent colors for differentiable volume rendering. Despite its powerful neural implicit representation, both speed and quality can be improved. Efficiency improvements mainly focus on optimizing queries on neural fields and minimizing queries [12, 13, 15, 16, 21, 22, 32, 39]. Quality enhancements involve anti-aliasing [3-5] or utilizing exterior supervision [14, 48], etc." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.554, + 0.785, + 0.749 + ], + "angle": 0, + "content": "Differentiable Point Based Rendering. Using points as rendering primitives was first proposed in [29]. To address the issue of resulting holey images when rasterizing discrete points directly, solutions fall into two categories. One approach directly employs points as rendering primitives and encodes the geometric and photometric features near the point using SIFT descriptors [37] or neural descriptors [1,38]. This leads to a feature image with gaps, which is then decoded to recover a hole-free RGB image. The other category models a point as a primitive occupying a specific space, like a surfel [36,47] or a 3D Gaussian [26]. The rendered image is then generated using a splatting technique. This technique saw significant development in the 2000s [36,57,58], and in the era of differentiable rendering, PointRF [52], DSS [47] and 3DGS [26] serve as notable examples. We assert that the point serves as a promising rendering primitive owing to its inherent ease of editing, as also corroborated in [54]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.784, + 0.84 + ], + "angle": 0, + "content": "Material and Lighting Estimation. Decomposing the materials and illumination of a scene from multi-view images presents a formidable challenge owing to its intrinsic high-dimensional complexity. Some methods simplified the problem under the assumption of a controllable light [2,6-8,18,34,35,40]. Subsequent research explores more complex lighting models to cope with realistic scenarios. NeRV [41] and PhySG [51] leverage an environmental map to manage arbitrary" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.358, + 0.128 + ], + "angle": 0, + "content": "J. Gao et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.146, + 0.788, + 0.451 + ], + "angle": 0, + "content": "lighting conditions. NeRD [9] addresses the challenge posed by images captured under varying illumination by attributing Spherical Gaussians to each image. IRON [50] introduces an innovative edge sampling algorithm tailored for neural SDFs. NeRFactor [53] exploits light visibility to achieve superior material and lighting decomposition. Further studies [19, 52] address the consideration of indirect lighting, leading to enhanced BRDF estimation quality. [20, 33] utilize differentiable marching tetrahedrons for direct optimization on mesh surfaces. [10, 28] deal with varying cameras, illumination, and backgrounds. Ref-NeRF [42] suggests the employment of integrated direction encoding to ameliorate the reconstruction fidelity of reflective objects. [30, 33] use split-sum approximation to approximate the shading effects. NeMF [55] represents the scene as a microflake volume. NeFII [44] integrates lights through path tracing with Monte Carlo sampling. InvRender [56] computes indirect illumination by directly leveraging the neural radiance field, rather than concurrently estimating it alongside the decomposition of direct lighting and materials. TensoIR [23] performs inverse rendering based on tensor factorization and neural fields. NeILF [45] and NeILF++ [49] expresses the incident lights as a neural incident light field, while NeILF++ integrates VolSDF [46] with NeILF through inter-reflection. However, existing schemes rarely delve into BRDF estimation for point clouds, and the estimation and rendering quality remain to be improved." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.467, + 0.502, + 0.485 + ], + "angle": 0, + "content": "3 Relightable 3D Gaussians" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.494, + 0.787, + 0.542 + ], + "angle": 0, + "content": "In this section, we introduce a novel pipeline to decompose materials and lighting from multi-view images based on 3D Gaussian Splatting (3DGS) [26]. An overview of our pipeline is shown in Fig. 2." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.558, + 0.362, + 0.574 + ], + "angle": 0, + "content": "3.1 Preliminary" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.577, + 0.787, + 0.623 + ], + "angle": 0, + "content": "Distinct from the widely adopted Neural Radiance Field, 3DGS [26] employs explicit 3D Gaussian points as its rendering primitives. A 3D Gaussian point is mathematically defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.363, + 0.628, + 0.786, + 0.657 + ], + "angle": 0, + "content": "\\[\nG (\\boldsymbol {x}) = \\exp \\left(- \\frac {1}{2} (\\boldsymbol {x} - \\boldsymbol {\\mu}) ^ {\\top} \\Sigma^ {- 1} (\\boldsymbol {x} - \\boldsymbol {\\mu})\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.66, + 0.786, + 0.69 + ], + "angle": 0, + "content": "where \\(\\pmb{\\mu}\\) and \\(\\Sigma\\) denote the 3D spatial mean and covariance matrix, respectively. Each Gaussian is also equipped with an opacity \\(o\\) and a view-dependent color \\(\\pmb{c}\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.69, + 0.787, + 0.797 + ], + "angle": 0, + "content": "The rendering process in 3DGS is divided into two main steps. Firstly, 3D Gaussians are projected to 2D Gaussians on the image plane. The 2D means are determined by accurate projection of 3D means, while the 2D covariance matrices are approximated by: \\(\\Sigma' = J\\boldsymbol{W} \\Sigma \\boldsymbol{W}^\\top \\boldsymbol{J}^\\top\\), where \\(\\boldsymbol{W}\\) and \\(\\boldsymbol{J}\\) denote the viewing transformation and the Jacobian of the affine approximation of perspective projection transformation [57]. Subsequently, the pixel color is derived by alpha blending \\(N\\) ordered 2D Gaussians from front to back:" + }, + { + "type": "equation", + "bbox": [ + 0.381, + 0.803, + 0.786, + 0.845 + ], + "angle": 0, + "content": "\\[\n\\mathcal {C} = \\sum_ {i \\in N} T _ {i} \\alpha_ {i} \\boldsymbol {c} _ {i}, \\quad T _ {i} = \\prod_ {j = 1} ^ {i - 1} (1 - \\alpha_ {j}). \\tag {2}\n\\]" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.559, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "Relightable 3D Gaussians" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "image", + "bbox": [ + 0.221, + 0.147, + 0.782, + 0.299 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.309, + 0.788, + 0.407 + ], + "angle": 0, + "content": "Fig. 2: The proposed differentiable rendering pipeline. Starting with a collection of 3D Gaussians that embody geometry, materials, and lighting attributes, we first execute the physically based rendering equation for each 3D Gaussian to determine its outgoing radiance, denoted as \\( c' \\). Following this, we perform rasterization and alpha blending to obtain vanilla color map \\( \\mathcal{C} \\), PBR color map \\( \\mathcal{C}' \\), depth map \\( \\mathcal{D} \\), normal map \\( \\mathcal{N} \\), etc. To optimize relightable 3D Gaussians, we utilize the ground truth image \\( \\mathcal{C}_{gt} \\) and the pseudo normal map derived from \\( \\mathcal{D} \\) for supervision." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.442, + 0.788, + 0.532 + ], + "angle": 0, + "content": "Here, \\(\\alpha\\) is obtained by multiplying the opacity \\(o\\) with the 2D covariance's contribution computed from \\(\\Sigma^{\\prime}\\) and pixel coordinate in image space [26]. In implementation details, the covariance matrix \\(\\Sigma\\) is parameterized as a unit quaternion \\(\\pmb{q}\\) and a 3D scaling vector \\(\\pmb{s}\\) to maintain its meaningful interpretation throughout optimization. Additionally, view-dependent color \\(\\pmb{c}_i\\) is represented through a set of Spherical Harmonics (SH)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.535, + 0.787, + 0.567 + ], + "angle": 0, + "content": "In summation, a 3D scene is represented by a collection of 3D Gaussians, with the \\(i^{th}\\) Gaussian \\(\\mathcal{P}_i\\) parameterized as \\(\\{\\pmb{\\mu}_i, \\pmb{q}_i, \\pmb{s}_i, o_i, \\pmb{c}_i\\}\\)." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.601, + 0.463, + 0.617 + ], + "angle": 0, + "content": "3.2 Geometry Enhancement" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.639, + 0.787, + 0.669 + ], + "angle": 0, + "content": "Scene geometry, specifically surface normal, is essential for realistic physically based rendering, which will be discussed in Sec. 3.3." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.672, + 0.788, + 0.778 + ], + "angle": 0, + "content": "Normal Estimation. We incorporate a normal attribute \\( \\pmb{n} \\) for each 3D Gaussian and try to solve it. A potential methodology treats the spatial mean of a 3D Gaussian as a conventional point, and executes normal estimation based on the local planar assumption. However, this approach cannot provide accurate normal estimation. The reasons are two fold: First, the reconstructed Gaussian point cloud is often sparse. More critically, the Gaussian points are naturally soft, which means these points are not precisely aligned with the object surface." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.841 + ], + "angle": 0, + "content": "To address these limitations, we propose to optimize \\(\\pmb{n}\\) from an initial random vector via back-propagation for each 3D Gaussian. We perceive the depth of all Gaussians along a ray as a distribution, and estimate the pixel depth as the expectation of this distribution. Similarly, we determine the normal for each" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.357, + 0.128 + ], + "angle": 0, + "content": "J. Gao et al." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.147, + 0.47, + 0.163 + ], + "angle": 0, + "content": "pixel. This process is described by:" + }, + { + "type": "equation", + "bbox": [ + 0.414, + 0.174, + 0.786, + 0.205 + ], + "angle": 0, + "content": "\\[\n\\{\\mathcal {D}, \\mathcal {N} \\} = \\sum_ {i \\in N} w _ {i} \\left\\{d _ {i}, \\boldsymbol {n} _ {i} \\right\\}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.216, + 0.789, + 0.278 + ], + "angle": 0, + "content": "where \\(d_{i}\\), \\(\\pmb{n}_i\\) and \\(w_{i} = T_{i}\\alpha_{i} / \\sum_{i\\in N}T_{i}\\alpha_{i}\\) denote the depth, normal and weight of the point. We then encourage the consistency between the rendered normal \\(\\mathcal{N}\\) and the pseudo normal \\(\\tilde{\\mathcal{N}}\\), which is computed from the rendered depth \\(\\mathcal{D}\\) under the local planarity assumption. The normal consistency is quantified as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.441, + 0.289, + 0.786, + 0.306 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {n} = \\| \\mathcal {N} - \\tilde {\\mathcal {N}} \\| _ {2}. \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.317, + 0.784, + 0.406 + ], + "angle": 0, + "content": "Normal Gradient Based Densification. To achieve superior rendering fidelity in detail regions, vanilla 3DGS [26] densifies 3D Gaussians through the gradient of view space points. Drawing upon this, to improve normal recovery in thin regions, we introduce an additional densification criterion on the gradient of normals. Specifically, we densify Gaussians whose normal gradient exceeds a threshold \\( T_{n} \\)." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.408, + 0.784, + 0.453 + ], + "angle": 0, + "content": "Constraint on Depth Distribution. Given our assumption of the object possessing an accurate surface, we enforce a constraint on the depth distribution by minimizing the uncertainty. The uncertainty is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.447, + 0.468, + 0.785, + 0.484 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {u} = \\mathcal {D} _ {s q} - \\mathcal {D} ^ {2}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.49, + 0.784, + 0.535 + ], + "angle": 0, + "content": "where \\(\\mathcal{D}_{sq} = \\sum_{i\\in N}w_id_i^2\\), and \\(\\mathcal{D}\\) is the depth estimation defined in Eq. 3. This constraint on the depth distribution drives Gaussian points to the object surface, thereby improving geometric reconstruction." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.536, + 0.784, + 0.567 + ], + "angle": 0, + "content": "Object Mask Constraint. If there is a mask indicating the object, we can constrain the optimization by a binary cross entropy loss [43]:" + }, + { + "type": "equation", + "bbox": [ + 0.364, + 0.581, + 0.785, + 0.596 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {O} = - M \\log O - (1 - M) \\log (1 - O), \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.609, + 0.784, + 0.64 + ], + "angle": 0, + "content": "where \\(M\\) is the object mask and \\(O = \\sum_{i\\in N}T_i\\alpha_i\\). With this constraint, \\(O\\) is forced to be aligned with the distribution of \\(M\\), and so we get opaque surface." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.662, + 0.486, + 0.679 + ], + "angle": 0, + "content": "3.3 BRDF and Light Modeling" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.687, + 0.784, + 0.734 + ], + "angle": 0, + "content": "Rendering Equation. We use the rendering equation [24] to model light interaction with surfaces, accounting for their material properties and geometry. It is given by:" + }, + { + "type": "equation", + "bbox": [ + 0.327, + 0.744, + 0.785, + 0.775 + ], + "angle": 0, + "content": "\\[\nL _ {o} \\left(\\boldsymbol {\\omega} _ {o}, \\boldsymbol {x}\\right) = \\int_ {\\Omega} f \\left(\\boldsymbol {\\omega} _ {o}, \\boldsymbol {\\omega} _ {i}, \\boldsymbol {x}\\right) L _ {i} \\left(\\boldsymbol {\\omega} _ {i}, \\boldsymbol {x}\\right) \\left(\\boldsymbol {\\omega} _ {i} \\cdot \\boldsymbol {n}\\right) d \\boldsymbol {\\omega} _ {i}, \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.78, + 0.786, + 0.84 + ], + "angle": 0, + "content": "where \\(\\pmb{x}\\) and \\(\\pmb{n}\\) are the surface point and its normal vector, \\(f\\) is the Bidirectional Reflectance Distribution Function (BRDF), and \\(L_{i}\\) and \\(L_{o}\\) denote the incoming and outgoing radiance in directions \\(\\omega_{i}\\) and \\(\\omega_{o}\\). \\(\\Omega\\) signifies the hemispherical domain above the surface." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.559, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "Relightable 3D Gaussians" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.787, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.282 + ], + "angle": 0, + "content": "Prior methods [49, 51] typically begin with the acquisition of intersection points between rays and surface through differentiable rendering. Then they apply the rendering equation at these points to facilitate Physical Based Rendering (PBR). However, these approach present significant challenges in the point-based framework. Certainly, it is feasible to extract surfaces in 3DGS [17]; but, it necessitates a coordinate-based incident light field for inverse PBR. The querying of this field at millions of points during every iteration imposes a substantial computational burden. Moreover, it is pertinent to acknowledge that accurately extracting geometric surfaces presents considerable challenges in 3DGS system." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.283, + 0.785, + 0.373 + ], + "angle": 0, + "content": "To tackle this issue, we propose to compute PBR color \\(\\{c_i^{\\prime}\\}_{i = 0}^{N}\\) for each 3D Gaussian, and then obtain the PBR image \\(\\mathcal{C}'\\) through alpha-blending, as shown in Fig. 2. This method is more efficient for two main reasons. First, PBR is performed on fewer 3D Gaussians rather than image pixels in all input views, as each Gaussian affects multiple pixels based on its scale. Second, by allocating discrete attributes for each Gaussian, we avoid the need for global neural fields." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.374, + 0.787, + 0.446 + ], + "angle": 0, + "content": "BRDF Parameterization. To make 3D Gaussians relightable, we assign BRDF properties to each Gaussian and adopt a simplified Disney BRDF model [11]. The BRDF properties include an albedo \\( \\pmb{b} \\in [0,1]^3 \\) and a roughness \\( r \\in [0,1] \\), and the BRDF \\( f \\) in Eq. 7 is divided into a diffuse term \\( f_{d} = \\frac{b}{\\pi} \\) and a specular term:" + }, + { + "type": "equation", + "bbox": [ + 0.331, + 0.446, + 0.786, + 0.478 + ], + "angle": 0, + "content": "\\[\nf _ {s} \\left(\\boldsymbol {\\omega} _ {o}, \\boldsymbol {\\omega} _ {i}\\right) = \\frac {D (\\boldsymbol {h} ; r) \\cdot F \\left(\\boldsymbol {\\omega} _ {o} , \\boldsymbol {h}\\right) \\cdot G \\left(\\boldsymbol {\\omega} _ {i} , \\boldsymbol {\\omega} _ {o} , h ; r\\right)}{\\left(\\boldsymbol {n} \\cdot \\boldsymbol {\\omega} _ {i}\\right) \\cdot \\left(\\boldsymbol {n} \\cdot \\boldsymbol {\\omega} _ {o}\\right)}, \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.482, + 0.784, + 0.51 + ], + "angle": 0, + "content": "where \\( \\pmb{h} = (\\pmb{\\omega}_i + \\pmb{\\omega}_o) / 2 \\) is the half vector, \\( D, F \\) and \\( G \\) denote the normal distribution function, Fresnel term and geometry term." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.513, + 0.784, + 0.603 + ], + "angle": 0, + "content": "Incident Light Modeling. Optimizing a NeILF for each Gaussian can be excessively unconstrained, leading to challenges in accurately decomposing incident lights from its appearance. To address this issue, we apply a prior by partitioning the incident light into globally shared direct components and individual per-Gaussian indirect components. The sampled incident light at a Gaussian from direction \\(\\omega_{i}\\) is represented as:" + }, + { + "type": "equation", + "bbox": [ + 0.345, + 0.616, + 0.785, + 0.631 + ], + "angle": 0, + "content": "\\[\nL _ {i} (\\boldsymbol {\\omega} _ {i}) = V (\\boldsymbol {\\omega} _ {i}) \\cdot L _ {\\text {d i r e c t}} (\\boldsymbol {\\omega} _ {i}) + L _ {\\text {i n d i r e c t}} (\\boldsymbol {\\omega} _ {i}), \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.641, + 0.785, + 0.746 + ], + "angle": 0, + "content": "where \\( V(\\omega_{i}) \\) represents the visibility term which will be further discussed in Sec. 4, the indirect light term \\( L_{\\text{indirect}} \\) is parameterized by 3-level SH, denoted as \\( l \\), while the direct light term \\( L_{\\text{direct}} \\) is parameterized as a globally shared 16x32 environment map, denoted as \\( l^{env} \\). Despite the utilization of relatively low-level Spherical Harmonics (SH), we can still capture some inter-reflection effects, as the rendering color of a given pixel arises from the collaboration of multiple relightable 3D Gaussians." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.747, + 0.784, + 0.793 + ], + "angle": 0, + "content": "For each 3D Gaussian, we sample \\(N_{s}\\) incident directions over the hemisphere space through Fibonacci sampling [45] to provide numerical integration for Eq. 7. The PBR color of each Gaussian is then given by:" + }, + { + "type": "equation", + "bbox": [ + 0.327, + 0.803, + 0.786, + 0.843 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {c} ^ {\\prime} \\left(\\boldsymbol {\\omega} _ {o}\\right) = \\sum_ {i = 0} ^ {N _ {s}} \\left(f _ {d} + f _ {s} \\left(\\boldsymbol {\\omega} _ {o}, \\boldsymbol {\\omega} _ {i}\\right)\\right) L _ {i} \\left(\\boldsymbol {\\omega} _ {i}\\right) \\left(\\boldsymbol {\\omega} _ {i} \\cdot \\boldsymbol {n}\\right) \\Delta \\boldsymbol {\\omega} _ {i}, \\tag {10}\n\\]" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.358, + 0.128 + ], + "angle": 0, + "content": "J. Gao et al." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.147, + 0.414, + 0.162 + ], + "angle": 0, + "content": "where \\(\\omega_{i}\\) is the solid angle." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.163, + 0.788, + 0.209 + ], + "angle": 0, + "content": "To summarize, our method represents a 3D scene as a set of relightable 3D Gaussians and a global environment light \\( l^{env} \\), where the \\( i^{th} \\) Gaussian \\( \\mathcal{P}_i \\) is parameterized as \\( \\{\\pmb{\\mu}_i, \\pmb{q}_i, \\pmb{s}_i, o_i, \\pmb{c}_i, \\pmb{n}_i, \\pmb{b}_i, r_i, l_i\\} \\)." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.231, + 0.391, + 0.246 + ], + "angle": 0, + "content": "3.4 Regularizations" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.256, + 0.784, + 0.286 + ], + "angle": 0, + "content": "To mitigate the materials-lighting ambiguity [45], proper regularizations are utilized to facilitate their plausible decomposition." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.286, + 0.785, + 0.317 + ], + "angle": 0, + "content": "Light Regularization. We apply a light regularization assuming a near-natural white incident light [30]:" + }, + { + "type": "equation", + "bbox": [ + 0.344, + 0.328, + 0.786, + 0.358 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {l i g h t}} = \\sum_ {c} \\left(L _ {c} - \\frac {1}{3} \\sum_ {c} L _ {c}\\right), c \\in \\{R, G, B \\}. \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.366, + 0.784, + 0.409 + ], + "angle": 0, + "content": "Smoothness Priors. We expect that the BRDF properties do not change drastically in homogeneous areas [45]. We define a smooth constraint on roughness as:" + }, + { + "type": "equation", + "bbox": [ + 0.4, + 0.412, + 0.785, + 0.429 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {s, r} = \\| \\nabla R \\| \\exp (- \\| \\nabla C _ {g t} \\|), \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.434, + 0.784, + 0.466 + ], + "angle": 0, + "content": "where \\( R = \\sum_{i\\in N}w_i r_i \\) is the rendered roughness map. Similarly, we also define smoothness constraints \\( \\mathcal{L}_{s,n} \\) and \\( \\mathcal{L}_{s,b} \\) on normal and albedo." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.488, + 0.492, + 0.507 + ], + "angle": 0, + "content": "4 Point-based Ray Tracing" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.521, + 0.784, + 0.551 + ], + "angle": 0, + "content": "For real-time realistic rendering with plausible shadow effects on relightable 3D Gaussians, we introduce a novel point-based ray tracing approach in this section." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.573, + 0.507, + 0.589 + ], + "angle": 0, + "content": "4.1 Ray Tracing on 3D Gaussians" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.599, + 0.785, + 0.704 + ], + "angle": 0, + "content": "Our proposed ray tracing technique on 3D Gaussians is built upon the Bounding Volume Hierarchy (BVH), enabling efficient visibility querying along a ray. Our method adopts the idea from [25], an in-place algorithm for constructing a binary radix tree that maximizes parallelism and facilitates real-time BVH construction. Specifically, we construct a binary radix tree from a given set of 3D Gaussians, where each leaf node represents the tight bounding box of a Gaussian, and each internal node denotes the bounding box of its two children." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.705, + 0.787, + 0.841 + ], + "angle": 0, + "content": "Unlike ray tracing with opaque polygonal meshes, where only the closest intersection point is required, ray tracing with semi-transparent Gaussians necessitates accounting for all Gaussians potentially influencing the ray's transmittance. The process of ray tracing on 3D Gaussian points can be described as follows: the ray travels from the camera center and accumulates transmittance as it passes through 3D Gaussian points until the transmittance is zero. The first key issue to be addressed is how to define the contribution of a single 3D Gaussian to the transmittance. As discussed in Sec.3.1, transforming a 3D Gaussian to a 2D Gaussian involves an approximation, which complicating the" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.559, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "Relightable 3D Gaussians" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.787, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "image", + "bbox": [ + 0.216, + 0.15, + 0.353, + 0.294 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.241, + 0.295, + 0.329, + 0.307 + ], + "angle": 0, + "content": "(a) Root Node" + }, + { + "type": "image", + "bbox": [ + 0.361, + 0.15, + 0.498, + 0.294 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.384, + 0.295, + 0.476, + 0.307 + ], + "angle": 0, + "content": "(b) Child Node" + }, + { + "type": "image", + "bbox": [ + 0.507, + 0.15, + 0.642, + 0.294 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.537, + 0.295, + 0.613, + 0.307 + ], + "angle": 0, + "content": "(c) Gaussian" + }, + { + "type": "image", + "bbox": [ + 0.652, + 0.15, + 0.788, + 0.294 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.681, + 0.295, + 0.758, + 0.307 + ], + "angle": 0, + "content": "(d) Gaussian" + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.319, + 0.788, + 0.389 + ], + "angle": 0, + "content": "Fig. 3: Intersection tests in point-based ray tracing. Intersection point between ray and Gaussian is obtained by three steps: (a) intersect the BVH root node; (b) dive into the intersected child nodes recursively until the leaf node; (c) perform Eq. 13 to get the equivalent intersection point. (d) shows that a 3D Gaussian actually has non-negligible effect on a segment of a ray." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.419, + 0.784, + 0.495 + ], + "angle": 0, + "content": "identification of accurate intersection between the Gaussian and a ray. As shown in Fig. 3d, a 3D Gaussian actually has non-negligible effect on a segment of a ray. Thus, drawing inspiration from [27], we approximate the intersection of ray with 3D Gaussian as a point where the 3D Gaussian's contribution peaks, as showed in Fig. 3(3). The equivalent intersection point is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.442, + 0.51, + 0.785, + 0.526 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {r} _ {\\boldsymbol {x}} = \\boldsymbol {r} _ {\\boldsymbol {o}} + t _ {j} \\boldsymbol {r} _ {\\boldsymbol {d}}, \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.532, + 0.784, + 0.564 + ], + "angle": 0, + "content": "where \\(\\pmb{r_o}\\) denotes the origin, \\(\\pmb{r_d}\\) is direction of the ray, corresponding to the previously mentioned \\(\\omega_{i}, t_{j}\\) is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.424, + 0.574, + 0.785, + 0.609 + ], + "angle": 0, + "content": "\\[\nt _ {j} = \\frac {\\left(\\boldsymbol {\\mu} - \\boldsymbol {r} _ {o}\\right) ^ {T} \\boldsymbol {\\Sigma} \\boldsymbol {r} _ {d}}{\\boldsymbol {r} _ {d} ^ {T} \\boldsymbol {\\Sigma} \\boldsymbol {r} _ {d}}. \\tag {14}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.617, + 0.784, + 0.647 + ], + "angle": 0, + "content": "Subsequently, we approximate the contribution of this 3D Gaussian to the ray as \\(\\alpha_{j}\\) at the equivalent intersection point \\(\\pmb{r}_{\\pmb{x}}\\)" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.649, + 0.788, + 0.769 + ], + "angle": 0, + "content": "Considering the transmittance along a ray: \\( T_{i} = \\prod_{j=1}^{i-1}(1 - \\alpha_{j}) \\), it is evident that the order of \\( \\alpha_{j} \\) does not affect \\( T_{i} \\). In other words, the order in which Gaussians are encountered along a ray does not impact the overall transmittance. As illustrated in Fig. 3, starting from the root node of the binary radix tree, intersection tests are recursively performed between the ray and the bounding volumes of each node's children. Upon reaching a leaf node, the associated Gaussian is identified. Through this traversal, the transmittance \\( T \\) is progressively attenuated:" + }, + { + "type": "equation", + "bbox": [ + 0.301, + 0.782, + 0.785, + 0.799 + ], + "angle": 0, + "content": "\\[\nT _ {i} = \\left(1 - \\alpha_ {i - 1}\\right) T _ {i - 1}, \\quad \\text {f o r} i = 1, \\dots , j - 1 \\text {w i t h} T _ {1} = 1. \\tag {15}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.784, + 0.841 + ], + "angle": 0, + "content": "To speed up ray tracing, the process is early terminated if a ray's transmittance drops below a certain threshold \\( T_{\\min} \\)." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.356, + 0.127 + ], + "angle": 0, + "content": "J. Gao et al." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.147, + 0.482, + 0.162 + ], + "angle": 0, + "content": "4.2 Visibility Pre-computation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.174, + 0.787, + 0.28 + ], + "angle": 0, + "content": "In Sec. 3.3, a potential ambiguity in the decomposition of indirect and direct lights has been discerned. While the essential visibility term can be derived via the proposed ray tracing on 3D Gaussians (Sec. 4.1), querying the visibility term through ray tracing online proves challenging due to the computational complexity. Nevertheless, given our exclusive focus on static scenes, we can pre-compute the visibility term \\( V(\\omega_{i}) \\) across the hemispherical domain determined by \\( \\pmb{n} \\) for each Gaussian, and subsequently integrate it into the rendering equation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.28, + 0.787, + 0.371 + ], + "angle": 0, + "content": "Hence, we divide the optimization process into two stages. In the first stage, we optimize a vanilla 3D Gaussian point cloud, augmented with an additional normal vector \\(\\pmb{n}\\) for each Gaussian. Immediately afterwards, we pre-compute the per-Gaussian visibility through the proposed ray tracing. In the second stage, we lock the geometry of 3D Gaussians and focus solely on optimizing the material and lighting parameters using the pipeline described in Fig. 2." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.395, + 0.527, + 0.411 + ], + "angle": 0, + "content": "4.3 Point Based Relighting Pipeline" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.422, + 0.787, + 0.543 + ], + "angle": 0, + "content": "Based on our relightable 3D Gaussians, we devise a point-based graphics pipeline that seamlessly integrates effortless scene editing and realistic relighting. To our knowledge, there is currently no point-based rendering pipeline that effectively accomplishes both tasks. Although explicit point representations make scene editing easy, achieving photo-realistic relighting proved nearly insurmountable in existing point-based rendering pipeline. Conversely, in inverse rendering approaches based on implicit representations, while highly realistic relighting is feasible, scene editing presents a challenging endeavor." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.543, + 0.788, + 0.695 + ], + "angle": 0, + "content": "As an illustration, we concentrate on relighting in a multi-object composition scene. Initially, our pipeline computes the visibility \\( V(\\omega_{i}) \\) for each Gaussian point via ray tracing (Sec. 4.1). Despite the challenging inter-object occlusions introduced by composition, our proposed point-based ray tracing method adeptly manages these complexities and ensures accurate occlusion updates within the novel scene. Subsequently, the rendering process unfolds (Sec. 3.3), starting with Gaussian-level PBR and ending with alpha blending. Throughout this process, the original indirect lighting of each object is discarded. Consequently, we achieve relighting with remarkably vivid shadow effects solely based on a discrete set of points, as illustrated in Fig. 1." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.72, + 0.377, + 0.736 + ], + "angle": 0, + "content": "5 Experiments" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.753, + 0.398, + 0.768 + ], + "angle": 0, + "content": "5.1 Training Details" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.787, + 0.841 + ], + "angle": 0, + "content": "As pointed out in Sec. 4.2, the training procedure is divided into two stages. The first stage follows the setting of 3DGS [26], along with the proposed normal gradient-based densification (Sec. 3.2) where \\( T_{\\mathbf{n}} \\) is set as \\( 2 \\times 10^{-9} \\). In the second stage, we sample \\( N_{s} = 64 \\) incident rays per Gaussian point for PBR. We train" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.559, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "Relightable 3D Gaussians" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "table_caption", + "bbox": [ + 0.229, + 0.145, + 0.774, + 0.159 + ], + "angle": 0, + "content": "Table 1: Quantitative results for novel view synthesis on NeRF synthetic dataset" + }, + { + "type": "table", + "bbox": [ + 0.333, + 0.172, + 0.671, + 0.293 + ], + "angle": 0, + "content": "
GeometryRelightablePSNR↑ SSIM↑ LPIPS
NPBG [1]point×28.100.9230.077
NPBG++ [38]point×28.120.9280.076
FreqPCR [54]point×31.240.9500.049
3DGS [26]point×33.880.9700.031
PhySG [51]neural✓18.910.8470.182
NeILF++ [49]neural✓26.370.9110.091
Nvdiffrec [33]mesh✓29.050.9390.081
R3DG(Ours)point✓31.220.9590.039
" + }, + { + "type": "image", + "bbox": [ + 0.229, + 0.304, + 0.396, + 0.447 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.289, + 0.449, + 0.343, + 0.461 + ], + "angle": 0, + "content": "(a) Lego" + }, + { + "type": "image", + "bbox": [ + 0.424, + 0.304, + 0.575, + 0.447 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.473, + 0.449, + 0.532, + 0.461 + ], + "angle": 0, + "content": "(b) Chair" + }, + { + "type": "image", + "bbox": [ + 0.594, + 0.314, + 0.782, + 0.44 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.656, + 0.449, + 0.723, + 0.461 + ], + "angle": 0, + "content": "(c) Hotdog" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.472, + 0.788, + 0.515 + ], + "angle": 0, + "content": "Fig. 4: Visualizations on NeRF synthetic dataset [31]. Each scene is displayed in an order from left to right and from top to bottom: Ground Truth, PBR Image, Normal Map and Ambient Occlusion Map." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.545, + 0.785, + 0.576 + ], + "angle": 0, + "content": "our model for 30,000 iterations in the initial stage and 10,000 iterations in the second stage. The loss in the first stage is calculated by:" + }, + { + "type": "equation", + "bbox": [ + 0.284, + 0.587, + 0.785, + 0.604 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\lambda_ {1} \\mathcal {L} _ {1} + \\lambda_ {s s i m} \\mathcal {L} _ {s s i m} + \\lambda_ {n} \\mathcal {L} _ {n} + \\lambda_ {s, n} \\mathcal {L} _ {s, n} + \\lambda_ {O} \\mathcal {L} _ {O} + \\lambda_ {u} \\mathcal {L} _ {u}, \\tag {16}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.615, + 0.787, + 0.647 + ], + "angle": 0, + "content": "where \\(\\{\\lambda_1, \\lambda_{ssim}, \\lambda_n, \\lambda_{s,n}, \\lambda_O, \\lambda_u\\}\\) are set to \\(\\{0.8, 0.2, 0.01, 0.01, 0.01, 0.01\\}\\), respectively. The loss in the second stage is given by:" + }, + { + "type": "equation", + "bbox": [ + 0.312, + 0.658, + 0.785, + 0.675 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\lambda_ {1} \\mathcal {L} _ {1} + \\lambda_ {s s i m} \\mathcal {L} _ {s s i m} + \\lambda_ {l} \\mathcal {L} _ {l} + \\lambda_ {s, b} \\mathcal {L} _ {s, b} + \\lambda_ {s, r} \\mathcal {L} _ {s, r}, \\tag {17}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.686, + 0.788, + 0.703 + ], + "angle": 0, + "content": "where \\(\\{\\lambda_1,\\lambda_{ssim},\\lambda_l,\\lambda_{s,b},\\lambda_{s,r}\\}\\) are set to \\(\\{0.8,0.2,0.0001,0.01,0.01\\}\\), respectively." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.724, + 0.578, + 0.74 + ], + "angle": 0, + "content": "5.2 Performance on Novel View Synthesis" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.841 + ], + "angle": 0, + "content": "We first evaluate the novel view synthesis (NVS) of the physically based rendering (PBR) on NeRF synthetic dataset [31]. We compare both point-based rendering methods and relightable rendering approaches. We report average metrics across all scenes in Tab. 1, including peak signal-to-noise ratio (PSNR), structural similarity index measure (SSIM), and learned perceptual image patch similarity (LPIPS). Compared with existing point-based rendering methods, our" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.358, + 0.127 + ], + "angle": 0, + "content": "J. Gao et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.288, + 0.145, + 0.715, + 0.16 + ], + "angle": 0, + "content": "Table 2: Quantitative results on Synthetic4Relight dataset [56]" + }, + { + "type": "table", + "bbox": [ + 0.22, + 0.172, + 0.785, + 0.257 + ], + "angle": 0, + "content": "
Novel View SynthesisRelightingAlbedoRoughnessTime
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓MSE↓(hours)
NerFactor [53]22.800.9160.15021.540.8750.17119.490.8640.206N/A>48
Nvdifferc-MC [19]34.290.9670.06824.220.9430.07829.610.9450.0750.0094.17
InvRender [56]30.740.9530.08628.670.9500.09128.280.9350.0720.00814.3
TensorIR [23]35.800.9780.04929.690.9510.07930.580.9460.0650.0153.24
R3DG (Ours)36.800.9820.02831.000.9640.05028.310.9510.0580.0130.90
" + }, + { + "type": "image", + "bbox": [ + 0.22, + 0.263, + 0.782, + 0.562 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.298, + 0.574, + 0.704, + 0.589 + ], + "angle": 0, + "content": "Fig. 5: Qualitative results on Synthetic4Relight dataset [56]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.615, + 0.788, + 0.691 + ], + "angle": 0, + "content": "R3DG demonstrates superiority over most of them [1,38,54]. While our performance slightly marginally trails that of vanilla 3DGS [26], it is noteworthy that our approach excels in relighting capability, presenting a significant advantage. Furthermore, in comparison to other relightable methods [33,49,51], our method showcases significantly better NVS quality." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.692, + 0.788, + 0.768 + ], + "angle": 0, + "content": "In Fig. 4, we visualize the our reconstruction results in NeRF synthetic dataset, including the PBR image, the normal map, and the pre-computed visibility illustrated in the form of the Ambient Occlusion (AO) map. As depicted in Fig. 4, our approach successfully achieves realistic PBR rendering, smooth normal estimation, and accurate visibility solving on discrete point clouds." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.787, + 0.485, + 0.803 + ], + "angle": 0, + "content": "5.3 Performance on Relighting" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.842 + ], + "angle": 0, + "content": "To comprehensively assess the relighting capabilities of our pipeline, we further undertook experiments using the Synthetic4Relight dataset [56]. Recognizing" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.559, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "Relightable 3D Gaussians" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.144, + 0.782, + 0.373 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.3, + 0.386, + 0.701, + 0.4 + ], + "angle": 0, + "content": "Fig. 6: Qualitative results of relighting on real-world scenes." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.43, + 0.784, + 0.476 + ], + "angle": 0, + "content": "the inherent scale ambiguity between the estimated albedo and lighting, we standardized the scale of the estimated albedo against the ground truth to relighting, consistent with previous studies [51, 56]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.476, + 0.785, + 0.581 + ], + "angle": 0, + "content": "Our evaluation encompasses the analysis of decomposed materials, the synthesis of novel views, and the relighting outcomes. The quantitative analysis is presented in Tab. 2. Our method outperforms existing approaches in terms of Novel View Synthesis (NVS) and relighting precision. Regarding material estimation, our method exhibits superior albedo accuracy in SSIM and LPIPS. Qualitatively, our method achieves visually pleasing material decomposition, facilitating a realistic relighting effect, which is shown in 5." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.581, + 0.785, + 0.627 + ], + "angle": 0, + "content": "Additionally, we conduct relighting on real-world scenes in the Mip-NeRF 360 dataset [4], as shown in Fig. 6. It is worth noting that our relighting in the kitchen with the second environment map, exhibits pleasing shadow effects." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.649, + 0.388, + 0.665 + ], + "angle": 0, + "content": "5.4 Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.788, + 0.841 + ], + "angle": 0, + "content": "We conduct ablation studies on three principal components of our method that significantly influence the quality of inverse rendering. The results are illustrated in Fig. 7. Initially, we investigate the impact of normal gradient densification (NGD). Fig. 7 (a, b) depict the rendering normal maps with and without the densification. It shows that the proposed densification markedly enhances the details of the normal map for *Lego*. In Fig. 7 (c, d, e), we examine the effects of our full light modeling. We simplify the full model to a global environment map in the optimization. The result demonstrates that the simplified model fails to generate shadow effects when relighting. Furthermore, we evaluate the significance of the proposed constraint on depth distribution, denoted as \\( \\mathcal{L}_u \\). The depth uncertainty is visualized in Fig. 7 (h, i), which illustrates the distribution" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.358, + 0.127 + ], + "angle": 0, + "content": "J. Gao et al." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.149, + 0.321, + 0.229 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.237, + 0.231, + 0.308, + 0.242 + ], + "angle": 0, + "content": "(a) w NGD." + }, + { + "type": "image", + "bbox": [ + 0.331, + 0.149, + 0.432, + 0.229 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.343, + 0.231, + 0.428, + 0.242 + ], + "angle": 0, + "content": "(b) w/o NGD." + }, + { + "type": "image", + "bbox": [ + 0.443, + 0.144, + 0.553, + 0.23 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.457, + 0.231, + 0.539, + 0.242 + ], + "angle": 0, + "content": "(c) Relit. GT." + }, + { + "type": "image", + "bbox": [ + 0.555, + 0.144, + 0.665, + 0.23 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.57, + 0.231, + 0.65, + 0.242 + ], + "angle": 0, + "content": "(d) Full light." + }, + { + "type": "image", + "bbox": [ + 0.668, + 0.144, + 0.779, + 0.23 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.681, + 0.231, + 0.765, + 0.242 + ], + "angle": 0, + "content": "(e) Simplified." + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.245, + 0.351, + 0.287 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.246, + 0.292, + 0.328, + 0.305 + ], + "angle": 0, + "content": "(f) AO w \\(\\mathcal{L}_u\\)" + }, + { + "type": "image", + "bbox": [ + 0.362, + 0.246, + 0.492, + 0.288 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.381, + 0.292, + 0.475, + 0.305 + ], + "angle": 0, + "content": "(g) AO w/o \\(\\mathcal{L}_u\\)" + }, + { + "type": "image", + "bbox": [ + 0.499, + 0.243, + 0.637, + 0.291 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.503, + 0.293, + 0.634, + 0.305 + ], + "angle": 0, + "content": "(h) Uncertainty w \\(\\mathcal{L}_u\\)" + }, + { + "type": "image", + "bbox": [ + 0.64, + 0.243, + 0.779, + 0.291 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.642, + 0.293, + 0.778, + 0.305 + ], + "angle": 0, + "content": "(i) Uncertainty w/o \\(\\mathcal{L}_u\\)" + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.315, + 0.788, + 0.4 + ], + "angle": 0, + "content": "Fig. 7: Ablation studies on main components of our method. (a) and (b) show normal maps with and without the proposed normal gradient densification (NGD). (d) and (e) display re-lighting results using the proposed full lighting model and only a simplified light modeling, i.e. a global environment map. (f) and (g) visualize ambient occlusion (AO) maps with and without the proposed constraint on depth distribution \\(\\mathcal{L}_u\\). (h) and (i) illustrate the depth uncertainty with and without \\(\\mathcal{L}_u\\)." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.435, + 0.785, + 0.482 + ], + "angle": 0, + "content": "constraint contribute significantly to reducing the depth uncertainty. Additionally, Fig. 7 (f, g) reveals that incorporating \\(\\mathcal{L}_u\\) leads to more plausible ambient occlusion (averaged visibility) which plays a crucial role in stage 2." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.519, + 0.359, + 0.535 + ], + "angle": 0, + "content": "6 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.565, + 0.788, + 0.747 + ], + "angle": 0, + "content": "In this paper, we have introduced a differentiable point-based rendering pipeline enabling effort less editing, accuracy ray tracing, and photo-realistic relighting. We present the scene as Relightable 3D Gaussians, an extension of traditional 3D Gaussians enriched with supplementary normals, BRDFs and indirect lighting. To reconstruct a relightable scene from multi-view images, we build a novel differentiable rendering pipeline based on the inverse rendering techniques and 3D Gaussian Splatting. Additionally, we introduce an innovative ray tracing scheme specifically designed for scenes represented by discrete points, providing precise visibility for realistic relighting. Quantitative and qualitative experiments confirm that our pipeline successfully reconstructs reasonable normals, materials for discrete point clouds, and achieves commendable accuracy in both novel view synthesis and scene relighting tasks." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Limitations and Future work. Our present pipeline targets the reconstruction of static objects. Certain design choices pose challenges in maintaining its performance with large scale scenes. The high density of points in large scene can slow down optimization as we sample rays and perform PBR at each point. This can be mitigated through deferred rendering technique. Regarding geometry, integrating MVS shows promise for achieving more accurate representations." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.559, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Relightable 3D Gaussians" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.403, + 0.163 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.174, + 0.826, + 0.205 + ], + "angle": 0, + "content": "This work was supported by National Key R&D Program of China (2023YFB3209702), and NSFC (62441204)." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.226, + 0.323, + 0.241 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.254, + 0.787, + 0.282 + ], + "angle": 0, + "content": "1. Aliev, K.A., Sevastopolsky, A., Kolos, M., Ulyanov, D., Lempitsky, V.: Neural point-based graphics. In: ECCV (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.283, + 0.786, + 0.309 + ], + "angle": 0, + "content": "2. Azinovic, D., Li, T.M., Kaplanyan, A., Nießner, M.: Inverse path tracing for joint material and lighting estimation. In: CVPR (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.31, + 0.787, + 0.35 + ], + "angle": 0, + "content": "3. Barron, J.T., Mildenhall, B., Tancik, M., Hedman, P., Martin-Brualla, R., Srinivasan, P.P.: Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In: ICCV (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.351, + 0.787, + 0.377 + ], + "angle": 0, + "content": "4. Barron, J.T., Mildenhall, B., Verbin, D., Srinivasan, P.P., Hedman, P.: Mip-nerf 360: Unbounded anti-aliased neural radiance fields. In: CVPR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.378, + 0.787, + 0.404 + ], + "angle": 0, + "content": "5. Barron, J.T., Mildenhall, B., Verbin, D., Srinivasan, P.P., Hedman, P.: Zip-nerf: Anti-aliased grid-based neural radiance fields. arXiv preprint (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.405, + 0.787, + 0.445 + ], + "angle": 0, + "content": "6. Bi, S., Xu, Z., Srinivasan, P., Mildenhall, B., Sunkavalli, K., Hašan, M., Hold-Geoffroy, Y., Kriegman, D., Ramamoorthi, R.: Neural reflectance fields for appearance acquisition. arXiv preprint (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.446, + 0.787, + 0.487 + ], + "angle": 0, + "content": "7. Bi, S., Xu, Z., Sunkavalli, K., Hašan, M., Hold-Geoffroy, Y., Kriegman, D., Ramamoorthi, R.: Deep reflectance volumes: Relightable reconstructions from multiview photometric images. In: ECCV (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.488, + 0.787, + 0.514 + ], + "angle": 0, + "content": "8. Bi, S., Xu, Z., Sunkavalli, K., Kriegman, D., Ramamoorthi, R.: Deep 3d capture: Geometry and reflectance from sparse multi-view images. In: CVPR (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.515, + 0.787, + 0.54 + ], + "angle": 0, + "content": "9. Boss, M., Braun, R., Jampani, V., Barron, J.T., Liu, C., Lensch, H.: Nerd: Neural reflectance decomposition from image collections. In: ICCV (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.541, + 0.787, + 0.581 + ], + "angle": 0, + "content": "0. Boss, M., Engelhardt, A., Kar, A., Li, Y., Sun, D., Barron, J., Lensch, H., Jampani, V.: Samurai: Shape and material from unconstrained real-world arbitrary image collections. In: NeurIPS (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.582, + 0.787, + 0.609 + ], + "angle": 0, + "content": "1. Burley, B., Studios, W.D.A.: Physically-based shading at disney. In: SIGGRAPH (2012)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.61, + 0.787, + 0.636 + ], + "angle": 0, + "content": "2. Chen, A., Xu, Z., Geiger, A., Yu, J., Su, H.: Tensorf: Tensorial radiance fields. In: ECCV (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.637, + 0.787, + 0.677 + ], + "angle": 0, + "content": "3. Chen, Z., Funkhouser, T., Hedman, P., Tagliasacchi, A.: Mobilenerf: Exploiting the polygon rasterization pipeline for efficient neural field rendering on mobile architectures. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.678, + 0.787, + 0.703 + ], + "angle": 0, + "content": "4. Deng, K., Liu, A., Zhu, J.Y., Ramanan, D.: Depth-supervised nef: Fewer views and faster training for free. In: CVPR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.704, + 0.787, + 0.731 + ], + "angle": 0, + "content": "5. Fridovich-Keil, S., Yu, A., Tancik, M., Chen, Q., Recht, B., Kanazawa, A.: Plenoxels: Radiance fields without neural networks. In: CVPR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.732, + 0.787, + 0.758 + ], + "angle": 0, + "content": "6. Garbin, S.J., Kowalski, M., Johnson, M., Shotton, J., Valentin, J.: Fastnerf: High-fidelity neural rendering at 200fps. In: ICCV (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.759, + 0.787, + 0.799 + ], + "angle": 0, + "content": "7. Guédon, A., Lepetit, V.: Sugar: Surface-aligned gaussian splatting for efficient 3d mesh reconstruction and high-quality mesh rendering. arXiv preprint arXiv:2311.12775 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.8, + 0.787, + 0.84 + ], + "angle": 0, + "content": "8. Guo, K., Lincoln, P., Davidson, P., Busch, J., Yu, X., Whalen, M., Harvey, G., Orts-Escalano, S., Pandey, R., Dourgarian, J., et al.: The relighttables: Volumetric performance capture of humans with realistic relighting. ACM TOG (2019)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.254, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.356, + 0.127 + ], + "angle": 0, + "content": "J. Gao et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.148, + 0.785, + 0.175 + ], + "angle": 0, + "content": "19. Hasselgren, J., Hofmann, N., Munkberg, J.: Shape, light, and material decomposition from images using monte carlo rendering and denoising. In: NeurIPS (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.178, + 0.785, + 0.218 + ], + "angle": 0, + "content": "20. Hasselgren, J., Hofmann, N., Munkberg, J.: Shape, Light, and Material Decomposition from Images using Monte Carlo Rendering and Denoising. arXiv:2206.03380 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.22, + 0.785, + 0.246 + ], + "angle": 0, + "content": "21. Hedman, P., Srinivasan, P.P., Mildenhall, B., Barron, J.T., Debevec, P.: Baking neural radiance fields for real-time view synthesis. In: ICCV (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.248, + 0.785, + 0.274 + ], + "angle": 0, + "content": "22. Hu, T., Liu, S., Chen, Y., Shen, T., Jia, J.: Efficientnerf efficient neural radiance fields. In: CVPR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.276, + 0.785, + 0.303 + ], + "angle": 0, + "content": "23. Jin, H., Liu, I., Xu, P., Zhang, X., Han, S., Bi, S., Zhou, X., Xu, Z., Su, H.: Tensoir: Tensorial inverse rendering. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.305, + 0.785, + 0.331 + ], + "angle": 0, + "content": "24. Kajiya, J.T.: The rendering equation. In: Proceedings of the 13th annual conference on Computer graphics and interactive techniques (1986)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.333, + 0.785, + 0.374 + ], + "angle": 0, + "content": "25. Karras, T.: Maximizing parallelism in the construction of bvhs, octrees, and k-d trees. In: Proceedings of the Fourth ACM SIGGRAPH/Eurographics conference on High-Performance Graphics (2012)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.376, + 0.785, + 0.402 + ], + "angle": 0, + "content": "26. Kerbl, B., Kopanas, G., Leimkuhler, T., Drettakis, G.: 3d gaussian splatting for real-time radiance field rendering. ACM TOG (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.404, + 0.785, + 0.43 + ], + "angle": 0, + "content": "27. Keselman, L., Hebert, M.: Flexible techniques for differentiable rendering with 3d gaussians. arXiv preprint (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.432, + 0.785, + 0.459 + ], + "angle": 0, + "content": "28. Kuang, Z., Olszewski, K., Chai, M., Huang, Z., Achlioptas, P., Tulyakov, S.: Neroic: Neural rendering of objects from online image collections. ACM TOG (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.461, + 0.717, + 0.473 + ], + "angle": 0, + "content": "29. Levoy, M., Whitted, T.: The use of points as a display primitive (1985)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.475, + 0.785, + 0.515 + ], + "angle": 0, + "content": "30. Liu, Y., Wang, P., Lin, C., Long, X., Wang, J., Liu, L., Komura, T., Wang, W.: Nano: Neural geometry and brdf reconstruction of reflective objects from multiview images. In: SIGGRAPH (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.517, + 0.785, + 0.557 + ], + "angle": 0, + "content": "31. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. In: ECCV (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.559, + 0.785, + 0.585 + ], + "angle": 0, + "content": "32. Müller, T., Evans, A., Schied, C., Keller, A.: Instant neural graphics primitives with a multiresolution hash encoding. ACM TOG (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.587, + 0.785, + 0.627 + ], + "angle": 0, + "content": "33. Munkberg, J., Hasselgren, J., Shen, T., Gao, J., Chen, W., Evans, A., Müller, T., Fidler, S.: Extracting triangular 3d models, materials, and lighting from images. In: CVPR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.629, + 0.785, + 0.656 + ], + "angle": 0, + "content": "34. Nam, G., Lee, J.H., Gutierrez, D., Kim, M.H.: Practical svbrdf acquisition of 3d objects with unstructured flash photography. ACM TOG (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.658, + 0.785, + 0.684 + ], + "angle": 0, + "content": "35. Park, J.J., Holynski, A., Seitz, S.M.: Seeing the world in a bag of chips. In: CVPR (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.686, + 0.785, + 0.727 + ], + "angle": 0, + "content": "36. Pfister, H., Zwicker, M., Van Baar, J., Gross, M.: Surfels: Surface elements as rendering primitives. In: Proceedings of the 27th annual conference on Computer graphics and interactive techniques (2000)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.728, + 0.785, + 0.754 + ], + "angle": 0, + "content": "37. Pittaluga, F., Koppal, S.J., Kang, S.B., Sinha, S.N.: Revealing scenes by inverting structure from motion reconstructions. In: CVPR (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.756, + 0.785, + 0.783 + ], + "angle": 0, + "content": "38. Rakhimov, R., Ardelean, A.T., Lempitsky, V., Burnaev, E.: Npbg++: Accelerating neural point-based graphics. In: CVPR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.785, + 0.785, + 0.812 + ], + "angle": 0, + "content": "39. Reiser, C., Peng, S., Liao, Y., Geiger, A.: Kilonerf: Speeding up neural radiance fields with thousands of tiny mlps. In: ICCV (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.813, + 0.785, + 0.84 + ], + "angle": 0, + "content": "40. Schmitt, C., Donne, S., Riegler, G., Koltun, V., Geiger, A.: On joint estimation of pose, geometry and svbrdf from a handheld scanner. In: CVPR (2020)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.148, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.559, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "Relightable 3D Gaussians" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.189 + ], + "angle": 0, + "content": "41. Srinivasan, P.P., Deng, B., Zhang, X., Tancik, M., Mildenhall, B., Barron, J.T.: Nerv: Neural reflectance and visibility fields for relighting and view synthesis. In: CVPR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.19, + 0.788, + 0.231 + ], + "angle": 0, + "content": "42. Verbin, D., Hedman, P., Mildenhall, B., Zickler, T., Barron, J.T., Srinivasan, P.P.: Ref-nerf: Structured view-dependent appearance for neural radiance fields. In: CVPR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.231, + 0.788, + 0.272 + ], + "angle": 0, + "content": "43. Wang, P., Liu, L., Liu, Y., Theobalt, C., Komura, T., Wang, W.: Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. arXiv preprint (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.273, + 0.788, + 0.3 + ], + "angle": 0, + "content": "44. Wu, H., Hu, Z., Li, L., Zhang, Y., Fan, C., Yu, X.: Nefii: Inverse rendering for reflectance decomposition with near-field indirect illumination. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.3, + 0.788, + 0.342 + ], + "angle": 0, + "content": "45. Yao, Y., Zhang, J., Liu, J., Qu, Y., Fang, T., McKinnon, D., Tsin, Y., Quan, L.: Neilf: Neural incident light field for physically-based material estimation. In: ECCV (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.342, + 0.788, + 0.37 + ], + "angle": 0, + "content": "46. Yariv, L., Gu, J., Kasten, Y., Lipman, Y.: Volume rendering of neural implicit surfaces. In: NeurIPS (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.37, + 0.788, + 0.398 + ], + "angle": 0, + "content": "47. Yifan, W., Serena, F., Wu, S., Öztireli, C., Sorkine-Hornung, O.: Differentiable surface splatting for point-based geometry processing. ACM TOG (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.398, + 0.788, + 0.424 + ], + "angle": 0, + "content": "48. Yu, Z., Peng, S., Niemeyer, M., Sattler, T., Geiger, A.: Monosdf: Exploring monocular geometric cues for neural implicit surface reconstruction. In: NeurIPS (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.424, + 0.788, + 0.467 + ], + "angle": 0, + "content": "49. Zhang, J., Yao, Y., Li, S., Liu, J., Fang, T., McKinnon, D., Tsin, Y., Quan, L.: Neil++: Inter-reflectable light fields for geometry and material estimation. In: ICCV (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.467, + 0.788, + 0.494 + ], + "angle": 0, + "content": "50. Zhang, K., Luan, F., Li, Z., Snavely, N.: Iron: Inverse rendering by optimizing neural sdfs and materials from photometric images. In: CVPR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.494, + 0.788, + 0.535 + ], + "angle": 0, + "content": "51. Zhang, K., Luan, F., Wang, Q., Bala, K., Snavely, N.: Physg: Inverse rendering with spherical gaussians for physics-based material editing and relighting. In: CVPR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.535, + 0.788, + 0.563 + ], + "angle": 0, + "content": "52. Zhang, Q., Baek, S.H., Rusinkiewicz, S., Heide, F.: Differentiable point-based radiance fields for efficient view synthesis. In: SIGGRAPH Asia (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.564, + 0.788, + 0.605 + ], + "angle": 0, + "content": "53. Zhang, X., Srinivasan, P.P., Deng, B., Debevec, P., Freeman, W.T., Barron, J.T.: Nerfactor: Neural factorization of shape and reflectance under an unknown illumination. ACM TOG (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.605, + 0.788, + 0.633 + ], + "angle": 0, + "content": "54. Zhang, Y., Huang, X., Ni, B., Li, T., Zhang, W.: Frequency-modulated point cloud rendering with easy editing. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.633, + 0.788, + 0.66 + ], + "angle": 0, + "content": "55. Zhang, Y., Xu, T., Yu, J., Ye, Y., Jing, Y., Wang, J., Yu, J., Yang, W.: Nemf: Inverse volume rendering with neural microflake field. In: ICCV (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.66, + 0.788, + 0.688 + ], + "angle": 0, + "content": "56. Zhang, Y., Sun, J., He, X., Fu, H., Jia, R., Zhou, X.: Modeling indirect illumination for inverse rendering. In: CVPR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.688, + 0.788, + 0.729 + ], + "angle": 0, + "content": "57. Zwicker, M., Pfister, H., Van Baar, J., Gross, M.: Surface splatting. In: Proceedings of the 28th annual conference on computer graphics and interactive techniques (2001)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.729, + 0.788, + 0.757 + ], + "angle": 0, + "content": "58. Zwicker, M., Pfister, H., Van Baar, J., Gross, M.: Ewa splatting. IEEE TVCG (2002)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.757 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/Relightable 3D Gaussians_ Realistic Point Cloud Relighting with BRDF Decomposition and Ray Tracing/8f8cf818-147d-408c-857b-28f5e4e05e99_origin.pdf b/2024/Relightable 3D Gaussians_ Realistic Point Cloud Relighting with BRDF Decomposition and Ray Tracing/8f8cf818-147d-408c-857b-28f5e4e05e99_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..63024814a6fa56bdc6342a74b6450e3d79a1e56d --- /dev/null +++ b/2024/Relightable 3D Gaussians_ Realistic Point Cloud Relighting with BRDF Decomposition and Ray Tracing/8f8cf818-147d-408c-857b-28f5e4e05e99_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:98dd8c66d73114daf0aa373268ef59c4d98e9d1db494a095d54fe5029eba316c +size 8614565 diff --git a/2024/Relightable 3D Gaussians_ Realistic Point Cloud Relighting with BRDF Decomposition and Ray Tracing/full.md b/2024/Relightable 3D Gaussians_ Realistic Point Cloud Relighting with BRDF Decomposition and Ray Tracing/full.md new file mode 100644 index 0000000000000000000000000000000000000000..eceed01077e94251d76a76ed46927463cf06eaa6 --- /dev/null +++ b/2024/Relightable 3D Gaussians_ Realistic Point Cloud Relighting with BRDF Decomposition and Ray Tracing/full.md @@ -0,0 +1,399 @@ +# Relightable 3D Gaussians: Realistic Point Cloud Relighting with BRDF Decomposition and Ray Tracing + +Jian $\mathrm{Gao}^{1*}$ , Chun $\mathrm{Gu}^{2*}$ , Youtian Lin $^{1}$ , Zhihao Li $^{3}$ , Hao Zhu $^{1}$ , Xun $\mathrm{Cao}^{1}$ , Li Zhang $^{2\boxtimes}$ , and Yao $\mathrm{Yao}^{1\boxtimes}$ + +$^{1}$ Nanjing University + +$^{2}$ Fudan University + +3 Huawei Noah's Ark Lab + +Abstract. In this paper, we present a novel differentiable point-based rendering framework to achieve photo-realistic relighting. To make the reconstructed scene relightable, we enhance vanilla 3D Gaussians by associating extra properties, including normal vectors, BRDF parameters, and incident lighting from various directions. From a collection of multiview images, the 3D scene is optimized through 3D Gaussian Splatting while BRDF and lighting are decomposed by physically based differentiable rendering. To produce plausible shadow effects in photo-realistic relighting, we introduce an innovative point-based ray tracing with the bounding volume hierarchies for efficient visibility pre-computation. Extensive experiments demonstrate our improved BRDF estimation, novel view synthesis and relighting results compared to state-of-the-art approaches. The proposed framework showcases the potential to revolutionize the mesh-based graphics pipeline with a point-based pipeline enabling editing, tracing, and relighting. + +Keywords: 3D Gaussian Splatting, Relighting, Point based rendering + +# 1 Introduction + +Reconstructing 3D scenes from multi-view images for photo-realistic rendering is a fundamental problem at the intersection of computer vision and graphics. Recently, 3D Gaussian Splatting (3DGS) [26] has been proposed and has gained significant attention from the community. The method employs a set of 3D Gaussian points to represent a 3D scene and projects these points onto a designated view through a tile-based rasterization. Attributes of each 3D Gaussian point are then optimized through the point-based differentiable rendering. Notably, 3DGS achieves real-time rendering with quality comparable to or even surpassing previous state-of-the-art methods (e.g., Mip-NeRF [3]), with a training speed on par with the most efficient Instant-NGP [32]. However, the current 3DGS is unable to reconstruct a scene that can be relighted under different lighting conditions, + +![](images/db90ca5bf82a52b5d6d04ed8397b42380b6f5d0b08db17516332ec2f153721dc.jpg) +(a) Point Cloud + +![](images/77fad46f9adc611e9ba0cf90993ce1f1e0aa703b8fead9cdd503726d1c334d57.jpg) +(b) Normal + +![](images/4764d770bb5551e8d11a38b78daafcf236583b1ee34141e43ba563b3f00de2e1.jpg) +(c) Ambient Occlusion + +![](images/0b357a25004a1d765e431fb52522809bbd1a4e14a34ff10c8b10278a46435edf.jpg) +(d) Physically Based Rendering +Fig. 1: Visual results of our pipeline on a multi-object composition scene. In our pipeline, we represent a scene as Relightable 3D Gaussians. From multi-view images, we recover the geometry and materials of individual objects with inverse rendering techniques (see Sec. 3). Then, objects are easily composited into a new scene, thanks to our explicit representation. After that, we solve the complex occlusions though point based ray tracing (see Sec. 4) and re-light the new scene. Ultimately, we achieve high-fidelity relighting with remarkably realistic shadow. + +making the method only applicable to the task of novel view synthesis. In addition, ray tracing, a crucial component in achieving realistic rendering, remains an unresolved challenge in point-based representation, limiting 3DGS to more dedicated rendering effects such as shadowing and light reflectance. + +In this paper, we aim to reconstruct a re烈table 3D Gaussian point cloud from multi-view images with a differentiable rendering framework and achieve realistic relighting. We make 3D Gaussian relightable by assigning it with additional normal, BRDF properties, and incident light information to model perpoint light reflectance. In contrast to the plain alpha blending in the original 3DGS [26], we apply physically based rendering (PBR) to get a PBR color for each 3D Gaussian point, which is then alpha-composed to obtain a rendered color for the corresponding image pixel. For robust material and lighting estimation, we split the incident light into a global environment map and an indirect incident light field. To capture accurate visibility for each 3D Gaussian, we propose a novel ray tracing method based on the bounding volume hierarchy (BVH), which enables efficient visibility pre-computing for real-time rendering. Additionally, the proposed ray tracing method can handle complex occlusion relationships in a novel multi-object composition scene, thus realizing realistic shadow effects. Moreover, proper regularizations are introduced to enhance the geometry and mitigate the material-lighting ambiguity during the optimization, including constraints on depth distribution, smoothness priors, and a lighting regularization. + +Extensive experiments conducted across diverse datasets demonstrate the improved material and lighting estimation results and novel view rendering quality. Additionally, we illustrate the relightable and editable capabilities of our system through multi-object composition in a novel lighting environment. To summarize, major contributions of the paper include: + +- We propose a material and lighting decomposition scheme for 3D Gaussian Splatting, where a normal vector, BRDF values, and incident lights are assigned and optimized for each 3D Gaussian point. +- We introduce a novel point-based ray tracing approach based on bounding volume hierarchy, enabling efficient visibility pre-computing for each 3D Gaussian point and rendering of a 3D scene with realistic shadow effect. +- We demonstrate a comprehensive graphics pipeline solely based on a discretized point representation, supporting relighting, editing, and ray tracing of a reconstructed 3D point cloud. + +# 2 Related Works + +Neural Radiance Field. Differentiable rendering techniques, exemplified by Neural Radiance Field (NeRF) [31], have garnered significant attention [3,12,32]. NeRF utilizes an implicit Multi-Layer Perceptron (MLP) that takes 3D positions and viewing directions as inputs to generate density and view-dependent colors for differentiable volume rendering. Despite its powerful neural implicit representation, both speed and quality can be improved. Efficiency improvements mainly focus on optimizing queries on neural fields and minimizing queries [12, 13, 15, 16, 21, 22, 32, 39]. Quality enhancements involve anti-aliasing [3-5] or utilizing exterior supervision [14, 48], etc. + +Differentiable Point Based Rendering. Using points as rendering primitives was first proposed in [29]. To address the issue of resulting holey images when rasterizing discrete points directly, solutions fall into two categories. One approach directly employs points as rendering primitives and encodes the geometric and photometric features near the point using SIFT descriptors [37] or neural descriptors [1,38]. This leads to a feature image with gaps, which is then decoded to recover a hole-free RGB image. The other category models a point as a primitive occupying a specific space, like a surfel [36,47] or a 3D Gaussian [26]. The rendered image is then generated using a splatting technique. This technique saw significant development in the 2000s [36,57,58], and in the era of differentiable rendering, PointRF [52], DSS [47] and 3DGS [26] serve as notable examples. We assert that the point serves as a promising rendering primitive owing to its inherent ease of editing, as also corroborated in [54]. + +Material and Lighting Estimation. Decomposing the materials and illumination of a scene from multi-view images presents a formidable challenge owing to its intrinsic high-dimensional complexity. Some methods simplified the problem under the assumption of a controllable light [2,6-8,18,34,35,40]. Subsequent research explores more complex lighting models to cope with realistic scenarios. NeRV [41] and PhySG [51] leverage an environmental map to manage arbitrary + +lighting conditions. NeRD [9] addresses the challenge posed by images captured under varying illumination by attributing Spherical Gaussians to each image. IRON [50] introduces an innovative edge sampling algorithm tailored for neural SDFs. NeRFactor [53] exploits light visibility to achieve superior material and lighting decomposition. Further studies [19, 52] address the consideration of indirect lighting, leading to enhanced BRDF estimation quality. [20, 33] utilize differentiable marching tetrahedrons for direct optimization on mesh surfaces. [10, 28] deal with varying cameras, illumination, and backgrounds. Ref-NeRF [42] suggests the employment of integrated direction encoding to ameliorate the reconstruction fidelity of reflective objects. [30, 33] use split-sum approximation to approximate the shading effects. NeMF [55] represents the scene as a microflake volume. NeFII [44] integrates lights through path tracing with Monte Carlo sampling. InvRender [56] computes indirect illumination by directly leveraging the neural radiance field, rather than concurrently estimating it alongside the decomposition of direct lighting and materials. TensoIR [23] performs inverse rendering based on tensor factorization and neural fields. NeILF [45] and NeILF++ [49] expresses the incident lights as a neural incident light field, while NeILF++ integrates VolSDF [46] with NeILF through inter-reflection. However, existing schemes rarely delve into BRDF estimation for point clouds, and the estimation and rendering quality remain to be improved. + +# 3 Relightable 3D Gaussians + +In this section, we introduce a novel pipeline to decompose materials and lighting from multi-view images based on 3D Gaussian Splatting (3DGS) [26]. An overview of our pipeline is shown in Fig. 2. + +# 3.1 Preliminary + +Distinct from the widely adopted Neural Radiance Field, 3DGS [26] employs explicit 3D Gaussian points as its rendering primitives. A 3D Gaussian point is mathematically defined as: + +$$ +G (\boldsymbol {x}) = \exp \left(- \frac {1}{2} (\boldsymbol {x} - \boldsymbol {\mu}) ^ {\top} \Sigma^ {- 1} (\boldsymbol {x} - \boldsymbol {\mu})\right), \tag {1} +$$ + +where $\pmb{\mu}$ and $\Sigma$ denote the 3D spatial mean and covariance matrix, respectively. Each Gaussian is also equipped with an opacity $o$ and a view-dependent color $\pmb{c}$ . + +The rendering process in 3DGS is divided into two main steps. Firstly, 3D Gaussians are projected to 2D Gaussians on the image plane. The 2D means are determined by accurate projection of 3D means, while the 2D covariance matrices are approximated by: $\Sigma' = J\boldsymbol{W} \Sigma \boldsymbol{W}^\top \boldsymbol{J}^\top$ , where $\boldsymbol{W}$ and $\boldsymbol{J}$ denote the viewing transformation and the Jacobian of the affine approximation of perspective projection transformation [57]. Subsequently, the pixel color is derived by alpha blending $N$ ordered 2D Gaussians from front to back: + +$$ +\mathcal {C} = \sum_ {i \in N} T _ {i} \alpha_ {i} \boldsymbol {c} _ {i}, \quad T _ {i} = \prod_ {j = 1} ^ {i - 1} (1 - \alpha_ {j}). \tag {2} +$$ + +![](images/f0101fd25b550b2fd16a3b6c60da2720ceae47ec9045536a575998c4f6bd5039.jpg) +Fig. 2: The proposed differentiable rendering pipeline. Starting with a collection of 3D Gaussians that embody geometry, materials, and lighting attributes, we first execute the physically based rendering equation for each 3D Gaussian to determine its outgoing radiance, denoted as $c'$ . Following this, we perform rasterization and alpha blending to obtain vanilla color map $\mathcal{C}$ , PBR color map $\mathcal{C}'$ , depth map $\mathcal{D}$ , normal map $\mathcal{N}$ , etc. To optimize relightable 3D Gaussians, we utilize the ground truth image $\mathcal{C}_{gt}$ and the pseudo normal map derived from $\mathcal{D}$ for supervision. + +Here, $\alpha$ is obtained by multiplying the opacity $o$ with the 2D covariance's contribution computed from $\Sigma^{\prime}$ and pixel coordinate in image space [26]. In implementation details, the covariance matrix $\Sigma$ is parameterized as a unit quaternion $\pmb{q}$ and a 3D scaling vector $\pmb{s}$ to maintain its meaningful interpretation throughout optimization. Additionally, view-dependent color $\pmb{c}_i$ is represented through a set of Spherical Harmonics (SH). + +In summation, a 3D scene is represented by a collection of 3D Gaussians, with the $i^{th}$ Gaussian $\mathcal{P}_i$ parameterized as $\{\pmb{\mu}_i, \pmb{q}_i, \pmb{s}_i, o_i, \pmb{c}_i\}$ . + +# 3.2 Geometry Enhancement + +Scene geometry, specifically surface normal, is essential for realistic physically based rendering, which will be discussed in Sec. 3.3. + +Normal Estimation. We incorporate a normal attribute $\pmb{n}$ for each 3D Gaussian and try to solve it. A potential methodology treats the spatial mean of a 3D Gaussian as a conventional point, and executes normal estimation based on the local planar assumption. However, this approach cannot provide accurate normal estimation. The reasons are two fold: First, the reconstructed Gaussian point cloud is often sparse. More critically, the Gaussian points are naturally soft, which means these points are not precisely aligned with the object surface. + +To address these limitations, we propose to optimize $\pmb{n}$ from an initial random vector via back-propagation for each 3D Gaussian. We perceive the depth of all Gaussians along a ray as a distribution, and estimate the pixel depth as the expectation of this distribution. Similarly, we determine the normal for each + +pixel. This process is described by: + +$$ +\{\mathcal {D}, \mathcal {N} \} = \sum_ {i \in N} w _ {i} \left\{d _ {i}, \boldsymbol {n} _ {i} \right\}, \tag {3} +$$ + +where $d_{i}$ , $\pmb{n}_i$ and $w_{i} = T_{i}\alpha_{i} / \sum_{i\in N}T_{i}\alpha_{i}$ denote the depth, normal and weight of the point. We then encourage the consistency between the rendered normal $\mathcal{N}$ and the pseudo normal $\tilde{\mathcal{N}}$ , which is computed from the rendered depth $\mathcal{D}$ under the local planarity assumption. The normal consistency is quantified as follows: + +$$ +\mathcal {L} _ {n} = \| \mathcal {N} - \tilde {\mathcal {N}} \| _ {2}. \tag {4} +$$ + +Normal Gradient Based Densification. To achieve superior rendering fidelity in detail regions, vanilla 3DGS [26] densifies 3D Gaussians through the gradient of view space points. Drawing upon this, to improve normal recovery in thin regions, we introduce an additional densification criterion on the gradient of normals. Specifically, we densify Gaussians whose normal gradient exceeds a threshold $T_{n}$ . + +Constraint on Depth Distribution. Given our assumption of the object possessing an accurate surface, we enforce a constraint on the depth distribution by minimizing the uncertainty. The uncertainty is defined as: + +$$ +\mathcal {L} _ {u} = \mathcal {D} _ {s q} - \mathcal {D} ^ {2}, \tag {5} +$$ + +where $\mathcal{D}_{sq} = \sum_{i\in N}w_id_i^2$ , and $\mathcal{D}$ is the depth estimation defined in Eq. 3. This constraint on the depth distribution drives Gaussian points to the object surface, thereby improving geometric reconstruction. + +Object Mask Constraint. If there is a mask indicating the object, we can constrain the optimization by a binary cross entropy loss [43]: + +$$ +\mathcal {L} _ {O} = - M \log O - (1 - M) \log (1 - O), \tag {6} +$$ + +where $M$ is the object mask and $O = \sum_{i\in N}T_i\alpha_i$ . With this constraint, $O$ is forced to be aligned with the distribution of $M$ , and so we get opaque surface. + +# 3.3 BRDF and Light Modeling + +Rendering Equation. We use the rendering equation [24] to model light interaction with surfaces, accounting for their material properties and geometry. It is given by: + +$$ +L _ {o} \left(\boldsymbol {\omega} _ {o}, \boldsymbol {x}\right) = \int_ {\Omega} f \left(\boldsymbol {\omega} _ {o}, \boldsymbol {\omega} _ {i}, \boldsymbol {x}\right) L _ {i} \left(\boldsymbol {\omega} _ {i}, \boldsymbol {x}\right) \left(\boldsymbol {\omega} _ {i} \cdot \boldsymbol {n}\right) d \boldsymbol {\omega} _ {i}, \tag {7} +$$ + +where $\pmb{x}$ and $\pmb{n}$ are the surface point and its normal vector, $f$ is the Bidirectional Reflectance Distribution Function (BRDF), and $L_{i}$ and $L_{o}$ denote the incoming and outgoing radiance in directions $\omega_{i}$ and $\omega_{o}$ . $\Omega$ signifies the hemispherical domain above the surface. + +Prior methods [49, 51] typically begin with the acquisition of intersection points between rays and surface through differentiable rendering. Then they apply the rendering equation at these points to facilitate Physical Based Rendering (PBR). However, these approach present significant challenges in the point-based framework. Certainly, it is feasible to extract surfaces in 3DGS [17]; but, it necessitates a coordinate-based incident light field for inverse PBR. The querying of this field at millions of points during every iteration imposes a substantial computational burden. Moreover, it is pertinent to acknowledge that accurately extracting geometric surfaces presents considerable challenges in 3DGS system. + +To tackle this issue, we propose to compute PBR color $\{c_i^{\prime}\}_{i = 0}^{N}$ for each 3D Gaussian, and then obtain the PBR image $\mathcal{C}'$ through alpha-blending, as shown in Fig. 2. This method is more efficient for two main reasons. First, PBR is performed on fewer 3D Gaussians rather than image pixels in all input views, as each Gaussian affects multiple pixels based on its scale. Second, by allocating discrete attributes for each Gaussian, we avoid the need for global neural fields. + +BRDF Parameterization. To make 3D Gaussians relightable, we assign BRDF properties to each Gaussian and adopt a simplified Disney BRDF model [11]. The BRDF properties include an albedo $\pmb{b} \in [0,1]^3$ and a roughness $r \in [0,1]$ , and the BRDF $f$ in Eq. 7 is divided into a diffuse term $f_{d} = \frac{b}{\pi}$ and a specular term: + +$$ +f _ {s} \left(\boldsymbol {\omega} _ {o}, \boldsymbol {\omega} _ {i}\right) = \frac {D (\boldsymbol {h} ; r) \cdot F \left(\boldsymbol {\omega} _ {o} , \boldsymbol {h}\right) \cdot G \left(\boldsymbol {\omega} _ {i} , \boldsymbol {\omega} _ {o} , h ; r\right)}{\left(\boldsymbol {n} \cdot \boldsymbol {\omega} _ {i}\right) \cdot \left(\boldsymbol {n} \cdot \boldsymbol {\omega} _ {o}\right)}, \tag {8} +$$ + +where $\pmb{h} = (\pmb{\omega}_i + \pmb{\omega}_o) / 2$ is the half vector, $D, F$ and $G$ denote the normal distribution function, Fresnel term and geometry term. + +Incident Light Modeling. Optimizing a NeILF for each Gaussian can be excessively unconstrained, leading to challenges in accurately decomposing incident lights from its appearance. To address this issue, we apply a prior by partitioning the incident light into globally shared direct components and individual per-Gaussian indirect components. The sampled incident light at a Gaussian from direction $\omega_{i}$ is represented as: + +$$ +L _ {i} (\boldsymbol {\omega} _ {i}) = V (\boldsymbol {\omega} _ {i}) \cdot L _ {\text {d i r e c t}} (\boldsymbol {\omega} _ {i}) + L _ {\text {i n d i r e c t}} (\boldsymbol {\omega} _ {i}), \tag {9} +$$ + +where $V(\omega_{i})$ represents the visibility term which will be further discussed in Sec. 4, the indirect light term $L_{\text{indirect}}$ is parameterized by 3-level SH, denoted as $l$ , while the direct light term $L_{\text{direct}}$ is parameterized as a globally shared 16x32 environment map, denoted as $l^{env}$ . Despite the utilization of relatively low-level Spherical Harmonics (SH), we can still capture some inter-reflection effects, as the rendering color of a given pixel arises from the collaboration of multiple relightable 3D Gaussians. + +For each 3D Gaussian, we sample $N_{s}$ incident directions over the hemisphere space through Fibonacci sampling [45] to provide numerical integration for Eq. 7. The PBR color of each Gaussian is then given by: + +$$ +\boldsymbol {c} ^ {\prime} \left(\boldsymbol {\omega} _ {o}\right) = \sum_ {i = 0} ^ {N _ {s}} \left(f _ {d} + f _ {s} \left(\boldsymbol {\omega} _ {o}, \boldsymbol {\omega} _ {i}\right)\right) L _ {i} \left(\boldsymbol {\omega} _ {i}\right) \left(\boldsymbol {\omega} _ {i} \cdot \boldsymbol {n}\right) \Delta \boldsymbol {\omega} _ {i}, \tag {10} +$$ + +where $\omega_{i}$ is the solid angle. + +To summarize, our method represents a 3D scene as a set of relightable 3D Gaussians and a global environment light $l^{env}$ , where the $i^{th}$ Gaussian $\mathcal{P}_i$ is parameterized as $\{\pmb{\mu}_i, \pmb{q}_i, \pmb{s}_i, o_i, \pmb{c}_i, \pmb{n}_i, \pmb{b}_i, r_i, l_i\}$ . + +# 3.4 Regularizations + +To mitigate the materials-lighting ambiguity [45], proper regularizations are utilized to facilitate their plausible decomposition. + +Light Regularization. We apply a light regularization assuming a near-natural white incident light [30]: + +$$ +\mathcal {L} _ {\text {l i g h t}} = \sum_ {c} \left(L _ {c} - \frac {1}{3} \sum_ {c} L _ {c}\right), c \in \{R, G, B \}. \tag {11} +$$ + +Smoothness Priors. We expect that the BRDF properties do not change drastically in homogeneous areas [45]. We define a smooth constraint on roughness as: + +$$ +\mathcal {L} _ {s, r} = \| \nabla R \| \exp (- \| \nabla C _ {g t} \|), \tag {12} +$$ + +where $R = \sum_{i\in N}w_i r_i$ is the rendered roughness map. Similarly, we also define smoothness constraints $\mathcal{L}_{s,n}$ and $\mathcal{L}_{s,b}$ on normal and albedo. + +# 4 Point-based Ray Tracing + +For real-time realistic rendering with plausible shadow effects on relightable 3D Gaussians, we introduce a novel point-based ray tracing approach in this section. + +# 4.1 Ray Tracing on 3D Gaussians + +Our proposed ray tracing technique on 3D Gaussians is built upon the Bounding Volume Hierarchy (BVH), enabling efficient visibility querying along a ray. Our method adopts the idea from [25], an in-place algorithm for constructing a binary radix tree that maximizes parallelism and facilitates real-time BVH construction. Specifically, we construct a binary radix tree from a given set of 3D Gaussians, where each leaf node represents the tight bounding box of a Gaussian, and each internal node denotes the bounding box of its two children. + +Unlike ray tracing with opaque polygonal meshes, where only the closest intersection point is required, ray tracing with semi-transparent Gaussians necessitates accounting for all Gaussians potentially influencing the ray's transmittance. The process of ray tracing on 3D Gaussian points can be described as follows: the ray travels from the camera center and accumulates transmittance as it passes through 3D Gaussian points until the transmittance is zero. The first key issue to be addressed is how to define the contribution of a single 3D Gaussian to the transmittance. As discussed in Sec.3.1, transforming a 3D Gaussian to a 2D Gaussian involves an approximation, which complicating the + +![](images/7dd81d7ea48234217c81c0676ca8bebe9540cb52587a70d919390505ba9f1bda.jpg) +(a) Root Node + +![](images/1ea1a93651801d076c5cd91dc0dd289887767b15d82ac2bfaec1c3d2d36157c0.jpg) +(b) Child Node + +![](images/71c5c1a930d868cf6692fce5aa54ed6ad19c3852eea3f3ba3c145f0051dbcc82.jpg) +(c) Gaussian +Fig. 3: Intersection tests in point-based ray tracing. Intersection point between ray and Gaussian is obtained by three steps: (a) intersect the BVH root node; (b) dive into the intersected child nodes recursively until the leaf node; (c) perform Eq. 13 to get the equivalent intersection point. (d) shows that a 3D Gaussian actually has non-negligible effect on a segment of a ray. + +![](images/f366c2a3d2b2fb588bf8a5819989d4f17b66aaf33f9e9f32e9464b3a32440977.jpg) +(d) Gaussian + +identification of accurate intersection between the Gaussian and a ray. As shown in Fig. 3d, a 3D Gaussian actually has non-negligible effect on a segment of a ray. Thus, drawing inspiration from [27], we approximate the intersection of ray with 3D Gaussian as a point where the 3D Gaussian's contribution peaks, as showed in Fig. 3(3). The equivalent intersection point is defined as: + +$$ +\boldsymbol {r} _ {\boldsymbol {x}} = \boldsymbol {r} _ {\boldsymbol {o}} + t _ {j} \boldsymbol {r} _ {\boldsymbol {d}}, \tag {13} +$$ + +where $\pmb{r_o}$ denotes the origin, $\pmb{r_d}$ is direction of the ray, corresponding to the previously mentioned $\omega_{i}, t_{j}$ is defined as: + +$$ +t _ {j} = \frac {\left(\boldsymbol {\mu} - \boldsymbol {r} _ {o}\right) ^ {T} \boldsymbol {\Sigma} \boldsymbol {r} _ {d}}{\boldsymbol {r} _ {d} ^ {T} \boldsymbol {\Sigma} \boldsymbol {r} _ {d}}. \tag {14} +$$ + +Subsequently, we approximate the contribution of this 3D Gaussian to the ray as $\alpha_{j}$ at the equivalent intersection point $\pmb{r}_{\pmb{x}}$ + +Considering the transmittance along a ray: $T_{i} = \prod_{j=1}^{i-1}(1 - \alpha_{j})$ , it is evident that the order of $\alpha_{j}$ does not affect $T_{i}$ . In other words, the order in which Gaussians are encountered along a ray does not impact the overall transmittance. As illustrated in Fig. 3, starting from the root node of the binary radix tree, intersection tests are recursively performed between the ray and the bounding volumes of each node's children. Upon reaching a leaf node, the associated Gaussian is identified. Through this traversal, the transmittance $T$ is progressively attenuated: + +$$ +T _ {i} = \left(1 - \alpha_ {i - 1}\right) T _ {i - 1}, \quad \text {f o r} i = 1, \dots , j - 1 \text {w i t h} T _ {1} = 1. \tag {15} +$$ + +To speed up ray tracing, the process is early terminated if a ray's transmittance drops below a certain threshold $T_{\min}$ . + +# 4.2 Visibility Pre-computation + +In Sec. 3.3, a potential ambiguity in the decomposition of indirect and direct lights has been discerned. While the essential visibility term can be derived via the proposed ray tracing on 3D Gaussians (Sec. 4.1), querying the visibility term through ray tracing online proves challenging due to the computational complexity. Nevertheless, given our exclusive focus on static scenes, we can pre-compute the visibility term $V(\omega_{i})$ across the hemispherical domain determined by $\pmb{n}$ for each Gaussian, and subsequently integrate it into the rendering equation. + +Hence, we divide the optimization process into two stages. In the first stage, we optimize a vanilla 3D Gaussian point cloud, augmented with an additional normal vector $\pmb{n}$ for each Gaussian. Immediately afterwards, we pre-compute the per-Gaussian visibility through the proposed ray tracing. In the second stage, we lock the geometry of 3D Gaussians and focus solely on optimizing the material and lighting parameters using the pipeline described in Fig. 2. + +# 4.3 Point Based Relighting Pipeline + +Based on our relightable 3D Gaussians, we devise a point-based graphics pipeline that seamlessly integrates effortless scene editing and realistic relighting. To our knowledge, there is currently no point-based rendering pipeline that effectively accomplishes both tasks. Although explicit point representations make scene editing easy, achieving photo-realistic relighting proved nearly insurmountable in existing point-based rendering pipeline. Conversely, in inverse rendering approaches based on implicit representations, while highly realistic relighting is feasible, scene editing presents a challenging endeavor. + +As an illustration, we concentrate on relighting in a multi-object composition scene. Initially, our pipeline computes the visibility $V(\omega_{i})$ for each Gaussian point via ray tracing (Sec. 4.1). Despite the challenging inter-object occlusions introduced by composition, our proposed point-based ray tracing method adeptly manages these complexities and ensures accurate occlusion updates within the novel scene. Subsequently, the rendering process unfolds (Sec. 3.3), starting with Gaussian-level PBR and ending with alpha blending. Throughout this process, the original indirect lighting of each object is discarded. Consequently, we achieve relighting with remarkably vivid shadow effects solely based on a discrete set of points, as illustrated in Fig. 1. + +# 5 Experiments + +# 5.1 Training Details + +As pointed out in Sec. 4.2, the training procedure is divided into two stages. The first stage follows the setting of 3DGS [26], along with the proposed normal gradient-based densification (Sec. 3.2) where $T_{\mathbf{n}}$ is set as $2 \times 10^{-9}$ . In the second stage, we sample $N_{s} = 64$ incident rays per Gaussian point for PBR. We train + +Table 1: Quantitative results for novel view synthesis on NeRF synthetic dataset + +
GeometryRelightablePSNR↑ SSIM↑ LPIPS
NPBG [1]point×28.100.9230.077
NPBG++ [38]point×28.120.9280.076
FreqPCR [54]point×31.240.9500.049
3DGS [26]point×33.880.9700.031
PhySG [51]neural✓18.910.8470.182
NeILF++ [49]neural✓26.370.9110.091
Nvdiffrec [33]mesh✓29.050.9390.081
R3DG(Ours)point✓31.220.9590.039
+ +![](images/691c40cc11a766ec29a9963794e6671aa4e0a4ef875e26e1e366231387f26dff.jpg) +(a) Lego + +![](images/d512835caae51e69be0bd4764e61a50ef532b082695459b4484b1db87d5a5b21.jpg) +(b) Chair + +![](images/5172a9fff6cd0cdfd0864a9e17ed364e479df8dd62e942fb360bd3a5a15597cb.jpg) +(c) Hotdog +Fig. 4: Visualizations on NeRF synthetic dataset [31]. Each scene is displayed in an order from left to right and from top to bottom: Ground Truth, PBR Image, Normal Map and Ambient Occlusion Map. + +our model for 30,000 iterations in the initial stage and 10,000 iterations in the second stage. The loss in the first stage is calculated by: + +$$ +\mathcal {L} = \lambda_ {1} \mathcal {L} _ {1} + \lambda_ {s s i m} \mathcal {L} _ {s s i m} + \lambda_ {n} \mathcal {L} _ {n} + \lambda_ {s, n} \mathcal {L} _ {s, n} + \lambda_ {O} \mathcal {L} _ {O} + \lambda_ {u} \mathcal {L} _ {u}, \tag {16} +$$ + +where $\{\lambda_1, \lambda_{ssim}, \lambda_n, \lambda_{s,n}, \lambda_O, \lambda_u\}$ are set to $\{0.8, 0.2, 0.01, 0.01, 0.01, 0.01\}$ , respectively. The loss in the second stage is given by: + +$$ +\mathcal {L} = \lambda_ {1} \mathcal {L} _ {1} + \lambda_ {s s i m} \mathcal {L} _ {s s i m} + \lambda_ {l} \mathcal {L} _ {l} + \lambda_ {s, b} \mathcal {L} _ {s, b} + \lambda_ {s, r} \mathcal {L} _ {s, r}, \tag {17} +$$ + +where $\{\lambda_1,\lambda_{ssim},\lambda_l,\lambda_{s,b},\lambda_{s,r}\}$ are set to $\{0.8,0.2,0.0001,0.01,0.01\}$ , respectively. + +# 5.2 Performance on Novel View Synthesis + +We first evaluate the novel view synthesis (NVS) of the physically based rendering (PBR) on NeRF synthetic dataset [31]. We compare both point-based rendering methods and relightable rendering approaches. We report average metrics across all scenes in Tab. 1, including peak signal-to-noise ratio (PSNR), structural similarity index measure (SSIM), and learned perceptual image patch similarity (LPIPS). Compared with existing point-based rendering methods, our + +Table 2: Quantitative results on Synthetic4Relight dataset [56] + +
Novel View SynthesisRelightingAlbedoRoughnessTime
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓MSE↓(hours)
NerFactor [53]22.800.9160.15021.540.8750.17119.490.8640.206N/A>48
Nvdifferc-MC [19]34.290.9670.06824.220.9430.07829.610.9450.0750.0094.17
InvRender [56]30.740.9530.08628.670.9500.09128.280.9350.0720.00814.3
TensorIR [23]35.800.9780.04929.690.9510.07930.580.9460.0650.0153.24
R3DG (Ours)36.800.9820.02831.000.9640.05028.310.9510.0580.0130.90
+ +![](images/eda3e19a9df0d153ec03f03ce255213cb4e3de1751b555409800b287014933c5.jpg) +Fig. 5: Qualitative results on Synthetic4Relight dataset [56]. + +R3DG demonstrates superiority over most of them [1,38,54]. While our performance slightly marginally trails that of vanilla 3DGS [26], it is noteworthy that our approach excels in relighting capability, presenting a significant advantage. Furthermore, in comparison to other relightable methods [33,49,51], our method showcases significantly better NVS quality. + +In Fig. 4, we visualize the our reconstruction results in NeRF synthetic dataset, including the PBR image, the normal map, and the pre-computed visibility illustrated in the form of the Ambient Occlusion (AO) map. As depicted in Fig. 4, our approach successfully achieves realistic PBR rendering, smooth normal estimation, and accurate visibility solving on discrete point clouds. + +# 5.3 Performance on Relighting + +To comprehensively assess the relighting capabilities of our pipeline, we further undertook experiments using the Synthetic4Relight dataset [56]. Recognizing + +![](images/fec47af15235fbfd3b6c09aae284aa837196eee7701c35dff80bb8a743ccc7dd.jpg) +Fig. 6: Qualitative results of relighting on real-world scenes. + +the inherent scale ambiguity between the estimated albedo and lighting, we standardized the scale of the estimated albedo against the ground truth to relighting, consistent with previous studies [51, 56]. + +Our evaluation encompasses the analysis of decomposed materials, the synthesis of novel views, and the relighting outcomes. The quantitative analysis is presented in Tab. 2. Our method outperforms existing approaches in terms of Novel View Synthesis (NVS) and relighting precision. Regarding material estimation, our method exhibits superior albedo accuracy in SSIM and LPIPS. Qualitatively, our method achieves visually pleasing material decomposition, facilitating a realistic relighting effect, which is shown in 5. + +Additionally, we conduct relighting on real-world scenes in the Mip-NeRF 360 dataset [4], as shown in Fig. 6. It is worth noting that our relighting in the kitchen with the second environment map, exhibits pleasing shadow effects. + +# 5.4 Ablation Study + +We conduct ablation studies on three principal components of our method that significantly influence the quality of inverse rendering. The results are illustrated in Fig. 7. Initially, we investigate the impact of normal gradient densification (NGD). Fig. 7 (a, b) depict the rendering normal maps with and without the densification. It shows that the proposed densification markedly enhances the details of the normal map for *Lego*. In Fig. 7 (c, d, e), we examine the effects of our full light modeling. We simplify the full model to a global environment map in the optimization. The result demonstrates that the simplified model fails to generate shadow effects when relighting. Furthermore, we evaluate the significance of the proposed constraint on depth distribution, denoted as $\mathcal{L}_u$ . The depth uncertainty is visualized in Fig. 7 (h, i), which illustrates the distribution + +![](images/5b3b21ad98f3987eb60bbdce21c626f4daf0bf4745a4fabeed8200109a7888a4.jpg) +(a) w NGD. + +![](images/6523353d9050e62fb8dc52cc480729a976107cbc8ccc0568fd9988778d08578a.jpg) +(b) w/o NGD. + +![](images/5758f5d0b48f1ef7bf0d2582aadb03b628d0e685558ec0a66b0068add60c40f0.jpg) +(c) Relit. GT. + +![](images/96deec950c0a1ce91651e02ca8ca1e20ae6b11f6c7e4cc0373e465b782a38585.jpg) +(d) Full light. + +![](images/ff8d7021dea44ee962f1e8c41b78f3df0d10a159cf53e5904436ca8b072ecbd7.jpg) +(e) Simplified. + +![](images/690f3d3716c469c06a0ee3cde99eb678c9216a4f1d55251ed51d3d4f0e96bec3.jpg) +(f) AO w $\mathcal{L}_u$ + +![](images/71cbe71748db2e43c57dbc9a9120d19c0f11c1cbff4d6c0a56c63cf57c92e180.jpg) +(g) AO w/o $\mathcal{L}_u$ + +![](images/2b3b550b97d69ba29525029c7a3c2296b1502eab4d62adcb05b70376006e012f.jpg) +(h) Uncertainty w $\mathcal{L}_u$ +Fig. 7: Ablation studies on main components of our method. (a) and (b) show normal maps with and without the proposed normal gradient densification (NGD). (d) and (e) display re-lighting results using the proposed full lighting model and only a simplified light modeling, i.e. a global environment map. (f) and (g) visualize ambient occlusion (AO) maps with and without the proposed constraint on depth distribution $\mathcal{L}_u$ . (h) and (i) illustrate the depth uncertainty with and without $\mathcal{L}_u$ . + +![](images/fa38bf3880ce90a9172ea6a6553efbb351ceff1352f95bee43ca8e6e386d8890.jpg) +(i) Uncertainty w/o $\mathcal{L}_u$ + +constraint contribute significantly to reducing the depth uncertainty. Additionally, Fig. 7 (f, g) reveals that incorporating $\mathcal{L}_u$ leads to more plausible ambient occlusion (averaged visibility) which plays a crucial role in stage 2. + +# 6 Conclusion + +In this paper, we have introduced a differentiable point-based rendering pipeline enabling effort less editing, accuracy ray tracing, and photo-realistic relighting. We present the scene as Relightable 3D Gaussians, an extension of traditional 3D Gaussians enriched with supplementary normals, BRDFs and indirect lighting. To reconstruct a relightable scene from multi-view images, we build a novel differentiable rendering pipeline based on the inverse rendering techniques and 3D Gaussian Splatting. Additionally, we introduce an innovative ray tracing scheme specifically designed for scenes represented by discrete points, providing precise visibility for realistic relighting. Quantitative and qualitative experiments confirm that our pipeline successfully reconstructs reasonable normals, materials for discrete point clouds, and achieves commendable accuracy in both novel view synthesis and scene relighting tasks. + +Limitations and Future work. Our present pipeline targets the reconstruction of static objects. Certain design choices pose challenges in maintaining its performance with large scale scenes. The high density of points in large scene can slow down optimization as we sample rays and perform PBR at each point. This can be mitigated through deferred rendering technique. Regarding geometry, integrating MVS shows promise for achieving more accurate representations. + +# Acknowledgements + +This work was supported by National Key R&D Program of China (2023YFB3209702), and NSFC (62441204). + +# References + +1. Aliev, K.A., Sevastopolsky, A., Kolos, M., Ulyanov, D., Lempitsky, V.: Neural point-based graphics. In: ECCV (2020) +2. Azinovic, D., Li, T.M., Kaplanyan, A., Nießner, M.: Inverse path tracing for joint material and lighting estimation. In: CVPR (2019) +3. Barron, J.T., Mildenhall, B., Tancik, M., Hedman, P., Martin-Brualla, R., Srinivasan, P.P.: Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In: ICCV (2021) +4. Barron, J.T., Mildenhall, B., Verbin, D., Srinivasan, P.P., Hedman, P.: Mip-nerf 360: Unbounded anti-aliased neural radiance fields. In: CVPR (2022) +5. Barron, J.T., Mildenhall, B., Verbin, D., Srinivasan, P.P., Hedman, P.: Zip-nerf: Anti-aliased grid-based neural radiance fields. arXiv preprint (2023) +6. Bi, S., Xu, Z., Srinivasan, P., Mildenhall, B., Sunkavalli, K., Hašan, M., Hold-Geoffroy, Y., Kriegman, D., Ramamoorthi, R.: Neural reflectance fields for appearance acquisition. arXiv preprint (2020) +7. Bi, S., Xu, Z., Sunkavalli, K., Hašan, M., Hold-Geoffroy, Y., Kriegman, D., Ramamoorthi, R.: Deep reflectance volumes: Relightable reconstructions from multiview photometric images. In: ECCV (2020) +8. Bi, S., Xu, Z., Sunkavalli, K., Kriegman, D., Ramamoorthi, R.: Deep 3d capture: Geometry and reflectance from sparse multi-view images. In: CVPR (2020) +9. Boss, M., Braun, R., Jampani, V., Barron, J.T., Liu, C., Lensch, H.: Nerd: Neural reflectance decomposition from image collections. In: ICCV (2021) +0. Boss, M., Engelhardt, A., Kar, A., Li, Y., Sun, D., Barron, J., Lensch, H., Jampani, V.: Samurai: Shape and material from unconstrained real-world arbitrary image collections. In: NeurIPS (2022) +1. Burley, B., Studios, W.D.A.: Physically-based shading at disney. In: SIGGRAPH (2012) +2. Chen, A., Xu, Z., Geiger, A., Yu, J., Su, H.: Tensorf: Tensorial radiance fields. In: ECCV (2022) +3. Chen, Z., Funkhouser, T., Hedman, P., Tagliasacchi, A.: Mobilenerf: Exploiting the polygon rasterization pipeline for efficient neural field rendering on mobile architectures. In: CVPR (2023) +4. Deng, K., Liu, A., Zhu, J.Y., Ramanan, D.: Depth-supervised nef: Fewer views and faster training for free. In: CVPR (2022) +5. Fridovich-Keil, S., Yu, A., Tancik, M., Chen, Q., Recht, B., Kanazawa, A.: Plenoxels: Radiance fields without neural networks. In: CVPR (2022) +6. Garbin, S.J., Kowalski, M., Johnson, M., Shotton, J., Valentin, J.: Fastnerf: High-fidelity neural rendering at 200fps. In: ICCV (2021) +7. Guédon, A., Lepetit, V.: Sugar: Surface-aligned gaussian splatting for efficient 3d mesh reconstruction and high-quality mesh rendering. arXiv preprint arXiv:2311.12775 (2023) +8. Guo, K., Lincoln, P., Davidson, P., Busch, J., Yu, X., Whalen, M., Harvey, G., Orts-Escalano, S., Pandey, R., Dourgarian, J., et al.: The relighttables: Volumetric performance capture of humans with realistic relighting. ACM TOG (2019) + +19. Hasselgren, J., Hofmann, N., Munkberg, J.: Shape, light, and material decomposition from images using monte carlo rendering and denoising. In: NeurIPS (2022) +20. Hasselgren, J., Hofmann, N., Munkberg, J.: Shape, Light, and Material Decomposition from Images using Monte Carlo Rendering and Denoising. arXiv:2206.03380 (2022) +21. Hedman, P., Srinivasan, P.P., Mildenhall, B., Barron, J.T., Debevec, P.: Baking neural radiance fields for real-time view synthesis. In: ICCV (2021) +22. Hu, T., Liu, S., Chen, Y., Shen, T., Jia, J.: Efficientnerf efficient neural radiance fields. In: CVPR (2022) +23. Jin, H., Liu, I., Xu, P., Zhang, X., Han, S., Bi, S., Zhou, X., Xu, Z., Su, H.: Tensoir: Tensorial inverse rendering. In: CVPR (2023) +24. Kajiya, J.T.: The rendering equation. In: Proceedings of the 13th annual conference on Computer graphics and interactive techniques (1986) +25. Karras, T.: Maximizing parallelism in the construction of bvhs, octrees, and k-d trees. In: Proceedings of the Fourth ACM SIGGRAPH/Eurographics conference on High-Performance Graphics (2012) +26. Kerbl, B., Kopanas, G., Leimkuhler, T., Drettakis, G.: 3d gaussian splatting for real-time radiance field rendering. ACM TOG (2023) +27. Keselman, L., Hebert, M.: Flexible techniques for differentiable rendering with 3d gaussians. arXiv preprint (2023) +28. Kuang, Z., Olszewski, K., Chai, M., Huang, Z., Achlioptas, P., Tulyakov, S.: Neroic: Neural rendering of objects from online image collections. ACM TOG (2022) +29. Levoy, M., Whitted, T.: The use of points as a display primitive (1985) +30. Liu, Y., Wang, P., Lin, C., Long, X., Wang, J., Liu, L., Komura, T., Wang, W.: Nano: Neural geometry and brdf reconstruction of reflective objects from multiview images. In: SIGGRAPH (2023) +31. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. In: ECCV (2020) +32. Müller, T., Evans, A., Schied, C., Keller, A.: Instant neural graphics primitives with a multiresolution hash encoding. ACM TOG (2022) +33. Munkberg, J., Hasselgren, J., Shen, T., Gao, J., Chen, W., Evans, A., Müller, T., Fidler, S.: Extracting triangular 3d models, materials, and lighting from images. In: CVPR (2022) +34. Nam, G., Lee, J.H., Gutierrez, D., Kim, M.H.: Practical svbrdf acquisition of 3d objects with unstructured flash photography. ACM TOG (2018) +35. Park, J.J., Holynski, A., Seitz, S.M.: Seeing the world in a bag of chips. In: CVPR (2020) +36. Pfister, H., Zwicker, M., Van Baar, J., Gross, M.: Surfels: Surface elements as rendering primitives. In: Proceedings of the 27th annual conference on Computer graphics and interactive techniques (2000) +37. Pittaluga, F., Koppal, S.J., Kang, S.B., Sinha, S.N.: Revealing scenes by inverting structure from motion reconstructions. In: CVPR (2019) +38. Rakhimov, R., Ardelean, A.T., Lempitsky, V., Burnaev, E.: Npbg++: Accelerating neural point-based graphics. In: CVPR (2022) +39. Reiser, C., Peng, S., Liao, Y., Geiger, A.: Kilonerf: Speeding up neural radiance fields with thousands of tiny mlps. In: ICCV (2021) +40. Schmitt, C., Donne, S., Riegler, G., Koltun, V., Geiger, A.: On joint estimation of pose, geometry and svbrdf from a handheld scanner. In: CVPR (2020) + +41. Srinivasan, P.P., Deng, B., Zhang, X., Tancik, M., Mildenhall, B., Barron, J.T.: Nerv: Neural reflectance and visibility fields for relighting and view synthesis. In: CVPR (2021) +42. Verbin, D., Hedman, P., Mildenhall, B., Zickler, T., Barron, J.T., Srinivasan, P.P.: Ref-nerf: Structured view-dependent appearance for neural radiance fields. In: CVPR (2022) +43. Wang, P., Liu, L., Liu, Y., Theobalt, C., Komura, T., Wang, W.: Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. arXiv preprint (2021) +44. Wu, H., Hu, Z., Li, L., Zhang, Y., Fan, C., Yu, X.: Nefii: Inverse rendering for reflectance decomposition with near-field indirect illumination. In: CVPR (2023) +45. Yao, Y., Zhang, J., Liu, J., Qu, Y., Fang, T., McKinnon, D., Tsin, Y., Quan, L.: Neilf: Neural incident light field for physically-based material estimation. In: ECCV (2022) +46. Yariv, L., Gu, J., Kasten, Y., Lipman, Y.: Volume rendering of neural implicit surfaces. In: NeurIPS (2021) +47. Yifan, W., Serena, F., Wu, S., Öztireli, C., Sorkine-Hornung, O.: Differentiable surface splatting for point-based geometry processing. ACM TOG (2019) +48. Yu, Z., Peng, S., Niemeyer, M., Sattler, T., Geiger, A.: Monosdf: Exploring monocular geometric cues for neural implicit surface reconstruction. In: NeurIPS (2022) +49. Zhang, J., Yao, Y., Li, S., Liu, J., Fang, T., McKinnon, D., Tsin, Y., Quan, L.: Neil++: Inter-reflectable light fields for geometry and material estimation. In: ICCV (2023) +50. Zhang, K., Luan, F., Li, Z., Snavely, N.: Iron: Inverse rendering by optimizing neural sdfs and materials from photometric images. In: CVPR (2022) +51. Zhang, K., Luan, F., Wang, Q., Bala, K., Snavely, N.: Physg: Inverse rendering with spherical gaussians for physics-based material editing and relighting. In: CVPR (2021) +52. Zhang, Q., Baek, S.H., Rusinkiewicz, S., Heide, F.: Differentiable point-based radiance fields for efficient view synthesis. In: SIGGRAPH Asia (2022) +53. Zhang, X., Srinivasan, P.P., Deng, B., Debevec, P., Freeman, W.T., Barron, J.T.: Nerfactor: Neural factorization of shape and reflectance under an unknown illumination. ACM TOG (2021) +54. Zhang, Y., Huang, X., Ni, B., Li, T., Zhang, W.: Frequency-modulated point cloud rendering with easy editing. In: CVPR (2023) +55. Zhang, Y., Xu, T., Yu, J., Ye, Y., Jing, Y., Wang, J., Yu, J., Yang, W.: Nemf: Inverse volume rendering with neural microflake field. In: ICCV (2023) +56. Zhang, Y., Sun, J., He, X., Fu, H., Jia, R., Zhou, X.: Modeling indirect illumination for inverse rendering. In: CVPR (2022) +57. Zwicker, M., Pfister, H., Van Baar, J., Gross, M.: Surface splatting. In: Proceedings of the 28th annual conference on computer graphics and interactive techniques (2001) +58. Zwicker, M., Pfister, H., Van Baar, J., Gross, M.: Ewa splatting. IEEE TVCG (2002) \ No newline at end of file diff --git a/2024/Relightable 3D Gaussians_ Realistic Point Cloud Relighting with BRDF Decomposition and Ray Tracing/images.zip b/2024/Relightable 3D Gaussians_ Realistic Point Cloud Relighting with BRDF Decomposition and Ray Tracing/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..e48532844164f1dfb7e57b7e049c42d2ee54f698 --- /dev/null +++ b/2024/Relightable 3D Gaussians_ Realistic Point Cloud Relighting with BRDF Decomposition and Ray Tracing/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6ee2f4c56c06482f34cf6edad2a15673391381333310f4c7fd61cf1cb5adff92 +size 523154 diff --git a/2024/Relightable 3D Gaussians_ Realistic Point Cloud Relighting with BRDF Decomposition and Ray Tracing/layout.json b/2024/Relightable 3D Gaussians_ Realistic Point Cloud Relighting with BRDF Decomposition and Ray Tracing/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..5af89325fd1ab2736b76c00bca879d2e1444485e --- /dev/null +++ b/2024/Relightable 3D Gaussians_ Realistic Point Cloud Relighting with BRDF Decomposition and Ray Tracing/layout.json @@ -0,0 +1,11222 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 138, + 111, + 477, + 166 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 111, + 477, + 166 + ], + "spans": [ + { + "bbox": [ + 138, + 111, + 477, + 166 + ], + "type": "text", + "content": "Relightable 3D Gaussians: Realistic Point Cloud Relighting with BRDF Decomposition and Ray Tracing" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 185, + 468, + 210 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 185, + 468, + 210 + ], + "spans": [ + { + "bbox": [ + 145, + 185, + 468, + 210 + ], + "type": "text", + "content": "Jian " + }, + { + "bbox": [ + 145, + 185, + 468, + 210 + ], + "type": "inline_equation", + "content": "\\mathrm{Gao}^{1*}" + }, + { + "bbox": [ + 145, + 185, + 468, + 210 + ], + "type": "text", + "content": ", Chun " + }, + { + "bbox": [ + 145, + 185, + 468, + 210 + ], + "type": "inline_equation", + "content": "\\mathrm{Gu}^{2*}" + }, + { + "bbox": [ + 145, + 185, + 468, + 210 + ], + "type": "text", + "content": ", Youtian Lin" + }, + { + "bbox": [ + 145, + 185, + 468, + 210 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 145, + 185, + 468, + 210 + ], + "type": "text", + "content": ", Zhihao Li" + }, + { + "bbox": [ + 145, + 185, + 468, + 210 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 145, + 185, + 468, + 210 + ], + "type": "text", + "content": ", Hao Zhu" + }, + { + "bbox": [ + 145, + 185, + 468, + 210 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 145, + 185, + 468, + 210 + ], + "type": "text", + "content": ", Xun " + }, + { + "bbox": [ + 145, + 185, + 468, + 210 + ], + "type": "inline_equation", + "content": "\\mathrm{Cao}^{1}" + }, + { + "bbox": [ + 145, + 185, + 468, + 210 + ], + "type": "text", + "content": ", Li Zhang" + }, + { + "bbox": [ + 145, + 185, + 468, + 210 + ], + "type": "inline_equation", + "content": "^{2\\boxtimes}" + }, + { + "bbox": [ + 145, + 185, + 468, + 210 + ], + "type": "text", + "content": ", and Yao " + }, + { + "bbox": [ + 145, + 185, + 468, + 210 + ], + "type": "inline_equation", + "content": "\\mathrm{Yao}^{1\\boxtimes}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 264, + 219, + 350, + 230 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 264, + 219, + 350, + 230 + ], + "spans": [ + { + "bbox": [ + 264, + 219, + 350, + 230 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 264, + 219, + 350, + 230 + ], + "type": "text", + "content": " Nanjing University" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 265, + 231, + 347, + 241 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 265, + 231, + 347, + 241 + ], + "spans": [ + { + "bbox": [ + 265, + 231, + 347, + 241 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 265, + 231, + 347, + 241 + ], + "type": "text", + "content": " Fudan University" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 253, + 242, + 361, + 252 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 253, + 242, + 361, + 252 + ], + "spans": [ + { + "bbox": [ + 253, + 242, + 361, + 252 + ], + "type": "text", + "content": "3 Huawei Noah's Ark Lab" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 159, + 283, + 455, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 283, + 455, + 449 + ], + "spans": [ + { + "bbox": [ + 159, + 283, + 455, + 449 + ], + "type": "text", + "content": "Abstract. In this paper, we present a novel differentiable point-based rendering framework to achieve photo-realistic relighting. To make the reconstructed scene relightable, we enhance vanilla 3D Gaussians by associating extra properties, including normal vectors, BRDF parameters, and incident lighting from various directions. From a collection of multiview images, the 3D scene is optimized through 3D Gaussian Splatting while BRDF and lighting are decomposed by physically based differentiable rendering. To produce plausible shadow effects in photo-realistic relighting, we introduce an innovative point-based ray tracing with the bounding volume hierarchies for efficient visibility pre-computation. Extensive experiments demonstrate our improved BRDF estimation, novel view synthesis and relighting results compared to state-of-the-art approaches. The proposed framework showcases the potential to revolutionize the mesh-based graphics pipeline with a point-based pipeline enabling editing, tracing, and relighting." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 160, + 458, + 449, + 470 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 458, + 449, + 470 + ], + "spans": [ + { + "bbox": [ + 160, + 458, + 449, + 470 + ], + "type": "text", + "content": "Keywords: 3D Gaussian Splatting, Relighting, Point based rendering" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 490, + 230, + 502 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 490, + 230, + 502 + ], + "spans": [ + { + "bbox": [ + 132, + 490, + 230, + 502 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 515, + 482, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 515, + 482, + 647 + ], + "spans": [ + { + "bbox": [ + 130, + 515, + 482, + 647 + ], + "type": "text", + "content": "Reconstructing 3D scenes from multi-view images for photo-realistic rendering is a fundamental problem at the intersection of computer vision and graphics. Recently, 3D Gaussian Splatting (3DGS) [26] has been proposed and has gained significant attention from the community. The method employs a set of 3D Gaussian points to represent a 3D scene and projects these points onto a designated view through a tile-based rasterization. Attributes of each 3D Gaussian point are then optimized through the point-based differentiable rendering. Notably, 3DGS achieves real-time rendering with quality comparable to or even surpassing previous state-of-the-art methods (e.g., Mip-NeRF [3]), with a training speed on par with the most efficient Instant-NGP [32]. However, the current 3DGS is unable to reconstruct a scene that can be relighted under different lighting conditions," + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 142, + 654, + 233, + 666 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 142, + 654, + 233, + 666 + ], + "spans": [ + { + "bbox": [ + 142, + 654, + 233, + 666 + ], + "type": "text", + "content": "*Equally contributed." + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 135, + 114, + 306, + 182 + ], + "blocks": [ + { + "bbox": [ + 135, + 114, + 306, + 182 + ], + "lines": [ + { + "bbox": [ + 135, + 114, + 306, + 182 + ], + "spans": [ + { + "bbox": [ + 135, + 114, + 306, + 182 + ], + "type": "image", + "image_path": "db90ca5bf82a52b5d6d04ed8397b42380b6f5d0b08db17516332ec2f153721dc.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 192, + 183, + 250, + 191 + ], + "lines": [ + { + "bbox": [ + 192, + 183, + 250, + 191 + ], + "spans": [ + { + "bbox": [ + 192, + 183, + 250, + 191 + ], + "type": "text", + "content": "(a) Point Cloud" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 307, + 114, + 479, + 182 + ], + "blocks": [ + { + "bbox": [ + 307, + 114, + 479, + 182 + ], + "lines": [ + { + "bbox": [ + 307, + 114, + 479, + 182 + ], + "spans": [ + { + "bbox": [ + 307, + 114, + 479, + 182 + ], + "type": "image", + "image_path": "77fad46f9adc611e9ba0cf90993ce1f1e0aa703b8fead9cdd503726d1c334d57.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 372, + 183, + 414, + 191 + ], + "lines": [ + { + "bbox": [ + 372, + 183, + 414, + 191 + ], + "spans": [ + { + "bbox": [ + 372, + 183, + 414, + 191 + ], + "type": "text", + "content": "(b) Normal" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 135, + 193, + 306, + 258 + ], + "blocks": [ + { + "bbox": [ + 135, + 193, + 306, + 258 + ], + "lines": [ + { + "bbox": [ + 135, + 193, + 306, + 258 + ], + "spans": [ + { + "bbox": [ + 135, + 193, + 306, + 258 + ], + "type": "image", + "image_path": "4764d770bb5551e8d11a38b78daafcf236583b1ee34141e43ba563b3f00de2e1.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 180, + 261, + 261, + 270 + ], + "lines": [ + { + "bbox": [ + 180, + 261, + 261, + 270 + ], + "spans": [ + { + "bbox": [ + 180, + 261, + 261, + 270 + ], + "type": "text", + "content": "(c) Ambient Occlusion" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 307, + 193, + 479, + 259 + ], + "blocks": [ + { + "bbox": [ + 307, + 193, + 479, + 259 + ], + "lines": [ + { + "bbox": [ + 307, + 193, + 479, + 259 + ], + "spans": [ + { + "bbox": [ + 307, + 193, + 479, + 259 + ], + "type": "image", + "image_path": "0b357a25004a1d765e431fb52522809bbd1a4e14a34ff10c8b10278a46435edf.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 336, + 261, + 450, + 270 + ], + "lines": [ + { + "bbox": [ + 336, + 261, + 450, + 270 + ], + "spans": [ + { + "bbox": [ + 336, + 261, + 450, + 270 + ], + "type": "text", + "content": "(d) Physically Based Rendering" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 279, + 482, + 357 + ], + "lines": [ + { + "bbox": [ + 130, + 279, + 482, + 357 + ], + "spans": [ + { + "bbox": [ + 130, + 279, + 482, + 357 + ], + "type": "text", + "content": "Fig. 1: Visual results of our pipeline on a multi-object composition scene. In our pipeline, we represent a scene as Relightable 3D Gaussians. From multi-view images, we recover the geometry and materials of individual objects with inverse rendering techniques (see Sec. 3). Then, objects are easily composited into a new scene, thanks to our explicit representation. After that, we solve the complex occlusions though point based ray tracing (see Sec. 4) and re-light the new scene. Ultimately, we achieve high-fidelity relighting with remarkably realistic shadow." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 395, + 480, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 395, + 480, + 443 + ], + "spans": [ + { + "bbox": [ + 130, + 395, + 480, + 443 + ], + "type": "text", + "content": "making the method only applicable to the task of novel view synthesis. In addition, ray tracing, a crucial component in achieving realistic rendering, remains an unresolved challenge in point-based representation, limiting 3DGS to more dedicated rendering effects such as shadowing and light reflectance." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 450, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 450, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 450, + 482, + 666 + ], + "type": "text", + "content": "In this paper, we aim to reconstruct a re烈table 3D Gaussian point cloud from multi-view images with a differentiable rendering framework and achieve realistic relighting. We make 3D Gaussian relightable by assigning it with additional normal, BRDF properties, and incident light information to model perpoint light reflectance. In contrast to the plain alpha blending in the original 3DGS [26], we apply physically based rendering (PBR) to get a PBR color for each 3D Gaussian point, which is then alpha-composed to obtain a rendered color for the corresponding image pixel. For robust material and lighting estimation, we split the incident light into a global environment map and an indirect incident light field. To capture accurate visibility for each 3D Gaussian, we propose a novel ray tracing method based on the bounding volume hierarchy (BVH), which enables efficient visibility pre-computing for real-time rendering. Additionally, the proposed ray tracing method can handle complex occlusion relationships in a novel multi-object composition scene, thus realizing realistic shadow effects. Moreover, proper regularizations are introduced to enhance the geometry and mitigate the material-lighting ambiguity during the optimization, including constraints on depth distribution, smoothness priors, and a lighting regularization." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "text", + "content": "J. Gao et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 175 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 175 + ], + "type": "text", + "content": "Extensive experiments conducted across diverse datasets demonstrate the improved material and lighting estimation results and novel view rendering quality. Additionally, we illustrate the relightable and editable capabilities of our system through multi-object composition in a novel lighting environment. To summarize, major contributions of the paper include:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 182, + 479, + 288 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 138, + 182, + 479, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 182, + 479, + 217 + ], + "spans": [ + { + "bbox": [ + 138, + 182, + 479, + 217 + ], + "type": "text", + "content": "- We propose a material and lighting decomposition scheme for 3D Gaussian Splatting, where a normal vector, BRDF values, and incident lights are assigned and optimized for each 3D Gaussian point." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 218, + 479, + 252 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 218, + 479, + 252 + ], + "spans": [ + { + "bbox": [ + 138, + 218, + 479, + 252 + ], + "type": "text", + "content": "- We introduce a novel point-based ray tracing approach based on bounding volume hierarchy, enabling efficient visibility pre-computing for each 3D Gaussian point and rendering of a 3D scene with realistic shadow effect." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 254, + 479, + 288 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 254, + 479, + 288 + ], + "spans": [ + { + "bbox": [ + 138, + 254, + 479, + 288 + ], + "type": "text", + "content": "- We demonstrate a comprehensive graphics pipeline solely based on a discretized point representation, supporting relighting, editing, and ray tracing of a reconstructed 3D point cloud." + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 132, + 306, + 242, + 319 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 306, + 242, + 319 + ], + "spans": [ + { + "bbox": [ + 132, + 306, + 242, + 319 + ], + "type": "text", + "content": "2 Related Works" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 330, + 480, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 330, + 480, + 437 + ], + "spans": [ + { + "bbox": [ + 130, + 330, + 480, + 437 + ], + "type": "text", + "content": "Neural Radiance Field. Differentiable rendering techniques, exemplified by Neural Radiance Field (NeRF) [31], have garnered significant attention [3,12,32]. NeRF utilizes an implicit Multi-Layer Perceptron (MLP) that takes 3D positions and viewing directions as inputs to generate density and view-dependent colors for differentiable volume rendering. Despite its powerful neural implicit representation, both speed and quality can be improved. Efficiency improvements mainly focus on optimizing queries on neural fields and minimizing queries [12, 13, 15, 16, 21, 22, 32, 39]. Quality enhancements involve anti-aliasing [3-5] or utilizing exterior supervision [14, 48], etc." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 438, + 480, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 438, + 480, + 593 + ], + "spans": [ + { + "bbox": [ + 130, + 438, + 480, + 593 + ], + "type": "text", + "content": "Differentiable Point Based Rendering. Using points as rendering primitives was first proposed in [29]. To address the issue of resulting holey images when rasterizing discrete points directly, solutions fall into two categories. One approach directly employs points as rendering primitives and encodes the geometric and photometric features near the point using SIFT descriptors [37] or neural descriptors [1,38]. This leads to a feature image with gaps, which is then decoded to recover a hole-free RGB image. The other category models a point as a primitive occupying a specific space, like a surfel [36,47] or a 3D Gaussian [26]. The rendered image is then generated using a splatting technique. This technique saw significant development in the 2000s [36,57,58], and in the era of differentiable rendering, PointRF [52], DSS [47] and 3DGS [26] serve as notable examples. We assert that the point serves as a promising rendering primitive owing to its inherent ease of editing, as also corroborated in [54]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 594, + 479, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 479, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 479, + 665 + ], + "type": "text", + "content": "Material and Lighting Estimation. Decomposing the materials and illumination of a scene from multi-view images presents a formidable challenge owing to its intrinsic high-dimensional complexity. Some methods simplified the problem under the assumption of a controllable light [2,6-8,18,34,35,40]. Subsequent research explores more complex lighting models to cope with realistic scenarios. NeRV [41] and PhySG [51] leverage an environmental map to manage arbitrary" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 342, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 342, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 342, + 91, + 447, + 102 + ], + "type": "text", + "content": "Relightable 3D Gaussians" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 115, + 482, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 115, + 482, + 357 + ], + "spans": [ + { + "bbox": [ + 130, + 115, + 482, + 357 + ], + "type": "text", + "content": "lighting conditions. NeRD [9] addresses the challenge posed by images captured under varying illumination by attributing Spherical Gaussians to each image. IRON [50] introduces an innovative edge sampling algorithm tailored for neural SDFs. NeRFactor [53] exploits light visibility to achieve superior material and lighting decomposition. Further studies [19, 52] address the consideration of indirect lighting, leading to enhanced BRDF estimation quality. [20, 33] utilize differentiable marching tetrahedrons for direct optimization on mesh surfaces. [10, 28] deal with varying cameras, illumination, and backgrounds. Ref-NeRF [42] suggests the employment of integrated direction encoding to ameliorate the reconstruction fidelity of reflective objects. [30, 33] use split-sum approximation to approximate the shading effects. NeMF [55] represents the scene as a microflake volume. NeFII [44] integrates lights through path tracing with Monte Carlo sampling. InvRender [56] computes indirect illumination by directly leveraging the neural radiance field, rather than concurrently estimating it alongside the decomposition of direct lighting and materials. TensoIR [23] performs inverse rendering based on tensor factorization and neural fields. NeILF [45] and NeILF++ [49] expresses the incident lights as a neural incident light field, while NeILF++ integrates VolSDF [46] with NeILF through inter-reflection. However, existing schemes rarely delve into BRDF estimation for point clouds, and the estimation and rendering quality remain to be improved." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 369, + 307, + 384 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 369, + 307, + 384 + ], + "spans": [ + { + "bbox": [ + 132, + 369, + 307, + 384 + ], + "type": "text", + "content": "3 Relightable 3D Gaussians" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 391, + 481, + 429 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 391, + 481, + 429 + ], + "spans": [ + { + "bbox": [ + 130, + 391, + 481, + 429 + ], + "type": "text", + "content": "In this section, we introduce a novel pipeline to decompose materials and lighting from multi-view images based on 3D Gaussian Splatting (3DGS) [26]. An overview of our pipeline is shown in Fig. 2." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 441, + 221, + 454 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 441, + 221, + 454 + ], + "spans": [ + { + "bbox": [ + 132, + 441, + 221, + 454 + ], + "type": "text", + "content": "3.1 Preliminary" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 456, + 481, + 493 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 456, + 481, + 493 + ], + "spans": [ + { + "bbox": [ + 130, + 456, + 481, + 493 + ], + "type": "text", + "content": "Distinct from the widely adopted Neural Radiance Field, 3DGS [26] employs explicit 3D Gaussian points as its rendering primitives. A 3D Gaussian point is mathematically defined as:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 222, + 497, + 481, + 520 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 222, + 497, + 481, + 520 + ], + "spans": [ + { + "bbox": [ + 222, + 497, + 481, + 520 + ], + "type": "interline_equation", + "content": "G (\\boldsymbol {x}) = \\exp \\left(- \\frac {1}{2} (\\boldsymbol {x} - \\boldsymbol {\\mu}) ^ {\\top} \\Sigma^ {- 1} (\\boldsymbol {x} - \\boldsymbol {\\mu})\\right), \\tag {1}", + "image_path": "682dd1de11255147d8f386188699de18bcc91ac0e488c0093edb6aa9ba871fda.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 522, + 481, + 546 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 522, + 481, + 546 + ], + "spans": [ + { + "bbox": [ + 130, + 522, + 481, + 546 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 522, + 481, + 546 + ], + "type": "inline_equation", + "content": "\\pmb{\\mu}" + }, + { + "bbox": [ + 130, + 522, + 481, + 546 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 522, + 481, + 546 + ], + "type": "inline_equation", + "content": "\\Sigma" + }, + { + "bbox": [ + 130, + 522, + 481, + 546 + ], + "type": "text", + "content": " denote the 3D spatial mean and covariance matrix, respectively. Each Gaussian is also equipped with an opacity " + }, + { + "bbox": [ + 130, + 522, + 481, + 546 + ], + "type": "inline_equation", + "content": "o" + }, + { + "bbox": [ + 130, + 522, + 481, + 546 + ], + "type": "text", + "content": " and a view-dependent color " + }, + { + "bbox": [ + 130, + 522, + 481, + 546 + ], + "type": "inline_equation", + "content": "\\pmb{c}" + }, + { + "bbox": [ + 130, + 522, + 481, + 546 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 546, + 481, + 631 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 546, + 481, + 631 + ], + "spans": [ + { + "bbox": [ + 130, + 546, + 481, + 631 + ], + "type": "text", + "content": "The rendering process in 3DGS is divided into two main steps. Firstly, 3D Gaussians are projected to 2D Gaussians on the image plane. The 2D means are determined by accurate projection of 3D means, while the 2D covariance matrices are approximated by: " + }, + { + "bbox": [ + 130, + 546, + 481, + 631 + ], + "type": "inline_equation", + "content": "\\Sigma' = J\\boldsymbol{W} \\Sigma \\boldsymbol{W}^\\top \\boldsymbol{J}^\\top" + }, + { + "bbox": [ + 130, + 546, + 481, + 631 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 546, + 481, + 631 + ], + "type": "inline_equation", + "content": "\\boldsymbol{W}" + }, + { + "bbox": [ + 130, + 546, + 481, + 631 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 546, + 481, + 631 + ], + "type": "inline_equation", + "content": "\\boldsymbol{J}" + }, + { + "bbox": [ + 130, + 546, + 481, + 631 + ], + "type": "text", + "content": " denote the viewing transformation and the Jacobian of the affine approximation of perspective projection transformation [57]. Subsequently, the pixel color is derived by alpha blending " + }, + { + "bbox": [ + 130, + 546, + 481, + 631 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 546, + 481, + 631 + ], + "type": "text", + "content": " ordered 2D Gaussians from front to back:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 233, + 635, + 481, + 669 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 233, + 635, + 481, + 669 + ], + "spans": [ + { + "bbox": [ + 233, + 635, + 481, + 669 + ], + "type": "interline_equation", + "content": "\\mathcal {C} = \\sum_ {i \\in N} T _ {i} \\alpha_ {i} \\boldsymbol {c} _ {i}, \\quad T _ {i} = \\prod_ {j = 1} ^ {i - 1} (1 - \\alpha_ {j}). \\tag {2}", + "image_path": "3fe475e70a490971919957a42b0380c9617e3b9b93e5722f806f891f587aef76.jpg" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 219, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 219, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 219, + 101 + ], + "type": "text", + "content": "J. Gao et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 135, + 116, + 478, + 236 + ], + "blocks": [ + { + "bbox": [ + 135, + 116, + 478, + 236 + ], + "lines": [ + { + "bbox": [ + 135, + 116, + 478, + 236 + ], + "spans": [ + { + "bbox": [ + 135, + 116, + 478, + 236 + ], + "type": "image", + "image_path": "f0101fd25b550b2fd16a3b6c60da2720ceae47ec9045536a575998c4f6bd5039.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 244, + 482, + 322 + ], + "lines": [ + { + "bbox": [ + 130, + 244, + 482, + 322 + ], + "spans": [ + { + "bbox": [ + 130, + 244, + 482, + 322 + ], + "type": "text", + "content": "Fig. 2: The proposed differentiable rendering pipeline. Starting with a collection of 3D Gaussians that embody geometry, materials, and lighting attributes, we first execute the physically based rendering equation for each 3D Gaussian to determine its outgoing radiance, denoted as " + }, + { + "bbox": [ + 130, + 244, + 482, + 322 + ], + "type": "inline_equation", + "content": "c'" + }, + { + "bbox": [ + 130, + 244, + 482, + 322 + ], + "type": "text", + "content": ". Following this, we perform rasterization and alpha blending to obtain vanilla color map " + }, + { + "bbox": [ + 130, + 244, + 482, + 322 + ], + "type": "inline_equation", + "content": "\\mathcal{C}" + }, + { + "bbox": [ + 130, + 244, + 482, + 322 + ], + "type": "text", + "content": ", PBR color map " + }, + { + "bbox": [ + 130, + 244, + 482, + 322 + ], + "type": "inline_equation", + "content": "\\mathcal{C}'" + }, + { + "bbox": [ + 130, + 244, + 482, + 322 + ], + "type": "text", + "content": ", depth map " + }, + { + "bbox": [ + 130, + 244, + 482, + 322 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 130, + 244, + 482, + 322 + ], + "type": "text", + "content": ", normal map " + }, + { + "bbox": [ + 130, + 244, + 482, + 322 + ], + "type": "inline_equation", + "content": "\\mathcal{N}" + }, + { + "bbox": [ + 130, + 244, + 482, + 322 + ], + "type": "text", + "content": ", etc. To optimize relightable 3D Gaussians, we utilize the ground truth image " + }, + { + "bbox": [ + 130, + 244, + 482, + 322 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_{gt}" + }, + { + "bbox": [ + 130, + 244, + 482, + 322 + ], + "type": "text", + "content": " and the pseudo normal map derived from " + }, + { + "bbox": [ + 130, + 244, + 482, + 322 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 130, + 244, + 482, + 322 + ], + "type": "text", + "content": " for supervision." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 350, + 482, + 421 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 350, + 482, + 421 + ], + "spans": [ + { + "bbox": [ + 130, + 350, + 482, + 421 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 130, + 350, + 482, + 421 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 130, + 350, + 482, + 421 + ], + "type": "text", + "content": " is obtained by multiplying the opacity " + }, + { + "bbox": [ + 130, + 350, + 482, + 421 + ], + "type": "inline_equation", + "content": "o" + }, + { + "bbox": [ + 130, + 350, + 482, + 421 + ], + "type": "text", + "content": " with the 2D covariance's contribution computed from " + }, + { + "bbox": [ + 130, + 350, + 482, + 421 + ], + "type": "inline_equation", + "content": "\\Sigma^{\\prime}" + }, + { + "bbox": [ + 130, + 350, + 482, + 421 + ], + "type": "text", + "content": " and pixel coordinate in image space [26]. In implementation details, the covariance matrix " + }, + { + "bbox": [ + 130, + 350, + 482, + 421 + ], + "type": "inline_equation", + "content": "\\Sigma" + }, + { + "bbox": [ + 130, + 350, + 482, + 421 + ], + "type": "text", + "content": " is parameterized as a unit quaternion " + }, + { + "bbox": [ + 130, + 350, + 482, + 421 + ], + "type": "inline_equation", + "content": "\\pmb{q}" + }, + { + "bbox": [ + 130, + 350, + 482, + 421 + ], + "type": "text", + "content": " and a 3D scaling vector " + }, + { + "bbox": [ + 130, + 350, + 482, + 421 + ], + "type": "inline_equation", + "content": "\\pmb{s}" + }, + { + "bbox": [ + 130, + 350, + 482, + 421 + ], + "type": "text", + "content": " to maintain its meaningful interpretation throughout optimization. Additionally, view-dependent color " + }, + { + "bbox": [ + 130, + 350, + 482, + 421 + ], + "type": "inline_equation", + "content": "\\pmb{c}_i" + }, + { + "bbox": [ + 130, + 350, + 482, + 421 + ], + "type": "text", + "content": " is represented through a set of Spherical Harmonics (SH)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 423, + 481, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 423, + 481, + 449 + ], + "spans": [ + { + "bbox": [ + 130, + 423, + 481, + 449 + ], + "type": "text", + "content": "In summation, a 3D scene is represented by a collection of 3D Gaussians, with the " + }, + { + "bbox": [ + 130, + 423, + 481, + 449 + ], + "type": "inline_equation", + "content": "i^{th}" + }, + { + "bbox": [ + 130, + 423, + 481, + 449 + ], + "type": "text", + "content": " Gaussian " + }, + { + "bbox": [ + 130, + 423, + 481, + 449 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_i" + }, + { + "bbox": [ + 130, + 423, + 481, + 449 + ], + "type": "text", + "content": " parameterized as " + }, + { + "bbox": [ + 130, + 423, + 481, + 449 + ], + "type": "inline_equation", + "content": "\\{\\pmb{\\mu}_i, \\pmb{q}_i, \\pmb{s}_i, o_i, \\pmb{c}_i\\}" + }, + { + "bbox": [ + 130, + 423, + 481, + 449 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 475, + 283, + 488 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 475, + 283, + 488 + ], + "spans": [ + { + "bbox": [ + 131, + 475, + 283, + 488 + ], + "type": "text", + "content": "3.2 Geometry Enhancement" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 506, + 481, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 506, + 481, + 529 + ], + "spans": [ + { + "bbox": [ + 130, + 506, + 481, + 529 + ], + "type": "text", + "content": "Scene geometry, specifically surface normal, is essential for realistic physically based rendering, which will be discussed in Sec. 3.3." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 532, + 482, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 532, + 482, + 616 + ], + "spans": [ + { + "bbox": [ + 130, + 532, + 482, + 616 + ], + "type": "text", + "content": "Normal Estimation. We incorporate a normal attribute " + }, + { + "bbox": [ + 130, + 532, + 482, + 616 + ], + "type": "inline_equation", + "content": "\\pmb{n}" + }, + { + "bbox": [ + 130, + 532, + 482, + 616 + ], + "type": "text", + "content": " for each 3D Gaussian and try to solve it. A potential methodology treats the spatial mean of a 3D Gaussian as a conventional point, and executes normal estimation based on the local planar assumption. However, this approach cannot provide accurate normal estimation. The reasons are two fold: First, the reconstructed Gaussian point cloud is often sparse. More critically, the Gaussian points are naturally soft, which means these points are not precisely aligned with the object surface." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "To address these limitations, we propose to optimize " + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\pmb{n}" + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": " from an initial random vector via back-propagation for each 3D Gaussian. We perceive the depth of all Gaussians along a ray as a distribution, and estimate the pixel depth as the expectation of this distribution. Similarly, we determine the normal for each" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 342, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 342, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 342, + 91, + 448, + 102 + ], + "type": "text", + "content": "Relightable 3D Gaussians" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 287, + 129 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 287, + 129 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 287, + 129 + ], + "type": "text", + "content": "pixel. This process is described by:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 253, + 137, + 481, + 162 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 253, + 137, + 481, + 162 + ], + "spans": [ + { + "bbox": [ + 253, + 137, + 481, + 162 + ], + "type": "interline_equation", + "content": "\\{\\mathcal {D}, \\mathcal {N} \\} = \\sum_ {i \\in N} w _ {i} \\left\\{d _ {i}, \\boldsymbol {n} _ {i} \\right\\}, \\tag {3}", + "image_path": "2bcdda79f88dcb1ce9983c489487834fb2baf47fe1a70f8734ae04db3239644a.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 171, + 482, + 220 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 171, + 482, + 220 + ], + "spans": [ + { + "bbox": [ + 132, + 171, + 482, + 220 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 132, + 171, + 482, + 220 + ], + "type": "inline_equation", + "content": "d_{i}" + }, + { + "bbox": [ + 132, + 171, + 482, + 220 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 132, + 171, + 482, + 220 + ], + "type": "inline_equation", + "content": "\\pmb{n}_i" + }, + { + "bbox": [ + 132, + 171, + 482, + 220 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 132, + 171, + 482, + 220 + ], + "type": "inline_equation", + "content": "w_{i} = T_{i}\\alpha_{i} / \\sum_{i\\in N}T_{i}\\alpha_{i}" + }, + { + "bbox": [ + 132, + 171, + 482, + 220 + ], + "type": "text", + "content": " denote the depth, normal and weight of the point. We then encourage the consistency between the rendered normal " + }, + { + "bbox": [ + 132, + 171, + 482, + 220 + ], + "type": "inline_equation", + "content": "\\mathcal{N}" + }, + { + "bbox": [ + 132, + 171, + 482, + 220 + ], + "type": "text", + "content": " and the pseudo normal " + }, + { + "bbox": [ + 132, + 171, + 482, + 220 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathcal{N}}" + }, + { + "bbox": [ + 132, + 171, + 482, + 220 + ], + "type": "text", + "content": ", which is computed from the rendered depth " + }, + { + "bbox": [ + 132, + 171, + 482, + 220 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 132, + 171, + 482, + 220 + ], + "type": "text", + "content": " under the local planarity assumption. The normal consistency is quantified as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 269, + 228, + 481, + 242 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 269, + 228, + 481, + 242 + ], + "spans": [ + { + "bbox": [ + 269, + 228, + 481, + 242 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {n} = \\| \\mathcal {N} - \\tilde {\\mathcal {N}} \\| _ {2}. \\tag {4}", + "image_path": "099e5661ab5f500d567582562444e2fa650488875912f1656a7a121f4a0db93e.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 251, + 479, + 321 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 251, + 479, + 321 + ], + "spans": [ + { + "bbox": [ + 132, + 251, + 479, + 321 + ], + "type": "text", + "content": "Normal Gradient Based Densification. To achieve superior rendering fidelity in detail regions, vanilla 3DGS [26] densifies 3D Gaussians through the gradient of view space points. Drawing upon this, to improve normal recovery in thin regions, we introduce an additional densification criterion on the gradient of normals. Specifically, we densify Gaussians whose normal gradient exceeds a threshold " + }, + { + "bbox": [ + 132, + 251, + 479, + 321 + ], + "type": "inline_equation", + "content": "T_{n}" + }, + { + "bbox": [ + 132, + 251, + 479, + 321 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 323, + 479, + 358 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 323, + 479, + 358 + ], + "spans": [ + { + "bbox": [ + 132, + 323, + 479, + 358 + ], + "type": "text", + "content": "Constraint on Depth Distribution. Given our assumption of the object possessing an accurate surface, we enforce a constraint on the depth distribution by minimizing the uncertainty. The uncertainty is defined as:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 273, + 370, + 480, + 383 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 273, + 370, + 480, + 383 + ], + "spans": [ + { + "bbox": [ + 273, + 370, + 480, + 383 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {u} = \\mathcal {D} _ {s q} - \\mathcal {D} ^ {2}, \\tag {5}", + "image_path": "708cebe1764b19a76ddd0930552839c53fa55bd3df3089a4a26d4135a48d3422.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 388, + 479, + 423 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 388, + 479, + 423 + ], + "spans": [ + { + "bbox": [ + 132, + 388, + 479, + 423 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 132, + 388, + 479, + 423 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{sq} = \\sum_{i\\in N}w_id_i^2" + }, + { + "bbox": [ + 132, + 388, + 479, + 423 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 132, + 388, + 479, + 423 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 132, + 388, + 479, + 423 + ], + "type": "text", + "content": " is the depth estimation defined in Eq. 3. This constraint on the depth distribution drives Gaussian points to the object surface, thereby improving geometric reconstruction." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 424, + 479, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 424, + 479, + 449 + ], + "spans": [ + { + "bbox": [ + 132, + 424, + 479, + 449 + ], + "type": "text", + "content": "Object Mask Constraint. If there is a mask indicating the object, we can constrain the optimization by a binary cross entropy loss [43]:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 222, + 460, + 480, + 472 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 222, + 460, + 480, + 472 + ], + "spans": [ + { + "bbox": [ + 222, + 460, + 480, + 472 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {O} = - M \\log O - (1 - M) \\log (1 - O), \\tag {6}", + "image_path": "09b02cb1d46dccd0cdfb8337f63127589c4155169383f28da70557314293c5bd.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 482, + 479, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 482, + 479, + 506 + ], + "spans": [ + { + "bbox": [ + 132, + 482, + 479, + 506 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 132, + 482, + 479, + 506 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 132, + 482, + 479, + 506 + ], + "type": "text", + "content": " is the object mask and " + }, + { + "bbox": [ + 132, + 482, + 479, + 506 + ], + "type": "inline_equation", + "content": "O = \\sum_{i\\in N}T_i\\alpha_i" + }, + { + "bbox": [ + 132, + 482, + 479, + 506 + ], + "type": "text", + "content": ". With this constraint, " + }, + { + "bbox": [ + 132, + 482, + 479, + 506 + ], + "type": "inline_equation", + "content": "O" + }, + { + "bbox": [ + 132, + 482, + 479, + 506 + ], + "type": "text", + "content": " is forced to be aligned with the distribution of " + }, + { + "bbox": [ + 132, + 482, + 479, + 506 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 132, + 482, + 479, + 506 + ], + "type": "text", + "content": ", and so we get opaque surface." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 524, + 297, + 537 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 524, + 297, + 537 + ], + "spans": [ + { + "bbox": [ + 132, + 524, + 297, + 537 + ], + "type": "text", + "content": "3.3 BRDF and Light Modeling" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 544, + 479, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 544, + 479, + 581 + ], + "spans": [ + { + "bbox": [ + 132, + 544, + 479, + 581 + ], + "type": "text", + "content": "Rendering Equation. We use the rendering equation [24] to model light interaction with surfaces, accounting for their material properties and geometry. It is given by:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 200, + 589, + 480, + 613 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 200, + 589, + 480, + 613 + ], + "spans": [ + { + "bbox": [ + 200, + 589, + 480, + 613 + ], + "type": "interline_equation", + "content": "L _ {o} \\left(\\boldsymbol {\\omega} _ {o}, \\boldsymbol {x}\\right) = \\int_ {\\Omega} f \\left(\\boldsymbol {\\omega} _ {o}, \\boldsymbol {\\omega} _ {i}, \\boldsymbol {x}\\right) L _ {i} \\left(\\boldsymbol {\\omega} _ {i}, \\boldsymbol {x}\\right) \\left(\\boldsymbol {\\omega} _ {i} \\cdot \\boldsymbol {n}\\right) d \\boldsymbol {\\omega} _ {i}, \\tag {7}", + "image_path": "90a065041578f5ef1f13068ab2660fac0c62556af6577bb0a5e34c6a9159dbee.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 617, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 617, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 617, + 481, + 665 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 132, + 617, + 481, + 665 + ], + "type": "inline_equation", + "content": "\\pmb{x}" + }, + { + "bbox": [ + 132, + 617, + 481, + 665 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 132, + 617, + 481, + 665 + ], + "type": "inline_equation", + "content": "\\pmb{n}" + }, + { + "bbox": [ + 132, + 617, + 481, + 665 + ], + "type": "text", + "content": " are the surface point and its normal vector, " + }, + { + "bbox": [ + 132, + 617, + 481, + 665 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 132, + 617, + 481, + 665 + ], + "type": "text", + "content": " is the Bidirectional Reflectance Distribution Function (BRDF), and " + }, + { + "bbox": [ + 132, + 617, + 481, + 665 + ], + "type": "inline_equation", + "content": "L_{i}" + }, + { + "bbox": [ + 132, + 617, + 481, + 665 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 132, + 617, + 481, + 665 + ], + "type": "inline_equation", + "content": "L_{o}" + }, + { + "bbox": [ + 132, + 617, + 481, + 665 + ], + "type": "text", + "content": " denote the incoming and outgoing radiance in directions " + }, + { + "bbox": [ + 132, + 617, + 481, + 665 + ], + "type": "inline_equation", + "content": "\\omega_{i}" + }, + { + "bbox": [ + 132, + 617, + 481, + 665 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 132, + 617, + 481, + 665 + ], + "type": "inline_equation", + "content": "\\omega_{o}" + }, + { + "bbox": [ + 132, + 617, + 481, + 665 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 132, + 617, + 481, + 665 + ], + "type": "inline_equation", + "content": "\\Omega" + }, + { + "bbox": [ + 132, + 617, + 481, + 665 + ], + "type": "text", + "content": " signifies the hemispherical domain above the surface." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 218, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 218, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 218, + 101 + ], + "type": "text", + "content": "J. Gao et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 223 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 223 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 223 + ], + "type": "text", + "content": "Prior methods [49, 51] typically begin with the acquisition of intersection points between rays and surface through differentiable rendering. Then they apply the rendering equation at these points to facilitate Physical Based Rendering (PBR). However, these approach present significant challenges in the point-based framework. Certainly, it is feasible to extract surfaces in 3DGS [17]; but, it necessitates a coordinate-based incident light field for inverse PBR. The querying of this field at millions of points during every iteration imposes a substantial computational burden. Moreover, it is pertinent to acknowledge that accurately extracting geometric surfaces presents considerable challenges in 3DGS system." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 224, + 480, + 295 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 224, + 480, + 295 + ], + "spans": [ + { + "bbox": [ + 130, + 224, + 480, + 295 + ], + "type": "text", + "content": "To tackle this issue, we propose to compute PBR color " + }, + { + "bbox": [ + 130, + 224, + 480, + 295 + ], + "type": "inline_equation", + "content": "\\{c_i^{\\prime}\\}_{i = 0}^{N}" + }, + { + "bbox": [ + 130, + 224, + 480, + 295 + ], + "type": "text", + "content": " for each 3D Gaussian, and then obtain the PBR image " + }, + { + "bbox": [ + 130, + 224, + 480, + 295 + ], + "type": "inline_equation", + "content": "\\mathcal{C}'" + }, + { + "bbox": [ + 130, + 224, + 480, + 295 + ], + "type": "text", + "content": " through alpha-blending, as shown in Fig. 2. This method is more efficient for two main reasons. First, PBR is performed on fewer 3D Gaussians rather than image pixels in all input views, as each Gaussian affects multiple pixels based on its scale. Second, by allocating discrete attributes for each Gaussian, we avoid the need for global neural fields." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 296, + 481, + 353 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 296, + 481, + 353 + ], + "spans": [ + { + "bbox": [ + 130, + 296, + 481, + 353 + ], + "type": "text", + "content": "BRDF Parameterization. To make 3D Gaussians relightable, we assign BRDF properties to each Gaussian and adopt a simplified Disney BRDF model [11]. The BRDF properties include an albedo " + }, + { + "bbox": [ + 130, + 296, + 481, + 353 + ], + "type": "inline_equation", + "content": "\\pmb{b} \\in [0,1]^3" + }, + { + "bbox": [ + 130, + 296, + 481, + 353 + ], + "type": "text", + "content": " and a roughness " + }, + { + "bbox": [ + 130, + 296, + 481, + 353 + ], + "type": "inline_equation", + "content": "r \\in [0,1]" + }, + { + "bbox": [ + 130, + 296, + 481, + 353 + ], + "type": "text", + "content": ", and the BRDF " + }, + { + "bbox": [ + 130, + 296, + 481, + 353 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 130, + 296, + 481, + 353 + ], + "type": "text", + "content": " in Eq. 7 is divided into a diffuse term " + }, + { + "bbox": [ + 130, + 296, + 481, + 353 + ], + "type": "inline_equation", + "content": "f_{d} = \\frac{b}{\\pi}" + }, + { + "bbox": [ + 130, + 296, + 481, + 353 + ], + "type": "text", + "content": " and a specular term:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 202, + 353, + 481, + 378 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 202, + 353, + 481, + 378 + ], + "spans": [ + { + "bbox": [ + 202, + 353, + 481, + 378 + ], + "type": "interline_equation", + "content": "f _ {s} \\left(\\boldsymbol {\\omega} _ {o}, \\boldsymbol {\\omega} _ {i}\\right) = \\frac {D (\\boldsymbol {h} ; r) \\cdot F \\left(\\boldsymbol {\\omega} _ {o} , \\boldsymbol {h}\\right) \\cdot G \\left(\\boldsymbol {\\omega} _ {i} , \\boldsymbol {\\omega} _ {o} , h ; r\\right)}{\\left(\\boldsymbol {n} \\cdot \\boldsymbol {\\omega} _ {i}\\right) \\cdot \\left(\\boldsymbol {n} \\cdot \\boldsymbol {\\omega} _ {o}\\right)}, \\tag {8}", + "image_path": "7993436ca67ece3be218a8f4081ef787962f7e12a1f9954bdfc90f2a92fd7a25.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 381, + 479, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 381, + 479, + 403 + ], + "spans": [ + { + "bbox": [ + 130, + 381, + 479, + 403 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 381, + 479, + 403 + ], + "type": "inline_equation", + "content": "\\pmb{h} = (\\pmb{\\omega}_i + \\pmb{\\omega}_o) / 2" + }, + { + "bbox": [ + 130, + 381, + 479, + 403 + ], + "type": "text", + "content": " is the half vector, " + }, + { + "bbox": [ + 130, + 381, + 479, + 403 + ], + "type": "inline_equation", + "content": "D, F" + }, + { + "bbox": [ + 130, + 381, + 479, + 403 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 381, + 479, + 403 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 130, + 381, + 479, + 403 + ], + "type": "text", + "content": " denote the normal distribution function, Fresnel term and geometry term." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 406, + 479, + 477 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 406, + 479, + 477 + ], + "spans": [ + { + "bbox": [ + 130, + 406, + 479, + 477 + ], + "type": "text", + "content": "Incident Light Modeling. Optimizing a NeILF for each Gaussian can be excessively unconstrained, leading to challenges in accurately decomposing incident lights from its appearance. To address this issue, we apply a prior by partitioning the incident light into globally shared direct components and individual per-Gaussian indirect components. The sampled incident light at a Gaussian from direction " + }, + { + "bbox": [ + 130, + 406, + 479, + 477 + ], + "type": "inline_equation", + "content": "\\omega_{i}" + }, + { + "bbox": [ + 130, + 406, + 479, + 477 + ], + "type": "text", + "content": " is represented as:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 211, + 487, + 480, + 499 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 211, + 487, + 480, + 499 + ], + "spans": [ + { + "bbox": [ + 211, + 487, + 480, + 499 + ], + "type": "interline_equation", + "content": "L _ {i} (\\boldsymbol {\\omega} _ {i}) = V (\\boldsymbol {\\omega} _ {i}) \\cdot L _ {\\text {d i r e c t}} (\\boldsymbol {\\omega} _ {i}) + L _ {\\text {i n d i r e c t}} (\\boldsymbol {\\omega} _ {i}), \\tag {9}", + "image_path": "c3f8f9eb301ea6d06498e7a6cd24467d818657720ce092358e957747a5006398.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 507, + 480, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 507, + 480, + 590 + ], + "spans": [ + { + "bbox": [ + 130, + 507, + 480, + 590 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 507, + 480, + 590 + ], + "type": "inline_equation", + "content": "V(\\omega_{i})" + }, + { + "bbox": [ + 130, + 507, + 480, + 590 + ], + "type": "text", + "content": " represents the visibility term which will be further discussed in Sec. 4, the indirect light term " + }, + { + "bbox": [ + 130, + 507, + 480, + 590 + ], + "type": "inline_equation", + "content": "L_{\\text{indirect}}" + }, + { + "bbox": [ + 130, + 507, + 480, + 590 + ], + "type": "text", + "content": " is parameterized by 3-level SH, denoted as " + }, + { + "bbox": [ + 130, + 507, + 480, + 590 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 130, + 507, + 480, + 590 + ], + "type": "text", + "content": ", while the direct light term " + }, + { + "bbox": [ + 130, + 507, + 480, + 590 + ], + "type": "inline_equation", + "content": "L_{\\text{direct}}" + }, + { + "bbox": [ + 130, + 507, + 480, + 590 + ], + "type": "text", + "content": " is parameterized as a globally shared 16x32 environment map, denoted as " + }, + { + "bbox": [ + 130, + 507, + 480, + 590 + ], + "type": "inline_equation", + "content": "l^{env}" + }, + { + "bbox": [ + 130, + 507, + 480, + 590 + ], + "type": "text", + "content": ". Despite the utilization of relatively low-level Spherical Harmonics (SH), we can still capture some inter-reflection effects, as the rendering color of a given pixel arises from the collaboration of multiple relightable 3D Gaussians." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 591, + 479, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 591, + 479, + 628 + ], + "spans": [ + { + "bbox": [ + 130, + 591, + 479, + 628 + ], + "type": "text", + "content": "For each 3D Gaussian, we sample " + }, + { + "bbox": [ + 130, + 591, + 479, + 628 + ], + "type": "inline_equation", + "content": "N_{s}" + }, + { + "bbox": [ + 130, + 591, + 479, + 628 + ], + "type": "text", + "content": " incident directions over the hemisphere space through Fibonacci sampling [45] to provide numerical integration for Eq. 7. The PBR color of each Gaussian is then given by:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 200, + 635, + 481, + 667 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 200, + 635, + 481, + 667 + ], + "spans": [ + { + "bbox": [ + 200, + 635, + 481, + 667 + ], + "type": "interline_equation", + "content": "\\boldsymbol {c} ^ {\\prime} \\left(\\boldsymbol {\\omega} _ {o}\\right) = \\sum_ {i = 0} ^ {N _ {s}} \\left(f _ {d} + f _ {s} \\left(\\boldsymbol {\\omega} _ {o}, \\boldsymbol {\\omega} _ {i}\\right)\\right) L _ {i} \\left(\\boldsymbol {\\omega} _ {i}\\right) \\left(\\boldsymbol {\\omega} _ {i} \\cdot \\boldsymbol {n}\\right) \\Delta \\boldsymbol {\\omega} _ {i}, \\tag {10}", + "image_path": "d13b29cc088851b817391684464a880a45a8b2bc1ae5413615222d34fb0a66af.jpg" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 342, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 342, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 342, + 91, + 448, + 102 + ], + "type": "text", + "content": "Relightable 3D Gaussians" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 253, + 128 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 253, + 128 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 253, + 128 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 132, + 116, + 253, + 128 + ], + "type": "inline_equation", + "content": "\\omega_{i}" + }, + { + "bbox": [ + 132, + 116, + 253, + 128 + ], + "type": "text", + "content": " is the solid angle." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 129, + 482, + 165 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 129, + 482, + 165 + ], + "spans": [ + { + "bbox": [ + 130, + 129, + 482, + 165 + ], + "type": "text", + "content": "To summarize, our method represents a 3D scene as a set of relightable 3D Gaussians and a global environment light " + }, + { + "bbox": [ + 130, + 129, + 482, + 165 + ], + "type": "inline_equation", + "content": "l^{env}" + }, + { + "bbox": [ + 130, + 129, + 482, + 165 + ], + "type": "text", + "content": ", where the " + }, + { + "bbox": [ + 130, + 129, + 482, + 165 + ], + "type": "inline_equation", + "content": "i^{th}" + }, + { + "bbox": [ + 130, + 129, + 482, + 165 + ], + "type": "text", + "content": " Gaussian " + }, + { + "bbox": [ + 130, + 129, + 482, + 165 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_i" + }, + { + "bbox": [ + 130, + 129, + 482, + 165 + ], + "type": "text", + "content": " is parameterized as " + }, + { + "bbox": [ + 130, + 129, + 482, + 165 + ], + "type": "inline_equation", + "content": "\\{\\pmb{\\mu}_i, \\pmb{q}_i, \\pmb{s}_i, o_i, \\pmb{c}_i, \\pmb{n}_i, \\pmb{b}_i, r_i, l_i\\}" + }, + { + "bbox": [ + 130, + 129, + 482, + 165 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 182, + 239, + 194 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 182, + 239, + 194 + ], + "spans": [ + { + "bbox": [ + 132, + 182, + 239, + 194 + ], + "type": "text", + "content": "3.4 Regularizations" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 202, + 479, + 226 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 202, + 479, + 226 + ], + "spans": [ + { + "bbox": [ + 130, + 202, + 479, + 226 + ], + "type": "text", + "content": "To mitigate the materials-lighting ambiguity [45], proper regularizations are utilized to facilitate their plausible decomposition." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 226, + 480, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 226, + 480, + 251 + ], + "spans": [ + { + "bbox": [ + 130, + 226, + 480, + 251 + ], + "type": "text", + "content": "Light Regularization. We apply a light regularization assuming a near-natural white incident light [30]:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 210, + 259, + 481, + 283 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 210, + 259, + 481, + 283 + ], + "spans": [ + { + "bbox": [ + 210, + 259, + 481, + 283 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {l i g h t}} = \\sum_ {c} \\left(L _ {c} - \\frac {1}{3} \\sum_ {c} L _ {c}\\right), c \\in \\{R, G, B \\}. \\tag {11}", + "image_path": "3e86aea4d39e6bd451ee817f320a7c093a2a11d57c787fea1ea7874944e78d8f.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 289, + 479, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 289, + 479, + 323 + ], + "spans": [ + { + "bbox": [ + 131, + 289, + 479, + 323 + ], + "type": "text", + "content": "Smoothness Priors. We expect that the BRDF properties do not change drastically in homogeneous areas [45]. We define a smooth constraint on roughness as:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 244, + 326, + 480, + 339 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 244, + 326, + 480, + 339 + ], + "spans": [ + { + "bbox": [ + 244, + 326, + 480, + 339 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {s, r} = \\| \\nabla R \\| \\exp (- \\| \\nabla C _ {g t} \\|), \\tag {12}", + "image_path": "3b54871a4ab99172b316bc93f681493e0a4d5eb00e7aea9a2df61e94d7a42b3e.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 343, + 479, + 369 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 343, + 479, + 369 + ], + "spans": [ + { + "bbox": [ + 130, + 343, + 479, + 369 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 343, + 479, + 369 + ], + "type": "inline_equation", + "content": "R = \\sum_{i\\in N}w_i r_i" + }, + { + "bbox": [ + 130, + 343, + 479, + 369 + ], + "type": "text", + "content": " is the rendered roughness map. Similarly, we also define smoothness constraints " + }, + { + "bbox": [ + 130, + 343, + 479, + 369 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{s,n}" + }, + { + "bbox": [ + 130, + 343, + 479, + 369 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 343, + 479, + 369 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{s,b}" + }, + { + "bbox": [ + 130, + 343, + 479, + 369 + ], + "type": "text", + "content": " on normal and albedo." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 386, + 301, + 401 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 386, + 301, + 401 + ], + "spans": [ + { + "bbox": [ + 132, + 386, + 301, + 401 + ], + "type": "text", + "content": "4 Point-based Ray Tracing" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 131, + 412, + 479, + 436 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 412, + 479, + 436 + ], + "spans": [ + { + "bbox": [ + 131, + 412, + 479, + 436 + ], + "type": "text", + "content": "For real-time realistic rendering with plausible shadow effects on relightable 3D Gaussians, we introduce a novel point-based ray tracing approach in this section." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 453, + 310, + 466 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 453, + 310, + 466 + ], + "spans": [ + { + "bbox": [ + 132, + 453, + 310, + 466 + ], + "type": "text", + "content": "4.1 Ray Tracing on 3D Gaussians" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 474, + 480, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 474, + 480, + 557 + ], + "spans": [ + { + "bbox": [ + 130, + 474, + 480, + 557 + ], + "type": "text", + "content": "Our proposed ray tracing technique on 3D Gaussians is built upon the Bounding Volume Hierarchy (BVH), enabling efficient visibility querying along a ray. Our method adopts the idea from [25], an in-place algorithm for constructing a binary radix tree that maximizes parallelism and facilitates real-time BVH construction. Specifically, we construct a binary radix tree from a given set of 3D Gaussians, where each leaf node represents the tight bounding box of a Gaussian, and each internal node denotes the bounding box of its two children." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 558, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 558, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 558, + 481, + 666 + ], + "type": "text", + "content": "Unlike ray tracing with opaque polygonal meshes, where only the closest intersection point is required, ray tracing with semi-transparent Gaussians necessitates accounting for all Gaussians potentially influencing the ray's transmittance. The process of ray tracing on 3D Gaussian points can be described as follows: the ray travels from the camera center and accumulates transmittance as it passes through 3D Gaussian points until the transmittance is zero. The first key issue to be addressed is how to define the contribution of a single 3D Gaussian to the transmittance. As discussed in Sec.3.1, transforming a 3D Gaussian to a 2D Gaussian involves an approximation, which complicating the" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 219, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 219, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 219, + 101 + ], + "type": "text", + "content": "J. Gao et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 132, + 118, + 216, + 232 + ], + "blocks": [ + { + "bbox": [ + 132, + 118, + 216, + 232 + ], + "lines": [ + { + "bbox": [ + 132, + 118, + 216, + 232 + ], + "spans": [ + { + "bbox": [ + 132, + 118, + 216, + 232 + ], + "type": "image", + "image_path": "7dd81d7ea48234217c81c0676ca8bebe9540cb52587a70d919390505ba9f1bda.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 147, + 233, + 201, + 243 + ], + "lines": [ + { + "bbox": [ + 147, + 233, + 201, + 243 + ], + "spans": [ + { + "bbox": [ + 147, + 233, + 201, + 243 + ], + "type": "text", + "content": "(a) Root Node" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 220, + 118, + 304, + 232 + ], + "blocks": [ + { + "bbox": [ + 220, + 118, + 304, + 232 + ], + "lines": [ + { + "bbox": [ + 220, + 118, + 304, + 232 + ], + "spans": [ + { + "bbox": [ + 220, + 118, + 304, + 232 + ], + "type": "image", + "image_path": "1ea1a93651801d076c5cd91dc0dd289887767b15d82ac2bfaec1c3d2d36157c0.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 235, + 233, + 291, + 243 + ], + "lines": [ + { + "bbox": [ + 235, + 233, + 291, + 243 + ], + "spans": [ + { + "bbox": [ + 235, + 233, + 291, + 243 + ], + "type": "text", + "content": "(b) Child Node" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 310, + 118, + 392, + 232 + ], + "blocks": [ + { + "bbox": [ + 310, + 118, + 392, + 232 + ], + "lines": [ + { + "bbox": [ + 310, + 118, + 392, + 232 + ], + "spans": [ + { + "bbox": [ + 310, + 118, + 392, + 232 + ], + "type": "image", + "image_path": "71c5c1a930d868cf6692fce5aa54ed6ad19c3852eea3f3ba3c145f0051dbcc82.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 328, + 233, + 375, + 243 + ], + "lines": [ + { + "bbox": [ + 328, + 233, + 375, + 243 + ], + "spans": [ + { + "bbox": [ + 328, + 233, + 375, + 243 + ], + "type": "text", + "content": "(c) Gaussian" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 131, + 252, + 482, + 308 + ], + "lines": [ + { + "bbox": [ + 131, + 252, + 482, + 308 + ], + "spans": [ + { + "bbox": [ + 131, + 252, + 482, + 308 + ], + "type": "text", + "content": "Fig. 3: Intersection tests in point-based ray tracing. Intersection point between ray and Gaussian is obtained by three steps: (a) intersect the BVH root node; (b) dive into the intersected child nodes recursively until the leaf node; (c) perform Eq. 13 to get the equivalent intersection point. (d) shows that a 3D Gaussian actually has non-negligible effect on a segment of a ray." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 399, + 118, + 482, + 232 + ], + "blocks": [ + { + "bbox": [ + 399, + 118, + 482, + 232 + ], + "lines": [ + { + "bbox": [ + 399, + 118, + 482, + 232 + ], + "spans": [ + { + "bbox": [ + 399, + 118, + 482, + 232 + ], + "type": "image", + "image_path": "f366c2a3d2b2fb588bf8a5819989d4f17b66aaf33f9e9f32e9464b3a32440977.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 416, + 233, + 463, + 243 + ], + "lines": [ + { + "bbox": [ + 416, + 233, + 463, + 243 + ], + "spans": [ + { + "bbox": [ + 416, + 233, + 463, + 243 + ], + "type": "text", + "content": "(d) Gaussian" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 331, + 479, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 331, + 479, + 392 + ], + "spans": [ + { + "bbox": [ + 130, + 331, + 479, + 392 + ], + "type": "text", + "content": "identification of accurate intersection between the Gaussian and a ray. As shown in Fig. 3d, a 3D Gaussian actually has non-negligible effect on a segment of a ray. Thus, drawing inspiration from [27], we approximate the intersection of ray with 3D Gaussian as a point where the 3D Gaussian's contribution peaks, as showed in Fig. 3(3). The equivalent intersection point is defined as:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 270, + 403, + 480, + 416 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 270, + 403, + 480, + 416 + ], + "spans": [ + { + "bbox": [ + 270, + 403, + 480, + 416 + ], + "type": "interline_equation", + "content": "\\boldsymbol {r} _ {\\boldsymbol {x}} = \\boldsymbol {r} _ {\\boldsymbol {o}} + t _ {j} \\boldsymbol {r} _ {\\boldsymbol {d}}, \\tag {13}", + "image_path": "8043f7a0d4066d7017b4df4120b81ae736baf03630d02f4fc30ebfb105947c98.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 421, + 479, + 446 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 421, + 479, + 446 + ], + "spans": [ + { + "bbox": [ + 130, + 421, + 479, + 446 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 421, + 479, + 446 + ], + "type": "inline_equation", + "content": "\\pmb{r_o}" + }, + { + "bbox": [ + 130, + 421, + 479, + 446 + ], + "type": "text", + "content": " denotes the origin, " + }, + { + "bbox": [ + 130, + 421, + 479, + 446 + ], + "type": "inline_equation", + "content": "\\pmb{r_d}" + }, + { + "bbox": [ + 130, + 421, + 479, + 446 + ], + "type": "text", + "content": " is direction of the ray, corresponding to the previously mentioned " + }, + { + "bbox": [ + 130, + 421, + 479, + 446 + ], + "type": "inline_equation", + "content": "\\omega_{i}, t_{j}" + }, + { + "bbox": [ + 130, + 421, + 479, + 446 + ], + "type": "text", + "content": " is defined as:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 259, + 454, + 480, + 482 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 259, + 454, + 480, + 482 + ], + "spans": [ + { + "bbox": [ + 259, + 454, + 480, + 482 + ], + "type": "interline_equation", + "content": "t _ {j} = \\frac {\\left(\\boldsymbol {\\mu} - \\boldsymbol {r} _ {o}\\right) ^ {T} \\boldsymbol {\\Sigma} \\boldsymbol {r} _ {d}}{\\boldsymbol {r} _ {d} ^ {T} \\boldsymbol {\\Sigma} \\boldsymbol {r} _ {d}}. \\tag {14}", + "image_path": "094d5cca1eb0e479ffbced1c592be1d04417eb2ac1b4d059f87e373a18aedd4c.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 488, + 479, + 512 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 488, + 479, + 512 + ], + "spans": [ + { + "bbox": [ + 130, + 488, + 479, + 512 + ], + "type": "text", + "content": "Subsequently, we approximate the contribution of this 3D Gaussian to the ray as " + }, + { + "bbox": [ + 130, + 488, + 479, + 512 + ], + "type": "inline_equation", + "content": "\\alpha_{j}" + }, + { + "bbox": [ + 130, + 488, + 479, + 512 + ], + "type": "text", + "content": " at the equivalent intersection point " + }, + { + "bbox": [ + 130, + 488, + 479, + 512 + ], + "type": "inline_equation", + "content": "\\pmb{r}_{\\pmb{x}}" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 514, + 482, + 609 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 514, + 482, + 609 + ], + "spans": [ + { + "bbox": [ + 130, + 514, + 482, + 609 + ], + "type": "text", + "content": "Considering the transmittance along a ray: " + }, + { + "bbox": [ + 130, + 514, + 482, + 609 + ], + "type": "inline_equation", + "content": "T_{i} = \\prod_{j=1}^{i-1}(1 - \\alpha_{j})" + }, + { + "bbox": [ + 130, + 514, + 482, + 609 + ], + "type": "text", + "content": ", it is evident that the order of " + }, + { + "bbox": [ + 130, + 514, + 482, + 609 + ], + "type": "inline_equation", + "content": "\\alpha_{j}" + }, + { + "bbox": [ + 130, + 514, + 482, + 609 + ], + "type": "text", + "content": " does not affect " + }, + { + "bbox": [ + 130, + 514, + 482, + 609 + ], + "type": "inline_equation", + "content": "T_{i}" + }, + { + "bbox": [ + 130, + 514, + 482, + 609 + ], + "type": "text", + "content": ". In other words, the order in which Gaussians are encountered along a ray does not impact the overall transmittance. As illustrated in Fig. 3, starting from the root node of the binary radix tree, intersection tests are recursively performed between the ray and the bounding volumes of each node's children. Upon reaching a leaf node, the associated Gaussian is identified. Through this traversal, the transmittance " + }, + { + "bbox": [ + 130, + 514, + 482, + 609 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 130, + 514, + 482, + 609 + ], + "type": "text", + "content": " is progressively attenuated:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 184, + 619, + 480, + 632 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 184, + 619, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 184, + 619, + 480, + 632 + ], + "type": "interline_equation", + "content": "T _ {i} = \\left(1 - \\alpha_ {i - 1}\\right) T _ {i - 1}, \\quad \\text {f o r} i = 1, \\dots , j - 1 \\text {w i t h} T _ {1} = 1. \\tag {15}", + "image_path": "af990c2c10f7aa9aa3e45608bdc4ac7b548744dc84e9f46735fc9b45cfcdc9bd.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 130, + 641, + 479, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 479, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 479, + 666 + ], + "type": "text", + "content": "To speed up ray tracing, the process is early terminated if a ray's transmittance drops below a certain threshold " + }, + { + "bbox": [ + 130, + 641, + 479, + 666 + ], + "type": "inline_equation", + "content": "T_{\\min}" + }, + { + "bbox": [ + 130, + 641, + 479, + 666 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 342, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 342, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 342, + 91, + 448, + 102 + ], + "type": "text", + "content": "Relightable 3D Gaussians" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 294, + 128 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 294, + 128 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 294, + 128 + ], + "type": "text", + "content": "4.2 Visibility Pre-computation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 137, + 481, + 221 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 137, + 481, + 221 + ], + "spans": [ + { + "bbox": [ + 130, + 137, + 481, + 221 + ], + "type": "text", + "content": "In Sec. 3.3, a potential ambiguity in the decomposition of indirect and direct lights has been discerned. While the essential visibility term can be derived via the proposed ray tracing on 3D Gaussians (Sec. 4.1), querying the visibility term through ray tracing online proves challenging due to the computational complexity. Nevertheless, given our exclusive focus on static scenes, we can pre-compute the visibility term " + }, + { + "bbox": [ + 130, + 137, + 481, + 221 + ], + "type": "inline_equation", + "content": "V(\\omega_{i})" + }, + { + "bbox": [ + 130, + 137, + 481, + 221 + ], + "type": "text", + "content": " across the hemispherical domain determined by " + }, + { + "bbox": [ + 130, + 137, + 481, + 221 + ], + "type": "inline_equation", + "content": "\\pmb{n}" + }, + { + "bbox": [ + 130, + 137, + 481, + 221 + ], + "type": "text", + "content": " for each Gaussian, and subsequently integrate it into the rendering equation." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 221, + 481, + 293 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 221, + 481, + 293 + ], + "spans": [ + { + "bbox": [ + 130, + 221, + 481, + 293 + ], + "type": "text", + "content": "Hence, we divide the optimization process into two stages. In the first stage, we optimize a vanilla 3D Gaussian point cloud, augmented with an additional normal vector " + }, + { + "bbox": [ + 130, + 221, + 481, + 293 + ], + "type": "inline_equation", + "content": "\\pmb{n}" + }, + { + "bbox": [ + 130, + 221, + 481, + 293 + ], + "type": "text", + "content": " for each Gaussian. Immediately afterwards, we pre-compute the per-Gaussian visibility through the proposed ray tracing. In the second stage, we lock the geometry of 3D Gaussians and focus solely on optimizing the material and lighting parameters using the pipeline described in Fig. 2." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 312, + 322, + 325 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 312, + 322, + 325 + ], + "spans": [ + { + "bbox": [ + 132, + 312, + 322, + 325 + ], + "type": "text", + "content": "4.3 Point Based Relighting Pipeline" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 334, + 481, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 334, + 481, + 430 + ], + "spans": [ + { + "bbox": [ + 130, + 334, + 481, + 430 + ], + "type": "text", + "content": "Based on our relightable 3D Gaussians, we devise a point-based graphics pipeline that seamlessly integrates effortless scene editing and realistic relighting. To our knowledge, there is currently no point-based rendering pipeline that effectively accomplishes both tasks. Although explicit point representations make scene editing easy, achieving photo-realistic relighting proved nearly insurmountable in existing point-based rendering pipeline. Conversely, in inverse rendering approaches based on implicit representations, while highly realistic relighting is feasible, scene editing presents a challenging endeavor." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 430, + 482, + 550 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 430, + 482, + 550 + ], + "spans": [ + { + "bbox": [ + 130, + 430, + 482, + 550 + ], + "type": "text", + "content": "As an illustration, we concentrate on relighting in a multi-object composition scene. Initially, our pipeline computes the visibility " + }, + { + "bbox": [ + 130, + 430, + 482, + 550 + ], + "type": "inline_equation", + "content": "V(\\omega_{i})" + }, + { + "bbox": [ + 130, + 430, + 482, + 550 + ], + "type": "text", + "content": " for each Gaussian point via ray tracing (Sec. 4.1). Despite the challenging inter-object occlusions introduced by composition, our proposed point-based ray tracing method adeptly manages these complexities and ensures accurate occlusion updates within the novel scene. Subsequently, the rendering process unfolds (Sec. 3.3), starting with Gaussian-level PBR and ending with alpha blending. Throughout this process, the original indirect lighting of each object is discarded. Consequently, we achieve relighting with remarkably vivid shadow effects solely based on a discrete set of points, as illustrated in Fig. 1." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 570, + 230, + 582 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 570, + 230, + 582 + ], + "spans": [ + { + "bbox": [ + 132, + 570, + 230, + 582 + ], + "type": "text", + "content": "5 Experiments" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 596, + 243, + 608 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 596, + 243, + 608 + ], + "spans": [ + { + "bbox": [ + 132, + 596, + 243, + 608 + ], + "type": "text", + "content": "5.1 Training Details" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "type": "text", + "content": "As pointed out in Sec. 4.2, the training procedure is divided into two stages. The first stage follows the setting of 3DGS [26], along with the proposed normal gradient-based densification (Sec. 3.2) where " + }, + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "type": "inline_equation", + "content": "T_{\\mathbf{n}}" + }, + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "type": "text", + "content": " is set as " + }, + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "type": "inline_equation", + "content": "2 \\times 10^{-9}" + }, + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "type": "text", + "content": ". In the second stage, we sample " + }, + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "type": "inline_equation", + "content": "N_{s} = 64" + }, + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "type": "text", + "content": " incident rays per Gaussian point for PBR. We train" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "text", + "content": "J. Gao et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 203, + 136, + 410, + 232 + ], + "blocks": [ + { + "bbox": [ + 140, + 114, + 473, + 125 + ], + "lines": [ + { + "bbox": [ + 140, + 114, + 473, + 125 + ], + "spans": [ + { + "bbox": [ + 140, + 114, + 473, + 125 + ], + "type": "text", + "content": "Table 1: Quantitative results for novel view synthesis on NeRF synthetic dataset" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 203, + 136, + 410, + 232 + ], + "lines": [ + { + "bbox": [ + 203, + 136, + 410, + 232 + ], + "spans": [ + { + "bbox": [ + 203, + 136, + 410, + 232 + ], + "type": "table", + "html": "
GeometryRelightablePSNR↑ SSIM↑ LPIPS
NPBG [1]point×28.100.9230.077
NPBG++ [38]point×28.120.9280.076
FreqPCR [54]point×31.240.9500.049
3DGS [26]point×33.880.9700.031
PhySG [51]neural✓18.910.8470.182
NeILF++ [49]neural✓26.370.9110.091
Nvdiffrec [33]mesh✓29.050.9390.081
R3DG(Ours)point✓31.220.9590.039
", + "image_path": "dfc6f587526f1d16cd3d279c42c0818c98e03b01f67b16d14002c5e0fff82942.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 140, + 240, + 242, + 354 + ], + "blocks": [ + { + "bbox": [ + 140, + 240, + 242, + 354 + ], + "lines": [ + { + "bbox": [ + 140, + 240, + 242, + 354 + ], + "spans": [ + { + "bbox": [ + 140, + 240, + 242, + 354 + ], + "type": "image", + "image_path": "691c40cc11a766ec29a9963794e6671aa4e0a4ef875e26e1e366231387f26dff.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 176, + 355, + 209, + 365 + ], + "lines": [ + { + "bbox": [ + 176, + 355, + 209, + 365 + ], + "spans": [ + { + "bbox": [ + 176, + 355, + 209, + 365 + ], + "type": "text", + "content": "(a) Lego" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 259, + 240, + 351, + 354 + ], + "blocks": [ + { + "bbox": [ + 259, + 240, + 351, + 354 + ], + "lines": [ + { + "bbox": [ + 259, + 240, + 351, + 354 + ], + "spans": [ + { + "bbox": [ + 259, + 240, + 351, + 354 + ], + "type": "image", + "image_path": "d512835caae51e69be0bd4764e61a50ef532b082695459b4484b1db87d5a5b21.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 289, + 355, + 325, + 365 + ], + "lines": [ + { + "bbox": [ + 289, + 355, + 325, + 365 + ], + "spans": [ + { + "bbox": [ + 289, + 355, + 325, + 365 + ], + "type": "text", + "content": "(b) Chair" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 363, + 248, + 478, + 348 + ], + "blocks": [ + { + "bbox": [ + 363, + 248, + 478, + 348 + ], + "lines": [ + { + "bbox": [ + 363, + 248, + 478, + 348 + ], + "spans": [ + { + "bbox": [ + 363, + 248, + 478, + 348 + ], + "type": "image", + "image_path": "5172a9fff6cd0cdfd0864a9e17ed364e479df8dd62e942fb360bd3a5a15597cb.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 401, + 355, + 442, + 365 + ], + "lines": [ + { + "bbox": [ + 401, + 355, + 442, + 365 + ], + "spans": [ + { + "bbox": [ + 401, + 355, + 442, + 365 + ], + "type": "text", + "content": "(c) Hotdog" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 373, + 482, + 407 + ], + "lines": [ + { + "bbox": [ + 130, + 373, + 482, + 407 + ], + "spans": [ + { + "bbox": [ + 130, + 373, + 482, + 407 + ], + "type": "text", + "content": "Fig. 4: Visualizations on NeRF synthetic dataset [31]. Each scene is displayed in an order from left to right and from top to bottom: Ground Truth, PBR Image, Normal Map and Ambient Occlusion Map." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 431, + 480, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 431, + 480, + 456 + ], + "spans": [ + { + "bbox": [ + 130, + 431, + 480, + 456 + ], + "type": "text", + "content": "our model for 30,000 iterations in the initial stage and 10,000 iterations in the second stage. The loss in the first stage is calculated by:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 173, + 464, + 480, + 478 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 173, + 464, + 480, + 478 + ], + "spans": [ + { + "bbox": [ + 173, + 464, + 480, + 478 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\lambda_ {1} \\mathcal {L} _ {1} + \\lambda_ {s s i m} \\mathcal {L} _ {s s i m} + \\lambda_ {n} \\mathcal {L} _ {n} + \\lambda_ {s, n} \\mathcal {L} _ {s, n} + \\lambda_ {O} \\mathcal {L} _ {O} + \\lambda_ {u} \\mathcal {L} _ {u}, \\tag {16}", + "image_path": "80ad13725c78d19ca58fdc611fa48a5a379f6afc75dd6dc1738f066a173fb1da.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 487, + 481, + 512 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 487, + 481, + 512 + ], + "spans": [ + { + "bbox": [ + 130, + 487, + 481, + 512 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 487, + 481, + 512 + ], + "type": "inline_equation", + "content": "\\{\\lambda_1, \\lambda_{ssim}, \\lambda_n, \\lambda_{s,n}, \\lambda_O, \\lambda_u\\}" + }, + { + "bbox": [ + 130, + 487, + 481, + 512 + ], + "type": "text", + "content": " are set to " + }, + { + "bbox": [ + 130, + 487, + 481, + 512 + ], + "type": "inline_equation", + "content": "\\{0.8, 0.2, 0.01, 0.01, 0.01, 0.01\\}" + }, + { + "bbox": [ + 130, + 487, + 481, + 512 + ], + "type": "text", + "content": ", respectively. The loss in the second stage is given by:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 190, + 521, + 480, + 534 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 190, + 521, + 480, + 534 + ], + "spans": [ + { + "bbox": [ + 190, + 521, + 480, + 534 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\lambda_ {1} \\mathcal {L} _ {1} + \\lambda_ {s s i m} \\mathcal {L} _ {s s i m} + \\lambda_ {l} \\mathcal {L} _ {l} + \\lambda_ {s, b} \\mathcal {L} _ {s, b} + \\lambda_ {s, r} \\mathcal {L} _ {s, r}, \\tag {17}", + "image_path": "a3aa5e6d64b650a89cce5ddba44971bc7e4b4f89fad870299c814c16587995f0.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 543, + 482, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 543, + 482, + 556 + ], + "spans": [ + { + "bbox": [ + 130, + 543, + 482, + 556 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 543, + 482, + 556 + ], + "type": "inline_equation", + "content": "\\{\\lambda_1,\\lambda_{ssim},\\lambda_l,\\lambda_{s,b},\\lambda_{s,r}\\}" + }, + { + "bbox": [ + 130, + 543, + 482, + 556 + ], + "type": "text", + "content": " are set to " + }, + { + "bbox": [ + 130, + 543, + 482, + 556 + ], + "type": "inline_equation", + "content": "\\{0.8,0.2,0.0001,0.01,0.01\\}" + }, + { + "bbox": [ + 130, + 543, + 482, + 556 + ], + "type": "text", + "content": ", respectively." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 131, + 573, + 353, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 573, + 353, + 586 + ], + "spans": [ + { + "bbox": [ + 131, + 573, + 353, + 586 + ], + "type": "text", + "content": "5.2 Performance on Novel View Synthesis" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": "We first evaluate the novel view synthesis (NVS) of the physically based rendering (PBR) on NeRF synthetic dataset [31]. We compare both point-based rendering methods and relightable rendering approaches. We report average metrics across all scenes in Tab. 1, including peak signal-to-noise ratio (PSNR), structural similarity index measure (SSIM), and learned perceptual image patch similarity (LPIPS). Compared with existing point-based rendering methods, our" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 342, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 342, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 342, + 91, + 448, + 102 + ], + "type": "text", + "content": "Relightable 3D Gaussians" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 134, + 136, + 480, + 203 + ], + "blocks": [ + { + "bbox": [ + 176, + 114, + 437, + 126 + ], + "lines": [ + { + "bbox": [ + 176, + 114, + 437, + 126 + ], + "spans": [ + { + "bbox": [ + 176, + 114, + 437, + 126 + ], + "type": "text", + "content": "Table 2: Quantitative results on Synthetic4Relight dataset [56]" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 134, + 136, + 480, + 203 + ], + "lines": [ + { + "bbox": [ + 134, + 136, + 480, + 203 + ], + "spans": [ + { + "bbox": [ + 134, + 136, + 480, + 203 + ], + "type": "table", + "html": "
Novel View SynthesisRelightingAlbedoRoughnessTime
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓MSE↓(hours)
NerFactor [53]22.800.9160.15021.540.8750.17119.490.8640.206N/A>48
Nvdifferc-MC [19]34.290.9670.06824.220.9430.07829.610.9450.0750.0094.17
InvRender [56]30.740.9530.08628.670.9500.09128.280.9350.0720.00814.3
TensorIR [23]35.800.9780.04929.690.9510.07930.580.9460.0650.0153.24
R3DG (Ours)36.800.9820.02831.000.9640.05028.310.9510.0580.0130.90
", + "image_path": "259d2722551fd50cceb1ade01cfc6d371ded05171ee9760515202be7a47bd4d7.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 134, + 208, + 478, + 445 + ], + "blocks": [ + { + "bbox": [ + 134, + 208, + 478, + 445 + ], + "lines": [ + { + "bbox": [ + 134, + 208, + 478, + 445 + ], + "spans": [ + { + "bbox": [ + 134, + 208, + 478, + 445 + ], + "type": "image", + "image_path": "eda3e19a9df0d153ec03f03ce255213cb4e3de1751b555409800b287014933c5.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 182, + 454, + 430, + 466 + ], + "lines": [ + { + "bbox": [ + 182, + 454, + 430, + 466 + ], + "spans": [ + { + "bbox": [ + 182, + 454, + 430, + 466 + ], + "type": "text", + "content": "Fig. 5: Qualitative results on Synthetic4Relight dataset [56]." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 487, + 482, + 547 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 487, + 482, + 547 + ], + "spans": [ + { + "bbox": [ + 130, + 487, + 482, + 547 + ], + "type": "text", + "content": "R3DG demonstrates superiority over most of them [1,38,54]. While our performance slightly marginally trails that of vanilla 3DGS [26], it is noteworthy that our approach excels in relighting capability, presenting a significant advantage. Furthermore, in comparison to other relightable methods [33,49,51], our method showcases significantly better NVS quality." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 548, + 482, + 608 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 548, + 482, + 608 + ], + "spans": [ + { + "bbox": [ + 130, + 548, + 482, + 608 + ], + "type": "text", + "content": "In Fig. 4, we visualize the our reconstruction results in NeRF synthetic dataset, including the PBR image, the normal map, and the pre-computed visibility illustrated in the form of the Ambient Occlusion (AO) map. As depicted in Fig. 4, our approach successfully achieves realistic PBR rendering, smooth normal estimation, and accurate visibility solving on discrete point clouds." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 623, + 296, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 623, + 296, + 635 + ], + "spans": [ + { + "bbox": [ + 132, + 623, + 296, + 635 + ], + "type": "text", + "content": "5.3 Performance on Relighting" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "content": "To comprehensively assess the relighting capabilities of our pipeline, we further undertook experiments using the Synthetic4Relight dataset [56]. Recognizing" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "text", + "content": "J. Gao et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 136, + 114, + 478, + 295 + ], + "blocks": [ + { + "bbox": [ + 136, + 114, + 478, + 295 + ], + "lines": [ + { + "bbox": [ + 136, + 114, + 478, + 295 + ], + "spans": [ + { + "bbox": [ + 136, + 114, + 478, + 295 + ], + "type": "image", + "image_path": "fec47af15235fbfd3b6c09aae284aa837196eee7701c35dff80bb8a743ccc7dd.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 183, + 305, + 429, + 316 + ], + "lines": [ + { + "bbox": [ + 183, + 305, + 429, + 316 + ], + "spans": [ + { + "bbox": [ + 183, + 305, + 429, + 316 + ], + "type": "text", + "content": "Fig. 6: Qualitative results of relighting on real-world scenes." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 340, + 479, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 340, + 479, + 376 + ], + "spans": [ + { + "bbox": [ + 130, + 340, + 479, + 376 + ], + "type": "text", + "content": "the inherent scale ambiguity between the estimated albedo and lighting, we standardized the scale of the estimated albedo against the ground truth to relighting, consistent with previous studies [51, 56]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 376, + 480, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 376, + 480, + 460 + ], + "spans": [ + { + "bbox": [ + 130, + 376, + 480, + 460 + ], + "type": "text", + "content": "Our evaluation encompasses the analysis of decomposed materials, the synthesis of novel views, and the relighting outcomes. The quantitative analysis is presented in Tab. 2. Our method outperforms existing approaches in terms of Novel View Synthesis (NVS) and relighting precision. Regarding material estimation, our method exhibits superior albedo accuracy in SSIM and LPIPS. Qualitatively, our method achieves visually pleasing material decomposition, facilitating a realistic relighting effect, which is shown in 5." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 460, + 480, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 460, + 480, + 496 + ], + "spans": [ + { + "bbox": [ + 130, + 460, + 480, + 496 + ], + "type": "text", + "content": "Additionally, we conduct relighting on real-world scenes in the Mip-NeRF 360 dataset [4], as shown in Fig. 6. It is worth noting that our relighting in the kitchen with the second environment map, exhibits pleasing shadow effects." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 514, + 237, + 526 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 514, + 237, + 526 + ], + "spans": [ + { + "bbox": [ + 132, + 514, + 237, + 526 + ], + "type": "text", + "content": "5.4 Ablation Study" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": "We conduct ablation studies on three principal components of our method that significantly influence the quality of inverse rendering. The results are illustrated in Fig. 7. Initially, we investigate the impact of normal gradient densification (NGD). Fig. 7 (a, b) depict the rendering normal maps with and without the densification. It shows that the proposed densification markedly enhances the details of the normal map for *Lego*. In Fig. 7 (c, d, e), we examine the effects of our full light modeling. We simplify the full model to a global environment map in the optimization. The result demonstrates that the simplified model fails to generate shadow effects when relighting. Furthermore, we evaluate the significance of the proposed constraint on depth distribution, denoted as " + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_u" + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": ". The depth uncertainty is visualized in Fig. 7 (h, i), which illustrates the distribution" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 342, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 342, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 342, + 91, + 447, + 101 + ], + "type": "text", + "content": "Relightable 3D Gaussians" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 118, + 196, + 181 + ], + "blocks": [ + { + "bbox": [ + 133, + 118, + 196, + 181 + ], + "lines": [ + { + "bbox": [ + 133, + 118, + 196, + 181 + ], + "spans": [ + { + "bbox": [ + 133, + 118, + 196, + 181 + ], + "type": "image", + "image_path": "5b3b21ad98f3987eb60bbdce21c626f4daf0bf4745a4fabeed8200109a7888a4.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 145, + 182, + 188, + 191 + ], + "lines": [ + { + "bbox": [ + 145, + 182, + 188, + 191 + ], + "spans": [ + { + "bbox": [ + 145, + 182, + 188, + 191 + ], + "type": "text", + "content": "(a) w NGD." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 202, + 118, + 264, + 181 + ], + "blocks": [ + { + "bbox": [ + 202, + 118, + 264, + 181 + ], + "lines": [ + { + "bbox": [ + 202, + 118, + 264, + 181 + ], + "spans": [ + { + "bbox": [ + 202, + 118, + 264, + 181 + ], + "type": "image", + "image_path": "6523353d9050e62fb8dc52cc480729a976107cbc8ccc0568fd9988778d08578a.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 209, + 182, + 261, + 191 + ], + "lines": [ + { + "bbox": [ + 209, + 182, + 261, + 191 + ], + "spans": [ + { + "bbox": [ + 209, + 182, + 261, + 191 + ], + "type": "text", + "content": "(b) w/o NGD." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 271, + 114, + 338, + 182 + ], + "blocks": [ + { + "bbox": [ + 271, + 114, + 338, + 182 + ], + "lines": [ + { + "bbox": [ + 271, + 114, + 338, + 182 + ], + "spans": [ + { + "bbox": [ + 271, + 114, + 338, + 182 + ], + "type": "image", + "image_path": "5758f5d0b48f1ef7bf0d2582aadb03b628d0e685558ec0a66b0068add60c40f0.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 279, + 182, + 329, + 191 + ], + "lines": [ + { + "bbox": [ + 279, + 182, + 329, + 191 + ], + "spans": [ + { + "bbox": [ + 279, + 182, + 329, + 191 + ], + "type": "text", + "content": "(c) Relit. GT." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 339, + 114, + 406, + 182 + ], + "blocks": [ + { + "bbox": [ + 339, + 114, + 406, + 182 + ], + "lines": [ + { + "bbox": [ + 339, + 114, + 406, + 182 + ], + "spans": [ + { + "bbox": [ + 339, + 114, + 406, + 182 + ], + "type": "image", + "image_path": "96deec950c0a1ce91651e02ca8ca1e20ae6b11f6c7e4cc0373e465b782a38585.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 348, + 182, + 397, + 191 + ], + "lines": [ + { + "bbox": [ + 348, + 182, + 397, + 191 + ], + "spans": [ + { + "bbox": [ + 348, + 182, + 397, + 191 + ], + "type": "text", + "content": "(d) Full light." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 408, + 114, + 476, + 182 + ], + "blocks": [ + { + "bbox": [ + 408, + 114, + 476, + 182 + ], + "lines": [ + { + "bbox": [ + 408, + 114, + 476, + 182 + ], + "spans": [ + { + "bbox": [ + 408, + 114, + 476, + 182 + ], + "type": "image", + "image_path": "ff8d7021dea44ee962f1e8c41b78f3df0d10a159cf53e5904436ca8b072ecbd7.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 416, + 182, + 468, + 191 + ], + "lines": [ + { + "bbox": [ + 416, + 182, + 468, + 191 + ], + "spans": [ + { + "bbox": [ + 416, + 182, + 468, + 191 + ], + "type": "text", + "content": "(e) Simplified." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 136, + 194, + 214, + 227 + ], + "blocks": [ + { + "bbox": [ + 136, + 194, + 214, + 227 + ], + "lines": [ + { + "bbox": [ + 136, + 194, + 214, + 227 + ], + "spans": [ + { + "bbox": [ + 136, + 194, + 214, + 227 + ], + "type": "image", + "image_path": "690f3d3716c469c06a0ee3cde99eb678c9216a4f1d55251ed51d3d4f0e96bec3.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 150, + 231, + 200, + 241 + ], + "lines": [ + { + "bbox": [ + 150, + 231, + 200, + 241 + ], + "spans": [ + { + "bbox": [ + 150, + 231, + 200, + 241 + ], + "type": "text", + "content": "(f) AO w " + }, + { + "bbox": [ + 150, + 231, + 200, + 241 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_u" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 221, + 194, + 301, + 228 + ], + "blocks": [ + { + "bbox": [ + 221, + 194, + 301, + 228 + ], + "lines": [ + { + "bbox": [ + 221, + 194, + 301, + 228 + ], + "spans": [ + { + "bbox": [ + 221, + 194, + 301, + 228 + ], + "type": "image", + "image_path": "71cbe71748db2e43c57dbc9a9120d19c0f11c1cbff4d6c0a56c63cf57c92e180.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 233, + 231, + 290, + 241 + ], + "lines": [ + { + "bbox": [ + 233, + 231, + 290, + 241 + ], + "spans": [ + { + "bbox": [ + 233, + 231, + 290, + 241 + ], + "type": "text", + "content": "(g) AO w/o " + }, + { + "bbox": [ + 233, + 231, + 290, + 241 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_u" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 305, + 192, + 389, + 230 + ], + "blocks": [ + { + "bbox": [ + 305, + 192, + 389, + 230 + ], + "lines": [ + { + "bbox": [ + 305, + 192, + 389, + 230 + ], + "spans": [ + { + "bbox": [ + 305, + 192, + 389, + 230 + ], + "type": "image", + "image_path": "2b3b550b97d69ba29525029c7a3c2296b1502eab4d62adcb05b70376006e012f.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 307, + 232, + 388, + 241 + ], + "lines": [ + { + "bbox": [ + 307, + 232, + 388, + 241 + ], + "spans": [ + { + "bbox": [ + 307, + 232, + 388, + 241 + ], + "type": "text", + "content": "(h) Uncertainty w " + }, + { + "bbox": [ + 307, + 232, + 388, + 241 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_u" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 131, + 249, + 482, + 316 + ], + "lines": [ + { + "bbox": [ + 131, + 249, + 482, + 316 + ], + "spans": [ + { + "bbox": [ + 131, + 249, + 482, + 316 + ], + "type": "text", + "content": "Fig. 7: Ablation studies on main components of our method. (a) and (b) show normal maps with and without the proposed normal gradient densification (NGD). (d) and (e) display re-lighting results using the proposed full lighting model and only a simplified light modeling, i.e. a global environment map. (f) and (g) visualize ambient occlusion (AO) maps with and without the proposed constraint on depth distribution " + }, + { + "bbox": [ + 131, + 249, + 482, + 316 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_u" + }, + { + "bbox": [ + 131, + 249, + 482, + 316 + ], + "type": "text", + "content": ". (h) and (i) illustrate the depth uncertainty with and without " + }, + { + "bbox": [ + 131, + 249, + 482, + 316 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_u" + }, + { + "bbox": [ + 131, + 249, + 482, + 316 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 391, + 192, + 476, + 230 + ], + "blocks": [ + { + "bbox": [ + 391, + 192, + 476, + 230 + ], + "lines": [ + { + "bbox": [ + 391, + 192, + 476, + 230 + ], + "spans": [ + { + "bbox": [ + 391, + 192, + 476, + 230 + ], + "type": "image", + "image_path": "fa38bf3880ce90a9172ea6a6553efbb351ceff1352f95bee43ca8e6e386d8890.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 392, + 232, + 476, + 241 + ], + "lines": [ + { + "bbox": [ + 392, + 232, + 476, + 241 + ], + "spans": [ + { + "bbox": [ + 392, + 232, + 476, + 241 + ], + "type": "text", + "content": "(i) Uncertainty w/o " + }, + { + "bbox": [ + 392, + 232, + 476, + 241 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_u" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "bbox": [ + 131, + 344, + 480, + 381 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 344, + 480, + 381 + ], + "spans": [ + { + "bbox": [ + 131, + 344, + 480, + 381 + ], + "type": "text", + "content": "constraint contribute significantly to reducing the depth uncertainty. Additionally, Fig. 7 (f, g) reveals that incorporating " + }, + { + "bbox": [ + 131, + 344, + 480, + 381 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_u" + }, + { + "bbox": [ + 131, + 344, + 480, + 381 + ], + "type": "text", + "content": " leads to more plausible ambient occlusion (averaged visibility) which plays a crucial role in stage 2." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 132, + 411, + 219, + 423 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 411, + 219, + 423 + ], + "spans": [ + { + "bbox": [ + 132, + 411, + 219, + 423 + ], + "type": "text", + "content": "6 Conclusion" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 130, + 447, + 482, + 591 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 447, + 482, + 591 + ], + "spans": [ + { + "bbox": [ + 130, + 447, + 482, + 591 + ], + "type": "text", + "content": "In this paper, we have introduced a differentiable point-based rendering pipeline enabling effort less editing, accuracy ray tracing, and photo-realistic relighting. We present the scene as Relightable 3D Gaussians, an extension of traditional 3D Gaussians enriched with supplementary normals, BRDFs and indirect lighting. To reconstruct a relightable scene from multi-view images, we build a novel differentiable rendering pipeline based on the inverse rendering techniques and 3D Gaussian Splatting. Additionally, we introduce an innovative ray tracing scheme specifically designed for scenes represented by discrete points, providing precise visibility for realistic relighting. Quantitative and qualitative experiments confirm that our pipeline successfully reconstructs reasonable normals, materials for discrete point clouds, and achieves commendable accuracy in both novel view synthesis and scene relighting tasks." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": "Limitations and Future work. Our present pipeline targets the reconstruction of static objects. Certain design choices pose challenges in maintaining its performance with large scale scenes. The high density of points in large scene can slow down optimization as we sample rays and perform PBR at each point. This can be mitigated through deferred rendering technique. Regarding geometry, integrating MVS shows promise for achieving more accurate representations." + } + ] + } + ], + "index": 24 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "text", + "content": "J. Gao et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 137, + 505, + 162 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 137, + 505, + 162 + ], + "spans": [ + { + "bbox": [ + 132, + 137, + 505, + 162 + ], + "type": "text", + "content": "This work was supported by National Key R&D Program of China (2023YFB3209702), and NSFC (62441204)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 178, + 197, + 190 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 178, + 197, + 190 + ], + "spans": [ + { + "bbox": [ + 133, + 178, + 197, + 190 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 201, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 23, + "blocks": [ + { + "bbox": [ + 138, + 201, + 481, + 223 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 201, + 481, + 223 + ], + "spans": [ + { + "bbox": [ + 138, + 201, + 481, + 223 + ], + "type": "text", + "content": "1. Aliev, K.A., Sevastopolsky, A., Kolos, M., Ulyanov, D., Lempitsky, V.: Neural point-based graphics. In: ECCV (2020)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 224, + 481, + 244 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 224, + 481, + 244 + ], + "spans": [ + { + "bbox": [ + 138, + 224, + 481, + 244 + ], + "type": "text", + "content": "2. Azinovic, D., Li, T.M., Kaplanyan, A., Nießner, M.: Inverse path tracing for joint material and lighting estimation. In: CVPR (2019)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 245, + 481, + 277 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 245, + 481, + 277 + ], + "spans": [ + { + "bbox": [ + 138, + 245, + 481, + 277 + ], + "type": "text", + "content": "3. Barron, J.T., Mildenhall, B., Tancik, M., Hedman, P., Martin-Brualla, R., Srinivasan, P.P.: Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In: ICCV (2021)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 277, + 481, + 298 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 277, + 481, + 298 + ], + "spans": [ + { + "bbox": [ + 138, + 277, + 481, + 298 + ], + "type": "text", + "content": "4. Barron, J.T., Mildenhall, B., Verbin, D., Srinivasan, P.P., Hedman, P.: Mip-nerf 360: Unbounded anti-aliased neural radiance fields. In: CVPR (2022)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 299, + 481, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 299, + 481, + 319 + ], + "spans": [ + { + "bbox": [ + 138, + 299, + 481, + 319 + ], + "type": "text", + "content": "5. Barron, J.T., Mildenhall, B., Verbin, D., Srinivasan, P.P., Hedman, P.: Zip-nerf: Anti-aliased grid-based neural radiance fields. arXiv preprint (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 320, + 481, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 320, + 481, + 352 + ], + "spans": [ + { + "bbox": [ + 138, + 320, + 481, + 352 + ], + "type": "text", + "content": "6. Bi, S., Xu, Z., Srinivasan, P., Mildenhall, B., Sunkavalli, K., Hašan, M., Hold-Geoffroy, Y., Kriegman, D., Ramamoorthi, R.: Neural reflectance fields for appearance acquisition. arXiv preprint (2020)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 353, + 481, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 353, + 481, + 385 + ], + "spans": [ + { + "bbox": [ + 138, + 353, + 481, + 385 + ], + "type": "text", + "content": "7. Bi, S., Xu, Z., Sunkavalli, K., Hašan, M., Hold-Geoffroy, Y., Kriegman, D., Ramamoorthi, R.: Deep reflectance volumes: Relightable reconstructions from multiview photometric images. In: ECCV (2020)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 386, + 481, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 386, + 481, + 407 + ], + "spans": [ + { + "bbox": [ + 138, + 386, + 481, + 407 + ], + "type": "text", + "content": "8. Bi, S., Xu, Z., Sunkavalli, K., Kriegman, D., Ramamoorthi, R.: Deep 3d capture: Geometry and reflectance from sparse multi-view images. In: CVPR (2020)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 407, + 481, + 427 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 407, + 481, + 427 + ], + "spans": [ + { + "bbox": [ + 138, + 407, + 481, + 427 + ], + "type": "text", + "content": "9. Boss, M., Braun, R., Jampani, V., Barron, J.T., Liu, C., Lensch, H.: Nerd: Neural reflectance decomposition from image collections. In: ICCV (2021)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 428, + 481, + 460 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 428, + 481, + 460 + ], + "spans": [ + { + "bbox": [ + 138, + 428, + 481, + 460 + ], + "type": "text", + "content": "0. Boss, M., Engelhardt, A., Kar, A., Li, Y., Sun, D., Barron, J., Lensch, H., Jampani, V.: Samurai: Shape and material from unconstrained real-world arbitrary image collections. In: NeurIPS (2022)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 460, + 481, + 482 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 460, + 481, + 482 + ], + "spans": [ + { + "bbox": [ + 138, + 460, + 481, + 482 + ], + "type": "text", + "content": "1. Burley, B., Studios, W.D.A.: Physically-based shading at disney. In: SIGGRAPH (2012)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 483, + 481, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 483, + 481, + 503 + ], + "spans": [ + { + "bbox": [ + 138, + 483, + 481, + 503 + ], + "type": "text", + "content": "2. Chen, A., Xu, Z., Geiger, A., Yu, J., Su, H.: Tensorf: Tensorial radiance fields. In: ECCV (2022)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 504, + 481, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 504, + 481, + 536 + ], + "spans": [ + { + "bbox": [ + 138, + 504, + 481, + 536 + ], + "type": "text", + "content": "3. Chen, Z., Funkhouser, T., Hedman, P., Tagliasacchi, A.: Mobilenerf: Exploiting the polygon rasterization pipeline for efficient neural field rendering on mobile architectures. In: CVPR (2023)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 536, + 481, + 556 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 536, + 481, + 556 + ], + "spans": [ + { + "bbox": [ + 138, + 536, + 481, + 556 + ], + "type": "text", + "content": "4. Deng, K., Liu, A., Zhu, J.Y., Ramanan, D.: Depth-supervised nef: Fewer views and faster training for free. In: CVPR (2022)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 138, + 557, + 481, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 557, + 481, + 578 + ], + "spans": [ + { + "bbox": [ + 138, + 557, + 481, + 578 + ], + "type": "text", + "content": "5. Fridovich-Keil, S., Yu, A., Tancik, M., Chen, Q., Recht, B., Kanazawa, A.: Plenoxels: Radiance fields without neural networks. In: CVPR (2022)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 138, + 579, + 481, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 579, + 481, + 600 + ], + "spans": [ + { + "bbox": [ + 138, + 579, + 481, + 600 + ], + "type": "text", + "content": "6. Garbin, S.J., Kowalski, M., Johnson, M., Shotton, J., Valentin, J.: Fastnerf: High-fidelity neural rendering at 200fps. In: ICCV (2021)" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 138, + 601, + 481, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 601, + 481, + 632 + ], + "spans": [ + { + "bbox": [ + 138, + 601, + 481, + 632 + ], + "type": "text", + "content": "7. Guédon, A., Lepetit, V.: Sugar: Surface-aligned gaussian splatting for efficient 3d mesh reconstruction and high-quality mesh rendering. arXiv preprint arXiv:2311.12775 (2023)" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 138, + 633, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 633, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 633, + 481, + 665 + ], + "type": "text", + "content": "8. Guo, K., Lincoln, P., Davidson, P., Busch, J., Yu, X., Whalen, M., Harvey, G., Orts-Escalano, S., Pandey, R., Dourgarian, J., et al.: The relighttables: Volumetric performance capture of humans with realistic relighting. ACM TOG (2019)" + } + ] + } + ], + "index": 22 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 342, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 342, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 342, + 91, + 447, + 102 + ], + "type": "text", + "content": "Relightable 3D Gaussians" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 117, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 24, + "blocks": [ + { + "bbox": [ + 133, + 117, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 117, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 133, + 117, + 480, + 138 + ], + "type": "text", + "content": "19. Hasselgren, J., Hofmann, N., Munkberg, J.: Shape, light, and material decomposition from images using monte carlo rendering and denoising. In: NeurIPS (2022)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 140, + 480, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 140, + 480, + 172 + ], + "spans": [ + { + "bbox": [ + 133, + 140, + 480, + 172 + ], + "type": "text", + "content": "20. Hasselgren, J., Hofmann, N., Munkberg, J.: Shape, Light, and Material Decomposition from Images using Monte Carlo Rendering and Denoising. arXiv:2206.03380 (2022)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 174, + 480, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 174, + 480, + 194 + ], + "spans": [ + { + "bbox": [ + 133, + 174, + 480, + 194 + ], + "type": "text", + "content": "21. Hedman, P., Srinivasan, P.P., Mildenhall, B., Barron, J.T., Debevec, P.: Baking neural radiance fields for real-time view synthesis. In: ICCV (2021)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 196, + 480, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 196, + 480, + 217 + ], + "spans": [ + { + "bbox": [ + 132, + 196, + 480, + 217 + ], + "type": "text", + "content": "22. Hu, T., Liu, S., Chen, Y., Shen, T., Jia, J.: Efficientnerf efficient neural radiance fields. In: CVPR (2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 218, + 480, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 218, + 480, + 239 + ], + "spans": [ + { + "bbox": [ + 132, + 218, + 480, + 239 + ], + "type": "text", + "content": "23. Jin, H., Liu, I., Xu, P., Zhang, X., Han, S., Bi, S., Zhou, X., Xu, Z., Su, H.: Tensoir: Tensorial inverse rendering. In: CVPR (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 241, + 480, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 241, + 480, + 262 + ], + "spans": [ + { + "bbox": [ + 132, + 241, + 480, + 262 + ], + "type": "text", + "content": "24. Kajiya, J.T.: The rendering equation. In: Proceedings of the 13th annual conference on Computer graphics and interactive techniques (1986)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 263, + 480, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 263, + 480, + 296 + ], + "spans": [ + { + "bbox": [ + 132, + 263, + 480, + 296 + ], + "type": "text", + "content": "25. Karras, T.: Maximizing parallelism in the construction of bvhs, octrees, and k-d trees. In: Proceedings of the Fourth ACM SIGGRAPH/Eurographics conference on High-Performance Graphics (2012)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 297, + 480, + 318 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 297, + 480, + 318 + ], + "spans": [ + { + "bbox": [ + 132, + 297, + 480, + 318 + ], + "type": "text", + "content": "26. Kerbl, B., Kopanas, G., Leimkuhler, T., Drettakis, G.: 3d gaussian splatting for real-time radiance field rendering. ACM TOG (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 319, + 480, + 340 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 319, + 480, + 340 + ], + "spans": [ + { + "bbox": [ + 132, + 319, + 480, + 340 + ], + "type": "text", + "content": "27. Keselman, L., Hebert, M.: Flexible techniques for differentiable rendering with 3d gaussians. arXiv preprint (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 342, + 480, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 342, + 480, + 363 + ], + "spans": [ + { + "bbox": [ + 132, + 342, + 480, + 363 + ], + "type": "text", + "content": "28. Kuang, Z., Olszewski, K., Chai, M., Huang, Z., Achlioptas, P., Tulyakov, S.: Neroic: Neural rendering of objects from online image collections. ACM TOG (2022)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 365, + 438, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 365, + 438, + 374 + ], + "spans": [ + { + "bbox": [ + 132, + 365, + 438, + 374 + ], + "type": "text", + "content": "29. Levoy, M., Whitted, T.: The use of points as a display primitive (1985)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 133, + 376, + 480, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 376, + 480, + 407 + ], + "spans": [ + { + "bbox": [ + 133, + 376, + 480, + 407 + ], + "type": "text", + "content": "30. Liu, Y., Wang, P., Lin, C., Long, X., Wang, J., Liu, L., Komura, T., Wang, W.: Nano: Neural geometry and brdf reconstruction of reflective objects from multiview images. In: SIGGRAPH (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 133, + 409, + 480, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 409, + 480, + 441 + ], + "spans": [ + { + "bbox": [ + 133, + 409, + 480, + 441 + ], + "type": "text", + "content": "31. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. In: ECCV (2020)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 442, + 480, + 463 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 442, + 480, + 463 + ], + "spans": [ + { + "bbox": [ + 132, + 442, + 480, + 463 + ], + "type": "text", + "content": "32. Müller, T., Evans, A., Schied, C., Keller, A.: Instant neural graphics primitives with a multiresolution hash encoding. ACM TOG (2022)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 133, + 464, + 480, + 496 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 464, + 480, + 496 + ], + "spans": [ + { + "bbox": [ + 133, + 464, + 480, + 496 + ], + "type": "text", + "content": "33. Munkberg, J., Hasselgren, J., Shen, T., Gao, J., Chen, W., Evans, A., Müller, T., Fidler, S.: Extracting triangular 3d models, materials, and lighting from images. In: CVPR (2022)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 498, + 480, + 519 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 498, + 480, + 519 + ], + "spans": [ + { + "bbox": [ + 132, + 498, + 480, + 519 + ], + "type": "text", + "content": "34. Nam, G., Lee, J.H., Gutierrez, D., Kim, M.H.: Practical svbrdf acquisition of 3d objects with unstructured flash photography. ACM TOG (2018)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 521, + 480, + 541 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 521, + 480, + 541 + ], + "spans": [ + { + "bbox": [ + 132, + 521, + 480, + 541 + ], + "type": "text", + "content": "35. Park, J.J., Holynski, A., Seitz, S.M.: Seeing the world in a bag of chips. In: CVPR (2020)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 133, + 543, + 480, + 575 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 543, + 480, + 575 + ], + "spans": [ + { + "bbox": [ + 133, + 543, + 480, + 575 + ], + "type": "text", + "content": "36. Pfister, H., Zwicker, M., Van Baar, J., Gross, M.: Surfels: Surface elements as rendering primitives. In: Proceedings of the 27th annual conference on Computer graphics and interactive techniques (2000)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 132, + 576, + 480, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 576, + 480, + 597 + ], + "spans": [ + { + "bbox": [ + 132, + 576, + 480, + 597 + ], + "type": "text", + "content": "37. Pittaluga, F., Koppal, S.J., Kang, S.B., Sinha, S.N.: Revealing scenes by inverting structure from motion reconstructions. In: CVPR (2019)" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 132, + 598, + 480, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 598, + 480, + 620 + ], + "spans": [ + { + "bbox": [ + 132, + 598, + 480, + 620 + ], + "type": "text", + "content": "38. Rakhimov, R., Ardelean, A.T., Lempitsky, V., Burnaev, E.: Npbg++: Accelerating neural point-based graphics. In: CVPR (2022)" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 132, + 621, + 480, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 621, + 480, + 643 + ], + "spans": [ + { + "bbox": [ + 132, + 621, + 480, + 643 + ], + "type": "text", + "content": "39. Reiser, C., Peng, S., Liao, Y., Geiger, A.: Kilonerf: Speeding up neural radiance fields with thousands of tiny mlps. In: ICCV (2021)" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 132, + 643, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 643, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 643, + 480, + 665 + ], + "type": "text", + "content": "40. Schmitt, C., Donne, S., Riegler, G., Koltun, V., Geiger, A.: On joint estimation of pose, geometry and svbrdf from a handheld scanner. In: CVPR (2020)" + } + ] + } + ], + "index": 23 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "text", + "content": "J. Gao et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 599 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "text", + "content": "41. Srinivasan, P.P., Deng, B., Zhang, X., Tancik, M., Mildenhall, B., Barron, J.T.: Nerv: Neural reflectance and visibility fields for relighting and view synthesis. In: CVPR (2021)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 150, + 482, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 150, + 482, + 182 + ], + "spans": [ + { + "bbox": [ + 130, + 150, + 482, + 182 + ], + "type": "text", + "content": "42. Verbin, D., Hedman, P., Mildenhall, B., Zickler, T., Barron, J.T., Srinivasan, P.P.: Ref-nerf: Structured view-dependent appearance for neural radiance fields. In: CVPR (2022)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 182, + 482, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 182, + 482, + 215 + ], + "spans": [ + { + "bbox": [ + 130, + 182, + 482, + 215 + ], + "type": "text", + "content": "43. Wang, P., Liu, L., Liu, Y., Theobalt, C., Komura, T., Wang, W.: Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. arXiv preprint (2021)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 216, + 482, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 216, + 482, + 237 + ], + "spans": [ + { + "bbox": [ + 130, + 216, + 482, + 237 + ], + "type": "text", + "content": "44. Wu, H., Hu, Z., Li, L., Zhang, Y., Fan, C., Yu, X.: Nefii: Inverse rendering for reflectance decomposition with near-field indirect illumination. In: CVPR (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 237, + 482, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 237, + 482, + 270 + ], + "spans": [ + { + "bbox": [ + 130, + 237, + 482, + 270 + ], + "type": "text", + "content": "45. Yao, Y., Zhang, J., Liu, J., Qu, Y., Fang, T., McKinnon, D., Tsin, Y., Quan, L.: Neilf: Neural incident light field for physically-based material estimation. In: ECCV (2022)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 270, + 482, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 270, + 482, + 293 + ], + "spans": [ + { + "bbox": [ + 130, + 270, + 482, + 293 + ], + "type": "text", + "content": "46. Yariv, L., Gu, J., Kasten, Y., Lipman, Y.: Volume rendering of neural implicit surfaces. In: NeurIPS (2021)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 293, + 482, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 293, + 482, + 315 + ], + "spans": [ + { + "bbox": [ + 130, + 293, + 482, + 315 + ], + "type": "text", + "content": "47. Yifan, W., Serena, F., Wu, S., Öztireli, C., Sorkine-Hornung, O.: Differentiable surface splatting for point-based geometry processing. ACM TOG (2019)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 315, + 482, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 315, + 482, + 335 + ], + "spans": [ + { + "bbox": [ + 130, + 315, + 482, + 335 + ], + "type": "text", + "content": "48. Yu, Z., Peng, S., Niemeyer, M., Sattler, T., Geiger, A.: Monosdf: Exploring monocular geometric cues for neural implicit surface reconstruction. In: NeurIPS (2022)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 335, + 482, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 335, + 482, + 369 + ], + "spans": [ + { + "bbox": [ + 130, + 335, + 482, + 369 + ], + "type": "text", + "content": "49. Zhang, J., Yao, Y., Li, S., Liu, J., Fang, T., McKinnon, D., Tsin, Y., Quan, L.: Neil++: Inter-reflectable light fields for geometry and material estimation. In: ICCV (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 369, + 482, + 391 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 369, + 482, + 391 + ], + "spans": [ + { + "bbox": [ + 130, + 369, + 482, + 391 + ], + "type": "text", + "content": "50. Zhang, K., Luan, F., Li, Z., Snavely, N.: Iron: Inverse rendering by optimizing neural sdfs and materials from photometric images. In: CVPR (2022)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 391, + 482, + 423 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 391, + 482, + 423 + ], + "spans": [ + { + "bbox": [ + 130, + 391, + 482, + 423 + ], + "type": "text", + "content": "51. Zhang, K., Luan, F., Wang, Q., Bala, K., Snavely, N.: Physg: Inverse rendering with spherical gaussians for physics-based material editing and relighting. In: CVPR (2021)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 423, + 482, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 423, + 482, + 445 + ], + "spans": [ + { + "bbox": [ + 130, + 423, + 482, + 445 + ], + "type": "text", + "content": "52. Zhang, Q., Baek, S.H., Rusinkiewicz, S., Heide, F.: Differentiable point-based radiance fields for efficient view synthesis. In: SIGGRAPH Asia (2022)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 446, + 482, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 446, + 482, + 479 + ], + "spans": [ + { + "bbox": [ + 130, + 446, + 482, + 479 + ], + "type": "text", + "content": "53. Zhang, X., Srinivasan, P.P., Deng, B., Debevec, P., Freeman, W.T., Barron, J.T.: Nerfactor: Neural factorization of shape and reflectance under an unknown illumination. ACM TOG (2021)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 479, + 482, + 501 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 479, + 482, + 501 + ], + "spans": [ + { + "bbox": [ + 130, + 479, + 482, + 501 + ], + "type": "text", + "content": "54. Zhang, Y., Huang, X., Ni, B., Li, T., Zhang, W.: Frequency-modulated point cloud rendering with easy editing. In: CVPR (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 501, + 482, + 522 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 501, + 482, + 522 + ], + "spans": [ + { + "bbox": [ + 130, + 501, + 482, + 522 + ], + "type": "text", + "content": "55. Zhang, Y., Xu, T., Yu, J., Ye, Y., Jing, Y., Wang, J., Yu, J., Yang, W.: Nemf: Inverse volume rendering with neural microflake field. In: ICCV (2023)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 130, + 522, + 482, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 522, + 482, + 544 + ], + "spans": [ + { + "bbox": [ + 130, + 522, + 482, + 544 + ], + "type": "text", + "content": "56. Zhang, Y., Sun, J., He, X., Fu, H., Jia, R., Zhou, X.: Modeling indirect illumination for inverse rendering. In: CVPR (2022)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 130, + 544, + 482, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 544, + 482, + 577 + ], + "spans": [ + { + "bbox": [ + 130, + 544, + 482, + 577 + ], + "type": "text", + "content": "57. Zwicker, M., Pfister, H., Van Baar, J., Gross, M.: Surface splatting. In: Proceedings of the 28th annual conference on computer graphics and interactive techniques (2001)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 130, + 577, + 482, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 577, + 482, + 599 + ], + "spans": [ + { + "bbox": [ + 130, + 577, + 482, + 599 + ], + "type": "text", + "content": "58. Zwicker, M., Pfister, H., Van Baar, J., Gross, M.: Ewa splatting. IEEE TVCG (2002)" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 342, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 342, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 342, + 91, + 447, + 101 + ], + "type": "text", + "content": "Relightable 3D Gaussians" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Relightable Neural Actor with Intrinsic Decomposition and Pose Control/b67bc924-ef1e-4a87-8446-f01e4ef31ad6_content_list.json b/2024/Relightable Neural Actor with Intrinsic Decomposition and Pose Control/b67bc924-ef1e-4a87-8446-f01e4ef31ad6_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..f23195d61934f45a32694350258c7fb73ea9f40a --- /dev/null +++ b/2024/Relightable Neural Actor with Intrinsic Decomposition and Pose Control/b67bc924-ef1e-4a87-8446-f01e4ef31ad6_content_list.json @@ -0,0 +1,1837 @@ +[ + { + "type": "text", + "text": "Relightable Neural Actor with Intrinsic Decomposition and Pose Control", + "text_level": 1, + "bbox": [ + 276, + 140, + 727, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Diogo Carbonera Luvizon $^{1,2}$ , Vladislav Golyanik $^{1}$ , Adam Kortylewski $^{1,3}$ , Marc Habermann $^{1,2}$ , and Christian Theobalt $^{1,2}$", + "bbox": [ + 222, + 210, + 779, + 243 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Max Planck Institute for Informatics, Saarland Informatics Campus \n $^{2}$ Saarbrücken Research Center for Visual Computing, Interaction and AI \n $^{3}$ University of Freiburg", + "bbox": [ + 254, + 253, + 746, + 296 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{dluvizon,golyanik,akortyle,mhaberma,theobalt}@mpi-inf.mpg.de https://vcai.mpi-inf.mpg.de/projects/RNA", + "bbox": [ + 254, + 296, + 746, + 325 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Creating a controllable and relightable digital avatar from multi-view video with fixed illumination is a very challenging problem since humans are highly articulated, creating pose-dependent appearance effects, and skin as well as clothing require space-varying BRDF modeling. Existing works on creating animatable avatars either do not focus on relighting at all, require controlled illumination setups, or try to recover a relightable avatar from very low cost setups, i.e. a single RGB video, at the cost of severely limited result quality, e.g. shadows not even being modeled. To address this, we propose Relightable Neural Actor, a new video-based method for learning a pose-driven neural human model that can be relighted, allows appearance editing, and models pose-dependent effects such as wrinkles and self-shadows. Importantly, for training, our method solely requires a multi-view recording of the human under a known, but static lighting condition. To tackle this challenging problem, we leverage an implicit geometry representation of the actor with a drivable density field that models pose-dependent deformations and derive a dynamic mapping between 3D and UV spaces, where normal, visibility, and materials are effectively encoded. To evaluate our approach in real-world scenarios, we collect a new dataset with four identities recorded under different light conditions, indoors and outdoors, providing the first benchmark of its kind for human relighting, and demonstrating state-of-the-art relighting results for novel human poses.", + "bbox": [ + 259, + 358, + 743, + 664 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 688, + 375, + 704 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Creating a relightable and controllable virtual actor of a real person solely from multi-view videos under static illumination is an exciting and challenging topic that has numerous applications in virtual and augmented reality, gaming, human-machine interaction, and content creation. However, it requires modeling the outgoing radiance of a particular surface point on the human, which depends on the pose-dependent and highly non-rigid surface deformation of the human body, the complex and spatially varying materials, as well as the scene lighting. Intrinsically decomposing these quantities solely from multi-view video and", + "bbox": [ + 212, + 718, + 787, + 840 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/1d796e40e690dcc337edceab4833f41ef7484c233c64e6670fc21aeb5b8c5b74.jpg", + "image_caption": [ + "Fig. 1: Our method learns a neural actor that is driven only by a 3D skeletal pose and allows rendering and editing of the actor's appearance under new lightning and poses not seen during training. Our approach models pose-dependent deformations, self-shadows, and performs intrinsic decomposition of dynamic humans through a pose-dependent mapping between 3D and UV spaces, which also enables editing the appearance and material properties at inference time. For training, our method only needs a multi-view video and an environment map." + ], + "image_footnote": [], + "bbox": [ + 215, + 142, + 480, + 275 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/2bfcf18ed57fbca0ab339fc63250a8c06953c6e79673d39f9e744e693239c609.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 491, + 142, + 679, + 268 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/b172c1837958adb9144e48e7917212a58ddb6ca8959a316cc51fb31ea2cefa16.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 679, + 142, + 785, + 268 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "enabling novel compositions that are of highly photorealistic quality defines an unsolved and challenging inverse problem in the literature.", + "bbox": [ + 212, + 414, + 784, + 444 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Relighting static scenes from multi-view imagery is a well-studied problem [22,33,49,62]. However, directly applying such methods to moving humans, which are dynamic, highly articulated, and have complex material composition, is not trivially possible, as these approaches do not account for dynamics. On the contrary, some works have focused on building drivable and dynamic full-body avatars from multi-view video [32,42,54] without considering relightability at all. Despite some works [4,11,18,36,66] assuming a well-calibrated lightstage settings, only few human-specific works [6,17,64] try to achieve human intrinsic decomposition and relighting from videos with static lights. While they are able to recover a relightable human representation from a single RGB video, they can solely replay the human motion under novel lighting conditions [6], or do not model self-shadows and space-varying BRDF [17], which results in low visual fidelity. In stark contrast, we assume a multi-view setting while accounting for high-frequency surface deformations as well as induced high-frequency lighting effects such as self-shadows and shading, which are indispensable for a truly immersive experience.", + "bbox": [ + 212, + 446, + 787, + 686 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address this, we propose Relightable Neural Actor, a new method that achieves intrinsic decomposition of dynamic human avatars, enabling pose-control, material editing, and realistic rendering with pose-dependent effects, while solely requiring a multi-view video of the real human under a single and static light condition during training (see Fig. 1). At inference time, we only need an arbitrary driving skeletal pose and an environment map.", + "bbox": [ + 212, + 688, + 790, + 777 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We solve this challenging problem by leveraging an implicit density field from a pose-driven neural actor and by proposing a novel intrinsic decomposition guided by the implicit density field. This strategy allows our method to model pose-dependent deformations in the geometry as dynamic maps, while", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "D.C. Luvizon et al.", + "bbox": [ + 271, + 114, + 400, + 126 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "representing the albedo and roughness properties of the actor as static maps. Our approach not only makes the problem computationally tractable for the long and dynamic human sequences, but also enables editing the neural actor's appearance and materials at inference time. Our key contribution is to sample normal and visibility from the posed density field and aggregate them into sparse UV maps, which are then densified by the proposed NormalNet and VisibilityNet models. This allows our approach to effectively model wrinkles and self-shadows even for long sequences. The albedo and roughness maps are optimized during training and sampled by the proposed UVDeltaNet, which accounts for pose-dependent deformations in the UV map and is intrinsically learned in our pipeline. Finally, given an environment map and a virtual camera position, our neural renderer produces the final image of the neural actor.", + "bbox": [ + 212, + 146, + 787, + 327 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To train and evaluate our method, we collected a real-world dataset for novel pose and novel lighting human synthesis, composed of four actors recorded under six different light conditions by a multi-view camera setup, including indoor and outdoor environments and a light probe for each sequence, providing the first real in-the-wild benchmark for controllable human relighting.", + "bbox": [ + 212, + 329, + 787, + 404 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To summarize, our main technical contributions are as follows: 1) The first method to learn a relightable neural actor from multi-view video under static lights that allows for pose control, appearance editing, and free-viewpoint rendering under arbitrary illumination while modeling pose-dependent effects. 2) We propose an intrinsic decomposition of diffuse and specular reflection based on a new approach to compute pose-dependent normals and visibility by sparsely sampling a neural field and densifying the results with NormalNet and VisibilityNet, making the problem of pose-driven human relighting tractable. Our approach relies on a mapping between 3D and UV space implicitly learned by UVDeltaNet, allowing us to learn an explicit and static material representation that is easy to edit at inference time.", + "bbox": [ + 212, + 405, + 787, + 569 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In our experimental evaluation, we compare our approach with two baselines in the proposed real data benchmark, showing the superiority of our approach in synthesizing relightable humans on challenging motion sequences. The effectiveness of our method is demonstrated by ablation studies and qualitative results.", + "bbox": [ + 212, + 571, + 787, + 633 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 214, + 657, + 388, + 674 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Since the goal of this work is to build a full-body relightable and animatable avatar, we will not discuss works [9, 10, 28, 44, 46, 47, 50, 56, 58, 59] focusing on face relighting as both settings have different requirements and challenges.", + "bbox": [ + 212, + 694, + 787, + 742 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.1 Relighting Static Scenes", + "text_level": 1, + "bbox": [ + 214, + 765, + 464, + 781 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Classic approaches for relighting and editing perform intrinsic decomposition through inverse rendering [43]. With recent advances in neural rendering [51], light transportation can be taken into account with fully differentiable rendering", + "bbox": [ + 212, + 794, + 787, + 842 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "Relightable Neural Actor with Intrinsic Decomposition and Pose Control", + "bbox": [ + 246, + 114, + 730, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "pipelines. For instance, Zhang et al. [62] address the problem of recovering geometry and spatially-varying BRDF with an implicit scene representation. However, this requires solving the light transport by evaluating the NeRF model for each light ray, which is very inefficient. NeRV [49] and Zhang et al. [63] mitigate this problem with a neural visibility field that is jointly optimized with the NeRF model, and NeRD [5] adds further relaxation by ignoring cast shadows. However, this assumption cannot be made for humans, which are highly articulated and often experience prominent self-casting shadows. Jin et al. [22] propose TensorIR, an efficient tensor-based representation that solves the inverse problem by casting camera and secondary rays. IRON [60] proposes a two-stage process for obtaining a mesh with material properties, and Lyu et al. [35] and NeRO [33] focus on multi-bounce illumination and reflective objects, which is usually not the case for humans. The methods mentioned above share the assumption of a static scene and cannot be trivially adapted to dynamic and controllable humans.", + "bbox": [ + 212, + 146, + 787, + 358 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.2 Human Models with Pose Control", + "text_level": 1, + "bbox": [ + 215, + 378, + 542, + 393 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Modeling humans is a long-standing topic. Traditionally, meshes from an actor can be automatically rigged to a skeleton [7] using Linear Blending Skinning (LBS) [27]. Parametric human models offer pose and shape control [34,41] and, more recently, implicit human bodies can be articulated [3, 37, 38]. Such models offer a generic and controllable representation, but are unable to represent individual details. Template-based methods synthesize a realistic level of details [14, 15, 21], but have to deal with non-linear pose-dependent deformations.", + "bbox": [ + 212, + 402, + 785, + 508 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Several approaches model appearance and geometry with an implicit field anchored on a coarse human mesh. Neural Body [42] optimizes latent vectors anchored in the vertices of the human mesh and Neural Actor [32] incorporates texture maps as a conditioning to break down the mapping from pose to dynamic effects (one-to-many mapping). ARAH [54] models the human surface with an SDF, Dual-Space NeRF [65] learns the color and light in separate spaces, and Kwon et al. [24] propose an approach that generalizes to new identities and poses. Shyshey et al. [48] jointly learn textures and a neural renderer but fall short in reconstructing the 3D human body. A few works [20,57] approach a restrictive scenario of reconstructing 3D humans from a monocular video but fail to model fine details under new poses. Common to all these type of methods [12,13,25, 40] is the entanglement of material properties and illumination, making them restricted to the same appearance as observed during training.", + "bbox": [ + 212, + 508, + 787, + 705 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.3 Relighting Human Models", + "text_level": 1, + "bbox": [ + 215, + 724, + 480, + 739 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Only a few methods in the literature tackle the problem of relighting humans from videos under static illumination. Li et al. [29] create a relightable human mesh by estimating a low-frequency environment using Spherical Harmonics. Some approaches target the problem of single-image [19] and static pose [64] relighting under unknown lights, which is an extremely ill-posed problem due to the high ambiguity between albedo and lighting. Relighting4D [6] extends", + "bbox": [ + 212, + 750, + 787, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "D.C. Luvizon et al.", + "bbox": [ + 271, + 114, + 400, + 127 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Neural Body [42] to estimate geometry and reflectance, but can only replay the same sequence observed during training under new lights and is not able to generate new poses. RANA [17] allows for pose control but assumes Lambertian reflectance, ignoring different material properties such as skin and clothes, and does not model cast shadows, resulting in unrealistic renderings for articulated humans. In contrast, our method performs intrinsic decomposition with space-varying BRDF, models illumination and self-shadows, allows for pose control and novel lights, and offers an easy manner to edit the material properties of the learned neural actor.", + "bbox": [ + 212, + 146, + 787, + 282 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3 Method", + "text_level": 1, + "bbox": [ + 214, + 305, + 330, + 321 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our goal is to learn a digital human avatar (neural actor) from RGB images that models pose-dependent effects and can be relighted, pose-controlled, and rendered from novel viewpoints at test time. To achieve relighting, we decompose the observed appearance into normals, visibility with respect to each light source, and BRDF represented as albedo and roughness. With these components and an environment map, we employ the microfacet model [52] to render the output image of our method, which is compared to the observed images. In addition—since we target a drivable actor—the geometry components need to be controllable by the skeletal pose.", + "bbox": [ + 212, + 338, + 787, + 474 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Fig. 2 shows an overview of our method divided into three main components. The first is a pose-driven implicit geometry model, where the goal is to obtain a skeleton-controllable implicit 3D density field that models the geometry of the actor with pose-dependent deformations. This model is adapted from Neural Actor [32] and is depicted here since our approach builds upon its internal design. Note that Neural Actor does not perform intrinsic decomposition and is unable to perform relighting or appearance editing, which is our main goal. The second component is our proposed intrinsic decomposition approach whose goal is to disentangle geometry represented as normal and visibility information, and material properties as albedo and roughness. Our key insight is to represent these components in UV space rather than in 3D coordinates, using the density field as guidance in the mapping function between 3D and UV spaces. This allows our method to model complex deformations that are pose-dependent in a compact UV map that can be easily computed and queried. Finally, we integrate the microfacet model [52] into our rendering pipeline that takes as input the disentangled components (i.e., normal, visibility, albedo, and roughness) and an environment map, and renders the image of the neural actor in an arbitrary pose and light condition from a free-viewpoint.", + "bbox": [ + 212, + 474, + 787, + 746 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1 Preliminaries", + "text_level": 1, + "bbox": [ + 214, + 768, + 372, + 782 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Human Pose and Environment Map. Our method takes as input a 3D articulated human skeleton pose $\\mathbf{P}_t = \\{\\pmb {\\theta}_t,\\mathbf{R}_t,\\mathbf{T}_t\\}$ , defined by the relative joint rotation $\\pmb {\\theta}\\in \\mathbb{R}^{J\\times 3}$ of $J$ body joints, the global rotation $\\mathbf{R}\\in \\mathbb{R}^{3\\times 3}$ , and global", + "bbox": [ + 212, + 794, + 785, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "Relightable Neural Actor with Intrinsic Decomposition and Pose Control", + "bbox": [ + 246, + 114, + 730, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/4ba019d3bcb19d26ed0bef08e626dd22de5b901e5002eed5f4402d39d5e00ab3.jpg", + "image_caption": [ + "Fig. 2: Our method takes as input a 3D skeletal pose and a static environment map, and renders the neural actor from a virtual camera position. The pose-driven geometry (Sec. 3.1) learns an implicit density function, from which we compute the normal and visibility information. Our intrinsic decomposition disentangles normal, visibility, albedo and roughness maps in UV space (Sec. 3.2). The neural renderer (Sec. 3.3) outputs our prediction, which is supervised with the reference image. Yellow represents learnable components and green non-learnable components." + ], + "image_footnote": [], + "bbox": [ + 217, + 145, + 787, + 299 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "translation $\\mathbf{T} \\in \\mathbb{R}^3$ of the person at frame $t$ . The 3D pose can be used to drive a coarse human mesh with linear blending skinning (LBS) [34]. We represent the global lights as an environment map $\\mathbf{L} \\in \\mathbb{R}^{16 \\times 32 \\times 3}$ , which can be obtained from HDR images of a mirror sphere [8]—see the supplement for more details. The environment map can be static during training and changed during inference to render results under novel illumination.", + "bbox": [ + 212, + 449, + 787, + 541 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Training Data. During training, our method is supervised with multi-view calibrated video streams $\\mathbf{V}_c = \\{\\mathbf{I}_{c,t}\\}$ , where $c \\in \\{1, \\dots, C\\}$ is the camera index and $\\mathbf{I}_{c,t}$ is the image from camera $c$ at frame $t$ . We obtain a foreground mask $\\mathbf{M}_{c,t}$ for each video frame [30]. The coarse human mesh and 3D skeletons can be obtained with multi-view pose tracking [1]. The input 3D skeletons drive the LBS model to obtain posed human meshes used to extract texture maps from the multi-view images [2]. These texture maps are used during training to learn pose-dependent deformations on the implicit human model. At inference time, a V2V [55] network predicts texture maps from posed normal maps [32].", + "bbox": [ + 212, + 561, + 787, + 700 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implicit Human Model. For realistically relighting articulated avatars, we need to model human geometry with deformations caused by dynamic and varied poses. Therefore, we follow Neural Actor [32] to represent the pose-driven geometry as an implicit field guided by the 3D skeletal pose parameters and conditioned by texture maps, which provide the ability to model high-frequency deformations. Specifically, a standard NeRF model [39] is optimized through volumetric rendering in the canonical space of the coarse human mesh and the canonical space is deformed based on ResNet [16] features extracted from the", + "bbox": [ + 212, + 719, + 787, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "D.C. Luvizon et al.", + "bbox": [ + 271, + 114, + 400, + 127 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "texture maps, which provide the ability to represent detailed wrinkles and deformations not modeled by the coarse mesh. During optimization, from each pixel in the reference images we cast a ray $\\mathbf{r}_n = \\mathbf{o} + \\delta_n\\mathbf{d}$ , where $n = 1,\\dots N$ is the index of each point in the ray, $\\mathbf{o} \\in \\mathbb{R}^3$ is the origin defined in the optical center of the camera, $\\mathbf{d} \\in \\mathbb{R}^3$ is the ray direction, and $\\delta_{n}$ is the depth (distance to the camera). The sampled points $\\mathbf{r}_n$ are projected to the mesh surface, resulting in the local coordinates $(u_{n},v_{n},h_{n})$ , where $(u,v)$ is the surface coordinate in the UV space and $h$ is the signed distance to the mesh.", + "bbox": [ + 212, + 146, + 787, + 268 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.2 Neural Intrinsic Decomposition", + "text_level": 1, + "bbox": [ + 214, + 294, + 522, + 311 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "At the core of our method, we perform intrinsic decomposition to obtain the normal, visibility, albedo, and roughness components, required for rendering the neural actor under new light conditions. Specifically, for each point in each ray cast into the scene we want to obtain the normal direction $\\mathbf{n} \\in \\mathbb{R}^3$ , the visibility coefficients $\\pmb{v} \\in \\mathbb{R}^{512}$ respective to each light source in the environment map, the RGB albedo $\\mathbf{a} \\in \\mathbb{R}^3$ , and the scalar roughness $\\rho$ . The normal and visibility values can be computed from the geometry model. However, computing the visibility of each point in 3D with respect to the light sources is computationally intractable for a dynamic and long sequence, since for every point in 3D and for every frame, several rays have to be cast towards each pixel of the environment map. Therefore, we propose a different approach.", + "bbox": [ + 212, + 325, + 787, + 491 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We notice that the intrinsic decomposition is only relevant for points close to the surface. Since the coarse human mesh provides a mapping from the mesh surface to UV space and the implicit human model provides a density field that models the geometry of the actor, we propose an aggregation strategy that samples normal and visibility from a given point of view (therefore, only a few points are sampled in 3D) and aggregates the obtained values in a per-frame UV map. Once we obtain the normal and visibility in UV space for a given pose, we can simply query the values by projecting the sampled 3D point $\\mathbf{r}_n$ into the UV space, preventing our method from the expensive computation of normal and visibility at every possible 3D coordinate over multiple frames. We next explain this approach in more detail.", + "bbox": [ + 212, + 492, + 787, + 657 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Normal Maps. For each point in a ray, we want to obtain the normal value that will be required for solving the rendering equation during relighting (will be introduced in Sec. 3.3). This value could be obtained by computing the gradient from the neural field [62]. However, this process requires back-propagating gradients throughout the entire network and often results in noisy normals. Instead, we explicitly compute the normal vector close to the visible surface by deriving the unbiased depth $\\hat{\\mathbf{D}}$ [53]:", + "bbox": [ + 212, + 686, + 787, + 792 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathbf {D}} (\\mathbf {r}) = \\frac {\\sum_ {n = 1} ^ {N} w _ {n} \\delta_ {n}}{\\sum_ {n = 1} ^ {N} w _ {n}}, \\quad \\text {w i t h} w _ {n} = T _ {n} \\left(1 - \\exp \\left(- \\sigma_ {n} \\bar {\\delta} _ {n}\\right)\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 297, + 805, + 785, + 844 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "Relightable Neural Actor with Intrinsic Decomposition and Pose Control", + "bbox": [ + 246, + 114, + 732, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $T_{n} = \\exp \\left(-\\sum_{j = 1}^{n - 1}\\sigma_{j}\\bar{\\delta}_{j}\\right)$ , $w_{n}$ is the density weight from volumetric rendering [39], and $\\sigma_{n}$ is the density predicted by the NeRF model at point $\\mathbf{r}_n$ .", + "bbox": [ + 212, + 143, + 782, + 181 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Then, as illustrated in Fig. 3, we convert the depth values relative to the camera to 3D coordinates and obtain the normal direction from the tangent surface computed at the intersection point, defined as the maximum value of the density weight $w_{n}$ . This process can be performed for multiple viewpoints to cover the surface of the human.", + "bbox": [ + 212, + 181, + 486, + 330 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Once the normal values are obtained in 3D, we project them to the UV map and aggregate them with the", + "bbox": [ + 212, + 333, + 486, + 378 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "obtained normals from different views. In this process, the $(u^{*},v^{*})$ coordinate where the obtained normal will be projected onto the UV map is defined by", + "bbox": [ + 212, + 378, + 784, + 409 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/3b45f527bf19b66200054d5f0242d5e8ba44f8178da9396b69d1696f62a4cf61.jpg", + "image_caption": [ + "Fig. 3: The unbiased depth from Eq. (1) is projected into 3D coordinates, where we approximate the tangent surface at the intersecting point and obtain the normal vector. Normals are sampled from multiple viewpoints and aggregated in the UV map." + ], + "image_footnote": [], + "bbox": [ + 500, + 193, + 620, + 270 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/b1acf57fc92ded197bad1307778668bf79e3040b092aa4643c58511589a67fd5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 620, + 194, + 696, + 268 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/a01c98550c215374a72a6f5002bd4bb3cd0459a6d0457b1852be1e8ec7b2426e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 697, + 194, + 779, + 268 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\left(u ^ {*}, v ^ {*}\\right) = \\Psi \\left(\\mathbf {r} _ {n} ^ {*}\\right), \\quad \\text {w i t h} \\mathbf {r} _ {n} ^ {*} = \\underset {\\mathbf {r} _ {n}} {\\arg \\max } w _ {n}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 349, + 419, + 784, + 445 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $\\varPsi$ is a function that projects a 3D point to the closest surface point in the human mesh. Note that $w_{n}$ depends on the sampled points $\\mathbf{r}_n$ (from $T_{n}$ in Eq. 1) and $(u^{*},v^{*})$ corresponds to the $(u_{n},v_{n})$ value obtained from the point $\\mathbf{r}_n$ on the ray that has the largest density weight $w_{n}$ . This process ensures that the normal map obtained aggregates the normals most likely close to the surface.", + "bbox": [ + 212, + 455, + 782, + 530 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this process, the resulting UV map can be sparse, which is undesirable for free-viewpoint rendering. Thus, the NormalNet is used to inpaint the missing normals in UV space. This lightweight CNN is implemented with 8 layers of partial convolutions [31], takes as input sparse normal maps with size $512 \\times 512 \\times 3$ , and predicts a dense normal map (see Fig. 2). During training, NormalNet is supervised with densely sampled normal maps from the training data.", + "bbox": [ + 212, + 531, + 784, + 621 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Visibility Maps. The visibility information is necessary to model pose-dependent shadows in the neural avatar, without which it is not possible to produce realistic relighting results. A naive approach for obtaining the visibility for each point in 3D would require querying the neural field thousands of times while casting rays from the intersecting point to all the light sources, which is prohibitive during training. In contrast, in our approach we efficiently encode visibility of each point in 3D close to the surface by applying an approximate mapping to UV space. Specifically, we pre-compute the visibility maps for each pose by sampling multiple camera views and aggregate the results in UV map following Eq. (2). The resulting visibility maps with size $256 \\times 256 \\times 512$ encode the visibility information w.r.t. each light source in the environment map (represented by 512 individual lights). Similarly to the computation of normal maps, this process results in sparse visibility maps, which are inpainted by the VisibilityNet (as", + "bbox": [ + 212, + 643, + 792, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "D.C. Luvizon et al.", + "bbox": [ + 271, + 114, + 401, + 127 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "shown in Fig. 2). This shallow CNN takes as input sparse visibility and normal maps, as the visibility highly depends on normals (e.g. the visibility is always zero for the backlight source), and is trained with densely sampled visibility from the training sequence. See the supplement for architecture details.", + "bbox": [ + 212, + 146, + 782, + 205 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Albedo and Roughness Maps. As discussed above, the normal and visibility components can be computed from the density field by using UV maps as proxies. However, the albedo and roughness properties cannot be directly obtained from the pre-trained Neural Actor model, since it predicts radiance which is the product of the rendering equation. Instead, we want to learn a decomposed representation. In intrinsic decomposition, the material properties are commonly optimized via differentiable rendering. Hence, we optimize the albedo and roughness of the neural actor by leveraging a differentiable renderer (Sec. 3.3) and by imposing priors through our design.", + "bbox": [ + 212, + 224, + 784, + 359 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "The material properties of an actor should not depend on the pose and should be efficiently encoded to enable fast computations and editing. Therefore, we model albedo and roughness as static UV maps, respectively defined as $\\mathbf{A} \\in \\mathbb{R}^{1024 \\times 1024 \\times 3}$ and $\\mathbf{B} \\in \\mathbb{R}^{1024 \\times 1024}$ . These tensors are jointly optimized for the whole training sequence and regularized with an L1 loss on the spatial gradients, which encourages smoothness in both maps:", + "bbox": [ + 212, + 359, + 784, + 444 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {s m o o t h}} = \\frac {1}{N} \\sum \\left| \\nabla_ {x} (\\mathbf {A}) \\right| + \\left| \\nabla_ {y} (\\mathbf {A}) \\right| + \\left| \\nabla_ {x} (\\mathbf {B}) \\right| + \\left| \\nabla_ {y} (\\mathbf {B}) \\right|, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 308, + 449, + 784, + 473 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where $\\nabla_{x}$ and $\\nabla_{y}$ represent the spatial gradient operators.", + "bbox": [ + 218, + 477, + 643, + 491 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Nonetheless, by assuming a material representation that does not depend on the skeletal pose, we still have to cope with intrinsic mesh registration errors. Such registration errors could cause blurry results or artifacts in the final renderings. To cope with this problem, we propose a corrective term to the UV mapping, which is predicted by the UVDeltaNet, as shown in Fig. 2. The UVDeltaNet is an MLP that takes as input the local coordinates $(u,v,h)$ for each point on the ray and the sampled texture features $\\psi$ , and outputs a corrective term $\\epsilon = (u',v')$ . The texture features are necessary to provide additional information, since the local coordinates alone cannot account for deviations in the mesh registration.", + "bbox": [ + 212, + 492, + 785, + 642 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Initially, UVDeltaNet is frozen and outputs zeros. After the initial convergence of $\\mathbf{A}$ and $\\mathbf{B}$ , UVDeltaNet is indirectly supervised by the rendering loss and regularized with:", + "bbox": [ + 212, + 643, + 782, + 686 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {u v}} = \\frac {1}{N} \\sum \\epsilon^ {2} + | \\nabla_ {x} (\\epsilon) | + | \\nabla_ {y} (\\epsilon) |, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 372, + 694, + 784, + 722 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "which minimizes the magnitude of the deformation and encourages its spatial gradient to be sparse, i.e., it makes the UV deformation spatially coherent.", + "bbox": [ + 212, + 726, + 782, + 756 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "3.3 Neural Renderer", + "text_level": 1, + "bbox": [ + 214, + 773, + 401, + 787 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "The third part of our method is our neural renderer, which takes the disentangled components as normal, visibility, albedo, and roughness; integrates the lights from the environment map, and outputs the rendered image.", + "bbox": [ + 212, + 794, + 782, + 839 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "Relightable Neural Actor with Intrinsic Decomposition and Pose Control", + "bbox": [ + 246, + 114, + 732, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Our neural renderer starts by recasting the ray $\\mathbf{r}$ with importance sampling around the estimated depth value $\\hat{\\mathbf{D}}$ . Thus, all the points $\\mathbf{r}_n$ are most likely close to the surface. For each sampled point in the ray, we obtain the normal direction $\\mathbf{n}_n$ and the visibility values $\\pmb{v}_n$ by interpolating the normal and visibility maps in UV around the point $(u_n,v_n) = \\varPsi(\\mathbf{r}_n)$ . The albedo $\\mathbf{a}_n$ and roughness $\\rho_n$ are obtained by interpolating the albedo and roughness maps in UV around the point $(u_n + u'_u,v_n + v'_n)$ , i.e., following the deformation predicted by UVDeltaNet.", + "bbox": [ + 212, + 146, + 787, + 252 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Before rendering the final image of the posed neural actor, we consider the light transport equation defined as", + "bbox": [ + 212, + 252, + 785, + 282 + ], + "page_idx": 9 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {L} _ {0} (\\mathbf {r} _ {n}, \\boldsymbol {\\omega} _ {0}) = \\sum_ {i} \\operatorname {R} (\\mathbf {a} _ {n}, \\rho_ {n}, \\boldsymbol {\\omega} _ {i}, \\boldsymbol {\\omega} _ {0}) \\boldsymbol {v} _ {n, i} \\mathbf {L} (\\boldsymbol {\\omega} _ {i}) (\\boldsymbol {\\omega} _ {i} \\mathbf {n} _ {n}), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 318, + 294, + 785, + 324 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "where $\\mathrm{R}(\\cdot)$ is the microfacet BRDF model [52], $\\omega_0$ and $\\omega_{i}$ are the camera and the light source directions, respectively, $\\pmb{v}_{n,i}$ is the visibility of point $\\mathbf{r}_n$ w.r.t. the ith light source, and $\\mathbf{L}(\\boldsymbol {\\omega}_i)$ is the incoming light from the environment map. Eq. (5) is defined for a single point $\\mathbf{r}_n$ on the ray and is integrated into our volume rendering as", + "bbox": [ + 212, + 334, + 787, + 409 + ], + "page_idx": 9 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {\\mathbf {I}} (\\mathbf {r}) = \\sum_ {n = 1} ^ {N} w _ {n} \\mathbf {L} _ {0} \\left(\\mathbf {r} _ {n}, \\omega_ {0}\\right), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 413, + 410, + 785, + 450 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "where $\\tilde{\\mathbf{I}}$ is the output image of the relighted neural actor.", + "bbox": [ + 214, + 457, + 629, + 472 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "During training, our relightable model is supervised with the following loss:", + "bbox": [ + 238, + 473, + 779, + 488 + ], + "page_idx": 9 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {r e l i g h t}} = \\lambda_ {\\mathrm {L 2}} \\mathcal {L} _ {\\mathrm {L 2}} + \\lambda_ {\\text {v g g}} \\mathcal {L} _ {\\text {v g g}} + \\lambda_ {\\text {s m o o t h}} \\mathcal {L} _ {\\text {s m o o t h}} + \\mathcal {L} _ {\\text {u v}}, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 310, + 501, + 785, + 517 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "where $\\mathcal{L}_{\\mathrm{L2}}$ and $\\mathcal{L}_{\\mathrm{vgg}}$ are the L2 and the perceptual losses [26] between $\\tilde{\\mathbf{I}} (\\mathbf{r})$ and the reference image. The hyperparameters are defined as $\\lambda_{\\mathrm{L2}} = 100$ , $\\lambda_{\\mathrm{vgg}} = 0.01$ , and $\\lambda_{\\mathrm{smooth}} = 0.1$ in all our experiments.", + "bbox": [ + 212, + 527, + 787, + 574 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4 Relightable Dynamic Actors Dataset", + "text_level": 1, + "bbox": [ + 214, + 595, + 606, + 614 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Existing datasets for human reenactment and relighting are often monocular [2, 17], contain very short sequences with only a few pose variations, and do not provide environment maps [2,17,32,42]. Hence, we cannot use them for evaluation under realistic scenarios with real ground truth. In addition, we observe that the only existing synthetic dataset for human relighting and pose control [17] suffers from artifacts in the geometry and salt-and-pepper noise in the renderings.", + "bbox": [ + 212, + 628, + 787, + 718 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "With these shortcomings in mind, we collect a new real dataset, which we call Relightable Dynamic Actors, with four identities performing a series of ten different actions (Fig. 4). Each subject is recorded six times wearing the same clothes but under different light conditions, including indoor and outdoor, where we used around 100 and eight cameras, respectively. For each sequence, we also capture HDR light probes, which can be used for training or testing. In total, our dataset contains around 90k multi-view video frames divided into 24 multi-view video sequences. We refer to the supplemental document for more details.", + "bbox": [ + 212, + 719, + 787, + 839 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "D.C. Luvizon et al.", + "bbox": [ + 271, + 114, + 400, + 127 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/a06b68f5e197c3c761b0ca2e1bf0c0a96d8720e1eb5027a3c6ac90444f51c1f3.jpg", + "image_caption": [ + "Fig. 4: Relightable Dynamic Actors dataset. For each actor, we recorded six sequences under different light conditions, including five indoor sequences and one outdoor. Environment maps were obtained from HDR pictures of a mirror sphere." + ], + "image_footnote": [], + "bbox": [ + 215, + 143, + 500, + 270 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/ea85ef744f3f6e88d203721ec7d31977792d3e0697e883150fee5d73e31dc5d1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 501, + 143, + 784, + 270 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "5 Experiments", + "text_level": 1, + "bbox": [ + 215, + 364, + 375, + 382 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "In this section, we provide an experimental evaluation of our method on the Reli-ightable Dynamic Actors dataset and show qualitative and quantitative results. Please refer to the supplementary video for dynamic results on long sequences.", + "bbox": [ + 214, + 397, + 782, + 444 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "5.1 Evaluation", + "text_level": 1, + "bbox": [ + 215, + 465, + 351, + 479 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Metrics. For quantitative evaluation, we report results on the PSNR, SSIM and LPIPS metrics [61], comparing the rendered relight results with the groundtruth images from our dataset. Importantly, our work is the first to be able to report such metrics on a real dataset since no previous dataset in the literature provided a real human recorded under different and known light conditions.", + "bbox": [ + 212, + 492, + 784, + 566 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Baselines. Since the only method in the literature capable of reposing a human under new light conditions does not provide implementation [17], and Relighting4D [6] is not able to perform inference for new poses, we defined two baselines for a fair comparison. Our baselines are derived from Neural Actor [32] and adapted to perform intrinsic decomposition based on the obtained normal and visibility information, as discussed in Sec. 3.2. However, instead of using NormalNet and VisibilityNet to obtain these components, we densely sample the neural field and perform morphological inpainting to obtain dense UV maps. The first baseline, referred to as Neural Actor $+IR$ , predicts the albedo and roughness components instead of RGB color directly from the deformed canonical space $\\mathbf{x}_n = (x,y,z)$ (see Fig. 2). The second baseline, Neural Actor $+IR + Tex$ , takes as input the texture features $\\psi$ as a conditioning for predicting the albedo and roughness values. In both cases, the predictor is a NeRF MLP with 8 layers.", + "bbox": [ + 214, + 568, + 787, + 763 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Implementation Details. Our model is trained in a three-stage process. First, we train the geometry component (Sec. 3.1) with the Adam [23] optimizer with a learning rate of 0.001 until convergence, which takes about two days on four NVidia A40 GPUs. Then, we pre-train the NormalNet and VisibilityNet with pre-computed normal and visibility UV maps sampled from the geometry model", + "bbox": [ + 214, + 765, + 787, + 840 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "Relightable Neural Actor with Intrinsic Decomposition and Pose Control", + "bbox": [ + 246, + 114, + 730, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 116, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/d89fd9d77195b1decc66845928fbb1097f4a8947c98467a015c03010339021a3.jpg", + "image_caption": [ + "Fig. 5: Qualitative results of our method on novel poses. Our approach generalizes to both indoor and outdoor illumination, and generates realistic renderings for arbitrary environment maps with detailed shadows and pose-dependent deformations." + ], + "image_footnote": [], + "bbox": [ + 290, + 143, + 712, + 285 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "from 40 different views. Both networks are trained in parallel using Adam with a learning rate of 0.01 for about 12 hours. Finally, we train our intrinsic decomposition model with the same optimizer until convergence, which takes about one day. Since the obtained environment maps are up to factor of scale [8], we apply global color correction during training and evaluation. The predicted texture maps are obtained with a V2V model [55] trained (in parallel to our model) to translate a sequence of SMPL normal maps into a sequence of texture maps. Please refer to our supplementary materials for architecture details.", + "bbox": [ + 212, + 356, + 784, + 477 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "5.2 Results", + "text_level": 1, + "bbox": [ + 215, + 497, + 321, + 511 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Qualitative Results. In Fig. 5, we show qualitative results of our method. Our results are all from a trained model driven by a new skeletal pose, new camera viewpoint, and new environment map not seen during training. Note how our approach produces realistic appearance and prominent cloth wrinkles, which can be easily observed under different light conditions. Importantly, notice how our results look realistic under novel lights, even though our model is trained on a single sequence under a static light condition.", + "bbox": [ + 212, + 521, + 784, + 626 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/579856259c9e3880467056a3905e461147e4ed8cd613e0b345aac11cc8301087.jpg", + "table_caption": [ + "Table 1: Quantitative results from our method on real data, subject S1, considering indoor and outdoor scenes. Our method consistently improves the baselines, especially in outdoor scenarios, where the gap between training and test lights is more pronounced." + ], + "table_footnote": [], + "table_body": "
MethodT2T3T4T5T6 (outdoors)
PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓
Neural Actor+IR17.5220.7500.18117.4730.7680.17817.7450.7670.17118.2440.7680.17518.3550.7100.203
Neural Actor+IR+Tex18.1750.7780.16417.7990.7900.16718.1320.7850.16718.5310.7890.16618.7040.7370.186
Ours18.6020.7920.16318.2400.8000.16918.3880.8000.16418.8200.8000.16519.4610.7530.173
", + "bbox": [ + 215, + 705, + 785, + 760 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "In Fig. 6, we show results from our intrinsic decomposition. Note that our normal representation nicely captures the geometry details and the visibility. This is important to obtain an albedo that does not contain baked-in lighting effects, e.g. shadows, as can be observed from our results.", + "bbox": [ + 212, + 776, + 784, + 835 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "D.C. Luvizon et al.", + "bbox": [ + 271, + 114, + 400, + 126 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/0dd55a64e873c05df6a67eec8ba9595a19834eecddf6d8249c71aa9f30b6c696.jpg", + "image_caption": [ + "Fig. 6: Results from our intrinsic decomposition. Our method provides detailed normals and visibility (averaged for visualization), and plausible albedo and roughness predictions that enable realistic relighting." + ], + "image_footnote": [], + "bbox": [ + 287, + 142, + 715, + 333 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Another important and unique characteristic of our method is the ability to model self-shadows and cloth wrinkles. In Fig. 7, we show results from two subjects, where a \"sunrise\" light probe was used to produce cast shadows. Note how pose-dependent effects are consistent with the mostly directional illumination.", + "bbox": [ + 212, + 411, + 542, + 516 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Comparisons. We evaluated our method on the test sequences from our dataset. Specifically, we compare against two baselines in the indoor sequences (T2-T5) and outdoor sequence (T6) of subject S1 in every 10th frame, which sums up to 4000 evaluation frames. In Tab. 1 we show, for the first time, quantitative results of real human relighting under new poses and new light conditions. Our approach consistently improves upon the two baselines.", + "bbox": [ + 212, + 518, + 542, + 670 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/5b5fbd96a5687646854daa587d130d438fb3408cdb371bc3824a1159e27170b2.jpg", + "image_caption": [ + "Fig. 7: Our method models pose-dependent effects and cast shadows, as can be observed from our results under the \"sunrise\" light." + ], + "image_footnote": [], + "bbox": [ + 562, + 383, + 782, + 594 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Ablation. In Tab. 2, we show an ablation with different components of our method. Compared to our baseline (Neural Actor+IR), learning the albedo and roughness as UV maps guided by the neural field significantly improves the performance on relighting and reposing. This is because, in the baselines, the albedo and roughness are under-constrained and can possibly embed shading effects, overfitting to the training data. The NormalNet and VisibilityNet provide additional capacity to the model to learn how to inpaint the sampled UV maps. Note that without these two models our results are computed using morphological region growth in the UV maps. Finally, the UVDeltaNet provides more flexibility to the albedo and roughness representation, accounting for fine details. This can be observed in the improvement in the LPIPS metric in the last", + "bbox": [ + 212, + 672, + 787, + 840 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "Relightable Neural Actor with Intrinsic Decomposition and Pose Control", + "bbox": [ + 246, + 114, + 730, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "row of Tab. 2. Nonetheless, this also means more capacity to overfit the training sequences, which translates into a small decrease in the PSNR metric.", + "bbox": [ + 212, + 146, + 782, + 176 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5.3 Applications", + "text_level": 1, + "bbox": [ + 215, + 200, + 366, + 215 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "An important characteristic of our approach is our static representation in UV for the albedo and roughness. This allows exciting and new applications such as color and material editing of an animatable and relightable neural actor, as shown in Fig. 8 and in Fig. 1. With a simple edit in the obtained albedo or rough-", + "bbox": [ + 212, + 226, + 431, + 391 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Table 2: Ablation study with different components of our method. The proposed intrinsic decomposition with static albedo and roughness maps significantly improves over the baseline, and the proposed NormalNet and VisibilityNet further improve the results while enabling free-viewpoint rendering. UVDeltaNet provides more realistic results (LPIPS metric) with a small sacrifice in the PSNR.", + "bbox": [ + 439, + 195, + 787, + 305 + ], + "page_idx": 13 + }, + { + "type": "table", + "img_path": "images/4f239530ef29c5d109f17ce1035e7d4cfad3d924dc0acf50596241a05ac81e4d.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodIndoors (S1,T2)Outdoors (S1,T6)
PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓
Baseline17.5220.7500.18118.3550.7100.203
+Albedo/Rough. Maps18.3340.7890.16819.3410.7470.181
+Normal/Vis.Net18.6340.7920.17319.7040.7600.180
+UVDeltaNet (Ours)18.6020.7920.16319.4610.7530.173
", + "bbox": [ + 444, + 306, + 784, + 378 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "ness 2D map, we can insert new textures or new materials and render the edited neural actor under new poses, new viewpoints, and new light conditions.", + "bbox": [ + 212, + 393, + 782, + 422 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "6 Limitations", + "text_level": 1, + "bbox": [ + 215, + 446, + 362, + 462 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "This work focuses on visual fidelity, and representations faster than NeRF could be tested in our framework in future. Moreover, like several previous techniques for the free-viewpoint rendering of humans and avatar generation [32,45], the proposed method can currently only deal with tight clothing due to the coarse geometry registration. The coarse mesh registration", + "bbox": [ + 212, + 479, + 486, + 630 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/bb58ca6ec11e0b637ecdbbdbccf9a05600ffbc19f4df560ea6561ae712b818da.jpg", + "image_caption": [ + "Fig.8: Our method allows editing the material properties and simultaneously performing skeletal pose control, free-viewpoint rendering, and relighting of the edited neural actor." + ], + "image_footnote": [], + "bbox": [ + 506, + 446, + 785, + 537 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "can also negatively affect the fine details in the face and hand synthesis.", + "bbox": [ + 212, + 631, + 733, + 646 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "7 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 670, + 359, + 686 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "We conclude that the proposed method for the free-viewpoint rendering of humans enables simultaneous relighting, material editing, and pose control at test time while requiring a single multi-view video sequence with static illumination during training. Our method demonstrates functionality that no competing method making the same assumptions at training time (i.e. from multi-view RGB video inputs only) can provide. We hope that our method and the proposed dataset will provoke further research in this exciting field, bringing the neural rendering of humans to a new qualitative level and paving the way for next-generation metaverse applications.", + "bbox": [ + 212, + 703, + 787, + 840 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "D.C. Luvizon et al.", + "bbox": [ + 271, + 114, + 400, + 127 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 217, + 145, + 401, + 162 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "This work was funded by the ERC Consolidator Grant 4DRepLy (770784), the Saarbrücken Research Center for Visual Computing, Interaction and AI, and the German Research Foundation (DFG) under Grant No. 468670075.", + "bbox": [ + 215, + 175, + 785, + 220 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 243, + 321, + 258 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Easymocap - make human motion capture easier. Github (2021), https://github.com/zju3dv/EasyMocap", + "2. Alldieck, T., Magnor, M., Xu, W., Theobalt, C., Pons-Moll, G.: Video based reconstruction of 3d people models. In: Computer Vision and Pattern Recognition (CVPR). pp. 8387-8397 (2018)", + "3. Alldieck, T., Xu, H., Sminchisescu, C.: imghum: Implicit generative models of 3d human shape and articulated pose. In: International Conference on Computer Vision (ICCV). pp. 5461-5470 (2021)", + "4. Bi, S., Lombardi, S., Saito, S., Simon, T., Wei, S.E., Mcphail, K., Ramamoorthi, R., Sheikh, Y., Saragih, J.: Deep relightable appearance models for animatable faces. ACM Transactions on Graphics (TOG) 40(4), 1-15 (2021)", + "5. Boss, M., Braun, R., Jampani, V., Barron, J.T., Liu, C., Lensch, H.: Nerd: Neural reflectance decomposition from image collections. In: International Conference on Computer Vision (ICCV). pp. 12684-12694 (2021)", + "6. Chen, Z., Liu, Z.: Relighting4d: Neural relightable human from videos. In: European Conference on Computer Vision (ECCV) (2022)", + "7. De Aguiar, E., Theobalt, C., Thrun, S., Seidel, H.P.: Automatic conversion of mesh animations into skeleton-based animations. In: Computer Graphics Forum (EUROGRAPHICS). vol. 27, pp. 389-397 (2008)", + "8. Debevec, P.E., Malik, J.: Recovering high dynamic range radiance maps from photographs. In: ACM SIGGRAPH 2008 classes, pp. 1-10 (2008)", + "9. Deng, B., Wang, Y., Wetzstein, G.: Unconditional generation of relightable 3d human faces. In: International Conference on 3D Vision (3DV). pp. 302-312. IEEE (2024)", + "0. d'Eon, E., Luebke, D., Enderton, E.: Efficient Rendering of Human Skin. In: Kautz, J., Pattanaik, S. (eds.) Rendering Techniques. The Eurographics Association (2007). https://doi.org/10.2312/EGWR/EGSR07/147-157", + "1. Guo, K., Lincoln, P., Davidson, P., Busch, J., Yu, X., Whalen, M., Harvey, G., Orts-Escalano, S., Pandey, R., Dourgarian, J., et al.: The relighttables: Volumetric performance capture of humans with realistic relighting. ACM Transactions on Graphics (TOG) 38(6), 1-19 (2019)", + "2. Habermann, M., Liu, L., Xu, W., Pons-Moll, G., Zollhoefer, M., Theobalt, C.: Hdhumans: A hybrid approach for high-fidelity digital humans. Proc. ACM Comput. Graph. Interact. Tech. 6(3) (aug 2023). https://doi.org/10.1145/3606927, https://doi.org/10.1145/3606927", + "3. Habermann, M., Liu, L., Xu, W., Zollhoefer, M., Pons-Moll, G., Theobalt, C.: Real-time deep dynamic characters. ACM Transactions on Graphics 40(4) (aug 2021)", + "4. Habermann, M., Xu, W., Zollhoefer, M., Pons-Moll, G., Theobalt, C.: Livecap: Real-time human performance capture from monocular video. ACM Transactions On Graphics (TOG) 38(2), 1-17 (2019)" + ], + "bbox": [ + 225, + 273, + 784, + 840 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "Relightable Neural Actor with Intrinsic Decomposition and Pose Control", + "bbox": [ + 246, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "15. Habermann, M., Xu, W., Zollhofer, M., Pons-Moll, G., Theobalt, C.: Deepcap: Monocular human performance capture using weak supervision. In: Computer Vision and Pattern Recognition (CVPR). pp. 5052-5063 (2020)", + "16. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Computer Vision and Pattern Recognition (CVPR). pp. 770-778 (2016)", + "17. Iqbal, U., Caliskan, A., Nagano, K., Khamis, S., Molchanov, P., Kautz, J.: Rana: Relightable articulated neural avatars. In: International Conference on Computer Vision (ICCV). pp. 23142-23153 (2023)", + "18. Iwase, S., Saito, S., Simon, T., Lombardi, S., Bagautdinov, T., Joshi, R., Prada, F., Shiratori, T., Sheikh, Y., Saragih, J.: Relightablehands: Efficient neural relighting of articulated hand models. In: Computer Vision and Pattern Recognition (CVPR). pp. 16663-16673 (2023)", + "19. Ji, C., Yu, T., Guo, K., Liu, J., Liu, Y.: Geometry-aware single-image full-body human relighting. In: European Conference on Computer Vision (ECCV). pp. 388-405. Springer (2022)", + "20. Jiang, W., Yi, K.M., Samei, G., Tuzel, O., Ranjan, A.: Neuman: Neural human radiance field from a single video. In: European Conference on Computer Vision (ECCV). pp. 402-418 (2022)", + "21. Jiang, Y., Habermann, M., Golyanik, V., Theobalt, C.: Hifecap: Monocular high-fidelity and expressive capture of human performances. In: British Machine Vision Conference (BMVC) (2022)", + "22. Jin, H., Liu, I., Xu, P., Zhang, X., Han, S., Bi, S., Zhou, X., Xu, Z., Su, H.: Tensoroir: Tensorial inverse rendering. In: Computer Vision and Pattern Recognition (CVPR). pp. 165-174 (2023)", + "23. Kingma, D.P., Ba, J.: Adam: A method for stochastic optimization. CoRR abs/1412.6980 (2014), https://api-semanticscholar.org/CorpusID:6628106", + "24. Kwon, Y., Kim, D., Ceylan, D., Fuchs, H.: Neural human performer: Learning generalizable radiance fields for human performance rendering. Advances in Neural Information Processing Systems (NeurIPS) 34, 24741-24752 (2021)", + "25. Kwon, Y., Liu, L., Fuchs, H., Habermann, M., Theobalt, C.: Deliffas: Deformable light fields for fast avatar synthesis. Advances in Neural Information Processing Systems (2023)", + "26. Ledig, C., Theis, L., Huszar, F., Caballero, J., Cunningham, A., Acosta, A., Aitken, A., Tejani, A., Totz, J., Wang, Z., et al.: Photo-realistic single image superresolution using a generative adversarial network. In: Computer Vision and Pattern Recognition (CVPR). pp. 4681-4690 (2017)", + "27. Lewis, J.P., Cordner, M., Fong, N.: Pose space deformation: A unified approach to shape interpolation and skeleton-driven deformation. In: SIGGRAPH. p. 165-172. ACM Press/Addison-Wesley Publishing Co., USA (2000). https://doi.org/10.1145/344779.344862, https://doi.org/10.1145/344779.344862", + "28. Li, G., Meka, A., Mueller, F., Buehler, M.C., Hilliges, O., Beeler, T.: Eyenerf: a hybrid representation for photorealistic synthesis, animation and relighting of human eyes. ACM Transactions on Graphics (TOG) 41(4), 1-16 (2022)", + "29. Li, G., Wu, C., Stoll, C., Liu, Y., Varanasi, K., Dai, Q., Theobalt, C.: Capturing relightable human performances under general uncontrolled illumination. In: Computer Graphics Forum. vol. 32, pp. 275-284. Wiley Online Library (2013)", + "30. Lin, S., Ryabtsev, A., Sengupta, S., Curless, B.L., Seitz, S.M., Kemelmacher-Shlizerman, I.: Real-time high-resolution background matting. In: Computer Vision and Pattern Recognition (CVPR). pp. 8762-8771 (2021)" + ], + "bbox": [ + 217, + 146, + 784, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "D.C. Luvizon et al.", + "bbox": [ + 271, + 114, + 400, + 126 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "31. Liu, G., Reda, F.A., Shih, K.J., Wang, T.C., Tao, A., Catanzaro, B.: Image inpainting for irregular holes using partial convolutions. In: European Conference on Computer Vision (ECCV) (2018)", + "32. Liu, L., Habermann, M., Rudnev, V., Sarkar, K., Gu, J., Theobalt, C.: Neural actor: Neural free-view synthesis of human actors with pose control. ACM Transactions on Graphics (TOG) 40(6), 1-16 (2021)", + "33. Liu, Y., Wang, P., Lin, C., Long, X., Wang, J., Liu, L., Komura, T., Wang, W.: Nano: Neural geometry and brdf reconstruction of reflective objects from multiview images. In: SIGGRAPH (2023)", + "34. Loper, M., Mahmood, N., Romero, J., Pons-Moll, G., Black, M.J.: SMPL: A skinned multi-person linear model. ACM Transactions on Graphics (SIGGRAPH Asia) 34(6), 248:1-248:16 (2015)", + "35. Lyu, L., Tewari, A., Leimkuehler, T., Habermann, M., Theobalt, C.: Neural radiance transfer fields for relightable novel-view synthesis with global illumination. In: European Conference on Computer Vision (ECCV) (2022)", + "36. Meka, A., Pandey, R., Hane, C., Orts-Escolano, S., Barnum, P., David-Son, P., Erickson, D., Zhang, Y., Taylor, J., Bouaziz, S., Legendre, C., Ma, W.C., Overbeck, R., Beeler, T., Debevec, P., Izadi, S., Theobalt, C., Rhemann, C., Fanello, S.: Deep relightable textures: Volumetric performance capture with neural rendering. ACM Transactions on Graphics (TOG) 39(6) (nov 2020). https://doi.org/10.1145/3414685.3417814, https://doi.org/10.1145/3414685.3417814", + "37. Mihajlovic, M., Saito, S., Bansal, A., Zollhoefer, M., Tang, S.: COAP: Compositional articulated occupancy of people. In: Computer Vision and Pattern Recognition (CVPR) (2022)", + "38. Mihajlovic, M., Zhang, Y., Black, M.J., Tang, S.: Leap: Learning articulated occupancy of people. In: Computer Vision and Pattern Recognition (CVPR). pp. 10461-10471 (2021)", + "39. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. In: European Conference on Computer Vision (ECCV) (2020)", + "40. Pang, H., Zhu, H., Kortylewski, A., Theobalt, C., Habermann, M.: Ash: Animatable gaussian splats for efficient and photoreal human rendering. In: Computer Vision and Pattern Recognition (CVPR). pp. 1165-1175 (June 2024)", + "41. Pavlakos, G., Choutas, V., Ghorbani, N., Bolkart, T., Osman, A.A.A., Tzionas, D., Black, M.J.: Expressive body capture: 3d hands, face, and body from a single image. In: Computer Vision and Pattern Recognition (CVPR) (2019)", + "42. Peng, S., Zhang, Y., Xu, Y., Wang, Q., Shuai, Q., Bao, H., Zhou, X.: Neural body: Implicit neural representations with structured latent codes for novel view synthesis of dynamic humans. In: Computer Vision and Pattern Recognition (CVPR) (2021)", + "43. Ramamoorthi, R., Hanrahan, P.: A signal-processing framework for inverse rendering. In: Conference on Computer Graphics and Interactive Techniques. p. 117-128. SIGGRAPH '01, Association for Computing Machinery, New York, NY, USA (2001). https://doi.org/10.1145/383259.383271, https://doi.org/10.1145/383259.383271", + "44. Ranjan, A., Yi, K.M., Chang, J.H.R., Tuzel, O.: Facelit: Neural 3d relightable faces. In: Computer Vision and Pattern Recognition (CVPR). pp. 8619-8628 (2023)", + "45. Remelli, E., Bagautdinov, T., Saito, S., Wu, C., Simon, T., Wei, S.E., Guo, K., Cao, Z., Prada, F., Saragih, J., et al.: Drivable volumetric avatars using texel-aligned features. In: ACM SIGGRAPH 2022 Conference Proceedings (2022)" + ], + "bbox": [ + 215, + 146, + 784, + 839 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "Relightable Neural Actor with Intrinsic Decomposition and Pose Control", + "bbox": [ + 246, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "46. Schwartz, G., Wei, S.E., Wang, T.L., Lombardi, S., Simon, T., Saragih, J., Sheikh, Y.: The eyes have it: An integrated eye and face model for photorealistic facial animation. ACM Transactions on Graphics (TOG) 39(4), 91-1 (2020)", + "47. Shu, Z., Sahasrabudhe, M., Guler, R.A., Samaras, D., Paragios, N., Kokkinos, I.: Deforming autoencoders: Unsupervised disentangling of shape and appearance. In: European Conference on Computer Vision (ECCV). pp. 650-665 (2018)", + "48. Shysheya, A., Zakharov, E., Aliev, K.A., Bashirov, R., Burkov, E., Iskakov, K., Ivakshtenko, A., Malkov, Y., Pasechnik, I., Ulyanov, D., et al.: Textured neural avatars. In: Computer Vision and Pattern Recognition (CVPR). pp. 2387-2397 (2019)", + "49. Srinivasan, P.P., Deng, B., Zhang, X., Tancik, M., Mildenhall, B., Barron, J.T.: Nerv: Neural reflectance and visibility fields for relighting and view synthesis. In: Computer Vision and Pattern Recognition (CVPR). pp. 7491-7500 (2021). https://doi.org/10.1109/CVPR46437.2021.00741", + "50. Tan, F., Fanello, S., Meka, A., Orts-Escalano, S., Tang, D., Pandey, R., Taylor, J., Tan, P., Zhang, Y.: Volux-gan: A generative model for 3d face synthesis with hdri relighting. In: ACM SIGGRAPH 2022 Conference Proceedings. pp. 1-9 (2022)", + "51. Tewari, A., Thies, J., Mildenhall, B., Srinivasan, P., Tretschk, E., Yifan, W., Lassner, C., Sitzmann, V., Martin-Brualla, R., Lombardi, S., Simon, T., Theobalt, C., Nießner, M., Barron, J.T., Wetzstein, G., Zollhöfer, M., Golyanik, V.: Advances in Neural Rendering. Computer Graphics Forum (EG STAR 2022) (2022)", + "52. Walter, B., Marschner, S.R., Li, H., Torrance, K.E.: Microfacet Models for Refraction through Rough Surfaces. In: Kautz, J., Pattanaik, S. (eds.) Rendering Techniques. The Eurographics Association (2007). https://doi.org/10.2312/EGWR/EGSR07/195-206", + "53. Wang, P., Liu, L., Liu, Y., Theobalt, C., Komura, T., Wang, W.: Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. In: Advances in Neural Information Processing Systems (NeurIPS) (2021)", + "54. Wang, S., Schwarz, K., Geiger, A., Tang, S.: Arah: Animatable volume rendering of articulated human sdfs. In: European Conference on Computer Vision (ECCV) (2022)", + "55. Wang, T.C., Liu, M.Y., Zhu, J.Y., Liu, G., Tao, A., Kautz, J., Catanzaro, B.: Video-to-video synthesis. In: Conference on Neural Information Processing Systems (NeurIPS) (2018)", + "56. Wang, Y., Holynski, A., Zhang, X., Zhang, X.: Sunstage: Portrait reconstruction and relighting using the sun as a light stage. In: Computer Vision and Pattern Recognition (CVPR). pp. 20792-20802 (2023)", + "57. Weng, C.Y., Curless, B., Srinivasan, P.P., Barron, J.T., Kemelmacher-Shlizerman, I.: Humannerf: Free-viewpoint rendering of moving people from monocular video. In: Computer Vision and Pattern Recognition (CVPR). pp. 16210-16220 (June 2022)", + "58. Wuu, C.h., Zheng, N., Ardisson, S., Bali, R., Belko, D., Brockmeyer, E., Evans, L., Godisart, T., Ha, H., Huang, X., Hypes, A., Koska, T., Krenn, S., Lombardi, S., Luo, X., McPhail, K., Millerschoen, L., Perdoch, M., Pitts, M., Richard, A., Saragih, J., Saragih, J., Shiratori, T., Simon, T., Stewart, M., Trimble, A., Weng, X., Whitewolf, D., Wu, C., Yu, S.I., Sheikh, Y.: Multiface: A dataset for neural face rendering. In: arXiv (2022). https://doi.org/10.48550/ARXIV.2207.11243, https://arxiv.org/abs/2207.11243", + "59. Yang, H., Zheng, M., Feng, W., Huang, H., Lai, Y.K., Wan, P., Wang, Z., Ma, C.: Towards practical capture of high-fidelity reconfigurable avatars. In: SIGGRAPH Asia 2023 Conference Papers. pp. 1-11 (2023)" + ], + "bbox": [ + 215, + 146, + 784, + 840 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "D.C. Luvizon et al.", + "bbox": [ + 271, + 114, + 400, + 126 + ], + "page_idx": 17 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "60. Zhang, K., Luan, F., Li, Z., Snavely, N.: IRON: Inverse rendering by optimizing neural sdfs and materials from photometric images. In: Computer Vision and Pattern Recognition (CVPR). pp. 5565-5574 (2022)", + "61. Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual metric. In: Computer Vision and Pattern Recognition (CVPR) (2018)", + "62. Zhang, X., Srinivasan, P.P., Deng, B., Debevec, P., Freeman, W.T., Barron, J.T.: NeRFactor: Neural factorization of shape and reflectance under an unknown illumination. ACM Transactions on Graphics (TOG) 40, 1-18 (2021)", + "63. Zhang, Y., Sun, J., He, X., Fu, H., Jia, R., Zhou, X.: Modeling indirect illumination for inverse rendering. In: Computer Vision and Pattern Recognition (CVPR). pp. 18643-18652 (2022)", + "64. Zheng, R., Li, P., Wang, H., Yu, T.: Learning visibility field for detailed 3d human reconstruction and relighting. In: Computer Vision and Pattern Recognition (CVPR). pp. 216-226 (2023)", + "65. Zhi, Y., Qian, S., Yan, X., Gao, S.: Dual-space nerf: Learning animatable avatars and scene lighting in separate spaces. In: International Conference on 3D Vision (3DV) (Sep 2022)", + "66. Zhou, T., He, K., Wu, D., Xu, T., Zhang, Q., Shao, K., Chen, W., Xu, L., Yu, J.: Relightable neural human assets from multi-view gradient illuminations. In: Computer Vision and Pattern Recognition (CVPR). pp. 4315-4327 (2023)" + ], + "bbox": [ + 215, + 146, + 787, + 439 + ], + "page_idx": 18 + }, + { + "type": "header", + "text": "Relightable Neural Actor with Intrinsic Decomposition and Pose Control", + "bbox": [ + 246, + 114, + 730, + 128 + ], + "page_idx": 18 + }, + { + "type": "page_number", + "text": "19", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 18 + } +] \ No newline at end of file diff --git a/2024/Relightable Neural Actor with Intrinsic Decomposition and Pose Control/b67bc924-ef1e-4a87-8446-f01e4ef31ad6_model.json b/2024/Relightable Neural Actor with Intrinsic Decomposition and Pose Control/b67bc924-ef1e-4a87-8446-f01e4ef31ad6_model.json new file mode 100644 index 0000000000000000000000000000000000000000..a1c57ca8dbb87baa4966b0f6c60da151d1166c65 --- /dev/null +++ b/2024/Relightable Neural Actor with Intrinsic Decomposition and Pose Control/b67bc924-ef1e-4a87-8446-f01e4ef31ad6_model.json @@ -0,0 +1,2548 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.277, + 0.141, + 0.728, + 0.187 + ], + "angle": 0, + "content": "Relightable Neural Actor with Intrinsic Decomposition and Pose Control" + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.212, + 0.78, + 0.244 + ], + "angle": 0, + "content": "Diogo Carbonera Luvizon\\(^{1,2}\\), Vladislav Golyanik\\(^{1}\\), Adam Kortylewski\\(^{1,3}\\), Marc Habermann\\(^{1,2}\\), and Christian Theobalt\\(^{1,2}\\)" + }, + { + "type": "text", + "bbox": [ + 0.255, + 0.255, + 0.748, + 0.297 + ], + "angle": 0, + "content": "\\(^{1}\\) Max Planck Institute for Informatics, Saarland Informatics Campus \n\\(^{2}\\) Saarbrücken Research Center for Visual Computing, Interaction and AI \n\\(^{3}\\) University of Freiburg" + }, + { + "type": "text", + "bbox": [ + 0.255, + 0.297, + 0.748, + 0.327 + ], + "angle": 0, + "content": "{dluvizon,golyanik,akortyle,mhaberma,theobalt}@mpi-inf.mpg.de https://vcai.mpi-inf.mpg.de/projects/RNA" + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.359, + 0.744, + 0.665 + ], + "angle": 0, + "content": "Abstract. Creating a controllable and relightable digital avatar from multi-view video with fixed illumination is a very challenging problem since humans are highly articulated, creating pose-dependent appearance effects, and skin as well as clothing require space-varying BRDF modeling. Existing works on creating animatable avatars either do not focus on relighting at all, require controlled illumination setups, or try to recover a relightable avatar from very low cost setups, i.e. a single RGB video, at the cost of severely limited result quality, e.g. shadows not even being modeled. To address this, we propose Relightable Neural Actor, a new video-based method for learning a pose-driven neural human model that can be relighted, allows appearance editing, and models pose-dependent effects such as wrinkles and self-shadows. Importantly, for training, our method solely requires a multi-view recording of the human under a known, but static lighting condition. To tackle this challenging problem, we leverage an implicit geometry representation of the actor with a drivable density field that models pose-dependent deformations and derive a dynamic mapping between 3D and UV spaces, where normal, visibility, and materials are effectively encoded. To evaluate our approach in real-world scenarios, we collect a new dataset with four identities recorded under different light conditions, indoors and outdoors, providing the first benchmark of its kind for human relighting, and demonstrating state-of-the-art relighting results for novel human poses." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.689, + 0.377, + 0.705 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.719, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Creating a relightable and controllable virtual actor of a real person solely from multi-view videos under static illumination is an exciting and challenging topic that has numerous applications in virtual and augmented reality, gaming, human-machine interaction, and content creation. However, it requires modeling the outgoing radiance of a particular surface point on the human, which depends on the pose-dependent and highly non-rigid surface deformation of the human body, the complex and spatially varying materials, as well as the scene lighting. Intrinsically decomposing these quantities solely from multi-view video and" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.401, + 0.127 + ], + "angle": 0, + "content": "D.C. Luvizon et al." + }, + { + "type": "image", + "bbox": [ + 0.216, + 0.143, + 0.482, + 0.276 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.493, + 0.143, + 0.68, + 0.27 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.68, + 0.143, + 0.787, + 0.27 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.286, + 0.788, + 0.385 + ], + "angle": 0, + "content": "Fig. 1: Our method learns a neural actor that is driven only by a 3D skeletal pose and allows rendering and editing of the actor's appearance under new lightning and poses not seen during training. Our approach models pose-dependent deformations, self-shadows, and performs intrinsic decomposition of dynamic humans through a pose-dependent mapping between 3D and UV spaces, which also enables editing the appearance and material properties at inference time. For training, our method only needs a multi-view video and an environment map." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.415, + 0.785, + 0.445 + ], + "angle": 0, + "content": "enabling novel compositions that are of highly photorealistic quality defines an unsolved and challenging inverse problem in the literature." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.447, + 0.789, + 0.688 + ], + "angle": 0, + "content": "Relighting static scenes from multi-view imagery is a well-studied problem [22,33,49,62]. However, directly applying such methods to moving humans, which are dynamic, highly articulated, and have complex material composition, is not trivially possible, as these approaches do not account for dynamics. On the contrary, some works have focused on building drivable and dynamic full-body avatars from multi-view video [32,42,54] without considering relightability at all. Despite some works [4,11,18,36,66] assuming a well-calibrated lightstage settings, only few human-specific works [6,17,64] try to achieve human intrinsic decomposition and relighting from videos with static lights. While they are able to recover a relightable human representation from a single RGB video, they can solely replay the human motion under novel lighting conditions [6], or do not model self-shadows and space-varying BRDF [17], which results in low visual fidelity. In stark contrast, we assume a multi-view setting while accounting for high-frequency surface deformations as well as induced high-frequency lighting effects such as self-shadows and shading, which are indispensable for a truly immersive experience." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.792, + 0.779 + ], + "angle": 0, + "content": "To address this, we propose Relightable Neural Actor, a new method that achieves intrinsic decomposition of dynamic human avatars, enabling pose-control, material editing, and realistic rendering with pose-dependent effects, while solely requiring a multi-view video of the real human under a single and static light condition during training (see Fig. 1). At inference time, we only need an arbitrary driving skeletal pose and an environment map." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.841 + ], + "angle": 0, + "content": "We solve this challenging problem by leveraging an implicit density field from a pose-driven neural actor and by proposing a novel intrinsic decomposition guided by the implicit density field. This strategy allows our method to model pose-dependent deformations in the geometry as dynamic maps, while" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.247, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Relightable Neural Actor with Intrinsic Decomposition and Pose Control" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.328 + ], + "angle": 0, + "content": "representing the albedo and roughness properties of the actor as static maps. Our approach not only makes the problem computationally tractable for the long and dynamic human sequences, but also enables editing the neural actor's appearance and materials at inference time. Our key contribution is to sample normal and visibility from the posed density field and aggregate them into sparse UV maps, which are then densified by the proposed NormalNet and VisibilityNet models. This allows our approach to effectively model wrinkles and self-shadows even for long sequences. The albedo and roughness maps are optimized during training and sampled by the proposed UVDeltaNet, which accounts for pose-dependent deformations in the UV map and is intrinsically learned in our pipeline. Finally, given an environment map and a virtual camera position, our neural renderer produces the final image of the neural actor." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.33, + 0.788, + 0.405 + ], + "angle": 0, + "content": "To train and evaluate our method, we collected a real-world dataset for novel pose and novel lighting human synthesis, composed of four actors recorded under six different light conditions by a multi-view camera setup, including indoor and outdoor environments and a light probe for each sequence, providing the first real in-the-wild benchmark for controllable human relighting." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.406, + 0.788, + 0.57 + ], + "angle": 0, + "content": "To summarize, our main technical contributions are as follows: 1) The first method to learn a relightable neural actor from multi-view video under static lights that allows for pose control, appearance editing, and free-viewpoint rendering under arbitrary illumination while modeling pose-dependent effects. 2) We propose an intrinsic decomposition of diffuse and specular reflection based on a new approach to compute pose-dependent normals and visibility by sparsely sampling a neural field and densifying the results with NormalNet and VisibilityNet, making the problem of pose-driven human relighting tractable. Our approach relies on a mapping between 3D and UV space implicitly learned by UVDeltaNet, allowing us to learn an explicit and static material representation that is easy to edit at inference time." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.572, + 0.788, + 0.634 + ], + "angle": 0, + "content": "In our experimental evaluation, we compare our approach with two baselines in the proposed real data benchmark, showing the superiority of our approach in synthesizing relightable humans on challenging motion sequences. The effectiveness of our method is demonstrated by ablation studies and qualitative results." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.659, + 0.389, + 0.675 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.695, + 0.788, + 0.743 + ], + "angle": 0, + "content": "Since the goal of this work is to build a full-body relightable and animatable avatar, we will not discuss works [9, 10, 28, 44, 46, 47, 50, 56, 58, 59] focusing on face relighting as both settings have different requirements and challenges." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.766, + 0.465, + 0.782 + ], + "angle": 0, + "content": "2.1 Relighting Static Scenes" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.788, + 0.843 + ], + "angle": 0, + "content": "Classic approaches for relighting and editing perform intrinsic decomposition through inverse rendering [43]. With recent advances in neural rendering [51], light transportation can be taken into account with fully differentiable rendering" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.401, + 0.128 + ], + "angle": 0, + "content": "D.C. Luvizon et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.359 + ], + "angle": 0, + "content": "pipelines. For instance, Zhang et al. [62] address the problem of recovering geometry and spatially-varying BRDF with an implicit scene representation. However, this requires solving the light transport by evaluating the NeRF model for each light ray, which is very inefficient. NeRV [49] and Zhang et al. [63] mitigate this problem with a neural visibility field that is jointly optimized with the NeRF model, and NeRD [5] adds further relaxation by ignoring cast shadows. However, this assumption cannot be made for humans, which are highly articulated and often experience prominent self-casting shadows. Jin et al. [22] propose TensorIR, an efficient tensor-based representation that solves the inverse problem by casting camera and secondary rays. IRON [60] proposes a two-stage process for obtaining a mesh with material properties, and Lyu et al. [35] and NeRO [33] focus on multi-bounce illumination and reflective objects, which is usually not the case for humans. The methods mentioned above share the assumption of a static scene and cannot be trivially adapted to dynamic and controllable humans." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.38, + 0.544, + 0.394 + ], + "angle": 0, + "content": "2.2 Human Models with Pose Control" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.403, + 0.787, + 0.509 + ], + "angle": 0, + "content": "Modeling humans is a long-standing topic. Traditionally, meshes from an actor can be automatically rigged to a skeleton [7] using Linear Blending Skinning (LBS) [27]. Parametric human models offer pose and shape control [34,41] and, more recently, implicit human bodies can be articulated [3, 37, 38]. Such models offer a generic and controllable representation, but are unable to represent individual details. Template-based methods synthesize a realistic level of details [14, 15, 21], but have to deal with non-linear pose-dependent deformations." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.509, + 0.788, + 0.706 + ], + "angle": 0, + "content": "Several approaches model appearance and geometry with an implicit field anchored on a coarse human mesh. Neural Body [42] optimizes latent vectors anchored in the vertices of the human mesh and Neural Actor [32] incorporates texture maps as a conditioning to break down the mapping from pose to dynamic effects (one-to-many mapping). ARAH [54] models the human surface with an SDF, Dual-Space NeRF [65] learns the color and light in separate spaces, and Kwon et al. [24] propose an approach that generalizes to new identities and poses. Shyshey et al. [48] jointly learn textures and a neural renderer but fall short in reconstructing the 3D human body. A few works [20,57] approach a restrictive scenario of reconstructing 3D humans from a monocular video but fail to model fine details under new poses. Common to all these type of methods [12,13,25, 40] is the entanglement of material properties and illumination, making them restricted to the same appearance as observed during training." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.726, + 0.482, + 0.741 + ], + "angle": 0, + "content": "2.3 Relighting Human Models" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Only a few methods in the literature tackle the problem of relighting humans from videos under static illumination. Li et al. [29] create a relightable human mesh by estimating a low-frequency environment using Spherical Harmonics. Some approaches target the problem of single-image [19] and static pose [64] relighting under unknown lights, which is an extremely ill-posed problem due to the high ambiguity between albedo and lighting. Relighting4D [6] extends" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.247, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Relightable Neural Actor with Intrinsic Decomposition and Pose Control" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.283 + ], + "angle": 0, + "content": "Neural Body [42] to estimate geometry and reflectance, but can only replay the same sequence observed during training under new lights and is not able to generate new poses. RANA [17] allows for pose control but assumes Lambertian reflectance, ignoring different material properties such as skin and clothes, and does not model cast shadows, resulting in unrealistic renderings for articulated humans. In contrast, our method performs intrinsic decomposition with space-varying BRDF, models illumination and self-shadows, allows for pose control and novel lights, and offers an easy manner to edit the material properties of the learned neural actor." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.306, + 0.331, + 0.323 + ], + "angle": 0, + "content": "3 Method" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.339, + 0.788, + 0.475 + ], + "angle": 0, + "content": "Our goal is to learn a digital human avatar (neural actor) from RGB images that models pose-dependent effects and can be relighted, pose-controlled, and rendered from novel viewpoints at test time. To achieve relighting, we decompose the observed appearance into normals, visibility with respect to each light source, and BRDF represented as albedo and roughness. With these components and an environment map, we employ the microfacet model [52] to render the output image of our method, which is compared to the observed images. In addition—since we target a drivable actor—the geometry components need to be controllable by the skeletal pose." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.476, + 0.789, + 0.747 + ], + "angle": 0, + "content": "Fig. 2 shows an overview of our method divided into three main components. The first is a pose-driven implicit geometry model, where the goal is to obtain a skeleton-controllable implicit 3D density field that models the geometry of the actor with pose-dependent deformations. This model is adapted from Neural Actor [32] and is depicted here since our approach builds upon its internal design. Note that Neural Actor does not perform intrinsic decomposition and is unable to perform relighting or appearance editing, which is our main goal. The second component is our proposed intrinsic decomposition approach whose goal is to disentangle geometry represented as normal and visibility information, and material properties as albedo and roughness. Our key insight is to represent these components in UV space rather than in 3D coordinates, using the density field as guidance in the mapping function between 3D and UV spaces. This allows our method to model complex deformations that are pose-dependent in a compact UV map that can be easily computed and queried. Finally, we integrate the microfacet model [52] into our rendering pipeline that takes as input the disentangled components (i.e., normal, visibility, albedo, and roughness) and an environment map, and renders the image of the neural actor in an arbitrary pose and light condition from a free-viewpoint." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.769, + 0.373, + 0.784 + ], + "angle": 0, + "content": "3.1 Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.787, + 0.842 + ], + "angle": 0, + "content": "Human Pose and Environment Map. Our method takes as input a 3D articulated human skeleton pose \\(\\mathbf{P}_t = \\{\\pmb {\\theta}_t,\\mathbf{R}_t,\\mathbf{T}_t\\}\\), defined by the relative joint rotation \\(\\pmb {\\theta}\\in \\mathbb{R}^{J\\times 3}\\) of \\(J\\) body joints, the global rotation \\(\\mathbf{R}\\in \\mathbb{R}^{3\\times 3}\\), and global" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.401, + 0.128 + ], + "angle": 0, + "content": "D.C. Luvizon et al." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.146, + 0.788, + 0.3 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.324, + 0.788, + 0.422 + ], + "angle": 0, + "content": "Fig. 2: Our method takes as input a 3D skeletal pose and a static environment map, and renders the neural actor from a virtual camera position. The pose-driven geometry (Sec. 3.1) learns an implicit density function, from which we compute the normal and visibility information. Our intrinsic decomposition disentangles normal, visibility, albedo and roughness maps in UV space (Sec. 3.2). The neural renderer (Sec. 3.3) outputs our prediction, which is supervised with the reference image. Yellow represents learnable components and green non-learnable components." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.45, + 0.788, + 0.542 + ], + "angle": 0, + "content": "translation \\(\\mathbf{T} \\in \\mathbb{R}^3\\) of the person at frame \\(t\\). The 3D pose can be used to drive a coarse human mesh with linear blending skinning (LBS) [34]. We represent the global lights as an environment map \\(\\mathbf{L} \\in \\mathbb{R}^{16 \\times 32 \\times 3}\\), which can be obtained from HDR images of a mirror sphere [8]—see the supplement for more details. The environment map can be static during training and changed during inference to render results under novel illumination." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.562, + 0.788, + 0.701 + ], + "angle": 0, + "content": "Training Data. During training, our method is supervised with multi-view calibrated video streams \\(\\mathbf{V}_c = \\{\\mathbf{I}_{c,t}\\}\\), where \\(c \\in \\{1, \\dots, C\\}\\) is the camera index and \\(\\mathbf{I}_{c,t}\\) is the image from camera \\(c\\) at frame \\(t\\). We obtain a foreground mask \\(\\mathbf{M}_{c,t}\\) for each video frame [30]. The coarse human mesh and 3D skeletons can be obtained with multi-view pose tracking [1]. The input 3D skeletons drive the LBS model to obtain posed human meshes used to extract texture maps from the multi-view images [2]. These texture maps are used during training to learn pose-dependent deformations on the implicit human model. At inference time, a V2V [55] network predicts texture maps from posed normal maps [32]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Implicit Human Model. For realistically relighting articulated avatars, we need to model human geometry with deformations caused by dynamic and varied poses. Therefore, we follow Neural Actor [32] to represent the pose-driven geometry as an implicit field guided by the 3D skeletal pose parameters and conditioned by texture maps, which provide the ability to model high-frequency deformations. Specifically, a standard NeRF model [39] is optimized through volumetric rendering in the canonical space of the coarse human mesh and the canonical space is deformed based on ResNet [16] features extracted from the" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.247, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "Relightable Neural Actor with Intrinsic Decomposition and Pose Control" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.269 + ], + "angle": 0, + "content": "texture maps, which provide the ability to represent detailed wrinkles and deformations not modeled by the coarse mesh. During optimization, from each pixel in the reference images we cast a ray \\(\\mathbf{r}_n = \\mathbf{o} + \\delta_n\\mathbf{d}\\), where \\(n = 1,\\dots N\\) is the index of each point in the ray, \\(\\mathbf{o} \\in \\mathbb{R}^3\\) is the origin defined in the optical center of the camera, \\(\\mathbf{d} \\in \\mathbb{R}^3\\) is the ray direction, and \\(\\delta_{n}\\) is the depth (distance to the camera). The sampled points \\(\\mathbf{r}_n\\) are projected to the mesh surface, resulting in the local coordinates \\((u_{n},v_{n},h_{n})\\), where \\((u,v)\\) is the surface coordinate in the UV space and \\(h\\) is the signed distance to the mesh." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.295, + 0.523, + 0.312 + ], + "angle": 0, + "content": "3.2 Neural Intrinsic Decomposition" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.326, + 0.789, + 0.492 + ], + "angle": 0, + "content": "At the core of our method, we perform intrinsic decomposition to obtain the normal, visibility, albedo, and roughness components, required for rendering the neural actor under new light conditions. Specifically, for each point in each ray cast into the scene we want to obtain the normal direction \\(\\mathbf{n} \\in \\mathbb{R}^3\\), the visibility coefficients \\(\\pmb{v} \\in \\mathbb{R}^{512}\\) respective to each light source in the environment map, the RGB albedo \\(\\mathbf{a} \\in \\mathbb{R}^3\\), and the scalar roughness \\(\\rho\\). The normal and visibility values can be computed from the geometry model. However, computing the visibility of each point in 3D with respect to the light sources is computationally intractable for a dynamic and long sequence, since for every point in 3D and for every frame, several rays have to be cast towards each pixel of the environment map. Therefore, we propose a different approach." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.493, + 0.789, + 0.659 + ], + "angle": 0, + "content": "We notice that the intrinsic decomposition is only relevant for points close to the surface. Since the coarse human mesh provides a mapping from the mesh surface to UV space and the implicit human model provides a density field that models the geometry of the actor, we propose an aggregation strategy that samples normal and visibility from a given point of view (therefore, only a few points are sampled in 3D) and aggregates the obtained values in a per-frame UV map. Once we obtain the normal and visibility in UV space for a given pose, we can simply query the values by projecting the sampled 3D point \\(\\mathbf{r}_n\\) into the UV space, preventing our method from the expensive computation of normal and visibility at every possible 3D coordinate over multiple frames. We next explain this approach in more detail." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.687, + 0.789, + 0.793 + ], + "angle": 0, + "content": "Normal Maps. For each point in a ray, we want to obtain the normal value that will be required for solving the rendering equation during relighting (will be introduced in Sec. 3.3). This value could be obtained by computing the gradient from the neural field [62]. However, this process requires back-propagating gradients throughout the entire network and often results in noisy normals. Instead, we explicitly compute the normal vector close to the visible surface by deriving the unbiased depth \\(\\hat{\\mathbf{D}}\\) [53]:" + }, + { + "type": "equation", + "bbox": [ + 0.299, + 0.806, + 0.787, + 0.845 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathbf {D}} (\\mathbf {r}) = \\frac {\\sum_ {n = 1} ^ {N} w _ {n} \\delta_ {n}}{\\sum_ {n = 1} ^ {N} w _ {n}}, \\quad \\text {w i t h} w _ {n} = T _ {n} \\left(1 - \\exp \\left(- \\sigma_ {n} \\bar {\\delta} _ {n}\\right)\\right), \\tag {1}\n\\]" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.402, + 0.128 + ], + "angle": 0, + "content": "D.C. Luvizon et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.145, + 0.784, + 0.183 + ], + "angle": 0, + "content": "where \\( T_{n} = \\exp \\left(-\\sum_{j = 1}^{n - 1}\\sigma_{j}\\bar{\\delta}_{j}\\right) \\), \\( w_{n} \\) is the density weight from volumetric rendering [39], and \\( \\sigma_{n} \\) is the density predicted by the NeRF model at point \\( \\mathbf{r}_n \\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.183, + 0.487, + 0.332 + ], + "angle": 0, + "content": "Then, as illustrated in Fig. 3, we convert the depth values relative to the camera to 3D coordinates and obtain the normal direction from the tangent surface computed at the intersection point, defined as the maximum value of the density weight \\( w_{n} \\). This process can be performed for multiple viewpoints to cover the surface of the human." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.334, + 0.487, + 0.379 + ], + "angle": 0, + "content": "Once the normal values are obtained in 3D, we project them to the UV map and aggregate them with the" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.379, + 0.785, + 0.41 + ], + "angle": 0, + "content": "obtained normals from different views. In this process, the \\((u^{*},v^{*})\\) coordinate where the obtained normal will be projected onto the UV map is defined by" + }, + { + "type": "image", + "bbox": [ + 0.5, + 0.194, + 0.621, + 0.271 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.622, + 0.195, + 0.697, + 0.269 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.698, + 0.195, + 0.78, + 0.27 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.497, + 0.281, + 0.788, + 0.365 + ], + "angle": 0, + "content": "Fig. 3: The unbiased depth from Eq. (1) is projected into 3D coordinates, where we approximate the tangent surface at the intersecting point and obtain the normal vector. Normals are sampled from multiple viewpoints and aggregated in the UV map." + }, + { + "type": "equation", + "bbox": [ + 0.35, + 0.42, + 0.785, + 0.446 + ], + "angle": 0, + "content": "\\[\n\\left(u ^ {*}, v ^ {*}\\right) = \\Psi \\left(\\mathbf {r} _ {n} ^ {*}\\right), \\quad \\text {w i t h} \\mathbf {r} _ {n} ^ {*} = \\underset {\\mathbf {r} _ {n}} {\\arg \\max } w _ {n}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.456, + 0.784, + 0.531 + ], + "angle": 0, + "content": "where \\(\\varPsi\\) is a function that projects a 3D point to the closest surface point in the human mesh. Note that \\(w_{n}\\) depends on the sampled points \\(\\mathbf{r}_n\\) (from \\(T_{n}\\) in Eq. 1) and \\((u^{*},v^{*})\\) corresponds to the \\((u_{n},v_{n})\\) value obtained from the point \\(\\mathbf{r}_n\\) on the ray that has the largest density weight \\(w_{n}\\). This process ensures that the normal map obtained aggregates the normals most likely close to the surface." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.532, + 0.785, + 0.622 + ], + "angle": 0, + "content": "In this process, the resulting UV map can be sparse, which is undesirable for free-viewpoint rendering. Thus, the NormalNet is used to inpaint the missing normals in UV space. This lightweight CNN is implemented with 8 layers of partial convolutions [31], takes as input sparse normal maps with size \\(512 \\times 512 \\times 3\\), and predicts a dense normal map (see Fig. 2). During training, NormalNet is supervised with densely sampled normal maps from the training data." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.644, + 0.793, + 0.842 + ], + "angle": 0, + "content": "Visibility Maps. The visibility information is necessary to model pose-dependent shadows in the neural avatar, without which it is not possible to produce realistic relighting results. A naive approach for obtaining the visibility for each point in 3D would require querying the neural field thousands of times while casting rays from the intersecting point to all the light sources, which is prohibitive during training. In contrast, in our approach we efficiently encode visibility of each point in 3D close to the surface by applying an approximate mapping to UV space. Specifically, we pre-compute the visibility maps for each pose by sampling multiple camera views and aggregate the results in UV map following Eq. (2). The resulting visibility maps with size \\(256 \\times 256 \\times 512\\) encode the visibility information w.r.t. each light source in the environment map (represented by 512 individual lights). Similarly to the computation of normal maps, this process results in sparse visibility maps, which are inpainted by the VisibilityNet (as" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.247, + 0.115, + 0.733, + 0.13 + ], + "angle": 0, + "content": "Relightable Neural Actor with Intrinsic Decomposition and Pose Control" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.207 + ], + "angle": 0, + "content": "shown in Fig. 2). This shallow CNN takes as input sparse visibility and normal maps, as the visibility highly depends on normals (e.g. the visibility is always zero for the backlight source), and is trained with densely sampled visibility from the training sequence. See the supplement for architecture details." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.225, + 0.785, + 0.36 + ], + "angle": 0, + "content": "Albedo and Roughness Maps. As discussed above, the normal and visibility components can be computed from the density field by using UV maps as proxies. However, the albedo and roughness properties cannot be directly obtained from the pre-trained Neural Actor model, since it predicts radiance which is the product of the rendering equation. Instead, we want to learn a decomposed representation. In intrinsic decomposition, the material properties are commonly optimized via differentiable rendering. Hence, we optimize the albedo and roughness of the neural actor by leveraging a differentiable renderer (Sec. 3.3) and by imposing priors through our design." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.361, + 0.785, + 0.445 + ], + "angle": 0, + "content": "The material properties of an actor should not depend on the pose and should be efficiently encoded to enable fast computations and editing. Therefore, we model albedo and roughness as static UV maps, respectively defined as \\(\\mathbf{A} \\in \\mathbb{R}^{1024 \\times 1024 \\times 3}\\) and \\(\\mathbf{B} \\in \\mathbb{R}^{1024 \\times 1024}\\). These tensors are jointly optimized for the whole training sequence and regularized with an L1 loss on the spatial gradients, which encourages smoothness in both maps:" + }, + { + "type": "equation", + "bbox": [ + 0.31, + 0.45, + 0.785, + 0.474 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {s m o o t h}} = \\frac {1}{N} \\sum \\left| \\nabla_ {x} (\\mathbf {A}) \\right| + \\left| \\nabla_ {y} (\\mathbf {A}) \\right| + \\left| \\nabla_ {x} (\\mathbf {B}) \\right| + \\left| \\nabla_ {y} (\\mathbf {B}) \\right|, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.22, + 0.478, + 0.644, + 0.492 + ], + "angle": 0, + "content": "where \\(\\nabla_{x}\\) and \\(\\nabla_{y}\\) represent the spatial gradient operators." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.493, + 0.786, + 0.643 + ], + "angle": 0, + "content": "Nonetheless, by assuming a material representation that does not depend on the skeletal pose, we still have to cope with intrinsic mesh registration errors. Such registration errors could cause blurry results or artifacts in the final renderings. To cope with this problem, we propose a corrective term to the UV mapping, which is predicted by the UVDeltaNet, as shown in Fig. 2. The UVDeltaNet is an MLP that takes as input the local coordinates \\((u,v,h)\\) for each point on the ray and the sampled texture features \\(\\psi\\), and outputs a corrective term \\(\\epsilon = (u',v')\\). The texture features are necessary to provide additional information, since the local coordinates alone cannot account for deviations in the mesh registration." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.644, + 0.784, + 0.688 + ], + "angle": 0, + "content": "Initially, UVDeltaNet is frozen and outputs zeros. After the initial convergence of \\(\\mathbf{A}\\) and \\(\\mathbf{B}\\), UVDeltaNet is indirectly supervised by the rendering loss and regularized with:" + }, + { + "type": "equation", + "bbox": [ + 0.373, + 0.695, + 0.785, + 0.723 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {u v}} = \\frac {1}{N} \\sum \\epsilon^ {2} + | \\nabla_ {x} (\\epsilon) | + | \\nabla_ {y} (\\epsilon) |, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.727, + 0.784, + 0.757 + ], + "angle": 0, + "content": "which minimizes the magnitude of the deformation and encourages its spatial gradient to be sparse, i.e., it makes the UV deformation spatially coherent." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.775, + 0.402, + 0.789 + ], + "angle": 0, + "content": "3.3 Neural Renderer" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.784, + 0.84 + ], + "angle": 0, + "content": "The third part of our method is our neural renderer, which takes the disentangled components as normal, visibility, albedo, and roughness; integrates the lights from the environment map, and outputs the rendered image." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.401, + 0.128 + ], + "angle": 0, + "content": "D.C. Luvizon et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.253 + ], + "angle": 0, + "content": "Our neural renderer starts by recasting the ray \\(\\mathbf{r}\\) with importance sampling around the estimated depth value \\(\\hat{\\mathbf{D}}\\). Thus, all the points \\(\\mathbf{r}_n\\) are most likely close to the surface. For each sampled point in the ray, we obtain the normal direction \\(\\mathbf{n}_n\\) and the visibility values \\(\\pmb{v}_n\\) by interpolating the normal and visibility maps in UV around the point \\((u_n,v_n) = \\varPsi(\\mathbf{r}_n)\\). The albedo \\(\\mathbf{a}_n\\) and roughness \\(\\rho_n\\) are obtained by interpolating the albedo and roughness maps in UV around the point \\((u_n + u'_u,v_n + v'_n)\\), i.e., following the deformation predicted by UVDeltaNet." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.253, + 0.787, + 0.283 + ], + "angle": 0, + "content": "Before rendering the final image of the posed neural actor, we consider the light transport equation defined as" + }, + { + "type": "equation", + "bbox": [ + 0.32, + 0.295, + 0.786, + 0.325 + ], + "angle": 0, + "content": "\\[\n\\mathbf {L} _ {0} (\\mathbf {r} _ {n}, \\boldsymbol {\\omega} _ {0}) = \\sum_ {i} \\operatorname {R} (\\mathbf {a} _ {n}, \\rho_ {n}, \\boldsymbol {\\omega} _ {i}, \\boldsymbol {\\omega} _ {0}) \\boldsymbol {v} _ {n, i} \\mathbf {L} (\\boldsymbol {\\omega} _ {i}) (\\boldsymbol {\\omega} _ {i} \\mathbf {n} _ {n}), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.335, + 0.788, + 0.41 + ], + "angle": 0, + "content": "where \\(\\mathrm{R}(\\cdot)\\) is the microfacet BRDF model [52], \\(\\omega_0\\) and \\(\\omega_{i}\\) are the camera and the light source directions, respectively, \\(\\pmb{v}_{n,i}\\) is the visibility of point \\(\\mathbf{r}_n\\) w.r.t. the ith light source, and \\(\\mathbf{L}(\\boldsymbol {\\omega}_i)\\) is the incoming light from the environment map. Eq. (5) is defined for a single point \\(\\mathbf{r}_n\\) on the ray and is integrated into our volume rendering as" + }, + { + "type": "equation", + "bbox": [ + 0.414, + 0.411, + 0.786, + 0.451 + ], + "angle": 0, + "content": "\\[\n\\tilde {\\mathbf {I}} (\\mathbf {r}) = \\sum_ {n = 1} ^ {N} w _ {n} \\mathbf {L} _ {0} \\left(\\mathbf {r} _ {n}, \\omega_ {0}\\right), \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.458, + 0.63, + 0.473 + ], + "angle": 0, + "content": "where \\(\\tilde{\\mathbf{I}}\\) is the output image of the relighted neural actor." + }, + { + "type": "text", + "bbox": [ + 0.239, + 0.474, + 0.78, + 0.489 + ], + "angle": 0, + "content": "During training, our relightable model is supervised with the following loss:" + }, + { + "type": "equation", + "bbox": [ + 0.312, + 0.502, + 0.786, + 0.518 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {r e l i g h t}} = \\lambda_ {\\mathrm {L 2}} \\mathcal {L} _ {\\mathrm {L 2}} + \\lambda_ {\\text {v g g}} \\mathcal {L} _ {\\text {v g g}} + \\lambda_ {\\text {s m o o t h}} \\mathcal {L} _ {\\text {s m o o t h}} + \\mathcal {L} _ {\\text {u v}}, \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.529, + 0.788, + 0.575 + ], + "angle": 0, + "content": "where \\(\\mathcal{L}_{\\mathrm{L2}}\\) and \\(\\mathcal{L}_{\\mathrm{vgg}}\\) are the L2 and the perceptual losses [26] between \\(\\tilde{\\mathbf{I}} (\\mathbf{r})\\) and the reference image. The hyperparameters are defined as \\(\\lambda_{\\mathrm{L2}} = 100\\), \\(\\lambda_{\\mathrm{vgg}} = 0.01\\), and \\(\\lambda_{\\mathrm{smooth}} = 0.1\\) in all our experiments." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.597, + 0.607, + 0.615 + ], + "angle": 0, + "content": "4 Relightable Dynamic Actors Dataset" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.629, + 0.788, + 0.719 + ], + "angle": 0, + "content": "Existing datasets for human reenactment and relighting are often monocular [2, 17], contain very short sequences with only a few pose variations, and do not provide environment maps [2,17,32,42]. Hence, we cannot use them for evaluation under realistic scenarios with real ground truth. In addition, we observe that the only existing synthetic dataset for human relighting and pose control [17] suffers from artifacts in the geometry and salt-and-pepper noise in the renderings." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.788, + 0.84 + ], + "angle": 0, + "content": "With these shortcomings in mind, we collect a new real dataset, which we call Relightable Dynamic Actors, with four identities performing a series of ten different actions (Fig. 4). Each subject is recorded six times wearing the same clothes but under different light conditions, including indoor and outdoor, where we used around 100 and eight cameras, respectively. For each sequence, we also capture HDR light probes, which can be used for training or testing. In total, our dataset contains around 90k multi-view video frames divided into 24 multi-view video sequences. We refer to the supplemental document for more details." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.248, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Relightable Neural Actor with Intrinsic Decomposition and Pose Control" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "image", + "bbox": [ + 0.217, + 0.144, + 0.5, + 0.271 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.144, + 0.785, + 0.271 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.293, + 0.788, + 0.337 + ], + "angle": 0, + "content": "Fig. 4: Relightable Dynamic Actors dataset. For each actor, we recorded six sequences under different light conditions, including five indoor sequences and one outdoor. Environment maps were obtained from HDR pictures of a mirror sphere." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.365, + 0.376, + 0.383 + ], + "angle": 0, + "content": "5 Experiments" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.398, + 0.784, + 0.445 + ], + "angle": 0, + "content": "In this section, we provide an experimental evaluation of our method on the Reli-ightable Dynamic Actors dataset and show qualitative and quantitative results. Please refer to the supplementary video for dynamic results on long sequences." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.467, + 0.352, + 0.48 + ], + "angle": 0, + "content": "5.1 Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.493, + 0.785, + 0.568 + ], + "angle": 0, + "content": "Metrics. For quantitative evaluation, we report results on the PSNR, SSIM and LPIPS metrics [61], comparing the rendered relight results with the groundtruth images from our dataset. Importantly, our work is the first to be able to report such metrics on a real dataset since no previous dataset in the literature provided a real human recorded under different and known light conditions." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.569, + 0.788, + 0.765 + ], + "angle": 0, + "content": "Baselines. Since the only method in the literature capable of reposing a human under new light conditions does not provide implementation [17], and Relighting4D [6] is not able to perform inference for new poses, we defined two baselines for a fair comparison. Our baselines are derived from Neural Actor [32] and adapted to perform intrinsic decomposition based on the obtained normal and visibility information, as discussed in Sec. 3.2. However, instead of using NormalNet and VisibilityNet to obtain these components, we densely sample the neural field and perform morphological inpainting to obtain dense UV maps. The first baseline, referred to as Neural Actor \\(+IR\\), predicts the albedo and roughness components instead of RGB color directly from the deformed canonical space \\(\\mathbf{x}_n = (x,y,z)\\) (see Fig. 2). The second baseline, Neural Actor \\(+IR + Tex\\), takes as input the texture features \\(\\psi\\) as a conditioning for predicting the albedo and roughness values. In both cases, the predictor is a NeRF MLP with 8 layers." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.766, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Implementation Details. Our model is trained in a three-stage process. First, we train the geometry component (Sec. 3.1) with the Adam [23] optimizer with a learning rate of 0.001 until convergence, which takes about two days on four NVidia A40 GPUs. Then, we pre-train the NormalNet and VisibilityNet with pre-computed normal and visibility UV maps sampled from the geometry model" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.401, + 0.127 + ], + "angle": 0, + "content": "D.C. Luvizon et al." + }, + { + "type": "image", + "bbox": [ + 0.292, + 0.145, + 0.713, + 0.286 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.288, + 0.785, + 0.329 + ], + "angle": 0, + "content": "Fig. 5: Qualitative results of our method on novel poses. Our approach generalizes to both indoor and outdoor illumination, and generates realistic renderings for arbitrary environment maps with detailed shadows and pose-dependent deformations." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.357, + 0.785, + 0.478 + ], + "angle": 0, + "content": "from 40 different views. Both networks are trained in parallel using Adam with a learning rate of 0.01 for about 12 hours. Finally, we train our intrinsic decomposition model with the same optimizer until convergence, which takes about one day. Since the obtained environment maps are up to factor of scale [8], we apply global color correction during training and evaluation. The predicted texture maps are obtained with a V2V model [55] trained (in parallel to our model) to translate a sequence of SMPL normal maps into a sequence of texture maps. Please refer to our supplementary materials for architecture details." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.498, + 0.323, + 0.512 + ], + "angle": 0, + "content": "5.2 Results" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.522, + 0.785, + 0.627 + ], + "angle": 0, + "content": "Qualitative Results. In Fig. 5, we show qualitative results of our method. Our results are all from a trained model driven by a new skeletal pose, new camera viewpoint, and new environment map not seen during training. Note how our approach produces realistic appearance and prominent cloth wrinkles, which can be easily observed under different light conditions. Importantly, notice how our results look realistic under novel lights, even though our model is trained on a single sequence under a static light condition." + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.652, + 0.785, + 0.695 + ], + "angle": 0, + "content": "Table 1: Quantitative results from our method on real data, subject S1, considering indoor and outdoor scenes. Our method consistently improves the baselines, especially in outdoor scenarios, where the gap between training and test lights is more pronounced." + }, + { + "type": "table", + "bbox": [ + 0.216, + 0.707, + 0.787, + 0.761 + ], + "angle": 0, + "content": "
MethodT2T3T4T5T6 (outdoors)
PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓
Neural Actor+IR17.5220.7500.18117.4730.7680.17817.7450.7670.17118.2440.7680.17518.3550.7100.203
Neural Actor+IR+Tex18.1750.7780.16417.7990.7900.16718.1320.7850.16718.5310.7890.16618.7040.7370.186
Ours18.6020.7920.16318.2400.8000.16918.3880.8000.16418.8200.8000.16519.4610.7530.173
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.777, + 0.785, + 0.837 + ], + "angle": 0, + "content": "In Fig. 6, we show results from our intrinsic decomposition. Note that our normal representation nicely captures the geometry details and the visibility. This is important to obtain an albedo that does not contain baked-in lighting effects, e.g. shadows, as can be observed from our results." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.248, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Relightable Neural Actor with Intrinsic Decomposition and Pose Control" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "image", + "bbox": [ + 0.288, + 0.143, + 0.717, + 0.334 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.334, + 0.785, + 0.377 + ], + "angle": 0, + "content": "Fig. 6: Results from our intrinsic decomposition. Our method provides detailed normals and visibility (averaged for visualization), and plausible albedo and roughness predictions that enable realistic relighting." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.412, + 0.544, + 0.517 + ], + "angle": 0, + "content": "Another important and unique characteristic of our method is the ability to model self-shadows and cloth wrinkles. In Fig. 7, we show results from two subjects, where a \"sunrise\" light probe was used to produce cast shadows. Note how pose-dependent effects are consistent with the mostly directional illumination." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.52, + 0.544, + 0.671 + ], + "angle": 0, + "content": "Comparisons. We evaluated our method on the test sequences from our dataset. Specifically, we compare against two baselines in the indoor sequences (T2-T5) and outdoor sequence (T6) of subject S1 in every 10th frame, which sums up to 4000 evaluation frames. In Tab. 1 we show, for the first time, quantitative results of real human relighting under new poses and new light conditions. Our approach consistently improves upon the two baselines." + }, + { + "type": "image", + "bbox": [ + 0.563, + 0.384, + 0.784, + 0.595 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.553, + 0.61, + 0.788, + 0.667 + ], + "angle": 0, + "content": "Fig. 7: Our method models pose-dependent effects and cast shadows, as can be observed from our results under the \"sunrise\" light." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Ablation. In Tab. 2, we show an ablation with different components of our method. Compared to our baseline (Neural Actor+IR), learning the albedo and roughness as UV maps guided by the neural field significantly improves the performance on relighting and reposing. This is because, in the baselines, the albedo and roughness are under-constrained and can possibly embed shading effects, overfitting to the training data. The NormalNet and VisibilityNet provide additional capacity to the model to learn how to inpaint the sampled UV maps. Note that without these two models our results are computed using morphological region growth in the UV maps. Finally, the UVDeltaNet provides more flexibility to the albedo and roughness representation, accounting for fine details. This can be observed in the improvement in the LPIPS metric in the last" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.401, + 0.128 + ], + "angle": 0, + "content": "D.C. Luvizon et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.178 + ], + "angle": 0, + "content": "row of Tab. 2. Nonetheless, this also means more capacity to overfit the training sequences, which translates into a small decrease in the PSNR metric." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.201, + 0.367, + 0.216 + ], + "angle": 0, + "content": "5.3 Applications" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.227, + 0.432, + 0.392 + ], + "angle": 0, + "content": "An important characteristic of our approach is our static representation in UV for the albedo and roughness. This allows exciting and new applications such as color and material editing of an animatable and relightable neural actor, as shown in Fig. 8 and in Fig. 1. With a simple edit in the obtained albedo or rough-" + }, + { + "type": "text", + "bbox": [ + 0.441, + 0.196, + 0.788, + 0.306 + ], + "angle": 0, + "content": "Table 2: Ablation study with different components of our method. The proposed intrinsic decomposition with static albedo and roughness maps significantly improves over the baseline, and the proposed NormalNet and VisibilityNet further improve the results while enabling free-viewpoint rendering. UVDeltaNet provides more realistic results (LPIPS metric) with a small sacrifice in the PSNR." + }, + { + "type": "table", + "bbox": [ + 0.445, + 0.308, + 0.785, + 0.379 + ], + "angle": 0, + "content": "
MethodIndoors (S1,T2)Outdoors (S1,T6)
PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓
Baseline17.5220.7500.18118.3550.7100.203
+Albedo/Rough. Maps18.3340.7890.16819.3410.7470.181
+Normal/Vis.Net18.6340.7920.17319.7040.7600.180
+UVDeltaNet (Ours)18.6020.7920.16319.4610.7530.173
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.394, + 0.784, + 0.424 + ], + "angle": 0, + "content": "ness 2D map, we can insert new textures or new materials and render the edited neural actor under new poses, new viewpoints, and new light conditions." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.448, + 0.364, + 0.463 + ], + "angle": 0, + "content": "6 Limitations" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.481, + 0.488, + 0.631 + ], + "angle": 0, + "content": "This work focuses on visual fidelity, and representations faster than NeRF could be tested in our framework in future. Moreover, like several previous techniques for the free-viewpoint rendering of humans and avatar generation [32,45], the proposed method can currently only deal with tight clothing due to the coarse geometry registration. The coarse mesh registration" + }, + { + "type": "image", + "bbox": [ + 0.507, + 0.448, + 0.787, + 0.539 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.497, + 0.552, + 0.788, + 0.621 + ], + "angle": 0, + "content": "Fig.8: Our method allows editing the material properties and simultaneously performing skeletal pose control, free-viewpoint rendering, and relighting of the edited neural actor." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.632, + 0.735, + 0.647 + ], + "angle": 0, + "content": "can also negatively affect the fine details in the face and hand synthesis." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.671, + 0.36, + 0.687 + ], + "angle": 0, + "content": "7 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.704, + 0.788, + 0.841 + ], + "angle": 0, + "content": "We conclude that the proposed method for the free-viewpoint rendering of humans enables simultaneous relighting, material editing, and pose control at test time while requiring a single multi-view video sequence with static illumination during training. Our method demonstrates functionality that no competing method making the same assumptions at training time (i.e. from multi-view RGB video inputs only) can provide. We hope that our method and the proposed dataset will provoke further research in this exciting field, bringing the neural rendering of humans to a new qualitative level and paving the way for next-generation metaverse applications." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.248, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Relightable Neural Actor with Intrinsic Decomposition and Pose Control" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.146, + 0.403, + 0.163 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.176, + 0.787, + 0.222 + ], + "angle": 0, + "content": "This work was funded by the ERC Consolidator Grant 4DRepLy (770784), the Saarbrücken Research Center for Visual Computing, Interaction and AI, and the German Research Foundation (DFG) under Grant No. 468670075." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.244, + 0.323, + 0.26 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.274, + 0.784, + 0.304 + ], + "angle": 0, + "content": "1. Easymocap - make human motion capture easier. Github (2021), https://github.com/zju3dv/EasyMocap" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.304, + 0.785, + 0.344 + ], + "angle": 0, + "content": "2. Alldieck, T., Magnor, M., Xu, W., Theobalt, C., Pons-Moll, G.: Video based reconstruction of 3d people models. In: Computer Vision and Pattern Recognition (CVPR). pp. 8387-8397 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.344, + 0.785, + 0.385 + ], + "angle": 0, + "content": "3. Alldieck, T., Xu, H., Sminchisescu, C.: imghum: Implicit generative models of 3d human shape and articulated pose. In: International Conference on Computer Vision (ICCV). pp. 5461-5470 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.385, + 0.785, + 0.427 + ], + "angle": 0, + "content": "4. Bi, S., Lombardi, S., Saito, S., Simon, T., Wei, S.E., Mcphail, K., Ramamoorthi, R., Sheikh, Y., Saragih, J.: Deep relightable appearance models for animatable faces. ACM Transactions on Graphics (TOG) 40(4), 1-15 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.427, + 0.785, + 0.468 + ], + "angle": 0, + "content": "5. Boss, M., Braun, R., Jampani, V., Barron, J.T., Liu, C., Lensch, H.: Nerd: Neural reflectance decomposition from image collections. In: International Conference on Computer Vision (ICCV). pp. 12684-12694 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.468, + 0.785, + 0.497 + ], + "angle": 0, + "content": "6. Chen, Z., Liu, Z.: Relighting4d: Neural relightable human from videos. In: European Conference on Computer Vision (ECCV) (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.497, + 0.785, + 0.537 + ], + "angle": 0, + "content": "7. De Aguiar, E., Theobalt, C., Thrun, S., Seidel, H.P.: Automatic conversion of mesh animations into skeleton-based animations. In: Computer Graphics Forum (EUROGRAPHICS). vol. 27, pp. 389-397 (2008)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.537, + 0.785, + 0.565 + ], + "angle": 0, + "content": "8. Debevec, P.E., Malik, J.: Recovering high dynamic range radiance maps from photographs. In: ACM SIGGRAPH 2008 classes, pp. 1-10 (2008)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.565, + 0.785, + 0.606 + ], + "angle": 0, + "content": "9. Deng, B., Wang, Y., Wetzstein, G.: Unconditional generation of relightable 3d human faces. In: International Conference on 3D Vision (3DV). pp. 302-312. IEEE (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.606, + 0.785, + 0.647 + ], + "angle": 0, + "content": "0. d'Eon, E., Luebke, D., Enderton, E.: Efficient Rendering of Human Skin. In: Kautz, J., Pattanaik, S. (eds.) Rendering Techniques. The Eurographics Association (2007). https://doi.org/10.2312/EGWR/EGSR07/147-157" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.647, + 0.785, + 0.702 + ], + "angle": 0, + "content": "1. Guo, K., Lincoln, P., Davidson, P., Busch, J., Yu, X., Whalen, M., Harvey, G., Orts-Escalano, S., Pandey, R., Dourgarian, J., et al.: The relighttables: Volumetric performance capture of humans with realistic relighting. ACM Transactions on Graphics (TOG) 38(6), 1-19 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.702, + 0.785, + 0.758 + ], + "angle": 0, + "content": "2. Habermann, M., Liu, L., Xu, W., Pons-Moll, G., Zollhoefer, M., Theobalt, C.: Hdhumans: A hybrid approach for high-fidelity digital humans. Proc. ACM Comput. Graph. Interact. Tech. 6(3) (aug 2023). https://doi.org/10.1145/3606927, https://doi.org/10.1145/3606927" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.758, + 0.785, + 0.799 + ], + "angle": 0, + "content": "3. Habermann, M., Liu, L., Xu, W., Zollhoefer, M., Pons-Moll, G., Theobalt, C.: Real-time deep dynamic characters. ACM Transactions on Graphics 40(4) (aug 2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.799, + 0.785, + 0.841 + ], + "angle": 0, + "content": "4. Habermann, M., Xu, W., Zollhoefer, M., Pons-Moll, G., Theobalt, C.: Livecap: Real-time human performance capture from monocular video. ACM Transactions On Graphics (TOG) 38(2), 1-17 (2019)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.274, + 0.785, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.401, + 0.127 + ], + "angle": 0, + "content": "D.C. Luvizon et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.785, + 0.189 + ], + "angle": 0, + "content": "15. Habermann, M., Xu, W., Zollhofer, M., Pons-Moll, G., Theobalt, C.: Deepcap: Monocular human performance capture using weak supervision. In: Computer Vision and Pattern Recognition (CVPR). pp. 5052-5063 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.191, + 0.785, + 0.218 + ], + "angle": 0, + "content": "16. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Computer Vision and Pattern Recognition (CVPR). pp. 770-778 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.22, + 0.785, + 0.261 + ], + "angle": 0, + "content": "17. Iqbal, U., Caliskan, A., Nagano, K., Khamis, S., Molchanov, P., Kautz, J.: Rana: Relightable articulated neural avatars. In: International Conference on Computer Vision (ICCV). pp. 23142-23153 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.262, + 0.785, + 0.317 + ], + "angle": 0, + "content": "18. Iwase, S., Saito, S., Simon, T., Lombardi, S., Bagautdinov, T., Joshi, R., Prada, F., Shiratori, T., Sheikh, Y., Saragih, J.: Relightablehands: Efficient neural relighting of articulated hand models. In: Computer Vision and Pattern Recognition (CVPR). pp. 16663-16673 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.318, + 0.785, + 0.359 + ], + "angle": 0, + "content": "19. Ji, C., Yu, T., Guo, K., Liu, J., Liu, Y.: Geometry-aware single-image full-body human relighting. In: European Conference on Computer Vision (ECCV). pp. 388-405. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.361, + 0.785, + 0.402 + ], + "angle": 0, + "content": "20. Jiang, W., Yi, K.M., Samei, G., Tuzel, O., Ranjan, A.: Neuman: Neural human radiance field from a single video. In: European Conference on Computer Vision (ECCV). pp. 402-418 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.403, + 0.785, + 0.445 + ], + "angle": 0, + "content": "21. Jiang, Y., Habermann, M., Golyanik, V., Theobalt, C.: Hifecap: Monocular high-fidelity and expressive capture of human performances. In: British Machine Vision Conference (BMVC) (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.446, + 0.785, + 0.487 + ], + "angle": 0, + "content": "22. Jin, H., Liu, I., Xu, P., Zhang, X., Han, S., Bi, S., Zhou, X., Xu, Z., Su, H.: Tensoroir: Tensorial inverse rendering. In: Computer Vision and Pattern Recognition (CVPR). pp. 165-174 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.488, + 0.785, + 0.516 + ], + "angle": 0, + "content": "23. Kingma, D.P., Ba, J.: Adam: A method for stochastic optimization. CoRR abs/1412.6980 (2014), https://api-semanticscholar.org/CorpusID:6628106" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.517, + 0.785, + 0.558 + ], + "angle": 0, + "content": "24. Kwon, Y., Kim, D., Ceylan, D., Fuchs, H.: Neural human performer: Learning generalizable radiance fields for human performance rendering. Advances in Neural Information Processing Systems (NeurIPS) 34, 24741-24752 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.559, + 0.785, + 0.601 + ], + "angle": 0, + "content": "25. Kwon, Y., Liu, L., Fuchs, H., Habermann, M., Theobalt, C.: Deliffas: Deformable light fields for fast avatar synthesis. Advances in Neural Information Processing Systems (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.602, + 0.785, + 0.657 + ], + "angle": 0, + "content": "26. Ledig, C., Theis, L., Huszar, F., Caballero, J., Cunningham, A., Acosta, A., Aitken, A., Tejani, A., Totz, J., Wang, Z., et al.: Photo-realistic single image superresolution using a generative adversarial network. In: Computer Vision and Pattern Recognition (CVPR). pp. 4681-4690 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.658, + 0.785, + 0.713 + ], + "angle": 0, + "content": "27. Lewis, J.P., Cordner, M., Fong, N.: Pose space deformation: A unified approach to shape interpolation and skeleton-driven deformation. In: SIGGRAPH. p. 165-172. ACM Press/Addison-Wesley Publishing Co., USA (2000). https://doi.org/10.1145/344779.344862, https://doi.org/10.1145/344779.344862" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.714, + 0.785, + 0.756 + ], + "angle": 0, + "content": "28. Li, G., Meka, A., Mueller, F., Buehler, M.C., Hilliges, O., Beeler, T.: Eyenerf: a hybrid representation for photorealistic synthesis, animation and relighting of human eyes. ACM Transactions on Graphics (TOG) 41(4), 1-16 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.757, + 0.785, + 0.798 + ], + "angle": 0, + "content": "29. Li, G., Wu, C., Stoll, C., Liu, Y., Varanasi, K., Dai, Q., Theobalt, C.: Capturing relightable human performances under general uncontrolled illumination. In: Computer Graphics Forum. vol. 32, pp. 275-284. Wiley Online Library (2013)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.799, + 0.785, + 0.84 + ], + "angle": 0, + "content": "30. Lin, S., Ryabtsev, A., Sengupta, S., Curless, B.L., Seitz, S.M., Kemelmacher-Shlizerman, I.: Real-time high-resolution background matting. In: Computer Vision and Pattern Recognition (CVPR). pp. 8762-8771 (2021)" + }, + { + "type": "list", + "bbox": [ + 0.218, + 0.147, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.248, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Relightable Neural Actor with Intrinsic Decomposition and Pose Control" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.189 + ], + "angle": 0, + "content": "31. Liu, G., Reda, F.A., Shih, K.J., Wang, T.C., Tao, A., Catanzaro, B.: Image inpainting for irregular holes using partial convolutions. In: European Conference on Computer Vision (ECCV) (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.191, + 0.785, + 0.233 + ], + "angle": 0, + "content": "32. Liu, L., Habermann, M., Rudnev, V., Sarkar, K., Gu, J., Theobalt, C.: Neural actor: Neural free-view synthesis of human actors with pose control. ACM Transactions on Graphics (TOG) 40(6), 1-16 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.234, + 0.785, + 0.274 + ], + "angle": 0, + "content": "33. Liu, Y., Wang, P., Lin, C., Long, X., Wang, J., Liu, L., Komura, T., Wang, W.: Nano: Neural geometry and brdf reconstruction of reflective objects from multiview images. In: SIGGRAPH (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.276, + 0.785, + 0.317 + ], + "angle": 0, + "content": "34. Loper, M., Mahmood, N., Romero, J., Pons-Moll, G., Black, M.J.: SMPL: A skinned multi-person linear model. ACM Transactions on Graphics (SIGGRAPH Asia) 34(6), 248:1-248:16 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.318, + 0.785, + 0.359 + ], + "angle": 0, + "content": "35. Lyu, L., Tewari, A., Leimkuehler, T., Habermann, M., Theobalt, C.: Neural radiance transfer fields for relightable novel-view synthesis with global illumination. In: European Conference on Computer Vision (ECCV) (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.361, + 0.785, + 0.443 + ], + "angle": 0, + "content": "36. Meka, A., Pandey, R., Hane, C., Orts-Escolano, S., Barnum, P., David-Son, P., Erickson, D., Zhang, Y., Taylor, J., Bouaziz, S., Legendre, C., Ma, W.C., Overbeck, R., Beeler, T., Debevec, P., Izadi, S., Theobalt, C., Rhemann, C., Fanello, S.: Deep relightable textures: Volumetric performance capture with neural rendering. ACM Transactions on Graphics (TOG) 39(6) (nov 2020). https://doi.org/10.1145/3414685.3417814, https://doi.org/10.1145/3414685.3417814" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.445, + 0.785, + 0.487 + ], + "angle": 0, + "content": "37. Mihajlovic, M., Saito, S., Bansal, A., Zollhoefer, M., Tang, S.: COAP: Compositional articulated occupancy of people. In: Computer Vision and Pattern Recognition (CVPR) (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.488, + 0.785, + 0.528 + ], + "angle": 0, + "content": "38. Mihajlovic, M., Zhang, Y., Black, M.J., Tang, S.: Leap: Learning articulated occupancy of people. In: Computer Vision and Pattern Recognition (CVPR). pp. 10461-10471 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.53, + 0.785, + 0.572 + ], + "angle": 0, + "content": "39. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. In: European Conference on Computer Vision (ECCV) (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.573, + 0.785, + 0.614 + ], + "angle": 0, + "content": "40. Pang, H., Zhu, H., Kortylewski, A., Theobalt, C., Habermann, M.: Ash: Animatable gaussian splats for efficient and photoreal human rendering. In: Computer Vision and Pattern Recognition (CVPR). pp. 1165-1175 (June 2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.615, + 0.785, + 0.657 + ], + "angle": 0, + "content": "41. Pavlakos, G., Choutas, V., Ghorbani, N., Bolkart, T., Osman, A.A.A., Tzionas, D., Black, M.J.: Expressive body capture: 3d hands, face, and body from a single image. In: Computer Vision and Pattern Recognition (CVPR) (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.658, + 0.785, + 0.699 + ], + "angle": 0, + "content": "42. Peng, S., Zhang, Y., Xu, Y., Wang, Q., Shuai, Q., Bao, H., Zhou, X.: Neural body: Implicit neural representations with structured latent codes for novel view synthesis of dynamic humans. In: Computer Vision and Pattern Recognition (CVPR) (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.7, + 0.785, + 0.768 + ], + "angle": 0, + "content": "43. Ramamoorthi, R., Hanrahan, P.: A signal-processing framework for inverse rendering. In: Conference on Computer Graphics and Interactive Techniques. p. 117-128. SIGGRAPH '01, Association for Computing Machinery, New York, NY, USA (2001). https://doi.org/10.1145/383259.383271, https://doi.org/10.1145/383259.383271" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.77, + 0.785, + 0.798 + ], + "angle": 0, + "content": "44. Ranjan, A., Yi, K.M., Chang, J.H.R., Tuzel, O.: Facelit: Neural 3d relightable faces. In: Computer Vision and Pattern Recognition (CVPR). pp. 8619-8628 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.799, + 0.785, + 0.84 + ], + "angle": 0, + "content": "45. Remelli, E., Bagautdinov, T., Saito, S., Wu, C., Simon, T., Wei, S.E., Guo, K., Cao, Z., Prada, F., Saragih, J., et al.: Drivable volumetric avatars using texel-aligned features. In: ACM SIGGRAPH 2022 Conference Proceedings (2022)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.401, + 0.127 + ], + "angle": 0, + "content": "D.C. Luvizon et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.785, + 0.189 + ], + "angle": 0, + "content": "46. Schwartz, G., Wei, S.E., Wang, T.L., Lombardi, S., Simon, T., Saragih, J., Sheikh, Y.: The eyes have it: An integrated eye and face model for photorealistic facial animation. ACM Transactions on Graphics (TOG) 39(4), 91-1 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.189, + 0.785, + 0.23 + ], + "angle": 0, + "content": "47. Shu, Z., Sahasrabudhe, M., Guler, R.A., Samaras, D., Paragios, N., Kokkinos, I.: Deforming autoencoders: Unsupervised disentangling of shape and appearance. In: European Conference on Computer Vision (ECCV). pp. 650-665 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.23, + 0.785, + 0.284 + ], + "angle": 0, + "content": "48. Shysheya, A., Zakharov, E., Aliev, K.A., Bashirov, R., Burkov, E., Iskakov, K., Ivakshtenko, A., Malkov, Y., Pasechnik, I., Ulyanov, D., et al.: Textured neural avatars. In: Computer Vision and Pattern Recognition (CVPR). pp. 2387-2397 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.284, + 0.785, + 0.338 + ], + "angle": 0, + "content": "49. Srinivasan, P.P., Deng, B., Zhang, X., Tancik, M., Mildenhall, B., Barron, J.T.: Nerv: Neural reflectance and visibility fields for relighting and view synthesis. In: Computer Vision and Pattern Recognition (CVPR). pp. 7491-7500 (2021). https://doi.org/10.1109/CVPR46437.2021.00741" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.338, + 0.785, + 0.379 + ], + "angle": 0, + "content": "50. Tan, F., Fanello, S., Meka, A., Orts-Escalano, S., Tang, D., Pandey, R., Taylor, J., Tan, P., Zhang, Y.: Volux-gan: A generative model for 3d face synthesis with hdri relighting. In: ACM SIGGRAPH 2022 Conference Proceedings. pp. 1-9 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.379, + 0.785, + 0.434 + ], + "angle": 0, + "content": "51. Tewari, A., Thies, J., Mildenhall, B., Srinivasan, P., Tretschk, E., Yifan, W., Lassner, C., Sitzmann, V., Martin-Brualla, R., Lombardi, S., Simon, T., Theobalt, C., Nießner, M., Barron, J.T., Wetzstein, G., Zollhöfer, M., Golyanik, V.: Advances in Neural Rendering. Computer Graphics Forum (EG STAR 2022) (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.434, + 0.785, + 0.487 + ], + "angle": 0, + "content": "52. Walter, B., Marschner, S.R., Li, H., Torrance, K.E.: Microfacet Models for Refraction through Rough Surfaces. In: Kautz, J., Pattanaik, S. (eds.) Rendering Techniques. The Eurographics Association (2007). https://doi.org/10.2312/EGWR/EGSR07/195-206" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.487, + 0.785, + 0.528 + ], + "angle": 0, + "content": "53. Wang, P., Liu, L., Liu, Y., Theobalt, C., Komura, T., Wang, W.: Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. In: Advances in Neural Information Processing Systems (NeurIPS) (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.528, + 0.785, + 0.568 + ], + "angle": 0, + "content": "54. Wang, S., Schwarz, K., Geiger, A., Tang, S.: Arah: Animatable volume rendering of articulated human sdfs. In: European Conference on Computer Vision (ECCV) (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.568, + 0.785, + 0.609 + ], + "angle": 0, + "content": "55. Wang, T.C., Liu, M.Y., Zhu, J.Y., Liu, G., Tao, A., Kautz, J., Catanzaro, B.: Video-to-video synthesis. In: Conference on Neural Information Processing Systems (NeurIPS) (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.609, + 0.785, + 0.65 + ], + "angle": 0, + "content": "56. Wang, Y., Holynski, A., Zhang, X., Zhang, X.: Sunstage: Portrait reconstruction and relighting using the sun as a light stage. In: Computer Vision and Pattern Recognition (CVPR). pp. 20792-20802 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.65, + 0.785, + 0.703 + ], + "angle": 0, + "content": "57. Weng, C.Y., Curless, B., Srinivasan, P.P., Barron, J.T., Kemelmacher-Shlizerman, I.: Humannerf: Free-viewpoint rendering of moving people from monocular video. In: Computer Vision and Pattern Recognition (CVPR). pp. 16210-16220 (June 2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.703, + 0.785, + 0.799 + ], + "angle": 0, + "content": "58. Wuu, C.h., Zheng, N., Ardisson, S., Bali, R., Belko, D., Brockmeyer, E., Evans, L., Godisart, T., Ha, H., Huang, X., Hypes, A., Koska, T., Krenn, S., Lombardi, S., Luo, X., McPhail, K., Millerschoen, L., Perdoch, M., Pitts, M., Richard, A., Saragih, J., Saragih, J., Shiratori, T., Simon, T., Stewart, M., Trimble, A., Weng, X., Whitewolf, D., Wu, C., Yu, S.I., Sheikh, Y.: Multiface: A dataset for neural face rendering. In: arXiv (2022). https://doi.org/10.48550/ARXIV.2207.11243, https://arxiv.org/abs/2207.11243" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.799, + 0.785, + 0.841 + ], + "angle": 0, + "content": "59. Yang, H., Zheng, M., Feng, W., Huang, H., Lai, Y.K., Wan, P., Wang, Z., Ma, C.: Towards practical capture of high-fidelity reconfigurable avatars. In: SIGGRAPH Asia 2023 Conference Papers. pp. 1-11 (2023)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.248, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Relightable Neural Actor with Intrinsic Decomposition and Pose Control" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "19" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.788, + 0.189 + ], + "angle": 0, + "content": "60. Zhang, K., Luan, F., Li, Z., Snavely, N.: IRON: Inverse rendering by optimizing neural sdfs and materials from photometric images. In: Computer Vision and Pattern Recognition (CVPR). pp. 5565-5574 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.19, + 0.788, + 0.231 + ], + "angle": 0, + "content": "61. Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual metric. In: Computer Vision and Pattern Recognition (CVPR) (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.231, + 0.787, + 0.273 + ], + "angle": 0, + "content": "62. Zhang, X., Srinivasan, P.P., Deng, B., Debevec, P., Freeman, W.T., Barron, J.T.: NeRFactor: Neural factorization of shape and reflectance under an unknown illumination. ACM Transactions on Graphics (TOG) 40, 1-18 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.273, + 0.787, + 0.313 + ], + "angle": 0, + "content": "63. Zhang, Y., Sun, J., He, X., Fu, H., Jia, R., Zhou, X.: Modeling indirect illumination for inverse rendering. In: Computer Vision and Pattern Recognition (CVPR). pp. 18643-18652 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.314, + 0.787, + 0.357 + ], + "angle": 0, + "content": "64. Zheng, R., Li, P., Wang, H., Yu, T.: Learning visibility field for detailed 3d human reconstruction and relighting. In: Computer Vision and Pattern Recognition (CVPR). pp. 216-226 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.357, + 0.787, + 0.398 + ], + "angle": 0, + "content": "65. Zhi, Y., Qian, S., Yan, X., Gao, S.: Dual-space nerf: Learning animatable avatars and scene lighting in separate spaces. In: International Conference on 3D Vision (3DV) (Sep 2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.398, + 0.787, + 0.44 + ], + "angle": 0, + "content": "66. Zhou, T., He, K., Wu, D., Xu, T., Zhang, Q., Shao, K., Chen, W., Xu, L., Yu, J.: Relightable neural human assets from multi-view gradient illuminations. In: Computer Vision and Pattern Recognition (CVPR). pp. 4315-4327 (2023)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.788, + 0.44 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/Relightable Neural Actor with Intrinsic Decomposition and Pose Control/b67bc924-ef1e-4a87-8446-f01e4ef31ad6_origin.pdf b/2024/Relightable Neural Actor with Intrinsic Decomposition and Pose Control/b67bc924-ef1e-4a87-8446-f01e4ef31ad6_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..92baf3ad16bdda7b8aa407df79f2f1be62438f3f --- /dev/null +++ b/2024/Relightable Neural Actor with Intrinsic Decomposition and Pose Control/b67bc924-ef1e-4a87-8446-f01e4ef31ad6_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:531da1fa904ea9cc2c5ce9824ce77d85d25b1a06c100d8bd558e3f572b60acae +size 9417940 diff --git a/2024/Relightable Neural Actor with Intrinsic Decomposition and Pose Control/full.md b/2024/Relightable Neural Actor with Intrinsic Decomposition and Pose Control/full.md new file mode 100644 index 0000000000000000000000000000000000000000..14b4ae0d657131b536aca7bd07336087df662b95 --- /dev/null +++ b/2024/Relightable Neural Actor with Intrinsic Decomposition and Pose Control/full.md @@ -0,0 +1,320 @@ +# Relightable Neural Actor with Intrinsic Decomposition and Pose Control + +Diogo Carbonera Luvizon $^{1,2}$ , Vladislav Golyanik $^{1}$ , Adam Kortylewski $^{1,3}$ , Marc Habermann $^{1,2}$ , and Christian Theobalt $^{1,2}$ + +$^{1}$ Max Planck Institute for Informatics, Saarland Informatics Campus + $^{2}$ Saarbrücken Research Center for Visual Computing, Interaction and AI + $^{3}$ University of Freiburg + +{dluvizon,golyanik,akortyle,mhaberma,theobalt}@mpi-inf.mpg.de https://vcai.mpi-inf.mpg.de/projects/RNA + +Abstract. Creating a controllable and relightable digital avatar from multi-view video with fixed illumination is a very challenging problem since humans are highly articulated, creating pose-dependent appearance effects, and skin as well as clothing require space-varying BRDF modeling. Existing works on creating animatable avatars either do not focus on relighting at all, require controlled illumination setups, or try to recover a relightable avatar from very low cost setups, i.e. a single RGB video, at the cost of severely limited result quality, e.g. shadows not even being modeled. To address this, we propose Relightable Neural Actor, a new video-based method for learning a pose-driven neural human model that can be relighted, allows appearance editing, and models pose-dependent effects such as wrinkles and self-shadows. Importantly, for training, our method solely requires a multi-view recording of the human under a known, but static lighting condition. To tackle this challenging problem, we leverage an implicit geometry representation of the actor with a drivable density field that models pose-dependent deformations and derive a dynamic mapping between 3D and UV spaces, where normal, visibility, and materials are effectively encoded. To evaluate our approach in real-world scenarios, we collect a new dataset with four identities recorded under different light conditions, indoors and outdoors, providing the first benchmark of its kind for human relighting, and demonstrating state-of-the-art relighting results for novel human poses. + +# 1 Introduction + +Creating a relightable and controllable virtual actor of a real person solely from multi-view videos under static illumination is an exciting and challenging topic that has numerous applications in virtual and augmented reality, gaming, human-machine interaction, and content creation. However, it requires modeling the outgoing radiance of a particular surface point on the human, which depends on the pose-dependent and highly non-rigid surface deformation of the human body, the complex and spatially varying materials, as well as the scene lighting. Intrinsically decomposing these quantities solely from multi-view video and + +![](images/1d796e40e690dcc337edceab4833f41ef7484c233c64e6670fc21aeb5b8c5b74.jpg) +Fig. 1: Our method learns a neural actor that is driven only by a 3D skeletal pose and allows rendering and editing of the actor's appearance under new lightning and poses not seen during training. Our approach models pose-dependent deformations, self-shadows, and performs intrinsic decomposition of dynamic humans through a pose-dependent mapping between 3D and UV spaces, which also enables editing the appearance and material properties at inference time. For training, our method only needs a multi-view video and an environment map. + +![](images/2bfcf18ed57fbca0ab339fc63250a8c06953c6e79673d39f9e744e693239c609.jpg) + +![](images/b172c1837958adb9144e48e7917212a58ddb6ca8959a316cc51fb31ea2cefa16.jpg) + +enabling novel compositions that are of highly photorealistic quality defines an unsolved and challenging inverse problem in the literature. + +Relighting static scenes from multi-view imagery is a well-studied problem [22,33,49,62]. However, directly applying such methods to moving humans, which are dynamic, highly articulated, and have complex material composition, is not trivially possible, as these approaches do not account for dynamics. On the contrary, some works have focused on building drivable and dynamic full-body avatars from multi-view video [32,42,54] without considering relightability at all. Despite some works [4,11,18,36,66] assuming a well-calibrated lightstage settings, only few human-specific works [6,17,64] try to achieve human intrinsic decomposition and relighting from videos with static lights. While they are able to recover a relightable human representation from a single RGB video, they can solely replay the human motion under novel lighting conditions [6], or do not model self-shadows and space-varying BRDF [17], which results in low visual fidelity. In stark contrast, we assume a multi-view setting while accounting for high-frequency surface deformations as well as induced high-frequency lighting effects such as self-shadows and shading, which are indispensable for a truly immersive experience. + +To address this, we propose Relightable Neural Actor, a new method that achieves intrinsic decomposition of dynamic human avatars, enabling pose-control, material editing, and realistic rendering with pose-dependent effects, while solely requiring a multi-view video of the real human under a single and static light condition during training (see Fig. 1). At inference time, we only need an arbitrary driving skeletal pose and an environment map. + +We solve this challenging problem by leveraging an implicit density field from a pose-driven neural actor and by proposing a novel intrinsic decomposition guided by the implicit density field. This strategy allows our method to model pose-dependent deformations in the geometry as dynamic maps, while + +representing the albedo and roughness properties of the actor as static maps. Our approach not only makes the problem computationally tractable for the long and dynamic human sequences, but also enables editing the neural actor's appearance and materials at inference time. Our key contribution is to sample normal and visibility from the posed density field and aggregate them into sparse UV maps, which are then densified by the proposed NormalNet and VisibilityNet models. This allows our approach to effectively model wrinkles and self-shadows even for long sequences. The albedo and roughness maps are optimized during training and sampled by the proposed UVDeltaNet, which accounts for pose-dependent deformations in the UV map and is intrinsically learned in our pipeline. Finally, given an environment map and a virtual camera position, our neural renderer produces the final image of the neural actor. + +To train and evaluate our method, we collected a real-world dataset for novel pose and novel lighting human synthesis, composed of four actors recorded under six different light conditions by a multi-view camera setup, including indoor and outdoor environments and a light probe for each sequence, providing the first real in-the-wild benchmark for controllable human relighting. + +To summarize, our main technical contributions are as follows: 1) The first method to learn a relightable neural actor from multi-view video under static lights that allows for pose control, appearance editing, and free-viewpoint rendering under arbitrary illumination while modeling pose-dependent effects. 2) We propose an intrinsic decomposition of diffuse and specular reflection based on a new approach to compute pose-dependent normals and visibility by sparsely sampling a neural field and densifying the results with NormalNet and VisibilityNet, making the problem of pose-driven human relighting tractable. Our approach relies on a mapping between 3D and UV space implicitly learned by UVDeltaNet, allowing us to learn an explicit and static material representation that is easy to edit at inference time. + +In our experimental evaluation, we compare our approach with two baselines in the proposed real data benchmark, showing the superiority of our approach in synthesizing relightable humans on challenging motion sequences. The effectiveness of our method is demonstrated by ablation studies and qualitative results. + +# 2 Related Work + +Since the goal of this work is to build a full-body relightable and animatable avatar, we will not discuss works [9, 10, 28, 44, 46, 47, 50, 56, 58, 59] focusing on face relighting as both settings have different requirements and challenges. + +# 2.1 Relighting Static Scenes + +Classic approaches for relighting and editing perform intrinsic decomposition through inverse rendering [43]. With recent advances in neural rendering [51], light transportation can be taken into account with fully differentiable rendering + +pipelines. For instance, Zhang et al. [62] address the problem of recovering geometry and spatially-varying BRDF with an implicit scene representation. However, this requires solving the light transport by evaluating the NeRF model for each light ray, which is very inefficient. NeRV [49] and Zhang et al. [63] mitigate this problem with a neural visibility field that is jointly optimized with the NeRF model, and NeRD [5] adds further relaxation by ignoring cast shadows. However, this assumption cannot be made for humans, which are highly articulated and often experience prominent self-casting shadows. Jin et al. [22] propose TensorIR, an efficient tensor-based representation that solves the inverse problem by casting camera and secondary rays. IRON [60] proposes a two-stage process for obtaining a mesh with material properties, and Lyu et al. [35] and NeRO [33] focus on multi-bounce illumination and reflective objects, which is usually not the case for humans. The methods mentioned above share the assumption of a static scene and cannot be trivially adapted to dynamic and controllable humans. + +# 2.2 Human Models with Pose Control + +Modeling humans is a long-standing topic. Traditionally, meshes from an actor can be automatically rigged to a skeleton [7] using Linear Blending Skinning (LBS) [27]. Parametric human models offer pose and shape control [34,41] and, more recently, implicit human bodies can be articulated [3, 37, 38]. Such models offer a generic and controllable representation, but are unable to represent individual details. Template-based methods synthesize a realistic level of details [14, 15, 21], but have to deal with non-linear pose-dependent deformations. + +Several approaches model appearance and geometry with an implicit field anchored on a coarse human mesh. Neural Body [42] optimizes latent vectors anchored in the vertices of the human mesh and Neural Actor [32] incorporates texture maps as a conditioning to break down the mapping from pose to dynamic effects (one-to-many mapping). ARAH [54] models the human surface with an SDF, Dual-Space NeRF [65] learns the color and light in separate spaces, and Kwon et al. [24] propose an approach that generalizes to new identities and poses. Shyshey et al. [48] jointly learn textures and a neural renderer but fall short in reconstructing the 3D human body. A few works [20,57] approach a restrictive scenario of reconstructing 3D humans from a monocular video but fail to model fine details under new poses. Common to all these type of methods [12,13,25, 40] is the entanglement of material properties and illumination, making them restricted to the same appearance as observed during training. + +# 2.3 Relighting Human Models + +Only a few methods in the literature tackle the problem of relighting humans from videos under static illumination. Li et al. [29] create a relightable human mesh by estimating a low-frequency environment using Spherical Harmonics. Some approaches target the problem of single-image [19] and static pose [64] relighting under unknown lights, which is an extremely ill-posed problem due to the high ambiguity between albedo and lighting. Relighting4D [6] extends + +Neural Body [42] to estimate geometry and reflectance, but can only replay the same sequence observed during training under new lights and is not able to generate new poses. RANA [17] allows for pose control but assumes Lambertian reflectance, ignoring different material properties such as skin and clothes, and does not model cast shadows, resulting in unrealistic renderings for articulated humans. In contrast, our method performs intrinsic decomposition with space-varying BRDF, models illumination and self-shadows, allows for pose control and novel lights, and offers an easy manner to edit the material properties of the learned neural actor. + +# 3 Method + +Our goal is to learn a digital human avatar (neural actor) from RGB images that models pose-dependent effects and can be relighted, pose-controlled, and rendered from novel viewpoints at test time. To achieve relighting, we decompose the observed appearance into normals, visibility with respect to each light source, and BRDF represented as albedo and roughness. With these components and an environment map, we employ the microfacet model [52] to render the output image of our method, which is compared to the observed images. In addition—since we target a drivable actor—the geometry components need to be controllable by the skeletal pose. + +Fig. 2 shows an overview of our method divided into three main components. The first is a pose-driven implicit geometry model, where the goal is to obtain a skeleton-controllable implicit 3D density field that models the geometry of the actor with pose-dependent deformations. This model is adapted from Neural Actor [32] and is depicted here since our approach builds upon its internal design. Note that Neural Actor does not perform intrinsic decomposition and is unable to perform relighting or appearance editing, which is our main goal. The second component is our proposed intrinsic decomposition approach whose goal is to disentangle geometry represented as normal and visibility information, and material properties as albedo and roughness. Our key insight is to represent these components in UV space rather than in 3D coordinates, using the density field as guidance in the mapping function between 3D and UV spaces. This allows our method to model complex deformations that are pose-dependent in a compact UV map that can be easily computed and queried. Finally, we integrate the microfacet model [52] into our rendering pipeline that takes as input the disentangled components (i.e., normal, visibility, albedo, and roughness) and an environment map, and renders the image of the neural actor in an arbitrary pose and light condition from a free-viewpoint. + +# 3.1 Preliminaries + +Human Pose and Environment Map. Our method takes as input a 3D articulated human skeleton pose $\mathbf{P}_t = \{\pmb {\theta}_t,\mathbf{R}_t,\mathbf{T}_t\}$ , defined by the relative joint rotation $\pmb {\theta}\in \mathbb{R}^{J\times 3}$ of $J$ body joints, the global rotation $\mathbf{R}\in \mathbb{R}^{3\times 3}$ , and global + +![](images/4ba019d3bcb19d26ed0bef08e626dd22de5b901e5002eed5f4402d39d5e00ab3.jpg) +Fig. 2: Our method takes as input a 3D skeletal pose and a static environment map, and renders the neural actor from a virtual camera position. The pose-driven geometry (Sec. 3.1) learns an implicit density function, from which we compute the normal and visibility information. Our intrinsic decomposition disentangles normal, visibility, albedo and roughness maps in UV space (Sec. 3.2). The neural renderer (Sec. 3.3) outputs our prediction, which is supervised with the reference image. Yellow represents learnable components and green non-learnable components. + +translation $\mathbf{T} \in \mathbb{R}^3$ of the person at frame $t$ . The 3D pose can be used to drive a coarse human mesh with linear blending skinning (LBS) [34]. We represent the global lights as an environment map $\mathbf{L} \in \mathbb{R}^{16 \times 32 \times 3}$ , which can be obtained from HDR images of a mirror sphere [8]—see the supplement for more details. The environment map can be static during training and changed during inference to render results under novel illumination. + +Training Data. During training, our method is supervised with multi-view calibrated video streams $\mathbf{V}_c = \{\mathbf{I}_{c,t}\}$ , where $c \in \{1, \dots, C\}$ is the camera index and $\mathbf{I}_{c,t}$ is the image from camera $c$ at frame $t$ . We obtain a foreground mask $\mathbf{M}_{c,t}$ for each video frame [30]. The coarse human mesh and 3D skeletons can be obtained with multi-view pose tracking [1]. The input 3D skeletons drive the LBS model to obtain posed human meshes used to extract texture maps from the multi-view images [2]. These texture maps are used during training to learn pose-dependent deformations on the implicit human model. At inference time, a V2V [55] network predicts texture maps from posed normal maps [32]. + +Implicit Human Model. For realistically relighting articulated avatars, we need to model human geometry with deformations caused by dynamic and varied poses. Therefore, we follow Neural Actor [32] to represent the pose-driven geometry as an implicit field guided by the 3D skeletal pose parameters and conditioned by texture maps, which provide the ability to model high-frequency deformations. Specifically, a standard NeRF model [39] is optimized through volumetric rendering in the canonical space of the coarse human mesh and the canonical space is deformed based on ResNet [16] features extracted from the + +texture maps, which provide the ability to represent detailed wrinkles and deformations not modeled by the coarse mesh. During optimization, from each pixel in the reference images we cast a ray $\mathbf{r}_n = \mathbf{o} + \delta_n\mathbf{d}$ , where $n = 1,\dots N$ is the index of each point in the ray, $\mathbf{o} \in \mathbb{R}^3$ is the origin defined in the optical center of the camera, $\mathbf{d} \in \mathbb{R}^3$ is the ray direction, and $\delta_{n}$ is the depth (distance to the camera). The sampled points $\mathbf{r}_n$ are projected to the mesh surface, resulting in the local coordinates $(u_{n},v_{n},h_{n})$ , where $(u,v)$ is the surface coordinate in the UV space and $h$ is the signed distance to the mesh. + +# 3.2 Neural Intrinsic Decomposition + +At the core of our method, we perform intrinsic decomposition to obtain the normal, visibility, albedo, and roughness components, required for rendering the neural actor under new light conditions. Specifically, for each point in each ray cast into the scene we want to obtain the normal direction $\mathbf{n} \in \mathbb{R}^3$ , the visibility coefficients $\pmb{v} \in \mathbb{R}^{512}$ respective to each light source in the environment map, the RGB albedo $\mathbf{a} \in \mathbb{R}^3$ , and the scalar roughness $\rho$ . The normal and visibility values can be computed from the geometry model. However, computing the visibility of each point in 3D with respect to the light sources is computationally intractable for a dynamic and long sequence, since for every point in 3D and for every frame, several rays have to be cast towards each pixel of the environment map. Therefore, we propose a different approach. + +We notice that the intrinsic decomposition is only relevant for points close to the surface. Since the coarse human mesh provides a mapping from the mesh surface to UV space and the implicit human model provides a density field that models the geometry of the actor, we propose an aggregation strategy that samples normal and visibility from a given point of view (therefore, only a few points are sampled in 3D) and aggregates the obtained values in a per-frame UV map. Once we obtain the normal and visibility in UV space for a given pose, we can simply query the values by projecting the sampled 3D point $\mathbf{r}_n$ into the UV space, preventing our method from the expensive computation of normal and visibility at every possible 3D coordinate over multiple frames. We next explain this approach in more detail. + +Normal Maps. For each point in a ray, we want to obtain the normal value that will be required for solving the rendering equation during relighting (will be introduced in Sec. 3.3). This value could be obtained by computing the gradient from the neural field [62]. However, this process requires back-propagating gradients throughout the entire network and often results in noisy normals. Instead, we explicitly compute the normal vector close to the visible surface by deriving the unbiased depth $\hat{\mathbf{D}}$ [53]: + +$$ +\hat {\mathbf {D}} (\mathbf {r}) = \frac {\sum_ {n = 1} ^ {N} w _ {n} \delta_ {n}}{\sum_ {n = 1} ^ {N} w _ {n}}, \quad \text {w i t h} w _ {n} = T _ {n} \left(1 - \exp \left(- \sigma_ {n} \bar {\delta} _ {n}\right)\right), \tag {1} +$$ + +where $T_{n} = \exp \left(-\sum_{j = 1}^{n - 1}\sigma_{j}\bar{\delta}_{j}\right)$ , $w_{n}$ is the density weight from volumetric rendering [39], and $\sigma_{n}$ is the density predicted by the NeRF model at point $\mathbf{r}_n$ . + +Then, as illustrated in Fig. 3, we convert the depth values relative to the camera to 3D coordinates and obtain the normal direction from the tangent surface computed at the intersection point, defined as the maximum value of the density weight $w_{n}$ . This process can be performed for multiple viewpoints to cover the surface of the human. + +Once the normal values are obtained in 3D, we project them to the UV map and aggregate them with the + +obtained normals from different views. In this process, the $(u^{*},v^{*})$ coordinate where the obtained normal will be projected onto the UV map is defined by + +![](images/3b45f527bf19b66200054d5f0242d5e8ba44f8178da9396b69d1696f62a4cf61.jpg) +Fig. 3: The unbiased depth from Eq. (1) is projected into 3D coordinates, where we approximate the tangent surface at the intersecting point and obtain the normal vector. Normals are sampled from multiple viewpoints and aggregated in the UV map. + +![](images/b1acf57fc92ded197bad1307778668bf79e3040b092aa4643c58511589a67fd5.jpg) + +![](images/a01c98550c215374a72a6f5002bd4bb3cd0459a6d0457b1852be1e8ec7b2426e.jpg) + +$$ +\left(u ^ {*}, v ^ {*}\right) = \Psi \left(\mathbf {r} _ {n} ^ {*}\right), \quad \text {w i t h} \mathbf {r} _ {n} ^ {*} = \underset {\mathbf {r} _ {n}} {\arg \max } w _ {n}, \tag {2} +$$ + +where $\varPsi$ is a function that projects a 3D point to the closest surface point in the human mesh. Note that $w_{n}$ depends on the sampled points $\mathbf{r}_n$ (from $T_{n}$ in Eq. 1) and $(u^{*},v^{*})$ corresponds to the $(u_{n},v_{n})$ value obtained from the point $\mathbf{r}_n$ on the ray that has the largest density weight $w_{n}$ . This process ensures that the normal map obtained aggregates the normals most likely close to the surface. + +In this process, the resulting UV map can be sparse, which is undesirable for free-viewpoint rendering. Thus, the NormalNet is used to inpaint the missing normals in UV space. This lightweight CNN is implemented with 8 layers of partial convolutions [31], takes as input sparse normal maps with size $512 \times 512 \times 3$ , and predicts a dense normal map (see Fig. 2). During training, NormalNet is supervised with densely sampled normal maps from the training data. + +Visibility Maps. The visibility information is necessary to model pose-dependent shadows in the neural avatar, without which it is not possible to produce realistic relighting results. A naive approach for obtaining the visibility for each point in 3D would require querying the neural field thousands of times while casting rays from the intersecting point to all the light sources, which is prohibitive during training. In contrast, in our approach we efficiently encode visibility of each point in 3D close to the surface by applying an approximate mapping to UV space. Specifically, we pre-compute the visibility maps for each pose by sampling multiple camera views and aggregate the results in UV map following Eq. (2). The resulting visibility maps with size $256 \times 256 \times 512$ encode the visibility information w.r.t. each light source in the environment map (represented by 512 individual lights). Similarly to the computation of normal maps, this process results in sparse visibility maps, which are inpainted by the VisibilityNet (as + +shown in Fig. 2). This shallow CNN takes as input sparse visibility and normal maps, as the visibility highly depends on normals (e.g. the visibility is always zero for the backlight source), and is trained with densely sampled visibility from the training sequence. See the supplement for architecture details. + +Albedo and Roughness Maps. As discussed above, the normal and visibility components can be computed from the density field by using UV maps as proxies. However, the albedo and roughness properties cannot be directly obtained from the pre-trained Neural Actor model, since it predicts radiance which is the product of the rendering equation. Instead, we want to learn a decomposed representation. In intrinsic decomposition, the material properties are commonly optimized via differentiable rendering. Hence, we optimize the albedo and roughness of the neural actor by leveraging a differentiable renderer (Sec. 3.3) and by imposing priors through our design. + +The material properties of an actor should not depend on the pose and should be efficiently encoded to enable fast computations and editing. Therefore, we model albedo and roughness as static UV maps, respectively defined as $\mathbf{A} \in \mathbb{R}^{1024 \times 1024 \times 3}$ and $\mathbf{B} \in \mathbb{R}^{1024 \times 1024}$ . These tensors are jointly optimized for the whole training sequence and regularized with an L1 loss on the spatial gradients, which encourages smoothness in both maps: + +$$ +\mathcal {L} _ {\text {s m o o t h}} = \frac {1}{N} \sum \left| \nabla_ {x} (\mathbf {A}) \right| + \left| \nabla_ {y} (\mathbf {A}) \right| + \left| \nabla_ {x} (\mathbf {B}) \right| + \left| \nabla_ {y} (\mathbf {B}) \right|, \tag {3} +$$ + +where $\nabla_{x}$ and $\nabla_{y}$ represent the spatial gradient operators. + +Nonetheless, by assuming a material representation that does not depend on the skeletal pose, we still have to cope with intrinsic mesh registration errors. Such registration errors could cause blurry results or artifacts in the final renderings. To cope with this problem, we propose a corrective term to the UV mapping, which is predicted by the UVDeltaNet, as shown in Fig. 2. The UVDeltaNet is an MLP that takes as input the local coordinates $(u,v,h)$ for each point on the ray and the sampled texture features $\psi$ , and outputs a corrective term $\epsilon = (u',v')$ . The texture features are necessary to provide additional information, since the local coordinates alone cannot account for deviations in the mesh registration. + +Initially, UVDeltaNet is frozen and outputs zeros. After the initial convergence of $\mathbf{A}$ and $\mathbf{B}$ , UVDeltaNet is indirectly supervised by the rendering loss and regularized with: + +$$ +\mathcal {L} _ {\mathrm {u v}} = \frac {1}{N} \sum \epsilon^ {2} + | \nabla_ {x} (\epsilon) | + | \nabla_ {y} (\epsilon) |, \tag {4} +$$ + +which minimizes the magnitude of the deformation and encourages its spatial gradient to be sparse, i.e., it makes the UV deformation spatially coherent. + +# 3.3 Neural Renderer + +The third part of our method is our neural renderer, which takes the disentangled components as normal, visibility, albedo, and roughness; integrates the lights from the environment map, and outputs the rendered image. + +Our neural renderer starts by recasting the ray $\mathbf{r}$ with importance sampling around the estimated depth value $\hat{\mathbf{D}}$ . Thus, all the points $\mathbf{r}_n$ are most likely close to the surface. For each sampled point in the ray, we obtain the normal direction $\mathbf{n}_n$ and the visibility values $\pmb{v}_n$ by interpolating the normal and visibility maps in UV around the point $(u_n,v_n) = \varPsi(\mathbf{r}_n)$ . The albedo $\mathbf{a}_n$ and roughness $\rho_n$ are obtained by interpolating the albedo and roughness maps in UV around the point $(u_n + u'_u,v_n + v'_n)$ , i.e., following the deformation predicted by UVDeltaNet. + +Before rendering the final image of the posed neural actor, we consider the light transport equation defined as + +$$ +\mathbf {L} _ {0} (\mathbf {r} _ {n}, \boldsymbol {\omega} _ {0}) = \sum_ {i} \operatorname {R} (\mathbf {a} _ {n}, \rho_ {n}, \boldsymbol {\omega} _ {i}, \boldsymbol {\omega} _ {0}) \boldsymbol {v} _ {n, i} \mathbf {L} (\boldsymbol {\omega} _ {i}) (\boldsymbol {\omega} _ {i} \mathbf {n} _ {n}), \tag {5} +$$ + +where $\mathrm{R}(\cdot)$ is the microfacet BRDF model [52], $\omega_0$ and $\omega_{i}$ are the camera and the light source directions, respectively, $\pmb{v}_{n,i}$ is the visibility of point $\mathbf{r}_n$ w.r.t. the ith light source, and $\mathbf{L}(\boldsymbol {\omega}_i)$ is the incoming light from the environment map. Eq. (5) is defined for a single point $\mathbf{r}_n$ on the ray and is integrated into our volume rendering as + +$$ +\tilde {\mathbf {I}} (\mathbf {r}) = \sum_ {n = 1} ^ {N} w _ {n} \mathbf {L} _ {0} \left(\mathbf {r} _ {n}, \omega_ {0}\right), \tag {6} +$$ + +where $\tilde{\mathbf{I}}$ is the output image of the relighted neural actor. + +During training, our relightable model is supervised with the following loss: + +$$ +\mathcal {L} _ {\text {r e l i g h t}} = \lambda_ {\mathrm {L 2}} \mathcal {L} _ {\mathrm {L 2}} + \lambda_ {\text {v g g}} \mathcal {L} _ {\text {v g g}} + \lambda_ {\text {s m o o t h}} \mathcal {L} _ {\text {s m o o t h}} + \mathcal {L} _ {\text {u v}}, \tag {7} +$$ + +where $\mathcal{L}_{\mathrm{L2}}$ and $\mathcal{L}_{\mathrm{vgg}}$ are the L2 and the perceptual losses [26] between $\tilde{\mathbf{I}} (\mathbf{r})$ and the reference image. The hyperparameters are defined as $\lambda_{\mathrm{L2}} = 100$ , $\lambda_{\mathrm{vgg}} = 0.01$ , and $\lambda_{\mathrm{smooth}} = 0.1$ in all our experiments. + +# 4 Relightable Dynamic Actors Dataset + +Existing datasets for human reenactment and relighting are often monocular [2, 17], contain very short sequences with only a few pose variations, and do not provide environment maps [2,17,32,42]. Hence, we cannot use them for evaluation under realistic scenarios with real ground truth. In addition, we observe that the only existing synthetic dataset for human relighting and pose control [17] suffers from artifacts in the geometry and salt-and-pepper noise in the renderings. + +With these shortcomings in mind, we collect a new real dataset, which we call Relightable Dynamic Actors, with four identities performing a series of ten different actions (Fig. 4). Each subject is recorded six times wearing the same clothes but under different light conditions, including indoor and outdoor, where we used around 100 and eight cameras, respectively. For each sequence, we also capture HDR light probes, which can be used for training or testing. In total, our dataset contains around 90k multi-view video frames divided into 24 multi-view video sequences. We refer to the supplemental document for more details. + +![](images/a06b68f5e197c3c761b0ca2e1bf0c0a96d8720e1eb5027a3c6ac90444f51c1f3.jpg) +Fig. 4: Relightable Dynamic Actors dataset. For each actor, we recorded six sequences under different light conditions, including five indoor sequences and one outdoor. Environment maps were obtained from HDR pictures of a mirror sphere. + +![](images/ea85ef744f3f6e88d203721ec7d31977792d3e0697e883150fee5d73e31dc5d1.jpg) + +# 5 Experiments + +In this section, we provide an experimental evaluation of our method on the Reli-ightable Dynamic Actors dataset and show qualitative and quantitative results. Please refer to the supplementary video for dynamic results on long sequences. + +# 5.1 Evaluation + +Metrics. For quantitative evaluation, we report results on the PSNR, SSIM and LPIPS metrics [61], comparing the rendered relight results with the groundtruth images from our dataset. Importantly, our work is the first to be able to report such metrics on a real dataset since no previous dataset in the literature provided a real human recorded under different and known light conditions. + +Baselines. Since the only method in the literature capable of reposing a human under new light conditions does not provide implementation [17], and Relighting4D [6] is not able to perform inference for new poses, we defined two baselines for a fair comparison. Our baselines are derived from Neural Actor [32] and adapted to perform intrinsic decomposition based on the obtained normal and visibility information, as discussed in Sec. 3.2. However, instead of using NormalNet and VisibilityNet to obtain these components, we densely sample the neural field and perform morphological inpainting to obtain dense UV maps. The first baseline, referred to as Neural Actor $+IR$ , predicts the albedo and roughness components instead of RGB color directly from the deformed canonical space $\mathbf{x}_n = (x,y,z)$ (see Fig. 2). The second baseline, Neural Actor $+IR + Tex$ , takes as input the texture features $\psi$ as a conditioning for predicting the albedo and roughness values. In both cases, the predictor is a NeRF MLP with 8 layers. + +Implementation Details. Our model is trained in a three-stage process. First, we train the geometry component (Sec. 3.1) with the Adam [23] optimizer with a learning rate of 0.001 until convergence, which takes about two days on four NVidia A40 GPUs. Then, we pre-train the NormalNet and VisibilityNet with pre-computed normal and visibility UV maps sampled from the geometry model + +![](images/d89fd9d77195b1decc66845928fbb1097f4a8947c98467a015c03010339021a3.jpg) +Fig. 5: Qualitative results of our method on novel poses. Our approach generalizes to both indoor and outdoor illumination, and generates realistic renderings for arbitrary environment maps with detailed shadows and pose-dependent deformations. + +from 40 different views. Both networks are trained in parallel using Adam with a learning rate of 0.01 for about 12 hours. Finally, we train our intrinsic decomposition model with the same optimizer until convergence, which takes about one day. Since the obtained environment maps are up to factor of scale [8], we apply global color correction during training and evaluation. The predicted texture maps are obtained with a V2V model [55] trained (in parallel to our model) to translate a sequence of SMPL normal maps into a sequence of texture maps. Please refer to our supplementary materials for architecture details. + +# 5.2 Results + +Qualitative Results. In Fig. 5, we show qualitative results of our method. Our results are all from a trained model driven by a new skeletal pose, new camera viewpoint, and new environment map not seen during training. Note how our approach produces realistic appearance and prominent cloth wrinkles, which can be easily observed under different light conditions. Importantly, notice how our results look realistic under novel lights, even though our model is trained on a single sequence under a static light condition. + +Table 1: Quantitative results from our method on real data, subject S1, considering indoor and outdoor scenes. Our method consistently improves the baselines, especially in outdoor scenarios, where the gap between training and test lights is more pronounced. + +
MethodT2T3T4T5T6 (outdoors)
PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓
Neural Actor+IR17.5220.7500.18117.4730.7680.17817.7450.7670.17118.2440.7680.17518.3550.7100.203
Neural Actor+IR+Tex18.1750.7780.16417.7990.7900.16718.1320.7850.16718.5310.7890.16618.7040.7370.186
Ours18.6020.7920.16318.2400.8000.16918.3880.8000.16418.8200.8000.16519.4610.7530.173
+ +In Fig. 6, we show results from our intrinsic decomposition. Note that our normal representation nicely captures the geometry details and the visibility. This is important to obtain an albedo that does not contain baked-in lighting effects, e.g. shadows, as can be observed from our results. + +![](images/0dd55a64e873c05df6a67eec8ba9595a19834eecddf6d8249c71aa9f30b6c696.jpg) +Fig. 6: Results from our intrinsic decomposition. Our method provides detailed normals and visibility (averaged for visualization), and plausible albedo and roughness predictions that enable realistic relighting. + +Another important and unique characteristic of our method is the ability to model self-shadows and cloth wrinkles. In Fig. 7, we show results from two subjects, where a "sunrise" light probe was used to produce cast shadows. Note how pose-dependent effects are consistent with the mostly directional illumination. + +Comparisons. We evaluated our method on the test sequences from our dataset. Specifically, we compare against two baselines in the indoor sequences (T2-T5) and outdoor sequence (T6) of subject S1 in every 10th frame, which sums up to 4000 evaluation frames. In Tab. 1 we show, for the first time, quantitative results of real human relighting under new poses and new light conditions. Our approach consistently improves upon the two baselines. + +![](images/5b5fbd96a5687646854daa587d130d438fb3408cdb371bc3824a1159e27170b2.jpg) +Fig. 7: Our method models pose-dependent effects and cast shadows, as can be observed from our results under the "sunrise" light. + +Ablation. In Tab. 2, we show an ablation with different components of our method. Compared to our baseline (Neural Actor+IR), learning the albedo and roughness as UV maps guided by the neural field significantly improves the performance on relighting and reposing. This is because, in the baselines, the albedo and roughness are under-constrained and can possibly embed shading effects, overfitting to the training data. The NormalNet and VisibilityNet provide additional capacity to the model to learn how to inpaint the sampled UV maps. Note that without these two models our results are computed using morphological region growth in the UV maps. Finally, the UVDeltaNet provides more flexibility to the albedo and roughness representation, accounting for fine details. This can be observed in the improvement in the LPIPS metric in the last + +row of Tab. 2. Nonetheless, this also means more capacity to overfit the training sequences, which translates into a small decrease in the PSNR metric. + +# 5.3 Applications + +An important characteristic of our approach is our static representation in UV for the albedo and roughness. This allows exciting and new applications such as color and material editing of an animatable and relightable neural actor, as shown in Fig. 8 and in Fig. 1. With a simple edit in the obtained albedo or rough- + +Table 2: Ablation study with different components of our method. The proposed intrinsic decomposition with static albedo and roughness maps significantly improves over the baseline, and the proposed NormalNet and VisibilityNet further improve the results while enabling free-viewpoint rendering. UVDeltaNet provides more realistic results (LPIPS metric) with a small sacrifice in the PSNR. + +
MethodIndoors (S1,T2)Outdoors (S1,T6)
PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓
Baseline17.5220.7500.18118.3550.7100.203
+Albedo/Rough. Maps18.3340.7890.16819.3410.7470.181
+Normal/Vis.Net18.6340.7920.17319.7040.7600.180
+UVDeltaNet (Ours)18.6020.7920.16319.4610.7530.173
+ +ness 2D map, we can insert new textures or new materials and render the edited neural actor under new poses, new viewpoints, and new light conditions. + +# 6 Limitations + +This work focuses on visual fidelity, and representations faster than NeRF could be tested in our framework in future. Moreover, like several previous techniques for the free-viewpoint rendering of humans and avatar generation [32,45], the proposed method can currently only deal with tight clothing due to the coarse geometry registration. The coarse mesh registration + +![](images/bb58ca6ec11e0b637ecdbbdbccf9a05600ffbc19f4df560ea6561ae712b818da.jpg) +Fig.8: Our method allows editing the material properties and simultaneously performing skeletal pose control, free-viewpoint rendering, and relighting of the edited neural actor. + +can also negatively affect the fine details in the face and hand synthesis. + +# 7 Conclusion + +We conclude that the proposed method for the free-viewpoint rendering of humans enables simultaneous relighting, material editing, and pose control at test time while requiring a single multi-view video sequence with static illumination during training. Our method demonstrates functionality that no competing method making the same assumptions at training time (i.e. from multi-view RGB video inputs only) can provide. We hope that our method and the proposed dataset will provoke further research in this exciting field, bringing the neural rendering of humans to a new qualitative level and paving the way for next-generation metaverse applications. + +# Acknowledgements + +This work was funded by the ERC Consolidator Grant 4DRepLy (770784), the Saarbrücken Research Center for Visual Computing, Interaction and AI, and the German Research Foundation (DFG) under Grant No. 468670075. + +# References + +1. Easymocap - make human motion capture easier. Github (2021), https://github.com/zju3dv/EasyMocap +2. Alldieck, T., Magnor, M., Xu, W., Theobalt, C., Pons-Moll, G.: Video based reconstruction of 3d people models. In: Computer Vision and Pattern Recognition (CVPR). pp. 8387-8397 (2018) +3. Alldieck, T., Xu, H., Sminchisescu, C.: imghum: Implicit generative models of 3d human shape and articulated pose. In: International Conference on Computer Vision (ICCV). pp. 5461-5470 (2021) +4. Bi, S., Lombardi, S., Saito, S., Simon, T., Wei, S.E., Mcphail, K., Ramamoorthi, R., Sheikh, Y., Saragih, J.: Deep relightable appearance models for animatable faces. ACM Transactions on Graphics (TOG) 40(4), 1-15 (2021) +5. Boss, M., Braun, R., Jampani, V., Barron, J.T., Liu, C., Lensch, H.: Nerd: Neural reflectance decomposition from image collections. In: International Conference on Computer Vision (ICCV). pp. 12684-12694 (2021) +6. Chen, Z., Liu, Z.: Relighting4d: Neural relightable human from videos. In: European Conference on Computer Vision (ECCV) (2022) +7. De Aguiar, E., Theobalt, C., Thrun, S., Seidel, H.P.: Automatic conversion of mesh animations into skeleton-based animations. In: Computer Graphics Forum (EUROGRAPHICS). vol. 27, pp. 389-397 (2008) +8. Debevec, P.E., Malik, J.: Recovering high dynamic range radiance maps from photographs. In: ACM SIGGRAPH 2008 classes, pp. 1-10 (2008) +9. Deng, B., Wang, Y., Wetzstein, G.: Unconditional generation of relightable 3d human faces. In: International Conference on 3D Vision (3DV). pp. 302-312. IEEE (2024) +0. d'Eon, E., Luebke, D., Enderton, E.: Efficient Rendering of Human Skin. In: Kautz, J., Pattanaik, S. (eds.) Rendering Techniques. The Eurographics Association (2007). https://doi.org/10.2312/EGWR/EGSR07/147-157 +1. Guo, K., Lincoln, P., Davidson, P., Busch, J., Yu, X., Whalen, M., Harvey, G., Orts-Escalano, S., Pandey, R., Dourgarian, J., et al.: The relighttables: Volumetric performance capture of humans with realistic relighting. ACM Transactions on Graphics (TOG) 38(6), 1-19 (2019) +2. Habermann, M., Liu, L., Xu, W., Pons-Moll, G., Zollhoefer, M., Theobalt, C.: Hdhumans: A hybrid approach for high-fidelity digital humans. Proc. ACM Comput. Graph. Interact. Tech. 6(3) (aug 2023). https://doi.org/10.1145/3606927, https://doi.org/10.1145/3606927 +3. Habermann, M., Liu, L., Xu, W., Zollhoefer, M., Pons-Moll, G., Theobalt, C.: Real-time deep dynamic characters. ACM Transactions on Graphics 40(4) (aug 2021) +4. Habermann, M., Xu, W., Zollhoefer, M., Pons-Moll, G., Theobalt, C.: Livecap: Real-time human performance capture from monocular video. ACM Transactions On Graphics (TOG) 38(2), 1-17 (2019) + +15. Habermann, M., Xu, W., Zollhofer, M., Pons-Moll, G., Theobalt, C.: Deepcap: Monocular human performance capture using weak supervision. In: Computer Vision and Pattern Recognition (CVPR). pp. 5052-5063 (2020) +16. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Computer Vision and Pattern Recognition (CVPR). pp. 770-778 (2016) +17. Iqbal, U., Caliskan, A., Nagano, K., Khamis, S., Molchanov, P., Kautz, J.: Rana: Relightable articulated neural avatars. In: International Conference on Computer Vision (ICCV). pp. 23142-23153 (2023) +18. Iwase, S., Saito, S., Simon, T., Lombardi, S., Bagautdinov, T., Joshi, R., Prada, F., Shiratori, T., Sheikh, Y., Saragih, J.: Relightablehands: Efficient neural relighting of articulated hand models. In: Computer Vision and Pattern Recognition (CVPR). pp. 16663-16673 (2023) +19. Ji, C., Yu, T., Guo, K., Liu, J., Liu, Y.: Geometry-aware single-image full-body human relighting. In: European Conference on Computer Vision (ECCV). pp. 388-405. Springer (2022) +20. Jiang, W., Yi, K.M., Samei, G., Tuzel, O., Ranjan, A.: Neuman: Neural human radiance field from a single video. In: European Conference on Computer Vision (ECCV). pp. 402-418 (2022) +21. Jiang, Y., Habermann, M., Golyanik, V., Theobalt, C.: Hifecap: Monocular high-fidelity and expressive capture of human performances. In: British Machine Vision Conference (BMVC) (2022) +22. Jin, H., Liu, I., Xu, P., Zhang, X., Han, S., Bi, S., Zhou, X., Xu, Z., Su, H.: Tensoroir: Tensorial inverse rendering. In: Computer Vision and Pattern Recognition (CVPR). pp. 165-174 (2023) +23. Kingma, D.P., Ba, J.: Adam: A method for stochastic optimization. CoRR abs/1412.6980 (2014), https://api-semanticscholar.org/CorpusID:6628106 +24. Kwon, Y., Kim, D., Ceylan, D., Fuchs, H.: Neural human performer: Learning generalizable radiance fields for human performance rendering. Advances in Neural Information Processing Systems (NeurIPS) 34, 24741-24752 (2021) +25. Kwon, Y., Liu, L., Fuchs, H., Habermann, M., Theobalt, C.: Deliffas: Deformable light fields for fast avatar synthesis. Advances in Neural Information Processing Systems (2023) +26. Ledig, C., Theis, L., Huszar, F., Caballero, J., Cunningham, A., Acosta, A., Aitken, A., Tejani, A., Totz, J., Wang, Z., et al.: Photo-realistic single image superresolution using a generative adversarial network. In: Computer Vision and Pattern Recognition (CVPR). pp. 4681-4690 (2017) +27. Lewis, J.P., Cordner, M., Fong, N.: Pose space deformation: A unified approach to shape interpolation and skeleton-driven deformation. In: SIGGRAPH. p. 165-172. ACM Press/Addison-Wesley Publishing Co., USA (2000). https://doi.org/10.1145/344779.344862, https://doi.org/10.1145/344779.344862 +28. Li, G., Meka, A., Mueller, F., Buehler, M.C., Hilliges, O., Beeler, T.: Eyenerf: a hybrid representation for photorealistic synthesis, animation and relighting of human eyes. ACM Transactions on Graphics (TOG) 41(4), 1-16 (2022) +29. Li, G., Wu, C., Stoll, C., Liu, Y., Varanasi, K., Dai, Q., Theobalt, C.: Capturing relightable human performances under general uncontrolled illumination. In: Computer Graphics Forum. vol. 32, pp. 275-284. Wiley Online Library (2013) +30. Lin, S., Ryabtsev, A., Sengupta, S., Curless, B.L., Seitz, S.M., Kemelmacher-Shlizerman, I.: Real-time high-resolution background matting. In: Computer Vision and Pattern Recognition (CVPR). pp. 8762-8771 (2021) + +31. Liu, G., Reda, F.A., Shih, K.J., Wang, T.C., Tao, A., Catanzaro, B.: Image inpainting for irregular holes using partial convolutions. In: European Conference on Computer Vision (ECCV) (2018) +32. Liu, L., Habermann, M., Rudnev, V., Sarkar, K., Gu, J., Theobalt, C.: Neural actor: Neural free-view synthesis of human actors with pose control. ACM Transactions on Graphics (TOG) 40(6), 1-16 (2021) +33. Liu, Y., Wang, P., Lin, C., Long, X., Wang, J., Liu, L., Komura, T., Wang, W.: Nano: Neural geometry and brdf reconstruction of reflective objects from multiview images. In: SIGGRAPH (2023) +34. Loper, M., Mahmood, N., Romero, J., Pons-Moll, G., Black, M.J.: SMPL: A skinned multi-person linear model. ACM Transactions on Graphics (SIGGRAPH Asia) 34(6), 248:1-248:16 (2015) +35. Lyu, L., Tewari, A., Leimkuehler, T., Habermann, M., Theobalt, C.: Neural radiance transfer fields for relightable novel-view synthesis with global illumination. In: European Conference on Computer Vision (ECCV) (2022) +36. Meka, A., Pandey, R., Hane, C., Orts-Escolano, S., Barnum, P., David-Son, P., Erickson, D., Zhang, Y., Taylor, J., Bouaziz, S., Legendre, C., Ma, W.C., Overbeck, R., Beeler, T., Debevec, P., Izadi, S., Theobalt, C., Rhemann, C., Fanello, S.: Deep relightable textures: Volumetric performance capture with neural rendering. ACM Transactions on Graphics (TOG) 39(6) (nov 2020). https://doi.org/10.1145/3414685.3417814, https://doi.org/10.1145/3414685.3417814 +37. Mihajlovic, M., Saito, S., Bansal, A., Zollhoefer, M., Tang, S.: COAP: Compositional articulated occupancy of people. In: Computer Vision and Pattern Recognition (CVPR) (2022) +38. Mihajlovic, M., Zhang, Y., Black, M.J., Tang, S.: Leap: Learning articulated occupancy of people. In: Computer Vision and Pattern Recognition (CVPR). pp. 10461-10471 (2021) +39. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. In: European Conference on Computer Vision (ECCV) (2020) +40. Pang, H., Zhu, H., Kortylewski, A., Theobalt, C., Habermann, M.: Ash: Animatable gaussian splats for efficient and photoreal human rendering. In: Computer Vision and Pattern Recognition (CVPR). pp. 1165-1175 (June 2024) +41. Pavlakos, G., Choutas, V., Ghorbani, N., Bolkart, T., Osman, A.A.A., Tzionas, D., Black, M.J.: Expressive body capture: 3d hands, face, and body from a single image. In: Computer Vision and Pattern Recognition (CVPR) (2019) +42. Peng, S., Zhang, Y., Xu, Y., Wang, Q., Shuai, Q., Bao, H., Zhou, X.: Neural body: Implicit neural representations with structured latent codes for novel view synthesis of dynamic humans. In: Computer Vision and Pattern Recognition (CVPR) (2021) +43. Ramamoorthi, R., Hanrahan, P.: A signal-processing framework for inverse rendering. In: Conference on Computer Graphics and Interactive Techniques. p. 117-128. SIGGRAPH '01, Association for Computing Machinery, New York, NY, USA (2001). https://doi.org/10.1145/383259.383271, https://doi.org/10.1145/383259.383271 +44. Ranjan, A., Yi, K.M., Chang, J.H.R., Tuzel, O.: Facelit: Neural 3d relightable faces. In: Computer Vision and Pattern Recognition (CVPR). pp. 8619-8628 (2023) +45. Remelli, E., Bagautdinov, T., Saito, S., Wu, C., Simon, T., Wei, S.E., Guo, K., Cao, Z., Prada, F., Saragih, J., et al.: Drivable volumetric avatars using texel-aligned features. In: ACM SIGGRAPH 2022 Conference Proceedings (2022) + +46. Schwartz, G., Wei, S.E., Wang, T.L., Lombardi, S., Simon, T., Saragih, J., Sheikh, Y.: The eyes have it: An integrated eye and face model for photorealistic facial animation. ACM Transactions on Graphics (TOG) 39(4), 91-1 (2020) +47. Shu, Z., Sahasrabudhe, M., Guler, R.A., Samaras, D., Paragios, N., Kokkinos, I.: Deforming autoencoders: Unsupervised disentangling of shape and appearance. In: European Conference on Computer Vision (ECCV). pp. 650-665 (2018) +48. Shysheya, A., Zakharov, E., Aliev, K.A., Bashirov, R., Burkov, E., Iskakov, K., Ivakshtenko, A., Malkov, Y., Pasechnik, I., Ulyanov, D., et al.: Textured neural avatars. In: Computer Vision and Pattern Recognition (CVPR). pp. 2387-2397 (2019) +49. Srinivasan, P.P., Deng, B., Zhang, X., Tancik, M., Mildenhall, B., Barron, J.T.: Nerv: Neural reflectance and visibility fields for relighting and view synthesis. In: Computer Vision and Pattern Recognition (CVPR). pp. 7491-7500 (2021). https://doi.org/10.1109/CVPR46437.2021.00741 +50. Tan, F., Fanello, S., Meka, A., Orts-Escalano, S., Tang, D., Pandey, R., Taylor, J., Tan, P., Zhang, Y.: Volux-gan: A generative model for 3d face synthesis with hdri relighting. In: ACM SIGGRAPH 2022 Conference Proceedings. pp. 1-9 (2022) +51. Tewari, A., Thies, J., Mildenhall, B., Srinivasan, P., Tretschk, E., Yifan, W., Lassner, C., Sitzmann, V., Martin-Brualla, R., Lombardi, S., Simon, T., Theobalt, C., Nießner, M., Barron, J.T., Wetzstein, G., Zollhöfer, M., Golyanik, V.: Advances in Neural Rendering. Computer Graphics Forum (EG STAR 2022) (2022) +52. Walter, B., Marschner, S.R., Li, H., Torrance, K.E.: Microfacet Models for Refraction through Rough Surfaces. In: Kautz, J., Pattanaik, S. (eds.) Rendering Techniques. The Eurographics Association (2007). https://doi.org/10.2312/EGWR/EGSR07/195-206 +53. Wang, P., Liu, L., Liu, Y., Theobalt, C., Komura, T., Wang, W.: Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. In: Advances in Neural Information Processing Systems (NeurIPS) (2021) +54. Wang, S., Schwarz, K., Geiger, A., Tang, S.: Arah: Animatable volume rendering of articulated human sdfs. In: European Conference on Computer Vision (ECCV) (2022) +55. Wang, T.C., Liu, M.Y., Zhu, J.Y., Liu, G., Tao, A., Kautz, J., Catanzaro, B.: Video-to-video synthesis. In: Conference on Neural Information Processing Systems (NeurIPS) (2018) +56. Wang, Y., Holynski, A., Zhang, X., Zhang, X.: Sunstage: Portrait reconstruction and relighting using the sun as a light stage. In: Computer Vision and Pattern Recognition (CVPR). pp. 20792-20802 (2023) +57. Weng, C.Y., Curless, B., Srinivasan, P.P., Barron, J.T., Kemelmacher-Shlizerman, I.: Humannerf: Free-viewpoint rendering of moving people from monocular video. In: Computer Vision and Pattern Recognition (CVPR). pp. 16210-16220 (June 2022) +58. Wuu, C.h., Zheng, N., Ardisson, S., Bali, R., Belko, D., Brockmeyer, E., Evans, L., Godisart, T., Ha, H., Huang, X., Hypes, A., Koska, T., Krenn, S., Lombardi, S., Luo, X., McPhail, K., Millerschoen, L., Perdoch, M., Pitts, M., Richard, A., Saragih, J., Saragih, J., Shiratori, T., Simon, T., Stewart, M., Trimble, A., Weng, X., Whitewolf, D., Wu, C., Yu, S.I., Sheikh, Y.: Multiface: A dataset for neural face rendering. In: arXiv (2022). https://doi.org/10.48550/ARXIV.2207.11243, https://arxiv.org/abs/2207.11243 +59. Yang, H., Zheng, M., Feng, W., Huang, H., Lai, Y.K., Wan, P., Wang, Z., Ma, C.: Towards practical capture of high-fidelity reconfigurable avatars. In: SIGGRAPH Asia 2023 Conference Papers. pp. 1-11 (2023) + +60. Zhang, K., Luan, F., Li, Z., Snavely, N.: IRON: Inverse rendering by optimizing neural sdfs and materials from photometric images. In: Computer Vision and Pattern Recognition (CVPR). pp. 5565-5574 (2022) +61. Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual metric. In: Computer Vision and Pattern Recognition (CVPR) (2018) +62. Zhang, X., Srinivasan, P.P., Deng, B., Debevec, P., Freeman, W.T., Barron, J.T.: NeRFactor: Neural factorization of shape and reflectance under an unknown illumination. ACM Transactions on Graphics (TOG) 40, 1-18 (2021) +63. Zhang, Y., Sun, J., He, X., Fu, H., Jia, R., Zhou, X.: Modeling indirect illumination for inverse rendering. In: Computer Vision and Pattern Recognition (CVPR). pp. 18643-18652 (2022) +64. Zheng, R., Li, P., Wang, H., Yu, T.: Learning visibility field for detailed 3d human reconstruction and relighting. In: Computer Vision and Pattern Recognition (CVPR). pp. 216-226 (2023) +65. Zhi, Y., Qian, S., Yan, X., Gao, S.: Dual-space nerf: Learning animatable avatars and scene lighting in separate spaces. In: International Conference on 3D Vision (3DV) (Sep 2022) +66. Zhou, T., He, K., Wu, D., Xu, T., Zhang, Q., Shao, K., Chen, W., Xu, L., Yu, J.: Relightable neural human assets from multi-view gradient illuminations. In: Computer Vision and Pattern Recognition (CVPR). pp. 4315-4327 (2023) \ No newline at end of file diff --git a/2024/Relightable Neural Actor with Intrinsic Decomposition and Pose Control/images.zip b/2024/Relightable Neural Actor with Intrinsic Decomposition and Pose Control/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..f753f680b4c8c10d27de2118aacedd2dbbb8a612 --- /dev/null +++ b/2024/Relightable Neural Actor with Intrinsic Decomposition and Pose Control/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:439908a0b696287fbcc383d0c6c3639610c7d81278d0e0f11f73571707af3b58 +size 387380 diff --git a/2024/Relightable Neural Actor with Intrinsic Decomposition and Pose Control/layout.json b/2024/Relightable Neural Actor with Intrinsic Decomposition and Pose Control/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..e146967fc012f14c26b5031133e7fdf6e2931a88 --- /dev/null +++ b/2024/Relightable Neural Actor with Intrinsic Decomposition and Pose Control/layout.json @@ -0,0 +1,9572 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 169, + 111, + 445, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 169, + 111, + 445, + 148 + ], + "spans": [ + { + "bbox": [ + 169, + 111, + 445, + 148 + ], + "type": "text", + "content": "Relightable Neural Actor with Intrinsic Decomposition and Pose Control" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 136, + 167, + 477, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 167, + 477, + 193 + ], + "spans": [ + { + "bbox": [ + 136, + 167, + 477, + 193 + ], + "type": "text", + "content": "Diogo Carbonera Luvizon" + }, + { + "bbox": [ + 136, + 167, + 477, + 193 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 136, + 167, + 477, + 193 + ], + "type": "text", + "content": ", Vladislav Golyanik" + }, + { + "bbox": [ + 136, + 167, + 477, + 193 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 136, + 167, + 477, + 193 + ], + "type": "text", + "content": ", Adam Kortylewski" + }, + { + "bbox": [ + 136, + 167, + 477, + 193 + ], + "type": "inline_equation", + "content": "^{1,3}" + }, + { + "bbox": [ + 136, + 167, + 477, + 193 + ], + "type": "text", + "content": ", Marc Habermann" + }, + { + "bbox": [ + 136, + 167, + 477, + 193 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 136, + 167, + 477, + 193 + ], + "type": "text", + "content": ", and Christian Theobalt" + }, + { + "bbox": [ + 136, + 167, + 477, + 193 + ], + "type": "inline_equation", + "content": "^{1,2}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 156, + 201, + 457, + 235 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 156, + 201, + 457, + 235 + ], + "spans": [ + { + "bbox": [ + 156, + 201, + 457, + 235 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 156, + 201, + 457, + 235 + ], + "type": "text", + "content": " Max Planck Institute for Informatics, Saarland Informatics Campus \n" + }, + { + "bbox": [ + 156, + 201, + 457, + 235 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 156, + 201, + 457, + 235 + ], + "type": "text", + "content": " Saarbrücken Research Center for Visual Computing, Interaction and AI \n" + }, + { + "bbox": [ + 156, + 201, + 457, + 235 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 156, + 201, + 457, + 235 + ], + "type": "text", + "content": " University of Freiburg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 156, + 235, + 457, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 156, + 235, + 457, + 258 + ], + "spans": [ + { + "bbox": [ + 156, + 235, + 457, + 258 + ], + "type": "text", + "content": "{dluvizon,golyanik,akortyle,mhaberma,theobalt}@mpi-inf.mpg.de https://vcai.mpi-inf.mpg.de/projects/RNA" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 159, + 284, + 455, + 526 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 284, + 455, + 526 + ], + "spans": [ + { + "bbox": [ + 159, + 284, + 455, + 526 + ], + "type": "text", + "content": "Abstract. Creating a controllable and relightable digital avatar from multi-view video with fixed illumination is a very challenging problem since humans are highly articulated, creating pose-dependent appearance effects, and skin as well as clothing require space-varying BRDF modeling. Existing works on creating animatable avatars either do not focus on relighting at all, require controlled illumination setups, or try to recover a relightable avatar from very low cost setups, i.e. a single RGB video, at the cost of severely limited result quality, e.g. shadows not even being modeled. To address this, we propose Relightable Neural Actor, a new video-based method for learning a pose-driven neural human model that can be relighted, allows appearance editing, and models pose-dependent effects such as wrinkles and self-shadows. Importantly, for training, our method solely requires a multi-view recording of the human under a known, but static lighting condition. To tackle this challenging problem, we leverage an implicit geometry representation of the actor with a drivable density field that models pose-dependent deformations and derive a dynamic mapping between 3D and UV spaces, where normal, visibility, and materials are effectively encoded. To evaluate our approach in real-world scenarios, we collect a new dataset with four identities recorded under different light conditions, indoors and outdoors, providing the first benchmark of its kind for human relighting, and demonstrating state-of-the-art relighting results for novel human poses." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 545, + 230, + 558 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 545, + 230, + 558 + ], + "spans": [ + { + "bbox": [ + 132, + 545, + 230, + 558 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "type": "text", + "content": "Creating a relightable and controllable virtual actor of a real person solely from multi-view videos under static illumination is an exciting and challenging topic that has numerous applications in virtual and augmented reality, gaming, human-machine interaction, and content creation. However, it requires modeling the outgoing radiance of a particular surface point on the human, which depends on the pose-dependent and highly non-rigid surface deformation of the human body, the complex and spatially varying materials, as well as the scene lighting. Intrinsically decomposing these quantities solely from multi-view video and" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 132, + 113, + 294, + 218 + ], + "blocks": [ + { + "bbox": [ + 132, + 113, + 294, + 218 + ], + "lines": [ + { + "bbox": [ + 132, + 113, + 294, + 218 + ], + "spans": [ + { + "bbox": [ + 132, + 113, + 294, + 218 + ], + "type": "image", + "image_path": "1d796e40e690dcc337edceab4833f41ef7484c233c64e6670fc21aeb5b8c5b74.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 226, + 482, + 304 + ], + "lines": [ + { + "bbox": [ + 130, + 226, + 482, + 304 + ], + "spans": [ + { + "bbox": [ + 130, + 226, + 482, + 304 + ], + "type": "text", + "content": "Fig. 1: Our method learns a neural actor that is driven only by a 3D skeletal pose and allows rendering and editing of the actor's appearance under new lightning and poses not seen during training. Our approach models pose-dependent deformations, self-shadows, and performs intrinsic decomposition of dynamic humans through a pose-dependent mapping between 3D and UV spaces, which also enables editing the appearance and material properties at inference time. For training, our method only needs a multi-view video and an environment map." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 301, + 113, + 416, + 213 + ], + "blocks": [ + { + "bbox": [ + 301, + 113, + 416, + 213 + ], + "lines": [ + { + "bbox": [ + 301, + 113, + 416, + 213 + ], + "spans": [ + { + "bbox": [ + 301, + 113, + 416, + 213 + ], + "type": "image", + "image_path": "2bfcf18ed57fbca0ab339fc63250a8c06953c6e79673d39f9e744e693239c609.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 416, + 113, + 481, + 213 + ], + "blocks": [ + { + "bbox": [ + 416, + 113, + 481, + 213 + ], + "lines": [ + { + "bbox": [ + 416, + 113, + 481, + 213 + ], + "spans": [ + { + "bbox": [ + 416, + 113, + 481, + 213 + ], + "type": "image", + "image_path": "b172c1837958adb9144e48e7917212a58ddb6ca8959a316cc51fb31ea2cefa16.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 328, + 480, + 352 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 328, + 480, + 352 + ], + "spans": [ + { + "bbox": [ + 130, + 328, + 480, + 352 + ], + "type": "text", + "content": "enabling novel compositions that are of highly photorealistic quality defines an unsolved and challenging inverse problem in the literature." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 354, + 482, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 354, + 482, + 544 + ], + "spans": [ + { + "bbox": [ + 130, + 354, + 482, + 544 + ], + "type": "text", + "content": "Relighting static scenes from multi-view imagery is a well-studied problem [22,33,49,62]. However, directly applying such methods to moving humans, which are dynamic, highly articulated, and have complex material composition, is not trivially possible, as these approaches do not account for dynamics. On the contrary, some works have focused on building drivable and dynamic full-body avatars from multi-view video [32,42,54] without considering relightability at all. Despite some works [4,11,18,36,66] assuming a well-calibrated lightstage settings, only few human-specific works [6,17,64] try to achieve human intrinsic decomposition and relighting from videos with static lights. While they are able to recover a relightable human representation from a single RGB video, they can solely replay the human motion under novel lighting conditions [6], or do not model self-shadows and space-varying BRDF [17], which results in low visual fidelity. In stark contrast, we assume a multi-view setting while accounting for high-frequency surface deformations as well as induced high-frequency lighting effects such as self-shadows and shading, which are indispensable for a truly immersive experience." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 545, + 484, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 484, + 616 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 484, + 616 + ], + "type": "text", + "content": "To address this, we propose Relightable Neural Actor, a new method that achieves intrinsic decomposition of dynamic human avatars, enabling pose-control, material editing, and realistic rendering with pose-dependent effects, while solely requiring a multi-view video of the real human under a single and static light condition during training (see Fig. 1). At inference time, we only need an arbitrary driving skeletal pose and an environment map." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "We solve this challenging problem by leveraging an implicit density field from a pose-driven neural actor and by proposing a novel intrinsic decomposition guided by the implicit density field. This strategy allows our method to model pose-dependent deformations in the geometry as dynamic maps, while" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 245, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 245, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 245, + 100 + ], + "type": "text", + "content": "D.C. Luvizon et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 259 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 259 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 259 + ], + "type": "text", + "content": "representing the albedo and roughness properties of the actor as static maps. Our approach not only makes the problem computationally tractable for the long and dynamic human sequences, but also enables editing the neural actor's appearance and materials at inference time. Our key contribution is to sample normal and visibility from the posed density field and aggregate them into sparse UV maps, which are then densified by the proposed NormalNet and VisibilityNet models. This allows our approach to effectively model wrinkles and self-shadows even for long sequences. The albedo and roughness maps are optimized during training and sampled by the proposed UVDeltaNet, which accounts for pose-dependent deformations in the UV map and is intrinsically learned in our pipeline. Finally, given an environment map and a virtual camera position, our neural renderer produces the final image of the neural actor." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 261, + 482, + 320 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 261, + 482, + 320 + ], + "spans": [ + { + "bbox": [ + 130, + 261, + 482, + 320 + ], + "type": "text", + "content": "To train and evaluate our method, we collected a real-world dataset for novel pose and novel lighting human synthesis, composed of four actors recorded under six different light conditions by a multi-view camera setup, including indoor and outdoor environments and a light probe for each sequence, providing the first real in-the-wild benchmark for controllable human relighting." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 321, + 482, + 451 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 321, + 482, + 451 + ], + "spans": [ + { + "bbox": [ + 130, + 321, + 482, + 451 + ], + "type": "text", + "content": "To summarize, our main technical contributions are as follows: 1) The first method to learn a relightable neural actor from multi-view video under static lights that allows for pose control, appearance editing, and free-viewpoint rendering under arbitrary illumination while modeling pose-dependent effects. 2) We propose an intrinsic decomposition of diffuse and specular reflection based on a new approach to compute pose-dependent normals and visibility by sparsely sampling a neural field and densifying the results with NormalNet and VisibilityNet, making the problem of pose-driven human relighting tractable. Our approach relies on a mapping between 3D and UV space implicitly learned by UVDeltaNet, allowing us to learn an explicit and static material representation that is easy to edit at inference time." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 453, + 482, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 453, + 482, + 502 + ], + "spans": [ + { + "bbox": [ + 130, + 453, + 482, + 502 + ], + "type": "text", + "content": "In our experimental evaluation, we compare our approach with two baselines in the proposed real data benchmark, showing the superiority of our approach in synthesizing relightable humans on challenging motion sequences. The effectiveness of our method is demonstrated by ablation studies and qualitative results." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 521, + 238, + 534 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 521, + 238, + 534 + ], + "spans": [ + { + "bbox": [ + 131, + 521, + 238, + 534 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 550, + 482, + 588 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 550, + 482, + 588 + ], + "spans": [ + { + "bbox": [ + 130, + 550, + 482, + 588 + ], + "type": "text", + "content": "Since the goal of this work is to build a full-body relightable and animatable avatar, we will not discuss works [9, 10, 28, 44, 46, 47, 50, 56, 58, 59] focusing on face relighting as both settings have different requirements and challenges." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 606, + 284, + 619 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 606, + 284, + 619 + ], + "spans": [ + { + "bbox": [ + 131, + 606, + 284, + 619 + ], + "type": "text", + "content": "2.1 Relighting Static Scenes" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "text", + "content": "Classic approaches for relighting and editing perform intrinsic decomposition through inverse rendering [43]. With recent advances in neural rendering [51], light transportation can be taken into account with fully differentiable rendering" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 151, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 151, + 91, + 447, + 102 + ], + "type": "text", + "content": "Relightable Neural Actor with Intrinsic Decomposition and Pose Control" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 284 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 284 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 284 + ], + "type": "text", + "content": "pipelines. For instance, Zhang et al. [62] address the problem of recovering geometry and spatially-varying BRDF with an implicit scene representation. However, this requires solving the light transport by evaluating the NeRF model for each light ray, which is very inefficient. NeRV [49] and Zhang et al. [63] mitigate this problem with a neural visibility field that is jointly optimized with the NeRF model, and NeRD [5] adds further relaxation by ignoring cast shadows. However, this assumption cannot be made for humans, which are highly articulated and often experience prominent self-casting shadows. Jin et al. [22] propose TensorIR, an efficient tensor-based representation that solves the inverse problem by casting camera and secondary rays. IRON [60] proposes a two-stage process for obtaining a mesh with material properties, and Lyu et al. [35] and NeRO [33] focus on multi-bounce illumination and reflective objects, which is usually not the case for humans. The methods mentioned above share the assumption of a static scene and cannot be trivially adapted to dynamic and controllable humans." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 300, + 332, + 312 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 300, + 332, + 312 + ], + "spans": [ + { + "bbox": [ + 132, + 300, + 332, + 312 + ], + "type": "text", + "content": "2.2 Human Models with Pose Control" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 319, + 481, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 319, + 481, + 403 + ], + "spans": [ + { + "bbox": [ + 130, + 319, + 481, + 403 + ], + "type": "text", + "content": "Modeling humans is a long-standing topic. Traditionally, meshes from an actor can be automatically rigged to a skeleton [7] using Linear Blending Skinning (LBS) [27]. Parametric human models offer pose and shape control [34,41] and, more recently, implicit human bodies can be articulated [3, 37, 38]. Such models offer a generic and controllable representation, but are unable to represent individual details. Template-based methods synthesize a realistic level of details [14, 15, 21], but have to deal with non-linear pose-dependent deformations." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 403, + 482, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 403, + 482, + 559 + ], + "spans": [ + { + "bbox": [ + 130, + 403, + 482, + 559 + ], + "type": "text", + "content": "Several approaches model appearance and geometry with an implicit field anchored on a coarse human mesh. Neural Body [42] optimizes latent vectors anchored in the vertices of the human mesh and Neural Actor [32] incorporates texture maps as a conditioning to break down the mapping from pose to dynamic effects (one-to-many mapping). ARAH [54] models the human surface with an SDF, Dual-Space NeRF [65] learns the color and light in separate spaces, and Kwon et al. [24] propose an approach that generalizes to new identities and poses. Shyshey et al. [48] jointly learn textures and a neural renderer but fall short in reconstructing the 3D human body. A few works [20,57] approach a restrictive scenario of reconstructing 3D humans from a monocular video but fail to model fine details under new poses. Common to all these type of methods [12,13,25, 40] is the entanglement of material properties and illumination, making them restricted to the same appearance as observed during training." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 574, + 294, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 574, + 294, + 586 + ], + "spans": [ + { + "bbox": [ + 132, + 574, + 294, + 586 + ], + "type": "text", + "content": "2.3 Relighting Human Models" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": "Only a few methods in the literature tackle the problem of relighting humans from videos under static illumination. Li et al. [29] create a relightable human mesh by estimating a low-frequency environment using Spherical Harmonics. Some approaches target the problem of single-image [19] and static pose [64] relighting under unknown lights, which is an extremely ill-posed problem due to the high ambiguity between albedo and lighting. Relighting4D [6] extends" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 245, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 245, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 245, + 101 + ], + "type": "text", + "content": "D.C. Luvizon et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "content": "Neural Body [42] to estimate geometry and reflectance, but can only replay the same sequence observed during training under new lights and is not able to generate new poses. RANA [17] allows for pose control but assumes Lambertian reflectance, ignoring different material properties such as skin and clothes, and does not model cast shadows, resulting in unrealistic renderings for articulated humans. In contrast, our method performs intrinsic decomposition with space-varying BRDF, models illumination and self-shadows, allows for pose control and novel lights, and offers an easy manner to edit the material properties of the learned neural actor." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 242, + 202, + 255 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 242, + 202, + 255 + ], + "spans": [ + { + "bbox": [ + 131, + 242, + 202, + 255 + ], + "type": "text", + "content": "3 Method" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 268, + 482, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 268, + 482, + 376 + ], + "spans": [ + { + "bbox": [ + 130, + 268, + 482, + 376 + ], + "type": "text", + "content": "Our goal is to learn a digital human avatar (neural actor) from RGB images that models pose-dependent effects and can be relighted, pose-controlled, and rendered from novel viewpoints at test time. To achieve relighting, we decompose the observed appearance into normals, visibility with respect to each light source, and BRDF represented as albedo and roughness. With these components and an environment map, we employ the microfacet model [52] to render the output image of our method, which is compared to the observed images. In addition—since we target a drivable actor—the geometry components need to be controllable by the skeletal pose." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 376, + 482, + 591 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 376, + 482, + 591 + ], + "spans": [ + { + "bbox": [ + 130, + 376, + 482, + 591 + ], + "type": "text", + "content": "Fig. 2 shows an overview of our method divided into three main components. The first is a pose-driven implicit geometry model, where the goal is to obtain a skeleton-controllable implicit 3D density field that models the geometry of the actor with pose-dependent deformations. This model is adapted from Neural Actor [32] and is depicted here since our approach builds upon its internal design. Note that Neural Actor does not perform intrinsic decomposition and is unable to perform relighting or appearance editing, which is our main goal. The second component is our proposed intrinsic decomposition approach whose goal is to disentangle geometry represented as normal and visibility information, and material properties as albedo and roughness. Our key insight is to represent these components in UV space rather than in 3D coordinates, using the density field as guidance in the mapping function between 3D and UV spaces. This allows our method to model complex deformations that are pose-dependent in a compact UV map that can be easily computed and queried. Finally, we integrate the microfacet model [52] into our rendering pipeline that takes as input the disentangled components (i.e., normal, visibility, albedo, and roughness) and an environment map, and renders the image of the neural actor in an arbitrary pose and light condition from a free-viewpoint." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 609, + 228, + 620 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 609, + 228, + 620 + ], + "spans": [ + { + "bbox": [ + 131, + 609, + 228, + 620 + ], + "type": "text", + "content": "3.1 Preliminaries" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "type": "text", + "content": "Human Pose and Environment Map. Our method takes as input a 3D articulated human skeleton pose " + }, + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_t = \\{\\pmb {\\theta}_t,\\mathbf{R}_t,\\mathbf{T}_t\\}" + }, + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "type": "text", + "content": ", defined by the relative joint rotation " + }, + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "type": "inline_equation", + "content": "\\pmb {\\theta}\\in \\mathbb{R}^{J\\times 3}" + }, + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "type": "inline_equation", + "content": "J" + }, + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "type": "text", + "content": " body joints, the global rotation " + }, + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{R}\\in \\mathbb{R}^{3\\times 3}" + }, + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "type": "text", + "content": ", and global" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 151, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 151, + 91, + 447, + 102 + ], + "type": "text", + "content": "Relightable Neural Actor with Intrinsic Decomposition and Pose Control" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 115, + 482, + 237 + ], + "blocks": [ + { + "bbox": [ + 133, + 115, + 482, + 237 + ], + "lines": [ + { + "bbox": [ + 133, + 115, + 482, + 237 + ], + "spans": [ + { + "bbox": [ + 133, + 115, + 482, + 237 + ], + "type": "image", + "image_path": "4ba019d3bcb19d26ed0bef08e626dd22de5b901e5002eed5f4402d39d5e00ab3.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 256, + 482, + 334 + ], + "lines": [ + { + "bbox": [ + 130, + 256, + 482, + 334 + ], + "spans": [ + { + "bbox": [ + 130, + 256, + 482, + 334 + ], + "type": "text", + "content": "Fig. 2: Our method takes as input a 3D skeletal pose and a static environment map, and renders the neural actor from a virtual camera position. The pose-driven geometry (Sec. 3.1) learns an implicit density function, from which we compute the normal and visibility information. Our intrinsic decomposition disentangles normal, visibility, albedo and roughness maps in UV space (Sec. 3.2). The neural renderer (Sec. 3.3) outputs our prediction, which is supervised with the reference image. Yellow represents learnable components and green non-learnable components." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 356, + 482, + 429 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 356, + 482, + 429 + ], + "spans": [ + { + "bbox": [ + 130, + 356, + 482, + 429 + ], + "type": "text", + "content": "translation " + }, + { + "bbox": [ + 130, + 356, + 482, + 429 + ], + "type": "inline_equation", + "content": "\\mathbf{T} \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 130, + 356, + 482, + 429 + ], + "type": "text", + "content": " of the person at frame " + }, + { + "bbox": [ + 130, + 356, + 482, + 429 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 356, + 482, + 429 + ], + "type": "text", + "content": ". The 3D pose can be used to drive a coarse human mesh with linear blending skinning (LBS) [34]. We represent the global lights as an environment map " + }, + { + "bbox": [ + 130, + 356, + 482, + 429 + ], + "type": "inline_equation", + "content": "\\mathbf{L} \\in \\mathbb{R}^{16 \\times 32 \\times 3}" + }, + { + "bbox": [ + 130, + 356, + 482, + 429 + ], + "type": "text", + "content": ", which can be obtained from HDR images of a mirror sphere [8]—see the supplement for more details. The environment map can be static during training and changed during inference to render results under novel illumination." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 445, + 482, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 445, + 482, + 555 + ], + "spans": [ + { + "bbox": [ + 130, + 445, + 482, + 555 + ], + "type": "text", + "content": "Training Data. During training, our method is supervised with multi-view calibrated video streams " + }, + { + "bbox": [ + 130, + 445, + 482, + 555 + ], + "type": "inline_equation", + "content": "\\mathbf{V}_c = \\{\\mathbf{I}_{c,t}\\}" + }, + { + "bbox": [ + 130, + 445, + 482, + 555 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 445, + 482, + 555 + ], + "type": "inline_equation", + "content": "c \\in \\{1, \\dots, C\\}" + }, + { + "bbox": [ + 130, + 445, + 482, + 555 + ], + "type": "text", + "content": " is the camera index and " + }, + { + "bbox": [ + 130, + 445, + 482, + 555 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_{c,t}" + }, + { + "bbox": [ + 130, + 445, + 482, + 555 + ], + "type": "text", + "content": " is the image from camera " + }, + { + "bbox": [ + 130, + 445, + 482, + 555 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 130, + 445, + 482, + 555 + ], + "type": "text", + "content": " at frame " + }, + { + "bbox": [ + 130, + 445, + 482, + 555 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 445, + 482, + 555 + ], + "type": "text", + "content": ". We obtain a foreground mask " + }, + { + "bbox": [ + 130, + 445, + 482, + 555 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_{c,t}" + }, + { + "bbox": [ + 130, + 445, + 482, + 555 + ], + "type": "text", + "content": " for each video frame [30]. The coarse human mesh and 3D skeletons can be obtained with multi-view pose tracking [1]. The input 3D skeletons drive the LBS model to obtain posed human meshes used to extract texture maps from the multi-view images [2]. These texture maps are used during training to learn pose-dependent deformations on the implicit human model. At inference time, a V2V [55] network predicts texture maps from posed normal maps [32]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": "Implicit Human Model. For realistically relighting articulated avatars, we need to model human geometry with deformations caused by dynamic and varied poses. Therefore, we follow Neural Actor [32] to represent the pose-driven geometry as an implicit field guided by the 3D skeletal pose parameters and conditioned by texture maps, which provide the ability to model high-frequency deformations. Specifically, a standard NeRF model [39] is optimized through volumetric rendering in the canonical space of the coarse human mesh and the canonical space is deformed based on ResNet [16] features extracted from the" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 245, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 245, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 245, + 101 + ], + "type": "text", + "content": "D.C. Luvizon et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "text", + "content": "texture maps, which provide the ability to represent detailed wrinkles and deformations not modeled by the coarse mesh. During optimization, from each pixel in the reference images we cast a ray " + }, + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "inline_equation", + "content": "\\mathbf{r}_n = \\mathbf{o} + \\delta_n\\mathbf{d}" + }, + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "inline_equation", + "content": "n = 1,\\dots N" + }, + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "text", + "content": " is the index of each point in the ray, " + }, + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "inline_equation", + "content": "\\mathbf{o} \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "text", + "content": " is the origin defined in the optical center of the camera, " + }, + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "inline_equation", + "content": "\\mathbf{d} \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "text", + "content": " is the ray direction, and " + }, + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "inline_equation", + "content": "\\delta_{n}" + }, + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "text", + "content": " is the depth (distance to the camera). The sampled points " + }, + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "inline_equation", + "content": "\\mathbf{r}_n" + }, + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "text", + "content": " are projected to the mesh surface, resulting in the local coordinates " + }, + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "inline_equation", + "content": "(u_{n},v_{n},h_{n})" + }, + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "inline_equation", + "content": "(u,v)" + }, + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "text", + "content": " is the surface coordinate in the UV space and " + }, + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "text", + "content": " is the signed distance to the mesh." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 233, + 320, + 247 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 233, + 320, + 247 + ], + "spans": [ + { + "bbox": [ + 131, + 233, + 320, + 247 + ], + "type": "text", + "content": "3.2 Neural Intrinsic Decomposition" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 258, + 482, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 258, + 482, + 389 + ], + "spans": [ + { + "bbox": [ + 130, + 258, + 482, + 389 + ], + "type": "text", + "content": "At the core of our method, we perform intrinsic decomposition to obtain the normal, visibility, albedo, and roughness components, required for rendering the neural actor under new light conditions. Specifically, for each point in each ray cast into the scene we want to obtain the normal direction " + }, + { + "bbox": [ + 130, + 258, + 482, + 389 + ], + "type": "inline_equation", + "content": "\\mathbf{n} \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 130, + 258, + 482, + 389 + ], + "type": "text", + "content": ", the visibility coefficients " + }, + { + "bbox": [ + 130, + 258, + 482, + 389 + ], + "type": "inline_equation", + "content": "\\pmb{v} \\in \\mathbb{R}^{512}" + }, + { + "bbox": [ + 130, + 258, + 482, + 389 + ], + "type": "text", + "content": " respective to each light source in the environment map, the RGB albedo " + }, + { + "bbox": [ + 130, + 258, + 482, + 389 + ], + "type": "inline_equation", + "content": "\\mathbf{a} \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 130, + 258, + 482, + 389 + ], + "type": "text", + "content": ", and the scalar roughness " + }, + { + "bbox": [ + 130, + 258, + 482, + 389 + ], + "type": "inline_equation", + "content": "\\rho" + }, + { + "bbox": [ + 130, + 258, + 482, + 389 + ], + "type": "text", + "content": ". The normal and visibility values can be computed from the geometry model. However, computing the visibility of each point in 3D with respect to the light sources is computationally intractable for a dynamic and long sequence, since for every point in 3D and for every frame, several rays have to be cast towards each pixel of the environment map. Therefore, we propose a different approach." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 390, + 482, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 390, + 482, + 521 + ], + "spans": [ + { + "bbox": [ + 130, + 390, + 482, + 521 + ], + "type": "text", + "content": "We notice that the intrinsic decomposition is only relevant for points close to the surface. Since the coarse human mesh provides a mapping from the mesh surface to UV space and the implicit human model provides a density field that models the geometry of the actor, we propose an aggregation strategy that samples normal and visibility from a given point of view (therefore, only a few points are sampled in 3D) and aggregates the obtained values in a per-frame UV map. Once we obtain the normal and visibility in UV space for a given pose, we can simply query the values by projecting the sampled 3D point " + }, + { + "bbox": [ + 130, + 390, + 482, + 521 + ], + "type": "inline_equation", + "content": "\\mathbf{r}_n" + }, + { + "bbox": [ + 130, + 390, + 482, + 521 + ], + "type": "text", + "content": " into the UV space, preventing our method from the expensive computation of normal and visibility at every possible 3D coordinate over multiple frames. We next explain this approach in more detail." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 544, + 482, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 544, + 482, + 628 + ], + "spans": [ + { + "bbox": [ + 130, + 544, + 482, + 628 + ], + "type": "text", + "content": "Normal Maps. For each point in a ray, we want to obtain the normal value that will be required for solving the rendering equation during relighting (will be introduced in Sec. 3.3). This value could be obtained by computing the gradient from the neural field [62]. However, this process requires back-propagating gradients throughout the entire network and often results in noisy normals. Instead, we explicitly compute the normal vector close to the visible surface by deriving the unbiased depth " + }, + { + "bbox": [ + 130, + 544, + 482, + 628 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{D}}" + }, + { + "bbox": [ + 130, + 544, + 482, + 628 + ], + "type": "text", + "content": " [53]:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 182, + 638, + 481, + 669 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 182, + 638, + 481, + 669 + ], + "spans": [ + { + "bbox": [ + 182, + 638, + 481, + 669 + ], + "type": "interline_equation", + "content": "\\hat {\\mathbf {D}} (\\mathbf {r}) = \\frac {\\sum_ {n = 1} ^ {N} w _ {n} \\delta_ {n}}{\\sum_ {n = 1} ^ {N} w _ {n}}, \\quad \\text {w i t h} w _ {n} = T _ {n} \\left(1 - \\exp \\left(- \\sigma_ {n} \\bar {\\delta} _ {n}\\right)\\right), \\tag {1}", + "image_path": "4247b2abf0e73d346cef6a900910804caa29d331a1f0aed13b0bb3cc936aaec4.jpg" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 151, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 151, + 91, + 448, + 102 + ], + "type": "text", + "content": "Relightable Neural Actor with Intrinsic Decomposition and Pose Control" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 114, + 479, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 114, + 479, + 144 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 479, + 144 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 114, + 479, + 144 + ], + "type": "inline_equation", + "content": "T_{n} = \\exp \\left(-\\sum_{j = 1}^{n - 1}\\sigma_{j}\\bar{\\delta}_{j}\\right)" + }, + { + "bbox": [ + 130, + 114, + 479, + 144 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 114, + 479, + 144 + ], + "type": "inline_equation", + "content": "w_{n}" + }, + { + "bbox": [ + 130, + 114, + 479, + 144 + ], + "type": "text", + "content": " is the density weight from volumetric rendering [39], and " + }, + { + "bbox": [ + 130, + 114, + 479, + 144 + ], + "type": "inline_equation", + "content": "\\sigma_{n}" + }, + { + "bbox": [ + 130, + 114, + 479, + 144 + ], + "type": "text", + "content": " is the density predicted by the NeRF model at point " + }, + { + "bbox": [ + 130, + 114, + 479, + 144 + ], + "type": "inline_equation", + "content": "\\mathbf{r}_n" + }, + { + "bbox": [ + 130, + 114, + 479, + 144 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 144, + 298, + 262 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 144, + 298, + 262 + ], + "spans": [ + { + "bbox": [ + 130, + 144, + 298, + 262 + ], + "type": "text", + "content": "Then, as illustrated in Fig. 3, we convert the depth values relative to the camera to 3D coordinates and obtain the normal direction from the tangent surface computed at the intersection point, defined as the maximum value of the density weight " + }, + { + "bbox": [ + 130, + 144, + 298, + 262 + ], + "type": "inline_equation", + "content": "w_{n}" + }, + { + "bbox": [ + 130, + 144, + 298, + 262 + ], + "type": "text", + "content": ". This process can be performed for multiple viewpoints to cover the surface of the human." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 264, + 298, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 264, + 298, + 300 + ], + "spans": [ + { + "bbox": [ + 130, + 264, + 298, + 300 + ], + "type": "text", + "content": "Once the normal values are obtained in 3D, we project them to the UV map and aggregate them with the" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 300, + 480, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 300, + 480, + 324 + ], + "spans": [ + { + "bbox": [ + 130, + 300, + 480, + 324 + ], + "type": "text", + "content": "obtained normals from different views. In this process, the " + }, + { + "bbox": [ + 130, + 300, + 480, + 324 + ], + "type": "inline_equation", + "content": "(u^{*},v^{*})" + }, + { + "bbox": [ + 130, + 300, + 480, + 324 + ], + "type": "text", + "content": " coordinate where the obtained normal will be projected onto the UV map is defined by" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 306, + 153, + 380, + 214 + ], + "blocks": [ + { + "bbox": [ + 306, + 153, + 380, + 214 + ], + "lines": [ + { + "bbox": [ + 306, + 153, + 380, + 214 + ], + "spans": [ + { + "bbox": [ + 306, + 153, + 380, + 214 + ], + "type": "image", + "image_path": "3b45f527bf19b66200054d5f0242d5e8ba44f8178da9396b69d1696f62a4cf61.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 222, + 482, + 289 + ], + "lines": [ + { + "bbox": [ + 304, + 222, + 482, + 289 + ], + "spans": [ + { + "bbox": [ + 304, + 222, + 482, + 289 + ], + "type": "text", + "content": "Fig. 3: The unbiased depth from Eq. (1) is projected into 3D coordinates, where we approximate the tangent surface at the intersecting point and obtain the normal vector. Normals are sampled from multiple viewpoints and aggregated in the UV map." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 380, + 154, + 426, + 213 + ], + "blocks": [ + { + "bbox": [ + 380, + 154, + 426, + 213 + ], + "lines": [ + { + "bbox": [ + 380, + 154, + 426, + 213 + ], + "spans": [ + { + "bbox": [ + 380, + 154, + 426, + 213 + ], + "type": "image", + "image_path": "b1acf57fc92ded197bad1307778668bf79e3040b092aa4643c58511589a67fd5.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 427, + 154, + 477, + 213 + ], + "blocks": [ + { + "bbox": [ + 427, + 154, + 477, + 213 + ], + "lines": [ + { + "bbox": [ + 427, + 154, + 477, + 213 + ], + "spans": [ + { + "bbox": [ + 427, + 154, + 477, + 213 + ], + "type": "image", + "image_path": "a01c98550c215374a72a6f5002bd4bb3cd0459a6d0457b1852be1e8ec7b2426e.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 214, + 332, + 480, + 353 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 214, + 332, + 480, + 353 + ], + "spans": [ + { + "bbox": [ + 214, + 332, + 480, + 353 + ], + "type": "interline_equation", + "content": "\\left(u ^ {*}, v ^ {*}\\right) = \\Psi \\left(\\mathbf {r} _ {n} ^ {*}\\right), \\quad \\text {w i t h} \\mathbf {r} _ {n} ^ {*} = \\underset {\\mathbf {r} _ {n}} {\\arg \\max } w _ {n}, \\tag {2}", + "image_path": "bd9b7a542c67bb77ccdbe02e20d5e5c0f10b55ea3cf485023dca01592a8a464e.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 361, + 479, + 420 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 361, + 479, + 420 + ], + "spans": [ + { + "bbox": [ + 130, + 361, + 479, + 420 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 361, + 479, + 420 + ], + "type": "inline_equation", + "content": "\\varPsi" + }, + { + "bbox": [ + 130, + 361, + 479, + 420 + ], + "type": "text", + "content": " is a function that projects a 3D point to the closest surface point in the human mesh. Note that " + }, + { + "bbox": [ + 130, + 361, + 479, + 420 + ], + "type": "inline_equation", + "content": "w_{n}" + }, + { + "bbox": [ + 130, + 361, + 479, + 420 + ], + "type": "text", + "content": " depends on the sampled points " + }, + { + "bbox": [ + 130, + 361, + 479, + 420 + ], + "type": "inline_equation", + "content": "\\mathbf{r}_n" + }, + { + "bbox": [ + 130, + 361, + 479, + 420 + ], + "type": "text", + "content": " (from " + }, + { + "bbox": [ + 130, + 361, + 479, + 420 + ], + "type": "inline_equation", + "content": "T_{n}" + }, + { + "bbox": [ + 130, + 361, + 479, + 420 + ], + "type": "text", + "content": " in Eq. 1) and " + }, + { + "bbox": [ + 130, + 361, + 479, + 420 + ], + "type": "inline_equation", + "content": "(u^{*},v^{*})" + }, + { + "bbox": [ + 130, + 361, + 479, + 420 + ], + "type": "text", + "content": " corresponds to the " + }, + { + "bbox": [ + 130, + 361, + 479, + 420 + ], + "type": "inline_equation", + "content": "(u_{n},v_{n})" + }, + { + "bbox": [ + 130, + 361, + 479, + 420 + ], + "type": "text", + "content": " value obtained from the point " + }, + { + "bbox": [ + 130, + 361, + 479, + 420 + ], + "type": "inline_equation", + "content": "\\mathbf{r}_n" + }, + { + "bbox": [ + 130, + 361, + 479, + 420 + ], + "type": "text", + "content": " on the ray that has the largest density weight " + }, + { + "bbox": [ + 130, + 361, + 479, + 420 + ], + "type": "inline_equation", + "content": "w_{n}" + }, + { + "bbox": [ + 130, + 361, + 479, + 420 + ], + "type": "text", + "content": ". This process ensures that the normal map obtained aggregates the normals most likely close to the surface." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 421, + 480, + 492 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 421, + 480, + 492 + ], + "spans": [ + { + "bbox": [ + 130, + 421, + 480, + 492 + ], + "type": "text", + "content": "In this process, the resulting UV map can be sparse, which is undesirable for free-viewpoint rendering. Thus, the NormalNet is used to inpaint the missing normals in UV space. This lightweight CNN is implemented with 8 layers of partial convolutions [31], takes as input sparse normal maps with size " + }, + { + "bbox": [ + 130, + 421, + 480, + 492 + ], + "type": "inline_equation", + "content": "512 \\times 512 \\times 3" + }, + { + "bbox": [ + 130, + 421, + 480, + 492 + ], + "type": "text", + "content": ", and predicts a dense normal map (see Fig. 2). During training, NormalNet is supervised with densely sampled normal maps from the training data." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 510, + 485, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 485, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 485, + 666 + ], + "type": "text", + "content": "Visibility Maps. The visibility information is necessary to model pose-dependent shadows in the neural avatar, without which it is not possible to produce realistic relighting results. A naive approach for obtaining the visibility for each point in 3D would require querying the neural field thousands of times while casting rays from the intersecting point to all the light sources, which is prohibitive during training. In contrast, in our approach we efficiently encode visibility of each point in 3D close to the surface by applying an approximate mapping to UV space. Specifically, we pre-compute the visibility maps for each pose by sampling multiple camera views and aggregate the results in UV map following Eq. (2). The resulting visibility maps with size " + }, + { + "bbox": [ + 130, + 510, + 485, + 666 + ], + "type": "inline_equation", + "content": "256 \\times 256 \\times 512" + }, + { + "bbox": [ + 130, + 510, + 485, + 666 + ], + "type": "text", + "content": " encode the visibility information w.r.t. each light source in the environment map (represented by 512 individual lights). Similarly to the computation of normal maps, this process results in sparse visibility maps, which are inpainted by the VisibilityNet (as" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 246, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 246, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 246, + 101 + ], + "type": "text", + "content": "D.C. Luvizon et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "text", + "content": "shown in Fig. 2). This shallow CNN takes as input sparse visibility and normal maps, as the visibility highly depends on normals (e.g. the visibility is always zero for the backlight source), and is trained with densely sampled visibility from the training sequence. See the supplement for architecture details." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 178, + 480, + 285 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 178, + 480, + 285 + ], + "spans": [ + { + "bbox": [ + 130, + 178, + 480, + 285 + ], + "type": "text", + "content": "Albedo and Roughness Maps. As discussed above, the normal and visibility components can be computed from the density field by using UV maps as proxies. However, the albedo and roughness properties cannot be directly obtained from the pre-trained Neural Actor model, since it predicts radiance which is the product of the rendering equation. Instead, we want to learn a decomposed representation. In intrinsic decomposition, the material properties are commonly optimized via differentiable rendering. Hence, we optimize the albedo and roughness of the neural actor by leveraging a differentiable renderer (Sec. 3.3) and by imposing priors through our design." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 285, + 480, + 352 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 285, + 480, + 352 + ], + "spans": [ + { + "bbox": [ + 130, + 285, + 480, + 352 + ], + "type": "text", + "content": "The material properties of an actor should not depend on the pose and should be efficiently encoded to enable fast computations and editing. Therefore, we model albedo and roughness as static UV maps, respectively defined as " + }, + { + "bbox": [ + 130, + 285, + 480, + 352 + ], + "type": "inline_equation", + "content": "\\mathbf{A} \\in \\mathbb{R}^{1024 \\times 1024 \\times 3}" + }, + { + "bbox": [ + 130, + 285, + 480, + 352 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 285, + 480, + 352 + ], + "type": "inline_equation", + "content": "\\mathbf{B} \\in \\mathbb{R}^{1024 \\times 1024}" + }, + { + "bbox": [ + 130, + 285, + 480, + 352 + ], + "type": "text", + "content": ". These tensors are jointly optimized for the whole training sequence and regularized with an L1 loss on the spatial gradients, which encourages smoothness in both maps:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 189, + 356, + 480, + 375 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 189, + 356, + 480, + 375 + ], + "spans": [ + { + "bbox": [ + 189, + 356, + 480, + 375 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {s m o o t h}} = \\frac {1}{N} \\sum \\left| \\nabla_ {x} (\\mathbf {A}) \\right| + \\left| \\nabla_ {y} (\\mathbf {A}) \\right| + \\left| \\nabla_ {x} (\\mathbf {B}) \\right| + \\left| \\nabla_ {y} (\\mathbf {B}) \\right|, \\tag {3}", + "image_path": "2e5f5a814efd85f26c42b17b75c85699b9b2e5a6e61775e4555454fd707f3faf.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 134, + 378, + 394, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 378, + 394, + 389 + ], + "spans": [ + { + "bbox": [ + 134, + 378, + 394, + 389 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 134, + 378, + 394, + 389 + ], + "type": "inline_equation", + "content": "\\nabla_{x}" + }, + { + "bbox": [ + 134, + 378, + 394, + 389 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 134, + 378, + 394, + 389 + ], + "type": "inline_equation", + "content": "\\nabla_{y}" + }, + { + "bbox": [ + 134, + 378, + 394, + 389 + ], + "type": "text", + "content": " represent the spatial gradient operators." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 390, + 481, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 390, + 481, + 509 + ], + "spans": [ + { + "bbox": [ + 130, + 390, + 481, + 509 + ], + "type": "text", + "content": "Nonetheless, by assuming a material representation that does not depend on the skeletal pose, we still have to cope with intrinsic mesh registration errors. Such registration errors could cause blurry results or artifacts in the final renderings. To cope with this problem, we propose a corrective term to the UV mapping, which is predicted by the UVDeltaNet, as shown in Fig. 2. The UVDeltaNet is an MLP that takes as input the local coordinates " + }, + { + "bbox": [ + 130, + 390, + 481, + 509 + ], + "type": "inline_equation", + "content": "(u,v,h)" + }, + { + "bbox": [ + 130, + 390, + 481, + 509 + ], + "type": "text", + "content": " for each point on the ray and the sampled texture features " + }, + { + "bbox": [ + 130, + 390, + 481, + 509 + ], + "type": "inline_equation", + "content": "\\psi" + }, + { + "bbox": [ + 130, + 390, + 481, + 509 + ], + "type": "text", + "content": ", and outputs a corrective term " + }, + { + "bbox": [ + 130, + 390, + 481, + 509 + ], + "type": "inline_equation", + "content": "\\epsilon = (u',v')" + }, + { + "bbox": [ + 130, + 390, + 481, + 509 + ], + "type": "text", + "content": ". The texture features are necessary to provide additional information, since the local coordinates alone cannot account for deviations in the mesh registration." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 510, + 479, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 479, + 544 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 479, + 544 + ], + "type": "text", + "content": "Initially, UVDeltaNet is frozen and outputs zeros. After the initial convergence of " + }, + { + "bbox": [ + 130, + 510, + 479, + 544 + ], + "type": "inline_equation", + "content": "\\mathbf{A}" + }, + { + "bbox": [ + 130, + 510, + 479, + 544 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 510, + 479, + 544 + ], + "type": "inline_equation", + "content": "\\mathbf{B}" + }, + { + "bbox": [ + 130, + 510, + 479, + 544 + ], + "type": "text", + "content": ", UVDeltaNet is indirectly supervised by the rendering loss and regularized with:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 228, + 550, + 480, + 572 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 228, + 550, + 480, + 572 + ], + "spans": [ + { + "bbox": [ + 228, + 550, + 480, + 572 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {u v}} = \\frac {1}{N} \\sum \\epsilon^ {2} + | \\nabla_ {x} (\\epsilon) | + | \\nabla_ {y} (\\epsilon) |, \\tag {4}", + "image_path": "2eb9e76dfd8920fbd3b2e4b56a33cd300e68c946a7ea7a8dd4dbb29a051ef20c.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 575, + 479, + 599 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 575, + 479, + 599 + ], + "spans": [ + { + "bbox": [ + 130, + 575, + 479, + 599 + ], + "type": "text", + "content": "which minimizes the magnitude of the deformation and encourages its spatial gradient to be sparse, i.e., it makes the UV deformation spatially coherent." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 131, + 613, + 246, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 613, + 246, + 624 + ], + "spans": [ + { + "bbox": [ + 131, + 613, + 246, + 624 + ], + "type": "text", + "content": "3.3 Neural Renderer" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 629, + 479, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 479, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 479, + 665 + ], + "type": "text", + "content": "The third part of our method is our neural renderer, which takes the disentangled components as normal, visibility, albedo, and roughness; integrates the lights from the environment map, and outputs the rendered image." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 151, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 151, + 91, + 448, + 102 + ], + "type": "text", + "content": "Relightable Neural Actor with Intrinsic Decomposition and Pose Control" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "text", + "content": "Our neural renderer starts by recasting the ray " + }, + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "inline_equation", + "content": "\\mathbf{r}" + }, + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "text", + "content": " with importance sampling around the estimated depth value " + }, + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{D}}" + }, + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "text", + "content": ". Thus, all the points " + }, + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "inline_equation", + "content": "\\mathbf{r}_n" + }, + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "text", + "content": " are most likely close to the surface. For each sampled point in the ray, we obtain the normal direction " + }, + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "inline_equation", + "content": "\\mathbf{n}_n" + }, + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "text", + "content": " and the visibility values " + }, + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "inline_equation", + "content": "\\pmb{v}_n" + }, + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "text", + "content": " by interpolating the normal and visibility maps in UV around the point " + }, + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "inline_equation", + "content": "(u_n,v_n) = \\varPsi(\\mathbf{r}_n)" + }, + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "text", + "content": ". The albedo " + }, + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "inline_equation", + "content": "\\mathbf{a}_n" + }, + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "text", + "content": " and roughness " + }, + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "inline_equation", + "content": "\\rho_n" + }, + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "text", + "content": " are obtained by interpolating the albedo and roughness maps in UV around the point " + }, + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "inline_equation", + "content": "(u_n + u'_u,v_n + v'_n)" + }, + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "text", + "content": ", i.e., following the deformation predicted by UVDeltaNet." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 200, + 481, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 200, + 481, + 224 + ], + "spans": [ + { + "bbox": [ + 130, + 200, + 481, + 224 + ], + "type": "text", + "content": "Before rendering the final image of the posed neural actor, we consider the light transport equation defined as" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 195, + 233, + 481, + 257 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 195, + 233, + 481, + 257 + ], + "spans": [ + { + "bbox": [ + 195, + 233, + 481, + 257 + ], + "type": "interline_equation", + "content": "\\mathbf {L} _ {0} (\\mathbf {r} _ {n}, \\boldsymbol {\\omega} _ {0}) = \\sum_ {i} \\operatorname {R} (\\mathbf {a} _ {n}, \\rho_ {n}, \\boldsymbol {\\omega} _ {i}, \\boldsymbol {\\omega} _ {0}) \\boldsymbol {v} _ {n, i} \\mathbf {L} (\\boldsymbol {\\omega} _ {i}) (\\boldsymbol {\\omega} _ {i} \\mathbf {n} _ {n}), \\tag {5}", + "image_path": "30820e1abc17af2d5c31dd0554406afdf0fd76076c28fa3d39b3d4b4047fe3c0.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 265, + 482, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 265, + 482, + 324 + ], + "spans": [ + { + "bbox": [ + 130, + 265, + 482, + 324 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 265, + 482, + 324 + ], + "type": "inline_equation", + "content": "\\mathrm{R}(\\cdot)" + }, + { + "bbox": [ + 130, + 265, + 482, + 324 + ], + "type": "text", + "content": " is the microfacet BRDF model [52], " + }, + { + "bbox": [ + 130, + 265, + 482, + 324 + ], + "type": "inline_equation", + "content": "\\omega_0" + }, + { + "bbox": [ + 130, + 265, + 482, + 324 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 265, + 482, + 324 + ], + "type": "inline_equation", + "content": "\\omega_{i}" + }, + { + "bbox": [ + 130, + 265, + 482, + 324 + ], + "type": "text", + "content": " are the camera and the light source directions, respectively, " + }, + { + "bbox": [ + 130, + 265, + 482, + 324 + ], + "type": "inline_equation", + "content": "\\pmb{v}_{n,i}" + }, + { + "bbox": [ + 130, + 265, + 482, + 324 + ], + "type": "text", + "content": " is the visibility of point " + }, + { + "bbox": [ + 130, + 265, + 482, + 324 + ], + "type": "inline_equation", + "content": "\\mathbf{r}_n" + }, + { + "bbox": [ + 130, + 265, + 482, + 324 + ], + "type": "text", + "content": " w.r.t. the ith light source, and " + }, + { + "bbox": [ + 130, + 265, + 482, + 324 + ], + "type": "inline_equation", + "content": "\\mathbf{L}(\\boldsymbol {\\omega}_i)" + }, + { + "bbox": [ + 130, + 265, + 482, + 324 + ], + "type": "text", + "content": " is the incoming light from the environment map. Eq. (5) is defined for a single point " + }, + { + "bbox": [ + 130, + 265, + 482, + 324 + ], + "type": "inline_equation", + "content": "\\mathbf{r}_n" + }, + { + "bbox": [ + 130, + 265, + 482, + 324 + ], + "type": "text", + "content": " on the ray and is integrated into our volume rendering as" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 253, + 325, + 481, + 357 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 253, + 325, + 481, + 357 + ], + "spans": [ + { + "bbox": [ + 253, + 325, + 481, + 357 + ], + "type": "interline_equation", + "content": "\\tilde {\\mathbf {I}} (\\mathbf {r}) = \\sum_ {n = 1} ^ {N} w _ {n} \\mathbf {L} _ {0} \\left(\\mathbf {r} _ {n}, \\omega_ {0}\\right), \\tag {6}", + "image_path": "effb2e352dd5d66b51480f5230a988ac950d77a24e86ce671163751cfce78419.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 362, + 385, + 374 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 362, + 385, + 374 + ], + "spans": [ + { + "bbox": [ + 131, + 362, + 385, + 374 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 362, + 385, + 374 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{I}}" + }, + { + "bbox": [ + 131, + 362, + 385, + 374 + ], + "type": "text", + "content": " is the output image of the relighted neural actor." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 146, + 375, + 477, + 387 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 375, + 477, + 387 + ], + "spans": [ + { + "bbox": [ + 146, + 375, + 477, + 387 + ], + "type": "text", + "content": "During training, our relightable model is supervised with the following loss:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 190, + 397, + 481, + 410 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 190, + 397, + 481, + 410 + ], + "spans": [ + { + "bbox": [ + 190, + 397, + 481, + 410 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {r e l i g h t}} = \\lambda_ {\\mathrm {L 2}} \\mathcal {L} _ {\\mathrm {L 2}} + \\lambda_ {\\text {v g g}} \\mathcal {L} _ {\\text {v g g}} + \\lambda_ {\\text {s m o o t h}} \\mathcal {L} _ {\\text {s m o o t h}} + \\mathcal {L} _ {\\text {u v}}, \\tag {7}", + "image_path": "8ed5e3df67419f4d524020a65d69b46b429fdc6362ad78e376f685d2a9f2ee3c.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 418, + 482, + 455 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 418, + 482, + 455 + ], + "spans": [ + { + "bbox": [ + 130, + 418, + 482, + 455 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 418, + 482, + 455 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{L2}}" + }, + { + "bbox": [ + 130, + 418, + 482, + 455 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 418, + 482, + 455 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{vgg}}" + }, + { + "bbox": [ + 130, + 418, + 482, + 455 + ], + "type": "text", + "content": " are the L2 and the perceptual losses [26] between " + }, + { + "bbox": [ + 130, + 418, + 482, + 455 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{I}} (\\mathbf{r})" + }, + { + "bbox": [ + 130, + 418, + 482, + 455 + ], + "type": "text", + "content": " and the reference image. The hyperparameters are defined as " + }, + { + "bbox": [ + 130, + 418, + 482, + 455 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{L2}} = 100" + }, + { + "bbox": [ + 130, + 418, + 482, + 455 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 418, + 482, + 455 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{vgg}} = 0.01" + }, + { + "bbox": [ + 130, + 418, + 482, + 455 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 418, + 482, + 455 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{smooth}} = 0.1" + }, + { + "bbox": [ + 130, + 418, + 482, + 455 + ], + "type": "text", + "content": " in all our experiments." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 131, + 472, + 371, + 487 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 472, + 371, + 487 + ], + "spans": [ + { + "bbox": [ + 131, + 472, + 371, + 487 + ], + "type": "text", + "content": "4 Relightable Dynamic Actors Dataset" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 498, + 482, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 498, + 482, + 569 + ], + "spans": [ + { + "bbox": [ + 130, + 498, + 482, + 569 + ], + "type": "text", + "content": "Existing datasets for human reenactment and relighting are often monocular [2, 17], contain very short sequences with only a few pose variations, and do not provide environment maps [2,17,32,42]. Hence, we cannot use them for evaluation under realistic scenarios with real ground truth. In addition, we observe that the only existing synthetic dataset for human relighting and pose control [17] suffers from artifacts in the geometry and salt-and-pepper noise in the renderings." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "text", + "content": "With these shortcomings in mind, we collect a new real dataset, which we call Relightable Dynamic Actors, with four identities performing a series of ten different actions (Fig. 4). Each subject is recorded six times wearing the same clothes but under different light conditions, including indoor and outdoor, where we used around 100 and eight cameras, respectively. For each sequence, we also capture HDR light probes, which can be used for training or testing. In total, our dataset contains around 90k multi-view video frames divided into 24 multi-view video sequences. We refer to the supplemental document for more details." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 245, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 245, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 245, + 101 + ], + "type": "text", + "content": "D.C. Luvizon et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 132, + 114, + 306, + 214 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 306, + 214 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 306, + 214 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 306, + 214 + ], + "type": "image", + "image_path": "a06b68f5e197c3c761b0ca2e1bf0c0a96d8720e1eb5027a3c6ac90444f51c1f3.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 232, + 482, + 266 + ], + "lines": [ + { + "bbox": [ + 131, + 232, + 482, + 266 + ], + "spans": [ + { + "bbox": [ + 131, + 232, + 482, + 266 + ], + "type": "text", + "content": "Fig. 4: Relightable Dynamic Actors dataset. For each actor, we recorded six sequences under different light conditions, including five indoor sequences and one outdoor. Environment maps were obtained from HDR pictures of a mirror sphere." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 307, + 114, + 480, + 214 + ], + "blocks": [ + { + "bbox": [ + 307, + 114, + 480, + 214 + ], + "lines": [ + { + "bbox": [ + 307, + 114, + 480, + 214 + ], + "spans": [ + { + "bbox": [ + 307, + 114, + 480, + 214 + ], + "type": "image", + "image_path": "ea85ef744f3f6e88d203721ec7d31977792d3e0697e883150fee5d73e31dc5d1.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 289, + 230, + 303 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 289, + 230, + 303 + ], + "spans": [ + { + "bbox": [ + 132, + 289, + 230, + 303 + ], + "type": "text", + "content": "5 Experiments" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 315, + 479, + 352 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 315, + 479, + 352 + ], + "spans": [ + { + "bbox": [ + 131, + 315, + 479, + 352 + ], + "type": "text", + "content": "In this section, we provide an experimental evaluation of our method on the Reli-ightable Dynamic Actors dataset and show qualitative and quantitative results. Please refer to the supplementary video for dynamic results on long sequences." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 369, + 215, + 380 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 369, + 215, + 380 + ], + "spans": [ + { + "bbox": [ + 132, + 369, + 215, + 380 + ], + "type": "text", + "content": "5.1 Evaluation" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 390, + 480, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 390, + 480, + 449 + ], + "spans": [ + { + "bbox": [ + 130, + 390, + 480, + 449 + ], + "type": "text", + "content": "Metrics. For quantitative evaluation, we report results on the PSNR, SSIM and LPIPS metrics [61], comparing the rendered relight results with the groundtruth images from our dataset. Importantly, our work is the first to be able to report such metrics on a real dataset since no previous dataset in the literature provided a real human recorded under different and known light conditions." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 450, + 482, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 450, + 482, + 605 + ], + "spans": [ + { + "bbox": [ + 131, + 450, + 482, + 605 + ], + "type": "text", + "content": "Baselines. Since the only method in the literature capable of reposing a human under new light conditions does not provide implementation [17], and Relighting4D [6] is not able to perform inference for new poses, we defined two baselines for a fair comparison. Our baselines are derived from Neural Actor [32] and adapted to perform intrinsic decomposition based on the obtained normal and visibility information, as discussed in Sec. 3.2. However, instead of using NormalNet and VisibilityNet to obtain these components, we densely sample the neural field and perform morphological inpainting to obtain dense UV maps. The first baseline, referred to as Neural Actor " + }, + { + "bbox": [ + 131, + 450, + 482, + 605 + ], + "type": "inline_equation", + "content": "+IR" + }, + { + "bbox": [ + 131, + 450, + 482, + 605 + ], + "type": "text", + "content": ", predicts the albedo and roughness components instead of RGB color directly from the deformed canonical space " + }, + { + "bbox": [ + 131, + 450, + 482, + 605 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_n = (x,y,z)" + }, + { + "bbox": [ + 131, + 450, + 482, + 605 + ], + "type": "text", + "content": " (see Fig. 2). The second baseline, Neural Actor " + }, + { + "bbox": [ + 131, + 450, + 482, + 605 + ], + "type": "inline_equation", + "content": "+IR + Tex" + }, + { + "bbox": [ + 131, + 450, + 482, + 605 + ], + "type": "text", + "content": ", takes as input the texture features " + }, + { + "bbox": [ + 131, + 450, + 482, + 605 + ], + "type": "inline_equation", + "content": "\\psi" + }, + { + "bbox": [ + 131, + 450, + 482, + 605 + ], + "type": "text", + "content": " as a conditioning for predicting the albedo and roughness values. In both cases, the predictor is a NeRF MLP with 8 layers." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 606, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 606, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 131, + 606, + 482, + 666 + ], + "type": "text", + "content": "Implementation Details. Our model is trained in a three-stage process. First, we train the geometry component (Sec. 3.1) with the Adam [23] optimizer with a learning rate of 0.001 until convergence, which takes about two days on four NVidia A40 GPUs. Then, we pre-train the NormalNet and VisibilityNet with pre-computed normal and visibility UV maps sampled from the geometry model" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 151, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 151, + 91, + 447, + 102 + ], + "type": "text", + "content": "Relightable Neural Actor with Intrinsic Decomposition and Pose Control" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 178, + 114, + 436, + 226 + ], + "blocks": [ + { + "bbox": [ + 178, + 114, + 436, + 226 + ], + "lines": [ + { + "bbox": [ + 178, + 114, + 436, + 226 + ], + "spans": [ + { + "bbox": [ + 178, + 114, + 436, + 226 + ], + "type": "image", + "image_path": "d89fd9d77195b1decc66845928fbb1097f4a8947c98467a015c03010339021a3.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 228, + 480, + 260 + ], + "lines": [ + { + "bbox": [ + 130, + 228, + 480, + 260 + ], + "spans": [ + { + "bbox": [ + 130, + 228, + 480, + 260 + ], + "type": "text", + "content": "Fig. 5: Qualitative results of our method on novel poses. Our approach generalizes to both indoor and outdoor illumination, and generates realistic renderings for arbitrary environment maps with detailed shadows and pose-dependent deformations." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 282, + 480, + 378 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 282, + 480, + 378 + ], + "spans": [ + { + "bbox": [ + 130, + 282, + 480, + 378 + ], + "type": "text", + "content": "from 40 different views. Both networks are trained in parallel using Adam with a learning rate of 0.01 for about 12 hours. Finally, we train our intrinsic decomposition model with the same optimizer until convergence, which takes about one day. Since the obtained environment maps are up to factor of scale [8], we apply global color correction during training and evaluation. The predicted texture maps are obtained with a V2V model [55] trained (in parallel to our model) to translate a sequence of SMPL normal maps into a sequence of texture maps. Please refer to our supplementary materials for architecture details." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 394, + 197, + 405 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 394, + 197, + 405 + ], + "spans": [ + { + "bbox": [ + 132, + 394, + 197, + 405 + ], + "type": "text", + "content": "5.2 Results" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 413, + 480, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 413, + 480, + 496 + ], + "spans": [ + { + "bbox": [ + 130, + 413, + 480, + 496 + ], + "type": "text", + "content": "Qualitative Results. In Fig. 5, we show qualitative results of our method. Our results are all from a trained model driven by a new skeletal pose, new camera viewpoint, and new environment map not seen during training. Note how our approach produces realistic appearance and prominent cloth wrinkles, which can be easily observed under different light conditions. Importantly, notice how our results look realistic under novel lights, even though our model is trained on a single sequence under a static light condition." + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 132, + 559, + 481, + 602 + ], + "blocks": [ + { + "bbox": [ + 130, + 516, + 480, + 550 + ], + "lines": [ + { + "bbox": [ + 130, + 516, + 480, + 550 + ], + "spans": [ + { + "bbox": [ + 130, + 516, + 480, + 550 + ], + "type": "text", + "content": "Table 1: Quantitative results from our method on real data, subject S1, considering indoor and outdoor scenes. Our method consistently improves the baselines, especially in outdoor scenarios, where the gap between training and test lights is more pronounced." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 132, + 559, + 481, + 602 + ], + "lines": [ + { + "bbox": [ + 132, + 559, + 481, + 602 + ], + "spans": [ + { + "bbox": [ + 132, + 559, + 481, + 602 + ], + "type": "table", + "html": "
MethodT2T3T4T5T6 (outdoors)
PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓
Neural Actor+IR17.5220.7500.18117.4730.7680.17817.7450.7670.17118.2440.7680.17518.3550.7100.203
Neural Actor+IR+Tex18.1750.7780.16417.7990.7900.16718.1320.7850.16718.5310.7890.16618.7040.7370.186
Ours18.6020.7920.16318.2400.8000.16918.3880.8000.16418.8200.8000.16519.4610.7530.173
", + "image_path": "579856259c9e3880467056a3905e461147e4ed8cd613e0b345aac11cc8301087.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 615, + 480, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 615, + 480, + 662 + ], + "spans": [ + { + "bbox": [ + 130, + 615, + 480, + 662 + ], + "type": "text", + "content": "In Fig. 6, we show results from our intrinsic decomposition. Note that our normal representation nicely captures the geometry details and the visibility. This is important to obtain an albedo that does not contain baked-in lighting effects, e.g. shadows, as can be observed from our results." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 245, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 245, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 245, + 100 + ], + "type": "text", + "content": "D.C. Luvizon et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 176, + 113, + 438, + 264 + ], + "blocks": [ + { + "bbox": [ + 176, + 113, + 438, + 264 + ], + "lines": [ + { + "bbox": [ + 176, + 113, + 438, + 264 + ], + "spans": [ + { + "bbox": [ + 176, + 113, + 438, + 264 + ], + "type": "image", + "image_path": "0dd55a64e873c05df6a67eec8ba9595a19834eecddf6d8249c71aa9f30b6c696.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 264, + 480, + 298 + ], + "lines": [ + { + "bbox": [ + 130, + 264, + 480, + 298 + ], + "spans": [ + { + "bbox": [ + 130, + 264, + 480, + 298 + ], + "type": "text", + "content": "Fig. 6: Results from our intrinsic decomposition. Our method provides detailed normals and visibility (averaged for visualization), and plausible albedo and roughness predictions that enable realistic relighting." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 326, + 332, + 409 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 326, + 332, + 409 + ], + "spans": [ + { + "bbox": [ + 130, + 326, + 332, + 409 + ], + "type": "text", + "content": "Another important and unique characteristic of our method is the ability to model self-shadows and cloth wrinkles. In Fig. 7, we show results from two subjects, where a \"sunrise\" light probe was used to produce cast shadows. Note how pose-dependent effects are consistent with the mostly directional illumination." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 411, + 332, + 531 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 411, + 332, + 531 + ], + "spans": [ + { + "bbox": [ + 130, + 411, + 332, + 531 + ], + "type": "text", + "content": "Comparisons. We evaluated our method on the test sequences from our dataset. Specifically, we compare against two baselines in the indoor sequences (T2-T5) and outdoor sequence (T6) of subject S1 in every 10th frame, which sums up to 4000 evaluation frames. In Tab. 1 we show, for the first time, quantitative results of real human relighting under new poses and new light conditions. Our approach consistently improves upon the two baselines." + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 344, + 304, + 479, + 471 + ], + "blocks": [ + { + "bbox": [ + 344, + 304, + 479, + 471 + ], + "lines": [ + { + "bbox": [ + 344, + 304, + 479, + 471 + ], + "spans": [ + { + "bbox": [ + 344, + 304, + 479, + 471 + ], + "type": "image", + "image_path": "5b5fbd96a5687646854daa587d130d438fb3408cdb371bc3824a1159e27170b2.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 338, + 483, + 482, + 528 + ], + "lines": [ + { + "bbox": [ + 338, + 483, + 482, + 528 + ], + "spans": [ + { + "bbox": [ + 338, + 483, + 482, + 528 + ], + "type": "text", + "content": "Fig. 7: Our method models pose-dependent effects and cast shadows, as can be observed from our results under the \"sunrise\" light." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": "Ablation. In Tab. 2, we show an ablation with different components of our method. Compared to our baseline (Neural Actor+IR), learning the albedo and roughness as UV maps guided by the neural field significantly improves the performance on relighting and reposing. This is because, in the baselines, the albedo and roughness are under-constrained and can possibly embed shading effects, overfitting to the training data. The NormalNet and VisibilityNet provide additional capacity to the model to learn how to inpaint the sampled UV maps. Note that without these two models our results are computed using morphological region growth in the UV maps. Finally, the UVDeltaNet provides more flexibility to the albedo and roughness representation, accounting for fine details. This can be observed in the improvement in the LPIPS metric in the last" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 151, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 151, + 91, + 447, + 102 + ], + "type": "text", + "content": "Relightable Neural Actor with Intrinsic Decomposition and Pose Control" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": "row of Tab. 2. Nonetheless, this also means more capacity to overfit the training sequences, which translates into a small decrease in the PSNR metric." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 159, + 224, + 171 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 159, + 224, + 171 + ], + "spans": [ + { + "bbox": [ + 132, + 159, + 224, + 171 + ], + "type": "text", + "content": "5.3 Applications" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 179, + 264, + 310 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 179, + 264, + 310 + ], + "spans": [ + { + "bbox": [ + 130, + 179, + 264, + 310 + ], + "type": "text", + "content": "An important characteristic of our approach is our static representation in UV for the albedo and roughness. This allows exciting and new applications such as color and material editing of an animatable and relightable neural actor, as shown in Fig. 8 and in Fig. 1. With a simple edit in the obtained albedo or rough-" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 269, + 155, + 482, + 242 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 269, + 155, + 482, + 242 + ], + "spans": [ + { + "bbox": [ + 269, + 155, + 482, + 242 + ], + "type": "text", + "content": "Table 2: Ablation study with different components of our method. The proposed intrinsic decomposition with static albedo and roughness maps significantly improves over the baseline, and the proposed NormalNet and VisibilityNet further improve the results while enabling free-viewpoint rendering. UVDeltaNet provides more realistic results (LPIPS metric) with a small sacrifice in the PSNR." + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 272, + 243, + 480, + 300 + ], + "blocks": [ + { + "bbox": [ + 272, + 243, + 480, + 300 + ], + "lines": [ + { + "bbox": [ + 272, + 243, + 480, + 300 + ], + "spans": [ + { + "bbox": [ + 272, + 243, + 480, + 300 + ], + "type": "table", + "html": "
MethodIndoors (S1,T2)Outdoors (S1,T6)
PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓
Baseline17.5220.7500.18118.3550.7100.203
+Albedo/Rough. Maps18.3340.7890.16819.3410.7470.181
+Normal/Vis.Net18.6340.7920.17319.7040.7600.180
+UVDeltaNet (Ours)18.6020.7920.16319.4610.7530.173
", + "image_path": "4f239530ef29c5d109f17ce1035e7d4cfad3d924dc0acf50596241a05ac81e4d.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 312, + 479, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 312, + 479, + 335 + ], + "spans": [ + { + "bbox": [ + 130, + 312, + 479, + 335 + ], + "type": "text", + "content": "ness 2D map, we can insert new textures or new materials and render the edited neural actor under new poses, new viewpoints, and new light conditions." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 354, + 222, + 366 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 354, + 222, + 366 + ], + "spans": [ + { + "bbox": [ + 132, + 354, + 222, + 366 + ], + "type": "text", + "content": "6 Limitations" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 380, + 298, + 499 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 380, + 298, + 499 + ], + "spans": [ + { + "bbox": [ + 130, + 380, + 298, + 499 + ], + "type": "text", + "content": "This work focuses on visual fidelity, and representations faster than NeRF could be tested in our framework in future. Moreover, like several previous techniques for the free-viewpoint rendering of humans and avatar generation [32,45], the proposed method can currently only deal with tight clothing due to the coarse geometry registration. The coarse mesh registration" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 310, + 354, + 481, + 426 + ], + "blocks": [ + { + "bbox": [ + 310, + 354, + 481, + 426 + ], + "lines": [ + { + "bbox": [ + 310, + 354, + 481, + 426 + ], + "spans": [ + { + "bbox": [ + 310, + 354, + 481, + 426 + ], + "type": "image", + "image_path": "bb58ca6ec11e0b637ecdbbdbccf9a05600ffbc19f4df560ea6561ae712b818da.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 437, + 482, + 491 + ], + "lines": [ + { + "bbox": [ + 304, + 437, + 482, + 491 + ], + "spans": [ + { + "bbox": [ + 304, + 437, + 482, + 491 + ], + "type": "text", + "content": "Fig.8: Our method allows editing the material properties and simultaneously performing skeletal pose control, free-viewpoint rendering, and relighting of the edited neural actor." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 500, + 449, + 512 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 500, + 449, + 512 + ], + "spans": [ + { + "bbox": [ + 130, + 500, + 449, + 512 + ], + "type": "text", + "content": "can also negatively affect the fine details in the face and hand synthesis." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 531, + 220, + 544 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 531, + 220, + 544 + ], + "spans": [ + { + "bbox": [ + 132, + 531, + 220, + 544 + ], + "type": "text", + "content": "7 Conclusion" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "text", + "content": "We conclude that the proposed method for the free-viewpoint rendering of humans enables simultaneous relighting, material editing, and pose control at test time while requiring a single multi-view video sequence with static illumination during training. Our method demonstrates functionality that no competing method making the same assumptions at training time (i.e. from multi-view RGB video inputs only) can provide. We hope that our method and the proposed dataset will provoke further research in this exciting field, bringing the neural rendering of humans to a new qualitative level and paving the way for next-generation metaverse applications." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 245, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 245, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 245, + 101 + ], + "type": "text", + "content": "D.C. Luvizon et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 115, + 246, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 115, + 246, + 129 + ], + "spans": [ + { + "bbox": [ + 133, + 115, + 246, + 129 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 139, + 481, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 139, + 481, + 175 + ], + "spans": [ + { + "bbox": [ + 132, + 139, + 481, + 175 + ], + "type": "text", + "content": "This work was funded by the ERC Consolidator Grant 4DRepLy (770784), the Saarbrücken Research Center for Visual Computing, Interaction and AI, and the German Research Foundation (DFG) under Grant No. 468670075." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 193, + 197, + 205 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 193, + 197, + 205 + ], + "spans": [ + { + "bbox": [ + 133, + 193, + 197, + 205 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 217, + 480, + 666 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 138, + 217, + 479, + 240 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 217, + 479, + 240 + ], + "spans": [ + { + "bbox": [ + 138, + 217, + 479, + 240 + ], + "type": "text", + "content": "1. Easymocap - make human motion capture easier. Github (2021), https://github.com/zju3dv/EasyMocap" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 240, + 480, + 272 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 240, + 480, + 272 + ], + "spans": [ + { + "bbox": [ + 138, + 240, + 480, + 272 + ], + "type": "text", + "content": "2. Alldieck, T., Magnor, M., Xu, W., Theobalt, C., Pons-Moll, G.: Video based reconstruction of 3d people models. In: Computer Vision and Pattern Recognition (CVPR). pp. 8387-8397 (2018)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 272, + 480, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 272, + 480, + 304 + ], + "spans": [ + { + "bbox": [ + 138, + 272, + 480, + 304 + ], + "type": "text", + "content": "3. Alldieck, T., Xu, H., Sminchisescu, C.: imghum: Implicit generative models of 3d human shape and articulated pose. In: International Conference on Computer Vision (ICCV). pp. 5461-5470 (2021)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 304, + 480, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 304, + 480, + 338 + ], + "spans": [ + { + "bbox": [ + 138, + 304, + 480, + 338 + ], + "type": "text", + "content": "4. Bi, S., Lombardi, S., Saito, S., Simon, T., Wei, S.E., Mcphail, K., Ramamoorthi, R., Sheikh, Y., Saragih, J.: Deep relightable appearance models for animatable faces. ACM Transactions on Graphics (TOG) 40(4), 1-15 (2021)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 338, + 480, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 338, + 480, + 370 + ], + "spans": [ + { + "bbox": [ + 138, + 338, + 480, + 370 + ], + "type": "text", + "content": "5. Boss, M., Braun, R., Jampani, V., Barron, J.T., Liu, C., Lensch, H.: Nerd: Neural reflectance decomposition from image collections. In: International Conference on Computer Vision (ICCV). pp. 12684-12694 (2021)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 370, + 480, + 393 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 370, + 480, + 393 + ], + "spans": [ + { + "bbox": [ + 138, + 370, + 480, + 393 + ], + "type": "text", + "content": "6. Chen, Z., Liu, Z.: Relighting4d: Neural relightable human from videos. In: European Conference on Computer Vision (ECCV) (2022)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 393, + 480, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 393, + 480, + 425 + ], + "spans": [ + { + "bbox": [ + 138, + 393, + 480, + 425 + ], + "type": "text", + "content": "7. De Aguiar, E., Theobalt, C., Thrun, S., Seidel, H.P.: Automatic conversion of mesh animations into skeleton-based animations. In: Computer Graphics Forum (EUROGRAPHICS). vol. 27, pp. 389-397 (2008)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 425, + 480, + 447 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 425, + 480, + 447 + ], + "spans": [ + { + "bbox": [ + 138, + 425, + 480, + 447 + ], + "type": "text", + "content": "8. Debevec, P.E., Malik, J.: Recovering high dynamic range radiance maps from photographs. In: ACM SIGGRAPH 2008 classes, pp. 1-10 (2008)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 447, + 480, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 447, + 480, + 479 + ], + "spans": [ + { + "bbox": [ + 138, + 447, + 480, + 479 + ], + "type": "text", + "content": "9. Deng, B., Wang, Y., Wetzstein, G.: Unconditional generation of relightable 3d human faces. In: International Conference on 3D Vision (3DV). pp. 302-312. IEEE (2024)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 479, + 480, + 512 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 479, + 480, + 512 + ], + "spans": [ + { + "bbox": [ + 138, + 479, + 480, + 512 + ], + "type": "text", + "content": "0. d'Eon, E., Luebke, D., Enderton, E.: Efficient Rendering of Human Skin. In: Kautz, J., Pattanaik, S. (eds.) Rendering Techniques. The Eurographics Association (2007). https://doi.org/10.2312/EGWR/EGSR07/147-157" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 512, + 480, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 512, + 480, + 555 + ], + "spans": [ + { + "bbox": [ + 138, + 512, + 480, + 555 + ], + "type": "text", + "content": "1. Guo, K., Lincoln, P., Davidson, P., Busch, J., Yu, X., Whalen, M., Harvey, G., Orts-Escalano, S., Pandey, R., Dourgarian, J., et al.: The relighttables: Volumetric performance capture of humans with realistic relighting. ACM Transactions on Graphics (TOG) 38(6), 1-19 (2019)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 555, + 480, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 555, + 480, + 600 + ], + "spans": [ + { + "bbox": [ + 138, + 555, + 480, + 600 + ], + "type": "text", + "content": "2. Habermann, M., Liu, L., Xu, W., Pons-Moll, G., Zollhoefer, M., Theobalt, C.: Hdhumans: A hybrid approach for high-fidelity digital humans. Proc. ACM Comput. Graph. Interact. Tech. 6(3) (aug 2023). https://doi.org/10.1145/3606927, https://doi.org/10.1145/3606927" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 600, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 600, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 138, + 600, + 480, + 632 + ], + "type": "text", + "content": "3. Habermann, M., Liu, L., Xu, W., Zollhoefer, M., Pons-Moll, G., Theobalt, C.: Real-time deep dynamic characters. ACM Transactions on Graphics 40(4) (aug 2021)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 632, + 480, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 632, + 480, + 666 + ], + "spans": [ + { + "bbox": [ + 138, + 632, + 480, + 666 + ], + "type": "text", + "content": "4. Habermann, M., Xu, W., Zollhoefer, M., Pons-Moll, G., Theobalt, C.: Livecap: Real-time human performance capture from monocular video. ACM Transactions On Graphics (TOG) 38(2), 1-17 (2019)" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 151, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 151, + 91, + 447, + 102 + ], + "type": "text", + "content": "Relightable Neural Actor with Intrinsic Decomposition and Pose Control" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 116, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "type": "text", + "content": "15. Habermann, M., Xu, W., Zollhofer, M., Pons-Moll, G., Theobalt, C.: Deepcap: Monocular human performance capture using weak supervision. In: Computer Vision and Pattern Recognition (CVPR). pp. 5052-5063 (2020)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 151, + 480, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 151, + 480, + 172 + ], + "spans": [ + { + "bbox": [ + 133, + 151, + 480, + 172 + ], + "type": "text", + "content": "16. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Computer Vision and Pattern Recognition (CVPR). pp. 770-778 (2016)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 174, + 480, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 174, + 480, + 206 + ], + "spans": [ + { + "bbox": [ + 133, + 174, + 480, + 206 + ], + "type": "text", + "content": "17. Iqbal, U., Caliskan, A., Nagano, K., Khamis, S., Molchanov, P., Kautz, J.: Rana: Relightable articulated neural avatars. In: International Conference on Computer Vision (ICCV). pp. 23142-23153 (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 134, + 207, + 480, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 207, + 480, + 251 + ], + "spans": [ + { + "bbox": [ + 134, + 207, + 480, + 251 + ], + "type": "text", + "content": "18. Iwase, S., Saito, S., Simon, T., Lombardi, S., Bagautdinov, T., Joshi, R., Prada, F., Shiratori, T., Sheikh, Y., Saragih, J.: Relightablehands: Efficient neural relighting of articulated hand models. In: Computer Vision and Pattern Recognition (CVPR). pp. 16663-16673 (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 134, + 251, + 480, + 284 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 251, + 480, + 284 + ], + "spans": [ + { + "bbox": [ + 134, + 251, + 480, + 284 + ], + "type": "text", + "content": "19. Ji, C., Yu, T., Guo, K., Liu, J., Liu, Y.: Geometry-aware single-image full-body human relighting. In: European Conference on Computer Vision (ECCV). pp. 388-405. Springer (2022)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 134, + 285, + 480, + 318 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 285, + 480, + 318 + ], + "spans": [ + { + "bbox": [ + 134, + 285, + 480, + 318 + ], + "type": "text", + "content": "20. Jiang, W., Yi, K.M., Samei, G., Tuzel, O., Ranjan, A.: Neuman: Neural human radiance field from a single video. In: European Conference on Computer Vision (ECCV). pp. 402-418 (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 133, + 319, + 480, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 319, + 480, + 352 + ], + "spans": [ + { + "bbox": [ + 133, + 319, + 480, + 352 + ], + "type": "text", + "content": "21. Jiang, Y., Habermann, M., Golyanik, V., Theobalt, C.: Hifecap: Monocular high-fidelity and expressive capture of human performances. In: British Machine Vision Conference (BMVC) (2022)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 133, + 353, + 480, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 353, + 480, + 385 + ], + "spans": [ + { + "bbox": [ + 133, + 353, + 480, + 385 + ], + "type": "text", + "content": "22. Jin, H., Liu, I., Xu, P., Zhang, X., Han, S., Bi, S., Zhou, X., Xu, Z., Su, H.: Tensoroir: Tensorial inverse rendering. In: Computer Vision and Pattern Recognition (CVPR). pp. 165-174 (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 133, + 386, + 480, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 386, + 480, + 408 + ], + "spans": [ + { + "bbox": [ + 133, + 386, + 480, + 408 + ], + "type": "text", + "content": "23. Kingma, D.P., Ba, J.: Adam: A method for stochastic optimization. CoRR abs/1412.6980 (2014), https://api-semanticscholar.org/CorpusID:6628106" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 133, + 409, + 480, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 409, + 480, + 441 + ], + "spans": [ + { + "bbox": [ + 133, + 409, + 480, + 441 + ], + "type": "text", + "content": "24. Kwon, Y., Kim, D., Ceylan, D., Fuchs, H.: Neural human performer: Learning generalizable radiance fields for human performance rendering. Advances in Neural Information Processing Systems (NeurIPS) 34, 24741-24752 (2021)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 133, + 442, + 480, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 442, + 480, + 475 + ], + "spans": [ + { + "bbox": [ + 133, + 442, + 480, + 475 + ], + "type": "text", + "content": "25. Kwon, Y., Liu, L., Fuchs, H., Habermann, M., Theobalt, C.: Deliffas: Deformable light fields for fast avatar synthesis. Advances in Neural Information Processing Systems (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 133, + 476, + 480, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 476, + 480, + 520 + ], + "spans": [ + { + "bbox": [ + 133, + 476, + 480, + 520 + ], + "type": "text", + "content": "26. Ledig, C., Theis, L., Huszar, F., Caballero, J., Cunningham, A., Acosta, A., Aitken, A., Tejani, A., Totz, J., Wang, Z., et al.: Photo-realistic single image superresolution using a generative adversarial network. In: Computer Vision and Pattern Recognition (CVPR). pp. 4681-4690 (2017)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 133, + 521, + 480, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 521, + 480, + 564 + ], + "spans": [ + { + "bbox": [ + 133, + 521, + 480, + 564 + ], + "type": "text", + "content": "27. Lewis, J.P., Cordner, M., Fong, N.: Pose space deformation: A unified approach to shape interpolation and skeleton-driven deformation. In: SIGGRAPH. p. 165-172. ACM Press/Addison-Wesley Publishing Co., USA (2000). https://doi.org/10.1145/344779.344862, https://doi.org/10.1145/344779.344862" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 133, + 565, + 480, + 598 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 565, + 480, + 598 + ], + "spans": [ + { + "bbox": [ + 133, + 565, + 480, + 598 + ], + "type": "text", + "content": "28. Li, G., Meka, A., Mueller, F., Buehler, M.C., Hilliges, O., Beeler, T.: Eyenerf: a hybrid representation for photorealistic synthesis, animation and relighting of human eyes. ACM Transactions on Graphics (TOG) 41(4), 1-16 (2022)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 133, + 599, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 599, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 133, + 599, + 480, + 632 + ], + "type": "text", + "content": "29. Li, G., Wu, C., Stoll, C., Liu, Y., Varanasi, K., Dai, Q., Theobalt, C.: Capturing relightable human performances under general uncontrolled illumination. In: Computer Graphics Forum. vol. 32, pp. 275-284. Wiley Online Library (2013)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 133, + 632, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 632, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 632, + 480, + 665 + ], + "type": "text", + "content": "30. Lin, S., Ryabtsev, A., Sengupta, S., Curless, B.L., Seitz, S.M., Kemelmacher-Shlizerman, I.: Real-time high-resolution background matting. In: Computer Vision and Pattern Recognition (CVPR). pp. 8762-8771 (2021)" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 245, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 245, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 245, + 100 + ], + "type": "text", + "content": "D.C. Luvizon et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 480, + 149 + ], + "type": "text", + "content": "31. Liu, G., Reda, F.A., Shih, K.J., Wang, T.C., Tao, A., Catanzaro, B.: Image inpainting for irregular holes using partial convolutions. In: European Conference on Computer Vision (ECCV) (2018)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 151, + 480, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 151, + 480, + 184 + ], + "spans": [ + { + "bbox": [ + 132, + 151, + 480, + 184 + ], + "type": "text", + "content": "32. Liu, L., Habermann, M., Rudnev, V., Sarkar, K., Gu, J., Theobalt, C.: Neural actor: Neural free-view synthesis of human actors with pose control. ACM Transactions on Graphics (TOG) 40(6), 1-16 (2021)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 185, + 480, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 185, + 480, + 217 + ], + "spans": [ + { + "bbox": [ + 132, + 185, + 480, + 217 + ], + "type": "text", + "content": "33. Liu, Y., Wang, P., Lin, C., Long, X., Wang, J., Liu, L., Komura, T., Wang, W.: Nano: Neural geometry and brdf reconstruction of reflective objects from multiview images. In: SIGGRAPH (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 218, + 480, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 218, + 480, + 251 + ], + "spans": [ + { + "bbox": [ + 132, + 218, + 480, + 251 + ], + "type": "text", + "content": "34. Loper, M., Mahmood, N., Romero, J., Pons-Moll, G., Black, M.J.: SMPL: A skinned multi-person linear model. ACM Transactions on Graphics (SIGGRAPH Asia) 34(6), 248:1-248:16 (2015)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 251, + 480, + 284 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 251, + 480, + 284 + ], + "spans": [ + { + "bbox": [ + 132, + 251, + 480, + 284 + ], + "type": "text", + "content": "35. Lyu, L., Tewari, A., Leimkuehler, T., Habermann, M., Theobalt, C.: Neural radiance transfer fields for relightable novel-view synthesis with global illumination. In: European Conference on Computer Vision (ECCV) (2022)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 285, + 480, + 350 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 285, + 480, + 350 + ], + "spans": [ + { + "bbox": [ + 132, + 285, + 480, + 350 + ], + "type": "text", + "content": "36. Meka, A., Pandey, R., Hane, C., Orts-Escolano, S., Barnum, P., David-Son, P., Erickson, D., Zhang, Y., Taylor, J., Bouaziz, S., Legendre, C., Ma, W.C., Overbeck, R., Beeler, T., Debevec, P., Izadi, S., Theobalt, C., Rhemann, C., Fanello, S.: Deep relightable textures: Volumetric performance capture with neural rendering. ACM Transactions on Graphics (TOG) 39(6) (nov 2020). https://doi.org/10.1145/3414685.3417814, https://doi.org/10.1145/3414685.3417814" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 352, + 480, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 352, + 480, + 385 + ], + "spans": [ + { + "bbox": [ + 132, + 352, + 480, + 385 + ], + "type": "text", + "content": "37. Mihajlovic, M., Saito, S., Bansal, A., Zollhoefer, M., Tang, S.: COAP: Compositional articulated occupancy of people. In: Computer Vision and Pattern Recognition (CVPR) (2022)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 386, + 480, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 386, + 480, + 418 + ], + "spans": [ + { + "bbox": [ + 132, + 386, + 480, + 418 + ], + "type": "text", + "content": "38. Mihajlovic, M., Zhang, Y., Black, M.J., Tang, S.: Leap: Learning articulated occupancy of people. In: Computer Vision and Pattern Recognition (CVPR). pp. 10461-10471 (2021)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 419, + 480, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 419, + 480, + 453 + ], + "spans": [ + { + "bbox": [ + 132, + 419, + 480, + 453 + ], + "type": "text", + "content": "39. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. In: European Conference on Computer Vision (ECCV) (2020)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 453, + 480, + 486 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 453, + 480, + 486 + ], + "spans": [ + { + "bbox": [ + 132, + 453, + 480, + 486 + ], + "type": "text", + "content": "40. Pang, H., Zhu, H., Kortylewski, A., Theobalt, C., Habermann, M.: Ash: Animatable gaussian splats for efficient and photoreal human rendering. In: Computer Vision and Pattern Recognition (CVPR). pp. 1165-1175 (June 2024)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 487, + 480, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 487, + 480, + 520 + ], + "spans": [ + { + "bbox": [ + 132, + 487, + 480, + 520 + ], + "type": "text", + "content": "41. Pavlakos, G., Choutas, V., Ghorbani, N., Bolkart, T., Osman, A.A.A., Tzionas, D., Black, M.J.: Expressive body capture: 3d hands, face, and body from a single image. In: Computer Vision and Pattern Recognition (CVPR) (2019)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 521, + 480, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 521, + 480, + 553 + ], + "spans": [ + { + "bbox": [ + 132, + 521, + 480, + 553 + ], + "type": "text", + "content": "42. Peng, S., Zhang, Y., Xu, Y., Wang, Q., Shuai, Q., Bao, H., Zhou, X.: Neural body: Implicit neural representations with structured latent codes for novel view synthesis of dynamic humans. In: Computer Vision and Pattern Recognition (CVPR) (2021)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 554, + 480, + 608 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 554, + 480, + 608 + ], + "spans": [ + { + "bbox": [ + 132, + 554, + 480, + 608 + ], + "type": "text", + "content": "43. Ramamoorthi, R., Hanrahan, P.: A signal-processing framework for inverse rendering. In: Conference on Computer Graphics and Interactive Techniques. p. 117-128. SIGGRAPH '01, Association for Computing Machinery, New York, NY, USA (2001). https://doi.org/10.1145/383259.383271, https://doi.org/10.1145/383259.383271" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 609, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 609, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 609, + 480, + 632 + ], + "type": "text", + "content": "44. Ranjan, A., Yi, K.M., Chang, J.H.R., Tuzel, O.: Facelit: Neural 3d relightable faces. In: Computer Vision and Pattern Recognition (CVPR). pp. 8619-8628 (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "type": "text", + "content": "45. Remelli, E., Bagautdinov, T., Saito, S., Wu, C., Simon, T., Wei, S.E., Guo, K., Cao, Z., Prada, F., Saragih, J., et al.: Drivable volumetric avatars using texel-aligned features. In: ACM SIGGRAPH 2022 Conference Proceedings (2022)" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 151, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 151, + 91, + 447, + 102 + ], + "type": "text", + "content": "Relightable Neural Actor with Intrinsic Decomposition and Pose Control" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 666 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "type": "text", + "content": "46. Schwartz, G., Wei, S.E., Wang, T.L., Lombardi, S., Simon, T., Saragih, J., Sheikh, Y.: The eyes have it: An integrated eye and face model for photorealistic facial animation. ACM Transactions on Graphics (TOG) 39(4), 91-1 (2020)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 149, + 480, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 149, + 480, + 182 + ], + "spans": [ + { + "bbox": [ + 133, + 149, + 480, + 182 + ], + "type": "text", + "content": "47. Shu, Z., Sahasrabudhe, M., Guler, R.A., Samaras, D., Paragios, N., Kokkinos, I.: Deforming autoencoders: Unsupervised disentangling of shape and appearance. In: European Conference on Computer Vision (ECCV). pp. 650-665 (2018)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 182, + 480, + 224 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 182, + 480, + 224 + ], + "spans": [ + { + "bbox": [ + 133, + 182, + 480, + 224 + ], + "type": "text", + "content": "48. Shysheya, A., Zakharov, E., Aliev, K.A., Bashirov, R., Burkov, E., Iskakov, K., Ivakshtenko, A., Malkov, Y., Pasechnik, I., Ulyanov, D., et al.: Textured neural avatars. In: Computer Vision and Pattern Recognition (CVPR). pp. 2387-2397 (2019)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 224, + 480, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 224, + 480, + 267 + ], + "spans": [ + { + "bbox": [ + 132, + 224, + 480, + 267 + ], + "type": "text", + "content": "49. Srinivasan, P.P., Deng, B., Zhang, X., Tancik, M., Mildenhall, B., Barron, J.T.: Nerv: Neural reflectance and visibility fields for relighting and view synthesis. In: Computer Vision and Pattern Recognition (CVPR). pp. 7491-7500 (2021). https://doi.org/10.1109/CVPR46437.2021.00741" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 267, + 480, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 267, + 480, + 300 + ], + "spans": [ + { + "bbox": [ + 132, + 267, + 480, + 300 + ], + "type": "text", + "content": "50. Tan, F., Fanello, S., Meka, A., Orts-Escalano, S., Tang, D., Pandey, R., Taylor, J., Tan, P., Zhang, Y.: Volux-gan: A generative model for 3d face synthesis with hdri relighting. In: ACM SIGGRAPH 2022 Conference Proceedings. pp. 1-9 (2022)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 300, + 480, + 343 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 300, + 480, + 343 + ], + "spans": [ + { + "bbox": [ + 132, + 300, + 480, + 343 + ], + "type": "text", + "content": "51. Tewari, A., Thies, J., Mildenhall, B., Srinivasan, P., Tretschk, E., Yifan, W., Lassner, C., Sitzmann, V., Martin-Brualla, R., Lombardi, S., Simon, T., Theobalt, C., Nießner, M., Barron, J.T., Wetzstein, G., Zollhöfer, M., Golyanik, V.: Advances in Neural Rendering. Computer Graphics Forum (EG STAR 2022) (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 343, + 480, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 343, + 480, + 385 + ], + "spans": [ + { + "bbox": [ + 132, + 343, + 480, + 385 + ], + "type": "text", + "content": "52. Walter, B., Marschner, S.R., Li, H., Torrance, K.E.: Microfacet Models for Refraction through Rough Surfaces. In: Kautz, J., Pattanaik, S. (eds.) Rendering Techniques. The Eurographics Association (2007). https://doi.org/10.2312/EGWR/EGSR07/195-206" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 385, + 480, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 385, + 480, + 418 + ], + "spans": [ + { + "bbox": [ + 132, + 385, + 480, + 418 + ], + "type": "text", + "content": "53. Wang, P., Liu, L., Liu, Y., Theobalt, C., Komura, T., Wang, W.: Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. In: Advances in Neural Information Processing Systems (NeurIPS) (2021)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 418, + 480, + 449 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 418, + 480, + 449 + ], + "spans": [ + { + "bbox": [ + 132, + 418, + 480, + 449 + ], + "type": "text", + "content": "54. Wang, S., Schwarz, K., Geiger, A., Tang, S.: Arah: Animatable volume rendering of articulated human sdfs. In: European Conference on Computer Vision (ECCV) (2022)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 449, + 480, + 482 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 449, + 480, + 482 + ], + "spans": [ + { + "bbox": [ + 132, + 449, + 480, + 482 + ], + "type": "text", + "content": "55. Wang, T.C., Liu, M.Y., Zhu, J.Y., Liu, G., Tao, A., Kautz, J., Catanzaro, B.: Video-to-video synthesis. In: Conference on Neural Information Processing Systems (NeurIPS) (2018)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 482, + 480, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 482, + 480, + 514 + ], + "spans": [ + { + "bbox": [ + 132, + 482, + 480, + 514 + ], + "type": "text", + "content": "56. Wang, Y., Holynski, A., Zhang, X., Zhang, X.: Sunstage: Portrait reconstruction and relighting using the sun as a light stage. In: Computer Vision and Pattern Recognition (CVPR). pp. 20792-20802 (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 514, + 480, + 556 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 514, + 480, + 556 + ], + "spans": [ + { + "bbox": [ + 132, + 514, + 480, + 556 + ], + "type": "text", + "content": "57. Weng, C.Y., Curless, B., Srinivasan, P.P., Barron, J.T., Kemelmacher-Shlizerman, I.: Humannerf: Free-viewpoint rendering of moving people from monocular video. In: Computer Vision and Pattern Recognition (CVPR). pp. 16210-16220 (June 2022)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 556, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 556, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 556, + 480, + 632 + ], + "type": "text", + "content": "58. Wuu, C.h., Zheng, N., Ardisson, S., Bali, R., Belko, D., Brockmeyer, E., Evans, L., Godisart, T., Ha, H., Huang, X., Hypes, A., Koska, T., Krenn, S., Lombardi, S., Luo, X., McPhail, K., Millerschoen, L., Perdoch, M., Pitts, M., Richard, A., Saragih, J., Saragih, J., Shiratori, T., Simon, T., Stewart, M., Trimble, A., Weng, X., Whitewolf, D., Wu, C., Yu, S.I., Sheikh, Y.: Multiface: A dataset for neural face rendering. In: arXiv (2022). https://doi.org/10.48550/ARXIV.2207.11243, https://arxiv.org/abs/2207.11243" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 632, + 480, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 480, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 480, + 666 + ], + "type": "text", + "content": "59. Yang, H., Zheng, M., Feng, W., Huang, H., Lai, Y.K., Wan, P., Wang, Z., Ma, C.: Towards practical capture of high-fidelity reconfigurable avatars. In: SIGGRAPH Asia 2023 Conference Papers. pp. 1-11 (2023)" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 245, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 245, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 245, + 100 + ], + "type": "text", + "content": "D.C. Luvizon et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 482, + 348 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 132, + 116, + 482, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 482, + 149 + ], + "type": "text", + "content": "60. Zhang, K., Luan, F., Li, Z., Snavely, N.: IRON: Inverse rendering by optimizing neural sdfs and materials from photometric images. In: Computer Vision and Pattern Recognition (CVPR). pp. 5565-5574 (2022)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 150, + 482, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 150, + 482, + 182 + ], + "spans": [ + { + "bbox": [ + 132, + 150, + 482, + 182 + ], + "type": "text", + "content": "61. Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual metric. In: Computer Vision and Pattern Recognition (CVPR) (2018)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 182, + 481, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 182, + 481, + 216 + ], + "spans": [ + { + "bbox": [ + 132, + 182, + 481, + 216 + ], + "type": "text", + "content": "62. Zhang, X., Srinivasan, P.P., Deng, B., Debevec, P., Freeman, W.T., Barron, J.T.: NeRFactor: Neural factorization of shape and reflectance under an unknown illumination. ACM Transactions on Graphics (TOG) 40, 1-18 (2021)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 216, + 481, + 247 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 216, + 481, + 247 + ], + "spans": [ + { + "bbox": [ + 132, + 216, + 481, + 247 + ], + "type": "text", + "content": "63. Zhang, Y., Sun, J., He, X., Fu, H., Jia, R., Zhou, X.: Modeling indirect illumination for inverse rendering. In: Computer Vision and Pattern Recognition (CVPR). pp. 18643-18652 (2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 248, + 481, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 248, + 481, + 282 + ], + "spans": [ + { + "bbox": [ + 132, + 248, + 481, + 282 + ], + "type": "text", + "content": "64. Zheng, R., Li, P., Wang, H., Yu, T.: Learning visibility field for detailed 3d human reconstruction and relighting. In: Computer Vision and Pattern Recognition (CVPR). pp. 216-226 (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 282, + 481, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 282, + 481, + 315 + ], + "spans": [ + { + "bbox": [ + 132, + 282, + 481, + 315 + ], + "type": "text", + "content": "65. Zhi, Y., Qian, S., Yan, X., Gao, S.: Dual-space nerf: Learning animatable avatars and scene lighting in separate spaces. In: International Conference on 3D Vision (3DV) (Sep 2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 315, + 481, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 315, + 481, + 348 + ], + "spans": [ + { + "bbox": [ + 132, + 315, + 481, + 348 + ], + "type": "text", + "content": "66. Zhou, T., He, K., Wu, D., Xu, T., Zhang, Q., Shao, K., Chen, W., Xu, L., Yu, J.: Relightable neural human assets from multi-view gradient illuminations. In: Computer Vision and Pattern Recognition (CVPR). pp. 4315-4327 (2023)" + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 151, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 151, + 91, + 447, + 102 + ], + "type": "text", + "content": "Relightable Neural Actor with Intrinsic Decomposition and Pose Control" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "19" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 18 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Remove Projective LiDAR Depthmap Artifacts via Exploiting Epipolar Geometry/0a1ce34c-ca2a-42ac-a914-2a125b92f644_content_list.json b/2024/Remove Projective LiDAR Depthmap Artifacts via Exploiting Epipolar Geometry/0a1ce34c-ca2a-42ac-a914-2a125b92f644_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..b24e250249f6f48516881bd90809d3f9fdde29cf --- /dev/null +++ b/2024/Remove Projective LiDAR Depthmap Artifacts via Exploiting Epipolar Geometry/0a1ce34c-ca2a-42ac-a914-2a125b92f644_content_list.json @@ -0,0 +1,2204 @@ +[ + { + "type": "text", + "text": "RePLAY: Remove Projective LiDAR Depthmap Artifacts via Exploiting Epipolar Geometry", + "text_level": 1, + "bbox": [ + 228, + 140, + 776, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Shengjie Zhu*, Girish Chandar Ganesan*, Abhinav Kumar, and Xiaoming Liu", + "bbox": [ + 217, + 212, + 782, + 228 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Department of Computer Science and Engineering, Michigan State University, East Lansing, MI, USA, 48824 [zhusheng, ganesang, kumarab6, liuxm]@msu.edu https://shngjz.github.io/RePLAy/", + "bbox": [ + 307, + 239, + 694, + 297 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/1f6301cf1efa3f044f19cbd5f11efd2bf5dea2a4937c3bd078663b43033c201a.jpg", + "image_caption": [ + "Fig. 1: Remove Projective LiDAR Depthmap Artifacts. The LiDAR projected depthmap has misalignment with the RGB camera. The background scans highlighted by blue arrows incorrectly overlay the foreground such as cars and pedestrians. The artifacts persist unattended in most AV datasets including KITTI-360 [41], nuScenes [10], Waymo [56], and DDAD [24]. We propose an easy-to-use, parameter-free, and analytical solution to remove the projective artifacts, shown by green arrows." + ], + "image_footnote": [], + "bbox": [ + 217, + 330, + 785, + 463 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. 3D sensing is a fundamental task for Autonomous Vehicles. Its deployment often relies on aligned RGB cameras and LiDAR. Despite meticulous synchronization and calibration, systematic misalignment persists in LiDAR projected depthmap. This is due to the physical baseline distance between the two sensors. The artifact is often reflected as background LiDAR incorrectly projected onto the foreground, such as cars and pedestrians. The KITTI dataset uses stereo cameras as a heuristic solution to remove artifacts. However most AV datasets, including nuScenes, Waymo, and DDAD, lack stereo images, making the KITTI solution inapplicable. We propose RePLAY, a parameter-free analytical solution to remove the projective artifacts. We construct a binocular vision system between a hypothesized virtual LiDAR camera and the RGB camera. We then remove the projective artifacts by determining the epipolar occlusion with the proposed analytical solution. We show unanimous improvement in the State-of-The-Art (SoTA) monocular depth estimators and 3D object detectors with the artifacts-free depthmaps.", + "bbox": [ + 259, + 571, + 743, + 792 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: LiDAR $\\cdot$ Autonomous Driving $\\cdot$ Epipolar Geometry", + "bbox": [ + 261, + 805, + 689, + 819 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Equal Contributions", + "bbox": [ + 220, + 824, + 372, + 839 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/abdc8bd1ff419c6ec89d5a731263d1e2f7eb1cfb61d83e346dc32a6e42a1911e.jpg", + "image_caption": [ + "(a) Autonomous Vehicles [10]" + ], + "image_footnote": [], + "bbox": [ + 228, + 145, + 454, + 181 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/a9b1390cf76e39cdf350736ed7cdef7fda9681647bf5b25148afe62f5121633d.jpg", + "image_caption": [ + "(b) Robotics Machines [14]", + "Fig. 2: LiDAR-RGB sensor sets have a physical baseline distance between the two sensors, resulting in Fig. 1 projective artifacts. (a) The background scanned by LiDAR incorrectly overlays on the foreground pedestrian observed by the RGB camera." + ], + "image_footnote": [], + "bbox": [ + 480, + 142, + 647, + 181 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/0c2984129ae310a10325f182ee878b2e9052d2ec106148698daaa0362979d1f5.jpg", + "image_caption": [ + "(c) Headset [26]" + ], + "image_footnote": [], + "bbox": [ + 669, + 142, + 779, + 181 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 265, + 374, + 280 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "LiDAR device scans the surrounding environment via emitting laser pulses into the 3D space. Its applications span across Autonomous Vehicles (AV) [40, 52, 56], AR/VR [27, 48], and Robotics [31]. Taking AV as an example, the LiDAR is synchronized and calibrated against the RGB camera. The correspondence between RGB pixels and LiDAR 3D points is determined by projecting the LiDAR 3D point onto the image plane, as shown with the Fig. 1 depthmaps.", + "bbox": [ + 212, + 296, + 784, + 386 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Despite ideal synchronization and calibration, there still exists systematic misalignment between the RGB camera and LiDAR. Fig. 1 highlights the projective artifacts in LiDAR depthmap with blue arrows, where the background scans incorrectly overlay on the foreground objects.", + "bbox": [ + 212, + 387, + 784, + 446 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The projective artifacts arise from the physical baseline distance between the RGB camera and LiDAR. This physical limitation is ubiquitous among different LiDAR-RGB sensor sets used in Robotics, AR/VR, and AV tasks (Fig. 2). The baseline distance (Fig. 2a) renders certain 3D scans exclusively visible to LiDAR, rather than the RGB camera. For those scans, their correspondence with image pixels is absent, causing Fig. 1 projective artifacts.", + "bbox": [ + 212, + 446, + 784, + 537 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The KITTI dataset [1] provides a heuristic solution to remove the projective artifacts, which we show in Fig. 3. However, the method requires stereo sensors, making it inapplicable to other AV datasets. Thus, the projective artifacts persist unattended in most AV datasets. Tab. 1 summarizes these datasets.", + "bbox": [ + 212, + 537, + 784, + 597 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We propose RePLAy: a parameter-free, non-learning, and analytical solution to resolve the projective artifacts. RePLAy only requires the relative pose between the LiDAR scanner and the RGB camera, i.e., their extrinsic matrix or calibration matrix. It applies to all AV datasets surveyed in Tab. 1 and diverse sensor settings included in Fig. 2. RePLAy does not even require RGB images, hence being intact at the asynchronization noise between the RGB camera and LiDAR. Fig. 3 provides a qualitative comparison with KITTI solution.", + "bbox": [ + 212, + 598, + 784, + 703 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "RePLAy first adopts an auto-calibration algorithm to calibrate a hypothesized virtual camera at the origin of the LiDAR 3D point cloud. The virtual LiDAR camera formulates a stereo system with the RGB camera. We attribute the projective artifact to the epipolar occlusion in stereo vision. We propose an analytical solution for detecting epipolar occlusion using epipolar geometry.", + "bbox": [ + 212, + 704, + 784, + 779 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "LiDAR projected depthmap is essential for diverse downstream applications. We select two to verify our effectiveness: monocular depth estimation (MonoDE) [5, 50, 59, 70] and monocular 3D object detection (Mono3DOD) that uses auxiliary depthmap [51, 63, 64, 66]. We benchmark their performance trained", + "bbox": [ + 212, + 779, + 785, + 839 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "S. Zhu et al.", + "bbox": [ + 271, + 114, + 354, + 126 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/18a9013cee8c87fcca991d968278b7c503f0d37051be215535c0a497eb477054.jpg", + "image_caption": [ + "(a) Raw Depthmap" + ], + "image_footnote": [], + "bbox": [ + 230, + 143, + 408, + 261 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/032acb6e0f9fc1ef93f835b5392559af2fcb93fe731881f96067f7ef798ce947.jpg", + "image_caption": [ + "(b) RePLAy", + "Fig. 3: Qualitative comparison with KITTI semi-dense depthmap. RePLAY only requires the LiDAR calibration matrix. KITTI [57] uses multi-frame fusion with the stereo camera to remove projective artifacts. The higher density of KITTI depthmap is from the additional sensor and temporal fusion." + ], + "image_footnote": [], + "bbox": [ + 416, + 143, + 596, + 262 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/48c1b008eaee63e26ab6e8dceea239a70cde5c0fbfd52bee5244b07ae1b6fbe6.jpg", + "image_caption": [ + "(c) KITTI [57] with Stereo" + ], + "image_footnote": [], + "bbox": [ + 602, + 143, + 779, + 262 + ], + "page_idx": 2 + }, + { + "type": "table", + "img_path": "images/6cfb7f2262d6cf9ee96285062dfe9bec8dffef196e55331a20605210b4e800ab.jpg", + "table_caption": [ + "Table 1: LiDAR Depthmap Projective Artifacts exist in most Autonomous Vehicle Datasets. KITTI uses stereo images as a heuristic solution." + ], + "table_footnote": [], + "table_body": "
KITTI [1]KITTI-360 [41]nuScenes [10]Waymo [56]DDAD [24]ApolloScape [30]Argoverse [62]
ArtifactsNoYesYesYesYesYesYes
", + "bbox": [ + 217, + 380, + 782, + 407 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "using LiDAR depthmap cleaned with multiple algorithms and RePLAy on AV datasets. We demonstrate unanimous improvement in State-of-The-Art (SoTA) methods [5,29,51] on both tasks. We additionally release the processed depthmap of five major AV datasets to benefit the community. Our contributions are:", + "bbox": [ + 212, + 440, + 784, + 501 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "$\\checkmark$ We utilize epipolar geometry to remove the projective artifacts in the LiDAR depthmap.", + "$\\checkmark$ We demonstrate unanimous improvement in the SoTA monocular depth and monocular 3D object detection methods.", + "$\\checkmark$ We provide processed depthmap in KITTI [1], KITTI-360 [41], nuScenes [10], Waymo [56], and DDAD [24] datasets." + ], + "bbox": [ + 215, + 505, + 782, + 594 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 625, + 387, + 640 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "LiDAR and Camera Misalignment. Multiple sources of misalignment exist between LiDAR and RGB camera. Foggy, rainy, and snowy weather negatively impacts both LiDAR [16, 37, 47] and RGB cameras [18-20, 54]. The slight drift of LiDAR and camera accumulates during vehicle operation. [3, 39, 67] investigate the robustness to the spatial misalignment. [67] benchmarks the temporal asynchronization between LiDAR and RGB camera. The projective artifact is another misalignment that universally exists in multiple AV datasets, as shown in Tab. 1. Our solution complements prior study [67] by addressing it.", + "bbox": [ + 212, + 657, + 784, + 777 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Heuristic Solution for Projective Artifacts. KITTI [1, 57] dataset uses the stereo image as a heuristic solution. [57] produces a stereo reconstruction through semi-global matching [28]. They additionally augment the stereo results with 11 neighboring temporal frames. The projective artifacts are removed via", + "bbox": [ + 212, + 779, + 785, + 839 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "RePLAY", + "bbox": [ + 671, + 114, + 730, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 114, + 785, + 126 + ], + "page_idx": 2 + }, + { + "type": "code", + "sub_type": "algorithm", + "code_caption": [ + "Algorithm 1 RePLAy: Remove Projective Artifacts" + ], + "code_body": "Input: LiDAR Point Cloud V, Calibration Matrix $\\mathbf{P} = [\\mathbf{R}\\mathbf{t}]$ , Camera Intrinsic $\\mathbf{K}_r$ \nOutput: Occluded Pixels $\\mathcal{O}$ \n1: Auto-Calibration ▷ Sec. 3.1 \n2: Set Virtual LiDAR Camera $\\mathbf{K}_l$ \n3: Produce Virtual LiDAR Depthmap $\\mathbf{D}_l$ \n4: Denseify Virtual LiDAR Depthmap $\\overline{\\mathbf{D}}_l$ \n5: for v ∈ V do \n6: p1, dp1 ← π(Rv | Kl) \n7: for p2 ∈ Lp1 do ▷ Eq. (14) and Fig. 8 \n8: dp2 ← bilinear interpolate $\\overline{\\mathbf{D}}_l$ at pixel location p2 ▷ Fig. 8 \n9: if g(p1, p2, dp1, dp2 | Kl, Kr, [I t]) ≤ 0 then ▷ Eq. (15) \n10: {p1} ∪ O \n11: break", + "bbox": [ + 215, + 164, + 785, + 349 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "consistency verification between the stereo depthmap and LiDAR depthmap. Compared to ours, depthmap from [57] has improved density. However, the requirement of stereo sensors limits its application, as summarized in Tab. 1. Additionally, [57] has a complicated pipeline, and is therefore, not used even in the KITTI-360 [41] dataset, which shares a similar setup.", + "bbox": [ + 212, + 382, + 784, + 458 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Stereo Occlusion Detection. Stereo co-planar cameras are classically formulated as the half-occlusion problem in stereo vision, i.e., the area occluded in one camera of a stereo vision system. There has been extensive research on detecting the half-occlusion region [2, 45, 58, 65]. Several works [6, 7, 32, 60, 72] jointly optimize occlusion and disparity. Others [17, 55, 61] apply left-right consistency check. The area receives different estimations from left and right views are excluded. RePLAy focuses on setups without stereo but where LiDAR provides rough 3D structures. Compared to the half-occlusion baseline which proposes a heuristic solution, RePLAy is an analytical solution.", + "bbox": [ + 212, + 460, + 785, + 595 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Downstream Tasks using LiDAR Depthmap. Multiple downstream tasks use LiDAR projected depthmaps. E.g., MonoDE methods [4, 5, 38, 68, 71] use depthmap as the GT supervision. Mono3DOD methods [36, 43, 63, 64, 66] use depthmap for supervising auxiliary task along with 3D object detection. When deployed to AV datasets other than KITTI, the aforementioned methods [42, 69] have diminished performance due to the projective artifacts. RePLAy improves their performance by resolving the projective artifacts in LiDAR depthmap.", + "bbox": [ + 212, + 598, + 785, + 704 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3 RePLAy", + "text_level": 1, + "bbox": [ + 215, + 737, + 333, + 753 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We now describe RePLAy in Algorithm 1. Sec. 3.1 attributes the projective artifacts to the epipolar occlusion induced by the binocular vision system between the virtual LiDAR camera and RGB camera. Sec. 3.2 outlines our analytical solution. Sec. 3.3 hold further discussions.", + "bbox": [ + 212, + 779, + 785, + 839 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "S. Zhu et al.", + "bbox": [ + 271, + 114, + 354, + 127 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/56ec7f48c4e539462254a9f00d7061d3cb87502dcea17c65b2d8e744a91ba4e9.jpg", + "image_caption": [ + "(a) LiDAR Point Cloud Bird-Eye View" + ], + "image_footnote": [], + "bbox": [ + 217, + 143, + 455, + 271 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/7359ba96e77b22761ddddcdd51b8855c34fb4771912147ebabb12351c17b518e.jpg", + "image_caption": [ + "(b) Depthmap $\\mathbf{D}_l$ of Virtual LiDAR Camera." + ], + "image_footnote": [], + "bbox": [ + 464, + 143, + 784, + 203 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/cc6005c98a3d65eba0f18a680a928fd4cf29a02759168a39e6eeb9f64c3e390b.jpg", + "image_caption": [ + "(c) Depthmap $\\mathbf{D}_r$ of RGB Camera.", + "Fig. 4: Virtual LiDAR camera and RGB Binocular System. In (a), we plot the bird-eye view of the point cloud $\\mathcal{V}$ from LiDAR. A binocular system is set between virtual LiDAR (epipolar left) and RGB camera (epipolar right). Subfigure (b) plots the depthmap $\\mathbf{D}_l$ from LiDAR view. As stated in Lemma 1, depthmap $\\mathbf{D}_l$ does not have projective artifacts. Compare (b) and (c), the artifacts arise when transfer from LiDAR to RGB camera, due to the epipolar occlusion among the binocular system. We formally define the projective artifacts (or epipolar occlusion) in Eq. (7)." + ], + "image_footnote": [], + "bbox": [ + 464, + 217, + 784, + 272 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We execute Sec. 3.1 auto-calibration over the entire dataset. We initialize LiDAR intrinsic $\\mathbf{K}_l$ the same as the RGB camera $\\mathbf{K}_r$ with slight augmentation to ensure a sufficient Field-of-View.", + "bbox": [ + 212, + 410, + 782, + 455 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1 LiDAR Depthmap Projective Artifacts as Epipolar Occlusion", + "text_level": 1, + "bbox": [ + 212, + 479, + 766, + 494 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Fig. 2 demonstrates the baseline distance between various LiDAR and RGB camera sensor sets. Fig. 2a shows that the background tree is solely visible to LiDAR. Intuitively, the background tree is occluded by the foreground pedestrian in a binocular vision system of the virtual LiDAR camera and RGB camera.", + "bbox": [ + 212, + 505, + 784, + 565 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Virtual LiDAR Camera. LiDAR produces a point cloud $\\mathcal{V}$ by emitting laser pulses into 3D space. Mathematically, we assume the LiDAR point cloud $\\mathcal{V}$ as a set of 3D rays originating from the coordinate origin:", + "bbox": [ + 212, + 566, + 784, + 612 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {V} = \\left\\{\\mathbf {v} _ {i} \\right\\} = \\left\\{\\mathbf {0} + \\bar {\\mathbf {r}} _ {i} \\cdot t _ {i} \\mid i \\in \\mathbb {Z}, t > 0, \\bar {\\mathbf {r}} _ {i} \\in \\mathbb {R} ^ {3} \\right\\}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 328, + 623, + 784, + 642 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathbf{v}_i$ is a 3D point. $\\overline{\\mathbf{r}}_i$ is a normalized direction vector, and $t_i$ is the euclidean distance between the 3D point and LiDAR. If we position a virtual LiDAR camera with arbitrary resolution and intrinsic $\\mathbf{K}_l$ at the LiDAR's location, then:", + "bbox": [ + 212, + 652, + 782, + 698 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Lemma 1. The depthmap of the virtual LiDAR camera is free of projective artifacts, i.e., no two scanned points $\\mathbf{v}_1$ and $\\mathbf{v}_2$ overlap after projection.", + "bbox": [ + 212, + 723, + 782, + 753 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\forall \\mathbf {v} _ {1}, \\mathbf {v} _ {2} \\in \\mathcal {V}, \\| \\pi (\\mathbf {v} _ {1} \\mid \\mathbf {K} _ {l}) - \\pi (\\mathbf {v} _ {2} \\mid \\mathbf {K} _ {l}) \\| _ {2} \\neq 0, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 336, + 766, + 784, + 782 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\pi (\\cdot)$ stands for the projection process. See proof in Supp. (Appendix A). We visualize the artifacts-free depthmap $\\mathbf{D}_l$ from the virtual LiDAR camera in Fig. 4b.", + "bbox": [ + 212, + 794, + 782, + 839 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "RePLAY", + "bbox": [ + 671, + 114, + 730, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/d98fbaebe05c71db5e6e056a53358e127082fa01709ca3d72d3161c4252fb17c.jpg", + "image_caption": [ + "Fig. 5: Auto-Calibration. The LiDAR point cloud $\\mathcal{V}$ in (a) is mapped to spherical coordinate $\\theta -\\phi$ with Eq. (3). We use a binary matrix $\\mathbf{S}$ to record the unique pixel locations $S$ after rasterization. Plots (b) and (c) are matrix $\\mathbf{S}$ before and after autocalibration where duplication (marked by arrows) is minimized with Eq. (5) loss $L$ . Plots (d) and (e) contrast the virtual LiDAR depthmap $\\mathbf{D}_l$ before and after autocalibration where Lemma 1 is fulfilled." + ], + "image_footnote": [], + "bbox": [ + 223, + 143, + 794, + 244 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Auto-Calibration. In practice, Lemma 1 may not hold in a realistic LiDAR point cloud due to possible bias in the point cloud coordinate origin $\\mathbf{0}$ . Fig. 5d, an example from the KITTI dataset shows persistent projective artifacts in the depthmap $\\mathbf{D}_l$ when without any treatment. We perform auto-calibration to fulfill Lemma 1 by re-adjusting the point cloud origin. Fig. 4b and Fig. 5e show artifacts-free depthmap after auto-calibration.", + "bbox": [ + 212, + 352, + 782, + 441 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Auto-calibration aims to resolve the artifacts in the virtual LiDAR camera depthmap $\\mathbf{D}_l$ . We convert each LiDAR point cloud to the spherical coordinate system. This hypothesizes an omnidirectional image $\\mathbf{S}$ of the LiDAR point cloud. Denote the origin to estimate as $\\mathbf{o}$ . Set the new 3D point $\\mathbf{u} = \\mathbf{v} + \\mathbf{o}$ , we have the polar angle $\\theta$ and azimuthal angle $\\phi$ after applying a rasterization operation $[\\cdot]:$", + "bbox": [ + 212, + 441, + 784, + 518 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\theta = \\left[ \\arccos \\left(u _ {x} / \\left(u _ {x} ^ {2} + u _ {z} ^ {2}\\right)\\right) \\right], \\phi = \\left[ \\arccos \\left(u _ {y} / \\left(u _ {x} ^ {2} + u _ {y} ^ {2} + u _ {z} ^ {2}\\right)\\right) \\right], \\left[ \\theta \\phi \\right] = f (\\mathbf {v} + \\mathbf {o}). \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 228, + 525, + 784, + 542 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Intuitively, function $f(\\cdot)$ converts a 3D point $\\mathbf{u} \\in \\mathcal{V}$ to a discrete 2D coordinate in $\\mathbf{S}$ . We denote the set of unique coordinates as $S$ :", + "bbox": [ + 212, + 551, + 782, + 582 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {S} = \\left\\{f (\\mathbf {v} + \\mathbf {o}) \\mid \\mathbf {v} \\in \\mathcal {V} \\right\\}. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 411, + 593, + 785, + 609 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We implement Eq. (4) with a binary matrix $\\mathbf{S}$ . Its elements are 1 at each discrete coordinate and 0 otherwise. The size of $\\mathcal{S}$ equals to summarizing the matrix $\\mathbf{S}$ . Artifacts become duplication in set $\\mathcal{S}$ . Removing artifacts is to minimize duplication. The loss $L$ is then defined as:", + "bbox": [ + 212, + 619, + 787, + 680 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nL = \\| \\mathcal {V} \\| - \\| \\mathcal {S} \\|. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 441, + 691, + 785, + 709 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The loss $L$ is not differentiable. In each iteration, we approximate its gradient by ablating origin $\\mathbf{o}$ with a small deviation and perform coordinate descent.", + "bbox": [ + 212, + 719, + 784, + 748 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Virtual LiDAR Camera and RGB Binocular System. In Fig. 4, we set up a binocular vision system, denoting the virtual LiDAR camera and RGB camera as the epipolar left and right camera respectively. Correspondingly, set their intrinsic as $\\mathbf{K}_l$ and $\\mathbf{K}_r$ , and projected depthmap as $\\mathbf{D}_l$ and $\\mathbf{D}_r$ . Thus the problem of removing the projective artifacts is translated to detecting the epipolar occlusion between the virtual LiDAR camera and RGB camera.", + "bbox": [ + 212, + 750, + 787, + 839 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "S. Zhu et al.", + "bbox": [ + 271, + 114, + 354, + 127 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/700d9a39658aced8eb8f08322e87f0c64546d4df343c2b682e6e549186f2c061.jpg", + "image_caption": [ + "Rotation" + ], + "image_footnote": [], + "bbox": [ + 217, + 143, + 473, + 188 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/812966440980895d7fe174e0a8d7d4de5547b26ea9239a85dcd0e2f4c612aef9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 478, + 143, + 524, + 172 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/9f24f8901b2622a3e8a412d866d44e0f1f584778dde342527c930495e1f5322c.jpg", + "image_caption": [ + "Fig. 6: Epipolar Occlusion with Pure Rotation Movement. We apply a rotation transformation along the $x-z$ axis in the camera coordinate system. As indicated by Lemma 2, a pure rotation movement does not lead to projective artifacts." + ], + "image_footnote": [], + "bbox": [ + 529, + 145, + 684, + 186 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Projective Artifacts are Epipolar Occlusion. We ensure that our virtual LiDAR has a larger field of view than the RGB camera by adjusting the resolution of the virtual LiDAR camera. Denote the relative movement between the LiDAR and RGB camera as $\\mathbf{P} = [\\mathbf{R}\\mathbf{t}]$ , where the auto-calibrated center o is integrated into t. Denote the projected pixels on epipolar left (LiDAR) and epipolar right (RGB) as p and q, and their pixel depths as $d$ and $d^{\\prime}$ . We have:", + "bbox": [ + 214, + 256, + 782, + 347 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {p} = \\pi (\\mathbf {v} \\mid \\mathbf {K} _ {l}), \\quad \\mathbf {q} = \\pi (\\mathbf {R v} + \\mathbf {t} \\mid \\mathbf {K} _ {r}). \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 362, + 358, + 785, + 376 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We then formally define the epipolar occlusion $\\mathcal{O}$ on RGB view as:", + "bbox": [ + 215, + 385, + 697, + 401 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {O} = \\left\\{\\mathbf {q} _ {1} \\mid \\forall \\mathbf {q} _ {1}, \\mathbf {q} _ {2} \\in \\mathcal {Q}, d _ {\\mathbf {q} _ {1}} ^ {\\prime} > d _ {\\mathbf {q} _ {2}} ^ {\\prime}, \\| \\mathbf {q} _ {1} - \\mathbf {q} _ {2} \\| _ {2} = 0 \\right\\}. \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 308, + 412, + 785, + 430 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Eq. (7) simply suggests if two pixels in RGB camera overlap with each other, then the one $\\mathbf{q}_1$ with larger depth is occluded by the one with smaller depth $\\mathbf{q}_2$ . Occlusion detection finds all occluded pixels $\\mathbf{q}_1$ , i.e., the set $\\mathcal{O}$ .", + "bbox": [ + 214, + 439, + 782, + 484 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.2 Epipolar Occlusion under Rotation and Translation", + "text_level": 1, + "bbox": [ + 214, + 506, + 684, + 522 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Suppose the relative camera motion among the Fig. 4a binocular system is an arbitrary camera pose $\\mathbf{P} = [\\mathbf{R}\\mathbf{t}]$ . Set $\\mathbf{I}$ as the identity matrix. RePLAy decomposes $\\mathbf{P}$ into a sequential pure rotation and pure translation movement:", + "bbox": [ + 214, + 531, + 782, + 578 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {P} _ {1} = \\left[ \\begin{array}{l l} \\mathbf {R} & \\mathbf {0} \\end{array} \\right], \\mathbf {P} _ {2} = \\left[ \\begin{array}{l l} \\mathbf {I} & \\mathbf {t} \\end{array} \\right], \\mathbf {P} = \\mathbf {P} _ {2} \\mathbf {P} _ {1}. \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 362, + 590, + 785, + 609 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We outline pure rotation movement in Lemma 2 and Fig. 6. The pure translation movement is illustrated in Lemma 3 and Fig. 7. Without loss of generality, we assume the epipolar left and right cameras share the same intrinsic, i.e., $\\mathbf{K}_l = \\mathbf{K}_r$ . When with different intrinsics, we apply a resize and crop operation defined by $\\mathbf{A} = \\mathbf{K}_l\\mathbf{K}_r^{-1}$ to unify the intrinsic.", + "bbox": [ + 214, + 614, + 782, + 689 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Lemma 2. Pure rotation movement does not cause epipolar occlusion.", + "text_level": 1, + "bbox": [ + 214, + 700, + 753, + 715 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We know from epipolar geometry [25] that pure rotation induces homography between two views. In other words, there is a one-to-one mapping between pixels before and after rotation. See Supp for more information. Fig. 6 illustrates Lemma 2 with a visual example.", + "bbox": [ + 214, + 726, + 782, + 786 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Lemma 3. Pure translation movement causes epipolar occlusion along the direction of the epipolar line.", + "bbox": [ + 214, + 796, + 785, + 825 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "RePLAY", + "bbox": [ + 671, + 114, + 730, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 114, + 785, + 126 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/9a4a4358b3ec92f4a432e9d33a295893771ff5265b38ee891a5d65c36701a703.jpg", + "image_caption": [ + "Fig. 7: Epipoplar Occlusion with Pure Translation Movement. A background pixel $\\mathbf{p}_1$ is gradually occluded by a foreground pixel $\\mathbf{p}_2$ under pure translation. We formulate three binocular systems at different baseline distances. For simplicity, we set the depth of $\\mathbf{p}_1$ to infinity, making its 2D location static. Direction vector $\\mathbf{n}$ is set to point from $\\mathbf{p}$ to $\\mathbf{q}$ . The scalar $t$ is marked by blue line segment. From Lemma 3, (a) The pixel $\\mathbf{p}_1$ is only occluded by pixel $\\mathbf{p}_2$ at the reverse direction of epipolar line (red line segment). (b) $\\mathbf{p}_1$ is not occluded. (c) $\\mathbf{p}_1$ is occluded following Eq. (7). (d) $\\mathbf{p}_1$ is considered occluded following Eq. (17). Intuitively, the occluded pixel $\\mathbf{p}_1$ (in a) is \"caught up\" (in c) and \"surpassed\" (in d) by $\\mathbf{p}_2$ while moving along the epipolar line." + ], + "image_footnote": [], + "bbox": [ + 217, + 143, + 787, + 344 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Proof. As pure rotation does not cause epipolar occlusion, we apply the translation movement over the artifacts-free depthmap after the pure rotation. With slight abuse of the notation, we set the pixels after pure rotation movement as $\\mathbf{p}$ and $\\mathbf{q}$ . The epipolar line $\\mathbf{l_p}$ on the right image from the left image pixel $\\mathbf{p}$ is:", + "bbox": [ + 212, + 496, + 787, + 556 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {l} _ {\\mathbf {p}} = \\mathbf {K} ^ {- \\top} \\left[ \\mathbf {t} _ {\\times} \\right] \\mathbf {R} \\mathbf {K} ^ {- 1} \\mathbf {p} = \\mathbf {K} ^ {- \\top} \\left[ \\mathbf {t} _ {\\times} \\right] \\mathbf {K} ^ {- 1} \\mathbf {p}, \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 367, + 565, + 784, + 583 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $[\\mathbf{t}_{\\times}]$ is the cross product in matrix form [71]. Due to a pure translation movement, the rotation $\\mathbf{R}$ is removed as being an identity matrix. Further, the left and right epipolar lines overlap with each other in pure translation:", + "bbox": [ + 212, + 590, + 787, + 638 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {l} _ {\\mathbf {q}} = - \\mathbf {K} ^ {- \\top} \\left[ \\mathbf {t} _ {\\times} \\right] \\mathbf {K} ^ {- 1} \\mathbf {q}, \\mathbf {p} ^ {\\top} \\mathbf {l} _ {\\mathbf {q}} = 0, \\mathbf {p} ^ {\\top} \\mathbf {l} _ {\\mathbf {p}} = 0. \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 349, + 646, + 784, + 664 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Also shown in Fig. 7, pixel $\\mathbf{q}$ is hence a function of pixel $\\mathbf{p}$ :", + "bbox": [ + 215, + 672, + 645, + 688 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {q} = \\mathbf {p} + t \\cdot \\mathbf {n}, t = \\| \\mathbf {q} - \\mathbf {p} \\| _ {2} = \\mathbf {n} ^ {\\intercal} (\\mathbf {q} - \\mathbf {p}), \\mathbf {n} = (\\mathbf {q} - \\mathbf {p}) / \\sqrt {\\| \\mathbf {q} - \\mathbf {p} \\| _ {2} ^ {2}}, \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 243, + 696, + 784, + 722 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $t$ is the moving distance, similar to the disparity in stereo vision. We define the vector $\\mathbf{n}$ pointed from $\\mathbf{p}$ to $\\mathbf{q}$ as the direction of the epipolar line $\\mathbf{l}$ . $\\mathbf{p}$ and $\\mathbf{q}$ are related by the equation $d^{\\prime}\\mathbf{q} = d\\mathbf{K}_{r}\\mathbf{R}\\mathbf{K}_{l}^{-1}\\mathbf{p} + \\mathbf{K}_{r}\\mathbf{t}$ . Expanding under pure translation movement gives us:", + "bbox": [ + 212, + 729, + 787, + 790 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\nt ^ {2} = \\left(p _ {x} - q _ {x}\\right) ^ {2} + \\left(p _ {y} - q _ {y}\\right) ^ {2} = \\frac {\\left(x + z \\cdot q _ {x}\\right) ^ {2} + \\left(y + z \\cdot q _ {y}\\right) ^ {2}}{\\left(d ^ {\\prime} - z\\right) ^ {2}} = \\frac {\\left(x + z \\cdot q _ {x}\\right) ^ {2} + \\left(y + z \\cdot q _ {y}\\right) ^ {2}}{d ^ {2}}. \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 212, + 796, + 785, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "S. Zhu et al.", + "bbox": [ + 269, + 114, + 354, + 127 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/3d39d62921456f7d1b58e19c6cfada73626e89c1d5bded49efc2025431470b30.jpg", + "image_caption": [ + "(a) Virtual LiDAR Depthmap $\\mathbf{D}_l$" + ], + "image_footnote": [], + "bbox": [ + 225, + 143, + 501, + 194 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/a4be020daa3bbc26cece9d94268f9f2879bc3a6eccb4f714eb5f2b0049a9348a.jpg", + "image_caption": [ + "(b) Densified Virtual LiDAR Depthmap $\\overline{\\mathbf{D}}_l$", + "Fig. 8: Per-pixel Occlusion Examination. To overcome LiDAR sparsity, for a query pixel, we verify inspection pixels along epipolar line $\\mathcal{L}_{\\mathbf{p}_1}$ following Eq. (17) to determine occlusion. Depth $\\overline{\\mathbf{D}}_l$ is densified with nearest-neighbor interpolation." + ], + "image_footnote": [], + "bbox": [ + 509, + 143, + 784, + 195 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Please see the derivation in Supp. (Appendix B). Note, we assume pixels $\\mathbf{p}$ and $\\mathbf{q}$ are respectively visible to the left and right camera. This suggests $d > 0$ and $d = d' - z > 0$ . Eq. (12) suggests moving distance $t^2$ monotonously decreases as the projected depth $d'$ increases. It fits the intuition where a foreground object has a larger movement than a background object. From Eq. (7), when with occlusion, we have $\\mathbf{q}_1 = \\mathbf{q}_2$ , and $d_{\\mathbf{q}_1} < d_{\\mathbf{q}_2}$ . Combine Eq. (11) and Eq. (12):", + "bbox": [ + 214, + 277, + 782, + 369 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {q} _ {1} = \\mathbf {q} _ {2}, \\text {a n d} d _ {\\mathbf {q} _ {1}} < d _ {\\mathbf {q} _ {2}} \\longrightarrow t _ {1} ^ {2} > t _ {2} ^ {2} \\longrightarrow \\mathbf {n} ^ {\\intercal} (\\mathbf {p} _ {1} - \\mathbf {p} _ {2}) < 0. \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 292, + 378, + 784, + 396 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "From Eq. (13), if pixel $\\mathbf{q}_1$ is occluded by $\\mathbf{q}_2$ in the right view, pixel $\\mathbf{q}_2$ must move a longer distance along the epipolar line than $\\mathbf{q}_1$ . This suggests, in the left view, $\\mathbf{p}_2$ resides at the \"back\" of $\\mathbf{p}_1$ w.r.t. the moving direction. In other words, $\\mathbf{p}_1$ is only occluded by other pixels along the reverse direction of epipolar line.", + "bbox": [ + 214, + 405, + 782, + 467 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Epipolar Occlusion Judgement. Combine Lemma 2 and 3, a pixel $\\mathbf{p}_1$ in LiDAR view is occluded by other pixels $\\mathbf{p}_2$ along the reverse of epipolar line direction. For pixel $\\mathbf{p}_1$ , the set of inspection pixels is:", + "bbox": [ + 214, + 474, + 782, + 521 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathbf {p} _ {1}} = \\left\\{\\mathbf {p} _ {2} \\mid \\mathbf {p} _ {2} = \\mathbf {p} _ {1} - \\mathbf {n} \\cdot k \\cdot \\Delta t, k \\in \\mathbb {Z} \\right\\}. \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 351, + 531, + 784, + 547 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "$k$ is an integer. At occlusion, we have $\\mathbf{q}_1 = \\mathbf{q}_2$ , combined with Eq. (11):", + "bbox": [ + 214, + 558, + 730, + 574 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {q} _ {1} = \\mathbf {q} _ {2} \\longrightarrow \\mathbf {p} _ {1} + t _ {1} \\mathbf {n} = \\mathbf {p} _ {2} + t _ {2} \\mathbf {n} \\longrightarrow \\mathbf {n} ^ {\\top} \\left(\\mathbf {p} _ {1} - \\mathbf {p} _ {2}\\right) + t _ {1} - t _ {2} = 0. \\tag {15}\n$$\n", + "text_format": "latex", + "bbox": [ + 253, + 584, + 784, + 599 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We use Fig. 7 to explain Eq. (15). In Fig. 7a, the pixel $\\mathbf{p}_1$ and $\\mathbf{p}_2$ has an initial distance $\\mathbf{n}^{\\top}(\\mathbf{p}_1 - \\mathbf{p}_2)$ . In Fig. 7c, after translation movement, each pixel moves a distance $t_1$ and $t_2$ along the epipolar line. Occlusion happens when pixel $\\mathbf{p}_2$ overlaps pixel $\\mathbf{p}_1$ . We summarize Eq. (15) into a function $g(\\cdot)$ as:", + "bbox": [ + 214, + 609, + 782, + 671 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\lambda \\leq g \\left(\\mathbf {p} _ {1}, \\mathbf {p} _ {2}, d _ {\\mathbf {p} _ {1}}, d _ {\\mathbf {p} _ {2}} \\mid \\mathbf {K}, \\mathbf {P}\\right) \\leq 0, \\tag {16}\n$$\n", + "text_format": "latex", + "bbox": [ + 377, + 681, + 784, + 699 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where the distance $t_1$ and $t_2$ are computed using intrinsic, camera motion, and depth values. In Fig. 7b, when $g(\\cdot) > 0$ , there is no occlusion. In Fig. 7c, when $g(\\cdot) = 0$ , two pixels overlap and occlusion happens. However, this is hard due to LiDAR sparsity, rasterization, and other noise sources. Hence, we relax with a threshold $\\lambda$ . In practice, we set $\\lambda$ to $-\\infty$ . This suggests the pixel $\\mathbf{p}_1$ is occluded if any pixel $\\mathbf{p}_2$ in the epipolar line \"surpasses\" it in the RGB view, as shown in Fig. 7d. Formally, we consider pixel $\\mathbf{p}_1$ occluded as:", + "bbox": [ + 214, + 708, + 782, + 813 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {O} = \\left\\{\\mathbf {p} _ {1} \\mid \\exists \\mathbf {p} _ {2} \\in \\mathcal {L} _ {\\mathbf {p} _ {1}}, g \\left(\\mathbf {p} _ {1}, \\mathbf {p} _ {2}, d _ {\\mathbf {p} _ {1}}, d _ {\\mathbf {p} _ {2}} \\mid \\mathbf {K}, \\mathbf {P}\\right) < 0 \\right\\}. \\tag {17}\n$$\n", + "text_format": "latex", + "bbox": [ + 310, + 824, + 784, + 842 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "RePLAY", + "bbox": [ + 671, + 114, + 730, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Eq. (17) is also a necessary condition of epipolar occlusion. Due to the Li-DAR sparsity, we apply a nearest-neighbor interpolation over virtual LiDAR depthmap $\\mathbf{D}_l$ to acquire $d_{\\mathbf{p}_2}$ whenever required, as shown in Fig. 8.", + "bbox": [ + 212, + 146, + 787, + 193 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "3.3 Discussions", + "text_level": 1, + "bbox": [ + 214, + 212, + 356, + 227 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Half-Occlusion (Depth Buffering) Baseline. Sec. 2 shows that half-occlusion algorithms solve projective artifacts in stereo vision, i.e., stereo/epipolar occlusion, without knowing LiDAR point cloud. Despite a different setting, they provide a baseline solution. When preparing RGB depthmap $\\mathbf{D}_r$ , only pixels with smaller depth are preserved if multiple 3D points $\\mathbf{v}$ have duplicate 2D locations after rasterization. The raw LiDAR depthmap uses half-occlusion. So, we refer Half-Occlusion as \"Raw\" following [23].", + "bbox": [ + 212, + 236, + 787, + 343 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Does RePLAy simply densify the sparse depthmap to remove occlusion? No. Densification alone is insufficient to remove occlusion. RePLAy enables robust densification by formulating a stereo system with an auto-calibration procedure to remove occlusion.", + "bbox": [ + 212, + 343, + 787, + 402 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Universal Existence of Projective Artifacts. Fig. 2 pictorially depicts that projective artifacts exist under all LiDAR and camera sensor settings, while Fig. 7 shows that these artifacts scale up with a larger baseline distance between sensors. We now show that this indeed is the case. The artifact width distribution plot on the outdoor KITTI and the indoor NYUv2 [53] datasets are shown in Fig. 10 in the Supp. The LiDAR and RGB camera in KITTI have a baseline distance of $0.4m$ [22]. We set the baseline distance to $0.1m$ following the AR headset configuration of Fig. 2. Despite the indoor image having a lesser resolution than outdoor ( $480 \\times 640$ compared to $375 \\times 1242$ ), we observe a similar distribution. Detailed definitions and evaluation protocols are in Supp. The analysis confirms that projective artifacts exist in both indoor and outdoor datasets.", + "bbox": [ + 212, + 402, + 787, + 570 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 214, + 590, + 375, + 609 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "We benchmark RePLAy on depthmap quality, MonoDE and Mono3DOD tasks. Sec. 4.1 assesses depthmap quality, while Sec. 4.2 and Sec. 4.3. report MonoDE and Mono3DOD results respectively.", + "bbox": [ + 212, + 622, + 787, + 667 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Datasets. We experiment on KITTI [21], nuScenes [10] and Waymo [56] datasets. KITTI provides stereo ground truth while the others do not.", + "bbox": [ + 212, + 667, + 787, + 696 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Data Splits. We use the following splits of the above datasets:", + "bbox": [ + 214, + 696, + 674, + 713 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- KITTI Eigen [15]: 23,158 training and 652 validation samples of raw KITTI.", + "- KITTI Stereo [46]: 200 training and 200 testing samples out of which 145 training scenes are mapped to the raw KITTI dataset.", + "- nuScenes Frontal Val: 28,130 training and 2,436 random validation samples (from 6,019 samples) from front camera and top LiDAR.", + "- Waymo Frontal Val: 52,386 training and 40,839 validation samples from the front camera and top LiDAR. We construct its training set by sampling every third frame from the training sequences as in [34]." + ], + "bbox": [ + 218, + 718, + 782, + 838 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "S. Zhu et al.", + "bbox": [ + 271, + 114, + 354, + 127 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/8a24d6b611b0f1ac654f4e5127ffc99c83e8b8c72521d63a447c226fda7a5398.jpg", + "table_caption": [ + "Table 2: Depthmap Quality Assessment with baselines and RePLAy. We adopt KITTI semi-dense depthmap and KITTI Stereo disparity map as the reference GT. [Key: Best, Second Best, Mod. = Modified]" + ], + "table_footnote": [], + "table_body": "
GTEval RegionDepthmapδ0.5↑δ1↑δ2↑RMS ↓RMSlog ↓ARel ↓SRel ↓log10 ↓SIlog ↓
Semi-DenseAllRaw0.9730.9830.9901.9350.0870.0290.3320.0108.660
Mod. Half-Occ0.9830.9910.9961.3470.0570.0190.1130.0075.712
RePLAy0.9850.9940.9981.2410.0460.0160.0630.0074.592
StereoAllRaw0.9500.9620.9704.1980.1690.0831.6040.02416.691
Mod. Half-Occ0.9710.9820.9882.5410.1050.0470.5460.01610.340
RePLAy0.9750.9860.9912.2240.0860.0340.3540.0148.362
ForegroundRaw0.8590.8710.8928.7260.3170.2285.9740.05829.772
Mod. Half-Occ0.9190.9310.9495.2850.1990.1542.3660.03319.250
RePLAy0.9250.9370.9564.8370.1790.1001.9290.03017.319
BackgroundRaw0.9670.9800.9862.7080.1130.0520.7740.01711.037
Mod. Half-Occ0.9800.9910.9951.7010.0700.0320.2100.0136.638
RePLAy0.9820.9930.9981.4390.0520.0270.0850.0124.722
", + "bbox": [ + 220, + 198, + 782, + 343 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/5e3e964925eb782e13d46115f93b022b7d2dcd7c4539d50a34cbdc014602cf62.jpg", + "table_caption": [ + "Table 3: Depthmap Quality Assessment on Removed Artifacts. RePLAy outperforms the Mod. half-occlusion baseline. For removed artifacts, a worse performance suggests a better algorithm. [Key: Best]" + ], + "table_footnote": [], + "table_body": "
GTEval RegionDepthmapδ0.5↓δ1↓δ2↓RMS↑RMSlog↑ARel↑SRel↑log10↑
Semi-DenseAllMod. Half-Occ0.8140.8460.8935.8500.2750.1963.7620.053
RePLAy0.2740.3630.54812.0840.5870.77815.8510.202
", + "bbox": [ + 220, + 409, + 782, + 446 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "- KITTI Detection Val: 3,748 training and 3,762 validation images from the left camera for the Mono3DOD task.", + "bbox": [ + 218, + 455, + 784, + 484 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Ground Truth (GT) Depthmaps. We use the following as GT depthmaps:", + "bbox": [ + 215, + 491, + 781, + 508 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- KITTI Eigen: We use the semi-dense depthmap available in KITTI.", + "- KITTI Stereo: We generate depthmaps from GT disparity maps, which are denser than semi-dense depthmaps.", + "- nuScenes and Waymo Frontal Val: We generate GT depthmaps by projecting only those LiDAR points present inside GT 3D bounding boxes. To prepare a reference depthmap, we follow [49] to filter LiDAR point cloud with GT 3D bounding boxes labels. Correspondingly, the evaluation is confined to the bounding box area, as only depth labels in that region are retained." + ], + "bbox": [ + 218, + 510, + 785, + 631 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4.1 Depthmap Quality Assessment", + "text_level": 1, + "bbox": [ + 215, + 651, + 516, + 666 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Baselines. We compare RePLAy to raw depthmap, which is the half-occlusion (depth buffering) method and a modified half-occlusion method. The modified half-occlusion modifies the half-occlusion method (Sec. 3.3) by back-projecting the estimated dense virtual depthmap $(\\overline{\\mathbf{D}}_l)$ onto the RGB camera and only preserving pixels in $\\mathbf{D}_r$ that are less than their back-projected values.", + "bbox": [ + 212, + 672, + 787, + 748 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Results. Tab. 2 compares the RePLAy depthmap with these baselines processed depthmaps. Tab. 2 verifies that RePLAy depthmap outperforms the raw and modified half-occlusion baselines. RePLAy depthmap agrees more with semidense or disparity depthmap after removing the projective artifacts. Fig. 3 shows qualitative comparisons. We believe projective artifacts persist in the half-occlusion method because of the LiDAR sparsity, leading to inferior depthmaps.", + "bbox": [ + 212, + 750, + 787, + 840 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "RePLAY", + "bbox": [ + 671, + 114, + 730, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 784, + 126 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/f6d8fe105746e7ed580956fc4a88f7c73fddf59f58ab52cf0fb5852352871a65.jpg", + "table_caption": [ + "Table 4: Auto-Calibration Ablation on Depthmap Quality Assessment with semidense GT depthmap. [Key: Best]" + ], + "table_footnote": [], + "table_body": "
DepthmapAutoCalibδ0.5↑δ1↑δ2↑RMS ↓RMSlog ↓ARel ↓SRel ↓log10 ↓SIlog ↓
RePLAy×0.9810.9900.9951.4730.0570.0190.1110.0075.680
✓0.9850.9940.9981.2410.0460.0160.0630.0074.592
", + "bbox": [ + 250, + 184, + 753, + 224 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/b438e377a79a854e06ec24d759262001669a2d44497d775e8d819040635487b3.jpg", + "table_caption": [ + "Table 5: MonoDE Results on KITTI Stereo. ZoeDepth [5] trained with RePLAy depthmaps outperforms raw and semi-dense [57] depthmaps on foreground. [Key: Best, Second Best]." + ], + "table_footnote": [], + "table_body": "
Eval RegionDepthmapδ1↑δ2↑δ3↑RMS ↓RMSlog ↓ARel ↓SRel ↓log10 ↓SIlog ↓
AllRaw0.9550.9820.9922.6470.1160.0670.4200.02611.324
Semi-Dense0.9770.9940.9972.2740.0820.0430.2010.0187.971
RePLAy0.9740.9920.9962.2880.0890.0490.2480.0208.588
ForegroundRaw0.8860.9470.9803.6920.1800.1271.1120.02216.518
Semi-Dense0.9760.9900.9942.0930.0910.0560.3440.0228.342
RePLAy0.9450.9780.9892.9390.1280.0810.6420.02911.798
BackgroundRaw0.9730.9930.9982.2940.0840.0480.1880.0207.981
Semi-Dense0.9780.9960.9982.2830.0760.0400.1670.0177.239
RePLAy0.9810.9970.9992.0840.0700.0390.1380.0176.681
", + "bbox": [ + 250, + 292, + 753, + 410 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Foreground Improvement. Tab. 2 rows 5 and 6 shows a significant improvement in foreground depthmap quality. On metric $\\delta_{0.5}$ , we reduce the errors by $46.8\\% = 1 - \\frac{1 - 0.925}{1 - 0.859}$ . This aligns with our observation that projective artifacts mostly occur over foreground objects, confirming Fig. 1.", + "bbox": [ + 215, + 439, + 784, + 500 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Comparison on Removed Artifacts. We next compare RePLAy on removed artifacts in Tab. 3. Tab. 3 shows that that in-addition to removing some artifacts, the modified half-occlusion also removes some non-artifacts (good points). Ideally, the removed points should be artifacts; therefore, the best algorithm should perform the worst on the quality assessment of removed points and hence the reversal in the directional indicator for Tab. 3. Qualitative comparison between raw, modified half-occlusion and RePLAy is provided in Supp. (Fig. 11).", + "bbox": [ + 215, + 500, + 784, + 606 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Auto-Calibration Ablation. We next conduct ablations of the Auto-Calibration module (Sec. 3.1) in Tab. 4. Tab. 4 confirms the effectiveness of Auto-Calibration module in RePLAy, since Auto-Calibration provides $15\\%$ and $19\\%$ relative improvement in ARel and $\\mathrm{SI}_{\\log}$ metrics respectively.", + "bbox": [ + 215, + 606, + 795, + 667 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "4.2 MonoDE: Monocular Depth Estimation", + "text_level": 1, + "bbox": [ + 215, + 691, + 586, + 707 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Baselines. We compare RePLAy to raw and Semi-Dense methods. We use ZoeDepth [5] as the depth model due to its superior generalization.", + "bbox": [ + 215, + 718, + 784, + 748 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "KITTI Stereo and Eigen Results. We train ZoeDepth [5] from scratch using depthmap with and without applying RePLAy on KITTI Eigen split. We evaluate on KITTI Stereo and Eigen split in Tabs. 5 and 6 respectively. Tab. 5 shows that ZoeDepth trained with RePLAy depthmap significantly outperforms the baselines. We observe despite semi-dense GT [57] using additional stereo cameras, RePLAy outperforms Semi-Dense especially in the background. This", + "bbox": [ + 215, + 750, + 785, + 839 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "S. Zhu et al.", + "bbox": [ + 271, + 114, + 354, + 127 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/6f220f610d772c75d69f6e1415071e503e51b4be7125aac7fa4a21db92d575bb.jpg", + "image_caption": [ + "(a) RGB Input" + ], + "image_footnote": [], + "bbox": [ + 217, + 143, + 405, + 207 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/de6d75883c0fb02382335191b6cf6e8f0c901164dc6d35b77bde25f1915bcea6.jpg", + "image_caption": [ + "(b) Raw depthmaps", + "Fig. 9: MonoDE Qualitative Results on KITTI Eigen dataset. ZoeDepth [5] trained with RePLAy depthmaps (c) outperforms the one with raw depthmaps (b), especially on foreground." + ], + "image_footnote": [], + "bbox": [ + 406, + 143, + 594, + 207 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/c4d84a85b31e9289af6581e81dc94bc23919ed25f0c87387e078c878bf636a4d.jpg", + "image_caption": [ + "(c) RePLAy depthmaps" + ], + "image_footnote": [], + "bbox": [ + 596, + 143, + 784, + 207 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/dfa1bd433fbd031f772b7a452b3f82c6e561ea6ce0f098a5f3e169461a96a8c9.jpg", + "table_caption": [ + "Table 6: MonoDE Results. ZoeDepth trained with RePLAy outperforms the baselines in foreground regions of all datasets. [Key: Best]" + ], + "table_footnote": [], + "table_body": "
Data SplitEval RegionDepthmapδ1↑δ2↑δ3↑RMS↓RMSlog↓ARel↓SRel↓log10↓SIlog↓
KITTI EigenAllRaw0.8690.9590.9872.9110.1630.1240.6060.04613.501
RePLAy0.9270.9850.9972.1850.1100.0820.2870.0338.674
nuScenes ValForegroundRaw0.9710.9870.9931.3470.0750.0490.2450.0186.339
RePLAy0.9810.9930.9961.1500.0590.0390.1370.0164.905
Waymo ValForegroundRaw0.9410.9880.9963.4680.1110.0790.4870.0329.116
RePLAy0.9620.9930.9983.0570.0910.0630.3670.0267.322
", + "bbox": [ + 220, + 332, + 781, + 414 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/38a4a5f90067e98159cd3569801399812b26c9ccd653b3971d3ec1a45e3ab633.jpg", + "table_caption": [ + "Table 7: MonoDE Analysis to Environmental Conditions on nuScenes Val split. ZoeDepth trained with RePLAY depthmaps shows consistent improvements on foreground under all environmental conditions. [Key: Best]" + ], + "table_footnote": [], + "table_body": "
ConditionSamplesDepthmapδ1↑δ2↑δ3↑RMS ↓RMSlog ↓ARel ↓SRel ↓log10 ↓SIlog ↓
Rain864Raw0.9840.9910.9951.3380.0580.0400.1630.0164.859
RePLAy0.9890.9930.9951.2220.0500.0340.1270.0154.110
Night533Raw0.9650.9820.9901.3470.0890.0590.3350.0217.481
RePLAy0.9780.9910.9951.0630.0670.0450.1720.0175.374
Sunny1039Raw0.9640.9870.9931.3520.0810.0530.2670.0196.985
RePLAy0.9760.9930.9961.1340.0620.0400.1280.0165.324
All Tab. 62436Raw0.9710.9870.9931.3470.0750.0490.2450.0186.339
RePLAy0.9810.9930.9961.1500.0590.0390.1370.0164.905
", + "bbox": [ + 220, + 484, + 781, + 589 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "may be due to [57] aggressively removing the background pixels around the foreground object boundaries. Fig. 9 shows this result visually. ZoeDepth trained with the raw depthmap learns strip-shaped artifacts (green boxes).", + "bbox": [ + 212, + 608, + 782, + 652 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Foreground Improvement. We additionally ablate depth performance on foreground, background, and entire image using the GT segmentation mask in Tab. 5. Tab. 5 shows that ZoeDepth trained with RePLAy depthmap outperforms in the foreground area, which agrees with the conclusion of Sec. 4.1.", + "bbox": [ + 212, + 655, + 782, + 715 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "nuScenes and Waymo Frontal Val Results. We benchmark MonoDE on two datasets without stereo sensors: nuScenes [10] and Waymo [56] in Tab. 6. Tab. 6 confirms that RePLAy achieves significant improvements on these datasets.", + "bbox": [ + 212, + 717, + 782, + 762 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Effect of Environmental Conditions. We next analyze the effect of environmental conditions on the nuScenes Val split in Tab. 7. We categorize the nuScenes [10] val set into rainy, night, and sunny following their description tag (nuScenes-devkit GitHub #124). Tab. 7 shows that RePLAy is robust across multiple conditions.", + "bbox": [ + 212, + 763, + 782, + 839 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "RePLAY", + "bbox": [ + 671, + 114, + 730, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "4.3 Mono3DOD: Monocular 3D Object Detection", + "text_level": 1, + "bbox": [ + 215, + 146, + 638, + 161 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "We finally verify the impact of RePLAy on the Mono3DOD task.", + "bbox": [ + 212, + 171, + 683, + 185 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Baseline. We compare RePLAy to raw method. We take SoTA 3D detectors, CaDDN [51] and MonoDTR [29], to showcase the impact of cleaning projected LiDAR depthmap with RePLAy, since these detectors use depthmap as an auxiliary training task. The two methods cover both convolutional and transformer architectures, making our benchmarking comprehensive. However, these two SoTA detectors use depth supervision in training differently.", + "bbox": [ + 212, + 186, + 785, + 276 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "CaDDN. CaDDN uses the densified raw depthmaps of [33]. For CaDDN experiments, we apply RePLAy to the projected depthmaps and then densify those maps using the same depth completion algorithm [33] and parameters of [51].", + "bbox": [ + 212, + 277, + 457, + 397 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "MonoDTR. MonoDTR [29] precomputes their GT depthmaps by $4\\times$ downsampling raw projected depthmaps. For MonoDTR experiments, we clean the projected depthmaps and then $4\\times$ down-sampling.", + "bbox": [ + 212, + 398, + 457, + 502 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Results. Tab. 8 shows results of retraining both detectors with our depthmaps on KITTI Val split. Cleaned depthmaps improves the performance of CaDDN and MonoDTR on most of the metrics. The biggest gains appear in the Easy", + "bbox": [ + 212, + 503, + 460, + 609 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "category. This is expected as Easy cars are closer to the ego camera and receive the maximum number of LiDAR points. Hence, improving the projected depthmap by RePLAy benefits the Easy category more.", + "bbox": [ + 212, + 609, + 785, + 655 + ], + "page_idx": 13 + }, + { + "type": "table", + "img_path": "images/8d5f92f42561477292d5a6d12bc3675a8608b4f49aec128cd148b47169fad918.jpg", + "table_caption": [ + "Table 8: KITTI Val Mono3DOD Results on Car category. Using RePLAy depthmaps in training boosts the Mono3DOD performance. [Key: Best, Second Best, $* =$ Reported, $\\dagger=$ Retrained, Mod= Moderate.]" + ], + "table_footnote": [], + "table_body": "
MethodDepthmapIoU3D > 0.7
AP 3D|R40[%]↑
EasyModHard
M3D-RPN [8]-14.5311.078.65
MonoPair [12]-16.2812.3010.42
MonoDLE [44]-17.4513.6611.68
Kinematic [9]-19.7614.1010.47
GrooMeD [35]-19.6714.3211.27
MonoRUn [11]-20.0214.6512.61
D4LCN [13]Raw*22.3216.2012.30
DFR-Net [73]Raw*24.8117.7814.41
CaDDN [51]Raw†21.6015.6713.24
RePLAy22.8815.7013.18
MonoDTR [29]Raw†24.5018.6615.69
RePLAy26.1518.7715.69
", + "bbox": [ + 478, + 375, + 787, + 570 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 678, + 359, + 694 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "We propose an analytical solution to remove projective artifacts in LiDAR depthmap. Our parameter-free method, RePLAy, applies to diverse LiDAR and RGB camera sensor settings. RePLAy only requires the relative position between LiDAR and RGB camera. We verify its benefits with unanimous improvements on SoTA MonoDE and Mono3DOD methods. We release processed depthmaps of five major AV datasets to benefit the community.", + "bbox": [ + 212, + 709, + 785, + 800 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Limitation. We use nearest-neighbor interpolation to densify virtual LiDAR depthmaps. Advanced strategies, e.g., depth completion, may improve accuracy.", + "bbox": [ + 212, + 800, + 785, + 832 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "S. Zhu et al.", + "bbox": [ + 271, + 114, + 354, + 127 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 143, + 323, + 160 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. The KITTI Vision Benchmark Suite. http://www.cvlabs.net/datasets/kitti/eval_object.php?obj_benchmark=3d (2022), accessed: 2022-07-03 2, 3", + "2. Arbelaez, P., Maire, M., Fowlkes, C., Malik, J.: Contour detection and hierarchical image segmentation. TPAMI (2010) 4", + "3. Bai, X., Hu, Z., Zhu, X., Huang, Q., Chen, Y., Fu, H., Tai, C.L.: Transfusion: Robust lidar-camera fusion for 3D object detection with transformers. In: CVPR (2022) 3", + "4. Bhat, S.F., Alhashim, I., Wonka, P.: Adabins: Depth estimation using adaptive bins. In: CVPR (2021) 4", + "5. Bhat, S.F., Birkl, R., Wofk, D., Wonka, P., Müller, M.: ZoeDepth: Zero-shot transfer by combining relative and metric depth. arXiv preprint arXiv:2302.12288 (2023) 2, 3, 4, 12, 13", + "6. Birchfield, S., Tomasi, C.: Depth discontinuities by pixel-to-pixel stereo. IJCV (1999) 4", + "7. Bobick, A.F., Intille, S.S.: Large occlusion stereo. IJCV (1999) 4", + "8. Brazil, G., Liu, X.: M3D-RPN: Monocular 3D region proposal network for object detection. In: ICCV (2019) 14", + "9. Brazil, G., Pons-Moll, G., Liu, X., Schiele, B.: Kinematic 3D object detection in monocular video. In: ECCV (2020) 14", + "0. Caesar, H., Bankiti, V., Lang, A., Vora, S., Liong, V., Xu, Q., Krishnan, A., Pan, Y., Baldan, G., Beijbom, O.: nuScenes: A multimodal dataset for autonomous driving. In: CVPR (2020) 1, 2, 3, 10, 13", + "1. Chen, H., Huang, Y., Tian, W., Gao, Z., Xiong, L.: MonoRun: Monocular 3D object detection by reconstruction and uncertainty propagation. In: CVPR (2021) 14", + "2. Chen, Y., Tai, L., Sun, K., Li, M.: MonoPair: Monocular 3D object detection using pairwise spatial relationships. In: CVPR (2020) 14", + "3. Ding, M., Huo, Y., Yi, H., Wang, Z., Shi, J., Lu, Z., Luo, P.: Learning depth-guided convolutions for monocular 3D object detection. In: CVPR Workshops (2020) 14", + "4. Ebadi, K., Bernreiter, L., Biggie, H., Catt, G., Chang, Y., Chatterjee, A., Denniston, C.E., Deschênes, S.P., Harlow, K., Khattak, S., et al.: Present and future of slam in extreme underground environments. arXiv preprint arXiv:2208.01787 (2022) 2", + "5. Eigen, D., Puhrsch, C., Fergus, R.: Depth map prediction from a single image using a multi-scale deep network. NeurIPS (2014) 10", + "6. Filgueira, A., González-Jorge, H., Lagüela, S., Díaz-Vilarino, L., Arias, P.: Quantifying the influence of rain in lidar performance. Measurement (2017) 3", + "7. Fua, P.: A parallel stereo algorithm that produces dense depth maps and preserves image features. Machine vision and applications (1993) 4", + "8. Garg, K., Nayar, S.K.: When does a camera see rain? In: ICCV (2005) 3", + "9. Garg, K., Nayar, S.K.: Vision and rain. IJCV (2007) 3", + "20. Garvelmann, J., Pohl, S., Weiler, M.: From observation to the quantification of snow processes with a time-lapse camera network. Hydrology and Earth Systems Sciences (2013) 3", + "21. Geiger, A., Lenz, P., Stiller, C., Urtasun, R.: Vision meets robotics: The KITTI dataset. IJRR (2013) 10", + "22. Geiger, A., Lenz, P., Urtasun, R.: Are we ready for autonomous driving? the KITTI vision benchmark suite. In: CVPR (2012) 10" + ], + "bbox": [ + 225, + 176, + 784, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "RePLAY", + "bbox": [ + 671, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 785, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "23. Godard, C., Mac Aodha, O., Firman, M., Brostow, G.J.: Digging into self-supervised monocular depth estimation. In: ICCV (2019) 10", + "24. Guizilini, V., Ambrus, R., Pillai, S., Raventos, A., Gaidon, A.: 3D packing for self-supervised monocular depth estimation. In: CVPR (2020) 1, 3", + "25. Hartley, R., Zisserman, A.: Multiple view geometry in computer vision. Cambridge university press (2003) 7", + "26. Herskovitz, J., Wu, J., White, S., Pavel, A., Reyes, G., Guo, A., Bigham, J.P.: Making mobile augmented reality applications accessible. In: SIGACCESS (2020) 2", + "27. Hillmann, C., Hillmann, C.: Comparing the gear vr, oculus go, and oculus quest. Unreal for Mobile and Standalone VR: Create Professional VR Apps Without Coding (2019) 2", + "28. Hirschmuller, H.: Stereo processing by semiglobal matching and mutual information. TPAMI (2007) 3", + "29. Huang, K.C., Wu, T.H., Su, H.T., Hsu, W.: MonoDTR: Monocular 3D object detection with depth-aware transformer. In: CVPR (2022) 3, 14", + "30. Huang, X., Cheng, X., Geng, Q., Cao, B., Zhou, D., Wang, P., Lin, Y., Yang, R.: The apolloscape dataset for autonomous driving. In: CVPR (2018) 3", + "31. Hutabarat, D., Rivai, M., Purwanto, D., Hutomo, H.: Lidar-based obstacle avoidance for the autonomous mobile robot. In: 2019 12th International Conference on Information & Communication Technology and System (ICTS). pp. 197-202. IEEE (2019) 2", + "32. Ishikawa, H., Geiger, D.: Occlusions, discontinuities, and epipolar lines in stereo. In: ECCV (1998) 4", + "33. Ku, J., Harakeh, A., Waslander, S.: In defense of classical image processing: Fast depth completion on the CPU. In: CRV (2018) 14", + "34. Kumar, A., Brazil, G., Corona, E., Parchami, A., Liu, X.: DEVIANT: Depth Equivariant Network for Monocular 3D Object Detection. In: ECCV (2022) 10", + "35. Kumar, A., Brazil, G., Liu, X.: GrooMeD-NMS: Grouped mathematically differentiable nms for monocular 3D object detection. In: CVPR (2021) 14", + "36. Kumar, A., Guo, Y., Huang, X., Ren, L., Liu, X.: SeaBird: Segmentation in Bird's View with Dice Loss Improves Monocular 3D Detection of Large Objects. In: CVPR (2024) 4", + "37. Kutila, M., Pyykonen, P., Holzhüter, H., Colomb, M., Duthon, P.: Automotive lidar performance verification in fog and rain. In: ITSC (2018) 3", + "38. Lee, J.H., Han, M.K., Ko, D.W., Suh, I.H.: From big to small: Multi-scale local planar guidance for monocular depth estimation. arXiv preprint arXiv:1907.10326 (2019) 4", + "39. Li, Y., Yu, A.W., Meng, T., Caine, B., Ngiam, J., Peng, D., Shen, J., Lu, Y., Zhou, D., Le, Q.V., et al.: Deepfusion: Lidar-camera deep fusion for multi-modal 3d object detection. In: CVPR (2022) 3", + "40. Li, Y., Ibanez-Guzman, J.: LiDAR for autonomous driving: The principles, challenges, and trends for automotive lidar and perception systems. IEEE Signal Processing Magazine (2020) 2", + "41. Liao, Y., Xie, J., Geiger, A.: KITTI-360: A novel dataset and benchmarks for urban scene understanding in 2D and 3D. TPAMI (2022) 1, 3, 4", + "42. Lin, J.T., Dai, D., Van Gool, L.: Depth estimation from monocular images and sparse radar data. In: IROS (2020) 4", + "43. Liu, F., Liu, X.: Voxel-based 3d detection and reconstruction of multiple objects from a single image. In: NeurIPS (2021) 4" + ], + "bbox": [ + 215, + 146, + 784, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "S. Zhu et al.", + "bbox": [ + 271, + 114, + 354, + 126 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "44. Ma, X., Zhang, Y., Xu, D., Zhou, D., Yi, S., Li, H., Ouyang, W.: Delving into localization errors for monocular 3D object detection. In: CVPR (2021) 14", + "45. Martin, D., Fowlkes, C., Malik, J.: Learning to detect natural image boundaries using local brightness, color, and texture cues. TPAMI (2004) 4", + "46. Menze, M., Geiger, A.: Object scene flow for autonomous vehicles. In: CVPR (2015) 10", + "47. Michaud, S., Lalonde, J.F., Giguere, P.: Towards characterizing the behavior of lidars in snowy conditions. In: IROS (2015) 3", + "48. Milanovic, V., Kasturi, A., Yang, J., Hu, F.: A fast single-pixel laser imager for VR/AR headset tracking. In: MOEMS and Miniaturized Systems XVI (2017) 2", + "49. Park, D., Ambrus, R., Guizilini, V., Li, J., Gaidon, A.: Is pseudo-lidar needed for monocular 3d object detection? In: ICCV (2021) 11", + "50. Piccinelli, L., Sakaridis, C., Yu, F.: iDisc: Internal discretization for monocular depth estimation. In: CVPR (2023) 2", + "51. Reading, C., Harakeh, A., Chae, J., Waslander, S.: Categorical depth distribution network for monocular 3D object detection. In: CVPR (2021) 2, 3, 14", + "52. Royo, S., Ballesta-Garcia, M.: An overview of LiDAR imaging systems for autonomous vehicles. Applied sciences (2019) 2", + "53. Silberman, N., Hoiem, D., Kohli, P., Fergus, R.: Indoor segmentation and support inference from rgbd images. ECCV (2012) 10", + "54. Spinneker, R., Koch, C., Park, S.B., Yoon, J.J.: Fast fog detection for camera based advanced driver assistance systems. In: ITSC (2014) 3", + "55. Sun, J., Li, Y., Kang, S.B., Shum, H.Y.: Symmetric stereo matching for occlusion handling. In: CVPR (2005) 4", + "56. Sun, P., Kretzschmar, H., Dotiwalla, X., Chouard, A., Patnaik, V., Tsui, P., Guo, J., Zhou, Y., Chai, Y., Caine, B., Vasudevan, V., Han, W., Ngiam, J., Zhao, H., Timofeev, A., Ettinger, S., Krivokon, M., Gao, A., Joshi, A., Zhang, Y., Shlens, J., Chen, Z., Anguelov, D.: Scalability in perception for autonomous driving: Waymo open dataset. In: CVPR (2020) 1, 2, 3, 10, 13", + "57. Uhrig, J., Schneider, N., Schneider, L., Franke, U., Brox, T., Geiger, A.: Sparsity invariant cnns. In: 3DV (2017) 3, 4, 12, 13", + "58. Wang, J., Zickler, T.: Local detection of stereo occlusion boundaries. In: CVPR (2019) 4", + "59. Wei, Y., Zhao, L., Zheng, W., Zhu, Z., Rao, Y., Huang, G., Lu, J., Zhou, J.: Surrounddepth: Entangling surrounding views for self-supervised multi-camera depth estimation. In: CoRL (2023) 2", + "60. Wei, Y., Quan, L.: Asymmetrical occlusion handling using graph cut for multi-view stereo. In: CVPR (2005) 4", + "61. Weng, J., Ahuja, N., Huang, T.S.: Two-view matching. In: ICCV (1988) 4", + "62. Wilson, B., Qi, W., Agarwal, T., Lambert, J., Singh, J., Khandelwal, S., Pan, B., Kumar, R., Hartnett, A., Pontes, J.K., et al.: Argoverse 2: Next generation datasets for self-driving perception and forecasting. arXiv preprint arXiv:2301.00493 (2023) 3", + "63. Wu, H., Wen, C., Li, W., Li, X., Yang, R., Wang, C.: Transformation-equivariant 3D object detection for autonomous driving. In: AAAI (2023) 2, 4", + "64. Wu, H., Wen, C., Shi, S., Li, X., Wang, C.: Virtual sparse convolution for multimodal 3D object detection. In: CVPR (2023) 2, 4", + "65. Xie, S., Tu, Z.: Holistically-nested edge detection. In: ICCV (2015) 4", + "66. Yang, Z., Chen, J., Miao, Z., Li, W., Zhu, X., Zhang, L.: Deepinteraction: 3D object detection via modality interaction. NeurIPS (2022) 2, 4" + ], + "bbox": [ + 212, + 146, + 787, + 839 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "RePLAY", + "bbox": [ + 671, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "67. Yu, K., Tao, T., Xie, H., Lin, Z., Liang, T., Wang, B., Chen, P., Hao, D., Wang, Y., Liang, X.: Benchmarking the robustness of lidar-camera fusion for 3D object detection. In: CVPR (2023) 3", + "68. Yuan, W., Gu, X., Dai, Z., Zhu, S., Tan, P.: Neural window fully-connected crfs for monocular depth estimation. In: CVPR (2022) 4", + "69. Zheng, K., Li, S., Qin, K., Li, Z., Zhao, Y., Peng, Z., Cheng, H.: Depth estimation via sparse radar prior and driving scene semantics. In: ACCV (2022) 4", + "70. Zhu, S., Brazil, G., Liu, X.: The edge of depth: Explicit constraints between segmentation and depth. In: CVPR (2020) 2", + "71. Zhu, S., Liu, X.: Lighteddepth: Video depth estimation in light of limited inference view angles. In: CVPR (2023) 4, 8", + "72. Zitnick, L., Kanade, T.: A cooperative algorithm for stereo matching and occlusion detection. TPAMI (2000) 4", + "73. Zou, Z., Ye, X., Du, L., Cheng, X., Tan, X., Zhang, L., Feng, J., Xue, X., Ding, E.: The devil is in the task: Exploiting reciprocal appearance-localization features for monocular 3D object detection. In: ICCV (2021) 14" + ], + "bbox": [ + 215, + 146, + 784, + 368 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "S. Zhu et al.", + "bbox": [ + 271, + 114, + 354, + 126 + ], + "page_idx": 17 + } +] \ No newline at end of file diff --git a/2024/Remove Projective LiDAR Depthmap Artifacts via Exploiting Epipolar Geometry/0a1ce34c-ca2a-42ac-a914-2a125b92f644_model.json b/2024/Remove Projective LiDAR Depthmap Artifacts via Exploiting Epipolar Geometry/0a1ce34c-ca2a-42ac-a914-2a125b92f644_model.json new file mode 100644 index 0000000000000000000000000000000000000000..7e3d11200cb6e95d6e7cc50a3ee5c1496dd3224e --- /dev/null +++ b/2024/Remove Projective LiDAR Depthmap Artifacts via Exploiting Epipolar Geometry/0a1ce34c-ca2a-42ac-a914-2a125b92f644_model.json @@ -0,0 +1,3261 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.23, + 0.141, + 0.777, + 0.187 + ], + "angle": 0, + "content": "RePLAY: Remove Projective LiDAR Depthmap Artifacts via Exploiting Epipolar Geometry" + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.213, + 0.784, + 0.229 + ], + "angle": 0, + "content": "Shengjie Zhu*, Girish Chandar Ganesan*, Abhinav Kumar, and Xiaoming Liu" + }, + { + "type": "text", + "bbox": [ + 0.308, + 0.24, + 0.696, + 0.298 + ], + "angle": 0, + "content": "Department of Computer Science and Engineering, Michigan State University, East Lansing, MI, USA, 48824 [zhusheng, ganesang, kumarab6, liuxm]@msu.edu https://shngjz.github.io/RePLAy/" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.331, + 0.787, + 0.464 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.468, + 0.788, + 0.553 + ], + "angle": 0, + "content": "Fig. 1: Remove Projective LiDAR Depthmap Artifacts. The LiDAR projected depthmap has misalignment with the RGB camera. The background scans highlighted by blue arrows incorrectly overlay the foreground such as cars and pedestrians. The artifacts persist unattended in most AV datasets including KITTI-360 [41], nuScenes [10], Waymo [56], and DDAD [24]. We propose an easy-to-use, parameter-free, and analytical solution to remove the projective artifacts, shown by green arrows." + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.572, + 0.744, + 0.794 + ], + "angle": 0, + "content": "Abstract. 3D sensing is a fundamental task for Autonomous Vehicles. Its deployment often relies on aligned RGB cameras and LiDAR. Despite meticulous synchronization and calibration, systematic misalignment persists in LiDAR projected depthmap. This is due to the physical baseline distance between the two sensors. The artifact is often reflected as background LiDAR incorrectly projected onto the foreground, such as cars and pedestrians. The KITTI dataset uses stereo cameras as a heuristic solution to remove artifacts. However most AV datasets, including nuScenes, Waymo, and DDAD, lack stereo images, making the KITTI solution inapplicable. We propose RePLAY, a parameter-free analytical solution to remove the projective artifacts. We construct a binocular vision system between a hypothesized virtual LiDAR camera and the RGB camera. We then remove the projective artifacts by determining the epipolar occlusion with the proposed analytical solution. We show unanimous improvement in the State-of-The-Art (SoTA) monocular depth estimators and 3D object detectors with the artifacts-free depthmaps." + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.806, + 0.691, + 0.82 + ], + "angle": 0, + "content": "Keywords: LiDAR \\(\\cdot\\) Autonomous Driving \\(\\cdot\\) Epipolar Geometry" + }, + { + "type": "page_footnote", + "bbox": [ + 0.221, + 0.825, + 0.373, + 0.84 + ], + "angle": 0, + "content": "* Equal Contributions" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.356, + 0.127 + ], + "angle": 0, + "content": "S. Zhu et al." + }, + { + "type": "image", + "bbox": [ + 0.229, + 0.146, + 0.455, + 0.183 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.256, + 0.184, + 0.428, + 0.197 + ], + "angle": 0, + "content": "(a) Autonomous Vehicles [10]" + }, + { + "type": "image", + "bbox": [ + 0.481, + 0.143, + 0.648, + 0.183 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.485, + 0.184, + 0.642, + 0.197 + ], + "angle": 0, + "content": "(b) Robotics Machines [14]" + }, + { + "type": "image", + "bbox": [ + 0.671, + 0.143, + 0.78, + 0.183 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.677, + 0.184, + 0.77, + 0.197 + ], + "angle": 0, + "content": "(c) Headset [26]" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.208, + 0.787, + 0.25 + ], + "angle": 0, + "content": "Fig. 2: LiDAR-RGB sensor sets have a physical baseline distance between the two sensors, resulting in Fig. 1 projective artifacts. (a) The background scanned by LiDAR incorrectly overlays on the foreground pedestrian observed by the RGB camera." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.266, + 0.375, + 0.281 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.297, + 0.785, + 0.387 + ], + "angle": 0, + "content": "LiDAR device scans the surrounding environment via emitting laser pulses into the 3D space. Its applications span across Autonomous Vehicles (AV) [40, 52, 56], AR/VR [27, 48], and Robotics [31]. Taking AV as an example, the LiDAR is synchronized and calibrated against the RGB camera. The correspondence between RGB pixels and LiDAR 3D points is determined by projecting the LiDAR 3D point onto the image plane, as shown with the Fig. 1 depthmaps." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.388, + 0.785, + 0.447 + ], + "angle": 0, + "content": "Despite ideal synchronization and calibration, there still exists systematic misalignment between the RGB camera and LiDAR. Fig. 1 highlights the projective artifacts in LiDAR depthmap with blue arrows, where the background scans incorrectly overlay on the foreground objects." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.448, + 0.785, + 0.538 + ], + "angle": 0, + "content": "The projective artifacts arise from the physical baseline distance between the RGB camera and LiDAR. This physical limitation is ubiquitous among different LiDAR-RGB sensor sets used in Robotics, AR/VR, and AV tasks (Fig. 2). The baseline distance (Fig. 2a) renders certain 3D scans exclusively visible to LiDAR, rather than the RGB camera. For those scans, their correspondence with image pixels is absent, causing Fig. 1 projective artifacts." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.539, + 0.785, + 0.598 + ], + "angle": 0, + "content": "The KITTI dataset [1] provides a heuristic solution to remove the projective artifacts, which we show in Fig. 3. However, the method requires stereo sensors, making it inapplicable to other AV datasets. Thus, the projective artifacts persist unattended in most AV datasets. Tab. 1 summarizes these datasets." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.599, + 0.785, + 0.704 + ], + "angle": 0, + "content": "We propose RePLAy: a parameter-free, non-learning, and analytical solution to resolve the projective artifacts. RePLAy only requires the relative pose between the LiDAR scanner and the RGB camera, i.e., their extrinsic matrix or calibration matrix. It applies to all AV datasets surveyed in Tab. 1 and diverse sensor settings included in Fig. 2. RePLAy does not even require RGB images, hence being intact at the asynchronization noise between the RGB camera and LiDAR. Fig. 3 provides a qualitative comparison with KITTI solution." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.705, + 0.785, + 0.78 + ], + "angle": 0, + "content": "RePLAy first adopts an auto-calibration algorithm to calibrate a hypothesized virtual camera at the origin of the LiDAR 3D point cloud. The virtual LiDAR camera formulates a stereo system with the RGB camera. We attribute the projective artifact to the epipolar occlusion in stereo vision. We propose an analytical solution for detecting epipolar occlusion using epipolar geometry." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.787, + 0.84 + ], + "angle": 0, + "content": "LiDAR projected depthmap is essential for diverse downstream applications. We select two to verify our effectiveness: monocular depth estimation (MonoDE) [5, 50, 59, 70] and monocular 3D object detection (Mono3DOD) that uses auxiliary depthmap [51, 63, 64, 66]. We benchmark their performance trained" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.673, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "RePLAY" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.787, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "image", + "bbox": [ + 0.231, + 0.145, + 0.41, + 0.262 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.26, + 0.264, + 0.376, + 0.276 + ], + "angle": 0, + "content": "(a) Raw Depthmap" + }, + { + "type": "image", + "bbox": [ + 0.418, + 0.145, + 0.597, + 0.263 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.47, + 0.265, + 0.545, + 0.276 + ], + "angle": 0, + "content": "(b) RePLAy" + }, + { + "type": "image", + "bbox": [ + 0.603, + 0.145, + 0.781, + 0.263 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.614, + 0.264, + 0.771, + 0.276 + ], + "angle": 0, + "content": "(c) KITTI [57] with Stereo" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.28, + 0.785, + 0.336 + ], + "angle": 0, + "content": "Fig. 3: Qualitative comparison with KITTI semi-dense depthmap. RePLAY only requires the LiDAR calibration matrix. KITTI [57] uses multi-frame fusion with the stereo camera to remove projective artifacts. The higher density of KITTI depthmap is from the additional sensor and temporal fusion." + }, + { + "type": "table_caption", + "bbox": [ + 0.215, + 0.341, + 0.785, + 0.369 + ], + "angle": 0, + "content": "Table 1: LiDAR Depthmap Projective Artifacts exist in most Autonomous Vehicle Datasets. KITTI uses stereo images as a heuristic solution." + }, + { + "type": "table", + "bbox": [ + 0.218, + 0.381, + 0.784, + 0.408 + ], + "angle": 0, + "content": "
KITTI [1]KITTI-360 [41]nuScenes [10]Waymo [56]DDAD [24]ApolloScape [30]Argoverse [62]
ArtifactsNoYesYesYesYesYesYes
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.441, + 0.785, + 0.502 + ], + "angle": 0, + "content": "using LiDAR depthmap cleaned with multiple algorithms and RePLAy on AV datasets. We demonstrate unanimous improvement in State-of-The-Art (SoTA) methods [5,29,51] on both tasks. We additionally release the processed depthmap of five major AV datasets to benefit the community. Our contributions are:" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.506, + 0.784, + 0.534 + ], + "angle": 0, + "content": "\\(\\checkmark\\) We utilize epipolar geometry to remove the projective artifacts in the LiDAR depthmap." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.536, + 0.784, + 0.563 + ], + "angle": 0, + "content": "\\(\\checkmark\\) We demonstrate unanimous improvement in the SoTA monocular depth and monocular 3D object detection methods." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.566, + 0.784, + 0.595 + ], + "angle": 0, + "content": "\\(\\checkmark\\) We provide processed depthmap in KITTI [1], KITTI-360 [41], nuScenes [10], Waymo [56], and DDAD [24] datasets." + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.506, + 0.784, + 0.595 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.625, + 0.388, + 0.641 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.658, + 0.785, + 0.779 + ], + "angle": 0, + "content": "LiDAR and Camera Misalignment. Multiple sources of misalignment exist between LiDAR and RGB camera. Foggy, rainy, and snowy weather negatively impacts both LiDAR [16, 37, 47] and RGB cameras [18-20, 54]. The slight drift of LiDAR and camera accumulates during vehicle operation. [3, 39, 67] investigate the robustness to the spatial misalignment. [67] benchmarks the temporal asynchronization between LiDAR and RGB camera. The projective artifact is another misalignment that universally exists in multiple AV datasets, as shown in Tab. 1. Our solution complements prior study [67] by addressing it." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Heuristic Solution for Projective Artifacts. KITTI [1, 57] dataset uses the stereo image as a heuristic solution. [57] produces a stereo reconstruction through semi-global matching [28]. They additionally augment the stereo results with 11 neighboring temporal frames. The projective artifacts are removed via" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.356, + 0.128 + ], + "angle": 0, + "content": "S. Zhu et al." + }, + { + "type": "code_caption", + "bbox": [ + 0.217, + 0.146, + 0.58, + 0.163 + ], + "angle": 0, + "content": "Algorithm 1 RePLAy: Remove Projective Artifacts" + }, + { + "type": "algorithm", + "bbox": [ + 0.216, + 0.165, + 0.787, + 0.35 + ], + "angle": 0, + "content": "Input: LiDAR Point Cloud V, Calibration Matrix \\(\\mathbf{P} = [\\mathbf{R}\\mathbf{t}]\\), Camera Intrinsic \\(\\mathbf{K}_r\\) \nOutput: Occluded Pixels \\(\\mathcal{O}\\) \n1: Auto-Calibration ▷ Sec. 3.1 \n2: Set Virtual LiDAR Camera \\(\\mathbf{K}_l\\) \n3: Produce Virtual LiDAR Depthmap \\(\\mathbf{D}_l\\) \n4: Denseify Virtual LiDAR Depthmap \\(\\overline{\\mathbf{D}}_l\\) \n5: for v ∈ V do \n6: p1, dp1 ← π(Rv | Kl) \n7: for p2 ∈ Lp1 do ▷ Eq. (14) and Fig. 8 \n8: dp2 ← bilinear interpolate \\(\\overline{\\mathbf{D}}_l\\) at pixel location p2 ▷ Fig. 8 \n9: if g(p1, p2, dp1, dp2 | Kl, Kr, [I t]) ≤ 0 then ▷ Eq. (15) \n10: {p1} ∪ O \n11: break" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.383, + 0.785, + 0.459 + ], + "angle": 0, + "content": "consistency verification between the stereo depthmap and LiDAR depthmap. Compared to ours, depthmap from [57] has improved density. However, the requirement of stereo sensors limits its application, as summarized in Tab. 1. Additionally, [57] has a complicated pipeline, and is therefore, not used even in the KITTI-360 [41] dataset, which shares a similar setup." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.461, + 0.787, + 0.597 + ], + "angle": 0, + "content": "Stereo Occlusion Detection. Stereo co-planar cameras are classically formulated as the half-occlusion problem in stereo vision, i.e., the area occluded in one camera of a stereo vision system. There has been extensive research on detecting the half-occlusion region [2, 45, 58, 65]. Several works [6, 7, 32, 60, 72] jointly optimize occlusion and disparity. Others [17, 55, 61] apply left-right consistency check. The area receives different estimations from left and right views are excluded. RePLAy focuses on setups without stereo but where LiDAR provides rough 3D structures. Compared to the half-occlusion baseline which proposes a heuristic solution, RePLAy is an analytical solution." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.599, + 0.787, + 0.705 + ], + "angle": 0, + "content": "Downstream Tasks using LiDAR Depthmap. Multiple downstream tasks use LiDAR projected depthmaps. E.g., MonoDE methods [4, 5, 38, 68, 71] use depthmap as the GT supervision. Mono3DOD methods [36, 43, 63, 64, 66] use depthmap for supervising auxiliary task along with 3D object detection. When deployed to AV datasets other than KITTI, the aforementioned methods [42, 69] have diminished performance due to the projective artifacts. RePLAy improves their performance by resolving the projective artifacts in LiDAR depthmap." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.738, + 0.334, + 0.755 + ], + "angle": 0, + "content": "3 RePLAy" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.787, + 0.84 + ], + "angle": 0, + "content": "We now describe RePLAy in Algorithm 1. Sec. 3.1 attributes the projective artifacts to the epipolar occlusion induced by the binocular vision system between the virtual LiDAR camera and RGB camera. Sec. 3.2 outlines our analytical solution. Sec. 3.3 hold further discussions." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.673, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "RePLAY" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.144, + 0.457, + 0.272 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.23, + 0.274, + 0.457, + 0.286 + ], + "angle": 0, + "content": "(a) LiDAR Point Cloud Bird-Eye View" + }, + { + "type": "image", + "bbox": [ + 0.465, + 0.144, + 0.785, + 0.204 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.495, + 0.205, + 0.758, + 0.216 + ], + "angle": 0, + "content": "(b) Depthmap \\(\\mathbf{D}_l\\) of Virtual LiDAR Camera." + }, + { + "type": "image", + "bbox": [ + 0.465, + 0.218, + 0.785, + 0.273 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.523, + 0.274, + 0.729, + 0.286 + ], + "angle": 0, + "content": "(c) Depthmap \\(\\mathbf{D}_r\\) of RGB Camera." + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.294, + 0.785, + 0.391 + ], + "angle": 0, + "content": "Fig. 4: Virtual LiDAR camera and RGB Binocular System. In (a), we plot the bird-eye view of the point cloud \\(\\mathcal{V}\\) from LiDAR. A binocular system is set between virtual LiDAR (epipolar left) and RGB camera (epipolar right). Subfigure (b) plots the depthmap \\(\\mathbf{D}_l\\) from LiDAR view. As stated in Lemma 1, depthmap \\(\\mathbf{D}_l\\) does not have projective artifacts. Compare (b) and (c), the artifacts arise when transfer from LiDAR to RGB camera, due to the epipolar occlusion among the binocular system. We formally define the projective artifacts (or epipolar occlusion) in Eq. (7)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.411, + 0.784, + 0.456 + ], + "angle": 0, + "content": "We execute Sec. 3.1 auto-calibration over the entire dataset. We initialize LiDAR intrinsic \\(\\mathbf{K}_l\\) the same as the RGB camera \\(\\mathbf{K}_r\\) with slight augmentation to ensure a sufficient Field-of-View." + }, + { + "type": "title", + "bbox": [ + 0.214, + 0.48, + 0.767, + 0.496 + ], + "angle": 0, + "content": "3.1 LiDAR Depthmap Projective Artifacts as Epipolar Occlusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.506, + 0.785, + 0.566 + ], + "angle": 0, + "content": "Fig. 2 demonstrates the baseline distance between various LiDAR and RGB camera sensor sets. Fig. 2a shows that the background tree is solely visible to LiDAR. Intuitively, the background tree is occluded by the foreground pedestrian in a binocular vision system of the virtual LiDAR camera and RGB camera." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.567, + 0.785, + 0.613 + ], + "angle": 0, + "content": "Virtual LiDAR Camera. LiDAR produces a point cloud \\(\\mathcal{V}\\) by emitting laser pulses into 3D space. Mathematically, we assume the LiDAR point cloud \\(\\mathcal{V}\\) as a set of 3D rays originating from the coordinate origin:" + }, + { + "type": "equation", + "bbox": [ + 0.33, + 0.624, + 0.785, + 0.643 + ], + "angle": 0, + "content": "\\[\n\\mathcal {V} = \\left\\{\\mathbf {v} _ {i} \\right\\} = \\left\\{\\mathbf {0} + \\bar {\\mathbf {r}} _ {i} \\cdot t _ {i} \\mid i \\in \\mathbb {Z}, t > 0, \\bar {\\mathbf {r}} _ {i} \\in \\mathbb {R} ^ {3} \\right\\}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.653, + 0.784, + 0.699 + ], + "angle": 0, + "content": "where \\(\\mathbf{v}_i\\) is a 3D point. \\(\\overline{\\mathbf{r}}_i\\) is a normalized direction vector, and \\(t_i\\) is the euclidean distance between the 3D point and LiDAR. If we position a virtual LiDAR camera with arbitrary resolution and intrinsic \\(\\mathbf{K}_l\\) at the LiDAR's location, then:" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.724, + 0.784, + 0.754 + ], + "angle": 0, + "content": "Lemma 1. The depthmap of the virtual LiDAR camera is free of projective artifacts, i.e., no two scanned points \\(\\mathbf{v}_1\\) and \\(\\mathbf{v}_2\\) overlap after projection." + }, + { + "type": "equation", + "bbox": [ + 0.338, + 0.767, + 0.785, + 0.784 + ], + "angle": 0, + "content": "\\[\n\\forall \\mathbf {v} _ {1}, \\mathbf {v} _ {2} \\in \\mathcal {V}, \\| \\pi (\\mathbf {v} _ {1} \\mid \\mathbf {K} _ {l}) - \\pi (\\mathbf {v} _ {2} \\mid \\mathbf {K} _ {l}) \\| _ {2} \\neq 0, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.784, + 0.84 + ], + "angle": 0, + "content": "where \\(\\pi (\\cdot)\\) stands for the projection process. See proof in Supp. (Appendix A). We visualize the artifacts-free depthmap \\(\\mathbf{D}_l\\) from the virtual LiDAR camera in Fig. 4b." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.356, + 0.128 + ], + "angle": 0, + "content": "S. Zhu et al." + }, + { + "type": "image", + "bbox": [ + 0.225, + 0.144, + 0.795, + 0.245 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.25, + 0.788, + 0.334 + ], + "angle": 0, + "content": "Fig. 5: Auto-Calibration. The LiDAR point cloud \\(\\mathcal{V}\\) in (a) is mapped to spherical coordinate \\(\\theta -\\phi\\) with Eq. (3). We use a binary matrix \\(\\mathbf{S}\\) to record the unique pixel locations \\(S\\) after rasterization. Plots (b) and (c) are matrix \\(\\mathbf{S}\\) before and after autocalibration where duplication (marked by arrows) is minimized with Eq. (5) loss \\(L\\). Plots (d) and (e) contrast the virtual LiDAR depthmap \\(\\mathbf{D}_l\\) before and after autocalibration where Lemma 1 is fulfilled." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.353, + 0.784, + 0.442 + ], + "angle": 0, + "content": "Auto-Calibration. In practice, Lemma 1 may not hold in a realistic LiDAR point cloud due to possible bias in the point cloud coordinate origin \\(\\mathbf{0}\\). Fig. 5d, an example from the KITTI dataset shows persistent projective artifacts in the depthmap \\(\\mathbf{D}_l\\) when without any treatment. We perform auto-calibration to fulfill Lemma 1 by re-adjusting the point cloud origin. Fig. 4b and Fig. 5e show artifacts-free depthmap after auto-calibration." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.443, + 0.785, + 0.52 + ], + "angle": 0, + "content": "Auto-calibration aims to resolve the artifacts in the virtual LiDAR camera depthmap \\(\\mathbf{D}_l\\). We convert each LiDAR point cloud to the spherical coordinate system. This hypothesizes an omnidirectional image \\(\\mathbf{S}\\) of the LiDAR point cloud. Denote the origin to estimate as \\(\\mathbf{o}\\). Set the new 3D point \\(\\mathbf{u} = \\mathbf{v} + \\mathbf{o}\\), we have the polar angle \\(\\theta\\) and azimuthal angle \\(\\phi\\) after applying a rasterization operation \\([\\cdot]:\\)" + }, + { + "type": "equation", + "bbox": [ + 0.23, + 0.526, + 0.785, + 0.544 + ], + "angle": 0, + "content": "\\[\n\\theta = \\left[ \\arccos \\left(u _ {x} / \\left(u _ {x} ^ {2} + u _ {z} ^ {2}\\right)\\right) \\right], \\phi = \\left[ \\arccos \\left(u _ {y} / \\left(u _ {x} ^ {2} + u _ {y} ^ {2} + u _ {z} ^ {2}\\right)\\right) \\right], \\left[ \\theta \\phi \\right] = f (\\mathbf {v} + \\mathbf {o}). \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.552, + 0.784, + 0.583 + ], + "angle": 0, + "content": "Intuitively, function \\( f(\\cdot) \\) converts a 3D point \\( \\mathbf{u} \\in \\mathcal{V} \\) to a discrete 2D coordinate in \\( \\mathbf{S} \\). We denote the set of unique coordinates as \\( S \\):" + }, + { + "type": "equation", + "bbox": [ + 0.412, + 0.594, + 0.787, + 0.611 + ], + "angle": 0, + "content": "\\[\n\\mathcal {S} = \\left\\{f (\\mathbf {v} + \\mathbf {o}) \\mid \\mathbf {v} \\in \\mathcal {V} \\right\\}. \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.621, + 0.788, + 0.681 + ], + "angle": 0, + "content": "We implement Eq. (4) with a binary matrix \\(\\mathbf{S}\\). Its elements are 1 at each discrete coordinate and 0 otherwise. The size of \\(\\mathcal{S}\\) equals to summarizing the matrix \\(\\mathbf{S}\\). Artifacts become duplication in set \\(\\mathcal{S}\\). Removing artifacts is to minimize duplication. The loss \\(L\\) is then defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.442, + 0.692, + 0.787, + 0.71 + ], + "angle": 0, + "content": "\\[\nL = \\| \\mathcal {V} \\| - \\| \\mathcal {S} \\|. \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.785, + 0.749 + ], + "angle": 0, + "content": "The loss \\( L \\) is not differentiable. In each iteration, we approximate its gradient by ablating origin \\( \\mathbf{o} \\) with a small deviation and perform coordinate descent." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Virtual LiDAR Camera and RGB Binocular System. In Fig. 4, we set up a binocular vision system, denoting the virtual LiDAR camera and RGB camera as the epipolar left and right camera respectively. Correspondingly, set their intrinsic as \\(\\mathbf{K}_l\\) and \\(\\mathbf{K}_r\\), and projected depthmap as \\(\\mathbf{D}_l\\) and \\(\\mathbf{D}_r\\). Thus the problem of removing the projective artifacts is translated to detecting the epipolar occlusion between the virtual LiDAR camera and RGB camera." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.673, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "RePLAY" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.787, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.145, + 0.475, + 0.189 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.48, + 0.145, + 0.526, + 0.174 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.48, + 0.176, + 0.525, + 0.185 + ], + "angle": 0, + "content": "Rotation" + }, + { + "type": "image", + "bbox": [ + 0.53, + 0.146, + 0.685, + 0.188 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.2, + 0.788, + 0.242 + ], + "angle": 0, + "content": "Fig. 6: Epipolar Occlusion with Pure Rotation Movement. We apply a rotation transformation along the \\(x-z\\) axis in the camera coordinate system. As indicated by Lemma 2, a pure rotation movement does not lead to projective artifacts." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.257, + 0.784, + 0.348 + ], + "angle": 0, + "content": "Projective Artifacts are Epipolar Occlusion. We ensure that our virtual LiDAR has a larger field of view than the RGB camera by adjusting the resolution of the virtual LiDAR camera. Denote the relative movement between the LiDAR and RGB camera as \\(\\mathbf{P} = [\\mathbf{R}\\mathbf{t}]\\), where the auto-calibrated center o is integrated into t. Denote the projected pixels on epipolar left (LiDAR) and epipolar right (RGB) as p and q, and their pixel depths as \\(d\\) and \\(d^{\\prime}\\). We have:" + }, + { + "type": "equation", + "bbox": [ + 0.364, + 0.359, + 0.786, + 0.377 + ], + "angle": 0, + "content": "\\[\n\\mathbf {p} = \\pi (\\mathbf {v} \\mid \\mathbf {K} _ {l}), \\quad \\mathbf {q} = \\pi (\\mathbf {R v} + \\mathbf {t} \\mid \\mathbf {K} _ {r}). \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.386, + 0.698, + 0.402 + ], + "angle": 0, + "content": "We then formally define the epipolar occlusion \\(\\mathcal{O}\\) on RGB view as:" + }, + { + "type": "equation", + "bbox": [ + 0.31, + 0.413, + 0.786, + 0.431 + ], + "angle": 0, + "content": "\\[\n\\mathcal {O} = \\left\\{\\mathbf {q} _ {1} \\mid \\forall \\mathbf {q} _ {1}, \\mathbf {q} _ {2} \\in \\mathcal {Q}, d _ {\\mathbf {q} _ {1}} ^ {\\prime} > d _ {\\mathbf {q} _ {2}} ^ {\\prime}, \\| \\mathbf {q} _ {1} - \\mathbf {q} _ {2} \\| _ {2} = 0 \\right\\}. \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.44, + 0.784, + 0.486 + ], + "angle": 0, + "content": "Eq. (7) simply suggests if two pixels in RGB camera overlap with each other, then the one \\(\\mathbf{q}_1\\) with larger depth is occluded by the one with smaller depth \\(\\mathbf{q}_2\\). Occlusion detection finds all occluded pixels \\(\\mathbf{q}_1\\), i.e., the set \\(\\mathcal{O}\\)." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.507, + 0.686, + 0.523 + ], + "angle": 0, + "content": "3.2 Epipolar Occlusion under Rotation and Translation" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.532, + 0.784, + 0.579 + ], + "angle": 0, + "content": "Suppose the relative camera motion among the Fig. 4a binocular system is an arbitrary camera pose \\(\\mathbf{P} = [\\mathbf{R}\\mathbf{t}]\\). Set \\(\\mathbf{I}\\) as the identity matrix. RePLAy decomposes \\(\\mathbf{P}\\) into a sequential pure rotation and pure translation movement:" + }, + { + "type": "equation", + "bbox": [ + 0.364, + 0.591, + 0.786, + 0.61 + ], + "angle": 0, + "content": "\\[\n\\mathbf {P} _ {1} = \\left[ \\begin{array}{l l} \\mathbf {R} & \\mathbf {0} \\end{array} \\right], \\mathbf {P} _ {2} = \\left[ \\begin{array}{l l} \\mathbf {I} & \\mathbf {t} \\end{array} \\right], \\mathbf {P} = \\mathbf {P} _ {2} \\mathbf {P} _ {1}. \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.615, + 0.784, + 0.69 + ], + "angle": 0, + "content": "We outline pure rotation movement in Lemma 2 and Fig. 6. The pure translation movement is illustrated in Lemma 3 and Fig. 7. Without loss of generality, we assume the epipolar left and right cameras share the same intrinsic, i.e., \\(\\mathbf{K}_l = \\mathbf{K}_r\\). When with different intrinsics, we apply a resize and crop operation defined by \\(\\mathbf{A} = \\mathbf{K}_l\\mathbf{K}_r^{-1}\\) to unify the intrinsic." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.701, + 0.754, + 0.716 + ], + "angle": 0, + "content": "Lemma 2. Pure rotation movement does not cause epipolar occlusion." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.727, + 0.784, + 0.787 + ], + "angle": 0, + "content": "We know from epipolar geometry [25] that pure rotation induces homography between two views. In other words, there is a one-to-one mapping between pixels before and after rotation. See Supp for more information. Fig. 6 illustrates Lemma 2 with a visual example." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.797, + 0.786, + 0.827 + ], + "angle": 0, + "content": "Lemma 3. Pure translation movement causes epipolar occlusion along the direction of the epipolar line." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.271, + 0.115, + 0.356, + 0.128 + ], + "angle": 0, + "content": "S. Zhu et al." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.145, + 0.788, + 0.345 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.354, + 0.788, + 0.48 + ], + "angle": 0, + "content": "Fig. 7: Epipoplar Occlusion with Pure Translation Movement. A background pixel \\(\\mathbf{p}_1\\) is gradually occluded by a foreground pixel \\(\\mathbf{p}_2\\) under pure translation. We formulate three binocular systems at different baseline distances. For simplicity, we set the depth of \\(\\mathbf{p}_1\\) to infinity, making its 2D location static. Direction vector \\(\\mathbf{n}\\) is set to point from \\(\\mathbf{p}\\) to \\(\\mathbf{q}\\). The scalar \\(t\\) is marked by blue line segment. From Lemma 3, (a) The pixel \\(\\mathbf{p}_1\\) is only occluded by pixel \\(\\mathbf{p}_2\\) at the reverse direction of epipolar line (red line segment). (b) \\(\\mathbf{p}_1\\) is not occluded. (c) \\(\\mathbf{p}_1\\) is occluded following Eq. (7). (d) \\(\\mathbf{p}_1\\) is considered occluded following Eq. (17). Intuitively, the occluded pixel \\(\\mathbf{p}_1\\) (in a) is \"caught up\" (in c) and \"surpassed\" (in d) by \\(\\mathbf{p}_2\\) while moving along the epipolar line." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.497, + 0.788, + 0.558 + ], + "angle": 0, + "content": "Proof. As pure rotation does not cause epipolar occlusion, we apply the translation movement over the artifacts-free depthmap after the pure rotation. With slight abuse of the notation, we set the pixels after pure rotation movement as \\(\\mathbf{p}\\) and \\(\\mathbf{q}\\). The epipolar line \\(\\mathbf{l_p}\\) on the right image from the left image pixel \\(\\mathbf{p}\\) is:" + }, + { + "type": "equation", + "bbox": [ + 0.368, + 0.566, + 0.785, + 0.584 + ], + "angle": 0, + "content": "\\[\n\\mathbf {l} _ {\\mathbf {p}} = \\mathbf {K} ^ {- \\top} \\left[ \\mathbf {t} _ {\\times} \\right] \\mathbf {R} \\mathbf {K} ^ {- 1} \\mathbf {p} = \\mathbf {K} ^ {- \\top} \\left[ \\mathbf {t} _ {\\times} \\right] \\mathbf {K} ^ {- 1} \\mathbf {p}, \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.592, + 0.788, + 0.639 + ], + "angle": 0, + "content": "where \\([\\mathbf{t}_{\\times}]\\) is the cross product in matrix form [71]. Due to a pure translation movement, the rotation \\(\\mathbf{R}\\) is removed as being an identity matrix. Further, the left and right epipolar lines overlap with each other in pure translation:" + }, + { + "type": "equation", + "bbox": [ + 0.351, + 0.647, + 0.785, + 0.665 + ], + "angle": 0, + "content": "\\[\n\\mathbf {l} _ {\\mathbf {q}} = - \\mathbf {K} ^ {- \\top} \\left[ \\mathbf {t} _ {\\times} \\right] \\mathbf {K} ^ {- 1} \\mathbf {q}, \\mathbf {p} ^ {\\top} \\mathbf {l} _ {\\mathbf {q}} = 0, \\mathbf {p} ^ {\\top} \\mathbf {l} _ {\\mathbf {p}} = 0. \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.673, + 0.646, + 0.689 + ], + "angle": 0, + "content": "Also shown in Fig. 7, pixel \\(\\mathbf{q}\\) is hence a function of pixel \\(\\mathbf{p}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.245, + 0.697, + 0.785, + 0.723 + ], + "angle": 0, + "content": "\\[\n\\mathbf {q} = \\mathbf {p} + t \\cdot \\mathbf {n}, t = \\| \\mathbf {q} - \\mathbf {p} \\| _ {2} = \\mathbf {n} ^ {\\intercal} (\\mathbf {q} - \\mathbf {p}), \\mathbf {n} = (\\mathbf {q} - \\mathbf {p}) / \\sqrt {\\| \\mathbf {q} - \\mathbf {p} \\| _ {2} ^ {2}}, \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.73, + 0.788, + 0.791 + ], + "angle": 0, + "content": "where \\(t\\) is the moving distance, similar to the disparity in stereo vision. We define the vector \\(\\mathbf{n}\\) pointed from \\(\\mathbf{p}\\) to \\(\\mathbf{q}\\) as the direction of the epipolar line \\(\\mathbf{l}\\). \\(\\mathbf{p}\\) and \\(\\mathbf{q}\\) are related by the equation \\(d^{\\prime}\\mathbf{q} = d\\mathbf{K}_{r}\\mathbf{R}\\mathbf{K}_{l}^{-1}\\mathbf{p} + \\mathbf{K}_{r}\\mathbf{t}\\). Expanding under pure translation movement gives us:" + }, + { + "type": "equation", + "bbox": [ + 0.214, + 0.797, + 0.787, + 0.842 + ], + "angle": 0, + "content": "\\[\nt ^ {2} = \\left(p _ {x} - q _ {x}\\right) ^ {2} + \\left(p _ {y} - q _ {y}\\right) ^ {2} = \\frac {\\left(x + z \\cdot q _ {x}\\right) ^ {2} + \\left(y + z \\cdot q _ {y}\\right) ^ {2}}{\\left(d ^ {\\prime} - z\\right) ^ {2}} = \\frac {\\left(x + z \\cdot q _ {x}\\right) ^ {2} + \\left(y + z \\cdot q _ {y}\\right) ^ {2}}{d ^ {2}}. \\tag {12}\n\\]" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.673, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "RePLAY" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "image", + "bbox": [ + 0.227, + 0.144, + 0.502, + 0.195 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.264, + 0.197, + 0.458, + 0.209 + ], + "angle": 0, + "content": "(a) Virtual LiDAR Depthmap \\(\\mathbf{D}_l\\)" + }, + { + "type": "image", + "bbox": [ + 0.51, + 0.144, + 0.785, + 0.196 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.521, + 0.198, + 0.773, + 0.21 + ], + "angle": 0, + "content": "(b) Densified Virtual LiDAR Depthmap \\(\\overline{\\mathbf{D}}_l\\)" + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.218, + 0.785, + 0.26 + ], + "angle": 0, + "content": "Fig. 8: Per-pixel Occlusion Examination. To overcome LiDAR sparsity, for a query pixel, we verify inspection pixels along epipolar line \\(\\mathcal{L}_{\\mathbf{p}_1}\\) following Eq. (17) to determine occlusion. Depth \\(\\overline{\\mathbf{D}}_l\\) is densified with nearest-neighbor interpolation." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.278, + 0.784, + 0.37 + ], + "angle": 0, + "content": "Please see the derivation in Supp. (Appendix B). Note, we assume pixels \\(\\mathbf{p}\\) and \\(\\mathbf{q}\\) are respectively visible to the left and right camera. This suggests \\(d > 0\\) and \\(d = d' - z > 0\\). Eq. (12) suggests moving distance \\(t^2\\) monotonously decreases as the projected depth \\(d'\\) increases. It fits the intuition where a foreground object has a larger movement than a background object. From Eq. (7), when with occlusion, we have \\(\\mathbf{q}_1 = \\mathbf{q}_2\\), and \\(d_{\\mathbf{q}_1} < d_{\\mathbf{q}_2}\\). Combine Eq. (11) and Eq. (12):" + }, + { + "type": "equation", + "bbox": [ + 0.294, + 0.379, + 0.785, + 0.397 + ], + "angle": 0, + "content": "\\[\n\\mathbf {q} _ {1} = \\mathbf {q} _ {2}, \\text {a n d} d _ {\\mathbf {q} _ {1}} < d _ {\\mathbf {q} _ {2}} \\longrightarrow t _ {1} ^ {2} > t _ {2} ^ {2} \\longrightarrow \\mathbf {n} ^ {\\intercal} (\\mathbf {p} _ {1} - \\mathbf {p} _ {2}) < 0. \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.406, + 0.784, + 0.468 + ], + "angle": 0, + "content": "From Eq. (13), if pixel \\(\\mathbf{q}_1\\) is occluded by \\(\\mathbf{q}_2\\) in the right view, pixel \\(\\mathbf{q}_2\\) must move a longer distance along the epipolar line than \\(\\mathbf{q}_1\\). This suggests, in the left view, \\(\\mathbf{p}_2\\) resides at the \"back\" of \\(\\mathbf{p}_1\\) w.r.t. the moving direction. In other words, \\(\\mathbf{p}_1\\) is only occluded by other pixels along the reverse direction of epipolar line." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.476, + 0.784, + 0.522 + ], + "angle": 0, + "content": "Epipolar Occlusion Judgement. Combine Lemma 2 and 3, a pixel \\(\\mathbf{p}_1\\) in LiDAR view is occluded by other pixels \\(\\mathbf{p}_2\\) along the reverse of epipolar line direction. For pixel \\(\\mathbf{p}_1\\), the set of inspection pixels is:" + }, + { + "type": "equation", + "bbox": [ + 0.352, + 0.532, + 0.785, + 0.549 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathbf {p} _ {1}} = \\left\\{\\mathbf {p} _ {2} \\mid \\mathbf {p} _ {2} = \\mathbf {p} _ {1} - \\mathbf {n} \\cdot k \\cdot \\Delta t, k \\in \\mathbb {Z} \\right\\}. \\tag {14}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.559, + 0.731, + 0.575 + ], + "angle": 0, + "content": "\\(k\\) is an integer. At occlusion, we have \\(\\mathbf{q}_1 = \\mathbf{q}_2\\), combined with Eq. (11):" + }, + { + "type": "equation", + "bbox": [ + 0.254, + 0.585, + 0.785, + 0.601 + ], + "angle": 0, + "content": "\\[\n\\mathbf {q} _ {1} = \\mathbf {q} _ {2} \\longrightarrow \\mathbf {p} _ {1} + t _ {1} \\mathbf {n} = \\mathbf {p} _ {2} + t _ {2} \\mathbf {n} \\longrightarrow \\mathbf {n} ^ {\\top} \\left(\\mathbf {p} _ {1} - \\mathbf {p} _ {2}\\right) + t _ {1} - t _ {2} = 0. \\tag {15}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.611, + 0.784, + 0.672 + ], + "angle": 0, + "content": "We use Fig. 7 to explain Eq. (15). In Fig. 7a, the pixel \\(\\mathbf{p}_1\\) and \\(\\mathbf{p}_2\\) has an initial distance \\(\\mathbf{n}^{\\top}(\\mathbf{p}_1 - \\mathbf{p}_2)\\). In Fig. 7c, after translation movement, each pixel moves a distance \\(t_1\\) and \\(t_2\\) along the epipolar line. Occlusion happens when pixel \\(\\mathbf{p}_2\\) overlaps pixel \\(\\mathbf{p}_1\\). We summarize Eq. (15) into a function \\(g(\\cdot)\\) as:" + }, + { + "type": "equation", + "bbox": [ + 0.378, + 0.682, + 0.785, + 0.7 + ], + "angle": 0, + "content": "\\[\n\\lambda \\leq g \\left(\\mathbf {p} _ {1}, \\mathbf {p} _ {2}, d _ {\\mathbf {p} _ {1}}, d _ {\\mathbf {p} _ {2}} \\mid \\mathbf {K}, \\mathbf {P}\\right) \\leq 0, \\tag {16}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.709, + 0.784, + 0.814 + ], + "angle": 0, + "content": "where the distance \\( t_1 \\) and \\( t_2 \\) are computed using intrinsic, camera motion, and depth values. In Fig. 7b, when \\( g(\\cdot) > 0 \\), there is no occlusion. In Fig. 7c, when \\( g(\\cdot) = 0 \\), two pixels overlap and occlusion happens. However, this is hard due to LiDAR sparsity, rasterization, and other noise sources. Hence, we relax with a threshold \\( \\lambda \\). In practice, we set \\( \\lambda \\) to \\( -\\infty \\). This suggests the pixel \\( \\mathbf{p}_1 \\) is occluded if any pixel \\( \\mathbf{p}_2 \\) in the epipolar line \"surpasses\" it in the RGB view, as shown in Fig. 7d. Formally, we consider pixel \\( \\mathbf{p}_1 \\) occluded as:" + }, + { + "type": "equation", + "bbox": [ + 0.311, + 0.825, + 0.785, + 0.843 + ], + "angle": 0, + "content": "\\[\n\\mathcal {O} = \\left\\{\\mathbf {p} _ {1} \\mid \\exists \\mathbf {p} _ {2} \\in \\mathcal {L} _ {\\mathbf {p} _ {1}}, g \\left(\\mathbf {p} _ {1}, \\mathbf {p} _ {2}, d _ {\\mathbf {p} _ {1}}, d _ {\\mathbf {p} _ {2}} \\mid \\mathbf {K}, \\mathbf {P}\\right) < 0 \\right\\}. \\tag {17}\n\\]" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.356, + 0.128 + ], + "angle": 0, + "content": "S. Zhu et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.194 + ], + "angle": 0, + "content": "Eq. (17) is also a necessary condition of epipolar occlusion. Due to the Li-DAR sparsity, we apply a nearest-neighbor interpolation over virtual LiDAR depthmap \\(\\mathbf{D}_l\\) to acquire \\(d_{\\mathbf{p}_2}\\) whenever required, as shown in Fig. 8." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.213, + 0.357, + 0.228 + ], + "angle": 0, + "content": "3.3 Discussions" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.237, + 0.788, + 0.344 + ], + "angle": 0, + "content": "Half-Occlusion (Depth Buffering) Baseline. Sec. 2 shows that half-occlusion algorithms solve projective artifacts in stereo vision, i.e., stereo/epipolar occlusion, without knowing LiDAR point cloud. Despite a different setting, they provide a baseline solution. When preparing RGB depthmap \\(\\mathbf{D}_r\\), only pixels with smaller depth are preserved if multiple 3D points \\(\\mathbf{v}\\) have duplicate 2D locations after rasterization. The raw LiDAR depthmap uses half-occlusion. So, we refer Half-Occlusion as \"Raw\" following [23]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.344, + 0.788, + 0.403 + ], + "angle": 0, + "content": "Does RePLAy simply densify the sparse depthmap to remove occlusion? No. Densification alone is insufficient to remove occlusion. RePLAy enables robust densification by formulating a stereo system with an auto-calibration procedure to remove occlusion." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.404, + 0.788, + 0.571 + ], + "angle": 0, + "content": "Universal Existence of Projective Artifacts. Fig. 2 pictorially depicts that projective artifacts exist under all LiDAR and camera sensor settings, while Fig. 7 shows that these artifacts scale up with a larger baseline distance between sensors. We now show that this indeed is the case. The artifact width distribution plot on the outdoor KITTI and the indoor NYUv2 [53] datasets are shown in Fig. 10 in the Supp. The LiDAR and RGB camera in KITTI have a baseline distance of \\(0.4m\\) [22]. We set the baseline distance to \\(0.1m\\) following the AR headset configuration of Fig. 2. Despite the indoor image having a lesser resolution than outdoor (\\(480 \\times 640\\) compared to \\(375 \\times 1242\\)), we observe a similar distribution. Detailed definitions and evaluation protocols are in Supp. The analysis confirms that projective artifacts exist in both indoor and outdoor datasets." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.592, + 0.377, + 0.61 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.623, + 0.788, + 0.668 + ], + "angle": 0, + "content": "We benchmark RePLAy on depthmap quality, MonoDE and Mono3DOD tasks. Sec. 4.1 assesses depthmap quality, while Sec. 4.2 and Sec. 4.3. report MonoDE and Mono3DOD results respectively." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.669, + 0.788, + 0.697 + ], + "angle": 0, + "content": "Datasets. We experiment on KITTI [21], nuScenes [10] and Waymo [56] datasets. KITTI provides stereo ground truth while the others do not." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.698, + 0.675, + 0.714 + ], + "angle": 0, + "content": "Data Splits. We use the following splits of the above datasets:" + }, + { + "type": "text", + "bbox": [ + 0.22, + 0.719, + 0.784, + 0.733 + ], + "angle": 0, + "content": "- KITTI Eigen [15]: 23,158 training and 652 validation samples of raw KITTI." + }, + { + "type": "text", + "bbox": [ + 0.22, + 0.735, + 0.784, + 0.762 + ], + "angle": 0, + "content": "- KITTI Stereo [46]: 200 training and 200 testing samples out of which 145 training scenes are mapped to the raw KITTI dataset." + }, + { + "type": "text", + "bbox": [ + 0.22, + 0.765, + 0.784, + 0.794 + ], + "angle": 0, + "content": "- nuScenes Frontal Val: 28,130 training and 2,436 random validation samples (from 6,019 samples) from front camera and top LiDAR." + }, + { + "type": "text", + "bbox": [ + 0.22, + 0.796, + 0.784, + 0.839 + ], + "angle": 0, + "content": "- Waymo Frontal Val: 52,386 training and 40,839 validation samples from the front camera and top LiDAR. We construct its training set by sampling every third frame from the training sequences as in [34]." + }, + { + "type": "list", + "bbox": [ + 0.22, + 0.719, + 0.784, + 0.839 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.673, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "RePLAY" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.788, + 0.188 + ], + "angle": 0, + "content": "Table 2: Depthmap Quality Assessment with baselines and RePLAy. We adopt KITTI semi-dense depthmap and KITTI Stereo disparity map as the reference GT. [Key: Best, Second Best, Mod. = Modified]" + }, + { + "type": "table", + "bbox": [ + 0.221, + 0.199, + 0.784, + 0.344 + ], + "angle": 0, + "content": "
GTEval RegionDepthmapδ0.5↑δ1↑δ2↑RMS ↓RMSlog ↓ARel ↓SRel ↓log10 ↓SIlog ↓
Semi-DenseAllRaw0.9730.9830.9901.9350.0870.0290.3320.0108.660
Mod. Half-Occ0.9830.9910.9961.3470.0570.0190.1130.0075.712
RePLAy0.9850.9940.9981.2410.0460.0160.0630.0074.592
StereoAllRaw0.9500.9620.9704.1980.1690.0831.6040.02416.691
Mod. Half-Occ0.9710.9820.9882.5410.1050.0470.5460.01610.340
RePLAy0.9750.9860.9912.2240.0860.0340.3540.0148.362
ForegroundRaw0.8590.8710.8928.7260.3170.2285.9740.05829.772
Mod. Half-Occ0.9190.9310.9495.2850.1990.1542.3660.03319.250
RePLAy0.9250.9370.9564.8370.1790.1001.9290.03017.319
BackgroundRaw0.9670.9800.9862.7080.1130.0520.7740.01711.037
Mod. Half-Occ0.9800.9910.9951.7010.0700.0320.2100.0136.638
RePLAy0.9820.9930.9981.4390.0520.0270.0850.0124.722
" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.355, + 0.788, + 0.398 + ], + "angle": 0, + "content": "Table 3: Depthmap Quality Assessment on Removed Artifacts. RePLAy outperforms the Mod. half-occlusion baseline. For removed artifacts, a worse performance suggests a better algorithm. [Key: Best]" + }, + { + "type": "table", + "bbox": [ + 0.221, + 0.41, + 0.784, + 0.448 + ], + "angle": 0, + "content": "
GTEval RegionDepthmapδ0.5↓δ1↓δ2↓RMS↑RMSlog↑ARel↑SRel↑log10↑
Semi-DenseAllMod. Half-Occ0.8140.8460.8935.8500.2750.1963.7620.053
RePLAy0.2740.3630.54812.0840.5870.77815.8510.202
" + }, + { + "type": "text", + "bbox": [ + 0.22, + 0.457, + 0.785, + 0.486 + ], + "angle": 0, + "content": "- KITTI Detection Val: 3,748 training and 3,762 validation images from the left camera for the Mono3DOD task." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.492, + 0.782, + 0.509 + ], + "angle": 0, + "content": "Ground Truth (GT) Depthmaps. We use the following as GT depthmaps:" + }, + { + "type": "text", + "bbox": [ + 0.22, + 0.511, + 0.731, + 0.526 + ], + "angle": 0, + "content": "- KITTI Eigen: We use the semi-dense depthmap available in KITTI." + }, + { + "type": "text", + "bbox": [ + 0.22, + 0.526, + 0.785, + 0.556 + ], + "angle": 0, + "content": "- KITTI Stereo: We generate depthmaps from GT disparity maps, which are denser than semi-dense depthmaps." + }, + { + "type": "text", + "bbox": [ + 0.221, + 0.557, + 0.787, + 0.632 + ], + "angle": 0, + "content": "- nuScenes and Waymo Frontal Val: We generate GT depthmaps by projecting only those LiDAR points present inside GT 3D bounding boxes. To prepare a reference depthmap, we follow [49] to filter LiDAR point cloud with GT 3D bounding boxes labels. Correspondingly, the evaluation is confined to the bounding box area, as only depth labels in that region are retained." + }, + { + "type": "list", + "bbox": [ + 0.22, + 0.511, + 0.787, + 0.632 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.652, + 0.517, + 0.667 + ], + "angle": 0, + "content": "4.1 Depthmap Quality Assessment" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.788, + 0.749 + ], + "angle": 0, + "content": "Baselines. We compare RePLAy to raw depthmap, which is the half-occlusion (depth buffering) method and a modified half-occlusion method. The modified half-occlusion modifies the half-occlusion method (Sec. 3.3) by back-projecting the estimated dense virtual depthmap \\((\\overline{\\mathbf{D}}_l)\\) onto the RGB camera and only preserving pixels in \\(\\mathbf{D}_r\\) that are less than their back-projected values." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Results. Tab. 2 compares the RePLAy depthmap with these baselines processed depthmaps. Tab. 2 verifies that RePLAy depthmap outperforms the raw and modified half-occlusion baselines. RePLAy depthmap agrees more with semidense or disparity depthmap after removing the projective artifacts. Fig. 3 shows qualitative comparisons. We believe projective artifacts persist in the half-occlusion method because of the LiDAR sparsity, leading to inferior depthmaps." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.356, + 0.128 + ], + "angle": 0, + "content": "S. Zhu et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.785, + 0.174 + ], + "angle": 0, + "content": "Table 4: Auto-Calibration Ablation on Depthmap Quality Assessment with semidense GT depthmap. [Key: Best]" + }, + { + "type": "table", + "bbox": [ + 0.25, + 0.185, + 0.754, + 0.226 + ], + "angle": 0, + "content": "
DepthmapAutoCalibδ0.5↑δ1↑δ2↑RMS ↓RMSlog ↓ARel ↓SRel ↓log10 ↓SIlog ↓
RePLAy×0.9810.9900.9951.4730.0570.0190.1110.0075.680
✓0.9850.9940.9981.2410.0460.0160.0630.0074.592
" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.238, + 0.785, + 0.281 + ], + "angle": 0, + "content": "Table 5: MonoDE Results on KITTI Stereo. ZoeDepth [5] trained with RePLAy depthmaps outperforms raw and semi-dense [57] depthmaps on foreground. [Key: Best, Second Best]." + }, + { + "type": "table", + "bbox": [ + 0.25, + 0.293, + 0.754, + 0.411 + ], + "angle": 0, + "content": "
Eval RegionDepthmapδ1↑δ2↑δ3↑RMS ↓RMSlog ↓ARel ↓SRel ↓log10 ↓SIlog ↓
AllRaw0.9550.9820.9922.6470.1160.0670.4200.02611.324
Semi-Dense0.9770.9940.9972.2740.0820.0430.2010.0187.971
RePLAy0.9740.9920.9962.2880.0890.0490.2480.0208.588
ForegroundRaw0.8860.9470.9803.6920.1800.1271.1120.02216.518
Semi-Dense0.9760.9900.9942.0930.0910.0560.3440.0228.342
RePLAy0.9450.9780.9892.9390.1280.0810.6420.02911.798
BackgroundRaw0.9730.9930.9982.2940.0840.0480.1880.0207.981
Semi-Dense0.9780.9960.9982.2830.0760.0400.1670.0177.239
RePLAy0.9810.9970.9992.0840.0700.0390.1380.0176.681
" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.44, + 0.785, + 0.501 + ], + "angle": 0, + "content": "Foreground Improvement. Tab. 2 rows 5 and 6 shows a significant improvement in foreground depthmap quality. On metric \\(\\delta_{0.5}\\), we reduce the errors by \\(46.8\\% = 1 - \\frac{1 - 0.925}{1 - 0.859}\\). This aligns with our observation that projective artifacts mostly occur over foreground objects, confirming Fig. 1." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.501, + 0.785, + 0.607 + ], + "angle": 0, + "content": "Comparison on Removed Artifacts. We next compare RePLAy on removed artifacts in Tab. 3. Tab. 3 shows that that in-addition to removing some artifacts, the modified half-occlusion also removes some non-artifacts (good points). Ideally, the removed points should be artifacts; therefore, the best algorithm should perform the worst on the quality assessment of removed points and hence the reversal in the directional indicator for Tab. 3. Qualitative comparison between raw, modified half-occlusion and RePLAy is provided in Supp. (Fig. 11)." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.607, + 0.796, + 0.669 + ], + "angle": 0, + "content": "Auto-Calibration Ablation. We next conduct ablations of the Auto-Calibration module (Sec. 3.1) in Tab. 4. Tab. 4 confirms the effectiveness of Auto-Calibration module in RePLAy, since Auto-Calibration provides \\(15\\%\\) and \\(19\\%\\) relative improvement in ARel and \\(\\mathrm{SI}_{\\log}\\) metrics respectively." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.692, + 0.588, + 0.708 + ], + "angle": 0, + "content": "4.2 MonoDE: Monocular Depth Estimation" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.719, + 0.785, + 0.749 + ], + "angle": 0, + "content": "Baselines. We compare RePLAy to raw and Semi-Dense methods. We use ZoeDepth [5] as the depth model due to its superior generalization." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.75, + 0.786, + 0.84 + ], + "angle": 0, + "content": "KITTI Stereo and Eigen Results. We train ZoeDepth [5] from scratch using depthmap with and without applying RePLAy on KITTI Eigen split. We evaluate on KITTI Stereo and Eigen split in Tabs. 5 and 6 respectively. Tab. 5 shows that ZoeDepth trained with RePLAy depthmap significantly outperforms the baselines. We observe despite semi-dense GT [57] using additional stereo cameras, RePLAy outperforms Semi-Dense especially in the background. This" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.673, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "RePLAY" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.144, + 0.406, + 0.208 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.269, + 0.209, + 0.358, + 0.221 + ], + "angle": 0, + "content": "(a) RGB Input" + }, + { + "type": "image", + "bbox": [ + 0.408, + 0.144, + 0.596, + 0.208 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.443, + 0.209, + 0.562, + 0.221 + ], + "angle": 0, + "content": "(b) Raw depthmaps" + }, + { + "type": "image", + "bbox": [ + 0.597, + 0.144, + 0.785, + 0.208 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.623, + 0.209, + 0.761, + 0.221 + ], + "angle": 0, + "content": "(c) RePLAy depthmaps" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.232, + 0.784, + 0.274 + ], + "angle": 0, + "content": "Fig. 9: MonoDE Qualitative Results on KITTI Eigen dataset. ZoeDepth [5] trained with RePLAy depthmaps (c) outperforms the one with raw depthmaps (b), especially on foreground." + }, + { + "type": "table_caption", + "bbox": [ + 0.215, + 0.292, + 0.784, + 0.321 + ], + "angle": 0, + "content": "Table 6: MonoDE Results. ZoeDepth trained with RePLAy outperforms the baselines in foreground regions of all datasets. [Key: Best]" + }, + { + "type": "table", + "bbox": [ + 0.221, + 0.333, + 0.782, + 0.415 + ], + "angle": 0, + "content": "
Data SplitEval RegionDepthmapδ1↑δ2↑δ3↑RMS↓RMSlog↓ARel↓SRel↓log10↓SIlog↓
KITTI EigenAllRaw0.8690.9590.9872.9110.1630.1240.6060.04613.501
RePLAy0.9270.9850.9972.1850.1100.0820.2870.0338.674
nuScenes ValForegroundRaw0.9710.9870.9931.3470.0750.0490.2450.0186.339
RePLAy0.9810.9930.9961.1500.0590.0390.1370.0164.905
Waymo ValForegroundRaw0.9410.9880.9963.4680.1110.0790.4870.0329.116
RePLAy0.9620.9930.9983.0570.0910.0630.3670.0267.322
" + }, + { + "type": "table_caption", + "bbox": [ + 0.215, + 0.431, + 0.784, + 0.473 + ], + "angle": 0, + "content": "Table 7: MonoDE Analysis to Environmental Conditions on nuScenes Val split. ZoeDepth trained with RePLAY depthmaps shows consistent improvements on foreground under all environmental conditions. [Key: Best]" + }, + { + "type": "table", + "bbox": [ + 0.221, + 0.486, + 0.782, + 0.59 + ], + "angle": 0, + "content": "
ConditionSamplesDepthmapδ1↑δ2↑δ3↑RMS ↓RMSlog ↓ARel ↓SRel ↓log10 ↓SIlog ↓
Rain864Raw0.9840.9910.9951.3380.0580.0400.1630.0164.859
RePLAy0.9890.9930.9951.2220.0500.0340.1270.0154.110
Night533Raw0.9650.9820.9901.3470.0890.0590.3350.0217.481
RePLAy0.9780.9910.9951.0630.0670.0450.1720.0175.374
Sunny1039Raw0.9640.9870.9931.3520.0810.0530.2670.0196.985
RePLAy0.9760.9930.9961.1340.0620.0400.1280.0165.324
All Tab. 62436Raw0.9710.9870.9931.3470.0750.0490.2450.0186.339
RePLAy0.9810.9930.9961.1500.0590.0390.1370.0164.905
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.609, + 0.784, + 0.654 + ], + "angle": 0, + "content": "may be due to [57] aggressively removing the background pixels around the foreground object boundaries. Fig. 9 shows this result visually. ZoeDepth trained with the raw depthmap learns strip-shaped artifacts (green boxes)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.656, + 0.784, + 0.716 + ], + "angle": 0, + "content": "Foreground Improvement. We additionally ablate depth performance on foreground, background, and entire image using the GT segmentation mask in Tab. 5. Tab. 5 shows that ZoeDepth trained with RePLAy depthmap outperforms in the foreground area, which agrees with the conclusion of Sec. 4.1." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.718, + 0.784, + 0.763 + ], + "angle": 0, + "content": "nuScenes and Waymo Frontal Val Results. We benchmark MonoDE on two datasets without stereo sensors: nuScenes [10] and Waymo [56] in Tab. 6. Tab. 6 confirms that RePLAy achieves significant improvements on these datasets." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.784, + 0.84 + ], + "angle": 0, + "content": "Effect of Environmental Conditions. We next analyze the effect of environmental conditions on the nuScenes Val split in Tab. 7. We categorize the nuScenes [10] val set into rainy, night, and sunny following their description tag (nuScenes-devkit GitHub #124). Tab. 7 shows that RePLAy is robust across multiple conditions." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.356, + 0.128 + ], + "angle": 0, + "content": "S. Zhu et al." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.147, + 0.64, + 0.162 + ], + "angle": 0, + "content": "4.3 Mono3DOD: Monocular 3D Object Detection" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.172, + 0.684, + 0.186 + ], + "angle": 0, + "content": "We finally verify the impact of RePLAy on the Mono3DOD task." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.187, + 0.786, + 0.277 + ], + "angle": 0, + "content": "Baseline. We compare RePLAy to raw method. We take SoTA 3D detectors, CaDDN [51] and MonoDTR [29], to showcase the impact of cleaning projected LiDAR depthmap with RePLAy, since these detectors use depthmap as an auxiliary training task. The two methods cover both convolutional and transformer architectures, making our benchmarking comprehensive. However, these two SoTA detectors use depth supervision in training differently." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.278, + 0.459, + 0.398 + ], + "angle": 0, + "content": "CaDDN. CaDDN uses the densified raw depthmaps of [33]. For CaDDN experiments, we apply RePLAy to the projected depthmaps and then densify those maps using the same depth completion algorithm [33] and parameters of [51]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.399, + 0.459, + 0.503 + ], + "angle": 0, + "content": "MonoDTR. MonoDTR [29] precomputes their GT depthmaps by \\(4\\times\\) downsampling raw projected depthmaps. For MonoDTR experiments, we clean the projected depthmaps and then \\(4\\times\\) down-sampling." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.505, + 0.461, + 0.61 + ], + "angle": 0, + "content": "Results. Tab. 8 shows results of retraining both detectors with our depthmaps on KITTI Val split. Cleaned depthmaps improves the performance of CaDDN and MonoDTR on most of the metrics. The biggest gains appear in the Easy" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.611, + 0.787, + 0.656 + ], + "angle": 0, + "content": "category. This is expected as Easy cars are closer to the ego camera and receive the maximum number of LiDAR points. Hence, improving the projected depthmap by RePLAy benefits the Easy category more." + }, + { + "type": "table_caption", + "bbox": [ + 0.47, + 0.304, + 0.788, + 0.374 + ], + "angle": 0, + "content": "Table 8: KITTI Val Mono3DOD Results on Car category. Using RePLAy depthmaps in training boosts the Mono3DOD performance. [Key: Best, Second Best, \\(* =\\) Reported, \\(\\dagger=\\) Retrained, Mod= Moderate.]" + }, + { + "type": "table", + "bbox": [ + 0.48, + 0.375, + 0.789, + 0.571 + ], + "angle": 0, + "content": "
MethodDepthmapIoU3D > 0.7
AP 3D|R40[%]↑
EasyModHard
M3D-RPN [8]-14.5311.078.65
MonoPair [12]-16.2812.3010.42
MonoDLE [44]-17.4513.6611.68
Kinematic [9]-19.7614.1010.47
GrooMeD [35]-19.6714.3211.27
MonoRUn [11]-20.0214.6512.61
D4LCN [13]Raw*22.3216.2012.30
DFR-Net [73]Raw*24.8117.7814.41
CaDDN [51]Raw†21.6015.6713.24
RePLAy22.8815.7013.18
MonoDTR [29]Raw†24.5018.6615.69
RePLAy26.1518.7715.69
" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.679, + 0.36, + 0.695 + ], + "angle": 0, + "content": "5 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.71, + 0.787, + 0.801 + ], + "angle": 0, + "content": "We propose an analytical solution to remove projective artifacts in LiDAR depthmap. Our parameter-free method, RePLAy, applies to diverse LiDAR and RGB camera sensor settings. RePLAy only requires the relative position between LiDAR and RGB camera. We verify its benefits with unanimous improvements on SoTA MonoDE and Mono3DOD methods. We release processed depthmaps of five major AV datasets to benefit the community." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.801, + 0.787, + 0.833 + ], + "angle": 0, + "content": "Limitation. We use nearest-neighbor interpolation to densify virtual LiDAR depthmaps. Advanced strategies, e.g., depth completion, may improve accuracy." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.673, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "RePLAY" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.324, + 0.161 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.177, + 0.785, + 0.204 + ], + "angle": 0, + "content": "1. The KITTI Vision Benchmark Suite. http://www.cvlabs.net/datasets/kitti/eval_object.php?obj_benchmark=3d (2022), accessed: 2022-07-03 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.205, + 0.785, + 0.232 + ], + "angle": 0, + "content": "2. Arbelaez, P., Maire, M., Fowlkes, C., Malik, J.: Contour detection and hierarchical image segmentation. TPAMI (2010) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.233, + 0.785, + 0.272 + ], + "angle": 0, + "content": "3. Bai, X., Hu, Z., Zhu, X., Huang, Q., Chen, Y., Fu, H., Tai, C.L.: Transfusion: Robust lidar-camera fusion for 3D object detection with transformers. In: CVPR (2022) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.273, + 0.785, + 0.3 + ], + "angle": 0, + "content": "4. Bhat, S.F., Alhashim, I., Wonka, P.: Adabins: Depth estimation using adaptive bins. In: CVPR (2021) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.301, + 0.785, + 0.342 + ], + "angle": 0, + "content": "5. Bhat, S.F., Birkl, R., Wofk, D., Wonka, P., Müller, M.: ZoeDepth: Zero-shot transfer by combining relative and metric depth. arXiv preprint arXiv:2302.12288 (2023) 2, 3, 4, 12, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.343, + 0.785, + 0.37 + ], + "angle": 0, + "content": "6. Birchfield, S., Tomasi, C.: Depth discontinuities by pixel-to-pixel stereo. IJCV (1999) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.371, + 0.674, + 0.384 + ], + "angle": 0, + "content": "7. Bobick, A.F., Intille, S.S.: Large occlusion stereo. IJCV (1999) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.385, + 0.785, + 0.411 + ], + "angle": 0, + "content": "8. Brazil, G., Liu, X.: M3D-RPN: Monocular 3D region proposal network for object detection. In: ICCV (2019) 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.412, + 0.785, + 0.439 + ], + "angle": 0, + "content": "9. Brazil, G., Pons-Moll, G., Liu, X., Schiele, B.: Kinematic 3D object detection in monocular video. In: ECCV (2020) 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.44, + 0.785, + 0.48 + ], + "angle": 0, + "content": "0. Caesar, H., Bankiti, V., Lang, A., Vora, S., Liong, V., Xu, Q., Krishnan, A., Pan, Y., Baldan, G., Beijbom, O.: nuScenes: A multimodal dataset for autonomous driving. In: CVPR (2020) 1, 2, 3, 10, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.481, + 0.785, + 0.52 + ], + "angle": 0, + "content": "1. Chen, H., Huang, Y., Tian, W., Gao, Z., Xiong, L.: MonoRun: Monocular 3D object detection by reconstruction and uncertainty propagation. In: CVPR (2021) 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.522, + 0.785, + 0.55 + ], + "angle": 0, + "content": "2. Chen, Y., Tai, L., Sun, K., Li, M.: MonoPair: Monocular 3D object detection using pairwise spatial relationships. In: CVPR (2020) 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.55, + 0.785, + 0.578 + ], + "angle": 0, + "content": "3. Ding, M., Huo, Y., Yi, H., Wang, Z., Shi, J., Lu, Z., Luo, P.: Learning depth-guided convolutions for monocular 3D object detection. In: CVPR Workshops (2020) 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.579, + 0.785, + 0.632 + ], + "angle": 0, + "content": "4. Ebadi, K., Bernreiter, L., Biggie, H., Catt, G., Chang, Y., Chatterjee, A., Denniston, C.E., Deschênes, S.P., Harlow, K., Khattak, S., et al.: Present and future of slam in extreme underground environments. arXiv preprint arXiv:2208.01787 (2022) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.633, + 0.785, + 0.66 + ], + "angle": 0, + "content": "5. Eigen, D., Puhrsch, C., Fergus, R.: Depth map prediction from a single image using a multi-scale deep network. NeurIPS (2014) 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.661, + 0.785, + 0.688 + ], + "angle": 0, + "content": "6. Filgueira, A., González-Jorge, H., Lagüela, S., Díaz-Vilarino, L., Arias, P.: Quantifying the influence of rain in lidar performance. Measurement (2017) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.689, + 0.785, + 0.716 + ], + "angle": 0, + "content": "7. Fua, P.: A parallel stereo algorithm that produces dense depth maps and preserves image features. Machine vision and applications (1993) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.717, + 0.729, + 0.73 + ], + "angle": 0, + "content": "8. Garg, K., Nayar, S.K.: When does a camera see rain? In: ICCV (2005) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.731, + 0.609, + 0.744 + ], + "angle": 0, + "content": "9. Garg, K., Nayar, S.K.: Vision and rain. IJCV (2007) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.745, + 0.785, + 0.784 + ], + "angle": 0, + "content": "20. Garvelmann, J., Pohl, S., Weiler, M.: From observation to the quantification of snow processes with a time-lapse camera network. Hydrology and Earth Systems Sciences (2013) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.785, + 0.785, + 0.812 + ], + "angle": 0, + "content": "21. Geiger, A., Lenz, P., Stiller, C., Urtasun, R.: Vision meets robotics: The KITTI dataset. IJRR (2013) 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.813, + 0.785, + 0.84 + ], + "angle": 0, + "content": "22. Geiger, A., Lenz, P., Urtasun, R.: Are we ready for autonomous driving? the KITTI vision benchmark suite. In: CVPR (2012) 10" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.177, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.356, + 0.127 + ], + "angle": 0, + "content": "S. Zhu et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.175 + ], + "angle": 0, + "content": "23. Godard, C., Mac Aodha, O., Firman, M., Brostow, G.J.: Digging into self-supervised monocular depth estimation. In: ICCV (2019) 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.176, + 0.785, + 0.203 + ], + "angle": 0, + "content": "24. Guizilini, V., Ambrus, R., Pillai, S., Raventos, A., Gaidon, A.: 3D packing for self-supervised monocular depth estimation. In: CVPR (2020) 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.204, + 0.785, + 0.231 + ], + "angle": 0, + "content": "25. Hartley, R., Zisserman, A.: Multiple view geometry in computer vision. Cambridge university press (2003) 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.232, + 0.785, + 0.271 + ], + "angle": 0, + "content": "26. Herskovitz, J., Wu, J., White, S., Pavel, A., Reyes, G., Guo, A., Bigham, J.P.: Making mobile augmented reality applications accessible. In: SIGACCESS (2020) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.273, + 0.785, + 0.313 + ], + "angle": 0, + "content": "27. Hillmann, C., Hillmann, C.: Comparing the gear vr, oculus go, and oculus quest. Unreal for Mobile and Standalone VR: Create Professional VR Apps Without Coding (2019) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.315, + 0.785, + 0.341 + ], + "angle": 0, + "content": "28. Hirschmuller, H.: Stereo processing by semiglobal matching and mutual information. TPAMI (2007) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.343, + 0.785, + 0.369 + ], + "angle": 0, + "content": "29. Huang, K.C., Wu, T.H., Su, H.T., Hsu, W.: MonoDTR: Monocular 3D object detection with depth-aware transformer. In: CVPR (2022) 3, 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.37, + 0.785, + 0.397 + ], + "angle": 0, + "content": "30. Huang, X., Cheng, X., Geng, Q., Cao, B., Zhou, D., Wang, P., Lin, Y., Yang, R.: The apolloscape dataset for autonomous driving. In: CVPR (2018) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.398, + 0.785, + 0.452 + ], + "angle": 0, + "content": "31. Hutabarat, D., Rivai, M., Purwanto, D., Hutomo, H.: Lidar-based obstacle avoidance for the autonomous mobile robot. In: 2019 12th International Conference on Information & Communication Technology and System (ICTS). pp. 197-202. IEEE (2019) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.453, + 0.785, + 0.479 + ], + "angle": 0, + "content": "32. Ishikawa, H., Geiger, D.: Occlusions, discontinuities, and epipolar lines in stereo. In: ECCV (1998) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.481, + 0.785, + 0.508 + ], + "angle": 0, + "content": "33. Ku, J., Harakeh, A., Waslander, S.: In defense of classical image processing: Fast depth completion on the CPU. In: CRV (2018) 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.509, + 0.785, + 0.536 + ], + "angle": 0, + "content": "34. Kumar, A., Brazil, G., Corona, E., Parchami, A., Liu, X.: DEVIANT: Depth Equivariant Network for Monocular 3D Object Detection. In: ECCV (2022) 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.537, + 0.785, + 0.563 + ], + "angle": 0, + "content": "35. Kumar, A., Brazil, G., Liu, X.: GrooMeD-NMS: Grouped mathematically differentiable nms for monocular 3D object detection. In: CVPR (2021) 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.565, + 0.785, + 0.604 + ], + "angle": 0, + "content": "36. Kumar, A., Guo, Y., Huang, X., Ren, L., Liu, X.: SeaBird: Segmentation in Bird's View with Dice Loss Improves Monocular 3D Detection of Large Objects. In: CVPR (2024) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.606, + 0.785, + 0.632 + ], + "angle": 0, + "content": "37. Kutila, M., Pyykonen, P., Holzhüter, H., Colomb, M., Duthon, P.: Automotive lidar performance verification in fog and rain. In: ITSC (2018) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.634, + 0.785, + 0.673 + ], + "angle": 0, + "content": "38. Lee, J.H., Han, M.K., Ko, D.W., Suh, I.H.: From big to small: Multi-scale local planar guidance for monocular depth estimation. arXiv preprint arXiv:1907.10326 (2019) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.675, + 0.785, + 0.715 + ], + "angle": 0, + "content": "39. Li, Y., Yu, A.W., Meng, T., Caine, B., Ngiam, J., Peng, D., Shen, J., Lu, Y., Zhou, D., Le, Q.V., et al.: Deepfusion: Lidar-camera deep fusion for multi-modal 3d object detection. In: CVPR (2022) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.717, + 0.785, + 0.757 + ], + "angle": 0, + "content": "40. Li, Y., Ibanez-Guzman, J.: LiDAR for autonomous driving: The principles, challenges, and trends for automotive lidar and perception systems. IEEE Signal Processing Magazine (2020) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.758, + 0.785, + 0.784 + ], + "angle": 0, + "content": "41. Liao, Y., Xie, J., Geiger, A.: KITTI-360: A novel dataset and benchmarks for urban scene understanding in 2D and 3D. TPAMI (2022) 1, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.786, + 0.785, + 0.812 + ], + "angle": 0, + "content": "42. Lin, J.T., Dai, D., Van Gool, L.: Depth estimation from monocular images and sparse radar data. In: IROS (2020) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.814, + 0.785, + 0.84 + ], + "angle": 0, + "content": "43. Liu, F., Liu, X.: Voxel-based 3d detection and reconstruction of multiple objects from a single image. In: NeurIPS (2021) 4" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.673, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "RePLAY" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.175 + ], + "angle": 0, + "content": "44. Ma, X., Zhang, Y., Xu, D., Zhou, D., Yi, S., Li, H., Ouyang, W.: Delving into localization errors for monocular 3D object detection. In: CVPR (2021) 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.177, + 0.787, + 0.203 + ], + "angle": 0, + "content": "45. Martin, D., Fowlkes, C., Malik, J.: Learning to detect natural image boundaries using local brightness, color, and texture cues. TPAMI (2004) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.204, + 0.786, + 0.229 + ], + "angle": 0, + "content": "46. Menze, M., Geiger, A.: Object scene flow for autonomous vehicles. In: CVPR (2015) 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.231, + 0.787, + 0.259 + ], + "angle": 0, + "content": "47. Michaud, S., Lalonde, J.F., Giguere, P.: Towards characterizing the behavior of lidars in snowy conditions. In: IROS (2015) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.26, + 0.786, + 0.286 + ], + "angle": 0, + "content": "48. Milanovic, V., Kasturi, A., Yang, J., Hu, F.: A fast single-pixel laser imager for VR/AR headset tracking. In: MOEMS and Miniaturized Systems XVI (2017) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.287, + 0.786, + 0.314 + ], + "angle": 0, + "content": "49. Park, D., Ambrus, R., Guizilini, V., Li, J., Gaidon, A.: Is pseudo-lidar needed for monocular 3d object detection? In: ICCV (2021) 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.315, + 0.786, + 0.342 + ], + "angle": 0, + "content": "50. Piccinelli, L., Sakaridis, C., Yu, F.: iDisc: Internal discretization for monocular depth estimation. In: CVPR (2023) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.343, + 0.786, + 0.37 + ], + "angle": 0, + "content": "51. Reading, C., Harakeh, A., Chae, J., Waslander, S.: Categorical depth distribution network for monocular 3D object detection. In: CVPR (2021) 2, 3, 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.371, + 0.786, + 0.397 + ], + "angle": 0, + "content": "52. Royo, S., Ballesta-Garcia, M.: An overview of LiDAR imaging systems for autonomous vehicles. Applied sciences (2019) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.398, + 0.786, + 0.424 + ], + "angle": 0, + "content": "53. Silberman, N., Hoiem, D., Kohli, P., Fergus, R.: Indoor segmentation and support inference from rgbd images. ECCV (2012) 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.425, + 0.786, + 0.452 + ], + "angle": 0, + "content": "54. Spinneker, R., Koch, C., Park, S.B., Yoon, J.J.: Fast fog detection for camera based advanced driver assistance systems. In: ITSC (2014) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.453, + 0.786, + 0.479 + ], + "angle": 0, + "content": "55. Sun, J., Li, Y., Kang, S.B., Shum, H.Y.: Symmetric stereo matching for occlusion handling. In: CVPR (2005) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.48, + 0.786, + 0.549 + ], + "angle": 0, + "content": "56. Sun, P., Kretzschmar, H., Dotiwalla, X., Chouard, A., Patnaik, V., Tsui, P., Guo, J., Zhou, Y., Chai, Y., Caine, B., Vasudevan, V., Han, W., Ngiam, J., Zhao, H., Timofeev, A., Ettinger, S., Krivokon, M., Gao, A., Joshi, A., Zhang, Y., Shlens, J., Chen, Z., Anguelov, D.: Scalability in perception for autonomous driving: Waymo open dataset. In: CVPR (2020) 1, 2, 3, 10, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.55, + 0.786, + 0.577 + ], + "angle": 0, + "content": "57. Uhrig, J., Schneider, N., Schneider, L., Franke, U., Brox, T., Geiger, A.: Sparsity invariant cnns. In: 3DV (2017) 3, 4, 12, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.578, + 0.786, + 0.604 + ], + "angle": 0, + "content": "58. Wang, J., Zickler, T.: Local detection of stereo occlusion boundaries. In: CVPR (2019) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.606, + 0.786, + 0.646 + ], + "angle": 0, + "content": "59. Wei, Y., Zhao, L., Zheng, W., Zhu, Z., Rao, Y., Huang, G., Lu, J., Zhou, J.: Surrounddepth: Entangling surrounding views for self-supervised multi-camera depth estimation. In: CoRL (2023) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.647, + 0.786, + 0.673 + ], + "angle": 0, + "content": "60. Wei, Y., Quan, L.: Asymmetrical occlusion handling using graph cut for multi-view stereo. In: CVPR (2005) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.675, + 0.74, + 0.688 + ], + "angle": 0, + "content": "61. Weng, J., Ahuja, N., Huang, T.S.: Two-view matching. In: ICCV (1988) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.689, + 0.786, + 0.741 + ], + "angle": 0, + "content": "62. Wilson, B., Qi, W., Agarwal, T., Lambert, J., Singh, J., Khandelwal, S., Pan, B., Kumar, R., Hartnett, A., Pontes, J.K., et al.: Argoverse 2: Next generation datasets for self-driving perception and forecasting. arXiv preprint arXiv:2301.00493 (2023) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.743, + 0.786, + 0.771 + ], + "angle": 0, + "content": "63. Wu, H., Wen, C., Li, W., Li, X., Yang, R., Wang, C.: Transformation-equivariant 3D object detection for autonomous driving. In: AAAI (2023) 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.772, + 0.786, + 0.798 + ], + "angle": 0, + "content": "64. Wu, H., Wen, C., Shi, S., Li, X., Wang, C.: Virtual sparse convolution for multimodal 3D object detection. In: CVPR (2023) 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.799, + 0.704, + 0.812 + ], + "angle": 0, + "content": "65. Xie, S., Tu, Z.: Holistically-nested edge detection. In: ICCV (2015) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.813, + 0.786, + 0.84 + ], + "angle": 0, + "content": "66. Yang, Z., Chen, J., Miao, Z., Li, W., Zhu, X., Zhang, L.: Deepinteraction: 3D object detection via modality interaction. NeurIPS (2022) 2, 4" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.356, + 0.127 + ], + "angle": 0, + "content": "S. Zhu et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.188 + ], + "angle": 0, + "content": "67. Yu, K., Tao, T., Xie, H., Lin, Z., Liang, T., Wang, B., Chen, P., Hao, D., Wang, Y., Liang, X.: Benchmarking the robustness of lidar-camera fusion for 3D object detection. In: CVPR (2023) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.189, + 0.785, + 0.217 + ], + "angle": 0, + "content": "68. Yuan, W., Gu, X., Dai, Z., Zhu, S., Tan, P.: Neural window fully-connected crfs for monocular depth estimation. In: CVPR (2022) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.218, + 0.785, + 0.244 + ], + "angle": 0, + "content": "69. Zheng, K., Li, S., Qin, K., Li, Z., Zhao, Y., Peng, Z., Cheng, H.: Depth estimation via sparse radar prior and driving scene semantics. In: ACCV (2022) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.245, + 0.785, + 0.272 + ], + "angle": 0, + "content": "70. Zhu, S., Brazil, G., Liu, X.: The edge of depth: Explicit constraints between segmentation and depth. In: CVPR (2020) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.273, + 0.785, + 0.299 + ], + "angle": 0, + "content": "71. Zhu, S., Liu, X.: Lighteddepth: Video depth estimation in light of limited inference view angles. In: CVPR (2023) 4, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.3, + 0.785, + 0.327 + ], + "angle": 0, + "content": "72. Zitnick, L., Kanade, T.: A cooperative algorithm for stereo matching and occlusion detection. TPAMI (2000) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.328, + 0.785, + 0.369 + ], + "angle": 0, + "content": "73. Zou, Z., Ye, X., Du, L., Cheng, X., Tan, X., Zhang, L., Feng, J., Xue, X., Ding, E.: The devil is in the task: Exploiting reciprocal appearance-localization features for monocular 3D object detection. In: ICCV (2021) 14" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.369 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/Remove Projective LiDAR Depthmap Artifacts via Exploiting Epipolar Geometry/0a1ce34c-ca2a-42ac-a914-2a125b92f644_origin.pdf b/2024/Remove Projective LiDAR Depthmap Artifacts via Exploiting Epipolar Geometry/0a1ce34c-ca2a-42ac-a914-2a125b92f644_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..261f572e4790ef3931aadb77c676080ffc91d2a0 --- /dev/null +++ b/2024/Remove Projective LiDAR Depthmap Artifacts via Exploiting Epipolar Geometry/0a1ce34c-ca2a-42ac-a914-2a125b92f644_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7c7b38470ccab6e39e4faa3678b4fd0c12346b33d0dfea0865a82d83dfa46250 +size 4924035 diff --git a/2024/Remove Projective LiDAR Depthmap Artifacts via Exploiting Epipolar Geometry/full.md b/2024/Remove Projective LiDAR Depthmap Artifacts via Exploiting Epipolar Geometry/full.md new file mode 100644 index 0000000000000000000000000000000000000000..b3e6a1d984ddbd84da7ebb8c010b1405599f009d --- /dev/null +++ b/2024/Remove Projective LiDAR Depthmap Artifacts via Exploiting Epipolar Geometry/full.md @@ -0,0 +1,446 @@ +# RePLAY: Remove Projective LiDAR Depthmap Artifacts via Exploiting Epipolar Geometry + +Shengjie Zhu*, Girish Chandar Ganesan*, Abhinav Kumar, and Xiaoming Liu + +Department of Computer Science and Engineering, Michigan State University, East Lansing, MI, USA, 48824 [zhusheng, ganesang, kumarab6, liuxm]@msu.edu https://shngjz.github.io/RePLAy/ + +![](images/1f6301cf1efa3f044f19cbd5f11efd2bf5dea2a4937c3bd078663b43033c201a.jpg) +Fig. 1: Remove Projective LiDAR Depthmap Artifacts. The LiDAR projected depthmap has misalignment with the RGB camera. The background scans highlighted by blue arrows incorrectly overlay the foreground such as cars and pedestrians. The artifacts persist unattended in most AV datasets including KITTI-360 [41], nuScenes [10], Waymo [56], and DDAD [24]. We propose an easy-to-use, parameter-free, and analytical solution to remove the projective artifacts, shown by green arrows. + +Abstract. 3D sensing is a fundamental task for Autonomous Vehicles. Its deployment often relies on aligned RGB cameras and LiDAR. Despite meticulous synchronization and calibration, systematic misalignment persists in LiDAR projected depthmap. This is due to the physical baseline distance between the two sensors. The artifact is often reflected as background LiDAR incorrectly projected onto the foreground, such as cars and pedestrians. The KITTI dataset uses stereo cameras as a heuristic solution to remove artifacts. However most AV datasets, including nuScenes, Waymo, and DDAD, lack stereo images, making the KITTI solution inapplicable. We propose RePLAY, a parameter-free analytical solution to remove the projective artifacts. We construct a binocular vision system between a hypothesized virtual LiDAR camera and the RGB camera. We then remove the projective artifacts by determining the epipolar occlusion with the proposed analytical solution. We show unanimous improvement in the State-of-The-Art (SoTA) monocular depth estimators and 3D object detectors with the artifacts-free depthmaps. + +Keywords: LiDAR $\cdot$ Autonomous Driving $\cdot$ Epipolar Geometry + +![](images/abdc8bd1ff419c6ec89d5a731263d1e2f7eb1cfb61d83e346dc32a6e42a1911e.jpg) +(a) Autonomous Vehicles [10] + +![](images/a9b1390cf76e39cdf350736ed7cdef7fda9681647bf5b25148afe62f5121633d.jpg) +(b) Robotics Machines [14] +Fig. 2: LiDAR-RGB sensor sets have a physical baseline distance between the two sensors, resulting in Fig. 1 projective artifacts. (a) The background scanned by LiDAR incorrectly overlays on the foreground pedestrian observed by the RGB camera. + +![](images/0c2984129ae310a10325f182ee878b2e9052d2ec106148698daaa0362979d1f5.jpg) +(c) Headset [26] + +# 1 Introduction + +LiDAR device scans the surrounding environment via emitting laser pulses into the 3D space. Its applications span across Autonomous Vehicles (AV) [40, 52, 56], AR/VR [27, 48], and Robotics [31]. Taking AV as an example, the LiDAR is synchronized and calibrated against the RGB camera. The correspondence between RGB pixels and LiDAR 3D points is determined by projecting the LiDAR 3D point onto the image plane, as shown with the Fig. 1 depthmaps. + +Despite ideal synchronization and calibration, there still exists systematic misalignment between the RGB camera and LiDAR. Fig. 1 highlights the projective artifacts in LiDAR depthmap with blue arrows, where the background scans incorrectly overlay on the foreground objects. + +The projective artifacts arise from the physical baseline distance between the RGB camera and LiDAR. This physical limitation is ubiquitous among different LiDAR-RGB sensor sets used in Robotics, AR/VR, and AV tasks (Fig. 2). The baseline distance (Fig. 2a) renders certain 3D scans exclusively visible to LiDAR, rather than the RGB camera. For those scans, their correspondence with image pixels is absent, causing Fig. 1 projective artifacts. + +The KITTI dataset [1] provides a heuristic solution to remove the projective artifacts, which we show in Fig. 3. However, the method requires stereo sensors, making it inapplicable to other AV datasets. Thus, the projective artifacts persist unattended in most AV datasets. Tab. 1 summarizes these datasets. + +We propose RePLAy: a parameter-free, non-learning, and analytical solution to resolve the projective artifacts. RePLAy only requires the relative pose between the LiDAR scanner and the RGB camera, i.e., their extrinsic matrix or calibration matrix. It applies to all AV datasets surveyed in Tab. 1 and diverse sensor settings included in Fig. 2. RePLAy does not even require RGB images, hence being intact at the asynchronization noise between the RGB camera and LiDAR. Fig. 3 provides a qualitative comparison with KITTI solution. + +RePLAy first adopts an auto-calibration algorithm to calibrate a hypothesized virtual camera at the origin of the LiDAR 3D point cloud. The virtual LiDAR camera formulates a stereo system with the RGB camera. We attribute the projective artifact to the epipolar occlusion in stereo vision. We propose an analytical solution for detecting epipolar occlusion using epipolar geometry. + +LiDAR projected depthmap is essential for diverse downstream applications. We select two to verify our effectiveness: monocular depth estimation (MonoDE) [5, 50, 59, 70] and monocular 3D object detection (Mono3DOD) that uses auxiliary depthmap [51, 63, 64, 66]. We benchmark their performance trained + +![](images/18a9013cee8c87fcca991d968278b7c503f0d37051be215535c0a497eb477054.jpg) +(a) Raw Depthmap + +![](images/032acb6e0f9fc1ef93f835b5392559af2fcb93fe731881f96067f7ef798ce947.jpg) +(b) RePLAy +Fig. 3: Qualitative comparison with KITTI semi-dense depthmap. RePLAY only requires the LiDAR calibration matrix. KITTI [57] uses multi-frame fusion with the stereo camera to remove projective artifacts. The higher density of KITTI depthmap is from the additional sensor and temporal fusion. + +![](images/48c1b008eaee63e26ab6e8dceea239a70cde5c0fbfd52bee5244b07ae1b6fbe6.jpg) +(c) KITTI [57] with Stereo + +Table 1: LiDAR Depthmap Projective Artifacts exist in most Autonomous Vehicle Datasets. KITTI uses stereo images as a heuristic solution. + +
KITTI [1]KITTI-360 [41]nuScenes [10]Waymo [56]DDAD [24]ApolloScape [30]Argoverse [62]
ArtifactsNoYesYesYesYesYesYes
+ +using LiDAR depthmap cleaned with multiple algorithms and RePLAy on AV datasets. We demonstrate unanimous improvement in State-of-The-Art (SoTA) methods [5,29,51] on both tasks. We additionally release the processed depthmap of five major AV datasets to benefit the community. Our contributions are: + +$\checkmark$ We utilize epipolar geometry to remove the projective artifacts in the LiDAR depthmap. +$\checkmark$ We demonstrate unanimous improvement in the SoTA monocular depth and monocular 3D object detection methods. +$\checkmark$ We provide processed depthmap in KITTI [1], KITTI-360 [41], nuScenes [10], Waymo [56], and DDAD [24] datasets. + +# 2 Related Work + +LiDAR and Camera Misalignment. Multiple sources of misalignment exist between LiDAR and RGB camera. Foggy, rainy, and snowy weather negatively impacts both LiDAR [16, 37, 47] and RGB cameras [18-20, 54]. The slight drift of LiDAR and camera accumulates during vehicle operation. [3, 39, 67] investigate the robustness to the spatial misalignment. [67] benchmarks the temporal asynchronization between LiDAR and RGB camera. The projective artifact is another misalignment that universally exists in multiple AV datasets, as shown in Tab. 1. Our solution complements prior study [67] by addressing it. + +Heuristic Solution for Projective Artifacts. KITTI [1, 57] dataset uses the stereo image as a heuristic solution. [57] produces a stereo reconstruction through semi-global matching [28]. They additionally augment the stereo results with 11 neighboring temporal frames. The projective artifacts are removed via + +Algorithm 1 RePLAy: Remove Projective Artifacts +Input: LiDAR Point Cloud V, Calibration Matrix $\mathbf{P} = [\mathbf{R}\mathbf{t}]$ , Camera Intrinsic $\mathbf{K}_r$ +Output: Occluded Pixels $\mathcal{O}$ +1: Auto-Calibration ▷ Sec. 3.1 +2: Set Virtual LiDAR Camera $\mathbf{K}_l$ +3: Produce Virtual LiDAR Depthmap $\mathbf{D}_l$ +4: Denseify Virtual LiDAR Depthmap $\overline{\mathbf{D}}_l$ +5: for v ∈ V do +6: p1, dp1 ← π(Rv | Kl) +7: for p2 ∈ Lp1 do ▷ Eq. (14) and Fig. 8 +8: dp2 ← bilinear interpolate $\overline{\mathbf{D}}_l$ at pixel location p2 ▷ Fig. 8 +9: if g(p1, p2, dp1, dp2 | Kl, Kr, [I t]) ≤ 0 then ▷ Eq. (15) +10: {p1} ∪ O +11: break + +consistency verification between the stereo depthmap and LiDAR depthmap. Compared to ours, depthmap from [57] has improved density. However, the requirement of stereo sensors limits its application, as summarized in Tab. 1. Additionally, [57] has a complicated pipeline, and is therefore, not used even in the KITTI-360 [41] dataset, which shares a similar setup. + +Stereo Occlusion Detection. Stereo co-planar cameras are classically formulated as the half-occlusion problem in stereo vision, i.e., the area occluded in one camera of a stereo vision system. There has been extensive research on detecting the half-occlusion region [2, 45, 58, 65]. Several works [6, 7, 32, 60, 72] jointly optimize occlusion and disparity. Others [17, 55, 61] apply left-right consistency check. The area receives different estimations from left and right views are excluded. RePLAy focuses on setups without stereo but where LiDAR provides rough 3D structures. Compared to the half-occlusion baseline which proposes a heuristic solution, RePLAy is an analytical solution. + +Downstream Tasks using LiDAR Depthmap. Multiple downstream tasks use LiDAR projected depthmaps. E.g., MonoDE methods [4, 5, 38, 68, 71] use depthmap as the GT supervision. Mono3DOD methods [36, 43, 63, 64, 66] use depthmap for supervising auxiliary task along with 3D object detection. When deployed to AV datasets other than KITTI, the aforementioned methods [42, 69] have diminished performance due to the projective artifacts. RePLAy improves their performance by resolving the projective artifacts in LiDAR depthmap. + +# 3 RePLAy + +We now describe RePLAy in Algorithm 1. Sec. 3.1 attributes the projective artifacts to the epipolar occlusion induced by the binocular vision system between the virtual LiDAR camera and RGB camera. Sec. 3.2 outlines our analytical solution. Sec. 3.3 hold further discussions. + +![](images/56ec7f48c4e539462254a9f00d7061d3cb87502dcea17c65b2d8e744a91ba4e9.jpg) +(a) LiDAR Point Cloud Bird-Eye View + +![](images/7359ba96e77b22761ddddcdd51b8855c34fb4771912147ebabb12351c17b518e.jpg) +(b) Depthmap $\mathbf{D}_l$ of Virtual LiDAR Camera. + +![](images/cc6005c98a3d65eba0f18a680a928fd4cf29a02759168a39e6eeb9f64c3e390b.jpg) +(c) Depthmap $\mathbf{D}_r$ of RGB Camera. +Fig. 4: Virtual LiDAR camera and RGB Binocular System. In (a), we plot the bird-eye view of the point cloud $\mathcal{V}$ from LiDAR. A binocular system is set between virtual LiDAR (epipolar left) and RGB camera (epipolar right). Subfigure (b) plots the depthmap $\mathbf{D}_l$ from LiDAR view. As stated in Lemma 1, depthmap $\mathbf{D}_l$ does not have projective artifacts. Compare (b) and (c), the artifacts arise when transfer from LiDAR to RGB camera, due to the epipolar occlusion among the binocular system. We formally define the projective artifacts (or epipolar occlusion) in Eq. (7). + +We execute Sec. 3.1 auto-calibration over the entire dataset. We initialize LiDAR intrinsic $\mathbf{K}_l$ the same as the RGB camera $\mathbf{K}_r$ with slight augmentation to ensure a sufficient Field-of-View. + +# 3.1 LiDAR Depthmap Projective Artifacts as Epipolar Occlusion + +Fig. 2 demonstrates the baseline distance between various LiDAR and RGB camera sensor sets. Fig. 2a shows that the background tree is solely visible to LiDAR. Intuitively, the background tree is occluded by the foreground pedestrian in a binocular vision system of the virtual LiDAR camera and RGB camera. + +Virtual LiDAR Camera. LiDAR produces a point cloud $\mathcal{V}$ by emitting laser pulses into 3D space. Mathematically, we assume the LiDAR point cloud $\mathcal{V}$ as a set of 3D rays originating from the coordinate origin: + +$$ +\mathcal {V} = \left\{\mathbf {v} _ {i} \right\} = \left\{\mathbf {0} + \bar {\mathbf {r}} _ {i} \cdot t _ {i} \mid i \in \mathbb {Z}, t > 0, \bar {\mathbf {r}} _ {i} \in \mathbb {R} ^ {3} \right\}, \tag {1} +$$ + +where $\mathbf{v}_i$ is a 3D point. $\overline{\mathbf{r}}_i$ is a normalized direction vector, and $t_i$ is the euclidean distance between the 3D point and LiDAR. If we position a virtual LiDAR camera with arbitrary resolution and intrinsic $\mathbf{K}_l$ at the LiDAR's location, then: + +Lemma 1. The depthmap of the virtual LiDAR camera is free of projective artifacts, i.e., no two scanned points $\mathbf{v}_1$ and $\mathbf{v}_2$ overlap after projection. + +$$ +\forall \mathbf {v} _ {1}, \mathbf {v} _ {2} \in \mathcal {V}, \| \pi (\mathbf {v} _ {1} \mid \mathbf {K} _ {l}) - \pi (\mathbf {v} _ {2} \mid \mathbf {K} _ {l}) \| _ {2} \neq 0, \tag {2} +$$ + +where $\pi (\cdot)$ stands for the projection process. See proof in Supp. (Appendix A). We visualize the artifacts-free depthmap $\mathbf{D}_l$ from the virtual LiDAR camera in Fig. 4b. + +![](images/d98fbaebe05c71db5e6e056a53358e127082fa01709ca3d72d3161c4252fb17c.jpg) +Fig. 5: Auto-Calibration. The LiDAR point cloud $\mathcal{V}$ in (a) is mapped to spherical coordinate $\theta -\phi$ with Eq. (3). We use a binary matrix $\mathbf{S}$ to record the unique pixel locations $S$ after rasterization. Plots (b) and (c) are matrix $\mathbf{S}$ before and after autocalibration where duplication (marked by arrows) is minimized with Eq. (5) loss $L$ . Plots (d) and (e) contrast the virtual LiDAR depthmap $\mathbf{D}_l$ before and after autocalibration where Lemma 1 is fulfilled. + +Auto-Calibration. In practice, Lemma 1 may not hold in a realistic LiDAR point cloud due to possible bias in the point cloud coordinate origin $\mathbf{0}$ . Fig. 5d, an example from the KITTI dataset shows persistent projective artifacts in the depthmap $\mathbf{D}_l$ when without any treatment. We perform auto-calibration to fulfill Lemma 1 by re-adjusting the point cloud origin. Fig. 4b and Fig. 5e show artifacts-free depthmap after auto-calibration. + +Auto-calibration aims to resolve the artifacts in the virtual LiDAR camera depthmap $\mathbf{D}_l$ . We convert each LiDAR point cloud to the spherical coordinate system. This hypothesizes an omnidirectional image $\mathbf{S}$ of the LiDAR point cloud. Denote the origin to estimate as $\mathbf{o}$ . Set the new 3D point $\mathbf{u} = \mathbf{v} + \mathbf{o}$ , we have the polar angle $\theta$ and azimuthal angle $\phi$ after applying a rasterization operation $[\cdot]:$ + +$$ +\theta = \left[ \arccos \left(u _ {x} / \left(u _ {x} ^ {2} + u _ {z} ^ {2}\right)\right) \right], \phi = \left[ \arccos \left(u _ {y} / \left(u _ {x} ^ {2} + u _ {y} ^ {2} + u _ {z} ^ {2}\right)\right) \right], \left[ \theta \phi \right] = f (\mathbf {v} + \mathbf {o}). \tag {3} +$$ + +Intuitively, function $f(\cdot)$ converts a 3D point $\mathbf{u} \in \mathcal{V}$ to a discrete 2D coordinate in $\mathbf{S}$ . We denote the set of unique coordinates as $S$ : + +$$ +\mathcal {S} = \left\{f (\mathbf {v} + \mathbf {o}) \mid \mathbf {v} \in \mathcal {V} \right\}. \tag {4} +$$ + +We implement Eq. (4) with a binary matrix $\mathbf{S}$ . Its elements are 1 at each discrete coordinate and 0 otherwise. The size of $\mathcal{S}$ equals to summarizing the matrix $\mathbf{S}$ . Artifacts become duplication in set $\mathcal{S}$ . Removing artifacts is to minimize duplication. The loss $L$ is then defined as: + +$$ +L = \| \mathcal {V} \| - \| \mathcal {S} \|. \tag {5} +$$ + +The loss $L$ is not differentiable. In each iteration, we approximate its gradient by ablating origin $\mathbf{o}$ with a small deviation and perform coordinate descent. + +Virtual LiDAR Camera and RGB Binocular System. In Fig. 4, we set up a binocular vision system, denoting the virtual LiDAR camera and RGB camera as the epipolar left and right camera respectively. Correspondingly, set their intrinsic as $\mathbf{K}_l$ and $\mathbf{K}_r$ , and projected depthmap as $\mathbf{D}_l$ and $\mathbf{D}_r$ . Thus the problem of removing the projective artifacts is translated to detecting the epipolar occlusion between the virtual LiDAR camera and RGB camera. + +![](images/700d9a39658aced8eb8f08322e87f0c64546d4df343c2b682e6e549186f2c061.jpg) +Rotation + +![](images/812966440980895d7fe174e0a8d7d4de5547b26ea9239a85dcd0e2f4c612aef9.jpg) + +![](images/9f24f8901b2622a3e8a412d866d44e0f1f584778dde342527c930495e1f5322c.jpg) +Fig. 6: Epipolar Occlusion with Pure Rotation Movement. We apply a rotation transformation along the $x-z$ axis in the camera coordinate system. As indicated by Lemma 2, a pure rotation movement does not lead to projective artifacts. + +Projective Artifacts are Epipolar Occlusion. We ensure that our virtual LiDAR has a larger field of view than the RGB camera by adjusting the resolution of the virtual LiDAR camera. Denote the relative movement between the LiDAR and RGB camera as $\mathbf{P} = [\mathbf{R}\mathbf{t}]$ , where the auto-calibrated center o is integrated into t. Denote the projected pixels on epipolar left (LiDAR) and epipolar right (RGB) as p and q, and their pixel depths as $d$ and $d^{\prime}$ . We have: + +$$ +\mathbf {p} = \pi (\mathbf {v} \mid \mathbf {K} _ {l}), \quad \mathbf {q} = \pi (\mathbf {R v} + \mathbf {t} \mid \mathbf {K} _ {r}). \tag {6} +$$ + +We then formally define the epipolar occlusion $\mathcal{O}$ on RGB view as: + +$$ +\mathcal {O} = \left\{\mathbf {q} _ {1} \mid \forall \mathbf {q} _ {1}, \mathbf {q} _ {2} \in \mathcal {Q}, d _ {\mathbf {q} _ {1}} ^ {\prime} > d _ {\mathbf {q} _ {2}} ^ {\prime}, \| \mathbf {q} _ {1} - \mathbf {q} _ {2} \| _ {2} = 0 \right\}. \tag {7} +$$ + +Eq. (7) simply suggests if two pixels in RGB camera overlap with each other, then the one $\mathbf{q}_1$ with larger depth is occluded by the one with smaller depth $\mathbf{q}_2$ . Occlusion detection finds all occluded pixels $\mathbf{q}_1$ , i.e., the set $\mathcal{O}$ . + +# 3.2 Epipolar Occlusion under Rotation and Translation + +Suppose the relative camera motion among the Fig. 4a binocular system is an arbitrary camera pose $\mathbf{P} = [\mathbf{R}\mathbf{t}]$ . Set $\mathbf{I}$ as the identity matrix. RePLAy decomposes $\mathbf{P}$ into a sequential pure rotation and pure translation movement: + +$$ +\mathbf {P} _ {1} = \left[ \begin{array}{l l} \mathbf {R} & \mathbf {0} \end{array} \right], \mathbf {P} _ {2} = \left[ \begin{array}{l l} \mathbf {I} & \mathbf {t} \end{array} \right], \mathbf {P} = \mathbf {P} _ {2} \mathbf {P} _ {1}. \tag {8} +$$ + +We outline pure rotation movement in Lemma 2 and Fig. 6. The pure translation movement is illustrated in Lemma 3 and Fig. 7. Without loss of generality, we assume the epipolar left and right cameras share the same intrinsic, i.e., $\mathbf{K}_l = \mathbf{K}_r$ . When with different intrinsics, we apply a resize and crop operation defined by $\mathbf{A} = \mathbf{K}_l\mathbf{K}_r^{-1}$ to unify the intrinsic. + +# Lemma 2. Pure rotation movement does not cause epipolar occlusion. + +We know from epipolar geometry [25] that pure rotation induces homography between two views. In other words, there is a one-to-one mapping between pixels before and after rotation. See Supp for more information. Fig. 6 illustrates Lemma 2 with a visual example. + +Lemma 3. Pure translation movement causes epipolar occlusion along the direction of the epipolar line. + +![](images/9a4a4358b3ec92f4a432e9d33a295893771ff5265b38ee891a5d65c36701a703.jpg) +Fig. 7: Epipoplar Occlusion with Pure Translation Movement. A background pixel $\mathbf{p}_1$ is gradually occluded by a foreground pixel $\mathbf{p}_2$ under pure translation. We formulate three binocular systems at different baseline distances. For simplicity, we set the depth of $\mathbf{p}_1$ to infinity, making its 2D location static. Direction vector $\mathbf{n}$ is set to point from $\mathbf{p}$ to $\mathbf{q}$ . The scalar $t$ is marked by blue line segment. From Lemma 3, (a) The pixel $\mathbf{p}_1$ is only occluded by pixel $\mathbf{p}_2$ at the reverse direction of epipolar line (red line segment). (b) $\mathbf{p}_1$ is not occluded. (c) $\mathbf{p}_1$ is occluded following Eq. (7). (d) $\mathbf{p}_1$ is considered occluded following Eq. (17). Intuitively, the occluded pixel $\mathbf{p}_1$ (in a) is "caught up" (in c) and "surpassed" (in d) by $\mathbf{p}_2$ while moving along the epipolar line. + +Proof. As pure rotation does not cause epipolar occlusion, we apply the translation movement over the artifacts-free depthmap after the pure rotation. With slight abuse of the notation, we set the pixels after pure rotation movement as $\mathbf{p}$ and $\mathbf{q}$ . The epipolar line $\mathbf{l_p}$ on the right image from the left image pixel $\mathbf{p}$ is: + +$$ +\mathbf {l} _ {\mathbf {p}} = \mathbf {K} ^ {- \top} \left[ \mathbf {t} _ {\times} \right] \mathbf {R} \mathbf {K} ^ {- 1} \mathbf {p} = \mathbf {K} ^ {- \top} \left[ \mathbf {t} _ {\times} \right] \mathbf {K} ^ {- 1} \mathbf {p}, \tag {9} +$$ + +where $[\mathbf{t}_{\times}]$ is the cross product in matrix form [71]. Due to a pure translation movement, the rotation $\mathbf{R}$ is removed as being an identity matrix. Further, the left and right epipolar lines overlap with each other in pure translation: + +$$ +\mathbf {l} _ {\mathbf {q}} = - \mathbf {K} ^ {- \top} \left[ \mathbf {t} _ {\times} \right] \mathbf {K} ^ {- 1} \mathbf {q}, \mathbf {p} ^ {\top} \mathbf {l} _ {\mathbf {q}} = 0, \mathbf {p} ^ {\top} \mathbf {l} _ {\mathbf {p}} = 0. \tag {10} +$$ + +Also shown in Fig. 7, pixel $\mathbf{q}$ is hence a function of pixel $\mathbf{p}$ : + +$$ +\mathbf {q} = \mathbf {p} + t \cdot \mathbf {n}, t = \| \mathbf {q} - \mathbf {p} \| _ {2} = \mathbf {n} ^ {\intercal} (\mathbf {q} - \mathbf {p}), \mathbf {n} = (\mathbf {q} - \mathbf {p}) / \sqrt {\| \mathbf {q} - \mathbf {p} \| _ {2} ^ {2}}, \tag {11} +$$ + +where $t$ is the moving distance, similar to the disparity in stereo vision. We define the vector $\mathbf{n}$ pointed from $\mathbf{p}$ to $\mathbf{q}$ as the direction of the epipolar line $\mathbf{l}$ . $\mathbf{p}$ and $\mathbf{q}$ are related by the equation $d^{\prime}\mathbf{q} = d\mathbf{K}_{r}\mathbf{R}\mathbf{K}_{l}^{-1}\mathbf{p} + \mathbf{K}_{r}\mathbf{t}$ . Expanding under pure translation movement gives us: + +$$ +t ^ {2} = \left(p _ {x} - q _ {x}\right) ^ {2} + \left(p _ {y} - q _ {y}\right) ^ {2} = \frac {\left(x + z \cdot q _ {x}\right) ^ {2} + \left(y + z \cdot q _ {y}\right) ^ {2}}{\left(d ^ {\prime} - z\right) ^ {2}} = \frac {\left(x + z \cdot q _ {x}\right) ^ {2} + \left(y + z \cdot q _ {y}\right) ^ {2}}{d ^ {2}}. \tag {12} +$$ + +![](images/3d39d62921456f7d1b58e19c6cfada73626e89c1d5bded49efc2025431470b30.jpg) +(a) Virtual LiDAR Depthmap $\mathbf{D}_l$ + +![](images/a4be020daa3bbc26cece9d94268f9f2879bc3a6eccb4f714eb5f2b0049a9348a.jpg) +(b) Densified Virtual LiDAR Depthmap $\overline{\mathbf{D}}_l$ +Fig. 8: Per-pixel Occlusion Examination. To overcome LiDAR sparsity, for a query pixel, we verify inspection pixels along epipolar line $\mathcal{L}_{\mathbf{p}_1}$ following Eq. (17) to determine occlusion. Depth $\overline{\mathbf{D}}_l$ is densified with nearest-neighbor interpolation. + +Please see the derivation in Supp. (Appendix B). Note, we assume pixels $\mathbf{p}$ and $\mathbf{q}$ are respectively visible to the left and right camera. This suggests $d > 0$ and $d = d' - z > 0$ . Eq. (12) suggests moving distance $t^2$ monotonously decreases as the projected depth $d'$ increases. It fits the intuition where a foreground object has a larger movement than a background object. From Eq. (7), when with occlusion, we have $\mathbf{q}_1 = \mathbf{q}_2$ , and $d_{\mathbf{q}_1} < d_{\mathbf{q}_2}$ . Combine Eq. (11) and Eq. (12): + +$$ +\mathbf {q} _ {1} = \mathbf {q} _ {2}, \text {a n d} d _ {\mathbf {q} _ {1}} < d _ {\mathbf {q} _ {2}} \longrightarrow t _ {1} ^ {2} > t _ {2} ^ {2} \longrightarrow \mathbf {n} ^ {\intercal} (\mathbf {p} _ {1} - \mathbf {p} _ {2}) < 0. \tag {13} +$$ + +From Eq. (13), if pixel $\mathbf{q}_1$ is occluded by $\mathbf{q}_2$ in the right view, pixel $\mathbf{q}_2$ must move a longer distance along the epipolar line than $\mathbf{q}_1$ . This suggests, in the left view, $\mathbf{p}_2$ resides at the "back" of $\mathbf{p}_1$ w.r.t. the moving direction. In other words, $\mathbf{p}_1$ is only occluded by other pixels along the reverse direction of epipolar line. + +Epipolar Occlusion Judgement. Combine Lemma 2 and 3, a pixel $\mathbf{p}_1$ in LiDAR view is occluded by other pixels $\mathbf{p}_2$ along the reverse of epipolar line direction. For pixel $\mathbf{p}_1$ , the set of inspection pixels is: + +$$ +\mathcal {L} _ {\mathbf {p} _ {1}} = \left\{\mathbf {p} _ {2} \mid \mathbf {p} _ {2} = \mathbf {p} _ {1} - \mathbf {n} \cdot k \cdot \Delta t, k \in \mathbb {Z} \right\}. \tag {14} +$$ + +$k$ is an integer. At occlusion, we have $\mathbf{q}_1 = \mathbf{q}_2$ , combined with Eq. (11): + +$$ +\mathbf {q} _ {1} = \mathbf {q} _ {2} \longrightarrow \mathbf {p} _ {1} + t _ {1} \mathbf {n} = \mathbf {p} _ {2} + t _ {2} \mathbf {n} \longrightarrow \mathbf {n} ^ {\top} \left(\mathbf {p} _ {1} - \mathbf {p} _ {2}\right) + t _ {1} - t _ {2} = 0. \tag {15} +$$ + +We use Fig. 7 to explain Eq. (15). In Fig. 7a, the pixel $\mathbf{p}_1$ and $\mathbf{p}_2$ has an initial distance $\mathbf{n}^{\top}(\mathbf{p}_1 - \mathbf{p}_2)$ . In Fig. 7c, after translation movement, each pixel moves a distance $t_1$ and $t_2$ along the epipolar line. Occlusion happens when pixel $\mathbf{p}_2$ overlaps pixel $\mathbf{p}_1$ . We summarize Eq. (15) into a function $g(\cdot)$ as: + +$$ +\lambda \leq g \left(\mathbf {p} _ {1}, \mathbf {p} _ {2}, d _ {\mathbf {p} _ {1}}, d _ {\mathbf {p} _ {2}} \mid \mathbf {K}, \mathbf {P}\right) \leq 0, \tag {16} +$$ + +where the distance $t_1$ and $t_2$ are computed using intrinsic, camera motion, and depth values. In Fig. 7b, when $g(\cdot) > 0$ , there is no occlusion. In Fig. 7c, when $g(\cdot) = 0$ , two pixels overlap and occlusion happens. However, this is hard due to LiDAR sparsity, rasterization, and other noise sources. Hence, we relax with a threshold $\lambda$ . In practice, we set $\lambda$ to $-\infty$ . This suggests the pixel $\mathbf{p}_1$ is occluded if any pixel $\mathbf{p}_2$ in the epipolar line "surpasses" it in the RGB view, as shown in Fig. 7d. Formally, we consider pixel $\mathbf{p}_1$ occluded as: + +$$ +\mathcal {O} = \left\{\mathbf {p} _ {1} \mid \exists \mathbf {p} _ {2} \in \mathcal {L} _ {\mathbf {p} _ {1}}, g \left(\mathbf {p} _ {1}, \mathbf {p} _ {2}, d _ {\mathbf {p} _ {1}}, d _ {\mathbf {p} _ {2}} \mid \mathbf {K}, \mathbf {P}\right) < 0 \right\}. \tag {17} +$$ + +Eq. (17) is also a necessary condition of epipolar occlusion. Due to the Li-DAR sparsity, we apply a nearest-neighbor interpolation over virtual LiDAR depthmap $\mathbf{D}_l$ to acquire $d_{\mathbf{p}_2}$ whenever required, as shown in Fig. 8. + +# 3.3 Discussions + +Half-Occlusion (Depth Buffering) Baseline. Sec. 2 shows that half-occlusion algorithms solve projective artifacts in stereo vision, i.e., stereo/epipolar occlusion, without knowing LiDAR point cloud. Despite a different setting, they provide a baseline solution. When preparing RGB depthmap $\mathbf{D}_r$ , only pixels with smaller depth are preserved if multiple 3D points $\mathbf{v}$ have duplicate 2D locations after rasterization. The raw LiDAR depthmap uses half-occlusion. So, we refer Half-Occlusion as "Raw" following [23]. + +Does RePLAy simply densify the sparse depthmap to remove occlusion? No. Densification alone is insufficient to remove occlusion. RePLAy enables robust densification by formulating a stereo system with an auto-calibration procedure to remove occlusion. + +Universal Existence of Projective Artifacts. Fig. 2 pictorially depicts that projective artifacts exist under all LiDAR and camera sensor settings, while Fig. 7 shows that these artifacts scale up with a larger baseline distance between sensors. We now show that this indeed is the case. The artifact width distribution plot on the outdoor KITTI and the indoor NYUv2 [53] datasets are shown in Fig. 10 in the Supp. The LiDAR and RGB camera in KITTI have a baseline distance of $0.4m$ [22]. We set the baseline distance to $0.1m$ following the AR headset configuration of Fig. 2. Despite the indoor image having a lesser resolution than outdoor ( $480 \times 640$ compared to $375 \times 1242$ ), we observe a similar distribution. Detailed definitions and evaluation protocols are in Supp. The analysis confirms that projective artifacts exist in both indoor and outdoor datasets. + +# 4 Experiments + +We benchmark RePLAy on depthmap quality, MonoDE and Mono3DOD tasks. Sec. 4.1 assesses depthmap quality, while Sec. 4.2 and Sec. 4.3. report MonoDE and Mono3DOD results respectively. + +Datasets. We experiment on KITTI [21], nuScenes [10] and Waymo [56] datasets. KITTI provides stereo ground truth while the others do not. + +Data Splits. We use the following splits of the above datasets: + +- KITTI Eigen [15]: 23,158 training and 652 validation samples of raw KITTI. +- KITTI Stereo [46]: 200 training and 200 testing samples out of which 145 training scenes are mapped to the raw KITTI dataset. +- nuScenes Frontal Val: 28,130 training and 2,436 random validation samples (from 6,019 samples) from front camera and top LiDAR. +- Waymo Frontal Val: 52,386 training and 40,839 validation samples from the front camera and top LiDAR. We construct its training set by sampling every third frame from the training sequences as in [34]. + +Table 2: Depthmap Quality Assessment with baselines and RePLAy. We adopt KITTI semi-dense depthmap and KITTI Stereo disparity map as the reference GT. [Key: Best, Second Best, Mod. = Modified] + +
GTEval RegionDepthmapδ0.5↑δ1↑δ2↑RMS ↓RMSlog ↓ARel ↓SRel ↓log10 ↓SIlog ↓
Semi-DenseAllRaw0.9730.9830.9901.9350.0870.0290.3320.0108.660
Mod. Half-Occ0.9830.9910.9961.3470.0570.0190.1130.0075.712
RePLAy0.9850.9940.9981.2410.0460.0160.0630.0074.592
StereoAllRaw0.9500.9620.9704.1980.1690.0831.6040.02416.691
Mod. Half-Occ0.9710.9820.9882.5410.1050.0470.5460.01610.340
RePLAy0.9750.9860.9912.2240.0860.0340.3540.0148.362
ForegroundRaw0.8590.8710.8928.7260.3170.2285.9740.05829.772
Mod. Half-Occ0.9190.9310.9495.2850.1990.1542.3660.03319.250
RePLAy0.9250.9370.9564.8370.1790.1001.9290.03017.319
BackgroundRaw0.9670.9800.9862.7080.1130.0520.7740.01711.037
Mod. Half-Occ0.9800.9910.9951.7010.0700.0320.2100.0136.638
RePLAy0.9820.9930.9981.4390.0520.0270.0850.0124.722
+ +Table 3: Depthmap Quality Assessment on Removed Artifacts. RePLAy outperforms the Mod. half-occlusion baseline. For removed artifacts, a worse performance suggests a better algorithm. [Key: Best] + +
GTEval RegionDepthmapδ0.5↓δ1↓δ2↓RMS↑RMSlog↑ARel↑SRel↑log10↑
Semi-DenseAllMod. Half-Occ0.8140.8460.8935.8500.2750.1963.7620.053
RePLAy0.2740.3630.54812.0840.5870.77815.8510.202
+ +- KITTI Detection Val: 3,748 training and 3,762 validation images from the left camera for the Mono3DOD task. + +Ground Truth (GT) Depthmaps. We use the following as GT depthmaps: + +- KITTI Eigen: We use the semi-dense depthmap available in KITTI. +- KITTI Stereo: We generate depthmaps from GT disparity maps, which are denser than semi-dense depthmaps. +- nuScenes and Waymo Frontal Val: We generate GT depthmaps by projecting only those LiDAR points present inside GT 3D bounding boxes. To prepare a reference depthmap, we follow [49] to filter LiDAR point cloud with GT 3D bounding boxes labels. Correspondingly, the evaluation is confined to the bounding box area, as only depth labels in that region are retained. + +# 4.1 Depthmap Quality Assessment + +Baselines. We compare RePLAy to raw depthmap, which is the half-occlusion (depth buffering) method and a modified half-occlusion method. The modified half-occlusion modifies the half-occlusion method (Sec. 3.3) by back-projecting the estimated dense virtual depthmap $(\overline{\mathbf{D}}_l)$ onto the RGB camera and only preserving pixels in $\mathbf{D}_r$ that are less than their back-projected values. + +Results. Tab. 2 compares the RePLAy depthmap with these baselines processed depthmaps. Tab. 2 verifies that RePLAy depthmap outperforms the raw and modified half-occlusion baselines. RePLAy depthmap agrees more with semidense or disparity depthmap after removing the projective artifacts. Fig. 3 shows qualitative comparisons. We believe projective artifacts persist in the half-occlusion method because of the LiDAR sparsity, leading to inferior depthmaps. + +Table 4: Auto-Calibration Ablation on Depthmap Quality Assessment with semidense GT depthmap. [Key: Best] + +
DepthmapAutoCalibδ0.5↑δ1↑δ2↑RMS ↓RMSlog ↓ARel ↓SRel ↓log10 ↓SIlog ↓
RePLAy×0.9810.9900.9951.4730.0570.0190.1110.0075.680
✓0.9850.9940.9981.2410.0460.0160.0630.0074.592
+ +Table 5: MonoDE Results on KITTI Stereo. ZoeDepth [5] trained with RePLAy depthmaps outperforms raw and semi-dense [57] depthmaps on foreground. [Key: Best, Second Best]. + +
Eval RegionDepthmapδ1↑δ2↑δ3↑RMS ↓RMSlog ↓ARel ↓SRel ↓log10 ↓SIlog ↓
AllRaw0.9550.9820.9922.6470.1160.0670.4200.02611.324
Semi-Dense0.9770.9940.9972.2740.0820.0430.2010.0187.971
RePLAy0.9740.9920.9962.2880.0890.0490.2480.0208.588
ForegroundRaw0.8860.9470.9803.6920.1800.1271.1120.02216.518
Semi-Dense0.9760.9900.9942.0930.0910.0560.3440.0228.342
RePLAy0.9450.9780.9892.9390.1280.0810.6420.02911.798
BackgroundRaw0.9730.9930.9982.2940.0840.0480.1880.0207.981
Semi-Dense0.9780.9960.9982.2830.0760.0400.1670.0177.239
RePLAy0.9810.9970.9992.0840.0700.0390.1380.0176.681
+ +Foreground Improvement. Tab. 2 rows 5 and 6 shows a significant improvement in foreground depthmap quality. On metric $\delta_{0.5}$ , we reduce the errors by $46.8\% = 1 - \frac{1 - 0.925}{1 - 0.859}$ . This aligns with our observation that projective artifacts mostly occur over foreground objects, confirming Fig. 1. + +Comparison on Removed Artifacts. We next compare RePLAy on removed artifacts in Tab. 3. Tab. 3 shows that that in-addition to removing some artifacts, the modified half-occlusion also removes some non-artifacts (good points). Ideally, the removed points should be artifacts; therefore, the best algorithm should perform the worst on the quality assessment of removed points and hence the reversal in the directional indicator for Tab. 3. Qualitative comparison between raw, modified half-occlusion and RePLAy is provided in Supp. (Fig. 11). + +Auto-Calibration Ablation. We next conduct ablations of the Auto-Calibration module (Sec. 3.1) in Tab. 4. Tab. 4 confirms the effectiveness of Auto-Calibration module in RePLAy, since Auto-Calibration provides $15\%$ and $19\%$ relative improvement in ARel and $\mathrm{SI}_{\log}$ metrics respectively. + +# 4.2 MonoDE: Monocular Depth Estimation + +Baselines. We compare RePLAy to raw and Semi-Dense methods. We use ZoeDepth [5] as the depth model due to its superior generalization. + +KITTI Stereo and Eigen Results. We train ZoeDepth [5] from scratch using depthmap with and without applying RePLAy on KITTI Eigen split. We evaluate on KITTI Stereo and Eigen split in Tabs. 5 and 6 respectively. Tab. 5 shows that ZoeDepth trained with RePLAy depthmap significantly outperforms the baselines. We observe despite semi-dense GT [57] using additional stereo cameras, RePLAy outperforms Semi-Dense especially in the background. This + +![](images/6f220f610d772c75d69f6e1415071e503e51b4be7125aac7fa4a21db92d575bb.jpg) +(a) RGB Input + +![](images/de6d75883c0fb02382335191b6cf6e8f0c901164dc6d35b77bde25f1915bcea6.jpg) +(b) Raw depthmaps +Fig. 9: MonoDE Qualitative Results on KITTI Eigen dataset. ZoeDepth [5] trained with RePLAy depthmaps (c) outperforms the one with raw depthmaps (b), especially on foreground. + +![](images/c4d84a85b31e9289af6581e81dc94bc23919ed25f0c87387e078c878bf636a4d.jpg) +(c) RePLAy depthmaps + +Table 6: MonoDE Results. ZoeDepth trained with RePLAy outperforms the baselines in foreground regions of all datasets. [Key: Best] + +
Data SplitEval RegionDepthmapδ1↑δ2↑δ3↑RMS↓RMSlog↓ARel↓SRel↓log10↓SIlog↓
KITTI EigenAllRaw0.8690.9590.9872.9110.1630.1240.6060.04613.501
RePLAy0.9270.9850.9972.1850.1100.0820.2870.0338.674
nuScenes ValForegroundRaw0.9710.9870.9931.3470.0750.0490.2450.0186.339
RePLAy0.9810.9930.9961.1500.0590.0390.1370.0164.905
Waymo ValForegroundRaw0.9410.9880.9963.4680.1110.0790.4870.0329.116
RePLAy0.9620.9930.9983.0570.0910.0630.3670.0267.322
+ +Table 7: MonoDE Analysis to Environmental Conditions on nuScenes Val split. ZoeDepth trained with RePLAY depthmaps shows consistent improvements on foreground under all environmental conditions. [Key: Best] + +
ConditionSamplesDepthmapδ1↑δ2↑δ3↑RMS ↓RMSlog ↓ARel ↓SRel ↓log10 ↓SIlog ↓
Rain864Raw0.9840.9910.9951.3380.0580.0400.1630.0164.859
RePLAy0.9890.9930.9951.2220.0500.0340.1270.0154.110
Night533Raw0.9650.9820.9901.3470.0890.0590.3350.0217.481
RePLAy0.9780.9910.9951.0630.0670.0450.1720.0175.374
Sunny1039Raw0.9640.9870.9931.3520.0810.0530.2670.0196.985
RePLAy0.9760.9930.9961.1340.0620.0400.1280.0165.324
All Tab. 62436Raw0.9710.9870.9931.3470.0750.0490.2450.0186.339
RePLAy0.9810.9930.9961.1500.0590.0390.1370.0164.905
+ +may be due to [57] aggressively removing the background pixels around the foreground object boundaries. Fig. 9 shows this result visually. ZoeDepth trained with the raw depthmap learns strip-shaped artifacts (green boxes). + +Foreground Improvement. We additionally ablate depth performance on foreground, background, and entire image using the GT segmentation mask in Tab. 5. Tab. 5 shows that ZoeDepth trained with RePLAy depthmap outperforms in the foreground area, which agrees with the conclusion of Sec. 4.1. + +nuScenes and Waymo Frontal Val Results. We benchmark MonoDE on two datasets without stereo sensors: nuScenes [10] and Waymo [56] in Tab. 6. Tab. 6 confirms that RePLAy achieves significant improvements on these datasets. + +Effect of Environmental Conditions. We next analyze the effect of environmental conditions on the nuScenes Val split in Tab. 7. We categorize the nuScenes [10] val set into rainy, night, and sunny following their description tag (nuScenes-devkit GitHub #124). Tab. 7 shows that RePLAy is robust across multiple conditions. + +# 4.3 Mono3DOD: Monocular 3D Object Detection + +We finally verify the impact of RePLAy on the Mono3DOD task. + +Baseline. We compare RePLAy to raw method. We take SoTA 3D detectors, CaDDN [51] and MonoDTR [29], to showcase the impact of cleaning projected LiDAR depthmap with RePLAy, since these detectors use depthmap as an auxiliary training task. The two methods cover both convolutional and transformer architectures, making our benchmarking comprehensive. However, these two SoTA detectors use depth supervision in training differently. + +CaDDN. CaDDN uses the densified raw depthmaps of [33]. For CaDDN experiments, we apply RePLAy to the projected depthmaps and then densify those maps using the same depth completion algorithm [33] and parameters of [51]. + +MonoDTR. MonoDTR [29] precomputes their GT depthmaps by $4\times$ downsampling raw projected depthmaps. For MonoDTR experiments, we clean the projected depthmaps and then $4\times$ down-sampling. + +Results. Tab. 8 shows results of retraining both detectors with our depthmaps on KITTI Val split. Cleaned depthmaps improves the performance of CaDDN and MonoDTR on most of the metrics. The biggest gains appear in the Easy + +category. This is expected as Easy cars are closer to the ego camera and receive the maximum number of LiDAR points. Hence, improving the projected depthmap by RePLAy benefits the Easy category more. + +Table 8: KITTI Val Mono3DOD Results on Car category. Using RePLAy depthmaps in training boosts the Mono3DOD performance. [Key: Best, Second Best, $* =$ Reported, $\dagger=$ Retrained, Mod= Moderate.] + +
MethodDepthmapIoU3D > 0.7
AP 3D|R40[%]↑
EasyModHard
M3D-RPN [8]-14.5311.078.65
MonoPair [12]-16.2812.3010.42
MonoDLE [44]-17.4513.6611.68
Kinematic [9]-19.7614.1010.47
GrooMeD [35]-19.6714.3211.27
MonoRUn [11]-20.0214.6512.61
D4LCN [13]Raw*22.3216.2012.30
DFR-Net [73]Raw*24.8117.7814.41
CaDDN [51]Raw†21.6015.6713.24
RePLAy22.8815.7013.18
MonoDTR [29]Raw†24.5018.6615.69
RePLAy26.1518.7715.69
+ +# 5 Conclusion + +We propose an analytical solution to remove projective artifacts in LiDAR depthmap. Our parameter-free method, RePLAy, applies to diverse LiDAR and RGB camera sensor settings. RePLAy only requires the relative position between LiDAR and RGB camera. We verify its benefits with unanimous improvements on SoTA MonoDE and Mono3DOD methods. We release processed depthmaps of five major AV datasets to benefit the community. + +Limitation. We use nearest-neighbor interpolation to densify virtual LiDAR depthmaps. Advanced strategies, e.g., depth completion, may improve accuracy. + +# References + +1. The KITTI Vision Benchmark Suite. http://www.cvlabs.net/datasets/kitti/eval_object.php?obj_benchmark=3d (2022), accessed: 2022-07-03 2, 3 +2. Arbelaez, P., Maire, M., Fowlkes, C., Malik, J.: Contour detection and hierarchical image segmentation. TPAMI (2010) 4 +3. Bai, X., Hu, Z., Zhu, X., Huang, Q., Chen, Y., Fu, H., Tai, C.L.: Transfusion: Robust lidar-camera fusion for 3D object detection with transformers. In: CVPR (2022) 3 +4. Bhat, S.F., Alhashim, I., Wonka, P.: Adabins: Depth estimation using adaptive bins. In: CVPR (2021) 4 +5. Bhat, S.F., Birkl, R., Wofk, D., Wonka, P., Müller, M.: ZoeDepth: Zero-shot transfer by combining relative and metric depth. arXiv preprint arXiv:2302.12288 (2023) 2, 3, 4, 12, 13 +6. Birchfield, S., Tomasi, C.: Depth discontinuities by pixel-to-pixel stereo. IJCV (1999) 4 +7. Bobick, A.F., Intille, S.S.: Large occlusion stereo. IJCV (1999) 4 +8. Brazil, G., Liu, X.: M3D-RPN: Monocular 3D region proposal network for object detection. In: ICCV (2019) 14 +9. Brazil, G., Pons-Moll, G., Liu, X., Schiele, B.: Kinematic 3D object detection in monocular video. In: ECCV (2020) 14 +0. Caesar, H., Bankiti, V., Lang, A., Vora, S., Liong, V., Xu, Q., Krishnan, A., Pan, Y., Baldan, G., Beijbom, O.: nuScenes: A multimodal dataset for autonomous driving. In: CVPR (2020) 1, 2, 3, 10, 13 +1. Chen, H., Huang, Y., Tian, W., Gao, Z., Xiong, L.: MonoRun: Monocular 3D object detection by reconstruction and uncertainty propagation. In: CVPR (2021) 14 +2. Chen, Y., Tai, L., Sun, K., Li, M.: MonoPair: Monocular 3D object detection using pairwise spatial relationships. In: CVPR (2020) 14 +3. Ding, M., Huo, Y., Yi, H., Wang, Z., Shi, J., Lu, Z., Luo, P.: Learning depth-guided convolutions for monocular 3D object detection. In: CVPR Workshops (2020) 14 +4. Ebadi, K., Bernreiter, L., Biggie, H., Catt, G., Chang, Y., Chatterjee, A., Denniston, C.E., Deschênes, S.P., Harlow, K., Khattak, S., et al.: Present and future of slam in extreme underground environments. arXiv preprint arXiv:2208.01787 (2022) 2 +5. Eigen, D., Puhrsch, C., Fergus, R.: Depth map prediction from a single image using a multi-scale deep network. NeurIPS (2014) 10 +6. Filgueira, A., González-Jorge, H., Lagüela, S., Díaz-Vilarino, L., Arias, P.: Quantifying the influence of rain in lidar performance. Measurement (2017) 3 +7. Fua, P.: A parallel stereo algorithm that produces dense depth maps and preserves image features. Machine vision and applications (1993) 4 +8. Garg, K., Nayar, S.K.: When does a camera see rain? In: ICCV (2005) 3 +9. Garg, K., Nayar, S.K.: Vision and rain. IJCV (2007) 3 +20. Garvelmann, J., Pohl, S., Weiler, M.: From observation to the quantification of snow processes with a time-lapse camera network. Hydrology and Earth Systems Sciences (2013) 3 +21. Geiger, A., Lenz, P., Stiller, C., Urtasun, R.: Vision meets robotics: The KITTI dataset. IJRR (2013) 10 +22. Geiger, A., Lenz, P., Urtasun, R.: Are we ready for autonomous driving? the KITTI vision benchmark suite. In: CVPR (2012) 10 + +23. Godard, C., Mac Aodha, O., Firman, M., Brostow, G.J.: Digging into self-supervised monocular depth estimation. In: ICCV (2019) 10 +24. Guizilini, V., Ambrus, R., Pillai, S., Raventos, A., Gaidon, A.: 3D packing for self-supervised monocular depth estimation. In: CVPR (2020) 1, 3 +25. Hartley, R., Zisserman, A.: Multiple view geometry in computer vision. Cambridge university press (2003) 7 +26. Herskovitz, J., Wu, J., White, S., Pavel, A., Reyes, G., Guo, A., Bigham, J.P.: Making mobile augmented reality applications accessible. In: SIGACCESS (2020) 2 +27. Hillmann, C., Hillmann, C.: Comparing the gear vr, oculus go, and oculus quest. Unreal for Mobile and Standalone VR: Create Professional VR Apps Without Coding (2019) 2 +28. Hirschmuller, H.: Stereo processing by semiglobal matching and mutual information. TPAMI (2007) 3 +29. Huang, K.C., Wu, T.H., Su, H.T., Hsu, W.: MonoDTR: Monocular 3D object detection with depth-aware transformer. In: CVPR (2022) 3, 14 +30. Huang, X., Cheng, X., Geng, Q., Cao, B., Zhou, D., Wang, P., Lin, Y., Yang, R.: The apolloscape dataset for autonomous driving. In: CVPR (2018) 3 +31. Hutabarat, D., Rivai, M., Purwanto, D., Hutomo, H.: Lidar-based obstacle avoidance for the autonomous mobile robot. In: 2019 12th International Conference on Information & Communication Technology and System (ICTS). pp. 197-202. IEEE (2019) 2 +32. Ishikawa, H., Geiger, D.: Occlusions, discontinuities, and epipolar lines in stereo. In: ECCV (1998) 4 +33. Ku, J., Harakeh, A., Waslander, S.: In defense of classical image processing: Fast depth completion on the CPU. In: CRV (2018) 14 +34. Kumar, A., Brazil, G., Corona, E., Parchami, A., Liu, X.: DEVIANT: Depth Equivariant Network for Monocular 3D Object Detection. In: ECCV (2022) 10 +35. Kumar, A., Brazil, G., Liu, X.: GrooMeD-NMS: Grouped mathematically differentiable nms for monocular 3D object detection. In: CVPR (2021) 14 +36. Kumar, A., Guo, Y., Huang, X., Ren, L., Liu, X.: SeaBird: Segmentation in Bird's View with Dice Loss Improves Monocular 3D Detection of Large Objects. In: CVPR (2024) 4 +37. Kutila, M., Pyykonen, P., Holzhüter, H., Colomb, M., Duthon, P.: Automotive lidar performance verification in fog and rain. In: ITSC (2018) 3 +38. Lee, J.H., Han, M.K., Ko, D.W., Suh, I.H.: From big to small: Multi-scale local planar guidance for monocular depth estimation. arXiv preprint arXiv:1907.10326 (2019) 4 +39. Li, Y., Yu, A.W., Meng, T., Caine, B., Ngiam, J., Peng, D., Shen, J., Lu, Y., Zhou, D., Le, Q.V., et al.: Deepfusion: Lidar-camera deep fusion for multi-modal 3d object detection. In: CVPR (2022) 3 +40. Li, Y., Ibanez-Guzman, J.: LiDAR for autonomous driving: The principles, challenges, and trends for automotive lidar and perception systems. IEEE Signal Processing Magazine (2020) 2 +41. Liao, Y., Xie, J., Geiger, A.: KITTI-360: A novel dataset and benchmarks for urban scene understanding in 2D and 3D. TPAMI (2022) 1, 3, 4 +42. Lin, J.T., Dai, D., Van Gool, L.: Depth estimation from monocular images and sparse radar data. In: IROS (2020) 4 +43. Liu, F., Liu, X.: Voxel-based 3d detection and reconstruction of multiple objects from a single image. In: NeurIPS (2021) 4 + +44. Ma, X., Zhang, Y., Xu, D., Zhou, D., Yi, S., Li, H., Ouyang, W.: Delving into localization errors for monocular 3D object detection. In: CVPR (2021) 14 +45. Martin, D., Fowlkes, C., Malik, J.: Learning to detect natural image boundaries using local brightness, color, and texture cues. TPAMI (2004) 4 +46. Menze, M., Geiger, A.: Object scene flow for autonomous vehicles. In: CVPR (2015) 10 +47. Michaud, S., Lalonde, J.F., Giguere, P.: Towards characterizing the behavior of lidars in snowy conditions. In: IROS (2015) 3 +48. Milanovic, V., Kasturi, A., Yang, J., Hu, F.: A fast single-pixel laser imager for VR/AR headset tracking. In: MOEMS and Miniaturized Systems XVI (2017) 2 +49. Park, D., Ambrus, R., Guizilini, V., Li, J., Gaidon, A.: Is pseudo-lidar needed for monocular 3d object detection? In: ICCV (2021) 11 +50. Piccinelli, L., Sakaridis, C., Yu, F.: iDisc: Internal discretization for monocular depth estimation. In: CVPR (2023) 2 +51. Reading, C., Harakeh, A., Chae, J., Waslander, S.: Categorical depth distribution network for monocular 3D object detection. In: CVPR (2021) 2, 3, 14 +52. Royo, S., Ballesta-Garcia, M.: An overview of LiDAR imaging systems for autonomous vehicles. Applied sciences (2019) 2 +53. Silberman, N., Hoiem, D., Kohli, P., Fergus, R.: Indoor segmentation and support inference from rgbd images. ECCV (2012) 10 +54. Spinneker, R., Koch, C., Park, S.B., Yoon, J.J.: Fast fog detection for camera based advanced driver assistance systems. In: ITSC (2014) 3 +55. Sun, J., Li, Y., Kang, S.B., Shum, H.Y.: Symmetric stereo matching for occlusion handling. In: CVPR (2005) 4 +56. Sun, P., Kretzschmar, H., Dotiwalla, X., Chouard, A., Patnaik, V., Tsui, P., Guo, J., Zhou, Y., Chai, Y., Caine, B., Vasudevan, V., Han, W., Ngiam, J., Zhao, H., Timofeev, A., Ettinger, S., Krivokon, M., Gao, A., Joshi, A., Zhang, Y., Shlens, J., Chen, Z., Anguelov, D.: Scalability in perception for autonomous driving: Waymo open dataset. In: CVPR (2020) 1, 2, 3, 10, 13 +57. Uhrig, J., Schneider, N., Schneider, L., Franke, U., Brox, T., Geiger, A.: Sparsity invariant cnns. In: 3DV (2017) 3, 4, 12, 13 +58. Wang, J., Zickler, T.: Local detection of stereo occlusion boundaries. In: CVPR (2019) 4 +59. Wei, Y., Zhao, L., Zheng, W., Zhu, Z., Rao, Y., Huang, G., Lu, J., Zhou, J.: Surrounddepth: Entangling surrounding views for self-supervised multi-camera depth estimation. In: CoRL (2023) 2 +60. Wei, Y., Quan, L.: Asymmetrical occlusion handling using graph cut for multi-view stereo. In: CVPR (2005) 4 +61. Weng, J., Ahuja, N., Huang, T.S.: Two-view matching. In: ICCV (1988) 4 +62. Wilson, B., Qi, W., Agarwal, T., Lambert, J., Singh, J., Khandelwal, S., Pan, B., Kumar, R., Hartnett, A., Pontes, J.K., et al.: Argoverse 2: Next generation datasets for self-driving perception and forecasting. arXiv preprint arXiv:2301.00493 (2023) 3 +63. Wu, H., Wen, C., Li, W., Li, X., Yang, R., Wang, C.: Transformation-equivariant 3D object detection for autonomous driving. In: AAAI (2023) 2, 4 +64. Wu, H., Wen, C., Shi, S., Li, X., Wang, C.: Virtual sparse convolution for multimodal 3D object detection. In: CVPR (2023) 2, 4 +65. Xie, S., Tu, Z.: Holistically-nested edge detection. In: ICCV (2015) 4 +66. Yang, Z., Chen, J., Miao, Z., Li, W., Zhu, X., Zhang, L.: Deepinteraction: 3D object detection via modality interaction. NeurIPS (2022) 2, 4 + +67. Yu, K., Tao, T., Xie, H., Lin, Z., Liang, T., Wang, B., Chen, P., Hao, D., Wang, Y., Liang, X.: Benchmarking the robustness of lidar-camera fusion for 3D object detection. In: CVPR (2023) 3 +68. Yuan, W., Gu, X., Dai, Z., Zhu, S., Tan, P.: Neural window fully-connected crfs for monocular depth estimation. In: CVPR (2022) 4 +69. Zheng, K., Li, S., Qin, K., Li, Z., Zhao, Y., Peng, Z., Cheng, H.: Depth estimation via sparse radar prior and driving scene semantics. In: ACCV (2022) 4 +70. Zhu, S., Brazil, G., Liu, X.: The edge of depth: Explicit constraints between segmentation and depth. In: CVPR (2020) 2 +71. Zhu, S., Liu, X.: Lighteddepth: Video depth estimation in light of limited inference view angles. In: CVPR (2023) 4, 8 +72. Zitnick, L., Kanade, T.: A cooperative algorithm for stereo matching and occlusion detection. TPAMI (2000) 4 +73. Zou, Z., Ye, X., Du, L., Cheng, X., Tan, X., Zhang, L., Feng, J., Xue, X., Ding, E.: The devil is in the task: Exploiting reciprocal appearance-localization features for monocular 3D object detection. In: ICCV (2021) 14 \ No newline at end of file diff --git a/2024/Remove Projective LiDAR Depthmap Artifacts via Exploiting Epipolar Geometry/images.zip b/2024/Remove Projective LiDAR Depthmap Artifacts via Exploiting Epipolar Geometry/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..1a259d4ea73dec40577750f6d2bf9b66cf17b2c8 --- /dev/null +++ b/2024/Remove Projective LiDAR Depthmap Artifacts via Exploiting Epipolar Geometry/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:76c7677b9b2f2fc50aef780e1cba49c7532eeaf12ca24317b95cca4c5a980458 +size 780695 diff --git a/2024/Remove Projective LiDAR Depthmap Artifacts via Exploiting Epipolar Geometry/layout.json b/2024/Remove Projective LiDAR Depthmap Artifacts via Exploiting Epipolar Geometry/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..e13e6e289182ba092e3011705d695972e6a4d37b --- /dev/null +++ b/2024/Remove Projective LiDAR Depthmap Artifacts via Exploiting Epipolar Geometry/layout.json @@ -0,0 +1,13344 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 140, + 111, + 475, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 140, + 111, + 475, + 148 + ], + "spans": [ + { + "bbox": [ + 140, + 111, + 475, + 148 + ], + "type": "text", + "content": "RePLAY: Remove Projective LiDAR Depthmap Artifacts via Exploiting Epipolar Geometry" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 133, + 168, + 479, + 181 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 168, + 479, + 181 + ], + "spans": [ + { + "bbox": [ + 133, + 168, + 479, + 181 + ], + "type": "text", + "content": "Shengjie Zhu*, Girish Chandar Ganesan*, Abhinav Kumar, and Xiaoming Liu" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 188, + 190, + 425, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 188, + 190, + 425, + 236 + ], + "spans": [ + { + "bbox": [ + 188, + 190, + 425, + 236 + ], + "type": "text", + "content": "Department of Computer Science and Engineering, Michigan State University, East Lansing, MI, USA, 48824 [zhusheng, ganesang, kumarab6, liuxm]@msu.edu https://shngjz.github.io/RePLAy/" + } + ] + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 133, + 262, + 481, + 367 + ], + "blocks": [ + { + "bbox": [ + 133, + 262, + 481, + 367 + ], + "lines": [ + { + "bbox": [ + 133, + 262, + 481, + 367 + ], + "spans": [ + { + "bbox": [ + 133, + 262, + 481, + 367 + ], + "type": "image", + "image_path": "1f6301cf1efa3f044f19cbd5f11efd2bf5dea2a4937c3bd078663b43033c201a.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 370, + 482, + 437 + ], + "lines": [ + { + "bbox": [ + 130, + 370, + 482, + 437 + ], + "spans": [ + { + "bbox": [ + 130, + 370, + 482, + 437 + ], + "type": "text", + "content": "Fig. 1: Remove Projective LiDAR Depthmap Artifacts. The LiDAR projected depthmap has misalignment with the RGB camera. The background scans highlighted by blue arrows incorrectly overlay the foreground such as cars and pedestrians. The artifacts persist unattended in most AV datasets including KITTI-360 [41], nuScenes [10], Waymo [56], and DDAD [24]. We propose an easy-to-use, parameter-free, and analytical solution to remove the projective artifacts, shown by green arrows." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 159, + 453, + 455, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 453, + 455, + 628 + ], + "spans": [ + { + "bbox": [ + 159, + 453, + 455, + 628 + ], + "type": "text", + "content": "Abstract. 3D sensing is a fundamental task for Autonomous Vehicles. Its deployment often relies on aligned RGB cameras and LiDAR. Despite meticulous synchronization and calibration, systematic misalignment persists in LiDAR projected depthmap. This is due to the physical baseline distance between the two sensors. The artifact is often reflected as background LiDAR incorrectly projected onto the foreground, such as cars and pedestrians. The KITTI dataset uses stereo cameras as a heuristic solution to remove artifacts. However most AV datasets, including nuScenes, Waymo, and DDAD, lack stereo images, making the KITTI solution inapplicable. We propose RePLAY, a parameter-free analytical solution to remove the projective artifacts. We construct a binocular vision system between a hypothesized virtual LiDAR camera and the RGB camera. We then remove the projective artifacts by determining the epipolar occlusion with the proposed analytical solution. We show unanimous improvement in the State-of-The-Art (SoTA) monocular depth estimators and 3D object detectors with the artifacts-free depthmaps." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 160, + 638, + 422, + 649 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 638, + 422, + 649 + ], + "spans": [ + { + "bbox": [ + 160, + 638, + 422, + 649 + ], + "type": "text", + "content": "Keywords: LiDAR " + }, + { + "bbox": [ + 160, + 638, + 422, + 649 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 638, + 422, + 649 + ], + "type": "text", + "content": " Autonomous Driving " + }, + { + "bbox": [ + 160, + 638, + 422, + 649 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 638, + 422, + 649 + ], + "type": "text", + "content": " Epipolar Geometry" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 135, + 653, + 228, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 653, + 228, + 665 + ], + "spans": [ + { + "bbox": [ + 135, + 653, + 228, + 665 + ], + "type": "text", + "content": "* Equal Contributions" + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 140, + 115, + 278, + 144 + ], + "blocks": [ + { + "bbox": [ + 140, + 115, + 278, + 144 + ], + "lines": [ + { + "bbox": [ + 140, + 115, + 278, + 144 + ], + "spans": [ + { + "bbox": [ + 140, + 115, + 278, + 144 + ], + "type": "image", + "image_path": "abdc8bd1ff419c6ec89d5a731263d1e2f7eb1cfb61d83e346dc32a6e42a1911e.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 156, + 145, + 261, + 156 + ], + "lines": [ + { + "bbox": [ + 156, + 145, + 261, + 156 + ], + "spans": [ + { + "bbox": [ + 156, + 145, + 261, + 156 + ], + "type": "text", + "content": "(a) Autonomous Vehicles [10]" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 294, + 113, + 396, + 144 + ], + "blocks": [ + { + "bbox": [ + 294, + 113, + 396, + 144 + ], + "lines": [ + { + "bbox": [ + 294, + 113, + 396, + 144 + ], + "spans": [ + { + "bbox": [ + 294, + 113, + 396, + 144 + ], + "type": "image", + "image_path": "a9b1390cf76e39cdf350736ed7cdef7fda9681647bf5b25148afe62f5121633d.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 296, + 145, + 392, + 156 + ], + "lines": [ + { + "bbox": [ + 296, + 145, + 392, + 156 + ], + "spans": [ + { + "bbox": [ + 296, + 145, + 392, + 156 + ], + "type": "text", + "content": "(b) Robotics Machines [14]" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 164, + 481, + 198 + ], + "lines": [ + { + "bbox": [ + 130, + 164, + 481, + 198 + ], + "spans": [ + { + "bbox": [ + 130, + 164, + 481, + 198 + ], + "type": "text", + "content": "Fig. 2: LiDAR-RGB sensor sets have a physical baseline distance between the two sensors, resulting in Fig. 1 projective artifacts. (a) The background scanned by LiDAR incorrectly overlays on the foreground pedestrian observed by the RGB camera." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 410, + 113, + 477, + 144 + ], + "blocks": [ + { + "bbox": [ + 410, + 113, + 477, + 144 + ], + "lines": [ + { + "bbox": [ + 410, + 113, + 477, + 144 + ], + "spans": [ + { + "bbox": [ + 410, + 113, + 477, + 144 + ], + "type": "image", + "image_path": "0c2984129ae310a10325f182ee878b2e9052d2ec106148698daaa0362979d1f5.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 414, + 145, + 471, + 156 + ], + "lines": [ + { + "bbox": [ + 414, + 145, + 471, + 156 + ], + "spans": [ + { + "bbox": [ + 414, + 145, + 471, + 156 + ], + "type": "text", + "content": "(c) Headset [26]" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 210, + 229, + 222 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 210, + 229, + 222 + ], + "spans": [ + { + "bbox": [ + 132, + 210, + 229, + 222 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 235, + 480, + 306 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 235, + 480, + 306 + ], + "spans": [ + { + "bbox": [ + 130, + 235, + 480, + 306 + ], + "type": "text", + "content": "LiDAR device scans the surrounding environment via emitting laser pulses into the 3D space. Its applications span across Autonomous Vehicles (AV) [40, 52, 56], AR/VR [27, 48], and Robotics [31]. Taking AV as an example, the LiDAR is synchronized and calibrated against the RGB camera. The correspondence between RGB pixels and LiDAR 3D points is determined by projecting the LiDAR 3D point onto the image plane, as shown with the Fig. 1 depthmaps." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 307, + 480, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 307, + 480, + 354 + ], + "spans": [ + { + "bbox": [ + 130, + 307, + 480, + 354 + ], + "type": "text", + "content": "Despite ideal synchronization and calibration, there still exists systematic misalignment between the RGB camera and LiDAR. Fig. 1 highlights the projective artifacts in LiDAR depthmap with blue arrows, where the background scans incorrectly overlay on the foreground objects." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 354, + 480, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 354, + 480, + 426 + ], + "spans": [ + { + "bbox": [ + 130, + 354, + 480, + 426 + ], + "type": "text", + "content": "The projective artifacts arise from the physical baseline distance between the RGB camera and LiDAR. This physical limitation is ubiquitous among different LiDAR-RGB sensor sets used in Robotics, AR/VR, and AV tasks (Fig. 2). The baseline distance (Fig. 2a) renders certain 3D scans exclusively visible to LiDAR, rather than the RGB camera. For those scans, their correspondence with image pixels is absent, causing Fig. 1 projective artifacts." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 426, + 480, + 473 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 426, + 480, + 473 + ], + "spans": [ + { + "bbox": [ + 130, + 426, + 480, + 473 + ], + "type": "text", + "content": "The KITTI dataset [1] provides a heuristic solution to remove the projective artifacts, which we show in Fig. 3. However, the method requires stereo sensors, making it inapplicable to other AV datasets. Thus, the projective artifacts persist unattended in most AV datasets. Tab. 1 summarizes these datasets." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 474, + 480, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 474, + 480, + 557 + ], + "spans": [ + { + "bbox": [ + 130, + 474, + 480, + 557 + ], + "type": "text", + "content": "We propose RePLAy: a parameter-free, non-learning, and analytical solution to resolve the projective artifacts. RePLAy only requires the relative pose between the LiDAR scanner and the RGB camera, i.e., their extrinsic matrix or calibration matrix. It applies to all AV datasets surveyed in Tab. 1 and diverse sensor settings included in Fig. 2. RePLAy does not even require RGB images, hence being intact at the asynchronization noise between the RGB camera and LiDAR. Fig. 3 provides a qualitative comparison with KITTI solution." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 558, + 480, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 558, + 480, + 617 + ], + "spans": [ + { + "bbox": [ + 130, + 558, + 480, + 617 + ], + "type": "text", + "content": "RePLAy first adopts an auto-calibration algorithm to calibrate a hypothesized virtual camera at the origin of the LiDAR 3D point cloud. The virtual LiDAR camera formulates a stereo system with the RGB camera. We attribute the projective artifact to the epipolar occlusion in stereo vision. We propose an analytical solution for detecting epipolar occlusion using epipolar geometry." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 617, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 481, + 665 + ], + "type": "text", + "content": "LiDAR projected depthmap is essential for diverse downstream applications. We select two to verify our effectiveness: monocular depth estimation (MonoDE) [5, 50, 59, 70] and monocular 3D object detection (Mono3DOD) that uses auxiliary depthmap [51, 63, 64, 66]. We benchmark their performance trained" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "text", + "content": "S. Zhu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 141, + 114, + 250, + 207 + ], + "blocks": [ + { + "bbox": [ + 141, + 114, + 250, + 207 + ], + "lines": [ + { + "bbox": [ + 141, + 114, + 250, + 207 + ], + "spans": [ + { + "bbox": [ + 141, + 114, + 250, + 207 + ], + "type": "image", + "image_path": "18a9013cee8c87fcca991d968278b7c503f0d37051be215535c0a497eb477054.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 159, + 209, + 230, + 218 + ], + "lines": [ + { + "bbox": [ + 159, + 209, + 230, + 218 + ], + "spans": [ + { + "bbox": [ + 159, + 209, + 230, + 218 + ], + "type": "text", + "content": "(a) Raw Depthmap" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 255, + 114, + 365, + 208 + ], + "blocks": [ + { + "bbox": [ + 255, + 114, + 365, + 208 + ], + "lines": [ + { + "bbox": [ + 255, + 114, + 365, + 208 + ], + "spans": [ + { + "bbox": [ + 255, + 114, + 365, + 208 + ], + "type": "image", + "image_path": "032acb6e0f9fc1ef93f835b5392559af2fcb93fe731881f96067f7ef798ce947.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 287, + 209, + 333, + 218 + ], + "lines": [ + { + "bbox": [ + 287, + 209, + 333, + 218 + ], + "spans": [ + { + "bbox": [ + 287, + 209, + 333, + 218 + ], + "type": "text", + "content": "(b) RePLAy" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 221, + 480, + 266 + ], + "lines": [ + { + "bbox": [ + 130, + 221, + 480, + 266 + ], + "spans": [ + { + "bbox": [ + 130, + 221, + 480, + 266 + ], + "type": "text", + "content": "Fig. 3: Qualitative comparison with KITTI semi-dense depthmap. RePLAY only requires the LiDAR calibration matrix. KITTI [57] uses multi-frame fusion with the stereo camera to remove projective artifacts. The higher density of KITTI depthmap is from the additional sensor and temporal fusion." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 369, + 114, + 477, + 208 + ], + "blocks": [ + { + "bbox": [ + 369, + 114, + 477, + 208 + ], + "lines": [ + { + "bbox": [ + 369, + 114, + 477, + 208 + ], + "spans": [ + { + "bbox": [ + 369, + 114, + 477, + 208 + ], + "type": "image", + "image_path": "48c1b008eaee63e26ab6e8dceea239a70cde5c0fbfd52bee5244b07ae1b6fbe6.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 375, + 209, + 471, + 218 + ], + "lines": [ + { + "bbox": [ + 375, + 209, + 471, + 218 + ], + "spans": [ + { + "bbox": [ + 375, + 209, + 471, + 218 + ], + "type": "text", + "content": "(c) KITTI [57] with Stereo" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 133, + 301, + 479, + 323 + ], + "blocks": [ + { + "bbox": [ + 131, + 270, + 480, + 292 + ], + "lines": [ + { + "bbox": [ + 131, + 270, + 480, + 292 + ], + "spans": [ + { + "bbox": [ + 131, + 270, + 480, + 292 + ], + "type": "text", + "content": "Table 1: LiDAR Depthmap Projective Artifacts exist in most Autonomous Vehicle Datasets. KITTI uses stereo images as a heuristic solution." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 133, + 301, + 479, + 323 + ], + "lines": [ + { + "bbox": [ + 133, + 301, + 479, + 323 + ], + "spans": [ + { + "bbox": [ + 133, + 301, + 479, + 323 + ], + "type": "table", + "html": "
KITTI [1]KITTI-360 [41]nuScenes [10]Waymo [56]DDAD [24]ApolloScape [30]Argoverse [62]
ArtifactsNoYesYesYesYesYesYes
", + "image_path": "6cfb7f2262d6cf9ee96285062dfe9bec8dffef196e55331a20605210b4e800ab.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 349, + 480, + 397 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 349, + 480, + 397 + ], + "spans": [ + { + "bbox": [ + 130, + 349, + 480, + 397 + ], + "type": "text", + "content": "using LiDAR depthmap cleaned with multiple algorithms and RePLAy on AV datasets. We demonstrate unanimous improvement in State-of-The-Art (SoTA) methods [5,29,51] on both tasks. We additionally release the processed depthmap of five major AV datasets to benefit the community. Our contributions are:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 400, + 479, + 471 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 132, + 400, + 479, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 400, + 479, + 422 + ], + "spans": [ + { + "bbox": [ + 132, + 400, + 479, + 422 + ], + "type": "inline_equation", + "content": "\\checkmark" + }, + { + "bbox": [ + 132, + 400, + 479, + 422 + ], + "type": "text", + "content": " We utilize epipolar geometry to remove the projective artifacts in the LiDAR depthmap." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 424, + 479, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 424, + 479, + 445 + ], + "spans": [ + { + "bbox": [ + 132, + 424, + 479, + 445 + ], + "type": "inline_equation", + "content": "\\checkmark" + }, + { + "bbox": [ + 132, + 424, + 479, + 445 + ], + "type": "text", + "content": " We demonstrate unanimous improvement in the SoTA monocular depth and monocular 3D object detection methods." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 448, + 479, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 448, + 479, + 471 + ], + "spans": [ + { + "bbox": [ + 132, + 448, + 479, + 471 + ], + "type": "inline_equation", + "content": "\\checkmark" + }, + { + "bbox": [ + 132, + 448, + 479, + 471 + ], + "type": "text", + "content": " We provide processed depthmap in KITTI [1], KITTI-360 [41], nuScenes [10], Waymo [56], and DDAD [24] datasets." + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 132, + 495, + 237, + 507 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 495, + 237, + 507 + ], + "spans": [ + { + "bbox": [ + 132, + 495, + 237, + 507 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 130, + 521, + 480, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 480, + 616 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 480, + 616 + ], + "type": "text", + "content": "LiDAR and Camera Misalignment. Multiple sources of misalignment exist between LiDAR and RGB camera. Foggy, rainy, and snowy weather negatively impacts both LiDAR [16, 37, 47] and RGB cameras [18-20, 54]. The slight drift of LiDAR and camera accumulates during vehicle operation. [3, 39, 67] investigate the robustness to the spatial misalignment. [67] benchmarks the temporal asynchronization between LiDAR and RGB camera. The projective artifact is another misalignment that universally exists in multiple AV datasets, as shown in Tab. 1. Our solution complements prior study [67] by addressing it." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 130, + 617, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 481, + 665 + ], + "type": "text", + "content": "Heuristic Solution for Projective Artifacts. KITTI [1, 57] dataset uses the stereo image as a heuristic solution. [57] produces a stereo reconstruction through semi-global matching [28]. They additionally augment the stereo results with 11 neighboring temporal frames. The projective artifacts are removed via" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 411, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 411, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 411, + 91, + 447, + 102 + ], + "type": "text", + "content": "RePLAY" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "code", + "bbox": [ + 132, + 130, + 481, + 277 + ], + "blocks": [ + { + "bbox": [ + 132, + 115, + 354, + 129 + ], + "lines": [ + { + "bbox": [ + 132, + 115, + 354, + 129 + ], + "spans": [ + { + "bbox": [ + 132, + 115, + 354, + 129 + ], + "type": "text", + "content": "Algorithm 1 RePLAy: Remove Projective Artifacts" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "code_caption" + }, + { + "bbox": [ + 132, + 130, + 481, + 277 + ], + "lines": [ + { + "bbox": [ + 132, + 130, + 481, + 277 + ], + "spans": [ + { + "bbox": [ + 132, + 130, + 481, + 277 + ], + "type": "text", + "content": "Input: LiDAR Point Cloud V, Calibration Matrix " + }, + { + "bbox": [ + 132, + 130, + 481, + 277 + ], + "type": "inline_equation", + "content": "\\mathbf{P} = [\\mathbf{R}\\mathbf{t}]" + }, + { + "bbox": [ + 132, + 130, + 481, + 277 + ], + "type": "text", + "content": ", Camera Intrinsic " + }, + { + "bbox": [ + 132, + 130, + 481, + 277 + ], + "type": "inline_equation", + "content": "\\mathbf{K}_r" + }, + { + "bbox": [ + 132, + 130, + 481, + 277 + ], + "type": "text", + "content": " \nOutput: Occluded Pixels " + }, + { + "bbox": [ + 132, + 130, + 481, + 277 + ], + "type": "inline_equation", + "content": "\\mathcal{O}" + }, + { + "bbox": [ + 132, + 130, + 481, + 277 + ], + "type": "text", + "content": " \n1: Auto-Calibration ▷ Sec. 3.1 \n2: Set Virtual LiDAR Camera " + }, + { + "bbox": [ + 132, + 130, + 481, + 277 + ], + "type": "inline_equation", + "content": "\\mathbf{K}_l" + }, + { + "bbox": [ + 132, + 130, + 481, + 277 + ], + "type": "text", + "content": " \n3: Produce Virtual LiDAR Depthmap " + }, + { + "bbox": [ + 132, + 130, + 481, + 277 + ], + "type": "inline_equation", + "content": "\\mathbf{D}_l" + }, + { + "bbox": [ + 132, + 130, + 481, + 277 + ], + "type": "text", + "content": " \n4: Denseify Virtual LiDAR Depthmap " + }, + { + "bbox": [ + 132, + 130, + 481, + 277 + ], + "type": "inline_equation", + "content": "\\overline{\\mathbf{D}}_l" + }, + { + "bbox": [ + 132, + 130, + 481, + 277 + ], + "type": "text", + "content": " \n5: for v ∈ V do \n6: p1, dp1 ← π(Rv | Kl) \n7: for p2 ∈ Lp1 do ▷ Eq. (14) and Fig. 8 \n8: dp2 ← bilinear interpolate " + }, + { + "bbox": [ + 132, + 130, + 481, + 277 + ], + "type": "inline_equation", + "content": "\\overline{\\mathbf{D}}_l" + }, + { + "bbox": [ + 132, + 130, + 481, + 277 + ], + "type": "text", + "content": " at pixel location p2 ▷ Fig. 8 \n9: if g(p1, p2, dp1, dp2 | Kl, Kr, [I t]) ≤ 0 then ▷ Eq. (15) \n10: {p1} ∪ O \n11: break" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "code_body" + } + ], + "index": 3, + "sub_type": "algorithm" + }, + { + "bbox": [ + 130, + 303, + 480, + 363 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 303, + 480, + 363 + ], + "spans": [ + { + "bbox": [ + 130, + 303, + 480, + 363 + ], + "type": "text", + "content": "consistency verification between the stereo depthmap and LiDAR depthmap. Compared to ours, depthmap from [57] has improved density. However, the requirement of stereo sensors limits its application, as summarized in Tab. 1. Additionally, [57] has a complicated pipeline, and is therefore, not used even in the KITTI-360 [41] dataset, which shares a similar setup." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 365, + 481, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 365, + 481, + 472 + ], + "spans": [ + { + "bbox": [ + 130, + 365, + 481, + 472 + ], + "type": "text", + "content": "Stereo Occlusion Detection. Stereo co-planar cameras are classically formulated as the half-occlusion problem in stereo vision, i.e., the area occluded in one camera of a stereo vision system. There has been extensive research on detecting the half-occlusion region [2, 45, 58, 65]. Several works [6, 7, 32, 60, 72] jointly optimize occlusion and disparity. Others [17, 55, 61] apply left-right consistency check. The area receives different estimations from left and right views are excluded. RePLAy focuses on setups without stereo but where LiDAR provides rough 3D structures. Compared to the half-occlusion baseline which proposes a heuristic solution, RePLAy is an analytical solution." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 474, + 481, + 558 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 474, + 481, + 558 + ], + "spans": [ + { + "bbox": [ + 130, + 474, + 481, + 558 + ], + "type": "text", + "content": "Downstream Tasks using LiDAR Depthmap. Multiple downstream tasks use LiDAR projected depthmaps. E.g., MonoDE methods [4, 5, 38, 68, 71] use depthmap as the GT supervision. Mono3DOD methods [36, 43, 63, 64, 66] use depthmap for supervising auxiliary task along with 3D object detection. When deployed to AV datasets other than KITTI, the aforementioned methods [42, 69] have diminished performance due to the projective artifacts. RePLAy improves their performance by resolving the projective artifacts in LiDAR depthmap." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 584, + 204, + 597 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 584, + 204, + 597 + ], + "spans": [ + { + "bbox": [ + 132, + 584, + 204, + 597 + ], + "type": "text", + "content": "3 RePLAy" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 617, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 481, + 665 + ], + "type": "text", + "content": "We now describe RePLAy in Algorithm 1. Sec. 3.1 attributes the projective artifacts to the epipolar occlusion induced by the binocular vision system between the virtual LiDAR camera and RGB camera. Sec. 3.2 outlines our analytical solution. Sec. 3.3 hold further discussions." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "type": "text", + "content": "S. Zhu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 114, + 279, + 215 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 279, + 215 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 279, + 215 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 279, + 215 + ], + "type": "image", + "image_path": "56ec7f48c4e539462254a9f00d7061d3cb87502dcea17c65b2d8e744a91ba4e9.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 140, + 217, + 279, + 226 + ], + "lines": [ + { + "bbox": [ + 140, + 217, + 279, + 226 + ], + "spans": [ + { + "bbox": [ + 140, + 217, + 279, + 226 + ], + "type": "text", + "content": "(a) LiDAR Point Cloud Bird-Eye View" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 284, + 114, + 480, + 161 + ], + "blocks": [ + { + "bbox": [ + 284, + 114, + 480, + 161 + ], + "lines": [ + { + "bbox": [ + 284, + 114, + 480, + 161 + ], + "spans": [ + { + "bbox": [ + 284, + 114, + 480, + 161 + ], + "type": "image", + "image_path": "7359ba96e77b22761ddddcdd51b8855c34fb4771912147ebabb12351c17b518e.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 302, + 162, + 463, + 171 + ], + "lines": [ + { + "bbox": [ + 302, + 162, + 463, + 171 + ], + "spans": [ + { + "bbox": [ + 302, + 162, + 463, + 171 + ], + "type": "text", + "content": "(b) Depthmap " + }, + { + "bbox": [ + 302, + 162, + 463, + 171 + ], + "type": "inline_equation", + "content": "\\mathbf{D}_l" + }, + { + "bbox": [ + 302, + 162, + 463, + 171 + ], + "type": "text", + "content": " of Virtual LiDAR Camera." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 284, + 172, + 480, + 216 + ], + "blocks": [ + { + "bbox": [ + 284, + 172, + 480, + 216 + ], + "lines": [ + { + "bbox": [ + 284, + 172, + 480, + 216 + ], + "spans": [ + { + "bbox": [ + 284, + 172, + 480, + 216 + ], + "type": "image", + "image_path": "cc6005c98a3d65eba0f18a680a928fd4cf29a02759168a39e6eeb9f64c3e390b.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 320, + 217, + 446, + 226 + ], + "lines": [ + { + "bbox": [ + 320, + 217, + 446, + 226 + ], + "spans": [ + { + "bbox": [ + 320, + 217, + 446, + 226 + ], + "type": "text", + "content": "(c) Depthmap " + }, + { + "bbox": [ + 320, + 217, + 446, + 226 + ], + "type": "inline_equation", + "content": "\\mathbf{D}_r" + }, + { + "bbox": [ + 320, + 217, + 446, + 226 + ], + "type": "text", + "content": " of RGB Camera." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 232, + 480, + 309 + ], + "lines": [ + { + "bbox": [ + 130, + 232, + 480, + 309 + ], + "spans": [ + { + "bbox": [ + 130, + 232, + 480, + 309 + ], + "type": "text", + "content": "Fig. 4: Virtual LiDAR camera and RGB Binocular System. In (a), we plot the bird-eye view of the point cloud " + }, + { + "bbox": [ + 130, + 232, + 480, + 309 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 130, + 232, + 480, + 309 + ], + "type": "text", + "content": " from LiDAR. A binocular system is set between virtual LiDAR (epipolar left) and RGB camera (epipolar right). Subfigure (b) plots the depthmap " + }, + { + "bbox": [ + 130, + 232, + 480, + 309 + ], + "type": "inline_equation", + "content": "\\mathbf{D}_l" + }, + { + "bbox": [ + 130, + 232, + 480, + 309 + ], + "type": "text", + "content": " from LiDAR view. As stated in Lemma 1, depthmap " + }, + { + "bbox": [ + 130, + 232, + 480, + 309 + ], + "type": "inline_equation", + "content": "\\mathbf{D}_l" + }, + { + "bbox": [ + 130, + 232, + 480, + 309 + ], + "type": "text", + "content": " does not have projective artifacts. Compare (b) and (c), the artifacts arise when transfer from LiDAR to RGB camera, due to the epipolar occlusion among the binocular system. We formally define the projective artifacts (or epipolar occlusion) in Eq. (7)." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 325, + 479, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 325, + 479, + 361 + ], + "spans": [ + { + "bbox": [ + 130, + 325, + 479, + 361 + ], + "type": "text", + "content": "We execute Sec. 3.1 auto-calibration over the entire dataset. We initialize LiDAR intrinsic " + }, + { + "bbox": [ + 130, + 325, + 479, + 361 + ], + "type": "inline_equation", + "content": "\\mathbf{K}_l" + }, + { + "bbox": [ + 130, + 325, + 479, + 361 + ], + "type": "text", + "content": " the same as the RGB camera " + }, + { + "bbox": [ + 130, + 325, + 479, + 361 + ], + "type": "inline_equation", + "content": "\\mathbf{K}_r" + }, + { + "bbox": [ + 130, + 325, + 479, + 361 + ], + "type": "text", + "content": " with slight augmentation to ensure a sufficient Field-of-View." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 380, + 469, + 392 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 380, + 469, + 392 + ], + "spans": [ + { + "bbox": [ + 130, + 380, + 469, + 392 + ], + "type": "text", + "content": "3.1 LiDAR Depthmap Projective Artifacts as Epipolar Occlusion" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 400, + 480, + 448 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 400, + 480, + 448 + ], + "spans": [ + { + "bbox": [ + 130, + 400, + 480, + 448 + ], + "type": "text", + "content": "Fig. 2 demonstrates the baseline distance between various LiDAR and RGB camera sensor sets. Fig. 2a shows that the background tree is solely visible to LiDAR. Intuitively, the background tree is occluded by the foreground pedestrian in a binocular vision system of the virtual LiDAR camera and RGB camera." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 449, + 480, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 449, + 480, + 485 + ], + "spans": [ + { + "bbox": [ + 130, + 449, + 480, + 485 + ], + "type": "text", + "content": "Virtual LiDAR Camera. LiDAR produces a point cloud " + }, + { + "bbox": [ + 130, + 449, + 480, + 485 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 130, + 449, + 480, + 485 + ], + "type": "text", + "content": " by emitting laser pulses into 3D space. Mathematically, we assume the LiDAR point cloud " + }, + { + "bbox": [ + 130, + 449, + 480, + 485 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 130, + 449, + 480, + 485 + ], + "type": "text", + "content": " as a set of 3D rays originating from the coordinate origin:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 201, + 494, + 480, + 509 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 201, + 494, + 480, + 509 + ], + "spans": [ + { + "bbox": [ + 201, + 494, + 480, + 509 + ], + "type": "interline_equation", + "content": "\\mathcal {V} = \\left\\{\\mathbf {v} _ {i} \\right\\} = \\left\\{\\mathbf {0} + \\bar {\\mathbf {r}} _ {i} \\cdot t _ {i} \\mid i \\in \\mathbb {Z}, t > 0, \\bar {\\mathbf {r}} _ {i} \\in \\mathbb {R} ^ {3} \\right\\}, \\tag {1}", + "image_path": "c48117f0cfdbbf30d096fbc377c375de45378c09708901d16a816444c19a560f.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 517, + 479, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 517, + 479, + 553 + ], + "spans": [ + { + "bbox": [ + 130, + 517, + 479, + 553 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 517, + 479, + 553 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_i" + }, + { + "bbox": [ + 130, + 517, + 479, + 553 + ], + "type": "text", + "content": " is a 3D point. " + }, + { + "bbox": [ + 130, + 517, + 479, + 553 + ], + "type": "inline_equation", + "content": "\\overline{\\mathbf{r}}_i" + }, + { + "bbox": [ + 130, + 517, + 479, + 553 + ], + "type": "text", + "content": " is a normalized direction vector, and " + }, + { + "bbox": [ + 130, + 517, + 479, + 553 + ], + "type": "inline_equation", + "content": "t_i" + }, + { + "bbox": [ + 130, + 517, + 479, + 553 + ], + "type": "text", + "content": " is the euclidean distance between the 3D point and LiDAR. If we position a virtual LiDAR camera with arbitrary resolution and intrinsic " + }, + { + "bbox": [ + 130, + 517, + 479, + 553 + ], + "type": "inline_equation", + "content": "\\mathbf{K}_l" + }, + { + "bbox": [ + 130, + 517, + 479, + 553 + ], + "type": "text", + "content": " at the LiDAR's location, then:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 573, + 479, + 597 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 573, + 479, + 597 + ], + "spans": [ + { + "bbox": [ + 130, + 573, + 479, + 597 + ], + "type": "text", + "content": "Lemma 1. The depthmap of the virtual LiDAR camera is free of projective artifacts, i.e., no two scanned points " + }, + { + "bbox": [ + 130, + 573, + 479, + 597 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_1" + }, + { + "bbox": [ + 130, + 573, + 479, + 597 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 573, + 479, + 597 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_2" + }, + { + "bbox": [ + 130, + 573, + 479, + 597 + ], + "type": "text", + "content": " overlap after projection." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 206, + 607, + 480, + 620 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 206, + 607, + 480, + 620 + ], + "spans": [ + { + "bbox": [ + 206, + 607, + 480, + 620 + ], + "type": "interline_equation", + "content": "\\forall \\mathbf {v} _ {1}, \\mathbf {v} _ {2} \\in \\mathcal {V}, \\| \\pi (\\mathbf {v} _ {1} \\mid \\mathbf {K} _ {l}) - \\pi (\\mathbf {v} _ {2} \\mid \\mathbf {K} _ {l}) \\| _ {2} \\neq 0, \\tag {2}", + "image_path": "6309c53e9e346b47df5b887c86e9128ea70f174980f99a73dae9baef92bbb0e9.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 130, + 629, + 479, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 479, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 479, + 665 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 629, + 479, + 665 + ], + "type": "inline_equation", + "content": "\\pi (\\cdot)" + }, + { + "bbox": [ + 130, + 629, + 479, + 665 + ], + "type": "text", + "content": " stands for the projection process. See proof in Supp. (Appendix A). We visualize the artifacts-free depthmap " + }, + { + "bbox": [ + 130, + 629, + 479, + 665 + ], + "type": "inline_equation", + "content": "\\mathbf{D}_l" + }, + { + "bbox": [ + 130, + 629, + 479, + 665 + ], + "type": "text", + "content": " from the virtual LiDAR camera in Fig. 4b." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 411, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 411, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 411, + 91, + 447, + 102 + ], + "type": "text", + "content": "RePLAY" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 137, + 114, + 486, + 194 + ], + "blocks": [ + { + "bbox": [ + 137, + 114, + 486, + 194 + ], + "lines": [ + { + "bbox": [ + 137, + 114, + 486, + 194 + ], + "spans": [ + { + "bbox": [ + 137, + 114, + 486, + 194 + ], + "type": "image", + "image_path": "d98fbaebe05c71db5e6e056a53358e127082fa01709ca3d72d3161c4252fb17c.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 198, + 482, + 264 + ], + "lines": [ + { + "bbox": [ + 130, + 198, + 482, + 264 + ], + "spans": [ + { + "bbox": [ + 130, + 198, + 482, + 264 + ], + "type": "text", + "content": "Fig. 5: Auto-Calibration. The LiDAR point cloud " + }, + { + "bbox": [ + 130, + 198, + 482, + 264 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 130, + 198, + 482, + 264 + ], + "type": "text", + "content": " in (a) is mapped to spherical coordinate " + }, + { + "bbox": [ + 130, + 198, + 482, + 264 + ], + "type": "inline_equation", + "content": "\\theta -\\phi" + }, + { + "bbox": [ + 130, + 198, + 482, + 264 + ], + "type": "text", + "content": " with Eq. (3). We use a binary matrix " + }, + { + "bbox": [ + 130, + 198, + 482, + 264 + ], + "type": "inline_equation", + "content": "\\mathbf{S}" + }, + { + "bbox": [ + 130, + 198, + 482, + 264 + ], + "type": "text", + "content": " to record the unique pixel locations " + }, + { + "bbox": [ + 130, + 198, + 482, + 264 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 130, + 198, + 482, + 264 + ], + "type": "text", + "content": " after rasterization. Plots (b) and (c) are matrix " + }, + { + "bbox": [ + 130, + 198, + 482, + 264 + ], + "type": "inline_equation", + "content": "\\mathbf{S}" + }, + { + "bbox": [ + 130, + 198, + 482, + 264 + ], + "type": "text", + "content": " before and after autocalibration where duplication (marked by arrows) is minimized with Eq. (5) loss " + }, + { + "bbox": [ + 130, + 198, + 482, + 264 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 130, + 198, + 482, + 264 + ], + "type": "text", + "content": ". Plots (d) and (e) contrast the virtual LiDAR depthmap " + }, + { + "bbox": [ + 130, + 198, + 482, + 264 + ], + "type": "inline_equation", + "content": "\\mathbf{D}_l" + }, + { + "bbox": [ + 130, + 198, + 482, + 264 + ], + "type": "text", + "content": " before and after autocalibration where Lemma 1 is fulfilled." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 279, + 479, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 279, + 479, + 350 + ], + "spans": [ + { + "bbox": [ + 130, + 279, + 479, + 350 + ], + "type": "text", + "content": "Auto-Calibration. In practice, Lemma 1 may not hold in a realistic LiDAR point cloud due to possible bias in the point cloud coordinate origin " + }, + { + "bbox": [ + 130, + 279, + 479, + 350 + ], + "type": "inline_equation", + "content": "\\mathbf{0}" + }, + { + "bbox": [ + 130, + 279, + 479, + 350 + ], + "type": "text", + "content": ". Fig. 5d, an example from the KITTI dataset shows persistent projective artifacts in the depthmap " + }, + { + "bbox": [ + 130, + 279, + 479, + 350 + ], + "type": "inline_equation", + "content": "\\mathbf{D}_l" + }, + { + "bbox": [ + 130, + 279, + 479, + 350 + ], + "type": "text", + "content": " when without any treatment. We perform auto-calibration to fulfill Lemma 1 by re-adjusting the point cloud origin. Fig. 4b and Fig. 5e show artifacts-free depthmap after auto-calibration." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 350, + 480, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 350, + 480, + 411 + ], + "spans": [ + { + "bbox": [ + 130, + 350, + 480, + 411 + ], + "type": "text", + "content": "Auto-calibration aims to resolve the artifacts in the virtual LiDAR camera depthmap " + }, + { + "bbox": [ + 130, + 350, + 480, + 411 + ], + "type": "inline_equation", + "content": "\\mathbf{D}_l" + }, + { + "bbox": [ + 130, + 350, + 480, + 411 + ], + "type": "text", + "content": ". We convert each LiDAR point cloud to the spherical coordinate system. This hypothesizes an omnidirectional image " + }, + { + "bbox": [ + 130, + 350, + 480, + 411 + ], + "type": "inline_equation", + "content": "\\mathbf{S}" + }, + { + "bbox": [ + 130, + 350, + 480, + 411 + ], + "type": "text", + "content": " of the LiDAR point cloud. Denote the origin to estimate as " + }, + { + "bbox": [ + 130, + 350, + 480, + 411 + ], + "type": "inline_equation", + "content": "\\mathbf{o}" + }, + { + "bbox": [ + 130, + 350, + 480, + 411 + ], + "type": "text", + "content": ". Set the new 3D point " + }, + { + "bbox": [ + 130, + 350, + 480, + 411 + ], + "type": "inline_equation", + "content": "\\mathbf{u} = \\mathbf{v} + \\mathbf{o}" + }, + { + "bbox": [ + 130, + 350, + 480, + 411 + ], + "type": "text", + "content": ", we have the polar angle " + }, + { + "bbox": [ + 130, + 350, + 480, + 411 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 130, + 350, + 480, + 411 + ], + "type": "text", + "content": " and azimuthal angle " + }, + { + "bbox": [ + 130, + 350, + 480, + 411 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 130, + 350, + 480, + 411 + ], + "type": "text", + "content": " after applying a rasterization operation " + }, + { + "bbox": [ + 130, + 350, + 480, + 411 + ], + "type": "inline_equation", + "content": "[\\cdot]:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 140, + 416, + 480, + 430 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 140, + 416, + 480, + 430 + ], + "spans": [ + { + "bbox": [ + 140, + 416, + 480, + 430 + ], + "type": "interline_equation", + "content": "\\theta = \\left[ \\arccos \\left(u _ {x} / \\left(u _ {x} ^ {2} + u _ {z} ^ {2}\\right)\\right) \\right], \\phi = \\left[ \\arccos \\left(u _ {y} / \\left(u _ {x} ^ {2} + u _ {y} ^ {2} + u _ {z} ^ {2}\\right)\\right) \\right], \\left[ \\theta \\phi \\right] = f (\\mathbf {v} + \\mathbf {o}). \\tag {3}", + "image_path": "e29844dcb41693dbdaeee81f3eac8934168c2cbbeb459d45b3f689e844b4c519.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 437, + 479, + 461 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 437, + 479, + 461 + ], + "spans": [ + { + "bbox": [ + 130, + 437, + 479, + 461 + ], + "type": "text", + "content": "Intuitively, function " + }, + { + "bbox": [ + 130, + 437, + 479, + 461 + ], + "type": "inline_equation", + "content": "f(\\cdot)" + }, + { + "bbox": [ + 130, + 437, + 479, + 461 + ], + "type": "text", + "content": " converts a 3D point " + }, + { + "bbox": [ + 130, + 437, + 479, + 461 + ], + "type": "inline_equation", + "content": "\\mathbf{u} \\in \\mathcal{V}" + }, + { + "bbox": [ + 130, + 437, + 479, + 461 + ], + "type": "text", + "content": " to a discrete 2D coordinate in " + }, + { + "bbox": [ + 130, + 437, + 479, + 461 + ], + "type": "inline_equation", + "content": "\\mathbf{S}" + }, + { + "bbox": [ + 130, + 437, + 479, + 461 + ], + "type": "text", + "content": ". We denote the set of unique coordinates as " + }, + { + "bbox": [ + 130, + 437, + 479, + 461 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 130, + 437, + 479, + 461 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 252, + 470, + 481, + 483 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 252, + 470, + 481, + 483 + ], + "spans": [ + { + "bbox": [ + 252, + 470, + 481, + 483 + ], + "type": "interline_equation", + "content": "\\mathcal {S} = \\left\\{f (\\mathbf {v} + \\mathbf {o}) \\mid \\mathbf {v} \\in \\mathcal {V} \\right\\}. \\tag {4}", + "image_path": "69bcd60cd8854275cf116d97361b7433a3880bf7fd8a3f4550c5af2f88e5b7bf.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 491, + 482, + 539 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 491, + 482, + 539 + ], + "spans": [ + { + "bbox": [ + 130, + 491, + 482, + 539 + ], + "type": "text", + "content": "We implement Eq. (4) with a binary matrix " + }, + { + "bbox": [ + 130, + 491, + 482, + 539 + ], + "type": "inline_equation", + "content": "\\mathbf{S}" + }, + { + "bbox": [ + 130, + 491, + 482, + 539 + ], + "type": "text", + "content": ". Its elements are 1 at each discrete coordinate and 0 otherwise. The size of " + }, + { + "bbox": [ + 130, + 491, + 482, + 539 + ], + "type": "inline_equation", + "content": "\\mathcal{S}" + }, + { + "bbox": [ + 130, + 491, + 482, + 539 + ], + "type": "text", + "content": " equals to summarizing the matrix " + }, + { + "bbox": [ + 130, + 491, + 482, + 539 + ], + "type": "inline_equation", + "content": "\\mathbf{S}" + }, + { + "bbox": [ + 130, + 491, + 482, + 539 + ], + "type": "text", + "content": ". Artifacts become duplication in set " + }, + { + "bbox": [ + 130, + 491, + 482, + 539 + ], + "type": "inline_equation", + "content": "\\mathcal{S}" + }, + { + "bbox": [ + 130, + 491, + 482, + 539 + ], + "type": "text", + "content": ". Removing artifacts is to minimize duplication. The loss " + }, + { + "bbox": [ + 130, + 491, + 482, + 539 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 130, + 491, + 482, + 539 + ], + "type": "text", + "content": " is then defined as:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 270, + 548, + 481, + 562 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 270, + 548, + 481, + 562 + ], + "spans": [ + { + "bbox": [ + 270, + 548, + 481, + 562 + ], + "type": "interline_equation", + "content": "L = \\| \\mathcal {V} \\| - \\| \\mathcal {S} \\|. \\tag {5}", + "image_path": "51477cbc0eacab107bbb2efc0aaf79881deb07c3ec743dbcf60b1ed532222f9a.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 570, + 480, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 480, + 593 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 480, + 593 + ], + "type": "text", + "content": "The loss " + }, + { + "bbox": [ + 130, + 570, + 480, + 593 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 130, + 570, + 480, + 593 + ], + "type": "text", + "content": " is not differentiable. In each iteration, we approximate its gradient by ablating origin " + }, + { + "bbox": [ + 130, + 570, + 480, + 593 + ], + "type": "inline_equation", + "content": "\\mathbf{o}" + }, + { + "bbox": [ + 130, + 570, + 480, + 593 + ], + "type": "text", + "content": " with a small deviation and perform coordinate descent." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 594, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 665 + ], + "type": "text", + "content": "Virtual LiDAR Camera and RGB Binocular System. In Fig. 4, we set up a binocular vision system, denoting the virtual LiDAR camera and RGB camera as the epipolar left and right camera respectively. Correspondingly, set their intrinsic as " + }, + { + "bbox": [ + 130, + 594, + 482, + 665 + ], + "type": "inline_equation", + "content": "\\mathbf{K}_l" + }, + { + "bbox": [ + 130, + 594, + 482, + 665 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 594, + 482, + 665 + ], + "type": "inline_equation", + "content": "\\mathbf{K}_r" + }, + { + "bbox": [ + 130, + 594, + 482, + 665 + ], + "type": "text", + "content": ", and projected depthmap as " + }, + { + "bbox": [ + 130, + 594, + 482, + 665 + ], + "type": "inline_equation", + "content": "\\mathbf{D}_l" + }, + { + "bbox": [ + 130, + 594, + 482, + 665 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 594, + 482, + 665 + ], + "type": "inline_equation", + "content": "\\mathbf{D}_r" + }, + { + "bbox": [ + 130, + 594, + 482, + 665 + ], + "type": "text", + "content": ". Thus the problem of removing the projective artifacts is translated to detecting the epipolar occlusion between the virtual LiDAR camera and RGB camera." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "type": "text", + "content": "S. Zhu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 114, + 290, + 149 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 290, + 149 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 290, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 290, + 149 + ], + "type": "image", + "image_path": "700d9a39658aced8eb8f08322e87f0c64546d4df343c2b682e6e549186f2c061.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 293, + 139, + 321, + 146 + ], + "lines": [ + { + "bbox": [ + 293, + 139, + 321, + 146 + ], + "spans": [ + { + "bbox": [ + 293, + 139, + 321, + 146 + ], + "type": "text", + "content": "Rotation" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 293, + 114, + 321, + 137 + ], + "blocks": [ + { + "bbox": [ + 293, + 114, + 321, + 137 + ], + "lines": [ + { + "bbox": [ + 293, + 114, + 321, + 137 + ], + "spans": [ + { + "bbox": [ + 293, + 114, + 321, + 137 + ], + "type": "image", + "image_path": "812966440980895d7fe174e0a8d7d4de5547b26ea9239a85dcd0e2f4c612aef9.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 324, + 115, + 419, + 148 + ], + "blocks": [ + { + "bbox": [ + 324, + 115, + 419, + 148 + ], + "lines": [ + { + "bbox": [ + 324, + 115, + 419, + 148 + ], + "spans": [ + { + "bbox": [ + 324, + 115, + 419, + 148 + ], + "type": "image", + "image_path": "9f24f8901b2622a3e8a412d866d44e0f1f584778dde342527c930495e1f5322c.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 158, + 482, + 191 + ], + "lines": [ + { + "bbox": [ + 131, + 158, + 482, + 191 + ], + "spans": [ + { + "bbox": [ + 131, + 158, + 482, + 191 + ], + "type": "text", + "content": "Fig. 6: Epipolar Occlusion with Pure Rotation Movement. We apply a rotation transformation along the " + }, + { + "bbox": [ + 131, + 158, + 482, + 191 + ], + "type": "inline_equation", + "content": "x-z" + }, + { + "bbox": [ + 131, + 158, + 482, + 191 + ], + "type": "text", + "content": " axis in the camera coordinate system. As indicated by Lemma 2, a pure rotation movement does not lead to projective artifacts." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 203, + 479, + 275 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 203, + 479, + 275 + ], + "spans": [ + { + "bbox": [ + 131, + 203, + 479, + 275 + ], + "type": "text", + "content": "Projective Artifacts are Epipolar Occlusion. We ensure that our virtual LiDAR has a larger field of view than the RGB camera by adjusting the resolution of the virtual LiDAR camera. Denote the relative movement between the LiDAR and RGB camera as " + }, + { + "bbox": [ + 131, + 203, + 479, + 275 + ], + "type": "inline_equation", + "content": "\\mathbf{P} = [\\mathbf{R}\\mathbf{t}]" + }, + { + "bbox": [ + 131, + 203, + 479, + 275 + ], + "type": "text", + "content": ", where the auto-calibrated center o is integrated into t. Denote the projected pixels on epipolar left (LiDAR) and epipolar right (RGB) as p and q, and their pixel depths as " + }, + { + "bbox": [ + 131, + 203, + 479, + 275 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 131, + 203, + 479, + 275 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 203, + 479, + 275 + ], + "type": "inline_equation", + "content": "d^{\\prime}" + }, + { + "bbox": [ + 131, + 203, + 479, + 275 + ], + "type": "text", + "content": ". We have:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 222, + 284, + 481, + 298 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 222, + 284, + 481, + 298 + ], + "spans": [ + { + "bbox": [ + 222, + 284, + 481, + 298 + ], + "type": "interline_equation", + "content": "\\mathbf {p} = \\pi (\\mathbf {v} \\mid \\mathbf {K} _ {l}), \\quad \\mathbf {q} = \\pi (\\mathbf {R v} + \\mathbf {t} \\mid \\mathbf {K} _ {r}). \\tag {6}", + "image_path": "c2a5725ad6a897c07f75b33c159208f1decd96cd87e8a10d1b0d070725ec575c.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 305, + 427, + 318 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 305, + 427, + 318 + ], + "spans": [ + { + "bbox": [ + 132, + 305, + 427, + 318 + ], + "type": "text", + "content": "We then formally define the epipolar occlusion " + }, + { + "bbox": [ + 132, + 305, + 427, + 318 + ], + "type": "inline_equation", + "content": "\\mathcal{O}" + }, + { + "bbox": [ + 132, + 305, + 427, + 318 + ], + "type": "text", + "content": " on RGB view as:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 189, + 327, + 481, + 341 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 189, + 327, + 481, + 341 + ], + "spans": [ + { + "bbox": [ + 189, + 327, + 481, + 341 + ], + "type": "interline_equation", + "content": "\\mathcal {O} = \\left\\{\\mathbf {q} _ {1} \\mid \\forall \\mathbf {q} _ {1}, \\mathbf {q} _ {2} \\in \\mathcal {Q}, d _ {\\mathbf {q} _ {1}} ^ {\\prime} > d _ {\\mathbf {q} _ {2}} ^ {\\prime}, \\| \\mathbf {q} _ {1} - \\mathbf {q} _ {2} \\| _ {2} = 0 \\right\\}. \\tag {7}", + "image_path": "2d1714336f9f71e487656ab48c4ada716c94eab3d199dcde5c1bd6fa156c5140.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 131, + 348, + 479, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 348, + 479, + 384 + ], + "spans": [ + { + "bbox": [ + 131, + 348, + 479, + 384 + ], + "type": "text", + "content": "Eq. (7) simply suggests if two pixels in RGB camera overlap with each other, then the one " + }, + { + "bbox": [ + 131, + 348, + 479, + 384 + ], + "type": "inline_equation", + "content": "\\mathbf{q}_1" + }, + { + "bbox": [ + 131, + 348, + 479, + 384 + ], + "type": "text", + "content": " with larger depth is occluded by the one with smaller depth " + }, + { + "bbox": [ + 131, + 348, + 479, + 384 + ], + "type": "inline_equation", + "content": "\\mathbf{q}_2" + }, + { + "bbox": [ + 131, + 348, + 479, + 384 + ], + "type": "text", + "content": ". Occlusion detection finds all occluded pixels " + }, + { + "bbox": [ + 131, + 348, + 479, + 384 + ], + "type": "inline_equation", + "content": "\\mathbf{q}_1" + }, + { + "bbox": [ + 131, + 348, + 479, + 384 + ], + "type": "text", + "content": ", i.e., the set " + }, + { + "bbox": [ + 131, + 348, + 479, + 384 + ], + "type": "inline_equation", + "content": "\\mathcal{O}" + }, + { + "bbox": [ + 131, + 348, + 479, + 384 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 131, + 401, + 419, + 414 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 401, + 419, + 414 + ], + "spans": [ + { + "bbox": [ + 131, + 401, + 419, + 414 + ], + "type": "text", + "content": "3.2 Epipolar Occlusion under Rotation and Translation" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 131, + 421, + 479, + 458 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 421, + 479, + 458 + ], + "spans": [ + { + "bbox": [ + 131, + 421, + 479, + 458 + ], + "type": "text", + "content": "Suppose the relative camera motion among the Fig. 4a binocular system is an arbitrary camera pose " + }, + { + "bbox": [ + 131, + 421, + 479, + 458 + ], + "type": "inline_equation", + "content": "\\mathbf{P} = [\\mathbf{R}\\mathbf{t}]" + }, + { + "bbox": [ + 131, + 421, + 479, + 458 + ], + "type": "text", + "content": ". Set " + }, + { + "bbox": [ + 131, + 421, + 479, + 458 + ], + "type": "inline_equation", + "content": "\\mathbf{I}" + }, + { + "bbox": [ + 131, + 421, + 479, + 458 + ], + "type": "text", + "content": " as the identity matrix. RePLAy decomposes " + }, + { + "bbox": [ + 131, + 421, + 479, + 458 + ], + "type": "inline_equation", + "content": "\\mathbf{P}" + }, + { + "bbox": [ + 131, + 421, + 479, + 458 + ], + "type": "text", + "content": " into a sequential pure rotation and pure translation movement:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 222, + 468, + 481, + 483 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 222, + 468, + 481, + 483 + ], + "spans": [ + { + "bbox": [ + 222, + 468, + 481, + 483 + ], + "type": "interline_equation", + "content": "\\mathbf {P} _ {1} = \\left[ \\begin{array}{l l} \\mathbf {R} & \\mathbf {0} \\end{array} \\right], \\mathbf {P} _ {2} = \\left[ \\begin{array}{l l} \\mathbf {I} & \\mathbf {t} \\end{array} \\right], \\mathbf {P} = \\mathbf {P} _ {2} \\mathbf {P} _ {1}. \\tag {8}", + "image_path": "898547236200323ea9f9339b25797f9b2a6638d9236b8da23549e4894eb2b8ec.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 131, + 487, + 479, + 546 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 487, + 479, + 546 + ], + "spans": [ + { + "bbox": [ + 131, + 487, + 479, + 546 + ], + "type": "text", + "content": "We outline pure rotation movement in Lemma 2 and Fig. 6. The pure translation movement is illustrated in Lemma 3 and Fig. 7. Without loss of generality, we assume the epipolar left and right cameras share the same intrinsic, i.e., " + }, + { + "bbox": [ + 131, + 487, + 479, + 546 + ], + "type": "inline_equation", + "content": "\\mathbf{K}_l = \\mathbf{K}_r" + }, + { + "bbox": [ + 131, + 487, + 479, + 546 + ], + "type": "text", + "content": ". When with different intrinsics, we apply a resize and crop operation defined by " + }, + { + "bbox": [ + 131, + 487, + 479, + 546 + ], + "type": "inline_equation", + "content": "\\mathbf{A} = \\mathbf{K}_l\\mathbf{K}_r^{-1}" + }, + { + "bbox": [ + 131, + 487, + 479, + 546 + ], + "type": "text", + "content": " to unify the intrinsic." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 131, + 555, + 461, + 567 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 555, + 461, + 567 + ], + "spans": [ + { + "bbox": [ + 131, + 555, + 461, + 567 + ], + "type": "text", + "content": "Lemma 2. Pure rotation movement does not cause epipolar occlusion." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 131, + 575, + 479, + 623 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 575, + 479, + 623 + ], + "spans": [ + { + "bbox": [ + 131, + 575, + 479, + 623 + ], + "type": "text", + "content": "We know from epipolar geometry [25] that pure rotation induces homography between two views. In other words, there is a one-to-one mapping between pixels before and after rotation. See Supp for more information. Fig. 6 illustrates Lemma 2 with a visual example." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 131, + 631, + 481, + 654 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 631, + 481, + 654 + ], + "spans": [ + { + "bbox": [ + 131, + 631, + 481, + 654 + ], + "type": "text", + "content": "Lemma 3. Pure translation movement causes epipolar occlusion along the direction of the epipolar line." + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 411, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 411, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 411, + 91, + 447, + 102 + ], + "type": "text", + "content": "RePLAY" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 114, + 482, + 273 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 482, + 273 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 482, + 273 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 482, + 273 + ], + "type": "image", + "image_path": "9a4a4358b3ec92f4a432e9d33a295893771ff5265b38ee891a5d65c36701a703.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 280, + 482, + 380 + ], + "lines": [ + { + "bbox": [ + 130, + 280, + 482, + 380 + ], + "spans": [ + { + "bbox": [ + 130, + 280, + 482, + 380 + ], + "type": "text", + "content": "Fig. 7: Epipoplar Occlusion with Pure Translation Movement. A background pixel " + }, + { + "bbox": [ + 130, + 280, + 482, + 380 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_1" + }, + { + "bbox": [ + 130, + 280, + 482, + 380 + ], + "type": "text", + "content": " is gradually occluded by a foreground pixel " + }, + { + "bbox": [ + 130, + 280, + 482, + 380 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_2" + }, + { + "bbox": [ + 130, + 280, + 482, + 380 + ], + "type": "text", + "content": " under pure translation. We formulate three binocular systems at different baseline distances. For simplicity, we set the depth of " + }, + { + "bbox": [ + 130, + 280, + 482, + 380 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_1" + }, + { + "bbox": [ + 130, + 280, + 482, + 380 + ], + "type": "text", + "content": " to infinity, making its 2D location static. Direction vector " + }, + { + "bbox": [ + 130, + 280, + 482, + 380 + ], + "type": "inline_equation", + "content": "\\mathbf{n}" + }, + { + "bbox": [ + 130, + 280, + 482, + 380 + ], + "type": "text", + "content": " is set to point from " + }, + { + "bbox": [ + 130, + 280, + 482, + 380 + ], + "type": "inline_equation", + "content": "\\mathbf{p}" + }, + { + "bbox": [ + 130, + 280, + 482, + 380 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 130, + 280, + 482, + 380 + ], + "type": "inline_equation", + "content": "\\mathbf{q}" + }, + { + "bbox": [ + 130, + 280, + 482, + 380 + ], + "type": "text", + "content": ". The scalar " + }, + { + "bbox": [ + 130, + 280, + 482, + 380 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 280, + 482, + 380 + ], + "type": "text", + "content": " is marked by blue line segment. From Lemma 3, (a) The pixel " + }, + { + "bbox": [ + 130, + 280, + 482, + 380 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_1" + }, + { + "bbox": [ + 130, + 280, + 482, + 380 + ], + "type": "text", + "content": " is only occluded by pixel " + }, + { + "bbox": [ + 130, + 280, + 482, + 380 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_2" + }, + { + "bbox": [ + 130, + 280, + 482, + 380 + ], + "type": "text", + "content": " at the reverse direction of epipolar line (red line segment). (b) " + }, + { + "bbox": [ + 130, + 280, + 482, + 380 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_1" + }, + { + "bbox": [ + 130, + 280, + 482, + 380 + ], + "type": "text", + "content": " is not occluded. (c) " + }, + { + "bbox": [ + 130, + 280, + 482, + 380 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_1" + }, + { + "bbox": [ + 130, + 280, + 482, + 380 + ], + "type": "text", + "content": " is occluded following Eq. (7). (d) " + }, + { + "bbox": [ + 130, + 280, + 482, + 380 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_1" + }, + { + "bbox": [ + 130, + 280, + 482, + 380 + ], + "type": "text", + "content": " is considered occluded following Eq. (17). Intuitively, the occluded pixel " + }, + { + "bbox": [ + 130, + 280, + 482, + 380 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_1" + }, + { + "bbox": [ + 130, + 280, + 482, + 380 + ], + "type": "text", + "content": " (in a) is \"caught up\" (in c) and \"surpassed\" (in d) by " + }, + { + "bbox": [ + 130, + 280, + 482, + 380 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_2" + }, + { + "bbox": [ + 130, + 280, + 482, + 380 + ], + "type": "text", + "content": " while moving along the epipolar line." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 393, + 482, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 393, + 482, + 441 + ], + "spans": [ + { + "bbox": [ + 130, + 393, + 482, + 441 + ], + "type": "text", + "content": "Proof. As pure rotation does not cause epipolar occlusion, we apply the translation movement over the artifacts-free depthmap after the pure rotation. With slight abuse of the notation, we set the pixels after pure rotation movement as " + }, + { + "bbox": [ + 130, + 393, + 482, + 441 + ], + "type": "inline_equation", + "content": "\\mathbf{p}" + }, + { + "bbox": [ + 130, + 393, + 482, + 441 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 393, + 482, + 441 + ], + "type": "inline_equation", + "content": "\\mathbf{q}" + }, + { + "bbox": [ + 130, + 393, + 482, + 441 + ], + "type": "text", + "content": ". The epipolar line " + }, + { + "bbox": [ + 130, + 393, + 482, + 441 + ], + "type": "inline_equation", + "content": "\\mathbf{l_p}" + }, + { + "bbox": [ + 130, + 393, + 482, + 441 + ], + "type": "text", + "content": " on the right image from the left image pixel " + }, + { + "bbox": [ + 130, + 393, + 482, + 441 + ], + "type": "inline_equation", + "content": "\\mathbf{p}" + }, + { + "bbox": [ + 130, + 393, + 482, + 441 + ], + "type": "text", + "content": " is:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 225, + 448, + 480, + 462 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 225, + 448, + 480, + 462 + ], + "spans": [ + { + "bbox": [ + 225, + 448, + 480, + 462 + ], + "type": "interline_equation", + "content": "\\mathbf {l} _ {\\mathbf {p}} = \\mathbf {K} ^ {- \\top} \\left[ \\mathbf {t} _ {\\times} \\right] \\mathbf {R} \\mathbf {K} ^ {- 1} \\mathbf {p} = \\mathbf {K} ^ {- \\top} \\left[ \\mathbf {t} _ {\\times} \\right] \\mathbf {K} ^ {- 1} \\mathbf {p}, \\tag {9}", + "image_path": "9ed3624e2bf0dc33d3e80a80738c86f00fa362f3ab623d39794202428f61625b.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 468, + 482, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 468, + 482, + 506 + ], + "spans": [ + { + "bbox": [ + 130, + 468, + 482, + 506 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 468, + 482, + 506 + ], + "type": "inline_equation", + "content": "[\\mathbf{t}_{\\times}]" + }, + { + "bbox": [ + 130, + 468, + 482, + 506 + ], + "type": "text", + "content": " is the cross product in matrix form [71]. Due to a pure translation movement, the rotation " + }, + { + "bbox": [ + 130, + 468, + 482, + 506 + ], + "type": "inline_equation", + "content": "\\mathbf{R}" + }, + { + "bbox": [ + 130, + 468, + 482, + 506 + ], + "type": "text", + "content": " is removed as being an identity matrix. Further, the left and right epipolar lines overlap with each other in pure translation:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 214, + 512, + 480, + 526 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 214, + 512, + 480, + 526 + ], + "spans": [ + { + "bbox": [ + 214, + 512, + 480, + 526 + ], + "type": "interline_equation", + "content": "\\mathbf {l} _ {\\mathbf {q}} = - \\mathbf {K} ^ {- \\top} \\left[ \\mathbf {t} _ {\\times} \\right] \\mathbf {K} ^ {- 1} \\mathbf {q}, \\mathbf {p} ^ {\\top} \\mathbf {l} _ {\\mathbf {q}} = 0, \\mathbf {p} ^ {\\top} \\mathbf {l} _ {\\mathbf {p}} = 0. \\tag {10}", + "image_path": "06f5354f69428cc204701488390b11c0d1ae68400c9452e0b27da357cde2640c.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 533, + 395, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 533, + 395, + 545 + ], + "spans": [ + { + "bbox": [ + 132, + 533, + 395, + 545 + ], + "type": "text", + "content": "Also shown in Fig. 7, pixel " + }, + { + "bbox": [ + 132, + 533, + 395, + 545 + ], + "type": "inline_equation", + "content": "\\mathbf{q}" + }, + { + "bbox": [ + 132, + 533, + 395, + 545 + ], + "type": "text", + "content": " is hence a function of pixel " + }, + { + "bbox": [ + 132, + 533, + 395, + 545 + ], + "type": "inline_equation", + "content": "\\mathbf{p}" + }, + { + "bbox": [ + 132, + 533, + 395, + 545 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 149, + 552, + 480, + 572 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 552, + 480, + 572 + ], + "spans": [ + { + "bbox": [ + 149, + 552, + 480, + 572 + ], + "type": "interline_equation", + "content": "\\mathbf {q} = \\mathbf {p} + t \\cdot \\mathbf {n}, t = \\| \\mathbf {q} - \\mathbf {p} \\| _ {2} = \\mathbf {n} ^ {\\intercal} (\\mathbf {q} - \\mathbf {p}), \\mathbf {n} = (\\mathbf {q} - \\mathbf {p}) / \\sqrt {\\| \\mathbf {q} - \\mathbf {p} \\| _ {2} ^ {2}}, \\tag {11}", + "image_path": "1dcd2543d7388705635cf6a653a46c0fa8d478aefc591af75ad9799883194c97.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 578, + 482, + 626 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 578, + 482, + 626 + ], + "spans": [ + { + "bbox": [ + 130, + 578, + 482, + 626 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 578, + 482, + 626 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 578, + 482, + 626 + ], + "type": "text", + "content": " is the moving distance, similar to the disparity in stereo vision. We define the vector " + }, + { + "bbox": [ + 130, + 578, + 482, + 626 + ], + "type": "inline_equation", + "content": "\\mathbf{n}" + }, + { + "bbox": [ + 130, + 578, + 482, + 626 + ], + "type": "text", + "content": " pointed from " + }, + { + "bbox": [ + 130, + 578, + 482, + 626 + ], + "type": "inline_equation", + "content": "\\mathbf{p}" + }, + { + "bbox": [ + 130, + 578, + 482, + 626 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 130, + 578, + 482, + 626 + ], + "type": "inline_equation", + "content": "\\mathbf{q}" + }, + { + "bbox": [ + 130, + 578, + 482, + 626 + ], + "type": "text", + "content": " as the direction of the epipolar line " + }, + { + "bbox": [ + 130, + 578, + 482, + 626 + ], + "type": "inline_equation", + "content": "\\mathbf{l}" + }, + { + "bbox": [ + 130, + 578, + 482, + 626 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 130, + 578, + 482, + 626 + ], + "type": "inline_equation", + "content": "\\mathbf{p}" + }, + { + "bbox": [ + 130, + 578, + 482, + 626 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 578, + 482, + 626 + ], + "type": "inline_equation", + "content": "\\mathbf{q}" + }, + { + "bbox": [ + 130, + 578, + 482, + 626 + ], + "type": "text", + "content": " are related by the equation " + }, + { + "bbox": [ + 130, + 578, + 482, + 626 + ], + "type": "inline_equation", + "content": "d^{\\prime}\\mathbf{q} = d\\mathbf{K}_{r}\\mathbf{R}\\mathbf{K}_{l}^{-1}\\mathbf{p} + \\mathbf{K}_{r}\\mathbf{t}" + }, + { + "bbox": [ + 130, + 578, + 482, + 626 + ], + "type": "text", + "content": ". Expanding under pure translation movement gives us:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 631, + 481, + 666 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 631, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 631, + 481, + 666 + ], + "type": "interline_equation", + "content": "t ^ {2} = \\left(p _ {x} - q _ {x}\\right) ^ {2} + \\left(p _ {y} - q _ {y}\\right) ^ {2} = \\frac {\\left(x + z \\cdot q _ {x}\\right) ^ {2} + \\left(y + z \\cdot q _ {y}\\right) ^ {2}}{\\left(d ^ {\\prime} - z\\right) ^ {2}} = \\frac {\\left(x + z \\cdot q _ {x}\\right) ^ {2} + \\left(y + z \\cdot q _ {y}\\right) ^ {2}}{d ^ {2}}. \\tag {12}", + "image_path": "8963f08e5f4b156ef1a264875e44fc8cbdf6d99900c02fcd3affe2e725a9bf07.jpg" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 165, + 91, + 217, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 165, + 91, + 217, + 101 + ], + "spans": [ + { + "bbox": [ + 165, + 91, + 217, + 101 + ], + "type": "text", + "content": "S. Zhu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 138, + 114, + 307, + 154 + ], + "blocks": [ + { + "bbox": [ + 138, + 114, + 307, + 154 + ], + "lines": [ + { + "bbox": [ + 138, + 114, + 307, + 154 + ], + "spans": [ + { + "bbox": [ + 138, + 114, + 307, + 154 + ], + "type": "image", + "image_path": "3d39d62921456f7d1b58e19c6cfada73626e89c1d5bded49efc2025431470b30.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 161, + 156, + 280, + 165 + ], + "lines": [ + { + "bbox": [ + 161, + 156, + 280, + 165 + ], + "spans": [ + { + "bbox": [ + 161, + 156, + 280, + 165 + ], + "type": "text", + "content": "(a) Virtual LiDAR Depthmap " + }, + { + "bbox": [ + 161, + 156, + 280, + 165 + ], + "type": "inline_equation", + "content": "\\mathbf{D}_l" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 312, + 114, + 480, + 155 + ], + "blocks": [ + { + "bbox": [ + 312, + 114, + 480, + 155 + ], + "lines": [ + { + "bbox": [ + 312, + 114, + 480, + 155 + ], + "spans": [ + { + "bbox": [ + 312, + 114, + 480, + 155 + ], + "type": "image", + "image_path": "a4be020daa3bbc26cece9d94268f9f2879bc3a6eccb4f714eb5f2b0049a9348a.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 318, + 156, + 473, + 166 + ], + "lines": [ + { + "bbox": [ + 318, + 156, + 473, + 166 + ], + "spans": [ + { + "bbox": [ + 318, + 156, + 473, + 166 + ], + "type": "text", + "content": "(b) Densified Virtual LiDAR Depthmap " + }, + { + "bbox": [ + 318, + 156, + 473, + 166 + ], + "type": "inline_equation", + "content": "\\overline{\\mathbf{D}}_l" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 131, + 172, + 480, + 205 + ], + "lines": [ + { + "bbox": [ + 131, + 172, + 480, + 205 + ], + "spans": [ + { + "bbox": [ + 131, + 172, + 480, + 205 + ], + "type": "text", + "content": "Fig. 8: Per-pixel Occlusion Examination. To overcome LiDAR sparsity, for a query pixel, we verify inspection pixels along epipolar line " + }, + { + "bbox": [ + 131, + 172, + 480, + 205 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathbf{p}_1}" + }, + { + "bbox": [ + 131, + 172, + 480, + 205 + ], + "type": "text", + "content": " following Eq. (17) to determine occlusion. Depth " + }, + { + "bbox": [ + 131, + 172, + 480, + 205 + ], + "type": "inline_equation", + "content": "\\overline{\\mathbf{D}}_l" + }, + { + "bbox": [ + 131, + 172, + 480, + 205 + ], + "type": "text", + "content": " is densified with nearest-neighbor interpolation." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 220, + 479, + 293 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 220, + 479, + 293 + ], + "spans": [ + { + "bbox": [ + 131, + 220, + 479, + 293 + ], + "type": "text", + "content": "Please see the derivation in Supp. (Appendix B). Note, we assume pixels " + }, + { + "bbox": [ + 131, + 220, + 479, + 293 + ], + "type": "inline_equation", + "content": "\\mathbf{p}" + }, + { + "bbox": [ + 131, + 220, + 479, + 293 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 220, + 479, + 293 + ], + "type": "inline_equation", + "content": "\\mathbf{q}" + }, + { + "bbox": [ + 131, + 220, + 479, + 293 + ], + "type": "text", + "content": " are respectively visible to the left and right camera. This suggests " + }, + { + "bbox": [ + 131, + 220, + 479, + 293 + ], + "type": "inline_equation", + "content": "d > 0" + }, + { + "bbox": [ + 131, + 220, + 479, + 293 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 220, + 479, + 293 + ], + "type": "inline_equation", + "content": "d = d' - z > 0" + }, + { + "bbox": [ + 131, + 220, + 479, + 293 + ], + "type": "text", + "content": ". Eq. (12) suggests moving distance " + }, + { + "bbox": [ + 131, + 220, + 479, + 293 + ], + "type": "inline_equation", + "content": "t^2" + }, + { + "bbox": [ + 131, + 220, + 479, + 293 + ], + "type": "text", + "content": " monotonously decreases as the projected depth " + }, + { + "bbox": [ + 131, + 220, + 479, + 293 + ], + "type": "inline_equation", + "content": "d'" + }, + { + "bbox": [ + 131, + 220, + 479, + 293 + ], + "type": "text", + "content": " increases. It fits the intuition where a foreground object has a larger movement than a background object. From Eq. (7), when with occlusion, we have " + }, + { + "bbox": [ + 131, + 220, + 479, + 293 + ], + "type": "inline_equation", + "content": "\\mathbf{q}_1 = \\mathbf{q}_2" + }, + { + "bbox": [ + 131, + 220, + 479, + 293 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 131, + 220, + 479, + 293 + ], + "type": "inline_equation", + "content": "d_{\\mathbf{q}_1} < d_{\\mathbf{q}_2}" + }, + { + "bbox": [ + 131, + 220, + 479, + 293 + ], + "type": "text", + "content": ". Combine Eq. (11) and Eq. (12):" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 179, + 300, + 480, + 314 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 179, + 300, + 480, + 314 + ], + "spans": [ + { + "bbox": [ + 179, + 300, + 480, + 314 + ], + "type": "interline_equation", + "content": "\\mathbf {q} _ {1} = \\mathbf {q} _ {2}, \\text {a n d} d _ {\\mathbf {q} _ {1}} < d _ {\\mathbf {q} _ {2}} \\longrightarrow t _ {1} ^ {2} > t _ {2} ^ {2} \\longrightarrow \\mathbf {n} ^ {\\intercal} (\\mathbf {p} _ {1} - \\mathbf {p} _ {2}) < 0. \\tag {13}", + "image_path": "493d1e209fc1b41a3eadd0abe26dcfb5e8ff818d40206963c39272cb23244445.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 321, + 479, + 370 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 321, + 479, + 370 + ], + "spans": [ + { + "bbox": [ + 131, + 321, + 479, + 370 + ], + "type": "text", + "content": "From Eq. (13), if pixel " + }, + { + "bbox": [ + 131, + 321, + 479, + 370 + ], + "type": "inline_equation", + "content": "\\mathbf{q}_1" + }, + { + "bbox": [ + 131, + 321, + 479, + 370 + ], + "type": "text", + "content": " is occluded by " + }, + { + "bbox": [ + 131, + 321, + 479, + 370 + ], + "type": "inline_equation", + "content": "\\mathbf{q}_2" + }, + { + "bbox": [ + 131, + 321, + 479, + 370 + ], + "type": "text", + "content": " in the right view, pixel " + }, + { + "bbox": [ + 131, + 321, + 479, + 370 + ], + "type": "inline_equation", + "content": "\\mathbf{q}_2" + }, + { + "bbox": [ + 131, + 321, + 479, + 370 + ], + "type": "text", + "content": " must move a longer distance along the epipolar line than " + }, + { + "bbox": [ + 131, + 321, + 479, + 370 + ], + "type": "inline_equation", + "content": "\\mathbf{q}_1" + }, + { + "bbox": [ + 131, + 321, + 479, + 370 + ], + "type": "text", + "content": ". This suggests, in the left view, " + }, + { + "bbox": [ + 131, + 321, + 479, + 370 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_2" + }, + { + "bbox": [ + 131, + 321, + 479, + 370 + ], + "type": "text", + "content": " resides at the \"back\" of " + }, + { + "bbox": [ + 131, + 321, + 479, + 370 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_1" + }, + { + "bbox": [ + 131, + 321, + 479, + 370 + ], + "type": "text", + "content": " w.r.t. the moving direction. In other words, " + }, + { + "bbox": [ + 131, + 321, + 479, + 370 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_1" + }, + { + "bbox": [ + 131, + 321, + 479, + 370 + ], + "type": "text", + "content": " is only occluded by other pixels along the reverse direction of epipolar line." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 376, + 479, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 376, + 479, + 413 + ], + "spans": [ + { + "bbox": [ + 131, + 376, + 479, + 413 + ], + "type": "text", + "content": "Epipolar Occlusion Judgement. Combine Lemma 2 and 3, a pixel " + }, + { + "bbox": [ + 131, + 376, + 479, + 413 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_1" + }, + { + "bbox": [ + 131, + 376, + 479, + 413 + ], + "type": "text", + "content": " in LiDAR view is occluded by other pixels " + }, + { + "bbox": [ + 131, + 376, + 479, + 413 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_2" + }, + { + "bbox": [ + 131, + 376, + 479, + 413 + ], + "type": "text", + "content": " along the reverse of epipolar line direction. For pixel " + }, + { + "bbox": [ + 131, + 376, + 479, + 413 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_1" + }, + { + "bbox": [ + 131, + 376, + 479, + 413 + ], + "type": "text", + "content": ", the set of inspection pixels is:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 215, + 421, + 480, + 434 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 215, + 421, + 480, + 434 + ], + "spans": [ + { + "bbox": [ + 215, + 421, + 480, + 434 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathbf {p} _ {1}} = \\left\\{\\mathbf {p} _ {2} \\mid \\mathbf {p} _ {2} = \\mathbf {p} _ {1} - \\mathbf {n} \\cdot k \\cdot \\Delta t, k \\in \\mathbb {Z} \\right\\}. \\tag {14}", + "image_path": "d072650a34662fed14c85251f618b202ee9a99afbd81ed5f23a50a287599278f.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 131, + 442, + 447, + 455 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 442, + 447, + 455 + ], + "spans": [ + { + "bbox": [ + 131, + 442, + 447, + 455 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 131, + 442, + 447, + 455 + ], + "type": "text", + "content": " is an integer. At occlusion, we have " + }, + { + "bbox": [ + 131, + 442, + 447, + 455 + ], + "type": "inline_equation", + "content": "\\mathbf{q}_1 = \\mathbf{q}_2" + }, + { + "bbox": [ + 131, + 442, + 447, + 455 + ], + "type": "text", + "content": ", combined with Eq. (11):" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 155, + 463, + 480, + 475 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 155, + 463, + 480, + 475 + ], + "spans": [ + { + "bbox": [ + 155, + 463, + 480, + 475 + ], + "type": "interline_equation", + "content": "\\mathbf {q} _ {1} = \\mathbf {q} _ {2} \\longrightarrow \\mathbf {p} _ {1} + t _ {1} \\mathbf {n} = \\mathbf {p} _ {2} + t _ {2} \\mathbf {n} \\longrightarrow \\mathbf {n} ^ {\\top} \\left(\\mathbf {p} _ {1} - \\mathbf {p} _ {2}\\right) + t _ {1} - t _ {2} = 0. \\tag {15}", + "image_path": "7cf29df3d78a237a8db6c4ec200594c336ed539b3cb1ad13107ba8a71bbc330b.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 131, + 483, + 479, + 532 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 483, + 479, + 532 + ], + "spans": [ + { + "bbox": [ + 131, + 483, + 479, + 532 + ], + "type": "text", + "content": "We use Fig. 7 to explain Eq. (15). In Fig. 7a, the pixel " + }, + { + "bbox": [ + 131, + 483, + 479, + 532 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_1" + }, + { + "bbox": [ + 131, + 483, + 479, + 532 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 483, + 479, + 532 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_2" + }, + { + "bbox": [ + 131, + 483, + 479, + 532 + ], + "type": "text", + "content": " has an initial distance " + }, + { + "bbox": [ + 131, + 483, + 479, + 532 + ], + "type": "inline_equation", + "content": "\\mathbf{n}^{\\top}(\\mathbf{p}_1 - \\mathbf{p}_2)" + }, + { + "bbox": [ + 131, + 483, + 479, + 532 + ], + "type": "text", + "content": ". In Fig. 7c, after translation movement, each pixel moves a distance " + }, + { + "bbox": [ + 131, + 483, + 479, + 532 + ], + "type": "inline_equation", + "content": "t_1" + }, + { + "bbox": [ + 131, + 483, + 479, + 532 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 483, + 479, + 532 + ], + "type": "inline_equation", + "content": "t_2" + }, + { + "bbox": [ + 131, + 483, + 479, + 532 + ], + "type": "text", + "content": " along the epipolar line. Occlusion happens when pixel " + }, + { + "bbox": [ + 131, + 483, + 479, + 532 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_2" + }, + { + "bbox": [ + 131, + 483, + 479, + 532 + ], + "type": "text", + "content": " overlaps pixel " + }, + { + "bbox": [ + 131, + 483, + 479, + 532 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_1" + }, + { + "bbox": [ + 131, + 483, + 479, + 532 + ], + "type": "text", + "content": ". We summarize Eq. (15) into a function " + }, + { + "bbox": [ + 131, + 483, + 479, + 532 + ], + "type": "inline_equation", + "content": "g(\\cdot)" + }, + { + "bbox": [ + 131, + 483, + 479, + 532 + ], + "type": "text", + "content": " as:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 231, + 540, + 480, + 554 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 231, + 540, + 480, + 554 + ], + "spans": [ + { + "bbox": [ + 231, + 540, + 480, + 554 + ], + "type": "interline_equation", + "content": "\\lambda \\leq g \\left(\\mathbf {p} _ {1}, \\mathbf {p} _ {2}, d _ {\\mathbf {p} _ {1}}, d _ {\\mathbf {p} _ {2}} \\mid \\mathbf {K}, \\mathbf {P}\\right) \\leq 0, \\tag {16}", + "image_path": "500c7d9c2ba456ce5ac4a914f69892d0899562b6299b008fc81acf41d8a206ae.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 131, + 561, + 479, + 644 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 561, + 479, + 644 + ], + "spans": [ + { + "bbox": [ + 131, + 561, + 479, + 644 + ], + "type": "text", + "content": "where the distance " + }, + { + "bbox": [ + 131, + 561, + 479, + 644 + ], + "type": "inline_equation", + "content": "t_1" + }, + { + "bbox": [ + 131, + 561, + 479, + 644 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 561, + 479, + 644 + ], + "type": "inline_equation", + "content": "t_2" + }, + { + "bbox": [ + 131, + 561, + 479, + 644 + ], + "type": "text", + "content": " are computed using intrinsic, camera motion, and depth values. In Fig. 7b, when " + }, + { + "bbox": [ + 131, + 561, + 479, + 644 + ], + "type": "inline_equation", + "content": "g(\\cdot) > 0" + }, + { + "bbox": [ + 131, + 561, + 479, + 644 + ], + "type": "text", + "content": ", there is no occlusion. In Fig. 7c, when " + }, + { + "bbox": [ + 131, + 561, + 479, + 644 + ], + "type": "inline_equation", + "content": "g(\\cdot) = 0" + }, + { + "bbox": [ + 131, + 561, + 479, + 644 + ], + "type": "text", + "content": ", two pixels overlap and occlusion happens. However, this is hard due to LiDAR sparsity, rasterization, and other noise sources. Hence, we relax with a threshold " + }, + { + "bbox": [ + 131, + 561, + 479, + 644 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 131, + 561, + 479, + 644 + ], + "type": "text", + "content": ". In practice, we set " + }, + { + "bbox": [ + 131, + 561, + 479, + 644 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 131, + 561, + 479, + 644 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 131, + 561, + 479, + 644 + ], + "type": "inline_equation", + "content": "-\\infty" + }, + { + "bbox": [ + 131, + 561, + 479, + 644 + ], + "type": "text", + "content": ". This suggests the pixel " + }, + { + "bbox": [ + 131, + 561, + 479, + 644 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_1" + }, + { + "bbox": [ + 131, + 561, + 479, + 644 + ], + "type": "text", + "content": " is occluded if any pixel " + }, + { + "bbox": [ + 131, + 561, + 479, + 644 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_2" + }, + { + "bbox": [ + 131, + 561, + 479, + 644 + ], + "type": "text", + "content": " in the epipolar line \"surpasses\" it in the RGB view, as shown in Fig. 7d. Formally, we consider pixel " + }, + { + "bbox": [ + 131, + 561, + 479, + 644 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_1" + }, + { + "bbox": [ + 131, + 561, + 479, + 644 + ], + "type": "text", + "content": " occluded as:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 190, + 653, + 480, + 667 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 190, + 653, + 480, + 667 + ], + "spans": [ + { + "bbox": [ + 190, + 653, + 480, + 667 + ], + "type": "interline_equation", + "content": "\\mathcal {O} = \\left\\{\\mathbf {p} _ {1} \\mid \\exists \\mathbf {p} _ {2} \\in \\mathcal {L} _ {\\mathbf {p} _ {1}}, g \\left(\\mathbf {p} _ {1}, \\mathbf {p} _ {2}, d _ {\\mathbf {p} _ {1}}, d _ {\\mathbf {p} _ {2}} \\mid \\mathbf {K}, \\mathbf {P}\\right) < 0 \\right\\}. \\tag {17}", + "image_path": "d644e233240096abadcd8b4744c633c6bf29799fd5d1b758d4ce10e37c4343a5.jpg" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 411, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 411, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 411, + 91, + 447, + 102 + ], + "type": "text", + "content": "RePLAY" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 153 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 153 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 153 + ], + "type": "text", + "content": "Eq. (17) is also a necessary condition of epipolar occlusion. Due to the Li-DAR sparsity, we apply a nearest-neighbor interpolation over virtual LiDAR depthmap " + }, + { + "bbox": [ + 130, + 116, + 482, + 153 + ], + "type": "inline_equation", + "content": "\\mathbf{D}_l" + }, + { + "bbox": [ + 130, + 116, + 482, + 153 + ], + "type": "text", + "content": " to acquire " + }, + { + "bbox": [ + 130, + 116, + 482, + 153 + ], + "type": "inline_equation", + "content": "d_{\\mathbf{p}_2}" + }, + { + "bbox": [ + 130, + 116, + 482, + 153 + ], + "type": "text", + "content": " whenever required, as shown in Fig. 8." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 168, + 218, + 180 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 168, + 218, + 180 + ], + "spans": [ + { + "bbox": [ + 131, + 168, + 218, + 180 + ], + "type": "text", + "content": "3.3 Discussions" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 187, + 482, + 272 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 187, + 482, + 272 + ], + "spans": [ + { + "bbox": [ + 130, + 187, + 482, + 272 + ], + "type": "text", + "content": "Half-Occlusion (Depth Buffering) Baseline. Sec. 2 shows that half-occlusion algorithms solve projective artifacts in stereo vision, i.e., stereo/epipolar occlusion, without knowing LiDAR point cloud. Despite a different setting, they provide a baseline solution. When preparing RGB depthmap " + }, + { + "bbox": [ + 130, + 187, + 482, + 272 + ], + "type": "inline_equation", + "content": "\\mathbf{D}_r" + }, + { + "bbox": [ + 130, + 187, + 482, + 272 + ], + "type": "text", + "content": ", only pixels with smaller depth are preserved if multiple 3D points " + }, + { + "bbox": [ + 130, + 187, + 482, + 272 + ], + "type": "inline_equation", + "content": "\\mathbf{v}" + }, + { + "bbox": [ + 130, + 187, + 482, + 272 + ], + "type": "text", + "content": " have duplicate 2D locations after rasterization. The raw LiDAR depthmap uses half-occlusion. So, we refer Half-Occlusion as \"Raw\" following [23]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 272, + 482, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 272, + 482, + 319 + ], + "spans": [ + { + "bbox": [ + 130, + 272, + 482, + 319 + ], + "type": "text", + "content": "Does RePLAy simply densify the sparse depthmap to remove occlusion? No. Densification alone is insufficient to remove occlusion. RePLAy enables robust densification by formulating a stereo system with an auto-calibration procedure to remove occlusion." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 319, + 482, + 452 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 319, + 482, + 452 + ], + "spans": [ + { + "bbox": [ + 130, + 319, + 482, + 452 + ], + "type": "text", + "content": "Universal Existence of Projective Artifacts. Fig. 2 pictorially depicts that projective artifacts exist under all LiDAR and camera sensor settings, while Fig. 7 shows that these artifacts scale up with a larger baseline distance between sensors. We now show that this indeed is the case. The artifact width distribution plot on the outdoor KITTI and the indoor NYUv2 [53] datasets are shown in Fig. 10 in the Supp. The LiDAR and RGB camera in KITTI have a baseline distance of " + }, + { + "bbox": [ + 130, + 319, + 482, + 452 + ], + "type": "inline_equation", + "content": "0.4m" + }, + { + "bbox": [ + 130, + 319, + 482, + 452 + ], + "type": "text", + "content": " [22]. We set the baseline distance to " + }, + { + "bbox": [ + 130, + 319, + 482, + 452 + ], + "type": "inline_equation", + "content": "0.1m" + }, + { + "bbox": [ + 130, + 319, + 482, + 452 + ], + "type": "text", + "content": " following the AR headset configuration of Fig. 2. Despite the indoor image having a lesser resolution than outdoor (" + }, + { + "bbox": [ + 130, + 319, + 482, + 452 + ], + "type": "inline_equation", + "content": "480 \\times 640" + }, + { + "bbox": [ + 130, + 319, + 482, + 452 + ], + "type": "text", + "content": " compared to " + }, + { + "bbox": [ + 130, + 319, + 482, + 452 + ], + "type": "inline_equation", + "content": "375 \\times 1242" + }, + { + "bbox": [ + 130, + 319, + 482, + 452 + ], + "type": "text", + "content": "), we observe a similar distribution. Detailed definitions and evaluation protocols are in Supp. The analysis confirms that projective artifacts exist in both indoor and outdoor datasets." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 468, + 230, + 483 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 468, + 230, + 483 + ], + "spans": [ + { + "bbox": [ + 131, + 468, + 230, + 483 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 493, + 482, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 493, + 482, + 529 + ], + "spans": [ + { + "bbox": [ + 130, + 493, + 482, + 529 + ], + "type": "text", + "content": "We benchmark RePLAy on depthmap quality, MonoDE and Mono3DOD tasks. Sec. 4.1 assesses depthmap quality, while Sec. 4.2 and Sec. 4.3. report MonoDE and Mono3DOD results respectively." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 529, + 482, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 529, + 482, + 552 + ], + "spans": [ + { + "bbox": [ + 130, + 529, + 482, + 552 + ], + "type": "text", + "content": "Datasets. We experiment on KITTI [21], nuScenes [10] and Waymo [56] datasets. KITTI provides stereo ground truth while the others do not." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 552, + 413, + 565 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 552, + 413, + 565 + ], + "spans": [ + { + "bbox": [ + 131, + 552, + 413, + 565 + ], + "type": "text", + "content": "Data Splits. We use the following splits of the above datasets:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 134, + 569, + 479, + 664 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 134, + 569, + 479, + 580 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 569, + 479, + 580 + ], + "spans": [ + { + "bbox": [ + 134, + 569, + 479, + 580 + ], + "type": "text", + "content": "- KITTI Eigen [15]: 23,158 training and 652 validation samples of raw KITTI." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 134, + 582, + 479, + 603 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 582, + 479, + 603 + ], + "spans": [ + { + "bbox": [ + 134, + 582, + 479, + 603 + ], + "type": "text", + "content": "- KITTI Stereo [46]: 200 training and 200 testing samples out of which 145 training scenes are mapped to the raw KITTI dataset." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 134, + 605, + 479, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 605, + 479, + 628 + ], + "spans": [ + { + "bbox": [ + 134, + 605, + 479, + 628 + ], + "type": "text", + "content": "- nuScenes Frontal Val: 28,130 training and 2,436 random validation samples (from 6,019 samples) from front camera and top LiDAR." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 134, + 630, + 479, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 630, + 479, + 664 + ], + "spans": [ + { + "bbox": [ + 134, + 630, + 479, + 664 + ], + "type": "text", + "content": "- Waymo Frontal Val: 52,386 training and 40,839 validation samples from the front camera and top LiDAR. We construct its training set by sampling every third frame from the training sequences as in [34]." + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "type": "text", + "content": "S. Zhu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 135, + 157, + 479, + 272 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "type": "text", + "content": "Table 2: Depthmap Quality Assessment with baselines and RePLAy. We adopt KITTI semi-dense depthmap and KITTI Stereo disparity map as the reference GT. [Key: Best, Second Best, Mod. = Modified]" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 135, + 157, + 479, + 272 + ], + "lines": [ + { + "bbox": [ + 135, + 157, + 479, + 272 + ], + "spans": [ + { + "bbox": [ + 135, + 157, + 479, + 272 + ], + "type": "table", + "html": "
GTEval RegionDepthmapδ0.5↑δ1↑δ2↑RMS ↓RMSlog ↓ARel ↓SRel ↓log10 ↓SIlog ↓
Semi-DenseAllRaw0.9730.9830.9901.9350.0870.0290.3320.0108.660
Mod. Half-Occ0.9830.9910.9961.3470.0570.0190.1130.0075.712
RePLAy0.9850.9940.9981.2410.0460.0160.0630.0074.592
StereoAllRaw0.9500.9620.9704.1980.1690.0831.6040.02416.691
Mod. Half-Occ0.9710.9820.9882.5410.1050.0470.5460.01610.340
RePLAy0.9750.9860.9912.2240.0860.0340.3540.0148.362
ForegroundRaw0.8590.8710.8928.7260.3170.2285.9740.05829.772
Mod. Half-Occ0.9190.9310.9495.2850.1990.1542.3660.03319.250
RePLAy0.9250.9370.9564.8370.1790.1001.9290.03017.319
BackgroundRaw0.9670.9800.9862.7080.1130.0520.7740.01711.037
Mod. Half-Occ0.9800.9910.9951.7010.0700.0320.2100.0136.638
RePLAy0.9820.9930.9981.4390.0520.0270.0850.0124.722
", + "image_path": "8a24d6b611b0f1ac654f4e5127ffc99c83e8b8c72521d63a447c226fda7a5398.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 135, + 324, + 479, + 354 + ], + "blocks": [ + { + "bbox": [ + 132, + 281, + 482, + 315 + ], + "lines": [ + { + "bbox": [ + 132, + 281, + 482, + 315 + ], + "spans": [ + { + "bbox": [ + 132, + 281, + 482, + 315 + ], + "type": "text", + "content": "Table 3: Depthmap Quality Assessment on Removed Artifacts. RePLAy outperforms the Mod. half-occlusion baseline. For removed artifacts, a worse performance suggests a better algorithm. [Key: Best]" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 135, + 324, + 479, + 354 + ], + "lines": [ + { + "bbox": [ + 135, + 324, + 479, + 354 + ], + "spans": [ + { + "bbox": [ + 135, + 324, + 479, + 354 + ], + "type": "table", + "html": "
GTEval RegionDepthmapδ0.5↓δ1↓δ2↓RMS↑RMSlog↑ARel↑SRel↑log10↑
Semi-DenseAllMod. Half-Occ0.8140.8460.8935.8500.2750.1963.7620.053
RePLAy0.2740.3630.54812.0840.5870.77815.8510.202
", + "image_path": "5e3e964925eb782e13d46115f93b022b7d2dcd7c4539d50a34cbdc014602cf62.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 134, + 361, + 480, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 361, + 480, + 384 + ], + "spans": [ + { + "bbox": [ + 134, + 361, + 480, + 384 + ], + "type": "text", + "content": "- KITTI Detection Val: 3,748 training and 3,762 validation images from the left camera for the Mono3DOD task." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 389, + 478, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 389, + 478, + 403 + ], + "spans": [ + { + "bbox": [ + 132, + 389, + 478, + 403 + ], + "type": "text", + "content": "Ground Truth (GT) Depthmaps. We use the following as GT depthmaps:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 134, + 404, + 481, + 500 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 134, + 404, + 447, + 416 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 404, + 447, + 416 + ], + "spans": [ + { + "bbox": [ + 134, + 404, + 447, + 416 + ], + "type": "text", + "content": "- KITTI Eigen: We use the semi-dense depthmap available in KITTI." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 134, + 416, + 480, + 440 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 416, + 480, + 440 + ], + "spans": [ + { + "bbox": [ + 134, + 416, + 480, + 440 + ], + "type": "text", + "content": "- KITTI Stereo: We generate depthmaps from GT disparity maps, which are denser than semi-dense depthmaps." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 135, + 441, + 481, + 500 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 441, + 481, + 500 + ], + "spans": [ + { + "bbox": [ + 135, + 441, + 481, + 500 + ], + "type": "text", + "content": "- nuScenes and Waymo Frontal Val: We generate GT depthmaps by projecting only those LiDAR points present inside GT 3D bounding boxes. To prepare a reference depthmap, we follow [49] to filter LiDAR point cloud with GT 3D bounding boxes labels. Correspondingly, the evaluation is confined to the bounding box area, as only depth labels in that region are retained." + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 132, + 516, + 316, + 528 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 516, + 316, + 528 + ], + "spans": [ + { + "bbox": [ + 132, + 516, + 316, + 528 + ], + "type": "text", + "content": "4.1 Depthmap Quality Assessment" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 533, + 482, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 482, + 593 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 482, + 593 + ], + "type": "text", + "content": "Baselines. We compare RePLAy to raw depthmap, which is the half-occlusion (depth buffering) method and a modified half-occlusion method. The modified half-occlusion modifies the half-occlusion method (Sec. 3.3) by back-projecting the estimated dense virtual depthmap " + }, + { + "bbox": [ + 130, + 533, + 482, + 593 + ], + "type": "inline_equation", + "content": "(\\overline{\\mathbf{D}}_l)" + }, + { + "bbox": [ + 130, + 533, + 482, + 593 + ], + "type": "text", + "content": " onto the RGB camera and only preserving pixels in " + }, + { + "bbox": [ + 130, + 533, + 482, + 593 + ], + "type": "inline_equation", + "content": "\\mathbf{D}_r" + }, + { + "bbox": [ + 130, + 533, + 482, + 593 + ], + "type": "text", + "content": " that are less than their back-projected values." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": "Results. Tab. 2 compares the RePLAy depthmap with these baselines processed depthmaps. Tab. 2 verifies that RePLAy depthmap outperforms the raw and modified half-occlusion baselines. RePLAy depthmap agrees more with semidense or disparity depthmap after removing the projective artifacts. Fig. 3 shows qualitative comparisons. We believe projective artifacts persist in the half-occlusion method because of the LiDAR sparsity, leading to inferior depthmaps." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 411, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 411, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 411, + 91, + 447, + 102 + ], + "type": "text", + "content": "RePLAY" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 153, + 146, + 461, + 178 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 480, + 137 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 480, + 137 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 480, + 137 + ], + "type": "text", + "content": "Table 4: Auto-Calibration Ablation on Depthmap Quality Assessment with semidense GT depthmap. [Key: Best]" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 153, + 146, + 461, + 178 + ], + "lines": [ + { + "bbox": [ + 153, + 146, + 461, + 178 + ], + "spans": [ + { + "bbox": [ + 153, + 146, + 461, + 178 + ], + "type": "table", + "html": "
DepthmapAutoCalibδ0.5↑δ1↑δ2↑RMS ↓RMSlog ↓ARel ↓SRel ↓log10 ↓SIlog ↓
RePLAy×0.9810.9900.9951.4730.0570.0190.1110.0075.680
✓0.9850.9940.9981.2410.0460.0160.0630.0074.592
", + "image_path": "f6d8fe105746e7ed580956fc4a88f7c73fddf59f58ab52cf0fb5852352871a65.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 153, + 232, + 461, + 325 + ], + "blocks": [ + { + "bbox": [ + 132, + 188, + 480, + 222 + ], + "lines": [ + { + "bbox": [ + 132, + 188, + 480, + 222 + ], + "spans": [ + { + "bbox": [ + 132, + 188, + 480, + 222 + ], + "type": "text", + "content": "Table 5: MonoDE Results on KITTI Stereo. ZoeDepth [5] trained with RePLAy depthmaps outperforms raw and semi-dense [57] depthmaps on foreground. [Key: Best, Second Best]." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 153, + 232, + 461, + 325 + ], + "lines": [ + { + "bbox": [ + 153, + 232, + 461, + 325 + ], + "spans": [ + { + "bbox": [ + 153, + 232, + 461, + 325 + ], + "type": "table", + "html": "
Eval RegionDepthmapδ1↑δ2↑δ3↑RMS ↓RMSlog ↓ARel ↓SRel ↓log10 ↓SIlog ↓
AllRaw0.9550.9820.9922.6470.1160.0670.4200.02611.324
Semi-Dense0.9770.9940.9972.2740.0820.0430.2010.0187.971
RePLAy0.9740.9920.9962.2880.0890.0490.2480.0208.588
ForegroundRaw0.8860.9470.9803.6920.1800.1271.1120.02216.518
Semi-Dense0.9760.9900.9942.0930.0910.0560.3440.0228.342
RePLAy0.9450.9780.9892.9390.1280.0810.6420.02911.798
BackgroundRaw0.9730.9930.9982.2940.0840.0480.1880.0207.981
Semi-Dense0.9780.9960.9982.2830.0760.0400.1670.0177.239
RePLAy0.9810.9970.9992.0840.0700.0390.1380.0176.681
", + "image_path": "b438e377a79a854e06ec24d759262001669a2d44497d775e8d819040635487b3.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 348, + 480, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 348, + 480, + 396 + ], + "spans": [ + { + "bbox": [ + 132, + 348, + 480, + 396 + ], + "type": "text", + "content": "Foreground Improvement. Tab. 2 rows 5 and 6 shows a significant improvement in foreground depthmap quality. On metric " + }, + { + "bbox": [ + 132, + 348, + 480, + 396 + ], + "type": "inline_equation", + "content": "\\delta_{0.5}" + }, + { + "bbox": [ + 132, + 348, + 480, + 396 + ], + "type": "text", + "content": ", we reduce the errors by " + }, + { + "bbox": [ + 132, + 348, + 480, + 396 + ], + "type": "inline_equation", + "content": "46.8\\% = 1 - \\frac{1 - 0.925}{1 - 0.859}" + }, + { + "bbox": [ + 132, + 348, + 480, + 396 + ], + "type": "text", + "content": ". This aligns with our observation that projective artifacts mostly occur over foreground objects, confirming Fig. 1." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 396, + 480, + 480 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 396, + 480, + 480 + ], + "spans": [ + { + "bbox": [ + 132, + 396, + 480, + 480 + ], + "type": "text", + "content": "Comparison on Removed Artifacts. We next compare RePLAy on removed artifacts in Tab. 3. Tab. 3 shows that that in-addition to removing some artifacts, the modified half-occlusion also removes some non-artifacts (good points). Ideally, the removed points should be artifacts; therefore, the best algorithm should perform the worst on the quality assessment of removed points and hence the reversal in the directional indicator for Tab. 3. Qualitative comparison between raw, modified half-occlusion and RePLAy is provided in Supp. (Fig. 11)." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 480, + 487, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 480, + 487, + 529 + ], + "spans": [ + { + "bbox": [ + 132, + 480, + 487, + 529 + ], + "type": "text", + "content": "Auto-Calibration Ablation. We next conduct ablations of the Auto-Calibration module (Sec. 3.1) in Tab. 4. Tab. 4 confirms the effectiveness of Auto-Calibration module in RePLAy, since Auto-Calibration provides " + }, + { + "bbox": [ + 132, + 480, + 487, + 529 + ], + "type": "inline_equation", + "content": "15\\%" + }, + { + "bbox": [ + 132, + 480, + 487, + 529 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 132, + 480, + 487, + 529 + ], + "type": "inline_equation", + "content": "19\\%" + }, + { + "bbox": [ + 132, + 480, + 487, + 529 + ], + "type": "text", + "content": " relative improvement in ARel and " + }, + { + "bbox": [ + 132, + 480, + 487, + 529 + ], + "type": "inline_equation", + "content": "\\mathrm{SI}_{\\log}" + }, + { + "bbox": [ + 132, + 480, + 487, + 529 + ], + "type": "text", + "content": " metrics respectively." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 548, + 359, + 560 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 548, + 359, + 560 + ], + "spans": [ + { + "bbox": [ + 132, + 548, + 359, + 560 + ], + "type": "text", + "content": "4.2 MonoDE: Monocular Depth Estimation" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 569, + 480, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 569, + 480, + 593 + ], + "spans": [ + { + "bbox": [ + 132, + 569, + 480, + 593 + ], + "type": "text", + "content": "Baselines. We compare RePLAy to raw and Semi-Dense methods. We use ZoeDepth [5] as the depth model due to its superior generalization." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 594, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 594, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 594, + 481, + 665 + ], + "type": "text", + "content": "KITTI Stereo and Eigen Results. We train ZoeDepth [5] from scratch using depthmap with and without applying RePLAy on KITTI Eigen split. We evaluate on KITTI Stereo and Eigen split in Tabs. 5 and 6 respectively. Tab. 5 shows that ZoeDepth trained with RePLAy depthmap significantly outperforms the baselines. We observe despite semi-dense GT [57] using additional stereo cameras, RePLAy outperforms Semi-Dense especially in the background. This" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "type": "text", + "content": "S. Zhu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 114, + 248, + 164 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 248, + 164 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 248, + 164 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 248, + 164 + ], + "type": "image", + "image_path": "6f220f610d772c75d69f6e1415071e503e51b4be7125aac7fa4a21db92d575bb.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 164, + 165, + 219, + 175 + ], + "lines": [ + { + "bbox": [ + 164, + 165, + 219, + 175 + ], + "spans": [ + { + "bbox": [ + 164, + 165, + 219, + 175 + ], + "type": "text", + "content": "(a) RGB Input" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 249, + 114, + 364, + 164 + ], + "blocks": [ + { + "bbox": [ + 249, + 114, + 364, + 164 + ], + "lines": [ + { + "bbox": [ + 249, + 114, + 364, + 164 + ], + "spans": [ + { + "bbox": [ + 249, + 114, + 364, + 164 + ], + "type": "image", + "image_path": "de6d75883c0fb02382335191b6cf6e8f0c901164dc6d35b77bde25f1915bcea6.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 271, + 165, + 343, + 175 + ], + "lines": [ + { + "bbox": [ + 271, + 165, + 343, + 175 + ], + "spans": [ + { + "bbox": [ + 271, + 165, + 343, + 175 + ], + "type": "text", + "content": "(b) Raw depthmaps" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 183, + 479, + 217 + ], + "lines": [ + { + "bbox": [ + 130, + 183, + 479, + 217 + ], + "spans": [ + { + "bbox": [ + 130, + 183, + 479, + 217 + ], + "type": "text", + "content": "Fig. 9: MonoDE Qualitative Results on KITTI Eigen dataset. ZoeDepth [5] trained with RePLAy depthmaps (c) outperforms the one with raw depthmaps (b), especially on foreground." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 365, + 114, + 480, + 164 + ], + "blocks": [ + { + "bbox": [ + 365, + 114, + 480, + 164 + ], + "lines": [ + { + "bbox": [ + 365, + 114, + 480, + 164 + ], + "spans": [ + { + "bbox": [ + 365, + 114, + 480, + 164 + ], + "type": "image", + "image_path": "c4d84a85b31e9289af6581e81dc94bc23919ed25f0c87387e078c878bf636a4d.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 381, + 165, + 465, + 175 + ], + "lines": [ + { + "bbox": [ + 381, + 165, + 465, + 175 + ], + "spans": [ + { + "bbox": [ + 381, + 165, + 465, + 175 + ], + "type": "text", + "content": "(c) RePLAy depthmaps" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 135, + 263, + 478, + 328 + ], + "blocks": [ + { + "bbox": [ + 131, + 231, + 479, + 254 + ], + "lines": [ + { + "bbox": [ + 131, + 231, + 479, + 254 + ], + "spans": [ + { + "bbox": [ + 131, + 231, + 479, + 254 + ], + "type": "text", + "content": "Table 6: MonoDE Results. ZoeDepth trained with RePLAy outperforms the baselines in foreground regions of all datasets. [Key: Best]" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 135, + 263, + 478, + 328 + ], + "lines": [ + { + "bbox": [ + 135, + 263, + 478, + 328 + ], + "spans": [ + { + "bbox": [ + 135, + 263, + 478, + 328 + ], + "type": "table", + "html": "
Data SplitEval RegionDepthmapδ1↑δ2↑δ3↑RMS↓RMSlog↓ARel↓SRel↓log10↓SIlog↓
KITTI EigenAllRaw0.8690.9590.9872.9110.1630.1240.6060.04613.501
RePLAy0.9270.9850.9972.1850.1100.0820.2870.0338.674
nuScenes ValForegroundRaw0.9710.9870.9931.3470.0750.0490.2450.0186.339
RePLAy0.9810.9930.9961.1500.0590.0390.1370.0164.905
Waymo ValForegroundRaw0.9410.9880.9963.4680.1110.0790.4870.0329.116
RePLAy0.9620.9930.9983.0570.0910.0630.3670.0267.322
", + "image_path": "dfa1bd433fbd031f772b7a452b3f82c6e561ea6ce0f098a5f3e169461a96a8c9.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 135, + 384, + 478, + 467 + ], + "blocks": [ + { + "bbox": [ + 131, + 341, + 479, + 374 + ], + "lines": [ + { + "bbox": [ + 131, + 341, + 479, + 374 + ], + "spans": [ + { + "bbox": [ + 131, + 341, + 479, + 374 + ], + "type": "text", + "content": "Table 7: MonoDE Analysis to Environmental Conditions on nuScenes Val split. ZoeDepth trained with RePLAY depthmaps shows consistent improvements on foreground under all environmental conditions. [Key: Best]" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 135, + 384, + 478, + 467 + ], + "lines": [ + { + "bbox": [ + 135, + 384, + 478, + 467 + ], + "spans": [ + { + "bbox": [ + 135, + 384, + 478, + 467 + ], + "type": "table", + "html": "
ConditionSamplesDepthmapδ1↑δ2↑δ3↑RMS ↓RMSlog ↓ARel ↓SRel ↓log10 ↓SIlog ↓
Rain864Raw0.9840.9910.9951.3380.0580.0400.1630.0164.859
RePLAy0.9890.9930.9951.2220.0500.0340.1270.0154.110
Night533Raw0.9650.9820.9901.3470.0890.0590.3350.0217.481
RePLAy0.9780.9910.9951.0630.0670.0450.1720.0175.374
Sunny1039Raw0.9640.9870.9931.3520.0810.0530.2670.0196.985
RePLAy0.9760.9930.9961.1340.0620.0400.1280.0165.324
All Tab. 62436Raw0.9710.9870.9931.3470.0750.0490.2450.0186.339
RePLAy0.9810.9930.9961.1500.0590.0390.1370.0164.905
", + "image_path": "38a4a5f90067e98159cd3569801399812b26c9ccd653b3971d3ec1a45e3ab633.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 482, + 479, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 482, + 479, + 517 + ], + "spans": [ + { + "bbox": [ + 130, + 482, + 479, + 517 + ], + "type": "text", + "content": "may be due to [57] aggressively removing the background pixels around the foreground object boundaries. Fig. 9 shows this result visually. ZoeDepth trained with the raw depthmap learns strip-shaped artifacts (green boxes)." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 519, + 479, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 519, + 479, + 567 + ], + "spans": [ + { + "bbox": [ + 130, + 519, + 479, + 567 + ], + "type": "text", + "content": "Foreground Improvement. We additionally ablate depth performance on foreground, background, and entire image using the GT segmentation mask in Tab. 5. Tab. 5 shows that ZoeDepth trained with RePLAy depthmap outperforms in the foreground area, which agrees with the conclusion of Sec. 4.1." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 568, + 479, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 568, + 479, + 604 + ], + "spans": [ + { + "bbox": [ + 130, + 568, + 479, + 604 + ], + "type": "text", + "content": "nuScenes and Waymo Frontal Val Results. We benchmark MonoDE on two datasets without stereo sensors: nuScenes [10] and Waymo [56] in Tab. 6. Tab. 6 confirms that RePLAy achieves significant improvements on these datasets." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 605, + 479, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 479, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 479, + 665 + ], + "type": "text", + "content": "Effect of Environmental Conditions. We next analyze the effect of environmental conditions on the nuScenes Val split in Tab. 7. We categorize the nuScenes [10] val set into rainy, night, and sunny following their description tag (nuScenes-devkit GitHub #124). Tab. 7 shows that RePLAy is robust across multiple conditions." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 411, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 411, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 411, + 91, + 447, + 102 + ], + "type": "text", + "content": "RePLAY" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 391, + 128 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 391, + 128 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 391, + 128 + ], + "type": "text", + "content": "4.3 Mono3DOD: Monocular 3D Object Detection" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 136, + 418, + 147 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 136, + 418, + 147 + ], + "spans": [ + { + "bbox": [ + 130, + 136, + 418, + 147 + ], + "type": "text", + "content": "We finally verify the impact of RePLAy on the Mono3DOD task." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 148, + 481, + 219 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 148, + 481, + 219 + ], + "spans": [ + { + "bbox": [ + 130, + 148, + 481, + 219 + ], + "type": "text", + "content": "Baseline. We compare RePLAy to raw method. We take SoTA 3D detectors, CaDDN [51] and MonoDTR [29], to showcase the impact of cleaning projected LiDAR depthmap with RePLAy, since these detectors use depthmap as an auxiliary training task. The two methods cover both convolutional and transformer architectures, making our benchmarking comprehensive. However, these two SoTA detectors use depth supervision in training differently." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 220, + 280, + 315 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 220, + 280, + 315 + ], + "spans": [ + { + "bbox": [ + 130, + 220, + 280, + 315 + ], + "type": "text", + "content": "CaDDN. CaDDN uses the densified raw depthmaps of [33]. For CaDDN experiments, we apply RePLAy to the projected depthmaps and then densify those maps using the same depth completion algorithm [33] and parameters of [51]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 316, + 280, + 398 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 316, + 280, + 398 + ], + "spans": [ + { + "bbox": [ + 130, + 316, + 280, + 398 + ], + "type": "text", + "content": "MonoDTR. MonoDTR [29] precomputes their GT depthmaps by " + }, + { + "bbox": [ + 130, + 316, + 280, + 398 + ], + "type": "inline_equation", + "content": "4\\times" + }, + { + "bbox": [ + 130, + 316, + 280, + 398 + ], + "type": "text", + "content": " downsampling raw projected depthmaps. For MonoDTR experiments, we clean the projected depthmaps and then " + }, + { + "bbox": [ + 130, + 316, + 280, + 398 + ], + "type": "inline_equation", + "content": "4\\times" + }, + { + "bbox": [ + 130, + 316, + 280, + 398 + ], + "type": "text", + "content": " down-sampling." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 399, + 282, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 399, + 282, + 483 + ], + "spans": [ + { + "bbox": [ + 130, + 399, + 282, + 483 + ], + "type": "text", + "content": "Results. Tab. 8 shows results of retraining both detectors with our depthmaps on KITTI Val split. Cleaned depthmaps improves the performance of CaDDN and MonoDTR on most of the metrics. The biggest gains appear in the Easy" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 483, + 481, + 519 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 483, + 481, + 519 + ], + "spans": [ + { + "bbox": [ + 130, + 483, + 481, + 519 + ], + "type": "text", + "content": "category. This is expected as Easy cars are closer to the ego camera and receive the maximum number of LiDAR points. Hence, improving the projected depthmap by RePLAy benefits the Easy category more." + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 293, + 297, + 482, + 452 + ], + "blocks": [ + { + "bbox": [ + 287, + 240, + 482, + 296 + ], + "lines": [ + { + "bbox": [ + 287, + 240, + 482, + 296 + ], + "spans": [ + { + "bbox": [ + 287, + 240, + 482, + 296 + ], + "type": "text", + "content": "Table 8: KITTI Val Mono3DOD Results on Car category. Using RePLAy depthmaps in training boosts the Mono3DOD performance. [Key: Best, Second Best, " + }, + { + "bbox": [ + 287, + 240, + 482, + 296 + ], + "type": "inline_equation", + "content": "* =" + }, + { + "bbox": [ + 287, + 240, + 482, + 296 + ], + "type": "text", + "content": " Reported, " + }, + { + "bbox": [ + 287, + 240, + 482, + 296 + ], + "type": "inline_equation", + "content": "\\dagger=" + }, + { + "bbox": [ + 287, + 240, + 482, + 296 + ], + "type": "text", + "content": " Retrained, Mod= Moderate.]" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 293, + 297, + 482, + 452 + ], + "lines": [ + { + "bbox": [ + 293, + 297, + 482, + 452 + ], + "spans": [ + { + "bbox": [ + 293, + 297, + 482, + 452 + ], + "type": "table", + "html": "
MethodDepthmapIoU3D > 0.7
AP 3D|R40[%]↑
EasyModHard
M3D-RPN [8]-14.5311.078.65
MonoPair [12]-16.2812.3010.42
MonoDLE [44]-17.4513.6611.68
Kinematic [9]-19.7614.1010.47
GrooMeD [35]-19.6714.3211.27
MonoRUn [11]-20.0214.6512.61
D4LCN [13]Raw*22.3216.2012.30
DFR-Net [73]Raw*24.8117.7814.41
CaDDN [51]Raw†21.6015.6713.24
RePLAy22.8815.7013.18
MonoDTR [29]Raw†24.5018.6615.69
RePLAy26.1518.7715.69
", + "image_path": "8d5f92f42561477292d5a6d12bc3675a8608b4f49aec128cd148b47169fad918.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 537, + 220, + 550 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 537, + 220, + 550 + ], + "spans": [ + { + "bbox": [ + 132, + 537, + 220, + 550 + ], + "type": "text", + "content": "5 Conclusion" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 562, + 481, + 634 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 562, + 481, + 634 + ], + "spans": [ + { + "bbox": [ + 130, + 562, + 481, + 634 + ], + "type": "text", + "content": "We propose an analytical solution to remove projective artifacts in LiDAR depthmap. Our parameter-free method, RePLAy, applies to diverse LiDAR and RGB camera sensor settings. RePLAy only requires the relative position between LiDAR and RGB camera. We verify its benefits with unanimous improvements on SoTA MonoDE and Mono3DOD methods. We release processed depthmaps of five major AV datasets to benefit the community." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 634, + 481, + 659 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 634, + 481, + 659 + ], + "spans": [ + { + "bbox": [ + 130, + 634, + 481, + 659 + ], + "type": "text", + "content": "Limitation. We use nearest-neighbor interpolation to densify virtual LiDAR depthmaps. Advanced strategies, e.g., depth completion, may improve accuracy." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "type": "text", + "content": "S. Zhu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 198, + 127 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 198, + 127 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 198, + 127 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 140, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 138, + 140, + 480, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 140, + 480, + 161 + ], + "spans": [ + { + "bbox": [ + 138, + 140, + 480, + 161 + ], + "type": "text", + "content": "1. The KITTI Vision Benchmark Suite. http://www.cvlabs.net/datasets/kitti/eval_object.php?obj_benchmark=3d (2022), accessed: 2022-07-03 2, 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 162, + 480, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 162, + 480, + 183 + ], + "spans": [ + { + "bbox": [ + 138, + 162, + 480, + 183 + ], + "type": "text", + "content": "2. Arbelaez, P., Maire, M., Fowlkes, C., Malik, J.: Contour detection and hierarchical image segmentation. TPAMI (2010) 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 184, + 480, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 184, + 480, + 215 + ], + "spans": [ + { + "bbox": [ + 138, + 184, + 480, + 215 + ], + "type": "text", + "content": "3. Bai, X., Hu, Z., Zhu, X., Huang, Q., Chen, Y., Fu, H., Tai, C.L.: Transfusion: Robust lidar-camera fusion for 3D object detection with transformers. In: CVPR (2022) 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 216, + 480, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 216, + 480, + 237 + ], + "spans": [ + { + "bbox": [ + 138, + 216, + 480, + 237 + ], + "type": "text", + "content": "4. Bhat, S.F., Alhashim, I., Wonka, P.: Adabins: Depth estimation using adaptive bins. In: CVPR (2021) 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 238, + 480, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 238, + 480, + 270 + ], + "spans": [ + { + "bbox": [ + 138, + 238, + 480, + 270 + ], + "type": "text", + "content": "5. Bhat, S.F., Birkl, R., Wofk, D., Wonka, P., Müller, M.: ZoeDepth: Zero-shot transfer by combining relative and metric depth. arXiv preprint arXiv:2302.12288 (2023) 2, 3, 4, 12, 13" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 271, + 480, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 271, + 480, + 293 + ], + "spans": [ + { + "bbox": [ + 138, + 271, + 480, + 293 + ], + "type": "text", + "content": "6. Birchfield, S., Tomasi, C.: Depth discontinuities by pixel-to-pixel stereo. IJCV (1999) 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 293, + 412, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 293, + 412, + 304 + ], + "spans": [ + { + "bbox": [ + 138, + 293, + 412, + 304 + ], + "type": "text", + "content": "7. Bobick, A.F., Intille, S.S.: Large occlusion stereo. IJCV (1999) 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 304, + 480, + 325 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 304, + 480, + 325 + ], + "spans": [ + { + "bbox": [ + 138, + 304, + 480, + 325 + ], + "type": "text", + "content": "8. Brazil, G., Liu, X.: M3D-RPN: Monocular 3D region proposal network for object detection. In: ICCV (2019) 14" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 326, + 480, + 347 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 326, + 480, + 347 + ], + "spans": [ + { + "bbox": [ + 138, + 326, + 480, + 347 + ], + "type": "text", + "content": "9. Brazil, G., Pons-Moll, G., Liu, X., Schiele, B.: Kinematic 3D object detection in monocular video. In: ECCV (2020) 14" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 348, + 480, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 348, + 480, + 380 + ], + "spans": [ + { + "bbox": [ + 138, + 348, + 480, + 380 + ], + "type": "text", + "content": "0. Caesar, H., Bankiti, V., Lang, A., Vora, S., Liong, V., Xu, Q., Krishnan, A., Pan, Y., Baldan, G., Beijbom, O.: nuScenes: A multimodal dataset for autonomous driving. In: CVPR (2020) 1, 2, 3, 10, 13" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 380, + 480, + 411 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 380, + 480, + 411 + ], + "spans": [ + { + "bbox": [ + 138, + 380, + 480, + 411 + ], + "type": "text", + "content": "1. Chen, H., Huang, Y., Tian, W., Gao, Z., Xiong, L.: MonoRun: Monocular 3D object detection by reconstruction and uncertainty propagation. In: CVPR (2021) 14" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 413, + 480, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 413, + 480, + 435 + ], + "spans": [ + { + "bbox": [ + 138, + 413, + 480, + 435 + ], + "type": "text", + "content": "2. Chen, Y., Tai, L., Sun, K., Li, M.: MonoPair: Monocular 3D object detection using pairwise spatial relationships. In: CVPR (2020) 14" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 435, + 480, + 457 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 435, + 480, + 457 + ], + "spans": [ + { + "bbox": [ + 138, + 435, + 480, + 457 + ], + "type": "text", + "content": "3. Ding, M., Huo, Y., Yi, H., Wang, Z., Shi, J., Lu, Z., Luo, P.: Learning depth-guided convolutions for monocular 3D object detection. In: CVPR Workshops (2020) 14" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 458, + 480, + 500 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 458, + 480, + 500 + ], + "spans": [ + { + "bbox": [ + 138, + 458, + 480, + 500 + ], + "type": "text", + "content": "4. Ebadi, K., Bernreiter, L., Biggie, H., Catt, G., Chang, Y., Chatterjee, A., Denniston, C.E., Deschênes, S.P., Harlow, K., Khattak, S., et al.: Present and future of slam in extreme underground environments. arXiv preprint arXiv:2208.01787 (2022) 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 501, + 480, + 522 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 501, + 480, + 522 + ], + "spans": [ + { + "bbox": [ + 138, + 501, + 480, + 522 + ], + "type": "text", + "content": "5. Eigen, D., Puhrsch, C., Fergus, R.: Depth map prediction from a single image using a multi-scale deep network. NeurIPS (2014) 10" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 523, + 480, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 523, + 480, + 544 + ], + "spans": [ + { + "bbox": [ + 138, + 523, + 480, + 544 + ], + "type": "text", + "content": "6. Filgueira, A., González-Jorge, H., Lagüela, S., Díaz-Vilarino, L., Arias, P.: Quantifying the influence of rain in lidar performance. Measurement (2017) 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 138, + 545, + 480, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 545, + 480, + 567 + ], + "spans": [ + { + "bbox": [ + 138, + 545, + 480, + 567 + ], + "type": "text", + "content": "7. Fua, P.: A parallel stereo algorithm that produces dense depth maps and preserves image features. Machine vision and applications (1993) 4" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 138, + 567, + 446, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 567, + 446, + 578 + ], + "spans": [ + { + "bbox": [ + 138, + 567, + 446, + 578 + ], + "type": "text", + "content": "8. Garg, K., Nayar, S.K.: When does a camera see rain? In: ICCV (2005) 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 138, + 578, + 372, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 578, + 372, + 589 + ], + "spans": [ + { + "bbox": [ + 138, + 578, + 372, + 589 + ], + "type": "text", + "content": "9. Garg, K., Nayar, S.K.: Vision and rain. IJCV (2007) 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 138, + 590, + 480, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 590, + 480, + 620 + ], + "spans": [ + { + "bbox": [ + 138, + 590, + 480, + 620 + ], + "type": "text", + "content": "20. Garvelmann, J., Pohl, S., Weiler, M.: From observation to the quantification of snow processes with a time-lapse camera network. Hydrology and Earth Systems Sciences (2013) 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 138, + 621, + 480, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 621, + 480, + 643 + ], + "spans": [ + { + "bbox": [ + 138, + 621, + 480, + 643 + ], + "type": "text", + "content": "21. Geiger, A., Lenz, P., Stiller, C., Urtasun, R.: Vision meets robotics: The KITTI dataset. IJRR (2013) 10" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 138, + 643, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 643, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 643, + 480, + 665 + ], + "type": "text", + "content": "22. Geiger, A., Lenz, P., Urtasun, R.: Are we ready for autonomous driving? the KITTI vision benchmark suite. In: CVPR (2012) 10" + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 411, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 411, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 411, + 91, + 447, + 102 + ], + "type": "text", + "content": "RePLAY" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 23, + "blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 480, + 138 + ], + "type": "text", + "content": "23. Godard, C., Mac Aodha, O., Firman, M., Brostow, G.J.: Digging into self-supervised monocular depth estimation. In: ICCV (2019) 10" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 139, + 480, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 139, + 480, + 160 + ], + "spans": [ + { + "bbox": [ + 132, + 139, + 480, + 160 + ], + "type": "text", + "content": "24. Guizilini, V., Ambrus, R., Pillai, S., Raventos, A., Gaidon, A.: 3D packing for self-supervised monocular depth estimation. In: CVPR (2020) 1, 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 161, + 480, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 161, + 480, + 182 + ], + "spans": [ + { + "bbox": [ + 132, + 161, + 480, + 182 + ], + "type": "text", + "content": "25. Hartley, R., Zisserman, A.: Multiple view geometry in computer vision. Cambridge university press (2003) 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 183, + 480, + 214 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 183, + 480, + 214 + ], + "spans": [ + { + "bbox": [ + 132, + 183, + 480, + 214 + ], + "type": "text", + "content": "26. Herskovitz, J., Wu, J., White, S., Pavel, A., Reyes, G., Guo, A., Bigham, J.P.: Making mobile augmented reality applications accessible. In: SIGACCESS (2020) 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 216, + 480, + 247 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 216, + 480, + 247 + ], + "spans": [ + { + "bbox": [ + 132, + 216, + 480, + 247 + ], + "type": "text", + "content": "27. Hillmann, C., Hillmann, C.: Comparing the gear vr, oculus go, and oculus quest. Unreal for Mobile and Standalone VR: Create Professional VR Apps Without Coding (2019) 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 249, + 480, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 249, + 480, + 270 + ], + "spans": [ + { + "bbox": [ + 132, + 249, + 480, + 270 + ], + "type": "text", + "content": "28. Hirschmuller, H.: Stereo processing by semiglobal matching and mutual information. TPAMI (2007) 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 271, + 480, + 292 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 271, + 480, + 292 + ], + "spans": [ + { + "bbox": [ + 132, + 271, + 480, + 292 + ], + "type": "text", + "content": "29. Huang, K.C., Wu, T.H., Su, H.T., Hsu, W.: MonoDTR: Monocular 3D object detection with depth-aware transformer. In: CVPR (2022) 3, 14" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 293, + 480, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 293, + 480, + 314 + ], + "spans": [ + { + "bbox": [ + 132, + 293, + 480, + 314 + ], + "type": "text", + "content": "30. Huang, X., Cheng, X., Geng, Q., Cao, B., Zhou, D., Wang, P., Lin, Y., Yang, R.: The apolloscape dataset for autonomous driving. In: CVPR (2018) 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 315, + 480, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 315, + 480, + 357 + ], + "spans": [ + { + "bbox": [ + 132, + 315, + 480, + 357 + ], + "type": "text", + "content": "31. Hutabarat, D., Rivai, M., Purwanto, D., Hutomo, H.: Lidar-based obstacle avoidance for the autonomous mobile robot. In: 2019 12th International Conference on Information & Communication Technology and System (ICTS). pp. 197-202. IEEE (2019) 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 358, + 480, + 379 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 358, + 480, + 379 + ], + "spans": [ + { + "bbox": [ + 132, + 358, + 480, + 379 + ], + "type": "text", + "content": "32. Ishikawa, H., Geiger, D.: Occlusions, discontinuities, and epipolar lines in stereo. In: ECCV (1998) 4" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 380, + 480, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 380, + 480, + 402 + ], + "spans": [ + { + "bbox": [ + 132, + 380, + 480, + 402 + ], + "type": "text", + "content": "33. Ku, J., Harakeh, A., Waslander, S.: In defense of classical image processing: Fast depth completion on the CPU. In: CRV (2018) 14" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 403, + 480, + 424 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 403, + 480, + 424 + ], + "spans": [ + { + "bbox": [ + 132, + 403, + 480, + 424 + ], + "type": "text", + "content": "34. Kumar, A., Brazil, G., Corona, E., Parchami, A., Liu, X.: DEVIANT: Depth Equivariant Network for Monocular 3D Object Detection. In: ECCV (2022) 10" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 425, + 480, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 425, + 480, + 445 + ], + "spans": [ + { + "bbox": [ + 132, + 425, + 480, + 445 + ], + "type": "text", + "content": "35. Kumar, A., Brazil, G., Liu, X.: GrooMeD-NMS: Grouped mathematically differentiable nms for monocular 3D object detection. In: CVPR (2021) 14" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 447, + 480, + 478 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 447, + 480, + 478 + ], + "spans": [ + { + "bbox": [ + 132, + 447, + 480, + 478 + ], + "type": "text", + "content": "36. Kumar, A., Guo, Y., Huang, X., Ren, L., Liu, X.: SeaBird: Segmentation in Bird's View with Dice Loss Improves Monocular 3D Detection of Large Objects. In: CVPR (2024) 4" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 479, + 480, + 500 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 479, + 480, + 500 + ], + "spans": [ + { + "bbox": [ + 132, + 479, + 480, + 500 + ], + "type": "text", + "content": "37. Kutila, M., Pyykonen, P., Holzhüter, H., Colomb, M., Duthon, P.: Automotive lidar performance verification in fog and rain. In: ITSC (2018) 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 502, + 480, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 502, + 480, + 533 + ], + "spans": [ + { + "bbox": [ + 132, + 502, + 480, + 533 + ], + "type": "text", + "content": "38. Lee, J.H., Han, M.K., Ko, D.W., Suh, I.H.: From big to small: Multi-scale local planar guidance for monocular depth estimation. arXiv preprint arXiv:1907.10326 (2019) 4" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 534, + 480, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 534, + 480, + 566 + ], + "spans": [ + { + "bbox": [ + 132, + 534, + 480, + 566 + ], + "type": "text", + "content": "39. Li, Y., Yu, A.W., Meng, T., Caine, B., Ngiam, J., Peng, D., Shen, J., Lu, Y., Zhou, D., Le, Q.V., et al.: Deepfusion: Lidar-camera deep fusion for multi-modal 3d object detection. In: CVPR (2022) 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 567, + 480, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 567, + 480, + 599 + ], + "spans": [ + { + "bbox": [ + 132, + 567, + 480, + 599 + ], + "type": "text", + "content": "40. Li, Y., Ibanez-Guzman, J.: LiDAR for autonomous driving: The principles, challenges, and trends for automotive lidar and perception systems. IEEE Signal Processing Magazine (2020) 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 132, + 600, + 480, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 600, + 480, + 620 + ], + "spans": [ + { + "bbox": [ + 132, + 600, + 480, + 620 + ], + "type": "text", + "content": "41. Liao, Y., Xie, J., Geiger, A.: KITTI-360: A novel dataset and benchmarks for urban scene understanding in 2D and 3D. TPAMI (2022) 1, 3, 4" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 132, + 622, + 480, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 622, + 480, + 643 + ], + "spans": [ + { + "bbox": [ + 132, + 622, + 480, + 643 + ], + "type": "text", + "content": "42. Lin, J.T., Dai, D., Van Gool, L.: Depth estimation from monocular images and sparse radar data. In: IROS (2020) 4" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 132, + 644, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 644, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 644, + 480, + 665 + ], + "type": "text", + "content": "43. Liu, F., Liu, X.: Voxel-based 3d detection and reconstruction of multiple objects from a single image. In: NeurIPS (2021) 4" + } + ] + } + ], + "index": 22 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "text", + "content": "S. Zhu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 665 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 138 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 138 + ], + "type": "text", + "content": "44. Ma, X., Zhang, Y., Xu, D., Zhou, D., Yi, S., Li, H., Ouyang, W.: Delving into localization errors for monocular 3D object detection. In: CVPR (2021) 14" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 140, + 481, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 140, + 481, + 160 + ], + "spans": [ + { + "bbox": [ + 130, + 140, + 481, + 160 + ], + "type": "text", + "content": "45. Martin, D., Fowlkes, C., Malik, J.: Learning to detect natural image boundaries using local brightness, color, and texture cues. TPAMI (2004) 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 161, + 481, + 181 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 161, + 481, + 181 + ], + "spans": [ + { + "bbox": [ + 132, + 161, + 481, + 181 + ], + "type": "text", + "content": "46. Menze, M., Geiger, A.: Object scene flow for autonomous vehicles. In: CVPR (2015) 10" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 182, + 481, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 182, + 481, + 205 + ], + "spans": [ + { + "bbox": [ + 132, + 182, + 481, + 205 + ], + "type": "text", + "content": "47. Michaud, S., Lalonde, J.F., Giguere, P.: Towards characterizing the behavior of lidars in snowy conditions. In: IROS (2015) 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 133, + 205, + 481, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 205, + 481, + 226 + ], + "spans": [ + { + "bbox": [ + 133, + 205, + 481, + 226 + ], + "type": "text", + "content": "48. Milanovic, V., Kasturi, A., Yang, J., Hu, F.: A fast single-pixel laser imager for VR/AR headset tracking. In: MOEMS and Miniaturized Systems XVI (2017) 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 133, + 227, + 481, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 227, + 481, + 248 + ], + "spans": [ + { + "bbox": [ + 133, + 227, + 481, + 248 + ], + "type": "text", + "content": "49. Park, D., Ambrus, R., Guizilini, V., Li, J., Gaidon, A.: Is pseudo-lidar needed for monocular 3d object detection? In: ICCV (2021) 11" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 249, + 481, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 249, + 481, + 270 + ], + "spans": [ + { + "bbox": [ + 132, + 249, + 481, + 270 + ], + "type": "text", + "content": "50. Piccinelli, L., Sakaridis, C., Yu, F.: iDisc: Internal discretization for monocular depth estimation. In: CVPR (2023) 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 271, + 481, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 271, + 481, + 293 + ], + "spans": [ + { + "bbox": [ + 132, + 271, + 481, + 293 + ], + "type": "text", + "content": "51. Reading, C., Harakeh, A., Chae, J., Waslander, S.: Categorical depth distribution network for monocular 3D object detection. In: CVPR (2021) 2, 3, 14" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 293, + 481, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 293, + 481, + 314 + ], + "spans": [ + { + "bbox": [ + 132, + 293, + 481, + 314 + ], + "type": "text", + "content": "52. Royo, S., Ballesta-Garcia, M.: An overview of LiDAR imaging systems for autonomous vehicles. Applied sciences (2019) 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 133, + 315, + 481, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 315, + 481, + 335 + ], + "spans": [ + { + "bbox": [ + 133, + 315, + 481, + 335 + ], + "type": "text", + "content": "53. Silberman, N., Hoiem, D., Kohli, P., Fergus, R.: Indoor segmentation and support inference from rgbd images. ECCV (2012) 10" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 133, + 336, + 481, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 336, + 481, + 357 + ], + "spans": [ + { + "bbox": [ + 133, + 336, + 481, + 357 + ], + "type": "text", + "content": "54. Spinneker, R., Koch, C., Park, S.B., Yoon, J.J.: Fast fog detection for camera based advanced driver assistance systems. In: ITSC (2014) 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 133, + 358, + 481, + 379 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 358, + 481, + 379 + ], + "spans": [ + { + "bbox": [ + 133, + 358, + 481, + 379 + ], + "type": "text", + "content": "55. Sun, J., Li, Y., Kang, S.B., Shum, H.Y.: Symmetric stereo matching for occlusion handling. In: CVPR (2005) 4" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 133, + 380, + 481, + 434 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 380, + 481, + 434 + ], + "spans": [ + { + "bbox": [ + 133, + 380, + 481, + 434 + ], + "type": "text", + "content": "56. Sun, P., Kretzschmar, H., Dotiwalla, X., Chouard, A., Patnaik, V., Tsui, P., Guo, J., Zhou, Y., Chai, Y., Caine, B., Vasudevan, V., Han, W., Ngiam, J., Zhao, H., Timofeev, A., Ettinger, S., Krivokon, M., Gao, A., Joshi, A., Zhang, Y., Shlens, J., Chen, Z., Anguelov, D.: Scalability in perception for autonomous driving: Waymo open dataset. In: CVPR (2020) 1, 2, 3, 10, 13" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 133, + 435, + 481, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 435, + 481, + 456 + ], + "spans": [ + { + "bbox": [ + 133, + 435, + 481, + 456 + ], + "type": "text", + "content": "57. Uhrig, J., Schneider, N., Schneider, L., Franke, U., Brox, T., Geiger, A.: Sparsity invariant cnns. In: 3DV (2017) 3, 4, 12, 13" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 133, + 457, + 481, + 478 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 457, + 481, + 478 + ], + "spans": [ + { + "bbox": [ + 133, + 457, + 481, + 478 + ], + "type": "text", + "content": "58. Wang, J., Zickler, T.: Local detection of stereo occlusion boundaries. In: CVPR (2019) 4" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 133, + 479, + 481, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 479, + 481, + 511 + ], + "spans": [ + { + "bbox": [ + 133, + 479, + 481, + 511 + ], + "type": "text", + "content": "59. Wei, Y., Zhao, L., Zheng, W., Zhu, Z., Rao, Y., Huang, G., Lu, J., Zhou, J.: Surrounddepth: Entangling surrounding views for self-supervised multi-camera depth estimation. In: CoRL (2023) 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 133, + 512, + 481, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 512, + 481, + 533 + ], + "spans": [ + { + "bbox": [ + 133, + 512, + 481, + 533 + ], + "type": "text", + "content": "60. Wei, Y., Quan, L.: Asymmetrical occlusion handling using graph cut for multi-view stereo. In: CVPR (2005) 4" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 133, + 534, + 452, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 534, + 452, + 544 + ], + "spans": [ + { + "bbox": [ + 133, + 534, + 452, + 544 + ], + "type": "text", + "content": "61. Weng, J., Ahuja, N., Huang, T.S.: Two-view matching. In: ICCV (1988) 4" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 133, + 545, + 481, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 545, + 481, + 586 + ], + "spans": [ + { + "bbox": [ + 133, + 545, + 481, + 586 + ], + "type": "text", + "content": "62. Wilson, B., Qi, W., Agarwal, T., Lambert, J., Singh, J., Khandelwal, S., Pan, B., Kumar, R., Hartnett, A., Pontes, J.K., et al.: Argoverse 2: Next generation datasets for self-driving perception and forecasting. arXiv preprint arXiv:2301.00493 (2023) 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 133, + 588, + 481, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 588, + 481, + 610 + ], + "spans": [ + { + "bbox": [ + 133, + 588, + 481, + 610 + ], + "type": "text", + "content": "63. Wu, H., Wen, C., Li, W., Li, X., Yang, R., Wang, C.: Transformation-equivariant 3D object detection for autonomous driving. In: AAAI (2023) 2, 4" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 133, + 611, + 481, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 611, + 481, + 632 + ], + "spans": [ + { + "bbox": [ + 133, + 611, + 481, + 632 + ], + "type": "text", + "content": "64. Wu, H., Wen, C., Shi, S., Li, X., Wang, C.: Virtual sparse convolution for multimodal 3D object detection. In: CVPR (2023) 2, 4" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 133, + 632, + 430, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 632, + 430, + 643 + ], + "spans": [ + { + "bbox": [ + 133, + 632, + 430, + 643 + ], + "type": "text", + "content": "65. Xie, S., Tu, Z.: Holistically-nested edge detection. In: ICCV (2015) 4" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 133, + 643, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 643, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 643, + 481, + 665 + ], + "type": "text", + "content": "66. Yang, Z., Chen, J., Miao, Z., Li, W., Zhu, X., Zhang, L.: Deepinteraction: 3D object detection via modality interaction. NeurIPS (2022) 2, 4" + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 411, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 411, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 411, + 91, + 447, + 102 + ], + "type": "text", + "content": "RePLAY" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 292 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 148 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 480, + 148 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 480, + 148 + ], + "type": "text", + "content": "67. Yu, K., Tao, T., Xie, H., Lin, Z., Liang, T., Wang, B., Chen, P., Hao, D., Wang, Y., Liang, X.: Benchmarking the robustness of lidar-camera fusion for 3D object detection. In: CVPR (2023) 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 149, + 480, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 149, + 480, + 171 + ], + "spans": [ + { + "bbox": [ + 132, + 149, + 480, + 171 + ], + "type": "text", + "content": "68. Yuan, W., Gu, X., Dai, Z., Zhu, S., Tan, P.: Neural window fully-connected crfs for monocular depth estimation. In: CVPR (2022) 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 172, + 480, + 193 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 172, + 480, + 193 + ], + "spans": [ + { + "bbox": [ + 132, + 172, + 480, + 193 + ], + "type": "text", + "content": "69. Zheng, K., Li, S., Qin, K., Li, Z., Zhao, Y., Peng, Z., Cheng, H.: Depth estimation via sparse radar prior and driving scene semantics. In: ACCV (2022) 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 194, + 480, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 194, + 480, + 215 + ], + "spans": [ + { + "bbox": [ + 132, + 194, + 480, + 215 + ], + "type": "text", + "content": "70. Zhu, S., Brazil, G., Liu, X.: The edge of depth: Explicit constraints between segmentation and depth. In: CVPR (2020) 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 216, + 480, + 236 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 216, + 480, + 236 + ], + "spans": [ + { + "bbox": [ + 132, + 216, + 480, + 236 + ], + "type": "text", + "content": "71. Zhu, S., Liu, X.: Lighteddepth: Video depth estimation in light of limited inference view angles. In: CVPR (2023) 4, 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 237, + 480, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 237, + 480, + 258 + ], + "spans": [ + { + "bbox": [ + 132, + 237, + 480, + 258 + ], + "type": "text", + "content": "72. Zitnick, L., Kanade, T.: A cooperative algorithm for stereo matching and occlusion detection. TPAMI (2000) 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 259, + 480, + 292 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 259, + 480, + 292 + ], + "spans": [ + { + "bbox": [ + 132, + 259, + 480, + 292 + ], + "type": "text", + "content": "73. Zou, Z., Ye, X., Du, L., Cheng, X., Tan, X., Zhang, L., Feng, J., Xue, X., Ding, E.: The devil is in the task: Exploiting reciprocal appearance-localization features for monocular 3D object detection. In: ICCV (2021) 14" + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "text", + "content": "S. Zhu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Removing Distributional Discrepancies in Captions Improves Image-Text Alignment/0409d8d6-db45-4397-97c7-eeaa2a4590eb_content_list.json b/2024/Removing Distributional Discrepancies in Captions Improves Image-Text Alignment/0409d8d6-db45-4397-97c7-eeaa2a4590eb_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..c71aae54565febf23b5df730ab8367050c561acb --- /dev/null +++ b/2024/Removing Distributional Discrepancies in Captions Improves Image-Text Alignment/0409d8d6-db45-4397-97c7-eeaa2a4590eb_content_list.json @@ -0,0 +1,1826 @@ +[ + { + "type": "text", + "text": "Removing Distributional Discrepancies in Captions Improves Image-Text Alignment", + "text_level": 1, + "bbox": [ + 261, + 140, + 738, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yuheng Li $^{1}$ , Haotian Liu $^{1}$ , Mu Cai $^{1}$ , Yijun Li $^{2}$ , Eli Shechtman $^{2}$ , Zhe Lin $^{2}$ , Yong Jae Lee $^{1}$ , and Krishna Kumar Singh $^{2}$", + "bbox": [ + 232, + 210, + 771, + 244 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 University of Wisconsin-Madison", + "bbox": [ + 382, + 253, + 617, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "2 Adobe Research", + "bbox": [ + 439, + 268, + 562, + 281 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. In this paper, we introduce a model designed to improve the prediction of image-text alignment, targeting the challenge of compositional understanding in current visual-language models. Our approach focuses on generating high-quality training datasets for the alignment task by producing mixed-type negative captions derived from positive ones. Critically, we address the distribution imbalance between positive and negative captions to ensure that the alignment model does not depend solely on textual information but also considers the associated images for predicting alignment accurately. By creating this enhanced training data, we fine-tune an existing leading visual-language model to boost its capability in understanding alignment. Our model significantly outperforms current top-performing methods across various datasets. We also demonstrate the applicability of our model by ranking the images generated by text-to-image models based on text alignment. Project page: https://yuheng-li.github.io/LLaVA-score/", + "bbox": [ + 261, + 318, + 743, + 527 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 217, + 551, + 375, + 566 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recent years have seen rapid advances in multimodal research, encompassing both visual generation [2, 29, 31] and visual understanding [1, 25, 27]. Training capable multimodal models generally requires extensive datasets of image-text pairs, e.g., LAION [32], CC12M [4], and others [11, 33], which are collected on a web-scale and thus tend to be noisy. This noise in the data contributes to certain challenges. For instance, vision-language models often struggle with hallucination [20] and face difficulties in mastering compositional reasoning [36, 40]. Similarly, text-to-image models frequently fail to generate accurate images when processing complex sentence prompts [10].", + "bbox": [ + 212, + 583, + 787, + 719 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Given these challenges, the ability to automatically assess whether an image and a caption are semantically aligned plays a crucial role. This capability is essential not only for cleaning the data used to pre-train these models but also for evaluating and enhancing the performance of both text-to-image and image-to-text generation models. The most frequently employed metric for this task is the CLIP score [13], which calculates the cosine distance between the CLIP [28] embeddings of the paired text and image. However, it has been observed that CLIP, along with other methods such as BLIP [19] and Flava [34], tends to", + "bbox": [ + 212, + 719, + 787, + 840 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/6acb0a818d1a1ae2c8cdce9e818099d9b54f72c81303aaaf2d1328cfa679878b.jpg", + "image_caption": [ + "Fig.1: Left: Qualitative examples for image-text alignment prediction, where our approach can distinguish fundamental concepts such as positioning, counting, and attributes. Right: Our approach shows superior performance in image-text alignment." + ], + "image_footnote": [], + "bbox": [ + 228, + 141, + 769, + 276 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "operate on a bag-of-words basis. They sometimes even cannot tell the difference between simple cases, such as horse eating grass versus grass eating horse [14, 23, 40].", + "bbox": [ + 212, + 340, + 784, + 385 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The conventional approach to training image-text alignment models involves generating negative captions that are misaligned with the images, which are paired with the original positive captions as the training data. For example, prior works randomly shuffle words to create negative captions [40], or employ language models to generate coherent negative sentences [39]. We follow the latter and leverage an LLM to generate a mix of different types of negative captions during training. Specifically, this includes the replacing type, where one linguistic element is substituted with an arbitrary counterpart (e.g. 'a knife is on the table' $\\rightarrow$ 'a spoon is on the table'), and the swapping type, where words within the same sentence are rearranged (e.g. 'an apple is to the left of a banana' $\\rightarrow$ 'a banana is to the left of an apple'). The former type can aid in enhancing an image-text alignment model's perceptual skills as it needs to distinguish between the original and replaced elements, while the latter type can help the model's reasoning capabilities as elements in the caption remain the same but the relationship between them changes.", + "bbox": [ + 212, + 387, + 785, + 612 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "However, critically, we find that this standard approach of generating negative captions, and ensuring that they have coherence (e.g., proper grammar) on a per-instance level, is insufficient. In particular, this approach cannot ensure consistency at the distribution level between positive and negative captions. Distributional biases may originate from the initial dataset or from the rules or models employed to generate negative captions. For instance, the COCO dataset [17] contains significantly more captions including the word giraffe compared to elephant. Yet, when using GPT to generate negative captions, we observe a tendency for GPT to substitute giraffe with elephant, resulting in a surplus of giraffe mentions in positive captions but more elephant in negative captions. Unfortunately, this means that an image-text alignment model trained on such data can be biased to predict sentences containing elephant to be a positive caption and those that contain giraffe to be a negative caption, independent of the paired image. To address this imbalance, our approach consists of fine-tuning a classifier on text captions only (without paired image inputs) to remove biased data that", + "bbox": [ + 212, + 614, + 785, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "Y. Li et al.", + "bbox": [ + 271, + 114, + 346, + 126 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "the text classifier can correctly predict with high confidence. It's important to note that this bias is not unique to the GPT model or how one prompts it, but can arise in other rule-based methods or pre-trained models as well.", + "bbox": [ + 212, + 146, + 782, + 191 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "While previous research [14] noticed distribution differences due to implausible and non-fluent negative captions, our work is the first to eliminate the dataset-level distribution differences using a language model to implicitly encompass all linguistic aspects, without the need for explicit identification. This approach allows us to incorporate considerations like word frequency which is not identified beforehand, thereby providing a more holistic optimization over the dataset's distribution.", + "bbox": [ + 212, + 191, + 784, + 296 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "For our image-text alignment model, we leverage a state-of-the-art vision-language model, LLaVA [24, 25], and finetune it with our curated data. Our model achieves significantly better results compared with other image-text alignment models, demonstrated in Figure 1. We also demonstrate that our curated data can improve other models like BLIP2 [18]. Finally, we further show that our image-text alignment model can help with other vision-language tasks like ranking generated images from T2I models [15]. In summary, we have four main contributions:", + "bbox": [ + 212, + 297, + 785, + 417 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We identify new dataset-level distribution differences between positive and negative captions that lead to biased image-text alignment models.", + "- To address this, we propose a method to maintain consistency between positive and negative caption distributions, which is critical to ensure that an image-text alignment model relies on both image and text to measure alignment instead of only the biases present in the text.", + "- We use our curated training data to finetune existing visual-language models like LLaVA, and obtain state-of-the-art results for image-text alignment.", + "- In addition, we demonstrate the application of our image-text alignment model in ranking the image generations produced by generative models." + ], + "bbox": [ + 225, + 431, + 784, + 584 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 609, + 387, + 625 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Challenges in Compositional Understanding. Image-text pairs form a crucial interface between visual and linguistic modalities, thus evaluating if a given image-text pair is aligned is important for both data curation and model performance evaluation. The pioneering work of CLIP [28] demonstrated this by leveraging an extensive corpus of such pairs for image-text contrastive training. More recent works such as BLIP [18, 19] and LLaVA [24, 25] further utilize large language models (LLMs) to achieve image-text alignment via the text generation objective. Such models usually inherit the frozen CLIP visual encoder to produce a set of visual tokens, and then feed such tokens and the language instructions into the LLM.", + "bbox": [ + 212, + 642, + 784, + 792 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "It has been observed that vision-language models like CLIP have limited capability in understanding compositionality [26,30,36,40,42]. Specifically, they find it challenging to recognize the permutation of words within sentences [36].", + "bbox": [ + 212, + 794, + 784, + 839 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "LLaVA-score", + "bbox": [ + 642, + 114, + 730, + 126 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Moreover, these models often struggle to identify the binding of attributes to multiple objects within a single sentence or to discern the relationships between objects [30,40,42]. While models built on top of large language models (LLMs), such as BLIP [19] and LLaVA [25], exhibit improved understanding capabilities, they still face difficulties with complex compositional understanding largely due to their lack of training on sufficiently challenging data.", + "bbox": [ + 212, + 146, + 782, + 234 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Enhancing Vision-Language Compositionality. Several studies focus on improving models' understanding of language compositionality [8,9,12,17,23,39, 40]. Some strategies [15,39] involve using language models to decompose text into multiple succinct assertion phrases, which are then evaluated by VQA models on an individual basis. [23] also finds that assessing the conditional probability of predicting text based on the input image offers more accurate outcomes than traditional discriminative approaches like contrastive or classification scores used in BLIP [19]. Still, the most widely used method remains the explicit fine-tuning of models to differentiate between hard negatives and correct captions [9, 39, 40]. [40] randomly shuffles words to generate negative captions, but subsequent analysis by [14] points out the approach's flaw: models could simply rely on textual cues (e.g., grammar correctness) for predictions. [39] attempts to employ LLMs to create negative captions, ensuring the negative captions are both fluent and meaningful. Nonetheless, we observed that current fine-tuning methods do not generate a diverse range of negative prompts. Moreover, merely evaluating the grammar or logical coherence of negative captions is insufficient to eliminate data bias in the distribution of negative captions.", + "bbox": [ + 212, + 237, + 787, + 494 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3 Approach", + "text_level": 1, + "bbox": [ + 215, + 522, + 346, + 539 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In our approach, we assume access to images accompanied by accurately labeled positive captions, similar to those found in the MSCOCO [21] dataset. Following [14], contrary to adopting rule-based techniques that often generate illogical sentences, our method utilizes large language models (LLMs) like GPT4 to transform positive captions into negative ones. We first outline our strategy for generating various types of negative captions and then present a straightforward technique to mitigate biases inherent in the distributions of positive and negative captions. We show the entire pipeline in Figure 2.", + "bbox": [ + 212, + 556, + 784, + 679 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1 Constructing Diverse Negative Captions", + "text_level": 1, + "bbox": [ + 215, + 705, + 594, + 720 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Suppose we have a dataset consisting of image-text pairs $\\{I, T_p\\}$ , where $I$ represents an image and $T_p$ is its associated positive caption. Previous research [39,40] showed that merely randomly shuffling image-text pairs to generate negative samples is insufficient to learn capable vision-language models that properly understand the structure of a caption and its relationship with the image. These studies have underscored the value of constructing hard negative captions to significantly improve the model's language compositional abilities. Following this", + "bbox": [ + 212, + 734, + 782, + 839 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "Y. Li et al.", + "bbox": [ + 271, + 114, + 344, + 126 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/5c5a4dbefb387151ee8e67fcb178f77af6bb8724c00034bc67ff276e97a68e60.jpg", + "image_caption": [ + "Fig. 2: Top: We feed the positive caption (green dots) into GPT to create two types of negative captions (red dots): substituting one linguistic element with any plausible alternative or swapping the positions of two components. The blue part in negative captions highlights the modifications. Bottom: we remove easy negative samples using only text data and utilize the remaining samples to fine-tune vision-language models." + ], + "image_footnote": [], + "bbox": [ + 248, + 145, + 754, + 276 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "insight [39], we utilize large language models (LLMs) to generate such hard negative captions. Specifically, we create two types of hard negative captions.", + "bbox": [ + 212, + 386, + 784, + 416 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We refer to the first method for creating negative captions as the replacing strategy, which identifies key components in a language and uses a language model to replace it with other plausible substitutes. The replaced component can be any linguistic part such as a noun, adjective, preposition, etc. For example, \"a photo of a broken down stop sign\" could be replaced with \"a photo of a brand new stop sign\"; \"a cute cat looking at a bird\" could be changed to \"a cute dog looking at a bird\". Conceptually, this type of negative caption aims to enhance the model's recognition capabilities. For executing this task, we engage GPT by providing it with specific instructions followed by some contextual examples, as illustrated in Figure 3.", + "bbox": [ + 212, + 417, + 787, + 566 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We refer to the second method for creating negative captions as the swapping strategy. This approach involves generating a new sentence by utilizing the original language components, typically resulting in the swapping of two or several identical linguistic elements. Specifically, we first ask GPT to break down the original positive sentence into its key components, and then request the model to construct a new sentence with these elements. For example, if the input caption is \"an airplane is flying in the blue sky\", GPT is tasked with identifying the key components including \"airplane\", \"flying\", \"blue\", and \"sky\". The newly crafted sentence could be \"a blue airplane is flying in the sky\". Please note that for some short captions, there may not be enough language elements to form a reasonably different sentence in meaning. Thus, we also need to judge if the new sentence makes sense or not. Please refer to the Figure 3 for our prompt used in GPT.", + "bbox": [ + 212, + 568, + 787, + 750 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2 Addressing Distribution Discrepancies", + "text_level": 1, + "bbox": [ + 214, + 770, + 578, + 786 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Previous studies [14] have demonstrated that artifact issues arising from incorrect construction in negative captions can lead to distribution disparities between positive and negative captions. To mitigate this, commonsense and grammar", + "bbox": [ + 212, + 794, + 785, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "LLaVA-score", + "bbox": [ + 642, + 114, + 730, + 126 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/979ada63a8130fe6e8e785fb495608efd50a0c44b36ea47277e1b5dfd9216ec9.jpg", + "image_caption": [ + "Fig. 3: Prompts used in GPT for generating two types of negative captions, with in-context examples shown in purple." + ], + "image_footnote": [], + "bbox": [ + 243, + 143, + 754, + 275 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/2a62d04cf16e045d8b6f645658bc1adedf3e2484b619fd58de450e9823cada58.jpg", + "image_caption": [ + "Fig. 4: Top prediction based on a text-only binary classifier. Top left: Negative captions generated through replacement strategy. Top right: Negative captions generated through swapping strategy. Bottom: Positive captions from the COCO dataset." + ], + "image_footnote": [], + "bbox": [ + 243, + 316, + 754, + 481 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "models are employed. Nonetheless, our findings indicate that merely assessing the sensibility of sentences is insufficient for aligning the distributions of positive and negative sentences because of two following two reasons.", + "bbox": [ + 215, + 563, + 787, + 608 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "First, each image-pair dataset inherently establishes its own unique distribution. For instance, the COCO dataset [21], despite its widespread use and diverse content, predominantly encompasses everyday scenes, objects, and activities. Common subjects such as people and motorcycles, along with frequent activities like skiing and surfing, characterize the distinct distribution of COCO. Second, when employing rule-based methods or prompting language models to generate negative captions, any strategy adopted will inevitably reflect biases from human preconceptions or pre-trained models. This leads to negative captions originating from a different distribution.", + "bbox": [ + 212, + 609, + 787, + 747 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To illustrate this concept, we develop a blind text-only binary classifier that processes positive and negative captions without seeing any images. Surprisingly, we find that despite the coherence and logical structure of negative captions, the classifier is adept at distinguishing between negative and positive captions based solely on the text. Figure 4 shows the most confidently correct predictions for both our replace and swap data categories constructed from COCO.", + "bbox": [ + 212, + 750, + 787, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "Y. Li et al.", + "bbox": [ + 271, + 114, + 346, + 126 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In the dataset with replaced negative data, it is observed that GPT frequently substitutes the terms \"boat\" and \"elephant\" for the original COCO captions' \"airplane\" and \"giraffe\", respectively. This pattern makes it straightforward for a text-only model to identify these captions as negative. In our swap data, the text-only classifier is able to identify captions with colorful animals interacting with a black object as negative. Note that these negative prompts are logically coherent and grammatically correct, thus cannot be easily detected by a grammar model. Conversely, the top accurately predicted positive captions often depict common activities like surfing and motorcycle riding, which align closely with the typical content style of COCO. It's important to recognize that data bias is not solely a characteristic of our method for generating data nor the manner in which we prompt GPT. In the experimental section, we demonstrate that such distribution discrepancy is also present in other approaches [14] that employ GPT for creating negative prompts.", + "bbox": [ + 212, + 145, + 787, + 357 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The presence of bias in the data can obstruct a vision-language model's ability to truly understand image structures and learn language compositions, as the model might rely solely on textual cues for making predictions. To address this issue, we propose a straightforward solution aimed at reducing data distribution differences by selectively removing data that is predicted with high confidence by a text-alone model.", + "bbox": [ + 212, + 357, + 787, + 446 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Figure 2 (bottom) depicts our conceptual goal of data filtering: to eliminate straightforward or biased positive and negative samples. This is done to ensure that the remaining text data cannot be distinguished by any text classifier, based solely on the text information. Essentially, our aim is to maximize the entropy of the text information within our dataset. In practice, we organize our dataset into $N$ partitions. For each iteration, one partition is designated as the test set while the remaining $N - 1$ partitions serve as the training set. We employ a pretrained Bert model [7] as our text-only classifier, and train it on the training set. Subsequently, this trained classifier is applied to the designated test set. We then rank the correctly predicted samples by the classifier's confidence level, removing the top $k\\%$ of these samples for both positive and negative class predictions. The rest of the data is retained as our refined dataset. This procedure is repeated for each partition, ensuring a comprehensive reduction of bias across the dataset.", + "bbox": [ + 212, + 449, + 787, + 657 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Note that while our observation and data filtering approach is applied to address the image-text alignment issue, it could be a more general problem for any multimodal data scenario where bias in one modality might negatively impact model training.", + "bbox": [ + 212, + 660, + 787, + 720 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.3 Finetuning VLMs for Image-Text Alignment Scoring", + "text_level": 1, + "bbox": [ + 214, + 739, + 697, + 757 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Once we have the unbiased data, we opt to fine-tune a vision-language model (VLM) to enhance its language compositional understanding capabilities with respect to images. In our study, we mainly select LLaVA-1.5 [24] due to its superior performance in image and text understanding. Since LLaVA-1.5 is designed to generate text, to adapt it for use as a image-text score calculator, we employ", + "bbox": [ + 212, + 763, + 787, + 840 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "LLaVA-score", + "bbox": [ + 642, + 114, + 730, + 126 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 114, + 784, + 125 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "the following prompt formatting: \"Does this image I match the following caption $T$ . Answer Yes or No directly.\" Given that LLaVA relies on a language model to produce subsequent words, we manually extract the logits associated with the responses Yes and No for the next word. We then define the matching score as:", + "bbox": [ + 212, + 146, + 787, + 207 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\frac {e ^ {\\mathbf {P} (\\text {Y e s} | \\text {p r o m p t})}}{e ^ {\\mathbf {P} (\\text {Y e s} | \\text {p r o m p t})} + e ^ {\\mathbf {P} (\\text {N o} | \\text {p r o m p t})}} \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 395, + 218, + 785, + 250 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We discover that this straightforward approach is quite effective and can already outperform many existing state-of-the-art baselines. However, LLaVA-1.5 was not specifically trained for this type of matching problem. Therefore, to enhance its performance, we finetune it with the same prompt formatting introduced above using our curated data. We assign the labels Yes and No to positive and negative pairs, respectively.", + "bbox": [ + 212, + 253, + 787, + 345 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Our curated dataset is not restricted to LLaVA-1.5. In our experiments, we also finetune a Q-former equipped with the Image-Text Matching (ITM) head in BLIP2 [18]. The ITM head is essentially a binary classification layer, identical to our dataset's structure. We thus fine-tune the model using the standard cross-entropy loss.", + "bbox": [ + 212, + 345, + 787, + 421 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4 Results", + "text_level": 1, + "bbox": [ + 215, + 444, + 326, + 459 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We evaluate our fine-tuned LLaVA-1.5 model against various baselines across different datasets. Additionally, we conduct ablation studies to evaluate the impact and importance of our dual-strategy approach for generating diverse negative captions and our method for addressing data distribution discrepancies. Since our main model is built upon LLaVA-1.5, we name our score as LLaVA-score.", + "bbox": [ + 212, + 474, + 787, + 551 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.1Baselines", + "text_level": 1, + "bbox": [ + 215, + 573, + 339, + 587 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We evaluate our model against a range of leading multimodal models:", + "bbox": [ + 212, + 598, + 718, + 612 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "CLIP-ViT-L/14, the largest variant of OpenAI CLIP models [28].", + "bbox": [ + 212, + 613, + 704, + 628 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "BLIP-2 [18], which incorporates both image-text matching and image-text contrastive learning approaches.", + "bbox": [ + 212, + 628, + 785, + 657 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "NegCLIP, as introduced in [40], fine-tuned on challenging negative captions generated by randomly shuffling words within sentences.", + "bbox": [ + 212, + 659, + 785, + 688 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "VisualGPTScore [22] proposes utilizing the probability of generating specific text given an image (i.e., $P(T|I)$ ) as an effective metric for calculating image-text alignment scores. For this, we utilize LLaVA-1.5 [24], the state-of-the-art vision-language model, to calculate their proposed VisualGPTScore, employing a reweighting technique as suggested by [22].", + "bbox": [ + 212, + 689, + 787, + 763 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Image-Reward [38]: A reward model, trained on human preferences of image-text pairs produced by Text-to-Image (T2I) models in DiffusionDB [37].", + "bbox": [ + 212, + 765, + 785, + 794 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "VQ2 [39] first extracts a set of candidate QA pairs from the text, then uses a VQA model to score each pair. For this baseline, we report results using both the PaLI model [6]—directly citing their paper as the model is not publicly", + "bbox": [ + 212, + 795, + 787, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "Y. Li et al.", + "bbox": [ + 271, + 114, + 346, + 127 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "available—and BLIP-T5 [18] which is accessible via their official GitHub repository [3]. Note that while TIFA [15] similarly utilizes a QA pairs approach, we exclude it from our results due to VQ2 demonstrating superior performance, aiming for simplicity in our comparison.", + "bbox": [ + 212, + 146, + 782, + 205 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "PaLI fine-tuned on the SeeTrue dataset [39]: A version of PaLI specifically fine-tuned for the alignment task, using a curated dataset. Performance metrics are cited directly from the original paper [39].", + "bbox": [ + 212, + 205, + 782, + 251 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "LLaVA-1.5 [24], Originally a text generation model, we employ a specific prompt as introduced in Sec. 3.3. By using Equation 1, we convert it into a scoring function. Through empirical testing, this prompt is found to be the most effective, and it is utilized to finetune our model.", + "bbox": [ + 212, + 252, + 784, + 310 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.2 Datasets and Metrics", + "text_level": 1, + "bbox": [ + 214, + 333, + 437, + 347 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We evaluate on the following datasets and metrics:", + "bbox": [ + 212, + 357, + 581, + 369 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Winoground [36]: This dataset uniquely comprises quartets, each including two images and two texts, necessitating a nuanced interpretation of both linguistic and visual elements for accurate matching. The metrics reported for this dataset encompass an image score, a text score, and a group score.", + "bbox": [ + 212, + 372, + 782, + 430 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "SeeTRUe [39]: A benchmark designed for assessing vision-language models, featuring a test set combining multiple sources. Current sources include Drawbench, EditBench, and COCO-t2i, pairing real texts with synthetic images. Following [39], we utilize ROC-AUC as the evaluation metric.", + "bbox": [ + 212, + 431, + 782, + 489 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "SugarCrepe [14]: A recently introduced benchmark focuses on generating creative negative captions using a language model and employs grammar and commonsense models for data cleaning, marking a novel approach in benchmark design.", + "bbox": [ + 212, + 493, + 797, + 535 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "MagicBrush [41]: This benchmark facilitates human-driven image editing, constructed using Dall-E 2, and includes captions for both the original and edited images. Featuring quartets similar to Winoground [36], we adopt the same methodology for calculating the group score with modification due to nature of data. Refer supp for details.", + "bbox": [ + 212, + 537, + 782, + 611 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.3 Implementation Details", + "text_level": 1, + "bbox": [ + 214, + 633, + 455, + 648 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "For our dataset curation, we select COCO, which provides positive image-text pairs. To diversify our image dataset and enhance our model's robustness to synthetic images, we incorporate a subset of the training images from SeeTRUE [39], specifically the coco_train_t2i set. To construct negative data, we use GPT to generate two types of negative captions (swap and replace), and we perform random sampling to maintain an equal amount of positive and negative data. In the process of curating our dataset, we ensure that neither images nor captions included in the training dataset are present within the test dataset.", + "bbox": [ + 212, + 657, + 782, + 777 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "For our model development, we choose to refine LLaVA-1.5 [24] using our curated data. We train the model with a batch size of 64 on 8 NVIDIA A100 GPUs for a single epoch, setting the learning rate at 2e-6. For our data filtering, we use $k$ to be $30\\%$ and $N$ to be 5. Find more details in the supp.", + "bbox": [ + 212, + 779, + 782, + 837 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "LLaVA-score", + "bbox": [ + 642, + 114, + 730, + 126 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 114, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/c8f8c582ee54b4bc1e659aa8ba51f1429d361684bccb249238f6f7a625c44398.jpg", + "table_caption": [ + "Table 1: We evaluate our model alongside baselines across multiple datasets. The best results are shown in bold. For three subsets of the SeeTRUE dataset, we present the ROC AUC scores following the original paper [39]. For the SugarCrepe dataset, accuracy is employed as the performance metric. For the MagicBrush [41] dataset, we report the group score. We use LLaVA-1.5 to calculate the VisualGPT score [23]." + ], + "table_footnote": [], + "table_body": "
Winoground image text groupSeeTRUE DrawBench EditBench COCO-T2ISugarCrepe replace swap addMagicBrush
Chance Performance25.0025.0016.6750.050.050.050.050.050.033.33
CLIP-ViT-L-14 [28]10.5028.507.7561.462.159.279.461.474.852.89
NegCLIP [40]11.7530.758.2563.266.062.885.375.387.261.12
BLIP2-ITM [18]24.2541.7519.0060.867.568.088.983.991.875.32
BLIP2-ITC [18]12.0028.508.5064.967.969.986.766.992.367.85
Image-Reward [38]15.2543.0012.7570.470.277.088.281.095.270.28
VisualGPT [23]37.0044.2527.5077.074.269.188.287.195.578.31
VQ2 (PaLI) [39]42.2547.0030.5082.673.683.4----
VQ2 (BLIP-T5) [39]34.0033.5023.2574.867.474.283.381.090.970.46
PaLI (ft on SeeTRUE) [39]38.0046.5028.7586.877.283.2----
LLaVA-1.5 [24]49.7551.0034.2586.978.384.593.588.395.882.61
LLaVA-score (Ours)68.0053.7547.2588.877.784.995.394.997.587.28
", + "bbox": [ + 217, + 220, + 784, + 383 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.4 Main results", + "text_level": 1, + "bbox": [ + 215, + 402, + 366, + 417 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Table 1 presents a comparison between our models and various strong baselines across different datasets. It is evident that our model outperforms others in nearly all datasets. Interestingly, utilizing Eq. 1, LLaVA-1.5 zero-shot performance is the second-best method overall. Particularly for Winoground, a benchmark well-known for its challenges in visual and linguistic reasoning, some models such as CLIP, BLIP2, and Image-Reward perform at or below chance level. Our fine-tuned LLaVA-1.5 substantially enhancing its reasoning capabilities showing the importance of our strategy for training data curation. For VQ2, the VQA models (BLIP or PaLI) underperform our model due to their lack of training on challenging negative examples. The finetuned PaLI model also falls short, attributed to its training dataset's lack of diversity and absence of data filtering. Our model also shows impressive results on synthetic image benchmarks like SeeTRUE and MagicBrush.", + "bbox": [ + 212, + 429, + 787, + 625 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.5 Performance on Attribute, Counting, Spatial Reasoning", + "text_level": 1, + "bbox": [ + 215, + 648, + 723, + 664 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "In the preceding section, we demonstrated that our model surpasses baseline models across various datasets. This section offers an alternative perspective on our model's performance, particularly in areas where both vision-language models and recent text-to-image (T2I) models encounter significant challenges. These challenges include multiple attributes binding, counting objects, and understanding spatial relationships between objects. Addressing these issues is crucial for both visual understanding and evaluating the capabilities of generative models.", + "bbox": [ + 212, + 672, + 787, + 777 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "To quantify the improvements our model achieves in addressing these common challenges, we construct three specialized datasets. Specifically, we prompt GPT to generate scenarios involving attribute binding, object counting, and spatial relationships, providing in-context examples to guide the generation process.", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "Y. Li et al.", + "bbox": [ + 271, + 114, + 346, + 127 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/c30d4847ba59f672eab4ab0bfa8eca7839c7beafbbd223880ca06e715e3ca888.jpg", + "table_caption": [ + "Table 2: Comparison of our model against baselines across three prevalent challenges, highlighting the highest average accuracy in bold." + ], + "table_footnote": [], + "table_body": "
AttributeCountingSpatial
avgposnegavgposnegavgposneg
Chance Performance505050505050505050
Threshold-Independent Models (with oracle)
CLIP-ViT-L-14 [28]635274586848531888
NegCLIP [40]657852596652574866
BLIP2-ITC [18]667260573678579024
VisualGPT [5]739056655278625668
Inherent Decision Models
BLIP2-ITM [18]5810016539610511002
Image-Reward [38]70100406110022579816
VQ2 (BLIP-T5) [39]669438568230549414
LLaVA-1.5 [24]719844629628579816
LLaVA-score (Ours)819066718656817686
", + "bbox": [ + 277, + 184, + 723, + 386 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Subsequently, the T2I model [35] is employed to generate 50 images for each prompt. From these, we manually select one positive and one negative image based on image-text alignment. We opt for synthetic images as it offers greater flexibility, as illustrated in Figure 5, it can incorporate diverse styles (e.g., painting styles for birds), create unconventional images (e.g., the examples involving soccer and cats), and has attribute/object merging capabilities of T2I models to enrich our dataset with varied negative cases (e.g., the negative example of mistakenly combining clock and apple).", + "bbox": [ + 212, + 417, + 784, + 541 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Table 2 presents our results. We report classification accuracy as the metric, alongside showing both positive and negative results in gray color. We categorize the baseline models into two distinct groups for comparison purposes. 1) Upper section of the table, consists of models that generate scores without an established classification threshold. Consequently, we calculate the accuracy by using an oracle threshold, which involves optimizing over all predicted scores to find the best cutoff (cheating on the test set). 2) Lower part of the table, operates with an inherent decision boundary for classification.", + "bbox": [ + 212, + 546, + 784, + 667 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Our model demonstrates superior performance across all three challenging cases, beating competitors that perform only marginally better than random guessing, particularly on tasks requiring spatial understanding. It's noteworthy that many models exhibit a bias, often predicting the same outcome across the test dataset. This tendency is underscored by their high accuracy on positive samples contrasted with poor performance on negative ones. Given that our images are synthesized using text-to-image (T2I) models, essential linguistic elements are likely to be present even in images labeled as negative (refer to Figure 5). This observation suggests that baseline models operate similarly to a bag-of-visual-words approach, predicting an image as matching the textual description as long as it contains certain key visual concepts.", + "bbox": [ + 212, + 672, + 785, + 840 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "LLaVA-score", + "bbox": [ + 642, + 114, + 730, + 126 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/4bb81f94169eb4a0073f657e01af448d9babb433f03df3bd08867e653360ea73.jpg", + "image_caption": [ + "An antique wooden clock beside a glossy red apple" + ], + "image_footnote": [], + "bbox": [ + 256, + 157, + 326, + 210 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/defbc6ffd1d28e4ef74d3908a2b5d9774449f89ea0b728b872406e055b848eed.jpg", + "image_caption": [ + "Attribute" + ], + "image_footnote": [], + "bbox": [ + 328, + 159, + 398, + 210 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/5783961fefc1047784d1ceee9ad8b089c093c3cbb3df20c4357d868f13d02fea.jpg", + "image_caption": [ + "Counting", + "A fishbowl with two goldfish" + ], + "image_footnote": [], + "bbox": [ + 429, + 159, + 496, + 210 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/3df66c9fae21c5d19b8333f1c2de6a46595c70639c389d09d66a86938f68d315.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 501, + 159, + 568, + 210 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/140f450c36c75dd216cfa816f271dc0116948543f0bf8ef8cb003a7ebf1d15c1.jpg", + "image_caption": [ + "Spatial", + "A soccer ball is embedded within a TV" + ], + "image_footnote": [], + "bbox": [ + 598, + 159, + 665, + 210 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/bd1a812394a82c2cb35f0fe3ed53f82414f82ed129b6a0ebcd3284f77fe1e60e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 669, + 159, + 736, + 210 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/a4aa1dab4ba65cffbeb0ab77bcfb50a4e214155315ec1b729b026dd972d344b3.jpg", + "image_caption": [ + "A towering red cat under a sleek black rose bush", + "Fig. 5: Our curated test datasets feature captions paired with one positive image and one negative image each. All the positive images are displayed on the left side." + ], + "image_footnote": [], + "bbox": [ + 258, + 228, + 325, + 279 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/4a1925bcbfdcec961c90dc671bdf47b0a0634854ccdb825522e27c61c8aa1bda.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 333, + 228, + 401, + 279 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/e67574e55fd1405d34d64ee845c7fb588e5f4fa65d830c645968ef72ca6df915.jpg", + "image_caption": [ + "A drawing of a bird perching on a branch with three leaves" + ], + "image_footnote": [], + "bbox": [ + 428, + 228, + 496, + 279 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/acaf7c4ff19ba7e0b34d962ecd7a6f5134d50536fece83fd46f6ffb9d2fdbe97.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 501, + 228, + 568, + 279 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/11d39e873742f7375db6fe9aef8e1bd92522319eade697bad63860142c29221b.jpg", + "image_caption": [ + "A cheeseburger on a plate with a soda on its right" + ], + "image_footnote": [], + "bbox": [ + 598, + 228, + 666, + 279 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/aed5b308a20204c690c9c5b0ba9c5bc4c9a14d306770935f492a092c68e45c46.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 669, + 228, + 736, + 279 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "4.6 Ablation Studies", + "text_level": 1, + "bbox": [ + 215, + 347, + 401, + 361 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Importance of different negative data. Table 3 begins with our baseline model, LLaVA-1.5 [24], outlining its performance. The following two rows illustrate the effects of finetuning the LLaVA-1.5 model with only replace and swap negative data, with filtering processes applied. As shown in the subcategories of the SugarCrepe dataset, replace and swap data enhance performance in their respective categories but do not significantly impact other cases. Nevertheless, our final model excels across almost all tests, indicating that combining replace and swap negative data has a complementary effect, particularly noticeable in the add category of the SugarCrepe dataset and other evaluated datasets.", + "bbox": [ + 212, + 371, + 787, + 507 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Importance of data filtering. The subsequent row presents outcomes from merging two types of negative data without any filtering. This combination has shown to be beneficial for specific datasets, like Winoground and SugarCrepe, although the results don't match those of the final model with filtering. However, biases in the text data cause the fine-tuned model to show no improvement on datasets like MagicBrush and SeeTRUE. Since data filtering reduces the amount of data, the fifth row displays results from randomly subsampling a training set to match the quantity of filtered data. This comparison highlights the effectiveness of our filtering technique as without it, the model may get biased to just rely on language distribution and ignore the image when making predictions.", + "bbox": [ + 212, + 508, + 787, + 657 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Effect of data filtering. Fig. 6 shows the impact of removing the top $k\\%$ of biased data on the performance across four test datasets. Here, $k$ varies from 0 to 90, representing the progression from no data removal to the exclusion of $90\\%$ of the data. The observed trend indicates that as the biased data is progressively removed, performance improves, peaking at approximately $30\\%$ to $40\\%$ . Beyond this point, performance declines due to the diminishing volume of training data.", + "bbox": [ + 212, + 657, + 787, + 750 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "We assess the quality of our filtered data by training a text-only model on $80\\%$ of the filtered data and testing it on the remaining $20\\%$ . Note that this evaluation is performed on filtered data, and a model should reach an accuracy close to $50\\%$ in ideal data indicating it cannot distinguish between positive and negative caption just using text. We vary the parameter $k$ from $0\\%$ to $90\\%$ , in increments of $10\\%$ , resulting in corresponding filtered data quality percentages", + "bbox": [ + 212, + 750, + 787, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "Y. Li et al.", + "bbox": [ + 271, + 114, + 346, + 126 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/15cfe004653eb3ff500a87a3e4063ae37d8367056ad90b0d96f837ac26fc11d4.jpg", + "table_caption": [ + "Table 3: Upper: Ablating our training data based on LLaVA-1.5 [24]. Lower: Finetuning a BLIP2 model with our data." + ], + "table_footnote": [], + "table_body": "
WinogroundSeeTRUESugarCrepeMagicBrush
imagetextgroupDrawBenchEditBenchCOCO-T2Ireplaceswapadd
LLaVA-1.5 [24]49.7551.0034.2586.978.384.593.588.395.882.61
Only replace w/ filter51.2554.2538.5087.477.683.795.288.895.385.70
Only swap w/ filter63.5049.2542.2581.074.079.292.695.591.376.63
w/o filter65.7551.5046.7588.476.481.194.591.795.082.99
Random subsample64.2550.0043.2588.276.881.894.592.093.683.12
LLaVA-score (Ours)68.0053.7547.2588.877.784.995.394.997.587.28
BLIP2-ITM24.2541.7519.0060.867.568.088.983.991.875.32
BLIP2-ITM (ft on our data)39.542.752887.576.282.894.391.496.079.62
", + "bbox": [ + 217, + 178, + 782, + 297 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/526aa469eb4024741f40c97100b02a1cb14eed3bcf28a9c8eee45fe6239b7081.jpg", + "image_caption": [ + "Fig. 6: Ablation on filter percentage of data. For most of the data, performance peaks around $30\\%$ and reduces after that due to decrease in training data size." + ], + "image_footnote": [], + "bbox": [ + 250, + 306, + 370, + 398 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/a8e31fa643745535ba26272c6d8b0122535148ebeb055f7665a4d5101b544d35.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 375, + 306, + 495, + 398 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/650a56a2d95f5561be9efc467fcc7c3624336a358d483f5e21ff185fe5623d2a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 501, + 306, + 622, + 398 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/c7dfa83c56298c6d5d9e5256c21db02319ad2bd9894d2556b6dd50397a80b702.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 627, + 306, + 746, + 398 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "of $75.9\\%$ , $68.2\\%$ , $60.7\\%$ , $56.4\\%$ , $53.8\\%$ , $51.3\\%$ , $50.6\\%$ , $51.0\\%$ , $49.8\\%$ , and $47.9\\%$ . Empirically, we find that data quality below $60\\%$ is satisfactory for our purposes. Consequently, we decide to discard $30\\%$ of the data in our experiment as it's a good trade-off between quality of filtered data and amount of training data.", + "bbox": [ + 212, + 448, + 782, + 508 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "To demonstrate that the distribution gap we identified also exists in datasets created by others, we apply the same evaluation method to SugarCrepe [14], a test benchmark also generated by GPT, that was refined using grammar and common sense models. Nevertheless, it still exhibits an accuracy of $69.0\\%$ , indicating the presence of the bias.", + "bbox": [ + 212, + 510, + 784, + 584 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Generalization on other model. The bottom section of Table 3 shows that when BLIP2 is fine-tuned with the ITM head (a binary classification head) using our curated data, there is an observable improvement in performance. It's important to highlight that we only finetune and use the Q-former in BLIP2 without including any LLMs, and this model surpasses nearly all baselines in benchmarks that do not utilize LLMs (refer table 1). As expected, the fine-tuned BLIP2 model does not outperform the model fine-tuned using LLaVA, particularly on the Winoground dataset, which is recognized for its challenging reasoning tasks. This suggests the importance of utilizing a more advanced vision-language model as the foundation model for achieving better results. Nonetheless, given that BLIP2's model size (180M) is significantly smaller compared to LLaVA-1.5 (13B), it offers a more lightweight option for simpler applications.", + "bbox": [ + 212, + 585, + 785, + 767 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "4.7 Application: Text Alignment Ranking for Image Generation", + "text_level": 1, + "bbox": [ + 215, + 786, + 754, + 801 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "In this subsection, we showcase an application of our model for image generation. A major challenge with T2I models is their inconsistent adherence to complex", + "bbox": [ + 212, + 809, + 782, + 839 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "LLaVA-score", + "bbox": [ + 642, + 114, + 730, + 126 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 784, + 126 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/ee0485ed87f1558edd4d3eb316beb9189231c49d98024acb91da0ed29d43b9cb.jpg", + "image_caption": [ + "Fig. 7: We show T2I generation results ranked by our model according to image-text alignment. The most aligned images are ranked higher." + ], + "image_footnote": [], + "bbox": [ + 263, + 143, + 743, + 315 + ], + "page_idx": 13 + }, + { + "type": "table", + "img_path": "images/643edbfeeb9a0f83e8507b11d1f70c1358bee5e00e0702bbdeb38b707cb4b8a0.jpg", + "table_caption": [ + "Table 4: Performance evaluation of ranking correlation on the TIFA [15] dataset." + ], + "table_footnote": [], + "table_body": "
CLIP-ViT-L-14 [28]NegCLIP [40]BLIP2-ITM [18]BLIP2-ITC [18]ImageReward [38]VisualGPT [23]VQ2(T5) [39]TIFA[15]LLaVA-1.5[24]Ours
Spearman ρ28.540.143.351.562.836.955.159.761.564.5
Kendall τ19.528.729.936.946.328.741.547.245.849.8
", + "bbox": [ + 222, + 388, + 782, + 438 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "compositional prompts [10]. Inspired by [16], one can generate a series of images from a T2I model and then employ our model to rerank them based on relevance, ultimately presenting the highest-ranked images to the user.", + "bbox": [ + 212, + 465, + 784, + 511 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "TIFA [15] introduced a benchmark with images generated using different T2I models from identical prompts with human rankings. Our evaluation on this dataset reveals our model's superior alignment with human rankings, as shown in Table 4. Spearman's $\\rho$ and Kendall's $\\tau$ are statistical measures that evaluate the similarity between two ranking orders, with higher values indicating greater alignment. Figure 7 illustrates qualitative ranking results using our model for the generation results from a T2I model [35].", + "bbox": [ + 212, + 511, + 785, + 617 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 638, + 359, + 652 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In this paper, we introduced an innovative method for generating high-quality training data for image-text alignment models. By implementing a mixed-type negative caption creation strategy and a novel filtering mechanism, we ensured a balanced distribution between positive and negative captions. This approach not only rectified the biases in existing models but also significantly enhanced the performance of our fine-tuned visual-language models, which achieve state-of-the-art results for image-text alignment. Our findings underscore the importance of high-quality, balanced training data in improving visual-language compositional reasoning and alignment. They also open up new avenues for exploring how these methods can be applied to other tasks or modalities in addition to the visual-language domain. In conclusion, our work offers an effective approach to improving image-text alignment, paving the road for future alignment studies.", + "bbox": [ + 212, + 657, + 787, + 840 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "Y. Li et al.", + "bbox": [ + 271, + 114, + 346, + 126 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "This work was supported in part by NSF CAREER IIS2150012, Adobe Data Science award, Microsoft Accelerate Foundation Models Research Program, and Institute of Information & communications Technology Planning & Evaluation (IITP) grants funded by the Korea government (MSIT) (No. 2022-0-00871, Development of AI Autonomy and Knowledge Enhancement for AI Agent Collaboration) and (No. RS-2022-00187238, Development of Large Korean Language Model Technology for Efficient Pre-training).", + "bbox": [ + 212, + 146, + 787, + 253 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 215, + 277, + 323, + 292 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Alayrac, J.B., Donahue, J., Luc, P., Miech, A., Barr, I., Hasson, Y., Lenc, K., Mensch, A., Millican, K., Reynolds, M., et al.: Flamingo: a visual language model for few-shot learning. NeurIPS 35, 23716-23736 (2022)", + "2. Betker, J., Goh, G., Jing, L., Brooks, T., Wang, J., Li, L., Ouyang, L., Zhuang, J., Lee, J., Guo, Y., Manassra, W., Dhariwal, P., Chu, C., Jiao, Y., Ramesh, A.: Improving image generation with better captions (2023), https://cdn.openai.com/papers/dall-e-3.pdf", + "3. Bitton, Y.: WYSIWYR repository. https://github.com/yonatanbitton/wysiwyr (2023), accessed: 2024-02-24", + "4. Changpinyo, S., Sharma, P., Ding, N., Soricut, R.: Conceptual 12M: Pushing web-scale image-text pre-training to recognize long-tail visual concepts. In: CVPR (2021)", + "5. Chen, J., Guo, H., Yi, K., Li, B., Elhoseiny, M.: Visualgpt: Data-efficient adaptation of pretrained language models for image captioning (June 2022)", + "6. Chen, X., Wang, X., Changpinyo, S., Piergiovanni, A.J., Padlewski, P., Salz, D.M., Goodman, S., Grycner, A., Mustafa, B., Beyer, L., Kolesnikov, A., Puigcerver, J., Ding, N., Rong, K., Akbari, H., Mishra, G., Xue, L., Thapliyal, A.V., Bradbury, J., Kuo, W., Seyedhosseini, M., Jia, C., Ayan, B.K., Riquelme, C., Steiner, A., Angelova, A., Zhai, X., Houlsby, N., Soricut, R.: Pali: A jointly-scaled multilingual language-image model. ArXiv abs/2209.06794 (2022), https://apisemantic scholar.org/CorpusID:252222320", + "7. Devlin, J., Chang, M.W., Lee, K., Toutanova, K.: Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805 (2018)", + "8. Doveh, S., Arbelle, A., Harary, S., Herzig, R., Kim, D., Cascante-Bonilla, P., Alfassy, A., Panda, R., Giryes, R., Feris, R., et al.: Dense and aligned captions (dac) promote compositional reasoning in vl models. Advances in Neural Information Processing Systems 36 (2024)", + "9. Doveh, S., Arbelle, A., Harary, S., Schwartz, E., Herzig, R., Giryes, R., Feris, R., Panda, R., Ullman, S., Karlinsky, L.: Teaching structured vision & language concepts to vision & language models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2657-2668 (2023)", + "10. Feng, W., He, X., Fu, T.J., Jampani, V., Akula, A.R., Narayana, P., Basu, S., Wang, X.E., Wang, W.Y.: Training-free structured diffusion guidance for compositional text-to-image synthesis. In: The Eleventh International Conference on Learning Representations (2023), https://openreview.net/forum?id=PUIqjT4rzq7", + "11. Gadre, S.Y., Ilharco, G., Fang, A., Hayase, J., Smyrnis, G., Nguyen, T., Marten, R., Wortsman, M., Ghosh, D., Zhang, J., et al.: Datacomp: In search of the next" + ], + "bbox": [ + 218, + 309, + 784, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "LLaVA-score", + "bbox": [ + 642, + 114, + 730, + 126 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "generation of multimodal datasets. Advances in Neural Information Processing Systems 36 (2024)", + "12. Herzig, R., Mendelson, A., Karlinsky, L., Arbelle, A., Feris, R., Darrell, T., Globerson, A.: Incorporating structured representations into pretrained vision & language models using scene graphs. EMNLP (2023)", + "13. Hessel, J., Holtzman, A., Forbes, M., Le Bras, R., Choi, Y.: CLIPScore: A reference-free evaluation metric for image captioning. In: Moens, M.F., Huang, X., Specia, L., Yih, S.W.t. (eds.) Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing. pp. 7514-7528. Association for Computational Linguistics, Online and Punta Cana, Dominican Republic (Nov 2021). https://doi.org/10.18653/v1/2021.emnlp-main.595, https://aclanthology.org/2021.emnlp-main.595", + "14. Hsieh, C.Y., Zhang, J., Ma, Z., Kembhavi, A., Krishna, R.: Sugarcrepe: Fixing hackable benchmarks for vision-language compositionality. In: Thirty-Seventh Conference on Neural Information Processing Systems Datasets and Benchmarks Track (2023)", + "15. Hu, Y., Liu, B., Kasai, J., Wang, Y., Ostendorf, M., Krishna, R., Smith, N.A.: Tifa: Accurate and interpretable text-to-image faithfulness evaluation with question answering. ICCV (2023)", + "16. Karthik, S., Roth, K., Mancini, M., Akata, Z.: If at first you don't succeed, try, try again: Faithful diffusion-based text-to-image generation by selection. arXiv preprint arXiv:2305.13308 (2023)", + "17. Le, T., Lal, V., Howard, P.: Coco-counterfactuals: Automatically constructed counterfactual examples for image-text pairs. Advances in Neural Information Processing Systems 36 (2024)", + "18. Li, J., Li, D., Savarese, S., Hoi, S.C.H.: Blip-2: Bootstrapping language-image pretraining with frozen image encoders and large language models. In: International Conference on Machine Learning (2023), https://apisemantic scholar.org/CorpusID:256390509", + "19. Li, J., Li, D., Xiong, C., Hoi, S.: Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In: International Conference on Machine Learning. pp. 12888-12900. PMLR (2022)", + "20. Li, Y., Du, Y., Zhou, K., Wang, J., Zhao, W.X., Wen, J.R.: Evaluating object hallucination in large vision-language models. EMNLP (2023)", + "21. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dólar, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13. pp. 740-755. Springer (2014)", + "22. Lin, Z., Chen, X., Pathak, D., Zhang, P., Ramanan, D.: Revisiting the role of language priors in vision-language models. arXiv preprint arXiv:2306.01879 (2023)", + "23. Lin, Z., Chen, X., Pathak, D., Zhang, P., Ramanan, D.: Visualgptscore: Visio-linguistic reasoning with multimodal generative pre-training scores. arXiv preprint arXiv:2306.01879 (2023)", + "24. Liu, H., Li, C., Li, Y., Lee, Y.J.: Improved baselines with visual instruction tuning (2023)", + "25. Liu, H., Li, C., Wu, Q., Lee, Y.J.: Visual instruction tuning. arXiv:2304.08485 (2023)", + "26. Ma, Z., Hong, J., Gul, M.O., Gandhi, M., Gao, I., Krishna, R.: Crepe: Can vision-language foundation models reason compositionally? arXiv preprint arXiv:2212.07796 (2022)" + ], + "bbox": [ + 217, + 147, + 784, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "Y. Li et al.", + "bbox": [ + 271, + 114, + 346, + 126 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "27. OpenAI: Gpt-4v(ision) system card (2023), https://cdn.openai.com/papers/GPTV_System_Card.pdf", + "28. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021)", + "29. Ramesh, A., Dhariwal, P., Nichol, A., Chu, C., Chen, M.: Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125 (2022)", + "30. Ray, A., Radenovic, F., Dubey, A., Plummer, B.A., Krishna, R., Saenko, K.: Cola: How to adapt vision-language models to compose objects localized with attributes? (2023)", + "31. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10684-10695 (2022)", + "32. Schuhmann, C., Beaumont, R., Vencu, R., Gordon, C., Wightman, R., Cherti, M., Coombes, T., Katta, A., Mullis, C., Wortsman, M., et al.: Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems 35, 25278-25294 (2022)", + "33. Sharma, P., Ding, N., Goodman, S., Soricut, R.: Conceptual captions: A cleaned, hypernymed, image alt-text dataset for automatic image captioning. In: Gurevych, I., Miyao, Y. (eds.) Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). pp. 2556-2565. Association for Computational Linguistics, Melbourne, Australia (Jul 2018). https://doi.org/10.18653/v1/P18-1238, https://aclanthology.org/P18-1238", + "34. Singh, A., Hu, R., Goswami, V., Couairon, G., Galuba, W., Rohrbach, M., Kiela, D.: Flava: A foundational language and vision alignment model. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15638-15650 (2022)", + "35. stabilityai: IF repository. https://github.com/deep-floyd/IF", + "36. Thrush, T., Jiang, R., Bartolo, M., Singh, A., Williams, A., Kiela, D., Ross, C.: Winoground: Probing vision and language models for visio-linguistic compositionality. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5238-5248 (2022)", + "37. Wang, Z.J., Montoya, E., Munechika, D., Yang, H., Hoover, B., Chau, D.H.: Large-scale prompt gallery dataset for text-to-image generative models. arXiv:2210.14896 [cs] (2022), https://arxiv.org/abs/2210.14896", + "38. Xu, J., Liu, X., Wu, Y., Tong, Y., Li, Q., Ding, M., Tang, J., Dong, Y.: Imagereward: Learning and evaluating human preferences for text-to-image generation. Advances in Neural Information Processing Systems 36 (2024)", + "39. Yarom, M., Bitton, Y., Changpinyo, S., Aharoni, R., Herzig, J., Lang, O., Ofek, E., Szpektor, I.: What you see is what you read? improving text-image alignment evaluation. arXiv preprint arXiv:2305.10400 (2023)", + "40. Yuksekgonul, M., Bianchi, F., Kalluri, P., Jurafsky, D., Zou, J.: When and why vision-language models behave like bags-of-words, and what to do about it? In: International Conference on Learning Representations (2023), https://openreview.net/forum?id=KRLUvxh8uaX", + "41. Zhang, K., Mo, L., Chen, W., Sun, H., Su, Y.: Magicbrush: A manually annotated dataset for instruction-guided image editing. In: Advances in Neural Information Processing Systems (2023)" + ], + "bbox": [ + 215, + 147, + 784, + 839 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "LLaVA-score", + "bbox": [ + 642, + 114, + 730, + 126 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + }, + { + "type": "ref_text", + "text": "42. Zhao, T., Zhang, T., Zhu, M., Shen, H., Lee, K., Lu, X., Yin, J.: Vl-checklist: Evaluating pre-trained vision-language models with objects, attributes and relations. arXiv preprint arXiv:2207.00221 (2022)", + "bbox": [ + 217, + 146, + 785, + 189 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "Y. Li et al.", + "bbox": [ + 271, + 114, + 344, + 126 + ], + "page_idx": 17 + } +] \ No newline at end of file diff --git a/2024/Removing Distributional Discrepancies in Captions Improves Image-Text Alignment/0409d8d6-db45-4397-97c7-eeaa2a4590eb_model.json b/2024/Removing Distributional Discrepancies in Captions Improves Image-Text Alignment/0409d8d6-db45-4397-97c7-eeaa2a4590eb_model.json new file mode 100644 index 0000000000000000000000000000000000000000..2c8997ab375f51afdcd4a1f92d733b9763e6fd12 --- /dev/null +++ b/2024/Removing Distributional Discrepancies in Captions Improves Image-Text Alignment/0409d8d6-db45-4397-97c7-eeaa2a4590eb_model.json @@ -0,0 +1,2425 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.263, + 0.141, + 0.74, + 0.187 + ], + "angle": 0, + "content": "Removing Distributional Discrepancies in Captions Improves Image-Text Alignment" + }, + { + "type": "text", + "bbox": [ + 0.233, + 0.212, + 0.772, + 0.245 + ], + "angle": 0, + "content": "Yuheng Li\\(^{1}\\), Haotian Liu\\(^{1}\\), Mu Cai\\(^{1}\\), Yijun Li\\(^{2}\\), Eli Shechtman\\(^{2}\\), Zhe Lin\\(^{2}\\), Yong Jae Lee\\(^{1}\\), and Krishna Kumar Singh\\(^{2}\\)" + }, + { + "type": "text", + "bbox": [ + 0.383, + 0.255, + 0.618, + 0.269 + ], + "angle": 0, + "content": "1 University of Wisconsin-Madison" + }, + { + "type": "text", + "bbox": [ + 0.44, + 0.269, + 0.563, + 0.282 + ], + "angle": 0, + "content": "2 Adobe Research" + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.319, + 0.744, + 0.528 + ], + "angle": 0, + "content": "Abstract. In this paper, we introduce a model designed to improve the prediction of image-text alignment, targeting the challenge of compositional understanding in current visual-language models. Our approach focuses on generating high-quality training datasets for the alignment task by producing mixed-type negative captions derived from positive ones. Critically, we address the distribution imbalance between positive and negative captions to ensure that the alignment model does not depend solely on textual information but also considers the associated images for predicting alignment accurately. By creating this enhanced training data, we fine-tune an existing leading visual-language model to boost its capability in understanding alignment. Our model significantly outperforms current top-performing methods across various datasets. We also demonstrate the applicability of our model by ranking the images generated by text-to-image models based on text alignment. Project page: https://yuheng-li.github.io/LLaVA-score/" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.552, + 0.377, + 0.568 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.584, + 0.788, + 0.72 + ], + "angle": 0, + "content": "Recent years have seen rapid advances in multimodal research, encompassing both visual generation [2, 29, 31] and visual understanding [1, 25, 27]. Training capable multimodal models generally requires extensive datasets of image-text pairs, e.g., LAION [32], CC12M [4], and others [11, 33], which are collected on a web-scale and thus tend to be noisy. This noise in the data contributes to certain challenges. For instance, vision-language models often struggle with hallucination [20] and face difficulties in mastering compositional reasoning [36, 40]. Similarly, text-to-image models frequently fail to generate accurate images when processing complex sentence prompts [10]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Given these challenges, the ability to automatically assess whether an image and a caption are semantically aligned plays a crucial role. This capability is essential not only for cleaning the data used to pre-train these models but also for evaluating and enhancing the performance of both text-to-image and image-to-text generation models. The most frequently employed metric for this task is the CLIP score [13], which calculates the cosine distance between the CLIP [28] embeddings of the paired text and image. However, it has been observed that CLIP, along with other methods such as BLIP [19] and Flava [34], tends to" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.347, + 0.127 + ], + "angle": 0, + "content": "Y. Li et al." + }, + { + "type": "image", + "bbox": [ + 0.23, + 0.142, + 0.77, + 0.277 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.279, + 0.785, + 0.321 + ], + "angle": 0, + "content": "Fig.1: Left: Qualitative examples for image-text alignment prediction, where our approach can distinguish fundamental concepts such as positioning, counting, and attributes. Right: Our approach shows superior performance in image-text alignment." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.342, + 0.785, + 0.386 + ], + "angle": 0, + "content": "operate on a bag-of-words basis. They sometimes even cannot tell the difference between simple cases, such as horse eating grass versus grass eating horse [14, 23, 40]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.388, + 0.787, + 0.613 + ], + "angle": 0, + "content": "The conventional approach to training image-text alignment models involves generating negative captions that are misaligned with the images, which are paired with the original positive captions as the training data. For example, prior works randomly shuffle words to create negative captions [40], or employ language models to generate coherent negative sentences [39]. We follow the latter and leverage an LLM to generate a mix of different types of negative captions during training. Specifically, this includes the replacing type, where one linguistic element is substituted with an arbitrary counterpart (e.g. 'a knife is on the table' \\(\\rightarrow\\) 'a spoon is on the table'), and the swapping type, where words within the same sentence are rearranged (e.g. 'an apple is to the left of a banana' \\(\\rightarrow\\) 'a banana is to the left of an apple'). The former type can aid in enhancing an image-text alignment model's perceptual skills as it needs to distinguish between the original and replaced elements, while the latter type can help the model's reasoning capabilities as elements in the caption remain the same but the relationship between them changes." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.615, + 0.787, + 0.841 + ], + "angle": 0, + "content": "However, critically, we find that this standard approach of generating negative captions, and ensuring that they have coherence (e.g., proper grammar) on a per-instance level, is insufficient. In particular, this approach cannot ensure consistency at the distribution level between positive and negative captions. Distributional biases may originate from the initial dataset or from the rules or models employed to generate negative captions. For instance, the COCO dataset [17] contains significantly more captions including the word giraffe compared to elephant. Yet, when using GPT to generate negative captions, we observe a tendency for GPT to substitute giraffe with elephant, resulting in a surplus of giraffe mentions in positive captions but more elephant in negative captions. Unfortunately, this means that an image-text alignment model trained on such data can be biased to predict sentences containing elephant to be a positive caption and those that contain giraffe to be a negative caption, independent of the paired image. To address this imbalance, our approach consists of fine-tuning a classifier on text captions only (without paired image inputs) to remove biased data that" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.643, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "LLaVA-score" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.192 + ], + "angle": 0, + "content": "the text classifier can correctly predict with high confidence. It's important to note that this bias is not unique to the GPT model or how one prompts it, but can arise in other rule-based methods or pre-trained models as well." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.193, + 0.785, + 0.297 + ], + "angle": 0, + "content": "While previous research [14] noticed distribution differences due to implausible and non-fluent negative captions, our work is the first to eliminate the dataset-level distribution differences using a language model to implicitly encompass all linguistic aspects, without the need for explicit identification. This approach allows us to incorporate considerations like word frequency which is not identified beforehand, thereby providing a more holistic optimization over the dataset's distribution." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.299, + 0.787, + 0.418 + ], + "angle": 0, + "content": "For our image-text alignment model, we leverage a state-of-the-art vision-language model, LLaVA [24, 25], and finetune it with our curated data. Our model achieves significantly better results compared with other image-text alignment models, demonstrated in Figure 1. We also demonstrate that our curated data can improve other models like BLIP2 [18]. Finally, we further show that our image-text alignment model can help with other vision-language tasks like ranking generated images from T2I models [15]. In summary, we have four main contributions:" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.433, + 0.784, + 0.463 + ], + "angle": 0, + "content": "- We identify new dataset-level distribution differences between positive and negative captions that lead to biased image-text alignment models." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.465, + 0.785, + 0.523 + ], + "angle": 0, + "content": "- To address this, we propose a method to maintain consistency between positive and negative caption distributions, which is critical to ensure that an image-text alignment model relies on both image and text to measure alignment instead of only the biases present in the text." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.525, + 0.784, + 0.554 + ], + "angle": 0, + "content": "- We use our curated training data to finetune existing visual-language models like LLaVA, and obtain state-of-the-art results for image-text alignment." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.556, + 0.784, + 0.585 + ], + "angle": 0, + "content": "- In addition, we demonstrate the application of our image-text alignment model in ranking the image generations produced by generative models." + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.433, + 0.785, + 0.585 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.61, + 0.388, + 0.626 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.643, + 0.785, + 0.793 + ], + "angle": 0, + "content": "Challenges in Compositional Understanding. Image-text pairs form a crucial interface between visual and linguistic modalities, thus evaluating if a given image-text pair is aligned is important for both data curation and model performance evaluation. The pioneering work of CLIP [28] demonstrated this by leveraging an extensive corpus of such pairs for image-text contrastive training. More recent works such as BLIP [18, 19] and LLaVA [24, 25] further utilize large language models (LLMs) to achieve image-text alignment via the text generation objective. Such models usually inherit the frozen CLIP visual encoder to produce a set of visual tokens, and then feed such tokens and the language instructions into the LLM." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.785, + 0.84 + ], + "angle": 0, + "content": "It has been observed that vision-language models like CLIP have limited capability in understanding compositionality [26,30,36,40,42]. Specifically, they find it challenging to recognize the permutation of words within sentences [36]." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.346, + 0.127 + ], + "angle": 0, + "content": "Y. Li et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.236 + ], + "angle": 0, + "content": "Moreover, these models often struggle to identify the binding of attributes to multiple objects within a single sentence or to discern the relationships between objects [30,40,42]. While models built on top of large language models (LLMs), such as BLIP [19] and LLaVA [25], exhibit improved understanding capabilities, they still face difficulties with complex compositional understanding largely due to their lack of training on sufficiently challenging data." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.238, + 0.788, + 0.496 + ], + "angle": 0, + "content": "Enhancing Vision-Language Compositionality. Several studies focus on improving models' understanding of language compositionality [8,9,12,17,23,39, 40]. Some strategies [15,39] involve using language models to decompose text into multiple succinct assertion phrases, which are then evaluated by VQA models on an individual basis. [23] also finds that assessing the conditional probability of predicting text based on the input image offers more accurate outcomes than traditional discriminative approaches like contrastive or classification scores used in BLIP [19]. Still, the most widely used method remains the explicit fine-tuning of models to differentiate between hard negatives and correct captions [9, 39, 40]. [40] randomly shuffles words to generate negative captions, but subsequent analysis by [14] points out the approach's flaw: models could simply rely on textual cues (e.g., grammar correctness) for predictions. [39] attempts to employ LLMs to create negative captions, ensuring the negative captions are both fluent and meaningful. Nonetheless, we observed that current fine-tuning methods do not generate a diverse range of negative prompts. Moreover, merely evaluating the grammar or logical coherence of negative captions is insufficient to eliminate data bias in the distribution of negative captions." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.523, + 0.347, + 0.54 + ], + "angle": 0, + "content": "3 Approach" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.558, + 0.785, + 0.68 + ], + "angle": 0, + "content": "In our approach, we assume access to images accompanied by accurately labeled positive captions, similar to those found in the MSCOCO [21] dataset. Following [14], contrary to adopting rule-based techniques that often generate illogical sentences, our method utilizes large language models (LLMs) like GPT4 to transform positive captions into negative ones. We first outline our strategy for generating various types of negative captions and then present a straightforward technique to mitigate biases inherent in the distributions of positive and negative captions. We show the entire pipeline in Figure 2." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.706, + 0.595, + 0.721 + ], + "angle": 0, + "content": "3.1 Constructing Diverse Negative Captions" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.784, + 0.84 + ], + "angle": 0, + "content": "Suppose we have a dataset consisting of image-text pairs \\(\\{I, T_p\\}\\), where \\(I\\) represents an image and \\(T_p\\) is its associated positive caption. Previous research [39,40] showed that merely randomly shuffling image-text pairs to generate negative samples is insufficient to learn capable vision-language models that properly understand the structure of a caption and its relationship with the image. These studies have underscored the value of constructing hard negative captions to significantly improve the model's language compositional abilities. Following this" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.643, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "LLaVA-score" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "image", + "bbox": [ + 0.249, + 0.146, + 0.756, + 0.277 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.288, + 0.788, + 0.359 + ], + "angle": 0, + "content": "Fig. 2: Top: We feed the positive caption (green dots) into GPT to create two types of negative captions (red dots): substituting one linguistic element with any plausible alternative or swapping the positions of two components. The blue part in negative captions highlights the modifications. Bottom: we remove easy negative samples using only text data and utilize the remaining samples to fine-tune vision-language models." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.387, + 0.785, + 0.417 + ], + "angle": 0, + "content": "insight [39], we utilize large language models (LLMs) to generate such hard negative captions. Specifically, we create two types of hard negative captions." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.418, + 0.788, + 0.568 + ], + "angle": 0, + "content": "We refer to the first method for creating negative captions as the replacing strategy, which identifies key components in a language and uses a language model to replace it with other plausible substitutes. The replaced component can be any linguistic part such as a noun, adjective, preposition, etc. For example, \"a photo of a broken down stop sign\" could be replaced with \"a photo of a brand new stop sign\"; \"a cute cat looking at a bird\" could be changed to \"a cute dog looking at a bird\". Conceptually, this type of negative caption aims to enhance the model's recognition capabilities. For executing this task, we engage GPT by providing it with specific instructions followed by some contextual examples, as illustrated in Figure 3." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.569, + 0.788, + 0.751 + ], + "angle": 0, + "content": "We refer to the second method for creating negative captions as the swapping strategy. This approach involves generating a new sentence by utilizing the original language components, typically resulting in the swapping of two or several identical linguistic elements. Specifically, we first ask GPT to break down the original positive sentence into its key components, and then request the model to construct a new sentence with these elements. For example, if the input caption is \"an airplane is flying in the blue sky\", GPT is tasked with identifying the key components including \"airplane\", \"flying\", \"blue\", and \"sky\". The newly crafted sentence could be \"a blue airplane is flying in the sky\". Please note that for some short captions, there may not be enough language elements to form a reasonably different sentence in meaning. Thus, we also need to judge if the new sentence makes sense or not. Please refer to the Figure 3 for our prompt used in GPT." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.771, + 0.579, + 0.787 + ], + "angle": 0, + "content": "3.2 Addressing Distribution Discrepancies" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.787, + 0.842 + ], + "angle": 0, + "content": "Previous studies [14] have demonstrated that artifact issues arising from incorrect construction in negative captions can lead to distribution disparities between positive and negative captions. To mitigate this, commonsense and grammar" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.347, + 0.127 + ], + "angle": 0, + "content": "Y. Li et al." + }, + { + "type": "image", + "bbox": [ + 0.244, + 0.145, + 0.756, + 0.276 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.281, + 0.785, + 0.309 + ], + "angle": 0, + "content": "Fig. 3: Prompts used in GPT for generating two types of negative captions, with in-context examples shown in purple." + }, + { + "type": "image", + "bbox": [ + 0.244, + 0.318, + 0.756, + 0.482 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.486, + 0.788, + 0.529 + ], + "angle": 0, + "content": "Fig. 4: Top prediction based on a text-only binary classifier. Top left: Negative captions generated through replacement strategy. Top right: Negative captions generated through swapping strategy. Bottom: Positive captions from the COCO dataset." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.564, + 0.789, + 0.609 + ], + "angle": 0, + "content": "models are employed. Nonetheless, our findings indicate that merely assessing the sensibility of sentences is insufficient for aligning the distributions of positive and negative sentences because of two following two reasons." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.611, + 0.788, + 0.748 + ], + "angle": 0, + "content": "First, each image-pair dataset inherently establishes its own unique distribution. For instance, the COCO dataset [21], despite its widespread use and diverse content, predominantly encompasses everyday scenes, objects, and activities. Common subjects such as people and motorcycles, along with frequent activities like skiing and surfing, characterize the distinct distribution of COCO. Second, when employing rule-based methods or prompting language models to generate negative captions, any strategy adopted will inevitably reflect biases from human preconceptions or pre-trained models. This leads to negative captions originating from a different distribution." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.841 + ], + "angle": 0, + "content": "To illustrate this concept, we develop a blind text-only binary classifier that processes positive and negative captions without seeing any images. Surprisingly, we find that despite the coherence and logical structure of negative captions, the classifier is adept at distinguishing between negative and positive captions based solely on the text. Figure 4 shows the most confidently correct predictions for both our replace and swap data categories constructed from COCO." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.643, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "LLaVA-score" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.126 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.146, + 0.788, + 0.358 + ], + "angle": 0, + "content": "In the dataset with replaced negative data, it is observed that GPT frequently substitutes the terms \"boat\" and \"elephant\" for the original COCO captions' \"airplane\" and \"giraffe\", respectively. This pattern makes it straightforward for a text-only model to identify these captions as negative. In our swap data, the text-only classifier is able to identify captions with colorful animals interacting with a black object as negative. Note that these negative prompts are logically coherent and grammatically correct, thus cannot be easily detected by a grammar model. Conversely, the top accurately predicted positive captions often depict common activities like surfing and motorcycle riding, which align closely with the typical content style of COCO. It's important to recognize that data bias is not solely a characteristic of our method for generating data nor the manner in which we prompt GPT. In the experimental section, we demonstrate that such distribution discrepancy is also present in other approaches [14] that employ GPT for creating negative prompts." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.358, + 0.788, + 0.448 + ], + "angle": 0, + "content": "The presence of bias in the data can obstruct a vision-language model's ability to truly understand image structures and learn language compositions, as the model might rely solely on textual cues for making predictions. To address this issue, we propose a straightforward solution aimed at reducing data distribution differences by selectively removing data that is predicted with high confidence by a text-alone model." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.45, + 0.788, + 0.659 + ], + "angle": 0, + "content": "Figure 2 (bottom) depicts our conceptual goal of data filtering: to eliminate straightforward or biased positive and negative samples. This is done to ensure that the remaining text data cannot be distinguished by any text classifier, based solely on the text information. Essentially, our aim is to maximize the entropy of the text information within our dataset. In practice, we organize our dataset into \\( N \\) partitions. For each iteration, one partition is designated as the test set while the remaining \\( N - 1 \\) partitions serve as the training set. We employ a pretrained Bert model [7] as our text-only classifier, and train it on the training set. Subsequently, this trained classifier is applied to the designated test set. We then rank the correctly predicted samples by the classifier's confidence level, removing the top \\( k\\% \\) of these samples for both positive and negative class predictions. The rest of the data is retained as our refined dataset. This procedure is repeated for each partition, ensuring a comprehensive reduction of bias across the dataset." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.661, + 0.788, + 0.721 + ], + "angle": 0, + "content": "Note that while our observation and data filtering approach is applied to address the image-text alignment issue, it could be a more general problem for any multimodal data scenario where bias in one modality might negatively impact model training." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.741, + 0.698, + 0.758 + ], + "angle": 0, + "content": "3.3 Finetuning VLMs for Image-Text Alignment Scoring" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Once we have the unbiased data, we opt to fine-tune a vision-language model (VLM) to enhance its language compositional understanding capabilities with respect to images. In our study, we mainly select LLaVA-1.5 [24] due to its superior performance in image and text understanding. Since LLaVA-1.5 is designed to generate text, to adapt it for use as a image-text score calculator, we employ" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.347, + 0.128 + ], + "angle": 0, + "content": "Y. Li et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.208 + ], + "angle": 0, + "content": "the following prompt formatting: \"Does this image I match the following caption \\(T\\). Answer Yes or No directly.\" Given that LLaVA relies on a language model to produce subsequent words, we manually extract the logits associated with the responses Yes and No for the next word. We then define the matching score as:" + }, + { + "type": "equation", + "bbox": [ + 0.397, + 0.219, + 0.786, + 0.251 + ], + "angle": 0, + "content": "\\[\n\\frac {e ^ {\\mathbf {P} (\\text {Y e s} | \\text {p r o m p t})}}{e ^ {\\mathbf {P} (\\text {Y e s} | \\text {p r o m p t})} + e ^ {\\mathbf {P} (\\text {N o} | \\text {p r o m p t})}} \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.255, + 0.788, + 0.346 + ], + "angle": 0, + "content": "We discover that this straightforward approach is quite effective and can already outperform many existing state-of-the-art baselines. However, LLaVA-1.5 was not specifically trained for this type of matching problem. Therefore, to enhance its performance, we finetune it with the same prompt formatting introduced above using our curated data. We assign the labels Yes and No to positive and negative pairs, respectively." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.347, + 0.788, + 0.422 + ], + "angle": 0, + "content": "Our curated dataset is not restricted to LLaVA-1.5. In our experiments, we also finetune a Q-former equipped with the Image-Text Matching (ITM) head in BLIP2 [18]. The ITM head is essentially a binary classification layer, identical to our dataset's structure. We thus fine-tune the model using the standard cross-entropy loss." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.445, + 0.327, + 0.46 + ], + "angle": 0, + "content": "4 Results" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.476, + 0.788, + 0.552 + ], + "angle": 0, + "content": "We evaluate our fine-tuned LLaVA-1.5 model against various baselines across different datasets. Additionally, we conduct ablation studies to evaluate the impact and importance of our dual-strategy approach for generating diverse negative captions and our method for addressing data distribution discrepancies. Since our main model is built upon LLaVA-1.5, we name our score as LLaVA-score." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.574, + 0.341, + 0.588 + ], + "angle": 0, + "content": "4.1Baselines" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.599, + 0.72, + 0.613 + ], + "angle": 0, + "content": "We evaluate our model against a range of leading multimodal models:" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.614, + 0.705, + 0.629 + ], + "angle": 0, + "content": "CLIP-ViT-L/14, the largest variant of OpenAI CLIP models [28]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.63, + 0.786, + 0.659 + ], + "angle": 0, + "content": "BLIP-2 [18], which incorporates both image-text matching and image-text contrastive learning approaches." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.66, + 0.786, + 0.689 + ], + "angle": 0, + "content": "NegCLIP, as introduced in [40], fine-tuned on challenging negative captions generated by randomly shuffling words within sentences." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.69, + 0.788, + 0.765 + ], + "angle": 0, + "content": "VisualGPTScore [22] proposes utilizing the probability of generating specific text given an image (i.e., \\( P(T|I) \\)) as an effective metric for calculating image-text alignment scores. For this, we utilize LLaVA-1.5 [24], the state-of-the-art vision-language model, to calculate their proposed VisualGPTScore, employing a reweighting technique as suggested by [22]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.766, + 0.786, + 0.795 + ], + "angle": 0, + "content": "Image-Reward [38]: A reward model, trained on human preferences of image-text pairs produced by Text-to-Image (T2I) models in DiffusionDB [37]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.796, + 0.788, + 0.842 + ], + "angle": 0, + "content": "VQ2 [39] first extracts a set of candidate QA pairs from the text, then uses a VQA model to score each pair. For this baseline, we report results using both the PaLI model [6]—directly citing their paper as the model is not publicly" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.643, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "LLaVA-score" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.783, + 0.206 + ], + "angle": 0, + "content": "available—and BLIP-T5 [18] which is accessible via their official GitHub repository [3]. Note that while TIFA [15] similarly utilizes a QA pairs approach, we exclude it from our results due to VQ2 demonstrating superior performance, aiming for simplicity in our comparison." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.207, + 0.784, + 0.252 + ], + "angle": 0, + "content": "PaLI fine-tuned on the SeeTrue dataset [39]: A version of PaLI specifically fine-tuned for the alignment task, using a curated dataset. Performance metrics are cited directly from the original paper [39]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.253, + 0.785, + 0.311 + ], + "angle": 0, + "content": "LLaVA-1.5 [24], Originally a text generation model, we employ a specific prompt as introduced in Sec. 3.3. By using Equation 1, we convert it into a scoring function. Through empirical testing, this prompt is found to be the most effective, and it is utilized to finetune our model." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.334, + 0.438, + 0.348 + ], + "angle": 0, + "content": "4.2 Datasets and Metrics" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.358, + 0.583, + 0.371 + ], + "angle": 0, + "content": "We evaluate on the following datasets and metrics:" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.373, + 0.784, + 0.431 + ], + "angle": 0, + "content": "Winoground [36]: This dataset uniquely comprises quartets, each including two images and two texts, necessitating a nuanced interpretation of both linguistic and visual elements for accurate matching. The metrics reported for this dataset encompass an image score, a text score, and a group score." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.433, + 0.784, + 0.491 + ], + "angle": 0, + "content": "SeeTRUe [39]: A benchmark designed for assessing vision-language models, featuring a test set combining multiple sources. Current sources include Drawbench, EditBench, and COCO-t2i, pairing real texts with synthetic images. Following [39], we utilize ROC-AUC as the evaluation metric." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.494, + 0.798, + 0.536 + ], + "angle": 0, + "content": "SugarCrepe [14]: A recently introduced benchmark focuses on generating creative negative captions using a language model and employs grammar and commonsense models for data cleaning, marking a novel approach in benchmark design." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.539, + 0.784, + 0.612 + ], + "angle": 0, + "content": "MagicBrush [41]: This benchmark facilitates human-driven image editing, constructed using Dall-E 2, and includes captions for both the original and edited images. Featuring quartets similar to Winoground [36], we adopt the same methodology for calculating the group score with modification due to nature of data. Refer supp for details." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.635, + 0.456, + 0.65 + ], + "angle": 0, + "content": "4.3 Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.659, + 0.784, + 0.778 + ], + "angle": 0, + "content": "For our dataset curation, we select COCO, which provides positive image-text pairs. To diversify our image dataset and enhance our model's robustness to synthetic images, we incorporate a subset of the training images from SeeTRUE [39], specifically the coco_train_t2i set. To construct negative data, we use GPT to generate two types of negative captions (swap and replace), and we perform random sampling to maintain an equal amount of positive and negative data. In the process of curating our dataset, we ensure that neither images nor captions included in the training dataset are present within the test dataset." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.784, + 0.838 + ], + "angle": 0, + "content": "For our model development, we choose to refine LLaVA-1.5 [24] using our curated data. We train the model with a batch size of 64 on 8 NVIDIA A100 GPUs for a single epoch, setting the learning rate at 2e-6. For our data filtering, we use \\( k \\) to be \\( 30\\% \\) and \\( N \\) to be 5. Find more details in the supp." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.348, + 0.128 + ], + "angle": 0, + "content": "Y. Li et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.145, + 0.788, + 0.217 + ], + "angle": 0, + "content": "Table 1: We evaluate our model alongside baselines across multiple datasets. The best results are shown in bold. For three subsets of the SeeTRUE dataset, we present the ROC AUC scores following the original paper [39]. For the SugarCrepe dataset, accuracy is employed as the performance metric. For the MagicBrush [41] dataset, we report the group score. We use LLaVA-1.5 to calculate the VisualGPT score [23]." + }, + { + "type": "table", + "bbox": [ + 0.218, + 0.221, + 0.785, + 0.384 + ], + "angle": 0, + "content": "
Winoground image text groupSeeTRUE DrawBench EditBench COCO-T2ISugarCrepe replace swap addMagicBrush
Chance Performance25.0025.0016.6750.050.050.050.050.050.033.33
CLIP-ViT-L-14 [28]10.5028.507.7561.462.159.279.461.474.852.89
NegCLIP [40]11.7530.758.2563.266.062.885.375.387.261.12
BLIP2-ITM [18]24.2541.7519.0060.867.568.088.983.991.875.32
BLIP2-ITC [18]12.0028.508.5064.967.969.986.766.992.367.85
Image-Reward [38]15.2543.0012.7570.470.277.088.281.095.270.28
VisualGPT [23]37.0044.2527.5077.074.269.188.287.195.578.31
VQ2 (PaLI) [39]42.2547.0030.5082.673.683.4----
VQ2 (BLIP-T5) [39]34.0033.5023.2574.867.474.283.381.090.970.46
PaLI (ft on SeeTRUE) [39]38.0046.5028.7586.877.283.2----
LLaVA-1.5 [24]49.7551.0034.2586.978.384.593.588.395.882.61
LLaVA-score (Ours)68.0053.7547.2588.877.784.995.394.997.587.28
" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.404, + 0.367, + 0.418 + ], + "angle": 0, + "content": "4.4 Main results" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.43, + 0.789, + 0.626 + ], + "angle": 0, + "content": "Table 1 presents a comparison between our models and various strong baselines across different datasets. It is evident that our model outperforms others in nearly all datasets. Interestingly, utilizing Eq. 1, LLaVA-1.5 zero-shot performance is the second-best method overall. Particularly for Winoground, a benchmark well-known for its challenges in visual and linguistic reasoning, some models such as CLIP, BLIP2, and Image-Reward perform at or below chance level. Our fine-tuned LLaVA-1.5 substantially enhancing its reasoning capabilities showing the importance of our strategy for training data curation. For VQ2, the VQA models (BLIP or PaLI) underperform our model due to their lack of training on challenging negative examples. The finetuned PaLI model also falls short, attributed to its training dataset's lack of diversity and absence of data filtering. Our model also shows impressive results on synthetic image benchmarks like SeeTRUE and MagicBrush." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.649, + 0.724, + 0.665 + ], + "angle": 0, + "content": "4.5 Performance on Attribute, Counting, Spatial Reasoning" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.788, + 0.779 + ], + "angle": 0, + "content": "In the preceding section, we demonstrated that our model surpasses baseline models across various datasets. This section offers an alternative perspective on our model's performance, particularly in areas where both vision-language models and recent text-to-image (T2I) models encounter significant challenges. These challenges include multiple attributes binding, counting objects, and understanding spatial relationships between objects. Addressing these issues is crucial for both visual understanding and evaluating the capabilities of generative models." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.841 + ], + "angle": 0, + "content": "To quantify the improvements our model achieves in addressing these common challenges, we construct three specialized datasets. Specifically, we prompt GPT to generate scenarios involving attribute binding, object counting, and spatial relationships, providing in-context examples to guide the generation process." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.643, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "LLaVA-score" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.784, + 0.174 + ], + "angle": 0, + "content": "Table 2: Comparison of our model against baselines across three prevalent challenges, highlighting the highest average accuracy in bold." + }, + { + "type": "table", + "bbox": [ + 0.279, + 0.185, + 0.724, + 0.387 + ], + "angle": 0, + "content": "
AttributeCountingSpatial
avgposnegavgposnegavgposneg
Chance Performance505050505050505050
Threshold-Independent Models (with oracle)
CLIP-ViT-L-14 [28]635274586848531888
NegCLIP [40]657852596652574866
BLIP2-ITC [18]667260573678579024
VisualGPT [5]739056655278625668
Inherent Decision Models
BLIP2-ITM [18]5810016539610511002
Image-Reward [38]70100406110022579816
VQ2 (BLIP-T5) [39]669438568230549414
LLaVA-1.5 [24]719844629628579816
LLaVA-score (Ours)819066718656817686
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.419, + 0.785, + 0.542 + ], + "angle": 0, + "content": "Subsequently, the T2I model [35] is employed to generate 50 images for each prompt. From these, we manually select one positive and one negative image based on image-text alignment. We opt for synthetic images as it offers greater flexibility, as illustrated in Figure 5, it can incorporate diverse styles (e.g., painting styles for birds), create unconventional images (e.g., the examples involving soccer and cats), and has attribute/object merging capabilities of T2I models to enrich our dataset with varied negative cases (e.g., the negative example of mistakenly combining clock and apple)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.547, + 0.785, + 0.669 + ], + "angle": 0, + "content": "Table 2 presents our results. We report classification accuracy as the metric, alongside showing both positive and negative results in gray color. We categorize the baseline models into two distinct groups for comparison purposes. 1) Upper section of the table, consists of models that generate scores without an established classification threshold. Consequently, we calculate the accuracy by using an oracle threshold, which involves optimizing over all predicted scores to find the best cutoff (cheating on the test set). 2) Lower part of the table, operates with an inherent decision boundary for classification." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.787, + 0.841 + ], + "angle": 0, + "content": "Our model demonstrates superior performance across all three challenging cases, beating competitors that perform only marginally better than random guessing, particularly on tasks requiring spatial understanding. It's noteworthy that many models exhibit a bias, often predicting the same outcome across the test dataset. This tendency is underscored by their high accuracy on positive samples contrasted with poor performance on negative ones. Given that our images are synthesized using text-to-image (T2I) models, essential linguistic elements are likely to be present even in images labeled as negative (refer to Figure 5). This observation suggests that baseline models operate similarly to a bag-of-visual-words approach, predicting an image as matching the textual description as long as it contains certain key visual concepts." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.347, + 0.127 + ], + "angle": 0, + "content": "Y. Li et al." + }, + { + "type": "image_caption", + "bbox": [ + 0.308, + 0.149, + 0.346, + 0.157 + ], + "angle": 0, + "content": "Attribute" + }, + { + "type": "image", + "bbox": [ + 0.258, + 0.159, + 0.327, + 0.211 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.33, + 0.16, + 0.399, + 0.211 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.275, + 0.212, + 0.385, + 0.226 + ], + "angle": 0, + "content": "An antique wooden clock beside a glossy red apple" + }, + { + "type": "image_caption", + "bbox": [ + 0.482, + 0.15, + 0.518, + 0.159 + ], + "angle": 0, + "content": "Counting" + }, + { + "type": "image", + "bbox": [ + 0.43, + 0.16, + 0.498, + 0.212 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.451, + 0.212, + 0.547, + 0.219 + ], + "angle": 0, + "content": "A fishbowl with two goldfish" + }, + { + "type": "image", + "bbox": [ + 0.502, + 0.16, + 0.569, + 0.212 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.652, + 0.15, + 0.68, + 0.158 + ], + "angle": 0, + "content": "Spatial" + }, + { + "type": "image", + "bbox": [ + 0.599, + 0.16, + 0.666, + 0.212 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.67, + 0.16, + 0.738, + 0.212 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.606, + 0.213, + 0.733, + 0.22 + ], + "angle": 0, + "content": "A soccer ball is embedded within a TV" + }, + { + "type": "image", + "bbox": [ + 0.259, + 0.229, + 0.326, + 0.28 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.285, + 0.281, + 0.374, + 0.295 + ], + "angle": 0, + "content": "A towering red cat under a sleek black rose bush" + }, + { + "type": "image", + "bbox": [ + 0.334, + 0.229, + 0.402, + 0.28 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.429, + 0.229, + 0.497, + 0.28 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.502, + 0.229, + 0.569, + 0.28 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.45, + 0.281, + 0.549, + 0.295 + ], + "angle": 0, + "content": "A drawing of a bird perching on a branch with three leaves" + }, + { + "type": "image", + "bbox": [ + 0.599, + 0.229, + 0.667, + 0.28 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.67, + 0.229, + 0.738, + 0.28 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.617, + 0.281, + 0.722, + 0.295 + ], + "angle": 0, + "content": "A cheeseburger on a plate with a soda on its right" + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.299, + 0.785, + 0.327 + ], + "angle": 0, + "content": "Fig. 5: Our curated test datasets feature captions paired with one positive image and one negative image each. All the positive images are displayed on the left side." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.348, + 0.402, + 0.362 + ], + "angle": 0, + "content": "4.6 Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.372, + 0.788, + 0.508 + ], + "angle": 0, + "content": "Importance of different negative data. Table 3 begins with our baseline model, LLaVA-1.5 [24], outlining its performance. The following two rows illustrate the effects of finetuning the LLaVA-1.5 model with only replace and swap negative data, with filtering processes applied. As shown in the subcategories of the SugarCrepe dataset, replace and swap data enhance performance in their respective categories but do not significantly impact other cases. Nevertheless, our final model excels across almost all tests, indicating that combining replace and swap negative data has a complementary effect, particularly noticeable in the add category of the SugarCrepe dataset and other evaluated datasets." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.509, + 0.788, + 0.659 + ], + "angle": 0, + "content": "Importance of data filtering. The subsequent row presents outcomes from merging two types of negative data without any filtering. This combination has shown to be beneficial for specific datasets, like Winoground and SugarCrepe, although the results don't match those of the final model with filtering. However, biases in the text data cause the fine-tuned model to show no improvement on datasets like MagicBrush and SeeTRUE. Since data filtering reduces the amount of data, the fifth row displays results from randomly subsampling a training set to match the quantity of filtered data. This comparison highlights the effectiveness of our filtering technique as without it, the model may get biased to just rely on language distribution and ignore the image when making predictions." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.659, + 0.788, + 0.75 + ], + "angle": 0, + "content": "Effect of data filtering. Fig. 6 shows the impact of removing the top \\( k\\% \\) of biased data on the performance across four test datasets. Here, \\( k \\) varies from 0 to 90, representing the progression from no data removal to the exclusion of \\( 90\\% \\) of the data. The observed trend indicates that as the biased data is progressively removed, performance improves, peaking at approximately \\( 30\\% \\) to \\( 40\\% \\). Beyond this point, performance declines due to the diminishing volume of training data." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.841 + ], + "angle": 0, + "content": "We assess the quality of our filtered data by training a text-only model on \\(80\\%\\) of the filtered data and testing it on the remaining \\(20\\%\\). Note that this evaluation is performed on filtered data, and a model should reach an accuracy close to \\(50\\%\\) in ideal data indicating it cannot distinguish between positive and negative caption just using text. We vary the parameter \\(k\\) from \\(0\\%\\) to \\(90\\%\\), in increments of \\(10\\%\\), resulting in corresponding filtered data quality percentages" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.643, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "LLaVA-score" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.784, + 0.173 + ], + "angle": 0, + "content": "Table 3: Upper: Ablating our training data based on LLaVA-1.5 [24]. Lower: Finetuning a BLIP2 model with our data." + }, + { + "type": "table", + "bbox": [ + 0.218, + 0.179, + 0.784, + 0.299 + ], + "angle": 0, + "content": "
WinogroundSeeTRUESugarCrepeMagicBrush
imagetextgroupDrawBenchEditBenchCOCO-T2Ireplaceswapadd
LLaVA-1.5 [24]49.7551.0034.2586.978.384.593.588.395.882.61
Only replace w/ filter51.2554.2538.5087.477.683.795.288.895.385.70
Only swap w/ filter63.5049.2542.2581.074.079.292.695.591.376.63
w/o filter65.7551.5046.7588.476.481.194.591.795.082.99
Random subsample64.2550.0043.2588.276.881.894.592.093.683.12
LLaVA-score (Ours)68.0053.7547.2588.877.784.995.394.997.587.28
BLIP2-ITM24.2541.7519.0060.867.568.088.983.991.875.32
BLIP2-ITM (ft on our data)39.542.752887.576.282.894.391.496.079.62
" + }, + { + "type": "image", + "bbox": [ + 0.251, + 0.307, + 0.371, + 0.4 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.376, + 0.307, + 0.496, + 0.4 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.502, + 0.308, + 0.623, + 0.4 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.629, + 0.308, + 0.747, + 0.4 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.404, + 0.784, + 0.431 + ], + "angle": 0, + "content": "Fig. 6: Ablation on filter percentage of data. For most of the data, performance peaks around \\(30\\%\\) and reduces after that due to decrease in training data size." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.449, + 0.784, + 0.51 + ], + "angle": 0, + "content": "of \\(75.9\\%\\), \\(68.2\\%\\), \\(60.7\\%\\), \\(56.4\\%\\), \\(53.8\\%\\), \\(51.3\\%\\), \\(50.6\\%\\), \\(51.0\\%\\), \\(49.8\\%\\), and \\(47.9\\%\\). Empirically, we find that data quality below \\(60\\%\\) is satisfactory for our purposes. Consequently, we decide to discard \\(30\\%\\) of the data in our experiment as it's a good trade-off between quality of filtered data and amount of training data." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.511, + 0.785, + 0.585 + ], + "angle": 0, + "content": "To demonstrate that the distribution gap we identified also exists in datasets created by others, we apply the same evaluation method to SugarCrepe [14], a test benchmark also generated by GPT, that was refined using grammar and common sense models. Nevertheless, it still exhibits an accuracy of \\(69.0\\%\\), indicating the presence of the bias." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.586, + 0.786, + 0.768 + ], + "angle": 0, + "content": "Generalization on other model. The bottom section of Table 3 shows that when BLIP2 is fine-tuned with the ITM head (a binary classification head) using our curated data, there is an observable improvement in performance. It's important to highlight that we only finetune and use the Q-former in BLIP2 without including any LLMs, and this model surpasses nearly all baselines in benchmarks that do not utilize LLMs (refer table 1). As expected, the fine-tuned BLIP2 model does not outperform the model fine-tuned using LLaVA, particularly on the Winoground dataset, which is recognized for its challenging reasoning tasks. This suggests the importance of utilizing a more advanced vision-language model as the foundation model for achieving better results. Nonetheless, given that BLIP2's model size (180M) is significantly smaller compared to LLaVA-1.5 (13B), it offers a more lightweight option for simpler applications." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.787, + 0.756, + 0.803 + ], + "angle": 0, + "content": "4.7 Application: Text Alignment Ranking for Image Generation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.784, + 0.84 + ], + "angle": 0, + "content": "In this subsection, we showcase an application of our model for image generation. A major challenge with T2I models is their inconsistent adherence to complex" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.347, + 0.127 + ], + "angle": 0, + "content": "Y. Li et al." + }, + { + "type": "image", + "bbox": [ + 0.264, + 0.145, + 0.744, + 0.316 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.328, + 0.785, + 0.357 + ], + "angle": 0, + "content": "Fig. 7: We show T2I generation results ranked by our model according to image-text alignment. The most aligned images are ranked higher." + }, + { + "type": "table_caption", + "bbox": [ + 0.227, + 0.37, + 0.773, + 0.384 + ], + "angle": 0, + "content": "Table 4: Performance evaluation of ranking correlation on the TIFA [15] dataset." + }, + { + "type": "table", + "bbox": [ + 0.223, + 0.39, + 0.784, + 0.439 + ], + "angle": 0, + "content": "
CLIP-ViT-L-14 [28]NegCLIP [40]BLIP2-ITM [18]BLIP2-ITC [18]ImageReward [38]VisualGPT [23]VQ2(T5) [39]TIFA[15]LLaVA-1.5[24]Ours
Spearman ρ28.540.143.351.562.836.955.159.761.564.5
Kendall τ19.528.729.936.946.328.741.547.245.849.8
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.467, + 0.785, + 0.512 + ], + "angle": 0, + "content": "compositional prompts [10]. Inspired by [16], one can generate a series of images from a T2I model and then employ our model to rerank them based on relevance, ultimately presenting the highest-ranked images to the user." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.512, + 0.787, + 0.618 + ], + "angle": 0, + "content": "TIFA [15] introduced a benchmark with images generated using different T2I models from identical prompts with human rankings. Our evaluation on this dataset reveals our model's superior alignment with human rankings, as shown in Table 4. Spearman's \\(\\rho\\) and Kendall's \\(\\tau\\) are statistical measures that evaluate the similarity between two ranking orders, with higher values indicating greater alignment. Figure 7 illustrates qualitative ranking results using our model for the generation results from a T2I model [35]." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.639, + 0.36, + 0.654 + ], + "angle": 0, + "content": "5 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.659, + 0.788, + 0.841 + ], + "angle": 0, + "content": "In this paper, we introduced an innovative method for generating high-quality training data for image-text alignment models. By implementing a mixed-type negative caption creation strategy and a novel filtering mechanism, we ensured a balanced distribution between positive and negative captions. This approach not only rectified the biases in existing models but also significantly enhanced the performance of our fine-tuned visual-language models, which achieve state-of-the-art results for image-text alignment. Our findings underscore the importance of high-quality, balanced training data in improving visual-language compositional reasoning and alignment. They also open up new avenues for exploring how these methods can be applied to other tasks or modalities in addition to the visual-language domain. In conclusion, our work offers an effective approach to improving image-text alignment, paving the road for future alignment studies." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.643, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "LLaVA-score" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.254 + ], + "angle": 0, + "content": "This work was supported in part by NSF CAREER IIS2150012, Adobe Data Science award, Microsoft Accelerate Foundation Models Research Program, and Institute of Information & communications Technology Planning & Evaluation (IITP) grants funded by the Korea government (MSIT) (No. 2022-0-00871, Development of AI Autonomy and Knowledge Enhancement for AI Agent Collaboration) and (No. RS-2022-00187238, Development of Large Korean Language Model Technology for Efficient Pre-training)." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.278, + 0.325, + 0.294 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.31, + 0.785, + 0.351 + ], + "angle": 0, + "content": "1. Alayrac, J.B., Donahue, J., Luc, P., Miech, A., Barr, I., Hasson, Y., Lenc, K., Mensch, A., Millican, K., Reynolds, M., et al.: Flamingo: a visual language model for few-shot learning. NeurIPS 35, 23716-23736 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.354, + 0.785, + 0.407 + ], + "angle": 0, + "content": "2. Betker, J., Goh, G., Jing, L., Brooks, T., Wang, J., Li, L., Ouyang, L., Zhuang, J., Lee, J., Guo, Y., Manassra, W., Dhariwal, P., Chu, C., Jiao, Y., Ramesh, A.: Improving image generation with better captions (2023), https://cdn.openai.com/papers/dall-e-3.pdf" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.409, + 0.785, + 0.435 + ], + "angle": 0, + "content": "3. Bitton, Y.: WYSIWYR repository. https://github.com/yonatanbitton/wysiwyr (2023), accessed: 2024-02-24" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.437, + 0.785, + 0.476 + ], + "angle": 0, + "content": "4. Changpinyo, S., Sharma, P., Ding, N., Soricut, R.: Conceptual 12M: Pushing web-scale image-text pre-training to recognize long-tail visual concepts. In: CVPR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.479, + 0.785, + 0.506 + ], + "angle": 0, + "content": "5. Chen, J., Guo, H., Yi, K., Li, B., Elhoseiny, M.: Visualgpt: Data-efficient adaptation of pretrained language models for image captioning (June 2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.508, + 0.785, + 0.603 + ], + "angle": 0, + "content": "6. Chen, X., Wang, X., Changpinyo, S., Piergiovanni, A.J., Padlewski, P., Salz, D.M., Goodman, S., Grycner, A., Mustafa, B., Beyer, L., Kolesnikov, A., Puigcerver, J., Ding, N., Rong, K., Akbari, H., Mishra, G., Xue, L., Thapliyal, A.V., Bradbury, J., Kuo, W., Seyedhosseini, M., Jia, C., Ayan, B.K., Riquelme, C., Steiner, A., Angelova, A., Zhai, X., Houlsby, N., Soricut, R.: Pali: A jointly-scaled multilingual language-image model. ArXiv abs/2209.06794 (2022), https://apisemantic scholar.org/CorpusID:252222320" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.605, + 0.785, + 0.644 + ], + "angle": 0, + "content": "7. Devlin, J., Chang, M.W., Lee, K., Toutanova, K.: Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.647, + 0.785, + 0.7 + ], + "angle": 0, + "content": "8. Doveh, S., Arbelle, A., Harary, S., Herzig, R., Kim, D., Cascante-Bonilla, P., Alfassy, A., Panda, R., Giryes, R., Feris, R., et al.: Dense and aligned captions (dac) promote compositional reasoning in vl models. Advances in Neural Information Processing Systems 36 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.702, + 0.785, + 0.757 + ], + "angle": 0, + "content": "9. Doveh, S., Arbelle, A., Harary, S., Schwartz, E., Herzig, R., Giryes, R., Feris, R., Panda, R., Ullman, S., Karlinsky, L.: Teaching structured vision & language concepts to vision & language models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2657-2668 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.758, + 0.785, + 0.813 + ], + "angle": 0, + "content": "10. Feng, W., He, X., Fu, T.J., Jampani, V., Akula, A.R., Narayana, P., Basu, S., Wang, X.E., Wang, W.Y.: Training-free structured diffusion guidance for compositional text-to-image synthesis. In: The Eleventh International Conference on Learning Representations (2023), https://openreview.net/forum?id=PUIqjT4rzq7" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.814, + 0.785, + 0.84 + ], + "angle": 0, + "content": "11. Gadre, S.Y., Ilharco, G., Fang, A., Hayase, J., Smyrnis, G., Nguyen, T., Marten, R., Wortsman, M., Ghosh, D., Zhang, J., et al.: Datacomp: In search of the next" + }, + { + "type": "list", + "bbox": [ + 0.22, + 0.31, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.347, + 0.127 + ], + "angle": 0, + "content": "Y. Li et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.245, + 0.148, + 0.785, + 0.175 + ], + "angle": 0, + "content": "generation of multimodal datasets. Advances in Neural Information Processing Systems 36 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.177, + 0.785, + 0.217 + ], + "angle": 0, + "content": "12. Herzig, R., Mendelson, A., Karlinsky, L., Arbelle, A., Feris, R., Darrell, T., Globerson, A.: Incorporating structured representations into pretrained vision & language models using scene graphs. EMNLP (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.218, + 0.785, + 0.313 + ], + "angle": 0, + "content": "13. Hessel, J., Holtzman, A., Forbes, M., Le Bras, R., Choi, Y.: CLIPScore: A reference-free evaluation metric for image captioning. In: Moens, M.F., Huang, X., Specia, L., Yih, S.W.t. (eds.) Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing. pp. 7514-7528. Association for Computational Linguistics, Online and Punta Cana, Dominican Republic (Nov 2021). https://doi.org/10.18653/v1/2021.emnlp-main.595, https://aclanthology.org/2021.emnlp-main.595" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.315, + 0.785, + 0.369 + ], + "angle": 0, + "content": "14. Hsieh, C.Y., Zhang, J., Ma, Z., Kembhavi, A., Krishna, R.: Sugarcrepe: Fixing hackable benchmarks for vision-language compositionality. In: Thirty-Seventh Conference on Neural Information Processing Systems Datasets and Benchmarks Track (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.37, + 0.785, + 0.411 + ], + "angle": 0, + "content": "15. Hu, Y., Liu, B., Kasai, J., Wang, Y., Ostendorf, M., Krishna, R., Smith, N.A.: Tifa: Accurate and interpretable text-to-image faithfulness evaluation with question answering. ICCV (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.412, + 0.785, + 0.452 + ], + "angle": 0, + "content": "16. Karthik, S., Roth, K., Mancini, M., Akata, Z.: If at first you don't succeed, try, try again: Faithful diffusion-based text-to-image generation by selection. arXiv preprint arXiv:2305.13308 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.453, + 0.785, + 0.493 + ], + "angle": 0, + "content": "17. Le, T., Lal, V., Howard, P.: Coco-counterfactuals: Automatically constructed counterfactual examples for image-text pairs. Advances in Neural Information Processing Systems 36 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.495, + 0.785, + 0.548 + ], + "angle": 0, + "content": "18. Li, J., Li, D., Savarese, S., Hoi, S.C.H.: Blip-2: Bootstrapping language-image pretraining with frozen image encoders and large language models. In: International Conference on Machine Learning (2023), https://apisemantic scholar.org/CorpusID:256390509" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.55, + 0.785, + 0.591 + ], + "angle": 0, + "content": "19. Li, J., Li, D., Xiong, C., Hoi, S.: Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In: International Conference on Machine Learning. pp. 12888-12900. PMLR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.592, + 0.785, + 0.619 + ], + "angle": 0, + "content": "20. Li, Y., Du, Y., Zhou, K., Wang, J., Zhao, W.X., Wen, J.R.: Evaluating object hallucination in large vision-language models. EMNLP (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.62, + 0.785, + 0.674 + ], + "angle": 0, + "content": "21. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dólar, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13. pp. 740-755. Springer (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.675, + 0.785, + 0.702 + ], + "angle": 0, + "content": "22. Lin, Z., Chen, X., Pathak, D., Zhang, P., Ramanan, D.: Revisiting the role of language priors in vision-language models. arXiv preprint arXiv:2306.01879 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.703, + 0.785, + 0.743 + ], + "angle": 0, + "content": "23. Lin, Z., Chen, X., Pathak, D., Zhang, P., Ramanan, D.: Visualgptscore: Visio-linguistic reasoning with multimodal generative pre-training scores. arXiv preprint arXiv:2306.01879 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.744, + 0.785, + 0.771 + ], + "angle": 0, + "content": "24. Liu, H., Li, C., Li, Y., Lee, Y.J.: Improved baselines with visual instruction tuning (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.772, + 0.785, + 0.799 + ], + "angle": 0, + "content": "25. Liu, H., Li, C., Wu, Q., Lee, Y.J.: Visual instruction tuning. arXiv:2304.08485 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.8, + 0.785, + 0.84 + ], + "angle": 0, + "content": "26. Ma, Z., Hong, J., Gul, M.O., Gandhi, M., Gao, I., Krishna, R.: Crepe: Can vision-language foundation models reason compositionally? arXiv preprint arXiv:2212.07796 (2022)" + }, + { + "type": "list", + "bbox": [ + 0.218, + 0.148, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.643, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "LLaVA-score" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.148, + 0.785, + 0.175 + ], + "angle": 0, + "content": "27. OpenAI: Gpt-4v(ision) system card (2023), https://cdn.openai.com/papers/GPTV_System_Card.pdf" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.177, + 0.785, + 0.23 + ], + "angle": 0, + "content": "28. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.232, + 0.785, + 0.272 + ], + "angle": 0, + "content": "29. Ramesh, A., Dhariwal, P., Nichol, A., Chu, C., Chen, M.: Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.273, + 0.785, + 0.313 + ], + "angle": 0, + "content": "30. Ray, A., Radenovic, F., Dubey, A., Plummer, B.A., Krishna, R., Saenko, K.: Cola: How to adapt vision-language models to compose objects localized with attributes? (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.315, + 0.785, + 0.355 + ], + "angle": 0, + "content": "31. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10684-10695 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.357, + 0.785, + 0.411 + ], + "angle": 0, + "content": "32. Schuhmann, C., Beaumont, R., Vencu, R., Gordon, C., Wightman, R., Cherti, M., Coombes, T., Katta, A., Mullis, C., Wortsman, M., et al.: Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems 35, 25278-25294 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.412, + 0.785, + 0.493 + ], + "angle": 0, + "content": "33. Sharma, P., Ding, N., Goodman, S., Soricut, R.: Conceptual captions: A cleaned, hypernymed, image alt-text dataset for automatic image captioning. In: Gurevych, I., Miyao, Y. (eds.) Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). pp. 2556-2565. Association for Computational Linguistics, Melbourne, Australia (Jul 2018). https://doi.org/10.18653/v1/P18-1238, https://aclanthology.org/P18-1238" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.495, + 0.785, + 0.549 + ], + "angle": 0, + "content": "34. Singh, A., Hu, R., Goswami, V., Couairon, G., Galuba, W., Rohrbach, M., Kiela, D.: Flava: A foundational language and vision alignment model. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15638-15650 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.55, + 0.663, + 0.563 + ], + "angle": 0, + "content": "35. stabilityai: IF repository. https://github.com/deep-floyd/IF" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.565, + 0.785, + 0.618 + ], + "angle": 0, + "content": "36. Thrush, T., Jiang, R., Bartolo, M., Singh, A., Williams, A., Kiela, D., Ross, C.: Winoground: Probing vision and language models for visio-linguistic compositionality. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5238-5248 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.62, + 0.785, + 0.66 + ], + "angle": 0, + "content": "37. Wang, Z.J., Montoya, E., Munechika, D., Yang, H., Hoover, B., Chau, D.H.: Large-scale prompt gallery dataset for text-to-image generative models. arXiv:2210.14896 [cs] (2022), https://arxiv.org/abs/2210.14896" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.661, + 0.785, + 0.702 + ], + "angle": 0, + "content": "38. Xu, J., Liu, X., Wu, Y., Tong, Y., Li, Q., Ding, M., Tang, J., Dong, Y.: Imagereward: Learning and evaluating human preferences for text-to-image generation. Advances in Neural Information Processing Systems 36 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.703, + 0.785, + 0.742 + ], + "angle": 0, + "content": "39. Yarom, M., Bitton, Y., Changpinyo, S., Aharoni, R., Herzig, J., Lang, O., Ofek, E., Szpektor, I.: What you see is what you read? improving text-image alignment evaluation. arXiv preprint arXiv:2305.10400 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.744, + 0.785, + 0.797 + ], + "angle": 0, + "content": "40. Yuksekgonul, M., Bianchi, F., Kalluri, P., Jurafsky, D., Zou, J.: When and why vision-language models behave like bags-of-words, and what to do about it? In: International Conference on Learning Representations (2023), https://openreview.net/forum?id=KRLUvxh8uaX" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.799, + 0.785, + 0.84 + ], + "angle": 0, + "content": "41. Zhang, K., Mo, L., Chen, W., Sun, H., Su, Y.: Magicbrush: A manually annotated dataset for instruction-guided image editing. In: Advances in Neural Information Processing Systems (2023)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.148, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.346, + 0.127 + ], + "angle": 0, + "content": "Y. Li et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.787, + 0.19 + ], + "angle": 0, + "content": "42. Zhao, T., Zhang, T., Zhu, M., Shen, H., Lee, K., Lu, X., Yin, J.: Vl-checklist: Evaluating pre-trained vision-language models with objects, attributes and relations. arXiv preprint arXiv:2207.00221 (2022)" + } + ] +] \ No newline at end of file diff --git a/2024/Removing Distributional Discrepancies in Captions Improves Image-Text Alignment/0409d8d6-db45-4397-97c7-eeaa2a4590eb_origin.pdf b/2024/Removing Distributional Discrepancies in Captions Improves Image-Text Alignment/0409d8d6-db45-4397-97c7-eeaa2a4590eb_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..887e42a6595c0b909e69e2ce3aa0c3b12a32ed5e --- /dev/null +++ b/2024/Removing Distributional Discrepancies in Captions Improves Image-Text Alignment/0409d8d6-db45-4397-97c7-eeaa2a4590eb_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4a20d5dc150f4110590d946bef7212b1199922861c81b47764e9a21de02d3a7c +size 3870812 diff --git a/2024/Removing Distributional Discrepancies in Captions Improves Image-Text Alignment/full.md b/2024/Removing Distributional Discrepancies in Captions Improves Image-Text Alignment/full.md new file mode 100644 index 0000000000000000000000000000000000000000..233015c0af24eb03d28c2d6eb271a36f894acccf --- /dev/null +++ b/2024/Removing Distributional Discrepancies in Captions Improves Image-Text Alignment/full.md @@ -0,0 +1,301 @@ +# Removing Distributional Discrepancies in Captions Improves Image-Text Alignment + +Yuheng Li $^{1}$ , Haotian Liu $^{1}$ , Mu Cai $^{1}$ , Yijun Li $^{2}$ , Eli Shechtman $^{2}$ , Zhe Lin $^{2}$ , Yong Jae Lee $^{1}$ , and Krishna Kumar Singh $^{2}$ + +1 University of Wisconsin-Madison + +2 Adobe Research + +Abstract. In this paper, we introduce a model designed to improve the prediction of image-text alignment, targeting the challenge of compositional understanding in current visual-language models. Our approach focuses on generating high-quality training datasets for the alignment task by producing mixed-type negative captions derived from positive ones. Critically, we address the distribution imbalance between positive and negative captions to ensure that the alignment model does not depend solely on textual information but also considers the associated images for predicting alignment accurately. By creating this enhanced training data, we fine-tune an existing leading visual-language model to boost its capability in understanding alignment. Our model significantly outperforms current top-performing methods across various datasets. We also demonstrate the applicability of our model by ranking the images generated by text-to-image models based on text alignment. Project page: https://yuheng-li.github.io/LLaVA-score/ + +# 1 Introduction + +Recent years have seen rapid advances in multimodal research, encompassing both visual generation [2, 29, 31] and visual understanding [1, 25, 27]. Training capable multimodal models generally requires extensive datasets of image-text pairs, e.g., LAION [32], CC12M [4], and others [11, 33], which are collected on a web-scale and thus tend to be noisy. This noise in the data contributes to certain challenges. For instance, vision-language models often struggle with hallucination [20] and face difficulties in mastering compositional reasoning [36, 40]. Similarly, text-to-image models frequently fail to generate accurate images when processing complex sentence prompts [10]. + +Given these challenges, the ability to automatically assess whether an image and a caption are semantically aligned plays a crucial role. This capability is essential not only for cleaning the data used to pre-train these models but also for evaluating and enhancing the performance of both text-to-image and image-to-text generation models. The most frequently employed metric for this task is the CLIP score [13], which calculates the cosine distance between the CLIP [28] embeddings of the paired text and image. However, it has been observed that CLIP, along with other methods such as BLIP [19] and Flava [34], tends to + +![](images/6acb0a818d1a1ae2c8cdce9e818099d9b54f72c81303aaaf2d1328cfa679878b.jpg) +Fig.1: Left: Qualitative examples for image-text alignment prediction, where our approach can distinguish fundamental concepts such as positioning, counting, and attributes. Right: Our approach shows superior performance in image-text alignment. + +operate on a bag-of-words basis. They sometimes even cannot tell the difference between simple cases, such as horse eating grass versus grass eating horse [14, 23, 40]. + +The conventional approach to training image-text alignment models involves generating negative captions that are misaligned with the images, which are paired with the original positive captions as the training data. For example, prior works randomly shuffle words to create negative captions [40], or employ language models to generate coherent negative sentences [39]. We follow the latter and leverage an LLM to generate a mix of different types of negative captions during training. Specifically, this includes the replacing type, where one linguistic element is substituted with an arbitrary counterpart (e.g. 'a knife is on the table' $\rightarrow$ 'a spoon is on the table'), and the swapping type, where words within the same sentence are rearranged (e.g. 'an apple is to the left of a banana' $\rightarrow$ 'a banana is to the left of an apple'). The former type can aid in enhancing an image-text alignment model's perceptual skills as it needs to distinguish between the original and replaced elements, while the latter type can help the model's reasoning capabilities as elements in the caption remain the same but the relationship between them changes. + +However, critically, we find that this standard approach of generating negative captions, and ensuring that they have coherence (e.g., proper grammar) on a per-instance level, is insufficient. In particular, this approach cannot ensure consistency at the distribution level between positive and negative captions. Distributional biases may originate from the initial dataset or from the rules or models employed to generate negative captions. For instance, the COCO dataset [17] contains significantly more captions including the word giraffe compared to elephant. Yet, when using GPT to generate negative captions, we observe a tendency for GPT to substitute giraffe with elephant, resulting in a surplus of giraffe mentions in positive captions but more elephant in negative captions. Unfortunately, this means that an image-text alignment model trained on such data can be biased to predict sentences containing elephant to be a positive caption and those that contain giraffe to be a negative caption, independent of the paired image. To address this imbalance, our approach consists of fine-tuning a classifier on text captions only (without paired image inputs) to remove biased data that + +the text classifier can correctly predict with high confidence. It's important to note that this bias is not unique to the GPT model or how one prompts it, but can arise in other rule-based methods or pre-trained models as well. + +While previous research [14] noticed distribution differences due to implausible and non-fluent negative captions, our work is the first to eliminate the dataset-level distribution differences using a language model to implicitly encompass all linguistic aspects, without the need for explicit identification. This approach allows us to incorporate considerations like word frequency which is not identified beforehand, thereby providing a more holistic optimization over the dataset's distribution. + +For our image-text alignment model, we leverage a state-of-the-art vision-language model, LLaVA [24, 25], and finetune it with our curated data. Our model achieves significantly better results compared with other image-text alignment models, demonstrated in Figure 1. We also demonstrate that our curated data can improve other models like BLIP2 [18]. Finally, we further show that our image-text alignment model can help with other vision-language tasks like ranking generated images from T2I models [15]. In summary, we have four main contributions: + +- We identify new dataset-level distribution differences between positive and negative captions that lead to biased image-text alignment models. +- To address this, we propose a method to maintain consistency between positive and negative caption distributions, which is critical to ensure that an image-text alignment model relies on both image and text to measure alignment instead of only the biases present in the text. +- We use our curated training data to finetune existing visual-language models like LLaVA, and obtain state-of-the-art results for image-text alignment. +- In addition, we demonstrate the application of our image-text alignment model in ranking the image generations produced by generative models. + +# 2 Related Work + +Challenges in Compositional Understanding. Image-text pairs form a crucial interface between visual and linguistic modalities, thus evaluating if a given image-text pair is aligned is important for both data curation and model performance evaluation. The pioneering work of CLIP [28] demonstrated this by leveraging an extensive corpus of such pairs for image-text contrastive training. More recent works such as BLIP [18, 19] and LLaVA [24, 25] further utilize large language models (LLMs) to achieve image-text alignment via the text generation objective. Such models usually inherit the frozen CLIP visual encoder to produce a set of visual tokens, and then feed such tokens and the language instructions into the LLM. + +It has been observed that vision-language models like CLIP have limited capability in understanding compositionality [26,30,36,40,42]. Specifically, they find it challenging to recognize the permutation of words within sentences [36]. + +Moreover, these models often struggle to identify the binding of attributes to multiple objects within a single sentence or to discern the relationships between objects [30,40,42]. While models built on top of large language models (LLMs), such as BLIP [19] and LLaVA [25], exhibit improved understanding capabilities, they still face difficulties with complex compositional understanding largely due to their lack of training on sufficiently challenging data. + +Enhancing Vision-Language Compositionality. Several studies focus on improving models' understanding of language compositionality [8,9,12,17,23,39, 40]. Some strategies [15,39] involve using language models to decompose text into multiple succinct assertion phrases, which are then evaluated by VQA models on an individual basis. [23] also finds that assessing the conditional probability of predicting text based on the input image offers more accurate outcomes than traditional discriminative approaches like contrastive or classification scores used in BLIP [19]. Still, the most widely used method remains the explicit fine-tuning of models to differentiate between hard negatives and correct captions [9, 39, 40]. [40] randomly shuffles words to generate negative captions, but subsequent analysis by [14] points out the approach's flaw: models could simply rely on textual cues (e.g., grammar correctness) for predictions. [39] attempts to employ LLMs to create negative captions, ensuring the negative captions are both fluent and meaningful. Nonetheless, we observed that current fine-tuning methods do not generate a diverse range of negative prompts. Moreover, merely evaluating the grammar or logical coherence of negative captions is insufficient to eliminate data bias in the distribution of negative captions. + +# 3 Approach + +In our approach, we assume access to images accompanied by accurately labeled positive captions, similar to those found in the MSCOCO [21] dataset. Following [14], contrary to adopting rule-based techniques that often generate illogical sentences, our method utilizes large language models (LLMs) like GPT4 to transform positive captions into negative ones. We first outline our strategy for generating various types of negative captions and then present a straightforward technique to mitigate biases inherent in the distributions of positive and negative captions. We show the entire pipeline in Figure 2. + +# 3.1 Constructing Diverse Negative Captions + +Suppose we have a dataset consisting of image-text pairs $\{I, T_p\}$ , where $I$ represents an image and $T_p$ is its associated positive caption. Previous research [39,40] showed that merely randomly shuffling image-text pairs to generate negative samples is insufficient to learn capable vision-language models that properly understand the structure of a caption and its relationship with the image. These studies have underscored the value of constructing hard negative captions to significantly improve the model's language compositional abilities. Following this + +![](images/5c5a4dbefb387151ee8e67fcb178f77af6bb8724c00034bc67ff276e97a68e60.jpg) +Fig. 2: Top: We feed the positive caption (green dots) into GPT to create two types of negative captions (red dots): substituting one linguistic element with any plausible alternative or swapping the positions of two components. The blue part in negative captions highlights the modifications. Bottom: we remove easy negative samples using only text data and utilize the remaining samples to fine-tune vision-language models. + +insight [39], we utilize large language models (LLMs) to generate such hard negative captions. Specifically, we create two types of hard negative captions. + +We refer to the first method for creating negative captions as the replacing strategy, which identifies key components in a language and uses a language model to replace it with other plausible substitutes. The replaced component can be any linguistic part such as a noun, adjective, preposition, etc. For example, "a photo of a broken down stop sign" could be replaced with "a photo of a brand new stop sign"; "a cute cat looking at a bird" could be changed to "a cute dog looking at a bird". Conceptually, this type of negative caption aims to enhance the model's recognition capabilities. For executing this task, we engage GPT by providing it with specific instructions followed by some contextual examples, as illustrated in Figure 3. + +We refer to the second method for creating negative captions as the swapping strategy. This approach involves generating a new sentence by utilizing the original language components, typically resulting in the swapping of two or several identical linguistic elements. Specifically, we first ask GPT to break down the original positive sentence into its key components, and then request the model to construct a new sentence with these elements. For example, if the input caption is "an airplane is flying in the blue sky", GPT is tasked with identifying the key components including "airplane", "flying", "blue", and "sky". The newly crafted sentence could be "a blue airplane is flying in the sky". Please note that for some short captions, there may not be enough language elements to form a reasonably different sentence in meaning. Thus, we also need to judge if the new sentence makes sense or not. Please refer to the Figure 3 for our prompt used in GPT. + +# 3.2 Addressing Distribution Discrepancies + +Previous studies [14] have demonstrated that artifact issues arising from incorrect construction in negative captions can lead to distribution disparities between positive and negative captions. To mitigate this, commonsense and grammar + +![](images/979ada63a8130fe6e8e785fb495608efd50a0c44b36ea47277e1b5dfd9216ec9.jpg) +Fig. 3: Prompts used in GPT for generating two types of negative captions, with in-context examples shown in purple. + +![](images/2a62d04cf16e045d8b6f645658bc1adedf3e2484b619fd58de450e9823cada58.jpg) +Fig. 4: Top prediction based on a text-only binary classifier. Top left: Negative captions generated through replacement strategy. Top right: Negative captions generated through swapping strategy. Bottom: Positive captions from the COCO dataset. + +models are employed. Nonetheless, our findings indicate that merely assessing the sensibility of sentences is insufficient for aligning the distributions of positive and negative sentences because of two following two reasons. + +First, each image-pair dataset inherently establishes its own unique distribution. For instance, the COCO dataset [21], despite its widespread use and diverse content, predominantly encompasses everyday scenes, objects, and activities. Common subjects such as people and motorcycles, along with frequent activities like skiing and surfing, characterize the distinct distribution of COCO. Second, when employing rule-based methods or prompting language models to generate negative captions, any strategy adopted will inevitably reflect biases from human preconceptions or pre-trained models. This leads to negative captions originating from a different distribution. + +To illustrate this concept, we develop a blind text-only binary classifier that processes positive and negative captions without seeing any images. Surprisingly, we find that despite the coherence and logical structure of negative captions, the classifier is adept at distinguishing between negative and positive captions based solely on the text. Figure 4 shows the most confidently correct predictions for both our replace and swap data categories constructed from COCO. + +In the dataset with replaced negative data, it is observed that GPT frequently substitutes the terms "boat" and "elephant" for the original COCO captions' "airplane" and "giraffe", respectively. This pattern makes it straightforward for a text-only model to identify these captions as negative. In our swap data, the text-only classifier is able to identify captions with colorful animals interacting with a black object as negative. Note that these negative prompts are logically coherent and grammatically correct, thus cannot be easily detected by a grammar model. Conversely, the top accurately predicted positive captions often depict common activities like surfing and motorcycle riding, which align closely with the typical content style of COCO. It's important to recognize that data bias is not solely a characteristic of our method for generating data nor the manner in which we prompt GPT. In the experimental section, we demonstrate that such distribution discrepancy is also present in other approaches [14] that employ GPT for creating negative prompts. + +The presence of bias in the data can obstruct a vision-language model's ability to truly understand image structures and learn language compositions, as the model might rely solely on textual cues for making predictions. To address this issue, we propose a straightforward solution aimed at reducing data distribution differences by selectively removing data that is predicted with high confidence by a text-alone model. + +Figure 2 (bottom) depicts our conceptual goal of data filtering: to eliminate straightforward or biased positive and negative samples. This is done to ensure that the remaining text data cannot be distinguished by any text classifier, based solely on the text information. Essentially, our aim is to maximize the entropy of the text information within our dataset. In practice, we organize our dataset into $N$ partitions. For each iteration, one partition is designated as the test set while the remaining $N - 1$ partitions serve as the training set. We employ a pretrained Bert model [7] as our text-only classifier, and train it on the training set. Subsequently, this trained classifier is applied to the designated test set. We then rank the correctly predicted samples by the classifier's confidence level, removing the top $k\%$ of these samples for both positive and negative class predictions. The rest of the data is retained as our refined dataset. This procedure is repeated for each partition, ensuring a comprehensive reduction of bias across the dataset. + +Note that while our observation and data filtering approach is applied to address the image-text alignment issue, it could be a more general problem for any multimodal data scenario where bias in one modality might negatively impact model training. + +# 3.3 Finetuning VLMs for Image-Text Alignment Scoring + +Once we have the unbiased data, we opt to fine-tune a vision-language model (VLM) to enhance its language compositional understanding capabilities with respect to images. In our study, we mainly select LLaVA-1.5 [24] due to its superior performance in image and text understanding. Since LLaVA-1.5 is designed to generate text, to adapt it for use as a image-text score calculator, we employ + +the following prompt formatting: "Does this image I match the following caption $T$ . Answer Yes or No directly." Given that LLaVA relies on a language model to produce subsequent words, we manually extract the logits associated with the responses Yes and No for the next word. We then define the matching score as: + +$$ +\frac {e ^ {\mathbf {P} (\text {Y e s} | \text {p r o m p t})}}{e ^ {\mathbf {P} (\text {Y e s} | \text {p r o m p t})} + e ^ {\mathbf {P} (\text {N o} | \text {p r o m p t})}} \tag {1} +$$ + +We discover that this straightforward approach is quite effective and can already outperform many existing state-of-the-art baselines. However, LLaVA-1.5 was not specifically trained for this type of matching problem. Therefore, to enhance its performance, we finetune it with the same prompt formatting introduced above using our curated data. We assign the labels Yes and No to positive and negative pairs, respectively. + +Our curated dataset is not restricted to LLaVA-1.5. In our experiments, we also finetune a Q-former equipped with the Image-Text Matching (ITM) head in BLIP2 [18]. The ITM head is essentially a binary classification layer, identical to our dataset's structure. We thus fine-tune the model using the standard cross-entropy loss. + +# 4 Results + +We evaluate our fine-tuned LLaVA-1.5 model against various baselines across different datasets. Additionally, we conduct ablation studies to evaluate the impact and importance of our dual-strategy approach for generating diverse negative captions and our method for addressing data distribution discrepancies. Since our main model is built upon LLaVA-1.5, we name our score as LLaVA-score. + +# 4.1Baselines + +We evaluate our model against a range of leading multimodal models: + +CLIP-ViT-L/14, the largest variant of OpenAI CLIP models [28]. + +BLIP-2 [18], which incorporates both image-text matching and image-text contrastive learning approaches. + +NegCLIP, as introduced in [40], fine-tuned on challenging negative captions generated by randomly shuffling words within sentences. + +VisualGPTScore [22] proposes utilizing the probability of generating specific text given an image (i.e., $P(T|I)$ ) as an effective metric for calculating image-text alignment scores. For this, we utilize LLaVA-1.5 [24], the state-of-the-art vision-language model, to calculate their proposed VisualGPTScore, employing a reweighting technique as suggested by [22]. + +Image-Reward [38]: A reward model, trained on human preferences of image-text pairs produced by Text-to-Image (T2I) models in DiffusionDB [37]. + +VQ2 [39] first extracts a set of candidate QA pairs from the text, then uses a VQA model to score each pair. For this baseline, we report results using both the PaLI model [6]—directly citing their paper as the model is not publicly + +available—and BLIP-T5 [18] which is accessible via their official GitHub repository [3]. Note that while TIFA [15] similarly utilizes a QA pairs approach, we exclude it from our results due to VQ2 demonstrating superior performance, aiming for simplicity in our comparison. + +PaLI fine-tuned on the SeeTrue dataset [39]: A version of PaLI specifically fine-tuned for the alignment task, using a curated dataset. Performance metrics are cited directly from the original paper [39]. + +LLaVA-1.5 [24], Originally a text generation model, we employ a specific prompt as introduced in Sec. 3.3. By using Equation 1, we convert it into a scoring function. Through empirical testing, this prompt is found to be the most effective, and it is utilized to finetune our model. + +# 4.2 Datasets and Metrics + +We evaluate on the following datasets and metrics: + +Winoground [36]: This dataset uniquely comprises quartets, each including two images and two texts, necessitating a nuanced interpretation of both linguistic and visual elements for accurate matching. The metrics reported for this dataset encompass an image score, a text score, and a group score. + +SeeTRUe [39]: A benchmark designed for assessing vision-language models, featuring a test set combining multiple sources. Current sources include Drawbench, EditBench, and COCO-t2i, pairing real texts with synthetic images. Following [39], we utilize ROC-AUC as the evaluation metric. + +SugarCrepe [14]: A recently introduced benchmark focuses on generating creative negative captions using a language model and employs grammar and commonsense models for data cleaning, marking a novel approach in benchmark design. + +MagicBrush [41]: This benchmark facilitates human-driven image editing, constructed using Dall-E 2, and includes captions for both the original and edited images. Featuring quartets similar to Winoground [36], we adopt the same methodology for calculating the group score with modification due to nature of data. Refer supp for details. + +# 4.3 Implementation Details + +For our dataset curation, we select COCO, which provides positive image-text pairs. To diversify our image dataset and enhance our model's robustness to synthetic images, we incorporate a subset of the training images from SeeTRUE [39], specifically the coco_train_t2i set. To construct negative data, we use GPT to generate two types of negative captions (swap and replace), and we perform random sampling to maintain an equal amount of positive and negative data. In the process of curating our dataset, we ensure that neither images nor captions included in the training dataset are present within the test dataset. + +For our model development, we choose to refine LLaVA-1.5 [24] using our curated data. We train the model with a batch size of 64 on 8 NVIDIA A100 GPUs for a single epoch, setting the learning rate at 2e-6. For our data filtering, we use $k$ to be $30\%$ and $N$ to be 5. Find more details in the supp. + +Table 1: We evaluate our model alongside baselines across multiple datasets. The best results are shown in bold. For three subsets of the SeeTRUE dataset, we present the ROC AUC scores following the original paper [39]. For the SugarCrepe dataset, accuracy is employed as the performance metric. For the MagicBrush [41] dataset, we report the group score. We use LLaVA-1.5 to calculate the VisualGPT score [23]. + +
Winoground image text groupSeeTRUE DrawBench EditBench COCO-T2ISugarCrepe replace swap addMagicBrush
Chance Performance25.0025.0016.6750.050.050.050.050.050.033.33
CLIP-ViT-L-14 [28]10.5028.507.7561.462.159.279.461.474.852.89
NegCLIP [40]11.7530.758.2563.266.062.885.375.387.261.12
BLIP2-ITM [18]24.2541.7519.0060.867.568.088.983.991.875.32
BLIP2-ITC [18]12.0028.508.5064.967.969.986.766.992.367.85
Image-Reward [38]15.2543.0012.7570.470.277.088.281.095.270.28
VisualGPT [23]37.0044.2527.5077.074.269.188.287.195.578.31
VQ2 (PaLI) [39]42.2547.0030.5082.673.683.4----
VQ2 (BLIP-T5) [39]34.0033.5023.2574.867.474.283.381.090.970.46
PaLI (ft on SeeTRUE) [39]38.0046.5028.7586.877.283.2----
LLaVA-1.5 [24]49.7551.0034.2586.978.384.593.588.395.882.61
LLaVA-score (Ours)68.0053.7547.2588.877.784.995.394.997.587.28
+ +# 4.4 Main results + +Table 1 presents a comparison between our models and various strong baselines across different datasets. It is evident that our model outperforms others in nearly all datasets. Interestingly, utilizing Eq. 1, LLaVA-1.5 zero-shot performance is the second-best method overall. Particularly for Winoground, a benchmark well-known for its challenges in visual and linguistic reasoning, some models such as CLIP, BLIP2, and Image-Reward perform at or below chance level. Our fine-tuned LLaVA-1.5 substantially enhancing its reasoning capabilities showing the importance of our strategy for training data curation. For VQ2, the VQA models (BLIP or PaLI) underperform our model due to their lack of training on challenging negative examples. The finetuned PaLI model also falls short, attributed to its training dataset's lack of diversity and absence of data filtering. Our model also shows impressive results on synthetic image benchmarks like SeeTRUE and MagicBrush. + +# 4.5 Performance on Attribute, Counting, Spatial Reasoning + +In the preceding section, we demonstrated that our model surpasses baseline models across various datasets. This section offers an alternative perspective on our model's performance, particularly in areas where both vision-language models and recent text-to-image (T2I) models encounter significant challenges. These challenges include multiple attributes binding, counting objects, and understanding spatial relationships between objects. Addressing these issues is crucial for both visual understanding and evaluating the capabilities of generative models. + +To quantify the improvements our model achieves in addressing these common challenges, we construct three specialized datasets. Specifically, we prompt GPT to generate scenarios involving attribute binding, object counting, and spatial relationships, providing in-context examples to guide the generation process. + +Table 2: Comparison of our model against baselines across three prevalent challenges, highlighting the highest average accuracy in bold. + +
AttributeCountingSpatial
avgposnegavgposnegavgposneg
Chance Performance505050505050505050
Threshold-Independent Models (with oracle)
CLIP-ViT-L-14 [28]635274586848531888
NegCLIP [40]657852596652574866
BLIP2-ITC [18]667260573678579024
VisualGPT [5]739056655278625668
Inherent Decision Models
BLIP2-ITM [18]5810016539610511002
Image-Reward [38]70100406110022579816
VQ2 (BLIP-T5) [39]669438568230549414
LLaVA-1.5 [24]719844629628579816
LLaVA-score (Ours)819066718656817686
+ +Subsequently, the T2I model [35] is employed to generate 50 images for each prompt. From these, we manually select one positive and one negative image based on image-text alignment. We opt for synthetic images as it offers greater flexibility, as illustrated in Figure 5, it can incorporate diverse styles (e.g., painting styles for birds), create unconventional images (e.g., the examples involving soccer and cats), and has attribute/object merging capabilities of T2I models to enrich our dataset with varied negative cases (e.g., the negative example of mistakenly combining clock and apple). + +Table 2 presents our results. We report classification accuracy as the metric, alongside showing both positive and negative results in gray color. We categorize the baseline models into two distinct groups for comparison purposes. 1) Upper section of the table, consists of models that generate scores without an established classification threshold. Consequently, we calculate the accuracy by using an oracle threshold, which involves optimizing over all predicted scores to find the best cutoff (cheating on the test set). 2) Lower part of the table, operates with an inherent decision boundary for classification. + +Our model demonstrates superior performance across all three challenging cases, beating competitors that perform only marginally better than random guessing, particularly on tasks requiring spatial understanding. It's noteworthy that many models exhibit a bias, often predicting the same outcome across the test dataset. This tendency is underscored by their high accuracy on positive samples contrasted with poor performance on negative ones. Given that our images are synthesized using text-to-image (T2I) models, essential linguistic elements are likely to be present even in images labeled as negative (refer to Figure 5). This observation suggests that baseline models operate similarly to a bag-of-visual-words approach, predicting an image as matching the textual description as long as it contains certain key visual concepts. + +![](images/4bb81f94169eb4a0073f657e01af448d9babb433f03df3bd08867e653360ea73.jpg) +An antique wooden clock beside a glossy red apple + +![](images/defbc6ffd1d28e4ef74d3908a2b5d9774449f89ea0b728b872406e055b848eed.jpg) +Attribute + +![](images/5783961fefc1047784d1ceee9ad8b089c093c3cbb3df20c4357d868f13d02fea.jpg) +Counting +A fishbowl with two goldfish + +![](images/3df66c9fae21c5d19b8333f1c2de6a46595c70639c389d09d66a86938f68d315.jpg) + +![](images/140f450c36c75dd216cfa816f271dc0116948543f0bf8ef8cb003a7ebf1d15c1.jpg) +Spatial +A soccer ball is embedded within a TV + +![](images/bd1a812394a82c2cb35f0fe3ed53f82414f82ed129b6a0ebcd3284f77fe1e60e.jpg) + +![](images/a4aa1dab4ba65cffbeb0ab77bcfb50a4e214155315ec1b729b026dd972d344b3.jpg) +A towering red cat under a sleek black rose bush +Fig. 5: Our curated test datasets feature captions paired with one positive image and one negative image each. All the positive images are displayed on the left side. + +![](images/4a1925bcbfdcec961c90dc671bdf47b0a0634854ccdb825522e27c61c8aa1bda.jpg) + +![](images/e67574e55fd1405d34d64ee845c7fb588e5f4fa65d830c645968ef72ca6df915.jpg) +A drawing of a bird perching on a branch with three leaves + +![](images/acaf7c4ff19ba7e0b34d962ecd7a6f5134d50536fece83fd46f6ffb9d2fdbe97.jpg) + +![](images/11d39e873742f7375db6fe9aef8e1bd92522319eade697bad63860142c29221b.jpg) +A cheeseburger on a plate with a soda on its right + +![](images/aed5b308a20204c690c9c5b0ba9c5bc4c9a14d306770935f492a092c68e45c46.jpg) + +# 4.6 Ablation Studies + +Importance of different negative data. Table 3 begins with our baseline model, LLaVA-1.5 [24], outlining its performance. The following two rows illustrate the effects of finetuning the LLaVA-1.5 model with only replace and swap negative data, with filtering processes applied. As shown in the subcategories of the SugarCrepe dataset, replace and swap data enhance performance in their respective categories but do not significantly impact other cases. Nevertheless, our final model excels across almost all tests, indicating that combining replace and swap negative data has a complementary effect, particularly noticeable in the add category of the SugarCrepe dataset and other evaluated datasets. + +Importance of data filtering. The subsequent row presents outcomes from merging two types of negative data without any filtering. This combination has shown to be beneficial for specific datasets, like Winoground and SugarCrepe, although the results don't match those of the final model with filtering. However, biases in the text data cause the fine-tuned model to show no improvement on datasets like MagicBrush and SeeTRUE. Since data filtering reduces the amount of data, the fifth row displays results from randomly subsampling a training set to match the quantity of filtered data. This comparison highlights the effectiveness of our filtering technique as without it, the model may get biased to just rely on language distribution and ignore the image when making predictions. + +Effect of data filtering. Fig. 6 shows the impact of removing the top $k\%$ of biased data on the performance across four test datasets. Here, $k$ varies from 0 to 90, representing the progression from no data removal to the exclusion of $90\%$ of the data. The observed trend indicates that as the biased data is progressively removed, performance improves, peaking at approximately $30\%$ to $40\%$ . Beyond this point, performance declines due to the diminishing volume of training data. + +We assess the quality of our filtered data by training a text-only model on $80\%$ of the filtered data and testing it on the remaining $20\%$ . Note that this evaluation is performed on filtered data, and a model should reach an accuracy close to $50\%$ in ideal data indicating it cannot distinguish between positive and negative caption just using text. We vary the parameter $k$ from $0\%$ to $90\%$ , in increments of $10\%$ , resulting in corresponding filtered data quality percentages + +Table 3: Upper: Ablating our training data based on LLaVA-1.5 [24]. Lower: Finetuning a BLIP2 model with our data. + +
WinogroundSeeTRUESugarCrepeMagicBrush
imagetextgroupDrawBenchEditBenchCOCO-T2Ireplaceswapadd
LLaVA-1.5 [24]49.7551.0034.2586.978.384.593.588.395.882.61
Only replace w/ filter51.2554.2538.5087.477.683.795.288.895.385.70
Only swap w/ filter63.5049.2542.2581.074.079.292.695.591.376.63
w/o filter65.7551.5046.7588.476.481.194.591.795.082.99
Random subsample64.2550.0043.2588.276.881.894.592.093.683.12
LLaVA-score (Ours)68.0053.7547.2588.877.784.995.394.997.587.28
BLIP2-ITM24.2541.7519.0060.867.568.088.983.991.875.32
BLIP2-ITM (ft on our data)39.542.752887.576.282.894.391.496.079.62
+ +![](images/526aa469eb4024741f40c97100b02a1cb14eed3bcf28a9c8eee45fe6239b7081.jpg) +Fig. 6: Ablation on filter percentage of data. For most of the data, performance peaks around $30\%$ and reduces after that due to decrease in training data size. + +![](images/a8e31fa643745535ba26272c6d8b0122535148ebeb055f7665a4d5101b544d35.jpg) + +![](images/650a56a2d95f5561be9efc467fcc7c3624336a358d483f5e21ff185fe5623d2a.jpg) + +![](images/c7dfa83c56298c6d5d9e5256c21db02319ad2bd9894d2556b6dd50397a80b702.jpg) + +of $75.9\%$ , $68.2\%$ , $60.7\%$ , $56.4\%$ , $53.8\%$ , $51.3\%$ , $50.6\%$ , $51.0\%$ , $49.8\%$ , and $47.9\%$ . Empirically, we find that data quality below $60\%$ is satisfactory for our purposes. Consequently, we decide to discard $30\%$ of the data in our experiment as it's a good trade-off between quality of filtered data and amount of training data. + +To demonstrate that the distribution gap we identified also exists in datasets created by others, we apply the same evaluation method to SugarCrepe [14], a test benchmark also generated by GPT, that was refined using grammar and common sense models. Nevertheless, it still exhibits an accuracy of $69.0\%$ , indicating the presence of the bias. + +Generalization on other model. The bottom section of Table 3 shows that when BLIP2 is fine-tuned with the ITM head (a binary classification head) using our curated data, there is an observable improvement in performance. It's important to highlight that we only finetune and use the Q-former in BLIP2 without including any LLMs, and this model surpasses nearly all baselines in benchmarks that do not utilize LLMs (refer table 1). As expected, the fine-tuned BLIP2 model does not outperform the model fine-tuned using LLaVA, particularly on the Winoground dataset, which is recognized for its challenging reasoning tasks. This suggests the importance of utilizing a more advanced vision-language model as the foundation model for achieving better results. Nonetheless, given that BLIP2's model size (180M) is significantly smaller compared to LLaVA-1.5 (13B), it offers a more lightweight option for simpler applications. + +# 4.7 Application: Text Alignment Ranking for Image Generation + +In this subsection, we showcase an application of our model for image generation. A major challenge with T2I models is their inconsistent adherence to complex + +![](images/ee0485ed87f1558edd4d3eb316beb9189231c49d98024acb91da0ed29d43b9cb.jpg) +Fig. 7: We show T2I generation results ranked by our model according to image-text alignment. The most aligned images are ranked higher. + +Table 4: Performance evaluation of ranking correlation on the TIFA [15] dataset. + +
CLIP-ViT-L-14 [28]NegCLIP [40]BLIP2-ITM [18]BLIP2-ITC [18]ImageReward [38]VisualGPT [23]VQ2(T5) [39]TIFA[15]LLaVA-1.5[24]Ours
Spearman ρ28.540.143.351.562.836.955.159.761.564.5
Kendall τ19.528.729.936.946.328.741.547.245.849.8
+ +compositional prompts [10]. Inspired by [16], one can generate a series of images from a T2I model and then employ our model to rerank them based on relevance, ultimately presenting the highest-ranked images to the user. + +TIFA [15] introduced a benchmark with images generated using different T2I models from identical prompts with human rankings. Our evaluation on this dataset reveals our model's superior alignment with human rankings, as shown in Table 4. Spearman's $\rho$ and Kendall's $\tau$ are statistical measures that evaluate the similarity between two ranking orders, with higher values indicating greater alignment. Figure 7 illustrates qualitative ranking results using our model for the generation results from a T2I model [35]. + +# 5 Conclusion + +In this paper, we introduced an innovative method for generating high-quality training data for image-text alignment models. By implementing a mixed-type negative caption creation strategy and a novel filtering mechanism, we ensured a balanced distribution between positive and negative captions. This approach not only rectified the biases in existing models but also significantly enhanced the performance of our fine-tuned visual-language models, which achieve state-of-the-art results for image-text alignment. Our findings underscore the importance of high-quality, balanced training data in improving visual-language compositional reasoning and alignment. They also open up new avenues for exploring how these methods can be applied to other tasks or modalities in addition to the visual-language domain. In conclusion, our work offers an effective approach to improving image-text alignment, paving the road for future alignment studies. + +This work was supported in part by NSF CAREER IIS2150012, Adobe Data Science award, Microsoft Accelerate Foundation Models Research Program, and Institute of Information & communications Technology Planning & Evaluation (IITP) grants funded by the Korea government (MSIT) (No. 2022-0-00871, Development of AI Autonomy and Knowledge Enhancement for AI Agent Collaboration) and (No. RS-2022-00187238, Development of Large Korean Language Model Technology for Efficient Pre-training). + +# References + +1. Alayrac, J.B., Donahue, J., Luc, P., Miech, A., Barr, I., Hasson, Y., Lenc, K., Mensch, A., Millican, K., Reynolds, M., et al.: Flamingo: a visual language model for few-shot learning. NeurIPS 35, 23716-23736 (2022) +2. Betker, J., Goh, G., Jing, L., Brooks, T., Wang, J., Li, L., Ouyang, L., Zhuang, J., Lee, J., Guo, Y., Manassra, W., Dhariwal, P., Chu, C., Jiao, Y., Ramesh, A.: Improving image generation with better captions (2023), https://cdn.openai.com/papers/dall-e-3.pdf +3. Bitton, Y.: WYSIWYR repository. https://github.com/yonatanbitton/wysiwyr (2023), accessed: 2024-02-24 +4. Changpinyo, S., Sharma, P., Ding, N., Soricut, R.: Conceptual 12M: Pushing web-scale image-text pre-training to recognize long-tail visual concepts. In: CVPR (2021) +5. Chen, J., Guo, H., Yi, K., Li, B., Elhoseiny, M.: Visualgpt: Data-efficient adaptation of pretrained language models for image captioning (June 2022) +6. Chen, X., Wang, X., Changpinyo, S., Piergiovanni, A.J., Padlewski, P., Salz, D.M., Goodman, S., Grycner, A., Mustafa, B., Beyer, L., Kolesnikov, A., Puigcerver, J., Ding, N., Rong, K., Akbari, H., Mishra, G., Xue, L., Thapliyal, A.V., Bradbury, J., Kuo, W., Seyedhosseini, M., Jia, C., Ayan, B.K., Riquelme, C., Steiner, A., Angelova, A., Zhai, X., Houlsby, N., Soricut, R.: Pali: A jointly-scaled multilingual language-image model. ArXiv abs/2209.06794 (2022), https://apisemantic scholar.org/CorpusID:252222320 +7. Devlin, J., Chang, M.W., Lee, K., Toutanova, K.: Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805 (2018) +8. Doveh, S., Arbelle, A., Harary, S., Herzig, R., Kim, D., Cascante-Bonilla, P., Alfassy, A., Panda, R., Giryes, R., Feris, R., et al.: Dense and aligned captions (dac) promote compositional reasoning in vl models. Advances in Neural Information Processing Systems 36 (2024) +9. Doveh, S., Arbelle, A., Harary, S., Schwartz, E., Herzig, R., Giryes, R., Feris, R., Panda, R., Ullman, S., Karlinsky, L.: Teaching structured vision & language concepts to vision & language models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2657-2668 (2023) +10. Feng, W., He, X., Fu, T.J., Jampani, V., Akula, A.R., Narayana, P., Basu, S., Wang, X.E., Wang, W.Y.: Training-free structured diffusion guidance for compositional text-to-image synthesis. In: The Eleventh International Conference on Learning Representations (2023), https://openreview.net/forum?id=PUIqjT4rzq7 +11. Gadre, S.Y., Ilharco, G., Fang, A., Hayase, J., Smyrnis, G., Nguyen, T., Marten, R., Wortsman, M., Ghosh, D., Zhang, J., et al.: Datacomp: In search of the next + +generation of multimodal datasets. Advances in Neural Information Processing Systems 36 (2024) +12. Herzig, R., Mendelson, A., Karlinsky, L., Arbelle, A., Feris, R., Darrell, T., Globerson, A.: Incorporating structured representations into pretrained vision & language models using scene graphs. EMNLP (2023) +13. Hessel, J., Holtzman, A., Forbes, M., Le Bras, R., Choi, Y.: CLIPScore: A reference-free evaluation metric for image captioning. In: Moens, M.F., Huang, X., Specia, L., Yih, S.W.t. (eds.) Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing. pp. 7514-7528. Association for Computational Linguistics, Online and Punta Cana, Dominican Republic (Nov 2021). https://doi.org/10.18653/v1/2021.emnlp-main.595, https://aclanthology.org/2021.emnlp-main.595 +14. Hsieh, C.Y., Zhang, J., Ma, Z., Kembhavi, A., Krishna, R.: Sugarcrepe: Fixing hackable benchmarks for vision-language compositionality. In: Thirty-Seventh Conference on Neural Information Processing Systems Datasets and Benchmarks Track (2023) +15. Hu, Y., Liu, B., Kasai, J., Wang, Y., Ostendorf, M., Krishna, R., Smith, N.A.: Tifa: Accurate and interpretable text-to-image faithfulness evaluation with question answering. ICCV (2023) +16. Karthik, S., Roth, K., Mancini, M., Akata, Z.: If at first you don't succeed, try, try again: Faithful diffusion-based text-to-image generation by selection. arXiv preprint arXiv:2305.13308 (2023) +17. Le, T., Lal, V., Howard, P.: Coco-counterfactuals: Automatically constructed counterfactual examples for image-text pairs. Advances in Neural Information Processing Systems 36 (2024) +18. Li, J., Li, D., Savarese, S., Hoi, S.C.H.: Blip-2: Bootstrapping language-image pretraining with frozen image encoders and large language models. In: International Conference on Machine Learning (2023), https://apisemantic scholar.org/CorpusID:256390509 +19. Li, J., Li, D., Xiong, C., Hoi, S.: Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In: International Conference on Machine Learning. pp. 12888-12900. PMLR (2022) +20. Li, Y., Du, Y., Zhou, K., Wang, J., Zhao, W.X., Wen, J.R.: Evaluating object hallucination in large vision-language models. EMNLP (2023) +21. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dólar, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13. pp. 740-755. Springer (2014) +22. Lin, Z., Chen, X., Pathak, D., Zhang, P., Ramanan, D.: Revisiting the role of language priors in vision-language models. arXiv preprint arXiv:2306.01879 (2023) +23. Lin, Z., Chen, X., Pathak, D., Zhang, P., Ramanan, D.: Visualgptscore: Visio-linguistic reasoning with multimodal generative pre-training scores. arXiv preprint arXiv:2306.01879 (2023) +24. Liu, H., Li, C., Li, Y., Lee, Y.J.: Improved baselines with visual instruction tuning (2023) +25. Liu, H., Li, C., Wu, Q., Lee, Y.J.: Visual instruction tuning. arXiv:2304.08485 (2023) +26. Ma, Z., Hong, J., Gul, M.O., Gandhi, M., Gao, I., Krishna, R.: Crepe: Can vision-language foundation models reason compositionally? arXiv preprint arXiv:2212.07796 (2022) + +27. OpenAI: Gpt-4v(ision) system card (2023), https://cdn.openai.com/papers/GPTV_System_Card.pdf +28. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021) +29. Ramesh, A., Dhariwal, P., Nichol, A., Chu, C., Chen, M.: Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125 (2022) +30. Ray, A., Radenovic, F., Dubey, A., Plummer, B.A., Krishna, R., Saenko, K.: Cola: How to adapt vision-language models to compose objects localized with attributes? (2023) +31. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10684-10695 (2022) +32. Schuhmann, C., Beaumont, R., Vencu, R., Gordon, C., Wightman, R., Cherti, M., Coombes, T., Katta, A., Mullis, C., Wortsman, M., et al.: Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems 35, 25278-25294 (2022) +33. Sharma, P., Ding, N., Goodman, S., Soricut, R.: Conceptual captions: A cleaned, hypernymed, image alt-text dataset for automatic image captioning. In: Gurevych, I., Miyao, Y. (eds.) Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). pp. 2556-2565. Association for Computational Linguistics, Melbourne, Australia (Jul 2018). https://doi.org/10.18653/v1/P18-1238, https://aclanthology.org/P18-1238 +34. Singh, A., Hu, R., Goswami, V., Couairon, G., Galuba, W., Rohrbach, M., Kiela, D.: Flava: A foundational language and vision alignment model. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15638-15650 (2022) +35. stabilityai: IF repository. https://github.com/deep-floyd/IF +36. Thrush, T., Jiang, R., Bartolo, M., Singh, A., Williams, A., Kiela, D., Ross, C.: Winoground: Probing vision and language models for visio-linguistic compositionality. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5238-5248 (2022) +37. Wang, Z.J., Montoya, E., Munechika, D., Yang, H., Hoover, B., Chau, D.H.: Large-scale prompt gallery dataset for text-to-image generative models. arXiv:2210.14896 [cs] (2022), https://arxiv.org/abs/2210.14896 +38. Xu, J., Liu, X., Wu, Y., Tong, Y., Li, Q., Ding, M., Tang, J., Dong, Y.: Imagereward: Learning and evaluating human preferences for text-to-image generation. Advances in Neural Information Processing Systems 36 (2024) +39. Yarom, M., Bitton, Y., Changpinyo, S., Aharoni, R., Herzig, J., Lang, O., Ofek, E., Szpektor, I.: What you see is what you read? improving text-image alignment evaluation. arXiv preprint arXiv:2305.10400 (2023) +40. Yuksekgonul, M., Bianchi, F., Kalluri, P., Jurafsky, D., Zou, J.: When and why vision-language models behave like bags-of-words, and what to do about it? In: International Conference on Learning Representations (2023), https://openreview.net/forum?id=KRLUvxh8uaX +41. Zhang, K., Mo, L., Chen, W., Sun, H., Su, Y.: Magicbrush: A manually annotated dataset for instruction-guided image editing. In: Advances in Neural Information Processing Systems (2023) + +42. Zhao, T., Zhang, T., Zhu, M., Shen, H., Lee, K., Lu, X., Yin, J.: Vl-checklist: Evaluating pre-trained vision-language models with objects, attributes and relations. arXiv preprint arXiv:2207.00221 (2022) \ No newline at end of file diff --git a/2024/Removing Distributional Discrepancies in Captions Improves Image-Text Alignment/images.zip b/2024/Removing Distributional Discrepancies in Captions Improves Image-Text Alignment/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..b402e3501b7302ecd4311c029e4a5f12602a667f --- /dev/null +++ b/2024/Removing Distributional Discrepancies in Captions Improves Image-Text Alignment/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:98b81c8403d762a7e096dd78db3cd3ad1d88c79dfd92723ea0751398122c6488 +size 509455 diff --git a/2024/Removing Distributional Discrepancies in Captions Improves Image-Text Alignment/layout.json b/2024/Removing Distributional Discrepancies in Captions Improves Image-Text Alignment/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..943809e1ca6bdc90c780f652420b58a3ee05f193 --- /dev/null +++ b/2024/Removing Distributional Discrepancies in Captions Improves Image-Text Alignment/layout.json @@ -0,0 +1,8563 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 160, + 111, + 452, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 111, + 452, + 148 + ], + "spans": [ + { + "bbox": [ + 160, + 111, + 452, + 148 + ], + "type": "text", + "content": "Removing Distributional Discrepancies in Captions Improves Image-Text Alignment" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 142, + 167, + 472, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 142, + 167, + 472, + 194 + ], + "spans": [ + { + "bbox": [ + 142, + 167, + 472, + 194 + ], + "type": "text", + "content": "Yuheng Li" + }, + { + "bbox": [ + 142, + 167, + 472, + 194 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 142, + 167, + 472, + 194 + ], + "type": "text", + "content": ", Haotian Liu" + }, + { + "bbox": [ + 142, + 167, + 472, + 194 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 142, + 167, + 472, + 194 + ], + "type": "text", + "content": ", Mu Cai" + }, + { + "bbox": [ + 142, + 167, + 472, + 194 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 142, + 167, + 472, + 194 + ], + "type": "text", + "content": ", Yijun Li" + }, + { + "bbox": [ + 142, + 167, + 472, + 194 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 142, + 167, + 472, + 194 + ], + "type": "text", + "content": ", Eli Shechtman" + }, + { + "bbox": [ + 142, + 167, + 472, + 194 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 142, + 167, + 472, + 194 + ], + "type": "text", + "content": ", Zhe Lin" + }, + { + "bbox": [ + 142, + 167, + 472, + 194 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 142, + 167, + 472, + 194 + ], + "type": "text", + "content": ", Yong Jae Lee" + }, + { + "bbox": [ + 142, + 167, + 472, + 194 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 142, + 167, + 472, + 194 + ], + "type": "text", + "content": ", and Krishna Kumar Singh" + }, + { + "bbox": [ + 142, + 167, + 472, + 194 + ], + "type": "inline_equation", + "content": "^{2}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 234, + 201, + 378, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 234, + 201, + 378, + 213 + ], + "spans": [ + { + "bbox": [ + 234, + 201, + 378, + 213 + ], + "type": "text", + "content": "1 University of Wisconsin-Madison" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 269, + 213, + 344, + 223 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 269, + 213, + 344, + 223 + ], + "spans": [ + { + "bbox": [ + 269, + 213, + 344, + 223 + ], + "type": "text", + "content": "2 Adobe Research" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 160, + 252, + 455, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 252, + 455, + 418 + ], + "spans": [ + { + "bbox": [ + 160, + 252, + 455, + 418 + ], + "type": "text", + "content": "Abstract. In this paper, we introduce a model designed to improve the prediction of image-text alignment, targeting the challenge of compositional understanding in current visual-language models. Our approach focuses on generating high-quality training datasets for the alignment task by producing mixed-type negative captions derived from positive ones. Critically, we address the distribution imbalance between positive and negative captions to ensure that the alignment model does not depend solely on textual information but also considers the associated images for predicting alignment accurately. By creating this enhanced training data, we fine-tune an existing leading visual-language model to boost its capability in understanding alignment. Our model significantly outperforms current top-performing methods across various datasets. We also demonstrate the applicability of our model by ranking the images generated by text-to-image models based on text alignment. Project page: https://yuheng-li.github.io/LLaVA-score/" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 437, + 230, + 449 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 437, + 230, + 449 + ], + "spans": [ + { + "bbox": [ + 133, + 437, + 230, + 449 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 462, + 482, + 570 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 462, + 482, + 570 + ], + "spans": [ + { + "bbox": [ + 130, + 462, + 482, + 570 + ], + "type": "text", + "content": "Recent years have seen rapid advances in multimodal research, encompassing both visual generation [2, 29, 31] and visual understanding [1, 25, 27]. Training capable multimodal models generally requires extensive datasets of image-text pairs, e.g., LAION [32], CC12M [4], and others [11, 33], which are collected on a web-scale and thus tend to be noisy. This noise in the data contributes to certain challenges. For instance, vision-language models often struggle with hallucination [20] and face difficulties in mastering compositional reasoning [36, 40]. Similarly, text-to-image models frequently fail to generate accurate images when processing complex sentence prompts [10]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": "Given these challenges, the ability to automatically assess whether an image and a caption are semantically aligned plays a crucial role. This capability is essential not only for cleaning the data used to pre-train these models but also for evaluating and enhancing the performance of both text-to-image and image-to-text generation models. The most frequently employed metric for this task is the CLIP score [13], which calculates the cosine distance between the CLIP [28] embeddings of the paired text and image. However, it has been observed that CLIP, along with other methods such as BLIP [19] and Flava [34], tends to" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 140, + 112, + 471, + 219 + ], + "blocks": [ + { + "bbox": [ + 140, + 112, + 471, + 219 + ], + "lines": [ + { + "bbox": [ + 140, + 112, + 471, + 219 + ], + "spans": [ + { + "bbox": [ + 140, + 112, + 471, + 219 + ], + "type": "image", + "image_path": "6acb0a818d1a1ae2c8cdce9e818099d9b54f72c81303aaaf2d1328cfa679878b.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 220, + 480, + 254 + ], + "lines": [ + { + "bbox": [ + 130, + 220, + 480, + 254 + ], + "spans": [ + { + "bbox": [ + 130, + 220, + 480, + 254 + ], + "type": "text", + "content": "Fig.1: Left: Qualitative examples for image-text alignment prediction, where our approach can distinguish fundamental concepts such as positioning, counting, and attributes. Right: Our approach shows superior performance in image-text alignment." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 270, + 480, + 305 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 270, + 480, + 305 + ], + "spans": [ + { + "bbox": [ + 130, + 270, + 480, + 305 + ], + "type": "text", + "content": "operate on a bag-of-words basis. They sometimes even cannot tell the difference between simple cases, such as horse eating grass versus grass eating horse [14, 23, 40]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 307, + 481, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 307, + 481, + 485 + ], + "spans": [ + { + "bbox": [ + 130, + 307, + 481, + 485 + ], + "type": "text", + "content": "The conventional approach to training image-text alignment models involves generating negative captions that are misaligned with the images, which are paired with the original positive captions as the training data. For example, prior works randomly shuffle words to create negative captions [40], or employ language models to generate coherent negative sentences [39]. We follow the latter and leverage an LLM to generate a mix of different types of negative captions during training. Specifically, this includes the replacing type, where one linguistic element is substituted with an arbitrary counterpart (e.g. 'a knife is on the table' " + }, + { + "bbox": [ + 130, + 307, + 481, + 485 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 130, + 307, + 481, + 485 + ], + "type": "text", + "content": " 'a spoon is on the table'), and the swapping type, where words within the same sentence are rearranged (e.g. 'an apple is to the left of a banana' " + }, + { + "bbox": [ + 130, + 307, + 481, + 485 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 130, + 307, + 481, + 485 + ], + "type": "text", + "content": " 'a banana is to the left of an apple'). The former type can aid in enhancing an image-text alignment model's perceptual skills as it needs to distinguish between the original and replaced elements, while the latter type can help the model's reasoning capabilities as elements in the caption remain the same but the relationship between them changes." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 487, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 487, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 487, + 481, + 666 + ], + "type": "text", + "content": "However, critically, we find that this standard approach of generating negative captions, and ensuring that they have coherence (e.g., proper grammar) on a per-instance level, is insufficient. In particular, this approach cannot ensure consistency at the distribution level between positive and negative captions. Distributional biases may originate from the initial dataset or from the rules or models employed to generate negative captions. For instance, the COCO dataset [17] contains significantly more captions including the word giraffe compared to elephant. Yet, when using GPT to generate negative captions, we observe a tendency for GPT to substitute giraffe with elephant, resulting in a surplus of giraffe mentions in positive captions but more elephant in negative captions. Unfortunately, this means that an image-text alignment model trained on such data can be biased to predict sentences containing elephant to be a positive caption and those that contain giraffe to be a negative caption, independent of the paired image. To address this imbalance, our approach consists of fine-tuning a classifier on text captions only (without paired image inputs) to remove biased data that" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 212, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 212, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 212, + 100 + ], + "type": "text", + "content": "Y. Li et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "content": "the text classifier can correctly predict with high confidence. It's important to note that this bias is not unique to the GPT model or how one prompts it, but can arise in other rule-based methods or pre-trained models as well." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 152, + 480, + 235 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 152, + 480, + 235 + ], + "spans": [ + { + "bbox": [ + 130, + 152, + 480, + 235 + ], + "type": "text", + "content": "While previous research [14] noticed distribution differences due to implausible and non-fluent negative captions, our work is the first to eliminate the dataset-level distribution differences using a language model to implicitly encompass all linguistic aspects, without the need for explicit identification. This approach allows us to incorporate considerations like word frequency which is not identified beforehand, thereby providing a more holistic optimization over the dataset's distribution." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 236, + 481, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 236, + 481, + 331 + ], + "spans": [ + { + "bbox": [ + 130, + 236, + 481, + 331 + ], + "type": "text", + "content": "For our image-text alignment model, we leverage a state-of-the-art vision-language model, LLaVA [24, 25], and finetune it with our curated data. Our model achieves significantly better results compared with other image-text alignment models, demonstrated in Figure 1. We also demonstrate that our curated data can improve other models like BLIP2 [18]. Finally, we further show that our image-text alignment model can help with other vision-language tasks like ranking generated images from T2I models [15]. In summary, we have four main contributions:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 342, + 480, + 463 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 138, + 342, + 479, + 366 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 342, + 479, + 366 + ], + "spans": [ + { + "bbox": [ + 138, + 342, + 479, + 366 + ], + "type": "text", + "content": "- We identify new dataset-level distribution differences between positive and negative captions that lead to biased image-text alignment models." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 368, + 480, + 414 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 368, + 480, + 414 + ], + "spans": [ + { + "bbox": [ + 138, + 368, + 480, + 414 + ], + "type": "text", + "content": "- To address this, we propose a method to maintain consistency between positive and negative caption distributions, which is critical to ensure that an image-text alignment model relies on both image and text to measure alignment instead of only the biases present in the text." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 415, + 479, + 438 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 415, + 479, + 438 + ], + "spans": [ + { + "bbox": [ + 138, + 415, + 479, + 438 + ], + "type": "text", + "content": "- We use our curated training data to finetune existing visual-language models like LLaVA, and obtain state-of-the-art results for image-text alignment." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 440, + 479, + 463 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 440, + 479, + 463 + ], + "spans": [ + { + "bbox": [ + 138, + 440, + 479, + 463 + ], + "type": "text", + "content": "- In addition, we demonstrate the application of our image-text alignment model in ranking the image generations produced by generative models." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 132, + 483, + 237, + 495 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 483, + 237, + 495 + ], + "spans": [ + { + "bbox": [ + 132, + 483, + 237, + 495 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 509, + 480, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 509, + 480, + 628 + ], + "spans": [ + { + "bbox": [ + 130, + 509, + 480, + 628 + ], + "type": "text", + "content": "Challenges in Compositional Understanding. Image-text pairs form a crucial interface between visual and linguistic modalities, thus evaluating if a given image-text pair is aligned is important for both data curation and model performance evaluation. The pioneering work of CLIP [28] demonstrated this by leveraging an extensive corpus of such pairs for image-text contrastive training. More recent works such as BLIP [18, 19] and LLaVA [24, 25] further utilize large language models (LLMs) to achieve image-text alignment via the text generation objective. Such models usually inherit the frozen CLIP visual encoder to produce a set of visual tokens, and then feed such tokens and the language instructions into the LLM." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 629, + 480, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 480, + 665 + ], + "type": "text", + "content": "It has been observed that vision-language models like CLIP have limited capability in understanding compositionality [26,30,36,40,42]. Specifically, they find it challenging to recognize the permutation of words within sentences [36]." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 393, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 393, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 393, + 91, + 447, + 100 + ], + "type": "text", + "content": "LLaVA-score" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 186 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 186 + ], + "type": "text", + "content": "Moreover, these models often struggle to identify the binding of attributes to multiple objects within a single sentence or to discern the relationships between objects [30,40,42]. While models built on top of large language models (LLMs), such as BLIP [19] and LLaVA [25], exhibit improved understanding capabilities, they still face difficulties with complex compositional understanding largely due to their lack of training on sufficiently challenging data." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 188, + 482, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 188, + 482, + 392 + ], + "spans": [ + { + "bbox": [ + 130, + 188, + 482, + 392 + ], + "type": "text", + "content": "Enhancing Vision-Language Compositionality. Several studies focus on improving models' understanding of language compositionality [8,9,12,17,23,39, 40]. Some strategies [15,39] involve using language models to decompose text into multiple succinct assertion phrases, which are then evaluated by VQA models on an individual basis. [23] also finds that assessing the conditional probability of predicting text based on the input image offers more accurate outcomes than traditional discriminative approaches like contrastive or classification scores used in BLIP [19]. Still, the most widely used method remains the explicit fine-tuning of models to differentiate between hard negatives and correct captions [9, 39, 40]. [40] randomly shuffles words to generate negative captions, but subsequent analysis by [14] points out the approach's flaw: models could simply rely on textual cues (e.g., grammar correctness) for predictions. [39] attempts to employ LLMs to create negative captions, ensuring the negative captions are both fluent and meaningful. Nonetheless, we observed that current fine-tuning methods do not generate a diverse range of negative prompts. Moreover, merely evaluating the grammar or logical coherence of negative captions is insufficient to eliminate data bias in the distribution of negative captions." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 414, + 212, + 427 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 414, + 212, + 427 + ], + "spans": [ + { + "bbox": [ + 132, + 414, + 212, + 427 + ], + "type": "text", + "content": "3 Approach" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 441, + 480, + 538 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 441, + 480, + 538 + ], + "spans": [ + { + "bbox": [ + 130, + 441, + 480, + 538 + ], + "type": "text", + "content": "In our approach, we assume access to images accompanied by accurately labeled positive captions, similar to those found in the MSCOCO [21] dataset. Following [14], contrary to adopting rule-based techniques that often generate illogical sentences, our method utilizes large language models (LLMs) like GPT4 to transform positive captions into negative ones. We first outline our strategy for generating various types of negative captions and then present a straightforward technique to mitigate biases inherent in the distributions of positive and negative captions. We show the entire pipeline in Figure 2." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 559, + 364, + 571 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 559, + 364, + 571 + ], + "spans": [ + { + "bbox": [ + 132, + 559, + 364, + 571 + ], + "type": "text", + "content": "3.1 Constructing Diverse Negative Captions" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 582, + 479, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 479, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 479, + 665 + ], + "type": "text", + "content": "Suppose we have a dataset consisting of image-text pairs " + }, + { + "bbox": [ + 130, + 582, + 479, + 665 + ], + "type": "inline_equation", + "content": "\\{I, T_p\\}" + }, + { + "bbox": [ + 130, + 582, + 479, + 665 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 582, + 479, + 665 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 130, + 582, + 479, + 665 + ], + "type": "text", + "content": " represents an image and " + }, + { + "bbox": [ + 130, + 582, + 479, + 665 + ], + "type": "inline_equation", + "content": "T_p" + }, + { + "bbox": [ + 130, + 582, + 479, + 665 + ], + "type": "text", + "content": " is its associated positive caption. Previous research [39,40] showed that merely randomly shuffling image-text pairs to generate negative samples is insufficient to learn capable vision-language models that properly understand the structure of a caption and its relationship with the image. These studies have underscored the value of constructing hard negative captions to significantly improve the model's language compositional abilities. Following this" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 211, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 211, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 211, + 100 + ], + "type": "text", + "content": "Y. Li et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 152, + 115, + 462, + 219 + ], + "blocks": [ + { + "bbox": [ + 152, + 115, + 462, + 219 + ], + "lines": [ + { + "bbox": [ + 152, + 115, + 462, + 219 + ], + "spans": [ + { + "bbox": [ + 152, + 115, + 462, + 219 + ], + "type": "image", + "image_path": "5c5a4dbefb387151ee8e67fcb178f77af6bb8724c00034bc67ff276e97a68e60.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 228, + 482, + 284 + ], + "lines": [ + { + "bbox": [ + 130, + 228, + 482, + 284 + ], + "spans": [ + { + "bbox": [ + 130, + 228, + 482, + 284 + ], + "type": "text", + "content": "Fig. 2: Top: We feed the positive caption (green dots) into GPT to create two types of negative captions (red dots): substituting one linguistic element with any plausible alternative or swapping the positions of two components. The blue part in negative captions highlights the modifications. Bottom: we remove easy negative samples using only text data and utilize the remaining samples to fine-tune vision-language models." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 306, + 480, + 330 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 306, + 480, + 330 + ], + "spans": [ + { + "bbox": [ + 130, + 306, + 480, + 330 + ], + "type": "text", + "content": "insight [39], we utilize large language models (LLMs) to generate such hard negative captions. Specifically, we create two types of hard negative captions." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 331, + 482, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 331, + 482, + 449 + ], + "spans": [ + { + "bbox": [ + 130, + 331, + 482, + 449 + ], + "type": "text", + "content": "We refer to the first method for creating negative captions as the replacing strategy, which identifies key components in a language and uses a language model to replace it with other plausible substitutes. The replaced component can be any linguistic part such as a noun, adjective, preposition, etc. For example, \"a photo of a broken down stop sign\" could be replaced with \"a photo of a brand new stop sign\"; \"a cute cat looking at a bird\" could be changed to \"a cute dog looking at a bird\". Conceptually, this type of negative caption aims to enhance the model's recognition capabilities. For executing this task, we engage GPT by providing it with specific instructions followed by some contextual examples, as illustrated in Figure 3." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 450, + 482, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 450, + 482, + 594 + ], + "spans": [ + { + "bbox": [ + 130, + 450, + 482, + 594 + ], + "type": "text", + "content": "We refer to the second method for creating negative captions as the swapping strategy. This approach involves generating a new sentence by utilizing the original language components, typically resulting in the swapping of two or several identical linguistic elements. Specifically, we first ask GPT to break down the original positive sentence into its key components, and then request the model to construct a new sentence with these elements. For example, if the input caption is \"an airplane is flying in the blue sky\", GPT is tasked with identifying the key components including \"airplane\", \"flying\", \"blue\", and \"sky\". The newly crafted sentence could be \"a blue airplane is flying in the sky\". Please note that for some short captions, there may not be enough language elements to form a reasonably different sentence in meaning. Thus, we also need to judge if the new sentence makes sense or not. Please refer to the Figure 3 for our prompt used in GPT." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 610, + 354, + 623 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 610, + 354, + 623 + ], + "spans": [ + { + "bbox": [ + 131, + 610, + 354, + 623 + ], + "type": "text", + "content": "3.2 Addressing Distribution Discrepancies" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "type": "text", + "content": "Previous studies [14] have demonstrated that artifact issues arising from incorrect construction in negative captions can lead to distribution disparities between positive and negative captions. To mitigate this, commonsense and grammar" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 393, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 393, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 393, + 91, + 447, + 100 + ], + "type": "text", + "content": "LLaVA-score" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 149, + 114, + 462, + 218 + ], + "blocks": [ + { + "bbox": [ + 149, + 114, + 462, + 218 + ], + "lines": [ + { + "bbox": [ + 149, + 114, + 462, + 218 + ], + "spans": [ + { + "bbox": [ + 149, + 114, + 462, + 218 + ], + "type": "image", + "image_path": "979ada63a8130fe6e8e785fb495608efd50a0c44b36ea47277e1b5dfd9216ec9.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 222, + 480, + 244 + ], + "lines": [ + { + "bbox": [ + 132, + 222, + 480, + 244 + ], + "spans": [ + { + "bbox": [ + 132, + 222, + 480, + 244 + ], + "type": "text", + "content": "Fig. 3: Prompts used in GPT for generating two types of negative captions, with in-context examples shown in purple." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 149, + 251, + 462, + 381 + ], + "blocks": [ + { + "bbox": [ + 149, + 251, + 462, + 381 + ], + "lines": [ + { + "bbox": [ + 149, + 251, + 462, + 381 + ], + "spans": [ + { + "bbox": [ + 149, + 251, + 462, + 381 + ], + "type": "image", + "image_path": "2a62d04cf16e045d8b6f645658bc1adedf3e2484b619fd58de450e9823cada58.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 384, + 482, + 418 + ], + "lines": [ + { + "bbox": [ + 132, + 384, + 482, + 418 + ], + "spans": [ + { + "bbox": [ + 132, + 384, + 482, + 418 + ], + "type": "text", + "content": "Fig. 4: Top prediction based on a text-only binary classifier. Top left: Negative captions generated through replacement strategy. Top right: Negative captions generated through swapping strategy. Bottom: Positive captions from the COCO dataset." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 446, + 482, + 482 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 446, + 482, + 482 + ], + "spans": [ + { + "bbox": [ + 132, + 446, + 482, + 482 + ], + "type": "text", + "content": "models are employed. Nonetheless, our findings indicate that merely assessing the sensibility of sentences is insufficient for aligning the distributions of positive and negative sentences because of two following two reasons." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 483, + 482, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 483, + 482, + 592 + ], + "spans": [ + { + "bbox": [ + 130, + 483, + 482, + 592 + ], + "type": "text", + "content": "First, each image-pair dataset inherently establishes its own unique distribution. For instance, the COCO dataset [21], despite its widespread use and diverse content, predominantly encompasses everyday scenes, objects, and activities. Common subjects such as people and motorcycles, along with frequent activities like skiing and surfing, characterize the distinct distribution of COCO. Second, when employing rule-based methods or prompting language models to generate negative captions, any strategy adopted will inevitably reflect biases from human preconceptions or pre-trained models. This leads to negative captions originating from a different distribution." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": "To illustrate this concept, we develop a blind text-only binary classifier that processes positive and negative captions without seeing any images. Surprisingly, we find that despite the coherence and logical structure of negative captions, the classifier is adept at distinguishing between negative and positive captions based solely on the text. Figure 4 shows the most confidently correct predictions for both our replace and swap data categories constructed from COCO." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 212, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 212, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 212, + 100 + ], + "type": "text", + "content": "Y. Li et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 115, + 482, + 283 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 115, + 482, + 283 + ], + "spans": [ + { + "bbox": [ + 130, + 115, + 482, + 283 + ], + "type": "text", + "content": "In the dataset with replaced negative data, it is observed that GPT frequently substitutes the terms \"boat\" and \"elephant\" for the original COCO captions' \"airplane\" and \"giraffe\", respectively. This pattern makes it straightforward for a text-only model to identify these captions as negative. In our swap data, the text-only classifier is able to identify captions with colorful animals interacting with a black object as negative. Note that these negative prompts are logically coherent and grammatically correct, thus cannot be easily detected by a grammar model. Conversely, the top accurately predicted positive captions often depict common activities like surfing and motorcycle riding, which align closely with the typical content style of COCO. It's important to recognize that data bias is not solely a characteristic of our method for generating data nor the manner in which we prompt GPT. In the experimental section, we demonstrate that such distribution discrepancy is also present in other approaches [14] that employ GPT for creating negative prompts." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 283, + 482, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 283, + 482, + 354 + ], + "spans": [ + { + "bbox": [ + 130, + 283, + 482, + 354 + ], + "type": "text", + "content": "The presence of bias in the data can obstruct a vision-language model's ability to truly understand image structures and learn language compositions, as the model might rely solely on textual cues for making predictions. To address this issue, we propose a straightforward solution aimed at reducing data distribution differences by selectively removing data that is predicted with high confidence by a text-alone model." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 356, + 482, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 356, + 482, + 521 + ], + "spans": [ + { + "bbox": [ + 130, + 356, + 482, + 521 + ], + "type": "text", + "content": "Figure 2 (bottom) depicts our conceptual goal of data filtering: to eliminate straightforward or biased positive and negative samples. This is done to ensure that the remaining text data cannot be distinguished by any text classifier, based solely on the text information. Essentially, our aim is to maximize the entropy of the text information within our dataset. In practice, we organize our dataset into " + }, + { + "bbox": [ + 130, + 356, + 482, + 521 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 356, + 482, + 521 + ], + "type": "text", + "content": " partitions. For each iteration, one partition is designated as the test set while the remaining " + }, + { + "bbox": [ + 130, + 356, + 482, + 521 + ], + "type": "inline_equation", + "content": "N - 1" + }, + { + "bbox": [ + 130, + 356, + 482, + 521 + ], + "type": "text", + "content": " partitions serve as the training set. We employ a pretrained Bert model [7] as our text-only classifier, and train it on the training set. Subsequently, this trained classifier is applied to the designated test set. We then rank the correctly predicted samples by the classifier's confidence level, removing the top " + }, + { + "bbox": [ + 130, + 356, + 482, + 521 + ], + "type": "inline_equation", + "content": "k\\%" + }, + { + "bbox": [ + 130, + 356, + 482, + 521 + ], + "type": "text", + "content": " of these samples for both positive and negative class predictions. The rest of the data is retained as our refined dataset. This procedure is repeated for each partition, ensuring a comprehensive reduction of bias across the dataset." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 523, + 482, + 571 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 523, + 482, + 571 + ], + "spans": [ + { + "bbox": [ + 130, + 523, + 482, + 571 + ], + "type": "text", + "content": "Note that while our observation and data filtering approach is applied to address the image-text alignment issue, it could be a more general problem for any multimodal data scenario where bias in one modality might negatively impact model training." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 586, + 427, + 600 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 586, + 427, + 600 + ], + "spans": [ + { + "bbox": [ + 131, + 586, + 427, + 600 + ], + "type": "text", + "content": "3.3 Finetuning VLMs for Image-Text Alignment Scoring" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": "Once we have the unbiased data, we opt to fine-tune a vision-language model (VLM) to enhance its language compositional understanding capabilities with respect to images. In our study, we mainly select LLaVA-1.5 [24] due to its superior performance in image and text understanding. Since LLaVA-1.5 is designed to generate text, to adapt it for use as a image-text score calculator, we employ" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 393, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 393, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 393, + 91, + 447, + 100 + ], + "type": "text", + "content": "LLaVA-score" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 99 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 99 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 99 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 164 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 164 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 164 + ], + "type": "text", + "content": "the following prompt formatting: \"Does this image I match the following caption " + }, + { + "bbox": [ + 130, + 116, + 482, + 164 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 130, + 116, + 482, + 164 + ], + "type": "text", + "content": ". Answer Yes or No directly.\" Given that LLaVA relies on a language model to produce subsequent words, we manually extract the logits associated with the responses Yes and No for the next word. We then define the matching score as:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 242, + 173, + 481, + 198 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 242, + 173, + 481, + 198 + ], + "spans": [ + { + "bbox": [ + 242, + 173, + 481, + 198 + ], + "type": "interline_equation", + "content": "\\frac {e ^ {\\mathbf {P} (\\text {Y e s} | \\text {p r o m p t})}}{e ^ {\\mathbf {P} (\\text {Y e s} | \\text {p r o m p t})} + e ^ {\\mathbf {P} (\\text {N o} | \\text {p r o m p t})}} \\tag {1}", + "image_path": "67abc2b3046e35fb61698c2547eb7f371aa8f51a4a1ad58e91f50e4d5b7c0fea.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 201, + 482, + 274 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 201, + 482, + 274 + ], + "spans": [ + { + "bbox": [ + 130, + 201, + 482, + 274 + ], + "type": "text", + "content": "We discover that this straightforward approach is quite effective and can already outperform many existing state-of-the-art baselines. However, LLaVA-1.5 was not specifically trained for this type of matching problem. Therefore, to enhance its performance, we finetune it with the same prompt formatting introduced above using our curated data. We assign the labels Yes and No to positive and negative pairs, respectively." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 274, + 482, + 334 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 274, + 482, + 334 + ], + "spans": [ + { + "bbox": [ + 130, + 274, + 482, + 334 + ], + "type": "text", + "content": "Our curated dataset is not restricted to LLaVA-1.5. In our experiments, we also finetune a Q-former equipped with the Image-Text Matching (ITM) head in BLIP2 [18]. The ITM head is essentially a binary classification layer, identical to our dataset's structure. We thus fine-tune the model using the standard cross-entropy loss." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 352, + 200, + 364 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 352, + 200, + 364 + ], + "spans": [ + { + "bbox": [ + 132, + 352, + 200, + 364 + ], + "type": "text", + "content": "4 Results" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 376, + 482, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 376, + 482, + 437 + ], + "spans": [ + { + "bbox": [ + 130, + 376, + 482, + 437 + ], + "type": "text", + "content": "We evaluate our fine-tuned LLaVA-1.5 model against various baselines across different datasets. Additionally, we conduct ablation studies to evaluate the impact and importance of our dual-strategy approach for generating diverse negative captions and our method for addressing data distribution discrepancies. Since our main model is built upon LLaVA-1.5, we name our score as LLaVA-score." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 454, + 208, + 465 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 454, + 208, + 465 + ], + "spans": [ + { + "bbox": [ + 132, + 454, + 208, + 465 + ], + "type": "text", + "content": "4.1Baselines" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 474, + 440, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 474, + 440, + 485 + ], + "spans": [ + { + "bbox": [ + 130, + 474, + 440, + 485 + ], + "type": "text", + "content": "We evaluate our model against a range of leading multimodal models:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 486, + 431, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 486, + 431, + 498 + ], + "spans": [ + { + "bbox": [ + 130, + 486, + 431, + 498 + ], + "type": "text", + "content": "CLIP-ViT-L/14, the largest variant of OpenAI CLIP models [28]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 498, + 481, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 498, + 481, + 521 + ], + "spans": [ + { + "bbox": [ + 130, + 498, + 481, + 521 + ], + "type": "text", + "content": "BLIP-2 [18], which incorporates both image-text matching and image-text contrastive learning approaches." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 522, + 481, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 522, + 481, + 545 + ], + "spans": [ + { + "bbox": [ + 130, + 522, + 481, + 545 + ], + "type": "text", + "content": "NegCLIP, as introduced in [40], fine-tuned on challenging negative captions generated by randomly shuffling words within sentences." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 546, + 482, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 546, + 482, + 605 + ], + "spans": [ + { + "bbox": [ + 130, + 546, + 482, + 605 + ], + "type": "text", + "content": "VisualGPTScore [22] proposes utilizing the probability of generating specific text given an image (i.e., " + }, + { + "bbox": [ + 130, + 546, + 482, + 605 + ], + "type": "inline_equation", + "content": "P(T|I)" + }, + { + "bbox": [ + 130, + 546, + 482, + 605 + ], + "type": "text", + "content": ") as an effective metric for calculating image-text alignment scores. For this, we utilize LLaVA-1.5 [24], the state-of-the-art vision-language model, to calculate their proposed VisualGPTScore, employing a reweighting technique as suggested by [22]." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 606, + 481, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 606, + 481, + 629 + ], + "spans": [ + { + "bbox": [ + 130, + 606, + 481, + 629 + ], + "type": "text", + "content": "Image-Reward [38]: A reward model, trained on human preferences of image-text pairs produced by Text-to-Image (T2I) models in DiffusionDB [37]." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 630, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 630, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 630, + 482, + 666 + ], + "type": "text", + "content": "VQ2 [39] first extracts a set of candidate QA pairs from the text, then uses a VQA model to score each pair. For this baseline, we report results using both the PaLI model [6]—directly citing their paper as the model is not publicly" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 212, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 212, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 212, + 101 + ], + "type": "text", + "content": "Y. Li et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "text", + "content": "available—and BLIP-T5 [18] which is accessible via their official GitHub repository [3]. Note that while TIFA [15] similarly utilizes a QA pairs approach, we exclude it from our results due to VQ2 demonstrating superior performance, aiming for simplicity in our comparison." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 163, + 479, + 199 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 163, + 479, + 199 + ], + "spans": [ + { + "bbox": [ + 130, + 163, + 479, + 199 + ], + "type": "text", + "content": "PaLI fine-tuned on the SeeTrue dataset [39]: A version of PaLI specifically fine-tuned for the alignment task, using a curated dataset. Performance metrics are cited directly from the original paper [39]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 200, + 480, + 246 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 200, + 480, + 246 + ], + "spans": [ + { + "bbox": [ + 130, + 200, + 480, + 246 + ], + "type": "text", + "content": "LLaVA-1.5 [24], Originally a text generation model, we employ a specific prompt as introduced in Sec. 3.3. By using Equation 1, we convert it into a scoring function. Through empirical testing, this prompt is found to be the most effective, and it is utilized to finetune our model." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 264, + 268, + 275 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 264, + 268, + 275 + ], + "spans": [ + { + "bbox": [ + 131, + 264, + 268, + 275 + ], + "type": "text", + "content": "4.2 Datasets and Metrics" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 283, + 356, + 293 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 283, + 356, + 293 + ], + "spans": [ + { + "bbox": [ + 130, + 283, + 356, + 293 + ], + "type": "text", + "content": "We evaluate on the following datasets and metrics:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 295, + 479, + 341 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 295, + 479, + 341 + ], + "spans": [ + { + "bbox": [ + 130, + 295, + 479, + 341 + ], + "type": "text", + "content": "Winoground [36]: This dataset uniquely comprises quartets, each including two images and two texts, necessitating a nuanced interpretation of both linguistic and visual elements for accurate matching. The metrics reported for this dataset encompass an image score, a text score, and a group score." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 342, + 479, + 388 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 342, + 479, + 388 + ], + "spans": [ + { + "bbox": [ + 130, + 342, + 479, + 388 + ], + "type": "text", + "content": "SeeTRUe [39]: A benchmark designed for assessing vision-language models, featuring a test set combining multiple sources. Current sources include Drawbench, EditBench, and COCO-t2i, pairing real texts with synthetic images. Following [39], we utilize ROC-AUC as the evaluation metric." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 391, + 488, + 424 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 391, + 488, + 424 + ], + "spans": [ + { + "bbox": [ + 130, + 391, + 488, + 424 + ], + "type": "text", + "content": "SugarCrepe [14]: A recently introduced benchmark focuses on generating creative negative captions using a language model and employs grammar and commonsense models for data cleaning, marking a novel approach in benchmark design." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 426, + 479, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 426, + 479, + 484 + ], + "spans": [ + { + "bbox": [ + 130, + 426, + 479, + 484 + ], + "type": "text", + "content": "MagicBrush [41]: This benchmark facilitates human-driven image editing, constructed using Dall-E 2, and includes captions for both the original and edited images. Featuring quartets similar to Winoground [36], we adopt the same methodology for calculating the group score with modification due to nature of data. Refer supp for details." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 131, + 502, + 279, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 502, + 279, + 514 + ], + "spans": [ + { + "bbox": [ + 131, + 502, + 279, + 514 + ], + "type": "text", + "content": "4.3 Implementation Details" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 521, + 479, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 479, + 616 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 479, + 616 + ], + "type": "text", + "content": "For our dataset curation, we select COCO, which provides positive image-text pairs. To diversify our image dataset and enhance our model's robustness to synthetic images, we incorporate a subset of the training images from SeeTRUE [39], specifically the coco_train_t2i set. To construct negative data, we use GPT to generate two types of negative captions (swap and replace), and we perform random sampling to maintain an equal amount of positive and negative data. In the process of curating our dataset, we ensure that neither images nor captions included in the training dataset are present within the test dataset." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 617, + 479, + 663 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 479, + 663 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 479, + 663 + ], + "type": "text", + "content": "For our model development, we choose to refine LLaVA-1.5 [24] using our curated data. We train the model with a batch size of 64 on 8 NVIDIA A100 GPUs for a single epoch, setting the learning rate at 2e-6. For our data filtering, we use " + }, + { + "bbox": [ + 130, + 617, + 479, + 663 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 617, + 479, + 663 + ], + "type": "text", + "content": " to be " + }, + { + "bbox": [ + 130, + 617, + 479, + 663 + ], + "type": "inline_equation", + "content": "30\\%" + }, + { + "bbox": [ + 130, + 617, + 479, + 663 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 617, + 479, + 663 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 617, + 479, + 663 + ], + "type": "text", + "content": " to be 5. Find more details in the supp." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 393, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 393, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 393, + 91, + 447, + 100 + ], + "type": "text", + "content": "LLaVA-score" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 133, + 175, + 480, + 304 + ], + "blocks": [ + { + "bbox": [ + 130, + 114, + 482, + 171 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 482, + 171 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 482, + 171 + ], + "type": "text", + "content": "Table 1: We evaluate our model alongside baselines across multiple datasets. The best results are shown in bold. For three subsets of the SeeTRUE dataset, we present the ROC AUC scores following the original paper [39]. For the SugarCrepe dataset, accuracy is employed as the performance metric. For the MagicBrush [41] dataset, we report the group score. We use LLaVA-1.5 to calculate the VisualGPT score [23]." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 133, + 175, + 480, + 304 + ], + "lines": [ + { + "bbox": [ + 133, + 175, + 480, + 304 + ], + "spans": [ + { + "bbox": [ + 133, + 175, + 480, + 304 + ], + "type": "table", + "html": "
Winoground image text groupSeeTRUE DrawBench EditBench COCO-T2ISugarCrepe replace swap addMagicBrush
Chance Performance25.0025.0016.6750.050.050.050.050.050.033.33
CLIP-ViT-L-14 [28]10.5028.507.7561.462.159.279.461.474.852.89
NegCLIP [40]11.7530.758.2563.266.062.885.375.387.261.12
BLIP2-ITM [18]24.2541.7519.0060.867.568.088.983.991.875.32
BLIP2-ITC [18]12.0028.508.5064.967.969.986.766.992.367.85
Image-Reward [38]15.2543.0012.7570.470.277.088.281.095.270.28
VisualGPT [23]37.0044.2527.5077.074.269.188.287.195.578.31
VQ2 (PaLI) [39]42.2547.0030.5082.673.683.4----
VQ2 (BLIP-T5) [39]34.0033.5023.2574.867.474.283.381.090.970.46
PaLI (ft on SeeTRUE) [39]38.0046.5028.7586.877.283.2----
LLaVA-1.5 [24]49.7551.0034.2586.978.384.593.588.395.882.61
LLaVA-score (Ours)68.0053.7547.2588.877.784.995.394.997.587.28
", + "image_path": "c8f8c582ee54b4bc1e659aa8ba51f1429d361684bccb249238f6f7a625c44398.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 319, + 224, + 331 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 319, + 224, + 331 + ], + "spans": [ + { + "bbox": [ + 132, + 319, + 224, + 331 + ], + "type": "text", + "content": "4.4 Main results" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 340, + 482, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 340, + 482, + 495 + ], + "spans": [ + { + "bbox": [ + 130, + 340, + 482, + 495 + ], + "type": "text", + "content": "Table 1 presents a comparison between our models and various strong baselines across different datasets. It is evident that our model outperforms others in nearly all datasets. Interestingly, utilizing Eq. 1, LLaVA-1.5 zero-shot performance is the second-best method overall. Particularly for Winoground, a benchmark well-known for its challenges in visual and linguistic reasoning, some models such as CLIP, BLIP2, and Image-Reward perform at or below chance level. Our fine-tuned LLaVA-1.5 substantially enhancing its reasoning capabilities showing the importance of our strategy for training data curation. For VQ2, the VQA models (BLIP or PaLI) underperform our model due to their lack of training on challenging negative examples. The finetuned PaLI model also falls short, attributed to its training dataset's lack of diversity and absence of data filtering. Our model also shows impressive results on synthetic image benchmarks like SeeTRUE and MagicBrush." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 514, + 443, + 526 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 514, + 443, + 526 + ], + "spans": [ + { + "bbox": [ + 132, + 514, + 443, + 526 + ], + "type": "text", + "content": "4.5 Performance on Attribute, Counting, Spatial Reasoning" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 533, + 482, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 482, + 616 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 482, + 616 + ], + "type": "text", + "content": "In the preceding section, we demonstrated that our model surpasses baseline models across various datasets. This section offers an alternative perspective on our model's performance, particularly in areas where both vision-language models and recent text-to-image (T2I) models encounter significant challenges. These challenges include multiple attributes binding, counting objects, and understanding spatial relationships between objects. Addressing these issues is crucial for both visual understanding and evaluating the capabilities of generative models." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "To quantify the improvements our model achieves in addressing these common challenges, we construct three specialized datasets. Specifically, we prompt GPT to generate scenarios involving attribute binding, object counting, and spatial relationships, providing in-context examples to guide the generation process." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 212, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 212, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 212, + 101 + ], + "type": "text", + "content": "Y. Li et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 170, + 146, + 443, + 306 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 479, + 137 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 479, + 137 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 479, + 137 + ], + "type": "text", + "content": "Table 2: Comparison of our model against baselines across three prevalent challenges, highlighting the highest average accuracy in bold." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 170, + 146, + 443, + 306 + ], + "lines": [ + { + "bbox": [ + 170, + 146, + 443, + 306 + ], + "spans": [ + { + "bbox": [ + 170, + 146, + 443, + 306 + ], + "type": "table", + "html": "
AttributeCountingSpatial
avgposnegavgposnegavgposneg
Chance Performance505050505050505050
Threshold-Independent Models (with oracle)
CLIP-ViT-L-14 [28]635274586848531888
NegCLIP [40]657852596652574866
BLIP2-ITC [18]667260573678579024
VisualGPT [5]739056655278625668
Inherent Decision Models
BLIP2-ITM [18]5810016539610511002
Image-Reward [38]70100406110022579816
VQ2 (BLIP-T5) [39]669438568230549414
LLaVA-1.5 [24]719844629628579816
LLaVA-score (Ours)819066718656817686
", + "image_path": "c30d4847ba59f672eab4ab0bfa8eca7839c7beafbbd223880ca06e715e3ca888.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 331, + 480, + 429 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 331, + 480, + 429 + ], + "spans": [ + { + "bbox": [ + 130, + 331, + 480, + 429 + ], + "type": "text", + "content": "Subsequently, the T2I model [35] is employed to generate 50 images for each prompt. From these, we manually select one positive and one negative image based on image-text alignment. We opt for synthetic images as it offers greater flexibility, as illustrated in Figure 5, it can incorporate diverse styles (e.g., painting styles for birds), create unconventional images (e.g., the examples involving soccer and cats), and has attribute/object merging capabilities of T2I models to enrich our dataset with varied negative cases (e.g., the negative example of mistakenly combining clock and apple)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 433, + 480, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 433, + 480, + 529 + ], + "spans": [ + { + "bbox": [ + 130, + 433, + 480, + 529 + ], + "type": "text", + "content": "Table 2 presents our results. We report classification accuracy as the metric, alongside showing both positive and negative results in gray color. We categorize the baseline models into two distinct groups for comparison purposes. 1) Upper section of the table, consists of models that generate scores without an established classification threshold. Consequently, we calculate the accuracy by using an oracle threshold, which involves optimizing over all predicted scores to find the best cutoff (cheating on the test set). 2) Lower part of the table, operates with an inherent decision boundary for classification." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 533, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 481, + 666 + ], + "type": "text", + "content": "Our model demonstrates superior performance across all three challenging cases, beating competitors that perform only marginally better than random guessing, particularly on tasks requiring spatial understanding. It's noteworthy that many models exhibit a bias, often predicting the same outcome across the test dataset. This tendency is underscored by their high accuracy on positive samples contrasted with poor performance on negative ones. Given that our images are synthesized using text-to-image (T2I) models, essential linguistic elements are likely to be present even in images labeled as negative (refer to Figure 5). This observation suggests that baseline models operate similarly to a bag-of-visual-words approach, predicting an image as matching the textual description as long as it contains certain key visual concepts." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 393, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 393, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 393, + 91, + 447, + 100 + ], + "type": "text", + "content": "LLaVA-score" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 157, + 125, + 200, + 167 + ], + "blocks": [ + { + "bbox": [ + 157, + 125, + 200, + 167 + ], + "lines": [ + { + "bbox": [ + 157, + 125, + 200, + 167 + ], + "spans": [ + { + "bbox": [ + 157, + 125, + 200, + 167 + ], + "type": "image", + "image_path": "4bb81f94169eb4a0073f657e01af448d9babb433f03df3bd08867e653360ea73.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 168, + 167, + 235, + 178 + ], + "lines": [ + { + "bbox": [ + 168, + 167, + 235, + 178 + ], + "spans": [ + { + "bbox": [ + 168, + 167, + 235, + 178 + ], + "type": "text", + "content": "An antique wooden clock beside a glossy red apple" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 201, + 126, + 244, + 167 + ], + "blocks": [ + { + "bbox": [ + 188, + 118, + 211, + 124 + ], + "lines": [ + { + "bbox": [ + 188, + 118, + 211, + 124 + ], + "spans": [ + { + "bbox": [ + 188, + 118, + 211, + 124 + ], + "type": "text", + "content": "Attribute" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 201, + 126, + 244, + 167 + ], + "lines": [ + { + "bbox": [ + 201, + 126, + 244, + 167 + ], + "spans": [ + { + "bbox": [ + 201, + 126, + 244, + 167 + ], + "type": "image", + "image_path": "defbc6ffd1d28e4ef74d3908a2b5d9774449f89ea0b728b872406e055b848eed.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 263, + 126, + 304, + 167 + ], + "blocks": [ + { + "bbox": [ + 294, + 118, + 317, + 125 + ], + "lines": [ + { + "bbox": [ + 294, + 118, + 317, + 125 + ], + "spans": [ + { + "bbox": [ + 294, + 118, + 317, + 125 + ], + "type": "text", + "content": "Counting" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 263, + 126, + 304, + 167 + ], + "lines": [ + { + "bbox": [ + 263, + 126, + 304, + 167 + ], + "spans": [ + { + "bbox": [ + 263, + 126, + 304, + 167 + ], + "type": "image", + "image_path": "5783961fefc1047784d1ceee9ad8b089c093c3cbb3df20c4357d868f13d02fea.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 276, + 167, + 334, + 173 + ], + "lines": [ + { + "bbox": [ + 276, + 167, + 334, + 173 + ], + "spans": [ + { + "bbox": [ + 276, + 167, + 334, + 173 + ], + "type": "text", + "content": "A fishbowl with two goldfish" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 307, + 126, + 348, + 167 + ], + "blocks": [ + { + "bbox": [ + 307, + 126, + 348, + 167 + ], + "lines": [ + { + "bbox": [ + 307, + 126, + 348, + 167 + ], + "spans": [ + { + "bbox": [ + 307, + 126, + 348, + 167 + ], + "type": "image", + "image_path": "3df66c9fae21c5d19b8333f1c2de6a46595c70639c389d09d66a86938f68d315.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 366, + 126, + 407, + 167 + ], + "blocks": [ + { + "bbox": [ + 399, + 118, + 416, + 125 + ], + "lines": [ + { + "bbox": [ + 399, + 118, + 416, + 125 + ], + "spans": [ + { + "bbox": [ + 399, + 118, + 416, + 125 + ], + "type": "text", + "content": "Spatial" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 366, + 126, + 407, + 167 + ], + "lines": [ + { + "bbox": [ + 366, + 126, + 407, + 167 + ], + "spans": [ + { + "bbox": [ + 366, + 126, + 407, + 167 + ], + "type": "image", + "image_path": "140f450c36c75dd216cfa816f271dc0116948543f0bf8ef8cb003a7ebf1d15c1.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 370, + 168, + 448, + 174 + ], + "lines": [ + { + "bbox": [ + 370, + 168, + 448, + 174 + ], + "spans": [ + { + "bbox": [ + 370, + 168, + 448, + 174 + ], + "type": "text", + "content": "A soccer ball is embedded within a TV" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 410, + 126, + 451, + 167 + ], + "blocks": [ + { + "bbox": [ + 410, + 126, + 451, + 167 + ], + "lines": [ + { + "bbox": [ + 410, + 126, + 451, + 167 + ], + "spans": [ + { + "bbox": [ + 410, + 126, + 451, + 167 + ], + "type": "image", + "image_path": "bd1a812394a82c2cb35f0fe3ed53f82414f82ed129b6a0ebcd3284f77fe1e60e.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 158, + 181, + 199, + 221 + ], + "blocks": [ + { + "bbox": [ + 158, + 181, + 199, + 221 + ], + "lines": [ + { + "bbox": [ + 158, + 181, + 199, + 221 + ], + "spans": [ + { + "bbox": [ + 158, + 181, + 199, + 221 + ], + "type": "image", + "image_path": "a4aa1dab4ba65cffbeb0ab77bcfb50a4e214155315ec1b729b026dd972d344b3.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 174, + 222, + 228, + 233 + ], + "lines": [ + { + "bbox": [ + 174, + 222, + 228, + 233 + ], + "spans": [ + { + "bbox": [ + 174, + 222, + 228, + 233 + ], + "type": "text", + "content": "A towering red cat under a sleek black rose bush" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 131, + 236, + 480, + 258 + ], + "lines": [ + { + "bbox": [ + 131, + 236, + 480, + 258 + ], + "spans": [ + { + "bbox": [ + 131, + 236, + 480, + 258 + ], + "type": "text", + "content": "Fig. 5: Our curated test datasets feature captions paired with one positive image and one negative image each. All the positive images are displayed on the left side." + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 204, + 181, + 246, + 221 + ], + "blocks": [ + { + "bbox": [ + 204, + 181, + 246, + 221 + ], + "lines": [ + { + "bbox": [ + 204, + 181, + 246, + 221 + ], + "spans": [ + { + "bbox": [ + 204, + 181, + 246, + 221 + ], + "type": "image", + "image_path": "4a1925bcbfdcec961c90dc671bdf47b0a0634854ccdb825522e27c61c8aa1bda.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 262, + 181, + 304, + 221 + ], + "blocks": [ + { + "bbox": [ + 262, + 181, + 304, + 221 + ], + "lines": [ + { + "bbox": [ + 262, + 181, + 304, + 221 + ], + "spans": [ + { + "bbox": [ + 262, + 181, + 304, + 221 + ], + "type": "image", + "image_path": "e67574e55fd1405d34d64ee845c7fb588e5f4fa65d830c645968ef72ca6df915.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 275, + 222, + 335, + 233 + ], + "lines": [ + { + "bbox": [ + 275, + 222, + 335, + 233 + ], + "spans": [ + { + "bbox": [ + 275, + 222, + 335, + 233 + ], + "type": "text", + "content": "A drawing of a bird perching on a branch with three leaves" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 307, + 181, + 348, + 221 + ], + "blocks": [ + { + "bbox": [ + 307, + 181, + 348, + 221 + ], + "lines": [ + { + "bbox": [ + 307, + 181, + 348, + 221 + ], + "spans": [ + { + "bbox": [ + 307, + 181, + 348, + 221 + ], + "type": "image", + "image_path": "acaf7c4ff19ba7e0b34d962ecd7a6f5134d50536fece83fd46f6ffb9d2fdbe97.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 366, + 181, + 408, + 221 + ], + "blocks": [ + { + "bbox": [ + 366, + 181, + 408, + 221 + ], + "lines": [ + { + "bbox": [ + 366, + 181, + 408, + 221 + ], + "spans": [ + { + "bbox": [ + 366, + 181, + 408, + 221 + ], + "type": "image", + "image_path": "11d39e873742f7375db6fe9aef8e1bd92522319eade697bad63860142c29221b.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 377, + 222, + 441, + 233 + ], + "lines": [ + { + "bbox": [ + 377, + 222, + 441, + 233 + ], + "spans": [ + { + "bbox": [ + 377, + 222, + 441, + 233 + ], + "type": "text", + "content": "A cheeseburger on a plate with a soda on its right" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 410, + 181, + 451, + 221 + ], + "blocks": [ + { + "bbox": [ + 410, + 181, + 451, + 221 + ], + "lines": [ + { + "bbox": [ + 410, + 181, + 451, + 221 + ], + "spans": [ + { + "bbox": [ + 410, + 181, + 451, + 221 + ], + "type": "image", + "image_path": "aed5b308a20204c690c9c5b0ba9c5bc4c9a14d306770935f492a092c68e45c46.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "bbox": [ + 132, + 275, + 246, + 286 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 275, + 246, + 286 + ], + "spans": [ + { + "bbox": [ + 132, + 275, + 246, + 286 + ], + "type": "text", + "content": "4.6 Ablation Studies" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 130, + 294, + 482, + 402 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 294, + 482, + 402 + ], + "spans": [ + { + "bbox": [ + 130, + 294, + 482, + 402 + ], + "type": "text", + "content": "Importance of different negative data. Table 3 begins with our baseline model, LLaVA-1.5 [24], outlining its performance. The following two rows illustrate the effects of finetuning the LLaVA-1.5 model with only replace and swap negative data, with filtering processes applied. As shown in the subcategories of the SugarCrepe dataset, replace and swap data enhance performance in their respective categories but do not significantly impact other cases. Nevertheless, our final model excels across almost all tests, indicating that combining replace and swap negative data has a complementary effect, particularly noticeable in the add category of the SugarCrepe dataset and other evaluated datasets." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 130, + 403, + 482, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 403, + 482, + 521 + ], + "spans": [ + { + "bbox": [ + 130, + 403, + 482, + 521 + ], + "type": "text", + "content": "Importance of data filtering. The subsequent row presents outcomes from merging two types of negative data without any filtering. This combination has shown to be beneficial for specific datasets, like Winoground and SugarCrepe, although the results don't match those of the final model with filtering. However, biases in the text data cause the fine-tuned model to show no improvement on datasets like MagicBrush and SeeTRUE. Since data filtering reduces the amount of data, the fifth row displays results from randomly subsampling a training set to match the quantity of filtered data. This comparison highlights the effectiveness of our filtering technique as without it, the model may get biased to just rely on language distribution and ignore the image when making predictions." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 130, + 521, + 482, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 482, + 594 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 482, + 594 + ], + "type": "text", + "content": "Effect of data filtering. Fig. 6 shows the impact of removing the top " + }, + { + "bbox": [ + 130, + 521, + 482, + 594 + ], + "type": "inline_equation", + "content": "k\\%" + }, + { + "bbox": [ + 130, + 521, + 482, + 594 + ], + "type": "text", + "content": " of biased data on the performance across four test datasets. Here, " + }, + { + "bbox": [ + 130, + 521, + 482, + 594 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 521, + 482, + 594 + ], + "type": "text", + "content": " varies from 0 to 90, representing the progression from no data removal to the exclusion of " + }, + { + "bbox": [ + 130, + 521, + 482, + 594 + ], + "type": "inline_equation", + "content": "90\\%" + }, + { + "bbox": [ + 130, + 521, + 482, + 594 + ], + "type": "text", + "content": " of the data. The observed trend indicates that as the biased data is progressively removed, performance improves, peaking at approximately " + }, + { + "bbox": [ + 130, + 521, + 482, + 594 + ], + "type": "inline_equation", + "content": "30\\%" + }, + { + "bbox": [ + 130, + 521, + 482, + 594 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 130, + 521, + 482, + 594 + ], + "type": "inline_equation", + "content": "40\\%" + }, + { + "bbox": [ + 130, + 521, + 482, + 594 + ], + "type": "text", + "content": ". Beyond this point, performance declines due to the diminishing volume of training data." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": "We assess the quality of our filtered data by training a text-only model on " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "80\\%" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": " of the filtered data and testing it on the remaining " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": ". Note that this evaluation is performed on filtered data, and a model should reach an accuracy close to " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": " in ideal data indicating it cannot distinguish between positive and negative caption just using text. We vary the parameter " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "0\\%" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "90\\%" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": ", in increments of " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": ", resulting in corresponding filtered data quality percentages" + } + ] + } + ], + "index": 28 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 212, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 212, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 212, + 100 + ], + "type": "text", + "content": "Y. Li et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 133, + 141, + 479, + 236 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 479, + 137 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 479, + 137 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 479, + 137 + ], + "type": "text", + "content": "Table 3: Upper: Ablating our training data based on LLaVA-1.5 [24]. Lower: Finetuning a BLIP2 model with our data." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 133, + 141, + 479, + 236 + ], + "lines": [ + { + "bbox": [ + 133, + 141, + 479, + 236 + ], + "spans": [ + { + "bbox": [ + 133, + 141, + 479, + 236 + ], + "type": "table", + "html": "
WinogroundSeeTRUESugarCrepeMagicBrush
imagetextgroupDrawBenchEditBenchCOCO-T2Ireplaceswapadd
LLaVA-1.5 [24]49.7551.0034.2586.978.384.593.588.395.882.61
Only replace w/ filter51.2554.2538.5087.477.683.795.288.895.385.70
Only swap w/ filter63.5049.2542.2581.074.079.292.695.591.376.63
w/o filter65.7551.5046.7588.476.481.194.591.795.082.99
Random subsample64.2550.0043.2588.276.881.894.592.093.683.12
LLaVA-score (Ours)68.0053.7547.2588.877.784.995.394.997.587.28
BLIP2-ITM24.2541.7519.0060.867.568.088.983.991.875.32
BLIP2-ITM (ft on our data)39.542.752887.576.282.894.391.496.079.62
", + "image_path": "15cfe004653eb3ff500a87a3e4063ae37d8367056ad90b0d96f837ac26fc11d4.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 153, + 243, + 227, + 316 + ], + "blocks": [ + { + "bbox": [ + 153, + 243, + 227, + 316 + ], + "lines": [ + { + "bbox": [ + 153, + 243, + 227, + 316 + ], + "spans": [ + { + "bbox": [ + 153, + 243, + 227, + 316 + ], + "type": "image", + "image_path": "526aa469eb4024741f40c97100b02a1cb14eed3bcf28a9c8eee45fe6239b7081.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 319, + 479, + 341 + ], + "lines": [ + { + "bbox": [ + 132, + 319, + 479, + 341 + ], + "spans": [ + { + "bbox": [ + 132, + 319, + 479, + 341 + ], + "type": "text", + "content": "Fig. 6: Ablation on filter percentage of data. For most of the data, performance peaks around " + }, + { + "bbox": [ + 132, + 319, + 479, + 341 + ], + "type": "inline_equation", + "content": "30\\%" + }, + { + "bbox": [ + 132, + 319, + 479, + 341 + ], + "type": "text", + "content": " and reduces after that due to decrease in training data size." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 230, + 243, + 303, + 316 + ], + "blocks": [ + { + "bbox": [ + 230, + 243, + 303, + 316 + ], + "lines": [ + { + "bbox": [ + 230, + 243, + 303, + 316 + ], + "spans": [ + { + "bbox": [ + 230, + 243, + 303, + 316 + ], + "type": "image", + "image_path": "a8e31fa643745535ba26272c6d8b0122535148ebeb055f7665a4d5101b544d35.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 307, + 243, + 381, + 316 + ], + "blocks": [ + { + "bbox": [ + 307, + 243, + 381, + 316 + ], + "lines": [ + { + "bbox": [ + 307, + 243, + 381, + 316 + ], + "spans": [ + { + "bbox": [ + 307, + 243, + 381, + 316 + ], + "type": "image", + "image_path": "650a56a2d95f5561be9efc467fcc7c3624336a358d483f5e21ff185fe5623d2a.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 384, + 243, + 457, + 316 + ], + "blocks": [ + { + "bbox": [ + 384, + 243, + 457, + 316 + ], + "lines": [ + { + "bbox": [ + 384, + 243, + 457, + 316 + ], + "spans": [ + { + "bbox": [ + 384, + 243, + 457, + 316 + ], + "type": "image", + "image_path": "c7dfa83c56298c6d5d9e5256c21db02319ad2bd9894d2556b6dd50397a80b702.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 355, + 479, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 355, + 479, + 403 + ], + "spans": [ + { + "bbox": [ + 130, + 355, + 479, + 403 + ], + "type": "text", + "content": "of " + }, + { + "bbox": [ + 130, + 355, + 479, + 403 + ], + "type": "inline_equation", + "content": "75.9\\%" + }, + { + "bbox": [ + 130, + 355, + 479, + 403 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 355, + 479, + 403 + ], + "type": "inline_equation", + "content": "68.2\\%" + }, + { + "bbox": [ + 130, + 355, + 479, + 403 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 355, + 479, + 403 + ], + "type": "inline_equation", + "content": "60.7\\%" + }, + { + "bbox": [ + 130, + 355, + 479, + 403 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 355, + 479, + 403 + ], + "type": "inline_equation", + "content": "56.4\\%" + }, + { + "bbox": [ + 130, + 355, + 479, + 403 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 355, + 479, + 403 + ], + "type": "inline_equation", + "content": "53.8\\%" + }, + { + "bbox": [ + 130, + 355, + 479, + 403 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 355, + 479, + 403 + ], + "type": "inline_equation", + "content": "51.3\\%" + }, + { + "bbox": [ + 130, + 355, + 479, + 403 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 355, + 479, + 403 + ], + "type": "inline_equation", + "content": "50.6\\%" + }, + { + "bbox": [ + 130, + 355, + 479, + 403 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 355, + 479, + 403 + ], + "type": "inline_equation", + "content": "51.0\\%" + }, + { + "bbox": [ + 130, + 355, + 479, + 403 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 355, + 479, + 403 + ], + "type": "inline_equation", + "content": "49.8\\%" + }, + { + "bbox": [ + 130, + 355, + 479, + 403 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 355, + 479, + 403 + ], + "type": "inline_equation", + "content": "47.9\\%" + }, + { + "bbox": [ + 130, + 355, + 479, + 403 + ], + "type": "text", + "content": ". Empirically, we find that data quality below " + }, + { + "bbox": [ + 130, + 355, + 479, + 403 + ], + "type": "inline_equation", + "content": "60\\%" + }, + { + "bbox": [ + 130, + 355, + 479, + 403 + ], + "type": "text", + "content": " is satisfactory for our purposes. Consequently, we decide to discard " + }, + { + "bbox": [ + 130, + 355, + 479, + 403 + ], + "type": "inline_equation", + "content": "30\\%" + }, + { + "bbox": [ + 130, + 355, + 479, + 403 + ], + "type": "text", + "content": " of the data in our experiment as it's a good trade-off between quality of filtered data and amount of training data." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 404, + 480, + 463 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 404, + 480, + 463 + ], + "spans": [ + { + "bbox": [ + 130, + 404, + 480, + 463 + ], + "type": "text", + "content": "To demonstrate that the distribution gap we identified also exists in datasets created by others, we apply the same evaluation method to SugarCrepe [14], a test benchmark also generated by GPT, that was refined using grammar and common sense models. Nevertheless, it still exhibits an accuracy of " + }, + { + "bbox": [ + 130, + 404, + 480, + 463 + ], + "type": "inline_equation", + "content": "69.0\\%" + }, + { + "bbox": [ + 130, + 404, + 480, + 463 + ], + "type": "text", + "content": ", indicating the presence of the bias." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 464, + 481, + 608 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 464, + 481, + 608 + ], + "spans": [ + { + "bbox": [ + 130, + 464, + 481, + 608 + ], + "type": "text", + "content": "Generalization on other model. The bottom section of Table 3 shows that when BLIP2 is fine-tuned with the ITM head (a binary classification head) using our curated data, there is an observable improvement in performance. It's important to highlight that we only finetune and use the Q-former in BLIP2 without including any LLMs, and this model surpasses nearly all baselines in benchmarks that do not utilize LLMs (refer table 1). As expected, the fine-tuned BLIP2 model does not outperform the model fine-tuned using LLaVA, particularly on the Winoground dataset, which is recognized for its challenging reasoning tasks. This suggests the importance of utilizing a more advanced vision-language model as the foundation model for achieving better results. Nonetheless, given that BLIP2's model size (180M) is significantly smaller compared to LLaVA-1.5 (13B), it offers a more lightweight option for simpler applications." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 623, + 462, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 623, + 462, + 635 + ], + "spans": [ + { + "bbox": [ + 132, + 623, + 462, + 635 + ], + "type": "text", + "content": "4.7 Application: Text Alignment Ranking for Image Generation" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 641, + 479, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 479, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 479, + 665 + ], + "type": "text", + "content": "In this subsection, we showcase an application of our model for image generation. A major challenge with T2I models is their inconsistent adherence to complex" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 393, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 393, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 393, + 91, + 447, + 100 + ], + "type": "text", + "content": "LLaVA-score" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 161, + 114, + 455, + 250 + ], + "blocks": [ + { + "bbox": [ + 161, + 114, + 455, + 250 + ], + "lines": [ + { + "bbox": [ + 161, + 114, + 455, + 250 + ], + "spans": [ + { + "bbox": [ + 161, + 114, + 455, + 250 + ], + "type": "image", + "image_path": "ee0485ed87f1558edd4d3eb316beb9189231c49d98024acb91da0ed29d43b9cb.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 259, + 480, + 282 + ], + "lines": [ + { + "bbox": [ + 132, + 259, + 480, + 282 + ], + "spans": [ + { + "bbox": [ + 132, + 259, + 480, + 282 + ], + "type": "text", + "content": "Fig. 7: We show T2I generation results ranked by our model according to image-text alignment. The most aligned images are ranked higher." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 136, + 308, + 479, + 347 + ], + "blocks": [ + { + "bbox": [ + 138, + 293, + 473, + 304 + ], + "lines": [ + { + "bbox": [ + 138, + 293, + 473, + 304 + ], + "spans": [ + { + "bbox": [ + 138, + 293, + 473, + 304 + ], + "type": "text", + "content": "Table 4: Performance evaluation of ranking correlation on the TIFA [15] dataset." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 136, + 308, + 479, + 347 + ], + "lines": [ + { + "bbox": [ + 136, + 308, + 479, + 347 + ], + "spans": [ + { + "bbox": [ + 136, + 308, + 479, + 347 + ], + "type": "table", + "html": "
CLIP-ViT-L-14 [28]NegCLIP [40]BLIP2-ITM [18]BLIP2-ITC [18]ImageReward [38]VisualGPT [23]VQ2(T5) [39]TIFA[15]LLaVA-1.5[24]Ours
Spearman ρ28.540.143.351.562.836.955.159.761.564.5
Kendall τ19.528.729.936.946.328.741.547.245.849.8
", + "image_path": "643edbfeeb9a0f83e8507b11d1f70c1358bee5e00e0702bbdeb38b707cb4b8a0.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 369, + 480, + 405 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 369, + 480, + 405 + ], + "spans": [ + { + "bbox": [ + 130, + 369, + 480, + 405 + ], + "type": "text", + "content": "compositional prompts [10]. Inspired by [16], one can generate a series of images from a T2I model and then employ our model to rerank them based on relevance, ultimately presenting the highest-ranked images to the user." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 405, + 481, + 489 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 405, + 481, + 489 + ], + "spans": [ + { + "bbox": [ + 130, + 405, + 481, + 489 + ], + "type": "text", + "content": "TIFA [15] introduced a benchmark with images generated using different T2I models from identical prompts with human rankings. Our evaluation on this dataset reveals our model's superior alignment with human rankings, as shown in Table 4. Spearman's " + }, + { + "bbox": [ + 130, + 405, + 481, + 489 + ], + "type": "inline_equation", + "content": "\\rho" + }, + { + "bbox": [ + 130, + 405, + 481, + 489 + ], + "type": "text", + "content": " and Kendall's " + }, + { + "bbox": [ + 130, + 405, + 481, + 489 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 130, + 405, + 481, + 489 + ], + "type": "text", + "content": " are statistical measures that evaluate the similarity between two ranking orders, with higher values indicating greater alignment. Figure 7 illustrates qualitative ranking results using our model for the generation results from a T2I model [35]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 506, + 220, + 517 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 506, + 220, + 517 + ], + "spans": [ + { + "bbox": [ + 132, + 506, + 220, + 517 + ], + "type": "text", + "content": "5 Conclusion" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "content": "In this paper, we introduced an innovative method for generating high-quality training data for image-text alignment models. By implementing a mixed-type negative caption creation strategy and a novel filtering mechanism, we ensured a balanced distribution between positive and negative captions. This approach not only rectified the biases in existing models but also significantly enhanced the performance of our fine-tuned visual-language models, which achieve state-of-the-art results for image-text alignment. Our findings underscore the importance of high-quality, balanced training data in improving visual-language compositional reasoning and alignment. They also open up new avenues for exploring how these methods can be applied to other tasks or modalities in addition to the visual-language domain. In conclusion, our work offers an effective approach to improving image-text alignment, paving the road for future alignment studies." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 212, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 212, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 212, + 100 + ], + "type": "text", + "content": "Y. Li et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 201 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 201 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 201 + ], + "type": "text", + "content": "This work was supported in part by NSF CAREER IIS2150012, Adobe Data Science award, Microsoft Accelerate Foundation Models Research Program, and Institute of Information & communications Technology Planning & Evaluation (IITP) grants funded by the Korea government (MSIT) (No. 2022-0-00871, Development of AI Autonomy and Knowledge Enhancement for AI Agent Collaboration) and (No. RS-2022-00187238, Development of Large Korean Language Model Technology for Efficient Pre-training)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 220, + 198, + 232 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 220, + 198, + 232 + ], + "spans": [ + { + "bbox": [ + 132, + 220, + 198, + 232 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 134, + 245, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 138, + 245, + 480, + 277 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 245, + 480, + 277 + ], + "spans": [ + { + "bbox": [ + 138, + 245, + 480, + 277 + ], + "type": "text", + "content": "1. Alayrac, J.B., Donahue, J., Luc, P., Miech, A., Barr, I., Hasson, Y., Lenc, K., Mensch, A., Millican, K., Reynolds, M., et al.: Flamingo: a visual language model for few-shot learning. NeurIPS 35, 23716-23736 (2022)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 280, + 480, + 322 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 280, + 480, + 322 + ], + "spans": [ + { + "bbox": [ + 138, + 280, + 480, + 322 + ], + "type": "text", + "content": "2. Betker, J., Goh, G., Jing, L., Brooks, T., Wang, J., Li, L., Ouyang, L., Zhuang, J., Lee, J., Guo, Y., Manassra, W., Dhariwal, P., Chu, C., Jiao, Y., Ramesh, A.: Improving image generation with better captions (2023), https://cdn.openai.com/papers/dall-e-3.pdf" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 323, + 480, + 344 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 323, + 480, + 344 + ], + "spans": [ + { + "bbox": [ + 138, + 323, + 480, + 344 + ], + "type": "text", + "content": "3. Bitton, Y.: WYSIWYR repository. https://github.com/yonatanbitton/wysiwyr (2023), accessed: 2024-02-24" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 346, + 480, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 346, + 480, + 376 + ], + "spans": [ + { + "bbox": [ + 138, + 346, + 480, + 376 + ], + "type": "text", + "content": "4. Changpinyo, S., Sharma, P., Ding, N., Soricut, R.: Conceptual 12M: Pushing web-scale image-text pre-training to recognize long-tail visual concepts. In: CVPR (2021)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 379, + 480, + 400 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 379, + 480, + 400 + ], + "spans": [ + { + "bbox": [ + 138, + 379, + 480, + 400 + ], + "type": "text", + "content": "5. Chen, J., Guo, H., Yi, K., Li, B., Elhoseiny, M.: Visualgpt: Data-efficient adaptation of pretrained language models for image captioning (June 2022)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 402, + 480, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 402, + 480, + 477 + ], + "spans": [ + { + "bbox": [ + 138, + 402, + 480, + 477 + ], + "type": "text", + "content": "6. Chen, X., Wang, X., Changpinyo, S., Piergiovanni, A.J., Padlewski, P., Salz, D.M., Goodman, S., Grycner, A., Mustafa, B., Beyer, L., Kolesnikov, A., Puigcerver, J., Ding, N., Rong, K., Akbari, H., Mishra, G., Xue, L., Thapliyal, A.V., Bradbury, J., Kuo, W., Seyedhosseini, M., Jia, C., Ayan, B.K., Riquelme, C., Steiner, A., Angelova, A., Zhai, X., Houlsby, N., Soricut, R.: Pali: A jointly-scaled multilingual language-image model. ArXiv abs/2209.06794 (2022), https://apisemantic scholar.org/CorpusID:252222320" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 479, + 480, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 479, + 480, + 510 + ], + "spans": [ + { + "bbox": [ + 138, + 479, + 480, + 510 + ], + "type": "text", + "content": "7. Devlin, J., Chang, M.W., Lee, K., Toutanova, K.: Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805 (2018)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 512, + 480, + 554 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 512, + 480, + 554 + ], + "spans": [ + { + "bbox": [ + 138, + 512, + 480, + 554 + ], + "type": "text", + "content": "8. Doveh, S., Arbelle, A., Harary, S., Herzig, R., Kim, D., Cascante-Bonilla, P., Alfassy, A., Panda, R., Giryes, R., Feris, R., et al.: Dense and aligned captions (dac) promote compositional reasoning in vl models. Advances in Neural Information Processing Systems 36 (2024)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 555, + 480, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 555, + 480, + 599 + ], + "spans": [ + { + "bbox": [ + 138, + 555, + 480, + 599 + ], + "type": "text", + "content": "9. Doveh, S., Arbelle, A., Harary, S., Schwartz, E., Herzig, R., Giryes, R., Feris, R., Panda, R., Ullman, S., Karlinsky, L.: Teaching structured vision & language concepts to vision & language models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2657-2668 (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 134, + 600, + 480, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 600, + 480, + 643 + ], + "spans": [ + { + "bbox": [ + 134, + 600, + 480, + 643 + ], + "type": "text", + "content": "10. Feng, W., He, X., Fu, T.J., Jampani, V., Akula, A.R., Narayana, P., Basu, S., Wang, X.E., Wang, W.Y.: Training-free structured diffusion guidance for compositional text-to-image synthesis. In: The Eleventh International Conference on Learning Representations (2023), https://openreview.net/forum?id=PUIqjT4rzq7" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 134, + 644, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 644, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 134, + 644, + 480, + 665 + ], + "type": "text", + "content": "11. Gadre, S.Y., Ilharco, G., Fang, A., Hayase, J., Smyrnis, G., Nguyen, T., Marten, R., Wortsman, M., Ghosh, D., Zhang, J., et al.: Datacomp: In search of the next" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 393, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 393, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 393, + 91, + 447, + 100 + ], + "type": "text", + "content": "LLaVA-score" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 117, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 149, + 117, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 117, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 149, + 117, + 480, + 138 + ], + "type": "text", + "content": "generation of multimodal datasets. Advances in Neural Information Processing Systems 36 (2024)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 140, + 480, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 140, + 480, + 171 + ], + "spans": [ + { + "bbox": [ + 133, + 140, + 480, + 171 + ], + "type": "text", + "content": "12. Herzig, R., Mendelson, A., Karlinsky, L., Arbelle, A., Feris, R., Darrell, T., Globerson, A.: Incorporating structured representations into pretrained vision & language models using scene graphs. EMNLP (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 172, + 480, + 247 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 172, + 480, + 247 + ], + "spans": [ + { + "bbox": [ + 133, + 172, + 480, + 247 + ], + "type": "text", + "content": "13. Hessel, J., Holtzman, A., Forbes, M., Le Bras, R., Choi, Y.: CLIPScore: A reference-free evaluation metric for image captioning. In: Moens, M.F., Huang, X., Specia, L., Yih, S.W.t. (eds.) Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing. pp. 7514-7528. Association for Computational Linguistics, Online and Punta Cana, Dominican Republic (Nov 2021). https://doi.org/10.18653/v1/2021.emnlp-main.595, https://aclanthology.org/2021.emnlp-main.595" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 249, + 480, + 292 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 249, + 480, + 292 + ], + "spans": [ + { + "bbox": [ + 133, + 249, + 480, + 292 + ], + "type": "text", + "content": "14. Hsieh, C.Y., Zhang, J., Ma, Z., Kembhavi, A., Krishna, R.: Sugarcrepe: Fixing hackable benchmarks for vision-language compositionality. In: Thirty-Seventh Conference on Neural Information Processing Systems Datasets and Benchmarks Track (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 133, + 293, + 480, + 325 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 293, + 480, + 325 + ], + "spans": [ + { + "bbox": [ + 133, + 293, + 480, + 325 + ], + "type": "text", + "content": "15. Hu, Y., Liu, B., Kasai, J., Wang, Y., Ostendorf, M., Krishna, R., Smith, N.A.: Tifa: Accurate and interpretable text-to-image faithfulness evaluation with question answering. ICCV (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 133, + 326, + 480, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 326, + 480, + 357 + ], + "spans": [ + { + "bbox": [ + 133, + 326, + 480, + 357 + ], + "type": "text", + "content": "16. Karthik, S., Roth, K., Mancini, M., Akata, Z.: If at first you don't succeed, try, try again: Faithful diffusion-based text-to-image generation by selection. arXiv preprint arXiv:2305.13308 (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 133, + 358, + 480, + 390 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 358, + 480, + 390 + ], + "spans": [ + { + "bbox": [ + 133, + 358, + 480, + 390 + ], + "type": "text", + "content": "17. Le, T., Lal, V., Howard, P.: Coco-counterfactuals: Automatically constructed counterfactual examples for image-text pairs. Advances in Neural Information Processing Systems 36 (2024)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 133, + 392, + 480, + 434 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 392, + 480, + 434 + ], + "spans": [ + { + "bbox": [ + 133, + 392, + 480, + 434 + ], + "type": "text", + "content": "18. Li, J., Li, D., Savarese, S., Hoi, S.C.H.: Blip-2: Bootstrapping language-image pretraining with frozen image encoders and large language models. In: International Conference on Machine Learning (2023), https://apisemantic scholar.org/CorpusID:256390509" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 133, + 435, + 480, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 435, + 480, + 468 + ], + "spans": [ + { + "bbox": [ + 133, + 435, + 480, + 468 + ], + "type": "text", + "content": "19. Li, J., Li, D., Xiong, C., Hoi, S.: Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In: International Conference on Machine Learning. pp. 12888-12900. PMLR (2022)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 133, + 468, + 480, + 490 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 468, + 480, + 490 + ], + "spans": [ + { + "bbox": [ + 133, + 468, + 480, + 490 + ], + "type": "text", + "content": "20. Li, Y., Du, Y., Zhou, K., Wang, J., Zhao, W.X., Wen, J.R.: Evaluating object hallucination in large vision-language models. EMNLP (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 133, + 491, + 480, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 491, + 480, + 533 + ], + "spans": [ + { + "bbox": [ + 133, + 491, + 480, + 533 + ], + "type": "text", + "content": "21. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dólar, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13. pp. 740-755. Springer (2014)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 133, + 534, + 480, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 534, + 480, + 555 + ], + "spans": [ + { + "bbox": [ + 133, + 534, + 480, + 555 + ], + "type": "text", + "content": "22. Lin, Z., Chen, X., Pathak, D., Zhang, P., Ramanan, D.: Revisiting the role of language priors in vision-language models. arXiv preprint arXiv:2306.01879 (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 133, + 556, + 480, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 556, + 480, + 588 + ], + "spans": [ + { + "bbox": [ + 133, + 556, + 480, + 588 + ], + "type": "text", + "content": "23. Lin, Z., Chen, X., Pathak, D., Zhang, P., Ramanan, D.: Visualgptscore: Visio-linguistic reasoning with multimodal generative pre-training scores. arXiv preprint arXiv:2306.01879 (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 133, + 589, + 480, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 589, + 480, + 610 + ], + "spans": [ + { + "bbox": [ + 133, + 589, + 480, + 610 + ], + "type": "text", + "content": "24. Liu, H., Li, C., Li, Y., Lee, Y.J.: Improved baselines with visual instruction tuning (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 133, + 611, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 611, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 133, + 611, + 480, + 632 + ], + "type": "text", + "content": "25. Liu, H., Li, C., Wu, Q., Lee, Y.J.: Visual instruction tuning. arXiv:2304.08485 (2023)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 133, + 633, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 633, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 633, + 480, + 665 + ], + "type": "text", + "content": "26. Ma, Z., Hong, J., Gul, M.O., Gandhi, M., Gao, I., Krishna, R.: Crepe: Can vision-language foundation models reason compositionally? arXiv preprint arXiv:2212.07796 (2022)" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 212, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 212, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 212, + 100 + ], + "type": "text", + "content": "Y. Li et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 117, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 132, + 117, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 117, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 132, + 117, + 480, + 138 + ], + "type": "text", + "content": "27. OpenAI: Gpt-4v(ision) system card (2023), https://cdn.openai.com/papers/GPTV_System_Card.pdf" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 140, + 480, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 140, + 480, + 182 + ], + "spans": [ + { + "bbox": [ + 132, + 140, + 480, + 182 + ], + "type": "text", + "content": "28. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 183, + 480, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 183, + 480, + 215 + ], + "spans": [ + { + "bbox": [ + 132, + 183, + 480, + 215 + ], + "type": "text", + "content": "29. Ramesh, A., Dhariwal, P., Nichol, A., Chu, C., Chen, M.: Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125 (2022)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 216, + 480, + 247 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 216, + 480, + 247 + ], + "spans": [ + { + "bbox": [ + 132, + 216, + 480, + 247 + ], + "type": "text", + "content": "30. Ray, A., Radenovic, F., Dubey, A., Plummer, B.A., Krishna, R., Saenko, K.: Cola: How to adapt vision-language models to compose objects localized with attributes? (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 249, + 480, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 249, + 480, + 281 + ], + "spans": [ + { + "bbox": [ + 132, + 249, + 480, + 281 + ], + "type": "text", + "content": "31. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10684-10695 (2022)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 282, + 480, + 325 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 282, + 480, + 325 + ], + "spans": [ + { + "bbox": [ + 132, + 282, + 480, + 325 + ], + "type": "text", + "content": "32. Schuhmann, C., Beaumont, R., Vencu, R., Gordon, C., Wightman, R., Cherti, M., Coombes, T., Katta, A., Mullis, C., Wortsman, M., et al.: Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems 35, 25278-25294 (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 326, + 480, + 390 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 326, + 480, + 390 + ], + "spans": [ + { + "bbox": [ + 132, + 326, + 480, + 390 + ], + "type": "text", + "content": "33. Sharma, P., Ding, N., Goodman, S., Soricut, R.: Conceptual captions: A cleaned, hypernymed, image alt-text dataset for automatic image captioning. In: Gurevych, I., Miyao, Y. (eds.) Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). pp. 2556-2565. Association for Computational Linguistics, Melbourne, Australia (Jul 2018). https://doi.org/10.18653/v1/P18-1238, https://aclanthology.org/P18-1238" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 392, + 480, + 434 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 392, + 480, + 434 + ], + "spans": [ + { + "bbox": [ + 132, + 392, + 480, + 434 + ], + "type": "text", + "content": "34. Singh, A., Hu, R., Goswami, V., Couairon, G., Galuba, W., Rohrbach, M., Kiela, D.: Flava: A foundational language and vision alignment model. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15638-15650 (2022)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 435, + 405, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 435, + 405, + 445 + ], + "spans": [ + { + "bbox": [ + 132, + 435, + 405, + 445 + ], + "type": "text", + "content": "35. stabilityai: IF repository. https://github.com/deep-floyd/IF" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 447, + 480, + 489 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 447, + 480, + 489 + ], + "spans": [ + { + "bbox": [ + 132, + 447, + 480, + 489 + ], + "type": "text", + "content": "36. Thrush, T., Jiang, R., Bartolo, M., Singh, A., Williams, A., Kiela, D., Ross, C.: Winoground: Probing vision and language models for visio-linguistic compositionality. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5238-5248 (2022)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 491, + 480, + 522 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 491, + 480, + 522 + ], + "spans": [ + { + "bbox": [ + 132, + 491, + 480, + 522 + ], + "type": "text", + "content": "37. Wang, Z.J., Montoya, E., Munechika, D., Yang, H., Hoover, B., Chau, D.H.: Large-scale prompt gallery dataset for text-to-image generative models. arXiv:2210.14896 [cs] (2022), https://arxiv.org/abs/2210.14896" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 523, + 480, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 523, + 480, + 555 + ], + "spans": [ + { + "bbox": [ + 132, + 523, + 480, + 555 + ], + "type": "text", + "content": "38. Xu, J., Liu, X., Wu, Y., Tong, Y., Li, Q., Ding, M., Tang, J., Dong, Y.: Imagereward: Learning and evaluating human preferences for text-to-image generation. Advances in Neural Information Processing Systems 36 (2024)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 556, + 480, + 587 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 556, + 480, + 587 + ], + "spans": [ + { + "bbox": [ + 132, + 556, + 480, + 587 + ], + "type": "text", + "content": "39. Yarom, M., Bitton, Y., Changpinyo, S., Aharoni, R., Herzig, J., Lang, O., Ofek, E., Szpektor, I.: What you see is what you read? improving text-image alignment evaluation. arXiv preprint arXiv:2305.10400 (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 589, + 480, + 631 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 589, + 480, + 631 + ], + "spans": [ + { + "bbox": [ + 132, + 589, + 480, + 631 + ], + "type": "text", + "content": "40. Yuksekgonul, M., Bianchi, F., Kalluri, P., Jurafsky, D., Zou, J.: When and why vision-language models behave like bags-of-words, and what to do about it? In: International Conference on Learning Representations (2023), https://openreview.net/forum?id=KRLUvxh8uaX" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "type": "text", + "content": "41. Zhang, K., Mo, L., Chen, W., Sun, H., Su, Y.: Magicbrush: A manually annotated dataset for instruction-guided image editing. In: Advances in Neural Information Processing Systems (2023)" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 393, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 393, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 393, + 91, + 447, + 100 + ], + "type": "text", + "content": "LLaVA-score" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 481, + 150 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 481, + 150 + ], + "type": "text", + "content": "42. Zhao, T., Zhang, T., Zhu, M., Shen, H., Lee, K., Lu, X., Yin, J.: Vl-checklist: Evaluating pre-trained vision-language models with objects, attributes and relations. arXiv preprint arXiv:2207.00221 (2022)" + } + ] + } + ], + "index": 2 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 211, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 211, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 211, + 100 + ], + "type": "text", + "content": "Y. Li et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Removing Rows and Columns of Tokens in Vision Transformer enables Faster Dense Prediction without Retraining/8e177374-55c7-4790-9a57-d4f7dcccc52f_content_list.json b/2024/Removing Rows and Columns of Tokens in Vision Transformer enables Faster Dense Prediction without Retraining/8e177374-55c7-4790-9a57-d4f7dcccc52f_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..4938d0f64c353aa4575eb968e09132b939f53a08 --- /dev/null +++ b/2024/Removing Rows and Columns of Tokens in Vision Transformer enables Faster Dense Prediction without Retraining/8e177374-55c7-4790-9a57-d4f7dcccc52f_content_list.json @@ -0,0 +1,1820 @@ +[ + { + "type": "text", + "text": "Removing Rows and Columns of Tokens in Vision Transformer enables Faster Dense Prediction without Retraining", + "text_level": 1, + "bbox": [ + 217, + 140, + 785, + 209 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Diwei $\\mathrm{Su}^{1}$ Cheng Fei, and Jianxu Luo", + "bbox": [ + 326, + 234, + 673, + 250 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "East China University of Science and Technology, Shanghai 200237, China dvsu@mail.ecust.edu.cn, cfei_jarvis@163.com, jxluo@ecust.edu.cn", + "bbox": [ + 253, + 262, + 750, + 291 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. In recent years, vision transformers based on self-attention mechanisms have demonstrated remarkable abilities in various tasks such as natural language processing, computer vision (CV), and multimodal applications. However, due to the high computational costs and the structural nature of images, the application of transformers to CV tasks faces challenges, particularly when handling ultra-high-resolution images. Recently, several token reduction methods have been proposed to improve the computational efficiency of transformers by reducing the number of tokens without the need for retraining. These methods primarily involve fusion based on matching or clustering. The former exhibits faster speed but suffers more accuracy loss compared to the latter. In this work, we propose a simple matching-based fusion method called Token Adapter, which achieves comparable accuracy to the clustering-based fusion method with faster speed and demonstrates higher potential in terms of robustness. Our method was applied to Segmenter, MaskDINO and SWAG, exhibiting promising performance on four tasks, including semantic segmentation, instance segmentation, panoptic segmentation, and image classification. Specifically, our method can be applied to Segmenter on ADE20k, providing $41\\%$ frames per second (FPS) acceleration while maintaining full performance without retraining or fine-tuning off-the-shelf weights. Our code will be released at https://github.com/MilknoCandy-Token-Adapter.", + "bbox": [ + 261, + 321, + 738, + 625 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Vision transformer $\\cdot$ Token reduction $\\cdot$ Dense prediction $\\cdot$ Efficient transformer", + "bbox": [ + 261, + 638, + 738, + 666 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 689, + 375, + 705 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Over the past few years, the computer vision field has been significantly advanced by the emergence of the Transformer architecture [31]. Vision Transformers [9] with global attention mechanism achieve state-of-the-art results on a wide range of vision tasks. To obtain better performance, researchers have scaled up the number of parameters of vision transformers to billions [15] through techniques such as self-supervised pre-training [14, 29] methods and adding visual prior knowledge [5, 22], etc. However, scaling up transformers brings performance improvements as well as huge computational costs due to the quadratic complexity", + "bbox": [ + 212, + 719, + 787, + 840 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/11284078890d1f4e17bd04b746a116476f5b9fa83092c7665867cb0034b02708.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 241, + 143, + 326, + 210 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/aff19d31763915ef52c1adbc17b1b77c1703f123491aca922a2010c494f27443.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 241, + 212, + 325, + 277 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/0e838eb5eca54fbb0f9687fd066b4d3eb6975387aa82cfb02ea75db3ccc68493.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 328, + 145, + 411, + 210 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/934332e7e73f9229188040a7940112769bd60c3dc3f68ed7bccb584f16870919.jpg", + "image_caption": [ + "Fig. 1: Visualizations of random pruning results from different methods on ADE20K. We randomly prune the same number of tokens using different methods. (a) Input Images. (b) Randomly remove tokens (c) Randomly remove specific rows and columns." + ], + "image_footnote": [], + "bbox": [ + 328, + 212, + 411, + 277 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/266a7240c18e3b8b09fd39f5288d1e8162dcd9a5ad549a91a78f7d5292a7d453.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 415, + 145, + 500, + 210 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/ca8ab54901f1f71dd1e6fb61e82be9bf7ccdcb4e954c8eb7468681d2d787d974.jpg", + "image_caption": [ + "(b)" + ], + "image_footnote": [], + "bbox": [ + 415, + 212, + 500, + 277 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/c2fab848ce3bf0583ce6b711a03d9aca0ecf50bc81d131ad4ede2fa888893143.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 500, + 145, + 586, + 210 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/363888ef80f14b88a682fc169a7183b73f3897f46f54ebd503ee60e876dba8bd.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 500, + 212, + 586, + 277 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/c34d7a81f395ec1c70d792e18796b20faf073be08fbae2afa28b156d2c644b78.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 589, + 145, + 674, + 210 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/686ac65b26deb62e8d9b0675da316d45bef98c4b20fffb5f125ab43a8c53fb88.jpg", + "image_caption": [ + "(c)" + ], + "image_footnote": [], + "bbox": [ + 589, + 212, + 674, + 277 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/a559b89f0ba42406227626182b87aacdfa1b529fc939e7e3f41ac8e797d81eca.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 676, + 145, + 761, + 210 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/85a5c927fd3dfbc2948526b57536194e3b96824df5546d0d8aa1e16f97b1e3fa.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 676, + 212, + 761, + 277 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "of self-attention mechanism. Vision transformers are extremely inefficient when processing high-resolution images. Designing efficient vision transformers has become a prominent research hotspot.", + "bbox": [ + 212, + 388, + 784, + 434 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Some researchers have reduced the computational complexity to linear by optimizing the computation of self-attention mechanism [24, 30, 39, 42], while others are focusing on reducing the number of tokens to increase the inference efficiency of vision transformers [3, 10, 32, 44]. Despite significantly improving the efficiency of the vision transformer, most of these methods necessitate the retraining or fine-tuning of the optimized Transformer, demanding substantial computational resources.", + "bbox": [ + 212, + 435, + 785, + 539 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "There are also methods that do not require fine-tuning or retraining, such as Expediting [19] and ToMe (Token Merging) [2]. Expediting integrates local information to fuse and reconstruct tokens using a clustering algorithm, and ToMe fuses the corresponding tokens based on a bipartite matching algorithm. Previous methods either selected the most representative tokens based on their importance scores or employed clustering algorithms to obtain more representative tokens. The former approach involves calculating importance scores for all tokens, introducing significant computation, and necessitating operations such as padding to ensure uniform dimensions across different batches for batch processing. Although the latter method effectively compresses token sequences, parameter tuning for clustering algorithms poses a challenge, and the iterative calculation approach incurs additional computational costs.", + "bbox": [ + 212, + 541, + 787, + 720 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this work, we present Token Adapter to merge rows and columns of tokens rather than merge tokens sparsely for the following three reasons:", + "bbox": [ + 214, + 722, + 785, + 753 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. Information Redundancy. In vision tasks, multi-scale features naturally provide more information than single-scale features, resulting in spatial information redundancy in multi-scale image features.", + "2. Image Structure. As image data is derived from sampling real-world optical information, semantic information is contained within consecutive pixel" + ], + "bbox": [ + 220, + 763, + 784, + 839 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "D. Su et al.", + "bbox": [ + 271, + 114, + 349, + 126 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "blocks rather than individual pixels. With the same number of missing pixels, removing entire rows or columns of pixels from an image could be less detrimental to the existing semantic information than the pruning partial pixels approach.", + "bbox": [ + 240, + 146, + 785, + 205 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Computational Efficiency. To speed up vision transformers, one simple idea is that using fewer tokens can make them more efficient. Wang et al. [32] has observed that simple images can be accurately predicted with only a few tokens. This observation highlights the inconsistency in the number of tokens required to make accurate predictions for images of varying content complexity.", + "bbox": [ + 222, + 207, + 787, + 297 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Fig. 1 shows the difference between our approach and the previous methods, i.e., the change in token selection strategy. It allows our method to compress token sequences efficiently and handle batch data directly without additional operations. Our experiments show that Token Adapter outperforms Expediting and ToMe on multiple segmentation datasets. To summarize, our main contributions can be outlined as follows: (i) We propose a simple yet powerful module that injects normal tokens into representative tokens while preserving unique tokens associated with dense prediction tasks; (ii) Our token reduction method preserves relative positional information among tokens, providing ViT-like models with substantial acceleration and slight accuracy loss in dense prediction tasks; (iii) Without the need for retraining or fine-tuning, our approach achieves performance comparable to clustering-based methods by employing matching-based techniques, demonstrating even better stability and speed.", + "bbox": [ + 212, + 304, + 787, + 503 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 522, + 387, + 539 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Efficient Vision Transformer. Some studies have focused on adapting the internal configuration of transformer models to improve their efficiency. These efforts include designing the architecture of the transformer [4,18,30,35], modifying attention mechanisms [13,21,34,38], approximating attention [6,11,12,26], etc.", + "bbox": [ + 212, + 551, + 787, + 627 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Token Reduction. Benefiting from the ability of Transformers to handle variable-length sequences, some recent work has attempted to optimize intermediate representations within transformers. Previous approaches could be categorized by their purpose into two groups: Pruning [16, 33, 36, 37] and Fusion [1, 23, 40, 41]. From a methodological perspective, these approaches can be divided into two categories: recognition-based and non-recognition-based methods. Recognition-based methods categorize all tokens into two groups, vital and minor, performing pruning or fusion operations on them separately based on specific rules. On the other hand, non-recognition-based methods employ clustering algorithms to cluster all tokens, reducing the number of tokens while preserving as much semantic information as possible.", + "bbox": [ + 212, + 628, + 787, + 794 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our Approach. We decompose token compression into row and column compression, providing a more efficient token reduction method. Additionally, we found that some rows or columns in tokens are unique and cannot be pruned", + "bbox": [ + 212, + 795, + 787, + 840 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "Token Adapter", + "bbox": [ + 629, + 114, + 730, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "or merged. In our experiments, retaining these rows or columns yields better results than not retaining them.", + "bbox": [ + 212, + 146, + 785, + 176 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3 Methodology", + "text_level": 1, + "bbox": [ + 215, + 199, + 380, + 217 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1 Overview", + "text_level": 1, + "bbox": [ + 215, + 229, + 341, + 244 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Designing for dense prediction tasks, our token adapter comprises two simple match-based approaches: the token injector and the token ejector. Token adapter utilize a subset of tokens for the purpose of representing another subset of tokens while maintaining the integrity of the original token set.", + "bbox": [ + 212, + 255, + 785, + 315 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Pipeline. As shown in Fig. 2, two modules of the token adapter are inserted in the middle of consecutive transformer blocks. Given a feature map $X \\in \\mathbb{R}^{H \\times W \\times C}$ , we compress rows and columns with reduction ratios $r_h$ and $r_w$ respectively, reducing the number of tokens by $n_r$ , where $n_r = H \\times W \\times (r_h + r_w - r_h \\times r_w)$ . Here, $r_h$ and $r_w$ are settable parameters. Finally, we rearrange the compressed tokens to rebuild the original token sequence. The details are introduced in the next section.", + "bbox": [ + 212, + 316, + 785, + 421 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2 Token Adapter", + "text_level": 1, + "bbox": [ + 215, + 443, + 385, + 459 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Token Injector. For sequence input $X_{s} \\in \\mathbb{R}^{B \\times N \\times C}$ , we first reshape it to $X \\in \\mathbb{R}^{B \\times H \\times W \\times C}$ , where $N = H \\times W$ represents the number of tokens. To identify which rows and columns to merge, importance scores are computed for each row and column. In images, areas with sharp pixel value changes usually correspond to inter-object or intra-object boundaries. Moreover, areas with larger pixel values compared to their surroundings are typically more significant.", + "bbox": [ + 212, + 467, + 785, + 559 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In our experiments, we perform two steps of dimension reduction on features to obtain their importance scores. For the input feature $X \\in \\mathbb{R}^{B \\times H \\times W \\times C}$ , we first calculate the maximum value of each token vector (you can also calculate its sum value or range value instead), thereby compressing the features into $X_{hw} \\in \\mathbb{R}^{B \\times H \\times W}$ . Subsequently, the same procedure is applied to each row and column, yielding $S_h \\in \\mathbb{R}^{B \\times H}$ and $S_w \\in \\mathbb{R}^{B \\times W}$ as the importance scores. Please note that we employed a straightforward method to calculate the importance scores, yet we believe that a better importance assessment algorithm could yield better results.", + "bbox": [ + 212, + 560, + 785, + 694 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Based on these scores, tokens are categorized into three groups: representative tokens, unique tokens, and normal tokens. Normal tokens are injected into representative tokens, while unique tokens remain intact. The compressed tokens are rearranged in their previous positions. A brief flow is delineated below:", + "bbox": [ + 212, + 695, + 785, + 756 + ], + "page_idx": 3 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. Calculate the importance scores for each row and column.", + "2. Group all tokens into: representative tokens, unique tokens and normal tokens by importance scores.", + "3. Match the most similar representative token for each normal tokens.", + "4. Merge normal tokens with corresponding representative tokens." + ], + "bbox": [ + 220, + 763, + 785, + 839 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "D. Su et al.", + "bbox": [ + 271, + 114, + 349, + 127 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/e855a7467c28b6fa09f0abbe5aedc80e37aab77478316a324c3b442234517fb2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 274, + 143, + 730, + 287 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/4119ace7158c446cc6f91d0c27bfd84d4d7fe52b11155638be2856f7e258471a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 274, + 289, + 728, + 373 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/d32acbea98682afb06ebaabd3abbe575d3adbfec9a1e3eef412538759f03c285.jpg", + "image_caption": [ + "Fig. 2: Token Adapter. (a) We use Token Injector module and Token ejector module to wrap $\\mathcal{L}_m$ Transformer blocks. (b) Token Injector mitigates the computational overhead of subsequent Transformer blocks by injecting normal tokens into representative tokens. (c) By leveraging token-to-token relationships, we represent the original sequence with compressed tokens." + ], + "image_footnote": [], + "bbox": [ + 274, + 373, + 728, + 518 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5. Rearrange new representative tokens and unique tokens in their original order.", + "bbox": [ + 220, + 630, + 784, + 659 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Within an image, consecutive pixel blocks form regions containing semantic information, and these associated semantic regions collectively constitute complete image objects. Merging entire rows or columns preserves the continuity of the original semantic regions while maintaining their relative positions, as depicted in Fig. 3. Our key thought involves representing each object through a refined set of representative and unique tokens. The results presented in Fig. 6 indicate that the retention of partial unique tokens generates better prediction accuracy.", + "bbox": [ + 212, + 672, + 787, + 792 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Token Ejector. After compression, each representative token contains a blend of information from the original representative tokens and the merged normal tokens. Therefore, we exploit the correlation (e.g., the distance matrix", + "bbox": [ + 212, + 795, + 785, + 839 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "Token Adapter", + "bbox": [ + 629, + 114, + 730, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/8ba4edbdf1fddf24fab397428548bebfd6c5b3f10c011b8b6b52b1d59a34d7a8.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 243, + 143, + 370, + 339 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/6c703140f6caf47edbaf3769f69371c4e1554b1865a177e99be2dce5093ea2db.jpg", + "image_caption": [ + "(b)" + ], + "image_footnote": [], + "bbox": [ + 372, + 143, + 500, + 339 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/409d5efa72c217c265f049798c9bac729e27cc5e644b06991b14ff099326e6e2.jpg", + "image_caption": [ + "(c)", + "Fig. 3: Token Selection Visualizations. Results obtained through the removal of specific rows and columns from the ADE20K validation dataset using Segmenter+ViT-L/16 equipped with Token Adapter ( $r_h = r_w = 0.3$ ). (a) Input Images. (b) Remove specific rows and columns. (c) Input images in which each patch is filled with the mean color of its respective region. (d) Remove specific rows and columns from (c)." + ], + "image_footnote": [], + "bbox": [ + 501, + 143, + 630, + 339 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/5522f4c3a7fbe32357c767eb7b6ada4c7e23ffb61a8ada1d60afc16a5b2e753c.jpg", + "image_caption": [ + "(d)" + ], + "image_footnote": [], + "bbox": [ + 632, + 143, + 759, + 339 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "$M_{d}$ ) between the reduced sequence and the original sequence to restore the original length. Illustrated in Fig. 2, the primary steps are as follows:", + "bbox": [ + 212, + 464, + 784, + 496 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. Compute the distance matrix $M_{d} \\in \\mathbb{R}^{N \\times N_{r}}$ between $X \\in \\mathbb{R}^{N \\times C}$ and $X' \\in \\mathbb{R}^{N_{r} \\times C}$ , with $N_{r} = N - n_{r}$ denoting the length of reduced sequence.", + "2. Perform token matching between tokens in $X$ and $X^{\\prime}$ according to $M_{d}$ .", + "3. Reconstruct the original sequence $\\tilde{X}$ by aggregating tokens in $X_r'$ in accordance with the alignment results." + ], + "bbox": [ + 220, + 503, + 784, + 580 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Note that the distance matrix is computed from the original sequence $X_{s}$ and the reduced sequence $X^{\\prime}$ that has not been refined by the subsequent transformer blocks. By following this approach, the reduced sequence can be efficiently reverted to its initial length.", + "bbox": [ + 212, + 590, + 787, + 652 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 215, + 676, + 375, + 694 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1 Setup", + "text_level": 1, + "bbox": [ + 215, + 708, + 310, + 724 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Dataset and evaluation metrics. We present task-specific metrics, including the mean Intersection-over-Union (mIoU) and average precision (AP), etc. Furthermore, we compare our method with others by analyzing model inference speed and floating-point operations (FLOPs). The datasets utilized for evaluation include ADE20K [43], PASCAL-Context [25], CityScapes [7], COCO [20], and ImageNet-1k [8], covering semantic segmentation, instance segmentation, panotic segmentation, object detection, and image classification tasks.", + "bbox": [ + 212, + 733, + 787, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "D. Su et al.", + "bbox": [ + 271, + 114, + 349, + 127 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation details. To assess the generalization of our method, we equipped two models with Token Adapter and conducted a series of experiments on four well-known datasets for dense prediction tasks. The proposed module is inserted into two classic models off-the-shelf with no re-training or fine-tuning required. All experiments are performed on a single NVIDIA GeForce RTX3090 GPU using Pytorch framework. Further details are presented in the following sections. Our code implementation is based on the open-source code of Expediting and ToMe.", + "bbox": [ + 212, + 146, + 787, + 268 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2 Main Results", + "text_level": 1, + "bbox": [ + 214, + 287, + 372, + 301 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this section, we delve into a comparative analysis of the proposed approach with other token reduction methods that do not require retraining and are applicable to dense prediction tasks. Since the number of GFLOPs does not provide a direct measure of model inference efficiency, our primary concern lies in accuracy metrics for dense prediction tasks and corresponding metrics for evaluating inference speed.", + "bbox": [ + 212, + 309, + 787, + 402 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d26110e380ee614d43021ab9cb1491e427a2bd2ec4287dd696f206de5ccd4e97.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 222, + 424, + 408, + 534 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/4f626b9755485388a28079aa8f711693e86678c1d3af5a3222513f36d5566951.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 408, + 425, + 593, + 534 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/5e39608380b77b7f6b34d0ed8520db5ad1bf13056eb6ee51d1af53eb9596c6d8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 596, + 425, + 779, + 534 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7d4f30a73a80dfa0133d4f816af59c8f22d6c376b380d198c82fb8a768510a19.jpg", + "image_caption": [ + "(a)", + "(d)", + "Fig. 4: Comparisons on Semantic Segmentation Tasks. A comprehensive comparison of our approach with Expediting and ToMe on ADE20K (1-st column), PACAL-Context (2-ed column), and Cityscapes (3-rd column) using Segmenter + ViT-L/16." + ], + "image_footnote": [], + "bbox": [ + 223, + 549, + 406, + 659 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/3b8bda187a68c6030b7991bd2325e2a906e952ed5659e422dbb389400d7879f8.jpg", + "image_caption": [ + "(b)", + "(e)" + ], + "image_footnote": [], + "bbox": [ + 410, + 551, + 591, + 659 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ddb848b8b2fa57243098d9e6e05b1490a466a0112fca2e734f47e78f9ba5fb95.jpg", + "image_caption": [ + "(c)", + "(f)" + ], + "image_footnote": [], + "bbox": [ + 596, + 551, + 777, + 659 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Comparisons on Semantic Segmentation Task. We use the ViT-like segmentation network model Segmenter [28] as our baseline for three benchmarks, including ADE20K, PASCAL-Context, and Cityscapes, reporting semantic segmentation results. We compare our approach with Expediting1 [19] and", + "bbox": [ + 212, + 756, + 787, + 818 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "Token Adapter", + "bbox": [ + 629, + 114, + 730, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 114, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "page_footnote", + "text": "$^{1}$ https://github.com/Expedit-LargeScale-Vision-Transformer", + "bbox": [ + 217, + 824, + 669, + 840 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/8c7995e057e6afae0f03c0a182dd415702d9b35a921950a80dc9235367e19d3d.jpg", + "table_caption": [ + "Table 1: Comparisons on Semantic Segmentation Tasks. Quantitative comparisons of our method with other token reduction approaches that do not require retraining on semantic segmentation tasks." + ], + "table_footnote": [], + "table_body": "
DatasetMethodmIoU(%)FPS(img/s)FLOPs(G)
ADE20KSegmenter + ViT-L/1651.8210.68658.98
+ Token Merging47.1217.15321.29
50.9714.02448.7
+ Expediting43.8622.4309.96
51.8514.94481.9
+ Token Adapter (Ours)45.7324.41190.74
52.0215.09273.0
PASCAL-ContextSegmenter + ViT-L/1658.0724.17338.74
+ Token Merging54.2421.84172.38
57.2321.39236.5
+ Expediting48.0350.2889.66
58.1434.50220.76
+ Token Adapter (Ours)47.8063.3465.29
57.8734.53141.5
CityscapesSegmenter + ViT-L/1679.147.42995.55
+ Token Merging71.2313.12448.5
77.818.88651.5
+ Expediting55.3916.68406.69
78.4610.17681.57
+ Token Adapter (Ours)72.0819.38239.65
78.4810.2376.16
", + "bbox": [ + 243, + 198, + 761, + 563 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "$\\mathrm{ToMe}^2$ [2], which are clustering-based and matching-based methods for token reduction that do not require retraining, respectively. The results are reported in Tab. 1. The two lines for each method in Tab. 1 represent the best result and the maximum speedup that can be obtained with the same token compression settings, respectively.", + "bbox": [ + 212, + 589, + 784, + 665 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To ensure fairness and comprehensive performance comparisons among different methods, we fine-tuned the hyper-parameters of each method starting from the default settings mentioned in their papers. The results are depicted in Fig. 4. Under maximum acceleration, Token Adapter exhibits superior stability to Expediting $(+6.11\\% \\mathrm{mIoU}$ in average) and achieves higher velocity compared to ToMe $(+93.33\\% \\mathrm{fps}$ in average). Refer to our supplementary material for detailed parameter settings.", + "bbox": [ + 212, + 665, + 787, + 770 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Comparisons on Instance/Panoptic Segmentation Task. To further explore the generalization capabilities of our method across various Transformer frameworks, we have selected the latest framework, MaskDINO [17]. It utilized", + "bbox": [ + 212, + 771, + 787, + 816 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "D. Su et al.", + "bbox": [ + 271, + 114, + 349, + 127 + ], + "page_idx": 7 + }, + { + "type": "page_footnote", + "text": "$^{2}$ https://github.com/facebookresearch/ToMe", + "bbox": [ + 217, + 824, + 542, + 839 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Swin Transformer as its backbone network and has exhibited remarkable performance in tasks like instance segmentation and panoptic segmentation on COCO. We employed MaskDINO as our baseline model to compare the overall performance of our approach with Expediting on the COCO dataset, as presented in Tab. 2.", + "bbox": [ + 212, + 146, + 787, + 220 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/88b76c1f5f7f31fbb2f7f6e22f61c5301f7168dff29d21a536e83a7bef9a41a7.jpg", + "table_caption": [ + "Table 2: Comparisons on Instance/Panoptic Segmentation Tasks. Inserting both the proposed algorithm and Expediting at different stages of MaskDINO + Swin-L, we compare their comprehensive performance on the COCO dataset." + ], + "table_footnote": [], + "table_body": "
MethodStagePQ(%)MaskAP(%)FPS(img/s)FLOPs(G)
MaskDINO + Swin-L-58.352.36.11936.86
+ Expediting130.722.676.05920.88
233.123.566.04918.68
357.9851.916.88810.94
+ Token Adapter (ours)458.2952.236.14917.03
155.9149.236.31915.74
256.0349.056.22916.11
357.5751.426.96809.66
458.4252.346.19916.38
", + "bbox": [ + 246, + 299, + 756, + 484 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Due to the window attention mechanism, segmenting all tokens into windows, we reduce the tokens within each window instead of the entire token sequence. To ensure a fair comparison, we gradually increase $\\mathcal{L}_b$ starting from the setting where Expediting balances efficiency and accuracy, to showcase the impact of applying this method at different stages of MaskDINO.", + "bbox": [ + 212, + 515, + 787, + 589 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "By observing Tab. 2, it is evident that our approach is more stable than Expediting. This could be attributed to the fact that we directly combine and reconstruct tokens through matching, whereas Expediting relies on clustering, which is sensitive to hyper-parameters for different scenarios. On top of that, with the integration of token adapter into the fourth stage of the backbone network of MaskDINO, we can observe a slight enhancement in performance $(+0.1\\% \\mathrm{PQ})$ .", + "bbox": [ + 212, + 590, + 787, + 696 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Comparisons on Image Classification Task. We applied three methods to SWAG[27] for testing on the ImageNet-1k dataset. The results are shown in Tab. 3, where $\\dagger$ indicates that the same fusion method is applied as Token Merging, i.e., layer-by-layer compression, and 32 indicates merging one row per layer. 'Size' indicates the resolution of the feature map. 'r' denotes the number of tokens compressed per layer.", + "bbox": [ + 212, + 696, + 787, + 787 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Visualization of the token injector. We visualize the token injection process in Fig. 5. Initially, we divide all tokens into three groups: representative", + "bbox": [ + 212, + 787, + 787, + 818 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "Token Adapter", + "bbox": [ + 629, + 114, + 730, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 114, + 785, + 126 + ], + "page_idx": 8 + }, + { + "type": "page_footnote", + "text": "3 https://github.com/facebookresearch/SWAG", + "bbox": [ + 217, + 824, + 545, + 839 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/70cd2c895d44e6b466ecbb0ce0faa79b61735849a62eece8e6038370ede23a6f.jpg", + "table_caption": [ + "Table 3: Comparisons on ImageNet1k validation dataset" + ], + "table_footnote": [], + "table_body": "
MethodSizerAcc(%)FPS(img/s)FLOPs(G)
SWAG + ViT-L/1632 × 32-88.0722.92364.77
+ Token Adapter23 × 23-87.4332.28270.65
+ Expediting23 × 23-87.829.21275.79
+ Token Adapter†-3287.8736.49224.02
+ Token Merging-3287.8038.58218.18
+ Token Adapter†-6485.3466.04130.36
+ Token Merging-6467.5679.99116.40
", + "bbox": [ + 272, + 170, + 730, + 303 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "tokens, unique tokens, and normal tokens. Subsequently, we inject the normal tokens into the representative tokens. In the third row of Fig. 5, we use different colored borders to represent the matching results of normal tokens and representative tokens. The last row in the figure represents the results of token injection. By observing this, we can notice that token injection effectively highlights the crucial pixels in the image while preserving the original semantic information. More results and hyper-parameter settings can be found in the supplementary material.", + "bbox": [ + 212, + 333, + 787, + 454 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.3 Ablation Studies", + "text_level": 1, + "bbox": [ + 215, + 479, + 401, + 494 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "We performed ablation experiments for each critical choice within our approach on ADE20K semantic segmentation task. The results can be found in Tab. 4. By default, we employed the official weights provided by Segmenter4 as a baseline model (Segmenter + ViT-L/16, mIoU: 51.82%, FPS: 10.68 img/s). The table cells marked in red and purple respectively denote the final choices for the token injector and token ejector.", + "bbox": [ + 212, + 508, + 787, + 599 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Token Reduction Strategy. We initially conducted ablation experiments on the token selection method, and the results are presented in Tab. 4a. In this table, random signifies the stochastic selection, alternating represents the periodic selection, and scores indicates the selection based on computed scores derived from features or attention scores. random and alternating represent the two selection strategies of ToMe. Selecting tokens based on calculated importance scores derived from features presents promising performance.", + "bbox": [ + 212, + 599, + 787, + 705 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Tab. 4b presents the influence of different evaluation metrics employed to calculate the distance among tokens. We utilize this distance matrix to perform token matching. Among these metrics, the cosine similarity distance metric demonstrates superior performance, hence we utilize cosine similarity to measure the distance between tokens. By comparing the pairwise distances between normal tokens and representative tokens, we can match each normal token with its closest representative token.", + "bbox": [ + 212, + 707, + 787, + 811 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "D. Su et al.", + "bbox": [ + 271, + 114, + 349, + 127 + ], + "page_idx": 9 + }, + { + "type": "page_footnote", + "text": "4 https://github.com/rstrudel/segmenter", + "bbox": [ + 217, + 824, + 521, + 839 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/d14396cbf5c1d9f4fde64417d905f21a6e8df91e0060f9782fd0b0c7417522a1.jpg", + "image_caption": [ + "Fig. 5: Token Injection Visualizations. Visualization results from Segmenter + ViT-L/16 applying Token Injector on ADE20K. The first row depicts the input image. The second row illustrates the grouping of all tokens, where gray represents representative tokens, orange represents unique tokens, and teal represents normal tokens. The third row displays the matching results between normal tokens and representative tokens. Different pairs are indicated by boxes of varying colors. The final row showcases the injected results." + ], + "image_footnote": [], + "bbox": [ + 215, + 142, + 787, + 364 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/76f898105d1f49a683be9f26cf0c238ab46e5cc10ff2225a68345de9f34884fc.jpg", + "image_caption": [ + "(a) compressed token" + ], + "image_footnote": [], + "bbox": [ + 217, + 500, + 485, + 626 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/e6c3b523fc076bd87025a383466d5652519ba9db59ff60e17a011cf8eb9ee0c1.jpg", + "image_caption": [ + "(b) unique token", + "Fig.6: Ablation studies on the ratio of rows and columns of tokens. (a) Ratio ablation of compressed rows or columns $(r_h$ and $r_w)$ . (b) Ablation of retention ratios for unique token rows or columns." + ], + "image_footnote": [], + "bbox": [ + 516, + 500, + 785, + 626 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "To obtain a reduced token sequence, we employ several methods, as shown in Tab. 4d. Keep denotes the preservation of representative tokens and unique tokens without merging normal tokens. Max pool and avg pool refer to pooling-based methods for token fusion, while bilinear represents resizing token sequence through bilinear interpolation. Average indicates averaging tokens within each matched group to merge them. Keep and average do not rely on pre-defined priors to adjust the tokens, hence yielding better results. However, keep leads", + "bbox": [ + 212, + 734, + 787, + 840 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "Token Adapter", + "bbox": [ + 629, + 114, + 730, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/28177be2e3ae1ad6f577eba021c481e991ce17f2766c274130c554eaa4f5ac42.jpg", + "table_caption": [ + "Table 4: Token Adapter ablation experiments using Segmenter + ViT/L-16 on ADE20K validation dataset with official checkpoint and no retraining. $r_h$ and $r_w$ are set to 0.3. Segmenter + ViT/L-16 without Token Adapter achieves 51.82% mIoU at 10.68 img/s. We provide the mean Intersection-over-Union (mIoU) and model inference speed (img/s)." + ], + "table_footnote": [], + "table_body": "
SelectionmIoUFPS
random51.3815.16
alternating51.4314.96
scores (feat.)52.0215.09
scores (attn.)51.1315.04
(a) Token Selection.
(b) Distance Metric.
(c) Feature Choice.
", + "bbox": [ + 235, + 224, + 776, + 325 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/00179f96dce25aa7d61948baa38a2ab90fb6102c253fd729b8cfc0aacaef087e.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodmIoUFPS
keep50.6715.19
max pool48.1514.97
avg pool48.1015.10
bilinear50.0815.12
average52.0215.09
", + "bbox": [ + 292, + 345, + 450, + 446 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/be24dc1a9fa078c2d3a5f06aa0fc649c57a6eb13431cec720743aaed28ce3c26.jpg", + "table_caption": [ + "(d) Merge Method." + ], + "table_footnote": [], + "table_body": "
ModemIoUFPS
max51.9315.13
m-threshold52.0215.09
", + "bbox": [ + 527, + 345, + 705, + 405 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "(e) Matching Mode.", + "bbox": [ + 550, + 406, + 683, + 419 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "to the loss of information from normal tokens, thus we ultimately opt for using average to merge the tokens.", + "bbox": [ + 212, + 503, + 784, + 532 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Restoration Method. Considering dependency on the correlation between the reduced sequence and its original counterpart to restore the sequence length, it is imperative to contemplate the selection of features utilized in calculating this relationship matrix. Within Tab. 4c, the superscript $u$ designates unmerged features, while $m$ signifies merged features. The variable $X_{p}$ represents representative tokens, while $X_{r}$ pertains to reduced tokens.", + "bbox": [ + 212, + 536, + 787, + 626 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "To facilitate understanding of the meaning of each symbol, we next summarize the algorithmic procedures of the token adapter, and omit the calculation process. For $X \\in \\mathbb{R}^{N \\times C}$ , we first partition tokens into $X_{p} \\in \\mathbb{R}^{N_{p} \\times C}$ , $X_{n} \\in \\mathbb{R}^{N_{n} \\times C}$ and $X_{u} \\in \\mathbb{R}^{N_{u} \\times C}$ . Here, we haven't merged $X_{n}$ into $X_{p}$ yet, so we can mark it as $X_{p}^{u}$ . $X_{r}$ is formed by combining $X_{p}^{u}$ and $X_{u}$ , representing the compressed sequence without merging. After merging operations, the new $X_{p}$ is marked as $X_{p}^{m}$ and the new $X_{r}$ is marked as $X_{r}^{m}$ . From the results, it shows that recovery using the merged and reduced token sequence $X_{r}^{m}$ works best.", + "bbox": [ + 212, + 627, + 787, + 748 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "A comparison result is presented for the two matching modes applied to the token ejector in Tab. 4e. In Sec. 3.2, we introduced the token ejector, which matches the most similar tokens for each token in the original token sequence according to the distance matrix $M_{d} \\in \\mathbb{R}^{N \\times N_{r}}$ . Next, the token vectors from each matched group are averaged to reconstruct the token sequence to its original size. Mode max stands for matching the most similar token for each token of", + "bbox": [ + 212, + 750, + 787, + 839 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "D. Su et al.", + "bbox": [ + 271, + 114, + 349, + 127 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "the original token sequence, i.e., one-to-one matching. Mode $m$ -threshold, on the other hand, involves filtering based on a set threshold and distance matrix after Mode $max$ matching. The filtered matching results are combined with the results from Mode $max$ , followed by executing the corresponding fusion process. The results indicate that the latter achieves more accurate recovery of the token sequence than the former.", + "bbox": [ + 212, + 146, + 787, + 236 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "We also conduct ablation experiments on the ratio of compressed rows, columns, and unique tokens, and the results are shown in Fig. 6. Although we determined the final parameter settings for our method based on the ablation experiments, we believe that there still remains a better solution. It will be discussed in the next section.", + "bbox": [ + 212, + 238, + 787, + 313 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "5 Discussion and Conclusion", + "text_level": 1, + "bbox": [ + 214, + 339, + 509, + 354 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "5.1 Limitation", + "text_level": 1, + "bbox": [ + 214, + 373, + 349, + 387 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "The experimental results mentioned earlier have already demonstrated the effectiveness of our method on dense prediction tasks for hierarchical transformer structures and ViT-like structures. However, there are still many aspects of our work that remain to be explored, which we will analyze one by one next.", + "bbox": [ + 212, + 400, + 785, + 460 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Generalizability. While our method has been validated for effectiveness in multiple segmentation tasks, it has not been validated on other more complex vision tasks. Previous methods essentially involved sparse sampling of token sequences, resulting in compressed token sequences being in sequence form only. The compressed sequence features can only be processed by a sequence processing module such as Transformer. Nevertheless, it remains to be examined whether discrete sparse sequence features are more effective than dense features for image tasks. Motivated by this, we designed the token adapter, a method that performs sparse sampling on the rows and columns of image features. Compared to previous methods, our approach yields more compact sampled features, which may be more effective for image tasks.", + "bbox": [ + 212, + 460, + 787, + 626 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Adaptability. Although our core idea is rather simple, the Token Adapter still takes time to tune the hyper-parameters. The current setting allows Token Adapter to be directly applied to existing models, but it also limits its further development. We chose not to retrain the model equipped with the token adapter to directly evaluate the effectiveness of our idea. In the future, we will continue to explore more efficient neural network acceleration schemes.", + "bbox": [ + 212, + 628, + 787, + 717 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Our core contribution is to compress and reconstruct rows and columns of image features in a way that conforms to the structure of the image. When applied to vision models, Token Adapter only compresses and reconstructs image features and thus can be applied to most vision models. For large language models, the concept of a Token Adapter can still be effectively transferred. By decomposing text embeddings based on the contextual information, and then leveraging the connections between the decomposed parts to compress them, it should be able to improve the efficiency of large language models.", + "bbox": [ + 212, + 719, + 787, + 839 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "Token Adapter", + "bbox": [ + 629, + 114, + 730, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "5.2 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 146, + 354, + 161 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In this work, we propose Token Adapter to expedite vision transformers used in dense prediction tasks. By injecting normal tokens into representative tokens, we enhance the efficiency of vision transformers. Notably, our method can be applied in off-the-shelf models directly without additional retraining or finetuning. Tab. 2 reveals that when applied to the early stages of MaskDINO, the clustering-based Expediting method results in a significant decline in accuracy, and the matching-based Token Adapter method demonstrates superior stability. Clustering algorithms typically require fine-tuning of multiple hyper-parameters, which remain fixed once adjusted. Our approach is expected to leverage mechanisms such as cross-attention to transform fixed hyper-parameters into dynamically learnable parameters, thereby enhancing the method's generalization capability. Theoretically, the design of the compression strategy and reconstruction strategy in our algorithm allows the token adapter to be inserted in the middle of any module designed for image tasks.", + "bbox": [ + 212, + 174, + 787, + 385 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In principle, Token Adapter bears resemblance to an unparameterized form of cross-attention mechanism, both blending information from one set of tokens into another set. Our thought lies in how we employ tokens to represent semantic entities within the image. Each entity in the image can be divided into distinct local semantic regions. By assigning several representative tokens to each local semantic region and injecting the remaining normal tokens into their corresponding tokens, we achieve a more refined fusion. After the fusion, the representative tokens representing different semantic regions will engage in information exchange within transformers, greatly enhancing computational efficiency and promoting diversity among tokens. We hope our research can offer a fresh perspective on image features and inspire the design of more efficient methods for extracting image features.", + "bbox": [ + 212, + 387, + 787, + 568 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 215, + 594, + 323, + 609 + ], + "page_idx": 13 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Bian, Z., Wang, Z., Han, W., Wang, K.: Multi-scale and token merging: Make your vit more efficient. arXiv preprint arXiv:2306.04897 (2023)", + "2. Bolya, D., Fu, C.Y., Dai, X., Zhang, P., Feichtenhofer, C., Hoffman, J.: Token merging: Your vit but faster. In: The Eleventh International Conference on Learning Representations (2022)", + "3. Bonnaerens, M., Dambre, J.: Learned thresholds token merging and pruning for vision transformers. arXiv preprint arXiv:2307.10780 (2023)", + "4. Carion, N., Massa, F., Synnaeve, G., Usunier, N., Kirillov, A., Zagoruyko, S.: End-to-end object detection with transformers. In: European conference on computer vision. pp. 213-229. Springer (2020)", + "5. Chen, Z., Duan, Y., Wang, W., He, J., Lu, T., Dai, J., Qiao, Y.: Vision transformer adapter for dense predictions. arXiv preprint arXiv:2205.08534 (2022)", + "6. Choromanski, K., Likhosherstov, V., Dohan, D., Song, X., Gane, A., Sarlos, T., Hawkins, P., Davis, J., Mohiuddin, A., Kaiser, L., et al.: Rethinking attention with performers. arXiv preprint arXiv:2009.14794 (2020)" + ], + "bbox": [ + 222, + 628, + 785, + 840 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "D. Su et al.", + "bbox": [ + 271, + 114, + 349, + 127 + ], + "page_idx": 13 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "7. Cordts, M., Omran, M., Ramos, S., Rehfeld, T., Enzweiler, M., Benenson, R., Franke, U., Roth, S., Schiele, B.: The cityscapes dataset for semantic urban scene understanding. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 3213-3223 (2016)", + "8. Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: Imagenet: A large-scale hierarchical image database. In: 2009 IEEE conference on computer vision and pattern recognition. pp. 248-255. IEEE (2009)", + "9. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929 (2020)", + "0. Fayyaz, M., Koohpayegani, S.A., Jafari, F.R., Sengupta, S., Joze, H.R.V., Sommerlade, E., Pirsiavash, H., Gall, J.: Adaptive token sampling for efficient vision transformers. In: European Conference on Computer Vision. pp. 396-414. Springer (2022)", + "1. Geng, Z., Guo, M.H., Chen, H., Li, X., Wei, K., Lin, Z.: Is attention better than matrix decomposition? arXiv preprint arXiv:2109.04553 (2021)", + "2. Guo, M.H., Liu, Z.N., Mu, T.J., Hu, S.M.: Beyond self-attention: External attention using two linear layers for visual tasks. IEEE Transactions on Pattern Analysis and Machine Intelligence 45(5), 5436-5447 (2022)", + "3. Hassani, A., Walton, S., Li, J., Li, S., Shi, H.: Neighborhood attention transformer. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6185-6194 (2023)", + "4. He, K., Chen, X., Xie, S., Li, Y., Dólar, P., Girshick, R.: Masked autoencoders are scalable vision learners. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 16000-16009 (2022)", + "5. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., et al.: Segment anything. arXiv preprint arXiv:2304.02643 (2023)", + "6. Kong, Z., Dong, P., Ma, X., Meng, X., Sun, M., Niu, W., Shen, X., Yuan, G., Ren, B., Qin, M., et al.: Spvit: Enabling faster vision transformers via soft token pruning. arXiv preprint arXiv:2112.13890 (2021)", + "7. Li, F., Zhang, H., Xu, H., Liu, S., Zhang, L., Ni, L.M., Shum, H.Y.: Mask dino: Towards a unified transformer-based framework for object detection and segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3041-3050 (2023)", + "8. Li, Y., Yuan, G., Wen, Y., Hu, J., Evangelidis, G., Tulyakov, S., Wang, Y., Ren, J.: Efficientformer: Vision transformers at mobilenet speed. Advances in Neural Information Processing Systems 35, 12934-12949 (2022)", + "9. Liang, W., Yuan, Y., Ding, H., Luo, X., Lin, W., Jia, D., Zhang, Z., Zhang, C., Hu, H.: Expediting large-scale vision transformer for dense prediction without fine-tuning. Advances in Neural Information Processing Systems 35, 35462-35477 (2022)", + "20. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollar, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13. pp. 740-755. Springer (2014)", + "21. Liu, Z., Hu, H., Lin, Y., Yao, Z., Xie, Z., Wei, Y., Ning, J., Cao, Y., Zhang, Z., Dong, L., et al.: Swin transformer v2: Scaling up capacity and resolution. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 12009-12019 (2022)" + ], + "bbox": [ + 225, + 146, + 785, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "Token Adapter", + "bbox": [ + 629, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 785, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "22. Liu, Z., Lin, Y., Cao, Y., Hu, H., Wei, Y., Zhang, Z., Lin, S., Guo, B.: Swin transformer: Hierarchical vision transformer using shifted windows. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 10012-10022 (2021)", + "23. Long, S., Zhao, Z., Pi, J., Wang, S., Wang, J.: Beyond attentive tokens: Incorporating token importance and diversity for efficient vision transformers. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10334-10343 (2023)", + "24. Lu, J., Yao, J., Zhang, J., Zhu, X., Xu, H., Gao, W., Xu, C., Xiang, T., Zhang, L.: Soft: Softmax-free transformer with linear complexity. Advances in Neural Information Processing Systems 34, 21297-21309 (2021)", + "25. Mottaghi, R., Chen, X., Liu, X., Cho, N.G., Lee, S.W., Fidler, S., Urtasun, R., Yuille, A.: The role of context for object detection and semantic segmentation in the wild. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 891-898 (2014)", + "26. Peng, H., Pappas, N., Yogatama, D., Schwartz, R., Smith, N.A., Kong, L.: Random feature attention. arXiv preprint arXiv:2103.02143 (2021)", + "27. Singh, M., Gustafson, L., Adcock, A., de Freitas Reis, V., Gedik, B., Kosaraju, R.P., Mahajan, D., Girshick, R., Dollár, P., Van Der Maaten, L.: Revisiting weakly supervised pre-training of visual perception models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 804-814 (2022)", + "28. Strudel, R., Garcia, R., Laptev, I., Schmid, C.: Segmenter: Transformer for semantic segmentation. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 7262-7272 (2021)", + "29. Tian, K., Jiang, Y., Dao, Q., Lin, C., Wang, L., Yuan, Z.: Designing bert for convolutional networks: Sparse and hierarchical masked modeling. arXiv preprint arXiv:2301.03580 (2023)", + "30. Tolstikhin, I.O., Houlsby, N., Kolesnikov, A., Beyer, L., Zhai, X., Unterthiner, T., Yung, J., Steiner, A., Keysers, D., Uszkoreit, J., et al.: Mlp-mixer: An allmlp architecture for vision. Advances in neural information processing systems 34, 24261-24272 (2021)", + "31. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. Advances in neural information processing systems 30 (2017)", + "32. Wang, Y., Huang, R., Song, S., Huang, Z., Huang, G.: Not all images are worth 16x16 words: Dynamic transformers for efficient image recognition. Advances in Neural Information Processing Systems 34, 11960-11973 (2021)", + "33. Wang, Z., Luo, H., Wang, P., Ding, F., Wang, F., Li, H.: Vtc-lfc: Vision transformer compression with low-frequency components. Advances in Neural Information Processing Systems 35, 13974-13988 (2022)", + "34. Yang, J., Li, C., Zhang, P., Dai, X., Xiao, B., Yuan, L., Gao, J.: Focal self-attention for local-global interactions in vision transformers. arXiv preprint arXiv:2107.00641 (2021)", + "35. Yang, R., Ma, H., Wu, J., Tang, Y., Xiao, X., Zheng, M., Li, X.: Scalablevit: Rethinking the context-oriented generalization of vision transformer. In: European Conference on Computer Vision. pp. 480-496. Springer (2022)", + "36. Ye, D., Lin, Y., Huang, Y., Sun, M.: Tr-bert: Dynamic token reduction for accelerating bert inference. arXiv preprint arXiv:2105.11618 (2021)", + "37. Yin, H., Vahdat, A., Alvarez, J.M., Mallya, A., Kautz, J., Molchanov, P.: A-vit: Adaptive tokens for efficient vision transformer. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10809-10818 (2022)" + ], + "bbox": [ + 215, + 146, + 785, + 840 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "D. Su et al.", + "bbox": [ + 271, + 114, + 349, + 127 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "38. Yu, Q., Xia, Y., Bai, Y., Lu, Y., Yuille, A.L., Shen, W.: Glance-and-gaze vision transformer. Advances in Neural Information Processing Systems 34, 12992-13003 (2021)", + "39. Yu, W., Luo, M., Zhou, P., Si, C., Zhou, Y., Wang, X., Feng, J., Yan, S.: Metaformer is actually what you need for vision. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10819-10829 (2022)", + "40. Zeng, W., Jin, S., Liu, W., Qian, C., Luo, P., Ouyang, W., Wang, X.: Not all tokens are equal: Human-centric visual analysis via token clustering transformer. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11101-11111 (2022)", + "41. Zeng, Z., Hawkins, C., Hong, M., Zhang, A., Pappas, N., Singh, V., Zheng, S.: Vcc: Scaling transformers to 128k tokens or more by prioritizing important tokens. arXiv preprint arXiv:2305.04241 (2023)", + "42. Zheng, M., Gao, P., Zhang, R., Li, K., Wang, X., Li, H., Dong, H.: End-to-end object detection with adaptive clustering transformer. arXiv preprint arXiv:2011.09315 (2020)", + "43. Zhou, B., Zhao, H., Puig, X., Xiao, T., Fidler, S., Barriuso, A., Torralba, A.: Semantic understanding of scenes through the ade20k dataset. International Journal of Computer Vision 127, 302-321 (2019)", + "44. Ziwen, C., Patnaik, K., Zhai, S., Wan, A., Ren, Z., Schwing, A.G., Colburn, A., Fuxin, L.: Autofocusformer: Image segmentation off the grid. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18227-18236 (2023)" + ], + "bbox": [ + 212, + 146, + 787, + 465 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "Token Adapter", + "bbox": [ + 629, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + } +] \ No newline at end of file diff --git a/2024/Removing Rows and Columns of Tokens in Vision Transformer enables Faster Dense Prediction without Retraining/8e177374-55c7-4790-9a57-d4f7dcccc52f_model.json b/2024/Removing Rows and Columns of Tokens in Vision Transformer enables Faster Dense Prediction without Retraining/8e177374-55c7-4790-9a57-d4f7dcccc52f_model.json new file mode 100644 index 0000000000000000000000000000000000000000..ff2c69f2adcc3152c2339f8e596acf76ce68fe79 --- /dev/null +++ b/2024/Removing Rows and Columns of Tokens in Vision Transformer enables Faster Dense Prediction without Retraining/8e177374-55c7-4790-9a57-d4f7dcccc52f_model.json @@ -0,0 +1,2522 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.218, + 0.141, + 0.787, + 0.21 + ], + "angle": 0, + "content": "Removing Rows and Columns of Tokens in Vision Transformer enables Faster Dense Prediction without Retraining" + }, + { + "type": "text", + "bbox": [ + 0.327, + 0.235, + 0.674, + 0.251 + ], + "angle": 0, + "content": "Diwei \\(\\mathrm{Su}^{1}\\) Cheng Fei, and Jianxu Luo" + }, + { + "type": "text", + "bbox": [ + 0.254, + 0.263, + 0.75, + 0.292 + ], + "angle": 0, + "content": "East China University of Science and Technology, Shanghai 200237, China dvsu@mail.ecust.edu.cn, cfei_jarvis@163.com, jxluo@ecust.edu.cn" + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.322, + 0.74, + 0.626 + ], + "angle": 0, + "content": "Abstract. In recent years, vision transformers based on self-attention mechanisms have demonstrated remarkable abilities in various tasks such as natural language processing, computer vision (CV), and multimodal applications. However, due to the high computational costs and the structural nature of images, the application of transformers to CV tasks faces challenges, particularly when handling ultra-high-resolution images. Recently, several token reduction methods have been proposed to improve the computational efficiency of transformers by reducing the number of tokens without the need for retraining. These methods primarily involve fusion based on matching or clustering. The former exhibits faster speed but suffers more accuracy loss compared to the latter. In this work, we propose a simple matching-based fusion method called Token Adapter, which achieves comparable accuracy to the clustering-based fusion method with faster speed and demonstrates higher potential in terms of robustness. Our method was applied to Segmenter, MaskDINO and SWAG, exhibiting promising performance on four tasks, including semantic segmentation, instance segmentation, panoptic segmentation, and image classification. Specifically, our method can be applied to Segmenter on ADE20k, providing \\(41\\%\\) frames per second (FPS) acceleration while maintaining full performance without retraining or fine-tuning off-the-shelf weights. Our code will be released at https://github.com/MilknoCandy-Token-Adapter." + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.64, + 0.74, + 0.667 + ], + "angle": 0, + "content": "Keywords: Vision transformer \\(\\cdot\\) Token reduction \\(\\cdot\\) Dense prediction \\(\\cdot\\) Efficient transformer" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.69, + 0.377, + 0.707 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Over the past few years, the computer vision field has been significantly advanced by the emergence of the Transformer architecture [31]. Vision Transformers [9] with global attention mechanism achieve state-of-the-art results on a wide range of vision tasks. To obtain better performance, researchers have scaled up the number of parameters of vision transformers to billions [15] through techniques such as self-supervised pre-training [14, 29] methods and adding visual prior knowledge [5, 22], etc. However, scaling up transformers brings performance improvements as well as huge computational costs due to the quadratic complexity" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.351, + 0.127 + ], + "angle": 0, + "content": "D. Su et al." + }, + { + "type": "image", + "bbox": [ + 0.242, + 0.145, + 0.327, + 0.211 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.242, + 0.213, + 0.326, + 0.278 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.329, + 0.146, + 0.413, + 0.211 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.329, + 0.213, + 0.413, + 0.278 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.318, + 0.28, + 0.338, + 0.293 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.416, + 0.146, + 0.501, + 0.211 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.416, + 0.213, + 0.501, + 0.278 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.491, + 0.28, + 0.513, + 0.293 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.146, + 0.587, + 0.211 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.213, + 0.587, + 0.278 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.591, + 0.146, + 0.676, + 0.211 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.591, + 0.213, + 0.676, + 0.278 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.678, + 0.146, + 0.762, + 0.211 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.678, + 0.213, + 0.762, + 0.278 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.668, + 0.28, + 0.686, + 0.292 + ], + "angle": 0, + "content": "(c)" + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.303, + 0.787, + 0.358 + ], + "angle": 0, + "content": "Fig. 1: Visualizations of random pruning results from different methods on ADE20K. We randomly prune the same number of tokens using different methods. (a) Input Images. (b) Randomly remove tokens (c) Randomly remove specific rows and columns." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.39, + 0.785, + 0.435 + ], + "angle": 0, + "content": "of self-attention mechanism. Vision transformers are extremely inefficient when processing high-resolution images. Designing efficient vision transformers has become a prominent research hotspot." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.436, + 0.787, + 0.54 + ], + "angle": 0, + "content": "Some researchers have reduced the computational complexity to linear by optimizing the computation of self-attention mechanism [24, 30, 39, 42], while others are focusing on reducing the number of tokens to increase the inference efficiency of vision transformers [3, 10, 32, 44]. Despite significantly improving the efficiency of the vision transformer, most of these methods necessitate the retraining or fine-tuning of the optimized Transformer, demanding substantial computational resources." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.542, + 0.788, + 0.722 + ], + "angle": 0, + "content": "There are also methods that do not require fine-tuning or retraining, such as Expediting [19] and ToMe (Token Merging) [2]. Expediting integrates local information to fuse and reconstruct tokens using a clustering algorithm, and ToMe fuses the corresponding tokens based on a bipartite matching algorithm. Previous methods either selected the most representative tokens based on their importance scores or employed clustering algorithms to obtain more representative tokens. The former approach involves calculating importance scores for all tokens, introducing significant computation, and necessitating operations such as padding to ensure uniform dimensions across different batches for batch processing. Although the latter method effectively compresses token sequences, parameter tuning for clustering algorithms poses a challenge, and the iterative calculation approach incurs additional computational costs." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.723, + 0.787, + 0.754 + ], + "angle": 0, + "content": "In this work, we present Token Adapter to merge rows and columns of tokens rather than merge tokens sparsely for the following three reasons:" + }, + { + "type": "text", + "bbox": [ + 0.222, + 0.765, + 0.785, + 0.809 + ], + "angle": 0, + "content": "1. Information Redundancy. In vision tasks, multi-scale features naturally provide more information than single-scale features, resulting in spatial information redundancy in multi-scale image features." + }, + { + "type": "text", + "bbox": [ + 0.222, + 0.81, + 0.785, + 0.84 + ], + "angle": 0, + "content": "2. Image Structure. As image data is derived from sampling real-world optical information, semantic information is contained within consecutive pixel" + }, + { + "type": "list", + "bbox": [ + 0.222, + 0.765, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.63, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Token Adapter" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.787, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.241, + 0.147, + 0.786, + 0.207 + ], + "angle": 0, + "content": "blocks rather than individual pixels. With the same number of missing pixels, removing entire rows or columns of pixels from an image could be less detrimental to the existing semantic information than the pruning partial pixels approach." + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.208, + 0.788, + 0.298 + ], + "angle": 0, + "content": "3. Computational Efficiency. To speed up vision transformers, one simple idea is that using fewer tokens can make them more efficient. Wang et al. [32] has observed that simple images can be accurately predicted with only a few tokens. This observation highlights the inconsistency in the number of tokens required to make accurate predictions for images of varying content complexity." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.305, + 0.789, + 0.504 + ], + "angle": 0, + "content": "Fig. 1 shows the difference between our approach and the previous methods, i.e., the change in token selection strategy. It allows our method to compress token sequences efficiently and handle batch data directly without additional operations. Our experiments show that Token Adapter outperforms Expediting and ToMe on multiple segmentation datasets. To summarize, our main contributions can be outlined as follows: (i) We propose a simple yet powerful module that injects normal tokens into representative tokens while preserving unique tokens associated with dense prediction tasks; (ii) Our token reduction method preserves relative positional information among tokens, providing ViT-like models with substantial acceleration and slight accuracy loss in dense prediction tasks; (iii) Without the need for retraining or fine-tuning, our approach achieves performance comparable to clustering-based methods by employing matching-based techniques, demonstrating even better stability and speed." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.523, + 0.388, + 0.54 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.553, + 0.788, + 0.628 + ], + "angle": 0, + "content": "Efficient Vision Transformer. Some studies have focused on adapting the internal configuration of transformer models to improve their efficiency. These efforts include designing the architecture of the transformer [4,18,30,35], modifying attention mechanisms [13,21,34,38], approximating attention [6,11,12,26], etc." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.63, + 0.788, + 0.795 + ], + "angle": 0, + "content": "Token Reduction. Benefiting from the ability of Transformers to handle variable-length sequences, some recent work has attempted to optimize intermediate representations within transformers. Previous approaches could be categorized by their purpose into two groups: Pruning [16, 33, 36, 37] and Fusion [1, 23, 40, 41]. From a methodological perspective, these approaches can be divided into two categories: recognition-based and non-recognition-based methods. Recognition-based methods categorize all tokens into two groups, vital and minor, performing pruning or fusion operations on them separately based on specific rules. On the other hand, non-recognition-based methods employ clustering algorithms to cluster all tokens, reducing the number of tokens while preserving as much semantic information as possible." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.796, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Our Approach. We decompose token compression into row and column compression, providing a more efficient token reduction method. Additionally, we found that some rows or columns in tokens are unique and cannot be pruned" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.351, + 0.128 + ], + "angle": 0, + "content": "D. Su et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.786, + 0.178 + ], + "angle": 0, + "content": "or merged. In our experiments, retaining these rows or columns yields better results than not retaining them." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.2, + 0.381, + 0.218 + ], + "angle": 0, + "content": "3 Methodology" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.231, + 0.342, + 0.246 + ], + "angle": 0, + "content": "3.1 Overview" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.256, + 0.786, + 0.316 + ], + "angle": 0, + "content": "Designing for dense prediction tasks, our token adapter comprises two simple match-based approaches: the token injector and the token ejector. Token adapter utilize a subset of tokens for the purpose of representing another subset of tokens while maintaining the integrity of the original token set." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.317, + 0.787, + 0.422 + ], + "angle": 0, + "content": "Pipeline. As shown in Fig. 2, two modules of the token adapter are inserted in the middle of consecutive transformer blocks. Given a feature map \\( X \\in \\mathbb{R}^{H \\times W \\times C} \\), we compress rows and columns with reduction ratios \\( r_h \\) and \\( r_w \\) respectively, reducing the number of tokens by \\( n_r \\), where \\( n_r = H \\times W \\times (r_h + r_w - r_h \\times r_w) \\). Here, \\( r_h \\) and \\( r_w \\) are settable parameters. Finally, we rearrange the compressed tokens to rebuild the original token sequence. The details are introduced in the next section." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.444, + 0.387, + 0.46 + ], + "angle": 0, + "content": "3.2 Token Adapter" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.468, + 0.787, + 0.56 + ], + "angle": 0, + "content": "Token Injector. For sequence input \\( X_{s} \\in \\mathbb{R}^{B \\times N \\times C} \\), we first reshape it to \\( X \\in \\mathbb{R}^{B \\times H \\times W \\times C} \\), where \\( N = H \\times W \\) represents the number of tokens. To identify which rows and columns to merge, importance scores are computed for each row and column. In images, areas with sharp pixel value changes usually correspond to inter-object or intra-object boundaries. Moreover, areas with larger pixel values compared to their surroundings are typically more significant." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.561, + 0.787, + 0.695 + ], + "angle": 0, + "content": "In our experiments, we perform two steps of dimension reduction on features to obtain their importance scores. For the input feature \\(X \\in \\mathbb{R}^{B \\times H \\times W \\times C}\\), we first calculate the maximum value of each token vector (you can also calculate its sum value or range value instead), thereby compressing the features into \\(X_{hw} \\in \\mathbb{R}^{B \\times H \\times W}\\). Subsequently, the same procedure is applied to each row and column, yielding \\(S_h \\in \\mathbb{R}^{B \\times H}\\) and \\(S_w \\in \\mathbb{R}^{B \\times W}\\) as the importance scores. Please note that we employed a straightforward method to calculate the importance scores, yet we believe that a better importance assessment algorithm could yield better results." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.696, + 0.787, + 0.757 + ], + "angle": 0, + "content": "Based on these scores, tokens are categorized into three groups: representative tokens, unique tokens, and normal tokens. Normal tokens are injected into representative tokens, while unique tokens remain intact. The compressed tokens are rearranged in their previous positions. A brief flow is delineated below:" + }, + { + "type": "text", + "bbox": [ + 0.222, + 0.765, + 0.663, + 0.779 + ], + "angle": 0, + "content": "1. Calculate the importance scores for each row and column." + }, + { + "type": "text", + "bbox": [ + 0.222, + 0.78, + 0.786, + 0.809 + ], + "angle": 0, + "content": "2. Group all tokens into: representative tokens, unique tokens and normal tokens by importance scores." + }, + { + "type": "text", + "bbox": [ + 0.222, + 0.811, + 0.737, + 0.824 + ], + "angle": 0, + "content": "3. Match the most similar representative token for each normal tokens." + }, + { + "type": "text", + "bbox": [ + 0.222, + 0.825, + 0.698, + 0.84 + ], + "angle": 0, + "content": "4. Merge normal tokens with corresponding representative tokens." + }, + { + "type": "list", + "bbox": [ + 0.222, + 0.765, + 0.786, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.63, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Token Adapter" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "image", + "bbox": [ + 0.276, + 0.144, + 0.731, + 0.289 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.275, + 0.29, + 0.729, + 0.374 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.275, + 0.374, + 0.729, + 0.519 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.529, + 0.788, + 0.6 + ], + "angle": 0, + "content": "Fig. 2: Token Adapter. (a) We use Token Injector module and Token ejector module to wrap \\(\\mathcal{L}_m\\) Transformer blocks. (b) Token Injector mitigates the computational overhead of subsequent Transformer blocks by injecting normal tokens into representative tokens. (c) By leveraging token-to-token relationships, we represent the original sequence with compressed tokens." + }, + { + "type": "text", + "bbox": [ + 0.222, + 0.631, + 0.785, + 0.66 + ], + "angle": 0, + "content": "5. Rearrange new representative tokens and unique tokens in their original order." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.788, + 0.794 + ], + "angle": 0, + "content": "Within an image, consecutive pixel blocks form regions containing semantic information, and these associated semantic regions collectively constitute complete image objects. Merging entire rows or columns preserves the continuity of the original semantic regions while maintaining their relative positions, as depicted in Fig. 3. Our key thought involves representing each object through a refined set of representative and unique tokens. The results presented in Fig. 6 indicate that the retention of partial unique tokens generates better prediction accuracy." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.796, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Token Ejector. After compression, each representative token contains a blend of information from the original representative tokens and the merged normal tokens. Therefore, we exploit the correlation (e.g., the distance matrix" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.351, + 0.128 + ], + "angle": 0, + "content": "D. Su et al." + }, + { + "type": "image", + "bbox": [ + 0.244, + 0.145, + 0.371, + 0.34 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.298, + 0.342, + 0.319, + 0.355 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.373, + 0.145, + 0.501, + 0.34 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.427, + 0.342, + 0.448, + 0.354 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.145, + 0.631, + 0.34 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.556, + 0.342, + 0.577, + 0.355 + ], + "angle": 0, + "content": "(c)" + }, + { + "type": "image", + "bbox": [ + 0.633, + 0.145, + 0.761, + 0.34 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.686, + 0.342, + 0.706, + 0.354 + ], + "angle": 0, + "content": "(d)" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.365, + 0.788, + 0.435 + ], + "angle": 0, + "content": "Fig. 3: Token Selection Visualizations. Results obtained through the removal of specific rows and columns from the ADE20K validation dataset using Segmenter+ViT-L/16 equipped with Token Adapter (\\(r_h = r_w = 0.3\\)). (a) Input Images. (b) Remove specific rows and columns. (c) Input images in which each patch is filled with the mean color of its respective region. (d) Remove specific rows and columns from (c)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.465, + 0.785, + 0.497 + ], + "angle": 0, + "content": "\\(M_{d}\\)) between the reduced sequence and the original sequence to restore the original length. Illustrated in Fig. 2, the primary steps are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.222, + 0.505, + 0.785, + 0.536 + ], + "angle": 0, + "content": "1. Compute the distance matrix \\( M_{d} \\in \\mathbb{R}^{N \\times N_{r}} \\) between \\( X \\in \\mathbb{R}^{N \\times C} \\) and \\( X' \\in \\mathbb{R}^{N_{r} \\times C} \\), with \\( N_{r} = N - n_{r} \\) denoting the length of reduced sequence." + }, + { + "type": "text", + "bbox": [ + 0.221, + 0.537, + 0.756, + 0.551 + ], + "angle": 0, + "content": "2. Perform token matching between tokens in \\(X\\) and \\(X^{\\prime}\\) according to \\(M_{d}\\)." + }, + { + "type": "text", + "bbox": [ + 0.222, + 0.552, + 0.785, + 0.581 + ], + "angle": 0, + "content": "3. Reconstruct the original sequence \\( \\tilde{X} \\) by aggregating tokens in \\( X_r' \\) in accordance with the alignment results." + }, + { + "type": "list", + "bbox": [ + 0.221, + 0.505, + 0.785, + 0.581 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.592, + 0.788, + 0.653 + ], + "angle": 0, + "content": "Note that the distance matrix is computed from the original sequence \\( X_{s} \\) and the reduced sequence \\( X^{\\prime} \\) that has not been refined by the subsequent transformer blocks. By following this approach, the reduced sequence can be efficiently reverted to its initial length." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.677, + 0.376, + 0.695 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.709, + 0.312, + 0.725 + ], + "angle": 0, + "content": "4.1 Setup" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.734, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Dataset and evaluation metrics. We present task-specific metrics, including the mean Intersection-over-Union (mIoU) and average precision (AP), etc. Furthermore, we compare our method with others by analyzing model inference speed and floating-point operations (FLOPs). The datasets utilized for evaluation include ADE20K [43], PASCAL-Context [25], CityScapes [7], COCO [20], and ImageNet-1k [8], covering semantic segmentation, instance segmentation, panotic segmentation, object detection, and image classification tasks." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.63, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Token Adapter" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.269 + ], + "angle": 0, + "content": "Implementation details. To assess the generalization of our method, we equipped two models with Token Adapter and conducted a series of experiments on four well-known datasets for dense prediction tasks. The proposed module is inserted into two classic models off-the-shelf with no re-training or fine-tuning required. All experiments are performed on a single NVIDIA GeForce RTX3090 GPU using Pytorch framework. Further details are presented in the following sections. Our code implementation is based on the open-source code of Expediting and ToMe." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.288, + 0.373, + 0.302 + ], + "angle": 0, + "content": "4.2 Main Results" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.31, + 0.788, + 0.403 + ], + "angle": 0, + "content": "In this section, we delve into a comparative analysis of the proposed approach with other token reduction methods that do not require retraining and are applicable to dense prediction tasks. Since the number of GFLOPs does not provide a direct measure of model inference efficiency, our primary concern lies in accuracy metrics for dense prediction tasks and corresponding metrics for evaluating inference speed." + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.425, + 0.409, + 0.535 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.306, + 0.538, + 0.326, + 0.55 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.41, + 0.426, + 0.594, + 0.535 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.491, + 0.538, + 0.512, + 0.55 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image", + "bbox": [ + 0.597, + 0.426, + 0.78, + 0.535 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.679, + 0.538, + 0.697, + 0.55 + ], + "angle": 0, + "content": "(c)" + }, + { + "type": "image", + "bbox": [ + 0.224, + 0.55, + 0.408, + 0.66 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.306, + 0.663, + 0.326, + 0.675 + ], + "angle": 0, + "content": "(d)" + }, + { + "type": "image", + "bbox": [ + 0.411, + 0.552, + 0.593, + 0.66 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.492, + 0.663, + 0.512, + 0.675 + ], + "angle": 0, + "content": "(e)" + }, + { + "type": "image", + "bbox": [ + 0.597, + 0.552, + 0.779, + 0.66 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.679, + 0.663, + 0.697, + 0.675 + ], + "angle": 0, + "content": "(f)" + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.687, + 0.788, + 0.73 + ], + "angle": 0, + "content": "Fig. 4: Comparisons on Semantic Segmentation Tasks. A comprehensive comparison of our approach with Expediting and ToMe on ADE20K (1-st column), PACAL-Context (2-ed column), and Cityscapes (3-rd column) using Segmenter + ViT-L/16." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.757, + 0.788, + 0.819 + ], + "angle": 0, + "content": "Comparisons on Semantic Segmentation Task. We use the ViT-like segmentation network model Segmenter [28] as our baseline for three benchmarks, including ADE20K, PASCAL-Context, and Cityscapes, reporting semantic segmentation results. We compare our approach with Expediting1 [19] and" + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.825, + 0.671, + 0.841 + ], + "angle": 0, + "content": "\\(^{1}\\) https://github.com/Expedit-LargeScale-Vision-Transformer" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.351, + 0.128 + ], + "angle": 0, + "content": "D. Su et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.788, + 0.187 + ], + "angle": 0, + "content": "Table 1: Comparisons on Semantic Segmentation Tasks. Quantitative comparisons of our method with other token reduction approaches that do not require retraining on semantic segmentation tasks." + }, + { + "type": "table", + "bbox": [ + 0.244, + 0.199, + 0.763, + 0.564 + ], + "angle": 0, + "content": "
DatasetMethodmIoU(%)FPS(img/s)FLOPs(G)
ADE20KSegmenter + ViT-L/1651.8210.68658.98
+ Token Merging47.1217.15321.29
50.9714.02448.7
+ Expediting43.8622.4309.96
51.8514.94481.9
+ Token Adapter (Ours)45.7324.41190.74
52.0215.09273.0
PASCAL-ContextSegmenter + ViT-L/1658.0724.17338.74
+ Token Merging54.2421.84172.38
57.2321.39236.5
+ Expediting48.0350.2889.66
58.1434.50220.76
+ Token Adapter (Ours)47.8063.3465.29
57.8734.53141.5
CityscapesSegmenter + ViT-L/1679.147.42995.55
+ Token Merging71.2313.12448.5
77.818.88651.5
+ Expediting55.3916.68406.69
78.4610.17681.57
+ Token Adapter (Ours)72.0819.38239.65
78.4810.2376.16
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.59, + 0.785, + 0.666 + ], + "angle": 0, + "content": "\\(\\mathrm{ToMe}^2\\) [2], which are clustering-based and matching-based methods for token reduction that do not require retraining, respectively. The results are reported in Tab. 1. The two lines for each method in Tab. 1 represent the best result and the maximum speedup that can be obtained with the same token compression settings, respectively." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.666, + 0.788, + 0.771 + ], + "angle": 0, + "content": "To ensure fairness and comprehensive performance comparisons among different methods, we fine-tuned the hyper-parameters of each method starting from the default settings mentioned in their papers. The results are depicted in Fig. 4. Under maximum acceleration, Token Adapter exhibits superior stability to Expediting \\((+6.11\\% \\mathrm{mIoU}\\) in average) and achieves higher velocity compared to ToMe \\((+93.33\\% \\mathrm{fps}\\) in average). Refer to our supplementary material for detailed parameter settings." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.772, + 0.788, + 0.817 + ], + "angle": 0, + "content": "Comparisons on Instance/Panoptic Segmentation Task. To further explore the generalization capabilities of our method across various Transformer frameworks, we have selected the latest framework, MaskDINO [17]. It utilized" + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.825, + 0.544, + 0.84 + ], + "angle": 0, + "content": "\\(^{2}\\) https://github.com/facebookresearch/ToMe" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.63, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Token Adapter" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.787, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.222 + ], + "angle": 0, + "content": "Swin Transformer as its backbone network and has exhibited remarkable performance in tasks like instance segmentation and panoptic segmentation on COCO. We employed MaskDINO as our baseline model to compare the overall performance of our approach with Expediting on the COCO dataset, as presented in Tab. 2." + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.246, + 0.79, + 0.288 + ], + "angle": 0, + "content": "Table 2: Comparisons on Instance/Panoptic Segmentation Tasks. Inserting both the proposed algorithm and Expediting at different stages of MaskDINO + Swin-L, we compare their comprehensive performance on the COCO dataset." + }, + { + "type": "table", + "bbox": [ + 0.248, + 0.3, + 0.758, + 0.486 + ], + "angle": 0, + "content": "
MethodStagePQ(%)MaskAP(%)FPS(img/s)FLOPs(G)
MaskDINO + Swin-L-58.352.36.11936.86
+ Expediting130.722.676.05920.88
233.123.566.04918.68
357.9851.916.88810.94
+ Token Adapter (ours)458.2952.236.14917.03
155.9149.236.31915.74
256.0349.056.22916.11
357.5751.426.96809.66
458.4252.346.19916.38
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.516, + 0.788, + 0.59 + ], + "angle": 0, + "content": "Due to the window attention mechanism, segmenting all tokens into windows, we reduce the tokens within each window instead of the entire token sequence. To ensure a fair comparison, we gradually increase \\(\\mathcal{L}_b\\) starting from the setting where Expediting balances efficiency and accuracy, to showcase the impact of applying this method at different stages of MaskDINO." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.592, + 0.788, + 0.697 + ], + "angle": 0, + "content": "By observing Tab. 2, it is evident that our approach is more stable than Expediting. This could be attributed to the fact that we directly combine and reconstruct tokens through matching, whereas Expediting relies on clustering, which is sensitive to hyper-parameters for different scenarios. On top of that, with the integration of token adapter into the fourth stage of the backbone network of MaskDINO, we can observe a slight enhancement in performance \\((+0.1\\% \\mathrm{PQ})\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.697, + 0.788, + 0.788 + ], + "angle": 0, + "content": "Comparisons on Image Classification Task. We applied three methods to SWAG[27] for testing on the ImageNet-1k dataset. The results are shown in Tab. 3, where \\(\\dagger\\) indicates that the same fusion method is applied as Token Merging, i.e., layer-by-layer compression, and 32 indicates merging one row per layer. 'Size' indicates the resolution of the feature map. 'r' denotes the number of tokens compressed per layer." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.788, + 0.788, + 0.819 + ], + "angle": 0, + "content": "Visualization of the token injector. We visualize the token injection process in Fig. 5. Initially, we divide all tokens into three groups: representative" + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.825, + 0.546, + 0.84 + ], + "angle": 0, + "content": "3 https://github.com/facebookresearch/SWAG" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.351, + 0.128 + ], + "angle": 0, + "content": "D. Su et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.31, + 0.145, + 0.694, + 0.16 + ], + "angle": 0, + "content": "Table 3: Comparisons on ImageNet1k validation dataset" + }, + { + "type": "table", + "bbox": [ + 0.274, + 0.171, + 0.731, + 0.304 + ], + "angle": 0, + "content": "
MethodSizerAcc(%)FPS(img/s)FLOPs(G)
SWAG + ViT-L/1632 × 32-88.0722.92364.77
+ Token Adapter23 × 23-87.4332.28270.65
+ Expediting23 × 23-87.829.21275.79
+ Token Adapter†-3287.8736.49224.02
+ Token Merging-3287.8038.58218.18
+ Token Adapter†-6485.3466.04130.36
+ Token Merging-6467.5679.99116.40
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.334, + 0.788, + 0.455 + ], + "angle": 0, + "content": "tokens, unique tokens, and normal tokens. Subsequently, we inject the normal tokens into the representative tokens. In the third row of Fig. 5, we use different colored borders to represent the matching results of normal tokens and representative tokens. The last row in the figure represents the results of token injection. By observing this, we can notice that token injection effectively highlights the crucial pixels in the image while preserving the original semantic information. More results and hyper-parameter settings can be found in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.481, + 0.403, + 0.495 + ], + "angle": 0, + "content": "4.3 Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.51, + 0.788, + 0.601 + ], + "angle": 0, + "content": "We performed ablation experiments for each critical choice within our approach on ADE20K semantic segmentation task. The results can be found in Tab. 4. By default, we employed the official weights provided by Segmenter4 as a baseline model (Segmenter + ViT-L/16, mIoU: 51.82%, FPS: 10.68 img/s). The table cells marked in red and purple respectively denote the final choices for the token injector and token ejector." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.601, + 0.788, + 0.707 + ], + "angle": 0, + "content": "Token Reduction Strategy. We initially conducted ablation experiments on the token selection method, and the results are presented in Tab. 4a. In this table, random signifies the stochastic selection, alternating represents the periodic selection, and scores indicates the selection based on computed scores derived from features or attention scores. random and alternating represent the two selection strategies of ToMe. Selecting tokens based on calculated importance scores derived from features presents promising performance." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.708, + 0.788, + 0.813 + ], + "angle": 0, + "content": "Tab. 4b presents the influence of different evaluation metrics employed to calculate the distance among tokens. We utilize this distance matrix to perform token matching. Among these metrics, the cosine similarity distance metric demonstrates superior performance, hence we utilize cosine similarity to measure the distance between tokens. By comparing the pairwise distances between normal tokens and representative tokens, we can match each normal token with its closest representative token." + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.825, + 0.522, + 0.84 + ], + "angle": 0, + "content": "4 https://github.com/rstrudel/segmenter" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.63, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Token Adapter" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "image", + "bbox": [ + 0.217, + 0.143, + 0.788, + 0.365 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.374, + 0.788, + 0.477 + ], + "angle": 0, + "content": "Fig. 5: Token Injection Visualizations. Visualization results from Segmenter + ViT-L/16 applying Token Injector on ADE20K. The first row depicts the input image. The second row illustrates the grouping of all tokens, where gray represents representative tokens, orange represents unique tokens, and teal represents normal tokens. The third row displays the matching results between normal tokens and representative tokens. Different pairs are indicated by boxes of varying colors. The final row showcases the injected results." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.5, + 0.486, + 0.627 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.289, + 0.63, + 0.416, + 0.642 + ], + "angle": 0, + "content": "(a) compressed token" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.5, + 0.787, + 0.627 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.602, + 0.63, + 0.702, + 0.642 + ], + "angle": 0, + "content": "(b) unique token" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.652, + 0.788, + 0.695 + ], + "angle": 0, + "content": "Fig.6: Ablation studies on the ratio of rows and columns of tokens. (a) Ratio ablation of compressed rows or columns \\((r_h\\) and \\(r_w)\\). (b) Ablation of retention ratios for unique token rows or columns." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.788, + 0.842 + ], + "angle": 0, + "content": "To obtain a reduced token sequence, we employ several methods, as shown in Tab. 4d. Keep denotes the preservation of representative tokens and unique tokens without merging normal tokens. Max pool and avg pool refer to pooling-based methods for token fusion, while bilinear represents resizing token sequence through bilinear interpolation. Average indicates averaging tokens within each matched group to merge them. Keep and average do not rely on pre-defined priors to adjust the tokens, hence yielding better results. However, keep leads" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.351, + 0.128 + ], + "angle": 0, + "content": "D. Su et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.145, + 0.788, + 0.215 + ], + "angle": 0, + "content": "Table 4: Token Adapter ablation experiments using Segmenter + ViT/L-16 on ADE20K validation dataset with official checkpoint and no retraining. \\( r_h \\) and \\( r_w \\) are set to 0.3. Segmenter + ViT/L-16 without Token Adapter achieves 51.82% mIoU at 10.68 img/s. We provide the mean Intersection-over-Union (mIoU) and model inference speed (img/s)." + }, + { + "type": "table", + "bbox": [ + 0.236, + 0.226, + 0.777, + 0.327 + ], + "angle": 0, + "content": "
SelectionmIoUFPS
random51.3815.16
alternating51.4314.96
scores (feat.)52.0215.09
scores (attn.)51.1315.04
(a) Token Selection.
(b) Distance Metric.
(c) Feature Choice.
" + }, + { + "type": "table", + "bbox": [ + 0.293, + 0.347, + 0.452, + 0.447 + ], + "angle": 0, + "content": "
MethodmIoUFPS
keep50.6715.19
max pool48.1514.97
avg pool48.1015.10
bilinear50.0815.12
average52.0215.09
" + }, + { + "type": "table_caption", + "bbox": [ + 0.308, + 0.449, + 0.436, + 0.461 + ], + "angle": 0, + "content": "(d) Merge Method." + }, + { + "type": "table", + "bbox": [ + 0.529, + 0.347, + 0.706, + 0.406 + ], + "angle": 0, + "content": "
ModemIoUFPS
max51.9315.13
m-threshold52.0215.09
" + }, + { + "type": "table_caption", + "bbox": [ + 0.551, + 0.407, + 0.684, + 0.42 + ], + "angle": 0, + "content": "(e) Matching Mode." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.505, + 0.785, + 0.534 + ], + "angle": 0, + "content": "to the loss of information from normal tokens, thus we ultimately opt for using average to merge the tokens." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.537, + 0.788, + 0.627 + ], + "angle": 0, + "content": "Restoration Method. Considering dependency on the correlation between the reduced sequence and its original counterpart to restore the sequence length, it is imperative to contemplate the selection of features utilized in calculating this relationship matrix. Within Tab. 4c, the superscript \\( u \\) designates unmerged features, while \\( m \\) signifies merged features. The variable \\( X_{p} \\) represents representative tokens, while \\( X_{r} \\) pertains to reduced tokens." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.628, + 0.788, + 0.749 + ], + "angle": 0, + "content": "To facilitate understanding of the meaning of each symbol, we next summarize the algorithmic procedures of the token adapter, and omit the calculation process. For \\( X \\in \\mathbb{R}^{N \\times C} \\), we first partition tokens into \\( X_{p} \\in \\mathbb{R}^{N_{p} \\times C} \\), \\( X_{n} \\in \\mathbb{R}^{N_{n} \\times C} \\) and \\( X_{u} \\in \\mathbb{R}^{N_{u} \\times C} \\). Here, we haven't merged \\( X_{n} \\) into \\( X_{p} \\) yet, so we can mark it as \\( X_{p}^{u} \\). \\( X_{r} \\) is formed by combining \\( X_{p}^{u} \\) and \\( X_{u} \\), representing the compressed sequence without merging. After merging operations, the new \\( X_{p} \\) is marked as \\( X_{p}^{m} \\) and the new \\( X_{r} \\) is marked as \\( X_{r}^{m} \\). From the results, it shows that recovery using the merged and reduced token sequence \\( X_{r}^{m} \\) works best." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.84 + ], + "angle": 0, + "content": "A comparison result is presented for the two matching modes applied to the token ejector in Tab. 4e. In Sec. 3.2, we introduced the token ejector, which matches the most similar tokens for each token in the original token sequence according to the distance matrix \\( M_{d} \\in \\mathbb{R}^{N \\times N_{r}} \\). Next, the token vectors from each matched group are averaged to reconstruct the token sequence to its original size. Mode max stands for matching the most similar token for each token of" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.63, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Token Adapter" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.237 + ], + "angle": 0, + "content": "the original token sequence, i.e., one-to-one matching. Mode \\( m \\)-threshold, on the other hand, involves filtering based on a set threshold and distance matrix after Mode \\( max \\) matching. The filtered matching results are combined with the results from Mode \\( max \\), followed by executing the corresponding fusion process. The results indicate that the latter achieves more accurate recovery of the token sequence than the former." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.239, + 0.788, + 0.314 + ], + "angle": 0, + "content": "We also conduct ablation experiments on the ratio of compressed rows, columns, and unique tokens, and the results are shown in Fig. 6. Although we determined the final parameter settings for our method based on the ablation experiments, we believe that there still remains a better solution. It will be discussed in the next section." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.34, + 0.511, + 0.356 + ], + "angle": 0, + "content": "5 Discussion and Conclusion" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.374, + 0.351, + 0.388 + ], + "angle": 0, + "content": "5.1 Limitation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.401, + 0.787, + 0.461 + ], + "angle": 0, + "content": "The experimental results mentioned earlier have already demonstrated the effectiveness of our method on dense prediction tasks for hierarchical transformer structures and ViT-like structures. However, there are still many aspects of our work that remain to be explored, which we will analyze one by one next." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.462, + 0.788, + 0.627 + ], + "angle": 0, + "content": "Generalizability. While our method has been validated for effectiveness in multiple segmentation tasks, it has not been validated on other more complex vision tasks. Previous methods essentially involved sparse sampling of token sequences, resulting in compressed token sequences being in sequence form only. The compressed sequence features can only be processed by a sequence processing module such as Transformer. Nevertheless, it remains to be examined whether discrete sparse sequence features are more effective than dense features for image tasks. Motivated by this, we designed the token adapter, a method that performs sparse sampling on the rows and columns of image features. Compared to previous methods, our approach yields more compact sampled features, which may be more effective for image tasks." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.629, + 0.788, + 0.718 + ], + "angle": 0, + "content": "Adaptability. Although our core idea is rather simple, the Token Adapter still takes time to tune the hyper-parameters. The current setting allows Token Adapter to be directly applied to existing models, but it also limits its further development. We chose not to retrain the model equipped with the token adapter to directly evaluate the effectiveness of our idea. In the future, we will continue to explore more efficient neural network acceleration schemes." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Our core contribution is to compress and reconstruct rows and columns of image features in a way that conforms to the structure of the image. When applied to vision models, Token Adapter only compresses and reconstructs image features and thus can be applied to most vision models. For large language models, the concept of a Token Adapter can still be effectively transferred. By decomposing text embeddings based on the contextual information, and then leveraging the connections between the decomposed parts to compress them, it should be able to improve the efficiency of large language models." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.351, + 0.128 + ], + "angle": 0, + "content": "D. Su et al." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.147, + 0.355, + 0.162 + ], + "angle": 0, + "content": "5.2 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.175, + 0.788, + 0.386 + ], + "angle": 0, + "content": "In this work, we propose Token Adapter to expedite vision transformers used in dense prediction tasks. By injecting normal tokens into representative tokens, we enhance the efficiency of vision transformers. Notably, our method can be applied in off-the-shelf models directly without additional retraining or finetuning. Tab. 2 reveals that when applied to the early stages of MaskDINO, the clustering-based Expediting method results in a significant decline in accuracy, and the matching-based Token Adapter method demonstrates superior stability. Clustering algorithms typically require fine-tuning of multiple hyper-parameters, which remain fixed once adjusted. Our approach is expected to leverage mechanisms such as cross-attention to transform fixed hyper-parameters into dynamically learnable parameters, thereby enhancing the method's generalization capability. Theoretically, the design of the compression strategy and reconstruction strategy in our algorithm allows the token adapter to be inserted in the middle of any module designed for image tasks." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.388, + 0.788, + 0.569 + ], + "angle": 0, + "content": "In principle, Token Adapter bears resemblance to an unparameterized form of cross-attention mechanism, both blending information from one set of tokens into another set. Our thought lies in how we employ tokens to represent semantic entities within the image. Each entity in the image can be divided into distinct local semantic regions. By assigning several representative tokens to each local semantic region and injecting the remaining normal tokens into their corresponding tokens, we achieve a more refined fusion. After the fusion, the representative tokens representing different semantic regions will engage in information exchange within transformers, greatly enhancing computational efficiency and promoting diversity among tokens. We hope our research can offer a fresh perspective on image features and inspire the design of more efficient methods for extracting image features." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.595, + 0.325, + 0.611 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.223, + 0.629, + 0.787, + 0.657 + ], + "angle": 0, + "content": "1. Bian, Z., Wang, Z., Han, W., Wang, K.: Multi-scale and token merging: Make your vit more efficient. arXiv preprint arXiv:2306.04897 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.223, + 0.658, + 0.787, + 0.699 + ], + "angle": 0, + "content": "2. Bolya, D., Fu, C.Y., Dai, X., Zhang, P., Feichtenhofer, C., Hoffman, J.: Token merging: Your vit but faster. In: The Eleventh International Conference on Learning Representations (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.223, + 0.7, + 0.787, + 0.728 + ], + "angle": 0, + "content": "3. Bonnaerens, M., Dambre, J.: Learned thresholds token merging and pruning for vision transformers. arXiv preprint arXiv:2307.10780 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.223, + 0.729, + 0.787, + 0.77 + ], + "angle": 0, + "content": "4. Carion, N., Massa, F., Synnaeve, G., Usunier, N., Kirillov, A., Zagoruyko, S.: End-to-end object detection with transformers. In: European conference on computer vision. pp. 213-229. Springer (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.223, + 0.771, + 0.787, + 0.799 + ], + "angle": 0, + "content": "5. Chen, Z., Duan, Y., Wang, W., He, J., Lu, T., Dai, J., Qiao, Y.: Vision transformer adapter for dense predictions. arXiv preprint arXiv:2205.08534 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.223, + 0.799, + 0.787, + 0.842 + ], + "angle": 0, + "content": "6. Choromanski, K., Likhosherstov, V., Dohan, D., Song, X., Gane, A., Sarlos, T., Hawkins, P., Davis, J., Mohiuddin, A., Kaiser, L., et al.: Rethinking attention with performers. arXiv preprint arXiv:2009.14794 (2020)" + }, + { + "type": "list", + "bbox": [ + 0.223, + 0.629, + 0.787, + 0.842 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.63, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Token Adapter" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.147, + 0.786, + 0.203 + ], + "angle": 0, + "content": "7. Cordts, M., Omran, M., Ramos, S., Rehfeld, T., Enzweiler, M., Benenson, R., Franke, U., Roth, S., Schiele, B.: The cityscapes dataset for semantic urban scene understanding. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 3213-3223 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.203, + 0.787, + 0.243 + ], + "angle": 0, + "content": "8. Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: Imagenet: A large-scale hierarchical image database. In: 2009 IEEE conference on computer vision and pattern recognition. pp. 248-255. IEEE (2009)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.243, + 0.787, + 0.298 + ], + "angle": 0, + "content": "9. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.298, + 0.787, + 0.352 + ], + "angle": 0, + "content": "0. Fayyaz, M., Koohpayegani, S.A., Jafari, F.R., Sengupta, S., Joze, H.R.V., Sommerlade, E., Pirsiavash, H., Gall, J.: Adaptive token sampling for efficient vision transformers. In: European Conference on Computer Vision. pp. 396-414. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.352, + 0.787, + 0.379 + ], + "angle": 0, + "content": "1. Geng, Z., Guo, M.H., Chen, H., Li, X., Wei, K., Lin, Z.: Is attention better than matrix decomposition? arXiv preprint arXiv:2109.04553 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.379, + 0.787, + 0.42 + ], + "angle": 0, + "content": "2. Guo, M.H., Liu, Z.N., Mu, T.J., Hu, S.M.: Beyond self-attention: External attention using two linear layers for visual tasks. IEEE Transactions on Pattern Analysis and Machine Intelligence 45(5), 5436-5447 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.42, + 0.787, + 0.46 + ], + "angle": 0, + "content": "3. Hassani, A., Walton, S., Li, J., Li, S., Shi, H.: Neighborhood attention transformer. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6185-6194 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.46, + 0.787, + 0.5 + ], + "angle": 0, + "content": "4. He, K., Chen, X., Xie, S., Li, Y., Dólar, P., Girshick, R.: Masked autoencoders are scalable vision learners. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 16000-16009 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.501, + 0.787, + 0.542 + ], + "angle": 0, + "content": "5. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., et al.: Segment anything. arXiv preprint arXiv:2304.02643 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.542, + 0.787, + 0.582 + ], + "angle": 0, + "content": "6. Kong, Z., Dong, P., Ma, X., Meng, X., Sun, M., Niu, W., Shen, X., Yuan, G., Ren, B., Qin, M., et al.: Spvit: Enabling faster vision transformers via soft token pruning. arXiv preprint arXiv:2112.13890 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.582, + 0.787, + 0.636 + ], + "angle": 0, + "content": "7. Li, F., Zhang, H., Xu, H., Liu, S., Zhang, L., Ni, L.M., Shum, H.Y.: Mask dino: Towards a unified transformer-based framework for object detection and segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3041-3050 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.636, + 0.787, + 0.677 + ], + "angle": 0, + "content": "8. Li, Y., Yuan, G., Wen, Y., Hu, J., Evangelidis, G., Tulyakov, S., Wang, Y., Ren, J.: Efficientformer: Vision transformers at mobilenet speed. Advances in Neural Information Processing Systems 35, 12934-12949 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.677, + 0.787, + 0.731 + ], + "angle": 0, + "content": "9. Liang, W., Yuan, Y., Ding, H., Luo, X., Lin, W., Jia, D., Zhang, Z., Zhang, C., Hu, H.: Expediting large-scale vision transformer for dense prediction without fine-tuning. Advances in Neural Information Processing Systems 35, 35462-35477 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.731, + 0.787, + 0.786 + ], + "angle": 0, + "content": "20. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollar, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13. pp. 740-755. Springer (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.786, + 0.787, + 0.84 + ], + "angle": 0, + "content": "21. Liu, Z., Hu, H., Lin, Y., Yao, Z., Xie, Z., Wei, Y., Ning, J., Cao, Y., Zhang, Z., Dong, L., et al.: Swin transformer v2: Scaling up capacity and resolution. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 12009-12019 (2022)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.147, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.351, + 0.128 + ], + "angle": 0, + "content": "D. Su et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.786, + 0.203 + ], + "angle": 0, + "content": "22. Liu, Z., Lin, Y., Cao, Y., Hu, H., Wei, Y., Zhang, Z., Lin, S., Guo, B.: Swin transformer: Hierarchical vision transformer using shifted windows. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 10012-10022 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.203, + 0.787, + 0.257 + ], + "angle": 0, + "content": "23. Long, S., Zhao, Z., Pi, J., Wang, S., Wang, J.: Beyond attentive tokens: Incorporating token importance and diversity for efficient vision transformers. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10334-10343 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.257, + 0.786, + 0.298 + ], + "angle": 0, + "content": "24. Lu, J., Yao, J., Zhang, J., Zhu, X., Xu, H., Gao, W., Xu, C., Xiang, T., Zhang, L.: Soft: Softmax-free transformer with linear complexity. Advances in Neural Information Processing Systems 34, 21297-21309 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.298, + 0.786, + 0.354 + ], + "angle": 0, + "content": "25. Mottaghi, R., Chen, X., Liu, X., Cho, N.G., Lee, S.W., Fidler, S., Urtasun, R., Yuille, A.: The role of context for object detection and semantic segmentation in the wild. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 891-898 (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.354, + 0.786, + 0.38 + ], + "angle": 0, + "content": "26. Peng, H., Pappas, N., Yogatama, D., Schwartz, R., Smith, N.A., Kong, L.: Random feature attention. arXiv preprint arXiv:2103.02143 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.379, + 0.786, + 0.435 + ], + "angle": 0, + "content": "27. Singh, M., Gustafson, L., Adcock, A., de Freitas Reis, V., Gedik, B., Kosaraju, R.P., Mahajan, D., Girshick, R., Dollár, P., Van Der Maaten, L.: Revisiting weakly supervised pre-training of visual perception models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 804-814 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.434, + 0.786, + 0.474 + ], + "angle": 0, + "content": "28. Strudel, R., Garcia, R., Laptev, I., Schmid, C.: Segmenter: Transformer for semantic segmentation. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 7262-7272 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.474, + 0.786, + 0.514 + ], + "angle": 0, + "content": "29. Tian, K., Jiang, Y., Dao, Q., Lin, C., Wang, L., Yuan, Z.: Designing bert for convolutional networks: Sparse and hierarchical masked modeling. arXiv preprint arXiv:2301.03580 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.514, + 0.786, + 0.569 + ], + "angle": 0, + "content": "30. Tolstikhin, I.O., Houlsby, N., Kolesnikov, A., Beyer, L., Zhai, X., Unterthiner, T., Yung, J., Steiner, A., Keysers, D., Uszkoreit, J., et al.: Mlp-mixer: An allmlp architecture for vision. Advances in neural information processing systems 34, 24261-24272 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.569, + 0.786, + 0.61 + ], + "angle": 0, + "content": "31. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. Advances in neural information processing systems 30 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.61, + 0.786, + 0.651 + ], + "angle": 0, + "content": "32. Wang, Y., Huang, R., Song, S., Huang, Z., Huang, G.: Not all images are worth 16x16 words: Dynamic transformers for efficient image recognition. Advances in Neural Information Processing Systems 34, 11960-11973 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.651, + 0.786, + 0.692 + ], + "angle": 0, + "content": "33. Wang, Z., Luo, H., Wang, P., Ding, F., Wang, F., Li, H.: Vtc-lfc: Vision transformer compression with low-frequency components. Advances in Neural Information Processing Systems 35, 13974-13988 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.692, + 0.786, + 0.732 + ], + "angle": 0, + "content": "34. Yang, J., Li, C., Zhang, P., Dai, X., Xiao, B., Yuan, L., Gao, J.: Focal self-attention for local-global interactions in vision transformers. arXiv preprint arXiv:2107.00641 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.732, + 0.786, + 0.773 + ], + "angle": 0, + "content": "35. Yang, R., Ma, H., Wu, J., Tang, Y., Xiao, X., Zheng, M., Li, X.: Scalablevit: Rethinking the context-oriented generalization of vision transformer. In: European Conference on Computer Vision. pp. 480-496. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.773, + 0.786, + 0.8 + ], + "angle": 0, + "content": "36. Ye, D., Lin, Y., Huang, Y., Sun, M.: Tr-bert: Dynamic token reduction for accelerating bert inference. arXiv preprint arXiv:2105.11618 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.8, + 0.786, + 0.841 + ], + "angle": 0, + "content": "37. Yin, H., Vahdat, A., Alvarez, J.M., Mallya, A., Kautz, J., Molchanov, P.: A-vit: Adaptive tokens for efficient vision transformer. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10809-10818 (2022)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.63, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Token Adapter" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.189 + ], + "angle": 0, + "content": "38. Yu, Q., Xia, Y., Bai, Y., Lu, Y., Yuille, A.L., Shen, W.: Glance-and-gaze vision transformer. Advances in Neural Information Processing Systems 34, 12992-13003 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.19, + 0.788, + 0.232 + ], + "angle": 0, + "content": "39. Yu, W., Luo, M., Zhou, P., Si, C., Zhou, Y., Wang, X., Feng, J., Yan, S.: Metaformer is actually what you need for vision. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10819-10829 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.232, + 0.788, + 0.286 + ], + "angle": 0, + "content": "40. Zeng, W., Jin, S., Liu, W., Qian, C., Luo, P., Ouyang, W., Wang, X.: Not all tokens are equal: Human-centric visual analysis via token clustering transformer. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11101-11111 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.287, + 0.788, + 0.328 + ], + "angle": 0, + "content": "41. Zeng, Z., Hawkins, C., Hong, M., Zhang, A., Pappas, N., Singh, V., Zheng, S.: Vcc: Scaling transformers to 128k tokens or more by prioritizing important tokens. arXiv preprint arXiv:2305.04241 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.328, + 0.788, + 0.37 + ], + "angle": 0, + "content": "42. Zheng, M., Gao, P., Zhang, R., Li, K., Wang, X., Li, H., Dong, H.: End-to-end object detection with adaptive clustering transformer. arXiv preprint arXiv:2011.09315 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.371, + 0.788, + 0.411 + ], + "angle": 0, + "content": "43. Zhou, B., Zhao, H., Puig, X., Xiao, T., Fidler, S., Barriuso, A., Torralba, A.: Semantic understanding of scenes through the ade20k dataset. International Journal of Computer Vision 127, 302-321 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.411, + 0.788, + 0.467 + ], + "angle": 0, + "content": "44. Ziwen, C., Patnaik, K., Zhai, S., Wan, A., Ren, Z., Schwing, A.G., Colburn, A., Fuxin, L.: Autofocusformer: Image segmentation off the grid. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18227-18236 (2023)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.467 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/Removing Rows and Columns of Tokens in Vision Transformer enables Faster Dense Prediction without Retraining/8e177374-55c7-4790-9a57-d4f7dcccc52f_origin.pdf b/2024/Removing Rows and Columns of Tokens in Vision Transformer enables Faster Dense Prediction without Retraining/8e177374-55c7-4790-9a57-d4f7dcccc52f_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..9b2720eab5b4b2bd26b7a9b0e58b6c581ca9936b --- /dev/null +++ b/2024/Removing Rows and Columns of Tokens in Vision Transformer enables Faster Dense Prediction without Retraining/8e177374-55c7-4790-9a57-d4f7dcccc52f_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f58dafc28c6aca7070a1d0330a578c4fa61138f19625c93e4eeb45d76531e525 +size 3211491 diff --git a/2024/Removing Rows and Columns of Tokens in Vision Transformer enables Faster Dense Prediction without Retraining/full.md b/2024/Removing Rows and Columns of Tokens in Vision Transformer enables Faster Dense Prediction without Retraining/full.md new file mode 100644 index 0000000000000000000000000000000000000000..d545023c824d22af6912fc75b4ea80042a1cbfc2 --- /dev/null +++ b/2024/Removing Rows and Columns of Tokens in Vision Transformer enables Faster Dense Prediction without Retraining/full.md @@ -0,0 +1,298 @@ +# Removing Rows and Columns of Tokens in Vision Transformer enables Faster Dense Prediction without Retraining + +Diwei $\mathrm{Su}^{1}$ Cheng Fei, and Jianxu Luo + +East China University of Science and Technology, Shanghai 200237, China dvsu@mail.ecust.edu.cn, cfei_jarvis@163.com, jxluo@ecust.edu.cn + +Abstract. In recent years, vision transformers based on self-attention mechanisms have demonstrated remarkable abilities in various tasks such as natural language processing, computer vision (CV), and multimodal applications. However, due to the high computational costs and the structural nature of images, the application of transformers to CV tasks faces challenges, particularly when handling ultra-high-resolution images. Recently, several token reduction methods have been proposed to improve the computational efficiency of transformers by reducing the number of tokens without the need for retraining. These methods primarily involve fusion based on matching or clustering. The former exhibits faster speed but suffers more accuracy loss compared to the latter. In this work, we propose a simple matching-based fusion method called Token Adapter, which achieves comparable accuracy to the clustering-based fusion method with faster speed and demonstrates higher potential in terms of robustness. Our method was applied to Segmenter, MaskDINO and SWAG, exhibiting promising performance on four tasks, including semantic segmentation, instance segmentation, panoptic segmentation, and image classification. Specifically, our method can be applied to Segmenter on ADE20k, providing $41\%$ frames per second (FPS) acceleration while maintaining full performance without retraining or fine-tuning off-the-shelf weights. Our code will be released at https://github.com/MilknoCandy-Token-Adapter. + +Keywords: Vision transformer $\cdot$ Token reduction $\cdot$ Dense prediction $\cdot$ Efficient transformer + +# 1 Introduction + +Over the past few years, the computer vision field has been significantly advanced by the emergence of the Transformer architecture [31]. Vision Transformers [9] with global attention mechanism achieve state-of-the-art results on a wide range of vision tasks. To obtain better performance, researchers have scaled up the number of parameters of vision transformers to billions [15] through techniques such as self-supervised pre-training [14, 29] methods and adding visual prior knowledge [5, 22], etc. However, scaling up transformers brings performance improvements as well as huge computational costs due to the quadratic complexity + +![](images/11284078890d1f4e17bd04b746a116476f5b9fa83092c7665867cb0034b02708.jpg) + +![](images/aff19d31763915ef52c1adbc17b1b77c1703f123491aca922a2010c494f27443.jpg) +(a) + +![](images/0e838eb5eca54fbb0f9687fd066b4d3eb6975387aa82cfb02ea75db3ccc68493.jpg) + +![](images/934332e7e73f9229188040a7940112769bd60c3dc3f68ed7bccb584f16870919.jpg) +Fig. 1: Visualizations of random pruning results from different methods on ADE20K. We randomly prune the same number of tokens using different methods. (a) Input Images. (b) Randomly remove tokens (c) Randomly remove specific rows and columns. + +![](images/266a7240c18e3b8b09fd39f5288d1e8162dcd9a5ad549a91a78f7d5292a7d453.jpg) + +![](images/ca8ab54901f1f71dd1e6fb61e82be9bf7ccdcb4e954c8eb7468681d2d787d974.jpg) +(b) + +![](images/c2fab848ce3bf0583ce6b711a03d9aca0ecf50bc81d131ad4ede2fa888893143.jpg) + +![](images/363888ef80f14b88a682fc169a7183b73f3897f46f54ebd503ee60e876dba8bd.jpg) + +![](images/c34d7a81f395ec1c70d792e18796b20faf073be08fbae2afa28b156d2c644b78.jpg) + +![](images/686ac65b26deb62e8d9b0675da316d45bef98c4b20fffb5f125ab43a8c53fb88.jpg) +(c) + +![](images/a559b89f0ba42406227626182b87aacdfa1b529fc939e7e3f41ac8e797d81eca.jpg) + +![](images/85a5c927fd3dfbc2948526b57536194e3b96824df5546d0d8aa1e16f97b1e3fa.jpg) + +of self-attention mechanism. Vision transformers are extremely inefficient when processing high-resolution images. Designing efficient vision transformers has become a prominent research hotspot. + +Some researchers have reduced the computational complexity to linear by optimizing the computation of self-attention mechanism [24, 30, 39, 42], while others are focusing on reducing the number of tokens to increase the inference efficiency of vision transformers [3, 10, 32, 44]. Despite significantly improving the efficiency of the vision transformer, most of these methods necessitate the retraining or fine-tuning of the optimized Transformer, demanding substantial computational resources. + +There are also methods that do not require fine-tuning or retraining, such as Expediting [19] and ToMe (Token Merging) [2]. Expediting integrates local information to fuse and reconstruct tokens using a clustering algorithm, and ToMe fuses the corresponding tokens based on a bipartite matching algorithm. Previous methods either selected the most representative tokens based on their importance scores or employed clustering algorithms to obtain more representative tokens. The former approach involves calculating importance scores for all tokens, introducing significant computation, and necessitating operations such as padding to ensure uniform dimensions across different batches for batch processing. Although the latter method effectively compresses token sequences, parameter tuning for clustering algorithms poses a challenge, and the iterative calculation approach incurs additional computational costs. + +In this work, we present Token Adapter to merge rows and columns of tokens rather than merge tokens sparsely for the following three reasons: + +1. Information Redundancy. In vision tasks, multi-scale features naturally provide more information than single-scale features, resulting in spatial information redundancy in multi-scale image features. +2. Image Structure. As image data is derived from sampling real-world optical information, semantic information is contained within consecutive pixel + +blocks rather than individual pixels. With the same number of missing pixels, removing entire rows or columns of pixels from an image could be less detrimental to the existing semantic information than the pruning partial pixels approach. + +3. Computational Efficiency. To speed up vision transformers, one simple idea is that using fewer tokens can make them more efficient. Wang et al. [32] has observed that simple images can be accurately predicted with only a few tokens. This observation highlights the inconsistency in the number of tokens required to make accurate predictions for images of varying content complexity. + +Fig. 1 shows the difference between our approach and the previous methods, i.e., the change in token selection strategy. It allows our method to compress token sequences efficiently and handle batch data directly without additional operations. Our experiments show that Token Adapter outperforms Expediting and ToMe on multiple segmentation datasets. To summarize, our main contributions can be outlined as follows: (i) We propose a simple yet powerful module that injects normal tokens into representative tokens while preserving unique tokens associated with dense prediction tasks; (ii) Our token reduction method preserves relative positional information among tokens, providing ViT-like models with substantial acceleration and slight accuracy loss in dense prediction tasks; (iii) Without the need for retraining or fine-tuning, our approach achieves performance comparable to clustering-based methods by employing matching-based techniques, demonstrating even better stability and speed. + +# 2 Related Work + +Efficient Vision Transformer. Some studies have focused on adapting the internal configuration of transformer models to improve their efficiency. These efforts include designing the architecture of the transformer [4,18,30,35], modifying attention mechanisms [13,21,34,38], approximating attention [6,11,12,26], etc. + +Token Reduction. Benefiting from the ability of Transformers to handle variable-length sequences, some recent work has attempted to optimize intermediate representations within transformers. Previous approaches could be categorized by their purpose into two groups: Pruning [16, 33, 36, 37] and Fusion [1, 23, 40, 41]. From a methodological perspective, these approaches can be divided into two categories: recognition-based and non-recognition-based methods. Recognition-based methods categorize all tokens into two groups, vital and minor, performing pruning or fusion operations on them separately based on specific rules. On the other hand, non-recognition-based methods employ clustering algorithms to cluster all tokens, reducing the number of tokens while preserving as much semantic information as possible. + +Our Approach. We decompose token compression into row and column compression, providing a more efficient token reduction method. Additionally, we found that some rows or columns in tokens are unique and cannot be pruned + +or merged. In our experiments, retaining these rows or columns yields better results than not retaining them. + +# 3 Methodology + +# 3.1 Overview + +Designing for dense prediction tasks, our token adapter comprises two simple match-based approaches: the token injector and the token ejector. Token adapter utilize a subset of tokens for the purpose of representing another subset of tokens while maintaining the integrity of the original token set. + +Pipeline. As shown in Fig. 2, two modules of the token adapter are inserted in the middle of consecutive transformer blocks. Given a feature map $X \in \mathbb{R}^{H \times W \times C}$ , we compress rows and columns with reduction ratios $r_h$ and $r_w$ respectively, reducing the number of tokens by $n_r$ , where $n_r = H \times W \times (r_h + r_w - r_h \times r_w)$ . Here, $r_h$ and $r_w$ are settable parameters. Finally, we rearrange the compressed tokens to rebuild the original token sequence. The details are introduced in the next section. + +# 3.2 Token Adapter + +Token Injector. For sequence input $X_{s} \in \mathbb{R}^{B \times N \times C}$ , we first reshape it to $X \in \mathbb{R}^{B \times H \times W \times C}$ , where $N = H \times W$ represents the number of tokens. To identify which rows and columns to merge, importance scores are computed for each row and column. In images, areas with sharp pixel value changes usually correspond to inter-object or intra-object boundaries. Moreover, areas with larger pixel values compared to their surroundings are typically more significant. + +In our experiments, we perform two steps of dimension reduction on features to obtain their importance scores. For the input feature $X \in \mathbb{R}^{B \times H \times W \times C}$ , we first calculate the maximum value of each token vector (you can also calculate its sum value or range value instead), thereby compressing the features into $X_{hw} \in \mathbb{R}^{B \times H \times W}$ . Subsequently, the same procedure is applied to each row and column, yielding $S_h \in \mathbb{R}^{B \times H}$ and $S_w \in \mathbb{R}^{B \times W}$ as the importance scores. Please note that we employed a straightforward method to calculate the importance scores, yet we believe that a better importance assessment algorithm could yield better results. + +Based on these scores, tokens are categorized into three groups: representative tokens, unique tokens, and normal tokens. Normal tokens are injected into representative tokens, while unique tokens remain intact. The compressed tokens are rearranged in their previous positions. A brief flow is delineated below: + +1. Calculate the importance scores for each row and column. +2. Group all tokens into: representative tokens, unique tokens and normal tokens by importance scores. +3. Match the most similar representative token for each normal tokens. +4. Merge normal tokens with corresponding representative tokens. + +![](images/e855a7467c28b6fa09f0abbe5aedc80e37aab77478316a324c3b442234517fb2.jpg) + +![](images/4119ace7158c446cc6f91d0c27bfd84d4d7fe52b11155638be2856f7e258471a.jpg) + +![](images/d32acbea98682afb06ebaabd3abbe575d3adbfec9a1e3eef412538759f03c285.jpg) +Fig. 2: Token Adapter. (a) We use Token Injector module and Token ejector module to wrap $\mathcal{L}_m$ Transformer blocks. (b) Token Injector mitigates the computational overhead of subsequent Transformer blocks by injecting normal tokens into representative tokens. (c) By leveraging token-to-token relationships, we represent the original sequence with compressed tokens. + +5. Rearrange new representative tokens and unique tokens in their original order. + +Within an image, consecutive pixel blocks form regions containing semantic information, and these associated semantic regions collectively constitute complete image objects. Merging entire rows or columns preserves the continuity of the original semantic regions while maintaining their relative positions, as depicted in Fig. 3. Our key thought involves representing each object through a refined set of representative and unique tokens. The results presented in Fig. 6 indicate that the retention of partial unique tokens generates better prediction accuracy. + +Token Ejector. After compression, each representative token contains a blend of information from the original representative tokens and the merged normal tokens. Therefore, we exploit the correlation (e.g., the distance matrix + +![](images/8ba4edbdf1fddf24fab397428548bebfd6c5b3f10c011b8b6b52b1d59a34d7a8.jpg) +(a) + +![](images/6c703140f6caf47edbaf3769f69371c4e1554b1865a177e99be2dce5093ea2db.jpg) +(b) + +![](images/409d5efa72c217c265f049798c9bac729e27cc5e644b06991b14ff099326e6e2.jpg) +(c) +Fig. 3: Token Selection Visualizations. Results obtained through the removal of specific rows and columns from the ADE20K validation dataset using Segmenter+ViT-L/16 equipped with Token Adapter ( $r_h = r_w = 0.3$ ). (a) Input Images. (b) Remove specific rows and columns. (c) Input images in which each patch is filled with the mean color of its respective region. (d) Remove specific rows and columns from (c). + +![](images/5522f4c3a7fbe32357c767eb7b6ada4c7e23ffb61a8ada1d60afc16a5b2e753c.jpg) +(d) + +$M_{d}$ ) between the reduced sequence and the original sequence to restore the original length. Illustrated in Fig. 2, the primary steps are as follows: + +1. Compute the distance matrix $M_{d} \in \mathbb{R}^{N \times N_{r}}$ between $X \in \mathbb{R}^{N \times C}$ and $X' \in \mathbb{R}^{N_{r} \times C}$ , with $N_{r} = N - n_{r}$ denoting the length of reduced sequence. +2. Perform token matching between tokens in $X$ and $X^{\prime}$ according to $M_{d}$ . +3. Reconstruct the original sequence $\tilde{X}$ by aggregating tokens in $X_r'$ in accordance with the alignment results. + +Note that the distance matrix is computed from the original sequence $X_{s}$ and the reduced sequence $X^{\prime}$ that has not been refined by the subsequent transformer blocks. By following this approach, the reduced sequence can be efficiently reverted to its initial length. + +# 4 Experiments + +# 4.1 Setup + +Dataset and evaluation metrics. We present task-specific metrics, including the mean Intersection-over-Union (mIoU) and average precision (AP), etc. Furthermore, we compare our method with others by analyzing model inference speed and floating-point operations (FLOPs). The datasets utilized for evaluation include ADE20K [43], PASCAL-Context [25], CityScapes [7], COCO [20], and ImageNet-1k [8], covering semantic segmentation, instance segmentation, panotic segmentation, object detection, and image classification tasks. + +Implementation details. To assess the generalization of our method, we equipped two models with Token Adapter and conducted a series of experiments on four well-known datasets for dense prediction tasks. The proposed module is inserted into two classic models off-the-shelf with no re-training or fine-tuning required. All experiments are performed on a single NVIDIA GeForce RTX3090 GPU using Pytorch framework. Further details are presented in the following sections. Our code implementation is based on the open-source code of Expediting and ToMe. + +# 4.2 Main Results + +In this section, we delve into a comparative analysis of the proposed approach with other token reduction methods that do not require retraining and are applicable to dense prediction tasks. Since the number of GFLOPs does not provide a direct measure of model inference efficiency, our primary concern lies in accuracy metrics for dense prediction tasks and corresponding metrics for evaluating inference speed. + +![](images/d26110e380ee614d43021ab9cb1491e427a2bd2ec4287dd696f206de5ccd4e97.jpg) + +![](images/4f626b9755485388a28079aa8f711693e86678c1d3af5a3222513f36d5566951.jpg) + +![](images/5e39608380b77b7f6b34d0ed8520db5ad1bf13056eb6ee51d1af53eb9596c6d8.jpg) + +![](images/7d4f30a73a80dfa0133d4f816af59c8f22d6c376b380d198c82fb8a768510a19.jpg) +(a) +(d) +Fig. 4: Comparisons on Semantic Segmentation Tasks. A comprehensive comparison of our approach with Expediting and ToMe on ADE20K (1-st column), PACAL-Context (2-ed column), and Cityscapes (3-rd column) using Segmenter + ViT-L/16. + +![](images/3b8bda187a68c6030b7991bd2325e2a906e952ed5659e422dbb389400d7879f8.jpg) +(b) +(e) + +![](images/ddb848b8b2fa57243098d9e6e05b1490a466a0112fca2e734f47e78f9ba5fb95.jpg) +(c) +(f) + +Comparisons on Semantic Segmentation Task. We use the ViT-like segmentation network model Segmenter [28] as our baseline for three benchmarks, including ADE20K, PASCAL-Context, and Cityscapes, reporting semantic segmentation results. We compare our approach with Expediting1 [19] and + +Table 1: Comparisons on Semantic Segmentation Tasks. Quantitative comparisons of our method with other token reduction approaches that do not require retraining on semantic segmentation tasks. + +
DatasetMethodmIoU(%)FPS(img/s)FLOPs(G)
ADE20KSegmenter + ViT-L/1651.8210.68658.98
+ Token Merging47.1217.15321.29
50.9714.02448.7
+ Expediting43.8622.4309.96
51.8514.94481.9
+ Token Adapter (Ours)45.7324.41190.74
52.0215.09273.0
PASCAL-ContextSegmenter + ViT-L/1658.0724.17338.74
+ Token Merging54.2421.84172.38
57.2321.39236.5
+ Expediting48.0350.2889.66
58.1434.50220.76
+ Token Adapter (Ours)47.8063.3465.29
57.8734.53141.5
CityscapesSegmenter + ViT-L/1679.147.42995.55
+ Token Merging71.2313.12448.5
77.818.88651.5
+ Expediting55.3916.68406.69
78.4610.17681.57
+ Token Adapter (Ours)72.0819.38239.65
78.4810.2376.16
+ +$\mathrm{ToMe}^2$ [2], which are clustering-based and matching-based methods for token reduction that do not require retraining, respectively. The results are reported in Tab. 1. The two lines for each method in Tab. 1 represent the best result and the maximum speedup that can be obtained with the same token compression settings, respectively. + +To ensure fairness and comprehensive performance comparisons among different methods, we fine-tuned the hyper-parameters of each method starting from the default settings mentioned in their papers. The results are depicted in Fig. 4. Under maximum acceleration, Token Adapter exhibits superior stability to Expediting $(+6.11\% \mathrm{mIoU}$ in average) and achieves higher velocity compared to ToMe $(+93.33\% \mathrm{fps}$ in average). Refer to our supplementary material for detailed parameter settings. + +Comparisons on Instance/Panoptic Segmentation Task. To further explore the generalization capabilities of our method across various Transformer frameworks, we have selected the latest framework, MaskDINO [17]. It utilized + +Swin Transformer as its backbone network and has exhibited remarkable performance in tasks like instance segmentation and panoptic segmentation on COCO. We employed MaskDINO as our baseline model to compare the overall performance of our approach with Expediting on the COCO dataset, as presented in Tab. 2. + +Table 2: Comparisons on Instance/Panoptic Segmentation Tasks. Inserting both the proposed algorithm and Expediting at different stages of MaskDINO + Swin-L, we compare their comprehensive performance on the COCO dataset. + +
MethodStagePQ(%)MaskAP(%)FPS(img/s)FLOPs(G)
MaskDINO + Swin-L-58.352.36.11936.86
+ Expediting130.722.676.05920.88
233.123.566.04918.68
357.9851.916.88810.94
+ Token Adapter (ours)458.2952.236.14917.03
155.9149.236.31915.74
256.0349.056.22916.11
357.5751.426.96809.66
458.4252.346.19916.38
+ +Due to the window attention mechanism, segmenting all tokens into windows, we reduce the tokens within each window instead of the entire token sequence. To ensure a fair comparison, we gradually increase $\mathcal{L}_b$ starting from the setting where Expediting balances efficiency and accuracy, to showcase the impact of applying this method at different stages of MaskDINO. + +By observing Tab. 2, it is evident that our approach is more stable than Expediting. This could be attributed to the fact that we directly combine and reconstruct tokens through matching, whereas Expediting relies on clustering, which is sensitive to hyper-parameters for different scenarios. On top of that, with the integration of token adapter into the fourth stage of the backbone network of MaskDINO, we can observe a slight enhancement in performance $(+0.1\% \mathrm{PQ})$ . + +Comparisons on Image Classification Task. We applied three methods to SWAG[27] for testing on the ImageNet-1k dataset. The results are shown in Tab. 3, where $\dagger$ indicates that the same fusion method is applied as Token Merging, i.e., layer-by-layer compression, and 32 indicates merging one row per layer. 'Size' indicates the resolution of the feature map. 'r' denotes the number of tokens compressed per layer. + +Visualization of the token injector. We visualize the token injection process in Fig. 5. Initially, we divide all tokens into three groups: representative + +Table 3: Comparisons on ImageNet1k validation dataset + +
MethodSizerAcc(%)FPS(img/s)FLOPs(G)
SWAG + ViT-L/1632 × 32-88.0722.92364.77
+ Token Adapter23 × 23-87.4332.28270.65
+ Expediting23 × 23-87.829.21275.79
+ Token Adapter†-3287.8736.49224.02
+ Token Merging-3287.8038.58218.18
+ Token Adapter†-6485.3466.04130.36
+ Token Merging-6467.5679.99116.40
+ +tokens, unique tokens, and normal tokens. Subsequently, we inject the normal tokens into the representative tokens. In the third row of Fig. 5, we use different colored borders to represent the matching results of normal tokens and representative tokens. The last row in the figure represents the results of token injection. By observing this, we can notice that token injection effectively highlights the crucial pixels in the image while preserving the original semantic information. More results and hyper-parameter settings can be found in the supplementary material. + +# 4.3 Ablation Studies + +We performed ablation experiments for each critical choice within our approach on ADE20K semantic segmentation task. The results can be found in Tab. 4. By default, we employed the official weights provided by Segmenter4 as a baseline model (Segmenter + ViT-L/16, mIoU: 51.82%, FPS: 10.68 img/s). The table cells marked in red and purple respectively denote the final choices for the token injector and token ejector. + +Token Reduction Strategy. We initially conducted ablation experiments on the token selection method, and the results are presented in Tab. 4a. In this table, random signifies the stochastic selection, alternating represents the periodic selection, and scores indicates the selection based on computed scores derived from features or attention scores. random and alternating represent the two selection strategies of ToMe. Selecting tokens based on calculated importance scores derived from features presents promising performance. + +Tab. 4b presents the influence of different evaluation metrics employed to calculate the distance among tokens. We utilize this distance matrix to perform token matching. Among these metrics, the cosine similarity distance metric demonstrates superior performance, hence we utilize cosine similarity to measure the distance between tokens. By comparing the pairwise distances between normal tokens and representative tokens, we can match each normal token with its closest representative token. + +![](images/d14396cbf5c1d9f4fde64417d905f21a6e8df91e0060f9782fd0b0c7417522a1.jpg) +Fig. 5: Token Injection Visualizations. Visualization results from Segmenter + ViT-L/16 applying Token Injector on ADE20K. The first row depicts the input image. The second row illustrates the grouping of all tokens, where gray represents representative tokens, orange represents unique tokens, and teal represents normal tokens. The third row displays the matching results between normal tokens and representative tokens. Different pairs are indicated by boxes of varying colors. The final row showcases the injected results. + +![](images/76f898105d1f49a683be9f26cf0c238ab46e5cc10ff2225a68345de9f34884fc.jpg) +(a) compressed token + +![](images/e6c3b523fc076bd87025a383466d5652519ba9db59ff60e17a011cf8eb9ee0c1.jpg) +(b) unique token +Fig.6: Ablation studies on the ratio of rows and columns of tokens. (a) Ratio ablation of compressed rows or columns $(r_h$ and $r_w)$ . (b) Ablation of retention ratios for unique token rows or columns. + +To obtain a reduced token sequence, we employ several methods, as shown in Tab. 4d. Keep denotes the preservation of representative tokens and unique tokens without merging normal tokens. Max pool and avg pool refer to pooling-based methods for token fusion, while bilinear represents resizing token sequence through bilinear interpolation. Average indicates averaging tokens within each matched group to merge them. Keep and average do not rely on pre-defined priors to adjust the tokens, hence yielding better results. However, keep leads + +Table 4: Token Adapter ablation experiments using Segmenter + ViT/L-16 on ADE20K validation dataset with official checkpoint and no retraining. $r_h$ and $r_w$ are set to 0.3. Segmenter + ViT/L-16 without Token Adapter achieves 51.82% mIoU at 10.68 img/s. We provide the mean Intersection-over-Union (mIoU) and model inference speed (img/s). + +
SelectionmIoUFPS
random51.3815.16
alternating51.4314.96
scores (feat.)52.0215.09
scores (attn.)51.1315.04
(a) Token Selection.
(b) Distance Metric.
(c) Feature Choice.
+ +
MethodmIoUFPS
keep50.6715.19
max pool48.1514.97
avg pool48.1015.10
bilinear50.0815.12
average52.0215.09
+ +(d) Merge Method. + +
ModemIoUFPS
max51.9315.13
m-threshold52.0215.09
+ +(e) Matching Mode. + +to the loss of information from normal tokens, thus we ultimately opt for using average to merge the tokens. + +Restoration Method. Considering dependency on the correlation between the reduced sequence and its original counterpart to restore the sequence length, it is imperative to contemplate the selection of features utilized in calculating this relationship matrix. Within Tab. 4c, the superscript $u$ designates unmerged features, while $m$ signifies merged features. The variable $X_{p}$ represents representative tokens, while $X_{r}$ pertains to reduced tokens. + +To facilitate understanding of the meaning of each symbol, we next summarize the algorithmic procedures of the token adapter, and omit the calculation process. For $X \in \mathbb{R}^{N \times C}$ , we first partition tokens into $X_{p} \in \mathbb{R}^{N_{p} \times C}$ , $X_{n} \in \mathbb{R}^{N_{n} \times C}$ and $X_{u} \in \mathbb{R}^{N_{u} \times C}$ . Here, we haven't merged $X_{n}$ into $X_{p}$ yet, so we can mark it as $X_{p}^{u}$ . $X_{r}$ is formed by combining $X_{p}^{u}$ and $X_{u}$ , representing the compressed sequence without merging. After merging operations, the new $X_{p}$ is marked as $X_{p}^{m}$ and the new $X_{r}$ is marked as $X_{r}^{m}$ . From the results, it shows that recovery using the merged and reduced token sequence $X_{r}^{m}$ works best. + +A comparison result is presented for the two matching modes applied to the token ejector in Tab. 4e. In Sec. 3.2, we introduced the token ejector, which matches the most similar tokens for each token in the original token sequence according to the distance matrix $M_{d} \in \mathbb{R}^{N \times N_{r}}$ . Next, the token vectors from each matched group are averaged to reconstruct the token sequence to its original size. Mode max stands for matching the most similar token for each token of + +the original token sequence, i.e., one-to-one matching. Mode $m$ -threshold, on the other hand, involves filtering based on a set threshold and distance matrix after Mode $max$ matching. The filtered matching results are combined with the results from Mode $max$ , followed by executing the corresponding fusion process. The results indicate that the latter achieves more accurate recovery of the token sequence than the former. + +We also conduct ablation experiments on the ratio of compressed rows, columns, and unique tokens, and the results are shown in Fig. 6. Although we determined the final parameter settings for our method based on the ablation experiments, we believe that there still remains a better solution. It will be discussed in the next section. + +# 5 Discussion and Conclusion + +# 5.1 Limitation + +The experimental results mentioned earlier have already demonstrated the effectiveness of our method on dense prediction tasks for hierarchical transformer structures and ViT-like structures. However, there are still many aspects of our work that remain to be explored, which we will analyze one by one next. + +Generalizability. While our method has been validated for effectiveness in multiple segmentation tasks, it has not been validated on other more complex vision tasks. Previous methods essentially involved sparse sampling of token sequences, resulting in compressed token sequences being in sequence form only. The compressed sequence features can only be processed by a sequence processing module such as Transformer. Nevertheless, it remains to be examined whether discrete sparse sequence features are more effective than dense features for image tasks. Motivated by this, we designed the token adapter, a method that performs sparse sampling on the rows and columns of image features. Compared to previous methods, our approach yields more compact sampled features, which may be more effective for image tasks. + +Adaptability. Although our core idea is rather simple, the Token Adapter still takes time to tune the hyper-parameters. The current setting allows Token Adapter to be directly applied to existing models, but it also limits its further development. We chose not to retrain the model equipped with the token adapter to directly evaluate the effectiveness of our idea. In the future, we will continue to explore more efficient neural network acceleration schemes. + +Our core contribution is to compress and reconstruct rows and columns of image features in a way that conforms to the structure of the image. When applied to vision models, Token Adapter only compresses and reconstructs image features and thus can be applied to most vision models. For large language models, the concept of a Token Adapter can still be effectively transferred. By decomposing text embeddings based on the contextual information, and then leveraging the connections between the decomposed parts to compress them, it should be able to improve the efficiency of large language models. + +# 5.2 Conclusion + +In this work, we propose Token Adapter to expedite vision transformers used in dense prediction tasks. By injecting normal tokens into representative tokens, we enhance the efficiency of vision transformers. Notably, our method can be applied in off-the-shelf models directly without additional retraining or finetuning. Tab. 2 reveals that when applied to the early stages of MaskDINO, the clustering-based Expediting method results in a significant decline in accuracy, and the matching-based Token Adapter method demonstrates superior stability. Clustering algorithms typically require fine-tuning of multiple hyper-parameters, which remain fixed once adjusted. Our approach is expected to leverage mechanisms such as cross-attention to transform fixed hyper-parameters into dynamically learnable parameters, thereby enhancing the method's generalization capability. Theoretically, the design of the compression strategy and reconstruction strategy in our algorithm allows the token adapter to be inserted in the middle of any module designed for image tasks. + +In principle, Token Adapter bears resemblance to an unparameterized form of cross-attention mechanism, both blending information from one set of tokens into another set. Our thought lies in how we employ tokens to represent semantic entities within the image. Each entity in the image can be divided into distinct local semantic regions. By assigning several representative tokens to each local semantic region and injecting the remaining normal tokens into their corresponding tokens, we achieve a more refined fusion. After the fusion, the representative tokens representing different semantic regions will engage in information exchange within transformers, greatly enhancing computational efficiency and promoting diversity among tokens. We hope our research can offer a fresh perspective on image features and inspire the design of more efficient methods for extracting image features. + +# References + +1. Bian, Z., Wang, Z., Han, W., Wang, K.: Multi-scale and token merging: Make your vit more efficient. arXiv preprint arXiv:2306.04897 (2023) +2. Bolya, D., Fu, C.Y., Dai, X., Zhang, P., Feichtenhofer, C., Hoffman, J.: Token merging: Your vit but faster. In: The Eleventh International Conference on Learning Representations (2022) +3. Bonnaerens, M., Dambre, J.: Learned thresholds token merging and pruning for vision transformers. arXiv preprint arXiv:2307.10780 (2023) +4. Carion, N., Massa, F., Synnaeve, G., Usunier, N., Kirillov, A., Zagoruyko, S.: End-to-end object detection with transformers. In: European conference on computer vision. pp. 213-229. Springer (2020) +5. Chen, Z., Duan, Y., Wang, W., He, J., Lu, T., Dai, J., Qiao, Y.: Vision transformer adapter for dense predictions. arXiv preprint arXiv:2205.08534 (2022) +6. Choromanski, K., Likhosherstov, V., Dohan, D., Song, X., Gane, A., Sarlos, T., Hawkins, P., Davis, J., Mohiuddin, A., Kaiser, L., et al.: Rethinking attention with performers. arXiv preprint arXiv:2009.14794 (2020) + +7. Cordts, M., Omran, M., Ramos, S., Rehfeld, T., Enzweiler, M., Benenson, R., Franke, U., Roth, S., Schiele, B.: The cityscapes dataset for semantic urban scene understanding. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 3213-3223 (2016) +8. Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: Imagenet: A large-scale hierarchical image database. In: 2009 IEEE conference on computer vision and pattern recognition. pp. 248-255. IEEE (2009) +9. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929 (2020) +0. Fayyaz, M., Koohpayegani, S.A., Jafari, F.R., Sengupta, S., Joze, H.R.V., Sommerlade, E., Pirsiavash, H., Gall, J.: Adaptive token sampling for efficient vision transformers. In: European Conference on Computer Vision. pp. 396-414. Springer (2022) +1. Geng, Z., Guo, M.H., Chen, H., Li, X., Wei, K., Lin, Z.: Is attention better than matrix decomposition? arXiv preprint arXiv:2109.04553 (2021) +2. Guo, M.H., Liu, Z.N., Mu, T.J., Hu, S.M.: Beyond self-attention: External attention using two linear layers for visual tasks. IEEE Transactions on Pattern Analysis and Machine Intelligence 45(5), 5436-5447 (2022) +3. Hassani, A., Walton, S., Li, J., Li, S., Shi, H.: Neighborhood attention transformer. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6185-6194 (2023) +4. He, K., Chen, X., Xie, S., Li, Y., Dólar, P., Girshick, R.: Masked autoencoders are scalable vision learners. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 16000-16009 (2022) +5. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., et al.: Segment anything. arXiv preprint arXiv:2304.02643 (2023) +6. Kong, Z., Dong, P., Ma, X., Meng, X., Sun, M., Niu, W., Shen, X., Yuan, G., Ren, B., Qin, M., et al.: Spvit: Enabling faster vision transformers via soft token pruning. arXiv preprint arXiv:2112.13890 (2021) +7. Li, F., Zhang, H., Xu, H., Liu, S., Zhang, L., Ni, L.M., Shum, H.Y.: Mask dino: Towards a unified transformer-based framework for object detection and segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3041-3050 (2023) +8. Li, Y., Yuan, G., Wen, Y., Hu, J., Evangelidis, G., Tulyakov, S., Wang, Y., Ren, J.: Efficientformer: Vision transformers at mobilenet speed. Advances in Neural Information Processing Systems 35, 12934-12949 (2022) +9. Liang, W., Yuan, Y., Ding, H., Luo, X., Lin, W., Jia, D., Zhang, Z., Zhang, C., Hu, H.: Expediting large-scale vision transformer for dense prediction without fine-tuning. Advances in Neural Information Processing Systems 35, 35462-35477 (2022) +20. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollar, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13. pp. 740-755. Springer (2014) +21. Liu, Z., Hu, H., Lin, Y., Yao, Z., Xie, Z., Wei, Y., Ning, J., Cao, Y., Zhang, Z., Dong, L., et al.: Swin transformer v2: Scaling up capacity and resolution. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 12009-12019 (2022) + +22. Liu, Z., Lin, Y., Cao, Y., Hu, H., Wei, Y., Zhang, Z., Lin, S., Guo, B.: Swin transformer: Hierarchical vision transformer using shifted windows. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 10012-10022 (2021) +23. Long, S., Zhao, Z., Pi, J., Wang, S., Wang, J.: Beyond attentive tokens: Incorporating token importance and diversity for efficient vision transformers. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10334-10343 (2023) +24. Lu, J., Yao, J., Zhang, J., Zhu, X., Xu, H., Gao, W., Xu, C., Xiang, T., Zhang, L.: Soft: Softmax-free transformer with linear complexity. Advances in Neural Information Processing Systems 34, 21297-21309 (2021) +25. Mottaghi, R., Chen, X., Liu, X., Cho, N.G., Lee, S.W., Fidler, S., Urtasun, R., Yuille, A.: The role of context for object detection and semantic segmentation in the wild. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 891-898 (2014) +26. Peng, H., Pappas, N., Yogatama, D., Schwartz, R., Smith, N.A., Kong, L.: Random feature attention. arXiv preprint arXiv:2103.02143 (2021) +27. Singh, M., Gustafson, L., Adcock, A., de Freitas Reis, V., Gedik, B., Kosaraju, R.P., Mahajan, D., Girshick, R., Dollár, P., Van Der Maaten, L.: Revisiting weakly supervised pre-training of visual perception models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 804-814 (2022) +28. Strudel, R., Garcia, R., Laptev, I., Schmid, C.: Segmenter: Transformer for semantic segmentation. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 7262-7272 (2021) +29. Tian, K., Jiang, Y., Dao, Q., Lin, C., Wang, L., Yuan, Z.: Designing bert for convolutional networks: Sparse and hierarchical masked modeling. arXiv preprint arXiv:2301.03580 (2023) +30. Tolstikhin, I.O., Houlsby, N., Kolesnikov, A., Beyer, L., Zhai, X., Unterthiner, T., Yung, J., Steiner, A., Keysers, D., Uszkoreit, J., et al.: Mlp-mixer: An allmlp architecture for vision. Advances in neural information processing systems 34, 24261-24272 (2021) +31. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. Advances in neural information processing systems 30 (2017) +32. Wang, Y., Huang, R., Song, S., Huang, Z., Huang, G.: Not all images are worth 16x16 words: Dynamic transformers for efficient image recognition. Advances in Neural Information Processing Systems 34, 11960-11973 (2021) +33. Wang, Z., Luo, H., Wang, P., Ding, F., Wang, F., Li, H.: Vtc-lfc: Vision transformer compression with low-frequency components. Advances in Neural Information Processing Systems 35, 13974-13988 (2022) +34. Yang, J., Li, C., Zhang, P., Dai, X., Xiao, B., Yuan, L., Gao, J.: Focal self-attention for local-global interactions in vision transformers. arXiv preprint arXiv:2107.00641 (2021) +35. Yang, R., Ma, H., Wu, J., Tang, Y., Xiao, X., Zheng, M., Li, X.: Scalablevit: Rethinking the context-oriented generalization of vision transformer. In: European Conference on Computer Vision. pp. 480-496. Springer (2022) +36. Ye, D., Lin, Y., Huang, Y., Sun, M.: Tr-bert: Dynamic token reduction for accelerating bert inference. arXiv preprint arXiv:2105.11618 (2021) +37. Yin, H., Vahdat, A., Alvarez, J.M., Mallya, A., Kautz, J., Molchanov, P.: A-vit: Adaptive tokens for efficient vision transformer. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10809-10818 (2022) + +38. Yu, Q., Xia, Y., Bai, Y., Lu, Y., Yuille, A.L., Shen, W.: Glance-and-gaze vision transformer. Advances in Neural Information Processing Systems 34, 12992-13003 (2021) +39. Yu, W., Luo, M., Zhou, P., Si, C., Zhou, Y., Wang, X., Feng, J., Yan, S.: Metaformer is actually what you need for vision. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10819-10829 (2022) +40. Zeng, W., Jin, S., Liu, W., Qian, C., Luo, P., Ouyang, W., Wang, X.: Not all tokens are equal: Human-centric visual analysis via token clustering transformer. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11101-11111 (2022) +41. Zeng, Z., Hawkins, C., Hong, M., Zhang, A., Pappas, N., Singh, V., Zheng, S.: Vcc: Scaling transformers to 128k tokens or more by prioritizing important tokens. arXiv preprint arXiv:2305.04241 (2023) +42. Zheng, M., Gao, P., Zhang, R., Li, K., Wang, X., Li, H., Dong, H.: End-to-end object detection with adaptive clustering transformer. arXiv preprint arXiv:2011.09315 (2020) +43. Zhou, B., Zhao, H., Puig, X., Xiao, T., Fidler, S., Barriuso, A., Torralba, A.: Semantic understanding of scenes through the ade20k dataset. International Journal of Computer Vision 127, 302-321 (2019) +44. Ziwen, C., Patnaik, K., Zhai, S., Wan, A., Ren, Z., Schwing, A.G., Colburn, A., Fuxin, L.: Autofocusformer: Image segmentation off the grid. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18227-18236 (2023) \ No newline at end of file diff --git a/2024/Removing Rows and Columns of Tokens in Vision Transformer enables Faster Dense Prediction without Retraining/images.zip b/2024/Removing Rows and Columns of Tokens in Vision Transformer enables Faster Dense Prediction without Retraining/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..2aca1b77deeb5446cb3632c6b807014aa2702cfb --- /dev/null +++ b/2024/Removing Rows and Columns of Tokens in Vision Transformer enables Faster Dense Prediction without Retraining/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c484097cd75887f62d74bdb719ccae5254477394bdf774357d45cd978f20ebfa +size 680558 diff --git a/2024/Removing Rows and Columns of Tokens in Vision Transformer enables Faster Dense Prediction without Retraining/layout.json b/2024/Removing Rows and Columns of Tokens in Vision Transformer enables Faster Dense Prediction without Retraining/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..2058c4fdb8f9882a264b30cef7b0e1df784d41a8 --- /dev/null +++ b/2024/Removing Rows and Columns of Tokens in Vision Transformer enables Faster Dense Prediction without Retraining/layout.json @@ -0,0 +1,9251 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 133, + 111, + 481, + 166 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 111, + 481, + 166 + ], + "spans": [ + { + "bbox": [ + 133, + 111, + 481, + 166 + ], + "type": "text", + "content": "Removing Rows and Columns of Tokens in Vision Transformer enables Faster Dense Prediction without Retraining" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 200, + 186, + 412, + 198 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 200, + 186, + 412, + 198 + ], + "spans": [ + { + "bbox": [ + 200, + 186, + 412, + 198 + ], + "type": "text", + "content": "Diwei " + }, + { + "bbox": [ + 200, + 186, + 412, + 198 + ], + "type": "inline_equation", + "content": "\\mathrm{Su}^{1}" + }, + { + "bbox": [ + 200, + 186, + 412, + 198 + ], + "type": "text", + "content": " Cheng Fei, and Jianxu Luo" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 155, + 208, + 459, + 231 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 155, + 208, + 459, + 231 + ], + "spans": [ + { + "bbox": [ + 155, + 208, + 459, + 231 + ], + "type": "text", + "content": "East China University of Science and Technology, Shanghai 200237, China dvsu@mail.ecust.edu.cn, cfei_jarvis@163.com, jxluo@ecust.edu.cn" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 160, + 255, + 452, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 255, + 452, + 495 + ], + "spans": [ + { + "bbox": [ + 160, + 255, + 452, + 495 + ], + "type": "text", + "content": "Abstract. In recent years, vision transformers based on self-attention mechanisms have demonstrated remarkable abilities in various tasks such as natural language processing, computer vision (CV), and multimodal applications. However, due to the high computational costs and the structural nature of images, the application of transformers to CV tasks faces challenges, particularly when handling ultra-high-resolution images. Recently, several token reduction methods have been proposed to improve the computational efficiency of transformers by reducing the number of tokens without the need for retraining. These methods primarily involve fusion based on matching or clustering. The former exhibits faster speed but suffers more accuracy loss compared to the latter. In this work, we propose a simple matching-based fusion method called Token Adapter, which achieves comparable accuracy to the clustering-based fusion method with faster speed and demonstrates higher potential in terms of robustness. Our method was applied to Segmenter, MaskDINO and SWAG, exhibiting promising performance on four tasks, including semantic segmentation, instance segmentation, panoptic segmentation, and image classification. Specifically, our method can be applied to Segmenter on ADE20k, providing " + }, + { + "bbox": [ + 160, + 255, + 452, + 495 + ], + "type": "inline_equation", + "content": "41\\%" + }, + { + "bbox": [ + 160, + 255, + 452, + 495 + ], + "type": "text", + "content": " frames per second (FPS) acceleration while maintaining full performance without retraining or fine-tuning off-the-shelf weights. Our code will be released at https://github.com/MilknoCandy-Token-Adapter." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 160, + 506, + 452, + 528 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 506, + 452, + 528 + ], + "spans": [ + { + "bbox": [ + 160, + 506, + 452, + 528 + ], + "type": "text", + "content": "Keywords: Vision transformer " + }, + { + "bbox": [ + 160, + 506, + 452, + 528 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 506, + 452, + 528 + ], + "type": "text", + "content": " Token reduction " + }, + { + "bbox": [ + 160, + 506, + 452, + 528 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 506, + 452, + 528 + ], + "type": "text", + "content": " Dense prediction " + }, + { + "bbox": [ + 160, + 506, + 452, + 528 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 506, + 452, + 528 + ], + "type": "text", + "content": " Efficient transformer" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 546, + 230, + 559 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 546, + 230, + 559 + ], + "spans": [ + { + "bbox": [ + 132, + 546, + 230, + 559 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": "Over the past few years, the computer vision field has been significantly advanced by the emergence of the Transformer architecture [31]. Vision Transformers [9] with global attention mechanism achieve state-of-the-art results on a wide range of vision tasks. To obtain better performance, researchers have scaled up the number of parameters of vision transformers to billions [15] through techniques such as self-supervised pre-training [14, 29] methods and adding visual prior knowledge [5, 22], etc. However, scaling up transformers brings performance improvements as well as huge computational costs due to the quadratic complexity" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 148, + 114, + 200, + 167 + ], + "blocks": [ + { + "bbox": [ + 148, + 114, + 200, + 167 + ], + "lines": [ + { + "bbox": [ + 148, + 114, + 200, + 167 + ], + "spans": [ + { + "bbox": [ + 148, + 114, + 200, + 167 + ], + "type": "image", + "image_path": "11284078890d1f4e17bd04b746a116476f5b9fa83092c7665867cb0034b02708.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 148, + 168, + 199, + 220 + ], + "blocks": [ + { + "bbox": [ + 148, + 168, + 199, + 220 + ], + "lines": [ + { + "bbox": [ + 148, + 168, + 199, + 220 + ], + "spans": [ + { + "bbox": [ + 148, + 168, + 199, + 220 + ], + "type": "image", + "image_path": "aff19d31763915ef52c1adbc17b1b77c1703f123491aca922a2010c494f27443.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 194, + 221, + 206, + 232 + ], + "lines": [ + { + "bbox": [ + 194, + 221, + 206, + 232 + ], + "spans": [ + { + "bbox": [ + 194, + 221, + 206, + 232 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 201, + 115, + 252, + 167 + ], + "blocks": [ + { + "bbox": [ + 201, + 115, + 252, + 167 + ], + "lines": [ + { + "bbox": [ + 201, + 115, + 252, + 167 + ], + "spans": [ + { + "bbox": [ + 201, + 115, + 252, + 167 + ], + "type": "image", + "image_path": "0e838eb5eca54fbb0f9687fd066b4d3eb6975387aa82cfb02ea75db3ccc68493.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 201, + 168, + 252, + 220 + ], + "blocks": [ + { + "bbox": [ + 201, + 168, + 252, + 220 + ], + "lines": [ + { + "bbox": [ + 201, + 168, + 252, + 220 + ], + "spans": [ + { + "bbox": [ + 201, + 168, + 252, + 220 + ], + "type": "image", + "image_path": "934332e7e73f9229188040a7940112769bd60c3dc3f68ed7bccb584f16870919.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 239, + 481, + 283 + ], + "lines": [ + { + "bbox": [ + 131, + 239, + 481, + 283 + ], + "spans": [ + { + "bbox": [ + 131, + 239, + 481, + 283 + ], + "type": "text", + "content": "Fig. 1: Visualizations of random pruning results from different methods on ADE20K. We randomly prune the same number of tokens using different methods. (a) Input Images. (b) Randomly remove tokens (c) Randomly remove specific rows and columns." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 254, + 115, + 306, + 167 + ], + "blocks": [ + { + "bbox": [ + 254, + 115, + 306, + 167 + ], + "lines": [ + { + "bbox": [ + 254, + 115, + 306, + 167 + ], + "spans": [ + { + "bbox": [ + 254, + 115, + 306, + 167 + ], + "type": "image", + "image_path": "266a7240c18e3b8b09fd39f5288d1e8162dcd9a5ad549a91a78f7d5292a7d453.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 254, + 168, + 306, + 220 + ], + "blocks": [ + { + "bbox": [ + 254, + 168, + 306, + 220 + ], + "lines": [ + { + "bbox": [ + 254, + 168, + 306, + 220 + ], + "spans": [ + { + "bbox": [ + 254, + 168, + 306, + 220 + ], + "type": "image", + "image_path": "ca8ab54901f1f71dd1e6fb61e82be9bf7ccdcb4e954c8eb7468681d2d787d974.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 300, + 221, + 313, + 232 + ], + "lines": [ + { + "bbox": [ + 300, + 221, + 313, + 232 + ], + "spans": [ + { + "bbox": [ + 300, + 221, + 313, + 232 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 306, + 115, + 359, + 167 + ], + "blocks": [ + { + "bbox": [ + 306, + 115, + 359, + 167 + ], + "lines": [ + { + "bbox": [ + 306, + 115, + 359, + 167 + ], + "spans": [ + { + "bbox": [ + 306, + 115, + 359, + 167 + ], + "type": "image", + "image_path": "c2fab848ce3bf0583ce6b711a03d9aca0ecf50bc81d131ad4ede2fa888893143.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 306, + 168, + 359, + 220 + ], + "blocks": [ + { + "bbox": [ + 306, + 168, + 359, + 220 + ], + "lines": [ + { + "bbox": [ + 306, + 168, + 359, + 220 + ], + "spans": [ + { + "bbox": [ + 306, + 168, + 359, + 220 + ], + "type": "image", + "image_path": "363888ef80f14b88a682fc169a7183b73f3897f46f54ebd503ee60e876dba8bd.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 361, + 115, + 413, + 167 + ], + "blocks": [ + { + "bbox": [ + 361, + 115, + 413, + 167 + ], + "lines": [ + { + "bbox": [ + 361, + 115, + 413, + 167 + ], + "spans": [ + { + "bbox": [ + 361, + 115, + 413, + 167 + ], + "type": "image", + "image_path": "c34d7a81f395ec1c70d792e18796b20faf073be08fbae2afa28b156d2c644b78.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 361, + 168, + 413, + 220 + ], + "blocks": [ + { + "bbox": [ + 361, + 168, + 413, + 220 + ], + "lines": [ + { + "bbox": [ + 361, + 168, + 413, + 220 + ], + "spans": [ + { + "bbox": [ + 361, + 168, + 413, + 220 + ], + "type": "image", + "image_path": "686ac65b26deb62e8d9b0675da316d45bef98c4b20fffb5f125ab43a8c53fb88.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 408, + 221, + 419, + 231 + ], + "lines": [ + { + "bbox": [ + 408, + 221, + 419, + 231 + ], + "spans": [ + { + "bbox": [ + 408, + 221, + 419, + 231 + ], + "type": "text", + "content": "(c)" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 414, + 115, + 466, + 167 + ], + "blocks": [ + { + "bbox": [ + 414, + 115, + 466, + 167 + ], + "lines": [ + { + "bbox": [ + 414, + 115, + 466, + 167 + ], + "spans": [ + { + "bbox": [ + 414, + 115, + 466, + 167 + ], + "type": "image", + "image_path": "a559b89f0ba42406227626182b87aacdfa1b529fc939e7e3f41ac8e797d81eca.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 414, + 168, + 466, + 220 + ], + "blocks": [ + { + "bbox": [ + 414, + 168, + 466, + 220 + ], + "lines": [ + { + "bbox": [ + 414, + 168, + 466, + 220 + ], + "spans": [ + { + "bbox": [ + 414, + 168, + 466, + 220 + ], + "type": "image", + "image_path": "85a5c927fd3dfbc2948526b57536194e3b96824df5546d0d8aa1e16f97b1e3fa.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 308, + 480, + 344 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 308, + 480, + 344 + ], + "spans": [ + { + "bbox": [ + 130, + 308, + 480, + 344 + ], + "type": "text", + "content": "of self-attention mechanism. Vision transformers are extremely inefficient when processing high-resolution images. Designing efficient vision transformers has become a prominent research hotspot." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 130, + 345, + 481, + 427 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 345, + 481, + 427 + ], + "spans": [ + { + "bbox": [ + 130, + 345, + 481, + 427 + ], + "type": "text", + "content": "Some researchers have reduced the computational complexity to linear by optimizing the computation of self-attention mechanism [24, 30, 39, 42], while others are focusing on reducing the number of tokens to increase the inference efficiency of vision transformers [3, 10, 32, 44]. Despite significantly improving the efficiency of the vision transformer, most of these methods necessitate the retraining or fine-tuning of the optimized Transformer, demanding substantial computational resources." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 130, + 429, + 482, + 571 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 429, + 482, + 571 + ], + "spans": [ + { + "bbox": [ + 130, + 429, + 482, + 571 + ], + "type": "text", + "content": "There are also methods that do not require fine-tuning or retraining, such as Expediting [19] and ToMe (Token Merging) [2]. Expediting integrates local information to fuse and reconstruct tokens using a clustering algorithm, and ToMe fuses the corresponding tokens based on a bipartite matching algorithm. Previous methods either selected the most representative tokens based on their importance scores or employed clustering algorithms to obtain more representative tokens. The former approach involves calculating importance scores for all tokens, introducing significant computation, and necessitating operations such as padding to ensure uniform dimensions across different batches for batch processing. Although the latter method effectively compresses token sequences, parameter tuning for clustering algorithms poses a challenge, and the iterative calculation approach incurs additional computational costs." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 131, + 572, + 481, + 597 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 572, + 481, + 597 + ], + "spans": [ + { + "bbox": [ + 131, + 572, + 481, + 597 + ], + "type": "text", + "content": "In this work, we present Token Adapter to merge rows and columns of tokens rather than merge tokens sparsely for the following three reasons:" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 135, + 605, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 24, + "blocks": [ + { + "bbox": [ + 135, + 605, + 480, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 605, + 480, + 640 + ], + "spans": [ + { + "bbox": [ + 135, + 605, + 480, + 640 + ], + "type": "text", + "content": "1. Information Redundancy. In vision tasks, multi-scale features naturally provide more information than single-scale features, resulting in spatial information redundancy in multi-scale image features." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 135, + 641, + 480, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 641, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 135, + 641, + 480, + 665 + ], + "type": "text", + "content": "2. Image Structure. As image data is derived from sampling real-world optical information, semantic information is contained within consecutive pixel" + } + ] + } + ], + "index": 23 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 214, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 214, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 214, + 100 + ], + "type": "text", + "content": "D. Su et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 147, + 116, + 481, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 147, + 116, + 481, + 163 + ], + "spans": [ + { + "bbox": [ + 147, + 116, + 481, + 163 + ], + "type": "text", + "content": "blocks rather than individual pixels. With the same number of missing pixels, removing entire rows or columns of pixels from an image could be less detrimental to the existing semantic information than the pruning partial pixels approach." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 136, + 164, + 482, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 164, + 482, + 236 + ], + "spans": [ + { + "bbox": [ + 136, + 164, + 482, + 236 + ], + "type": "text", + "content": "3. Computational Efficiency. To speed up vision transformers, one simple idea is that using fewer tokens can make them more efficient. Wang et al. [32] has observed that simple images can be accurately predicted with only a few tokens. This observation highlights the inconsistency in the number of tokens required to make accurate predictions for images of varying content complexity." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 241, + 482, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 241, + 482, + 399 + ], + "spans": [ + { + "bbox": [ + 130, + 241, + 482, + 399 + ], + "type": "text", + "content": "Fig. 1 shows the difference between our approach and the previous methods, i.e., the change in token selection strategy. It allows our method to compress token sequences efficiently and handle batch data directly without additional operations. Our experiments show that Token Adapter outperforms Expediting and ToMe on multiple segmentation datasets. To summarize, our main contributions can be outlined as follows: (i) We propose a simple yet powerful module that injects normal tokens into representative tokens while preserving unique tokens associated with dense prediction tasks; (ii) Our token reduction method preserves relative positional information among tokens, providing ViT-like models with substantial acceleration and slight accuracy loss in dense prediction tasks; (iii) Without the need for retraining or fine-tuning, our approach achieves performance comparable to clustering-based methods by employing matching-based techniques, demonstrating even better stability and speed." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 414, + 237, + 427 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 414, + 237, + 427 + ], + "spans": [ + { + "bbox": [ + 132, + 414, + 237, + 427 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 437, + 482, + 497 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 437, + 482, + 497 + ], + "spans": [ + { + "bbox": [ + 130, + 437, + 482, + 497 + ], + "type": "text", + "content": "Efficient Vision Transformer. Some studies have focused on adapting the internal configuration of transformer models to improve their efficiency. These efforts include designing the architecture of the transformer [4,18,30,35], modifying attention mechanisms [13,21,34,38], approximating attention [6,11,12,26], etc." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 498, + 482, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 498, + 482, + 629 + ], + "spans": [ + { + "bbox": [ + 130, + 498, + 482, + 629 + ], + "type": "text", + "content": "Token Reduction. Benefiting from the ability of Transformers to handle variable-length sequences, some recent work has attempted to optimize intermediate representations within transformers. Previous approaches could be categorized by their purpose into two groups: Pruning [16, 33, 36, 37] and Fusion [1, 23, 40, 41]. From a methodological perspective, these approaches can be divided into two categories: recognition-based and non-recognition-based methods. Recognition-based methods categorize all tokens into two groups, vital and minor, performing pruning or fusion operations on them separately based on specific rules. On the other hand, non-recognition-based methods employ clustering algorithms to cluster all tokens, reducing the number of tokens while preserving as much semantic information as possible." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 630, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 630, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 630, + 482, + 666 + ], + "type": "text", + "content": "Our Approach. We decompose token compression into row and column compression, providing a more efficient token reduction method. Additionally, we found that some rows or columns in tokens are unique and cannot be pruned" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 385, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 385, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 385, + 91, + 447, + 102 + ], + "type": "text", + "content": "Token Adapter" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 481, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 481, + 140 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 481, + 140 + ], + "type": "text", + "content": "or merged. In our experiments, retaining these rows or columns yields better results than not retaining them." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 158, + 233, + 172 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 158, + 233, + 172 + ], + "spans": [ + { + "bbox": [ + 132, + 158, + 233, + 172 + ], + "type": "text", + "content": "3 Methodology" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 182, + 209, + 194 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 182, + 209, + 194 + ], + "spans": [ + { + "bbox": [ + 132, + 182, + 209, + 194 + ], + "type": "text", + "content": "3.1 Overview" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 202, + 481, + 250 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 202, + 481, + 250 + ], + "spans": [ + { + "bbox": [ + 130, + 202, + 481, + 250 + ], + "type": "text", + "content": "Designing for dense prediction tasks, our token adapter comprises two simple match-based approaches: the token injector and the token ejector. Token adapter utilize a subset of tokens for the purpose of representing another subset of tokens while maintaining the integrity of the original token set." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 251, + 481, + 334 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 251, + 481, + 334 + ], + "spans": [ + { + "bbox": [ + 130, + 251, + 481, + 334 + ], + "type": "text", + "content": "Pipeline. As shown in Fig. 2, two modules of the token adapter are inserted in the middle of consecutive transformer blocks. Given a feature map " + }, + { + "bbox": [ + 130, + 251, + 481, + 334 + ], + "type": "inline_equation", + "content": "X \\in \\mathbb{R}^{H \\times W \\times C}" + }, + { + "bbox": [ + 130, + 251, + 481, + 334 + ], + "type": "text", + "content": ", we compress rows and columns with reduction ratios " + }, + { + "bbox": [ + 130, + 251, + 481, + 334 + ], + "type": "inline_equation", + "content": "r_h" + }, + { + "bbox": [ + 130, + 251, + 481, + 334 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 251, + 481, + 334 + ], + "type": "inline_equation", + "content": "r_w" + }, + { + "bbox": [ + 130, + 251, + 481, + 334 + ], + "type": "text", + "content": " respectively, reducing the number of tokens by " + }, + { + "bbox": [ + 130, + 251, + 481, + 334 + ], + "type": "inline_equation", + "content": "n_r" + }, + { + "bbox": [ + 130, + 251, + 481, + 334 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 251, + 481, + 334 + ], + "type": "inline_equation", + "content": "n_r = H \\times W \\times (r_h + r_w - r_h \\times r_w)" + }, + { + "bbox": [ + 130, + 251, + 481, + 334 + ], + "type": "text", + "content": ". Here, " + }, + { + "bbox": [ + 130, + 251, + 481, + 334 + ], + "type": "inline_equation", + "content": "r_h" + }, + { + "bbox": [ + 130, + 251, + 481, + 334 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 251, + 481, + 334 + ], + "type": "inline_equation", + "content": "r_w" + }, + { + "bbox": [ + 130, + 251, + 481, + 334 + ], + "type": "text", + "content": " are settable parameters. Finally, we rearrange the compressed tokens to rebuild the original token sequence. The details are introduced in the next section." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 351, + 236, + 364 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 351, + 236, + 364 + ], + "spans": [ + { + "bbox": [ + 132, + 351, + 236, + 364 + ], + "type": "text", + "content": "3.2 Token Adapter" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 370, + 481, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 370, + 481, + 443 + ], + "spans": [ + { + "bbox": [ + 130, + 370, + 481, + 443 + ], + "type": "text", + "content": "Token Injector. For sequence input " + }, + { + "bbox": [ + 130, + 370, + 481, + 443 + ], + "type": "inline_equation", + "content": "X_{s} \\in \\mathbb{R}^{B \\times N \\times C}" + }, + { + "bbox": [ + 130, + 370, + 481, + 443 + ], + "type": "text", + "content": ", we first reshape it to " + }, + { + "bbox": [ + 130, + 370, + 481, + 443 + ], + "type": "inline_equation", + "content": "X \\in \\mathbb{R}^{B \\times H \\times W \\times C}" + }, + { + "bbox": [ + 130, + 370, + 481, + 443 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 370, + 481, + 443 + ], + "type": "inline_equation", + "content": "N = H \\times W" + }, + { + "bbox": [ + 130, + 370, + 481, + 443 + ], + "type": "text", + "content": " represents the number of tokens. To identify which rows and columns to merge, importance scores are computed for each row and column. In images, areas with sharp pixel value changes usually correspond to inter-object or intra-object boundaries. Moreover, areas with larger pixel values compared to their surroundings are typically more significant." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 444, + 481, + 550 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 444, + 481, + 550 + ], + "spans": [ + { + "bbox": [ + 130, + 444, + 481, + 550 + ], + "type": "text", + "content": "In our experiments, we perform two steps of dimension reduction on features to obtain their importance scores. For the input feature " + }, + { + "bbox": [ + 130, + 444, + 481, + 550 + ], + "type": "inline_equation", + "content": "X \\in \\mathbb{R}^{B \\times H \\times W \\times C}" + }, + { + "bbox": [ + 130, + 444, + 481, + 550 + ], + "type": "text", + "content": ", we first calculate the maximum value of each token vector (you can also calculate its sum value or range value instead), thereby compressing the features into " + }, + { + "bbox": [ + 130, + 444, + 481, + 550 + ], + "type": "inline_equation", + "content": "X_{hw} \\in \\mathbb{R}^{B \\times H \\times W}" + }, + { + "bbox": [ + 130, + 444, + 481, + 550 + ], + "type": "text", + "content": ". Subsequently, the same procedure is applied to each row and column, yielding " + }, + { + "bbox": [ + 130, + 444, + 481, + 550 + ], + "type": "inline_equation", + "content": "S_h \\in \\mathbb{R}^{B \\times H}" + }, + { + "bbox": [ + 130, + 444, + 481, + 550 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 444, + 481, + 550 + ], + "type": "inline_equation", + "content": "S_w \\in \\mathbb{R}^{B \\times W}" + }, + { + "bbox": [ + 130, + 444, + 481, + 550 + ], + "type": "text", + "content": " as the importance scores. Please note that we employed a straightforward method to calculate the importance scores, yet we believe that a better importance assessment algorithm could yield better results." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 551, + 481, + 599 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 551, + 481, + 599 + ], + "spans": [ + { + "bbox": [ + 130, + 551, + 481, + 599 + ], + "type": "text", + "content": "Based on these scores, tokens are categorized into three groups: representative tokens, unique tokens, and normal tokens. Normal tokens are injected into representative tokens, while unique tokens remain intact. The compressed tokens are rearranged in their previous positions. A brief flow is delineated below:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 135, + 605, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 135, + 605, + 405, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 605, + 405, + 616 + ], + "spans": [ + { + "bbox": [ + 135, + 605, + 405, + 616 + ], + "type": "text", + "content": "1. Calculate the importance scores for each row and column." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 135, + 617, + 481, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 617, + 481, + 640 + ], + "spans": [ + { + "bbox": [ + 135, + 617, + 481, + 640 + ], + "type": "text", + "content": "2. Group all tokens into: representative tokens, unique tokens and normal tokens by importance scores." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 135, + 642, + 451, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 642, + 451, + 652 + ], + "spans": [ + { + "bbox": [ + 135, + 642, + 451, + 652 + ], + "type": "text", + "content": "3. Match the most similar representative token for each normal tokens." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 135, + 653, + 427, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 653, + 427, + 665 + ], + "spans": [ + { + "bbox": [ + 135, + 653, + 427, + 665 + ], + "type": "text", + "content": "4. Merge normal tokens with corresponding representative tokens." + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 214, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 214, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 214, + 101 + ], + "type": "text", + "content": "D. Su et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 168, + 114, + 447, + 228 + ], + "blocks": [ + { + "bbox": [ + 168, + 114, + 447, + 228 + ], + "lines": [ + { + "bbox": [ + 168, + 114, + 447, + 228 + ], + "spans": [ + { + "bbox": [ + 168, + 114, + 447, + 228 + ], + "type": "image", + "image_path": "e855a7467c28b6fa09f0abbe5aedc80e37aab77478316a324c3b442234517fb2.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 168, + 229, + 446, + 296 + ], + "blocks": [ + { + "bbox": [ + 168, + 229, + 446, + 296 + ], + "lines": [ + { + "bbox": [ + 168, + 229, + 446, + 296 + ], + "spans": [ + { + "bbox": [ + 168, + 229, + 446, + 296 + ], + "type": "image", + "image_path": "4119ace7158c446cc6f91d0c27bfd84d4d7fe52b11155638be2856f7e258471a.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 168, + 296, + 446, + 411 + ], + "blocks": [ + { + "bbox": [ + 168, + 296, + 446, + 411 + ], + "lines": [ + { + "bbox": [ + 168, + 296, + 446, + 411 + ], + "spans": [ + { + "bbox": [ + 168, + 296, + 446, + 411 + ], + "type": "image", + "image_path": "d32acbea98682afb06ebaabd3abbe575d3adbfec9a1e3eef412538759f03c285.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 418, + 482, + 475 + ], + "lines": [ + { + "bbox": [ + 130, + 418, + 482, + 475 + ], + "spans": [ + { + "bbox": [ + 130, + 418, + 482, + 475 + ], + "type": "text", + "content": "Fig. 2: Token Adapter. (a) We use Token Injector module and Token ejector module to wrap " + }, + { + "bbox": [ + 130, + 418, + 482, + 475 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_m" + }, + { + "bbox": [ + 130, + 418, + 482, + 475 + ], + "type": "text", + "content": " Transformer blocks. (b) Token Injector mitigates the computational overhead of subsequent Transformer blocks by injecting normal tokens into representative tokens. (c) By leveraging token-to-token relationships, we represent the original sequence with compressed tokens." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 135, + 499, + 480, + 522 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 499, + 480, + 522 + ], + "spans": [ + { + "bbox": [ + 135, + 499, + 480, + 522 + ], + "type": "text", + "content": "5. Rearrange new representative tokens and unique tokens in their original order." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 533, + 482, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 482, + 628 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 482, + 628 + ], + "type": "text", + "content": "Within an image, consecutive pixel blocks form regions containing semantic information, and these associated semantic regions collectively constitute complete image objects. Merging entire rows or columns preserves the continuity of the original semantic regions while maintaining their relative positions, as depicted in Fig. 3. Our key thought involves representing each object through a refined set of representative and unique tokens. The results presented in Fig. 6 indicate that the retention of partial unique tokens generates better prediction accuracy." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 630, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 630, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 630, + 481, + 665 + ], + "type": "text", + "content": "Token Ejector. After compression, each representative token contains a blend of information from the original representative tokens and the merged normal tokens. Therefore, we exploit the correlation (e.g., the distance matrix" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 385, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 385, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 385, + 91, + 447, + 102 + ], + "type": "text", + "content": "Token Adapter" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 149, + 114, + 227, + 269 + ], + "blocks": [ + { + "bbox": [ + 149, + 114, + 227, + 269 + ], + "lines": [ + { + "bbox": [ + 149, + 114, + 227, + 269 + ], + "spans": [ + { + "bbox": [ + 149, + 114, + 227, + 269 + ], + "type": "image", + "image_path": "8ba4edbdf1fddf24fab397428548bebfd6c5b3f10c011b8b6b52b1d59a34d7a8.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 182, + 270, + 195, + 281 + ], + "lines": [ + { + "bbox": [ + 182, + 270, + 195, + 281 + ], + "spans": [ + { + "bbox": [ + 182, + 270, + 195, + 281 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 228, + 114, + 306, + 269 + ], + "blocks": [ + { + "bbox": [ + 228, + 114, + 306, + 269 + ], + "lines": [ + { + "bbox": [ + 228, + 114, + 306, + 269 + ], + "spans": [ + { + "bbox": [ + 228, + 114, + 306, + 269 + ], + "type": "image", + "image_path": "6c703140f6caf47edbaf3769f69371c4e1554b1865a177e99be2dce5093ea2db.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 261, + 270, + 274, + 280 + ], + "lines": [ + { + "bbox": [ + 261, + 270, + 274, + 280 + ], + "spans": [ + { + "bbox": [ + 261, + 270, + 274, + 280 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 307, + 114, + 386, + 269 + ], + "blocks": [ + { + "bbox": [ + 307, + 114, + 386, + 269 + ], + "lines": [ + { + "bbox": [ + 307, + 114, + 386, + 269 + ], + "spans": [ + { + "bbox": [ + 307, + 114, + 386, + 269 + ], + "type": "image", + "image_path": "409d5efa72c217c265f049798c9bac729e27cc5e644b06991b14ff099326e6e2.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 340, + 270, + 353, + 281 + ], + "lines": [ + { + "bbox": [ + 340, + 270, + 353, + 281 + ], + "spans": [ + { + "bbox": [ + 340, + 270, + 353, + 281 + ], + "type": "text", + "content": "(c)" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 289, + 482, + 344 + ], + "lines": [ + { + "bbox": [ + 130, + 289, + 482, + 344 + ], + "spans": [ + { + "bbox": [ + 130, + 289, + 482, + 344 + ], + "type": "text", + "content": "Fig. 3: Token Selection Visualizations. Results obtained through the removal of specific rows and columns from the ADE20K validation dataset using Segmenter+ViT-L/16 equipped with Token Adapter (" + }, + { + "bbox": [ + 130, + 289, + 482, + 344 + ], + "type": "inline_equation", + "content": "r_h = r_w = 0.3" + }, + { + "bbox": [ + 130, + 289, + 482, + 344 + ], + "type": "text", + "content": "). (a) Input Images. (b) Remove specific rows and columns. (c) Input images in which each patch is filled with the mean color of its respective region. (d) Remove specific rows and columns from (c)." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 387, + 114, + 465, + 269 + ], + "blocks": [ + { + "bbox": [ + 387, + 114, + 465, + 269 + ], + "lines": [ + { + "bbox": [ + 387, + 114, + 465, + 269 + ], + "spans": [ + { + "bbox": [ + 387, + 114, + 465, + 269 + ], + "type": "image", + "image_path": "5522f4c3a7fbe32357c767eb7b6ada4c7e23ffb61a8ada1d60afc16a5b2e753c.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 419, + 270, + 432, + 280 + ], + "lines": [ + { + "bbox": [ + 419, + 270, + 432, + 280 + ], + "spans": [ + { + "bbox": [ + 419, + 270, + 432, + 280 + ], + "type": "text", + "content": "(d)" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 368, + 480, + 393 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 368, + 480, + 393 + ], + "spans": [ + { + "bbox": [ + 130, + 368, + 480, + 393 + ], + "type": "inline_equation", + "content": "M_{d}" + }, + { + "bbox": [ + 130, + 368, + 480, + 393 + ], + "type": "text", + "content": ") between the reduced sequence and the original sequence to restore the original length. Illustrated in Fig. 2, the primary steps are as follows:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 135, + 399, + 480, + 460 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 135, + 399, + 480, + 424 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 399, + 480, + 424 + ], + "spans": [ + { + "bbox": [ + 135, + 399, + 480, + 424 + ], + "type": "text", + "content": "1. Compute the distance matrix " + }, + { + "bbox": [ + 135, + 399, + 480, + 424 + ], + "type": "inline_equation", + "content": "M_{d} \\in \\mathbb{R}^{N \\times N_{r}}" + }, + { + "bbox": [ + 135, + 399, + 480, + 424 + ], + "type": "text", + "content": " between " + }, + { + "bbox": [ + 135, + 399, + 480, + 424 + ], + "type": "inline_equation", + "content": "X \\in \\mathbb{R}^{N \\times C}" + }, + { + "bbox": [ + 135, + 399, + 480, + 424 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 135, + 399, + 480, + 424 + ], + "type": "inline_equation", + "content": "X' \\in \\mathbb{R}^{N_{r} \\times C}" + }, + { + "bbox": [ + 135, + 399, + 480, + 424 + ], + "type": "text", + "content": ", with " + }, + { + "bbox": [ + 135, + 399, + 480, + 424 + ], + "type": "inline_equation", + "content": "N_{r} = N - n_{r}" + }, + { + "bbox": [ + 135, + 399, + 480, + 424 + ], + "type": "text", + "content": " denoting the length of reduced sequence." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 135, + 425, + 462, + 436 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 425, + 462, + 436 + ], + "spans": [ + { + "bbox": [ + 135, + 425, + 462, + 436 + ], + "type": "text", + "content": "2. Perform token matching between tokens in " + }, + { + "bbox": [ + 135, + 425, + 462, + 436 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 135, + 425, + 462, + 436 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 135, + 425, + 462, + 436 + ], + "type": "inline_equation", + "content": "X^{\\prime}" + }, + { + "bbox": [ + 135, + 425, + 462, + 436 + ], + "type": "text", + "content": " according to " + }, + { + "bbox": [ + 135, + 425, + 462, + 436 + ], + "type": "inline_equation", + "content": "M_{d}" + }, + { + "bbox": [ + 135, + 425, + 462, + 436 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 135, + 437, + 480, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 437, + 480, + 460 + ], + "spans": [ + { + "bbox": [ + 135, + 437, + 480, + 460 + ], + "type": "text", + "content": "3. Reconstruct the original sequence " + }, + { + "bbox": [ + 135, + 437, + 480, + 460 + ], + "type": "inline_equation", + "content": "\\tilde{X}" + }, + { + "bbox": [ + 135, + 437, + 480, + 460 + ], + "type": "text", + "content": " by aggregating tokens in " + }, + { + "bbox": [ + 135, + 437, + 480, + 460 + ], + "type": "inline_equation", + "content": "X_r'" + }, + { + "bbox": [ + 135, + 437, + 480, + 460 + ], + "type": "text", + "content": " in accordance with the alignment results." + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 130, + 468, + 482, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 468, + 482, + 517 + ], + "spans": [ + { + "bbox": [ + 130, + 468, + 482, + 517 + ], + "type": "text", + "content": "Note that the distance matrix is computed from the original sequence " + }, + { + "bbox": [ + 130, + 468, + 482, + 517 + ], + "type": "inline_equation", + "content": "X_{s}" + }, + { + "bbox": [ + 130, + 468, + 482, + 517 + ], + "type": "text", + "content": " and the reduced sequence " + }, + { + "bbox": [ + 130, + 468, + 482, + 517 + ], + "type": "inline_equation", + "content": "X^{\\prime}" + }, + { + "bbox": [ + 130, + 468, + 482, + 517 + ], + "type": "text", + "content": " that has not been refined by the subsequent transformer blocks. By following this approach, the reduced sequence can be efficiently reverted to its initial length." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 536, + 230, + 550 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 536, + 230, + 550 + ], + "spans": [ + { + "bbox": [ + 132, + 536, + 230, + 550 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 561, + 190, + 574 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 561, + 190, + 574 + ], + "spans": [ + { + "bbox": [ + 132, + 561, + 190, + 574 + ], + "type": "text", + "content": "4.1 Setup" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "content": "Dataset and evaluation metrics. We present task-specific metrics, including the mean Intersection-over-Union (mIoU) and average precision (AP), etc. Furthermore, we compare our method with others by analyzing model inference speed and floating-point operations (FLOPs). The datasets utilized for evaluation include ADE20K [43], PASCAL-Context [25], CityScapes [7], COCO [20], and ImageNet-1k [8], covering semantic segmentation, instance segmentation, panotic segmentation, object detection, and image classification tasks." + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 214, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 214, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 214, + 101 + ], + "type": "text", + "content": "D. Su et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 213 + ], + "type": "text", + "content": "Implementation details. To assess the generalization of our method, we equipped two models with Token Adapter and conducted a series of experiments on four well-known datasets for dense prediction tasks. The proposed module is inserted into two classic models off-the-shelf with no re-training or fine-tuning required. All experiments are performed on a single NVIDIA GeForce RTX3090 GPU using Pytorch framework. Further details are presented in the following sections. Our code implementation is based on the open-source code of Expediting and ToMe." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 228, + 228, + 239 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 228, + 228, + 239 + ], + "spans": [ + { + "bbox": [ + 131, + 228, + 228, + 239 + ], + "type": "text", + "content": "4.2 Main Results" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 245, + 482, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 245, + 482, + 319 + ], + "spans": [ + { + "bbox": [ + 130, + 245, + 482, + 319 + ], + "type": "text", + "content": "In this section, we delve into a comparative analysis of the proposed approach with other token reduction methods that do not require retraining and are applicable to dense prediction tasks. Since the number of GFLOPs does not provide a direct measure of model inference efficiency, our primary concern lies in accuracy metrics for dense prediction tasks and corresponding metrics for evaluating inference speed." + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 136, + 336, + 250, + 423 + ], + "blocks": [ + { + "bbox": [ + 136, + 336, + 250, + 423 + ], + "lines": [ + { + "bbox": [ + 136, + 336, + 250, + 423 + ], + "spans": [ + { + "bbox": [ + 136, + 336, + 250, + 423 + ], + "type": "image", + "image_path": "d26110e380ee614d43021ab9cb1491e427a2bd2ec4287dd696f206de5ccd4e97.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 250, + 337, + 363, + 423 + ], + "blocks": [ + { + "bbox": [ + 250, + 337, + 363, + 423 + ], + "lines": [ + { + "bbox": [ + 250, + 337, + 363, + 423 + ], + "spans": [ + { + "bbox": [ + 250, + 337, + 363, + 423 + ], + "type": "image", + "image_path": "4f626b9755485388a28079aa8f711693e86678c1d3af5a3222513f36d5566951.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 365, + 337, + 477, + 423 + ], + "blocks": [ + { + "bbox": [ + 365, + 337, + 477, + 423 + ], + "lines": [ + { + "bbox": [ + 365, + 337, + 477, + 423 + ], + "spans": [ + { + "bbox": [ + 365, + 337, + 477, + 423 + ], + "type": "image", + "image_path": "5e39608380b77b7f6b34d0ed8520db5ad1bf13056eb6ee51d1af53eb9596c6d8.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 137, + 435, + 249, + 522 + ], + "blocks": [ + { + "bbox": [ + 187, + 426, + 199, + 435 + ], + "lines": [ + { + "bbox": [ + 187, + 426, + 199, + 435 + ], + "spans": [ + { + "bbox": [ + 187, + 426, + 199, + 435 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 137, + 435, + 249, + 522 + ], + "lines": [ + { + "bbox": [ + 137, + 435, + 249, + 522 + ], + "spans": [ + { + "bbox": [ + 137, + 435, + 249, + 522 + ], + "type": "image", + "image_path": "7d4f30a73a80dfa0133d4f816af59c8f22d6c376b380d198c82fb8a768510a19.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 187, + 525, + 199, + 534 + ], + "lines": [ + { + "bbox": [ + 187, + 525, + 199, + 534 + ], + "spans": [ + { + "bbox": [ + 187, + 525, + 199, + 534 + ], + "type": "text", + "content": "(d)" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 131, + 544, + 482, + 578 + ], + "lines": [ + { + "bbox": [ + 131, + 544, + 482, + 578 + ], + "spans": [ + { + "bbox": [ + 131, + 544, + 482, + 578 + ], + "type": "text", + "content": "Fig. 4: Comparisons on Semantic Segmentation Tasks. A comprehensive comparison of our approach with Expediting and ToMe on ADE20K (1-st column), PACAL-Context (2-ed column), and Cityscapes (3-rd column) using Segmenter + ViT-L/16." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 251, + 437, + 362, + 522 + ], + "blocks": [ + { + "bbox": [ + 300, + 426, + 313, + 435 + ], + "lines": [ + { + "bbox": [ + 300, + 426, + 313, + 435 + ], + "spans": [ + { + "bbox": [ + 300, + 426, + 313, + 435 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 251, + 437, + 362, + 522 + ], + "lines": [ + { + "bbox": [ + 251, + 437, + 362, + 522 + ], + "spans": [ + { + "bbox": [ + 251, + 437, + 362, + 522 + ], + "type": "image", + "image_path": "3b8bda187a68c6030b7991bd2325e2a906e952ed5659e422dbb389400d7879f8.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 301, + 525, + 313, + 534 + ], + "lines": [ + { + "bbox": [ + 301, + 525, + 313, + 534 + ], + "spans": [ + { + "bbox": [ + 301, + 525, + 313, + 534 + ], + "type": "text", + "content": "(e)" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 365, + 437, + 476, + 522 + ], + "blocks": [ + { + "bbox": [ + 415, + 426, + 426, + 435 + ], + "lines": [ + { + "bbox": [ + 415, + 426, + 426, + 435 + ], + "spans": [ + { + "bbox": [ + 415, + 426, + 426, + 435 + ], + "type": "text", + "content": "(c)" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 365, + 437, + 476, + 522 + ], + "lines": [ + { + "bbox": [ + 365, + 437, + 476, + 522 + ], + "spans": [ + { + "bbox": [ + 365, + 437, + 476, + 522 + ], + "type": "image", + "image_path": "ddb848b8b2fa57243098d9e6e05b1490a466a0112fca2e734f47e78f9ba5fb95.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 415, + 525, + 426, + 534 + ], + "lines": [ + { + "bbox": [ + 415, + 525, + 426, + 534 + ], + "spans": [ + { + "bbox": [ + 415, + 525, + 426, + 534 + ], + "type": "text", + "content": "(f)" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 599, + 482, + 648 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 599, + 482, + 648 + ], + "spans": [ + { + "bbox": [ + 130, + 599, + 482, + 648 + ], + "type": "text", + "content": "Comparisons on Semantic Segmentation Task. We use the ViT-like segmentation network model Segmenter [28] as our baseline for three benchmarks, including ADE20K, PASCAL-Context, and Cityscapes, reporting semantic segmentation results. We compare our approach with Expediting1 [19] and" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 385, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 385, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 385, + 91, + 447, + 102 + ], + "type": "text", + "content": "Token Adapter" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 133, + 653, + 410, + 666 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 653, + 410, + 666 + ], + "spans": [ + { + "bbox": [ + 133, + 653, + 410, + 666 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 133, + 653, + 410, + 666 + ], + "type": "text", + "content": " https://github.com/Expedit-LargeScale-Vision-Transformer" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 149, + 157, + 466, + 446 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "type": "text", + "content": "Table 1: Comparisons on Semantic Segmentation Tasks. Quantitative comparisons of our method with other token reduction approaches that do not require retraining on semantic segmentation tasks." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 149, + 157, + 466, + 446 + ], + "lines": [ + { + "bbox": [ + 149, + 157, + 466, + 446 + ], + "spans": [ + { + "bbox": [ + 149, + 157, + 466, + 446 + ], + "type": "table", + "html": "
DatasetMethodmIoU(%)FPS(img/s)FLOPs(G)
ADE20KSegmenter + ViT-L/1651.8210.68658.98
+ Token Merging47.1217.15321.29
50.9714.02448.7
+ Expediting43.8622.4309.96
51.8514.94481.9
+ Token Adapter (Ours)45.7324.41190.74
52.0215.09273.0
PASCAL-ContextSegmenter + ViT-L/1658.0724.17338.74
+ Token Merging54.2421.84172.38
57.2321.39236.5
+ Expediting48.0350.2889.66
58.1434.50220.76
+ Token Adapter (Ours)47.8063.3465.29
57.8734.53141.5
CityscapesSegmenter + ViT-L/1679.147.42995.55
+ Token Merging71.2313.12448.5
77.818.88651.5
+ Expediting55.3916.68406.69
78.4610.17681.57
+ Token Adapter (Ours)72.0819.38239.65
78.4810.2376.16
", + "image_path": "8c7995e057e6afae0f03c0a182dd415702d9b35a921950a80dc9235367e19d3d.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 467, + 480, + 527 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 467, + 480, + 527 + ], + "spans": [ + { + "bbox": [ + 130, + 467, + 480, + 527 + ], + "type": "inline_equation", + "content": "\\mathrm{ToMe}^2" + }, + { + "bbox": [ + 130, + 467, + 480, + 527 + ], + "type": "text", + "content": " [2], which are clustering-based and matching-based methods for token reduction that do not require retraining, respectively. The results are reported in Tab. 1. The two lines for each method in Tab. 1 represent the best result and the maximum speedup that can be obtained with the same token compression settings, respectively." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 527, + 482, + 610 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 527, + 482, + 610 + ], + "spans": [ + { + "bbox": [ + 130, + 527, + 482, + 610 + ], + "type": "text", + "content": "To ensure fairness and comprehensive performance comparisons among different methods, we fine-tuned the hyper-parameters of each method starting from the default settings mentioned in their papers. The results are depicted in Fig. 4. Under maximum acceleration, Token Adapter exhibits superior stability to Expediting " + }, + { + "bbox": [ + 130, + 527, + 482, + 610 + ], + "type": "inline_equation", + "content": "(+6.11\\% \\mathrm{mIoU}" + }, + { + "bbox": [ + 130, + 527, + 482, + 610 + ], + "type": "text", + "content": " in average) and achieves higher velocity compared to ToMe " + }, + { + "bbox": [ + 130, + 527, + 482, + 610 + ], + "type": "inline_equation", + "content": "(+93.33\\% \\mathrm{fps}" + }, + { + "bbox": [ + 130, + 527, + 482, + 610 + ], + "type": "text", + "content": " in average). Refer to our supplementary material for detailed parameter settings." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 611, + 482, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 611, + 482, + 647 + ], + "spans": [ + { + "bbox": [ + 130, + 611, + 482, + 647 + ], + "type": "text", + "content": "Comparisons on Instance/Panoptic Segmentation Task. To further explore the generalization capabilities of our method across various Transformer frameworks, we have selected the latest framework, MaskDINO [17]. It utilized" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 214, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 214, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 214, + 101 + ], + "type": "text", + "content": "D. Su et al." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 133, + 653, + 332, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 653, + 332, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 653, + 332, + 665 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 133, + 653, + 332, + 665 + ], + "type": "text", + "content": " https://github.com/facebookresearch/ToMe" + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 175 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 175 + ], + "type": "text", + "content": "Swin Transformer as its backbone network and has exhibited remarkable performance in tasks like instance segmentation and panoptic segmentation on COCO. We employed MaskDINO as our baseline model to compare the overall performance of our approach with Expediting on the COCO dataset, as presented in Tab. 2." + } + ] + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 151, + 237, + 463, + 384 + ], + "blocks": [ + { + "bbox": [ + 130, + 194, + 483, + 228 + ], + "lines": [ + { + "bbox": [ + 130, + 194, + 483, + 228 + ], + "spans": [ + { + "bbox": [ + 130, + 194, + 483, + 228 + ], + "type": "text", + "content": "Table 2: Comparisons on Instance/Panoptic Segmentation Tasks. Inserting both the proposed algorithm and Expediting at different stages of MaskDINO + Swin-L, we compare their comprehensive performance on the COCO dataset." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 151, + 237, + 463, + 384 + ], + "lines": [ + { + "bbox": [ + 151, + 237, + 463, + 384 + ], + "spans": [ + { + "bbox": [ + 151, + 237, + 463, + 384 + ], + "type": "table", + "html": "
MethodStagePQ(%)MaskAP(%)FPS(img/s)FLOPs(G)
MaskDINO + Swin-L-58.352.36.11936.86
+ Expediting130.722.676.05920.88
233.123.566.04918.68
357.9851.916.88810.94
+ Token Adapter (ours)458.2952.236.14917.03
155.9149.236.31915.74
256.0349.056.22916.11
357.5751.426.96809.66
458.4252.346.19916.38
", + "image_path": "88b76c1f5f7f31fbb2f7f6e22f61c5301f7168dff29d21a536e83a7bef9a41a7.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 408, + 482, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 408, + 482, + 467 + ], + "spans": [ + { + "bbox": [ + 130, + 408, + 482, + 467 + ], + "type": "text", + "content": "Due to the window attention mechanism, segmenting all tokens into windows, we reduce the tokens within each window instead of the entire token sequence. To ensure a fair comparison, we gradually increase " + }, + { + "bbox": [ + 130, + 408, + 482, + 467 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_b" + }, + { + "bbox": [ + 130, + 408, + 482, + 467 + ], + "type": "text", + "content": " starting from the setting where Expediting balances efficiency and accuracy, to showcase the impact of applying this method at different stages of MaskDINO." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 468, + 482, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 468, + 482, + 552 + ], + "spans": [ + { + "bbox": [ + 130, + 468, + 482, + 552 + ], + "type": "text", + "content": "By observing Tab. 2, it is evident that our approach is more stable than Expediting. This could be attributed to the fact that we directly combine and reconstruct tokens through matching, whereas Expediting relies on clustering, which is sensitive to hyper-parameters for different scenarios. On top of that, with the integration of token adapter into the fourth stage of the backbone network of MaskDINO, we can observe a slight enhancement in performance " + }, + { + "bbox": [ + 130, + 468, + 482, + 552 + ], + "type": "inline_equation", + "content": "(+0.1\\% \\mathrm{PQ})" + }, + { + "bbox": [ + 130, + 468, + 482, + 552 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 552, + 482, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 552, + 482, + 624 + ], + "spans": [ + { + "bbox": [ + 130, + 552, + 482, + 624 + ], + "type": "text", + "content": "Comparisons on Image Classification Task. We applied three methods to SWAG[27] for testing on the ImageNet-1k dataset. The results are shown in Tab. 3, where " + }, + { + "bbox": [ + 130, + 552, + 482, + 624 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 130, + 552, + 482, + 624 + ], + "type": "text", + "content": " indicates that the same fusion method is applied as Token Merging, i.e., layer-by-layer compression, and 32 indicates merging one row per layer. 'Size' indicates the resolution of the feature map. 'r' denotes the number of tokens compressed per layer." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 624, + 482, + 648 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 624, + 482, + 648 + ], + "spans": [ + { + "bbox": [ + 130, + 624, + 482, + 648 + ], + "type": "text", + "content": "Visualization of the token injector. We visualize the token injection process in Fig. 5. Initially, we divide all tokens into three groups: representative" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 385, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 385, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 385, + 91, + 447, + 102 + ], + "type": "text", + "content": "Token Adapter" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 133, + 653, + 334, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 653, + 334, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 653, + 334, + 665 + ], + "type": "text", + "content": "3 https://github.com/facebookresearch/SWAG" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 167, + 135, + 447, + 240 + ], + "blocks": [ + { + "bbox": [ + 189, + 114, + 424, + 126 + ], + "lines": [ + { + "bbox": [ + 189, + 114, + 424, + 126 + ], + "spans": [ + { + "bbox": [ + 189, + 114, + 424, + 126 + ], + "type": "text", + "content": "Table 3: Comparisons on ImageNet1k validation dataset" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 167, + 135, + 447, + 240 + ], + "lines": [ + { + "bbox": [ + 167, + 135, + 447, + 240 + ], + "spans": [ + { + "bbox": [ + 167, + 135, + 447, + 240 + ], + "type": "table", + "html": "
MethodSizerAcc(%)FPS(img/s)FLOPs(G)
SWAG + ViT-L/1632 × 32-88.0722.92364.77
+ Token Adapter23 × 23-87.4332.28270.65
+ Expediting23 × 23-87.829.21275.79
+ Token Adapter†-3287.8736.49224.02
+ Token Merging-3287.8038.58218.18
+ Token Adapter†-6485.3466.04130.36
+ Token Merging-6467.5679.99116.40
", + "image_path": "70cd2c895d44e6b466ecbb0ce0faa79b61735849a62eece8e6038370ede23a6f.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 264, + 482, + 360 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 264, + 482, + 360 + ], + "spans": [ + { + "bbox": [ + 130, + 264, + 482, + 360 + ], + "type": "text", + "content": "tokens, unique tokens, and normal tokens. Subsequently, we inject the normal tokens into the representative tokens. In the third row of Fig. 5, we use different colored borders to represent the matching results of normal tokens and representative tokens. The last row in the figure represents the results of token injection. By observing this, we can notice that token injection effectively highlights the crucial pixels in the image while preserving the original semantic information. More results and hyper-parameter settings can be found in the supplementary material." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 380, + 246, + 392 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 380, + 246, + 392 + ], + "spans": [ + { + "bbox": [ + 132, + 380, + 246, + 392 + ], + "type": "text", + "content": "4.3 Ablation Studies" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 403, + 482, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 403, + 482, + 475 + ], + "spans": [ + { + "bbox": [ + 130, + 403, + 482, + 475 + ], + "type": "text", + "content": "We performed ablation experiments for each critical choice within our approach on ADE20K semantic segmentation task. The results can be found in Tab. 4. By default, we employed the official weights provided by Segmenter4 as a baseline model (Segmenter + ViT-L/16, mIoU: 51.82%, FPS: 10.68 img/s). The table cells marked in red and purple respectively denote the final choices for the token injector and token ejector." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 475, + 482, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 475, + 482, + 559 + ], + "spans": [ + { + "bbox": [ + 130, + 475, + 482, + 559 + ], + "type": "text", + "content": "Token Reduction Strategy. We initially conducted ablation experiments on the token selection method, and the results are presented in Tab. 4a. In this table, random signifies the stochastic selection, alternating represents the periodic selection, and scores indicates the selection based on computed scores derived from features or attention scores. random and alternating represent the two selection strategies of ToMe. Selecting tokens based on calculated importance scores derived from features presents promising performance." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 560, + 482, + 643 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 560, + 482, + 643 + ], + "spans": [ + { + "bbox": [ + 130, + 560, + 482, + 643 + ], + "type": "text", + "content": "Tab. 4b presents the influence of different evaluation metrics employed to calculate the distance among tokens. We utilize this distance matrix to perform token matching. Among these metrics, the cosine similarity distance metric demonstrates superior performance, hence we utilize cosine similarity to measure the distance between tokens. By comparing the pairwise distances between normal tokens and representative tokens, we can match each normal token with its closest representative token." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 214, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 214, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 214, + 101 + ], + "type": "text", + "content": "D. Su et al." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 133, + 653, + 319, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 653, + 319, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 653, + 319, + 665 + ], + "type": "text", + "content": "4 https://github.com/rstrudel/segmenter" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 132, + 113, + 482, + 289 + ], + "blocks": [ + { + "bbox": [ + 132, + 113, + 482, + 289 + ], + "lines": [ + { + "bbox": [ + 132, + 113, + 482, + 289 + ], + "spans": [ + { + "bbox": [ + 132, + 113, + 482, + 289 + ], + "type": "image", + "image_path": "d14396cbf5c1d9f4fde64417d905f21a6e8df91e0060f9782fd0b0c7417522a1.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 296, + 482, + 377 + ], + "lines": [ + { + "bbox": [ + 130, + 296, + 482, + 377 + ], + "spans": [ + { + "bbox": [ + 130, + 296, + 482, + 377 + ], + "type": "text", + "content": "Fig. 5: Token Injection Visualizations. Visualization results from Segmenter + ViT-L/16 applying Token Injector on ADE20K. The first row depicts the input image. The second row illustrates the grouping of all tokens, where gray represents representative tokens, orange represents unique tokens, and teal represents normal tokens. The third row displays the matching results between normal tokens and representative tokens. Different pairs are indicated by boxes of varying colors. The final row showcases the injected results." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 133, + 396, + 297, + 496 + ], + "blocks": [ + { + "bbox": [ + 133, + 396, + 297, + 496 + ], + "lines": [ + { + "bbox": [ + 133, + 396, + 297, + 496 + ], + "spans": [ + { + "bbox": [ + 133, + 396, + 297, + 496 + ], + "type": "image", + "image_path": "76f898105d1f49a683be9f26cf0c238ab46e5cc10ff2225a68345de9f34884fc.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 176, + 498, + 254, + 508 + ], + "lines": [ + { + "bbox": [ + 176, + 498, + 254, + 508 + ], + "spans": [ + { + "bbox": [ + 176, + 498, + 254, + 508 + ], + "type": "text", + "content": "(a) compressed token" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 316, + 396, + 481, + 496 + ], + "blocks": [ + { + "bbox": [ + 316, + 396, + 481, + 496 + ], + "lines": [ + { + "bbox": [ + 316, + 396, + 481, + 496 + ], + "spans": [ + { + "bbox": [ + 316, + 396, + 481, + 496 + ], + "type": "image", + "image_path": "e6c3b523fc076bd87025a383466d5652519ba9db59ff60e17a011cf8eb9ee0c1.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 368, + 498, + 429, + 508 + ], + "lines": [ + { + "bbox": [ + 368, + 498, + 429, + 508 + ], + "spans": [ + { + "bbox": [ + 368, + 498, + 429, + 508 + ], + "type": "text", + "content": "(b) unique token" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 516, + 482, + 550 + ], + "lines": [ + { + "bbox": [ + 130, + 516, + 482, + 550 + ], + "spans": [ + { + "bbox": [ + 130, + 516, + 482, + 550 + ], + "type": "text", + "content": "Fig.6: Ablation studies on the ratio of rows and columns of tokens. (a) Ratio ablation of compressed rows or columns " + }, + { + "bbox": [ + 130, + 516, + 482, + 550 + ], + "type": "inline_equation", + "content": "(r_h" + }, + { + "bbox": [ + 130, + 516, + 482, + 550 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 516, + 482, + 550 + ], + "type": "inline_equation", + "content": "r_w)" + }, + { + "bbox": [ + 130, + 516, + 482, + 550 + ], + "type": "text", + "content": ". (b) Ablation of retention ratios for unique token rows or columns." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "content": "To obtain a reduced token sequence, we employ several methods, as shown in Tab. 4d. Keep denotes the preservation of representative tokens and unique tokens without merging normal tokens. Max pool and avg pool refer to pooling-based methods for token fusion, while bilinear represents resizing token sequence through bilinear interpolation. Average indicates averaging tokens within each matched group to merge them. Keep and average do not rely on pre-defined priors to adjust the tokens, hence yielding better results. However, keep leads" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 385, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 385, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 385, + 91, + 447, + 102 + ], + "type": "text", + "content": "Token Adapter" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 144, + 178, + 475, + 258 + ], + "blocks": [ + { + "bbox": [ + 130, + 114, + 482, + 170 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 482, + 170 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 482, + 170 + ], + "type": "text", + "content": "Table 4: Token Adapter ablation experiments using Segmenter + ViT/L-16 on ADE20K validation dataset with official checkpoint and no retraining. " + }, + { + "bbox": [ + 130, + 114, + 482, + 170 + ], + "type": "inline_equation", + "content": "r_h" + }, + { + "bbox": [ + 130, + 114, + 482, + 170 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 114, + 482, + 170 + ], + "type": "inline_equation", + "content": "r_w" + }, + { + "bbox": [ + 130, + 114, + 482, + 170 + ], + "type": "text", + "content": " are set to 0.3. Segmenter + ViT/L-16 without Token Adapter achieves 51.82% mIoU at 10.68 img/s. We provide the mean Intersection-over-Union (mIoU) and model inference speed (img/s)." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 144, + 178, + 475, + 258 + ], + "lines": [ + { + "bbox": [ + 144, + 178, + 475, + 258 + ], + "spans": [ + { + "bbox": [ + 144, + 178, + 475, + 258 + ], + "type": "table", + "html": "
SelectionmIoUFPS
random51.3815.16
alternating51.4314.96
scores (feat.)52.0215.09
scores (attn.)51.1315.04
(a) Token Selection.
(b) Distance Metric.
(c) Feature Choice.
", + "image_path": "28177be2e3ae1ad6f577eba021c481e991ce17f2766c274130c554eaa4f5ac42.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 179, + 274, + 276, + 354 + ], + "blocks": [ + { + "bbox": [ + 179, + 274, + 276, + 354 + ], + "lines": [ + { + "bbox": [ + 179, + 274, + 276, + 354 + ], + "spans": [ + { + "bbox": [ + 179, + 274, + 276, + 354 + ], + "type": "table", + "html": "
MethodmIoUFPS
keep50.6715.19
max pool48.1514.97
avg pool48.1015.10
bilinear50.0815.12
average52.0215.09
", + "image_path": "00179f96dce25aa7d61948baa38a2ab90fb6102c253fd729b8cfc0aacaef087e.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 323, + 274, + 432, + 321 + ], + "blocks": [ + { + "bbox": [ + 188, + 355, + 266, + 365 + ], + "lines": [ + { + "bbox": [ + 188, + 355, + 266, + 365 + ], + "spans": [ + { + "bbox": [ + 188, + 355, + 266, + 365 + ], + "type": "text", + "content": "(d) Merge Method." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 323, + 274, + 432, + 321 + ], + "lines": [ + { + "bbox": [ + 323, + 274, + 432, + 321 + ], + "spans": [ + { + "bbox": [ + 323, + 274, + 432, + 321 + ], + "type": "table", + "html": "
ModemIoUFPS
max51.9315.13
m-threshold52.0215.09
", + "image_path": "be24dc1a9fa078c2d3a5f06aa0fc649c57a6eb13431cec720743aaed28ce3c26.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 337, + 322, + 418, + 332 + ], + "lines": [ + { + "bbox": [ + 337, + 322, + 418, + 332 + ], + "spans": [ + { + "bbox": [ + 337, + 322, + 418, + 332 + ], + "type": "text", + "content": "(e) Matching Mode." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 130, + 399, + 480, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 399, + 480, + 422 + ], + "spans": [ + { + "bbox": [ + 130, + 399, + 480, + 422 + ], + "type": "text", + "content": "to the loss of information from normal tokens, thus we ultimately opt for using average to merge the tokens." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 425, + 482, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 425, + 482, + 496 + ], + "spans": [ + { + "bbox": [ + 130, + 425, + 482, + 496 + ], + "type": "text", + "content": "Restoration Method. Considering dependency on the correlation between the reduced sequence and its original counterpart to restore the sequence length, it is imperative to contemplate the selection of features utilized in calculating this relationship matrix. Within Tab. 4c, the superscript " + }, + { + "bbox": [ + 130, + 425, + 482, + 496 + ], + "type": "inline_equation", + "content": "u" + }, + { + "bbox": [ + 130, + 425, + 482, + 496 + ], + "type": "text", + "content": " designates unmerged features, while " + }, + { + "bbox": [ + 130, + 425, + 482, + 496 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 130, + 425, + 482, + 496 + ], + "type": "text", + "content": " signifies merged features. The variable " + }, + { + "bbox": [ + 130, + 425, + 482, + 496 + ], + "type": "inline_equation", + "content": "X_{p}" + }, + { + "bbox": [ + 130, + 425, + 482, + 496 + ], + "type": "text", + "content": " represents representative tokens, while " + }, + { + "bbox": [ + 130, + 425, + 482, + 496 + ], + "type": "inline_equation", + "content": "X_{r}" + }, + { + "bbox": [ + 130, + 425, + 482, + 496 + ], + "type": "text", + "content": " pertains to reduced tokens." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 497, + 482, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 497, + 482, + 593 + ], + "spans": [ + { + "bbox": [ + 130, + 497, + 482, + 593 + ], + "type": "text", + "content": "To facilitate understanding of the meaning of each symbol, we next summarize the algorithmic procedures of the token adapter, and omit the calculation process. For " + }, + { + "bbox": [ + 130, + 497, + 482, + 593 + ], + "type": "inline_equation", + "content": "X \\in \\mathbb{R}^{N \\times C}" + }, + { + "bbox": [ + 130, + 497, + 482, + 593 + ], + "type": "text", + "content": ", we first partition tokens into " + }, + { + "bbox": [ + 130, + 497, + 482, + 593 + ], + "type": "inline_equation", + "content": "X_{p} \\in \\mathbb{R}^{N_{p} \\times C}" + }, + { + "bbox": [ + 130, + 497, + 482, + 593 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 497, + 482, + 593 + ], + "type": "inline_equation", + "content": "X_{n} \\in \\mathbb{R}^{N_{n} \\times C}" + }, + { + "bbox": [ + 130, + 497, + 482, + 593 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 497, + 482, + 593 + ], + "type": "inline_equation", + "content": "X_{u} \\in \\mathbb{R}^{N_{u} \\times C}" + }, + { + "bbox": [ + 130, + 497, + 482, + 593 + ], + "type": "text", + "content": ". Here, we haven't merged " + }, + { + "bbox": [ + 130, + 497, + 482, + 593 + ], + "type": "inline_equation", + "content": "X_{n}" + }, + { + "bbox": [ + 130, + 497, + 482, + 593 + ], + "type": "text", + "content": " into " + }, + { + "bbox": [ + 130, + 497, + 482, + 593 + ], + "type": "inline_equation", + "content": "X_{p}" + }, + { + "bbox": [ + 130, + 497, + 482, + 593 + ], + "type": "text", + "content": " yet, so we can mark it as " + }, + { + "bbox": [ + 130, + 497, + 482, + 593 + ], + "type": "inline_equation", + "content": "X_{p}^{u}" + }, + { + "bbox": [ + 130, + 497, + 482, + 593 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 130, + 497, + 482, + 593 + ], + "type": "inline_equation", + "content": "X_{r}" + }, + { + "bbox": [ + 130, + 497, + 482, + 593 + ], + "type": "text", + "content": " is formed by combining " + }, + { + "bbox": [ + 130, + 497, + 482, + 593 + ], + "type": "inline_equation", + "content": "X_{p}^{u}" + }, + { + "bbox": [ + 130, + 497, + 482, + 593 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 497, + 482, + 593 + ], + "type": "inline_equation", + "content": "X_{u}" + }, + { + "bbox": [ + 130, + 497, + 482, + 593 + ], + "type": "text", + "content": ", representing the compressed sequence without merging. After merging operations, the new " + }, + { + "bbox": [ + 130, + 497, + 482, + 593 + ], + "type": "inline_equation", + "content": "X_{p}" + }, + { + "bbox": [ + 130, + 497, + 482, + 593 + ], + "type": "text", + "content": " is marked as " + }, + { + "bbox": [ + 130, + 497, + 482, + 593 + ], + "type": "inline_equation", + "content": "X_{p}^{m}" + }, + { + "bbox": [ + 130, + 497, + 482, + 593 + ], + "type": "text", + "content": " and the new " + }, + { + "bbox": [ + 130, + 497, + 482, + 593 + ], + "type": "inline_equation", + "content": "X_{r}" + }, + { + "bbox": [ + 130, + 497, + 482, + 593 + ], + "type": "text", + "content": " is marked as " + }, + { + "bbox": [ + 130, + 497, + 482, + 593 + ], + "type": "inline_equation", + "content": "X_{r}^{m}" + }, + { + "bbox": [ + 130, + 497, + 482, + 593 + ], + "type": "text", + "content": ". From the results, it shows that recovery using the merged and reduced token sequence " + }, + { + "bbox": [ + 130, + 497, + 482, + 593 + ], + "type": "inline_equation", + "content": "X_{r}^{m}" + }, + { + "bbox": [ + 130, + 497, + 482, + 593 + ], + "type": "text", + "content": " works best." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 594, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 665 + ], + "type": "text", + "content": "A comparison result is presented for the two matching modes applied to the token ejector in Tab. 4e. In Sec. 3.2, we introduced the token ejector, which matches the most similar tokens for each token in the original token sequence according to the distance matrix " + }, + { + "bbox": [ + 130, + 594, + 482, + 665 + ], + "type": "inline_equation", + "content": "M_{d} \\in \\mathbb{R}^{N \\times N_{r}}" + }, + { + "bbox": [ + 130, + 594, + 482, + 665 + ], + "type": "text", + "content": ". Next, the token vectors from each matched group are averaged to reconstruct the token sequence to its original size. Mode max stands for matching the most similar token for each token of" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 214, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 214, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 214, + 101 + ], + "type": "text", + "content": "D. Su et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 187 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 187 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 187 + ], + "type": "text", + "content": "the original token sequence, i.e., one-to-one matching. Mode " + }, + { + "bbox": [ + 130, + 116, + 482, + 187 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 130, + 116, + 482, + 187 + ], + "type": "text", + "content": "-threshold, on the other hand, involves filtering based on a set threshold and distance matrix after Mode " + }, + { + "bbox": [ + 130, + 116, + 482, + 187 + ], + "type": "inline_equation", + "content": "max" + }, + { + "bbox": [ + 130, + 116, + 482, + 187 + ], + "type": "text", + "content": " matching. The filtered matching results are combined with the results from Mode " + }, + { + "bbox": [ + 130, + 116, + 482, + 187 + ], + "type": "inline_equation", + "content": "max" + }, + { + "bbox": [ + 130, + 116, + 482, + 187 + ], + "type": "text", + "content": ", followed by executing the corresponding fusion process. The results indicate that the latter achieves more accurate recovery of the token sequence than the former." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 189, + 482, + 248 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 189, + 482, + 248 + ], + "spans": [ + { + "bbox": [ + 130, + 189, + 482, + 248 + ], + "type": "text", + "content": "We also conduct ablation experiments on the ratio of compressed rows, columns, and unique tokens, and the results are shown in Fig. 6. Although we determined the final parameter settings for our method based on the ablation experiments, we believe that there still remains a better solution. It will be discussed in the next section." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 269, + 312, + 281 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 269, + 312, + 281 + ], + "spans": [ + { + "bbox": [ + 131, + 269, + 312, + 281 + ], + "type": "text", + "content": "5 Discussion and Conclusion" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 296, + 214, + 307 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 296, + 214, + 307 + ], + "spans": [ + { + "bbox": [ + 131, + 296, + 214, + 307 + ], + "type": "text", + "content": "5.1 Limitation" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 317, + 481, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 317, + 481, + 365 + ], + "spans": [ + { + "bbox": [ + 130, + 317, + 481, + 365 + ], + "type": "text", + "content": "The experimental results mentioned earlier have already demonstrated the effectiveness of our method on dense prediction tasks for hierarchical transformer structures and ViT-like structures. However, there are still many aspects of our work that remain to be explored, which we will analyze one by one next." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 365, + 482, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 365, + 482, + 496 + ], + "spans": [ + { + "bbox": [ + 130, + 365, + 482, + 496 + ], + "type": "text", + "content": "Generalizability. While our method has been validated for effectiveness in multiple segmentation tasks, it has not been validated on other more complex vision tasks. Previous methods essentially involved sparse sampling of token sequences, resulting in compressed token sequences being in sequence form only. The compressed sequence features can only be processed by a sequence processing module such as Transformer. Nevertheless, it remains to be examined whether discrete sparse sequence features are more effective than dense features for image tasks. Motivated by this, we designed the token adapter, a method that performs sparse sampling on the rows and columns of image features. Compared to previous methods, our approach yields more compact sampled features, which may be more effective for image tasks." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 498, + 482, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 498, + 482, + 568 + ], + "spans": [ + { + "bbox": [ + 130, + 498, + 482, + 568 + ], + "type": "text", + "content": "Adaptability. Although our core idea is rather simple, the Token Adapter still takes time to tune the hyper-parameters. The current setting allows Token Adapter to be directly applied to existing models, but it also limits its further development. We chose not to retrain the model equipped with the token adapter to directly evaluate the effectiveness of our idea. In the future, we will continue to explore more efficient neural network acceleration schemes." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "text", + "content": "Our core contribution is to compress and reconstruct rows and columns of image features in a way that conforms to the structure of the image. When applied to vision models, Token Adapter only compresses and reconstructs image features and thus can be applied to most vision models. For large language models, the concept of a Token Adapter can still be effectively transferred. By decomposing text embeddings based on the contextual information, and then leveraging the connections between the decomposed parts to compress them, it should be able to improve the efficiency of large language models." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 385, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 385, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 385, + 91, + 447, + 102 + ], + "type": "text", + "content": "Token Adapter" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 217, + 128 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 217, + 128 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 217, + 128 + ], + "type": "text", + "content": "5.2 Conclusion" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 138, + 482, + 305 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 138, + 482, + 305 + ], + "spans": [ + { + "bbox": [ + 130, + 138, + 482, + 305 + ], + "type": "text", + "content": "In this work, we propose Token Adapter to expedite vision transformers used in dense prediction tasks. By injecting normal tokens into representative tokens, we enhance the efficiency of vision transformers. Notably, our method can be applied in off-the-shelf models directly without additional retraining or finetuning. Tab. 2 reveals that when applied to the early stages of MaskDINO, the clustering-based Expediting method results in a significant decline in accuracy, and the matching-based Token Adapter method demonstrates superior stability. Clustering algorithms typically require fine-tuning of multiple hyper-parameters, which remain fixed once adjusted. Our approach is expected to leverage mechanisms such as cross-attention to transform fixed hyper-parameters into dynamically learnable parameters, thereby enhancing the method's generalization capability. Theoretically, the design of the compression strategy and reconstruction strategy in our algorithm allows the token adapter to be inserted in the middle of any module designed for image tasks." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 307, + 482, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 307, + 482, + 450 + ], + "spans": [ + { + "bbox": [ + 130, + 307, + 482, + 450 + ], + "type": "text", + "content": "In principle, Token Adapter bears resemblance to an unparameterized form of cross-attention mechanism, both blending information from one set of tokens into another set. Our thought lies in how we employ tokens to represent semantic entities within the image. Each entity in the image can be divided into distinct local semantic regions. By assigning several representative tokens to each local semantic region and injecting the remaining normal tokens into their corresponding tokens, we achieve a more refined fusion. After the fusion, the representative tokens representing different semantic regions will engage in information exchange within transformers, greatly enhancing computational efficiency and promoting diversity among tokens. We hope our research can offer a fresh perspective on image features and inspire the design of more efficient methods for extracting image features." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 471, + 198, + 483 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 471, + 198, + 483 + ], + "spans": [ + { + "bbox": [ + 132, + 471, + 198, + 483 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 136, + 498, + 481, + 666 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 136, + 498, + 481, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 498, + 481, + 520 + ], + "spans": [ + { + "bbox": [ + 136, + 498, + 481, + 520 + ], + "type": "text", + "content": "1. Bian, Z., Wang, Z., Han, W., Wang, K.: Multi-scale and token merging: Make your vit more efficient. arXiv preprint arXiv:2306.04897 (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 136, + 521, + 481, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 521, + 481, + 553 + ], + "spans": [ + { + "bbox": [ + 136, + 521, + 481, + 553 + ], + "type": "text", + "content": "2. Bolya, D., Fu, C.Y., Dai, X., Zhang, P., Feichtenhofer, C., Hoffman, J.: Token merging: Your vit but faster. In: The Eleventh International Conference on Learning Representations (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 136, + 554, + 481, + 576 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 554, + 481, + 576 + ], + "spans": [ + { + "bbox": [ + 136, + 554, + 481, + 576 + ], + "type": "text", + "content": "3. Bonnaerens, M., Dambre, J.: Learned thresholds token merging and pruning for vision transformers. arXiv preprint arXiv:2307.10780 (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 136, + 577, + 481, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 577, + 481, + 609 + ], + "spans": [ + { + "bbox": [ + 136, + 577, + 481, + 609 + ], + "type": "text", + "content": "4. Carion, N., Massa, F., Synnaeve, G., Usunier, N., Kirillov, A., Zagoruyko, S.: End-to-end object detection with transformers. In: European conference on computer vision. pp. 213-229. Springer (2020)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 136, + 610, + 481, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 610, + 481, + 632 + ], + "spans": [ + { + "bbox": [ + 136, + 610, + 481, + 632 + ], + "type": "text", + "content": "5. Chen, Z., Duan, Y., Wang, W., He, J., Lu, T., Dai, J., Qiao, Y.: Vision transformer adapter for dense predictions. arXiv preprint arXiv:2205.08534 (2022)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 136, + 632, + 481, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 632, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 136, + 632, + 481, + 666 + ], + "type": "text", + "content": "6. Choromanski, K., Likhosherstov, V., Dohan, D., Song, X., Gane, A., Sarlos, T., Hawkins, P., Davis, J., Mohiuddin, A., Kaiser, L., et al.: Rethinking attention with performers. arXiv preprint arXiv:2009.14794 (2020)" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 214, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 214, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 214, + 101 + ], + "type": "text", + "content": "D. Su et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 138, + 116, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 138, + 116, + 481, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 116, + 481, + 160 + ], + "spans": [ + { + "bbox": [ + 138, + 116, + 481, + 160 + ], + "type": "text", + "content": "7. Cordts, M., Omran, M., Ramos, S., Rehfeld, T., Enzweiler, M., Benenson, R., Franke, U., Roth, S., Schiele, B.: The cityscapes dataset for semantic urban scene understanding. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 3213-3223 (2016)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 160, + 481, + 192 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 160, + 481, + 192 + ], + "spans": [ + { + "bbox": [ + 138, + 160, + 481, + 192 + ], + "type": "text", + "content": "8. Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: Imagenet: A large-scale hierarchical image database. In: 2009 IEEE conference on computer vision and pattern recognition. pp. 248-255. IEEE (2009)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 192, + 481, + 236 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 192, + 481, + 236 + ], + "spans": [ + { + "bbox": [ + 138, + 192, + 481, + 236 + ], + "type": "text", + "content": "9. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929 (2020)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 236, + 481, + 278 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 236, + 481, + 278 + ], + "spans": [ + { + "bbox": [ + 138, + 236, + 481, + 278 + ], + "type": "text", + "content": "0. Fayyaz, M., Koohpayegani, S.A., Jafari, F.R., Sengupta, S., Joze, H.R.V., Sommerlade, E., Pirsiavash, H., Gall, J.: Adaptive token sampling for efficient vision transformers. In: European Conference on Computer Vision. pp. 396-414. Springer (2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 278, + 481, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 278, + 481, + 300 + ], + "spans": [ + { + "bbox": [ + 138, + 278, + 481, + 300 + ], + "type": "text", + "content": "1. Geng, Z., Guo, M.H., Chen, H., Li, X., Wei, K., Lin, Z.: Is attention better than matrix decomposition? arXiv preprint arXiv:2109.04553 (2021)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 300, + 481, + 332 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 300, + 481, + 332 + ], + "spans": [ + { + "bbox": [ + 138, + 300, + 481, + 332 + ], + "type": "text", + "content": "2. Guo, M.H., Liu, Z.N., Mu, T.J., Hu, S.M.: Beyond self-attention: External attention using two linear layers for visual tasks. IEEE Transactions on Pattern Analysis and Machine Intelligence 45(5), 5436-5447 (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 332, + 481, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 332, + 481, + 364 + ], + "spans": [ + { + "bbox": [ + 138, + 332, + 481, + 364 + ], + "type": "text", + "content": "3. Hassani, A., Walton, S., Li, J., Li, S., Shi, H.: Neighborhood attention transformer. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6185-6194 (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 364, + 481, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 364, + 481, + 396 + ], + "spans": [ + { + "bbox": [ + 138, + 364, + 481, + 396 + ], + "type": "text", + "content": "4. He, K., Chen, X., Xie, S., Li, Y., Dólar, P., Girshick, R.: Masked autoencoders are scalable vision learners. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 16000-16009 (2022)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 396, + 481, + 429 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 396, + 481, + 429 + ], + "spans": [ + { + "bbox": [ + 138, + 396, + 481, + 429 + ], + "type": "text", + "content": "5. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., et al.: Segment anything. arXiv preprint arXiv:2304.02643 (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 429, + 481, + 460 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 429, + 481, + 460 + ], + "spans": [ + { + "bbox": [ + 138, + 429, + 481, + 460 + ], + "type": "text", + "content": "6. Kong, Z., Dong, P., Ma, X., Meng, X., Sun, M., Niu, W., Shen, X., Yuan, G., Ren, B., Qin, M., et al.: Spvit: Enabling faster vision transformers via soft token pruning. arXiv preprint arXiv:2112.13890 (2021)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 460, + 481, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 460, + 481, + 503 + ], + "spans": [ + { + "bbox": [ + 138, + 460, + 481, + 503 + ], + "type": "text", + "content": "7. Li, F., Zhang, H., Xu, H., Liu, S., Zhang, L., Ni, L.M., Shum, H.Y.: Mask dino: Towards a unified transformer-based framework for object detection and segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3041-3050 (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 503, + 481, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 503, + 481, + 536 + ], + "spans": [ + { + "bbox": [ + 138, + 503, + 481, + 536 + ], + "type": "text", + "content": "8. Li, Y., Yuan, G., Wen, Y., Hu, J., Evangelidis, G., Tulyakov, S., Wang, Y., Ren, J.: Efficientformer: Vision transformers at mobilenet speed. Advances in Neural Information Processing Systems 35, 12934-12949 (2022)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 536, + 481, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 536, + 481, + 578 + ], + "spans": [ + { + "bbox": [ + 138, + 536, + 481, + 578 + ], + "type": "text", + "content": "9. Liang, W., Yuan, Y., Ding, H., Luo, X., Lin, W., Jia, D., Zhang, Z., Zhang, C., Hu, H.: Expediting large-scale vision transformer for dense prediction without fine-tuning. Advances in Neural Information Processing Systems 35, 35462-35477 (2022)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 578, + 481, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 578, + 481, + 622 + ], + "spans": [ + { + "bbox": [ + 138, + 578, + 481, + 622 + ], + "type": "text", + "content": "20. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollar, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13. pp. 740-755. Springer (2014)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 622, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 622, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 622, + 481, + 665 + ], + "type": "text", + "content": "21. Liu, Z., Hu, H., Lin, Y., Yao, Z., Xie, Z., Wei, Y., Ning, J., Cao, Y., Zhang, Z., Dong, L., et al.: Swin transformer v2: Scaling up capacity and resolution. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 12009-12019 (2022)" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 385, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 385, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 385, + 91, + 447, + 102 + ], + "type": "text", + "content": "Token Adapter" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 666 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 481, + 160 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 481, + 160 + ], + "type": "text", + "content": "22. Liu, Z., Lin, Y., Cao, Y., Hu, H., Wei, Y., Zhang, Z., Lin, S., Guo, B.: Swin transformer: Hierarchical vision transformer using shifted windows. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 10012-10022 (2021)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 160, + 481, + 203 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 160, + 481, + 203 + ], + "spans": [ + { + "bbox": [ + 132, + 160, + 481, + 203 + ], + "type": "text", + "content": "23. Long, S., Zhao, Z., Pi, J., Wang, S., Wang, J.: Beyond attentive tokens: Incorporating token importance and diversity for efficient vision transformers. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10334-10343 (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 203, + 481, + 236 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 203, + 481, + 236 + ], + "spans": [ + { + "bbox": [ + 132, + 203, + 481, + 236 + ], + "type": "text", + "content": "24. Lu, J., Yao, J., Zhang, J., Zhu, X., Xu, H., Gao, W., Xu, C., Xiang, T., Zhang, L.: Soft: Softmax-free transformer with linear complexity. Advances in Neural Information Processing Systems 34, 21297-21309 (2021)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 236, + 481, + 280 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 236, + 481, + 280 + ], + "spans": [ + { + "bbox": [ + 132, + 236, + 481, + 280 + ], + "type": "text", + "content": "25. Mottaghi, R., Chen, X., Liu, X., Cho, N.G., Lee, S.W., Fidler, S., Urtasun, R., Yuille, A.: The role of context for object detection and semantic segmentation in the wild. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 891-898 (2014)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 280, + 481, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 280, + 481, + 300 + ], + "spans": [ + { + "bbox": [ + 132, + 280, + 481, + 300 + ], + "type": "text", + "content": "26. Peng, H., Pappas, N., Yogatama, D., Schwartz, R., Smith, N.A., Kong, L.: Random feature attention. arXiv preprint arXiv:2103.02143 (2021)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 300, + 481, + 344 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 300, + 481, + 344 + ], + "spans": [ + { + "bbox": [ + 132, + 300, + 481, + 344 + ], + "type": "text", + "content": "27. Singh, M., Gustafson, L., Adcock, A., de Freitas Reis, V., Gedik, B., Kosaraju, R.P., Mahajan, D., Girshick, R., Dollár, P., Van Der Maaten, L.: Revisiting weakly supervised pre-training of visual perception models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 804-814 (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 343, + 481, + 375 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 343, + 481, + 375 + ], + "spans": [ + { + "bbox": [ + 132, + 343, + 481, + 375 + ], + "type": "text", + "content": "28. Strudel, R., Garcia, R., Laptev, I., Schmid, C.: Segmenter: Transformer for semantic segmentation. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 7262-7272 (2021)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 375, + 481, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 375, + 481, + 407 + ], + "spans": [ + { + "bbox": [ + 132, + 375, + 481, + 407 + ], + "type": "text", + "content": "29. Tian, K., Jiang, Y., Dao, Q., Lin, C., Wang, L., Yuan, Z.: Designing bert for convolutional networks: Sparse and hierarchical masked modeling. arXiv preprint arXiv:2301.03580 (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 407, + 481, + 450 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 407, + 481, + 450 + ], + "spans": [ + { + "bbox": [ + 132, + 407, + 481, + 450 + ], + "type": "text", + "content": "30. Tolstikhin, I.O., Houlsby, N., Kolesnikov, A., Beyer, L., Zhai, X., Unterthiner, T., Yung, J., Steiner, A., Keysers, D., Uszkoreit, J., et al.: Mlp-mixer: An allmlp architecture for vision. Advances in neural information processing systems 34, 24261-24272 (2021)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 450, + 481, + 483 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 450, + 481, + 483 + ], + "spans": [ + { + "bbox": [ + 132, + 450, + 481, + 483 + ], + "type": "text", + "content": "31. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. Advances in neural information processing systems 30 (2017)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 483, + 481, + 515 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 483, + 481, + 515 + ], + "spans": [ + { + "bbox": [ + 132, + 483, + 481, + 515 + ], + "type": "text", + "content": "32. Wang, Y., Huang, R., Song, S., Huang, Z., Huang, G.: Not all images are worth 16x16 words: Dynamic transformers for efficient image recognition. Advances in Neural Information Processing Systems 34, 11960-11973 (2021)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 515, + 481, + 548 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 515, + 481, + 548 + ], + "spans": [ + { + "bbox": [ + 132, + 515, + 481, + 548 + ], + "type": "text", + "content": "33. Wang, Z., Luo, H., Wang, P., Ding, F., Wang, F., Li, H.: Vtc-lfc: Vision transformer compression with low-frequency components. Advances in Neural Information Processing Systems 35, 13974-13988 (2022)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 548, + 481, + 579 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 548, + 481, + 579 + ], + "spans": [ + { + "bbox": [ + 132, + 548, + 481, + 579 + ], + "type": "text", + "content": "34. Yang, J., Li, C., Zhang, P., Dai, X., Xiao, B., Yuan, L., Gao, J.: Focal self-attention for local-global interactions in vision transformers. arXiv preprint arXiv:2107.00641 (2021)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 579, + 481, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 579, + 481, + 612 + ], + "spans": [ + { + "bbox": [ + 132, + 579, + 481, + 612 + ], + "type": "text", + "content": "35. Yang, R., Ma, H., Wu, J., Tang, Y., Xiao, X., Zheng, M., Li, X.: Scalablevit: Rethinking the context-oriented generalization of vision transformer. In: European Conference on Computer Vision. pp. 480-496. Springer (2022)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 612, + 481, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 612, + 481, + 633 + ], + "spans": [ + { + "bbox": [ + 132, + 612, + 481, + 633 + ], + "type": "text", + "content": "36. Ye, D., Lin, Y., Huang, Y., Sun, M.: Tr-bert: Dynamic token reduction for accelerating bert inference. arXiv preprint arXiv:2105.11618 (2021)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 633, + 481, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 633, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 633, + 481, + 666 + ], + "type": "text", + "content": "37. Yin, H., Vahdat, A., Alvarez, J.M., Mallya, A., Kautz, J., Molchanov, P.: A-vit: Adaptive tokens for efficient vision transformer. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10809-10818 (2022)" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 214, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 214, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 214, + 101 + ], + "type": "text", + "content": "D. Su et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 369 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "text", + "content": "38. Yu, Q., Xia, Y., Bai, Y., Lu, Y., Yuille, A.L., Shen, W.: Glance-and-gaze vision transformer. Advances in Neural Information Processing Systems 34, 12992-13003 (2021)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 150, + 482, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 150, + 482, + 183 + ], + "spans": [ + { + "bbox": [ + 130, + 150, + 482, + 183 + ], + "type": "text", + "content": "39. Yu, W., Luo, M., Zhou, P., Si, C., Zhou, Y., Wang, X., Feng, J., Yan, S.: Metaformer is actually what you need for vision. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10819-10829 (2022)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 183, + 482, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 183, + 482, + 226 + ], + "spans": [ + { + "bbox": [ + 130, + 183, + 482, + 226 + ], + "type": "text", + "content": "40. Zeng, W., Jin, S., Liu, W., Qian, C., Luo, P., Ouyang, W., Wang, X.: Not all tokens are equal: Human-centric visual analysis via token clustering transformer. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11101-11111 (2022)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 227, + 482, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 227, + 482, + 259 + ], + "spans": [ + { + "bbox": [ + 130, + 227, + 482, + 259 + ], + "type": "text", + "content": "41. Zeng, Z., Hawkins, C., Hong, M., Zhang, A., Pappas, N., Singh, V., Zheng, S.: Vcc: Scaling transformers to 128k tokens or more by prioritizing important tokens. arXiv preprint arXiv:2305.04241 (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 259, + 482, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 259, + 482, + 293 + ], + "spans": [ + { + "bbox": [ + 130, + 259, + 482, + 293 + ], + "type": "text", + "content": "42. Zheng, M., Gao, P., Zhang, R., Li, K., Wang, X., Li, H., Dong, H.: End-to-end object detection with adaptive clustering transformer. arXiv preprint arXiv:2011.09315 (2020)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 293, + 482, + 325 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 293, + 482, + 325 + ], + "spans": [ + { + "bbox": [ + 130, + 293, + 482, + 325 + ], + "type": "text", + "content": "43. Zhou, B., Zhao, H., Puig, X., Xiao, T., Fidler, S., Barriuso, A., Torralba, A.: Semantic understanding of scenes through the ade20k dataset. International Journal of Computer Vision 127, 302-321 (2019)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 325, + 482, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 325, + 482, + 369 + ], + "spans": [ + { + "bbox": [ + 130, + 325, + 482, + 369 + ], + "type": "text", + "content": "44. Ziwen, C., Patnaik, K., Zhai, S., Wan, A., Ren, Z., Schwing, A.G., Colburn, A., Fuxin, L.: Autofocusformer: Image segmentation off the grid. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18227-18236 (2023)" + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 385, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 385, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 385, + 91, + 447, + 102 + ], + "type": "text", + "content": "Token Adapter" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/RepVF_ A Unified Vector Fields Representation for Multi-task 3D Perception/9c289191-ecfe-49ca-a348-d2cad5492a03_content_list.json b/2024/RepVF_ A Unified Vector Fields Representation for Multi-task 3D Perception/9c289191-ecfe-49ca-a348-d2cad5492a03_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..7b26f3441de28e129b4503eb23a063e5cdfd1b5f --- /dev/null +++ b/2024/RepVF_ A Unified Vector Fields Representation for Multi-task 3D Perception/9c289191-ecfe-49ca-a348-d2cad5492a03_content_list.json @@ -0,0 +1,1774 @@ +[ + { + "type": "text", + "text": "RepVF: A Unified Vector Fields Representation for Multi-task 3D Perception", + "text_level": 1, + "bbox": [ + 228, + 140, + 776, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Chunliang Li $^{1*}$ , Wencheng Han $^{2*}$ , Junbo Yin $^{1}$ , Sanyuan Zhao $^{1}$ , and Jianbing Shen $^{2}$", + "bbox": [ + 313, + 210, + 687, + 244 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ School of Computer Science, Beijing Institute of Technology \n $^{2}$ SKL-IOTSC, Computer and Information Science, University of Macau, China \n{jbji, sanyuanzhao}@bit.edu.cn \n{wencheng256, yinjunbocn}@gmail.com \njianbingshen@um.edu.mo", + "bbox": [ + 233, + 253, + 766, + 324 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Concurrent processing of multiple autonomous driving 3D perception tasks within the same spatiotemporal scene poses a significant challenge, in particular due to the computational inefficiencies and feature competition between tasks when using traditional multi-task learning approaches. This paper addresses these issues by proposing a novel unified representation, RepVF, which harmonizes the representation of various perception tasks such as 3D object detection and 3D lane detection within a single framework. RepVF characterizes the structure of different targets in the scene through a vector field, enabling a single-head, multi-task learning model that significantly reduces computational redundancy and feature competition. Building upon RepVF, we introduce RFTR, a network designed to exploit the inherent connections between different tasks by utilizing a hierarchical structure of queries that implicitly model the relationships both between and within tasks. This approach eliminates the need for task-specific heads and parameters, fundamentally reducing the conflicts inherent in traditional multi-task learning paradigms. We validate our approach by combining labels from the OpenLane dataset with the Waymo Open dataset. Our work presents a significant advancement in the efficiency and effectiveness of multi-task perception in autonomous driving, offering a new perspective on handling multiple 3D perception tasks synchronously and in parallel. The code will be available at: https://github.com/jbji/RepVF.", + "bbox": [ + 261, + 348, + 743, + 654 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: 3D Lane Detection $\\cdot$ 3D Object Detection $\\cdot$ Multi-task Method.", + "bbox": [ + 261, + 666, + 758, + 681 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 733, + 375, + 750 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In autonomous driving, multiple 3D perception tasks often need to be processed synchronously, in real-time, and in parallel. These tasks, which exist within the", + "bbox": [ + 212, + 760, + 787, + 791 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal contribution. †Corresponding author. This work was supported in part by the FDCT grants 0102/2023/RIA2, 0154/2022/A3, and 001/2024/SKL, the MYRG-CRG2022-00013-IOTSC-ICI grant and the SRG2022-00023-IOTSC grant.", + "bbox": [ + 230, + 797, + 787, + 839 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/2bfbbd98ca04d688ded301c786f4503e6fd2dcc793e811e4c5a6dc7ca31f9204.jpg", + "image_caption": [ + "(a) Multi-tasking paradigm comparison." + ], + "image_footnote": [], + "bbox": [ + 243, + 152, + 500, + 313 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/c98f0f250df4db5e11794a7c0ede71ed5720cd8dd1abc95d2d1f2b098d52c4aa.jpg", + "image_caption": [ + "(b) $RepVF$ , conversion and supervision.", + "Fig. 1: $RepVF$ is proposed as a unified representation for 3D perception to achieve single-head multi-tasking. It consists of a set of vector fields learned to represent road elements, and is trained utilizing existing labels through differentiable conversions. The proposed single-head multi-tasking paradigm reduces task conflict and competition." + ], + "image_footnote": [], + "bbox": [ + 514, + 157, + 769, + 311 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "same spatiotemporal scene, are often complementary to each other in both geometric and semantic aspects [19]. For example, 3D object detection [12] and 3D lane detection [11] define foreground objects and background lanes within the same 3D scene. Vehicles usually travel following the guiding semantics of lanes that do not spatially overlap and are partially obscured by other vehicles. Some previous industrial solutions [1,2] utilize a multi-model paradigm for multi-task perceptions, as depicted in Figure 1a(1). These separate models cannot share common features across different tasks, resulting in computational waste. Furthermore, this paradigm does not take advantage of the inherent connections between different tasks. Some works [19,28,38,39] attempt to overcome this limitation by using a common backbone and separate head networks for each task, as shown in Figure 1a(2). This approach can considerably reduce the computational cost when handling a large number of tasks, but it might lead to a task balancing issue. Since different task-aware heads focus on various aspects of the scene, they may compete for the backbone features, resulting in instability during training [19,39].", + "bbox": [ + 212, + 430, + 787, + 672 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The competition is attributed to the variation in representation between different tasks. For example, the 3D detection task [12,26,28,37,38,60,62] guides the head network to predict 3D bounding boxes, whereas the lane detection task [6,11] necessitates the corresponding head to regress lane shapes. Specifically, 3D bounding boxes describe a cubic area in 3D space, representing different category foregrounds using position, size, and orientation. In contrast, 3D lanes are represented using anchor [6,11,18] or parametric [15] methods to describe one or more 1-D lines. These representations focus on different aspects of the scenes, which ultimately leads to competition.", + "bbox": [ + 212, + 672, + 787, + 808 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we propose a unified representation, $RepVF$ (Representative Vector Fields), for perception tasks in autonomous driving. Using this represent", + "bbox": [ + 214, + 809, + 787, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "C. Li et al.", + "bbox": [ + 271, + 114, + 346, + 126 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "tation, we design a single-headed, unified model that can perform tasks simultaneously. As illustrated in Fig. 1b, $RepVF$ represents 3D vector fields that each assigns vectors to spatial locations, denoted as $(S, \\mathcal{F}(S)) \\subseteq \\mathbb{R}^{3 + d}$ in Sec. 3.2. Sub-vector fields representing different targets in the 3D perception scene progressively adhere to their spatial extent, accurately characterizing the anisotropic structure in their respective spatial vicinity. $RepVF$ can be differentiably transformed into task-specific elements, hence requiring no special supervision and can utilize existing labels. With $RepVF$ , our network only processes one single type of fundamental perception element. As shown in Fig. 1a, unlike the traditional multi-task approach that employs multiple task heads [19, 28, 30, 38, 72], we follow an unprecedented single-head multi-task paradigm. The single-head multi-task eliminates task-specific parameters and obviates the need to explicitly model task interactions.", + "bbox": [ + 212, + 146, + 787, + 343 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "RepVF representation serve as a geometrical, cross-scale 3D perception representation alternative. Unlike perception elements capable of handling only single-scale tasks [6,11,12,62], it breaks down the scene into smaller perceptual units while retaining the capability for target-level perception at larger scales. Based on the RepVF representation, we further propose RFTR to leverage this relationship, utilizing queries with a hierarchical structure that implicitly models the relationships both between and within different tasks. Following the single-head multi-task architecture, with only shared parameters across tasks, RFTR does not depend on existing multi-task optimization strategies [21,33,35,52,73]. The gradient discrepancies between tasks across different training iterations have been mitigated, fundamentally reducing the competition and conflict inherent in multi-tasking 3D perception. By employing RepVF to replace the fundamental elements for different tasks, and by combining 3D lane labels from the OpenLane dataset [6] with the Waymo Open dataset [53] for simultaneous multi-task training and inference, our approach achieves comparable or even better multi-task performance than single-task counterparts with only one unified model.", + "bbox": [ + 212, + 343, + 787, + 584 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In summary, we introduce $RepVF$ as a unified representation for multiple perception tasks in autonomous driving, effectively reducing the competition among different task representations during training and enhancing model convergence. We further develop a single-head multi-task framework RFTR based on $RepVF$ , fully exploiting the intrinsic connections among different tasks within a scene, improving feature interaction with queries, and enabling gradient-balanced multi-task training. Our approach achieves impressive performance across two key tasks in autonomous driving—3D object detection and 3D lane detection—showcasing the advantages of a unified task representation in the context of autonomous driving applications.", + "bbox": [ + 212, + 584, + 787, + 737 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 756, + 388, + 772 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.1 Camera-based 3D Perception in Autonomous Driving", + "text_level": 1, + "bbox": [ + 212, + 786, + 705, + 801 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3D object detection [7,9,12,62,70] and 3D lane detection [11,14] are two common 3D perception tasks, and previous approaches usually perform the two tasks", + "bbox": [ + 212, + 809, + 785, + 840 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "RepVF: Unified Multi-task 3D Perception", + "bbox": [ + 450, + 114, + 730, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "independently. 3D object detection has been extensively studied either in model architectures [7, 37, 38, 59] or feature representations [17, 25, 27, 28, 49, 50, 71, 76]. Early methods [20, 22, 45] predict 3D bounding boxes using 2D detectors, and Mono3D [7] is one of the first methods directly make predictions from 3D representations. Motivated by DETR [11], DETR3D [62] uses sparse 3D object centers as queries, and PETR [37] further enhances its feature representations by embedding 3D positions. Some recent approaches [17, 28, 38, 58] also focus on incorporating temporal information into 3D object detection, but they all follow the transformer-based [56] paradigm introduced by DETR3D [62].", + "bbox": [ + 212, + 146, + 787, + 282 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Only recently have DETR-based methods been introduced for 3D lane detection [3, 38, 42, 63, 64]. Previous methods [6, 11, 14, 34, 57] prefer to project image features into BEV space using Inverse Perspective Mapping (IPM), but this introduces prohibitive height prediction error. Some work [67] avoids IPM via depth estimation, and Anchor-3DLane [18] reverses this projection process. Following DETR [3, 38, 61] introduce 3d lane anchors as queries, and [42] adopts finer lane prior queries and is end-to-end. With the advances of DETR [61, 62] networks in both tasks [3, 28, 37, 42], it is architecturally ready to unify the two tasks without standalone models.", + "bbox": [ + 212, + 282, + 787, + 417 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.2 Multi-task 3D Perception", + "text_level": 1, + "bbox": [ + 215, + 446, + 472, + 462 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The problem of multi-tasking in autonomous driving perception is still in its infancy. Recent advances [19, 28, 37-39, 48, 65, 75] elaborate on universal perception features and multiple tasks are handled with different heads. Moreover, there is no consensus on the required tasks. For example, MMF [29] addresses both depth completion and object detection issues, while [19, 28, 38, 39] performs object detection and map segmentation, with [38] performing additional 3D lane detection separately. However, most of these studies only use multiple tasks to demonstrate the universality of their learned features, and the tasks are trained and evaluated separately [28, 38] rather than simultaneously [19, 39], not to mention dataset discrepancies.", + "bbox": [ + 212, + 476, + 787, + 626 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In addition, multitask learning (MTL) typically involves task conflict. Previous MTL studies [55] addresses this either through parameter sharing [13, 44, 47, 51, 54, 66] or optimization [8, 21, 33, 35, 52, 73]. Parameter sharing approaches are conceptually straightforward, and are further classified into hard-sharing (with shared and task-specific parameters) and soft-sharing (with a cross-task talk mechanism). Optimization-based approaches, e.g. IMTL [33] and DWA [35], attribute MTL imbalance to the gradient and loss imbalance and explore optimization calibration methods. Taking advantage of these advances, FULLER [19] stands as the pioneering work in autonomous driving perception that analyzes the performance of modern optimization-based MTL calibration methods [8, 33, 35]. With a unified task representation, we follow a different paradigm from previous ones, that eliminate task-specific heads and mitigate the gradient imbalance. Our approach can be seen as a hard-sharing technique without task-specific parameters and does not require calibration.", + "bbox": [ + 212, + 628, + 787, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "C. Li et al.", + "bbox": [ + 271, + 114, + 344, + 126 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3 Representative Vector Fields", + "text_level": 1, + "bbox": [ + 215, + 143, + 532, + 162 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We begin by revisiting the representations of two common 3D perception tasks: 3D object detection [4, 12, 53] and 3D lane detection [6, 67]. By exploring how task-specific representations have evolved into task-specific designs [3, 6, 7, 11, 37, 38, 42, 50, 62], we propose the task-agnostic $RepVF$ (Representative Vector Fields), which arise from the geometric commonality in the tasks and effectively reduce task competition via unified representations.", + "bbox": [ + 212, + 176, + 787, + 267 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1 Task-specific Representations and Single-tasking", + "text_level": 1, + "bbox": [ + 214, + 287, + 661, + 304 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Representation Formulations. Task-specific representations vary due to the geometric features of the task targets, which differ in dimensions and scales. As a key task for understanding 3D space, 3D object detection deals with objects irregularly shaped and scaled from one to several metres, defined by coarse outer boundaries as 3D bounding boxes [12], formally a 7-d coordinate $B = (x,y,z,l,w,h,\\theta)$ encoding object center $(x,y,z)$ , box dimensions $(l,w,h)$ and heading $\\theta$ . Conversely, due to their linear shape and ability to cover long distances, 3d lanes are not described by boundaries but by direct geometries, typically ordered sequences of $N_{L}$ 3D point coordinates [6]:", + "bbox": [ + 212, + 311, + 787, + 449 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL = \\left[ \\left(x _ {1}, y _ {1}, z _ {1}\\right), \\left(x _ {2}, y _ {2}, z _ {2}\\right), \\dots , \\left(x _ {N _ {l}}, z _ {N _ {l}}, z _ {N _ {l}}\\right) \\right] \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 334, + 459, + 785, + 476 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "This may look familiar to an unordered point cloud, but they are ordered along the direction of the road. In addition, 3D perception problems typically involve classifications that require category scores $\\mathcal{C}_i$ , varying from 1, 3 or 23 [4,12,53] classes (3D object) to 21 or 22 [6] classes (3D lane).", + "bbox": [ + 212, + 484, + 787, + 547 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Evolving into Task-specific Designs. Pioneering 3D object detection from RGB images is proposal based, e.g. Mono3D [7] reduces proposal space dimension of each class to $(x,y,z,\\theta,t)$ with representative templates, and only predicts $\\varDelta_{\\mathrm{pos}}$ ; OFT-Net [50] predicts offsets of all bounding box parameters relative to the top-view grid, i.e. $\\varDelta_{\\mathrm{pos}},\\varDelta_{\\mathrm{dim}},\\varDelta_{\\mathrm{ang}}$ . Taking advantage of the center coordinates, DETR-based methods [37,62] regress $\\varDelta_{\\mathrm{pos}}$ along with bounding box parameters. However, constrained by the limited spatial extent of center points, extracted features around them capture only local spatial areas [77], failing to capture the finer pose and structure of objects.", + "bbox": [ + 212, + 566, + 787, + 703 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3D lines are suitable for anchor-based representations [6, 11, 38], resulting in equally spaced longitudinal lines and anchor-based predictions $\\varDelta_{\\mathrm{xz}}$ . Anchor representations reduce the representation dimension, but also restrict the degrees of freedom, limiting the maximum model capacity, which has led some works [6, 18] focus on more sophisticated anchor representations. Recent DETR-based methods [3,38,42] use anchor coordinates as queries, exploring the full potential of anchors. Unfortunately, since it's not trivial and stable to directly predict parameter-driven curves [10, 15], mainstream methods stick to predicting anchor-based offsets $\\varDelta_{\\mathrm{xz}}$ , and nearly no one ever worked directly on the raw lane format.", + "bbox": [ + 212, + 704, + 787, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "RepVF: Unified Multi-task 3D Perception", + "bbox": [ + 450, + 114, + 730, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Intersection of Task-specific Representations. We believe that it's already been architecturally ready to unify the two (and possibly more) perception tasks, with the population and success of DETR practices [19,28,38,61]. However, they still rely on multi-head design due to the discrepancy in task representation, which consumes more computational resources. Despite the significant differences in the dynamics and scales of the described targets, these task-specific representations only encapsulate descriptions of shapes, structures, and directions (object heading and lane ordinality). It is plausible to represent these task targets universally, taking into account perception hierarchies.", + "bbox": [ + 212, + 146, + 787, + 282 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.2 The Task-Agnostic Representation through RepVF", + "text_level": 1, + "bbox": [ + 214, + 300, + 683, + 316 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To facilitate efficient design in multitasking frameworks, we introduce $RepVF$ , a vector fields representation inspired by 2D point set object representation [43,43]. Distinct from conventional isotropic points, $RepVF$ assigns a vector to each location making itself anisotropic. We define $RepVF$ as a collection of vector fields, each field denoted by $R_{i}$ as a perception target, whose semantic vector assignment mapping $\\mathcal{F}_i$ is defined on $D = \\bigcup_{i=1}^{n_s} S_i \\subseteq \\mathbb{R}^3$ :", + "bbox": [ + 212, + 321, + 787, + 412 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\{R _ {i} \\}, \\quad R _ {i} = \\left(S _ {i}, \\mathcal {F} _ {i} \\left(S _ {i}\\right)\\right) \\subseteq \\mathbb {R} ^ {3 + d}, \\quad \\mathcal {F}: D \\rightarrow \\mathbb {R} ^ {d} \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 330, + 417, + 784, + 435 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $d$ is the dimension of the target vector domain. Therefore, for any point $x$ within $D$ , $\\mathcal{F}$ assigns it a vector $\\mathcal{F}(x) \\in \\mathbb{R}^d$ that encapsulates dimensional and positional attributes of the point. This definition allows us to cover the whole space: The sampling domain $D$ is partitioned into $n_s$ disjoint subsets $S_i$ , with each subset containing $n$ representative sampling points: $S_i = \\{(x_k, y_k, z_k)\\}_{k=1}^n$ . Accordingly, the collective vector field $\\mathcal{F}$ sprawls over $D = \\bigcup_{i=1}^{n_s} S_i$ , articulated into $n_s$ sub-vector fields $\\mathcal{F}_i$ catering to specific tasks.", + "bbox": [ + 212, + 440, + 787, + 546 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Intuitively, $S_{i}$ directly outlines entity-level geometric structures, and $\\mathcal{F}_i(S_i)$ delineates contextual characteristics (i.e. $\\theta$ for bounding box $B$ , or adjacent point directions for lane $L$ ). Considering the directionality, we use unit-length vectors parallel to the x-y plane to capture angular behavior, resulting in behavior akin to scalars and $d = 1$ . To represent proper $\\mathcal{F}_i$ for 3d lanes, directions of the penultimate points are employed for terminal lane line points.", + "bbox": [ + 212, + 547, + 787, + 638 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Converting RepVF into Task-Specific Representation. To leverage existing task labels for training and evaluation, similar to object detection representations based on point sets [43, 68], we convert predictions $RepVF \\hat{R}_i$ , category $\\hat{C}_i$ to task-specific representations. We define geometric interpretation processes as differentiable, parameter-free transformation functions:", + "bbox": [ + 212, + 655, + 787, + 731 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {T} ^ {(\\cdot)}: \\hat {R} _ {i} \\rightarrow \\hat {P} _ {i}, \\quad \\text {s . t .} \\hat {P} _ {i} \\in \\{\\hat {B}, \\hat {L} \\} _ {i} \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 377, + 736, + 784, + 753 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Specifically, we have $\\mathcal{T}^l:\\hat{R}_i\\to \\hat{L}_i$ for 3d lanes and $\\mathcal{T}^b:\\hat{R}_i\\rightarrow \\hat{B}_i$ for 3d objects. For semantic parts, to align $\\hat{\\mathcal{C}}_i$ with task-specific classes, we split it into $n_{(\\cdot)}$ bins and take max score within each bin, where $n_{(\\cdot)}$ is the count of task categories. $\\hat{R}_i$ and $\\hat{R}_j$ remain formally equivalent until contextualized into $\\hat{P}_i$ and $\\hat{P}_j$ , thus maintaining task agnosticism.", + "bbox": [ + 212, + 758, + 787, + 842 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "C. Li et al.", + "bbox": [ + 271, + 114, + 346, + 127 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Conversion Function Choice. In our implementation, $\\mathcal{T}^l$ sorts the unordered points of $\\hat{S}_i$ by the road direction $(+X$ in Waymo [53]). Unlike 2D boxes, 3D boxes are not strictly aligned with to $XYZ$ axes due to headings. Therefore, $\\mathcal{T}^b$ consists three parts that respectively compute heading, box center and dimensions:", + "bbox": [ + 214, + 146, + 787, + 207 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {T} _ {h} ^ {b}: \\hat {\\mathcal {F}} _ {i} \\rightarrow \\hat {\\theta} _ {i}; \\quad \\mathcal {T} _ {x y z} ^ {b}: \\hat {S} _ {i} \\rightarrow (\\hat {x} _ {i}, \\hat {y} _ {i}, \\hat {z} _ {i}); \\quad \\mathcal {T} _ {l w h} ^ {b}: \\hat {S} _ {i} \\rightarrow (\\hat {l} _ {i}, \\hat {w} _ {i}, \\hat {h} _ {i}) \\qquad (4)\n$$\n", + "text_format": "latex", + "bbox": [ + 279, + 215, + 785, + 234 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Here, $\\mathcal{T}_h^b$ and $\\mathcal{T}_{xyz}^b$ are simple arithmetic averages. With $\\hat{\\theta}_i$ given by $\\mathcal{T}_h^b$ , by projecting $\\hat{S}_i$ to the main orientation axis $\\hat{\\theta}_i$ , the predictions are aligned with the predicted directions. We then proceed with two candidates of $\\mathcal{T}_{lwh}^b$ similar to 2D practices [68]: (1) $\\mathcal{T}_{lwh}^b = \\mathcal{T}_1$ , the min-max function, performing min/max operations projection dimensions for dimensions; (2) $\\mathcal{T}_{lwh}^b = \\mathcal{T}_2$ employs a momentum-based function taking the variance of predicted points as the boundary box dimensions. In our practice, $\\mathcal{T}_2$ works slightly better.", + "bbox": [ + 214, + 244, + 787, + 354 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Learning RepVF. RepVF is supervised by loss functions universally across tasks without special losses. Here we use $\\mathbf{L}^1$ regression loss on $S_{i}$ and $\\mathcal{F}_i$ , and focal loss [32] on $\\mathcal{C}_i$ . Smaller structures are described by points in $S_{i}$ , and $R_{i}$ itself is set-level, extending RepVF across different perception hierarchies. To capture this cross-level property, we adopt DETR-based structure [37,61,62] and aligned query design to use raw representations motivated by [3, 36]. With multiple decoder layers [62, 77], the learning process of RepVF can be characterized as iterative update $\\varOmega_l$ on queries $Q_{l}$ and inferring $\\hat{R}_l,\\hat{C}_l$ from them with linear layer $\\varPhi$ and decoder layer $l$ :", + "bbox": [ + 214, + 375, + 787, + 512 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\nQ _ {l} = \\Omega_ {l} \\left(Q _ {l - 1}\\right), \\quad \\hat {R} _ {l}, \\hat {\\mathcal {C}} _ {l} = \\Phi (Q _ {l}) \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 385, + 521, + 784, + 537 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We will explain our network in detail in the next section.", + "bbox": [ + 215, + 547, + 627, + 561 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4 RFTR: The Single-head Multi-tasker", + "text_level": 1, + "bbox": [ + 215, + 584, + 609, + 601 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this section, we introduce the single-head multi-task framework RFTR (Representative Vector Fields Transformer), which uses $RepVF$ representation to unify different tasks and exploit intrinsic connections among them. We also observed one interesting behavior of RFTR in the balance of multi-task gradients. RFTR is built upon DETR networks [5,37,62] and is end-to-end on all tasks.", + "bbox": [ + 214, + 614, + 787, + 689 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.1 Network Structure", + "text_level": 1, + "bbox": [ + 215, + 710, + 418, + 724 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Figure 2 shows our RFTR architecture built upon $RepVF$ . It inputs images $I$ from $N$ cameras with known camera parameters, and outputs unified representative vector fields $\\mathcal{F}$ for various perception tasks in the 3D scene. RFTR contains four main components that extract features from images, generate set-level queries for hierarchical representation structure, solve unified representative vector fields and multiple tasks with one single head, and convert results to task representations in a differentiable way to utilize existing task labels.", + "bbox": [ + 214, + 733, + 787, + 840 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "RepVF: Unified Multi-task 3D Perception", + "bbox": [ + 450, + 114, + 730, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/867f1b847c5b1f9e3c5f9394f6700d4d85a853583e8c6e4af81047ad62423d23.jpg", + "image_caption": [ + "Fig. 2: Overview of our RFTR (Representative Vector Fields Transformer) built upon RepVF. Multi-view image features with 3D position embeds are extracted by image backbone and fed into the decoder. We generate set-level query embeds from 3D space sampling sets, each representing a perception target. One single unified task head is then used to predict unified representative vector fields. Finally, predictions are transformed in a differentiable manner into task specific representations to utilize existing labels for supervision." + ], + "image_footnote": [], + "bbox": [ + 218, + 143, + 782, + 335 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Feature Extraction. Following common practices in utilizing 2D backbones to extract features [28,37,62], we adopt ResNet [16] (ResNet-50) and FPN [31] to produce 2D features $F^{2d}$ . They're encoded with 3D position embedding in PETR [37] into 3D position aware features $F^{3d}$ .", + "bbox": [ + 212, + 474, + 787, + 537 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Generating Set-level Perception Queries. As $RepVF$ is sampled from $n_s$ point sets $S_i$ with $n$ points each and is two-level, our aim is to generate queries that are consistent with this hierarchical point-set structure. Hence provide the decoder with sufficient spatial information, improving feature interactions. Unlike former approaches [37, 38, 42, 62] that encode one single 3D location, each query in RFTR encodes multiple locations simultaneously.", + "bbox": [ + 212, + 559, + 787, + 648 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We generate set-level queries $Q_0 \\in \\mathbb{R}^{n_s \\times d_q}$ from initial point sets corresponding to $S_i$ , where $d_q$ is the dimension of each query. The generation process $\\mathcal{G}$ consists of encoding initialized positions from the point level, and embedding them at the set level:", + "bbox": [ + 212, + 650, + 787, + 709 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {G} = \\operatorname {e m b e d} (\\text {F l a t t e n (P E)}) \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 406, + 710, + 785, + 727 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where PE is the positional encoding [56]. Point-level PE is flattened into set-level and embedded by a small MLP with two layers. To avoid confusing the model with expansive locations falling on multiple targets in the early stages of training, we use $n$ identical 3D locations sampled with 0-1 uniform distribution for all $n_s$ initial point sets.", + "bbox": [ + 212, + 733, + 787, + 809 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To our best knowledge, the closest generation approach to ours is [3], which generates query from an ordered sequence of anchor points motivated by [36].", + "bbox": [ + 212, + 809, + 787, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "C. Li et al.", + "bbox": [ + 271, + 114, + 346, + 126 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "However, ours is from an unordered sampling point set and is not task-specific. Moreover, we use a tiny 16-dimensional PE for each coordination, divert from former methods [37, 62] with a much larger 256-dimension. In our multi-task setting, queries are equivalently initialized, thus task-agnostic. We divide the first $80\\%$ as 3D objects and the last $20\\%$ as 3D lanes.", + "bbox": [ + 212, + 146, + 787, + 220 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Single-head Multi-tasking. Similar to all previous approaches [5, 37, 62], queries are learnable and updated in decoder layers iteratively. The unified head $\\varPhi$ of RFTR is composed by a multi-layer transformer decoder [56], and two branches $\\varPhi_{g}$ and $\\varPhi_{s}$ dedicated for geometric regression and semantic prediction respectively to predict $RepVF$ . We use the same $n_{L} = 6$ decoder layers with previous methods [5, 37, 62] to exploit intrinsic tasks connections implicitly. The interaction process within these layers can be formally expressed as:", + "bbox": [ + 212, + 239, + 787, + 347 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\nQ _ {l} = \\Omega_ {l} \\left(F ^ {3 d}, Q _ {l - 1}\\right), \\quad l = 1, \\dots , N _ {L} \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 372, + 352, + 785, + 369 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "With unified head $\\varPhi=\\{\\varPhi_{g},\\varPhi_{s}\\}$ , predictions $\\hat{R}_{li}$ and $\\hat{C}_{li}$ is decoded from $i$ -th query at $l$ -th layer:", + "bbox": [ + 212, + 375, + 785, + 404 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {R} _ {l i} = \\Phi_ {g} \\left(Q _ {l i}\\right), \\quad \\hat {\\mathcal {C}} _ {l i} = \\Phi_ {s} \\left(Q _ {l i}\\right) \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 395, + 405, + 785, + 421 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "For $\\varPhi_{g}$ , additional offsets are added to produce $S_{li}$ . Unified geometric predictions $\\hat{R}_l = (\\hat{S}_l,\\hat{F}_l)\\subset \\mathbb{R}^{n_s\\times (3 + d_p)}$ are converted to task-specific representations:", + "bbox": [ + 214, + 424, + 785, + 455 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {L} _ {l i} = \\mathcal {T} ^ {l} (\\hat {R} _ {l i}), \\quad \\hat {B} _ {l i} = \\left(\\mathcal {T} _ {x y z} ^ {b}, \\mathcal {T} _ {\\theta} ^ {b}, \\mathcal {T} _ {l w h} ^ {b}\\right) (\\hat {R} _ {l i}) \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 348, + 462, + 785, + 481 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Here, $\\mathcal{T}^{(\\cdot)}$ are differentiable functions predefined in section 3.2. Each head branch $\\varPhi_{(\\cdot)}$ is a simple small MLP network with three fully connected layers. Following DETR3D [62], we weight the loss for the predictions from each decoder layer $\\Omega_l$ during training, and only outputs from the last layer is used during inference.", + "bbox": [ + 212, + 487, + 787, + 547 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Loss Function. RFTR uses existing task labels for supervision, and its loss is calculated between converted predictions and task ground truths. It consists of the following components for coordinate regression, vector fields regression and classification respectively:", + "bbox": [ + 212, + 566, + 787, + 626 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} ^ {(\\cdot)} = w _ {r} ^ {(\\cdot)} \\mathcal {L} _ {r} ^ {(\\cdot)} + w _ {f} ^ {(\\cdot)} \\mathcal {L} _ {f} ^ {(\\cdot)} + w _ {c} ^ {(\\cdot)} \\mathcal {L} _ {c} ^ {(\\cdot)}, \\quad \\mathcal {L} = \\sum_ {\\text {t a s k s}} \\mathcal {L} ^ {(\\cdot)} \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 321, + 633, + 785, + 664 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where $w^{(\\cdot)}$ represent loss component weights for task $(\\cdot)$ , i.e. 3D object detection and 3D lane detection. We use the Hungarian algorithm [23] for label assignment after cost calculation between ground truth and converted predictions $\\hat{B}_{li}, \\hat{P}_{li}$ .", + "bbox": [ + 212, + 671, + 785, + 715 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "In practice, we choose $w_{c}^{b} = w_{c}^{l} = 2.0$ , $w_{r}^{l} = 0.003$ , $w_{r}^{b} = 0.1$ , $w_{f}^{b} = 0.2$ and $w_{f}^{l} = 0.032$ . Coordinate regression and classification weights are selected empirically to equalize the scales and contributions of each loss term [74], fields weights are empirically selected. We use $\\mathrm{L}^1$ loss for $R_{i}$ regression, and focal loss [32] with $\\gamma = 2.0$ and $\\alpha = 0.25$ for classification. Since we predict a degenerated scalar field, we supervise the angular prediction by its sine and cosine values. All tasks share the same losses and similar weights, no additional, task-specific loss measures are adopted.", + "bbox": [ + 212, + 715, + 787, + 840 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "RepVF: Unified Multi-task 3D Perception", + "bbox": [ + 450, + 114, + 730, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 114, + 785, + 127 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/9f6a331ed5cd25b2a4805ca119f81d2ff5f3e2f6bfe46e5fbadcb08a203d58fe.jpg", + "image_caption": [ + "(a) Gradient disparities by iteration index." + ], + "image_footnote": [], + "bbox": [ + 254, + 148, + 558, + 284 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/2bd0552c0fb3bfb1dbd0f2bdaa5704f10ab8a1549f59c09b67bea00dafe52298.jpg", + "image_caption": [ + "(b) Average gradient disparities.", + "Fig. 3: Our single-head RFTR model shows a reduced gradient imbalance and better disparity stability (mean: 2.47, variance: 0.90) compared to the multi-head baseline (mean: 2.98 variance: 1.86) or the multi-head baseline with RepVF (mean: 2.72, variance: 1.57). Ideally balanced gradient disparity through training iterations should approach 2. For ease of trend observation, we have clipped the top $2\\%$ values to the mean, with curves smoothed through a window size of 50 1-d convolution." + ], + "image_footnote": [], + "bbox": [ + 571, + 148, + 746, + 282 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.2 Gradient Balance in Single-head Multi-tasking", + "text_level": 1, + "bbox": [ + 214, + 422, + 645, + 438 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "The multi-head multi-task approach often leads to task conflict, attributed in prior studies [21, 33, 46, 52] to gradient imbalance. These studies [33, 46] and more recent work [19] have explored appropriate gradient calibration strategies to address this issue. As we will show, our single-head RFTR naturally achieves better balanced multi-task gradients thus calibration may no longer be required.", + "bbox": [ + 212, + 444, + 782, + 518 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "To elucidate the advantage of our single-head multi-task paradigm over multi-head multi-task settings, we use an improved measure of gradient disparity adapted from FULLER [19], by taking into account the symmetry disparity:", + "bbox": [ + 212, + 520, + 784, + 565 + ], + "page_idx": 9 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {d i f f} \\left(\\nabla \\mathcal {L} ^ {l}, \\nabla \\mathcal {L} ^ {b}\\right) = \\frac {\\| \\nabla \\mathcal {L} ^ {l} \\|}{\\| \\nabla \\mathcal {L} ^ {b} \\|} + \\frac {\\| \\nabla \\mathcal {L} ^ {b} \\|}{\\| \\nabla \\mathcal {L} ^ {l} \\|} \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 372, + 571, + 784, + 606 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "where $\\| \\cdot \\|$ is the Frobenius norm and $\\nabla \\mathcal{L}^{(\\cdot)}$ denotes gradients. Given that $\\| \\nabla \\mathcal{L}^{(\\cdot)}\\|$ is always greater than 0, the value of diff function is always greater than or equal to 2. Therefore, the more stable and closer to 2 the gradient disparity is during model training, the better the multi-task stability indicated.", + "bbox": [ + 212, + 612, + 782, + 672 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Figure 3a illustrates the gradient disparity curves through training iterations. The multi-head multi-task method [37,38] is under the same task settings with ours. In the sampling 1k iterations, the stability and absolute scale of the gradient disparity curve for the multi-head configuration are inferior to our single-head multi-task structure. The mean and variance of the multi-head disparities are about $108.5\\%$ and $107.8\\%$ higher, respectively, than our single-headed RFTR approach, as shown in Fig. 3b. We have also observed that unified representation and single head design contributed to the improvement about equally. More importantly, this means that our single-head RFTR mitigates, though not eliminating, the gradient imbalance between tasks, which is fundamentally different from task gradient balancing strategies that do not change the gradient itself.", + "bbox": [ + 212, + 674, + 787, + 840 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "C. Li et al.", + "bbox": [ + 271, + 114, + 346, + 126 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "5 Experiments", + "text_level": 1, + "bbox": [ + 215, + 143, + 375, + 162 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "We evaluate our method on the Waymo Open Dataset [53] extended by 3D lanes annotated in OpenLane [6], so that we train and evaluate tasks simultaneously.", + "bbox": [ + 214, + 176, + 787, + 208 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "5.1 Datasets", + "text_level": 1, + "bbox": [ + 215, + 229, + 334, + 243 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Waymo Open Dataset and OpenLane. Waymo Open Dataset (Perception) [53] is a large-scale multimodal dataset composed of data from 5 cameras(front and sides), 1 mid-range lidar, and 4 short-range lidars. It contains 1000 sequences in total(about 198k samples), a training set of 798 sequences and a validation set of 202 sequences. OpenLane [6] is a comprehensive real-world 3D lane detection dataset built upon the Waymo Open dataset [53]. It includes a vast array of 200K frames and over 880K carefully annotated lanes, making it one of the largest of its kind to date. The dataset encompasses complex lane structures and features a variety of scene tags, such as weather and locations. Designed to emulate real-world scenarios, OpenLane provides a challenging benchmark for advanced lane detection algorithms. We integrate its 3D lane labels with the original Waymo Open sensor data for multi-task annotation.", + "bbox": [ + 214, + 255, + 787, + 436 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Alignment of Datasets. To achieve both tasks simultaneously, we align both datasets in terms of coordinate systems and data splits. (1) Coordinate systems. For the vehicle frame, OpenLane [6] uses a vehicle coordinate system that corresponds to the \"y-front, x-right, z-up\" positioning of 3D-LaneNet [11]. We then align this system to coincide with the commonly utilized \"x-front, y-right, z-up\" axis for lidar-based 3D object detection. With respect to the sensor frame, we apply transformations to the camera intrinsics so that the sensor frames for all cameras conform to a uniform \"z-depth, x-right, y-down\" frame. The predictions are then transformed back into their respective task-based coordinate systems for evaluation. (2) Data splitting. The OpenLane [6] dataset provides a 1000 segment full version and a 300 segment smaller subset ( $30\\%$ ). The partitioning of the larger version complements the original train/Validation split of the Waymo Open Dataset. The subset version partitioning is aligned with OpenLane. We perform ablation studies on the smaller subset.", + "bbox": [ + 214, + 459, + 789, + 672 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "5.2 Implementation Details", + "text_level": 1, + "bbox": [ + 215, + 693, + 457, + 709 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Data Processing. We use the v1.4.2 of Waymo Open Dataset [53], and v1.2 for OpenLane [6]. We set the perception range to $[-84.88m, 84.88m]$ for the X and Y axis, and $[-10m, 10m]$ for the Z axis. 3D objects that do not fall within the perception range, are invisible to the cameras, or do not have LiDAR points, are filtered. For 3D lanes, visibility filtering is first applied after axis transformation, then 3D lanes outside the perception range are pruned and then resampled to a fixed size. Approximately $2.07\\%$ of the frames without annotations after filtering are excluded from use.", + "bbox": [ + 214, + 718, + 787, + 839 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "RepVF: Unified Multi-task 3D Perception", + "bbox": [ + 450, + 114, + 730, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 785, + 127 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/c175dacc54a44eaf68865ae90af4b670a220dca536cc682e1b7ab238e6533791.jpg", + "table_caption": [ + "Table 1: Comparison of recent models on 3D lane detection using the OpenLane [6] validation set. The down arrow indicates that lower metric values correspond to better performance, and vice versa. Red indicates the best result, and blue the second-best." + ], + "table_footnote": [], + "table_body": "
MethodsF-Score ↑Category Accuracy ↑X error (m) ↓Z error (m) ↓
nearfarnearfar
3D-LaneNet [11] [ICCV19]44.1-0.4790.5720.3670.443
Gen-LaneNet [14] [ECCV20]32.3-0.5910.6840.4110.521
PersFormer [6] [ECCV22]50.592.30.4850.5530.3640.431
Cond-IPM [6]36.6-0.5631.0800.4210.892
CurveFormer [3] [ICRA23]50.5-0.3400.7720.2070.651
PETRv2-E [38] [ICCV23]51.9-0.4930.6430.3220.463
PETRv2-V-10 [38]57.8-0.4270.5820.2930.421
PETRv2-V-400 [38]61.2-0.4000.5730.2650.413
BEV-LaneDet [57] [CVPR23]58.4-0.3090.6590.2440.631
Anchor3DLane [18] [CVPR23]53.190.00.3000.3110.1030.139
SPG [69] [ICCV23]52.3-0.4680.5140.3710.418
RFTR (Ours)61.891.60.3410.4500.0730.107
", + "bbox": [ + 251, + 198, + 759, + 358 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/470b00319f5b3fe4016db9e10ea2529dd7b8e791cc8c8133195cf7040d40507c.jpg", + "table_caption": [ + "Table 2: Comparison of recent works of 3D object detection on the Waymo Open Dataset [53] val set. mAPs are LEVEL_1 in Waymo metrics." + ], + "table_footnote": [], + "table_body": "
MethodsOverallVehicleCyclistPedestrian
mAPmAPHmAPmAPHmAPmAPHmAPmAPH
DETR3D [62]10.19.69.29.115.815.55.34.1
DETR3D [62]8.68.018.118.02.41.65.34.4
PETR [37]20.919.731.130.819.517.612.110.6
RFTR (Ours)19.518.722.622.427.826.38.07.2
", + "bbox": [ + 251, + 419, + 759, + 502 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Model and Training. We used ResNet 50 [16] as the 2D image backbone for fair comparisons. Following PETR [37], the same upsampled and fused P4 feature with 1/16 input resolution is used as the 2D feature. A CPFPN [31] is used as the image neck. All 3D coordinates are normalized to $[0,1]$ . The same weights are used for both the loss computation and the bipartite matching. We adopt progressive weighting for the loss of each decoder layer to improve the convergence speed. For 3D object detection, we compute matching costs based on the center displacement and class score weightings. Dimensions are refined post-matching using the ground truth headings to increase training robustness. We use a total of $n_s = 750$ queries with $n = 15$ sample locations each for the two tasks, 600 for 3D objects and 150 for 3D lanes. We trained our model using the AdamW [41] optimizer with a 0.01 weight decay. The learning rate started at $2.0 \\times 10^{-4}$ and was decayed with a cosine annealing policy [40]. We adopted a multi-scale training strategy, with the shorter side chosen randomly from [640,900] and the longer side not exceeding 1600. We applied rotation augmentations between $[-\\pi /18,\\pi /18]$ . We resized images of sizes $1920 \\times 1080$ and $1920 \\times 1280$ from different views to $1600 \\times 900$ before augmenting. The final resolution for network inputs was $1408 \\times 512$ . All experiments on the full dataset are trained for 24 epochs on 4 GeForce RTX 4090 GPUs with a batch size of 8. No test time augmentation methods are used during inference.", + "bbox": [ + 212, + 537, + 787, + 843 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "C. Li et al.", + "bbox": [ + 271, + 114, + 344, + 126 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/06f1d69f3e44123ec45862776d7fee29e3761a7f5836c7cf9a9f1598cc45f104.jpg", + "image_caption": [ + "Fig. 4: Visualization of the RFTR model's performance and the learned $RepVF$ ." + ], + "image_footnote": [], + "bbox": [ + 220, + 145, + 782, + 313 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "5.3 Comparison with SOTA", + "text_level": 1, + "bbox": [ + 215, + 368, + 460, + 383 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "3D Lane Detection. As shown in Tab. 1, our RFTR model showcases significant achievements in 3D lane detection on the OpenLane validation set, demonstrating its competitive edge with a high F-score of 61.8 and category accuracy of $91.6\\%$ . It matches or exceeds state-of-the-art models such as PETRv2-V-400 [38] (more than $26\\times$ denser representation than ours) when assessed on the OpenLane [6] validation set. Notably, while our model has a slightly higher X error in the \"far\" class than [18], it manages to lead with a higher F-score and near accurate Z errors. Moreover, our method yields notable performance improvements over the Persformer [6] baseline, reflecting the advantageous adaptation of our unified RepVF representation and the RFTR architecture.", + "bbox": [ + 212, + 392, + 787, + 544 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "3D Object Detection. The performance of our RFTR model in 3D object detection, as demonstrated in Tab. 2, shows comparable promising results on the Waymo Open Dataset [53] validation set. While our approach shows a commanding lead in detecting vehicles and cyclists with a solid L1 mAP and mAPH, it shows challenges in detecting smaller entities such as pedestrians. It should be noted, however, that these preliminary results indicate the great potential of using our unified perceptual model for complex tasks, which will require further fine-tuning for object classes with less prominent features.", + "bbox": [ + 212, + 565, + 787, + 686 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "5.4 Ablation Studies", + "text_level": 1, + "bbox": [ + 215, + 709, + 401, + 723 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Head design. We begin by finding out how head designs and representations contribute to the performance, in Tab. 3. The $RepVF$ representation with the single-head design, i.e. RFTR, performs the best in general with 37.4M parameters. Unified representation brings performance improvements in multi-task learning and reduces task competition. In contrast, task-specific multi-head design produces inferior and unstably biased performance due to task competition, even with more parameters.", + "bbox": [ + 212, + 733, + 787, + 839 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "RepVF: Unified Multi-task 3D Perception", + "bbox": [ + 450, + 114, + 730, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/6c7d99a2902f34c4f5f7a219f2fd8fe02fa3a6dcdbef81ba2d8fc22e549538e8.jpg", + "table_caption": [ + "Table 3: Multi-task head design and representation ablation. Tasks are trained jointly on $30\\%$ data of Waymo Open Dataset (L1 mAP) and OpenLane (F-Score). TS denotes task-specific representation. $\\dagger$ : lightweight models." + ], + "table_footnote": [], + "table_body": "
RepVFTSHeadParams↓F-Score↑Vehicle↑
✓Multi†37.1M49.620.1
✓Multi46.7M58.717.8
✓Multi†39.7M63.418.2
✓Multi49.2M66.422.1
✓Single37.4M66.525.3
", + "bbox": [ + 233, + 226, + 509, + 320 + ], + "page_idx": 13 + }, + { + "type": "table", + "img_path": "images/e31ad472859c4307f83241df8d3cd7020a364f3a42a9a43ece5d08cda42fdcf7.jpg", + "table_caption": [ + "Table 4: Results with different representations on 3D lane. 'Acc' for category accuracy and 'Error' for X error near. MT means simultaneous multitask 3D lane and 3D object detection." + ], + "table_footnote": [], + "table_body": "
RepresentationF-Score↑Acc↑Error↓
Anchor53.486.00.476
PS58.188.00.545
RepVF66.091.40.439
RepVF + MT66.591.70.420
", + "bbox": [ + 542, + 231, + 777, + 316 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Representation. Moreover, we have evaluated the effectiveness of different representations for 3D lane detection in Tab. 4, including traditional anchor representations [6,37], simple randomly sampled point set representation (PS), and our proposed representative vector fields (RepVF). This study reveals that RepVF significantly outperforms traditional anchor and PS methods, and incorporating multi-task learning marginally improves these metrics.", + "bbox": [ + 212, + 349, + 787, + 441 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5.5 Qualitative Results", + "text_level": 1, + "bbox": [ + 214, + 469, + 421, + 484 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Figure 4 demonstrates the capability of RFTR to simultaneously achieve 3D lane detection and 3D object detection on the extended [24] Waymo open dataset [53]. The figure clearly illustrates that RFTR can accurately predict RepVF and perform both tasks at once. In terms of 3D lane detection, RFTR is capable of predicting the shape of lane lines in areas obscured by vehicles, demonstrating its robustness against occlusion. Regarding 3D object detection, the unified RepVF representation accurately captures the 3D bounding boxes. This dual achievement highlights RFTR's effectiveness in handling complex driving scenarios, enhancing both navigation and safety features in autonomous driving systems.", + "bbox": [ + 212, + 500, + 787, + 652 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "6 Conclusion", + "text_level": 1, + "bbox": [ + 214, + 680, + 359, + 696 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In this paper, we propose $RepVF$ , a novel unified vector fields representation that characterizes the geometric and semantic structure of different 3D perception targets in the scene. Built upon $RepVF$ , our proposed RFTR exploits the intrinsic connections between tasks, mitigates multi-task gradient imbalance and feature competition, and improves model convergence and multi-task expression. Our work not only contributes a novel single-head perspective to the multi-task learning paradigm in autonomous driving but also sets a promising direction for future research in efficient and effective 3D perception task handling.", + "bbox": [ + 212, + 719, + 787, + 840 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "C. Li et al.", + "bbox": [ + 271, + 114, + 346, + 127 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 215, + 143, + 323, + 159 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Mobileye at CES 2024. https://www-mobileye.com/ces-2024/", + "2. NVIDIA DRIVE Solutions. https://developer.nvidia.com/drive", + "3. Bai, Y., Chen, Z., Fu, Z., Peng, L., Liang, P., Cheng, E.: CurveFormer: 3D Lane Detection by Curve Propagation with Curve Queries and Attention. In: 2023 IEEE International Conference on Robotics and Automation (ICRA). pp. 7062-7068 (May 2023). https://doi.org/10.1109/ICRA48891.2023.10161160", + "4. Caesar, H., Bankiti, V., Lang, A.H., Vora, S., Liong, V.E., Xu, Q., Krishnan, A., Pan, Y., Baldan, G., Beijbom, O.: nuScenes: A Multimodal Dataset for Autonomous Driving. In: 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 11618-11628. IEEE, Seattle, WA, USA (Jun 2020). https://doi.org/10.1109/CVPR42600.2020.01164", + "5. Carion, N., Massa, F., Synnaeve, G., Usunier, N., Kirillov, A., Zagoruyko, S.: End-to-End Object Detection with Transformers. In: Vedaldi, A., Bischof, H., Brox, T., Frahm, J.M. (eds.) Computer Vision - ECCV 2020. pp. 213-229. Lecture Notes in Computer Science, Springer International Publishing, Cham (2020). https://doi.org/10.1007/978-3-030-58452-8_13", + "6. Chen, L., Sima, C., Li, Y., Zheng, Z., Xu, J., Geng, X., Li, H., He, C., Shi, J., Qiao, Y., Yan, J.: PersFormer: 3D Lane Detection via Perspective Transformer and the OpenLane Benchmark. In: Avidan, S., Brostow, G., Cisse, M., Farinella, G.M., Hassner, T. (eds.) Computer Vision - ECCV 2022, vol. 13698, pp. 550-567. Springer Nature Switzerland, Cham (2022). https://doi.org/10.1007/978-3-031-19839-7_32", + "7. Chen, X., Kundu, K., Zhang, Z., Ma, H., Fidler, S., Urtasun, R.: Monocular 3D Object Detection for Autonomous Driving. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 2147-2156 (2016)", + "8. Chen, Z., Badrinarayanan, V., Lee, C.Y., Rabinovich, A.: GradNorm: Gradient Normalization for Adaptive Loss Balancing in Deep Multitask Networks. In: Proceedings of the 35th International Conference on Machine Learning. pp. 794-803. PMLR (Jul 2018)", + "9. Cheng, W., Yin, J., Li, W., Yang, R., Shen, J.: Language-Guided 3D Object Detection in Point Cloud for Autonomous Driving (May 2023)", + "0. Feng, Z., Guo, S., Tan, X., Xu, K., Wang, M., Ma, L.: Rethinking Efficient Lane Detection via Curve Modeling (May 2023). https://doi.org/10.48550/arXiv.2203.02431", + "1. Garnett, N., Cohen, R., Pe'er, T., Lahav, R., Levi, D.: 3D-LaneNet: End-to-End 3D Multiple Lane Detection. In: 2019 IEEE/CVF International Conference on Computer Vision (ICCV). pp. 2921–2930. IEEE, Seoul, Korea (South) (Oct 2019). https://doi.org/10.1109/ICCV.2019.00301", + "12. Geiger, A., Lenz, P., Urtasun, R.: Are we ready for autonomous driving? The KITTI vision benchmark suite. In: 2012 IEEE Conference on Computer Vision and Pattern Recognition. pp. 3354-3361 (Jun 2012). https://doi.org/10.1109/CVPR.2012.6248074", + "13. Guo, P., Lee, C.Y., Ulbricht, D.: Learning to Branch for Multi-Task Learning. In: Proceedings of the 37th International Conference on Machine Learning. pp. 3854-3863. PMLR (Nov 2020)", + "14. Guo, Y., Chen, G., Zhao, P., Zhang, W., Miao, J., Wang, J., Choe, T.E.: GenLaneNet: A Generalized and Scalable Approach for 3D Lane Detection. In: Vedaldi, A., Bischof, H., Brox, T., Frahm, J.M. (eds.) Computer Vision - ECCV 2020. pp." + ], + "bbox": [ + 218, + 175, + 785, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "RepVF: Unified Multi-task 3D Perception", + "bbox": [ + 450, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "666-681. Lecture Notes in Computer Science, Springer International Publishing, Cham (2020). https://doi.org/10.1007/978-3-030-58589-1_40", + "15. Han, W., Shen, J.: Decoupling the Curve Modeling and Pavement Regression for Lane Detection (Sep 2023)", + "16. He, K., Zhang, X., Ren, S., Sun, J.: Deep Residual Learning for Image Recognition. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 770-778 (2016)", + "17. Huang, J., Huang, G.: BEVDet4D: Exploit Temporal Cues in Multi-camera 3D Object Detection (Jun 2022)", + "18. Huang, S., Shen, Z., Huang, Z., Ding, Z.h., Dai, J., Han, J., Wang, N., Liu, S.: Anchor3DLane: Learning To Regress 3D Anchors for Monocular 3D Lane Detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 17451-17460 (2023)", + "19. Huang, Z., Lin, S., Liu, G., Luo, M., Ye, C., Xu, H., Chang, X., Liang, X.: FULLER: Unified Multi-modality Multi-task 3D Perception via Multi-level Gradient Calibration. In: 2023 IEEE/CVF International Conference on Computer Vision (ICCV). pp. 3479-3488. IEEE, Paris, France (Oct 2023). https://doi.org/10.1109/ICCV51070.2023.00324", + "20. Kehl, W., Manhardt, F., Tombari, F., Ilic, S., Navab, N.: SSD-6D: Making RGB-Based 3D Detection and 6D Pose Estimation Great Again. In: Proceedings of the IEEE International Conference on Computer Vision. pp. 1521–1529 (2017)", + "21. Kendall, A., Gal, Y., Cipolla, R.: Multi-Task Learning Using Uncertainty to Weigh Losses for Scene Geometry and Semantics. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 7482-7491 (2018)", + "22. Ku, J., Pon, A.D., Waslander, S.L.: Monocular 3D Object Detection Leveraging Accurate Proposals and Shape Reconstruction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11867-11876 (2019)", + "23. Kuhn, H.W.: The Hungarian method for the assignment problem. Naval Research Logistics Quarterly 2(1-2), 83-97 (1955). https://doi.org/10.1002/nav.3800020109", + "24. Li, L.H., Yatskar, M., Yin, D., Hsieh, C.J., Chang, K.W.: VisualBERT: A Simple and Performant Baseline for Vision and Language (Aug 2019). https://doi.org/10.48550/arXiv.1908.03557", + "25. Li, X., Yin, J., Li, W., Xu, C., Yang, R., Shen, J.: Di-v2x: Learning domain-invariant representation for vehicle-infrastructure collaborative 3d object detection. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 38, pp. 3208-3215 (2024)", + "26. Li, X., Yin, J., Shi, B., Li, Y., Yang, R., Shen, J.: Lwsis: Lidar-guided weakly supervised instance segmentation for autonomous driving. Proceedings of the AAAI Conference on Artificial Intelligence 37(2), 1433-1441 (Jun 2023). https://doi.org/10.1609/aaii.v37i2.25228, https://ojs.aaii.org/index.php/AAAI/article/view/25228", + "27. Li, Y., Chen, Y., Qi, X., Li, Z., Sun, J., Jia, J.: Unifying Voxel-based Representation with Transformer for 3D Object Detection. Advances in Neural Information Processing Systems 35, 18442-18455 (Dec 2022)", + "28. Li, Z., Wang, W., Li, H., Xie, E., Sima, C., Lu, T., Qiao, Y., Dai, J.: BEVFormer: Learning Bird's-Eye-View Representation from Multi-camera Images via Spatiotemporal Transformers. In: Avidan, S., Brostow, G., Cisse, M., Farinella, G.M., Hassner, T. (eds.) Computer Vision - ECCV 2022. pp. 1-18. Lecture Notes in Computer Science, Springer Nature Switzerland, Cham (2022). https://doi.org/10.1007/978-3-031-20077-9_1" + ], + "bbox": [ + 215, + 146, + 784, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "C. Li et al.", + "bbox": [ + 271, + 114, + 346, + 126 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "29. Liang, M., Yang, B., Chen, Y., Hu, R., Urtasun, R.: Multi-Task Multi-Sensor Fusion for 3D Object Detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7345-7353 (2019)", + "30. Liang, T., Xie, H., Yu, K., Xia, Z., Lin, Z., Wang, Y., Tang, T., Wang, B., Tang, Z.: BEVFusion: A Simple and Robust LiDAR-Camera Fusion Framework. Advances in Neural Information Processing Systems 35, 10421-10434 (Dec 2022)", + "31. Lin, T.Y., Dollar, P., Girshick, R., He, K., Hariharan, B., Belongie, S.: Feature Pyramid Networks for Object Detection. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 2117-2125 (2017)", + "32. Lin, T.Y., Goyal, P., Girshick, R., He, K., Dollar, P.: Focal Loss for Dense Object Detection. In: Proceedings of the IEEE International Conference on Computer Vision. pp. 2980-2988 (2017)", + "33. Liu, L., Li, Y., Kuang, Z., Xue, J.H., Chen, Y., Yang, W., Liao, Q., Zhang, W.: Towards Impartial Multi-task Learning. In: International Conference on Learning Representations (Oct 2020)", + "34. Liu, R., Chen, D., Liu, T., Xiong, Z., Yuan, Z.: Learning to Predict 3D Lane Shape and Camera Pose from a Single Image via Geometry Constraints. Proceedings of the AAAI Conference on Artificial Intelligence 36(2), 1765-1772 (Jun 2022). https://doi.org/10.1609/aaai.v36i2.20069", + "35. Liu, S., Johns, E., Davison, A.J.: End-to-End Multi-Task Learning with Attention (Apr 2019). https://doi.org/10.48550/arXiv.1803.10704", + "36. Liu, S., Li, F., Zhang, H., Yang, X., Qi, X., Su, H., Zhu, J., Zhang, L.: DAB-DETR: DYNAMIC ANCHOR BOXES ARE BETTER QUERIES FOR DETR (2022)", + "37. Liu, Y., Wang, T., Zhang, X., Sun, J.: PETR: Position Embedding Transformation for Multi-view 3D Object Detection. In: Avidan, S., Brostow, G., Cisse, M., Farinella, G.M., Hassner, T. (eds.) Computer Vision - ECCV 2022, vol. 13687, pp. 531-548. Springer Nature Switzerland, Cham (2022). https://doi.org/10.1007/978-3-031-19812-0_31", + "38. Liu, Y., Yan, J., Jia, F., Li, S., Gao, A., Wang, T., Zhang, X.: PETRv2: A Unified Framework for 3D Perception from Multi-Camera Images", + "39. Liu, Z., Tang, H., Amini, A., Yang, X., Mao, H., Rus, D.L., Han, S.: BEVFusion: Multi-Task Multi-Sensor Fusion with Unified Bird's-Eye View Representation. In: 2023 IEEE International Conference on Robotics and Automation (ICRA). pp. 2774-2781 (May 2023). https://doi.org/10.1109/ICRA48891.2023.10160968", + "40. Loshchilov, I., Hutter, F.: SGDR: Stochastic Gradient Descent with Warm Restarts. In: International Conference on Learning Representations (Nov 2016)", + "41. Loshchilov, I., Hutter, F.: DECOUPLED WEIGHT DECAY REGULARIZATION (2019)", + "42. Luo, Y., Zheng, C., Yan, X., Kun, T., Zheng, C., Cui, S., Li, Z.: LATR: 3D Lane Detection from Monocular Images with Transformer. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 7941-7952 (2023)", + "43. Ma, Z., Wang, L., Zhang, H., Lu, W., Yin, J.: RPT: Learning Point Set Representation for Siamese Visual Tracking. In: Bartoli, A., Fusiello, A. (eds.) Computer Vision - ECCV 2020 Workshops. pp. 653-665. Lecture Notes in Computer Science, Springer International Publishing, Cham (2020). https://doi.org/10.1007/978-3-030-68238-5_43", + "44. Misra, I., Shrivastava, A., Gupta, A., Hebert, M.: Cross-Stitch Networks for Multi-Task Learning. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 3994–4003 (2016)" + ], + "bbox": [ + 212, + 146, + 787, + 839 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "RepVF: Unified Multi-task 3D Perception", + "bbox": [ + 450, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "45. Mousavian, A., Anguelov, D., Flynn, J., Kosecka, J.: 3D Bounding Box Estimation Using Deep Learning and Geometry. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 7074-7082 (2017)", + "46. Navon, A., Shamsian, A., Achituve, I., Maron, H., Kawaguchi, K., Chechik, G., Fetaya, E.: Multi-Task Learning as a Bargaining Game (Feb 2022)", + "47. Neven, D., De Brabandere, B., Georgoulis, S., Proesmans, M., Van Gool, L.: Fast Scene Understanding for Autonomous Driving (Aug 2017). https://doi.org/10.48550/arXiv.1708.02550", + "48. Philion, J., Fidler, S.: Lift, Splat, Shoot: Encoding Images from Arbitrary Camera Rigs by Implicitly Unprojecting to 3D. In: Vedaldi, A., Bischof, H., Brox, T., Frahm, J.M. (eds.) Computer Vision - ECCV 2020. pp. 194-210. Lecture Notes in Computer Science, Springer International Publishing, Cham (2020). https://doi.org/10.1007/978-3-030-58568-6_12", + "49. Reading, C., Harakeh, A., Chae, J., Waslander, S.L.: Categorical Depth Distribution Network for Monocular 3D Object Detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8555-8564 (2021)", + "50. Roddick, T., Kendall, A., Cipolla, R.: Orthographic Feature Transform for Monocular 3D Object Detection (Nov 2018). https://doi.org/10.48550/arXiv.1811.08188", + "51. Ruder, S., Bingel, J., Augenstein, I., Søgaard, A.: Latent Multi-Task Architecture Learning. Proceedings of the AAAI Conference on Artificial Intelligence 33(01), 4822-4829 (Jul 2019). https://doi.org/10.1609/aaai.v33i01.33014822", + "52. Sener, O., Koltun, V.: Multi-Task Learning as Multi-Objective Optimization. In: Advances in Neural Information Processing Systems. vol. 31. Curran Associates, Inc. (2018)", + "53. Sun, P., Kretzschmar, H., Dotiwalla, X., Chouard, A., Patnaik, V., Tsui, P., Guo, J., Zhou, Y., Chai, Y., Caine, B., Vasudevan, V., Han, W., Ngiam, J., Zhao, H., Timofeev, A., Ettinger, S., Krivokon, M., Gao, A., Joshi, A., Zhang, Y., Shlens, J., Chen, Z., Anguelov, D.: Scalability in Perception for Autonomous Driving: Waymo Open Dataset", + "54. Teichmann, M., Weber, M., Zöllner, M., Cipolla, R., Urtasun, R.: MultiNet: Realtime Joint Semantic Reasoning for Autonomous Driving. In: 2018 IEEE Intelligent Vehicles Symposium (IV). pp. 1013-1020 (Jun 2018). https://doi.org/10.1109/IVS.2018.8500504", + "55. Vandenhende, S., Georgoulis, S., Van Gansbeke, W., Proesmans, M., Dai, D., Van Gool, L.: Multi-Task Learning for Dense Prediction Tasks: A Survey. IEEE Transactions on Pattern Analysis and Machine Intelligence 44(7), 3614-3633 (Jul 2022). https://doi.org/10.1109/TPAMI.2021.3054719", + "56. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is All you Need. In: Advances in Neural Information Processing Systems. vol. 30. Curran Associates, Inc. (2017)", + "57. Wang, R., Qin, J., Li, K., Li, Y., Cao, D., Xu, J.: BEV-LaneDet: An Efficient 3D Lane Detection Based on Virtual Camera via Key-Points. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1002-1011 (2023)", + "58. Wang, S., Liu, Y., Wang, T., Li, Y., Zhang, X.: Exploring Object-Centric Temporal Modeling for Efficient Multi-View 3D Object Detection (Jun 2023)", + "59. Wang, T., Zhu, X., Pang, J., Lin, D.: FCOS3D: Fully Convolutional One-Stage Monocular 3D Object Detection. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 913–922 (2021)" + ], + "bbox": [ + 215, + 147, + 784, + 839 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "C. Li et al.", + "bbox": [ + 271, + 114, + 346, + 126 + ], + "page_idx": 17 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "60. Wang, Y., Yin, J., Li, W., Frossard, P., Yang, R., Shen, J.: Ssda3d: Semi-supervised domain adaptation for 3d object detection from point cloud. Proceedings of the AAAI Conference on Artificial Intelligence 37(3), 2707-2715 (Jun 2023). https://doi.org/10.1609/aaai.v37i3.25370, https://ojs(aaai.org/index.php/AAAI/article/view/25370", + "61. Wang, Y., Zhang, X., Yang, T., Sun, J.: Anchor DETR: Query Design for Transformer-Based Detector. Proceedings of the AAAI Conference on Artificial Intelligence 36(3), 2567-2575 (Jun 2022). https://doi.org/10.1609/aaai.v36i3.20158", + "62. Wang, Y., Guizilini, V.C., Zhang, T., Wang, Y., Zhao, H., Solomon, J.: DETR3D: 3D Object Detection from Multi-view Images via 3D-to-2D Queries. In: Proceedings of the 5th Conference on Robot Learning. pp. 180-191. PMLR (Jan 2022)", + "63. Wu, D., Chang, J., Jia, F., Liu, Y., Wang, T., Shen, J.: Topomlp: An simple yet strong pipeline for driving topology reasoning. arXiv preprint arXiv:2310.06753 (2023)", + "64. Wu, D., Jia, F., Chang, J., Li, Z., Sun, J., Han, C., Li, S., Liu, Y., Ge, Z., Wang, T.: The 1st-place solution for cvpr 2023 openlane topology in autonomous driving challenge. arXiv preprint arXiv:2306.09590 (2023)", + "65. Xie, E., Yu, Z., Zhou, D., Philion, J., Anandkumar, A., Fidler, S., Luo, P., Alvarez, J.M.: M$^2$BEV: Multi-Camera Joint 3D Detection and Segmentation with Unified Birds-Eye View Representation (Apr 2022). https://doi.org/10.48550/arXiv.2204.05088", + "66. Xu, D., Ouyang, W., Wang, X., Sebe, N.: PAD-Net: Multi-Tasks Guided Prediction-and-Distillation Network for Simultaneous Depth Estimation and Scene Parsing. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 675-684 (2018)", + "67. Yan, F., Nie, M., Cai, X., Han, J., Xu, H., Yang, Z., Ye, C., Fu, Y., Mi, M.B., Zhang, L.: ONCE-3DLanes: Building Monocular 3D Lane Detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 17143-17152 (2022)", + "68. Yang, Z., Liu, S., Hu, H., Wang, L., Lin, S.: RepPoints: Point Set Representation for Object Detection. In: 2019 IEEE/CVF International Conference on Computer Vision (ICCV). pp. 9656-9665. IEEE, Seoul, Korea (South) (Oct 2019). https://doi.org/10.1109/ICCV.2019.00975", + "69. Yao, C., Yu, L., Wu, Y., Jia, Y.: Sparse Point Guided 3D Lane Detection", + "70. Yin, J., Fang, J., Zhou, D., Zhang, L., Xu, C.Z., Shen, J., Wang, W.: Semi-supervised 3D Object Detection with Proficient Teachers. In: Avidan, S., Bros-tow, G., Cisse, M., Farinella, G.M., Hassner, T. (eds.) Computer Vision - ECCV 2022, vol. 13698, pp. 727-743. Springer Nature Switzerland, Cham (2022). https://doi.org/10.1007/978-3-031-19839-7_42", + "71. Yin, J., Shen, J., Chen, R., Li, W., Yang, R., Frossard, P., Wang, W.: Is-fusion: Instance-scene collaborative fusion for multimodal 3d object detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14905-14915 (2024)", + "72. Yin, J., Wang, W., Meng, Q., Yang, R., Shen, J.: A unified object motion and affinity model for online multi-object tracking. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6768-6777 (2020)", + "73. Yu, T., Kumar, S., Gupta, A., Levine, S., Hausman, K., Finn, C.: Gradient Surgery for Multi-Task Learning (Dec 2020). https://doi.org/10.48550/arXiv.2001.06782" + ], + "bbox": [ + 212, + 146, + 787, + 838 + ], + "page_idx": 18 + }, + { + "type": "header", + "text": "RepVF: Unified Multi-task 3D Perception", + "bbox": [ + 450, + 114, + 730, + 128 + ], + "page_idx": 18 + }, + { + "type": "page_number", + "text": "19", + "bbox": [ + 767, + 116, + 785, + 126 + ], + "page_idx": 18 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "74. Zhang, Y., Wang, C., Wang, X., Zeng, W., Liu, W.: FairMOT: On the Fairness of Detection and Re-Identification in Multiple Object Tracking. International Journal of Computer Vision 129(11), 3069–3087 (Nov 2021). https://doi.org/10.1007/s11263-021-01513-4", + "75. Zhou, D., Fang, J., Song, X., Liu, L., Yin, J., Dai, Y., Li, H., Yang, R.: Joint 3d instance segmentation and object detection for autonomous driving. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1839-1849 (2020)", + "76. Zhou, D., Song, X., Dai, Y., Yin, J., Lu, F., Liao, M., Fang, J., Zhang, L.: Iafa: Instance-aware feature aggregation for 3d object detection from a single image. In: Proceedings of the Asian Conference on Computer Vision (2020)", + "77. Zhu, X., Su, W., Lu, L., Li, B., Wang, X., Dai, J.: Deformable DETR: Deformable Transformers for End-to-End Object Detection (Mar 2021). https://doi.org/10.48550/arXiv.2010.04159" + ], + "bbox": [ + 215, + 146, + 787, + 340 + ], + "page_idx": 19 + }, + { + "type": "page_number", + "text": "20", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 19 + }, + { + "type": "header", + "text": "C. Li et al.", + "bbox": [ + 271, + 114, + 344, + 126 + ], + "page_idx": 19 + } +] \ No newline at end of file diff --git a/2024/RepVF_ A Unified Vector Fields Representation for Multi-task 3D Perception/9c289191-ecfe-49ca-a348-d2cad5492a03_model.json b/2024/RepVF_ A Unified Vector Fields Representation for Multi-task 3D Perception/9c289191-ecfe-49ca-a348-d2cad5492a03_model.json new file mode 100644 index 0000000000000000000000000000000000000000..ca15670f9df76f1ff0ee609e4c798e0d40684139 --- /dev/null +++ b/2024/RepVF_ A Unified Vector Fields Representation for Multi-task 3D Perception/9c289191-ecfe-49ca-a348-d2cad5492a03_model.json @@ -0,0 +1,2638 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.229, + 0.141, + 0.777, + 0.187 + ], + "angle": 0, + "content": "RepVF: A Unified Vector Fields Representation for Multi-task 3D Perception" + }, + { + "type": "text", + "bbox": [ + 0.315, + 0.212, + 0.689, + 0.245 + ], + "angle": 0, + "content": "Chunliang Li\\(^{1*}\\), Wencheng Han\\(^{2*}\\), Junbo Yin\\(^{1}\\), Sanyuan Zhao\\(^{1}\\), and Jianbing Shen\\(^{2}\\)" + }, + { + "type": "text", + "bbox": [ + 0.235, + 0.255, + 0.767, + 0.325 + ], + "angle": 0, + "content": "\\(^{1}\\) School of Computer Science, Beijing Institute of Technology \n\\(^{2}\\) SKL-IOTSC, Computer and Information Science, University of Macau, China \n{jbji, sanyuanzhao}@bit.edu.cn \n{wencheng256, yinjunbocn}@gmail.com \njianbingshen@um.edu.mo" + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.349, + 0.744, + 0.655 + ], + "angle": 0, + "content": "Abstract. Concurrent processing of multiple autonomous driving 3D perception tasks within the same spatiotemporal scene poses a significant challenge, in particular due to the computational inefficiencies and feature competition between tasks when using traditional multi-task learning approaches. This paper addresses these issues by proposing a novel unified representation, RepVF, which harmonizes the representation of various perception tasks such as 3D object detection and 3D lane detection within a single framework. RepVF characterizes the structure of different targets in the scene through a vector field, enabling a single-head, multi-task learning model that significantly reduces computational redundancy and feature competition. Building upon RepVF, we introduce RFTR, a network designed to exploit the inherent connections between different tasks by utilizing a hierarchical structure of queries that implicitly model the relationships both between and within tasks. This approach eliminates the need for task-specific heads and parameters, fundamentally reducing the conflicts inherent in traditional multi-task learning paradigms. We validate our approach by combining labels from the OpenLane dataset with the Waymo Open dataset. Our work presents a significant advancement in the efficiency and effectiveness of multi-task perception in autonomous driving, offering a new perspective on handling multiple 3D perception tasks synchronously and in parallel. The code will be available at: https://github.com/jbji/RepVF." + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.667, + 0.759, + 0.682 + ], + "angle": 0, + "content": "Keywords: 3D Lane Detection \\(\\cdot\\) 3D Object Detection \\(\\cdot\\) Multi-task Method." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.734, + 0.376, + 0.75 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.761, + 0.788, + 0.792 + ], + "angle": 0, + "content": "In autonomous driving, multiple 3D perception tasks often need to be processed synchronously, in real-time, and in parallel. These tasks, which exist within the" + }, + { + "type": "page_footnote", + "bbox": [ + 0.231, + 0.798, + 0.788, + 0.84 + ], + "angle": 0, + "content": "*Equal contribution. †Corresponding author. This work was supported in part by the FDCT grants 0102/2023/RIA2, 0154/2022/A3, and 001/2024/SKL, the MYRG-CRG2022-00013-IOTSC-ICI grant and the SRG2022-00023-IOTSC grant." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.347, + 0.127 + ], + "angle": 0, + "content": "C. Li et al." + }, + { + "type": "image", + "bbox": [ + 0.244, + 0.154, + 0.5, + 0.314 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.254, + 0.319, + 0.484, + 0.332 + ], + "angle": 0, + "content": "(a) Multi-tasking paradigm comparison." + }, + { + "type": "image", + "bbox": [ + 0.515, + 0.159, + 0.77, + 0.312 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.526, + 0.319, + 0.754, + 0.333 + ], + "angle": 0, + "content": "(b) \\( RepVF \\), conversion and supervision." + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.342, + 0.789, + 0.4 + ], + "angle": 0, + "content": "Fig. 1: \\( RepVF \\) is proposed as a unified representation for 3D perception to achieve single-head multi-tasking. It consists of a set of vector fields learned to represent road elements, and is trained utilizing existing labels through differentiable conversions. The proposed single-head multi-tasking paradigm reduces task conflict and competition." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.431, + 0.789, + 0.674 + ], + "angle": 0, + "content": "same spatiotemporal scene, are often complementary to each other in both geometric and semantic aspects [19]. For example, 3D object detection [12] and 3D lane detection [11] define foreground objects and background lanes within the same 3D scene. Vehicles usually travel following the guiding semantics of lanes that do not spatially overlap and are partially obscured by other vehicles. Some previous industrial solutions [1,2] utilize a multi-model paradigm for multi-task perceptions, as depicted in Figure 1a(1). These separate models cannot share common features across different tasks, resulting in computational waste. Furthermore, this paradigm does not take advantage of the inherent connections between different tasks. Some works [19,28,38,39] attempt to overcome this limitation by using a common backbone and separate head networks for each task, as shown in Figure 1a(2). This approach can considerably reduce the computational cost when handling a large number of tasks, but it might lead to a task balancing issue. Since different task-aware heads focus on various aspects of the scene, they may compete for the backbone features, resulting in instability during training [19,39]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.673, + 0.789, + 0.809 + ], + "angle": 0, + "content": "The competition is attributed to the variation in representation between different tasks. For example, the 3D detection task [12,26,28,37,38,60,62] guides the head network to predict 3D bounding boxes, whereas the lane detection task [6,11] necessitates the corresponding head to regress lane shapes. Specifically, 3D bounding boxes describe a cubic area in 3D space, representing different category foregrounds using position, size, and orientation. In contrast, 3D lanes are represented using anchor [6,11,18] or parametric [15] methods to describe one or more 1-D lines. These representations focus on different aspects of the scenes, which ultimately leads to competition." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.81, + 0.789, + 0.842 + ], + "angle": 0, + "content": "In this paper, we propose a unified representation, \\( RepVF \\) (Representative Vector Fields), for perception tasks in autonomous driving. Using this represent" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.451, + 0.115, + 0.732, + 0.13 + ], + "angle": 0, + "content": "RepVF: Unified Multi-task 3D Perception" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.787, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.344 + ], + "angle": 0, + "content": "tation, we design a single-headed, unified model that can perform tasks simultaneously. As illustrated in Fig. 1b, \\( RepVF \\) represents 3D vector fields that each assigns vectors to spatial locations, denoted as \\( (S, \\mathcal{F}(S)) \\subseteq \\mathbb{R}^{3 + d} \\) in Sec. 3.2. Sub-vector fields representing different targets in the 3D perception scene progressively adhere to their spatial extent, accurately characterizing the anisotropic structure in their respective spatial vicinity. \\( RepVF \\) can be differentiably transformed into task-specific elements, hence requiring no special supervision and can utilize existing labels. With \\( RepVF \\), our network only processes one single type of fundamental perception element. As shown in Fig. 1a, unlike the traditional multi-task approach that employs multiple task heads [19, 28, 30, 38, 72], we follow an unprecedented single-head multi-task paradigm. The single-head multi-task eliminates task-specific parameters and obviates the need to explicitly model task interactions." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.344, + 0.788, + 0.585 + ], + "angle": 0, + "content": "RepVF representation serve as a geometrical, cross-scale 3D perception representation alternative. Unlike perception elements capable of handling only single-scale tasks [6,11,12,62], it breaks down the scene into smaller perceptual units while retaining the capability for target-level perception at larger scales. Based on the RepVF representation, we further propose RFTR to leverage this relationship, utilizing queries with a hierarchical structure that implicitly models the relationships both between and within different tasks. Following the single-head multi-task architecture, with only shared parameters across tasks, RFTR does not depend on existing multi-task optimization strategies [21,33,35,52,73]. The gradient discrepancies between tasks across different training iterations have been mitigated, fundamentally reducing the competition and conflict inherent in multi-tasking 3D perception. By employing RepVF to replace the fundamental elements for different tasks, and by combining 3D lane labels from the OpenLane dataset [6] with the Waymo Open dataset [53] for simultaneous multi-task training and inference, our approach achieves comparable or even better multi-task performance than single-task counterparts with only one unified model." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.585, + 0.788, + 0.738 + ], + "angle": 0, + "content": "In summary, we introduce \\( RepVF \\) as a unified representation for multiple perception tasks in autonomous driving, effectively reducing the competition among different task representations during training and enhancing model convergence. We further develop a single-head multi-task framework RFTR based on \\( RepVF \\), fully exploiting the intrinsic connections among different tasks within a scene, improving feature interaction with queries, and enabling gradient-balanced multi-task training. Our approach achieves impressive performance across two key tasks in autonomous driving—3D object detection and 3D lane detection—showcasing the advantages of a unified task representation in the context of autonomous driving applications." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.757, + 0.39, + 0.773 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "title", + "bbox": [ + 0.214, + 0.787, + 0.707, + 0.803 + ], + "angle": 0, + "content": "2.1 Camera-based 3D Perception in Autonomous Driving" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.842 + ], + "angle": 0, + "content": "3D object detection [7,9,12,62,70] and 3D lane detection [11,14] are two common 3D perception tasks, and previous approaches usually perform the two tasks" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.346, + 0.127 + ], + "angle": 0, + "content": "C. Li et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.283 + ], + "angle": 0, + "content": "independently. 3D object detection has been extensively studied either in model architectures [7, 37, 38, 59] or feature representations [17, 25, 27, 28, 49, 50, 71, 76]. Early methods [20, 22, 45] predict 3D bounding boxes using 2D detectors, and Mono3D [7] is one of the first methods directly make predictions from 3D representations. Motivated by DETR [11], DETR3D [62] uses sparse 3D object centers as queries, and PETR [37] further enhances its feature representations by embedding 3D positions. Some recent approaches [17, 28, 38, 58] also focus on incorporating temporal information into 3D object detection, but they all follow the transformer-based [56] paradigm introduced by DETR3D [62]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.284, + 0.788, + 0.419 + ], + "angle": 0, + "content": "Only recently have DETR-based methods been introduced for 3D lane detection [3, 38, 42, 63, 64]. Previous methods [6, 11, 14, 34, 57] prefer to project image features into BEV space using Inverse Perspective Mapping (IPM), but this introduces prohibitive height prediction error. Some work [67] avoids IPM via depth estimation, and Anchor-3DLane [18] reverses this projection process. Following DETR [3, 38, 61] introduce 3d lane anchors as queries, and [42] adopts finer lane prior queries and is end-to-end. With the advances of DETR [61, 62] networks in both tasks [3, 28, 37, 42], it is architecturally ready to unify the two tasks without standalone models." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.447, + 0.473, + 0.463 + ], + "angle": 0, + "content": "2.2 Multi-task 3D Perception" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.477, + 0.788, + 0.627 + ], + "angle": 0, + "content": "The problem of multi-tasking in autonomous driving perception is still in its infancy. Recent advances [19, 28, 37-39, 48, 65, 75] elaborate on universal perception features and multiple tasks are handled with different heads. Moreover, there is no consensus on the required tasks. For example, MMF [29] addresses both depth completion and object detection issues, while [19, 28, 38, 39] performs object detection and map segmentation, with [38] performing additional 3D lane detection separately. However, most of these studies only use multiple tasks to demonstrate the universality of their learned features, and the tasks are trained and evaluated separately [28, 38] rather than simultaneously [19, 39], not to mention dataset discrepancies." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.629, + 0.788, + 0.841 + ], + "angle": 0, + "content": "In addition, multitask learning (MTL) typically involves task conflict. Previous MTL studies [55] addresses this either through parameter sharing [13, 44, 47, 51, 54, 66] or optimization [8, 21, 33, 35, 52, 73]. Parameter sharing approaches are conceptually straightforward, and are further classified into hard-sharing (with shared and task-specific parameters) and soft-sharing (with a cross-task talk mechanism). Optimization-based approaches, e.g. IMTL [33] and DWA [35], attribute MTL imbalance to the gradient and loss imbalance and explore optimization calibration methods. Taking advantage of these advances, FULLER [19] stands as the pioneering work in autonomous driving perception that analyzes the performance of modern optimization-based MTL calibration methods [8, 33, 35]. With a unified task representation, we follow a different paradigm from previous ones, that eliminate task-specific heads and mitigate the gradient imbalance. Our approach can be seen as a hard-sharing technique without task-specific parameters and does not require calibration." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.451, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "RepVF: Unified Multi-task 3D Perception" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.145, + 0.533, + 0.163 + ], + "angle": 0, + "content": "3 Representative Vector Fields" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.177, + 0.788, + 0.268 + ], + "angle": 0, + "content": "We begin by revisiting the representations of two common 3D perception tasks: 3D object detection [4, 12, 53] and 3D lane detection [6, 67]. By exploring how task-specific representations have evolved into task-specific designs [3, 6, 7, 11, 37, 38, 42, 50, 62], we propose the task-agnostic \\( RepVF \\) (Representative Vector Fields), which arise from the geometric commonality in the tasks and effectively reduce task competition via unified representations." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.289, + 0.662, + 0.305 + ], + "angle": 0, + "content": "3.1 Task-specific Representations and Single-tasking" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.313, + 0.789, + 0.45 + ], + "angle": 0, + "content": "Representation Formulations. Task-specific representations vary due to the geometric features of the task targets, which differ in dimensions and scales. As a key task for understanding 3D space, 3D object detection deals with objects irregularly shaped and scaled from one to several metres, defined by coarse outer boundaries as 3D bounding boxes [12], formally a 7-d coordinate \\( B = (x,y,z,l,w,h,\\theta) \\) encoding object center \\( (x,y,z) \\), box dimensions \\( (l,w,h) \\) and heading \\( \\theta \\). Conversely, due to their linear shape and ability to cover long distances, 3d lanes are not described by boundaries but by direct geometries, typically ordered sequences of \\( N_{L} \\) 3D point coordinates [6]:" + }, + { + "type": "equation", + "bbox": [ + 0.335, + 0.46, + 0.786, + 0.477 + ], + "angle": 0, + "content": "\\[\nL = \\left[ \\left(x _ {1}, y _ {1}, z _ {1}\\right), \\left(x _ {2}, y _ {2}, z _ {2}\\right), \\dots , \\left(x _ {N _ {l}}, z _ {N _ {l}}, z _ {N _ {l}}\\right) \\right] \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.486, + 0.788, + 0.548 + ], + "angle": 0, + "content": "This may look familiar to an unordered point cloud, but they are ordered along the direction of the road. In addition, 3D perception problems typically involve classifications that require category scores \\( \\mathcal{C}_i \\), varying from 1, 3 or 23 [4,12,53] classes (3D object) to 21 or 22 [6] classes (3D lane)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.568, + 0.788, + 0.704 + ], + "angle": 0, + "content": "Evolving into Task-specific Designs. Pioneering 3D object detection from RGB images is proposal based, e.g. Mono3D [7] reduces proposal space dimension of each class to \\((x,y,z,\\theta,t)\\) with representative templates, and only predicts \\(\\varDelta_{\\mathrm{pos}}\\); OFT-Net [50] predicts offsets of all bounding box parameters relative to the top-view grid, i.e. \\(\\varDelta_{\\mathrm{pos}},\\varDelta_{\\mathrm{dim}},\\varDelta_{\\mathrm{ang}}\\). Taking advantage of the center coordinates, DETR-based methods [37,62] regress \\(\\varDelta_{\\mathrm{pos}}\\) along with bounding box parameters. However, constrained by the limited spatial extent of center points, extracted features around them capture only local spatial areas [77], failing to capture the finer pose and structure of objects." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.705, + 0.789, + 0.841 + ], + "angle": 0, + "content": "3D lines are suitable for anchor-based representations [6, 11, 38], resulting in equally spaced longitudinal lines and anchor-based predictions \\(\\varDelta_{\\mathrm{xz}}\\). Anchor representations reduce the representation dimension, but also restrict the degrees of freedom, limiting the maximum model capacity, which has led some works [6, 18] focus on more sophisticated anchor representations. Recent DETR-based methods [3,38,42] use anchor coordinates as queries, exploring the full potential of anchors. Unfortunately, since it's not trivial and stable to directly predict parameter-driven curves [10, 15], mainstream methods stick to predicting anchor-based offsets \\(\\varDelta_{\\mathrm{xz}}\\), and nearly no one ever worked directly on the raw lane format." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.347, + 0.128 + ], + "angle": 0, + "content": "C. Li et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.284 + ], + "angle": 0, + "content": "Intersection of Task-specific Representations. We believe that it's already been architecturally ready to unify the two (and possibly more) perception tasks, with the population and success of DETR practices [19,28,38,61]. However, they still rely on multi-head design due to the discrepancy in task representation, which consumes more computational resources. Despite the significant differences in the dynamics and scales of the described targets, these task-specific representations only encapsulate descriptions of shapes, structures, and directions (object heading and lane ordinality). It is plausible to represent these task targets universally, taking into account perception hierarchies." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.301, + 0.684, + 0.317 + ], + "angle": 0, + "content": "3.2 The Task-Agnostic Representation through RepVF" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.322, + 0.788, + 0.414 + ], + "angle": 0, + "content": "To facilitate efficient design in multitasking frameworks, we introduce \\( RepVF \\), a vector fields representation inspired by 2D point set object representation [43,43]. Distinct from conventional isotropic points, \\( RepVF \\) assigns a vector to each location making itself anisotropic. We define \\( RepVF \\) as a collection of vector fields, each field denoted by \\( R_{i} \\) as a perception target, whose semantic vector assignment mapping \\( \\mathcal{F}_i \\) is defined on \\( D = \\bigcup_{i=1}^{n_s} S_i \\subseteq \\mathbb{R}^3 \\):" + }, + { + "type": "equation", + "bbox": [ + 0.331, + 0.419, + 0.785, + 0.436 + ], + "angle": 0, + "content": "\\[\n\\{R _ {i} \\}, \\quad R _ {i} = \\left(S _ {i}, \\mathcal {F} _ {i} \\left(S _ {i}\\right)\\right) \\subseteq \\mathbb {R} ^ {3 + d}, \\quad \\mathcal {F}: D \\rightarrow \\mathbb {R} ^ {d} \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.441, + 0.788, + 0.547 + ], + "angle": 0, + "content": "where \\(d\\) is the dimension of the target vector domain. Therefore, for any point \\(x\\) within \\(D\\), \\(\\mathcal{F}\\) assigns it a vector \\(\\mathcal{F}(x) \\in \\mathbb{R}^d\\) that encapsulates dimensional and positional attributes of the point. This definition allows us to cover the whole space: The sampling domain \\(D\\) is partitioned into \\(n_s\\) disjoint subsets \\(S_i\\), with each subset containing \\(n\\) representative sampling points: \\(S_i = \\{(x_k, y_k, z_k)\\}_{k=1}^n\\). Accordingly, the collective vector field \\(\\mathcal{F}\\) sprawls over \\(D = \\bigcup_{i=1}^{n_s} S_i\\), articulated into \\(n_s\\) sub-vector fields \\(\\mathcal{F}_i\\) catering to specific tasks." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.548, + 0.788, + 0.64 + ], + "angle": 0, + "content": "Intuitively, \\( S_{i} \\) directly outlines entity-level geometric structures, and \\( \\mathcal{F}_i(S_i) \\) delineates contextual characteristics (i.e. \\( \\theta \\) for bounding box \\( B \\), or adjacent point directions for lane \\( L \\)). Considering the directionality, we use unit-length vectors parallel to the x-y plane to capture angular behavior, resulting in behavior akin to scalars and \\( d = 1 \\). To represent proper \\( \\mathcal{F}_i \\) for 3d lanes, directions of the penultimate points are employed for terminal lane line points." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.656, + 0.788, + 0.732 + ], + "angle": 0, + "content": "Converting RepVF into Task-Specific Representation. To leverage existing task labels for training and evaluation, similar to object detection representations based on point sets [43, 68], we convert predictions \\( RepVF \\hat{R}_i \\), category \\( \\hat{C}_i \\) to task-specific representations. We define geometric interpretation processes as differentiable, parameter-free transformation functions:" + }, + { + "type": "equation", + "bbox": [ + 0.379, + 0.737, + 0.785, + 0.755 + ], + "angle": 0, + "content": "\\[\n\\mathcal {T} ^ {(\\cdot)}: \\hat {R} _ {i} \\rightarrow \\hat {P} _ {i}, \\quad \\text {s . t .} \\hat {P} _ {i} \\in \\{\\hat {B}, \\hat {L} \\} _ {i} \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.759, + 0.788, + 0.843 + ], + "angle": 0, + "content": "Specifically, we have \\(\\mathcal{T}^l:\\hat{R}_i\\to \\hat{L}_i\\) for 3d lanes and \\(\\mathcal{T}^b:\\hat{R}_i\\rightarrow \\hat{B}_i\\) for 3d objects. For semantic parts, to align \\(\\hat{\\mathcal{C}}_i\\) with task-specific classes, we split it into \\(n_{(\\cdot)}\\) bins and take max score within each bin, where \\(n_{(\\cdot)}\\) is the count of task categories. \\(\\hat{R}_i\\) and \\(\\hat{R}_j\\) remain formally equivalent until contextualized into \\(\\hat{P}_i\\) and \\(\\hat{P}_j\\), thus maintaining task agnosticism." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.451, + 0.115, + 0.732, + 0.13 + ], + "angle": 0, + "content": "RepVF: Unified Multi-task 3D Perception" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.147, + 0.788, + 0.208 + ], + "angle": 0, + "content": "Conversion Function Choice. In our implementation, \\(\\mathcal{T}^l\\) sorts the unordered points of \\(\\hat{S}_i\\) by the road direction \\((+X\\) in Waymo [53]). Unlike 2D boxes, 3D boxes are not strictly aligned with to \\(XYZ\\) axes due to headings. Therefore, \\(\\mathcal{T}^b\\) consists three parts that respectively compute heading, box center and dimensions:" + }, + { + "type": "equation", + "bbox": [ + 0.28, + 0.216, + 0.786, + 0.235 + ], + "angle": 0, + "content": "\\[\n\\mathcal {T} _ {h} ^ {b}: \\hat {\\mathcal {F}} _ {i} \\rightarrow \\hat {\\theta} _ {i}; \\quad \\mathcal {T} _ {x y z} ^ {b}: \\hat {S} _ {i} \\rightarrow (\\hat {x} _ {i}, \\hat {y} _ {i}, \\hat {z} _ {i}); \\quad \\mathcal {T} _ {l w h} ^ {b}: \\hat {S} _ {i} \\rightarrow (\\hat {l} _ {i}, \\hat {w} _ {i}, \\hat {h} _ {i}) \\qquad (4)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.245, + 0.788, + 0.356 + ], + "angle": 0, + "content": "Here, \\(\\mathcal{T}_h^b\\) and \\(\\mathcal{T}_{xyz}^b\\) are simple arithmetic averages. With \\(\\hat{\\theta}_i\\) given by \\(\\mathcal{T}_h^b\\), by projecting \\(\\hat{S}_i\\) to the main orientation axis \\(\\hat{\\theta}_i\\), the predictions are aligned with the predicted directions. We then proceed with two candidates of \\(\\mathcal{T}_{lwh}^b\\) similar to 2D practices [68]: (1) \\(\\mathcal{T}_{lwh}^b = \\mathcal{T}_1\\), the min-max function, performing min/max operations projection dimensions for dimensions; (2) \\(\\mathcal{T}_{lwh}^b = \\mathcal{T}_2\\) employs a momentum-based function taking the variance of predicted points as the boundary box dimensions. In our practice, \\(\\mathcal{T}_2\\) works slightly better." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.376, + 0.789, + 0.513 + ], + "angle": 0, + "content": "Learning RepVF. RepVF is supervised by loss functions universally across tasks without special losses. Here we use \\(\\mathbf{L}^1\\) regression loss on \\(S_{i}\\) and \\(\\mathcal{F}_i\\), and focal loss [32] on \\(\\mathcal{C}_i\\). Smaller structures are described by points in \\(S_{i}\\), and \\(R_{i}\\) itself is set-level, extending RepVF across different perception hierarchies. To capture this cross-level property, we adopt DETR-based structure [37,61,62] and aligned query design to use raw representations motivated by [3, 36]. With multiple decoder layers [62, 77], the learning process of RepVF can be characterized as iterative update \\(\\varOmega_l\\) on queries \\(Q_{l}\\) and inferring \\(\\hat{R}_l,\\hat{C}_l\\) from them with linear layer \\(\\varPhi\\) and decoder layer \\(l\\):" + }, + { + "type": "equation", + "bbox": [ + 0.387, + 0.522, + 0.785, + 0.539 + ], + "angle": 0, + "content": "\\[\nQ _ {l} = \\Omega_ {l} \\left(Q _ {l - 1}\\right), \\quad \\hat {R} _ {l}, \\hat {\\mathcal {C}} _ {l} = \\Phi (Q _ {l}) \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.548, + 0.629, + 0.563 + ], + "angle": 0, + "content": "We will explain our network in detail in the next section." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.585, + 0.611, + 0.602 + ], + "angle": 0, + "content": "4 RFTR: The Single-head Multi-tasker" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.615, + 0.788, + 0.69 + ], + "angle": 0, + "content": "In this section, we introduce the single-head multi-task framework RFTR (Representative Vector Fields Transformer), which uses \\( RepVF \\) representation to unify different tasks and exploit intrinsic connections among them. We also observed one interesting behavior of RFTR in the balance of multi-task gradients. RFTR is built upon DETR networks [5,37,62] and is end-to-end on all tasks." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.712, + 0.419, + 0.725 + ], + "angle": 0, + "content": "4.1 Network Structure" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.734, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Figure 2 shows our RFTR architecture built upon \\( RepVF \\). It inputs images \\( I \\) from \\( N \\) cameras with known camera parameters, and outputs unified representative vector fields \\( \\mathcal{F} \\) for various perception tasks in the 3D scene. RFTR contains four main components that extract features from images, generate set-level queries for hierarchical representation structure, solve unified representative vector fields and multiple tasks with one single head, and convert results to task representations in a differentiable way to utilize existing task labels." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.347, + 0.127 + ], + "angle": 0, + "content": "C. Li et al." + }, + { + "type": "image", + "bbox": [ + 0.219, + 0.145, + 0.784, + 0.337 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.348, + 0.788, + 0.446 + ], + "angle": 0, + "content": "Fig. 2: Overview of our RFTR (Representative Vector Fields Transformer) built upon RepVF. Multi-view image features with 3D position embeds are extracted by image backbone and fed into the decoder. We generate set-level query embeds from 3D space sampling sets, each representing a perception target. One single unified task head is then used to predict unified representative vector fields. Finally, predictions are transformed in a differentiable manner into task specific representations to utilize existing labels for supervision." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.476, + 0.788, + 0.538 + ], + "angle": 0, + "content": "Feature Extraction. Following common practices in utilizing 2D backbones to extract features [28,37,62], we adopt ResNet [16] (ResNet-50) and FPN [31] to produce 2D features \\(F^{2d}\\). They're encoded with 3D position embedding in PETR [37] into 3D position aware features \\(F^{3d}\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.56, + 0.788, + 0.65 + ], + "angle": 0, + "content": "Generating Set-level Perception Queries. As \\( RepVF \\) is sampled from \\( n_s \\) point sets \\( S_i \\) with \\( n \\) points each and is two-level, our aim is to generate queries that are consistent with this hierarchical point-set structure. Hence provide the decoder with sufficient spatial information, improving feature interactions. Unlike former approaches [37, 38, 42, 62] that encode one single 3D location, each query in RFTR encodes multiple locations simultaneously." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.651, + 0.788, + 0.71 + ], + "angle": 0, + "content": "We generate set-level queries \\( Q_0 \\in \\mathbb{R}^{n_s \\times d_q} \\) from initial point sets corresponding to \\( S_i \\), where \\( d_q \\) is the dimension of each query. The generation process \\( \\mathcal{G} \\) consists of encoding initialized positions from the point level, and embedding them at the set level:" + }, + { + "type": "equation", + "bbox": [ + 0.408, + 0.711, + 0.786, + 0.728 + ], + "angle": 0, + "content": "\\[\n\\mathcal {G} = \\operatorname {e m b e d} (\\text {F l a t t e n (P E)}) \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.734, + 0.788, + 0.81 + ], + "angle": 0, + "content": "where PE is the positional encoding [56]. Point-level PE is flattened into set-level and embedded by a small MLP with two layers. To avoid confusing the model with expansive locations falling on multiple targets in the early stages of training, we use \\( n \\) identical 3D locations sampled with 0-1 uniform distribution for all \\( n_s \\) initial point sets." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.788, + 0.842 + ], + "angle": 0, + "content": "To our best knowledge, the closest generation approach to ours is [3], which generates query from an ordered sequence of anchor points motivated by [36]." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.451, + 0.115, + 0.732, + 0.13 + ], + "angle": 0, + "content": "RepVF: Unified Multi-task 3D Perception" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.787, + 0.128 + ], + "angle": 0, + "content": "9" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.222 + ], + "angle": 0, + "content": "However, ours is from an unordered sampling point set and is not task-specific. Moreover, we use a tiny 16-dimensional PE for each coordination, divert from former methods [37, 62] with a much larger 256-dimension. In our multi-task setting, queries are equivalently initialized, thus task-agnostic. We divide the first \\(80\\%\\) as 3D objects and the last \\(20\\%\\) as 3D lanes." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.241, + 0.788, + 0.348 + ], + "angle": 0, + "content": "Single-head Multi-tasking. Similar to all previous approaches [5, 37, 62], queries are learnable and updated in decoder layers iteratively. The unified head \\(\\varPhi\\) of RFTR is composed by a multi-layer transformer decoder [56], and two branches \\(\\varPhi_{g}\\) and \\(\\varPhi_{s}\\) dedicated for geometric regression and semantic prediction respectively to predict \\(RepVF\\). We use the same \\(n_{L} = 6\\) decoder layers with previous methods [5, 37, 62] to exploit intrinsic tasks connections implicitly. The interaction process within these layers can be formally expressed as:" + }, + { + "type": "equation", + "bbox": [ + 0.374, + 0.353, + 0.786, + 0.371 + ], + "angle": 0, + "content": "\\[\nQ _ {l} = \\Omega_ {l} \\left(F ^ {3 d}, Q _ {l - 1}\\right), \\quad l = 1, \\dots , N _ {L} \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.375, + 0.786, + 0.405 + ], + "angle": 0, + "content": "With unified head \\(\\varPhi=\\{\\varPhi_{g},\\varPhi_{s}\\}\\), predictions \\(\\hat{R}_{li}\\) and \\(\\hat{C}_{li}\\) is decoded from \\(i\\)-th query at \\(l\\)-th layer:" + }, + { + "type": "equation", + "bbox": [ + 0.397, + 0.406, + 0.786, + 0.422 + ], + "angle": 0, + "content": "\\[\n\\hat {R} _ {l i} = \\Phi_ {g} \\left(Q _ {l i}\\right), \\quad \\hat {\\mathcal {C}} _ {l i} = \\Phi_ {s} \\left(Q _ {l i}\\right) \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.425, + 0.787, + 0.457 + ], + "angle": 0, + "content": "For \\(\\varPhi_{g}\\), additional offsets are added to produce \\(S_{li}\\). Unified geometric predictions \\(\\hat{R}_l = (\\hat{S}_l,\\hat{F}_l)\\subset \\mathbb{R}^{n_s\\times (3 + d_p)}\\) are converted to task-specific representations:" + }, + { + "type": "equation", + "bbox": [ + 0.349, + 0.463, + 0.786, + 0.482 + ], + "angle": 0, + "content": "\\[\n\\hat {L} _ {l i} = \\mathcal {T} ^ {l} (\\hat {R} _ {l i}), \\quad \\hat {B} _ {l i} = \\left(\\mathcal {T} _ {x y z} ^ {b}, \\mathcal {T} _ {\\theta} ^ {b}, \\mathcal {T} _ {l w h} ^ {b}\\right) (\\hat {R} _ {l i}) \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.488, + 0.788, + 0.549 + ], + "angle": 0, + "content": "Here, \\(\\mathcal{T}^{(\\cdot)}\\) are differentiable functions predefined in section 3.2. Each head branch \\(\\varPhi_{(\\cdot)}\\) is a simple small MLP network with three fully connected layers. Following DETR3D [62], we weight the loss for the predictions from each decoder layer \\(\\Omega_l\\) during training, and only outputs from the last layer is used during inference." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.567, + 0.788, + 0.627 + ], + "angle": 0, + "content": "Loss Function. RFTR uses existing task labels for supervision, and its loss is calculated between converted predictions and task ground truths. It consists of the following components for coordinate regression, vector fields regression and classification respectively:" + }, + { + "type": "equation", + "bbox": [ + 0.323, + 0.634, + 0.786, + 0.665 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} ^ {(\\cdot)} = w _ {r} ^ {(\\cdot)} \\mathcal {L} _ {r} ^ {(\\cdot)} + w _ {f} ^ {(\\cdot)} \\mathcal {L} _ {f} ^ {(\\cdot)} + w _ {c} ^ {(\\cdot)} \\mathcal {L} _ {c} ^ {(\\cdot)}, \\quad \\mathcal {L} = \\sum_ {\\text {t a s k s}} \\mathcal {L} ^ {(\\cdot)} \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.672, + 0.787, + 0.717 + ], + "angle": 0, + "content": "where \\( w^{(\\cdot)} \\) represent loss component weights for task \\( (\\cdot) \\), i.e. 3D object detection and 3D lane detection. We use the Hungarian algorithm [23] for label assignment after cost calculation between ground truth and converted predictions \\( \\hat{B}_{li}, \\hat{P}_{li} \\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.717, + 0.788, + 0.841 + ], + "angle": 0, + "content": "In practice, we choose \\( w_{c}^{b} = w_{c}^{l} = 2.0 \\), \\( w_{r}^{l} = 0.003 \\), \\( w_{r}^{b} = 0.1 \\), \\( w_{f}^{b} = 0.2 \\) and \\( w_{f}^{l} = 0.032 \\). Coordinate regression and classification weights are selected empirically to equalize the scales and contributions of each loss term [74], fields weights are empirically selected. We use \\( \\mathrm{L}^1 \\) loss for \\( R_{i} \\) regression, and focal loss [32] with \\( \\gamma = 2.0 \\) and \\( \\alpha = 0.25 \\) for classification. Since we predict a degenerated scalar field, we supervise the angular prediction by its sine and cosine values. All tasks share the same losses and similar weights, no additional, task-specific loss measures are adopted." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.347, + 0.127 + ], + "angle": 0, + "content": "C. Li et al." + }, + { + "type": "image", + "bbox": [ + 0.256, + 0.15, + 0.56, + 0.285 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.293, + 0.291, + 0.52, + 0.303 + ], + "angle": 0, + "content": "(a) Gradient disparities by iteration index." + }, + { + "type": "image", + "bbox": [ + 0.572, + 0.15, + 0.747, + 0.284 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.573, + 0.291, + 0.744, + 0.303 + ], + "angle": 0, + "content": "(b) Average gradient disparities." + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.313, + 0.788, + 0.397 + ], + "angle": 0, + "content": "Fig. 3: Our single-head RFTR model shows a reduced gradient imbalance and better disparity stability (mean: 2.47, variance: 0.90) compared to the multi-head baseline (mean: 2.98 variance: 1.86) or the multi-head baseline with RepVF (mean: 2.72, variance: 1.57). Ideally balanced gradient disparity through training iterations should approach 2. For ease of trend observation, we have clipped the top \\(2\\%\\) values to the mean, with curves smoothed through a window size of 50 1-d convolution." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.424, + 0.646, + 0.439 + ], + "angle": 0, + "content": "4.2 Gradient Balance in Single-head Multi-tasking" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.445, + 0.784, + 0.52 + ], + "angle": 0, + "content": "The multi-head multi-task approach often leads to task conflict, attributed in prior studies [21, 33, 46, 52] to gradient imbalance. These studies [33, 46] and more recent work [19] have explored appropriate gradient calibration strategies to address this issue. As we will show, our single-head RFTR naturally achieves better balanced multi-task gradients thus calibration may no longer be required." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.521, + 0.785, + 0.566 + ], + "angle": 0, + "content": "To elucidate the advantage of our single-head multi-task paradigm over multi-head multi-task settings, we use an improved measure of gradient disparity adapted from FULLER [19], by taking into account the symmetry disparity:" + }, + { + "type": "equation", + "bbox": [ + 0.373, + 0.572, + 0.785, + 0.607 + ], + "angle": 0, + "content": "\\[\n\\operatorname {d i f f} \\left(\\nabla \\mathcal {L} ^ {l}, \\nabla \\mathcal {L} ^ {b}\\right) = \\frac {\\| \\nabla \\mathcal {L} ^ {l} \\|}{\\| \\nabla \\mathcal {L} ^ {b} \\|} + \\frac {\\| \\nabla \\mathcal {L} ^ {b} \\|}{\\| \\nabla \\mathcal {L} ^ {l} \\|} \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.613, + 0.784, + 0.674 + ], + "angle": 0, + "content": "where \\(\\| \\cdot \\|\\) is the Frobenius norm and \\(\\nabla \\mathcal{L}^{(\\cdot)}\\) denotes gradients. Given that \\(\\| \\nabla \\mathcal{L}^{(\\cdot)}\\|\\) is always greater than 0, the value of diff function is always greater than or equal to 2. Therefore, the more stable and closer to 2 the gradient disparity is during model training, the better the multi-task stability indicated." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.675, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Figure 3a illustrates the gradient disparity curves through training iterations. The multi-head multi-task method [37,38] is under the same task settings with ours. In the sampling 1k iterations, the stability and absolute scale of the gradient disparity curve for the multi-head configuration are inferior to our single-head multi-task structure. The mean and variance of the multi-head disparities are about \\(108.5\\%\\) and \\(107.8\\%\\) higher, respectively, than our single-headed RFTR approach, as shown in Fig. 3b. We have also observed that unified representation and single head design contributed to the improvement about equally. More importantly, this means that our single-head RFTR mitigates, though not eliminating, the gradient imbalance between tasks, which is fundamentally different from task gradient balancing strategies that do not change the gradient itself." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.451, + 0.115, + 0.732, + 0.13 + ], + "angle": 0, + "content": "RepVF: Unified Multi-task 3D Perception" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.128 + ], + "angle": 0, + "content": "11" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.145, + 0.377, + 0.163 + ], + "angle": 0, + "content": "5 Experiments" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.177, + 0.788, + 0.209 + ], + "angle": 0, + "content": "We evaluate our method on the Waymo Open Dataset [53] extended by 3D lanes annotated in OpenLane [6], so that we train and evaluate tasks simultaneously." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.23, + 0.335, + 0.244 + ], + "angle": 0, + "content": "5.1 Datasets" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.256, + 0.789, + 0.438 + ], + "angle": 0, + "content": "Waymo Open Dataset and OpenLane. Waymo Open Dataset (Perception) [53] is a large-scale multimodal dataset composed of data from 5 cameras(front and sides), 1 mid-range lidar, and 4 short-range lidars. It contains 1000 sequences in total(about 198k samples), a training set of 798 sequences and a validation set of 202 sequences. OpenLane [6] is a comprehensive real-world 3D lane detection dataset built upon the Waymo Open dataset [53]. It includes a vast array of 200K frames and over 880K carefully annotated lanes, making it one of the largest of its kind to date. The dataset encompasses complex lane structures and features a variety of scene tags, such as weather and locations. Designed to emulate real-world scenarios, OpenLane provides a challenging benchmark for advanced lane detection algorithms. We integrate its 3D lane labels with the original Waymo Open sensor data for multi-task annotation." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.46, + 0.79, + 0.673 + ], + "angle": 0, + "content": "Alignment of Datasets. To achieve both tasks simultaneously, we align both datasets in terms of coordinate systems and data splits. (1) Coordinate systems. For the vehicle frame, OpenLane [6] uses a vehicle coordinate system that corresponds to the \"y-front, x-right, z-up\" positioning of 3D-LaneNet [11]. We then align this system to coincide with the commonly utilized \"x-front, y-right, z-up\" axis for lidar-based 3D object detection. With respect to the sensor frame, we apply transformations to the camera intrinsics so that the sensor frames for all cameras conform to a uniform \"z-depth, x-right, y-down\" frame. The predictions are then transformed back into their respective task-based coordinate systems for evaluation. (2) Data splitting. The OpenLane [6] dataset provides a 1000 segment full version and a 300 segment smaller subset (\\(30\\%\\)). The partitioning of the larger version complements the original train/Validation split of the Waymo Open Dataset. The subset version partitioning is aligned with OpenLane. We perform ablation studies on the smaller subset." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.694, + 0.458, + 0.71 + ], + "angle": 0, + "content": "5.2 Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.719, + 0.789, + 0.84 + ], + "angle": 0, + "content": "Data Processing. We use the v1.4.2 of Waymo Open Dataset [53], and v1.2 for OpenLane [6]. We set the perception range to \\([-84.88m, 84.88m]\\) for the X and Y axis, and \\([-10m, 10m]\\) for the Z axis. 3D objects that do not fall within the perception range, are invisible to the cameras, or do not have LiDAR points, are filtered. For 3D lanes, visibility filtering is first applied after axis transformation, then 3D lanes outside the perception range are pruned and then resampled to a fixed size. Approximately \\(2.07\\%\\) of the frames without annotations after filtering are excluded from use." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.346, + 0.127 + ], + "angle": 0, + "content": "C. Li et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.788, + 0.187 + ], + "angle": 0, + "content": "Table 1: Comparison of recent models on 3D lane detection using the OpenLane [6] validation set. The down arrow indicates that lower metric values correspond to better performance, and vice versa. Red indicates the best result, and blue the second-best." + }, + { + "type": "table", + "bbox": [ + 0.252, + 0.199, + 0.761, + 0.359 + ], + "angle": 0, + "content": "
MethodsF-Score ↑Category Accuracy ↑X error (m) ↓Z error (m) ↓
nearfarnearfar
3D-LaneNet [11] [ICCV19]44.1-0.4790.5720.3670.443
Gen-LaneNet [14] [ECCV20]32.3-0.5910.6840.4110.521
PersFormer [6] [ECCV22]50.592.30.4850.5530.3640.431
Cond-IPM [6]36.6-0.5631.0800.4210.892
CurveFormer [3] [ICRA23]50.5-0.3400.7720.2070.651
PETRv2-E [38] [ICCV23]51.9-0.4930.6430.3220.463
PETRv2-V-10 [38]57.8-0.4270.5820.2930.421
PETRv2-V-400 [38]61.2-0.4000.5730.2650.413
BEV-LaneDet [57] [CVPR23]58.4-0.3090.6590.2440.631
Anchor3DLane [18] [CVPR23]53.190.00.3000.3110.1030.139
SPG [69] [ICCV23]52.3-0.4680.5140.3710.418
RFTR (Ours)61.891.60.3410.4500.0730.107
" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.379, + 0.787, + 0.408 + ], + "angle": 0, + "content": "Table 2: Comparison of recent works of 3D object detection on the Waymo Open Dataset [53] val set. mAPs are LEVEL_1 in Waymo metrics." + }, + { + "type": "table", + "bbox": [ + 0.253, + 0.42, + 0.761, + 0.503 + ], + "angle": 0, + "content": "
MethodsOverallVehicleCyclistPedestrian
mAPmAPHmAPmAPHmAPmAPHmAPmAPH
DETR3D [62]10.19.69.29.115.815.55.34.1
DETR3D [62]8.68.018.118.02.41.65.34.4
PETR [37]20.919.731.130.819.517.612.110.6
RFTR (Ours)19.518.722.622.427.826.38.07.2
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.538, + 0.789, + 0.844 + ], + "angle": 0, + "content": "Model and Training. We used ResNet 50 [16] as the 2D image backbone for fair comparisons. Following PETR [37], the same upsampled and fused P4 feature with 1/16 input resolution is used as the 2D feature. A CPFPN [31] is used as the image neck. All 3D coordinates are normalized to \\([0,1]\\). The same weights are used for both the loss computation and the bipartite matching. We adopt progressive weighting for the loss of each decoder layer to improve the convergence speed. For 3D object detection, we compute matching costs based on the center displacement and class score weightings. Dimensions are refined post-matching using the ground truth headings to increase training robustness. We use a total of \\( n_s = 750 \\) queries with \\( n = 15 \\) sample locations each for the two tasks, 600 for 3D objects and 150 for 3D lanes. We trained our model using the AdamW [41] optimizer with a 0.01 weight decay. The learning rate started at \\( 2.0 \\times 10^{-4} \\) and was decayed with a cosine annealing policy [40]. We adopted a multi-scale training strategy, with the shorter side chosen randomly from [640,900] and the longer side not exceeding 1600. We applied rotation augmentations between \\( [-\\pi /18,\\pi /18] \\). We resized images of sizes \\( 1920 \\times 1080 \\) and \\( 1920 \\times 1280 \\) from different views to \\( 1600 \\times 900 \\) before augmenting. The final resolution for network inputs was \\( 1408 \\times 512 \\). All experiments on the full dataset are trained for 24 epochs on 4 GeForce RTX 4090 GPUs with a batch size of 8. No test time augmentation methods are used during inference." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.451, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "RepVF: Unified Multi-task 3D Perception" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "image", + "bbox": [ + 0.221, + 0.146, + 0.784, + 0.314 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.232, + 0.325, + 0.772, + 0.34 + ], + "angle": 0, + "content": "Fig. 4: Visualization of the RFTR model's performance and the learned \\( RepVF \\)." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.369, + 0.461, + 0.384 + ], + "angle": 0, + "content": "5.3 Comparison with SOTA" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.393, + 0.789, + 0.545 + ], + "angle": 0, + "content": "3D Lane Detection. As shown in Tab. 1, our RFTR model showcases significant achievements in 3D lane detection on the OpenLane validation set, demonstrating its competitive edge with a high F-score of 61.8 and category accuracy of \\(91.6\\%\\). It matches or exceeds state-of-the-art models such as PETRv2-V-400 [38] (more than \\(26\\times\\) denser representation than ours) when assessed on the OpenLane [6] validation set. Notably, while our model has a slightly higher X error in the \"far\" class than [18], it manages to lead with a higher F-score and near accurate Z errors. Moreover, our method yields notable performance improvements over the Persformer [6] baseline, reflecting the advantageous adaptation of our unified RepVF representation and the RFTR architecture." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.566, + 0.789, + 0.688 + ], + "angle": 0, + "content": "3D Object Detection. The performance of our RFTR model in 3D object detection, as demonstrated in Tab. 2, shows comparable promising results on the Waymo Open Dataset [53] validation set. While our approach shows a commanding lead in detecting vehicles and cyclists with a solid L1 mAP and mAPH, it shows challenges in detecting smaller entities such as pedestrians. It should be noted, however, that these preliminary results indicate the great potential of using our unified perceptual model for complex tasks, which will require further fine-tuning for object classes with less prominent features." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.71, + 0.403, + 0.724 + ], + "angle": 0, + "content": "5.4 Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.734, + 0.789, + 0.84 + ], + "angle": 0, + "content": "Head design. We begin by finding out how head designs and representations contribute to the performance, in Tab. 3. The \\( RepVF \\) representation with the single-head design, i.e. RFTR, performs the best in general with 37.4M parameters. Unified representation brings performance improvements in multi-task learning and reduces task competition. In contrast, task-specific multi-head design produces inferior and unstably biased performance due to task competition, even with more parameters." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.347, + 0.128 + ], + "angle": 0, + "content": "C. Li et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.145, + 0.526, + 0.215 + ], + "angle": 0, + "content": "Table 3: Multi-task head design and representation ablation. Tasks are trained jointly on \\(30\\%\\) data of Waymo Open Dataset (L1 mAP) and OpenLane (F-Score). TS denotes task-specific representation. \\(\\dagger\\): lightweight models." + }, + { + "type": "table", + "bbox": [ + 0.234, + 0.227, + 0.511, + 0.321 + ], + "angle": 0, + "content": "
RepVFTSHeadParams↓F-Score↑Vehicle↑
✓Multi†37.1M49.620.1
✓Multi46.7M58.717.8
✓Multi†39.7M63.418.2
✓Multi49.2M66.422.1
✓Single37.4M66.525.3
" + }, + { + "type": "table_caption", + "bbox": [ + 0.531, + 0.149, + 0.788, + 0.22 + ], + "angle": 0, + "content": "Table 4: Results with different representations on 3D lane. 'Acc' for category accuracy and 'Error' for X error near. MT means simultaneous multitask 3D lane and 3D object detection." + }, + { + "type": "table", + "bbox": [ + 0.543, + 0.232, + 0.778, + 0.317 + ], + "angle": 0, + "content": "
RepresentationF-Score↑Acc↑Error↓
Anchor53.486.00.476
PS58.188.00.545
RepVF66.091.40.439
RepVF + MT66.591.70.420
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.351, + 0.788, + 0.442 + ], + "angle": 0, + "content": "Representation. Moreover, we have evaluated the effectiveness of different representations for 3D lane detection in Tab. 4, including traditional anchor representations [6,37], simple randomly sampled point set representation (PS), and our proposed representative vector fields (RepVF). This study reveals that RepVF significantly outperforms traditional anchor and PS methods, and incorporating multi-task learning marginally improves these metrics." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.47, + 0.422, + 0.485 + ], + "angle": 0, + "content": "5.5 Qualitative Results" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.501, + 0.789, + 0.653 + ], + "angle": 0, + "content": "Figure 4 demonstrates the capability of RFTR to simultaneously achieve 3D lane detection and 3D object detection on the extended [24] Waymo open dataset [53]. The figure clearly illustrates that RFTR can accurately predict RepVF and perform both tasks at once. In terms of 3D lane detection, RFTR is capable of predicting the shape of lane lines in areas obscured by vehicles, demonstrating its robustness against occlusion. Regarding 3D object detection, the unified RepVF representation accurately captures the 3D bounding boxes. This dual achievement highlights RFTR's effectiveness in handling complex driving scenarios, enhancing both navigation and safety features in autonomous driving systems." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.681, + 0.36, + 0.698 + ], + "angle": 0, + "content": "6 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.788, + 0.841 + ], + "angle": 0, + "content": "In this paper, we propose \\( RepVF \\), a novel unified vector fields representation that characterizes the geometric and semantic structure of different 3D perception targets in the scene. Built upon \\( RepVF \\), our proposed RFTR exploits the intrinsic connections between tasks, mitigates multi-task gradient imbalance and feature competition, and improves model convergence and multi-task expression. Our work not only contributes a novel single-head perspective to the multi-task learning paradigm in autonomous driving but also sets a promising direction for future research in efficient and effective 3D perception task handling." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.451, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "RepVF: Unified Multi-task 3D Perception" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.145, + 0.324, + 0.16 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.176, + 0.652, + 0.19 + ], + "angle": 0, + "content": "1. Mobileye at CES 2024. https://www-mobileye.com/ces-2024/" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.191, + 0.664, + 0.204 + ], + "angle": 0, + "content": "2. NVIDIA DRIVE Solutions. https://developer.nvidia.com/drive" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.205, + 0.787, + 0.259 + ], + "angle": 0, + "content": "3. Bai, Y., Chen, Z., Fu, Z., Peng, L., Liang, P., Cheng, E.: CurveFormer: 3D Lane Detection by Curve Propagation with Curve Queries and Attention. In: 2023 IEEE International Conference on Robotics and Automation (ICRA). pp. 7062-7068 (May 2023). https://doi.org/10.1109/ICRA48891.2023.10161160" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.261, + 0.786, + 0.327 + ], + "angle": 0, + "content": "4. Caesar, H., Bankiti, V., Lang, A.H., Vora, S., Liong, V.E., Xu, Q., Krishnan, A., Pan, Y., Baldan, G., Beijbom, O.: nuScenes: A Multimodal Dataset for Autonomous Driving. In: 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 11618-11628. IEEE, Seattle, WA, USA (Jun 2020). https://doi.org/10.1109/CVPR42600.2020.01164" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.329, + 0.786, + 0.398 + ], + "angle": 0, + "content": "5. Carion, N., Massa, F., Synnaeve, G., Usunier, N., Kirillov, A., Zagoruyko, S.: End-to-End Object Detection with Transformers. In: Vedaldi, A., Bischof, H., Brox, T., Frahm, J.M. (eds.) Computer Vision - ECCV 2020. pp. 213-229. Lecture Notes in Computer Science, Springer International Publishing, Cham (2020). https://doi.org/10.1007/978-3-030-58452-8_13" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.399, + 0.786, + 0.48 + ], + "angle": 0, + "content": "6. Chen, L., Sima, C., Li, Y., Zheng, Z., Xu, J., Geng, X., Li, H., He, C., Shi, J., Qiao, Y., Yan, J.: PersFormer: 3D Lane Detection via Perspective Transformer and the OpenLane Benchmark. In: Avidan, S., Brostow, G., Cisse, M., Farinella, G.M., Hassner, T. (eds.) Computer Vision - ECCV 2022, vol. 13698, pp. 550-567. Springer Nature Switzerland, Cham (2022). https://doi.org/10.1007/978-3-031-19839-7_32" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.481, + 0.786, + 0.522 + ], + "angle": 0, + "content": "7. Chen, X., Kundu, K., Zhang, Z., Ma, H., Fidler, S., Urtasun, R.: Monocular 3D Object Detection for Autonomous Driving. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 2147-2156 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.523, + 0.786, + 0.578 + ], + "angle": 0, + "content": "8. Chen, Z., Badrinarayanan, V., Lee, C.Y., Rabinovich, A.: GradNorm: Gradient Normalization for Adaptive Loss Balancing in Deep Multitask Networks. In: Proceedings of the 35th International Conference on Machine Learning. pp. 794-803. PMLR (Jul 2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.579, + 0.786, + 0.605 + ], + "angle": 0, + "content": "9. Cheng, W., Yin, J., Li, W., Yang, R., Shen, J.: Language-Guided 3D Object Detection in Point Cloud for Autonomous Driving (May 2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.606, + 0.786, + 0.645 + ], + "angle": 0, + "content": "0. Feng, Z., Guo, S., Tan, X., Xu, K., Wang, M., Ma, L.: Rethinking Efficient Lane Detection via Curve Modeling (May 2023). https://doi.org/10.48550/arXiv.2203.02431" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.647, + 0.786, + 0.701 + ], + "angle": 0, + "content": "1. Garnett, N., Cohen, R., Pe'er, T., Lahav, R., Levi, D.: 3D-LaneNet: End-to-End 3D Multiple Lane Detection. In: 2019 IEEE/CVF International Conference on Computer Vision (ICCV). pp. 2921–2930. IEEE, Seoul, Korea (South) (Oct 2019). https://doi.org/10.1109/ICCV.2019.00301" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.702, + 0.786, + 0.756 + ], + "angle": 0, + "content": "12. Geiger, A., Lenz, P., Urtasun, R.: Are we ready for autonomous driving? The KITTI vision benchmark suite. In: 2012 IEEE Conference on Computer Vision and Pattern Recognition. pp. 3354-3361 (Jun 2012). https://doi.org/10.1109/CVPR.2012.6248074" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.757, + 0.786, + 0.799 + ], + "angle": 0, + "content": "13. Guo, P., Lee, C.Y., Ulbricht, D.: Learning to Branch for Multi-Task Learning. In: Proceedings of the 37th International Conference on Machine Learning. pp. 3854-3863. PMLR (Nov 2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.799, + 0.786, + 0.84 + ], + "angle": 0, + "content": "14. Guo, Y., Chen, G., Zhao, P., Zhang, W., Miao, J., Wang, J., Choe, T.E.: GenLaneNet: A Generalized and Scalable Approach for 3D Lane Detection. In: Vedaldi, A., Bischof, H., Brox, T., Frahm, J.M. (eds.) Computer Vision - ECCV 2020. pp." + }, + { + "type": "list", + "bbox": [ + 0.22, + 0.176, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.347, + 0.127 + ], + "angle": 0, + "content": "C. Li et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.241, + 0.147, + 0.785, + 0.175 + ], + "angle": 0, + "content": "666-681. Lecture Notes in Computer Science, Springer International Publishing, Cham (2020). https://doi.org/10.1007/978-3-030-58589-1_40" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.175, + 0.785, + 0.202 + ], + "angle": 0, + "content": "15. Han, W., Shen, J.: Decoupling the Curve Modeling and Pavement Regression for Lane Detection (Sep 2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.203, + 0.785, + 0.242 + ], + "angle": 0, + "content": "16. He, K., Zhang, X., Ren, S., Sun, J.: Deep Residual Learning for Image Recognition. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 770-778 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.243, + 0.785, + 0.269 + ], + "angle": 0, + "content": "17. Huang, J., Huang, G.: BEVDet4D: Exploit Temporal Cues in Multi-camera 3D Object Detection (Jun 2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.27, + 0.785, + 0.323 + ], + "angle": 0, + "content": "18. Huang, S., Shen, Z., Huang, Z., Ding, Z.h., Dai, J., Han, J., Wang, N., Liu, S.: Anchor3DLane: Learning To Regress 3D Anchors for Monocular 3D Lane Detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 17451-17460 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.324, + 0.785, + 0.39 + ], + "angle": 0, + "content": "19. Huang, Z., Lin, S., Liu, G., Luo, M., Ye, C., Xu, H., Chang, X., Liang, X.: FULLER: Unified Multi-modality Multi-task 3D Perception via Multi-level Gradient Calibration. In: 2023 IEEE/CVF International Conference on Computer Vision (ICCV). pp. 3479-3488. IEEE, Paris, France (Oct 2023). https://doi.org/10.1109/ICCV51070.2023.00324" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.392, + 0.785, + 0.433 + ], + "angle": 0, + "content": "20. Kehl, W., Manhardt, F., Tombari, F., Ilic, S., Navab, N.: SSD-6D: Making RGB-Based 3D Detection and 6D Pose Estimation Great Again. In: Proceedings of the IEEE International Conference on Computer Vision. pp. 1521–1529 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.434, + 0.785, + 0.473 + ], + "angle": 0, + "content": "21. Kendall, A., Gal, Y., Cipolla, R.: Multi-Task Learning Using Uncertainty to Weigh Losses for Scene Geometry and Semantics. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 7482-7491 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.474, + 0.785, + 0.514 + ], + "angle": 0, + "content": "22. Ku, J., Pon, A.D., Waslander, S.L.: Monocular 3D Object Detection Leveraging Accurate Proposals and Shape Reconstruction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11867-11876 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.515, + 0.785, + 0.553 + ], + "angle": 0, + "content": "23. Kuhn, H.W.: The Hungarian method for the assignment problem. Naval Research Logistics Quarterly 2(1-2), 83-97 (1955). https://doi.org/10.1002/nav.3800020109" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.554, + 0.785, + 0.593 + ], + "angle": 0, + "content": "24. Li, L.H., Yatskar, M., Yin, D., Hsieh, C.J., Chang, K.W.: VisualBERT: A Simple and Performant Baseline for Vision and Language (Aug 2019). https://doi.org/10.48550/arXiv.1908.03557" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.595, + 0.785, + 0.649 + ], + "angle": 0, + "content": "25. Li, X., Yin, J., Li, W., Xu, C., Yang, R., Shen, J.: Di-v2x: Learning domain-invariant representation for vehicle-infrastructure collaborative 3d object detection. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 38, pp. 3208-3215 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.649, + 0.785, + 0.716 + ], + "angle": 0, + "content": "26. Li, X., Yin, J., Shi, B., Li, Y., Yang, R., Shen, J.: Lwsis: Lidar-guided weakly supervised instance segmentation for autonomous driving. Proceedings of the AAAI Conference on Artificial Intelligence 37(2), 1433-1441 (Jun 2023). https://doi.org/10.1609/aaii.v37i2.25228, https://ojs.aaii.org/index.php/AAAI/article/view/25228" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.717, + 0.785, + 0.757 + ], + "angle": 0, + "content": "27. Li, Y., Chen, Y., Qi, X., Li, Z., Sun, J., Jia, J.: Unifying Voxel-based Representation with Transformer for 3D Object Detection. Advances in Neural Information Processing Systems 35, 18442-18455 (Dec 2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.758, + 0.785, + 0.84 + ], + "angle": 0, + "content": "28. Li, Z., Wang, W., Li, H., Xie, E., Sima, C., Lu, T., Qiao, Y., Dai, J.: BEVFormer: Learning Bird's-Eye-View Representation from Multi-camera Images via Spatiotemporal Transformers. In: Avidan, S., Brostow, G., Cisse, M., Farinella, G.M., Hassner, T. (eds.) Computer Vision - ECCV 2022. pp. 1-18. Lecture Notes in Computer Science, Springer Nature Switzerland, Cham (2022). https://doi.org/10.1007/978-3-031-20077-9_1" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.451, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "RepVF: Unified Multi-task 3D Perception" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.19 + ], + "angle": 0, + "content": "29. Liang, M., Yang, B., Chen, Y., Hu, R., Urtasun, R.: Multi-Task Multi-Sensor Fusion for 3D Object Detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7345-7353 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.19, + 0.788, + 0.232 + ], + "angle": 0, + "content": "30. Liang, T., Xie, H., Yu, K., Xia, Z., Lin, Z., Wang, Y., Tang, T., Wang, B., Tang, Z.: BEVFusion: A Simple and Robust LiDAR-Camera Fusion Framework. Advances in Neural Information Processing Systems 35, 10421-10434 (Dec 2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.232, + 0.788, + 0.275 + ], + "angle": 0, + "content": "31. Lin, T.Y., Dollar, P., Girshick, R., He, K., Hariharan, B., Belongie, S.: Feature Pyramid Networks for Object Detection. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 2117-2125 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.276, + 0.788, + 0.317 + ], + "angle": 0, + "content": "32. Lin, T.Y., Goyal, P., Girshick, R., He, K., Dollar, P.: Focal Loss for Dense Object Detection. In: Proceedings of the IEEE International Conference on Computer Vision. pp. 2980-2988 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.318, + 0.788, + 0.36 + ], + "angle": 0, + "content": "33. Liu, L., Li, Y., Kuang, Z., Xue, J.H., Chen, Y., Yang, W., Liao, Q., Zhang, W.: Towards Impartial Multi-task Learning. In: International Conference on Learning Representations (Oct 2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.36, + 0.788, + 0.416 + ], + "angle": 0, + "content": "34. Liu, R., Chen, D., Liu, T., Xiong, Z., Yuan, Z.: Learning to Predict 3D Lane Shape and Camera Pose from a Single Image via Geometry Constraints. Proceedings of the AAAI Conference on Artificial Intelligence 36(2), 1765-1772 (Jun 2022). https://doi.org/10.1609/aaai.v36i2.20069" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.416, + 0.788, + 0.445 + ], + "angle": 0, + "content": "35. Liu, S., Johns, E., Davison, A.J.: End-to-End Multi-Task Learning with Attention (Apr 2019). https://doi.org/10.48550/arXiv.1803.10704" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.445, + 0.788, + 0.473 + ], + "angle": 0, + "content": "36. Liu, S., Li, F., Zhang, H., Yang, X., Qi, X., Su, H., Zhu, J., Zhang, L.: DAB-DETR: DYNAMIC ANCHOR BOXES ARE BETTER QUERIES FOR DETR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.473, + 0.788, + 0.543 + ], + "angle": 0, + "content": "37. Liu, Y., Wang, T., Zhang, X., Sun, J.: PETR: Position Embedding Transformation for Multi-view 3D Object Detection. In: Avidan, S., Brostow, G., Cisse, M., Farinella, G.M., Hassner, T. (eds.) Computer Vision - ECCV 2022, vol. 13687, pp. 531-548. Springer Nature Switzerland, Cham (2022). https://doi.org/10.1007/978-3-031-19812-0_31" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.543, + 0.788, + 0.572 + ], + "angle": 0, + "content": "38. Liu, Y., Yan, J., Jia, F., Li, S., Gao, A., Wang, T., Zhang, X.: PETRv2: A Unified Framework for 3D Perception from Multi-Camera Images" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.572, + 0.788, + 0.628 + ], + "angle": 0, + "content": "39. Liu, Z., Tang, H., Amini, A., Yang, X., Mao, H., Rus, D.L., Han, S.: BEVFusion: Multi-Task Multi-Sensor Fusion with Unified Bird's-Eye View Representation. In: 2023 IEEE International Conference on Robotics and Automation (ICRA). pp. 2774-2781 (May 2023). https://doi.org/10.1109/ICRA48891.2023.10160968" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.628, + 0.788, + 0.657 + ], + "angle": 0, + "content": "40. Loshchilov, I., Hutter, F.: SGDR: Stochastic Gradient Descent with Warm Restarts. In: International Conference on Learning Representations (Nov 2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.657, + 0.788, + 0.685 + ], + "angle": 0, + "content": "41. Loshchilov, I., Hutter, F.: DECOUPLED WEIGHT DECAY REGULARIZATION (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.686, + 0.788, + 0.728 + ], + "angle": 0, + "content": "42. Luo, Y., Zheng, C., Yan, X., Kun, T., Zheng, C., Cui, S., Li, Z.: LATR: 3D Lane Detection from Monocular Images with Transformer. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 7941-7952 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.728, + 0.788, + 0.797 + ], + "angle": 0, + "content": "43. Ma, Z., Wang, L., Zhang, H., Lu, W., Yin, J.: RPT: Learning Point Set Representation for Siamese Visual Tracking. In: Bartoli, A., Fusiello, A. (eds.) Computer Vision - ECCV 2020 Workshops. pp. 653-665. Lecture Notes in Computer Science, Springer International Publishing, Cham (2020). https://doi.org/10.1007/978-3-030-68238-5_43" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.798, + 0.788, + 0.84 + ], + "angle": 0, + "content": "44. Misra, I., Shrivastava, A., Gupta, A., Hebert, M.: Cross-Stitch Networks for Multi-Task Learning. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 3994–4003 (2016)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.347, + 0.127 + ], + "angle": 0, + "content": "C. Li et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.148, + 0.785, + 0.189 + ], + "angle": 0, + "content": "45. Mousavian, A., Anguelov, D., Flynn, J., Kosecka, J.: 3D Bounding Box Estimation Using Deep Learning and Geometry. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 7074-7082 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.19, + 0.785, + 0.216 + ], + "angle": 0, + "content": "46. Navon, A., Shamsian, A., Achituve, I., Maron, H., Kawaguchi, K., Chechik, G., Fetaya, E.: Multi-Task Learning as a Bargaining Game (Feb 2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.216, + 0.785, + 0.255 + ], + "angle": 0, + "content": "47. Neven, D., De Brabandere, B., Georgoulis, S., Proesmans, M., Van Gool, L.: Fast Scene Understanding for Autonomous Driving (Aug 2017). https://doi.org/10.48550/arXiv.1708.02550" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.257, + 0.785, + 0.324 + ], + "angle": 0, + "content": "48. Philion, J., Fidler, S.: Lift, Splat, Shoot: Encoding Images from Arbitrary Camera Rigs by Implicitly Unprojecting to 3D. In: Vedaldi, A., Bischof, H., Brox, T., Frahm, J.M. (eds.) Computer Vision - ECCV 2020. pp. 194-210. Lecture Notes in Computer Science, Springer International Publishing, Cham (2020). https://doi.org/10.1007/978-3-030-58568-6_12" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.325, + 0.785, + 0.379 + ], + "angle": 0, + "content": "49. Reading, C., Harakeh, A., Chae, J., Waslander, S.L.: Categorical Depth Distribution Network for Monocular 3D Object Detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8555-8564 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.379, + 0.785, + 0.417 + ], + "angle": 0, + "content": "50. Roddick, T., Kendall, A., Cipolla, R.: Orthographic Feature Transform for Monocular 3D Object Detection (Nov 2018). https://doi.org/10.48550/arXiv.1811.08188" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.42, + 0.785, + 0.46 + ], + "angle": 0, + "content": "51. Ruder, S., Bingel, J., Augenstein, I., Søgaard, A.: Latent Multi-Task Architecture Learning. Proceedings of the AAAI Conference on Artificial Intelligence 33(01), 4822-4829 (Jul 2019). https://doi.org/10.1609/aaai.v33i01.33014822" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.461, + 0.785, + 0.5 + ], + "angle": 0, + "content": "52. Sener, O., Koltun, V.: Multi-Task Learning as Multi-Objective Optimization. In: Advances in Neural Information Processing Systems. vol. 31. Curran Associates, Inc. (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.501, + 0.785, + 0.569 + ], + "angle": 0, + "content": "53. Sun, P., Kretzschmar, H., Dotiwalla, X., Chouard, A., Patnaik, V., Tsui, P., Guo, J., Zhou, Y., Chai, Y., Caine, B., Vasudevan, V., Han, W., Ngiam, J., Zhao, H., Timofeev, A., Ettinger, S., Krivokon, M., Gao, A., Joshi, A., Zhang, Y., Shlens, J., Chen, Z., Anguelov, D.: Scalability in Perception for Autonomous Driving: Waymo Open Dataset" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.569, + 0.785, + 0.622 + ], + "angle": 0, + "content": "54. Teichmann, M., Weber, M., Zöllner, M., Cipolla, R., Urtasun, R.: MultiNet: Realtime Joint Semantic Reasoning for Autonomous Driving. In: 2018 IEEE Intelligent Vehicles Symposium (IV). pp. 1013-1020 (Jun 2018). https://doi.org/10.1109/IVS.2018.8500504" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.623, + 0.785, + 0.678 + ], + "angle": 0, + "content": "55. Vandenhende, S., Georgoulis, S., Van Gansbeke, W., Proesmans, M., Dai, D., Van Gool, L.: Multi-Task Learning for Dense Prediction Tasks: A Survey. IEEE Transactions on Pattern Analysis and Machine Intelligence 44(7), 3614-3633 (Jul 2022). https://doi.org/10.1109/TPAMI.2021.3054719" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.679, + 0.785, + 0.718 + ], + "angle": 0, + "content": "56. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is All you Need. In: Advances in Neural Information Processing Systems. vol. 30. Curran Associates, Inc. (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.719, + 0.785, + 0.772 + ], + "angle": 0, + "content": "57. Wang, R., Qin, J., Li, K., Li, Y., Cao, D., Xu, J.: BEV-LaneDet: An Efficient 3D Lane Detection Based on Virtual Camera via Key-Points. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1002-1011 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.773, + 0.785, + 0.8 + ], + "angle": 0, + "content": "58. Wang, S., Liu, Y., Wang, T., Li, Y., Zhang, X.: Exploring Object-Centric Temporal Modeling for Efficient Multi-View 3D Object Detection (Jun 2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.8, + 0.785, + 0.84 + ], + "angle": 0, + "content": "59. Wang, T., Zhu, X., Pang, J., Lin, D.: FCOS3D: Fully Convolutional One-Stage Monocular 3D Object Detection. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 913–922 (2021)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.148, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.451, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "RepVF: Unified Multi-task 3D Perception" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "19" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.216 + ], + "angle": 0, + "content": "60. Wang, Y., Yin, J., Li, W., Frossard, P., Yang, R., Shen, J.: Ssda3d: Semi-supervised domain adaptation for 3d object detection from point cloud. Proceedings of the AAAI Conference on Artificial Intelligence 37(3), 2707-2715 (Jun 2023). https://doi.org/10.1609/aaai.v37i3.25370, https://ojs(aaai.org/index.php/AAAI/article/view/25370" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.218, + 0.787, + 0.271 + ], + "angle": 0, + "content": "61. Wang, Y., Zhang, X., Yang, T., Sun, J.: Anchor DETR: Query Design for Transformer-Based Detector. Proceedings of the AAAI Conference on Artificial Intelligence 36(3), 2567-2575 (Jun 2022). https://doi.org/10.1609/aaai.v36i3.20158" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.272, + 0.787, + 0.314 + ], + "angle": 0, + "content": "62. Wang, Y., Guizilini, V.C., Zhang, T., Wang, Y., Zhao, H., Solomon, J.: DETR3D: 3D Object Detection from Multi-view Images via 3D-to-2D Queries. In: Proceedings of the 5th Conference on Robot Learning. pp. 180-191. PMLR (Jan 2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.315, + 0.787, + 0.355 + ], + "angle": 0, + "content": "63. Wu, D., Chang, J., Jia, F., Liu, Y., Wang, T., Shen, J.: Topomlp: An simple yet strong pipeline for driving topology reasoning. arXiv preprint arXiv:2310.06753 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.356, + 0.787, + 0.397 + ], + "angle": 0, + "content": "64. Wu, D., Jia, F., Chang, J., Li, Z., Sun, J., Han, C., Li, S., Liu, Y., Ge, Z., Wang, T.: The 1st-place solution for cvpr 2023 openlane topology in autonomous driving challenge. arXiv preprint arXiv:2306.09590 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.398, + 0.787, + 0.451 + ], + "angle": 0, + "content": "65. Xie, E., Yu, Z., Zhou, D., Philion, J., Anandkumar, A., Fidler, S., Luo, P., Alvarez, J.M.: M$^2$BEV: Multi-Camera Joint 3D Detection and Segmentation with Unified Birds-Eye View Representation (Apr 2022). https://doi.org/10.48550/arXiv.2204.05088" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.452, + 0.787, + 0.508 + ], + "angle": 0, + "content": "66. Xu, D., Ouyang, W., Wang, X., Sebe, N.: PAD-Net: Multi-Tasks Guided Prediction-and-Distillation Network for Simultaneous Depth Estimation and Scene Parsing. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 675-684 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.508, + 0.787, + 0.563 + ], + "angle": 0, + "content": "67. Yan, F., Nie, M., Cai, X., Han, J., Xu, H., Yang, Z., Ye, C., Fu, Y., Mi, M.B., Zhang, L.: ONCE-3DLanes: Building Monocular 3D Lane Detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 17143-17152 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.564, + 0.787, + 0.618 + ], + "angle": 0, + "content": "68. Yang, Z., Liu, S., Hu, H., Wang, L., Lin, S.: RepPoints: Point Set Representation for Object Detection. In: 2019 IEEE/CVF International Conference on Computer Vision (ICCV). pp. 9656-9665. IEEE, Seoul, Korea (South) (Oct 2019). https://doi.org/10.1109/ICCV.2019.00975" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.619, + 0.733, + 0.632 + ], + "angle": 0, + "content": "69. Yao, C., Yu, L., Wu, Y., Jia, Y.: Sparse Point Guided 3D Lane Detection" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.633, + 0.787, + 0.701 + ], + "angle": 0, + "content": "70. Yin, J., Fang, J., Zhou, D., Zhang, L., Xu, C.Z., Shen, J., Wang, W.: Semi-supervised 3D Object Detection with Proficient Teachers. In: Avidan, S., Bros-tow, G., Cisse, M., Farinella, G.M., Hassner, T. (eds.) Computer Vision - ECCV 2022, vol. 13698, pp. 727-743. Springer Nature Switzerland, Cham (2022). https://doi.org/10.1007/978-3-031-19839-7_42" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.702, + 0.787, + 0.757 + ], + "angle": 0, + "content": "71. Yin, J., Shen, J., Chen, R., Li, W., Yang, R., Frossard, P., Wang, W.: Is-fusion: Instance-scene collaborative fusion for multimodal 3d object detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14905-14915 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.758, + 0.787, + 0.799 + ], + "angle": 0, + "content": "72. Yin, J., Wang, W., Meng, Q., Yang, R., Shen, J.: A unified object motion and affinity model for online multi-object tracking. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6768-6777 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.799, + 0.787, + 0.839 + ], + "angle": 0, + "content": "73. Yu, T., Kumar, S., Gupta, A., Levine, S., Hausman, K., Finn, C.: Gradient Surgery for Multi-Task Learning (Dec 2020). https://doi.org/10.48550/arXiv.2001.06782" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.839 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "20" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.346, + 0.127 + ], + "angle": 0, + "content": "C. Li et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.788, + 0.202 + ], + "angle": 0, + "content": "74. Zhang, Y., Wang, C., Wang, X., Zeng, W., Liu, W.: FairMOT: On the Fairness of Detection and Re-Identification in Multiple Object Tracking. International Journal of Computer Vision 129(11), 3069–3087 (Nov 2021). https://doi.org/10.1007/s11263-021-01513-4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.203, + 0.787, + 0.259 + ], + "angle": 0, + "content": "75. Zhou, D., Fang, J., Song, X., Liu, L., Yin, J., Dai, Y., Li, H., Yang, R.: Joint 3d instance segmentation and object detection for autonomous driving. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1839-1849 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.259, + 0.787, + 0.3 + ], + "angle": 0, + "content": "76. Zhou, D., Song, X., Dai, Y., Yin, J., Lu, F., Liao, M., Fang, J., Zhang, L.: Iafa: Instance-aware feature aggregation for 3d object detection from a single image. In: Proceedings of the Asian Conference on Computer Vision (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.3, + 0.787, + 0.341 + ], + "angle": 0, + "content": "77. Zhu, X., Su, W., Lu, L., Li, B., Wang, X., Dai, J.: Deformable DETR: Deformable Transformers for End-to-End Object Detection (Mar 2021). https://doi.org/10.48550/arXiv.2010.04159" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.788, + 0.341 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/RepVF_ A Unified Vector Fields Representation for Multi-task 3D Perception/9c289191-ecfe-49ca-a348-d2cad5492a03_origin.pdf b/2024/RepVF_ A Unified Vector Fields Representation for Multi-task 3D Perception/9c289191-ecfe-49ca-a348-d2cad5492a03_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..822addbf844d41669f6cb8c8b32c0b9076d1f914 --- /dev/null +++ b/2024/RepVF_ A Unified Vector Fields Representation for Multi-task 3D Perception/9c289191-ecfe-49ca-a348-d2cad5492a03_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:079171bd8928b036196352883fb1f63b72246f02acdf58018230450ef0cc36b8 +size 4784417 diff --git a/2024/RepVF_ A Unified Vector Fields Representation for Multi-task 3D Perception/full.md b/2024/RepVF_ A Unified Vector Fields Representation for Multi-task 3D Perception/full.md new file mode 100644 index 0000000000000000000000000000000000000000..57c55b877b4cbe073b00966e51644b594b5f7aa4 --- /dev/null +++ b/2024/RepVF_ A Unified Vector Fields Representation for Multi-task 3D Perception/full.md @@ -0,0 +1,327 @@ +# RepVF: A Unified Vector Fields Representation for Multi-task 3D Perception + +Chunliang Li $^{1*}$ , Wencheng Han $^{2*}$ , Junbo Yin $^{1}$ , Sanyuan Zhao $^{1}$ , and Jianbing Shen $^{2}$ + +$^{1}$ School of Computer Science, Beijing Institute of Technology + $^{2}$ SKL-IOTSC, Computer and Information Science, University of Macau, China +{jbji, sanyuanzhao}@bit.edu.cn +{wencheng256, yinjunbocn}@gmail.com +jianbingshen@um.edu.mo + +Abstract. Concurrent processing of multiple autonomous driving 3D perception tasks within the same spatiotemporal scene poses a significant challenge, in particular due to the computational inefficiencies and feature competition between tasks when using traditional multi-task learning approaches. This paper addresses these issues by proposing a novel unified representation, RepVF, which harmonizes the representation of various perception tasks such as 3D object detection and 3D lane detection within a single framework. RepVF characterizes the structure of different targets in the scene through a vector field, enabling a single-head, multi-task learning model that significantly reduces computational redundancy and feature competition. Building upon RepVF, we introduce RFTR, a network designed to exploit the inherent connections between different tasks by utilizing a hierarchical structure of queries that implicitly model the relationships both between and within tasks. This approach eliminates the need for task-specific heads and parameters, fundamentally reducing the conflicts inherent in traditional multi-task learning paradigms. We validate our approach by combining labels from the OpenLane dataset with the Waymo Open dataset. Our work presents a significant advancement in the efficiency and effectiveness of multi-task perception in autonomous driving, offering a new perspective on handling multiple 3D perception tasks synchronously and in parallel. The code will be available at: https://github.com/jbji/RepVF. + +Keywords: 3D Lane Detection $\cdot$ 3D Object Detection $\cdot$ Multi-task Method. + +# 1 Introduction + +In autonomous driving, multiple 3D perception tasks often need to be processed synchronously, in real-time, and in parallel. These tasks, which exist within the + +![](images/2bfbbd98ca04d688ded301c786f4503e6fd2dcc793e811e4c5a6dc7ca31f9204.jpg) +(a) Multi-tasking paradigm comparison. + +![](images/c98f0f250df4db5e11794a7c0ede71ed5720cd8dd1abc95d2d1f2b098d52c4aa.jpg) +(b) $RepVF$ , conversion and supervision. +Fig. 1: $RepVF$ is proposed as a unified representation for 3D perception to achieve single-head multi-tasking. It consists of a set of vector fields learned to represent road elements, and is trained utilizing existing labels through differentiable conversions. The proposed single-head multi-tasking paradigm reduces task conflict and competition. + +same spatiotemporal scene, are often complementary to each other in both geometric and semantic aspects [19]. For example, 3D object detection [12] and 3D lane detection [11] define foreground objects and background lanes within the same 3D scene. Vehicles usually travel following the guiding semantics of lanes that do not spatially overlap and are partially obscured by other vehicles. Some previous industrial solutions [1,2] utilize a multi-model paradigm for multi-task perceptions, as depicted in Figure 1a(1). These separate models cannot share common features across different tasks, resulting in computational waste. Furthermore, this paradigm does not take advantage of the inherent connections between different tasks. Some works [19,28,38,39] attempt to overcome this limitation by using a common backbone and separate head networks for each task, as shown in Figure 1a(2). This approach can considerably reduce the computational cost when handling a large number of tasks, but it might lead to a task balancing issue. Since different task-aware heads focus on various aspects of the scene, they may compete for the backbone features, resulting in instability during training [19,39]. + +The competition is attributed to the variation in representation between different tasks. For example, the 3D detection task [12,26,28,37,38,60,62] guides the head network to predict 3D bounding boxes, whereas the lane detection task [6,11] necessitates the corresponding head to regress lane shapes. Specifically, 3D bounding boxes describe a cubic area in 3D space, representing different category foregrounds using position, size, and orientation. In contrast, 3D lanes are represented using anchor [6,11,18] or parametric [15] methods to describe one or more 1-D lines. These representations focus on different aspects of the scenes, which ultimately leads to competition. + +In this paper, we propose a unified representation, $RepVF$ (Representative Vector Fields), for perception tasks in autonomous driving. Using this represent + +tation, we design a single-headed, unified model that can perform tasks simultaneously. As illustrated in Fig. 1b, $RepVF$ represents 3D vector fields that each assigns vectors to spatial locations, denoted as $(S, \mathcal{F}(S)) \subseteq \mathbb{R}^{3 + d}$ in Sec. 3.2. Sub-vector fields representing different targets in the 3D perception scene progressively adhere to their spatial extent, accurately characterizing the anisotropic structure in their respective spatial vicinity. $RepVF$ can be differentiably transformed into task-specific elements, hence requiring no special supervision and can utilize existing labels. With $RepVF$ , our network only processes one single type of fundamental perception element. As shown in Fig. 1a, unlike the traditional multi-task approach that employs multiple task heads [19, 28, 30, 38, 72], we follow an unprecedented single-head multi-task paradigm. The single-head multi-task eliminates task-specific parameters and obviates the need to explicitly model task interactions. + +RepVF representation serve as a geometrical, cross-scale 3D perception representation alternative. Unlike perception elements capable of handling only single-scale tasks [6,11,12,62], it breaks down the scene into smaller perceptual units while retaining the capability for target-level perception at larger scales. Based on the RepVF representation, we further propose RFTR to leverage this relationship, utilizing queries with a hierarchical structure that implicitly models the relationships both between and within different tasks. Following the single-head multi-task architecture, with only shared parameters across tasks, RFTR does not depend on existing multi-task optimization strategies [21,33,35,52,73]. The gradient discrepancies between tasks across different training iterations have been mitigated, fundamentally reducing the competition and conflict inherent in multi-tasking 3D perception. By employing RepVF to replace the fundamental elements for different tasks, and by combining 3D lane labels from the OpenLane dataset [6] with the Waymo Open dataset [53] for simultaneous multi-task training and inference, our approach achieves comparable or even better multi-task performance than single-task counterparts with only one unified model. + +In summary, we introduce $RepVF$ as a unified representation for multiple perception tasks in autonomous driving, effectively reducing the competition among different task representations during training and enhancing model convergence. We further develop a single-head multi-task framework RFTR based on $RepVF$ , fully exploiting the intrinsic connections among different tasks within a scene, improving feature interaction with queries, and enabling gradient-balanced multi-task training. Our approach achieves impressive performance across two key tasks in autonomous driving—3D object detection and 3D lane detection—showcasing the advantages of a unified task representation in the context of autonomous driving applications. + +# 2 Related Work + +# 2.1 Camera-based 3D Perception in Autonomous Driving + +3D object detection [7,9,12,62,70] and 3D lane detection [11,14] are two common 3D perception tasks, and previous approaches usually perform the two tasks + +independently. 3D object detection has been extensively studied either in model architectures [7, 37, 38, 59] or feature representations [17, 25, 27, 28, 49, 50, 71, 76]. Early methods [20, 22, 45] predict 3D bounding boxes using 2D detectors, and Mono3D [7] is one of the first methods directly make predictions from 3D representations. Motivated by DETR [11], DETR3D [62] uses sparse 3D object centers as queries, and PETR [37] further enhances its feature representations by embedding 3D positions. Some recent approaches [17, 28, 38, 58] also focus on incorporating temporal information into 3D object detection, but they all follow the transformer-based [56] paradigm introduced by DETR3D [62]. + +Only recently have DETR-based methods been introduced for 3D lane detection [3, 38, 42, 63, 64]. Previous methods [6, 11, 14, 34, 57] prefer to project image features into BEV space using Inverse Perspective Mapping (IPM), but this introduces prohibitive height prediction error. Some work [67] avoids IPM via depth estimation, and Anchor-3DLane [18] reverses this projection process. Following DETR [3, 38, 61] introduce 3d lane anchors as queries, and [42] adopts finer lane prior queries and is end-to-end. With the advances of DETR [61, 62] networks in both tasks [3, 28, 37, 42], it is architecturally ready to unify the two tasks without standalone models. + +# 2.2 Multi-task 3D Perception + +The problem of multi-tasking in autonomous driving perception is still in its infancy. Recent advances [19, 28, 37-39, 48, 65, 75] elaborate on universal perception features and multiple tasks are handled with different heads. Moreover, there is no consensus on the required tasks. For example, MMF [29] addresses both depth completion and object detection issues, while [19, 28, 38, 39] performs object detection and map segmentation, with [38] performing additional 3D lane detection separately. However, most of these studies only use multiple tasks to demonstrate the universality of their learned features, and the tasks are trained and evaluated separately [28, 38] rather than simultaneously [19, 39], not to mention dataset discrepancies. + +In addition, multitask learning (MTL) typically involves task conflict. Previous MTL studies [55] addresses this either through parameter sharing [13, 44, 47, 51, 54, 66] or optimization [8, 21, 33, 35, 52, 73]. Parameter sharing approaches are conceptually straightforward, and are further classified into hard-sharing (with shared and task-specific parameters) and soft-sharing (with a cross-task talk mechanism). Optimization-based approaches, e.g. IMTL [33] and DWA [35], attribute MTL imbalance to the gradient and loss imbalance and explore optimization calibration methods. Taking advantage of these advances, FULLER [19] stands as the pioneering work in autonomous driving perception that analyzes the performance of modern optimization-based MTL calibration methods [8, 33, 35]. With a unified task representation, we follow a different paradigm from previous ones, that eliminate task-specific heads and mitigate the gradient imbalance. Our approach can be seen as a hard-sharing technique without task-specific parameters and does not require calibration. + +# 3 Representative Vector Fields + +We begin by revisiting the representations of two common 3D perception tasks: 3D object detection [4, 12, 53] and 3D lane detection [6, 67]. By exploring how task-specific representations have evolved into task-specific designs [3, 6, 7, 11, 37, 38, 42, 50, 62], we propose the task-agnostic $RepVF$ (Representative Vector Fields), which arise from the geometric commonality in the tasks and effectively reduce task competition via unified representations. + +# 3.1 Task-specific Representations and Single-tasking + +Representation Formulations. Task-specific representations vary due to the geometric features of the task targets, which differ in dimensions and scales. As a key task for understanding 3D space, 3D object detection deals with objects irregularly shaped and scaled from one to several metres, defined by coarse outer boundaries as 3D bounding boxes [12], formally a 7-d coordinate $B = (x,y,z,l,w,h,\theta)$ encoding object center $(x,y,z)$ , box dimensions $(l,w,h)$ and heading $\theta$ . Conversely, due to their linear shape and ability to cover long distances, 3d lanes are not described by boundaries but by direct geometries, typically ordered sequences of $N_{L}$ 3D point coordinates [6]: + +$$ +L = \left[ \left(x _ {1}, y _ {1}, z _ {1}\right), \left(x _ {2}, y _ {2}, z _ {2}\right), \dots , \left(x _ {N _ {l}}, z _ {N _ {l}}, z _ {N _ {l}}\right) \right] \tag {1} +$$ + +This may look familiar to an unordered point cloud, but they are ordered along the direction of the road. In addition, 3D perception problems typically involve classifications that require category scores $\mathcal{C}_i$ , varying from 1, 3 or 23 [4,12,53] classes (3D object) to 21 or 22 [6] classes (3D lane). + +Evolving into Task-specific Designs. Pioneering 3D object detection from RGB images is proposal based, e.g. Mono3D [7] reduces proposal space dimension of each class to $(x,y,z,\theta,t)$ with representative templates, and only predicts $\varDelta_{\mathrm{pos}}$ ; OFT-Net [50] predicts offsets of all bounding box parameters relative to the top-view grid, i.e. $\varDelta_{\mathrm{pos}},\varDelta_{\mathrm{dim}},\varDelta_{\mathrm{ang}}$ . Taking advantage of the center coordinates, DETR-based methods [37,62] regress $\varDelta_{\mathrm{pos}}$ along with bounding box parameters. However, constrained by the limited spatial extent of center points, extracted features around them capture only local spatial areas [77], failing to capture the finer pose and structure of objects. + +3D lines are suitable for anchor-based representations [6, 11, 38], resulting in equally spaced longitudinal lines and anchor-based predictions $\varDelta_{\mathrm{xz}}$ . Anchor representations reduce the representation dimension, but also restrict the degrees of freedom, limiting the maximum model capacity, which has led some works [6, 18] focus on more sophisticated anchor representations. Recent DETR-based methods [3,38,42] use anchor coordinates as queries, exploring the full potential of anchors. Unfortunately, since it's not trivial and stable to directly predict parameter-driven curves [10, 15], mainstream methods stick to predicting anchor-based offsets $\varDelta_{\mathrm{xz}}$ , and nearly no one ever worked directly on the raw lane format. + +Intersection of Task-specific Representations. We believe that it's already been architecturally ready to unify the two (and possibly more) perception tasks, with the population and success of DETR practices [19,28,38,61]. However, they still rely on multi-head design due to the discrepancy in task representation, which consumes more computational resources. Despite the significant differences in the dynamics and scales of the described targets, these task-specific representations only encapsulate descriptions of shapes, structures, and directions (object heading and lane ordinality). It is plausible to represent these task targets universally, taking into account perception hierarchies. + +# 3.2 The Task-Agnostic Representation through RepVF + +To facilitate efficient design in multitasking frameworks, we introduce $RepVF$ , a vector fields representation inspired by 2D point set object representation [43,43]. Distinct from conventional isotropic points, $RepVF$ assigns a vector to each location making itself anisotropic. We define $RepVF$ as a collection of vector fields, each field denoted by $R_{i}$ as a perception target, whose semantic vector assignment mapping $\mathcal{F}_i$ is defined on $D = \bigcup_{i=1}^{n_s} S_i \subseteq \mathbb{R}^3$ : + +$$ +\{R _ {i} \}, \quad R _ {i} = \left(S _ {i}, \mathcal {F} _ {i} \left(S _ {i}\right)\right) \subseteq \mathbb {R} ^ {3 + d}, \quad \mathcal {F}: D \rightarrow \mathbb {R} ^ {d} \tag {2} +$$ + +where $d$ is the dimension of the target vector domain. Therefore, for any point $x$ within $D$ , $\mathcal{F}$ assigns it a vector $\mathcal{F}(x) \in \mathbb{R}^d$ that encapsulates dimensional and positional attributes of the point. This definition allows us to cover the whole space: The sampling domain $D$ is partitioned into $n_s$ disjoint subsets $S_i$ , with each subset containing $n$ representative sampling points: $S_i = \{(x_k, y_k, z_k)\}_{k=1}^n$ . Accordingly, the collective vector field $\mathcal{F}$ sprawls over $D = \bigcup_{i=1}^{n_s} S_i$ , articulated into $n_s$ sub-vector fields $\mathcal{F}_i$ catering to specific tasks. + +Intuitively, $S_{i}$ directly outlines entity-level geometric structures, and $\mathcal{F}_i(S_i)$ delineates contextual characteristics (i.e. $\theta$ for bounding box $B$ , or adjacent point directions for lane $L$ ). Considering the directionality, we use unit-length vectors parallel to the x-y plane to capture angular behavior, resulting in behavior akin to scalars and $d = 1$ . To represent proper $\mathcal{F}_i$ for 3d lanes, directions of the penultimate points are employed for terminal lane line points. + +Converting RepVF into Task-Specific Representation. To leverage existing task labels for training and evaluation, similar to object detection representations based on point sets [43, 68], we convert predictions $RepVF \hat{R}_i$ , category $\hat{C}_i$ to task-specific representations. We define geometric interpretation processes as differentiable, parameter-free transformation functions: + +$$ +\mathcal {T} ^ {(\cdot)}: \hat {R} _ {i} \rightarrow \hat {P} _ {i}, \quad \text {s . t .} \hat {P} _ {i} \in \{\hat {B}, \hat {L} \} _ {i} \tag {3} +$$ + +Specifically, we have $\mathcal{T}^l:\hat{R}_i\to \hat{L}_i$ for 3d lanes and $\mathcal{T}^b:\hat{R}_i\rightarrow \hat{B}_i$ for 3d objects. For semantic parts, to align $\hat{\mathcal{C}}_i$ with task-specific classes, we split it into $n_{(\cdot)}$ bins and take max score within each bin, where $n_{(\cdot)}$ is the count of task categories. $\hat{R}_i$ and $\hat{R}_j$ remain formally equivalent until contextualized into $\hat{P}_i$ and $\hat{P}_j$ , thus maintaining task agnosticism. + +Conversion Function Choice. In our implementation, $\mathcal{T}^l$ sorts the unordered points of $\hat{S}_i$ by the road direction $(+X$ in Waymo [53]). Unlike 2D boxes, 3D boxes are not strictly aligned with to $XYZ$ axes due to headings. Therefore, $\mathcal{T}^b$ consists three parts that respectively compute heading, box center and dimensions: + +$$ +\mathcal {T} _ {h} ^ {b}: \hat {\mathcal {F}} _ {i} \rightarrow \hat {\theta} _ {i}; \quad \mathcal {T} _ {x y z} ^ {b}: \hat {S} _ {i} \rightarrow (\hat {x} _ {i}, \hat {y} _ {i}, \hat {z} _ {i}); \quad \mathcal {T} _ {l w h} ^ {b}: \hat {S} _ {i} \rightarrow (\hat {l} _ {i}, \hat {w} _ {i}, \hat {h} _ {i}) \qquad (4) +$$ + +Here, $\mathcal{T}_h^b$ and $\mathcal{T}_{xyz}^b$ are simple arithmetic averages. With $\hat{\theta}_i$ given by $\mathcal{T}_h^b$ , by projecting $\hat{S}_i$ to the main orientation axis $\hat{\theta}_i$ , the predictions are aligned with the predicted directions. We then proceed with two candidates of $\mathcal{T}_{lwh}^b$ similar to 2D practices [68]: (1) $\mathcal{T}_{lwh}^b = \mathcal{T}_1$ , the min-max function, performing min/max operations projection dimensions for dimensions; (2) $\mathcal{T}_{lwh}^b = \mathcal{T}_2$ employs a momentum-based function taking the variance of predicted points as the boundary box dimensions. In our practice, $\mathcal{T}_2$ works slightly better. + +Learning RepVF. RepVF is supervised by loss functions universally across tasks without special losses. Here we use $\mathbf{L}^1$ regression loss on $S_{i}$ and $\mathcal{F}_i$ , and focal loss [32] on $\mathcal{C}_i$ . Smaller structures are described by points in $S_{i}$ , and $R_{i}$ itself is set-level, extending RepVF across different perception hierarchies. To capture this cross-level property, we adopt DETR-based structure [37,61,62] and aligned query design to use raw representations motivated by [3, 36]. With multiple decoder layers [62, 77], the learning process of RepVF can be characterized as iterative update $\varOmega_l$ on queries $Q_{l}$ and inferring $\hat{R}_l,\hat{C}_l$ from them with linear layer $\varPhi$ and decoder layer $l$ : + +$$ +Q _ {l} = \Omega_ {l} \left(Q _ {l - 1}\right), \quad \hat {R} _ {l}, \hat {\mathcal {C}} _ {l} = \Phi (Q _ {l}) \tag {5} +$$ + +We will explain our network in detail in the next section. + +# 4 RFTR: The Single-head Multi-tasker + +In this section, we introduce the single-head multi-task framework RFTR (Representative Vector Fields Transformer), which uses $RepVF$ representation to unify different tasks and exploit intrinsic connections among them. We also observed one interesting behavior of RFTR in the balance of multi-task gradients. RFTR is built upon DETR networks [5,37,62] and is end-to-end on all tasks. + +# 4.1 Network Structure + +Figure 2 shows our RFTR architecture built upon $RepVF$ . It inputs images $I$ from $N$ cameras with known camera parameters, and outputs unified representative vector fields $\mathcal{F}$ for various perception tasks in the 3D scene. RFTR contains four main components that extract features from images, generate set-level queries for hierarchical representation structure, solve unified representative vector fields and multiple tasks with one single head, and convert results to task representations in a differentiable way to utilize existing task labels. + +![](images/867f1b847c5b1f9e3c5f9394f6700d4d85a853583e8c6e4af81047ad62423d23.jpg) +Fig. 2: Overview of our RFTR (Representative Vector Fields Transformer) built upon RepVF. Multi-view image features with 3D position embeds are extracted by image backbone and fed into the decoder. We generate set-level query embeds from 3D space sampling sets, each representing a perception target. One single unified task head is then used to predict unified representative vector fields. Finally, predictions are transformed in a differentiable manner into task specific representations to utilize existing labels for supervision. + +Feature Extraction. Following common practices in utilizing 2D backbones to extract features [28,37,62], we adopt ResNet [16] (ResNet-50) and FPN [31] to produce 2D features $F^{2d}$ . They're encoded with 3D position embedding in PETR [37] into 3D position aware features $F^{3d}$ . + +Generating Set-level Perception Queries. As $RepVF$ is sampled from $n_s$ point sets $S_i$ with $n$ points each and is two-level, our aim is to generate queries that are consistent with this hierarchical point-set structure. Hence provide the decoder with sufficient spatial information, improving feature interactions. Unlike former approaches [37, 38, 42, 62] that encode one single 3D location, each query in RFTR encodes multiple locations simultaneously. + +We generate set-level queries $Q_0 \in \mathbb{R}^{n_s \times d_q}$ from initial point sets corresponding to $S_i$ , where $d_q$ is the dimension of each query. The generation process $\mathcal{G}$ consists of encoding initialized positions from the point level, and embedding them at the set level: + +$$ +\mathcal {G} = \operatorname {e m b e d} (\text {F l a t t e n (P E)}) \tag {6} +$$ + +where PE is the positional encoding [56]. Point-level PE is flattened into set-level and embedded by a small MLP with two layers. To avoid confusing the model with expansive locations falling on multiple targets in the early stages of training, we use $n$ identical 3D locations sampled with 0-1 uniform distribution for all $n_s$ initial point sets. + +To our best knowledge, the closest generation approach to ours is [3], which generates query from an ordered sequence of anchor points motivated by [36]. + +However, ours is from an unordered sampling point set and is not task-specific. Moreover, we use a tiny 16-dimensional PE for each coordination, divert from former methods [37, 62] with a much larger 256-dimension. In our multi-task setting, queries are equivalently initialized, thus task-agnostic. We divide the first $80\%$ as 3D objects and the last $20\%$ as 3D lanes. + +Single-head Multi-tasking. Similar to all previous approaches [5, 37, 62], queries are learnable and updated in decoder layers iteratively. The unified head $\varPhi$ of RFTR is composed by a multi-layer transformer decoder [56], and two branches $\varPhi_{g}$ and $\varPhi_{s}$ dedicated for geometric regression and semantic prediction respectively to predict $RepVF$ . We use the same $n_{L} = 6$ decoder layers with previous methods [5, 37, 62] to exploit intrinsic tasks connections implicitly. The interaction process within these layers can be formally expressed as: + +$$ +Q _ {l} = \Omega_ {l} \left(F ^ {3 d}, Q _ {l - 1}\right), \quad l = 1, \dots , N _ {L} \tag {7} +$$ + +With unified head $\varPhi=\{\varPhi_{g},\varPhi_{s}\}$ , predictions $\hat{R}_{li}$ and $\hat{C}_{li}$ is decoded from $i$ -th query at $l$ -th layer: + +$$ +\hat {R} _ {l i} = \Phi_ {g} \left(Q _ {l i}\right), \quad \hat {\mathcal {C}} _ {l i} = \Phi_ {s} \left(Q _ {l i}\right) \tag {8} +$$ + +For $\varPhi_{g}$ , additional offsets are added to produce $S_{li}$ . Unified geometric predictions $\hat{R}_l = (\hat{S}_l,\hat{F}_l)\subset \mathbb{R}^{n_s\times (3 + d_p)}$ are converted to task-specific representations: + +$$ +\hat {L} _ {l i} = \mathcal {T} ^ {l} (\hat {R} _ {l i}), \quad \hat {B} _ {l i} = \left(\mathcal {T} _ {x y z} ^ {b}, \mathcal {T} _ {\theta} ^ {b}, \mathcal {T} _ {l w h} ^ {b}\right) (\hat {R} _ {l i}) \tag {9} +$$ + +Here, $\mathcal{T}^{(\cdot)}$ are differentiable functions predefined in section 3.2. Each head branch $\varPhi_{(\cdot)}$ is a simple small MLP network with three fully connected layers. Following DETR3D [62], we weight the loss for the predictions from each decoder layer $\Omega_l$ during training, and only outputs from the last layer is used during inference. + +Loss Function. RFTR uses existing task labels for supervision, and its loss is calculated between converted predictions and task ground truths. It consists of the following components for coordinate regression, vector fields regression and classification respectively: + +$$ +\mathcal {L} ^ {(\cdot)} = w _ {r} ^ {(\cdot)} \mathcal {L} _ {r} ^ {(\cdot)} + w _ {f} ^ {(\cdot)} \mathcal {L} _ {f} ^ {(\cdot)} + w _ {c} ^ {(\cdot)} \mathcal {L} _ {c} ^ {(\cdot)}, \quad \mathcal {L} = \sum_ {\text {t a s k s}} \mathcal {L} ^ {(\cdot)} \tag {10} +$$ + +where $w^{(\cdot)}$ represent loss component weights for task $(\cdot)$ , i.e. 3D object detection and 3D lane detection. We use the Hungarian algorithm [23] for label assignment after cost calculation between ground truth and converted predictions $\hat{B}_{li}, \hat{P}_{li}$ . + +In practice, we choose $w_{c}^{b} = w_{c}^{l} = 2.0$ , $w_{r}^{l} = 0.003$ , $w_{r}^{b} = 0.1$ , $w_{f}^{b} = 0.2$ and $w_{f}^{l} = 0.032$ . Coordinate regression and classification weights are selected empirically to equalize the scales and contributions of each loss term [74], fields weights are empirically selected. We use $\mathrm{L}^1$ loss for $R_{i}$ regression, and focal loss [32] with $\gamma = 2.0$ and $\alpha = 0.25$ for classification. Since we predict a degenerated scalar field, we supervise the angular prediction by its sine and cosine values. All tasks share the same losses and similar weights, no additional, task-specific loss measures are adopted. + +![](images/9f6a331ed5cd25b2a4805ca119f81d2ff5f3e2f6bfe46e5fbadcb08a203d58fe.jpg) +(a) Gradient disparities by iteration index. + +![](images/2bd0552c0fb3bfb1dbd0f2bdaa5704f10ab8a1549f59c09b67bea00dafe52298.jpg) +(b) Average gradient disparities. +Fig. 3: Our single-head RFTR model shows a reduced gradient imbalance and better disparity stability (mean: 2.47, variance: 0.90) compared to the multi-head baseline (mean: 2.98 variance: 1.86) or the multi-head baseline with RepVF (mean: 2.72, variance: 1.57). Ideally balanced gradient disparity through training iterations should approach 2. For ease of trend observation, we have clipped the top $2\%$ values to the mean, with curves smoothed through a window size of 50 1-d convolution. + +# 4.2 Gradient Balance in Single-head Multi-tasking + +The multi-head multi-task approach often leads to task conflict, attributed in prior studies [21, 33, 46, 52] to gradient imbalance. These studies [33, 46] and more recent work [19] have explored appropriate gradient calibration strategies to address this issue. As we will show, our single-head RFTR naturally achieves better balanced multi-task gradients thus calibration may no longer be required. + +To elucidate the advantage of our single-head multi-task paradigm over multi-head multi-task settings, we use an improved measure of gradient disparity adapted from FULLER [19], by taking into account the symmetry disparity: + +$$ +\operatorname {d i f f} \left(\nabla \mathcal {L} ^ {l}, \nabla \mathcal {L} ^ {b}\right) = \frac {\| \nabla \mathcal {L} ^ {l} \|}{\| \nabla \mathcal {L} ^ {b} \|} + \frac {\| \nabla \mathcal {L} ^ {b} \|}{\| \nabla \mathcal {L} ^ {l} \|} \tag {11} +$$ + +where $\| \cdot \|$ is the Frobenius norm and $\nabla \mathcal{L}^{(\cdot)}$ denotes gradients. Given that $\| \nabla \mathcal{L}^{(\cdot)}\|$ is always greater than 0, the value of diff function is always greater than or equal to 2. Therefore, the more stable and closer to 2 the gradient disparity is during model training, the better the multi-task stability indicated. + +Figure 3a illustrates the gradient disparity curves through training iterations. The multi-head multi-task method [37,38] is under the same task settings with ours. In the sampling 1k iterations, the stability and absolute scale of the gradient disparity curve for the multi-head configuration are inferior to our single-head multi-task structure. The mean and variance of the multi-head disparities are about $108.5\%$ and $107.8\%$ higher, respectively, than our single-headed RFTR approach, as shown in Fig. 3b. We have also observed that unified representation and single head design contributed to the improvement about equally. More importantly, this means that our single-head RFTR mitigates, though not eliminating, the gradient imbalance between tasks, which is fundamentally different from task gradient balancing strategies that do not change the gradient itself. + +# 5 Experiments + +We evaluate our method on the Waymo Open Dataset [53] extended by 3D lanes annotated in OpenLane [6], so that we train and evaluate tasks simultaneously. + +# 5.1 Datasets + +Waymo Open Dataset and OpenLane. Waymo Open Dataset (Perception) [53] is a large-scale multimodal dataset composed of data from 5 cameras(front and sides), 1 mid-range lidar, and 4 short-range lidars. It contains 1000 sequences in total(about 198k samples), a training set of 798 sequences and a validation set of 202 sequences. OpenLane [6] is a comprehensive real-world 3D lane detection dataset built upon the Waymo Open dataset [53]. It includes a vast array of 200K frames and over 880K carefully annotated lanes, making it one of the largest of its kind to date. The dataset encompasses complex lane structures and features a variety of scene tags, such as weather and locations. Designed to emulate real-world scenarios, OpenLane provides a challenging benchmark for advanced lane detection algorithms. We integrate its 3D lane labels with the original Waymo Open sensor data for multi-task annotation. + +Alignment of Datasets. To achieve both tasks simultaneously, we align both datasets in terms of coordinate systems and data splits. (1) Coordinate systems. For the vehicle frame, OpenLane [6] uses a vehicle coordinate system that corresponds to the "y-front, x-right, z-up" positioning of 3D-LaneNet [11]. We then align this system to coincide with the commonly utilized "x-front, y-right, z-up" axis for lidar-based 3D object detection. With respect to the sensor frame, we apply transformations to the camera intrinsics so that the sensor frames for all cameras conform to a uniform "z-depth, x-right, y-down" frame. The predictions are then transformed back into their respective task-based coordinate systems for evaluation. (2) Data splitting. The OpenLane [6] dataset provides a 1000 segment full version and a 300 segment smaller subset ( $30\%$ ). The partitioning of the larger version complements the original train/Validation split of the Waymo Open Dataset. The subset version partitioning is aligned with OpenLane. We perform ablation studies on the smaller subset. + +# 5.2 Implementation Details + +Data Processing. We use the v1.4.2 of Waymo Open Dataset [53], and v1.2 for OpenLane [6]. We set the perception range to $[-84.88m, 84.88m]$ for the X and Y axis, and $[-10m, 10m]$ for the Z axis. 3D objects that do not fall within the perception range, are invisible to the cameras, or do not have LiDAR points, are filtered. For 3D lanes, visibility filtering is first applied after axis transformation, then 3D lanes outside the perception range are pruned and then resampled to a fixed size. Approximately $2.07\%$ of the frames without annotations after filtering are excluded from use. + +Table 1: Comparison of recent models on 3D lane detection using the OpenLane [6] validation set. The down arrow indicates that lower metric values correspond to better performance, and vice versa. Red indicates the best result, and blue the second-best. + +
MethodsF-Score ↑Category Accuracy ↑X error (m) ↓Z error (m) ↓
nearfarnearfar
3D-LaneNet [11] [ICCV19]44.1-0.4790.5720.3670.443
Gen-LaneNet [14] [ECCV20]32.3-0.5910.6840.4110.521
PersFormer [6] [ECCV22]50.592.30.4850.5530.3640.431
Cond-IPM [6]36.6-0.5631.0800.4210.892
CurveFormer [3] [ICRA23]50.5-0.3400.7720.2070.651
PETRv2-E [38] [ICCV23]51.9-0.4930.6430.3220.463
PETRv2-V-10 [38]57.8-0.4270.5820.2930.421
PETRv2-V-400 [38]61.2-0.4000.5730.2650.413
BEV-LaneDet [57] [CVPR23]58.4-0.3090.6590.2440.631
Anchor3DLane [18] [CVPR23]53.190.00.3000.3110.1030.139
SPG [69] [ICCV23]52.3-0.4680.5140.3710.418
RFTR (Ours)61.891.60.3410.4500.0730.107
+ +Table 2: Comparison of recent works of 3D object detection on the Waymo Open Dataset [53] val set. mAPs are LEVEL_1 in Waymo metrics. + +
MethodsOverallVehicleCyclistPedestrian
mAPmAPHmAPmAPHmAPmAPHmAPmAPH
DETR3D [62]10.19.69.29.115.815.55.34.1
DETR3D [62]8.68.018.118.02.41.65.34.4
PETR [37]20.919.731.130.819.517.612.110.6
RFTR (Ours)19.518.722.622.427.826.38.07.2
+ +Model and Training. We used ResNet 50 [16] as the 2D image backbone for fair comparisons. Following PETR [37], the same upsampled and fused P4 feature with 1/16 input resolution is used as the 2D feature. A CPFPN [31] is used as the image neck. All 3D coordinates are normalized to $[0,1]$ . The same weights are used for both the loss computation and the bipartite matching. We adopt progressive weighting for the loss of each decoder layer to improve the convergence speed. For 3D object detection, we compute matching costs based on the center displacement and class score weightings. Dimensions are refined post-matching using the ground truth headings to increase training robustness. We use a total of $n_s = 750$ queries with $n = 15$ sample locations each for the two tasks, 600 for 3D objects and 150 for 3D lanes. We trained our model using the AdamW [41] optimizer with a 0.01 weight decay. The learning rate started at $2.0 \times 10^{-4}$ and was decayed with a cosine annealing policy [40]. We adopted a multi-scale training strategy, with the shorter side chosen randomly from [640,900] and the longer side not exceeding 1600. We applied rotation augmentations between $[-\pi /18,\pi /18]$ . We resized images of sizes $1920 \times 1080$ and $1920 \times 1280$ from different views to $1600 \times 900$ before augmenting. The final resolution for network inputs was $1408 \times 512$ . All experiments on the full dataset are trained for 24 epochs on 4 GeForce RTX 4090 GPUs with a batch size of 8. No test time augmentation methods are used during inference. + +![](images/06f1d69f3e44123ec45862776d7fee29e3761a7f5836c7cf9a9f1598cc45f104.jpg) +Fig. 4: Visualization of the RFTR model's performance and the learned $RepVF$ . + +# 5.3 Comparison with SOTA + +3D Lane Detection. As shown in Tab. 1, our RFTR model showcases significant achievements in 3D lane detection on the OpenLane validation set, demonstrating its competitive edge with a high F-score of 61.8 and category accuracy of $91.6\%$ . It matches or exceeds state-of-the-art models such as PETRv2-V-400 [38] (more than $26\times$ denser representation than ours) when assessed on the OpenLane [6] validation set. Notably, while our model has a slightly higher X error in the "far" class than [18], it manages to lead with a higher F-score and near accurate Z errors. Moreover, our method yields notable performance improvements over the Persformer [6] baseline, reflecting the advantageous adaptation of our unified RepVF representation and the RFTR architecture. + +3D Object Detection. The performance of our RFTR model in 3D object detection, as demonstrated in Tab. 2, shows comparable promising results on the Waymo Open Dataset [53] validation set. While our approach shows a commanding lead in detecting vehicles and cyclists with a solid L1 mAP and mAPH, it shows challenges in detecting smaller entities such as pedestrians. It should be noted, however, that these preliminary results indicate the great potential of using our unified perceptual model for complex tasks, which will require further fine-tuning for object classes with less prominent features. + +# 5.4 Ablation Studies + +Head design. We begin by finding out how head designs and representations contribute to the performance, in Tab. 3. The $RepVF$ representation with the single-head design, i.e. RFTR, performs the best in general with 37.4M parameters. Unified representation brings performance improvements in multi-task learning and reduces task competition. In contrast, task-specific multi-head design produces inferior and unstably biased performance due to task competition, even with more parameters. + +Table 3: Multi-task head design and representation ablation. Tasks are trained jointly on $30\%$ data of Waymo Open Dataset (L1 mAP) and OpenLane (F-Score). TS denotes task-specific representation. $\dagger$ : lightweight models. + +
RepVFTSHeadParams↓F-Score↑Vehicle↑
✓Multi†37.1M49.620.1
✓Multi46.7M58.717.8
✓Multi†39.7M63.418.2
✓Multi49.2M66.422.1
✓Single37.4M66.525.3
+ +Table 4: Results with different representations on 3D lane. 'Acc' for category accuracy and 'Error' for X error near. MT means simultaneous multitask 3D lane and 3D object detection. + +
RepresentationF-Score↑Acc↑Error↓
Anchor53.486.00.476
PS58.188.00.545
RepVF66.091.40.439
RepVF + MT66.591.70.420
+ +Representation. Moreover, we have evaluated the effectiveness of different representations for 3D lane detection in Tab. 4, including traditional anchor representations [6,37], simple randomly sampled point set representation (PS), and our proposed representative vector fields (RepVF). This study reveals that RepVF significantly outperforms traditional anchor and PS methods, and incorporating multi-task learning marginally improves these metrics. + +# 5.5 Qualitative Results + +Figure 4 demonstrates the capability of RFTR to simultaneously achieve 3D lane detection and 3D object detection on the extended [24] Waymo open dataset [53]. The figure clearly illustrates that RFTR can accurately predict RepVF and perform both tasks at once. In terms of 3D lane detection, RFTR is capable of predicting the shape of lane lines in areas obscured by vehicles, demonstrating its robustness against occlusion. Regarding 3D object detection, the unified RepVF representation accurately captures the 3D bounding boxes. This dual achievement highlights RFTR's effectiveness in handling complex driving scenarios, enhancing both navigation and safety features in autonomous driving systems. + +# 6 Conclusion + +In this paper, we propose $RepVF$ , a novel unified vector fields representation that characterizes the geometric and semantic structure of different 3D perception targets in the scene. Built upon $RepVF$ , our proposed RFTR exploits the intrinsic connections between tasks, mitigates multi-task gradient imbalance and feature competition, and improves model convergence and multi-task expression. Our work not only contributes a novel single-head perspective to the multi-task learning paradigm in autonomous driving but also sets a promising direction for future research in efficient and effective 3D perception task handling. + +# References + +1. Mobileye at CES 2024. https://www-mobileye.com/ces-2024/ +2. NVIDIA DRIVE Solutions. https://developer.nvidia.com/drive +3. Bai, Y., Chen, Z., Fu, Z., Peng, L., Liang, P., Cheng, E.: CurveFormer: 3D Lane Detection by Curve Propagation with Curve Queries and Attention. In: 2023 IEEE International Conference on Robotics and Automation (ICRA). pp. 7062-7068 (May 2023). https://doi.org/10.1109/ICRA48891.2023.10161160 +4. Caesar, H., Bankiti, V., Lang, A.H., Vora, S., Liong, V.E., Xu, Q., Krishnan, A., Pan, Y., Baldan, G., Beijbom, O.: nuScenes: A Multimodal Dataset for Autonomous Driving. In: 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 11618-11628. IEEE, Seattle, WA, USA (Jun 2020). https://doi.org/10.1109/CVPR42600.2020.01164 +5. Carion, N., Massa, F., Synnaeve, G., Usunier, N., Kirillov, A., Zagoruyko, S.: End-to-End Object Detection with Transformers. In: Vedaldi, A., Bischof, H., Brox, T., Frahm, J.M. (eds.) Computer Vision - ECCV 2020. pp. 213-229. Lecture Notes in Computer Science, Springer International Publishing, Cham (2020). https://doi.org/10.1007/978-3-030-58452-8_13 +6. Chen, L., Sima, C., Li, Y., Zheng, Z., Xu, J., Geng, X., Li, H., He, C., Shi, J., Qiao, Y., Yan, J.: PersFormer: 3D Lane Detection via Perspective Transformer and the OpenLane Benchmark. In: Avidan, S., Brostow, G., Cisse, M., Farinella, G.M., Hassner, T. (eds.) Computer Vision - ECCV 2022, vol. 13698, pp. 550-567. Springer Nature Switzerland, Cham (2022). https://doi.org/10.1007/978-3-031-19839-7_32 +7. Chen, X., Kundu, K., Zhang, Z., Ma, H., Fidler, S., Urtasun, R.: Monocular 3D Object Detection for Autonomous Driving. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 2147-2156 (2016) +8. Chen, Z., Badrinarayanan, V., Lee, C.Y., Rabinovich, A.: GradNorm: Gradient Normalization for Adaptive Loss Balancing in Deep Multitask Networks. In: Proceedings of the 35th International Conference on Machine Learning. pp. 794-803. PMLR (Jul 2018) +9. Cheng, W., Yin, J., Li, W., Yang, R., Shen, J.: Language-Guided 3D Object Detection in Point Cloud for Autonomous Driving (May 2023) +0. Feng, Z., Guo, S., Tan, X., Xu, K., Wang, M., Ma, L.: Rethinking Efficient Lane Detection via Curve Modeling (May 2023). https://doi.org/10.48550/arXiv.2203.02431 +1. Garnett, N., Cohen, R., Pe'er, T., Lahav, R., Levi, D.: 3D-LaneNet: End-to-End 3D Multiple Lane Detection. In: 2019 IEEE/CVF International Conference on Computer Vision (ICCV). pp. 2921–2930. IEEE, Seoul, Korea (South) (Oct 2019). https://doi.org/10.1109/ICCV.2019.00301 +12. Geiger, A., Lenz, P., Urtasun, R.: Are we ready for autonomous driving? The KITTI vision benchmark suite. In: 2012 IEEE Conference on Computer Vision and Pattern Recognition. pp. 3354-3361 (Jun 2012). https://doi.org/10.1109/CVPR.2012.6248074 +13. Guo, P., Lee, C.Y., Ulbricht, D.: Learning to Branch for Multi-Task Learning. In: Proceedings of the 37th International Conference on Machine Learning. pp. 3854-3863. PMLR (Nov 2020) +14. Guo, Y., Chen, G., Zhao, P., Zhang, W., Miao, J., Wang, J., Choe, T.E.: GenLaneNet: A Generalized and Scalable Approach for 3D Lane Detection. In: Vedaldi, A., Bischof, H., Brox, T., Frahm, J.M. (eds.) Computer Vision - ECCV 2020. pp. + +666-681. Lecture Notes in Computer Science, Springer International Publishing, Cham (2020). https://doi.org/10.1007/978-3-030-58589-1_40 +15. Han, W., Shen, J.: Decoupling the Curve Modeling and Pavement Regression for Lane Detection (Sep 2023) +16. He, K., Zhang, X., Ren, S., Sun, J.: Deep Residual Learning for Image Recognition. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 770-778 (2016) +17. Huang, J., Huang, G.: BEVDet4D: Exploit Temporal Cues in Multi-camera 3D Object Detection (Jun 2022) +18. Huang, S., Shen, Z., Huang, Z., Ding, Z.h., Dai, J., Han, J., Wang, N., Liu, S.: Anchor3DLane: Learning To Regress 3D Anchors for Monocular 3D Lane Detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 17451-17460 (2023) +19. Huang, Z., Lin, S., Liu, G., Luo, M., Ye, C., Xu, H., Chang, X., Liang, X.: FULLER: Unified Multi-modality Multi-task 3D Perception via Multi-level Gradient Calibration. In: 2023 IEEE/CVF International Conference on Computer Vision (ICCV). pp. 3479-3488. IEEE, Paris, France (Oct 2023). https://doi.org/10.1109/ICCV51070.2023.00324 +20. Kehl, W., Manhardt, F., Tombari, F., Ilic, S., Navab, N.: SSD-6D: Making RGB-Based 3D Detection and 6D Pose Estimation Great Again. In: Proceedings of the IEEE International Conference on Computer Vision. pp. 1521–1529 (2017) +21. Kendall, A., Gal, Y., Cipolla, R.: Multi-Task Learning Using Uncertainty to Weigh Losses for Scene Geometry and Semantics. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 7482-7491 (2018) +22. Ku, J., Pon, A.D., Waslander, S.L.: Monocular 3D Object Detection Leveraging Accurate Proposals and Shape Reconstruction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11867-11876 (2019) +23. Kuhn, H.W.: The Hungarian method for the assignment problem. Naval Research Logistics Quarterly 2(1-2), 83-97 (1955). https://doi.org/10.1002/nav.3800020109 +24. Li, L.H., Yatskar, M., Yin, D., Hsieh, C.J., Chang, K.W.: VisualBERT: A Simple and Performant Baseline for Vision and Language (Aug 2019). https://doi.org/10.48550/arXiv.1908.03557 +25. Li, X., Yin, J., Li, W., Xu, C., Yang, R., Shen, J.: Di-v2x: Learning domain-invariant representation for vehicle-infrastructure collaborative 3d object detection. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 38, pp. 3208-3215 (2024) +26. Li, X., Yin, J., Shi, B., Li, Y., Yang, R., Shen, J.: Lwsis: Lidar-guided weakly supervised instance segmentation for autonomous driving. Proceedings of the AAAI Conference on Artificial Intelligence 37(2), 1433-1441 (Jun 2023). https://doi.org/10.1609/aaii.v37i2.25228, https://ojs.aaii.org/index.php/AAAI/article/view/25228 +27. Li, Y., Chen, Y., Qi, X., Li, Z., Sun, J., Jia, J.: Unifying Voxel-based Representation with Transformer for 3D Object Detection. Advances in Neural Information Processing Systems 35, 18442-18455 (Dec 2022) +28. Li, Z., Wang, W., Li, H., Xie, E., Sima, C., Lu, T., Qiao, Y., Dai, J.: BEVFormer: Learning Bird's-Eye-View Representation from Multi-camera Images via Spatiotemporal Transformers. In: Avidan, S., Brostow, G., Cisse, M., Farinella, G.M., Hassner, T. (eds.) Computer Vision - ECCV 2022. pp. 1-18. Lecture Notes in Computer Science, Springer Nature Switzerland, Cham (2022). https://doi.org/10.1007/978-3-031-20077-9_1 + +29. Liang, M., Yang, B., Chen, Y., Hu, R., Urtasun, R.: Multi-Task Multi-Sensor Fusion for 3D Object Detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7345-7353 (2019) +30. Liang, T., Xie, H., Yu, K., Xia, Z., Lin, Z., Wang, Y., Tang, T., Wang, B., Tang, Z.: BEVFusion: A Simple and Robust LiDAR-Camera Fusion Framework. Advances in Neural Information Processing Systems 35, 10421-10434 (Dec 2022) +31. Lin, T.Y., Dollar, P., Girshick, R., He, K., Hariharan, B., Belongie, S.: Feature Pyramid Networks for Object Detection. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 2117-2125 (2017) +32. Lin, T.Y., Goyal, P., Girshick, R., He, K., Dollar, P.: Focal Loss for Dense Object Detection. In: Proceedings of the IEEE International Conference on Computer Vision. pp. 2980-2988 (2017) +33. Liu, L., Li, Y., Kuang, Z., Xue, J.H., Chen, Y., Yang, W., Liao, Q., Zhang, W.: Towards Impartial Multi-task Learning. In: International Conference on Learning Representations (Oct 2020) +34. Liu, R., Chen, D., Liu, T., Xiong, Z., Yuan, Z.: Learning to Predict 3D Lane Shape and Camera Pose from a Single Image via Geometry Constraints. Proceedings of the AAAI Conference on Artificial Intelligence 36(2), 1765-1772 (Jun 2022). https://doi.org/10.1609/aaai.v36i2.20069 +35. Liu, S., Johns, E., Davison, A.J.: End-to-End Multi-Task Learning with Attention (Apr 2019). https://doi.org/10.48550/arXiv.1803.10704 +36. Liu, S., Li, F., Zhang, H., Yang, X., Qi, X., Su, H., Zhu, J., Zhang, L.: DAB-DETR: DYNAMIC ANCHOR BOXES ARE BETTER QUERIES FOR DETR (2022) +37. Liu, Y., Wang, T., Zhang, X., Sun, J.: PETR: Position Embedding Transformation for Multi-view 3D Object Detection. In: Avidan, S., Brostow, G., Cisse, M., Farinella, G.M., Hassner, T. (eds.) Computer Vision - ECCV 2022, vol. 13687, pp. 531-548. Springer Nature Switzerland, Cham (2022). https://doi.org/10.1007/978-3-031-19812-0_31 +38. Liu, Y., Yan, J., Jia, F., Li, S., Gao, A., Wang, T., Zhang, X.: PETRv2: A Unified Framework for 3D Perception from Multi-Camera Images +39. Liu, Z., Tang, H., Amini, A., Yang, X., Mao, H., Rus, D.L., Han, S.: BEVFusion: Multi-Task Multi-Sensor Fusion with Unified Bird's-Eye View Representation. In: 2023 IEEE International Conference on Robotics and Automation (ICRA). pp. 2774-2781 (May 2023). https://doi.org/10.1109/ICRA48891.2023.10160968 +40. Loshchilov, I., Hutter, F.: SGDR: Stochastic Gradient Descent with Warm Restarts. In: International Conference on Learning Representations (Nov 2016) +41. Loshchilov, I., Hutter, F.: DECOUPLED WEIGHT DECAY REGULARIZATION (2019) +42. Luo, Y., Zheng, C., Yan, X., Kun, T., Zheng, C., Cui, S., Li, Z.: LATR: 3D Lane Detection from Monocular Images with Transformer. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 7941-7952 (2023) +43. Ma, Z., Wang, L., Zhang, H., Lu, W., Yin, J.: RPT: Learning Point Set Representation for Siamese Visual Tracking. In: Bartoli, A., Fusiello, A. (eds.) Computer Vision - ECCV 2020 Workshops. pp. 653-665. Lecture Notes in Computer Science, Springer International Publishing, Cham (2020). https://doi.org/10.1007/978-3-030-68238-5_43 +44. Misra, I., Shrivastava, A., Gupta, A., Hebert, M.: Cross-Stitch Networks for Multi-Task Learning. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 3994–4003 (2016) + +45. Mousavian, A., Anguelov, D., Flynn, J., Kosecka, J.: 3D Bounding Box Estimation Using Deep Learning and Geometry. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 7074-7082 (2017) +46. Navon, A., Shamsian, A., Achituve, I., Maron, H., Kawaguchi, K., Chechik, G., Fetaya, E.: Multi-Task Learning as a Bargaining Game (Feb 2022) +47. Neven, D., De Brabandere, B., Georgoulis, S., Proesmans, M., Van Gool, L.: Fast Scene Understanding for Autonomous Driving (Aug 2017). https://doi.org/10.48550/arXiv.1708.02550 +48. Philion, J., Fidler, S.: Lift, Splat, Shoot: Encoding Images from Arbitrary Camera Rigs by Implicitly Unprojecting to 3D. In: Vedaldi, A., Bischof, H., Brox, T., Frahm, J.M. (eds.) Computer Vision - ECCV 2020. pp. 194-210. Lecture Notes in Computer Science, Springer International Publishing, Cham (2020). https://doi.org/10.1007/978-3-030-58568-6_12 +49. Reading, C., Harakeh, A., Chae, J., Waslander, S.L.: Categorical Depth Distribution Network for Monocular 3D Object Detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8555-8564 (2021) +50. Roddick, T., Kendall, A., Cipolla, R.: Orthographic Feature Transform for Monocular 3D Object Detection (Nov 2018). https://doi.org/10.48550/arXiv.1811.08188 +51. Ruder, S., Bingel, J., Augenstein, I., Søgaard, A.: Latent Multi-Task Architecture Learning. Proceedings of the AAAI Conference on Artificial Intelligence 33(01), 4822-4829 (Jul 2019). https://doi.org/10.1609/aaai.v33i01.33014822 +52. Sener, O., Koltun, V.: Multi-Task Learning as Multi-Objective Optimization. In: Advances in Neural Information Processing Systems. vol. 31. Curran Associates, Inc. (2018) +53. Sun, P., Kretzschmar, H., Dotiwalla, X., Chouard, A., Patnaik, V., Tsui, P., Guo, J., Zhou, Y., Chai, Y., Caine, B., Vasudevan, V., Han, W., Ngiam, J., Zhao, H., Timofeev, A., Ettinger, S., Krivokon, M., Gao, A., Joshi, A., Zhang, Y., Shlens, J., Chen, Z., Anguelov, D.: Scalability in Perception for Autonomous Driving: Waymo Open Dataset +54. Teichmann, M., Weber, M., Zöllner, M., Cipolla, R., Urtasun, R.: MultiNet: Realtime Joint Semantic Reasoning for Autonomous Driving. In: 2018 IEEE Intelligent Vehicles Symposium (IV). pp. 1013-1020 (Jun 2018). https://doi.org/10.1109/IVS.2018.8500504 +55. Vandenhende, S., Georgoulis, S., Van Gansbeke, W., Proesmans, M., Dai, D., Van Gool, L.: Multi-Task Learning for Dense Prediction Tasks: A Survey. IEEE Transactions on Pattern Analysis and Machine Intelligence 44(7), 3614-3633 (Jul 2022). https://doi.org/10.1109/TPAMI.2021.3054719 +56. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is All you Need. In: Advances in Neural Information Processing Systems. vol. 30. Curran Associates, Inc. (2017) +57. Wang, R., Qin, J., Li, K., Li, Y., Cao, D., Xu, J.: BEV-LaneDet: An Efficient 3D Lane Detection Based on Virtual Camera via Key-Points. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1002-1011 (2023) +58. Wang, S., Liu, Y., Wang, T., Li, Y., Zhang, X.: Exploring Object-Centric Temporal Modeling for Efficient Multi-View 3D Object Detection (Jun 2023) +59. Wang, T., Zhu, X., Pang, J., Lin, D.: FCOS3D: Fully Convolutional One-Stage Monocular 3D Object Detection. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 913–922 (2021) + +60. Wang, Y., Yin, J., Li, W., Frossard, P., Yang, R., Shen, J.: Ssda3d: Semi-supervised domain adaptation for 3d object detection from point cloud. Proceedings of the AAAI Conference on Artificial Intelligence 37(3), 2707-2715 (Jun 2023). https://doi.org/10.1609/aaai.v37i3.25370, https://ojs(aaai.org/index.php/AAAI/article/view/25370 +61. Wang, Y., Zhang, X., Yang, T., Sun, J.: Anchor DETR: Query Design for Transformer-Based Detector. Proceedings of the AAAI Conference on Artificial Intelligence 36(3), 2567-2575 (Jun 2022). https://doi.org/10.1609/aaai.v36i3.20158 +62. Wang, Y., Guizilini, V.C., Zhang, T., Wang, Y., Zhao, H., Solomon, J.: DETR3D: 3D Object Detection from Multi-view Images via 3D-to-2D Queries. In: Proceedings of the 5th Conference on Robot Learning. pp. 180-191. PMLR (Jan 2022) +63. Wu, D., Chang, J., Jia, F., Liu, Y., Wang, T., Shen, J.: Topomlp: An simple yet strong pipeline for driving topology reasoning. arXiv preprint arXiv:2310.06753 (2023) +64. Wu, D., Jia, F., Chang, J., Li, Z., Sun, J., Han, C., Li, S., Liu, Y., Ge, Z., Wang, T.: The 1st-place solution for cvpr 2023 openlane topology in autonomous driving challenge. arXiv preprint arXiv:2306.09590 (2023) +65. Xie, E., Yu, Z., Zhou, D., Philion, J., Anandkumar, A., Fidler, S., Luo, P., Alvarez, J.M.: M$^2$BEV: Multi-Camera Joint 3D Detection and Segmentation with Unified Birds-Eye View Representation (Apr 2022). https://doi.org/10.48550/arXiv.2204.05088 +66. Xu, D., Ouyang, W., Wang, X., Sebe, N.: PAD-Net: Multi-Tasks Guided Prediction-and-Distillation Network for Simultaneous Depth Estimation and Scene Parsing. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 675-684 (2018) +67. Yan, F., Nie, M., Cai, X., Han, J., Xu, H., Yang, Z., Ye, C., Fu, Y., Mi, M.B., Zhang, L.: ONCE-3DLanes: Building Monocular 3D Lane Detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 17143-17152 (2022) +68. Yang, Z., Liu, S., Hu, H., Wang, L., Lin, S.: RepPoints: Point Set Representation for Object Detection. In: 2019 IEEE/CVF International Conference on Computer Vision (ICCV). pp. 9656-9665. IEEE, Seoul, Korea (South) (Oct 2019). https://doi.org/10.1109/ICCV.2019.00975 +69. Yao, C., Yu, L., Wu, Y., Jia, Y.: Sparse Point Guided 3D Lane Detection +70. Yin, J., Fang, J., Zhou, D., Zhang, L., Xu, C.Z., Shen, J., Wang, W.: Semi-supervised 3D Object Detection with Proficient Teachers. In: Avidan, S., Bros-tow, G., Cisse, M., Farinella, G.M., Hassner, T. (eds.) Computer Vision - ECCV 2022, vol. 13698, pp. 727-743. Springer Nature Switzerland, Cham (2022). https://doi.org/10.1007/978-3-031-19839-7_42 +71. Yin, J., Shen, J., Chen, R., Li, W., Yang, R., Frossard, P., Wang, W.: Is-fusion: Instance-scene collaborative fusion for multimodal 3d object detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14905-14915 (2024) +72. Yin, J., Wang, W., Meng, Q., Yang, R., Shen, J.: A unified object motion and affinity model for online multi-object tracking. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6768-6777 (2020) +73. Yu, T., Kumar, S., Gupta, A., Levine, S., Hausman, K., Finn, C.: Gradient Surgery for Multi-Task Learning (Dec 2020). https://doi.org/10.48550/arXiv.2001.06782 + +74. Zhang, Y., Wang, C., Wang, X., Zeng, W., Liu, W.: FairMOT: On the Fairness of Detection and Re-Identification in Multiple Object Tracking. International Journal of Computer Vision 129(11), 3069–3087 (Nov 2021). https://doi.org/10.1007/s11263-021-01513-4 +75. Zhou, D., Fang, J., Song, X., Liu, L., Yin, J., Dai, Y., Li, H., Yang, R.: Joint 3d instance segmentation and object detection for autonomous driving. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1839-1849 (2020) +76. Zhou, D., Song, X., Dai, Y., Yin, J., Lu, F., Liao, M., Fang, J., Zhang, L.: Iafa: Instance-aware feature aggregation for 3d object detection from a single image. In: Proceedings of the Asian Conference on Computer Vision (2020) +77. Zhu, X., Su, W., Lu, L., Li, B., Wang, X., Dai, J.: Deformable DETR: Deformable Transformers for End-to-End Object Detection (Mar 2021). https://doi.org/10.48550/arXiv.2010.04159 \ No newline at end of file diff --git a/2024/RepVF_ A Unified Vector Fields Representation for Multi-task 3D Perception/images.zip b/2024/RepVF_ A Unified Vector Fields Representation for Multi-task 3D Perception/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..46fc9fc5f4b8869a3f37ea96dab5c81f1cb97d23 --- /dev/null +++ b/2024/RepVF_ A Unified Vector Fields Representation for Multi-task 3D Perception/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4218d2829bf3cbeab9792e9bb08339d29307f53ff80f4485808b33c6a6ee1079 +size 388415 diff --git a/2024/RepVF_ A Unified Vector Fields Representation for Multi-task 3D Perception/layout.json b/2024/RepVF_ A Unified Vector Fields Representation for Multi-task 3D Perception/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..1a743a0aa67572f48d5d393fadd6fe5656b7271a --- /dev/null +++ b/2024/RepVF_ A Unified Vector Fields Representation for Multi-task 3D Perception/layout.json @@ -0,0 +1,11395 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 140, + 111, + 475, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 140, + 111, + 475, + 148 + ], + "spans": [ + { + "bbox": [ + 140, + 111, + 475, + 148 + ], + "type": "text", + "content": "RepVF: A Unified Vector Fields Representation for Multi-task 3D Perception" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 192, + 167, + 421, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 192, + 167, + 421, + 194 + ], + "spans": [ + { + "bbox": [ + 192, + 167, + 421, + 194 + ], + "type": "text", + "content": "Chunliang Li" + }, + { + "bbox": [ + 192, + 167, + 421, + 194 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 192, + 167, + 421, + 194 + ], + "type": "text", + "content": ", Wencheng Han" + }, + { + "bbox": [ + 192, + 167, + 421, + 194 + ], + "type": "inline_equation", + "content": "^{2*}" + }, + { + "bbox": [ + 192, + 167, + 421, + 194 + ], + "type": "text", + "content": ", Junbo Yin" + }, + { + "bbox": [ + 192, + 167, + 421, + 194 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 192, + 167, + 421, + 194 + ], + "type": "text", + "content": ", Sanyuan Zhao" + }, + { + "bbox": [ + 192, + 167, + 421, + 194 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 192, + 167, + 421, + 194 + ], + "type": "text", + "content": ", and Jianbing Shen" + }, + { + "bbox": [ + 192, + 167, + 421, + 194 + ], + "type": "inline_equation", + "content": "^{2}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 143, + 201, + 469, + 257 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 201, + 469, + 257 + ], + "spans": [ + { + "bbox": [ + 143, + 201, + 469, + 257 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 143, + 201, + 469, + 257 + ], + "type": "text", + "content": " School of Computer Science, Beijing Institute of Technology \n" + }, + { + "bbox": [ + 143, + 201, + 469, + 257 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 143, + 201, + 469, + 257 + ], + "type": "text", + "content": " SKL-IOTSC, Computer and Information Science, University of Macau, China \n{jbji, sanyuanzhao}@bit.edu.cn \n{wencheng256, yinjunbocn}@gmail.com \njianbingshen@um.edu.mo" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 160, + 276, + 455, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 276, + 455, + 518 + ], + "spans": [ + { + "bbox": [ + 160, + 276, + 455, + 518 + ], + "type": "text", + "content": "Abstract. Concurrent processing of multiple autonomous driving 3D perception tasks within the same spatiotemporal scene poses a significant challenge, in particular due to the computational inefficiencies and feature competition between tasks when using traditional multi-task learning approaches. This paper addresses these issues by proposing a novel unified representation, RepVF, which harmonizes the representation of various perception tasks such as 3D object detection and 3D lane detection within a single framework. RepVF characterizes the structure of different targets in the scene through a vector field, enabling a single-head, multi-task learning model that significantly reduces computational redundancy and feature competition. Building upon RepVF, we introduce RFTR, a network designed to exploit the inherent connections between different tasks by utilizing a hierarchical structure of queries that implicitly model the relationships both between and within tasks. This approach eliminates the need for task-specific heads and parameters, fundamentally reducing the conflicts inherent in traditional multi-task learning paradigms. We validate our approach by combining labels from the OpenLane dataset with the Waymo Open dataset. Our work presents a significant advancement in the efficiency and effectiveness of multi-task perception in autonomous driving, offering a new perspective on handling multiple 3D perception tasks synchronously and in parallel. The code will be available at: https://github.com/jbji/RepVF." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 160, + 528, + 464, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 528, + 464, + 540 + ], + "spans": [ + { + "bbox": [ + 160, + 528, + 464, + 540 + ], + "type": "text", + "content": "Keywords: 3D Lane Detection " + }, + { + "bbox": [ + 160, + 528, + 464, + 540 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 528, + 464, + 540 + ], + "type": "text", + "content": " 3D Object Detection " + }, + { + "bbox": [ + 160, + 528, + 464, + 540 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 528, + 464, + 540 + ], + "type": "text", + "content": " Multi-task Method." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 581, + 230, + 594 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 581, + 230, + 594 + ], + "spans": [ + { + "bbox": [ + 132, + 581, + 230, + 594 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 602, + 482, + 627 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 602, + 482, + 627 + ], + "spans": [ + { + "bbox": [ + 130, + 602, + 482, + 627 + ], + "type": "text", + "content": "In autonomous driving, multiple 3D perception tasks often need to be processed synchronously, in real-time, and in parallel. These tasks, which exist within the" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 141, + 632, + 482, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 632, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 141, + 632, + 482, + 665 + ], + "type": "text", + "content": "*Equal contribution. †Corresponding author. This work was supported in part by the FDCT grants 0102/2023/RIA2, 0154/2022/A3, and 001/2024/SKL, the MYRG-CRG2022-00013-IOTSC-ICI grant and the SRG2022-00023-IOTSC grant." + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 149, + 121, + 306, + 248 + ], + "blocks": [ + { + "bbox": [ + 149, + 121, + 306, + 248 + ], + "lines": [ + { + "bbox": [ + 149, + 121, + 306, + 248 + ], + "spans": [ + { + "bbox": [ + 149, + 121, + 306, + 248 + ], + "type": "image", + "image_path": "2bfbbd98ca04d688ded301c786f4503e6fd2dcc793e811e4c5a6dc7ca31f9204.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 155, + 252, + 296, + 262 + ], + "lines": [ + { + "bbox": [ + 155, + 252, + 296, + 262 + ], + "spans": [ + { + "bbox": [ + 155, + 252, + 296, + 262 + ], + "type": "text", + "content": "(a) Multi-tasking paradigm comparison." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 315, + 125, + 471, + 247 + ], + "blocks": [ + { + "bbox": [ + 315, + 125, + 471, + 247 + ], + "lines": [ + { + "bbox": [ + 315, + 125, + 471, + 247 + ], + "spans": [ + { + "bbox": [ + 315, + 125, + 471, + 247 + ], + "type": "image", + "image_path": "c98f0f250df4db5e11794a7c0ede71ed5720cd8dd1abc95d2d1f2b098d52c4aa.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 321, + 252, + 461, + 263 + ], + "lines": [ + { + "bbox": [ + 321, + 252, + 461, + 263 + ], + "spans": [ + { + "bbox": [ + 321, + 252, + 461, + 263 + ], + "type": "text", + "content": "(b) " + }, + { + "bbox": [ + 321, + 252, + 461, + 263 + ], + "type": "inline_equation", + "content": "RepVF" + }, + { + "bbox": [ + 321, + 252, + 461, + 263 + ], + "type": "text", + "content": ", conversion and supervision." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 270, + 482, + 316 + ], + "lines": [ + { + "bbox": [ + 130, + 270, + 482, + 316 + ], + "spans": [ + { + "bbox": [ + 130, + 270, + 482, + 316 + ], + "type": "text", + "content": "Fig. 1: " + }, + { + "bbox": [ + 130, + 270, + 482, + 316 + ], + "type": "inline_equation", + "content": "RepVF" + }, + { + "bbox": [ + 130, + 270, + 482, + 316 + ], + "type": "text", + "content": " is proposed as a unified representation for 3D perception to achieve single-head multi-tasking. It consists of a set of vector fields learned to represent road elements, and is trained utilizing existing labels through differentiable conversions. The proposed single-head multi-tasking paradigm reduces task conflict and competition." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 341, + 482, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 341, + 482, + 533 + ], + "spans": [ + { + "bbox": [ + 130, + 341, + 482, + 533 + ], + "type": "text", + "content": "same spatiotemporal scene, are often complementary to each other in both geometric and semantic aspects [19]. For example, 3D object detection [12] and 3D lane detection [11] define foreground objects and background lanes within the same 3D scene. Vehicles usually travel following the guiding semantics of lanes that do not spatially overlap and are partially obscured by other vehicles. Some previous industrial solutions [1,2] utilize a multi-model paradigm for multi-task perceptions, as depicted in Figure 1a(1). These separate models cannot share common features across different tasks, resulting in computational waste. Furthermore, this paradigm does not take advantage of the inherent connections between different tasks. Some works [19,28,38,39] attempt to overcome this limitation by using a common backbone and separate head networks for each task, as shown in Figure 1a(2). This approach can considerably reduce the computational cost when handling a large number of tasks, but it might lead to a task balancing issue. Since different task-aware heads focus on various aspects of the scene, they may compete for the backbone features, resulting in instability during training [19,39]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 533, + 482, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 482, + 640 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 482, + 640 + ], + "type": "text", + "content": "The competition is attributed to the variation in representation between different tasks. For example, the 3D detection task [12,26,28,37,38,60,62] guides the head network to predict 3D bounding boxes, whereas the lane detection task [6,11] necessitates the corresponding head to regress lane shapes. Specifically, 3D bounding boxes describe a cubic area in 3D space, representing different category foregrounds using position, size, and orientation. In contrast, 3D lanes are represented using anchor [6,11,18] or parametric [15] methods to describe one or more 1-D lines. These representations focus on different aspects of the scenes, which ultimately leads to competition." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 641, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 641, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 131, + 641, + 482, + 666 + ], + "type": "text", + "content": "In this paper, we propose a unified representation, " + }, + { + "bbox": [ + 131, + 641, + 482, + 666 + ], + "type": "inline_equation", + "content": "RepVF" + }, + { + "bbox": [ + 131, + 641, + 482, + 666 + ], + "type": "text", + "content": " (Representative Vector Fields), for perception tasks in autonomous driving. Using this represent" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 212, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 212, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 212, + 100 + ], + "type": "text", + "content": "C. Li et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 272 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 272 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 272 + ], + "type": "text", + "content": "tation, we design a single-headed, unified model that can perform tasks simultaneously. As illustrated in Fig. 1b, " + }, + { + "bbox": [ + 130, + 116, + 482, + 272 + ], + "type": "inline_equation", + "content": "RepVF" + }, + { + "bbox": [ + 130, + 116, + 482, + 272 + ], + "type": "text", + "content": " represents 3D vector fields that each assigns vectors to spatial locations, denoted as " + }, + { + "bbox": [ + 130, + 116, + 482, + 272 + ], + "type": "inline_equation", + "content": "(S, \\mathcal{F}(S)) \\subseteq \\mathbb{R}^{3 + d}" + }, + { + "bbox": [ + 130, + 116, + 482, + 272 + ], + "type": "text", + "content": " in Sec. 3.2. Sub-vector fields representing different targets in the 3D perception scene progressively adhere to their spatial extent, accurately characterizing the anisotropic structure in their respective spatial vicinity. " + }, + { + "bbox": [ + 130, + 116, + 482, + 272 + ], + "type": "inline_equation", + "content": "RepVF" + }, + { + "bbox": [ + 130, + 116, + 482, + 272 + ], + "type": "text", + "content": " can be differentiably transformed into task-specific elements, hence requiring no special supervision and can utilize existing labels. With " + }, + { + "bbox": [ + 130, + 116, + 482, + 272 + ], + "type": "inline_equation", + "content": "RepVF" + }, + { + "bbox": [ + 130, + 116, + 482, + 272 + ], + "type": "text", + "content": ", our network only processes one single type of fundamental perception element. As shown in Fig. 1a, unlike the traditional multi-task approach that employs multiple task heads [19, 28, 30, 38, 72], we follow an unprecedented single-head multi-task paradigm. The single-head multi-task eliminates task-specific parameters and obviates the need to explicitly model task interactions." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 272, + 482, + 463 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 272, + 482, + 463 + ], + "spans": [ + { + "bbox": [ + 130, + 272, + 482, + 463 + ], + "type": "text", + "content": "RepVF representation serve as a geometrical, cross-scale 3D perception representation alternative. Unlike perception elements capable of handling only single-scale tasks [6,11,12,62], it breaks down the scene into smaller perceptual units while retaining the capability for target-level perception at larger scales. Based on the RepVF representation, we further propose RFTR to leverage this relationship, utilizing queries with a hierarchical structure that implicitly models the relationships both between and within different tasks. Following the single-head multi-task architecture, with only shared parameters across tasks, RFTR does not depend on existing multi-task optimization strategies [21,33,35,52,73]. The gradient discrepancies between tasks across different training iterations have been mitigated, fundamentally reducing the competition and conflict inherent in multi-tasking 3D perception. By employing RepVF to replace the fundamental elements for different tasks, and by combining 3D lane labels from the OpenLane dataset [6] with the Waymo Open dataset [53] for simultaneous multi-task training and inference, our approach achieves comparable or even better multi-task performance than single-task counterparts with only one unified model." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 463, + 482, + 584 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 463, + 482, + 584 + ], + "spans": [ + { + "bbox": [ + 130, + 463, + 482, + 584 + ], + "type": "text", + "content": "In summary, we introduce " + }, + { + "bbox": [ + 130, + 463, + 482, + 584 + ], + "type": "inline_equation", + "content": "RepVF" + }, + { + "bbox": [ + 130, + 463, + 482, + 584 + ], + "type": "text", + "content": " as a unified representation for multiple perception tasks in autonomous driving, effectively reducing the competition among different task representations during training and enhancing model convergence. We further develop a single-head multi-task framework RFTR based on " + }, + { + "bbox": [ + 130, + 463, + 482, + 584 + ], + "type": "inline_equation", + "content": "RepVF" + }, + { + "bbox": [ + 130, + 463, + 482, + 584 + ], + "type": "text", + "content": ", fully exploiting the intrinsic connections among different tasks within a scene, improving feature interaction with queries, and enabling gradient-balanced multi-task training. Our approach achieves impressive performance across two key tasks in autonomous driving—3D object detection and 3D lane detection—showcasing the advantages of a unified task representation in the context of autonomous driving applications." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 599, + 238, + 612 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 599, + 238, + 612 + ], + "spans": [ + { + "bbox": [ + 132, + 599, + 238, + 612 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 623, + 432, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 623, + 432, + 635 + ], + "spans": [ + { + "bbox": [ + 130, + 623, + 432, + 635 + ], + "type": "text", + "content": "2.1 Camera-based 3D Perception in Autonomous Driving" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "content": "3D object detection [7,9,12,62,70] and 3D lane detection [11,14] are two common 3D perception tasks, and previous approaches usually perform the two tasks" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 276, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 276, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 276, + 91, + 447, + 102 + ], + "type": "text", + "content": "RepVF: Unified Multi-task 3D Perception" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "content": "independently. 3D object detection has been extensively studied either in model architectures [7, 37, 38, 59] or feature representations [17, 25, 27, 28, 49, 50, 71, 76]. Early methods [20, 22, 45] predict 3D bounding boxes using 2D detectors, and Mono3D [7] is one of the first methods directly make predictions from 3D representations. Motivated by DETR [11], DETR3D [62] uses sparse 3D object centers as queries, and PETR [37] further enhances its feature representations by embedding 3D positions. Some recent approaches [17, 28, 38, 58] also focus on incorporating temporal information into 3D object detection, but they all follow the transformer-based [56] paradigm introduced by DETR3D [62]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 224, + 482, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 224, + 482, + 331 + ], + "spans": [ + { + "bbox": [ + 130, + 224, + 482, + 331 + ], + "type": "text", + "content": "Only recently have DETR-based methods been introduced for 3D lane detection [3, 38, 42, 63, 64]. Previous methods [6, 11, 14, 34, 57] prefer to project image features into BEV space using Inverse Perspective Mapping (IPM), but this introduces prohibitive height prediction error. Some work [67] avoids IPM via depth estimation, and Anchor-3DLane [18] reverses this projection process. Following DETR [3, 38, 61] introduce 3d lane anchors as queries, and [42] adopts finer lane prior queries and is end-to-end. With the advances of DETR [61, 62] networks in both tasks [3, 28, 37, 42], it is architecturally ready to unify the two tasks without standalone models." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 354, + 289, + 366 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 354, + 289, + 366 + ], + "spans": [ + { + "bbox": [ + 132, + 354, + 289, + 366 + ], + "type": "text", + "content": "2.2 Multi-task 3D Perception" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 377, + 482, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 377, + 482, + 496 + ], + "spans": [ + { + "bbox": [ + 130, + 377, + 482, + 496 + ], + "type": "text", + "content": "The problem of multi-tasking in autonomous driving perception is still in its infancy. Recent advances [19, 28, 37-39, 48, 65, 75] elaborate on universal perception features and multiple tasks are handled with different heads. Moreover, there is no consensus on the required tasks. For example, MMF [29] addresses both depth completion and object detection issues, while [19, 28, 38, 39] performs object detection and map segmentation, with [38] performing additional 3D lane detection separately. However, most of these studies only use multiple tasks to demonstrate the universality of their learned features, and the tasks are trained and evaluated separately [28, 38] rather than simultaneously [19, 39], not to mention dataset discrepancies." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "text", + "content": "In addition, multitask learning (MTL) typically involves task conflict. Previous MTL studies [55] addresses this either through parameter sharing [13, 44, 47, 51, 54, 66] or optimization [8, 21, 33, 35, 52, 73]. Parameter sharing approaches are conceptually straightforward, and are further classified into hard-sharing (with shared and task-specific parameters) and soft-sharing (with a cross-task talk mechanism). Optimization-based approaches, e.g. IMTL [33] and DWA [35], attribute MTL imbalance to the gradient and loss imbalance and explore optimization calibration methods. Taking advantage of these advances, FULLER [19] stands as the pioneering work in autonomous driving perception that analyzes the performance of modern optimization-based MTL calibration methods [8, 33, 35]. With a unified task representation, we follow a different paradigm from previous ones, that eliminate task-specific heads and mitigate the gradient imbalance. Our approach can be seen as a hard-sharing technique without task-specific parameters and does not require calibration." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 211, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 211, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 211, + 100 + ], + "type": "text", + "content": "C. Li et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 326, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 326, + 129 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 326, + 129 + ], + "type": "text", + "content": "3 Representative Vector Fields" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 140, + 482, + 212 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 140, + 482, + 212 + ], + "spans": [ + { + "bbox": [ + 130, + 140, + 482, + 212 + ], + "type": "text", + "content": "We begin by revisiting the representations of two common 3D perception tasks: 3D object detection [4, 12, 53] and 3D lane detection [6, 67]. By exploring how task-specific representations have evolved into task-specific designs [3, 6, 7, 11, 37, 38, 42, 50, 62], we propose the task-agnostic " + }, + { + "bbox": [ + 130, + 140, + 482, + 212 + ], + "type": "inline_equation", + "content": "RepVF" + }, + { + "bbox": [ + 130, + 140, + 482, + 212 + ], + "type": "text", + "content": " (Representative Vector Fields), which arise from the geometric commonality in the tasks and effectively reduce task competition via unified representations." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 228, + 405, + 241 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 228, + 405, + 241 + ], + "spans": [ + { + "bbox": [ + 131, + 228, + 405, + 241 + ], + "type": "text", + "content": "3.1 Task-specific Representations and Single-tasking" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 247, + 482, + 356 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 247, + 482, + 356 + ], + "spans": [ + { + "bbox": [ + 130, + 247, + 482, + 356 + ], + "type": "text", + "content": "Representation Formulations. Task-specific representations vary due to the geometric features of the task targets, which differ in dimensions and scales. As a key task for understanding 3D space, 3D object detection deals with objects irregularly shaped and scaled from one to several metres, defined by coarse outer boundaries as 3D bounding boxes [12], formally a 7-d coordinate " + }, + { + "bbox": [ + 130, + 247, + 482, + 356 + ], + "type": "inline_equation", + "content": "B = (x,y,z,l,w,h,\\theta)" + }, + { + "bbox": [ + 130, + 247, + 482, + 356 + ], + "type": "text", + "content": " encoding object center " + }, + { + "bbox": [ + 130, + 247, + 482, + 356 + ], + "type": "inline_equation", + "content": "(x,y,z)" + }, + { + "bbox": [ + 130, + 247, + 482, + 356 + ], + "type": "text", + "content": ", box dimensions " + }, + { + "bbox": [ + 130, + 247, + 482, + 356 + ], + "type": "inline_equation", + "content": "(l,w,h)" + }, + { + "bbox": [ + 130, + 247, + 482, + 356 + ], + "type": "text", + "content": " and heading " + }, + { + "bbox": [ + 130, + 247, + 482, + 356 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 130, + 247, + 482, + 356 + ], + "type": "text", + "content": ". Conversely, due to their linear shape and ability to cover long distances, 3d lanes are not described by boundaries but by direct geometries, typically ordered sequences of " + }, + { + "bbox": [ + 130, + 247, + 482, + 356 + ], + "type": "inline_equation", + "content": "N_{L}" + }, + { + "bbox": [ + 130, + 247, + 482, + 356 + ], + "type": "text", + "content": " 3D point coordinates [6]:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 205, + 364, + 481, + 377 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 205, + 364, + 481, + 377 + ], + "spans": [ + { + "bbox": [ + 205, + 364, + 481, + 377 + ], + "type": "interline_equation", + "content": "L = \\left[ \\left(x _ {1}, y _ {1}, z _ {1}\\right), \\left(x _ {2}, y _ {2}, z _ {2}\\right), \\dots , \\left(x _ {N _ {l}}, z _ {N _ {l}}, z _ {N _ {l}}\\right) \\right] \\tag {1}", + "image_path": "b1a49831d3eaef49a7ff218125529dfa9bee68350ae155394cfffca265f0c00e.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 384, + 482, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 384, + 482, + 434 + ], + "spans": [ + { + "bbox": [ + 130, + 384, + 482, + 434 + ], + "type": "text", + "content": "This may look familiar to an unordered point cloud, but they are ordered along the direction of the road. In addition, 3D perception problems typically involve classifications that require category scores " + }, + { + "bbox": [ + 130, + 384, + 482, + 434 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_i" + }, + { + "bbox": [ + 130, + 384, + 482, + 434 + ], + "type": "text", + "content": ", varying from 1, 3 or 23 [4,12,53] classes (3D object) to 21 or 22 [6] classes (3D lane)." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 449, + 482, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 449, + 482, + 557 + ], + "spans": [ + { + "bbox": [ + 130, + 449, + 482, + 557 + ], + "type": "text", + "content": "Evolving into Task-specific Designs. Pioneering 3D object detection from RGB images is proposal based, e.g. Mono3D [7] reduces proposal space dimension of each class to " + }, + { + "bbox": [ + 130, + 449, + 482, + 557 + ], + "type": "inline_equation", + "content": "(x,y,z,\\theta,t)" + }, + { + "bbox": [ + 130, + 449, + 482, + 557 + ], + "type": "text", + "content": " with representative templates, and only predicts " + }, + { + "bbox": [ + 130, + 449, + 482, + 557 + ], + "type": "inline_equation", + "content": "\\varDelta_{\\mathrm{pos}}" + }, + { + "bbox": [ + 130, + 449, + 482, + 557 + ], + "type": "text", + "content": "; OFT-Net [50] predicts offsets of all bounding box parameters relative to the top-view grid, i.e. " + }, + { + "bbox": [ + 130, + 449, + 482, + 557 + ], + "type": "inline_equation", + "content": "\\varDelta_{\\mathrm{pos}},\\varDelta_{\\mathrm{dim}},\\varDelta_{\\mathrm{ang}}" + }, + { + "bbox": [ + 130, + 449, + 482, + 557 + ], + "type": "text", + "content": ". Taking advantage of the center coordinates, DETR-based methods [37,62] regress " + }, + { + "bbox": [ + 130, + 449, + 482, + 557 + ], + "type": "inline_equation", + "content": "\\varDelta_{\\mathrm{pos}}" + }, + { + "bbox": [ + 130, + 449, + 482, + 557 + ], + "type": "text", + "content": " along with bounding box parameters. However, constrained by the limited spatial extent of center points, extracted features around them capture only local spatial areas [77], failing to capture the finer pose and structure of objects." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "text", + "content": "3D lines are suitable for anchor-based representations [6, 11, 38], resulting in equally spaced longitudinal lines and anchor-based predictions " + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\varDelta_{\\mathrm{xz}}" + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "text", + "content": ". Anchor representations reduce the representation dimension, but also restrict the degrees of freedom, limiting the maximum model capacity, which has led some works [6, 18] focus on more sophisticated anchor representations. Recent DETR-based methods [3,38,42] use anchor coordinates as queries, exploring the full potential of anchors. Unfortunately, since it's not trivial and stable to directly predict parameter-driven curves [10, 15], mainstream methods stick to predicting anchor-based offsets " + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\varDelta_{\\mathrm{xz}}" + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "text", + "content": ", and nearly no one ever worked directly on the raw lane format." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 276, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 276, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 276, + 91, + 447, + 102 + ], + "type": "text", + "content": "RepVF: Unified Multi-task 3D Perception" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "content": "Intersection of Task-specific Representations. We believe that it's already been architecturally ready to unify the two (and possibly more) perception tasks, with the population and success of DETR practices [19,28,38,61]. However, they still rely on multi-head design due to the discrepancy in task representation, which consumes more computational resources. Despite the significant differences in the dynamics and scales of the described targets, these task-specific representations only encapsulate descriptions of shapes, structures, and directions (object heading and lane ordinality). It is plausible to represent these task targets universally, taking into account perception hierarchies." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 238, + 418, + 251 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 238, + 418, + 251 + ], + "spans": [ + { + "bbox": [ + 131, + 238, + 418, + 251 + ], + "type": "text", + "content": "3.2 The Task-Agnostic Representation through RepVF" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 255, + 482, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 255, + 482, + 327 + ], + "spans": [ + { + "bbox": [ + 130, + 255, + 482, + 327 + ], + "type": "text", + "content": "To facilitate efficient design in multitasking frameworks, we introduce " + }, + { + "bbox": [ + 130, + 255, + 482, + 327 + ], + "type": "inline_equation", + "content": "RepVF" + }, + { + "bbox": [ + 130, + 255, + 482, + 327 + ], + "type": "text", + "content": ", a vector fields representation inspired by 2D point set object representation [43,43]. Distinct from conventional isotropic points, " + }, + { + "bbox": [ + 130, + 255, + 482, + 327 + ], + "type": "inline_equation", + "content": "RepVF" + }, + { + "bbox": [ + 130, + 255, + 482, + 327 + ], + "type": "text", + "content": " assigns a vector to each location making itself anisotropic. We define " + }, + { + "bbox": [ + 130, + 255, + 482, + 327 + ], + "type": "inline_equation", + "content": "RepVF" + }, + { + "bbox": [ + 130, + 255, + 482, + 327 + ], + "type": "text", + "content": " as a collection of vector fields, each field denoted by " + }, + { + "bbox": [ + 130, + 255, + 482, + 327 + ], + "type": "inline_equation", + "content": "R_{i}" + }, + { + "bbox": [ + 130, + 255, + 482, + 327 + ], + "type": "text", + "content": " as a perception target, whose semantic vector assignment mapping " + }, + { + "bbox": [ + 130, + 255, + 482, + 327 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_i" + }, + { + "bbox": [ + 130, + 255, + 482, + 327 + ], + "type": "text", + "content": " is defined on " + }, + { + "bbox": [ + 130, + 255, + 482, + 327 + ], + "type": "inline_equation", + "content": "D = \\bigcup_{i=1}^{n_s} S_i \\subseteq \\mathbb{R}^3" + }, + { + "bbox": [ + 130, + 255, + 482, + 327 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 202, + 331, + 480, + 345 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 202, + 331, + 480, + 345 + ], + "spans": [ + { + "bbox": [ + 202, + 331, + 480, + 345 + ], + "type": "interline_equation", + "content": "\\{R _ {i} \\}, \\quad R _ {i} = \\left(S _ {i}, \\mathcal {F} _ {i} \\left(S _ {i}\\right)\\right) \\subseteq \\mathbb {R} ^ {3 + d}, \\quad \\mathcal {F}: D \\rightarrow \\mathbb {R} ^ {d} \\tag {2}", + "image_path": "6c75ccc4335903d820be4fc7f32c50b42c588da996bae68fe0a7e574ab4b149b.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 349, + 482, + 433 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 349, + 482, + 433 + ], + "spans": [ + { + "bbox": [ + 130, + 349, + 482, + 433 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 349, + 482, + 433 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 130, + 349, + 482, + 433 + ], + "type": "text", + "content": " is the dimension of the target vector domain. Therefore, for any point " + }, + { + "bbox": [ + 130, + 349, + 482, + 433 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 130, + 349, + 482, + 433 + ], + "type": "text", + "content": " within " + }, + { + "bbox": [ + 130, + 349, + 482, + 433 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 130, + 349, + 482, + 433 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 349, + 482, + 433 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 130, + 349, + 482, + 433 + ], + "type": "text", + "content": " assigns it a vector " + }, + { + "bbox": [ + 130, + 349, + 482, + 433 + ], + "type": "inline_equation", + "content": "\\mathcal{F}(x) \\in \\mathbb{R}^d" + }, + { + "bbox": [ + 130, + 349, + 482, + 433 + ], + "type": "text", + "content": " that encapsulates dimensional and positional attributes of the point. This definition allows us to cover the whole space: The sampling domain " + }, + { + "bbox": [ + 130, + 349, + 482, + 433 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 130, + 349, + 482, + 433 + ], + "type": "text", + "content": " is partitioned into " + }, + { + "bbox": [ + 130, + 349, + 482, + 433 + ], + "type": "inline_equation", + "content": "n_s" + }, + { + "bbox": [ + 130, + 349, + 482, + 433 + ], + "type": "text", + "content": " disjoint subsets " + }, + { + "bbox": [ + 130, + 349, + 482, + 433 + ], + "type": "inline_equation", + "content": "S_i" + }, + { + "bbox": [ + 130, + 349, + 482, + 433 + ], + "type": "text", + "content": ", with each subset containing " + }, + { + "bbox": [ + 130, + 349, + 482, + 433 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 130, + 349, + 482, + 433 + ], + "type": "text", + "content": " representative sampling points: " + }, + { + "bbox": [ + 130, + 349, + 482, + 433 + ], + "type": "inline_equation", + "content": "S_i = \\{(x_k, y_k, z_k)\\}_{k=1}^n" + }, + { + "bbox": [ + 130, + 349, + 482, + 433 + ], + "type": "text", + "content": ". Accordingly, the collective vector field " + }, + { + "bbox": [ + 130, + 349, + 482, + 433 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 130, + 349, + 482, + 433 + ], + "type": "text", + "content": " sprawls over " + }, + { + "bbox": [ + 130, + 349, + 482, + 433 + ], + "type": "inline_equation", + "content": "D = \\bigcup_{i=1}^{n_s} S_i" + }, + { + "bbox": [ + 130, + 349, + 482, + 433 + ], + "type": "text", + "content": ", articulated into " + }, + { + "bbox": [ + 130, + 349, + 482, + 433 + ], + "type": "inline_equation", + "content": "n_s" + }, + { + "bbox": [ + 130, + 349, + 482, + 433 + ], + "type": "text", + "content": " sub-vector fields " + }, + { + "bbox": [ + 130, + 349, + 482, + 433 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_i" + }, + { + "bbox": [ + 130, + 349, + 482, + 433 + ], + "type": "text", + "content": " catering to specific tasks." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 434, + 482, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 434, + 482, + 506 + ], + "spans": [ + { + "bbox": [ + 130, + 434, + 482, + 506 + ], + "type": "text", + "content": "Intuitively, " + }, + { + "bbox": [ + 130, + 434, + 482, + 506 + ], + "type": "inline_equation", + "content": "S_{i}" + }, + { + "bbox": [ + 130, + 434, + 482, + 506 + ], + "type": "text", + "content": " directly outlines entity-level geometric structures, and " + }, + { + "bbox": [ + 130, + 434, + 482, + 506 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_i(S_i)" + }, + { + "bbox": [ + 130, + 434, + 482, + 506 + ], + "type": "text", + "content": " delineates contextual characteristics (i.e. " + }, + { + "bbox": [ + 130, + 434, + 482, + 506 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 130, + 434, + 482, + 506 + ], + "type": "text", + "content": " for bounding box " + }, + { + "bbox": [ + 130, + 434, + 482, + 506 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 130, + 434, + 482, + 506 + ], + "type": "text", + "content": ", or adjacent point directions for lane " + }, + { + "bbox": [ + 130, + 434, + 482, + 506 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 130, + 434, + 482, + 506 + ], + "type": "text", + "content": "). Considering the directionality, we use unit-length vectors parallel to the x-y plane to capture angular behavior, resulting in behavior akin to scalars and " + }, + { + "bbox": [ + 130, + 434, + 482, + 506 + ], + "type": "inline_equation", + "content": "d = 1" + }, + { + "bbox": [ + 130, + 434, + 482, + 506 + ], + "type": "text", + "content": ". To represent proper " + }, + { + "bbox": [ + 130, + 434, + 482, + 506 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_i" + }, + { + "bbox": [ + 130, + 434, + 482, + 506 + ], + "type": "text", + "content": " for 3d lanes, directions of the penultimate points are employed for terminal lane line points." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 519, + 482, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 519, + 482, + 579 + ], + "spans": [ + { + "bbox": [ + 130, + 519, + 482, + 579 + ], + "type": "text", + "content": "Converting RepVF into Task-Specific Representation. To leverage existing task labels for training and evaluation, similar to object detection representations based on point sets [43, 68], we convert predictions " + }, + { + "bbox": [ + 130, + 519, + 482, + 579 + ], + "type": "inline_equation", + "content": "RepVF \\hat{R}_i" + }, + { + "bbox": [ + 130, + 519, + 482, + 579 + ], + "type": "text", + "content": ", category " + }, + { + "bbox": [ + 130, + 519, + 482, + 579 + ], + "type": "inline_equation", + "content": "\\hat{C}_i" + }, + { + "bbox": [ + 130, + 519, + 482, + 579 + ], + "type": "text", + "content": " to task-specific representations. We define geometric interpretation processes as differentiable, parameter-free transformation functions:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 231, + 583, + 480, + 597 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 231, + 583, + 480, + 597 + ], + "spans": [ + { + "bbox": [ + 231, + 583, + 480, + 597 + ], + "type": "interline_equation", + "content": "\\mathcal {T} ^ {(\\cdot)}: \\hat {R} _ {i} \\rightarrow \\hat {P} _ {i}, \\quad \\text {s . t .} \\hat {P} _ {i} \\in \\{\\hat {B}, \\hat {L} \\} _ {i} \\tag {3}", + "image_path": "8c5e51430cce3b7d2f1e38f0f1f88f0a0f8513c506be484bacaf1e2808632359.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 601, + 482, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 601, + 482, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 601, + 482, + 667 + ], + "type": "text", + "content": "Specifically, we have " + }, + { + "bbox": [ + 130, + 601, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\mathcal{T}^l:\\hat{R}_i\\to \\hat{L}_i" + }, + { + "bbox": [ + 130, + 601, + 482, + 667 + ], + "type": "text", + "content": " for 3d lanes and " + }, + { + "bbox": [ + 130, + 601, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\mathcal{T}^b:\\hat{R}_i\\rightarrow \\hat{B}_i" + }, + { + "bbox": [ + 130, + 601, + 482, + 667 + ], + "type": "text", + "content": " for 3d objects. For semantic parts, to align " + }, + { + "bbox": [ + 130, + 601, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{C}}_i" + }, + { + "bbox": [ + 130, + 601, + 482, + 667 + ], + "type": "text", + "content": " with task-specific classes, we split it into " + }, + { + "bbox": [ + 130, + 601, + 482, + 667 + ], + "type": "inline_equation", + "content": "n_{(\\cdot)}" + }, + { + "bbox": [ + 130, + 601, + 482, + 667 + ], + "type": "text", + "content": " bins and take max score within each bin, where " + }, + { + "bbox": [ + 130, + 601, + 482, + 667 + ], + "type": "inline_equation", + "content": "n_{(\\cdot)}" + }, + { + "bbox": [ + 130, + 601, + 482, + 667 + ], + "type": "text", + "content": " is the count of task categories. " + }, + { + "bbox": [ + 130, + 601, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\hat{R}_i" + }, + { + "bbox": [ + 130, + 601, + 482, + 667 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 601, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\hat{R}_j" + }, + { + "bbox": [ + 130, + 601, + 482, + 667 + ], + "type": "text", + "content": " remain formally equivalent until contextualized into " + }, + { + "bbox": [ + 130, + 601, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\hat{P}_i" + }, + { + "bbox": [ + 130, + 601, + 482, + 667 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 601, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\hat{P}_j" + }, + { + "bbox": [ + 130, + 601, + 482, + 667 + ], + "type": "text", + "content": ", thus maintaining task agnosticism." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 212, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 212, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 212, + 101 + ], + "type": "text", + "content": "C. Li et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 131, + 116, + 482, + 164 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 116, + 482, + 164 + ], + "spans": [ + { + "bbox": [ + 131, + 116, + 482, + 164 + ], + "type": "text", + "content": "Conversion Function Choice. In our implementation, " + }, + { + "bbox": [ + 131, + 116, + 482, + 164 + ], + "type": "inline_equation", + "content": "\\mathcal{T}^l" + }, + { + "bbox": [ + 131, + 116, + 482, + 164 + ], + "type": "text", + "content": " sorts the unordered points of " + }, + { + "bbox": [ + 131, + 116, + 482, + 164 + ], + "type": "inline_equation", + "content": "\\hat{S}_i" + }, + { + "bbox": [ + 131, + 116, + 482, + 164 + ], + "type": "text", + "content": " by the road direction " + }, + { + "bbox": [ + 131, + 116, + 482, + 164 + ], + "type": "inline_equation", + "content": "(+X" + }, + { + "bbox": [ + 131, + 116, + 482, + 164 + ], + "type": "text", + "content": " in Waymo [53]). Unlike 2D boxes, 3D boxes are not strictly aligned with to " + }, + { + "bbox": [ + 131, + 116, + 482, + 164 + ], + "type": "inline_equation", + "content": "XYZ" + }, + { + "bbox": [ + 131, + 116, + 482, + 164 + ], + "type": "text", + "content": " axes due to headings. Therefore, " + }, + { + "bbox": [ + 131, + 116, + 482, + 164 + ], + "type": "inline_equation", + "content": "\\mathcal{T}^b" + }, + { + "bbox": [ + 131, + 116, + 482, + 164 + ], + "type": "text", + "content": " consists three parts that respectively compute heading, box center and dimensions:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 171, + 171, + 481, + 186 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 171, + 171, + 481, + 186 + ], + "spans": [ + { + "bbox": [ + 171, + 171, + 481, + 186 + ], + "type": "interline_equation", + "content": "\\mathcal {T} _ {h} ^ {b}: \\hat {\\mathcal {F}} _ {i} \\rightarrow \\hat {\\theta} _ {i}; \\quad \\mathcal {T} _ {x y z} ^ {b}: \\hat {S} _ {i} \\rightarrow (\\hat {x} _ {i}, \\hat {y} _ {i}, \\hat {z} _ {i}); \\quad \\mathcal {T} _ {l w h} ^ {b}: \\hat {S} _ {i} \\rightarrow (\\hat {l} _ {i}, \\hat {w} _ {i}, \\hat {h} _ {i}) \\qquad (4)", + "image_path": "7fd8ae5b024aac3f338f8e7b5e59c959b012cb0f5771d909c5c27d438f27c0c2.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 194, + 482, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 194, + 482, + 281 + ], + "spans": [ + { + "bbox": [ + 131, + 194, + 482, + 281 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 131, + 194, + 482, + 281 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_h^b" + }, + { + "bbox": [ + 131, + 194, + 482, + 281 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 194, + 482, + 281 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_{xyz}^b" + }, + { + "bbox": [ + 131, + 194, + 482, + 281 + ], + "type": "text", + "content": " are simple arithmetic averages. With " + }, + { + "bbox": [ + 131, + 194, + 482, + 281 + ], + "type": "inline_equation", + "content": "\\hat{\\theta}_i" + }, + { + "bbox": [ + 131, + 194, + 482, + 281 + ], + "type": "text", + "content": " given by " + }, + { + "bbox": [ + 131, + 194, + 482, + 281 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_h^b" + }, + { + "bbox": [ + 131, + 194, + 482, + 281 + ], + "type": "text", + "content": ", by projecting " + }, + { + "bbox": [ + 131, + 194, + 482, + 281 + ], + "type": "inline_equation", + "content": "\\hat{S}_i" + }, + { + "bbox": [ + 131, + 194, + 482, + 281 + ], + "type": "text", + "content": " to the main orientation axis " + }, + { + "bbox": [ + 131, + 194, + 482, + 281 + ], + "type": "inline_equation", + "content": "\\hat{\\theta}_i" + }, + { + "bbox": [ + 131, + 194, + 482, + 281 + ], + "type": "text", + "content": ", the predictions are aligned with the predicted directions. We then proceed with two candidates of " + }, + { + "bbox": [ + 131, + 194, + 482, + 281 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_{lwh}^b" + }, + { + "bbox": [ + 131, + 194, + 482, + 281 + ], + "type": "text", + "content": " similar to 2D practices [68]: (1) " + }, + { + "bbox": [ + 131, + 194, + 482, + 281 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_{lwh}^b = \\mathcal{T}_1" + }, + { + "bbox": [ + 131, + 194, + 482, + 281 + ], + "type": "text", + "content": ", the min-max function, performing min/max operations projection dimensions for dimensions; (2) " + }, + { + "bbox": [ + 131, + 194, + 482, + 281 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_{lwh}^b = \\mathcal{T}_2" + }, + { + "bbox": [ + 131, + 194, + 482, + 281 + ], + "type": "text", + "content": " employs a momentum-based function taking the variance of predicted points as the boundary box dimensions. In our practice, " + }, + { + "bbox": [ + 131, + 194, + 482, + 281 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_2" + }, + { + "bbox": [ + 131, + 194, + 482, + 281 + ], + "type": "text", + "content": " works slightly better." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 297, + 482, + 406 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 297, + 482, + 406 + ], + "spans": [ + { + "bbox": [ + 131, + 297, + 482, + 406 + ], + "type": "text", + "content": "Learning RepVF. RepVF is supervised by loss functions universally across tasks without special losses. Here we use " + }, + { + "bbox": [ + 131, + 297, + 482, + 406 + ], + "type": "inline_equation", + "content": "\\mathbf{L}^1" + }, + { + "bbox": [ + 131, + 297, + 482, + 406 + ], + "type": "text", + "content": " regression loss on " + }, + { + "bbox": [ + 131, + 297, + 482, + 406 + ], + "type": "inline_equation", + "content": "S_{i}" + }, + { + "bbox": [ + 131, + 297, + 482, + 406 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 297, + 482, + 406 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_i" + }, + { + "bbox": [ + 131, + 297, + 482, + 406 + ], + "type": "text", + "content": ", and focal loss [32] on " + }, + { + "bbox": [ + 131, + 297, + 482, + 406 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_i" + }, + { + "bbox": [ + 131, + 297, + 482, + 406 + ], + "type": "text", + "content": ". Smaller structures are described by points in " + }, + { + "bbox": [ + 131, + 297, + 482, + 406 + ], + "type": "inline_equation", + "content": "S_{i}" + }, + { + "bbox": [ + 131, + 297, + 482, + 406 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 131, + 297, + 482, + 406 + ], + "type": "inline_equation", + "content": "R_{i}" + }, + { + "bbox": [ + 131, + 297, + 482, + 406 + ], + "type": "text", + "content": " itself is set-level, extending RepVF across different perception hierarchies. To capture this cross-level property, we adopt DETR-based structure [37,61,62] and aligned query design to use raw representations motivated by [3, 36]. With multiple decoder layers [62, 77], the learning process of RepVF can be characterized as iterative update " + }, + { + "bbox": [ + 131, + 297, + 482, + 406 + ], + "type": "inline_equation", + "content": "\\varOmega_l" + }, + { + "bbox": [ + 131, + 297, + 482, + 406 + ], + "type": "text", + "content": " on queries " + }, + { + "bbox": [ + 131, + 297, + 482, + 406 + ], + "type": "inline_equation", + "content": "Q_{l}" + }, + { + "bbox": [ + 131, + 297, + 482, + 406 + ], + "type": "text", + "content": " and inferring " + }, + { + "bbox": [ + 131, + 297, + 482, + 406 + ], + "type": "inline_equation", + "content": "\\hat{R}_l,\\hat{C}_l" + }, + { + "bbox": [ + 131, + 297, + 482, + 406 + ], + "type": "text", + "content": " from them with linear layer " + }, + { + "bbox": [ + 131, + 297, + 482, + 406 + ], + "type": "inline_equation", + "content": "\\varPhi" + }, + { + "bbox": [ + 131, + 297, + 482, + 406 + ], + "type": "text", + "content": " and decoder layer " + }, + { + "bbox": [ + 131, + 297, + 482, + 406 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 131, + 297, + 482, + 406 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 236, + 413, + 480, + 426 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 236, + 413, + 480, + 426 + ], + "spans": [ + { + "bbox": [ + 236, + 413, + 480, + 426 + ], + "type": "interline_equation", + "content": "Q _ {l} = \\Omega_ {l} \\left(Q _ {l - 1}\\right), \\quad \\hat {R} _ {l}, \\hat {\\mathcal {C}} _ {l} = \\Phi (Q _ {l}) \\tag {5}", + "image_path": "b73255bc37b5d6a4f55739d734cb6beb1f924eff2061c844bcf47688bc9aa1b8.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 434, + 384, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 434, + 384, + 445 + ], + "spans": [ + { + "bbox": [ + 132, + 434, + 384, + 445 + ], + "type": "text", + "content": "We will explain our network in detail in the next section." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 463, + 373, + 476 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 463, + 373, + 476 + ], + "spans": [ + { + "bbox": [ + 132, + 463, + 373, + 476 + ], + "type": "text", + "content": "4 RFTR: The Single-head Multi-tasker" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 487, + 482, + 546 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 487, + 482, + 546 + ], + "spans": [ + { + "bbox": [ + 131, + 487, + 482, + 546 + ], + "type": "text", + "content": "In this section, we introduce the single-head multi-task framework RFTR (Representative Vector Fields Transformer), which uses " + }, + { + "bbox": [ + 131, + 487, + 482, + 546 + ], + "type": "inline_equation", + "content": "RepVF" + }, + { + "bbox": [ + 131, + 487, + 482, + 546 + ], + "type": "text", + "content": " representation to unify different tasks and exploit intrinsic connections among them. We also observed one interesting behavior of RFTR in the balance of multi-task gradients. RFTR is built upon DETR networks [5,37,62] and is end-to-end on all tasks." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 563, + 256, + 574 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 563, + 256, + 574 + ], + "spans": [ + { + "bbox": [ + 132, + 563, + 256, + 574 + ], + "type": "text", + "content": "4.1 Network Structure" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 131, + 581, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 581, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 131, + 581, + 482, + 666 + ], + "type": "text", + "content": "Figure 2 shows our RFTR architecture built upon " + }, + { + "bbox": [ + 131, + 581, + 482, + 666 + ], + "type": "inline_equation", + "content": "RepVF" + }, + { + "bbox": [ + 131, + 581, + 482, + 666 + ], + "type": "text", + "content": ". It inputs images " + }, + { + "bbox": [ + 131, + 581, + 482, + 666 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 131, + 581, + 482, + 666 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 131, + 581, + 482, + 666 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 131, + 581, + 482, + 666 + ], + "type": "text", + "content": " cameras with known camera parameters, and outputs unified representative vector fields " + }, + { + "bbox": [ + 131, + 581, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 131, + 581, + 482, + 666 + ], + "type": "text", + "content": " for various perception tasks in the 3D scene. RFTR contains four main components that extract features from images, generate set-level queries for hierarchical representation structure, solve unified representative vector fields and multiple tasks with one single head, and convert results to task representations in a differentiable way to utilize existing task labels." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 276, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 276, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 276, + 91, + 447, + 102 + ], + "type": "text", + "content": "RepVF: Unified Multi-task 3D Perception" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 134, + 114, + 479, + 266 + ], + "blocks": [ + { + "bbox": [ + 134, + 114, + 479, + 266 + ], + "lines": [ + { + "bbox": [ + 134, + 114, + 479, + 266 + ], + "spans": [ + { + "bbox": [ + 134, + 114, + 479, + 266 + ], + "type": "image", + "image_path": "867f1b847c5b1f9e3c5f9394f6700d4d85a853583e8c6e4af81047ad62423d23.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 275, + 482, + 353 + ], + "lines": [ + { + "bbox": [ + 130, + 275, + 482, + 353 + ], + "spans": [ + { + "bbox": [ + 130, + 275, + 482, + 353 + ], + "type": "text", + "content": "Fig. 2: Overview of our RFTR (Representative Vector Fields Transformer) built upon RepVF. Multi-view image features with 3D position embeds are extracted by image backbone and fed into the decoder. We generate set-level query embeds from 3D space sampling sets, each representing a perception target. One single unified task head is then used to predict unified representative vector fields. Finally, predictions are transformed in a differentiable manner into task specific representations to utilize existing labels for supervision." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 376, + 482, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 376, + 482, + 426 + ], + "spans": [ + { + "bbox": [ + 130, + 376, + 482, + 426 + ], + "type": "text", + "content": "Feature Extraction. Following common practices in utilizing 2D backbones to extract features [28,37,62], we adopt ResNet [16] (ResNet-50) and FPN [31] to produce 2D features " + }, + { + "bbox": [ + 130, + 376, + 482, + 426 + ], + "type": "inline_equation", + "content": "F^{2d}" + }, + { + "bbox": [ + 130, + 376, + 482, + 426 + ], + "type": "text", + "content": ". They're encoded with 3D position embedding in PETR [37] into 3D position aware features " + }, + { + "bbox": [ + 130, + 376, + 482, + 426 + ], + "type": "inline_equation", + "content": "F^{3d}" + }, + { + "bbox": [ + 130, + 376, + 482, + 426 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 443, + 482, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 443, + 482, + 514 + ], + "spans": [ + { + "bbox": [ + 130, + 443, + 482, + 514 + ], + "type": "text", + "content": "Generating Set-level Perception Queries. As " + }, + { + "bbox": [ + 130, + 443, + 482, + 514 + ], + "type": "inline_equation", + "content": "RepVF" + }, + { + "bbox": [ + 130, + 443, + 482, + 514 + ], + "type": "text", + "content": " is sampled from " + }, + { + "bbox": [ + 130, + 443, + 482, + 514 + ], + "type": "inline_equation", + "content": "n_s" + }, + { + "bbox": [ + 130, + 443, + 482, + 514 + ], + "type": "text", + "content": " point sets " + }, + { + "bbox": [ + 130, + 443, + 482, + 514 + ], + "type": "inline_equation", + "content": "S_i" + }, + { + "bbox": [ + 130, + 443, + 482, + 514 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 130, + 443, + 482, + 514 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 130, + 443, + 482, + 514 + ], + "type": "text", + "content": " points each and is two-level, our aim is to generate queries that are consistent with this hierarchical point-set structure. Hence provide the decoder with sufficient spatial information, improving feature interactions. Unlike former approaches [37, 38, 42, 62] that encode one single 3D location, each query in RFTR encodes multiple locations simultaneously." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 515, + 482, + 562 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 515, + 482, + 562 + ], + "spans": [ + { + "bbox": [ + 130, + 515, + 482, + 562 + ], + "type": "text", + "content": "We generate set-level queries " + }, + { + "bbox": [ + 130, + 515, + 482, + 562 + ], + "type": "inline_equation", + "content": "Q_0 \\in \\mathbb{R}^{n_s \\times d_q}" + }, + { + "bbox": [ + 130, + 515, + 482, + 562 + ], + "type": "text", + "content": " from initial point sets corresponding to " + }, + { + "bbox": [ + 130, + 515, + 482, + 562 + ], + "type": "inline_equation", + "content": "S_i" + }, + { + "bbox": [ + 130, + 515, + 482, + 562 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 515, + 482, + 562 + ], + "type": "inline_equation", + "content": "d_q" + }, + { + "bbox": [ + 130, + 515, + 482, + 562 + ], + "type": "text", + "content": " is the dimension of each query. The generation process " + }, + { + "bbox": [ + 130, + 515, + 482, + 562 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 130, + 515, + 482, + 562 + ], + "type": "text", + "content": " consists of encoding initialized positions from the point level, and embedding them at the set level:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 249, + 563, + 481, + 576 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 249, + 563, + 481, + 576 + ], + "spans": [ + { + "bbox": [ + 249, + 563, + 481, + 576 + ], + "type": "interline_equation", + "content": "\\mathcal {G} = \\operatorname {e m b e d} (\\text {F l a t t e n (P E)}) \\tag {6}", + "image_path": "48a14691b71e99f1635c4cafc3420e7cdb102499952a7a70739c20f13370b2c9.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 581, + 482, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 581, + 482, + 641 + ], + "spans": [ + { + "bbox": [ + 130, + 581, + 482, + 641 + ], + "type": "text", + "content": "where PE is the positional encoding [56]. Point-level PE is flattened into set-level and embedded by a small MLP with two layers. To avoid confusing the model with expansive locations falling on multiple targets in the early stages of training, we use " + }, + { + "bbox": [ + 130, + 581, + 482, + 641 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 130, + 581, + 482, + 641 + ], + "type": "text", + "content": " identical 3D locations sampled with 0-1 uniform distribution for all " + }, + { + "bbox": [ + 130, + 581, + 482, + 641 + ], + "type": "inline_equation", + "content": "n_s" + }, + { + "bbox": [ + 130, + 581, + 482, + 641 + ], + "type": "text", + "content": " initial point sets." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "content": "To our best knowledge, the closest generation approach to ours is [3], which generates query from an ordered sequence of anchor points motivated by [36]." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 212, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 212, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 212, + 100 + ], + "type": "text", + "content": "C. Li et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 175 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 175 + ], + "type": "text", + "content": "However, ours is from an unordered sampling point set and is not task-specific. Moreover, we use a tiny 16-dimensional PE for each coordination, divert from former methods [37, 62] with a much larger 256-dimension. In our multi-task setting, queries are equivalently initialized, thus task-agnostic. We divide the first " + }, + { + "bbox": [ + 130, + 116, + 482, + 175 + ], + "type": "inline_equation", + "content": "80\\%" + }, + { + "bbox": [ + 130, + 116, + 482, + 175 + ], + "type": "text", + "content": " as 3D objects and the last " + }, + { + "bbox": [ + 130, + 116, + 482, + 175 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 130, + 116, + 482, + 175 + ], + "type": "text", + "content": " as 3D lanes." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 190, + 482, + 275 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 190, + 482, + 275 + ], + "spans": [ + { + "bbox": [ + 130, + 190, + 482, + 275 + ], + "type": "text", + "content": "Single-head Multi-tasking. Similar to all previous approaches [5, 37, 62], queries are learnable and updated in decoder layers iteratively. The unified head " + }, + { + "bbox": [ + 130, + 190, + 482, + 275 + ], + "type": "inline_equation", + "content": "\\varPhi" + }, + { + "bbox": [ + 130, + 190, + 482, + 275 + ], + "type": "text", + "content": " of RFTR is composed by a multi-layer transformer decoder [56], and two branches " + }, + { + "bbox": [ + 130, + 190, + 482, + 275 + ], + "type": "inline_equation", + "content": "\\varPhi_{g}" + }, + { + "bbox": [ + 130, + 190, + 482, + 275 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 190, + 482, + 275 + ], + "type": "inline_equation", + "content": "\\varPhi_{s}" + }, + { + "bbox": [ + 130, + 190, + 482, + 275 + ], + "type": "text", + "content": " dedicated for geometric regression and semantic prediction respectively to predict " + }, + { + "bbox": [ + 130, + 190, + 482, + 275 + ], + "type": "inline_equation", + "content": "RepVF" + }, + { + "bbox": [ + 130, + 190, + 482, + 275 + ], + "type": "text", + "content": ". We use the same " + }, + { + "bbox": [ + 130, + 190, + 482, + 275 + ], + "type": "inline_equation", + "content": "n_{L} = 6" + }, + { + "bbox": [ + 130, + 190, + 482, + 275 + ], + "type": "text", + "content": " decoder layers with previous methods [5, 37, 62] to exploit intrinsic tasks connections implicitly. The interaction process within these layers can be formally expressed as:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 228, + 279, + 481, + 293 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 228, + 279, + 481, + 293 + ], + "spans": [ + { + "bbox": [ + 228, + 279, + 481, + 293 + ], + "type": "interline_equation", + "content": "Q _ {l} = \\Omega_ {l} \\left(F ^ {3 d}, Q _ {l - 1}\\right), \\quad l = 1, \\dots , N _ {L} \\tag {7}", + "image_path": "d3ba2b376d70d25b8636ea0b590b047a53e5624b36192380ac7d643a8ecf8fed.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 297, + 481, + 320 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 297, + 481, + 320 + ], + "spans": [ + { + "bbox": [ + 130, + 297, + 481, + 320 + ], + "type": "text", + "content": "With unified head " + }, + { + "bbox": [ + 130, + 297, + 481, + 320 + ], + "type": "inline_equation", + "content": "\\varPhi=\\{\\varPhi_{g},\\varPhi_{s}\\}" + }, + { + "bbox": [ + 130, + 297, + 481, + 320 + ], + "type": "text", + "content": ", predictions " + }, + { + "bbox": [ + 130, + 297, + 481, + 320 + ], + "type": "inline_equation", + "content": "\\hat{R}_{li}" + }, + { + "bbox": [ + 130, + 297, + 481, + 320 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 297, + 481, + 320 + ], + "type": "inline_equation", + "content": "\\hat{C}_{li}" + }, + { + "bbox": [ + 130, + 297, + 481, + 320 + ], + "type": "text", + "content": " is decoded from " + }, + { + "bbox": [ + 130, + 297, + 481, + 320 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 297, + 481, + 320 + ], + "type": "text", + "content": "-th query at " + }, + { + "bbox": [ + 130, + 297, + 481, + 320 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 130, + 297, + 481, + 320 + ], + "type": "text", + "content": "-th layer:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 242, + 321, + 481, + 334 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 242, + 321, + 481, + 334 + ], + "spans": [ + { + "bbox": [ + 242, + 321, + 481, + 334 + ], + "type": "interline_equation", + "content": "\\hat {R} _ {l i} = \\Phi_ {g} \\left(Q _ {l i}\\right), \\quad \\hat {\\mathcal {C}} _ {l i} = \\Phi_ {s} \\left(Q _ {l i}\\right) \\tag {8}", + "image_path": "83e330216d511594583d8ab985ea4af7bc3c6a8888bec317ca2c790652b160d9.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 336, + 481, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 336, + 481, + 361 + ], + "spans": [ + { + "bbox": [ + 131, + 336, + 481, + 361 + ], + "type": "text", + "content": "For " + }, + { + "bbox": [ + 131, + 336, + 481, + 361 + ], + "type": "inline_equation", + "content": "\\varPhi_{g}" + }, + { + "bbox": [ + 131, + 336, + 481, + 361 + ], + "type": "text", + "content": ", additional offsets are added to produce " + }, + { + "bbox": [ + 131, + 336, + 481, + 361 + ], + "type": "inline_equation", + "content": "S_{li}" + }, + { + "bbox": [ + 131, + 336, + 481, + 361 + ], + "type": "text", + "content": ". Unified geometric predictions " + }, + { + "bbox": [ + 131, + 336, + 481, + 361 + ], + "type": "inline_equation", + "content": "\\hat{R}_l = (\\hat{S}_l,\\hat{F}_l)\\subset \\mathbb{R}^{n_s\\times (3 + d_p)}" + }, + { + "bbox": [ + 131, + 336, + 481, + 361 + ], + "type": "text", + "content": " are converted to task-specific representations:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 213, + 366, + 481, + 381 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 213, + 366, + 481, + 381 + ], + "spans": [ + { + "bbox": [ + 213, + 366, + 481, + 381 + ], + "type": "interline_equation", + "content": "\\hat {L} _ {l i} = \\mathcal {T} ^ {l} (\\hat {R} _ {l i}), \\quad \\hat {B} _ {l i} = \\left(\\mathcal {T} _ {x y z} ^ {b}, \\mathcal {T} _ {\\theta} ^ {b}, \\mathcal {T} _ {l w h} ^ {b}\\right) (\\hat {R} _ {l i}) \\tag {9}", + "image_path": "92341abae289403b4946e84437e8455b50cec868ae85064247f6b3ee80bf60c2.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 386, + 482, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 386, + 482, + 434 + ], + "spans": [ + { + "bbox": [ + 130, + 386, + 482, + 434 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 130, + 386, + 482, + 434 + ], + "type": "inline_equation", + "content": "\\mathcal{T}^{(\\cdot)}" + }, + { + "bbox": [ + 130, + 386, + 482, + 434 + ], + "type": "text", + "content": " are differentiable functions predefined in section 3.2. Each head branch " + }, + { + "bbox": [ + 130, + 386, + 482, + 434 + ], + "type": "inline_equation", + "content": "\\varPhi_{(\\cdot)}" + }, + { + "bbox": [ + 130, + 386, + 482, + 434 + ], + "type": "text", + "content": " is a simple small MLP network with three fully connected layers. Following DETR3D [62], we weight the loss for the predictions from each decoder layer " + }, + { + "bbox": [ + 130, + 386, + 482, + 434 + ], + "type": "inline_equation", + "content": "\\Omega_l" + }, + { + "bbox": [ + 130, + 386, + 482, + 434 + ], + "type": "text", + "content": " during training, and only outputs from the last layer is used during inference." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 449, + 482, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 449, + 482, + 496 + ], + "spans": [ + { + "bbox": [ + 130, + 449, + 482, + 496 + ], + "type": "text", + "content": "Loss Function. RFTR uses existing task labels for supervision, and its loss is calculated between converted predictions and task ground truths. It consists of the following components for coordinate regression, vector fields regression and classification respectively:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 197, + 502, + 481, + 526 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 197, + 502, + 481, + 526 + ], + "spans": [ + { + "bbox": [ + 197, + 502, + 481, + 526 + ], + "type": "interline_equation", + "content": "\\mathcal {L} ^ {(\\cdot)} = w _ {r} ^ {(\\cdot)} \\mathcal {L} _ {r} ^ {(\\cdot)} + w _ {f} ^ {(\\cdot)} \\mathcal {L} _ {f} ^ {(\\cdot)} + w _ {c} ^ {(\\cdot)} \\mathcal {L} _ {c} ^ {(\\cdot)}, \\quad \\mathcal {L} = \\sum_ {\\text {t a s k s}} \\mathcal {L} ^ {(\\cdot)} \\tag {10}", + "image_path": "8a7114dd13374b3d7eb321444fe72c8a5b4602a8c81e363f415258f441846837.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 532, + 481, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 532, + 481, + 567 + ], + "spans": [ + { + "bbox": [ + 130, + 532, + 481, + 567 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 532, + 481, + 567 + ], + "type": "inline_equation", + "content": "w^{(\\cdot)}" + }, + { + "bbox": [ + 130, + 532, + 481, + 567 + ], + "type": "text", + "content": " represent loss component weights for task " + }, + { + "bbox": [ + 130, + 532, + 481, + 567 + ], + "type": "inline_equation", + "content": "(\\cdot)" + }, + { + "bbox": [ + 130, + 532, + 481, + 567 + ], + "type": "text", + "content": ", i.e. 3D object detection and 3D lane detection. We use the Hungarian algorithm [23] for label assignment after cost calculation between ground truth and converted predictions " + }, + { + "bbox": [ + 130, + 532, + 481, + 567 + ], + "type": "inline_equation", + "content": "\\hat{B}_{li}, \\hat{P}_{li}" + }, + { + "bbox": [ + 130, + 532, + 481, + 567 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 567, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 567, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 567, + 482, + 666 + ], + "type": "text", + "content": "In practice, we choose " + }, + { + "bbox": [ + 130, + 567, + 482, + 666 + ], + "type": "inline_equation", + "content": "w_{c}^{b} = w_{c}^{l} = 2.0" + }, + { + "bbox": [ + 130, + 567, + 482, + 666 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 567, + 482, + 666 + ], + "type": "inline_equation", + "content": "w_{r}^{l} = 0.003" + }, + { + "bbox": [ + 130, + 567, + 482, + 666 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 567, + 482, + 666 + ], + "type": "inline_equation", + "content": "w_{r}^{b} = 0.1" + }, + { + "bbox": [ + 130, + 567, + 482, + 666 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 567, + 482, + 666 + ], + "type": "inline_equation", + "content": "w_{f}^{b} = 0.2" + }, + { + "bbox": [ + 130, + 567, + 482, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 567, + 482, + 666 + ], + "type": "inline_equation", + "content": "w_{f}^{l} = 0.032" + }, + { + "bbox": [ + 130, + 567, + 482, + 666 + ], + "type": "text", + "content": ". Coordinate regression and classification weights are selected empirically to equalize the scales and contributions of each loss term [74], fields weights are empirically selected. We use " + }, + { + "bbox": [ + 130, + 567, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathrm{L}^1" + }, + { + "bbox": [ + 130, + 567, + 482, + 666 + ], + "type": "text", + "content": " loss for " + }, + { + "bbox": [ + 130, + 567, + 482, + 666 + ], + "type": "inline_equation", + "content": "R_{i}" + }, + { + "bbox": [ + 130, + 567, + 482, + 666 + ], + "type": "text", + "content": " regression, and focal loss [32] with " + }, + { + "bbox": [ + 130, + 567, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\gamma = 2.0" + }, + { + "bbox": [ + 130, + 567, + 482, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 567, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\alpha = 0.25" + }, + { + "bbox": [ + 130, + 567, + 482, + 666 + ], + "type": "text", + "content": " for classification. Since we predict a degenerated scalar field, we supervise the angular prediction by its sine and cosine values. All tasks share the same losses and similar weights, no additional, task-specific loss measures are adopted." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 276, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 276, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 276, + 91, + 447, + 102 + ], + "type": "text", + "content": "RepVF: Unified Multi-task 3D Perception" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 101 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 101 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 101 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 156, + 118, + 342, + 225 + ], + "blocks": [ + { + "bbox": [ + 156, + 118, + 342, + 225 + ], + "lines": [ + { + "bbox": [ + 156, + 118, + 342, + 225 + ], + "spans": [ + { + "bbox": [ + 156, + 118, + 342, + 225 + ], + "type": "image", + "image_path": "9f6a331ed5cd25b2a4805ca119f81d2ff5f3e2f6bfe46e5fbadcb08a203d58fe.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 179, + 230, + 318, + 239 + ], + "lines": [ + { + "bbox": [ + 179, + 230, + 318, + 239 + ], + "spans": [ + { + "bbox": [ + 179, + 230, + 318, + 239 + ], + "type": "text", + "content": "(a) Gradient disparities by iteration index." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 350, + 118, + 457, + 224 + ], + "blocks": [ + { + "bbox": [ + 350, + 118, + 457, + 224 + ], + "lines": [ + { + "bbox": [ + 350, + 118, + 457, + 224 + ], + "spans": [ + { + "bbox": [ + 350, + 118, + 457, + 224 + ], + "type": "image", + "image_path": "2bd0552c0fb3bfb1dbd0f2bdaa5704f10ab8a1549f59c09b67bea00dafe52298.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 350, + 230, + 455, + 239 + ], + "lines": [ + { + "bbox": [ + 350, + 230, + 455, + 239 + ], + "spans": [ + { + "bbox": [ + 350, + 230, + 455, + 239 + ], + "type": "text", + "content": "(b) Average gradient disparities." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 247, + 482, + 314 + ], + "lines": [ + { + "bbox": [ + 130, + 247, + 482, + 314 + ], + "spans": [ + { + "bbox": [ + 130, + 247, + 482, + 314 + ], + "type": "text", + "content": "Fig. 3: Our single-head RFTR model shows a reduced gradient imbalance and better disparity stability (mean: 2.47, variance: 0.90) compared to the multi-head baseline (mean: 2.98 variance: 1.86) or the multi-head baseline with RepVF (mean: 2.72, variance: 1.57). Ideally balanced gradient disparity through training iterations should approach 2. For ease of trend observation, we have clipped the top " + }, + { + "bbox": [ + 130, + 247, + 482, + 314 + ], + "type": "inline_equation", + "content": "2\\%" + }, + { + "bbox": [ + 130, + 247, + 482, + 314 + ], + "type": "text", + "content": " values to the mean, with curves smoothed through a window size of 50 1-d convolution." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 335, + 395, + 347 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 335, + 395, + 347 + ], + "spans": [ + { + "bbox": [ + 131, + 335, + 395, + 347 + ], + "type": "text", + "content": "4.2 Gradient Balance in Single-head Multi-tasking" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 352, + 479, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 352, + 479, + 411 + ], + "spans": [ + { + "bbox": [ + 130, + 352, + 479, + 411 + ], + "type": "text", + "content": "The multi-head multi-task approach often leads to task conflict, attributed in prior studies [21, 33, 46, 52] to gradient imbalance. These studies [33, 46] and more recent work [19] have explored appropriate gradient calibration strategies to address this issue. As we will show, our single-head RFTR naturally achieves better balanced multi-task gradients thus calibration may no longer be required." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 412, + 480, + 448 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 412, + 480, + 448 + ], + "spans": [ + { + "bbox": [ + 130, + 412, + 480, + 448 + ], + "type": "text", + "content": "To elucidate the advantage of our single-head multi-task paradigm over multi-head multi-task settings, we use an improved measure of gradient disparity adapted from FULLER [19], by taking into account the symmetry disparity:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 228, + 453, + 480, + 480 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 228, + 453, + 480, + 480 + ], + "spans": [ + { + "bbox": [ + 228, + 453, + 480, + 480 + ], + "type": "interline_equation", + "content": "\\operatorname {d i f f} \\left(\\nabla \\mathcal {L} ^ {l}, \\nabla \\mathcal {L} ^ {b}\\right) = \\frac {\\| \\nabla \\mathcal {L} ^ {l} \\|}{\\| \\nabla \\mathcal {L} ^ {b} \\|} + \\frac {\\| \\nabla \\mathcal {L} ^ {b} \\|}{\\| \\nabla \\mathcal {L} ^ {l} \\|} \\tag {11}", + "image_path": "3c544036c169895c2a5e3bc045a3a183bd388fcbe0bcbc7c98be1979a857f924.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 485, + 479, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 485, + 479, + 533 + ], + "spans": [ + { + "bbox": [ + 130, + 485, + 479, + 533 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 485, + 479, + 533 + ], + "type": "inline_equation", + "content": "\\| \\cdot \\|" + }, + { + "bbox": [ + 130, + 485, + 479, + 533 + ], + "type": "text", + "content": " is the Frobenius norm and " + }, + { + "bbox": [ + 130, + 485, + 479, + 533 + ], + "type": "inline_equation", + "content": "\\nabla \\mathcal{L}^{(\\cdot)}" + }, + { + "bbox": [ + 130, + 485, + 479, + 533 + ], + "type": "text", + "content": " denotes gradients. Given that " + }, + { + "bbox": [ + 130, + 485, + 479, + 533 + ], + "type": "inline_equation", + "content": "\\| \\nabla \\mathcal{L}^{(\\cdot)}\\|" + }, + { + "bbox": [ + 130, + 485, + 479, + 533 + ], + "type": "text", + "content": " is always greater than 0, the value of diff function is always greater than or equal to 2. Therefore, the more stable and closer to 2 the gradient disparity is during model training, the better the multi-task stability indicated." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 534, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 534, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 534, + 482, + 666 + ], + "type": "text", + "content": "Figure 3a illustrates the gradient disparity curves through training iterations. The multi-head multi-task method [37,38] is under the same task settings with ours. In the sampling 1k iterations, the stability and absolute scale of the gradient disparity curve for the multi-head configuration are inferior to our single-head multi-task structure. The mean and variance of the multi-head disparities are about " + }, + { + "bbox": [ + 130, + 534, + 482, + 666 + ], + "type": "inline_equation", + "content": "108.5\\%" + }, + { + "bbox": [ + 130, + 534, + 482, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 534, + 482, + 666 + ], + "type": "inline_equation", + "content": "107.8\\%" + }, + { + "bbox": [ + 130, + 534, + 482, + 666 + ], + "type": "text", + "content": " higher, respectively, than our single-headed RFTR approach, as shown in Fig. 3b. We have also observed that unified representation and single head design contributed to the improvement about equally. More importantly, this means that our single-head RFTR mitigates, though not eliminating, the gradient imbalance between tasks, which is fundamentally different from task gradient balancing strategies that do not change the gradient itself." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 212, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 212, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 212, + 100 + ], + "type": "text", + "content": "C. Li et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 230, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 230, + 129 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 230, + 129 + ], + "type": "text", + "content": "5 Experiments" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 140, + 482, + 165 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 140, + 482, + 165 + ], + "spans": [ + { + "bbox": [ + 131, + 140, + 482, + 165 + ], + "type": "text", + "content": "We evaluate our method on the Waymo Open Dataset [53] extended by 3D lanes annotated in OpenLane [6], so that we train and evaluate tasks simultaneously." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 182, + 205, + 193 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 182, + 205, + 193 + ], + "spans": [ + { + "bbox": [ + 132, + 182, + 205, + 193 + ], + "type": "text", + "content": "5.1 Datasets" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 202, + 482, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 202, + 482, + 346 + ], + "spans": [ + { + "bbox": [ + 131, + 202, + 482, + 346 + ], + "type": "text", + "content": "Waymo Open Dataset and OpenLane. Waymo Open Dataset (Perception) [53] is a large-scale multimodal dataset composed of data from 5 cameras(front and sides), 1 mid-range lidar, and 4 short-range lidars. It contains 1000 sequences in total(about 198k samples), a training set of 798 sequences and a validation set of 202 sequences. OpenLane [6] is a comprehensive real-world 3D lane detection dataset built upon the Waymo Open dataset [53]. It includes a vast array of 200K frames and over 880K carefully annotated lanes, making it one of the largest of its kind to date. The dataset encompasses complex lane structures and features a variety of scene tags, such as weather and locations. Designed to emulate real-world scenarios, OpenLane provides a challenging benchmark for advanced lane detection algorithms. We integrate its 3D lane labels with the original Waymo Open sensor data for multi-task annotation." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 364, + 483, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 364, + 483, + 533 + ], + "spans": [ + { + "bbox": [ + 131, + 364, + 483, + 533 + ], + "type": "text", + "content": "Alignment of Datasets. To achieve both tasks simultaneously, we align both datasets in terms of coordinate systems and data splits. (1) Coordinate systems. For the vehicle frame, OpenLane [6] uses a vehicle coordinate system that corresponds to the \"y-front, x-right, z-up\" positioning of 3D-LaneNet [11]. We then align this system to coincide with the commonly utilized \"x-front, y-right, z-up\" axis for lidar-based 3D object detection. With respect to the sensor frame, we apply transformations to the camera intrinsics so that the sensor frames for all cameras conform to a uniform \"z-depth, x-right, y-down\" frame. The predictions are then transformed back into their respective task-based coordinate systems for evaluation. (2) Data splitting. The OpenLane [6] dataset provides a 1000 segment full version and a 300 segment smaller subset (" + }, + { + "bbox": [ + 131, + 364, + 483, + 533 + ], + "type": "inline_equation", + "content": "30\\%" + }, + { + "bbox": [ + 131, + 364, + 483, + 533 + ], + "type": "text", + "content": "). The partitioning of the larger version complements the original train/Validation split of the Waymo Open Dataset. The subset version partitioning is aligned with OpenLane. We perform ablation studies on the smaller subset." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 549, + 280, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 549, + 280, + 562 + ], + "spans": [ + { + "bbox": [ + 132, + 549, + 280, + 562 + ], + "type": "text", + "content": "5.2 Implementation Details" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 569, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 569, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 131, + 569, + 482, + 665 + ], + "type": "text", + "content": "Data Processing. We use the v1.4.2 of Waymo Open Dataset [53], and v1.2 for OpenLane [6]. We set the perception range to " + }, + { + "bbox": [ + 131, + 569, + 482, + 665 + ], + "type": "inline_equation", + "content": "[-84.88m, 84.88m]" + }, + { + "bbox": [ + 131, + 569, + 482, + 665 + ], + "type": "text", + "content": " for the X and Y axis, and " + }, + { + "bbox": [ + 131, + 569, + 482, + 665 + ], + "type": "inline_equation", + "content": "[-10m, 10m]" + }, + { + "bbox": [ + 131, + 569, + 482, + 665 + ], + "type": "text", + "content": " for the Z axis. 3D objects that do not fall within the perception range, are invisible to the cameras, or do not have LiDAR points, are filtered. For 3D lanes, visibility filtering is first applied after axis transformation, then 3D lanes outside the perception range are pruned and then resampled to a fixed size. Approximately " + }, + { + "bbox": [ + 131, + 569, + 482, + 665 + ], + "type": "inline_equation", + "content": "2.07\\%" + }, + { + "bbox": [ + 131, + 569, + 482, + 665 + ], + "type": "text", + "content": " of the frames without annotations after filtering are excluded from use." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 276, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 276, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 276, + 91, + 447, + 102 + ], + "type": "text", + "content": "RepVF: Unified Multi-task 3D Perception" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 101 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 101 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 101 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 154, + 157, + 465, + 284 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "type": "text", + "content": "Table 1: Comparison of recent models on 3D lane detection using the OpenLane [6] validation set. The down arrow indicates that lower metric values correspond to better performance, and vice versa. Red indicates the best result, and blue the second-best." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 154, + 157, + 465, + 284 + ], + "lines": [ + { + "bbox": [ + 154, + 157, + 465, + 284 + ], + "spans": [ + { + "bbox": [ + 154, + 157, + 465, + 284 + ], + "type": "table", + "html": "
MethodsF-Score ↑Category Accuracy ↑X error (m) ↓Z error (m) ↓
nearfarnearfar
3D-LaneNet [11] [ICCV19]44.1-0.4790.5720.3670.443
Gen-LaneNet [14] [ECCV20]32.3-0.5910.6840.4110.521
PersFormer [6] [ECCV22]50.592.30.4850.5530.3640.431
Cond-IPM [6]36.6-0.5631.0800.4210.892
CurveFormer [3] [ICRA23]50.5-0.3400.7720.2070.651
PETRv2-E [38] [ICCV23]51.9-0.4930.6430.3220.463
PETRv2-V-10 [38]57.8-0.4270.5820.2930.421
PETRv2-V-400 [38]61.2-0.4000.5730.2650.413
BEV-LaneDet [57] [CVPR23]58.4-0.3090.6590.2440.631
Anchor3DLane [18] [CVPR23]53.190.00.3000.3110.1030.139
SPG [69] [ICCV23]52.3-0.4680.5140.3710.418
RFTR (Ours)61.891.60.3410.4500.0730.107
", + "image_path": "c175dacc54a44eaf68865ae90af4b670a220dca536cc682e1b7ab238e6533791.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 154, + 332, + 465, + 398 + ], + "blocks": [ + { + "bbox": [ + 132, + 300, + 481, + 323 + ], + "lines": [ + { + "bbox": [ + 132, + 300, + 481, + 323 + ], + "spans": [ + { + "bbox": [ + 132, + 300, + 481, + 323 + ], + "type": "text", + "content": "Table 2: Comparison of recent works of 3D object detection on the Waymo Open Dataset [53] val set. mAPs are LEVEL_1 in Waymo metrics." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 154, + 332, + 465, + 398 + ], + "lines": [ + { + "bbox": [ + 154, + 332, + 465, + 398 + ], + "spans": [ + { + "bbox": [ + 154, + 332, + 465, + 398 + ], + "type": "table", + "html": "
MethodsOverallVehicleCyclistPedestrian
mAPmAPHmAPmAPHmAPmAPHmAPmAPH
DETR3D [62]10.19.69.29.115.815.55.34.1
DETR3D [62]8.68.018.118.02.41.65.34.4
PETR [37]20.919.731.130.819.517.612.110.6
RFTR (Ours)19.518.722.622.427.826.38.07.2
", + "image_path": "470b00319f5b3fe4016db9e10ea2529dd7b8e791cc8c8133195cf7040d40507c.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 426, + 482, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 426, + 482, + 668 + ], + "spans": [ + { + "bbox": [ + 130, + 426, + 482, + 668 + ], + "type": "text", + "content": "Model and Training. We used ResNet 50 [16] as the 2D image backbone for fair comparisons. Following PETR [37], the same upsampled and fused P4 feature with 1/16 input resolution is used as the 2D feature. A CPFPN [31] is used as the image neck. All 3D coordinates are normalized to " + }, + { + "bbox": [ + 130, + 426, + 482, + 668 + ], + "type": "inline_equation", + "content": "[0,1]" + }, + { + "bbox": [ + 130, + 426, + 482, + 668 + ], + "type": "text", + "content": ". The same weights are used for both the loss computation and the bipartite matching. We adopt progressive weighting for the loss of each decoder layer to improve the convergence speed. For 3D object detection, we compute matching costs based on the center displacement and class score weightings. Dimensions are refined post-matching using the ground truth headings to increase training robustness. We use a total of " + }, + { + "bbox": [ + 130, + 426, + 482, + 668 + ], + "type": "inline_equation", + "content": "n_s = 750" + }, + { + "bbox": [ + 130, + 426, + 482, + 668 + ], + "type": "text", + "content": " queries with " + }, + { + "bbox": [ + 130, + 426, + 482, + 668 + ], + "type": "inline_equation", + "content": "n = 15" + }, + { + "bbox": [ + 130, + 426, + 482, + 668 + ], + "type": "text", + "content": " sample locations each for the two tasks, 600 for 3D objects and 150 for 3D lanes. We trained our model using the AdamW [41] optimizer with a 0.01 weight decay. The learning rate started at " + }, + { + "bbox": [ + 130, + 426, + 482, + 668 + ], + "type": "inline_equation", + "content": "2.0 \\times 10^{-4}" + }, + { + "bbox": [ + 130, + 426, + 482, + 668 + ], + "type": "text", + "content": " and was decayed with a cosine annealing policy [40]. We adopted a multi-scale training strategy, with the shorter side chosen randomly from [640,900] and the longer side not exceeding 1600. We applied rotation augmentations between " + }, + { + "bbox": [ + 130, + 426, + 482, + 668 + ], + "type": "inline_equation", + "content": "[-\\pi /18,\\pi /18]" + }, + { + "bbox": [ + 130, + 426, + 482, + 668 + ], + "type": "text", + "content": ". We resized images of sizes " + }, + { + "bbox": [ + 130, + 426, + 482, + 668 + ], + "type": "inline_equation", + "content": "1920 \\times 1080" + }, + { + "bbox": [ + 130, + 426, + 482, + 668 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 426, + 482, + 668 + ], + "type": "inline_equation", + "content": "1920 \\times 1280" + }, + { + "bbox": [ + 130, + 426, + 482, + 668 + ], + "type": "text", + "content": " from different views to " + }, + { + "bbox": [ + 130, + 426, + 482, + 668 + ], + "type": "inline_equation", + "content": "1600 \\times 900" + }, + { + "bbox": [ + 130, + 426, + 482, + 668 + ], + "type": "text", + "content": " before augmenting. The final resolution for network inputs was " + }, + { + "bbox": [ + 130, + 426, + 482, + 668 + ], + "type": "inline_equation", + "content": "1408 \\times 512" + }, + { + "bbox": [ + 130, + 426, + 482, + 668 + ], + "type": "text", + "content": ". All experiments on the full dataset are trained for 24 epochs on 4 GeForce RTX 4090 GPUs with a batch size of 8. No test time augmentation methods are used during inference." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 211, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 211, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 211, + 100 + ], + "type": "text", + "content": "C. Li et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 135, + 115, + 479, + 248 + ], + "blocks": [ + { + "bbox": [ + 135, + 115, + 479, + 248 + ], + "lines": [ + { + "bbox": [ + 135, + 115, + 479, + 248 + ], + "spans": [ + { + "bbox": [ + 135, + 115, + 479, + 248 + ], + "type": "image", + "image_path": "06f1d69f3e44123ec45862776d7fee29e3761a7f5836c7cf9a9f1598cc45f104.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 141, + 257, + 472, + 269 + ], + "lines": [ + { + "bbox": [ + 141, + 257, + 472, + 269 + ], + "spans": [ + { + "bbox": [ + 141, + 257, + 472, + 269 + ], + "type": "text", + "content": "Fig. 4: Visualization of the RFTR model's performance and the learned " + }, + { + "bbox": [ + 141, + 257, + 472, + 269 + ], + "type": "inline_equation", + "content": "RepVF" + }, + { + "bbox": [ + 141, + 257, + 472, + 269 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 292, + 282, + 304 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 292, + 282, + 304 + ], + "spans": [ + { + "bbox": [ + 132, + 292, + 282, + 304 + ], + "type": "text", + "content": "5.3 Comparison with SOTA" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 311, + 482, + 431 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 311, + 482, + 431 + ], + "spans": [ + { + "bbox": [ + 130, + 311, + 482, + 431 + ], + "type": "text", + "content": "3D Lane Detection. As shown in Tab. 1, our RFTR model showcases significant achievements in 3D lane detection on the OpenLane validation set, demonstrating its competitive edge with a high F-score of 61.8 and category accuracy of " + }, + { + "bbox": [ + 130, + 311, + 482, + 431 + ], + "type": "inline_equation", + "content": "91.6\\%" + }, + { + "bbox": [ + 130, + 311, + 482, + 431 + ], + "type": "text", + "content": ". It matches or exceeds state-of-the-art models such as PETRv2-V-400 [38] (more than " + }, + { + "bbox": [ + 130, + 311, + 482, + 431 + ], + "type": "inline_equation", + "content": "26\\times" + }, + { + "bbox": [ + 130, + 311, + 482, + 431 + ], + "type": "text", + "content": " denser representation than ours) when assessed on the OpenLane [6] validation set. Notably, while our model has a slightly higher X error in the \"far\" class than [18], it manages to lead with a higher F-score and near accurate Z errors. Moreover, our method yields notable performance improvements over the Persformer [6] baseline, reflecting the advantageous adaptation of our unified RepVF representation and the RFTR architecture." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 448, + 482, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 448, + 482, + 544 + ], + "spans": [ + { + "bbox": [ + 130, + 448, + 482, + 544 + ], + "type": "text", + "content": "3D Object Detection. The performance of our RFTR model in 3D object detection, as demonstrated in Tab. 2, shows comparable promising results on the Waymo Open Dataset [53] validation set. While our approach shows a commanding lead in detecting vehicles and cyclists with a solid L1 mAP and mAPH, it shows challenges in detecting smaller entities such as pedestrians. It should be noted, however, that these preliminary results indicate the great potential of using our unified perceptual model for complex tasks, which will require further fine-tuning for object classes with less prominent features." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 562, + 246, + 573 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 562, + 246, + 573 + ], + "spans": [ + { + "bbox": [ + 132, + 562, + 246, + 573 + ], + "type": "text", + "content": "5.4 Ablation Studies" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 581, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 581, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 581, + 482, + 665 + ], + "type": "text", + "content": "Head design. We begin by finding out how head designs and representations contribute to the performance, in Tab. 3. The " + }, + { + "bbox": [ + 130, + 581, + 482, + 665 + ], + "type": "inline_equation", + "content": "RepVF" + }, + { + "bbox": [ + 130, + 581, + 482, + 665 + ], + "type": "text", + "content": " representation with the single-head design, i.e. RFTR, performs the best in general with 37.4M parameters. Unified representation brings performance improvements in multi-task learning and reduces task competition. In contrast, task-specific multi-head design produces inferior and unstably biased performance due to task competition, even with more parameters." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 276, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 276, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 276, + 91, + 447, + 102 + ], + "type": "text", + "content": "RepVF: Unified Multi-task 3D Perception" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 143, + 179, + 312, + 254 + ], + "blocks": [ + { + "bbox": [ + 130, + 114, + 321, + 170 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 321, + 170 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 321, + 170 + ], + "type": "text", + "content": "Table 3: Multi-task head design and representation ablation. Tasks are trained jointly on " + }, + { + "bbox": [ + 130, + 114, + 321, + 170 + ], + "type": "inline_equation", + "content": "30\\%" + }, + { + "bbox": [ + 130, + 114, + 321, + 170 + ], + "type": "text", + "content": " data of Waymo Open Dataset (L1 mAP) and OpenLane (F-Score). TS denotes task-specific representation. " + }, + { + "bbox": [ + 130, + 114, + 321, + 170 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 130, + 114, + 321, + 170 + ], + "type": "text", + "content": ": lightweight models." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 143, + 179, + 312, + 254 + ], + "lines": [ + { + "bbox": [ + 143, + 179, + 312, + 254 + ], + "spans": [ + { + "bbox": [ + 143, + 179, + 312, + 254 + ], + "type": "table", + "html": "
RepVFTSHeadParams↓F-Score↑Vehicle↑
✓Multi†37.1M49.620.1
✓Multi46.7M58.717.8
✓Multi†39.7M63.418.2
✓Multi49.2M66.422.1
✓Single37.4M66.525.3
", + "image_path": "6c7d99a2902f34c4f5f7a219f2fd8fe02fa3a6dcdbef81ba2d8fc22e549538e8.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 332, + 183, + 476, + 251 + ], + "blocks": [ + { + "bbox": [ + 324, + 118, + 482, + 174 + ], + "lines": [ + { + "bbox": [ + 324, + 118, + 482, + 174 + ], + "spans": [ + { + "bbox": [ + 324, + 118, + 482, + 174 + ], + "type": "text", + "content": "Table 4: Results with different representations on 3D lane. 'Acc' for category accuracy and 'Error' for X error near. MT means simultaneous multitask 3D lane and 3D object detection." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 332, + 183, + 476, + 251 + ], + "lines": [ + { + "bbox": [ + 332, + 183, + 476, + 251 + ], + "spans": [ + { + "bbox": [ + 332, + 183, + 476, + 251 + ], + "type": "table", + "html": "
RepresentationF-Score↑Acc↑Error↓
Anchor53.486.00.476
PS58.188.00.545
RepVF66.091.40.439
RepVF + MT66.591.70.420
", + "image_path": "e31ad472859c4307f83241df8d3cd7020a364f3a42a9a43ece5d08cda42fdcf7.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 277, + 482, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 277, + 482, + 350 + ], + "spans": [ + { + "bbox": [ + 130, + 277, + 482, + 350 + ], + "type": "text", + "content": "Representation. Moreover, we have evaluated the effectiveness of different representations for 3D lane detection in Tab. 4, including traditional anchor representations [6,37], simple randomly sampled point set representation (PS), and our proposed representative vector fields (RepVF). This study reveals that RepVF significantly outperforms traditional anchor and PS methods, and incorporating multi-task learning marginally improves these metrics." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 372, + 258, + 384 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 372, + 258, + 384 + ], + "spans": [ + { + "bbox": [ + 131, + 372, + 258, + 384 + ], + "type": "text", + "content": "5.5 Qualitative Results" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 396, + 482, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 396, + 482, + 517 + ], + "spans": [ + { + "bbox": [ + 130, + 396, + 482, + 517 + ], + "type": "text", + "content": "Figure 4 demonstrates the capability of RFTR to simultaneously achieve 3D lane detection and 3D object detection on the extended [24] Waymo open dataset [53]. The figure clearly illustrates that RFTR can accurately predict RepVF and perform both tasks at once. In terms of 3D lane detection, RFTR is capable of predicting the shape of lane lines in areas obscured by vehicles, demonstrating its robustness against occlusion. Regarding 3D object detection, the unified RepVF representation accurately captures the 3D bounding boxes. This dual achievement highlights RFTR's effectiveness in handling complex driving scenarios, enhancing both navigation and safety features in autonomous driving systems." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 539, + 220, + 552 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 539, + 220, + 552 + ], + "spans": [ + { + "bbox": [ + 131, + 539, + 220, + 552 + ], + "type": "text", + "content": "6 Conclusion" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": "In this paper, we propose " + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "inline_equation", + "content": "RepVF" + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": ", a novel unified vector fields representation that characterizes the geometric and semantic structure of different 3D perception targets in the scene. Built upon " + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "inline_equation", + "content": "RepVF" + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": ", our proposed RFTR exploits the intrinsic connections between tasks, mitigates multi-task gradient imbalance and feature competition, and improves model convergence and multi-task expression. Our work not only contributes a novel single-head perspective to the multi-task learning paradigm in autonomous driving but also sets a promising direction for future research in efficient and effective 3D perception task handling." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 212, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 212, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 212, + 101 + ], + "type": "text", + "content": "C. Li et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 198, + 126 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 198, + 126 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 198, + 126 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 134, + 139, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 138, + 139, + 399, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 139, + 399, + 150 + ], + "spans": [ + { + "bbox": [ + 138, + 139, + 399, + 150 + ], + "type": "text", + "content": "1. Mobileye at CES 2024. https://www-mobileye.com/ces-2024/" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 151, + 406, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 151, + 406, + 161 + ], + "spans": [ + { + "bbox": [ + 138, + 151, + 406, + 161 + ], + "type": "text", + "content": "2. NVIDIA DRIVE Solutions. https://developer.nvidia.com/drive" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 162, + 481, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 162, + 481, + 205 + ], + "spans": [ + { + "bbox": [ + 138, + 162, + 481, + 205 + ], + "type": "text", + "content": "3. Bai, Y., Chen, Z., Fu, Z., Peng, L., Liang, P., Cheng, E.: CurveFormer: 3D Lane Detection by Curve Propagation with Curve Queries and Attention. In: 2023 IEEE International Conference on Robotics and Automation (ICRA). pp. 7062-7068 (May 2023). https://doi.org/10.1109/ICRA48891.2023.10161160" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 206, + 481, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 206, + 481, + 258 + ], + "spans": [ + { + "bbox": [ + 138, + 206, + 481, + 258 + ], + "type": "text", + "content": "4. Caesar, H., Bankiti, V., Lang, A.H., Vora, S., Liong, V.E., Xu, Q., Krishnan, A., Pan, Y., Baldan, G., Beijbom, O.: nuScenes: A Multimodal Dataset for Autonomous Driving. In: 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 11618-11628. IEEE, Seattle, WA, USA (Jun 2020). https://doi.org/10.1109/CVPR42600.2020.01164" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 260, + 481, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 260, + 481, + 315 + ], + "spans": [ + { + "bbox": [ + 138, + 260, + 481, + 315 + ], + "type": "text", + "content": "5. Carion, N., Massa, F., Synnaeve, G., Usunier, N., Kirillov, A., Zagoruyko, S.: End-to-End Object Detection with Transformers. In: Vedaldi, A., Bischof, H., Brox, T., Frahm, J.M. (eds.) Computer Vision - ECCV 2020. pp. 213-229. Lecture Notes in Computer Science, Springer International Publishing, Cham (2020). https://doi.org/10.1007/978-3-030-58452-8_13" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 316, + 481, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 316, + 481, + 380 + ], + "spans": [ + { + "bbox": [ + 138, + 316, + 481, + 380 + ], + "type": "text", + "content": "6. Chen, L., Sima, C., Li, Y., Zheng, Z., Xu, J., Geng, X., Li, H., He, C., Shi, J., Qiao, Y., Yan, J.: PersFormer: 3D Lane Detection via Perspective Transformer and the OpenLane Benchmark. In: Avidan, S., Brostow, G., Cisse, M., Farinella, G.M., Hassner, T. (eds.) Computer Vision - ECCV 2022, vol. 13698, pp. 550-567. Springer Nature Switzerland, Cham (2022). https://doi.org/10.1007/978-3-031-19839-7_32" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 380, + 481, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 380, + 481, + 413 + ], + "spans": [ + { + "bbox": [ + 138, + 380, + 481, + 413 + ], + "type": "text", + "content": "7. Chen, X., Kundu, K., Zhang, Z., Ma, H., Fidler, S., Urtasun, R.: Monocular 3D Object Detection for Autonomous Driving. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 2147-2156 (2016)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 414, + 481, + 457 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 414, + 481, + 457 + ], + "spans": [ + { + "bbox": [ + 138, + 414, + 481, + 457 + ], + "type": "text", + "content": "8. Chen, Z., Badrinarayanan, V., Lee, C.Y., Rabinovich, A.: GradNorm: Gradient Normalization for Adaptive Loss Balancing in Deep Multitask Networks. In: Proceedings of the 35th International Conference on Machine Learning. pp. 794-803. PMLR (Jul 2018)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 458, + 481, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 458, + 481, + 479 + ], + "spans": [ + { + "bbox": [ + 138, + 458, + 481, + 479 + ], + "type": "text", + "content": "9. Cheng, W., Yin, J., Li, W., Yang, R., Shen, J.: Language-Guided 3D Object Detection in Point Cloud for Autonomous Driving (May 2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 479, + 481, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 479, + 481, + 510 + ], + "spans": [ + { + "bbox": [ + 138, + 479, + 481, + 510 + ], + "type": "text", + "content": "0. Feng, Z., Guo, S., Tan, X., Xu, K., Wang, M., Ma, L.: Rethinking Efficient Lane Detection via Curve Modeling (May 2023). https://doi.org/10.48550/arXiv.2203.02431" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 512, + 481, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 512, + 481, + 555 + ], + "spans": [ + { + "bbox": [ + 138, + 512, + 481, + 555 + ], + "type": "text", + "content": "1. Garnett, N., Cohen, R., Pe'er, T., Lahav, R., Levi, D.: 3D-LaneNet: End-to-End 3D Multiple Lane Detection. In: 2019 IEEE/CVF International Conference on Computer Vision (ICCV). pp. 2921–2930. IEEE, Seoul, Korea (South) (Oct 2019). https://doi.org/10.1109/ICCV.2019.00301" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 134, + 555, + 481, + 598 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 555, + 481, + 598 + ], + "spans": [ + { + "bbox": [ + 134, + 555, + 481, + 598 + ], + "type": "text", + "content": "12. Geiger, A., Lenz, P., Urtasun, R.: Are we ready for autonomous driving? The KITTI vision benchmark suite. In: 2012 IEEE Conference on Computer Vision and Pattern Recognition. pp. 3354-3361 (Jun 2012). https://doi.org/10.1109/CVPR.2012.6248074" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 134, + 599, + 481, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 599, + 481, + 632 + ], + "spans": [ + { + "bbox": [ + 134, + 599, + 481, + 632 + ], + "type": "text", + "content": "13. Guo, P., Lee, C.Y., Ulbricht, D.: Learning to Branch for Multi-Task Learning. In: Proceedings of the 37th International Conference on Machine Learning. pp. 3854-3863. PMLR (Nov 2020)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 134, + 632, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 632, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 134, + 632, + 481, + 665 + ], + "type": "text", + "content": "14. Guo, Y., Chen, G., Zhao, P., Zhang, W., Miao, J., Wang, J., Choe, T.E.: GenLaneNet: A Generalized and Scalable Approach for 3D Lane Detection. In: Vedaldi, A., Bischof, H., Brox, T., Frahm, J.M. (eds.) Computer Vision - ECCV 2020. pp." + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 276, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 276, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 276, + 91, + 447, + 102 + ], + "type": "text", + "content": "RepVF: Unified Multi-task 3D Perception" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 147, + 116, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 147, + 116, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 147, + 116, + 480, + 138 + ], + "type": "text", + "content": "666-681. Lecture Notes in Computer Science, Springer International Publishing, Cham (2020). https://doi.org/10.1007/978-3-030-58589-1_40" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 138, + 480, + 159 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 138, + 480, + 159 + ], + "spans": [ + { + "bbox": [ + 133, + 138, + 480, + 159 + ], + "type": "text", + "content": "15. Han, W., Shen, J.: Decoupling the Curve Modeling and Pavement Regression for Lane Detection (Sep 2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 160, + 480, + 191 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 160, + 480, + 191 + ], + "spans": [ + { + "bbox": [ + 133, + 160, + 480, + 191 + ], + "type": "text", + "content": "16. He, K., Zhang, X., Ren, S., Sun, J.: Deep Residual Learning for Image Recognition. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 770-778 (2016)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 192, + 480, + 213 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 192, + 480, + 213 + ], + "spans": [ + { + "bbox": [ + 133, + 192, + 480, + 213 + ], + "type": "text", + "content": "17. Huang, J., Huang, G.: BEVDet4D: Exploit Temporal Cues in Multi-camera 3D Object Detection (Jun 2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 133, + 213, + 480, + 255 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 213, + 480, + 255 + ], + "spans": [ + { + "bbox": [ + 133, + 213, + 480, + 255 + ], + "type": "text", + "content": "18. Huang, S., Shen, Z., Huang, Z., Ding, Z.h., Dai, J., Han, J., Wang, N., Liu, S.: Anchor3DLane: Learning To Regress 3D Anchors for Monocular 3D Lane Detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 17451-17460 (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 133, + 256, + 480, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 256, + 480, + 308 + ], + "spans": [ + { + "bbox": [ + 133, + 256, + 480, + 308 + ], + "type": "text", + "content": "19. Huang, Z., Lin, S., Liu, G., Luo, M., Ye, C., Xu, H., Chang, X., Liang, X.: FULLER: Unified Multi-modality Multi-task 3D Perception via Multi-level Gradient Calibration. In: 2023 IEEE/CVF International Conference on Computer Vision (ICCV). pp. 3479-3488. IEEE, Paris, France (Oct 2023). https://doi.org/10.1109/ICCV51070.2023.00324" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 310, + 480, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 310, + 480, + 342 + ], + "spans": [ + { + "bbox": [ + 132, + 310, + 480, + 342 + ], + "type": "text", + "content": "20. Kehl, W., Manhardt, F., Tombari, F., Ilic, S., Navab, N.: SSD-6D: Making RGB-Based 3D Detection and 6D Pose Estimation Great Again. In: Proceedings of the IEEE International Conference on Computer Vision. pp. 1521–1529 (2017)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 343, + 480, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 343, + 480, + 374 + ], + "spans": [ + { + "bbox": [ + 132, + 343, + 480, + 374 + ], + "type": "text", + "content": "21. Kendall, A., Gal, Y., Cipolla, R.: Multi-Task Learning Using Uncertainty to Weigh Losses for Scene Geometry and Semantics. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 7482-7491 (2018)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 375, + 480, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 375, + 480, + 407 + ], + "spans": [ + { + "bbox": [ + 132, + 375, + 480, + 407 + ], + "type": "text", + "content": "22. Ku, J., Pon, A.D., Waslander, S.L.: Monocular 3D Object Detection Leveraging Accurate Proposals and Shape Reconstruction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11867-11876 (2019)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 407, + 480, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 407, + 480, + 437 + ], + "spans": [ + { + "bbox": [ + 132, + 407, + 480, + 437 + ], + "type": "text", + "content": "23. Kuhn, H.W.: The Hungarian method for the assignment problem. Naval Research Logistics Quarterly 2(1-2), 83-97 (1955). https://doi.org/10.1002/nav.3800020109" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 438, + 480, + 469 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 438, + 480, + 469 + ], + "spans": [ + { + "bbox": [ + 132, + 438, + 480, + 469 + ], + "type": "text", + "content": "24. Li, L.H., Yatskar, M., Yin, D., Hsieh, C.J., Chang, K.W.: VisualBERT: A Simple and Performant Baseline for Vision and Language (Aug 2019). https://doi.org/10.48550/arXiv.1908.03557" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 471, + 480, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 471, + 480, + 514 + ], + "spans": [ + { + "bbox": [ + 132, + 471, + 480, + 514 + ], + "type": "text", + "content": "25. Li, X., Yin, J., Li, W., Xu, C., Yang, R., Shen, J.: Di-v2x: Learning domain-invariant representation for vehicle-infrastructure collaborative 3d object detection. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 38, pp. 3208-3215 (2024)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 514, + 480, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 514, + 480, + 567 + ], + "spans": [ + { + "bbox": [ + 132, + 514, + 480, + 567 + ], + "type": "text", + "content": "26. Li, X., Yin, J., Shi, B., Li, Y., Yang, R., Shen, J.: Lwsis: Lidar-guided weakly supervised instance segmentation for autonomous driving. Proceedings of the AAAI Conference on Artificial Intelligence 37(2), 1433-1441 (Jun 2023). https://doi.org/10.1609/aaii.v37i2.25228, https://ojs.aaii.org/index.php/AAAI/article/view/25228" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 567, + 480, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 567, + 480, + 599 + ], + "spans": [ + { + "bbox": [ + 132, + 567, + 480, + 599 + ], + "type": "text", + "content": "27. Li, Y., Chen, Y., Qi, X., Li, Z., Sun, J., Jia, J.: Unifying Voxel-based Representation with Transformer for 3D Object Detection. Advances in Neural Information Processing Systems 35, 18442-18455 (Dec 2022)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 600, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 600, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 600, + 480, + 665 + ], + "type": "text", + "content": "28. Li, Z., Wang, W., Li, H., Xie, E., Sima, C., Lu, T., Qiao, Y., Dai, J.: BEVFormer: Learning Bird's-Eye-View Representation from Multi-camera Images via Spatiotemporal Transformers. In: Avidan, S., Brostow, G., Cisse, M., Farinella, G.M., Hassner, T. (eds.) Computer Vision - ECCV 2022. pp. 1-18. Lecture Notes in Computer Science, Springer Nature Switzerland, Cham (2022). https://doi.org/10.1007/978-3-031-20077-9_1" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 212, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 212, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 212, + 100 + ], + "type": "text", + "content": "C. Li et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 665 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 150 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 150 + ], + "type": "text", + "content": "29. Liang, M., Yang, B., Chen, Y., Hu, R., Urtasun, R.: Multi-Task Multi-Sensor Fusion for 3D Object Detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7345-7353 (2019)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 150, + 482, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 150, + 482, + 183 + ], + "spans": [ + { + "bbox": [ + 130, + 150, + 482, + 183 + ], + "type": "text", + "content": "30. Liang, T., Xie, H., Yu, K., Xia, Z., Lin, Z., Wang, Y., Tang, T., Wang, B., Tang, Z.: BEVFusion: A Simple and Robust LiDAR-Camera Fusion Framework. Advances in Neural Information Processing Systems 35, 10421-10434 (Dec 2022)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 183, + 482, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 183, + 482, + 217 + ], + "spans": [ + { + "bbox": [ + 130, + 183, + 482, + 217 + ], + "type": "text", + "content": "31. Lin, T.Y., Dollar, P., Girshick, R., He, K., Hariharan, B., Belongie, S.: Feature Pyramid Networks for Object Detection. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 2117-2125 (2017)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 218, + 482, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 218, + 482, + 251 + ], + "spans": [ + { + "bbox": [ + 130, + 218, + 482, + 251 + ], + "type": "text", + "content": "32. Lin, T.Y., Goyal, P., Girshick, R., He, K., Dollar, P.: Focal Loss for Dense Object Detection. In: Proceedings of the IEEE International Conference on Computer Vision. pp. 2980-2988 (2017)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 251, + 482, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 251, + 482, + 285 + ], + "spans": [ + { + "bbox": [ + 130, + 251, + 482, + 285 + ], + "type": "text", + "content": "33. Liu, L., Li, Y., Kuang, Z., Xue, J.H., Chen, Y., Yang, W., Liao, Q., Zhang, W.: Towards Impartial Multi-task Learning. In: International Conference on Learning Representations (Oct 2020)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 285, + 482, + 329 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 285, + 482, + 329 + ], + "spans": [ + { + "bbox": [ + 130, + 285, + 482, + 329 + ], + "type": "text", + "content": "34. Liu, R., Chen, D., Liu, T., Xiong, Z., Yuan, Z.: Learning to Predict 3D Lane Shape and Camera Pose from a Single Image via Geometry Constraints. Proceedings of the AAAI Conference on Artificial Intelligence 36(2), 1765-1772 (Jun 2022). https://doi.org/10.1609/aaai.v36i2.20069" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 329, + 482, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 329, + 482, + 352 + ], + "spans": [ + { + "bbox": [ + 130, + 329, + 482, + 352 + ], + "type": "text", + "content": "35. Liu, S., Johns, E., Davison, A.J.: End-to-End Multi-Task Learning with Attention (Apr 2019). https://doi.org/10.48550/arXiv.1803.10704" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 352, + 482, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 352, + 482, + 374 + ], + "spans": [ + { + "bbox": [ + 130, + 352, + 482, + 374 + ], + "type": "text", + "content": "36. Liu, S., Li, F., Zhang, H., Yang, X., Qi, X., Su, H., Zhu, J., Zhang, L.: DAB-DETR: DYNAMIC ANCHOR BOXES ARE BETTER QUERIES FOR DETR (2022)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 374, + 482, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 374, + 482, + 430 + ], + "spans": [ + { + "bbox": [ + 130, + 374, + 482, + 430 + ], + "type": "text", + "content": "37. Liu, Y., Wang, T., Zhang, X., Sun, J.: PETR: Position Embedding Transformation for Multi-view 3D Object Detection. In: Avidan, S., Brostow, G., Cisse, M., Farinella, G.M., Hassner, T. (eds.) Computer Vision - ECCV 2022, vol. 13687, pp. 531-548. Springer Nature Switzerland, Cham (2022). https://doi.org/10.1007/978-3-031-19812-0_31" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 430, + 482, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 430, + 482, + 453 + ], + "spans": [ + { + "bbox": [ + 130, + 430, + 482, + 453 + ], + "type": "text", + "content": "38. Liu, Y., Yan, J., Jia, F., Li, S., Gao, A., Wang, T., Zhang, X.: PETRv2: A Unified Framework for 3D Perception from Multi-Camera Images" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 453, + 482, + 497 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 453, + 482, + 497 + ], + "spans": [ + { + "bbox": [ + 130, + 453, + 482, + 497 + ], + "type": "text", + "content": "39. Liu, Z., Tang, H., Amini, A., Yang, X., Mao, H., Rus, D.L., Han, S.: BEVFusion: Multi-Task Multi-Sensor Fusion with Unified Bird's-Eye View Representation. In: 2023 IEEE International Conference on Robotics and Automation (ICRA). pp. 2774-2781 (May 2023). https://doi.org/10.1109/ICRA48891.2023.10160968" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 497, + 482, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 497, + 482, + 520 + ], + "spans": [ + { + "bbox": [ + 130, + 497, + 482, + 520 + ], + "type": "text", + "content": "40. Loshchilov, I., Hutter, F.: SGDR: Stochastic Gradient Descent with Warm Restarts. In: International Conference on Learning Representations (Nov 2016)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 520, + 482, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 520, + 482, + 542 + ], + "spans": [ + { + "bbox": [ + 130, + 520, + 482, + 542 + ], + "type": "text", + "content": "41. Loshchilov, I., Hutter, F.: DECOUPLED WEIGHT DECAY REGULARIZATION (2019)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 543, + 482, + 576 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 543, + 482, + 576 + ], + "spans": [ + { + "bbox": [ + 130, + 543, + 482, + 576 + ], + "type": "text", + "content": "42. Luo, Y., Zheng, C., Yan, X., Kun, T., Zheng, C., Cui, S., Li, Z.: LATR: 3D Lane Detection from Monocular Images with Transformer. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 7941-7952 (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 576, + 482, + 631 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 576, + 482, + 631 + ], + "spans": [ + { + "bbox": [ + 130, + 576, + 482, + 631 + ], + "type": "text", + "content": "43. Ma, Z., Wang, L., Zhang, H., Lu, W., Yin, J.: RPT: Learning Point Set Representation for Siamese Visual Tracking. In: Bartoli, A., Fusiello, A. (eds.) Computer Vision - ECCV 2020 Workshops. pp. 653-665. Lecture Notes in Computer Science, Springer International Publishing, Cham (2020). https://doi.org/10.1007/978-3-030-68238-5_43" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 130, + 632, + 482, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 632, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 632, + 482, + 665 + ], + "type": "text", + "content": "44. Misra, I., Shrivastava, A., Gupta, A., Hebert, M.: Cross-Stitch Networks for Multi-Task Learning. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 3994–4003 (2016)" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 276, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 276, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 276, + 91, + 447, + 102 + ], + "type": "text", + "content": "RepVF: Unified Multi-task 3D Perception" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 117, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 133, + 117, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 117, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 117, + 480, + 149 + ], + "type": "text", + "content": "45. Mousavian, A., Anguelov, D., Flynn, J., Kosecka, J.: 3D Bounding Box Estimation Using Deep Learning and Geometry. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 7074-7082 (2017)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 150, + 480, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 150, + 480, + 171 + ], + "spans": [ + { + "bbox": [ + 133, + 150, + 480, + 171 + ], + "type": "text", + "content": "46. Navon, A., Shamsian, A., Achituve, I., Maron, H., Kawaguchi, K., Chechik, G., Fetaya, E.: Multi-Task Learning as a Bargaining Game (Feb 2022)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 171, + 480, + 201 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 171, + 480, + 201 + ], + "spans": [ + { + "bbox": [ + 133, + 171, + 480, + 201 + ], + "type": "text", + "content": "47. Neven, D., De Brabandere, B., Georgoulis, S., Proesmans, M., Van Gool, L.: Fast Scene Understanding for Autonomous Driving (Aug 2017). https://doi.org/10.48550/arXiv.1708.02550" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 134, + 203, + 480, + 256 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 203, + 480, + 256 + ], + "spans": [ + { + "bbox": [ + 134, + 203, + 480, + 256 + ], + "type": "text", + "content": "48. Philion, J., Fidler, S.: Lift, Splat, Shoot: Encoding Images from Arbitrary Camera Rigs by Implicitly Unprojecting to 3D. In: Vedaldi, A., Bischof, H., Brox, T., Frahm, J.M. (eds.) Computer Vision - ECCV 2020. pp. 194-210. Lecture Notes in Computer Science, Springer International Publishing, Cham (2020). https://doi.org/10.1007/978-3-030-58568-6_12" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 133, + 257, + 480, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 257, + 480, + 300 + ], + "spans": [ + { + "bbox": [ + 133, + 257, + 480, + 300 + ], + "type": "text", + "content": "49. Reading, C., Harakeh, A., Chae, J., Waslander, S.L.: Categorical Depth Distribution Network for Monocular 3D Object Detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8555-8564 (2021)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 300, + 480, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 300, + 480, + 330 + ], + "spans": [ + { + "bbox": [ + 132, + 300, + 480, + 330 + ], + "type": "text", + "content": "50. Roddick, T., Kendall, A., Cipolla, R.: Orthographic Feature Transform for Monocular 3D Object Detection (Nov 2018). https://doi.org/10.48550/arXiv.1811.08188" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 332, + 480, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 332, + 480, + 364 + ], + "spans": [ + { + "bbox": [ + 132, + 332, + 480, + 364 + ], + "type": "text", + "content": "51. Ruder, S., Bingel, J., Augenstein, I., Søgaard, A.: Latent Multi-Task Architecture Learning. Proceedings of the AAAI Conference on Artificial Intelligence 33(01), 4822-4829 (Jul 2019). https://doi.org/10.1609/aaai.v33i01.33014822" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 365, + 480, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 365, + 480, + 396 + ], + "spans": [ + { + "bbox": [ + 132, + 365, + 480, + 396 + ], + "type": "text", + "content": "52. Sener, O., Koltun, V.: Multi-Task Learning as Multi-Objective Optimization. In: Advances in Neural Information Processing Systems. vol. 31. Curran Associates, Inc. (2018)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 133, + 396, + 480, + 450 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 396, + 480, + 450 + ], + "spans": [ + { + "bbox": [ + 133, + 396, + 480, + 450 + ], + "type": "text", + "content": "53. Sun, P., Kretzschmar, H., Dotiwalla, X., Chouard, A., Patnaik, V., Tsui, P., Guo, J., Zhou, Y., Chai, Y., Caine, B., Vasudevan, V., Han, W., Ngiam, J., Zhao, H., Timofeev, A., Ettinger, S., Krivokon, M., Gao, A., Joshi, A., Zhang, Y., Shlens, J., Chen, Z., Anguelov, D.: Scalability in Perception for Autonomous Driving: Waymo Open Dataset" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 133, + 450, + 480, + 492 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 450, + 480, + 492 + ], + "spans": [ + { + "bbox": [ + 133, + 450, + 480, + 492 + ], + "type": "text", + "content": "54. Teichmann, M., Weber, M., Zöllner, M., Cipolla, R., Urtasun, R.: MultiNet: Realtime Joint Semantic Reasoning for Autonomous Driving. In: 2018 IEEE Intelligent Vehicles Symposium (IV). pp. 1013-1020 (Jun 2018). https://doi.org/10.1109/IVS.2018.8500504" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 133, + 493, + 480, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 493, + 480, + 536 + ], + "spans": [ + { + "bbox": [ + 133, + 493, + 480, + 536 + ], + "type": "text", + "content": "55. Vandenhende, S., Georgoulis, S., Van Gansbeke, W., Proesmans, M., Dai, D., Van Gool, L.: Multi-Task Learning for Dense Prediction Tasks: A Survey. IEEE Transactions on Pattern Analysis and Machine Intelligence 44(7), 3614-3633 (Jul 2022). https://doi.org/10.1109/TPAMI.2021.3054719" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 133, + 537, + 480, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 537, + 480, + 568 + ], + "spans": [ + { + "bbox": [ + 133, + 537, + 480, + 568 + ], + "type": "text", + "content": "56. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is All you Need. In: Advances in Neural Information Processing Systems. vol. 30. Curran Associates, Inc. (2017)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 133, + 569, + 480, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 569, + 480, + 611 + ], + "spans": [ + { + "bbox": [ + 133, + 569, + 480, + 611 + ], + "type": "text", + "content": "57. Wang, R., Qin, J., Li, K., Li, Y., Cao, D., Xu, J.: BEV-LaneDet: An Efficient 3D Lane Detection Based on Virtual Camera via Key-Points. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1002-1011 (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 133, + 612, + 480, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 612, + 480, + 633 + ], + "spans": [ + { + "bbox": [ + 133, + 612, + 480, + 633 + ], + "type": "text", + "content": "58. Wang, S., Liu, Y., Wang, T., Li, Y., Zhang, X.: Exploring Object-Centric Temporal Modeling for Efficient Multi-View 3D Object Detection (Jun 2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 133, + 633, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 633, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 633, + 480, + 665 + ], + "type": "text", + "content": "59. Wang, T., Zhu, X., Pang, J., Lin, D.: FCOS3D: Fully Convolutional One-Stage Monocular 3D Object Detection. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 913–922 (2021)" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 212, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 212, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 212, + 100 + ], + "type": "text", + "content": "C. Li et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 664 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 171 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 171 + ], + "type": "text", + "content": "60. Wang, Y., Yin, J., Li, W., Frossard, P., Yang, R., Shen, J.: Ssda3d: Semi-supervised domain adaptation for 3d object detection from point cloud. Proceedings of the AAAI Conference on Artificial Intelligence 37(3), 2707-2715 (Jun 2023). https://doi.org/10.1609/aaai.v37i3.25370, https://ojs(aaai.org/index.php/AAAI/article/view/25370" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 172, + 481, + 214 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 172, + 481, + 214 + ], + "spans": [ + { + "bbox": [ + 132, + 172, + 481, + 214 + ], + "type": "text", + "content": "61. Wang, Y., Zhang, X., Yang, T., Sun, J.: Anchor DETR: Query Design for Transformer-Based Detector. Proceedings of the AAAI Conference on Artificial Intelligence 36(3), 2567-2575 (Jun 2022). https://doi.org/10.1609/aaai.v36i3.20158" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 215, + 481, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 215, + 481, + 248 + ], + "spans": [ + { + "bbox": [ + 132, + 215, + 481, + 248 + ], + "type": "text", + "content": "62. Wang, Y., Guizilini, V.C., Zhang, T., Wang, Y., Zhao, H., Solomon, J.: DETR3D: 3D Object Detection from Multi-view Images via 3D-to-2D Queries. In: Proceedings of the 5th Conference on Robot Learning. pp. 180-191. PMLR (Jan 2022)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 249, + 481, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 249, + 481, + 281 + ], + "spans": [ + { + "bbox": [ + 132, + 249, + 481, + 281 + ], + "type": "text", + "content": "63. Wu, D., Chang, J., Jia, F., Liu, Y., Wang, T., Shen, J.: Topomlp: An simple yet strong pipeline for driving topology reasoning. arXiv preprint arXiv:2310.06753 (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 281, + 481, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 281, + 481, + 314 + ], + "spans": [ + { + "bbox": [ + 132, + 281, + 481, + 314 + ], + "type": "text", + "content": "64. Wu, D., Jia, F., Chang, J., Li, Z., Sun, J., Han, C., Li, S., Liu, Y., Ge, Z., Wang, T.: The 1st-place solution for cvpr 2023 openlane topology in autonomous driving challenge. arXiv preprint arXiv:2306.09590 (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 315, + 481, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 315, + 481, + 357 + ], + "spans": [ + { + "bbox": [ + 132, + 315, + 481, + 357 + ], + "type": "text", + "content": "65. Xie, E., Yu, Z., Zhou, D., Philion, J., Anandkumar, A., Fidler, S., Luo, P., Alvarez, J.M.: M$^2$BEV: Multi-Camera Joint 3D Detection and Segmentation with Unified Birds-Eye View Representation (Apr 2022). https://doi.org/10.48550/arXiv.2204.05088" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 357, + 481, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 357, + 481, + 402 + ], + "spans": [ + { + "bbox": [ + 132, + 357, + 481, + 402 + ], + "type": "text", + "content": "66. Xu, D., Ouyang, W., Wang, X., Sebe, N.: PAD-Net: Multi-Tasks Guided Prediction-and-Distillation Network for Simultaneous Depth Estimation and Scene Parsing. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 675-684 (2018)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 402, + 481, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 402, + 481, + 445 + ], + "spans": [ + { + "bbox": [ + 132, + 402, + 481, + 445 + ], + "type": "text", + "content": "67. Yan, F., Nie, M., Cai, X., Han, J., Xu, H., Yang, Z., Ye, C., Fu, Y., Mi, M.B., Zhang, L.: ONCE-3DLanes: Building Monocular 3D Lane Detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 17143-17152 (2022)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 446, + 481, + 489 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 446, + 481, + 489 + ], + "spans": [ + { + "bbox": [ + 132, + 446, + 481, + 489 + ], + "type": "text", + "content": "68. Yang, Z., Liu, S., Hu, H., Wang, L., Lin, S.: RepPoints: Point Set Representation for Object Detection. In: 2019 IEEE/CVF International Conference on Computer Vision (ICCV). pp. 9656-9665. IEEE, Seoul, Korea (South) (Oct 2019). https://doi.org/10.1109/ICCV.2019.00975" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 490, + 448, + 500 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 490, + 448, + 500 + ], + "spans": [ + { + "bbox": [ + 132, + 490, + 448, + 500 + ], + "type": "text", + "content": "69. Yao, C., Yu, L., Wu, Y., Jia, Y.: Sparse Point Guided 3D Lane Detection" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 501, + 481, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 501, + 481, + 555 + ], + "spans": [ + { + "bbox": [ + 132, + 501, + 481, + 555 + ], + "type": "text", + "content": "70. Yin, J., Fang, J., Zhou, D., Zhang, L., Xu, C.Z., Shen, J., Wang, W.: Semi-supervised 3D Object Detection with Proficient Teachers. In: Avidan, S., Bros-tow, G., Cisse, M., Farinella, G.M., Hassner, T. (eds.) Computer Vision - ECCV 2022, vol. 13698, pp. 727-743. Springer Nature Switzerland, Cham (2022). https://doi.org/10.1007/978-3-031-19839-7_42" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 555, + 481, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 555, + 481, + 599 + ], + "spans": [ + { + "bbox": [ + 132, + 555, + 481, + 599 + ], + "type": "text", + "content": "71. Yin, J., Shen, J., Chen, R., Li, W., Yang, R., Frossard, P., Wang, W.: Is-fusion: Instance-scene collaborative fusion for multimodal 3d object detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14905-14915 (2024)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 600, + 481, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 600, + 481, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 600, + 481, + 632 + ], + "type": "text", + "content": "72. Yin, J., Wang, W., Meng, Q., Yang, R., Shen, J.: A unified object motion and affinity model for online multi-object tracking. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6768-6777 (2020)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 632, + 481, + 664 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 481, + 664 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 481, + 664 + ], + "type": "text", + "content": "73. Yu, T., Kumar, S., Gupta, A., Levine, S., Hausman, K., Finn, C.: Gradient Surgery for Multi-Task Learning (Dec 2020). https://doi.org/10.48550/arXiv.2001.06782" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 276, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 276, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 276, + 91, + 447, + 102 + ], + "type": "text", + "content": "RepVF: Unified Multi-task 3D Perception" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "text", + "content": "19" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 18 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 482, + 270 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 132, + 116, + 482, + 159 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 482, + 159 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 482, + 159 + ], + "type": "text", + "content": "74. Zhang, Y., Wang, C., Wang, X., Zeng, W., Liu, W.: FairMOT: On the Fairness of Detection and Re-Identification in Multiple Object Tracking. International Journal of Computer Vision 129(11), 3069–3087 (Nov 2021). https://doi.org/10.1007/s11263-021-01513-4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 160, + 481, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 160, + 481, + 205 + ], + "spans": [ + { + "bbox": [ + 132, + 160, + 481, + 205 + ], + "type": "text", + "content": "75. Zhou, D., Fang, J., Song, X., Liu, L., Yin, J., Dai, Y., Li, H., Yang, R.: Joint 3d instance segmentation and object detection for autonomous driving. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1839-1849 (2020)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 205, + 481, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 205, + 481, + 237 + ], + "spans": [ + { + "bbox": [ + 132, + 205, + 481, + 237 + ], + "type": "text", + "content": "76. Zhou, D., Song, X., Dai, Y., Yin, J., Lu, F., Liao, M., Fang, J., Zhang, L.: Iafa: Instance-aware feature aggregation for 3d object detection from a single image. In: Proceedings of the Asian Conference on Computer Vision (2020)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 237, + 481, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 237, + 481, + 270 + ], + "spans": [ + { + "bbox": [ + 132, + 237, + 481, + 270 + ], + "type": "text", + "content": "77. Zhu, X., Su, W., Lu, L., Li, B., Wang, X., Dai, J.: Deformable DETR: Deformable Transformers for End-to-End Object Detection (Mar 2021). https://doi.org/10.48550/arXiv.2010.04159" + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "20" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 211, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 211, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 211, + 100 + ], + "type": "text", + "content": "C. Li et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 19 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Repaint123_ Fast and High-quality One Image to 3D Generation with Progressive Controllable Repainting/a8257f3f-735a-49a4-ab19-4aa6628f29e5_content_list.json b/2024/Repaint123_ Fast and High-quality One Image to 3D Generation with Progressive Controllable Repainting/a8257f3f-735a-49a4-ab19-4aa6628f29e5_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..0c3d531f99a6e2f0761a07610b3c89b619457744 --- /dev/null +++ b/2024/Repaint123_ Fast and High-quality One Image to 3D Generation with Progressive Controllable Repainting/a8257f3f-735a-49a4-ab19-4aa6628f29e5_content_list.json @@ -0,0 +1,1596 @@ +[ + { + "type": "text", + "text": "Repaint123: Fast and High-quality One Image to 3D Generation with Progressive Controllable Repainting", + "text_level": 1, + "bbox": [ + 222, + 140, + 781, + 209 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Junwu Zhang1* Zhenyu Tang1* Yatian Pang1,3 Xinhua Cheng1 Peng Jin1 Yida Wei4 Xing Zhou5 Munan Ning1,2 Li Yuan1,2* \n1Peking University 2Pengcheng Laboratory 3National University of Singapore \n45RabbitPre", + "bbox": [ + 222, + 234, + 781, + 297 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/7dfab4bc6fd25e686d0908268ecef8871845a43d7f8ede6eaf58479ad8b584e7.jpg", + "image_caption": [ + "Fig. 1: Repaint123 generates high-quality 3D content with detailed texture from a single image in 2 minutes. Repaint123 adopts 3D Gaussian Splatting in the coarse stage and then utilizes a progressively repairing strategy with the diffusion model for high-quality 3D content generation with efficiency." + ], + "image_footnote": [], + "bbox": [ + 217, + 339, + 784, + 523 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Recent image-to-3D methods achieve impressive results with plausible 3D geometry due to the development of diffusion models and optimization techniques. However, existing image-to-3D methods suffer from texture deficiencies in novel views, including multi-view inconsistency and quality degradation. To alleviate multi-view bias and enhance image quality in novel-view textures, we present Repaint123, a fast image-to-3D approach for creating high-quality 3D content with detailed textures. Repaint123 proposes a progressively repairing strategy to simultaneously enhance the consistency and quality of textures across different views, generating invisible regions according to visible textures, with the visibility map calculated by the depth alignment across views.", + "bbox": [ + 259, + 647, + 743, + 801 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Equal contribution.", + "bbox": [ + 217, + 811, + 364, + 824 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "\\*\\* Corresponding author.", + "bbox": [ + 215, + 825, + 383, + 839 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Furthermore, multiple control techniques, including reference-driven information injection and coarse-based depth guidance, are introduced to alleviate the texture bias accumulated during the repainting process for improved consistency and quality. For novel-view texture refinement with short-term view consistency, our method progressively repairs novel-view images with adaptive strengths based on visibility, enhancing the balance of image quality and view consistency. To alleviate the accumulated bias as progressively repainting, we control the repainting process by depth-guided geometry and attention-driven reference-view textures. Extensive experiments demonstrate the superior ability of our method to create 3D content with consistent and detailed textures in 2 minutes.", + "bbox": [ + 259, + 147, + 743, + 299 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Keywords: Image-to-3D $\\cdot$ Texture Enhancement $\\cdot$ Diffusion model", + "bbox": [ + 261, + 313, + 710, + 327 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 354, + 374, + 369 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Generating 3D content from one given reference image plays a key role at the intersection of computer vision and computer graphics [11, 17, 24, 25, 33, 35], which is desired by users for innovative applications across fields including robotics, virtual reality, and augmented reality. However, the image-to-3D task is quite challenging due to the generated 3D content is expected with reasonable geometry and consistent textures, while the single image input is insufficient in multi-view information. Recent studies [24, 28, 35, 50, 51] employ 2D diffusion models [14, 39] to guide 3D generation from the reference image with Score Distillation Sampling (SDS) [34]. While preliminary attempts [28, 51] leveraging view-independent diffusion models (e.g. Stable Diffusion [39]) faced challenges with over-saturated textures and multi-face geometry, more", + "bbox": [ + 212, + 386, + 488, + 748 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "recent efforts [24, 26, 35, 46, 50] utilize view-conditioned diffusion models [24, 26, 46]. Although these methods utilizing view-conditioned diffusion models mitigate the multi-face problem and generate 3D content with plausible geometry, they still encounter challenges in generating high-quality textures. This limitation stems from training these models on small-scale synthetic 3D datasets, hindering their ability to produce detailed images from novel views.", + "bbox": [ + 212, + 750, + 787, + 840 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/de3c7f701ab1e3f72825b2f035b8dab4103ed427b498cf030a640de73bdc38cd.jpg", + "image_caption": [ + "Fig.2: Motivation. Current image-to-3D methods adopt SDS loss for generating, resulting in inconsistent and poor textures. Repaint123 proposes a progressively repainting strategy based on visible content to efficiently create invisible content with MSE loss, achieving consistent and high-quality textures." + ], + "image_footnote": [], + "bbox": [ + 500, + 417, + 787, + 599 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "Zhang et al.", + "bbox": [ + 271, + 114, + 354, + 128 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address textural deficiencies including inconsistency and low quality mentioned above, we propose a novel approach named Repaint123 for fast and high-quality image-to-3d generation by generating consistent and high-quality novel-view images with a controllable content repaint process. As shown in Figure 2, the core of Repaint123 is to progressively repaint the invisible content with a view-independent diffusion model for adjacent-view consistency, incorporating multiple controls of geometry and texture for enhanced quality and multi-view consistency. Specifically, our method adopts a two-stage optimization framework. In the coarse stage, we utilize 3D Gaussian Splatting [18] as the representation to efficiently obtain a coarse 3D model in 1 minute. In the fine stage, we progressively sample the camera views around and repaint invisible content with control techniques including using novel-view depth map as geometry prior, injecting reference-view attention features as texture prior [61] and employing reference-view image prompt as semantic prior. To further enhance image quality and preserve multi-view consistency, we propose a visibility-based adaptive strategy that refines the visible regions with different strengths. After achieving high-qualityainted images with multi-view consistency, we directly leverage simple MSE loss to efficiently refine 3D content textures instead of the time-consuming and over-smoothed SDS.", + "bbox": [ + 217, + 146, + 785, + 431 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We conduct extensive experiments on multiple datasets and demonstrate that our method significantly advances image-to-3D generation, producing high-quality, multi-view consistent 3D content with detailed textures in approximately 2 minutes from scratch (as shown in Figure 1), outperforming current state-of-the-art NeRF-based and Gaussian-Splatting-based methods in both consistency and texture quality. Our contributions can be summarized as follows:", + "bbox": [ + 212, + 433, + 782, + 523 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose a progressively repairing strategy based on visibility for the image-to-3D task to generate 3D content with consistency and high quality by repairing invisible parts according to visible textures.", + "- We introduce multiple control techniques including reference-driven information injection and coarse-based depth guidance to alleviate the texture bias accumulated during the repainting process for further quality enhancement.", + "- Our experiments demonstrate Repaint123 efficiently generates high-quality 3D content with consistent and detailed textures in 2 minutes." + ], + "bbox": [ + 225, + 532, + 784, + 652 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Works", + "text_level": 1, + "bbox": [ + 215, + 676, + 395, + 691 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.1 Diffusion Models for 3D Generation", + "text_level": 1, + "bbox": [ + 215, + 709, + 558, + 723 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The recent notable achievements in view-independent 2D diffusion models [14,39] have brought about exciting prospects for generating 3D objects. Pioneering studies [34, 53] have introduced the concept of distilling a 2D text-to-image generation model to generate 3D shapes. Subsequent works [1, 6, 7, 10, 16, 22, 28, 35, 37, 42-44, 51, 52, 55-57, 60, 61, 64] have adopted a similar per-object optimization approach, building upon these initial works. Nevertheless, the majority of these techniques consistently experience low efficiency and multi-face", + "bbox": [ + 212, + 733, + 784, + 839 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "Repaint123: Progressive Controllable Repainting for 3D Generation", + "bbox": [ + 282, + 114, + 730, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "issues. Unlike the previous study HiFi-123 [61], which employed similar inversion and attention injection techniques for image-to-3D generation, our approach diverges primarily in our proposed repairing strategy, as well as many variations in optimization objectives, image prompting, and 3D representations. Recently, some works [23, 25, 26, 48] extend view-independent diffusion model to view-conditioned diffusion model to generate multi-view images for reconstruction, while these methods usually suffer from low-quality textures as the multi-view diffusion models are trained on limited and synthesized data. To optimize both consistency and quality, we integrate a view-conditioned diffusion model for initial consistency in the coarse stage and a view-independent model to enhance texture quality in the fine stage.", + "bbox": [ + 212, + 146, + 787, + 313 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.2 Controllable Image Synthesis", + "text_level": 1, + "bbox": [ + 214, + 335, + 503, + 351 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "A major challenge in the field of image generation is achieving controllability. Many works have been done recently to increase the controllability of images generated by diffusion models. ControlNet [62] and T2I-Adapter [31] attempt to control the creation of images by utilizing data from different modalities. Besides, some optimization-based methods [12,30,40] learn new parameters or fine-tune the diffusion model in order to enhance control over the generation process. Other methods [3,58] leverage multi-view attention to introduce information from other-view images for gaining superior control. We incorporate multiple control techniques into the view-independent diffusion model for maintaining view consistency in the fine stage.", + "bbox": [ + 212, + 361, + 787, + 512 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.3 3D Representations", + "text_level": 1, + "bbox": [ + 214, + 535, + 426, + 550 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Neural Radiance Fields (NeRF) [29], as a volumetric rendering method, has gained popularity for its ability to enable 3D optimization [2,4,8,13,21] under 2D supervision, while NeRF optimization can be time-consuming. Numerous efforts [32,41] for spatial pruning have been dedicated to accelerating the training process of NeRF on the reconstruction setting. However, they fail in the generation setting. Recently, 3D Gaussian Splatting [5,9,18,50,59] has emerged as an alternative 3D representation to NeRF and has shown remarkable advancements in terms of both quality and speed, offering a promising avenue. Therefore, we adopt Gaussian Splatting for efficiently generating coarse 3D model.", + "bbox": [ + 212, + 560, + 787, + 698 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3 Preliminary", + "text_level": 1, + "bbox": [ + 214, + 720, + 370, + 739 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1 DDIM Inversion", + "text_level": 1, + "bbox": [ + 214, + 753, + 401, + 768 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "DDIM [47] transforms random noise $\\pmb{x}_T$ into clean data $\\pmb{x}_0$ over a series of time steps, by using the deterministic DDIM sampling in the reverse process, i.e., $\\pmb{x}_{t - 1} = (\\alpha_{t - 1} / \\alpha_t)(\\pmb{x}_t - \\sigma_t\\epsilon_\\phi) + \\sigma_{t - 1}\\epsilon_\\phi$ . On the contrary, DDIM inversion progressively converts clean data to a noisy state $\\pmb{x}_T$ , i.e., $\\pmb{x}_t = (\\alpha_t / \\alpha_{t - 1})(\\pmb{x}_{t - 1} - \\sigma_t\\epsilon_\\phi)$ .", + "bbox": [ + 212, + 779, + 787, + 842 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "Zhang et al.", + "bbox": [ + 271, + 114, + 354, + 128 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/f803aef4bd19c57067c8f2e5580f4c5a5ecb736b5efa92be62e662497d20c5be.jpg", + "image_caption": [ + "Fig. 3: Overview of Repaint123. In the coarse stage, we leverage 3D Gaussian Splatting to represent the coarse content, which is optimized by MSE loss at the reference and SDS loss at the invisible views. In the fine stage, we convert coarse content to mesh and progressively sample and generate invisible views from the reference view bidirectionally. Concretely, the current view is repainted based on the visibility map from bidirectional neighbor views with control techniques including attention injection and depth guidance, leading to consistent and detailed textures." + ], + "image_footnote": [], + "bbox": [ + 218, + 148, + 782, + 296 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "$\\sigma_{t-1}\\epsilon_{\\phi}) + \\sigma_t\\epsilon_{\\phi}$ , here $\\epsilon_{\\phi}$ is the predicted noise by the UNet. This method can precisely reconstruct the original clean data while greatly speeding up the process by skipping many intermediate diffusion steps.", + "bbox": [ + 212, + 436, + 787, + 484 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2 3D Gaussian Splitting", + "text_level": 1, + "bbox": [ + 214, + 505, + 449, + 522 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Gaussian Splatting [18] presents a novel method for synthesizing new views and reconstructing 3D scenes, achieving real-time speed. Unlike NeRF, Gaussian Splatting uses a set of anisotropic 3D Gaussians defined by their locations, covariances, colors, and opacities to represent the scene. To compute the color of each pixel $\\mathbf{p}$ in the image, it utilizes a typical neural point-based rendering [19,20], The rendering process is as follows:", + "bbox": [ + 212, + 530, + 787, + 621 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nC (\\mathbf {p}) = \\sum_ {i \\in \\mathcal {N}} c _ {i} \\alpha_ {i} \\prod_ {j = 1} ^ {i - 1} (1 - \\alpha_ {j}), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 393, + 633, + 785, + 676 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {w h e r e}, \\alpha_ {i} = o _ {i} e ^ {- \\frac {1}{2} (\\mathbf {p} - \\boldsymbol {\\mu} _ {i}) ^ {T} \\boldsymbol {\\Sigma} _ {i} ^ {- 1} (\\mathbf {p} - \\boldsymbol {\\mu} _ {i})},\n$$\n", + "text_format": "latex", + "bbox": [ + 372, + 678, + 630, + 698 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $c_{i}$ , $o_{i}$ , $\\mu_{i}$ , and $\\Sigma_{i}$ represent the color, opacity, position, and covariance of the $i$ -th Gaussian respectively, and $\\mathcal{N}$ denotes the number of related Gaussians.", + "bbox": [ + 214, + 709, + 787, + 739 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4 Method", + "text_level": 1, + "bbox": [ + 214, + 762, + 330, + 777 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we introduce our two-stage framework named Repaint123 for fast and high-quality image-to-3D generation, as illustrated in Figure 3. In the coarse stage, we adopt 3D Gaussian Splitting [18] representation optimized by", + "bbox": [ + 212, + 794, + 785, + 842 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "Repaint123: Progressive Controllable Repainting for 3D Generation", + "bbox": [ + 282, + 114, + 730, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "SDS loss [34] with the view-conditioned diffusion model [24] to learn a coarse geometry and texture (see Section 4.1). Inspired by [50], we convert the coarse 3D Gaussians to textured mesh to facilitate subsequent texture enhancement in UV space. In the fine stage, we progressively paint invisible content in novel views by rotating camera viewpoints bi-directionally from the reference view (see Section 4.2). To mitigate texture bias accumulated during the gradual painting across wide angles, we incorporate multiple control techniques including depth guidance, injection of reference-view attention, and the use of reference-view image prompts (see Section 4.2). Additionally, we propose a visibility-based adaptive strategy to refine visible regions when observed from a superior perspective (see Section 4.2). After obtaining a high-quality painted image that remains consistent across different views, we efficiently optimize the texture map by simple Mean Square Error (MSE) loss (see Section 4.2). This process is repeated for each sampled view, alternating between painting and optimization, until the texture is completely reconstructed to accommodate a 360-degree perspective.", + "bbox": [ + 212, + 146, + 787, + 375 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1 Coarse Stage: Gaussian Splatting with 3D Diffusion Prior", + "text_level": 1, + "bbox": [ + 214, + 395, + 736, + 411 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In the coarse stage, we adopt 3D Gaussian Splatting for efficient initialization. Through SDS [34], 3D diffusion priors are back-propagated to the 3D Gaussians. At each diffusion step, we sample a random current camera view $v_{c}$ and its camera pose $p_{c}$ , which orbits around the center of the object. Subsequently, the RGB image, denoted as $I_{\\mathrm{rgb}}^{c}$ , and the alpha image, represented by $I_{\\mathrm{a}}^{c}$ , are rendered from the current viewpoint $p$ . Given an image $\\tilde{I}_{\\mathrm{rgb}}^{r}$ and an alpha mask $\\tilde{I}_{\\mathrm{a}}^{r}$ as input to the reference view, the SDS loss can be formulated as:", + "bbox": [ + 212, + 419, + 787, + 532 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {S D S}} = \\mathbb {E} _ {t, c, \\epsilon} \\left[ (\\epsilon_ {\\phi} (I _ {\\mathrm {r g b}} ^ {c}; t, \\tilde {I} _ {\\mathrm {r g b}} ^ {r}, \\Delta p _ {c}) - \\epsilon) \\frac {\\partial I _ {\\mathrm {r g b}} ^ {c}}{\\partial \\Theta} \\right] \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 334, + 542, + 785, + 578 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\epsilon_{\\phi}(\\cdot)$ is the predicted noise by the diffusion $\\phi$ , $\\Delta p_{c}$ is the relative camera pose and $\\Theta$ represents the estimizable Gaussian parameters. Additionally, we optimize the RGB image and alpha image rendered from the reference view to align with the input reference image using reconstruction loss:", + "bbox": [ + 212, + 585, + 787, + 648 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {R e f}} = \\lambda_ {\\mathrm {r g b}} | | I _ {\\mathrm {r g b}} ^ {r} - \\tilde {I} _ {\\mathrm {r g b}} ^ {r} | | _ {2} ^ {2} + | | I _ {\\mathrm {a}} ^ {r} - \\tilde {I} _ {\\mathrm {a}} ^ {r} | | _ {2} ^ {2}. \\qquad (3)\n$$\n", + "text_format": "latex", + "bbox": [ + 357, + 657, + 785, + 676 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The final loss function is the weighted sum of $\\mathcal{L}_{\\mathrm{SDS}}$ and $\\mathcal{L}_{\\mathrm{Ref}}$ .", + "bbox": [ + 214, + 686, + 658, + 702 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2 Fine Stage: Visibility-based Controllable Repainting", + "text_level": 1, + "bbox": [ + 214, + 724, + 692, + 739 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Progressive Invisible Content Repainting. We repaint previously unseen content with progressively sampling novel camera viewpoints, beginning from the reference view and increasing a certain angular interval with each iteration. To obtain invisible regions in the image from a newly sampled view $v_{c}$ , we use the back-projection technique based on the depth map $D^{p}\\in \\mathbb{R}^{H\\times W}$ in the previously seen view $v_{p}$ . Formally, given the camera intrinsics $K\\in \\mathbb{R}^{3\\times 3}$ , we first remap the", + "bbox": [ + 212, + 750, + 787, + 842 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "Zhang et al.", + "bbox": [ + 271, + 114, + 354, + 128 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "pixels of the image rendered from $v_{p}$ into 3D points $X^{p}$ expressed in the camera coordinate frame, which can be obtained as $X_{i,j} = K^{-1}[iD_{i,j},jD_{i,j},D_{i,j}]^{\\top}$ . With $P_{p},P_{c}\\in \\mathbb{R}^{3\\times 4}$ the world-to-camera poses for images from $v_{p}$ and $v_{c}$ , we can transform $X^{p}$ into camera coordinate frame of $v_{c}$ to obtain $X^{p,c}$ :", + "bbox": [ + 214, + 146, + 787, + 209 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\nX ^ {p, c} = P _ {p} P _ {c} ^ {- 1} h \\left(X ^ {p}\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 426, + 220, + 785, + 239 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $h:(x,y,z)\\to (x,y,z,1)$ is the homogeneous mapping. Then we can render a depth map $D^{p,c}$ from $X^{p,c}$ as $D_{i,j}^{p,c} = KX_{i,j,2}^{p,c}$ . Regions with dissimilar depth values between $D^c$ and $D^{p,c}$ are recognized as invisible regions, forming a repainting mask $M$ where areas with zeros are invisible and areas with ones are visible. Therefore, we repaint invisible content and maintain visible content by:", + "bbox": [ + 214, + 250, + 787, + 325 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\nx _ {t} = x _ {t} ^ {\\mathrm {i n v}} \\odot M + x _ {t} ^ {\\mathrm {r e v}} \\odot (1 - M), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 385, + 338, + 785, + 356 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $\\odot$ is element-wise matrix multiplication, $t$ is the current denoising timestep, $x_{t}^{\\mathrm{inv}}$ is the DDIM [47] inverted latent, and $x_{t}^{\\mathrm{rev}} \\sim \\mathcal{N}(\\mu_{\\phi}(x_{t + 1},t),\\Sigma_{\\phi}(x_{t + 1},t))$ is the diffusion model output.", + "bbox": [ + 214, + 367, + 787, + 412 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Additionally, to mitigate accumulated bias for distant camera views during gradual repainting, as shown in the fine stage in Figure 3, we sample camera viewpoints bi-directionally, alternating between clockwise and counterclockwise directions. This bidirectional sampling allows us to paint up to 180 degrees from two opposite directions, instead of painting across the entire 360 degrees from one direction. Moreover, to paint bidirectionally invisible content and ensure smooth transitions at the junction of the two directions, we replace the previous unidirectional painting mask $M$ by a bidirectional painting mask. This mask merges two unidirectional masks from bidirectional neighbor views (red cameras in Figure 3) by taking the maximum value of each pixel.", + "bbox": [ + 214, + 412, + 787, + 564 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Control-based Content Enhancement. While the progressive painting maintains consistency between adjacent views, views that are farther apart may become inconsistent and experience a decline in quality due to cumulative errors. To tackle these issues, we improve the painting process by incorporating various control techniques to enhance the consistency and quality of synthesized images. Specifically, for geometry control, we employ ControlNet [62] to condition the diffusion model with coarse depth maps, ensuring geometric consistency across images. To transfer reference textures and mitigate the cumulative texture bias during the progressive painting, inspired by HiFi-123 [61], we incorporate an attention injection mechanism [3] that injects reference attention features into the diffusion denoising process. By replacing the novel-view content features (Key features $K_{t}$ and Value features $V_{t}$ ) with reference-view attention features ( $K_{r}$ and $V_{r}$ ) during each denoising step, the novel-view image features can directly query the high-quality reference features by:", + "bbox": [ + 214, + 587, + 787, + 800 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {A t t e n t i o n} \\left(Q _ {t}, K _ {r}, V _ {r}\\right) = \\operatorname {S o f t m a x} \\left(\\frac {Q _ {t} K _ {r} ^ {T}}{\\sqrt {d}}\\right) V _ {r}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 336, + 809, + 785, + 844 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "Repaint123: Progressive Controllable Repainting for 3D Generation", + "bbox": [ + 282, + 114, + 732, + 130 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c0c41a3644127407014e0bd4b849a1abf8447e2b634353c776f0f94ade11bebe.jpg", + "image_caption": [ + "Fig. 4: Visibility-based Controllable Repainting Pipeline. Our method repairs invisible content (black areas in $M_t$ ) of novel-view rendered image $I_{\\mathrm{rgb}}$ according to the inverted latent $x_t^{inv}$ and the visibility map $V$ to obtain refined image $I_{\\mathrm{rgb}}^{\\mathrm{fine}}$ . We introduce multiple control techniques to enhance the consistency and quality of generated textures, including attention feature injection, depth guidance by ControlNet, and semantic prompt by CLIP encoder. Noticing that the repainting mask $M_t$ is simultaneously changed with denoising timesteps for adaptive refinement." + ], + "image_footnote": [], + "bbox": [ + 220, + 148, + 784, + 333 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $Q_{t}$ represents the novel-view query features. This enhancement facilitates the transfer of texture details and improves the consistency between the reference-view image and novel-view images.", + "bbox": [ + 212, + 470, + 784, + 516 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "During our experiments, we observed a decline in image quality during repainting, likely due to not using text prompts for performing classifier-free guidance [15], crucial for producing high-quality images in diffusion models. To improve this, we adopt IP-Adapter [58], which uses a CLIP encoder to convert reference images into semantic features. These features are then projected into image prompts, providing visual cues that enhance image generation through classifier-free guidance.", + "bbox": [ + 212, + 517, + 787, + 623 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Adaptive Visible Content Refinement. Repainting only the invisible content often leads to distortion in the visible area As shown in Figure 5, when the previous view is an oblique view, it leads to a low-resolution update on the texture maps, resulting in high distortion when rendering from a better observing view. Therefore, the visible content currently observed in a superior view should be refined to replenish details. However, the selection of refinement strength for these visible content is tricky [38, 54], as excessive strength produces unfaithful results leading to inconsistency with previously repainted views, while insufficient strength limits quality improvement. We propose a visibility-based adaptive refinement strategy to refine these previously seen regions with different strengths, aiming to achieve improved quality-consistency trade-off. As visualized in Figure 5, we view painting for visible content as a process similar to superresolution that replenishes detailed information. According to the Orthographic", + "bbox": [ + 212, + 643, + 787, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 116, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "Zhang et al.", + "bbox": [ + 271, + 114, + 354, + 128 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Projection Theorem, which asserts that the projected resolution of a surface is directly proportional to the view obliqueness ( $\\cos\\theta$ as shown in Figure 5), we can assume that the repainting strength is equal to $(1 - \\cos\\theta_{1} / \\cos\\theta_{2})$ , where $\\cos\\theta_{2} / \\cos\\theta_{1}$ represents the upsampling scale in terms of super-resolution and visibility variations in the context of 3D rendering. Therefore, we can define a visibility map $V$ as $V_{i,j} = \\cos\\theta_{i,j}^{p,c} / \\cos\\theta_{i,j}^{c}$ , where $\\cos\\theta_{i,j}^{p,c}$ and $\\cos\\theta_{i,j}^{c}$ are obtained by multiplying the current-view normal map with camera rays of the previous view $v_{p}$ and the current view $v_{c}$ respectively. Higher visibility value indicates less repainting strength, while invisible areas require full strength.", + "bbox": [ + 212, + 146, + 486, + 431 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "By this association between repairing strength and the visibility map $V$ , we can binarize the visibility map to the timestep-aware adaptive repairing mask $M_{t}$ during each denoising step, visualized in the green box \"Timestep-aware binarization\" in Figure 4:", + "bbox": [ + 212, + 433, + 787, + 494 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\nM _ {t} ^ {i, j} = \\left\\{ \\begin{array}{l l} 1, & \\text {i f} V ^ {i, j} > 1 - t / T \\\\ 0, & \\text {e l s e}, \\end{array} \\right. \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 388, + 511, + 785, + 551 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where $i$ , and $j$ are the 2D position of pixels in visibility map $V$ , and $T$ is the total number of timesteps of the diffusion model. By doing this, we can adaptively refine visible content with a proper strength.", + "bbox": [ + 212, + 566, + 787, + 613 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Efficient Optimization. Using a progressive painting strategy and various control techniques to refine the inverted latent, we obtain a high-quality, view-consistent refined image, $I_{\\mathrm{rgb}}^{\\mathrm{fine}}$ . This process allows us to use simple MSE loss for efficient texture map optimization:", + "bbox": [ + 212, + 651, + 787, + 712 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {f i n e}} = \\left| \\left| I _ {\\mathrm {r g b}} ^ {\\text {f i n e}} - I _ {\\mathrm {r g b}} \\right| \\right| _ {2} ^ {2}, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 416, + 728, + 785, + 748 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where $I_{\\mathrm{rgb}}$ represents the rendered image in the fine stage. MSE loss is faster and more deterministic to optimize than the traditional SDS loss, speeding up the optimization process. After sampling a novel view, we first conduct image repainting and then proceed with efficient texture optimization, repeating this process until we cover a 360-degree view range.", + "bbox": [ + 212, + 763, + 787, + 842 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/3c926188a9f39f0ddd010635916674dffc9e5b8ea128ee7aed973698f09f7502.jpg", + "image_caption": [ + "Fig. 5: Relation between view obliqueness and refinement strength. Red box shows the same texture part in different views. Texture details in generated views might be insufficient due to the variation of angles." + ], + "image_footnote": [], + "bbox": [ + 509, + 150, + 777, + 318 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "Repaint123: Progressive Controllable Repainting for 3D Generation", + "bbox": [ + 282, + 114, + 730, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 114, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "5 Experiment", + "text_level": 1, + "bbox": [ + 215, + 143, + 366, + 162 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "5.1 Implementation Details", + "text_level": 1, + "bbox": [ + 215, + 183, + 455, + 198 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "In our experiment, we consistently apply the same hyperparameters across all our method's results. We progressively turn around the viewpoints by 40 degrees each time, consequently obtaining 8 views per object. We use a 50-step DDIM schedule and perform a 30-step latent diffusion inversion. Stable Diffusion 1.5 is utilized for all methods. During the coarse stage training, we set $\\lambda_{\\mathrm{rgb}}$ to 10. Generating a single 3D object takes only 2 minutes on a single 40G A100 GPU, approximately 1 minute for the coarse stage and mesh extraction, and another 1 minute for refinement.", + "bbox": [ + 215, + 215, + 785, + 335 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "5.2 Baselines", + "text_level": 1, + "bbox": [ + 215, + 366, + 338, + 382 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "We adopt RealFusion [28], Make-It-3D [51], and Zero123-XL [24], Magic123 [35] as our NeRF-based baselines and DreamGaussian [50] as our Gaussian-Splatting-based baseline. RealFusion presents a single-stage algorithm for NeRF generation leveraging 2D SDS loss for novel views. Make-It-3D is a two-stage approach that shares similar objectives with RealFusion but employs a point cloud representation for refinement at the second stage. For Zero123-XL, we adopt the implementation [49] and add a mesh fine-tuning stage for fair comparison. Integrating Zero123 and RealFusion, Magic123 incorporates a 2D SDS loss with 3D SDS loss provided by Zero123 to balance geometry and texture and adopts DMTet [45] representation at the second stage. DreamGaussian integrates 3D Gaussian Splatting into 3D generation and greatly improves the speed.", + "bbox": [ + 215, + 398, + 785, + 566 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "5.3 Evaluation Protocol", + "text_level": 1, + "bbox": [ + 215, + 595, + 426, + 609 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Datasets. Based on previous research, we utilized the Realfusion15 dataset [28] and test-alpha dataset collected by Make-It-3D [51], which comprises many common objects from diverse styles.", + "bbox": [ + 215, + 627, + 785, + 672 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Evaluation metrics. An effective 3D generation approach should produce 3D content which closely resemble the reference view, and maintain consistency of semantics and textures with the reference image when observed from new views. Therefore, to evaluate the overall quality of the generated 3D object, we choose the following metrics from two aspects: 1) PSNR and LPIPS [63], which measure pixel-level and perceptual reconstruction quality respectively at the reference view; 2) CLIP Similarity [36] and Contextual Distance [27], which assess the similarity of semantics and detailed textures respectively between the novel perspective and the reference view.", + "bbox": [ + 215, + 703, + 785, + 839 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "Zhang et al.", + "bbox": [ + 271, + 114, + 354, + 128 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/4c34e05eed042a781c9a901ee81c2129140d5ad78431f0f79ad645c6d86c11e9.jpg", + "image_caption": [ + "Fig. 6: Qualitative comparisons on image-to-3D generation. Zoom in for details." + ], + "image_footnote": [], + "bbox": [ + 241, + 157, + 763, + 777 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "Repaint123: Progressive Controllable Repainting for 3D Generation", + "bbox": [ + 282, + 114, + 730, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/ae0fa86480ddcf1a5826b7e8f8a59ea30eda7ef1b76cf60210058224d4bf32bf.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
DatasetMetrics\\MethodsNeRF-basedGaussian-Splatting-based
RealFusionMake-it-3DZero-123-XL*Magic123DreamGaussianRepaint123
RealFusion15CLIP-Similarity↑0.710.810.830.820.770.85
Context-Dis↓2.201.821.591.641.611.55
PSNR↑19.2416.5619.5619.6818.9419.00
LPIPS↓0.1940.1770.1080.1070.1110.101
Test-alphaCLIP-Similarity↑0.680.760.840.840.790.88
Context-Dis↓2.201.731.521.571.621.50
PSNR↑22.9117.2124.3924.6922.3322.38
LPIPS↓0.1050.2370.0500.0460.0570.048
Optimization time20min1h30min1h (+2h)2min2 min
", + "bbox": [ + 215, + 143, + 782, + 285 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Table 1: We show quantitative results in terms of CLIP-Similarity↑ / Contextual-Distance↓ / PSNR↑ / LPIPS↓. The results are shown on the RealFusion15 and test-alpha datasets, while bold reflects the best for all methods and the underline represents the best for Gaussian-Splatting-based methods. * indicates that Zero123-XL incorporates a mesh fine-tuning stage for further quality improvement. The time required by textual inversion is indicated in parentheses.", + "bbox": [ + 212, + 287, + 787, + 372 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "5.4 Comparisons", + "text_level": 1, + "bbox": [ + 215, + 412, + 370, + 428 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Comparisons with NeRF-based Methods. As shown in Table 1, we evaluate the quality of generated 3D objects across various NeRF-based methods. Our method achieves superior 3D consistency in generating 3D objects, as evidenced by the best performance of CLIP-similarity and Contextual-distance. Regarding reference-view reconstruction quality, there is a gap compared with NeRF-based approaches as shown in Table 1, which we attribute to the immaturity of current Gaussian-Splatting-based methods. Compared to NeRF-based methods for the optimization time, our approach reaches a significant acceleration of over 10 times and simultaneously achieves high quality, due to the 3D Gaussian-Splatting representation in coarse stage and efficient texture refinement in fine stage. As shown in Figure 6, Repaint123 achieves the best visual results in terms of texture consistency and generation quality as opposed to other NeRF-based methods. From the visual comparison, our method achieves consistent and detailed textures in invisible areas, while Zero123-XL results in over-smooth textures, Magic123 produces inconsistent and oversaturated colors, Realfusion and Make-It-3D fail to generate full geometry and consistent textures.", + "bbox": [ + 212, + 438, + 787, + 681 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Comparisons with GS-based methods. We conduct comparisons with other works based on 3D Gaussian Splatting in the last column of Table 1. Within a comparable generation time, our method demonstrates superiority over existing Gaussian-based approach in aspects of texture consistency and the quality of reference-view reconstruction. The superiority of our proposed Repaint123 is evidenced by the evaluation of four distinct metrics from the Table 1. As shown in Figure 6, DreamGaussian usually leads to over-smooth texture inconsistent with reference view, while our method can produce high-quality 3D content with view-consistent and detailed textures.", + "bbox": [ + 212, + 703, + 787, + 839 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "Zhang et al.", + "bbox": [ + 271, + 114, + 354, + 128 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/46cee30e8c7b0c143b6339a8dd8e9bcd57a50eaae9fc1070caf5cb9b894a8502.jpg", + "image_caption": [ + "Fig. 7: Qualitative ablation study by removing one component at a time from the overall method. The presence of artifacts is highlighted by red boxes. Inconsistencies such as multi-face problems and mismatches in content and style are evident without using repainting and attention injection strategies. Absence of image prompting and adaptive refinement notably degrades quality." + ], + "image_footnote": [], + "bbox": [ + 250, + 147, + 754, + 452 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "5.5 Ablation and Analysis", + "text_level": 1, + "bbox": [ + 215, + 564, + 444, + 579 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "This section details both qualitative and quantitative analyses to highlight the effectiveness of our proposed methods, as illustrated in Figure 7 and Table 2.", + "bbox": [ + 212, + 595, + 785, + 625 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Effectiveness of Repainting. Figure 7 illustrates that without repainting strategy, the generated novel-view image tends to resemble the reference image, leading to inconsistencies in the shared visible areas between different views, such as the content misalignment (observed in the Ice Cream example) and issues with multiple faces (evident in the Anya example). These issues stem from the absence of alignment constraints for shared visible regions, leading to conflicts and quality degradation of the reconstructed 3D texture.", + "bbox": [ + 212, + 627, + 787, + 733 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Impact of Attention Injection. Table 2 shows that injecting attention features from the reference-view image markedly enhances the consistency of both semantics and fine-grained textures, as evidenced by increased CLIP similarity and reduced Contextual distance. Without this injection strategy, as depicted in Figure 7, the synthesized images retain basic semantics but fail to accurately transfer detailed textures and styles from the reference view, resulting in inconsistencies across multiple views.", + "bbox": [ + 212, + 734, + 787, + 839 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "Repaint123: Progressive Controllable Repainting for 3D Generation", + "bbox": [ + 282, + 114, + 730, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 784, + 126 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/1771ae6aec9201acec78ecf9651eee077f8c384bc9a468f34982f0f6e02f10f9.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Method \\ MetricCLIP↑Contextual↓PSNR↑LPIPS↓
Coarse0.711.7821.170.133
vanilla painting0.711.6222.410.049
+attention injection0.781.5622.420.048
+image prompt0.841.5222.400.048
+adaptive refinemnet (Ours)0.881.5022.380.048
", + "bbox": [ + 277, + 143, + 723, + 241 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Table 2: Quantitative ablation study on Test-alpha dataset by progressively adding our proposed components. The last three lines show the cumulative effects of the proposed modules. Our approach, utilizing all strategies, delivered the best performance.", + "bbox": [ + 215, + 242, + 787, + 284 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Role of Image Prompt. The introduction of using the reference image as an image prompt, as shown in Table 2 and Figure 7, significantly boosts both multi-view consistency and image quality of the generated images. Without this technique to perform classifier-free guidance, the generation of detailed and consistent textures across views is compromised.", + "bbox": [ + 212, + 327, + 782, + 402 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Benefits of Adaptive Refinement. The necessity of adaptive refinement is clear from Figure 7, where its absence leads to artifacts and blurriness in obliquely viewing areas of previous views due to low-resolution updates, as mentioned in Section 4.2. Table 2 also demonstrates its benefits through improved CLIP similarity and Contextual distance.", + "bbox": [ + 212, + 402, + 784, + 479 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "6 Limitations", + "text_level": 1, + "bbox": [ + 215, + 503, + 362, + 520 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Despite the promising results, our method still has some limitations. Our method is based on 3D Gaussian Splatting representation in the coarse stage during training. While 3D Gaussian Splatting accelerates the training process, it may exhibit geometry artifacts during mesh extraction, such as holes, and inferior results compared to NeRF-based methods for the reconstruction of reference view, due to immaturity in generation tasks. This is supported by results in the Appendix, where NeRF is considered as an alternative to 3D Gaussian Splatting for the coarse stage. We expect these limitations can be mitigated with the development of 3D Gaussian Splatting in the future.", + "bbox": [ + 212, + 537, + 787, + 676 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "7 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 700, + 359, + 715 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "This work presents Repaint123 for generating high-quality 3D content from a single image in about 2 minutes. By leveraging progressive controllable paint, our approach overcomes the limitations of existing studies and achieves state-of-the-art results in terms of both texture quality and multi-view consistency, paving the way for future progress in 3D content generation from one image. Furthermore, we validate the effectiveness of our proposed method through a comprehensive set of experiments.", + "bbox": [ + 212, + 733, + 787, + 840 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "Zhang et al.", + "bbox": [ + 271, + 114, + 354, + 128 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 217, + 143, + 401, + 162 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "This work was supported in part by Natural Science Foundation of China (No. 62332002, 62202014), and Shenzhen Basic Research Program (No.JCYJ20220813151736001)", + "bbox": [ + 212, + 178, + 862, + 209 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 233, + 321, + 250 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Armandpour, M., Zheng, H., Sadeghian, A., Sadeghian, A., Zhou, M.: Re-imagine the negative prompt algorithm: Transform 2d diffusion into 3d, alleviate janus problem and beyond. arXiv preprint arXiv:2304.04968 (2023)", + "2. Barron, J.T., Mildenhall, B., Verbin, D., Srinivasan, P.P., Hedman, P.: Mip-nerf 360: Unbounded anti-aliased neural radiance fields. CVPR (2022)", + "3. Cao, M., Wang, X., Qi, Z., Shan, Y., Qie, X., Zheng, Y.: Masactrl: Tuning-free mutual self-attention control for consistent image synthesis and editing. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 22560-22570 (October 2023)", + "4. Chan, E.R., Lin, C.Z., Chan, M.A., Nagano, K., Pan, B., Mello, S.D., Gallo, O., Guibas, L., Tremblay, J., Khamis, S., Karras, T., Wetzstein, G.: Efficient geometry-aware 3D generative adversarial networks. In: CVPR (2022)", + "5. Chen, G., Wang, W.: A survey on 3d gaussian splatting. arXiv preprint arXiv:2401.03890 (2024)", + "6. Chen, R., Chen, Y., Jiao, N., Jia, K.: *Fantasia3d: Disentangling geometry and appearance for high-quality text-to-3d content creation*. arXiv preprint arXiv:2303.13873 (2023)", + "7. Chen, Y., Zhang, C., Yang, X., Cai, Z., Yu, G., Yang, L., Lin, G.: It3d: Improved text-to-3d generation with explicit view synthesis. arXiv preprint arXiv:2308.11473 (2023)", + "8. Chen, Z., Funkhouser, T., Hedman, P., Tagliasacchi, A.: Mobilenerf: Exploiting the polygon rasterization pipeline for efficient neural field rendering on mobile architectures. arXiv preprint arXiv:2208.00277 (2022)", + "9. Chen, Z., Wang, F., Liu, H.: Text-to-3d using gaussian splatting. arXiv preprint arXiv:2309.16585 (2023)", + "0. Cheng, X., Yang, T., Wang, J., Li, Y., Zhang, L., Zhang, J., Yuan, L.: Progressive3d: Progressively local editing for text-to-3d content creation with complex semantic prompts. arXiv preprint arXiv:2310.11784 (2023)", + "1. Dou, Z., Wu, Q., Lin, C., Cao, Z., Wu, Q., Wan, W., Komura, T., Wang, W.: Tore: Token reduction for efficient human mesh recovery with transformer. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 15143-15155 (2023)", + "2. Gal, R., Alaluf, Y., Atzmon, Y., Patashnik, O., Bermano, A.H., Chechik, G., Cohen-Or, D.: An image is worth one word: Personalizing text-to-image generation using textual inversion. arXiv preprint arXiv:2208.01618 (2022)", + "3. Hedman, P., Srinivasan, P.P., Mildenhall, B., Barron, J.T., Debevec, P.: Baking neural radiance fields for real-time view synthesis. ICCV (2021)", + "4. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. arXiv preprint arxiv:2006.11239 (2020)", + "5. Ho, J., Salimans, T.: Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598 (2022)" + ], + "bbox": [ + 225, + 266, + 785, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "Repaint123: Progressive Controllable Repainting for 3D Generation", + "bbox": [ + 282, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "16. Huang, Y., Wang, J., Shi, Y., Qi, X., Zha, Z.J., Zhang, L.: Dreamtime: An improved optimization strategy for text-to-3d content creation. arXiv preprint arXiv:2306.12422 (2023)", + "17. Jun, H., Nichol, A.: Shap-e: Generating conditional 3d implicit functions. arXiv preprint arXiv:2305.02463 (2023)", + "18. Kerbl, B., Kopanas, G., Leimkuhler, T., Drettakis, G.: 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics (ToG) 42(4), 1-14 (2023)", + "19. Kopanas, G., Leimkuhler, T., Rainer, G., Jambon, C., Drettakis, G.: Neural point catacaustics for novel-view synthesis of reflections. ACM Transactions on Graphics (TOG) 41(6), 1-15 (2022)", + "20. Kopanas, G., Philip, J., Leimkuhler, T., Drettakis, G.: Point-based neural rendering with per-view optimization. In: Computer Graphics Forum. vol. 40, pp. 29-43. Wiley Online Library (2021)", + "21. Li, Z., Müller, T., Evans, A., Taylor, R.H., Unberath, M., Liu, M.Y., Lin, C.H.: Neuralangelo: High-fidelity neural surface reconstruction. In: CVPR (2023)", + "22. Lin, C.H., Gao, J., Tang, L., Takikawa, T., Zeng, X., Huang, X., Kreis, K., Fidler, S., Liu, M.Y., Lin, T.Y.: Magic3d: High-resolution text-to-3d content creation. In: CVPR (2023)", + "23. Liu, M., Xu, C., Jin, H., Chen, L., Xu, Z., Su, H., et al.: One-2-3-45: Any single image to 3d mesh in 45 seconds without per-shape optimization. arXiv preprint arXiv:2306.16928 (2023)", + "24. Liu, R., Wu, R., Van Hoorick, B., Tokmakov, P., Zakharov, S., Vondrick, C.: Zero-1-to-3: Zero-shot one image to 3d object. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 9298-9309 (2023)", + "25. Liu, Y., Lin, C., Zeng, Z., Long, X., Liu, L., Komura, T., Wang, W.: Syncdreamer: Generating multiview-consistent images from a single-view image. arXiv preprint arXiv:2309.03453 (2023)", + "26. Long, X., Guo, Y.C., Lin, C., Liu, Y., Dou, Z., Liu, L., Ma, Y., Zhang, S.H., Habermann, M., Theobalt, C., et al.: Wonder3d: Single image to 3d using cross-domain diffusion. arXiv preprint arXiv:2310.15008 (2023)", + "27. Mechrez, R., Talmi, I., Zelnik-Manor, L.: The contextual loss for image transformation with non-aligned data. In: Proceedings of the European conference on computer vision (ECCV). pp. 768-783 (2018)", + "28. Melas-Kyriazi, L., Laina, I., Rupprecht, C., Vedaldi, A.: Realfusion: 360deg reconstruction of any object from a single image. In: CVPR (2023)", + "29. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. In: ECCV (2020)", + "30. Mokady, R., Hertz, A., Aberman, K., Pritch, Y., Cohen-Or, D.: Null-text inversion for editing real images using guided diffusion models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6038-6047 (2023)", + "31. Mou, C., Wang, X., Xie, L., Zhang, J., Qi, Z., Shan, Y., Qie, X.: T2i-adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models. arXiv preprint arXiv:2302.08453 (2023)", + "32. Müller, T., Evans, A., Schied, C., Keller, A.: Instant neural graphics primitives with a multiresolution hash encoding. ACM TOG (2022)", + "33. Nichol, A., Jun, H., Dhariwal, P., Mishkin, P., Chen, M.: Point-e: A system for generating 3d point clouds from complex prompts. arXiv preprint arXiv:2212.08751 (2022)" + ], + "bbox": [ + 215, + 146, + 784, + 840 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "Zhang et al.", + "bbox": [ + 271, + 114, + 352, + 128 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "34. Poole, B., Jain, A., Barron, J.T., Mildenhall, B.: Dreamfusion: Text-to-3d using 2d diffusion. arXiv preprint arXiv:2209.14988 (2022)", + "35. Qian, G., Mai, J., Hamdi, A., Ren, J., Siarohin, A., Li, B., Lee, H.Y., Skorokhodov, I., Wonka, P., Tulyakov, S., et al.: Magic123: One image to high-quality 3d object generation using both 2d and 3d diffusion priors. arXiv preprint arXiv:2306.17843 (2023)", + "36. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021)", + "37. Raj, A., Kaza, S., Poole, B., Niemeyer, M., Ruiz, N., Mildenhall, B., Zada, S., Aberman, K., Rubinstein, M., Barron, J., et al.: Dreambooth3d: Subject-driven text-to-3d generation. arXiv preprint arXiv:2303.13508 (2023)", + "38. Richardson, E., Metzer, G., Alaluf, Y., Giryes, R., Cohen-Or, D.: Texture: Text-guided texturing of 3d shapes. arXiv preprint arXiv:2302.01721 (2023)", + "39. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10684-10695 (2022)", + "40. Ruiz, N., Li, Y., Jampani, V., Pritch, Y., Rubinstein, M., Aberman, K.: Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 22500-22510 (2023)", + "41. Sara Fridovich-Keil and Alex Yu, Tancik, M., Chen, Q., Recht, B., Kanazawa, A.: Plenoxels: Radiance fields without neural networks. In: CVPR (2022)", + "42. Seo, H., Kim, H., Kim, G., Chun, S.Y.: Ditto-nerf: Diffusion-based iterative text to omni-directional 3d model. arXiv preprint arXiv:2304.02827 (2023)", + "43. Seo, J., Jang, W., Kwak, M.S., Ko, J., Kim, H., Kim, J., Kim, J.H., Lee, J., Kim, S.: Let 2d diffusion model know 3d-consistency for robust text-to-3d generation. arXiv preprint arXiv:2303.07937 (2023)", + "44. Shen, Q., Yang, X., Wang, X.: Anything-3d: Towards single-view anything reconstruction in the wild. arXiv preprint arXiv:2304.10261 (2023)", + "45. Shen, T., Gao, J., Yin, K., Liu, M.Y., Fidler, S.: Deep marching tetrahedra: a hybrid representation for high-resolution 3d shape synthesis. In: Advances in Neural Information Processing Systems (NeurIPS) (2021)", + "46. Shi, Y., Wang, P., Ye, J., Long, M., Li, K., Yang, X.: Mvdream: Multi-view diffusion for 3d generation. arXiv preprint arXiv:2308.16512 (2023)", + "47. Song, J., Meng, C., Ermon, S.: Denoising diffusion implicit models. arXiv:2010.02502 (October 2020), https://arxiv.org/abs/2010.02502", + "48. Szymanowicz, S., Rupprecht, C., Vedaldi, A.: Viewset diffusion:(0-) image-conditioned 3d generative models from 2d data. arXiv preprint arXiv:2306.07881 (2023)", + "49. Tang, J.: Stable-dreamfusion: Text-to-3d with stable-diffusion (2022), https://github.com/ashawkey/stable-dreamfusion", + "50. Tang, J., Ren, J., Zhou, H., Liu, Z., Zeng, G.: Dreamgaussian: Generative gaussian splatting for efficient 3d content creation. arXiv preprint arXiv:2309.16653 (2023)", + "51. Tang, J., Wang, T., Zhang, B., Zhang, T., Yi, R., Ma, L., Chen, D.: Make-it-3d: High-fidelity 3d creation from a single image with diffusion prior. In: ICCV (2023)", + "52. Tsalicoglou, C., Manhardt, F., Tonioni, A., Niemeyer, M., Tombari, F.: Textmesh: Generation of realistic 3d meshes from text prompts. arXiv preprint arXiv:2304.12439 (2023)" + ], + "bbox": [ + 212, + 146, + 787, + 839 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "Repaint123: Progressive Controllable Repainting for 3D Generation", + "bbox": [ + 282, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "53. Wang, H., Du, X., Li, J., Yeh, R.A., Shakhnarovich, G.: Score jacobian chaining: Lifting pretrained 2d diffusion models for 3d generation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12619-12629 (2023)", + "54. Wang, T., Kanakis, M., Schindler, K., Van Gool, L., Obukhov, A.: Breathing new life into 3d assets with generative repainting. arXiv preprint arXiv:2309.08523 (2023)", + "55. Wang, Z., Lu, C., Wang, Y., Bao, F., Li, C., Su, H., Zhu, J.: Prolificdreamer: High-fidelity and diverse text-to-3d generation with variational score distillation. arXiv preprint arXiv:2305.16213 (2023)", + "56. Wu, J., Gao, X., Liu, X., Shen, Z., Zhao, C., Feng, H., Liu, J., Ding, E.: Hd-fusion: Detailed text-to-3d generation leveraging multiple noise estimation. arXiv preprint arXiv:2307.16183 (2023)", + "57. Xu, D., Jiang, Y., Wang, P., Fan, Z., Wang, Y., Wang, Z.: Neurallift-360: Lifting an in-the-wild 2d photo to a 3d object with 360 views. arXiv e-prints pp. arXiv-2211 (2022)", + "58. Ye, H., Zhang, J., Liu, S., Han, X., Yang, W.: Ip-adapter: Text compatible image prompt adapter for text-to-image diffusion models (2023)", + "59. Yi, T., Fang, J., Wu, G., Xie, L., Zhang, X., Liu, W., Tian, Q., Wang, X.: Gaussian dreamer: Fast generation from text to 3d gaussian splatting with point cloud priors. arXiv preprint arXiv:2310.08529 (2023)", + "60. Yu, C., Zhou, Q., Li, J., Zhang, Z., Wang, Z., Wang, F.: Points-to-3d: Bridging the gap between sparse points and shape-controllable text-to-3d generation. arXiv preprint arXiv:2307.13908 (2023)", + "61. Yu, W., Yuan, L., Cao, Y.P., Gao, X., Li, X., Quan, L., Shan, Y., Tian, Y.: Hifi-123: Towards high-fidelity one image to 3d content generation. arXiv preprint arXiv:2310.06744 (2023)", + "62. Zhang, L., Rao, A., Agrawala, M.: Adding conditional control to text-to-image diffusion models. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 3836-3847 (2023)", + "63. Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual metric. In: CVPR (2018)", + "64. Zhu, J., Zhuang, P.: Hifa: High-fidelity text-to-3d with advanced diffusion guidance. arXiv preprint arXiv:2305.18766 (2023)" + ], + "bbox": [ + 215, + 146, + 785, + 619 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "Zhang et al.", + "bbox": [ + 271, + 114, + 354, + 128 + ], + "page_idx": 17 + } +] \ No newline at end of file diff --git a/2024/Repaint123_ Fast and High-quality One Image to 3D Generation with Progressive Controllable Repainting/a8257f3f-735a-49a4-ab19-4aa6628f29e5_model.json b/2024/Repaint123_ Fast and High-quality One Image to 3D Generation with Progressive Controllable Repainting/a8257f3f-735a-49a4-ab19-4aa6628f29e5_model.json new file mode 100644 index 0000000000000000000000000000000000000000..d5e1de6e2f54797ca04b1f105f5d30ff1f553904 --- /dev/null +++ b/2024/Repaint123_ Fast and High-quality One Image to 3D Generation with Progressive Controllable Repainting/a8257f3f-735a-49a4-ab19-4aa6628f29e5_model.json @@ -0,0 +1,2315 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.223, + 0.141, + 0.782, + 0.21 + ], + "angle": 0, + "content": "Repaint123: Fast and High-quality One Image to 3D Generation with Progressive Controllable Repainting" + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.235, + 0.782, + 0.298 + ], + "angle": 0, + "content": "Junwu Zhang1* Zhenyu Tang1* Yatian Pang1,3 Xinhua Cheng1 Peng Jin1 Yida Wei4 Xing Zhou5 Munan Ning1,2 Li Yuan1,2* \n1Peking University 2Pengcheng Laboratory 3National University of Singapore \n45RabbitPre" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.34, + 0.785, + 0.525 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.535, + 0.788, + 0.592 + ], + "angle": 0, + "content": "Fig. 1: Repaint123 generates high-quality 3D content with detailed texture from a single image in 2 minutes. Repaint123 adopts 3D Gaussian Splatting in the coarse stage and then utilizes a progressively repairing strategy with the diffusion model for high-quality 3D content generation with efficiency." + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.648, + 0.744, + 0.803 + ], + "angle": 0, + "content": "Abstract. Recent image-to-3D methods achieve impressive results with plausible 3D geometry due to the development of diffusion models and optimization techniques. However, existing image-to-3D methods suffer from texture deficiencies in novel views, including multi-view inconsistency and quality degradation. To alleviate multi-view bias and enhance image quality in novel-view textures, we present Repaint123, a fast image-to-3D approach for creating high-quality 3D content with detailed textures. Repaint123 proposes a progressively repairing strategy to simultaneously enhance the consistency and quality of textures across different views, generating invisible regions according to visible textures, with the visibility map calculated by the depth alignment across views." + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.812, + 0.365, + 0.825 + ], + "angle": 0, + "content": "* Equal contribution." + }, + { + "type": "page_footnote", + "bbox": [ + 0.217, + 0.826, + 0.384, + 0.84 + ], + "angle": 0, + "content": "\\*\\* Corresponding author." + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.812, + 0.384, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.355, + 0.129 + ], + "angle": 0, + "content": "Zhang et al." + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.148, + 0.744, + 0.3 + ], + "angle": 0, + "content": "Furthermore, multiple control techniques, including reference-driven information injection and coarse-based depth guidance, are introduced to alleviate the texture bias accumulated during the repainting process for improved consistency and quality. For novel-view texture refinement with short-term view consistency, our method progressively repairs novel-view images with adaptive strengths based on visibility, enhancing the balance of image quality and view consistency. To alleviate the accumulated bias as progressively repainting, we control the repainting process by depth-guided geometry and attention-driven reference-view textures. Extensive experiments demonstrate the superior ability of our method to create 3D content with consistent and detailed textures in 2 minutes." + }, + { + "type": "text", + "bbox": [ + 0.263, + 0.314, + 0.711, + 0.328 + ], + "angle": 0, + "content": "Keywords: Image-to-3D \\(\\cdot\\) Texture Enhancement \\(\\cdot\\) Diffusion model" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.355, + 0.375, + 0.37 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.387, + 0.489, + 0.749 + ], + "angle": 0, + "content": "Generating 3D content from one given reference image plays a key role at the intersection of computer vision and computer graphics [11, 17, 24, 25, 33, 35], which is desired by users for innovative applications across fields including robotics, virtual reality, and augmented reality. However, the image-to-3D task is quite challenging due to the generated 3D content is expected with reasonable geometry and consistent textures, while the single image input is insufficient in multi-view information. Recent studies [24, 28, 35, 50, 51] employ 2D diffusion models [14, 39] to guide 3D generation from the reference image with Score Distillation Sampling (SDS) [34]. While preliminary attempts [28, 51] leveraging view-independent diffusion models (e.g. Stable Diffusion [39]) faced challenges with over-saturated textures and multi-face geometry, more" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.841 + ], + "angle": 0, + "content": "recent efforts [24, 26, 35, 46, 50] utilize view-conditioned diffusion models [24, 26, 46]. Although these methods utilizing view-conditioned diffusion models mitigate the multi-face problem and generate 3D content with plausible geometry, they still encounter challenges in generating high-quality textures. This limitation stems from training these models on small-scale synthetic 3D datasets, hindering their ability to produce detailed images from novel views." + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.419, + 0.789, + 0.601 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.497, + 0.61, + 0.788, + 0.722 + ], + "angle": 0, + "content": "Fig.2: Motivation. Current image-to-3D methods adopt SDS loss for generating, resulting in inconsistent and poor textures. Repaint123 proposes a progressively repainting strategy based on visible content to efficiently create invisible content with MSE loss, achieving consistent and high-quality textures." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.283, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Repaint123: Progressive Controllable Repainting for 3D Generation" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.147, + 0.786, + 0.433 + ], + "angle": 0, + "content": "To address textural deficiencies including inconsistency and low quality mentioned above, we propose a novel approach named Repaint123 for fast and high-quality image-to-3d generation by generating consistent and high-quality novel-view images with a controllable content repaint process. As shown in Figure 2, the core of Repaint123 is to progressively repaint the invisible content with a view-independent diffusion model for adjacent-view consistency, incorporating multiple controls of geometry and texture for enhanced quality and multi-view consistency. Specifically, our method adopts a two-stage optimization framework. In the coarse stage, we utilize 3D Gaussian Splatting [18] as the representation to efficiently obtain a coarse 3D model in 1 minute. In the fine stage, we progressively sample the camera views around and repaint invisible content with control techniques including using novel-view depth map as geometry prior, injecting reference-view attention features as texture prior [61] and employing reference-view image prompt as semantic prior. To further enhance image quality and preserve multi-view consistency, we propose a visibility-based adaptive strategy that refines the visible regions with different strengths. After achieving high-qualityainted images with multi-view consistency, we directly leverage simple MSE loss to efficiently refine 3D content textures instead of the time-consuming and over-smoothed SDS." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.434, + 0.784, + 0.525 + ], + "angle": 0, + "content": "We conduct extensive experiments on multiple datasets and demonstrate that our method significantly advances image-to-3D generation, producing high-quality, multi-view consistent 3D content with detailed textures in approximately 2 minutes from scratch (as shown in Figure 1), outperforming current state-of-the-art NeRF-based and Gaussian-Splatting-based methods in both consistency and texture quality. Our contributions can be summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.534, + 0.784, + 0.579 + ], + "angle": 0, + "content": "- We propose a progressively repairing strategy based on visibility for the image-to-3D task to generate 3D content with consistency and high quality by repairing invisible parts according to visible textures." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.58, + 0.785, + 0.624 + ], + "angle": 0, + "content": "- We introduce multiple control techniques including reference-driven information injection and coarse-based depth guidance to alleviate the texture bias accumulated during the repainting process for further quality enhancement." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.625, + 0.784, + 0.653 + ], + "angle": 0, + "content": "- Our experiments demonstrate Repaint123 efficiently generates high-quality 3D content with consistent and detailed textures in 2 minutes." + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.534, + 0.785, + 0.653 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.678, + 0.396, + 0.693 + ], + "angle": 0, + "content": "2 Related Works" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.71, + 0.559, + 0.724 + ], + "angle": 0, + "content": "2.1 Diffusion Models for 3D Generation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.734, + 0.785, + 0.84 + ], + "angle": 0, + "content": "The recent notable achievements in view-independent 2D diffusion models [14,39] have brought about exciting prospects for generating 3D objects. Pioneering studies [34, 53] have introduced the concept of distilling a 2D text-to-image generation model to generate 3D shapes. Subsequent works [1, 6, 7, 10, 16, 22, 28, 35, 37, 42-44, 51, 52, 55-57, 60, 61, 64] have adopted a similar per-object optimization approach, building upon these initial works. Nevertheless, the majority of these techniques consistently experience low efficiency and multi-face" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.355, + 0.129 + ], + "angle": 0, + "content": "Zhang et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.314 + ], + "angle": 0, + "content": "issues. Unlike the previous study HiFi-123 [61], which employed similar inversion and attention injection techniques for image-to-3D generation, our approach diverges primarily in our proposed repairing strategy, as well as many variations in optimization objectives, image prompting, and 3D representations. Recently, some works [23, 25, 26, 48] extend view-independent diffusion model to view-conditioned diffusion model to generate multi-view images for reconstruction, while these methods usually suffer from low-quality textures as the multi-view diffusion models are trained on limited and synthesized data. To optimize both consistency and quality, we integrate a view-conditioned diffusion model for initial consistency in the coarse stage and a view-independent model to enhance texture quality in the fine stage." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.336, + 0.504, + 0.352 + ], + "angle": 0, + "content": "2.2 Controllable Image Synthesis" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.362, + 0.788, + 0.513 + ], + "angle": 0, + "content": "A major challenge in the field of image generation is achieving controllability. Many works have been done recently to increase the controllability of images generated by diffusion models. ControlNet [62] and T2I-Adapter [31] attempt to control the creation of images by utilizing data from different modalities. Besides, some optimization-based methods [12,30,40] learn new parameters or fine-tune the diffusion model in order to enhance control over the generation process. Other methods [3,58] leverage multi-view attention to introduce information from other-view images for gaining superior control. We incorporate multiple control techniques into the view-independent diffusion model for maintaining view consistency in the fine stage." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.536, + 0.427, + 0.551 + ], + "angle": 0, + "content": "2.3 3D Representations" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.561, + 0.788, + 0.699 + ], + "angle": 0, + "content": "Neural Radiance Fields (NeRF) [29], as a volumetric rendering method, has gained popularity for its ability to enable 3D optimization [2,4,8,13,21] under 2D supervision, while NeRF optimization can be time-consuming. Numerous efforts [32,41] for spatial pruning have been dedicated to accelerating the training process of NeRF on the reconstruction setting. However, they fail in the generation setting. Recently, 3D Gaussian Splatting [5,9,18,50,59] has emerged as an alternative 3D representation to NeRF and has shown remarkable advancements in terms of both quality and speed, offering a promising avenue. Therefore, we adopt Gaussian Splatting for efficiently generating coarse 3D model." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.722, + 0.371, + 0.74 + ], + "angle": 0, + "content": "3 Preliminary" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.754, + 0.402, + 0.769 + ], + "angle": 0, + "content": "3.1 DDIM Inversion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.789, + 0.843 + ], + "angle": 0, + "content": "DDIM [47] transforms random noise \\(\\pmb{x}_T\\) into clean data \\(\\pmb{x}_0\\) over a series of time steps, by using the deterministic DDIM sampling in the reverse process, i.e., \\(\\pmb{x}_{t - 1} = (\\alpha_{t - 1} / \\alpha_t)(\\pmb{x}_t - \\sigma_t\\epsilon_\\phi) + \\sigma_{t - 1}\\epsilon_\\phi\\). On the contrary, DDIM inversion progressively converts clean data to a noisy state \\(\\pmb{x}_T\\), i.e., \\(\\pmb{x}_t = (\\alpha_t / \\alpha_{t - 1})(\\pmb{x}_{t - 1} - \\sigma_t\\epsilon_\\phi)\\)." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.283, + 0.115, + 0.732, + 0.13 + ], + "angle": 0, + "content": "Repaint123: Progressive Controllable Repainting for 3D Generation" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "image", + "bbox": [ + 0.219, + 0.15, + 0.784, + 0.297 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.31, + 0.788, + 0.409 + ], + "angle": 0, + "content": "Fig. 3: Overview of Repaint123. In the coarse stage, we leverage 3D Gaussian Splatting to represent the coarse content, which is optimized by MSE loss at the reference and SDS loss at the invisible views. In the fine stage, we convert coarse content to mesh and progressively sample and generate invisible views from the reference view bidirectionally. Concretely, the current view is repainted based on the visibility map from bidirectional neighbor views with control techniques including attention injection and depth guidance, leading to consistent and detailed textures." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.438, + 0.788, + 0.485 + ], + "angle": 0, + "content": "\\(\\sigma_{t-1}\\epsilon_{\\phi}) + \\sigma_t\\epsilon_{\\phi}\\), here \\(\\epsilon_{\\phi}\\) is the predicted noise by the UNet. This method can precisely reconstruct the original clean data while greatly speeding up the process by skipping many intermediate diffusion steps." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.506, + 0.45, + 0.523 + ], + "angle": 0, + "content": "3.2 3D Gaussian Splitting" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.531, + 0.788, + 0.622 + ], + "angle": 0, + "content": "Gaussian Splatting [18] presents a novel method for synthesizing new views and reconstructing 3D scenes, achieving real-time speed. Unlike NeRF, Gaussian Splatting uses a set of anisotropic 3D Gaussians defined by their locations, covariances, colors, and opacities to represent the scene. To compute the color of each pixel \\(\\mathbf{p}\\) in the image, it utilizes a typical neural point-based rendering [19,20], The rendering process is as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.394, + 0.634, + 0.786, + 0.677 + ], + "angle": 0, + "content": "\\[\nC (\\mathbf {p}) = \\sum_ {i \\in \\mathcal {N}} c _ {i} \\alpha_ {i} \\prod_ {j = 1} ^ {i - 1} (1 - \\alpha_ {j}), \\tag {1}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.373, + 0.679, + 0.631, + 0.699 + ], + "angle": 0, + "content": "\\[\n\\mathrm {w h e r e}, \\alpha_ {i} = o _ {i} e ^ {- \\frac {1}{2} (\\mathbf {p} - \\boldsymbol {\\mu} _ {i}) ^ {T} \\boldsymbol {\\Sigma} _ {i} ^ {- 1} (\\mathbf {p} - \\boldsymbol {\\mu} _ {i})},\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.71, + 0.788, + 0.74 + ], + "angle": 0, + "content": "where \\(c_{i}\\), \\(o_{i}\\), \\(\\mu_{i}\\), and \\(\\Sigma_{i}\\) represent the color, opacity, position, and covariance of the \\(i\\)-th Gaussian respectively, and \\(\\mathcal{N}\\) denotes the number of related Gaussians." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.763, + 0.331, + 0.779 + ], + "angle": 0, + "content": "4 Method" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.787, + 0.843 + ], + "angle": 0, + "content": "In this section, we introduce our two-stage framework named Repaint123 for fast and high-quality image-to-3D generation, as illustrated in Figure 3. In the coarse stage, we adopt 3D Gaussian Splitting [18] representation optimized by" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.355, + 0.129 + ], + "angle": 0, + "content": "Zhang et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.375 + ], + "angle": 0, + "content": "SDS loss [34] with the view-conditioned diffusion model [24] to learn a coarse geometry and texture (see Section 4.1). Inspired by [50], we convert the coarse 3D Gaussians to textured mesh to facilitate subsequent texture enhancement in UV space. In the fine stage, we progressively paint invisible content in novel views by rotating camera viewpoints bi-directionally from the reference view (see Section 4.2). To mitigate texture bias accumulated during the gradual painting across wide angles, we incorporate multiple control techniques including depth guidance, injection of reference-view attention, and the use of reference-view image prompts (see Section 4.2). Additionally, we propose a visibility-based adaptive strategy to refine visible regions when observed from a superior perspective (see Section 4.2). After obtaining a high-quality painted image that remains consistent across different views, we efficiently optimize the texture map by simple Mean Square Error (MSE) loss (see Section 4.2). This process is repeated for each sampled view, alternating between painting and optimization, until the texture is completely reconstructed to accommodate a 360-degree perspective." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.396, + 0.738, + 0.412 + ], + "angle": 0, + "content": "4.1 Coarse Stage: Gaussian Splatting with 3D Diffusion Prior" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.42, + 0.788, + 0.533 + ], + "angle": 0, + "content": "In the coarse stage, we adopt 3D Gaussian Splatting for efficient initialization. Through SDS [34], 3D diffusion priors are back-propagated to the 3D Gaussians. At each diffusion step, we sample a random current camera view \\( v_{c} \\) and its camera pose \\( p_{c} \\), which orbits around the center of the object. Subsequently, the RGB image, denoted as \\( I_{\\mathrm{rgb}}^{c} \\), and the alpha image, represented by \\( I_{\\mathrm{a}}^{c} \\), are rendered from the current viewpoint \\( p \\). Given an image \\( \\tilde{I}_{\\mathrm{rgb}}^{r} \\) and an alpha mask \\( \\tilde{I}_{\\mathrm{a}}^{r} \\) as input to the reference view, the SDS loss can be formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.335, + 0.544, + 0.786, + 0.579 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {S D S}} = \\mathbb {E} _ {t, c, \\epsilon} \\left[ (\\epsilon_ {\\phi} (I _ {\\mathrm {r g b}} ^ {c}; t, \\tilde {I} _ {\\mathrm {r g b}} ^ {r}, \\Delta p _ {c}) - \\epsilon) \\frac {\\partial I _ {\\mathrm {r g b}} ^ {c}}{\\partial \\Theta} \\right] \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.587, + 0.788, + 0.649 + ], + "angle": 0, + "content": "where \\(\\epsilon_{\\phi}(\\cdot)\\) is the predicted noise by the diffusion \\(\\phi\\), \\(\\Delta p_{c}\\) is the relative camera pose and \\(\\Theta\\) represents the estimizable Gaussian parameters. Additionally, we optimize the RGB image and alpha image rendered from the reference view to align with the input reference image using reconstruction loss:" + }, + { + "type": "equation", + "bbox": [ + 0.359, + 0.658, + 0.787, + 0.678 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {R e f}} = \\lambda_ {\\mathrm {r g b}} | | I _ {\\mathrm {r g b}} ^ {r} - \\tilde {I} _ {\\mathrm {r g b}} ^ {r} | | _ {2} ^ {2} + | | I _ {\\mathrm {a}} ^ {r} - \\tilde {I} _ {\\mathrm {a}} ^ {r} | | _ {2} ^ {2}. \\qquad (3)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.687, + 0.66, + 0.703 + ], + "angle": 0, + "content": "The final loss function is the weighted sum of \\(\\mathcal{L}_{\\mathrm{SDS}}\\) and \\(\\mathcal{L}_{\\mathrm{Ref}}\\)." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.725, + 0.694, + 0.74 + ], + "angle": 0, + "content": "4.2 Fine Stage: Visibility-based Controllable Repainting" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.843 + ], + "angle": 0, + "content": "Progressive Invisible Content Repainting. We repaint previously unseen content with progressively sampling novel camera viewpoints, beginning from the reference view and increasing a certain angular interval with each iteration. To obtain invisible regions in the image from a newly sampled view \\( v_{c} \\), we use the back-projection technique based on the depth map \\( D^{p}\\in \\mathbb{R}^{H\\times W} \\) in the previously seen view \\( v_{p} \\). Formally, given the camera intrinsics \\( K\\in \\mathbb{R}^{3\\times 3} \\), we first remap the" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.283, + 0.115, + 0.733, + 0.131 + ], + "angle": 0, + "content": "Repaint123: Progressive Controllable Repainting for 3D Generation" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.147, + 0.788, + 0.21 + ], + "angle": 0, + "content": "pixels of the image rendered from \\( v_{p} \\) into 3D points \\( X^{p} \\) expressed in the camera coordinate frame, which can be obtained as \\( X_{i,j} = K^{-1}[iD_{i,j},jD_{i,j},D_{i,j}]^{\\top} \\). With \\( P_{p},P_{c}\\in \\mathbb{R}^{3\\times 4} \\) the world-to-camera poses for images from \\( v_{p} \\) and \\( v_{c} \\), we can transform \\( X^{p} \\) into camera coordinate frame of \\( v_{c} \\) to obtain \\( X^{p,c} \\):" + }, + { + "type": "equation", + "bbox": [ + 0.428, + 0.222, + 0.786, + 0.24 + ], + "angle": 0, + "content": "\\[\nX ^ {p, c} = P _ {p} P _ {c} ^ {- 1} h \\left(X ^ {p}\\right), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.251, + 0.788, + 0.327 + ], + "angle": 0, + "content": "where \\( h:(x,y,z)\\to (x,y,z,1) \\) is the homogeneous mapping. Then we can render a depth map \\( D^{p,c} \\) from \\( X^{p,c} \\) as \\( D_{i,j}^{p,c} = KX_{i,j,2}^{p,c} \\). Regions with dissimilar depth values between \\( D^c \\) and \\( D^{p,c} \\) are recognized as invisible regions, forming a repainting mask \\( M \\) where areas with zeros are invisible and areas with ones are visible. Therefore, we repaint invisible content and maintain visible content by:" + }, + { + "type": "equation", + "bbox": [ + 0.386, + 0.339, + 0.786, + 0.357 + ], + "angle": 0, + "content": "\\[\nx _ {t} = x _ {t} ^ {\\mathrm {i n v}} \\odot M + x _ {t} ^ {\\mathrm {r e v}} \\odot (1 - M), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.368, + 0.789, + 0.413 + ], + "angle": 0, + "content": "where \\( \\odot \\) is element-wise matrix multiplication, \\( t \\) is the current denoising timestep, \\( x_{t}^{\\mathrm{inv}} \\) is the DDIM [47] inverted latent, and \\( x_{t}^{\\mathrm{rev}} \\sim \\mathcal{N}(\\mu_{\\phi}(x_{t + 1},t),\\Sigma_{\\phi}(x_{t + 1},t)) \\) is the diffusion model output." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.413, + 0.789, + 0.565 + ], + "angle": 0, + "content": "Additionally, to mitigate accumulated bias for distant camera views during gradual repainting, as shown in the fine stage in Figure 3, we sample camera viewpoints bi-directionally, alternating between clockwise and counterclockwise directions. This bidirectional sampling allows us to paint up to 180 degrees from two opposite directions, instead of painting across the entire 360 degrees from one direction. Moreover, to paint bidirectionally invisible content and ensure smooth transitions at the junction of the two directions, we replace the previous unidirectional painting mask \\( M \\) by a bidirectional painting mask. This mask merges two unidirectional masks from bidirectional neighbor views (red cameras in Figure 3) by taking the maximum value of each pixel." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.588, + 0.789, + 0.801 + ], + "angle": 0, + "content": "Control-based Content Enhancement. While the progressive painting maintains consistency between adjacent views, views that are farther apart may become inconsistent and experience a decline in quality due to cumulative errors. To tackle these issues, we improve the painting process by incorporating various control techniques to enhance the consistency and quality of synthesized images. Specifically, for geometry control, we employ ControlNet [62] to condition the diffusion model with coarse depth maps, ensuring geometric consistency across images. To transfer reference textures and mitigate the cumulative texture bias during the progressive painting, inspired by HiFi-123 [61], we incorporate an attention injection mechanism [3] that injects reference attention features into the diffusion denoising process. By replacing the novel-view content features (Key features \\( K_{t} \\) and Value features \\( V_{t} \\)) with reference-view attention features (\\( K_{r} \\) and \\( V_{r} \\)) during each denoising step, the novel-view image features can directly query the high-quality reference features by:" + }, + { + "type": "equation", + "bbox": [ + 0.337, + 0.81, + 0.787, + 0.845 + ], + "angle": 0, + "content": "\\[\n\\operatorname {A t t e n t i o n} \\left(Q _ {t}, K _ {r}, V _ {r}\\right) = \\operatorname {S o f t m a x} \\left(\\frac {Q _ {t} K _ {r} ^ {T}}{\\sqrt {d}}\\right) V _ {r}, \\tag {6}\n\\]" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.117, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.116, + 0.355, + 0.129 + ], + "angle": 0, + "content": "Zhang et al." + }, + { + "type": "image", + "bbox": [ + 0.221, + 0.149, + 0.785, + 0.334 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.346, + 0.788, + 0.444 + ], + "angle": 0, + "content": "Fig. 4: Visibility-based Controllable Repainting Pipeline. Our method repairs invisible content (black areas in \\( M_t \\)) of novel-view rendered image \\( I_{\\mathrm{rgb}} \\) according to the inverted latent \\( x_t^{inv} \\) and the visibility map \\( V \\) to obtain refined image \\( I_{\\mathrm{rgb}}^{\\mathrm{fine}} \\). We introduce multiple control techniques to enhance the consistency and quality of generated textures, including attention feature injection, depth guidance by ControlNet, and semantic prompt by CLIP encoder. Noticing that the repainting mask \\( M_t \\) is simultaneously changed with denoising timesteps for adaptive refinement." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.472, + 0.785, + 0.517 + ], + "angle": 0, + "content": "where \\( Q_{t} \\) represents the novel-view query features. This enhancement facilitates the transfer of texture details and improves the consistency between the reference-view image and novel-view images." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.518, + 0.788, + 0.624 + ], + "angle": 0, + "content": "During our experiments, we observed a decline in image quality during repainting, likely due to not using text prompts for performing classifier-free guidance [15], crucial for producing high-quality images in diffusion models. To improve this, we adopt IP-Adapter [58], which uses a CLIP encoder to convert reference images into semantic features. These features are then projected into image prompts, providing visual cues that enhance image generation through classifier-free guidance." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.644, + 0.789, + 0.842 + ], + "angle": 0, + "content": "Adaptive Visible Content Refinement. Repainting only the invisible content often leads to distortion in the visible area As shown in Figure 5, when the previous view is an oblique view, it leads to a low-resolution update on the texture maps, resulting in high distortion when rendering from a better observing view. Therefore, the visible content currently observed in a superior view should be refined to replenish details. However, the selection of refinement strength for these visible content is tricky [38, 54], as excessive strength produces unfaithful results leading to inconsistency with previously repainted views, while insufficient strength limits quality improvement. We propose a visibility-based adaptive refinement strategy to refine these previously seen regions with different strengths, aiming to achieve improved quality-consistency trade-off. As visualized in Figure 5, we view painting for visible content as a process similar to superresolution that replenishes detailed information. According to the Orthographic" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.283, + 0.115, + 0.731, + 0.129 + ], + "angle": 0, + "content": "Repaint123: Progressive Controllable Repainting for 3D Generation" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.488, + 0.433 + ], + "angle": 0, + "content": "Projection Theorem, which asserts that the projected resolution of a surface is directly proportional to the view obliqueness (\\( \\cos\\theta \\) as shown in Figure 5), we can assume that the repainting strength is equal to \\( (1 - \\cos\\theta_{1} / \\cos\\theta_{2}) \\), where \\( \\cos\\theta_{2} / \\cos\\theta_{1} \\) represents the upsampling scale in terms of super-resolution and visibility variations in the context of 3D rendering. Therefore, we can define a visibility map \\( V \\) as \\( V_{i,j} = \\cos\\theta_{i,j}^{p,c} / \\cos\\theta_{i,j}^{c} \\), where \\( \\cos\\theta_{i,j}^{p,c} \\) and \\( \\cos\\theta_{i,j}^{c} \\) are obtained by multiplying the current-view normal map with camera rays of the previous view \\( v_{p} \\) and the current view \\( v_{c} \\) respectively. Higher visibility value indicates less repainting strength, while invisible areas require full strength." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.434, + 0.788, + 0.495 + ], + "angle": 0, + "content": "By this association between repairing strength and the visibility map \\( V \\), we can binarize the visibility map to the timestep-aware adaptive repairing mask \\( M_{t} \\) during each denoising step, visualized in the green box \"Timestep-aware binarization\" in Figure 4:" + }, + { + "type": "equation", + "bbox": [ + 0.389, + 0.512, + 0.787, + 0.553 + ], + "angle": 0, + "content": "\\[\nM _ {t} ^ {i, j} = \\left\\{ \\begin{array}{l l} 1, & \\text {i f} V ^ {i, j} > 1 - t / T \\\\ 0, & \\text {e l s e}, \\end{array} \\right. \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.568, + 0.788, + 0.614 + ], + "angle": 0, + "content": "where \\( i \\), and \\( j \\) are the 2D position of pixels in visibility map \\( V \\), and \\( T \\) is the total number of timesteps of the diffusion model. By doing this, we can adaptively refine visible content with a proper strength." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.652, + 0.788, + 0.713 + ], + "angle": 0, + "content": "Efficient Optimization. Using a progressive painting strategy and various control techniques to refine the inverted latent, we obtain a high-quality, view-consistent refined image, \\( I_{\\mathrm{rgb}}^{\\mathrm{fine}} \\). This process allows us to use simple MSE loss for efficient texture map optimization:" + }, + { + "type": "equation", + "bbox": [ + 0.418, + 0.729, + 0.787, + 0.749 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {f i n e}} = \\left| \\left| I _ {\\mathrm {r g b}} ^ {\\text {f i n e}} - I _ {\\mathrm {r g b}} \\right| \\right| _ {2} ^ {2}, \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.788, + 0.843 + ], + "angle": 0, + "content": "where \\( I_{\\mathrm{rgb}} \\) represents the rendered image in the fine stage. MSE loss is faster and more deterministic to optimize than the traditional SDS loss, speeding up the optimization process. After sampling a novel view, we first conduct image repainting and then proceed with efficient texture optimization, repeating this process until we cover a 360-degree view range." + }, + { + "type": "image", + "bbox": [ + 0.51, + 0.151, + 0.779, + 0.319 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.497, + 0.329, + 0.788, + 0.399 + ], + "angle": 0, + "content": "Fig. 5: Relation between view obliqueness and refinement strength. Red box shows the same texture part in different views. Texture details in generated views might be insufficient due to the variation of angles." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.355, + 0.129 + ], + "angle": 0, + "content": "Zhang et al." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.145, + 0.367, + 0.163 + ], + "angle": 0, + "content": "5 Experiment" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.184, + 0.457, + 0.199 + ], + "angle": 0, + "content": "5.1 Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.217, + 0.787, + 0.337 + ], + "angle": 0, + "content": "In our experiment, we consistently apply the same hyperparameters across all our method's results. We progressively turn around the viewpoints by 40 degrees each time, consequently obtaining 8 views per object. We use a 50-step DDIM schedule and perform a 30-step latent diffusion inversion. Stable Diffusion 1.5 is utilized for all methods. During the coarse stage training, we set \\(\\lambda_{\\mathrm{rgb}}\\) to 10. Generating a single 3D object takes only 2 minutes on a single 40G A100 GPU, approximately 1 minute for the coarse stage and mesh extraction, and another 1 minute for refinement." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.367, + 0.339, + 0.383 + ], + "angle": 0, + "content": "5.2 Baselines" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.399, + 0.787, + 0.567 + ], + "angle": 0, + "content": "We adopt RealFusion [28], Make-It-3D [51], and Zero123-XL [24], Magic123 [35] as our NeRF-based baselines and DreamGaussian [50] as our Gaussian-Splatting-based baseline. RealFusion presents a single-stage algorithm for NeRF generation leveraging 2D SDS loss for novel views. Make-It-3D is a two-stage approach that shares similar objectives with RealFusion but employs a point cloud representation for refinement at the second stage. For Zero123-XL, we adopt the implementation [49] and add a mesh fine-tuning stage for fair comparison. Integrating Zero123 and RealFusion, Magic123 incorporates a 2D SDS loss with 3D SDS loss provided by Zero123 to balance geometry and texture and adopts DMTet [45] representation at the second stage. DreamGaussian integrates 3D Gaussian Splatting into 3D generation and greatly improves the speed." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.596, + 0.427, + 0.611 + ], + "angle": 0, + "content": "5.3 Evaluation Protocol" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.628, + 0.787, + 0.674 + ], + "angle": 0, + "content": "Datasets. Based on previous research, we utilized the Realfusion15 dataset [28] and test-alpha dataset collected by Make-It-3D [51], which comprises many common objects from diverse styles." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.704, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Evaluation metrics. An effective 3D generation approach should produce 3D content which closely resemble the reference view, and maintain consistency of semantics and textures with the reference image when observed from new views. Therefore, to evaluate the overall quality of the generated 3D object, we choose the following metrics from two aspects: 1) PSNR and LPIPS [63], which measure pixel-level and perceptual reconstruction quality respectively at the reference view; 2) CLIP Similarity [36] and Contextual Distance [27], which assess the similarity of semantics and detailed textures respectively between the novel perspective and the reference view." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.283, + 0.115, + 0.731, + 0.129 + ], + "angle": 0, + "content": "Repaint123: Progressive Controllable Repainting for 3D Generation" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "image", + "bbox": [ + 0.242, + 0.158, + 0.764, + 0.778 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.234, + 0.79, + 0.767, + 0.807 + ], + "angle": 0, + "content": "Fig. 6: Qualitative comparisons on image-to-3D generation. Zoom in for details." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.355, + 0.129 + ], + "angle": 0, + "content": "Zhang et al." + }, + { + "type": "table", + "bbox": [ + 0.217, + 0.144, + 0.784, + 0.286 + ], + "angle": 0, + "content": "
DatasetMetrics\\MethodsNeRF-basedGaussian-Splatting-based
RealFusionMake-it-3DZero-123-XL*Magic123DreamGaussianRepaint123
RealFusion15CLIP-Similarity↑0.710.810.830.820.770.85
Context-Dis↓2.201.821.591.641.611.55
PSNR↑19.2416.5619.5619.6818.9419.00
LPIPS↓0.1940.1770.1080.1070.1110.101
Test-alphaCLIP-Similarity↑0.680.760.840.840.790.88
Context-Dis↓2.201.731.521.571.621.50
PSNR↑22.9117.2124.3924.6922.3322.38
LPIPS↓0.1050.2370.0500.0460.0570.048
Optimization time20min1h30min1h (+2h)2min2 min
" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.288, + 0.788, + 0.373 + ], + "angle": 0, + "content": "Table 1: We show quantitative results in terms of CLIP-Similarity↑ / Contextual-Distance↓ / PSNR↑ / LPIPS↓. The results are shown on the RealFusion15 and test-alpha datasets, while bold reflects the best for all methods and the underline represents the best for Gaussian-Splatting-based methods. * indicates that Zero123-XL incorporates a mesh fine-tuning stage for further quality improvement. The time required by textual inversion is indicated in parentheses." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.414, + 0.371, + 0.429 + ], + "angle": 0, + "content": "5.4 Comparisons" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.439, + 0.788, + 0.683 + ], + "angle": 0, + "content": "Comparisons with NeRF-based Methods. As shown in Table 1, we evaluate the quality of generated 3D objects across various NeRF-based methods. Our method achieves superior 3D consistency in generating 3D objects, as evidenced by the best performance of CLIP-similarity and Contextual-distance. Regarding reference-view reconstruction quality, there is a gap compared with NeRF-based approaches as shown in Table 1, which we attribute to the immaturity of current Gaussian-Splatting-based methods. Compared to NeRF-based methods for the optimization time, our approach reaches a significant acceleration of over 10 times and simultaneously achieves high quality, due to the 3D Gaussian-Splatting representation in coarse stage and efficient texture refinement in fine stage. As shown in Figure 6, Repaint123 achieves the best visual results in terms of texture consistency and generation quality as opposed to other NeRF-based methods. From the visual comparison, our method achieves consistent and detailed textures in invisible areas, while Zero123-XL results in over-smooth textures, Magic123 produces inconsistent and oversaturated colors, Realfusion and Make-It-3D fail to generate full geometry and consistent textures." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.704, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Comparisons with GS-based methods. We conduct comparisons with other works based on 3D Gaussian Splatting in the last column of Table 1. Within a comparable generation time, our method demonstrates superiority over existing Gaussian-based approach in aspects of texture consistency and the quality of reference-view reconstruction. The superiority of our proposed Repaint123 is evidenced by the evaluation of four distinct metrics from the Table 1. As shown in Figure 6, DreamGaussian usually leads to over-smooth texture inconsistent with reference view, while our method can produce high-quality 3D content with view-consistent and detailed textures." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.283, + 0.115, + 0.731, + 0.129 + ], + "angle": 0, + "content": "Repaint123: Progressive Controllable Repainting for 3D Generation" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "image", + "bbox": [ + 0.251, + 0.148, + 0.756, + 0.453 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.463, + 0.788, + 0.534 + ], + "angle": 0, + "content": "Fig. 7: Qualitative ablation study by removing one component at a time from the overall method. The presence of artifacts is highlighted by red boxes. Inconsistencies such as multi-face problems and mismatches in content and style are evident without using repainting and attention injection strategies. Absence of image prompting and adaptive refinement notably degrades quality." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.565, + 0.446, + 0.58 + ], + "angle": 0, + "content": "5.5 Ablation and Analysis" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.596, + 0.787, + 0.626 + ], + "angle": 0, + "content": "This section details both qualitative and quantitative analyses to highlight the effectiveness of our proposed methods, as illustrated in Figure 7 and Table 2." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.628, + 0.788, + 0.734 + ], + "angle": 0, + "content": "Effectiveness of Repainting. Figure 7 illustrates that without repainting strategy, the generated novel-view image tends to resemble the reference image, leading to inconsistencies in the shared visible areas between different views, such as the content misalignment (observed in the Ice Cream example) and issues with multiple faces (evident in the Anya example). These issues stem from the absence of alignment constraints for shared visible regions, leading to conflicts and quality degradation of the reconstructed 3D texture." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Impact of Attention Injection. Table 2 shows that injecting attention features from the reference-view image markedly enhances the consistency of both semantics and fine-grained textures, as evidenced by increased CLIP similarity and reduced Contextual distance. Without this injection strategy, as depicted in Figure 7, the synthesized images retain basic semantics but fail to accurately transfer detailed textures and styles from the reference view, resulting in inconsistencies across multiple views." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.355, + 0.129 + ], + "angle": 0, + "content": "Zhang et al." + }, + { + "type": "table", + "bbox": [ + 0.278, + 0.144, + 0.725, + 0.242 + ], + "angle": 0, + "content": "
Method \\ MetricCLIP↑Contextual↓PSNR↑LPIPS↓
Coarse0.711.7821.170.133
vanilla painting0.711.6222.410.049
+attention injection0.781.5622.420.048
+image prompt0.841.5222.400.048
+adaptive refinemnet (Ours)0.881.5022.380.048
" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.243, + 0.788, + 0.285 + ], + "angle": 0, + "content": "Table 2: Quantitative ablation study on Test-alpha dataset by progressively adding our proposed components. The last three lines show the cumulative effects of the proposed modules. Our approach, utilizing all strategies, delivered the best performance." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.328, + 0.784, + 0.403 + ], + "angle": 0, + "content": "Role of Image Prompt. The introduction of using the reference image as an image prompt, as shown in Table 2 and Figure 7, significantly boosts both multi-view consistency and image quality of the generated images. Without this technique to perform classifier-free guidance, the generation of detailed and consistent textures across views is compromised." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.404, + 0.785, + 0.48 + ], + "angle": 0, + "content": "Benefits of Adaptive Refinement. The necessity of adaptive refinement is clear from Figure 7, where its absence leads to artifacts and blurriness in obliquely viewing areas of previous views due to low-resolution updates, as mentioned in Section 4.2. Table 2 also demonstrates its benefits through improved CLIP similarity and Contextual distance." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.505, + 0.364, + 0.521 + ], + "angle": 0, + "content": "6 Limitations" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.539, + 0.788, + 0.678 + ], + "angle": 0, + "content": "Despite the promising results, our method still has some limitations. Our method is based on 3D Gaussian Splatting representation in the coarse stage during training. While 3D Gaussian Splatting accelerates the training process, it may exhibit geometry artifacts during mesh extraction, such as holes, and inferior results compared to NeRF-based methods for the reconstruction of reference view, due to immaturity in generation tasks. This is supported by results in the Appendix, where NeRF is considered as an alternative to 3D Gaussian Splatting for the coarse stage. We expect these limitations can be mitigated with the development of 3D Gaussian Splatting in the future." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.701, + 0.36, + 0.716 + ], + "angle": 0, + "content": "7 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.734, + 0.788, + 0.841 + ], + "angle": 0, + "content": "This work presents Repaint123 for generating high-quality 3D content from a single image in about 2 minutes. By leveraging progressive controllable paint, our approach overcomes the limitations of existing studies and achieves state-of-the-art results in terms of both texture quality and multi-view consistency, paving the way for future progress in 3D content generation from one image. Furthermore, we validate the effectiveness of our proposed method through a comprehensive set of experiments." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.283, + 0.115, + 0.732, + 0.13 + ], + "angle": 0, + "content": "Repaint123: Progressive Controllable Repainting for 3D Generation" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.403, + 0.163 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.179, + 0.864, + 0.21 + ], + "angle": 0, + "content": "This work was supported in part by Natural Science Foundation of China (No. 62332002, 62202014), and Shenzhen Basic Research Program (No.JCYJ20220813151736001)" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.234, + 0.323, + 0.25 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.267, + 0.786, + 0.309 + ], + "angle": 0, + "content": "1. Armandpour, M., Zheng, H., Sadeghian, A., Sadeghian, A., Zhou, M.: Re-imagine the negative prompt algorithm: Transform 2d diffusion into 3d, alleviate janus problem and beyond. arXiv preprint arXiv:2304.04968 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.311, + 0.786, + 0.337 + ], + "angle": 0, + "content": "2. Barron, J.T., Mildenhall, B., Verbin, D., Srinivasan, P.P., Hedman, P.: Mip-nerf 360: Unbounded anti-aliased neural radiance fields. CVPR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.339, + 0.786, + 0.393 + ], + "angle": 0, + "content": "3. Cao, M., Wang, X., Qi, Z., Shan, Y., Qie, X., Zheng, Y.: Masactrl: Tuning-free mutual self-attention control for consistent image synthesis and editing. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 22560-22570 (October 2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.395, + 0.786, + 0.435 + ], + "angle": 0, + "content": "4. Chan, E.R., Lin, C.Z., Chan, M.A., Nagano, K., Pan, B., Mello, S.D., Gallo, O., Guibas, L., Tremblay, J., Khamis, S., Karras, T., Wetzstein, G.: Efficient geometry-aware 3D generative adversarial networks. In: CVPR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.437, + 0.786, + 0.463 + ], + "angle": 0, + "content": "5. Chen, G., Wang, W.: A survey on 3d gaussian splatting. arXiv preprint arXiv:2401.03890 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.465, + 0.786, + 0.505 + ], + "angle": 0, + "content": "6. Chen, R., Chen, Y., Jiao, N., Jia, K.: *Fantasia3d: Disentangling geometry and appearance for high-quality text-to-3d content creation*. arXiv preprint arXiv:2303.13873 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.507, + 0.786, + 0.546 + ], + "angle": 0, + "content": "7. Chen, Y., Zhang, C., Yang, X., Cai, Z., Yu, G., Yang, L., Lin, G.: It3d: Improved text-to-3d generation with explicit view synthesis. arXiv preprint arXiv:2308.11473 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.548, + 0.786, + 0.589 + ], + "angle": 0, + "content": "8. Chen, Z., Funkhouser, T., Hedman, P., Tagliasacchi, A.: Mobilenerf: Exploiting the polygon rasterization pipeline for efficient neural field rendering on mobile architectures. arXiv preprint arXiv:2208.00277 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.591, + 0.786, + 0.617 + ], + "angle": 0, + "content": "9. Chen, Z., Wang, F., Liu, H.: Text-to-3d using gaussian splatting. arXiv preprint arXiv:2309.16585 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.619, + 0.786, + 0.658 + ], + "angle": 0, + "content": "0. Cheng, X., Yang, T., Wang, J., Li, Y., Zhang, L., Zhang, J., Yuan, L.: Progressive3d: Progressively local editing for text-to-3d content creation with complex semantic prompts. arXiv preprint arXiv:2310.11784 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.66, + 0.786, + 0.714 + ], + "angle": 0, + "content": "1. Dou, Z., Wu, Q., Lin, C., Cao, Z., Wu, Q., Wan, W., Komura, T., Wang, W.: Tore: Token reduction for efficient human mesh recovery with transformer. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 15143-15155 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.716, + 0.786, + 0.756 + ], + "angle": 0, + "content": "2. Gal, R., Alaluf, Y., Atzmon, Y., Patashnik, O., Bermano, A.H., Chechik, G., Cohen-Or, D.: An image is worth one word: Personalizing text-to-image generation using textual inversion. arXiv preprint arXiv:2208.01618 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.758, + 0.786, + 0.784 + ], + "angle": 0, + "content": "3. Hedman, P., Srinivasan, P.P., Mildenhall, B., Barron, J.T., Debevec, P.: Baking neural radiance fields for real-time view synthesis. ICCV (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.786, + 0.786, + 0.812 + ], + "angle": 0, + "content": "4. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. arXiv preprint arxiv:2006.11239 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.814, + 0.786, + 0.84 + ], + "angle": 0, + "content": "5. Ho, J., Salimans, T.: Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598 (2022)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.267, + 0.786, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.354, + 0.129 + ], + "angle": 0, + "content": "Zhang et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.785, + 0.189 + ], + "angle": 0, + "content": "16. Huang, Y., Wang, J., Shi, Y., Qi, X., Zha, Z.J., Zhang, L.: Dreamtime: An improved optimization strategy for text-to-3d content creation. arXiv preprint arXiv:2306.12422 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.189, + 0.785, + 0.217 + ], + "angle": 0, + "content": "17. Jun, H., Nichol, A.: Shap-e: Generating conditional 3d implicit functions. arXiv preprint arXiv:2305.02463 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.217, + 0.785, + 0.257 + ], + "angle": 0, + "content": "18. Kerbl, B., Kopanas, G., Leimkuhler, T., Drettakis, G.: 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics (ToG) 42(4), 1-14 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.257, + 0.785, + 0.298 + ], + "angle": 0, + "content": "19. Kopanas, G., Leimkuhler, T., Rainer, G., Jambon, C., Drettakis, G.: Neural point catacaustics for novel-view synthesis of reflections. ACM Transactions on Graphics (TOG) 41(6), 1-15 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.298, + 0.785, + 0.338 + ], + "angle": 0, + "content": "20. Kopanas, G., Philip, J., Leimkuhler, T., Drettakis, G.: Point-based neural rendering with per-view optimization. In: Computer Graphics Forum. vol. 40, pp. 29-43. Wiley Online Library (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.338, + 0.785, + 0.365 + ], + "angle": 0, + "content": "21. Li, Z., Müller, T., Evans, A., Taylor, R.H., Unberath, M., Liu, M.Y., Lin, C.H.: Neuralangelo: High-fidelity neural surface reconstruction. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.365, + 0.785, + 0.405 + ], + "angle": 0, + "content": "22. Lin, C.H., Gao, J., Tang, L., Takikawa, T., Zeng, X., Huang, X., Kreis, K., Fidler, S., Liu, M.Y., Lin, T.Y.: Magic3d: High-resolution text-to-3d content creation. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.405, + 0.785, + 0.446 + ], + "angle": 0, + "content": "23. Liu, M., Xu, C., Jin, H., Chen, L., Xu, Z., Su, H., et al.: One-2-3-45: Any single image to 3d mesh in 45 seconds without per-shape optimization. arXiv preprint arXiv:2306.16928 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.446, + 0.785, + 0.488 + ], + "angle": 0, + "content": "24. Liu, R., Wu, R., Van Hoorick, B., Tokmakov, P., Zakharov, S., Vondrick, C.: Zero-1-to-3: Zero-shot one image to 3d object. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 9298-9309 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.488, + 0.785, + 0.528 + ], + "angle": 0, + "content": "25. Liu, Y., Lin, C., Zeng, Z., Long, X., Liu, L., Komura, T., Wang, W.: Syncdreamer: Generating multiview-consistent images from a single-view image. arXiv preprint arXiv:2309.03453 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.528, + 0.785, + 0.569 + ], + "angle": 0, + "content": "26. Long, X., Guo, Y.C., Lin, C., Liu, Y., Dou, Z., Liu, L., Ma, Y., Zhang, S.H., Habermann, M., Theobalt, C., et al.: Wonder3d: Single image to 3d using cross-domain diffusion. arXiv preprint arXiv:2310.15008 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.569, + 0.785, + 0.61 + ], + "angle": 0, + "content": "27. Mechrez, R., Talmi, I., Zelnik-Manor, L.: The contextual loss for image transformation with non-aligned data. In: Proceedings of the European conference on computer vision (ECCV). pp. 768-783 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.61, + 0.785, + 0.637 + ], + "angle": 0, + "content": "28. Melas-Kyriazi, L., Laina, I., Rupprecht, C., Vedaldi, A.: Realfusion: 360deg reconstruction of any object from a single image. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.637, + 0.785, + 0.677 + ], + "angle": 0, + "content": "29. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. In: ECCV (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.677, + 0.785, + 0.732 + ], + "angle": 0, + "content": "30. Mokady, R., Hertz, A., Aberman, K., Pritch, Y., Cohen-Or, D.: Null-text inversion for editing real images using guided diffusion models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6038-6047 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.732, + 0.785, + 0.772 + ], + "angle": 0, + "content": "31. Mou, C., Wang, X., Xie, L., Zhang, J., Qi, Z., Shan, Y., Qie, X.: T2i-adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models. arXiv preprint arXiv:2302.08453 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.772, + 0.785, + 0.799 + ], + "angle": 0, + "content": "32. Müller, T., Evans, A., Schied, C., Keller, A.: Instant neural graphics primitives with a multiresolution hash encoding. ACM TOG (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.799, + 0.785, + 0.841 + ], + "angle": 0, + "content": "33. Nichol, A., Jun, H., Dhariwal, P., Mishkin, P., Chen, M.: Point-e: A system for generating 3d point clouds from complex prompts. arXiv preprint arXiv:2212.08751 (2022)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.283, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Repaint123: Progressive Controllable Repainting for 3D Generation" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.175 + ], + "angle": 0, + "content": "34. Poole, B., Jain, A., Barron, J.T., Mildenhall, B.: Dreamfusion: Text-to-3d using 2d diffusion. arXiv preprint arXiv:2209.14988 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.176, + 0.788, + 0.232 + ], + "angle": 0, + "content": "35. Qian, G., Mai, J., Hamdi, A., Ren, J., Siarohin, A., Li, B., Lee, H.Y., Skorokhodov, I., Wonka, P., Tulyakov, S., et al.: Magic123: One image to high-quality 3d object generation using both 2d and 3d diffusion priors. arXiv preprint arXiv:2306.17843 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.233, + 0.787, + 0.286 + ], + "angle": 0, + "content": "36. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.288, + 0.787, + 0.328 + ], + "angle": 0, + "content": "37. Raj, A., Kaza, S., Poole, B., Niemeyer, M., Ruiz, N., Mildenhall, B., Zada, S., Aberman, K., Rubinstein, M., Barron, J., et al.: Dreambooth3d: Subject-driven text-to-3d generation. arXiv preprint arXiv:2303.13508 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.328, + 0.787, + 0.357 + ], + "angle": 0, + "content": "38. Richardson, E., Metzer, G., Alaluf, Y., Giryes, R., Cohen-Or, D.: Texture: Text-guided texturing of 3d shapes. arXiv preprint arXiv:2302.01721 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.357, + 0.787, + 0.398 + ], + "angle": 0, + "content": "39. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10684-10695 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.399, + 0.787, + 0.453 + ], + "angle": 0, + "content": "40. Ruiz, N., Li, Y., Jampani, V., Pritch, Y., Rubinstein, M., Aberman, K.: Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 22500-22510 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.453, + 0.787, + 0.479 + ], + "angle": 0, + "content": "41. Sara Fridovich-Keil and Alex Yu, Tancik, M., Chen, Q., Recht, B., Kanazawa, A.: Plenoxels: Radiance fields without neural networks. In: CVPR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.481, + 0.787, + 0.508 + ], + "angle": 0, + "content": "42. Seo, H., Kim, H., Kim, G., Chun, S.Y.: Ditto-nerf: Diffusion-based iterative text to omni-directional 3d model. arXiv preprint arXiv:2304.02827 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.509, + 0.787, + 0.55 + ], + "angle": 0, + "content": "43. Seo, J., Jang, W., Kwak, M.S., Ko, J., Kim, H., Kim, J., Kim, J.H., Lee, J., Kim, S.: Let 2d diffusion model know 3d-consistency for robust text-to-3d generation. arXiv preprint arXiv:2303.07937 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.55, + 0.787, + 0.577 + ], + "angle": 0, + "content": "44. Shen, Q., Yang, X., Wang, X.: Anything-3d: Towards single-view anything reconstruction in the wild. arXiv preprint arXiv:2304.10261 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.578, + 0.787, + 0.619 + ], + "angle": 0, + "content": "45. Shen, T., Gao, J., Yin, K., Liu, M.Y., Fidler, S.: Deep marching tetrahedra: a hybrid representation for high-resolution 3d shape synthesis. In: Advances in Neural Information Processing Systems (NeurIPS) (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.62, + 0.787, + 0.646 + ], + "angle": 0, + "content": "46. Shi, Y., Wang, P., Ye, J., Long, M., Li, K., Yang, X.: Mvdream: Multi-view diffusion for 3d generation. arXiv preprint arXiv:2308.16512 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.647, + 0.787, + 0.674 + ], + "angle": 0, + "content": "47. Song, J., Meng, C., Ermon, S.: Denoising diffusion implicit models. arXiv:2010.02502 (October 2020), https://arxiv.org/abs/2010.02502" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.675, + 0.787, + 0.715 + ], + "angle": 0, + "content": "48. Szymanowicz, S., Rupprecht, C., Vedaldi, A.: Viewset diffusion:(0-) image-conditioned 3d generative models from 2d data. arXiv preprint arXiv:2306.07881 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.716, + 0.787, + 0.743 + ], + "angle": 0, + "content": "49. Tang, J.: Stable-dreamfusion: Text-to-3d with stable-diffusion (2022), https://github.com/ashawkey/stable-dreamfusion" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.744, + 0.787, + 0.771 + ], + "angle": 0, + "content": "50. Tang, J., Ren, J., Zhou, H., Liu, Z., Zeng, G.: Dreamgaussian: Generative gaussian splatting for efficient 3d content creation. arXiv preprint arXiv:2309.16653 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.772, + 0.787, + 0.799 + ], + "angle": 0, + "content": "51. Tang, J., Wang, T., Zhang, B., Zhang, T., Yi, R., Ma, L., Chen, D.: Make-it-3d: High-fidelity 3d creation from a single image with diffusion prior. In: ICCV (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.799, + 0.787, + 0.84 + ], + "angle": 0, + "content": "52. Tsalicoglou, C., Manhardt, F., Tonioni, A., Niemeyer, M., Tombari, F.: Textmesh: Generation of realistic 3d meshes from text prompts. arXiv preprint arXiv:2304.12439 (2023)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.355, + 0.129 + ], + "angle": 0, + "content": "Zhang et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.787, + 0.204 + ], + "angle": 0, + "content": "53. Wang, H., Du, X., Li, J., Yeh, R.A., Shakhnarovich, G.: Score jacobian chaining: Lifting pretrained 2d diffusion models for 3d generation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12619-12629 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.205, + 0.786, + 0.245 + ], + "angle": 0, + "content": "54. Wang, T., Kanakis, M., Schindler, K., Van Gool, L., Obukhov, A.: Breathing new life into 3d assets with generative repainting. arXiv preprint arXiv:2309.08523 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.246, + 0.787, + 0.286 + ], + "angle": 0, + "content": "55. Wang, Z., Lu, C., Wang, Y., Bao, F., Li, C., Su, H., Zhu, J.: Prolificdreamer: High-fidelity and diverse text-to-3d generation with variational score distillation. arXiv preprint arXiv:2305.16213 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.287, + 0.787, + 0.327 + ], + "angle": 0, + "content": "56. Wu, J., Gao, X., Liu, X., Shen, Z., Zhao, C., Feng, H., Liu, J., Ding, E.: Hd-fusion: Detailed text-to-3d generation leveraging multiple noise estimation. arXiv preprint arXiv:2307.16183 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.328, + 0.787, + 0.37 + ], + "angle": 0, + "content": "57. Xu, D., Jiang, Y., Wang, P., Fan, Z., Wang, Y., Wang, Z.: Neurallift-360: Lifting an in-the-wild 2d photo to a 3d object with 360 views. arXiv e-prints pp. arXiv-2211 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.371, + 0.787, + 0.397 + ], + "angle": 0, + "content": "58. Ye, H., Zhang, J., Liu, S., Han, X., Yang, W.: Ip-adapter: Text compatible image prompt adapter for text-to-image diffusion models (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.398, + 0.787, + 0.439 + ], + "angle": 0, + "content": "59. Yi, T., Fang, J., Wu, G., Xie, L., Zhang, X., Liu, W., Tian, Q., Wang, X.: Gaussian dreamer: Fast generation from text to 3d gaussian splatting with point cloud priors. arXiv preprint arXiv:2310.08529 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.44, + 0.787, + 0.481 + ], + "angle": 0, + "content": "60. Yu, C., Zhou, Q., Li, J., Zhang, Z., Wang, Z., Wang, F.: Points-to-3d: Bridging the gap between sparse points and shape-controllable text-to-3d generation. arXiv preprint arXiv:2307.13908 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.481, + 0.787, + 0.522 + ], + "angle": 0, + "content": "61. Yu, W., Yuan, L., Cao, Y.P., Gao, X., Li, X., Quan, L., Shan, Y., Tian, Y.: Hifi-123: Towards high-fidelity one image to 3d content generation. arXiv preprint arXiv:2310.06744 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.523, + 0.787, + 0.564 + ], + "angle": 0, + "content": "62. Zhang, L., Rao, A., Agrawala, M.: Adding conditional control to text-to-image diffusion models. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 3836-3847 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.565, + 0.787, + 0.592 + ], + "angle": 0, + "content": "63. Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual metric. In: CVPR (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.592, + 0.787, + 0.62 + ], + "angle": 0, + "content": "64. Zhu, J., Zhuang, P.: Hifa: High-fidelity text-to-3d with advanced diffusion guidance. arXiv preprint arXiv:2305.18766 (2023)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.62 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/Repaint123_ Fast and High-quality One Image to 3D Generation with Progressive Controllable Repainting/a8257f3f-735a-49a4-ab19-4aa6628f29e5_origin.pdf b/2024/Repaint123_ Fast and High-quality One Image to 3D Generation with Progressive Controllable Repainting/a8257f3f-735a-49a4-ab19-4aa6628f29e5_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..8b6786e1af432f6957e96d2198af72a6c6d72137 --- /dev/null +++ b/2024/Repaint123_ Fast and High-quality One Image to 3D Generation with Progressive Controllable Repainting/a8257f3f-735a-49a4-ab19-4aa6628f29e5_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2a2ef424389c56c6499eb8d72451b7856478d817458e9bb8586a866a40a879d0 +size 9434381 diff --git a/2024/Repaint123_ Fast and High-quality One Image to 3D Generation with Progressive Controllable Repainting/full.md b/2024/Repaint123_ Fast and High-quality One Image to 3D Generation with Progressive Controllable Repainting/full.md new file mode 100644 index 0000000000000000000000000000000000000000..f05f883788a7879371eca9aae2c3c1cf2953666b --- /dev/null +++ b/2024/Repaint123_ Fast and High-quality One Image to 3D Generation with Progressive Controllable Repainting/full.md @@ -0,0 +1,280 @@ +# Repaint123: Fast and High-quality One Image to 3D Generation with Progressive Controllable Repainting + +Junwu Zhang1* Zhenyu Tang1* Yatian Pang1,3 Xinhua Cheng1 Peng Jin1 Yida Wei4 Xing Zhou5 Munan Ning1,2 Li Yuan1,2* +1Peking University 2Pengcheng Laboratory 3National University of Singapore +45RabbitPre + +![](images/7dfab4bc6fd25e686d0908268ecef8871845a43d7f8ede6eaf58479ad8b584e7.jpg) +Fig. 1: Repaint123 generates high-quality 3D content with detailed texture from a single image in 2 minutes. Repaint123 adopts 3D Gaussian Splatting in the coarse stage and then utilizes a progressively repairing strategy with the diffusion model for high-quality 3D content generation with efficiency. + +Abstract. Recent image-to-3D methods achieve impressive results with plausible 3D geometry due to the development of diffusion models and optimization techniques. However, existing image-to-3D methods suffer from texture deficiencies in novel views, including multi-view inconsistency and quality degradation. To alleviate multi-view bias and enhance image quality in novel-view textures, we present Repaint123, a fast image-to-3D approach for creating high-quality 3D content with detailed textures. Repaint123 proposes a progressively repairing strategy to simultaneously enhance the consistency and quality of textures across different views, generating invisible regions according to visible textures, with the visibility map calculated by the depth alignment across views. + +Furthermore, multiple control techniques, including reference-driven information injection and coarse-based depth guidance, are introduced to alleviate the texture bias accumulated during the repainting process for improved consistency and quality. For novel-view texture refinement with short-term view consistency, our method progressively repairs novel-view images with adaptive strengths based on visibility, enhancing the balance of image quality and view consistency. To alleviate the accumulated bias as progressively repainting, we control the repainting process by depth-guided geometry and attention-driven reference-view textures. Extensive experiments demonstrate the superior ability of our method to create 3D content with consistent and detailed textures in 2 minutes. + +Keywords: Image-to-3D $\cdot$ Texture Enhancement $\cdot$ Diffusion model + +# 1 Introduction + +Generating 3D content from one given reference image plays a key role at the intersection of computer vision and computer graphics [11, 17, 24, 25, 33, 35], which is desired by users for innovative applications across fields including robotics, virtual reality, and augmented reality. However, the image-to-3D task is quite challenging due to the generated 3D content is expected with reasonable geometry and consistent textures, while the single image input is insufficient in multi-view information. Recent studies [24, 28, 35, 50, 51] employ 2D diffusion models [14, 39] to guide 3D generation from the reference image with Score Distillation Sampling (SDS) [34]. While preliminary attempts [28, 51] leveraging view-independent diffusion models (e.g. Stable Diffusion [39]) faced challenges with over-saturated textures and multi-face geometry, more + +recent efforts [24, 26, 35, 46, 50] utilize view-conditioned diffusion models [24, 26, 46]. Although these methods utilizing view-conditioned diffusion models mitigate the multi-face problem and generate 3D content with plausible geometry, they still encounter challenges in generating high-quality textures. This limitation stems from training these models on small-scale synthetic 3D datasets, hindering their ability to produce detailed images from novel views. + +![](images/de3c7f701ab1e3f72825b2f035b8dab4103ed427b498cf030a640de73bdc38cd.jpg) +Fig.2: Motivation. Current image-to-3D methods adopt SDS loss for generating, resulting in inconsistent and poor textures. Repaint123 proposes a progressively repainting strategy based on visible content to efficiently create invisible content with MSE loss, achieving consistent and high-quality textures. + +To address textural deficiencies including inconsistency and low quality mentioned above, we propose a novel approach named Repaint123 for fast and high-quality image-to-3d generation by generating consistent and high-quality novel-view images with a controllable content repaint process. As shown in Figure 2, the core of Repaint123 is to progressively repaint the invisible content with a view-independent diffusion model for adjacent-view consistency, incorporating multiple controls of geometry and texture for enhanced quality and multi-view consistency. Specifically, our method adopts a two-stage optimization framework. In the coarse stage, we utilize 3D Gaussian Splatting [18] as the representation to efficiently obtain a coarse 3D model in 1 minute. In the fine stage, we progressively sample the camera views around and repaint invisible content with control techniques including using novel-view depth map as geometry prior, injecting reference-view attention features as texture prior [61] and employing reference-view image prompt as semantic prior. To further enhance image quality and preserve multi-view consistency, we propose a visibility-based adaptive strategy that refines the visible regions with different strengths. After achieving high-qualityainted images with multi-view consistency, we directly leverage simple MSE loss to efficiently refine 3D content textures instead of the time-consuming and over-smoothed SDS. + +We conduct extensive experiments on multiple datasets and demonstrate that our method significantly advances image-to-3D generation, producing high-quality, multi-view consistent 3D content with detailed textures in approximately 2 minutes from scratch (as shown in Figure 1), outperforming current state-of-the-art NeRF-based and Gaussian-Splatting-based methods in both consistency and texture quality. Our contributions can be summarized as follows: + +- We propose a progressively repairing strategy based on visibility for the image-to-3D task to generate 3D content with consistency and high quality by repairing invisible parts according to visible textures. +- We introduce multiple control techniques including reference-driven information injection and coarse-based depth guidance to alleviate the texture bias accumulated during the repainting process for further quality enhancement. +- Our experiments demonstrate Repaint123 efficiently generates high-quality 3D content with consistent and detailed textures in 2 minutes. + +# 2 Related Works + +# 2.1 Diffusion Models for 3D Generation + +The recent notable achievements in view-independent 2D diffusion models [14,39] have brought about exciting prospects for generating 3D objects. Pioneering studies [34, 53] have introduced the concept of distilling a 2D text-to-image generation model to generate 3D shapes. Subsequent works [1, 6, 7, 10, 16, 22, 28, 35, 37, 42-44, 51, 52, 55-57, 60, 61, 64] have adopted a similar per-object optimization approach, building upon these initial works. Nevertheless, the majority of these techniques consistently experience low efficiency and multi-face + +issues. Unlike the previous study HiFi-123 [61], which employed similar inversion and attention injection techniques for image-to-3D generation, our approach diverges primarily in our proposed repairing strategy, as well as many variations in optimization objectives, image prompting, and 3D representations. Recently, some works [23, 25, 26, 48] extend view-independent diffusion model to view-conditioned diffusion model to generate multi-view images for reconstruction, while these methods usually suffer from low-quality textures as the multi-view diffusion models are trained on limited and synthesized data. To optimize both consistency and quality, we integrate a view-conditioned diffusion model for initial consistency in the coarse stage and a view-independent model to enhance texture quality in the fine stage. + +# 2.2 Controllable Image Synthesis + +A major challenge in the field of image generation is achieving controllability. Many works have been done recently to increase the controllability of images generated by diffusion models. ControlNet [62] and T2I-Adapter [31] attempt to control the creation of images by utilizing data from different modalities. Besides, some optimization-based methods [12,30,40] learn new parameters or fine-tune the diffusion model in order to enhance control over the generation process. Other methods [3,58] leverage multi-view attention to introduce information from other-view images for gaining superior control. We incorporate multiple control techniques into the view-independent diffusion model for maintaining view consistency in the fine stage. + +# 2.3 3D Representations + +Neural Radiance Fields (NeRF) [29], as a volumetric rendering method, has gained popularity for its ability to enable 3D optimization [2,4,8,13,21] under 2D supervision, while NeRF optimization can be time-consuming. Numerous efforts [32,41] for spatial pruning have been dedicated to accelerating the training process of NeRF on the reconstruction setting. However, they fail in the generation setting. Recently, 3D Gaussian Splatting [5,9,18,50,59] has emerged as an alternative 3D representation to NeRF and has shown remarkable advancements in terms of both quality and speed, offering a promising avenue. Therefore, we adopt Gaussian Splatting for efficiently generating coarse 3D model. + +# 3 Preliminary + +# 3.1 DDIM Inversion + +DDIM [47] transforms random noise $\pmb{x}_T$ into clean data $\pmb{x}_0$ over a series of time steps, by using the deterministic DDIM sampling in the reverse process, i.e., $\pmb{x}_{t - 1} = (\alpha_{t - 1} / \alpha_t)(\pmb{x}_t - \sigma_t\epsilon_\phi) + \sigma_{t - 1}\epsilon_\phi$ . On the contrary, DDIM inversion progressively converts clean data to a noisy state $\pmb{x}_T$ , i.e., $\pmb{x}_t = (\alpha_t / \alpha_{t - 1})(\pmb{x}_{t - 1} - \sigma_t\epsilon_\phi)$ . + +![](images/f803aef4bd19c57067c8f2e5580f4c5a5ecb736b5efa92be62e662497d20c5be.jpg) +Fig. 3: Overview of Repaint123. In the coarse stage, we leverage 3D Gaussian Splatting to represent the coarse content, which is optimized by MSE loss at the reference and SDS loss at the invisible views. In the fine stage, we convert coarse content to mesh and progressively sample and generate invisible views from the reference view bidirectionally. Concretely, the current view is repainted based on the visibility map from bidirectional neighbor views with control techniques including attention injection and depth guidance, leading to consistent and detailed textures. + +$\sigma_{t-1}\epsilon_{\phi}) + \sigma_t\epsilon_{\phi}$ , here $\epsilon_{\phi}$ is the predicted noise by the UNet. This method can precisely reconstruct the original clean data while greatly speeding up the process by skipping many intermediate diffusion steps. + +# 3.2 3D Gaussian Splitting + +Gaussian Splatting [18] presents a novel method for synthesizing new views and reconstructing 3D scenes, achieving real-time speed. Unlike NeRF, Gaussian Splatting uses a set of anisotropic 3D Gaussians defined by their locations, covariances, colors, and opacities to represent the scene. To compute the color of each pixel $\mathbf{p}$ in the image, it utilizes a typical neural point-based rendering [19,20], The rendering process is as follows: + +$$ +C (\mathbf {p}) = \sum_ {i \in \mathcal {N}} c _ {i} \alpha_ {i} \prod_ {j = 1} ^ {i - 1} (1 - \alpha_ {j}), \tag {1} +$$ + +$$ +\mathrm {w h e r e}, \alpha_ {i} = o _ {i} e ^ {- \frac {1}{2} (\mathbf {p} - \boldsymbol {\mu} _ {i}) ^ {T} \boldsymbol {\Sigma} _ {i} ^ {- 1} (\mathbf {p} - \boldsymbol {\mu} _ {i})}, +$$ + +where $c_{i}$ , $o_{i}$ , $\mu_{i}$ , and $\Sigma_{i}$ represent the color, opacity, position, and covariance of the $i$ -th Gaussian respectively, and $\mathcal{N}$ denotes the number of related Gaussians. + +# 4 Method + +In this section, we introduce our two-stage framework named Repaint123 for fast and high-quality image-to-3D generation, as illustrated in Figure 3. In the coarse stage, we adopt 3D Gaussian Splitting [18] representation optimized by + +SDS loss [34] with the view-conditioned diffusion model [24] to learn a coarse geometry and texture (see Section 4.1). Inspired by [50], we convert the coarse 3D Gaussians to textured mesh to facilitate subsequent texture enhancement in UV space. In the fine stage, we progressively paint invisible content in novel views by rotating camera viewpoints bi-directionally from the reference view (see Section 4.2). To mitigate texture bias accumulated during the gradual painting across wide angles, we incorporate multiple control techniques including depth guidance, injection of reference-view attention, and the use of reference-view image prompts (see Section 4.2). Additionally, we propose a visibility-based adaptive strategy to refine visible regions when observed from a superior perspective (see Section 4.2). After obtaining a high-quality painted image that remains consistent across different views, we efficiently optimize the texture map by simple Mean Square Error (MSE) loss (see Section 4.2). This process is repeated for each sampled view, alternating between painting and optimization, until the texture is completely reconstructed to accommodate a 360-degree perspective. + +# 4.1 Coarse Stage: Gaussian Splatting with 3D Diffusion Prior + +In the coarse stage, we adopt 3D Gaussian Splatting for efficient initialization. Through SDS [34], 3D diffusion priors are back-propagated to the 3D Gaussians. At each diffusion step, we sample a random current camera view $v_{c}$ and its camera pose $p_{c}$ , which orbits around the center of the object. Subsequently, the RGB image, denoted as $I_{\mathrm{rgb}}^{c}$ , and the alpha image, represented by $I_{\mathrm{a}}^{c}$ , are rendered from the current viewpoint $p$ . Given an image $\tilde{I}_{\mathrm{rgb}}^{r}$ and an alpha mask $\tilde{I}_{\mathrm{a}}^{r}$ as input to the reference view, the SDS loss can be formulated as: + +$$ +\mathcal {L} _ {\mathrm {S D S}} = \mathbb {E} _ {t, c, \epsilon} \left[ (\epsilon_ {\phi} (I _ {\mathrm {r g b}} ^ {c}; t, \tilde {I} _ {\mathrm {r g b}} ^ {r}, \Delta p _ {c}) - \epsilon) \frac {\partial I _ {\mathrm {r g b}} ^ {c}}{\partial \Theta} \right] \tag {2} +$$ + +where $\epsilon_{\phi}(\cdot)$ is the predicted noise by the diffusion $\phi$ , $\Delta p_{c}$ is the relative camera pose and $\Theta$ represents the estimizable Gaussian parameters. Additionally, we optimize the RGB image and alpha image rendered from the reference view to align with the input reference image using reconstruction loss: + +$$ +\mathcal {L} _ {\mathrm {R e f}} = \lambda_ {\mathrm {r g b}} | | I _ {\mathrm {r g b}} ^ {r} - \tilde {I} _ {\mathrm {r g b}} ^ {r} | | _ {2} ^ {2} + | | I _ {\mathrm {a}} ^ {r} - \tilde {I} _ {\mathrm {a}} ^ {r} | | _ {2} ^ {2}. \qquad (3) +$$ + +The final loss function is the weighted sum of $\mathcal{L}_{\mathrm{SDS}}$ and $\mathcal{L}_{\mathrm{Ref}}$ . + +# 4.2 Fine Stage: Visibility-based Controllable Repainting + +Progressive Invisible Content Repainting. We repaint previously unseen content with progressively sampling novel camera viewpoints, beginning from the reference view and increasing a certain angular interval with each iteration. To obtain invisible regions in the image from a newly sampled view $v_{c}$ , we use the back-projection technique based on the depth map $D^{p}\in \mathbb{R}^{H\times W}$ in the previously seen view $v_{p}$ . Formally, given the camera intrinsics $K\in \mathbb{R}^{3\times 3}$ , we first remap the + +pixels of the image rendered from $v_{p}$ into 3D points $X^{p}$ expressed in the camera coordinate frame, which can be obtained as $X_{i,j} = K^{-1}[iD_{i,j},jD_{i,j},D_{i,j}]^{\top}$ . With $P_{p},P_{c}\in \mathbb{R}^{3\times 4}$ the world-to-camera poses for images from $v_{p}$ and $v_{c}$ , we can transform $X^{p}$ into camera coordinate frame of $v_{c}$ to obtain $X^{p,c}$ : + +$$ +X ^ {p, c} = P _ {p} P _ {c} ^ {- 1} h \left(X ^ {p}\right), \tag {4} +$$ + +where $h:(x,y,z)\to (x,y,z,1)$ is the homogeneous mapping. Then we can render a depth map $D^{p,c}$ from $X^{p,c}$ as $D_{i,j}^{p,c} = KX_{i,j,2}^{p,c}$ . Regions with dissimilar depth values between $D^c$ and $D^{p,c}$ are recognized as invisible regions, forming a repainting mask $M$ where areas with zeros are invisible and areas with ones are visible. Therefore, we repaint invisible content and maintain visible content by: + +$$ +x _ {t} = x _ {t} ^ {\mathrm {i n v}} \odot M + x _ {t} ^ {\mathrm {r e v}} \odot (1 - M), \tag {5} +$$ + +where $\odot$ is element-wise matrix multiplication, $t$ is the current denoising timestep, $x_{t}^{\mathrm{inv}}$ is the DDIM [47] inverted latent, and $x_{t}^{\mathrm{rev}} \sim \mathcal{N}(\mu_{\phi}(x_{t + 1},t),\Sigma_{\phi}(x_{t + 1},t))$ is the diffusion model output. + +Additionally, to mitigate accumulated bias for distant camera views during gradual repainting, as shown in the fine stage in Figure 3, we sample camera viewpoints bi-directionally, alternating between clockwise and counterclockwise directions. This bidirectional sampling allows us to paint up to 180 degrees from two opposite directions, instead of painting across the entire 360 degrees from one direction. Moreover, to paint bidirectionally invisible content and ensure smooth transitions at the junction of the two directions, we replace the previous unidirectional painting mask $M$ by a bidirectional painting mask. This mask merges two unidirectional masks from bidirectional neighbor views (red cameras in Figure 3) by taking the maximum value of each pixel. + +Control-based Content Enhancement. While the progressive painting maintains consistency between adjacent views, views that are farther apart may become inconsistent and experience a decline in quality due to cumulative errors. To tackle these issues, we improve the painting process by incorporating various control techniques to enhance the consistency and quality of synthesized images. Specifically, for geometry control, we employ ControlNet [62] to condition the diffusion model with coarse depth maps, ensuring geometric consistency across images. To transfer reference textures and mitigate the cumulative texture bias during the progressive painting, inspired by HiFi-123 [61], we incorporate an attention injection mechanism [3] that injects reference attention features into the diffusion denoising process. By replacing the novel-view content features (Key features $K_{t}$ and Value features $V_{t}$ ) with reference-view attention features ( $K_{r}$ and $V_{r}$ ) during each denoising step, the novel-view image features can directly query the high-quality reference features by: + +$$ +\operatorname {A t t e n t i o n} \left(Q _ {t}, K _ {r}, V _ {r}\right) = \operatorname {S o f t m a x} \left(\frac {Q _ {t} K _ {r} ^ {T}}{\sqrt {d}}\right) V _ {r}, \tag {6} +$$ + +![](images/c0c41a3644127407014e0bd4b849a1abf8447e2b634353c776f0f94ade11bebe.jpg) +Fig. 4: Visibility-based Controllable Repainting Pipeline. Our method repairs invisible content (black areas in $M_t$ ) of novel-view rendered image $I_{\mathrm{rgb}}$ according to the inverted latent $x_t^{inv}$ and the visibility map $V$ to obtain refined image $I_{\mathrm{rgb}}^{\mathrm{fine}}$ . We introduce multiple control techniques to enhance the consistency and quality of generated textures, including attention feature injection, depth guidance by ControlNet, and semantic prompt by CLIP encoder. Noticing that the repainting mask $M_t$ is simultaneously changed with denoising timesteps for adaptive refinement. + +where $Q_{t}$ represents the novel-view query features. This enhancement facilitates the transfer of texture details and improves the consistency between the reference-view image and novel-view images. + +During our experiments, we observed a decline in image quality during repainting, likely due to not using text prompts for performing classifier-free guidance [15], crucial for producing high-quality images in diffusion models. To improve this, we adopt IP-Adapter [58], which uses a CLIP encoder to convert reference images into semantic features. These features are then projected into image prompts, providing visual cues that enhance image generation through classifier-free guidance. + +Adaptive Visible Content Refinement. Repainting only the invisible content often leads to distortion in the visible area As shown in Figure 5, when the previous view is an oblique view, it leads to a low-resolution update on the texture maps, resulting in high distortion when rendering from a better observing view. Therefore, the visible content currently observed in a superior view should be refined to replenish details. However, the selection of refinement strength for these visible content is tricky [38, 54], as excessive strength produces unfaithful results leading to inconsistency with previously repainted views, while insufficient strength limits quality improvement. We propose a visibility-based adaptive refinement strategy to refine these previously seen regions with different strengths, aiming to achieve improved quality-consistency trade-off. As visualized in Figure 5, we view painting for visible content as a process similar to superresolution that replenishes detailed information. According to the Orthographic + +Projection Theorem, which asserts that the projected resolution of a surface is directly proportional to the view obliqueness ( $\cos\theta$ as shown in Figure 5), we can assume that the repainting strength is equal to $(1 - \cos\theta_{1} / \cos\theta_{2})$ , where $\cos\theta_{2} / \cos\theta_{1}$ represents the upsampling scale in terms of super-resolution and visibility variations in the context of 3D rendering. Therefore, we can define a visibility map $V$ as $V_{i,j} = \cos\theta_{i,j}^{p,c} / \cos\theta_{i,j}^{c}$ , where $\cos\theta_{i,j}^{p,c}$ and $\cos\theta_{i,j}^{c}$ are obtained by multiplying the current-view normal map with camera rays of the previous view $v_{p}$ and the current view $v_{c}$ respectively. Higher visibility value indicates less repainting strength, while invisible areas require full strength. + +By this association between repairing strength and the visibility map $V$ , we can binarize the visibility map to the timestep-aware adaptive repairing mask $M_{t}$ during each denoising step, visualized in the green box "Timestep-aware binarization" in Figure 4: + +$$ +M _ {t} ^ {i, j} = \left\{ \begin{array}{l l} 1, & \text {i f} V ^ {i, j} > 1 - t / T \\ 0, & \text {e l s e}, \end{array} \right. \tag {7} +$$ + +where $i$ , and $j$ are the 2D position of pixels in visibility map $V$ , and $T$ is the total number of timesteps of the diffusion model. By doing this, we can adaptively refine visible content with a proper strength. + +Efficient Optimization. Using a progressive painting strategy and various control techniques to refine the inverted latent, we obtain a high-quality, view-consistent refined image, $I_{\mathrm{rgb}}^{\mathrm{fine}}$ . This process allows us to use simple MSE loss for efficient texture map optimization: + +$$ +\mathcal {L} _ {\text {f i n e}} = \left| \left| I _ {\mathrm {r g b}} ^ {\text {f i n e}} - I _ {\mathrm {r g b}} \right| \right| _ {2} ^ {2}, \tag {8} +$$ + +where $I_{\mathrm{rgb}}$ represents the rendered image in the fine stage. MSE loss is faster and more deterministic to optimize than the traditional SDS loss, speeding up the optimization process. After sampling a novel view, we first conduct image repainting and then proceed with efficient texture optimization, repeating this process until we cover a 360-degree view range. + +![](images/3c926188a9f39f0ddd010635916674dffc9e5b8ea128ee7aed973698f09f7502.jpg) +Fig. 5: Relation between view obliqueness and refinement strength. Red box shows the same texture part in different views. Texture details in generated views might be insufficient due to the variation of angles. + +# 5 Experiment + +# 5.1 Implementation Details + +In our experiment, we consistently apply the same hyperparameters across all our method's results. We progressively turn around the viewpoints by 40 degrees each time, consequently obtaining 8 views per object. We use a 50-step DDIM schedule and perform a 30-step latent diffusion inversion. Stable Diffusion 1.5 is utilized for all methods. During the coarse stage training, we set $\lambda_{\mathrm{rgb}}$ to 10. Generating a single 3D object takes only 2 minutes on a single 40G A100 GPU, approximately 1 minute for the coarse stage and mesh extraction, and another 1 minute for refinement. + +# 5.2 Baselines + +We adopt RealFusion [28], Make-It-3D [51], and Zero123-XL [24], Magic123 [35] as our NeRF-based baselines and DreamGaussian [50] as our Gaussian-Splatting-based baseline. RealFusion presents a single-stage algorithm for NeRF generation leveraging 2D SDS loss for novel views. Make-It-3D is a two-stage approach that shares similar objectives with RealFusion but employs a point cloud representation for refinement at the second stage. For Zero123-XL, we adopt the implementation [49] and add a mesh fine-tuning stage for fair comparison. Integrating Zero123 and RealFusion, Magic123 incorporates a 2D SDS loss with 3D SDS loss provided by Zero123 to balance geometry and texture and adopts DMTet [45] representation at the second stage. DreamGaussian integrates 3D Gaussian Splatting into 3D generation and greatly improves the speed. + +# 5.3 Evaluation Protocol + +Datasets. Based on previous research, we utilized the Realfusion15 dataset [28] and test-alpha dataset collected by Make-It-3D [51], which comprises many common objects from diverse styles. + +Evaluation metrics. An effective 3D generation approach should produce 3D content which closely resemble the reference view, and maintain consistency of semantics and textures with the reference image when observed from new views. Therefore, to evaluate the overall quality of the generated 3D object, we choose the following metrics from two aspects: 1) PSNR and LPIPS [63], which measure pixel-level and perceptual reconstruction quality respectively at the reference view; 2) CLIP Similarity [36] and Contextual Distance [27], which assess the similarity of semantics and detailed textures respectively between the novel perspective and the reference view. + +![](images/4c34e05eed042a781c9a901ee81c2129140d5ad78431f0f79ad645c6d86c11e9.jpg) +Fig. 6: Qualitative comparisons on image-to-3D generation. Zoom in for details. + +
DatasetMetrics\MethodsNeRF-basedGaussian-Splatting-based
RealFusionMake-it-3DZero-123-XL*Magic123DreamGaussianRepaint123
RealFusion15CLIP-Similarity↑0.710.810.830.820.770.85
Context-Dis↓2.201.821.591.641.611.55
PSNR↑19.2416.5619.5619.6818.9419.00
LPIPS↓0.1940.1770.1080.1070.1110.101
Test-alphaCLIP-Similarity↑0.680.760.840.840.790.88
Context-Dis↓2.201.731.521.571.621.50
PSNR↑22.9117.2124.3924.6922.3322.38
LPIPS↓0.1050.2370.0500.0460.0570.048
Optimization time20min1h30min1h (+2h)2min2 min
+ +Table 1: We show quantitative results in terms of CLIP-Similarity↑ / Contextual-Distance↓ / PSNR↑ / LPIPS↓. The results are shown on the RealFusion15 and test-alpha datasets, while bold reflects the best for all methods and the underline represents the best for Gaussian-Splatting-based methods. * indicates that Zero123-XL incorporates a mesh fine-tuning stage for further quality improvement. The time required by textual inversion is indicated in parentheses. + +# 5.4 Comparisons + +Comparisons with NeRF-based Methods. As shown in Table 1, we evaluate the quality of generated 3D objects across various NeRF-based methods. Our method achieves superior 3D consistency in generating 3D objects, as evidenced by the best performance of CLIP-similarity and Contextual-distance. Regarding reference-view reconstruction quality, there is a gap compared with NeRF-based approaches as shown in Table 1, which we attribute to the immaturity of current Gaussian-Splatting-based methods. Compared to NeRF-based methods for the optimization time, our approach reaches a significant acceleration of over 10 times and simultaneously achieves high quality, due to the 3D Gaussian-Splatting representation in coarse stage and efficient texture refinement in fine stage. As shown in Figure 6, Repaint123 achieves the best visual results in terms of texture consistency and generation quality as opposed to other NeRF-based methods. From the visual comparison, our method achieves consistent and detailed textures in invisible areas, while Zero123-XL results in over-smooth textures, Magic123 produces inconsistent and oversaturated colors, Realfusion and Make-It-3D fail to generate full geometry and consistent textures. + +Comparisons with GS-based methods. We conduct comparisons with other works based on 3D Gaussian Splatting in the last column of Table 1. Within a comparable generation time, our method demonstrates superiority over existing Gaussian-based approach in aspects of texture consistency and the quality of reference-view reconstruction. The superiority of our proposed Repaint123 is evidenced by the evaluation of four distinct metrics from the Table 1. As shown in Figure 6, DreamGaussian usually leads to over-smooth texture inconsistent with reference view, while our method can produce high-quality 3D content with view-consistent and detailed textures. + +![](images/46cee30e8c7b0c143b6339a8dd8e9bcd57a50eaae9fc1070caf5cb9b894a8502.jpg) +Fig. 7: Qualitative ablation study by removing one component at a time from the overall method. The presence of artifacts is highlighted by red boxes. Inconsistencies such as multi-face problems and mismatches in content and style are evident without using repainting and attention injection strategies. Absence of image prompting and adaptive refinement notably degrades quality. + +# 5.5 Ablation and Analysis + +This section details both qualitative and quantitative analyses to highlight the effectiveness of our proposed methods, as illustrated in Figure 7 and Table 2. + +Effectiveness of Repainting. Figure 7 illustrates that without repainting strategy, the generated novel-view image tends to resemble the reference image, leading to inconsistencies in the shared visible areas between different views, such as the content misalignment (observed in the Ice Cream example) and issues with multiple faces (evident in the Anya example). These issues stem from the absence of alignment constraints for shared visible regions, leading to conflicts and quality degradation of the reconstructed 3D texture. + +Impact of Attention Injection. Table 2 shows that injecting attention features from the reference-view image markedly enhances the consistency of both semantics and fine-grained textures, as evidenced by increased CLIP similarity and reduced Contextual distance. Without this injection strategy, as depicted in Figure 7, the synthesized images retain basic semantics but fail to accurately transfer detailed textures and styles from the reference view, resulting in inconsistencies across multiple views. + +
Method \ MetricCLIP↑Contextual↓PSNR↑LPIPS↓
Coarse0.711.7821.170.133
vanilla painting0.711.6222.410.049
+attention injection0.781.5622.420.048
+image prompt0.841.5222.400.048
+adaptive refinemnet (Ours)0.881.5022.380.048
+ +Table 2: Quantitative ablation study on Test-alpha dataset by progressively adding our proposed components. The last three lines show the cumulative effects of the proposed modules. Our approach, utilizing all strategies, delivered the best performance. + +Role of Image Prompt. The introduction of using the reference image as an image prompt, as shown in Table 2 and Figure 7, significantly boosts both multi-view consistency and image quality of the generated images. Without this technique to perform classifier-free guidance, the generation of detailed and consistent textures across views is compromised. + +Benefits of Adaptive Refinement. The necessity of adaptive refinement is clear from Figure 7, where its absence leads to artifacts and blurriness in obliquely viewing areas of previous views due to low-resolution updates, as mentioned in Section 4.2. Table 2 also demonstrates its benefits through improved CLIP similarity and Contextual distance. + +# 6 Limitations + +Despite the promising results, our method still has some limitations. Our method is based on 3D Gaussian Splatting representation in the coarse stage during training. While 3D Gaussian Splatting accelerates the training process, it may exhibit geometry artifacts during mesh extraction, such as holes, and inferior results compared to NeRF-based methods for the reconstruction of reference view, due to immaturity in generation tasks. This is supported by results in the Appendix, where NeRF is considered as an alternative to 3D Gaussian Splatting for the coarse stage. We expect these limitations can be mitigated with the development of 3D Gaussian Splatting in the future. + +# 7 Conclusion + +This work presents Repaint123 for generating high-quality 3D content from a single image in about 2 minutes. By leveraging progressive controllable paint, our approach overcomes the limitations of existing studies and achieves state-of-the-art results in terms of both texture quality and multi-view consistency, paving the way for future progress in 3D content generation from one image. Furthermore, we validate the effectiveness of our proposed method through a comprehensive set of experiments. + +# Acknowledgements + +This work was supported in part by Natural Science Foundation of China (No. 62332002, 62202014), and Shenzhen Basic Research Program (No.JCYJ20220813151736001) + +# References + +1. Armandpour, M., Zheng, H., Sadeghian, A., Sadeghian, A., Zhou, M.: Re-imagine the negative prompt algorithm: Transform 2d diffusion into 3d, alleviate janus problem and beyond. arXiv preprint arXiv:2304.04968 (2023) +2. Barron, J.T., Mildenhall, B., Verbin, D., Srinivasan, P.P., Hedman, P.: Mip-nerf 360: Unbounded anti-aliased neural radiance fields. CVPR (2022) +3. Cao, M., Wang, X., Qi, Z., Shan, Y., Qie, X., Zheng, Y.: Masactrl: Tuning-free mutual self-attention control for consistent image synthesis and editing. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 22560-22570 (October 2023) +4. Chan, E.R., Lin, C.Z., Chan, M.A., Nagano, K., Pan, B., Mello, S.D., Gallo, O., Guibas, L., Tremblay, J., Khamis, S., Karras, T., Wetzstein, G.: Efficient geometry-aware 3D generative adversarial networks. In: CVPR (2022) +5. Chen, G., Wang, W.: A survey on 3d gaussian splatting. arXiv preprint arXiv:2401.03890 (2024) +6. Chen, R., Chen, Y., Jiao, N., Jia, K.: *Fantasia3d: Disentangling geometry and appearance for high-quality text-to-3d content creation*. arXiv preprint arXiv:2303.13873 (2023) +7. Chen, Y., Zhang, C., Yang, X., Cai, Z., Yu, G., Yang, L., Lin, G.: It3d: Improved text-to-3d generation with explicit view synthesis. arXiv preprint arXiv:2308.11473 (2023) +8. Chen, Z., Funkhouser, T., Hedman, P., Tagliasacchi, A.: Mobilenerf: Exploiting the polygon rasterization pipeline for efficient neural field rendering on mobile architectures. arXiv preprint arXiv:2208.00277 (2022) +9. Chen, Z., Wang, F., Liu, H.: Text-to-3d using gaussian splatting. arXiv preprint arXiv:2309.16585 (2023) +0. Cheng, X., Yang, T., Wang, J., Li, Y., Zhang, L., Zhang, J., Yuan, L.: Progressive3d: Progressively local editing for text-to-3d content creation with complex semantic prompts. arXiv preprint arXiv:2310.11784 (2023) +1. Dou, Z., Wu, Q., Lin, C., Cao, Z., Wu, Q., Wan, W., Komura, T., Wang, W.: Tore: Token reduction for efficient human mesh recovery with transformer. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 15143-15155 (2023) +2. Gal, R., Alaluf, Y., Atzmon, Y., Patashnik, O., Bermano, A.H., Chechik, G., Cohen-Or, D.: An image is worth one word: Personalizing text-to-image generation using textual inversion. arXiv preprint arXiv:2208.01618 (2022) +3. Hedman, P., Srinivasan, P.P., Mildenhall, B., Barron, J.T., Debevec, P.: Baking neural radiance fields for real-time view synthesis. ICCV (2021) +4. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. arXiv preprint arxiv:2006.11239 (2020) +5. Ho, J., Salimans, T.: Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598 (2022) + +16. Huang, Y., Wang, J., Shi, Y., Qi, X., Zha, Z.J., Zhang, L.: Dreamtime: An improved optimization strategy for text-to-3d content creation. arXiv preprint arXiv:2306.12422 (2023) +17. Jun, H., Nichol, A.: Shap-e: Generating conditional 3d implicit functions. arXiv preprint arXiv:2305.02463 (2023) +18. Kerbl, B., Kopanas, G., Leimkuhler, T., Drettakis, G.: 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics (ToG) 42(4), 1-14 (2023) +19. Kopanas, G., Leimkuhler, T., Rainer, G., Jambon, C., Drettakis, G.: Neural point catacaustics for novel-view synthesis of reflections. ACM Transactions on Graphics (TOG) 41(6), 1-15 (2022) +20. Kopanas, G., Philip, J., Leimkuhler, T., Drettakis, G.: Point-based neural rendering with per-view optimization. In: Computer Graphics Forum. vol. 40, pp. 29-43. Wiley Online Library (2021) +21. Li, Z., Müller, T., Evans, A., Taylor, R.H., Unberath, M., Liu, M.Y., Lin, C.H.: Neuralangelo: High-fidelity neural surface reconstruction. In: CVPR (2023) +22. Lin, C.H., Gao, J., Tang, L., Takikawa, T., Zeng, X., Huang, X., Kreis, K., Fidler, S., Liu, M.Y., Lin, T.Y.: Magic3d: High-resolution text-to-3d content creation. In: CVPR (2023) +23. Liu, M., Xu, C., Jin, H., Chen, L., Xu, Z., Su, H., et al.: One-2-3-45: Any single image to 3d mesh in 45 seconds without per-shape optimization. arXiv preprint arXiv:2306.16928 (2023) +24. Liu, R., Wu, R., Van Hoorick, B., Tokmakov, P., Zakharov, S., Vondrick, C.: Zero-1-to-3: Zero-shot one image to 3d object. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 9298-9309 (2023) +25. Liu, Y., Lin, C., Zeng, Z., Long, X., Liu, L., Komura, T., Wang, W.: Syncdreamer: Generating multiview-consistent images from a single-view image. arXiv preprint arXiv:2309.03453 (2023) +26. Long, X., Guo, Y.C., Lin, C., Liu, Y., Dou, Z., Liu, L., Ma, Y., Zhang, S.H., Habermann, M., Theobalt, C., et al.: Wonder3d: Single image to 3d using cross-domain diffusion. arXiv preprint arXiv:2310.15008 (2023) +27. Mechrez, R., Talmi, I., Zelnik-Manor, L.: The contextual loss for image transformation with non-aligned data. In: Proceedings of the European conference on computer vision (ECCV). pp. 768-783 (2018) +28. Melas-Kyriazi, L., Laina, I., Rupprecht, C., Vedaldi, A.: Realfusion: 360deg reconstruction of any object from a single image. In: CVPR (2023) +29. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. In: ECCV (2020) +30. Mokady, R., Hertz, A., Aberman, K., Pritch, Y., Cohen-Or, D.: Null-text inversion for editing real images using guided diffusion models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6038-6047 (2023) +31. Mou, C., Wang, X., Xie, L., Zhang, J., Qi, Z., Shan, Y., Qie, X.: T2i-adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models. arXiv preprint arXiv:2302.08453 (2023) +32. Müller, T., Evans, A., Schied, C., Keller, A.: Instant neural graphics primitives with a multiresolution hash encoding. ACM TOG (2022) +33. Nichol, A., Jun, H., Dhariwal, P., Mishkin, P., Chen, M.: Point-e: A system for generating 3d point clouds from complex prompts. arXiv preprint arXiv:2212.08751 (2022) + +34. Poole, B., Jain, A., Barron, J.T., Mildenhall, B.: Dreamfusion: Text-to-3d using 2d diffusion. arXiv preprint arXiv:2209.14988 (2022) +35. Qian, G., Mai, J., Hamdi, A., Ren, J., Siarohin, A., Li, B., Lee, H.Y., Skorokhodov, I., Wonka, P., Tulyakov, S., et al.: Magic123: One image to high-quality 3d object generation using both 2d and 3d diffusion priors. arXiv preprint arXiv:2306.17843 (2023) +36. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021) +37. Raj, A., Kaza, S., Poole, B., Niemeyer, M., Ruiz, N., Mildenhall, B., Zada, S., Aberman, K., Rubinstein, M., Barron, J., et al.: Dreambooth3d: Subject-driven text-to-3d generation. arXiv preprint arXiv:2303.13508 (2023) +38. Richardson, E., Metzer, G., Alaluf, Y., Giryes, R., Cohen-Or, D.: Texture: Text-guided texturing of 3d shapes. arXiv preprint arXiv:2302.01721 (2023) +39. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10684-10695 (2022) +40. Ruiz, N., Li, Y., Jampani, V., Pritch, Y., Rubinstein, M., Aberman, K.: Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 22500-22510 (2023) +41. Sara Fridovich-Keil and Alex Yu, Tancik, M., Chen, Q., Recht, B., Kanazawa, A.: Plenoxels: Radiance fields without neural networks. In: CVPR (2022) +42. Seo, H., Kim, H., Kim, G., Chun, S.Y.: Ditto-nerf: Diffusion-based iterative text to omni-directional 3d model. arXiv preprint arXiv:2304.02827 (2023) +43. Seo, J., Jang, W., Kwak, M.S., Ko, J., Kim, H., Kim, J., Kim, J.H., Lee, J., Kim, S.: Let 2d diffusion model know 3d-consistency for robust text-to-3d generation. arXiv preprint arXiv:2303.07937 (2023) +44. Shen, Q., Yang, X., Wang, X.: Anything-3d: Towards single-view anything reconstruction in the wild. arXiv preprint arXiv:2304.10261 (2023) +45. Shen, T., Gao, J., Yin, K., Liu, M.Y., Fidler, S.: Deep marching tetrahedra: a hybrid representation for high-resolution 3d shape synthesis. In: Advances in Neural Information Processing Systems (NeurIPS) (2021) +46. Shi, Y., Wang, P., Ye, J., Long, M., Li, K., Yang, X.: Mvdream: Multi-view diffusion for 3d generation. arXiv preprint arXiv:2308.16512 (2023) +47. Song, J., Meng, C., Ermon, S.: Denoising diffusion implicit models. arXiv:2010.02502 (October 2020), https://arxiv.org/abs/2010.02502 +48. Szymanowicz, S., Rupprecht, C., Vedaldi, A.: Viewset diffusion:(0-) image-conditioned 3d generative models from 2d data. arXiv preprint arXiv:2306.07881 (2023) +49. Tang, J.: Stable-dreamfusion: Text-to-3d with stable-diffusion (2022), https://github.com/ashawkey/stable-dreamfusion +50. Tang, J., Ren, J., Zhou, H., Liu, Z., Zeng, G.: Dreamgaussian: Generative gaussian splatting for efficient 3d content creation. arXiv preprint arXiv:2309.16653 (2023) +51. Tang, J., Wang, T., Zhang, B., Zhang, T., Yi, R., Ma, L., Chen, D.: Make-it-3d: High-fidelity 3d creation from a single image with diffusion prior. In: ICCV (2023) +52. Tsalicoglou, C., Manhardt, F., Tonioni, A., Niemeyer, M., Tombari, F.: Textmesh: Generation of realistic 3d meshes from text prompts. arXiv preprint arXiv:2304.12439 (2023) + +53. Wang, H., Du, X., Li, J., Yeh, R.A., Shakhnarovich, G.: Score jacobian chaining: Lifting pretrained 2d diffusion models for 3d generation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12619-12629 (2023) +54. Wang, T., Kanakis, M., Schindler, K., Van Gool, L., Obukhov, A.: Breathing new life into 3d assets with generative repainting. arXiv preprint arXiv:2309.08523 (2023) +55. Wang, Z., Lu, C., Wang, Y., Bao, F., Li, C., Su, H., Zhu, J.: Prolificdreamer: High-fidelity and diverse text-to-3d generation with variational score distillation. arXiv preprint arXiv:2305.16213 (2023) +56. Wu, J., Gao, X., Liu, X., Shen, Z., Zhao, C., Feng, H., Liu, J., Ding, E.: Hd-fusion: Detailed text-to-3d generation leveraging multiple noise estimation. arXiv preprint arXiv:2307.16183 (2023) +57. Xu, D., Jiang, Y., Wang, P., Fan, Z., Wang, Y., Wang, Z.: Neurallift-360: Lifting an in-the-wild 2d photo to a 3d object with 360 views. arXiv e-prints pp. arXiv-2211 (2022) +58. Ye, H., Zhang, J., Liu, S., Han, X., Yang, W.: Ip-adapter: Text compatible image prompt adapter for text-to-image diffusion models (2023) +59. Yi, T., Fang, J., Wu, G., Xie, L., Zhang, X., Liu, W., Tian, Q., Wang, X.: Gaussian dreamer: Fast generation from text to 3d gaussian splatting with point cloud priors. arXiv preprint arXiv:2310.08529 (2023) +60. Yu, C., Zhou, Q., Li, J., Zhang, Z., Wang, Z., Wang, F.: Points-to-3d: Bridging the gap between sparse points and shape-controllable text-to-3d generation. arXiv preprint arXiv:2307.13908 (2023) +61. Yu, W., Yuan, L., Cao, Y.P., Gao, X., Li, X., Quan, L., Shan, Y., Tian, Y.: Hifi-123: Towards high-fidelity one image to 3d content generation. arXiv preprint arXiv:2310.06744 (2023) +62. Zhang, L., Rao, A., Agrawala, M.: Adding conditional control to text-to-image diffusion models. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 3836-3847 (2023) +63. Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual metric. In: CVPR (2018) +64. Zhu, J., Zhuang, P.: Hifa: High-fidelity text-to-3d with advanced diffusion guidance. arXiv preprint arXiv:2305.18766 (2023) \ No newline at end of file diff --git a/2024/Repaint123_ Fast and High-quality One Image to 3D Generation with Progressive Controllable Repainting/images.zip b/2024/Repaint123_ Fast and High-quality One Image to 3D Generation with Progressive Controllable Repainting/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..5e1a2ca1d2a7390133d6a78ba7cbdb2aca06d1aa --- /dev/null +++ b/2024/Repaint123_ Fast and High-quality One Image to 3D Generation with Progressive Controllable Repainting/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5efef3d31330ed58e1ac0e35d047a9b139c0d0006a1f7fa3f894290d5fa799f5 +size 561283 diff --git a/2024/Repaint123_ Fast and High-quality One Image to 3D Generation with Progressive Controllable Repainting/layout.json b/2024/Repaint123_ Fast and High-quality One Image to 3D Generation with Progressive Controllable Repainting/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..21528dec1c7e3e490f14df4827ff62e5974bb0c2 --- /dev/null +++ b/2024/Repaint123_ Fast and High-quality One Image to 3D Generation with Progressive Controllable Repainting/layout.json @@ -0,0 +1,8676 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 136, + 111, + 478, + 166 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 111, + 478, + 166 + ], + "spans": [ + { + "bbox": [ + 136, + 111, + 478, + 166 + ], + "type": "text", + "content": "Repaint123: Fast and High-quality One Image to 3D Generation with Progressive Controllable Repainting" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 136, + 186, + 478, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 186, + 478, + 236 + ], + "spans": [ + { + "bbox": [ + 136, + 186, + 478, + 236 + ], + "type": "text", + "content": "Junwu Zhang1* Zhenyu Tang1* Yatian Pang1,3 Xinhua Cheng1 Peng Jin1 Yida Wei4 Xing Zhou5 Munan Ning1,2 Li Yuan1,2* \n1Peking University 2Pengcheng Laboratory 3National University of Singapore \n45RabbitPre" + } + ] + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 133, + 269, + 480, + 415 + ], + "blocks": [ + { + "bbox": [ + 133, + 269, + 480, + 415 + ], + "lines": [ + { + "bbox": [ + 133, + 269, + 480, + 415 + ], + "spans": [ + { + "bbox": [ + 133, + 269, + 480, + 415 + ], + "type": "image", + "image_path": "7dfab4bc6fd25e686d0908268ecef8871845a43d7f8ede6eaf58479ad8b584e7.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 423, + 482, + 468 + ], + "lines": [ + { + "bbox": [ + 130, + 423, + 482, + 468 + ], + "spans": [ + { + "bbox": [ + 130, + 423, + 482, + 468 + ], + "type": "text", + "content": "Fig. 1: Repaint123 generates high-quality 3D content with detailed texture from a single image in 2 minutes. Repaint123 adopts 3D Gaussian Splatting in the coarse stage and then utilizes a progressively repairing strategy with the diffusion model for high-quality 3D content generation with efficiency." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 159, + 513, + 455, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 513, + 455, + 635 + ], + "spans": [ + { + "bbox": [ + 159, + 513, + 455, + 635 + ], + "type": "text", + "content": "Abstract. Recent image-to-3D methods achieve impressive results with plausible 3D geometry due to the development of diffusion models and optimization techniques. However, existing image-to-3D methods suffer from texture deficiencies in novel views, including multi-view inconsistency and quality degradation. To alleviate multi-view bias and enhance image quality in novel-view textures, we present Repaint123, a fast image-to-3D approach for creating high-quality 3D content with detailed textures. Repaint123 proposes a progressively repairing strategy to simultaneously enhance the consistency and quality of textures across different views, generating invisible regions according to visible textures, with the visibility map calculated by the depth alignment across views." + } + ] + } + ], + "index": 4 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 643, + 223, + 653 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 643, + 223, + 653 + ], + "spans": [ + { + "bbox": [ + 133, + 643, + 223, + 653 + ], + "type": "text", + "content": "* Equal contribution." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 654, + 235, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 654, + 235, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 654, + 235, + 665 + ], + "type": "text", + "content": "\\*\\* Corresponding author." + } + ] + } + ], + "index": 6 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 159, + 117, + 455, + 237 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 117, + 455, + 237 + ], + "spans": [ + { + "bbox": [ + 159, + 117, + 455, + 237 + ], + "type": "text", + "content": "Furthermore, multiple control techniques, including reference-driven information injection and coarse-based depth guidance, are introduced to alleviate the texture bias accumulated during the repainting process for improved consistency and quality. For novel-view texture refinement with short-term view consistency, our method progressively repairs novel-view images with adaptive strengths based on visibility, enhancing the balance of image quality and view consistency. To alleviate the accumulated bias as progressively repainting, we control the repainting process by depth-guided geometry and attention-driven reference-view textures. Extensive experiments demonstrate the superior ability of our method to create 3D content with consistent and detailed textures in 2 minutes." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 160, + 248, + 435, + 259 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 248, + 435, + 259 + ], + "spans": [ + { + "bbox": [ + 160, + 248, + 435, + 259 + ], + "type": "text", + "content": "Keywords: Image-to-3D " + }, + { + "bbox": [ + 160, + 248, + 435, + 259 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 248, + 435, + 259 + ], + "type": "text", + "content": " Texture Enhancement " + }, + { + "bbox": [ + 160, + 248, + 435, + 259 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 248, + 435, + 259 + ], + "type": "text", + "content": " Diffusion model" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 281, + 229, + 293 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 281, + 229, + 293 + ], + "spans": [ + { + "bbox": [ + 132, + 281, + 229, + 293 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 306, + 299, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 306, + 299, + 593 + ], + "spans": [ + { + "bbox": [ + 130, + 306, + 299, + 593 + ], + "type": "text", + "content": "Generating 3D content from one given reference image plays a key role at the intersection of computer vision and computer graphics [11, 17, 24, 25, 33, 35], which is desired by users for innovative applications across fields including robotics, virtual reality, and augmented reality. However, the image-to-3D task is quite challenging due to the generated 3D content is expected with reasonable geometry and consistent textures, while the single image input is insufficient in multi-view information. Recent studies [24, 28, 35, 50, 51] employ 2D diffusion models [14, 39] to guide 3D generation from the reference image with Score Distillation Sampling (SDS) [34]. While preliminary attempts [28, 51] leveraging view-independent diffusion models (e.g. Stable Diffusion [39]) faced challenges with over-saturated textures and multi-face geometry, more" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": "recent efforts [24, 26, 35, 46, 50] utilize view-conditioned diffusion models [24, 26, 46]. Although these methods utilizing view-conditioned diffusion models mitigate the multi-face problem and generate 3D content with plausible geometry, they still encounter challenges in generating high-quality textures. This limitation stems from training these models on small-scale synthetic 3D datasets, hindering their ability to produce detailed images from novel views." + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 306, + 331, + 482, + 475 + ], + "blocks": [ + { + "bbox": [ + 306, + 331, + 482, + 475 + ], + "lines": [ + { + "bbox": [ + 306, + 331, + 482, + 475 + ], + "spans": [ + { + "bbox": [ + 306, + 331, + 482, + 475 + ], + "type": "image", + "image_path": "de3c7f701ab1e3f72825b2f035b8dab4103ed427b498cf030a640de73bdc38cd.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 483, + 482, + 571 + ], + "lines": [ + { + "bbox": [ + 304, + 483, + 482, + 571 + ], + "spans": [ + { + "bbox": [ + 304, + 483, + 482, + 571 + ], + "type": "text", + "content": "Fig.2: Motivation. Current image-to-3D methods adopt SDS loss for generating, resulting in inconsistent and poor textures. Repaint123 proposes a progressively repainting strategy based on visible content to efficiently create invisible content with MSE loss, achieving consistent and high-quality textures." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "type": "text", + "content": "Zhang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 481, + 342 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 481, + 342 + ], + "type": "text", + "content": "To address textural deficiencies including inconsistency and low quality mentioned above, we propose a novel approach named Repaint123 for fast and high-quality image-to-3d generation by generating consistent and high-quality novel-view images with a controllable content repaint process. As shown in Figure 2, the core of Repaint123 is to progressively repaint the invisible content with a view-independent diffusion model for adjacent-view consistency, incorporating multiple controls of geometry and texture for enhanced quality and multi-view consistency. Specifically, our method adopts a two-stage optimization framework. In the coarse stage, we utilize 3D Gaussian Splatting [18] as the representation to efficiently obtain a coarse 3D model in 1 minute. In the fine stage, we progressively sample the camera views around and repaint invisible content with control techniques including using novel-view depth map as geometry prior, injecting reference-view attention features as texture prior [61] and employing reference-view image prompt as semantic prior. To further enhance image quality and preserve multi-view consistency, we propose a visibility-based adaptive strategy that refines the visible regions with different strengths. After achieving high-qualityainted images with multi-view consistency, we directly leverage simple MSE loss to efficiently refine 3D content textures instead of the time-consuming and over-smoothed SDS." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 343, + 479, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 343, + 479, + 415 + ], + "spans": [ + { + "bbox": [ + 130, + 343, + 479, + 415 + ], + "type": "text", + "content": "We conduct extensive experiments on multiple datasets and demonstrate that our method significantly advances image-to-3D generation, producing high-quality, multi-view consistent 3D content with detailed textures in approximately 2 minutes from scratch (as shown in Figure 1), outperforming current state-of-the-art NeRF-based and Gaussian-Splatting-based methods in both consistency and texture quality. Our contributions can be summarized as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 422, + 480, + 517 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 138, + 422, + 479, + 458 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 422, + 479, + 458 + ], + "spans": [ + { + "bbox": [ + 138, + 422, + 479, + 458 + ], + "type": "text", + "content": "- We propose a progressively repairing strategy based on visibility for the image-to-3D task to generate 3D content with consistency and high quality by repairing invisible parts according to visible textures." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 459, + 480, + 494 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 459, + 480, + 494 + ], + "spans": [ + { + "bbox": [ + 138, + 459, + 480, + 494 + ], + "type": "text", + "content": "- We introduce multiple control techniques including reference-driven information injection and coarse-based depth guidance to alleviate the texture bias accumulated during the repainting process for further quality enhancement." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 495, + 479, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 495, + 479, + 517 + ], + "spans": [ + { + "bbox": [ + 138, + 495, + 479, + 517 + ], + "type": "text", + "content": "- Our experiments demonstrate Repaint123 efficiently generates high-quality 3D content with consistent and detailed textures in 2 minutes." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 132, + 536, + 242, + 548 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 536, + 242, + 548 + ], + "spans": [ + { + "bbox": [ + 132, + 536, + 242, + 548 + ], + "type": "text", + "content": "2 Related Works" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 562, + 342, + 573 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 562, + 342, + 573 + ], + "spans": [ + { + "bbox": [ + 132, + 562, + 342, + 573 + ], + "type": "text", + "content": "2.1 Diffusion Models for 3D Generation" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 581, + 480, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 581, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 581, + 480, + 665 + ], + "type": "text", + "content": "The recent notable achievements in view-independent 2D diffusion models [14,39] have brought about exciting prospects for generating 3D objects. Pioneering studies [34, 53] have introduced the concept of distilling a 2D text-to-image generation model to generate 3D shapes. Subsequent works [1, 6, 7, 10, 16, 22, 28, 35, 37, 42-44, 51, 52, 55-57, 60, 61, 64] have adopted a similar per-object optimization approach, building upon these initial works. Nevertheless, the majority of these techniques consistently experience low efficiency and multi-face" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 173, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 173, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 173, + 91, + 447, + 102 + ], + "type": "text", + "content": "Repaint123: Progressive Controllable Repainting for 3D Generation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 248 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 248 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 248 + ], + "type": "text", + "content": "issues. Unlike the previous study HiFi-123 [61], which employed similar inversion and attention injection techniques for image-to-3D generation, our approach diverges primarily in our proposed repairing strategy, as well as many variations in optimization objectives, image prompting, and 3D representations. Recently, some works [23, 25, 26, 48] extend view-independent diffusion model to view-conditioned diffusion model to generate multi-view images for reconstruction, while these methods usually suffer from low-quality textures as the multi-view diffusion models are trained on limited and synthesized data. To optimize both consistency and quality, we integrate a view-conditioned diffusion model for initial consistency in the coarse stage and a view-independent model to enhance texture quality in the fine stage." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 266, + 308, + 278 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 266, + 308, + 278 + ], + "spans": [ + { + "bbox": [ + 131, + 266, + 308, + 278 + ], + "type": "text", + "content": "2.2 Controllable Image Synthesis" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 286, + 482, + 406 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 286, + 482, + 406 + ], + "spans": [ + { + "bbox": [ + 130, + 286, + 482, + 406 + ], + "type": "text", + "content": "A major challenge in the field of image generation is achieving controllability. Many works have been done recently to increase the controllability of images generated by diffusion models. ControlNet [62] and T2I-Adapter [31] attempt to control the creation of images by utilizing data from different modalities. Besides, some optimization-based methods [12,30,40] learn new parameters or fine-tune the diffusion model in order to enhance control over the generation process. Other methods [3,58] leverage multi-view attention to introduce information from other-view images for gaining superior control. We incorporate multiple control techniques into the view-independent diffusion model for maintaining view consistency in the fine stage." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 424, + 261, + 436 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 424, + 261, + 436 + ], + "spans": [ + { + "bbox": [ + 131, + 424, + 261, + 436 + ], + "type": "text", + "content": "2.3 3D Representations" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 444, + 482, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 444, + 482, + 553 + ], + "spans": [ + { + "bbox": [ + 130, + 444, + 482, + 553 + ], + "type": "text", + "content": "Neural Radiance Fields (NeRF) [29], as a volumetric rendering method, has gained popularity for its ability to enable 3D optimization [2,4,8,13,21] under 2D supervision, while NeRF optimization can be time-consuming. Numerous efforts [32,41] for spatial pruning have been dedicated to accelerating the training process of NeRF on the reconstruction setting. However, they fail in the generation setting. Recently, 3D Gaussian Splatting [5,9,18,50,59] has emerged as an alternative 3D representation to NeRF and has shown remarkable advancements in terms of both quality and speed, offering a promising avenue. Therefore, we adopt Gaussian Splatting for efficiently generating coarse 3D model." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 571, + 227, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 571, + 227, + 586 + ], + "spans": [ + { + "bbox": [ + 131, + 571, + 227, + 586 + ], + "type": "text", + "content": "3 Preliminary" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 597, + 246, + 609 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 597, + 246, + 609 + ], + "spans": [ + { + "bbox": [ + 131, + 597, + 246, + 609 + ], + "type": "text", + "content": "3.1 DDIM Inversion" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 617, + 482, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 667 + ], + "type": "text", + "content": "DDIM [47] transforms random noise " + }, + { + "bbox": [ + 130, + 617, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\pmb{x}_T" + }, + { + "bbox": [ + 130, + 617, + 482, + 667 + ], + "type": "text", + "content": " into clean data " + }, + { + "bbox": [ + 130, + 617, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\pmb{x}_0" + }, + { + "bbox": [ + 130, + 617, + 482, + 667 + ], + "type": "text", + "content": " over a series of time steps, by using the deterministic DDIM sampling in the reverse process, i.e., " + }, + { + "bbox": [ + 130, + 617, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\pmb{x}_{t - 1} = (\\alpha_{t - 1} / \\alpha_t)(\\pmb{x}_t - \\sigma_t\\epsilon_\\phi) + \\sigma_{t - 1}\\epsilon_\\phi" + }, + { + "bbox": [ + 130, + 617, + 482, + 667 + ], + "type": "text", + "content": ". On the contrary, DDIM inversion progressively converts clean data to a noisy state " + }, + { + "bbox": [ + 130, + 617, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\pmb{x}_T" + }, + { + "bbox": [ + 130, + 617, + 482, + 667 + ], + "type": "text", + "content": ", i.e., " + }, + { + "bbox": [ + 130, + 617, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\pmb{x}_t = (\\alpha_t / \\alpha_{t - 1})(\\pmb{x}_{t - 1} - \\sigma_t\\epsilon_\\phi)" + }, + { + "bbox": [ + 130, + 617, + 482, + 667 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "type": "text", + "content": "Zhang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 134, + 118, + 479, + 235 + ], + "blocks": [ + { + "bbox": [ + 134, + 118, + 479, + 235 + ], + "lines": [ + { + "bbox": [ + 134, + 118, + 479, + 235 + ], + "spans": [ + { + "bbox": [ + 134, + 118, + 479, + 235 + ], + "type": "image", + "image_path": "f803aef4bd19c57067c8f2e5580f4c5a5ecb736b5efa92be62e662497d20c5be.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 245, + 482, + 323 + ], + "lines": [ + { + "bbox": [ + 130, + 245, + 482, + 323 + ], + "spans": [ + { + "bbox": [ + 130, + 245, + 482, + 323 + ], + "type": "text", + "content": "Fig. 3: Overview of Repaint123. In the coarse stage, we leverage 3D Gaussian Splatting to represent the coarse content, which is optimized by MSE loss at the reference and SDS loss at the invisible views. In the fine stage, we convert coarse content to mesh and progressively sample and generate invisible views from the reference view bidirectionally. Concretely, the current view is repainted based on the visibility map from bidirectional neighbor views with control techniques including attention injection and depth guidance, leading to consistent and detailed textures." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 346, + 482, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 346, + 482, + 384 + ], + "spans": [ + { + "bbox": [ + 130, + 346, + 482, + 384 + ], + "type": "inline_equation", + "content": "\\sigma_{t-1}\\epsilon_{\\phi}) + \\sigma_t\\epsilon_{\\phi}" + }, + { + "bbox": [ + 130, + 346, + 482, + 384 + ], + "type": "text", + "content": ", here " + }, + { + "bbox": [ + 130, + 346, + 482, + 384 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\phi}" + }, + { + "bbox": [ + 130, + 346, + 482, + 384 + ], + "type": "text", + "content": " is the predicted noise by the UNet. This method can precisely reconstruct the original clean data while greatly speeding up the process by skipping many intermediate diffusion steps." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 400, + 275, + 414 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 400, + 275, + 414 + ], + "spans": [ + { + "bbox": [ + 131, + 400, + 275, + 414 + ], + "type": "text", + "content": "3.2 3D Gaussian Splitting" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 420, + 482, + 492 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 420, + 482, + 492 + ], + "spans": [ + { + "bbox": [ + 130, + 420, + 482, + 492 + ], + "type": "text", + "content": "Gaussian Splatting [18] presents a novel method for synthesizing new views and reconstructing 3D scenes, achieving real-time speed. Unlike NeRF, Gaussian Splatting uses a set of anisotropic 3D Gaussians defined by their locations, covariances, colors, and opacities to represent the scene. To compute the color of each pixel " + }, + { + "bbox": [ + 130, + 420, + 482, + 492 + ], + "type": "inline_equation", + "content": "\\mathbf{p}" + }, + { + "bbox": [ + 130, + 420, + 482, + 492 + ], + "type": "text", + "content": " in the image, it utilizes a typical neural point-based rendering [19,20], The rendering process is as follows:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 241, + 502, + 481, + 536 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 241, + 502, + 481, + 536 + ], + "spans": [ + { + "bbox": [ + 241, + 502, + 481, + 536 + ], + "type": "interline_equation", + "content": "C (\\mathbf {p}) = \\sum_ {i \\in \\mathcal {N}} c _ {i} \\alpha_ {i} \\prod_ {j = 1} ^ {i - 1} (1 - \\alpha_ {j}), \\tag {1}", + "image_path": "c5910d09407f008264f8f03015acf505056d975c29569b5f3964659d0ff7f829.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 228, + 537, + 386, + 553 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 228, + 537, + 386, + 553 + ], + "spans": [ + { + "bbox": [ + 228, + 537, + 386, + 553 + ], + "type": "interline_equation", + "content": "\\mathrm {w h e r e}, \\alpha_ {i} = o _ {i} e ^ {- \\frac {1}{2} (\\mathbf {p} - \\boldsymbol {\\mu} _ {i}) ^ {T} \\boldsymbol {\\Sigma} _ {i} ^ {- 1} (\\mathbf {p} - \\boldsymbol {\\mu} _ {i})},", + "image_path": "5855e90ca11d8e728e2aca61264b27185eba87a8e78d4b56709a0612f4fd7f59.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 562, + 482, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 562, + 482, + 586 + ], + "spans": [ + { + "bbox": [ + 131, + 562, + 482, + 586 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 562, + 482, + 586 + ], + "type": "inline_equation", + "content": "c_{i}" + }, + { + "bbox": [ + 131, + 562, + 482, + 586 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 131, + 562, + 482, + 586 + ], + "type": "inline_equation", + "content": "o_{i}" + }, + { + "bbox": [ + 131, + 562, + 482, + 586 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 131, + 562, + 482, + 586 + ], + "type": "inline_equation", + "content": "\\mu_{i}" + }, + { + "bbox": [ + 131, + 562, + 482, + 586 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 131, + 562, + 482, + 586 + ], + "type": "inline_equation", + "content": "\\Sigma_{i}" + }, + { + "bbox": [ + 131, + 562, + 482, + 586 + ], + "type": "text", + "content": " represent the color, opacity, position, and covariance of the " + }, + { + "bbox": [ + 131, + 562, + 482, + 586 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 131, + 562, + 482, + 586 + ], + "type": "text", + "content": "-th Gaussian respectively, and " + }, + { + "bbox": [ + 131, + 562, + 482, + 586 + ], + "type": "inline_equation", + "content": "\\mathcal{N}" + }, + { + "bbox": [ + 131, + 562, + 482, + 586 + ], + "type": "text", + "content": " denotes the number of related Gaussians." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 604, + 202, + 616 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 604, + 202, + 616 + ], + "spans": [ + { + "bbox": [ + 131, + 604, + 202, + 616 + ], + "type": "text", + "content": "4 Method" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 629, + 481, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 481, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 481, + 667 + ], + "type": "text", + "content": "In this section, we introduce our two-stage framework named Repaint123 for fast and high-quality image-to-3D generation, as illustrated in Figure 3. In the coarse stage, we adopt 3D Gaussian Splitting [18] representation optimized by" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 173, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 173, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 173, + 91, + 447, + 102 + ], + "type": "text", + "content": "Repaint123: Progressive Controllable Repainting for 3D Generation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 297 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 297 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 297 + ], + "type": "text", + "content": "SDS loss [34] with the view-conditioned diffusion model [24] to learn a coarse geometry and texture (see Section 4.1). Inspired by [50], we convert the coarse 3D Gaussians to textured mesh to facilitate subsequent texture enhancement in UV space. In the fine stage, we progressively paint invisible content in novel views by rotating camera viewpoints bi-directionally from the reference view (see Section 4.2). To mitigate texture bias accumulated during the gradual painting across wide angles, we incorporate multiple control techniques including depth guidance, injection of reference-view attention, and the use of reference-view image prompts (see Section 4.2). Additionally, we propose a visibility-based adaptive strategy to refine visible regions when observed from a superior perspective (see Section 4.2). After obtaining a high-quality painted image that remains consistent across different views, we efficiently optimize the texture map by simple Mean Square Error (MSE) loss (see Section 4.2). This process is repeated for each sampled view, alternating between painting and optimization, until the texture is completely reconstructed to accommodate a 360-degree perspective." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 313, + 451, + 326 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 313, + 451, + 326 + ], + "spans": [ + { + "bbox": [ + 131, + 313, + 451, + 326 + ], + "type": "text", + "content": "4.1 Coarse Stage: Gaussian Splatting with 3D Diffusion Prior" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 332, + 482, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 332, + 482, + 422 + ], + "spans": [ + { + "bbox": [ + 130, + 332, + 482, + 422 + ], + "type": "text", + "content": "In the coarse stage, we adopt 3D Gaussian Splatting for efficient initialization. Through SDS [34], 3D diffusion priors are back-propagated to the 3D Gaussians. At each diffusion step, we sample a random current camera view " + }, + { + "bbox": [ + 130, + 332, + 482, + 422 + ], + "type": "inline_equation", + "content": "v_{c}" + }, + { + "bbox": [ + 130, + 332, + 482, + 422 + ], + "type": "text", + "content": " and its camera pose " + }, + { + "bbox": [ + 130, + 332, + 482, + 422 + ], + "type": "inline_equation", + "content": "p_{c}" + }, + { + "bbox": [ + 130, + 332, + 482, + 422 + ], + "type": "text", + "content": ", which orbits around the center of the object. Subsequently, the RGB image, denoted as " + }, + { + "bbox": [ + 130, + 332, + 482, + 422 + ], + "type": "inline_equation", + "content": "I_{\\mathrm{rgb}}^{c}" + }, + { + "bbox": [ + 130, + 332, + 482, + 422 + ], + "type": "text", + "content": ", and the alpha image, represented by " + }, + { + "bbox": [ + 130, + 332, + 482, + 422 + ], + "type": "inline_equation", + "content": "I_{\\mathrm{a}}^{c}" + }, + { + "bbox": [ + 130, + 332, + 482, + 422 + ], + "type": "text", + "content": ", are rendered from the current viewpoint " + }, + { + "bbox": [ + 130, + 332, + 482, + 422 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 130, + 332, + 482, + 422 + ], + "type": "text", + "content": ". Given an image " + }, + { + "bbox": [ + 130, + 332, + 482, + 422 + ], + "type": "inline_equation", + "content": "\\tilde{I}_{\\mathrm{rgb}}^{r}" + }, + { + "bbox": [ + 130, + 332, + 482, + 422 + ], + "type": "text", + "content": " and an alpha mask " + }, + { + "bbox": [ + 130, + 332, + 482, + 422 + ], + "type": "inline_equation", + "content": "\\tilde{I}_{\\mathrm{a}}^{r}" + }, + { + "bbox": [ + 130, + 332, + 482, + 422 + ], + "type": "text", + "content": " as input to the reference view, the SDS loss can be formulated as:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 205, + 430, + 481, + 458 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 205, + 430, + 481, + 458 + ], + "spans": [ + { + "bbox": [ + 205, + 430, + 481, + 458 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {S D S}} = \\mathbb {E} _ {t, c, \\epsilon} \\left[ (\\epsilon_ {\\phi} (I _ {\\mathrm {r g b}} ^ {c}; t, \\tilde {I} _ {\\mathrm {r g b}} ^ {r}, \\Delta p _ {c}) - \\epsilon) \\frac {\\partial I _ {\\mathrm {r g b}} ^ {c}}{\\partial \\Theta} \\right] \\tag {2}", + "image_path": "9b4bdd685b5c964200931336ecd686e5c3341f2f5fa3bed2c57dd8db0e2b1ea6.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 464, + 482, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 464, + 482, + 514 + ], + "spans": [ + { + "bbox": [ + 130, + 464, + 482, + 514 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 464, + 482, + 514 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\phi}(\\cdot)" + }, + { + "bbox": [ + 130, + 464, + 482, + 514 + ], + "type": "text", + "content": " is the predicted noise by the diffusion " + }, + { + "bbox": [ + 130, + 464, + 482, + 514 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 130, + 464, + 482, + 514 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 464, + 482, + 514 + ], + "type": "inline_equation", + "content": "\\Delta p_{c}" + }, + { + "bbox": [ + 130, + 464, + 482, + 514 + ], + "type": "text", + "content": " is the relative camera pose and " + }, + { + "bbox": [ + 130, + 464, + 482, + 514 + ], + "type": "inline_equation", + "content": "\\Theta" + }, + { + "bbox": [ + 130, + 464, + 482, + 514 + ], + "type": "text", + "content": " represents the estimizable Gaussian parameters. Additionally, we optimize the RGB image and alpha image rendered from the reference view to align with the input reference image using reconstruction loss:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 219, + 521, + 481, + 536 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 219, + 521, + 481, + 536 + ], + "spans": [ + { + "bbox": [ + 219, + 521, + 481, + 536 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {R e f}} = \\lambda_ {\\mathrm {r g b}} | | I _ {\\mathrm {r g b}} ^ {r} - \\tilde {I} _ {\\mathrm {r g b}} ^ {r} | | _ {2} ^ {2} + | | I _ {\\mathrm {a}} ^ {r} - \\tilde {I} _ {\\mathrm {a}} ^ {r} | | _ {2} ^ {2}. \\qquad (3)", + "image_path": "e01d255ea3579c1e8e020241b5454b108f86158e3a69388fb4addffcfff95c04.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 544, + 403, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 544, + 403, + 556 + ], + "spans": [ + { + "bbox": [ + 131, + 544, + 403, + 556 + ], + "type": "text", + "content": "The final loss function is the weighted sum of " + }, + { + "bbox": [ + 131, + 544, + 403, + 556 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{SDS}}" + }, + { + "bbox": [ + 131, + 544, + 403, + 556 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 544, + 403, + 556 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{Ref}}" + }, + { + "bbox": [ + 131, + 544, + 403, + 556 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 574, + 424, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 574, + 424, + 586 + ], + "spans": [ + { + "bbox": [ + 131, + 574, + 424, + 586 + ], + "type": "text", + "content": "4.2 Fine Stage: Visibility-based Controllable Repainting" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 594, + 482, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 667 + ], + "type": "text", + "content": "Progressive Invisible Content Repainting. We repaint previously unseen content with progressively sampling novel camera viewpoints, beginning from the reference view and increasing a certain angular interval with each iteration. To obtain invisible regions in the image from a newly sampled view " + }, + { + "bbox": [ + 130, + 594, + 482, + 667 + ], + "type": "inline_equation", + "content": "v_{c}" + }, + { + "bbox": [ + 130, + 594, + 482, + 667 + ], + "type": "text", + "content": ", we use the back-projection technique based on the depth map " + }, + { + "bbox": [ + 130, + 594, + 482, + 667 + ], + "type": "inline_equation", + "content": "D^{p}\\in \\mathbb{R}^{H\\times W}" + }, + { + "bbox": [ + 130, + 594, + 482, + 667 + ], + "type": "text", + "content": " in the previously seen view " + }, + { + "bbox": [ + 130, + 594, + 482, + 667 + ], + "type": "inline_equation", + "content": "v_{p}" + }, + { + "bbox": [ + 130, + 594, + 482, + 667 + ], + "type": "text", + "content": ". Formally, given the camera intrinsics " + }, + { + "bbox": [ + 130, + 594, + 482, + 667 + ], + "type": "inline_equation", + "content": "K\\in \\mathbb{R}^{3\\times 3}" + }, + { + "bbox": [ + 130, + 594, + 482, + 667 + ], + "type": "text", + "content": ", we first remap the" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "type": "text", + "content": "Zhang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 131, + 116, + 482, + 166 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 116, + 482, + 166 + ], + "spans": [ + { + "bbox": [ + 131, + 116, + 482, + 166 + ], + "type": "text", + "content": "pixels of the image rendered from " + }, + { + "bbox": [ + 131, + 116, + 482, + 166 + ], + "type": "inline_equation", + "content": "v_{p}" + }, + { + "bbox": [ + 131, + 116, + 482, + 166 + ], + "type": "text", + "content": " into 3D points " + }, + { + "bbox": [ + 131, + 116, + 482, + 166 + ], + "type": "inline_equation", + "content": "X^{p}" + }, + { + "bbox": [ + 131, + 116, + 482, + 166 + ], + "type": "text", + "content": " expressed in the camera coordinate frame, which can be obtained as " + }, + { + "bbox": [ + 131, + 116, + 482, + 166 + ], + "type": "inline_equation", + "content": "X_{i,j} = K^{-1}[iD_{i,j},jD_{i,j},D_{i,j}]^{\\top}" + }, + { + "bbox": [ + 131, + 116, + 482, + 166 + ], + "type": "text", + "content": ". With " + }, + { + "bbox": [ + 131, + 116, + 482, + 166 + ], + "type": "inline_equation", + "content": "P_{p},P_{c}\\in \\mathbb{R}^{3\\times 4}" + }, + { + "bbox": [ + 131, + 116, + 482, + 166 + ], + "type": "text", + "content": " the world-to-camera poses for images from " + }, + { + "bbox": [ + 131, + 116, + 482, + 166 + ], + "type": "inline_equation", + "content": "v_{p}" + }, + { + "bbox": [ + 131, + 116, + 482, + 166 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 116, + 482, + 166 + ], + "type": "inline_equation", + "content": "v_{c}" + }, + { + "bbox": [ + 131, + 116, + 482, + 166 + ], + "type": "text", + "content": ", we can transform " + }, + { + "bbox": [ + 131, + 116, + 482, + 166 + ], + "type": "inline_equation", + "content": "X^{p}" + }, + { + "bbox": [ + 131, + 116, + 482, + 166 + ], + "type": "text", + "content": " into camera coordinate frame of " + }, + { + "bbox": [ + 131, + 116, + 482, + 166 + ], + "type": "inline_equation", + "content": "v_{c}" + }, + { + "bbox": [ + 131, + 116, + 482, + 166 + ], + "type": "text", + "content": " to obtain " + }, + { + "bbox": [ + 131, + 116, + 482, + 166 + ], + "type": "inline_equation", + "content": "X^{p,c}" + }, + { + "bbox": [ + 131, + 116, + 482, + 166 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 261, + 175, + 481, + 190 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 261, + 175, + 481, + 190 + ], + "spans": [ + { + "bbox": [ + 261, + 175, + 481, + 190 + ], + "type": "interline_equation", + "content": "X ^ {p, c} = P _ {p} P _ {c} ^ {- 1} h \\left(X ^ {p}\\right), \\tag {4}", + "image_path": "01048800d820eafe419384938a1803917d4db9955553decfeb2728f92f04df7f.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 198, + 482, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 198, + 482, + 258 + ], + "spans": [ + { + "bbox": [ + 131, + 198, + 482, + 258 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 198, + 482, + 258 + ], + "type": "inline_equation", + "content": "h:(x,y,z)\\to (x,y,z,1)" + }, + { + "bbox": [ + 131, + 198, + 482, + 258 + ], + "type": "text", + "content": " is the homogeneous mapping. Then we can render a depth map " + }, + { + "bbox": [ + 131, + 198, + 482, + 258 + ], + "type": "inline_equation", + "content": "D^{p,c}" + }, + { + "bbox": [ + 131, + 198, + 482, + 258 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 131, + 198, + 482, + 258 + ], + "type": "inline_equation", + "content": "X^{p,c}" + }, + { + "bbox": [ + 131, + 198, + 482, + 258 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 131, + 198, + 482, + 258 + ], + "type": "inline_equation", + "content": "D_{i,j}^{p,c} = KX_{i,j,2}^{p,c}" + }, + { + "bbox": [ + 131, + 198, + 482, + 258 + ], + "type": "text", + "content": ". Regions with dissimilar depth values between " + }, + { + "bbox": [ + 131, + 198, + 482, + 258 + ], + "type": "inline_equation", + "content": "D^c" + }, + { + "bbox": [ + 131, + 198, + 482, + 258 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 198, + 482, + 258 + ], + "type": "inline_equation", + "content": "D^{p,c}" + }, + { + "bbox": [ + 131, + 198, + 482, + 258 + ], + "type": "text", + "content": " are recognized as invisible regions, forming a repainting mask " + }, + { + "bbox": [ + 131, + 198, + 482, + 258 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 131, + 198, + 482, + 258 + ], + "type": "text", + "content": " where areas with zeros are invisible and areas with ones are visible. Therefore, we repaint invisible content and maintain visible content by:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 236, + 268, + 481, + 282 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 236, + 268, + 481, + 282 + ], + "spans": [ + { + "bbox": [ + 236, + 268, + 481, + 282 + ], + "type": "interline_equation", + "content": "x _ {t} = x _ {t} ^ {\\mathrm {i n v}} \\odot M + x _ {t} ^ {\\mathrm {r e v}} \\odot (1 - M), \\tag {5}", + "image_path": "e05e63936ea53b586af6b98e2b3378f56815c0c34f7d47db2e35322d8b45ea53.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 291, + 482, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 291, + 482, + 327 + ], + "spans": [ + { + "bbox": [ + 131, + 291, + 482, + 327 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 291, + 482, + 327 + ], + "type": "inline_equation", + "content": "\\odot" + }, + { + "bbox": [ + 131, + 291, + 482, + 327 + ], + "type": "text", + "content": " is element-wise matrix multiplication, " + }, + { + "bbox": [ + 131, + 291, + 482, + 327 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 131, + 291, + 482, + 327 + ], + "type": "text", + "content": " is the current denoising timestep, " + }, + { + "bbox": [ + 131, + 291, + 482, + 327 + ], + "type": "inline_equation", + "content": "x_{t}^{\\mathrm{inv}}" + }, + { + "bbox": [ + 131, + 291, + 482, + 327 + ], + "type": "text", + "content": " is the DDIM [47] inverted latent, and " + }, + { + "bbox": [ + 131, + 291, + 482, + 327 + ], + "type": "inline_equation", + "content": "x_{t}^{\\mathrm{rev}} \\sim \\mathcal{N}(\\mu_{\\phi}(x_{t + 1},t),\\Sigma_{\\phi}(x_{t + 1},t))" + }, + { + "bbox": [ + 131, + 291, + 482, + 327 + ], + "type": "text", + "content": " is the diffusion model output." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 327, + 482, + 447 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 327, + 482, + 447 + ], + "spans": [ + { + "bbox": [ + 131, + 327, + 482, + 447 + ], + "type": "text", + "content": "Additionally, to mitigate accumulated bias for distant camera views during gradual repainting, as shown in the fine stage in Figure 3, we sample camera viewpoints bi-directionally, alternating between clockwise and counterclockwise directions. This bidirectional sampling allows us to paint up to 180 degrees from two opposite directions, instead of painting across the entire 360 degrees from one direction. Moreover, to paint bidirectionally invisible content and ensure smooth transitions at the junction of the two directions, we replace the previous unidirectional painting mask " + }, + { + "bbox": [ + 131, + 327, + 482, + 447 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 131, + 327, + 482, + 447 + ], + "type": "text", + "content": " by a bidirectional painting mask. This mask merges two unidirectional masks from bidirectional neighbor views (red cameras in Figure 3) by taking the maximum value of each pixel." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 465, + 482, + 634 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 465, + 482, + 634 + ], + "spans": [ + { + "bbox": [ + 131, + 465, + 482, + 634 + ], + "type": "text", + "content": "Control-based Content Enhancement. While the progressive painting maintains consistency between adjacent views, views that are farther apart may become inconsistent and experience a decline in quality due to cumulative errors. To tackle these issues, we improve the painting process by incorporating various control techniques to enhance the consistency and quality of synthesized images. Specifically, for geometry control, we employ ControlNet [62] to condition the diffusion model with coarse depth maps, ensuring geometric consistency across images. To transfer reference textures and mitigate the cumulative texture bias during the progressive painting, inspired by HiFi-123 [61], we incorporate an attention injection mechanism [3] that injects reference attention features into the diffusion denoising process. By replacing the novel-view content features (Key features " + }, + { + "bbox": [ + 131, + 465, + 482, + 634 + ], + "type": "inline_equation", + "content": "K_{t}" + }, + { + "bbox": [ + 131, + 465, + 482, + 634 + ], + "type": "text", + "content": " and Value features " + }, + { + "bbox": [ + 131, + 465, + 482, + 634 + ], + "type": "inline_equation", + "content": "V_{t}" + }, + { + "bbox": [ + 131, + 465, + 482, + 634 + ], + "type": "text", + "content": ") with reference-view attention features (" + }, + { + "bbox": [ + 131, + 465, + 482, + 634 + ], + "type": "inline_equation", + "content": "K_{r}" + }, + { + "bbox": [ + 131, + 465, + 482, + 634 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 465, + 482, + 634 + ], + "type": "inline_equation", + "content": "V_{r}" + }, + { + "bbox": [ + 131, + 465, + 482, + 634 + ], + "type": "text", + "content": ") during each denoising step, the novel-view image features can directly query the high-quality reference features by:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 206, + 641, + 481, + 669 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 206, + 641, + 481, + 669 + ], + "spans": [ + { + "bbox": [ + 206, + 641, + 481, + 669 + ], + "type": "interline_equation", + "content": "\\operatorname {A t t e n t i o n} \\left(Q _ {t}, K _ {r}, V _ {r}\\right) = \\operatorname {S o f t m a x} \\left(\\frac {Q _ {t} K _ {r} ^ {T}}{\\sqrt {d}}\\right) V _ {r}, \\tag {6}", + "image_path": "ff1cef407fddf9e9e702de22a93c288dee8efcd60a9d4da60e68b0cae11fab6e.jpg" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 173, + 91, + 448, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 173, + 91, + 448, + 103 + ], + "spans": [ + { + "bbox": [ + 173, + 91, + 448, + 103 + ], + "type": "text", + "content": "Repaint123: Progressive Controllable Repainting for 3D Generation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 135, + 118, + 480, + 264 + ], + "blocks": [ + { + "bbox": [ + 135, + 118, + 480, + 264 + ], + "lines": [ + { + "bbox": [ + 135, + 118, + 480, + 264 + ], + "spans": [ + { + "bbox": [ + 135, + 118, + 480, + 264 + ], + "type": "image", + "image_path": "c0c41a3644127407014e0bd4b849a1abf8447e2b634353c776f0f94ade11bebe.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 274, + 482, + 351 + ], + "lines": [ + { + "bbox": [ + 130, + 274, + 482, + 351 + ], + "spans": [ + { + "bbox": [ + 130, + 274, + 482, + 351 + ], + "type": "text", + "content": "Fig. 4: Visibility-based Controllable Repainting Pipeline. Our method repairs invisible content (black areas in " + }, + { + "bbox": [ + 130, + 274, + 482, + 351 + ], + "type": "inline_equation", + "content": "M_t" + }, + { + "bbox": [ + 130, + 274, + 482, + 351 + ], + "type": "text", + "content": ") of novel-view rendered image " + }, + { + "bbox": [ + 130, + 274, + 482, + 351 + ], + "type": "inline_equation", + "content": "I_{\\mathrm{rgb}}" + }, + { + "bbox": [ + 130, + 274, + 482, + 351 + ], + "type": "text", + "content": " according to the inverted latent " + }, + { + "bbox": [ + 130, + 274, + 482, + 351 + ], + "type": "inline_equation", + "content": "x_t^{inv}" + }, + { + "bbox": [ + 130, + 274, + 482, + 351 + ], + "type": "text", + "content": " and the visibility map " + }, + { + "bbox": [ + 130, + 274, + 482, + 351 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 130, + 274, + 482, + 351 + ], + "type": "text", + "content": " to obtain refined image " + }, + { + "bbox": [ + 130, + 274, + 482, + 351 + ], + "type": "inline_equation", + "content": "I_{\\mathrm{rgb}}^{\\mathrm{fine}}" + }, + { + "bbox": [ + 130, + 274, + 482, + 351 + ], + "type": "text", + "content": ". We introduce multiple control techniques to enhance the consistency and quality of generated textures, including attention feature injection, depth guidance by ControlNet, and semantic prompt by CLIP encoder. Noticing that the repainting mask " + }, + { + "bbox": [ + 130, + 274, + 482, + 351 + ], + "type": "inline_equation", + "content": "M_t" + }, + { + "bbox": [ + 130, + 274, + 482, + 351 + ], + "type": "text", + "content": " is simultaneously changed with denoising timesteps for adaptive refinement." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 373, + 480, + 409 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 373, + 480, + 409 + ], + "spans": [ + { + "bbox": [ + 130, + 373, + 480, + 409 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 373, + 480, + 409 + ], + "type": "inline_equation", + "content": "Q_{t}" + }, + { + "bbox": [ + 130, + 373, + 480, + 409 + ], + "type": "text", + "content": " represents the novel-view query features. This enhancement facilitates the transfer of texture details and improves the consistency between the reference-view image and novel-view images." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 410, + 482, + 494 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 410, + 482, + 494 + ], + "spans": [ + { + "bbox": [ + 130, + 410, + 482, + 494 + ], + "type": "text", + "content": "During our experiments, we observed a decline in image quality during repainting, likely due to not using text prompts for performing classifier-free guidance [15], crucial for producing high-quality images in diffusion models. To improve this, we adopt IP-Adapter [58], which uses a CLIP encoder to convert reference images into semantic features. These features are then projected into image prompts, providing visual cues that enhance image generation through classifier-free guidance." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "content": "Adaptive Visible Content Refinement. Repainting only the invisible content often leads to distortion in the visible area As shown in Figure 5, when the previous view is an oblique view, it leads to a low-resolution update on the texture maps, resulting in high distortion when rendering from a better observing view. Therefore, the visible content currently observed in a superior view should be refined to replenish details. However, the selection of refinement strength for these visible content is tricky [38, 54], as excessive strength produces unfaithful results leading to inconsistency with previously repainted views, while insufficient strength limits quality improvement. We propose a visibility-based adaptive refinement strategy to refine these previously seen regions with different strengths, aiming to achieve improved quality-consistency trade-off. As visualized in Figure 5, we view painting for visible content as a process similar to superresolution that replenishes detailed information. According to the Orthographic" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 92, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 92, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 92, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "type": "text", + "content": "Zhang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 298, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 298, + 342 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 298, + 342 + ], + "type": "text", + "content": "Projection Theorem, which asserts that the projected resolution of a surface is directly proportional to the view obliqueness (" + }, + { + "bbox": [ + 130, + 116, + 298, + 342 + ], + "type": "inline_equation", + "content": "\\cos\\theta" + }, + { + "bbox": [ + 130, + 116, + 298, + 342 + ], + "type": "text", + "content": " as shown in Figure 5), we can assume that the repainting strength is equal to " + }, + { + "bbox": [ + 130, + 116, + 298, + 342 + ], + "type": "inline_equation", + "content": "(1 - \\cos\\theta_{1} / \\cos\\theta_{2})" + }, + { + "bbox": [ + 130, + 116, + 298, + 342 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 116, + 298, + 342 + ], + "type": "inline_equation", + "content": "\\cos\\theta_{2} / \\cos\\theta_{1}" + }, + { + "bbox": [ + 130, + 116, + 298, + 342 + ], + "type": "text", + "content": " represents the upsampling scale in terms of super-resolution and visibility variations in the context of 3D rendering. Therefore, we can define a visibility map " + }, + { + "bbox": [ + 130, + 116, + 298, + 342 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 130, + 116, + 298, + 342 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 130, + 116, + 298, + 342 + ], + "type": "inline_equation", + "content": "V_{i,j} = \\cos\\theta_{i,j}^{p,c} / \\cos\\theta_{i,j}^{c}" + }, + { + "bbox": [ + 130, + 116, + 298, + 342 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 116, + 298, + 342 + ], + "type": "inline_equation", + "content": "\\cos\\theta_{i,j}^{p,c}" + }, + { + "bbox": [ + 130, + 116, + 298, + 342 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 116, + 298, + 342 + ], + "type": "inline_equation", + "content": "\\cos\\theta_{i,j}^{c}" + }, + { + "bbox": [ + 130, + 116, + 298, + 342 + ], + "type": "text", + "content": " are obtained by multiplying the current-view normal map with camera rays of the previous view " + }, + { + "bbox": [ + 130, + 116, + 298, + 342 + ], + "type": "inline_equation", + "content": "v_{p}" + }, + { + "bbox": [ + 130, + 116, + 298, + 342 + ], + "type": "text", + "content": " and the current view " + }, + { + "bbox": [ + 130, + 116, + 298, + 342 + ], + "type": "inline_equation", + "content": "v_{c}" + }, + { + "bbox": [ + 130, + 116, + 298, + 342 + ], + "type": "text", + "content": " respectively. Higher visibility value indicates less repainting strength, while invisible areas require full strength." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 343, + 482, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 343, + 482, + 392 + ], + "spans": [ + { + "bbox": [ + 130, + 343, + 482, + 392 + ], + "type": "text", + "content": "By this association between repairing strength and the visibility map " + }, + { + "bbox": [ + 130, + 343, + 482, + 392 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 130, + 343, + 482, + 392 + ], + "type": "text", + "content": ", we can binarize the visibility map to the timestep-aware adaptive repairing mask " + }, + { + "bbox": [ + 130, + 343, + 482, + 392 + ], + "type": "inline_equation", + "content": "M_{t}" + }, + { + "bbox": [ + 130, + 343, + 482, + 392 + ], + "type": "text", + "content": " during each denoising step, visualized in the green box \"Timestep-aware binarization\" in Figure 4:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 238, + 405, + 481, + 437 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 238, + 405, + 481, + 437 + ], + "spans": [ + { + "bbox": [ + 238, + 405, + 481, + 437 + ], + "type": "interline_equation", + "content": "M _ {t} ^ {i, j} = \\left\\{ \\begin{array}{l l} 1, & \\text {i f} V ^ {i, j} > 1 - t / T \\\\ 0, & \\text {e l s e}, \\end{array} \\right. \\tag {7}", + "image_path": "2503360f55e9d5fd78475c3828781d6afd815a130a18015425de42e7f789c49a.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 449, + 482, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 449, + 482, + 486 + ], + "spans": [ + { + "bbox": [ + 130, + 449, + 482, + 486 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 449, + 482, + 486 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 449, + 482, + 486 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 449, + 482, + 486 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 130, + 449, + 482, + 486 + ], + "type": "text", + "content": " are the 2D position of pixels in visibility map " + }, + { + "bbox": [ + 130, + 449, + 482, + 486 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 130, + 449, + 482, + 486 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 449, + 482, + 486 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 130, + 449, + 482, + 486 + ], + "type": "text", + "content": " is the total number of timesteps of the diffusion model. By doing this, we can adaptively refine visible content with a proper strength." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 516, + 482, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 516, + 482, + 564 + ], + "spans": [ + { + "bbox": [ + 130, + 516, + 482, + 564 + ], + "type": "text", + "content": "Efficient Optimization. Using a progressive painting strategy and various control techniques to refine the inverted latent, we obtain a high-quality, view-consistent refined image, " + }, + { + "bbox": [ + 130, + 516, + 482, + 564 + ], + "type": "inline_equation", + "content": "I_{\\mathrm{rgb}}^{\\mathrm{fine}}" + }, + { + "bbox": [ + 130, + 516, + 482, + 564 + ], + "type": "text", + "content": ". This process allows us to use simple MSE loss for efficient texture map optimization:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 255, + 577, + 481, + 593 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 255, + 577, + 481, + 593 + ], + "spans": [ + { + "bbox": [ + 255, + 577, + 481, + 593 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {f i n e}} = \\left| \\left| I _ {\\mathrm {r g b}} ^ {\\text {f i n e}} - I _ {\\mathrm {r g b}} \\right| \\right| _ {2} ^ {2}, \\tag {8}", + "image_path": "0eb9081d40d4f88dd1457d4fa800d3a2a6643ba4ac47e3d83fd22a1237a3b1b4.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "inline_equation", + "content": "I_{\\mathrm{rgb}}" + }, + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "text", + "content": " represents the rendered image in the fine stage. MSE loss is faster and more deterministic to optimize than the traditional SDS loss, speeding up the optimization process. After sampling a novel view, we first conduct image repainting and then proceed with efficient texture optimization, repeating this process until we cover a 360-degree view range." + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 312, + 119, + 476, + 252 + ], + "blocks": [ + { + "bbox": [ + 312, + 119, + 476, + 252 + ], + "lines": [ + { + "bbox": [ + 312, + 119, + 476, + 252 + ], + "spans": [ + { + "bbox": [ + 312, + 119, + 476, + 252 + ], + "type": "image", + "image_path": "3c926188a9f39f0ddd010635916674dffc9e5b8ea128ee7aed973698f09f7502.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 260, + 482, + 316 + ], + "lines": [ + { + "bbox": [ + 304, + 260, + 482, + 316 + ], + "spans": [ + { + "bbox": [ + 304, + 260, + 482, + 316 + ], + "type": "text", + "content": "Fig. 5: Relation between view obliqueness and refinement strength. Red box shows the same texture part in different views. Texture details in generated views might be insufficient due to the variation of angles." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 173, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 173, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 173, + 91, + 447, + 102 + ], + "type": "text", + "content": "Repaint123: Progressive Controllable Repainting for 3D Generation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 224, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 224, + 129 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 224, + 129 + ], + "type": "text", + "content": "5 Experiment" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 145, + 279, + 157 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 145, + 279, + 157 + ], + "spans": [ + { + "bbox": [ + 132, + 145, + 279, + 157 + ], + "type": "text", + "content": "5.1 Implementation Details" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 171, + 481, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 171, + 481, + 266 + ], + "spans": [ + { + "bbox": [ + 132, + 171, + 481, + 266 + ], + "type": "text", + "content": "In our experiment, we consistently apply the same hyperparameters across all our method's results. We progressively turn around the viewpoints by 40 degrees each time, consequently obtaining 8 views per object. We use a 50-step DDIM schedule and perform a 30-step latent diffusion inversion. Stable Diffusion 1.5 is utilized for all methods. During the coarse stage training, we set " + }, + { + "bbox": [ + 132, + 171, + 481, + 266 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{rgb}}" + }, + { + "bbox": [ + 132, + 171, + 481, + 266 + ], + "type": "text", + "content": " to 10. Generating a single 3D object takes only 2 minutes on a single 40G A100 GPU, approximately 1 minute for the coarse stage and mesh extraction, and another 1 minute for refinement." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 290, + 207, + 303 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 290, + 207, + 303 + ], + "spans": [ + { + "bbox": [ + 132, + 290, + 207, + 303 + ], + "type": "text", + "content": "5.2 Baselines" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 316, + 481, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 316, + 481, + 449 + ], + "spans": [ + { + "bbox": [ + 132, + 316, + 481, + 449 + ], + "type": "text", + "content": "We adopt RealFusion [28], Make-It-3D [51], and Zero123-XL [24], Magic123 [35] as our NeRF-based baselines and DreamGaussian [50] as our Gaussian-Splatting-based baseline. RealFusion presents a single-stage algorithm for NeRF generation leveraging 2D SDS loss for novel views. Make-It-3D is a two-stage approach that shares similar objectives with RealFusion but employs a point cloud representation for refinement at the second stage. For Zero123-XL, we adopt the implementation [49] and add a mesh fine-tuning stage for fair comparison. Integrating Zero123 and RealFusion, Magic123 incorporates a 2D SDS loss with 3D SDS loss provided by Zero123 to balance geometry and texture and adopts DMTet [45] representation at the second stage. DreamGaussian integrates 3D Gaussian Splatting into 3D generation and greatly improves the speed." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 472, + 261, + 483 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 472, + 261, + 483 + ], + "spans": [ + { + "bbox": [ + 132, + 472, + 261, + 483 + ], + "type": "text", + "content": "5.3 Evaluation Protocol" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 497, + 481, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 497, + 481, + 533 + ], + "spans": [ + { + "bbox": [ + 132, + 497, + 481, + 533 + ], + "type": "text", + "content": "Datasets. Based on previous research, we utilized the Realfusion15 dataset [28] and test-alpha dataset collected by Make-It-3D [51], which comprises many common objects from diverse styles." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 557, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 557, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 557, + 481, + 665 + ], + "type": "text", + "content": "Evaluation metrics. An effective 3D generation approach should produce 3D content which closely resemble the reference view, and maintain consistency of semantics and textures with the reference image when observed from new views. Therefore, to evaluate the overall quality of the generated 3D object, we choose the following metrics from two aspects: 1) PSNR and LPIPS [63], which measure pixel-level and perceptual reconstruction quality respectively at the reference view; 2) CLIP Similarity [36] and Contextual Distance [27], which assess the similarity of semantics and detailed textures respectively between the novel perspective and the reference view." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "type": "text", + "content": "Zhang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 148, + 125, + 467, + 616 + ], + "blocks": [ + { + "bbox": [ + 148, + 125, + 467, + 616 + ], + "lines": [ + { + "bbox": [ + 148, + 125, + 467, + 616 + ], + "spans": [ + { + "bbox": [ + 148, + 125, + 467, + 616 + ], + "type": "image", + "image_path": "4c34e05eed042a781c9a901ee81c2129140d5ad78431f0f79ad645c6d86c11e9.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 143, + 625, + 469, + 639 + ], + "lines": [ + { + "bbox": [ + 143, + 625, + 469, + 639 + ], + "spans": [ + { + "bbox": [ + 143, + 625, + 469, + 639 + ], + "type": "text", + "content": "Fig. 6: Qualitative comparisons on image-to-3D generation. Zoom in for details." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 173, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 173, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 173, + 91, + 447, + 102 + ], + "type": "text", + "content": "Repaint123: Progressive Controllable Repainting for 3D Generation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 132, + 114, + 479, + 226 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 479, + 226 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 479, + 226 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 479, + 226 + ], + "type": "table", + "html": "
DatasetMetrics\\MethodsNeRF-basedGaussian-Splatting-based
RealFusionMake-it-3DZero-123-XL*Magic123DreamGaussianRepaint123
RealFusion15CLIP-Similarity↑0.710.810.830.820.770.85
Context-Dis↓2.201.821.591.641.611.55
PSNR↑19.2416.5619.5619.6818.9419.00
LPIPS↓0.1940.1770.1080.1070.1110.101
Test-alphaCLIP-Similarity↑0.680.760.840.840.790.88
Context-Dis↓2.201.731.521.571.621.50
PSNR↑22.9117.2124.3924.6922.3322.38
LPIPS↓0.1050.2370.0500.0460.0570.048
Optimization time20min1h30min1h (+2h)2min2 min
", + "image_path": "ae0fa86480ddcf1a5826b7e8f8a59ea30eda7ef1b76cf60210058224d4bf32bf.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 228, + 482, + 295 + ], + "lines": [ + { + "bbox": [ + 130, + 228, + 482, + 295 + ], + "spans": [ + { + "bbox": [ + 130, + 228, + 482, + 295 + ], + "type": "text", + "content": "Table 1: We show quantitative results in terms of CLIP-Similarity↑ / Contextual-Distance↓ / PSNR↑ / LPIPS↓. The results are shown on the RealFusion15 and test-alpha datasets, while bold reflects the best for all methods and the underline represents the best for Gaussian-Splatting-based methods. * indicates that Zero123-XL incorporates a mesh fine-tuning stage for further quality improvement. The time required by textual inversion is indicated in parentheses." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 132, + 327, + 227, + 339 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 327, + 227, + 339 + ], + "spans": [ + { + "bbox": [ + 132, + 327, + 227, + 339 + ], + "type": "text", + "content": "5.4 Comparisons" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 347, + 482, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 347, + 482, + 540 + ], + "spans": [ + { + "bbox": [ + 130, + 347, + 482, + 540 + ], + "type": "text", + "content": "Comparisons with NeRF-based Methods. As shown in Table 1, we evaluate the quality of generated 3D objects across various NeRF-based methods. Our method achieves superior 3D consistency in generating 3D objects, as evidenced by the best performance of CLIP-similarity and Contextual-distance. Regarding reference-view reconstruction quality, there is a gap compared with NeRF-based approaches as shown in Table 1, which we attribute to the immaturity of current Gaussian-Splatting-based methods. Compared to NeRF-based methods for the optimization time, our approach reaches a significant acceleration of over 10 times and simultaneously achieves high quality, due to the 3D Gaussian-Splatting representation in coarse stage and efficient texture refinement in fine stage. As shown in Figure 6, Repaint123 achieves the best visual results in terms of texture consistency and generation quality as opposed to other NeRF-based methods. From the visual comparison, our method achieves consistent and detailed textures in invisible areas, while Zero123-XL results in over-smooth textures, Magic123 produces inconsistent and oversaturated colors, Realfusion and Make-It-3D fail to generate full geometry and consistent textures." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 557, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 557, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 557, + 482, + 665 + ], + "type": "text", + "content": "Comparisons with GS-based methods. We conduct comparisons with other works based on 3D Gaussian Splatting in the last column of Table 1. Within a comparable generation time, our method demonstrates superiority over existing Gaussian-based approach in aspects of texture consistency and the quality of reference-view reconstruction. The superiority of our proposed Repaint123 is evidenced by the evaluation of four distinct metrics from the Table 1. As shown in Figure 6, DreamGaussian usually leads to over-smooth texture inconsistent with reference view, while our method can produce high-quality 3D content with view-consistent and detailed textures." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "type": "text", + "content": "Zhang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 153, + 117, + 462, + 358 + ], + "blocks": [ + { + "bbox": [ + 153, + 117, + 462, + 358 + ], + "lines": [ + { + "bbox": [ + 153, + 117, + 462, + 358 + ], + "spans": [ + { + "bbox": [ + 153, + 117, + 462, + 358 + ], + "type": "image", + "image_path": "46cee30e8c7b0c143b6339a8dd8e9bcd57a50eaae9fc1070caf5cb9b894a8502.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 366, + 482, + 422 + ], + "lines": [ + { + "bbox": [ + 130, + 366, + 482, + 422 + ], + "spans": [ + { + "bbox": [ + 130, + 366, + 482, + 422 + ], + "type": "text", + "content": "Fig. 7: Qualitative ablation study by removing one component at a time from the overall method. The presence of artifacts is highlighted by red boxes. Inconsistencies such as multi-face problems and mismatches in content and style are evident without using repainting and attention injection strategies. Absence of image prompting and adaptive refinement notably degrades quality." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 447, + 272, + 459 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 447, + 272, + 459 + ], + "spans": [ + { + "bbox": [ + 132, + 447, + 272, + 459 + ], + "type": "text", + "content": "5.5 Ablation and Analysis" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 472, + 481, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 472, + 481, + 495 + ], + "spans": [ + { + "bbox": [ + 130, + 472, + 481, + 495 + ], + "type": "text", + "content": "This section details both qualitative and quantitative analyses to highlight the effectiveness of our proposed methods, as illustrated in Figure 7 and Table 2." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 497, + 482, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 497, + 482, + 581 + ], + "spans": [ + { + "bbox": [ + 130, + 497, + 482, + 581 + ], + "type": "text", + "content": "Effectiveness of Repainting. Figure 7 illustrates that without repainting strategy, the generated novel-view image tends to resemble the reference image, leading to inconsistencies in the shared visible areas between different views, such as the content misalignment (observed in the Ice Cream example) and issues with multiple faces (evident in the Anya example). These issues stem from the absence of alignment constraints for shared visible regions, leading to conflicts and quality degradation of the reconstructed 3D texture." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 582, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 482, + 665 + ], + "type": "text", + "content": "Impact of Attention Injection. Table 2 shows that injecting attention features from the reference-view image markedly enhances the consistency of both semantics and fine-grained textures, as evidenced by increased CLIP similarity and reduced Contextual distance. Without this injection strategy, as depicted in Figure 7, the synthesized images retain basic semantics but fail to accurately transfer detailed textures and styles from the reference view, resulting in inconsistencies across multiple views." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 173, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 173, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 173, + 91, + 447, + 102 + ], + "type": "text", + "content": "Repaint123: Progressive Controllable Repainting for 3D Generation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 170, + 114, + 443, + 191 + ], + "blocks": [ + { + "bbox": [ + 170, + 114, + 443, + 191 + ], + "lines": [ + { + "bbox": [ + 170, + 114, + 443, + 191 + ], + "spans": [ + { + "bbox": [ + 170, + 114, + 443, + 191 + ], + "type": "table", + "html": "
Method \\ MetricCLIP↑Contextual↓PSNR↑LPIPS↓
Coarse0.711.7821.170.133
vanilla painting0.711.6222.410.049
+attention injection0.781.5622.420.048
+image prompt0.841.5222.400.048
+adaptive refinemnet (Ours)0.881.5022.380.048
", + "image_path": "1771ae6aec9201acec78ecf9651eee077f8c384bc9a468f34982f0f6e02f10f9.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 192, + 482, + 225 + ], + "lines": [ + { + "bbox": [ + 132, + 192, + 482, + 225 + ], + "spans": [ + { + "bbox": [ + 132, + 192, + 482, + 225 + ], + "type": "text", + "content": "Table 2: Quantitative ablation study on Test-alpha dataset by progressively adding our proposed components. The last three lines show the cumulative effects of the proposed modules. Our approach, utilizing all strategies, delivered the best performance." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 130, + 259, + 479, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 259, + 479, + 319 + ], + "spans": [ + { + "bbox": [ + 130, + 259, + 479, + 319 + ], + "type": "text", + "content": "Role of Image Prompt. The introduction of using the reference image as an image prompt, as shown in Table 2 and Figure 7, significantly boosts both multi-view consistency and image quality of the generated images. Without this technique to perform classifier-free guidance, the generation of detailed and consistent textures across views is compromised." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 319, + 480, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 319, + 480, + 380 + ], + "spans": [ + { + "bbox": [ + 130, + 319, + 480, + 380 + ], + "type": "text", + "content": "Benefits of Adaptive Refinement. The necessity of adaptive refinement is clear from Figure 7, where its absence leads to artifacts and blurriness in obliquely viewing areas of previous views due to low-resolution updates, as mentioned in Section 4.2. Table 2 also demonstrates its benefits through improved CLIP similarity and Contextual distance." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 399, + 222, + 412 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 399, + 222, + 412 + ], + "spans": [ + { + "bbox": [ + 132, + 399, + 222, + 412 + ], + "type": "text", + "content": "6 Limitations" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 426, + 482, + 536 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 426, + 482, + 536 + ], + "spans": [ + { + "bbox": [ + 130, + 426, + 482, + 536 + ], + "type": "text", + "content": "Despite the promising results, our method still has some limitations. Our method is based on 3D Gaussian Splatting representation in the coarse stage during training. While 3D Gaussian Splatting accelerates the training process, it may exhibit geometry artifacts during mesh extraction, such as holes, and inferior results compared to NeRF-based methods for the reconstruction of reference view, due to immaturity in generation tasks. This is supported by results in the Appendix, where NeRF is considered as an alternative to 3D Gaussian Splatting for the coarse stage. We expect these limitations can be mitigated with the development of 3D Gaussian Splatting in the future." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 555, + 220, + 567 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 555, + 220, + 567 + ], + "spans": [ + { + "bbox": [ + 132, + 555, + 220, + 567 + ], + "type": "text", + "content": "7 Conclusion" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "content": "This work presents Repaint123 for generating high-quality 3D content from a single image in about 2 minutes. By leveraging progressive controllable paint, our approach overcomes the limitations of existing studies and achieves state-of-the-art results in terms of both texture quality and multi-view consistency, paving the way for future progress in 3D content generation from one image. Furthermore, we validate the effectiveness of our proposed method through a comprehensive set of experiments." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "type": "text", + "content": "Zhang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 141, + 528, + 166 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 141, + 528, + 166 + ], + "spans": [ + { + "bbox": [ + 130, + 141, + 528, + 166 + ], + "type": "text", + "content": "This work was supported in part by Natural Science Foundation of China (No. 62332002, 62202014), and Shenzhen Basic Research Program (No.JCYJ20220813151736001)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 185, + 197, + 198 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 185, + 197, + 198 + ], + "spans": [ + { + "bbox": [ + 133, + 185, + 197, + 198 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 211, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 138, + 211, + 481, + 244 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 211, + 481, + 244 + ], + "spans": [ + { + "bbox": [ + 138, + 211, + 481, + 244 + ], + "type": "text", + "content": "1. Armandpour, M., Zheng, H., Sadeghian, A., Sadeghian, A., Zhou, M.: Re-imagine the negative prompt algorithm: Transform 2d diffusion into 3d, alleviate janus problem and beyond. arXiv preprint arXiv:2304.04968 (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 246, + 481, + 266 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 246, + 481, + 266 + ], + "spans": [ + { + "bbox": [ + 138, + 246, + 481, + 266 + ], + "type": "text", + "content": "2. Barron, J.T., Mildenhall, B., Verbin, D., Srinivasan, P.P., Hedman, P.: Mip-nerf 360: Unbounded anti-aliased neural radiance fields. CVPR (2022)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 268, + 481, + 311 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 268, + 481, + 311 + ], + "spans": [ + { + "bbox": [ + 138, + 268, + 481, + 311 + ], + "type": "text", + "content": "3. Cao, M., Wang, X., Qi, Z., Shan, Y., Qie, X., Zheng, Y.: Masactrl: Tuning-free mutual self-attention control for consistent image synthesis and editing. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 22560-22570 (October 2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 312, + 481, + 344 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 312, + 481, + 344 + ], + "spans": [ + { + "bbox": [ + 138, + 312, + 481, + 344 + ], + "type": "text", + "content": "4. Chan, E.R., Lin, C.Z., Chan, M.A., Nagano, K., Pan, B., Mello, S.D., Gallo, O., Guibas, L., Tremblay, J., Khamis, S., Karras, T., Wetzstein, G.: Efficient geometry-aware 3D generative adversarial networks. In: CVPR (2022)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 346, + 481, + 366 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 346, + 481, + 366 + ], + "spans": [ + { + "bbox": [ + 138, + 346, + 481, + 366 + ], + "type": "text", + "content": "5. Chen, G., Wang, W.: A survey on 3d gaussian splatting. arXiv preprint arXiv:2401.03890 (2024)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 368, + 481, + 399 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 368, + 481, + 399 + ], + "spans": [ + { + "bbox": [ + 138, + 368, + 481, + 399 + ], + "type": "text", + "content": "6. Chen, R., Chen, Y., Jiao, N., Jia, K.: *Fantasia3d: Disentangling geometry and appearance for high-quality text-to-3d content creation*. arXiv preprint arXiv:2303.13873 (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 401, + 481, + 432 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 401, + 481, + 432 + ], + "spans": [ + { + "bbox": [ + 138, + 401, + 481, + 432 + ], + "type": "text", + "content": "7. Chen, Y., Zhang, C., Yang, X., Cai, Z., Yu, G., Yang, L., Lin, G.: It3d: Improved text-to-3d generation with explicit view synthesis. arXiv preprint arXiv:2308.11473 (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 434, + 481, + 466 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 434, + 481, + 466 + ], + "spans": [ + { + "bbox": [ + 138, + 434, + 481, + 466 + ], + "type": "text", + "content": "8. Chen, Z., Funkhouser, T., Hedman, P., Tagliasacchi, A.: Mobilenerf: Exploiting the polygon rasterization pipeline for efficient neural field rendering on mobile architectures. arXiv preprint arXiv:2208.00277 (2022)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 468, + 481, + 488 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 468, + 481, + 488 + ], + "spans": [ + { + "bbox": [ + 138, + 468, + 481, + 488 + ], + "type": "text", + "content": "9. Chen, Z., Wang, F., Liu, H.: Text-to-3d using gaussian splatting. arXiv preprint arXiv:2309.16585 (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 490, + 481, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 490, + 481, + 521 + ], + "spans": [ + { + "bbox": [ + 138, + 490, + 481, + 521 + ], + "type": "text", + "content": "0. Cheng, X., Yang, T., Wang, J., Li, Y., Zhang, L., Zhang, J., Yuan, L.: Progressive3d: Progressively local editing for text-to-3d content creation with complex semantic prompts. arXiv preprint arXiv:2310.11784 (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 522, + 481, + 565 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 522, + 481, + 565 + ], + "spans": [ + { + "bbox": [ + 138, + 522, + 481, + 565 + ], + "type": "text", + "content": "1. Dou, Z., Wu, Q., Lin, C., Cao, Z., Wu, Q., Wan, W., Komura, T., Wang, W.: Tore: Token reduction for efficient human mesh recovery with transformer. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 15143-15155 (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 567, + 481, + 598 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 567, + 481, + 598 + ], + "spans": [ + { + "bbox": [ + 138, + 567, + 481, + 598 + ], + "type": "text", + "content": "2. Gal, R., Alaluf, Y., Atzmon, Y., Patashnik, O., Bermano, A.H., Chechik, G., Cohen-Or, D.: An image is worth one word: Personalizing text-to-image generation using textual inversion. arXiv preprint arXiv:2208.01618 (2022)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 600, + 481, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 600, + 481, + 620 + ], + "spans": [ + { + "bbox": [ + 138, + 600, + 481, + 620 + ], + "type": "text", + "content": "3. Hedman, P., Srinivasan, P.P., Mildenhall, B., Barron, J.T., Debevec, P.: Baking neural radiance fields for real-time view synthesis. ICCV (2021)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 622, + 481, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 622, + 481, + 643 + ], + "spans": [ + { + "bbox": [ + 138, + 622, + 481, + 643 + ], + "type": "text", + "content": "4. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. arXiv preprint arxiv:2006.11239 (2020)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 138, + 644, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 644, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 644, + 481, + 665 + ], + "type": "text", + "content": "5. Ho, J., Salimans, T.: Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598 (2022)" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 173, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 173, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 173, + 91, + 447, + 102 + ], + "type": "text", + "content": "Repaint123: Progressive Controllable Repainting for 3D Generation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 666 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "type": "text", + "content": "16. Huang, Y., Wang, J., Shi, Y., Qi, X., Zha, Z.J., Zhang, L.: Dreamtime: An improved optimization strategy for text-to-3d content creation. arXiv preprint arXiv:2306.12422 (2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 149, + 480, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 149, + 480, + 171 + ], + "spans": [ + { + "bbox": [ + 133, + 149, + 480, + 171 + ], + "type": "text", + "content": "17. Jun, H., Nichol, A.: Shap-e: Generating conditional 3d implicit functions. arXiv preprint arXiv:2305.02463 (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 171, + 480, + 203 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 171, + 480, + 203 + ], + "spans": [ + { + "bbox": [ + 133, + 171, + 480, + 203 + ], + "type": "text", + "content": "18. Kerbl, B., Kopanas, G., Leimkuhler, T., Drettakis, G.: 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics (ToG) 42(4), 1-14 (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 203, + 480, + 236 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 203, + 480, + 236 + ], + "spans": [ + { + "bbox": [ + 133, + 203, + 480, + 236 + ], + "type": "text", + "content": "19. Kopanas, G., Leimkuhler, T., Rainer, G., Jambon, C., Drettakis, G.: Neural point catacaustics for novel-view synthesis of reflections. ACM Transactions on Graphics (TOG) 41(6), 1-15 (2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 236, + 480, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 236, + 480, + 267 + ], + "spans": [ + { + "bbox": [ + 132, + 236, + 480, + 267 + ], + "type": "text", + "content": "20. Kopanas, G., Philip, J., Leimkuhler, T., Drettakis, G.: Point-based neural rendering with per-view optimization. In: Computer Graphics Forum. vol. 40, pp. 29-43. Wiley Online Library (2021)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 267, + 480, + 289 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 267, + 480, + 289 + ], + "spans": [ + { + "bbox": [ + 132, + 267, + 480, + 289 + ], + "type": "text", + "content": "21. Li, Z., Müller, T., Evans, A., Taylor, R.H., Unberath, M., Liu, M.Y., Lin, C.H.: Neuralangelo: High-fidelity neural surface reconstruction. In: CVPR (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 289, + 480, + 320 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 289, + 480, + 320 + ], + "spans": [ + { + "bbox": [ + 132, + 289, + 480, + 320 + ], + "type": "text", + "content": "22. Lin, C.H., Gao, J., Tang, L., Takikawa, T., Zeng, X., Huang, X., Kreis, K., Fidler, S., Liu, M.Y., Lin, T.Y.: Magic3d: High-resolution text-to-3d content creation. In: CVPR (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 320, + 480, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 320, + 480, + 353 + ], + "spans": [ + { + "bbox": [ + 132, + 320, + 480, + 353 + ], + "type": "text", + "content": "23. Liu, M., Xu, C., Jin, H., Chen, L., Xu, Z., Su, H., et al.: One-2-3-45: Any single image to 3d mesh in 45 seconds without per-shape optimization. arXiv preprint arXiv:2306.16928 (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 353, + 480, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 353, + 480, + 386 + ], + "spans": [ + { + "bbox": [ + 132, + 353, + 480, + 386 + ], + "type": "text", + "content": "24. Liu, R., Wu, R., Van Hoorick, B., Tokmakov, P., Zakharov, S., Vondrick, C.: Zero-1-to-3: Zero-shot one image to 3d object. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 9298-9309 (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 386, + 480, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 386, + 480, + 418 + ], + "spans": [ + { + "bbox": [ + 132, + 386, + 480, + 418 + ], + "type": "text", + "content": "25. Liu, Y., Lin, C., Zeng, Z., Long, X., Liu, L., Komura, T., Wang, W.: Syncdreamer: Generating multiview-consistent images from a single-view image. arXiv preprint arXiv:2309.03453 (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 418, + 480, + 450 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 418, + 480, + 450 + ], + "spans": [ + { + "bbox": [ + 132, + 418, + 480, + 450 + ], + "type": "text", + "content": "26. Long, X., Guo, Y.C., Lin, C., Liu, Y., Dou, Z., Liu, L., Ma, Y., Zhang, S.H., Habermann, M., Theobalt, C., et al.: Wonder3d: Single image to 3d using cross-domain diffusion. arXiv preprint arXiv:2310.15008 (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 450, + 480, + 483 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 450, + 480, + 483 + ], + "spans": [ + { + "bbox": [ + 132, + 450, + 480, + 483 + ], + "type": "text", + "content": "27. Mechrez, R., Talmi, I., Zelnik-Manor, L.: The contextual loss for image transformation with non-aligned data. In: Proceedings of the European conference on computer vision (ECCV). pp. 768-783 (2018)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 483, + 480, + 504 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 483, + 480, + 504 + ], + "spans": [ + { + "bbox": [ + 132, + 483, + 480, + 504 + ], + "type": "text", + "content": "28. Melas-Kyriazi, L., Laina, I., Rupprecht, C., Vedaldi, A.: Realfusion: 360deg reconstruction of any object from a single image. In: CVPR (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 504, + 480, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 504, + 480, + 536 + ], + "spans": [ + { + "bbox": [ + 132, + 504, + 480, + 536 + ], + "type": "text", + "content": "29. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. In: ECCV (2020)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 536, + 480, + 579 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 536, + 480, + 579 + ], + "spans": [ + { + "bbox": [ + 132, + 536, + 480, + 579 + ], + "type": "text", + "content": "30. Mokady, R., Hertz, A., Aberman, K., Pritch, Y., Cohen-Or, D.: Null-text inversion for editing real images using guided diffusion models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6038-6047 (2023)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 579, + 480, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 579, + 480, + 611 + ], + "spans": [ + { + "bbox": [ + 132, + 579, + 480, + 611 + ], + "type": "text", + "content": "31. Mou, C., Wang, X., Xie, L., Zhang, J., Qi, Z., Shan, Y., Qie, X.: T2i-adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models. arXiv preprint arXiv:2302.08453 (2023)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 611, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 611, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 611, + 480, + 632 + ], + "type": "text", + "content": "32. Müller, T., Evans, A., Schied, C., Keller, A.: Instant neural graphics primitives with a multiresolution hash encoding. ACM TOG (2022)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 632, + 480, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 480, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 480, + 666 + ], + "type": "text", + "content": "33. Nichol, A., Jun, H., Dhariwal, P., Mishkin, P., Chen, M.: Point-e: A system for generating 3d point clouds from complex prompts. arXiv preprint arXiv:2212.08751 (2022)" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 216, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 216, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 216, + 102 + ], + "type": "text", + "content": "Zhang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 665 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 138 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 138 + ], + "type": "text", + "content": "34. Poole, B., Jain, A., Barron, J.T., Mildenhall, B.: Dreamfusion: Text-to-3d using 2d diffusion. arXiv preprint arXiv:2209.14988 (2022)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 139, + 482, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 139, + 482, + 183 + ], + "spans": [ + { + "bbox": [ + 130, + 139, + 482, + 183 + ], + "type": "text", + "content": "35. Qian, G., Mai, J., Hamdi, A., Ren, J., Siarohin, A., Li, B., Lee, H.Y., Skorokhodov, I., Wonka, P., Tulyakov, S., et al.: Magic123: One image to high-quality 3d object generation using both 2d and 3d diffusion priors. arXiv preprint arXiv:2306.17843 (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 184, + 481, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 184, + 481, + 226 + ], + "spans": [ + { + "bbox": [ + 132, + 184, + 481, + 226 + ], + "type": "text", + "content": "36. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 228, + 481, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 228, + 481, + 259 + ], + "spans": [ + { + "bbox": [ + 132, + 228, + 481, + 259 + ], + "type": "text", + "content": "37. Raj, A., Kaza, S., Poole, B., Niemeyer, M., Ruiz, N., Mildenhall, B., Zada, S., Aberman, K., Rubinstein, M., Barron, J., et al.: Dreambooth3d: Subject-driven text-to-3d generation. arXiv preprint arXiv:2303.13508 (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 259, + 481, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 259, + 481, + 282 + ], + "spans": [ + { + "bbox": [ + 132, + 259, + 481, + 282 + ], + "type": "text", + "content": "38. Richardson, E., Metzer, G., Alaluf, Y., Giryes, R., Cohen-Or, D.: Texture: Text-guided texturing of 3d shapes. arXiv preprint arXiv:2302.01721 (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 282, + 481, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 282, + 481, + 315 + ], + "spans": [ + { + "bbox": [ + 132, + 282, + 481, + 315 + ], + "type": "text", + "content": "39. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10684-10695 (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 316, + 481, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 316, + 481, + 358 + ], + "spans": [ + { + "bbox": [ + 132, + 316, + 481, + 358 + ], + "type": "text", + "content": "40. Ruiz, N., Li, Y., Jampani, V., Pritch, Y., Rubinstein, M., Aberman, K.: Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 22500-22510 (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 358, + 481, + 379 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 358, + 481, + 379 + ], + "spans": [ + { + "bbox": [ + 132, + 358, + 481, + 379 + ], + "type": "text", + "content": "41. Sara Fridovich-Keil and Alex Yu, Tancik, M., Chen, Q., Recht, B., Kanazawa, A.: Plenoxels: Radiance fields without neural networks. In: CVPR (2022)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 380, + 481, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 380, + 481, + 402 + ], + "spans": [ + { + "bbox": [ + 132, + 380, + 481, + 402 + ], + "type": "text", + "content": "42. Seo, H., Kim, H., Kim, G., Chun, S.Y.: Ditto-nerf: Diffusion-based iterative text to omni-directional 3d model. arXiv preprint arXiv:2304.02827 (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 403, + 481, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 403, + 481, + 435 + ], + "spans": [ + { + "bbox": [ + 132, + 403, + 481, + 435 + ], + "type": "text", + "content": "43. Seo, J., Jang, W., Kwak, M.S., Ko, J., Kim, H., Kim, J., Kim, J.H., Lee, J., Kim, S.: Let 2d diffusion model know 3d-consistency for robust text-to-3d generation. arXiv preprint arXiv:2303.07937 (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 435, + 481, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 435, + 481, + 456 + ], + "spans": [ + { + "bbox": [ + 132, + 435, + 481, + 456 + ], + "type": "text", + "content": "44. Shen, Q., Yang, X., Wang, X.: Anything-3d: Towards single-view anything reconstruction in the wild. arXiv preprint arXiv:2304.10261 (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 457, + 481, + 490 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 457, + 481, + 490 + ], + "spans": [ + { + "bbox": [ + 132, + 457, + 481, + 490 + ], + "type": "text", + "content": "45. Shen, T., Gao, J., Yin, K., Liu, M.Y., Fidler, S.: Deep marching tetrahedra: a hybrid representation for high-resolution 3d shape synthesis. In: Advances in Neural Information Processing Systems (NeurIPS) (2021)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 491, + 481, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 491, + 481, + 511 + ], + "spans": [ + { + "bbox": [ + 132, + 491, + 481, + 511 + ], + "type": "text", + "content": "46. Shi, Y., Wang, P., Ye, J., Long, M., Li, K., Yang, X.: Mvdream: Multi-view diffusion for 3d generation. arXiv preprint arXiv:2308.16512 (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 512, + 481, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 512, + 481, + 533 + ], + "spans": [ + { + "bbox": [ + 132, + 512, + 481, + 533 + ], + "type": "text", + "content": "47. Song, J., Meng, C., Ermon, S.: Denoising diffusion implicit models. arXiv:2010.02502 (October 2020), https://arxiv.org/abs/2010.02502" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 534, + 481, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 534, + 481, + 566 + ], + "spans": [ + { + "bbox": [ + 132, + 534, + 481, + 566 + ], + "type": "text", + "content": "48. Szymanowicz, S., Rupprecht, C., Vedaldi, A.: Viewset diffusion:(0-) image-conditioned 3d generative models from 2d data. arXiv preprint arXiv:2306.07881 (2023)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 567, + 481, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 567, + 481, + 588 + ], + "spans": [ + { + "bbox": [ + 132, + 567, + 481, + 588 + ], + "type": "text", + "content": "49. Tang, J.: Stable-dreamfusion: Text-to-3d with stable-diffusion (2022), https://github.com/ashawkey/stable-dreamfusion" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 589, + 481, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 589, + 481, + 610 + ], + "spans": [ + { + "bbox": [ + 132, + 589, + 481, + 610 + ], + "type": "text", + "content": "50. Tang, J., Ren, J., Zhou, H., Liu, Z., Zeng, G.: Dreamgaussian: Generative gaussian splatting for efficient 3d content creation. arXiv preprint arXiv:2309.16653 (2023)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 611, + 481, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 611, + 481, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 611, + 481, + 632 + ], + "type": "text", + "content": "51. Tang, J., Wang, T., Zhang, B., Zhang, T., Yi, R., Ma, L., Chen, D.: Make-it-3d: High-fidelity 3d creation from a single image with diffusion prior. In: ICCV (2023)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 132, + 632, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 481, + 665 + ], + "type": "text", + "content": "52. Tsalicoglou, C., Manhardt, F., Tonioni, A., Niemeyer, M., Tombari, F.: Textmesh: Generation of realistic 3d meshes from text prompts. arXiv preprint arXiv:2304.12439 (2023)" + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 173, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 173, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 173, + 91, + 447, + 102 + ], + "type": "text", + "content": "Repaint123: Progressive Controllable Repainting for 3D Generation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 491 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 481, + 161 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 481, + 161 + ], + "type": "text", + "content": "53. Wang, H., Du, X., Li, J., Yeh, R.A., Shakhnarovich, G.: Score jacobian chaining: Lifting pretrained 2d diffusion models for 3d generation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12619-12629 (2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 162, + 481, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 162, + 481, + 194 + ], + "spans": [ + { + "bbox": [ + 132, + 162, + 481, + 194 + ], + "type": "text", + "content": "54. Wang, T., Kanakis, M., Schindler, K., Van Gool, L., Obukhov, A.: Breathing new life into 3d assets with generative repainting. arXiv preprint arXiv:2309.08523 (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 194, + 481, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 194, + 481, + 226 + ], + "spans": [ + { + "bbox": [ + 132, + 194, + 481, + 226 + ], + "type": "text", + "content": "55. Wang, Z., Lu, C., Wang, Y., Bao, F., Li, C., Su, H., Zhu, J.: Prolificdreamer: High-fidelity and diverse text-to-3d generation with variational score distillation. arXiv preprint arXiv:2305.16213 (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 227, + 481, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 227, + 481, + 258 + ], + "spans": [ + { + "bbox": [ + 132, + 227, + 481, + 258 + ], + "type": "text", + "content": "56. Wu, J., Gao, X., Liu, X., Shen, Z., Zhao, C., Feng, H., Liu, J., Ding, E.: Hd-fusion: Detailed text-to-3d generation leveraging multiple noise estimation. arXiv preprint arXiv:2307.16183 (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 259, + 481, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 259, + 481, + 293 + ], + "spans": [ + { + "bbox": [ + 132, + 259, + 481, + 293 + ], + "type": "text", + "content": "57. Xu, D., Jiang, Y., Wang, P., Fan, Z., Wang, Y., Wang, Z.: Neurallift-360: Lifting an in-the-wild 2d photo to a 3d object with 360 views. arXiv e-prints pp. arXiv-2211 (2022)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 293, + 481, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 293, + 481, + 314 + ], + "spans": [ + { + "bbox": [ + 132, + 293, + 481, + 314 + ], + "type": "text", + "content": "58. Ye, H., Zhang, J., Liu, S., Han, X., Yang, W.: Ip-adapter: Text compatible image prompt adapter for text-to-image diffusion models (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 315, + 481, + 347 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 315, + 481, + 347 + ], + "spans": [ + { + "bbox": [ + 132, + 315, + 481, + 347 + ], + "type": "text", + "content": "59. Yi, T., Fang, J., Wu, G., Xie, L., Zhang, X., Liu, W., Tian, Q., Wang, X.: Gaussian dreamer: Fast generation from text to 3d gaussian splatting with point cloud priors. arXiv preprint arXiv:2310.08529 (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 348, + 481, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 348, + 481, + 380 + ], + "spans": [ + { + "bbox": [ + 132, + 348, + 481, + 380 + ], + "type": "text", + "content": "60. Yu, C., Zhou, Q., Li, J., Zhang, Z., Wang, Z., Wang, F.: Points-to-3d: Bridging the gap between sparse points and shape-controllable text-to-3d generation. arXiv preprint arXiv:2307.13908 (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 380, + 481, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 380, + 481, + 413 + ], + "spans": [ + { + "bbox": [ + 132, + 380, + 481, + 413 + ], + "type": "text", + "content": "61. Yu, W., Yuan, L., Cao, Y.P., Gao, X., Li, X., Quan, L., Shan, Y., Tian, Y.: Hifi-123: Towards high-fidelity one image to 3d content generation. arXiv preprint arXiv:2310.06744 (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 414, + 481, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 414, + 481, + 446 + ], + "spans": [ + { + "bbox": [ + 132, + 414, + 481, + 446 + ], + "type": "text", + "content": "62. Zhang, L., Rao, A., Agrawala, M.: Adding conditional control to text-to-image diffusion models. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 3836-3847 (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 447, + 481, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 447, + 481, + 468 + ], + "spans": [ + { + "bbox": [ + 132, + 447, + 481, + 468 + ], + "type": "text", + "content": "63. Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual metric. In: CVPR (2018)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 468, + 481, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 468, + 481, + 491 + ], + "spans": [ + { + "bbox": [ + 132, + 468, + 481, + 491 + ], + "type": "text", + "content": "64. Zhu, J., Zhuang, P.: Hifa: High-fidelity text-to-3d with advanced diffusion guidance. arXiv preprint arXiv:2305.18766 (2023)" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "type": "text", + "content": "Zhang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Representation Enhancement-Stabilization_ Reducing Bias-Variance of Domain Generalization/f28a3a3b-4152-48f4-99b3-107ae4c33ff2_content_list.json b/2024/Representation Enhancement-Stabilization_ Reducing Bias-Variance of Domain Generalization/f28a3a3b-4152-48f4-99b3-107ae4c33ff2_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..57434e0d3bd7cc907f975be0c40e9f8a02401294 --- /dev/null +++ b/2024/Representation Enhancement-Stabilization_ Reducing Bias-Variance of Domain Generalization/f28a3a3b-4152-48f4-99b3-107ae4c33ff2_content_list.json @@ -0,0 +1,2032 @@ +[ + { + "type": "text", + "text": "Representation Enhancement-Stabilization: Reducing Bias-Variance of Domain Generalization", + "text_level": 1, + "bbox": [ + 79, + 48, + 917, + 107 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Wei Huang1, Yilei Shi1, Zhitong Xiong1, and Xiao Xiang Zhu1,2", + "bbox": [ + 125, + 139, + 871, + 160 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Technical University of Munich, Arcisstraße 21, 80333 München, Germany \n2. Munich Center for Machine Learning, 80333 Munich \n{w2wei.huang, yilei.shi, zhitong.xiong, xiaoxiang.zhu}@tum.de", + "bbox": [ + 123, + 175, + 874, + 229 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Domain Generalization (DG) focuses on enhancing the generalization of deep learning models trained on multiple source domains to adapt to unseen target domains. This paper explores DG through the lens of bias-variance decomposition, uncovering that test errors in DG predominantly arise from cross-domain bias and variance. Inspired by this insight, we introduce a Representation Enhancement-Stabilization (RES) framework, comprising a Representation Enhancement (RE) module and a Representation Stabilization (RS) module. In RE, a novel set of feature frequency augmentation techniques is used to progressively reduce cross-domain bias during feature extraction. Furthermore, in RS, a novel Mutual Exponential Moving Average (MEMA) strategy is designed to stabilize model optimization for diminishing cross-domain variance during training. Collectively, the whole RES method can significantly enhance model generalization. We evaluate RES on five benchmark datasets and the results show that it outperforms multiple advanced DG methods. Our code will be available at https://github.com/zhu-xlab/DG-RES.", + "bbox": [ + 147, + 269, + 849, + 554 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Domain generalization $\\cdot$ Frequency Domain $\\cdot$ Data Augmentation $\\cdot$ Bias-variance Decomposition", + "bbox": [ + 147, + 572, + 849, + 606 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 79, + 637, + 309, + 656 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In recent years, deep learning models have achieved remarkable advancements in computer vision tasks, primarily under the assumption that training and test data are independent and identically distributed. However, in real-world scenarios, this assumption often breaks down due to domain shift/bias, where the distribution of test data markedly differs from that of training data. Domain shift can significantly degrade the performance of deep models on unfamiliar, unseen target domains. For instance, self-driving models trained under daylight conditions might fail to perform effectively in nighttime environments.", + "bbox": [ + 79, + 676, + 917, + 830 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To address these challenges, Domain Generalization (DG) has been proposed as a robust solution. DG focuses on training models on multiple diverse yet related source domains for robust performance on arbitrary unseen target domains. Numerous DG approaches have been developed, including adversarial training [23, 48], disentangled representation learning [37], meta-learning [7], model ensemble [2], flatness optimization [4], and domain data augmentation [8, 41, 47].", + "bbox": [ + 79, + 832, + 917, + 949 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this study we highlight the cross-domain generalization of deep models from the perspective of bias-variance decomposition. Geman et al. [12] first used bias-variance to decompose the mean square error (MSE) of neural networks. Building on this, Yang et al. [43] developed the bias-variance decomposition for the cross-entropy (CE) loss in classification tasks, which was extended by Arpit et al. [2] to analyze the impact of variance in DG, as illustrated below:", + "bbox": [ + 72, + 56, + 920, + 172 + ], + "page_idx": 1 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbb {E} _ {x, y \\sim D _ {s}} [ \\mathrm {C E} (y, f (x; D _ {s})) ] = \\underbrace {\\mathbb {E} _ {x , y \\sim D _ {s}} [ \\mathrm {C E} (y , \\bar {f} (x)) ]} _ {\\text {B i a s} ^ {2}} + \\underbrace {\\mathbb {E} _ {x , D _ {s}} [ \\mathrm {K L} (\\bar {f} (x) , f (x ; D _ {s})) ]} _ {\\text {V a r i a n c e}},\n$$\n", + "text_format": "latex", + "bbox": [ + 87, + 185, + 915, + 230 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "(1)", + "bbox": [ + 726, + 230, + 915, + 247 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "where $D_{s}$ represents source domains used for training. The formula decomposes the expected source CE loss into two components: bias and variance. Bias evaluates the discrepancy between model's average predictions and actual outcomes, while variance measures sensitivity to fluctuations during training.", + "bbox": [ + 72, + 248, + 917, + 325 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Motivated by them, we extend the bias-variance decomposition to DG to comprehensively analysis the cross-domain prediction errors as the following:", + "bbox": [ + 72, + 326, + 917, + 365 + ], + "page_idx": 1 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathbb {E} _ {x, y \\sim D _ {t}} [ \\mathrm {C E} (y, f (x _ {t}; D _ {t})) ] = \\mathbb {E} _ {x, y \\sim D _ {s}} [ \\mathrm {C E} (y, f (x _ {s} + \\delta ; D _ {s})) ] \\\\ = \\underbrace {\\mathbb {E} _ {x , y \\sim D _ {s}} [ \\mathrm {C E} (y , \\bar {f} (x _ {s} + \\delta)) ]} \\\\ \\text {C r o s s - D o m a i n B i a s} ^ {2} \\\\ + \\underbrace {\\mathbb {E} _ {x \\sim D _ {s}} \\left[ \\mathrm {K L} \\left(\\bar {f} \\left(x _ {s} + \\delta\\right) , f \\left(x _ {s} + \\delta ; D _ {s}\\right)\\right) \\right]} _ {\\text {C r o s s - D o m a i n V a r i a n c e}} \\tag {2} \\\\ + \\Delta_ {d o m a i n} (\\text {B i a s} ^ {2}, \\text {V a r i a n c e}). \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 147, + 378, + 915, + 518 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "This formula analyzes the expected CE loss when a model trained in source domains $D_{s}$ , denoted as $f(x_{s};D_{s})$ , is applied to new target domains $D_{t}$ , with respect to domain shift $\\delta$ between source and target images, $x_{s}$ and $x_{t}$ . The formula is decomposed into three components:", + "bbox": [ + 72, + 531, + 917, + 609 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- (1) Cross-Domain Bias, as shown in the left of Fig. 1. It quantifies the discrepancy between the model's average prediction in the target domain and the actual labels. High cross-domain bias dramatically enlarges the divergence between the distribution learnt from source domains and the true underlying distribution of target domains. By progressively reducing bias in feature extraction during training, i.e., $f(x_{s}) \\gets f(x_{s} + \\delta)$ , there is an approximate bias error $\\mathbb{E}_{x,y\\sim D_s}[\\mathrm{CE}(y,\\bar{f} (x))]\\gets \\mathbb{E}_{x,y\\sim D_s}[\\mathrm{CE}(y,\\bar{f} (x_s + \\delta))]$ .", + "- (2) Cross-Domain Variance, as shown in the right of Fig. 1. It measures the variability and sensitivity in the model's predictions from source to target domains. The model trained on the source domain tends to generate unstable predictions when confronted with the target domain data, with elevated cross-domain variance. Based on cross-domain bias reduction in (1), further decreasing cross-domain variance results in an approximate variance error $\\mathbb{E}_{x\\sim D_s}[\\mathrm{KL}(\\bar{f} (x_s),f(x_s;D_s))]\\gets \\mathbb{E}_{x\\sim D_s}[\\mathrm{KL}(\\bar{f} (x_s + \\delta),f(x_s + \\delta ;D_s))]$ .", + "- (3) $\\varDelta_{domain}(\\mathrm{Bias}^2$ , Variance). This term denotes the extra test error of the intertwined effects of the cross-domain bias and variance. The decrease of both bias and variance can contribute to its reduction." + ], + "bbox": [ + 89, + 619, + 920, + 946 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 77, + 16, + 94, + 29 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "W. Huang, Y. Shi et al.", + "bbox": [ + 157, + 14, + 394, + 32 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Overall, diminishing cross-domain bias and variance can narrow the gap between Eq. (1) and Eq. (2), enabling models trained on source domains to effectively adapt to target domains. This analytical framework provides actionable insights for training models capable of generalizing across diverse data distributions.", + "bbox": [ + 72, + 56, + 922, + 136 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/6679d66906eec54370f44285981249f7ead866cd92a0972158a3749a3e8918cb.jpg", + "image_caption": [ + "(a) Challenge of cross-domain bias" + ], + "image_footnote": [], + "bbox": [ + 79, + 191, + 525, + 318 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/0871b06d20bf67577b0aedd804006b951b44147dcbcf1d76f223eab837d39919.jpg", + "image_caption": [ + "(b) Challenge of cross-domain variance", + "Fig. 1: Two challenges impeding DG from the view of bias-variance decomposition." + ], + "image_footnote": [], + "bbox": [ + 561, + 186, + 920, + 315 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Based on this analytical framework, we proposed a joint Representation Enhancement-Stabilization (RES) method, which consists of a Representation Enhancement (RE) module and a Representation Stabilization (RS) module to alleviate cross-domain bias ans cross-domain variance, respectively. In RE, we propose a novel set of feature augmentation techniques within the frequency domain extracted through Fast Fourier Transform (FFT), including random noise, random dropout, and mixup. RE effectively broadens cross-domain representation space and thereby enhances robust representation learning, leading to a reduction in domain bias. Notably, the augmentations of RE target only the amplitude spectrum of high-dimensional features, with their phase spectrum remaining unchanged for maintaining domain-irrelevant semantics. In RS, we present a novel Mutual Exponential Meaning Average (MEMA) model parameter optimization strategy. This approach facilitates a dynamic exchange and integration of parameters between teacher and student models, enhancing model stability in out-of-domain performance throughout training process.", + "bbox": [ + 72, + 424, + 922, + 720 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We evaluate the proposed RES method on five DG benchmark datasets, including PACS [22], VLCS [10], OfficeHome [36], TerraIncognita [3], and DomainNet [30], and comparison experiment results indicate our RES outperforms multiple state-of-the-art DG methods.", + "bbox": [ + 72, + 726, + 934, + 806 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our contributions are summarized as follows: (1) We rethink the prediction error of DG from an extended bias-variance decomposition analytical framework. (2) Based on the bias-variance decomposition, we propose a RES method for DG, consisting of a Representation Enhancement (RE) module and a Representation Stabilization (RS) module, to respectively reduce cross-domain bias and variance. (3) Comparison experiments demonstrate that our RES method outperforms multiple state-of-the-art DG methods, verifying its effectiveness.", + "bbox": [ + 72, + 811, + 924, + 949 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "RES", + "bbox": [ + 789, + 14, + 840, + 30 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 900, + 16, + 920, + 30 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 79, + 52, + 329, + 74 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this section, we first give a brief review of the mainstreams DG methods. Then we delve into a detailed introduction to data augmentation methods and flatness-aware methods, which are relevant to our RES.", + "bbox": [ + 79, + 92, + 917, + 149 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.1 Domain Generalization", + "text_level": 1, + "bbox": [ + 79, + 178, + 423, + 196 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Various DG methods have been developed. One type of DG methods focus on distribution alignment among source domains to learn domain-invariant representations via all kinds of strategies, including domain alignment [23, 28, 46], adversarial training [23, 48], causal learning [25, 26], disentangled representation learning [31, 37], low-rank decomposition [33], self-supervised learning [1, 19], meta-learning [6, 7, 44], and data normalization [20, 34, 34].", + "bbox": [ + 79, + 208, + 917, + 326 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.2 Data Augmentation Methods", + "text_level": 1, + "bbox": [ + 79, + 352, + 498, + 373 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our RE module aligns with another paradigm of data augmentation methods, which enhance model generalization to unseen domains by diversifying source domain data at two distinct levels.", + "bbox": [ + 79, + 383, + 917, + 440 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "(1) Image level [8,41,42,47]. Yang et al. [42] and Zhou et al. [47] have used cross-domain image-to-image translation based on domain-adversarial learning, further applying these augmented images in model training. Xu et al. [41] introduced a novel Fourier-based image augmentation strategy for DG, based on the assumption that phase information in the frequency domain contains high-level semantics and is less affected by domain shift.", + "bbox": [ + 79, + 442, + 917, + 557 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "(2) Feature level [14, 24, 39]. Techniques such as the straightforward perturbation of high-dimensional feature embeddings with Gaussian noise by Li et al. [24], the use of Adaptive Instance Normalization (AdaIN) [17] for feature-level style randomization augmentation by Wang et al. [39], and the DomainDrop framework by Guo et al. [14] which drops high domain-activated channels to enhance feature channel robustness against domain shift, have been proposed.", + "bbox": [ + 79, + 559, + 917, + 674 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.3 Flatness-aware Methods", + "text_level": 1, + "bbox": [ + 79, + 702, + 435, + 721 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Flatness-aware methods aiming at avoiding unstable sharp optimization draws rising interest under the situation where a validation set has a different distribution from the test data, leading to a non-i.i.d. scenario. Cha et al. [4] hypothesized that in the non-i.i.d scenario the generalization gap between flat and sharp minima is more pronounced, and then proposed a Stochastic Weight Averaging Densely (SWAD) method aiming to locate flat minima and reduce domain gap. Moreover, Arpit et al. [2] argued that models trained on specific training domains often display erratic performance on test domains with distribution shifts. To mitigate this, they suggested the use of ensembling moving average (EoA) models, thereby decreasing uncertainty in test domains for enhanced domain generalization. Sharpness-Aware Minimization (SAM) [11] and an improved", + "bbox": [ + 79, + 733, + 917, + 947 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 79, + 17, + 94, + 29 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "W. Huang, Y. Shi et al.", + "bbox": [ + 159, + 14, + 392, + 30 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Sharpness-Aware Gradient Matching (SGAM) [38] were further proposed to enhance model generalization by simultaneously minimizing the loss value and its sharpness, leading to more robust models.", + "bbox": [ + 72, + 55, + 920, + 115 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3 RES-based Domain Generalization", + "text_level": 1, + "bbox": [ + 75, + 144, + 627, + 165 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1 Notations and Overview", + "text_level": 1, + "bbox": [ + 75, + 186, + 443, + 206 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Assuming there are $K$ distinct yet semantic-related source domains represented as $D_{s} = \\{D_{s}^{1}, D_{s}^{2}, \\ldots, D_{s}^{K}\\}$ , each originating from different distributions and comprising $N_{k}$ image-label pairs, denoted as $D_{s}^{k} = \\{(x_{i}, y_{i})\\}_{i=1}^{N_{k}}$ . In this study, the model training is independent of domain labels, so there is no domain subscript in $(x, y)$ . The primary objective of DG is to leverage these source domains to train a model, parameterized by $\\theta$ , that exhibits strong generalization capabilities and can effectively adapt to unseen target domains. In RES, there are two models sharing the same architecture, a student model $M^{S}$ parameterized with $\\theta^{S}$ and a teacher model $M^{T}$ parameterized with $\\theta^{T}$ . $M^{S}$ is used for model training and $M^{T}$ is utilized for model validation and test.", + "bbox": [ + 72, + 221, + 920, + 413 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The proposed RES-based DG framework is shown in Fig. 2 with its training process summarized in Algorithm 1. RES consists of: (1) a RE module that augments amplitude spectrum of feature maps while maintaining the phase spectrum, to expand cross-domain feature representation; and (2) a RS module that mutually fuses the parameters of the teacher and student models to each other.", + "bbox": [ + 72, + 416, + 917, + 512 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/5b1cd187f3bbe4313188096cb4604f8d9caeee0635f6c54a1e19f2fb70851cf7.jpg", + "image_caption": [ + "Fig. 2: Workflow of the proposed RES-based domain generalization framework, which consists of Representation Enhancement (RE) and Representation Stabilization (RS). RE aims to enhance feature representation to reduce cross-domain bias via a set of augmentations performing on the amplitude spectrum of frequency domain, while RS aims to further stabilize cross-domain performance to reduce cross-domain variance through a novel MEMA optimization strategy between student and teacher models." + ], + "image_footnote": [], + "bbox": [ + 82, + 547, + 917, + 814 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "RES", + "bbox": [ + 789, + 14, + 840, + 30 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 903, + 16, + 917, + 30 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2 Feature Frequency Augmentation-Based RE", + "text_level": 1, + "bbox": [ + 77, + 53, + 680, + 74 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Our RE modules augments source data at the feature level. During training, the student model $M^S$ is split into two parts: $M^{S-}$ which includes all layers up to a middle layer $l$ for feature map extraction, and $M^{S+}$ which consists of the remaining layers for the remaining feature extraction and classification.", + "bbox": [ + 77, + 92, + 920, + 170 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Decomposing Phase and Amplitude From Features. Given an image $x \\in \\mathcal{R}^{3 \\times \\bar{H} \\times \\bar{W}}$ where $\\bar{H} \\times \\bar{W}$ represents image spatial size, a corresponding multilayer feature map $f \\in \\mathcal{R}^{C \\times H \\times W}$ can be extracted by $M^{S^{-}}$ , where $C$ signifies the number of channels, and $H, W$ denote the height and width of the feature maps respectively. It is formulated as", + "bbox": [ + 77, + 204, + 920, + 302 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nf = M ^ {S -} (x). \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 428, + 321, + 912, + 341 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To delve into the frequency properties of these feature maps, we employ a 2D FFT across the spatial dimensions of each channel, ensuring that every layer of $f$ undergoes this transformative process. This approach allows us to comprehensively analyze the frequency content within each individual feature channel. For channel $c$ , its FFT formulation is as follows:", + "bbox": [ + 77, + 359, + 917, + 456 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nF ^ {c} (u, v) = \\sum_ {h = 0} ^ {H - 1} \\sum_ {w = 0} ^ {W - 1} f ^ {c} (h, w) \\cdot e ^ {- j 2 \\pi \\left(\\frac {u h}{H} + \\frac {v w}{W}\\right)}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 266, + 476, + 912, + 526 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $f^{c}(h,w)$ represents the pixel value at the spatial location $(h,w)$ within the range of $[H - 1,W - 1]$ in channel $c$ , and $F^{c}(u,v)$ denotes the complex frequency spectrum at the frequency coordinates $(u,v)$ of channel $c$ . The term $j$ is the imaginary unit.", + "bbox": [ + 77, + 541, + 917, + 621 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Then we compute a multi-layer phase spectrum denoted as $\\mathcal{P}$ and a multi-layer amplitude spectrum represented as $\\mathcal{A}$ , from the complex frequency spectrum $F$ , as", + "bbox": [ + 77, + 622, + 917, + 679 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {P} \\left(F ^ {c} (u, v)\\right) = \\arctan 2 \\left(\\operatorname {I m a g} \\left(F ^ {c} (u, v)\\right), \\operatorname {R e a l} \\left(F ^ {c} (u, v)\\right)\\right), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 208, + 700, + 912, + 718 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {A} \\left(F ^ {c} (u, v)\\right) = \\sqrt {\\operatorname {R e a l} \\left(F ^ {c} (u , v)\\right) ^ {2} + \\operatorname {I m a g} \\left(F ^ {c} (u , v)\\right) ^ {2}}. \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 227, + 760, + 912, + 780 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Here, $\\operatorname{Real}(F^c(u,v))$ and $\\operatorname{Imag}(F^c(u,v))$ indicate the real and imaginary parts of the channel- $c$ frequency spectrum $F^c$ , respectively. $\\mathcal{P}$ and $\\mathcal{A}$ have the same size of $C \\times H \\times W$ as $f$ . Drawing inspiration from [15,29,32,41], there is a common assumption that the phase component of frequency spectrum maintains the high-level semantics of the original data while the amplitude component contains its low-level statistics. In this study, we extend it from the image level to feature level, where we regard feature maps as high-dimension signals with the same characteristic of phase-amplitude decomposition as images.", + "bbox": [ + 77, + 791, + 917, + 947 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 77, + 16, + 94, + 30 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "W. Huang, Y. Shi et al.", + "bbox": [ + 157, + 14, + 392, + 32 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Frequency Augmentations on Features. To avoid damaging the semantics of features, augmentations are only implemented on the amplitude $\\mathcal{A}$ to obtain an augmented $\\mathcal{A}'$ while the phase $\\mathcal{P}$ keeps unchanged, as", + "bbox": [ + 72, + 56, + 920, + 115 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {A} ^ {\\prime} = \\operatorname {A u g} (\\mathcal {A}). \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 426, + 126, + 917, + 147 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Given a batch of images $\\{x_{1}, x_{2}, \\ldots, x_{B}\\}$ where $B$ is batch size, a corresponding batch of amplitudes $\\{\\mathcal{A}_1, \\mathcal{A}_2, \\ldots, \\mathcal{A}_B\\}$ can be obtained by Eqs. (3), (4), and (6). During each training iteration, we randomly select and apply one augmentation strategy from a pre-defined augmentation list, including none, random noise, random dropout, and mixup. Here none represents no augmentation, and the details of the rest three strategies are as follows:", + "bbox": [ + 72, + 156, + 920, + 271 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "- Random noise, denoted as $\\mathrm{RE}^{ns}$ . Given a multi-layer amplitude $\\mathcal{A}_i \\in \\mathcal{R}^{C \\times H \\times W}$ , we add Gaussian noise $G \\in \\mathcal{R}^{1 \\times H \\times W}$ to it. The elements of $G$ adhere to a Gaussian distribution characterized by mean $\\mu$ and standard deviation $\\sigma$ , i.e., $G \\sim \\mathcal{N}(\\mu, \\sigma^2)$ . Leveraging the broadcasting mechanism, this noise addition is equally conducted to each layer of $\\mathcal{A}_i$ , as:", + "bbox": [ + 89, + 279, + 922, + 380 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {A} _ {i} ^ {\\prime} = \\mathcal {A} _ {i} * (I + G) = \\mathcal {A} _ {i} * (I + \\mathcal {N} (\\mu , \\sigma^ {2})), \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 302, + 388, + 917, + 411 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $I$ represents the Identity Matrix. This approach guarantees that the same frequency components throughout all channels of the feature map are enhanced with the same noise intensity.", + "bbox": [ + 113, + 419, + 917, + 476 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "- Random dropout, denoted as $\\mathrm{RE}^{dp}$ . We randomly generate a binary dropout mask, $M \\in \\mathcal{R}^{1 \\times H \\times W}$ , where each element is set either to 0 with probability $p$ or to 1 with a reversal probability $1 - p$ . The dimension of $\\mathcal{M}$ is also expanded to $C \\times H \\times W$ by broadcasting. The dropout-augmented amplitude $\\mathcal{A}'$ is calculated by:", + "bbox": [ + 89, + 476, + 922, + 575 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {A} _ {i} ^ {\\prime} = \\mathcal {A} \\cdot M. \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 452, + 585, + 917, + 606 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "This operation introduces sparsity of amplitudes, aiming at reducing their overfitting for better robustness. It is worth mentioning that the element of $M$ at the location $[0,0]$ always equals 1, which represents the mean value of the feature map, to avoid over-masking on features.", + "bbox": [ + 113, + 616, + 922, + 694 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "- Mixup, denoted as $\\mathrm{RE}^{mix}$ . Unlike random noise and random dropout which operate on single amplitudes extracted from individual samples, mixup engages in a cross-sample amplitude interaction by linear interpolation on different amplitudes extracted from different samples. Specifically, for a given amplitude spectrum $\\mathcal{A}_i$ , mixup randomly selects another amplitude $\\mathcal{A}_j$ from the set $\\{\\mathcal{A}_1, \\mathcal{A}_2, \\dots, \\mathcal{A}_B\\}$ , and then blends $\\mathcal{A}_i$ with $\\mathcal{A}_j$ , resulting in a new, hybrid amplitude spectrum $\\mathcal{A}_i'$ as follows:", + "bbox": [ + 89, + 692, + 929, + 830 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {A} _ {i} ^ {\\prime} = \\alpha \\mathcal {A} _ {i} + (1 - \\alpha) \\mathcal {A} _ {j}, \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 397, + 840, + 917, + 861 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $\\alpha$ is a uniform random coefficient between [0,1]. This procedure enriches the amplitude characteristics by promoting a more diverse fusion of feature styles across domains. It boosts the model's generalization by augmenting intricate cross-domain amplitude patterns effectively.", + "bbox": [ + 113, + 869, + 922, + 949 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "RES", + "bbox": [ + 789, + 14, + 840, + 30 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 903, + 16, + 920, + 30 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Reconstructing Features. For each pair of phase and augmented amplitude components, $\\mathcal{P}$ and $\\mathcal{A}'$ , an augmented multi-layer feature map, $f'$ , can be reconstructed via inverse Fast Fourier Transform (iFFT) as", + "bbox": [ + 72, + 56, + 920, + 115 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\nf ^ {\\prime c} (h, w) = \\frac {1}{H W} \\sum_ {u = 0} ^ {H - 1} \\sum_ {v = 0} ^ {W - 1} \\mathcal {A} ^ {\\prime} \\left(F ^ {c} (u, v)\\right) \\cdot e ^ {j 2 \\pi \\left(\\frac {u h}{H} + \\frac {v w}{W}\\right)} \\cdot e ^ {j \\mathcal {P} \\left(F ^ {c} (u, v)\\right)} \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 133, + 147, + 917, + 200 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Here, $e^{j2\\pi \\left(\\frac{uh}{H} + \\frac{vw}{W}\\right)}$ is a complex exponential term representing the frequency component in the inverse Fourier transform, and $e^{j\\mathcal{P}(F^c(u,v))}$ reintroduces the phase information into the transform.", + "bbox": [ + 72, + 217, + 917, + 278 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Finally, $f'$ is used for further feature extraction and class classification by the remaining layers of the student model, $M^{S+}$ , as", + "bbox": [ + 72, + 279, + 917, + 318 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\np = M ^ {S +} \\left(f ^ {\\prime}\\right). \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 423, + 334, + 917, + 357 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "3.3 MEMA-based RS", + "text_level": 1, + "bbox": [ + 75, + 386, + 360, + 404 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "As shown in Fig. 2, there are two models, one student model $M^S$ with parameters $\\theta^S$ and one teacher model $M^T$ with parameters $\\theta^T$ . Their parameters undergo mutual updates to provide stabilization to each other, referred to as $\\mathrm{RE}^T$ and $\\mathrm{RE}^S$ for the teacher and student models, respectively. Their processes are formulated as", + "bbox": [ + 72, + 421, + 917, + 518 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\theta^ {T} \\leftarrow \\tau^ {T} \\theta^ {T} + (1 - \\tau^ {T}) \\theta^ {S}, \\quad \\tau^ {T} \\in [ 0, 1), \\tag {13} \\\\ \\theta^ {S} \\leftarrow (1 - \\tau^ {S}) \\theta^ {S} + \\tau^ {S} \\theta^ {T}, \\quad \\tau^ {S} \\in [ 0, 1). \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 290, + 531, + 912, + 578 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Here, only the student model $M^S$ is actively trained, while the teacher model $M^T$ is responsible solely for parameter updates. The student model $M^S$ is the primary learner, actively absorbing new knowledge and adapting to training data. In contrast, the teacher model $M^T$ acts as a robust and stable reference using moving average updating strategy, providing more robust and consistent parameters by reducing the variance of training batches [2].", + "bbox": [ + 72, + 596, + 917, + 713 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Both the values of $\\tau^T$ and $\\tau^S$ are small and they play a mutual role in this process: (1) Stability from the teacher model. With a small $\\tau^T$ , the teacher model's parameters change slowly, ensuring that they remain stable and consistent.(2) Gradual integration for the student model. A small $\\tau^S$ implies that the student model integrates the teacher's parameters slowly and steadily. This allows the student model to gradually assimilate the stable characteristics of the teacher model. To simplify the hyperparameter tuning process in experiments, $\\tau^T$ is set to a small constant value of 1e-3, while $\\tau^S$ is variable for different datasets.", + "bbox": [ + 72, + 713, + 917, + 869 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In summary, the student model acts as an active learner, continually evolving and assimilating new information, whereas the teacher model functions as a reliable guide, providing a stable parameter reference to shield the student model from the effects of cross-domain variance in training batches.", + "bbox": [ + 72, + 871, + 917, + 947 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 77, + 16, + 94, + 29 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "W. Huang, Y. Shi et al.", + "bbox": [ + 157, + 14, + 392, + 32 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/be812eeca85e4691a769c0d932d077986d2b813362b7fd860d5b88d1b4c3c762.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Algorithm 1 Training Process of our RES-based DG
1:Student model MS with initialized parameters θS, teacher model MT with initialized parameters θT, all training sample pairs {(xi,yi)}N, and batch size B
2:for iter = 1 to Niter do
3:Sample Batch Data: {(xi,yi)}B
4:Extract Feature: f← MS-(x)
5:Decompose Phase and Amplitude: P, A← FFT(f)
6:Augmentation Selection: Aug ← [random noise, random dropout, mixup, none]
7:Augment Amplitude: A' ← Aug(A)
8:Reconstruct feature: f' = iFFT(P, A')
9:Compute prediction: p ← MS+(f')
10:Compute Loss: E(θS) ← l(p, y)
11:Update Student Model with Loss: θS ← θS - η∇θL(θS)
12:Integrate Student Model to Teacher Model: θT = τTθT + (1 - τT)θS
13:Stabilize Student Model with Teacher Model: θS = (1 - τS)θS + τSθT
14:end for
", + "bbox": [ + 77, + 53, + 917, + 354 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "3.4 Supervision Loss", + "text_level": 1, + "bbox": [ + 75, + 390, + 348, + 409 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "During each training iteration with a batch size of $B$ , the empirical risk minimization (ERM) loss function is used to optimize the student model $M^S$ across all the training domains as", + "bbox": [ + 72, + 422, + 922, + 481 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {E} \\left(\\theta^ {S}\\right) = \\frac {1}{B} \\sum_ {i = 1} ^ {B} \\ell \\left(p _ {i}, y _ {i}\\right) = \\frac {1}{B} \\sum_ {i = 1} ^ {B} \\ell \\left(M ^ {S} \\left(x _ {i}; \\theta^ {S}\\right), y _ {i}\\right), \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 230, + 495, + 917, + 549 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where the detailed progress of $M^S (x_i;\\theta^S)$ come from Eqs. (3)-(7) and (12) and $\\ell$ represents the CE loss. Here $\\mathcal{E}(\\theta^S)$ represents the basic ERM loss.", + "bbox": [ + 75, + 562, + 917, + 604 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 75, + 632, + 312, + 655 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This section outlines benchmark datasets and experimental settings, conducts an ablation study on RES, compares it with advanced DG methods, examines the impact of $\\tau^S$ , and quantitatively assesses RES's effects in reducing cross-domain bias and variance and improving accuracy across domains. In Supplementary Material, we compare the effects of data augmentation from two perspectives: image-level vs. feature-level frequency augmentation, and normal feature augmentation vs. feature frequency augmentation. This comparison demonstrates the superiority of our feature frequency augmentation.", + "bbox": [ + 72, + 673, + 920, + 830 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.1 Benchmark Datasets", + "text_level": 1, + "bbox": [ + 75, + 858, + 397, + 876 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We evaluate the proposed RES method on five DG datasets: (1) PACS [22], which contains 9,991 images in 7 object categories across 4 diverse domains; (2) VLCS [10], which comprises 10,729 examples from 5 categories, collected from 4", + "bbox": [ + 72, + 889, + 920, + 949 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "RES", + "bbox": [ + 789, + 14, + 840, + 30 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 900, + 16, + 917, + 30 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "domains; (3) OfficeHome [36], which encompasses around 15,500 samples in 65 categories from four domains. (4) TerraIncognita [3], which contains 24,788 images within 10 categories from four domains; and (5) DomainNet [30], a vast dataset with 586,575 images across 345 categories, spanning 6 domains.", + "bbox": [ + 77, + 53, + 917, + 133 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.2 Experimental Settings", + "text_level": 1, + "bbox": [ + 79, + 160, + 414, + 180 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Basic settings. We utilize ResNet-50 [16], pretrained on ImageNet-1K [9], as the backbone. Our codes are constructed on Domainbed's framework [13], allocating $80\\%$ of in-domain data for training and the remaining $20\\%$ for validation, with test performed on an unseen domain. PACS, VLCS, and OfficeHome are trained for 10K iterations and validated every 300 iterations, TerraIncognita is trained for 3K iterations and validated every 300 iterations, and DomainNet undergoes 400K iterations of training and 1K iterations of validation. Adam is used as the optimizer, with a batch size of 128 and a learning rate of 5e-5.", + "bbox": [ + 79, + 191, + 917, + 349 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Settings of RES. In every training iteration, we apply the frequency augmentation of RE randomly to one of the four potential middle-layer positions within ResNet-50. These positions are strategically located at the ends of the $1^{st}$ , $2^{nd}$ , $3^{rd}$ , and $4^{th}$ residual blocks, ensuring diverse augmentation effects across different layers of the model. There are some hyperparameters in RES summarized in Table 1, where $\\tau^S$ is variable and tuned based on the validation performance as discussed in Section 4.4. All the experiments of RES are conducted for three times and the mean values and standard deviations are reported.", + "bbox": [ + 79, + 375, + 917, + 533 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/3a525fc0b294942b7369d68d06ba65b4b7d7c25da1b49d5d3eb558725375ce36.jpg", + "table_caption": [ + "Table 1: Hyperparameters of RES." + ], + "table_footnote": [], + "table_body": "
REnsREdpREmixRSS→TRST→S
G ~ N(0.75, 0.752)p = 0.5α ~ U(0,1)τT = 0.001τS is variable
", + "bbox": [ + 101, + 601, + 886, + 640 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.3 Ablation Study", + "text_level": 1, + "bbox": [ + 79, + 720, + 329, + 741 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "The ablation study of RES's modules, including sub-strategies of RE and RS, on PACS are provided in Table 2 and Fig. 3. RE employs random noise $(\\mathrm{RE}^{ns})$ , random dropout $(\\mathrm{RE}^{dp})$ , and mixup $(\\mathrm{RE}^{mix})$ techniques for feature augmentation, and RS employs MEMA containing representation stabilization for both student models $(\\mathrm{RS}^s)$ and teacher models $(\\mathrm{RS}^t)$ .", + "bbox": [ + 79, + 752, + 917, + 850 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "After applying all three RE strategies, there is significant performance improvement from ERM's 85.5 to (ERM+RE)'s 88.8 as shown in Table 2. with their training trajectory shown in Fig. 3. The results verify the effect of RE in reducing cross-domain bias. Additionally, the application of RS strategies reduces their standard deviations from ERM's 1.1 to (ERM+RS)'s 0.5, demonstrating their", + "bbox": [ + 79, + 850, + 917, + 947 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 79, + 16, + 104, + 30 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "W. Huang, Y. Shi et al.", + "bbox": [ + 157, + 14, + 392, + 32 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "effectiveness in reducing cross-domain variance. When RE and RS strategies are applied in conjunction, as shown in the last highlighted row of the table, the model achieves optimal test performance, with an average accuracy increase to 90.0 with the smallest standard deviation of 0.3, significantly surpassing ERM.", + "bbox": [ + 72, + 56, + 920, + 134 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Through this ablation study, the complementary effects of RE and RS strategies in enhancing model generalization across unseen domains are clearly demonstrated, underscoring the significance of reducing cross-domain bias-variance.", + "bbox": [ + 72, + 136, + 917, + 195 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/ea1f6c289a81a75e0e8b6d5407df0e263855d07301679d41534e096aa44e8a0f.jpg", + "table_caption": [ + "Table 2: Ablation study of the proposed RES on PACS." + ], + "table_footnote": [], + "table_body": "
ERMREnsREdpREmixRStRSsPACSAvg
ArtClipartPaintingSketch
✓84.4±0.681.1±0.796.3±0.580.2±2.585.5±1.1
✓✓88.8±1.282.2±0.597.0±0.681.2±0.687.3±0.7
✓✓87.6±0.382.7±1.096.8±0.882.1±1.687.3±0.9
✓✓89.4±0.281.4±1.196.6±0.485.1±0.988.5±0.7
✓✓✓✓89.0±0.983.0±0.497.0±0.485.1±0.788.8±0.6
✓✓87.9±0.183.1±0.296.0±0.482.2±1.287.3±0.5
✓✓✓90.9±0.482.0±0.597.9±0.382.2±0.888.2±0.5
✓✓✓✓✓✓91.6±0.484.2±0.398.1±0.186.1±0.490.0±0.3
", + "bbox": [ + 77, + 281, + 920, + 463 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/08137a2f85d6c20c90f34abf5ff7bf3125ae05adadf886a224b498eba15e1bb0.jpg", + "image_caption": [ + "Fig. 3: Training convergence of RES's modules on PACS." + ], + "image_footnote": [], + "bbox": [ + 79, + 556, + 290, + 666 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/3e7d83a1c426e5e53f56d75ba58d7fd4828ab466af565363c02db4e145af166a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 292, + 556, + 498, + 666 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/d7dae5339d1af2a7e57ddcf594eb8fb4c21397d097ababc5751992cc946f6dd9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 501, + 556, + 707, + 666 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/6530f02b979b7854895fed0920e83bed28e7ec39f543cea4f41de33364c2cb0d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 709, + 556, + 915, + 666 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4.4 Values of $\\tau^S$ in MEMA", + "text_level": 1, + "bbox": [ + 75, + 783, + 430, + 804 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "In RES, the MEMA's $\\tau^S$ is the sole hyperparameter requiring adjustment across various benchmarks. We explored five distinct values: [0, 1e-3, 5e-3, 1e-2, 5e-2] for each benchmark, with results shown in Fig. 4. The results indicate that the optimal $\\tau^S$ value varies by dataset, attributed to differences in their convergence speed. Subsequent comparison experiments will report out-of-domain accuracies for $\\tau^S$ corresponding to the highest in-domain accuracies observed.", + "bbox": [ + 72, + 830, + 920, + 949 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "RES", + "bbox": [ + 789, + 14, + 840, + 30 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 893, + 16, + 917, + 30 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/0f20f1dd4586800170ead541a7e393ac55aa4fa61987021677557514a810dd7e.jpg", + "image_caption": [ + "Fig. 4: In-domain (validation) and out-of-domain (test) accuracies vary with $\\tau^S$ ." + ], + "image_footnote": [], + "bbox": [ + 79, + 52, + 244, + 159 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/c0647528d9cb92293ebe9c318fbb5f9500179dbf9bd0705beea3990f4bc71b8e.jpg", + "image_caption": [], + "image_footnote": [ + "In-domain Accuracy Out-of-domain Accuracy" + ], + "bbox": [ + 246, + 52, + 411, + 159 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/5750340daef3e26b6ecded5b8076874aee4552a1b83e2bc1590faadbfed360e6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 411, + 52, + 576, + 159 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/d754bad40a69e14f967c846c05c2a1e028772096c5264ff0647e3cfeb6c2a31b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 578, + 52, + 743, + 159 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/7d738aba32c1e28d31f67b15a48b4c51be8761854e05b5eae2c6942b1267ec95.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 745, + 52, + 917, + 159 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "4.5 In Comparison With Advanced DG Methods", + "text_level": 1, + "bbox": [ + 75, + 243, + 692, + 265 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "To objectively assess the effectiveness of the proposed RES, we conducted comparative analyses against several state-of-the-art DG methods. This comparison spans the baseline ERM, various data augmentation strategies (e.g., Mixup [40], CCFP [21], DomainDrop [14]), flatness-aware methods (such as SWAD [4], SMA [2], SAGM [38], FAD [45]), and some other DG approaches (including FISH [35], RSC [18], DAC-SC [20], GVRT [27], MIRO [5]). These evaluations were performed across five benchmarks utilizing the ResNet-50 backbone, with results detailed in Table 3. The domain-wise performance of the datasets of our RES are provided in Supplementary Material. Notably, RES incorporates both data augmentation and flatness-aware strategies, positioning it uniquely to tackle the dual challenges of cross-domain bias and variance simultaneously.", + "bbox": [ + 72, + 276, + 922, + 489 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/93b0374640c31070a923ef58504389a5c6bb1c72d4bdd2ad7a99439b3f9cf8d5.jpg", + "table_caption": [ + "Table 3: Performance comparisons between some SOTA methods and our RES." + ], + "table_footnote": [], + "table_body": "
MethodPACSVLCSOfficeHomeTerraIncDomainNetAvg
ERM [ICLR'20] [13]85.7±0.577.4±0.367.5±0.547.2±0.441.2±0.263.8
RSC [ECCV'20] [18]85.2±0.977.1±0.565.5±0.946.6±1.038.9±0.562.7
Mixup [AAAI'20] [40]84.3±0.577.7±0.469.0±0.148.9±0.839.6±0.163.9
FISH [ICLR'21] [35]85.5±0.377.8±0.368.6±0.445.1±1.342.7±0.263.9
GVRT [ECCV'22] [27]85.1±0.379.0±0.270.1±0.148.0±0.244.1±0.165.2
MIRO [ECCV'22] [5]85.4±0.479.0±0.070.5±0.450.4±1.144.3±0.265.9
SMA [NeurIPS'22] [2]87.5±0.278.2±0.270.6±0.150.3±0.546.0±0.166.5
SWAD [NeurIPS'21] [4]88.1±0.179.1±0.170.6±0.250.0±0.346.5±0.166.9
CCFP [ICCV'23] [21]86.6±0.278.9±0.368.9±0.148.6±0.441.2±0.164.8
DAC-SC [CVPR'23] [20]87.5±0.178.7±0.370.3±0.246.5±0.344.9±0.165.6
FAD [ICCV'23] [45]88.2±0.278.9±0.869.2±0.545.7±1.044.4±0.165.3
SAGM [CVPR'23] [38]86.6±0.280.0±0.370.1±0.248.8±0.945.0±0.266.1
DomainDrop [ICCV'23] [14]87.9±0.379.8±0.368.7±0.151.5±0.444.4±0.566.5
Our RES90.0±0.379.8±0.271.8±0.351.4±0.646.7±0.267.9
", + "bbox": [ + 77, + 570, + 917, + 819 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "The table showcases the RES algorithm's performance across five benchmarks, with \"Avg\" representing the average performance across them. This provides a comprehensive view of each method's generalizability and effectiveness in DG. The proposed RES demonstrates superior performance, particularly high-", + "bbox": [ + 72, + 871, + 922, + 949 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 79, + 16, + 104, + 30 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "W. Huang, Y. Shi et al.", + "bbox": [ + 157, + 14, + 394, + 32 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "lighted by its leading average score, indicating its robustness and adaptability across a diverse set of domains. The comparative analysis confirms the superiority and effectiveness of the RES in mitigating cross-domain bias and variance.", + "bbox": [ + 72, + 55, + 920, + 115 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "4.6 Effects of RES on Domain Generalization", + "text_level": 1, + "bbox": [ + 75, + 144, + 648, + 165 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "This subsection quantitatively describes the advantages of RES on DG from three perspectives, including reducing cross-domain bias, reducing cross-domain variance, and improving both in-domain and out-of-domain performance.", + "bbox": [ + 72, + 177, + 920, + 237 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Reducing Cross-domain Bias.", + "text_level": 1, + "bbox": [ + 72, + 266, + 476, + 284 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "First, we measure the cosine similarity between the mean feature embeddings obtained from the global average pooling layer of ResNet-50 across both source and target datasets within PACS and OfficeHome. This assessment encompasses comparisons between the baseline ERM and our RES, as illustrated in Fig. 5. Our findings indicate that RES significantly enhances the cosine similarity between source and target data compared to ERM, demonstrating its capability in diminishing cross-domain l", + "bbox": [ + 72, + 286, + 476, + 559 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/98537107b1eac575a6fa9f9a64829580caa55ad8d6d66e01c521662b59d6c3a9.jpg", + "image_caption": [ + "Source-target Similarity of ERM" + ], + "image_footnote": [], + "bbox": [ + 503, + 305, + 704, + 406 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/3d30c3a52669d4733368fb3698831c99d9c06b62a9fa49ed93762eec97ee1fdf.jpg", + "image_caption": [ + "Source-target Similarity of RES", + "Fig. 5: Cosine similarity between mean features of source domains and target domain. The higher the similarity is, the lower the cross-domain bias is." + ], + "image_footnote": [], + "bbox": [ + 709, + 305, + 907, + 406 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Reducing Cross-domain Variance. Secondly, we present the training convergence trajectories alongside the corresponding cumulative instability for four DG scenarios within the PACS dataset, as illustrated in Fig. 6. It is observable that, in comparison to the baseline method ERM represented by the blue color, the proposed RES depicted in red markedly reduces the fluctuations in both in-domain and out-of-domain accuracies with almost no instability increase after the initial increase stage during training, which spans approximately the first 2000 iterations. Benefiting from the stable and narrower performance gap between in-domain (validation) and out-of-domain (test) sets, the best models chosen based on validation sets demonstrate consistent performance on test sets.", + "bbox": [ + 72, + 588, + 920, + 785 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Improving Both In-domain and Out-of-domain Performance. As shown in Table 4, the comparison between ERM and RES on PACS illustrates significant enhancements in both in-domain and out-of-domain accuracies achieved by RES. While ERM delivers reliable in-domain performance, it falls short in adapting to unseen domains. In contrast, RES not only boosts in-domain accuracy from ERM's 96.5 to its 98.0 but also lifts the out-of-domain accuracy", + "bbox": [ + 72, + 830, + 920, + 949 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "RES", + "bbox": [ + 789, + 14, + 840, + 30 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 893, + 16, + 917, + 30 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/39fe8bbb5f1e067860c76e62dfea3015376927079bca55e272d0b05d37440c84.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 79, + 53, + 288, + 157 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/f3a31b1b8db827223e6e43aad1ce85273232cfc0491c1404b9177188d29cffe2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 290, + 53, + 493, + 156 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/a0521d58739d241ef42273e6b53f1ebcee8198a2033df73d3a97f6dcb2d0ab35.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 496, + 53, + 704, + 156 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/027a54cdeccde8d259cee14db751fb3daa804cbdfa50d12968278f4223fd0561.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 709, + 53, + 915, + 156 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/e595c659d3bd6cb12a812db44ae5b0ddf8c57f3a4112414f030e7072d7a8836a.jpg", + "image_caption": [ + "Fig. 6: Training stability of ERM and RES on PACS, evaluated by the absolute first derivative of accuracy every 300 iterations $(|f'(Acc)| = |Acc_t - Acc_{t-1}|)$ , capturing accuracy change rate. Accumulative instability aggregates $|f'(Acc)|$ up to iteration $t$ , with lower values indicating more stable performance and less variance." + ], + "image_footnote": [], + "bbox": [ + 75, + 160, + 285, + 287 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/fc040e88401742c649beef52436be9131fdd48af3115187c790327d92a5dc5e2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 290, + 160, + 496, + 286 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/59be9bedf7b27b910332f53db8bd723ff6bad11f9a825a99fabf887ac72633c9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 498, + 160, + 707, + 286 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/a4afb372df88a610c22aa9610c66d5cd420b3c56ff77c58e817fcea3176f6180.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 711, + 160, + 917, + 286 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "from $85.5 \\pm 1.1$ to $90.0 \\pm 0.3$ . This robust improvement underscores RES's robust capability to refine model performance across various distributions.", + "bbox": [ + 72, + 411, + 917, + 450 + ], + "page_idx": 13 + }, + { + "type": "table", + "img_path": "images/04a18eb3e252cd809c5273fcc125c6b7359511727a5e405b209c7a1e1e98a54d.jpg", + "table_caption": [ + "Table 4: In-domain and out-of-domain accuracies of PACS between ERM and RES." + ], + "table_footnote": [], + "table_body": "
SplitMethodPACSAvg
ArtClipartPaintingSketch
In-domain (Validation Set)ERM97.2±0.097.0±0.495.4±0.296.3±0.496.5±0.3
Our RES98.4±0.298.0±0.097.4±0.398.2±0.398.0±0.2
Out-of-domain (Test Set)ERM84.4±0.681.1±0.796.3±0.580.2±2.585.5±1.1
Our RES91.6±0.484.2±0.398.1±0.186.1±0.490.0±0.3
", + "bbox": [ + 79, + 518, + 915, + 632 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Conclusion", + "text_level": 1, + "bbox": [ + 75, + 713, + 288, + 734 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In this study, we explored DG through bias-variance decomposition, an analytical framework that offers deeper insights into the challenges and solutions for cross-domain generalization in deep learning models. Our approach quantified the detrimental effects of cross-domain bias and variance on model performance and therefore introduced the Representation Enhancement-Stabilization (RES) method as a potent solution. The RES method effectively minimizes the cross-domain bias and variance, enhancing model adaptability and generalization across unseen target domains. Empirical evaluations on five DG benchmarks reveals that the proposed RES markedly outperforms multiple advanced DG methods in improving model generalization.", + "bbox": [ + 72, + 754, + 917, + 947 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 79, + 16, + 104, + 30 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "W. Huang, Y. Shi et al.", + "bbox": [ + 157, + 14, + 394, + 32 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 82, + 53, + 348, + 74 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "The work was supported by the German Research Foundation (DFG GZ: ZH 498/18-1; Project number: 519016653). The work of Z. Xiong was supported by the German Federal Ministry for the Environment, Nature Conservation, Nuclear Safety and Consumer Protection (BMUV) based on a resolution of the German Bundestag (grant number: 67KI32002B; Acronym: $EKAPEx$ ). The work of X. Zhu was also supported by the German Federal Ministry of Education and Research (BMBF) in the framework of the international future AI lab \"AI4EO - Artificial Intelligence for Earth Observation: Reasoning, Uncertainties, Ethics and Beyond\" (grant number: 01DD20001), by German Federal Ministry for Economic Affairs and Climate Action in the framework of the \"national center of excellence ML4Earth\" (grant number: 50EE2201C) and by the Munich Center for Machine Learning.", + "bbox": [ + 79, + 94, + 917, + 326 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "RES", + "bbox": [ + 791, + 16, + 837, + 29 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 893, + 17, + 915, + 29 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 77, + 53, + 237, + 73 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Albuquerque, I., Naik, N., Li, J., Keskar, N., Socher, R.: Improving out-of-distribution generalization via multi-task self-supervised pretraining. arXiv preprint arXiv:2003.13525 (2020)", + "2. Arpit, D., Wang, H., Zhou, Y., Xiong, C.: Ensemble of averages: Improving model selection and boosting performance in domain generalization. Advances in Neural Information Processing Systems 35, 8265-8277 (2022)", + "3. Beery, S., Van Horn, G., Perona, P.: Recognition in terra incognita. In: Proceedings of the European conference on computer vision (ECCV). pp. 456-473 (2018)", + "4. Cha, J., Chun, S., Lee, K., Cho, H.C., Park, S., Lee, Y., Park, S.: Swad: Domain generalization by seeking flat minima. Advances in Neural Information Processing Systems 34, 22405-22418 (2021)", + "5. Cha, J., Lee, K., Park, S., Chun, S.: Domain generalization by mutual-information regularization with pre-trained models. In: European Conference on Computer Vision. pp. 440-457. Springer (2022)", + "6. Chen, C., Li, J., Han, X., Liu, X., Yu, Y.: Compound domain generalization via meta-knowledge encoding. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7119-7129 (2022)", + "7. Chen, J., Gao, Z., Wu, X., Luo, J.: Meta-causal learning for single domain generalization. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7683-7692 (2023)", + "8. Cubuk, E.D., Zoph, B., Shlens, J., Le, Q.V.: Randaugment: Practical automated data augmentation with a reduced search space. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops. pp. 702-703 (2020)", + "9. Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: Imagenet: A large-scale hierarchical image database. In: 2009 IEEE conference on computer vision and pattern recognition. pp. 248-255. IEEE (2009)", + "10. Fang, C., Xu, Y., Rockmore, D.N.: Unbiased metric learning: On the utilization of multiple datasets and web images for softening bias. In: Proceedings of the IEEE International Conference on Computer Vision. pp. 1657-1664 (2013)", + "11. Foret, P., Kleiner, A., Mobahi, H., Neyshabur, B.: Sharpness-aware minimization for efficiently improving generalization. arXiv preprint arXiv:2010.01412 (2020)", + "12. Geman, S., Bienenstock, E., Doursat, R.: Neural networks and the bias/variance dilemma. Neural computation 4(1), 1-58 (1992)", + "13. Gulrajani, I., Lopez-Paz, D.: In search of lost domain generalization. In: International Conference on Learning Representations (2020)", + "14. Guo, J., Qi, L., Shi, Y.: Domaindrop: Suppressing domain-sensitive channels for domain generalization. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 19114-19124 (2023)", + "15. Hansen, B.C., Hess, R.F.: Structural sparseness and spatial phase alignment in natural scenes. JOSA A 24(7), 1873-1885 (2007)", + "16. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 770-778 (2016)", + "17. Huang, X., Belongie, S.: Arbitrary style transfer in real-time with adaptive instance normalization. In: Proceedings of the IEEE international conference on computer vision. pp. 1501-1510 (2017)" + ], + "bbox": [ + 79, + 97, + 920, + 947 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 79, + 16, + 104, + 30 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "W. Huang, Y. Shi et al.", + "bbox": [ + 157, + 14, + 392, + 32 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "18. Huang, Z., Wang, H., Xing, E.P., Huang, D.: Self-challenging improves cross-domain generalization. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part II 16. pp. 124-140. Springer (2020)", + "19. Kim, D., Yoo, Y., Park, S., Kim, J., Lee, J.: Selfreg: Self-supervised contrastive regularization for domain generalization. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 9619-9628 (2021)", + "20. Lee, S., Bae, J., Kim, H.Y.: Decompose, adjust, compose: Effective normalization by playing with frequency for domain generalization. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11776-11785 (2023)", + "21. Li, C., Zhang, D., Huang, W., Zhang, J.: Cross contrasting feature perturbation for domain generalization. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 1327-1337 (2023)", + "22. Li, D., Yang, Y., Song, Y.Z., Hospedales, T.M.: Deeper, broader and artier domain generalization. In: Proceedings of the IEEE international conference on computer vision. pp. 5542-5550 (2017)", + "23. Li, H., Pan, S.J., Wang, S., Kot, A.C.: Domain generalization with adversarial feature learning. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 5400-5409 (2018)", + "24. Li, P., Li, D., Li, W., Gong, S., Fu, Y., Hospedales, T.M.: A simple feature augmentation for domain generalization. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 8886-8895 (2021)", + "25. Lv, F., Liang, J., Li, S., Zang, B., Liu, C.H., Wang, Z., Liu, D.: Causality inspired representation learning for domain generalization. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8046-8056 (2022)", + "26. Mahajan, D., Tople, S., Sharma, A.: Domain generalization using causal matching. In: International Conference on Machine Learning. pp. 7313-7324. PMLR (2021)", + "27. Min, S., Park, N., Kim, S., Park, S., Kim, J.: Grounding visual representations with texts for domain generalization. In: European Conference on Computer Vision. pp. 37-53. Springer (2022)", + "28. Muandet, K., Balduzzi, D., Scholkopf, B.: Domain generalization via invariant feature representation. In: International conference on machine learning. pp. 10-18. PMLR (2013)", + "29. Oppenheim, A., Lim, J., Kopec, G., Pohlig, S.: Phase in speech and pictures. In: ICASSP'79. IEEE International Conference on Acoustics, Speech, and Signal Processing. vol. 4, pp. 632-637. IEEE (1979)", + "30. Peng, X., Bai, Q., Xia, X., Huang, Z., Saenko, K., Wang, B.: Moment matching for multi-source domain adaptation. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 1406-1415 (2019)", + "31. Peng, X., Huang, Z., Sun, X., Saenko, K.: Domain agnostic learning with disentangled representations. In: International Conference on Machine Learning. pp. 5102-5112. PMLR (2019)", + "32. Piotrowski, L.N., Campbell, F.W.: A demonstration of the visual importance and flexibility of spatial-frequency amplitude and phase. Perception 11(3), 337-346 (1982)", + "33. Piratla, V., Netrapalli, P., Sarawagi, S.: Efficient domain generalization via common-specific low-rank decomposition. In: International Conference on Machine Learning. pp. 7728-7738. PMLR (2020)" + ], + "bbox": [ + 79, + 56, + 920, + 947 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "RES", + "bbox": [ + 789, + 14, + 840, + 30 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 893, + 16, + 917, + 30 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "34. Seo, S., Suh, Y., Kim, D., Kim, G., Han, J., Han, B.: Learning to optimize domain specific normalization for domain generalization. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXII 16. pp. 68-83. Springer (2020)", + "35. Shi, Y., Seely, J., Torr, P.H., Siddharth, N., Hannun, A., Usunier, N., Synnaeve, G.: Gradient matching for domain generalization. arXiv preprint arXiv:2104.09937 (2021)", + "36. Venkateswara, H., Eusebio, J., Chakraborty, S., Panchanathan, S.: Deep hashing network for unsupervised domain adaptation. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 5018-5027 (2017)", + "37. Wang, G., Han, H., Shan, S., Chen, X.: Cross-domain face presentation attack detection via multi-domain disentangled representation learning. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 6678-6687 (2020)", + "38. Wang, P., Zhang, Z., Lei, Z., Zhang, L.: Sharpness-aware gradient matching for domain generalization. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3769-3778 (2023)", + "39. Wang, Y., Qi, L., Shi, Y., Gao, Y.: Feature-based style randomization for domain generalization. IEEE Transactions on Circuits and Systems for Video Technology 32(8), 5495-5509 (2022)", + "40. Xu, M., Zhang, J., Ni, B., Li, T., Wang, C., Tian, Q., Zhang, W.: Adversarial domain adaptation with domain mixup. In: Proceedings of the AAAI conference on artificial intelligence. vol. 34, pp. 6502-6509 (2020)", + "41. Xu, Q., Zhang, R., Zhang, Y., Wang, Y., Tian, Q.: A fourier-based framework for domain generalization. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14383-14392 (2021)", + "42. Yang, F.E., Cheng, Y.C., Shiau, Z.Y., Wang, Y.C.F.: Adversarial teacher-student representation learning for domain generalization. Advances in Neural Information Processing Systems 34, 19448-19460 (2021)", + "43. Yang, Z., Yu, Y., You, C., Steinhardt, J., Ma, Y.: Rethinking bias-variance tradeoff for generalization of neural networks. In: International Conference on Machine Learning. pp. 10767-10777. PMLR (2020)", + "44. Zhang, J., Qi, L., Shi, Y., Gao, Y.: Mvdg: A unified multi-view framework for domain generalization. In: European Conference on Computer Vision. pp. 161-177. Springer (2022)", + "45. Zhang, X., Xu, R., Yu, H., Dong, Y., Tian, P., Cui, P.: Flatness-aware minimization for domain generalization. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5189-5202 (2023)", + "46. Zhao, S., Gong, M., Liu, T., Fu, H., Tao, D.: Domain generalization via entropy regularization. Advances in Neural Information Processing Systems 33, 16096-16107 (2020)", + "47. Zhou, K., Yang, Y., Hospedales, T., Xiang, T.: Deep domain-adversarial image generation for domain generalisation. In: Proceedings of the AAAI conference on artificial intelligence. vol. 34, pp. 13025-13032 (2020)", + "48. Zhu, W., Lu, L., Xiao, J., Han, M., Luo, J., Harrison, A.P.: Localized adversarial domain generalization. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7108-7118 (2022)" + ], + "bbox": [ + 79, + 56, + 917, + 895 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 79, + 16, + 104, + 29 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "W. Huang, Y. Shi et al.", + "bbox": [ + 157, + 14, + 392, + 32 + ], + "page_idx": 17 + } +] \ No newline at end of file diff --git a/2024/Representation Enhancement-Stabilization_ Reducing Bias-Variance of Domain Generalization/f28a3a3b-4152-48f4-99b3-107ae4c33ff2_model.json b/2024/Representation Enhancement-Stabilization_ Reducing Bias-Variance of Domain Generalization/f28a3a3b-4152-48f4-99b3-107ae4c33ff2_model.json new file mode 100644 index 0000000000000000000000000000000000000000..0fe3814981fad36427e59d0637509df159a6282e --- /dev/null +++ b/2024/Representation Enhancement-Stabilization_ Reducing Bias-Variance of Domain Generalization/f28a3a3b-4152-48f4-99b3-107ae4c33ff2_model.json @@ -0,0 +1,2612 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.05, + 0.92, + 0.108 + ], + "angle": 0, + "content": "Representation Enhancement-Stabilization: Reducing Bias-Variance of Domain Generalization" + }, + { + "type": "text", + "bbox": [ + 0.127, + 0.14, + 0.872, + 0.161 + ], + "angle": 0, + "content": "Wei Huang1, Yilei Shi1, Zhitong Xiong1, and Xiao Xiang Zhu1,2" + }, + { + "type": "text", + "bbox": [ + 0.124, + 0.177, + 0.876, + 0.23 + ], + "angle": 0, + "content": "1. Technical University of Munich, Arcisstraße 21, 80333 München, Germany \n2. Munich Center for Machine Learning, 80333 Munich \n{w2wei.huang, yilei.shi, zhitong.xiong, xiaoxiang.zhu}@tum.de" + }, + { + "type": "text", + "bbox": [ + 0.149, + 0.271, + 0.851, + 0.555 + ], + "angle": 0, + "content": "Abstract. Domain Generalization (DG) focuses on enhancing the generalization of deep learning models trained on multiple source domains to adapt to unseen target domains. This paper explores DG through the lens of bias-variance decomposition, uncovering that test errors in DG predominantly arise from cross-domain bias and variance. Inspired by this insight, we introduce a Representation Enhancement-Stabilization (RES) framework, comprising a Representation Enhancement (RE) module and a Representation Stabilization (RS) module. In RE, a novel set of feature frequency augmentation techniques is used to progressively reduce cross-domain bias during feature extraction. Furthermore, in RS, a novel Mutual Exponential Moving Average (MEMA) strategy is designed to stabilize model optimization for diminishing cross-domain variance during training. Collectively, the whole RES method can significantly enhance model generalization. We evaluate RES on five benchmark datasets and the results show that it outperforms multiple advanced DG methods. Our code will be available at https://github.com/zhu-xlab/DG-RES." + }, + { + "type": "text", + "bbox": [ + 0.148, + 0.573, + 0.851, + 0.608 + ], + "angle": 0, + "content": "Keywords: Domain generalization \\(\\cdot\\) Frequency Domain \\(\\cdot\\) Data Augmentation \\(\\cdot\\) Bias-variance Decomposition" + }, + { + "type": "title", + "bbox": [ + 0.081, + 0.639, + 0.31, + 0.658 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.08, + 0.677, + 0.918, + 0.832 + ], + "angle": 0, + "content": "In recent years, deep learning models have achieved remarkable advancements in computer vision tasks, primarily under the assumption that training and test data are independent and identically distributed. However, in real-world scenarios, this assumption often breaks down due to domain shift/bias, where the distribution of test data markedly differs from that of training data. Domain shift can significantly degrade the performance of deep models on unfamiliar, unseen target domains. For instance, self-driving models trained under daylight conditions might fail to perform effectively in nighttime environments." + }, + { + "type": "text", + "bbox": [ + 0.08, + 0.833, + 0.918, + 0.95 + ], + "angle": 0, + "content": "To address these challenges, Domain Generalization (DG) has been proposed as a robust solution. DG focuses on training models on multiple diverse yet related source domains for robust performance on arbitrary unseen target domains. Numerous DG approaches have been developed, including adversarial training [23, 48], disentangled representation learning [37], meta-learning [7], model ensemble [2], flatness optimization [4], and domain data augmentation [8, 41, 47]." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.079, + 0.017, + 0.096, + 0.03 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.159, + 0.016, + 0.395, + 0.033 + ], + "angle": 0, + "content": "W. Huang, Y. Shi et al." + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.057, + 0.921, + 0.173 + ], + "angle": 0, + "content": "In this study we highlight the cross-domain generalization of deep models from the perspective of bias-variance decomposition. Geman et al. [12] first used bias-variance to decompose the mean square error (MSE) of neural networks. Building on this, Yang et al. [43] developed the bias-variance decomposition for the cross-entropy (CE) loss in classification tasks, which was extended by Arpit et al. [2] to analyze the impact of variance in DG, as illustrated below:" + }, + { + "type": "equation", + "bbox": [ + 0.089, + 0.186, + 0.916, + 0.231 + ], + "angle": 0, + "content": "\\[\n\\mathbb {E} _ {x, y \\sim D _ {s}} [ \\mathrm {C E} (y, f (x; D _ {s})) ] = \\underbrace {\\mathbb {E} _ {x , y \\sim D _ {s}} [ \\mathrm {C E} (y , \\bar {f} (x)) ]} _ {\\text {B i a s} ^ {2}} + \\underbrace {\\mathbb {E} _ {x , D _ {s}} [ \\mathrm {K L} (\\bar {f} (x) , f (x ; D _ {s})) ]} _ {\\text {V a r i a n c e}},\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.728, + 0.231, + 0.916, + 0.248 + ], + "angle": 0, + "content": "(1)" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.25, + 0.92, + 0.326 + ], + "angle": 0, + "content": "where \\( D_{s} \\) represents source domains used for training. The formula decomposes the expected source CE loss into two components: bias and variance. Bias evaluates the discrepancy between model's average predictions and actual outcomes, while variance measures sensitivity to fluctuations during training." + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.327, + 0.919, + 0.366 + ], + "angle": 0, + "content": "Motivated by them, we extend the bias-variance decomposition to DG to comprehensively analysis the cross-domain prediction errors as the following:" + }, + { + "type": "equation", + "bbox": [ + 0.149, + 0.379, + 0.916, + 0.52 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathbb {E} _ {x, y \\sim D _ {t}} [ \\mathrm {C E} (y, f (x _ {t}; D _ {t})) ] = \\mathbb {E} _ {x, y \\sim D _ {s}} [ \\mathrm {C E} (y, f (x _ {s} + \\delta ; D _ {s})) ] \\\\ = \\underbrace {\\mathbb {E} _ {x , y \\sim D _ {s}} [ \\mathrm {C E} (y , \\bar {f} (x _ {s} + \\delta)) ]} \\\\ \\text {C r o s s - D o m a i n B i a s} ^ {2} \\\\ + \\underbrace {\\mathbb {E} _ {x \\sim D _ {s}} \\left[ \\mathrm {K L} \\left(\\bar {f} \\left(x _ {s} + \\delta\\right) , f \\left(x _ {s} + \\delta ; D _ {s}\\right)\\right) \\right]} _ {\\text {C r o s s - D o m a i n V a r i a n c e}} \\tag {2} \\\\ + \\Delta_ {d o m a i n} (\\text {B i a s} ^ {2}, \\text {V a r i a n c e}). \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.532, + 0.92, + 0.61 + ], + "angle": 0, + "content": "This formula analyzes the expected CE loss when a model trained in source domains \\( D_{s} \\), denoted as \\( f(x_{s};D_{s}) \\), is applied to new target domains \\( D_{t} \\), with respect to domain shift \\( \\delta \\) between source and target images, \\( x_{s} \\) and \\( x_{t} \\). The formula is decomposed into three components:" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.62, + 0.92, + 0.757 + ], + "angle": 0, + "content": "- (1) Cross-Domain Bias, as shown in the left of Fig. 1. It quantifies the discrepancy between the model's average prediction in the target domain and the actual labels. High cross-domain bias dramatically enlarges the divergence between the distribution learnt from source domains and the true underlying distribution of target domains. By progressively reducing bias in feature extraction during training, i.e., \\( f(x_{s}) \\gets f(x_{s} + \\delta) \\), there is an approximate bias error \\( \\mathbb{E}_{x,y\\sim D_s}[\\mathrm{CE}(y,\\bar{f} (x))]\\gets \\mathbb{E}_{x,y\\sim D_s}[\\mathrm{CE}(y,\\bar{f} (x_s + \\delta))] \\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.757, + 0.922, + 0.892 + ], + "angle": 0, + "content": "- (2) Cross-Domain Variance, as shown in the right of Fig. 1. It measures the variability and sensitivity in the model's predictions from source to target domains. The model trained on the source domain tends to generate unstable predictions when confronted with the target domain data, with elevated cross-domain variance. Based on cross-domain bias reduction in (1), further decreasing cross-domain variance results in an approximate variance error \\(\\mathbb{E}_{x\\sim D_s}[\\mathrm{KL}(\\bar{f} (x_s),f(x_s;D_s))]\\gets \\mathbb{E}_{x\\sim D_s}[\\mathrm{KL}(\\bar{f} (x_s + \\delta),f(x_s + \\delta ;D_s))]\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.892, + 0.922, + 0.947 + ], + "angle": 0, + "content": "- (3) \\(\\varDelta_{domain}(\\mathrm{Bias}^2\\), Variance). This term denotes the extra test error of the intertwined effects of the cross-domain bias and variance. The decrease of both bias and variance can contribute to its reduction." + }, + { + "type": "list", + "bbox": [ + 0.09, + 0.62, + 0.922, + 0.947 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.79, + 0.016, + 0.842, + 0.032 + ], + "angle": 0, + "content": "RES" + }, + { + "type": "page_number", + "bbox": [ + 0.903, + 0.017, + 0.921, + 0.032 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.057, + 0.923, + 0.138 + ], + "angle": 0, + "content": "Overall, diminishing cross-domain bias and variance can narrow the gap between Eq. (1) and Eq. (2), enabling models trained on source domains to effectively adapt to target domains. This analytical framework provides actionable insights for training models capable of generalizing across diverse data distributions." + }, + { + "type": "image", + "bbox": [ + 0.08, + 0.192, + 0.526, + 0.319 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.152, + 0.321, + 0.453, + 0.336 + ], + "angle": 0, + "content": "(a) Challenge of cross-domain bias" + }, + { + "type": "image", + "bbox": [ + 0.563, + 0.188, + 0.922, + 0.317 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.574, + 0.32, + 0.911, + 0.337 + ], + "angle": 0, + "content": "(b) Challenge of cross-domain variance" + }, + { + "type": "image_caption", + "bbox": [ + 0.086, + 0.351, + 0.908, + 0.371 + ], + "angle": 0, + "content": "Fig. 1: Two challenges impeding DG from the view of bias-variance decomposition." + }, + { + "type": "text", + "bbox": [ + 0.074, + 0.426, + 0.923, + 0.721 + ], + "angle": 0, + "content": "Based on this analytical framework, we proposed a joint Representation Enhancement-Stabilization (RES) method, which consists of a Representation Enhancement (RE) module and a Representation Stabilization (RS) module to alleviate cross-domain bias ans cross-domain variance, respectively. In RE, we propose a novel set of feature augmentation techniques within the frequency domain extracted through Fast Fourier Transform (FFT), including random noise, random dropout, and mixup. RE effectively broadens cross-domain representation space and thereby enhances robust representation learning, leading to a reduction in domain bias. Notably, the augmentations of RE target only the amplitude spectrum of high-dimensional features, with their phase spectrum remaining unchanged for maintaining domain-irrelevant semantics. In RS, we present a novel Mutual Exponential Meaning Average (MEMA) model parameter optimization strategy. This approach facilitates a dynamic exchange and integration of parameters between teacher and student models, enhancing model stability in out-of-domain performance throughout training process." + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.727, + 0.935, + 0.807 + ], + "angle": 0, + "content": "We evaluate the proposed RES method on five DG benchmark datasets, including PACS [22], VLCS [10], OfficeHome [36], TerraIncognita [3], and DomainNet [30], and comparison experiment results indicate our RES outperforms multiple state-of-the-art DG methods." + }, + { + "type": "text", + "bbox": [ + 0.074, + 0.813, + 0.926, + 0.951 + ], + "angle": 0, + "content": "Our contributions are summarized as follows: (1) We rethink the prediction error of DG from an extended bias-variance decomposition analytical framework. (2) Based on the bias-variance decomposition, we propose a RES method for DG, consisting of a Representation Enhancement (RE) module and a Representation Stabilization (RS) module, to respectively reduce cross-domain bias and variance. (3) Comparison experiments demonstrate that our RES method outperforms multiple state-of-the-art DG methods, verifying its effectiveness." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.08, + 0.018, + 0.095, + 0.03 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.16, + 0.016, + 0.393, + 0.032 + ], + "angle": 0, + "content": "W. Huang, Y. Shi et al." + }, + { + "type": "title", + "bbox": [ + 0.081, + 0.053, + 0.331, + 0.075 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "text", + "bbox": [ + 0.081, + 0.093, + 0.918, + 0.15 + ], + "angle": 0, + "content": "In this section, we first give a brief review of the mainstreams DG methods. Then we delve into a detailed introduction to data augmentation methods and flatness-aware methods, which are relevant to our RES." + }, + { + "type": "title", + "bbox": [ + 0.081, + 0.179, + 0.424, + 0.198 + ], + "angle": 0, + "content": "2.1 Domain Generalization" + }, + { + "type": "text", + "bbox": [ + 0.081, + 0.209, + 0.918, + 0.327 + ], + "angle": 0, + "content": "Various DG methods have been developed. One type of DG methods focus on distribution alignment among source domains to learn domain-invariant representations via all kinds of strategies, including domain alignment [23, 28, 46], adversarial training [23, 48], causal learning [25, 26], disentangled representation learning [31, 37], low-rank decomposition [33], self-supervised learning [1, 19], meta-learning [6, 7, 44], and data normalization [20, 34, 34]." + }, + { + "type": "title", + "bbox": [ + 0.081, + 0.354, + 0.5, + 0.374 + ], + "angle": 0, + "content": "2.2 Data Augmentation Methods" + }, + { + "type": "text", + "bbox": [ + 0.081, + 0.384, + 0.918, + 0.441 + ], + "angle": 0, + "content": "Our RE module aligns with another paradigm of data augmentation methods, which enhance model generalization to unseen domains by diversifying source domain data at two distinct levels." + }, + { + "type": "text", + "bbox": [ + 0.08, + 0.443, + 0.918, + 0.559 + ], + "angle": 0, + "content": "(1) Image level [8,41,42,47]. Yang et al. [42] and Zhou et al. [47] have used cross-domain image-to-image translation based on domain-adversarial learning, further applying these augmented images in model training. Xu et al. [41] introduced a novel Fourier-based image augmentation strategy for DG, based on the assumption that phase information in the frequency domain contains high-level semantics and is less affected by domain shift." + }, + { + "type": "text", + "bbox": [ + 0.08, + 0.56, + 0.918, + 0.676 + ], + "angle": 0, + "content": "(2) Feature level [14, 24, 39]. Techniques such as the straightforward perturbation of high-dimensional feature embeddings with Gaussian noise by Li et al. [24], the use of Adaptive Instance Normalization (AdaIN) [17] for feature-level style randomization augmentation by Wang et al. [39], and the DomainDrop framework by Guo et al. [14] which drops high domain-activated channels to enhance feature channel robustness against domain shift, have been proposed." + }, + { + "type": "title", + "bbox": [ + 0.081, + 0.704, + 0.437, + 0.723 + ], + "angle": 0, + "content": "2.3 Flatness-aware Methods" + }, + { + "type": "text", + "bbox": [ + 0.081, + 0.734, + 0.918, + 0.948 + ], + "angle": 0, + "content": "Flatness-aware methods aiming at avoiding unstable sharp optimization draws rising interest under the situation where a validation set has a different distribution from the test data, leading to a non-i.i.d. scenario. Cha et al. [4] hypothesized that in the non-i.i.d scenario the generalization gap between flat and sharp minima is more pronounced, and then proposed a Stochastic Weight Averaging Densely (SWAD) method aiming to locate flat minima and reduce domain gap. Moreover, Arpit et al. [2] argued that models trained on specific training domains often display erratic performance on test domains with distribution shifts. To mitigate this, they suggested the use of ensembling moving average (EoA) models, thereby decreasing uncertainty in test domains for enhanced domain generalization. Sharpness-Aware Minimization (SAM) [11] and an improved" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.79, + 0.016, + 0.842, + 0.032 + ], + "angle": 0, + "content": "RES" + }, + { + "type": "page_number", + "bbox": [ + 0.904, + 0.017, + 0.92, + 0.031 + ], + "angle": 0, + "content": "5" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.056, + 0.921, + 0.116 + ], + "angle": 0, + "content": "Sharpness-Aware Gradient Matching (SGAM) [38] were further proposed to enhance model generalization by simultaneously minimizing the loss value and its sharpness, leading to more robust models." + }, + { + "type": "title", + "bbox": [ + 0.076, + 0.146, + 0.628, + 0.166 + ], + "angle": 0, + "content": "3 RES-based Domain Generalization" + }, + { + "type": "title", + "bbox": [ + 0.076, + 0.188, + 0.444, + 0.208 + ], + "angle": 0, + "content": "3.1 Notations and Overview" + }, + { + "type": "text", + "bbox": [ + 0.074, + 0.222, + 0.921, + 0.414 + ], + "angle": 0, + "content": "Assuming there are \\(K\\) distinct yet semantic-related source domains represented as \\(D_{s} = \\{D_{s}^{1}, D_{s}^{2}, \\ldots, D_{s}^{K}\\}\\), each originating from different distributions and comprising \\(N_{k}\\) image-label pairs, denoted as \\(D_{s}^{k} = \\{(x_{i}, y_{i})\\}_{i=1}^{N_{k}}\\). In this study, the model training is independent of domain labels, so there is no domain subscript in \\((x, y)\\). The primary objective of DG is to leverage these source domains to train a model, parameterized by \\(\\theta\\), that exhibits strong generalization capabilities and can effectively adapt to unseen target domains. In RES, there are two models sharing the same architecture, a student model \\(M^{S}\\) parameterized with \\(\\theta^{S}\\) and a teacher model \\(M^{T}\\) parameterized with \\(\\theta^{T}\\). \\(M^{S}\\) is used for model training and \\(M^{T}\\) is utilized for model validation and test." + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.417, + 0.92, + 0.513 + ], + "angle": 0, + "content": "The proposed RES-based DG framework is shown in Fig. 2 with its training process summarized in Algorithm 1. RES consists of: (1) a RE module that augments amplitude spectrum of feature maps while maintaining the phase spectrum, to expand cross-domain feature representation; and (2) a RS module that mutually fuses the parameters of the teacher and student models to each other." + }, + { + "type": "image", + "bbox": [ + 0.083, + 0.548, + 0.918, + 0.815 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.075, + 0.836, + 0.923, + 0.945 + ], + "angle": 0, + "content": "Fig. 2: Workflow of the proposed RES-based domain generalization framework, which consists of Representation Enhancement (RE) and Representation Stabilization (RS). RE aims to enhance feature representation to reduce cross-domain bias via a set of augmentations performing on the amplitude spectrum of frequency domain, while RS aims to further stabilize cross-domain performance to reduce cross-domain variance through a novel MEMA optimization strategy between student and teacher models." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.078, + 0.017, + 0.096, + 0.031 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.159, + 0.016, + 0.394, + 0.033 + ], + "angle": 0, + "content": "W. Huang, Y. Shi et al." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.055, + 0.682, + 0.075 + ], + "angle": 0, + "content": "3.2 Feature Frequency Augmentation-Based RE" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.093, + 0.921, + 0.172 + ], + "angle": 0, + "content": "Our RE modules augments source data at the feature level. During training, the student model \\( M^S \\) is split into two parts: \\( M^{S-} \\) which includes all layers up to a middle layer \\( l \\) for feature map extraction, and \\( M^{S+} \\) which consists of the remaining layers for the remaining feature extraction and classification." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.205, + 0.921, + 0.304 + ], + "angle": 0, + "content": "Decomposing Phase and Amplitude From Features. Given an image \\( x \\in \\mathcal{R}^{3 \\times \\bar{H} \\times \\bar{W}} \\) where \\( \\bar{H} \\times \\bar{W} \\) represents image spatial size, a corresponding multilayer feature map \\( f \\in \\mathcal{R}^{C \\times H \\times W} \\) can be extracted by \\( M^{S^{-}} \\), where \\( C \\) signifies the number of channels, and \\( H, W \\) denote the height and width of the feature maps respectively. It is formulated as" + }, + { + "type": "equation", + "bbox": [ + 0.429, + 0.322, + 0.915, + 0.342 + ], + "angle": 0, + "content": "\\[\nf = M ^ {S -} (x). \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.36, + 0.92, + 0.458 + ], + "angle": 0, + "content": "To delve into the frequency properties of these feature maps, we employ a 2D FFT across the spatial dimensions of each channel, ensuring that every layer of \\( f \\) undergoes this transformative process. This approach allows us to comprehensively analyze the frequency content within each individual feature channel. For channel \\( c \\), its FFT formulation is as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.268, + 0.477, + 0.915, + 0.527 + ], + "angle": 0, + "content": "\\[\nF ^ {c} (u, v) = \\sum_ {h = 0} ^ {H - 1} \\sum_ {w = 0} ^ {W - 1} f ^ {c} (h, w) \\cdot e ^ {- j 2 \\pi \\left(\\frac {u h}{H} + \\frac {v w}{W}\\right)}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.543, + 0.92, + 0.622 + ], + "angle": 0, + "content": "where \\( f^{c}(h,w) \\) represents the pixel value at the spatial location \\( (h,w) \\) within the range of \\( [H - 1,W - 1] \\) in channel \\( c \\), and \\( F^{c}(u,v) \\) denotes the complex frequency spectrum at the frequency coordinates \\( (u,v) \\) of channel \\( c \\). The term \\( j \\) is the imaginary unit." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.623, + 0.919, + 0.681 + ], + "angle": 0, + "content": "Then we compute a multi-layer phase spectrum denoted as \\(\\mathcal{P}\\) and a multi-layer amplitude spectrum represented as \\(\\mathcal{A}\\), from the complex frequency spectrum \\(F\\), as" + }, + { + "type": "equation", + "bbox": [ + 0.209, + 0.702, + 0.915, + 0.72 + ], + "angle": 0, + "content": "\\[\n\\mathcal {P} \\left(F ^ {c} (u, v)\\right) = \\arctan 2 \\left(\\operatorname {I m a g} \\left(F ^ {c} (u, v)\\right), \\operatorname {R e a l} \\left(F ^ {c} (u, v)\\right)\\right), \\tag {5}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.228, + 0.761, + 0.915, + 0.781 + ], + "angle": 0, + "content": "\\[\n\\mathcal {A} \\left(F ^ {c} (u, v)\\right) = \\sqrt {\\operatorname {R e a l} \\left(F ^ {c} (u , v)\\right) ^ {2} + \\operatorname {I m a g} \\left(F ^ {c} (u , v)\\right) ^ {2}}. \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.792, + 0.92, + 0.949 + ], + "angle": 0, + "content": "Here, \\(\\operatorname{Real}(F^c(u,v))\\) and \\(\\operatorname{Imag}(F^c(u,v))\\) indicate the real and imaginary parts of the channel- \\(c\\) frequency spectrum \\(F^c\\), respectively. \\(\\mathcal{P}\\) and \\(\\mathcal{A}\\) have the same size of \\(C \\times H \\times W\\) as \\(f\\). Drawing inspiration from [15,29,32,41], there is a common assumption that the phase component of frequency spectrum maintains the high-level semantics of the original data while the amplitude component contains its low-level statistics. In this study, we extend it from the image level to feature level, where we regard feature maps as high-dimension signals with the same characteristic of phase-amplitude decomposition as images." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.79, + 0.016, + 0.842, + 0.032 + ], + "angle": 0, + "content": "RES" + }, + { + "type": "page_number", + "bbox": [ + 0.904, + 0.017, + 0.921, + 0.031 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.057, + 0.921, + 0.116 + ], + "angle": 0, + "content": "Frequency Augmentations on Features. To avoid damaging the semantics of features, augmentations are only implemented on the amplitude \\(\\mathcal{A}\\) to obtain an augmented \\(\\mathcal{A}'\\) while the phase \\(\\mathcal{P}\\) keeps unchanged, as" + }, + { + "type": "equation", + "bbox": [ + 0.428, + 0.127, + 0.919, + 0.148 + ], + "angle": 0, + "content": "\\[\n\\mathcal {A} ^ {\\prime} = \\operatorname {A u g} (\\mathcal {A}). \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.157, + 0.922, + 0.273 + ], + "angle": 0, + "content": "Given a batch of images \\(\\{x_{1}, x_{2}, \\ldots, x_{B}\\}\\) where \\(B\\) is batch size, a corresponding batch of amplitudes \\(\\{\\mathcal{A}_1, \\mathcal{A}_2, \\ldots, \\mathcal{A}_B\\}\\) can be obtained by Eqs. (3), (4), and (6). During each training iteration, we randomly select and apply one augmentation strategy from a pre-defined augmentation list, including none, random noise, random dropout, and mixup. Here none represents no augmentation, and the details of the rest three strategies are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.281, + 0.924, + 0.381 + ], + "angle": 0, + "content": "- Random noise, denoted as \\(\\mathrm{RE}^{ns}\\). Given a multi-layer amplitude \\(\\mathcal{A}_i \\in \\mathcal{R}^{C \\times H \\times W}\\), we add Gaussian noise \\(G \\in \\mathcal{R}^{1 \\times H \\times W}\\) to it. The elements of \\(G\\) adhere to a Gaussian distribution characterized by mean \\(\\mu\\) and standard deviation \\(\\sigma\\), i.e., \\(G \\sim \\mathcal{N}(\\mu, \\sigma^2)\\). Leveraging the broadcasting mechanism, this noise addition is equally conducted to each layer of \\(\\mathcal{A}_i\\), as:" + }, + { + "type": "equation", + "bbox": [ + 0.303, + 0.389, + 0.919, + 0.412 + ], + "angle": 0, + "content": "\\[\n\\mathcal {A} _ {i} ^ {\\prime} = \\mathcal {A} _ {i} * (I + G) = \\mathcal {A} _ {i} * (I + \\mathcal {N} (\\mu , \\sigma^ {2})), \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.115, + 0.421, + 0.92, + 0.478 + ], + "angle": 0, + "content": "where \\( I \\) represents the Identity Matrix. This approach guarantees that the same frequency components throughout all channels of the feature map are enhanced with the same noise intensity." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.478, + 0.924, + 0.577 + ], + "angle": 0, + "content": "- Random dropout, denoted as \\(\\mathrm{RE}^{dp}\\). We randomly generate a binary dropout mask, \\(M \\in \\mathcal{R}^{1 \\times H \\times W}\\), where each element is set either to 0 with probability \\(p\\) or to 1 with a reversal probability \\(1 - p\\). The dimension of \\(\\mathcal{M}\\) is also expanded to \\(C \\times H \\times W\\) by broadcasting. The dropout-augmented amplitude \\(\\mathcal{A}'\\) is calculated by:" + }, + { + "type": "equation", + "bbox": [ + 0.455, + 0.586, + 0.919, + 0.607 + ], + "angle": 0, + "content": "\\[\n\\mathcal {A} _ {i} ^ {\\prime} = \\mathcal {A} \\cdot M. \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.115, + 0.617, + 0.924, + 0.695 + ], + "angle": 0, + "content": "This operation introduces sparsity of amplitudes, aiming at reducing their overfitting for better robustness. It is worth mentioning that the element of \\( M \\) at the location \\( [0,0] \\) always equals 1, which represents the mean value of the feature map, to avoid over-masking on features." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.694, + 0.931, + 0.832 + ], + "angle": 0, + "content": "- Mixup, denoted as \\(\\mathrm{RE}^{mix}\\). Unlike random noise and random dropout which operate on single amplitudes extracted from individual samples, mixup engages in a cross-sample amplitude interaction by linear interpolation on different amplitudes extracted from different samples. Specifically, for a given amplitude spectrum \\(\\mathcal{A}_i\\), mixup randomly selects another amplitude \\(\\mathcal{A}_j\\) from the set \\(\\{\\mathcal{A}_1, \\mathcal{A}_2, \\dots, \\mathcal{A}_B\\}\\), and then blends \\(\\mathcal{A}_i\\) with \\(\\mathcal{A}_j\\), resulting in a new, hybrid amplitude spectrum \\(\\mathcal{A}_i'\\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.398, + 0.842, + 0.919, + 0.863 + ], + "angle": 0, + "content": "\\[\n\\mathcal {A} _ {i} ^ {\\prime} = \\alpha \\mathcal {A} _ {i} + (1 - \\alpha) \\mathcal {A} _ {j}, \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.115, + 0.871, + 0.924, + 0.95 + ], + "angle": 0, + "content": "where \\(\\alpha\\) is a uniform random coefficient between [0,1]. This procedure enriches the amplitude characteristics by promoting a more diverse fusion of feature styles across domains. It boosts the model's generalization by augmenting intricate cross-domain amplitude patterns effectively." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.079, + 0.017, + 0.096, + 0.03 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.159, + 0.016, + 0.394, + 0.033 + ], + "angle": 0, + "content": "W. Huang, Y. Shi et al." + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.057, + 0.921, + 0.116 + ], + "angle": 0, + "content": "Reconstructing Features. For each pair of phase and augmented amplitude components, \\(\\mathcal{P}\\) and \\(\\mathcal{A}'\\), an augmented multi-layer feature map, \\(f'\\), can be reconstructed via inverse Fast Fourier Transform (iFFT) as" + }, + { + "type": "equation", + "bbox": [ + 0.135, + 0.149, + 0.919, + 0.2 + ], + "angle": 0, + "content": "\\[\nf ^ {\\prime c} (h, w) = \\frac {1}{H W} \\sum_ {u = 0} ^ {H - 1} \\sum_ {v = 0} ^ {W - 1} \\mathcal {A} ^ {\\prime} \\left(F ^ {c} (u, v)\\right) \\cdot e ^ {j 2 \\pi \\left(\\frac {u h}{H} + \\frac {v w}{W}\\right)} \\cdot e ^ {j \\mathcal {P} \\left(F ^ {c} (u, v)\\right)} \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.218, + 0.92, + 0.279 + ], + "angle": 0, + "content": "Here, \\( e^{j2\\pi \\left(\\frac{uh}{H} + \\frac{vw}{W}\\right)} \\) is a complex exponential term representing the frequency component in the inverse Fourier transform, and \\( e^{j\\mathcal{P}(F^c(u,v))} \\) reintroduces the phase information into the transform." + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.28, + 0.919, + 0.319 + ], + "angle": 0, + "content": "Finally, \\( f' \\) is used for further feature extraction and class classification by the remaining layers of the student model, \\( M^{S+} \\), as" + }, + { + "type": "equation", + "bbox": [ + 0.425, + 0.335, + 0.918, + 0.358 + ], + "angle": 0, + "content": "\\[\np = M ^ {S +} \\left(f ^ {\\prime}\\right). \\tag {12}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.076, + 0.387, + 0.361, + 0.405 + ], + "angle": 0, + "content": "3.3 MEMA-based RS" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.422, + 0.92, + 0.519 + ], + "angle": 0, + "content": "As shown in Fig. 2, there are two models, one student model \\( M^S \\) with parameters \\( \\theta^S \\) and one teacher model \\( M^T \\) with parameters \\( \\theta^T \\). Their parameters undergo mutual updates to provide stabilization to each other, referred to as \\( \\mathrm{RE}^T \\) and \\( \\mathrm{RE}^S \\) for the teacher and student models, respectively. Their processes are formulated as" + }, + { + "type": "equation", + "bbox": [ + 0.292, + 0.532, + 0.915, + 0.58 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\theta^ {T} \\leftarrow \\tau^ {T} \\theta^ {T} + (1 - \\tau^ {T}) \\theta^ {S}, \\quad \\tau^ {T} \\in [ 0, 1), \\tag {13} \\\\ \\theta^ {S} \\leftarrow (1 - \\tau^ {S}) \\theta^ {S} + \\tau^ {S} \\theta^ {T}, \\quad \\tau^ {S} \\in [ 0, 1). \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.598, + 0.92, + 0.715 + ], + "angle": 0, + "content": "Here, only the student model \\( M^S \\) is actively trained, while the teacher model \\( M^T \\) is responsible solely for parameter updates. The student model \\( M^S \\) is the primary learner, actively absorbing new knowledge and adapting to training data. In contrast, the teacher model \\( M^T \\) acts as a robust and stable reference using moving average updating strategy, providing more robust and consistent parameters by reducing the variance of training batches [2]." + }, + { + "type": "text", + "bbox": [ + 0.074, + 0.715, + 0.92, + 0.87 + ], + "angle": 0, + "content": "Both the values of \\(\\tau^T\\) and \\(\\tau^S\\) are small and they play a mutual role in this process: (1) Stability from the teacher model. With a small \\(\\tau^T\\), the teacher model's parameters change slowly, ensuring that they remain stable and consistent.(2) Gradual integration for the student model. A small \\(\\tau^S\\) implies that the student model integrates the teacher's parameters slowly and steadily. This allows the student model to gradually assimilate the stable characteristics of the teacher model. To simplify the hyperparameter tuning process in experiments, \\(\\tau^T\\) is set to a small constant value of 1e-3, while \\(\\tau^S\\) is variable for different datasets." + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.872, + 0.92, + 0.949 + ], + "angle": 0, + "content": "In summary, the student model acts as an active learner, continually evolving and assimilating new information, whereas the teacher model functions as a reliable guide, providing a stable parameter reference to shield the student model from the effects of cross-domain variance in training batches." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.79, + 0.016, + 0.842, + 0.032 + ], + "angle": 0, + "content": "RES" + }, + { + "type": "page_number", + "bbox": [ + 0.903, + 0.017, + 0.92, + 0.031 + ], + "angle": 0, + "content": "9" + }, + { + "type": "table", + "bbox": [ + 0.078, + 0.055, + 0.92, + 0.355 + ], + "angle": 0, + "content": "
Algorithm 1 Training Process of our RES-based DG
1:Student model MS with initialized parameters θS, teacher model MT with initialized parameters θT, all training sample pairs {(xi,yi)}N, and batch size B
2:for iter = 1 to Niter do
3:Sample Batch Data: {(xi,yi)}B
4:Extract Feature: f← MS-(x)
5:Decompose Phase and Amplitude: P, A← FFT(f)
6:Augmentation Selection: Aug ← [random noise, random dropout, mixup, none]
7:Augment Amplitude: A' ← Aug(A)
8:Reconstruct feature: f' = iFFT(P, A')
9:Compute prediction: p ← MS+(f')
10:Compute Loss: E(θS) ← l(p, y)
11:Update Student Model with Loss: θS ← θS - η∇θL(θS)
12:Integrate Student Model to Teacher Model: θT = τTθT + (1 - τT)θS
13:Stabilize Student Model with Teacher Model: θS = (1 - τS)θS + τSθT
14:end for
" + }, + { + "type": "title", + "bbox": [ + 0.076, + 0.391, + 0.351, + 0.41 + ], + "angle": 0, + "content": "3.4 Supervision Loss" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.423, + 0.923, + 0.482 + ], + "angle": 0, + "content": "During each training iteration with a batch size of \\( B \\), the empirical risk minimization (ERM) loss function is used to optimize the student model \\( M^S \\) across all the training domains as" + }, + { + "type": "equation", + "bbox": [ + 0.231, + 0.496, + 0.92, + 0.55 + ], + "angle": 0, + "content": "\\[\n\\mathcal {E} \\left(\\theta^ {S}\\right) = \\frac {1}{B} \\sum_ {i = 1} ^ {B} \\ell \\left(p _ {i}, y _ {i}\\right) = \\frac {1}{B} \\sum_ {i = 1} ^ {B} \\ell \\left(M ^ {S} \\left(x _ {i}; \\theta^ {S}\\right), y _ {i}\\right), \\tag {14}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.563, + 0.92, + 0.605 + ], + "angle": 0, + "content": "where the detailed progress of \\( M^S (x_i;\\theta^S) \\) come from Eqs. (3)-(7) and (12) and \\( \\ell \\) represents the CE loss. Here \\( \\mathcal{E}(\\theta^S) \\) represents the basic ERM loss." + }, + { + "type": "title", + "bbox": [ + 0.076, + 0.634, + 0.314, + 0.656 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.674, + 0.922, + 0.831 + ], + "angle": 0, + "content": "This section outlines benchmark datasets and experimental settings, conducts an ablation study on RES, compares it with advanced DG methods, examines the impact of \\(\\tau^S\\), and quantitatively assesses RES's effects in reducing cross-domain bias and variance and improving accuracy across domains. In Supplementary Material, we compare the effects of data augmentation from two perspectives: image-level vs. feature-level frequency augmentation, and normal feature augmentation vs. feature frequency augmentation. This comparison demonstrates the superiority of our feature frequency augmentation." + }, + { + "type": "title", + "bbox": [ + 0.076, + 0.859, + 0.399, + 0.877 + ], + "angle": 0, + "content": "4.1 Benchmark Datasets" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.891, + 0.922, + 0.951 + ], + "angle": 0, + "content": "We evaluate the proposed RES method on five DG datasets: (1) PACS [22], which contains 9,991 images in 7 object categories across 4 diverse domains; (2) VLCS [10], which comprises 10,729 examples from 5 categories, collected from 4" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.08, + 0.017, + 0.106, + 0.032 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.159, + 0.015, + 0.394, + 0.033 + ], + "angle": 0, + "content": "W. Huang, Y. Shi et al." + }, + { + "type": "text", + "bbox": [ + 0.079, + 0.055, + 0.92, + 0.134 + ], + "angle": 0, + "content": "domains; (3) OfficeHome [36], which encompasses around 15,500 samples in 65 categories from four domains. (4) TerraIncognita [3], which contains 24,788 images within 10 categories from four domains; and (5) DomainNet [30], a vast dataset with 586,575 images across 345 categories, spanning 6 domains." + }, + { + "type": "title", + "bbox": [ + 0.08, + 0.162, + 0.416, + 0.182 + ], + "angle": 0, + "content": "4.2 Experimental Settings" + }, + { + "type": "text", + "bbox": [ + 0.08, + 0.193, + 0.92, + 0.35 + ], + "angle": 0, + "content": "Basic settings. We utilize ResNet-50 [16], pretrained on ImageNet-1K [9], as the backbone. Our codes are constructed on Domainbed's framework [13], allocating \\(80\\%\\) of in-domain data for training and the remaining \\(20\\%\\) for validation, with test performed on an unseen domain. PACS, VLCS, and OfficeHome are trained for 10K iterations and validated every 300 iterations, TerraIncognita is trained for 3K iterations and validated every 300 iterations, and DomainNet undergoes 400K iterations of training and 1K iterations of validation. Adam is used as the optimizer, with a batch size of 128 and a learning rate of 5e-5." + }, + { + "type": "text", + "bbox": [ + 0.08, + 0.377, + 0.92, + 0.534 + ], + "angle": 0, + "content": "Settings of RES. In every training iteration, we apply the frequency augmentation of RE randomly to one of the four potential middle-layer positions within ResNet-50. These positions are strategically located at the ends of the \\(1^{st}\\), \\(2^{nd}\\), \\(3^{rd}\\), and \\(4^{th}\\) residual blocks, ensuring diverse augmentation effects across different layers of the model. There are some hyperparameters in RES summarized in Table 1, where \\(\\tau^S\\) is variable and tuned based on the validation performance as discussed in Section 4.4. All the experiments of RES are conducted for three times and the mean values and standard deviations are reported." + }, + { + "type": "table_caption", + "bbox": [ + 0.329, + 0.567, + 0.669, + 0.586 + ], + "angle": 0, + "content": "Table 1: Hyperparameters of RES." + }, + { + "type": "table", + "bbox": [ + 0.104, + 0.602, + 0.887, + 0.642 + ], + "angle": 0, + "content": "
REnsREdpREmixRSS→TRST→S
G ~ N(0.75, 0.752)p = 0.5α ~ U(0,1)τT = 0.001τS is variable
" + }, + { + "type": "title", + "bbox": [ + 0.08, + 0.721, + 0.331, + 0.743 + ], + "angle": 0, + "content": "4.3 Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.08, + 0.753, + 0.919, + 0.851 + ], + "angle": 0, + "content": "The ablation study of RES's modules, including sub-strategies of RE and RS, on PACS are provided in Table 2 and Fig. 3. RE employs random noise \\((\\mathrm{RE}^{ns})\\), random dropout \\((\\mathrm{RE}^{dp})\\), and mixup \\((\\mathrm{RE}^{mix})\\) techniques for feature augmentation, and RS employs MEMA containing representation stabilization for both student models \\((\\mathrm{RS}^s)\\) and teacher models \\((\\mathrm{RS}^t)\\)." + }, + { + "type": "text", + "bbox": [ + 0.08, + 0.852, + 0.92, + 0.949 + ], + "angle": 0, + "content": "After applying all three RE strategies, there is significant performance improvement from ERM's 85.5 to (ERM+RE)'s 88.8 as shown in Table 2. with their training trajectory shown in Fig. 3. The results verify the effect of RE in reducing cross-domain bias. Additionally, the application of RS strategies reduces their standard deviations from ERM's 1.1 to (ERM+RS)'s 0.5, demonstrating their" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.79, + 0.016, + 0.842, + 0.032 + ], + "angle": 0, + "content": "RES" + }, + { + "type": "page_number", + "bbox": [ + 0.894, + 0.017, + 0.919, + 0.031 + ], + "angle": 0, + "content": "11" + }, + { + "type": "text", + "bbox": [ + 0.074, + 0.057, + 0.921, + 0.135 + ], + "angle": 0, + "content": "effectiveness in reducing cross-domain variance. When RE and RS strategies are applied in conjunction, as shown in the last highlighted row of the table, the model achieves optimal test performance, with an average accuracy increase to 90.0 with the smallest standard deviation of 0.3, significantly surpassing ERM." + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.137, + 0.92, + 0.196 + ], + "angle": 0, + "content": "Through this ablation study, the complementary effects of RE and RS strategies in enhancing model generalization across unseen domains are clearly demonstrated, underscoring the significance of reducing cross-domain bias-variance." + }, + { + "type": "table_caption", + "bbox": [ + 0.217, + 0.235, + 0.78, + 0.253 + ], + "angle": 0, + "content": "Table 2: Ablation study of the proposed RES on PACS." + }, + { + "type": "table", + "bbox": [ + 0.079, + 0.282, + 0.921, + 0.465 + ], + "angle": 0, + "content": "
ERMREnsREdpREmixRStRSsPACSAvg
ArtClipartPaintingSketch
✓84.4±0.681.1±0.796.3±0.580.2±2.585.5±1.1
✓✓88.8±1.282.2±0.597.0±0.681.2±0.687.3±0.7
✓✓87.6±0.382.7±1.096.8±0.882.1±1.687.3±0.9
✓✓89.4±0.281.4±1.196.6±0.485.1±0.988.5±0.7
✓✓✓✓89.0±0.983.0±0.497.0±0.485.1±0.788.8±0.6
✓✓87.9±0.183.1±0.296.0±0.482.2±1.287.3±0.5
✓✓✓90.9±0.482.0±0.597.9±0.382.2±0.888.2±0.5
✓✓✓✓✓✓91.6±0.484.2±0.398.1±0.186.1±0.490.0±0.3
" + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.557, + 0.291, + 0.668 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.293, + 0.557, + 0.501, + 0.668 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.557, + 0.709, + 0.668 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.71, + 0.557, + 0.917, + 0.668 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.212, + 0.683, + 0.783, + 0.702 + ], + "angle": 0, + "content": "Fig. 3: Training convergence of RES's modules on PACS." + }, + { + "type": "title", + "bbox": [ + 0.076, + 0.785, + 0.431, + 0.806 + ], + "angle": 0, + "content": "4.4 Values of \\(\\tau^S\\) in MEMA" + }, + { + "type": "text", + "bbox": [ + 0.074, + 0.832, + 0.922, + 0.95 + ], + "angle": 0, + "content": "In RES, the MEMA's \\(\\tau^S\\) is the sole hyperparameter requiring adjustment across various benchmarks. We explored five distinct values: [0, 1e-3, 5e-3, 1e-2, 5e-2] for each benchmark, with results shown in Fig. 4. The results indicate that the optimal \\(\\tau^S\\) value varies by dataset, attributed to differences in their convergence speed. Subsequent comparison experiments will report out-of-domain accuracies for \\(\\tau^S\\) corresponding to the highest in-domain accuracies observed." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.08, + 0.017, + 0.105, + 0.031 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.159, + 0.016, + 0.396, + 0.034 + ], + "angle": 0, + "content": "W. Huang, Y. Shi et al." + }, + { + "type": "image", + "bbox": [ + 0.08, + 0.053, + 0.245, + 0.16 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.247, + 0.053, + 0.412, + 0.16 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.414, + 0.053, + 0.577, + 0.16 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.58, + 0.053, + 0.744, + 0.16 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.746, + 0.053, + 0.918, + 0.16 + ], + "angle": 0, + "content": null + }, + { + "type": "image_footnote", + "bbox": [ + 0.376, + 0.163, + 0.653, + 0.173 + ], + "angle": 0, + "content": "In-domain Accuracy Out-of-domain Accuracy" + }, + { + "type": "image_caption", + "bbox": [ + 0.1, + 0.187, + 0.895, + 0.208 + ], + "angle": 0, + "content": "Fig. 4: In-domain (validation) and out-of-domain (test) accuracies vary with \\(\\tau^S\\)." + }, + { + "type": "title", + "bbox": [ + 0.076, + 0.245, + 0.694, + 0.266 + ], + "angle": 0, + "content": "4.5 In Comparison With Advanced DG Methods" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.277, + 0.923, + 0.491 + ], + "angle": 0, + "content": "To objectively assess the effectiveness of the proposed RES, we conducted comparative analyses against several state-of-the-art DG methods. This comparison spans the baseline ERM, various data augmentation strategies (e.g., Mixup [40], CCFP [21], DomainDrop [14]), flatness-aware methods (such as SWAD [4], SMA [2], SAGM [38], FAD [45]), and some other DG approaches (including FISH [35], RSC [18], DAC-SC [20], GVRT [27], MIRO [5]). These evaluations were performed across five benchmarks utilizing the ResNet-50 backbone, with results detailed in Table 3. The domain-wise performance of the datasets of our RES are provided in Supplementary Material. Notably, RES incorporates both data augmentation and flatness-aware strategies, positioning it uniquely to tackle the dual challenges of cross-domain bias and variance simultaneously." + }, + { + "type": "table_caption", + "bbox": [ + 0.101, + 0.525, + 0.894, + 0.544 + ], + "angle": 0, + "content": "Table 3: Performance comparisons between some SOTA methods and our RES." + }, + { + "type": "table", + "bbox": [ + 0.079, + 0.572, + 0.92, + 0.82 + ], + "angle": 0, + "content": "
MethodPACSVLCSOfficeHomeTerraIncDomainNetAvg
ERM [ICLR'20] [13]85.7±0.577.4±0.367.5±0.547.2±0.441.2±0.263.8
RSC [ECCV'20] [18]85.2±0.977.1±0.565.5±0.946.6±1.038.9±0.562.7
Mixup [AAAI'20] [40]84.3±0.577.7±0.469.0±0.148.9±0.839.6±0.163.9
FISH [ICLR'21] [35]85.5±0.377.8±0.368.6±0.445.1±1.342.7±0.263.9
GVRT [ECCV'22] [27]85.1±0.379.0±0.270.1±0.148.0±0.244.1±0.165.2
MIRO [ECCV'22] [5]85.4±0.479.0±0.070.5±0.450.4±1.144.3±0.265.9
SMA [NeurIPS'22] [2]87.5±0.278.2±0.270.6±0.150.3±0.546.0±0.166.5
SWAD [NeurIPS'21] [4]88.1±0.179.1±0.170.6±0.250.0±0.346.5±0.166.9
CCFP [ICCV'23] [21]86.6±0.278.9±0.368.9±0.148.6±0.441.2±0.164.8
DAC-SC [CVPR'23] [20]87.5±0.178.7±0.370.3±0.246.5±0.344.9±0.165.6
FAD [ICCV'23] [45]88.2±0.278.9±0.869.2±0.545.7±1.044.4±0.165.3
SAGM [CVPR'23] [38]86.6±0.280.0±0.370.1±0.248.8±0.945.0±0.266.1
DomainDrop [ICCV'23] [14]87.9±0.379.8±0.368.7±0.151.5±0.444.4±0.566.5
Our RES90.0±0.379.8±0.271.8±0.351.4±0.646.7±0.267.9
" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.872, + 0.924, + 0.951 + ], + "angle": 0, + "content": "The table showcases the RES algorithm's performance across five benchmarks, with \"Avg\" representing the average performance across them. This provides a comprehensive view of each method's generalizability and effectiveness in DG. The proposed RES demonstrates superior performance, particularly high-" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.79, + 0.016, + 0.841, + 0.032 + ], + "angle": 0, + "content": "RES" + }, + { + "type": "page_number", + "bbox": [ + 0.894, + 0.017, + 0.92, + 0.031 + ], + "angle": 0, + "content": "13" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.056, + 0.921, + 0.117 + ], + "angle": 0, + "content": "lighted by its leading average score, indicating its robustness and adaptability across a diverse set of domains. The comparative analysis confirms the superiority and effectiveness of the RES in mitigating cross-domain bias and variance." + }, + { + "type": "title", + "bbox": [ + 0.076, + 0.145, + 0.651, + 0.166 + ], + "angle": 0, + "content": "4.6 Effects of RES on Domain Generalization" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.178, + 0.921, + 0.239 + ], + "angle": 0, + "content": "This subsection quantitatively describes the advantages of RES on DG from three perspectives, including reducing cross-domain bias, reducing cross-domain variance, and improving both in-domain and out-of-domain performance." + }, + { + "type": "title", + "bbox": [ + 0.075, + 0.267, + 0.479, + 0.286 + ], + "angle": 0, + "content": "Reducing Cross-domain Bias." + }, + { + "type": "text", + "bbox": [ + 0.074, + 0.287, + 0.479, + 0.56 + ], + "angle": 0, + "content": "First, we measure the cosine similarity between the mean feature embeddings obtained from the global average pooling layer of ResNet-50 across both source and target datasets within PACS and OfficeHome. This assessment encompasses comparisons between the baseline ERM and our RES, as illustrated in Fig. 5. Our findings indicate that RES significantly enhances the cosine similarity between source and target data compared to ERM, demonstrating its capability in diminishing cross-domain l" + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.306, + 0.707, + 0.407 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.541, + 0.409, + 0.7, + 0.419 + ], + "angle": 0, + "content": "Source-target Similarity of ERM" + }, + { + "type": "image", + "bbox": [ + 0.71, + 0.307, + 0.908, + 0.407 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.717, + 0.409, + 0.875, + 0.419 + ], + "angle": 0, + "content": "Source-target Similarity of RES" + }, + { + "type": "image_caption", + "bbox": [ + 0.494, + 0.433, + 0.922, + 0.504 + ], + "angle": 0, + "content": "Fig. 5: Cosine similarity between mean features of source domains and target domain. The higher the similarity is, the lower the cross-domain bias is." + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.589, + 0.922, + 0.786 + ], + "angle": 0, + "content": "Reducing Cross-domain Variance. Secondly, we present the training convergence trajectories alongside the corresponding cumulative instability for four DG scenarios within the PACS dataset, as illustrated in Fig. 6. It is observable that, in comparison to the baseline method ERM represented by the blue color, the proposed RES depicted in red markedly reduces the fluctuations in both in-domain and out-of-domain accuracies with almost no instability increase after the initial increase stage during training, which spans approximately the first 2000 iterations. Benefiting from the stable and narrower performance gap between in-domain (validation) and out-of-domain (test) sets, the best models chosen based on validation sets demonstrate consistent performance on test sets." + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.832, + 0.922, + 0.951 + ], + "angle": 0, + "content": "Improving Both In-domain and Out-of-domain Performance. As shown in Table 4, the comparison between ERM and RES on PACS illustrates significant enhancements in both in-domain and out-of-domain accuracies achieved by RES. While ERM delivers reliable in-domain performance, it falls short in adapting to unseen domains. In contrast, RES not only boosts in-domain accuracy from ERM's 96.5 to its 98.0 but also lifts the out-of-domain accuracy" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.08, + 0.017, + 0.105, + 0.031 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.159, + 0.016, + 0.396, + 0.034 + ], + "angle": 0, + "content": "W. Huang, Y. Shi et al." + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.054, + 0.289, + 0.158 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.291, + 0.054, + 0.496, + 0.157 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.498, + 0.054, + 0.707, + 0.157 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.71, + 0.054, + 0.917, + 0.157 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.077, + 0.162, + 0.288, + 0.288 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.291, + 0.162, + 0.498, + 0.287 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.162, + 0.709, + 0.287 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.713, + 0.162, + 0.92, + 0.287 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.075, + 0.302, + 0.921, + 0.374 + ], + "angle": 0, + "content": "Fig. 6: Training stability of ERM and RES on PACS, evaluated by the absolute first derivative of accuracy every 300 iterations \\((|f'(Acc)| = |Acc_t - Acc_{t-1}|)\\), capturing accuracy change rate. Accumulative instability aggregates \\(|f'(Acc)|\\) up to iteration \\(t\\), with lower values indicating more stable performance and less variance." + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.412, + 0.918, + 0.452 + ], + "angle": 0, + "content": "from \\(85.5 \\pm 1.1\\) to \\(90.0 \\pm 0.3\\). This robust improvement underscores RES's robust capability to refine model performance across various distributions." + }, + { + "type": "table_caption", + "bbox": [ + 0.08, + 0.484, + 0.914, + 0.503 + ], + "angle": 0, + "content": "Table 4: In-domain and out-of-domain accuracies of PACS between ERM and RES." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.519, + 0.917, + 0.634 + ], + "angle": 0, + "content": "
SplitMethodPACSAvg
ArtClipartPaintingSketch
In-domain (Validation Set)ERM97.2±0.097.0±0.495.4±0.296.3±0.496.5±0.3
Our RES98.4±0.298.0±0.097.4±0.398.2±0.398.0±0.2
Out-of-domain (Test Set)ERM84.4±0.681.1±0.796.3±0.580.2±2.585.5±1.1
Our RES91.6±0.484.2±0.398.1±0.186.1±0.490.0±0.3
" + }, + { + "type": "title", + "bbox": [ + 0.076, + 0.715, + 0.289, + 0.736 + ], + "angle": 0, + "content": "5 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.074, + 0.755, + 0.92, + 0.949 + ], + "angle": 0, + "content": "In this study, we explored DG through bias-variance decomposition, an analytical framework that offers deeper insights into the challenges and solutions for cross-domain generalization in deep learning models. Our approach quantified the detrimental effects of cross-domain bias and variance on model performance and therefore introduced the Representation Enhancement-Stabilization (RES) method as a potent solution. The RES method effectively minimizes the cross-domain bias and variance, enhancing model adaptability and generalization across unseen target domains. Empirical evaluations on five DG benchmarks reveals that the proposed RES markedly outperforms multiple advanced DG methods in improving model generalization." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.792, + 0.017, + 0.838, + 0.03 + ], + "angle": 0, + "content": "RES" + }, + { + "type": "page_number", + "bbox": [ + 0.895, + 0.018, + 0.917, + 0.03 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.083, + 0.054, + 0.35, + 0.075 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.081, + 0.095, + 0.919, + 0.328 + ], + "angle": 0, + "content": "The work was supported by the German Research Foundation (DFG GZ: ZH 498/18-1; Project number: 519016653). The work of Z. Xiong was supported by the German Federal Ministry for the Environment, Nature Conservation, Nuclear Safety and Consumer Protection (BMUV) based on a resolution of the German Bundestag (grant number: 67KI32002B; Acronym: \\(EKAPEx\\)). The work of X. Zhu was also supported by the German Federal Ministry of Education and Research (BMBF) in the framework of the international future AI lab \"AI4EO - Artificial Intelligence for Earth Observation: Reasoning, Uncertainties, Ethics and Beyond\" (grant number: 01DD20001), by German Federal Ministry for Economic Affairs and Climate Action in the framework of the \"national center of excellence ML4Earth\" (grant number: 50EE2201C) and by the Munich Center for Machine Learning." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.08, + 0.017, + 0.106, + 0.031 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.159, + 0.016, + 0.394, + 0.033 + ], + "angle": 0, + "content": "W. Huang, Y. Shi et al." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.054, + 0.238, + 0.074 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.089, + 0.098, + 0.92, + 0.152 + ], + "angle": 0, + "content": "1. Albuquerque, I., Naik, N., Li, J., Keskar, N., Socher, R.: Improving out-of-distribution generalization via multi-task self-supervised pretraining. arXiv preprint arXiv:2003.13525 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.153, + 0.92, + 0.207 + ], + "angle": 0, + "content": "2. Arpit, D., Wang, H., Zhou, Y., Xiong, C.: Ensemble of averages: Improving model selection and boosting performance in domain generalization. Advances in Neural Information Processing Systems 35, 8265-8277 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.089, + 0.208, + 0.918, + 0.242 + ], + "angle": 0, + "content": "3. Beery, S., Van Horn, G., Perona, P.: Recognition in terra incognita. In: Proceedings of the European conference on computer vision (ECCV). pp. 456-473 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.09, + 0.244, + 0.921, + 0.297 + ], + "angle": 0, + "content": "4. Cha, J., Chun, S., Lee, K., Cho, H.C., Park, S., Lee, Y., Park, S.: Swad: Domain generalization by seeking flat minima. Advances in Neural Information Processing Systems 34, 22405-22418 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.091, + 0.298, + 0.92, + 0.352 + ], + "angle": 0, + "content": "5. Cha, J., Lee, K., Park, S., Chun, S.: Domain generalization by mutual-information regularization with pre-trained models. In: European Conference on Computer Vision. pp. 440-457. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.09, + 0.353, + 0.918, + 0.405 + ], + "angle": 0, + "content": "6. Chen, C., Li, J., Han, X., Liu, X., Yu, Y.: Compound domain generalization via meta-knowledge encoding. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7119-7129 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.091, + 0.407, + 0.918, + 0.46 + ], + "angle": 0, + "content": "7. Chen, J., Gao, Z., Wu, X., Luo, J.: Meta-causal learning for single domain generalization. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7683-7692 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.09, + 0.461, + 0.92, + 0.532 + ], + "angle": 0, + "content": "8. Cubuk, E.D., Zoph, B., Shlens, J., Le, Q.V.: Randaugment: Practical automated data augmentation with a reduced search space. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops. pp. 702-703 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.09, + 0.533, + 0.918, + 0.587 + ], + "angle": 0, + "content": "9. Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: Imagenet: A large-scale hierarchical image database. In: 2009 IEEE conference on computer vision and pattern recognition. pp. 248-255. IEEE (2009)" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.588, + 0.92, + 0.641 + ], + "angle": 0, + "content": "10. Fang, C., Xu, Y., Rockmore, D.N.: Unbiased metric learning: On the utilization of multiple datasets and web images for softening bias. In: Proceedings of the IEEE International Conference on Computer Vision. pp. 1657-1664 (2013)" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.642, + 0.918, + 0.678 + ], + "angle": 0, + "content": "11. Foret, P., Kleiner, A., Mobahi, H., Neyshabur, B.: Sharpness-aware minimization for efficiently improving generalization. arXiv preprint arXiv:2010.01412 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.679, + 0.918, + 0.714 + ], + "angle": 0, + "content": "12. Geman, S., Bienenstock, E., Doursat, R.: Neural networks and the bias/variance dilemma. Neural computation 4(1), 1-58 (1992)" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.715, + 0.918, + 0.75 + ], + "angle": 0, + "content": "13. Gulrajani, I., Lopez-Paz, D.: In search of lost domain generalization. In: International Conference on Learning Representations (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.751, + 0.918, + 0.804 + ], + "angle": 0, + "content": "14. Guo, J., Qi, L., Shi, Y.: Domaindrop: Suppressing domain-sensitive channels for domain generalization. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 19114-19124 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.805, + 0.918, + 0.841 + ], + "angle": 0, + "content": "15. Hansen, B.C., Hess, R.F.: Structural sparseness and spatial phase alignment in natural scenes. JOSA A 24(7), 1873-1885 (2007)" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.842, + 0.918, + 0.895 + ], + "angle": 0, + "content": "16. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 770-778 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.896, + 0.918, + 0.949 + ], + "angle": 0, + "content": "17. Huang, X., Belongie, S.: Arbitrary style transfer in real-time with adaptive instance normalization. In: Proceedings of the IEEE international conference on computer vision. pp. 1501-1510 (2017)" + }, + { + "type": "list", + "bbox": [ + 0.081, + 0.098, + 0.921, + 0.949 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.79, + 0.016, + 0.842, + 0.032 + ], + "angle": 0, + "content": "RES" + }, + { + "type": "page_number", + "bbox": [ + 0.894, + 0.017, + 0.92, + 0.031 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.057, + 0.92, + 0.129 + ], + "angle": 0, + "content": "18. Huang, Z., Wang, H., Xing, E.P., Huang, D.: Self-challenging improves cross-domain generalization. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part II 16. pp. 124-140. Springer (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.13, + 0.92, + 0.183 + ], + "angle": 0, + "content": "19. Kim, D., Yoo, Y., Park, S., Kim, J., Lee, J.: Selfreg: Self-supervised contrastive regularization for domain generalization. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 9619-9628 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.184, + 0.921, + 0.254 + ], + "angle": 0, + "content": "20. Lee, S., Bae, J., Kim, H.Y.: Decompose, adjust, compose: Effective normalization by playing with frequency for domain generalization. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11776-11785 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.255, + 0.918, + 0.307 + ], + "angle": 0, + "content": "21. Li, C., Zhang, D., Huang, W., Zhang, J.: Cross contrasting feature perturbation for domain generalization. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 1327-1337 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.308, + 0.918, + 0.361 + ], + "angle": 0, + "content": "22. Li, D., Yang, Y., Song, Y.Z., Hospedales, T.M.: Deeper, broader and artier domain generalization. In: Proceedings of the IEEE international conference on computer vision. pp. 5542-5550 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.362, + 0.918, + 0.414 + ], + "angle": 0, + "content": "23. Li, H., Pan, S.J., Wang, S., Kot, A.C.: Domain generalization with adversarial feature learning. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 5400-5409 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.415, + 0.918, + 0.468 + ], + "angle": 0, + "content": "24. Li, P., Li, D., Li, W., Gong, S., Fu, Y., Hospedales, T.M.: A simple feature augmentation for domain generalization. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 8886-8895 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.469, + 0.918, + 0.539 + ], + "angle": 0, + "content": "25. Lv, F., Liang, J., Li, S., Zang, B., Liu, C.H., Wang, Z., Liu, D.: Causality inspired representation learning for domain generalization. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8046-8056 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.54, + 0.917, + 0.575 + ], + "angle": 0, + "content": "26. Mahajan, D., Tople, S., Sharma, A.: Domain generalization using causal matching. In: International Conference on Machine Learning. pp. 7313-7324. PMLR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.576, + 0.918, + 0.628 + ], + "angle": 0, + "content": "27. Min, S., Park, N., Kim, S., Park, S., Kim, J.: Grounding visual representations with texts for domain generalization. In: European Conference on Computer Vision. pp. 37-53. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.629, + 0.918, + 0.682 + ], + "angle": 0, + "content": "28. Muandet, K., Balduzzi, D., Scholkopf, B.: Domain generalization via invariant feature representation. In: International conference on machine learning. pp. 10-18. PMLR (2013)" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.683, + 0.918, + 0.735 + ], + "angle": 0, + "content": "29. Oppenheim, A., Lim, J., Kopec, G., Pohlig, S.: Phase in speech and pictures. In: ICASSP'79. IEEE International Conference on Acoustics, Speech, and Signal Processing. vol. 4, pp. 632-637. IEEE (1979)" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.736, + 0.918, + 0.789 + ], + "angle": 0, + "content": "30. Peng, X., Bai, Q., Xia, X., Huang, Z., Saenko, K., Wang, B.: Moment matching for multi-source domain adaptation. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 1406-1415 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.79, + 0.918, + 0.842 + ], + "angle": 0, + "content": "31. Peng, X., Huang, Z., Sun, X., Saenko, K.: Domain agnostic learning with disentangled representations. In: International Conference on Machine Learning. pp. 5102-5112. PMLR (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.843, + 0.918, + 0.895 + ], + "angle": 0, + "content": "32. Piotrowski, L.N., Campbell, F.W.: A demonstration of the visual importance and flexibility of spatial-frequency amplitude and phase. Perception 11(3), 337-346 (1982)" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.896, + 0.918, + 0.949 + ], + "angle": 0, + "content": "33. Piratla, V., Netrapalli, P., Sarawagi, S.: Efficient domain generalization via common-specific low-rank decomposition. In: International Conference on Machine Learning. pp. 7728-7738. PMLR (2020)" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.057, + 0.921, + 0.949 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.08, + 0.017, + 0.105, + 0.03 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.159, + 0.016, + 0.394, + 0.033 + ], + "angle": 0, + "content": "W. Huang, Y. Shi et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.057, + 0.918, + 0.129 + ], + "angle": 0, + "content": "34. Seo, S., Suh, Y., Kim, D., Kim, G., Han, J., Han, B.: Learning to optimize domain specific normalization for domain generalization. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXII 16. pp. 68-83. Springer (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.13, + 0.918, + 0.183 + ], + "angle": 0, + "content": "35. Shi, Y., Seely, J., Torr, P.H., Siddharth, N., Hannun, A., Usunier, N., Synnaeve, G.: Gradient matching for domain generalization. arXiv preprint arXiv:2104.09937 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.184, + 0.918, + 0.236 + ], + "angle": 0, + "content": "36. Venkateswara, H., Eusebio, J., Chakraborty, S., Panchanathan, S.: Deep hashing network for unsupervised domain adaptation. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 5018-5027 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.237, + 0.918, + 0.307 + ], + "angle": 0, + "content": "37. Wang, G., Han, H., Shan, S., Chen, X.: Cross-domain face presentation attack detection via multi-domain disentangled representation learning. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 6678-6687 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.308, + 0.918, + 0.361 + ], + "angle": 0, + "content": "38. Wang, P., Zhang, Z., Lei, Z., Zhang, L.: Sharpness-aware gradient matching for domain generalization. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3769-3778 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.362, + 0.918, + 0.414 + ], + "angle": 0, + "content": "39. Wang, Y., Qi, L., Shi, Y., Gao, Y.: Feature-based style randomization for domain generalization. IEEE Transactions on Circuits and Systems for Video Technology 32(8), 5495-5509 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.415, + 0.918, + 0.468 + ], + "angle": 0, + "content": "40. Xu, M., Zhang, J., Ni, B., Li, T., Wang, C., Tian, Q., Zhang, W.: Adversarial domain adaptation with domain mixup. In: Proceedings of the AAAI conference on artificial intelligence. vol. 34, pp. 6502-6509 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.469, + 0.918, + 0.522 + ], + "angle": 0, + "content": "41. Xu, Q., Zhang, R., Zhang, Y., Wang, Y., Tian, Q.: A fourier-based framework for domain generalization. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14383-14392 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.522, + 0.918, + 0.575 + ], + "angle": 0, + "content": "42. Yang, F.E., Cheng, Y.C., Shiau, Z.Y., Wang, Y.C.F.: Adversarial teacher-student representation learning for domain generalization. Advances in Neural Information Processing Systems 34, 19448-19460 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.576, + 0.918, + 0.628 + ], + "angle": 0, + "content": "43. Yang, Z., Yu, Y., You, C., Steinhardt, J., Ma, Y.: Rethinking bias-variance tradeoff for generalization of neural networks. In: International Conference on Machine Learning. pp. 10767-10777. PMLR (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.629, + 0.918, + 0.682 + ], + "angle": 0, + "content": "44. Zhang, J., Qi, L., Shi, Y., Gao, Y.: Mvdg: A unified multi-view framework for domain generalization. In: European Conference on Computer Vision. pp. 161-177. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.683, + 0.918, + 0.735 + ], + "angle": 0, + "content": "45. Zhang, X., Xu, R., Yu, H., Dong, Y., Tian, P., Cui, P.: Flatness-aware minimization for domain generalization. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5189-5202 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.736, + 0.918, + 0.788 + ], + "angle": 0, + "content": "46. Zhao, S., Gong, M., Liu, T., Fu, H., Tao, D.: Domain generalization via entropy regularization. Advances in Neural Information Processing Systems 33, 16096-16107 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.789, + 0.918, + 0.842 + ], + "angle": 0, + "content": "47. Zhou, K., Yang, Y., Hospedales, T., Xiang, T.: Deep domain-adversarial image generation for domain generalisation. In: Proceedings of the AAAI conference on artificial intelligence. vol. 34, pp. 13025-13032 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.843, + 0.918, + 0.896 + ], + "angle": 0, + "content": "48. Zhu, W., Lu, L., Xiao, J., Han, M., Luo, J., Harrison, A.P.: Localized adversarial domain generalization. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7108-7118 (2022)" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.057, + 0.918, + 0.896 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/Representation Enhancement-Stabilization_ Reducing Bias-Variance of Domain Generalization/f28a3a3b-4152-48f4-99b3-107ae4c33ff2_origin.pdf b/2024/Representation Enhancement-Stabilization_ Reducing Bias-Variance of Domain Generalization/f28a3a3b-4152-48f4-99b3-107ae4c33ff2_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..f0dce421c3de0897a8a80ca4b7f930501413dad7 --- /dev/null +++ b/2024/Representation Enhancement-Stabilization_ Reducing Bias-Variance of Domain Generalization/f28a3a3b-4152-48f4-99b3-107ae4c33ff2_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d5f53c26beabad337d0ec834b700b86503c757cc7909118f5950b69619e26322 +size 3487358 diff --git a/2024/Representation Enhancement-Stabilization_ Reducing Bias-Variance of Domain Generalization/full.md b/2024/Representation Enhancement-Stabilization_ Reducing Bias-Variance of Domain Generalization/full.md new file mode 100644 index 0000000000000000000000000000000000000000..e7319cee0e5e908fcebb004b64990b86932f93e4 --- /dev/null +++ b/2024/Representation Enhancement-Stabilization_ Reducing Bias-Variance of Domain Generalization/full.md @@ -0,0 +1,363 @@ +# Representation Enhancement-Stabilization: Reducing Bias-Variance of Domain Generalization + +Wei Huang1, Yilei Shi1, Zhitong Xiong1, and Xiao Xiang Zhu1,2 + +1. Technical University of Munich, Arcisstraße 21, 80333 München, Germany +2. Munich Center for Machine Learning, 80333 Munich +{w2wei.huang, yilei.shi, zhitong.xiong, xiaoxiang.zhu}@tum.de + +Abstract. Domain Generalization (DG) focuses on enhancing the generalization of deep learning models trained on multiple source domains to adapt to unseen target domains. This paper explores DG through the lens of bias-variance decomposition, uncovering that test errors in DG predominantly arise from cross-domain bias and variance. Inspired by this insight, we introduce a Representation Enhancement-Stabilization (RES) framework, comprising a Representation Enhancement (RE) module and a Representation Stabilization (RS) module. In RE, a novel set of feature frequency augmentation techniques is used to progressively reduce cross-domain bias during feature extraction. Furthermore, in RS, a novel Mutual Exponential Moving Average (MEMA) strategy is designed to stabilize model optimization for diminishing cross-domain variance during training. Collectively, the whole RES method can significantly enhance model generalization. We evaluate RES on five benchmark datasets and the results show that it outperforms multiple advanced DG methods. Our code will be available at https://github.com/zhu-xlab/DG-RES. + +Keywords: Domain generalization $\cdot$ Frequency Domain $\cdot$ Data Augmentation $\cdot$ Bias-variance Decomposition + +# 1 Introduction + +In recent years, deep learning models have achieved remarkable advancements in computer vision tasks, primarily under the assumption that training and test data are independent and identically distributed. However, in real-world scenarios, this assumption often breaks down due to domain shift/bias, where the distribution of test data markedly differs from that of training data. Domain shift can significantly degrade the performance of deep models on unfamiliar, unseen target domains. For instance, self-driving models trained under daylight conditions might fail to perform effectively in nighttime environments. + +To address these challenges, Domain Generalization (DG) has been proposed as a robust solution. DG focuses on training models on multiple diverse yet related source domains for robust performance on arbitrary unseen target domains. Numerous DG approaches have been developed, including adversarial training [23, 48], disentangled representation learning [37], meta-learning [7], model ensemble [2], flatness optimization [4], and domain data augmentation [8, 41, 47]. + +In this study we highlight the cross-domain generalization of deep models from the perspective of bias-variance decomposition. Geman et al. [12] first used bias-variance to decompose the mean square error (MSE) of neural networks. Building on this, Yang et al. [43] developed the bias-variance decomposition for the cross-entropy (CE) loss in classification tasks, which was extended by Arpit et al. [2] to analyze the impact of variance in DG, as illustrated below: + +$$ +\mathbb {E} _ {x, y \sim D _ {s}} [ \mathrm {C E} (y, f (x; D _ {s})) ] = \underbrace {\mathbb {E} _ {x , y \sim D _ {s}} [ \mathrm {C E} (y , \bar {f} (x)) ]} _ {\text {B i a s} ^ {2}} + \underbrace {\mathbb {E} _ {x , D _ {s}} [ \mathrm {K L} (\bar {f} (x) , f (x ; D _ {s})) ]} _ {\text {V a r i a n c e}}, +$$ + +(1) + +where $D_{s}$ represents source domains used for training. The formula decomposes the expected source CE loss into two components: bias and variance. Bias evaluates the discrepancy between model's average predictions and actual outcomes, while variance measures sensitivity to fluctuations during training. + +Motivated by them, we extend the bias-variance decomposition to DG to comprehensively analysis the cross-domain prediction errors as the following: + +$$ +\begin{array}{l} \mathbb {E} _ {x, y \sim D _ {t}} [ \mathrm {C E} (y, f (x _ {t}; D _ {t})) ] = \mathbb {E} _ {x, y \sim D _ {s}} [ \mathrm {C E} (y, f (x _ {s} + \delta ; D _ {s})) ] \\ = \underbrace {\mathbb {E} _ {x , y \sim D _ {s}} [ \mathrm {C E} (y , \bar {f} (x _ {s} + \delta)) ]} \\ \text {C r o s s - D o m a i n B i a s} ^ {2} \\ + \underbrace {\mathbb {E} _ {x \sim D _ {s}} \left[ \mathrm {K L} \left(\bar {f} \left(x _ {s} + \delta\right) , f \left(x _ {s} + \delta ; D _ {s}\right)\right) \right]} _ {\text {C r o s s - D o m a i n V a r i a n c e}} \tag {2} \\ + \Delta_ {d o m a i n} (\text {B i a s} ^ {2}, \text {V a r i a n c e}). \\ \end{array} +$$ + +This formula analyzes the expected CE loss when a model trained in source domains $D_{s}$ , denoted as $f(x_{s};D_{s})$ , is applied to new target domains $D_{t}$ , with respect to domain shift $\delta$ between source and target images, $x_{s}$ and $x_{t}$ . The formula is decomposed into three components: + +- (1) Cross-Domain Bias, as shown in the left of Fig. 1. It quantifies the discrepancy between the model's average prediction in the target domain and the actual labels. High cross-domain bias dramatically enlarges the divergence between the distribution learnt from source domains and the true underlying distribution of target domains. By progressively reducing bias in feature extraction during training, i.e., $f(x_{s}) \gets f(x_{s} + \delta)$ , there is an approximate bias error $\mathbb{E}_{x,y\sim D_s}[\mathrm{CE}(y,\bar{f} (x))]\gets \mathbb{E}_{x,y\sim D_s}[\mathrm{CE}(y,\bar{f} (x_s + \delta))]$ . +- (2) Cross-Domain Variance, as shown in the right of Fig. 1. It measures the variability and sensitivity in the model's predictions from source to target domains. The model trained on the source domain tends to generate unstable predictions when confronted with the target domain data, with elevated cross-domain variance. Based on cross-domain bias reduction in (1), further decreasing cross-domain variance results in an approximate variance error $\mathbb{E}_{x\sim D_s}[\mathrm{KL}(\bar{f} (x_s),f(x_s;D_s))]\gets \mathbb{E}_{x\sim D_s}[\mathrm{KL}(\bar{f} (x_s + \delta),f(x_s + \delta ;D_s))]$ . +- (3) $\varDelta_{domain}(\mathrm{Bias}^2$ , Variance). This term denotes the extra test error of the intertwined effects of the cross-domain bias and variance. The decrease of both bias and variance can contribute to its reduction. + +Overall, diminishing cross-domain bias and variance can narrow the gap between Eq. (1) and Eq. (2), enabling models trained on source domains to effectively adapt to target domains. This analytical framework provides actionable insights for training models capable of generalizing across diverse data distributions. + +![](images/6679d66906eec54370f44285981249f7ead866cd92a0972158a3749a3e8918cb.jpg) +(a) Challenge of cross-domain bias + +![](images/0871b06d20bf67577b0aedd804006b951b44147dcbcf1d76f223eab837d39919.jpg) +(b) Challenge of cross-domain variance +Fig. 1: Two challenges impeding DG from the view of bias-variance decomposition. + +Based on this analytical framework, we proposed a joint Representation Enhancement-Stabilization (RES) method, which consists of a Representation Enhancement (RE) module and a Representation Stabilization (RS) module to alleviate cross-domain bias ans cross-domain variance, respectively. In RE, we propose a novel set of feature augmentation techniques within the frequency domain extracted through Fast Fourier Transform (FFT), including random noise, random dropout, and mixup. RE effectively broadens cross-domain representation space and thereby enhances robust representation learning, leading to a reduction in domain bias. Notably, the augmentations of RE target only the amplitude spectrum of high-dimensional features, with their phase spectrum remaining unchanged for maintaining domain-irrelevant semantics. In RS, we present a novel Mutual Exponential Meaning Average (MEMA) model parameter optimization strategy. This approach facilitates a dynamic exchange and integration of parameters between teacher and student models, enhancing model stability in out-of-domain performance throughout training process. + +We evaluate the proposed RES method on five DG benchmark datasets, including PACS [22], VLCS [10], OfficeHome [36], TerraIncognita [3], and DomainNet [30], and comparison experiment results indicate our RES outperforms multiple state-of-the-art DG methods. + +Our contributions are summarized as follows: (1) We rethink the prediction error of DG from an extended bias-variance decomposition analytical framework. (2) Based on the bias-variance decomposition, we propose a RES method for DG, consisting of a Representation Enhancement (RE) module and a Representation Stabilization (RS) module, to respectively reduce cross-domain bias and variance. (3) Comparison experiments demonstrate that our RES method outperforms multiple state-of-the-art DG methods, verifying its effectiveness. + +# 2 Related Work + +In this section, we first give a brief review of the mainstreams DG methods. Then we delve into a detailed introduction to data augmentation methods and flatness-aware methods, which are relevant to our RES. + +# 2.1 Domain Generalization + +Various DG methods have been developed. One type of DG methods focus on distribution alignment among source domains to learn domain-invariant representations via all kinds of strategies, including domain alignment [23, 28, 46], adversarial training [23, 48], causal learning [25, 26], disentangled representation learning [31, 37], low-rank decomposition [33], self-supervised learning [1, 19], meta-learning [6, 7, 44], and data normalization [20, 34, 34]. + +# 2.2 Data Augmentation Methods + +Our RE module aligns with another paradigm of data augmentation methods, which enhance model generalization to unseen domains by diversifying source domain data at two distinct levels. + +(1) Image level [8,41,42,47]. Yang et al. [42] and Zhou et al. [47] have used cross-domain image-to-image translation based on domain-adversarial learning, further applying these augmented images in model training. Xu et al. [41] introduced a novel Fourier-based image augmentation strategy for DG, based on the assumption that phase information in the frequency domain contains high-level semantics and is less affected by domain shift. + +(2) Feature level [14, 24, 39]. Techniques such as the straightforward perturbation of high-dimensional feature embeddings with Gaussian noise by Li et al. [24], the use of Adaptive Instance Normalization (AdaIN) [17] for feature-level style randomization augmentation by Wang et al. [39], and the DomainDrop framework by Guo et al. [14] which drops high domain-activated channels to enhance feature channel robustness against domain shift, have been proposed. + +# 2.3 Flatness-aware Methods + +Flatness-aware methods aiming at avoiding unstable sharp optimization draws rising interest under the situation where a validation set has a different distribution from the test data, leading to a non-i.i.d. scenario. Cha et al. [4] hypothesized that in the non-i.i.d scenario the generalization gap between flat and sharp minima is more pronounced, and then proposed a Stochastic Weight Averaging Densely (SWAD) method aiming to locate flat minima and reduce domain gap. Moreover, Arpit et al. [2] argued that models trained on specific training domains often display erratic performance on test domains with distribution shifts. To mitigate this, they suggested the use of ensembling moving average (EoA) models, thereby decreasing uncertainty in test domains for enhanced domain generalization. Sharpness-Aware Minimization (SAM) [11] and an improved + +Sharpness-Aware Gradient Matching (SGAM) [38] were further proposed to enhance model generalization by simultaneously minimizing the loss value and its sharpness, leading to more robust models. + +# 3 RES-based Domain Generalization + +# 3.1 Notations and Overview + +Assuming there are $K$ distinct yet semantic-related source domains represented as $D_{s} = \{D_{s}^{1}, D_{s}^{2}, \ldots, D_{s}^{K}\}$ , each originating from different distributions and comprising $N_{k}$ image-label pairs, denoted as $D_{s}^{k} = \{(x_{i}, y_{i})\}_{i=1}^{N_{k}}$ . In this study, the model training is independent of domain labels, so there is no domain subscript in $(x, y)$ . The primary objective of DG is to leverage these source domains to train a model, parameterized by $\theta$ , that exhibits strong generalization capabilities and can effectively adapt to unseen target domains. In RES, there are two models sharing the same architecture, a student model $M^{S}$ parameterized with $\theta^{S}$ and a teacher model $M^{T}$ parameterized with $\theta^{T}$ . $M^{S}$ is used for model training and $M^{T}$ is utilized for model validation and test. + +The proposed RES-based DG framework is shown in Fig. 2 with its training process summarized in Algorithm 1. RES consists of: (1) a RE module that augments amplitude spectrum of feature maps while maintaining the phase spectrum, to expand cross-domain feature representation; and (2) a RS module that mutually fuses the parameters of the teacher and student models to each other. + +![](images/5b1cd187f3bbe4313188096cb4604f8d9caeee0635f6c54a1e19f2fb70851cf7.jpg) +Fig. 2: Workflow of the proposed RES-based domain generalization framework, which consists of Representation Enhancement (RE) and Representation Stabilization (RS). RE aims to enhance feature representation to reduce cross-domain bias via a set of augmentations performing on the amplitude spectrum of frequency domain, while RS aims to further stabilize cross-domain performance to reduce cross-domain variance through a novel MEMA optimization strategy between student and teacher models. + +# 3.2 Feature Frequency Augmentation-Based RE + +Our RE modules augments source data at the feature level. During training, the student model $M^S$ is split into two parts: $M^{S-}$ which includes all layers up to a middle layer $l$ for feature map extraction, and $M^{S+}$ which consists of the remaining layers for the remaining feature extraction and classification. + +Decomposing Phase and Amplitude From Features. Given an image $x \in \mathcal{R}^{3 \times \bar{H} \times \bar{W}}$ where $\bar{H} \times \bar{W}$ represents image spatial size, a corresponding multilayer feature map $f \in \mathcal{R}^{C \times H \times W}$ can be extracted by $M^{S^{-}}$ , where $C$ signifies the number of channels, and $H, W$ denote the height and width of the feature maps respectively. It is formulated as + +$$ +f = M ^ {S -} (x). \tag {3} +$$ + +To delve into the frequency properties of these feature maps, we employ a 2D FFT across the spatial dimensions of each channel, ensuring that every layer of $f$ undergoes this transformative process. This approach allows us to comprehensively analyze the frequency content within each individual feature channel. For channel $c$ , its FFT formulation is as follows: + +$$ +F ^ {c} (u, v) = \sum_ {h = 0} ^ {H - 1} \sum_ {w = 0} ^ {W - 1} f ^ {c} (h, w) \cdot e ^ {- j 2 \pi \left(\frac {u h}{H} + \frac {v w}{W}\right)}, \tag {4} +$$ + +where $f^{c}(h,w)$ represents the pixel value at the spatial location $(h,w)$ within the range of $[H - 1,W - 1]$ in channel $c$ , and $F^{c}(u,v)$ denotes the complex frequency spectrum at the frequency coordinates $(u,v)$ of channel $c$ . The term $j$ is the imaginary unit. + +Then we compute a multi-layer phase spectrum denoted as $\mathcal{P}$ and a multi-layer amplitude spectrum represented as $\mathcal{A}$ , from the complex frequency spectrum $F$ , as + +$$ +\mathcal {P} \left(F ^ {c} (u, v)\right) = \arctan 2 \left(\operatorname {I m a g} \left(F ^ {c} (u, v)\right), \operatorname {R e a l} \left(F ^ {c} (u, v)\right)\right), \tag {5} +$$ + +$$ +\mathcal {A} \left(F ^ {c} (u, v)\right) = \sqrt {\operatorname {R e a l} \left(F ^ {c} (u , v)\right) ^ {2} + \operatorname {I m a g} \left(F ^ {c} (u , v)\right) ^ {2}}. \tag {6} +$$ + +Here, $\operatorname{Real}(F^c(u,v))$ and $\operatorname{Imag}(F^c(u,v))$ indicate the real and imaginary parts of the channel- $c$ frequency spectrum $F^c$ , respectively. $\mathcal{P}$ and $\mathcal{A}$ have the same size of $C \times H \times W$ as $f$ . Drawing inspiration from [15,29,32,41], there is a common assumption that the phase component of frequency spectrum maintains the high-level semantics of the original data while the amplitude component contains its low-level statistics. In this study, we extend it from the image level to feature level, where we regard feature maps as high-dimension signals with the same characteristic of phase-amplitude decomposition as images. + +Frequency Augmentations on Features. To avoid damaging the semantics of features, augmentations are only implemented on the amplitude $\mathcal{A}$ to obtain an augmented $\mathcal{A}'$ while the phase $\mathcal{P}$ keeps unchanged, as + +$$ +\mathcal {A} ^ {\prime} = \operatorname {A u g} (\mathcal {A}). \tag {7} +$$ + +Given a batch of images $\{x_{1}, x_{2}, \ldots, x_{B}\}$ where $B$ is batch size, a corresponding batch of amplitudes $\{\mathcal{A}_1, \mathcal{A}_2, \ldots, \mathcal{A}_B\}$ can be obtained by Eqs. (3), (4), and (6). During each training iteration, we randomly select and apply one augmentation strategy from a pre-defined augmentation list, including none, random noise, random dropout, and mixup. Here none represents no augmentation, and the details of the rest three strategies are as follows: + +- Random noise, denoted as $\mathrm{RE}^{ns}$ . Given a multi-layer amplitude $\mathcal{A}_i \in \mathcal{R}^{C \times H \times W}$ , we add Gaussian noise $G \in \mathcal{R}^{1 \times H \times W}$ to it. The elements of $G$ adhere to a Gaussian distribution characterized by mean $\mu$ and standard deviation $\sigma$ , i.e., $G \sim \mathcal{N}(\mu, \sigma^2)$ . Leveraging the broadcasting mechanism, this noise addition is equally conducted to each layer of $\mathcal{A}_i$ , as: + +$$ +\mathcal {A} _ {i} ^ {\prime} = \mathcal {A} _ {i} * (I + G) = \mathcal {A} _ {i} * (I + \mathcal {N} (\mu , \sigma^ {2})), \tag {8} +$$ + +where $I$ represents the Identity Matrix. This approach guarantees that the same frequency components throughout all channels of the feature map are enhanced with the same noise intensity. + +- Random dropout, denoted as $\mathrm{RE}^{dp}$ . We randomly generate a binary dropout mask, $M \in \mathcal{R}^{1 \times H \times W}$ , where each element is set either to 0 with probability $p$ or to 1 with a reversal probability $1 - p$ . The dimension of $\mathcal{M}$ is also expanded to $C \times H \times W$ by broadcasting. The dropout-augmented amplitude $\mathcal{A}'$ is calculated by: + +$$ +\mathcal {A} _ {i} ^ {\prime} = \mathcal {A} \cdot M. \tag {9} +$$ + +This operation introduces sparsity of amplitudes, aiming at reducing their overfitting for better robustness. It is worth mentioning that the element of $M$ at the location $[0,0]$ always equals 1, which represents the mean value of the feature map, to avoid over-masking on features. + +- Mixup, denoted as $\mathrm{RE}^{mix}$ . Unlike random noise and random dropout which operate on single amplitudes extracted from individual samples, mixup engages in a cross-sample amplitude interaction by linear interpolation on different amplitudes extracted from different samples. Specifically, for a given amplitude spectrum $\mathcal{A}_i$ , mixup randomly selects another amplitude $\mathcal{A}_j$ from the set $\{\mathcal{A}_1, \mathcal{A}_2, \dots, \mathcal{A}_B\}$ , and then blends $\mathcal{A}_i$ with $\mathcal{A}_j$ , resulting in a new, hybrid amplitude spectrum $\mathcal{A}_i'$ as follows: + +$$ +\mathcal {A} _ {i} ^ {\prime} = \alpha \mathcal {A} _ {i} + (1 - \alpha) \mathcal {A} _ {j}, \tag {10} +$$ + +where $\alpha$ is a uniform random coefficient between [0,1]. This procedure enriches the amplitude characteristics by promoting a more diverse fusion of feature styles across domains. It boosts the model's generalization by augmenting intricate cross-domain amplitude patterns effectively. + +Reconstructing Features. For each pair of phase and augmented amplitude components, $\mathcal{P}$ and $\mathcal{A}'$ , an augmented multi-layer feature map, $f'$ , can be reconstructed via inverse Fast Fourier Transform (iFFT) as + +$$ +f ^ {\prime c} (h, w) = \frac {1}{H W} \sum_ {u = 0} ^ {H - 1} \sum_ {v = 0} ^ {W - 1} \mathcal {A} ^ {\prime} \left(F ^ {c} (u, v)\right) \cdot e ^ {j 2 \pi \left(\frac {u h}{H} + \frac {v w}{W}\right)} \cdot e ^ {j \mathcal {P} \left(F ^ {c} (u, v)\right)} \tag {11} +$$ + +Here, $e^{j2\pi \left(\frac{uh}{H} + \frac{vw}{W}\right)}$ is a complex exponential term representing the frequency component in the inverse Fourier transform, and $e^{j\mathcal{P}(F^c(u,v))}$ reintroduces the phase information into the transform. + +Finally, $f'$ is used for further feature extraction and class classification by the remaining layers of the student model, $M^{S+}$ , as + +$$ +p = M ^ {S +} \left(f ^ {\prime}\right). \tag {12} +$$ + +# 3.3 MEMA-based RS + +As shown in Fig. 2, there are two models, one student model $M^S$ with parameters $\theta^S$ and one teacher model $M^T$ with parameters $\theta^T$ . Their parameters undergo mutual updates to provide stabilization to each other, referred to as $\mathrm{RE}^T$ and $\mathrm{RE}^S$ for the teacher and student models, respectively. Their processes are formulated as + +$$ +\begin{array}{l} \theta^ {T} \leftarrow \tau^ {T} \theta^ {T} + (1 - \tau^ {T}) \theta^ {S}, \quad \tau^ {T} \in [ 0, 1), \tag {13} \\ \theta^ {S} \leftarrow (1 - \tau^ {S}) \theta^ {S} + \tau^ {S} \theta^ {T}, \quad \tau^ {S} \in [ 0, 1). \\ \end{array} +$$ + +Here, only the student model $M^S$ is actively trained, while the teacher model $M^T$ is responsible solely for parameter updates. The student model $M^S$ is the primary learner, actively absorbing new knowledge and adapting to training data. In contrast, the teacher model $M^T$ acts as a robust and stable reference using moving average updating strategy, providing more robust and consistent parameters by reducing the variance of training batches [2]. + +Both the values of $\tau^T$ and $\tau^S$ are small and they play a mutual role in this process: (1) Stability from the teacher model. With a small $\tau^T$ , the teacher model's parameters change slowly, ensuring that they remain stable and consistent.(2) Gradual integration for the student model. A small $\tau^S$ implies that the student model integrates the teacher's parameters slowly and steadily. This allows the student model to gradually assimilate the stable characteristics of the teacher model. To simplify the hyperparameter tuning process in experiments, $\tau^T$ is set to a small constant value of 1e-3, while $\tau^S$ is variable for different datasets. + +In summary, the student model acts as an active learner, continually evolving and assimilating new information, whereas the teacher model functions as a reliable guide, providing a stable parameter reference to shield the student model from the effects of cross-domain variance in training batches. + +
Algorithm 1 Training Process of our RES-based DG
1:Student model MS with initialized parameters θS, teacher model MT with initialized parameters θT, all training sample pairs {(xi,yi)}N, and batch size B
2:for iter = 1 to Niter do
3:Sample Batch Data: {(xi,yi)}B
4:Extract Feature: f← MS-(x)
5:Decompose Phase and Amplitude: P, A← FFT(f)
6:Augmentation Selection: Aug ← [random noise, random dropout, mixup, none]
7:Augment Amplitude: A' ← Aug(A)
8:Reconstruct feature: f' = iFFT(P, A')
9:Compute prediction: p ← MS+(f')
10:Compute Loss: E(θS) ← l(p, y)
11:Update Student Model with Loss: θS ← θS - η∇θL(θS)
12:Integrate Student Model to Teacher Model: θT = τTθT + (1 - τT)θS
13:Stabilize Student Model with Teacher Model: θS = (1 - τS)θS + τSθT
14:end for
+ +# 3.4 Supervision Loss + +During each training iteration with a batch size of $B$ , the empirical risk minimization (ERM) loss function is used to optimize the student model $M^S$ across all the training domains as + +$$ +\mathcal {E} \left(\theta^ {S}\right) = \frac {1}{B} \sum_ {i = 1} ^ {B} \ell \left(p _ {i}, y _ {i}\right) = \frac {1}{B} \sum_ {i = 1} ^ {B} \ell \left(M ^ {S} \left(x _ {i}; \theta^ {S}\right), y _ {i}\right), \tag {14} +$$ + +where the detailed progress of $M^S (x_i;\theta^S)$ come from Eqs. (3)-(7) and (12) and $\ell$ represents the CE loss. Here $\mathcal{E}(\theta^S)$ represents the basic ERM loss. + +# 4 Experiments + +This section outlines benchmark datasets and experimental settings, conducts an ablation study on RES, compares it with advanced DG methods, examines the impact of $\tau^S$ , and quantitatively assesses RES's effects in reducing cross-domain bias and variance and improving accuracy across domains. In Supplementary Material, we compare the effects of data augmentation from two perspectives: image-level vs. feature-level frequency augmentation, and normal feature augmentation vs. feature frequency augmentation. This comparison demonstrates the superiority of our feature frequency augmentation. + +# 4.1 Benchmark Datasets + +We evaluate the proposed RES method on five DG datasets: (1) PACS [22], which contains 9,991 images in 7 object categories across 4 diverse domains; (2) VLCS [10], which comprises 10,729 examples from 5 categories, collected from 4 + +domains; (3) OfficeHome [36], which encompasses around 15,500 samples in 65 categories from four domains. (4) TerraIncognita [3], which contains 24,788 images within 10 categories from four domains; and (5) DomainNet [30], a vast dataset with 586,575 images across 345 categories, spanning 6 domains. + +# 4.2 Experimental Settings + +Basic settings. We utilize ResNet-50 [16], pretrained on ImageNet-1K [9], as the backbone. Our codes are constructed on Domainbed's framework [13], allocating $80\%$ of in-domain data for training and the remaining $20\%$ for validation, with test performed on an unseen domain. PACS, VLCS, and OfficeHome are trained for 10K iterations and validated every 300 iterations, TerraIncognita is trained for 3K iterations and validated every 300 iterations, and DomainNet undergoes 400K iterations of training and 1K iterations of validation. Adam is used as the optimizer, with a batch size of 128 and a learning rate of 5e-5. + +Settings of RES. In every training iteration, we apply the frequency augmentation of RE randomly to one of the four potential middle-layer positions within ResNet-50. These positions are strategically located at the ends of the $1^{st}$ , $2^{nd}$ , $3^{rd}$ , and $4^{th}$ residual blocks, ensuring diverse augmentation effects across different layers of the model. There are some hyperparameters in RES summarized in Table 1, where $\tau^S$ is variable and tuned based on the validation performance as discussed in Section 4.4. All the experiments of RES are conducted for three times and the mean values and standard deviations are reported. + +Table 1: Hyperparameters of RES. + +
REnsREdpREmixRSS→TRST→S
G ~ N(0.75, 0.752)p = 0.5α ~ U(0,1)τT = 0.001τS is variable
+ +# 4.3 Ablation Study + +The ablation study of RES's modules, including sub-strategies of RE and RS, on PACS are provided in Table 2 and Fig. 3. RE employs random noise $(\mathrm{RE}^{ns})$ , random dropout $(\mathrm{RE}^{dp})$ , and mixup $(\mathrm{RE}^{mix})$ techniques for feature augmentation, and RS employs MEMA containing representation stabilization for both student models $(\mathrm{RS}^s)$ and teacher models $(\mathrm{RS}^t)$ . + +After applying all three RE strategies, there is significant performance improvement from ERM's 85.5 to (ERM+RE)'s 88.8 as shown in Table 2. with their training trajectory shown in Fig. 3. The results verify the effect of RE in reducing cross-domain bias. Additionally, the application of RS strategies reduces their standard deviations from ERM's 1.1 to (ERM+RS)'s 0.5, demonstrating their + +effectiveness in reducing cross-domain variance. When RE and RS strategies are applied in conjunction, as shown in the last highlighted row of the table, the model achieves optimal test performance, with an average accuracy increase to 90.0 with the smallest standard deviation of 0.3, significantly surpassing ERM. + +Through this ablation study, the complementary effects of RE and RS strategies in enhancing model generalization across unseen domains are clearly demonstrated, underscoring the significance of reducing cross-domain bias-variance. + +Table 2: Ablation study of the proposed RES on PACS. + +
ERMREnsREdpREmixRStRSsPACSAvg
ArtClipartPaintingSketch
✓84.4±0.681.1±0.796.3±0.580.2±2.585.5±1.1
✓✓88.8±1.282.2±0.597.0±0.681.2±0.687.3±0.7
✓✓87.6±0.382.7±1.096.8±0.882.1±1.687.3±0.9
✓✓89.4±0.281.4±1.196.6±0.485.1±0.988.5±0.7
✓✓✓✓89.0±0.983.0±0.497.0±0.485.1±0.788.8±0.6
✓✓87.9±0.183.1±0.296.0±0.482.2±1.287.3±0.5
✓✓✓90.9±0.482.0±0.597.9±0.382.2±0.888.2±0.5
✓✓✓✓✓✓91.6±0.484.2±0.398.1±0.186.1±0.490.0±0.3
+ +![](images/08137a2f85d6c20c90f34abf5ff7bf3125ae05adadf886a224b498eba15e1bb0.jpg) +Fig. 3: Training convergence of RES's modules on PACS. + +![](images/3e7d83a1c426e5e53f56d75ba58d7fd4828ab466af565363c02db4e145af166a.jpg) + +![](images/d7dae5339d1af2a7e57ddcf594eb8fb4c21397d097ababc5751992cc946f6dd9.jpg) + +![](images/6530f02b979b7854895fed0920e83bed28e7ec39f543cea4f41de33364c2cb0d.jpg) + +# 4.4 Values of $\tau^S$ in MEMA + +In RES, the MEMA's $\tau^S$ is the sole hyperparameter requiring adjustment across various benchmarks. We explored five distinct values: [0, 1e-3, 5e-3, 1e-2, 5e-2] for each benchmark, with results shown in Fig. 4. The results indicate that the optimal $\tau^S$ value varies by dataset, attributed to differences in their convergence speed. Subsequent comparison experiments will report out-of-domain accuracies for $\tau^S$ corresponding to the highest in-domain accuracies observed. + +![](images/0f20f1dd4586800170ead541a7e393ac55aa4fa61987021677557514a810dd7e.jpg) +Fig. 4: In-domain (validation) and out-of-domain (test) accuracies vary with $\tau^S$ . + +![](images/c0647528d9cb92293ebe9c318fbb5f9500179dbf9bd0705beea3990f4bc71b8e.jpg) +In-domain Accuracy Out-of-domain Accuracy + +![](images/5750340daef3e26b6ecded5b8076874aee4552a1b83e2bc1590faadbfed360e6.jpg) + +![](images/d754bad40a69e14f967c846c05c2a1e028772096c5264ff0647e3cfeb6c2a31b.jpg) + +![](images/7d738aba32c1e28d31f67b15a48b4c51be8761854e05b5eae2c6942b1267ec95.jpg) + +# 4.5 In Comparison With Advanced DG Methods + +To objectively assess the effectiveness of the proposed RES, we conducted comparative analyses against several state-of-the-art DG methods. This comparison spans the baseline ERM, various data augmentation strategies (e.g., Mixup [40], CCFP [21], DomainDrop [14]), flatness-aware methods (such as SWAD [4], SMA [2], SAGM [38], FAD [45]), and some other DG approaches (including FISH [35], RSC [18], DAC-SC [20], GVRT [27], MIRO [5]). These evaluations were performed across five benchmarks utilizing the ResNet-50 backbone, with results detailed in Table 3. The domain-wise performance of the datasets of our RES are provided in Supplementary Material. Notably, RES incorporates both data augmentation and flatness-aware strategies, positioning it uniquely to tackle the dual challenges of cross-domain bias and variance simultaneously. + +Table 3: Performance comparisons between some SOTA methods and our RES. + +
MethodPACSVLCSOfficeHomeTerraIncDomainNetAvg
ERM [ICLR'20] [13]85.7±0.577.4±0.367.5±0.547.2±0.441.2±0.263.8
RSC [ECCV'20] [18]85.2±0.977.1±0.565.5±0.946.6±1.038.9±0.562.7
Mixup [AAAI'20] [40]84.3±0.577.7±0.469.0±0.148.9±0.839.6±0.163.9
FISH [ICLR'21] [35]85.5±0.377.8±0.368.6±0.445.1±1.342.7±0.263.9
GVRT [ECCV'22] [27]85.1±0.379.0±0.270.1±0.148.0±0.244.1±0.165.2
MIRO [ECCV'22] [5]85.4±0.479.0±0.070.5±0.450.4±1.144.3±0.265.9
SMA [NeurIPS'22] [2]87.5±0.278.2±0.270.6±0.150.3±0.546.0±0.166.5
SWAD [NeurIPS'21] [4]88.1±0.179.1±0.170.6±0.250.0±0.346.5±0.166.9
CCFP [ICCV'23] [21]86.6±0.278.9±0.368.9±0.148.6±0.441.2±0.164.8
DAC-SC [CVPR'23] [20]87.5±0.178.7±0.370.3±0.246.5±0.344.9±0.165.6
FAD [ICCV'23] [45]88.2±0.278.9±0.869.2±0.545.7±1.044.4±0.165.3
SAGM [CVPR'23] [38]86.6±0.280.0±0.370.1±0.248.8±0.945.0±0.266.1
DomainDrop [ICCV'23] [14]87.9±0.379.8±0.368.7±0.151.5±0.444.4±0.566.5
Our RES90.0±0.379.8±0.271.8±0.351.4±0.646.7±0.267.9
+ +The table showcases the RES algorithm's performance across five benchmarks, with "Avg" representing the average performance across them. This provides a comprehensive view of each method's generalizability and effectiveness in DG. The proposed RES demonstrates superior performance, particularly high- + +lighted by its leading average score, indicating its robustness and adaptability across a diverse set of domains. The comparative analysis confirms the superiority and effectiveness of the RES in mitigating cross-domain bias and variance. + +# 4.6 Effects of RES on Domain Generalization + +This subsection quantitatively describes the advantages of RES on DG from three perspectives, including reducing cross-domain bias, reducing cross-domain variance, and improving both in-domain and out-of-domain performance. + +# Reducing Cross-domain Bias. + +First, we measure the cosine similarity between the mean feature embeddings obtained from the global average pooling layer of ResNet-50 across both source and target datasets within PACS and OfficeHome. This assessment encompasses comparisons between the baseline ERM and our RES, as illustrated in Fig. 5. Our findings indicate that RES significantly enhances the cosine similarity between source and target data compared to ERM, demonstrating its capability in diminishing cross-domain l + +![](images/98537107b1eac575a6fa9f9a64829580caa55ad8d6d66e01c521662b59d6c3a9.jpg) +Source-target Similarity of ERM + +![](images/3d30c3a52669d4733368fb3698831c99d9c06b62a9fa49ed93762eec97ee1fdf.jpg) +Source-target Similarity of RES +Fig. 5: Cosine similarity between mean features of source domains and target domain. The higher the similarity is, the lower the cross-domain bias is. + +Reducing Cross-domain Variance. Secondly, we present the training convergence trajectories alongside the corresponding cumulative instability for four DG scenarios within the PACS dataset, as illustrated in Fig. 6. It is observable that, in comparison to the baseline method ERM represented by the blue color, the proposed RES depicted in red markedly reduces the fluctuations in both in-domain and out-of-domain accuracies with almost no instability increase after the initial increase stage during training, which spans approximately the first 2000 iterations. Benefiting from the stable and narrower performance gap between in-domain (validation) and out-of-domain (test) sets, the best models chosen based on validation sets demonstrate consistent performance on test sets. + +Improving Both In-domain and Out-of-domain Performance. As shown in Table 4, the comparison between ERM and RES on PACS illustrates significant enhancements in both in-domain and out-of-domain accuracies achieved by RES. While ERM delivers reliable in-domain performance, it falls short in adapting to unseen domains. In contrast, RES not only boosts in-domain accuracy from ERM's 96.5 to its 98.0 but also lifts the out-of-domain accuracy + +![](images/39fe8bbb5f1e067860c76e62dfea3015376927079bca55e272d0b05d37440c84.jpg) + +![](images/f3a31b1b8db827223e6e43aad1ce85273232cfc0491c1404b9177188d29cffe2.jpg) + +![](images/a0521d58739d241ef42273e6b53f1ebcee8198a2033df73d3a97f6dcb2d0ab35.jpg) + +![](images/027a54cdeccde8d259cee14db751fb3daa804cbdfa50d12968278f4223fd0561.jpg) + +![](images/e595c659d3bd6cb12a812db44ae5b0ddf8c57f3a4112414f030e7072d7a8836a.jpg) +Fig. 6: Training stability of ERM and RES on PACS, evaluated by the absolute first derivative of accuracy every 300 iterations $(|f'(Acc)| = |Acc_t - Acc_{t-1}|)$ , capturing accuracy change rate. Accumulative instability aggregates $|f'(Acc)|$ up to iteration $t$ , with lower values indicating more stable performance and less variance. + +![](images/fc040e88401742c649beef52436be9131fdd48af3115187c790327d92a5dc5e2.jpg) + +![](images/59be9bedf7b27b910332f53db8bd723ff6bad11f9a825a99fabf887ac72633c9.jpg) + +![](images/a4afb372df88a610c22aa9610c66d5cd420b3c56ff77c58e817fcea3176f6180.jpg) + +from $85.5 \pm 1.1$ to $90.0 \pm 0.3$ . This robust improvement underscores RES's robust capability to refine model performance across various distributions. + +Table 4: In-domain and out-of-domain accuracies of PACS between ERM and RES. + +
SplitMethodPACSAvg
ArtClipartPaintingSketch
In-domain (Validation Set)ERM97.2±0.097.0±0.495.4±0.296.3±0.496.5±0.3
Our RES98.4±0.298.0±0.097.4±0.398.2±0.398.0±0.2
Out-of-domain (Test Set)ERM84.4±0.681.1±0.796.3±0.580.2±2.585.5±1.1
Our RES91.6±0.484.2±0.398.1±0.186.1±0.490.0±0.3
+ +# 5 Conclusion + +In this study, we explored DG through bias-variance decomposition, an analytical framework that offers deeper insights into the challenges and solutions for cross-domain generalization in deep learning models. Our approach quantified the detrimental effects of cross-domain bias and variance on model performance and therefore introduced the Representation Enhancement-Stabilization (RES) method as a potent solution. The RES method effectively minimizes the cross-domain bias and variance, enhancing model adaptability and generalization across unseen target domains. Empirical evaluations on five DG benchmarks reveals that the proposed RES markedly outperforms multiple advanced DG methods in improving model generalization. + +# Acknowledgements + +The work was supported by the German Research Foundation (DFG GZ: ZH 498/18-1; Project number: 519016653). The work of Z. Xiong was supported by the German Federal Ministry for the Environment, Nature Conservation, Nuclear Safety and Consumer Protection (BMUV) based on a resolution of the German Bundestag (grant number: 67KI32002B; Acronym: $EKAPEx$ ). The work of X. Zhu was also supported by the German Federal Ministry of Education and Research (BMBF) in the framework of the international future AI lab "AI4EO - Artificial Intelligence for Earth Observation: Reasoning, Uncertainties, Ethics and Beyond" (grant number: 01DD20001), by German Federal Ministry for Economic Affairs and Climate Action in the framework of the "national center of excellence ML4Earth" (grant number: 50EE2201C) and by the Munich Center for Machine Learning. + +# References + +1. Albuquerque, I., Naik, N., Li, J., Keskar, N., Socher, R.: Improving out-of-distribution generalization via multi-task self-supervised pretraining. arXiv preprint arXiv:2003.13525 (2020) +2. Arpit, D., Wang, H., Zhou, Y., Xiong, C.: Ensemble of averages: Improving model selection and boosting performance in domain generalization. Advances in Neural Information Processing Systems 35, 8265-8277 (2022) +3. Beery, S., Van Horn, G., Perona, P.: Recognition in terra incognita. In: Proceedings of the European conference on computer vision (ECCV). pp. 456-473 (2018) +4. Cha, J., Chun, S., Lee, K., Cho, H.C., Park, S., Lee, Y., Park, S.: Swad: Domain generalization by seeking flat minima. Advances in Neural Information Processing Systems 34, 22405-22418 (2021) +5. Cha, J., Lee, K., Park, S., Chun, S.: Domain generalization by mutual-information regularization with pre-trained models. In: European Conference on Computer Vision. pp. 440-457. Springer (2022) +6. Chen, C., Li, J., Han, X., Liu, X., Yu, Y.: Compound domain generalization via meta-knowledge encoding. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7119-7129 (2022) +7. Chen, J., Gao, Z., Wu, X., Luo, J.: Meta-causal learning for single domain generalization. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7683-7692 (2023) +8. Cubuk, E.D., Zoph, B., Shlens, J., Le, Q.V.: Randaugment: Practical automated data augmentation with a reduced search space. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops. pp. 702-703 (2020) +9. Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: Imagenet: A large-scale hierarchical image database. In: 2009 IEEE conference on computer vision and pattern recognition. pp. 248-255. IEEE (2009) +10. Fang, C., Xu, Y., Rockmore, D.N.: Unbiased metric learning: On the utilization of multiple datasets and web images for softening bias. In: Proceedings of the IEEE International Conference on Computer Vision. pp. 1657-1664 (2013) +11. Foret, P., Kleiner, A., Mobahi, H., Neyshabur, B.: Sharpness-aware minimization for efficiently improving generalization. arXiv preprint arXiv:2010.01412 (2020) +12. Geman, S., Bienenstock, E., Doursat, R.: Neural networks and the bias/variance dilemma. Neural computation 4(1), 1-58 (1992) +13. Gulrajani, I., Lopez-Paz, D.: In search of lost domain generalization. In: International Conference on Learning Representations (2020) +14. Guo, J., Qi, L., Shi, Y.: Domaindrop: Suppressing domain-sensitive channels for domain generalization. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 19114-19124 (2023) +15. Hansen, B.C., Hess, R.F.: Structural sparseness and spatial phase alignment in natural scenes. JOSA A 24(7), 1873-1885 (2007) +16. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 770-778 (2016) +17. Huang, X., Belongie, S.: Arbitrary style transfer in real-time with adaptive instance normalization. In: Proceedings of the IEEE international conference on computer vision. pp. 1501-1510 (2017) + +18. Huang, Z., Wang, H., Xing, E.P., Huang, D.: Self-challenging improves cross-domain generalization. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part II 16. pp. 124-140. Springer (2020) +19. Kim, D., Yoo, Y., Park, S., Kim, J., Lee, J.: Selfreg: Self-supervised contrastive regularization for domain generalization. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 9619-9628 (2021) +20. Lee, S., Bae, J., Kim, H.Y.: Decompose, adjust, compose: Effective normalization by playing with frequency for domain generalization. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11776-11785 (2023) +21. Li, C., Zhang, D., Huang, W., Zhang, J.: Cross contrasting feature perturbation for domain generalization. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 1327-1337 (2023) +22. Li, D., Yang, Y., Song, Y.Z., Hospedales, T.M.: Deeper, broader and artier domain generalization. In: Proceedings of the IEEE international conference on computer vision. pp. 5542-5550 (2017) +23. Li, H., Pan, S.J., Wang, S., Kot, A.C.: Domain generalization with adversarial feature learning. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 5400-5409 (2018) +24. Li, P., Li, D., Li, W., Gong, S., Fu, Y., Hospedales, T.M.: A simple feature augmentation for domain generalization. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 8886-8895 (2021) +25. Lv, F., Liang, J., Li, S., Zang, B., Liu, C.H., Wang, Z., Liu, D.: Causality inspired representation learning for domain generalization. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8046-8056 (2022) +26. Mahajan, D., Tople, S., Sharma, A.: Domain generalization using causal matching. In: International Conference on Machine Learning. pp. 7313-7324. PMLR (2021) +27. Min, S., Park, N., Kim, S., Park, S., Kim, J.: Grounding visual representations with texts for domain generalization. In: European Conference on Computer Vision. pp. 37-53. Springer (2022) +28. Muandet, K., Balduzzi, D., Scholkopf, B.: Domain generalization via invariant feature representation. In: International conference on machine learning. pp. 10-18. PMLR (2013) +29. Oppenheim, A., Lim, J., Kopec, G., Pohlig, S.: Phase in speech and pictures. In: ICASSP'79. IEEE International Conference on Acoustics, Speech, and Signal Processing. vol. 4, pp. 632-637. IEEE (1979) +30. Peng, X., Bai, Q., Xia, X., Huang, Z., Saenko, K., Wang, B.: Moment matching for multi-source domain adaptation. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 1406-1415 (2019) +31. Peng, X., Huang, Z., Sun, X., Saenko, K.: Domain agnostic learning with disentangled representations. In: International Conference on Machine Learning. pp. 5102-5112. PMLR (2019) +32. Piotrowski, L.N., Campbell, F.W.: A demonstration of the visual importance and flexibility of spatial-frequency amplitude and phase. Perception 11(3), 337-346 (1982) +33. Piratla, V., Netrapalli, P., Sarawagi, S.: Efficient domain generalization via common-specific low-rank decomposition. In: International Conference on Machine Learning. pp. 7728-7738. PMLR (2020) + +34. Seo, S., Suh, Y., Kim, D., Kim, G., Han, J., Han, B.: Learning to optimize domain specific normalization for domain generalization. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXII 16. pp. 68-83. Springer (2020) +35. Shi, Y., Seely, J., Torr, P.H., Siddharth, N., Hannun, A., Usunier, N., Synnaeve, G.: Gradient matching for domain generalization. arXiv preprint arXiv:2104.09937 (2021) +36. Venkateswara, H., Eusebio, J., Chakraborty, S., Panchanathan, S.: Deep hashing network for unsupervised domain adaptation. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 5018-5027 (2017) +37. Wang, G., Han, H., Shan, S., Chen, X.: Cross-domain face presentation attack detection via multi-domain disentangled representation learning. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 6678-6687 (2020) +38. Wang, P., Zhang, Z., Lei, Z., Zhang, L.: Sharpness-aware gradient matching for domain generalization. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3769-3778 (2023) +39. Wang, Y., Qi, L., Shi, Y., Gao, Y.: Feature-based style randomization for domain generalization. IEEE Transactions on Circuits and Systems for Video Technology 32(8), 5495-5509 (2022) +40. Xu, M., Zhang, J., Ni, B., Li, T., Wang, C., Tian, Q., Zhang, W.: Adversarial domain adaptation with domain mixup. In: Proceedings of the AAAI conference on artificial intelligence. vol. 34, pp. 6502-6509 (2020) +41. Xu, Q., Zhang, R., Zhang, Y., Wang, Y., Tian, Q.: A fourier-based framework for domain generalization. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14383-14392 (2021) +42. Yang, F.E., Cheng, Y.C., Shiau, Z.Y., Wang, Y.C.F.: Adversarial teacher-student representation learning for domain generalization. Advances in Neural Information Processing Systems 34, 19448-19460 (2021) +43. Yang, Z., Yu, Y., You, C., Steinhardt, J., Ma, Y.: Rethinking bias-variance tradeoff for generalization of neural networks. In: International Conference on Machine Learning. pp. 10767-10777. PMLR (2020) +44. Zhang, J., Qi, L., Shi, Y., Gao, Y.: Mvdg: A unified multi-view framework for domain generalization. In: European Conference on Computer Vision. pp. 161-177. Springer (2022) +45. Zhang, X., Xu, R., Yu, H., Dong, Y., Tian, P., Cui, P.: Flatness-aware minimization for domain generalization. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5189-5202 (2023) +46. Zhao, S., Gong, M., Liu, T., Fu, H., Tao, D.: Domain generalization via entropy regularization. Advances in Neural Information Processing Systems 33, 16096-16107 (2020) +47. Zhou, K., Yang, Y., Hospedales, T., Xiang, T.: Deep domain-adversarial image generation for domain generalisation. In: Proceedings of the AAAI conference on artificial intelligence. vol. 34, pp. 13025-13032 (2020) +48. Zhu, W., Lu, L., Xiao, J., Han, M., Luo, J., Harrison, A.P.: Localized adversarial domain generalization. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7108-7118 (2022) \ No newline at end of file diff --git a/2024/Representation Enhancement-Stabilization_ Reducing Bias-Variance of Domain Generalization/images.zip b/2024/Representation Enhancement-Stabilization_ Reducing Bias-Variance of Domain Generalization/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..059e4dc6ce5aa38a23d56a8278f892d287a43c06 --- /dev/null +++ b/2024/Representation Enhancement-Stabilization_ Reducing Bias-Variance of Domain Generalization/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3b79d38315300c4e46423542adc0b546af83990f3810a1e1db1f437bb014a86d +size 614959 diff --git a/2024/Representation Enhancement-Stabilization_ Reducing Bias-Variance of Domain Generalization/layout.json b/2024/Representation Enhancement-Stabilization_ Reducing Bias-Variance of Domain Generalization/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..c146876998c0fc19a3b54a2a43bdb5c3acec6308 --- /dev/null +++ b/2024/Representation Enhancement-Stabilization_ Reducing Bias-Variance of Domain Generalization/layout.json @@ -0,0 +1,11012 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 33, + 30, + 379, + 66 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 30, + 379, + 66 + ], + "spans": [ + { + "bbox": [ + 33, + 30, + 379, + 66 + ], + "type": "text", + "content": "Representation Enhancement-Stabilization: Reducing Bias-Variance of Domain Generalization" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 52, + 86, + 360, + 99 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 86, + 360, + 99 + ], + "spans": [ + { + "bbox": [ + 52, + 86, + 360, + 99 + ], + "type": "text", + "content": "Wei Huang1, Yilei Shi1, Zhitong Xiong1, and Xiao Xiang Zhu1,2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 51, + 108, + 361, + 141 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 51, + 108, + 361, + 141 + ], + "spans": [ + { + "bbox": [ + 51, + 108, + 361, + 141 + ], + "type": "text", + "content": "1. Technical University of Munich, Arcisstraße 21, 80333 München, Germany \n2. Munich Center for Machine Learning, 80333 Munich \n{w2wei.huang, yilei.shi, zhitong.xiong, xiaoxiang.zhu}@tum.de" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 166, + 351, + 341 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 166, + 351, + 341 + ], + "spans": [ + { + "bbox": [ + 61, + 166, + 351, + 341 + ], + "type": "text", + "content": "Abstract. Domain Generalization (DG) focuses on enhancing the generalization of deep learning models trained on multiple source domains to adapt to unseen target domains. This paper explores DG through the lens of bias-variance decomposition, uncovering that test errors in DG predominantly arise from cross-domain bias and variance. Inspired by this insight, we introduce a Representation Enhancement-Stabilization (RES) framework, comprising a Representation Enhancement (RE) module and a Representation Stabilization (RS) module. In RE, a novel set of feature frequency augmentation techniques is used to progressively reduce cross-domain bias during feature extraction. Furthermore, in RS, a novel Mutual Exponential Moving Average (MEMA) strategy is designed to stabilize model optimization for diminishing cross-domain variance during training. Collectively, the whole RES method can significantly enhance model generalization. We evaluate RES on five benchmark datasets and the results show that it outperforms multiple advanced DG methods. Our code will be available at https://github.com/zhu-xlab/DG-RES." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 352, + 351, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 352, + 351, + 373 + ], + "spans": [ + { + "bbox": [ + 61, + 352, + 351, + 373 + ], + "type": "text", + "content": "Keywords: Domain generalization " + }, + { + "bbox": [ + 61, + 352, + 351, + 373 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 61, + 352, + 351, + 373 + ], + "type": "text", + "content": " Frequency Domain " + }, + { + "bbox": [ + 61, + 352, + 351, + 373 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 61, + 352, + 351, + 373 + ], + "type": "text", + "content": " Data Augmentation " + }, + { + "bbox": [ + 61, + 352, + 351, + 373 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 61, + 352, + 351, + 373 + ], + "type": "text", + "content": " Bias-variance Decomposition" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 33, + 392, + 128, + 404 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 392, + 128, + 404 + ], + "spans": [ + { + "bbox": [ + 33, + 392, + 128, + 404 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 33, + 416, + 379, + 511 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 416, + 379, + 511 + ], + "spans": [ + { + "bbox": [ + 33, + 416, + 379, + 511 + ], + "type": "text", + "content": "In recent years, deep learning models have achieved remarkable advancements in computer vision tasks, primarily under the assumption that training and test data are independent and identically distributed. However, in real-world scenarios, this assumption often breaks down due to domain shift/bias, where the distribution of test data markedly differs from that of training data. Domain shift can significantly degrade the performance of deep models on unfamiliar, unseen target domains. For instance, self-driving models trained under daylight conditions might fail to perform effectively in nighttime environments." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 33, + 512, + 379, + 584 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 512, + 379, + 584 + ], + "spans": [ + { + "bbox": [ + 33, + 512, + 379, + 584 + ], + "type": "text", + "content": "To address these challenges, Domain Generalization (DG) has been proposed as a robust solution. DG focuses on training models on multiple diverse yet related source domains for robust performance on arbitrary unseen target domains. Numerous DG approaches have been developed, including adversarial training [23, 48], disentangled representation learning [37], meta-learning [7], model ensemble [2], flatness optimization [4], and domain data augmentation [8, 41, 47]." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [], + "page_size": [ + 413, + 615 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 30, + 35, + 380, + 106 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 35, + 380, + 106 + ], + "spans": [ + { + "bbox": [ + 30, + 35, + 380, + 106 + ], + "type": "text", + "content": "In this study we highlight the cross-domain generalization of deep models from the perspective of bias-variance decomposition. Geman et al. [12] first used bias-variance to decompose the mean square error (MSE) of neural networks. Building on this, Yang et al. [43] developed the bias-variance decomposition for the cross-entropy (CE) loss in classification tasks, which was extended by Arpit et al. [2] to analyze the impact of variance in DG, as illustrated below:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 36, + 114, + 378, + 142 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 36, + 114, + 378, + 142 + ], + "spans": [ + { + "bbox": [ + 36, + 114, + 378, + 142 + ], + "type": "interline_equation", + "content": "\\mathbb {E} _ {x, y \\sim D _ {s}} [ \\mathrm {C E} (y, f (x; D _ {s})) ] = \\underbrace {\\mathbb {E} _ {x , y \\sim D _ {s}} [ \\mathrm {C E} (y , \\bar {f} (x)) ]} _ {\\text {B i a s} ^ {2}} + \\underbrace {\\mathbb {E} _ {x , D _ {s}} [ \\mathrm {K L} (\\bar {f} (x) , f (x ; D _ {s})) ]} _ {\\text {V a r i a n c e}},", + "image_path": "54114314cfcf947f80fc5a307a4e868f232ed11cfd0e52acd40a757d6f124f59.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 300, + 142, + 378, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 300, + 142, + 378, + 152 + ], + "spans": [ + { + "bbox": [ + 300, + 142, + 378, + 152 + ], + "type": "text", + "content": "(1)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 30, + 153, + 379, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 153, + 379, + 200 + ], + "spans": [ + { + "bbox": [ + 30, + 153, + 379, + 200 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 30, + 153, + 379, + 200 + ], + "type": "inline_equation", + "content": "D_{s}" + }, + { + "bbox": [ + 30, + 153, + 379, + 200 + ], + "type": "text", + "content": " represents source domains used for training. The formula decomposes the expected source CE loss into two components: bias and variance. Bias evaluates the discrepancy between model's average predictions and actual outcomes, while variance measures sensitivity to fluctuations during training." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 30, + 201, + 379, + 225 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 201, + 379, + 225 + ], + "spans": [ + { + "bbox": [ + 30, + 201, + 379, + 225 + ], + "type": "text", + "content": "Motivated by them, we extend the bias-variance decomposition to DG to comprehensively analysis the cross-domain prediction errors as the following:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 61, + 233, + 378, + 319 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 233, + 378, + 319 + ], + "spans": [ + { + "bbox": [ + 61, + 233, + 378, + 319 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathbb {E} _ {x, y \\sim D _ {t}} [ \\mathrm {C E} (y, f (x _ {t}; D _ {t})) ] = \\mathbb {E} _ {x, y \\sim D _ {s}} [ \\mathrm {C E} (y, f (x _ {s} + \\delta ; D _ {s})) ] \\\\ = \\underbrace {\\mathbb {E} _ {x , y \\sim D _ {s}} [ \\mathrm {C E} (y , \\bar {f} (x _ {s} + \\delta)) ]} \\\\ \\text {C r o s s - D o m a i n B i a s} ^ {2} \\\\ + \\underbrace {\\mathbb {E} _ {x \\sim D _ {s}} \\left[ \\mathrm {K L} \\left(\\bar {f} \\left(x _ {s} + \\delta\\right) , f \\left(x _ {s} + \\delta ; D _ {s}\\right)\\right) \\right]} _ {\\text {C r o s s - D o m a i n V a r i a n c e}} \\tag {2} \\\\ + \\Delta_ {d o m a i n} (\\text {B i a s} ^ {2}, \\text {V a r i a n c e}). \\\\ \\end{array}", + "image_path": "5cffff631d65c4d37e7ba71259ae0c1fbaafc9e65db9264282e704fa581fd83b.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 30, + 327, + 379, + 375 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 327, + 379, + 375 + ], + "spans": [ + { + "bbox": [ + 30, + 327, + 379, + 375 + ], + "type": "text", + "content": "This formula analyzes the expected CE loss when a model trained in source domains " + }, + { + "bbox": [ + 30, + 327, + 379, + 375 + ], + "type": "inline_equation", + "content": "D_{s}" + }, + { + "bbox": [ + 30, + 327, + 379, + 375 + ], + "type": "text", + "content": ", denoted as " + }, + { + "bbox": [ + 30, + 327, + 379, + 375 + ], + "type": "inline_equation", + "content": "f(x_{s};D_{s})" + }, + { + "bbox": [ + 30, + 327, + 379, + 375 + ], + "type": "text", + "content": ", is applied to new target domains " + }, + { + "bbox": [ + 30, + 327, + 379, + 375 + ], + "type": "inline_equation", + "content": "D_{t}" + }, + { + "bbox": [ + 30, + 327, + 379, + 375 + ], + "type": "text", + "content": ", with respect to domain shift " + }, + { + "bbox": [ + 30, + 327, + 379, + 375 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 30, + 327, + 379, + 375 + ], + "type": "text", + "content": " between source and target images, " + }, + { + "bbox": [ + 30, + 327, + 379, + 375 + ], + "type": "inline_equation", + "content": "x_{s}" + }, + { + "bbox": [ + 30, + 327, + 379, + 375 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 30, + 327, + 379, + 375 + ], + "type": "inline_equation", + "content": "x_{t}" + }, + { + "bbox": [ + 30, + 327, + 379, + 375 + ], + "type": "text", + "content": ". The formula is decomposed into three components:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 37, + 381, + 380, + 582 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 37, + 381, + 379, + 465 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 37, + 381, + 379, + 465 + ], + "spans": [ + { + "bbox": [ + 37, + 381, + 379, + 465 + ], + "type": "text", + "content": "- (1) Cross-Domain Bias, as shown in the left of Fig. 1. It quantifies the discrepancy between the model's average prediction in the target domain and the actual labels. High cross-domain bias dramatically enlarges the divergence between the distribution learnt from source domains and the true underlying distribution of target domains. By progressively reducing bias in feature extraction during training, i.e., " + }, + { + "bbox": [ + 37, + 381, + 379, + 465 + ], + "type": "inline_equation", + "content": "f(x_{s}) \\gets f(x_{s} + \\delta)" + }, + { + "bbox": [ + 37, + 381, + 379, + 465 + ], + "type": "text", + "content": ", there is an approximate bias error " + }, + { + "bbox": [ + 37, + 381, + 379, + 465 + ], + "type": "inline_equation", + "content": "\\mathbb{E}_{x,y\\sim D_s}[\\mathrm{CE}(y,\\bar{f} (x))]\\gets \\mathbb{E}_{x,y\\sim D_s}[\\mathrm{CE}(y,\\bar{f} (x_s + \\delta))]" + }, + { + "bbox": [ + 37, + 381, + 379, + 465 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 37, + 465, + 380, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 37, + 465, + 380, + 548 + ], + "spans": [ + { + "bbox": [ + 37, + 465, + 380, + 548 + ], + "type": "text", + "content": "- (2) Cross-Domain Variance, as shown in the right of Fig. 1. It measures the variability and sensitivity in the model's predictions from source to target domains. The model trained on the source domain tends to generate unstable predictions when confronted with the target domain data, with elevated cross-domain variance. Based on cross-domain bias reduction in (1), further decreasing cross-domain variance results in an approximate variance error " + }, + { + "bbox": [ + 37, + 465, + 380, + 548 + ], + "type": "inline_equation", + "content": "\\mathbb{E}_{x\\sim D_s}[\\mathrm{KL}(\\bar{f} (x_s),f(x_s;D_s))]\\gets \\mathbb{E}_{x\\sim D_s}[\\mathrm{KL}(\\bar{f} (x_s + \\delta),f(x_s + \\delta ;D_s))]" + }, + { + "bbox": [ + 37, + 465, + 380, + 548 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 37, + 548, + 380, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 37, + 548, + 380, + 582 + ], + "spans": [ + { + "bbox": [ + 37, + 548, + 380, + 582 + ], + "type": "text", + "content": "- (3) " + }, + { + "bbox": [ + 37, + 548, + 380, + 582 + ], + "type": "inline_equation", + "content": "\\varDelta_{domain}(\\mathrm{Bias}^2" + }, + { + "bbox": [ + 37, + 548, + 380, + 582 + ], + "type": "text", + "content": ", Variance). This term denotes the extra test error of the intertwined effects of the cross-domain bias and variance. The decrease of both bias and variance can contribute to its reduction." + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 32, + 10, + 39, + 18 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 32, + 10, + 39, + 18 + ], + "spans": [ + { + "bbox": [ + 32, + 10, + 39, + 18 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 65, + 9, + 163, + 20 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 9, + 163, + 20 + ], + "spans": [ + { + "bbox": [ + 65, + 9, + 163, + 20 + ], + "type": "text", + "content": "W. Huang, Y. Shi et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 413, + 615 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 30, + 35, + 381, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 35, + 381, + 84 + ], + "spans": [ + { + "bbox": [ + 30, + 35, + 381, + 84 + ], + "type": "text", + "content": "Overall, diminishing cross-domain bias and variance can narrow the gap between Eq. (1) and Eq. (2), enabling models trained on source domains to effectively adapt to target domains. This analytical framework provides actionable insights for training models capable of generalizing across diverse data distributions." + } + ] + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 33, + 118, + 217, + 196 + ], + "blocks": [ + { + "bbox": [ + 33, + 118, + 217, + 196 + ], + "lines": [ + { + "bbox": [ + 33, + 118, + 217, + 196 + ], + "spans": [ + { + "bbox": [ + 33, + 118, + 217, + 196 + ], + "type": "image", + "image_path": "6679d66906eec54370f44285981249f7ead866cd92a0972158a3749a3e8918cb.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 62, + 197, + 187, + 206 + ], + "lines": [ + { + "bbox": [ + 62, + 197, + 187, + 206 + ], + "spans": [ + { + "bbox": [ + 62, + 197, + 187, + 206 + ], + "type": "text", + "content": "(a) Challenge of cross-domain bias" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 232, + 115, + 380, + 194 + ], + "blocks": [ + { + "bbox": [ + 232, + 115, + 380, + 194 + ], + "lines": [ + { + "bbox": [ + 232, + 115, + 380, + 194 + ], + "spans": [ + { + "bbox": [ + 232, + 115, + 380, + 194 + ], + "type": "image", + "image_path": "0871b06d20bf67577b0aedd804006b951b44147dcbcf1d76f223eab837d39919.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 237, + 196, + 376, + 207 + ], + "lines": [ + { + "bbox": [ + 237, + 196, + 376, + 207 + ], + "spans": [ + { + "bbox": [ + 237, + 196, + 376, + 207 + ], + "type": "text", + "content": "(b) Challenge of cross-domain variance" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 35, + 215, + 375, + 228 + ], + "lines": [ + { + "bbox": [ + 35, + 215, + 375, + 228 + ], + "spans": [ + { + "bbox": [ + 35, + 215, + 375, + 228 + ], + "type": "text", + "content": "Fig. 1: Two challenges impeding DG from the view of bias-variance decomposition." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 30, + 261, + 381, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 261, + 381, + 443 + ], + "spans": [ + { + "bbox": [ + 30, + 261, + 381, + 443 + ], + "type": "text", + "content": "Based on this analytical framework, we proposed a joint Representation Enhancement-Stabilization (RES) method, which consists of a Representation Enhancement (RE) module and a Representation Stabilization (RS) module to alleviate cross-domain bias ans cross-domain variance, respectively. In RE, we propose a novel set of feature augmentation techniques within the frequency domain extracted through Fast Fourier Transform (FFT), including random noise, random dropout, and mixup. RE effectively broadens cross-domain representation space and thereby enhances robust representation learning, leading to a reduction in domain bias. Notably, the augmentations of RE target only the amplitude spectrum of high-dimensional features, with their phase spectrum remaining unchanged for maintaining domain-irrelevant semantics. In RS, we present a novel Mutual Exponential Meaning Average (MEMA) model parameter optimization strategy. This approach facilitates a dynamic exchange and integration of parameters between teacher and student models, enhancing model stability in out-of-domain performance throughout training process." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 30, + 447, + 386, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 447, + 386, + 496 + ], + "spans": [ + { + "bbox": [ + 30, + 447, + 386, + 496 + ], + "type": "text", + "content": "We evaluate the proposed RES method on five DG benchmark datasets, including PACS [22], VLCS [10], OfficeHome [36], TerraIncognita [3], and DomainNet [30], and comparison experiment results indicate our RES outperforms multiple state-of-the-art DG methods." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 30, + 499, + 382, + 584 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 499, + 382, + 584 + ], + "spans": [ + { + "bbox": [ + 30, + 499, + 382, + 584 + ], + "type": "text", + "content": "Our contributions are summarized as follows: (1) We rethink the prediction error of DG from an extended bias-variance decomposition analytical framework. (2) Based on the bias-variance decomposition, we propose a RES method for DG, consisting of a Representation Enhancement (RE) module and a Representation Stabilization (RS) module, to respectively reduce cross-domain bias and variance. (3) Comparison experiments demonstrate that our RES method outperforms multiple state-of-the-art DG methods, verifying its effectiveness." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 326, + 9, + 347, + 19 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 9, + 347, + 19 + ], + "spans": [ + { + "bbox": [ + 326, + 9, + 347, + 19 + ], + "type": "text", + "content": "RES" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 372, + 10, + 380, + 19 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 372, + 10, + 380, + 19 + ], + "spans": [ + { + "bbox": [ + 372, + 10, + 380, + 19 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 413, + 615 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 33, + 32, + 136, + 46 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 32, + 136, + 46 + ], + "spans": [ + { + "bbox": [ + 33, + 32, + 136, + 46 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 33, + 57, + 379, + 92 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 57, + 379, + 92 + ], + "spans": [ + { + "bbox": [ + 33, + 57, + 379, + 92 + ], + "type": "text", + "content": "In this section, we first give a brief review of the mainstreams DG methods. Then we delve into a detailed introduction to data augmentation methods and flatness-aware methods, which are relevant to our RES." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 33, + 110, + 175, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 110, + 175, + 121 + ], + "spans": [ + { + "bbox": [ + 33, + 110, + 175, + 121 + ], + "type": "text", + "content": "2.1 Domain Generalization" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 33, + 128, + 379, + 201 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 128, + 379, + 201 + ], + "spans": [ + { + "bbox": [ + 33, + 128, + 379, + 201 + ], + "type": "text", + "content": "Various DG methods have been developed. One type of DG methods focus on distribution alignment among source domains to learn domain-invariant representations via all kinds of strategies, including domain alignment [23, 28, 46], adversarial training [23, 48], causal learning [25, 26], disentangled representation learning [31, 37], low-rank decomposition [33], self-supervised learning [1, 19], meta-learning [6, 7, 44], and data normalization [20, 34, 34]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 33, + 217, + 206, + 230 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 217, + 206, + 230 + ], + "spans": [ + { + "bbox": [ + 33, + 217, + 206, + 230 + ], + "type": "text", + "content": "2.2 Data Augmentation Methods" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 33, + 236, + 379, + 271 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 236, + 379, + 271 + ], + "spans": [ + { + "bbox": [ + 33, + 236, + 379, + 271 + ], + "type": "text", + "content": "Our RE module aligns with another paradigm of data augmentation methods, which enhance model generalization to unseen domains by diversifying source domain data at two distinct levels." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 33, + 272, + 379, + 343 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 272, + 379, + 343 + ], + "spans": [ + { + "bbox": [ + 33, + 272, + 379, + 343 + ], + "type": "text", + "content": "(1) Image level [8,41,42,47]. Yang et al. [42] and Zhou et al. [47] have used cross-domain image-to-image translation based on domain-adversarial learning, further applying these augmented images in model training. Xu et al. [41] introduced a novel Fourier-based image augmentation strategy for DG, based on the assumption that phase information in the frequency domain contains high-level semantics and is less affected by domain shift." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 33, + 344, + 379, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 344, + 379, + 415 + ], + "spans": [ + { + "bbox": [ + 33, + 344, + 379, + 415 + ], + "type": "text", + "content": "(2) Feature level [14, 24, 39]. Techniques such as the straightforward perturbation of high-dimensional feature embeddings with Gaussian noise by Li et al. [24], the use of Adaptive Instance Normalization (AdaIN) [17] for feature-level style randomization augmentation by Wang et al. [39], and the DomainDrop framework by Guo et al. [14] which drops high domain-activated channels to enhance feature channel robustness against domain shift, have been proposed." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 33, + 432, + 180, + 444 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 432, + 180, + 444 + ], + "spans": [ + { + "bbox": [ + 33, + 432, + 180, + 444 + ], + "type": "text", + "content": "2.3 Flatness-aware Methods" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 33, + 451, + 379, + 583 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 451, + 379, + 583 + ], + "spans": [ + { + "bbox": [ + 33, + 451, + 379, + 583 + ], + "type": "text", + "content": "Flatness-aware methods aiming at avoiding unstable sharp optimization draws rising interest under the situation where a validation set has a different distribution from the test data, leading to a non-i.i.d. scenario. Cha et al. [4] hypothesized that in the non-i.i.d scenario the generalization gap between flat and sharp minima is more pronounced, and then proposed a Stochastic Weight Averaging Densely (SWAD) method aiming to locate flat minima and reduce domain gap. Moreover, Arpit et al. [2] argued that models trained on specific training domains often display erratic performance on test domains with distribution shifts. To mitigate this, they suggested the use of ensembling moving average (EoA) models, thereby decreasing uncertainty in test domains for enhanced domain generalization. Sharpness-Aware Minimization (SAM) [11] and an improved" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 33, + 11, + 39, + 18 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 11, + 39, + 18 + ], + "spans": [ + { + "bbox": [ + 33, + 11, + 39, + 18 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 66, + 9, + 162, + 19 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 9, + 162, + 19 + ], + "spans": [ + { + "bbox": [ + 66, + 9, + 162, + 19 + ], + "type": "text", + "content": "W. Huang, Y. Shi et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 413, + 615 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 30, + 34, + 380, + 71 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 34, + 380, + 71 + ], + "spans": [ + { + "bbox": [ + 30, + 34, + 380, + 71 + ], + "type": "text", + "content": "Sharpness-Aware Gradient Matching (SGAM) [38] were further proposed to enhance model generalization by simultaneously minimizing the loss value and its sharpness, leading to more robust models." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 31, + 89, + 259, + 102 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 31, + 89, + 259, + 102 + ], + "spans": [ + { + "bbox": [ + 31, + 89, + 259, + 102 + ], + "type": "text", + "content": "3 RES-based Domain Generalization" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 31, + 115, + 183, + 127 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 31, + 115, + 183, + 127 + ], + "spans": [ + { + "bbox": [ + 31, + 115, + 183, + 127 + ], + "type": "text", + "content": "3.1 Notations and Overview" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 30, + 136, + 380, + 254 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 136, + 380, + 254 + ], + "spans": [ + { + "bbox": [ + 30, + 136, + 380, + 254 + ], + "type": "text", + "content": "Assuming there are " + }, + { + "bbox": [ + 30, + 136, + 380, + 254 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 30, + 136, + 380, + 254 + ], + "type": "text", + "content": " distinct yet semantic-related source domains represented as " + }, + { + "bbox": [ + 30, + 136, + 380, + 254 + ], + "type": "inline_equation", + "content": "D_{s} = \\{D_{s}^{1}, D_{s}^{2}, \\ldots, D_{s}^{K}\\}" + }, + { + "bbox": [ + 30, + 136, + 380, + 254 + ], + "type": "text", + "content": ", each originating from different distributions and comprising " + }, + { + "bbox": [ + 30, + 136, + 380, + 254 + ], + "type": "inline_equation", + "content": "N_{k}" + }, + { + "bbox": [ + 30, + 136, + 380, + 254 + ], + "type": "text", + "content": " image-label pairs, denoted as " + }, + { + "bbox": [ + 30, + 136, + 380, + 254 + ], + "type": "inline_equation", + "content": "D_{s}^{k} = \\{(x_{i}, y_{i})\\}_{i=1}^{N_{k}}" + }, + { + "bbox": [ + 30, + 136, + 380, + 254 + ], + "type": "text", + "content": ". In this study, the model training is independent of domain labels, so there is no domain subscript in " + }, + { + "bbox": [ + 30, + 136, + 380, + 254 + ], + "type": "inline_equation", + "content": "(x, y)" + }, + { + "bbox": [ + 30, + 136, + 380, + 254 + ], + "type": "text", + "content": ". The primary objective of DG is to leverage these source domains to train a model, parameterized by " + }, + { + "bbox": [ + 30, + 136, + 380, + 254 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 30, + 136, + 380, + 254 + ], + "type": "text", + "content": ", that exhibits strong generalization capabilities and can effectively adapt to unseen target domains. In RES, there are two models sharing the same architecture, a student model " + }, + { + "bbox": [ + 30, + 136, + 380, + 254 + ], + "type": "inline_equation", + "content": "M^{S}" + }, + { + "bbox": [ + 30, + 136, + 380, + 254 + ], + "type": "text", + "content": " parameterized with " + }, + { + "bbox": [ + 30, + 136, + 380, + 254 + ], + "type": "inline_equation", + "content": "\\theta^{S}" + }, + { + "bbox": [ + 30, + 136, + 380, + 254 + ], + "type": "text", + "content": " and a teacher model " + }, + { + "bbox": [ + 30, + 136, + 380, + 254 + ], + "type": "inline_equation", + "content": "M^{T}" + }, + { + "bbox": [ + 30, + 136, + 380, + 254 + ], + "type": "text", + "content": " parameterized with " + }, + { + "bbox": [ + 30, + 136, + 380, + 254 + ], + "type": "inline_equation", + "content": "\\theta^{T}" + }, + { + "bbox": [ + 30, + 136, + 380, + 254 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 30, + 136, + 380, + 254 + ], + "type": "inline_equation", + "content": "M^{S}" + }, + { + "bbox": [ + 30, + 136, + 380, + 254 + ], + "type": "text", + "content": " is used for model training and " + }, + { + "bbox": [ + 30, + 136, + 380, + 254 + ], + "type": "inline_equation", + "content": "M^{T}" + }, + { + "bbox": [ + 30, + 136, + 380, + 254 + ], + "type": "text", + "content": " is utilized for model validation and test." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 30, + 256, + 379, + 315 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 256, + 379, + 315 + ], + "spans": [ + { + "bbox": [ + 30, + 256, + 379, + 315 + ], + "type": "text", + "content": "The proposed RES-based DG framework is shown in Fig. 2 with its training process summarized in Algorithm 1. RES consists of: (1) a RE module that augments amplitude spectrum of feature maps while maintaining the phase spectrum, to expand cross-domain feature representation; and (2) a RS module that mutually fuses the parameters of the teacher and student models to each other." + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 34, + 337, + 379, + 501 + ], + "blocks": [ + { + "bbox": [ + 34, + 337, + 379, + 501 + ], + "lines": [ + { + "bbox": [ + 34, + 337, + 379, + 501 + ], + "spans": [ + { + "bbox": [ + 34, + 337, + 379, + 501 + ], + "type": "image", + "image_path": "5b1cd187f3bbe4313188096cb4604f8d9caeee0635f6c54a1e19f2fb70851cf7.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 30, + 514, + 381, + 581 + ], + "lines": [ + { + "bbox": [ + 30, + 514, + 381, + 581 + ], + "spans": [ + { + "bbox": [ + 30, + 514, + 381, + 581 + ], + "type": "text", + "content": "Fig. 2: Workflow of the proposed RES-based domain generalization framework, which consists of Representation Enhancement (RE) and Representation Stabilization (RS). RE aims to enhance feature representation to reduce cross-domain bias via a set of augmentations performing on the amplitude spectrum of frequency domain, while RS aims to further stabilize cross-domain performance to reduce cross-domain variance through a novel MEMA optimization strategy between student and teacher models." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 326, + 9, + 347, + 19 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 9, + 347, + 19 + ], + "spans": [ + { + "bbox": [ + 326, + 9, + 347, + 19 + ], + "type": "text", + "content": "RES" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 373, + 10, + 379, + 19 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 373, + 10, + 379, + 19 + ], + "spans": [ + { + "bbox": [ + 373, + 10, + 379, + 19 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 413, + 615 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 32, + 33, + 281, + 46 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 32, + 33, + 281, + 46 + ], + "spans": [ + { + "bbox": [ + 32, + 33, + 281, + 46 + ], + "type": "text", + "content": "3.2 Feature Frequency Augmentation-Based RE" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 32, + 57, + 380, + 105 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 32, + 57, + 380, + 105 + ], + "spans": [ + { + "bbox": [ + 32, + 57, + 380, + 105 + ], + "type": "text", + "content": "Our RE modules augments source data at the feature level. During training, the student model " + }, + { + "bbox": [ + 32, + 57, + 380, + 105 + ], + "type": "inline_equation", + "content": "M^S" + }, + { + "bbox": [ + 32, + 57, + 380, + 105 + ], + "type": "text", + "content": " is split into two parts: " + }, + { + "bbox": [ + 32, + 57, + 380, + 105 + ], + "type": "inline_equation", + "content": "M^{S-}" + }, + { + "bbox": [ + 32, + 57, + 380, + 105 + ], + "type": "text", + "content": " which includes all layers up to a middle layer " + }, + { + "bbox": [ + 32, + 57, + 380, + 105 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 32, + 57, + 380, + 105 + ], + "type": "text", + "content": " for feature map extraction, and " + }, + { + "bbox": [ + 32, + 57, + 380, + 105 + ], + "type": "inline_equation", + "content": "M^{S+}" + }, + { + "bbox": [ + 32, + 57, + 380, + 105 + ], + "type": "text", + "content": " which consists of the remaining layers for the remaining feature extraction and classification." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 32, + 126, + 380, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 32, + 126, + 380, + 186 + ], + "spans": [ + { + "bbox": [ + 32, + 126, + 380, + 186 + ], + "type": "text", + "content": "Decomposing Phase and Amplitude From Features. Given an image " + }, + { + "bbox": [ + 32, + 126, + 380, + 186 + ], + "type": "inline_equation", + "content": "x \\in \\mathcal{R}^{3 \\times \\bar{H} \\times \\bar{W}}" + }, + { + "bbox": [ + 32, + 126, + 380, + 186 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 32, + 126, + 380, + 186 + ], + "type": "inline_equation", + "content": "\\bar{H} \\times \\bar{W}" + }, + { + "bbox": [ + 32, + 126, + 380, + 186 + ], + "type": "text", + "content": " represents image spatial size, a corresponding multilayer feature map " + }, + { + "bbox": [ + 32, + 126, + 380, + 186 + ], + "type": "inline_equation", + "content": "f \\in \\mathcal{R}^{C \\times H \\times W}" + }, + { + "bbox": [ + 32, + 126, + 380, + 186 + ], + "type": "text", + "content": " can be extracted by " + }, + { + "bbox": [ + 32, + 126, + 380, + 186 + ], + "type": "inline_equation", + "content": "M^{S^{-}}" + }, + { + "bbox": [ + 32, + 126, + 380, + 186 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 32, + 126, + 380, + 186 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 32, + 126, + 380, + 186 + ], + "type": "text", + "content": " signifies the number of channels, and " + }, + { + "bbox": [ + 32, + 126, + 380, + 186 + ], + "type": "inline_equation", + "content": "H, W" + }, + { + "bbox": [ + 32, + 126, + 380, + 186 + ], + "type": "text", + "content": " denote the height and width of the feature maps respectively. It is formulated as" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 177, + 198, + 377, + 210 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 177, + 198, + 377, + 210 + ], + "spans": [ + { + "bbox": [ + 177, + 198, + 377, + 210 + ], + "type": "interline_equation", + "content": "f = M ^ {S -} (x). \\tag {3}", + "image_path": "b3dc9dd36602ea0f08c17a0120a45c2d60a9ddef1f7f04d94bbf40dd7f274d0c.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 32, + 221, + 379, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 32, + 221, + 379, + 281 + ], + "spans": [ + { + "bbox": [ + 32, + 221, + 379, + 281 + ], + "type": "text", + "content": "To delve into the frequency properties of these feature maps, we employ a 2D FFT across the spatial dimensions of each channel, ensuring that every layer of " + }, + { + "bbox": [ + 32, + 221, + 379, + 281 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 32, + 221, + 379, + 281 + ], + "type": "text", + "content": " undergoes this transformative process. This approach allows us to comprehensively analyze the frequency content within each individual feature channel. For channel " + }, + { + "bbox": [ + 32, + 221, + 379, + 281 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 32, + 221, + 379, + 281 + ], + "type": "text", + "content": ", its FFT formulation is as follows:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 110, + 293, + 377, + 324 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 110, + 293, + 377, + 324 + ], + "spans": [ + { + "bbox": [ + 110, + 293, + 377, + 324 + ], + "type": "interline_equation", + "content": "F ^ {c} (u, v) = \\sum_ {h = 0} ^ {H - 1} \\sum_ {w = 0} ^ {W - 1} f ^ {c} (h, w) \\cdot e ^ {- j 2 \\pi \\left(\\frac {u h}{H} + \\frac {v w}{W}\\right)}, \\tag {4}", + "image_path": "d64775a501c71268f6360503273154957fce73c28e8ee8d7a3854692577c5625.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 32, + 333, + 379, + 382 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 32, + 333, + 379, + 382 + ], + "spans": [ + { + "bbox": [ + 32, + 333, + 379, + 382 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 32, + 333, + 379, + 382 + ], + "type": "inline_equation", + "content": "f^{c}(h,w)" + }, + { + "bbox": [ + 32, + 333, + 379, + 382 + ], + "type": "text", + "content": " represents the pixel value at the spatial location " + }, + { + "bbox": [ + 32, + 333, + 379, + 382 + ], + "type": "inline_equation", + "content": "(h,w)" + }, + { + "bbox": [ + 32, + 333, + 379, + 382 + ], + "type": "text", + "content": " within the range of " + }, + { + "bbox": [ + 32, + 333, + 379, + 382 + ], + "type": "inline_equation", + "content": "[H - 1,W - 1]" + }, + { + "bbox": [ + 32, + 333, + 379, + 382 + ], + "type": "text", + "content": " in channel " + }, + { + "bbox": [ + 32, + 333, + 379, + 382 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 32, + 333, + 379, + 382 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 32, + 333, + 379, + 382 + ], + "type": "inline_equation", + "content": "F^{c}(u,v)" + }, + { + "bbox": [ + 32, + 333, + 379, + 382 + ], + "type": "text", + "content": " denotes the complex frequency spectrum at the frequency coordinates " + }, + { + "bbox": [ + 32, + 333, + 379, + 382 + ], + "type": "inline_equation", + "content": "(u,v)" + }, + { + "bbox": [ + 32, + 333, + 379, + 382 + ], + "type": "text", + "content": " of channel " + }, + { + "bbox": [ + 32, + 333, + 379, + 382 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 32, + 333, + 379, + 382 + ], + "type": "text", + "content": ". The term " + }, + { + "bbox": [ + 32, + 333, + 379, + 382 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 32, + 333, + 379, + 382 + ], + "type": "text", + "content": " is the imaginary unit." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 32, + 383, + 379, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 32, + 383, + 379, + 418 + ], + "spans": [ + { + "bbox": [ + 32, + 383, + 379, + 418 + ], + "type": "text", + "content": "Then we compute a multi-layer phase spectrum denoted as " + }, + { + "bbox": [ + 32, + 383, + 379, + 418 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 32, + 383, + 379, + 418 + ], + "type": "text", + "content": " and a multi-layer amplitude spectrum represented as " + }, + { + "bbox": [ + 32, + 383, + 379, + 418 + ], + "type": "inline_equation", + "content": "\\mathcal{A}" + }, + { + "bbox": [ + 32, + 383, + 379, + 418 + ], + "type": "text", + "content": ", from the complex frequency spectrum " + }, + { + "bbox": [ + 32, + 383, + 379, + 418 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 32, + 383, + 379, + 418 + ], + "type": "text", + "content": ", as" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 86, + 431, + 377, + 442 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 86, + 431, + 377, + 442 + ], + "spans": [ + { + "bbox": [ + 86, + 431, + 377, + 442 + ], + "type": "interline_equation", + "content": "\\mathcal {P} \\left(F ^ {c} (u, v)\\right) = \\arctan 2 \\left(\\operatorname {I m a g} \\left(F ^ {c} (u, v)\\right), \\operatorname {R e a l} \\left(F ^ {c} (u, v)\\right)\\right), \\tag {5}", + "image_path": "9d72a45f99bb6f0dc2686416bb93940e488b4f59eb26866a368117b229d3432d.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 94, + 468, + 377, + 480 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 94, + 468, + 377, + 480 + ], + "spans": [ + { + "bbox": [ + 94, + 468, + 377, + 480 + ], + "type": "interline_equation", + "content": "\\mathcal {A} \\left(F ^ {c} (u, v)\\right) = \\sqrt {\\operatorname {R e a l} \\left(F ^ {c} (u , v)\\right) ^ {2} + \\operatorname {I m a g} \\left(F ^ {c} (u , v)\\right) ^ {2}}. \\tag {6}", + "image_path": "1f098d6598b5d82a23e762deb0f3109f6b0ca3b5c9b961478498cc424b47df04.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 32, + 487, + 379, + 583 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 32, + 487, + 379, + 583 + ], + "spans": [ + { + "bbox": [ + 32, + 487, + 379, + 583 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 32, + 487, + 379, + 583 + ], + "type": "inline_equation", + "content": "\\operatorname{Real}(F^c(u,v))" + }, + { + "bbox": [ + 32, + 487, + 379, + 583 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 32, + 487, + 379, + 583 + ], + "type": "inline_equation", + "content": "\\operatorname{Imag}(F^c(u,v))" + }, + { + "bbox": [ + 32, + 487, + 379, + 583 + ], + "type": "text", + "content": " indicate the real and imaginary parts of the channel- " + }, + { + "bbox": [ + 32, + 487, + 379, + 583 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 32, + 487, + 379, + 583 + ], + "type": "text", + "content": " frequency spectrum " + }, + { + "bbox": [ + 32, + 487, + 379, + 583 + ], + "type": "inline_equation", + "content": "F^c" + }, + { + "bbox": [ + 32, + 487, + 379, + 583 + ], + "type": "text", + "content": ", respectively. " + }, + { + "bbox": [ + 32, + 487, + 379, + 583 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 32, + 487, + 379, + 583 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 32, + 487, + 379, + 583 + ], + "type": "inline_equation", + "content": "\\mathcal{A}" + }, + { + "bbox": [ + 32, + 487, + 379, + 583 + ], + "type": "text", + "content": " have the same size of " + }, + { + "bbox": [ + 32, + 487, + 379, + 583 + ], + "type": "inline_equation", + "content": "C \\times H \\times W" + }, + { + "bbox": [ + 32, + 487, + 379, + 583 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 32, + 487, + 379, + 583 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 32, + 487, + 379, + 583 + ], + "type": "text", + "content": ". Drawing inspiration from [15,29,32,41], there is a common assumption that the phase component of frequency spectrum maintains the high-level semantics of the original data while the amplitude component contains its low-level statistics. In this study, we extend it from the image level to feature level, where we regard feature maps as high-dimension signals with the same characteristic of phase-amplitude decomposition as images." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 32, + 10, + 39, + 19 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 32, + 10, + 39, + 19 + ], + "spans": [ + { + "bbox": [ + 32, + 10, + 39, + 19 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 65, + 9, + 162, + 20 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 9, + 162, + 20 + ], + "spans": [ + { + "bbox": [ + 65, + 9, + 162, + 20 + ], + "type": "text", + "content": "W. Huang, Y. Shi et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 413, + 615 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 30, + 35, + 380, + 71 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 35, + 380, + 71 + ], + "spans": [ + { + "bbox": [ + 30, + 35, + 380, + 71 + ], + "type": "text", + "content": "Frequency Augmentations on Features. To avoid damaging the semantics of features, augmentations are only implemented on the amplitude " + }, + { + "bbox": [ + 30, + 35, + 380, + 71 + ], + "type": "inline_equation", + "content": "\\mathcal{A}" + }, + { + "bbox": [ + 30, + 35, + 380, + 71 + ], + "type": "text", + "content": " to obtain an augmented " + }, + { + "bbox": [ + 30, + 35, + 380, + 71 + ], + "type": "inline_equation", + "content": "\\mathcal{A}'" + }, + { + "bbox": [ + 30, + 35, + 380, + 71 + ], + "type": "text", + "content": " while the phase " + }, + { + "bbox": [ + 30, + 35, + 380, + 71 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 30, + 35, + 380, + 71 + ], + "type": "text", + "content": " keeps unchanged, as" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 176, + 78, + 379, + 91 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 176, + 78, + 379, + 91 + ], + "spans": [ + { + "bbox": [ + 176, + 78, + 379, + 91 + ], + "type": "interline_equation", + "content": "\\mathcal {A} ^ {\\prime} = \\operatorname {A u g} (\\mathcal {A}). \\tag {7}", + "image_path": "c8da36017da7ab0ed94f6511a431a6c31ae3dc921bccc2f4ee26a01a3426952c.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 30, + 96, + 380, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 96, + 380, + 167 + ], + "spans": [ + { + "bbox": [ + 30, + 96, + 380, + 167 + ], + "type": "text", + "content": "Given a batch of images " + }, + { + "bbox": [ + 30, + 96, + 380, + 167 + ], + "type": "inline_equation", + "content": "\\{x_{1}, x_{2}, \\ldots, x_{B}\\}" + }, + { + "bbox": [ + 30, + 96, + 380, + 167 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 30, + 96, + 380, + 167 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 30, + 96, + 380, + 167 + ], + "type": "text", + "content": " is batch size, a corresponding batch of amplitudes " + }, + { + "bbox": [ + 30, + 96, + 380, + 167 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{A}_1, \\mathcal{A}_2, \\ldots, \\mathcal{A}_B\\}" + }, + { + "bbox": [ + 30, + 96, + 380, + 167 + ], + "type": "text", + "content": " can be obtained by Eqs. (3), (4), and (6). During each training iteration, we randomly select and apply one augmentation strategy from a pre-defined augmentation list, including none, random noise, random dropout, and mixup. Here none represents no augmentation, and the details of the rest three strategies are as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 37, + 172, + 381, + 234 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 37, + 172, + 381, + 234 + ], + "spans": [ + { + "bbox": [ + 37, + 172, + 381, + 234 + ], + "type": "text", + "content": "- Random noise, denoted as " + }, + { + "bbox": [ + 37, + 172, + 381, + 234 + ], + "type": "inline_equation", + "content": "\\mathrm{RE}^{ns}" + }, + { + "bbox": [ + 37, + 172, + 381, + 234 + ], + "type": "text", + "content": ". Given a multi-layer amplitude " + }, + { + "bbox": [ + 37, + 172, + 381, + 234 + ], + "type": "inline_equation", + "content": "\\mathcal{A}_i \\in \\mathcal{R}^{C \\times H \\times W}" + }, + { + "bbox": [ + 37, + 172, + 381, + 234 + ], + "type": "text", + "content": ", we add Gaussian noise " + }, + { + "bbox": [ + 37, + 172, + 381, + 234 + ], + "type": "inline_equation", + "content": "G \\in \\mathcal{R}^{1 \\times H \\times W}" + }, + { + "bbox": [ + 37, + 172, + 381, + 234 + ], + "type": "text", + "content": " to it. The elements of " + }, + { + "bbox": [ + 37, + 172, + 381, + 234 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 37, + 172, + 381, + 234 + ], + "type": "text", + "content": " adhere to a Gaussian distribution characterized by mean " + }, + { + "bbox": [ + 37, + 172, + 381, + 234 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 37, + 172, + 381, + 234 + ], + "type": "text", + "content": " and standard deviation " + }, + { + "bbox": [ + 37, + 172, + 381, + 234 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 37, + 172, + 381, + 234 + ], + "type": "text", + "content": ", i.e., " + }, + { + "bbox": [ + 37, + 172, + 381, + 234 + ], + "type": "inline_equation", + "content": "G \\sim \\mathcal{N}(\\mu, \\sigma^2)" + }, + { + "bbox": [ + 37, + 172, + 381, + 234 + ], + "type": "text", + "content": ". Leveraging the broadcasting mechanism, this noise addition is equally conducted to each layer of " + }, + { + "bbox": [ + 37, + 172, + 381, + 234 + ], + "type": "inline_equation", + "content": "\\mathcal{A}_i" + }, + { + "bbox": [ + 37, + 172, + 381, + 234 + ], + "type": "text", + "content": ", as:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 125, + 239, + 379, + 253 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 125, + 239, + 379, + 253 + ], + "spans": [ + { + "bbox": [ + 125, + 239, + 379, + 253 + ], + "type": "interline_equation", + "content": "\\mathcal {A} _ {i} ^ {\\prime} = \\mathcal {A} _ {i} * (I + G) = \\mathcal {A} _ {i} * (I + \\mathcal {N} (\\mu , \\sigma^ {2})), \\tag {8}", + "image_path": "2a071830f5401262cdea926e2f9c63e8cf2b3b8ce923ddf2690f9952e4d66a76.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 258, + 379, + 293 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 258, + 379, + 293 + ], + "spans": [ + { + "bbox": [ + 47, + 258, + 379, + 293 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 258, + 379, + 293 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 47, + 258, + 379, + 293 + ], + "type": "text", + "content": " represents the Identity Matrix. This approach guarantees that the same frequency components throughout all channels of the feature map are enhanced with the same noise intensity." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 37, + 293, + 381, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 37, + 293, + 381, + 354 + ], + "spans": [ + { + "bbox": [ + 37, + 293, + 381, + 354 + ], + "type": "text", + "content": "- Random dropout, denoted as " + }, + { + "bbox": [ + 37, + 293, + 381, + 354 + ], + "type": "inline_equation", + "content": "\\mathrm{RE}^{dp}" + }, + { + "bbox": [ + 37, + 293, + 381, + 354 + ], + "type": "text", + "content": ". We randomly generate a binary dropout mask, " + }, + { + "bbox": [ + 37, + 293, + 381, + 354 + ], + "type": "inline_equation", + "content": "M \\in \\mathcal{R}^{1 \\times H \\times W}" + }, + { + "bbox": [ + 37, + 293, + 381, + 354 + ], + "type": "text", + "content": ", where each element is set either to 0 with probability " + }, + { + "bbox": [ + 37, + 293, + 381, + 354 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 37, + 293, + 381, + 354 + ], + "type": "text", + "content": " or to 1 with a reversal probability " + }, + { + "bbox": [ + 37, + 293, + 381, + 354 + ], + "type": "inline_equation", + "content": "1 - p" + }, + { + "bbox": [ + 37, + 293, + 381, + 354 + ], + "type": "text", + "content": ". The dimension of " + }, + { + "bbox": [ + 37, + 293, + 381, + 354 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 37, + 293, + 381, + 354 + ], + "type": "text", + "content": " is also expanded to " + }, + { + "bbox": [ + 37, + 293, + 381, + 354 + ], + "type": "inline_equation", + "content": "C \\times H \\times W" + }, + { + "bbox": [ + 37, + 293, + 381, + 354 + ], + "type": "text", + "content": " by broadcasting. The dropout-augmented amplitude " + }, + { + "bbox": [ + 37, + 293, + 381, + 354 + ], + "type": "inline_equation", + "content": "\\mathcal{A}'" + }, + { + "bbox": [ + 37, + 293, + 381, + 354 + ], + "type": "text", + "content": " is calculated by:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 187, + 360, + 379, + 373 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 187, + 360, + 379, + 373 + ], + "spans": [ + { + "bbox": [ + 187, + 360, + 379, + 373 + ], + "type": "interline_equation", + "content": "\\mathcal {A} _ {i} ^ {\\prime} = \\mathcal {A} \\cdot M. \\tag {9}", + "image_path": "905124054a8e4d7820838ca5c2caa0a2956eaed8588f36a79cb4ed55837cbecc.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 379, + 381, + 427 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 379, + 381, + 427 + ], + "spans": [ + { + "bbox": [ + 47, + 379, + 381, + 427 + ], + "type": "text", + "content": "This operation introduces sparsity of amplitudes, aiming at reducing their overfitting for better robustness. It is worth mentioning that the element of " + }, + { + "bbox": [ + 47, + 379, + 381, + 427 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 47, + 379, + 381, + 427 + ], + "type": "text", + "content": " at the location " + }, + { + "bbox": [ + 47, + 379, + 381, + 427 + ], + "type": "inline_equation", + "content": "[0,0]" + }, + { + "bbox": [ + 47, + 379, + 381, + 427 + ], + "type": "text", + "content": " always equals 1, which represents the mean value of the feature map, to avoid over-masking on features." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 37, + 426, + 384, + 511 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 37, + 426, + 384, + 511 + ], + "spans": [ + { + "bbox": [ + 37, + 426, + 384, + 511 + ], + "type": "text", + "content": "- Mixup, denoted as " + }, + { + "bbox": [ + 37, + 426, + 384, + 511 + ], + "type": "inline_equation", + "content": "\\mathrm{RE}^{mix}" + }, + { + "bbox": [ + 37, + 426, + 384, + 511 + ], + "type": "text", + "content": ". Unlike random noise and random dropout which operate on single amplitudes extracted from individual samples, mixup engages in a cross-sample amplitude interaction by linear interpolation on different amplitudes extracted from different samples. Specifically, for a given amplitude spectrum " + }, + { + "bbox": [ + 37, + 426, + 384, + 511 + ], + "type": "inline_equation", + "content": "\\mathcal{A}_i" + }, + { + "bbox": [ + 37, + 426, + 384, + 511 + ], + "type": "text", + "content": ", mixup randomly selects another amplitude " + }, + { + "bbox": [ + 37, + 426, + 384, + 511 + ], + "type": "inline_equation", + "content": "\\mathcal{A}_j" + }, + { + "bbox": [ + 37, + 426, + 384, + 511 + ], + "type": "text", + "content": " from the set " + }, + { + "bbox": [ + 37, + 426, + 384, + 511 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{A}_1, \\mathcal{A}_2, \\dots, \\mathcal{A}_B\\}" + }, + { + "bbox": [ + 37, + 426, + 384, + 511 + ], + "type": "text", + "content": ", and then blends " + }, + { + "bbox": [ + 37, + 426, + 384, + 511 + ], + "type": "inline_equation", + "content": "\\mathcal{A}_i" + }, + { + "bbox": [ + 37, + 426, + 384, + 511 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 37, + 426, + 384, + 511 + ], + "type": "inline_equation", + "content": "\\mathcal{A}_j" + }, + { + "bbox": [ + 37, + 426, + 384, + 511 + ], + "type": "text", + "content": ", resulting in a new, hybrid amplitude spectrum " + }, + { + "bbox": [ + 37, + 426, + 384, + 511 + ], + "type": "inline_equation", + "content": "\\mathcal{A}_i'" + }, + { + "bbox": [ + 37, + 426, + 384, + 511 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 164, + 517, + 379, + 530 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 164, + 517, + 379, + 530 + ], + "spans": [ + { + "bbox": [ + 164, + 517, + 379, + 530 + ], + "type": "interline_equation", + "content": "\\mathcal {A} _ {i} ^ {\\prime} = \\alpha \\mathcal {A} _ {i} + (1 - \\alpha) \\mathcal {A} _ {j}, \\tag {10}", + "image_path": "074221983697b45824e9cf2dffce628dba5ab92bf7d65e634d352d249b20f2aa.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 47, + 535, + 381, + 584 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 535, + 381, + 584 + ], + "spans": [ + { + "bbox": [ + 47, + 535, + 381, + 584 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 535, + 381, + 584 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 47, + 535, + 381, + 584 + ], + "type": "text", + "content": " is a uniform random coefficient between [0,1]. This procedure enriches the amplitude characteristics by promoting a more diverse fusion of feature styles across domains. It boosts the model's generalization by augmenting intricate cross-domain amplitude patterns effectively." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 326, + 9, + 347, + 19 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 9, + 347, + 19 + ], + "spans": [ + { + "bbox": [ + 326, + 9, + 347, + 19 + ], + "type": "text", + "content": "RES" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 373, + 10, + 380, + 19 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 373, + 10, + 380, + 19 + ], + "spans": [ + { + "bbox": [ + 373, + 10, + 380, + 19 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 413, + 615 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 30, + 35, + 380, + 71 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 35, + 380, + 71 + ], + "spans": [ + { + "bbox": [ + 30, + 35, + 380, + 71 + ], + "type": "text", + "content": "Reconstructing Features. For each pair of phase and augmented amplitude components, " + }, + { + "bbox": [ + 30, + 35, + 380, + 71 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 30, + 35, + 380, + 71 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 30, + 35, + 380, + 71 + ], + "type": "inline_equation", + "content": "\\mathcal{A}'" + }, + { + "bbox": [ + 30, + 35, + 380, + 71 + ], + "type": "text", + "content": ", an augmented multi-layer feature map, " + }, + { + "bbox": [ + 30, + 35, + 380, + 71 + ], + "type": "inline_equation", + "content": "f'" + }, + { + "bbox": [ + 30, + 35, + 380, + 71 + ], + "type": "text", + "content": ", can be reconstructed via inverse Fast Fourier Transform (iFFT) as" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 91, + 379, + 123 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 379, + 123 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 379, + 123 + ], + "type": "interline_equation", + "content": "f ^ {\\prime c} (h, w) = \\frac {1}{H W} \\sum_ {u = 0} ^ {H - 1} \\sum_ {v = 0} ^ {W - 1} \\mathcal {A} ^ {\\prime} \\left(F ^ {c} (u, v)\\right) \\cdot e ^ {j 2 \\pi \\left(\\frac {u h}{H} + \\frac {v w}{W}\\right)} \\cdot e ^ {j \\mathcal {P} \\left(F ^ {c} (u, v)\\right)} \\tag {11}", + "image_path": "bfafc7b4dd7368b66146cd4e4d87c4c7eb288369efc79b59d966fd08825e966c.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 30, + 134, + 379, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 134, + 379, + 171 + ], + "spans": [ + { + "bbox": [ + 30, + 134, + 379, + 171 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 30, + 134, + 379, + 171 + ], + "type": "inline_equation", + "content": "e^{j2\\pi \\left(\\frac{uh}{H} + \\frac{vw}{W}\\right)}" + }, + { + "bbox": [ + 30, + 134, + 379, + 171 + ], + "type": "text", + "content": " is a complex exponential term representing the frequency component in the inverse Fourier transform, and " + }, + { + "bbox": [ + 30, + 134, + 379, + 171 + ], + "type": "inline_equation", + "content": "e^{j\\mathcal{P}(F^c(u,v))}" + }, + { + "bbox": [ + 30, + 134, + 379, + 171 + ], + "type": "text", + "content": " reintroduces the phase information into the transform." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 30, + 172, + 379, + 196 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 172, + 379, + 196 + ], + "spans": [ + { + "bbox": [ + 30, + 172, + 379, + 196 + ], + "type": "text", + "content": "Finally, " + }, + { + "bbox": [ + 30, + 172, + 379, + 196 + ], + "type": "inline_equation", + "content": "f'" + }, + { + "bbox": [ + 30, + 172, + 379, + 196 + ], + "type": "text", + "content": " is used for further feature extraction and class classification by the remaining layers of the student model, " + }, + { + "bbox": [ + 30, + 172, + 379, + 196 + ], + "type": "inline_equation", + "content": "M^{S+}" + }, + { + "bbox": [ + 30, + 172, + 379, + 196 + ], + "type": "text", + "content": ", as" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 175, + 206, + 379, + 220 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 175, + 206, + 379, + 220 + ], + "spans": [ + { + "bbox": [ + 175, + 206, + 379, + 220 + ], + "type": "interline_equation", + "content": "p = M ^ {S +} \\left(f ^ {\\prime}\\right). \\tag {12}", + "image_path": "6bcbe1390e468e5cacbd3228233a15eaf2acdcb0ac2bf0b15b723272dc4ccdaf.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 31, + 238, + 149, + 249 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 31, + 238, + 149, + 249 + ], + "spans": [ + { + "bbox": [ + 31, + 238, + 149, + 249 + ], + "type": "text", + "content": "3.3 MEMA-based RS" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 30, + 259, + 379, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 259, + 379, + 319 + ], + "spans": [ + { + "bbox": [ + 30, + 259, + 379, + 319 + ], + "type": "text", + "content": "As shown in Fig. 2, there are two models, one student model " + }, + { + "bbox": [ + 30, + 259, + 379, + 319 + ], + "type": "inline_equation", + "content": "M^S" + }, + { + "bbox": [ + 30, + 259, + 379, + 319 + ], + "type": "text", + "content": " with parameters " + }, + { + "bbox": [ + 30, + 259, + 379, + 319 + ], + "type": "inline_equation", + "content": "\\theta^S" + }, + { + "bbox": [ + 30, + 259, + 379, + 319 + ], + "type": "text", + "content": " and one teacher model " + }, + { + "bbox": [ + 30, + 259, + 379, + 319 + ], + "type": "inline_equation", + "content": "M^T" + }, + { + "bbox": [ + 30, + 259, + 379, + 319 + ], + "type": "text", + "content": " with parameters " + }, + { + "bbox": [ + 30, + 259, + 379, + 319 + ], + "type": "inline_equation", + "content": "\\theta^T" + }, + { + "bbox": [ + 30, + 259, + 379, + 319 + ], + "type": "text", + "content": ". Their parameters undergo mutual updates to provide stabilization to each other, referred to as " + }, + { + "bbox": [ + 30, + 259, + 379, + 319 + ], + "type": "inline_equation", + "content": "\\mathrm{RE}^T" + }, + { + "bbox": [ + 30, + 259, + 379, + 319 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 30, + 259, + 379, + 319 + ], + "type": "inline_equation", + "content": "\\mathrm{RE}^S" + }, + { + "bbox": [ + 30, + 259, + 379, + 319 + ], + "type": "text", + "content": " for the teacher and student models, respectively. Their processes are formulated as" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 120, + 327, + 377, + 356 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 120, + 327, + 377, + 356 + ], + "spans": [ + { + "bbox": [ + 120, + 327, + 377, + 356 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\theta^ {T} \\leftarrow \\tau^ {T} \\theta^ {T} + (1 - \\tau^ {T}) \\theta^ {S}, \\quad \\tau^ {T} \\in [ 0, 1), \\tag {13} \\\\ \\theta^ {S} \\leftarrow (1 - \\tau^ {S}) \\theta^ {S} + \\tau^ {S} \\theta^ {T}, \\quad \\tau^ {S} \\in [ 0, 1). \\\\ \\end{array}", + "image_path": "8536c372e49182d0a868b0bed4ef36f90965716ef4da58f41615bd17c7643a43.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 30, + 367, + 379, + 439 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 367, + 379, + 439 + ], + "spans": [ + { + "bbox": [ + 30, + 367, + 379, + 439 + ], + "type": "text", + "content": "Here, only the student model " + }, + { + "bbox": [ + 30, + 367, + 379, + 439 + ], + "type": "inline_equation", + "content": "M^S" + }, + { + "bbox": [ + 30, + 367, + 379, + 439 + ], + "type": "text", + "content": " is actively trained, while the teacher model " + }, + { + "bbox": [ + 30, + 367, + 379, + 439 + ], + "type": "inline_equation", + "content": "M^T" + }, + { + "bbox": [ + 30, + 367, + 379, + 439 + ], + "type": "text", + "content": " is responsible solely for parameter updates. The student model " + }, + { + "bbox": [ + 30, + 367, + 379, + 439 + ], + "type": "inline_equation", + "content": "M^S" + }, + { + "bbox": [ + 30, + 367, + 379, + 439 + ], + "type": "text", + "content": " is the primary learner, actively absorbing new knowledge and adapting to training data. In contrast, the teacher model " + }, + { + "bbox": [ + 30, + 367, + 379, + 439 + ], + "type": "inline_equation", + "content": "M^T" + }, + { + "bbox": [ + 30, + 367, + 379, + 439 + ], + "type": "text", + "content": " acts as a robust and stable reference using moving average updating strategy, providing more robust and consistent parameters by reducing the variance of training batches [2]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 30, + 439, + 379, + 535 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 439, + 379, + 535 + ], + "spans": [ + { + "bbox": [ + 30, + 439, + 379, + 535 + ], + "type": "text", + "content": "Both the values of " + }, + { + "bbox": [ + 30, + 439, + 379, + 535 + ], + "type": "inline_equation", + "content": "\\tau^T" + }, + { + "bbox": [ + 30, + 439, + 379, + 535 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 30, + 439, + 379, + 535 + ], + "type": "inline_equation", + "content": "\\tau^S" + }, + { + "bbox": [ + 30, + 439, + 379, + 535 + ], + "type": "text", + "content": " are small and they play a mutual role in this process: (1) Stability from the teacher model. With a small " + }, + { + "bbox": [ + 30, + 439, + 379, + 535 + ], + "type": "inline_equation", + "content": "\\tau^T" + }, + { + "bbox": [ + 30, + 439, + 379, + 535 + ], + "type": "text", + "content": ", the teacher model's parameters change slowly, ensuring that they remain stable and consistent.(2) Gradual integration for the student model. A small " + }, + { + "bbox": [ + 30, + 439, + 379, + 535 + ], + "type": "inline_equation", + "content": "\\tau^S" + }, + { + "bbox": [ + 30, + 439, + 379, + 535 + ], + "type": "text", + "content": " implies that the student model integrates the teacher's parameters slowly and steadily. This allows the student model to gradually assimilate the stable characteristics of the teacher model. To simplify the hyperparameter tuning process in experiments, " + }, + { + "bbox": [ + 30, + 439, + 379, + 535 + ], + "type": "inline_equation", + "content": "\\tau^T" + }, + { + "bbox": [ + 30, + 439, + 379, + 535 + ], + "type": "text", + "content": " is set to a small constant value of 1e-3, while " + }, + { + "bbox": [ + 30, + 439, + 379, + 535 + ], + "type": "inline_equation", + "content": "\\tau^S" + }, + { + "bbox": [ + 30, + 439, + 379, + 535 + ], + "type": "text", + "content": " is variable for different datasets." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 30, + 536, + 379, + 583 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 536, + 379, + 583 + ], + "spans": [ + { + "bbox": [ + 30, + 536, + 379, + 583 + ], + "type": "text", + "content": "In summary, the student model acts as an active learner, continually evolving and assimilating new information, whereas the teacher model functions as a reliable guide, providing a stable parameter reference to shield the student model from the effects of cross-domain variance in training batches." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 32, + 10, + 39, + 18 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 32, + 10, + 39, + 18 + ], + "spans": [ + { + "bbox": [ + 32, + 10, + 39, + 18 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 65, + 9, + 162, + 20 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 9, + 162, + 20 + ], + "spans": [ + { + "bbox": [ + 65, + 9, + 162, + 20 + ], + "type": "text", + "content": "W. Huang, Y. Shi et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 413, + 615 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 32, + 33, + 379, + 218 + ], + "blocks": [ + { + "bbox": [ + 32, + 33, + 379, + 218 + ], + "lines": [ + { + "bbox": [ + 32, + 33, + 379, + 218 + ], + "spans": [ + { + "bbox": [ + 32, + 33, + 379, + 218 + ], + "type": "table", + "html": "
Algorithm 1 Training Process of our RES-based DG
1:Student model MS with initialized parameters θS, teacher model MT with initialized parameters θT, all training sample pairs {(xi,yi)}N, and batch size B
2:for iter = 1 to Niter do
3:Sample Batch Data: {(xi,yi)}B
4:Extract Feature: f← MS-(x)
5:Decompose Phase and Amplitude: P, A← FFT(f)
6:Augmentation Selection: Aug ← [random noise, random dropout, mixup, none]
7:Augment Amplitude: A' ← Aug(A)
8:Reconstruct feature: f' = iFFT(P, A')
9:Compute prediction: p ← MS+(f')
10:Compute Loss: E(θS) ← l(p, y)
11:Update Student Model with Loss: θS ← θS - η∇θL(θS)
12:Integrate Student Model to Teacher Model: θT = τTθT + (1 - τT)θS
13:Stabilize Student Model with Teacher Model: θS = (1 - τS)θS + τSθT
14:end for
", + "image_path": "be812eeca85e4691a769c0d932d077986d2b813362b7fd860d5b88d1b4c3c762.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 31, + 240, + 144, + 252 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 31, + 240, + 144, + 252 + ], + "spans": [ + { + "bbox": [ + 31, + 240, + 144, + 252 + ], + "type": "text", + "content": "3.4 Supervision Loss" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 30, + 260, + 381, + 296 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 260, + 381, + 296 + ], + "spans": [ + { + "bbox": [ + 30, + 260, + 381, + 296 + ], + "type": "text", + "content": "During each training iteration with a batch size of " + }, + { + "bbox": [ + 30, + 260, + 381, + 296 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 30, + 260, + 381, + 296 + ], + "type": "text", + "content": ", the empirical risk minimization (ERM) loss function is used to optimize the student model " + }, + { + "bbox": [ + 30, + 260, + 381, + 296 + ], + "type": "inline_equation", + "content": "M^S" + }, + { + "bbox": [ + 30, + 260, + 381, + 296 + ], + "type": "text", + "content": " across all the training domains as" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 95, + 305, + 379, + 338 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 95, + 305, + 379, + 338 + ], + "spans": [ + { + "bbox": [ + 95, + 305, + 379, + 338 + ], + "type": "interline_equation", + "content": "\\mathcal {E} \\left(\\theta^ {S}\\right) = \\frac {1}{B} \\sum_ {i = 1} ^ {B} \\ell \\left(p _ {i}, y _ {i}\\right) = \\frac {1}{B} \\sum_ {i = 1} ^ {B} \\ell \\left(M ^ {S} \\left(x _ {i}; \\theta^ {S}\\right), y _ {i}\\right), \\tag {14}", + "image_path": "3707ff8837f8f118f652f06b49772ed155d64b95cd8cb88b55ddf113a73077fc.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 31, + 346, + 379, + 372 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 31, + 346, + 379, + 372 + ], + "spans": [ + { + "bbox": [ + 31, + 346, + 379, + 372 + ], + "type": "text", + "content": "where the detailed progress of " + }, + { + "bbox": [ + 31, + 346, + 379, + 372 + ], + "type": "inline_equation", + "content": "M^S (x_i;\\theta^S)" + }, + { + "bbox": [ + 31, + 346, + 379, + 372 + ], + "type": "text", + "content": " come from Eqs. (3)-(7) and (12) and " + }, + { + "bbox": [ + 31, + 346, + 379, + 372 + ], + "type": "inline_equation", + "content": "\\ell" + }, + { + "bbox": [ + 31, + 346, + 379, + 372 + ], + "type": "text", + "content": " represents the CE loss. Here " + }, + { + "bbox": [ + 31, + 346, + 379, + 372 + ], + "type": "inline_equation", + "content": "\\mathcal{E}(\\theta^S)" + }, + { + "bbox": [ + 31, + 346, + 379, + 372 + ], + "type": "text", + "content": " represents the basic ERM loss." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 31, + 389, + 129, + 403 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 31, + 389, + 129, + 403 + ], + "spans": [ + { + "bbox": [ + 31, + 389, + 129, + 403 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 30, + 414, + 380, + 511 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 414, + 380, + 511 + ], + "spans": [ + { + "bbox": [ + 30, + 414, + 380, + 511 + ], + "type": "text", + "content": "This section outlines benchmark datasets and experimental settings, conducts an ablation study on RES, compares it with advanced DG methods, examines the impact of " + }, + { + "bbox": [ + 30, + 414, + 380, + 511 + ], + "type": "inline_equation", + "content": "\\tau^S" + }, + { + "bbox": [ + 30, + 414, + 380, + 511 + ], + "type": "text", + "content": ", and quantitatively assesses RES's effects in reducing cross-domain bias and variance and improving accuracy across domains. In Supplementary Material, we compare the effects of data augmentation from two perspectives: image-level vs. feature-level frequency augmentation, and normal feature augmentation vs. feature frequency augmentation. This comparison demonstrates the superiority of our feature frequency augmentation." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 31, + 528, + 164, + 539 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 31, + 528, + 164, + 539 + ], + "spans": [ + { + "bbox": [ + 31, + 528, + 164, + 539 + ], + "type": "text", + "content": "4.1 Benchmark Datasets" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 30, + 547, + 380, + 584 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 547, + 380, + 584 + ], + "spans": [ + { + "bbox": [ + 30, + 547, + 380, + 584 + ], + "type": "text", + "content": "We evaluate the proposed RES method on five DG datasets: (1) PACS [22], which contains 9,991 images in 7 object categories across 4 diverse domains; (2) VLCS [10], which comprises 10,729 examples from 5 categories, collected from 4" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 326, + 9, + 347, + 19 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 9, + 347, + 19 + ], + "spans": [ + { + "bbox": [ + 326, + 9, + 347, + 19 + ], + "type": "text", + "content": "RES" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 372, + 10, + 379, + 19 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 372, + 10, + 379, + 19 + ], + "spans": [ + { + "bbox": [ + 372, + 10, + 379, + 19 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 413, + 615 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 32, + 33, + 379, + 82 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 32, + 33, + 379, + 82 + ], + "spans": [ + { + "bbox": [ + 32, + 33, + 379, + 82 + ], + "type": "text", + "content": "domains; (3) OfficeHome [36], which encompasses around 15,500 samples in 65 categories from four domains. (4) TerraIncognita [3], which contains 24,788 images within 10 categories from four domains; and (5) DomainNet [30], a vast dataset with 586,575 images across 345 categories, spanning 6 domains." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 33, + 99, + 171, + 111 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 99, + 171, + 111 + ], + "spans": [ + { + "bbox": [ + 33, + 99, + 171, + 111 + ], + "type": "text", + "content": "4.2 Experimental Settings" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 33, + 118, + 379, + 215 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 118, + 379, + 215 + ], + "spans": [ + { + "bbox": [ + 33, + 118, + 379, + 215 + ], + "type": "text", + "content": "Basic settings. We utilize ResNet-50 [16], pretrained on ImageNet-1K [9], as the backbone. Our codes are constructed on Domainbed's framework [13], allocating " + }, + { + "bbox": [ + 33, + 118, + 379, + 215 + ], + "type": "inline_equation", + "content": "80\\%" + }, + { + "bbox": [ + 33, + 118, + 379, + 215 + ], + "type": "text", + "content": " of in-domain data for training and the remaining " + }, + { + "bbox": [ + 33, + 118, + 379, + 215 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 33, + 118, + 379, + 215 + ], + "type": "text", + "content": " for validation, with test performed on an unseen domain. PACS, VLCS, and OfficeHome are trained for 10K iterations and validated every 300 iterations, TerraIncognita is trained for 3K iterations and validated every 300 iterations, and DomainNet undergoes 400K iterations of training and 1K iterations of validation. Adam is used as the optimizer, with a batch size of 128 and a learning rate of 5e-5." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 33, + 231, + 379, + 328 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 231, + 379, + 328 + ], + "spans": [ + { + "bbox": [ + 33, + 231, + 379, + 328 + ], + "type": "text", + "content": "Settings of RES. In every training iteration, we apply the frequency augmentation of RE randomly to one of the four potential middle-layer positions within ResNet-50. These positions are strategically located at the ends of the " + }, + { + "bbox": [ + 33, + 231, + 379, + 328 + ], + "type": "inline_equation", + "content": "1^{st}" + }, + { + "bbox": [ + 33, + 231, + 379, + 328 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 33, + 231, + 379, + 328 + ], + "type": "inline_equation", + "content": "2^{nd}" + }, + { + "bbox": [ + 33, + 231, + 379, + 328 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 33, + 231, + 379, + 328 + ], + "type": "inline_equation", + "content": "3^{rd}" + }, + { + "bbox": [ + 33, + 231, + 379, + 328 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 33, + 231, + 379, + 328 + ], + "type": "inline_equation", + "content": "4^{th}" + }, + { + "bbox": [ + 33, + 231, + 379, + 328 + ], + "type": "text", + "content": " residual blocks, ensuring diverse augmentation effects across different layers of the model. There are some hyperparameters in RES summarized in Table 1, where " + }, + { + "bbox": [ + 33, + 231, + 379, + 328 + ], + "type": "inline_equation", + "content": "\\tau^S" + }, + { + "bbox": [ + 33, + 231, + 379, + 328 + ], + "type": "text", + "content": " is variable and tuned based on the validation performance as discussed in Section 4.4. All the experiments of RES are conducted for three times and the mean values and standard deviations are reported." + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 42, + 370, + 366, + 394 + ], + "blocks": [ + { + "bbox": [ + 135, + 348, + 276, + 360 + ], + "lines": [ + { + "bbox": [ + 135, + 348, + 276, + 360 + ], + "spans": [ + { + "bbox": [ + 135, + 348, + 276, + 360 + ], + "type": "text", + "content": "Table 1: Hyperparameters of RES." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 42, + 370, + 366, + 394 + ], + "lines": [ + { + "bbox": [ + 42, + 370, + 366, + 394 + ], + "spans": [ + { + "bbox": [ + 42, + 370, + 366, + 394 + ], + "type": "table", + "html": "
REnsREdpREmixRSS→TRST→S
G ~ N(0.75, 0.752)p = 0.5α ~ U(0,1)τT = 0.001τS is variable
", + "image_path": "3a525fc0b294942b7369d68d06ba65b4b7d7c25da1b49d5d3eb558725375ce36.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 33, + 443, + 136, + 456 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 443, + 136, + 456 + ], + "spans": [ + { + "bbox": [ + 33, + 443, + 136, + 456 + ], + "type": "text", + "content": "4.3 Ablation Study" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 33, + 463, + 379, + 523 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 463, + 379, + 523 + ], + "spans": [ + { + "bbox": [ + 33, + 463, + 379, + 523 + ], + "type": "text", + "content": "The ablation study of RES's modules, including sub-strategies of RE and RS, on PACS are provided in Table 2 and Fig. 3. RE employs random noise " + }, + { + "bbox": [ + 33, + 463, + 379, + 523 + ], + "type": "inline_equation", + "content": "(\\mathrm{RE}^{ns})" + }, + { + "bbox": [ + 33, + 463, + 379, + 523 + ], + "type": "text", + "content": ", random dropout " + }, + { + "bbox": [ + 33, + 463, + 379, + 523 + ], + "type": "inline_equation", + "content": "(\\mathrm{RE}^{dp})" + }, + { + "bbox": [ + 33, + 463, + 379, + 523 + ], + "type": "text", + "content": ", and mixup " + }, + { + "bbox": [ + 33, + 463, + 379, + 523 + ], + "type": "inline_equation", + "content": "(\\mathrm{RE}^{mix})" + }, + { + "bbox": [ + 33, + 463, + 379, + 523 + ], + "type": "text", + "content": " techniques for feature augmentation, and RS employs MEMA containing representation stabilization for both student models " + }, + { + "bbox": [ + 33, + 463, + 379, + 523 + ], + "type": "inline_equation", + "content": "(\\mathrm{RS}^s)" + }, + { + "bbox": [ + 33, + 463, + 379, + 523 + ], + "type": "text", + "content": " and teacher models " + }, + { + "bbox": [ + 33, + 463, + 379, + 523 + ], + "type": "inline_equation", + "content": "(\\mathrm{RS}^t)" + }, + { + "bbox": [ + 33, + 463, + 379, + 523 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 33, + 523, + 379, + 583 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 523, + 379, + 583 + ], + "spans": [ + { + "bbox": [ + 33, + 523, + 379, + 583 + ], + "type": "text", + "content": "After applying all three RE strategies, there is significant performance improvement from ERM's 85.5 to (ERM+RE)'s 88.8 as shown in Table 2. with their training trajectory shown in Fig. 3. The results verify the effect of RE in reducing cross-domain bias. Additionally, the application of RS strategies reduces their standard deviations from ERM's 1.1 to (ERM+RS)'s 0.5, demonstrating their" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 33, + 10, + 43, + 19 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 10, + 43, + 19 + ], + "spans": [ + { + "bbox": [ + 33, + 10, + 43, + 19 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 65, + 9, + 162, + 20 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 9, + 162, + 20 + ], + "spans": [ + { + "bbox": [ + 65, + 9, + 162, + 20 + ], + "type": "text", + "content": "W. Huang, Y. Shi et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 413, + 615 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 30, + 35, + 380, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 35, + 380, + 83 + ], + "spans": [ + { + "bbox": [ + 30, + 35, + 380, + 83 + ], + "type": "text", + "content": "effectiveness in reducing cross-domain variance. When RE and RS strategies are applied in conjunction, as shown in the last highlighted row of the table, the model achieves optimal test performance, with an average accuracy increase to 90.0 with the smallest standard deviation of 0.3, significantly surpassing ERM." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 30, + 84, + 379, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 84, + 379, + 120 + ], + "spans": [ + { + "bbox": [ + 30, + 84, + 379, + 120 + ], + "type": "text", + "content": "Through this ablation study, the complementary effects of RE and RS strategies in enhancing model generalization across unseen domains are clearly demonstrated, underscoring the significance of reducing cross-domain bias-variance." + } + ] + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 32, + 173, + 380, + 285 + ], + "blocks": [ + { + "bbox": [ + 89, + 144, + 322, + 155 + ], + "lines": [ + { + "bbox": [ + 89, + 144, + 322, + 155 + ], + "spans": [ + { + "bbox": [ + 89, + 144, + 322, + 155 + ], + "type": "text", + "content": "Table 2: Ablation study of the proposed RES on PACS." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 32, + 173, + 380, + 285 + ], + "lines": [ + { + "bbox": [ + 32, + 173, + 380, + 285 + ], + "spans": [ + { + "bbox": [ + 32, + 173, + 380, + 285 + ], + "type": "table", + "html": "
ERMREnsREdpREmixRStRSsPACSAvg
ArtClipartPaintingSketch
✓84.4±0.681.1±0.796.3±0.580.2±2.585.5±1.1
✓✓88.8±1.282.2±0.597.0±0.681.2±0.687.3±0.7
✓✓87.6±0.382.7±1.096.8±0.882.1±1.687.3±0.9
✓✓89.4±0.281.4±1.196.6±0.485.1±0.988.5±0.7
✓✓✓✓89.0±0.983.0±0.497.0±0.485.1±0.788.8±0.6
✓✓87.9±0.183.1±0.296.0±0.482.2±1.287.3±0.5
✓✓✓90.9±0.482.0±0.597.9±0.382.2±0.888.2±0.5
✓✓✓✓✓✓91.6±0.484.2±0.398.1±0.186.1±0.490.0±0.3
", + "image_path": "ea1f6c289a81a75e0e8b6d5407df0e263855d07301679d41534e096aa44e8a0f.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 33, + 342, + 120, + 410 + ], + "blocks": [ + { + "bbox": [ + 33, + 342, + 120, + 410 + ], + "lines": [ + { + "bbox": [ + 33, + 342, + 120, + 410 + ], + "spans": [ + { + "bbox": [ + 33, + 342, + 120, + 410 + ], + "type": "image", + "image_path": "08137a2f85d6c20c90f34abf5ff7bf3125ae05adadf886a224b498eba15e1bb0.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 87, + 420, + 323, + 431 + ], + "lines": [ + { + "bbox": [ + 87, + 420, + 323, + 431 + ], + "spans": [ + { + "bbox": [ + 87, + 420, + 323, + 431 + ], + "type": "text", + "content": "Fig. 3: Training convergence of RES's modules on PACS." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 121, + 342, + 206, + 410 + ], + "blocks": [ + { + "bbox": [ + 121, + 342, + 206, + 410 + ], + "lines": [ + { + "bbox": [ + 121, + 342, + 206, + 410 + ], + "spans": [ + { + "bbox": [ + 121, + 342, + 206, + 410 + ], + "type": "image", + "image_path": "3e7d83a1c426e5e53f56d75ba58d7fd4828ab466af565363c02db4e145af166a.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 207, + 342, + 292, + 410 + ], + "blocks": [ + { + "bbox": [ + 207, + 342, + 292, + 410 + ], + "lines": [ + { + "bbox": [ + 207, + 342, + 292, + 410 + ], + "spans": [ + { + "bbox": [ + 207, + 342, + 292, + 410 + ], + "type": "image", + "image_path": "d7dae5339d1af2a7e57ddcf594eb8fb4c21397d097ababc5751992cc946f6dd9.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 293, + 342, + 378, + 410 + ], + "blocks": [ + { + "bbox": [ + 293, + 342, + 378, + 410 + ], + "lines": [ + { + "bbox": [ + 293, + 342, + 378, + 410 + ], + "spans": [ + { + "bbox": [ + 293, + 342, + 378, + 410 + ], + "type": "image", + "image_path": "6530f02b979b7854895fed0920e83bed28e7ec39f543cea4f41de33364c2cb0d.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 31, + 482, + 178, + 495 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 31, + 482, + 178, + 495 + ], + "spans": [ + { + "bbox": [ + 31, + 482, + 178, + 495 + ], + "type": "text", + "content": "4.4 Values of " + }, + { + "bbox": [ + 31, + 482, + 178, + 495 + ], + "type": "inline_equation", + "content": "\\tau^S" + }, + { + "bbox": [ + 31, + 482, + 178, + 495 + ], + "type": "text", + "content": " in MEMA" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 30, + 511, + 380, + 584 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 511, + 380, + 584 + ], + "spans": [ + { + "bbox": [ + 30, + 511, + 380, + 584 + ], + "type": "text", + "content": "In RES, the MEMA's " + }, + { + "bbox": [ + 30, + 511, + 380, + 584 + ], + "type": "inline_equation", + "content": "\\tau^S" + }, + { + "bbox": [ + 30, + 511, + 380, + 584 + ], + "type": "text", + "content": " is the sole hyperparameter requiring adjustment across various benchmarks. We explored five distinct values: [0, 1e-3, 5e-3, 1e-2, 5e-2] for each benchmark, with results shown in Fig. 4. The results indicate that the optimal " + }, + { + "bbox": [ + 30, + 511, + 380, + 584 + ], + "type": "inline_equation", + "content": "\\tau^S" + }, + { + "bbox": [ + 30, + 511, + 380, + 584 + ], + "type": "text", + "content": " value varies by dataset, attributed to differences in their convergence speed. Subsequent comparison experiments will report out-of-domain accuracies for " + }, + { + "bbox": [ + 30, + 511, + 380, + 584 + ], + "type": "inline_equation", + "content": "\\tau^S" + }, + { + "bbox": [ + 30, + 511, + 380, + 584 + ], + "type": "text", + "content": " corresponding to the highest in-domain accuracies observed." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 326, + 9, + 347, + 19 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 9, + 347, + 19 + ], + "spans": [ + { + "bbox": [ + 326, + 9, + 347, + 19 + ], + "type": "text", + "content": "RES" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 369, + 10, + 379, + 19 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 369, + 10, + 379, + 19 + ], + "spans": [ + { + "bbox": [ + 369, + 10, + 379, + 19 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 413, + 615 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 33, + 32, + 101, + 98 + ], + "blocks": [ + { + "bbox": [ + 33, + 32, + 101, + 98 + ], + "lines": [ + { + "bbox": [ + 33, + 32, + 101, + 98 + ], + "spans": [ + { + "bbox": [ + 33, + 32, + 101, + 98 + ], + "type": "image", + "image_path": "0f20f1dd4586800170ead541a7e393ac55aa4fa61987021677557514a810dd7e.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 41, + 115, + 369, + 127 + ], + "lines": [ + { + "bbox": [ + 41, + 115, + 369, + 127 + ], + "spans": [ + { + "bbox": [ + 41, + 115, + 369, + 127 + ], + "type": "text", + "content": "Fig. 4: In-domain (validation) and out-of-domain (test) accuracies vary with " + }, + { + "bbox": [ + 41, + 115, + 369, + 127 + ], + "type": "inline_equation", + "content": "\\tau^S" + }, + { + "bbox": [ + 41, + 115, + 369, + 127 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 102, + 32, + 170, + 98 + ], + "blocks": [ + { + "bbox": [ + 102, + 32, + 170, + 98 + ], + "lines": [ + { + "bbox": [ + 102, + 32, + 170, + 98 + ], + "spans": [ + { + "bbox": [ + 102, + 32, + 170, + 98 + ], + "type": "image", + "image_path": "c0647528d9cb92293ebe9c318fbb5f9500179dbf9bd0705beea3990f4bc71b8e.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 155, + 100, + 269, + 106 + ], + "lines": [ + { + "bbox": [ + 155, + 100, + 269, + 106 + ], + "spans": [ + { + "bbox": [ + 155, + 100, + 269, + 106 + ], + "type": "text", + "content": "In-domain Accuracy Out-of-domain Accuracy" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_footnote" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 170, + 32, + 238, + 98 + ], + "blocks": [ + { + "bbox": [ + 170, + 32, + 238, + 98 + ], + "lines": [ + { + "bbox": [ + 170, + 32, + 238, + 98 + ], + "spans": [ + { + "bbox": [ + 170, + 32, + 238, + 98 + ], + "type": "image", + "image_path": "5750340daef3e26b6ecded5b8076874aee4552a1b83e2bc1590faadbfed360e6.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 239, + 32, + 307, + 98 + ], + "blocks": [ + { + "bbox": [ + 239, + 32, + 307, + 98 + ], + "lines": [ + { + "bbox": [ + 239, + 32, + 307, + 98 + ], + "spans": [ + { + "bbox": [ + 239, + 32, + 307, + 98 + ], + "type": "image", + "image_path": "d754bad40a69e14f967c846c05c2a1e028772096c5264ff0647e3cfeb6c2a31b.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 308, + 32, + 379, + 98 + ], + "blocks": [ + { + "bbox": [ + 308, + 32, + 379, + 98 + ], + "lines": [ + { + "bbox": [ + 308, + 32, + 379, + 98 + ], + "spans": [ + { + "bbox": [ + 308, + 32, + 379, + 98 + ], + "type": "image", + "image_path": "7d738aba32c1e28d31f67b15a48b4c51be8761854e05b5eae2c6942b1267ec95.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 31, + 150, + 286, + 163 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 31, + 150, + 286, + 163 + ], + "spans": [ + { + "bbox": [ + 31, + 150, + 286, + 163 + ], + "type": "text", + "content": "4.5 In Comparison With Advanced DG Methods" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 30, + 170, + 381, + 301 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 170, + 381, + 301 + ], + "spans": [ + { + "bbox": [ + 30, + 170, + 381, + 301 + ], + "type": "text", + "content": "To objectively assess the effectiveness of the proposed RES, we conducted comparative analyses against several state-of-the-art DG methods. This comparison spans the baseline ERM, various data augmentation strategies (e.g., Mixup [40], CCFP [21], DomainDrop [14]), flatness-aware methods (such as SWAD [4], SMA [2], SAGM [38], FAD [45]), and some other DG approaches (including FISH [35], RSC [18], DAC-SC [20], GVRT [27], MIRO [5]). These evaluations were performed across five benchmarks utilizing the ResNet-50 backbone, with results detailed in Table 3. The domain-wise performance of the datasets of our RES are provided in Supplementary Material. Notably, RES incorporates both data augmentation and flatness-aware strategies, positioning it uniquely to tackle the dual challenges of cross-domain bias and variance simultaneously." + } + ] + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 32, + 351, + 379, + 504 + ], + "blocks": [ + { + "bbox": [ + 41, + 322, + 369, + 334 + ], + "lines": [ + { + "bbox": [ + 41, + 322, + 369, + 334 + ], + "spans": [ + { + "bbox": [ + 41, + 322, + 369, + 334 + ], + "type": "text", + "content": "Table 3: Performance comparisons between some SOTA methods and our RES." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 32, + 351, + 379, + 504 + ], + "lines": [ + { + "bbox": [ + 32, + 351, + 379, + 504 + ], + "spans": [ + { + "bbox": [ + 32, + 351, + 379, + 504 + ], + "type": "table", + "html": "
MethodPACSVLCSOfficeHomeTerraIncDomainNetAvg
ERM [ICLR'20] [13]85.7±0.577.4±0.367.5±0.547.2±0.441.2±0.263.8
RSC [ECCV'20] [18]85.2±0.977.1±0.565.5±0.946.6±1.038.9±0.562.7
Mixup [AAAI'20] [40]84.3±0.577.7±0.469.0±0.148.9±0.839.6±0.163.9
FISH [ICLR'21] [35]85.5±0.377.8±0.368.6±0.445.1±1.342.7±0.263.9
GVRT [ECCV'22] [27]85.1±0.379.0±0.270.1±0.148.0±0.244.1±0.165.2
MIRO [ECCV'22] [5]85.4±0.479.0±0.070.5±0.450.4±1.144.3±0.265.9
SMA [NeurIPS'22] [2]87.5±0.278.2±0.270.6±0.150.3±0.546.0±0.166.5
SWAD [NeurIPS'21] [4]88.1±0.179.1±0.170.6±0.250.0±0.346.5±0.166.9
CCFP [ICCV'23] [21]86.6±0.278.9±0.368.9±0.148.6±0.441.2±0.164.8
DAC-SC [CVPR'23] [20]87.5±0.178.7±0.370.3±0.246.5±0.344.9±0.165.6
FAD [ICCV'23] [45]88.2±0.278.9±0.869.2±0.545.7±1.044.4±0.165.3
SAGM [CVPR'23] [38]86.6±0.280.0±0.370.1±0.248.8±0.945.0±0.266.1
DomainDrop [ICCV'23] [14]87.9±0.379.8±0.368.7±0.151.5±0.444.4±0.566.5
Our RES90.0±0.379.8±0.271.8±0.351.4±0.646.7±0.267.9
", + "image_path": "93b0374640c31070a923ef58504389a5c6bb1c72d4bdd2ad7a99439b3f9cf8d5.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 30, + 536, + 381, + 584 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 536, + 381, + 584 + ], + "spans": [ + { + "bbox": [ + 30, + 536, + 381, + 584 + ], + "type": "text", + "content": "The table showcases the RES algorithm's performance across five benchmarks, with \"Avg\" representing the average performance across them. This provides a comprehensive view of each method's generalizability and effectiveness in DG. The proposed RES demonstrates superior performance, particularly high-" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 33, + 10, + 43, + 19 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 10, + 43, + 19 + ], + "spans": [ + { + "bbox": [ + 33, + 10, + 43, + 19 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 65, + 9, + 163, + 20 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 9, + 163, + 20 + ], + "spans": [ + { + "bbox": [ + 65, + 9, + 163, + 20 + ], + "type": "text", + "content": "W. Huang, Y. Shi et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 413, + 615 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "bbox": [ + 30, + 34, + 380, + 71 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 34, + 380, + 71 + ], + "spans": [ + { + "bbox": [ + 30, + 34, + 380, + 71 + ], + "type": "text", + "content": "lighted by its leading average score, indicating its robustness and adaptability across a diverse set of domains. The comparative analysis confirms the superiority and effectiveness of the RES in mitigating cross-domain bias and variance." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 31, + 89, + 268, + 102 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 31, + 89, + 268, + 102 + ], + "spans": [ + { + "bbox": [ + 31, + 89, + 268, + 102 + ], + "type": "text", + "content": "4.6 Effects of RES on Domain Generalization" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 30, + 109, + 380, + 146 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 109, + 380, + 146 + ], + "spans": [ + { + "bbox": [ + 30, + 109, + 380, + 146 + ], + "type": "text", + "content": "This subsection quantitatively describes the advantages of RES on DG from three perspectives, including reducing cross-domain bias, reducing cross-domain variance, and improving both in-domain and out-of-domain performance." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 30, + 164, + 197, + 175 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 164, + 197, + 175 + ], + "spans": [ + { + "bbox": [ + 30, + 164, + 197, + 175 + ], + "type": "text", + "content": "Reducing Cross-domain Bias." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 30, + 176, + 197, + 344 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 176, + 197, + 344 + ], + "spans": [ + { + "bbox": [ + 30, + 176, + 197, + 344 + ], + "type": "text", + "content": "First, we measure the cosine similarity between the mean feature embeddings obtained from the global average pooling layer of ResNet-50 across both source and target datasets within PACS and OfficeHome. This assessment encompasses comparisons between the baseline ERM and our RES, as illustrated in Fig. 5. Our findings indicate that RES significantly enhances the cosine similarity between source and target data compared to ERM, demonstrating its capability in diminishing cross-domain l" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 208, + 188, + 291, + 250 + ], + "blocks": [ + { + "bbox": [ + 208, + 188, + 291, + 250 + ], + "lines": [ + { + "bbox": [ + 208, + 188, + 291, + 250 + ], + "spans": [ + { + "bbox": [ + 208, + 188, + 291, + 250 + ], + "type": "image", + "image_path": "98537107b1eac575a6fa9f9a64829580caa55ad8d6d66e01c521662b59d6c3a9.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 223, + 251, + 289, + 257 + ], + "lines": [ + { + "bbox": [ + 223, + 251, + 289, + 257 + ], + "spans": [ + { + "bbox": [ + 223, + 251, + 289, + 257 + ], + "type": "text", + "content": "Source-target Similarity of ERM" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 293, + 188, + 375, + 250 + ], + "blocks": [ + { + "bbox": [ + 293, + 188, + 375, + 250 + ], + "lines": [ + { + "bbox": [ + 293, + 188, + 375, + 250 + ], + "spans": [ + { + "bbox": [ + 293, + 188, + 375, + 250 + ], + "type": "image", + "image_path": "3d30c3a52669d4733368fb3698831c99d9c06b62a9fa49ed93762eec97ee1fdf.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 296, + 251, + 361, + 257 + ], + "lines": [ + { + "bbox": [ + 296, + 251, + 361, + 257 + ], + "spans": [ + { + "bbox": [ + 296, + 251, + 361, + 257 + ], + "type": "text", + "content": "Source-target Similarity of RES" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 204, + 266, + 380, + 309 + ], + "lines": [ + { + "bbox": [ + 204, + 266, + 380, + 309 + ], + "spans": [ + { + "bbox": [ + 204, + 266, + 380, + 309 + ], + "type": "text", + "content": "Fig. 5: Cosine similarity between mean features of source domains and target domain. The higher the similarity is, the lower the cross-domain bias is." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 30, + 362, + 380, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 362, + 380, + 483 + ], + "spans": [ + { + "bbox": [ + 30, + 362, + 380, + 483 + ], + "type": "text", + "content": "Reducing Cross-domain Variance. Secondly, we present the training convergence trajectories alongside the corresponding cumulative instability for four DG scenarios within the PACS dataset, as illustrated in Fig. 6. It is observable that, in comparison to the baseline method ERM represented by the blue color, the proposed RES depicted in red markedly reduces the fluctuations in both in-domain and out-of-domain accuracies with almost no instability increase after the initial increase stage during training, which spans approximately the first 2000 iterations. Benefiting from the stable and narrower performance gap between in-domain (validation) and out-of-domain (test) sets, the best models chosen based on validation sets demonstrate consistent performance on test sets." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 30, + 511, + 380, + 584 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 511, + 380, + 584 + ], + "spans": [ + { + "bbox": [ + 30, + 511, + 380, + 584 + ], + "type": "text", + "content": "Improving Both In-domain and Out-of-domain Performance. As shown in Table 4, the comparison between ERM and RES on PACS illustrates significant enhancements in both in-domain and out-of-domain accuracies achieved by RES. While ERM delivers reliable in-domain performance, it falls short in adapting to unseen domains. In contrast, RES not only boosts in-domain accuracy from ERM's 96.5 to its 98.0 but also lifts the out-of-domain accuracy" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 326, + 9, + 347, + 19 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 9, + 347, + 19 + ], + "spans": [ + { + "bbox": [ + 326, + 9, + 347, + 19 + ], + "type": "text", + "content": "RES" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 369, + 10, + 379, + 19 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 369, + 10, + 379, + 19 + ], + "spans": [ + { + "bbox": [ + 369, + 10, + 379, + 19 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 413, + 615 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 33, + 33, + 119, + 97 + ], + "blocks": [ + { + "bbox": [ + 33, + 33, + 119, + 97 + ], + "lines": [ + { + "bbox": [ + 33, + 33, + 119, + 97 + ], + "spans": [ + { + "bbox": [ + 33, + 33, + 119, + 97 + ], + "type": "image", + "image_path": "39fe8bbb5f1e067860c76e62dfea3015376927079bca55e272d0b05d37440c84.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 120, + 33, + 204, + 96 + ], + "blocks": [ + { + "bbox": [ + 120, + 33, + 204, + 96 + ], + "lines": [ + { + "bbox": [ + 120, + 33, + 204, + 96 + ], + "spans": [ + { + "bbox": [ + 120, + 33, + 204, + 96 + ], + "type": "image", + "image_path": "f3a31b1b8db827223e6e43aad1ce85273232cfc0491c1404b9177188d29cffe2.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 205, + 33, + 291, + 96 + ], + "blocks": [ + { + "bbox": [ + 205, + 33, + 291, + 96 + ], + "lines": [ + { + "bbox": [ + 205, + 33, + 291, + 96 + ], + "spans": [ + { + "bbox": [ + 205, + 33, + 291, + 96 + ], + "type": "image", + "image_path": "a0521d58739d241ef42273e6b53f1ebcee8198a2033df73d3a97f6dcb2d0ab35.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 293, + 33, + 378, + 96 + ], + "blocks": [ + { + "bbox": [ + 293, + 33, + 378, + 96 + ], + "lines": [ + { + "bbox": [ + 293, + 33, + 378, + 96 + ], + "spans": [ + { + "bbox": [ + 293, + 33, + 378, + 96 + ], + "type": "image", + "image_path": "027a54cdeccde8d259cee14db751fb3daa804cbdfa50d12968278f4223fd0561.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 31, + 99, + 118, + 177 + ], + "blocks": [ + { + "bbox": [ + 31, + 99, + 118, + 177 + ], + "lines": [ + { + "bbox": [ + 31, + 99, + 118, + 177 + ], + "spans": [ + { + "bbox": [ + 31, + 99, + 118, + 177 + ], + "type": "image", + "image_path": "e595c659d3bd6cb12a812db44ae5b0ddf8c57f3a4112414f030e7072d7a8836a.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 30, + 185, + 380, + 230 + ], + "lines": [ + { + "bbox": [ + 30, + 185, + 380, + 230 + ], + "spans": [ + { + "bbox": [ + 30, + 185, + 380, + 230 + ], + "type": "text", + "content": "Fig. 6: Training stability of ERM and RES on PACS, evaluated by the absolute first derivative of accuracy every 300 iterations " + }, + { + "bbox": [ + 30, + 185, + 380, + 230 + ], + "type": "inline_equation", + "content": "(|f'(Acc)| = |Acc_t - Acc_{t-1}|)" + }, + { + "bbox": [ + 30, + 185, + 380, + 230 + ], + "type": "text", + "content": ", capturing accuracy change rate. Accumulative instability aggregates " + }, + { + "bbox": [ + 30, + 185, + 380, + 230 + ], + "type": "inline_equation", + "content": "|f'(Acc)|" + }, + { + "bbox": [ + 30, + 185, + 380, + 230 + ], + "type": "text", + "content": " up to iteration " + }, + { + "bbox": [ + 30, + 185, + 380, + 230 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 30, + 185, + 380, + 230 + ], + "type": "text", + "content": ", with lower values indicating more stable performance and less variance." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 120, + 99, + 205, + 176 + ], + "blocks": [ + { + "bbox": [ + 120, + 99, + 205, + 176 + ], + "lines": [ + { + "bbox": [ + 120, + 99, + 205, + 176 + ], + "spans": [ + { + "bbox": [ + 120, + 99, + 205, + 176 + ], + "type": "image", + "image_path": "fc040e88401742c649beef52436be9131fdd48af3115187c790327d92a5dc5e2.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 206, + 99, + 292, + 176 + ], + "blocks": [ + { + "bbox": [ + 206, + 99, + 292, + 176 + ], + "lines": [ + { + "bbox": [ + 206, + 99, + 292, + 176 + ], + "spans": [ + { + "bbox": [ + 206, + 99, + 292, + 176 + ], + "type": "image", + "image_path": "59be9bedf7b27b910332f53db8bd723ff6bad11f9a825a99fabf887ac72633c9.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 294, + 99, + 379, + 176 + ], + "blocks": [ + { + "bbox": [ + 294, + 99, + 379, + 176 + ], + "lines": [ + { + "bbox": [ + 294, + 99, + 379, + 176 + ], + "spans": [ + { + "bbox": [ + 294, + 99, + 379, + 176 + ], + "type": "image", + "image_path": "a4afb372df88a610c22aa9610c66d5cd420b3c56ff77c58e817fcea3176f6180.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 30, + 253, + 379, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 253, + 379, + 277 + ], + "spans": [ + { + "bbox": [ + 30, + 253, + 379, + 277 + ], + "type": "text", + "content": "from " + }, + { + "bbox": [ + 30, + 253, + 379, + 277 + ], + "type": "inline_equation", + "content": "85.5 \\pm 1.1" + }, + { + "bbox": [ + 30, + 253, + 379, + 277 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 30, + 253, + 379, + 277 + ], + "type": "inline_equation", + "content": "90.0 \\pm 0.3" + }, + { + "bbox": [ + 30, + 253, + 379, + 277 + ], + "type": "text", + "content": ". This robust improvement underscores RES's robust capability to refine model performance across various distributions." + } + ] + } + ], + "index": 11 + }, + { + "type": "table", + "bbox": [ + 33, + 319, + 378, + 389 + ], + "blocks": [ + { + "bbox": [ + 33, + 297, + 377, + 309 + ], + "lines": [ + { + "bbox": [ + 33, + 297, + 377, + 309 + ], + "spans": [ + { + "bbox": [ + 33, + 297, + 377, + 309 + ], + "type": "text", + "content": "Table 4: In-domain and out-of-domain accuracies of PACS between ERM and RES." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 33, + 319, + 378, + 389 + ], + "lines": [ + { + "bbox": [ + 33, + 319, + 378, + 389 + ], + "spans": [ + { + "bbox": [ + 33, + 319, + 378, + 389 + ], + "type": "table", + "html": "
SplitMethodPACSAvg
ArtClipartPaintingSketch
In-domain (Validation Set)ERM97.2±0.097.0±0.495.4±0.296.3±0.496.5±0.3
Our RES98.4±0.298.0±0.097.4±0.398.2±0.398.0±0.2
Out-of-domain (Test Set)ERM84.4±0.681.1±0.796.3±0.580.2±2.585.5±1.1
Our RES91.6±0.484.2±0.398.1±0.186.1±0.490.0±0.3
", + "image_path": "04a18eb3e252cd809c5273fcc125c6b7359511727a5e405b209c7a1e1e98a54d.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_body" + } + ], + "index": 13 + }, + { + "bbox": [ + 31, + 439, + 119, + 452 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 31, + 439, + 119, + 452 + ], + "spans": [ + { + "bbox": [ + 31, + 439, + 119, + 452 + ], + "type": "text", + "content": "5 Conclusion" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 30, + 464, + 379, + 583 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 30, + 464, + 379, + 583 + ], + "spans": [ + { + "bbox": [ + 30, + 464, + 379, + 583 + ], + "type": "text", + "content": "In this study, we explored DG through bias-variance decomposition, an analytical framework that offers deeper insights into the challenges and solutions for cross-domain generalization in deep learning models. Our approach quantified the detrimental effects of cross-domain bias and variance on model performance and therefore introduced the Representation Enhancement-Stabilization (RES) method as a potent solution. The RES method effectively minimizes the cross-domain bias and variance, enhancing model adaptability and generalization across unseen target domains. Empirical evaluations on five DG benchmarks reveals that the proposed RES markedly outperforms multiple advanced DG methods in improving model generalization." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 33, + 10, + 43, + 19 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 10, + 43, + 19 + ], + "spans": [ + { + "bbox": [ + 33, + 10, + 43, + 19 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 65, + 9, + 163, + 20 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 9, + 163, + 20 + ], + "spans": [ + { + "bbox": [ + 65, + 9, + 163, + 20 + ], + "type": "text", + "content": "W. Huang, Y. Shi et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 413, + 615 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 34, + 33, + 144, + 46 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 34, + 33, + 144, + 46 + ], + "spans": [ + { + "bbox": [ + 34, + 33, + 144, + 46 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 33, + 58, + 379, + 201 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 58, + 379, + 201 + ], + "spans": [ + { + "bbox": [ + 33, + 58, + 379, + 201 + ], + "type": "text", + "content": "The work was supported by the German Research Foundation (DFG GZ: ZH 498/18-1; Project number: 519016653). The work of Z. Xiong was supported by the German Federal Ministry for the Environment, Nature Conservation, Nuclear Safety and Consumer Protection (BMUV) based on a resolution of the German Bundestag (grant number: 67KI32002B; Acronym: " + }, + { + "bbox": [ + 33, + 58, + 379, + 201 + ], + "type": "inline_equation", + "content": "EKAPEx" + }, + { + "bbox": [ + 33, + 58, + 379, + 201 + ], + "type": "text", + "content": "). The work of X. Zhu was also supported by the German Federal Ministry of Education and Research (BMBF) in the framework of the international future AI lab \"AI4EO - Artificial Intelligence for Earth Observation: Reasoning, Uncertainties, Ethics and Beyond\" (grant number: 01DD20001), by German Federal Ministry for Economic Affairs and Climate Action in the framework of the \"national center of excellence ML4Earth\" (grant number: 50EE2201C) and by the Munich Center for Machine Learning." + } + ] + } + ], + "index": 3 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 327, + 10, + 346, + 18 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 10, + 346, + 18 + ], + "spans": [ + { + "bbox": [ + 327, + 10, + 346, + 18 + ], + "type": "text", + "content": "RES" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 369, + 11, + 378, + 18 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 369, + 11, + 378, + 18 + ], + "spans": [ + { + "bbox": [ + 369, + 11, + 378, + 18 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 413, + 615 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 32, + 33, + 98, + 45 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 32, + 33, + 98, + 45 + ], + "spans": [ + { + "bbox": [ + 32, + 33, + 98, + 45 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 33, + 60, + 380, + 583 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 36, + 60, + 379, + 93 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 36, + 60, + 379, + 93 + ], + "spans": [ + { + "bbox": [ + 36, + 60, + 379, + 93 + ], + "type": "text", + "content": "1. Albuquerque, I., Naik, N., Li, J., Keskar, N., Socher, R.: Improving out-of-distribution generalization via multi-task self-supervised pretraining. arXiv preprint arXiv:2003.13525 (2020)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 36, + 94, + 379, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 36, + 94, + 379, + 127 + ], + "spans": [ + { + "bbox": [ + 36, + 94, + 379, + 127 + ], + "type": "text", + "content": "2. Arpit, D., Wang, H., Zhou, Y., Xiong, C.: Ensemble of averages: Improving model selection and boosting performance in domain generalization. Advances in Neural Information Processing Systems 35, 8265-8277 (2022)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 36, + 127, + 379, + 148 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 36, + 127, + 379, + 148 + ], + "spans": [ + { + "bbox": [ + 36, + 127, + 379, + 148 + ], + "type": "text", + "content": "3. Beery, S., Van Horn, G., Perona, P.: Recognition in terra incognita. In: Proceedings of the European conference on computer vision (ECCV). pp. 456-473 (2018)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 37, + 150, + 380, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 37, + 150, + 380, + 182 + ], + "spans": [ + { + "bbox": [ + 37, + 150, + 380, + 182 + ], + "type": "text", + "content": "4. Cha, J., Chun, S., Lee, K., Cho, H.C., Park, S., Lee, Y., Park, S.: Swad: Domain generalization by seeking flat minima. Advances in Neural Information Processing Systems 34, 22405-22418 (2021)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 37, + 183, + 379, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 37, + 183, + 379, + 216 + ], + "spans": [ + { + "bbox": [ + 37, + 183, + 379, + 216 + ], + "type": "text", + "content": "5. Cha, J., Lee, K., Park, S., Chun, S.: Domain generalization by mutual-information regularization with pre-trained models. In: European Conference on Computer Vision. pp. 440-457. Springer (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 37, + 217, + 379, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 37, + 217, + 379, + 249 + ], + "spans": [ + { + "bbox": [ + 37, + 217, + 379, + 249 + ], + "type": "text", + "content": "6. Chen, C., Li, J., Han, X., Liu, X., Yu, Y.: Compound domain generalization via meta-knowledge encoding. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7119-7129 (2022)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 37, + 250, + 379, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 37, + 250, + 379, + 282 + ], + "spans": [ + { + "bbox": [ + 37, + 250, + 379, + 282 + ], + "type": "text", + "content": "7. Chen, J., Gao, Z., Wu, X., Luo, J.: Meta-causal learning for single domain generalization. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7683-7692 (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 37, + 283, + 379, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 37, + 283, + 379, + 327 + ], + "spans": [ + { + "bbox": [ + 37, + 283, + 379, + 327 + ], + "type": "text", + "content": "8. Cubuk, E.D., Zoph, B., Shlens, J., Le, Q.V.: Randaugment: Practical automated data augmentation with a reduced search space. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops. pp. 702-703 (2020)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 37, + 327, + 379, + 361 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 37, + 327, + 379, + 361 + ], + "spans": [ + { + "bbox": [ + 37, + 327, + 379, + 361 + ], + "type": "text", + "content": "9. Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: Imagenet: A large-scale hierarchical image database. In: 2009 IEEE conference on computer vision and pattern recognition. pp. 248-255. IEEE (2009)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 33, + 361, + 379, + 394 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 361, + 379, + 394 + ], + "spans": [ + { + "bbox": [ + 33, + 361, + 379, + 394 + ], + "type": "text", + "content": "10. Fang, C., Xu, Y., Rockmore, D.N.: Unbiased metric learning: On the utilization of multiple datasets and web images for softening bias. In: Proceedings of the IEEE International Conference on Computer Vision. pp. 1657-1664 (2013)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 33, + 394, + 379, + 416 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 394, + 379, + 416 + ], + "spans": [ + { + "bbox": [ + 33, + 394, + 379, + 416 + ], + "type": "text", + "content": "11. Foret, P., Kleiner, A., Mobahi, H., Neyshabur, B.: Sharpness-aware minimization for efficiently improving generalization. arXiv preprint arXiv:2010.01412 (2020)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 33, + 417, + 379, + 439 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 417, + 379, + 439 + ], + "spans": [ + { + "bbox": [ + 33, + 417, + 379, + 439 + ], + "type": "text", + "content": "12. Geman, S., Bienenstock, E., Doursat, R.: Neural networks and the bias/variance dilemma. Neural computation 4(1), 1-58 (1992)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 33, + 439, + 379, + 461 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 439, + 379, + 461 + ], + "spans": [ + { + "bbox": [ + 33, + 439, + 379, + 461 + ], + "type": "text", + "content": "13. Gulrajani, I., Lopez-Paz, D.: In search of lost domain generalization. In: International Conference on Learning Representations (2020)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 33, + 461, + 379, + 494 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 461, + 379, + 494 + ], + "spans": [ + { + "bbox": [ + 33, + 461, + 379, + 494 + ], + "type": "text", + "content": "14. Guo, J., Qi, L., Shi, Y.: Domaindrop: Suppressing domain-sensitive channels for domain generalization. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 19114-19124 (2023)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 33, + 495, + 379, + 517 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 495, + 379, + 517 + ], + "spans": [ + { + "bbox": [ + 33, + 495, + 379, + 517 + ], + "type": "text", + "content": "15. Hansen, B.C., Hess, R.F.: Structural sparseness and spatial phase alignment in natural scenes. JOSA A 24(7), 1873-1885 (2007)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 33, + 517, + 379, + 550 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 517, + 379, + 550 + ], + "spans": [ + { + "bbox": [ + 33, + 517, + 379, + 550 + ], + "type": "text", + "content": "16. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 770-778 (2016)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 33, + 551, + 379, + 583 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 551, + 379, + 583 + ], + "spans": [ + { + "bbox": [ + 33, + 551, + 379, + 583 + ], + "type": "text", + "content": "17. Huang, X., Belongie, S.: Arbitrary style transfer in real-time with adaptive instance normalization. In: Proceedings of the IEEE international conference on computer vision. pp. 1501-1510 (2017)" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 33, + 10, + 43, + 19 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 10, + 43, + 19 + ], + "spans": [ + { + "bbox": [ + 33, + 10, + 43, + 19 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 65, + 9, + 162, + 20 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 9, + 162, + 20 + ], + "spans": [ + { + "bbox": [ + 65, + 9, + 162, + 20 + ], + "type": "text", + "content": "W. Huang, Y. Shi et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 413, + 615 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 33, + 35, + 380, + 583 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 33, + 35, + 379, + 79 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 35, + 379, + 79 + ], + "spans": [ + { + "bbox": [ + 33, + 35, + 379, + 79 + ], + "type": "text", + "content": "18. Huang, Z., Wang, H., Xing, E.P., Huang, D.: Self-challenging improves cross-domain generalization. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part II 16. pp. 124-140. Springer (2020)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 33, + 79, + 379, + 112 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 79, + 379, + 112 + ], + "spans": [ + { + "bbox": [ + 33, + 79, + 379, + 112 + ], + "type": "text", + "content": "19. Kim, D., Yoo, Y., Park, S., Kim, J., Lee, J.: Selfreg: Self-supervised contrastive regularization for domain generalization. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 9619-9628 (2021)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 33, + 113, + 380, + 156 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 113, + 380, + 156 + ], + "spans": [ + { + "bbox": [ + 33, + 113, + 380, + 156 + ], + "type": "text", + "content": "20. Lee, S., Bae, J., Kim, H.Y.: Decompose, adjust, compose: Effective normalization by playing with frequency for domain generalization. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11776-11785 (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 33, + 156, + 379, + 188 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 156, + 379, + 188 + ], + "spans": [ + { + "bbox": [ + 33, + 156, + 379, + 188 + ], + "type": "text", + "content": "21. Li, C., Zhang, D., Huang, W., Zhang, J.: Cross contrasting feature perturbation for domain generalization. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 1327-1337 (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 33, + 189, + 379, + 222 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 189, + 379, + 222 + ], + "spans": [ + { + "bbox": [ + 33, + 189, + 379, + 222 + ], + "type": "text", + "content": "22. Li, D., Yang, Y., Song, Y.Z., Hospedales, T.M.: Deeper, broader and artier domain generalization. In: Proceedings of the IEEE international conference on computer vision. pp. 5542-5550 (2017)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 33, + 222, + 379, + 254 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 222, + 379, + 254 + ], + "spans": [ + { + "bbox": [ + 33, + 222, + 379, + 254 + ], + "type": "text", + "content": "23. Li, H., Pan, S.J., Wang, S., Kot, A.C.: Domain generalization with adversarial feature learning. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 5400-5409 (2018)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 33, + 255, + 379, + 287 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 255, + 379, + 287 + ], + "spans": [ + { + "bbox": [ + 33, + 255, + 379, + 287 + ], + "type": "text", + "content": "24. Li, P., Li, D., Li, W., Gong, S., Fu, Y., Hospedales, T.M.: A simple feature augmentation for domain generalization. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 8886-8895 (2021)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 33, + 288, + 379, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 288, + 379, + 331 + ], + "spans": [ + { + "bbox": [ + 33, + 288, + 379, + 331 + ], + "type": "text", + "content": "25. Lv, F., Liang, J., Li, S., Zang, B., Liu, C.H., Wang, Z., Liu, D.: Causality inspired representation learning for domain generalization. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8046-8056 (2022)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 33, + 332, + 378, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 332, + 378, + 353 + ], + "spans": [ + { + "bbox": [ + 33, + 332, + 378, + 353 + ], + "type": "text", + "content": "26. Mahajan, D., Tople, S., Sharma, A.: Domain generalization using causal matching. In: International Conference on Machine Learning. pp. 7313-7324. PMLR (2021)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 33, + 354, + 379, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 354, + 379, + 386 + ], + "spans": [ + { + "bbox": [ + 33, + 354, + 379, + 386 + ], + "type": "text", + "content": "27. Min, S., Park, N., Kim, S., Park, S., Kim, J.: Grounding visual representations with texts for domain generalization. In: European Conference on Computer Vision. pp. 37-53. Springer (2022)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 33, + 386, + 379, + 419 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 386, + 379, + 419 + ], + "spans": [ + { + "bbox": [ + 33, + 386, + 379, + 419 + ], + "type": "text", + "content": "28. Muandet, K., Balduzzi, D., Scholkopf, B.: Domain generalization via invariant feature representation. In: International conference on machine learning. pp. 10-18. PMLR (2013)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 33, + 420, + 379, + 452 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 420, + 379, + 452 + ], + "spans": [ + { + "bbox": [ + 33, + 420, + 379, + 452 + ], + "type": "text", + "content": "29. Oppenheim, A., Lim, J., Kopec, G., Pohlig, S.: Phase in speech and pictures. In: ICASSP'79. IEEE International Conference on Acoustics, Speech, and Signal Processing. vol. 4, pp. 632-637. IEEE (1979)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 33, + 452, + 379, + 485 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 452, + 379, + 485 + ], + "spans": [ + { + "bbox": [ + 33, + 452, + 379, + 485 + ], + "type": "text", + "content": "30. Peng, X., Bai, Q., Xia, X., Huang, Z., Saenko, K., Wang, B.: Moment matching for multi-source domain adaptation. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 1406-1415 (2019)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 33, + 485, + 379, + 517 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 485, + 379, + 517 + ], + "spans": [ + { + "bbox": [ + 33, + 485, + 379, + 517 + ], + "type": "text", + "content": "31. Peng, X., Huang, Z., Sun, X., Saenko, K.: Domain agnostic learning with disentangled representations. In: International Conference on Machine Learning. pp. 5102-5112. PMLR (2019)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 33, + 518, + 379, + 550 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 518, + 379, + 550 + ], + "spans": [ + { + "bbox": [ + 33, + 518, + 379, + 550 + ], + "type": "text", + "content": "32. Piotrowski, L.N., Campbell, F.W.: A demonstration of the visual importance and flexibility of spatial-frequency amplitude and phase. Perception 11(3), 337-346 (1982)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 33, + 551, + 379, + 583 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 551, + 379, + 583 + ], + "spans": [ + { + "bbox": [ + 33, + 551, + 379, + 583 + ], + "type": "text", + "content": "33. Piratla, V., Netrapalli, P., Sarawagi, S.: Efficient domain generalization via common-specific low-rank decomposition. In: International Conference on Machine Learning. pp. 7728-7738. PMLR (2020)" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 326, + 9, + 347, + 19 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 9, + 347, + 19 + ], + "spans": [ + { + "bbox": [ + 326, + 9, + 347, + 19 + ], + "type": "text", + "content": "RES" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 369, + 10, + 379, + 19 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 369, + 10, + 379, + 19 + ], + "spans": [ + { + "bbox": [ + 369, + 10, + 379, + 19 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 413, + 615 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 33, + 35, + 379, + 551 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 33, + 35, + 379, + 79 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 35, + 379, + 79 + ], + "spans": [ + { + "bbox": [ + 33, + 35, + 379, + 79 + ], + "type": "text", + "content": "34. Seo, S., Suh, Y., Kim, D., Kim, G., Han, J., Han, B.: Learning to optimize domain specific normalization for domain generalization. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXII 16. pp. 68-83. Springer (2020)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 33, + 79, + 379, + 112 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 79, + 379, + 112 + ], + "spans": [ + { + "bbox": [ + 33, + 79, + 379, + 112 + ], + "type": "text", + "content": "35. Shi, Y., Seely, J., Torr, P.H., Siddharth, N., Hannun, A., Usunier, N., Synnaeve, G.: Gradient matching for domain generalization. arXiv preprint arXiv:2104.09937 (2021)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 33, + 113, + 379, + 145 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 113, + 379, + 145 + ], + "spans": [ + { + "bbox": [ + 33, + 113, + 379, + 145 + ], + "type": "text", + "content": "36. Venkateswara, H., Eusebio, J., Chakraborty, S., Panchanathan, S.: Deep hashing network for unsupervised domain adaptation. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 5018-5027 (2017)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 33, + 145, + 379, + 188 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 145, + 379, + 188 + ], + "spans": [ + { + "bbox": [ + 33, + 145, + 379, + 188 + ], + "type": "text", + "content": "37. Wang, G., Han, H., Shan, S., Chen, X.: Cross-domain face presentation attack detection via multi-domain disentangled representation learning. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 6678-6687 (2020)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 33, + 189, + 379, + 222 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 189, + 379, + 222 + ], + "spans": [ + { + "bbox": [ + 33, + 189, + 379, + 222 + ], + "type": "text", + "content": "38. Wang, P., Zhang, Z., Lei, Z., Zhang, L.: Sharpness-aware gradient matching for domain generalization. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3769-3778 (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 33, + 222, + 379, + 254 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 222, + 379, + 254 + ], + "spans": [ + { + "bbox": [ + 33, + 222, + 379, + 254 + ], + "type": "text", + "content": "39. Wang, Y., Qi, L., Shi, Y., Gao, Y.: Feature-based style randomization for domain generalization. IEEE Transactions on Circuits and Systems for Video Technology 32(8), 5495-5509 (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 33, + 255, + 379, + 287 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 255, + 379, + 287 + ], + "spans": [ + { + "bbox": [ + 33, + 255, + 379, + 287 + ], + "type": "text", + "content": "40. Xu, M., Zhang, J., Ni, B., Li, T., Wang, C., Tian, Q., Zhang, W.: Adversarial domain adaptation with domain mixup. In: Proceedings of the AAAI conference on artificial intelligence. vol. 34, pp. 6502-6509 (2020)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 33, + 288, + 379, + 321 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 288, + 379, + 321 + ], + "spans": [ + { + "bbox": [ + 33, + 288, + 379, + 321 + ], + "type": "text", + "content": "41. Xu, Q., Zhang, R., Zhang, Y., Wang, Y., Tian, Q.: A fourier-based framework for domain generalization. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14383-14392 (2021)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 33, + 321, + 379, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 321, + 379, + 353 + ], + "spans": [ + { + "bbox": [ + 33, + 321, + 379, + 353 + ], + "type": "text", + "content": "42. Yang, F.E., Cheng, Y.C., Shiau, Z.Y., Wang, Y.C.F.: Adversarial teacher-student representation learning for domain generalization. Advances in Neural Information Processing Systems 34, 19448-19460 (2021)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 33, + 354, + 379, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 354, + 379, + 386 + ], + "spans": [ + { + "bbox": [ + 33, + 354, + 379, + 386 + ], + "type": "text", + "content": "43. Yang, Z., Yu, Y., You, C., Steinhardt, J., Ma, Y.: Rethinking bias-variance tradeoff for generalization of neural networks. In: International Conference on Machine Learning. pp. 10767-10777. PMLR (2020)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 33, + 386, + 379, + 419 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 386, + 379, + 419 + ], + "spans": [ + { + "bbox": [ + 33, + 386, + 379, + 419 + ], + "type": "text", + "content": "44. Zhang, J., Qi, L., Shi, Y., Gao, Y.: Mvdg: A unified multi-view framework for domain generalization. In: European Conference on Computer Vision. pp. 161-177. Springer (2022)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 33, + 420, + 379, + 452 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 420, + 379, + 452 + ], + "spans": [ + { + "bbox": [ + 33, + 420, + 379, + 452 + ], + "type": "text", + "content": "45. Zhang, X., Xu, R., Yu, H., Dong, Y., Tian, P., Cui, P.: Flatness-aware minimization for domain generalization. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5189-5202 (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 33, + 452, + 379, + 484 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 452, + 379, + 484 + ], + "spans": [ + { + "bbox": [ + 33, + 452, + 379, + 484 + ], + "type": "text", + "content": "46. Zhao, S., Gong, M., Liu, T., Fu, H., Tao, D.: Domain generalization via entropy regularization. Advances in Neural Information Processing Systems 33, 16096-16107 (2020)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 33, + 485, + 379, + 517 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 485, + 379, + 517 + ], + "spans": [ + { + "bbox": [ + 33, + 485, + 379, + 517 + ], + "type": "text", + "content": "47. Zhou, K., Yang, Y., Hospedales, T., Xiang, T.: Deep domain-adversarial image generation for domain generalisation. In: Proceedings of the AAAI conference on artificial intelligence. vol. 34, pp. 13025-13032 (2020)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 33, + 518, + 379, + 551 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 518, + 379, + 551 + ], + "spans": [ + { + "bbox": [ + 33, + 518, + 379, + 551 + ], + "type": "text", + "content": "48. Zhu, W., Lu, L., Xiao, J., Han, M., Luo, J., Harrison, A.P.: Localized adversarial domain generalization. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7108-7118 (2022)" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 33, + 10, + 43, + 18 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 33, + 10, + 43, + 18 + ], + "spans": [ + { + "bbox": [ + 33, + 10, + 43, + 18 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 65, + 9, + 162, + 20 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 9, + 162, + 20 + ], + "spans": [ + { + "bbox": [ + 65, + 9, + 162, + 20 + ], + "type": "text", + "content": "W. Huang, Y. Shi et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 413, + 615 + ], + "page_idx": 17 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Representing Topological Self-Similarity Using Fractal Feature Maps for Accurate Segmentation of Tubular Structures/494e6709-0fe6-4415-a992-88fda17780d8_content_list.json b/2024/Representing Topological Self-Similarity Using Fractal Feature Maps for Accurate Segmentation of Tubular Structures/494e6709-0fe6-4415-a992-88fda17780d8_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..551253b131f115e13b6b5b53b286953d0ae0b7ce --- /dev/null +++ b/2024/Representing Topological Self-Similarity Using Fractal Feature Maps for Accurate Segmentation of Tubular Structures/494e6709-0fe6-4415-a992-88fda17780d8_content_list.json @@ -0,0 +1,2015 @@ +[ + { + "type": "text", + "text": "Representing Topological Self-Similarity Using Fractal Feature Maps for Accurate Segmentation of Tubular Structures", + "text_level": 1, + "bbox": [ + 222, + 140, + 781, + 205 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jiaxing Huang $^{1,2}$ , Yanfeng Zhou $^{1,2}$ , Yaoru Luo $^{1,2}$ , Guole Liu $^{1,2}$ , Heng Guo $^{3,4}$ , and Ge Yang $^{1,2}$", + "bbox": [ + 250, + 233, + 754, + 267 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Institute of Automation, Chinese Academy of Sciences, 100190, Beijing, China", + "bbox": [ + 232, + 276, + 771, + 291 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ School of Artificial Intelligence, University of Chinese Academy of Sciences", + "bbox": [ + 243, + 291, + 756, + 304 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3 DAMO Academy, Alibaba Group", + "bbox": [ + 382, + 304, + 620, + 319 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "4 Hupan Lab, 310023, Hangzhou, China", + "bbox": [ + 366, + 319, + 635, + 333 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{huangjiaxing2021, ge.yang}@ia.ac.cn", + "bbox": [ + 361, + 333, + 640, + 347 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Accurate segmentation of long and thin tubular structures is required in a wide variety of areas such as biology, medicine, and remote sensing. The complex topology and geometry of such structures often pose significant technical challenges. A fundamental property of such structures is their topological self-similarity, which can be quantified by fractal features such as fractal dimension (FD). In this study, we incorporate fractal features into a deep learning model by extending FD to the pixel-level using a sliding window technique. The resulting fractal feature maps (FFMs) are then incorporated as additional input to the model and additional weight in the loss function to enhance segmentation performance by utilizing the topological self-similarity. Moreover, we extend the U-Net architecture by incorporating an edge decoder and a skeleton decoder to improve boundary accuracy and skeletal continuity of segmentation, respectively. Extensive experiments on five tubular structure datasets validate the effectiveness and robustness of our approach. Furthermore, the integration of FFMs with other popular segmentation models such as HR-Net also yields performance enhancement, suggesting FFM can be incorporated as a plug-in module with different model architectures. Code and data are openly accessible at https://github.com/cbmi-group/FFM-Multi-Decoder-Network.", + "bbox": [ + 259, + 386, + 743, + 662 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Tubular structures $\\cdot$ Topological self-similarity $\\cdot$ Fractal feature map $\\cdot$ Deep learning $\\cdot$ Plug-in module", + "bbox": [ + 259, + 676, + 743, + 705 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 731, + 375, + 748 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Accurate segmentation of tubular structures is of significant importance across a wide variety of areas. In the area of biological research, for example, the accurate segmentation of tubular structures such as the endoplasmic reticulum (ER, Fig. 1) is critical to the study of related human disease mechanisms [16,28]. In the area of clinical research, the accurate segmentation of blood vessels (Fig. 1) is", + "bbox": [ + 212, + 763, + 787, + 842 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/28bfbbe558905be1d30148d324d73501d360da3aa552c19ecb45b9edb57b7b1d.jpg", + "image_caption": [ + "Fig. 1: Topological self-similarity in tubular structures. If we consider \"one junction with multiple edges\" as a basic component, complex tubular structures have similar components at different scales. The images on the right are magnified views of the rectangular regions on the left." + ], + "image_footnote": [], + "bbox": [ + 305, + 143, + 702, + 297 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "essential to the early diagnosis of diseases such as retinopathy and stroke [39,58]. Similarly, in the area of remote sensing, the accurate extraction of roads from aerial imagery is essential for navigation and route planning [27]. However, accurate segmentation of tubular structures from images remains challenging due to factors such as their complex morphology and geometry, low image signal-to-noise ratio, and poor image contrast.", + "bbox": [ + 212, + 383, + 787, + 474 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "A wide variety of techniques have been developed for the segmentation of tubular structures. Classical methods rely on manually crafted features such as intensity, texture, and shape. For example, previous studies [2,7,40,56] have utilized deformable shape models to fit tubular structures, leveraging their geometric properties. However, these techniques often cannot handle challenges posed by factors such as poor contrast, high noise and complex background. Deep learning methods [13, 25, 34, 44, 59] have revolutionized image segmentation and achieved substantial improvements in segmentation performance. Recent studies on using deep learning models for segmentation of tubular structures have focused on optimization of loss functions [3, 26, 46, 49, 51] and refinement of model structures [10, 12, 18, 37, 42]. However, these studies primarily aim to achieve high segmentation performance utilizing a limited input of images without providing additional information to their segmentation models. Tubular structures exhibit distinct topology and geometry that are vital for segmentation. In this study, we explore using characteristics of their structure to assist deep learning models in segmentation.", + "bbox": [ + 212, + 476, + 787, + 717 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "One notable characteristic of tubular structures is their topological self-similarity, namely large and complex tubular structures exhibit similar topological patterns at different scales. For example, Fig. 1 shows that if \"one junction with multiple edges\" is considered a primary structural component, tubular structure entities such as the ER network and the retinal blood vessel network exhibit similar topology at both global and local scales. To quantify the topological self-similarity of tubular structures, we utilize the fractal theory, which characterizes self-similarity of intricate structures at different scales [35,61]. Cen-", + "bbox": [ + 212, + 719, + 787, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "J. Huang et al.", + "bbox": [ + 271, + 114, + 372, + 128 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Intral to this theory is the fractal dimension (FD), a key parameter used previously to describe the textural attributes of images.", + "bbox": [ + 212, + 146, + 782, + 176 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In addition to topology and geometry, edges and skeletons are important characteristics in defining tubular structures. In the segmentation of these structures, boundary accuracy and skeletal continuity are crucial for downstream tasks. For example, a commonly encountered problem in the segmentation of interconnected tubular structures (Fig. 1) is the breakages in segmentation results due to factors such as low-contrast or blurring.", + "bbox": [ + 212, + 176, + 784, + 267 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To enhance the segmentation quality of tubular structures, we exploit the topological self-similarity as well as edge and skeletal characteristics of tubular structures. The main research contributions of this study are as follows:", + "bbox": [ + 212, + 268, + 784, + 311 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "1) We have developed a strategy to incorporate fractal features into deep learning networks. Specifically, we extend the fractal dimension from the image-level to the pixel-level and generate the fractal feature map (FFM), which characterizes topological self-similarity and textural complexity of each region within an image or its associated label. The FFM computed from the image is denoted as $FFM_{image}$ , while the FFM derived from the label is indicated as $FFM_{label}$ . Utilization of $FFM_{image}$ as an additional model input and $FFM_{label}$ as an additional loss function weight substantially enhances segmentation performance.", + "bbox": [ + 212, + 314, + 784, + 434 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2) We develop the multi-decoder network (MD-Net) by extending the U-Net [44] architecture with an edge decoder and a skeleton decoder. These decoders enable the model to simultaneously predict the boundaries and skeletons of image objects, in addition to the primary segmentation masks. By incorporating related constraints within the loss function, our model focuses not only on achieving accurate target segmentation but also allocates increased attention to boundary delineation and skeleton preservation, thereby enhancing the overall prediction quality.", + "bbox": [ + 212, + 435, + 784, + 556 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3) We have demonstrated the versatility and robustness of FFMs. Incorporation of $FFM_{image}$ into the vanilla U-Net [44] and HR-Net [50] enhances segmentation performance, indicating that FFM can be used as a plug-in for different models.", + "bbox": [ + 212, + 556, + 784, + 614 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 214, + 642, + 387, + 657 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.1 Tubular Structure Segmentation", + "text_level": 1, + "bbox": [ + 214, + 676, + 529, + 691 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Classical methods have been proposed to improve the performance of segmenting tubular structures by taking into account their geometric characteristics. Firstly, various methods utilize active contours and compute geodesics or minimal distance curves to approximate the contours of tubular structures, thereby effectively delineating their boundaries [1,6]. Secondly, tree structure-based methods utilize intrinsic shape priors to assist segmentation of different tubular structures. These methods employ a bottom-up method to identify tubular objects and a top-down grouping strategy to recognize tree structures, generating corresponding shape priors [4]. Lastly, various centerline-based methods", + "bbox": [ + 212, + 704, + 785, + 839 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "Fractal Feature Maps for Segmentation of Tubular Structures", + "bbox": [ + 321, + 114, + 730, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "such as the one developed in [47] use multiscale detection strategies to accurately identify centerlines so that distance transform can be used to provide valuable information for segmentation of tubular structures.", + "bbox": [ + 212, + 146, + 782, + 191 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Deep learning methods have also been proposed to integrate topological and geometrical prior knowledge of tubular structures to enhance the performance of their segmentation. The integration is primarily achieved in three ways.", + "bbox": [ + 212, + 191, + 784, + 251 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "1) Convolutional kernel design. The popular deformable convolution [9] and dilated convolution [57] aim to overcome the limitations of geometric transformations in CNNs and have demonstrated exceptional performance in complex segmentation tasks. Additionally, DSC-Net utilizes dynamic snake convolution to accurately capture the distinct features of tubular structures [42]. By adaptively focusing on slender and winding local structures, DSC-Net achieves improved performance in capturing the intricacy of tubular structures.", + "bbox": [ + 212, + 252, + 785, + 357 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2) Model architecture design. Various architecture designs have been proposed to learn the topological and geometrical features of tubular structures. In [31], a global transformer and dual local attention network are employed to simultaneously capture global and local features to effectively learn the complex geometric properties of tubular structures. Dong et al. [10] propose an enhanced Deformable U-Net that exploits flexible deformable convolutional layers to better generate clear boundaries for 3D cardiac cine MRI. High-frequency components that have strong capabilities to perceive thin structures are fused in [14] to enhance the performance of segmenting thin structures.", + "bbox": [ + 212, + 358, + 787, + 492 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3) Loss function design. Various loss functions have been explored for the segmentation of tubular structures. In [46], a similarity measure centerlineDice (clDice) based on the intersection of segmentation masks and their respective skeletons is introduced. A loss function based on clDice is proposed to enable the networks to generate segmentations with more accurate connectivity information and topology preservation. Araujo et al. [3] design a loss function based on the morphological closing operator that allows models to produce more topologically coherent masks and consistent vascular trees. Wong et al. [52] apply a novel Persistence Diagram Loss that quantifies topological correctness of segmentation over fine-grained structures.", + "bbox": [ + 212, + 493, + 787, + 643 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this study, instead of relying on intricate designs of model architectures or loss functions, we incorporate FFMs as a model input and a loss function weight to enhance the perception of topological self-similarity and textures of images.", + "bbox": [ + 212, + 645, + 785, + 690 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.2 Fractal Theory and Applications", + "text_level": 1, + "bbox": [ + 214, + 710, + 532, + 726 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Despite their topological and geometrical complexity, tubular structures often exhibit topological similarities across different spatial scales. This observation suggests that their complex spatial patterns can be effectively described using simple texture features. Fractal geometry provides a means to describe the irregular or fragmented shapes of natural features and other intricate objects that traditional Euclidean geometry struggles to analyze [29]. Specifically, fractal features offer the capacity to describe and characterize the topological and", + "bbox": [ + 212, + 734, + 787, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "J. Huang et al.", + "bbox": [ + 271, + 114, + 372, + 128 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "geometrical complexity as well as textural composition of tubular image objects. Thus, fractal geometry has been applied to image classification and segmentation tasks.", + "bbox": [ + 212, + 146, + 782, + 190 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For classification, Roberto et al. [43] and Lin et al. [32] utilize the fractional Brownian motion (FBM) model [23] to extract fractal features of images. These features are then fed into a support vector machine or a convolutional neural network (CNN) classifier to differentiate between different objects. For segmentation, the FBM model is utilized in [33, 61] to extract fractal features from images. Such features are combined with classical methods such as thresholding and region growth techniques for image object segmentation.", + "bbox": [ + 212, + 191, + 784, + 297 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Although existing methods have demonstrated good performance by leveraging fractal features, there is still a gap in exploring the integration of fractal features with deep learning for segmentation tasks. The inherent self-similarity observed in tubular structures aligns well with the fractal theory. In this study, we address this gap by incorporating FFMs into the segmentation model, aiming to provide a new and reliable source of information to enhance segmentation performance.", + "bbox": [ + 212, + 297, + 785, + 404 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3 Method", + "text_level": 1, + "bbox": [ + 215, + 431, + 330, + 446 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1 Fractal Feature Map", + "text_level": 1, + "bbox": [ + 215, + 465, + 431, + 482 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In fractal geometry, the fractal dimension (FD) provides a quantitative measure of an image's degree of self-similarity and roughness. FD can be estimated via the property of self-similarity [41]. Given an image $A$ , it is self-similar if $A$ comprises $N_r$ distinct copies of itself scaled down by a factor of $r$ . Consequently, for an image, the FD is defined as:", + "bbox": [ + 212, + 493, + 784, + 570 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nF D = \\lim _ {r \\rightarrow 0} \\frac {\\log N _ {r}}{\\log (1 / r)} \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 429, + 580, + 785, + 613 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Although the definition of FD based on self-similarity is simple and concise, its direct estimation becomes impractical when dealing with irregular images. To overcome this challenge and estimate the FDs of images, the box-counting method [24, 30] is employed.", + "bbox": [ + 212, + 625, + 782, + 686 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Box-counting Method: Consider a grayscale image $I$ with dimensions $M \\times M$ , where $L$ denotes the maximum gray level (typically $L = 255$ ). We can model the image $I$ as a three-dimensional space with $(x, y)$ indicating the two-dimensional position and a third coordinate $(z)$ denoting the gray value. Then the 3-D space is subdivided into smaller cubic regions, or \"boxes\", each with dimensions $k \\times k \\times h$ . Here, $k$ is a given scale used to be a multiple of the sidelength of a pixel in $(x, y)$ and $h$ can be a multiple of the gray level unit in $z$ -direction. Given $L$ and $k$ , the value of $h$ calculated using the following formula:", + "bbox": [ + 212, + 686, + 785, + 808 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nh = (L - 1) \\times k / M \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 433, + 825, + 784, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "Fractal Feature Maps for Segmentation of Tubular Structures", + "bbox": [ + 321, + 114, + 732, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Given a $k \\times k$ grid located at position $(i,j)$ , suppose that the minimum gray value is contained within the $m_{th}$ box $(m = \\lceil minG_{i,j} / h \\rceil)$ , and the maximum gray value within the $l_{th}$ box $(l = \\lceil maxG_{i,j} / h \\rceil)$ . The minimum number of boxes required to encompass all gray levels within the grid at $(i,j)$ is computed as:", + "bbox": [ + 212, + 146, + 787, + 208 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nn _ {r} (i, j) = l - m + 1 \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 431, + 224, + 785, + 239 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Considering all grids, the number of boxes that can cover all the patches is expressed as:", + "bbox": [ + 212, + 248, + 785, + 277 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nN _ {r} = \\sum_ {i, j} n _ {r} (i, j) \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 442, + 279, + 785, + 310 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $r = k / M$ . We can obtain a series of $N_{r}$ using differing values of $k$ . Finally, the fractal dimension can be estimated from the least-squares linear fitting of $\\log N_{r}$ versus $\\log (1 / r)$ , as illustrated in Equation (1). The flow of box-counting method is summarized in the supplementary material.", + "bbox": [ + 212, + 318, + 782, + 378 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Fractal Feature Map: Although fractal dimension and fractal features have been previously applied to classification tasks [32, 43], the inherent differences between classification and segmentation tasks preclude the direct application of image-level fractal dimension to segmentation endeavors. We extend the calculation of FD from the image-level to the pixel-level to generate the FFM of an image. As depicted in Fig. 2, the process begins with the utilization of a sliding window technique. Within a $5 \\times 5$ window, for example, the FD of this region is computed using the box-counting method. Subsequently, the window is shifted along both the horizontal and vertical directions with a step size of 1, resulting in the calculation of the FFM for the entire image. The algorithm for generating the FFM is summarized in Algorithm 1.", + "bbox": [ + 212, + 378, + 787, + 545 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The FFMs of the images (denoted as $FFM_{image}$ ) are incorporated into the segmentation model as additional input channels (Image, $FFM_{image}$ ), enhancing the focus on fractal structure and self-similarity after the normalization process. To mitigate the impact of image noise on the calculation of FD, we adopt the box-counting method proposed in [29] as they substitute the gray value with the standard deviation to make the method more robust.", + "bbox": [ + 212, + 546, + 472, + 757 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Algorithm 1 Generation of FFM.", + "text_level": 1, + "bbox": [ + 482, + 550, + 735, + 565 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Input: Image $I$ with size $M \\times N$ , window size $w \\times w$ .", + "bbox": [ + 496, + 570, + 785, + 595 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Output: Fractal feature map (FFM) of $I$ .", + "bbox": [ + 496, + 597, + 785, + 609 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Assign $FFM \\gets$ Float array of size $(M, N)$", + "bbox": [ + 496, + 611, + 785, + 625 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Assign padding size $p \\gets \\lfloor w / 2 \\rfloor$", + "bbox": [ + 496, + 625, + 705, + 638 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "$Pad_{I}\\gets$ Linear Padding $(I,p)$", + "bbox": [ + 496, + 638, + 692, + 652 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For $i\\gets 0$ to $M - 1$ do", + "bbox": [ + 496, + 652, + 658, + 665 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For $j\\gets 0$ to $N - 1$ do", + "bbox": [ + 527, + 667, + 689, + 680 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Selected region $R_{i,j}$", + "bbox": [ + 558, + 681, + 689, + 695 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nR _ {i, j} \\leftarrow P a d _ {I} [ i: i + w, j: j + w ]\n$$\n", + "text_format": "latex", + "bbox": [ + 558, + 695, + 767, + 708 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nF D \\leftarrow \\operatorname {B o x - c o u n t i n g} \\left(R _ {i, j}\\right)\n$$\n", + "text_format": "latex", + "bbox": [ + 558, + 708, + 732, + 722 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nF F M [ i, j ] \\leftarrow F D\n$$\n", + "text_format": "latex", + "bbox": [ + 558, + 722, + 674, + 734 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nF F M \\leftarrow \\operatorname {N o r m a l i z a t i o n} (F F M)\n$$\n", + "text_format": "latex", + "bbox": [ + 496, + 736, + 707, + 750 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.2 Multi-Decoder Network", + "text_level": 1, + "bbox": [ + 214, + 781, + 460, + 796 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In segmenting tubular structures, accurate boundary detection and preservation of global topological connectivity are crucial requirements. To meet these re", + "bbox": [ + 212, + 809, + 785, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "J. Huang et al.", + "bbox": [ + 271, + 114, + 372, + 128 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/d3dfcc36229c8ebbc942c24040f3ee904a819d0e6fa09cd93cb59b54bcabaa61.jpg", + "image_caption": [ + "Fig. 2: Workflow of computing FFM of an image." + ], + "image_footnote": [], + "bbox": [ + 303, + 146, + 702, + 247 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "requirements, we propose a new model that we refer to as Multi-Decoder Network (MD-Net).", + "bbox": [ + 212, + 273, + 782, + 303 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In addition to predicting the segmentation mask, we introduce an edge decoder and a skeleton decoder to generate the edge and skeleton of the tubular structure, respectively, as depicted in Fig. 3. When (Image, $FFM_{image}$ ) are fed into MD-Net, the Encoder employs convolution layers to extract a series of low-level to high-level image features. These features are simultaneously transmitted to the three Decoders by skip connections for prediction. Specifically, in the Encoder, each convolution layer involves the repeated application of two 3x3 convolutions, followed by a rectified linear unit (ReLU), and a 2x2 max pooling operation with a stride of 2 for down-sampling. As for the Decoders in MD-Net, each convolution layer includes an up-sampling of the feature map, followed by a 2x2 convolution that reduces the number of feature channels by half. Subsequently, the feature is concatenated with the corresponding feature copied from the Encoder. This concatenated feature is then processed using two 3x3 convolutions, each followed by a ReLU activation function.", + "bbox": [ + 212, + 303, + 784, + 513 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "During training, the ground truths for edges and skeletons are derived from the annotated masks by employing the findContours function from the OpenCV library and the skeletonization algorithms in the scikit-image library. We have visualized the boundaries and skeletons in the supplementary material to corroborate the accuracy of the extraction techniques. During the inference stage, the final structural segmentation is the output of the object decoder. Edge and skeleton predictions are only used in model training.", + "bbox": [ + 212, + 515, + 784, + 619 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Loss Function: Given an image $I$ with $N$ pixels, the segmentation ground truth is denoted as $y$ , where foreground and background pixels are labeled as 1 and 0, respectively. The prediction of the model is represented as $\\hat{y} \\in [0,1]$ , indicating the probabilities of individual pixels being classified as foreground. To compute the loss, we utilize the differentiable Soft IOU Loss [19], denoted as $\\mathcal{L}_{object}$ in Equation (5), where $\\eta$ is a smoothing coefficient.", + "bbox": [ + 212, + 619, + 784, + 712 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {o b j e c t}} = \\mathcal {L} _ {i o u} = 1 - \\frac {\\sum_ {i = 1} ^ {N} y _ {i} \\hat {y} _ {i}}{\\sum_ {i = 1} ^ {N} y _ {i} + \\hat {y} _ {i} - y _ {i} \\hat {y} _ {i} + \\eta} \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 339, + 719, + 784, + 758 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The BCE loss function is applied separately to the edge and skeleton segmentation tasks, denoted as $\\mathcal{L}_{edge},\\mathcal{L}_{skeleton}$ .", + "bbox": [ + 212, + 763, + 782, + 795 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {e d g e}}, \\mathcal {L} _ {\\text {s k e l e t o n}} = \\mathcal {L} _ {\\text {b c e}} = - \\frac {1}{N} \\sum_ {i = 1} ^ {N} y _ {i} \\cdot \\log (\\hat {y} _ {i}) + (1 - y _ {i}) \\cdot \\log (1 - \\hat {y} _ {i}) \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 256, + 801, + 784, + 842 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "Fractal Feature Maps for Segmentation of Tubular Structures", + "bbox": [ + 321, + 114, + 732, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "For MD-Net, which has three segmentation tasks, the loss function is composed as in Equation (8). The weights $\\alpha, \\beta$ and $\\gamma$ are typically set to 1.0, 0.5 and 0.5, respectively.", + "bbox": [ + 212, + 146, + 787, + 191 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {g l o b a l}} = \\alpha \\mathcal {L} _ {\\text {o b j e c t}} + \\beta \\mathcal {L} _ {\\text {e d g e}} + \\gamma \\mathcal {L} _ {\\text {s k e l e t o n}} \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 354, + 191, + 785, + 209 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "3.3 Fractal Feature Constrained Loss", + "text_level": 1, + "bbox": [ + 215, + 228, + 537, + 242 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The value of FD serves as an indicator of the texture complexity present in a given image region. Hence, it is advantageous to incorporate pixel-level FFM as weights of the loss function. This allows for a higher weight to be assigned to regions that exhibit greater complexity, as accurately segmenting such regions poses a greater challenge. Consequently, we calculate the FFM for each label associated with the image, denoted as $FFM_{label}$ , and employ $FFM_{label}$ as pixel-level weights within the loss function to guide the convergence of the model. Considering the specificity of the edge decoder and skeleton decoder, we have thus restricted the application of $FFM_{label}$ exclusively to the object decoder. The Constrained Loss of MD-Net is as follows:", + "bbox": [ + 212, + 252, + 787, + 402 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {c o n s t r a i n e d}} = \\alpha \\mathcal {L} _ {\\text {o b j e c t}} \\cdot F F M _ {\\text {l a b e l}} + \\beta \\mathcal {L} _ {\\text {e d g e}} + \\gamma \\mathcal {L} _ {\\text {s k e l e t o n}} \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 295, + 415, + 785, + 431 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 215, + 452, + 375, + 469 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.1 Datasets", + "text_level": 1, + "bbox": [ + 215, + 483, + 334, + 497 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We evaluate the FFM and MD-Net using five publicly available tubular datasets including the endoplasmic reticulum (ER) and mitochondrial (MITO) [11, 36] organelle network datasets, the Retinal OCT-Angiography vessel Segmentation (ROSE) and Structured Analysis of the Retina (STARE) [15, 37] retinal vessel", + "bbox": [ + 212, + 507, + 787, + 570 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/61582cd15f228a0900dcfb446a58ef3b1def5fb8657444d24a1edcf323d681ff.jpg", + "image_caption": [ + "Fig. 3: Overview and details of our proposed model MD-Net. The left part illustrates the overall structure of MD-Net. (Image, $FFM_{image}$ ) is processed as input by the Encoder, which extracts features of varying sizes from the input. These features are transmitted simultaneously to three Decoders through Skip Connection. The right part is a detailed description of the Encoder and Decoders. Each Decoder performs upsampling and concatenation operations similar to U-Net to obtain the predictions, comprising the edge, skeleton, and object." + ], + "image_footnote": [], + "bbox": [ + 225, + 595, + 782, + 743 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "J. Huang et al.", + "bbox": [ + 271, + 114, + 372, + 128 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "datasets, and the Massachusetts Roads (ROAD) [38] remote sensing dataset. By selecting these datasets, our aim is to comprehensively evaluate the performance of our methods across different types of structures and domains, providing a robust assessment of their capabilities. To further examine the influence of FFM, we included the non-tubular dataset NUCLEUS [5].", + "bbox": [ + 212, + 146, + 782, + 220 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "All images in ER, MITO, STARE, and NUCLEUS are randomly sampled patches with size of $256 \\times 256$ for training and sliding window with size of $256 \\times 256$ for validation and testing. A crop size of $256 \\times 256$ is used randomly during training and regularly during testing for the ROAD dataset as in [42]. Horizontal and vertical flipping as well as $90^{\\circ} / 180^{\\circ} / 270^{\\circ}$ rotation are used for data augmentation. In addition, data augmentation for ROSE is conducted by rotation of an angle from $-10^{\\circ}$ to $10^{\\circ}$ as described in [37]. Detailed partitions of the six datasets and experimental setup are shown in the supplementary material.", + "bbox": [ + 212, + 222, + 784, + 356 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.2 Performance Comparison", + "text_level": 1, + "bbox": [ + 215, + 378, + 470, + 393 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We select several segmentation networks for comparison, including the vanilla U-Net [44], U-Net++ [60], nnU-Net [20], and HR-Net [50]. Additionally, we compare our methods with SOTA models for diverse datasets. For the task of retinal vessel segmentation, we choose three-stage [53], OCTA-Net [37], GT-DLA [31], and AF-Net [45], as these networks are tailored for this task. Furthermore, we compare the performance of MD-Net in the ROAD dataset with DCU-Net [54], Dconn-Net [55], and the DSC-Net [42], a tubular structure segmentation network with specially designed loss $\\mathcal{L}_{tc}$ . To assess the generalization of FFM, we apply $FFM_{image}$ to both U-Net and HR-Net. All models are trained on the same dataset with the same hyperparameter settings.", + "bbox": [ + 212, + 402, + 784, + 551 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.3 Evaluation Metrics", + "text_level": 1, + "bbox": [ + 215, + 574, + 418, + 588 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "The models are assessed using three types of metrics: volumetric, topology, and distance.", + "bbox": [ + 212, + 598, + 782, + 626 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. Volumetric scores, including intersection-over-union (IoU), accuracy (ACC), centerlineDice (clDice) [46], and AUC. These metrics quantify the overlap and agreement between the prediction and ground truth.", + "2. Topology scores. To evaluate the topological correctness of the segmentations, we calculate the Betti Error [17] $\\beta$ for the sum of Betti Numbers $\\beta_0$ and $\\beta_{1}$ . In the evaluation of ROSE and STARE, $\\beta$ Error represents the $\\beta_{1}$ only.", + "3. Distance score: Hausdorff Distance (HD) [48] quantifies the accuracy of the boundary delineation and provides information about the spatial closeness of the predicted and ground truth boundaries." + ], + "bbox": [ + 212, + 628, + 789, + 760 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Since the nucleus is oval in shape and not tubular, no Betti Error and clDice are used in segmentation performance evaluation and no skeletal decoder is used in the MD-Net. In this study, the evaluation metrics IoU, ACC, AUC, and clDice, are expressed as percentages (\\%). The Hausdorff Distance (HD) is measured in pixels (px). All performance metrics are calculated for each image and averaged.", + "bbox": [ + 212, + 763, + 782, + 839 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "Fractal Feature Maps for Segmentation of Tubular Structures", + "bbox": [ + 321, + 114, + 730, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/47ba4810f869f35cebaf1d614893c6789fd89c6459f4ecc1821904b10c893231.jpg", + "table_caption": [ + "Table 1: Bold numbers indicate the best performance and italicized numbers represent the next best performance. The result marked with a star (*) represents the model input replaced by (Image) with (Image, $FFM_{image}$ ). $\\mathcal{L}_{tc}$ in (b) is Wasserstein-distance-based TCLoss proposed in [42]. $\\mathcal{L}_{\\text{constrained}}$ represents the integration of $FFM_{label}$ as a loss function weight, which leads to further improvement." + ], + "table_footnote": [], + "table_body": "
ModelLossIoU↑clDice↑ACC↑AUC↑βError↓HD↓IoU↑clDice↑ACC↑AUC↑βError↓HD↓
ERMITO
U-Net++ [60]Liou75.0294.6791.0497.3126.027.0479.7097.3098.1999.592.304.24
nnU-Net [20]Liou73.4194.5191.2987.9525.156.9679.2497.1698.0794.623.404.39
DSC-Net [42]Liou75.5194.4491.5697.0934.516.9280.3297.1698.1699.652.704.36
Dconn-Net [55]Liou75.9595.2491.5996.6020.316.9779.8297.2098.0399.542.904.32
AF-Net [45]Liou76.0194.5891.8197.3729.926.7980.3696.9498.0899.642.604.18
GT-DLA [31]Liou75.8994.9291.8297.3120.266.8480.2597.1698.1999.602.704.21
U-Net [44]Liou75.4494.6391.8297.3528.726.8779.7796.9198.0799.612.804.56
HR-Net [50]Liou75.8395.0891.7197.3622.576.9079.6397.2798.1499.622.904.25
U-Net*Liou76.5995.4392.0297.5620.786.8180.7197.4298.2199.632.704.27
HR-Net*Liou76.4395.4791.9597.5520.526.8380.6297.2998.1799.623.304.18
MD-Net*Lglobal77.0195.7892.0697.5919.106.7781.1197.7298.2599.662.204.16
MD-Net*Lconstrained77.0995.7492.1497.6519.526.7281.1897.6198.2699.652.804.14
ROSESTARE
three-stage [53]Liou62.1167.0791.7993.419.137.2866.8376.7194.8696.433.456.57
OCTA-Net [37]Liou63.2766.3292.3494.5310.227.1165.3776.7494.6295.443.226.47
DSC-Net [42]Liou63.1567.5991.9894.437.007.3667.0176.2694.9796.503.156.31
Dconn-Net [55]Liou60.3764.5188.5793.265.738.1866.1476.3394.7896.653.776.48
AF-Net [45]Liou62.3367.0591.7392.558.887.3267.2777.1794.9196.473.186.49
GT-DLA [31]Liou63.5766.3192.2794.189.897.1667.1276.3294.7295.793.126.63
U-Net [44]Liou61.5267.5391.3394.048.227.4266.1576.0594.6896.393.226.67
HR-Net [50]Liou63.0967.7192.0894.388.337.2667.0376.9294.9296.292.736.53
U-Net*Liou64.0767.9592.2594.477.887.1068.0777.3995.1596.632.776.33
HR-Net*Liou64.6869.4292.1494.776.017.2168.2277.4995.1396.523.216.32
MD-Net*Lglobal65.0769.7892.3694.884.227.1068.4977.7995.2096.912.576.29
MD-Net*Lconstrained65.1969.5892.4294.884.897.0668.7378.2095.2696.992.606.34
", + "bbox": [ + 218, + 226, + 785, + 487 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/eea3febbcb64f67f2d5bf3fa9e71808749737b63d2ea7bbf84c60110bb96dd1f.jpg", + "table_caption": [ + "(a) Quantitative experimental results for the ER, MITO, ROSE, and STARE dataset." + ], + "table_footnote": [], + "table_body": "
ModelLossIoU↑cDice↑ACC↑AUC↑βError↓HD↓ModelLossIoU↑ACC↑AUC↑HD↓
DCU-Net [54]Lce62.9286.9898.0398.342.568.04CS-Net [39]Liou77.9698.1298.584.13
DSC-Net [42]Lce63.9987.7498.0598.392.567.96Transnet [8]Liou78.2198.3498.644.32
DSC-Net [42]Ltc64.2287.6498.0598.462.457.34DU-Net [21]Liou78.0998.4698.944.19
Dconn-Net [55]Ltc65.1487.6197.3098.292.256.97U-Net++[60]Liou78.3398.5999.694.09
AF-Net [45]Ltc65.0387.2797.4198.382.217.29nnU-Net [20]Liou78.2898.5794.934.09
GT-DLA [31]Ltc64.7686.8097.4397.942.386.94OCTA-Net [37]Liou75.3098.3499.264.43
U-Net [44]Lce62.4786.8797.9798.292.618.11U-Net [44]Liou78.0298.5499.654.23
HR-Net [50]Lce62.8986.7498.0698.392.588.06HR-Net [50]Liou77.7098.5099.654.15
U-Net*Lce65.7487.7498.3898.722.486.98U-Net*Liou78.4698.5899.684.15
HR-Net*Lce65.6087.6198.3298.792.316.98HR-Net*Liou77.8398.5399.664.15
MD-Net*Lglobal66.0788.0898.4398.802.196.92MD-Net*Lglobal78.6698.5999.684.08
MD-Net*Lconstrained66.1688.0798.4298.772.066.97MD-Net*Lconstrained78.7798.5999.664.06
", + "bbox": [ + 220, + 512, + 781, + 628 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "(b) Results for ROAD dataset.", + "bbox": [ + 287, + 630, + 467, + 641 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "(c) Results for NUCLEUS dataset.", + "bbox": [ + 558, + 630, + 759, + 642 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.4 Configurations", + "text_level": 1, + "bbox": [ + 215, + 666, + 383, + 681 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "All models are implemented using PyTorch (version 1.12.0) and executed on 4 NVIDIA 3090 GPU cards. During the training phase, we employed the SGD optimizer with an initial learning rate of 0.05. To dynamically adjust the learning rate as the training progressed, we incorporated warm-up and exponential decay techniques. For all datasets, a fixed batch size of 32 is employed, ensuring consistency in the training process. Additionally, to prevent overfitting, a regularization weight of 0.0005 is applied. The entire training process spanned 50 epochs, providing sufficient iterations for model convergence and learning.", + "bbox": [ + 212, + 718, + 787, + 840 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "J. Huang et al.", + "bbox": [ + 271, + 114, + 372, + 128 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.5 Quantitative Evaluation", + "text_level": 1, + "bbox": [ + 215, + 146, + 460, + 161 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Based on the results presented in Tab. 1, several conclusions can be drawn.", + "bbox": [ + 212, + 172, + 756, + 186 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Performance of MD-Net: Our proposed model, MD-Net, demonstrates superior performance in terms of segmentation accuracy and topological continuity compared to the other models. This superiority is observed across five tubular datasets as well as one non-tubular dataset. We have also performed t-test to check whether improvements of our methods over competing methods in performance are statistically significant. Results are provided in supplementary.", + "bbox": [ + 212, + 188, + 785, + 279 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "In tubular datasets, MD-Net outperforms other methods in terms of segmentation results. Specifically, in the ER, MITO, ROSE, STARE, and ROAD datasets, MD-Net achieves improvements in IoU of $1.08\\%$ , $0.82\\%$ , $1.62\\%$ , $1.46\\%$ , and $1.02\\%$ , respectively, compared to existing SOTA methods. Furthermore, when considering topological continuity and boundary extraction, MD-Net demonstrates the best performance with the minimum $\\beta$ error and Hausdorff Distance. These results highlight the ability of MD-Net, equipped with an edge decoder and a skeleton decoder, to effectively capture edge and skeleton features of thin tubular structures, resulting in more accurate and continuous topology segmentation outcomes.", + "bbox": [ + 212, + 279, + 792, + 430 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Even in the non-tubular dataset NUCLEUS, MD-Net achieves the best segmentation results with an IoU of $78.77\\%$ , ACC of $98.59\\%$ , and HD of 4.06. This demonstrates the effectiveness of MD-Net and FFM in datasets with simpler structures.", + "bbox": [ + 212, + 430, + 785, + 489 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Generalization of FFM: $FFM_{image}$ is incorporated into U-Net and HR-Net as an additional input channel without modifying the training parameters. The results in Tab. 1 indicate that the inclusion of $FFM_{image}$ leads to improved segmentation performance for both U-Net and HR-Net.", + "bbox": [ + 212, + 491, + 785, + 551 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "For U-Net, the segmentation performance (IoU) shows improvements of $1.14\\%$ , $0.94\\%$ , $2.55\\%$ , $1.92\\%$ , $3.27\\%$ , and $0.44\\%$ on the ER, MITO, ROSE, STARE, ROAD, and NUCLEUS datasets with the assistance of $FFM_{image}$ , respectively. Similarly, for HR-Net, the IoU improves by $0.60\\%$ , $0.99\\%$ , $1.59\\%$ , $1.19\\%$ , $2.71\\%$ , and $0.13\\%$ on the same datasets, respectively. On average, $FFM_{image}$ brings an improvement of $1.96\\%$ and $1.42\\%$ in IoU for U-Net and HR-Net, respectively, in the segmentation task of tubular datasets.", + "bbox": [ + 212, + 551, + 789, + 657 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "In conclusion, FFM demonstrates notable generalization capabilities, leading to enhanced performance across various models. Particularly in datasets with intricate structures, FFM consistently outperforms the original models in segmentation tasks. This improvement can be attributed to the inherent characteristics of FD involved in FFM.", + "bbox": [ + 212, + 657, + 785, + 734 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "$FFM_{label}$ and $\\mathcal{L}_{global}$ : Fractal Dimension is a significant metric used to quantify the complexity of image textures. In the context of model training, we leverage the $FFMs_{label}$ as weights for the loss function. This method enables us to guide the model's focus towards regions with higher complexity, as indicated by a higher FD. Through extensive experimentation, we have observed that incorporating FFM-based weights into the loss function yields improvements in both the training performance and efficiency of the model, as depicted in Tab. 1.", + "bbox": [ + 212, + 734, + 785, + 840 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "Fractal Feature Maps for Segmentation of Tubular Structures", + "bbox": [ + 321, + 114, + 730, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 116, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/baed5fb96d32065d30b487232642e5f700fec30761cd65502c23c33f097548b3.jpg", + "image_caption": [ + "Fig. 4: Comparison of segmentation results. (a) Image. (b) Label. (c) Results of U-Net. (d) Results of existing SOTA approaches. From top to bottom, it's AF-Net, AF-Net, GT-DLA, AF-Net, and Dconn-Net. (e) Results of U-Net*. (f) Results of MD-Net*. Red: true positive. Green: false negative. Blue: false positive. In the ER, MITO, ROSE, STAR, and ROAD rows, columns (b) through (f) display the magnified results of the regions demarcated by yellow squares in the corresponding images of column (a)." + ], + "image_footnote": [], + "bbox": [ + 292, + 146, + 709, + 409 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "These findings highlight the tangible benefits of including $FFM_{label}$ as a weight parameter in optimizing the training process.", + "bbox": [ + 212, + 518, + 787, + 550 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "4.6 Qualitative Evaluation", + "text_level": 1, + "bbox": [ + 215, + 570, + 447, + 585 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "To facilitate a more comprehensive comparison of the experimental outcomes across various models, we employed visualizations to present the segmentation results, as depicted in Fig. 4. The visual analysis provides a clear and intuitive representation of the model's performance. The observed results validate that the integration of FFM leads to enhanced segmentation outcomes in terms of both edge accuracy and topological continuity. More visualization results can be found in the supplementary material.", + "bbox": [ + 212, + 595, + 787, + 702 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "5 Ablation Study", + "text_level": 1, + "bbox": [ + 215, + 723, + 403, + 742 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "5.1 Effectiveness of Fractal Feature Maps", + "text_level": 1, + "bbox": [ + 215, + 753, + 571, + 771 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "To ascertain that the observed performance improvement of the model is attributed to the FFM and not the increase of input channels, we replace the $FFM_{image}$ with the image, thereby transforming the input (image, $FFM_{image}$ ) to (image, image). Additionally, we computed the image's Hurst feature (HF,", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "J. Huang et al.", + "bbox": [ + 271, + 114, + 372, + 128 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "classical fractal analysis), mean feature (MF, directional analysis), and contrast feature (CF, directional analysis) as introduced by [22]. These features replaced the $FFM_{image}$ as input and were trained and evaluated on U-Net and MD-Net. Experimental results demonstrated the advantages of the $FFM_{image}$ as shown in Tab. 2. This finding indicates that the inclusion of the FFM plays an effective role in achieving better performance in the segmentation task.", + "bbox": [ + 212, + 146, + 787, + 238 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "5.2 Robustness of Fractal Feature Maps", + "text_level": 1, + "bbox": [ + 214, + 257, + 560, + 272 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "The computation of the FFM is influenced by the selection of window size and step size. We used different window and step sizes to generate different $FFMs_{image}$ and tested their performance. The results in Tab. 3 demonstrate the robustness of FFM. When used for training the U-Net and MD-Net, the $FFMs_{image}$ calculated using different window sizes consistently yield results with a variation range of $0.3\\%$ on the ER and STARE datasets. Similarly, applying different steps during the calculation of FFMs does not lead to significant changes in performance. It should be noted that the efficiency of computing can be substantially improved by increasing the step size.", + "bbox": [ + 212, + 279, + 787, + 414 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "The numerical results in Tab. 2 and Tab. 3 are calculated by volumetric score IoU. Results with more metrics can be found in the supplementary material.", + "bbox": [ + 212, + 415, + 785, + 445 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "5.3 Decoders of MD-Net", + "text_level": 1, + "bbox": [ + 214, + 465, + 434, + 479 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "To evaluate the effectiveness of the edge decoder and the skeleton decoder, a comparative ablation analysis was conducted between MD-Net and U-Net on the ER and STARE datasets. Subsequently, we proceeded to remove either the edge decoder or the skeleton decoder from MD-Net and trained the modified models. The results shown in Tab. 4 confirm that both the edge decoder and the skeleton decoder contribute to improving the segmentation performance. It is noteworthy that the best performance is achieved when all components are utilized simultaneously.", + "bbox": [ + 212, + 488, + 787, + 609 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "5.4 Limitations of Fractal Feature Map", + "text_level": 1, + "bbox": [ + 214, + 628, + 553, + 645 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "The time complexity of FFMs' computation is associated with the size of image $(M\\times N)$ , window size $w$ , and step size $S$ . Its time complexity is represented", + "bbox": [ + 214, + 651, + 787, + 681 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/3b9e74d68964cac9481a8c3df221aa9b538a5caebf82814c86c997491c7a7fa0.jpg", + "table_caption": [ + "Table 2: Ablation study of FFM in U-Net and MD-Net. The values within the parentheses in the (image, $FFM_{image}$ ) column represent the percentage improvement of IoU compared to the baseline results listed in the (image) column." + ], + "table_footnote": [], + "table_body": "
ModelDatasetInput
(image)(image,image)(image,HF)(image,MF)(image,CF)(image,FFM image)
U-NetER75.4475.6175.8975.8675.1076.59(+1.14%)
MITO79.7780.2880.5180.2380.3880.71(+0.94%)
ROSE61.5262.5762.6162.5862.3764.07(+2.55%)
STARE66.1566.0865.5466.1466.8368.07(+1.92%)
ROAD62.4763.4464.9364.5864.2565.74(+3.27%)
MD-NetER76.2876.3476.2475.9775.5277.01(+0.73%)
MITO80.2880.3179.9780.4180.3281.11(+0.83%)
ROSE63.3163.3062.7762.8962.9065.07(+1.76%)
STARE66.4666.5766.4367.0367.2568.49(+2.03%)
ROAD64.7965.0465.2365.1265.1566.07(+1.28%)
", + "bbox": [ + 285, + 742, + 715, + 857 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "Fractal Feature Maps for Segmentation of Tubular Structures", + "bbox": [ + 321, + 114, + 730, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 784, + 126 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/c574c61e4b5e72fdde153e1297964d6f7b3e109af3c68c3037aa279d582d3cd6.jpg", + "table_caption": [ + "Table 3: Ablation of step size and window size in U-Net and MD-Net." + ], + "table_footnote": [], + "table_body": "
ModelStep SizeWindow SizeERSTAREModelStep SizeWindow SizeERSTARE
U-Net*11176.7567.87MD-Net*11177.0368.41
1976.7367.911977.0168.21
1776.6567.891776.9268.50
1576.5968.071577.0168.49
2576.4967.872576.9168.46
3576.6867.783576.8868.32
4576.6167.784576.9068.30
", + "bbox": [ + 285, + 170, + 715, + 244 + ], + "page_idx": 13 + }, + { + "type": "table", + "img_path": "images/25730273b3ded7e0de4638e010f5b5611241aa8e04bb93a0284508e168ec2945.jpg", + "table_caption": [ + "Table 4: Ablation study of MD-Net." + ], + "table_footnote": [], + "table_body": "
DatasetModelFFMEdge DecoderSkeleton DecoderIoU↑clDice↑ACC↑AUC↑βError↓HD↓
ERU-Net75.4494.6391.8297.3528.726.87
MD-Net✓✓76.2895.2792.0097.3922.556.78
U-Net✓76.5995.4392.0297.5620.786.81
MD-Net✓✓76.9795.6092.0997.5922.426.71
MD-Net✓✓76.8595.5592.1397.5120.636.74
MD-Net✓✓✓77.0995.7492.1497.6519.526.72
STAREU-Net66.1576.0594.6896.393.226.67
MD-Net✓✓66.4676.5494.7696.423.276.47
U-Net✓68.0777.3995.1596.632.776.33
MD-Net✓✓68.2677.8095.2096.772.656.29
MD-Net✓✓68.3977.7695.2096.893.226.31
MD-Net✓✓✓68.7378.2095.2696.992.606.34
", + "bbox": [ + 285, + 280, + 712, + 429 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "as: $O\\left(\\frac{M \\cdot N}{S^2} \\cdot \\left(w^2 \\cdot \\log(w)\\right)\\right)$ . Calculation of FFM indeed incurs computational overhead but is optimized by using multithreading and larger $S$ . Currently, the generation of FFMs for a $256 \\times 256$ image takes about 150 milliseconds on a Xeon 8336C CPU, given a step size of 1 and a window size of 5. Detailed information on training and inference time can be found in the supplementary material.", + "bbox": [ + 212, + 441, + 787, + 521 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "6 Conclusions", + "text_level": 1, + "bbox": [ + 215, + 554, + 369, + 570 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In this study, we propose a method that uses fractal feature maps (FFMs) along with a multi-decoder network (MD-Net) for the semantic segmentation of tubular structures. FFMs are used to capture the texture and self-similarity of image regions or label regions at the pixel-level through the fractal dimension. $FFM_{image}$ are used as an additional input to enhance the segmentation model's perception of tubular structures and $FFM_{label}$ are used as a weight for the loss function to guide the model training. MD-Net improves the quality of segmentation by simultaneously predicting edges and skeletons through the incorporation of an edge decoder and a skeleton decoder. The proposed method is evaluated on five datasets of tubular structures and one dataset of non-tubular structures. The results demonstrate the superiority of MD-Net with FFM in the segmentation of tubular structures and stable performance in the segmentation of non-tubular structures. Furthermore, the performance improvement observed when incorporating the FFM module into vanilla U-Net and HR-Net underscores the broad applicability and potential of incorporating fractal information into deep learning models.", + "bbox": [ + 212, + 598, + 787, + 843 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "J. Huang et al.", + "bbox": [ + 271, + 114, + 372, + 128 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 217, + 143, + 401, + 162 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "This work was supported in part by the National Natural Science Foundation of China (grants 92354307, 91954201, 31971289, 32101216), the Strategic Priority Research Program of the Chinese Academy of Sciences (grant XDB37040402) and the Fundamental Research Funds for the Central Universities (grant E3E45201X2).", + "bbox": [ + 212, + 180, + 787, + 256 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 215, + 282, + 321, + 299 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Alvarez, L., Trujillo, A., Cuenca, C., González, E., Esclarín, J., Gomez, L., Mazarra, L., Aleman-Flores, M., Tahoces, P.G., Carreira, J.M.: Tracking the aortic lumen geometry by optimizing the 3d orientation of its cross-sections. In: Medical Image Computing and Computer-Assisted Intervention- MICCAI 2017: 20th International Conference, Quebec City, QC, Canada, September 11-13, 2017, Proceedings, Part II 20. pp. 174-181. Springer (2017)", + "2. Antiga, L., Ene-Iordache, B., Remuzzi, A.: Computational geometry for patient-specific reconstruction and meshing of blood vessels from mr and ct angiography. IEEE Transactions on Medical Imaging 22(5), 674-684 (2003)", + "3. Araújo, R.J., Cardoso, J.S., Oliveira, H.P.: Topological similarity index and loss function for blood vessel segmentation. arXiv preprint arXiv:2107.14531 (2021)", + "4. Bauer, C., Pock, T., Sorantin, E., Bischof, H., Beichel, R.: Segmentation of interwoven 3d tubular tree structures utilizing shape priors and graph cuts. Medical Image Analysis 14(2), 172-184 (2010)", + "5. Caicedo, J.C., Goodman, A., Karhohs, K.W., Cimini, B.A., Ackerman, J., Haghighi, M., Heng, C., Becker, T., Doan, M., McQuin, C., et al.: Nucleus segmentation across imaging experiments: the 2018 data science bowl. Nature Methods 16(12), 1247-1253 (2019)", + "6. Caselles, V., Kimmel, R., Sapiro, G.: Geodesic active contours. International Journal of Computer Vision 22, 61-79 (1997)", + "7. Challoo, M., Gao, Y., Busch, A.: Distinctive phase interdependency model for retinal vasculature delineation in oct-angiography images. IEEE Transactions on Medical Imaging (2023)", + "8. Chen, J., Lu, Y., Yu, Q., Luo, X., Adeli, E., Wang, Y., Lu, L., Yuille, A.L., Zhou, Y.: Transunet: Transformers make strong encoders for medical image segmentation. arXiv preprint arXiv:2102.04306 (2021)", + "9. Dai, J., Qi, H., Xiong, Y., Li, Y., Zhang, G., Hu, H., Wei, Y.: Deformable convolutional networks. In: Proceedings of the IEEE international conference on computer vision. pp. 764-773 (2017)", + "0. Dong, S., Pan, Z., Fu, Y., Yang, Q., Gao, Y., Yu, T., Shi, Y., Zhuo, C.: Deu-net 2.0: Enhanced deformable u-net for 3d cardiac cine migraine segmentation. Medical Image Analysis 78, 102389 (2022)", + "11. Guo, Y., Huang, J., Zhou, Y., Luo, Y., Li, W., Yang, G.: Segmentation of intracellular structures in fluorescence microscopy images by fusing low-level features. In: Pattern Recognition and Computer Vision: 4th Chinese Conference, PRCV 2021, Beijing, China, October 29–November 1, 2021, Proceedings, Part III 4. pp. 386–397. Springer (2021)" + ], + "bbox": [ + 218, + 318, + 785, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "Fractal Feature Maps for Segmentation of Tubular Structures", + "bbox": [ + 321, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 785, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "12. Gupta, S., Zhang, Y., Hu, X., Prasanna, P., Chen, C.: Topology-aware uncertainty for image segmentation. Advances in Neural Information Processing Systems 36 (2024)", + "13. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 770-778 (2016)", + "14. He, Y., Ge, R., Wu, J., Coatrieux, J.L., Shu, H., Chen, Y., Yang, G., Li, S.: Thin semantics enhancement via high-frequency priori rule for thin structures segmentation. In: 2021 IEEE International Conference on Data Mining (ICDM). pp. 1096-1101. IEEE (2021)", + "15. Hoover, A., Goldbaum, M.: Locating the optic nerve in a retinal image using the fuzzy convergence of the blood vessels. IEEE Transactions on Medical Imaging 22(8), 951-958 (2003)", + "16. Hotamisligil, G.S.: Endoplasmic reticulum stress and the inflammatory basis of metabolic disease. Cell 140(6), 900-917 (2010)", + "17. Hu, X., Li, F., Samaras, D., Chen, C.: Topology-preserving deep image segmentation. Advances in neural information processing systems 32 (2019)", + "18. Hu, X., Wang, Y., Fuxin, L., Samaras, D., Chen, C.: Topology-aware segmentation using discrete Morse theory. arXiv preprint arXiv:2103.09992 (2021)", + "19. Huang, Y., Tang, Z., Chen, D., Su, K., Chen, C.: Batching soft iou for training semantic segmentation networks. IEEE Signal Processing Letters 27, 66-70 (2019)", + "20. Isensee, F., Jaeger, P.F., Kohl, S.A., Petersen, J., Maier-Hein, K.H.: nnu-net: a self-configuring method for deep learning-based biomedical image segmentation. Nature Methods 18(2), 203-211 (2021)", + "21. Jin, Q., Meng, Z., Pham, T.D., Chen, Q., Wei, L., Su, R.: Dunet: A deformable network for retinal vessel segmentation. Knowledge-Based Systems 178, 149-162 (2019)", + "22. Kaplan, L.M.: Extended fractal analysis for texture classification and segmentation. IEEE transactions on image processing 8(11), 1572-1585 (1999)", + "23. Kaplan, L.M., Kuo, C.C.: Extending self-similarity for fractional brownian motion. IEEE Transactions on Signal Processing 42(12), 3526-3530 (1994)", + "24. Konatar, I., Popovic, T., Popovic, N.: Box-counting method in python for fractal analysis of biomedical images. In: 2020 24th International Conference on Information Technology (IT). pp. 1-4. IEEE (2020)", + "25. Kreitner, L., Paetzold, J.C., Rauch, N., Chen, C., Hagag, A.M., Fayed, A.E., Sivaprasad, S., Rausch, S., Weichsel, J., Menze, B.H., et al.: Synthetic optical coherence tomography angiographs for detailed retinal vessel segmentation without human annotations. IEEE Transactions on Medical Imaging (2024)", + "26. La Barbera, G., Rouet, L., Boussaid, H., Lubet, A., Kassir, R., Sarnacki, S., Gori, P., Bloch, I.: Tubular structures segmentation of pediatric abdominal-visceral cect images with renal tumors: assessment, comparison and improvement. Medical Image Analysis 90, 102986 (2023)", + "27. Laptev, I., Mayer, H., Lindeberg, T., Eckstein, W., Steger, C., Baumgartner, A.: Automatic extraction of roads from aerial images based on scale space and snakes. Machine Vision and Applications 12, 23-31 (2000)", + "28. Lee, C.A., Blackstone, C.: Er morphology and endo-lysosomal crosstalk: Functions and disease implications. Biochimica et Biophysica Acta (BBA)-Molecular and Cell Biology of Lipids 1865(1), 158544 (2020)", + "29. Lee, W.L., Hsieh, K.S.: A robust algorithm for the fractal dimension of images and its applications to the classification of natural images and ultrasonic liver images. Signal Processing 90(6), 1894-1904 (2010)" + ], + "bbox": [ + 215, + 146, + 785, + 840 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "J. Huang et al.", + "bbox": [ + 271, + 114, + 372, + 128 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "30. Li, J., Du, Q., Sun, C.: An improved box-counting method for image fractal dimension estimation. Pattern Recognition 42(11), 2460-2469 (2009)", + "31. Li, Y., Zhang, Y., Liu, J.Y., Wang, K., Zhang, K., Zhang, G.S., Liao, X.F., Yang, G.: Global transformer and dual local attention network via deep-shallow hierarchical feature fusion for retinal vessel segmentation. IEEE Transactions on Cybernetics 53(9), 5826-5839 (2023)", + "32. Lin, P.L., Huang, P.W., Lee, C.H., Wu, M.T.: Automatic classification for solitary pulmonary nodule in ct image by fractal analysis based on fractional brownian motion model. Pattern Recognition 46(12), 3279-3287 (2013)", + "33. Lin, P., Huang, P., Huang, P., Hsu, H.: Alveolar bone-loss area localization in periodontitis radiographs based on threshold segmentation with a hybrid feature fused of intensity and the h-value of fractional brownian motion model. Computer Methods and Programs in Biomedicine 121(3), 117-126 (2015)", + "34. Long, J., Shelhamer, E., Darrell, T.: Fully convolutional networks for semantic segmentation. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 3431-3440 (2015)", + "35. Lopes, R., Betrouni, N.: Fractal and multifractal analysis: A review. Medical Image Analysis 13(4), 634-649 (2009)", + "36. Luo, Y., Guo, Y., Li, W., Liu, G., Yang, G.: Fluorescence microscopy image datasets for deep learning segmentation of intracellular organelle networks (2020). https://doi.org/10.21227/t2he-zn97, https://dx.doi.org/10.21227/t2he-zn97", + "37. Ma, Y., Hao, H., Xie, J., Fu, H., Zhang, J., Yang, J., Wang, Z., Liu, J., Zheng, Y., Zhao, Y.: Rose: a retinal oct-angiography vessel segmentation dataset and new model. IEEE Transactions on Medical Imaging 40(3), 928–939 (2020)", + "38. Mnih, V.: Machine learning for aerial image labeling. University of Toronto (Canada) (2013)", + "39. Mou, L., Zhao, Y., Chen, L., Cheng, J., Gu, Z., Hao, H., Qi, H., Zheng, Y., Frangi, A., Liu, J.: Cs-net: Channel and spatial attention network for curvilinear structure segmentation. In: Medical Image Computing and Computer Assisted Intervention-MICCAI 2019: 22nd International Conference, Shenzhen, China, October 13-17, 2019, Proceedings, Part I 22. pp. 721-730. Springer (2019)", + "40. Nain, D., Yezzi, A., Turk, G.: Vessel segmentation using a shape driven flow. In: Medical Image Computing and Computer-Assisted Intervention-MICCAI 2004: 7th International Conference, Saint-Malo, France, September 26-29, 2004. Proceedings, Part I 7. pp. 51-59. Springer (2004)", + "41. Pentland, A.P.: Fractal-based description of natural scenes. IEEE Transactions on Pattern Analysis and Machine Intelligence (6), 661-674 (1984)", + "42. Qi, Y., He, Y., Qi, X., Zhang, Y., Yang, G.: Dynamic snake convolution based on topological geometric constraints for tubular structure segmentation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 6070-6079 (2023)", + "43. Roberto, G.F., Lumini, A., Neves, L.A., do Nascimento, M.Z.: Fractal neural network: A new ensemble of fractal geometry and convolutional neural networks for the classification of histology images. Expert Systems with Applications 166, 114103 (2021)", + "44. Ronneberger, O., Fischer, P., Brox, T.: U-net: Convolutional networks for biomedical image segmentation. In: Medical Image Computing and Computer-Assisted Intervention-MICCAI 2015: 18th International Conference, Munich, Germany, October 5-9, 2015, Proceedings, Part III 18. pp. 234-241. Springer (2015)" + ], + "bbox": [ + 212, + 146, + 787, + 839 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "Fractal Feature Maps for Segmentation of Tubular Structures", + "bbox": [ + 321, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "45. Shi, T., Ding, X., Zhou, W., Pan, F., Yan, Z., Bai, X., Yang, X.: Affinity feature strengthening for accurate, complete and robust vessel segmentation. IEEE Journal of Biomedical and Health Informatics 27(8), 4006-4017 (2023)", + "46. Shit, S., Paetzold, J.C., Sekuboyina, A., Ezhov, I., Unger, A., Zhylka, A., Pluim, J.P., Bauer, U., Menze, B.H.: cldice-a novel topology-preserving loss function for tubular structure segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16560-16569 (2021)", + "47. Sironi, A., Lepetit, V., Fua, P.: Multiscale centerline detection by learning a scale-space distance transform. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 2697-2704 (2014)", + "48. Taha, A.A., Hanbury, A.: Metrics for evaluating 3d medical image segmentation: analysis, selection, and tool. BMC Medical Imaging 15(1), 1-28 (2015)", + "49. Wang, C., Hayashi, Y., Oda, M., Itoh, H., Kitasaka, T., Frangi, A.F., Mori, K.: Tubular structure segmentation using spatially connected network with radial distance loss for 3d medical images. In: Medical Image Computing and Computer Assisted Intervention-MICCAI 2019: 22nd International Conference, Shenzhen, China, October 13-17, 2019, Proceedings, Part VI 22. pp. 348-356. Springer (2019)", + "50. Wang, J., Sun, K., Cheng, T., Jiang, B., Deng, C., Zhao, Y., Liu, D., Mu, Y., Tan, M., Wang, X., et al.: Deep high-resolution representation learning for visual recognition. IEEE Transactions on Pattern Analysis and Machine Intelligence 43(10), 3349-3364 (2020)", + "51. Wang, Y., Wei, X., Liu, F., Chen, J., Zhou, Y., Shen, W., Fishman, E.K., Yuille, A.L.: Deep distance transform for tubular structure segmentation in ct scans. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3833-3842 (2020)", + "52. Wong, C.C., Vong, C.M.: Persistent homology based graph convolution network for fine-grained 3d shape segmentation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 7098-7107 (2021)", + "53. Yan, Z., Yang, X., Cheng, K.T.: A three-stage deep learning model for accurate retinal vessel segmentation. IEEE Journal of Biomedical and Health Informatics 23(4), 1427-1436 (2018)", + "54. Yang, X., Li, Z., Guo, Y., Zhou, D.: Dcu-net: A deformable convolutional neural network based on cascade u-net for retinal vessel segmentation. Multimedia Tools and Applications 81(11), 15593-15607 (2022)", + "55. Yang, Z., Farsiu, S.: Directional connectivity-based segmentation of medical images. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11525-11535 (2023)", + "56. Yim, P.J., Cebral, J.J., Mullick, R., Marcos, H.B., Choyke, P.L.: Vessel surface reconstruction with a tubular deformable model. IEEE Transactions on Medical Imaging 20(12), 1411-1421 (2001)", + "57. Yu, F., Koltun, V., Funkhouser, T.: Dilated residual networks. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 472-480 (2017)", + "58. Zhao, Y., Zheng, Y., Liu, Y., Zhao, Y., Luo, L., Yang, S., Na, T., Wang, Y., Liu, J.: Automatic 2-d/3-d vessel enhancement in multiple modality images using a weighted symmetry filter. IEEE Transactions on Medical Imaging 37(2), 438-450 (2017)", + "59. Zhou, Y., Huang, J., Wang, C., Song, L., Yang, G.: Xnet: Wavelet-based low and high frequency fusion networks for fully- and semi-supervised semantic segmentation of biomedical images. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 21085-21096 (October 2023)" + ], + "bbox": [ + 215, + 146, + 785, + 839 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "J. Huang et al.", + "bbox": [ + 271, + 114, + 372, + 128 + ], + "page_idx": 17 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "60. Zhou, Z., Rahman Siddiquee, M.M., Tajbakhsh, N., Liang, J.: Unet++: A nested u-net architecture for medical image segmentation. In: Deep Learning in Medical Image Analysis and Multimodal Learning for Clinical Decision Support: 4th International Workshop, DLMIA 2018, and 8th International Workshop, ML-CDS 2018, Held in Conjunction with MICCAI 2018, Granada, Spain, September 20, 2018, Proceedings 4. pp. 3-11. Springer (2018)", + "61. Zhuang, Z., Lei, N., Joseph Raj, A.N., Qiu, S.: Application of fractal theory and fuzzy enhancement in ultrasound image segmentation. Medical & Biological Engineering & Computing 57, 623-632 (2019)" + ], + "bbox": [ + 215, + 146, + 787, + 272 + ], + "page_idx": 18 + }, + { + "type": "header", + "text": "Fractal Feature Maps for Segmentation of Tubular Structures", + "bbox": [ + 321, + 114, + 730, + 128 + ], + "page_idx": 18 + }, + { + "type": "page_number", + "text": "19", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 18 + } +] \ No newline at end of file diff --git a/2024/Representing Topological Self-Similarity Using Fractal Feature Maps for Accurate Segmentation of Tubular Structures/494e6709-0fe6-4415-a992-88fda17780d8_model.json b/2024/Representing Topological Self-Similarity Using Fractal Feature Maps for Accurate Segmentation of Tubular Structures/494e6709-0fe6-4415-a992-88fda17780d8_model.json new file mode 100644 index 0000000000000000000000000000000000000000..fb03415b892769212ffc2c890991248665f7fb71 --- /dev/null +++ b/2024/Representing Topological Self-Similarity Using Fractal Feature Maps for Accurate Segmentation of Tubular Structures/494e6709-0fe6-4415-a992-88fda17780d8_model.json @@ -0,0 +1,2702 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.223, + 0.141, + 0.782, + 0.207 + ], + "angle": 0, + "content": "Representing Topological Self-Similarity Using Fractal Feature Maps for Accurate Segmentation of Tubular Structures" + }, + { + "type": "text", + "bbox": [ + 0.25, + 0.234, + 0.756, + 0.268 + ], + "angle": 0, + "content": "Jiaxing Huang\\(^{1,2}\\), Yanfeng Zhou\\(^{1,2}\\), Yaoru Luo\\(^{1,2}\\), Guole Liu\\(^{1,2}\\), Heng Guo\\(^{3,4}\\), and Ge Yang\\(^{1,2}\\)" + }, + { + "type": "text", + "bbox": [ + 0.233, + 0.277, + 0.772, + 0.292 + ], + "angle": 0, + "content": "\\(^{1}\\) Institute of Automation, Chinese Academy of Sciences, 100190, Beijing, China" + }, + { + "type": "text", + "bbox": [ + 0.245, + 0.292, + 0.758, + 0.305 + ], + "angle": 0, + "content": "\\(^{2}\\) School of Artificial Intelligence, University of Chinese Academy of Sciences" + }, + { + "type": "text", + "bbox": [ + 0.383, + 0.305, + 0.621, + 0.32 + ], + "angle": 0, + "content": "3 DAMO Academy, Alibaba Group" + }, + { + "type": "text", + "bbox": [ + 0.367, + 0.32, + 0.637, + 0.334 + ], + "angle": 0, + "content": "4 Hupan Lab, 310023, Hangzhou, China" + }, + { + "type": "text", + "bbox": [ + 0.362, + 0.334, + 0.642, + 0.348 + ], + "angle": 0, + "content": "{huangjiaxing2021, ge.yang}@ia.ac.cn" + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.387, + 0.744, + 0.664 + ], + "angle": 0, + "content": "Abstract. Accurate segmentation of long and thin tubular structures is required in a wide variety of areas such as biology, medicine, and remote sensing. The complex topology and geometry of such structures often pose significant technical challenges. A fundamental property of such structures is their topological self-similarity, which can be quantified by fractal features such as fractal dimension (FD). In this study, we incorporate fractal features into a deep learning model by extending FD to the pixel-level using a sliding window technique. The resulting fractal feature maps (FFMs) are then incorporated as additional input to the model and additional weight in the loss function to enhance segmentation performance by utilizing the topological self-similarity. Moreover, we extend the U-Net architecture by incorporating an edge decoder and a skeleton decoder to improve boundary accuracy and skeletal continuity of segmentation, respectively. Extensive experiments on five tubular structure datasets validate the effectiveness and robustness of our approach. Furthermore, the integration of FFMs with other popular segmentation models such as HR-Net also yields performance enhancement, suggesting FFM can be incorporated as a plug-in module with different model architectures. Code and data are openly accessible at https://github.com/cbmi-group/FFM-Multi-Decoder-Network." + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.678, + 0.744, + 0.706 + ], + "angle": 0, + "content": "Keywords: Tubular structures \\(\\cdot\\) Topological self-similarity \\(\\cdot\\) Fractal feature map \\(\\cdot\\) Deep learning \\(\\cdot\\) Plug-in module" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.732, + 0.377, + 0.749 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.788, + 0.843 + ], + "angle": 0, + "content": "Accurate segmentation of tubular structures is of significant importance across a wide variety of areas. In the area of biological research, for example, the accurate segmentation of tubular structures such as the endoplasmic reticulum (ER, Fig. 1) is critical to the study of related human disease mechanisms [16,28]. In the area of clinical research, the accurate segmentation of blood vessels (Fig. 1) is" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.374, + 0.129 + ], + "angle": 0, + "content": "J. Huang et al." + }, + { + "type": "image", + "bbox": [ + 0.306, + 0.145, + 0.703, + 0.299 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.31, + 0.789, + 0.368 + ], + "angle": 0, + "content": "Fig. 1: Topological self-similarity in tubular structures. If we consider \"one junction with multiple edges\" as a basic component, complex tubular structures have similar components at different scales. The images on the right are magnified views of the rectangular regions on the left." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.384, + 0.789, + 0.476 + ], + "angle": 0, + "content": "essential to the early diagnosis of diseases such as retinopathy and stroke [39,58]. Similarly, in the area of remote sensing, the accurate extraction of roads from aerial imagery is essential for navigation and route planning [27]. However, accurate segmentation of tubular structures from images remains challenging due to factors such as their complex morphology and geometry, low image signal-to-noise ratio, and poor image contrast." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.477, + 0.789, + 0.718 + ], + "angle": 0, + "content": "A wide variety of techniques have been developed for the segmentation of tubular structures. Classical methods rely on manually crafted features such as intensity, texture, and shape. For example, previous studies [2,7,40,56] have utilized deformable shape models to fit tubular structures, leveraging their geometric properties. However, these techniques often cannot handle challenges posed by factors such as poor contrast, high noise and complex background. Deep learning methods [13, 25, 34, 44, 59] have revolutionized image segmentation and achieved substantial improvements in segmentation performance. Recent studies on using deep learning models for segmentation of tubular structures have focused on optimization of loss functions [3, 26, 46, 49, 51] and refinement of model structures [10, 12, 18, 37, 42]. However, these studies primarily aim to achieve high segmentation performance utilizing a limited input of images without providing additional information to their segmentation models. Tubular structures exhibit distinct topology and geometry that are vital for segmentation. In this study, we explore using characteristics of their structure to assist deep learning models in segmentation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.789, + 0.842 + ], + "angle": 0, + "content": "One notable characteristic of tubular structures is their topological self-similarity, namely large and complex tubular structures exhibit similar topological patterns at different scales. For example, Fig. 1 shows that if \"one junction with multiple edges\" is considered a primary structural component, tubular structure entities such as the ER network and the retinal blood vessel network exhibit similar topology at both global and local scales. To quantify the topological self-similarity of tubular structures, we utilize the fractal theory, which characterizes self-similarity of intricate structures at different scales [35,61]. Cen-" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.322, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Fractal Feature Maps for Segmentation of Tubular Structures" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.177 + ], + "angle": 0, + "content": "Intral to this theory is the fractal dimension (FD), a key parameter used previously to describe the textural attributes of images." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.178, + 0.785, + 0.268 + ], + "angle": 0, + "content": "In addition to topology and geometry, edges and skeletons are important characteristics in defining tubular structures. In the segmentation of these structures, boundary accuracy and skeletal continuity are crucial for downstream tasks. For example, a commonly encountered problem in the segmentation of interconnected tubular structures (Fig. 1) is the breakages in segmentation results due to factors such as low-contrast or blurring." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.269, + 0.785, + 0.313 + ], + "angle": 0, + "content": "To enhance the segmentation quality of tubular structures, we exploit the topological self-similarity as well as edge and skeletal characteristics of tubular structures. The main research contributions of this study are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.315, + 0.785, + 0.435 + ], + "angle": 0, + "content": "1) We have developed a strategy to incorporate fractal features into deep learning networks. Specifically, we extend the fractal dimension from the image-level to the pixel-level and generate the fractal feature map (FFM), which characterizes topological self-similarity and textural complexity of each region within an image or its associated label. The FFM computed from the image is denoted as \\( FFM_{image} \\), while the FFM derived from the label is indicated as \\( FFM_{label} \\). Utilization of \\( FFM_{image} \\) as an additional model input and \\( FFM_{label} \\) as an additional loss function weight substantially enhances segmentation performance." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.436, + 0.785, + 0.557 + ], + "angle": 0, + "content": "2) We develop the multi-decoder network (MD-Net) by extending the U-Net [44] architecture with an edge decoder and a skeleton decoder. These decoders enable the model to simultaneously predict the boundaries and skeletons of image objects, in addition to the primary segmentation masks. By incorporating related constraints within the loss function, our model focuses not only on achieving accurate target segmentation but also allocates increased attention to boundary delineation and skeleton preservation, thereby enhancing the overall prediction quality." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.557, + 0.785, + 0.616 + ], + "angle": 0, + "content": "3) We have demonstrated the versatility and robustness of FFMs. Incorporation of \\(FFM_{image}\\) into the vanilla U-Net [44] and HR-Net [50] enhances segmentation performance, indicating that FFM can be used as a plug-in for different models." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.643, + 0.388, + 0.658 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.677, + 0.53, + 0.692 + ], + "angle": 0, + "content": "2.1 Tubular Structure Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.705, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Classical methods have been proposed to improve the performance of segmenting tubular structures by taking into account their geometric characteristics. Firstly, various methods utilize active contours and compute geodesics or minimal distance curves to approximate the contours of tubular structures, thereby effectively delineating their boundaries [1,6]. Secondly, tree structure-based methods utilize intrinsic shape priors to assist segmentation of different tubular structures. These methods employ a bottom-up method to identify tubular objects and a top-down grouping strategy to recognize tree structures, generating corresponding shape priors [4]. Lastly, various centerline-based methods" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.374, + 0.129 + ], + "angle": 0, + "content": "J. Huang et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.192 + ], + "angle": 0, + "content": "such as the one developed in [47] use multiscale detection strategies to accurately identify centerlines so that distance transform can be used to provide valuable information for segmentation of tubular structures." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.193, + 0.785, + 0.252 + ], + "angle": 0, + "content": "Deep learning methods have also been proposed to integrate topological and geometrical prior knowledge of tubular structures to enhance the performance of their segmentation. The integration is primarily achieved in three ways." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.253, + 0.787, + 0.358 + ], + "angle": 0, + "content": "1) Convolutional kernel design. The popular deformable convolution [9] and dilated convolution [57] aim to overcome the limitations of geometric transformations in CNNs and have demonstrated exceptional performance in complex segmentation tasks. Additionally, DSC-Net utilizes dynamic snake convolution to accurately capture the distinct features of tubular structures [42]. By adaptively focusing on slender and winding local structures, DSC-Net achieves improved performance in capturing the intricacy of tubular structures." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.359, + 0.788, + 0.493 + ], + "angle": 0, + "content": "2) Model architecture design. Various architecture designs have been proposed to learn the topological and geometrical features of tubular structures. In [31], a global transformer and dual local attention network are employed to simultaneously capture global and local features to effectively learn the complex geometric properties of tubular structures. Dong et al. [10] propose an enhanced Deformable U-Net that exploits flexible deformable convolutional layers to better generate clear boundaries for 3D cardiac cine MRI. High-frequency components that have strong capabilities to perceive thin structures are fused in [14] to enhance the performance of segmenting thin structures." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.494, + 0.788, + 0.645 + ], + "angle": 0, + "content": "3) Loss function design. Various loss functions have been explored for the segmentation of tubular structures. In [46], a similarity measure centerlineDice (clDice) based on the intersection of segmentation masks and their respective skeletons is introduced. A loss function based on clDice is proposed to enable the networks to generate segmentations with more accurate connectivity information and topology preservation. Araujo et al. [3] design a loss function based on the morphological closing operator that allows models to produce more topologically coherent masks and consistent vascular trees. Wong et al. [52] apply a novel Persistence Diagram Loss that quantifies topological correctness of segmentation over fine-grained structures." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.646, + 0.787, + 0.691 + ], + "angle": 0, + "content": "In this study, instead of relying on intricate designs of model architectures or loss functions, we incorporate FFMs as a model input and a loss function weight to enhance the perception of topological self-similarity and textures of images." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.711, + 0.533, + 0.727 + ], + "angle": 0, + "content": "2.2 Fractal Theory and Applications" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Despite their topological and geometrical complexity, tubular structures often exhibit topological similarities across different spatial scales. This observation suggests that their complex spatial patterns can be effectively described using simple texture features. Fractal geometry provides a means to describe the irregular or fragmented shapes of natural features and other intricate objects that traditional Euclidean geometry struggles to analyze [29]. Specifically, fractal features offer the capacity to describe and characterize the topological and" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.322, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "Fractal Feature Maps for Segmentation of Tubular Structures" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.191 + ], + "angle": 0, + "content": "geometrical complexity as well as textural composition of tubular image objects. Thus, fractal geometry has been applied to image classification and segmentation tasks." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.192, + 0.785, + 0.299 + ], + "angle": 0, + "content": "For classification, Roberto et al. [43] and Lin et al. [32] utilize the fractional Brownian motion (FBM) model [23] to extract fractal features of images. These features are then fed into a support vector machine or a convolutional neural network (CNN) classifier to differentiate between different objects. For segmentation, the FBM model is utilized in [33, 61] to extract fractal features from images. Such features are combined with classical methods such as thresholding and region growth techniques for image object segmentation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.299, + 0.787, + 0.405 + ], + "angle": 0, + "content": "Although existing methods have demonstrated good performance by leveraging fractal features, there is still a gap in exploring the integration of fractal features with deep learning for segmentation tasks. The inherent self-similarity observed in tubular structures aligns well with the fractal theory. In this study, we address this gap by incorporating FFMs into the segmentation model, aiming to provide a new and reliable source of information to enhance segmentation performance." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.432, + 0.331, + 0.447 + ], + "angle": 0, + "content": "3 Method" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.466, + 0.433, + 0.483 + ], + "angle": 0, + "content": "3.1 Fractal Feature Map" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.494, + 0.785, + 0.571 + ], + "angle": 0, + "content": "In fractal geometry, the fractal dimension (FD) provides a quantitative measure of an image's degree of self-similarity and roughness. FD can be estimated via the property of self-similarity [41]. Given an image \\( A \\), it is self-similar if \\( A \\) comprises \\( N_r \\) distinct copies of itself scaled down by a factor of \\( r \\). Consequently, for an image, the FD is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.431, + 0.582, + 0.786, + 0.614 + ], + "angle": 0, + "content": "\\[\nF D = \\lim _ {r \\rightarrow 0} \\frac {\\log N _ {r}}{\\log (1 / r)} \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.626, + 0.784, + 0.687 + ], + "angle": 0, + "content": "Although the definition of FD based on self-similarity is simple and concise, its direct estimation becomes impractical when dealing with irregular images. To overcome this challenge and estimate the FDs of images, the box-counting method [24, 30] is employed." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.687, + 0.786, + 0.809 + ], + "angle": 0, + "content": "Box-counting Method: Consider a grayscale image \\( I \\) with dimensions \\( M \\times M \\), where \\( L \\) denotes the maximum gray level (typically \\( L = 255 \\)). We can model the image \\( I \\) as a three-dimensional space with \\( (x, y) \\) indicating the two-dimensional position and a third coordinate \\( (z) \\) denoting the gray value. Then the 3-D space is subdivided into smaller cubic regions, or \"boxes\", each with dimensions \\( k \\times k \\times h \\). Here, \\( k \\) is a given scale used to be a multiple of the sidelength of a pixel in \\( (x, y) \\) and \\( h \\) can be a multiple of the gray level unit in \\( z \\)-direction. Given \\( L \\) and \\( k \\), the value of \\( h \\) calculated using the following formula:" + }, + { + "type": "equation", + "bbox": [ + 0.434, + 0.826, + 0.785, + 0.842 + ], + "angle": 0, + "content": "\\[\nh = (L - 1) \\times k / M \\tag {2}\n\\]" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.373, + 0.129 + ], + "angle": 0, + "content": "J. Huang et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.209 + ], + "angle": 0, + "content": "Given a \\( k \\times k \\) grid located at position \\( (i,j) \\), suppose that the minimum gray value is contained within the \\( m_{th} \\) box \\( (m = \\lceil minG_{i,j} / h \\rceil) \\), and the maximum gray value within the \\( l_{th} \\) box \\( (l = \\lceil maxG_{i,j} / h \\rceil) \\). The minimum number of boxes required to encompass all gray levels within the grid at \\( (i,j) \\) is computed as:" + }, + { + "type": "equation", + "bbox": [ + 0.433, + 0.225, + 0.786, + 0.241 + ], + "angle": 0, + "content": "\\[\nn _ {r} (i, j) = l - m + 1 \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.249, + 0.786, + 0.278 + ], + "angle": 0, + "content": "Considering all grids, the number of boxes that can cover all the patches is expressed as:" + }, + { + "type": "equation", + "bbox": [ + 0.443, + 0.28, + 0.786, + 0.311 + ], + "angle": 0, + "content": "\\[\nN _ {r} = \\sum_ {i, j} n _ {r} (i, j) \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.319, + 0.784, + 0.379 + ], + "angle": 0, + "content": "where \\( r = k / M \\). We can obtain a series of \\( N_{r} \\) using differing values of \\( k \\). Finally, the fractal dimension can be estimated from the least-squares linear fitting of \\( \\log N_{r} \\) versus \\( \\log (1 / r) \\), as illustrated in Equation (1). The flow of box-counting method is summarized in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.38, + 0.788, + 0.546 + ], + "angle": 0, + "content": "Fractal Feature Map: Although fractal dimension and fractal features have been previously applied to classification tasks [32, 43], the inherent differences between classification and segmentation tasks preclude the direct application of image-level fractal dimension to segmentation endeavors. We extend the calculation of FD from the image-level to the pixel-level to generate the FFM of an image. As depicted in Fig. 2, the process begins with the utilization of a sliding window technique. Within a \\(5 \\times 5\\) window, for example, the FD of this region is computed using the box-counting method. Subsequently, the window is shifted along both the horizontal and vertical directions with a step size of 1, resulting in the calculation of the FFM for the entire image. The algorithm for generating the FFM is summarized in Algorithm 1." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.547, + 0.473, + 0.758 + ], + "angle": 0, + "content": "The FFMs of the images (denoted as \\(FFM_{image}\\)) are incorporated into the segmentation model as additional input channels (Image, \\(FFM_{image}\\)), enhancing the focus on fractal structure and self-similarity after the normalization process. To mitigate the impact of image noise on the calculation of FD, we adopt the box-counting method proposed in [29] as they substitute the gray value with the standard deviation to make the method more robust." + }, + { + "type": "title", + "bbox": [ + 0.483, + 0.551, + 0.736, + 0.566 + ], + "angle": 0, + "content": "Algorithm 1 Generation of FFM." + }, + { + "type": "text", + "bbox": [ + 0.497, + 0.571, + 0.786, + 0.597 + ], + "angle": 0, + "content": "Input: Image \\(I\\) with size \\(M \\times N\\), window size \\(w \\times w\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.598, + 0.786, + 0.611 + ], + "angle": 0, + "content": "Output: Fractal feature map (FFM) of \\(I\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.612, + 0.786, + 0.626 + ], + "angle": 0, + "content": "Assign \\( FFM \\gets \\) Float array of size \\( (M, N) \\)" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.626, + 0.706, + 0.64 + ], + "angle": 0, + "content": "Assign padding size \\( p \\gets \\lfloor w / 2 \\rfloor \\)" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.64, + 0.694, + 0.654 + ], + "angle": 0, + "content": "\\(Pad_{I}\\gets\\) Linear Padding \\((I,p)\\)" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.654, + 0.66, + 0.666 + ], + "angle": 0, + "content": "For \\(i\\gets 0\\) to \\(M - 1\\) do" + }, + { + "type": "text", + "bbox": [ + 0.529, + 0.668, + 0.69, + 0.681 + ], + "angle": 0, + "content": "For \\(j\\gets 0\\) to \\(N - 1\\) do" + }, + { + "type": "text", + "bbox": [ + 0.559, + 0.682, + 0.69, + 0.696 + ], + "angle": 0, + "content": "Selected region \\(R_{i,j}\\)" + }, + { + "type": "equation", + "bbox": [ + 0.56, + 0.696, + 0.769, + 0.709 + ], + "angle": 0, + "content": "\\[\nR _ {i, j} \\leftarrow P a d _ {I} [ i: i + w, j: j + w ]\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.56, + 0.709, + 0.733, + 0.723 + ], + "angle": 0, + "content": "\\[\nF D \\leftarrow \\operatorname {B o x - c o u n t i n g} \\left(R _ {i, j}\\right)\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.56, + 0.723, + 0.676, + 0.736 + ], + "angle": 0, + "content": "\\[\nF F M [ i, j ] \\leftarrow F D\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.498, + 0.737, + 0.709, + 0.751 + ], + "angle": 0, + "content": "\\[\nF F M \\leftarrow \\operatorname {N o r m a l i z a t i o n} (F F M)\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.782, + 0.461, + 0.797 + ], + "angle": 0, + "content": "3.2 Multi-Decoder Network" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.841 + ], + "angle": 0, + "content": "In segmenting tubular structures, accurate boundary detection and preservation of global topological connectivity are crucial requirements. To meet these re" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.322, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "Fractal Feature Maps for Segmentation of Tubular Structures" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "image", + "bbox": [ + 0.304, + 0.147, + 0.703, + 0.248 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.333, + 0.258, + 0.668, + 0.272 + ], + "angle": 0, + "content": "Fig. 2: Workflow of computing FFM of an image." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.274, + 0.784, + 0.304 + ], + "angle": 0, + "content": "requirements, we propose a new model that we refer to as Multi-Decoder Network (MD-Net)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.304, + 0.785, + 0.514 + ], + "angle": 0, + "content": "In addition to predicting the segmentation mask, we introduce an edge decoder and a skeleton decoder to generate the edge and skeleton of the tubular structure, respectively, as depicted in Fig. 3. When (Image, \\( FFM_{image} \\)) are fed into MD-Net, the Encoder employs convolution layers to extract a series of low-level to high-level image features. These features are simultaneously transmitted to the three Decoders by skip connections for prediction. Specifically, in the Encoder, each convolution layer involves the repeated application of two 3x3 convolutions, followed by a rectified linear unit (ReLU), and a 2x2 max pooling operation with a stride of 2 for down-sampling. As for the Decoders in MD-Net, each convolution layer includes an up-sampling of the feature map, followed by a 2x2 convolution that reduces the number of feature channels by half. Subsequently, the feature is concatenated with the corresponding feature copied from the Encoder. This concatenated feature is then processed using two 3x3 convolutions, each followed by a ReLU activation function." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.516, + 0.785, + 0.621 + ], + "angle": 0, + "content": "During training, the ground truths for edges and skeletons are derived from the annotated masks by employing the findContours function from the OpenCV library and the skeletonization algorithms in the scikit-image library. We have visualized the boundaries and skeletons in the supplementary material to corroborate the accuracy of the extraction techniques. During the inference stage, the final structural segmentation is the output of the object decoder. Edge and skeleton predictions are only used in model training." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.621, + 0.785, + 0.713 + ], + "angle": 0, + "content": "Loss Function: Given an image \\(I\\) with \\(N\\) pixels, the segmentation ground truth is denoted as \\(y\\), where foreground and background pixels are labeled as 1 and 0, respectively. The prediction of the model is represented as \\(\\hat{y} \\in [0,1]\\), indicating the probabilities of individual pixels being classified as foreground. To compute the loss, we utilize the differentiable Soft IOU Loss [19], denoted as \\(\\mathcal{L}_{object}\\) in Equation (5), where \\(\\eta\\) is a smoothing coefficient." + }, + { + "type": "equation", + "bbox": [ + 0.34, + 0.72, + 0.785, + 0.759 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {o b j e c t}} = \\mathcal {L} _ {i o u} = 1 - \\frac {\\sum_ {i = 1} ^ {N} y _ {i} \\hat {y} _ {i}}{\\sum_ {i = 1} ^ {N} y _ {i} + \\hat {y} _ {i} - y _ {i} \\hat {y} _ {i} + \\eta} \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.764, + 0.784, + 0.796 + ], + "angle": 0, + "content": "The BCE loss function is applied separately to the edge and skeleton segmentation tasks, denoted as \\(\\mathcal{L}_{edge},\\mathcal{L}_{skeleton}\\)." + }, + { + "type": "equation", + "bbox": [ + 0.258, + 0.803, + 0.785, + 0.843 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {e d g e}}, \\mathcal {L} _ {\\text {s k e l e t o n}} = \\mathcal {L} _ {\\text {b c e}} = - \\frac {1}{N} \\sum_ {i = 1} ^ {N} y _ {i} \\cdot \\log (\\hat {y} _ {i}) + (1 - y _ {i}) \\cdot \\log (1 - \\hat {y} _ {i}) \\tag {6}\n\\]" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.374, + 0.129 + ], + "angle": 0, + "content": "J. Huang et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.193 + ], + "angle": 0, + "content": "For MD-Net, which has three segmentation tasks, the loss function is composed as in Equation (8). The weights \\(\\alpha, \\beta\\) and \\(\\gamma\\) are typically set to 1.0, 0.5 and 0.5, respectively." + }, + { + "type": "equation", + "bbox": [ + 0.356, + 0.193, + 0.786, + 0.21 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {g l o b a l}} = \\alpha \\mathcal {L} _ {\\text {o b j e c t}} + \\beta \\mathcal {L} _ {\\text {e d g e}} + \\gamma \\mathcal {L} _ {\\text {s k e l e t o n}} \\tag {7}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.229, + 0.538, + 0.243 + ], + "angle": 0, + "content": "3.3 Fractal Feature Constrained Loss" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.253, + 0.789, + 0.404 + ], + "angle": 0, + "content": "The value of FD serves as an indicator of the texture complexity present in a given image region. Hence, it is advantageous to incorporate pixel-level FFM as weights of the loss function. This allows for a higher weight to be assigned to regions that exhibit greater complexity, as accurately segmenting such regions poses a greater challenge. Consequently, we calculate the FFM for each label associated with the image, denoted as \\(FFM_{label}\\), and employ \\(FFM_{label}\\) as pixel-level weights within the loss function to guide the convergence of the model. Considering the specificity of the edge decoder and skeleton decoder, we have thus restricted the application of \\(FFM_{label}\\) exclusively to the object decoder. The Constrained Loss of MD-Net is as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.297, + 0.416, + 0.786, + 0.433 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {c o n s t r a i n e d}} = \\alpha \\mathcal {L} _ {\\text {o b j e c t}} \\cdot F F M _ {\\text {l a b e l}} + \\beta \\mathcal {L} _ {\\text {e d g e}} + \\gamma \\mathcal {L} _ {\\text {s k e l e t o n}} \\tag {8}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.453, + 0.377, + 0.47 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.484, + 0.336, + 0.498 + ], + "angle": 0, + "content": "4.1 Datasets" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.508, + 0.789, + 0.571 + ], + "angle": 0, + "content": "We evaluate the FFM and MD-Net using five publicly available tubular datasets including the endoplasmic reticulum (ER) and mitochondrial (MITO) [11, 36] organelle network datasets, the Retinal OCT-Angiography vessel Segmentation (ROSE) and Structured Analysis of the Retina (STARE) [15, 37] retinal vessel" + }, + { + "type": "image", + "bbox": [ + 0.226, + 0.596, + 0.784, + 0.744 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.757, + 0.789, + 0.855 + ], + "angle": 0, + "content": "Fig. 3: Overview and details of our proposed model MD-Net. The left part illustrates the overall structure of MD-Net. (Image, \\( FFM_{image} \\)) is processed as input by the Encoder, which extracts features of varying sizes from the input. These features are transmitted simultaneously to three Decoders through Skip Connection. The right part is a detailed description of the Encoder and Decoders. Each Decoder performs upsampling and concatenation operations similar to U-Net to obtain the predictions, comprising the edge, skeleton, and object." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.322, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Fractal Feature Maps for Segmentation of Tubular Structures" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.221 + ], + "angle": 0, + "content": "datasets, and the Massachusetts Roads (ROAD) [38] remote sensing dataset. By selecting these datasets, our aim is to comprehensively evaluate the performance of our methods across different types of structures and domains, providing a robust assessment of their capabilities. To further examine the influence of FFM, we included the non-tubular dataset NUCLEUS [5]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.223, + 0.785, + 0.357 + ], + "angle": 0, + "content": "All images in ER, MITO, STARE, and NUCLEUS are randomly sampled patches with size of \\(256 \\times 256\\) for training and sliding window with size of \\(256 \\times 256\\) for validation and testing. A crop size of \\(256 \\times 256\\) is used randomly during training and regularly during testing for the ROAD dataset as in [42]. Horizontal and vertical flipping as well as \\(90^{\\circ} / 180^{\\circ} / 270^{\\circ}\\) rotation are used for data augmentation. In addition, data augmentation for ROSE is conducted by rotation of an angle from \\(-10^{\\circ}\\) to \\(10^{\\circ}\\) as described in [37]. Detailed partitions of the six datasets and experimental setup are shown in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.38, + 0.471, + 0.394 + ], + "angle": 0, + "content": "4.2 Performance Comparison" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.403, + 0.785, + 0.553 + ], + "angle": 0, + "content": "We select several segmentation networks for comparison, including the vanilla U-Net [44], U-Net++ [60], nnU-Net [20], and HR-Net [50]. Additionally, we compare our methods with SOTA models for diverse datasets. For the task of retinal vessel segmentation, we choose three-stage [53], OCTA-Net [37], GT-DLA [31], and AF-Net [45], as these networks are tailored for this task. Furthermore, we compare the performance of MD-Net in the ROAD dataset with DCU-Net [54], Dconn-Net [55], and the DSC-Net [42], a tubular structure segmentation network with specially designed loss \\(\\mathcal{L}_{tc}\\). To assess the generalization of FFM, we apply \\(FFM_{image}\\) to both U-Net and HR-Net. All models are trained on the same dataset with the same hyperparameter settings." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.575, + 0.419, + 0.589 + ], + "angle": 0, + "content": "4.3 Evaluation Metrics" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.599, + 0.784, + 0.627 + ], + "angle": 0, + "content": "The models are assessed using three types of metrics: volumetric, topology, and distance." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.629, + 0.79, + 0.673 + ], + "angle": 0, + "content": "1. Volumetric scores, including intersection-over-union (IoU), accuracy (ACC), centerlineDice (clDice) [46], and AUC. These metrics quantify the overlap and agreement between the prediction and ground truth." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.784, + 0.717 + ], + "angle": 0, + "content": "2. Topology scores. To evaluate the topological correctness of the segmentations, we calculate the Betti Error [17] \\(\\beta\\) for the sum of Betti Numbers \\(\\beta_0\\) and \\(\\beta_{1}\\). In the evaluation of ROSE and STARE, \\(\\beta\\) Error represents the \\(\\beta_{1}\\) only." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.784, + 0.761 + ], + "angle": 0, + "content": "3. Distance score: Hausdorff Distance (HD) [48] quantifies the accuracy of the boundary delineation and provides information about the spatial closeness of the predicted and ground truth boundaries." + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.629, + 0.79, + 0.761 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.784, + 0.84 + ], + "angle": 0, + "content": "Since the nucleus is oval in shape and not tubular, no Betti Error and clDice are used in segmentation performance evaluation and no skeletal decoder is used in the MD-Net. In this study, the evaluation metrics IoU, ACC, AUC, and clDice, are expressed as percentages (\\%). The Hausdorff Distance (HD) is measured in pixels (px). All performance metrics are calculated for each image and averaged." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.373, + 0.129 + ], + "angle": 0, + "content": "J. Huang et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.788, + 0.215 + ], + "angle": 0, + "content": "Table 1: Bold numbers indicate the best performance and italicized numbers represent the next best performance. The result marked with a star (*) represents the model input replaced by (Image) with (Image, \\( FFM_{image} \\)). \\( \\mathcal{L}_{tc} \\) in (b) is Wasserstein-distance-based TCLoss proposed in [42]. \\( \\mathcal{L}_{\\text{constrained}} \\) represents the integration of \\( FFM_{label} \\) as a loss function weight, which leads to further improvement." + }, + { + "type": "table", + "bbox": [ + 0.22, + 0.227, + 0.787, + 0.488 + ], + "angle": 0, + "content": "
ModelLossIoU↑clDice↑ACC↑AUC↑βError↓HD↓IoU↑clDice↑ACC↑AUC↑βError↓HD↓
ERMITO
U-Net++ [60]Liou75.0294.6791.0497.3126.027.0479.7097.3098.1999.592.304.24
nnU-Net [20]Liou73.4194.5191.2987.9525.156.9679.2497.1698.0794.623.404.39
DSC-Net [42]Liou75.5194.4491.5697.0934.516.9280.3297.1698.1699.652.704.36
Dconn-Net [55]Liou75.9595.2491.5996.6020.316.9779.8297.2098.0399.542.904.32
AF-Net [45]Liou76.0194.5891.8197.3729.926.7980.3696.9498.0899.642.604.18
GT-DLA [31]Liou75.8994.9291.8297.3120.266.8480.2597.1698.1999.602.704.21
U-Net [44]Liou75.4494.6391.8297.3528.726.8779.7796.9198.0799.612.804.56
HR-Net [50]Liou75.8395.0891.7197.3622.576.9079.6397.2798.1499.622.904.25
U-Net*Liou76.5995.4392.0297.5620.786.8180.7197.4298.2199.632.704.27
HR-Net*Liou76.4395.4791.9597.5520.526.8380.6297.2998.1799.623.304.18
MD-Net*Lglobal77.0195.7892.0697.5919.106.7781.1197.7298.2599.662.204.16
MD-Net*Lconstrained77.0995.7492.1497.6519.526.7281.1897.6198.2699.652.804.14
ROSESTARE
three-stage [53]Liou62.1167.0791.7993.419.137.2866.8376.7194.8696.433.456.57
OCTA-Net [37]Liou63.2766.3292.3494.5310.227.1165.3776.7494.6295.443.226.47
DSC-Net [42]Liou63.1567.5991.9894.437.007.3667.0176.2694.9796.503.156.31
Dconn-Net [55]Liou60.3764.5188.5793.265.738.1866.1476.3394.7896.653.776.48
AF-Net [45]Liou62.3367.0591.7392.558.887.3267.2777.1794.9196.473.186.49
GT-DLA [31]Liou63.5766.3192.2794.189.897.1667.1276.3294.7295.793.126.63
U-Net [44]Liou61.5267.5391.3394.048.227.4266.1576.0594.6896.393.226.67
HR-Net [50]Liou63.0967.7192.0894.388.337.2667.0376.9294.9296.292.736.53
U-Net*Liou64.0767.9592.2594.477.887.1068.0777.3995.1596.632.776.33
HR-Net*Liou64.6869.4292.1494.776.017.2168.2277.4995.1396.523.216.32
MD-Net*Lglobal65.0769.7892.3694.884.227.1068.4977.7995.2096.912.576.29
MD-Net*Lconstrained65.1969.5892.4294.884.897.0668.7378.2095.2696.992.606.34
" + }, + { + "type": "table_caption", + "bbox": [ + 0.254, + 0.489, + 0.748, + 0.501 + ], + "angle": 0, + "content": "(a) Quantitative experimental results for the ER, MITO, ROSE, and STARE dataset." + }, + { + "type": "table", + "bbox": [ + 0.221, + 0.513, + 0.782, + 0.63 + ], + "angle": 0, + "content": "
ModelLossIoU↑cDice↑ACC↑AUC↑βError↓HD↓ModelLossIoU↑ACC↑AUC↑HD↓
DCU-Net [54]Lce62.9286.9898.0398.342.568.04CS-Net [39]Liou77.9698.1298.584.13
DSC-Net [42]Lce63.9987.7498.0598.392.567.96Transnet [8]Liou78.2198.3498.644.32
DSC-Net [42]Ltc64.2287.6498.0598.462.457.34DU-Net [21]Liou78.0998.4698.944.19
Dconn-Net [55]Ltc65.1487.6197.3098.292.256.97U-Net++[60]Liou78.3398.5999.694.09
AF-Net [45]Ltc65.0387.2797.4198.382.217.29nnU-Net [20]Liou78.2898.5794.934.09
GT-DLA [31]Ltc64.7686.8097.4397.942.386.94OCTA-Net [37]Liou75.3098.3499.264.43
U-Net [44]Lce62.4786.8797.9798.292.618.11U-Net [44]Liou78.0298.5499.654.23
HR-Net [50]Lce62.8986.7498.0698.392.588.06HR-Net [50]Liou77.7098.5099.654.15
U-Net*Lce65.7487.7498.3898.722.486.98U-Net*Liou78.4698.5899.684.15
HR-Net*Lce65.6087.6198.3298.792.316.98HR-Net*Liou77.8398.5399.664.15
MD-Net*Lglobal66.0788.0898.4398.802.196.92MD-Net*Lglobal78.6698.5999.684.08
MD-Net*Lconstrained66.1688.0798.4298.772.066.97MD-Net*Lconstrained78.7798.5999.664.06
" + }, + { + "type": "table_caption", + "bbox": [ + 0.288, + 0.631, + 0.468, + 0.642 + ], + "angle": 0, + "content": "(b) Results for ROAD dataset." + }, + { + "type": "table_caption", + "bbox": [ + 0.559, + 0.631, + 0.761, + 0.643 + ], + "angle": 0, + "content": "(c) Results for NUCLEUS dataset." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.667, + 0.384, + 0.683 + ], + "angle": 0, + "content": "4.4 Configurations" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.719, + 0.788, + 0.841 + ], + "angle": 0, + "content": "All models are implemented using PyTorch (version 1.12.0) and executed on 4 NVIDIA 3090 GPU cards. During the training phase, we employed the SGD optimizer with an initial learning rate of 0.05. To dynamically adjust the learning rate as the training progressed, we incorporated warm-up and exponential decay techniques. For all datasets, a fixed batch size of 32 is employed, ensuring consistency in the training process. Additionally, to prevent overfitting, a regularization weight of 0.0005 is applied. The entire training process spanned 50 epochs, providing sufficient iterations for model convergence and learning." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.322, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Fractal Feature Maps for Segmentation of Tubular Structures" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.117, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.147, + 0.461, + 0.162 + ], + "angle": 0, + "content": "4.5 Quantitative Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.174, + 0.757, + 0.188 + ], + "angle": 0, + "content": "Based on the results presented in Tab. 1, several conclusions can be drawn." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.189, + 0.787, + 0.28 + ], + "angle": 0, + "content": "Performance of MD-Net: Our proposed model, MD-Net, demonstrates superior performance in terms of segmentation accuracy and topological continuity compared to the other models. This superiority is observed across five tubular datasets as well as one non-tubular dataset. We have also performed t-test to check whether improvements of our methods over competing methods in performance are statistically significant. Results are provided in supplementary." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.28, + 0.794, + 0.431 + ], + "angle": 0, + "content": "In tubular datasets, MD-Net outperforms other methods in terms of segmentation results. Specifically, in the ER, MITO, ROSE, STARE, and ROAD datasets, MD-Net achieves improvements in IoU of \\(1.08\\%\\), \\(0.82\\%\\), \\(1.62\\%\\), \\(1.46\\%\\), and \\(1.02\\%\\), respectively, compared to existing SOTA methods. Furthermore, when considering topological continuity and boundary extraction, MD-Net demonstrates the best performance with the minimum \\(\\beta\\) error and Hausdorff Distance. These results highlight the ability of MD-Net, equipped with an edge decoder and a skeleton decoder, to effectively capture edge and skeleton features of thin tubular structures, resulting in more accurate and continuous topology segmentation outcomes." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.431, + 0.787, + 0.491 + ], + "angle": 0, + "content": "Even in the non-tubular dataset NUCLEUS, MD-Net achieves the best segmentation results with an IoU of \\(78.77\\%\\), ACC of \\(98.59\\%\\), and HD of 4.06. This demonstrates the effectiveness of MD-Net and FFM in datasets with simpler structures." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.492, + 0.787, + 0.553 + ], + "angle": 0, + "content": "Generalization of FFM: \\(FFM_{image}\\) is incorporated into U-Net and HR-Net as an additional input channel without modifying the training parameters. The results in Tab. 1 indicate that the inclusion of \\(FFM_{image}\\) leads to improved segmentation performance for both U-Net and HR-Net." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.553, + 0.79, + 0.658 + ], + "angle": 0, + "content": "For U-Net, the segmentation performance (IoU) shows improvements of \\(1.14\\%\\), \\(0.94\\%\\), \\(2.55\\%\\), \\(1.92\\%\\), \\(3.27\\%\\), and \\(0.44\\%\\) on the ER, MITO, ROSE, STARE, ROAD, and NUCLEUS datasets with the assistance of \\(FFM_{image}\\), respectively. Similarly, for HR-Net, the IoU improves by \\(0.60\\%\\), \\(0.99\\%\\), \\(1.59\\%\\), \\(1.19\\%\\), \\(2.71\\%\\), and \\(0.13\\%\\) on the same datasets, respectively. On average, \\(FFM_{image}\\) brings an improvement of \\(1.96\\%\\) and \\(1.42\\%\\) in IoU for U-Net and HR-Net, respectively, in the segmentation task of tubular datasets." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.659, + 0.787, + 0.735 + ], + "angle": 0, + "content": "In conclusion, FFM demonstrates notable generalization capabilities, leading to enhanced performance across various models. Particularly in datasets with intricate structures, FFM consistently outperforms the original models in segmentation tasks. This improvement can be attributed to the inherent characteristics of FD involved in FFM." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.787, + 0.842 + ], + "angle": 0, + "content": "\\(FFM_{label}\\) and \\(\\mathcal{L}_{global}\\): Fractal Dimension is a significant metric used to quantify the complexity of image textures. In the context of model training, we leverage the \\(FFMs_{label}\\) as weights for the loss function. This method enables us to guide the model's focus towards regions with higher complexity, as indicated by a higher FD. Through extensive experimentation, we have observed that incorporating FFM-based weights into the loss function yields improvements in both the training performance and efficiency of the model, as depicted in Tab. 1." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.374, + 0.129 + ], + "angle": 0, + "content": "J. Huang et al." + }, + { + "type": "image", + "bbox": [ + 0.293, + 0.147, + 0.71, + 0.41 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.424, + 0.788, + 0.51 + ], + "angle": 0, + "content": "Fig. 4: Comparison of segmentation results. (a) Image. (b) Label. (c) Results of U-Net. (d) Results of existing SOTA approaches. From top to bottom, it's AF-Net, AF-Net, GT-DLA, AF-Net, and Dconn-Net. (e) Results of U-Net*. (f) Results of MD-Net*. Red: true positive. Green: false negative. Blue: false positive. In the ER, MITO, ROSE, STAR, and ROAD rows, columns (b) through (f) display the magnified results of the regions demarcated by yellow squares in the corresponding images of column (a)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.519, + 0.788, + 0.551 + ], + "angle": 0, + "content": "These findings highlight the tangible benefits of including \\( FFM_{label} \\) as a weight parameter in optimizing the training process." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.571, + 0.449, + 0.587 + ], + "angle": 0, + "content": "4.6 Qualitative Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.596, + 0.789, + 0.703 + ], + "angle": 0, + "content": "To facilitate a more comprehensive comparison of the experimental outcomes across various models, we employed visualizations to present the segmentation results, as depicted in Fig. 4. The visual analysis provides a clear and intuitive representation of the model's performance. The observed results validate that the integration of FFM leads to enhanced segmentation outcomes in terms of both edge accuracy and topological continuity. More visualization results can be found in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.724, + 0.405, + 0.743 + ], + "angle": 0, + "content": "5 Ablation Study" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.755, + 0.573, + 0.772 + ], + "angle": 0, + "content": "5.1 Effectiveness of Fractal Feature Maps" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.842 + ], + "angle": 0, + "content": "To ascertain that the observed performance improvement of the model is attributed to the FFM and not the increase of input channels, we replace the \\(FFM_{image}\\) with the image, thereby transforming the input (image, \\(FFM_{image}\\)) to (image, image). Additionally, we computed the image's Hurst feature (HF," + } + ], + [ + { + "type": "header", + "bbox": [ + 0.322, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Fractal Feature Maps for Segmentation of Tubular Structures" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.239 + ], + "angle": 0, + "content": "classical fractal analysis), mean feature (MF, directional analysis), and contrast feature (CF, directional analysis) as introduced by [22]. These features replaced the \\( FFM_{image} \\) as input and were trained and evaluated on U-Net and MD-Net. Experimental results demonstrated the advantages of the \\( FFM_{image} \\) as shown in Tab. 2. This finding indicates that the inclusion of the FFM plays an effective role in achieving better performance in the segmentation task." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.258, + 0.561, + 0.273 + ], + "angle": 0, + "content": "5.2 Robustness of Fractal Feature Maps" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.28, + 0.788, + 0.415 + ], + "angle": 0, + "content": "The computation of the FFM is influenced by the selection of window size and step size. We used different window and step sizes to generate different \\(FFMs_{image}\\) and tested their performance. The results in Tab. 3 demonstrate the robustness of FFM. When used for training the U-Net and MD-Net, the \\(FFMs_{image}\\) calculated using different window sizes consistently yield results with a variation range of \\(0.3\\%\\) on the ER and STARE datasets. Similarly, applying different steps during the calculation of FFMs does not lead to significant changes in performance. It should be noted that the efficiency of computing can be substantially improved by increasing the step size." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.416, + 0.787, + 0.446 + ], + "angle": 0, + "content": "The numerical results in Tab. 2 and Tab. 3 are calculated by volumetric score IoU. Results with more metrics can be found in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.466, + 0.436, + 0.48 + ], + "angle": 0, + "content": "5.3 Decoders of MD-Net" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.489, + 0.788, + 0.61 + ], + "angle": 0, + "content": "To evaluate the effectiveness of the edge decoder and the skeleton decoder, a comparative ablation analysis was conducted between MD-Net and U-Net on the ER and STARE datasets. Subsequently, we proceeded to remove either the edge decoder or the skeleton decoder from MD-Net and trained the modified models. The results shown in Tab. 4 confirm that both the edge decoder and the skeleton decoder contribute to improving the segmentation performance. It is noteworthy that the best performance is achieved when all components are utilized simultaneously." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.63, + 0.554, + 0.646 + ], + "angle": 0, + "content": "5.4 Limitations of Fractal Feature Map" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.652, + 0.788, + 0.683 + ], + "angle": 0, + "content": "The time complexity of FFMs' computation is associated with the size of image \\((M\\times N)\\), window size \\(w\\), and step size \\(S\\). Its time complexity is represented" + }, + { + "type": "table_caption", + "bbox": [ + 0.215, + 0.688, + 0.788, + 0.731 + ], + "angle": 0, + "content": "Table 2: Ablation study of FFM in U-Net and MD-Net. The values within the parentheses in the (image, \\(FFM_{image}\\)) column represent the percentage improvement of IoU compared to the baseline results listed in the (image) column." + }, + { + "type": "table", + "bbox": [ + 0.286, + 0.743, + 0.716, + 0.858 + ], + "angle": 0, + "content": "
ModelDatasetInput
(image)(image,image)(image,HF)(image,MF)(image,CF)(image,FFM image)
U-NetER75.4475.6175.8975.8675.1076.59(+1.14%)
MITO79.7780.2880.5180.2380.3880.71(+0.94%)
ROSE61.5262.5762.6162.5862.3764.07(+2.55%)
STARE66.1566.0865.5466.1466.8368.07(+1.92%)
ROAD62.4763.4464.9364.5864.2565.74(+3.27%)
MD-NetER76.2876.3476.2475.9775.5277.01(+0.73%)
MITO80.2880.3179.9780.4180.3281.11(+0.83%)
ROSE63.3163.3062.7762.8962.9065.07(+1.76%)
STARE66.4666.5766.4367.0367.2568.49(+2.03%)
ROAD64.7965.0465.2365.1265.1566.07(+1.28%)
" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.374, + 0.129 + ], + "angle": 0, + "content": "J. Huang et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.265, + 0.145, + 0.737, + 0.159 + ], + "angle": 0, + "content": "Table 3: Ablation of step size and window size in U-Net and MD-Net." + }, + { + "type": "table", + "bbox": [ + 0.286, + 0.171, + 0.716, + 0.245 + ], + "angle": 0, + "content": "
ModelStep SizeWindow SizeERSTAREModelStep SizeWindow SizeERSTARE
U-Net*11176.7567.87MD-Net*11177.0368.41
1976.7367.911977.0168.21
1776.6567.891776.9268.50
1576.5968.071577.0168.49
2576.4967.872576.9168.46
3576.6867.783576.8868.32
4576.6167.784576.9068.30
" + }, + { + "type": "table_caption", + "bbox": [ + 0.377, + 0.255, + 0.625, + 0.269 + ], + "angle": 0, + "content": "Table 4: Ablation study of MD-Net." + }, + { + "type": "table", + "bbox": [ + 0.286, + 0.281, + 0.714, + 0.43 + ], + "angle": 0, + "content": "
DatasetModelFFMEdge DecoderSkeleton DecoderIoU↑clDice↑ACC↑AUC↑βError↓HD↓
ERU-Net75.4494.6391.8297.3528.726.87
MD-Net✓✓76.2895.2792.0097.3922.556.78
U-Net✓76.5995.4392.0297.5620.786.81
MD-Net✓✓76.9795.6092.0997.5922.426.71
MD-Net✓✓76.8595.5592.1397.5120.636.74
MD-Net✓✓✓77.0995.7492.1497.6519.526.72
STAREU-Net66.1576.0594.6896.393.226.67
MD-Net✓✓66.4676.5494.7696.423.276.47
U-Net✓68.0777.3995.1596.632.776.33
MD-Net✓✓68.2677.8095.2096.772.656.29
MD-Net✓✓68.3977.7695.2096.893.226.31
MD-Net✓✓✓68.7378.2095.2696.992.606.34
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.443, + 0.788, + 0.522 + ], + "angle": 0, + "content": "as: \\( O\\left(\\frac{M \\cdot N}{S^2} \\cdot \\left(w^2 \\cdot \\log(w)\\right)\\right) \\). Calculation of FFM indeed incurs computational overhead but is optimized by using multithreading and larger \\( S \\). Currently, the generation of FFMs for a \\( 256 \\times 256 \\) image takes about 150 milliseconds on a Xeon 8336C CPU, given a step size of 1 and a window size of 5. Detailed information on training and inference time can be found in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.555, + 0.37, + 0.571 + ], + "angle": 0, + "content": "6 Conclusions" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.599, + 0.789, + 0.844 + ], + "angle": 0, + "content": "In this study, we propose a method that uses fractal feature maps (FFMs) along with a multi-decoder network (MD-Net) for the semantic segmentation of tubular structures. FFMs are used to capture the texture and self-similarity of image regions or label regions at the pixel-level through the fractal dimension. \\( FFM_{image} \\) are used as an additional input to enhance the segmentation model's perception of tubular structures and \\( FFM_{label} \\) are used as a weight for the loss function to guide the model training. MD-Net improves the quality of segmentation by simultaneously predicting edges and skeletons through the incorporation of an edge decoder and a skeleton decoder. The proposed method is evaluated on five datasets of tubular structures and one dataset of non-tubular structures. The results demonstrate the superiority of MD-Net with FFM in the segmentation of tubular structures and stable performance in the segmentation of non-tubular structures. Furthermore, the performance improvement observed when incorporating the FFM module into vanilla U-Net and HR-Net underscores the broad applicability and potential of incorporating fractal information into deep learning models." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.322, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Fractal Feature Maps for Segmentation of Tubular Structures" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.403, + 0.163 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.181, + 0.788, + 0.257 + ], + "angle": 0, + "content": "This work was supported in part by the National Natural Science Foundation of China (grants 92354307, 91954201, 31971289, 32101216), the Strategic Priority Research Program of the Chinese Academy of Sciences (grant XDB37040402) and the Fundamental Research Funds for the Central Universities (grant E3E45201X2)." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.284, + 0.323, + 0.3 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.319, + 0.787, + 0.404 + ], + "angle": 0, + "content": "1. Alvarez, L., Trujillo, A., Cuenca, C., González, E., Esclarín, J., Gomez, L., Mazarra, L., Aleman-Flores, M., Tahoces, P.G., Carreira, J.M.: Tracking the aortic lumen geometry by optimizing the 3d orientation of its cross-sections. In: Medical Image Computing and Computer-Assisted Intervention- MICCAI 2017: 20th International Conference, Quebec City, QC, Canada, September 11-13, 2017, Proceedings, Part II 20. pp. 174-181. Springer (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.405, + 0.787, + 0.445 + ], + "angle": 0, + "content": "2. Antiga, L., Ene-Iordache, B., Remuzzi, A.: Computational geometry for patient-specific reconstruction and meshing of blood vessels from mr and ct angiography. IEEE Transactions on Medical Imaging 22(5), 674-684 (2003)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.447, + 0.787, + 0.474 + ], + "angle": 0, + "content": "3. Araújo, R.J., Cardoso, J.S., Oliveira, H.P.: Topological similarity index and loss function for blood vessel segmentation. arXiv preprint arXiv:2107.14531 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.476, + 0.787, + 0.516 + ], + "angle": 0, + "content": "4. Bauer, C., Pock, T., Sorantin, E., Bischof, H., Beichel, R.: Segmentation of interwoven 3d tubular tree structures utilizing shape priors and graph cuts. Medical Image Analysis 14(2), 172-184 (2010)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.518, + 0.787, + 0.572 + ], + "angle": 0, + "content": "5. Caicedo, J.C., Goodman, A., Karhohs, K.W., Cimini, B.A., Ackerman, J., Haghighi, M., Heng, C., Becker, T., Doan, M., McQuin, C., et al.: Nucleus segmentation across imaging experiments: the 2018 data science bowl. Nature Methods 16(12), 1247-1253 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.574, + 0.787, + 0.601 + ], + "angle": 0, + "content": "6. Caselles, V., Kimmel, R., Sapiro, G.: Geodesic active contours. International Journal of Computer Vision 22, 61-79 (1997)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.602, + 0.787, + 0.643 + ], + "angle": 0, + "content": "7. Challoo, M., Gao, Y., Busch, A.: Distinctive phase interdependency model for retinal vasculature delineation in oct-angiography images. IEEE Transactions on Medical Imaging (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.645, + 0.787, + 0.685 + ], + "angle": 0, + "content": "8. Chen, J., Lu, Y., Yu, Q., Luo, X., Adeli, E., Wang, Y., Lu, L., Yuille, A.L., Zhou, Y.: Transunet: Transformers make strong encoders for medical image segmentation. arXiv preprint arXiv:2102.04306 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.687, + 0.787, + 0.728 + ], + "angle": 0, + "content": "9. Dai, J., Qi, H., Xiong, Y., Li, Y., Zhang, G., Hu, H., Wei, Y.: Deformable convolutional networks. In: Proceedings of the IEEE international conference on computer vision. pp. 764-773 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.729, + 0.787, + 0.77 + ], + "angle": 0, + "content": "0. Dong, S., Pan, Z., Fu, Y., Yang, Q., Gao, Y., Yu, T., Shi, Y., Zhuo, C.: Deu-net 2.0: Enhanced deformable u-net for 3d cardiac cine migraine segmentation. Medical Image Analysis 78, 102389 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.772, + 0.787, + 0.84 + ], + "angle": 0, + "content": "11. Guo, Y., Huang, J., Zhou, Y., Luo, Y., Li, W., Yang, G.: Segmentation of intracellular structures in fluorescence microscopy images by fusing low-level features. In: Pattern Recognition and Computer Vision: 4th Chinese Conference, PRCV 2021, Beijing, China, October 29–November 1, 2021, Proceedings, Part III 4. pp. 386–397. Springer (2021)" + }, + { + "type": "list", + "bbox": [ + 0.22, + 0.319, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.373, + 0.129 + ], + "angle": 0, + "content": "J. Huang et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.787, + 0.189 + ], + "angle": 0, + "content": "12. Gupta, S., Zhang, Y., Hu, X., Prasanna, P., Chen, C.: Topology-aware uncertainty for image segmentation. Advances in Neural Information Processing Systems 36 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.189, + 0.787, + 0.23 + ], + "angle": 0, + "content": "13. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 770-778 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.23, + 0.787, + 0.285 + ], + "angle": 0, + "content": "14. He, Y., Ge, R., Wu, J., Coatrieux, J.L., Shu, H., Chen, Y., Yang, G., Li, S.: Thin semantics enhancement via high-frequency priori rule for thin structures segmentation. In: 2021 IEEE International Conference on Data Mining (ICDM). pp. 1096-1101. IEEE (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.285, + 0.787, + 0.325 + ], + "angle": 0, + "content": "15. Hoover, A., Goldbaum, M.: Locating the optic nerve in a retinal image using the fuzzy convergence of the blood vessels. IEEE Transactions on Medical Imaging 22(8), 951-958 (2003)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.325, + 0.787, + 0.352 + ], + "angle": 0, + "content": "16. Hotamisligil, G.S.: Endoplasmic reticulum stress and the inflammatory basis of metabolic disease. Cell 140(6), 900-917 (2010)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.352, + 0.787, + 0.379 + ], + "angle": 0, + "content": "17. Hu, X., Li, F., Samaras, D., Chen, C.: Topology-preserving deep image segmentation. Advances in neural information processing systems 32 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.379, + 0.787, + 0.406 + ], + "angle": 0, + "content": "18. Hu, X., Wang, Y., Fuxin, L., Samaras, D., Chen, C.: Topology-aware segmentation using discrete Morse theory. arXiv preprint arXiv:2103.09992 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.406, + 0.787, + 0.433 + ], + "angle": 0, + "content": "19. Huang, Y., Tang, Z., Chen, D., Su, K., Chen, C.: Batching soft iou for training semantic segmentation networks. IEEE Signal Processing Letters 27, 66-70 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.433, + 0.787, + 0.473 + ], + "angle": 0, + "content": "20. Isensee, F., Jaeger, P.F., Kohl, S.A., Petersen, J., Maier-Hein, K.H.: nnu-net: a self-configuring method for deep learning-based biomedical image segmentation. Nature Methods 18(2), 203-211 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.473, + 0.787, + 0.514 + ], + "angle": 0, + "content": "21. Jin, Q., Meng, Z., Pham, T.D., Chen, Q., Wei, L., Su, R.: Dunet: A deformable network for retinal vessel segmentation. Knowledge-Based Systems 178, 149-162 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.514, + 0.787, + 0.542 + ], + "angle": 0, + "content": "22. Kaplan, L.M.: Extended fractal analysis for texture classification and segmentation. IEEE transactions on image processing 8(11), 1572-1585 (1999)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.542, + 0.787, + 0.568 + ], + "angle": 0, + "content": "23. Kaplan, L.M., Kuo, C.C.: Extending self-similarity for fractional brownian motion. IEEE Transactions on Signal Processing 42(12), 3526-3530 (1994)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.568, + 0.787, + 0.609 + ], + "angle": 0, + "content": "24. Konatar, I., Popovic, T., Popovic, N.: Box-counting method in python for fractal analysis of biomedical images. In: 2020 24th International Conference on Information Technology (IT). pp. 1-4. IEEE (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.609, + 0.787, + 0.663 + ], + "angle": 0, + "content": "25. Kreitner, L., Paetzold, J.C., Rauch, N., Chen, C., Hagag, A.M., Fayed, A.E., Sivaprasad, S., Rausch, S., Weichsel, J., Menze, B.H., et al.: Synthetic optical coherence tomography angiographs for detailed retinal vessel segmentation without human annotations. IEEE Transactions on Medical Imaging (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.663, + 0.787, + 0.718 + ], + "angle": 0, + "content": "26. La Barbera, G., Rouet, L., Boussaid, H., Lubet, A., Kassir, R., Sarnacki, S., Gori, P., Bloch, I.: Tubular structures segmentation of pediatric abdominal-visceral cect images with renal tumors: assessment, comparison and improvement. Medical Image Analysis 90, 102986 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.718, + 0.787, + 0.759 + ], + "angle": 0, + "content": "27. Laptev, I., Mayer, H., Lindeberg, T., Eckstein, W., Steger, C., Baumgartner, A.: Automatic extraction of roads from aerial images based on scale space and snakes. Machine Vision and Applications 12, 23-31 (2000)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.759, + 0.787, + 0.8 + ], + "angle": 0, + "content": "28. Lee, C.A., Blackstone, C.: Er morphology and endo-lysosomal crosstalk: Functions and disease implications. Biochimica et Biophysica Acta (BBA)-Molecular and Cell Biology of Lipids 1865(1), 158544 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.8, + 0.787, + 0.841 + ], + "angle": 0, + "content": "29. Lee, W.L., Hsieh, K.S.: A robust algorithm for the fractal dimension of images and its applications to the classification of natural images and ultrasonic liver images. Signal Processing 90(6), 1894-1904 (2010)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.322, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Fractal Feature Maps for Segmentation of Tubular Structures" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.175 + ], + "angle": 0, + "content": "30. Li, J., Du, Q., Sun, C.: An improved box-counting method for image fractal dimension estimation. Pattern Recognition 42(11), 2460-2469 (2009)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.175, + 0.788, + 0.231 + ], + "angle": 0, + "content": "31. Li, Y., Zhang, Y., Liu, J.Y., Wang, K., Zhang, K., Zhang, G.S., Liao, X.F., Yang, G.: Global transformer and dual local attention network via deep-shallow hierarchical feature fusion for retinal vessel segmentation. IEEE Transactions on Cybernetics 53(9), 5826-5839 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.231, + 0.788, + 0.272 + ], + "angle": 0, + "content": "32. Lin, P.L., Huang, P.W., Lee, C.H., Wu, M.T.: Automatic classification for solitary pulmonary nodule in ct image by fractal analysis based on fractional brownian motion model. Pattern Recognition 46(12), 3279-3287 (2013)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.272, + 0.788, + 0.327 + ], + "angle": 0, + "content": "33. Lin, P., Huang, P., Huang, P., Hsu, H.: Alveolar bone-loss area localization in periodontitis radiographs based on threshold segmentation with a hybrid feature fused of intensity and the h-value of fractional brownian motion model. Computer Methods and Programs in Biomedicine 121(3), 117-126 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.328, + 0.788, + 0.37 + ], + "angle": 0, + "content": "34. Long, J., Shelhamer, E., Darrell, T.: Fully convolutional networks for semantic segmentation. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 3431-3440 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.37, + 0.788, + 0.397 + ], + "angle": 0, + "content": "35. Lopes, R., Betrouni, N.: Fractal and multifractal analysis: A review. Medical Image Analysis 13(4), 634-649 (2009)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.397, + 0.788, + 0.451 + ], + "angle": 0, + "content": "36. Luo, Y., Guo, Y., Li, W., Liu, G., Yang, G.: Fluorescence microscopy image datasets for deep learning segmentation of intracellular organelle networks (2020). https://doi.org/10.21227/t2he-zn97, https://dx.doi.org/10.21227/t2he-zn97" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.452, + 0.788, + 0.493 + ], + "angle": 0, + "content": "37. Ma, Y., Hao, H., Xie, J., Fu, H., Zhang, J., Yang, J., Wang, Z., Liu, J., Zheng, Y., Zhao, Y.: Rose: a retinal oct-angiography vessel segmentation dataset and new model. IEEE Transactions on Medical Imaging 40(3), 928–939 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.493, + 0.788, + 0.522 + ], + "angle": 0, + "content": "38. Mnih, V.: Machine learning for aerial image labeling. University of Toronto (Canada) (2013)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.522, + 0.788, + 0.59 + ], + "angle": 0, + "content": "39. Mou, L., Zhao, Y., Chen, L., Cheng, J., Gu, Z., Hao, H., Qi, H., Zheng, Y., Frangi, A., Liu, J.: Cs-net: Channel and spatial attention network for curvilinear structure segmentation. In: Medical Image Computing and Computer Assisted Intervention-MICCAI 2019: 22nd International Conference, Shenzhen, China, October 13-17, 2019, Proceedings, Part I 22. pp. 721-730. Springer (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.591, + 0.788, + 0.646 + ], + "angle": 0, + "content": "40. Nain, D., Yezzi, A., Turk, G.: Vessel segmentation using a shape driven flow. In: Medical Image Computing and Computer-Assisted Intervention-MICCAI 2004: 7th International Conference, Saint-Malo, France, September 26-29, 2004. Proceedings, Part I 7. pp. 51-59. Springer (2004)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.646, + 0.788, + 0.674 + ], + "angle": 0, + "content": "41. Pentland, A.P.: Fractal-based description of natural scenes. IEEE Transactions on Pattern Analysis and Machine Intelligence (6), 661-674 (1984)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.674, + 0.788, + 0.729 + ], + "angle": 0, + "content": "42. Qi, Y., He, Y., Qi, X., Zhang, Y., Yang, G.: Dynamic snake convolution based on topological geometric constraints for tubular structure segmentation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 6070-6079 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.729, + 0.788, + 0.784 + ], + "angle": 0, + "content": "43. Roberto, G.F., Lumini, A., Neves, L.A., do Nascimento, M.Z.: Fractal neural network: A new ensemble of fractal geometry and convolutional neural networks for the classification of histology images. Expert Systems with Applications 166, 114103 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.784, + 0.788, + 0.84 + ], + "angle": 0, + "content": "44. Ronneberger, O., Fischer, P., Brox, T.: U-net: Convolutional networks for biomedical image segmentation. In: Medical Image Computing and Computer-Assisted Intervention-MICCAI 2015: 18th International Conference, Munich, Germany, October 5-9, 2015, Proceedings, Part III 18. pp. 234-241. Springer (2015)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.373, + 0.129 + ], + "angle": 0, + "content": "J. Huang et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.787, + 0.189 + ], + "angle": 0, + "content": "45. Shi, T., Ding, X., Zhou, W., Pan, F., Yan, Z., Bai, X., Yang, X.: Affinity feature strengthening for accurate, complete and robust vessel segmentation. IEEE Journal of Biomedical and Health Informatics 27(8), 4006-4017 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.189, + 0.787, + 0.243 + ], + "angle": 0, + "content": "46. Shit, S., Paetzold, J.C., Sekuboyina, A., Ezhov, I., Unger, A., Zhylka, A., Pluim, J.P., Bauer, U., Menze, B.H.: cldice-a novel topology-preserving loss function for tubular structure segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16560-16569 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.243, + 0.787, + 0.284 + ], + "angle": 0, + "content": "47. Sironi, A., Lepetit, V., Fua, P.: Multiscale centerline detection by learning a scale-space distance transform. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 2697-2704 (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.284, + 0.787, + 0.311 + ], + "angle": 0, + "content": "48. Taha, A.A., Hanbury, A.: Metrics for evaluating 3d medical image segmentation: analysis, selection, and tool. BMC Medical Imaging 15(1), 1-28 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.311, + 0.787, + 0.379 + ], + "angle": 0, + "content": "49. Wang, C., Hayashi, Y., Oda, M., Itoh, H., Kitasaka, T., Frangi, A.F., Mori, K.: Tubular structure segmentation using spatially connected network with radial distance loss for 3d medical images. In: Medical Image Computing and Computer Assisted Intervention-MICCAI 2019: 22nd International Conference, Shenzhen, China, October 13-17, 2019, Proceedings, Part VI 22. pp. 348-356. Springer (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.379, + 0.787, + 0.433 + ], + "angle": 0, + "content": "50. Wang, J., Sun, K., Cheng, T., Jiang, B., Deng, C., Zhao, Y., Liu, D., Mu, Y., Tan, M., Wang, X., et al.: Deep high-resolution representation learning for visual recognition. IEEE Transactions on Pattern Analysis and Machine Intelligence 43(10), 3349-3364 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.433, + 0.787, + 0.488 + ], + "angle": 0, + "content": "51. Wang, Y., Wei, X., Liu, F., Chen, J., Zhou, Y., Shen, W., Fishman, E.K., Yuille, A.L.: Deep distance transform for tubular structure segmentation in ct scans. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3833-3842 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.488, + 0.787, + 0.528 + ], + "angle": 0, + "content": "52. Wong, C.C., Vong, C.M.: Persistent homology based graph convolution network for fine-grained 3d shape segmentation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 7098-7107 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.528, + 0.787, + 0.569 + ], + "angle": 0, + "content": "53. Yan, Z., Yang, X., Cheng, K.T.: A three-stage deep learning model for accurate retinal vessel segmentation. IEEE Journal of Biomedical and Health Informatics 23(4), 1427-1436 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.569, + 0.787, + 0.609 + ], + "angle": 0, + "content": "54. Yang, X., Li, Z., Guo, Y., Zhou, D.: Dcu-net: A deformable convolutional neural network based on cascade u-net for retinal vessel segmentation. Multimedia Tools and Applications 81(11), 15593-15607 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.609, + 0.787, + 0.65 + ], + "angle": 0, + "content": "55. Yang, Z., Farsiu, S.: Directional connectivity-based segmentation of medical images. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11525-11535 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.65, + 0.787, + 0.69 + ], + "angle": 0, + "content": "56. Yim, P.J., Cebral, J.J., Mullick, R., Marcos, H.B., Choyke, P.L.: Vessel surface reconstruction with a tubular deformable model. IEEE Transactions on Medical Imaging 20(12), 1411-1421 (2001)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.69, + 0.787, + 0.731 + ], + "angle": 0, + "content": "57. Yu, F., Koltun, V., Funkhouser, T.: Dilated residual networks. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 472-480 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.731, + 0.787, + 0.785 + ], + "angle": 0, + "content": "58. Zhao, Y., Zheng, Y., Liu, Y., Zhao, Y., Luo, L., Yang, S., Na, T., Wang, Y., Liu, J.: Automatic 2-d/3-d vessel enhancement in multiple modality images using a weighted symmetry filter. IEEE Transactions on Medical Imaging 37(2), 438-450 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.785, + 0.787, + 0.84 + ], + "angle": 0, + "content": "59. Zhou, Y., Huang, J., Wang, C., Song, L., Yang, G.: Xnet: Wavelet-based low and high frequency fusion networks for fully- and semi-supervised semantic segmentation of biomedical images. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 21085-21096 (October 2023)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.322, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Fractal Feature Maps for Segmentation of Tubular Structures" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "19" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.788, + 0.231 + ], + "angle": 0, + "content": "60. Zhou, Z., Rahman Siddiquee, M.M., Tajbakhsh, N., Liang, J.: Unet++: A nested u-net architecture for medical image segmentation. In: Deep Learning in Medical Image Analysis and Multimodal Learning for Clinical Decision Support: 4th International Workshop, DLMIA 2018, and 8th International Workshop, ML-CDS 2018, Held in Conjunction with MICCAI 2018, Granada, Spain, September 20, 2018, Proceedings 4. pp. 3-11. Springer (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.232, + 0.787, + 0.273 + ], + "angle": 0, + "content": "61. Zhuang, Z., Lei, N., Joseph Raj, A.N., Qiu, S.: Application of fractal theory and fuzzy enhancement in ultrasound image segmentation. Medical & Biological Engineering & Computing 57, 623-632 (2019)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.788, + 0.273 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/Representing Topological Self-Similarity Using Fractal Feature Maps for Accurate Segmentation of Tubular Structures/494e6709-0fe6-4415-a992-88fda17780d8_origin.pdf b/2024/Representing Topological Self-Similarity Using Fractal Feature Maps for Accurate Segmentation of Tubular Structures/494e6709-0fe6-4415-a992-88fda17780d8_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..faccd200e9961c435293bdd9ce418ac237796502 --- /dev/null +++ b/2024/Representing Topological Self-Similarity Using Fractal Feature Maps for Accurate Segmentation of Tubular Structures/494e6709-0fe6-4415-a992-88fda17780d8_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:138ad3c89b8cb5feafb2bd5d7acba8a4be0cf0d0a30911ef160cbe736e33badd +size 1816839 diff --git a/2024/Representing Topological Self-Similarity Using Fractal Feature Maps for Accurate Segmentation of Tubular Structures/full.md b/2024/Representing Topological Self-Similarity Using Fractal Feature Maps for Accurate Segmentation of Tubular Structures/full.md new file mode 100644 index 0000000000000000000000000000000000000000..84097bd5c70325ff0a6a110435fc8c5dc4f79ec2 --- /dev/null +++ b/2024/Representing Topological Self-Similarity Using Fractal Feature Maps for Accurate Segmentation of Tubular Structures/full.md @@ -0,0 +1,361 @@ +# Representing Topological Self-Similarity Using Fractal Feature Maps for Accurate Segmentation of Tubular Structures + +Jiaxing Huang $^{1,2}$ , Yanfeng Zhou $^{1,2}$ , Yaoru Luo $^{1,2}$ , Guole Liu $^{1,2}$ , Heng Guo $^{3,4}$ , and Ge Yang $^{1,2}$ + +$^{1}$ Institute of Automation, Chinese Academy of Sciences, 100190, Beijing, China + +$^{2}$ School of Artificial Intelligence, University of Chinese Academy of Sciences + +3 DAMO Academy, Alibaba Group + +4 Hupan Lab, 310023, Hangzhou, China + +{huangjiaxing2021, ge.yang}@ia.ac.cn + +Abstract. Accurate segmentation of long and thin tubular structures is required in a wide variety of areas such as biology, medicine, and remote sensing. The complex topology and geometry of such structures often pose significant technical challenges. A fundamental property of such structures is their topological self-similarity, which can be quantified by fractal features such as fractal dimension (FD). In this study, we incorporate fractal features into a deep learning model by extending FD to the pixel-level using a sliding window technique. The resulting fractal feature maps (FFMs) are then incorporated as additional input to the model and additional weight in the loss function to enhance segmentation performance by utilizing the topological self-similarity. Moreover, we extend the U-Net architecture by incorporating an edge decoder and a skeleton decoder to improve boundary accuracy and skeletal continuity of segmentation, respectively. Extensive experiments on five tubular structure datasets validate the effectiveness and robustness of our approach. Furthermore, the integration of FFMs with other popular segmentation models such as HR-Net also yields performance enhancement, suggesting FFM can be incorporated as a plug-in module with different model architectures. Code and data are openly accessible at https://github.com/cbmi-group/FFM-Multi-Decoder-Network. + +Keywords: Tubular structures $\cdot$ Topological self-similarity $\cdot$ Fractal feature map $\cdot$ Deep learning $\cdot$ Plug-in module + +# 1 Introduction + +Accurate segmentation of tubular structures is of significant importance across a wide variety of areas. In the area of biological research, for example, the accurate segmentation of tubular structures such as the endoplasmic reticulum (ER, Fig. 1) is critical to the study of related human disease mechanisms [16,28]. In the area of clinical research, the accurate segmentation of blood vessels (Fig. 1) is + +![](images/28bfbbe558905be1d30148d324d73501d360da3aa552c19ecb45b9edb57b7b1d.jpg) +Fig. 1: Topological self-similarity in tubular structures. If we consider "one junction with multiple edges" as a basic component, complex tubular structures have similar components at different scales. The images on the right are magnified views of the rectangular regions on the left. + +essential to the early diagnosis of diseases such as retinopathy and stroke [39,58]. Similarly, in the area of remote sensing, the accurate extraction of roads from aerial imagery is essential for navigation and route planning [27]. However, accurate segmentation of tubular structures from images remains challenging due to factors such as their complex morphology and geometry, low image signal-to-noise ratio, and poor image contrast. + +A wide variety of techniques have been developed for the segmentation of tubular structures. Classical methods rely on manually crafted features such as intensity, texture, and shape. For example, previous studies [2,7,40,56] have utilized deformable shape models to fit tubular structures, leveraging their geometric properties. However, these techniques often cannot handle challenges posed by factors such as poor contrast, high noise and complex background. Deep learning methods [13, 25, 34, 44, 59] have revolutionized image segmentation and achieved substantial improvements in segmentation performance. Recent studies on using deep learning models for segmentation of tubular structures have focused on optimization of loss functions [3, 26, 46, 49, 51] and refinement of model structures [10, 12, 18, 37, 42]. However, these studies primarily aim to achieve high segmentation performance utilizing a limited input of images without providing additional information to their segmentation models. Tubular structures exhibit distinct topology and geometry that are vital for segmentation. In this study, we explore using characteristics of their structure to assist deep learning models in segmentation. + +One notable characteristic of tubular structures is their topological self-similarity, namely large and complex tubular structures exhibit similar topological patterns at different scales. For example, Fig. 1 shows that if "one junction with multiple edges" is considered a primary structural component, tubular structure entities such as the ER network and the retinal blood vessel network exhibit similar topology at both global and local scales. To quantify the topological self-similarity of tubular structures, we utilize the fractal theory, which characterizes self-similarity of intricate structures at different scales [35,61]. Cen- + +Intral to this theory is the fractal dimension (FD), a key parameter used previously to describe the textural attributes of images. + +In addition to topology and geometry, edges and skeletons are important characteristics in defining tubular structures. In the segmentation of these structures, boundary accuracy and skeletal continuity are crucial for downstream tasks. For example, a commonly encountered problem in the segmentation of interconnected tubular structures (Fig. 1) is the breakages in segmentation results due to factors such as low-contrast or blurring. + +To enhance the segmentation quality of tubular structures, we exploit the topological self-similarity as well as edge and skeletal characteristics of tubular structures. The main research contributions of this study are as follows: + +1) We have developed a strategy to incorporate fractal features into deep learning networks. Specifically, we extend the fractal dimension from the image-level to the pixel-level and generate the fractal feature map (FFM), which characterizes topological self-similarity and textural complexity of each region within an image or its associated label. The FFM computed from the image is denoted as $FFM_{image}$ , while the FFM derived from the label is indicated as $FFM_{label}$ . Utilization of $FFM_{image}$ as an additional model input and $FFM_{label}$ as an additional loss function weight substantially enhances segmentation performance. + +2) We develop the multi-decoder network (MD-Net) by extending the U-Net [44] architecture with an edge decoder and a skeleton decoder. These decoders enable the model to simultaneously predict the boundaries and skeletons of image objects, in addition to the primary segmentation masks. By incorporating related constraints within the loss function, our model focuses not only on achieving accurate target segmentation but also allocates increased attention to boundary delineation and skeleton preservation, thereby enhancing the overall prediction quality. + +3) We have demonstrated the versatility and robustness of FFMs. Incorporation of $FFM_{image}$ into the vanilla U-Net [44] and HR-Net [50] enhances segmentation performance, indicating that FFM can be used as a plug-in for different models. + +# 2 Related Work + +# 2.1 Tubular Structure Segmentation + +Classical methods have been proposed to improve the performance of segmenting tubular structures by taking into account their geometric characteristics. Firstly, various methods utilize active contours and compute geodesics or minimal distance curves to approximate the contours of tubular structures, thereby effectively delineating their boundaries [1,6]. Secondly, tree structure-based methods utilize intrinsic shape priors to assist segmentation of different tubular structures. These methods employ a bottom-up method to identify tubular objects and a top-down grouping strategy to recognize tree structures, generating corresponding shape priors [4]. Lastly, various centerline-based methods + +such as the one developed in [47] use multiscale detection strategies to accurately identify centerlines so that distance transform can be used to provide valuable information for segmentation of tubular structures. + +Deep learning methods have also been proposed to integrate topological and geometrical prior knowledge of tubular structures to enhance the performance of their segmentation. The integration is primarily achieved in three ways. + +1) Convolutional kernel design. The popular deformable convolution [9] and dilated convolution [57] aim to overcome the limitations of geometric transformations in CNNs and have demonstrated exceptional performance in complex segmentation tasks. Additionally, DSC-Net utilizes dynamic snake convolution to accurately capture the distinct features of tubular structures [42]. By adaptively focusing on slender and winding local structures, DSC-Net achieves improved performance in capturing the intricacy of tubular structures. + +2) Model architecture design. Various architecture designs have been proposed to learn the topological and geometrical features of tubular structures. In [31], a global transformer and dual local attention network are employed to simultaneously capture global and local features to effectively learn the complex geometric properties of tubular structures. Dong et al. [10] propose an enhanced Deformable U-Net that exploits flexible deformable convolutional layers to better generate clear boundaries for 3D cardiac cine MRI. High-frequency components that have strong capabilities to perceive thin structures are fused in [14] to enhance the performance of segmenting thin structures. + +3) Loss function design. Various loss functions have been explored for the segmentation of tubular structures. In [46], a similarity measure centerlineDice (clDice) based on the intersection of segmentation masks and their respective skeletons is introduced. A loss function based on clDice is proposed to enable the networks to generate segmentations with more accurate connectivity information and topology preservation. Araujo et al. [3] design a loss function based on the morphological closing operator that allows models to produce more topologically coherent masks and consistent vascular trees. Wong et al. [52] apply a novel Persistence Diagram Loss that quantifies topological correctness of segmentation over fine-grained structures. + +In this study, instead of relying on intricate designs of model architectures or loss functions, we incorporate FFMs as a model input and a loss function weight to enhance the perception of topological self-similarity and textures of images. + +# 2.2 Fractal Theory and Applications + +Despite their topological and geometrical complexity, tubular structures often exhibit topological similarities across different spatial scales. This observation suggests that their complex spatial patterns can be effectively described using simple texture features. Fractal geometry provides a means to describe the irregular or fragmented shapes of natural features and other intricate objects that traditional Euclidean geometry struggles to analyze [29]. Specifically, fractal features offer the capacity to describe and characterize the topological and + +geometrical complexity as well as textural composition of tubular image objects. Thus, fractal geometry has been applied to image classification and segmentation tasks. + +For classification, Roberto et al. [43] and Lin et al. [32] utilize the fractional Brownian motion (FBM) model [23] to extract fractal features of images. These features are then fed into a support vector machine or a convolutional neural network (CNN) classifier to differentiate between different objects. For segmentation, the FBM model is utilized in [33, 61] to extract fractal features from images. Such features are combined with classical methods such as thresholding and region growth techniques for image object segmentation. + +Although existing methods have demonstrated good performance by leveraging fractal features, there is still a gap in exploring the integration of fractal features with deep learning for segmentation tasks. The inherent self-similarity observed in tubular structures aligns well with the fractal theory. In this study, we address this gap by incorporating FFMs into the segmentation model, aiming to provide a new and reliable source of information to enhance segmentation performance. + +# 3 Method + +# 3.1 Fractal Feature Map + +In fractal geometry, the fractal dimension (FD) provides a quantitative measure of an image's degree of self-similarity and roughness. FD can be estimated via the property of self-similarity [41]. Given an image $A$ , it is self-similar if $A$ comprises $N_r$ distinct copies of itself scaled down by a factor of $r$ . Consequently, for an image, the FD is defined as: + +$$ +F D = \lim _ {r \rightarrow 0} \frac {\log N _ {r}}{\log (1 / r)} \tag {1} +$$ + +Although the definition of FD based on self-similarity is simple and concise, its direct estimation becomes impractical when dealing with irregular images. To overcome this challenge and estimate the FDs of images, the box-counting method [24, 30] is employed. + +Box-counting Method: Consider a grayscale image $I$ with dimensions $M \times M$ , where $L$ denotes the maximum gray level (typically $L = 255$ ). We can model the image $I$ as a three-dimensional space with $(x, y)$ indicating the two-dimensional position and a third coordinate $(z)$ denoting the gray value. Then the 3-D space is subdivided into smaller cubic regions, or "boxes", each with dimensions $k \times k \times h$ . Here, $k$ is a given scale used to be a multiple of the sidelength of a pixel in $(x, y)$ and $h$ can be a multiple of the gray level unit in $z$ -direction. Given $L$ and $k$ , the value of $h$ calculated using the following formula: + +$$ +h = (L - 1) \times k / M \tag {2} +$$ + +Given a $k \times k$ grid located at position $(i,j)$ , suppose that the minimum gray value is contained within the $m_{th}$ box $(m = \lceil minG_{i,j} / h \rceil)$ , and the maximum gray value within the $l_{th}$ box $(l = \lceil maxG_{i,j} / h \rceil)$ . The minimum number of boxes required to encompass all gray levels within the grid at $(i,j)$ is computed as: + +$$ +n _ {r} (i, j) = l - m + 1 \tag {3} +$$ + +Considering all grids, the number of boxes that can cover all the patches is expressed as: + +$$ +N _ {r} = \sum_ {i, j} n _ {r} (i, j) \tag {4} +$$ + +where $r = k / M$ . We can obtain a series of $N_{r}$ using differing values of $k$ . Finally, the fractal dimension can be estimated from the least-squares linear fitting of $\log N_{r}$ versus $\log (1 / r)$ , as illustrated in Equation (1). The flow of box-counting method is summarized in the supplementary material. + +Fractal Feature Map: Although fractal dimension and fractal features have been previously applied to classification tasks [32, 43], the inherent differences between classification and segmentation tasks preclude the direct application of image-level fractal dimension to segmentation endeavors. We extend the calculation of FD from the image-level to the pixel-level to generate the FFM of an image. As depicted in Fig. 2, the process begins with the utilization of a sliding window technique. Within a $5 \times 5$ window, for example, the FD of this region is computed using the box-counting method. Subsequently, the window is shifted along both the horizontal and vertical directions with a step size of 1, resulting in the calculation of the FFM for the entire image. The algorithm for generating the FFM is summarized in Algorithm 1. + +The FFMs of the images (denoted as $FFM_{image}$ ) are incorporated into the segmentation model as additional input channels (Image, $FFM_{image}$ ), enhancing the focus on fractal structure and self-similarity after the normalization process. To mitigate the impact of image noise on the calculation of FD, we adopt the box-counting method proposed in [29] as they substitute the gray value with the standard deviation to make the method more robust. + +# Algorithm 1 Generation of FFM. + +Input: Image $I$ with size $M \times N$ , window size $w \times w$ . + +Output: Fractal feature map (FFM) of $I$ . + +Assign $FFM \gets$ Float array of size $(M, N)$ + +Assign padding size $p \gets \lfloor w / 2 \rfloor$ + +$Pad_{I}\gets$ Linear Padding $(I,p)$ + +For $i\gets 0$ to $M - 1$ do + +For $j\gets 0$ to $N - 1$ do + +Selected region $R_{i,j}$ + +$$ +R _ {i, j} \leftarrow P a d _ {I} [ i: i + w, j: j + w ] +$$ + +$$ +F D \leftarrow \operatorname {B o x - c o u n t i n g} \left(R _ {i, j}\right) +$$ + +$$ +F F M [ i, j ] \leftarrow F D +$$ + +$$ +F F M \leftarrow \operatorname {N o r m a l i z a t i o n} (F F M) +$$ + +# 3.2 Multi-Decoder Network + +In segmenting tubular structures, accurate boundary detection and preservation of global topological connectivity are crucial requirements. To meet these re + +![](images/d3dfcc36229c8ebbc942c24040f3ee904a819d0e6fa09cd93cb59b54bcabaa61.jpg) +Fig. 2: Workflow of computing FFM of an image. + +requirements, we propose a new model that we refer to as Multi-Decoder Network (MD-Net). + +In addition to predicting the segmentation mask, we introduce an edge decoder and a skeleton decoder to generate the edge and skeleton of the tubular structure, respectively, as depicted in Fig. 3. When (Image, $FFM_{image}$ ) are fed into MD-Net, the Encoder employs convolution layers to extract a series of low-level to high-level image features. These features are simultaneously transmitted to the three Decoders by skip connections for prediction. Specifically, in the Encoder, each convolution layer involves the repeated application of two 3x3 convolutions, followed by a rectified linear unit (ReLU), and a 2x2 max pooling operation with a stride of 2 for down-sampling. As for the Decoders in MD-Net, each convolution layer includes an up-sampling of the feature map, followed by a 2x2 convolution that reduces the number of feature channels by half. Subsequently, the feature is concatenated with the corresponding feature copied from the Encoder. This concatenated feature is then processed using two 3x3 convolutions, each followed by a ReLU activation function. + +During training, the ground truths for edges and skeletons are derived from the annotated masks by employing the findContours function from the OpenCV library and the skeletonization algorithms in the scikit-image library. We have visualized the boundaries and skeletons in the supplementary material to corroborate the accuracy of the extraction techniques. During the inference stage, the final structural segmentation is the output of the object decoder. Edge and skeleton predictions are only used in model training. + +Loss Function: Given an image $I$ with $N$ pixels, the segmentation ground truth is denoted as $y$ , where foreground and background pixels are labeled as 1 and 0, respectively. The prediction of the model is represented as $\hat{y} \in [0,1]$ , indicating the probabilities of individual pixels being classified as foreground. To compute the loss, we utilize the differentiable Soft IOU Loss [19], denoted as $\mathcal{L}_{object}$ in Equation (5), where $\eta$ is a smoothing coefficient. + +$$ +\mathcal {L} _ {\text {o b j e c t}} = \mathcal {L} _ {i o u} = 1 - \frac {\sum_ {i = 1} ^ {N} y _ {i} \hat {y} _ {i}}{\sum_ {i = 1} ^ {N} y _ {i} + \hat {y} _ {i} - y _ {i} \hat {y} _ {i} + \eta} \tag {5} +$$ + +The BCE loss function is applied separately to the edge and skeleton segmentation tasks, denoted as $\mathcal{L}_{edge},\mathcal{L}_{skeleton}$ . + +$$ +\mathcal {L} _ {\text {e d g e}}, \mathcal {L} _ {\text {s k e l e t o n}} = \mathcal {L} _ {\text {b c e}} = - \frac {1}{N} \sum_ {i = 1} ^ {N} y _ {i} \cdot \log (\hat {y} _ {i}) + (1 - y _ {i}) \cdot \log (1 - \hat {y} _ {i}) \tag {6} +$$ + +For MD-Net, which has three segmentation tasks, the loss function is composed as in Equation (8). The weights $\alpha, \beta$ and $\gamma$ are typically set to 1.0, 0.5 and 0.5, respectively. + +$$ +\mathcal {L} _ {\text {g l o b a l}} = \alpha \mathcal {L} _ {\text {o b j e c t}} + \beta \mathcal {L} _ {\text {e d g e}} + \gamma \mathcal {L} _ {\text {s k e l e t o n}} \tag {7} +$$ + +# 3.3 Fractal Feature Constrained Loss + +The value of FD serves as an indicator of the texture complexity present in a given image region. Hence, it is advantageous to incorporate pixel-level FFM as weights of the loss function. This allows for a higher weight to be assigned to regions that exhibit greater complexity, as accurately segmenting such regions poses a greater challenge. Consequently, we calculate the FFM for each label associated with the image, denoted as $FFM_{label}$ , and employ $FFM_{label}$ as pixel-level weights within the loss function to guide the convergence of the model. Considering the specificity of the edge decoder and skeleton decoder, we have thus restricted the application of $FFM_{label}$ exclusively to the object decoder. The Constrained Loss of MD-Net is as follows: + +$$ +\mathcal {L} _ {\text {c o n s t r a i n e d}} = \alpha \mathcal {L} _ {\text {o b j e c t}} \cdot F F M _ {\text {l a b e l}} + \beta \mathcal {L} _ {\text {e d g e}} + \gamma \mathcal {L} _ {\text {s k e l e t o n}} \tag {8} +$$ + +# 4 Experiments + +# 4.1 Datasets + +We evaluate the FFM and MD-Net using five publicly available tubular datasets including the endoplasmic reticulum (ER) and mitochondrial (MITO) [11, 36] organelle network datasets, the Retinal OCT-Angiography vessel Segmentation (ROSE) and Structured Analysis of the Retina (STARE) [15, 37] retinal vessel + +![](images/61582cd15f228a0900dcfb446a58ef3b1def5fb8657444d24a1edcf323d681ff.jpg) +Fig. 3: Overview and details of our proposed model MD-Net. The left part illustrates the overall structure of MD-Net. (Image, $FFM_{image}$ ) is processed as input by the Encoder, which extracts features of varying sizes from the input. These features are transmitted simultaneously to three Decoders through Skip Connection. The right part is a detailed description of the Encoder and Decoders. Each Decoder performs upsampling and concatenation operations similar to U-Net to obtain the predictions, comprising the edge, skeleton, and object. + +datasets, and the Massachusetts Roads (ROAD) [38] remote sensing dataset. By selecting these datasets, our aim is to comprehensively evaluate the performance of our methods across different types of structures and domains, providing a robust assessment of their capabilities. To further examine the influence of FFM, we included the non-tubular dataset NUCLEUS [5]. + +All images in ER, MITO, STARE, and NUCLEUS are randomly sampled patches with size of $256 \times 256$ for training and sliding window with size of $256 \times 256$ for validation and testing. A crop size of $256 \times 256$ is used randomly during training and regularly during testing for the ROAD dataset as in [42]. Horizontal and vertical flipping as well as $90^{\circ} / 180^{\circ} / 270^{\circ}$ rotation are used for data augmentation. In addition, data augmentation for ROSE is conducted by rotation of an angle from $-10^{\circ}$ to $10^{\circ}$ as described in [37]. Detailed partitions of the six datasets and experimental setup are shown in the supplementary material. + +# 4.2 Performance Comparison + +We select several segmentation networks for comparison, including the vanilla U-Net [44], U-Net++ [60], nnU-Net [20], and HR-Net [50]. Additionally, we compare our methods with SOTA models for diverse datasets. For the task of retinal vessel segmentation, we choose three-stage [53], OCTA-Net [37], GT-DLA [31], and AF-Net [45], as these networks are tailored for this task. Furthermore, we compare the performance of MD-Net in the ROAD dataset with DCU-Net [54], Dconn-Net [55], and the DSC-Net [42], a tubular structure segmentation network with specially designed loss $\mathcal{L}_{tc}$ . To assess the generalization of FFM, we apply $FFM_{image}$ to both U-Net and HR-Net. All models are trained on the same dataset with the same hyperparameter settings. + +# 4.3 Evaluation Metrics + +The models are assessed using three types of metrics: volumetric, topology, and distance. + +1. Volumetric scores, including intersection-over-union (IoU), accuracy (ACC), centerlineDice (clDice) [46], and AUC. These metrics quantify the overlap and agreement between the prediction and ground truth. +2. Topology scores. To evaluate the topological correctness of the segmentations, we calculate the Betti Error [17] $\beta$ for the sum of Betti Numbers $\beta_0$ and $\beta_{1}$ . In the evaluation of ROSE and STARE, $\beta$ Error represents the $\beta_{1}$ only. +3. Distance score: Hausdorff Distance (HD) [48] quantifies the accuracy of the boundary delineation and provides information about the spatial closeness of the predicted and ground truth boundaries. + +Since the nucleus is oval in shape and not tubular, no Betti Error and clDice are used in segmentation performance evaluation and no skeletal decoder is used in the MD-Net. In this study, the evaluation metrics IoU, ACC, AUC, and clDice, are expressed as percentages (\%). The Hausdorff Distance (HD) is measured in pixels (px). All performance metrics are calculated for each image and averaged. + +Table 1: Bold numbers indicate the best performance and italicized numbers represent the next best performance. The result marked with a star (*) represents the model input replaced by (Image) with (Image, $FFM_{image}$ ). $\mathcal{L}_{tc}$ in (b) is Wasserstein-distance-based TCLoss proposed in [42]. $\mathcal{L}_{\text{constrained}}$ represents the integration of $FFM_{label}$ as a loss function weight, which leads to further improvement. + +
ModelLossIoU↑clDice↑ACC↑AUC↑βError↓HD↓IoU↑clDice↑ACC↑AUC↑βError↓HD↓
ERMITO
U-Net++ [60]Liou75.0294.6791.0497.3126.027.0479.7097.3098.1999.592.304.24
nnU-Net [20]Liou73.4194.5191.2987.9525.156.9679.2497.1698.0794.623.404.39
DSC-Net [42]Liou75.5194.4491.5697.0934.516.9280.3297.1698.1699.652.704.36
Dconn-Net [55]Liou75.9595.2491.5996.6020.316.9779.8297.2098.0399.542.904.32
AF-Net [45]Liou76.0194.5891.8197.3729.926.7980.3696.9498.0899.642.604.18
GT-DLA [31]Liou75.8994.9291.8297.3120.266.8480.2597.1698.1999.602.704.21
U-Net [44]Liou75.4494.6391.8297.3528.726.8779.7796.9198.0799.612.804.56
HR-Net [50]Liou75.8395.0891.7197.3622.576.9079.6397.2798.1499.622.904.25
U-Net*Liou76.5995.4392.0297.5620.786.8180.7197.4298.2199.632.704.27
HR-Net*Liou76.4395.4791.9597.5520.526.8380.6297.2998.1799.623.304.18
MD-Net*Lglobal77.0195.7892.0697.5919.106.7781.1197.7298.2599.662.204.16
MD-Net*Lconstrained77.0995.7492.1497.6519.526.7281.1897.6198.2699.652.804.14
ROSESTARE
three-stage [53]Liou62.1167.0791.7993.419.137.2866.8376.7194.8696.433.456.57
OCTA-Net [37]Liou63.2766.3292.3494.5310.227.1165.3776.7494.6295.443.226.47
DSC-Net [42]Liou63.1567.5991.9894.437.007.3667.0176.2694.9796.503.156.31
Dconn-Net [55]Liou60.3764.5188.5793.265.738.1866.1476.3394.7896.653.776.48
AF-Net [45]Liou62.3367.0591.7392.558.887.3267.2777.1794.9196.473.186.49
GT-DLA [31]Liou63.5766.3192.2794.189.897.1667.1276.3294.7295.793.126.63
U-Net [44]Liou61.5267.5391.3394.048.227.4266.1576.0594.6896.393.226.67
HR-Net [50]Liou63.0967.7192.0894.388.337.2667.0376.9294.9296.292.736.53
U-Net*Liou64.0767.9592.2594.477.887.1068.0777.3995.1596.632.776.33
HR-Net*Liou64.6869.4292.1494.776.017.2168.2277.4995.1396.523.216.32
MD-Net*Lglobal65.0769.7892.3694.884.227.1068.4977.7995.2096.912.576.29
MD-Net*Lconstrained65.1969.5892.4294.884.897.0668.7378.2095.2696.992.606.34
+ +(a) Quantitative experimental results for the ER, MITO, ROSE, and STARE dataset. + +
ModelLossIoU↑cDice↑ACC↑AUC↑βError↓HD↓ModelLossIoU↑ACC↑AUC↑HD↓
DCU-Net [54]Lce62.9286.9898.0398.342.568.04CS-Net [39]Liou77.9698.1298.584.13
DSC-Net [42]Lce63.9987.7498.0598.392.567.96Transnet [8]Liou78.2198.3498.644.32
DSC-Net [42]Ltc64.2287.6498.0598.462.457.34DU-Net [21]Liou78.0998.4698.944.19
Dconn-Net [55]Ltc65.1487.6197.3098.292.256.97U-Net++[60]Liou78.3398.5999.694.09
AF-Net [45]Ltc65.0387.2797.4198.382.217.29nnU-Net [20]Liou78.2898.5794.934.09
GT-DLA [31]Ltc64.7686.8097.4397.942.386.94OCTA-Net [37]Liou75.3098.3499.264.43
U-Net [44]Lce62.4786.8797.9798.292.618.11U-Net [44]Liou78.0298.5499.654.23
HR-Net [50]Lce62.8986.7498.0698.392.588.06HR-Net [50]Liou77.7098.5099.654.15
U-Net*Lce65.7487.7498.3898.722.486.98U-Net*Liou78.4698.5899.684.15
HR-Net*Lce65.6087.6198.3298.792.316.98HR-Net*Liou77.8398.5399.664.15
MD-Net*Lglobal66.0788.0898.4398.802.196.92MD-Net*Lglobal78.6698.5999.684.08
MD-Net*Lconstrained66.1688.0798.4298.772.066.97MD-Net*Lconstrained78.7798.5999.664.06
+ +(b) Results for ROAD dataset. + +(c) Results for NUCLEUS dataset. + +# 4.4 Configurations + +All models are implemented using PyTorch (version 1.12.0) and executed on 4 NVIDIA 3090 GPU cards. During the training phase, we employed the SGD optimizer with an initial learning rate of 0.05. To dynamically adjust the learning rate as the training progressed, we incorporated warm-up and exponential decay techniques. For all datasets, a fixed batch size of 32 is employed, ensuring consistency in the training process. Additionally, to prevent overfitting, a regularization weight of 0.0005 is applied. The entire training process spanned 50 epochs, providing sufficient iterations for model convergence and learning. + +# 4.5 Quantitative Evaluation + +Based on the results presented in Tab. 1, several conclusions can be drawn. + +Performance of MD-Net: Our proposed model, MD-Net, demonstrates superior performance in terms of segmentation accuracy and topological continuity compared to the other models. This superiority is observed across five tubular datasets as well as one non-tubular dataset. We have also performed t-test to check whether improvements of our methods over competing methods in performance are statistically significant. Results are provided in supplementary. + +In tubular datasets, MD-Net outperforms other methods in terms of segmentation results. Specifically, in the ER, MITO, ROSE, STARE, and ROAD datasets, MD-Net achieves improvements in IoU of $1.08\%$ , $0.82\%$ , $1.62\%$ , $1.46\%$ , and $1.02\%$ , respectively, compared to existing SOTA methods. Furthermore, when considering topological continuity and boundary extraction, MD-Net demonstrates the best performance with the minimum $\beta$ error and Hausdorff Distance. These results highlight the ability of MD-Net, equipped with an edge decoder and a skeleton decoder, to effectively capture edge and skeleton features of thin tubular structures, resulting in more accurate and continuous topology segmentation outcomes. + +Even in the non-tubular dataset NUCLEUS, MD-Net achieves the best segmentation results with an IoU of $78.77\%$ , ACC of $98.59\%$ , and HD of 4.06. This demonstrates the effectiveness of MD-Net and FFM in datasets with simpler structures. + +Generalization of FFM: $FFM_{image}$ is incorporated into U-Net and HR-Net as an additional input channel without modifying the training parameters. The results in Tab. 1 indicate that the inclusion of $FFM_{image}$ leads to improved segmentation performance for both U-Net and HR-Net. + +For U-Net, the segmentation performance (IoU) shows improvements of $1.14\%$ , $0.94\%$ , $2.55\%$ , $1.92\%$ , $3.27\%$ , and $0.44\%$ on the ER, MITO, ROSE, STARE, ROAD, and NUCLEUS datasets with the assistance of $FFM_{image}$ , respectively. Similarly, for HR-Net, the IoU improves by $0.60\%$ , $0.99\%$ , $1.59\%$ , $1.19\%$ , $2.71\%$ , and $0.13\%$ on the same datasets, respectively. On average, $FFM_{image}$ brings an improvement of $1.96\%$ and $1.42\%$ in IoU for U-Net and HR-Net, respectively, in the segmentation task of tubular datasets. + +In conclusion, FFM demonstrates notable generalization capabilities, leading to enhanced performance across various models. Particularly in datasets with intricate structures, FFM consistently outperforms the original models in segmentation tasks. This improvement can be attributed to the inherent characteristics of FD involved in FFM. + +$FFM_{label}$ and $\mathcal{L}_{global}$ : Fractal Dimension is a significant metric used to quantify the complexity of image textures. In the context of model training, we leverage the $FFMs_{label}$ as weights for the loss function. This method enables us to guide the model's focus towards regions with higher complexity, as indicated by a higher FD. Through extensive experimentation, we have observed that incorporating FFM-based weights into the loss function yields improvements in both the training performance and efficiency of the model, as depicted in Tab. 1. + +![](images/baed5fb96d32065d30b487232642e5f700fec30761cd65502c23c33f097548b3.jpg) +Fig. 4: Comparison of segmentation results. (a) Image. (b) Label. (c) Results of U-Net. (d) Results of existing SOTA approaches. From top to bottom, it's AF-Net, AF-Net, GT-DLA, AF-Net, and Dconn-Net. (e) Results of U-Net*. (f) Results of MD-Net*. Red: true positive. Green: false negative. Blue: false positive. In the ER, MITO, ROSE, STAR, and ROAD rows, columns (b) through (f) display the magnified results of the regions demarcated by yellow squares in the corresponding images of column (a). + +These findings highlight the tangible benefits of including $FFM_{label}$ as a weight parameter in optimizing the training process. + +# 4.6 Qualitative Evaluation + +To facilitate a more comprehensive comparison of the experimental outcomes across various models, we employed visualizations to present the segmentation results, as depicted in Fig. 4. The visual analysis provides a clear and intuitive representation of the model's performance. The observed results validate that the integration of FFM leads to enhanced segmentation outcomes in terms of both edge accuracy and topological continuity. More visualization results can be found in the supplementary material. + +# 5 Ablation Study + +# 5.1 Effectiveness of Fractal Feature Maps + +To ascertain that the observed performance improvement of the model is attributed to the FFM and not the increase of input channels, we replace the $FFM_{image}$ with the image, thereby transforming the input (image, $FFM_{image}$ ) to (image, image). Additionally, we computed the image's Hurst feature (HF, + +classical fractal analysis), mean feature (MF, directional analysis), and contrast feature (CF, directional analysis) as introduced by [22]. These features replaced the $FFM_{image}$ as input and were trained and evaluated on U-Net and MD-Net. Experimental results demonstrated the advantages of the $FFM_{image}$ as shown in Tab. 2. This finding indicates that the inclusion of the FFM plays an effective role in achieving better performance in the segmentation task. + +# 5.2 Robustness of Fractal Feature Maps + +The computation of the FFM is influenced by the selection of window size and step size. We used different window and step sizes to generate different $FFMs_{image}$ and tested their performance. The results in Tab. 3 demonstrate the robustness of FFM. When used for training the U-Net and MD-Net, the $FFMs_{image}$ calculated using different window sizes consistently yield results with a variation range of $0.3\%$ on the ER and STARE datasets. Similarly, applying different steps during the calculation of FFMs does not lead to significant changes in performance. It should be noted that the efficiency of computing can be substantially improved by increasing the step size. + +The numerical results in Tab. 2 and Tab. 3 are calculated by volumetric score IoU. Results with more metrics can be found in the supplementary material. + +# 5.3 Decoders of MD-Net + +To evaluate the effectiveness of the edge decoder and the skeleton decoder, a comparative ablation analysis was conducted between MD-Net and U-Net on the ER and STARE datasets. Subsequently, we proceeded to remove either the edge decoder or the skeleton decoder from MD-Net and trained the modified models. The results shown in Tab. 4 confirm that both the edge decoder and the skeleton decoder contribute to improving the segmentation performance. It is noteworthy that the best performance is achieved when all components are utilized simultaneously. + +# 5.4 Limitations of Fractal Feature Map + +The time complexity of FFMs' computation is associated with the size of image $(M\times N)$ , window size $w$ , and step size $S$ . Its time complexity is represented + +Table 2: Ablation study of FFM in U-Net and MD-Net. The values within the parentheses in the (image, $FFM_{image}$ ) column represent the percentage improvement of IoU compared to the baseline results listed in the (image) column. + +
ModelDatasetInput
(image)(image,image)(image,HF)(image,MF)(image,CF)(image,FFM image)
U-NetER75.4475.6175.8975.8675.1076.59(+1.14%)
MITO79.7780.2880.5180.2380.3880.71(+0.94%)
ROSE61.5262.5762.6162.5862.3764.07(+2.55%)
STARE66.1566.0865.5466.1466.8368.07(+1.92%)
ROAD62.4763.4464.9364.5864.2565.74(+3.27%)
MD-NetER76.2876.3476.2475.9775.5277.01(+0.73%)
MITO80.2880.3179.9780.4180.3281.11(+0.83%)
ROSE63.3163.3062.7762.8962.9065.07(+1.76%)
STARE66.4666.5766.4367.0367.2568.49(+2.03%)
ROAD64.7965.0465.2365.1265.1566.07(+1.28%)
+ +Table 3: Ablation of step size and window size in U-Net and MD-Net. + +
ModelStep SizeWindow SizeERSTAREModelStep SizeWindow SizeERSTARE
U-Net*11176.7567.87MD-Net*11177.0368.41
1976.7367.911977.0168.21
1776.6567.891776.9268.50
1576.5968.071577.0168.49
2576.4967.872576.9168.46
3576.6867.783576.8868.32
4576.6167.784576.9068.30
+ +Table 4: Ablation study of MD-Net. + +
DatasetModelFFMEdge DecoderSkeleton DecoderIoU↑clDice↑ACC↑AUC↑βError↓HD↓
ERU-Net75.4494.6391.8297.3528.726.87
MD-Net✓✓76.2895.2792.0097.3922.556.78
U-Net✓76.5995.4392.0297.5620.786.81
MD-Net✓✓76.9795.6092.0997.5922.426.71
MD-Net✓✓76.8595.5592.1397.5120.636.74
MD-Net✓✓✓77.0995.7492.1497.6519.526.72
STAREU-Net66.1576.0594.6896.393.226.67
MD-Net✓✓66.4676.5494.7696.423.276.47
U-Net✓68.0777.3995.1596.632.776.33
MD-Net✓✓68.2677.8095.2096.772.656.29
MD-Net✓✓68.3977.7695.2096.893.226.31
MD-Net✓✓✓68.7378.2095.2696.992.606.34
+ +as: $O\left(\frac{M \cdot N}{S^2} \cdot \left(w^2 \cdot \log(w)\right)\right)$ . Calculation of FFM indeed incurs computational overhead but is optimized by using multithreading and larger $S$ . Currently, the generation of FFMs for a $256 \times 256$ image takes about 150 milliseconds on a Xeon 8336C CPU, given a step size of 1 and a window size of 5. Detailed information on training and inference time can be found in the supplementary material. + +# 6 Conclusions + +In this study, we propose a method that uses fractal feature maps (FFMs) along with a multi-decoder network (MD-Net) for the semantic segmentation of tubular structures. FFMs are used to capture the texture and self-similarity of image regions or label regions at the pixel-level through the fractal dimension. $FFM_{image}$ are used as an additional input to enhance the segmentation model's perception of tubular structures and $FFM_{label}$ are used as a weight for the loss function to guide the model training. MD-Net improves the quality of segmentation by simultaneously predicting edges and skeletons through the incorporation of an edge decoder and a skeleton decoder. The proposed method is evaluated on five datasets of tubular structures and one dataset of non-tubular structures. The results demonstrate the superiority of MD-Net with FFM in the segmentation of tubular structures and stable performance in the segmentation of non-tubular structures. Furthermore, the performance improvement observed when incorporating the FFM module into vanilla U-Net and HR-Net underscores the broad applicability and potential of incorporating fractal information into deep learning models. + +# Acknowledgements + +This work was supported in part by the National Natural Science Foundation of China (grants 92354307, 91954201, 31971289, 32101216), the Strategic Priority Research Program of the Chinese Academy of Sciences (grant XDB37040402) and the Fundamental Research Funds for the Central Universities (grant E3E45201X2). + +# References + +1. Alvarez, L., Trujillo, A., Cuenca, C., González, E., Esclarín, J., Gomez, L., Mazarra, L., Aleman-Flores, M., Tahoces, P.G., Carreira, J.M.: Tracking the aortic lumen geometry by optimizing the 3d orientation of its cross-sections. In: Medical Image Computing and Computer-Assisted Intervention- MICCAI 2017: 20th International Conference, Quebec City, QC, Canada, September 11-13, 2017, Proceedings, Part II 20. pp. 174-181. Springer (2017) +2. Antiga, L., Ene-Iordache, B., Remuzzi, A.: Computational geometry for patient-specific reconstruction and meshing of blood vessels from mr and ct angiography. IEEE Transactions on Medical Imaging 22(5), 674-684 (2003) +3. Araújo, R.J., Cardoso, J.S., Oliveira, H.P.: Topological similarity index and loss function for blood vessel segmentation. arXiv preprint arXiv:2107.14531 (2021) +4. Bauer, C., Pock, T., Sorantin, E., Bischof, H., Beichel, R.: Segmentation of interwoven 3d tubular tree structures utilizing shape priors and graph cuts. Medical Image Analysis 14(2), 172-184 (2010) +5. Caicedo, J.C., Goodman, A., Karhohs, K.W., Cimini, B.A., Ackerman, J., Haghighi, M., Heng, C., Becker, T., Doan, M., McQuin, C., et al.: Nucleus segmentation across imaging experiments: the 2018 data science bowl. Nature Methods 16(12), 1247-1253 (2019) +6. Caselles, V., Kimmel, R., Sapiro, G.: Geodesic active contours. International Journal of Computer Vision 22, 61-79 (1997) +7. Challoo, M., Gao, Y., Busch, A.: Distinctive phase interdependency model for retinal vasculature delineation in oct-angiography images. IEEE Transactions on Medical Imaging (2023) +8. Chen, J., Lu, Y., Yu, Q., Luo, X., Adeli, E., Wang, Y., Lu, L., Yuille, A.L., Zhou, Y.: Transunet: Transformers make strong encoders for medical image segmentation. arXiv preprint arXiv:2102.04306 (2021) +9. Dai, J., Qi, H., Xiong, Y., Li, Y., Zhang, G., Hu, H., Wei, Y.: Deformable convolutional networks. In: Proceedings of the IEEE international conference on computer vision. pp. 764-773 (2017) +0. Dong, S., Pan, Z., Fu, Y., Yang, Q., Gao, Y., Yu, T., Shi, Y., Zhuo, C.: Deu-net 2.0: Enhanced deformable u-net for 3d cardiac cine migraine segmentation. Medical Image Analysis 78, 102389 (2022) +11. Guo, Y., Huang, J., Zhou, Y., Luo, Y., Li, W., Yang, G.: Segmentation of intracellular structures in fluorescence microscopy images by fusing low-level features. In: Pattern Recognition and Computer Vision: 4th Chinese Conference, PRCV 2021, Beijing, China, October 29–November 1, 2021, Proceedings, Part III 4. pp. 386–397. Springer (2021) + +12. Gupta, S., Zhang, Y., Hu, X., Prasanna, P., Chen, C.: Topology-aware uncertainty for image segmentation. Advances in Neural Information Processing Systems 36 (2024) +13. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 770-778 (2016) +14. He, Y., Ge, R., Wu, J., Coatrieux, J.L., Shu, H., Chen, Y., Yang, G., Li, S.: Thin semantics enhancement via high-frequency priori rule for thin structures segmentation. In: 2021 IEEE International Conference on Data Mining (ICDM). pp. 1096-1101. IEEE (2021) +15. Hoover, A., Goldbaum, M.: Locating the optic nerve in a retinal image using the fuzzy convergence of the blood vessels. IEEE Transactions on Medical Imaging 22(8), 951-958 (2003) +16. Hotamisligil, G.S.: Endoplasmic reticulum stress and the inflammatory basis of metabolic disease. Cell 140(6), 900-917 (2010) +17. Hu, X., Li, F., Samaras, D., Chen, C.: Topology-preserving deep image segmentation. Advances in neural information processing systems 32 (2019) +18. Hu, X., Wang, Y., Fuxin, L., Samaras, D., Chen, C.: Topology-aware segmentation using discrete Morse theory. arXiv preprint arXiv:2103.09992 (2021) +19. Huang, Y., Tang, Z., Chen, D., Su, K., Chen, C.: Batching soft iou for training semantic segmentation networks. IEEE Signal Processing Letters 27, 66-70 (2019) +20. Isensee, F., Jaeger, P.F., Kohl, S.A., Petersen, J., Maier-Hein, K.H.: nnu-net: a self-configuring method for deep learning-based biomedical image segmentation. Nature Methods 18(2), 203-211 (2021) +21. Jin, Q., Meng, Z., Pham, T.D., Chen, Q., Wei, L., Su, R.: Dunet: A deformable network for retinal vessel segmentation. Knowledge-Based Systems 178, 149-162 (2019) +22. Kaplan, L.M.: Extended fractal analysis for texture classification and segmentation. IEEE transactions on image processing 8(11), 1572-1585 (1999) +23. Kaplan, L.M., Kuo, C.C.: Extending self-similarity for fractional brownian motion. IEEE Transactions on Signal Processing 42(12), 3526-3530 (1994) +24. Konatar, I., Popovic, T., Popovic, N.: Box-counting method in python for fractal analysis of biomedical images. In: 2020 24th International Conference on Information Technology (IT). pp. 1-4. IEEE (2020) +25. Kreitner, L., Paetzold, J.C., Rauch, N., Chen, C., Hagag, A.M., Fayed, A.E., Sivaprasad, S., Rausch, S., Weichsel, J., Menze, B.H., et al.: Synthetic optical coherence tomography angiographs for detailed retinal vessel segmentation without human annotations. IEEE Transactions on Medical Imaging (2024) +26. La Barbera, G., Rouet, L., Boussaid, H., Lubet, A., Kassir, R., Sarnacki, S., Gori, P., Bloch, I.: Tubular structures segmentation of pediatric abdominal-visceral cect images with renal tumors: assessment, comparison and improvement. Medical Image Analysis 90, 102986 (2023) +27. Laptev, I., Mayer, H., Lindeberg, T., Eckstein, W., Steger, C., Baumgartner, A.: Automatic extraction of roads from aerial images based on scale space and snakes. Machine Vision and Applications 12, 23-31 (2000) +28. Lee, C.A., Blackstone, C.: Er morphology and endo-lysosomal crosstalk: Functions and disease implications. Biochimica et Biophysica Acta (BBA)-Molecular and Cell Biology of Lipids 1865(1), 158544 (2020) +29. Lee, W.L., Hsieh, K.S.: A robust algorithm for the fractal dimension of images and its applications to the classification of natural images and ultrasonic liver images. Signal Processing 90(6), 1894-1904 (2010) + +30. Li, J., Du, Q., Sun, C.: An improved box-counting method for image fractal dimension estimation. Pattern Recognition 42(11), 2460-2469 (2009) +31. Li, Y., Zhang, Y., Liu, J.Y., Wang, K., Zhang, K., Zhang, G.S., Liao, X.F., Yang, G.: Global transformer and dual local attention network via deep-shallow hierarchical feature fusion for retinal vessel segmentation. IEEE Transactions on Cybernetics 53(9), 5826-5839 (2023) +32. Lin, P.L., Huang, P.W., Lee, C.H., Wu, M.T.: Automatic classification for solitary pulmonary nodule in ct image by fractal analysis based on fractional brownian motion model. Pattern Recognition 46(12), 3279-3287 (2013) +33. Lin, P., Huang, P., Huang, P., Hsu, H.: Alveolar bone-loss area localization in periodontitis radiographs based on threshold segmentation with a hybrid feature fused of intensity and the h-value of fractional brownian motion model. Computer Methods and Programs in Biomedicine 121(3), 117-126 (2015) +34. Long, J., Shelhamer, E., Darrell, T.: Fully convolutional networks for semantic segmentation. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 3431-3440 (2015) +35. Lopes, R., Betrouni, N.: Fractal and multifractal analysis: A review. Medical Image Analysis 13(4), 634-649 (2009) +36. Luo, Y., Guo, Y., Li, W., Liu, G., Yang, G.: Fluorescence microscopy image datasets for deep learning segmentation of intracellular organelle networks (2020). https://doi.org/10.21227/t2he-zn97, https://dx.doi.org/10.21227/t2he-zn97 +37. Ma, Y., Hao, H., Xie, J., Fu, H., Zhang, J., Yang, J., Wang, Z., Liu, J., Zheng, Y., Zhao, Y.: Rose: a retinal oct-angiography vessel segmentation dataset and new model. IEEE Transactions on Medical Imaging 40(3), 928–939 (2020) +38. Mnih, V.: Machine learning for aerial image labeling. University of Toronto (Canada) (2013) +39. Mou, L., Zhao, Y., Chen, L., Cheng, J., Gu, Z., Hao, H., Qi, H., Zheng, Y., Frangi, A., Liu, J.: Cs-net: Channel and spatial attention network for curvilinear structure segmentation. In: Medical Image Computing and Computer Assisted Intervention-MICCAI 2019: 22nd International Conference, Shenzhen, China, October 13-17, 2019, Proceedings, Part I 22. pp. 721-730. Springer (2019) +40. Nain, D., Yezzi, A., Turk, G.: Vessel segmentation using a shape driven flow. In: Medical Image Computing and Computer-Assisted Intervention-MICCAI 2004: 7th International Conference, Saint-Malo, France, September 26-29, 2004. Proceedings, Part I 7. pp. 51-59. Springer (2004) +41. Pentland, A.P.: Fractal-based description of natural scenes. IEEE Transactions on Pattern Analysis and Machine Intelligence (6), 661-674 (1984) +42. Qi, Y., He, Y., Qi, X., Zhang, Y., Yang, G.: Dynamic snake convolution based on topological geometric constraints for tubular structure segmentation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 6070-6079 (2023) +43. Roberto, G.F., Lumini, A., Neves, L.A., do Nascimento, M.Z.: Fractal neural network: A new ensemble of fractal geometry and convolutional neural networks for the classification of histology images. Expert Systems with Applications 166, 114103 (2021) +44. Ronneberger, O., Fischer, P., Brox, T.: U-net: Convolutional networks for biomedical image segmentation. In: Medical Image Computing and Computer-Assisted Intervention-MICCAI 2015: 18th International Conference, Munich, Germany, October 5-9, 2015, Proceedings, Part III 18. pp. 234-241. Springer (2015) + +45. Shi, T., Ding, X., Zhou, W., Pan, F., Yan, Z., Bai, X., Yang, X.: Affinity feature strengthening for accurate, complete and robust vessel segmentation. IEEE Journal of Biomedical and Health Informatics 27(8), 4006-4017 (2023) +46. Shit, S., Paetzold, J.C., Sekuboyina, A., Ezhov, I., Unger, A., Zhylka, A., Pluim, J.P., Bauer, U., Menze, B.H.: cldice-a novel topology-preserving loss function for tubular structure segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16560-16569 (2021) +47. Sironi, A., Lepetit, V., Fua, P.: Multiscale centerline detection by learning a scale-space distance transform. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 2697-2704 (2014) +48. Taha, A.A., Hanbury, A.: Metrics for evaluating 3d medical image segmentation: analysis, selection, and tool. BMC Medical Imaging 15(1), 1-28 (2015) +49. Wang, C., Hayashi, Y., Oda, M., Itoh, H., Kitasaka, T., Frangi, A.F., Mori, K.: Tubular structure segmentation using spatially connected network with radial distance loss for 3d medical images. In: Medical Image Computing and Computer Assisted Intervention-MICCAI 2019: 22nd International Conference, Shenzhen, China, October 13-17, 2019, Proceedings, Part VI 22. pp. 348-356. Springer (2019) +50. Wang, J., Sun, K., Cheng, T., Jiang, B., Deng, C., Zhao, Y., Liu, D., Mu, Y., Tan, M., Wang, X., et al.: Deep high-resolution representation learning for visual recognition. IEEE Transactions on Pattern Analysis and Machine Intelligence 43(10), 3349-3364 (2020) +51. Wang, Y., Wei, X., Liu, F., Chen, J., Zhou, Y., Shen, W., Fishman, E.K., Yuille, A.L.: Deep distance transform for tubular structure segmentation in ct scans. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3833-3842 (2020) +52. Wong, C.C., Vong, C.M.: Persistent homology based graph convolution network for fine-grained 3d shape segmentation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 7098-7107 (2021) +53. Yan, Z., Yang, X., Cheng, K.T.: A three-stage deep learning model for accurate retinal vessel segmentation. IEEE Journal of Biomedical and Health Informatics 23(4), 1427-1436 (2018) +54. Yang, X., Li, Z., Guo, Y., Zhou, D.: Dcu-net: A deformable convolutional neural network based on cascade u-net for retinal vessel segmentation. Multimedia Tools and Applications 81(11), 15593-15607 (2022) +55. Yang, Z., Farsiu, S.: Directional connectivity-based segmentation of medical images. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11525-11535 (2023) +56. Yim, P.J., Cebral, J.J., Mullick, R., Marcos, H.B., Choyke, P.L.: Vessel surface reconstruction with a tubular deformable model. IEEE Transactions on Medical Imaging 20(12), 1411-1421 (2001) +57. Yu, F., Koltun, V., Funkhouser, T.: Dilated residual networks. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 472-480 (2017) +58. Zhao, Y., Zheng, Y., Liu, Y., Zhao, Y., Luo, L., Yang, S., Na, T., Wang, Y., Liu, J.: Automatic 2-d/3-d vessel enhancement in multiple modality images using a weighted symmetry filter. IEEE Transactions on Medical Imaging 37(2), 438-450 (2017) +59. Zhou, Y., Huang, J., Wang, C., Song, L., Yang, G.: Xnet: Wavelet-based low and high frequency fusion networks for fully- and semi-supervised semantic segmentation of biomedical images. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 21085-21096 (October 2023) + +60. Zhou, Z., Rahman Siddiquee, M.M., Tajbakhsh, N., Liang, J.: Unet++: A nested u-net architecture for medical image segmentation. In: Deep Learning in Medical Image Analysis and Multimodal Learning for Clinical Decision Support: 4th International Workshop, DLMIA 2018, and 8th International Workshop, ML-CDS 2018, Held in Conjunction with MICCAI 2018, Granada, Spain, September 20, 2018, Proceedings 4. pp. 3-11. Springer (2018) +61. Zhuang, Z., Lei, N., Joseph Raj, A.N., Qiu, S.: Application of fractal theory and fuzzy enhancement in ultrasound image segmentation. Medical & Biological Engineering & Computing 57, 623-632 (2019) \ No newline at end of file diff --git a/2024/Representing Topological Self-Similarity Using Fractal Feature Maps for Accurate Segmentation of Tubular Structures/images.zip b/2024/Representing Topological Self-Similarity Using Fractal Feature Maps for Accurate Segmentation of Tubular Structures/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..5bdc3b2c80eb9fbfe1227f0b3b1fa1488eab268d --- /dev/null +++ b/2024/Representing Topological Self-Similarity Using Fractal Feature Maps for Accurate Segmentation of Tubular Structures/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3457432aa8b75176906c35e28618ef1fece0c32f8acc84073cf11fa2aae36415 +size 624569 diff --git a/2024/Representing Topological Self-Similarity Using Fractal Feature Maps for Accurate Segmentation of Tubular Structures/layout.json b/2024/Representing Topological Self-Similarity Using Fractal Feature Maps for Accurate Segmentation of Tubular Structures/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..ce483a707e50f446162f24c6d8bfd77e47b0b4c9 --- /dev/null +++ b/2024/Representing Topological Self-Similarity Using Fractal Feature Maps for Accurate Segmentation of Tubular Structures/layout.json @@ -0,0 +1,10961 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 136, + 111, + 478, + 163 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 111, + 478, + 163 + ], + "spans": [ + { + "bbox": [ + 136, + 111, + 478, + 163 + ], + "type": "text", + "content": "Representing Topological Self-Similarity Using Fractal Feature Maps for Accurate Segmentation of Tubular Structures" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 153, + 185, + 462, + 212 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 153, + 185, + 462, + 212 + ], + "spans": [ + { + "bbox": [ + 153, + 185, + 462, + 212 + ], + "type": "text", + "content": "Jiaxing Huang" + }, + { + "bbox": [ + 153, + 185, + 462, + 212 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 153, + 185, + 462, + 212 + ], + "type": "text", + "content": ", Yanfeng Zhou" + }, + { + "bbox": [ + 153, + 185, + 462, + 212 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 153, + 185, + 462, + 212 + ], + "type": "text", + "content": ", Yaoru Luo" + }, + { + "bbox": [ + 153, + 185, + 462, + 212 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 153, + 185, + 462, + 212 + ], + "type": "text", + "content": ", Guole Liu" + }, + { + "bbox": [ + 153, + 185, + 462, + 212 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 153, + 185, + 462, + 212 + ], + "type": "text", + "content": ", Heng Guo" + }, + { + "bbox": [ + 153, + 185, + 462, + 212 + ], + "type": "inline_equation", + "content": "^{3,4}" + }, + { + "bbox": [ + 153, + 185, + 462, + 212 + ], + "type": "text", + "content": ", and Ge Yang" + }, + { + "bbox": [ + 153, + 185, + 462, + 212 + ], + "type": "inline_equation", + "content": "^{1,2}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 142, + 219, + 472, + 231 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 142, + 219, + 472, + 231 + ], + "spans": [ + { + "bbox": [ + 142, + 219, + 472, + 231 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 142, + 219, + 472, + 231 + ], + "type": "text", + "content": " Institute of Automation, Chinese Academy of Sciences, 100190, Beijing, China" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 149, + 231, + 463, + 241 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 231, + 463, + 241 + ], + "spans": [ + { + "bbox": [ + 149, + 231, + 463, + 241 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 149, + 231, + 463, + 241 + ], + "type": "text", + "content": " School of Artificial Intelligence, University of Chinese Academy of Sciences" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 234, + 241, + 380, + 253 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 234, + 241, + 380, + 253 + ], + "spans": [ + { + "bbox": [ + 234, + 241, + 380, + 253 + ], + "type": "text", + "content": "3 DAMO Academy, Alibaba Group" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 224, + 253, + 389, + 264 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 224, + 253, + 389, + 264 + ], + "spans": [ + { + "bbox": [ + 224, + 253, + 389, + 264 + ], + "type": "text", + "content": "4 Hupan Lab, 310023, Hangzhou, China" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 221, + 264, + 392, + 275 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 221, + 264, + 392, + 275 + ], + "spans": [ + { + "bbox": [ + 221, + 264, + 392, + 275 + ], + "type": "text", + "content": "{huangjiaxing2021, ge.yang}@ia.ac.cn" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 159, + 306, + 455, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 306, + 455, + 525 + ], + "spans": [ + { + "bbox": [ + 159, + 306, + 455, + 525 + ], + "type": "text", + "content": "Abstract. Accurate segmentation of long and thin tubular structures is required in a wide variety of areas such as biology, medicine, and remote sensing. The complex topology and geometry of such structures often pose significant technical challenges. A fundamental property of such structures is their topological self-similarity, which can be quantified by fractal features such as fractal dimension (FD). In this study, we incorporate fractal features into a deep learning model by extending FD to the pixel-level using a sliding window technique. The resulting fractal feature maps (FFMs) are then incorporated as additional input to the model and additional weight in the loss function to enhance segmentation performance by utilizing the topological self-similarity. Moreover, we extend the U-Net architecture by incorporating an edge decoder and a skeleton decoder to improve boundary accuracy and skeletal continuity of segmentation, respectively. Extensive experiments on five tubular structure datasets validate the effectiveness and robustness of our approach. Furthermore, the integration of FFMs with other popular segmentation models such as HR-Net also yields performance enhancement, suggesting FFM can be incorporated as a plug-in module with different model architectures. Code and data are openly accessible at https://github.com/cbmi-group/FFM-Multi-Decoder-Network." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 159, + 536, + 455, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 536, + 455, + 559 + ], + "spans": [ + { + "bbox": [ + 159, + 536, + 455, + 559 + ], + "type": "text", + "content": "Keywords: Tubular structures " + }, + { + "bbox": [ + 159, + 536, + 455, + 559 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 536, + 455, + 559 + ], + "type": "text", + "content": " Topological self-similarity " + }, + { + "bbox": [ + 159, + 536, + 455, + 559 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 536, + 455, + 559 + ], + "type": "text", + "content": " Fractal feature map " + }, + { + "bbox": [ + 159, + 536, + 455, + 559 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 536, + 455, + 559 + ], + "type": "text", + "content": " Deep learning " + }, + { + "bbox": [ + 159, + 536, + 455, + 559 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 536, + 455, + 559 + ], + "type": "text", + "content": " Plug-in module" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 579, + 230, + 593 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 579, + 230, + 593 + ], + "spans": [ + { + "bbox": [ + 132, + 579, + 230, + 593 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "text", + "content": "Accurate segmentation of tubular structures is of significant importance across a wide variety of areas. In the area of biological research, for example, the accurate segmentation of tubular structures such as the endoplasmic reticulum (ER, Fig. 1) is critical to the study of related human disease mechanisms [16,28]. In the area of clinical research, the accurate segmentation of blood vessels (Fig. 1) is" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 187, + 114, + 430, + 236 + ], + "blocks": [ + { + "bbox": [ + 187, + 114, + 430, + 236 + ], + "lines": [ + { + "bbox": [ + 187, + 114, + 430, + 236 + ], + "spans": [ + { + "bbox": [ + 187, + 114, + 430, + 236 + ], + "type": "image", + "image_path": "28bfbbe558905be1d30148d324d73501d360da3aa552c19ecb45b9edb57b7b1d.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 245, + 482, + 291 + ], + "lines": [ + { + "bbox": [ + 131, + 245, + 482, + 291 + ], + "spans": [ + { + "bbox": [ + 131, + 245, + 482, + 291 + ], + "type": "text", + "content": "Fig. 1: Topological self-similarity in tubular structures. If we consider \"one junction with multiple edges\" as a basic component, complex tubular structures have similar components at different scales. The images on the right are magnified views of the rectangular regions on the left." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 304, + 482, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 304, + 482, + 376 + ], + "spans": [ + { + "bbox": [ + 130, + 304, + 482, + 376 + ], + "type": "text", + "content": "essential to the early diagnosis of diseases such as retinopathy and stroke [39,58]. Similarly, in the area of remote sensing, the accurate extraction of roads from aerial imagery is essential for navigation and route planning [27]. However, accurate segmentation of tubular structures from images remains challenging due to factors such as their complex morphology and geometry, low image signal-to-noise ratio, and poor image contrast." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 377, + 482, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 377, + 482, + 568 + ], + "spans": [ + { + "bbox": [ + 130, + 377, + 482, + 568 + ], + "type": "text", + "content": "A wide variety of techniques have been developed for the segmentation of tubular structures. Classical methods rely on manually crafted features such as intensity, texture, and shape. For example, previous studies [2,7,40,56] have utilized deformable shape models to fit tubular structures, leveraging their geometric properties. However, these techniques often cannot handle challenges posed by factors such as poor contrast, high noise and complex background. Deep learning methods [13, 25, 34, 44, 59] have revolutionized image segmentation and achieved substantial improvements in segmentation performance. Recent studies on using deep learning models for segmentation of tubular structures have focused on optimization of loss functions [3, 26, 46, 49, 51] and refinement of model structures [10, 12, 18, 37, 42]. However, these studies primarily aim to achieve high segmentation performance utilizing a limited input of images without providing additional information to their segmentation models. Tubular structures exhibit distinct topology and geometry that are vital for segmentation. In this study, we explore using characteristics of their structure to assist deep learning models in segmentation." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": "One notable characteristic of tubular structures is their topological self-similarity, namely large and complex tubular structures exhibit similar topological patterns at different scales. For example, Fig. 1 shows that if \"one junction with multiple edges\" is considered a primary structural component, tubular structure entities such as the ER network and the retinal blood vessel network exhibit similar topology at both global and local scales. To quantify the topological self-similarity of tubular structures, we utilize the fractal theory, which characterizes self-similarity of intricate structures at different scales [35,61]. Cen-" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 228, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 228, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 228, + 102 + ], + "type": "text", + "content": "J. Huang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": "Intral to this theory is the fractal dimension (FD), a key parameter used previously to describe the textural attributes of images." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 140, + 480, + 212 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 140, + 480, + 212 + ], + "spans": [ + { + "bbox": [ + 130, + 140, + 480, + 212 + ], + "type": "text", + "content": "In addition to topology and geometry, edges and skeletons are important characteristics in defining tubular structures. In the segmentation of these structures, boundary accuracy and skeletal continuity are crucial for downstream tasks. For example, a commonly encountered problem in the segmentation of interconnected tubular structures (Fig. 1) is the breakages in segmentation results due to factors such as low-contrast or blurring." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 213, + 480, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 213, + 480, + 247 + ], + "spans": [ + { + "bbox": [ + 130, + 213, + 480, + 247 + ], + "type": "text", + "content": "To enhance the segmentation quality of tubular structures, we exploit the topological self-similarity as well as edge and skeletal characteristics of tubular structures. The main research contributions of this study are as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 249, + 480, + 344 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 249, + 480, + 344 + ], + "spans": [ + { + "bbox": [ + 130, + 249, + 480, + 344 + ], + "type": "text", + "content": "1) We have developed a strategy to incorporate fractal features into deep learning networks. Specifically, we extend the fractal dimension from the image-level to the pixel-level and generate the fractal feature map (FFM), which characterizes topological self-similarity and textural complexity of each region within an image or its associated label. The FFM computed from the image is denoted as " + }, + { + "bbox": [ + 130, + 249, + 480, + 344 + ], + "type": "inline_equation", + "content": "FFM_{image}" + }, + { + "bbox": [ + 130, + 249, + 480, + 344 + ], + "type": "text", + "content": ", while the FFM derived from the label is indicated as " + }, + { + "bbox": [ + 130, + 249, + 480, + 344 + ], + "type": "inline_equation", + "content": "FFM_{label}" + }, + { + "bbox": [ + 130, + 249, + 480, + 344 + ], + "type": "text", + "content": ". Utilization of " + }, + { + "bbox": [ + 130, + 249, + 480, + 344 + ], + "type": "inline_equation", + "content": "FFM_{image}" + }, + { + "bbox": [ + 130, + 249, + 480, + 344 + ], + "type": "text", + "content": " as an additional model input and " + }, + { + "bbox": [ + 130, + 249, + 480, + 344 + ], + "type": "inline_equation", + "content": "FFM_{label}" + }, + { + "bbox": [ + 130, + 249, + 480, + 344 + ], + "type": "text", + "content": " as an additional loss function weight substantially enhances segmentation performance." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 345, + 480, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 345, + 480, + 441 + ], + "spans": [ + { + "bbox": [ + 130, + 345, + 480, + 441 + ], + "type": "text", + "content": "2) We develop the multi-decoder network (MD-Net) by extending the U-Net [44] architecture with an edge decoder and a skeleton decoder. These decoders enable the model to simultaneously predict the boundaries and skeletons of image objects, in addition to the primary segmentation masks. By incorporating related constraints within the loss function, our model focuses not only on achieving accurate target segmentation but also allocates increased attention to boundary delineation and skeleton preservation, thereby enhancing the overall prediction quality." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 441, + 480, + 487 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 441, + 480, + 487 + ], + "spans": [ + { + "bbox": [ + 130, + 441, + 480, + 487 + ], + "type": "text", + "content": "3) We have demonstrated the versatility and robustness of FFMs. Incorporation of " + }, + { + "bbox": [ + 130, + 441, + 480, + 487 + ], + "type": "inline_equation", + "content": "FFM_{image}" + }, + { + "bbox": [ + 130, + 441, + 480, + 487 + ], + "type": "text", + "content": " into the vanilla U-Net [44] and HR-Net [50] enhances segmentation performance, indicating that FFM can be used as a plug-in for different models." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 509, + 237, + 521 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 509, + 237, + 521 + ], + "spans": [ + { + "bbox": [ + 131, + 509, + 237, + 521 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 536, + 324, + 548 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 536, + 324, + 548 + ], + "spans": [ + { + "bbox": [ + 131, + 536, + 324, + 548 + ], + "type": "text", + "content": "2.1 Tubular Structure Segmentation" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 558, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 558, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 558, + 481, + 665 + ], + "type": "text", + "content": "Classical methods have been proposed to improve the performance of segmenting tubular structures by taking into account their geometric characteristics. Firstly, various methods utilize active contours and compute geodesics or minimal distance curves to approximate the contours of tubular structures, thereby effectively delineating their boundaries [1,6]. Secondly, tree structure-based methods utilize intrinsic shape priors to assist segmentation of different tubular structures. These methods employ a bottom-up method to identify tubular objects and a top-down grouping strategy to recognize tree structures, generating corresponding shape priors [4]. Lastly, various centerline-based methods" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 197, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 197, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 197, + 91, + 447, + 102 + ], + "type": "text", + "content": "Fractal Feature Maps for Segmentation of Tubular Structures" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "content": "such as the one developed in [47] use multiscale detection strategies to accurately identify centerlines so that distance transform can be used to provide valuable information for segmentation of tubular structures." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 152, + 480, + 199 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 152, + 480, + 199 + ], + "spans": [ + { + "bbox": [ + 130, + 152, + 480, + 199 + ], + "type": "text", + "content": "Deep learning methods have also been proposed to integrate topological and geometrical prior knowledge of tubular structures to enhance the performance of their segmentation. The integration is primarily achieved in three ways." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 200, + 481, + 283 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 200, + 481, + 283 + ], + "spans": [ + { + "bbox": [ + 130, + 200, + 481, + 283 + ], + "type": "text", + "content": "1) Convolutional kernel design. The popular deformable convolution [9] and dilated convolution [57] aim to overcome the limitations of geometric transformations in CNNs and have demonstrated exceptional performance in complex segmentation tasks. Additionally, DSC-Net utilizes dynamic snake convolution to accurately capture the distinct features of tubular structures [42]. By adaptively focusing on slender and winding local structures, DSC-Net achieves improved performance in capturing the intricacy of tubular structures." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 284, + 482, + 390 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 284, + 482, + 390 + ], + "spans": [ + { + "bbox": [ + 130, + 284, + 482, + 390 + ], + "type": "text", + "content": "2) Model architecture design. Various architecture designs have been proposed to learn the topological and geometrical features of tubular structures. In [31], a global transformer and dual local attention network are employed to simultaneously capture global and local features to effectively learn the complex geometric properties of tubular structures. Dong et al. [10] propose an enhanced Deformable U-Net that exploits flexible deformable convolutional layers to better generate clear boundaries for 3D cardiac cine MRI. High-frequency components that have strong capabilities to perceive thin structures are fused in [14] to enhance the performance of segmenting thin structures." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 391, + 482, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 391, + 482, + 510 + ], + "spans": [ + { + "bbox": [ + 130, + 391, + 482, + 510 + ], + "type": "text", + "content": "3) Loss function design. Various loss functions have been explored for the segmentation of tubular structures. In [46], a similarity measure centerlineDice (clDice) based on the intersection of segmentation masks and their respective skeletons is introduced. A loss function based on clDice is proposed to enable the networks to generate segmentations with more accurate connectivity information and topology preservation. Araujo et al. [3] design a loss function based on the morphological closing operator that allows models to produce more topologically coherent masks and consistent vascular trees. Wong et al. [52] apply a novel Persistence Diagram Loss that quantifies topological correctness of segmentation over fine-grained structures." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 511, + 481, + 547 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 511, + 481, + 547 + ], + "spans": [ + { + "bbox": [ + 130, + 511, + 481, + 547 + ], + "type": "text", + "content": "In this study, instead of relying on intricate designs of model architectures or loss functions, we incorporate FFMs as a model input and a loss function weight to enhance the perception of topological self-similarity and textures of images." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 563, + 326, + 575 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 563, + 326, + 575 + ], + "spans": [ + { + "bbox": [ + 131, + 563, + 326, + 575 + ], + "type": "text", + "content": "2.2 Fractal Theory and Applications" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "content": "Despite their topological and geometrical complexity, tubular structures often exhibit topological similarities across different spatial scales. This observation suggests that their complex spatial patterns can be effectively described using simple texture features. Fractal geometry provides a means to describe the irregular or fragmented shapes of natural features and other intricate objects that traditional Euclidean geometry struggles to analyze [29]. Specifically, fractal features offer the capacity to describe and characterize the topological and" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 228, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 228, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 228, + 102 + ], + "type": "text", + "content": "J. Huang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 151 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 151 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 151 + ], + "type": "text", + "content": "geometrical complexity as well as textural composition of tubular image objects. Thus, fractal geometry has been applied to image classification and segmentation tasks." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 152, + 480, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 152, + 480, + 236 + ], + "spans": [ + { + "bbox": [ + 130, + 152, + 480, + 236 + ], + "type": "text", + "content": "For classification, Roberto et al. [43] and Lin et al. [32] utilize the fractional Brownian motion (FBM) model [23] to extract fractal features of images. These features are then fed into a support vector machine or a convolutional neural network (CNN) classifier to differentiate between different objects. For segmentation, the FBM model is utilized in [33, 61] to extract fractal features from images. Such features are combined with classical methods such as thresholding and region growth techniques for image object segmentation." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 236, + 481, + 320 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 236, + 481, + 320 + ], + "spans": [ + { + "bbox": [ + 130, + 236, + 481, + 320 + ], + "type": "text", + "content": "Although existing methods have demonstrated good performance by leveraging fractal features, there is still a gap in exploring the integration of fractal features with deep learning for segmentation tasks. The inherent self-similarity observed in tubular structures aligns well with the fractal theory. In this study, we address this gap by incorporating FFMs into the segmentation model, aiming to provide a new and reliable source of information to enhance segmentation performance." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 342, + 202, + 354 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 342, + 202, + 354 + ], + "spans": [ + { + "bbox": [ + 132, + 342, + 202, + 354 + ], + "type": "text", + "content": "3 Method" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 369, + 264, + 382 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 369, + 264, + 382 + ], + "spans": [ + { + "bbox": [ + 132, + 369, + 264, + 382 + ], + "type": "text", + "content": "3.1 Fractal Feature Map" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 391, + 480, + 452 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 391, + 480, + 452 + ], + "spans": [ + { + "bbox": [ + 130, + 391, + 480, + 452 + ], + "type": "text", + "content": "In fractal geometry, the fractal dimension (FD) provides a quantitative measure of an image's degree of self-similarity and roughness. FD can be estimated via the property of self-similarity [41]. Given an image " + }, + { + "bbox": [ + 130, + 391, + 480, + 452 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 130, + 391, + 480, + 452 + ], + "type": "text", + "content": ", it is self-similar if " + }, + { + "bbox": [ + 130, + 391, + 480, + 452 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 130, + 391, + 480, + 452 + ], + "type": "text", + "content": " comprises " + }, + { + "bbox": [ + 130, + 391, + 480, + 452 + ], + "type": "inline_equation", + "content": "N_r" + }, + { + "bbox": [ + 130, + 391, + 480, + 452 + ], + "type": "text", + "content": " distinct copies of itself scaled down by a factor of " + }, + { + "bbox": [ + 130, + 391, + 480, + 452 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 130, + 391, + 480, + 452 + ], + "type": "text", + "content": ". Consequently, for an image, the FD is defined as:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 263, + 460, + 481, + 486 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 263, + 460, + 481, + 486 + ], + "spans": [ + { + "bbox": [ + 263, + 460, + 481, + 486 + ], + "type": "interline_equation", + "content": "F D = \\lim _ {r \\rightarrow 0} \\frac {\\log N _ {r}}{\\log (1 / r)} \\tag {1}", + "image_path": "95fc01480f88a9da3d2b92d8480e26266b5fe5575d2e284cdd48aa4e855bc1c0.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 495, + 479, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 495, + 479, + 544 + ], + "spans": [ + { + "bbox": [ + 130, + 495, + 479, + 544 + ], + "type": "text", + "content": "Although the definition of FD based on self-similarity is simple and concise, its direct estimation becomes impractical when dealing with irregular images. To overcome this challenge and estimate the FDs of images, the box-counting method [24, 30] is employed." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 544, + 481, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 544, + 481, + 640 + ], + "spans": [ + { + "bbox": [ + 130, + 544, + 481, + 640 + ], + "type": "text", + "content": "Box-counting Method: Consider a grayscale image " + }, + { + "bbox": [ + 130, + 544, + 481, + 640 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 130, + 544, + 481, + 640 + ], + "type": "text", + "content": " with dimensions " + }, + { + "bbox": [ + 130, + 544, + 481, + 640 + ], + "type": "inline_equation", + "content": "M \\times M" + }, + { + "bbox": [ + 130, + 544, + 481, + 640 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 544, + 481, + 640 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 130, + 544, + 481, + 640 + ], + "type": "text", + "content": " denotes the maximum gray level (typically " + }, + { + "bbox": [ + 130, + 544, + 481, + 640 + ], + "type": "inline_equation", + "content": "L = 255" + }, + { + "bbox": [ + 130, + 544, + 481, + 640 + ], + "type": "text", + "content": "). We can model the image " + }, + { + "bbox": [ + 130, + 544, + 481, + 640 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 130, + 544, + 481, + 640 + ], + "type": "text", + "content": " as a three-dimensional space with " + }, + { + "bbox": [ + 130, + 544, + 481, + 640 + ], + "type": "inline_equation", + "content": "(x, y)" + }, + { + "bbox": [ + 130, + 544, + 481, + 640 + ], + "type": "text", + "content": " indicating the two-dimensional position and a third coordinate " + }, + { + "bbox": [ + 130, + 544, + 481, + 640 + ], + "type": "inline_equation", + "content": "(z)" + }, + { + "bbox": [ + 130, + 544, + 481, + 640 + ], + "type": "text", + "content": " denoting the gray value. Then the 3-D space is subdivided into smaller cubic regions, or \"boxes\", each with dimensions " + }, + { + "bbox": [ + 130, + 544, + 481, + 640 + ], + "type": "inline_equation", + "content": "k \\times k \\times h" + }, + { + "bbox": [ + 130, + 544, + 481, + 640 + ], + "type": "text", + "content": ". Here, " + }, + { + "bbox": [ + 130, + 544, + 481, + 640 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 544, + 481, + 640 + ], + "type": "text", + "content": " is a given scale used to be a multiple of the sidelength of a pixel in " + }, + { + "bbox": [ + 130, + 544, + 481, + 640 + ], + "type": "inline_equation", + "content": "(x, y)" + }, + { + "bbox": [ + 130, + 544, + 481, + 640 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 544, + 481, + 640 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 130, + 544, + 481, + 640 + ], + "type": "text", + "content": " can be a multiple of the gray level unit in " + }, + { + "bbox": [ + 130, + 544, + 481, + 640 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 130, + 544, + 481, + 640 + ], + "type": "text", + "content": "-direction. Given " + }, + { + "bbox": [ + 130, + 544, + 481, + 640 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 130, + 544, + 481, + 640 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 544, + 481, + 640 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 544, + 481, + 640 + ], + "type": "text", + "content": ", the value of " + }, + { + "bbox": [ + 130, + 544, + 481, + 640 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 130, + 544, + 481, + 640 + ], + "type": "text", + "content": " calculated using the following formula:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 265, + 654, + 480, + 666 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 265, + 654, + 480, + 666 + ], + "spans": [ + { + "bbox": [ + 265, + 654, + 480, + 666 + ], + "type": "interline_equation", + "content": "h = (L - 1) \\times k / M \\tag {2}", + "image_path": "dfc5e1ffe7f24b8099e0e6869af67728bb63b60bbf171a73f068fe605e27b50d.jpg" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 197, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 197, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 197, + 91, + 448, + 102 + ], + "type": "text", + "content": "Fractal Feature Maps for Segmentation of Tubular Structures" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 165 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 165 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 165 + ], + "type": "text", + "content": "Given a " + }, + { + "bbox": [ + 130, + 116, + 482, + 165 + ], + "type": "inline_equation", + "content": "k \\times k" + }, + { + "bbox": [ + 130, + 116, + 482, + 165 + ], + "type": "text", + "content": " grid located at position " + }, + { + "bbox": [ + 130, + 116, + 482, + 165 + ], + "type": "inline_equation", + "content": "(i,j)" + }, + { + "bbox": [ + 130, + 116, + 482, + 165 + ], + "type": "text", + "content": ", suppose that the minimum gray value is contained within the " + }, + { + "bbox": [ + 130, + 116, + 482, + 165 + ], + "type": "inline_equation", + "content": "m_{th}" + }, + { + "bbox": [ + 130, + 116, + 482, + 165 + ], + "type": "text", + "content": " box " + }, + { + "bbox": [ + 130, + 116, + 482, + 165 + ], + "type": "inline_equation", + "content": "(m = \\lceil minG_{i,j} / h \\rceil)" + }, + { + "bbox": [ + 130, + 116, + 482, + 165 + ], + "type": "text", + "content": ", and the maximum gray value within the " + }, + { + "bbox": [ + 130, + 116, + 482, + 165 + ], + "type": "inline_equation", + "content": "l_{th}" + }, + { + "bbox": [ + 130, + 116, + 482, + 165 + ], + "type": "text", + "content": " box " + }, + { + "bbox": [ + 130, + 116, + 482, + 165 + ], + "type": "inline_equation", + "content": "(l = \\lceil maxG_{i,j} / h \\rceil)" + }, + { + "bbox": [ + 130, + 116, + 482, + 165 + ], + "type": "text", + "content": ". The minimum number of boxes required to encompass all gray levels within the grid at " + }, + { + "bbox": [ + 130, + 116, + 482, + 165 + ], + "type": "inline_equation", + "content": "(i,j)" + }, + { + "bbox": [ + 130, + 116, + 482, + 165 + ], + "type": "text", + "content": " is computed as:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 264, + 178, + 481, + 190 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 264, + 178, + 481, + 190 + ], + "spans": [ + { + "bbox": [ + 264, + 178, + 481, + 190 + ], + "type": "interline_equation", + "content": "n _ {r} (i, j) = l - m + 1 \\tag {3}", + "image_path": "f19b484ccd8f4dc7cff33776c0beaced729649af51f9603523e6c90f6dbf87bc.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 197, + 481, + 220 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 197, + 481, + 220 + ], + "spans": [ + { + "bbox": [ + 130, + 197, + 481, + 220 + ], + "type": "text", + "content": "Considering all grids, the number of boxes that can cover all the patches is expressed as:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 271, + 221, + 481, + 246 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 271, + 221, + 481, + 246 + ], + "spans": [ + { + "bbox": [ + 271, + 221, + 481, + 246 + ], + "type": "interline_equation", + "content": "N _ {r} = \\sum_ {i, j} n _ {r} (i, j) \\tag {4}", + "image_path": "b9d457438f3f2a1674a248d3bbff47450dcee1bb3db7794dfb0b609cea7a0dba.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 252, + 479, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 252, + 479, + 300 + ], + "spans": [ + { + "bbox": [ + 130, + 252, + 479, + 300 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 252, + 479, + 300 + ], + "type": "inline_equation", + "content": "r = k / M" + }, + { + "bbox": [ + 130, + 252, + 479, + 300 + ], + "type": "text", + "content": ". We can obtain a series of " + }, + { + "bbox": [ + 130, + 252, + 479, + 300 + ], + "type": "inline_equation", + "content": "N_{r}" + }, + { + "bbox": [ + 130, + 252, + 479, + 300 + ], + "type": "text", + "content": " using differing values of " + }, + { + "bbox": [ + 130, + 252, + 479, + 300 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 252, + 479, + 300 + ], + "type": "text", + "content": ". Finally, the fractal dimension can be estimated from the least-squares linear fitting of " + }, + { + "bbox": [ + 130, + 252, + 479, + 300 + ], + "type": "inline_equation", + "content": "\\log N_{r}" + }, + { + "bbox": [ + 130, + 252, + 479, + 300 + ], + "type": "text", + "content": " versus " + }, + { + "bbox": [ + 130, + 252, + 479, + 300 + ], + "type": "inline_equation", + "content": "\\log (1 / r)" + }, + { + "bbox": [ + 130, + 252, + 479, + 300 + ], + "type": "text", + "content": ", as illustrated in Equation (1). The flow of box-counting method is summarized in the supplementary material." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 300, + 482, + 432 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 300, + 482, + 432 + ], + "spans": [ + { + "bbox": [ + 130, + 300, + 482, + 432 + ], + "type": "text", + "content": "Fractal Feature Map: Although fractal dimension and fractal features have been previously applied to classification tasks [32, 43], the inherent differences between classification and segmentation tasks preclude the direct application of image-level fractal dimension to segmentation endeavors. We extend the calculation of FD from the image-level to the pixel-level to generate the FFM of an image. As depicted in Fig. 2, the process begins with the utilization of a sliding window technique. Within a " + }, + { + "bbox": [ + 130, + 300, + 482, + 432 + ], + "type": "inline_equation", + "content": "5 \\times 5" + }, + { + "bbox": [ + 130, + 300, + 482, + 432 + ], + "type": "text", + "content": " window, for example, the FD of this region is computed using the box-counting method. Subsequently, the window is shifted along both the horizontal and vertical directions with a step size of 1, resulting in the calculation of the FFM for the entire image. The algorithm for generating the FFM is summarized in Algorithm 1." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 433, + 289, + 600 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 433, + 289, + 600 + ], + "spans": [ + { + "bbox": [ + 130, + 433, + 289, + 600 + ], + "type": "text", + "content": "The FFMs of the images (denoted as " + }, + { + "bbox": [ + 130, + 433, + 289, + 600 + ], + "type": "inline_equation", + "content": "FFM_{image}" + }, + { + "bbox": [ + 130, + 433, + 289, + 600 + ], + "type": "text", + "content": ") are incorporated into the segmentation model as additional input channels (Image, " + }, + { + "bbox": [ + 130, + 433, + 289, + 600 + ], + "type": "inline_equation", + "content": "FFM_{image}" + }, + { + "bbox": [ + 130, + 433, + 289, + 600 + ], + "type": "text", + "content": "), enhancing the focus on fractal structure and self-similarity after the normalization process. To mitigate the impact of image noise on the calculation of FD, we adopt the box-counting method proposed in [29] as they substitute the gray value with the standard deviation to make the method more robust." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 295, + 436, + 450, + 448 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 436, + 450, + 448 + ], + "spans": [ + { + "bbox": [ + 295, + 436, + 450, + 448 + ], + "type": "text", + "content": "Algorithm 1 Generation of FFM." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 452, + 481, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 452, + 481, + 472 + ], + "spans": [ + { + "bbox": [ + 304, + 452, + 481, + 472 + ], + "type": "text", + "content": "Input: Image " + }, + { + "bbox": [ + 304, + 452, + 481, + 472 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 304, + 452, + 481, + 472 + ], + "type": "text", + "content": " with size " + }, + { + "bbox": [ + 304, + 452, + 481, + 472 + ], + "type": "inline_equation", + "content": "M \\times N" + }, + { + "bbox": [ + 304, + 452, + 481, + 472 + ], + "type": "text", + "content": ", window size " + }, + { + "bbox": [ + 304, + 452, + 481, + 472 + ], + "type": "inline_equation", + "content": "w \\times w" + }, + { + "bbox": [ + 304, + 452, + 481, + 472 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 473, + 481, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 473, + 481, + 483 + ], + "spans": [ + { + "bbox": [ + 304, + 473, + 481, + 483 + ], + "type": "text", + "content": "Output: Fractal feature map (FFM) of " + }, + { + "bbox": [ + 304, + 473, + 481, + 483 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 304, + 473, + 481, + 483 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 484, + 481, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 484, + 481, + 495 + ], + "spans": [ + { + "bbox": [ + 304, + 484, + 481, + 495 + ], + "type": "text", + "content": "Assign " + }, + { + "bbox": [ + 304, + 484, + 481, + 495 + ], + "type": "inline_equation", + "content": "FFM \\gets" + }, + { + "bbox": [ + 304, + 484, + 481, + 495 + ], + "type": "text", + "content": " Float array of size " + }, + { + "bbox": [ + 304, + 484, + 481, + 495 + ], + "type": "inline_equation", + "content": "(M, N)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 495, + 432, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 495, + 432, + 506 + ], + "spans": [ + { + "bbox": [ + 304, + 495, + 432, + 506 + ], + "type": "text", + "content": "Assign padding size " + }, + { + "bbox": [ + 304, + 495, + 432, + 506 + ], + "type": "inline_equation", + "content": "p \\gets \\lfloor w / 2 \\rfloor" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 506, + 424, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 506, + 424, + 517 + ], + "spans": [ + { + "bbox": [ + 304, + 506, + 424, + 517 + ], + "type": "inline_equation", + "content": "Pad_{I}\\gets" + }, + { + "bbox": [ + 304, + 506, + 424, + 517 + ], + "type": "text", + "content": " Linear Padding " + }, + { + "bbox": [ + 304, + 506, + 424, + 517 + ], + "type": "inline_equation", + "content": "(I,p)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 517, + 403, + 527 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 517, + 403, + 527 + ], + "spans": [ + { + "bbox": [ + 304, + 517, + 403, + 527 + ], + "type": "text", + "content": "For " + }, + { + "bbox": [ + 304, + 517, + 403, + 527 + ], + "type": "inline_equation", + "content": "i\\gets 0" + }, + { + "bbox": [ + 304, + 517, + 403, + 527 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 304, + 517, + 403, + 527 + ], + "type": "inline_equation", + "content": "M - 1" + }, + { + "bbox": [ + 304, + 517, + 403, + 527 + ], + "type": "text", + "content": " do" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 323, + 529, + 422, + 539 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 323, + 529, + 422, + 539 + ], + "spans": [ + { + "bbox": [ + 323, + 529, + 422, + 539 + ], + "type": "text", + "content": "For " + }, + { + "bbox": [ + 323, + 529, + 422, + 539 + ], + "type": "inline_equation", + "content": "j\\gets 0" + }, + { + "bbox": [ + 323, + 529, + 422, + 539 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 323, + 529, + 422, + 539 + ], + "type": "inline_equation", + "content": "N - 1" + }, + { + "bbox": [ + 323, + 529, + 422, + 539 + ], + "type": "text", + "content": " do" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 342, + 540, + 422, + 551 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 342, + 540, + 422, + 551 + ], + "spans": [ + { + "bbox": [ + 342, + 540, + 422, + 551 + ], + "type": "text", + "content": "Selected region " + }, + { + "bbox": [ + 342, + 540, + 422, + 551 + ], + "type": "inline_equation", + "content": "R_{i,j}" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 342, + 551, + 470, + 561 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 342, + 551, + 470, + 561 + ], + "spans": [ + { + "bbox": [ + 342, + 551, + 470, + 561 + ], + "type": "interline_equation", + "content": "R _ {i, j} \\leftarrow P a d _ {I} [ i: i + w, j: j + w ]", + "image_path": "3f3f9ed26ac373a3590104f63156ba4eaa4137404f01a36e06da03ff9d70f1e2.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 342, + 561, + 448, + 572 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 342, + 561, + 448, + 572 + ], + "spans": [ + { + "bbox": [ + 342, + 561, + 448, + 572 + ], + "type": "interline_equation", + "content": "F D \\leftarrow \\operatorname {B o x - c o u n t i n g} \\left(R _ {i, j}\\right)", + "image_path": "18c8b83abb64b9eb68eccc4ffe95717210fd48b3694d484959be4aebc9cdc8b7.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 342, + 572, + 413, + 582 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 342, + 572, + 413, + 582 + ], + "spans": [ + { + "bbox": [ + 342, + 572, + 413, + 582 + ], + "type": "interline_equation", + "content": "F F M [ i, j ] \\leftarrow F D", + "image_path": "03a4fc70c98d8d17cd2b9a20b6eb76a2ba1deee422c42cc16aa7ca4dea5792b9.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 583, + 433, + 594 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 583, + 433, + 594 + ], + "spans": [ + { + "bbox": [ + 304, + 583, + 433, + 594 + ], + "type": "interline_equation", + "content": "F F M \\leftarrow \\operatorname {N o r m a l i z a t i o n} (F F M)", + "image_path": "433b075af0b19935d7e37b64ec40a700a2f816eaf45b263b8bd718b6edf28d75.jpg" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 131, + 619, + 282, + 631 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 619, + 282, + 631 + ], + "spans": [ + { + "bbox": [ + 131, + 619, + 282, + 631 + ], + "type": "text", + "content": "3.2 Multi-Decoder Network" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "content": "In segmenting tubular structures, accurate boundary detection and preservation of global topological connectivity are crucial requirements. To meet these re" + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 228, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 228, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 228, + 102 + ], + "type": "text", + "content": "J. Huang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 186, + 116, + 430, + 196 + ], + "blocks": [ + { + "bbox": [ + 186, + 116, + 430, + 196 + ], + "lines": [ + { + "bbox": [ + 186, + 116, + 430, + 196 + ], + "spans": [ + { + "bbox": [ + 186, + 116, + 430, + 196 + ], + "type": "image", + "image_path": "d3dfcc36229c8ebbc942c24040f3ee904a819d0e6fa09cd93cb59b54bcabaa61.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 203, + 204, + 408, + 215 + ], + "lines": [ + { + "bbox": [ + 203, + 204, + 408, + 215 + ], + "spans": [ + { + "bbox": [ + 203, + 204, + 408, + 215 + ], + "type": "text", + "content": "Fig. 2: Workflow of computing FFM of an image." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 217, + 479, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 217, + 479, + 240 + ], + "spans": [ + { + "bbox": [ + 130, + 217, + 479, + 240 + ], + "type": "text", + "content": "requirements, we propose a new model that we refer to as Multi-Decoder Network (MD-Net)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 240, + 480, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 240, + 480, + 407 + ], + "spans": [ + { + "bbox": [ + 130, + 240, + 480, + 407 + ], + "type": "text", + "content": "In addition to predicting the segmentation mask, we introduce an edge decoder and a skeleton decoder to generate the edge and skeleton of the tubular structure, respectively, as depicted in Fig. 3. When (Image, " + }, + { + "bbox": [ + 130, + 240, + 480, + 407 + ], + "type": "inline_equation", + "content": "FFM_{image}" + }, + { + "bbox": [ + 130, + 240, + 480, + 407 + ], + "type": "text", + "content": ") are fed into MD-Net, the Encoder employs convolution layers to extract a series of low-level to high-level image features. These features are simultaneously transmitted to the three Decoders by skip connections for prediction. Specifically, in the Encoder, each convolution layer involves the repeated application of two 3x3 convolutions, followed by a rectified linear unit (ReLU), and a 2x2 max pooling operation with a stride of 2 for down-sampling. As for the Decoders in MD-Net, each convolution layer includes an up-sampling of the feature map, followed by a 2x2 convolution that reduces the number of feature channels by half. Subsequently, the feature is concatenated with the corresponding feature copied from the Encoder. This concatenated feature is then processed using two 3x3 convolutions, each followed by a ReLU activation function." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 408, + 480, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 408, + 480, + 491 + ], + "spans": [ + { + "bbox": [ + 130, + 408, + 480, + 491 + ], + "type": "text", + "content": "During training, the ground truths for edges and skeletons are derived from the annotated masks by employing the findContours function from the OpenCV library and the skeletonization algorithms in the scikit-image library. We have visualized the boundaries and skeletons in the supplementary material to corroborate the accuracy of the extraction techniques. During the inference stage, the final structural segmentation is the output of the object decoder. Edge and skeleton predictions are only used in model training." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 491, + 480, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 491, + 480, + 564 + ], + "spans": [ + { + "bbox": [ + 130, + 491, + 480, + 564 + ], + "type": "text", + "content": "Loss Function: Given an image " + }, + { + "bbox": [ + 130, + 491, + 480, + 564 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 130, + 491, + 480, + 564 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 130, + 491, + 480, + 564 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 491, + 480, + 564 + ], + "type": "text", + "content": " pixels, the segmentation ground truth is denoted as " + }, + { + "bbox": [ + 130, + 491, + 480, + 564 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 130, + 491, + 480, + 564 + ], + "type": "text", + "content": ", where foreground and background pixels are labeled as 1 and 0, respectively. The prediction of the model is represented as " + }, + { + "bbox": [ + 130, + 491, + 480, + 564 + ], + "type": "inline_equation", + "content": "\\hat{y} \\in [0,1]" + }, + { + "bbox": [ + 130, + 491, + 480, + 564 + ], + "type": "text", + "content": ", indicating the probabilities of individual pixels being classified as foreground. To compute the loss, we utilize the differentiable Soft IOU Loss [19], denoted as " + }, + { + "bbox": [ + 130, + 491, + 480, + 564 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{object}" + }, + { + "bbox": [ + 130, + 491, + 480, + 564 + ], + "type": "text", + "content": " in Equation (5), where " + }, + { + "bbox": [ + 130, + 491, + 480, + 564 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 130, + 491, + 480, + 564 + ], + "type": "text", + "content": " is a smoothing coefficient." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 208, + 570, + 480, + 601 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 208, + 570, + 480, + 601 + ], + "spans": [ + { + "bbox": [ + 208, + 570, + 480, + 601 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {o b j e c t}} = \\mathcal {L} _ {i o u} = 1 - \\frac {\\sum_ {i = 1} ^ {N} y _ {i} \\hat {y} _ {i}}{\\sum_ {i = 1} ^ {N} y _ {i} + \\hat {y} _ {i} - y _ {i} \\hat {y} _ {i} + \\eta} \\tag {5}", + "image_path": "dc911c4148e7222f725d11487da4b7fbffa8b24db033b749c041955b39dd9a4f.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 605, + 479, + 630 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 479, + 630 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 479, + 630 + ], + "type": "text", + "content": "The BCE loss function is applied separately to the edge and skeleton segmentation tasks, denoted as " + }, + { + "bbox": [ + 130, + 605, + 479, + 630 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{edge},\\mathcal{L}_{skeleton}" + }, + { + "bbox": [ + 130, + 605, + 479, + 630 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 157, + 635, + 480, + 667 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 157, + 635, + 480, + 667 + ], + "spans": [ + { + "bbox": [ + 157, + 635, + 480, + 667 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {e d g e}}, \\mathcal {L} _ {\\text {s k e l e t o n}} = \\mathcal {L} _ {\\text {b c e}} = - \\frac {1}{N} \\sum_ {i = 1} ^ {N} y _ {i} \\cdot \\log (\\hat {y} _ {i}) + (1 - y _ {i}) \\cdot \\log (1 - \\hat {y} _ {i}) \\tag {6}", + "image_path": "35f52307093994228df6d40465aa3d0278f6add469dde8b37279c7df855c61df.jpg" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 197, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 197, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 197, + 91, + 448, + 102 + ], + "type": "text", + "content": "Fractal Feature Maps for Segmentation of Tubular Structures" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 152 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 152 + ], + "type": "text", + "content": "For MD-Net, which has three segmentation tasks, the loss function is composed as in Equation (8). The weights " + }, + { + "bbox": [ + 130, + 116, + 482, + 152 + ], + "type": "inline_equation", + "content": "\\alpha, \\beta" + }, + { + "bbox": [ + 130, + 116, + 482, + 152 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 116, + 482, + 152 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 130, + 116, + 482, + 152 + ], + "type": "text", + "content": " are typically set to 1.0, 0.5 and 0.5, respectively." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 217, + 152, + 481, + 166 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 217, + 152, + 481, + 166 + ], + "spans": [ + { + "bbox": [ + 217, + 152, + 481, + 166 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {g l o b a l}} = \\alpha \\mathcal {L} _ {\\text {o b j e c t}} + \\beta \\mathcal {L} _ {\\text {e d g e}} + \\gamma \\mathcal {L} _ {\\text {s k e l e t o n}} \\tag {7}", + "image_path": "c196ca7e433ec3ad001b9615a45f5fbeb4811d6f8fdfc4d2cf4117f121dd2538.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 181, + 329, + 192 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 181, + 329, + 192 + ], + "spans": [ + { + "bbox": [ + 132, + 181, + 329, + 192 + ], + "type": "text", + "content": "3.3 Fractal Feature Constrained Loss" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 200, + 482, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 200, + 482, + 319 + ], + "spans": [ + { + "bbox": [ + 130, + 200, + 482, + 319 + ], + "type": "text", + "content": "The value of FD serves as an indicator of the texture complexity present in a given image region. Hence, it is advantageous to incorporate pixel-level FFM as weights of the loss function. This allows for a higher weight to be assigned to regions that exhibit greater complexity, as accurately segmenting such regions poses a greater challenge. Consequently, we calculate the FFM for each label associated with the image, denoted as " + }, + { + "bbox": [ + 130, + 200, + 482, + 319 + ], + "type": "inline_equation", + "content": "FFM_{label}" + }, + { + "bbox": [ + 130, + 200, + 482, + 319 + ], + "type": "text", + "content": ", and employ " + }, + { + "bbox": [ + 130, + 200, + 482, + 319 + ], + "type": "inline_equation", + "content": "FFM_{label}" + }, + { + "bbox": [ + 130, + 200, + 482, + 319 + ], + "type": "text", + "content": " as pixel-level weights within the loss function to guide the convergence of the model. Considering the specificity of the edge decoder and skeleton decoder, we have thus restricted the application of " + }, + { + "bbox": [ + 130, + 200, + 482, + 319 + ], + "type": "inline_equation", + "content": "FFM_{label}" + }, + { + "bbox": [ + 130, + 200, + 482, + 319 + ], + "type": "text", + "content": " exclusively to the object decoder. The Constrained Loss of MD-Net is as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 181, + 329, + 481, + 342 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 181, + 329, + 481, + 342 + ], + "spans": [ + { + "bbox": [ + 181, + 329, + 481, + 342 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {c o n s t r a i n e d}} = \\alpha \\mathcal {L} _ {\\text {o b j e c t}} \\cdot F F M _ {\\text {l a b e l}} + \\beta \\mathcal {L} _ {\\text {e d g e}} + \\gamma \\mathcal {L} _ {\\text {s k e l e t o n}} \\tag {8}", + "image_path": "05d1f57ef1ba82ed0e65fb79f37ab8788e0bd48eaa9e5d38420b2752b648f590.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 358, + 230, + 372 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 358, + 230, + 372 + ], + "spans": [ + { + "bbox": [ + 132, + 358, + 230, + 372 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 383, + 205, + 394 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 383, + 205, + 394 + ], + "spans": [ + { + "bbox": [ + 132, + 383, + 205, + 394 + ], + "type": "text", + "content": "4.1 Datasets" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 402, + 482, + 452 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 402, + 482, + 452 + ], + "spans": [ + { + "bbox": [ + 130, + 402, + 482, + 452 + ], + "type": "text", + "content": "We evaluate the FFM and MD-Net using five publicly available tubular datasets including the endoplasmic reticulum (ER) and mitochondrial (MITO) [11, 36] organelle network datasets, the Retinal OCT-Angiography vessel Segmentation (ROSE) and Structured Analysis of the Retina (STARE) [15, 37] retinal vessel" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 138, + 472, + 479, + 589 + ], + "blocks": [ + { + "bbox": [ + 138, + 472, + 479, + 589 + ], + "lines": [ + { + "bbox": [ + 138, + 472, + 479, + 589 + ], + "spans": [ + { + "bbox": [ + 138, + 472, + 479, + 589 + ], + "type": "image", + "image_path": "61582cd15f228a0900dcfb446a58ef3b1def5fb8657444d24a1edcf323d681ff.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 599, + 482, + 677 + ], + "lines": [ + { + "bbox": [ + 130, + 599, + 482, + 677 + ], + "spans": [ + { + "bbox": [ + 130, + 599, + 482, + 677 + ], + "type": "text", + "content": "Fig. 3: Overview and details of our proposed model MD-Net. The left part illustrates the overall structure of MD-Net. (Image, " + }, + { + "bbox": [ + 130, + 599, + 482, + 677 + ], + "type": "inline_equation", + "content": "FFM_{image}" + }, + { + "bbox": [ + 130, + 599, + 482, + 677 + ], + "type": "text", + "content": ") is processed as input by the Encoder, which extracts features of varying sizes from the input. These features are transmitted simultaneously to three Decoders through Skip Connection. The right part is a detailed description of the Encoder and Decoders. Each Decoder performs upsampling and concatenation operations similar to U-Net to obtain the predictions, comprising the edge, skeleton, and object." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 228, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 228, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 228, + 102 + ], + "type": "text", + "content": "J. Huang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 175 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 175 + ], + "type": "text", + "content": "datasets, and the Massachusetts Roads (ROAD) [38] remote sensing dataset. By selecting these datasets, our aim is to comprehensively evaluate the performance of our methods across different types of structures and domains, providing a robust assessment of their capabilities. To further examine the influence of FFM, we included the non-tubular dataset NUCLEUS [5]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 176, + 480, + 282 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 176, + 480, + 282 + ], + "spans": [ + { + "bbox": [ + 130, + 176, + 480, + 282 + ], + "type": "text", + "content": "All images in ER, MITO, STARE, and NUCLEUS are randomly sampled patches with size of " + }, + { + "bbox": [ + 130, + 176, + 480, + 282 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 130, + 176, + 480, + 282 + ], + "type": "text", + "content": " for training and sliding window with size of " + }, + { + "bbox": [ + 130, + 176, + 480, + 282 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 130, + 176, + 480, + 282 + ], + "type": "text", + "content": " for validation and testing. A crop size of " + }, + { + "bbox": [ + 130, + 176, + 480, + 282 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 130, + 176, + 480, + 282 + ], + "type": "text", + "content": " is used randomly during training and regularly during testing for the ROAD dataset as in [42]. Horizontal and vertical flipping as well as " + }, + { + "bbox": [ + 130, + 176, + 480, + 282 + ], + "type": "inline_equation", + "content": "90^{\\circ} / 180^{\\circ} / 270^{\\circ}" + }, + { + "bbox": [ + 130, + 176, + 480, + 282 + ], + "type": "text", + "content": " rotation are used for data augmentation. In addition, data augmentation for ROSE is conducted by rotation of an angle from " + }, + { + "bbox": [ + 130, + 176, + 480, + 282 + ], + "type": "inline_equation", + "content": "-10^{\\circ}" + }, + { + "bbox": [ + 130, + 176, + 480, + 282 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 130, + 176, + 480, + 282 + ], + "type": "inline_equation", + "content": "10^{\\circ}" + }, + { + "bbox": [ + 130, + 176, + 480, + 282 + ], + "type": "text", + "content": " as described in [37]. Detailed partitions of the six datasets and experimental setup are shown in the supplementary material." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 300, + 288, + 312 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 300, + 288, + 312 + ], + "spans": [ + { + "bbox": [ + 132, + 300, + 288, + 312 + ], + "type": "text", + "content": "4.2 Performance Comparison" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 319, + 480, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 319, + 480, + 437 + ], + "spans": [ + { + "bbox": [ + 130, + 319, + 480, + 437 + ], + "type": "text", + "content": "We select several segmentation networks for comparison, including the vanilla U-Net [44], U-Net++ [60], nnU-Net [20], and HR-Net [50]. Additionally, we compare our methods with SOTA models for diverse datasets. For the task of retinal vessel segmentation, we choose three-stage [53], OCTA-Net [37], GT-DLA [31], and AF-Net [45], as these networks are tailored for this task. Furthermore, we compare the performance of MD-Net in the ROAD dataset with DCU-Net [54], Dconn-Net [55], and the DSC-Net [42], a tubular structure segmentation network with specially designed loss " + }, + { + "bbox": [ + 130, + 319, + 480, + 437 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{tc}" + }, + { + "bbox": [ + 130, + 319, + 480, + 437 + ], + "type": "text", + "content": ". To assess the generalization of FFM, we apply " + }, + { + "bbox": [ + 130, + 319, + 480, + 437 + ], + "type": "inline_equation", + "content": "FFM_{image}" + }, + { + "bbox": [ + 130, + 319, + 480, + 437 + ], + "type": "text", + "content": " to both U-Net and HR-Net. All models are trained on the same dataset with the same hyperparameter settings." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 455, + 256, + 466 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 455, + 256, + 466 + ], + "spans": [ + { + "bbox": [ + 132, + 455, + 256, + 466 + ], + "type": "text", + "content": "4.3 Evaluation Metrics" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 474, + 479, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 474, + 479, + 496 + ], + "spans": [ + { + "bbox": [ + 130, + 474, + 479, + 496 + ], + "type": "text", + "content": "The models are assessed using three types of metrics: volumetric, topology, and distance." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 498, + 483, + 602 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 130, + 498, + 483, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 498, + 483, + 533 + ], + "spans": [ + { + "bbox": [ + 130, + 498, + 483, + 533 + ], + "type": "text", + "content": "1. Volumetric scores, including intersection-over-union (IoU), accuracy (ACC), centerlineDice (clDice) [46], and AUC. These metrics quantify the overlap and agreement between the prediction and ground truth." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 533, + 479, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 479, + 567 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 479, + 567 + ], + "type": "text", + "content": "2. Topology scores. To evaluate the topological correctness of the segmentations, we calculate the Betti Error [17] " + }, + { + "bbox": [ + 130, + 533, + 479, + 567 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 130, + 533, + 479, + 567 + ], + "type": "text", + "content": " for the sum of Betti Numbers " + }, + { + "bbox": [ + 130, + 533, + 479, + 567 + ], + "type": "inline_equation", + "content": "\\beta_0" + }, + { + "bbox": [ + 130, + 533, + 479, + 567 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 533, + 479, + 567 + ], + "type": "inline_equation", + "content": "\\beta_{1}" + }, + { + "bbox": [ + 130, + 533, + 479, + 567 + ], + "type": "text", + "content": ". In the evaluation of ROSE and STARE, " + }, + { + "bbox": [ + 130, + 533, + 479, + 567 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 130, + 533, + 479, + 567 + ], + "type": "text", + "content": " Error represents the " + }, + { + "bbox": [ + 130, + 533, + 479, + 567 + ], + "type": "inline_equation", + "content": "\\beta_{1}" + }, + { + "bbox": [ + 130, + 533, + 479, + 567 + ], + "type": "text", + "content": " only." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 570, + 479, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 479, + 602 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 479, + 602 + ], + "type": "text", + "content": "3. Distance score: Hausdorff Distance (HD) [48] quantifies the accuracy of the boundary delineation and provides information about the spatial closeness of the predicted and ground truth boundaries." + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 130, + 605, + 479, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 479, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 479, + 665 + ], + "type": "text", + "content": "Since the nucleus is oval in shape and not tubular, no Betti Error and clDice are used in segmentation performance evaluation and no skeletal decoder is used in the MD-Net. In this study, the evaluation metrics IoU, ACC, AUC, and clDice, are expressed as percentages (\\%). The Hausdorff Distance (HD) is measured in pixels (px). All performance metrics are calculated for each image and averaged." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 197, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 197, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 197, + 91, + 447, + 102 + ], + "type": "text", + "content": "Fractal Feature Maps for Segmentation of Tubular Structures" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 134, + 179, + 481, + 386 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 482, + 170 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 482, + 170 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 482, + 170 + ], + "type": "text", + "content": "Table 1: Bold numbers indicate the best performance and italicized numbers represent the next best performance. The result marked with a star (*) represents the model input replaced by (Image) with (Image, " + }, + { + "bbox": [ + 132, + 114, + 482, + 170 + ], + "type": "inline_equation", + "content": "FFM_{image}" + }, + { + "bbox": [ + 132, + 114, + 482, + 170 + ], + "type": "text", + "content": "). " + }, + { + "bbox": [ + 132, + 114, + 482, + 170 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{tc}" + }, + { + "bbox": [ + 132, + 114, + 482, + 170 + ], + "type": "text", + "content": " in (b) is Wasserstein-distance-based TCLoss proposed in [42]. " + }, + { + "bbox": [ + 132, + 114, + 482, + 170 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\text{constrained}}" + }, + { + "bbox": [ + 132, + 114, + 482, + 170 + ], + "type": "text", + "content": " represents the integration of " + }, + { + "bbox": [ + 132, + 114, + 482, + 170 + ], + "type": "inline_equation", + "content": "FFM_{label}" + }, + { + "bbox": [ + 132, + 114, + 482, + 170 + ], + "type": "text", + "content": " as a loss function weight, which leads to further improvement." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 134, + 179, + 481, + 386 + ], + "lines": [ + { + "bbox": [ + 134, + 179, + 481, + 386 + ], + "spans": [ + { + "bbox": [ + 134, + 179, + 481, + 386 + ], + "type": "table", + "html": "
ModelLossIoU↑clDice↑ACC↑AUC↑βError↓HD↓IoU↑clDice↑ACC↑AUC↑βError↓HD↓
ERMITO
U-Net++ [60]Liou75.0294.6791.0497.3126.027.0479.7097.3098.1999.592.304.24
nnU-Net [20]Liou73.4194.5191.2987.9525.156.9679.2497.1698.0794.623.404.39
DSC-Net [42]Liou75.5194.4491.5697.0934.516.9280.3297.1698.1699.652.704.36
Dconn-Net [55]Liou75.9595.2491.5996.6020.316.9779.8297.2098.0399.542.904.32
AF-Net [45]Liou76.0194.5891.8197.3729.926.7980.3696.9498.0899.642.604.18
GT-DLA [31]Liou75.8994.9291.8297.3120.266.8480.2597.1698.1999.602.704.21
U-Net [44]Liou75.4494.6391.8297.3528.726.8779.7796.9198.0799.612.804.56
HR-Net [50]Liou75.8395.0891.7197.3622.576.9079.6397.2798.1499.622.904.25
U-Net*Liou76.5995.4392.0297.5620.786.8180.7197.4298.2199.632.704.27
HR-Net*Liou76.4395.4791.9597.5520.526.8380.6297.2998.1799.623.304.18
MD-Net*Lglobal77.0195.7892.0697.5919.106.7781.1197.7298.2599.662.204.16
MD-Net*Lconstrained77.0995.7492.1497.6519.526.7281.1897.6198.2699.652.804.14
ROSESTARE
three-stage [53]Liou62.1167.0791.7993.419.137.2866.8376.7194.8696.433.456.57
OCTA-Net [37]Liou63.2766.3292.3494.5310.227.1165.3776.7494.6295.443.226.47
DSC-Net [42]Liou63.1567.5991.9894.437.007.3667.0176.2694.9796.503.156.31
Dconn-Net [55]Liou60.3764.5188.5793.265.738.1866.1476.3394.7896.653.776.48
AF-Net [45]Liou62.3367.0591.7392.558.887.3267.2777.1794.9196.473.186.49
GT-DLA [31]Liou63.5766.3192.2794.189.897.1667.1276.3294.7295.793.126.63
U-Net [44]Liou61.5267.5391.3394.048.227.4266.1576.0594.6896.393.226.67
HR-Net [50]Liou63.0967.7192.0894.388.337.2667.0376.9294.9296.292.736.53
U-Net*Liou64.0767.9592.2594.477.887.1068.0777.3995.1596.632.776.33
HR-Net*Liou64.6869.4292.1494.776.017.2168.2277.4995.1396.523.216.32
MD-Net*Lglobal65.0769.7892.3694.884.227.1068.4977.7995.2096.912.576.29
MD-Net*Lconstrained65.1969.5892.4294.884.897.0668.7378.2095.2696.992.606.34
", + "image_path": "47ba4810f869f35cebaf1d614893c6789fd89c6459f4ecc1821904b10c893231.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 135, + 406, + 478, + 498 + ], + "blocks": [ + { + "bbox": [ + 155, + 387, + 457, + 396 + ], + "lines": [ + { + "bbox": [ + 155, + 387, + 457, + 396 + ], + "spans": [ + { + "bbox": [ + 155, + 387, + 457, + 396 + ], + "type": "text", + "content": "(a) Quantitative experimental results for the ER, MITO, ROSE, and STARE dataset." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 135, + 406, + 478, + 498 + ], + "lines": [ + { + "bbox": [ + 135, + 406, + 478, + 498 + ], + "spans": [ + { + "bbox": [ + 135, + 406, + 478, + 498 + ], + "type": "table", + "html": "
ModelLossIoU↑cDice↑ACC↑AUC↑βError↓HD↓ModelLossIoU↑ACC↑AUC↑HD↓
DCU-Net [54]Lce62.9286.9898.0398.342.568.04CS-Net [39]Liou77.9698.1298.584.13
DSC-Net [42]Lce63.9987.7498.0598.392.567.96Transnet [8]Liou78.2198.3498.644.32
DSC-Net [42]Ltc64.2287.6498.0598.462.457.34DU-Net [21]Liou78.0998.4698.944.19
Dconn-Net [55]Ltc65.1487.6197.3098.292.256.97U-Net++[60]Liou78.3398.5999.694.09
AF-Net [45]Ltc65.0387.2797.4198.382.217.29nnU-Net [20]Liou78.2898.5794.934.09
GT-DLA [31]Ltc64.7686.8097.4397.942.386.94OCTA-Net [37]Liou75.3098.3499.264.43
U-Net [44]Lce62.4786.8797.9798.292.618.11U-Net [44]Liou78.0298.5499.654.23
HR-Net [50]Lce62.8986.7498.0698.392.588.06HR-Net [50]Liou77.7098.5099.654.15
U-Net*Lce65.7487.7498.3898.722.486.98U-Net*Liou78.4698.5899.684.15
HR-Net*Lce65.6087.6198.3298.792.316.98HR-Net*Liou77.8398.5399.664.15
MD-Net*Lglobal66.0788.0898.4398.802.196.92MD-Net*Lglobal78.6698.5999.684.08
MD-Net*Lconstrained66.1688.0798.4298.772.066.97MD-Net*Lconstrained78.7798.5999.664.06
", + "image_path": "eea3febbcb64f67f2d5bf3fa9e71808749737b63d2ea7bbf84c60110bb96dd1f.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 176, + 499, + 286, + 508 + ], + "lines": [ + { + "bbox": [ + 176, + 499, + 286, + 508 + ], + "spans": [ + { + "bbox": [ + 176, + 499, + 286, + 508 + ], + "type": "text", + "content": "(b) Results for ROAD dataset." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 342, + 499, + 465, + 509 + ], + "lines": [ + { + "bbox": [ + 342, + 499, + 465, + 509 + ], + "spans": [ + { + "bbox": [ + 342, + 499, + 465, + 509 + ], + "type": "text", + "content": "(c) Results for NUCLEUS dataset." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 132, + 528, + 235, + 540 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 528, + 235, + 540 + ], + "spans": [ + { + "bbox": [ + 132, + 528, + 235, + 540 + ], + "type": "text", + "content": "4.4 Configurations" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "type": "text", + "content": "All models are implemented using PyTorch (version 1.12.0) and executed on 4 NVIDIA 3090 GPU cards. During the training phase, we employed the SGD optimizer with an initial learning rate of 0.05. To dynamically adjust the learning rate as the training progressed, we incorporated warm-up and exponential decay techniques. For all datasets, a fixed batch size of 32 is employed, ensuring consistency in the training process. Additionally, to prevent overfitting, a regularization weight of 0.0005 is applied. The entire training process spanned 50 epochs, providing sufficient iterations for model convergence and learning." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 228, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 228, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 228, + 102 + ], + "type": "text", + "content": "J. Huang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 282, + 128 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 282, + 128 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 282, + 128 + ], + "type": "text", + "content": "4.5 Quantitative Evaluation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 137, + 463, + 148 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 137, + 463, + 148 + ], + "spans": [ + { + "bbox": [ + 130, + 137, + 463, + 148 + ], + "type": "text", + "content": "Based on the results presented in Tab. 1, several conclusions can be drawn." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 149, + 481, + 221 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 149, + 481, + 221 + ], + "spans": [ + { + "bbox": [ + 130, + 149, + 481, + 221 + ], + "type": "text", + "content": "Performance of MD-Net: Our proposed model, MD-Net, demonstrates superior performance in terms of segmentation accuracy and topological continuity compared to the other models. This superiority is observed across five tubular datasets as well as one non-tubular dataset. We have also performed t-test to check whether improvements of our methods over competing methods in performance are statistically significant. Results are provided in supplementary." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 221, + 485, + 341 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 221, + 485, + 341 + ], + "spans": [ + { + "bbox": [ + 130, + 221, + 485, + 341 + ], + "type": "text", + "content": "In tubular datasets, MD-Net outperforms other methods in terms of segmentation results. Specifically, in the ER, MITO, ROSE, STARE, and ROAD datasets, MD-Net achieves improvements in IoU of " + }, + { + "bbox": [ + 130, + 221, + 485, + 341 + ], + "type": "inline_equation", + "content": "1.08\\%" + }, + { + "bbox": [ + 130, + 221, + 485, + 341 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 221, + 485, + 341 + ], + "type": "inline_equation", + "content": "0.82\\%" + }, + { + "bbox": [ + 130, + 221, + 485, + 341 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 221, + 485, + 341 + ], + "type": "inline_equation", + "content": "1.62\\%" + }, + { + "bbox": [ + 130, + 221, + 485, + 341 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 221, + 485, + 341 + ], + "type": "inline_equation", + "content": "1.46\\%" + }, + { + "bbox": [ + 130, + 221, + 485, + 341 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 221, + 485, + 341 + ], + "type": "inline_equation", + "content": "1.02\\%" + }, + { + "bbox": [ + 130, + 221, + 485, + 341 + ], + "type": "text", + "content": ", respectively, compared to existing SOTA methods. Furthermore, when considering topological continuity and boundary extraction, MD-Net demonstrates the best performance with the minimum " + }, + { + "bbox": [ + 130, + 221, + 485, + 341 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 130, + 221, + 485, + 341 + ], + "type": "text", + "content": " error and Hausdorff Distance. These results highlight the ability of MD-Net, equipped with an edge decoder and a skeleton decoder, to effectively capture edge and skeleton features of thin tubular structures, resulting in more accurate and continuous topology segmentation outcomes." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 341, + 481, + 388 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 341, + 481, + 388 + ], + "spans": [ + { + "bbox": [ + 130, + 341, + 481, + 388 + ], + "type": "text", + "content": "Even in the non-tubular dataset NUCLEUS, MD-Net achieves the best segmentation results with an IoU of " + }, + { + "bbox": [ + 130, + 341, + 481, + 388 + ], + "type": "inline_equation", + "content": "78.77\\%" + }, + { + "bbox": [ + 130, + 341, + 481, + 388 + ], + "type": "text", + "content": ", ACC of " + }, + { + "bbox": [ + 130, + 341, + 481, + 388 + ], + "type": "inline_equation", + "content": "98.59\\%" + }, + { + "bbox": [ + 130, + 341, + 481, + 388 + ], + "type": "text", + "content": ", and HD of 4.06. This demonstrates the effectiveness of MD-Net and FFM in datasets with simpler structures." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 389, + 481, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 389, + 481, + 437 + ], + "spans": [ + { + "bbox": [ + 130, + 389, + 481, + 437 + ], + "type": "text", + "content": "Generalization of FFM: " + }, + { + "bbox": [ + 130, + 389, + 481, + 437 + ], + "type": "inline_equation", + "content": "FFM_{image}" + }, + { + "bbox": [ + 130, + 389, + 481, + 437 + ], + "type": "text", + "content": " is incorporated into U-Net and HR-Net as an additional input channel without modifying the training parameters. The results in Tab. 1 indicate that the inclusion of " + }, + { + "bbox": [ + 130, + 389, + 481, + 437 + ], + "type": "inline_equation", + "content": "FFM_{image}" + }, + { + "bbox": [ + 130, + 389, + 481, + 437 + ], + "type": "text", + "content": " leads to improved segmentation performance for both U-Net and HR-Net." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 437, + 483, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 437, + 483, + 521 + ], + "spans": [ + { + "bbox": [ + 130, + 437, + 483, + 521 + ], + "type": "text", + "content": "For U-Net, the segmentation performance (IoU) shows improvements of " + }, + { + "bbox": [ + 130, + 437, + 483, + 521 + ], + "type": "inline_equation", + "content": "1.14\\%" + }, + { + "bbox": [ + 130, + 437, + 483, + 521 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 437, + 483, + 521 + ], + "type": "inline_equation", + "content": "0.94\\%" + }, + { + "bbox": [ + 130, + 437, + 483, + 521 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 437, + 483, + 521 + ], + "type": "inline_equation", + "content": "2.55\\%" + }, + { + "bbox": [ + 130, + 437, + 483, + 521 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 437, + 483, + 521 + ], + "type": "inline_equation", + "content": "1.92\\%" + }, + { + "bbox": [ + 130, + 437, + 483, + 521 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 437, + 483, + 521 + ], + "type": "inline_equation", + "content": "3.27\\%" + }, + { + "bbox": [ + 130, + 437, + 483, + 521 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 437, + 483, + 521 + ], + "type": "inline_equation", + "content": "0.44\\%" + }, + { + "bbox": [ + 130, + 437, + 483, + 521 + ], + "type": "text", + "content": " on the ER, MITO, ROSE, STARE, ROAD, and NUCLEUS datasets with the assistance of " + }, + { + "bbox": [ + 130, + 437, + 483, + 521 + ], + "type": "inline_equation", + "content": "FFM_{image}" + }, + { + "bbox": [ + 130, + 437, + 483, + 521 + ], + "type": "text", + "content": ", respectively. Similarly, for HR-Net, the IoU improves by " + }, + { + "bbox": [ + 130, + 437, + 483, + 521 + ], + "type": "inline_equation", + "content": "0.60\\%" + }, + { + "bbox": [ + 130, + 437, + 483, + 521 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 437, + 483, + 521 + ], + "type": "inline_equation", + "content": "0.99\\%" + }, + { + "bbox": [ + 130, + 437, + 483, + 521 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 437, + 483, + 521 + ], + "type": "inline_equation", + "content": "1.59\\%" + }, + { + "bbox": [ + 130, + 437, + 483, + 521 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 437, + 483, + 521 + ], + "type": "inline_equation", + "content": "1.19\\%" + }, + { + "bbox": [ + 130, + 437, + 483, + 521 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 437, + 483, + 521 + ], + "type": "inline_equation", + "content": "2.71\\%" + }, + { + "bbox": [ + 130, + 437, + 483, + 521 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 437, + 483, + 521 + ], + "type": "inline_equation", + "content": "0.13\\%" + }, + { + "bbox": [ + 130, + 437, + 483, + 521 + ], + "type": "text", + "content": " on the same datasets, respectively. On average, " + }, + { + "bbox": [ + 130, + 437, + 483, + 521 + ], + "type": "inline_equation", + "content": "FFM_{image}" + }, + { + "bbox": [ + 130, + 437, + 483, + 521 + ], + "type": "text", + "content": " brings an improvement of " + }, + { + "bbox": [ + 130, + 437, + 483, + 521 + ], + "type": "inline_equation", + "content": "1.96\\%" + }, + { + "bbox": [ + 130, + 437, + 483, + 521 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 437, + 483, + 521 + ], + "type": "inline_equation", + "content": "1.42\\%" + }, + { + "bbox": [ + 130, + 437, + 483, + 521 + ], + "type": "text", + "content": " in IoU for U-Net and HR-Net, respectively, in the segmentation task of tubular datasets." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 521, + 481, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 481, + 582 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 481, + 582 + ], + "type": "text", + "content": "In conclusion, FFM demonstrates notable generalization capabilities, leading to enhanced performance across various models. Particularly in datasets with intricate structures, FFM consistently outperforms the original models in segmentation tasks. This improvement can be attributed to the inherent characteristics of FD involved in FFM." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 582, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 481, + 666 + ], + "type": "inline_equation", + "content": "FFM_{label}" + }, + { + "bbox": [ + 130, + 582, + 481, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 582, + 481, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{global}" + }, + { + "bbox": [ + 130, + 582, + 481, + 666 + ], + "type": "text", + "content": ": Fractal Dimension is a significant metric used to quantify the complexity of image textures. In the context of model training, we leverage the " + }, + { + "bbox": [ + 130, + 582, + 481, + 666 + ], + "type": "inline_equation", + "content": "FFMs_{label}" + }, + { + "bbox": [ + 130, + 582, + 481, + 666 + ], + "type": "text", + "content": " as weights for the loss function. This method enables us to guide the model's focus towards regions with higher complexity, as indicated by a higher FD. Through extensive experimentation, we have observed that incorporating FFM-based weights into the loss function yields improvements in both the training performance and efficiency of the model, as depicted in Tab. 1." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 197, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 197, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 197, + 91, + 447, + 102 + ], + "type": "text", + "content": "Fractal Feature Maps for Segmentation of Tubular Structures" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 179, + 116, + 434, + 324 + ], + "blocks": [ + { + "bbox": [ + 179, + 116, + 434, + 324 + ], + "lines": [ + { + "bbox": [ + 179, + 116, + 434, + 324 + ], + "spans": [ + { + "bbox": [ + 179, + 116, + 434, + 324 + ], + "type": "image", + "image_path": "baed5fb96d32065d30b487232642e5f700fec30761cd65502c23c33f097548b3.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 335, + 482, + 403 + ], + "lines": [ + { + "bbox": [ + 130, + 335, + 482, + 403 + ], + "spans": [ + { + "bbox": [ + 130, + 335, + 482, + 403 + ], + "type": "text", + "content": "Fig. 4: Comparison of segmentation results. (a) Image. (b) Label. (c) Results of U-Net. (d) Results of existing SOTA approaches. From top to bottom, it's AF-Net, AF-Net, GT-DLA, AF-Net, and Dconn-Net. (e) Results of U-Net*. (f) Results of MD-Net*. Red: true positive. Green: false negative. Blue: false positive. In the ER, MITO, ROSE, STAR, and ROAD rows, columns (b) through (f) display the magnified results of the regions demarcated by yellow squares in the corresponding images of column (a)." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 411, + 482, + 436 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 411, + 482, + 436 + ], + "spans": [ + { + "bbox": [ + 130, + 411, + 482, + 436 + ], + "type": "text", + "content": "These findings highlight the tangible benefits of including " + }, + { + "bbox": [ + 130, + 411, + 482, + 436 + ], + "type": "inline_equation", + "content": "FFM_{label}" + }, + { + "bbox": [ + 130, + 411, + 482, + 436 + ], + "type": "text", + "content": " as a weight parameter in optimizing the training process." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 452, + 274, + 464 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 452, + 274, + 464 + ], + "spans": [ + { + "bbox": [ + 132, + 452, + 274, + 464 + ], + "type": "text", + "content": "4.6 Qualitative Evaluation" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 472, + 482, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 472, + 482, + 556 + ], + "spans": [ + { + "bbox": [ + 130, + 472, + 482, + 556 + ], + "type": "text", + "content": "To facilitate a more comprehensive comparison of the experimental outcomes across various models, we employed visualizations to present the segmentation results, as depicted in Fig. 4. The visual analysis provides a clear and intuitive representation of the model's performance. The observed results validate that the integration of FFM leads to enhanced segmentation outcomes in terms of both edge accuracy and topological continuity. More visualization results can be found in the supplementary material." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 573, + 247, + 588 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 573, + 247, + 588 + ], + "spans": [ + { + "bbox": [ + 132, + 573, + 247, + 588 + ], + "type": "text", + "content": "5 Ablation Study" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 597, + 350, + 611 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 597, + 350, + 611 + ], + "spans": [ + { + "bbox": [ + 132, + 597, + 350, + 611 + ], + "type": "text", + "content": "5.1 Effectiveness of Fractal Feature Maps" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "To ascertain that the observed performance improvement of the model is attributed to the FFM and not the increase of input channels, we replace the " + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "inline_equation", + "content": "FFM_{image}" + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": " with the image, thereby transforming the input (image, " + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "inline_equation", + "content": "FFM_{image}" + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": ") to (image, image). Additionally, we computed the image's Hurst feature (HF," + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 228, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 228, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 228, + 102 + ], + "type": "text", + "content": "J. Huang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "text", + "content": "classical fractal analysis), mean feature (MF, directional analysis), and contrast feature (CF, directional analysis) as introduced by [22]. These features replaced the " + }, + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "inline_equation", + "content": "FFM_{image}" + }, + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "text", + "content": " as input and were trained and evaluated on U-Net and MD-Net. Experimental results demonstrated the advantages of the " + }, + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "inline_equation", + "content": "FFM_{image}" + }, + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "text", + "content": " as shown in Tab. 2. This finding indicates that the inclusion of the FFM plays an effective role in achieving better performance in the segmentation task." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 204, + 343, + 216 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 204, + 343, + 216 + ], + "spans": [ + { + "bbox": [ + 131, + 204, + 343, + 216 + ], + "type": "text", + "content": "5.2 Robustness of Fractal Feature Maps" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 221, + 482, + 328 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 221, + 482, + 328 + ], + "spans": [ + { + "bbox": [ + 130, + 221, + 482, + 328 + ], + "type": "text", + "content": "The computation of the FFM is influenced by the selection of window size and step size. We used different window and step sizes to generate different " + }, + { + "bbox": [ + 130, + 221, + 482, + 328 + ], + "type": "inline_equation", + "content": "FFMs_{image}" + }, + { + "bbox": [ + 130, + 221, + 482, + 328 + ], + "type": "text", + "content": " and tested their performance. The results in Tab. 3 demonstrate the robustness of FFM. When used for training the U-Net and MD-Net, the " + }, + { + "bbox": [ + 130, + 221, + 482, + 328 + ], + "type": "inline_equation", + "content": "FFMs_{image}" + }, + { + "bbox": [ + 130, + 221, + 482, + 328 + ], + "type": "text", + "content": " calculated using different window sizes consistently yield results with a variation range of " + }, + { + "bbox": [ + 130, + 221, + 482, + 328 + ], + "type": "inline_equation", + "content": "0.3\\%" + }, + { + "bbox": [ + 130, + 221, + 482, + 328 + ], + "type": "text", + "content": " on the ER and STARE datasets. Similarly, applying different steps during the calculation of FFMs does not lead to significant changes in performance. It should be noted that the efficiency of computing can be substantially improved by increasing the step size." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 329, + 481, + 353 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 329, + 481, + 353 + ], + "spans": [ + { + "bbox": [ + 130, + 329, + 481, + 353 + ], + "type": "text", + "content": "The numerical results in Tab. 2 and Tab. 3 are calculated by volumetric score IoU. Results with more metrics can be found in the supplementary material." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 369, + 266, + 380 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 369, + 266, + 380 + ], + "spans": [ + { + "bbox": [ + 131, + 369, + 266, + 380 + ], + "type": "text", + "content": "5.3 Decoders of MD-Net" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 387, + 482, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 387, + 482, + 483 + ], + "spans": [ + { + "bbox": [ + 130, + 387, + 482, + 483 + ], + "type": "text", + "content": "To evaluate the effectiveness of the edge decoder and the skeleton decoder, a comparative ablation analysis was conducted between MD-Net and U-Net on the ER and STARE datasets. Subsequently, we proceeded to remove either the edge decoder or the skeleton decoder from MD-Net and trained the modified models. The results shown in Tab. 4 confirm that both the edge decoder and the skeleton decoder contribute to improving the segmentation performance. It is noteworthy that the best performance is achieved when all components are utilized simultaneously." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 498, + 339, + 511 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 498, + 339, + 511 + ], + "spans": [ + { + "bbox": [ + 131, + 498, + 339, + 511 + ], + "type": "text", + "content": "5.4 Limitations of Fractal Feature Map" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 516, + 482, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 516, + 482, + 540 + ], + "spans": [ + { + "bbox": [ + 131, + 516, + 482, + 540 + ], + "type": "text", + "content": "The time complexity of FFMs' computation is associated with the size of image " + }, + { + "bbox": [ + 131, + 516, + 482, + 540 + ], + "type": "inline_equation", + "content": "(M\\times N)" + }, + { + "bbox": [ + 131, + 516, + 482, + 540 + ], + "type": "text", + "content": ", window size " + }, + { + "bbox": [ + 131, + 516, + 482, + 540 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 131, + 516, + 482, + 540 + ], + "type": "text", + "content": ", and step size " + }, + { + "bbox": [ + 131, + 516, + 482, + 540 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 131, + 516, + 482, + 540 + ], + "type": "text", + "content": ". Its time complexity is represented" + } + ] + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 175, + 588, + 438, + 679 + ], + "blocks": [ + { + "bbox": [ + 131, + 544, + 482, + 578 + ], + "lines": [ + { + "bbox": [ + 131, + 544, + 482, + 578 + ], + "spans": [ + { + "bbox": [ + 131, + 544, + 482, + 578 + ], + "type": "text", + "content": "Table 2: Ablation study of FFM in U-Net and MD-Net. The values within the parentheses in the (image, " + }, + { + "bbox": [ + 131, + 544, + 482, + 578 + ], + "type": "inline_equation", + "content": "FFM_{image}" + }, + { + "bbox": [ + 131, + 544, + 482, + 578 + ], + "type": "text", + "content": ") column represent the percentage improvement of IoU compared to the baseline results listed in the (image) column." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 175, + 588, + 438, + 679 + ], + "lines": [ + { + "bbox": [ + 175, + 588, + 438, + 679 + ], + "spans": [ + { + "bbox": [ + 175, + 588, + 438, + 679 + ], + "type": "table", + "html": "
ModelDatasetInput
(image)(image,image)(image,HF)(image,MF)(image,CF)(image,FFM image)
U-NetER75.4475.6175.8975.8675.1076.59(+1.14%)
MITO79.7780.2880.5180.2380.3880.71(+0.94%)
ROSE61.5262.5762.6162.5862.3764.07(+2.55%)
STARE66.1566.0865.5466.1466.8368.07(+1.92%)
ROAD62.4763.4464.9364.5864.2565.74(+3.27%)
MD-NetER76.2876.3476.2475.9775.5277.01(+0.73%)
MITO80.2880.3179.9780.4180.3281.11(+0.83%)
ROSE63.3163.3062.7762.8962.9065.07(+1.76%)
STARE66.4666.5766.4367.0367.2568.49(+2.03%)
ROAD64.7965.0465.2365.1265.1566.07(+1.28%)
", + "image_path": "3b9e74d68964cac9481a8c3df221aa9b538a5caebf82814c86c997491c7a7fa0.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 197, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 197, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 197, + 91, + 447, + 102 + ], + "type": "text", + "content": "Fractal Feature Maps for Segmentation of Tubular Structures" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 175, + 135, + 438, + 194 + ], + "blocks": [ + { + "bbox": [ + 162, + 114, + 451, + 125 + ], + "lines": [ + { + "bbox": [ + 162, + 114, + 451, + 125 + ], + "spans": [ + { + "bbox": [ + 162, + 114, + 451, + 125 + ], + "type": "text", + "content": "Table 3: Ablation of step size and window size in U-Net and MD-Net." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 175, + 135, + 438, + 194 + ], + "lines": [ + { + "bbox": [ + 175, + 135, + 438, + 194 + ], + "spans": [ + { + "bbox": [ + 175, + 135, + 438, + 194 + ], + "type": "table", + "html": "
ModelStep SizeWindow SizeERSTAREModelStep SizeWindow SizeERSTARE
U-Net*11176.7567.87MD-Net*11177.0368.41
1976.7367.911977.0168.21
1776.6567.891776.9268.50
1576.5968.071577.0168.49
2576.4967.872576.9168.46
3576.6867.783576.8868.32
4576.6167.784576.9068.30
", + "image_path": "c574c61e4b5e72fdde153e1297964d6f7b3e109af3c68c3037aa279d582d3cd6.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 175, + 222, + 436, + 340 + ], + "blocks": [ + { + "bbox": [ + 230, + 201, + 382, + 213 + ], + "lines": [ + { + "bbox": [ + 230, + 201, + 382, + 213 + ], + "spans": [ + { + "bbox": [ + 230, + 201, + 382, + 213 + ], + "type": "text", + "content": "Table 4: Ablation study of MD-Net." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 175, + 222, + 436, + 340 + ], + "lines": [ + { + "bbox": [ + 175, + 222, + 436, + 340 + ], + "spans": [ + { + "bbox": [ + 175, + 222, + 436, + 340 + ], + "type": "table", + "html": "
DatasetModelFFMEdge DecoderSkeleton DecoderIoU↑clDice↑ACC↑AUC↑βError↓HD↓
ERU-Net75.4494.6391.8297.3528.726.87
MD-Net✓✓76.2895.2792.0097.3922.556.78
U-Net✓76.5995.4392.0297.5620.786.81
MD-Net✓✓76.9795.6092.0997.5922.426.71
MD-Net✓✓76.8595.5592.1397.5120.636.74
MD-Net✓✓✓77.0995.7492.1497.6519.526.72
STAREU-Net66.1576.0594.6896.393.226.67
MD-Net✓✓66.4676.5494.7696.423.276.47
U-Net✓68.0777.3995.1596.632.776.33
MD-Net✓✓68.2677.8095.2096.772.656.29
MD-Net✓✓68.3977.7695.2096.893.226.31
MD-Net✓✓✓68.7378.2095.2696.992.606.34
", + "image_path": "25730273b3ded7e0de4638e010f5b5611241aa8e04bb93a0284508e168ec2945.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 350, + 482, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 350, + 482, + 413 + ], + "spans": [ + { + "bbox": [ + 130, + 350, + 482, + 413 + ], + "type": "text", + "content": "as: " + }, + { + "bbox": [ + 130, + 350, + 482, + 413 + ], + "type": "inline_equation", + "content": "O\\left(\\frac{M \\cdot N}{S^2} \\cdot \\left(w^2 \\cdot \\log(w)\\right)\\right)" + }, + { + "bbox": [ + 130, + 350, + 482, + 413 + ], + "type": "text", + "content": ". Calculation of FFM indeed incurs computational overhead but is optimized by using multithreading and larger " + }, + { + "bbox": [ + 130, + 350, + 482, + 413 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 130, + 350, + 482, + 413 + ], + "type": "text", + "content": ". Currently, the generation of FFMs for a " + }, + { + "bbox": [ + 130, + 350, + 482, + 413 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 130, + 350, + 482, + 413 + ], + "type": "text", + "content": " image takes about 150 milliseconds on a Xeon 8336C CPU, given a step size of 1 and a window size of 5. Detailed information on training and inference time can be found in the supplementary material." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 439, + 226, + 452 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 439, + 226, + 452 + ], + "spans": [ + { + "bbox": [ + 132, + 439, + 226, + 452 + ], + "type": "text", + "content": "6 Conclusions" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 474, + 482, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 474, + 482, + 668 + ], + "spans": [ + { + "bbox": [ + 130, + 474, + 482, + 668 + ], + "type": "text", + "content": "In this study, we propose a method that uses fractal feature maps (FFMs) along with a multi-decoder network (MD-Net) for the semantic segmentation of tubular structures. FFMs are used to capture the texture and self-similarity of image regions or label regions at the pixel-level through the fractal dimension. " + }, + { + "bbox": [ + 130, + 474, + 482, + 668 + ], + "type": "inline_equation", + "content": "FFM_{image}" + }, + { + "bbox": [ + 130, + 474, + 482, + 668 + ], + "type": "text", + "content": " are used as an additional input to enhance the segmentation model's perception of tubular structures and " + }, + { + "bbox": [ + 130, + 474, + 482, + 668 + ], + "type": "inline_equation", + "content": "FFM_{label}" + }, + { + "bbox": [ + 130, + 474, + 482, + 668 + ], + "type": "text", + "content": " are used as a weight for the loss function to guide the model training. MD-Net improves the quality of segmentation by simultaneously predicting edges and skeletons through the incorporation of an edge decoder and a skeleton decoder. The proposed method is evaluated on five datasets of tubular structures and one dataset of non-tubular structures. The results demonstrate the superiority of MD-Net with FFM in the segmentation of tubular structures and stable performance in the segmentation of non-tubular structures. Furthermore, the performance improvement observed when incorporating the FFM module into vanilla U-Net and HR-Net underscores the broad applicability and potential of incorporating fractal information into deep learning models." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 228, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 228, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 228, + 102 + ], + "type": "text", + "content": "J. Huang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 143, + 482, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 143, + 482, + 203 + ], + "spans": [ + { + "bbox": [ + 130, + 143, + 482, + 203 + ], + "type": "text", + "content": "This work was supported in part by the National Natural Science Foundation of China (grants 92354307, 91954201, 31971289, 32101216), the Strategic Priority Research Program of the Chinese Academy of Sciences (grant XDB37040402) and the Fundamental Research Funds for the Central Universities (grant E3E45201X2)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 224, + 197, + 237 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 224, + 197, + 237 + ], + "spans": [ + { + "bbox": [ + 132, + 224, + 197, + 237 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 134, + 252, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 138, + 252, + 481, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 252, + 481, + 319 + ], + "spans": [ + { + "bbox": [ + 138, + 252, + 481, + 319 + ], + "type": "text", + "content": "1. Alvarez, L., Trujillo, A., Cuenca, C., González, E., Esclarín, J., Gomez, L., Mazarra, L., Aleman-Flores, M., Tahoces, P.G., Carreira, J.M.: Tracking the aortic lumen geometry by optimizing the 3d orientation of its cross-sections. In: Medical Image Computing and Computer-Assisted Intervention- MICCAI 2017: 20th International Conference, Quebec City, QC, Canada, September 11-13, 2017, Proceedings, Part II 20. pp. 174-181. Springer (2017)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 320, + 481, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 320, + 481, + 352 + ], + "spans": [ + { + "bbox": [ + 138, + 320, + 481, + 352 + ], + "type": "text", + "content": "2. Antiga, L., Ene-Iordache, B., Remuzzi, A.: Computational geometry for patient-specific reconstruction and meshing of blood vessels from mr and ct angiography. IEEE Transactions on Medical Imaging 22(5), 674-684 (2003)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 354, + 481, + 375 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 354, + 481, + 375 + ], + "spans": [ + { + "bbox": [ + 138, + 354, + 481, + 375 + ], + "type": "text", + "content": "3. Araújo, R.J., Cardoso, J.S., Oliveira, H.P.: Topological similarity index and loss function for blood vessel segmentation. arXiv preprint arXiv:2107.14531 (2021)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 376, + 481, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 376, + 481, + 408 + ], + "spans": [ + { + "bbox": [ + 138, + 376, + 481, + 408 + ], + "type": "text", + "content": "4. Bauer, C., Pock, T., Sorantin, E., Bischof, H., Beichel, R.: Segmentation of interwoven 3d tubular tree structures utilizing shape priors and graph cuts. Medical Image Analysis 14(2), 172-184 (2010)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 410, + 481, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 410, + 481, + 453 + ], + "spans": [ + { + "bbox": [ + 138, + 410, + 481, + 453 + ], + "type": "text", + "content": "5. Caicedo, J.C., Goodman, A., Karhohs, K.W., Cimini, B.A., Ackerman, J., Haghighi, M., Heng, C., Becker, T., Doan, M., McQuin, C., et al.: Nucleus segmentation across imaging experiments: the 2018 data science bowl. Nature Methods 16(12), 1247-1253 (2019)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 454, + 481, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 454, + 481, + 475 + ], + "spans": [ + { + "bbox": [ + 138, + 454, + 481, + 475 + ], + "type": "text", + "content": "6. Caselles, V., Kimmel, R., Sapiro, G.: Geodesic active contours. International Journal of Computer Vision 22, 61-79 (1997)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 476, + 481, + 509 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 476, + 481, + 509 + ], + "spans": [ + { + "bbox": [ + 138, + 476, + 481, + 509 + ], + "type": "text", + "content": "7. Challoo, M., Gao, Y., Busch, A.: Distinctive phase interdependency model for retinal vasculature delineation in oct-angiography images. IEEE Transactions on Medical Imaging (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 510, + 481, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 510, + 481, + 542 + ], + "spans": [ + { + "bbox": [ + 138, + 510, + 481, + 542 + ], + "type": "text", + "content": "8. Chen, J., Lu, Y., Yu, Q., Luo, X., Adeli, E., Wang, Y., Lu, L., Yuille, A.L., Zhou, Y.: Transunet: Transformers make strong encoders for medical image segmentation. arXiv preprint arXiv:2102.04306 (2021)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 544, + 481, + 576 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 544, + 481, + 576 + ], + "spans": [ + { + "bbox": [ + 138, + 544, + 481, + 576 + ], + "type": "text", + "content": "9. Dai, J., Qi, H., Xiong, Y., Li, Y., Zhang, G., Hu, H., Wei, Y.: Deformable convolutional networks. In: Proceedings of the IEEE international conference on computer vision. pp. 764-773 (2017)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 577, + 481, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 577, + 481, + 609 + ], + "spans": [ + { + "bbox": [ + 138, + 577, + 481, + 609 + ], + "type": "text", + "content": "0. Dong, S., Pan, Z., Fu, Y., Yang, Q., Gao, Y., Yu, T., Shi, Y., Zhuo, C.: Deu-net 2.0: Enhanced deformable u-net for 3d cardiac cine migraine segmentation. Medical Image Analysis 78, 102389 (2022)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 134, + 611, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 611, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 134, + 611, + 481, + 665 + ], + "type": "text", + "content": "11. Guo, Y., Huang, J., Zhou, Y., Luo, Y., Li, W., Yang, G.: Segmentation of intracellular structures in fluorescence microscopy images by fusing low-level features. In: Pattern Recognition and Computer Vision: 4th Chinese Conference, PRCV 2021, Beijing, China, October 29–November 1, 2021, Proceedings, Part III 4. pp. 386–397. Springer (2021)" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 197, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 197, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 197, + 91, + 447, + 102 + ], + "type": "text", + "content": "Fractal Feature Maps for Segmentation of Tubular Structures" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 666 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "type": "text", + "content": "12. Gupta, S., Zhang, Y., Hu, X., Prasanna, P., Chen, C.: Topology-aware uncertainty for image segmentation. Advances in Neural Information Processing Systems 36 (2024)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 149, + 481, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 149, + 481, + 182 + ], + "spans": [ + { + "bbox": [ + 133, + 149, + 481, + 182 + ], + "type": "text", + "content": "13. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 770-778 (2016)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 182, + 481, + 225 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 182, + 481, + 225 + ], + "spans": [ + { + "bbox": [ + 133, + 182, + 481, + 225 + ], + "type": "text", + "content": "14. He, Y., Ge, R., Wu, J., Coatrieux, J.L., Shu, H., Chen, Y., Yang, G., Li, S.: Thin semantics enhancement via high-frequency priori rule for thin structures segmentation. In: 2021 IEEE International Conference on Data Mining (ICDM). pp. 1096-1101. IEEE (2021)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 225, + 481, + 257 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 225, + 481, + 257 + ], + "spans": [ + { + "bbox": [ + 133, + 225, + 481, + 257 + ], + "type": "text", + "content": "15. Hoover, A., Goldbaum, M.: Locating the optic nerve in a retinal image using the fuzzy convergence of the blood vessels. IEEE Transactions on Medical Imaging 22(8), 951-958 (2003)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 133, + 257, + 481, + 278 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 257, + 481, + 278 + ], + "spans": [ + { + "bbox": [ + 133, + 257, + 481, + 278 + ], + "type": "text", + "content": "16. Hotamisligil, G.S.: Endoplasmic reticulum stress and the inflammatory basis of metabolic disease. Cell 140(6), 900-917 (2010)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 133, + 278, + 481, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 278, + 481, + 300 + ], + "spans": [ + { + "bbox": [ + 133, + 278, + 481, + 300 + ], + "type": "text", + "content": "17. Hu, X., Li, F., Samaras, D., Chen, C.: Topology-preserving deep image segmentation. Advances in neural information processing systems 32 (2019)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 133, + 300, + 481, + 321 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 300, + 481, + 321 + ], + "spans": [ + { + "bbox": [ + 133, + 300, + 481, + 321 + ], + "type": "text", + "content": "18. Hu, X., Wang, Y., Fuxin, L., Samaras, D., Chen, C.: Topology-aware segmentation using discrete Morse theory. arXiv preprint arXiv:2103.09992 (2021)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 133, + 321, + 481, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 321, + 481, + 342 + ], + "spans": [ + { + "bbox": [ + 133, + 321, + 481, + 342 + ], + "type": "text", + "content": "19. Huang, Y., Tang, Z., Chen, D., Su, K., Chen, C.: Batching soft iou for training semantic segmentation networks. IEEE Signal Processing Letters 27, 66-70 (2019)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 133, + 342, + 481, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 342, + 481, + 374 + ], + "spans": [ + { + "bbox": [ + 133, + 342, + 481, + 374 + ], + "type": "text", + "content": "20. Isensee, F., Jaeger, P.F., Kohl, S.A., Petersen, J., Maier-Hein, K.H.: nnu-net: a self-configuring method for deep learning-based biomedical image segmentation. Nature Methods 18(2), 203-211 (2021)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 374, + 481, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 374, + 481, + 407 + ], + "spans": [ + { + "bbox": [ + 132, + 374, + 481, + 407 + ], + "type": "text", + "content": "21. Jin, Q., Meng, Z., Pham, T.D., Chen, Q., Wei, L., Su, R.: Dunet: A deformable network for retinal vessel segmentation. Knowledge-Based Systems 178, 149-162 (2019)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 407, + 481, + 429 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 407, + 481, + 429 + ], + "spans": [ + { + "bbox": [ + 132, + 407, + 481, + 429 + ], + "type": "text", + "content": "22. Kaplan, L.M.: Extended fractal analysis for texture classification and segmentation. IEEE transactions on image processing 8(11), 1572-1585 (1999)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 429, + 481, + 449 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 429, + 481, + 449 + ], + "spans": [ + { + "bbox": [ + 132, + 429, + 481, + 449 + ], + "type": "text", + "content": "23. Kaplan, L.M., Kuo, C.C.: Extending self-similarity for fractional brownian motion. IEEE Transactions on Signal Processing 42(12), 3526-3530 (1994)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 449, + 481, + 482 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 449, + 481, + 482 + ], + "spans": [ + { + "bbox": [ + 132, + 449, + 481, + 482 + ], + "type": "text", + "content": "24. Konatar, I., Popovic, T., Popovic, N.: Box-counting method in python for fractal analysis of biomedical images. In: 2020 24th International Conference on Information Technology (IT). pp. 1-4. IEEE (2020)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 482, + 481, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 482, + 481, + 525 + ], + "spans": [ + { + "bbox": [ + 132, + 482, + 481, + 525 + ], + "type": "text", + "content": "25. Kreitner, L., Paetzold, J.C., Rauch, N., Chen, C., Hagag, A.M., Fayed, A.E., Sivaprasad, S., Rausch, S., Weichsel, J., Menze, B.H., et al.: Synthetic optical coherence tomography angiographs for detailed retinal vessel segmentation without human annotations. IEEE Transactions on Medical Imaging (2024)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 525, + 481, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 525, + 481, + 568 + ], + "spans": [ + { + "bbox": [ + 132, + 525, + 481, + 568 + ], + "type": "text", + "content": "26. La Barbera, G., Rouet, L., Boussaid, H., Lubet, A., Kassir, R., Sarnacki, S., Gori, P., Bloch, I.: Tubular structures segmentation of pediatric abdominal-visceral cect images with renal tumors: assessment, comparison and improvement. Medical Image Analysis 90, 102986 (2023)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 568, + 481, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 568, + 481, + 601 + ], + "spans": [ + { + "bbox": [ + 132, + 568, + 481, + 601 + ], + "type": "text", + "content": "27. Laptev, I., Mayer, H., Lindeberg, T., Eckstein, W., Steger, C., Baumgartner, A.: Automatic extraction of roads from aerial images based on scale space and snakes. Machine Vision and Applications 12, 23-31 (2000)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 601, + 481, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 601, + 481, + 633 + ], + "spans": [ + { + "bbox": [ + 132, + 601, + 481, + 633 + ], + "type": "text", + "content": "28. Lee, C.A., Blackstone, C.: Er morphology and endo-lysosomal crosstalk: Functions and disease implications. Biochimica et Biophysica Acta (BBA)-Molecular and Cell Biology of Lipids 1865(1), 158544 (2020)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 633, + 481, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 633, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 633, + 481, + 666 + ], + "type": "text", + "content": "29. Lee, W.L., Hsieh, K.S.: A robust algorithm for the fractal dimension of images and its applications to the classification of natural images and ultrasonic liver images. Signal Processing 90(6), 1894-1904 (2010)" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 228, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 228, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 228, + 102 + ], + "type": "text", + "content": "J. Huang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 665 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 138 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 138 + ], + "type": "text", + "content": "30. Li, J., Du, Q., Sun, C.: An improved box-counting method for image fractal dimension estimation. Pattern Recognition 42(11), 2460-2469 (2009)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 138, + 482, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 138, + 482, + 182 + ], + "spans": [ + { + "bbox": [ + 130, + 138, + 482, + 182 + ], + "type": "text", + "content": "31. Li, Y., Zhang, Y., Liu, J.Y., Wang, K., Zhang, K., Zhang, G.S., Liao, X.F., Yang, G.: Global transformer and dual local attention network via deep-shallow hierarchical feature fusion for retinal vessel segmentation. IEEE Transactions on Cybernetics 53(9), 5826-5839 (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 182, + 482, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 182, + 482, + 215 + ], + "spans": [ + { + "bbox": [ + 130, + 182, + 482, + 215 + ], + "type": "text", + "content": "32. Lin, P.L., Huang, P.W., Lee, C.H., Wu, M.T.: Automatic classification for solitary pulmonary nodule in ct image by fractal analysis based on fractional brownian motion model. Pattern Recognition 46(12), 3279-3287 (2013)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 215, + 482, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 215, + 482, + 258 + ], + "spans": [ + { + "bbox": [ + 130, + 215, + 482, + 258 + ], + "type": "text", + "content": "33. Lin, P., Huang, P., Huang, P., Hsu, H.: Alveolar bone-loss area localization in periodontitis radiographs based on threshold segmentation with a hybrid feature fused of intensity and the h-value of fractional brownian motion model. Computer Methods and Programs in Biomedicine 121(3), 117-126 (2015)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 259, + 482, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 259, + 482, + 293 + ], + "spans": [ + { + "bbox": [ + 130, + 259, + 482, + 293 + ], + "type": "text", + "content": "34. Long, J., Shelhamer, E., Darrell, T.: Fully convolutional networks for semantic segmentation. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 3431-3440 (2015)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 293, + 482, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 293, + 482, + 314 + ], + "spans": [ + { + "bbox": [ + 130, + 293, + 482, + 314 + ], + "type": "text", + "content": "35. Lopes, R., Betrouni, N.: Fractal and multifractal analysis: A review. Medical Image Analysis 13(4), 634-649 (2009)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 314, + 482, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 314, + 482, + 357 + ], + "spans": [ + { + "bbox": [ + 130, + 314, + 482, + 357 + ], + "type": "text", + "content": "36. Luo, Y., Guo, Y., Li, W., Liu, G., Yang, G.: Fluorescence microscopy image datasets for deep learning segmentation of intracellular organelle networks (2020). https://doi.org/10.21227/t2he-zn97, https://dx.doi.org/10.21227/t2he-zn97" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 357, + 482, + 390 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 357, + 482, + 390 + ], + "spans": [ + { + "bbox": [ + 130, + 357, + 482, + 390 + ], + "type": "text", + "content": "37. Ma, Y., Hao, H., Xie, J., Fu, H., Zhang, J., Yang, J., Wang, Z., Liu, J., Zheng, Y., Zhao, Y.: Rose: a retinal oct-angiography vessel segmentation dataset and new model. IEEE Transactions on Medical Imaging 40(3), 928–939 (2020)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 390, + 482, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 390, + 482, + 413 + ], + "spans": [ + { + "bbox": [ + 130, + 390, + 482, + 413 + ], + "type": "text", + "content": "38. Mnih, V.: Machine learning for aerial image labeling. University of Toronto (Canada) (2013)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 413, + 482, + 467 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 413, + 482, + 467 + ], + "spans": [ + { + "bbox": [ + 130, + 413, + 482, + 467 + ], + "type": "text", + "content": "39. Mou, L., Zhao, Y., Chen, L., Cheng, J., Gu, Z., Hao, H., Qi, H., Zheng, Y., Frangi, A., Liu, J.: Cs-net: Channel and spatial attention network for curvilinear structure segmentation. In: Medical Image Computing and Computer Assisted Intervention-MICCAI 2019: 22nd International Conference, Shenzhen, China, October 13-17, 2019, Proceedings, Part I 22. pp. 721-730. Springer (2019)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 468, + 482, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 468, + 482, + 511 + ], + "spans": [ + { + "bbox": [ + 130, + 468, + 482, + 511 + ], + "type": "text", + "content": "40. Nain, D., Yezzi, A., Turk, G.: Vessel segmentation using a shape driven flow. In: Medical Image Computing and Computer-Assisted Intervention-MICCAI 2004: 7th International Conference, Saint-Malo, France, September 26-29, 2004. Proceedings, Part I 7. pp. 51-59. Springer (2004)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 511, + 482, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 511, + 482, + 533 + ], + "spans": [ + { + "bbox": [ + 130, + 511, + 482, + 533 + ], + "type": "text", + "content": "41. Pentland, A.P.: Fractal-based description of natural scenes. IEEE Transactions on Pattern Analysis and Machine Intelligence (6), 661-674 (1984)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 533, + 482, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 482, + 577 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 482, + 577 + ], + "type": "text", + "content": "42. Qi, Y., He, Y., Qi, X., Zhang, Y., Yang, G.: Dynamic snake convolution based on topological geometric constraints for tubular structure segmentation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 6070-6079 (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 577, + 482, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 577, + 482, + 620 + ], + "spans": [ + { + "bbox": [ + 130, + 577, + 482, + 620 + ], + "type": "text", + "content": "43. Roberto, G.F., Lumini, A., Neves, L.A., do Nascimento, M.Z.: Fractal neural network: A new ensemble of fractal geometry and convolutional neural networks for the classification of histology images. Expert Systems with Applications 166, 114103 (2021)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 620, + 482, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 620, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 620, + 482, + 665 + ], + "type": "text", + "content": "44. Ronneberger, O., Fischer, P., Brox, T.: U-net: Convolutional networks for biomedical image segmentation. In: Medical Image Computing and Computer-Assisted Intervention-MICCAI 2015: 18th International Conference, Munich, Germany, October 5-9, 2015, Proceedings, Part III 18. pp. 234-241. Springer (2015)" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 197, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 197, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 197, + 91, + 447, + 102 + ], + "type": "text", + "content": "Fractal Feature Maps for Segmentation of Tubular Structures" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "type": "text", + "content": "45. Shi, T., Ding, X., Zhou, W., Pan, F., Yan, Z., Bai, X., Yang, X.: Affinity feature strengthening for accurate, complete and robust vessel segmentation. IEEE Journal of Biomedical and Health Informatics 27(8), 4006-4017 (2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 149, + 481, + 192 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 149, + 481, + 192 + ], + "spans": [ + { + "bbox": [ + 133, + 149, + 481, + 192 + ], + "type": "text", + "content": "46. Shit, S., Paetzold, J.C., Sekuboyina, A., Ezhov, I., Unger, A., Zhylka, A., Pluim, J.P., Bauer, U., Menze, B.H.: cldice-a novel topology-preserving loss function for tubular structure segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16560-16569 (2021)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 192, + 481, + 224 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 192, + 481, + 224 + ], + "spans": [ + { + "bbox": [ + 132, + 192, + 481, + 224 + ], + "type": "text", + "content": "47. Sironi, A., Lepetit, V., Fua, P.: Multiscale centerline detection by learning a scale-space distance transform. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 2697-2704 (2014)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 224, + 481, + 246 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 224, + 481, + 246 + ], + "spans": [ + { + "bbox": [ + 132, + 224, + 481, + 246 + ], + "type": "text", + "content": "48. Taha, A.A., Hanbury, A.: Metrics for evaluating 3d medical image segmentation: analysis, selection, and tool. BMC Medical Imaging 15(1), 1-28 (2015)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 246, + 481, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 246, + 481, + 300 + ], + "spans": [ + { + "bbox": [ + 132, + 246, + 481, + 300 + ], + "type": "text", + "content": "49. Wang, C., Hayashi, Y., Oda, M., Itoh, H., Kitasaka, T., Frangi, A.F., Mori, K.: Tubular structure segmentation using spatially connected network with radial distance loss for 3d medical images. In: Medical Image Computing and Computer Assisted Intervention-MICCAI 2019: 22nd International Conference, Shenzhen, China, October 13-17, 2019, Proceedings, Part VI 22. pp. 348-356. Springer (2019)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 300, + 481, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 300, + 481, + 342 + ], + "spans": [ + { + "bbox": [ + 132, + 300, + 481, + 342 + ], + "type": "text", + "content": "50. Wang, J., Sun, K., Cheng, T., Jiang, B., Deng, C., Zhao, Y., Liu, D., Mu, Y., Tan, M., Wang, X., et al.: Deep high-resolution representation learning for visual recognition. IEEE Transactions on Pattern Analysis and Machine Intelligence 43(10), 3349-3364 (2020)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 342, + 481, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 342, + 481, + 386 + ], + "spans": [ + { + "bbox": [ + 132, + 342, + 481, + 386 + ], + "type": "text", + "content": "51. Wang, Y., Wei, X., Liu, F., Chen, J., Zhou, Y., Shen, W., Fishman, E.K., Yuille, A.L.: Deep distance transform for tubular structure segmentation in ct scans. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3833-3842 (2020)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 386, + 481, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 386, + 481, + 418 + ], + "spans": [ + { + "bbox": [ + 132, + 386, + 481, + 418 + ], + "type": "text", + "content": "52. Wong, C.C., Vong, C.M.: Persistent homology based graph convolution network for fine-grained 3d shape segmentation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 7098-7107 (2021)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 418, + 481, + 450 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 418, + 481, + 450 + ], + "spans": [ + { + "bbox": [ + 132, + 418, + 481, + 450 + ], + "type": "text", + "content": "53. Yan, Z., Yang, X., Cheng, K.T.: A three-stage deep learning model for accurate retinal vessel segmentation. IEEE Journal of Biomedical and Health Informatics 23(4), 1427-1436 (2018)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 450, + 481, + 482 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 450, + 481, + 482 + ], + "spans": [ + { + "bbox": [ + 132, + 450, + 481, + 482 + ], + "type": "text", + "content": "54. Yang, X., Li, Z., Guo, Y., Zhou, D.: Dcu-net: A deformable convolutional neural network based on cascade u-net for retinal vessel segmentation. Multimedia Tools and Applications 81(11), 15593-15607 (2022)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 482, + 481, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 482, + 481, + 514 + ], + "spans": [ + { + "bbox": [ + 132, + 482, + 481, + 514 + ], + "type": "text", + "content": "55. Yang, Z., Farsiu, S.: Directional connectivity-based segmentation of medical images. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11525-11535 (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 514, + 481, + 546 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 514, + 481, + 546 + ], + "spans": [ + { + "bbox": [ + 132, + 514, + 481, + 546 + ], + "type": "text", + "content": "56. Yim, P.J., Cebral, J.J., Mullick, R., Marcos, H.B., Choyke, P.L.: Vessel surface reconstruction with a tubular deformable model. IEEE Transactions on Medical Imaging 20(12), 1411-1421 (2001)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 546, + 481, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 546, + 481, + 578 + ], + "spans": [ + { + "bbox": [ + 132, + 546, + 481, + 578 + ], + "type": "text", + "content": "57. Yu, F., Koltun, V., Funkhouser, T.: Dilated residual networks. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 472-480 (2017)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 578, + 481, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 578, + 481, + 621 + ], + "spans": [ + { + "bbox": [ + 132, + 578, + 481, + 621 + ], + "type": "text", + "content": "58. Zhao, Y., Zheng, Y., Liu, Y., Zhao, Y., Luo, L., Yang, S., Na, T., Wang, Y., Liu, J.: Automatic 2-d/3-d vessel enhancement in multiple modality images using a weighted symmetry filter. IEEE Transactions on Medical Imaging 37(2), 438-450 (2017)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 621, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 621, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 621, + 481, + 665 + ], + "type": "text", + "content": "59. Zhou, Y., Huang, J., Wang, C., Song, L., Yang, G.: Xnet: Wavelet-based low and high frequency fusion networks for fully- and semi-supervised semantic segmentation of biomedical images. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 21085-21096 (October 2023)" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 228, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 228, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 228, + 102 + ], + "type": "text", + "content": "J. Huang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 482, + 216 + ], + "type": "list", + "angle": 0, + "index": 4, + "blocks": [ + { + "bbox": [ + 132, + 116, + 482, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 482, + 182 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 482, + 182 + ], + "type": "text", + "content": "60. Zhou, Z., Rahman Siddiquee, M.M., Tajbakhsh, N., Liang, J.: Unet++: A nested u-net architecture for medical image segmentation. In: Deep Learning in Medical Image Analysis and Multimodal Learning for Clinical Decision Support: 4th International Workshop, DLMIA 2018, and 8th International Workshop, ML-CDS 2018, Held in Conjunction with MICCAI 2018, Granada, Spain, September 20, 2018, Proceedings 4. pp. 3-11. Springer (2018)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 183, + 481, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 183, + 481, + 216 + ], + "spans": [ + { + "bbox": [ + 132, + 183, + 481, + 216 + ], + "type": "text", + "content": "61. Zhuang, Z., Lei, N., Joseph Raj, A.N., Qiu, S.: Application of fractal theory and fuzzy enhancement in ultrasound image segmentation. Medical & Biological Engineering & Computing 57, 623-632 (2019)" + } + ] + } + ], + "index": 3 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 197, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 197, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 197, + 91, + 447, + 102 + ], + "type": "text", + "content": "Fractal Feature Maps for Segmentation of Tubular Structures" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "19" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 18 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Reprojection Errors as Prompts for Efficient Scene Coordinate Regression/91ff6348-fcc0-42f1-b6d0-08cddca0de4a_content_list.json b/2024/Reprojection Errors as Prompts for Efficient Scene Coordinate Regression/91ff6348-fcc0-42f1-b6d0-08cddca0de4a_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..4be9f541cc22e252d1719b3e2c8f5dcc856a9f3d --- /dev/null +++ b/2024/Reprojection Errors as Prompts for Efficient Scene Coordinate Regression/91ff6348-fcc0-42f1-b6d0-08cddca0de4a_content_list.json @@ -0,0 +1,1487 @@ +[ + { + "type": "text", + "text": "Reprojection Errors as Prompts for Efficient Scene Coordinate Regression", + "text_level": 1, + "bbox": [ + 248, + 140, + 754, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Ting-Ru Liu $^{1}$ , Hsuan-Kung Yang $^{2}$ , Jou-Min Liu $^{1*}$ , Chun-Wei Huang $^{1*}$ , Tsung-Chih Chiang $^{1*}$ , Quan Kong $^{2}$ , Norimasa Kobori $^{2}$ , and Chun-Yi Lee $^{1}$", + "bbox": [ + 233, + 210, + 767, + 243 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ National Tsing Hua University $^{2}$ Woven by Toyota, Inc. {tingru, diasimui, alan, george, cymaxwelllee}@elsa.cs.nthu.edu.tw {hsuan-kung.yang, quan.kong, norimasa.kobori} @woven toyota https://tingru0203.github.io/egfs", + "bbox": [ + 245, + 253, + 756, + 311 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Scene coordinate regression (SCR) methods have emerged as a promising area of research due to their potential for accurate visual localization. However, many existing SCR approaches train on samples from all image regions, including dynamic objects and texture-less areas. Utilizing these areas for optimization during training can potentially hamper the overall performance and efficiency of the model. In this study, we first perform an in-depth analysis to validate the adverse impacts of these areas. Drawing inspiration from our analysis, we then introduce an error-guided feature selection (EGFS) mechanism, in tandem with the use of the Segment Anything Model (SAM) [1]. This mechanism seeds low reprojection areas as prompts and expands them into error-guided masks, and then utilizes these masks to sample points and filter out problematic areas in an iterative manner. The experiments demonstrate that our method outperforms existing SCR approaches that do not rely on 3D information on the Cambridge Landmarks and Indoor6 datasets.", + "bbox": [ + 261, + 349, + 743, + 559 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 217, + 585, + 375, + 599 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The objective of visual localization is to estimate a 6-DoF camera pose from images, a key component in fields such as Augmented Reality, Virtual Reality, and autonomous driving. Contemporary leading methods in visual localization typically involve establishing 2D-3D correspondences and then utilizing Perspective-Point (PnP) [2] with RANSAC [3] for camera pose estimation. These methods can be broadly classified into two main directions: feature-matching [4] and scene coordinate regression (SCR) [5]. Feature-matching approaches reconstruct a 3D scene using Structure from Motion (SfM), identify and describe key points in 2D images [6,7], and link these to 3D coordinates [8,9]. Nevertheless, they may encounter challenges such as high computational demands, significant storage requirements, and potential privacy concerns [10]. On the other hand, SCR methods [11-17] employ deep neural networks (DNNs) to predict the 3D coordinates of pixels and then utilize PnP with RANSAC for camera pose estimation, which", + "bbox": [ + 212, + 617, + 787, + 815 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* indicates equal contribution.", + "bbox": [ + 217, + 824, + 423, + 839 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/81109ee5448de08a173c3c5cdd9c5f07af530f24c707302e16cb5843c84ea9a8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 217, + 179, + 351, + 251 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/883e196dd6e558123f277c64d848b1800e7d3b73c6cb98ec4f45f7f1ca1cc253.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 352, + 178, + 483, + 251 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/dcd7b4046dfa769bf8dec1e588187ab33e4d3e3f807d1a020d7117ab844414ba.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 483, + 178, + 625, + 251 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/5b35f737699fb2b972c91d3bf48aa007781d148bd1689ce057f968a67c4533fd.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 629, + 179, + 785, + 251 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/03932a13732059a84a631165e27a977cf78cf596ee03c4ee09b3871fbb483151.jpg", + "image_caption": [ + "Fig. 1: Visualization of the primary components (i.e., (d)-(h)) introduced in the proposed visual localization scheme. (d) illustrates the point prompts selected from (b) with low reprojection errors, while (e) presents an error-guided mask expanded from the prompted points in (d) using SAM. (f) displays the proposed error-guided feature selection (EGFS), which refines the mask from (e) with the predicted confidence map (c) to ensure high-quality scene coordinates are sampled for estimating the final camera pose. The point cloud constructed from the predicted scene coordinates is shown on the right-hand side (i.e., (g)-(h)), with the confidence (yellow parts) and the refined EGFS mask (green for selected areas; red for rejected areas)." + ], + "image_footnote": [], + "bbox": [ + 217, + 260, + 625, + 333 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/ce4fd49ac586c0410b143e90782785e1f88e82b72dfbfb28fb472af3e0afaee2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 630, + 261, + 785, + 333 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "provide benefits such as accuracy in smaller scenes, reduced training times, as well as minimized storage requirements. Given these advantages, SCR is thus the primary focus of this study and presents potential for further enhancements.", + "bbox": [ + 212, + 532, + 787, + 579 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Albeit effective, previous SCR methods face two primary challenges to be addressed: the presence of dynamic objects and texture-less regions. Dynamic objects, such as pedestrians and cars, pose difficulties for these techniques due to their changing nature and the unsuitable features extracted from these regions. As illustrated in Fig. 1, these regions could lead to high reprojection errors during training. The second challenge is that current methods struggle with flat, texture-less surfaces. Despite their apparent simplicity, these areas often result in inaccurate scene coordinates due to difficulties in feature extraction. While previous methods [14, 15, 17] employ techniques like RANSAC to filter out outliers and differentiable end-to-end optimization approaches to disregard these outliers, they demand significant computational time for end-to-end processing and do not account for the semantic meaning of the selected areas, which may still result in the selection of outlier areas. Thus, they might not effectively prevent the models from unstable weight updates and can lead to training instability.", + "bbox": [ + 212, + 580, + 787, + 792 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In light of the aforementioned issues, this study aims to explore the possibility of guiding the sampling process to favor regions with low reprojection errors, and seeks to leverage visual information in an image to expand masks", + "bbox": [ + 212, + 795, + 787, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "T.-R. Liu et al.", + "bbox": [ + 271, + 114, + 375, + 127 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "that bear semantic regional meanings. The core philosophy of our methodology focuses on selecting robust regions for training without the need to manually define explicit areas or categories. Such robust regions might differ across scenes, which make it inappropriate for manual definition, as elaborated in Section 4. More specifically, we propose a strategy, named Error-Guided Feature Selection (EGFS), for deriving low reprojection error samples as point prompts, and expanding them to encompass a complete mask with similar semantic meanings. The masks are iteratively updated after being frozen for a specified period, which enable the proposed model to dynamically renew the focused regions. To achieve this prompt-based semantic regions expansion, we employ the Segment Anything Model (SAM) [1], a vision foundation model capable of providing a general understanding of the scene. This error-guided scheme ensures that our training process concentrates on regions with low reprojection errors and adaptively reduces the variability introduced by dynamic or texture-less objects. Moreover, EGPS adopts a predicted confidence map to refine the expanded error-guided mask for updates, which ensures the regions marked by the masks are sufficiently reliable. An illustrations of each component is provided in Fig. 1.", + "bbox": [ + 217, + 146, + 785, + 402 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To validate the effectiveness of our proposed method, we conduct evaluations on both outdoor and indoor datasets, including the Cambridge Landmarks dataset [18] and the Indoor6 [19] dataset, for evaluating the performance of our proposed SCR methodology. The experimental results suggest that our method can be broadly applied to previous methods and provides benefits, and leads to state-of-the-art (SOTA) performance when compared with SCR approaches without leveraging any 3D information in terms of estimated translational and rotational errors, while requiring less training time and smaller model sizes. We further perform a series of detailed analyses and ablation studies to confirm the effectiveness of our error-guided sampled regions and the proposed methodology.", + "bbox": [ + 217, + 404, + 785, + 553 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 217, + 573, + 385, + 589 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Scene Coordinate Regression. Conventional SCR approaches typically begin by establishing 2D-3D correspondences and then utilize RANSAC-based optimization to estimate the camera pose. To establish 2D-3D correspondences, previous endeavors have explored the use of random forests [5, 20-22] or convolutional neural networks (CNNs) [11-15,17,23] to regress scene coordinates from images. Since SCR typically requires scene-specific training, training time has emerged as a critical concern in practical applications. The ACE [17] method, which can be trained within five minutes, stands out as a promising solution. Moreover, unlike methods that map every 2D pixel to a 3D coordinate, some works [19,24,25] have aimed at learning to detect 3D landmarks to establish 2D-3D correspondences. However, they often require the reconstruction of a 3D model, which incurs additional computational time and costs. Our method seeks to utilize SAM to derive error-guided masks from 2D images, while maintaining the training efficiency.", + "bbox": [ + 217, + 601, + 785, + 797 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Emphasis on Robust Features for Localization. Several prior studies [24,26-31] in visual localization have discovered that not all regions within an image contribute", + "bbox": [ + 217, + 809, + 785, + 839 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "Reprojection Errors as Prompts for Efficient Scene Coordinate Regression", + "bbox": [ + 240, + 114, + 730, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "equally to localization accuracy. As a result, these approaches have attempted to distinguish robust and invariant features, found that prioritizing the recognition and comprehension of those regions can enhance visual localization tasks more efficiently and effectively. However, a common limitation of these methods is their reliance on pre-defined semantics or the necessity of a 3D model, which could hinder their generalizability and practicality in real-world applications.", + "bbox": [ + 212, + 146, + 787, + 238 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3 Preliminary of Scene Coordinate Regression (SCR)", + "text_level": 1, + "bbox": [ + 214, + 258, + 748, + 277 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "SCR-based visual localization determines the camera pose of an RGB image $I \\in \\mathbb{R}^{H \\times W \\times 3}$ by predicting 3D scene coordinates $\\mathcal{V}$ for a set of pixels and establishing the 2D-3D correspondences between pixel coordinates in the image and those predicted 3D scene coordinates. The primary objective of the training process is to establish a mapping function $f(\\cdot)$ such that $\\mathcal{V} = f(I)$ . The training data comprise RGB images paired with their respective camera poses, but do not necessarily include depth information. To generate accurate $\\mathcal{V}$ , SCR methods often involve optimizing the reprojection error $\\mathcal{R}$ , which is measured by projecting a 3D predicted scene coordinate $\\mathrm{y_i} \\in \\mathcal{V}$ back onto its 2D pixel coordinate and calculating the discrepancy from the actual coordinate of the image patch, where i denotes the patch index. The reprojection error for each patch $\\mathrm{r_i} \\in \\mathcal{R}$ can be formulated as $r(\\mathrm{p_i}, \\mathrm{y_i}, \\mathrm{h^*}) = \\| \\mathrm{p_i} - K\\mathrm{h^{*-1}}\\mathrm{y_i}\\|$ , where $K$ denotes the camera intrinsic matrix and $\\mathrm{h^*}$ denotes the ground truth camera pose. After obtaining $\\mathcal{V}$ , the camera pose h can be estimated according to the following equation:", + "bbox": [ + 212, + 290, + 789, + 503 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {h} = g (\\mathcal {C}) = g ^ {P n P} \\left(\\mathcal {C} _ {\\mathcal {I}}\\right), \\text {w i t h} \\mathcal {C} = \\left\\{\\left(\\mathrm {p} _ {\\mathrm {i}}, \\mathrm {y} _ {\\mathrm {i}}\\right) | \\mathrm {p} _ {\\mathrm {i}} \\in I, \\mathrm {y} _ {\\mathrm {i}} \\in \\mathcal {Y} \\right\\}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 303, + 511, + 785, + 530 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $g(\\cdot)$ involves PnP with RANSAC, followed by Levenberg-Marquardt-based refinement, $\\mathcal{C}$ represents the set of all 2D-3D correspondences and $\\mathcal{C}_{\\mathcal{I}}$ signifies inlier correspondences after refinement, which is a subset of $\\mathcal{C}$ . Inlier correspondences denote those that are better consistent with the estimated camera pose, while correspondences that do not align well with the estimated camera pose are termed outliers. The final camera pose estimation is determined by the inlier correspondences $\\mathcal{C}_{\\mathcal{I}}$ . In the following paragraphs, we introduce two scene coordinate regression methods that have demonstrated superior performance.", + "bbox": [ + 212, + 539, + 787, + 661 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "DSAC Variants. DSAC [13] introduces a differentiable variant of RANSAC, and enables end-to-end training of the entire pipeline. This design combines the benefits of RANSAC for handling outliers with the power of deep neural networks to learn complex patterns. In addition, DSAC++ [14] removes the requirement of depth information for SCR training and adopts robust loss terms to downweight outliers. DSAC* [15] further simplifies and refines DSAC++, and encourages the network to focus on reliable scene structures while ignoring outlier predictions.", + "bbox": [ + 212, + 672, + 787, + 782 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "ACE. ACE [17] is a fast SCR method that can achieve promising localization accuracy within a short training time of only five minutes. ACE splits its regression network into a pre-trained scene-agnostic backbone $f_{B}$ , and a scene-specific", + "bbox": [ + 214, + 794, + 787, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "T.-R. Liu et al.", + "bbox": [ + 271, + 114, + 375, + 128 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/5b6845a7066e395903dd28812392eed4dcb5061a06f9c0fc41e6f052db6cf123.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 233, + 143, + 498, + 267 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/020b608dc5d31389981574950ff0eb001465e1efe841fe1f00aa377e45ea2d5b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 501, + 143, + 767, + 266 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/f613d160fd3f192fc7bde8f6fc65622be02fecfd61f9fad868e0a16619c39ef8.jpg", + "image_caption": [ + "Fig. 2: Analysis between reprojection error and semantic meaning. The analysis result indicates the regions with low reprojection errors tend to have higher inlier ratios, while the errors do not always align with specific semantic categories, e.g., \"tree\" and \"rug\"." + ], + "image_footnote": [], + "bbox": [ + 233, + 268, + 498, + 391 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/6606dbd3678efa9350336af18d56201d63984931797c331948b4e48336e946f9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 501, + 268, + 767, + 391 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "multi-layer perceptron (MLP) $f_{H}$ . The backbone $f_{B}$ extracts feature vectors $\\mathbf{f}_i$ from image patches $p_i$ and the MLP $f_{H}$ predicts the scene coordinates $y_{i}$ based on the extracted $\\mathrm{f}_i$ . Between $f_{B}$ and $f_{H}$ , the features are shuffled to decorrelate gradients within a batch, which enables ACE to enhance its training efficiency.", + "bbox": [ + 212, + 479, + 787, + 542 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4 In-Depth Evaluation of Scene Coordinate Regression", + "text_level": 1, + "bbox": [ + 214, + 566, + 754, + 585 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1 Challenges in Scene Coordinate Regression", + "text_level": 1, + "bbox": [ + 214, + 601, + 617, + 617 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Several challenging factors have been identified that may compromise the accuracy of SCR. Two key challenges among them are dynamic objects and textureless surfaces [27, 32-34]. Dynamic objects exhibit characteristics of motion or disappearance across different frames, while textureless surfaces lack distinctive feature points or salient characteristics. These challenges highlight the complexity of achieving precise SCR in complex visual scenarios. To address them, previous SCR approaches [13-15, 17] have employed RANSAC to filter out outliers, and some researches [14, 15, 17] have attempted to utilize end-to-end learning to concentrate on reliable scene structures while down-weighting outlier predictions during training. Unfortunately, these prior endeavors failed to fully exploit the rich semantic information in the visual content, which can be crucial for enhancing the robustness and accuracy of camera pose estimation in complex scenes. Furthermore, RANSAC-based approaches for camera pose estimation could potentially fail in scenarios with insufficient number of inlier correspondences [35].", + "bbox": [ + 212, + 628, + 787, + 843 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "Reprojection Errors as Prompts for Efficient Scene Coordinate Regression", + "bbox": [ + 238, + 114, + 730, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/907825aa10e8fb607d365151a0f99a17944d3280758c178917bcbc4ec8b63e63.jpg", + "image_caption": [ + "Fig. 3: An overview of the training framework." + ], + "image_footnote": [], + "bbox": [ + 215, + 142, + 787, + 263 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2 Analysis between Reprojection Error and Semantic Meaning", + "text_level": 1, + "bbox": [ + 215, + 316, + 763, + 333 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For a more comprehensive understanding of the correlation and relationship between reprojection errors and semantic meaning, we investigate the performance of ACE on the reprojection errors and the inlier correspondences selected by RANSAC for estimating the final camera pose during the inference process. First, image pixels are labeled by a semantic segmentation model ViT-Adapter [36]. Subsequently, the reprojection errors and inlier correspondences of ACE for each semantic class are calculated. Fig. 2 presents the median reprojection errors and the ratios of inlier correspondences for the top ten most frequent classes in each scene. The left two figures correspond to the outdoor dataset Cambridge Landmarks [18], while the right two figures pertain to the indoor dataset Indoor6 [19]. It is observed that regions with low reprojection errors tend to have relatively higher inlier ratios, indicating a greater impact on the estimated camera poses. As a result, improving the accuracy of low-error areas could lead to more precise camera pose predictions. Nevertheless, it is also observed that the reprojection errors for the same class can vary significantly across different scenes. For example, in the Cambridge Landmarks dataset, the error for \"tree\" is low in King's College but high in Old Hospital. Similarly, in the Indoor6 dataset, the error for \"rug\" is high in scene1 but low in scene3. This indicates that low reprojection errors do not always align with specific semantic categories, which presents a challenge in directly using certain specific semantic segmentation labels for area selection. Motivated by these insights, the primary objective of this study is to develop a methodology that does not rely on pre-defined semantic categories to identify low reprojection error areas for enhancing localization accuracy in SCR.", + "bbox": [ + 212, + 340, + 787, + 690 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5 Methodology", + "text_level": 1, + "bbox": [ + 215, + 710, + 380, + 728 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1 Problem Definition and Framework Overview", + "text_level": 1, + "bbox": [ + 215, + 739, + 637, + 755 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Fig. 3 provides an overview of the proposed training framework for SCR, which iteratively samples features to train a scene-specific MLP. The scene-specific MLP consists of a scene coordinate head and a confidence head. In each iteration, the model is trained for $k$ epochs. During the initial iteration, features are randomly sampled from all parts of images in order to derive the first set", + "bbox": [ + 212, + 763, + 787, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "T.-R. Liu et al.", + "bbox": [ + 271, + 114, + 375, + 127 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "of reprojection errors. In subsequent iterations, features are selected based on error-guided feature selection (EGFS) masks generated according to reprojection errors and a confidence map. This iterative mechanism enables the model to dynamically update its focus areas throughout the training process. During the inference phase, the network estimates scene coordinates $\\mathcal{V}$ and a confidence map for the refinement, while the masks generated during training are not required for the inference phase. The inference procedure is depicted in Fig. 4.", + "bbox": [ + 212, + 146, + 787, + 253 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.2 Error-Guided Feature Selection (EGFS) with SAM", + "text_level": 1, + "bbox": [ + 214, + 277, + 683, + 292 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In alignment with the discussion in Section 4 on identifying potential areas contributing to the final camera pose estimation during the training stage, we employ the Segmentation Anything Model (SAM) [1], which is capable of leveraging diverse visual cues to generate high-quality object masks. Specifically, in this study, SAM is utilized to extend point prompts with low reprojection errors to form a complete mask sharing a similar semantic context. The rationale behind this approach is the high likelihood that these comprehensive semantic regions correspond to low error areas and can contribute to final pose estimation. Such a concept harnesses the capabilities of the foundational SAM model to uncover more viable areas, and hence, enabling a thorough understanding of the image context. In our experiments, we utilize EfficientViT-SAM-L0 [37], a lightweight and efficient variant of SAM, to identify salient regions by selecting the $\\tau$ percentage of points with the lowest reprojection errors as point prompts, where $\\tau$ is adjustable and set to ten in our case. After obtaining the error-guided mask from the SAM model, the predicted confidence map $c$ is utilized to refine the error-guided mask to ensure the points marked by the mask is sufficiently reliable. Specifically, in each image, the points with confidence scores below a threshold $\\sigma$ are filtered out. The design for confidence refinement and its optimization are described in Section 5.3. In each iteration, the refined error-guided masks serve as guidance for training the scene-specific MLP. Only the features corresponding to the regions selected by the masks are sampled into the training buffer. This selective inclusion is facilitated by the nature of $f_{H}$ [17], which employs $1\\times 1$ convolutions to treat and process each selected sample independently using the same set of shared weights. The proposed feature selection mechanism ensures that the training process focuses on the crucial regions identified in the current iteration, and therefore enables the model to enhance its overall performance.", + "bbox": [ + 212, + 304, + 787, + 698 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.3 Scene Coordinate and EGFS Refinement with Confidence", + "text_level": 1, + "bbox": [ + 214, + 720, + 738, + 736 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In order to further refine the error-guided mask and ensure the selected points are of sufficiently high quality for predicting accurate camera pose, we incorporate the design of a confidence map [38, 39]. Specifically, we predict a confidence score pixel-wise from the confidence head and optimize it jointly with the reprojection error. This confidence head replaces the last three $1 \\times 1$ convolution layers of the original scene-specific MLP with two $1 \\times 1$ convolution layers for", + "bbox": [ + 212, + 750, + 787, + 840 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "Reprojection Errors as Prompts for Efficient Scene Coordinate Regression", + "bbox": [ + 238, + 114, + 730, + 130 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ca54741ca98085209b7b2368001f7d1a5da12a6c2833602c004768ea58b3ea06.jpg", + "image_caption": [ + "Fig. 4: An overview of the inference procedure." + ], + "image_footnote": [], + "bbox": [ + 300, + 143, + 702, + 247 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "confidence score prediction, thus enabling adaptive weights for the reprojection loss at each position. The training loss used by our method is formulated as:", + "bbox": [ + 212, + 297, + 782, + 328 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\ell \\left(\\mathrm {p} _ {\\mathrm {i}}, \\mathrm {y} _ {\\mathrm {i}}, \\mathrm {h} _ {\\mathrm {i}} ^ {*}\\right) = \\left\\{ \\begin{array}{l l} \\mathrm {c} _ {\\mathrm {i}} \\cdot \\hat {r} \\left(\\mathrm {p} _ {\\mathrm {i}}, \\mathrm {y} _ {\\mathrm {i}}, \\mathrm {h} _ {\\mathrm {i}} ^ {*}\\right) - \\alpha \\log \\mathrm {c} _ {\\mathrm {i}} & \\text {i f y} _ {\\mathrm {i}} \\in \\mathcal {V} \\\\ \\| \\mathrm {y} _ {\\mathrm {i}} - \\bar {\\mathrm {y}} _ {\\mathrm {i}} \\| _ {0} - \\alpha \\log (1 - \\mathrm {c} _ {\\mathrm {i}}) & \\text {o t h e r w i s e} \\end{array} , \\right. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 305, + 333, + 785, + 375 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $\\hat{r} (\\mathrm{p_i},\\mathrm{y_i},\\mathrm{h_i^*})$ represents a tanh clamping of the reprojection error that changes over time, $\\mathcal{V}$ denotes the set of 2D pixels with valid scene coordinate predictions, as explained in [15, 17], $c_{i}$ is the confidence score at pixel i, $\\alpha$ is a hyperparameter that balances the confidence regularization term, and $\\bar{\\mathbf{y}}_{\\mathrm{i}}$ is a dummy scene coordinate derived from the ground truth camera pose, assuming a constant image depth of $10\\mathrm{m}$ . When the scene coordinate prediction is valid, the model down-weights the reprojection errors in challenging regions using confidence scores and focuses more on regions with reliable predictions. On the other hand, when the scene coordinate prediction is invalid, the model is incentivized to lower the confidence scores for inaccurate predictions. The confidence score is also used in the inference phase to select reliable scene coordinates. For each query image, the median of the confidence scores is calculated, and only scene coordinates with confidence greater than this median are selected. This thresholding ensures that only the most reliable 2D-3D correspondences are used to generate the camera pose by the PnP with RANSAC pose solver. By focusing on high-confidence correspondences, a more accurate camera pose can be estimated.", + "bbox": [ + 212, + 380, + 787, + 623 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6 Experimental Results", + "text_level": 1, + "bbox": [ + 215, + 642, + 460, + 660 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6.1 Experimental Setups", + "text_level": 1, + "bbox": [ + 215, + 670, + 434, + 686 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Datasets. Our evaluation of the proposed method was conducted on two representative datasets: the outdoor dataset Cambridge Landmarks [18] and the indoor dataset Indoor6 [19]. The details of the datasets are described as follows:", + "bbox": [ + 212, + 691, + 784, + 739 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Cambridge Landmarks. The Cambridge Landmarks Dataset [18] is a renowned outdoor dataset extensively utilized for visual localization tasks, which includes five distinct scenes. Every scene in this dataset is composed of RGB images along with their respective ground truth camera poses reconstructed using the SfM technique. It includes a broad spectrum of conditions and various viewpoints, and is ideal for evaluating the resilience and effectiveness of various SCR methods.", + "bbox": [ + 212, + 748, + 787, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "T.-R. Liu et al.", + "bbox": [ + 271, + 114, + 375, + 127 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/f81fed9dc585897b3db20ba12e8f4f43ec2b4d1907206b9c88bcc9b70f462cad.jpg", + "table_caption": [ + "Table 1: A comparison of model sizes, training times, and median translation and rotation errors (in $\\mathrm{cm} / {}^{\\circ}$ ) evaluated on the Cambridge Landmarks dataset. 'dual' and 'quad.' denote the ensemble versions of two and four models, respectively." + ], + "table_footnote": [], + "table_body": "
MethodSizeMapping TimeKing's College Great Court Old Hospital Shop Facade St Mary's ChurchAverage
FMhLoc (SP+SG) [4,6,8]~800MB~35 min12/0.216/0.115/0.34/0.27/0.211/0.2
pixLoc [26]~600MB30/0.114/0.216/0.35/0.210/0.315/0.2
SCR (w/ 3D model)DSAC* [15]28MB15 hr15/0.349/0.321/0.45/0.313/0.421/0.3
FocusTune [31]4MB6 min19/0.338/0.118/0.46/0.315/0.519/0.3
FocusTune (quad.) [31]16MB24 min15/0.329/0.117/0.45/0.29/0.315/0.3
NeuMap [40]--14/0.26/0.119/0.36/0.217/0.512/0.3
SACReg-L [41]--11/0.213/0.113/0.26/0.25/0.310/0.2
SCRDSAC* [15]28MB15 hr18/0.334/0.221/0.45/0.315/0.619/0.4
ACE [17]4MB5 min28/0.443/0.231/0.65/0.318/0.625/0.4
ACE (quad.) [17]16MB20 min18/0.328/0.125/0.55/0.39/0.317/0.3
EGFS4.5MB12 min14/0.331/0.121/0.45/0.315/0.517/0.3
EGFS (dual)9MB21 min14/0.328/0.119/0.45/0.210/0.315/0.3
", + "bbox": [ + 218, + 198, + 785, + 330 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/f289044e04b916e259c0c1456ee4cdea2a70a38057fcae22a37da31f29416728.jpg", + "table_caption": [ + "Table 2: A comparison of model sizes, training times, and the proportions of translation and rotation errors that are below $5\\mathrm{cm} / 5^{\\circ}$ evaluated on the Indoor6 dataset." + ], + "table_footnote": [], + "table_body": "
MethodSizeMapping Timescene1scene2ascene3scene4ascene5scene6Average
FMhLoc [4]~1.5GB~3.3 hr70.5%52.1%86.0%75.3%58.0%86.7%71.4%
SCR (w/ 3D model)DSAC* [15]28MB15 hr18.7%28.0%19.7%60.8%10.6%44.3%30.4%
SLD* (300 landmarks) [25]15MB~5.5 hr47.2%48.2%56.2%67.7%33.7%52.0%50.8%
SLD* (1000 landmarks) [25]120MB~44 hr68.5%62.6%76.2%77.2%57.8%78.0%70.1%
SCRDSAC* [15]28MB15 hr23.0%33.9%26.0%67.1%10.6%50.2%35.1%
ACE [17]4MB5 min26.0%32.3%31.4%62.0%14.2%47.4%35.6%
ACE (quad.) [17]16MB20 min52.9%52.5%62.9%69.6%31.1%82.4%58.6%
EGFS4.5MB21 min46.4%60.6%56.4%78.7%22.8%71.6%56.1%
EGFS (dual)9MB30 min58.5%59.1%67.0%76.1%30.6%75.9%61.2%
", + "bbox": [ + 220, + 383, + 785, + 507 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Indoor6. The Indoor6 [19] dataset comprises six distinct indoor scenes captured over several days, with ground truth camera poses computed using COLMAP [42]. Each scene includes multiple rooms and contains illumination variations, making it challenging for different types of visual localization tasks.", + "bbox": [ + 215, + 534, + 787, + 595 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Implementation Details. The training process of our methodology includes 20 epochs, with the generation of masks every five epochs. We set the confidence regularization parameter $\\alpha$ to ten and the reprojection error threshold $\\tau$ to ten. For mask refinement, the parameter $\\sigma$ is set to the median confidence score of each image. Our method builds upon the ACE [17] architecture as the backbone, and thus, the remaining hyperparameters are kept the same as those used in [17].", + "bbox": [ + 215, + 614, + 787, + 708 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "6.2 Results on the Cambridge Landmarks and Indoor6 Dataset", + "text_level": 1, + "bbox": [ + 215, + 726, + 750, + 742 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Cambridge Landmarks Dataset. Table 1 presents the experimental results evaluated on the Cambridge Landmarks dataset, which affirm the effectiveness and efficiency of our proposed methodology. It can be observed that EGFS maintains a model size similar to ACE, with only a 0.5MB increase for the confidence head, while reducing average translational and rotational errors. The results further reveal that EGFS not only surpasses DSAC* in performance and performs", + "bbox": [ + 215, + 750, + 787, + 840 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "Reprojection Errors as Prompts for Efficient Scene Coordinate Regression", + "bbox": [ + 238, + 114, + 730, + 130 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/c0fd9b1afb6c3f6140384331b3e6d0a852d42ce7d1205ae890d91ae97097bfbe.jpg", + "table_caption": [ + "Table 3: Comparison of the effectiveness between (a) directly sampling solely based on pixel points with low reprojection errors, and (b) sampling through the EGFS masks." + ], + "table_footnote": [], + "table_body": "
DatasetSceneri < Q0.3(R) ri < Q0.4(R) ri < Q0.5(R) ri < Q0.6(R) ri < Q0.7(R)EGFS Masks
CambridgeKing's College20/0.319/0.320/0.320/0.320/0.314/0.3
Great Court39/0.237/0.136/0.238/0.239/0.231/0.1
Old Hospital26/0.525/0.523/0.525/0.524/0.521/0.4
Shop Facade6/0.36/0.36/0.36/0.35/0.35/0.3
St Mary's Church17/0.517/0.516/0.517/0.516/0.515/0.5
Average22/0.421/0.420/0.421/0.421/0.417/0.3
Indoor6scene134.4%34.9%34.9%32.0%30.2%46.4%
scene2a39.8%40.6%46.1%41.3%42.5%60.6%
scene345.2%43.9%44.9%41.3%42.3%56.4%
scene4a63.2%63.9%62.6%63.2%67.1%78.7%
scene519.5%18.1%17.1%18.3%16.9%22.8%
scene659.7%60.3%60.0%57.2%58.1%71.6%
Average43.6%43.6%44.3%42.2%42.9%56.1%
", + "bbox": [ + 232, + 184, + 772, + 353 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "on par with the ensemble version of ACE (i.e., ACE (quad.)), but also features a smaller model size and reduced training time. Furthermore, our model's ensemble version (i.e., EGFS (dual.)) is able to exceed the performance of DSAC* and ACE (quad.) while requiring a smaller model size and less training time.", + "bbox": [ + 212, + 380, + 787, + 444 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Indoor6 Dataset. Table 2 presents the evaluation results on Indoor6. EGFS significantly outperforms DSAC* and ACE across all scenes, while also achieving comparable performance to ACE (quad.) with a much smaller model size (4.5MB compared to 16MB). Moreover, EGFS (dual.) demonstrates superior performance relative to ACE (quad.), which further substantiates the effectiveness of EGFS in optimization through error-guided masks and confidence refinement.", + "bbox": [ + 212, + 467, + 789, + 560 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "6.3 Effectiveness of Error-Guided Feature Selection", + "text_level": 1, + "bbox": [ + 214, + 584, + 653, + 599 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "To further substantiate the advantages of employing the EGFS approach, this experiment compares several baseline schemes that sample points with different reprojection error thresholds against EGFS. The results are summarized in Table 3. Specifically, these baseline schemes select sample points based on various quantiles of the reprojection error map, ranging from $30\\%$ to $70\\%$ , denoted as $Q_{0.3}(R)$ through $Q_{0.7}(R)$ , respectively. In each baseline scheme, a point is chosen if its reprojection error $r$ falls below the threshold specified by $Q_{\\text{quantile}}(R)$ . The results suggest that relying solely on reprojection errors does not guarantee optimal learning of scene coordinates. This is due to the fact that areas of low reprojection errors can be scattered and may not encompass entire semantic objects. In contrast, the proposed EGFS expands the points into complete semantic masks followed by confidence refinement, which enables our model to determine scene coordinates with minimal transitional and rotational errors. This experiment thus confirms the effectiveness of leveraging SAM's ability to interpret image context alongside confidence maps for enhanced sample point selection.", + "bbox": [ + 212, + 613, + 787, + 840 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "T.-R. Liu et al.", + "bbox": [ + 271, + 114, + 375, + 128 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/cdd7b19eca1bfb808f3163421798963e98e839761a8fc757fb66044a3767915a.jpg", + "image_caption": [ + "Fig. 5: Visualization of the EGFS mask refinement process at every five epochs, which depicts the reprojection errors at the beginning (epoch 5) and the end (epoch 20), as well as the refined error-guided masks used throughout training. The red dots represent low reprojection errors that serve as prompts, while the light green overlay denotes the refined EGFS masks. It can be observed that the EGFS masks enhances over epochs." + ], + "image_footnote": [], + "bbox": [ + 243, + 143, + 756, + 311 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/f5e2853b2307b1a8e4bf9ebd579a2e93d56e6ab2d967d8d51020e89a3c472809.jpg", + "image_caption": [ + "Fig. 6: Visualization of the estimated camera pose trajectories from testing sequences with the camera frustums colored based on translational errors. Pose errors denoted." + ], + "image_footnote": [], + "bbox": [ + 256, + 412, + 759, + 579 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "6.4 Qualitative Results", + "text_level": 1, + "bbox": [ + 214, + 657, + 419, + 672 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Iterative Error-guided Feature Selection Mask. Fig. 5 illustrates the refinement of EGFS masks every five epochs and compares the reprojection errors at the beginning and the end. It can be observed that EGFS is not only capable of expanding points with low reprojection errors (i.e., $\\mathrm{r} < Q_{0.1}(R)$ ) into regions that bear similar semantic meanings, but also refines the masks with a confidence map to exclude uncertain areas. The masks are observed to be iteratively enhanced from coarse to fine. In the beginning (i.e., at epoch five), the reprojection errors appear noisy, whereas by the end (i.e., at epoch 20), the EGFS masks become more refined and concentrated on key regions beneficial for SCR.", + "bbox": [ + 212, + 703, + 787, + 840 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "Reprojection Errors as Prompts for Efficient Scene Coordinate Regression", + "bbox": [ + 240, + 114, + 730, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/bbeac4a9f655ceae17f344e5c6b90cfeb9afe3b17e0ce04d08638b8591f46760.jpg", + "image_caption": [ + "Fig. 7: Visualization of point clouds reconstructed from estimated scene coordinates of the training sequence. The point clouds derived from the scene coordinates estimated by the model trained with the proposed EGFS are clearer compared to those from the ACE model, even without the application of refined EGFS masks at inference time." + ], + "image_footnote": [], + "bbox": [ + 246, + 146, + 748, + 378 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Visualization of the Estimated Camera Poses. In this section, we present the estimated camera poses from the testing sequences, and compare the proposed EGFS with the ACE baseline. The results are illustrated in Fig. 6. To visualize camera trajectories, we connect consecutive camera positions with lines and indicate camera orientations with frustums. Each sampled camera pose is associated with a color to represent the translation error. It is observed that our proposed EGFS can estimate more accurate camera poses as compared to ACE.", + "bbox": [ + 214, + 482, + 787, + 590 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Scene Coordinates and 3D Point Clouds w/ and w/o EFGS. To validate the effectiveness of EGFS qualitatively and its impact on the quality of estimated scene coordinates that learned from training sequence, we depict the 3D point cloud generated from the scene coordinates from the training sequence, and apply colors based on the corresponding queried pixels in Fig. 7. To compare the 3D point clouds reconstructed from the scene coordinates, we compare the baseline ACE approach with the proposed EGFS approach. For our approach, we visualize the point clouds both before and after the application of EGFS masks and confidence maps when collecting the scene coordinates. Please note that both before and after versions which shown in the figure are trained with EGFS refinement. The 3D point cloud constructed by the proposed approach appears clearer, in contrast to the ACE-generated point cloud, which exhibits blurriness due to inaccurate scene coordinate estimation. In addition, the point cloud reconstructed from raw scene coordinates predicted without EGFS from our approach contains noise and floaters due to the inclusion of scene coordinates", + "bbox": [ + 214, + 613, + 787, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "T.-R. Liu et al.", + "bbox": [ + 271, + 114, + 375, + 127 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/2e91c655bb39e7fbc02620fda7ca4f6f5a8bd629fd83d7ebe782d25be6719296.jpg", + "image_caption": [ + "Fig. 8: Visualization of the estimated scene coordinates and the confidence maps on unseen samples. The RGB color in the scene coordinates represents the XYZ coordinates. The point clouds are colored green for selected areas and red for rejected areas." + ], + "image_footnote": [], + "bbox": [ + 248, + 145, + 751, + 362 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "with low confidence scores. On the other hand, the point cloud reconstructed from the scene coordinates with EGFS refinement exhibits clearer contours and less noise. This justifies EGFS's ability to eliminate areas of low confidence values and produce high-quality dense scene coordinates.", + "bbox": [ + 212, + 445, + 787, + 506 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Visualization of Estimated Confidence Map on Unseen Samples. In this section, we qualitatively evaluate the confidence maps estimated from the testing sequence by visualizing them in both the 2D image plane and as 3D projections onto the point cloud. The visualized point cloud is reconstructed from the training sequence, whereas the plotted samples are selected from the testing sequence and were not included in the training procedure. These visualizations are provided in Fig. 8. It can be observed that the estimated confidence map assists in rejecting scene coordinates located in areas where accurate prediction is challenging, such as in the air or on texture-less surfaces.", + "bbox": [ + 212, + 530, + 787, + 667 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "6.5 Ablation Study", + "text_level": 1, + "bbox": [ + 214, + 691, + 390, + 707 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Error-Guided Feature Selection (EGFS) and Confidence Refinement. We first present an ablation analysis to validate the effectiveness of the EGFS masks and confidence refinement in enhancing our method's performance. Table 4 reports the average results on the Cambridge Landmarks and Indoor6 datasets. The results demonstrate that the model trained with the proposed EGFS mechanism outperforms those trained without it, which indicates that focusing on training with robust features leads to improved performance. Furthermore, the incorporation of the confidence map also enhances the overall", + "bbox": [ + 212, + 718, + 787, + 840 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "Reprojection Errors as Prompts for Efficient Scene Coordinate Regression", + "bbox": [ + 240, + 114, + 730, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 784, + 126 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/94d6a9c992b71ded29d6724838fc6f1dc2efef7b1389e6316ca305bd5b08622e.jpg", + "table_caption": [ + "Table 4: Ablation study on the impact of each proposed components." + ], + "table_footnote": [], + "table_body": "
Error-Guided MasksConfidence RefinementCambridge (cm/°)Indoor6 (%)
XX25/0.435.6
✓X19/0.341.6
X✓19/0.450.9
✓✓17/0.356.1
", + "bbox": [ + 220, + 184, + 522, + 280 + ], + "page_idx": 13 + }, + { + "type": "table", + "img_path": "images/9e398625b170aa06b86272d7f6d6f3cfe5f3caa56e979858ebf7a208e86d0bb4.jpg", + "table_caption": [ + "Table 5: Impact of varying proportions of point prompts $(\\tau)$" + ], + "table_footnote": [], + "table_body": "
Proportions of point prompts (τ)Cambridge (cm/°)Indoor6 (%)
517/0.354.3
1017/0.356.1
1517/0.355.4
2017/0.355.3
", + "bbox": [ + 542, + 184, + 787, + 280 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "performance. The final row of Table 4 demonstrates that combining both techniques (i.e., our EGFS) achieves the best results, which further confirms the effectiveness of EGFS in improving localization accuracy.", + "bbox": [ + 212, + 304, + 785, + 351 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Analysis of Proportions of Point Prompts. We evaluate the performance of different proportions of point prompts used for expanding into EGFS masks, and the results are reported in Table 5. It can be observed that the proportions of point prompts do not significantly impact the performance on the Cambridge Landmark dataset. This may be attributed to the fact that the scenes in the dataset typically feature a primary architectural structure with fewer complicated details in the surroundings. On the other hand, for the Indoor6 dataset, the proportions of point prompts significantly affect the performance. The rationale behind these observations is that the indoor scenes are more complicated, and feature various small items that necessitate a larger number of low-error point prompts to comprehensively capture the entire scene structure. Please note that we employ $\\tau = 10$ as our setting for all the experiments presented.", + "bbox": [ + 212, + 369, + 787, + 551 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "7 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 569, + 359, + 585 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "This paper addressed key challenges in SCR for visual localization, and specifically focused on the impacts of dynamic objects and texture-less regions. Our approach introduced an innovative EGFS mechanism through the use of SAM and confidence maps to enhance the performance of SCR. This technique effectively filtered out problematic areas by concentrating on regions with low reprojection errors. Moreover, we used confidence maps to further refine the pixels selected by EGFS and perform this process iteratively to enable dynamic updates of the focused regions. The experimental results on the Cambridge Landmarks and Indoor6 datasets suggest that our method can provide improvements in terms of training efficiency, model size, and accuracy. Our findings highlighted the importance of carefully selecting low-reprojection error pixels by taking into account semantic information and confidence scores. Furthermore, we quantitatively and qualitatively demonstrated that dynamically updating the masks enables more robust selection of error points, thus enabling better training efficiency and effectiveness. Our ablation studies validated the effectiveness of the techniques adopted in our method, and solidified their roles in enhancing the performance.", + "bbox": [ + 212, + 597, + 787, + 840 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "T.-R. Liu et al.", + "bbox": [ + 271, + 114, + 375, + 127 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 215, + 143, + 401, + 162 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "The authors gratefully acknowledge the support from the National Science and Technology Council (NSTC) in Taiwan under grant numbers MOST 111-2223-E007-004-MY3, 113-2221-E-007-122-MY3, 113-2640-E-002-003, 113-2221-E-007-104-MY3, as well as the financial support from Woven by Toyota, Inc., Japan. The authors would also like to express their appreciation for the donation of the GPUs from NVIDIA Corporation and NVIDIA AI Technology Center (NVAITC) used in this work. Furthermore, the authors extend their gratitude to the National Center for High-Performance Computing (NCHC) for providing the necessary computational and storage resources.", + "bbox": [ + 212, + 174, + 787, + 310 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 215, + 330, + 323, + 347 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Dlarr, and Ross Girshick. Segment anything. arXiv:2304.02643, 2023.", + "2. Xiao-Shan Gao, Xiao-Rong Hou, Jianliang Tang, and Hang-Fei Cheng. Complete solution classification for the perspective-three-point problem. IEEE transactions on pattern analysis and machine intelligence, 25(8):930-943, 2003.", + "3. Martin A Fischler and Robert C Bolles. Random sample consensus: a paradigm for model fitting with applications to image analysis and automated cartography. Communications of the ACM, 1981.", + "4. Paul-Edouard Sarlin, Cesar Cadena, Roland Siegwart, and Marcin Dymczyk. From coarse to fine: Robust hierarchical localization at large scale. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2019.", + "5. Jamie Shotton, Ben Glocker, Christopher Zach, Shahram Izadi, Antonio Criminisi, and Andrew Fitzgibbon. Scene coordinate regression forests for camera relocalization in RGB-D images. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2013.", + "6. Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Superpoint: Self-supervised interest point detection and description. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), pages 224-236, 2018.", + "7. Jiaming Sun, Zehong Shen, Yuang Wang, Hujun Bao, and Xiaowei Zhou. LoFTR: Detector-free local feature matching with transformers. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), pages 8922-8931, 2021.", + "8. Paul-Edouard Sarlin, Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. SuperGlue: Learning feature matching with graph neural networks. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2020.", + "9. Philipp Lindenberger, Paul-Edouard Sarlin, and Marc Pollefeys. LightGlue: Local feature matching at light speed. arXiv preprint arXiv:2306.13643, 2023.", + "10. Pablo Speciale, Johannes L Schonberger, Sing Bing Kang, Sudipta N Sinha, and Marc Pollefeys. Privacy preserving image-based localization. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), pages 5493-5503, 2019.", + "11. Xiaotian Li, Juha Ylioinas, Jakob Verbeek, and Juho Kannala. Scene coordinate regression with angle-based reprojection loss for camera relocalization. In Proc. European Conf. on Computer Vision Workshop (ECCVW), 2018.", + "12. Xiaotian Li, Juha Ylioinas, and Juho Kannala. Full-frame scene coordinate regression for image-based localization. RSS, 2018." + ], + "bbox": [ + 217, + 359, + 785, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "Reprojection Errors as Prompts for Efficient Scene Coordinate Regression", + "bbox": [ + 238, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "13. Eric Brachmann, Alexander Krull, Sebastian Nowozin, Jamie Shotton, Frank Michel, Stefan Gumhold, and Carsten Rother. DSAC-Differentiable RANSAC for camera localization. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2017.", + "14. Eric Brachmann and Carsten Rother. Learning less is more - 6D camera localization via 3D surface regression. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2018.", + "15. Eric Brachmann and Carsten Rother. Visual camera re-localization from RGB and RGB-D images using DSAC. TPAMI, 2021.", + "16. Eric Brachmann and Carsten Rother. Expert sample consensus applied to camera re-localization. In Proc. IEEE Int. Conf. on Computer Vision (ICCV), 2019.", + "17. Eric Brachmann, Tommaso Cavallari, and Victor Adrian Prisacariu. Accelerated coordinate encoding: Learning to relocalize in minutes using RGB and poses. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2023.", + "18. Alex Kendall, Matthew Grimes, and Roberto Cipolla. PoseNet: A convolutional network for real-time 6-DOF camera relocalization. In Proc. IEEE Int. Conf. on Computer Vision (ICCV), December 2015.", + "19. Tien Do, Ondrej Miksik, Joseph DeGol, Hyun Soo Park, and Sudipta N. Sinha. Learning to detect scene landmarks for camera localization. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2022.", + "20. Abner Guzman-Rivera, Pushmeet Kohli, Ben Glocker, Jamie Shotton, Toby Sharp, Andrew Fitzgibbon, and Shahram Izadi. Multi-output learning for camera localization. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2014.", + "21. Eric Brachmann, Frank Michel, Alexander Krull, Michael Ying Yang, Stefan Gumhold, and carsten Rother. Uncertainty-driven 6D pose estimation of objects and scenes from a single RGB image. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2016.", + "22. Julien Valentin, Matthias Niessner, Jamie Shotton, Andrew Fitzgibbon, Shahram Izadi, and Philip H. S. Torr. Exploiting uncertainty in regression forests for accurate camera relocalization. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2015.", + "23. Tao Xie, Kun Dai, Siyi Lu, Ke Wang, Zhiqiang Jiang, Jinghan Gao, Dedong Liu, Jie Xu, Lijun Zhao, and Ruifeng Li. OFVL-MS: Once for visual localization across multiple indoor scenes. In Proc. IEEE Int. Conf. on Computer Vision (ICCV), 2023.", + "24. Zhaoyang Huang, Han Zhou, Yijin Li, Bangbang Yang, Yan Xu, Xiaowei Zhou, Hujun Bao, Guofeng Zhang, and Hongsheng Li. VS-Net: Voting with segmentation for visual localization. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2021.", + "25. Tien Do and Sudipta N. Sinha. Improved scene landmark detection for camera localization. In Proceedings of the International Conference on 3D Vision (3DV), 2024.", + "26. Paul-Edouard Sarlin, Ajaykumar Unagar, Mans Larsson, Hugo Germain, Carl Toft, Victor Larsson, Marc Pollefeys, Vincent Lepetit, Lars Hammarstrand, Fredrik Kahl, and Torsten Sattler. Back to the Feature: Learning Robust Camera Localization from Pixels to Pose. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2021.", + "27. Mohammad Altillawi. PixSelect: Less but reliable pixels for accurate and efficient localization. In Proc. IEEE Int. Conf. on Robotics and Automation (ICRA), pages 4156-4162. IEEE, 2022." + ], + "bbox": [ + 215, + 146, + 785, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "T.-R. Liu et al.", + "bbox": [ + 271, + 114, + 374, + 126 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "28. Fei Xue, Ignas Budvytis, Daniel Olmeda Reino, and Roberto Cipolla. Efficient large-scale localization by global instance recognition. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), pages 17348-17357, 2022.", + "29. Fei Xue, Ignas Budvytis, and Roberto Cipolla. SFD2: Semantic-guided feature detection and description. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), pages 5206-5216, 2023.", + "30. Bach-Thuan Bui, Dinh-Tuan Tran, and Joo-Ho Lee. D2S: Representing local descriptors and global scene coordinates for camera relocalization. arXiv preprint arXiv:2307.15250, 2023.", + "31. Son Tung Nguyen, Alejandro Fontan, Michael Milford, and Tobias Fischer. Focus-Tune: Tuning visual localization through focus-guided sampling. In Proc. IEEE Winter Conf. on Applications of Computer Vision (WACV), pages 3606-3615, January 2024.", + "32. Johanna Wald, Torsten Sattler, Stuart Golodetz, Tommaso Cavallari, and Federico Tombari. Beyond controlled environments: 3d camera re-localization in changing indoor scenes. In Proc. European Conf. on Computer Vision (ECCV), pages 467-487, 2020.", + "33. Siyan Dong, Qingnan Fan, He Wang, Ji Shi, Li Yi, Thomas Funkhouser, Baoquan Chen, and Leonidas J Guibas. Robust neural routing through space partitions for camera relocalization in dynamic indoor environments. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2021.", + "34. Donghwan Lee, Soohyun Ryu, Suyong Yeon, Yonghan Lee, Deokhwa Kim, Cheolho Han, Yohann Cabon, Philippe Weinzaepfel, Nicolas Guerin, Gabriela Csurka, and Martin Humenberger. Large-scale localization datasets in crowded indoor spaces. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), pages 3227-3236, June 2021.", + "35. Hongyi Fan, Joe Kileel, and Benjamin Kimia. On the instability of relative pose estimation and RANSAC's role. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), pages 8935-8943, June 2022.", + "36. Zhe Chen, Yuchen Duan, Wenhai Wang, Junjun He, Tong Lu, Jifeng Dai, and Yu Qiao. Vision transformer adapter for dense predictions. arXiv preprint arXiv:2205.08534, 2022.", + "37. Han Cai, Junyan Li, Muyan Hu, Chuang Gan, and Song Han. EfficientViT: Lightweight multi-scale attention for high-resolution dense prediction. In Proc. IEEE Int. Conf. on Computer Vision (ICCV), pages 17302-17313, 2023.", + "38. Alex Kendall and Yarin Gal. What uncertainties do we need in bayesian deep learning for computer vision? Advances in neural information processing systems, 30, 2017.", + "39. Sheng Wan, Tung-Yu Wu, Wing H Wong, and Chen-Yi Lee. Confnet: predict with confidence. In 2018 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 2921-2925. IEEE, 2018.", + "40. Shitao Tang, Sicong Tang, Andrea Tagliasacchi, Ping Tan, and Yasutaka Furukawa. Neumap: Neural coordinate mapping by auto-transdecoder for camera localization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 929-939, June 2023.", + "41. Revaud, Jérôme and Cabon, Yohann and Brégier, Romain and Lee, JongMin and Weinzaepfel, Philippe. SACReg: Scene-Agnostic Coordinate Regression for Visual Localization, 2023.", + "42. Johannes L. Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), June 2016." + ], + "bbox": [ + 215, + 146, + 787, + 838 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "Reprojection Errors as Prompts for Efficient Scene Coordinate Regression", + "bbox": [ + 238, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 16 + } +] \ No newline at end of file diff --git a/2024/Reprojection Errors as Prompts for Efficient Scene Coordinate Regression/91ff6348-fcc0-42f1-b6d0-08cddca0de4a_model.json b/2024/Reprojection Errors as Prompts for Efficient Scene Coordinate Regression/91ff6348-fcc0-42f1-b6d0-08cddca0de4a_model.json new file mode 100644 index 0000000000000000000000000000000000000000..687fcd149817a552b5557f97d4f1c5ecba6fe3e3 --- /dev/null +++ b/2024/Reprojection Errors as Prompts for Efficient Scene Coordinate Regression/91ff6348-fcc0-42f1-b6d0-08cddca0de4a_model.json @@ -0,0 +1,1983 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.249, + 0.141, + 0.756, + 0.187 + ], + "angle": 0, + "content": "Reprojection Errors as Prompts for Efficient Scene Coordinate Regression" + }, + { + "type": "text", + "bbox": [ + 0.235, + 0.212, + 0.768, + 0.244 + ], + "angle": 0, + "content": "Ting-Ru Liu\\(^{1}\\), Hsuan-Kung Yang\\(^{2}\\), Jou-Min Liu\\(^{1*}\\), Chun-Wei Huang\\(^{1*}\\), Tsung-Chih Chiang\\(^{1*}\\), Quan Kong\\(^{2}\\), Norimasa Kobori\\(^{2}\\), and Chun-Yi Lee\\(^{1}\\)" + }, + { + "type": "text", + "bbox": [ + 0.246, + 0.255, + 0.758, + 0.313 + ], + "angle": 0, + "content": "\\(^{1}\\)National Tsing Hua University \\(^{2}\\)Woven by Toyota, Inc. {tingru, diasimui, alan, george, cymaxwelllee}@elsa.cs.nthu.edu.tw {hsuan-kung.yang, quan.kong, norimasa.kobori} @woven toyota https://tingru0203.github.io/egfs" + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.351, + 0.744, + 0.56 + ], + "angle": 0, + "content": "Abstract. Scene coordinate regression (SCR) methods have emerged as a promising area of research due to their potential for accurate visual localization. However, many existing SCR approaches train on samples from all image regions, including dynamic objects and texture-less areas. Utilizing these areas for optimization during training can potentially hamper the overall performance and efficiency of the model. In this study, we first perform an in-depth analysis to validate the adverse impacts of these areas. Drawing inspiration from our analysis, we then introduce an error-guided feature selection (EGFS) mechanism, in tandem with the use of the Segment Anything Model (SAM) [1]. This mechanism seeds low reprojection areas as prompts and expands them into error-guided masks, and then utilizes these masks to sample points and filter out problematic areas in an iterative manner. The experiments demonstrate that our method outperforms existing SCR approaches that do not rely on 3D information on the Cambridge Landmarks and Indoor6 datasets." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.586, + 0.377, + 0.601 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.618, + 0.788, + 0.816 + ], + "angle": 0, + "content": "The objective of visual localization is to estimate a 6-DoF camera pose from images, a key component in fields such as Augmented Reality, Virtual Reality, and autonomous driving. Contemporary leading methods in visual localization typically involve establishing 2D-3D correspondences and then utilizing Perspective-Point (PnP) [2] with RANSAC [3] for camera pose estimation. These methods can be broadly classified into two main directions: feature-matching [4] and scene coordinate regression (SCR) [5]. Feature-matching approaches reconstruct a 3D scene using Structure from Motion (SfM), identify and describe key points in 2D images [6,7], and link these to 3D coordinates [8,9]. Nevertheless, they may encounter challenges such as high computational demands, significant storage requirements, and potential privacy concerns [10]. On the other hand, SCR methods [11-17] employ deep neural networks (DNNs) to predict the 3D coordinates of pixels and then utilize PnP with RANSAC for camera pose estimation, which" + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.825, + 0.424, + 0.84 + ], + "angle": 0, + "content": "* indicates equal contribution." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.377, + 0.128 + ], + "angle": 0, + "content": "T.-R. Liu et al." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.18, + 0.352, + 0.252 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.353, + 0.179, + 0.485, + 0.252 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.485, + 0.179, + 0.627, + 0.252 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.63, + 0.18, + 0.787, + 0.252 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.261, + 0.627, + 0.334 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.631, + 0.262, + 0.787, + 0.334 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.374, + 0.788, + 0.501 + ], + "angle": 0, + "content": "Fig. 1: Visualization of the primary components (i.e., (d)-(h)) introduced in the proposed visual localization scheme. (d) illustrates the point prompts selected from (b) with low reprojection errors, while (e) presents an error-guided mask expanded from the prompted points in (d) using SAM. (f) displays the proposed error-guided feature selection (EGFS), which refines the mask from (e) with the predicted confidence map (c) to ensure high-quality scene coordinates are sampled for estimating the final camera pose. The point cloud constructed from the predicted scene coordinates is shown on the right-hand side (i.e., (g)-(h)), with the confidence (yellow parts) and the refined EGFS mask (green for selected areas; red for rejected areas)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.534, + 0.788, + 0.58 + ], + "angle": 0, + "content": "provide benefits such as accuracy in smaller scenes, reduced training times, as well as minimized storage requirements. Given these advantages, SCR is thus the primary focus of this study and presents potential for further enhancements." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.582, + 0.788, + 0.793 + ], + "angle": 0, + "content": "Albeit effective, previous SCR methods face two primary challenges to be addressed: the presence of dynamic objects and texture-less regions. Dynamic objects, such as pedestrians and cars, pose difficulties for these techniques due to their changing nature and the unsuitable features extracted from these regions. As illustrated in Fig. 1, these regions could lead to high reprojection errors during training. The second challenge is that current methods struggle with flat, texture-less surfaces. Despite their apparent simplicity, these areas often result in inaccurate scene coordinates due to difficulties in feature extraction. While previous methods [14, 15, 17] employ techniques like RANSAC to filter out outliers and differentiable end-to-end optimization approaches to disregard these outliers, they demand significant computational time for end-to-end processing and do not account for the semantic meaning of the selected areas, which may still result in the selection of outlier areas. Thus, they might not effectively prevent the models from unstable weight updates and can lead to training instability." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.796, + 0.788, + 0.842 + ], + "angle": 0, + "content": "In light of the aforementioned issues, this study aims to explore the possibility of guiding the sampling process to favor regions with low reprojection errors, and seeks to leverage visual information in an image to expand masks" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.241, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Reprojection Errors as Prompts for Efficient Scene Coordinate Regression" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.147, + 0.786, + 0.403 + ], + "angle": 0, + "content": "that bear semantic regional meanings. The core philosophy of our methodology focuses on selecting robust regions for training without the need to manually define explicit areas or categories. Such robust regions might differ across scenes, which make it inappropriate for manual definition, as elaborated in Section 4. More specifically, we propose a strategy, named Error-Guided Feature Selection (EGFS), for deriving low reprojection error samples as point prompts, and expanding them to encompass a complete mask with similar semantic meanings. The masks are iteratively updated after being frozen for a specified period, which enable the proposed model to dynamically renew the focused regions. To achieve this prompt-based semantic regions expansion, we employ the Segment Anything Model (SAM) [1], a vision foundation model capable of providing a general understanding of the scene. This error-guided scheme ensures that our training process concentrates on regions with low reprojection errors and adaptively reduces the variability introduced by dynamic or texture-less objects. Moreover, EGPS adopts a predicted confidence map to refine the expanded error-guided mask for updates, which ensures the regions marked by the masks are sufficiently reliable. An illustrations of each component is provided in Fig. 1." + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.405, + 0.786, + 0.554 + ], + "angle": 0, + "content": "To validate the effectiveness of our proposed method, we conduct evaluations on both outdoor and indoor datasets, including the Cambridge Landmarks dataset [18] and the Indoor6 [19] dataset, for evaluating the performance of our proposed SCR methodology. The experimental results suggest that our method can be broadly applied to previous methods and provides benefits, and leads to state-of-the-art (SOTA) performance when compared with SCR approaches without leveraging any 3D information in terms of estimated translational and rotational errors, while requiring less training time and smaller model sizes. We further perform a series of detailed analyses and ablation studies to confirm the effectiveness of our error-guided sampled regions and the proposed methodology." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.574, + 0.387, + 0.59 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.602, + 0.786, + 0.798 + ], + "angle": 0, + "content": "Scene Coordinate Regression. Conventional SCR approaches typically begin by establishing 2D-3D correspondences and then utilize RANSAC-based optimization to estimate the camera pose. To establish 2D-3D correspondences, previous endeavors have explored the use of random forests [5, 20-22] or convolutional neural networks (CNNs) [11-15,17,23] to regress scene coordinates from images. Since SCR typically requires scene-specific training, training time has emerged as a critical concern in practical applications. The ACE [17] method, which can be trained within five minutes, stands out as a promising solution. Moreover, unlike methods that map every 2D pixel to a 3D coordinate, some works [19,24,25] have aimed at learning to detect 3D landmarks to establish 2D-3D correspondences. However, they often require the reconstruction of a 3D model, which incurs additional computational time and costs. Our method seeks to utilize SAM to derive error-guided masks from 2D images, while maintaining the training efficiency." + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.81, + 0.786, + 0.84 + ], + "angle": 0, + "content": "Emphasis on Robust Features for Localization. Several prior studies [24,26-31] in visual localization have discovered that not all regions within an image contribute" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.377, + 0.129 + ], + "angle": 0, + "content": "T.-R. Liu et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.239 + ], + "angle": 0, + "content": "equally to localization accuracy. As a result, these approaches have attempted to distinguish robust and invariant features, found that prioritizing the recognition and comprehension of those regions can enhance visual localization tasks more efficiently and effectively. However, a common limitation of these methods is their reliance on pre-defined semantics or the necessity of a 3D model, which could hinder their generalizability and practicality in real-world applications." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.26, + 0.749, + 0.279 + ], + "angle": 0, + "content": "3 Preliminary of Scene Coordinate Regression (SCR)" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.291, + 0.79, + 0.504 + ], + "angle": 0, + "content": "SCR-based visual localization determines the camera pose of an RGB image \\(I \\in \\mathbb{R}^{H \\times W \\times 3}\\) by predicting 3D scene coordinates \\(\\mathcal{V}\\) for a set of pixels and establishing the 2D-3D correspondences between pixel coordinates in the image and those predicted 3D scene coordinates. The primary objective of the training process is to establish a mapping function \\(f(\\cdot)\\) such that \\(\\mathcal{V} = f(I)\\). The training data comprise RGB images paired with their respective camera poses, but do not necessarily include depth information. To generate accurate \\(\\mathcal{V}\\), SCR methods often involve optimizing the reprojection error \\(\\mathcal{R}\\), which is measured by projecting a 3D predicted scene coordinate \\(\\mathrm{y_i} \\in \\mathcal{V}\\) back onto its 2D pixel coordinate and calculating the discrepancy from the actual coordinate of the image patch, where i denotes the patch index. The reprojection error for each patch \\(\\mathrm{r_i} \\in \\mathcal{R}\\) can be formulated as \\(r(\\mathrm{p_i}, \\mathrm{y_i}, \\mathrm{h^*}) = \\| \\mathrm{p_i} - K\\mathrm{h^{*-1}}\\mathrm{y_i}\\|\\), where \\(K\\) denotes the camera intrinsic matrix and \\(\\mathrm{h^*}\\) denotes the ground truth camera pose. After obtaining \\(\\mathcal{V}\\), the camera pose h can be estimated according to the following equation:" + }, + { + "type": "equation", + "bbox": [ + 0.305, + 0.512, + 0.786, + 0.531 + ], + "angle": 0, + "content": "\\[\n\\mathrm {h} = g (\\mathcal {C}) = g ^ {P n P} \\left(\\mathcal {C} _ {\\mathcal {I}}\\right), \\text {w i t h} \\mathcal {C} = \\left\\{\\left(\\mathrm {p} _ {\\mathrm {i}}, \\mathrm {y} _ {\\mathrm {i}}\\right) | \\mathrm {p} _ {\\mathrm {i}} \\in I, \\mathrm {y} _ {\\mathrm {i}} \\in \\mathcal {Y} \\right\\}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.54, + 0.788, + 0.662 + ], + "angle": 0, + "content": "where \\( g(\\cdot) \\) involves PnP with RANSAC, followed by Levenberg-Marquardt-based refinement, \\( \\mathcal{C} \\) represents the set of all 2D-3D correspondences and \\( \\mathcal{C}_{\\mathcal{I}} \\) signifies inlier correspondences after refinement, which is a subset of \\( \\mathcal{C} \\). Inlier correspondences denote those that are better consistent with the estimated camera pose, while correspondences that do not align well with the estimated camera pose are termed outliers. The final camera pose estimation is determined by the inlier correspondences \\( \\mathcal{C}_{\\mathcal{I}} \\). In the following paragraphs, we introduce two scene coordinate regression methods that have demonstrated superior performance." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.788, + 0.783 + ], + "angle": 0, + "content": "DSAC Variants. DSAC [13] introduces a differentiable variant of RANSAC, and enables end-to-end training of the entire pipeline. This design combines the benefits of RANSAC for handling outliers with the power of deep neural networks to learn complex patterns. In addition, DSAC++ [14] removes the requirement of depth information for SCR training and adopts robust loss terms to downweight outliers. DSAC* [15] further simplifies and refines DSAC++, and encourages the network to focus on reliable scene structures while ignoring outlier predictions." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.795, + 0.788, + 0.842 + ], + "angle": 0, + "content": "ACE. ACE [17] is a fast SCR method that can achieve promising localization accuracy within a short training time of only five minutes. ACE splits its regression network into a pre-trained scene-agnostic backbone \\( f_{B} \\), and a scene-specific" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.24, + 0.115, + 0.732, + 0.13 + ], + "angle": 0, + "content": "Reprojection Errors as Prompts for Efficient Scene Coordinate Regression" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "image", + "bbox": [ + 0.234, + 0.145, + 0.499, + 0.268 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.502, + 0.145, + 0.768, + 0.267 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.234, + 0.27, + 0.499, + 0.392 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.502, + 0.27, + 0.768, + 0.392 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.409, + 0.788, + 0.452 + ], + "angle": 0, + "content": "Fig. 2: Analysis between reprojection error and semantic meaning. The analysis result indicates the regions with low reprojection errors tend to have higher inlier ratios, while the errors do not always align with specific semantic categories, e.g., \"tree\" and \"rug\"." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.481, + 0.788, + 0.543 + ], + "angle": 0, + "content": "multi-layer perceptron (MLP) \\( f_{H} \\). The backbone \\( f_{B} \\) extracts feature vectors \\( \\mathbf{f}_i \\) from image patches \\( p_i \\) and the MLP \\( f_{H} \\) predicts the scene coordinates \\( y_{i} \\) based on the extracted \\( \\mathrm{f}_i \\). Between \\( f_{B} \\) and \\( f_{H} \\), the features are shuffled to decorrelate gradients within a batch, which enables ACE to enhance its training efficiency." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.567, + 0.755, + 0.586 + ], + "angle": 0, + "content": "4 In-Depth Evaluation of Scene Coordinate Regression" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.602, + 0.618, + 0.618 + ], + "angle": 0, + "content": "4.1 Challenges in Scene Coordinate Regression" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.629, + 0.789, + 0.844 + ], + "angle": 0, + "content": "Several challenging factors have been identified that may compromise the accuracy of SCR. Two key challenges among them are dynamic objects and textureless surfaces [27, 32-34]. Dynamic objects exhibit characteristics of motion or disappearance across different frames, while textureless surfaces lack distinctive feature points or salient characteristics. These challenges highlight the complexity of achieving precise SCR in complex visual scenarios. To address them, previous SCR approaches [13-15, 17] have employed RANSAC to filter out outliers, and some researches [14, 15, 17] have attempted to utilize end-to-end learning to concentrate on reliable scene structures while down-weighting outlier predictions during training. Unfortunately, these prior endeavors failed to fully exploit the rich semantic information in the visual content, which can be crucial for enhancing the robustness and accuracy of camera pose estimation in complex scenes. Furthermore, RANSAC-based approaches for camera pose estimation could potentially fail in scenarios with insufficient number of inlier correspondences [35]." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.377, + 0.128 + ], + "angle": 0, + "content": "T.-R. Liu et al." + }, + { + "type": "image", + "bbox": [ + 0.216, + 0.143, + 0.788, + 0.265 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.343, + 0.274, + 0.66, + 0.289 + ], + "angle": 0, + "content": "Fig. 3: An overview of the training framework." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.317, + 0.764, + 0.334 + ], + "angle": 0, + "content": "4.2 Analysis between Reprojection Error and Semantic Meaning" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.341, + 0.789, + 0.691 + ], + "angle": 0, + "content": "For a more comprehensive understanding of the correlation and relationship between reprojection errors and semantic meaning, we investigate the performance of ACE on the reprojection errors and the inlier correspondences selected by RANSAC for estimating the final camera pose during the inference process. First, image pixels are labeled by a semantic segmentation model ViT-Adapter [36]. Subsequently, the reprojection errors and inlier correspondences of ACE for each semantic class are calculated. Fig. 2 presents the median reprojection errors and the ratios of inlier correspondences for the top ten most frequent classes in each scene. The left two figures correspond to the outdoor dataset Cambridge Landmarks [18], while the right two figures pertain to the indoor dataset Indoor6 [19]. It is observed that regions with low reprojection errors tend to have relatively higher inlier ratios, indicating a greater impact on the estimated camera poses. As a result, improving the accuracy of low-error areas could lead to more precise camera pose predictions. Nevertheless, it is also observed that the reprojection errors for the same class can vary significantly across different scenes. For example, in the Cambridge Landmarks dataset, the error for \"tree\" is low in King's College but high in Old Hospital. Similarly, in the Indoor6 dataset, the error for \"rug\" is high in scene1 but low in scene3. This indicates that low reprojection errors do not always align with specific semantic categories, which presents a challenge in directly using certain specific semantic segmentation labels for area selection. Motivated by these insights, the primary objective of this study is to develop a methodology that does not rely on pre-defined semantic categories to identify low reprojection error areas for enhancing localization accuracy in SCR." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.711, + 0.381, + 0.729 + ], + "angle": 0, + "content": "5 Methodology" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.741, + 0.638, + 0.756 + ], + "angle": 0, + "content": "5.1 Problem Definition and Framework Overview" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Fig. 3 provides an overview of the proposed training framework for SCR, which iteratively samples features to train a scene-specific MLP. The scene-specific MLP consists of a scene coordinate head and a confidence head. In each iteration, the model is trained for \\( k \\) epochs. During the initial iteration, features are randomly sampled from all parts of images in order to derive the first set" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.24, + 0.115, + 0.732, + 0.131 + ], + "angle": 0, + "content": "Reprojection Errors as Prompts for Efficient Scene Coordinate Regression" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.254 + ], + "angle": 0, + "content": "of reprojection errors. In subsequent iterations, features are selected based on error-guided feature selection (EGFS) masks generated according to reprojection errors and a confidence map. This iterative mechanism enables the model to dynamically update its focus areas throughout the training process. During the inference phase, the network estimates scene coordinates \\(\\mathcal{V}\\) and a confidence map for the refinement, while the masks generated during training are not required for the inference phase. The inference procedure is depicted in Fig. 4." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.278, + 0.684, + 0.294 + ], + "angle": 0, + "content": "5.2 Error-Guided Feature Selection (EGFS) with SAM" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.305, + 0.789, + 0.699 + ], + "angle": 0, + "content": "In alignment with the discussion in Section 4 on identifying potential areas contributing to the final camera pose estimation during the training stage, we employ the Segmentation Anything Model (SAM) [1], which is capable of leveraging diverse visual cues to generate high-quality object masks. Specifically, in this study, SAM is utilized to extend point prompts with low reprojection errors to form a complete mask sharing a similar semantic context. The rationale behind this approach is the high likelihood that these comprehensive semantic regions correspond to low error areas and can contribute to final pose estimation. Such a concept harnesses the capabilities of the foundational SAM model to uncover more viable areas, and hence, enabling a thorough understanding of the image context. In our experiments, we utilize EfficientViT-SAM-L0 [37], a lightweight and efficient variant of SAM, to identify salient regions by selecting the \\(\\tau\\) percentage of points with the lowest reprojection errors as point prompts, where \\(\\tau\\) is adjustable and set to ten in our case. After obtaining the error-guided mask from the SAM model, the predicted confidence map \\(c\\) is utilized to refine the error-guided mask to ensure the points marked by the mask is sufficiently reliable. Specifically, in each image, the points with confidence scores below a threshold \\(\\sigma\\) are filtered out. The design for confidence refinement and its optimization are described in Section 5.3. In each iteration, the refined error-guided masks serve as guidance for training the scene-specific MLP. Only the features corresponding to the regions selected by the masks are sampled into the training buffer. This selective inclusion is facilitated by the nature of \\(f_{H}\\) [17], which employs \\(1\\times 1\\) convolutions to treat and process each selected sample independently using the same set of shared weights. The proposed feature selection mechanism ensures that the training process focuses on the crucial regions identified in the current iteration, and therefore enables the model to enhance its overall performance." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.722, + 0.74, + 0.737 + ], + "angle": 0, + "content": "5.3 Scene Coordinate and EGFS Refinement with Confidence" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.841 + ], + "angle": 0, + "content": "In order to further refine the error-guided mask and ensure the selected points are of sufficiently high quality for predicting accurate camera pose, we incorporate the design of a confidence map [38, 39]. Specifically, we predict a confidence score pixel-wise from the confidence head and optimize it jointly with the reprojection error. This confidence head replaces the last three \\(1 \\times 1\\) convolution layers of the original scene-specific MLP with two \\(1 \\times 1\\) convolution layers for" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.377, + 0.128 + ], + "angle": 0, + "content": "T.-R. Liu et al." + }, + { + "type": "image", + "bbox": [ + 0.302, + 0.144, + 0.703, + 0.248 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.342, + 0.257, + 0.661, + 0.273 + ], + "angle": 0, + "content": "Fig. 4: An overview of the inference procedure." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.298, + 0.784, + 0.329 + ], + "angle": 0, + "content": "confidence score prediction, thus enabling adaptive weights for the reprojection loss at each position. The training loss used by our method is formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.307, + 0.334, + 0.786, + 0.375 + ], + "angle": 0, + "content": "\\[\n\\ell \\left(\\mathrm {p} _ {\\mathrm {i}}, \\mathrm {y} _ {\\mathrm {i}}, \\mathrm {h} _ {\\mathrm {i}} ^ {*}\\right) = \\left\\{ \\begin{array}{l l} \\mathrm {c} _ {\\mathrm {i}} \\cdot \\hat {r} \\left(\\mathrm {p} _ {\\mathrm {i}}, \\mathrm {y} _ {\\mathrm {i}}, \\mathrm {h} _ {\\mathrm {i}} ^ {*}\\right) - \\alpha \\log \\mathrm {c} _ {\\mathrm {i}} & \\text {i f y} _ {\\mathrm {i}} \\in \\mathcal {V} \\\\ \\| \\mathrm {y} _ {\\mathrm {i}} - \\bar {\\mathrm {y}} _ {\\mathrm {i}} \\| _ {0} - \\alpha \\log (1 - \\mathrm {c} _ {\\mathrm {i}}) & \\text {o t h e r w i s e} \\end{array} , \\right. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.381, + 0.788, + 0.624 + ], + "angle": 0, + "content": "where \\(\\hat{r} (\\mathrm{p_i},\\mathrm{y_i},\\mathrm{h_i^*})\\) represents a tanh clamping of the reprojection error that changes over time, \\(\\mathcal{V}\\) denotes the set of 2D pixels with valid scene coordinate predictions, as explained in [15, 17], \\(c_{i}\\) is the confidence score at pixel i, \\(\\alpha\\) is a hyperparameter that balances the confidence regularization term, and \\(\\bar{\\mathbf{y}}_{\\mathrm{i}}\\) is a dummy scene coordinate derived from the ground truth camera pose, assuming a constant image depth of \\(10\\mathrm{m}\\). When the scene coordinate prediction is valid, the model down-weights the reprojection errors in challenging regions using confidence scores and focuses more on regions with reliable predictions. On the other hand, when the scene coordinate prediction is invalid, the model is incentivized to lower the confidence scores for inaccurate predictions. The confidence score is also used in the inference phase to select reliable scene coordinates. For each query image, the median of the confidence scores is calculated, and only scene coordinates with confidence greater than this median are selected. This thresholding ensures that only the most reliable 2D-3D correspondences are used to generate the camera pose by the PnP with RANSAC pose solver. By focusing on high-confidence correspondences, a more accurate camera pose can be estimated." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.643, + 0.462, + 0.661 + ], + "angle": 0, + "content": "6 Experimental Results" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.671, + 0.436, + 0.687 + ], + "angle": 0, + "content": "6.1 Experimental Setups" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.693, + 0.785, + 0.74 + ], + "angle": 0, + "content": "Datasets. Our evaluation of the proposed method was conducted on two representative datasets: the outdoor dataset Cambridge Landmarks [18] and the indoor dataset Indoor6 [19]. The details of the datasets are described as follows:" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.749, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Cambridge Landmarks. The Cambridge Landmarks Dataset [18] is a renowned outdoor dataset extensively utilized for visual localization tasks, which includes five distinct scenes. Every scene in this dataset is composed of RGB images along with their respective ground truth camera poses reconstructed using the SfM technique. It includes a broad spectrum of conditions and various viewpoints, and is ideal for evaluating the resilience and effectiveness of various SCR methods." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.24, + 0.115, + 0.732, + 0.131 + ], + "angle": 0, + "content": "Reprojection Errors as Prompts for Efficient Scene Coordinate Regression" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.788, + 0.189 + ], + "angle": 0, + "content": "Table 1: A comparison of model sizes, training times, and median translation and rotation errors (in \\(\\mathrm{cm} / {}^{\\circ}\\)) evaluated on the Cambridge Landmarks dataset. 'dual' and 'quad.' denote the ensemble versions of two and four models, respectively." + }, + { + "type": "table", + "bbox": [ + 0.22, + 0.199, + 0.787, + 0.332 + ], + "angle": 0, + "content": "
MethodSizeMapping TimeKing's College Great Court Old Hospital Shop Facade St Mary's ChurchAverage
FMhLoc (SP+SG) [4,6,8]~800MB~35 min12/0.216/0.115/0.34/0.27/0.211/0.2
pixLoc [26]~600MB30/0.114/0.216/0.35/0.210/0.315/0.2
SCR (w/ 3D model)DSAC* [15]28MB15 hr15/0.349/0.321/0.45/0.313/0.421/0.3
FocusTune [31]4MB6 min19/0.338/0.118/0.46/0.315/0.519/0.3
FocusTune (quad.) [31]16MB24 min15/0.329/0.117/0.45/0.29/0.315/0.3
NeuMap [40]--14/0.26/0.119/0.36/0.217/0.512/0.3
SACReg-L [41]--11/0.213/0.113/0.26/0.25/0.310/0.2
SCRDSAC* [15]28MB15 hr18/0.334/0.221/0.45/0.315/0.619/0.4
ACE [17]4MB5 min28/0.443/0.231/0.65/0.318/0.625/0.4
ACE (quad.) [17]16MB20 min18/0.328/0.125/0.55/0.39/0.317/0.3
EGFS4.5MB12 min14/0.331/0.121/0.45/0.315/0.517/0.3
EGFS (dual)9MB21 min14/0.328/0.119/0.45/0.210/0.315/0.3
" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.345, + 0.785, + 0.373 + ], + "angle": 0, + "content": "Table 2: A comparison of model sizes, training times, and the proportions of translation and rotation errors that are below \\(5\\mathrm{cm} / 5^{\\circ}\\) evaluated on the Indoor6 dataset." + }, + { + "type": "table", + "bbox": [ + 0.221, + 0.385, + 0.787, + 0.508 + ], + "angle": 0, + "content": "
MethodSizeMapping Timescene1scene2ascene3scene4ascene5scene6Average
FMhLoc [4]~1.5GB~3.3 hr70.5%52.1%86.0%75.3%58.0%86.7%71.4%
SCR (w/ 3D model)DSAC* [15]28MB15 hr18.7%28.0%19.7%60.8%10.6%44.3%30.4%
SLD* (300 landmarks) [25]15MB~5.5 hr47.2%48.2%56.2%67.7%33.7%52.0%50.8%
SLD* (1000 landmarks) [25]120MB~44 hr68.5%62.6%76.2%77.2%57.8%78.0%70.1%
SCRDSAC* [15]28MB15 hr23.0%33.9%26.0%67.1%10.6%50.2%35.1%
ACE [17]4MB5 min26.0%32.3%31.4%62.0%14.2%47.4%35.6%
ACE (quad.) [17]16MB20 min52.9%52.5%62.9%69.6%31.1%82.4%58.6%
EGFS4.5MB21 min46.4%60.6%56.4%78.7%22.8%71.6%56.1%
EGFS (dual)9MB30 min58.5%59.1%67.0%76.1%30.6%75.9%61.2%
" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.535, + 0.788, + 0.596 + ], + "angle": 0, + "content": "Indoor6. The Indoor6 [19] dataset comprises six distinct indoor scenes captured over several days, with ground truth camera poses computed using COLMAP [42]. Each scene includes multiple rooms and contains illumination variations, making it challenging for different types of visual localization tasks." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.616, + 0.788, + 0.709 + ], + "angle": 0, + "content": "Implementation Details. The training process of our methodology includes 20 epochs, with the generation of masks every five epochs. We set the confidence regularization parameter \\(\\alpha\\) to ten and the reprojection error threshold \\(\\tau\\) to ten. For mask refinement, the parameter \\(\\sigma\\) is set to the median confidence score of each image. Our method builds upon the ACE [17] architecture as the backbone, and thus, the remaining hyperparameters are kept the same as those used in [17]." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.727, + 0.75, + 0.743 + ], + "angle": 0, + "content": "6.2 Results on the Cambridge Landmarks and Indoor6 Dataset" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.75, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Cambridge Landmarks Dataset. Table 1 presents the experimental results evaluated on the Cambridge Landmarks dataset, which affirm the effectiveness and efficiency of our proposed methodology. It can be observed that EGFS maintains a model size similar to ACE, with only a 0.5MB increase for the confidence head, while reducing average translational and rotational errors. The results further reveal that EGFS not only surpasses DSAC* in performance and performs" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.377, + 0.129 + ], + "angle": 0, + "content": "T.-R. Liu et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.788, + 0.175 + ], + "angle": 0, + "content": "Table 3: Comparison of the effectiveness between (a) directly sampling solely based on pixel points with low reprojection errors, and (b) sampling through the EGFS masks." + }, + { + "type": "table", + "bbox": [ + 0.233, + 0.185, + 0.773, + 0.354 + ], + "angle": 0, + "content": "
DatasetSceneri < Q0.3(R) ri < Q0.4(R) ri < Q0.5(R) ri < Q0.6(R) ri < Q0.7(R)EGFS Masks
CambridgeKing's College20/0.319/0.320/0.320/0.320/0.314/0.3
Great Court39/0.237/0.136/0.238/0.239/0.231/0.1
Old Hospital26/0.525/0.523/0.525/0.524/0.521/0.4
Shop Facade6/0.36/0.36/0.36/0.35/0.35/0.3
St Mary's Church17/0.517/0.516/0.517/0.516/0.515/0.5
Average22/0.421/0.420/0.421/0.421/0.417/0.3
Indoor6scene134.4%34.9%34.9%32.0%30.2%46.4%
scene2a39.8%40.6%46.1%41.3%42.5%60.6%
scene345.2%43.9%44.9%41.3%42.3%56.4%
scene4a63.2%63.9%62.6%63.2%67.1%78.7%
scene519.5%18.1%17.1%18.3%16.9%22.8%
scene659.7%60.3%60.0%57.2%58.1%71.6%
Average43.6%43.6%44.3%42.2%42.9%56.1%
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.381, + 0.788, + 0.445 + ], + "angle": 0, + "content": "on par with the ensemble version of ACE (i.e., ACE (quad.)), but also features a smaller model size and reduced training time. Furthermore, our model's ensemble version (i.e., EGFS (dual.)) is able to exceed the performance of DSAC* and ACE (quad.) while requiring a smaller model size and less training time." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.468, + 0.79, + 0.561 + ], + "angle": 0, + "content": "Indoor6 Dataset. Table 2 presents the evaluation results on Indoor6. EGFS significantly outperforms DSAC* and ACE across all scenes, while also achieving comparable performance to ACE (quad.) with a much smaller model size (4.5MB compared to 16MB). Moreover, EGFS (dual.) demonstrates superior performance relative to ACE (quad.), which further substantiates the effectiveness of EGFS in optimization through error-guided masks and confidence refinement." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.585, + 0.655, + 0.601 + ], + "angle": 0, + "content": "6.3 Effectiveness of Error-Guided Feature Selection" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.614, + 0.789, + 0.842 + ], + "angle": 0, + "content": "To further substantiate the advantages of employing the EGFS approach, this experiment compares several baseline schemes that sample points with different reprojection error thresholds against EGFS. The results are summarized in Table 3. Specifically, these baseline schemes select sample points based on various quantiles of the reprojection error map, ranging from \\(30\\%\\) to \\(70\\%\\), denoted as \\(Q_{0.3}(R)\\) through \\(Q_{0.7}(R)\\), respectively. In each baseline scheme, a point is chosen if its reprojection error \\(r\\) falls below the threshold specified by \\(Q_{\\text{quantile}}(R)\\). The results suggest that relying solely on reprojection errors does not guarantee optimal learning of scene coordinates. This is due to the fact that areas of low reprojection errors can be scattered and may not encompass entire semantic objects. In contrast, the proposed EGFS expands the points into complete semantic masks followed by confidence refinement, which enables our model to determine scene coordinates with minimal transitional and rotational errors. This experiment thus confirms the effectiveness of leveraging SAM's ability to interpret image context alongside confidence maps for enhanced sample point selection." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.241, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Reprojection Errors as Prompts for Efficient Scene Coordinate Regression" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "image", + "bbox": [ + 0.245, + 0.145, + 0.758, + 0.312 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.322, + 0.788, + 0.392 + ], + "angle": 0, + "content": "Fig. 5: Visualization of the EGFS mask refinement process at every five epochs, which depicts the reprojection errors at the beginning (epoch 5) and the end (epoch 20), as well as the refined error-guided masks used throughout training. The red dots represent low reprojection errors that serve as prompts, while the light green overlay denotes the refined EGFS masks. It can be observed that the EGFS masks enhances over epochs." + }, + { + "type": "image", + "bbox": [ + 0.258, + 0.414, + 0.76, + 0.58 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.592, + 0.788, + 0.62 + ], + "angle": 0, + "content": "Fig. 6: Visualization of the estimated camera pose trajectories from testing sequences with the camera frustums colored based on translational errors. Pose errors denoted." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.658, + 0.421, + 0.673 + ], + "angle": 0, + "content": "6.4 Qualitative Results" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.704, + 0.789, + 0.841 + ], + "angle": 0, + "content": "Iterative Error-guided Feature Selection Mask. Fig. 5 illustrates the refinement of EGFS masks every five epochs and compares the reprojection errors at the beginning and the end. It can be observed that EGFS is not only capable of expanding points with low reprojection errors (i.e., \\( \\mathrm{r} < Q_{0.1}(R) \\)) into regions that bear similar semantic meanings, but also refines the masks with a confidence map to exclude uncertain areas. The masks are observed to be iteratively enhanced from coarse to fine. In the beginning (i.e., at epoch five), the reprojection errors appear noisy, whereas by the end (i.e., at epoch 20), the EGFS masks become more refined and concentrated on key regions beneficial for SCR." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.377, + 0.128 + ], + "angle": 0, + "content": "T.-R. Liu et al." + }, + { + "type": "image", + "bbox": [ + 0.248, + 0.147, + 0.749, + 0.379 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.396, + 0.789, + 0.454 + ], + "angle": 0, + "content": "Fig. 7: Visualization of point clouds reconstructed from estimated scene coordinates of the training sequence. The point clouds derived from the scene coordinates estimated by the model trained with the proposed EGFS are clearer compared to those from the ACE model, even without the application of refined EGFS masks at inference time." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.483, + 0.789, + 0.591 + ], + "angle": 0, + "content": "Visualization of the Estimated Camera Poses. In this section, we present the estimated camera poses from the testing sequences, and compare the proposed EGFS with the ACE baseline. The results are illustrated in Fig. 6. To visualize camera trajectories, we connect consecutive camera positions with lines and indicate camera orientations with frustums. Each sampled camera pose is associated with a color to represent the translation error. It is observed that our proposed EGFS can estimate more accurate camera poses as compared to ACE." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.614, + 0.789, + 0.842 + ], + "angle": 0, + "content": "Scene Coordinates and 3D Point Clouds w/ and w/o EFGS. To validate the effectiveness of EGFS qualitatively and its impact on the quality of estimated scene coordinates that learned from training sequence, we depict the 3D point cloud generated from the scene coordinates from the training sequence, and apply colors based on the corresponding queried pixels in Fig. 7. To compare the 3D point clouds reconstructed from the scene coordinates, we compare the baseline ACE approach with the proposed EGFS approach. For our approach, we visualize the point clouds both before and after the application of EGFS masks and confidence maps when collecting the scene coordinates. Please note that both before and after versions which shown in the figure are trained with EGFS refinement. The 3D point cloud constructed by the proposed approach appears clearer, in contrast to the ACE-generated point cloud, which exhibits blurriness due to inaccurate scene coordinate estimation. In addition, the point cloud reconstructed from raw scene coordinates predicted without EGFS from our approach contains noise and floaters due to the inclusion of scene coordinates" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.241, + 0.115, + 0.732, + 0.13 + ], + "angle": 0, + "content": "Reprojection Errors as Prompts for Efficient Scene Coordinate Regression" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "image", + "bbox": [ + 0.249, + 0.146, + 0.752, + 0.363 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.373, + 0.788, + 0.416 + ], + "angle": 0, + "content": "Fig. 8: Visualization of the estimated scene coordinates and the confidence maps on unseen samples. The RGB color in the scene coordinates represents the XYZ coordinates. The point clouds are colored green for selected areas and red for rejected areas." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.446, + 0.788, + 0.507 + ], + "angle": 0, + "content": "with low confidence scores. On the other hand, the point cloud reconstructed from the scene coordinates with EGFS refinement exhibits clearer contours and less noise. This justifies EGFS's ability to eliminate areas of low confidence values and produce high-quality dense scene coordinates." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.531, + 0.788, + 0.668 + ], + "angle": 0, + "content": "Visualization of Estimated Confidence Map on Unseen Samples. In this section, we qualitatively evaluate the confidence maps estimated from the testing sequence by visualizing them in both the 2D image plane and as 3D projections onto the point cloud. The visualized point cloud is reconstructed from the training sequence, whereas the plotted samples are selected from the testing sequence and were not included in the training procedure. These visualizations are provided in Fig. 8. It can be observed that the estimated confidence map assists in rejecting scene coordinates located in areas where accurate prediction is challenging, such as in the air or on texture-less surfaces." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.692, + 0.391, + 0.708 + ], + "angle": 0, + "content": "6.5 Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.719, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Error-Guided Feature Selection (EGFS) and Confidence Refinement. We first present an ablation analysis to validate the effectiveness of the EGFS masks and confidence refinement in enhancing our method's performance. Table 4 reports the average results on the Cambridge Landmarks and Indoor6 datasets. The results demonstrate that the model trained with the proposed EGFS mechanism outperforms those trained without it, which indicates that focusing on training with robust features leads to improved performance. Furthermore, the incorporation of the confidence map also enhances the overall" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.377, + 0.128 + ], + "angle": 0, + "content": "T.-R. Liu et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.522, + 0.174 + ], + "angle": 0, + "content": "Table 4: Ablation study on the impact of each proposed components." + }, + { + "type": "table", + "bbox": [ + 0.222, + 0.185, + 0.524, + 0.281 + ], + "angle": 0, + "content": "
Error-Guided MasksConfidence RefinementCambridge (cm/°)Indoor6 (%)
XX25/0.435.6
✓X19/0.341.6
X✓19/0.450.9
✓✓17/0.356.1
" + }, + { + "type": "table_caption", + "bbox": [ + 0.538, + 0.145, + 0.788, + 0.174 + ], + "angle": 0, + "content": "Table 5: Impact of varying proportions of point prompts \\((\\tau)\\)" + }, + { + "type": "table", + "bbox": [ + 0.543, + 0.185, + 0.789, + 0.281 + ], + "angle": 0, + "content": "
Proportions of point prompts (τ)Cambridge (cm/°)Indoor6 (%)
517/0.354.3
1017/0.356.1
1517/0.355.4
2017/0.355.3
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.305, + 0.786, + 0.352 + ], + "angle": 0, + "content": "performance. The final row of Table 4 demonstrates that combining both techniques (i.e., our EGFS) achieves the best results, which further confirms the effectiveness of EGFS in improving localization accuracy." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.37, + 0.788, + 0.552 + ], + "angle": 0, + "content": "Analysis of Proportions of Point Prompts. We evaluate the performance of different proportions of point prompts used for expanding into EGFS masks, and the results are reported in Table 5. It can be observed that the proportions of point prompts do not significantly impact the performance on the Cambridge Landmark dataset. This may be attributed to the fact that the scenes in the dataset typically feature a primary architectural structure with fewer complicated details in the surroundings. On the other hand, for the Indoor6 dataset, the proportions of point prompts significantly affect the performance. The rationale behind these observations is that the indoor scenes are more complicated, and feature various small items that necessitate a larger number of low-error point prompts to comprehensively capture the entire scene structure. Please note that we employ \\(\\tau = 10\\) as our setting for all the experiments presented." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.57, + 0.36, + 0.587 + ], + "angle": 0, + "content": "7 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.598, + 0.789, + 0.842 + ], + "angle": 0, + "content": "This paper addressed key challenges in SCR for visual localization, and specifically focused on the impacts of dynamic objects and texture-less regions. Our approach introduced an innovative EGFS mechanism through the use of SAM and confidence maps to enhance the performance of SCR. This technique effectively filtered out problematic areas by concentrating on regions with low reprojection errors. Moreover, we used confidence maps to further refine the pixels selected by EGFS and perform this process iteratively to enable dynamic updates of the focused regions. The experimental results on the Cambridge Landmarks and Indoor6 datasets suggest that our method can provide improvements in terms of training efficiency, model size, and accuracy. Our findings highlighted the importance of carefully selecting low-reprojection error pixels by taking into account semantic information and confidence scores. Furthermore, we quantitatively and qualitatively demonstrated that dynamically updating the masks enables more robust selection of error points, thus enabling better training efficiency and effectiveness. Our ablation studies validated the effectiveness of the techniques adopted in our method, and solidified their roles in enhancing the performance." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.24, + 0.115, + 0.732, + 0.13 + ], + "angle": 0, + "content": "Reprojection Errors as Prompts for Efficient Scene Coordinate Regression" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.145, + 0.403, + 0.163 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.175, + 0.788, + 0.311 + ], + "angle": 0, + "content": "The authors gratefully acknowledge the support from the National Science and Technology Council (NSTC) in Taiwan under grant numbers MOST 111-2223-E007-004-MY3, 113-2221-E-007-122-MY3, 113-2640-E-002-003, 113-2221-E-007-104-MY3, as well as the financial support from Woven by Toyota, Inc., Japan. The authors would also like to express their appreciation for the donation of the GPUs from NVIDIA Corporation and NVIDIA AI Technology Center (NVAITC) used in this work. Furthermore, the authors extend their gratitude to the National Center for High-Performance Computing (NCHC) for providing the necessary computational and storage resources." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.332, + 0.324, + 0.348 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.223, + 0.361, + 0.787, + 0.403 + ], + "angle": 0, + "content": "1. Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Dlarr, and Ross Girshick. Segment anything. arXiv:2304.02643, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.223, + 0.403, + 0.787, + 0.443 + ], + "angle": 0, + "content": "2. Xiao-Shan Gao, Xiao-Rong Hou, Jianliang Tang, and Hang-Fei Cheng. Complete solution classification for the perspective-three-point problem. IEEE transactions on pattern analysis and machine intelligence, 25(8):930-943, 2003." + }, + { + "type": "ref_text", + "bbox": [ + 0.223, + 0.443, + 0.787, + 0.485 + ], + "angle": 0, + "content": "3. Martin A Fischler and Robert C Bolles. Random sample consensus: a paradigm for model fitting with applications to image analysis and automated cartography. Communications of the ACM, 1981." + }, + { + "type": "ref_text", + "bbox": [ + 0.223, + 0.485, + 0.787, + 0.526 + ], + "angle": 0, + "content": "4. Paul-Edouard Sarlin, Cesar Cadena, Roland Siegwart, and Marcin Dymczyk. From coarse to fine: Robust hierarchical localization at large scale. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.223, + 0.526, + 0.787, + 0.58 + ], + "angle": 0, + "content": "5. Jamie Shotton, Ben Glocker, Christopher Zach, Shahram Izadi, Antonio Criminisi, and Andrew Fitzgibbon. Scene coordinate regression forests for camera relocalization in RGB-D images. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2013." + }, + { + "type": "ref_text", + "bbox": [ + 0.223, + 0.58, + 0.787, + 0.622 + ], + "angle": 0, + "content": "6. Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Superpoint: Self-supervised interest point detection and description. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), pages 224-236, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.223, + 0.622, + 0.787, + 0.663 + ], + "angle": 0, + "content": "7. Jiaming Sun, Zehong Shen, Yuang Wang, Hujun Bao, and Xiaowei Zhou. LoFTR: Detector-free local feature matching with transformers. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), pages 8922-8931, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.223, + 0.663, + 0.787, + 0.704 + ], + "angle": 0, + "content": "8. Paul-Edouard Sarlin, Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. SuperGlue: Learning feature matching with graph neural networks. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.223, + 0.704, + 0.787, + 0.731 + ], + "angle": 0, + "content": "9. Philipp Lindenberger, Paul-Edouard Sarlin, and Marc Pollefeys. LightGlue: Local feature matching at light speed. arXiv preprint arXiv:2306.13643, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.731, + 0.787, + 0.772 + ], + "angle": 0, + "content": "10. Pablo Speciale, Johannes L Schonberger, Sing Bing Kang, Sudipta N Sinha, and Marc Pollefeys. Privacy preserving image-based localization. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), pages 5493-5503, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.772, + 0.787, + 0.813 + ], + "angle": 0, + "content": "11. Xiaotian Li, Juha Ylioinas, Jakob Verbeek, and Juho Kannala. Scene coordinate regression with angle-based reprojection loss for camera relocalization. In Proc. European Conf. on Computer Vision Workshop (ECCVW), 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.813, + 0.787, + 0.84 + ], + "angle": 0, + "content": "12. Xiaotian Li, Juha Ylioinas, and Juho Kannala. Full-frame scene coordinate regression for image-based localization. RSS, 2018." + }, + { + "type": "list", + "bbox": [ + 0.218, + 0.361, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.375, + 0.127 + ], + "angle": 0, + "content": "T.-R. Liu et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.787, + 0.203 + ], + "angle": 0, + "content": "13. Eric Brachmann, Alexander Krull, Sebastian Nowozin, Jamie Shotton, Frank Michel, Stefan Gumhold, and Carsten Rother. DSAC-Differentiable RANSAC for camera localization. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.203, + 0.787, + 0.243 + ], + "angle": 0, + "content": "14. Eric Brachmann and Carsten Rother. Learning less is more - 6D camera localization via 3D surface regression. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.243, + 0.787, + 0.27 + ], + "angle": 0, + "content": "15. Eric Brachmann and Carsten Rother. Visual camera re-localization from RGB and RGB-D images using DSAC. TPAMI, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.27, + 0.787, + 0.297 + ], + "angle": 0, + "content": "16. Eric Brachmann and Carsten Rother. Expert sample consensus applied to camera re-localization. In Proc. IEEE Int. Conf. on Computer Vision (ICCV), 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.297, + 0.787, + 0.338 + ], + "angle": 0, + "content": "17. Eric Brachmann, Tommaso Cavallari, and Victor Adrian Prisacariu. Accelerated coordinate encoding: Learning to relocalize in minutes using RGB and poses. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.338, + 0.787, + 0.378 + ], + "angle": 0, + "content": "18. Alex Kendall, Matthew Grimes, and Roberto Cipolla. PoseNet: A convolutional network for real-time 6-DOF camera relocalization. In Proc. IEEE Int. Conf. on Computer Vision (ICCV), December 2015." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.378, + 0.787, + 0.418 + ], + "angle": 0, + "content": "19. Tien Do, Ondrej Miksik, Joseph DeGol, Hyun Soo Park, and Sudipta N. Sinha. Learning to detect scene landmarks for camera localization. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.418, + 0.787, + 0.473 + ], + "angle": 0, + "content": "20. Abner Guzman-Rivera, Pushmeet Kohli, Ben Glocker, Jamie Shotton, Toby Sharp, Andrew Fitzgibbon, and Shahram Izadi. Multi-output learning for camera localization. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2014." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.473, + 0.787, + 0.527 + ], + "angle": 0, + "content": "21. Eric Brachmann, Frank Michel, Alexander Krull, Michael Ying Yang, Stefan Gumhold, and carsten Rother. Uncertainty-driven 6D pose estimation of objects and scenes from a single RGB image. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2016." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.527, + 0.787, + 0.581 + ], + "angle": 0, + "content": "22. Julien Valentin, Matthias Niessner, Jamie Shotton, Andrew Fitzgibbon, Shahram Izadi, and Philip H. S. Torr. Exploiting uncertainty in regression forests for accurate camera relocalization. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2015." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.581, + 0.787, + 0.635 + ], + "angle": 0, + "content": "23. Tao Xie, Kun Dai, Siyi Lu, Ke Wang, Zhiqiang Jiang, Jinghan Gao, Dedong Liu, Jie Xu, Lijun Zhao, and Ruifeng Li. OFVL-MS: Once for visual localization across multiple indoor scenes. In Proc. IEEE Int. Conf. on Computer Vision (ICCV), 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.635, + 0.787, + 0.69 + ], + "angle": 0, + "content": "24. Zhaoyang Huang, Han Zhou, Yijin Li, Bangbang Yang, Yan Xu, Xiaowei Zhou, Hujun Bao, Guofeng Zhang, and Hongsheng Li. VS-Net: Voting with segmentation for visual localization. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.69, + 0.787, + 0.731 + ], + "angle": 0, + "content": "25. Tien Do and Sudipta N. Sinha. Improved scene landmark detection for camera localization. In Proceedings of the International Conference on 3D Vision (3DV), 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.731, + 0.787, + 0.799 + ], + "angle": 0, + "content": "26. Paul-Edouard Sarlin, Ajaykumar Unagar, Mans Larsson, Hugo Germain, Carl Toft, Victor Larsson, Marc Pollefeys, Vincent Lepetit, Lars Hammarstrand, Fredrik Kahl, and Torsten Sattler. Back to the Feature: Learning Robust Camera Localization from Pixels to Pose. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.799, + 0.787, + 0.84 + ], + "angle": 0, + "content": "27. Mohammad Altillawi. PixSelect: Less but reliable pixels for accurate and efficient localization. In Proc. IEEE Int. Conf. on Robotics and Automation (ICRA), pages 4156-4162. IEEE, 2022." + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.24, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Reprojection Errors as Prompts for Efficient Scene Coordinate Regression" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.788, + 0.189 + ], + "angle": 0, + "content": "28. Fei Xue, Ignas Budvytis, Daniel Olmeda Reino, and Roberto Cipolla. Efficient large-scale localization by global instance recognition. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), pages 17348-17357, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.189, + 0.788, + 0.229 + ], + "angle": 0, + "content": "29. Fei Xue, Ignas Budvytis, and Roberto Cipolla. SFD2: Semantic-guided feature detection and description. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), pages 5206-5216, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.229, + 0.788, + 0.27 + ], + "angle": 0, + "content": "30. Bach-Thuan Bui, Dinh-Tuan Tran, and Joo-Ho Lee. D2S: Representing local descriptors and global scene coordinates for camera relocalization. arXiv preprint arXiv:2307.15250, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.27, + 0.788, + 0.324 + ], + "angle": 0, + "content": "31. Son Tung Nguyen, Alejandro Fontan, Michael Milford, and Tobias Fischer. Focus-Tune: Tuning visual localization through focus-guided sampling. In Proc. IEEE Winter Conf. on Applications of Computer Vision (WACV), pages 3606-3615, January 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.324, + 0.788, + 0.378 + ], + "angle": 0, + "content": "32. Johanna Wald, Torsten Sattler, Stuart Golodetz, Tommaso Cavallari, and Federico Tombari. Beyond controlled environments: 3d camera re-localization in changing indoor scenes. In Proc. European Conf. on Computer Vision (ECCV), pages 467-487, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.379, + 0.788, + 0.434 + ], + "angle": 0, + "content": "33. Siyan Dong, Qingnan Fan, He Wang, Ji Shi, Li Yi, Thomas Funkhouser, Baoquan Chen, and Leonidas J Guibas. Robust neural routing through space partitions for camera relocalization in dynamic indoor environments. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.434, + 0.788, + 0.501 + ], + "angle": 0, + "content": "34. Donghwan Lee, Soohyun Ryu, Suyong Yeon, Yonghan Lee, Deokhwa Kim, Cheolho Han, Yohann Cabon, Philippe Weinzaepfel, Nicolas Guerin, Gabriela Csurka, and Martin Humenberger. Large-scale localization datasets in crowded indoor spaces. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), pages 3227-3236, June 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.501, + 0.788, + 0.543 + ], + "angle": 0, + "content": "35. Hongyi Fan, Joe Kileel, and Benjamin Kimia. On the instability of relative pose estimation and RANSAC's role. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), pages 8935-8943, June 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.543, + 0.788, + 0.582 + ], + "angle": 0, + "content": "36. Zhe Chen, Yuchen Duan, Wenhai Wang, Junjun He, Tong Lu, Jifeng Dai, and Yu Qiao. Vision transformer adapter for dense predictions. arXiv preprint arXiv:2205.08534, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.582, + 0.788, + 0.624 + ], + "angle": 0, + "content": "37. Han Cai, Junyan Li, Muyan Hu, Chuang Gan, and Song Han. EfficientViT: Lightweight multi-scale attention for high-resolution dense prediction. In Proc. IEEE Int. Conf. on Computer Vision (ICCV), pages 17302-17313, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.624, + 0.788, + 0.663 + ], + "angle": 0, + "content": "38. Alex Kendall and Yarin Gal. What uncertainties do we need in bayesian deep learning for computer vision? Advances in neural information processing systems, 30, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.663, + 0.788, + 0.704 + ], + "angle": 0, + "content": "39. Sheng Wan, Tung-Yu Wu, Wing H Wong, and Chen-Yi Lee. Confnet: predict with confidence. In 2018 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 2921-2925. IEEE, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.704, + 0.788, + 0.759 + ], + "angle": 0, + "content": "40. Shitao Tang, Sicong Tang, Andrea Tagliasacchi, Ping Tan, and Yasutaka Furukawa. Neumap: Neural coordinate mapping by auto-transdecoder for camera localization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 929-939, June 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.759, + 0.788, + 0.799 + ], + "angle": 0, + "content": "41. Revaud, Jérôme and Cabon, Yohann and Brégier, Romain and Lee, JongMin and Weinzaepfel, Philippe. SACReg: Scene-Agnostic Coordinate Regression for Visual Localization, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.799, + 0.788, + 0.839 + ], + "angle": 0, + "content": "42. Johannes L. Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), June 2016." + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.788, + 0.839 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/Reprojection Errors as Prompts for Efficient Scene Coordinate Regression/91ff6348-fcc0-42f1-b6d0-08cddca0de4a_origin.pdf b/2024/Reprojection Errors as Prompts for Efficient Scene Coordinate Regression/91ff6348-fcc0-42f1-b6d0-08cddca0de4a_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..9a77f585100074ecfc327bb1c43e9bb8c2e38647 --- /dev/null +++ b/2024/Reprojection Errors as Prompts for Efficient Scene Coordinate Regression/91ff6348-fcc0-42f1-b6d0-08cddca0de4a_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3bb8d3865abf8cea67d7cd0fc7bc76f732cf5de942cf3c535defc7d9f2d73a53 +size 15938794 diff --git a/2024/Reprojection Errors as Prompts for Efficient Scene Coordinate Regression/full.md b/2024/Reprojection Errors as Prompts for Efficient Scene Coordinate Regression/full.md new file mode 100644 index 0000000000000000000000000000000000000000..02ba4b47554b3791dceee494ffb59672a439a73f --- /dev/null +++ b/2024/Reprojection Errors as Prompts for Efficient Scene Coordinate Regression/full.md @@ -0,0 +1,238 @@ +# Reprojection Errors as Prompts for Efficient Scene Coordinate Regression + +Ting-Ru Liu $^{1}$ , Hsuan-Kung Yang $^{2}$ , Jou-Min Liu $^{1*}$ , Chun-Wei Huang $^{1*}$ , Tsung-Chih Chiang $^{1*}$ , Quan Kong $^{2}$ , Norimasa Kobori $^{2}$ , and Chun-Yi Lee $^{1}$ + +$^{1}$ National Tsing Hua University $^{2}$ Woven by Toyota, Inc. {tingru, diasimui, alan, george, cymaxwelllee}@elsa.cs.nthu.edu.tw {hsuan-kung.yang, quan.kong, norimasa.kobori} @woven toyota https://tingru0203.github.io/egfs + +Abstract. Scene coordinate regression (SCR) methods have emerged as a promising area of research due to their potential for accurate visual localization. However, many existing SCR approaches train on samples from all image regions, including dynamic objects and texture-less areas. Utilizing these areas for optimization during training can potentially hamper the overall performance and efficiency of the model. In this study, we first perform an in-depth analysis to validate the adverse impacts of these areas. Drawing inspiration from our analysis, we then introduce an error-guided feature selection (EGFS) mechanism, in tandem with the use of the Segment Anything Model (SAM) [1]. This mechanism seeds low reprojection areas as prompts and expands them into error-guided masks, and then utilizes these masks to sample points and filter out problematic areas in an iterative manner. The experiments demonstrate that our method outperforms existing SCR approaches that do not rely on 3D information on the Cambridge Landmarks and Indoor6 datasets. + +# 1 Introduction + +The objective of visual localization is to estimate a 6-DoF camera pose from images, a key component in fields such as Augmented Reality, Virtual Reality, and autonomous driving. Contemporary leading methods in visual localization typically involve establishing 2D-3D correspondences and then utilizing Perspective-Point (PnP) [2] with RANSAC [3] for camera pose estimation. These methods can be broadly classified into two main directions: feature-matching [4] and scene coordinate regression (SCR) [5]. Feature-matching approaches reconstruct a 3D scene using Structure from Motion (SfM), identify and describe key points in 2D images [6,7], and link these to 3D coordinates [8,9]. Nevertheless, they may encounter challenges such as high computational demands, significant storage requirements, and potential privacy concerns [10]. On the other hand, SCR methods [11-17] employ deep neural networks (DNNs) to predict the 3D coordinates of pixels and then utilize PnP with RANSAC for camera pose estimation, which + +![](images/81109ee5448de08a173c3c5cdd9c5f07af530f24c707302e16cb5843c84ea9a8.jpg) + +![](images/883e196dd6e558123f277c64d848b1800e7d3b73c6cb98ec4f45f7f1ca1cc253.jpg) + +![](images/dcd7b4046dfa769bf8dec1e588187ab33e4d3e3f807d1a020d7117ab844414ba.jpg) + +![](images/5b35f737699fb2b972c91d3bf48aa007781d148bd1689ce057f968a67c4533fd.jpg) + +![](images/03932a13732059a84a631165e27a977cf78cf596ee03c4ee09b3871fbb483151.jpg) +Fig. 1: Visualization of the primary components (i.e., (d)-(h)) introduced in the proposed visual localization scheme. (d) illustrates the point prompts selected from (b) with low reprojection errors, while (e) presents an error-guided mask expanded from the prompted points in (d) using SAM. (f) displays the proposed error-guided feature selection (EGFS), which refines the mask from (e) with the predicted confidence map (c) to ensure high-quality scene coordinates are sampled for estimating the final camera pose. The point cloud constructed from the predicted scene coordinates is shown on the right-hand side (i.e., (g)-(h)), with the confidence (yellow parts) and the refined EGFS mask (green for selected areas; red for rejected areas). + +![](images/ce4fd49ac586c0410b143e90782785e1f88e82b72dfbfb28fb472af3e0afaee2.jpg) + +provide benefits such as accuracy in smaller scenes, reduced training times, as well as minimized storage requirements. Given these advantages, SCR is thus the primary focus of this study and presents potential for further enhancements. + +Albeit effective, previous SCR methods face two primary challenges to be addressed: the presence of dynamic objects and texture-less regions. Dynamic objects, such as pedestrians and cars, pose difficulties for these techniques due to their changing nature and the unsuitable features extracted from these regions. As illustrated in Fig. 1, these regions could lead to high reprojection errors during training. The second challenge is that current methods struggle with flat, texture-less surfaces. Despite their apparent simplicity, these areas often result in inaccurate scene coordinates due to difficulties in feature extraction. While previous methods [14, 15, 17] employ techniques like RANSAC to filter out outliers and differentiable end-to-end optimization approaches to disregard these outliers, they demand significant computational time for end-to-end processing and do not account for the semantic meaning of the selected areas, which may still result in the selection of outlier areas. Thus, they might not effectively prevent the models from unstable weight updates and can lead to training instability. + +In light of the aforementioned issues, this study aims to explore the possibility of guiding the sampling process to favor regions with low reprojection errors, and seeks to leverage visual information in an image to expand masks + +that bear semantic regional meanings. The core philosophy of our methodology focuses on selecting robust regions for training without the need to manually define explicit areas or categories. Such robust regions might differ across scenes, which make it inappropriate for manual definition, as elaborated in Section 4. More specifically, we propose a strategy, named Error-Guided Feature Selection (EGFS), for deriving low reprojection error samples as point prompts, and expanding them to encompass a complete mask with similar semantic meanings. The masks are iteratively updated after being frozen for a specified period, which enable the proposed model to dynamically renew the focused regions. To achieve this prompt-based semantic regions expansion, we employ the Segment Anything Model (SAM) [1], a vision foundation model capable of providing a general understanding of the scene. This error-guided scheme ensures that our training process concentrates on regions with low reprojection errors and adaptively reduces the variability introduced by dynamic or texture-less objects. Moreover, EGPS adopts a predicted confidence map to refine the expanded error-guided mask for updates, which ensures the regions marked by the masks are sufficiently reliable. An illustrations of each component is provided in Fig. 1. + +To validate the effectiveness of our proposed method, we conduct evaluations on both outdoor and indoor datasets, including the Cambridge Landmarks dataset [18] and the Indoor6 [19] dataset, for evaluating the performance of our proposed SCR methodology. The experimental results suggest that our method can be broadly applied to previous methods and provides benefits, and leads to state-of-the-art (SOTA) performance when compared with SCR approaches without leveraging any 3D information in terms of estimated translational and rotational errors, while requiring less training time and smaller model sizes. We further perform a series of detailed analyses and ablation studies to confirm the effectiveness of our error-guided sampled regions and the proposed methodology. + +# 2 Related Work + +Scene Coordinate Regression. Conventional SCR approaches typically begin by establishing 2D-3D correspondences and then utilize RANSAC-based optimization to estimate the camera pose. To establish 2D-3D correspondences, previous endeavors have explored the use of random forests [5, 20-22] or convolutional neural networks (CNNs) [11-15,17,23] to regress scene coordinates from images. Since SCR typically requires scene-specific training, training time has emerged as a critical concern in practical applications. The ACE [17] method, which can be trained within five minutes, stands out as a promising solution. Moreover, unlike methods that map every 2D pixel to a 3D coordinate, some works [19,24,25] have aimed at learning to detect 3D landmarks to establish 2D-3D correspondences. However, they often require the reconstruction of a 3D model, which incurs additional computational time and costs. Our method seeks to utilize SAM to derive error-guided masks from 2D images, while maintaining the training efficiency. + +Emphasis on Robust Features for Localization. Several prior studies [24,26-31] in visual localization have discovered that not all regions within an image contribute + +equally to localization accuracy. As a result, these approaches have attempted to distinguish robust and invariant features, found that prioritizing the recognition and comprehension of those regions can enhance visual localization tasks more efficiently and effectively. However, a common limitation of these methods is their reliance on pre-defined semantics or the necessity of a 3D model, which could hinder their generalizability and practicality in real-world applications. + +# 3 Preliminary of Scene Coordinate Regression (SCR) + +SCR-based visual localization determines the camera pose of an RGB image $I \in \mathbb{R}^{H \times W \times 3}$ by predicting 3D scene coordinates $\mathcal{V}$ for a set of pixels and establishing the 2D-3D correspondences between pixel coordinates in the image and those predicted 3D scene coordinates. The primary objective of the training process is to establish a mapping function $f(\cdot)$ such that $\mathcal{V} = f(I)$ . The training data comprise RGB images paired with their respective camera poses, but do not necessarily include depth information. To generate accurate $\mathcal{V}$ , SCR methods often involve optimizing the reprojection error $\mathcal{R}$ , which is measured by projecting a 3D predicted scene coordinate $\mathrm{y_i} \in \mathcal{V}$ back onto its 2D pixel coordinate and calculating the discrepancy from the actual coordinate of the image patch, where i denotes the patch index. The reprojection error for each patch $\mathrm{r_i} \in \mathcal{R}$ can be formulated as $r(\mathrm{p_i}, \mathrm{y_i}, \mathrm{h^*}) = \| \mathrm{p_i} - K\mathrm{h^{*-1}}\mathrm{y_i}\|$ , where $K$ denotes the camera intrinsic matrix and $\mathrm{h^*}$ denotes the ground truth camera pose. After obtaining $\mathcal{V}$ , the camera pose h can be estimated according to the following equation: + +$$ +\mathrm {h} = g (\mathcal {C}) = g ^ {P n P} \left(\mathcal {C} _ {\mathcal {I}}\right), \text {w i t h} \mathcal {C} = \left\{\left(\mathrm {p} _ {\mathrm {i}}, \mathrm {y} _ {\mathrm {i}}\right) | \mathrm {p} _ {\mathrm {i}} \in I, \mathrm {y} _ {\mathrm {i}} \in \mathcal {Y} \right\}, \tag {1} +$$ + +where $g(\cdot)$ involves PnP with RANSAC, followed by Levenberg-Marquardt-based refinement, $\mathcal{C}$ represents the set of all 2D-3D correspondences and $\mathcal{C}_{\mathcal{I}}$ signifies inlier correspondences after refinement, which is a subset of $\mathcal{C}$ . Inlier correspondences denote those that are better consistent with the estimated camera pose, while correspondences that do not align well with the estimated camera pose are termed outliers. The final camera pose estimation is determined by the inlier correspondences $\mathcal{C}_{\mathcal{I}}$ . In the following paragraphs, we introduce two scene coordinate regression methods that have demonstrated superior performance. + +DSAC Variants. DSAC [13] introduces a differentiable variant of RANSAC, and enables end-to-end training of the entire pipeline. This design combines the benefits of RANSAC for handling outliers with the power of deep neural networks to learn complex patterns. In addition, DSAC++ [14] removes the requirement of depth information for SCR training and adopts robust loss terms to downweight outliers. DSAC* [15] further simplifies and refines DSAC++, and encourages the network to focus on reliable scene structures while ignoring outlier predictions. + +ACE. ACE [17] is a fast SCR method that can achieve promising localization accuracy within a short training time of only five minutes. ACE splits its regression network into a pre-trained scene-agnostic backbone $f_{B}$ , and a scene-specific + +![](images/5b6845a7066e395903dd28812392eed4dcb5061a06f9c0fc41e6f052db6cf123.jpg) + +![](images/020b608dc5d31389981574950ff0eb001465e1efe841fe1f00aa377e45ea2d5b.jpg) + +![](images/f613d160fd3f192fc7bde8f6fc65622be02fecfd61f9fad868e0a16619c39ef8.jpg) +Fig. 2: Analysis between reprojection error and semantic meaning. The analysis result indicates the regions with low reprojection errors tend to have higher inlier ratios, while the errors do not always align with specific semantic categories, e.g., "tree" and "rug". + +![](images/6606dbd3678efa9350336af18d56201d63984931797c331948b4e48336e946f9.jpg) + +multi-layer perceptron (MLP) $f_{H}$ . The backbone $f_{B}$ extracts feature vectors $\mathbf{f}_i$ from image patches $p_i$ and the MLP $f_{H}$ predicts the scene coordinates $y_{i}$ based on the extracted $\mathrm{f}_i$ . Between $f_{B}$ and $f_{H}$ , the features are shuffled to decorrelate gradients within a batch, which enables ACE to enhance its training efficiency. + +# 4 In-Depth Evaluation of Scene Coordinate Regression + +# 4.1 Challenges in Scene Coordinate Regression + +Several challenging factors have been identified that may compromise the accuracy of SCR. Two key challenges among them are dynamic objects and textureless surfaces [27, 32-34]. Dynamic objects exhibit characteristics of motion or disappearance across different frames, while textureless surfaces lack distinctive feature points or salient characteristics. These challenges highlight the complexity of achieving precise SCR in complex visual scenarios. To address them, previous SCR approaches [13-15, 17] have employed RANSAC to filter out outliers, and some researches [14, 15, 17] have attempted to utilize end-to-end learning to concentrate on reliable scene structures while down-weighting outlier predictions during training. Unfortunately, these prior endeavors failed to fully exploit the rich semantic information in the visual content, which can be crucial for enhancing the robustness and accuracy of camera pose estimation in complex scenes. Furthermore, RANSAC-based approaches for camera pose estimation could potentially fail in scenarios with insufficient number of inlier correspondences [35]. + +![](images/907825aa10e8fb607d365151a0f99a17944d3280758c178917bcbc4ec8b63e63.jpg) +Fig. 3: An overview of the training framework. + +# 4.2 Analysis between Reprojection Error and Semantic Meaning + +For a more comprehensive understanding of the correlation and relationship between reprojection errors and semantic meaning, we investigate the performance of ACE on the reprojection errors and the inlier correspondences selected by RANSAC for estimating the final camera pose during the inference process. First, image pixels are labeled by a semantic segmentation model ViT-Adapter [36]. Subsequently, the reprojection errors and inlier correspondences of ACE for each semantic class are calculated. Fig. 2 presents the median reprojection errors and the ratios of inlier correspondences for the top ten most frequent classes in each scene. The left two figures correspond to the outdoor dataset Cambridge Landmarks [18], while the right two figures pertain to the indoor dataset Indoor6 [19]. It is observed that regions with low reprojection errors tend to have relatively higher inlier ratios, indicating a greater impact on the estimated camera poses. As a result, improving the accuracy of low-error areas could lead to more precise camera pose predictions. Nevertheless, it is also observed that the reprojection errors for the same class can vary significantly across different scenes. For example, in the Cambridge Landmarks dataset, the error for "tree" is low in King's College but high in Old Hospital. Similarly, in the Indoor6 dataset, the error for "rug" is high in scene1 but low in scene3. This indicates that low reprojection errors do not always align with specific semantic categories, which presents a challenge in directly using certain specific semantic segmentation labels for area selection. Motivated by these insights, the primary objective of this study is to develop a methodology that does not rely on pre-defined semantic categories to identify low reprojection error areas for enhancing localization accuracy in SCR. + +# 5 Methodology + +# 5.1 Problem Definition and Framework Overview + +Fig. 3 provides an overview of the proposed training framework for SCR, which iteratively samples features to train a scene-specific MLP. The scene-specific MLP consists of a scene coordinate head and a confidence head. In each iteration, the model is trained for $k$ epochs. During the initial iteration, features are randomly sampled from all parts of images in order to derive the first set + +of reprojection errors. In subsequent iterations, features are selected based on error-guided feature selection (EGFS) masks generated according to reprojection errors and a confidence map. This iterative mechanism enables the model to dynamically update its focus areas throughout the training process. During the inference phase, the network estimates scene coordinates $\mathcal{V}$ and a confidence map for the refinement, while the masks generated during training are not required for the inference phase. The inference procedure is depicted in Fig. 4. + +# 5.2 Error-Guided Feature Selection (EGFS) with SAM + +In alignment with the discussion in Section 4 on identifying potential areas contributing to the final camera pose estimation during the training stage, we employ the Segmentation Anything Model (SAM) [1], which is capable of leveraging diverse visual cues to generate high-quality object masks. Specifically, in this study, SAM is utilized to extend point prompts with low reprojection errors to form a complete mask sharing a similar semantic context. The rationale behind this approach is the high likelihood that these comprehensive semantic regions correspond to low error areas and can contribute to final pose estimation. Such a concept harnesses the capabilities of the foundational SAM model to uncover more viable areas, and hence, enabling a thorough understanding of the image context. In our experiments, we utilize EfficientViT-SAM-L0 [37], a lightweight and efficient variant of SAM, to identify salient regions by selecting the $\tau$ percentage of points with the lowest reprojection errors as point prompts, where $\tau$ is adjustable and set to ten in our case. After obtaining the error-guided mask from the SAM model, the predicted confidence map $c$ is utilized to refine the error-guided mask to ensure the points marked by the mask is sufficiently reliable. Specifically, in each image, the points with confidence scores below a threshold $\sigma$ are filtered out. The design for confidence refinement and its optimization are described in Section 5.3. In each iteration, the refined error-guided masks serve as guidance for training the scene-specific MLP. Only the features corresponding to the regions selected by the masks are sampled into the training buffer. This selective inclusion is facilitated by the nature of $f_{H}$ [17], which employs $1\times 1$ convolutions to treat and process each selected sample independently using the same set of shared weights. The proposed feature selection mechanism ensures that the training process focuses on the crucial regions identified in the current iteration, and therefore enables the model to enhance its overall performance. + +# 5.3 Scene Coordinate and EGFS Refinement with Confidence + +In order to further refine the error-guided mask and ensure the selected points are of sufficiently high quality for predicting accurate camera pose, we incorporate the design of a confidence map [38, 39]. Specifically, we predict a confidence score pixel-wise from the confidence head and optimize it jointly with the reprojection error. This confidence head replaces the last three $1 \times 1$ convolution layers of the original scene-specific MLP with two $1 \times 1$ convolution layers for + +![](images/ca54741ca98085209b7b2368001f7d1a5da12a6c2833602c004768ea58b3ea06.jpg) +Fig. 4: An overview of the inference procedure. + +confidence score prediction, thus enabling adaptive weights for the reprojection loss at each position. The training loss used by our method is formulated as: + +$$ +\ell \left(\mathrm {p} _ {\mathrm {i}}, \mathrm {y} _ {\mathrm {i}}, \mathrm {h} _ {\mathrm {i}} ^ {*}\right) = \left\{ \begin{array}{l l} \mathrm {c} _ {\mathrm {i}} \cdot \hat {r} \left(\mathrm {p} _ {\mathrm {i}}, \mathrm {y} _ {\mathrm {i}}, \mathrm {h} _ {\mathrm {i}} ^ {*}\right) - \alpha \log \mathrm {c} _ {\mathrm {i}} & \text {i f y} _ {\mathrm {i}} \in \mathcal {V} \\ \| \mathrm {y} _ {\mathrm {i}} - \bar {\mathrm {y}} _ {\mathrm {i}} \| _ {0} - \alpha \log (1 - \mathrm {c} _ {\mathrm {i}}) & \text {o t h e r w i s e} \end{array} , \right. \tag {2} +$$ + +where $\hat{r} (\mathrm{p_i},\mathrm{y_i},\mathrm{h_i^*})$ represents a tanh clamping of the reprojection error that changes over time, $\mathcal{V}$ denotes the set of 2D pixels with valid scene coordinate predictions, as explained in [15, 17], $c_{i}$ is the confidence score at pixel i, $\alpha$ is a hyperparameter that balances the confidence regularization term, and $\bar{\mathbf{y}}_{\mathrm{i}}$ is a dummy scene coordinate derived from the ground truth camera pose, assuming a constant image depth of $10\mathrm{m}$ . When the scene coordinate prediction is valid, the model down-weights the reprojection errors in challenging regions using confidence scores and focuses more on regions with reliable predictions. On the other hand, when the scene coordinate prediction is invalid, the model is incentivized to lower the confidence scores for inaccurate predictions. The confidence score is also used in the inference phase to select reliable scene coordinates. For each query image, the median of the confidence scores is calculated, and only scene coordinates with confidence greater than this median are selected. This thresholding ensures that only the most reliable 2D-3D correspondences are used to generate the camera pose by the PnP with RANSAC pose solver. By focusing on high-confidence correspondences, a more accurate camera pose can be estimated. + +# 6 Experimental Results + +# 6.1 Experimental Setups + +Datasets. Our evaluation of the proposed method was conducted on two representative datasets: the outdoor dataset Cambridge Landmarks [18] and the indoor dataset Indoor6 [19]. The details of the datasets are described as follows: + +Cambridge Landmarks. The Cambridge Landmarks Dataset [18] is a renowned outdoor dataset extensively utilized for visual localization tasks, which includes five distinct scenes. Every scene in this dataset is composed of RGB images along with their respective ground truth camera poses reconstructed using the SfM technique. It includes a broad spectrum of conditions and various viewpoints, and is ideal for evaluating the resilience and effectiveness of various SCR methods. + +Table 1: A comparison of model sizes, training times, and median translation and rotation errors (in $\mathrm{cm} / {}^{\circ}$ ) evaluated on the Cambridge Landmarks dataset. 'dual' and 'quad.' denote the ensemble versions of two and four models, respectively. + +
MethodSizeMapping TimeKing's College Great Court Old Hospital Shop Facade St Mary's ChurchAverage
FMhLoc (SP+SG) [4,6,8]~800MB~35 min12/0.216/0.115/0.34/0.27/0.211/0.2
pixLoc [26]~600MB30/0.114/0.216/0.35/0.210/0.315/0.2
SCR (w/ 3D model)DSAC* [15]28MB15 hr15/0.349/0.321/0.45/0.313/0.421/0.3
FocusTune [31]4MB6 min19/0.338/0.118/0.46/0.315/0.519/0.3
FocusTune (quad.) [31]16MB24 min15/0.329/0.117/0.45/0.29/0.315/0.3
NeuMap [40]--14/0.26/0.119/0.36/0.217/0.512/0.3
SACReg-L [41]--11/0.213/0.113/0.26/0.25/0.310/0.2
SCRDSAC* [15]28MB15 hr18/0.334/0.221/0.45/0.315/0.619/0.4
ACE [17]4MB5 min28/0.443/0.231/0.65/0.318/0.625/0.4
ACE (quad.) [17]16MB20 min18/0.328/0.125/0.55/0.39/0.317/0.3
EGFS4.5MB12 min14/0.331/0.121/0.45/0.315/0.517/0.3
EGFS (dual)9MB21 min14/0.328/0.119/0.45/0.210/0.315/0.3
+ +Table 2: A comparison of model sizes, training times, and the proportions of translation and rotation errors that are below $5\mathrm{cm} / 5^{\circ}$ evaluated on the Indoor6 dataset. + +
MethodSizeMapping Timescene1scene2ascene3scene4ascene5scene6Average
FMhLoc [4]~1.5GB~3.3 hr70.5%52.1%86.0%75.3%58.0%86.7%71.4%
SCR (w/ 3D model)DSAC* [15]28MB15 hr18.7%28.0%19.7%60.8%10.6%44.3%30.4%
SLD* (300 landmarks) [25]15MB~5.5 hr47.2%48.2%56.2%67.7%33.7%52.0%50.8%
SLD* (1000 landmarks) [25]120MB~44 hr68.5%62.6%76.2%77.2%57.8%78.0%70.1%
SCRDSAC* [15]28MB15 hr23.0%33.9%26.0%67.1%10.6%50.2%35.1%
ACE [17]4MB5 min26.0%32.3%31.4%62.0%14.2%47.4%35.6%
ACE (quad.) [17]16MB20 min52.9%52.5%62.9%69.6%31.1%82.4%58.6%
EGFS4.5MB21 min46.4%60.6%56.4%78.7%22.8%71.6%56.1%
EGFS (dual)9MB30 min58.5%59.1%67.0%76.1%30.6%75.9%61.2%
+ +Indoor6. The Indoor6 [19] dataset comprises six distinct indoor scenes captured over several days, with ground truth camera poses computed using COLMAP [42]. Each scene includes multiple rooms and contains illumination variations, making it challenging for different types of visual localization tasks. + +Implementation Details. The training process of our methodology includes 20 epochs, with the generation of masks every five epochs. We set the confidence regularization parameter $\alpha$ to ten and the reprojection error threshold $\tau$ to ten. For mask refinement, the parameter $\sigma$ is set to the median confidence score of each image. Our method builds upon the ACE [17] architecture as the backbone, and thus, the remaining hyperparameters are kept the same as those used in [17]. + +# 6.2 Results on the Cambridge Landmarks and Indoor6 Dataset + +Cambridge Landmarks Dataset. Table 1 presents the experimental results evaluated on the Cambridge Landmarks dataset, which affirm the effectiveness and efficiency of our proposed methodology. It can be observed that EGFS maintains a model size similar to ACE, with only a 0.5MB increase for the confidence head, while reducing average translational and rotational errors. The results further reveal that EGFS not only surpasses DSAC* in performance and performs + +Table 3: Comparison of the effectiveness between (a) directly sampling solely based on pixel points with low reprojection errors, and (b) sampling through the EGFS masks. + +
DatasetSceneri < Q0.3(R) ri < Q0.4(R) ri < Q0.5(R) ri < Q0.6(R) ri < Q0.7(R)EGFS Masks
CambridgeKing's College20/0.319/0.320/0.320/0.320/0.314/0.3
Great Court39/0.237/0.136/0.238/0.239/0.231/0.1
Old Hospital26/0.525/0.523/0.525/0.524/0.521/0.4
Shop Facade6/0.36/0.36/0.36/0.35/0.35/0.3
St Mary's Church17/0.517/0.516/0.517/0.516/0.515/0.5
Average22/0.421/0.420/0.421/0.421/0.417/0.3
Indoor6scene134.4%34.9%34.9%32.0%30.2%46.4%
scene2a39.8%40.6%46.1%41.3%42.5%60.6%
scene345.2%43.9%44.9%41.3%42.3%56.4%
scene4a63.2%63.9%62.6%63.2%67.1%78.7%
scene519.5%18.1%17.1%18.3%16.9%22.8%
scene659.7%60.3%60.0%57.2%58.1%71.6%
Average43.6%43.6%44.3%42.2%42.9%56.1%
+ +on par with the ensemble version of ACE (i.e., ACE (quad.)), but also features a smaller model size and reduced training time. Furthermore, our model's ensemble version (i.e., EGFS (dual.)) is able to exceed the performance of DSAC* and ACE (quad.) while requiring a smaller model size and less training time. + +Indoor6 Dataset. Table 2 presents the evaluation results on Indoor6. EGFS significantly outperforms DSAC* and ACE across all scenes, while also achieving comparable performance to ACE (quad.) with a much smaller model size (4.5MB compared to 16MB). Moreover, EGFS (dual.) demonstrates superior performance relative to ACE (quad.), which further substantiates the effectiveness of EGFS in optimization through error-guided masks and confidence refinement. + +# 6.3 Effectiveness of Error-Guided Feature Selection + +To further substantiate the advantages of employing the EGFS approach, this experiment compares several baseline schemes that sample points with different reprojection error thresholds against EGFS. The results are summarized in Table 3. Specifically, these baseline schemes select sample points based on various quantiles of the reprojection error map, ranging from $30\%$ to $70\%$ , denoted as $Q_{0.3}(R)$ through $Q_{0.7}(R)$ , respectively. In each baseline scheme, a point is chosen if its reprojection error $r$ falls below the threshold specified by $Q_{\text{quantile}}(R)$ . The results suggest that relying solely on reprojection errors does not guarantee optimal learning of scene coordinates. This is due to the fact that areas of low reprojection errors can be scattered and may not encompass entire semantic objects. In contrast, the proposed EGFS expands the points into complete semantic masks followed by confidence refinement, which enables our model to determine scene coordinates with minimal transitional and rotational errors. This experiment thus confirms the effectiveness of leveraging SAM's ability to interpret image context alongside confidence maps for enhanced sample point selection. + +![](images/cdd7b19eca1bfb808f3163421798963e98e839761a8fc757fb66044a3767915a.jpg) +Fig. 5: Visualization of the EGFS mask refinement process at every five epochs, which depicts the reprojection errors at the beginning (epoch 5) and the end (epoch 20), as well as the refined error-guided masks used throughout training. The red dots represent low reprojection errors that serve as prompts, while the light green overlay denotes the refined EGFS masks. It can be observed that the EGFS masks enhances over epochs. + +![](images/f5e2853b2307b1a8e4bf9ebd579a2e93d56e6ab2d967d8d51020e89a3c472809.jpg) +Fig. 6: Visualization of the estimated camera pose trajectories from testing sequences with the camera frustums colored based on translational errors. Pose errors denoted. + +# 6.4 Qualitative Results + +Iterative Error-guided Feature Selection Mask. Fig. 5 illustrates the refinement of EGFS masks every five epochs and compares the reprojection errors at the beginning and the end. It can be observed that EGFS is not only capable of expanding points with low reprojection errors (i.e., $\mathrm{r} < Q_{0.1}(R)$ ) into regions that bear similar semantic meanings, but also refines the masks with a confidence map to exclude uncertain areas. The masks are observed to be iteratively enhanced from coarse to fine. In the beginning (i.e., at epoch five), the reprojection errors appear noisy, whereas by the end (i.e., at epoch 20), the EGFS masks become more refined and concentrated on key regions beneficial for SCR. + +![](images/bbeac4a9f655ceae17f344e5c6b90cfeb9afe3b17e0ce04d08638b8591f46760.jpg) +Fig. 7: Visualization of point clouds reconstructed from estimated scene coordinates of the training sequence. The point clouds derived from the scene coordinates estimated by the model trained with the proposed EGFS are clearer compared to those from the ACE model, even without the application of refined EGFS masks at inference time. + +Visualization of the Estimated Camera Poses. In this section, we present the estimated camera poses from the testing sequences, and compare the proposed EGFS with the ACE baseline. The results are illustrated in Fig. 6. To visualize camera trajectories, we connect consecutive camera positions with lines and indicate camera orientations with frustums. Each sampled camera pose is associated with a color to represent the translation error. It is observed that our proposed EGFS can estimate more accurate camera poses as compared to ACE. + +Scene Coordinates and 3D Point Clouds w/ and w/o EFGS. To validate the effectiveness of EGFS qualitatively and its impact on the quality of estimated scene coordinates that learned from training sequence, we depict the 3D point cloud generated from the scene coordinates from the training sequence, and apply colors based on the corresponding queried pixels in Fig. 7. To compare the 3D point clouds reconstructed from the scene coordinates, we compare the baseline ACE approach with the proposed EGFS approach. For our approach, we visualize the point clouds both before and after the application of EGFS masks and confidence maps when collecting the scene coordinates. Please note that both before and after versions which shown in the figure are trained with EGFS refinement. The 3D point cloud constructed by the proposed approach appears clearer, in contrast to the ACE-generated point cloud, which exhibits blurriness due to inaccurate scene coordinate estimation. In addition, the point cloud reconstructed from raw scene coordinates predicted without EGFS from our approach contains noise and floaters due to the inclusion of scene coordinates + +![](images/2e91c655bb39e7fbc02620fda7ca4f6f5a8bd629fd83d7ebe782d25be6719296.jpg) +Fig. 8: Visualization of the estimated scene coordinates and the confidence maps on unseen samples. The RGB color in the scene coordinates represents the XYZ coordinates. The point clouds are colored green for selected areas and red for rejected areas. + +with low confidence scores. On the other hand, the point cloud reconstructed from the scene coordinates with EGFS refinement exhibits clearer contours and less noise. This justifies EGFS's ability to eliminate areas of low confidence values and produce high-quality dense scene coordinates. + +Visualization of Estimated Confidence Map on Unseen Samples. In this section, we qualitatively evaluate the confidence maps estimated from the testing sequence by visualizing them in both the 2D image plane and as 3D projections onto the point cloud. The visualized point cloud is reconstructed from the training sequence, whereas the plotted samples are selected from the testing sequence and were not included in the training procedure. These visualizations are provided in Fig. 8. It can be observed that the estimated confidence map assists in rejecting scene coordinates located in areas where accurate prediction is challenging, such as in the air or on texture-less surfaces. + +# 6.5 Ablation Study + +Error-Guided Feature Selection (EGFS) and Confidence Refinement. We first present an ablation analysis to validate the effectiveness of the EGFS masks and confidence refinement in enhancing our method's performance. Table 4 reports the average results on the Cambridge Landmarks and Indoor6 datasets. The results demonstrate that the model trained with the proposed EGFS mechanism outperforms those trained without it, which indicates that focusing on training with robust features leads to improved performance. Furthermore, the incorporation of the confidence map also enhances the overall + +Table 4: Ablation study on the impact of each proposed components. + +
Error-Guided MasksConfidence RefinementCambridge (cm/°)Indoor6 (%)
XX25/0.435.6
✓X19/0.341.6
X✓19/0.450.9
✓✓17/0.356.1
+ +Table 5: Impact of varying proportions of point prompts $(\tau)$ + +
Proportions of point prompts (τ)Cambridge (cm/°)Indoor6 (%)
517/0.354.3
1017/0.356.1
1517/0.355.4
2017/0.355.3
+ +performance. The final row of Table 4 demonstrates that combining both techniques (i.e., our EGFS) achieves the best results, which further confirms the effectiveness of EGFS in improving localization accuracy. + +Analysis of Proportions of Point Prompts. We evaluate the performance of different proportions of point prompts used for expanding into EGFS masks, and the results are reported in Table 5. It can be observed that the proportions of point prompts do not significantly impact the performance on the Cambridge Landmark dataset. This may be attributed to the fact that the scenes in the dataset typically feature a primary architectural structure with fewer complicated details in the surroundings. On the other hand, for the Indoor6 dataset, the proportions of point prompts significantly affect the performance. The rationale behind these observations is that the indoor scenes are more complicated, and feature various small items that necessitate a larger number of low-error point prompts to comprehensively capture the entire scene structure. Please note that we employ $\tau = 10$ as our setting for all the experiments presented. + +# 7 Conclusion + +This paper addressed key challenges in SCR for visual localization, and specifically focused on the impacts of dynamic objects and texture-less regions. Our approach introduced an innovative EGFS mechanism through the use of SAM and confidence maps to enhance the performance of SCR. This technique effectively filtered out problematic areas by concentrating on regions with low reprojection errors. Moreover, we used confidence maps to further refine the pixels selected by EGFS and perform this process iteratively to enable dynamic updates of the focused regions. The experimental results on the Cambridge Landmarks and Indoor6 datasets suggest that our method can provide improvements in terms of training efficiency, model size, and accuracy. Our findings highlighted the importance of carefully selecting low-reprojection error pixels by taking into account semantic information and confidence scores. Furthermore, we quantitatively and qualitatively demonstrated that dynamically updating the masks enables more robust selection of error points, thus enabling better training efficiency and effectiveness. Our ablation studies validated the effectiveness of the techniques adopted in our method, and solidified their roles in enhancing the performance. + +# Acknowledgements + +The authors gratefully acknowledge the support from the National Science and Technology Council (NSTC) in Taiwan under grant numbers MOST 111-2223-E007-004-MY3, 113-2221-E-007-122-MY3, 113-2640-E-002-003, 113-2221-E-007-104-MY3, as well as the financial support from Woven by Toyota, Inc., Japan. The authors would also like to express their appreciation for the donation of the GPUs from NVIDIA Corporation and NVIDIA AI Technology Center (NVAITC) used in this work. Furthermore, the authors extend their gratitude to the National Center for High-Performance Computing (NCHC) for providing the necessary computational and storage resources. + +# References + +1. Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Dlarr, and Ross Girshick. Segment anything. arXiv:2304.02643, 2023. +2. Xiao-Shan Gao, Xiao-Rong Hou, Jianliang Tang, and Hang-Fei Cheng. Complete solution classification for the perspective-three-point problem. IEEE transactions on pattern analysis and machine intelligence, 25(8):930-943, 2003. +3. Martin A Fischler and Robert C Bolles. Random sample consensus: a paradigm for model fitting with applications to image analysis and automated cartography. Communications of the ACM, 1981. +4. Paul-Edouard Sarlin, Cesar Cadena, Roland Siegwart, and Marcin Dymczyk. From coarse to fine: Robust hierarchical localization at large scale. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2019. +5. Jamie Shotton, Ben Glocker, Christopher Zach, Shahram Izadi, Antonio Criminisi, and Andrew Fitzgibbon. Scene coordinate regression forests for camera relocalization in RGB-D images. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2013. +6. Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Superpoint: Self-supervised interest point detection and description. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), pages 224-236, 2018. +7. Jiaming Sun, Zehong Shen, Yuang Wang, Hujun Bao, and Xiaowei Zhou. LoFTR: Detector-free local feature matching with transformers. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), pages 8922-8931, 2021. +8. Paul-Edouard Sarlin, Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. SuperGlue: Learning feature matching with graph neural networks. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2020. +9. Philipp Lindenberger, Paul-Edouard Sarlin, and Marc Pollefeys. LightGlue: Local feature matching at light speed. arXiv preprint arXiv:2306.13643, 2023. +10. Pablo Speciale, Johannes L Schonberger, Sing Bing Kang, Sudipta N Sinha, and Marc Pollefeys. Privacy preserving image-based localization. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), pages 5493-5503, 2019. +11. Xiaotian Li, Juha Ylioinas, Jakob Verbeek, and Juho Kannala. Scene coordinate regression with angle-based reprojection loss for camera relocalization. In Proc. European Conf. on Computer Vision Workshop (ECCVW), 2018. +12. Xiaotian Li, Juha Ylioinas, and Juho Kannala. Full-frame scene coordinate regression for image-based localization. RSS, 2018. + +13. Eric Brachmann, Alexander Krull, Sebastian Nowozin, Jamie Shotton, Frank Michel, Stefan Gumhold, and Carsten Rother. DSAC-Differentiable RANSAC for camera localization. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2017. +14. Eric Brachmann and Carsten Rother. Learning less is more - 6D camera localization via 3D surface regression. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2018. +15. Eric Brachmann and Carsten Rother. Visual camera re-localization from RGB and RGB-D images using DSAC. TPAMI, 2021. +16. Eric Brachmann and Carsten Rother. Expert sample consensus applied to camera re-localization. In Proc. IEEE Int. Conf. on Computer Vision (ICCV), 2019. +17. Eric Brachmann, Tommaso Cavallari, and Victor Adrian Prisacariu. Accelerated coordinate encoding: Learning to relocalize in minutes using RGB and poses. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2023. +18. Alex Kendall, Matthew Grimes, and Roberto Cipolla. PoseNet: A convolutional network for real-time 6-DOF camera relocalization. In Proc. IEEE Int. Conf. on Computer Vision (ICCV), December 2015. +19. Tien Do, Ondrej Miksik, Joseph DeGol, Hyun Soo Park, and Sudipta N. Sinha. Learning to detect scene landmarks for camera localization. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2022. +20. Abner Guzman-Rivera, Pushmeet Kohli, Ben Glocker, Jamie Shotton, Toby Sharp, Andrew Fitzgibbon, and Shahram Izadi. Multi-output learning for camera localization. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2014. +21. Eric Brachmann, Frank Michel, Alexander Krull, Michael Ying Yang, Stefan Gumhold, and carsten Rother. Uncertainty-driven 6D pose estimation of objects and scenes from a single RGB image. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2016. +22. Julien Valentin, Matthias Niessner, Jamie Shotton, Andrew Fitzgibbon, Shahram Izadi, and Philip H. S. Torr. Exploiting uncertainty in regression forests for accurate camera relocalization. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2015. +23. Tao Xie, Kun Dai, Siyi Lu, Ke Wang, Zhiqiang Jiang, Jinghan Gao, Dedong Liu, Jie Xu, Lijun Zhao, and Ruifeng Li. OFVL-MS: Once for visual localization across multiple indoor scenes. In Proc. IEEE Int. Conf. on Computer Vision (ICCV), 2023. +24. Zhaoyang Huang, Han Zhou, Yijin Li, Bangbang Yang, Yan Xu, Xiaowei Zhou, Hujun Bao, Guofeng Zhang, and Hongsheng Li. VS-Net: Voting with segmentation for visual localization. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2021. +25. Tien Do and Sudipta N. Sinha. Improved scene landmark detection for camera localization. In Proceedings of the International Conference on 3D Vision (3DV), 2024. +26. Paul-Edouard Sarlin, Ajaykumar Unagar, Mans Larsson, Hugo Germain, Carl Toft, Victor Larsson, Marc Pollefeys, Vincent Lepetit, Lars Hammarstrand, Fredrik Kahl, and Torsten Sattler. Back to the Feature: Learning Robust Camera Localization from Pixels to Pose. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2021. +27. Mohammad Altillawi. PixSelect: Less but reliable pixels for accurate and efficient localization. In Proc. IEEE Int. Conf. on Robotics and Automation (ICRA), pages 4156-4162. IEEE, 2022. + +28. Fei Xue, Ignas Budvytis, Daniel Olmeda Reino, and Roberto Cipolla. Efficient large-scale localization by global instance recognition. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), pages 17348-17357, 2022. +29. Fei Xue, Ignas Budvytis, and Roberto Cipolla. SFD2: Semantic-guided feature detection and description. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), pages 5206-5216, 2023. +30. Bach-Thuan Bui, Dinh-Tuan Tran, and Joo-Ho Lee. D2S: Representing local descriptors and global scene coordinates for camera relocalization. arXiv preprint arXiv:2307.15250, 2023. +31. Son Tung Nguyen, Alejandro Fontan, Michael Milford, and Tobias Fischer. Focus-Tune: Tuning visual localization through focus-guided sampling. In Proc. IEEE Winter Conf. on Applications of Computer Vision (WACV), pages 3606-3615, January 2024. +32. Johanna Wald, Torsten Sattler, Stuart Golodetz, Tommaso Cavallari, and Federico Tombari. Beyond controlled environments: 3d camera re-localization in changing indoor scenes. In Proc. European Conf. on Computer Vision (ECCV), pages 467-487, 2020. +33. Siyan Dong, Qingnan Fan, He Wang, Ji Shi, Li Yi, Thomas Funkhouser, Baoquan Chen, and Leonidas J Guibas. Robust neural routing through space partitions for camera relocalization in dynamic indoor environments. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2021. +34. Donghwan Lee, Soohyun Ryu, Suyong Yeon, Yonghan Lee, Deokhwa Kim, Cheolho Han, Yohann Cabon, Philippe Weinzaepfel, Nicolas Guerin, Gabriela Csurka, and Martin Humenberger. Large-scale localization datasets in crowded indoor spaces. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), pages 3227-3236, June 2021. +35. Hongyi Fan, Joe Kileel, and Benjamin Kimia. On the instability of relative pose estimation and RANSAC's role. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), pages 8935-8943, June 2022. +36. Zhe Chen, Yuchen Duan, Wenhai Wang, Junjun He, Tong Lu, Jifeng Dai, and Yu Qiao. Vision transformer adapter for dense predictions. arXiv preprint arXiv:2205.08534, 2022. +37. Han Cai, Junyan Li, Muyan Hu, Chuang Gan, and Song Han. EfficientViT: Lightweight multi-scale attention for high-resolution dense prediction. In Proc. IEEE Int. Conf. on Computer Vision (ICCV), pages 17302-17313, 2023. +38. Alex Kendall and Yarin Gal. What uncertainties do we need in bayesian deep learning for computer vision? Advances in neural information processing systems, 30, 2017. +39. Sheng Wan, Tung-Yu Wu, Wing H Wong, and Chen-Yi Lee. Confnet: predict with confidence. In 2018 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 2921-2925. IEEE, 2018. +40. Shitao Tang, Sicong Tang, Andrea Tagliasacchi, Ping Tan, and Yasutaka Furukawa. Neumap: Neural coordinate mapping by auto-transdecoder for camera localization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 929-939, June 2023. +41. Revaud, Jérôme and Cabon, Yohann and Brégier, Romain and Lee, JongMin and Weinzaepfel, Philippe. SACReg: Scene-Agnostic Coordinate Regression for Visual Localization, 2023. +42. Johannes L. Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), June 2016. \ No newline at end of file diff --git a/2024/Reprojection Errors as Prompts for Efficient Scene Coordinate Regression/images.zip b/2024/Reprojection Errors as Prompts for Efficient Scene Coordinate Regression/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..f997af06444c699e9fac6a9037bc816d6cc9669f --- /dev/null +++ b/2024/Reprojection Errors as Prompts for Efficient Scene Coordinate Regression/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e021d6f9d212dd06e4e518bca0f4fbe878edbf0062070a9c997d958ac2298e71 +size 667701 diff --git a/2024/Reprojection Errors as Prompts for Efficient Scene Coordinate Regression/layout.json b/2024/Reprojection Errors as Prompts for Efficient Scene Coordinate Regression/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..3bd52325b442b61347284a880df8dba99b3d19d7 --- /dev/null +++ b/2024/Reprojection Errors as Prompts for Efficient Scene Coordinate Regression/layout.json @@ -0,0 +1,7546 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 152, + 111, + 462, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 111, + 462, + 148 + ], + "spans": [ + { + "bbox": [ + 152, + 111, + 462, + 148 + ], + "type": "text", + "content": "Reprojection Errors as Prompts for Efficient Scene Coordinate Regression" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 143, + 167, + 470, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 167, + 470, + 193 + ], + "spans": [ + { + "bbox": [ + 143, + 167, + 470, + 193 + ], + "type": "text", + "content": "Ting-Ru Liu" + }, + { + "bbox": [ + 143, + 167, + 470, + 193 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 143, + 167, + 470, + 193 + ], + "type": "text", + "content": ", Hsuan-Kung Yang" + }, + { + "bbox": [ + 143, + 167, + 470, + 193 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 143, + 167, + 470, + 193 + ], + "type": "text", + "content": ", Jou-Min Liu" + }, + { + "bbox": [ + 143, + 167, + 470, + 193 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 143, + 167, + 470, + 193 + ], + "type": "text", + "content": ", Chun-Wei Huang" + }, + { + "bbox": [ + 143, + 167, + 470, + 193 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 143, + 167, + 470, + 193 + ], + "type": "text", + "content": ", Tsung-Chih Chiang" + }, + { + "bbox": [ + 143, + 167, + 470, + 193 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 143, + 167, + 470, + 193 + ], + "type": "text", + "content": ", Quan Kong" + }, + { + "bbox": [ + 143, + 167, + 470, + 193 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 143, + 167, + 470, + 193 + ], + "type": "text", + "content": ", Norimasa Kobori" + }, + { + "bbox": [ + 143, + 167, + 470, + 193 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 143, + 167, + 470, + 193 + ], + "type": "text", + "content": ", and Chun-Yi Lee" + }, + { + "bbox": [ + 143, + 167, + 470, + 193 + ], + "type": "inline_equation", + "content": "^{1}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 150, + 201, + 463, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 150, + 201, + 463, + 247 + ], + "spans": [ + { + "bbox": [ + 150, + 201, + 463, + 247 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 150, + 201, + 463, + 247 + ], + "type": "text", + "content": "National Tsing Hua University " + }, + { + "bbox": [ + 150, + 201, + 463, + 247 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 150, + 201, + 463, + 247 + ], + "type": "text", + "content": "Woven by Toyota, Inc. {tingru, diasimui, alan, george, cymaxwelllee}@elsa.cs.nthu.edu.tw {hsuan-kung.yang, quan.kong, norimasa.kobori} @woven toyota https://tingru0203.github.io/egfs" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 160, + 277, + 455, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 277, + 455, + 443 + ], + "spans": [ + { + "bbox": [ + 160, + 277, + 455, + 443 + ], + "type": "text", + "content": "Abstract. Scene coordinate regression (SCR) methods have emerged as a promising area of research due to their potential for accurate visual localization. However, many existing SCR approaches train on samples from all image regions, including dynamic objects and texture-less areas. Utilizing these areas for optimization during training can potentially hamper the overall performance and efficiency of the model. In this study, we first perform an in-depth analysis to validate the adverse impacts of these areas. Drawing inspiration from our analysis, we then introduce an error-guided feature selection (EGFS) mechanism, in tandem with the use of the Segment Anything Model (SAM) [1]. This mechanism seeds low reprojection areas as prompts and expands them into error-guided masks, and then utilizes these masks to sample points and filter out problematic areas in an iterative manner. The experiments demonstrate that our method outperforms existing SCR approaches that do not rely on 3D information on the Cambridge Landmarks and Indoor6 datasets." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 464, + 230, + 475 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 464, + 230, + 475 + ], + "spans": [ + { + "bbox": [ + 133, + 464, + 230, + 475 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 489, + 482, + 646 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 489, + 482, + 646 + ], + "spans": [ + { + "bbox": [ + 130, + 489, + 482, + 646 + ], + "type": "text", + "content": "The objective of visual localization is to estimate a 6-DoF camera pose from images, a key component in fields such as Augmented Reality, Virtual Reality, and autonomous driving. Contemporary leading methods in visual localization typically involve establishing 2D-3D correspondences and then utilizing Perspective-Point (PnP) [2] with RANSAC [3] for camera pose estimation. These methods can be broadly classified into two main directions: feature-matching [4] and scene coordinate regression (SCR) [5]. Feature-matching approaches reconstruct a 3D scene using Structure from Motion (SfM), identify and describe key points in 2D images [6,7], and link these to 3D coordinates [8,9]. Nevertheless, they may encounter challenges such as high computational demands, significant storage requirements, and potential privacy concerns [10]. On the other hand, SCR methods [11-17] employ deep neural networks (DNNs) to predict the 3D coordinates of pixels and then utilize PnP with RANSAC for camera pose estimation, which" + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 653, + 259, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 653, + 259, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 653, + 259, + 665 + ], + "type": "text", + "content": "* indicates equal contribution." + } + ] + } + ], + "index": 6 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 142, + 215, + 199 + ], + "blocks": [ + { + "bbox": [ + 133, + 142, + 215, + 199 + ], + "lines": [ + { + "bbox": [ + 133, + 142, + 215, + 199 + ], + "spans": [ + { + "bbox": [ + 133, + 142, + 215, + 199 + ], + "type": "image", + "image_path": "81109ee5448de08a173c3c5cdd9c5f07af530f24c707302e16cb5843c84ea9a8.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 216, + 141, + 296, + 199 + ], + "blocks": [ + { + "bbox": [ + 216, + 141, + 296, + 199 + ], + "lines": [ + { + "bbox": [ + 216, + 141, + 296, + 199 + ], + "spans": [ + { + "bbox": [ + 216, + 141, + 296, + 199 + ], + "type": "image", + "image_path": "883e196dd6e558123f277c64d848b1800e7d3b73c6cb98ec4f45f7f1ca1cc253.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 296, + 141, + 383, + 199 + ], + "blocks": [ + { + "bbox": [ + 296, + 141, + 383, + 199 + ], + "lines": [ + { + "bbox": [ + 296, + 141, + 383, + 199 + ], + "spans": [ + { + "bbox": [ + 296, + 141, + 383, + 199 + ], + "type": "image", + "image_path": "dcd7b4046dfa769bf8dec1e588187ab33e4d3e3f807d1a020d7117ab844414ba.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 385, + 142, + 481, + 199 + ], + "blocks": [ + { + "bbox": [ + 385, + 142, + 481, + 199 + ], + "lines": [ + { + "bbox": [ + 385, + 142, + 481, + 199 + ], + "spans": [ + { + "bbox": [ + 385, + 142, + 481, + 199 + ], + "type": "image", + "image_path": "5b35f737699fb2b972c91d3bf48aa007781d148bd1689ce057f968a67c4533fd.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 133, + 206, + 383, + 264 + ], + "blocks": [ + { + "bbox": [ + 133, + 206, + 383, + 264 + ], + "lines": [ + { + "bbox": [ + 133, + 206, + 383, + 264 + ], + "spans": [ + { + "bbox": [ + 133, + 206, + 383, + 264 + ], + "type": "image", + "image_path": "03932a13732059a84a631165e27a977cf78cf596ee03c4ee09b3871fbb483151.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 296, + 482, + 396 + ], + "lines": [ + { + "bbox": [ + 130, + 296, + 482, + 396 + ], + "spans": [ + { + "bbox": [ + 130, + 296, + 482, + 396 + ], + "type": "text", + "content": "Fig. 1: Visualization of the primary components (i.e., (d)-(h)) introduced in the proposed visual localization scheme. (d) illustrates the point prompts selected from (b) with low reprojection errors, while (e) presents an error-guided mask expanded from the prompted points in (d) using SAM. (f) displays the proposed error-guided feature selection (EGFS), which refines the mask from (e) with the predicted confidence map (c) to ensure high-quality scene coordinates are sampled for estimating the final camera pose. The point cloud constructed from the predicted scene coordinates is shown on the right-hand side (i.e., (g)-(h)), with the confidence (yellow parts) and the refined EGFS mask (green for selected areas; red for rejected areas)." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 386, + 207, + 481, + 264 + ], + "blocks": [ + { + "bbox": [ + 386, + 207, + 481, + 264 + ], + "lines": [ + { + "bbox": [ + 386, + 207, + 481, + 264 + ], + "spans": [ + { + "bbox": [ + 386, + 207, + 481, + 264 + ], + "type": "image", + "image_path": "ce4fd49ac586c0410b143e90782785e1f88e82b72dfbfb28fb472af3e0afaee2.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 422, + 482, + 459 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 422, + 482, + 459 + ], + "spans": [ + { + "bbox": [ + 130, + 422, + 482, + 459 + ], + "type": "text", + "content": "provide benefits such as accuracy in smaller scenes, reduced training times, as well as minimized storage requirements. Given these advantages, SCR is thus the primary focus of this study and presents potential for further enhancements." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 460, + 482, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 460, + 482, + 628 + ], + "spans": [ + { + "bbox": [ + 130, + 460, + 482, + 628 + ], + "type": "text", + "content": "Albeit effective, previous SCR methods face two primary challenges to be addressed: the presence of dynamic objects and texture-less regions. Dynamic objects, such as pedestrians and cars, pose difficulties for these techniques due to their changing nature and the unsuitable features extracted from these regions. As illustrated in Fig. 1, these regions could lead to high reprojection errors during training. The second challenge is that current methods struggle with flat, texture-less surfaces. Despite their apparent simplicity, these areas often result in inaccurate scene coordinates due to difficulties in feature extraction. While previous methods [14, 15, 17] employ techniques like RANSAC to filter out outliers and differentiable end-to-end optimization approaches to disregard these outliers, they demand significant computational time for end-to-end processing and do not account for the semantic meaning of the selected areas, which may still result in the selection of outlier areas. Thus, they might not effectively prevent the models from unstable weight updates and can lead to training instability." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 630, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 630, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 630, + 482, + 666 + ], + "type": "text", + "content": "In light of the aforementioned issues, this study aims to explore the possibility of guiding the sampling process to favor regions with low reprojection errors, and seeks to leverage visual information in an image to expand masks" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 230, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 230, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 230, + 101 + ], + "type": "text", + "content": "T.-R. Liu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 481, + 319 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 481, + 319 + ], + "type": "text", + "content": "that bear semantic regional meanings. The core philosophy of our methodology focuses on selecting robust regions for training without the need to manually define explicit areas or categories. Such robust regions might differ across scenes, which make it inappropriate for manual definition, as elaborated in Section 4. More specifically, we propose a strategy, named Error-Guided Feature Selection (EGFS), for deriving low reprojection error samples as point prompts, and expanding them to encompass a complete mask with similar semantic meanings. The masks are iteratively updated after being frozen for a specified period, which enable the proposed model to dynamically renew the focused regions. To achieve this prompt-based semantic regions expansion, we employ the Segment Anything Model (SAM) [1], a vision foundation model capable of providing a general understanding of the scene. This error-guided scheme ensures that our training process concentrates on regions with low reprojection errors and adaptively reduces the variability introduced by dynamic or texture-less objects. Moreover, EGPS adopts a predicted confidence map to refine the expanded error-guided mask for updates, which ensures the regions marked by the masks are sufficiently reliable. An illustrations of each component is provided in Fig. 1." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 320, + 481, + 438 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 320, + 481, + 438 + ], + "spans": [ + { + "bbox": [ + 133, + 320, + 481, + 438 + ], + "type": "text", + "content": "To validate the effectiveness of our proposed method, we conduct evaluations on both outdoor and indoor datasets, including the Cambridge Landmarks dataset [18] and the Indoor6 [19] dataset, for evaluating the performance of our proposed SCR methodology. The experimental results suggest that our method can be broadly applied to previous methods and provides benefits, and leads to state-of-the-art (SOTA) performance when compared with SCR approaches without leveraging any 3D information in terms of estimated translational and rotational errors, while requiring less training time and smaller model sizes. We further perform a series of detailed analyses and ablation studies to confirm the effectiveness of our error-guided sampled regions and the proposed methodology." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 454, + 236, + 467 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 454, + 236, + 467 + ], + "spans": [ + { + "bbox": [ + 133, + 454, + 236, + 467 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 476, + 481, + 632 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 476, + 481, + 632 + ], + "spans": [ + { + "bbox": [ + 133, + 476, + 481, + 632 + ], + "type": "text", + "content": "Scene Coordinate Regression. Conventional SCR approaches typically begin by establishing 2D-3D correspondences and then utilize RANSAC-based optimization to estimate the camera pose. To establish 2D-3D correspondences, previous endeavors have explored the use of random forests [5, 20-22] or convolutional neural networks (CNNs) [11-15,17,23] to regress scene coordinates from images. Since SCR typically requires scene-specific training, training time has emerged as a critical concern in practical applications. The ACE [17] method, which can be trained within five minutes, stands out as a promising solution. Moreover, unlike methods that map every 2D pixel to a 3D coordinate, some works [19,24,25] have aimed at learning to detect 3D landmarks to establish 2D-3D correspondences. However, they often require the reconstruction of a 3D model, which incurs additional computational time and costs. Our method seeks to utilize SAM to derive error-guided masks from 2D images, while maintaining the training efficiency." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 133, + 641, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 641, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 641, + 481, + 665 + ], + "type": "text", + "content": "Emphasis on Robust Features for Localization. Several prior studies [24,26-31] in visual localization have discovered that not all regions within an image contribute" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 147, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 147, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 147, + 91, + 447, + 102 + ], + "type": "text", + "content": "Reprojection Errors as Prompts for Efficient Scene Coordinate Regression" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "text", + "content": "equally to localization accuracy. As a result, these approaches have attempted to distinguish robust and invariant features, found that prioritizing the recognition and comprehension of those regions can enhance visual localization tasks more efficiently and effectively. However, a common limitation of these methods is their reliance on pre-defined semantics or the necessity of a 3D model, which could hinder their generalizability and practicality in real-world applications." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 205, + 458, + 220 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 205, + 458, + 220 + ], + "spans": [ + { + "bbox": [ + 131, + 205, + 458, + 220 + ], + "type": "text", + "content": "3 Preliminary of Scene Coordinate Regression (SCR)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 230, + 483, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 230, + 483, + 399 + ], + "spans": [ + { + "bbox": [ + 130, + 230, + 483, + 399 + ], + "type": "text", + "content": "SCR-based visual localization determines the camera pose of an RGB image " + }, + { + "bbox": [ + 130, + 230, + 483, + 399 + ], + "type": "inline_equation", + "content": "I \\in \\mathbb{R}^{H \\times W \\times 3}" + }, + { + "bbox": [ + 130, + 230, + 483, + 399 + ], + "type": "text", + "content": " by predicting 3D scene coordinates " + }, + { + "bbox": [ + 130, + 230, + 483, + 399 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 130, + 230, + 483, + 399 + ], + "type": "text", + "content": " for a set of pixels and establishing the 2D-3D correspondences between pixel coordinates in the image and those predicted 3D scene coordinates. The primary objective of the training process is to establish a mapping function " + }, + { + "bbox": [ + 130, + 230, + 483, + 399 + ], + "type": "inline_equation", + "content": "f(\\cdot)" + }, + { + "bbox": [ + 130, + 230, + 483, + 399 + ], + "type": "text", + "content": " such that " + }, + { + "bbox": [ + 130, + 230, + 483, + 399 + ], + "type": "inline_equation", + "content": "\\mathcal{V} = f(I)" + }, + { + "bbox": [ + 130, + 230, + 483, + 399 + ], + "type": "text", + "content": ". The training data comprise RGB images paired with their respective camera poses, but do not necessarily include depth information. To generate accurate " + }, + { + "bbox": [ + 130, + 230, + 483, + 399 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 130, + 230, + 483, + 399 + ], + "type": "text", + "content": ", SCR methods often involve optimizing the reprojection error " + }, + { + "bbox": [ + 130, + 230, + 483, + 399 + ], + "type": "inline_equation", + "content": "\\mathcal{R}" + }, + { + "bbox": [ + 130, + 230, + 483, + 399 + ], + "type": "text", + "content": ", which is measured by projecting a 3D predicted scene coordinate " + }, + { + "bbox": [ + 130, + 230, + 483, + 399 + ], + "type": "inline_equation", + "content": "\\mathrm{y_i} \\in \\mathcal{V}" + }, + { + "bbox": [ + 130, + 230, + 483, + 399 + ], + "type": "text", + "content": " back onto its 2D pixel coordinate and calculating the discrepancy from the actual coordinate of the image patch, where i denotes the patch index. The reprojection error for each patch " + }, + { + "bbox": [ + 130, + 230, + 483, + 399 + ], + "type": "inline_equation", + "content": "\\mathrm{r_i} \\in \\mathcal{R}" + }, + { + "bbox": [ + 130, + 230, + 483, + 399 + ], + "type": "text", + "content": " can be formulated as " + }, + { + "bbox": [ + 130, + 230, + 483, + 399 + ], + "type": "inline_equation", + "content": "r(\\mathrm{p_i}, \\mathrm{y_i}, \\mathrm{h^*}) = \\| \\mathrm{p_i} - K\\mathrm{h^{*-1}}\\mathrm{y_i}\\|" + }, + { + "bbox": [ + 130, + 230, + 483, + 399 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 230, + 483, + 399 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 230, + 483, + 399 + ], + "type": "text", + "content": " denotes the camera intrinsic matrix and " + }, + { + "bbox": [ + 130, + 230, + 483, + 399 + ], + "type": "inline_equation", + "content": "\\mathrm{h^*}" + }, + { + "bbox": [ + 130, + 230, + 483, + 399 + ], + "type": "text", + "content": " denotes the ground truth camera pose. After obtaining " + }, + { + "bbox": [ + 130, + 230, + 483, + 399 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 130, + 230, + 483, + 399 + ], + "type": "text", + "content": ", the camera pose h can be estimated according to the following equation:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 186, + 405, + 481, + 420 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 186, + 405, + 481, + 420 + ], + "spans": [ + { + "bbox": [ + 186, + 405, + 481, + 420 + ], + "type": "interline_equation", + "content": "\\mathrm {h} = g (\\mathcal {C}) = g ^ {P n P} \\left(\\mathcal {C} _ {\\mathcal {I}}\\right), \\text {w i t h} \\mathcal {C} = \\left\\{\\left(\\mathrm {p} _ {\\mathrm {i}}, \\mathrm {y} _ {\\mathrm {i}}\\right) | \\mathrm {p} _ {\\mathrm {i}} \\in I, \\mathrm {y} _ {\\mathrm {i}} \\in \\mathcal {Y} \\right\\}, \\tag {1}", + "image_path": "c0623f201511d7fe3cf494ff7b4ee5e0c5d3632eb6d684ac632aa80259f0479c.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 427, + 482, + 524 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 427, + 482, + 524 + ], + "spans": [ + { + "bbox": [ + 130, + 427, + 482, + 524 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 427, + 482, + 524 + ], + "type": "inline_equation", + "content": "g(\\cdot)" + }, + { + "bbox": [ + 130, + 427, + 482, + 524 + ], + "type": "text", + "content": " involves PnP with RANSAC, followed by Levenberg-Marquardt-based refinement, " + }, + { + "bbox": [ + 130, + 427, + 482, + 524 + ], + "type": "inline_equation", + "content": "\\mathcal{C}" + }, + { + "bbox": [ + 130, + 427, + 482, + 524 + ], + "type": "text", + "content": " represents the set of all 2D-3D correspondences and " + }, + { + "bbox": [ + 130, + 427, + 482, + 524 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_{\\mathcal{I}}" + }, + { + "bbox": [ + 130, + 427, + 482, + 524 + ], + "type": "text", + "content": " signifies inlier correspondences after refinement, which is a subset of " + }, + { + "bbox": [ + 130, + 427, + 482, + 524 + ], + "type": "inline_equation", + "content": "\\mathcal{C}" + }, + { + "bbox": [ + 130, + 427, + 482, + 524 + ], + "type": "text", + "content": ". Inlier correspondences denote those that are better consistent with the estimated camera pose, while correspondences that do not align well with the estimated camera pose are termed outliers. The final camera pose estimation is determined by the inlier correspondences " + }, + { + "bbox": [ + 130, + 427, + 482, + 524 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_{\\mathcal{I}}" + }, + { + "bbox": [ + 130, + 427, + 482, + 524 + ], + "type": "text", + "content": ". In the following paragraphs, we introduce two scene coordinate regression methods that have demonstrated superior performance." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 533, + 482, + 620 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 482, + 620 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 482, + 620 + ], + "type": "text", + "content": "DSAC Variants. DSAC [13] introduces a differentiable variant of RANSAC, and enables end-to-end training of the entire pipeline. This design combines the benefits of RANSAC for handling outliers with the power of deep neural networks to learn complex patterns. In addition, DSAC++ [14] removes the requirement of depth information for SCR training and adopts robust loss terms to downweight outliers. DSAC* [15] further simplifies and refines DSAC++, and encourages the network to focus on reliable scene structures while ignoring outlier predictions." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 629, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 629, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 131, + 629, + 482, + 666 + ], + "type": "text", + "content": "ACE. ACE [17] is a fast SCR method that can achieve promising localization accuracy within a short training time of only five minutes. ACE splits its regression network into a pre-trained scene-agnostic backbone " + }, + { + "bbox": [ + 131, + 629, + 482, + 666 + ], + "type": "inline_equation", + "content": "f_{B}" + }, + { + "bbox": [ + 131, + 629, + 482, + 666 + ], + "type": "text", + "content": ", and a scene-specific" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 230, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 230, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 230, + 102 + ], + "type": "text", + "content": "T.-R. Liu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 143, + 114, + 305, + 212 + ], + "blocks": [ + { + "bbox": [ + 143, + 114, + 305, + 212 + ], + "lines": [ + { + "bbox": [ + 143, + 114, + 305, + 212 + ], + "spans": [ + { + "bbox": [ + 143, + 114, + 305, + 212 + ], + "type": "image", + "image_path": "5b6845a7066e395903dd28812392eed4dcb5061a06f9c0fc41e6f052db6cf123.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 307, + 114, + 470, + 211 + ], + "blocks": [ + { + "bbox": [ + 307, + 114, + 470, + 211 + ], + "lines": [ + { + "bbox": [ + 307, + 114, + 470, + 211 + ], + "spans": [ + { + "bbox": [ + 307, + 114, + 470, + 211 + ], + "type": "image", + "image_path": "020b608dc5d31389981574950ff0eb001465e1efe841fe1f00aa377e45ea2d5b.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 143, + 213, + 305, + 310 + ], + "blocks": [ + { + "bbox": [ + 143, + 213, + 305, + 310 + ], + "lines": [ + { + "bbox": [ + 143, + 213, + 305, + 310 + ], + "spans": [ + { + "bbox": [ + 143, + 213, + 305, + 310 + ], + "type": "image", + "image_path": "f613d160fd3f192fc7bde8f6fc65622be02fecfd61f9fad868e0a16619c39ef8.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 323, + 482, + 357 + ], + "lines": [ + { + "bbox": [ + 131, + 323, + 482, + 357 + ], + "spans": [ + { + "bbox": [ + 131, + 323, + 482, + 357 + ], + "type": "text", + "content": "Fig. 2: Analysis between reprojection error and semantic meaning. The analysis result indicates the regions with low reprojection errors tend to have higher inlier ratios, while the errors do not always align with specific semantic categories, e.g., \"tree\" and \"rug\"." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 307, + 213, + 470, + 310 + ], + "blocks": [ + { + "bbox": [ + 307, + 213, + 470, + 310 + ], + "lines": [ + { + "bbox": [ + 307, + 213, + 470, + 310 + ], + "spans": [ + { + "bbox": [ + 307, + 213, + 470, + 310 + ], + "type": "image", + "image_path": "6606dbd3678efa9350336af18d56201d63984931797c331948b4e48336e946f9.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 380, + 482, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 380, + 482, + 430 + ], + "spans": [ + { + "bbox": [ + 130, + 380, + 482, + 430 + ], + "type": "text", + "content": "multi-layer perceptron (MLP) " + }, + { + "bbox": [ + 130, + 380, + 482, + 430 + ], + "type": "inline_equation", + "content": "f_{H}" + }, + { + "bbox": [ + 130, + 380, + 482, + 430 + ], + "type": "text", + "content": ". The backbone " + }, + { + "bbox": [ + 130, + 380, + 482, + 430 + ], + "type": "inline_equation", + "content": "f_{B}" + }, + { + "bbox": [ + 130, + 380, + 482, + 430 + ], + "type": "text", + "content": " extracts feature vectors " + }, + { + "bbox": [ + 130, + 380, + 482, + 430 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_i" + }, + { + "bbox": [ + 130, + 380, + 482, + 430 + ], + "type": "text", + "content": " from image patches " + }, + { + "bbox": [ + 130, + 380, + 482, + 430 + ], + "type": "inline_equation", + "content": "p_i" + }, + { + "bbox": [ + 130, + 380, + 482, + 430 + ], + "type": "text", + "content": " and the MLP " + }, + { + "bbox": [ + 130, + 380, + 482, + 430 + ], + "type": "inline_equation", + "content": "f_{H}" + }, + { + "bbox": [ + 130, + 380, + 482, + 430 + ], + "type": "text", + "content": " predicts the scene coordinates " + }, + { + "bbox": [ + 130, + 380, + 482, + 430 + ], + "type": "inline_equation", + "content": "y_{i}" + }, + { + "bbox": [ + 130, + 380, + 482, + 430 + ], + "type": "text", + "content": " based on the extracted " + }, + { + "bbox": [ + 130, + 380, + 482, + 430 + ], + "type": "inline_equation", + "content": "\\mathrm{f}_i" + }, + { + "bbox": [ + 130, + 380, + 482, + 430 + ], + "type": "text", + "content": ". Between " + }, + { + "bbox": [ + 130, + 380, + 482, + 430 + ], + "type": "inline_equation", + "content": "f_{B}" + }, + { + "bbox": [ + 130, + 380, + 482, + 430 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 380, + 482, + 430 + ], + "type": "inline_equation", + "content": "f_{H}" + }, + { + "bbox": [ + 130, + 380, + 482, + 430 + ], + "type": "text", + "content": ", the features are shuffled to decorrelate gradients within a batch, which enables ACE to enhance its training efficiency." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 449, + 462, + 464 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 449, + 462, + 464 + ], + "spans": [ + { + "bbox": [ + 131, + 449, + 462, + 464 + ], + "type": "text", + "content": "4 In-Depth Evaluation of Scene Coordinate Regression" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 476, + 378, + 489 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 476, + 378, + 489 + ], + "spans": [ + { + "bbox": [ + 131, + 476, + 378, + 489 + ], + "type": "text", + "content": "4.1 Challenges in Scene Coordinate Regression" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 498, + 482, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 498, + 482, + 668 + ], + "spans": [ + { + "bbox": [ + 130, + 498, + 482, + 668 + ], + "type": "text", + "content": "Several challenging factors have been identified that may compromise the accuracy of SCR. Two key challenges among them are dynamic objects and textureless surfaces [27, 32-34]. Dynamic objects exhibit characteristics of motion or disappearance across different frames, while textureless surfaces lack distinctive feature points or salient characteristics. These challenges highlight the complexity of achieving precise SCR in complex visual scenarios. To address them, previous SCR approaches [13-15, 17] have employed RANSAC to filter out outliers, and some researches [14, 15, 17] have attempted to utilize end-to-end learning to concentrate on reliable scene structures while down-weighting outlier predictions during training. Unfortunately, these prior endeavors failed to fully exploit the rich semantic information in the visual content, which can be crucial for enhancing the robustness and accuracy of camera pose estimation in complex scenes. Furthermore, RANSAC-based approaches for camera pose estimation could potentially fail in scenarios with insufficient number of inlier correspondences [35]." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 146, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 146, + 91, + 447, + 102 + ], + "type": "text", + "content": "Reprojection Errors as Prompts for Efficient Scene Coordinate Regression" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 132, + 113, + 482, + 209 + ], + "blocks": [ + { + "bbox": [ + 132, + 113, + 482, + 209 + ], + "lines": [ + { + "bbox": [ + 132, + 113, + 482, + 209 + ], + "spans": [ + { + "bbox": [ + 132, + 113, + 482, + 209 + ], + "type": "image", + "image_path": "907825aa10e8fb607d365151a0f99a17944d3280758c178917bcbc4ec8b63e63.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 209, + 217, + 403, + 228 + ], + "lines": [ + { + "bbox": [ + 209, + 217, + 403, + 228 + ], + "spans": [ + { + "bbox": [ + 209, + 217, + 403, + 228 + ], + "type": "text", + "content": "Fig. 3: An overview of the training framework." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 251, + 467, + 264 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 251, + 467, + 264 + ], + "spans": [ + { + "bbox": [ + 132, + 251, + 467, + 264 + ], + "type": "text", + "content": "4.2 Analysis between Reprojection Error and Semantic Meaning" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 270, + 482, + 547 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 270, + 482, + 547 + ], + "spans": [ + { + "bbox": [ + 130, + 270, + 482, + 547 + ], + "type": "text", + "content": "For a more comprehensive understanding of the correlation and relationship between reprojection errors and semantic meaning, we investigate the performance of ACE on the reprojection errors and the inlier correspondences selected by RANSAC for estimating the final camera pose during the inference process. First, image pixels are labeled by a semantic segmentation model ViT-Adapter [36]. Subsequently, the reprojection errors and inlier correspondences of ACE for each semantic class are calculated. Fig. 2 presents the median reprojection errors and the ratios of inlier correspondences for the top ten most frequent classes in each scene. The left two figures correspond to the outdoor dataset Cambridge Landmarks [18], while the right two figures pertain to the indoor dataset Indoor6 [19]. It is observed that regions with low reprojection errors tend to have relatively higher inlier ratios, indicating a greater impact on the estimated camera poses. As a result, improving the accuracy of low-error areas could lead to more precise camera pose predictions. Nevertheless, it is also observed that the reprojection errors for the same class can vary significantly across different scenes. For example, in the Cambridge Landmarks dataset, the error for \"tree\" is low in King's College but high in Old Hospital. Similarly, in the Indoor6 dataset, the error for \"rug\" is high in scene1 but low in scene3. This indicates that low reprojection errors do not always align with specific semantic categories, which presents a challenge in directly using certain specific semantic segmentation labels for area selection. Motivated by these insights, the primary objective of this study is to develop a methodology that does not rely on pre-defined semantic categories to identify low reprojection error areas for enhancing localization accuracy in SCR." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 563, + 233, + 577 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 563, + 233, + 577 + ], + "spans": [ + { + "bbox": [ + 132, + 563, + 233, + 577 + ], + "type": "text", + "content": "5 Methodology" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 586, + 390, + 598 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 586, + 390, + 598 + ], + "spans": [ + { + "bbox": [ + 132, + 586, + 390, + 598 + ], + "type": "text", + "content": "5.1 Problem Definition and Framework Overview" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": "Fig. 3 provides an overview of the proposed training framework for SCR, which iteratively samples features to train a scene-specific MLP. The scene-specific MLP consists of a scene coordinate head and a confidence head. In each iteration, the model is trained for " + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": " epochs. During the initial iteration, features are randomly sampled from all parts of images in order to derive the first set" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 230, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 230, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 230, + 101 + ], + "type": "text", + "content": "T.-R. Liu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 201 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 201 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 201 + ], + "type": "text", + "content": "of reprojection errors. In subsequent iterations, features are selected based on error-guided feature selection (EGFS) masks generated according to reprojection errors and a confidence map. This iterative mechanism enables the model to dynamically update its focus areas throughout the training process. During the inference phase, the network estimates scene coordinates " + }, + { + "bbox": [ + 130, + 116, + 482, + 201 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 130, + 116, + 482, + 201 + ], + "type": "text", + "content": " and a confidence map for the refinement, while the masks generated during training are not required for the inference phase. The inference procedure is depicted in Fig. 4." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 220, + 418, + 232 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 220, + 418, + 232 + ], + "spans": [ + { + "bbox": [ + 131, + 220, + 418, + 232 + ], + "type": "text", + "content": "5.2 Error-Guided Feature Selection (EGFS) with SAM" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 241, + 482, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 241, + 482, + 553 + ], + "spans": [ + { + "bbox": [ + 130, + 241, + 482, + 553 + ], + "type": "text", + "content": "In alignment with the discussion in Section 4 on identifying potential areas contributing to the final camera pose estimation during the training stage, we employ the Segmentation Anything Model (SAM) [1], which is capable of leveraging diverse visual cues to generate high-quality object masks. Specifically, in this study, SAM is utilized to extend point prompts with low reprojection errors to form a complete mask sharing a similar semantic context. The rationale behind this approach is the high likelihood that these comprehensive semantic regions correspond to low error areas and can contribute to final pose estimation. Such a concept harnesses the capabilities of the foundational SAM model to uncover more viable areas, and hence, enabling a thorough understanding of the image context. In our experiments, we utilize EfficientViT-SAM-L0 [37], a lightweight and efficient variant of SAM, to identify salient regions by selecting the " + }, + { + "bbox": [ + 130, + 241, + 482, + 553 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 130, + 241, + 482, + 553 + ], + "type": "text", + "content": " percentage of points with the lowest reprojection errors as point prompts, where " + }, + { + "bbox": [ + 130, + 241, + 482, + 553 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 130, + 241, + 482, + 553 + ], + "type": "text", + "content": " is adjustable and set to ten in our case. After obtaining the error-guided mask from the SAM model, the predicted confidence map " + }, + { + "bbox": [ + 130, + 241, + 482, + 553 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 130, + 241, + 482, + 553 + ], + "type": "text", + "content": " is utilized to refine the error-guided mask to ensure the points marked by the mask is sufficiently reliable. Specifically, in each image, the points with confidence scores below a threshold " + }, + { + "bbox": [ + 130, + 241, + 482, + 553 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 130, + 241, + 482, + 553 + ], + "type": "text", + "content": " are filtered out. The design for confidence refinement and its optimization are described in Section 5.3. In each iteration, the refined error-guided masks serve as guidance for training the scene-specific MLP. Only the features corresponding to the regions selected by the masks are sampled into the training buffer. This selective inclusion is facilitated by the nature of " + }, + { + "bbox": [ + 130, + 241, + 482, + 553 + ], + "type": "inline_equation", + "content": "f_{H}" + }, + { + "bbox": [ + 130, + 241, + 482, + 553 + ], + "type": "text", + "content": " [17], which employs " + }, + { + "bbox": [ + 130, + 241, + 482, + 553 + ], + "type": "inline_equation", + "content": "1\\times 1" + }, + { + "bbox": [ + 130, + 241, + 482, + 553 + ], + "type": "text", + "content": " convolutions to treat and process each selected sample independently using the same set of shared weights. The proposed feature selection mechanism ensures that the training process focuses on the crucial regions identified in the current iteration, and therefore enables the model to enhance its overall performance." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 571, + 452, + 583 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 571, + 452, + 583 + ], + "spans": [ + { + "bbox": [ + 131, + 571, + 452, + 583 + ], + "type": "text", + "content": "5.3 Scene Coordinate and EGFS Refinement with Confidence" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": "In order to further refine the error-guided mask and ensure the selected points are of sufficiently high quality for predicting accurate camera pose, we incorporate the design of a confidence map [38, 39]. Specifically, we predict a confidence score pixel-wise from the confidence head and optimize it jointly with the reprojection error. This confidence head replaces the last three " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "1 \\times 1" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": " convolution layers of the original scene-specific MLP with two " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "1 \\times 1" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": " convolution layers for" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 146, + 91, + 447, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 91, + 447, + 103 + ], + "spans": [ + { + "bbox": [ + 146, + 91, + 447, + 103 + ], + "type": "text", + "content": "Reprojection Errors as Prompts for Efficient Scene Coordinate Regression" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 184, + 114, + 430, + 196 + ], + "blocks": [ + { + "bbox": [ + 184, + 114, + 430, + 196 + ], + "lines": [ + { + "bbox": [ + 184, + 114, + 430, + 196 + ], + "spans": [ + { + "bbox": [ + 184, + 114, + 430, + 196 + ], + "type": "image", + "image_path": "ca54741ca98085209b7b2368001f7d1a5da12a6c2833602c004768ea58b3ea06.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 209, + 203, + 404, + 216 + ], + "lines": [ + { + "bbox": [ + 209, + 203, + 404, + 216 + ], + "spans": [ + { + "bbox": [ + 209, + 203, + 404, + 216 + ], + "type": "text", + "content": "Fig. 4: An overview of the inference procedure." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 236, + 479, + 260 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 236, + 479, + 260 + ], + "spans": [ + { + "bbox": [ + 130, + 236, + 479, + 260 + ], + "type": "text", + "content": "confidence score prediction, thus enabling adaptive weights for the reprojection loss at each position. The training loss used by our method is formulated as:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 187, + 264, + 481, + 297 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 187, + 264, + 481, + 297 + ], + "spans": [ + { + "bbox": [ + 187, + 264, + 481, + 297 + ], + "type": "interline_equation", + "content": "\\ell \\left(\\mathrm {p} _ {\\mathrm {i}}, \\mathrm {y} _ {\\mathrm {i}}, \\mathrm {h} _ {\\mathrm {i}} ^ {*}\\right) = \\left\\{ \\begin{array}{l l} \\mathrm {c} _ {\\mathrm {i}} \\cdot \\hat {r} \\left(\\mathrm {p} _ {\\mathrm {i}}, \\mathrm {y} _ {\\mathrm {i}}, \\mathrm {h} _ {\\mathrm {i}} ^ {*}\\right) - \\alpha \\log \\mathrm {c} _ {\\mathrm {i}} & \\text {i f y} _ {\\mathrm {i}} \\in \\mathcal {V} \\\\ \\| \\mathrm {y} _ {\\mathrm {i}} - \\bar {\\mathrm {y}} _ {\\mathrm {i}} \\| _ {0} - \\alpha \\log (1 - \\mathrm {c} _ {\\mathrm {i}}) & \\text {o t h e r w i s e} \\end{array} , \\right. \\tag {2}", + "image_path": "45268d6f4b34c57fdd2810a8e87e604a9042c3c5812c8722d29eee2d7b29cbce.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 301, + 482, + 494 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 301, + 482, + 494 + ], + "spans": [ + { + "bbox": [ + 130, + 301, + 482, + 494 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 301, + 482, + 494 + ], + "type": "inline_equation", + "content": "\\hat{r} (\\mathrm{p_i},\\mathrm{y_i},\\mathrm{h_i^*})" + }, + { + "bbox": [ + 130, + 301, + 482, + 494 + ], + "type": "text", + "content": " represents a tanh clamping of the reprojection error that changes over time, " + }, + { + "bbox": [ + 130, + 301, + 482, + 494 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 130, + 301, + 482, + 494 + ], + "type": "text", + "content": " denotes the set of 2D pixels with valid scene coordinate predictions, as explained in [15, 17], " + }, + { + "bbox": [ + 130, + 301, + 482, + 494 + ], + "type": "inline_equation", + "content": "c_{i}" + }, + { + "bbox": [ + 130, + 301, + 482, + 494 + ], + "type": "text", + "content": " is the confidence score at pixel i, " + }, + { + "bbox": [ + 130, + 301, + 482, + 494 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 130, + 301, + 482, + 494 + ], + "type": "text", + "content": " is a hyperparameter that balances the confidence regularization term, and " + }, + { + "bbox": [ + 130, + 301, + 482, + 494 + ], + "type": "inline_equation", + "content": "\\bar{\\mathbf{y}}_{\\mathrm{i}}" + }, + { + "bbox": [ + 130, + 301, + 482, + 494 + ], + "type": "text", + "content": " is a dummy scene coordinate derived from the ground truth camera pose, assuming a constant image depth of " + }, + { + "bbox": [ + 130, + 301, + 482, + 494 + ], + "type": "inline_equation", + "content": "10\\mathrm{m}" + }, + { + "bbox": [ + 130, + 301, + 482, + 494 + ], + "type": "text", + "content": ". When the scene coordinate prediction is valid, the model down-weights the reprojection errors in challenging regions using confidence scores and focuses more on regions with reliable predictions. On the other hand, when the scene coordinate prediction is invalid, the model is incentivized to lower the confidence scores for inaccurate predictions. The confidence score is also used in the inference phase to select reliable scene coordinates. For each query image, the median of the confidence scores is calculated, and only scene coordinates with confidence greater than this median are selected. This thresholding ensures that only the most reliable 2D-3D correspondences are used to generate the camera pose by the PnP with RANSAC pose solver. By focusing on high-confidence correspondences, a more accurate camera pose can be estimated." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 509, + 282, + 523 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 509, + 282, + 523 + ], + "spans": [ + { + "bbox": [ + 132, + 509, + 282, + 523 + ], + "type": "text", + "content": "6 Experimental Results" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 531, + 266, + 544 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 531, + 266, + 544 + ], + "spans": [ + { + "bbox": [ + 132, + 531, + 266, + 544 + ], + "type": "text", + "content": "6.1 Experimental Setups" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 548, + 480, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 548, + 480, + 586 + ], + "spans": [ + { + "bbox": [ + 130, + 548, + 480, + 586 + ], + "type": "text", + "content": "Datasets. Our evaluation of the proposed method was conducted on two representative datasets: the outdoor dataset Cambridge Landmarks [18] and the indoor dataset Indoor6 [19]. The details of the datasets are described as follows:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 593, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 593, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 593, + 482, + 666 + ], + "type": "text", + "content": "Cambridge Landmarks. The Cambridge Landmarks Dataset [18] is a renowned outdoor dataset extensively utilized for visual localization tasks, which includes five distinct scenes. Every scene in this dataset is composed of RGB images along with their respective ground truth camera poses reconstructed using the SfM technique. It includes a broad spectrum of conditions and various viewpoints, and is ideal for evaluating the resilience and effectiveness of various SCR methods." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 230, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 230, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 230, + 101 + ], + "type": "text", + "content": "T.-R. Liu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 134, + 157, + 481, + 262 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 482, + 149 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 482, + 149 + ], + "type": "text", + "content": "Table 1: A comparison of model sizes, training times, and median translation and rotation errors (in " + }, + { + "bbox": [ + 132, + 114, + 482, + 149 + ], + "type": "inline_equation", + "content": "\\mathrm{cm} / {}^{\\circ}" + }, + { + "bbox": [ + 132, + 114, + 482, + 149 + ], + "type": "text", + "content": ") evaluated on the Cambridge Landmarks dataset. 'dual' and 'quad.' denote the ensemble versions of two and four models, respectively." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 134, + 157, + 481, + 262 + ], + "lines": [ + { + "bbox": [ + 134, + 157, + 481, + 262 + ], + "spans": [ + { + "bbox": [ + 134, + 157, + 481, + 262 + ], + "type": "table", + "html": "
MethodSizeMapping TimeKing's College Great Court Old Hospital Shop Facade St Mary's ChurchAverage
FMhLoc (SP+SG) [4,6,8]~800MB~35 min12/0.216/0.115/0.34/0.27/0.211/0.2
pixLoc [26]~600MB30/0.114/0.216/0.35/0.210/0.315/0.2
SCR (w/ 3D model)DSAC* [15]28MB15 hr15/0.349/0.321/0.45/0.313/0.421/0.3
FocusTune [31]4MB6 min19/0.338/0.118/0.46/0.315/0.519/0.3
FocusTune (quad.) [31]16MB24 min15/0.329/0.117/0.45/0.29/0.315/0.3
NeuMap [40]--14/0.26/0.119/0.36/0.217/0.512/0.3
SACReg-L [41]--11/0.213/0.113/0.26/0.25/0.310/0.2
SCRDSAC* [15]28MB15 hr18/0.334/0.221/0.45/0.315/0.619/0.4
ACE [17]4MB5 min28/0.443/0.231/0.65/0.318/0.625/0.4
ACE (quad.) [17]16MB20 min18/0.328/0.125/0.55/0.39/0.317/0.3
EGFS4.5MB12 min14/0.331/0.121/0.45/0.315/0.517/0.3
EGFS (dual)9MB21 min14/0.328/0.119/0.45/0.210/0.315/0.3
", + "image_path": "f81fed9dc585897b3db20ba12e8f4f43ec2b4d1907206b9c88bcc9b70f462cad.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 135, + 304, + 481, + 402 + ], + "blocks": [ + { + "bbox": [ + 132, + 273, + 480, + 295 + ], + "lines": [ + { + "bbox": [ + 132, + 273, + 480, + 295 + ], + "spans": [ + { + "bbox": [ + 132, + 273, + 480, + 295 + ], + "type": "text", + "content": "Table 2: A comparison of model sizes, training times, and the proportions of translation and rotation errors that are below " + }, + { + "bbox": [ + 132, + 273, + 480, + 295 + ], + "type": "inline_equation", + "content": "5\\mathrm{cm} / 5^{\\circ}" + }, + { + "bbox": [ + 132, + 273, + 480, + 295 + ], + "type": "text", + "content": " evaluated on the Indoor6 dataset." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 135, + 304, + 481, + 402 + ], + "lines": [ + { + "bbox": [ + 135, + 304, + 481, + 402 + ], + "spans": [ + { + "bbox": [ + 135, + 304, + 481, + 402 + ], + "type": "table", + "html": "
MethodSizeMapping Timescene1scene2ascene3scene4ascene5scene6Average
FMhLoc [4]~1.5GB~3.3 hr70.5%52.1%86.0%75.3%58.0%86.7%71.4%
SCR (w/ 3D model)DSAC* [15]28MB15 hr18.7%28.0%19.7%60.8%10.6%44.3%30.4%
SLD* (300 landmarks) [25]15MB~5.5 hr47.2%48.2%56.2%67.7%33.7%52.0%50.8%
SLD* (1000 landmarks) [25]120MB~44 hr68.5%62.6%76.2%77.2%57.8%78.0%70.1%
SCRDSAC* [15]28MB15 hr23.0%33.9%26.0%67.1%10.6%50.2%35.1%
ACE [17]4MB5 min26.0%32.3%31.4%62.0%14.2%47.4%35.6%
ACE (quad.) [17]16MB20 min52.9%52.5%62.9%69.6%31.1%82.4%58.6%
EGFS4.5MB21 min46.4%60.6%56.4%78.7%22.8%71.6%56.1%
EGFS (dual)9MB30 min58.5%59.1%67.0%76.1%30.6%75.9%61.2%
", + "image_path": "f289044e04b916e259c0c1456ee4cdea2a70a38057fcae22a37da31f29416728.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 423, + 482, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 423, + 482, + 472 + ], + "spans": [ + { + "bbox": [ + 132, + 423, + 482, + 472 + ], + "type": "text", + "content": "Indoor6. The Indoor6 [19] dataset comprises six distinct indoor scenes captured over several days, with ground truth camera poses computed using COLMAP [42]. Each scene includes multiple rooms and contains illumination variations, making it challenging for different types of visual localization tasks." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 487, + 482, + 561 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 487, + 482, + 561 + ], + "spans": [ + { + "bbox": [ + 132, + 487, + 482, + 561 + ], + "type": "text", + "content": "Implementation Details. The training process of our methodology includes 20 epochs, with the generation of masks every five epochs. We set the confidence regularization parameter " + }, + { + "bbox": [ + 132, + 487, + 482, + 561 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 132, + 487, + 482, + 561 + ], + "type": "text", + "content": " to ten and the reprojection error threshold " + }, + { + "bbox": [ + 132, + 487, + 482, + 561 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 132, + 487, + 482, + 561 + ], + "type": "text", + "content": " to ten. For mask refinement, the parameter " + }, + { + "bbox": [ + 132, + 487, + 482, + 561 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 132, + 487, + 482, + 561 + ], + "type": "text", + "content": " is set to the median confidence score of each image. Our method builds upon the ACE [17] architecture as the backbone, and thus, the remaining hyperparameters are kept the same as those used in [17]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 575, + 459, + 588 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 575, + 459, + 588 + ], + "spans": [ + { + "bbox": [ + 132, + 575, + 459, + 588 + ], + "type": "text", + "content": "6.2 Results on the Cambridge Landmarks and Indoor6 Dataset" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 594, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 594, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 594, + 482, + 666 + ], + "type": "text", + "content": "Cambridge Landmarks Dataset. Table 1 presents the experimental results evaluated on the Cambridge Landmarks dataset, which affirm the effectiveness and efficiency of our proposed methodology. It can be observed that EGFS maintains a model size similar to ACE, with only a 0.5MB increase for the confidence head, while reducing average translational and rotational errors. The results further reveal that EGFS not only surpasses DSAC* in performance and performs" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 146, + 91, + 447, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 91, + 447, + 103 + ], + "spans": [ + { + "bbox": [ + 146, + 91, + 447, + 103 + ], + "type": "text", + "content": "Reprojection Errors as Prompts for Efficient Scene Coordinate Regression" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 142, + 146, + 473, + 280 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 482, + 138 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 482, + 138 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 482, + 138 + ], + "type": "text", + "content": "Table 3: Comparison of the effectiveness between (a) directly sampling solely based on pixel points with low reprojection errors, and (b) sampling through the EGFS masks." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 142, + 146, + 473, + 280 + ], + "lines": [ + { + "bbox": [ + 142, + 146, + 473, + 280 + ], + "spans": [ + { + "bbox": [ + 142, + 146, + 473, + 280 + ], + "type": "table", + "html": "
DatasetSceneri < Q0.3(R) ri < Q0.4(R) ri < Q0.5(R) ri < Q0.6(R) ri < Q0.7(R)EGFS Masks
CambridgeKing's College20/0.319/0.320/0.320/0.320/0.314/0.3
Great Court39/0.237/0.136/0.238/0.239/0.231/0.1
Old Hospital26/0.525/0.523/0.525/0.524/0.521/0.4
Shop Facade6/0.36/0.36/0.36/0.35/0.35/0.3
St Mary's Church17/0.517/0.516/0.517/0.516/0.515/0.5
Average22/0.421/0.420/0.421/0.421/0.417/0.3
Indoor6scene134.4%34.9%34.9%32.0%30.2%46.4%
scene2a39.8%40.6%46.1%41.3%42.5%60.6%
scene345.2%43.9%44.9%41.3%42.3%56.4%
scene4a63.2%63.9%62.6%63.2%67.1%78.7%
scene519.5%18.1%17.1%18.3%16.9%22.8%
scene659.7%60.3%60.0%57.2%58.1%71.6%
Average43.6%43.6%44.3%42.2%42.9%56.1%
", + "image_path": "c0fd9b1afb6c3f6140384331b3e6d0a852d42ce7d1205ae890d91ae97097bfbe.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 301, + 482, + 352 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 301, + 482, + 352 + ], + "spans": [ + { + "bbox": [ + 130, + 301, + 482, + 352 + ], + "type": "text", + "content": "on par with the ensemble version of ACE (i.e., ACE (quad.)), but also features a smaller model size and reduced training time. Furthermore, our model's ensemble version (i.e., EGFS (dual.)) is able to exceed the performance of DSAC* and ACE (quad.) while requiring a smaller model size and less training time." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 370, + 483, + 444 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 370, + 483, + 444 + ], + "spans": [ + { + "bbox": [ + 130, + 370, + 483, + 444 + ], + "type": "text", + "content": "Indoor6 Dataset. Table 2 presents the evaluation results on Indoor6. EGFS significantly outperforms DSAC* and ACE across all scenes, while also achieving comparable performance to ACE (quad.) with a much smaller model size (4.5MB compared to 16MB). Moreover, EGFS (dual.) demonstrates superior performance relative to ACE (quad.), which further substantiates the effectiveness of EGFS in optimization through error-guided masks and confidence refinement." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 463, + 400, + 475 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 463, + 400, + 475 + ], + "spans": [ + { + "bbox": [ + 131, + 463, + 400, + 475 + ], + "type": "text", + "content": "6.3 Effectiveness of Error-Guided Feature Selection" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "type": "text", + "content": "To further substantiate the advantages of employing the EGFS approach, this experiment compares several baseline schemes that sample points with different reprojection error thresholds against EGFS. The results are summarized in Table 3. Specifically, these baseline schemes select sample points based on various quantiles of the reprojection error map, ranging from " + }, + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "type": "inline_equation", + "content": "30\\%" + }, + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "type": "inline_equation", + "content": "70\\%" + }, + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "type": "text", + "content": ", denoted as " + }, + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "type": "inline_equation", + "content": "Q_{0.3}(R)" + }, + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "type": "text", + "content": " through " + }, + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "type": "inline_equation", + "content": "Q_{0.7}(R)" + }, + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "type": "text", + "content": ", respectively. In each baseline scheme, a point is chosen if its reprojection error " + }, + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "type": "text", + "content": " falls below the threshold specified by " + }, + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "type": "inline_equation", + "content": "Q_{\\text{quantile}}(R)" + }, + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "type": "text", + "content": ". The results suggest that relying solely on reprojection errors does not guarantee optimal learning of scene coordinates. This is due to the fact that areas of low reprojection errors can be scattered and may not encompass entire semantic objects. In contrast, the proposed EGFS expands the points into complete semantic masks followed by confidence refinement, which enables our model to determine scene coordinates with minimal transitional and rotational errors. This experiment thus confirms the effectiveness of leveraging SAM's ability to interpret image context alongside confidence maps for enhanced sample point selection." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 230, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 230, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 230, + 102 + ], + "type": "text", + "content": "T.-R. Liu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 149, + 114, + 463, + 247 + ], + "blocks": [ + { + "bbox": [ + 149, + 114, + 463, + 247 + ], + "lines": [ + { + "bbox": [ + 149, + 114, + 463, + 247 + ], + "spans": [ + { + "bbox": [ + 149, + 114, + 463, + 247 + ], + "type": "image", + "image_path": "cdd7b19eca1bfb808f3163421798963e98e839761a8fc757fb66044a3767915a.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 255, + 482, + 310 + ], + "lines": [ + { + "bbox": [ + 131, + 255, + 482, + 310 + ], + "spans": [ + { + "bbox": [ + 131, + 255, + 482, + 310 + ], + "type": "text", + "content": "Fig. 5: Visualization of the EGFS mask refinement process at every five epochs, which depicts the reprojection errors at the beginning (epoch 5) and the end (epoch 20), as well as the refined error-guided masks used throughout training. The red dots represent low reprojection errors that serve as prompts, while the light green overlay denotes the refined EGFS masks. It can be observed that the EGFS masks enhances over epochs." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 157, + 327, + 465, + 459 + ], + "blocks": [ + { + "bbox": [ + 157, + 327, + 465, + 459 + ], + "lines": [ + { + "bbox": [ + 157, + 327, + 465, + 459 + ], + "spans": [ + { + "bbox": [ + 157, + 327, + 465, + 459 + ], + "type": "image", + "image_path": "f5e2853b2307b1a8e4bf9ebd579a2e93d56e6ab2d967d8d51020e89a3c472809.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 468, + 482, + 491 + ], + "lines": [ + { + "bbox": [ + 131, + 468, + 482, + 491 + ], + "spans": [ + { + "bbox": [ + 131, + 468, + 482, + 491 + ], + "type": "text", + "content": "Fig. 6: Visualization of the estimated camera pose trajectories from testing sequences with the camera frustums colored based on translational errors. Pose errors denoted." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 521, + 257, + 533 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 521, + 257, + 533 + ], + "spans": [ + { + "bbox": [ + 131, + 521, + 257, + 533 + ], + "type": "text", + "content": "6.4 Qualitative Results" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "text", + "content": "Iterative Error-guided Feature Selection Mask. Fig. 5 illustrates the refinement of EGFS masks every five epochs and compares the reprojection errors at the beginning and the end. It can be observed that EGFS is not only capable of expanding points with low reprojection errors (i.e., " + }, + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathrm{r} < Q_{0.1}(R)" + }, + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "text", + "content": ") into regions that bear similar semantic meanings, but also refines the masks with a confidence map to exclude uncertain areas. The masks are observed to be iteratively enhanced from coarse to fine. In the beginning (i.e., at epoch five), the reprojection errors appear noisy, whereas by the end (i.e., at epoch 20), the EGFS masks become more refined and concentrated on key regions beneficial for SCR." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 147, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 147, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 147, + 91, + 447, + 102 + ], + "type": "text", + "content": "Reprojection Errors as Prompts for Efficient Scene Coordinate Regression" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 151, + 116, + 458, + 300 + ], + "blocks": [ + { + "bbox": [ + 151, + 116, + 458, + 300 + ], + "lines": [ + { + "bbox": [ + 151, + 116, + 458, + 300 + ], + "spans": [ + { + "bbox": [ + 151, + 116, + 458, + 300 + ], + "type": "image", + "image_path": "bbeac4a9f655ceae17f344e5c6b90cfeb9afe3b17e0ce04d08638b8591f46760.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 313, + 482, + 359 + ], + "lines": [ + { + "bbox": [ + 131, + 313, + 482, + 359 + ], + "spans": [ + { + "bbox": [ + 131, + 313, + 482, + 359 + ], + "type": "text", + "content": "Fig. 7: Visualization of point clouds reconstructed from estimated scene coordinates of the training sequence. The point clouds derived from the scene coordinates estimated by the model trained with the proposed EGFS are clearer compared to those from the ACE model, even without the application of refined EGFS masks at inference time." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 382, + 482, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 382, + 482, + 468 + ], + "spans": [ + { + "bbox": [ + 131, + 382, + 482, + 468 + ], + "type": "text", + "content": "Visualization of the Estimated Camera Poses. In this section, we present the estimated camera poses from the testing sequences, and compare the proposed EGFS with the ACE baseline. The results are illustrated in Fig. 6. To visualize camera trajectories, we connect consecutive camera positions with lines and indicate camera orientations with frustums. Each sampled camera pose is associated with a color to represent the translation error. It is observed that our proposed EGFS can estimate more accurate camera poses as compared to ACE." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 486, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 486, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 131, + 486, + 482, + 666 + ], + "type": "text", + "content": "Scene Coordinates and 3D Point Clouds w/ and w/o EFGS. To validate the effectiveness of EGFS qualitatively and its impact on the quality of estimated scene coordinates that learned from training sequence, we depict the 3D point cloud generated from the scene coordinates from the training sequence, and apply colors based on the corresponding queried pixels in Fig. 7. To compare the 3D point clouds reconstructed from the scene coordinates, we compare the baseline ACE approach with the proposed EGFS approach. For our approach, we visualize the point clouds both before and after the application of EGFS masks and confidence maps when collecting the scene coordinates. Please note that both before and after versions which shown in the figure are trained with EGFS refinement. The 3D point cloud constructed by the proposed approach appears clearer, in contrast to the ACE-generated point cloud, which exhibits blurriness due to inaccurate scene coordinate estimation. In addition, the point cloud reconstructed from raw scene coordinates predicted without EGFS from our approach contains noise and floaters due to the inclusion of scene coordinates" + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 230, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 230, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 230, + 101 + ], + "type": "text", + "content": "T.-R. Liu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 152, + 115, + 460, + 287 + ], + "blocks": [ + { + "bbox": [ + 152, + 115, + 460, + 287 + ], + "lines": [ + { + "bbox": [ + 152, + 115, + 460, + 287 + ], + "spans": [ + { + "bbox": [ + 152, + 115, + 460, + 287 + ], + "type": "image", + "image_path": "2e91c655bb39e7fbc02620fda7ca4f6f5a8bd629fd83d7ebe782d25be6719296.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 295, + 482, + 329 + ], + "lines": [ + { + "bbox": [ + 131, + 295, + 482, + 329 + ], + "spans": [ + { + "bbox": [ + 131, + 295, + 482, + 329 + ], + "type": "text", + "content": "Fig. 8: Visualization of the estimated scene coordinates and the confidence maps on unseen samples. The RGB color in the scene coordinates represents the XYZ coordinates. The point clouds are colored green for selected areas and red for rejected areas." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 353, + 482, + 401 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 353, + 482, + 401 + ], + "spans": [ + { + "bbox": [ + 130, + 353, + 482, + 401 + ], + "type": "text", + "content": "with low confidence scores. On the other hand, the point cloud reconstructed from the scene coordinates with EGFS refinement exhibits clearer contours and less noise. This justifies EGFS's ability to eliminate areas of low confidence values and produce high-quality dense scene coordinates." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 420, + 482, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 420, + 482, + 529 + ], + "spans": [ + { + "bbox": [ + 130, + 420, + 482, + 529 + ], + "type": "text", + "content": "Visualization of Estimated Confidence Map on Unseen Samples. In this section, we qualitatively evaluate the confidence maps estimated from the testing sequence by visualizing them in both the 2D image plane and as 3D projections onto the point cloud. The visualized point cloud is reconstructed from the training sequence, whereas the plotted samples are selected from the testing sequence and were not included in the training procedure. These visualizations are provided in Fig. 8. It can be observed that the estimated confidence map assists in rejecting scene coordinates located in areas where accurate prediction is challenging, such as in the air or on texture-less surfaces." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 548, + 239, + 560 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 548, + 239, + 560 + ], + "spans": [ + { + "bbox": [ + 131, + 548, + 239, + 560 + ], + "type": "text", + "content": "6.5 Ablation Study" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "type": "text", + "content": "Error-Guided Feature Selection (EGFS) and Confidence Refinement. We first present an ablation analysis to validate the effectiveness of the EGFS masks and confidence refinement in enhancing our method's performance. Table 4 reports the average results on the Cambridge Landmarks and Indoor6 datasets. The results demonstrate that the model trained with the proposed EGFS mechanism outperforms those trained without it, which indicates that focusing on training with robust features leads to improved performance. Furthermore, the incorporation of the confidence map also enhances the overall" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 147, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 147, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 147, + 91, + 447, + 102 + ], + "type": "text", + "content": "Reprojection Errors as Prompts for Efficient Scene Coordinate Regression" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 135, + 146, + 320, + 222 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 319, + 137 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 319, + 137 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 319, + 137 + ], + "type": "text", + "content": "Table 4: Ablation study on the impact of each proposed components." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 135, + 146, + 320, + 222 + ], + "lines": [ + { + "bbox": [ + 135, + 146, + 320, + 222 + ], + "spans": [ + { + "bbox": [ + 135, + 146, + 320, + 222 + ], + "type": "table", + "html": "
Error-Guided MasksConfidence RefinementCambridge (cm/°)Indoor6 (%)
XX25/0.435.6
✓X19/0.341.6
X✓19/0.450.9
✓✓17/0.356.1
", + "image_path": "94d6a9c992b71ded29d6724838fc6f1dc2efef7b1389e6316ca305bd5b08622e.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 332, + 146, + 482, + 222 + ], + "blocks": [ + { + "bbox": [ + 329, + 114, + 482, + 137 + ], + "lines": [ + { + "bbox": [ + 329, + 114, + 482, + 137 + ], + "spans": [ + { + "bbox": [ + 329, + 114, + 482, + 137 + ], + "type": "text", + "content": "Table 5: Impact of varying proportions of point prompts " + }, + { + "bbox": [ + 329, + 114, + 482, + 137 + ], + "type": "inline_equation", + "content": "(\\tau)" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 332, + 146, + 482, + 222 + ], + "lines": [ + { + "bbox": [ + 332, + 146, + 482, + 222 + ], + "spans": [ + { + "bbox": [ + 332, + 146, + 482, + 222 + ], + "type": "table", + "html": "
Proportions of point prompts (τ)Cambridge (cm/°)Indoor6 (%)
517/0.354.3
1017/0.356.1
1517/0.355.4
2017/0.355.3
", + "image_path": "9e398625b170aa06b86272d7f6d6f3cfe5f3caa56e979858ebf7a208e86d0bb4.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 241, + 481, + 278 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 241, + 481, + 278 + ], + "spans": [ + { + "bbox": [ + 130, + 241, + 481, + 278 + ], + "type": "text", + "content": "performance. The final row of Table 4 demonstrates that combining both techniques (i.e., our EGFS) achieves the best results, which further confirms the effectiveness of EGFS in improving localization accuracy." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 293, + 482, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 293, + 482, + 437 + ], + "spans": [ + { + "bbox": [ + 130, + 293, + 482, + 437 + ], + "type": "text", + "content": "Analysis of Proportions of Point Prompts. We evaluate the performance of different proportions of point prompts used for expanding into EGFS masks, and the results are reported in Table 5. It can be observed that the proportions of point prompts do not significantly impact the performance on the Cambridge Landmark dataset. This may be attributed to the fact that the scenes in the dataset typically feature a primary architectural structure with fewer complicated details in the surroundings. On the other hand, for the Indoor6 dataset, the proportions of point prompts significantly affect the performance. The rationale behind these observations is that the indoor scenes are more complicated, and feature various small items that necessitate a larger number of low-error point prompts to comprehensively capture the entire scene structure. Please note that we employ " + }, + { + "bbox": [ + 130, + 293, + 482, + 437 + ], + "type": "inline_equation", + "content": "\\tau = 10" + }, + { + "bbox": [ + 130, + 293, + 482, + 437 + ], + "type": "text", + "content": " as our setting for all the experiments presented." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 451, + 220, + 464 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 451, + 220, + 464 + ], + "spans": [ + { + "bbox": [ + 132, + 451, + 220, + 464 + ], + "type": "text", + "content": "7 Conclusion" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 473, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 473, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 473, + 482, + 666 + ], + "type": "text", + "content": "This paper addressed key challenges in SCR for visual localization, and specifically focused on the impacts of dynamic objects and texture-less regions. Our approach introduced an innovative EGFS mechanism through the use of SAM and confidence maps to enhance the performance of SCR. This technique effectively filtered out problematic areas by concentrating on regions with low reprojection errors. Moreover, we used confidence maps to further refine the pixels selected by EGFS and perform this process iteratively to enable dynamic updates of the focused regions. The experimental results on the Cambridge Landmarks and Indoor6 datasets suggest that our method can provide improvements in terms of training efficiency, model size, and accuracy. Our findings highlighted the importance of carefully selecting low-reprojection error pixels by taking into account semantic information and confidence scores. Furthermore, we quantitatively and qualitatively demonstrated that dynamically updating the masks enables more robust selection of error points, thus enabling better training efficiency and effectiveness. Our ablation studies validated the effectiveness of the techniques adopted in our method, and solidified their roles in enhancing the performance." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 230, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 230, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 230, + 101 + ], + "type": "text", + "content": "T.-R. Liu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 246, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 246, + 129 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 246, + 129 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 138, + 482, + 246 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 138, + 482, + 246 + ], + "spans": [ + { + "bbox": [ + 130, + 138, + 482, + 246 + ], + "type": "text", + "content": "The authors gratefully acknowledge the support from the National Science and Technology Council (NSTC) in Taiwan under grant numbers MOST 111-2223-E007-004-MY3, 113-2221-E-007-122-MY3, 113-2640-E-002-003, 113-2221-E-007-104-MY3, as well as the financial support from Woven by Toyota, Inc., Japan. The authors would also like to express their appreciation for the donation of the GPUs from NVIDIA Corporation and NVIDIA AI Technology Center (NVAITC) used in this work. Furthermore, the authors extend their gratitude to the National Center for High-Performance Computing (NCHC) for providing the necessary computational and storage resources." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 262, + 198, + 275 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 262, + 198, + 275 + ], + "spans": [ + { + "bbox": [ + 132, + 262, + 198, + 275 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 285, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 136, + 285, + 481, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 285, + 481, + 319 + ], + "spans": [ + { + "bbox": [ + 136, + 285, + 481, + 319 + ], + "type": "text", + "content": "1. Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Dlarr, and Ross Girshick. Segment anything. arXiv:2304.02643, 2023." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 136, + 319, + 481, + 350 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 319, + 481, + 350 + ], + "spans": [ + { + "bbox": [ + 136, + 319, + 481, + 350 + ], + "type": "text", + "content": "2. Xiao-Shan Gao, Xiao-Rong Hou, Jianliang Tang, and Hang-Fei Cheng. Complete solution classification for the perspective-three-point problem. IEEE transactions on pattern analysis and machine intelligence, 25(8):930-943, 2003." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 136, + 350, + 481, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 350, + 481, + 384 + ], + "spans": [ + { + "bbox": [ + 136, + 350, + 481, + 384 + ], + "type": "text", + "content": "3. Martin A Fischler and Robert C Bolles. Random sample consensus: a paradigm for model fitting with applications to image analysis and automated cartography. Communications of the ACM, 1981." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 136, + 384, + 481, + 416 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 384, + 481, + 416 + ], + "spans": [ + { + "bbox": [ + 136, + 384, + 481, + 416 + ], + "type": "text", + "content": "4. Paul-Edouard Sarlin, Cesar Cadena, Roland Siegwart, and Marcin Dymczyk. From coarse to fine: Robust hierarchical localization at large scale. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2019." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 136, + 416, + 481, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 416, + 481, + 459 + ], + "spans": [ + { + "bbox": [ + 136, + 416, + 481, + 459 + ], + "type": "text", + "content": "5. Jamie Shotton, Ben Glocker, Christopher Zach, Shahram Izadi, Antonio Criminisi, and Andrew Fitzgibbon. Scene coordinate regression forests for camera relocalization in RGB-D images. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2013." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 136, + 459, + 481, + 492 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 459, + 481, + 492 + ], + "spans": [ + { + "bbox": [ + 136, + 459, + 481, + 492 + ], + "type": "text", + "content": "6. Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Superpoint: Self-supervised interest point detection and description. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), pages 224-236, 2018." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 136, + 492, + 481, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 492, + 481, + 525 + ], + "spans": [ + { + "bbox": [ + 136, + 492, + 481, + 525 + ], + "type": "text", + "content": "7. Jiaming Sun, Zehong Shen, Yuang Wang, Hujun Bao, and Xiaowei Zhou. LoFTR: Detector-free local feature matching with transformers. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), pages 8922-8931, 2021." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 136, + 525, + 481, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 525, + 481, + 557 + ], + "spans": [ + { + "bbox": [ + 136, + 525, + 481, + 557 + ], + "type": "text", + "content": "8. Paul-Edouard Sarlin, Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. SuperGlue: Learning feature matching with graph neural networks. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2020." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 136, + 557, + 481, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 557, + 481, + 578 + ], + "spans": [ + { + "bbox": [ + 136, + 557, + 481, + 578 + ], + "type": "text", + "content": "9. Philipp Lindenberger, Paul-Edouard Sarlin, and Marc Pollefeys. LightGlue: Local feature matching at light speed. arXiv preprint arXiv:2306.13643, 2023." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 133, + 578, + 481, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 578, + 481, + 611 + ], + "spans": [ + { + "bbox": [ + 133, + 578, + 481, + 611 + ], + "type": "text", + "content": "10. Pablo Speciale, Johannes L Schonberger, Sing Bing Kang, Sudipta N Sinha, and Marc Pollefeys. Privacy preserving image-based localization. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), pages 5493-5503, 2019." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 133, + 611, + 481, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 611, + 481, + 643 + ], + "spans": [ + { + "bbox": [ + 133, + 611, + 481, + 643 + ], + "type": "text", + "content": "11. Xiaotian Li, Juha Ylioinas, Jakob Verbeek, and Juho Kannala. Scene coordinate regression with angle-based reprojection loss for camera relocalization. In Proc. European Conf. on Computer Vision Workshop (ECCVW), 2018." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 133, + 643, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 643, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 643, + 481, + 665 + ], + "type": "text", + "content": "12. Xiaotian Li, Juha Ylioinas, and Juho Kannala. Full-frame scene coordinate regression for image-based localization. RSS, 2018." + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 146, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 146, + 91, + 447, + 102 + ], + "type": "text", + "content": "Reprojection Errors as Prompts for Efficient Scene Coordinate Regression" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 481, + 160 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 481, + 160 + ], + "type": "text", + "content": "13. Eric Brachmann, Alexander Krull, Sebastian Nowozin, Jamie Shotton, Frank Michel, Stefan Gumhold, and Carsten Rother. DSAC-Differentiable RANSAC for camera localization. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2017." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 160, + 481, + 192 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 160, + 481, + 192 + ], + "spans": [ + { + "bbox": [ + 133, + 160, + 481, + 192 + ], + "type": "text", + "content": "14. Eric Brachmann and Carsten Rother. Learning less is more - 6D camera localization via 3D surface regression. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2018." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 192, + 481, + 213 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 192, + 481, + 213 + ], + "spans": [ + { + "bbox": [ + 133, + 192, + 481, + 213 + ], + "type": "text", + "content": "15. Eric Brachmann and Carsten Rother. Visual camera re-localization from RGB and RGB-D images using DSAC. TPAMI, 2021." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 213, + 481, + 235 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 213, + 481, + 235 + ], + "spans": [ + { + "bbox": [ + 133, + 213, + 481, + 235 + ], + "type": "text", + "content": "16. Eric Brachmann and Carsten Rother. Expert sample consensus applied to camera re-localization. In Proc. IEEE Int. Conf. on Computer Vision (ICCV), 2019." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 133, + 235, + 481, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 235, + 481, + 267 + ], + "spans": [ + { + "bbox": [ + 133, + 235, + 481, + 267 + ], + "type": "text", + "content": "17. Eric Brachmann, Tommaso Cavallari, and Victor Adrian Prisacariu. Accelerated coordinate encoding: Learning to relocalize in minutes using RGB and poses. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2023." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 133, + 267, + 481, + 299 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 267, + 481, + 299 + ], + "spans": [ + { + "bbox": [ + 133, + 267, + 481, + 299 + ], + "type": "text", + "content": "18. Alex Kendall, Matthew Grimes, and Roberto Cipolla. PoseNet: A convolutional network for real-time 6-DOF camera relocalization. In Proc. IEEE Int. Conf. on Computer Vision (ICCV), December 2015." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 133, + 299, + 481, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 299, + 481, + 331 + ], + "spans": [ + { + "bbox": [ + 133, + 299, + 481, + 331 + ], + "type": "text", + "content": "19. Tien Do, Ondrej Miksik, Joseph DeGol, Hyun Soo Park, and Sudipta N. Sinha. Learning to detect scene landmarks for camera localization. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2022." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 133, + 331, + 481, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 331, + 481, + 374 + ], + "spans": [ + { + "bbox": [ + 133, + 331, + 481, + 374 + ], + "type": "text", + "content": "20. Abner Guzman-Rivera, Pushmeet Kohli, Ben Glocker, Jamie Shotton, Toby Sharp, Andrew Fitzgibbon, and Shahram Izadi. Multi-output learning for camera localization. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2014." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 374, + 481, + 417 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 374, + 481, + 417 + ], + "spans": [ + { + "bbox": [ + 132, + 374, + 481, + 417 + ], + "type": "text", + "content": "21. Eric Brachmann, Frank Michel, Alexander Krull, Michael Ying Yang, Stefan Gumhold, and carsten Rother. Uncertainty-driven 6D pose estimation of objects and scenes from a single RGB image. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2016." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 417, + 481, + 460 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 417, + 481, + 460 + ], + "spans": [ + { + "bbox": [ + 132, + 417, + 481, + 460 + ], + "type": "text", + "content": "22. Julien Valentin, Matthias Niessner, Jamie Shotton, Andrew Fitzgibbon, Shahram Izadi, and Philip H. S. Torr. Exploiting uncertainty in regression forests for accurate camera relocalization. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2015." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 460, + 481, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 460, + 481, + 502 + ], + "spans": [ + { + "bbox": [ + 132, + 460, + 481, + 502 + ], + "type": "text", + "content": "23. Tao Xie, Kun Dai, Siyi Lu, Ke Wang, Zhiqiang Jiang, Jinghan Gao, Dedong Liu, Jie Xu, Lijun Zhao, and Ruifeng Li. OFVL-MS: Once for visual localization across multiple indoor scenes. In Proc. IEEE Int. Conf. on Computer Vision (ICCV), 2023." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 502, + 481, + 546 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 502, + 481, + 546 + ], + "spans": [ + { + "bbox": [ + 132, + 502, + 481, + 546 + ], + "type": "text", + "content": "24. Zhaoyang Huang, Han Zhou, Yijin Li, Bangbang Yang, Yan Xu, Xiaowei Zhou, Hujun Bao, Guofeng Zhang, and Hongsheng Li. VS-Net: Voting with segmentation for visual localization. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2021." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 546, + 481, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 546, + 481, + 578 + ], + "spans": [ + { + "bbox": [ + 132, + 546, + 481, + 578 + ], + "type": "text", + "content": "25. Tien Do and Sudipta N. Sinha. Improved scene landmark detection for camera localization. In Proceedings of the International Conference on 3D Vision (3DV), 2024." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 578, + 481, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 578, + 481, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 578, + 481, + 632 + ], + "type": "text", + "content": "26. Paul-Edouard Sarlin, Ajaykumar Unagar, Mans Larsson, Hugo Germain, Carl Toft, Victor Larsson, Marc Pollefeys, Vincent Lepetit, Lars Hammarstrand, Fredrik Kahl, and Torsten Sattler. Back to the Feature: Learning Robust Camera Localization from Pixels to Pose. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2021." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 632, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 481, + 665 + ], + "type": "text", + "content": "27. Mohammad Altillawi. PixSelect: Less but reliable pixels for accurate and efficient localization. In Proc. IEEE Int. Conf. on Robotics and Automation (ICRA), pages 4156-4162. IEEE, 2022." + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 229, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 229, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 229, + 100 + ], + "type": "text", + "content": "T.-R. Liu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 482, + 664 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 132, + 116, + 482, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 482, + 149 + ], + "type": "text", + "content": "28. Fei Xue, Ignas Budvytis, Daniel Olmeda Reino, and Roberto Cipolla. Efficient large-scale localization by global instance recognition. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), pages 17348-17357, 2022." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 149, + 482, + 181 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 149, + 482, + 181 + ], + "spans": [ + { + "bbox": [ + 132, + 149, + 482, + 181 + ], + "type": "text", + "content": "29. Fei Xue, Ignas Budvytis, and Roberto Cipolla. SFD2: Semantic-guided feature detection and description. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), pages 5206-5216, 2023." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 181, + 482, + 213 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 181, + 482, + 213 + ], + "spans": [ + { + "bbox": [ + 132, + 181, + 482, + 213 + ], + "type": "text", + "content": "30. Bach-Thuan Bui, Dinh-Tuan Tran, and Joo-Ho Lee. D2S: Representing local descriptors and global scene coordinates for camera relocalization. arXiv preprint arXiv:2307.15250, 2023." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 213, + 482, + 256 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 213, + 482, + 256 + ], + "spans": [ + { + "bbox": [ + 132, + 213, + 482, + 256 + ], + "type": "text", + "content": "31. Son Tung Nguyen, Alejandro Fontan, Michael Milford, and Tobias Fischer. Focus-Tune: Tuning visual localization through focus-guided sampling. In Proc. IEEE Winter Conf. on Applications of Computer Vision (WACV), pages 3606-3615, January 2024." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 256, + 482, + 299 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 256, + 482, + 299 + ], + "spans": [ + { + "bbox": [ + 132, + 256, + 482, + 299 + ], + "type": "text", + "content": "32. Johanna Wald, Torsten Sattler, Stuart Golodetz, Tommaso Cavallari, and Federico Tombari. Beyond controlled environments: 3d camera re-localization in changing indoor scenes. In Proc. European Conf. on Computer Vision (ECCV), pages 467-487, 2020." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 300, + 482, + 343 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 300, + 482, + 343 + ], + "spans": [ + { + "bbox": [ + 132, + 300, + 482, + 343 + ], + "type": "text", + "content": "33. Siyan Dong, Qingnan Fan, He Wang, Ji Shi, Li Yi, Thomas Funkhouser, Baoquan Chen, and Leonidas J Guibas. Robust neural routing through space partitions for camera relocalization in dynamic indoor environments. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2021." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 343, + 482, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 343, + 482, + 396 + ], + "spans": [ + { + "bbox": [ + 132, + 343, + 482, + 396 + ], + "type": "text", + "content": "34. Donghwan Lee, Soohyun Ryu, Suyong Yeon, Yonghan Lee, Deokhwa Kim, Cheolho Han, Yohann Cabon, Philippe Weinzaepfel, Nicolas Guerin, Gabriela Csurka, and Martin Humenberger. Large-scale localization datasets in crowded indoor spaces. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), pages 3227-3236, June 2021." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 396, + 482, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 396, + 482, + 430 + ], + "spans": [ + { + "bbox": [ + 132, + 396, + 482, + 430 + ], + "type": "text", + "content": "35. Hongyi Fan, Joe Kileel, and Benjamin Kimia. On the instability of relative pose estimation and RANSAC's role. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), pages 8935-8943, June 2022." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 430, + 482, + 460 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 430, + 482, + 460 + ], + "spans": [ + { + "bbox": [ + 132, + 430, + 482, + 460 + ], + "type": "text", + "content": "36. Zhe Chen, Yuchen Duan, Wenhai Wang, Junjun He, Tong Lu, Jifeng Dai, and Yu Qiao. Vision transformer adapter for dense predictions. arXiv preprint arXiv:2205.08534, 2022." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 460, + 482, + 494 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 460, + 482, + 494 + ], + "spans": [ + { + "bbox": [ + 132, + 460, + 482, + 494 + ], + "type": "text", + "content": "37. Han Cai, Junyan Li, Muyan Hu, Chuang Gan, and Song Han. EfficientViT: Lightweight multi-scale attention for high-resolution dense prediction. In Proc. IEEE Int. Conf. on Computer Vision (ICCV), pages 17302-17313, 2023." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 494, + 482, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 494, + 482, + 525 + ], + "spans": [ + { + "bbox": [ + 132, + 494, + 482, + 525 + ], + "type": "text", + "content": "38. Alex Kendall and Yarin Gal. What uncertainties do we need in bayesian deep learning for computer vision? Advances in neural information processing systems, 30, 2017." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 525, + 482, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 525, + 482, + 557 + ], + "spans": [ + { + "bbox": [ + 132, + 525, + 482, + 557 + ], + "type": "text", + "content": "39. Sheng Wan, Tung-Yu Wu, Wing H Wong, and Chen-Yi Lee. Confnet: predict with confidence. In 2018 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 2921-2925. IEEE, 2018." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 557, + 482, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 557, + 482, + 601 + ], + "spans": [ + { + "bbox": [ + 132, + 557, + 482, + 601 + ], + "type": "text", + "content": "40. Shitao Tang, Sicong Tang, Andrea Tagliasacchi, Ping Tan, and Yasutaka Furukawa. Neumap: Neural coordinate mapping by auto-transdecoder for camera localization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 929-939, June 2023." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 601, + 482, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 601, + 482, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 601, + 482, + 632 + ], + "type": "text", + "content": "41. Revaud, Jérôme and Cabon, Yohann and Brégier, Romain and Lee, JongMin and Weinzaepfel, Philippe. SACReg: Scene-Agnostic Coordinate Regression for Visual Localization, 2023." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 632, + 482, + 664 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 482, + 664 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 482, + 664 + ], + "type": "text", + "content": "42. Johannes L. Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), June 2016." + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 146, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 146, + 91, + 447, + 102 + ], + "type": "text", + "content": "Reprojection Errors as Prompts for Efficient Scene Coordinate Regression" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Reshaping the Online Data Buffering and Organizing Mechanism for Continual Test-Time Adaptation/69885bc8-fe2a-47b1-a255-a982ec605708_content_list.json b/2024/Reshaping the Online Data Buffering and Organizing Mechanism for Continual Test-Time Adaptation/69885bc8-fe2a-47b1-a255-a982ec605708_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..96a94712d4cca347575bed16258706c6c1fe2120 --- /dev/null +++ b/2024/Reshaping the Online Data Buffering and Organizing Mechanism for Continual Test-Time Adaptation/69885bc8-fe2a-47b1-a255-a982ec605708_content_list.json @@ -0,0 +1,1648 @@ +[ + { + "type": "text", + "text": "Reshaping the Online Data Buffering and Organizing Mechanism for Continual Test-Time Adaptation", + "text_level": 1, + "bbox": [ + 228, + 140, + 774, + 209 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zhilin Zhu $^{1,2}$ , Xiaopeng Hong $^{1,2\\star}$ , Zhiheng Ma $^{3,4,5}$ , Weijun Zhuang $^{1,2}$ , Yaohui Ma $^{1,5}$ , Yong Dai $^{2}$ , and Yaowei Wang $^{2,1}$", + "bbox": [ + 220, + 233, + 781, + 265 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Harbin Institute of Technology", + "bbox": [ + 390, + 276, + 611, + 290 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "2 Pengcheng Laboratory", + "bbox": [ + 418, + 291, + 583, + 305 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3 Shenzhen University of Advanced Technology", + "bbox": [ + 343, + 305, + 658, + 318 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "4 Guangdong Provincial Key Laboratory of Computility Microelectronics", + "bbox": [ + 256, + 319, + 745, + 332 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Shenzhen Institute of Advanced Technology, Chinese Academy of Science {zhuzhilin, weijunzhuang}@stu.hit.edu.cn hongxiaopeng@ieee.org ma@siat.ac.cn yh.ma@siat.ac.cn chd-dy@foxmail.com wangyw@pcl.ac.cn", + "bbox": [ + 256, + 333, + 766, + 375 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Continual Test-Time Adaptation (CTTA) involves adapting a pre-trained source model to continually changing unsupervised target domains. In this paper, we systematically analyze the challenges of this task: online environment, unsupervised nature, and the risks of error accumulation and catastrophic forgetting under continual domain shifts. To address these challenges, we reshape the online data buffering and organizing mechanism for CTTA. We propose an uncertainty-aware buffering approach to identify and aggregate significant samples with high certainty from the unsupervised, single-pass data stream. Based on this, we propose a graph-based class relation preservation constraint to overcome catastrophic forgetting. Furthermore, a pseudo-target replay objective is used to mitigate error accumulation. Extensive experiments demonstrate the superiority of our method in both segmentation and classification CTTA tasks. Code is available at https://github.com/z1358/OBAO.", + "bbox": [ + 259, + 400, + 740, + 595 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Continual test-time adaptation $\\cdot$ Unsupervised learning $\\cdot$ Continual learning $\\cdot$ Catastrophic forgetting", + "bbox": [ + 259, + 606, + 740, + 635 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 655, + 375, + 671 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Deep learning models have exhibited remarkable advancements in scenarios characterized by a consistent data distribution between training and test environments. However, the deployment of machine perception systems in the real world frequently exposes them to non-stationary and constantly changing environments [14, 21, 47, 53, 60]. When such phenomena occur (e.g., unexpected domain shifts over time), these models are prone to severe performance degradation [18, 37, 51]. Consequently, the development of techniques for continual adaptation during the unsupervised deployment phase is essential for enhancing the reliability of machine perception systems.", + "bbox": [ + 212, + 681, + 787, + 819 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Corresponding author", + "bbox": [ + 217, + 824, + 382, + 839 + ], + "page_idx": 0 + }, + { + "type": "table", + "img_path": "images/f693bf9a99303dbedc915dd5f53517af4c9c315f3edfb13b6ea0fa886d1055c9.jpg", + "table_caption": [ + "Table 1: Comparison of different adaptation settings. We differentiate along four dimensions: the available form of the source domain data and target domain data, the distribution and access mode of the target domain data stream during adaptation." + ], + "table_footnote": [], + "table_body": "
SettingDataTarget domain data stream
SourceTargetDistributionAccess mode
Fine-tuning-xT,yTstationaryoffline (multi-epoch)
Domain GeneralizationxS,yS---
Domain AdaptationxS,ySxTstationaryoffline (multi-epoch)
Test-Time TrainingxS,ySxTstationaryonline (single-pass)
Test-Time Adaptation-xTstationaryoffline or online
Continual Test-Time Adaptation-x1T→···→xnTcontinually changingonline (single-pass)
", + "bbox": [ + 230, + 198, + 777, + 309 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "With sufficient source labeled data accessible, the fields of domain generalization [22, 23, 63, 66] and domain adaptation [20, 24, 25, 28] have witnessed considerable success. However, a particularly challenging scenario arises when no source data is available, and unsupervised adaptation must happen at test time when unlabeled test inputs arrive. This task is known as test-time adaptation (TTA) [5, 7, 45]. Given the rarity of encountering only a singular domain shift in practical applications, the community has recently been increasingly focusing on continual test-time adaptation (CTTA) [2, 12, 30, 51, 58], i.e., evaluating the ability of methods to adapt to continual domain shifts. We summarize the differences between existing adaptation settings in Tab. 1. Obviously, the considered CTTA setting is the most practical and challenging.", + "bbox": [ + 212, + 339, + 787, + 505 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Within this setting, we expose the key challenges as follows: (1) Online environment. This denotes a scenario where data is processed in a single pass, necessitating immediate predictions from the model. Due to the characteristics of the online environment, it is difficult for the model to obtain a sufficient number of informative samples, thereby impeding the efficacy of adaptation; (2) Unsupervised nature. Throughout adaptation, ground-truth labels for input data are unavailable. Moreover, due to privacy concerns or legal restrictions, the source domain data may no longer be available and we only have access to the source pre-trained model; (3) Error accumulation and catastrophic forgetting. The dynamic distribution shift when adapting to continually changing target domains can render pseudo-labels unreliable and impede knowledge retention about the source domain, leading to error accumulation [51, 54] and triggering catastrophic forgetting [6, 30, 58].", + "bbox": [ + 212, + 506, + 787, + 703 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Addressing the first challenge, an intuitive solution is to employ a buffer to store high-value samples. However, existing methods that utilize such a buffer store source domain samples with ground-truth labels [12, 43], which is difficult to satisfy in real-world scenarios and conflicts with the source-free constraint in the second challenge. Therefore, the issue of effectively buffering and organizing high-value samples in an unsupervised single-pass data stream has not been fully explored. Current CTTA methods primarily address the third challenge, focusing mainly on alignments in either the parameter space [2, 35, 40, 51] or the feature space [4, 12] of the source model to mitigate knowledge forgetting.", + "bbox": [ + 212, + 704, + 787, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "Z. Zhu et al.", + "bbox": [ + 271, + 114, + 357, + 127 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/fad5669ba277ebfa7db9e8f59138252a4fb8af44d55f35bf7fdc27916c238c29.jpg", + "image_caption": [ + "(a) W/o a buffer" + ], + "image_footnote": [], + "bbox": [ + 240, + 146, + 390, + 210 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/70dbd4dbd6ff3daec0498c281c295a40103b6c089478258ce7fb34141148fdeb.jpg", + "image_caption": [ + "(b) Static source buffer" + ], + "image_footnote": [], + "bbox": [ + 424, + 146, + 576, + 227 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/2f4e1d8c73e641ba711b0b528813684ae73945523f6b4b69581c6490f3e639c7.jpg", + "image_caption": [ + "(c) Our dynamic target buffer", + "Fig. 1: Comparison of different buffering mechanisms. (a) Without a buffer, only unreliable signals from the current unsupervised input data are available. (b) Some methods [12, 43] use a static source buffer that stores source samples with ground-truth labels to provide supervised signals. (c) We work on a dynamic target buffer, which can provide reliable signals for the adaptation process." + ], + "image_footnote": [], + "bbox": [ + 602, + 146, + 771, + 234 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this paper, we endeavor to address these three challenges simultaneously. To this end, we reshape the online data buffering and organizing mechanism for CTTA. As shown in Fig. 1, unlike previous methods which directly sample from the source domain [12, 43], we are targeted at source-free, unsupervised, and online buffering. Through a specially designed uncertainty-aware buffering approach, we dynamically evaluate and aggregate reliable information in the unsupervised, single-pass data stream (Challenges 1 and 2). Based on this, we efficiently organize these samples to assist the model's unsupervised adaptation process. Utilizing pseudo-labels generated from these low-uncertainty samples, we consider them as anchors of classes and organize them into a class relation graph (CRG). To combat catastrophic forgetting under continually changing domains, we introduce a novel graph-based class relation preservation constraint to maintain the intrinsic class relations obtained on the source domain (Challenge 3). Specifically, we penalize topological changes within the CRG to mitigate knowledge forgetting, while allowing reasonable vertex movements to enhance adaptability. Additionally, these samples are incorporated into a pseudo-target replay objective to reduce the risk of error accumulation from misclassified samples (Challenge 3). In summary, our main contributions are as follows:", + "bbox": [ + 212, + 364, + 787, + 638 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- To address the three revealed challenges simultaneously, we reshape the online data buffering and organizing mechanism for CTTA.", + "- To frame the buffering mechanism, we design an uncertainty-aware approach to identify and aggregate reliable data from unsupervised online data stream.", + "- Through a well-designed graph-based class relation preservation constraint and a pseudo-target replay objective, we effectively mitigate catastrophic forgetting and error accumulation in CTTA.", + "- We extensively evaluate our proposed method and the results on four challenging datasets show that the method clearly improves the CTTA performance on both segmentation and classification benchmarks, indicating its effectiveness and universality." + ], + "bbox": [ + 223, + 664, + 784, + 838 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "Reshaping the Online Data Buffering and Organizing Mechanism for CTTA", + "bbox": [ + 225, + 114, + 730, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 143, + 388, + 160 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.1 Domain Adaptation", + "text_level": 1, + "bbox": [ + 215, + 181, + 426, + 196 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Domain adaptation (DA) [38,65] aims to generalize a model from a well-annotated source domain to an unlabeled target domain following the transductive learning principle. The underlying assumption in DA is that both labeled source domain and unlabeled target domain data are concurrently accessible. Consequently, the prevalent methods in DA encompass the utilization of domain adversarial training [8,11,16,19,29,49] or the minimization of domain discrepancy [32,33,39]. These methods work towards harmonizing the feature distribution [16,32,49] or the input space [19,59] between the two domains.", + "bbox": [ + 215, + 210, + 789, + 332 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.2 Continual Test-Time Adaptation", + "text_level": 1, + "bbox": [ + 215, + 359, + 529, + 375 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Test-Time Adaptation (TTA). TTA focuses on a more challenging setting after model deployment where only a pre-trained source model and unlabeled target data are available. Some approaches [5, 27] focus on the offline setting, where multiple epochs of access to the entire set of test data are provided for adaptation, while others [1, 7, 31, 52] focus on the online setting with a single-pass data stream. To name a few, TENT [50] adapts the model by using entropy minimization loss to update a few trainable parameters in batch normalization layers. This approach has also inspired subsequent works such as EATA [36] and RoTTA [61] to investigate the robustness of normalized layers. Due to the domain gap, TSD [52] promotes feature alignment and uniformity from the perspective of feature revision, while TIPI [34] proposes invariance regularizers to find the transformation space that can simulate the domain shift.", + "bbox": [ + 215, + 388, + 787, + 570 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Continual Test-Time Adaptation (CTTA). In recent years, CTTA has received increasing research attention [15,44,51,58] due to its practicality. CoTTA [51] stands out as the first method specifically tailored to address this task. It utilizes weight-averaged and augmentation-averaged predictions to reduce error accumulation, while stochastically restoring the weights of a small fraction of neurons to the source model during each iteration to mitigate catastrophic forgetting. Building upon this, PETAL [2] further introduces a data-driven parameter recovery technique to align with the source model in the parameter space, while RMT [12] and SANTA [4] seek alignment with the source model in the feature space by optimizing an additional contrastive learning objective. Additionally, DSS [54] considers sample security to mitigate the risk of using incorrect information. Some approaches [15, 30, 44] also prevent forgetting by freezing a majority or all of the model parameters, but they require optimization along with the model during training using the source domain data, thereby relaxing the unsupervised challenge and not being truly source-free. Under our proposal, we simultaneously address three challenges in CTTA.", + "bbox": [ + 215, + 597, + 787, + 839 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 230, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "Z. Zhu et al.", + "bbox": [ + 271, + 114, + 357, + 127 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/c625ea5d503b5153c2d15d12f3ebce3d9c91ec6f5d8c0cc585af3169df2bf172.jpg", + "image_caption": [ + "Fig. 2: The framework of our proposed method. At time step $t$ , the incoming data and sampled buffer data are used for adaptation. The proposed class relation preservation loss $\\mathcal{L}_{CRP}$ ensures topological consistency between the Class Relation Graph (CRG) $G_{t}$ estimated at the current time step $t$ and $\\hat{G}$ from the source domain, thereby effectively preserving intrinsic semantic information. Concurrently, we integrate a pseudo-target replay loss $\\mathcal{L}_{PCE}$ to mitigate error accumulation." + ], + "image_footnote": [], + "bbox": [ + 256, + 147, + 743, + 358 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3 Methodology", + "text_level": 1, + "bbox": [ + 215, + 482, + 380, + 500 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1 Preliminaries and Overall Framework", + "text_level": 1, + "bbox": [ + 215, + 513, + 571, + 527 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In continual test-time adaptation (CTTA), we are given a source model $f_{\\theta_0}$ pretrained from the labeled source domain $(\\mathcal{X}^S, \\mathcal{Y}^S)$ and a series of sequentially changing unlabeled target domains such as $[\\mathcal{X}_1^T, \\mathcal{X}_2^T, \\dots, \\mathcal{X}_n^T]$ . Each target domain shares the same label space as the source domain (for close-set adaptation). Note that the model has no information about when the domain changes happen during the adaptation process. The goal of CTTA is to adapt the model to a continually changing unsupervised online data stream of target domains without source data. To be more specific, at inference time step $t$ , the unlabeled target test data $x(t)$ is given to the model $f_{\\theta_t}$ , and the model needs to make predictions and adapt itself accordingly $(f_{\\theta_t} \\rightarrow f_{\\theta_{t+1}})$ for the upcoming test inputs $x(t+1)$ .", + "bbox": [ + 212, + 537, + 787, + 688 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We adopt a teacher-student framework following previous works [2,12,51] as illustrated in Fig. 2. Leveraging the proposed uncertainty-aware buffering mechanism, our method dynamically evaluates and buffers significant high-certainty samples in the unsupervised online data stream, providing reliable information for the model's adaptation process. Building upon this, we further introduce a novel class relation preservation constraint $(\\mathcal{L}_{CRP})$ for preserving domain-invariant semantic information in dynamic environments, effectively mitigating the catastrophic forgetting problem. Additionally, we address error accumulation through a pseudo-target replay objective $(\\mathcal{L}_{PCE})$ to minimize the impact of miscalibration. A self-training loss $(\\mathcal{L}_{ST})$ is used to stabilize the adaptation", + "bbox": [ + 212, + 688, + 787, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "Reshaping the Online Data Buffering and Organizing Mechanism for CTTA", + "bbox": [ + 225, + 113, + 730, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "process of the student model (adapted model). In the adopted framework, both the teacher model $f_{\\theta'}$ and the student model $f_{\\theta}$ are initialized with the source model $f_{\\theta_0}$ . Subsequently, the teacher model is continuously updated by the exponential moving average of the student model. We elaborate our method in the following subsections.", + "bbox": [ + 212, + 146, + 787, + 220 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.2 Uncertainty-aware Buffering", + "text_level": 1, + "bbox": [ + 215, + 243, + 496, + 258 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In the unsupervised online data stream within CTTA, the model necessitates adaptation utilizing the data at hand, typically characterized by a scarcity of information. While certain methods [12, 43] endeavor to incorporate additional supervised information by formulating a static buffer from the source domain, fulfilling this requirement proves challenging within practical application scenarios. Conversely, our proposition involves sampling and buffering throughout the unsupervised adaptation process by a designed uncertainty-aware approach.", + "bbox": [ + 212, + 267, + 787, + 372 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "At the outset, the dynamic buffer is initialized as empty. As the adaptation progresses, the buffer undergoes continual updates to reflect the evolving nature of the data stream encountered. Specifically, at any given inference time step $t$ , the model calculates the uncertainty associated with the current batch of samples. The quantification of sample uncertainty manifests as the entropy of the sample, denoted as:", + "bbox": [ + 212, + 373, + 789, + 463 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nH (x (t)) = - f _ {\\theta_ {t}} (x (t)) ^ {\\top} (\\log f _ {\\theta_ {t}} (x (t))), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 370, + 473, + 785, + 491 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $f_{\\theta_t}(x(t))$ corresponds the softmax predictions of the student model for the current input data. Samples exhibiting an entropy value less than a predefined threshold $H_0$ are deemed to possess a high degree of certainty and are thus considered candidates for inclusion in the dynamic buffer. $H_0$ is defined as:", + "bbox": [ + 212, + 500, + 787, + 561 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nH _ {0} = \\alpha \\ln C, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 457, + 573, + 785, + 588 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\alpha$ serves as a threshold coefficient and $C$ denotes the total number of classes in the dataset. The process of storing samples into the buffer, or updating existing entries, is contingent upon the buffer's capacity. High-certainty samples identified from the present batch, along with their derived pseudo-labels from the teacher model, are added to the buffer if space permits. Conversely, in scenarios where the buffer is at capacity, these new samples replace the ones exhibiting the highest uncertainty to maintain an optimal blend of timeliness and reliability. Upon completion of these updates, the resulting buffer $\\mathcal{M}_t$ stands ready to facilitate the model's adaptation in subsequent iterations.", + "bbox": [ + 212, + 598, + 787, + 733 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Notably, our approach to buffering is entirely devoid of reliance on source data, underscoring its source-free nature. This strategy enables the model to discern and prioritize data based on their uncertainty. By dynamically buffering high-certainty samples within the data stream, our approach not only circumvents the limitations posed by reliance on static source buffers but also enriches the adaptation process with valuable, reliable data. The intuitive idea is to utilize CE loss to replay samples in the buffer, thereby emphasizing reliable data", + "bbox": [ + 212, + 734, + 787, + 839 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "Z. Zhu et al.", + "bbox": [ + 271, + 114, + 357, + 127 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "and effectively suppressing the risk and impact of miscalibration due to error accumulation. In each test-time adaptation iteration, we randomly sample examples and labels $(\\tilde{x}(t), \\tilde{y}(t))$ from the buffer. The number of sampled examples is equal to the batch size and the corresponding pseudo-target replay loss is:", + "bbox": [ + 212, + 146, + 782, + 207 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {P C E} = - \\sum_ {c = 1} ^ {C} \\tilde {y} _ {c} (t) \\log \\tilde {p} _ {c}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 410, + 220, + 785, + 260 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $\\tilde{p}$ is the softmax prediction of the student model for $\\tilde{x}(t)$ . For simplicity, we omit the summation of all samples in the batch. Moreover, these sampled examples also serve as anchors and are organized in a class relation graph. On this basis, we introduce a novel class relation preservation constraint in Sec. 3.3 to preserve domain-invariant knowledge and overcome forgetting.", + "bbox": [ + 214, + 272, + 784, + 348 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.3 Class Relation Preservation Constraint", + "text_level": 1, + "bbox": [ + 215, + 372, + 583, + 386 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In online scenarios where domains are constantly changing, neural networks tend to \"take shortcuts\" and focus on simplistic features [17, 55] from the single-pass data stream. This phenomenon results in the model's forgetting of intrinsic semantic information. In this work, we propose to overcome catastrophic forgetting by preserving the domain-invariant class relation, which is consistent across domains regardless of distribution discrepancy [64]. Therefore, it is reasonable to utilize the intrinsic class relation obtained on the source domain with large labeled samples to constrain the model adaptation process on target domains.", + "bbox": [ + 212, + 398, + 784, + 518 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We describe here how to obtain the defined class relation graph (CRG). For computational stability, we normalize the feature space and adopt the cosine similarity metric. Let $\\bar{\\cdot}$ denotes the normalization operation, where $\\bar{f} = f / \\| f\\|$ . Given the normalized feature space $\\bar{\\mathcal{F}}$ , the CRG $G$ is defined as $G = \\langle V,E\\rangle$ , where $V = \\{\\bar{v}_1,\\dots ,\\bar{v}_C|\\bar{v}_i\\in \\bar{\\mathcal{F}}\\}$ is the set of $C$ vertices representative of $\\bar{\\mathcal{F}}$ , and $E$ is the edge set describing the neighborhood relations of vertices in $V$ . Each edge $e_{ij}$ is assigned with a weight $s_{ij}$ , which is the similarity between $\\bar{v}_i$ and $\\bar{v}_j$ : $s_{ij} = \\bar{v}_i^\\top \\bar{v}_j$ . Each vertex $\\bar{v}_i$ is the centroid vector of class $i$ . The intrinsic CRG $\\hat{G}$ is obtained by modeling the source-domain class relation. We demonstrate in experiments that the construction of our CRG $\\hat{G}$ can be derived from different data conditions, including class prototypes and classifier weights.", + "bbox": [ + 212, + 521, + 785, + 686 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "At time step $t$ , given CRG $G_{t} = \\langle V_{t}, E_{t} \\rangle$ , when catastrophic forgetting occurs, $G_{t}$ is distorted as the semantic relations among categories are confused. To mitigate forgetting, the model should be adapted to strive towards preserving the topology in $G_{t}$ the same as that in $\\hat{G}$ . This is achieved by restricting the neighboring relations of vertices described by the weights of edges (i.e., similarity between classes) in $E_{t}$ . One way to do this is to keep the ranking of the edge weights constant during the adaptation process. However optimizing a non-smooth global ranking is difficult and inefficient [3,46]. An alternative and more feasible approach involves assessing the extent of change in neighboring relations [13,48] by evaluating the correlation between the initial edge weights", + "bbox": [ + 212, + 688, + 785, + 840 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "Reshaping the Online Data Buffering and Organizing Mechanism for CTTA", + "bbox": [ + 225, + 114, + 730, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "and their current observation. A lower correlation indicates a higher probability that the rank of an edge has changed during adaptation, and this should be penalized. On this basis, we define the Class Relation Preservation constraint (CRP) as follows:", + "bbox": [ + 212, + 146, + 787, + 205 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {C R P} = - \\frac {\\sum_ {i , j} ^ {C} s _ {i j} \\tilde {s} _ {i j}}{\\sqrt {\\sum_ {i , j} ^ {C} s _ {i j} ^ {2}} \\sqrt {\\sum_ {i , j} \\tilde {s} _ {i j} ^ {2}}}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 388, + 218, + 785, + 267 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $S = \\{s_{ij}|1\\leq i,j\\leq C\\}$ and $\\tilde{S} = \\{\\tilde{s}_{ij}|1\\leq i,j\\leq C\\}$ are the sets of the initial and observation values of the edge weights in $\\hat{E}$ and $E_{t}$ , respectively.", + "bbox": [ + 212, + 272, + 785, + 306 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To achieve end-to-end optimization, the ideal situation is using source-domain images with ground-truth labels to obtain an estimate of the current CRG $G_{t}$ under time step $t$ . That is, extracting features from these images by $f_{\\theta_t}$ , which in turn yields the class prototypes and ultimately the class relation. However, it is often impractical to store a certain number of source-domain samples in the source-free CTTA setting. In this work, we leverage reliable data from our proposed uncertainty-aware buffering mechanism to accurately estimate CRG $G_{t}$ and achieve end-to-end optimization in an unsupervised test data stream.", + "bbox": [ + 212, + 306, + 785, + 426 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "From the perspective of viewing $S$ as a vector, the proposed CRP term constrains the consistency of the direction between two vectors during the adaptation process. Opting for the origin moment instead of the central moment allows the CRP term to prioritize direction consistency between vectors. This makes it more robust to scale variations or magnitude disparities in edge weights, which may arise from shifts in data distribution. Some previous methods [4,12] directly penalize the movement of target domain samples in the feature space relative to the source prototypes. This can be broadly interpreted as penalizing the movement of corresponding elements between $\\hat{V}$ and $V_{t}$ in our defined CRG. However, in addition to suffering from classification errors when assigning source prototypes, constraints on features can also impede the effectiveness of adaptation in target domains. Instead of penalizing the movement of vertices in $G$ in the feature space, the proposed CRP term penalizes the change of topological relations between vertices, while allowing reasonable movement of vertices to enhance adaptation. To summarize, by preserving the domain-invariant class relation during the model adaptation process, the proposed CRP term effectively overcomes catastrophic forgetting caused by semantic confusion.", + "bbox": [ + 212, + 426, + 787, + 684 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "3.4 Optimization Objective", + "text_level": 1, + "bbox": [ + 215, + 705, + 455, + 722 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "At time step $t$ , the teacher model $f_{\\theta_t'}$ first generates pseudo-labels to help the learning process of the student model $f_{\\theta_t}$ . Specifically, we compute the symmetric cross-entropy loss [12] between the outputs of the student model and the pseudo-labels. The self-training loss can be formulated as follows:", + "bbox": [ + 212, + 729, + 787, + 791 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {S T} = - \\sum_ {c = 1} ^ {C} q _ {c} \\log p _ {c} - \\sum_ {c = 1} ^ {C} p _ {c} \\log q _ {c}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 372, + 803, + 785, + 843 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "Z. Zhu et al.", + "bbox": [ + 271, + 114, + 357, + 127 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $q$ is the softmax prediction of the teacher model, and $p$ is the softmax prediction of the student model, with $C$ being the total number of classes. Finally, the total loss function $\\mathcal{L}_T$ can be formulated as follows:", + "bbox": [ + 212, + 146, + 782, + 191 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {T} = \\mathcal {L} _ {S T} + \\mathcal {L} _ {P C E} + \\lambda_ {C R P} \\mathcal {L} _ {C R P}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 383, + 208, + 785, + 224 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where $\\lambda_{CRP}$ is a trade-off hyperparameter.", + "bbox": [ + 214, + 238, + 524, + 253 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 214, + 284, + 375, + 301 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We conduct comprehensive experiments under the CTTA setting as in [12, 51], encompassing three mainstream datasets ImageNet-to-ImageNet-C, CIFAR10-to-CIFAR10C, and CIFAR100-to-CIFAR100C for image classification, as well as Cityscapes-to-ACDC for semantic segmentation.", + "bbox": [ + 212, + 323, + 782, + 383 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.1 Datasets and Experimental Settings", + "text_level": 1, + "bbox": [ + 214, + 414, + 558, + 430 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "CIFAR10C, CIFAR100C, ImageNet-C. These datasets are originally developed to evaluate the robustness of classification networks [18]. Each dataset contains 15 distinct types of corruption at varying severity levels ranging from 1 to 5. These corruptions are applied to the test and validation images of CIFAR and ImageNet, respectively. We employ the clean training sets on CIFAR10, CIFAR100, and ImageNet as source domain datasets, while CIFAR10C, CIFAR100C, and ImageNet-C serve as the corresponding target domain datasets. Following the continuous benchmark setup in [51], we sequentially adapt the source model to 15 target domains under the largest severity level 5. The entire adaptation process is performed in an online manner, without assuming knowledge of when the domain changes. For each corruption, we utilize 10,000 images for datasets derived from CIFAR and 5,000 images for ImageNet-C.", + "bbox": [ + 212, + 446, + 787, + 628 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Cityscapes-to-ACDC. This dataset is designed for the continual semantic segmentation task [51] in autonomous driving. The Cityscapes dataset [9] is used as the source domain to provide an off-the-shelf pre-trained segmentation model, while the Adverse Conditions Dataset (ACDC) [41], which contains images collected under four different weather conditions: Fog, Night, Rain, and Snow, is used to represent the changing target domains. 400 unlabeled images from each condition are used for adaptation. Note that the ACDC dataset shares the same semantic classes as the evaluation classes of Cityscapes. To simulate continual distribution shifts encountered in real-world scenarios, we cyclically repeat the same sequence of target domains ten times (e.g., in total 40: Fog $\\rightarrow$ Night $\\rightarrow$ Rain $\\rightarrow$ Snow $\\rightarrow$ Fog $\\rightarrow$ ...). This also provides a long-term perspective on the performance evaluation of different adaptation methods.", + "bbox": [ + 212, + 657, + 787, + 840 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "Reshaping the Online Data Buffering and Organizing Mechanism for CTTA", + "bbox": [ + 225, + 114, + 732, + 130 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/421f46a30db28a1dc069f981a4402528b43c197fea074ea37d0c68dd4b3566c8.jpg", + "table_caption": [ + "Table 2: Semantic segmentation results (mIoU in %) on the Cityscapes-to-ACDC CTTA task. The four test conditions are repeated ten times to evaluate the long-term adaptation performance. For brevity, we only show the continual adaptation results solely in the first, fourth, seventh, and final rounds. All results are evaluated based on the Segformer-B5 architecture. Bold text indicates the best performance." + ], + "table_footnote": [], + "table_body": "
Timet
Round17
ConditionFegNightRainSnowMeanFegNightRainSnowMeanFegNightRainSnowMeanFegNightRainSnowMeanFegNightRainSnowMeanFegNightRainSnowMeanFegNightRainSnowMeanFegNightRainSnowMeanMean↑
Source69.140.359.757.856.769.140.359.757.856.769.140.359.757.856.769.140.359.757.856.769.140.359.757.846.769.140.359.757.856.769.140.359.757.856.756.7
BN Stats Adapt62.338.054.653.052.062.338.054.653.052.062.338.054.653.052.062.338.054.653.052.062.338.054.653.046.762.338.054.653.052.052.0
TENT-cont. [50]69.040.260.157.356.766.536.358.754.053.064.232.855.350.950.864.232.855.350.950.861.829.851.947.847.852.3
CoTTA [51]70.941.262.459.758.670.941.262.459.758.670.941.262.459.758.670.941.262.459.758.670.941.262.459.747.858.6
Ours71.242.365.062.060.172.843.666.763.361.672.542.566.863.361.372.542.566.863.361.372.542.966.763.061.361.3
", + "bbox": [ + 233, + 226, + 769, + 325 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Implementation Details. To ensure consistency and comparability, we follow the implementation details from previous works [12, 51]. We adopt pre-trained WideResNet-28 [62] model for CIFAR10-to-CIFAR10C, pre-trained ResNeXt-29 [57] model for CIFAR100-to-CIFAR100C, and standard pre-trained ResNet-50 model for ImageNet-to-ImagenetC from Robustbench [10] for all methods. Our batch size is set to 64 for ImageNet-C and 200 for other classification experiments. We evaluate the model using online predictions obtained immediately as the data is encountered. For segmentation CTTA, we employ a transformer-based architecture Segformer-B5 [56] trained on Cityscapes as our segmentation model, and use down-sampled images from ACDC with a resolution of $960 \\times 540$ as network inputs. We utilize the Adam optimizer with a learning rate of 1e-4 and set the batch size to 1. To generate robust pseudo-labels, we use multi-scale input with flipping as the augmentation method following [51] for the segmentation experiments. When obtaining the intrinsic class relation $\\hat{G}$ , we utilize the weights of the classifier as proxies for classes, whereas in classification scenarios where feature prototypes are easily accessible [4, 12], we utilize the prototypes to do so. All experiments are conducted using an NVIDIA RTX4090 GPU.", + "bbox": [ + 212, + 352, + 787, + 609 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.2 Experiments on Segmentation CTTA", + "text_level": 1, + "bbox": [ + 214, + 633, + 568, + 648 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "We report results based on the mean intersection over union (mIoU) metric for the complex continual test-time semantic segmentation Cityscapes-to-ACDC task. As detailed in Tab. 2, a discernible decline in the performance of TENT across long-term sequences is observed, highlighting the pervasive issue of model degradation over time. Conversely, our proposed method effectively mitigates catastrophic forgetting and error accumulation, thereby consistently maintaining a high mean mIoU level across repeated sequences of target domains. Particularly noteworthy is our method's achievement of the highest mIoU score, showcasing an impressive relative improvement of $4.6\\%$ in mIoU compared to the previous sota method CoTTA [51], averaged across ten rounds. This significant improvement demonstrates the ability of our method to adapt to dynamic environments within dense prediction tasks over the long term. Additionally,", + "bbox": [ + 212, + 657, + 787, + 840 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "Z. Zhu et al.", + "bbox": [ + 271, + 114, + 357, + 127 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/e859bf18adab27983f5c49330d18085db872a801e43f91cd0e67f3f6e5cb5eea.jpg", + "image_caption": [ + "Fig. 3: Qualitative comparison of semantic segmentation on the Cityscapes-to-ACDC CTTA task. The results for all methods come from the final round." + ], + "image_footnote": [], + "bbox": [ + 222, + 145, + 777, + 334 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "there is a trend of incremental improvement in the average mIoU metric during the initial rounds, indicating the capacity of our method to extract and leverage the reliable information embedded within the unsupervised data stream. This capability facilitates further refinement of the adaptation performance. Comprehensive results are available in the supplementary material.", + "bbox": [ + 212, + 407, + 782, + 482 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "For a more intuitive validation of the effectiveness of our method, we conduct qualitative comparison experiments and show the results obtained through the CTTA process in Fig. 3. Compared to other state-of-the-art methods, our results exhibit superior clarity and accuracy, yielding the best segmentation maps in all four target domains. Our method provides consistent improvements for most of the categories (as shown by the white boxes), while the comparison methods mostly suffer from severe semantic confusion. These observations demonstrate the effectiveness of our method in complex real-world environments.", + "bbox": [ + 212, + 484, + 787, + 604 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4.3 Experiments on Classification CTTA", + "text_level": 1, + "bbox": [ + 215, + 630, + 565, + 645 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "ImageNet-to-ImageNet-C. Given the source model pre-trained on ImageNet, we sequentially execute CTTA on the 15 domains of ImageNet-C. As shown in Tab. 3, when directly testing the source model on target domains, the average classification error rate is as high as $82.0\\%$ , which demonstrates the need for adaptation. Leveraging the batch normalization statistics from the input data of the current iteration yielded the BN Stats Adapt method [26,42], a relatively straightforward approach that reduces the average classification error rate to $68.6\\%$ . While the TENT-based method [50] initially helps the model to continuously adapt to target domains, it is prone to severe catastrophic forgetting and error accumulation over long periods. CoTTA [51] and RMT [12] align to the source domain either in parameter space or feature space, respectively, intending to mitigate the problem of catastrophic forgetting, and both have made some", + "bbox": [ + 212, + 657, + 787, + 840 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "Reshaping the Online Data Buffering and Organizing Mechanism for CTTA", + "bbox": [ + 225, + 114, + 730, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/02fba1b0e3b3b2259050a851d9d2c5d274f77b832a456fc7a3201d406681f582.jpg", + "table_caption": [ + "Table 3: Classification error rate (\\%, lower is better) for the standard ImageNet-to-ImageNet-C CTTA task. All results are evaluated with the largest corruption severity level 5 in an online manner. Bold text indicates the best performance." + ], + "table_footnote": [], + "table_body": "
Timet
MethodGaussian shotinputsedefocusglassmotionzoomsnowfrostfogbrightnesscontrastelasticpixelatejpegMean↓
Source97.897.198.281.789.885.278.083.577.175.941.394.582.579.368.682.0
BN Stats Adapt85.083.785.084.784.373.761.266.068.252.134.982.755.951.359.868.6
TENT-cont. [50]81.674.672.777.673.865.555.361.663.051.738.272.150.847.453.362.6
CoTTA [51]84.782.180.681.379.068.657.560.360.548.336.666.147.241.246.062.7
RMT [12]80.276.474.577.174.466.257.657.059.148.039.160.647.342.543.460.2
DSS [54]84.680.478.783.979.874.962.962.862.949.737.471.049.542.948.264.6
Ours78.575.373.075.773.164.556.055.858.147.638.558.546.142.043.459.0
", + "bbox": [ + 233, + 198, + 769, + 345 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "progress. DSS [54] has noted the existence of high-quality and low-quality samples in the data stream, but it lacks further exploration on how to organize and utilize this information.", + "bbox": [ + 212, + 372, + 782, + 417 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "In our method, we reshape the online data buffering mechanism in CTTA, enabling dynamic aggregation of high-value samples from the unsupervised data stream. Based on this, we introduce a novel graph-based class relation preservation constraint alongside a pseudo-target replay objective. As a result, our method achieves a substantial reduction in the average classification error rate from $64.6\\%$ observed in the DSS method to $59.0\\%$ . Notably, our method outperforms all previous methods and reduces the error rate by a relative $2.0\\%$ compared to the previous sota method, averaged across 15 domains. Moreover, our method demonstrates the best performance on the vast majority of corruption types, highlighting its effectiveness in mitigating catastrophic forgetting and error accumulation.", + "bbox": [ + 212, + 417, + 787, + 583 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "CIFAR10-to-CIFAR10C, CIFAR100-to-CIFAR100C. To further demonstrate the effectiveness of the proposed method, we evaluate it on CIFAR10-to-CIFAR10C and CIFAR100-to-CIFAR100C CTTA tasks that have different numbers of categories. The experimental results are summarized in Tab. 4. Once again, our method outperforms all previous competing methods and reduces the average error rate to $15.8\\%$ and $29.0\\%$ , respectively. Compared with both the previous sota parameter space recovery methods (PETAL [2]) and feature space alignment methods (SANTA [4]), our method shows substantial improvements, demonstrating the superiority of the proposed class relation preservation mechanism in overcoming catastrophic forgetting. The consistency enhancement across datasets with varying levels of complexity proves the effectiveness and universality of our method.", + "bbox": [ + 212, + 606, + 787, + 787 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Experiments on Gradually Changing Setup. Following [4, 12, 51], we also experiment with a gradually changing setup. Different from the previous stan-", + "bbox": [ + 214, + 808, + 785, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "Z. Zhu et al.", + "bbox": [ + 271, + 114, + 357, + 127 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/c00ec0b0ae34b598cfb369fc718c5e1a1d4c2662f88a9cee7df966cea81f539b.jpg", + "table_caption": [ + "Table 4: Classification error rate (\\%, lower is better) for the standard CTTA tasks on CIFAR10-to-CIFAR10C, CIFAR100-to-CIFAR100C. All results are evaluated with the largest corruption severity level 5 in an online manner. Bold text indicates the best." + ], + "table_footnote": [], + "table_body": "
Timet
MethodGaussianMean↓
Source72.365.772.946.954.334.842.025.141.326.09.346.726.658.530.343.5
", + "bbox": [ + 233, + 198, + 769, + 468 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/5f7b8907d2c7c3cfff21a18ef470c8dee062f67bfad0ff3d0e7fbe8824b8767b.jpg", + "table_caption": [ + "Table 5: Gradually changing CIFAR10-to-CIFAR10C results. The severity level changes gradually between the lowest and the highest." + ], + "table_footnote": [], + "table_body": "
Avg. Error (%)SourceBN Stats AdaptTENT-cont. [50]CoTTA [51]RMT [12]SANTA [4]Our
CIFAR10C24.713.720.410.910.410.79.9
", + "bbox": [ + 233, + 521, + 769, + 550 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "dard setup where the corruption type changes abruptly at the largest severity level, the change process under the gradually changing setup can be represented as follows: $\\underbrace{..\\to 2\\to 1}_{\\text{before}}\\xrightarrow[{\\text{type}}]{\\text{change}}\\underbrace{1\\to 2\\to 3\\to 4\\to 5\\to 4\\to 3\\to 2\\to 1}_{\\text{current type, gradually changing severity}}\\xrightarrow[{\\text{type}}]{\\text{change}}$", + "bbox": [ + 212, + 578, + 787, + 643 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "$\\underbrace{1 \\rightarrow 2 \\rightarrow \\ldots}_{\\text {after }}$ , where each number denotes the severity of corruption. The exper", + "bbox": [ + 212, + 643, + 787, + 672 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Experimental results are shown in Tab. 5, and our proposed method also achieves superior performance in this setup, reducing the average error rate to $9.9\\%$ on the gradual CIFAR10-to-CIFAR10C benchmark.", + "bbox": [ + 212, + 670, + 787, + 715 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "4.4 Ablation Study", + "text_level": 1, + "bbox": [ + 214, + 739, + 390, + 753 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Effectiveness of Each Component. We conduct the ablation study on the CIFAR100-to-CIFAR100C scenario and evaluate the effectiveness of each component in Tab. 6. Ex1 denotes the performance using only self-training loss in the teacher-student framework, yielding a relatively high average error rate. From Ex2 to Ex4, we demonstrate the importance of our proposed uncertainty-aware", + "bbox": [ + 212, + 763, + 787, + 840 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "Reshaping the Online Data Buffering and Organizing Mechanism for CTTA", + "bbox": [ + 225, + 114, + 730, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 784, + 126 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "buffering mechanism. In contrast, the reservoir and FIFO buffering mechanisms do not have the ability to actively perform information aggregation, and thus they provide only a negligible improvement when combined with our pseudotarget replay loss. $Ex5$ demonstrates the effectiveness of our proposed class relation preservation constraint. $Ex6$ shows a significant overall improvement, suggesting that our proposed components are complementary.", + "bbox": [ + 212, + 146, + 787, + 237 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/088c7945566c5ef7bb306c0a1fc671551b35338ab370eaa56aae7bb2d079ce50.jpg", + "image_caption": [ + "Fig. 4: Hyperparameter sensitivity analysis." + ], + "image_footnote": [], + "bbox": [ + 243, + 237, + 545, + 330 + ], + "page_idx": 13 + }, + { + "type": "table", + "img_path": "images/5539fe919ec6c0661794d1d16554b5ef95414af6ff59c1f4c76cc9f658fa6c8e.jpg", + "table_caption": [ + "Table 6: Ablation experiments." + ], + "table_footnote": [], + "table_body": "
BufferingL_PCEL_CRPMean↓
Ex1---30.7
Ex2Reservoir✓-30.5
Ex3FIFO✓-30.6
Ex4Uncertainty✓-29.7
Ex5--✓30.0
Ex6Uncertainty✓✓29.0
", + "bbox": [ + 565, + 265, + 776, + 353 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Hyperparameter Sensitivity. Here we analyze the sensitivity of hyperparameters within the CIFAR100-to-CIFAR100C scenario. We consider three hyperparameters as illustrated in Fig. 4. Loss coefficient $\\lambda_{CRP}$ . The hyperparameter $\\lambda_{CRP}$ in Eq. (6) controls the strength of the proposed CRP constraint. We vary it within the range of $\\{0,50,150,200,250,300\\}$ . It can be observed that increasing $\\lambda_{CRP}$ over a reasonably broad range leads to a reduction in the average classification error rate of CTTA across 15 domains, indicating the effectiveness of CRP. While excessively large values of $\\lambda_{CRP}$ (e.g., 300) may compromise the model's adaptability. Threshold coefficient $\\alpha$ . The hyperparameter $\\alpha$ in Eq. (2) is utilized to adjust the threshold for screening low-uncertainty samples in the streaming data. It is evident that it is stable over specific ranges and achieves the lowest average error rate among all domains at a value of 0.1. Buffer capacity. Here we examine the capacity of the buffer, which is employed to dynamically aggregate high-value samples from the online unsupervised data stream. As demonstrated in Fig. 4, our method achieves optimal results with a buffer capacity of 200, imposing only a negligible storage overhead.", + "bbox": [ + 212, + 377, + 787, + 619 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 642, + 359, + 657 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In this paper, we revisit the key challenges in CTTA and address all of them simultaneously. To this end, we reshape the online data buffering and organizing mechanism, endowing our dynamic target buffer with the ability to aggregate reliable information in the unsupervised data stream. Through our designed class relation preservation constraint and a pseudo-target replay objective, the problem of catastrophic forgetting and error accumulation is significantly alleviated. Extensive experiments demonstrate the effectiveness of the proposed method. Regarding the adoption of uncertainty, our method employs the mainstream practice of sample entropy. In our framework, there is potential for investigating more diverse forms of uncertainty modeling, such as those based on energy functions. We leave this exploration to future work.", + "bbox": [ + 212, + 672, + 787, + 840 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "Z. Zhu et al.", + "bbox": [ + 271, + 114, + 357, + 127 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgement", + "text_level": 1, + "bbox": [ + 217, + 143, + 392, + 162 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "This work was funded in part by the National Natural Science Foundation of China (62076195, 62376070, 62206271, U20B2052), in part by the Fundamental Research Funds for the Central Universities (AUGA5710011522), and by the Pengcheng Laboratory Research Project No. PCL2023A08.", + "bbox": [ + 215, + 175, + 785, + 234 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 258, + 321, + 273 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Boudiaf, M., Mueller, R., Ben Ayed, I., Bertinetto, L.: Parameter-free online test-time adaptation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8344-8353 (2022)", + "2. Brahma, D., Rai, P.: A probabilistic framework for lifelong test-time adaptation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 3582-3591 (June 2023)", + "3. Burges, C., Ragno, R., Le, Q.: Learning to rank with nonsmooth cost functions. Advances in neural information processing systems 19 (2006)", + "4. Chakrabarty, G., Sreenivas, M., Biswas, S.: SANTA: Source anchoring network and target alignment for continual test time adaptation. Transactions on Machine Learning Research (2023), https://openreview.net/forum?id=V7guVYzvE4", + "5. Chen, D., Wang, D., Darrell, T., Ebrahimi, S.: Contrastive test-time adaptation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 295-305 (2022)", + "6. Chen, H., Wang, Y., Hu, Q.: Multi-granularity regularized re-balancing for class incremental learning. IEEE Transactions on Knowledge and Data Engineering 35(7), 7263-7277 (2022)", + "7. Choi, S., Yang, S., Choi, S., Yun, S.: Improving test-time adaptation via shift-agnostic weight regularization and nearest source prototypes. In: European Conference on Computer Vision. pp. 440-458. Springer (2022)", + "8. Cicek, S., Soatto, S.: Unsupervised domain adaptation via regularized conditional alignment. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 1416-1425 (2019)", + "9. Cordts, M., Omran, M., Ramos, S., Rehfeld, T., Enzweiler, M., Benenson, R., Franke, U., Roth, S., Schiele, B.: The cityscapes dataset for semantic urban scene understanding. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 3213-3223 (June 2016)", + "10. Croce, F., Andriushchenko, M., Sehwag, V., Debenedetti, E., Flammarion, N., Chiang, M., Mittal, P., Hein, M.: Robustbench: a standardized adversarial robustness benchmark. In: Thirty-fifth Conference on Neural Information Processing Systems Datasets and Benchmarks Track (Round 2) (2021)", + "11. Cui, S., Wang, S., Zhuo, J., Su, C., Huang, Q., Tian, Q.: Gradually vanishing bridge for adversarial domain adaptation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 12455-12464 (2020)", + "12. Döbler, M., Marsden, R.A., Yang, B.: Robust mean teacher for continual and gradual test-time adaptation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 7704-7714 (June 2023)", + "13. Dong, S., Hong, X., Tao, X., Chang, X., Wei, X., Gong, Y.: Few-shot class incremental learning via relation knowledge distillation. In: Proceedings of the AAAI Conference on Artificial Intelligence. pp. 1255-1263 (2021)" + ], + "bbox": [ + 218, + 287, + 784, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "Reshaping the Online Data Buffering and Organizing Mechanism for CTTA", + "bbox": [ + 225, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "14. Fan, Y., Wang, Y., Zhu, P., Hu, Q.: Dynamic sub-graph distillation for robust semi-supervised continual learning. In: Proceedings of the AAAI Conference on Artificial Intelligence. pp. 11927-11935 (2024)", + "15. Gan, Y., Bai, Y., Lou, Y., Ma, X., Zhang, R., Shi, N., Luo, L.: Decorate the newcomers: Visual domain prompt for continual test time adaptation. In: Proceedings of the AAAI Conference on Artificial Intelligence. pp. 7595-7603 (2023)", + "16. Ganin, Y., Lempitsky, V.: Unsupervised domain adaptation by backpropagation. In: International conference on machine learning. pp. 1180-1189. PMLR (2015)", + "17. Geirhos, R., Jacobsen, J.H., Michaelis, C., Zemel, R., Brendel, W., Bethge, M., Wichmann, F.A.: Shortcut learning in deep neural networks. Nature Machine Intelligence 2(11), 665-673 (2020)", + "18. Hendrycks, D., Dietterich, T.: Benchmarking neural network robustness to common corruptions and perturbations. In: International Conference on Learning Representations (2019), https://openreview.net/forum?id=HJz6tiCqYm", + "19. Hoffman, J., Tzeng, E., Park, T., Zhu, J.Y., Isola, P., Saenko, K., Efros, A., Darrell, T.: Cycada: Cycle-consistent adversarial domain adaptation. In: International conference on machine learning. pp. 1989-1998. Pmlr (2018)", + "20. Hwang, S., Lee, S., Kim, S., Ok, J., Kwak, S.: Combating label distribution shift for active domain adaptation. In: European Conference on Computer Vision. pp. 549-566. Springer (2022)", + "21. Koh, P.W., Sagawa, S., Marklund, H., Xie, S.M., Zhang, M., Balsubramani, A., Hu, W., Yasunaga, M., Phillips, R.L., Gao, I., et al.: Wilds: A benchmark of inthe-wild distribution shifts. In: International Conference on Machine Learning. pp. 5637-5664. PMLR (2021)", + "22. Lee, K., Kim, S., Kwak, S.: Cross-domain ensemble distillation for domain generalization. In: European Conference on Computer Vision. pp. 1-20. Springer (2022)", + "23. Li, H., Pan, S.J., Wang, S., Kot, A.C.: Domain generalization with adversarial feature learning. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 5400-5409 (2018)", + "24. Li, S., Xie, B., Lin, Q., Liu, C.H., Huang, G., Wang, G.: Generalized domain conditioned adaptation network. IEEE Transactions on Pattern Analysis and Machine Intelligence 44(8), 4093-4109 (2021)", + "25. Li, S., Xie, M., Gong, K., Liu, C.H., Wang, Y., Li, W.: Transferable semantic augmentation for domain adaptation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11516-11525 (2021)", + "26. Li, Y., Wang, N., Shi, J., Liu, J., Hou, X.: Revisiting batch normalization for practical domain adaptation. arXiv preprint arXiv:1603.04779 (2016)", + "27. Liang, J., Hu, D., Feng, J.: Do we really need to access the source data? source hypothesis transfer for unsupervised domain adaptation. In: International conference on machine learning. pp. 6028-6039. PMLR (2020)", + "28. Lin, H., Zhang, Y., Qiu, Z., Niu, S., Gan, C., Liu, Y., Tan, M.: Prototype-guided continual adaptation for class-incremental unsupervised domain adaptation. In: European Conference on Computer Vision. pp. 351-368. Springer (2022)", + "29. Liu, H., Long, M., Wang, J., Jordan, M.: Transferable adversarial training: A general approach to adapting deep classifiers. In: International Conference on Machine Learning. pp. 4013-4022. PMLR (2019)", + "30. Liu, J., Yang, S., Jia, P., Zhang, R., Lu, M., Guo, Y., Xue, W., Zhang, S.: ViDA: Homeostatic visual domain adapter for continual test time adaptation. In: The Twelfth International Conference on Learning Representations (2024), https://openreview.net/forum?id=sJ88Wg5Bp5" + ], + "bbox": [ + 217, + 146, + 784, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "Z. Zhu et al.", + "bbox": [ + 271, + 114, + 354, + 126 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "31. Liu, Y., Kothari, P., Van Delft, B., Bellot-Gurlet, B., Mordan, T., Alahi, A.: Ttt++: When does self-supervised test-time training fail or thrive? Advances in Neural Information Processing Systems 34, 21808-21820 (2021)", + "32. Long, M., Cao, Y., Wang, J., Jordan, M.: Learning transferable features with deep adaptation networks. In: International conference on machine learning. pp. 97-105. PMLR (2015)", + "33. Long, M., Zhu, H., Wang, J., Jordan, M.I.: Deep transfer learning with joint adaptation networks. In: International conference on machine learning. pp. 2208-2217. PMLR (2017)", + "34. Nguyen, A.T., Nguyen-Tang, T., Lim, S.N., Torr, P.H.: Tipi: Test time adaptation with transformation invariance. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 24162-24171 (2023)", + "35. Niloy, F.F., Ahmed, S.M., Raychaudhuri, D.S., Oymak, S., Roy-Chowdhury, A.K.: Effective restoration of source knowledge in continual test time adaptation. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 2091-2100 (2024)", + "36. Niu, S., Wu, J., Zhang, Y., Chen, Y., Zheng, S., Zhao, P., Tan, M.: Efficient test-time model adaptation without forgetting. In: International conference on machine learning. pp. 16888-16905. PMLR (2022)", + "37. Niu, S., Wu, J., Zhang, Y., Wen, Z., Chen, Y., Zhao, P., Tan, M.: Towards stable test-time adaptation in dynamic wild world. In: The Eleventh International Conference on Learning Representations (2023), https://openreview.net/forum?id=g2YraF75Tj", + "38. Patel, V.M., Gopalan, R., Li, R., Chellappa, R.: Visual domain adaptation: A survey of recent advances. IEEE signal processing magazine 32(3), 53-69 (2015)", + "39. Peng, X., Bai, Q., Xia, X., Huang, Z., Saenko, K., Wang, B.: Moment matching for multi-source domain adaptation. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 1406-1415 (2019)", + "40. Press, O., Schneider, S., Kümmerer, M., Bethge, M.: Rdumb: A simple approach that questions our progress in continual test-time adaptation. Advances in Neural Information Processing Systems 36 (2023)", + "41. Sakaridis, C., Dai, D., Van Gool, L.: Acdc: The adverse conditions dataset with correspondences for semantic driving scene understanding. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 10765-10775 (2021)", + "42. Schneider, S., Rusak, E., Eck, L., Bringmann, O., Brendel, W., Bethge, M.: Improving robustness against common corruptions by covariate shift adaptation. Advances in neural information processing systems 33, 11539-11551 (2020)", + "43. Sojka, D., Cygert, S., Twardowski, B., Trzciński, T.: Ar-tta: A simple method for real-world continual test-time adaptation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) Workshops. pp. 3491-3495 (October 2023)", + "44. Song, J., Lee, J., Kweon, I.S., Choi, S.: Ecotta: Memory-efficient continual test-time adaptation via self-distilled regularization. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11920-11929 (2023)", + "45. Tan, M., Chen, G., Wu, J., Zhang, Y., Chen, Y., Zhao, P., Niu, S.: Uncertainty-calibrated test-time model adaptation without forgetting. arXiv preprint arXiv:2403.11491 (2024)", + "46. Tao, X., Chang, X., Hong, X., Wei, X., Gong, Y.: Topology-preserving class-incremental learning. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XIX 16. pp. 254-270. Springer (2020)" + ], + "bbox": [ + 215, + 146, + 784, + 839 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "Reshaping the Online Data Buffering and Organizing Mechanism for CTTA", + "bbox": [ + 225, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "47. Tao, X., Hong, X., Chang, X., Dong, S., Wei, X., Gong, Y.: Few-shot class incremental learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (June 2020)", + "48. Tao, X., Hong, X., Chang, X., Gong, Y.: Bi-objective continual learning: Learning 'new' while consolidating 'known'. In: Proceedings of the AAAI Conference on Artificial Intelligence. pp. 5989-5996 (2020)", + "49. Tsai, Y.H., Hung, W.C., Schulter, S., Sohn, K., Yang, M.H., Chandraker, M.: Learning to adapt structured output space for semantic segmentation. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 7472-7481 (2018)", + "50. Wang, D., Shelhamer, E., Liu, S., Olshausen, B., Darrell, T.: Tent: Fully test-time adaptation by entropy minimization. In: International Conference on Learning Representations (2021), https://openreview.net/forum?id=uX13bZLkr3c", + "51. Wang, Q., Fink, O., Van Gool, L., Dai, D.: Continual test-time domain adaptation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 7201-7211 (June 2022)", + "52. Wang, S., Zhang, D., Yan, Z., Zhang, J., Li, R.: Feature alignment and uniformity for test time adaptation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 20050-20060 (2023)", + "53. Wang, Y., Ma, Z., Huang, Z., Wang, Y., Su, Z., Hong, X.: Isolation and impartial aggregation: A paradigm of incremental learning without interference. In: Proceedings of the AAAI Conference on Artificial Intelligence. pp. 10209-10217 (2023)", + "54. Wang, Y., Hong, J., Cheraghian, A., Rahman, S., Ahmedt-Aristizabal, D., Petersson, L., Harandi, M.: Continual test-time domain adaptation via dynamic sample selection. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV). pp. 1701-1710 (January 2024)", + "55. Wei, Y., Ye, J., Huang, Z., Zhang, J., Shan, H.: Online prototype learning for online continual learning. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 18764-18774 (October 2023)", + "56. Xie, E., Wang, W., Yu, Z., Anandkumar, A., Alvarez, J.M., Luo, P.: Segformer: Simple and efficient design for semantic segmentation with transformers. In: Beygelzimer, A., Dauphin, Y., Liang, P., Vaughan, J.W. (eds.) Advances in Neural Information Processing Systems (2021), https://openreview.net/forum?id=OG18MI5TRL", + "57. Xie, S., Girshick, R., Dollar, P., Tu, Z., He, K.: Aggregated residual transformations for deep neural networks. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1492-1500 (2017)", + "58. Yang, X., Gu, Y., Wei, K., Deng, C.: Exploring safety supervision for continual test-time domain adaptation. In: Proceedings of the Thirty-Second International Joint Conference on Artificial Intelligence, IJCAI-23. pp. 1649-1657. International Joint Conferences on Artificial Intelligence Organization (8 2023). https://doi.org/10.24963/ijcai.2023/183", + "59. Yang, Y., Soatto, S.: Fda: Fourier domain adaptation for semantic segmentation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 4085-4095 (2020)", + "60. Yao, X., Wang, Y., Zhu, P., Lin, W., Li, J., Li, W., Hu, Q.: Socialized learning: Making each other better through multi-agent collaboration. In: Forty-first International Conference on Machine Learning (2024)", + "61. Yuan, L., Xie, B., Li, S.: Robust test-time adaptation in dynamic scenarios. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15922-15932 (2023)" + ], + "bbox": [ + 215, + 146, + 784, + 840 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "Z. Zhu et al.", + "bbox": [ + 271, + 114, + 354, + 126 + ], + "page_idx": 17 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "62. Zagoruyko, S., Komodakis, N.: Wide residual networks. In: British Machine Vision Conference 2016. British Machine Vision Association (2016)", + "63. Zhang, J., Qi, L., Shi, Y., Gao, Y.: Mvdg: A unified multi-view framework for domain generalization. In: European Conference on Computer Vision. pp. 161-177. Springer (2022)", + "64. Zhang, Y., Wang, Z., He, W.: Class relationship embedded learning for source-free unsupervised domain adaptation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7619-7629 (2023)", + "65. Zhang, Y., Wang, Z., Li, J., Zhuang, J., Lin, Z.: Towards effective instance discrimination contrastive loss for unsupervised domain adaptation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 11388-11399 (2023)", + "66. Zhou, K., Liu, Z., Qiao, Y., Xiang, T., Loy, C.C.: Domain generalization: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence (2022)" + ], + "bbox": [ + 215, + 146, + 785, + 340 + ], + "page_idx": 18 + }, + { + "type": "header", + "text": "Reshaping the Online Data Buffering and Organizing Mechanism for CTTA", + "bbox": [ + 225, + 114, + 730, + 128 + ], + "page_idx": 18 + }, + { + "type": "page_number", + "text": "19", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 18 + } +] \ No newline at end of file diff --git a/2024/Reshaping the Online Data Buffering and Organizing Mechanism for Continual Test-Time Adaptation/69885bc8-fe2a-47b1-a255-a982ec605708_model.json b/2024/Reshaping the Online Data Buffering and Organizing Mechanism for Continual Test-Time Adaptation/69885bc8-fe2a-47b1-a255-a982ec605708_model.json new file mode 100644 index 0000000000000000000000000000000000000000..1aaf289ed2454babef0a133fe7295b1a4ccde165 --- /dev/null +++ b/2024/Reshaping the Online Data Buffering and Organizing Mechanism for Continual Test-Time Adaptation/69885bc8-fe2a-47b1-a255-a982ec605708_model.json @@ -0,0 +1,2438 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.23, + 0.141, + 0.776, + 0.21 + ], + "angle": 0, + "content": "Reshaping the Online Data Buffering and Organizing Mechanism for Continual Test-Time Adaptation" + }, + { + "type": "text", + "bbox": [ + 0.221, + 0.234, + 0.782, + 0.266 + ], + "angle": 0, + "content": "Zhilin Zhu\\(^{1,2}\\), Xiaopeng Hong\\(^{1,2\\star}\\), Zhiheng Ma\\(^{3,4,5}\\), Weijun Zhuang\\(^{1,2}\\), Yaohui Ma\\(^{1,5}\\), Yong Dai\\(^{2}\\), and Yaowei Wang\\(^{2,1}\\)" + }, + { + "type": "text", + "bbox": [ + 0.391, + 0.277, + 0.612, + 0.291 + ], + "angle": 0, + "content": "1 Harbin Institute of Technology" + }, + { + "type": "text", + "bbox": [ + 0.419, + 0.292, + 0.584, + 0.306 + ], + "angle": 0, + "content": "2 Pengcheng Laboratory" + }, + { + "type": "text", + "bbox": [ + 0.344, + 0.306, + 0.66, + 0.319 + ], + "angle": 0, + "content": "3 Shenzhen University of Advanced Technology" + }, + { + "type": "text", + "bbox": [ + 0.258, + 0.32, + 0.746, + 0.333 + ], + "angle": 0, + "content": "4 Guangdong Provincial Key Laboratory of Computility Microelectronics" + }, + { + "type": "text", + "bbox": [ + 0.257, + 0.334, + 0.767, + 0.375 + ], + "angle": 0, + "content": "Shenzhen Institute of Advanced Technology, Chinese Academy of Science {zhuzhilin, weijunzhuang}@stu.hit.edu.cn hongxiaopeng@ieee.org ma@siat.ac.cn yh.ma@siat.ac.cn chd-dy@foxmail.com wangyw@pcl.ac.cn" + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.401, + 0.741, + 0.596 + ], + "angle": 0, + "content": "Abstract. Continual Test-Time Adaptation (CTTA) involves adapting a pre-trained source model to continually changing unsupervised target domains. In this paper, we systematically analyze the challenges of this task: online environment, unsupervised nature, and the risks of error accumulation and catastrophic forgetting under continual domain shifts. To address these challenges, we reshape the online data buffering and organizing mechanism for CTTA. We propose an uncertainty-aware buffering approach to identify and aggregate significant samples with high certainty from the unsupervised, single-pass data stream. Based on this, we propose a graph-based class relation preservation constraint to overcome catastrophic forgetting. Furthermore, a pseudo-target replay objective is used to mitigate error accumulation. Extensive experiments demonstrate the superiority of our method in both segmentation and classification CTTA tasks. Code is available at https://github.com/z1358/OBAO." + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.607, + 0.741, + 0.636 + ], + "angle": 0, + "content": "Keywords: Continual test-time adaptation \\(\\cdot\\) Unsupervised learning \\(\\cdot\\) Continual learning \\(\\cdot\\) Catastrophic forgetting" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.656, + 0.377, + 0.672 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.683, + 0.788, + 0.82 + ], + "angle": 0, + "content": "Deep learning models have exhibited remarkable advancements in scenarios characterized by a consistent data distribution between training and test environments. However, the deployment of machine perception systems in the real world frequently exposes them to non-stationary and constantly changing environments [14, 21, 47, 53, 60]. When such phenomena occur (e.g., unexpected domain shifts over time), these models are prone to severe performance degradation [18, 37, 51]. Consequently, the development of techniques for continual adaptation during the unsupervised deployment phase is essential for enhancing the reliability of machine perception systems." + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.825, + 0.383, + 0.84 + ], + "angle": 0, + "content": "* Corresponding author" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.358, + 0.128 + ], + "angle": 0, + "content": "Z. Zhu et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.788, + 0.188 + ], + "angle": 0, + "content": "Table 1: Comparison of different adaptation settings. We differentiate along four dimensions: the available form of the source domain data and target domain data, the distribution and access mode of the target domain data stream during adaptation." + }, + { + "type": "table", + "bbox": [ + 0.232, + 0.199, + 0.778, + 0.31 + ], + "angle": 0, + "content": "
SettingDataTarget domain data stream
SourceTargetDistributionAccess mode
Fine-tuning-xT,yTstationaryoffline (multi-epoch)
Domain GeneralizationxS,yS---
Domain AdaptationxS,ySxTstationaryoffline (multi-epoch)
Test-Time TrainingxS,ySxTstationaryonline (single-pass)
Test-Time Adaptation-xTstationaryoffline or online
Continual Test-Time Adaptation-x1T→···→xnTcontinually changingonline (single-pass)
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.34, + 0.788, + 0.506 + ], + "angle": 0, + "content": "With sufficient source labeled data accessible, the fields of domain generalization [22, 23, 63, 66] and domain adaptation [20, 24, 25, 28] have witnessed considerable success. However, a particularly challenging scenario arises when no source data is available, and unsupervised adaptation must happen at test time when unlabeled test inputs arrive. This task is known as test-time adaptation (TTA) [5, 7, 45]. Given the rarity of encountering only a singular domain shift in practical applications, the community has recently been increasingly focusing on continual test-time adaptation (CTTA) [2, 12, 30, 51, 58], i.e., evaluating the ability of methods to adapt to continual domain shifts. We summarize the differences between existing adaptation settings in Tab. 1. Obviously, the considered CTTA setting is the most practical and challenging." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.507, + 0.788, + 0.704 + ], + "angle": 0, + "content": "Within this setting, we expose the key challenges as follows: (1) Online environment. This denotes a scenario where data is processed in a single pass, necessitating immediate predictions from the model. Due to the characteristics of the online environment, it is difficult for the model to obtain a sufficient number of informative samples, thereby impeding the efficacy of adaptation; (2) Unsupervised nature. Throughout adaptation, ground-truth labels for input data are unavailable. Moreover, due to privacy concerns or legal restrictions, the source domain data may no longer be available and we only have access to the source pre-trained model; (3) Error accumulation and catastrophic forgetting. The dynamic distribution shift when adapting to continually changing target domains can render pseudo-labels unreliable and impede knowledge retention about the source domain, leading to error accumulation [51, 54] and triggering catastrophic forgetting [6, 30, 58]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.705, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Addressing the first challenge, an intuitive solution is to employ a buffer to store high-value samples. However, existing methods that utilize such a buffer store source domain samples with ground-truth labels [12, 43], which is difficult to satisfy in real-world scenarios and conflicts with the source-free constraint in the second challenge. Therefore, the issue of effectively buffering and organizing high-value samples in an unsupervised single-pass data stream has not been fully explored. Current CTTA methods primarily address the third challenge, focusing mainly on alignments in either the parameter space [2, 35, 40, 51] or the feature space [4, 12] of the source model to mitigate knowledge forgetting." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.227, + 0.115, + 0.731, + 0.129 + ], + "angle": 0, + "content": "Reshaping the Online Data Buffering and Organizing Mechanism for CTTA" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "image", + "bbox": [ + 0.241, + 0.147, + 0.391, + 0.212 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.267, + 0.237, + 0.366, + 0.249 + ], + "angle": 0, + "content": "(a) W/o a buffer" + }, + { + "type": "image", + "bbox": [ + 0.425, + 0.147, + 0.578, + 0.228 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.434, + 0.238, + 0.57, + 0.249 + ], + "angle": 0, + "content": "(b) Static source buffer" + }, + { + "type": "image", + "bbox": [ + 0.604, + 0.147, + 0.772, + 0.236 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.601, + 0.238, + 0.776, + 0.249 + ], + "angle": 0, + "content": "(c) Our dynamic target buffer" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.261, + 0.785, + 0.331 + ], + "angle": 0, + "content": "Fig. 1: Comparison of different buffering mechanisms. (a) Without a buffer, only unreliable signals from the current unsupervised input data are available. (b) Some methods [12, 43] use a static source buffer that stores source samples with ground-truth labels to provide supervised signals. (c) We work on a dynamic target buffer, which can provide reliable signals for the adaptation process." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.366, + 0.788, + 0.64 + ], + "angle": 0, + "content": "In this paper, we endeavor to address these three challenges simultaneously. To this end, we reshape the online data buffering and organizing mechanism for CTTA. As shown in Fig. 1, unlike previous methods which directly sample from the source domain [12, 43], we are targeted at source-free, unsupervised, and online buffering. Through a specially designed uncertainty-aware buffering approach, we dynamically evaluate and aggregate reliable information in the unsupervised, single-pass data stream (Challenges 1 and 2). Based on this, we efficiently organize these samples to assist the model's unsupervised adaptation process. Utilizing pseudo-labels generated from these low-uncertainty samples, we consider them as anchors of classes and organize them into a class relation graph (CRG). To combat catastrophic forgetting under continually changing domains, we introduce a novel graph-based class relation preservation constraint to maintain the intrinsic class relations obtained on the source domain (Challenge 3). Specifically, we penalize topological changes within the CRG to mitigate knowledge forgetting, while allowing reasonable vertex movements to enhance adaptability. Additionally, these samples are incorporated into a pseudo-target replay objective to reduce the risk of error accumulation from misclassified samples (Challenge 3). In summary, our main contributions are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.665, + 0.784, + 0.695 + ], + "angle": 0, + "content": "- To address the three revealed challenges simultaneously, we reshape the online data buffering and organizing mechanism for CTTA." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.699, + 0.785, + 0.728 + ], + "angle": 0, + "content": "- To frame the buffering mechanism, we design an uncertainty-aware approach to identify and aggregate reliable data from unsupervised online data stream." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.732, + 0.784, + 0.775 + ], + "angle": 0, + "content": "- Through a well-designed graph-based class relation preservation constraint and a pseudo-target replay objective, we effectively mitigate catastrophic forgetting and error accumulation in CTTA." + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.78, + 0.784, + 0.839 + ], + "angle": 0, + "content": "- We extensively evaluate our proposed method and the results on four challenging datasets show that the method clearly improves the CTTA performance on both segmentation and classification benchmarks, indicating its effectiveness and universality." + }, + { + "type": "list", + "bbox": [ + 0.225, + 0.665, + 0.785, + 0.839 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.232, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.358, + 0.128 + ], + "angle": 0, + "content": "Z. Zhu et al." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.145, + 0.389, + 0.161 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.182, + 0.427, + 0.198 + ], + "angle": 0, + "content": "2.1 Domain Adaptation" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.212, + 0.79, + 0.333 + ], + "angle": 0, + "content": "Domain adaptation (DA) [38,65] aims to generalize a model from a well-annotated source domain to an unlabeled target domain following the transductive learning principle. The underlying assumption in DA is that both labeled source domain and unlabeled target domain data are concurrently accessible. Consequently, the prevalent methods in DA encompass the utilization of domain adversarial training [8,11,16,19,29,49] or the minimization of domain discrepancy [32,33,39]. These methods work towards harmonizing the feature distribution [16,32,49] or the input space [19,59] between the two domains." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.36, + 0.531, + 0.375 + ], + "angle": 0, + "content": "2.2 Continual Test-Time Adaptation" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.39, + 0.788, + 0.571 + ], + "angle": 0, + "content": "Test-Time Adaptation (TTA). TTA focuses on a more challenging setting after model deployment where only a pre-trained source model and unlabeled target data are available. Some approaches [5, 27] focus on the offline setting, where multiple epochs of access to the entire set of test data are provided for adaptation, while others [1, 7, 31, 52] focus on the online setting with a single-pass data stream. To name a few, TENT [50] adapts the model by using entropy minimization loss to update a few trainable parameters in batch normalization layers. This approach has also inspired subsequent works such as EATA [36] and RoTTA [61] to investigate the robustness of normalized layers. Due to the domain gap, TSD [52] promotes feature alignment and uniformity from the perspective of feature revision, while TIPI [34] proposes invariance regularizers to find the transformation space that can simulate the domain shift." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.598, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Continual Test-Time Adaptation (CTTA). In recent years, CTTA has received increasing research attention [15,44,51,58] due to its practicality. CoTTA [51] stands out as the first method specifically tailored to address this task. It utilizes weight-averaged and augmentation-averaged predictions to reduce error accumulation, while stochastically restoring the weights of a small fraction of neurons to the source model during each iteration to mitigate catastrophic forgetting. Building upon this, PETAL [2] further introduces a data-driven parameter recovery technique to align with the source model in the parameter space, while RMT [12] and SANTA [4] seek alignment with the source model in the feature space by optimizing an additional contrastive learning objective. Additionally, DSS [54] considers sample security to mitigate the risk of using incorrect information. Some approaches [15, 30, 44] also prevent forgetting by freezing a majority or all of the model parameters, but they require optimization along with the model during training using the source domain data, thereby relaxing the unsupervised challenge and not being truly source-free. Under our proposal, we simultaneously address three challenges in CTTA." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.227, + 0.114, + 0.731, + 0.129 + ], + "angle": 0, + "content": "Reshaping the Online Data Buffering and Organizing Mechanism for CTTA" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "image", + "bbox": [ + 0.258, + 0.148, + 0.744, + 0.359 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.372, + 0.788, + 0.456 + ], + "angle": 0, + "content": "Fig. 2: The framework of our proposed method. At time step \\( t \\), the incoming data and sampled buffer data are used for adaptation. The proposed class relation preservation loss \\( \\mathcal{L}_{CRP} \\) ensures topological consistency between the Class Relation Graph (CRG) \\( G_{t} \\) estimated at the current time step \\( t \\) and \\( \\hat{G} \\) from the source domain, thereby effectively preserving intrinsic semantic information. Concurrently, we integrate a pseudo-target replay loss \\( \\mathcal{L}_{PCE} \\) to mitigate error accumulation." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.483, + 0.381, + 0.501 + ], + "angle": 0, + "content": "3 Methodology" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.514, + 0.573, + 0.529 + ], + "angle": 0, + "content": "3.1 Preliminaries and Overall Framework" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.538, + 0.788, + 0.689 + ], + "angle": 0, + "content": "In continual test-time adaptation (CTTA), we are given a source model \\( f_{\\theta_0} \\) pretrained from the labeled source domain \\( (\\mathcal{X}^S, \\mathcal{Y}^S) \\) and a series of sequentially changing unlabeled target domains such as \\( [\\mathcal{X}_1^T, \\mathcal{X}_2^T, \\dots, \\mathcal{X}_n^T] \\). Each target domain shares the same label space as the source domain (for close-set adaptation). Note that the model has no information about when the domain changes happen during the adaptation process. The goal of CTTA is to adapt the model to a continually changing unsupervised online data stream of target domains without source data. To be more specific, at inference time step \\( t \\), the unlabeled target test data \\( x(t) \\) is given to the model \\( f_{\\theta_t} \\), and the model needs to make predictions and adapt itself accordingly \\( (f_{\\theta_t} \\rightarrow f_{\\theta_{t+1}}) \\) for the upcoming test inputs \\( x(t+1) \\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.788, + 0.841 + ], + "angle": 0, + "content": "We adopt a teacher-student framework following previous works [2,12,51] as illustrated in Fig. 2. Leveraging the proposed uncertainty-aware buffering mechanism, our method dynamically evaluates and buffers significant high-certainty samples in the unsupervised online data stream, providing reliable information for the model's adaptation process. Building upon this, we further introduce a novel class relation preservation constraint \\((\\mathcal{L}_{CRP})\\) for preserving domain-invariant semantic information in dynamic environments, effectively mitigating the catastrophic forgetting problem. Additionally, we address error accumulation through a pseudo-target replay objective \\((\\mathcal{L}_{PCE})\\) to minimize the impact of miscalibration. A self-training loss \\((\\mathcal{L}_{ST})\\) is used to stabilize the adaptation" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.358, + 0.128 + ], + "angle": 0, + "content": "Z. Zhu et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.222 + ], + "angle": 0, + "content": "process of the student model (adapted model). In the adopted framework, both the teacher model \\( f_{\\theta'} \\) and the student model \\( f_{\\theta} \\) are initialized with the source model \\( f_{\\theta_0} \\). Subsequently, the teacher model is continuously updated by the exponential moving average of the student model. We elaborate our method in the following subsections." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.244, + 0.498, + 0.26 + ], + "angle": 0, + "content": "3.2 Uncertainty-aware Buffering" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.268, + 0.789, + 0.373 + ], + "angle": 0, + "content": "In the unsupervised online data stream within CTTA, the model necessitates adaptation utilizing the data at hand, typically characterized by a scarcity of information. While certain methods [12, 43] endeavor to incorporate additional supervised information by formulating a static buffer from the source domain, fulfilling this requirement proves challenging within practical application scenarios. Conversely, our proposition involves sampling and buffering throughout the unsupervised adaptation process by a designed uncertainty-aware approach." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.374, + 0.79, + 0.464 + ], + "angle": 0, + "content": "At the outset, the dynamic buffer is initialized as empty. As the adaptation progresses, the buffer undergoes continual updates to reflect the evolving nature of the data stream encountered. Specifically, at any given inference time step \\( t \\), the model calculates the uncertainty associated with the current batch of samples. The quantification of sample uncertainty manifests as the entropy of the sample, denoted as:" + }, + { + "type": "equation", + "bbox": [ + 0.371, + 0.474, + 0.787, + 0.492 + ], + "angle": 0, + "content": "\\[\nH (x (t)) = - f _ {\\theta_ {t}} (x (t)) ^ {\\top} (\\log f _ {\\theta_ {t}} (x (t))), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.501, + 0.788, + 0.562 + ], + "angle": 0, + "content": "where \\( f_{\\theta_t}(x(t)) \\) corresponds the softmax predictions of the student model for the current input data. Samples exhibiting an entropy value less than a predefined threshold \\( H_0 \\) are deemed to possess a high degree of certainty and are thus considered candidates for inclusion in the dynamic buffer. \\( H_0 \\) is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.458, + 0.574, + 0.786, + 0.589 + ], + "angle": 0, + "content": "\\[\nH _ {0} = \\alpha \\ln C, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.599, + 0.788, + 0.734 + ], + "angle": 0, + "content": "where \\(\\alpha\\) serves as a threshold coefficient and \\(C\\) denotes the total number of classes in the dataset. The process of storing samples into the buffer, or updating existing entries, is contingent upon the buffer's capacity. High-certainty samples identified from the present batch, along with their derived pseudo-labels from the teacher model, are added to the buffer if space permits. Conversely, in scenarios where the buffer is at capacity, these new samples replace the ones exhibiting the highest uncertainty to maintain an optimal blend of timeliness and reliability. Upon completion of these updates, the resulting buffer \\(\\mathcal{M}_t\\) stands ready to facilitate the model's adaptation in subsequent iterations." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Notably, our approach to buffering is entirely devoid of reliance on source data, underscoring its source-free nature. This strategy enables the model to discern and prioritize data based on their uncertainty. By dynamically buffering high-certainty samples within the data stream, our approach not only circumvents the limitations posed by reliance on static source buffers but also enriches the adaptation process with valuable, reliable data. The intuitive idea is to utilize CE loss to replay samples in the buffer, thereby emphasizing reliable data" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.227, + 0.115, + 0.731, + 0.129 + ], + "angle": 0, + "content": "Reshaping the Online Data Buffering and Organizing Mechanism for CTTA" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.208 + ], + "angle": 0, + "content": "and effectively suppressing the risk and impact of miscalibration due to error accumulation. In each test-time adaptation iteration, we randomly sample examples and labels \\((\\tilde{x}(t), \\tilde{y}(t))\\) from the buffer. The number of sampled examples is equal to the batch size and the corresponding pseudo-target replay loss is:" + }, + { + "type": "equation", + "bbox": [ + 0.411, + 0.221, + 0.786, + 0.261 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {P C E} = - \\sum_ {c = 1} ^ {C} \\tilde {y} _ {c} (t) \\log \\tilde {p} _ {c}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.273, + 0.785, + 0.349 + ], + "angle": 0, + "content": "where \\(\\tilde{p}\\) is the softmax prediction of the student model for \\(\\tilde{x}(t)\\). For simplicity, we omit the summation of all samples in the batch. Moreover, these sampled examples also serve as anchors and are organized in a class relation graph. On this basis, we introduce a novel class relation preservation constraint in Sec. 3.3 to preserve domain-invariant knowledge and overcome forgetting." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.373, + 0.584, + 0.387 + ], + "angle": 0, + "content": "3.3 Class Relation Preservation Constraint" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.4, + 0.785, + 0.52 + ], + "angle": 0, + "content": "In online scenarios where domains are constantly changing, neural networks tend to \"take shortcuts\" and focus on simplistic features [17, 55] from the single-pass data stream. This phenomenon results in the model's forgetting of intrinsic semantic information. In this work, we propose to overcome catastrophic forgetting by preserving the domain-invariant class relation, which is consistent across domains regardless of distribution discrepancy [64]. Therefore, it is reasonable to utilize the intrinsic class relation obtained on the source domain with large labeled samples to constrain the model adaptation process on target domains." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.522, + 0.786, + 0.688 + ], + "angle": 0, + "content": "We describe here how to obtain the defined class relation graph (CRG). For computational stability, we normalize the feature space and adopt the cosine similarity metric. Let \\(\\bar{\\cdot}\\) denotes the normalization operation, where \\(\\bar{f} = f / \\| f\\|\\). Given the normalized feature space \\(\\bar{\\mathcal{F}}\\), the CRG \\(G\\) is defined as \\(G = \\langle V,E\\rangle\\), where \\(V = \\{\\bar{v}_1,\\dots ,\\bar{v}_C|\\bar{v}_i\\in \\bar{\\mathcal{F}}\\}\\) is the set of \\(C\\) vertices representative of \\(\\bar{\\mathcal{F}}\\), and \\(E\\) is the edge set describing the neighborhood relations of vertices in \\(V\\). Each edge \\(e_{ij}\\) is assigned with a weight \\(s_{ij}\\), which is the similarity between \\(\\bar{v}_i\\) and \\(\\bar{v}_j\\): \\(s_{ij} = \\bar{v}_i^\\top \\bar{v}_j\\). Each vertex \\(\\bar{v}_i\\) is the centroid vector of class \\(i\\). The intrinsic CRG \\(\\hat{G}\\) is obtained by modeling the source-domain class relation. We demonstrate in experiments that the construction of our CRG \\(\\hat{G}\\) can be derived from different data conditions, including class prototypes and classifier weights." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.787, + 0.841 + ], + "angle": 0, + "content": "At time step \\( t \\), given CRG \\( G_{t} = \\langle V_{t}, E_{t} \\rangle \\), when catastrophic forgetting occurs, \\( G_{t} \\) is distorted as the semantic relations among categories are confused. To mitigate forgetting, the model should be adapted to strive towards preserving the topology in \\( G_{t} \\) the same as that in \\( \\hat{G} \\). This is achieved by restricting the neighboring relations of vertices described by the weights of edges (i.e., similarity between classes) in \\( E_{t} \\). One way to do this is to keep the ranking of the edge weights constant during the adaptation process. However optimizing a non-smooth global ranking is difficult and inefficient [3,46]. An alternative and more feasible approach involves assessing the extent of change in neighboring relations [13,48] by evaluating the correlation between the initial edge weights" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.358, + 0.128 + ], + "angle": 0, + "content": "Z. Zhu et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.207 + ], + "angle": 0, + "content": "and their current observation. A lower correlation indicates a higher probability that the rank of an edge has changed during adaptation, and this should be penalized. On this basis, we define the Class Relation Preservation constraint (CRP) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.389, + 0.219, + 0.786, + 0.268 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {C R P} = - \\frac {\\sum_ {i , j} ^ {C} s _ {i j} \\tilde {s} _ {i j}}{\\sqrt {\\sum_ {i , j} ^ {C} s _ {i j} ^ {2}} \\sqrt {\\sum_ {i , j} \\tilde {s} _ {i j} ^ {2}}}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.273, + 0.786, + 0.307 + ], + "angle": 0, + "content": "where \\( S = \\{s_{ij}|1\\leq i,j\\leq C\\} \\) and \\( \\tilde{S} = \\{\\tilde{s}_{ij}|1\\leq i,j\\leq C\\} \\) are the sets of the initial and observation values of the edge weights in \\( \\hat{E} \\) and \\( E_{t} \\), respectively." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.307, + 0.787, + 0.427 + ], + "angle": 0, + "content": "To achieve end-to-end optimization, the ideal situation is using source-domain images with ground-truth labels to obtain an estimate of the current CRG \\( G_{t} \\) under time step \\( t \\). That is, extracting features from these images by \\( f_{\\theta_t} \\), which in turn yields the class prototypes and ultimately the class relation. However, it is often impractical to store a certain number of source-domain samples in the source-free CTTA setting. In this work, we leverage reliable data from our proposed uncertainty-aware buffering mechanism to accurately estimate CRG \\( G_{t} \\) and achieve end-to-end optimization in an unsupervised test data stream." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.428, + 0.789, + 0.685 + ], + "angle": 0, + "content": "From the perspective of viewing \\( S \\) as a vector, the proposed CRP term constrains the consistency of the direction between two vectors during the adaptation process. Opting for the origin moment instead of the central moment allows the CRP term to prioritize direction consistency between vectors. This makes it more robust to scale variations or magnitude disparities in edge weights, which may arise from shifts in data distribution. Some previous methods [4,12] directly penalize the movement of target domain samples in the feature space relative to the source prototypes. This can be broadly interpreted as penalizing the movement of corresponding elements between \\( \\hat{V} \\) and \\( V_{t} \\) in our defined CRG. However, in addition to suffering from classification errors when assigning source prototypes, constraints on features can also impede the effectiveness of adaptation in target domains. Instead of penalizing the movement of vertices in \\( G \\) in the feature space, the proposed CRP term penalizes the change of topological relations between vertices, while allowing reasonable movement of vertices to enhance adaptation. To summarize, by preserving the domain-invariant class relation during the model adaptation process, the proposed CRP term effectively overcomes catastrophic forgetting caused by semantic confusion." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.707, + 0.457, + 0.723 + ], + "angle": 0, + "content": "3.4 Optimization Objective" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.731, + 0.788, + 0.792 + ], + "angle": 0, + "content": "At time step \\( t \\), the teacher model \\( f_{\\theta_t'} \\) first generates pseudo-labels to help the learning process of the student model \\( f_{\\theta_t} \\). Specifically, we compute the symmetric cross-entropy loss [12] between the outputs of the student model and the pseudo-labels. The self-training loss can be formulated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.374, + 0.804, + 0.786, + 0.844 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {S T} = - \\sum_ {c = 1} ^ {C} q _ {c} \\log p _ {c} - \\sum_ {c = 1} ^ {C} p _ {c} \\log q _ {c}, \\tag {5}\n\\]" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.226, + 0.115, + 0.733, + 0.131 + ], + "angle": 0, + "content": "Reshaping the Online Data Buffering and Organizing Mechanism for CTTA" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.193 + ], + "angle": 0, + "content": "where \\( q \\) is the softmax prediction of the teacher model, and \\( p \\) is the softmax prediction of the student model, with \\( C \\) being the total number of classes. Finally, the total loss function \\( \\mathcal{L}_T \\) can be formulated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.384, + 0.209, + 0.786, + 0.226 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {T} = \\mathcal {L} _ {S T} + \\mathcal {L} _ {P C E} + \\lambda_ {C R P} \\mathcal {L} _ {C R P}, \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.239, + 0.525, + 0.254 + ], + "angle": 0, + "content": "where \\(\\lambda_{CRP}\\) is a trade-off hyperparameter." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.285, + 0.376, + 0.303 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.324, + 0.784, + 0.385 + ], + "angle": 0, + "content": "We conduct comprehensive experiments under the CTTA setting as in [12, 51], encompassing three mainstream datasets ImageNet-to-ImageNet-C, CIFAR10-to-CIFAR10C, and CIFAR100-to-CIFAR100C for image classification, as well as Cityscapes-to-ACDC for semantic segmentation." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.415, + 0.56, + 0.431 + ], + "angle": 0, + "content": "4.1 Datasets and Experimental Settings" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.447, + 0.788, + 0.63 + ], + "angle": 0, + "content": "CIFAR10C, CIFAR100C, ImageNet-C. These datasets are originally developed to evaluate the robustness of classification networks [18]. Each dataset contains 15 distinct types of corruption at varying severity levels ranging from 1 to 5. These corruptions are applied to the test and validation images of CIFAR and ImageNet, respectively. We employ the clean training sets on CIFAR10, CIFAR100, and ImageNet as source domain datasets, while CIFAR10C, CIFAR100C, and ImageNet-C serve as the corresponding target domain datasets. Following the continuous benchmark setup in [51], we sequentially adapt the source model to 15 target domains under the largest severity level 5. The entire adaptation process is performed in an online manner, without assuming knowledge of when the domain changes. For each corruption, we utilize 10,000 images for datasets derived from CIFAR and 5,000 images for ImageNet-C." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.659, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Cityscapes-to-ACDC. This dataset is designed for the continual semantic segmentation task [51] in autonomous driving. The Cityscapes dataset [9] is used as the source domain to provide an off-the-shelf pre-trained segmentation model, while the Adverse Conditions Dataset (ACDC) [41], which contains images collected under four different weather conditions: Fog, Night, Rain, and Snow, is used to represent the changing target domains. 400 unlabeled images from each condition are used for adaptation. Note that the ACDC dataset shares the same semantic classes as the evaluation classes of Cityscapes. To simulate continual distribution shifts encountered in real-world scenarios, we cyclically repeat the same sequence of target domains ten times (e.g., in total 40: Fog \\(\\rightarrow\\) Night \\(\\rightarrow\\) Rain \\(\\rightarrow\\) Snow \\(\\rightarrow\\) Fog \\(\\rightarrow\\) ...). This also provides a long-term perspective on the performance evaluation of different adaptation methods." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.358, + 0.128 + ], + "angle": 0, + "content": "Z. Zhu et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.145, + 0.788, + 0.215 + ], + "angle": 0, + "content": "Table 2: Semantic segmentation results (mIoU in %) on the Cityscapes-to-ACDC CTTA task. The four test conditions are repeated ten times to evaluate the long-term adaptation performance. For brevity, we only show the continual adaptation results solely in the first, fourth, seventh, and final rounds. All results are evaluated based on the Segformer-B5 architecture. Bold text indicates the best performance." + }, + { + "type": "table", + "bbox": [ + 0.234, + 0.227, + 0.77, + 0.326 + ], + "angle": 0, + "content": "
Timet
Round17
ConditionFegNightRainSnowMeanFegNightRainSnowMeanFegNightRainSnowMeanFegNightRainSnowMeanFegNightRainSnowMeanFegNightRainSnowMeanFegNightRainSnowMeanFegNightRainSnowMeanMean↑
Source69.140.359.757.856.769.140.359.757.856.769.140.359.757.856.769.140.359.757.856.769.140.359.757.846.769.140.359.757.856.769.140.359.757.856.756.7
BN Stats Adapt62.338.054.653.052.062.338.054.653.052.062.338.054.653.052.062.338.054.653.052.062.338.054.653.046.762.338.054.653.052.052.0
TENT-cont. [50]69.040.260.157.356.766.536.358.754.053.064.232.855.350.950.864.232.855.350.950.861.829.851.947.847.852.3
CoTTA [51]70.941.262.459.758.670.941.262.459.758.670.941.262.459.758.670.941.262.459.758.670.941.262.459.747.858.6
Ours71.242.365.062.060.172.843.666.763.361.672.542.566.863.361.372.542.566.863.361.372.542.966.763.061.361.3
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.353, + 0.788, + 0.611 + ], + "angle": 0, + "content": "Implementation Details. To ensure consistency and comparability, we follow the implementation details from previous works [12, 51]. We adopt pre-trained WideResNet-28 [62] model for CIFAR10-to-CIFAR10C, pre-trained ResNeXt-29 [57] model for CIFAR100-to-CIFAR100C, and standard pre-trained ResNet-50 model for ImageNet-to-ImagenetC from Robustbench [10] for all methods. Our batch size is set to 64 for ImageNet-C and 200 for other classification experiments. We evaluate the model using online predictions obtained immediately as the data is encountered. For segmentation CTTA, we employ a transformer-based architecture Segformer-B5 [56] trained on Cityscapes as our segmentation model, and use down-sampled images from ACDC with a resolution of \\(960 \\times 540\\) as network inputs. We utilize the Adam optimizer with a learning rate of 1e-4 and set the batch size to 1. To generate robust pseudo-labels, we use multi-scale input with flipping as the augmentation method following [51] for the segmentation experiments. When obtaining the intrinsic class relation \\(\\hat{G}\\), we utilize the weights of the classifier as proxies for classes, whereas in classification scenarios where feature prototypes are easily accessible [4, 12], we utilize the prototypes to do so. All experiments are conducted using an NVIDIA RTX4090 GPU." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.634, + 0.57, + 0.649 + ], + "angle": 0, + "content": "4.2 Experiments on Segmentation CTTA" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.659, + 0.788, + 0.841 + ], + "angle": 0, + "content": "We report results based on the mean intersection over union (mIoU) metric for the complex continual test-time semantic segmentation Cityscapes-to-ACDC task. As detailed in Tab. 2, a discernible decline in the performance of TENT across long-term sequences is observed, highlighting the pervasive issue of model degradation over time. Conversely, our proposed method effectively mitigates catastrophic forgetting and error accumulation, thereby consistently maintaining a high mean mIoU level across repeated sequences of target domains. Particularly noteworthy is our method's achievement of the highest mIoU score, showcasing an impressive relative improvement of \\(4.6\\%\\) in mIoU compared to the previous sota method CoTTA [51], averaged across ten rounds. This significant improvement demonstrates the ability of our method to adapt to dynamic environments within dense prediction tasks over the long term. Additionally," + } + ], + [ + { + "type": "header", + "bbox": [ + 0.227, + 0.115, + 0.731, + 0.129 + ], + "angle": 0, + "content": "Reshaping the Online Data Buffering and Organizing Mechanism for CTTA" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.146, + 0.778, + 0.335 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.349, + 0.785, + 0.377 + ], + "angle": 0, + "content": "Fig. 3: Qualitative comparison of semantic segmentation on the Cityscapes-to-ACDC CTTA task. The results for all methods come from the final round." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.409, + 0.784, + 0.483 + ], + "angle": 0, + "content": "there is a trend of incremental improvement in the average mIoU metric during the initial rounds, indicating the capacity of our method to extract and leverage the reliable information embedded within the unsupervised data stream. This capability facilitates further refinement of the adaptation performance. Comprehensive results are available in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.485, + 0.788, + 0.606 + ], + "angle": 0, + "content": "For a more intuitive validation of the effectiveness of our method, we conduct qualitative comparison experiments and show the results obtained through the CTTA process in Fig. 3. Compared to other state-of-the-art methods, our results exhibit superior clarity and accuracy, yielding the best segmentation maps in all four target domains. Our method provides consistent improvements for most of the categories (as shown by the white boxes), while the comparison methods mostly suffer from severe semantic confusion. These observations demonstrate the effectiveness of our method in complex real-world environments." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.631, + 0.566, + 0.646 + ], + "angle": 0, + "content": "4.3 Experiments on Classification CTTA" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.659, + 0.788, + 0.841 + ], + "angle": 0, + "content": "ImageNet-to-ImageNet-C. Given the source model pre-trained on ImageNet, we sequentially execute CTTA on the 15 domains of ImageNet-C. As shown in Tab. 3, when directly testing the source model on target domains, the average classification error rate is as high as \\(82.0\\%\\), which demonstrates the need for adaptation. Leveraging the batch normalization statistics from the input data of the current iteration yielded the BN Stats Adapt method [26,42], a relatively straightforward approach that reduces the average classification error rate to \\(68.6\\%\\). While the TENT-based method [50] initially helps the model to continuously adapt to target domains, it is prone to severe catastrophic forgetting and error accumulation over long periods. CoTTA [51] and RMT [12] align to the source domain either in parameter space or feature space, respectively, intending to mitigate the problem of catastrophic forgetting, and both have made some" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.358, + 0.128 + ], + "angle": 0, + "content": "Z. Zhu et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.788, + 0.187 + ], + "angle": 0, + "content": "Table 3: Classification error rate (\\%, lower is better) for the standard ImageNet-to-ImageNet-C CTTA task. All results are evaluated with the largest corruption severity level 5 in an online manner. Bold text indicates the best performance." + }, + { + "type": "table", + "bbox": [ + 0.235, + 0.199, + 0.77, + 0.347 + ], + "angle": 0, + "content": "
Timet
MethodGaussian shotinputsedefocusglassmotionzoomsnowfrostfogbrightnesscontrastelasticpixelatejpegMean↓
Source97.897.198.281.789.885.278.083.577.175.941.394.582.579.368.682.0
BN Stats Adapt85.083.785.084.784.373.761.266.068.252.134.982.755.951.359.868.6
TENT-cont. [50]81.674.672.777.673.865.555.361.663.051.738.272.150.847.453.362.6
CoTTA [51]84.782.180.681.379.068.657.560.360.548.336.666.147.241.246.062.7
RMT [12]80.276.474.577.174.466.257.657.059.148.039.160.647.342.543.460.2
DSS [54]84.680.478.783.979.874.962.962.862.949.737.471.049.542.948.264.6
Ours78.575.373.075.773.164.556.055.858.147.638.558.546.142.043.459.0
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.373, + 0.784, + 0.418 + ], + "angle": 0, + "content": "progress. DSS [54] has noted the existence of high-quality and low-quality samples in the data stream, but it lacks further exploration on how to organize and utilize this information." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.419, + 0.788, + 0.584 + ], + "angle": 0, + "content": "In our method, we reshape the online data buffering mechanism in CTTA, enabling dynamic aggregation of high-value samples from the unsupervised data stream. Based on this, we introduce a novel graph-based class relation preservation constraint alongside a pseudo-target replay objective. As a result, our method achieves a substantial reduction in the average classification error rate from \\(64.6\\%\\) observed in the DSS method to \\(59.0\\%\\). Notably, our method outperforms all previous methods and reduces the error rate by a relative \\(2.0\\%\\) compared to the previous sota method, averaged across 15 domains. Moreover, our method demonstrates the best performance on the vast majority of corruption types, highlighting its effectiveness in mitigating catastrophic forgetting and error accumulation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.607, + 0.788, + 0.789 + ], + "angle": 0, + "content": "CIFAR10-to-CIFAR10C, CIFAR100-to-CIFAR100C. To further demonstrate the effectiveness of the proposed method, we evaluate it on CIFAR10-to-CIFAR10C and CIFAR100-to-CIFAR100C CTTA tasks that have different numbers of categories. The experimental results are summarized in Tab. 4. Once again, our method outperforms all previous competing methods and reduces the average error rate to \\(15.8\\%\\) and \\(29.0\\%\\), respectively. Compared with both the previous sota parameter space recovery methods (PETAL [2]) and feature space alignment methods (SANTA [4]), our method shows substantial improvements, demonstrating the superiority of the proposed class relation preservation mechanism in overcoming catastrophic forgetting. The consistency enhancement across datasets with varying levels of complexity proves the effectiveness and universality of our method." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.809, + 0.787, + 0.842 + ], + "angle": 0, + "content": "Experiments on Gradually Changing Setup. Following [4, 12, 51], we also experiment with a gradually changing setup. Different from the previous stan-" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.227, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Reshaping the Online Data Buffering and Organizing Mechanism for CTTA" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "table_caption", + "bbox": [ + 0.215, + 0.145, + 0.788, + 0.187 + ], + "angle": 0, + "content": "Table 4: Classification error rate (\\%, lower is better) for the standard CTTA tasks on CIFAR10-to-CIFAR10C, CIFAR100-to-CIFAR100C. All results are evaluated with the largest corruption severity level 5 in an online manner. Bold text indicates the best." + }, + { + "type": "table", + "bbox": [ + 0.235, + 0.199, + 0.77, + 0.469 + ], + "angle": 0, + "content": "
Timet
MethodGaussianMean↓
Source72.365.772.946.954.334.842.025.141.326.09.346.726.658.530.343.5
" + }, + { + "type": "table_caption", + "bbox": [ + 0.215, + 0.481, + 0.788, + 0.51 + ], + "angle": 0, + "content": "Table 5: Gradually changing CIFAR10-to-CIFAR10C results. The severity level changes gradually between the lowest and the highest." + }, + { + "type": "table", + "bbox": [ + 0.235, + 0.522, + 0.77, + 0.551 + ], + "angle": 0, + "content": "
Avg. Error (%)SourceBN Stats AdaptTENT-cont. [50]CoTTA [51]RMT [12]SANTA [4]Our
CIFAR10C24.713.720.410.910.410.79.9
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.579, + 0.788, + 0.645 + ], + "angle": 0, + "content": "dard setup where the corruption type changes abruptly at the largest severity level, the change process under the gradually changing setup can be represented as follows: \\(\\underbrace{..\\to 2\\to 1}_{\\text{before}}\\xrightarrow[{\\text{type}}]{\\text{change}}\\underbrace{1\\to 2\\to 3\\to 4\\to 5\\to 4\\to 3\\to 2\\to 1}_{\\text{current type, gradually changing severity}}\\xrightarrow[{\\text{type}}]{\\text{change}}\\)" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.644, + 0.788, + 0.674 + ], + "angle": 0, + "content": "\\(\\underbrace{1 \\rightarrow 2 \\rightarrow \\ldots}_{\\text {after }}\\), where each number denotes the severity of corruption. The exper" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.671, + 0.788, + 0.716 + ], + "angle": 0, + "content": "Experimental results are shown in Tab. 5, and our proposed method also achieves superior performance in this setup, reducing the average error rate to \\(9.9\\%\\) on the gradual CIFAR10-to-CIFAR10C benchmark." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.74, + 0.391, + 0.755 + ], + "angle": 0, + "content": "4.4 Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Effectiveness of Each Component. We conduct the ablation study on the CIFAR100-to-CIFAR100C scenario and evaluate the effectiveness of each component in Tab. 6. Ex1 denotes the performance using only self-training loss in the teacher-student framework, yielding a relatively high average error rate. From Ex2 to Ex4, we demonstrate the importance of our proposed uncertainty-aware" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.358, + 0.128 + ], + "angle": 0, + "content": "Z. Zhu et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.238 + ], + "angle": 0, + "content": "buffering mechanism. In contrast, the reservoir and FIFO buffering mechanisms do not have the ability to actively perform information aggregation, and thus they provide only a negligible improvement when combined with our pseudotarget replay loss. \\( Ex5 \\) demonstrates the effectiveness of our proposed class relation preservation constraint. \\( Ex6 \\) shows a significant overall improvement, suggesting that our proposed components are complementary." + }, + { + "type": "image", + "bbox": [ + 0.245, + 0.238, + 0.547, + 0.332 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.248, + 0.342, + 0.543, + 0.357 + ], + "angle": 0, + "content": "Fig. 4: Hyperparameter sensitivity analysis." + }, + { + "type": "table_caption", + "bbox": [ + 0.563, + 0.241, + 0.779, + 0.256 + ], + "angle": 0, + "content": "Table 6: Ablation experiments." + }, + { + "type": "table", + "bbox": [ + 0.566, + 0.266, + 0.777, + 0.354 + ], + "angle": 0, + "content": "
BufferingL_PCEL_CRPMean↓
Ex1---30.7
Ex2Reservoir✓-30.5
Ex3FIFO✓-30.6
Ex4Uncertainty✓-29.7
Ex5--✓30.0
Ex6Uncertainty✓✓29.0
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.378, + 0.788, + 0.621 + ], + "angle": 0, + "content": "Hyperparameter Sensitivity. Here we analyze the sensitivity of hyperparameters within the CIFAR100-to-CIFAR100C scenario. We consider three hyperparameters as illustrated in Fig. 4. Loss coefficient \\(\\lambda_{CRP}\\). The hyperparameter \\(\\lambda_{CRP}\\) in Eq. (6) controls the strength of the proposed CRP constraint. We vary it within the range of \\(\\{0,50,150,200,250,300\\}\\). It can be observed that increasing \\(\\lambda_{CRP}\\) over a reasonably broad range leads to a reduction in the average classification error rate of CTTA across 15 domains, indicating the effectiveness of CRP. While excessively large values of \\(\\lambda_{CRP}\\) (e.g., 300) may compromise the model's adaptability. Threshold coefficient \\(\\alpha\\). The hyperparameter \\(\\alpha\\) in Eq. (2) is utilized to adjust the threshold for screening low-uncertainty samples in the streaming data. It is evident that it is stable over specific ranges and achieves the lowest average error rate among all domains at a value of 0.1. Buffer capacity. Here we examine the capacity of the buffer, which is employed to dynamically aggregate high-value samples from the online unsupervised data stream. As demonstrated in Fig. 4, our method achieves optimal results with a buffer capacity of 200, imposing only a negligible storage overhead." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.643, + 0.36, + 0.658 + ], + "angle": 0, + "content": "5 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.788, + 0.841 + ], + "angle": 0, + "content": "In this paper, we revisit the key challenges in CTTA and address all of them simultaneously. To this end, we reshape the online data buffering and organizing mechanism, endowing our dynamic target buffer with the ability to aggregate reliable information in the unsupervised data stream. Through our designed class relation preservation constraint and a pseudo-target replay objective, the problem of catastrophic forgetting and error accumulation is significantly alleviated. Extensive experiments demonstrate the effectiveness of the proposed method. Regarding the adoption of uncertainty, our method employs the mainstream practice of sample entropy. In our framework, there is potential for investigating more diverse forms of uncertainty modeling, such as those based on energy functions. We leave this exploration to future work." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.227, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Reshaping the Online Data Buffering and Organizing Mechanism for CTTA" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.393, + 0.163 + ], + "angle": 0, + "content": "Acknowledgement" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.176, + 0.787, + 0.236 + ], + "angle": 0, + "content": "This work was funded in part by the National Natural Science Foundation of China (62076195, 62376070, 62206271, U20B2052), in part by the Fundamental Research Funds for the Central Universities (AUGA5710011522), and by the Pengcheng Laboratory Research Project No. PCL2023A08." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.259, + 0.323, + 0.275 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.289, + 0.785, + 0.331 + ], + "angle": 0, + "content": "1. Boudiaf, M., Mueller, R., Ben Ayed, I., Bertinetto, L.: Parameter-free online test-time adaptation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8344-8353 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.332, + 0.785, + 0.372 + ], + "angle": 0, + "content": "2. Brahma, D., Rai, P.: A probabilistic framework for lifelong test-time adaptation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 3582-3591 (June 2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.373, + 0.785, + 0.4 + ], + "angle": 0, + "content": "3. Burges, C., Ragno, R., Le, Q.: Learning to rank with nonsmooth cost functions. Advances in neural information processing systems 19 (2006)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.401, + 0.785, + 0.441 + ], + "angle": 0, + "content": "4. Chakrabarty, G., Sreenivas, M., Biswas, S.: SANTA: Source anchoring network and target alignment for continual test time adaptation. Transactions on Machine Learning Research (2023), https://openreview.net/forum?id=V7guVYzvE4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.442, + 0.785, + 0.482 + ], + "angle": 0, + "content": "5. Chen, D., Wang, D., Darrell, T., Ebrahimi, S.: Contrastive test-time adaptation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 295-305 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.484, + 0.785, + 0.523 + ], + "angle": 0, + "content": "6. Chen, H., Wang, Y., Hu, Q.: Multi-granularity regularized re-balancing for class incremental learning. IEEE Transactions on Knowledge and Data Engineering 35(7), 7263-7277 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.525, + 0.785, + 0.565 + ], + "angle": 0, + "content": "7. Choi, S., Yang, S., Choi, S., Yun, S.: Improving test-time adaptation via shift-agnostic weight regularization and nearest source prototypes. In: European Conference on Computer Vision. pp. 440-458. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.567, + 0.785, + 0.606 + ], + "angle": 0, + "content": "8. Cicek, S., Soatto, S.: Unsupervised domain adaptation via regularized conditional alignment. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 1416-1425 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.608, + 0.785, + 0.661 + ], + "angle": 0, + "content": "9. Cordts, M., Omran, M., Ramos, S., Rehfeld, T., Enzweiler, M., Benenson, R., Franke, U., Roth, S., Schiele, B.: The cityscapes dataset for semantic urban scene understanding. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 3213-3223 (June 2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.662, + 0.785, + 0.716 + ], + "angle": 0, + "content": "10. Croce, F., Andriushchenko, M., Sehwag, V., Debenedetti, E., Flammarion, N., Chiang, M., Mittal, P., Hein, M.: Robustbench: a standardized adversarial robustness benchmark. In: Thirty-fifth Conference on Neural Information Processing Systems Datasets and Benchmarks Track (Round 2) (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.717, + 0.785, + 0.758 + ], + "angle": 0, + "content": "11. Cui, S., Wang, S., Zhuo, J., Su, C., Huang, Q., Tian, Q.: Gradually vanishing bridge for adversarial domain adaptation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 12455-12464 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.759, + 0.785, + 0.799 + ], + "angle": 0, + "content": "12. Döbler, M., Marsden, R.A., Yang, B.: Robust mean teacher for continual and gradual test-time adaptation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 7704-7714 (June 2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.8, + 0.785, + 0.84 + ], + "angle": 0, + "content": "13. Dong, S., Hong, X., Tao, X., Chang, X., Wei, X., Gong, Y.: Few-shot class incremental learning via relation knowledge distillation. In: Proceedings of the AAAI Conference on Artificial Intelligence. pp. 1255-1263 (2021)" + }, + { + "type": "list", + "bbox": [ + 0.22, + 0.289, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.356, + 0.127 + ], + "angle": 0, + "content": "Z. Zhu et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.785, + 0.189 + ], + "angle": 0, + "content": "14. Fan, Y., Wang, Y., Zhu, P., Hu, Q.: Dynamic sub-graph distillation for robust semi-supervised continual learning. In: Proceedings of the AAAI Conference on Artificial Intelligence. pp. 11927-11935 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.19, + 0.785, + 0.232 + ], + "angle": 0, + "content": "15. Gan, Y., Bai, Y., Lou, Y., Ma, X., Zhang, R., Shi, N., Luo, L.: Decorate the newcomers: Visual domain prompt for continual test time adaptation. In: Proceedings of the AAAI Conference on Artificial Intelligence. pp. 7595-7603 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.232, + 0.785, + 0.259 + ], + "angle": 0, + "content": "16. Ganin, Y., Lempitsky, V.: Unsupervised domain adaptation by backpropagation. In: International conference on machine learning. pp. 1180-1189. PMLR (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.26, + 0.785, + 0.3 + ], + "angle": 0, + "content": "17. Geirhos, R., Jacobsen, J.H., Michaelis, C., Zemel, R., Brendel, W., Bethge, M., Wichmann, F.A.: Shortcut learning in deep neural networks. Nature Machine Intelligence 2(11), 665-673 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.301, + 0.785, + 0.342 + ], + "angle": 0, + "content": "18. Hendrycks, D., Dietterich, T.: Benchmarking neural network robustness to common corruptions and perturbations. In: International Conference on Learning Representations (2019), https://openreview.net/forum?id=HJz6tiCqYm" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.343, + 0.785, + 0.384 + ], + "angle": 0, + "content": "19. Hoffman, J., Tzeng, E., Park, T., Zhu, J.Y., Isola, P., Saenko, K., Efros, A., Darrell, T.: Cycada: Cycle-consistent adversarial domain adaptation. In: International conference on machine learning. pp. 1989-1998. Pmlr (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.384, + 0.785, + 0.425 + ], + "angle": 0, + "content": "20. Hwang, S., Lee, S., Kim, S., Ok, J., Kwak, S.: Combating label distribution shift for active domain adaptation. In: European Conference on Computer Vision. pp. 549-566. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.425, + 0.785, + 0.48 + ], + "angle": 0, + "content": "21. Koh, P.W., Sagawa, S., Marklund, H., Xie, S.M., Zhang, M., Balsubramani, A., Hu, W., Yasunaga, M., Phillips, R.L., Gao, I., et al.: Wilds: A benchmark of inthe-wild distribution shifts. In: International Conference on Machine Learning. pp. 5637-5664. PMLR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.481, + 0.785, + 0.508 + ], + "angle": 0, + "content": "22. Lee, K., Kim, S., Kwak, S.: Cross-domain ensemble distillation for domain generalization. In: European Conference on Computer Vision. pp. 1-20. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.509, + 0.785, + 0.55 + ], + "angle": 0, + "content": "23. Li, H., Pan, S.J., Wang, S., Kot, A.C.: Domain generalization with adversarial feature learning. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 5400-5409 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.55, + 0.785, + 0.591 + ], + "angle": 0, + "content": "24. Li, S., Xie, B., Lin, Q., Liu, C.H., Huang, G., Wang, G.: Generalized domain conditioned adaptation network. IEEE Transactions on Pattern Analysis and Machine Intelligence 44(8), 4093-4109 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.592, + 0.785, + 0.633 + ], + "angle": 0, + "content": "25. Li, S., Xie, M., Gong, K., Liu, C.H., Wang, Y., Li, W.: Transferable semantic augmentation for domain adaptation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11516-11525 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.633, + 0.785, + 0.661 + ], + "angle": 0, + "content": "26. Li, Y., Wang, N., Shi, J., Liu, J., Hou, X.: Revisiting batch normalization for practical domain adaptation. arXiv preprint arXiv:1603.04779 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.661, + 0.785, + 0.702 + ], + "angle": 0, + "content": "27. Liang, J., Hu, D., Feng, J.: Do we really need to access the source data? source hypothesis transfer for unsupervised domain adaptation. In: International conference on machine learning. pp. 6028-6039. PMLR (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.702, + 0.785, + 0.744 + ], + "angle": 0, + "content": "28. Lin, H., Zhang, Y., Qiu, Z., Niu, S., Gan, C., Liu, Y., Tan, M.: Prototype-guided continual adaptation for class-incremental unsupervised domain adaptation. In: European Conference on Computer Vision. pp. 351-368. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.744, + 0.785, + 0.785 + ], + "angle": 0, + "content": "29. Liu, H., Long, M., Wang, J., Jordan, M.: Transferable adversarial training: A general approach to adapting deep classifiers. In: International Conference on Machine Learning. pp. 4013-4022. PMLR (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.785, + 0.785, + 0.84 + ], + "angle": 0, + "content": "30. Liu, J., Yang, S., Jia, P., Zhang, R., Lu, M., Guo, Y., Xue, W., Zhang, S.: ViDA: Homeostatic visual domain adapter for continual test time adaptation. In: The Twelfth International Conference on Learning Representations (2024), https://openreview.net/forum?id=sJ88Wg5Bp5" + }, + { + "type": "list", + "bbox": [ + 0.218, + 0.147, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.227, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Reshaping the Online Data Buffering and Organizing Mechanism for CTTA" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.189 + ], + "angle": 0, + "content": "31. Liu, Y., Kothari, P., Van Delft, B., Bellot-Gurlet, B., Mordan, T., Alahi, A.: Ttt++: When does self-supervised test-time training fail or thrive? Advances in Neural Information Processing Systems 34, 21808-21820 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.19, + 0.785, + 0.23 + ], + "angle": 0, + "content": "32. Long, M., Cao, Y., Wang, J., Jordan, M.: Learning transferable features with deep adaptation networks. In: International conference on machine learning. pp. 97-105. PMLR (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.23, + 0.785, + 0.271 + ], + "angle": 0, + "content": "33. Long, M., Zhu, H., Wang, J., Jordan, M.I.: Deep transfer learning with joint adaptation networks. In: International conference on machine learning. pp. 2208-2217. PMLR (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.271, + 0.785, + 0.311 + ], + "angle": 0, + "content": "34. Nguyen, A.T., Nguyen-Tang, T., Lim, S.N., Torr, P.H.: Tipi: Test time adaptation with transformation invariance. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 24162-24171 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.311, + 0.785, + 0.366 + ], + "angle": 0, + "content": "35. Niloy, F.F., Ahmed, S.M., Raychaudhuri, D.S., Oymak, S., Roy-Chowdhury, A.K.: Effective restoration of source knowledge in continual test time adaptation. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 2091-2100 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.366, + 0.785, + 0.406 + ], + "angle": 0, + "content": "36. Niu, S., Wu, J., Zhang, Y., Chen, Y., Zheng, S., Zhao, P., Tan, M.: Efficient test-time model adaptation without forgetting. In: International conference on machine learning. pp. 16888-16905. PMLR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.406, + 0.785, + 0.462 + ], + "angle": 0, + "content": "37. Niu, S., Wu, J., Zhang, Y., Wen, Z., Chen, Y., Zhao, P., Tan, M.: Towards stable test-time adaptation in dynamic wild world. In: The Eleventh International Conference on Learning Representations (2023), https://openreview.net/forum?id=g2YraF75Tj" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.462, + 0.785, + 0.488 + ], + "angle": 0, + "content": "38. Patel, V.M., Gopalan, R., Li, R., Chellappa, R.: Visual domain adaptation: A survey of recent advances. IEEE signal processing magazine 32(3), 53-69 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.488, + 0.785, + 0.528 + ], + "angle": 0, + "content": "39. Peng, X., Bai, Q., Xia, X., Huang, Z., Saenko, K., Wang, B.: Moment matching for multi-source domain adaptation. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 1406-1415 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.528, + 0.785, + 0.569 + ], + "angle": 0, + "content": "40. Press, O., Schneider, S., Kümmerer, M., Bethge, M.: Rdumb: A simple approach that questions our progress in continual test-time adaptation. Advances in Neural Information Processing Systems 36 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.569, + 0.785, + 0.609 + ], + "angle": 0, + "content": "41. Sakaridis, C., Dai, D., Van Gool, L.: Acdc: The adverse conditions dataset with correspondences for semantic driving scene understanding. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 10765-10775 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.609, + 0.785, + 0.65 + ], + "angle": 0, + "content": "42. Schneider, S., Rusak, E., Eck, L., Bringmann, O., Brendel, W., Bethge, M.: Improving robustness against common corruptions by covariate shift adaptation. Advances in neural information processing systems 33, 11539-11551 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.65, + 0.785, + 0.704 + ], + "angle": 0, + "content": "43. Sojka, D., Cygert, S., Twardowski, B., Trzciński, T.: Ar-tta: A simple method for real-world continual test-time adaptation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) Workshops. pp. 3491-3495 (October 2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.704, + 0.785, + 0.744 + ], + "angle": 0, + "content": "44. Song, J., Lee, J., Kweon, I.S., Choi, S.: Ecotta: Memory-efficient continual test-time adaptation via self-distilled regularization. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11920-11929 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.744, + 0.785, + 0.786 + ], + "angle": 0, + "content": "45. Tan, M., Chen, G., Wu, J., Zhang, Y., Chen, Y., Zhao, P., Niu, S.: Uncertainty-calibrated test-time model adaptation without forgetting. arXiv preprint arXiv:2403.11491 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.786, + 0.785, + 0.84 + ], + "angle": 0, + "content": "46. Tao, X., Chang, X., Hong, X., Wei, X., Gong, Y.: Topology-preserving class-incremental learning. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XIX 16. pp. 254-270. Springer (2020)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.356, + 0.127 + ], + "angle": 0, + "content": "Z. Zhu et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.785, + 0.189 + ], + "angle": 0, + "content": "47. Tao, X., Hong, X., Chang, X., Dong, S., Wei, X., Gong, Y.: Few-shot class incremental learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (June 2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.189, + 0.785, + 0.229 + ], + "angle": 0, + "content": "48. Tao, X., Hong, X., Chang, X., Gong, Y.: Bi-objective continual learning: Learning 'new' while consolidating 'known'. In: Proceedings of the AAAI Conference on Artificial Intelligence. pp. 5989-5996 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.229, + 0.785, + 0.284 + ], + "angle": 0, + "content": "49. Tsai, Y.H., Hung, W.C., Schulter, S., Sohn, K., Yang, M.H., Chandraker, M.: Learning to adapt structured output space for semantic segmentation. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 7472-7481 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.284, + 0.785, + 0.325 + ], + "angle": 0, + "content": "50. Wang, D., Shelhamer, E., Liu, S., Olshausen, B., Darrell, T.: Tent: Fully test-time adaptation by entropy minimization. In: International Conference on Learning Representations (2021), https://openreview.net/forum?id=uX13bZLkr3c" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.325, + 0.785, + 0.365 + ], + "angle": 0, + "content": "51. Wang, Q., Fink, O., Van Gool, L., Dai, D.: Continual test-time domain adaptation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 7201-7211 (June 2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.365, + 0.785, + 0.405 + ], + "angle": 0, + "content": "52. Wang, S., Zhang, D., Yan, Z., Zhang, J., Li, R.: Feature alignment and uniformity for test time adaptation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 20050-20060 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.405, + 0.785, + 0.447 + ], + "angle": 0, + "content": "53. Wang, Y., Ma, Z., Huang, Z., Wang, Y., Su, Z., Hong, X.: Isolation and impartial aggregation: A paradigm of incremental learning without interference. In: Proceedings of the AAAI Conference on Artificial Intelligence. pp. 10209-10217 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.447, + 0.785, + 0.5 + ], + "angle": 0, + "content": "54. Wang, Y., Hong, J., Cheraghian, A., Rahman, S., Ahmedt-Aristizabal, D., Petersson, L., Harandi, M.: Continual test-time domain adaptation via dynamic sample selection. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV). pp. 1701-1710 (January 2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.501, + 0.785, + 0.542 + ], + "angle": 0, + "content": "55. Wei, Y., Ye, J., Huang, Z., Zhang, J., Shan, H.: Online prototype learning for online continual learning. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 18764-18774 (October 2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.542, + 0.785, + 0.609 + ], + "angle": 0, + "content": "56. Xie, E., Wang, W., Yu, Z., Anandkumar, A., Alvarez, J.M., Luo, P.: Segformer: Simple and efficient design for semantic segmentation with transformers. In: Beygelzimer, A., Dauphin, Y., Liang, P., Vaughan, J.W. (eds.) Advances in Neural Information Processing Systems (2021), https://openreview.net/forum?id=OG18MI5TRL" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.609, + 0.785, + 0.651 + ], + "angle": 0, + "content": "57. Xie, S., Girshick, R., Dollar, P., Tu, Z., He, K.: Aggregated residual transformations for deep neural networks. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1492-1500 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.651, + 0.785, + 0.718 + ], + "angle": 0, + "content": "58. Yang, X., Gu, Y., Wei, K., Deng, C.: Exploring safety supervision for continual test-time domain adaptation. In: Proceedings of the Thirty-Second International Joint Conference on Artificial Intelligence, IJCAI-23. pp. 1649-1657. International Joint Conferences on Artificial Intelligence Organization (8 2023). https://doi.org/10.24963/ijcai.2023/183" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.718, + 0.785, + 0.759 + ], + "angle": 0, + "content": "59. Yang, Y., Soatto, S.: Fda: Fourier domain adaptation for semantic segmentation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 4085-4095 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.759, + 0.785, + 0.8 + ], + "angle": 0, + "content": "60. Yao, X., Wang, Y., Zhu, P., Lin, W., Li, J., Li, W., Hu, Q.: Socialized learning: Making each other better through multi-agent collaboration. In: Forty-first International Conference on Machine Learning (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.8, + 0.785, + 0.841 + ], + "angle": 0, + "content": "61. Yuan, L., Xie, B., Li, S.: Robust test-time adaptation in dynamic scenarios. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15922-15932 (2023)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.227, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Reshaping the Online Data Buffering and Organizing Mechanism for CTTA" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "19" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.175 + ], + "angle": 0, + "content": "62. Zagoruyko, S., Komodakis, N.: Wide residual networks. In: British Machine Vision Conference 2016. British Machine Vision Association (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.175, + 0.787, + 0.217 + ], + "angle": 0, + "content": "63. Zhang, J., Qi, L., Shi, Y., Gao, Y.: Mvdg: A unified multi-view framework for domain generalization. In: European Conference on Computer Vision. pp. 161-177. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.218, + 0.785, + 0.259 + ], + "angle": 0, + "content": "64. Zhang, Y., Wang, Z., He, W.: Class relationship embedded learning for source-free unsupervised domain adaptation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7619-7629 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.259, + 0.785, + 0.313 + ], + "angle": 0, + "content": "65. Zhang, Y., Wang, Z., Li, J., Zhuang, J., Lin, Z.: Towards effective instance discrimination contrastive loss for unsupervised domain adaptation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 11388-11399 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.314, + 0.785, + 0.342 + ], + "angle": 0, + "content": "66. Zhou, K., Liu, Z., Qiao, Y., Xiang, T., Loy, C.C.: Domain generalization: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence (2022)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.342 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/Reshaping the Online Data Buffering and Organizing Mechanism for Continual Test-Time Adaptation/69885bc8-fe2a-47b1-a255-a982ec605708_origin.pdf b/2024/Reshaping the Online Data Buffering and Organizing Mechanism for Continual Test-Time Adaptation/69885bc8-fe2a-47b1-a255-a982ec605708_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..15ce617b7f0279c1180262d0772c15f6fca326f2 --- /dev/null +++ b/2024/Reshaping the Online Data Buffering and Organizing Mechanism for Continual Test-Time Adaptation/69885bc8-fe2a-47b1-a255-a982ec605708_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:144d6f6618d4c9e47bf96bb01014f7df08a63b46f193ac9aaa4826717ff26617 +size 1392993 diff --git a/2024/Reshaping the Online Data Buffering and Organizing Mechanism for Continual Test-Time Adaptation/full.md b/2024/Reshaping the Online Data Buffering and Organizing Mechanism for Continual Test-Time Adaptation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..84c339ff6c1b71118cae553abdb4ae4505c8dd15 --- /dev/null +++ b/2024/Reshaping the Online Data Buffering and Organizing Mechanism for Continual Test-Time Adaptation/full.md @@ -0,0 +1,290 @@ +# Reshaping the Online Data Buffering and Organizing Mechanism for Continual Test-Time Adaptation + +Zhilin Zhu $^{1,2}$ , Xiaopeng Hong $^{1,2\star}$ , Zhiheng Ma $^{3,4,5}$ , Weijun Zhuang $^{1,2}$ , Yaohui Ma $^{1,5}$ , Yong Dai $^{2}$ , and Yaowei Wang $^{2,1}$ + +1 Harbin Institute of Technology + +2 Pengcheng Laboratory + +3 Shenzhen University of Advanced Technology + +4 Guangdong Provincial Key Laboratory of Computility Microelectronics + +Shenzhen Institute of Advanced Technology, Chinese Academy of Science {zhuzhilin, weijunzhuang}@stu.hit.edu.cn hongxiaopeng@ieee.org ma@siat.ac.cn yh.ma@siat.ac.cn chd-dy@foxmail.com wangyw@pcl.ac.cn + +Abstract. Continual Test-Time Adaptation (CTTA) involves adapting a pre-trained source model to continually changing unsupervised target domains. In this paper, we systematically analyze the challenges of this task: online environment, unsupervised nature, and the risks of error accumulation and catastrophic forgetting under continual domain shifts. To address these challenges, we reshape the online data buffering and organizing mechanism for CTTA. We propose an uncertainty-aware buffering approach to identify and aggregate significant samples with high certainty from the unsupervised, single-pass data stream. Based on this, we propose a graph-based class relation preservation constraint to overcome catastrophic forgetting. Furthermore, a pseudo-target replay objective is used to mitigate error accumulation. Extensive experiments demonstrate the superiority of our method in both segmentation and classification CTTA tasks. Code is available at https://github.com/z1358/OBAO. + +Keywords: Continual test-time adaptation $\cdot$ Unsupervised learning $\cdot$ Continual learning $\cdot$ Catastrophic forgetting + +# 1 Introduction + +Deep learning models have exhibited remarkable advancements in scenarios characterized by a consistent data distribution between training and test environments. However, the deployment of machine perception systems in the real world frequently exposes them to non-stationary and constantly changing environments [14, 21, 47, 53, 60]. When such phenomena occur (e.g., unexpected domain shifts over time), these models are prone to severe performance degradation [18, 37, 51]. Consequently, the development of techniques for continual adaptation during the unsupervised deployment phase is essential for enhancing the reliability of machine perception systems. + +Table 1: Comparison of different adaptation settings. We differentiate along four dimensions: the available form of the source domain data and target domain data, the distribution and access mode of the target domain data stream during adaptation. + +
SettingDataTarget domain data stream
SourceTargetDistributionAccess mode
Fine-tuning-xT,yTstationaryoffline (multi-epoch)
Domain GeneralizationxS,yS---
Domain AdaptationxS,ySxTstationaryoffline (multi-epoch)
Test-Time TrainingxS,ySxTstationaryonline (single-pass)
Test-Time Adaptation-xTstationaryoffline or online
Continual Test-Time Adaptation-x1T→···→xnTcontinually changingonline (single-pass)
+ +With sufficient source labeled data accessible, the fields of domain generalization [22, 23, 63, 66] and domain adaptation [20, 24, 25, 28] have witnessed considerable success. However, a particularly challenging scenario arises when no source data is available, and unsupervised adaptation must happen at test time when unlabeled test inputs arrive. This task is known as test-time adaptation (TTA) [5, 7, 45]. Given the rarity of encountering only a singular domain shift in practical applications, the community has recently been increasingly focusing on continual test-time adaptation (CTTA) [2, 12, 30, 51, 58], i.e., evaluating the ability of methods to adapt to continual domain shifts. We summarize the differences between existing adaptation settings in Tab. 1. Obviously, the considered CTTA setting is the most practical and challenging. + +Within this setting, we expose the key challenges as follows: (1) Online environment. This denotes a scenario where data is processed in a single pass, necessitating immediate predictions from the model. Due to the characteristics of the online environment, it is difficult for the model to obtain a sufficient number of informative samples, thereby impeding the efficacy of adaptation; (2) Unsupervised nature. Throughout adaptation, ground-truth labels for input data are unavailable. Moreover, due to privacy concerns or legal restrictions, the source domain data may no longer be available and we only have access to the source pre-trained model; (3) Error accumulation and catastrophic forgetting. The dynamic distribution shift when adapting to continually changing target domains can render pseudo-labels unreliable and impede knowledge retention about the source domain, leading to error accumulation [51, 54] and triggering catastrophic forgetting [6, 30, 58]. + +Addressing the first challenge, an intuitive solution is to employ a buffer to store high-value samples. However, existing methods that utilize such a buffer store source domain samples with ground-truth labels [12, 43], which is difficult to satisfy in real-world scenarios and conflicts with the source-free constraint in the second challenge. Therefore, the issue of effectively buffering and organizing high-value samples in an unsupervised single-pass data stream has not been fully explored. Current CTTA methods primarily address the third challenge, focusing mainly on alignments in either the parameter space [2, 35, 40, 51] or the feature space [4, 12] of the source model to mitigate knowledge forgetting. + +![](images/fad5669ba277ebfa7db9e8f59138252a4fb8af44d55f35bf7fdc27916c238c29.jpg) +(a) W/o a buffer + +![](images/70dbd4dbd6ff3daec0498c281c295a40103b6c089478258ce7fb34141148fdeb.jpg) +(b) Static source buffer + +![](images/2f4e1d8c73e641ba711b0b528813684ae73945523f6b4b69581c6490f3e639c7.jpg) +(c) Our dynamic target buffer +Fig. 1: Comparison of different buffering mechanisms. (a) Without a buffer, only unreliable signals from the current unsupervised input data are available. (b) Some methods [12, 43] use a static source buffer that stores source samples with ground-truth labels to provide supervised signals. (c) We work on a dynamic target buffer, which can provide reliable signals for the adaptation process. + +In this paper, we endeavor to address these three challenges simultaneously. To this end, we reshape the online data buffering and organizing mechanism for CTTA. As shown in Fig. 1, unlike previous methods which directly sample from the source domain [12, 43], we are targeted at source-free, unsupervised, and online buffering. Through a specially designed uncertainty-aware buffering approach, we dynamically evaluate and aggregate reliable information in the unsupervised, single-pass data stream (Challenges 1 and 2). Based on this, we efficiently organize these samples to assist the model's unsupervised adaptation process. Utilizing pseudo-labels generated from these low-uncertainty samples, we consider them as anchors of classes and organize them into a class relation graph (CRG). To combat catastrophic forgetting under continually changing domains, we introduce a novel graph-based class relation preservation constraint to maintain the intrinsic class relations obtained on the source domain (Challenge 3). Specifically, we penalize topological changes within the CRG to mitigate knowledge forgetting, while allowing reasonable vertex movements to enhance adaptability. Additionally, these samples are incorporated into a pseudo-target replay objective to reduce the risk of error accumulation from misclassified samples (Challenge 3). In summary, our main contributions are as follows: + +- To address the three revealed challenges simultaneously, we reshape the online data buffering and organizing mechanism for CTTA. +- To frame the buffering mechanism, we design an uncertainty-aware approach to identify and aggregate reliable data from unsupervised online data stream. +- Through a well-designed graph-based class relation preservation constraint and a pseudo-target replay objective, we effectively mitigate catastrophic forgetting and error accumulation in CTTA. +- We extensively evaluate our proposed method and the results on four challenging datasets show that the method clearly improves the CTTA performance on both segmentation and classification benchmarks, indicating its effectiveness and universality. + +# 2 Related Work + +# 2.1 Domain Adaptation + +Domain adaptation (DA) [38,65] aims to generalize a model from a well-annotated source domain to an unlabeled target domain following the transductive learning principle. The underlying assumption in DA is that both labeled source domain and unlabeled target domain data are concurrently accessible. Consequently, the prevalent methods in DA encompass the utilization of domain adversarial training [8,11,16,19,29,49] or the minimization of domain discrepancy [32,33,39]. These methods work towards harmonizing the feature distribution [16,32,49] or the input space [19,59] between the two domains. + +# 2.2 Continual Test-Time Adaptation + +Test-Time Adaptation (TTA). TTA focuses on a more challenging setting after model deployment where only a pre-trained source model and unlabeled target data are available. Some approaches [5, 27] focus on the offline setting, where multiple epochs of access to the entire set of test data are provided for adaptation, while others [1, 7, 31, 52] focus on the online setting with a single-pass data stream. To name a few, TENT [50] adapts the model by using entropy minimization loss to update a few trainable parameters in batch normalization layers. This approach has also inspired subsequent works such as EATA [36] and RoTTA [61] to investigate the robustness of normalized layers. Due to the domain gap, TSD [52] promotes feature alignment and uniformity from the perspective of feature revision, while TIPI [34] proposes invariance regularizers to find the transformation space that can simulate the domain shift. + +Continual Test-Time Adaptation (CTTA). In recent years, CTTA has received increasing research attention [15,44,51,58] due to its practicality. CoTTA [51] stands out as the first method specifically tailored to address this task. It utilizes weight-averaged and augmentation-averaged predictions to reduce error accumulation, while stochastically restoring the weights of a small fraction of neurons to the source model during each iteration to mitigate catastrophic forgetting. Building upon this, PETAL [2] further introduces a data-driven parameter recovery technique to align with the source model in the parameter space, while RMT [12] and SANTA [4] seek alignment with the source model in the feature space by optimizing an additional contrastive learning objective. Additionally, DSS [54] considers sample security to mitigate the risk of using incorrect information. Some approaches [15, 30, 44] also prevent forgetting by freezing a majority or all of the model parameters, but they require optimization along with the model during training using the source domain data, thereby relaxing the unsupervised challenge and not being truly source-free. Under our proposal, we simultaneously address three challenges in CTTA. + +![](images/c625ea5d503b5153c2d15d12f3ebce3d9c91ec6f5d8c0cc585af3169df2bf172.jpg) +Fig. 2: The framework of our proposed method. At time step $t$ , the incoming data and sampled buffer data are used for adaptation. The proposed class relation preservation loss $\mathcal{L}_{CRP}$ ensures topological consistency between the Class Relation Graph (CRG) $G_{t}$ estimated at the current time step $t$ and $\hat{G}$ from the source domain, thereby effectively preserving intrinsic semantic information. Concurrently, we integrate a pseudo-target replay loss $\mathcal{L}_{PCE}$ to mitigate error accumulation. + +# 3 Methodology + +# 3.1 Preliminaries and Overall Framework + +In continual test-time adaptation (CTTA), we are given a source model $f_{\theta_0}$ pretrained from the labeled source domain $(\mathcal{X}^S, \mathcal{Y}^S)$ and a series of sequentially changing unlabeled target domains such as $[\mathcal{X}_1^T, \mathcal{X}_2^T, \dots, \mathcal{X}_n^T]$ . Each target domain shares the same label space as the source domain (for close-set adaptation). Note that the model has no information about when the domain changes happen during the adaptation process. The goal of CTTA is to adapt the model to a continually changing unsupervised online data stream of target domains without source data. To be more specific, at inference time step $t$ , the unlabeled target test data $x(t)$ is given to the model $f_{\theta_t}$ , and the model needs to make predictions and adapt itself accordingly $(f_{\theta_t} \rightarrow f_{\theta_{t+1}})$ for the upcoming test inputs $x(t+1)$ . + +We adopt a teacher-student framework following previous works [2,12,51] as illustrated in Fig. 2. Leveraging the proposed uncertainty-aware buffering mechanism, our method dynamically evaluates and buffers significant high-certainty samples in the unsupervised online data stream, providing reliable information for the model's adaptation process. Building upon this, we further introduce a novel class relation preservation constraint $(\mathcal{L}_{CRP})$ for preserving domain-invariant semantic information in dynamic environments, effectively mitigating the catastrophic forgetting problem. Additionally, we address error accumulation through a pseudo-target replay objective $(\mathcal{L}_{PCE})$ to minimize the impact of miscalibration. A self-training loss $(\mathcal{L}_{ST})$ is used to stabilize the adaptation + +process of the student model (adapted model). In the adopted framework, both the teacher model $f_{\theta'}$ and the student model $f_{\theta}$ are initialized with the source model $f_{\theta_0}$ . Subsequently, the teacher model is continuously updated by the exponential moving average of the student model. We elaborate our method in the following subsections. + +# 3.2 Uncertainty-aware Buffering + +In the unsupervised online data stream within CTTA, the model necessitates adaptation utilizing the data at hand, typically characterized by a scarcity of information. While certain methods [12, 43] endeavor to incorporate additional supervised information by formulating a static buffer from the source domain, fulfilling this requirement proves challenging within practical application scenarios. Conversely, our proposition involves sampling and buffering throughout the unsupervised adaptation process by a designed uncertainty-aware approach. + +At the outset, the dynamic buffer is initialized as empty. As the adaptation progresses, the buffer undergoes continual updates to reflect the evolving nature of the data stream encountered. Specifically, at any given inference time step $t$ , the model calculates the uncertainty associated with the current batch of samples. The quantification of sample uncertainty manifests as the entropy of the sample, denoted as: + +$$ +H (x (t)) = - f _ {\theta_ {t}} (x (t)) ^ {\top} (\log f _ {\theta_ {t}} (x (t))), \tag {1} +$$ + +where $f_{\theta_t}(x(t))$ corresponds the softmax predictions of the student model for the current input data. Samples exhibiting an entropy value less than a predefined threshold $H_0$ are deemed to possess a high degree of certainty and are thus considered candidates for inclusion in the dynamic buffer. $H_0$ is defined as: + +$$ +H _ {0} = \alpha \ln C, \tag {2} +$$ + +where $\alpha$ serves as a threshold coefficient and $C$ denotes the total number of classes in the dataset. The process of storing samples into the buffer, or updating existing entries, is contingent upon the buffer's capacity. High-certainty samples identified from the present batch, along with their derived pseudo-labels from the teacher model, are added to the buffer if space permits. Conversely, in scenarios where the buffer is at capacity, these new samples replace the ones exhibiting the highest uncertainty to maintain an optimal blend of timeliness and reliability. Upon completion of these updates, the resulting buffer $\mathcal{M}_t$ stands ready to facilitate the model's adaptation in subsequent iterations. + +Notably, our approach to buffering is entirely devoid of reliance on source data, underscoring its source-free nature. This strategy enables the model to discern and prioritize data based on their uncertainty. By dynamically buffering high-certainty samples within the data stream, our approach not only circumvents the limitations posed by reliance on static source buffers but also enriches the adaptation process with valuable, reliable data. The intuitive idea is to utilize CE loss to replay samples in the buffer, thereby emphasizing reliable data + +and effectively suppressing the risk and impact of miscalibration due to error accumulation. In each test-time adaptation iteration, we randomly sample examples and labels $(\tilde{x}(t), \tilde{y}(t))$ from the buffer. The number of sampled examples is equal to the batch size and the corresponding pseudo-target replay loss is: + +$$ +\mathcal {L} _ {P C E} = - \sum_ {c = 1} ^ {C} \tilde {y} _ {c} (t) \log \tilde {p} _ {c}, \tag {3} +$$ + +where $\tilde{p}$ is the softmax prediction of the student model for $\tilde{x}(t)$ . For simplicity, we omit the summation of all samples in the batch. Moreover, these sampled examples also serve as anchors and are organized in a class relation graph. On this basis, we introduce a novel class relation preservation constraint in Sec. 3.3 to preserve domain-invariant knowledge and overcome forgetting. + +# 3.3 Class Relation Preservation Constraint + +In online scenarios where domains are constantly changing, neural networks tend to "take shortcuts" and focus on simplistic features [17, 55] from the single-pass data stream. This phenomenon results in the model's forgetting of intrinsic semantic information. In this work, we propose to overcome catastrophic forgetting by preserving the domain-invariant class relation, which is consistent across domains regardless of distribution discrepancy [64]. Therefore, it is reasonable to utilize the intrinsic class relation obtained on the source domain with large labeled samples to constrain the model adaptation process on target domains. + +We describe here how to obtain the defined class relation graph (CRG). For computational stability, we normalize the feature space and adopt the cosine similarity metric. Let $\bar{\cdot}$ denotes the normalization operation, where $\bar{f} = f / \| f\|$ . Given the normalized feature space $\bar{\mathcal{F}}$ , the CRG $G$ is defined as $G = \langle V,E\rangle$ , where $V = \{\bar{v}_1,\dots ,\bar{v}_C|\bar{v}_i\in \bar{\mathcal{F}}\}$ is the set of $C$ vertices representative of $\bar{\mathcal{F}}$ , and $E$ is the edge set describing the neighborhood relations of vertices in $V$ . Each edge $e_{ij}$ is assigned with a weight $s_{ij}$ , which is the similarity between $\bar{v}_i$ and $\bar{v}_j$ : $s_{ij} = \bar{v}_i^\top \bar{v}_j$ . Each vertex $\bar{v}_i$ is the centroid vector of class $i$ . The intrinsic CRG $\hat{G}$ is obtained by modeling the source-domain class relation. We demonstrate in experiments that the construction of our CRG $\hat{G}$ can be derived from different data conditions, including class prototypes and classifier weights. + +At time step $t$ , given CRG $G_{t} = \langle V_{t}, E_{t} \rangle$ , when catastrophic forgetting occurs, $G_{t}$ is distorted as the semantic relations among categories are confused. To mitigate forgetting, the model should be adapted to strive towards preserving the topology in $G_{t}$ the same as that in $\hat{G}$ . This is achieved by restricting the neighboring relations of vertices described by the weights of edges (i.e., similarity between classes) in $E_{t}$ . One way to do this is to keep the ranking of the edge weights constant during the adaptation process. However optimizing a non-smooth global ranking is difficult and inefficient [3,46]. An alternative and more feasible approach involves assessing the extent of change in neighboring relations [13,48] by evaluating the correlation between the initial edge weights + +and their current observation. A lower correlation indicates a higher probability that the rank of an edge has changed during adaptation, and this should be penalized. On this basis, we define the Class Relation Preservation constraint (CRP) as follows: + +$$ +\mathcal {L} _ {C R P} = - \frac {\sum_ {i , j} ^ {C} s _ {i j} \tilde {s} _ {i j}}{\sqrt {\sum_ {i , j} ^ {C} s _ {i j} ^ {2}} \sqrt {\sum_ {i , j} \tilde {s} _ {i j} ^ {2}}}, \tag {4} +$$ + +where $S = \{s_{ij}|1\leq i,j\leq C\}$ and $\tilde{S} = \{\tilde{s}_{ij}|1\leq i,j\leq C\}$ are the sets of the initial and observation values of the edge weights in $\hat{E}$ and $E_{t}$ , respectively. + +To achieve end-to-end optimization, the ideal situation is using source-domain images with ground-truth labels to obtain an estimate of the current CRG $G_{t}$ under time step $t$ . That is, extracting features from these images by $f_{\theta_t}$ , which in turn yields the class prototypes and ultimately the class relation. However, it is often impractical to store a certain number of source-domain samples in the source-free CTTA setting. In this work, we leverage reliable data from our proposed uncertainty-aware buffering mechanism to accurately estimate CRG $G_{t}$ and achieve end-to-end optimization in an unsupervised test data stream. + +From the perspective of viewing $S$ as a vector, the proposed CRP term constrains the consistency of the direction between two vectors during the adaptation process. Opting for the origin moment instead of the central moment allows the CRP term to prioritize direction consistency between vectors. This makes it more robust to scale variations or magnitude disparities in edge weights, which may arise from shifts in data distribution. Some previous methods [4,12] directly penalize the movement of target domain samples in the feature space relative to the source prototypes. This can be broadly interpreted as penalizing the movement of corresponding elements between $\hat{V}$ and $V_{t}$ in our defined CRG. However, in addition to suffering from classification errors when assigning source prototypes, constraints on features can also impede the effectiveness of adaptation in target domains. Instead of penalizing the movement of vertices in $G$ in the feature space, the proposed CRP term penalizes the change of topological relations between vertices, while allowing reasonable movement of vertices to enhance adaptation. To summarize, by preserving the domain-invariant class relation during the model adaptation process, the proposed CRP term effectively overcomes catastrophic forgetting caused by semantic confusion. + +# 3.4 Optimization Objective + +At time step $t$ , the teacher model $f_{\theta_t'}$ first generates pseudo-labels to help the learning process of the student model $f_{\theta_t}$ . Specifically, we compute the symmetric cross-entropy loss [12] between the outputs of the student model and the pseudo-labels. The self-training loss can be formulated as follows: + +$$ +\mathcal {L} _ {S T} = - \sum_ {c = 1} ^ {C} q _ {c} \log p _ {c} - \sum_ {c = 1} ^ {C} p _ {c} \log q _ {c}, \tag {5} +$$ + +where $q$ is the softmax prediction of the teacher model, and $p$ is the softmax prediction of the student model, with $C$ being the total number of classes. Finally, the total loss function $\mathcal{L}_T$ can be formulated as follows: + +$$ +\mathcal {L} _ {T} = \mathcal {L} _ {S T} + \mathcal {L} _ {P C E} + \lambda_ {C R P} \mathcal {L} _ {C R P}, \tag {6} +$$ + +where $\lambda_{CRP}$ is a trade-off hyperparameter. + +# 4 Experiments + +We conduct comprehensive experiments under the CTTA setting as in [12, 51], encompassing three mainstream datasets ImageNet-to-ImageNet-C, CIFAR10-to-CIFAR10C, and CIFAR100-to-CIFAR100C for image classification, as well as Cityscapes-to-ACDC for semantic segmentation. + +# 4.1 Datasets and Experimental Settings + +CIFAR10C, CIFAR100C, ImageNet-C. These datasets are originally developed to evaluate the robustness of classification networks [18]. Each dataset contains 15 distinct types of corruption at varying severity levels ranging from 1 to 5. These corruptions are applied to the test and validation images of CIFAR and ImageNet, respectively. We employ the clean training sets on CIFAR10, CIFAR100, and ImageNet as source domain datasets, while CIFAR10C, CIFAR100C, and ImageNet-C serve as the corresponding target domain datasets. Following the continuous benchmark setup in [51], we sequentially adapt the source model to 15 target domains under the largest severity level 5. The entire adaptation process is performed in an online manner, without assuming knowledge of when the domain changes. For each corruption, we utilize 10,000 images for datasets derived from CIFAR and 5,000 images for ImageNet-C. + +Cityscapes-to-ACDC. This dataset is designed for the continual semantic segmentation task [51] in autonomous driving. The Cityscapes dataset [9] is used as the source domain to provide an off-the-shelf pre-trained segmentation model, while the Adverse Conditions Dataset (ACDC) [41], which contains images collected under four different weather conditions: Fog, Night, Rain, and Snow, is used to represent the changing target domains. 400 unlabeled images from each condition are used for adaptation. Note that the ACDC dataset shares the same semantic classes as the evaluation classes of Cityscapes. To simulate continual distribution shifts encountered in real-world scenarios, we cyclically repeat the same sequence of target domains ten times (e.g., in total 40: Fog $\rightarrow$ Night $\rightarrow$ Rain $\rightarrow$ Snow $\rightarrow$ Fog $\rightarrow$ ...). This also provides a long-term perspective on the performance evaluation of different adaptation methods. + +Table 2: Semantic segmentation results (mIoU in %) on the Cityscapes-to-ACDC CTTA task. The four test conditions are repeated ten times to evaluate the long-term adaptation performance. For brevity, we only show the continual adaptation results solely in the first, fourth, seventh, and final rounds. All results are evaluated based on the Segformer-B5 architecture. Bold text indicates the best performance. + +
Timet
Round17
ConditionFegNightRainSnowMeanFegNightRainSnowMeanFegNightRainSnowMeanFegNightRainSnowMeanFegNightRainSnowMeanFegNightRainSnowMeanFegNightRainSnowMeanFegNightRainSnowMeanMean↑
Source69.140.359.757.856.769.140.359.757.856.769.140.359.757.856.769.140.359.757.856.769.140.359.757.846.769.140.359.757.856.769.140.359.757.856.756.7
BN Stats Adapt62.338.054.653.052.062.338.054.653.052.062.338.054.653.052.062.338.054.653.052.062.338.054.653.046.762.338.054.653.052.052.0
TENT-cont. [50]69.040.260.157.356.766.536.358.754.053.064.232.855.350.950.864.232.855.350.950.861.829.851.947.847.852.3
CoTTA [51]70.941.262.459.758.670.941.262.459.758.670.941.262.459.758.670.941.262.459.758.670.941.262.459.747.858.6
Ours71.242.365.062.060.172.843.666.763.361.672.542.566.863.361.372.542.566.863.361.372.542.966.763.061.361.3
+ +Implementation Details. To ensure consistency and comparability, we follow the implementation details from previous works [12, 51]. We adopt pre-trained WideResNet-28 [62] model for CIFAR10-to-CIFAR10C, pre-trained ResNeXt-29 [57] model for CIFAR100-to-CIFAR100C, and standard pre-trained ResNet-50 model for ImageNet-to-ImagenetC from Robustbench [10] for all methods. Our batch size is set to 64 for ImageNet-C and 200 for other classification experiments. We evaluate the model using online predictions obtained immediately as the data is encountered. For segmentation CTTA, we employ a transformer-based architecture Segformer-B5 [56] trained on Cityscapes as our segmentation model, and use down-sampled images from ACDC with a resolution of $960 \times 540$ as network inputs. We utilize the Adam optimizer with a learning rate of 1e-4 and set the batch size to 1. To generate robust pseudo-labels, we use multi-scale input with flipping as the augmentation method following [51] for the segmentation experiments. When obtaining the intrinsic class relation $\hat{G}$ , we utilize the weights of the classifier as proxies for classes, whereas in classification scenarios where feature prototypes are easily accessible [4, 12], we utilize the prototypes to do so. All experiments are conducted using an NVIDIA RTX4090 GPU. + +# 4.2 Experiments on Segmentation CTTA + +We report results based on the mean intersection over union (mIoU) metric for the complex continual test-time semantic segmentation Cityscapes-to-ACDC task. As detailed in Tab. 2, a discernible decline in the performance of TENT across long-term sequences is observed, highlighting the pervasive issue of model degradation over time. Conversely, our proposed method effectively mitigates catastrophic forgetting and error accumulation, thereby consistently maintaining a high mean mIoU level across repeated sequences of target domains. Particularly noteworthy is our method's achievement of the highest mIoU score, showcasing an impressive relative improvement of $4.6\%$ in mIoU compared to the previous sota method CoTTA [51], averaged across ten rounds. This significant improvement demonstrates the ability of our method to adapt to dynamic environments within dense prediction tasks over the long term. Additionally, + +![](images/e859bf18adab27983f5c49330d18085db872a801e43f91cd0e67f3f6e5cb5eea.jpg) +Fig. 3: Qualitative comparison of semantic segmentation on the Cityscapes-to-ACDC CTTA task. The results for all methods come from the final round. + +there is a trend of incremental improvement in the average mIoU metric during the initial rounds, indicating the capacity of our method to extract and leverage the reliable information embedded within the unsupervised data stream. This capability facilitates further refinement of the adaptation performance. Comprehensive results are available in the supplementary material. + +For a more intuitive validation of the effectiveness of our method, we conduct qualitative comparison experiments and show the results obtained through the CTTA process in Fig. 3. Compared to other state-of-the-art methods, our results exhibit superior clarity and accuracy, yielding the best segmentation maps in all four target domains. Our method provides consistent improvements for most of the categories (as shown by the white boxes), while the comparison methods mostly suffer from severe semantic confusion. These observations demonstrate the effectiveness of our method in complex real-world environments. + +# 4.3 Experiments on Classification CTTA + +ImageNet-to-ImageNet-C. Given the source model pre-trained on ImageNet, we sequentially execute CTTA on the 15 domains of ImageNet-C. As shown in Tab. 3, when directly testing the source model on target domains, the average classification error rate is as high as $82.0\%$ , which demonstrates the need for adaptation. Leveraging the batch normalization statistics from the input data of the current iteration yielded the BN Stats Adapt method [26,42], a relatively straightforward approach that reduces the average classification error rate to $68.6\%$ . While the TENT-based method [50] initially helps the model to continuously adapt to target domains, it is prone to severe catastrophic forgetting and error accumulation over long periods. CoTTA [51] and RMT [12] align to the source domain either in parameter space or feature space, respectively, intending to mitigate the problem of catastrophic forgetting, and both have made some + +Table 3: Classification error rate (\%, lower is better) for the standard ImageNet-to-ImageNet-C CTTA task. All results are evaluated with the largest corruption severity level 5 in an online manner. Bold text indicates the best performance. + +
Timet
MethodGaussian shotinputsedefocusglassmotionzoomsnowfrostfogbrightnesscontrastelasticpixelatejpegMean↓
Source97.897.198.281.789.885.278.083.577.175.941.394.582.579.368.682.0
BN Stats Adapt85.083.785.084.784.373.761.266.068.252.134.982.755.951.359.868.6
TENT-cont. [50]81.674.672.777.673.865.555.361.663.051.738.272.150.847.453.362.6
CoTTA [51]84.782.180.681.379.068.657.560.360.548.336.666.147.241.246.062.7
RMT [12]80.276.474.577.174.466.257.657.059.148.039.160.647.342.543.460.2
DSS [54]84.680.478.783.979.874.962.962.862.949.737.471.049.542.948.264.6
Ours78.575.373.075.773.164.556.055.858.147.638.558.546.142.043.459.0
+ +progress. DSS [54] has noted the existence of high-quality and low-quality samples in the data stream, but it lacks further exploration on how to organize and utilize this information. + +In our method, we reshape the online data buffering mechanism in CTTA, enabling dynamic aggregation of high-value samples from the unsupervised data stream. Based on this, we introduce a novel graph-based class relation preservation constraint alongside a pseudo-target replay objective. As a result, our method achieves a substantial reduction in the average classification error rate from $64.6\%$ observed in the DSS method to $59.0\%$ . Notably, our method outperforms all previous methods and reduces the error rate by a relative $2.0\%$ compared to the previous sota method, averaged across 15 domains. Moreover, our method demonstrates the best performance on the vast majority of corruption types, highlighting its effectiveness in mitigating catastrophic forgetting and error accumulation. + +CIFAR10-to-CIFAR10C, CIFAR100-to-CIFAR100C. To further demonstrate the effectiveness of the proposed method, we evaluate it on CIFAR10-to-CIFAR10C and CIFAR100-to-CIFAR100C CTTA tasks that have different numbers of categories. The experimental results are summarized in Tab. 4. Once again, our method outperforms all previous competing methods and reduces the average error rate to $15.8\%$ and $29.0\%$ , respectively. Compared with both the previous sota parameter space recovery methods (PETAL [2]) and feature space alignment methods (SANTA [4]), our method shows substantial improvements, demonstrating the superiority of the proposed class relation preservation mechanism in overcoming catastrophic forgetting. The consistency enhancement across datasets with varying levels of complexity proves the effectiveness and universality of our method. + +Experiments on Gradually Changing Setup. Following [4, 12, 51], we also experiment with a gradually changing setup. Different from the previous stan- + +Table 4: Classification error rate (\%, lower is better) for the standard CTTA tasks on CIFAR10-to-CIFAR10C, CIFAR100-to-CIFAR100C. All results are evaluated with the largest corruption severity level 5 in an online manner. Bold text indicates the best. + +
Timet
MethodGaussianMean↓
Source72.365.772.946.954.334.842.025.141.326.09.346.726.658.530.343.5
+ +Table 5: Gradually changing CIFAR10-to-CIFAR10C results. The severity level changes gradually between the lowest and the highest. + +
Avg. Error (%)SourceBN Stats AdaptTENT-cont. [50]CoTTA [51]RMT [12]SANTA [4]Our
CIFAR10C24.713.720.410.910.410.79.9
+ +dard setup where the corruption type changes abruptly at the largest severity level, the change process under the gradually changing setup can be represented as follows: $\underbrace{..\to 2\to 1}_{\text{before}}\xrightarrow[{\text{type}}]{\text{change}}\underbrace{1\to 2\to 3\to 4\to 5\to 4\to 3\to 2\to 1}_{\text{current type, gradually changing severity}}\xrightarrow[{\text{type}}]{\text{change}}$ + +$\underbrace{1 \rightarrow 2 \rightarrow \ldots}_{\text {after }}$ , where each number denotes the severity of corruption. The exper + +Experimental results are shown in Tab. 5, and our proposed method also achieves superior performance in this setup, reducing the average error rate to $9.9\%$ on the gradual CIFAR10-to-CIFAR10C benchmark. + +# 4.4 Ablation Study + +Effectiveness of Each Component. We conduct the ablation study on the CIFAR100-to-CIFAR100C scenario and evaluate the effectiveness of each component in Tab. 6. Ex1 denotes the performance using only self-training loss in the teacher-student framework, yielding a relatively high average error rate. From Ex2 to Ex4, we demonstrate the importance of our proposed uncertainty-aware + +buffering mechanism. In contrast, the reservoir and FIFO buffering mechanisms do not have the ability to actively perform information aggregation, and thus they provide only a negligible improvement when combined with our pseudotarget replay loss. $Ex5$ demonstrates the effectiveness of our proposed class relation preservation constraint. $Ex6$ shows a significant overall improvement, suggesting that our proposed components are complementary. + +![](images/088c7945566c5ef7bb306c0a1fc671551b35338ab370eaa56aae7bb2d079ce50.jpg) +Fig. 4: Hyperparameter sensitivity analysis. + +Table 6: Ablation experiments. + +
BufferingL_PCEL_CRPMean↓
Ex1---30.7
Ex2Reservoir✓-30.5
Ex3FIFO✓-30.6
Ex4Uncertainty✓-29.7
Ex5--✓30.0
Ex6Uncertainty✓✓29.0
+ +Hyperparameter Sensitivity. Here we analyze the sensitivity of hyperparameters within the CIFAR100-to-CIFAR100C scenario. We consider three hyperparameters as illustrated in Fig. 4. Loss coefficient $\lambda_{CRP}$ . The hyperparameter $\lambda_{CRP}$ in Eq. (6) controls the strength of the proposed CRP constraint. We vary it within the range of $\{0,50,150,200,250,300\}$ . It can be observed that increasing $\lambda_{CRP}$ over a reasonably broad range leads to a reduction in the average classification error rate of CTTA across 15 domains, indicating the effectiveness of CRP. While excessively large values of $\lambda_{CRP}$ (e.g., 300) may compromise the model's adaptability. Threshold coefficient $\alpha$ . The hyperparameter $\alpha$ in Eq. (2) is utilized to adjust the threshold for screening low-uncertainty samples in the streaming data. It is evident that it is stable over specific ranges and achieves the lowest average error rate among all domains at a value of 0.1. Buffer capacity. Here we examine the capacity of the buffer, which is employed to dynamically aggregate high-value samples from the online unsupervised data stream. As demonstrated in Fig. 4, our method achieves optimal results with a buffer capacity of 200, imposing only a negligible storage overhead. + +# 5 Conclusion + +In this paper, we revisit the key challenges in CTTA and address all of them simultaneously. To this end, we reshape the online data buffering and organizing mechanism, endowing our dynamic target buffer with the ability to aggregate reliable information in the unsupervised data stream. Through our designed class relation preservation constraint and a pseudo-target replay objective, the problem of catastrophic forgetting and error accumulation is significantly alleviated. Extensive experiments demonstrate the effectiveness of the proposed method. Regarding the adoption of uncertainty, our method employs the mainstream practice of sample entropy. In our framework, there is potential for investigating more diverse forms of uncertainty modeling, such as those based on energy functions. We leave this exploration to future work. + +# Acknowledgement + +This work was funded in part by the National Natural Science Foundation of China (62076195, 62376070, 62206271, U20B2052), in part by the Fundamental Research Funds for the Central Universities (AUGA5710011522), and by the Pengcheng Laboratory Research Project No. PCL2023A08. + +# References + +1. Boudiaf, M., Mueller, R., Ben Ayed, I., Bertinetto, L.: Parameter-free online test-time adaptation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8344-8353 (2022) +2. Brahma, D., Rai, P.: A probabilistic framework for lifelong test-time adaptation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 3582-3591 (June 2023) +3. Burges, C., Ragno, R., Le, Q.: Learning to rank with nonsmooth cost functions. Advances in neural information processing systems 19 (2006) +4. Chakrabarty, G., Sreenivas, M., Biswas, S.: SANTA: Source anchoring network and target alignment for continual test time adaptation. Transactions on Machine Learning Research (2023), https://openreview.net/forum?id=V7guVYzvE4 +5. Chen, D., Wang, D., Darrell, T., Ebrahimi, S.: Contrastive test-time adaptation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 295-305 (2022) +6. Chen, H., Wang, Y., Hu, Q.: Multi-granularity regularized re-balancing for class incremental learning. IEEE Transactions on Knowledge and Data Engineering 35(7), 7263-7277 (2022) +7. Choi, S., Yang, S., Choi, S., Yun, S.: Improving test-time adaptation via shift-agnostic weight regularization and nearest source prototypes. In: European Conference on Computer Vision. pp. 440-458. Springer (2022) +8. Cicek, S., Soatto, S.: Unsupervised domain adaptation via regularized conditional alignment. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 1416-1425 (2019) +9. Cordts, M., Omran, M., Ramos, S., Rehfeld, T., Enzweiler, M., Benenson, R., Franke, U., Roth, S., Schiele, B.: The cityscapes dataset for semantic urban scene understanding. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 3213-3223 (June 2016) +10. Croce, F., Andriushchenko, M., Sehwag, V., Debenedetti, E., Flammarion, N., Chiang, M., Mittal, P., Hein, M.: Robustbench: a standardized adversarial robustness benchmark. In: Thirty-fifth Conference on Neural Information Processing Systems Datasets and Benchmarks Track (Round 2) (2021) +11. Cui, S., Wang, S., Zhuo, J., Su, C., Huang, Q., Tian, Q.: Gradually vanishing bridge for adversarial domain adaptation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 12455-12464 (2020) +12. Döbler, M., Marsden, R.A., Yang, B.: Robust mean teacher for continual and gradual test-time adaptation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 7704-7714 (June 2023) +13. Dong, S., Hong, X., Tao, X., Chang, X., Wei, X., Gong, Y.: Few-shot class incremental learning via relation knowledge distillation. In: Proceedings of the AAAI Conference on Artificial Intelligence. pp. 1255-1263 (2021) + +14. Fan, Y., Wang, Y., Zhu, P., Hu, Q.: Dynamic sub-graph distillation for robust semi-supervised continual learning. In: Proceedings of the AAAI Conference on Artificial Intelligence. pp. 11927-11935 (2024) +15. Gan, Y., Bai, Y., Lou, Y., Ma, X., Zhang, R., Shi, N., Luo, L.: Decorate the newcomers: Visual domain prompt for continual test time adaptation. In: Proceedings of the AAAI Conference on Artificial Intelligence. pp. 7595-7603 (2023) +16. Ganin, Y., Lempitsky, V.: Unsupervised domain adaptation by backpropagation. In: International conference on machine learning. pp. 1180-1189. PMLR (2015) +17. Geirhos, R., Jacobsen, J.H., Michaelis, C., Zemel, R., Brendel, W., Bethge, M., Wichmann, F.A.: Shortcut learning in deep neural networks. Nature Machine Intelligence 2(11), 665-673 (2020) +18. Hendrycks, D., Dietterich, T.: Benchmarking neural network robustness to common corruptions and perturbations. In: International Conference on Learning Representations (2019), https://openreview.net/forum?id=HJz6tiCqYm +19. Hoffman, J., Tzeng, E., Park, T., Zhu, J.Y., Isola, P., Saenko, K., Efros, A., Darrell, T.: Cycada: Cycle-consistent adversarial domain adaptation. In: International conference on machine learning. pp. 1989-1998. Pmlr (2018) +20. Hwang, S., Lee, S., Kim, S., Ok, J., Kwak, S.: Combating label distribution shift for active domain adaptation. In: European Conference on Computer Vision. pp. 549-566. Springer (2022) +21. Koh, P.W., Sagawa, S., Marklund, H., Xie, S.M., Zhang, M., Balsubramani, A., Hu, W., Yasunaga, M., Phillips, R.L., Gao, I., et al.: Wilds: A benchmark of inthe-wild distribution shifts. In: International Conference on Machine Learning. pp. 5637-5664. PMLR (2021) +22. Lee, K., Kim, S., Kwak, S.: Cross-domain ensemble distillation for domain generalization. In: European Conference on Computer Vision. pp. 1-20. Springer (2022) +23. Li, H., Pan, S.J., Wang, S., Kot, A.C.: Domain generalization with adversarial feature learning. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 5400-5409 (2018) +24. Li, S., Xie, B., Lin, Q., Liu, C.H., Huang, G., Wang, G.: Generalized domain conditioned adaptation network. IEEE Transactions on Pattern Analysis and Machine Intelligence 44(8), 4093-4109 (2021) +25. Li, S., Xie, M., Gong, K., Liu, C.H., Wang, Y., Li, W.: Transferable semantic augmentation for domain adaptation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11516-11525 (2021) +26. Li, Y., Wang, N., Shi, J., Liu, J., Hou, X.: Revisiting batch normalization for practical domain adaptation. arXiv preprint arXiv:1603.04779 (2016) +27. Liang, J., Hu, D., Feng, J.: Do we really need to access the source data? source hypothesis transfer for unsupervised domain adaptation. In: International conference on machine learning. pp. 6028-6039. PMLR (2020) +28. Lin, H., Zhang, Y., Qiu, Z., Niu, S., Gan, C., Liu, Y., Tan, M.: Prototype-guided continual adaptation for class-incremental unsupervised domain adaptation. In: European Conference on Computer Vision. pp. 351-368. Springer (2022) +29. Liu, H., Long, M., Wang, J., Jordan, M.: Transferable adversarial training: A general approach to adapting deep classifiers. In: International Conference on Machine Learning. pp. 4013-4022. PMLR (2019) +30. Liu, J., Yang, S., Jia, P., Zhang, R., Lu, M., Guo, Y., Xue, W., Zhang, S.: ViDA: Homeostatic visual domain adapter for continual test time adaptation. In: The Twelfth International Conference on Learning Representations (2024), https://openreview.net/forum?id=sJ88Wg5Bp5 + +31. Liu, Y., Kothari, P., Van Delft, B., Bellot-Gurlet, B., Mordan, T., Alahi, A.: Ttt++: When does self-supervised test-time training fail or thrive? Advances in Neural Information Processing Systems 34, 21808-21820 (2021) +32. Long, M., Cao, Y., Wang, J., Jordan, M.: Learning transferable features with deep adaptation networks. In: International conference on machine learning. pp. 97-105. PMLR (2015) +33. Long, M., Zhu, H., Wang, J., Jordan, M.I.: Deep transfer learning with joint adaptation networks. In: International conference on machine learning. pp. 2208-2217. PMLR (2017) +34. Nguyen, A.T., Nguyen-Tang, T., Lim, S.N., Torr, P.H.: Tipi: Test time adaptation with transformation invariance. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 24162-24171 (2023) +35. Niloy, F.F., Ahmed, S.M., Raychaudhuri, D.S., Oymak, S., Roy-Chowdhury, A.K.: Effective restoration of source knowledge in continual test time adaptation. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 2091-2100 (2024) +36. Niu, S., Wu, J., Zhang, Y., Chen, Y., Zheng, S., Zhao, P., Tan, M.: Efficient test-time model adaptation without forgetting. In: International conference on machine learning. pp. 16888-16905. PMLR (2022) +37. Niu, S., Wu, J., Zhang, Y., Wen, Z., Chen, Y., Zhao, P., Tan, M.: Towards stable test-time adaptation in dynamic wild world. In: The Eleventh International Conference on Learning Representations (2023), https://openreview.net/forum?id=g2YraF75Tj +38. Patel, V.M., Gopalan, R., Li, R., Chellappa, R.: Visual domain adaptation: A survey of recent advances. IEEE signal processing magazine 32(3), 53-69 (2015) +39. Peng, X., Bai, Q., Xia, X., Huang, Z., Saenko, K., Wang, B.: Moment matching for multi-source domain adaptation. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 1406-1415 (2019) +40. Press, O., Schneider, S., Kümmerer, M., Bethge, M.: Rdumb: A simple approach that questions our progress in continual test-time adaptation. Advances in Neural Information Processing Systems 36 (2023) +41. Sakaridis, C., Dai, D., Van Gool, L.: Acdc: The adverse conditions dataset with correspondences for semantic driving scene understanding. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 10765-10775 (2021) +42. Schneider, S., Rusak, E., Eck, L., Bringmann, O., Brendel, W., Bethge, M.: Improving robustness against common corruptions by covariate shift adaptation. Advances in neural information processing systems 33, 11539-11551 (2020) +43. Sojka, D., Cygert, S., Twardowski, B., Trzciński, T.: Ar-tta: A simple method for real-world continual test-time adaptation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) Workshops. pp. 3491-3495 (October 2023) +44. Song, J., Lee, J., Kweon, I.S., Choi, S.: Ecotta: Memory-efficient continual test-time adaptation via self-distilled regularization. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11920-11929 (2023) +45. Tan, M., Chen, G., Wu, J., Zhang, Y., Chen, Y., Zhao, P., Niu, S.: Uncertainty-calibrated test-time model adaptation without forgetting. arXiv preprint arXiv:2403.11491 (2024) +46. Tao, X., Chang, X., Hong, X., Wei, X., Gong, Y.: Topology-preserving class-incremental learning. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XIX 16. pp. 254-270. Springer (2020) + +47. Tao, X., Hong, X., Chang, X., Dong, S., Wei, X., Gong, Y.: Few-shot class incremental learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (June 2020) +48. Tao, X., Hong, X., Chang, X., Gong, Y.: Bi-objective continual learning: Learning 'new' while consolidating 'known'. In: Proceedings of the AAAI Conference on Artificial Intelligence. pp. 5989-5996 (2020) +49. Tsai, Y.H., Hung, W.C., Schulter, S., Sohn, K., Yang, M.H., Chandraker, M.: Learning to adapt structured output space for semantic segmentation. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 7472-7481 (2018) +50. Wang, D., Shelhamer, E., Liu, S., Olshausen, B., Darrell, T.: Tent: Fully test-time adaptation by entropy minimization. In: International Conference on Learning Representations (2021), https://openreview.net/forum?id=uX13bZLkr3c +51. Wang, Q., Fink, O., Van Gool, L., Dai, D.: Continual test-time domain adaptation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 7201-7211 (June 2022) +52. Wang, S., Zhang, D., Yan, Z., Zhang, J., Li, R.: Feature alignment and uniformity for test time adaptation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 20050-20060 (2023) +53. Wang, Y., Ma, Z., Huang, Z., Wang, Y., Su, Z., Hong, X.: Isolation and impartial aggregation: A paradigm of incremental learning without interference. In: Proceedings of the AAAI Conference on Artificial Intelligence. pp. 10209-10217 (2023) +54. Wang, Y., Hong, J., Cheraghian, A., Rahman, S., Ahmedt-Aristizabal, D., Petersson, L., Harandi, M.: Continual test-time domain adaptation via dynamic sample selection. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV). pp. 1701-1710 (January 2024) +55. Wei, Y., Ye, J., Huang, Z., Zhang, J., Shan, H.: Online prototype learning for online continual learning. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 18764-18774 (October 2023) +56. Xie, E., Wang, W., Yu, Z., Anandkumar, A., Alvarez, J.M., Luo, P.: Segformer: Simple and efficient design for semantic segmentation with transformers. In: Beygelzimer, A., Dauphin, Y., Liang, P., Vaughan, J.W. (eds.) Advances in Neural Information Processing Systems (2021), https://openreview.net/forum?id=OG18MI5TRL +57. Xie, S., Girshick, R., Dollar, P., Tu, Z., He, K.: Aggregated residual transformations for deep neural networks. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1492-1500 (2017) +58. Yang, X., Gu, Y., Wei, K., Deng, C.: Exploring safety supervision for continual test-time domain adaptation. In: Proceedings of the Thirty-Second International Joint Conference on Artificial Intelligence, IJCAI-23. pp. 1649-1657. International Joint Conferences on Artificial Intelligence Organization (8 2023). https://doi.org/10.24963/ijcai.2023/183 +59. Yang, Y., Soatto, S.: Fda: Fourier domain adaptation for semantic segmentation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 4085-4095 (2020) +60. Yao, X., Wang, Y., Zhu, P., Lin, W., Li, J., Li, W., Hu, Q.: Socialized learning: Making each other better through multi-agent collaboration. In: Forty-first International Conference on Machine Learning (2024) +61. Yuan, L., Xie, B., Li, S.: Robust test-time adaptation in dynamic scenarios. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15922-15932 (2023) + +62. Zagoruyko, S., Komodakis, N.: Wide residual networks. In: British Machine Vision Conference 2016. British Machine Vision Association (2016) +63. Zhang, J., Qi, L., Shi, Y., Gao, Y.: Mvdg: A unified multi-view framework for domain generalization. In: European Conference on Computer Vision. pp. 161-177. Springer (2022) +64. Zhang, Y., Wang, Z., He, W.: Class relationship embedded learning for source-free unsupervised domain adaptation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7619-7629 (2023) +65. Zhang, Y., Wang, Z., Li, J., Zhuang, J., Lin, Z.: Towards effective instance discrimination contrastive loss for unsupervised domain adaptation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 11388-11399 (2023) +66. Zhou, K., Liu, Z., Qiao, Y., Xiang, T., Loy, C.C.: Domain generalization: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence (2022) \ No newline at end of file diff --git a/2024/Reshaping the Online Data Buffering and Organizing Mechanism for Continual Test-Time Adaptation/images.zip b/2024/Reshaping the Online Data Buffering and Organizing Mechanism for Continual Test-Time Adaptation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..d7ae7f735b0965f11b5f7f55a3a01c0b91cde695 --- /dev/null +++ b/2024/Reshaping the Online Data Buffering and Organizing Mechanism for Continual Test-Time Adaptation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a73572d2154b8498df6687d2fd61c4d22396b1890a688f4c5f60c742db937f3d +size 519866 diff --git a/2024/Reshaping the Online Data Buffering and Organizing Mechanism for Continual Test-Time Adaptation/layout.json b/2024/Reshaping the Online Data Buffering and Organizing Mechanism for Continual Test-Time Adaptation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..4728fef8648aa9223d228d0d60196f20686dfd12 --- /dev/null +++ b/2024/Reshaping the Online Data Buffering and Organizing Mechanism for Continual Test-Time Adaptation/layout.json @@ -0,0 +1,9661 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 140, + 111, + 474, + 166 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 140, + 111, + 474, + 166 + ], + "spans": [ + { + "bbox": [ + 140, + 111, + 474, + 166 + ], + "type": "text", + "content": "Reshaping the Online Data Buffering and Organizing Mechanism for Continual Test-Time Adaptation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 135, + 185, + 478, + 210 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 185, + 478, + 210 + ], + "spans": [ + { + "bbox": [ + 135, + 185, + 478, + 210 + ], + "type": "text", + "content": "Zhilin Zhu" + }, + { + "bbox": [ + 135, + 185, + 478, + 210 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 135, + 185, + 478, + 210 + ], + "type": "text", + "content": ", Xiaopeng Hong" + }, + { + "bbox": [ + 135, + 185, + 478, + 210 + ], + "type": "inline_equation", + "content": "^{1,2\\star}" + }, + { + "bbox": [ + 135, + 185, + 478, + 210 + ], + "type": "text", + "content": ", Zhiheng Ma" + }, + { + "bbox": [ + 135, + 185, + 478, + 210 + ], + "type": "inline_equation", + "content": "^{3,4,5}" + }, + { + "bbox": [ + 135, + 185, + 478, + 210 + ], + "type": "text", + "content": ", Weijun Zhuang" + }, + { + "bbox": [ + 135, + 185, + 478, + 210 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 135, + 185, + 478, + 210 + ], + "type": "text", + "content": ", Yaohui Ma" + }, + { + "bbox": [ + 135, + 185, + 478, + 210 + ], + "type": "inline_equation", + "content": "^{1,5}" + }, + { + "bbox": [ + 135, + 185, + 478, + 210 + ], + "type": "text", + "content": ", Yong Dai" + }, + { + "bbox": [ + 135, + 185, + 478, + 210 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 135, + 185, + 478, + 210 + ], + "type": "text", + "content": ", and Yaowei Wang" + }, + { + "bbox": [ + 135, + 185, + 478, + 210 + ], + "type": "inline_equation", + "content": "^{2,1}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 239, + 219, + 374, + 230 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 239, + 219, + 374, + 230 + ], + "spans": [ + { + "bbox": [ + 239, + 219, + 374, + 230 + ], + "type": "text", + "content": "1 Harbin Institute of Technology" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 256, + 231, + 357, + 242 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 256, + 231, + 357, + 242 + ], + "spans": [ + { + "bbox": [ + 256, + 231, + 357, + 242 + ], + "type": "text", + "content": "2 Pengcheng Laboratory" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 210, + 242, + 403, + 252 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 210, + 242, + 403, + 252 + ], + "spans": [ + { + "bbox": [ + 210, + 242, + 403, + 252 + ], + "type": "text", + "content": "3 Shenzhen University of Advanced Technology" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 157, + 253, + 456, + 263 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 157, + 253, + 456, + 263 + ], + "spans": [ + { + "bbox": [ + 157, + 253, + 456, + 263 + ], + "type": "text", + "content": "4 Guangdong Provincial Key Laboratory of Computility Microelectronics" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 157, + 264, + 469, + 297 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 157, + 264, + 469, + 297 + ], + "spans": [ + { + "bbox": [ + 157, + 264, + 469, + 297 + ], + "type": "text", + "content": "Shenzhen Institute of Advanced Technology, Chinese Academy of Science {zhuzhilin, weijunzhuang}@stu.hit.edu.cn hongxiaopeng@ieee.org ma@siat.ac.cn yh.ma@siat.ac.cn chd-dy@foxmail.com wangyw@pcl.ac.cn" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 159, + 317, + 453, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 317, + 453, + 472 + ], + "spans": [ + { + "bbox": [ + 159, + 317, + 453, + 472 + ], + "type": "text", + "content": "Abstract. Continual Test-Time Adaptation (CTTA) involves adapting a pre-trained source model to continually changing unsupervised target domains. In this paper, we systematically analyze the challenges of this task: online environment, unsupervised nature, and the risks of error accumulation and catastrophic forgetting under continual domain shifts. To address these challenges, we reshape the online data buffering and organizing mechanism for CTTA. We propose an uncertainty-aware buffering approach to identify and aggregate significant samples with high certainty from the unsupervised, single-pass data stream. Based on this, we propose a graph-based class relation preservation constraint to overcome catastrophic forgetting. Furthermore, a pseudo-target replay objective is used to mitigate error accumulation. Extensive experiments demonstrate the superiority of our method in both segmentation and classification CTTA tasks. Code is available at https://github.com/z1358/OBAO." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 159, + 480, + 453, + 503 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 480, + 453, + 503 + ], + "spans": [ + { + "bbox": [ + 159, + 480, + 453, + 503 + ], + "type": "text", + "content": "Keywords: Continual test-time adaptation " + }, + { + "bbox": [ + 159, + 480, + 453, + 503 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 480, + 453, + 503 + ], + "type": "text", + "content": " Unsupervised learning " + }, + { + "bbox": [ + 159, + 480, + 453, + 503 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 480, + 453, + 503 + ], + "type": "text", + "content": " Continual learning " + }, + { + "bbox": [ + 159, + 480, + 453, + 503 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 480, + 453, + 503 + ], + "type": "text", + "content": " Catastrophic forgetting" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 519, + 230, + 532 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 519, + 230, + 532 + ], + "spans": [ + { + "bbox": [ + 132, + 519, + 230, + 532 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 540, + 482, + 649 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 540, + 482, + 649 + ], + "spans": [ + { + "bbox": [ + 130, + 540, + 482, + 649 + ], + "type": "text", + "content": "Deep learning models have exhibited remarkable advancements in scenarios characterized by a consistent data distribution between training and test environments. However, the deployment of machine perception systems in the real world frequently exposes them to non-stationary and constantly changing environments [14, 21, 47, 53, 60]. When such phenomena occur (e.g., unexpected domain shifts over time), these models are prone to severe performance degradation [18, 37, 51]. Consequently, the development of techniques for continual adaptation during the unsupervised deployment phase is essential for enhancing the reliability of machine perception systems." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 653, + 234, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 653, + 234, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 653, + 234, + 665 + ], + "type": "text", + "content": "* Corresponding author" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 141, + 157, + 476, + 245 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "type": "text", + "content": "Table 1: Comparison of different adaptation settings. We differentiate along four dimensions: the available form of the source domain data and target domain data, the distribution and access mode of the target domain data stream during adaptation." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 141, + 157, + 476, + 245 + ], + "lines": [ + { + "bbox": [ + 141, + 157, + 476, + 245 + ], + "spans": [ + { + "bbox": [ + 141, + 157, + 476, + 245 + ], + "type": "table", + "html": "
SettingDataTarget domain data stream
SourceTargetDistributionAccess mode
Fine-tuning-xT,yTstationaryoffline (multi-epoch)
Domain GeneralizationxS,yS---
Domain AdaptationxS,ySxTstationaryoffline (multi-epoch)
Test-Time TrainingxS,ySxTstationaryonline (single-pass)
Test-Time Adaptation-xTstationaryoffline or online
Continual Test-Time Adaptation-x1T→···→xnTcontinually changingonline (single-pass)
", + "image_path": "f693bf9a99303dbedc915dd5f53517af4c9c315f3edfb13b6ea0fa886d1055c9.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 269, + 482, + 400 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 269, + 482, + 400 + ], + "spans": [ + { + "bbox": [ + 130, + 269, + 482, + 400 + ], + "type": "text", + "content": "With sufficient source labeled data accessible, the fields of domain generalization [22, 23, 63, 66] and domain adaptation [20, 24, 25, 28] have witnessed considerable success. However, a particularly challenging scenario arises when no source data is available, and unsupervised adaptation must happen at test time when unlabeled test inputs arrive. This task is known as test-time adaptation (TTA) [5, 7, 45]. Given the rarity of encountering only a singular domain shift in practical applications, the community has recently been increasingly focusing on continual test-time adaptation (CTTA) [2, 12, 30, 51, 58], i.e., evaluating the ability of methods to adapt to continual domain shifts. We summarize the differences between existing adaptation settings in Tab. 1. Obviously, the considered CTTA setting is the most practical and challenging." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 401, + 482, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 401, + 482, + 557 + ], + "spans": [ + { + "bbox": [ + 130, + 401, + 482, + 557 + ], + "type": "text", + "content": "Within this setting, we expose the key challenges as follows: (1) Online environment. This denotes a scenario where data is processed in a single pass, necessitating immediate predictions from the model. Due to the characteristics of the online environment, it is difficult for the model to obtain a sufficient number of informative samples, thereby impeding the efficacy of adaptation; (2) Unsupervised nature. Throughout adaptation, ground-truth labels for input data are unavailable. Moreover, due to privacy concerns or legal restrictions, the source domain data may no longer be available and we only have access to the source pre-trained model; (3) Error accumulation and catastrophic forgetting. The dynamic distribution shift when adapting to continually changing target domains can render pseudo-labels unreliable and impede knowledge retention about the source domain, leading to error accumulation [51, 54] and triggering catastrophic forgetting [6, 30, 58]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "text", + "content": "Addressing the first challenge, an intuitive solution is to employ a buffer to store high-value samples. However, existing methods that utilize such a buffer store source domain samples with ground-truth labels [12, 43], which is difficult to satisfy in real-world scenarios and conflicts with the source-free constraint in the second challenge. Therefore, the issue of effectively buffering and organizing high-value samples in an unsupervised single-pass data stream has not been fully explored. Current CTTA methods primarily address the third challenge, focusing mainly on alignments in either the parameter space [2, 35, 40, 51] or the feature space [4, 12] of the source model to mitigate knowledge forgetting." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 219, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 219, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 219, + 101 + ], + "type": "text", + "content": "Z. Zhu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 147, + 116, + 239, + 167 + ], + "blocks": [ + { + "bbox": [ + 147, + 116, + 239, + 167 + ], + "lines": [ + { + "bbox": [ + 147, + 116, + 239, + 167 + ], + "spans": [ + { + "bbox": [ + 147, + 116, + 239, + 167 + ], + "type": "image", + "image_path": "fad5669ba277ebfa7db9e8f59138252a4fb8af44d55f35bf7fdc27916c238c29.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 163, + 187, + 223, + 197 + ], + "lines": [ + { + "bbox": [ + 163, + 187, + 223, + 197 + ], + "spans": [ + { + "bbox": [ + 163, + 187, + 223, + 197 + ], + "type": "text", + "content": "(a) W/o a buffer" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 260, + 116, + 353, + 180 + ], + "blocks": [ + { + "bbox": [ + 260, + 116, + 353, + 180 + ], + "lines": [ + { + "bbox": [ + 260, + 116, + 353, + 180 + ], + "spans": [ + { + "bbox": [ + 260, + 116, + 353, + 180 + ], + "type": "image", + "image_path": "70dbd4dbd6ff3daec0498c281c295a40103b6c089478258ce7fb34141148fdeb.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 265, + 188, + 348, + 197 + ], + "lines": [ + { + "bbox": [ + 265, + 188, + 348, + 197 + ], + "spans": [ + { + "bbox": [ + 265, + 188, + 348, + 197 + ], + "type": "text", + "content": "(b) Static source buffer" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 369, + 116, + 472, + 186 + ], + "blocks": [ + { + "bbox": [ + 369, + 116, + 472, + 186 + ], + "lines": [ + { + "bbox": [ + 369, + 116, + 472, + 186 + ], + "spans": [ + { + "bbox": [ + 369, + 116, + 472, + 186 + ], + "type": "image", + "image_path": "2f4e1d8c73e641ba711b0b528813684ae73945523f6b4b69581c6490f3e639c7.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 367, + 188, + 474, + 197 + ], + "lines": [ + { + "bbox": [ + 367, + 188, + 474, + 197 + ], + "spans": [ + { + "bbox": [ + 367, + 188, + 474, + 197 + ], + "type": "text", + "content": "(c) Our dynamic target buffer" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 206, + 480, + 262 + ], + "lines": [ + { + "bbox": [ + 130, + 206, + 480, + 262 + ], + "spans": [ + { + "bbox": [ + 130, + 206, + 480, + 262 + ], + "type": "text", + "content": "Fig. 1: Comparison of different buffering mechanisms. (a) Without a buffer, only unreliable signals from the current unsupervised input data are available. (b) Some methods [12, 43] use a static source buffer that stores source samples with ground-truth labels to provide supervised signals. (c) We work on a dynamic target buffer, which can provide reliable signals for the adaptation process." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 289, + 482, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 289, + 482, + 506 + ], + "spans": [ + { + "bbox": [ + 130, + 289, + 482, + 506 + ], + "type": "text", + "content": "In this paper, we endeavor to address these three challenges simultaneously. To this end, we reshape the online data buffering and organizing mechanism for CTTA. As shown in Fig. 1, unlike previous methods which directly sample from the source domain [12, 43], we are targeted at source-free, unsupervised, and online buffering. Through a specially designed uncertainty-aware buffering approach, we dynamically evaluate and aggregate reliable information in the unsupervised, single-pass data stream (Challenges 1 and 2). Based on this, we efficiently organize these samples to assist the model's unsupervised adaptation process. Utilizing pseudo-labels generated from these low-uncertainty samples, we consider them as anchors of classes and organize them into a class relation graph (CRG). To combat catastrophic forgetting under continually changing domains, we introduce a novel graph-based class relation preservation constraint to maintain the intrinsic class relations obtained on the source domain (Challenge 3). Specifically, we penalize topological changes within the CRG to mitigate knowledge forgetting, while allowing reasonable vertex movements to enhance adaptability. Additionally, these samples are incorporated into a pseudo-target replay objective to reduce the risk of error accumulation from misclassified samples (Challenge 3). In summary, our main contributions are as follows:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 137, + 526, + 480, + 664 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 138, + 526, + 479, + 550 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 526, + 479, + 550 + ], + "spans": [ + { + "bbox": [ + 138, + 526, + 479, + 550 + ], + "type": "text", + "content": "- To address the three revealed challenges simultaneously, we reshape the online data buffering and organizing mechanism for CTTA." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 553, + 480, + 576 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 553, + 480, + 576 + ], + "spans": [ + { + "bbox": [ + 138, + 553, + 480, + 576 + ], + "type": "text", + "content": "- To frame the buffering mechanism, we design an uncertainty-aware approach to identify and aggregate reliable data from unsupervised online data stream." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 579, + 479, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 579, + 479, + 613 + ], + "spans": [ + { + "bbox": [ + 138, + 579, + 479, + 613 + ], + "type": "text", + "content": "- Through a well-designed graph-based class relation preservation constraint and a pseudo-target replay objective, we effectively mitigate catastrophic forgetting and error accumulation in CTTA." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 137, + 617, + 479, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 617, + 479, + 664 + ], + "spans": [ + { + "bbox": [ + 137, + 617, + 479, + 664 + ], + "type": "text", + "content": "- We extensively evaluate our proposed method and the results on four challenging datasets show that the method clearly improves the CTTA performance on both segmentation and classification benchmarks, indicating its effectiveness and universality." + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 138, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 138, + 91, + 447, + 102 + ], + "type": "text", + "content": "Reshaping the Online Data Buffering and Organizing Mechanism for CTTA" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 238, + 127 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 238, + 127 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 238, + 127 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 144, + 261, + 156 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 144, + 261, + 156 + ], + "spans": [ + { + "bbox": [ + 132, + 144, + 261, + 156 + ], + "type": "text", + "content": "2.1 Domain Adaptation" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 167, + 483, + 263 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 167, + 483, + 263 + ], + "spans": [ + { + "bbox": [ + 132, + 167, + 483, + 263 + ], + "type": "text", + "content": "Domain adaptation (DA) [38,65] aims to generalize a model from a well-annotated source domain to an unlabeled target domain following the transductive learning principle. The underlying assumption in DA is that both labeled source domain and unlabeled target domain data are concurrently accessible. Consequently, the prevalent methods in DA encompass the utilization of domain adversarial training [8,11,16,19,29,49] or the minimization of domain discrepancy [32,33,39]. These methods work towards harmonizing the feature distribution [16,32,49] or the input space [19,59] between the two domains." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 285, + 324, + 297 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 285, + 324, + 297 + ], + "spans": [ + { + "bbox": [ + 132, + 285, + 324, + 297 + ], + "type": "text", + "content": "2.2 Continual Test-Time Adaptation" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 308, + 482, + 452 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 308, + 482, + 452 + ], + "spans": [ + { + "bbox": [ + 132, + 308, + 482, + 452 + ], + "type": "text", + "content": "Test-Time Adaptation (TTA). TTA focuses on a more challenging setting after model deployment where only a pre-trained source model and unlabeled target data are available. Some approaches [5, 27] focus on the offline setting, where multiple epochs of access to the entire set of test data are provided for adaptation, while others [1, 7, 31, 52] focus on the online setting with a single-pass data stream. To name a few, TENT [50] adapts the model by using entropy minimization loss to update a few trainable parameters in batch normalization layers. This approach has also inspired subsequent works such as EATA [36] and RoTTA [61] to investigate the robustness of normalized layers. Due to the domain gap, TSD [52] promotes feature alignment and uniformity from the perspective of feature revision, while TIPI [34] proposes invariance regularizers to find the transformation space that can simulate the domain shift." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 473, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 473, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 473, + 482, + 665 + ], + "type": "text", + "content": "Continual Test-Time Adaptation (CTTA). In recent years, CTTA has received increasing research attention [15,44,51,58] due to its practicality. CoTTA [51] stands out as the first method specifically tailored to address this task. It utilizes weight-averaged and augmentation-averaged predictions to reduce error accumulation, while stochastically restoring the weights of a small fraction of neurons to the source model during each iteration to mitigate catastrophic forgetting. Building upon this, PETAL [2] further introduces a data-driven parameter recovery technique to align with the source model in the parameter space, while RMT [12] and SANTA [4] seek alignment with the source model in the feature space by optimizing an additional contrastive learning objective. Additionally, DSS [54] considers sample security to mitigate the risk of using incorrect information. Some approaches [15, 30, 44] also prevent forgetting by freezing a majority or all of the model parameters, but they require optimization along with the model during training using the source domain data, thereby relaxing the unsupervised challenge and not being truly source-free. Under our proposal, we simultaneously address three challenges in CTTA." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 141, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 141, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 141, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 219, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 219, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 219, + 101 + ], + "type": "text", + "content": "Z. Zhu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 157, + 117, + 455, + 284 + ], + "blocks": [ + { + "bbox": [ + 157, + 117, + 455, + 284 + ], + "lines": [ + { + "bbox": [ + 157, + 117, + 455, + 284 + ], + "spans": [ + { + "bbox": [ + 157, + 117, + 455, + 284 + ], + "type": "image", + "image_path": "c625ea5d503b5153c2d15d12f3ebce3d9c91ec6f5d8c0cc585af3169df2bf172.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 294, + 482, + 361 + ], + "lines": [ + { + "bbox": [ + 130, + 294, + 482, + 361 + ], + "spans": [ + { + "bbox": [ + 130, + 294, + 482, + 361 + ], + "type": "text", + "content": "Fig. 2: The framework of our proposed method. At time step " + }, + { + "bbox": [ + 130, + 294, + 482, + 361 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 294, + 482, + 361 + ], + "type": "text", + "content": ", the incoming data and sampled buffer data are used for adaptation. The proposed class relation preservation loss " + }, + { + "bbox": [ + 130, + 294, + 482, + 361 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{CRP}" + }, + { + "bbox": [ + 130, + 294, + 482, + 361 + ], + "type": "text", + "content": " ensures topological consistency between the Class Relation Graph (CRG) " + }, + { + "bbox": [ + 130, + 294, + 482, + 361 + ], + "type": "inline_equation", + "content": "G_{t}" + }, + { + "bbox": [ + 130, + 294, + 482, + 361 + ], + "type": "text", + "content": " estimated at the current time step " + }, + { + "bbox": [ + 130, + 294, + 482, + 361 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 294, + 482, + 361 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 294, + 482, + 361 + ], + "type": "inline_equation", + "content": "\\hat{G}" + }, + { + "bbox": [ + 130, + 294, + 482, + 361 + ], + "type": "text", + "content": " from the source domain, thereby effectively preserving intrinsic semantic information. Concurrently, we integrate a pseudo-target replay loss " + }, + { + "bbox": [ + 130, + 294, + 482, + 361 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{PCE}" + }, + { + "bbox": [ + 130, + 294, + 482, + 361 + ], + "type": "text", + "content": " to mitigate error accumulation." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 382, + 233, + 396 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 382, + 233, + 396 + ], + "spans": [ + { + "bbox": [ + 132, + 382, + 233, + 396 + ], + "type": "text", + "content": "3 Methodology" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 407, + 350, + 418 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 407, + 350, + 418 + ], + "spans": [ + { + "bbox": [ + 132, + 407, + 350, + 418 + ], + "type": "text", + "content": "3.1 Preliminaries and Overall Framework" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 426, + 482, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 426, + 482, + 545 + ], + "spans": [ + { + "bbox": [ + 130, + 426, + 482, + 545 + ], + "type": "text", + "content": "In continual test-time adaptation (CTTA), we are given a source model " + }, + { + "bbox": [ + 130, + 426, + 482, + 545 + ], + "type": "inline_equation", + "content": "f_{\\theta_0}" + }, + { + "bbox": [ + 130, + 426, + 482, + 545 + ], + "type": "text", + "content": " pretrained from the labeled source domain " + }, + { + "bbox": [ + 130, + 426, + 482, + 545 + ], + "type": "inline_equation", + "content": "(\\mathcal{X}^S, \\mathcal{Y}^S)" + }, + { + "bbox": [ + 130, + 426, + 482, + 545 + ], + "type": "text", + "content": " and a series of sequentially changing unlabeled target domains such as " + }, + { + "bbox": [ + 130, + 426, + 482, + 545 + ], + "type": "inline_equation", + "content": "[\\mathcal{X}_1^T, \\mathcal{X}_2^T, \\dots, \\mathcal{X}_n^T]" + }, + { + "bbox": [ + 130, + 426, + 482, + 545 + ], + "type": "text", + "content": ". Each target domain shares the same label space as the source domain (for close-set adaptation). Note that the model has no information about when the domain changes happen during the adaptation process. The goal of CTTA is to adapt the model to a continually changing unsupervised online data stream of target domains without source data. To be more specific, at inference time step " + }, + { + "bbox": [ + 130, + 426, + 482, + 545 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 426, + 482, + 545 + ], + "type": "text", + "content": ", the unlabeled target test data " + }, + { + "bbox": [ + 130, + 426, + 482, + 545 + ], + "type": "inline_equation", + "content": "x(t)" + }, + { + "bbox": [ + 130, + 426, + 482, + 545 + ], + "type": "text", + "content": " is given to the model " + }, + { + "bbox": [ + 130, + 426, + 482, + 545 + ], + "type": "inline_equation", + "content": "f_{\\theta_t}" + }, + { + "bbox": [ + 130, + 426, + 482, + 545 + ], + "type": "text", + "content": ", and the model needs to make predictions and adapt itself accordingly " + }, + { + "bbox": [ + 130, + 426, + 482, + 545 + ], + "type": "inline_equation", + "content": "(f_{\\theta_t} \\rightarrow f_{\\theta_{t+1}})" + }, + { + "bbox": [ + 130, + 426, + 482, + 545 + ], + "type": "text", + "content": " for the upcoming test inputs " + }, + { + "bbox": [ + 130, + 426, + 482, + 545 + ], + "type": "inline_equation", + "content": "x(t+1)" + }, + { + "bbox": [ + 130, + 426, + 482, + 545 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": "We adopt a teacher-student framework following previous works [2,12,51] as illustrated in Fig. 2. Leveraging the proposed uncertainty-aware buffering mechanism, our method dynamically evaluates and buffers significant high-certainty samples in the unsupervised online data stream, providing reliable information for the model's adaptation process. Building upon this, we further introduce a novel class relation preservation constraint " + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "inline_equation", + "content": "(\\mathcal{L}_{CRP})" + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": " for preserving domain-invariant semantic information in dynamic environments, effectively mitigating the catastrophic forgetting problem. Additionally, we address error accumulation through a pseudo-target replay objective " + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "inline_equation", + "content": "(\\mathcal{L}_{PCE})" + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": " to minimize the impact of miscalibration. A self-training loss " + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "inline_equation", + "content": "(\\mathcal{L}_{ST})" + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": " is used to stabilize the adaptation" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 138, + 90, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 90, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 138, + 90, + 447, + 102 + ], + "type": "text", + "content": "Reshaping the Online Data Buffering and Organizing Mechanism for CTTA" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 175 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 175 + ], + "type": "text", + "content": "process of the student model (adapted model). In the adopted framework, both the teacher model " + }, + { + "bbox": [ + 130, + 116, + 482, + 175 + ], + "type": "inline_equation", + "content": "f_{\\theta'}" + }, + { + "bbox": [ + 130, + 116, + 482, + 175 + ], + "type": "text", + "content": " and the student model " + }, + { + "bbox": [ + 130, + 116, + 482, + 175 + ], + "type": "inline_equation", + "content": "f_{\\theta}" + }, + { + "bbox": [ + 130, + 116, + 482, + 175 + ], + "type": "text", + "content": " are initialized with the source model " + }, + { + "bbox": [ + 130, + 116, + 482, + 175 + ], + "type": "inline_equation", + "content": "f_{\\theta_0}" + }, + { + "bbox": [ + 130, + 116, + 482, + 175 + ], + "type": "text", + "content": ". Subsequently, the teacher model is continuously updated by the exponential moving average of the student model. We elaborate our method in the following subsections." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 193, + 304, + 205 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 193, + 304, + 205 + ], + "spans": [ + { + "bbox": [ + 132, + 193, + 304, + 205 + ], + "type": "text", + "content": "3.2 Uncertainty-aware Buffering" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 212, + 482, + 295 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 212, + 482, + 295 + ], + "spans": [ + { + "bbox": [ + 130, + 212, + 482, + 295 + ], + "type": "text", + "content": "In the unsupervised online data stream within CTTA, the model necessitates adaptation utilizing the data at hand, typically characterized by a scarcity of information. While certain methods [12, 43] endeavor to incorporate additional supervised information by formulating a static buffer from the source domain, fulfilling this requirement proves challenging within practical application scenarios. Conversely, our proposition involves sampling and buffering throughout the unsupervised adaptation process by a designed uncertainty-aware approach." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 296, + 483, + 367 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 296, + 483, + 367 + ], + "spans": [ + { + "bbox": [ + 130, + 296, + 483, + 367 + ], + "type": "text", + "content": "At the outset, the dynamic buffer is initialized as empty. As the adaptation progresses, the buffer undergoes continual updates to reflect the evolving nature of the data stream encountered. Specifically, at any given inference time step " + }, + { + "bbox": [ + 130, + 296, + 483, + 367 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 296, + 483, + 367 + ], + "type": "text", + "content": ", the model calculates the uncertainty associated with the current batch of samples. The quantification of sample uncertainty manifests as the entropy of the sample, denoted as:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 227, + 375, + 481, + 389 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 227, + 375, + 481, + 389 + ], + "spans": [ + { + "bbox": [ + 227, + 375, + 481, + 389 + ], + "type": "interline_equation", + "content": "H (x (t)) = - f _ {\\theta_ {t}} (x (t)) ^ {\\top} (\\log f _ {\\theta_ {t}} (x (t))), \\tag {1}", + "image_path": "e9d407416ba0fd79a90bd679c473a796e34389a3dc2c91e2a50ade3467b9454f.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 396, + 482, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 396, + 482, + 445 + ], + "spans": [ + { + "bbox": [ + 130, + 396, + 482, + 445 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 396, + 482, + 445 + ], + "type": "inline_equation", + "content": "f_{\\theta_t}(x(t))" + }, + { + "bbox": [ + 130, + 396, + 482, + 445 + ], + "type": "text", + "content": " corresponds the softmax predictions of the student model for the current input data. Samples exhibiting an entropy value less than a predefined threshold " + }, + { + "bbox": [ + 130, + 396, + 482, + 445 + ], + "type": "inline_equation", + "content": "H_0" + }, + { + "bbox": [ + 130, + 396, + 482, + 445 + ], + "type": "text", + "content": " are deemed to possess a high degree of certainty and are thus considered candidates for inclusion in the dynamic buffer. " + }, + { + "bbox": [ + 130, + 396, + 482, + 445 + ], + "type": "inline_equation", + "content": "H_0" + }, + { + "bbox": [ + 130, + 396, + 482, + 445 + ], + "type": "text", + "content": " is defined as:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 280, + 454, + 481, + 466 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 280, + 454, + 481, + 466 + ], + "spans": [ + { + "bbox": [ + 280, + 454, + 481, + 466 + ], + "type": "interline_equation", + "content": "H _ {0} = \\alpha \\ln C, \\tag {2}", + "image_path": "4171818d1811a58359e62a8c0dd815160df782e74dbb019604e7debdc8305447.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 474, + 482, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 474, + 482, + 581 + ], + "spans": [ + { + "bbox": [ + 130, + 474, + 482, + 581 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 474, + 482, + 581 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 130, + 474, + 482, + 581 + ], + "type": "text", + "content": " serves as a threshold coefficient and " + }, + { + "bbox": [ + 130, + 474, + 482, + 581 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 130, + 474, + 482, + 581 + ], + "type": "text", + "content": " denotes the total number of classes in the dataset. The process of storing samples into the buffer, or updating existing entries, is contingent upon the buffer's capacity. High-certainty samples identified from the present batch, along with their derived pseudo-labels from the teacher model, are added to the buffer if space permits. Conversely, in scenarios where the buffer is at capacity, these new samples replace the ones exhibiting the highest uncertainty to maintain an optimal blend of timeliness and reliability. Upon completion of these updates, the resulting buffer " + }, + { + "bbox": [ + 130, + 474, + 482, + 581 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_t" + }, + { + "bbox": [ + 130, + 474, + 482, + 581 + ], + "type": "text", + "content": " stands ready to facilitate the model's adaptation in subsequent iterations." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 582, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 482, + 665 + ], + "type": "text", + "content": "Notably, our approach to buffering is entirely devoid of reliance on source data, underscoring its source-free nature. This strategy enables the model to discern and prioritize data based on their uncertainty. By dynamically buffering high-certainty samples within the data stream, our approach not only circumvents the limitations posed by reliance on static source buffers but also enriches the adaptation process with valuable, reliable data. The intuitive idea is to utilize CE loss to replay samples in the buffer, thereby emphasizing reliable data" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 219, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 219, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 219, + 101 + ], + "type": "text", + "content": "Z. Zhu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 164 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 164 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 164 + ], + "type": "text", + "content": "and effectively suppressing the risk and impact of miscalibration due to error accumulation. In each test-time adaptation iteration, we randomly sample examples and labels " + }, + { + "bbox": [ + 130, + 116, + 479, + 164 + ], + "type": "inline_equation", + "content": "(\\tilde{x}(t), \\tilde{y}(t))" + }, + { + "bbox": [ + 130, + 116, + 479, + 164 + ], + "type": "text", + "content": " from the buffer. The number of sampled examples is equal to the batch size and the corresponding pseudo-target replay loss is:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 251, + 175, + 481, + 206 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 251, + 175, + 481, + 206 + ], + "spans": [ + { + "bbox": [ + 251, + 175, + 481, + 206 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {P C E} = - \\sum_ {c = 1} ^ {C} \\tilde {y} _ {c} (t) \\log \\tilde {p} _ {c}, \\tag {3}", + "image_path": "da8efb25c9e4bea5ce30b69ceccbb37fa20a5b90886864dbc002ad8e178e8a2b.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 216, + 480, + 276 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 216, + 480, + 276 + ], + "spans": [ + { + "bbox": [ + 131, + 216, + 480, + 276 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 216, + 480, + 276 + ], + "type": "inline_equation", + "content": "\\tilde{p}" + }, + { + "bbox": [ + 131, + 216, + 480, + 276 + ], + "type": "text", + "content": " is the softmax prediction of the student model for " + }, + { + "bbox": [ + 131, + 216, + 480, + 276 + ], + "type": "inline_equation", + "content": "\\tilde{x}(t)" + }, + { + "bbox": [ + 131, + 216, + 480, + 276 + ], + "type": "text", + "content": ". For simplicity, we omit the summation of all samples in the batch. Moreover, these sampled examples also serve as anchors and are organized in a class relation graph. On this basis, we introduce a novel class relation preservation constraint in Sec. 3.3 to preserve domain-invariant knowledge and overcome forgetting." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 295, + 357, + 306 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 295, + 357, + 306 + ], + "spans": [ + { + "bbox": [ + 132, + 295, + 357, + 306 + ], + "type": "text", + "content": "3.3 Class Relation Preservation Constraint" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 316, + 480, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 316, + 480, + 411 + ], + "spans": [ + { + "bbox": [ + 130, + 316, + 480, + 411 + ], + "type": "text", + "content": "In online scenarios where domains are constantly changing, neural networks tend to \"take shortcuts\" and focus on simplistic features [17, 55] from the single-pass data stream. This phenomenon results in the model's forgetting of intrinsic semantic information. In this work, we propose to overcome catastrophic forgetting by preserving the domain-invariant class relation, which is consistent across domains regardless of distribution discrepancy [64]. Therefore, it is reasonable to utilize the intrinsic class relation obtained on the source domain with large labeled samples to constrain the model adaptation process on target domains." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 413, + 481, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 413, + 481, + 544 + ], + "spans": [ + { + "bbox": [ + 130, + 413, + 481, + 544 + ], + "type": "text", + "content": "We describe here how to obtain the defined class relation graph (CRG). For computational stability, we normalize the feature space and adopt the cosine similarity metric. Let " + }, + { + "bbox": [ + 130, + 413, + 481, + 544 + ], + "type": "inline_equation", + "content": "\\bar{\\cdot}" + }, + { + "bbox": [ + 130, + 413, + 481, + 544 + ], + "type": "text", + "content": " denotes the normalization operation, where " + }, + { + "bbox": [ + 130, + 413, + 481, + 544 + ], + "type": "inline_equation", + "content": "\\bar{f} = f / \\| f\\|" + }, + { + "bbox": [ + 130, + 413, + 481, + 544 + ], + "type": "text", + "content": ". Given the normalized feature space " + }, + { + "bbox": [ + 130, + 413, + 481, + 544 + ], + "type": "inline_equation", + "content": "\\bar{\\mathcal{F}}" + }, + { + "bbox": [ + 130, + 413, + 481, + 544 + ], + "type": "text", + "content": ", the CRG " + }, + { + "bbox": [ + 130, + 413, + 481, + 544 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 130, + 413, + 481, + 544 + ], + "type": "text", + "content": " is defined as " + }, + { + "bbox": [ + 130, + 413, + 481, + 544 + ], + "type": "inline_equation", + "content": "G = \\langle V,E\\rangle" + }, + { + "bbox": [ + 130, + 413, + 481, + 544 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 413, + 481, + 544 + ], + "type": "inline_equation", + "content": "V = \\{\\bar{v}_1,\\dots ,\\bar{v}_C|\\bar{v}_i\\in \\bar{\\mathcal{F}}\\}" + }, + { + "bbox": [ + 130, + 413, + 481, + 544 + ], + "type": "text", + "content": " is the set of " + }, + { + "bbox": [ + 130, + 413, + 481, + 544 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 130, + 413, + 481, + 544 + ], + "type": "text", + "content": " vertices representative of " + }, + { + "bbox": [ + 130, + 413, + 481, + 544 + ], + "type": "inline_equation", + "content": "\\bar{\\mathcal{F}}" + }, + { + "bbox": [ + 130, + 413, + 481, + 544 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 413, + 481, + 544 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 130, + 413, + 481, + 544 + ], + "type": "text", + "content": " is the edge set describing the neighborhood relations of vertices in " + }, + { + "bbox": [ + 130, + 413, + 481, + 544 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 130, + 413, + 481, + 544 + ], + "type": "text", + "content": ". Each edge " + }, + { + "bbox": [ + 130, + 413, + 481, + 544 + ], + "type": "inline_equation", + "content": "e_{ij}" + }, + { + "bbox": [ + 130, + 413, + 481, + 544 + ], + "type": "text", + "content": " is assigned with a weight " + }, + { + "bbox": [ + 130, + 413, + 481, + 544 + ], + "type": "inline_equation", + "content": "s_{ij}" + }, + { + "bbox": [ + 130, + 413, + 481, + 544 + ], + "type": "text", + "content": ", which is the similarity between " + }, + { + "bbox": [ + 130, + 413, + 481, + 544 + ], + "type": "inline_equation", + "content": "\\bar{v}_i" + }, + { + "bbox": [ + 130, + 413, + 481, + 544 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 413, + 481, + 544 + ], + "type": "inline_equation", + "content": "\\bar{v}_j" + }, + { + "bbox": [ + 130, + 413, + 481, + 544 + ], + "type": "text", + "content": ": " + }, + { + "bbox": [ + 130, + 413, + 481, + 544 + ], + "type": "inline_equation", + "content": "s_{ij} = \\bar{v}_i^\\top \\bar{v}_j" + }, + { + "bbox": [ + 130, + 413, + 481, + 544 + ], + "type": "text", + "content": ". Each vertex " + }, + { + "bbox": [ + 130, + 413, + 481, + 544 + ], + "type": "inline_equation", + "content": "\\bar{v}_i" + }, + { + "bbox": [ + 130, + 413, + 481, + 544 + ], + "type": "text", + "content": " is the centroid vector of class " + }, + { + "bbox": [ + 130, + 413, + 481, + 544 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 413, + 481, + 544 + ], + "type": "text", + "content": ". The intrinsic CRG " + }, + { + "bbox": [ + 130, + 413, + 481, + 544 + ], + "type": "inline_equation", + "content": "\\hat{G}" + }, + { + "bbox": [ + 130, + 413, + 481, + 544 + ], + "type": "text", + "content": " is obtained by modeling the source-domain class relation. We demonstrate in experiments that the construction of our CRG " + }, + { + "bbox": [ + 130, + 413, + 481, + 544 + ], + "type": "inline_equation", + "content": "\\hat{G}" + }, + { + "bbox": [ + 130, + 413, + 481, + 544 + ], + "type": "text", + "content": " can be derived from different data conditions, including class prototypes and classifier weights." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 545, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 481, + 666 + ], + "type": "text", + "content": "At time step " + }, + { + "bbox": [ + 130, + 545, + 481, + 666 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 545, + 481, + 666 + ], + "type": "text", + "content": ", given CRG " + }, + { + "bbox": [ + 130, + 545, + 481, + 666 + ], + "type": "inline_equation", + "content": "G_{t} = \\langle V_{t}, E_{t} \\rangle" + }, + { + "bbox": [ + 130, + 545, + 481, + 666 + ], + "type": "text", + "content": ", when catastrophic forgetting occurs, " + }, + { + "bbox": [ + 130, + 545, + 481, + 666 + ], + "type": "inline_equation", + "content": "G_{t}" + }, + { + "bbox": [ + 130, + 545, + 481, + 666 + ], + "type": "text", + "content": " is distorted as the semantic relations among categories are confused. To mitigate forgetting, the model should be adapted to strive towards preserving the topology in " + }, + { + "bbox": [ + 130, + 545, + 481, + 666 + ], + "type": "inline_equation", + "content": "G_{t}" + }, + { + "bbox": [ + 130, + 545, + 481, + 666 + ], + "type": "text", + "content": " the same as that in " + }, + { + "bbox": [ + 130, + 545, + 481, + 666 + ], + "type": "inline_equation", + "content": "\\hat{G}" + }, + { + "bbox": [ + 130, + 545, + 481, + 666 + ], + "type": "text", + "content": ". This is achieved by restricting the neighboring relations of vertices described by the weights of edges (i.e., similarity between classes) in " + }, + { + "bbox": [ + 130, + 545, + 481, + 666 + ], + "type": "inline_equation", + "content": "E_{t}" + }, + { + "bbox": [ + 130, + 545, + 481, + 666 + ], + "type": "text", + "content": ". One way to do this is to keep the ranking of the edge weights constant during the adaptation process. However optimizing a non-smooth global ranking is difficult and inefficient [3,46]. An alternative and more feasible approach involves assessing the extent of change in neighboring relations [13,48] by evaluating the correlation between the initial edge weights" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 138, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 138, + 91, + 447, + 102 + ], + "type": "text", + "content": "Reshaping the Online Data Buffering and Organizing Mechanism for CTTA" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 163 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 163 + ], + "type": "text", + "content": "and their current observation. A lower correlation indicates a higher probability that the rank of an edge has changed during adaptation, and this should be penalized. On this basis, we define the Class Relation Preservation constraint (CRP) as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 238, + 173, + 481, + 212 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 238, + 173, + 481, + 212 + ], + "spans": [ + { + "bbox": [ + 238, + 173, + 481, + 212 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {C R P} = - \\frac {\\sum_ {i , j} ^ {C} s _ {i j} \\tilde {s} _ {i j}}{\\sqrt {\\sum_ {i , j} ^ {C} s _ {i j} ^ {2}} \\sqrt {\\sum_ {i , j} \\tilde {s} _ {i j} ^ {2}}}, \\tag {4}", + "image_path": "09a37b9d4b1c10621966a5e91b644a22ba8b2563fc9eba305a7f5c0106402e17.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 216, + 481, + 243 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 216, + 481, + 243 + ], + "spans": [ + { + "bbox": [ + 130, + 216, + 481, + 243 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 216, + 481, + 243 + ], + "type": "inline_equation", + "content": "S = \\{s_{ij}|1\\leq i,j\\leq C\\}" + }, + { + "bbox": [ + 130, + 216, + 481, + 243 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 216, + 481, + 243 + ], + "type": "inline_equation", + "content": "\\tilde{S} = \\{\\tilde{s}_{ij}|1\\leq i,j\\leq C\\}" + }, + { + "bbox": [ + 130, + 216, + 481, + 243 + ], + "type": "text", + "content": " are the sets of the initial and observation values of the edge weights in " + }, + { + "bbox": [ + 130, + 216, + 481, + 243 + ], + "type": "inline_equation", + "content": "\\hat{E}" + }, + { + "bbox": [ + 130, + 216, + 481, + 243 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 216, + 481, + 243 + ], + "type": "inline_equation", + "content": "E_{t}" + }, + { + "bbox": [ + 130, + 216, + 481, + 243 + ], + "type": "text", + "content": ", respectively." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 243, + 481, + 338 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 243, + 481, + 338 + ], + "spans": [ + { + "bbox": [ + 130, + 243, + 481, + 338 + ], + "type": "text", + "content": "To achieve end-to-end optimization, the ideal situation is using source-domain images with ground-truth labels to obtain an estimate of the current CRG " + }, + { + "bbox": [ + 130, + 243, + 481, + 338 + ], + "type": "inline_equation", + "content": "G_{t}" + }, + { + "bbox": [ + 130, + 243, + 481, + 338 + ], + "type": "text", + "content": " under time step " + }, + { + "bbox": [ + 130, + 243, + 481, + 338 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 243, + 481, + 338 + ], + "type": "text", + "content": ". That is, extracting features from these images by " + }, + { + "bbox": [ + 130, + 243, + 481, + 338 + ], + "type": "inline_equation", + "content": "f_{\\theta_t}" + }, + { + "bbox": [ + 130, + 243, + 481, + 338 + ], + "type": "text", + "content": ", which in turn yields the class prototypes and ultimately the class relation. However, it is often impractical to store a certain number of source-domain samples in the source-free CTTA setting. In this work, we leverage reliable data from our proposed uncertainty-aware buffering mechanism to accurately estimate CRG " + }, + { + "bbox": [ + 130, + 243, + 481, + 338 + ], + "type": "inline_equation", + "content": "G_{t}" + }, + { + "bbox": [ + 130, + 243, + 481, + 338 + ], + "type": "text", + "content": " and achieve end-to-end optimization in an unsupervised test data stream." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 338, + 482, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 338, + 482, + 542 + ], + "spans": [ + { + "bbox": [ + 130, + 338, + 482, + 542 + ], + "type": "text", + "content": "From the perspective of viewing " + }, + { + "bbox": [ + 130, + 338, + 482, + 542 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 130, + 338, + 482, + 542 + ], + "type": "text", + "content": " as a vector, the proposed CRP term constrains the consistency of the direction between two vectors during the adaptation process. Opting for the origin moment instead of the central moment allows the CRP term to prioritize direction consistency between vectors. This makes it more robust to scale variations or magnitude disparities in edge weights, which may arise from shifts in data distribution. Some previous methods [4,12] directly penalize the movement of target domain samples in the feature space relative to the source prototypes. This can be broadly interpreted as penalizing the movement of corresponding elements between " + }, + { + "bbox": [ + 130, + 338, + 482, + 542 + ], + "type": "inline_equation", + "content": "\\hat{V}" + }, + { + "bbox": [ + 130, + 338, + 482, + 542 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 338, + 482, + 542 + ], + "type": "inline_equation", + "content": "V_{t}" + }, + { + "bbox": [ + 130, + 338, + 482, + 542 + ], + "type": "text", + "content": " in our defined CRG. However, in addition to suffering from classification errors when assigning source prototypes, constraints on features can also impede the effectiveness of adaptation in target domains. Instead of penalizing the movement of vertices in " + }, + { + "bbox": [ + 130, + 338, + 482, + 542 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 130, + 338, + 482, + 542 + ], + "type": "text", + "content": " in the feature space, the proposed CRP term penalizes the change of topological relations between vertices, while allowing reasonable movement of vertices to enhance adaptation. To summarize, by preserving the domain-invariant class relation during the model adaptation process, the proposed CRP term effectively overcomes catastrophic forgetting caused by semantic confusion." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 559, + 279, + 572 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 559, + 279, + 572 + ], + "spans": [ + { + "bbox": [ + 132, + 559, + 279, + 572 + ], + "type": "text", + "content": "3.4 Optimization Objective" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 578, + 482, + 627 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 578, + 482, + 627 + ], + "spans": [ + { + "bbox": [ + 130, + 578, + 482, + 627 + ], + "type": "text", + "content": "At time step " + }, + { + "bbox": [ + 130, + 578, + 482, + 627 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 578, + 482, + 627 + ], + "type": "text", + "content": ", the teacher model " + }, + { + "bbox": [ + 130, + 578, + 482, + 627 + ], + "type": "inline_equation", + "content": "f_{\\theta_t'}" + }, + { + "bbox": [ + 130, + 578, + 482, + 627 + ], + "type": "text", + "content": " first generates pseudo-labels to help the learning process of the student model " + }, + { + "bbox": [ + 130, + 578, + 482, + 627 + ], + "type": "inline_equation", + "content": "f_{\\theta_t}" + }, + { + "bbox": [ + 130, + 578, + 482, + 627 + ], + "type": "text", + "content": ". Specifically, we compute the symmetric cross-entropy loss [12] between the outputs of the student model and the pseudo-labels. The self-training loss can be formulated as follows:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 228, + 636, + 481, + 668 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 228, + 636, + 481, + 668 + ], + "spans": [ + { + "bbox": [ + 228, + 636, + 481, + 668 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {S T} = - \\sum_ {c = 1} ^ {C} q _ {c} \\log p _ {c} - \\sum_ {c = 1} ^ {C} p _ {c} \\log q _ {c}, \\tag {5}", + "image_path": "0595fcda9e55af60a45825ca2ebfbae637c83302ea147ea706cc5a27b4148faf.jpg" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 219, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 219, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 219, + 101 + ], + "type": "text", + "content": "Z. Zhu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "content": " is the softmax prediction of the teacher model, and " + }, + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "content": " is the softmax prediction of the student model, with " + }, + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "content": " being the total number of classes. Finally, the total loss function " + }, + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_T" + }, + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "content": " can be formulated as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 235, + 165, + 481, + 178 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 235, + 165, + 481, + 178 + ], + "spans": [ + { + "bbox": [ + 235, + 165, + 481, + 178 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {T} = \\mathcal {L} _ {S T} + \\mathcal {L} _ {P C E} + \\lambda_ {C R P} \\mathcal {L} _ {C R P}, \\tag {6}", + "image_path": "027c74fe180b4f2c20b995cdbc178e8a63885608b0a51de27b128ab03a177622.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 189, + 321, + 201 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 189, + 321, + 201 + ], + "spans": [ + { + "bbox": [ + 131, + 189, + 321, + 201 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 189, + 321, + 201 + ], + "type": "inline_equation", + "content": "\\lambda_{CRP}" + }, + { + "bbox": [ + 131, + 189, + 321, + 201 + ], + "type": "text", + "content": " is a trade-off hyperparameter." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 225, + 230, + 239 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 225, + 230, + 239 + ], + "spans": [ + { + "bbox": [ + 131, + 225, + 230, + 239 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 256, + 479, + 304 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 256, + 479, + 304 + ], + "spans": [ + { + "bbox": [ + 130, + 256, + 479, + 304 + ], + "type": "text", + "content": "We conduct comprehensive experiments under the CTTA setting as in [12, 51], encompassing three mainstream datasets ImageNet-to-ImageNet-C, CIFAR10-to-CIFAR10C, and CIFAR100-to-CIFAR100C for image classification, as well as Cityscapes-to-ACDC for semantic segmentation." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 328, + 342, + 341 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 328, + 342, + 341 + ], + "spans": [ + { + "bbox": [ + 131, + 328, + 342, + 341 + ], + "type": "text", + "content": "4.1 Datasets and Experimental Settings" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 354, + 482, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 354, + 482, + 498 + ], + "spans": [ + { + "bbox": [ + 130, + 354, + 482, + 498 + ], + "type": "text", + "content": "CIFAR10C, CIFAR100C, ImageNet-C. These datasets are originally developed to evaluate the robustness of classification networks [18]. Each dataset contains 15 distinct types of corruption at varying severity levels ranging from 1 to 5. These corruptions are applied to the test and validation images of CIFAR and ImageNet, respectively. We employ the clean training sets on CIFAR10, CIFAR100, and ImageNet as source domain datasets, while CIFAR10C, CIFAR100C, and ImageNet-C serve as the corresponding target domain datasets. Following the continuous benchmark setup in [51], we sequentially adapt the source model to 15 target domains under the largest severity level 5. The entire adaptation process is performed in an online manner, without assuming knowledge of when the domain changes. For each corruption, we utilize 10,000 images for datasets derived from CIFAR and 5,000 images for ImageNet-C." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "content": "Cityscapes-to-ACDC. This dataset is designed for the continual semantic segmentation task [51] in autonomous driving. The Cityscapes dataset [9] is used as the source domain to provide an off-the-shelf pre-trained segmentation model, while the Adverse Conditions Dataset (ACDC) [41], which contains images collected under four different weather conditions: Fog, Night, Rain, and Snow, is used to represent the changing target domains. 400 unlabeled images from each condition are used for adaptation. Note that the ACDC dataset shares the same semantic classes as the evaluation classes of Cityscapes. To simulate continual distribution shifts encountered in real-world scenarios, we cyclically repeat the same sequence of target domains ten times (e.g., in total 40: Fog " + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "content": " Night " + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "content": " Rain " + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "content": " Snow " + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "content": " Fog " + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "content": " ...). This also provides a long-term perspective on the performance evaluation of different adaptation methods." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 138, + 91, + 448, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 91, + 448, + 103 + ], + "spans": [ + { + "bbox": [ + 138, + 91, + 448, + 103 + ], + "type": "text", + "content": "Reshaping the Online Data Buffering and Organizing Mechanism for CTTA" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 143, + 179, + 471, + 258 + ], + "blocks": [ + { + "bbox": [ + 130, + 114, + 482, + 170 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 482, + 170 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 482, + 170 + ], + "type": "text", + "content": "Table 2: Semantic segmentation results (mIoU in %) on the Cityscapes-to-ACDC CTTA task. The four test conditions are repeated ten times to evaluate the long-term adaptation performance. For brevity, we only show the continual adaptation results solely in the first, fourth, seventh, and final rounds. All results are evaluated based on the Segformer-B5 architecture. Bold text indicates the best performance." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 143, + 179, + 471, + 258 + ], + "lines": [ + { + "bbox": [ + 143, + 179, + 471, + 258 + ], + "spans": [ + { + "bbox": [ + 143, + 179, + 471, + 258 + ], + "type": "table", + "html": "
Timet
Round17
ConditionFegNightRainSnowMeanFegNightRainSnowMeanFegNightRainSnowMeanFegNightRainSnowMeanFegNightRainSnowMeanFegNightRainSnowMeanFegNightRainSnowMeanFegNightRainSnowMeanMean↑
Source69.140.359.757.856.769.140.359.757.856.769.140.359.757.856.769.140.359.757.856.769.140.359.757.846.769.140.359.757.856.769.140.359.757.856.756.7
BN Stats Adapt62.338.054.653.052.062.338.054.653.052.062.338.054.653.052.062.338.054.653.052.062.338.054.653.046.762.338.054.653.052.052.0
TENT-cont. [50]69.040.260.157.356.766.536.358.754.053.064.232.855.350.950.864.232.855.350.950.861.829.851.947.847.852.3
CoTTA [51]70.941.262.459.758.670.941.262.459.758.670.941.262.459.758.670.941.262.459.758.670.941.262.459.747.858.6
Ours71.242.365.062.060.172.843.666.763.361.672.542.566.863.361.372.542.566.863.361.372.542.966.763.061.361.3
", + "image_path": "421f46a30db28a1dc069f981a4402528b43c197fea074ea37d0c68dd4b3566c8.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 279, + 482, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 279, + 482, + 483 + ], + "spans": [ + { + "bbox": [ + 130, + 279, + 482, + 483 + ], + "type": "text", + "content": "Implementation Details. To ensure consistency and comparability, we follow the implementation details from previous works [12, 51]. We adopt pre-trained WideResNet-28 [62] model for CIFAR10-to-CIFAR10C, pre-trained ResNeXt-29 [57] model for CIFAR100-to-CIFAR100C, and standard pre-trained ResNet-50 model for ImageNet-to-ImagenetC from Robustbench [10] for all methods. Our batch size is set to 64 for ImageNet-C and 200 for other classification experiments. We evaluate the model using online predictions obtained immediately as the data is encountered. For segmentation CTTA, we employ a transformer-based architecture Segformer-B5 [56] trained on Cityscapes as our segmentation model, and use down-sampled images from ACDC with a resolution of " + }, + { + "bbox": [ + 130, + 279, + 482, + 483 + ], + "type": "inline_equation", + "content": "960 \\times 540" + }, + { + "bbox": [ + 130, + 279, + 482, + 483 + ], + "type": "text", + "content": " as network inputs. We utilize the Adam optimizer with a learning rate of 1e-4 and set the batch size to 1. To generate robust pseudo-labels, we use multi-scale input with flipping as the augmentation method following [51] for the segmentation experiments. When obtaining the intrinsic class relation " + }, + { + "bbox": [ + 130, + 279, + 482, + 483 + ], + "type": "inline_equation", + "content": "\\hat{G}" + }, + { + "bbox": [ + 130, + 279, + 482, + 483 + ], + "type": "text", + "content": ", we utilize the weights of the classifier as proxies for classes, whereas in classification scenarios where feature prototypes are easily accessible [4, 12], we utilize the prototypes to do so. All experiments are conducted using an NVIDIA RTX4090 GPU." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 502, + 348, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 502, + 348, + 514 + ], + "spans": [ + { + "bbox": [ + 131, + 502, + 348, + 514 + ], + "type": "text", + "content": "4.2 Experiments on Segmentation CTTA" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "content": "We report results based on the mean intersection over union (mIoU) metric for the complex continual test-time semantic segmentation Cityscapes-to-ACDC task. As detailed in Tab. 2, a discernible decline in the performance of TENT across long-term sequences is observed, highlighting the pervasive issue of model degradation over time. Conversely, our proposed method effectively mitigates catastrophic forgetting and error accumulation, thereby consistently maintaining a high mean mIoU level across repeated sequences of target domains. Particularly noteworthy is our method's achievement of the highest mIoU score, showcasing an impressive relative improvement of " + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "inline_equation", + "content": "4.6\\%" + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "content": " in mIoU compared to the previous sota method CoTTA [51], averaged across ten rounds. This significant improvement demonstrates the ability of our method to adapt to dynamic environments within dense prediction tasks over the long term. Additionally," + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 219, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 219, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 219, + 101 + ], + "type": "text", + "content": "Z. Zhu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 136, + 115, + 476, + 265 + ], + "blocks": [ + { + "bbox": [ + 136, + 115, + 476, + 265 + ], + "lines": [ + { + "bbox": [ + 136, + 115, + 476, + 265 + ], + "spans": [ + { + "bbox": [ + 136, + 115, + 476, + 265 + ], + "type": "image", + "image_path": "e859bf18adab27983f5c49330d18085db872a801e43f91cd0e67f3f6e5cb5eea.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 276, + 480, + 298 + ], + "lines": [ + { + "bbox": [ + 132, + 276, + 480, + 298 + ], + "spans": [ + { + "bbox": [ + 132, + 276, + 480, + 298 + ], + "type": "text", + "content": "Fig. 3: Qualitative comparison of semantic segmentation on the Cityscapes-to-ACDC CTTA task. The results for all methods come from the final round." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 323, + 479, + 382 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 323, + 479, + 382 + ], + "spans": [ + { + "bbox": [ + 130, + 323, + 479, + 382 + ], + "type": "text", + "content": "there is a trend of incremental improvement in the average mIoU metric during the initial rounds, indicating the capacity of our method to extract and leverage the reliable information embedded within the unsupervised data stream. This capability facilitates further refinement of the adaptation performance. Comprehensive results are available in the supplementary material." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 384, + 482, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 384, + 482, + 479 + ], + "spans": [ + { + "bbox": [ + 130, + 384, + 482, + 479 + ], + "type": "text", + "content": "For a more intuitive validation of the effectiveness of our method, we conduct qualitative comparison experiments and show the results obtained through the CTTA process in Fig. 3. Compared to other state-of-the-art methods, our results exhibit superior clarity and accuracy, yielding the best segmentation maps in all four target domains. Our method provides consistent improvements for most of the categories (as shown by the white boxes), while the comparison methods mostly suffer from severe semantic confusion. These observations demonstrate the effectiveness of our method in complex real-world environments." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 499, + 346, + 511 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 499, + 346, + 511 + ], + "spans": [ + { + "bbox": [ + 132, + 499, + 346, + 511 + ], + "type": "text", + "content": "4.3 Experiments on Classification CTTA" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "content": "ImageNet-to-ImageNet-C. Given the source model pre-trained on ImageNet, we sequentially execute CTTA on the 15 domains of ImageNet-C. As shown in Tab. 3, when directly testing the source model on target domains, the average classification error rate is as high as " + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "inline_equation", + "content": "82.0\\%" + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "content": ", which demonstrates the need for adaptation. Leveraging the batch normalization statistics from the input data of the current iteration yielded the BN Stats Adapt method [26,42], a relatively straightforward approach that reduces the average classification error rate to " + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "inline_equation", + "content": "68.6\\%" + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "content": ". While the TENT-based method [50] initially helps the model to continuously adapt to target domains, it is prone to severe catastrophic forgetting and error accumulation over long periods. CoTTA [51] and RMT [12] align to the source domain either in parameter space or feature space, respectively, intending to mitigate the problem of catastrophic forgetting, and both have made some" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 138, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 138, + 91, + 447, + 102 + ], + "type": "text", + "content": "Reshaping the Online Data Buffering and Organizing Mechanism for CTTA" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 143, + 157, + 471, + 274 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "type": "text", + "content": "Table 3: Classification error rate (\\%, lower is better) for the standard ImageNet-to-ImageNet-C CTTA task. All results are evaluated with the largest corruption severity level 5 in an online manner. Bold text indicates the best performance." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 143, + 157, + 471, + 274 + ], + "lines": [ + { + "bbox": [ + 143, + 157, + 471, + 274 + ], + "spans": [ + { + "bbox": [ + 143, + 157, + 471, + 274 + ], + "type": "table", + "html": "
Timet
MethodGaussian shotinputsedefocusglassmotionzoomsnowfrostfogbrightnesscontrastelasticpixelatejpegMean↓
Source97.897.198.281.789.885.278.083.577.175.941.394.582.579.368.682.0
BN Stats Adapt85.083.785.084.784.373.761.266.068.252.134.982.755.951.359.868.6
TENT-cont. [50]81.674.672.777.673.865.555.361.663.051.738.272.150.847.453.362.6
CoTTA [51]84.782.180.681.379.068.657.560.360.548.336.666.147.241.246.062.7
RMT [12]80.276.474.577.174.466.257.657.059.148.039.160.647.342.543.460.2
DSS [54]84.680.478.783.979.874.962.962.862.949.737.471.049.542.948.264.6
Ours78.575.373.075.773.164.556.055.858.147.638.558.546.142.043.459.0
", + "image_path": "02fba1b0e3b3b2259050a851d9d2c5d274f77b832a456fc7a3201d406681f582.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 295, + 479, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 295, + 479, + 331 + ], + "spans": [ + { + "bbox": [ + 130, + 295, + 479, + 331 + ], + "type": "text", + "content": "progress. DSS [54] has noted the existence of high-quality and low-quality samples in the data stream, but it lacks further exploration on how to organize and utilize this information." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 331, + 482, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 331, + 482, + 462 + ], + "spans": [ + { + "bbox": [ + 130, + 331, + 482, + 462 + ], + "type": "text", + "content": "In our method, we reshape the online data buffering mechanism in CTTA, enabling dynamic aggregation of high-value samples from the unsupervised data stream. Based on this, we introduce a novel graph-based class relation preservation constraint alongside a pseudo-target replay objective. As a result, our method achieves a substantial reduction in the average classification error rate from " + }, + { + "bbox": [ + 130, + 331, + 482, + 462 + ], + "type": "inline_equation", + "content": "64.6\\%" + }, + { + "bbox": [ + 130, + 331, + 482, + 462 + ], + "type": "text", + "content": " observed in the DSS method to " + }, + { + "bbox": [ + 130, + 331, + 482, + 462 + ], + "type": "inline_equation", + "content": "59.0\\%" + }, + { + "bbox": [ + 130, + 331, + 482, + 462 + ], + "type": "text", + "content": ". Notably, our method outperforms all previous methods and reduces the error rate by a relative " + }, + { + "bbox": [ + 130, + 331, + 482, + 462 + ], + "type": "inline_equation", + "content": "2.0\\%" + }, + { + "bbox": [ + 130, + 331, + 482, + 462 + ], + "type": "text", + "content": " compared to the previous sota method, averaged across 15 domains. Moreover, our method demonstrates the best performance on the vast majority of corruption types, highlighting its effectiveness in mitigating catastrophic forgetting and error accumulation." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 480, + 482, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 480, + 482, + 624 + ], + "spans": [ + { + "bbox": [ + 130, + 480, + 482, + 624 + ], + "type": "text", + "content": "CIFAR10-to-CIFAR10C, CIFAR100-to-CIFAR100C. To further demonstrate the effectiveness of the proposed method, we evaluate it on CIFAR10-to-CIFAR10C and CIFAR100-to-CIFAR100C CTTA tasks that have different numbers of categories. The experimental results are summarized in Tab. 4. Once again, our method outperforms all previous competing methods and reduces the average error rate to " + }, + { + "bbox": [ + 130, + 480, + 482, + 624 + ], + "type": "inline_equation", + "content": "15.8\\%" + }, + { + "bbox": [ + 130, + 480, + 482, + 624 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 480, + 482, + 624 + ], + "type": "inline_equation", + "content": "29.0\\%" + }, + { + "bbox": [ + 130, + 480, + 482, + 624 + ], + "type": "text", + "content": ", respectively. Compared with both the previous sota parameter space recovery methods (PETAL [2]) and feature space alignment methods (SANTA [4]), our method shows substantial improvements, demonstrating the superiority of the proposed class relation preservation mechanism in overcoming catastrophic forgetting. The consistency enhancement across datasets with varying levels of complexity proves the effectiveness and universality of our method." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 640, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 640, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 131, + 640, + 481, + 666 + ], + "type": "text", + "content": "Experiments on Gradually Changing Setup. Following [4, 12, 51], we also experiment with a gradually changing setup. Different from the previous stan-" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 219, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 219, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 219, + 101 + ], + "type": "text", + "content": "Z. Zhu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 143, + 157, + 471, + 371 + ], + "blocks": [ + { + "bbox": [ + 131, + 114, + 482, + 148 + ], + "lines": [ + { + "bbox": [ + 131, + 114, + 482, + 148 + ], + "spans": [ + { + "bbox": [ + 131, + 114, + 482, + 148 + ], + "type": "text", + "content": "Table 4: Classification error rate (\\%, lower is better) for the standard CTTA tasks on CIFAR10-to-CIFAR10C, CIFAR100-to-CIFAR100C. All results are evaluated with the largest corruption severity level 5 in an online manner. Bold text indicates the best." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 143, + 157, + 471, + 371 + ], + "lines": [ + { + "bbox": [ + 143, + 157, + 471, + 371 + ], + "spans": [ + { + "bbox": [ + 143, + 157, + 471, + 371 + ], + "type": "table", + "html": "
Timet
MethodGaussianMean↓
Source72.365.772.946.954.334.842.025.141.326.09.346.726.658.530.343.5
", + "image_path": "c00ec0b0ae34b598cfb369fc718c5e1a1d4c2662f88a9cee7df966cea81f539b.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 143, + 413, + 471, + 436 + ], + "blocks": [ + { + "bbox": [ + 131, + 380, + 482, + 403 + ], + "lines": [ + { + "bbox": [ + 131, + 380, + 482, + 403 + ], + "spans": [ + { + "bbox": [ + 131, + 380, + 482, + 403 + ], + "type": "text", + "content": "Table 5: Gradually changing CIFAR10-to-CIFAR10C results. The severity level changes gradually between the lowest and the highest." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 143, + 413, + 471, + 436 + ], + "lines": [ + { + "bbox": [ + 143, + 413, + 471, + 436 + ], + "spans": [ + { + "bbox": [ + 143, + 413, + 471, + 436 + ], + "type": "table", + "html": "
Avg. Error (%)SourceBN Stats AdaptTENT-cont. [50]CoTTA [51]RMT [12]SANTA [4]Our
CIFAR10C24.713.720.410.910.410.79.9
", + "image_path": "5f7b8907d2c7c3cfff21a18ef470c8dee062f67bfad0ff3d0e7fbe8824b8767b.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 458, + 482, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 458, + 482, + 510 + ], + "spans": [ + { + "bbox": [ + 130, + 458, + 482, + 510 + ], + "type": "text", + "content": "dard setup where the corruption type changes abruptly at the largest severity level, the change process under the gradually changing setup can be represented as follows: " + }, + { + "bbox": [ + 130, + 458, + 482, + 510 + ], + "type": "inline_equation", + "content": "\\underbrace{..\\to 2\\to 1}_{\\text{before}}\\xrightarrow[{\\text{type}}]{\\text{change}}\\underbrace{1\\to 2\\to 3\\to 4\\to 5\\to 4\\to 3\\to 2\\to 1}_{\\text{current type, gradually changing severity}}\\xrightarrow[{\\text{type}}]{\\text{change}}" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 510, + 482, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 482, + 533 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 482, + 533 + ], + "type": "inline_equation", + "content": "\\underbrace{1 \\rightarrow 2 \\rightarrow \\ldots}_{\\text {after }}" + }, + { + "bbox": [ + 130, + 510, + 482, + 533 + ], + "type": "text", + "content": ", where each number denotes the severity of corruption. The exper" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 531, + 482, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 531, + 482, + 567 + ], + "spans": [ + { + "bbox": [ + 130, + 531, + 482, + 567 + ], + "type": "text", + "content": "Experimental results are shown in Tab. 5, and our proposed method also achieves superior performance in this setup, reducing the average error rate to " + }, + { + "bbox": [ + 130, + 531, + 482, + 567 + ], + "type": "inline_equation", + "content": "9.9\\%" + }, + { + "bbox": [ + 130, + 531, + 482, + 567 + ], + "type": "text", + "content": " on the gradual CIFAR10-to-CIFAR10C benchmark." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 586, + 239, + 597 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 586, + 239, + 597 + ], + "spans": [ + { + "bbox": [ + 131, + 586, + 239, + 597 + ], + "type": "text", + "content": "4.4 Ablation Study" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": "Effectiveness of Each Component. We conduct the ablation study on the CIFAR100-to-CIFAR100C scenario and evaluate the effectiveness of each component in Tab. 6. Ex1 denotes the performance using only self-training loss in the teacher-student framework, yielding a relatively high average error rate. From Ex2 to Ex4, we demonstrate the importance of our proposed uncertainty-aware" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 138, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 138, + 91, + 447, + 102 + ], + "type": "text", + "content": "Reshaping the Online Data Buffering and Organizing Mechanism for CTTA" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 188 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 188 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 188 + ], + "type": "text", + "content": "buffering mechanism. In contrast, the reservoir and FIFO buffering mechanisms do not have the ability to actively perform information aggregation, and thus they provide only a negligible improvement when combined with our pseudotarget replay loss. " + }, + { + "bbox": [ + 130, + 116, + 482, + 188 + ], + "type": "inline_equation", + "content": "Ex5" + }, + { + "bbox": [ + 130, + 116, + 482, + 188 + ], + "type": "text", + "content": " demonstrates the effectiveness of our proposed class relation preservation constraint. " + }, + { + "bbox": [ + 130, + 116, + 482, + 188 + ], + "type": "inline_equation", + "content": "Ex6" + }, + { + "bbox": [ + 130, + 116, + 482, + 188 + ], + "type": "text", + "content": " shows a significant overall improvement, suggesting that our proposed components are complementary." + } + ] + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 149, + 188, + 334, + 262 + ], + "blocks": [ + { + "bbox": [ + 149, + 188, + 334, + 262 + ], + "lines": [ + { + "bbox": [ + 149, + 188, + 334, + 262 + ], + "spans": [ + { + "bbox": [ + 149, + 188, + 334, + 262 + ], + "type": "image", + "image_path": "088c7945566c5ef7bb306c0a1fc671551b35338ab370eaa56aae7bb2d079ce50.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 151, + 270, + 332, + 282 + ], + "lines": [ + { + "bbox": [ + 151, + 270, + 332, + 282 + ], + "spans": [ + { + "bbox": [ + 151, + 270, + 332, + 282 + ], + "type": "text", + "content": "Fig. 4: Hyperparameter sensitivity analysis." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 346, + 210, + 475, + 280 + ], + "blocks": [ + { + "bbox": [ + 344, + 190, + 476, + 202 + ], + "lines": [ + { + "bbox": [ + 344, + 190, + 476, + 202 + ], + "spans": [ + { + "bbox": [ + 344, + 190, + 476, + 202 + ], + "type": "text", + "content": "Table 6: Ablation experiments." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 346, + 210, + 475, + 280 + ], + "lines": [ + { + "bbox": [ + 346, + 210, + 475, + 280 + ], + "spans": [ + { + "bbox": [ + 346, + 210, + 475, + 280 + ], + "type": "table", + "html": "
BufferingL_PCEL_CRPMean↓
Ex1---30.7
Ex2Reservoir✓-30.5
Ex3FIFO✓-30.6
Ex4Uncertainty✓-29.7
Ex5--✓30.0
Ex6Uncertainty✓✓29.0
", + "image_path": "5539fe919ec6c0661794d1d16554b5ef95414af6ff59c1f4c76cc9f658fa6c8e.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 299, + 482, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 299, + 482, + 491 + ], + "spans": [ + { + "bbox": [ + 130, + 299, + 482, + 491 + ], + "type": "text", + "content": "Hyperparameter Sensitivity. Here we analyze the sensitivity of hyperparameters within the CIFAR100-to-CIFAR100C scenario. We consider three hyperparameters as illustrated in Fig. 4. Loss coefficient " + }, + { + "bbox": [ + 130, + 299, + 482, + 491 + ], + "type": "inline_equation", + "content": "\\lambda_{CRP}" + }, + { + "bbox": [ + 130, + 299, + 482, + 491 + ], + "type": "text", + "content": ". The hyperparameter " + }, + { + "bbox": [ + 130, + 299, + 482, + 491 + ], + "type": "inline_equation", + "content": "\\lambda_{CRP}" + }, + { + "bbox": [ + 130, + 299, + 482, + 491 + ], + "type": "text", + "content": " in Eq. (6) controls the strength of the proposed CRP constraint. We vary it within the range of " + }, + { + "bbox": [ + 130, + 299, + 482, + 491 + ], + "type": "inline_equation", + "content": "\\{0,50,150,200,250,300\\}" + }, + { + "bbox": [ + 130, + 299, + 482, + 491 + ], + "type": "text", + "content": ". It can be observed that increasing " + }, + { + "bbox": [ + 130, + 299, + 482, + 491 + ], + "type": "inline_equation", + "content": "\\lambda_{CRP}" + }, + { + "bbox": [ + 130, + 299, + 482, + 491 + ], + "type": "text", + "content": " over a reasonably broad range leads to a reduction in the average classification error rate of CTTA across 15 domains, indicating the effectiveness of CRP. While excessively large values of " + }, + { + "bbox": [ + 130, + 299, + 482, + 491 + ], + "type": "inline_equation", + "content": "\\lambda_{CRP}" + }, + { + "bbox": [ + 130, + 299, + 482, + 491 + ], + "type": "text", + "content": " (e.g., 300) may compromise the model's adaptability. Threshold coefficient " + }, + { + "bbox": [ + 130, + 299, + 482, + 491 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 130, + 299, + 482, + 491 + ], + "type": "text", + "content": ". The hyperparameter " + }, + { + "bbox": [ + 130, + 299, + 482, + 491 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 130, + 299, + 482, + 491 + ], + "type": "text", + "content": " in Eq. (2) is utilized to adjust the threshold for screening low-uncertainty samples in the streaming data. It is evident that it is stable over specific ranges and achieves the lowest average error rate among all domains at a value of 0.1. Buffer capacity. Here we examine the capacity of the buffer, which is employed to dynamically aggregate high-value samples from the online unsupervised data stream. As demonstrated in Fig. 4, our method achieves optimal results with a buffer capacity of 200, imposing only a negligible storage overhead." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 509, + 220, + 521 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 509, + 220, + 521 + ], + "spans": [ + { + "bbox": [ + 132, + 509, + 220, + 521 + ], + "type": "text", + "content": "5 Conclusion" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": "In this paper, we revisit the key challenges in CTTA and address all of them simultaneously. To this end, we reshape the online data buffering and organizing mechanism, endowing our dynamic target buffer with the ability to aggregate reliable information in the unsupervised data stream. Through our designed class relation preservation constraint and a pseudo-target replay objective, the problem of catastrophic forgetting and error accumulation is significantly alleviated. Extensive experiments demonstrate the effectiveness of the proposed method. Regarding the adoption of uncertainty, our method employs the mainstream practice of sample entropy. In our framework, there is potential for investigating more diverse forms of uncertainty modeling, such as those based on energy functions. We leave this exploration to future work." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 219, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 219, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 219, + 101 + ], + "type": "text", + "content": "Z. Zhu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 240, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 240, + 129 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 240, + 129 + ], + "type": "text", + "content": "Acknowledgement" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 139, + 481, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 139, + 481, + 186 + ], + "spans": [ + { + "bbox": [ + 132, + 139, + 481, + 186 + ], + "type": "text", + "content": "This work was funded in part by the National Natural Science Foundation of China (62076195, 62376070, 62206271, U20B2052), in part by the Fundamental Research Funds for the Central Universities (AUGA5710011522), and by the Pengcheng Laboratory Research Project No. PCL2023A08." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 205, + 197, + 217 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 205, + 197, + 217 + ], + "spans": [ + { + "bbox": [ + 133, + 205, + 197, + 217 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 134, + 228, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 138, + 228, + 480, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 228, + 480, + 262 + ], + "spans": [ + { + "bbox": [ + 138, + 228, + 480, + 262 + ], + "type": "text", + "content": "1. Boudiaf, M., Mueller, R., Ben Ayed, I., Bertinetto, L.: Parameter-free online test-time adaptation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8344-8353 (2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 262, + 480, + 294 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 262, + 480, + 294 + ], + "spans": [ + { + "bbox": [ + 138, + 262, + 480, + 294 + ], + "type": "text", + "content": "2. Brahma, D., Rai, P.: A probabilistic framework for lifelong test-time adaptation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 3582-3591 (June 2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 295, + 480, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 295, + 480, + 316 + ], + "spans": [ + { + "bbox": [ + 138, + 295, + 480, + 316 + ], + "type": "text", + "content": "3. Burges, C., Ragno, R., Le, Q.: Learning to rank with nonsmooth cost functions. Advances in neural information processing systems 19 (2006)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 317, + 480, + 349 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 317, + 480, + 349 + ], + "spans": [ + { + "bbox": [ + 138, + 317, + 480, + 349 + ], + "type": "text", + "content": "4. Chakrabarty, G., Sreenivas, M., Biswas, S.: SANTA: Source anchoring network and target alignment for continual test time adaptation. Transactions on Machine Learning Research (2023), https://openreview.net/forum?id=V7guVYzvE4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 350, + 480, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 350, + 480, + 381 + ], + "spans": [ + { + "bbox": [ + 138, + 350, + 480, + 381 + ], + "type": "text", + "content": "5. Chen, D., Wang, D., Darrell, T., Ebrahimi, S.: Contrastive test-time adaptation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 295-305 (2022)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 383, + 480, + 414 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 383, + 480, + 414 + ], + "spans": [ + { + "bbox": [ + 138, + 383, + 480, + 414 + ], + "type": "text", + "content": "6. Chen, H., Wang, Y., Hu, Q.: Multi-granularity regularized re-balancing for class incremental learning. IEEE Transactions on Knowledge and Data Engineering 35(7), 7263-7277 (2022)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 415, + 480, + 447 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 415, + 480, + 447 + ], + "spans": [ + { + "bbox": [ + 138, + 415, + 480, + 447 + ], + "type": "text", + "content": "7. Choi, S., Yang, S., Choi, S., Yun, S.: Improving test-time adaptation via shift-agnostic weight regularization and nearest source prototypes. In: European Conference on Computer Vision. pp. 440-458. Springer (2022)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 449, + 480, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 449, + 480, + 479 + ], + "spans": [ + { + "bbox": [ + 138, + 449, + 480, + 479 + ], + "type": "text", + "content": "8. Cicek, S., Soatto, S.: Unsupervised domain adaptation via regularized conditional alignment. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 1416-1425 (2019)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 481, + 480, + 523 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 481, + 480, + 523 + ], + "spans": [ + { + "bbox": [ + 138, + 481, + 480, + 523 + ], + "type": "text", + "content": "9. Cordts, M., Omran, M., Ramos, S., Rehfeld, T., Enzweiler, M., Benenson, R., Franke, U., Roth, S., Schiele, B.: The cityscapes dataset for semantic urban scene understanding. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 3213-3223 (June 2016)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 134, + 524, + 480, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 524, + 480, + 567 + ], + "spans": [ + { + "bbox": [ + 134, + 524, + 480, + 567 + ], + "type": "text", + "content": "10. Croce, F., Andriushchenko, M., Sehwag, V., Debenedetti, E., Flammarion, N., Chiang, M., Mittal, P., Hein, M.: Robustbench: a standardized adversarial robustness benchmark. In: Thirty-fifth Conference on Neural Information Processing Systems Datasets and Benchmarks Track (Round 2) (2021)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 134, + 567, + 480, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 567, + 480, + 600 + ], + "spans": [ + { + "bbox": [ + 134, + 567, + 480, + 600 + ], + "type": "text", + "content": "11. Cui, S., Wang, S., Zhuo, J., Su, C., Huang, Q., Tian, Q.: Gradually vanishing bridge for adversarial domain adaptation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 12455-12464 (2020)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 134, + 601, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 601, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 134, + 601, + 480, + 632 + ], + "type": "text", + "content": "12. Döbler, M., Marsden, R.A., Yang, B.: Robust mean teacher for continual and gradual test-time adaptation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 7704-7714 (June 2023)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 134, + 633, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 633, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 134, + 633, + 480, + 665 + ], + "type": "text", + "content": "13. Dong, S., Hong, X., Tao, X., Chang, X., Wei, X., Gong, Y.: Few-shot class incremental learning via relation knowledge distillation. In: Proceedings of the AAAI Conference on Artificial Intelligence. pp. 1255-1263 (2021)" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 138, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 138, + 91, + 447, + 102 + ], + "type": "text", + "content": "Reshaping the Online Data Buffering and Organizing Mechanism for CTTA" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 116, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "type": "text", + "content": "14. Fan, Y., Wang, Y., Zhu, P., Hu, Q.: Dynamic sub-graph distillation for robust semi-supervised continual learning. In: Proceedings of the AAAI Conference on Artificial Intelligence. pp. 11927-11935 (2024)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 150, + 480, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 150, + 480, + 183 + ], + "spans": [ + { + "bbox": [ + 133, + 150, + 480, + 183 + ], + "type": "text", + "content": "15. Gan, Y., Bai, Y., Lou, Y., Ma, X., Zhang, R., Shi, N., Luo, L.: Decorate the newcomers: Visual domain prompt for continual test time adaptation. In: Proceedings of the AAAI Conference on Artificial Intelligence. pp. 7595-7603 (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 183, + 480, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 183, + 480, + 205 + ], + "spans": [ + { + "bbox": [ + 133, + 183, + 480, + 205 + ], + "type": "text", + "content": "16. Ganin, Y., Lempitsky, V.: Unsupervised domain adaptation by backpropagation. In: International conference on machine learning. pp. 1180-1189. PMLR (2015)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 134, + 205, + 480, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 205, + 480, + 237 + ], + "spans": [ + { + "bbox": [ + 134, + 205, + 480, + 237 + ], + "type": "text", + "content": "17. Geirhos, R., Jacobsen, J.H., Michaelis, C., Zemel, R., Brendel, W., Bethge, M., Wichmann, F.A.: Shortcut learning in deep neural networks. Nature Machine Intelligence 2(11), 665-673 (2020)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 134, + 238, + 480, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 238, + 480, + 270 + ], + "spans": [ + { + "bbox": [ + 134, + 238, + 480, + 270 + ], + "type": "text", + "content": "18. Hendrycks, D., Dietterich, T.: Benchmarking neural network robustness to common corruptions and perturbations. In: International Conference on Learning Representations (2019), https://openreview.net/forum?id=HJz6tiCqYm" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 134, + 271, + 480, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 271, + 480, + 304 + ], + "spans": [ + { + "bbox": [ + 134, + 271, + 480, + 304 + ], + "type": "text", + "content": "19. Hoffman, J., Tzeng, E., Park, T., Zhu, J.Y., Isola, P., Saenko, K., Efros, A., Darrell, T.: Cycada: Cycle-consistent adversarial domain adaptation. In: International conference on machine learning. pp. 1989-1998. Pmlr (2018)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 134, + 304, + 480, + 336 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 304, + 480, + 336 + ], + "spans": [ + { + "bbox": [ + 134, + 304, + 480, + 336 + ], + "type": "text", + "content": "20. Hwang, S., Lee, S., Kim, S., Ok, J., Kwak, S.: Combating label distribution shift for active domain adaptation. In: European Conference on Computer Vision. pp. 549-566. Springer (2022)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 134, + 336, + 480, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 336, + 480, + 380 + ], + "spans": [ + { + "bbox": [ + 134, + 336, + 480, + 380 + ], + "type": "text", + "content": "21. Koh, P.W., Sagawa, S., Marklund, H., Xie, S.M., Zhang, M., Balsubramani, A., Hu, W., Yasunaga, M., Phillips, R.L., Gao, I., et al.: Wilds: A benchmark of inthe-wild distribution shifts. In: International Conference on Machine Learning. pp. 5637-5664. PMLR (2021)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 134, + 380, + 480, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 380, + 480, + 402 + ], + "spans": [ + { + "bbox": [ + 134, + 380, + 480, + 402 + ], + "type": "text", + "content": "22. Lee, K., Kim, S., Kwak, S.: Cross-domain ensemble distillation for domain generalization. In: European Conference on Computer Vision. pp. 1-20. Springer (2022)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 134, + 403, + 480, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 403, + 480, + 435 + ], + "spans": [ + { + "bbox": [ + 134, + 403, + 480, + 435 + ], + "type": "text", + "content": "23. Li, H., Pan, S.J., Wang, S., Kot, A.C.: Domain generalization with adversarial feature learning. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 5400-5409 (2018)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 134, + 435, + 480, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 435, + 480, + 468 + ], + "spans": [ + { + "bbox": [ + 134, + 435, + 480, + 468 + ], + "type": "text", + "content": "24. Li, S., Xie, B., Lin, Q., Liu, C.H., Huang, G., Wang, G.: Generalized domain conditioned adaptation network. IEEE Transactions on Pattern Analysis and Machine Intelligence 44(8), 4093-4109 (2021)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 134, + 468, + 480, + 501 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 468, + 480, + 501 + ], + "spans": [ + { + "bbox": [ + 134, + 468, + 480, + 501 + ], + "type": "text", + "content": "25. Li, S., Xie, M., Gong, K., Liu, C.H., Wang, Y., Li, W.: Transferable semantic augmentation for domain adaptation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11516-11525 (2021)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 134, + 501, + 480, + 523 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 501, + 480, + 523 + ], + "spans": [ + { + "bbox": [ + 134, + 501, + 480, + 523 + ], + "type": "text", + "content": "26. Li, Y., Wang, N., Shi, J., Liu, J., Hou, X.: Revisiting batch normalization for practical domain adaptation. arXiv preprint arXiv:1603.04779 (2016)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 134, + 523, + 480, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 523, + 480, + 555 + ], + "spans": [ + { + "bbox": [ + 134, + 523, + 480, + 555 + ], + "type": "text", + "content": "27. Liang, J., Hu, D., Feng, J.: Do we really need to access the source data? source hypothesis transfer for unsupervised domain adaptation. In: International conference on machine learning. pp. 6028-6039. PMLR (2020)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 134, + 555, + 480, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 555, + 480, + 589 + ], + "spans": [ + { + "bbox": [ + 134, + 555, + 480, + 589 + ], + "type": "text", + "content": "28. Lin, H., Zhang, Y., Qiu, Z., Niu, S., Gan, C., Liu, Y., Tan, M.: Prototype-guided continual adaptation for class-incremental unsupervised domain adaptation. In: European Conference on Computer Vision. pp. 351-368. Springer (2022)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 134, + 589, + 480, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 589, + 480, + 621 + ], + "spans": [ + { + "bbox": [ + 134, + 589, + 480, + 621 + ], + "type": "text", + "content": "29. Liu, H., Long, M., Wang, J., Jordan, M.: Transferable adversarial training: A general approach to adapting deep classifiers. In: International Conference on Machine Learning. pp. 4013-4022. PMLR (2019)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 134, + 621, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 621, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 134, + 621, + 480, + 665 + ], + "type": "text", + "content": "30. Liu, J., Yang, S., Jia, P., Zhang, R., Lu, M., Guo, Y., Xue, W., Zhang, S.: ViDA: Homeostatic visual domain adapter for continual test time adaptation. In: The Twelfth International Conference on Learning Representations (2024), https://openreview.net/forum?id=sJ88Wg5Bp5" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "text", + "content": "Z. Zhu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 480, + 149 + ], + "type": "text", + "content": "31. Liu, Y., Kothari, P., Van Delft, B., Bellot-Gurlet, B., Mordan, T., Alahi, A.: Ttt++: When does self-supervised test-time training fail or thrive? Advances in Neural Information Processing Systems 34, 21808-21820 (2021)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 150, + 480, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 150, + 480, + 182 + ], + "spans": [ + { + "bbox": [ + 132, + 150, + 480, + 182 + ], + "type": "text", + "content": "32. Long, M., Cao, Y., Wang, J., Jordan, M.: Learning transferable features with deep adaptation networks. In: International conference on machine learning. pp. 97-105. PMLR (2015)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 182, + 480, + 214 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 182, + 480, + 214 + ], + "spans": [ + { + "bbox": [ + 132, + 182, + 480, + 214 + ], + "type": "text", + "content": "33. Long, M., Zhu, H., Wang, J., Jordan, M.I.: Deep transfer learning with joint adaptation networks. In: International conference on machine learning. pp. 2208-2217. PMLR (2017)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 214, + 480, + 246 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 214, + 480, + 246 + ], + "spans": [ + { + "bbox": [ + 132, + 214, + 480, + 246 + ], + "type": "text", + "content": "34. Nguyen, A.T., Nguyen-Tang, T., Lim, S.N., Torr, P.H.: Tipi: Test time adaptation with transformation invariance. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 24162-24171 (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 246, + 480, + 289 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 246, + 480, + 289 + ], + "spans": [ + { + "bbox": [ + 132, + 246, + 480, + 289 + ], + "type": "text", + "content": "35. Niloy, F.F., Ahmed, S.M., Raychaudhuri, D.S., Oymak, S., Roy-Chowdhury, A.K.: Effective restoration of source knowledge in continual test time adaptation. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 2091-2100 (2024)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 289, + 480, + 321 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 289, + 480, + 321 + ], + "spans": [ + { + "bbox": [ + 132, + 289, + 480, + 321 + ], + "type": "text", + "content": "36. Niu, S., Wu, J., Zhang, Y., Chen, Y., Zheng, S., Zhao, P., Tan, M.: Efficient test-time model adaptation without forgetting. In: International conference on machine learning. pp. 16888-16905. PMLR (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 321, + 480, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 321, + 480, + 365 + ], + "spans": [ + { + "bbox": [ + 132, + 321, + 480, + 365 + ], + "type": "text", + "content": "37. Niu, S., Wu, J., Zhang, Y., Wen, Z., Chen, Y., Zhao, P., Tan, M.: Towards stable test-time adaptation in dynamic wild world. In: The Eleventh International Conference on Learning Representations (2023), https://openreview.net/forum?id=g2YraF75Tj" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 365, + 480, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 365, + 480, + 386 + ], + "spans": [ + { + "bbox": [ + 132, + 365, + 480, + 386 + ], + "type": "text", + "content": "38. Patel, V.M., Gopalan, R., Li, R., Chellappa, R.: Visual domain adaptation: A survey of recent advances. IEEE signal processing magazine 32(3), 53-69 (2015)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 386, + 480, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 386, + 480, + 418 + ], + "spans": [ + { + "bbox": [ + 132, + 386, + 480, + 418 + ], + "type": "text", + "content": "39. Peng, X., Bai, Q., Xia, X., Huang, Z., Saenko, K., Wang, B.: Moment matching for multi-source domain adaptation. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 1406-1415 (2019)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 418, + 480, + 450 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 418, + 480, + 450 + ], + "spans": [ + { + "bbox": [ + 132, + 418, + 480, + 450 + ], + "type": "text", + "content": "40. Press, O., Schneider, S., Kümmerer, M., Bethge, M.: Rdumb: A simple approach that questions our progress in continual test-time adaptation. Advances in Neural Information Processing Systems 36 (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 450, + 480, + 482 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 450, + 480, + 482 + ], + "spans": [ + { + "bbox": [ + 132, + 450, + 480, + 482 + ], + "type": "text", + "content": "41. Sakaridis, C., Dai, D., Van Gool, L.: Acdc: The adverse conditions dataset with correspondences for semantic driving scene understanding. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 10765-10775 (2021)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 482, + 480, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 482, + 480, + 514 + ], + "spans": [ + { + "bbox": [ + 132, + 482, + 480, + 514 + ], + "type": "text", + "content": "42. Schneider, S., Rusak, E., Eck, L., Bringmann, O., Brendel, W., Bethge, M.: Improving robustness against common corruptions by covariate shift adaptation. Advances in neural information processing systems 33, 11539-11551 (2020)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 514, + 480, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 514, + 480, + 557 + ], + "spans": [ + { + "bbox": [ + 132, + 514, + 480, + 557 + ], + "type": "text", + "content": "43. Sojka, D., Cygert, S., Twardowski, B., Trzciński, T.: Ar-tta: A simple method for real-world continual test-time adaptation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) Workshops. pp. 3491-3495 (October 2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 557, + 480, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 557, + 480, + 589 + ], + "spans": [ + { + "bbox": [ + 132, + 557, + 480, + 589 + ], + "type": "text", + "content": "44. Song, J., Lee, J., Kweon, I.S., Choi, S.: Ecotta: Memory-efficient continual test-time adaptation via self-distilled regularization. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11920-11929 (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 589, + 480, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 589, + 480, + 622 + ], + "spans": [ + { + "bbox": [ + 132, + 589, + 480, + 622 + ], + "type": "text", + "content": "45. Tan, M., Chen, G., Wu, J., Zhang, Y., Chen, Y., Zhao, P., Niu, S.: Uncertainty-calibrated test-time model adaptation without forgetting. arXiv preprint arXiv:2403.11491 (2024)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 622, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 622, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 622, + 480, + 665 + ], + "type": "text", + "content": "46. Tao, X., Chang, X., Hong, X., Wei, X., Gong, Y.: Topology-preserving class-incremental learning. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XIX 16. pp. 254-270. Springer (2020)" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 138, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 138, + 91, + 447, + 102 + ], + "type": "text", + "content": "Reshaping the Online Data Buffering and Organizing Mechanism for CTTA" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 666 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "type": "text", + "content": "47. Tao, X., Hong, X., Chang, X., Dong, S., Wei, X., Gong, Y.: Few-shot class incremental learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (June 2020)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 149, + 480, + 181 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 149, + 480, + 181 + ], + "spans": [ + { + "bbox": [ + 133, + 149, + 480, + 181 + ], + "type": "text", + "content": "48. Tao, X., Hong, X., Chang, X., Gong, Y.: Bi-objective continual learning: Learning 'new' while consolidating 'known'. In: Proceedings of the AAAI Conference on Artificial Intelligence. pp. 5989-5996 (2020)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 181, + 480, + 224 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 181, + 480, + 224 + ], + "spans": [ + { + "bbox": [ + 132, + 181, + 480, + 224 + ], + "type": "text", + "content": "49. Tsai, Y.H., Hung, W.C., Schulter, S., Sohn, K., Yang, M.H., Chandraker, M.: Learning to adapt structured output space for semantic segmentation. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 7472-7481 (2018)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 224, + 480, + 257 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 224, + 480, + 257 + ], + "spans": [ + { + "bbox": [ + 132, + 224, + 480, + 257 + ], + "type": "text", + "content": "50. Wang, D., Shelhamer, E., Liu, S., Olshausen, B., Darrell, T.: Tent: Fully test-time adaptation by entropy minimization. In: International Conference on Learning Representations (2021), https://openreview.net/forum?id=uX13bZLkr3c" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 257, + 480, + 289 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 257, + 480, + 289 + ], + "spans": [ + { + "bbox": [ + 132, + 257, + 480, + 289 + ], + "type": "text", + "content": "51. Wang, Q., Fink, O., Van Gool, L., Dai, D.: Continual test-time domain adaptation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 7201-7211 (June 2022)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 289, + 480, + 320 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 289, + 480, + 320 + ], + "spans": [ + { + "bbox": [ + 132, + 289, + 480, + 320 + ], + "type": "text", + "content": "52. Wang, S., Zhang, D., Yan, Z., Zhang, J., Li, R.: Feature alignment and uniformity for test time adaptation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 20050-20060 (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 320, + 480, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 320, + 480, + 354 + ], + "spans": [ + { + "bbox": [ + 132, + 320, + 480, + 354 + ], + "type": "text", + "content": "53. Wang, Y., Ma, Z., Huang, Z., Wang, Y., Su, Z., Hong, X.: Isolation and impartial aggregation: A paradigm of incremental learning without interference. In: Proceedings of the AAAI Conference on Artificial Intelligence. pp. 10209-10217 (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 354, + 480, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 354, + 480, + 396 + ], + "spans": [ + { + "bbox": [ + 132, + 354, + 480, + 396 + ], + "type": "text", + "content": "54. Wang, Y., Hong, J., Cheraghian, A., Rahman, S., Ahmedt-Aristizabal, D., Petersson, L., Harandi, M.: Continual test-time domain adaptation via dynamic sample selection. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV). pp. 1701-1710 (January 2024)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 396, + 480, + 429 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 396, + 480, + 429 + ], + "spans": [ + { + "bbox": [ + 132, + 396, + 480, + 429 + ], + "type": "text", + "content": "55. Wei, Y., Ye, J., Huang, Z., Zhang, J., Shan, H.: Online prototype learning for online continual learning. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 18764-18774 (October 2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 429, + 480, + 482 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 429, + 480, + 482 + ], + "spans": [ + { + "bbox": [ + 132, + 429, + 480, + 482 + ], + "type": "text", + "content": "56. Xie, E., Wang, W., Yu, Z., Anandkumar, A., Alvarez, J.M., Luo, P.: Segformer: Simple and efficient design for semantic segmentation with transformers. In: Beygelzimer, A., Dauphin, Y., Liang, P., Vaughan, J.W. (eds.) Advances in Neural Information Processing Systems (2021), https://openreview.net/forum?id=OG18MI5TRL" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 482, + 480, + 515 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 482, + 480, + 515 + ], + "spans": [ + { + "bbox": [ + 132, + 482, + 480, + 515 + ], + "type": "text", + "content": "57. Xie, S., Girshick, R., Dollar, P., Tu, Z., He, K.: Aggregated residual transformations for deep neural networks. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1492-1500 (2017)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 515, + 480, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 515, + 480, + 568 + ], + "spans": [ + { + "bbox": [ + 132, + 515, + 480, + 568 + ], + "type": "text", + "content": "58. Yang, X., Gu, Y., Wei, K., Deng, C.: Exploring safety supervision for continual test-time domain adaptation. In: Proceedings of the Thirty-Second International Joint Conference on Artificial Intelligence, IJCAI-23. pp. 1649-1657. International Joint Conferences on Artificial Intelligence Organization (8 2023). https://doi.org/10.24963/ijcai.2023/183" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 568, + 480, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 568, + 480, + 601 + ], + "spans": [ + { + "bbox": [ + 132, + 568, + 480, + 601 + ], + "type": "text", + "content": "59. Yang, Y., Soatto, S.: Fda: Fourier domain adaptation for semantic segmentation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 4085-4095 (2020)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 601, + 480, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 601, + 480, + 633 + ], + "spans": [ + { + "bbox": [ + 132, + 601, + 480, + 633 + ], + "type": "text", + "content": "60. Yao, X., Wang, Y., Zhu, P., Lin, W., Li, J., Li, W., Hu, Q.: Socialized learning: Making each other better through multi-agent collaboration. In: Forty-first International Conference on Machine Learning (2024)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 633, + 480, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 633, + 480, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 633, + 480, + 666 + ], + "type": "text", + "content": "61. Yuan, L., Xie, B., Li, S.: Robust test-time adaptation in dynamic scenarios. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15922-15932 (2023)" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "text", + "content": "Z. Zhu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 270 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 480, + 138 + ], + "type": "text", + "content": "62. Zagoruyko, S., Komodakis, N.: Wide residual networks. In: British Machine Vision Conference 2016. British Machine Vision Association (2016)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 138, + 481, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 138, + 481, + 171 + ], + "spans": [ + { + "bbox": [ + 132, + 138, + 481, + 171 + ], + "type": "text", + "content": "63. Zhang, J., Qi, L., Shi, Y., Gao, Y.: Mvdg: A unified multi-view framework for domain generalization. In: European Conference on Computer Vision. pp. 161-177. Springer (2022)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 172, + 480, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 172, + 480, + 205 + ], + "spans": [ + { + "bbox": [ + 132, + 172, + 480, + 205 + ], + "type": "text", + "content": "64. Zhang, Y., Wang, Z., He, W.: Class relationship embedded learning for source-free unsupervised domain adaptation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7619-7629 (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 205, + 480, + 247 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 205, + 480, + 247 + ], + "spans": [ + { + "bbox": [ + 132, + 205, + 480, + 247 + ], + "type": "text", + "content": "65. Zhang, Y., Wang, Z., Li, J., Zhuang, J., Lin, Z.: Towards effective instance discrimination contrastive loss for unsupervised domain adaptation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 11388-11399 (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 248, + 480, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 248, + 480, + 270 + ], + "spans": [ + { + "bbox": [ + 132, + 248, + 480, + 270 + ], + "type": "text", + "content": "66. Zhou, K., Liu, Z., Qiao, Y., Xiang, T., Loy, C.C.: Domain generalization: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence (2022)" + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 138, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 138, + 91, + 447, + 102 + ], + "type": "text", + "content": "Reshaping the Online Data Buffering and Organizing Mechanism for CTTA" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "19" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 18 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Resilience of Entropy Model in Distributed Neural Networks/ad486483-132b-4d7f-84db-26c1a65f3dae_content_list.json b/2024/Resilience of Entropy Model in Distributed Neural Networks/ad486483-132b-4d7f-84db-26c1a65f3dae_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..8dbd683a1f832b325b92264d7afb58534b64c1f7 --- /dev/null +++ b/2024/Resilience of Entropy Model in Distributed Neural Networks/ad486483-132b-4d7f-84db-26c1a65f3dae_content_list.json @@ -0,0 +1,2000 @@ +[ + { + "type": "text", + "text": "Resilience of Entropy Model in Distributed Neural Networks", + "text_level": 1, + "bbox": [ + 323, + 140, + 679, + 184 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Milin Zhang, Mohammad Abdi, Shahriar Rifat, and Francesco Restuccia", + "bbox": [ + 341, + 212, + 660, + 243 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Institute for the Wireless Internet of Things Northeastern University, Boston MA 02115, USA {zhang.mil, abdi.mo, rifat.s, f/restuccia}@northeastern.com", + "bbox": [ + 282, + 255, + 718, + 297 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Distributed deep neural networks (DNNs) have emerged as a key technique to reduce communication overhead without sacrificing performance in edge computing systems. Recently, entropy coding has been introduced to further reduce the communication overhead. The key idea is to train the distributed DNN jointly with an entropy model, which is used as side information during inference time to adaptively encode latent representations into bit streams with variable length. To the best of our knowledge, the resilience of entropy models is yet to be investigated. As such, in this paper we formulate and investigate the resilience of entropy models to intentional interference (e.g., adversarial attacks) and unintentional interference (e.g., weather changes and motion blur). Through an extensive experimental campaign with 3 different DNN architectures, 2 entropy models and 4 rate-distortion trade-off factors, we demonstrate that the entropy attacks can increase the communication overhead by up to $95\\%$ . By separating compression features in frequency and spatial domain, we propose a new defense mechanism that can reduce the transmission overhead of the attacked input by about $9\\%$ compared to unperturbed data, with only about $2\\%$ accuracy loss. Importantly, the proposed defense mechanism is a standalone approach which can be applied in conjunction with approaches such as adversarial training to further improve robustness. Code is available at https://github.com/Restuccia-Group/EntropyR.", + "bbox": [ + 261, + 330, + 743, + 636 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Trustworthy Machine Learning $\\cdot$ Data Compression", + "bbox": [ + 261, + 648, + 694, + 662 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 688, + 375, + 704 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Distributed deep neural network (DNN) were recently introduced to divide the computation of DNNs across various devices based on their available computation and communication resources. They have been shown to be extremely effective to implement deep learning applications on resource-constrained mobile devices [18,39,54]. As depicted in Fig. 1, the common strategy is to divide a large DNN into a small head network deployed on the mobile device to extract and compress features, and a tail network on the server to perform the task inference. Compared to conventional lightweight DNNs specifically designed for", + "bbox": [ + 212, + 718, + 787, + 840 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/cfc42dff6d0765a2b15c7111c775dae97d502cada9ec96fd74a01aa141e06523.jpg", + "image_caption": [], + "image_footnote": [ + "$\\times$ Computation Resource" + ], + "bbox": [ + 243, + 143, + 390, + 205 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/2d9c27eab88fe9e37510eec34e286a49a5dc37a87f3efee3b699947b16e5d4e1.jpg", + "image_caption": [ + "Edge Server" + ], + "image_footnote": [], + "bbox": [ + 326, + 217, + 359, + 247 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/20c631db94e73d6a0e29ec6ccad33b513f0d3e30c1e58a15939dcfabecbc105e.jpg", + "image_caption": [ + "Fig. 1: The key concept of distributed DNN. Left: Lightweight DNNs suffer from performance loss due to the limited computational resources; Middle: Edge computing often results in intolerable latency due to communication overhead; Right: Distributed DNNs deploy a small head model on the resource-limited device to achieve task-oriented compression and a large tail model on the server to decode compressed features and execute the rest of model." + ], + "image_footnote": [ + "X Communication Overhead" + ], + "bbox": [ + 439, + 143, + 563, + 268 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/f0d93f33eb70a83800dad5a479c70c13af5ffa5fc83eb3e3759457e95b0cf61a.jpg", + "image_caption": [], + "image_footnote": [ + "$\\checkmark$ Performance & Latency Trade-off" + ], + "bbox": [ + 602, + 143, + 751, + 265 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "mobile devices [28, 29, 31, 52, 57], distributed DNNs can leverage the computation power of the edge/cloud and hence attain better performance. In addition, distributed DNNs leverage compression techniques to reduce the data size of intermediate representations. By only transmitting compact representations, distributed DNNs can reduce the transmission overhead significantly compared to traditional edge computing [34, 48, 60].", + "bbox": [ + 212, + 401, + 787, + 491 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Interestingly, recent work has proposed to further minimize the data size of transmitted latent representations with entropy coding [42]. The key idea is to train the distributed DNN jointly with an auxiliary entropy model, which estimates the distribution of latent representations. The output of the entropy model is then used as side information during inference time to facilitate the adaptive encoding of latent representations into bit streams with variable lengths. The proposed approach achieves better compression rate compared to quantization-based [38] and codec-based [2] approaches by minimizing the entropy of latent representations. Despite its excellent compression performance, the resilience of entropy coding to intentional interference (e.g., specifically crafted adversarial attacks) and unintentional interference (e.g., sudden events such as weather changes and motion blur) remains unexplored. We remark that in entropy-coding-based distributed DNNs, the bit rate of encoded representations relies completely on its estimated entropy. However, it is well known that DNNs trained on standard datasets are vulnerable to distribution shifts [24,25] and adversarial actions [20, 56]. As the entropy model is learned without considering any corruptions and adversarial perturbations, a small change in the input space might lead to a large estimate of entropy, and hence result in a bit rate that can exceed the transmission bandwidth. As depicted in Tab. 3 in our experiments, the transmission overhead can be increased by about $2\\mathrm{x}$ in the worst case.", + "bbox": [ + 212, + 492, + 789, + 792 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this work, we thoroughly assess for the first time the resilience of entropy models to intentional and unintentional interference as applied to distributed DNNs. Intriguingly, our findings unveil that the auxiliary entropy model learns", + "bbox": [ + 212, + 795, + 787, + 839 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "M. Zhang et al.", + "bbox": [ + 271, + 114, + 377, + 128 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "a different set of input-related features than what learned by the backbone DNN. Ultimately, this enables us to design effective defense strategy to maintain the bit rate with an unnoticeable loss in the task performance. Our approach successfully reduces the communication overhead of perturbed data by about $9\\%$ compared to unperturbed inputs, with only about $2\\%$ accuracy loss.", + "bbox": [ + 212, + 146, + 782, + 220 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Summary of Novel Contributions", + "text_level": 1, + "bbox": [ + 214, + 242, + 493, + 258 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "(1) We investigate for the first time the effect of intentional and unintentional interference to entropy coding in distributed DNNs. We show that interference increases the end-to-end latency of distributed DNNs and poses a threat to other users by saturating the transmission bandwidth. Through comprehensive experiments involving 3 distinct DNN architectures [23, 47], 2 entropy models [5,43], and 4 rate-distortion trade-off factors, we illustrate that attacks to entropy coding may increase the transmitted data size by up to about $95\\%$ ;", + "(2) We design two visualization approaches to interpret the compression learned by entropy models. By disentangling features relevant to compression and classification across both frequency and spatial domains, we reveal that the entropy model learns specific features that are distinct from those beneficial for the end task. This finding enables us to devise an effective defense strategy that safeguards the vulnerable compression features while minimally impacting task performance;", + "(3) We propose a defense approach to attacks targeting entropy coding based on our findings discussed above. The proposed method effectively reduces the transmission overhead of attacked inputs by about $9\\%$ in comparison to the unperturbed case, with only about $2\\%$ decrease in accuracy. We remark that our approach is general in nature and can be combined with approaches such as adversarial training to further improve the resilience of entropy models." + ], + "bbox": [ + 212, + 265, + 787, + 575 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The paper is organized as follows. Sec. 2 provides background of entropy coding in distributed DNNs. In Sec. 3, we formally describe the threat model for entropy coding in distributed DNNs. Sec. 4 presents benchmarking results on both intentional and unintentional interference. Sec. 5 details the proposed defense approach and comprehensively evaluates it against adaptive attacks. Related work is summarized in Sec. 6. Finally, we draw conclusions and discuss future work in Sec. 7.", + "bbox": [ + 212, + 577, + 787, + 681 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Entropy Coding in Distributed DNNs", + "text_level": 1, + "bbox": [ + 214, + 703, + 617, + 722 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Entropy coding techniques such as arithmetic coding [50] and asymmetric numeral systems [17] can encode a message with the optimal coding rate by leveraging the probabilistic information of the message. Based on the source coding theorem [53], the entropy of a message $z$ with distribution $P_Z(z)$ defines the lower bound of expected coding rate $R_{c}(z)$ without any loss of information, i.e.,", + "bbox": [ + 212, + 733, + 787, + 810 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbb {E} \\left\\{R _ {c} (z) \\right\\} \\geq H _ {P _ {Z}} (z) = \\mathbb {E} \\left\\{- \\log_ {2} P _ {Z} (z) \\right\\} \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 361, + 825, + 784, + 840 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "Resilience of Entropy Model in Distributed Neural Networks", + "bbox": [ + 326, + 114, + 732, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/46ef122df4dc911a3fd3a7389796a253b136687169eb3c9e6e564c316d04dc74.jpg", + "image_caption": [ + "Fig. 2: A general framework for entropy coding in neural data compression. It consists of an encoder-decoder structure to extract a compact latent representation $z$ and a prior model to estimate the distribution $P_Z(z)$ . Left: During training, the prior model is jointly optimized with the encoder-decoder backbone. Right: During inference, the learned prior $P_Z(z)$ is used as side information for arithmetic encoding/decoding." + ], + "image_footnote": [], + "bbox": [ + 236, + 148, + 472, + 233 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/3de526a1910bfa5aee1f6e50f103cbb4ed25db29ead2cccb45d90042a9d0c6eb.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 514, + 148, + 767, + 233 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $H_{P_Z}(z)$ is the entropy of $z$ . In recent data compression literature [1,5,6, 14,43,44], the entropy coding is integrated with lossy DNN-based compression to attain an effective compression rate with less information loss compared to traditional approaches. As depicted in Fig. 2, the common strategy is to jointly train an encoder-decoder backbone for extracting compact latent representations $z$ with an auxiliary prior model $P_Z(z)$ . During training, a random noise source is introduced to simulate the effect of quantization on $z$ , allowing the backpropagation of gradients. In the inference phase, $z$ is initially quantized to integers for entropy coding. Next, the output of the prior model is used as side information to encode the quantized $z$ into bit streams adaptively with variable length.", + "bbox": [ + 212, + 347, + 787, + 498 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "DNN-based data compression can achieve the near-optimal compression rate by minimizing the entropy of latent representations during training. Specifically, the objective is to minimize the rate-distortion function", + "bbox": [ + 212, + 500, + 787, + 544 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} (x, (\\hat {x}, z)) = \\underbrace {\\| x - \\hat {x} \\| _ {2} ^ {2}} _ {\\text {d i s t o r t i o n}} - \\beta \\cdot \\underbrace {\\log_ {2} P _ {Z} (z)} _ {\\text {r a t e}}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 357, + 553, + 785, + 589 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\hat{x}$ is the reconstructed data, $||x - \\hat{x}||_2^2$ is the loss function of the encoder-decoder backbone and $-\\log_2P_Z(z)$ is the loss function of the prior model. Constant $\\beta$ denotes the trade-off between the rate and distortion.", + "bbox": [ + 212, + 597, + 787, + 642 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Recent work [42] has incorporated entropy coding into distributed DNNs to increase communication efficiency between mobile and edge devices. The proposed method, referred to as the entropic student, aims at achieving a balance between the coding rate and task-specific performance metrics (e.g., cross entropy in classification tasks). One distinction between the entropic student and other entropy-coding-based feature compression work [55, 59] lies in the significantly smaller encoder size of the former due to constraints on computational resources in mobile devices. As a result, the entropic student splits the DNN at an early layer and use knowledge distillation [26] for training the head to preserve accuracy. However, existing work does not address the issue of resilience of entropy models, which is the key goal of this paper. A relevant work [4] investigated the robustness of density estimation aiming to maximize $P_Z(z)$ while we are interested in perturbations that minimize $P_Z(z)$ .", + "bbox": [ + 212, + 643, + 789, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "M. Zhang et al.", + "bbox": [ + 271, + 114, + 377, + 128 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3 Modeling Interference to Entropy Coding", + "text_level": 1, + "bbox": [ + 215, + 143, + 655, + 162 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "While the resilience of DNNs to distribution shifts [7,19,24,25,58] and adversarial actions [3,11,12,30,35,56] has been extensively investigated, we investigate types of interference that can alter the length of the encoded bit stream, which ultimately leads to increased bandwidth utilization. In this section, we formulate the resilience of compression in entropy-coding-driven distributed DNNs.", + "bbox": [ + 212, + 175, + 787, + 251 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Threat Model for Accuracy. Let $D = \\{x, y\\}$ denotes a dataset for classification problem where $x \\in \\mathcal{X}$ and $y \\in \\mathcal{Y}$ are specific input and corresponding label samples respectively, while $\\mathcal{X}, \\mathcal{Y}$ are input and output space, respectively. There exists a prior yet unknown distribution $P(y|x)$ such that", + "bbox": [ + 212, + 271, + 787, + 333 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nP (y | x) \\geq P (\\hat {y} | x) \\quad \\forall \\hat {y} \\in \\mathcal {Y} \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 401, + 343, + 785, + 359 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\hat{y}$ denotes incorrect labels. The DNN classifier, is trained to approximate the distribution $\\tilde{P}(y|x) \\approx P(y|x)$ with empirical risk minimization approaches based on an assumption that in-sample data and out-of-sample data shared a similar distribution $P(y|x)$ . However, the unintentional interference denoted by $\\delta$ , which can be modeled as an additive noise to $x$ , creates a new dataset $D' = \\{x + \\delta, y\\}$ . It introduces a covariate shift that $P(x + \\delta) \\neq P(x)$ but $P(y|x + \\delta) = P(y|x)$ , resulting in classification errors such that", + "bbox": [ + 214, + 368, + 787, + 474 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {P} (\\hat {y} | x + \\delta) \\geq \\tilde {P} (y | x + \\delta) \\quad \\exists \\hat {y} \\in \\mathcal {Y} \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 372, + 484, + 785, + 501 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In the intentional interference scenario, algorithms attempt to find the minimum covariate shift that leads to Eq. (4). In this setting, a small perturbation in $l_{p}$ distance $||\\delta||_{p} \\leq \\epsilon$ is intentionally crafted to mislead the DNN classifier where $\\epsilon$ denotes the constraint. Attackers formulate Eq. (4) as an optimization problem over the input space $\\mathcal{X}$ , and hence can apply the gradient information of the loss function w.r.t. $x$ to find the $\\delta$ . Projected Gradient Descent (PGD) [35] provided a unified view on iterative gradient attacks under $l_{p}$ constraint:", + "bbox": [ + 212, + 511, + 787, + 617 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nx ^ {i + 1} = x ^ {i} + \\alpha \\cdot \\epsilon \\cdot \\frac {\\nabla_ {x} \\mathcal {L}}{| | \\nabla_ {x} \\mathcal {L} | | _ {p}} \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 406, + 627, + 785, + 659 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $x^{i}$ is the adversarial sample at $i$ step, $\\alpha$ is the learning rate, and $\\frac{\\nabla_{x}\\mathcal{L}}{||\\nabla_{x}\\mathcal{L}||_{p}}$ is the normalized gradient of a loss function $\\mathcal{L}$ (e.g., a sign function in $l_{\\infty}$ space). $\\mathcal{L}$ can be cross entropy or other more advanced loss functions [11].", + "bbox": [ + 214, + 667, + 787, + 719 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Threat Model for Data Rate. Let $H(x) = z$ denotes the head network where $z \\in \\mathcal{Z}$ is the output of the head and $\\mathcal{Z}$ is the latent space. The goal is to achieve a minimum compression rate $-\\log_2 P_Z(z)$ . As $z$ is dependent on $x$ , the distribution shift introduced by unintentional interference can result in a difference in coding rate:", + "bbox": [ + 212, + 738, + 787, + 814 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n- \\log_ {2} P _ {Z} (H (x + \\delta)) \\neq - \\log_ {2} P _ {Z} (H (x)) \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 359, + 825, + 785, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "Resilience of Entropy Model in Distributed Neural Networks", + "bbox": [ + 326, + 114, + 732, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/9a36f0770cb7ccb5b113b0f91198b1d751ff5b6f93b752db241ec5351f2081d3.jpg", + "table_caption": [ + "Table 1: DNNs, entropy models and datasets utilized in experiments." + ], + "table_footnote": [], + "table_body": "
DNN Architectures
ResNet-50 [23]A standard baseline in distributed DNNs [38-40, 42]
ResNet-101 [23]A deeper model in the ResNet family
RegNetY-6.4GF [47]An advanced design with better performance
Entropy Models
FP [5]A fully factorized prior model
MSHP [43]An effective learnable entropy model
Datasets
ImageNet [16]The standard ImageNet validation dataset
ImageNet-C [25]A synthetic dataset with 15 corruptions and 5 severities
Random NoiseNoisy images as a baseline for adversarial robustness
PGD-Acc [35]PGD targeting classification
PGD-EPGD targeting entropy
", + "bbox": [ + 243, + 171, + 759, + 393 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Note that the unintentional distribution shift of $x$ does not necessarily result in a larger bit rate. However, in the adversarial setting, attackers intend to maximize the entropy of $z$ as follows:", + "bbox": [ + 212, + 421, + 787, + 465 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {\\delta} \\log_ {2} P _ {Z} (H (x + \\delta)) \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 411, + 477, + 785, + 505 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\ns. t. \\quad | | \\delta | | _ {p} \\leq \\epsilon\n$$\n", + "text_format": "latex", + "bbox": [ + 418, + 503, + 519, + 518 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Thus, to correctly assess the robustness of entropy models, we use the same PGD algorithm in Eq. (5) yet with the entropy as its loss function $\\mathcal{L} = -\\log_2P_Z(H(x))$", + "bbox": [ + 212, + 527, + 792, + 560 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4 Understanding the Resilience of Entropy Model", + "text_level": 1, + "bbox": [ + 212, + 580, + 715, + 599 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1 Experimental Setup", + "text_level": 1, + "bbox": [ + 214, + 612, + 428, + 628 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this section, we outline the experimental setup for evaluating the resilience of entropy models. We first specify the DNNs and entropy models under consideration, followed by a description of the metrics and datasets used in our experiments. A summary of our experimental setup is provided in Tab. 1.", + "bbox": [ + 212, + 636, + 787, + 698 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Ablation of DNNs and Entropy Models. We consider three distinct DNNs: ResNet-50 [23], ResNet-101 [23] and RegNetY-6.4GF [47]. ResNet-50 serves as a common baseline in distributed DNN literature [38-40,42] while ResNet-101 is a deeper variant with the same design as former. RegNetY-6.4GF has an advanced design with better classification performance. The early layers are replaced by a specific tailored head architecture as proposed in [42] and incorporated with different entropy models. For the sake of brevity, we consider ResNet-50 as the default DNN architecture without additional specification. We employ two entropy", + "bbox": [ + 212, + 718, + 787, + 842 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "M. Zhang et al.", + "bbox": [ + 271, + 114, + 377, + 128 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/749e0e30e507829f0b3bba2cddea1397c49d6d7a843a8db6bd6add6a20458096.jpg", + "image_caption": [ + "(a) resilience w.r.t. $\\beta$" + ], + "image_footnote": [], + "bbox": [ + 248, + 143, + 493, + 296 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e54d2744cd2b3bda3702db2cc58fa4f3936dcf54b086ee1492e3bec496597862.jpg", + "image_caption": [ + "(b) resilience w.r.t. severities", + "Fig. 3: Resilience of accuracy and data rate to 4 common corruptions." + ], + "image_footnote": [], + "bbox": [ + 496, + 146, + 754, + 296 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "models: Factorized Prior (FP) [5] and Mean Scale Hyper Prior (MSHP) [43]. FP is the earliest design of entropy model and a basic component in many other advanced designs, while MSHP provides more powerful compression performance by injecting a hierarchical learnable block before FP to extract the entropy information. For conciseness, we consider MSHP as the default entropy model. In addition, since models are trained with different rate-distortion trade-off $\\beta$ , we also investigate the compression robustness w.r.t. different $\\beta$ .", + "bbox": [ + 212, + 364, + 787, + 472 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Datasets and Metrics. Conversely from existing work that measures the compression performance with bits per pixel (BPP) [5,6,14,43,44], we evaluate the entropy model with the data size of the whole bit stream after encoding as in [37]. Compared to BPP, data size quantifies more directly the networking traffic between the head and tail. To assess the resilience to unintentional interference, we use the ImageNet-C dataset proposed by [25] comprising 15 common corruptions classified into 4 categories (noise, blur, digital, and weather) with 5 different severities. We also report the data size and accuracy on the clean ImageNet [16] validation set as a baseline. To assess the resilience to intentional interference, we implement the PGD targeting both the classification and compression performance in $l_{\\infty}$ space. For clarity, we denote the conventional PGD as PGD-Acc and the PGD targeting entropy models as PGD-E. Meanwhile, random noise with the same perturbation level is added to the clean ImageNet as a baseline comparison for adversarial robustness.", + "bbox": [ + 212, + 492, + 787, + 705 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2 Resilience to Unintentional Interference", + "text_level": 1, + "bbox": [ + 215, + 724, + 591, + 739 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Resilience w.r.t. $\\beta$ . We select one corruption as a representative from each category in ImageNet-C [25] to investigate the resilience as a function of ratedistortion trade-off $\\beta$ . Fig. 3a shows the compression and classification performance on defocus blur, shot noise, snow and contrast dataset, each with the severity of 5. Compared to the clean ImageNet, all corruptions decrease the accuracy significantly. However, not all corruptions increase the data size. Only", + "bbox": [ + 212, + 750, + 787, + 840 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "Resilience of Entropy Model in Distributed Neural Networks", + "bbox": [ + 328, + 114, + 732, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/8be872c73f70ba0a8c4e4513c940128f241a103bbc273e23017aba14c4602d5b.jpg", + "table_caption": [ + "Table 2: Resilience w.r.t. prior models. The average data size of FP and MSHP for clean ImageNet are $11.65 \\pm 1.02$ and $9.62 \\pm 1.70$ KBBytes, respectively. Red for larger data size and blue for lower compared to the baseline." + ], + "table_footnote": [], + "table_body": "
Prior \nModelGaussian \nSize[KB]Noise \nAcc[%]Motion \nSize[KB]Blur \nAcc[%]Impulse \nSize[KB]Noise \nAcc[%]Glass \nSize[KB]Blur \nAcc[%]
FP13.19±0.4059.5110.51±0.7062.2213.36±0.3450.9310.43±0.6653.63
MSHP11.96±0.4759.367.88±1.4661.5812.52±0.3551.127.78±1.3953.67
", + "bbox": [ + 222, + 198, + 781, + 272 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "shot noise increases the data size by $65.31\\%$ on average. In contrast, Defocus blur and contrast decrease the data size by $53.31\\%$ and $67.45\\%$ on average, respectively. Snow has little affect to compression, resulting in only $4.42\\%$ decrease of data size. This indicates that the entropy model learns a different set of features than the classification model. Since it shows the same trend for different rate-distortion trade-off factors, we only consider $\\beta = 0.08$ in next experiments.", + "bbox": [ + 212, + 297, + 787, + 388 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Resilience w.r.t. Severities. Next, we explore resilience w.r.t. severity levels using the same corruptions mentioned earlier. As illustrated in Fig. 3b, the data size of shot noise rises with increasing severity, whereas defocus blur and contrast consistently decrease the data size. Snow shows minimal variation across different severity levels. This can be attributed to defocus blur and contrast removing high-frequency components in images, while shot noise introduces new patterns in high-frequency space. Snow, on the other hand, lacks a specific pattern in the frequency domain. Thus, the compression is particularly sensitive to high-frequency information in the input space.", + "bbox": [ + 212, + 410, + 787, + 547 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Resilience w.r.t. Prior Models. Tab. 2 shows the impact of additional 4 corruptions at severity level 1 on both FP and MSHP model. To further validate the previous observation, we choose 2 noise corruptions (Gaussian noise and impulse noise), which introduce high-frequency noise to images, and two blur corruptions (motion blur and glass blur), which eliminate high-frequency components from images. As shown in Tab. 2, Gaussian noise increases $13.22\\%$ and $24.32\\%$ of data size for FP and MSHP, respectively. Impulse noise increases $14.42\\%$ and $30.15\\%$ of data size for FP and MSHP, respectively. On the other hand, motion blur reduces $9.79\\%$ and $18.09\\%$ of data size for FP and MSHP, respectively. Glass blur reduces $10.47\\%$ and $19.13\\%$ of data size for FP and MSHP, respectively. Thus, the previous finding that compressive features reside in the high-frequency space remains consistent for different entropy models.", + "bbox": [ + 212, + 566, + 787, + 750 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.3 Resilience to Intentional Interference", + "text_level": 1, + "bbox": [ + 215, + 770, + 568, + 784 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Resilience w.r.t. DNNs Fig. 4a shows the resilience of different DNN architectures to accuracy and entropy attacks with the $l_{\\infty}$ constraint $\\epsilon = 4 / 255$ . As evident in the figure, the attacks show similar patterns regardless of their", + "bbox": [ + 212, + 794, + 787, + 839 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "M. Zhang et al.", + "bbox": [ + 271, + 114, + 377, + 128 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c1422fe811d72e379e187f90ad9483c30699ed87429a3cf3260ddb21859c5d72.jpg", + "image_caption": [ + "(a) resilience w.r.t. DNN architectures" + ], + "image_footnote": [], + "bbox": [ + 243, + 143, + 498, + 301 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/2fdb8b2a921a1e5acb43a4edba73095d1962eed910e547599bc7c07565b2c889.jpg", + "image_caption": [ + "(b) resilience w.r.t. prior models", + "Fig. 4: Resilience of accuracy and data rate to intentional interference" + ], + "image_footnote": [], + "bbox": [ + 501, + 146, + 759, + 300 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "victim models. While PGD-Acc reduces $42.88\\%$ accuracy on average, it only increases by $7.28\\%$ of data size on average compared to random noise. On the other hand, PGD-E decreases $2.34\\%$ accuracy on average compared to random noise but increases the bit rate 1.35x times on average. This indicates that PGD-E and PGD-Acc are targeting separate features in the input space which has minor impact on each other.", + "bbox": [ + 212, + 371, + 787, + 460 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Resilience w.r.t. Prior Models. Fig. 4b illustrates the effect of adversarial attacks with $\\epsilon = 8 / 255$ on victims using different prior models. As shown in the figure, both entropy models show considerable vulnerability to PGD-E compared to PGD-Acc. The data size are increased by $21.93\\%$ and $46.82\\%$ for FP and MSHP respectively while the performance loss are $9.91\\%$ and $6.62\\%$ respectively. On the other hand, PGD-Acc successfully reduces $72.10\\%$ and $57.10\\%$ classification performance of FP and MSHP compared to random noise respectively. However, it only increases $5.66\\%$ and $10.19\\%$ bit rate for FP and MSHP respectively. Hence, attackers targeting compression tend to perturb a different set of features than attackers targeting discriminative tasks regardless of the entropy models. In addition, while the FP has a better robustness in compression, it has significantly worse robustness in accuracy.", + "bbox": [ + 212, + 484, + 787, + 667 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "5 Proposed Defense Mechanism", + "text_level": 1, + "bbox": [ + 214, + 690, + 540, + 708 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "5.1 Disentangle Compression Features in Frequency Domain", + "text_level": 1, + "bbox": [ + 214, + 723, + 728, + 739 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "As shown in Sec. 4.2, interference that introduces high-frequency noise (e.g., shot noise) can effectively increase the data size while interference that removes high-frequency information (e.g., defocus blur) reduces the data size. Intuitively this is a consequence of the small head in distributed DNNs as semantic information is usually captured in deeper layers while early layers tend to capture low level information [8]. To validate our intuition, we introduce total variation, defined", + "bbox": [ + 212, + 750, + 787, + 840 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "Resilience of Entropy Model in Distributed Neural Networks", + "bbox": [ + 328, + 114, + 732, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/8cd8faf40432ed1024ee9ac561f449d17f469e27643b54f9ea85ec2200e8c119.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 251, + 146, + 331, + 208 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/d0636016d8cacc5ccf4eb783d04afd96f3e7ad8320ae5d84c517527343444d60.jpg", + "image_caption": [ + "Fig. 5: Entropy models are sensitive to high-frequency features. From left to right: images from ImageNet validation set, bit rate maps and total variation maps." + ], + "image_footnote": [], + "bbox": [ + 251, + 210, + 331, + 271 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/be791e5b2e7d3b8a3a3024edacae1bae51cc2b4333197a34a697c1a2042fe377.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 334, + 146, + 415, + 208 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/b7634bdb8b1a114c0539bd8177bf72e1cec1176c0bf8e3781ca928186e8a3dfb.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 334, + 210, + 415, + 271 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/3006b3bbba65f4a2dd721196d2cd4fae009fca8f33c317c9682df03784ea6d6c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 421, + 146, + 500, + 208 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/d31bd2c70df5c30176ed942294fe5476b0fc9e7d7c1731b84f8791d293d2b878.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 421, + 210, + 500, + 271 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/3ae55c9591f3169f26560aa464c58fd4bcd61295e1a0316061af2c26227f1df0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 504, + 146, + 584, + 208 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/7d7c7698af0229855a4a33345826e806f3880bf4ba5a76002209c13e29b5be08.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 504, + 210, + 584, + 271 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/e2331d1baeff47774cf21ff0223a1803f1f80f68192ad5c4b04b6c6b6432331b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 589, + 146, + 668, + 208 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/59b7d051e5b9660d0b36ec34203cecec281c924a50f04b35f49d529116c67fc4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 589, + 210, + 668, + 271 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/f09587c113ad775d597b521159ddf19a1f4bfc4bf9757a2d0a514404d9be0c54.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 673, + 146, + 753, + 208 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/180fe7f9967c9d145a06778637bf0f9f8693da31097a41269fe68c12ed915f5d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 673, + 210, + 753, + 271 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "as the integral of image gradient magnitude, which is first proposed for image denoising in [51]. The anitrosopic total variation of a 2-D image is defined as", + "bbox": [ + 214, + 344, + 782, + 376 + ], + "page_idx": 9 + }, + { + "type": "equation", + "text": "\n$$\nT V (x) = \\sum_ {i, j} \\left| x _ {i + 1, j} - x _ {i, j} \\right| + \\left| x _ {i, j + 1} - x _ {i, j} \\right| \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 344, + 386, + 785, + 419 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "where $x$ denotes an image sample, $i,j$ denotes the pixel position in width and height, respectively. The image gradient magnitude $|x_{i+1,j} - x_{i,j}|$ and $|x_{i,j+1} - x_{i,j}|$ describe the sharpness of pixel changes, where larger magnitude indicates high-frequency information such as edges and textures in images.", + "bbox": [ + 214, + 429, + 782, + 491 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "We compare the total variation map of an image with its bit rate map generated by the entropy model. To plot the total variation map, we first slice an image into many small patches and then compute the total variation for each patch. Meanwhile, the bit rate map is the estimated entropy of the latent representation $-\\log_2P_Z(z)$ . As shown in Fig. 5, there exists a significant correlation between the bit rate map and the total variation map, which indicates that the entropy model is sensitive to high-frequency features in the input space.", + "bbox": [ + 212, + 489, + 784, + 597 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "5.2 Disentangle Compression Features in Spatial Space", + "text_level": 1, + "bbox": [ + 214, + 618, + 683, + 633 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "The work in [42] has shown that entropy model in distributed DNNs tends to assign a larger number of bits to task-oriented information, such as objects, while compressing irrelevant information, like the background. Consequently, it becomes intuitively easier for an adversary to increase the bit rate in the background area to attack the entropy model.", + "bbox": [ + 212, + 643, + 782, + 718 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "To interpret the action conducted by attackers, we generate the bit rate comparison map between adversarial samples generated by PGD-E and PGD-Acc. This comparison map represents the estimated entropy of adversarial samples generated by PGD-E subtracted from the bit rate of PGD-Acc-perturbed samples. In this context, the red area indicates that PGD-E tends to allocate more bits compared to PGD-Acc, whereas the blue region suggests that PGD-E tends to allocate fewer bits. As depicted in Fig. 6, PGD-E puts more efforts to increase the bit rate in background region, whereas PGD-Acc focuses on altering", + "bbox": [ + 212, + 719, + 784, + 839 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "M. Zhang et al.", + "bbox": [ + 271, + 114, + 377, + 128 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/f4b643126eed029e66abdfbf9d68d09b122c66ab69117af0bf015a834d40eddb.jpg", + "image_caption": [ + "Fig. 6: Entropy models are more vulnerable to perturbation in background region. From left to right: images from ImageNet validation set, bit rate comparison between PGD-E and PGD-Acc (PGD-E allocate less bit in blue area and more bit in red area)." + ], + "image_footnote": [], + "bbox": [ + 251, + 146, + 326, + 205 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/5e54bca6224bab3090b457b056900a27ea21b7a0ef4fe1d7bcd69af0154c4e99.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 334, + 146, + 413, + 205 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/fd4b0018db2a647c063c2563f557969b90e7d2619cea840d2880685b297bdbda.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 421, + 146, + 498, + 205 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/83b8dd48844c0a14b260f22c8f8b706d357945e0ac4d712130a7d2d9f3d48562.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 506, + 146, + 583, + 205 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/25da3dd17cf80b197cf2144dd097b4648a714051e107f57cfd08bd79aed2d207.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 589, + 146, + 668, + 205 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/2669ece1f428339965f782c3128b1b89cad20ebd9895799225f4316d6befe9c1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 676, + 146, + 753, + 205 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "task-oriented information. This is in line with the experiments in Sec. 4.3 that PGD-E and PGD-Acc target different features in the input space. On one hand, increasing the bit rate in the background region will affect little accuracy as they are irrelevant information for the discriminative task. On the other hand, modifying object information can significantly degrade the performance, yet it has less impact on compression.", + "bbox": [ + 212, + 290, + 782, + 380 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "5.3 Object-Aware Total Variation Denoising", + "text_level": 1, + "bbox": [ + 214, + 398, + 591, + 415 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Given the observation in Sec. 5.1, we propose a denoising technique based on total variation to remove high-frequency noise in adversarial images where we solve the following optimization problem:", + "bbox": [ + 212, + 420, + 782, + 465 + ], + "page_idx": 10 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {x} \\frac {1}{2} \\| x - x ^ {\\prime} \\| _ {2} ^ {2} + \\lambda \\cdot T V (x), \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 395, + 472, + 784, + 501 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "where $x'$ is the distorted image, $x$ is the denoised image, $TV(x)$ is the total variation of $x$ and $\\lambda$ is a regularization factor to control the degree of smoothing. Eq. (9) means the image $x$ after denoising should keep most of the information in $x'$ (i.e., minimize $||x - x'||_2^2$ ) but discard the high-frequency information (i.e., minimize $TV(x)$ ). Optimizing Eq. (9) is not trivial as total variation is non-differentiable. As a result, we use a sub-gradient descent:", + "bbox": [ + 212, + 506, + 782, + 595 + ], + "page_idx": 10 + }, + { + "type": "equation", + "text": "\n$$\nx ^ {i + 1} = x ^ {i} - \\alpha \\cdot \\left(\\left(x ^ {i} - x ^ {\\prime}\\right) + \\lambda \\cdot g \\left(x ^ {i}\\right)\\right) \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 366, + 603, + 784, + 621 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "where $x^i$ is the denoised image at the step $i$ , $\\alpha$ is the step size and $g(x^i)$ is the image gradient of $x^i$ [9].", + "bbox": [ + 212, + 627, + 782, + 657 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Simply applying the total variation denoising will also decrease the task-oriented performance as it also removes useful information in high-frequency domain. Since adversarial attacks targeting entropy models attempt to increase bit rate in background region as demonstrated in Sec. 5.2, we add a mask to Eq. (10) to force the denoising algorithm to only remove high-frequency information in non-object space, that is,", + "bbox": [ + 212, + 657, + 784, + 748 + ], + "page_idx": 10 + }, + { + "type": "equation", + "text": "\n$$\nx ^ {i + 1} = x ^ {i} - \\alpha \\cdot m \\cdot \\left(\\left(x ^ {i} - x ^ {\\prime}\\right) + \\lambda \\cdot g \\left(x ^ {i}\\right)\\right) \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 352, + 753, + 784, + 772 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "where $m$ is the mask to control the denoising level. In practice, we interpolate the output of entropy model $P_Z(z)$ as the soft mask. This is because the higher bit rate $-\\log_2P_Z(z)$ corresponding to object area has a smaller value of $P_Z(z)$ , making it naturally a soft mask to avoid smoothing the object region.", + "bbox": [ + 212, + 779, + 782, + 839 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "Resilience of Entropy Model in Distributed Neural Networks", + "bbox": [ + 328, + 114, + 732, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 116, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/7f72dae5f3f2da2c179f62d944675f625c1808786897677a5f915f32c6dd20b4.jpg", + "table_caption": [ + "Table 3: Compression and classification performance before and after defense. The average data size for clean ImageNet is $9.62 \\pm 1.70$ KBytes." + ], + "table_footnote": [], + "table_body": "
ε = 2/255ε = 4/255ε = 8/255ε = 16/255
Size[KB]Acc[%]Size[KB]Acc[%]Size[KB]Acc[%]Size[KB]Acc[%]
before12.35±0.6373.2614.11±0.3871.3316.15±0.2965.6218.79±0.3752.13
after8.76±1.3870.829.55±1.0770.1511.02±0.6567.2213.44±0.4859.87
", + "bbox": [ + 215, + 184, + 789, + 258 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Experimental Results. We investigate the proposed defense method as a function of perturbation budget $\\epsilon$ . As shown in Tab. 3, attackers targeting entropy significantly increase the data size, by up to $95.32\\%$ when $\\epsilon = 16/255$ compared to the clean images (9.62 KBytes). On the contrary, our denoising approach significantly reduce the data size for all $\\epsilon$ . For $\\epsilon = 2/255, 4/255$ , the average data size is even $8.94\\%$ and $0.73\\%$ smaller than the clean images with only $2.44\\%$ and $1.18\\%$ accuracy loss, respectively. For $\\epsilon = 8/255, 16/255$ , the data size is reduced by $31.80\\%$ and $28.47\\%$ respectively, while the accuracy improved slightly after denoising ( $1.60\\%$ for $\\epsilon = 8/255$ and $7.74\\%$ for $\\epsilon = 16/255$ ).", + "bbox": [ + 212, + 285, + 785, + 420 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "We remark that the proposed defense is a standalone approach which can be incorporated into other approaches such as adversarial training to further improve the resilience. The accuracy loss incurred in small perturbation scenarios $(\\epsilon = 2 / 255, 4 / 255)$ can be mitigated by training with augmented data.", + "bbox": [ + 212, + 421, + 785, + 482 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "5.4 Adaptive Attacks", + "text_level": 1, + "bbox": [ + 215, + 502, + 405, + 517 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "To thoroughly evaluate the effectiveness of a defense strategy, [10] proposed to design adaptive attacks that counter the defense mechanism with perfect knowledge of both DNN and defense in the white-box setting. To this end, we assess our defense with two adaptive attacks.", + "bbox": [ + 212, + 525, + 785, + 585 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Low Frequency Attack. As the total variation denoising is proposed to remove high-frequency components, the first adaptive attack that we considered is to only add perturbations in low frequency space. Following the methodology outlined in [21], we first convert the gradient to frequency space with discrete cosine transform and mask out the high-frequency components. The gradient is then transformed back and Eq. (5) is applied to create the adversarial samples.", + "bbox": [ + 212, + 607, + 785, + 699 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Regional Attack. Given that the defense exclusively denoises the background region, the second adaptive attack we considered is to force the adversary put greater effort in increasing bit rate in the object region. Similar to Eq. (11), a mask is multiplied to the loss function before backpropagation. In practice, we choose $1 - P_Z(z)$ as the soft mask where $P_Z(z)$ is the probability of $z$ ranging between 0 and 1. A larger $P_Z(z)$ indicates a smaller bit rate $-\\log_2P_Z(z)$ corresponding to the background region. Therefore, $1 - P_Z(z)$ encourages the adversary to take more attention to object region.", + "bbox": [ + 212, + 719, + 787, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "M. Zhang et al.", + "bbox": [ + 271, + 114, + 377, + 128 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/9fe1f34475416dab50c98311d15694be740c94249ca47763259021abb6e9ab3b.jpg", + "table_caption": [ + "Table 4: Resilience to adaptive attacks with $\\epsilon = 4 / 255$" + ], + "table_footnote": [], + "table_body": "
Adaptive AttacksBefore DenoisingAfter Denoising
Size[KB]Acc[%]Size[KB]Acc[%]
Frequency14.11±0.3871.389.56±1.0770.17
Regional14.05±0.4171.639.50±1.1070.22
", + "bbox": [ + 339, + 171, + 658, + 243 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Resilience to Adaptive Attacks. Tab. 4 shows results of adaptive attacks with perturbation budget $\\epsilon = 4 / 255$ . Our defense successfully reduces the average data size by $67.75\\%$ and $67.62\\%$ for low frequency attack and regional attack respectively while the accuracy only decreases $1.21\\%$ and $1.41\\%$ respectively. Thus, the proposed approach is resilient to adaptive attacks.", + "bbox": [ + 212, + 268, + 787, + 347 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "6 Related Work", + "text_level": 1, + "bbox": [ + 214, + 366, + 387, + 383 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Distributed Deep Neural Network. Distributed DNN is proposed to meet the challenge of deploying artificial intelligence to resource-constrained platforms such as mobile devices and Internet of Things (IoT) devices [41]. To minimize the end-to-end latency across devices, dimension reduction designs (i.e., bottlenecks) are introduced to compress the data size of intermediate representations that need to be sent [18, 54].", + "bbox": [ + 212, + 396, + 785, + 487 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "However, unlike autoencoders [32,36,49] that have symmetric designs on both encoder and decoder sides, distributed DNNs usually have asymmetric designs due to the limit of computation resources, and thus inject the bottleneck in early layers [18, 38-40, 54]. As a result, naively end-to-end trained distributed DNNs have noticeable performance loss [18, 54]. To preserve accuracy, [38] proposed to use knowledge distillation to train the distributed DNN separately while [39] proposed a multi-stage training approach for each part of the DNN.", + "bbox": [ + 212, + 487, + 785, + 593 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Along with bottleneck-based distributed DNNs, different coding-based approaches are also applied to reduce the data size of the latent representations. [2] proposed to apply codec-based compression such as JPEG to latent representations and [15] adopted spatio-temporal coding such as HEVC to compress the streaming latent features in video tasks. [42,55,59] use entropy coding to achieve a higher compression ratio of latent representations.", + "bbox": [ + 212, + 593, + 785, + 684 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Entropy Coding in Deep Data Compression. [5] first introduced a fully factorized prior to integrate entropy coding with variational autoencoder, showcasing superior quality enhancement over conventional codec-based methods like JPEG for image compression. Building upon this, [6] extended the conventional factorized prior model to a learnable hierarchical hyper prior, resulting in improved performance. [43] proposed a joint auto-regressive and hyper prior design while [44] proposed a channel-aware entropy coding scheme. In addition, [14] introduced the attention mechanism to prior models and [1] extended the technique to video compression.", + "bbox": [ + 212, + 703, + 787, + 840 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "Resilience of Entropy Model in Distributed Neural Networks", + "bbox": [ + 326, + 113, + 730, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 784, + 126 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "To minimize the communication overhead in distributed DNN scenarios, [55] for the first time adopted the entropy coding to further compress the latent representations after the bottleneck while [59] extended the approach to object detection. However, these approaches overlooked resource constraints, resulting in over-complex network designs for mobile devices. Conversely, [42] integrated entropy coding with a multi-stage training strategy [39], aiming to optimize a tripartite trade-off encompassing computation resources in mobile devices, communication overhead between mobile and edge devices, and end-to-end performance of neural networks.", + "bbox": [ + 212, + 146, + 787, + 282 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Resilience of DNN efficiency. While the reliability of DNNs have been extensively investigated, limited attention has been devoted to interference that undermine their efficiency. [22] explored this direction by demonstrating that imperceptible perturbations to input, leveraging intermediate representations of DNNs, could nullify the computation savings achieved by dynamic depth neural networks, which adapt their depth based on input complexity. [46] extended these attacks to target both dynamic depth and dynamic width neural networks. Moreover, in [45], the authors proposed simultaneous adjustments to the direction and magnitude of attacks, enhancing their effectiveness. [27] revealed that early exit dynamic DNNs could be tricked into late-stage inferences, significantly slowing down inference speed.", + "bbox": [ + 212, + 303, + 787, + 469 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "The study of efficiency vulnerabilities in DNNs has also ventured into real-world deployment scenarios. [33] focused on attacking LiDAR-based detection systems, introducing latency in detections and exposing vulnerabilities in critical contexts such as self-driving cars. [13] demonstrated that neural image caption generation models could be manipulated to incur increased computation costs by inducing unnecessary decoder calls for token generation.", + "bbox": [ + 212, + 470, + 787, + 560 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In star contrast to existing research, our work investigates a novel threat emerging in distributed DNNs. Here, adversaries not only compromise the communication efficiency of entropy-coded distributed DNNs but also pose a threat to other users by saturating the transmission bandwidth.", + "bbox": [ + 212, + 560, + 787, + 621 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "7 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 643, + 359, + 659 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "This paper has investigated the resilience of entropy models in distributed DNNs against both intentional and unintentional interference. We conducted thorough evaluations using 3 different DNN architectures, 2 entropy model designs, and 4 rate-distortion trade-off factors with common corruption datasets and adversarial attacks. Our analysis disentangled compression features in both spatial and frequency domains, revealing vulnerabilities of the entropy model to specific types of perturbations. Building on these findings, we proposed a standalone defense strategy aimed at reducing data size with minimal task-oriented performance loss. Our future work will focus on designing more advanced defense approaches for distributed DNNs that are resilient in both compression and classification tasks.", + "bbox": [ + 212, + 672, + 787, + 839 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "M. Zhang et al.", + "bbox": [ + 271, + 114, + 377, + 128 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 217, + 143, + 401, + 162 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "This work is funded in part by the Air Force Office of Scientific Research under contract number FA9550-23-1-0261, by the Office of Naval Research under contract number N00014-23-1-2221, as well as by National Science Foundation grants CNS-2134973 and CNS-2312875. Any opinions, findings and conclusions or recommendations expressed in this material are those of the authors and do not necessarily reflect the views of the U.S. Government.", + "bbox": [ + 212, + 172, + 787, + 263 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 215, + 282, + 321, + 299 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Agustsson, E., Minnen, D., Johnston, N., Balle, J., Hwang, S.J., Toderici, G.: Scale-space flow for end-to-end optimized video compression. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8503-8512 (2020)", + "2. Alvar, S.R., Bajic, I.V.: Pareto-optimal bit allocation for collaborative intelligence. IEEE Transactions on Image Processing 30, 3348-3361 (2021)", + "3. Andriushchenko, M., Croce, F., Flammarion, N., Hein, M.: Square attack: a query-efficient black-box adversarial attack via random search. In: European conference on computer vision. pp. 484-501. Springer (2020)", + "4. Arvinte, M., Cornelius, C., Martin, J., Himayat, N.: Investigating the adversarial robustness of density estimation using the probability flow ode. arXiv preprint arXiv:2310.07084 (2023)", + "5. Balle, J., Laparra, V., Simoncelli, E.P.: End-to-end optimized image compression. In: 5th International Conference on Learning Representations, ICLR 2017 (2017)", + "6. Balle, J., Minnen, D., Singh, S., Hwang, S.J., Johnston, N.: Variational image compression with a scale hyperprior. In: International Conference on Learning Representations (2018)", + "7. Barbu, A., Mayo, D., Alverio, J., Luo, W., Wang, C., Gutfreund, D., Tenenbaum, J., Katz, B.: Objectnet: A large-scale bias-controlled dataset for pushing the limits of object recognition models. Advances in neural information processing systems 32 (2019)", + "8. Bau, D., Zhou, B., Khosla, A., Oliva, A., Torralba, A.: Network dissection: Quantifying interpretability of deep visual representations. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 6541-6549 (2017)", + "9. Boigné, E., Parkinson, D.Y., Ihme, M.: Towards data-informed motion artifact reduction in quantitative ct using piecewise linear interpolation. IEEE Transactions on Computational Imaging 8, 917-932 (2022)", + "10. Carlini, N., Wagner, D.: Adversarial examples are not easily detected: Bypassing ten detection methods. In: Proceedings of the 10th ACM workshop on artificial intelligence and security. pp. 3-14 (2017)", + "1. Carlini, N., Wagner, D.: Towards evaluating the robustness of neural networks. In: 2017 ieee symposium on security and privacy (sp). pp. 39-57. IEEE (2017)", + "2. Chen, J., Jordan, M.I., Wainwright, M.J.: Hopskipjumpattack: A query-efficient decision-based attack. In: 2020 ieee symposium on security and privacy (sp). pp. 1277-1294. IEEE (2020)", + "3. Chen, S., Song, Z., Haque, M., Liu, C., Yang, W.: Nicgslowdown: Evaluating the efficiency robustness of neural image caption generation models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15365-15374 (2022)" + ], + "bbox": [ + 225, + 309, + 785, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "Resilience of Entropy Model in Distributed Neural Networks", + "bbox": [ + 326, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "14. Cheng, Z., Sun, H., Takeuchi, M., Katto, J.: Learned image compression with discretized gaussian mixture likelihoods and attention modules. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 7939-7948 (2020)", + "15. Choi, H., Bajic, I.V.: Deep feature compression for collaborative object detection. In: 2018 25th IEEE International Conference on Image Processing (ICIP). pp. 3743-3747. IEEE (2018)", + "16. Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: Imagenet: A large-scale hierarchical image database. In: 2009 IEEE conference on computer vision and pattern recognition. pp. 248-255. IEEE (2009)", + "17. Duda, J.: Asymmetric numeral systems: entropy coding combining speed of huffman coding with compression rate of arithmetic coding. arXiv preprint arXiv:1311.2540 (2013)", + "18. Eshratifar, A.E., Esmaili, A., Pedram, M.: Bottlenet: A deep learning architecture for intelligent mobile cloud computing services. In: 2019 IEEE/ACM International Symposium on Low Power Electronics and Design (ISLPED). pp. 1-6. IEEE (2019)", + "19. Geirhos, R., Rubisch, P., Michaelis, C., Bethge, M., Wichmann, F.A., Brendel, W.: Imagenet-trained cnns are biased towards texture; increasing shape bias improves accuracy and robustness. In: International Conference on Learning Representations (2018)", + "20. Goodfellow, I.J., Shlens, J., Szegedy, C.: Explaining and harnessing adversarial examples. arXiv preprint arXiv:1412.6572 (2014)", + "21. Guo, C., Frank, J.S., Weinberger, K.Q.: Low frequency adversarial perturbation. arXiv preprint arXiv:1809.08758 (2018)", + "22. Haque, M., Chauhan, A., Liu, C., Yang, W.: Ilfo: Adversarial attack on adaptive neural networks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14264-14273 (2020)", + "23. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 770-778 (2016)", + "24. Hendrycks, D., Basart, S., Mu, N., Kadavath, S., Wang, F., Dorundo, E., Desai, R., Zhu, T., Parajuli, S., Guo, M., et al.: The many faces of robustness: A critical analysis of out-of-distribution generalization. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 8340-8349 (2021)", + "25. Hendrycks, D., Dietterich, T.: Benchmarking neural network robustness to common corruptions and perturbations. In: International Conference on Learning Representations (2018)", + "26. Hinton, G., Vinyals, O., Dean, J.: Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531 (2015)", + "27. Hong, S., Kaya, Y., Modoranu, I.V., Dumitraş, T.: A panda? no, it's a sloth: Slowdown attacks on adaptive multi-exit neural network inference. arXiv preprint arXiv:2010.02432 (2020)", + "28. Howard, A., Sandler, M., Chu, G., Chen, L.C., Chen, B., Tan, M., Wang, W., Zhu, Y., Pang, R., Vasudevan, V., et al.: Searching for mobilenetv3. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 1314-1324 (2019)", + "29. Howard, A.G., Zhu, M., Chen, B., Kalenichenko, D., Wang, W., Weyand, T., Andreetto, M., Adam, H.: Mobilenets: Efficient convolutional neural networks for mobile vision applications. arXiv preprint arXiv:1704.04861 (2017)", + "30. Ilyas, A., Engstrom, L., Athalye, A., Lin, J.: Black-box adversarial attacks with limited queries and information. In: International conference on machine learning. pp. 2137-2146. PMLR (2018)" + ], + "bbox": [ + 215, + 146, + 785, + 840 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "M. Zhang et al.", + "bbox": [ + 271, + 114, + 377, + 128 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "31. Jacob, B., Kligys, S., Chen, B., Zhu, M., Tang, M., Howard, A., Adam, H., Kalenichenko, D.: Quantization and training of neural networks for efficient integer-arithmetic-only inference. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 2704-2713 (2018)", + "32. Kingma, D.P., Welling, M.: Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114 (2013)", + "33. Liu, H., Wu, Y., Yu, Z., Vorobeychik, Y., Zhang, N.: Slowlidar: Increasing the latency of lidar-based detection using adversarial examples. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5146-5155 (2023)", + "34. Liu, P., Qi, B., Banerjee, S.: Edgeeye: An edge service framework for real-time intelligent video analytics. In: Proceedings of the 1st international workshop on edge systems, analytics and networking. pp. 1-6 (2018)", + "35. Madry, A., Makelov, A., Schmidt, L., Tsipras, D., Vladu, A.: Towards deep learning models resistant to adversarial attacks. In: International Conference on Learning Representations (2018)", + "36. Masci, J., Meier, U., Ciresan, D., Schmidhuber, J.: Stacked convolutional autoencoders for hierarchical feature extraction. In: Artificial Neural Networks and Machine Learning-ICANN 2011: 21st International Conference on Artificial Neural Networks, Espoo, Finland, June 14-17, 2011, Proceedings, Part I 21. pp. 52-59. Springer (2011)", + "37. Matsubara, Y., Yang, R., Levorato, M., Mandt, S.: Sc2 benchmark: Supervised compression for split computing. Transactions on machine learning research (2023)", + "38. Matsubara, Y., Callegaro, D., Baidya, S., Levorato, M., Singh, S.: Head network distillation: Splitting distilled deep neural networks for resource-constrained edge computing systems. IEEE Access 8, 212177-212193 (2020)", + "39. Matsubara, Y., Callegaro, D., Singh, S., Levorato, M., Restuccia, F.: Bottlefit: Learning compressed representations in deep neural networks for effective and efficient split computing. In: 2022 IEEE 23rd International Symposium on a World of Wireless, Mobile and Multimedia Networks (WoWMoM). pp. 337-346. IEEE (2022)", + "40. Matsubara, Y., Levorato, M.: Neural compression and filtering for edge-assisted real-time object detection in challenged networks. In: 2020 25th International Conference on Pattern Recognition (ICPR). pp. 2272-2279. IEEE (2021)", + "41. Matsubara, Y., Levorato, M., Restuccia, F.: Split computing and early exiting for deep learning applications: Survey and research challenges. ACM Computing Surveys 55(5), 1-30 (2022)", + "42. Matsubara, Y., Yang, R., Levorato, M., Mandt, S.: Supervised compression for resource-constrained edge computing systems. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 2685-2695 (2022)", + "43. Minnen, D., Balle, J., Toderici, G.D.: Joint autoregressive and hierarchical priors for learned image compression. Advances in neural information processing systems 31 (2018)", + "44. Minnen, D., Singh, S.: Channel-wise autoregressive entropy models for learned image compression. In: 2020 IEEE International Conference on Image Processing (ICIP). pp. 3339-3343. IEEE (2020)", + "45. Pan, J., Foo, L.G., Zheng, Q., Fan, Z., Rahmani, H., Ke, Q., Liu, J.: Gradmdm: Adversarial attack on dynamic networks. IEEE Transactions on Pattern Analysis and Machine Intelligence (2023)" + ], + "bbox": [ + 212, + 146, + 787, + 839 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "Resilience of Entropy Model in Distributed Neural Networks", + "bbox": [ + 326, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "46. Pan, J., Zheng, Q., Fan, Z., Rahmani, H., Ke, Q., Liu, J.: Gradauto: Energy-oriented attack on dynamic neural networks. In: European Conference on Computer Vision. pp. 637-653. Springer (2022)", + "47. Radosavovic, I., Kosaraju, R.P., Girshick, R., He, K., Dóllár, P.: Designing network design spaces. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10428-10436 (2020)", + "48. Ran, X., Chen, H., Zhu, X., Liu, Z., Chen, J.: Deepdecision: A mobile deep learning framework for edge video analytics. In: IEEE INFOCOM 2018-IEEE conference on computer communications. pp. 1421-1429. IEEE (2018)", + "49. Rifai, S., Vincent, P., Muller, X., Glorot, X., Bengio, Y.: Contractive auto-encoders: Explicit invariance during feature extraction. In: Proceedings of the 28th international conference on international conference on machine learning. pp. 833-840 (2011)", + "50. Rissanen, J., Langdon, G.: Universal modeling and coding. IEEE Transactions on Information Theory 27(1), 12-23 (1981)", + "51. Rudin, L.I., Osher, S., Fatemi, E.: Nonlinear total variation based noise removal algorithms. Physica D: nonlinear phenomena 60(1-4), 259-268 (1992)", + "52. Sandler, M., Howard, A., Zhu, M., Zhmoginov, A., Chen, L.C.: Mobilenetv2: Inverted residuals and linear bottlenecks. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 4510-4520 (2018)", + "53. Shannon, C.E.: A mathematical theory of communication. The Bell system technical journal 27(3), 379-423 (1948)", + "54. Shao, J., Zhang, J.: Bottlenet++: An end-to-end approach for feature compression in device-edge co-inference systems. In: 2020 IEEE International Conference on Communications Workshops (ICC Workshops). pp. 1-6. IEEE (2020)", + "55. Singh, S., Abu-El-Haija, S., Johnston, N., Balle, J., Shrivastava, A., Toderici, G.: End-to-end learning of compressible features. In: 2020 IEEE International Conference on Image Processing (ICIP). pp. 3349-3353. IEEE (2020)", + "56. Szegedy, C., Zaremba, W., Sutskever, I., Bruna, J., Erhan, D., Goodfellow, I., Fergus, R.: Intriguing properties of neural networks. In: 2nd International Conference on Learning Representations, ICLR 2014 (2014)", + "57. Tan, M., Chen, B., Pang, R., Vasudevan, V., Sandler, M., Howard, A., Le, Q.V.: Mnasnet: Platform-aware neural architecture search for mobile. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 2820-2828 (2019)", + "58. Wang, H., Ge, S., Lipton, Z., Xing, E.P.: Learning robust global representations by penalizing local predictive power. Advances in Neural Information Processing Systems 32 (2019)", + "59. Yuan, Z., Rawlekar, S., Garg, S., Erkip, E., Wang, Y.: Feature compression for rate constrained object detection on the edge. In: 2022 IEEE 5th International Conference on Multimedia Information Processing and Retrieval (MIPR). pp. 1-6. IEEE (2022)", + "60. Zhang, W., Zhang, Z., Zeadally, S., Chao, H.C., Leung, V.C.: Masm: A multiple-algorithm service model for energy-delay optimization in edge artificial intelligence. IEEE Transactions on Industrial Informatics 15(7), 4216-4224 (2019)" + ], + "bbox": [ + 215, + 147, + 784, + 770 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "M. Zhang et al.", + "bbox": [ + 271, + 114, + 377, + 128 + ], + "page_idx": 17 + } +] \ No newline at end of file diff --git a/2024/Resilience of Entropy Model in Distributed Neural Networks/ad486483-132b-4d7f-84db-26c1a65f3dae_model.json b/2024/Resilience of Entropy Model in Distributed Neural Networks/ad486483-132b-4d7f-84db-26c1a65f3dae_model.json new file mode 100644 index 0000000000000000000000000000000000000000..a5822a977512024157d48a4fba00b8c54bbd478b --- /dev/null +++ b/2024/Resilience of Entropy Model in Distributed Neural Networks/ad486483-132b-4d7f-84db-26c1a65f3dae_model.json @@ -0,0 +1,2722 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.325, + 0.141, + 0.681, + 0.185 + ], + "angle": 0, + "content": "Resilience of Entropy Model in Distributed Neural Networks" + }, + { + "type": "text", + "bbox": [ + 0.342, + 0.213, + 0.661, + 0.244 + ], + "angle": 0, + "content": "Milin Zhang, Mohammad Abdi, Shahriar Rifat, and Francesco Restuccia" + }, + { + "type": "text", + "bbox": [ + 0.284, + 0.256, + 0.72, + 0.298 + ], + "angle": 0, + "content": "Institute for the Wireless Internet of Things Northeastern University, Boston MA 02115, USA {zhang.mil, abdi.mo, rifat.s, f/restuccia}@northeastern.com" + }, + { + "type": "text", + "bbox": [ + 0.263, + 0.332, + 0.744, + 0.637 + ], + "angle": 0, + "content": "Abstract. Distributed deep neural networks (DNNs) have emerged as a key technique to reduce communication overhead without sacrificing performance in edge computing systems. Recently, entropy coding has been introduced to further reduce the communication overhead. The key idea is to train the distributed DNN jointly with an entropy model, which is used as side information during inference time to adaptively encode latent representations into bit streams with variable length. To the best of our knowledge, the resilience of entropy models is yet to be investigated. As such, in this paper we formulate and investigate the resilience of entropy models to intentional interference (e.g., adversarial attacks) and unintentional interference (e.g., weather changes and motion blur). Through an extensive experimental campaign with 3 different DNN architectures, 2 entropy models and 4 rate-distortion trade-off factors, we demonstrate that the entropy attacks can increase the communication overhead by up to \\(95\\%\\). By separating compression features in frequency and spatial domain, we propose a new defense mechanism that can reduce the transmission overhead of the attacked input by about \\(9\\%\\) compared to unperturbed data, with only about \\(2\\%\\) accuracy loss. Importantly, the proposed defense mechanism is a standalone approach which can be applied in conjunction with approaches such as adversarial training to further improve robustness. Code is available at https://github.com/Restuccia-Group/EntropyR." + }, + { + "type": "text", + "bbox": [ + 0.263, + 0.65, + 0.695, + 0.664 + ], + "angle": 0, + "content": "Keywords: Trustworthy Machine Learning \\(\\cdot\\) Data Compression" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.689, + 0.377, + 0.705 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.719, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Distributed deep neural network (DNN) were recently introduced to divide the computation of DNNs across various devices based on their available computation and communication resources. They have been shown to be extremely effective to implement deep learning applications on resource-constrained mobile devices [18,39,54]. As depicted in Fig. 1, the common strategy is to divide a large DNN into a small head network deployed on the mobile device to extract and compress features, and a tail network on the server to perform the task inference. Compared to conventional lightweight DNNs specifically designed for" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.378, + 0.129 + ], + "angle": 0, + "content": "M. Zhang et al." + }, + { + "type": "image", + "bbox": [ + 0.245, + 0.145, + 0.391, + 0.207 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.328, + 0.218, + 0.361, + 0.248 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.328, + 0.249, + 0.362, + 0.269 + ], + "angle": 0, + "content": "Edge Server" + }, + { + "type": "image_footnote", + "bbox": [ + 0.251, + 0.271, + 0.398, + 0.285 + ], + "angle": 0, + "content": "\\(\\times\\) Computation Resource" + }, + { + "type": "image", + "bbox": [ + 0.441, + 0.145, + 0.564, + 0.269 + ], + "angle": 0, + "content": null + }, + { + "type": "image_footnote", + "bbox": [ + 0.413, + 0.272, + 0.58, + 0.285 + ], + "angle": 0, + "content": "X Communication Overhead" + }, + { + "type": "image", + "bbox": [ + 0.603, + 0.145, + 0.753, + 0.266 + ], + "angle": 0, + "content": null + }, + { + "type": "image_footnote", + "bbox": [ + 0.595, + 0.271, + 0.757, + 0.293 + ], + "angle": 0, + "content": "\\(\\checkmark\\) Performance & Latency Trade-off" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.305, + 0.789, + 0.389 + ], + "angle": 0, + "content": "Fig. 1: The key concept of distributed DNN. Left: Lightweight DNNs suffer from performance loss due to the limited computational resources; Middle: Edge computing often results in intolerable latency due to communication overhead; Right: Distributed DNNs deploy a small head model on the resource-limited device to achieve task-oriented compression and a large tail model on the server to decode compressed features and execute the rest of model." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.402, + 0.789, + 0.492 + ], + "angle": 0, + "content": "mobile devices [28, 29, 31, 52, 57], distributed DNNs can leverage the computation power of the edge/cloud and hence attain better performance. In addition, distributed DNNs leverage compression techniques to reduce the data size of intermediate representations. By only transmitting compact representations, distributed DNNs can reduce the transmission overhead significantly compared to traditional edge computing [34, 48, 60]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.493, + 0.79, + 0.794 + ], + "angle": 0, + "content": "Interestingly, recent work has proposed to further minimize the data size of transmitted latent representations with entropy coding [42]. The key idea is to train the distributed DNN jointly with an auxiliary entropy model, which estimates the distribution of latent representations. The output of the entropy model is then used as side information during inference time to facilitate the adaptive encoding of latent representations into bit streams with variable lengths. The proposed approach achieves better compression rate compared to quantization-based [38] and codec-based [2] approaches by minimizing the entropy of latent representations. Despite its excellent compression performance, the resilience of entropy coding to intentional interference (e.g., specifically crafted adversarial attacks) and unintentional interference (e.g., sudden events such as weather changes and motion blur) remains unexplored. We remark that in entropy-coding-based distributed DNNs, the bit rate of encoded representations relies completely on its estimated entropy. However, it is well known that DNNs trained on standard datasets are vulnerable to distribution shifts [24,25] and adversarial actions [20, 56]. As the entropy model is learned without considering any corruptions and adversarial perturbations, a small change in the input space might lead to a large estimate of entropy, and hence result in a bit rate that can exceed the transmission bandwidth. As depicted in Tab. 3 in our experiments, the transmission overhead can be increased by about \\(2\\mathrm{x}\\) in the worst case." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.796, + 0.789, + 0.84 + ], + "angle": 0, + "content": "In this work, we thoroughly assess for the first time the resilience of entropy models to intentional and unintentional interference as applied to distributed DNNs. Intriguingly, our findings unveil that the auxiliary entropy model learns" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.328, + 0.115, + 0.733, + 0.13 + ], + "angle": 0, + "content": "Resilience of Entropy Model in Distributed Neural Networks" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.222 + ], + "angle": 0, + "content": "a different set of input-related features than what learned by the backbone DNN. Ultimately, this enables us to design effective defense strategy to maintain the bit rate with an unnoticeable loss in the task performance. Our approach successfully reduces the communication overhead of perturbed data by about \\(9\\%\\) compared to unperturbed inputs, with only about \\(2\\%\\) accuracy loss." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.243, + 0.494, + 0.259 + ], + "angle": 0, + "content": "Summary of Novel Contributions" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.266, + 0.788, + 0.372 + ], + "angle": 0, + "content": "(1) We investigate for the first time the effect of intentional and unintentional interference to entropy coding in distributed DNNs. We show that interference increases the end-to-end latency of distributed DNNs and poses a threat to other users by saturating the transmission bandwidth. Through comprehensive experiments involving 3 distinct DNN architectures [23, 47], 2 entropy models [5,43], and 4 rate-distortion trade-off factors, we illustrate that attacks to entropy coding may increase the transmitted data size by up to about \\(95\\%\\);" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.375, + 0.788, + 0.481 + ], + "angle": 0, + "content": "(2) We design two visualization approaches to interpret the compression learned by entropy models. By disentangling features relevant to compression and classification across both frequency and spatial domains, we reveal that the entropy model learns specific features that are distinct from those beneficial for the end task. This finding enables us to devise an effective defense strategy that safeguards the vulnerable compression features while minimally impacting task performance;" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.484, + 0.788, + 0.576 + ], + "angle": 0, + "content": "(3) We propose a defense approach to attacks targeting entropy coding based on our findings discussed above. The proposed method effectively reduces the transmission overhead of attacked inputs by about \\(9\\%\\) in comparison to the unperturbed case, with only about \\(2\\%\\) decrease in accuracy. We remark that our approach is general in nature and can be combined with approaches such as adversarial training to further improve the resilience of entropy models." + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.266, + 0.788, + 0.576 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.578, + 0.788, + 0.683 + ], + "angle": 0, + "content": "The paper is organized as follows. Sec. 2 provides background of entropy coding in distributed DNNs. In Sec. 3, we formally describe the threat model for entropy coding in distributed DNNs. Sec. 4 presents benchmarking results on both intentional and unintentional interference. Sec. 5 details the proposed defense approach and comprehensively evaluates it against adaptive attacks. Related work is summarized in Sec. 6. Finally, we draw conclusions and discuss future work in Sec. 7." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.704, + 0.618, + 0.723 + ], + "angle": 0, + "content": "2 Entropy Coding in Distributed DNNs" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.734, + 0.788, + 0.811 + ], + "angle": 0, + "content": "Entropy coding techniques such as arithmetic coding [50] and asymmetric numeral systems [17] can encode a message with the optimal coding rate by leveraging the probabilistic information of the message. Based on the source coding theorem [53], the entropy of a message \\( z \\) with distribution \\( P_Z(z) \\) defines the lower bound of expected coding rate \\( R_{c}(z) \\) without any loss of information, i.e.," + }, + { + "type": "equation", + "bbox": [ + 0.362, + 0.826, + 0.785, + 0.842 + ], + "angle": 0, + "content": "\\[\n\\mathbb {E} \\left\\{R _ {c} (z) \\right\\} \\geq H _ {P _ {Z}} (z) = \\mathbb {E} \\left\\{- \\log_ {2} P _ {Z} (z) \\right\\} \\tag {1}\n\\]" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.378, + 0.129 + ], + "angle": 0, + "content": "M. Zhang et al." + }, + { + "type": "image", + "bbox": [ + 0.237, + 0.149, + 0.473, + 0.234 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.516, + 0.149, + 0.768, + 0.234 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.25, + 0.789, + 0.322 + ], + "angle": 0, + "content": "Fig. 2: A general framework for entropy coding in neural data compression. It consists of an encoder-decoder structure to extract a compact latent representation \\( z \\) and a prior model to estimate the distribution \\( P_Z(z) \\). Left: During training, the prior model is jointly optimized with the encoder-decoder backbone. Right: During inference, the learned prior \\( P_Z(z) \\) is used as side information for arithmetic encoding/decoding." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.348, + 0.789, + 0.499 + ], + "angle": 0, + "content": "where \\( H_{P_Z}(z) \\) is the entropy of \\( z \\). In recent data compression literature [1,5,6, 14,43,44], the entropy coding is integrated with lossy DNN-based compression to attain an effective compression rate with less information loss compared to traditional approaches. As depicted in Fig. 2, the common strategy is to jointly train an encoder-decoder backbone for extracting compact latent representations \\( z \\) with an auxiliary prior model \\( P_Z(z) \\). During training, a random noise source is introduced to simulate the effect of quantization on \\( z \\), allowing the backpropagation of gradients. In the inference phase, \\( z \\) is initially quantized to integers for entropy coding. Next, the output of the prior model is used as side information to encode the quantized \\( z \\) into bit streams adaptively with variable length." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.5, + 0.789, + 0.545 + ], + "angle": 0, + "content": "DNN-based data compression can achieve the near-optimal compression rate by minimizing the entropy of latent representations during training. Specifically, the objective is to minimize the rate-distortion function" + }, + { + "type": "equation", + "bbox": [ + 0.358, + 0.554, + 0.787, + 0.59 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} (x, (\\hat {x}, z)) = \\underbrace {\\| x - \\hat {x} \\| _ {2} ^ {2}} _ {\\text {d i s t o r t i o n}} - \\beta \\cdot \\underbrace {\\log_ {2} P _ {Z} (z)} _ {\\text {r a t e}}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.598, + 0.789, + 0.643 + ], + "angle": 0, + "content": "where \\(\\hat{x}\\) is the reconstructed data, \\(||x - \\hat{x}||_2^2\\) is the loss function of the encoder-decoder backbone and \\(-\\log_2P_Z(z)\\) is the loss function of the prior model. Constant \\(\\beta\\) denotes the trade-off between the rate and distortion." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.644, + 0.79, + 0.842 + ], + "angle": 0, + "content": "Recent work [42] has incorporated entropy coding into distributed DNNs to increase communication efficiency between mobile and edge devices. The proposed method, referred to as the entropic student, aims at achieving a balance between the coding rate and task-specific performance metrics (e.g., cross entropy in classification tasks). One distinction between the entropic student and other entropy-coding-based feature compression work [55, 59] lies in the significantly smaller encoder size of the former due to constraints on computational resources in mobile devices. As a result, the entropic student splits the DNN at an early layer and use knowledge distillation [26] for training the head to preserve accuracy. However, existing work does not address the issue of resilience of entropy models, which is the key goal of this paper. A relevant work [4] investigated the robustness of density estimation aiming to maximize \\( P_Z(z) \\) while we are interested in perturbations that minimize \\( P_Z(z) \\)." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.328, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "Resilience of Entropy Model in Distributed Neural Networks" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.145, + 0.656, + 0.163 + ], + "angle": 0, + "content": "3 Modeling Interference to Entropy Coding" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.176, + 0.788, + 0.252 + ], + "angle": 0, + "content": "While the resilience of DNNs to distribution shifts [7,19,24,25,58] and adversarial actions [3,11,12,30,35,56] has been extensively investigated, we investigate types of interference that can alter the length of the encoded bit stream, which ultimately leads to increased bandwidth utilization. In this section, we formulate the resilience of compression in entropy-coding-driven distributed DNNs." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.272, + 0.788, + 0.334 + ], + "angle": 0, + "content": "Threat Model for Accuracy. Let \\( D = \\{x, y\\} \\) denotes a dataset for classification problem where \\( x \\in \\mathcal{X} \\) and \\( y \\in \\mathcal{Y} \\) are specific input and corresponding label samples respectively, while \\( \\mathcal{X}, \\mathcal{Y} \\) are input and output space, respectively. There exists a prior yet unknown distribution \\( P(y|x) \\) such that" + }, + { + "type": "equation", + "bbox": [ + 0.403, + 0.344, + 0.786, + 0.361 + ], + "angle": 0, + "content": "\\[\nP (y | x) \\geq P (\\hat {y} | x) \\quad \\forall \\hat {y} \\in \\mathcal {Y} \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.369, + 0.789, + 0.476 + ], + "angle": 0, + "content": "where \\(\\hat{y}\\) denotes incorrect labels. The DNN classifier, is trained to approximate the distribution \\(\\tilde{P}(y|x) \\approx P(y|x)\\) with empirical risk minimization approaches based on an assumption that in-sample data and out-of-sample data shared a similar distribution \\(P(y|x)\\). However, the unintentional interference denoted by \\(\\delta\\), which can be modeled as an additive noise to \\(x\\), creates a new dataset \\(D' = \\{x + \\delta, y\\}\\). It introduces a covariate shift that \\(P(x + \\delta) \\neq P(x)\\) but \\(P(y|x + \\delta) = P(y|x)\\), resulting in classification errors such that" + }, + { + "type": "equation", + "bbox": [ + 0.374, + 0.485, + 0.786, + 0.502 + ], + "angle": 0, + "content": "\\[\n\\tilde {P} (\\hat {y} | x + \\delta) \\geq \\tilde {P} (y | x + \\delta) \\quad \\exists \\hat {y} \\in \\mathcal {Y} \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.512, + 0.788, + 0.618 + ], + "angle": 0, + "content": "In the intentional interference scenario, algorithms attempt to find the minimum covariate shift that leads to Eq. (4). In this setting, a small perturbation in \\( l_{p} \\) distance \\( ||\\delta||_{p} \\leq \\epsilon \\) is intentionally crafted to mislead the DNN classifier where \\( \\epsilon \\) denotes the constraint. Attackers formulate Eq. (4) as an optimization problem over the input space \\( \\mathcal{X} \\), and hence can apply the gradient information of the loss function w.r.t. \\( x \\) to find the \\( \\delta \\). Projected Gradient Descent (PGD) [35] provided a unified view on iterative gradient attacks under \\( l_{p} \\) constraint:" + }, + { + "type": "equation", + "bbox": [ + 0.407, + 0.628, + 0.786, + 0.66 + ], + "angle": 0, + "content": "\\[\nx ^ {i + 1} = x ^ {i} + \\alpha \\cdot \\epsilon \\cdot \\frac {\\nabla_ {x} \\mathcal {L}}{| | \\nabla_ {x} \\mathcal {L} | | _ {p}} \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.669, + 0.788, + 0.72 + ], + "angle": 0, + "content": "where \\( x^{i} \\) is the adversarial sample at \\( i \\) step, \\( \\alpha \\) is the learning rate, and \\( \\frac{\\nabla_{x}\\mathcal{L}}{||\\nabla_{x}\\mathcal{L}||_{p}} \\) is the normalized gradient of a loss function \\( \\mathcal{L} \\) (e.g., a sign function in \\( l_{\\infty} \\) space). \\( \\mathcal{L} \\) can be cross entropy or other more advanced loss functions [11]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.739, + 0.788, + 0.815 + ], + "angle": 0, + "content": "Threat Model for Data Rate. Let \\( H(x) = z \\) denotes the head network where \\( z \\in \\mathcal{Z} \\) is the output of the head and \\( \\mathcal{Z} \\) is the latent space. The goal is to achieve a minimum compression rate \\( -\\log_2 P_Z(z) \\). As \\( z \\) is dependent on \\( x \\), the distribution shift introduced by unintentional interference can result in a difference in coding rate:" + }, + { + "type": "equation", + "bbox": [ + 0.361, + 0.826, + 0.786, + 0.842 + ], + "angle": 0, + "content": "\\[\n- \\log_ {2} P _ {Z} (H (x + \\delta)) \\neq - \\log_ {2} P _ {Z} (H (x)) \\tag {6}\n\\]" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.378, + 0.129 + ], + "angle": 0, + "content": "M. Zhang et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.267, + 0.145, + 0.735, + 0.16 + ], + "angle": 0, + "content": "Table 1: DNNs, entropy models and datasets utilized in experiments." + }, + { + "type": "table", + "bbox": [ + 0.244, + 0.172, + 0.76, + 0.394 + ], + "angle": 0, + "content": "
DNN Architectures
ResNet-50 [23]A standard baseline in distributed DNNs [38-40, 42]
ResNet-101 [23]A deeper model in the ResNet family
RegNetY-6.4GF [47]An advanced design with better performance
Entropy Models
FP [5]A fully factorized prior model
MSHP [43]An effective learnable entropy model
Datasets
ImageNet [16]The standard ImageNet validation dataset
ImageNet-C [25]A synthetic dataset with 15 corruptions and 5 severities
Random NoiseNoisy images as a baseline for adversarial robustness
PGD-Acc [35]PGD targeting classification
PGD-EPGD targeting entropy
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.422, + 0.789, + 0.467 + ], + "angle": 0, + "content": "Note that the unintentional distribution shift of \\( x \\) does not necessarily result in a larger bit rate. However, in the adversarial setting, attackers intend to maximize the entropy of \\( z \\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.413, + 0.478, + 0.786, + 0.506 + ], + "angle": 0, + "content": "\\[\n\\min _ {\\delta} \\log_ {2} P _ {Z} (H (x + \\delta)) \\tag {7}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.419, + 0.504, + 0.521, + 0.519 + ], + "angle": 0, + "content": "\\[\ns. t. \\quad | | \\delta | | _ {p} \\leq \\epsilon\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.529, + 0.793, + 0.561 + ], + "angle": 0, + "content": "Thus, to correctly assess the robustness of entropy models, we use the same PGD algorithm in Eq. (5) yet with the entropy as its loss function \\(\\mathcal{L} = -\\log_2P_Z(H(x))\\)" + }, + { + "type": "title", + "bbox": [ + 0.214, + 0.581, + 0.717, + 0.6 + ], + "angle": 0, + "content": "4 Understanding the Resilience of Entropy Model" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.613, + 0.429, + 0.629 + ], + "angle": 0, + "content": "4.1 Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.637, + 0.788, + 0.699 + ], + "angle": 0, + "content": "In this section, we outline the experimental setup for evaluating the resilience of entropy models. We first specify the DNNs and entropy models under consideration, followed by a description of the metrics and datasets used in our experiments. A summary of our experimental setup is provided in Tab. 1." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.719, + 0.789, + 0.843 + ], + "angle": 0, + "content": "Ablation of DNNs and Entropy Models. We consider three distinct DNNs: ResNet-50 [23], ResNet-101 [23] and RegNetY-6.4GF [47]. ResNet-50 serves as a common baseline in distributed DNN literature [38-40,42] while ResNet-101 is a deeper variant with the same design as former. RegNetY-6.4GF has an advanced design with better classification performance. The early layers are replaced by a specific tailored head architecture as proposed in [42] and incorporated with different entropy models. For the sake of brevity, we consider ResNet-50 as the default DNN architecture without additional specification. We employ two entropy" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.329, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "Resilience of Entropy Model in Distributed Neural Networks" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "image", + "bbox": [ + 0.249, + 0.145, + 0.495, + 0.297 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.31, + 0.3, + 0.435, + 0.311 + ], + "angle": 0, + "content": "(a) resilience w.r.t. \\(\\beta\\)" + }, + { + "type": "image", + "bbox": [ + 0.498, + 0.147, + 0.756, + 0.297 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.541, + 0.3, + 0.711, + 0.311 + ], + "angle": 0, + "content": "(b) resilience w.r.t. severities" + }, + { + "type": "image_caption", + "bbox": [ + 0.266, + 0.323, + 0.735, + 0.338 + ], + "angle": 0, + "content": "Fig. 3: Resilience of accuracy and data rate to 4 common corruptions." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.365, + 0.788, + 0.473 + ], + "angle": 0, + "content": "models: Factorized Prior (FP) [5] and Mean Scale Hyper Prior (MSHP) [43]. FP is the earliest design of entropy model and a basic component in many other advanced designs, while MSHP provides more powerful compression performance by injecting a hierarchical learnable block before FP to extract the entropy information. For conciseness, we consider MSHP as the default entropy model. In addition, since models are trained with different rate-distortion trade-off \\(\\beta\\), we also investigate the compression robustness w.r.t. different \\(\\beta\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.493, + 0.789, + 0.706 + ], + "angle": 0, + "content": "Datasets and Metrics. Conversely from existing work that measures the compression performance with bits per pixel (BPP) [5,6,14,43,44], we evaluate the entropy model with the data size of the whole bit stream after encoding as in [37]. Compared to BPP, data size quantifies more directly the networking traffic between the head and tail. To assess the resilience to unintentional interference, we use the ImageNet-C dataset proposed by [25] comprising 15 common corruptions classified into 4 categories (noise, blur, digital, and weather) with 5 different severities. We also report the data size and accuracy on the clean ImageNet [16] validation set as a baseline. To assess the resilience to intentional interference, we implement the PGD targeting both the classification and compression performance in \\( l_{\\infty} \\) space. For clarity, we denote the conventional PGD as PGD-Acc and the PGD targeting entropy models as PGD-E. Meanwhile, random noise with the same perturbation level is added to the clean ImageNet as a baseline comparison for adversarial robustness." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.726, + 0.593, + 0.74 + ], + "angle": 0, + "content": "4.2 Resilience to Unintentional Interference" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Resilience w.r.t. \\(\\beta\\). We select one corruption as a representative from each category in ImageNet-C [25] to investigate the resilience as a function of ratedistortion trade-off \\(\\beta\\). Fig. 3a shows the compression and classification performance on defocus blur, shot noise, snow and contrast dataset, each with the severity of 5. Compared to the clean ImageNet, all corruptions decrease the accuracy significantly. However, not all corruptions increase the data size. Only" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.378, + 0.129 + ], + "angle": 0, + "content": "M. Zhang et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.788, + 0.187 + ], + "angle": 0, + "content": "Table 2: Resilience w.r.t. prior models. The average data size of FP and MSHP for clean ImageNet are \\(11.65 \\pm 1.02\\) and \\(9.62 \\pm 1.70\\) KBBytes, respectively. Red for larger data size and blue for lower compared to the baseline." + }, + { + "type": "table", + "bbox": [ + 0.223, + 0.199, + 0.782, + 0.273 + ], + "angle": 0, + "content": "
Prior \nModelGaussian \nSize[KB]Noise \nAcc[%]Motion \nSize[KB]Blur \nAcc[%]Impulse \nSize[KB]Noise \nAcc[%]Glass \nSize[KB]Blur \nAcc[%]
FP13.19±0.4059.5110.51±0.7062.2213.36±0.3450.9310.43±0.6653.63
MSHP11.96±0.4759.367.88±1.4661.5812.52±0.3551.127.78±1.3953.67
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.299, + 0.788, + 0.39 + ], + "angle": 0, + "content": "shot noise increases the data size by \\(65.31\\%\\) on average. In contrast, Defocus blur and contrast decrease the data size by \\(53.31\\%\\) and \\(67.45\\%\\) on average, respectively. Snow has little affect to compression, resulting in only \\(4.42\\%\\) decrease of data size. This indicates that the entropy model learns a different set of features than the classification model. Since it shows the same trend for different rate-distortion trade-off factors, we only consider \\(\\beta = 0.08\\) in next experiments." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.411, + 0.789, + 0.548 + ], + "angle": 0, + "content": "Resilience w.r.t. Severities. Next, we explore resilience w.r.t. severity levels using the same corruptions mentioned earlier. As illustrated in Fig. 3b, the data size of shot noise rises with increasing severity, whereas defocus blur and contrast consistently decrease the data size. Snow shows minimal variation across different severity levels. This can be attributed to defocus blur and contrast removing high-frequency components in images, while shot noise introduces new patterns in high-frequency space. Snow, on the other hand, lacks a specific pattern in the frequency domain. Thus, the compression is particularly sensitive to high-frequency information in the input space." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.568, + 0.789, + 0.751 + ], + "angle": 0, + "content": "Resilience w.r.t. Prior Models. Tab. 2 shows the impact of additional 4 corruptions at severity level 1 on both FP and MSHP model. To further validate the previous observation, we choose 2 noise corruptions (Gaussian noise and impulse noise), which introduce high-frequency noise to images, and two blur corruptions (motion blur and glass blur), which eliminate high-frequency components from images. As shown in Tab. 2, Gaussian noise increases \\(13.22\\%\\) and \\(24.32\\%\\) of data size for FP and MSHP, respectively. Impulse noise increases \\(14.42\\%\\) and \\(30.15\\%\\) of data size for FP and MSHP, respectively. On the other hand, motion blur reduces \\(9.79\\%\\) and \\(18.09\\%\\) of data size for FP and MSHP, respectively. Glass blur reduces \\(10.47\\%\\) and \\(19.13\\%\\) of data size for FP and MSHP, respectively. Thus, the previous finding that compressive features reside in the high-frequency space remains consistent for different entropy models." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.771, + 0.57, + 0.785 + ], + "angle": 0, + "content": "4.3 Resilience to Intentional Interference" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Resilience w.r.t. DNNs Fig. 4a shows the resilience of different DNN architectures to accuracy and entropy attacks with the \\( l_{\\infty} \\) constraint \\( \\epsilon = 4 / 255 \\). As evident in the figure, the attacks show similar patterns regardless of their" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.329, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "Resilience of Entropy Model in Distributed Neural Networks" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "image", + "bbox": [ + 0.245, + 0.145, + 0.499, + 0.302 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.261, + 0.304, + 0.484, + 0.316 + ], + "angle": 0, + "content": "(a) resilience w.r.t. DNN architectures" + }, + { + "type": "image", + "bbox": [ + 0.502, + 0.147, + 0.761, + 0.301 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.538, + 0.304, + 0.726, + 0.316 + ], + "angle": 0, + "content": "(b) resilience w.r.t. prior models" + }, + { + "type": "image_caption", + "bbox": [ + 0.266, + 0.328, + 0.736, + 0.342 + ], + "angle": 0, + "content": "Fig. 4: Resilience of accuracy and data rate to intentional interference" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.372, + 0.789, + 0.462 + ], + "angle": 0, + "content": "victim models. While PGD-Acc reduces \\(42.88\\%\\) accuracy on average, it only increases by \\(7.28\\%\\) of data size on average compared to random noise. On the other hand, PGD-E decreases \\(2.34\\%\\) accuracy on average compared to random noise but increases the bit rate 1.35x times on average. This indicates that PGD-E and PGD-Acc are targeting separate features in the input space which has minor impact on each other." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.486, + 0.789, + 0.669 + ], + "angle": 0, + "content": "Resilience w.r.t. Prior Models. Fig. 4b illustrates the effect of adversarial attacks with \\(\\epsilon = 8 / 255\\) on victims using different prior models. As shown in the figure, both entropy models show considerable vulnerability to PGD-E compared to PGD-Acc. The data size are increased by \\(21.93\\%\\) and \\(46.82\\%\\) for FP and MSHP respectively while the performance loss are \\(9.91\\%\\) and \\(6.62\\%\\) respectively. On the other hand, PGD-Acc successfully reduces \\(72.10\\%\\) and \\(57.10\\%\\) classification performance of FP and MSHP compared to random noise respectively. However, it only increases \\(5.66\\%\\) and \\(10.19\\%\\) bit rate for FP and MSHP respectively. Hence, attackers targeting compression tend to perturb a different set of features than attackers targeting discriminative tasks regardless of the entropy models. In addition, while the FP has a better robustness in compression, it has significantly worse robustness in accuracy." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.691, + 0.542, + 0.709 + ], + "angle": 0, + "content": "5 Proposed Defense Mechanism" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.724, + 0.729, + 0.74 + ], + "angle": 0, + "content": "5.1 Disentangle Compression Features in Frequency Domain" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.789, + 0.841 + ], + "angle": 0, + "content": "As shown in Sec. 4.2, interference that introduces high-frequency noise (e.g., shot noise) can effectively increase the data size while interference that removes high-frequency information (e.g., defocus blur) reduces the data size. Intuitively this is a consequence of the small head in distributed DNNs as semantic information is usually captured in deeper layers while early layers tend to capture low level information [8]. To validate our intuition, we introduce total variation, defined" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.378, + 0.129 + ], + "angle": 0, + "content": "M. Zhang et al." + }, + { + "type": "image", + "bbox": [ + 0.252, + 0.147, + 0.332, + 0.209 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.252, + 0.211, + 0.332, + 0.272 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.336, + 0.147, + 0.416, + 0.209 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.336, + 0.211, + 0.416, + 0.272 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.422, + 0.147, + 0.501, + 0.209 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.422, + 0.211, + 0.501, + 0.272 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.147, + 0.585, + 0.209 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.506, + 0.211, + 0.585, + 0.272 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.59, + 0.147, + 0.669, + 0.209 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.59, + 0.211, + 0.669, + 0.272 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.674, + 0.147, + 0.754, + 0.209 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.674, + 0.211, + 0.754, + 0.272 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.287, + 0.785, + 0.317 + ], + "angle": 0, + "content": "Fig. 5: Entropy models are sensitive to high-frequency features. From left to right: images from ImageNet validation set, bit rate maps and total variation maps." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.345, + 0.784, + 0.377 + ], + "angle": 0, + "content": "as the integral of image gradient magnitude, which is first proposed for image denoising in [51]. The anitrosopic total variation of a 2-D image is defined as" + }, + { + "type": "equation", + "bbox": [ + 0.345, + 0.387, + 0.786, + 0.42 + ], + "angle": 0, + "content": "\\[\nT V (x) = \\sum_ {i, j} \\left| x _ {i + 1, j} - x _ {i, j} \\right| + \\left| x _ {i, j + 1} - x _ {i, j} \\right| \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.43, + 0.784, + 0.492 + ], + "angle": 0, + "content": "where \\( x \\) denotes an image sample, \\( i,j \\) denotes the pixel position in width and height, respectively. The image gradient magnitude \\( |x_{i+1,j} - x_{i,j}| \\) and \\( |x_{i,j+1} - x_{i,j}| \\) describe the sharpness of pixel changes, where larger magnitude indicates high-frequency information such as edges and textures in images." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.491, + 0.785, + 0.598 + ], + "angle": 0, + "content": "We compare the total variation map of an image with its bit rate map generated by the entropy model. To plot the total variation map, we first slice an image into many small patches and then compute the total variation for each patch. Meanwhile, the bit rate map is the estimated entropy of the latent representation \\(-\\log_2P_Z(z)\\). As shown in Fig. 5, there exists a significant correlation between the bit rate map and the total variation map, which indicates that the entropy model is sensitive to high-frequency features in the input space." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.619, + 0.684, + 0.635 + ], + "angle": 0, + "content": "5.2 Disentangle Compression Features in Spatial Space" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.644, + 0.784, + 0.719 + ], + "angle": 0, + "content": "The work in [42] has shown that entropy model in distributed DNNs tends to assign a larger number of bits to task-oriented information, such as objects, while compressing irrelevant information, like the background. Consequently, it becomes intuitively easier for an adversary to increase the bit rate in the background area to attack the entropy model." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.785, + 0.84 + ], + "angle": 0, + "content": "To interpret the action conducted by attackers, we generate the bit rate comparison map between adversarial samples generated by PGD-E and PGD-Acc. This comparison map represents the estimated entropy of adversarial samples generated by PGD-E subtracted from the bit rate of PGD-Acc-perturbed samples. In this context, the red area indicates that PGD-E tends to allocate more bits compared to PGD-Acc, whereas the blue region suggests that PGD-E tends to allocate fewer bits. As depicted in Fig. 6, PGD-E puts more efforts to increase the bit rate in background region, whereas PGD-Acc focuses on altering" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.329, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "Resilience of Entropy Model in Distributed Neural Networks" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "image", + "bbox": [ + 0.252, + 0.147, + 0.328, + 0.206 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.336, + 0.147, + 0.414, + 0.206 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.422, + 0.147, + 0.499, + 0.206 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.507, + 0.147, + 0.584, + 0.206 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.591, + 0.147, + 0.669, + 0.207 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.677, + 0.147, + 0.754, + 0.207 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.222, + 0.785, + 0.264 + ], + "angle": 0, + "content": "Fig. 6: Entropy models are more vulnerable to perturbation in background region. From left to right: images from ImageNet validation set, bit rate comparison between PGD-E and PGD-Acc (PGD-E allocate less bit in blue area and more bit in red area)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.291, + 0.784, + 0.381 + ], + "angle": 0, + "content": "task-oriented information. This is in line with the experiments in Sec. 4.3 that PGD-E and PGD-Acc target different features in the input space. On one hand, increasing the bit rate in the background region will affect little accuracy as they are irrelevant information for the discriminative task. On the other hand, modifying object information can significantly degrade the performance, yet it has less impact on compression." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.4, + 0.593, + 0.416 + ], + "angle": 0, + "content": "5.3 Object-Aware Total Variation Denoising" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.421, + 0.784, + 0.467 + ], + "angle": 0, + "content": "Given the observation in Sec. 5.1, we propose a denoising technique based on total variation to remove high-frequency noise in adversarial images where we solve the following optimization problem:" + }, + { + "type": "equation", + "bbox": [ + 0.396, + 0.473, + 0.785, + 0.502 + ], + "angle": 0, + "content": "\\[\n\\min _ {x} \\frac {1}{2} \\| x - x ^ {\\prime} \\| _ {2} ^ {2} + \\lambda \\cdot T V (x), \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.507, + 0.784, + 0.597 + ], + "angle": 0, + "content": "where \\( x' \\) is the distorted image, \\( x \\) is the denoised image, \\( TV(x) \\) is the total variation of \\( x \\) and \\( \\lambda \\) is a regularization factor to control the degree of smoothing. Eq. (9) means the image \\( x \\) after denoising should keep most of the information in \\( x' \\) (i.e., minimize \\( ||x - x'||_2^2 \\)) but discard the high-frequency information (i.e., minimize \\( TV(x) \\)). Optimizing Eq. (9) is not trivial as total variation is non-differentiable. As a result, we use a sub-gradient descent:" + }, + { + "type": "equation", + "bbox": [ + 0.367, + 0.604, + 0.785, + 0.622 + ], + "angle": 0, + "content": "\\[\nx ^ {i + 1} = x ^ {i} - \\alpha \\cdot \\left(\\left(x ^ {i} - x ^ {\\prime}\\right) + \\lambda \\cdot g \\left(x ^ {i}\\right)\\right) \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.628, + 0.784, + 0.658 + ], + "angle": 0, + "content": "where \\( x^i \\) is the denoised image at the step \\( i \\), \\( \\alpha \\) is the step size and \\( g(x^i) \\) is the image gradient of \\( x^i \\) [9]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.659, + 0.785, + 0.749 + ], + "angle": 0, + "content": "Simply applying the total variation denoising will also decrease the task-oriented performance as it also removes useful information in high-frequency domain. Since adversarial attacks targeting entropy models attempt to increase bit rate in background region as demonstrated in Sec. 5.2, we add a mask to Eq. (10) to force the denoising algorithm to only remove high-frequency information in non-object space, that is," + }, + { + "type": "equation", + "bbox": [ + 0.354, + 0.755, + 0.785, + 0.773 + ], + "angle": 0, + "content": "\\[\nx ^ {i + 1} = x ^ {i} - \\alpha \\cdot m \\cdot \\left(\\left(x ^ {i} - x ^ {\\prime}\\right) + \\lambda \\cdot g \\left(x ^ {i}\\right)\\right) \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.784, + 0.84 + ], + "angle": 0, + "content": "where \\( m \\) is the mask to control the denoising level. In practice, we interpolate the output of entropy model \\( P_Z(z) \\) as the soft mask. This is because the higher bit rate \\( -\\log_2P_Z(z) \\) corresponding to object area has a smaller value of \\( P_Z(z) \\), making it naturally a soft mask to avoid smoothing the object region." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.378, + 0.129 + ], + "angle": 0, + "content": "M. Zhang et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.788, + 0.174 + ], + "angle": 0, + "content": "Table 3: Compression and classification performance before and after defense. The average data size for clean ImageNet is \\(9.62 \\pm 1.70\\) KBytes." + }, + { + "type": "table", + "bbox": [ + 0.217, + 0.185, + 0.79, + 0.259 + ], + "angle": 0, + "content": "
ε = 2/255ε = 4/255ε = 8/255ε = 16/255
Size[KB]Acc[%]Size[KB]Acc[%]Size[KB]Acc[%]Size[KB]Acc[%]
before12.35±0.6373.2614.11±0.3871.3316.15±0.2965.6218.79±0.3752.13
after8.76±1.3870.829.55±1.0770.1511.02±0.6567.2213.44±0.4859.87
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.286, + 0.787, + 0.421 + ], + "angle": 0, + "content": "Experimental Results. We investigate the proposed defense method as a function of perturbation budget \\(\\epsilon\\). As shown in Tab. 3, attackers targeting entropy significantly increase the data size, by up to \\(95.32\\%\\) when \\(\\epsilon = 16/255\\) compared to the clean images (9.62 KBytes). On the contrary, our denoising approach significantly reduce the data size for all \\(\\epsilon\\). For \\(\\epsilon = 2/255, 4/255\\), the average data size is even \\(8.94\\%\\) and \\(0.73\\%\\) smaller than the clean images with only \\(2.44\\%\\) and \\(1.18\\%\\) accuracy loss, respectively. For \\(\\epsilon = 8/255, 16/255\\), the data size is reduced by \\(31.80\\%\\) and \\(28.47\\%\\) respectively, while the accuracy improved slightly after denoising (\\(1.60\\%\\) for \\(\\epsilon = 8/255\\) and \\(7.74\\%\\) for \\(\\epsilon = 16/255\\))." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.422, + 0.787, + 0.483 + ], + "angle": 0, + "content": "We remark that the proposed defense is a standalone approach which can be incorporated into other approaches such as adversarial training to further improve the resilience. The accuracy loss incurred in small perturbation scenarios \\((\\epsilon = 2 / 255, 4 / 255)\\) can be mitigated by training with augmented data." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.503, + 0.406, + 0.518 + ], + "angle": 0, + "content": "5.4 Adaptive Attacks" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.526, + 0.787, + 0.587 + ], + "angle": 0, + "content": "To thoroughly evaluate the effectiveness of a defense strategy, [10] proposed to design adaptive attacks that counter the defense mechanism with perfect knowledge of both DNN and defense in the white-box setting. To this end, we assess our defense with two adaptive attacks." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.608, + 0.787, + 0.7 + ], + "angle": 0, + "content": "Low Frequency Attack. As the total variation denoising is proposed to remove high-frequency components, the first adaptive attack that we considered is to only add perturbations in low frequency space. Following the methodology outlined in [21], we first convert the gradient to frequency space with discrete cosine transform and mask out the high-frequency components. The gradient is then transformed back and Eq. (5) is applied to create the adversarial samples." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Regional Attack. Given that the defense exclusively denoises the background region, the second adaptive attack we considered is to force the adversary put greater effort in increasing bit rate in the object region. Similar to Eq. (11), a mask is multiplied to the loss function before backpropagation. In practice, we choose \\( 1 - P_Z(z) \\) as the soft mask where \\( P_Z(z) \\) is the probability of \\( z \\) ranging between 0 and 1. A larger \\( P_Z(z) \\) indicates a smaller bit rate \\( -\\log_2P_Z(z) \\) corresponding to the background region. Therefore, \\( 1 - P_Z(z) \\) encourages the adversary to take more attention to object region." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.328, + 0.114, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Resilience of Entropy Model in Distributed Neural Networks" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "table_caption", + "bbox": [ + 0.314, + 0.145, + 0.688, + 0.159 + ], + "angle": 0, + "content": "Table 4: Resilience to adaptive attacks with \\(\\epsilon = 4 / 255\\)" + }, + { + "type": "table", + "bbox": [ + 0.34, + 0.172, + 0.659, + 0.244 + ], + "angle": 0, + "content": "
Adaptive AttacksBefore DenoisingAfter Denoising
Size[KB]Acc[%]Size[KB]Acc[%]
Frequency14.11±0.3871.389.56±1.0770.17
Regional14.05±0.4171.639.50±1.1070.22
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.27, + 0.788, + 0.348 + ], + "angle": 0, + "content": "Resilience to Adaptive Attacks. Tab. 4 shows results of adaptive attacks with perturbation budget \\(\\epsilon = 4 / 255\\). Our defense successfully reduces the average data size by \\(67.75\\%\\) and \\(67.62\\%\\) for low frequency attack and regional attack respectively while the accuracy only decreases \\(1.21\\%\\) and \\(1.41\\%\\) respectively. Thus, the proposed approach is resilient to adaptive attacks." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.367, + 0.388, + 0.384 + ], + "angle": 0, + "content": "6 Related Work" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.397, + 0.787, + 0.488 + ], + "angle": 0, + "content": "Distributed Deep Neural Network. Distributed DNN is proposed to meet the challenge of deploying artificial intelligence to resource-constrained platforms such as mobile devices and Internet of Things (IoT) devices [41]. To minimize the end-to-end latency across devices, dimension reduction designs (i.e., bottlenecks) are introduced to compress the data size of intermediate representations that need to be sent [18, 54]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.488, + 0.787, + 0.594 + ], + "angle": 0, + "content": "However, unlike autoencoders [32,36,49] that have symmetric designs on both encoder and decoder sides, distributed DNNs usually have asymmetric designs due to the limit of computation resources, and thus inject the bottleneck in early layers [18, 38-40, 54]. As a result, naively end-to-end trained distributed DNNs have noticeable performance loss [18, 54]. To preserve accuracy, [38] proposed to use knowledge distillation to train the distributed DNN separately while [39] proposed a multi-stage training approach for each part of the DNN." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.594, + 0.787, + 0.685 + ], + "angle": 0, + "content": "Along with bottleneck-based distributed DNNs, different coding-based approaches are also applied to reduce the data size of the latent representations. [2] proposed to apply codec-based compression such as JPEG to latent representations and [15] adopted spatio-temporal coding such as HEVC to compress the streaming latent features in video tasks. [42,55,59] use entropy coding to achieve a higher compression ratio of latent representations." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.704, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Entropy Coding in Deep Data Compression. [5] first introduced a fully factorized prior to integrate entropy coding with variational autoencoder, showcasing superior quality enhancement over conventional codec-based methods like JPEG for image compression. Building upon this, [6] extended the conventional factorized prior model to a learnable hierarchical hyper prior, resulting in improved performance. [43] proposed a joint auto-regressive and hyper prior design while [44] proposed a channel-aware entropy coding scheme. In addition, [14] introduced the attention mechanism to prior models and [1] extended the technique to video compression." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.378, + 0.129 + ], + "angle": 0, + "content": "M. Zhang et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.283 + ], + "angle": 0, + "content": "To minimize the communication overhead in distributed DNN scenarios, [55] for the first time adopted the entropy coding to further compress the latent representations after the bottleneck while [59] extended the approach to object detection. However, these approaches overlooked resource constraints, resulting in over-complex network designs for mobile devices. Conversely, [42] integrated entropy coding with a multi-stage training strategy [39], aiming to optimize a tripartite trade-off encompassing computation resources in mobile devices, communication overhead between mobile and edge devices, and end-to-end performance of neural networks." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.304, + 0.788, + 0.47 + ], + "angle": 0, + "content": "Resilience of DNN efficiency. While the reliability of DNNs have been extensively investigated, limited attention has been devoted to interference that undermine their efficiency. [22] explored this direction by demonstrating that imperceptible perturbations to input, leveraging intermediate representations of DNNs, could nullify the computation savings achieved by dynamic depth neural networks, which adapt their depth based on input complexity. [46] extended these attacks to target both dynamic depth and dynamic width neural networks. Moreover, in [45], the authors proposed simultaneous adjustments to the direction and magnitude of attacks, enhancing their effectiveness. [27] revealed that early exit dynamic DNNs could be tricked into late-stage inferences, significantly slowing down inference speed." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.471, + 0.788, + 0.561 + ], + "angle": 0, + "content": "The study of efficiency vulnerabilities in DNNs has also ventured into real-world deployment scenarios. [33] focused on attacking LiDAR-based detection systems, introducing latency in detections and exposing vulnerabilities in critical contexts such as self-driving cars. [13] demonstrated that neural image caption generation models could be manipulated to incur increased computation costs by inducing unnecessary decoder calls for token generation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.561, + 0.788, + 0.622 + ], + "angle": 0, + "content": "In star contrast to existing research, our work investigates a novel threat emerging in distributed DNNs. Here, adversaries not only compromise the communication efficiency of entropy-coded distributed DNNs but also pose a threat to other users by saturating the transmission bandwidth." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.644, + 0.36, + 0.66 + ], + "angle": 0, + "content": "7 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.788, + 0.84 + ], + "angle": 0, + "content": "This paper has investigated the resilience of entropy models in distributed DNNs against both intentional and unintentional interference. We conducted thorough evaluations using 3 different DNN architectures, 2 entropy model designs, and 4 rate-distortion trade-off factors with common corruption datasets and adversarial attacks. Our analysis disentangled compression features in both spatial and frequency domains, revealing vulnerabilities of the entropy model to specific types of perturbations. Building on these findings, we proposed a standalone defense strategy aimed at reducing data size with minimal task-oriented performance loss. Our future work will focus on designing more advanced defense approaches for distributed DNNs that are resilient in both compression and classification tasks." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.328, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Resilience of Entropy Model in Distributed Neural Networks" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.403, + 0.163 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.173, + 0.788, + 0.264 + ], + "angle": 0, + "content": "This work is funded in part by the Air Force Office of Scientific Research under contract number FA9550-23-1-0261, by the Office of Naval Research under contract number N00014-23-1-2221, as well as by National Science Foundation grants CNS-2134973 and CNS-2312875. Any opinions, findings and conclusions or recommendations expressed in this material are those of the authors and do not necessarily reflect the views of the U.S. Government." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.284, + 0.323, + 0.3 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.31, + 0.787, + 0.367 + ], + "angle": 0, + "content": "1. Agustsson, E., Minnen, D., Johnston, N., Balle, J., Hwang, S.J., Toderici, G.: Scale-space flow for end-to-end optimized video compression. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8503-8512 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.366, + 0.787, + 0.393 + ], + "angle": 0, + "content": "2. Alvar, S.R., Bajic, I.V.: Pareto-optimal bit allocation for collaborative intelligence. IEEE Transactions on Image Processing 30, 3348-3361 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.393, + 0.787, + 0.434 + ], + "angle": 0, + "content": "3. Andriushchenko, M., Croce, F., Flammarion, N., Hein, M.: Square attack: a query-efficient black-box adversarial attack via random search. In: European conference on computer vision. pp. 484-501. Springer (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.434, + 0.787, + 0.473 + ], + "angle": 0, + "content": "4. Arvinte, M., Cornelius, C., Martin, J., Himayat, N.: Investigating the adversarial robustness of density estimation using the probability flow ode. arXiv preprint arXiv:2310.07084 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.473, + 0.787, + 0.502 + ], + "angle": 0, + "content": "5. Balle, J., Laparra, V., Simoncelli, E.P.: End-to-end optimized image compression. In: 5th International Conference on Learning Representations, ICLR 2017 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.502, + 0.787, + 0.542 + ], + "angle": 0, + "content": "6. Balle, J., Minnen, D., Singh, S., Hwang, S.J., Johnston, N.: Variational image compression with a scale hyperprior. In: International Conference on Learning Representations (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.542, + 0.787, + 0.596 + ], + "angle": 0, + "content": "7. Barbu, A., Mayo, D., Alverio, J., Luo, W., Wang, C., Gutfreund, D., Tenenbaum, J., Katz, B.: Objectnet: A large-scale bias-controlled dataset for pushing the limits of object recognition models. Advances in neural information processing systems 32 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.596, + 0.787, + 0.637 + ], + "angle": 0, + "content": "8. Bau, D., Zhou, B., Khosla, A., Oliva, A., Torralba, A.: Network dissection: Quantifying interpretability of deep visual representations. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 6541-6549 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.637, + 0.787, + 0.678 + ], + "angle": 0, + "content": "9. Boigné, E., Parkinson, D.Y., Ihme, M.: Towards data-informed motion artifact reduction in quantitative ct using piecewise linear interpolation. IEEE Transactions on Computational Imaging 8, 917-932 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.678, + 0.787, + 0.718 + ], + "angle": 0, + "content": "10. Carlini, N., Wagner, D.: Adversarial examples are not easily detected: Bypassing ten detection methods. In: Proceedings of the 10th ACM workshop on artificial intelligence and security. pp. 3-14 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.718, + 0.787, + 0.746 + ], + "angle": 0, + "content": "1. Carlini, N., Wagner, D.: Towards evaluating the robustness of neural networks. In: 2017 ieee symposium on security and privacy (sp). pp. 39-57. IEEE (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.746, + 0.787, + 0.786 + ], + "angle": 0, + "content": "2. Chen, J., Jordan, M.I., Wainwright, M.J.: Hopskipjumpattack: A query-efficient decision-based attack. In: 2020 ieee symposium on security and privacy (sp). pp. 1277-1294. IEEE (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.786, + 0.787, + 0.84 + ], + "angle": 0, + "content": "3. Chen, S., Song, Z., Haque, M., Liu, C., Yang, W.: Nicgslowdown: Evaluating the efficiency robustness of neural image caption generation models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15365-15374 (2022)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.31, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.378, + 0.129 + ], + "angle": 0, + "content": "M. Zhang et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.787, + 0.203 + ], + "angle": 0, + "content": "14. Cheng, Z., Sun, H., Takeuchi, M., Katto, J.: Learned image compression with discretized gaussian mixture likelihoods and attention modules. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 7939-7948 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.203, + 0.787, + 0.243 + ], + "angle": 0, + "content": "15. Choi, H., Bajic, I.V.: Deep feature compression for collaborative object detection. In: 2018 25th IEEE International Conference on Image Processing (ICIP). pp. 3743-3747. IEEE (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.243, + 0.787, + 0.285 + ], + "angle": 0, + "content": "16. Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: Imagenet: A large-scale hierarchical image database. In: 2009 IEEE conference on computer vision and pattern recognition. pp. 248-255. IEEE (2009)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.285, + 0.787, + 0.325 + ], + "angle": 0, + "content": "17. Duda, J.: Asymmetric numeral systems: entropy coding combining speed of huffman coding with compression rate of arithmetic coding. arXiv preprint arXiv:1311.2540 (2013)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.325, + 0.787, + 0.367 + ], + "angle": 0, + "content": "18. Eshratifar, A.E., Esmaili, A., Pedram, M.: Bottlenet: A deep learning architecture for intelligent mobile cloud computing services. In: 2019 IEEE/ACM International Symposium on Low Power Electronics and Design (ISLPED). pp. 1-6. IEEE (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.367, + 0.787, + 0.42 + ], + "angle": 0, + "content": "19. Geirhos, R., Rubisch, P., Michaelis, C., Bethge, M., Wichmann, F.A., Brendel, W.: Imagenet-trained cnns are biased towards texture; increasing shape bias improves accuracy and robustness. In: International Conference on Learning Representations (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.42, + 0.787, + 0.447 + ], + "angle": 0, + "content": "20. Goodfellow, I.J., Shlens, J., Szegedy, C.: Explaining and harnessing adversarial examples. arXiv preprint arXiv:1412.6572 (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.447, + 0.787, + 0.473 + ], + "angle": 0, + "content": "21. Guo, C., Frank, J.S., Weinberger, K.Q.: Low frequency adversarial perturbation. arXiv preprint arXiv:1809.08758 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.473, + 0.787, + 0.514 + ], + "angle": 0, + "content": "22. Haque, M., Chauhan, A., Liu, C., Yang, W.: Ilfo: Adversarial attack on adaptive neural networks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14264-14273 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.514, + 0.787, + 0.555 + ], + "angle": 0, + "content": "23. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 770-778 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.555, + 0.787, + 0.61 + ], + "angle": 0, + "content": "24. Hendrycks, D., Basart, S., Mu, N., Kadavath, S., Wang, F., Dorundo, E., Desai, R., Zhu, T., Parajuli, S., Guo, M., et al.: The many faces of robustness: A critical analysis of out-of-distribution generalization. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 8340-8349 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.61, + 0.787, + 0.65 + ], + "angle": 0, + "content": "25. Hendrycks, D., Dietterich, T.: Benchmarking neural network robustness to common corruptions and perturbations. In: International Conference on Learning Representations (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.65, + 0.787, + 0.677 + ], + "angle": 0, + "content": "26. Hinton, G., Vinyals, O., Dean, J.: Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.677, + 0.787, + 0.717 + ], + "angle": 0, + "content": "27. Hong, S., Kaya, Y., Modoranu, I.V., Dumitraş, T.: A panda? no, it's a sloth: Slowdown attacks on adaptive multi-exit neural network inference. arXiv preprint arXiv:2010.02432 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.718, + 0.787, + 0.759 + ], + "angle": 0, + "content": "28. Howard, A., Sandler, M., Chu, G., Chen, L.C., Chen, B., Tan, M., Wang, W., Zhu, Y., Pang, R., Vasudevan, V., et al.: Searching for mobilenetv3. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 1314-1324 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.759, + 0.787, + 0.8 + ], + "angle": 0, + "content": "29. Howard, A.G., Zhu, M., Chen, B., Kalenichenko, D., Wang, W., Weyand, T., Andreetto, M., Adam, H.: Mobilenets: Efficient convolutional neural networks for mobile vision applications. arXiv preprint arXiv:1704.04861 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.8, + 0.787, + 0.841 + ], + "angle": 0, + "content": "30. Ilyas, A., Engstrom, L., Athalye, A., Lin, J.: Black-box adversarial attacks with limited queries and information. In: International conference on machine learning. pp. 2137-2146. PMLR (2018)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.328, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Resilience of Entropy Model in Distributed Neural Networks" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.204 + ], + "angle": 0, + "content": "31. Jacob, B., Kligys, S., Chen, B., Zhu, M., Tang, M., Howard, A., Adam, H., Kalenichenko, D.: Quantization and training of neural networks for efficient integer-arithmetic-only inference. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 2704-2713 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.205, + 0.787, + 0.233 + ], + "angle": 0, + "content": "32. Kingma, D.P., Welling, M.: Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114 (2013)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.234, + 0.787, + 0.288 + ], + "angle": 0, + "content": "33. Liu, H., Wu, Y., Yu, Z., Vorobeychik, Y., Zhang, N.: Slowlidar: Increasing the latency of lidar-based detection using adversarial examples. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5146-5155 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.29, + 0.787, + 0.331 + ], + "angle": 0, + "content": "34. Liu, P., Qi, B., Banerjee, S.: Edgeeye: An edge service framework for real-time intelligent video analytics. In: Proceedings of the 1st international workshop on edge systems, analytics and networking. pp. 1-6 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.333, + 0.787, + 0.374 + ], + "angle": 0, + "content": "35. Madry, A., Makelov, A., Schmidt, L., Tsipras, D., Vladu, A.: Towards deep learning models resistant to adversarial attacks. In: International Conference on Learning Representations (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.375, + 0.787, + 0.444 + ], + "angle": 0, + "content": "36. Masci, J., Meier, U., Ciresan, D., Schmidhuber, J.: Stacked convolutional autoencoders for hierarchical feature extraction. In: Artificial Neural Networks and Machine Learning-ICANN 2011: 21st International Conference on Artificial Neural Networks, Espoo, Finland, June 14-17, 2011, Proceedings, Part I 21. pp. 52-59. Springer (2011)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.446, + 0.787, + 0.473 + ], + "angle": 0, + "content": "37. Matsubara, Y., Yang, R., Levorato, M., Mandt, S.: Sc2 benchmark: Supervised compression for split computing. Transactions on machine learning research (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.475, + 0.787, + 0.515 + ], + "angle": 0, + "content": "38. Matsubara, Y., Callegaro, D., Baidya, S., Levorato, M., Singh, S.: Head network distillation: Splitting distilled deep neural networks for resource-constrained edge computing systems. IEEE Access 8, 212177-212193 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.517, + 0.787, + 0.584 + ], + "angle": 0, + "content": "39. Matsubara, Y., Callegaro, D., Singh, S., Levorato, M., Restuccia, F.: Bottlefit: Learning compressed representations in deep neural networks for effective and efficient split computing. In: 2022 IEEE 23rd International Symposium on a World of Wireless, Mobile and Multimedia Networks (WoWMoM). pp. 337-346. IEEE (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.587, + 0.787, + 0.628 + ], + "angle": 0, + "content": "40. Matsubara, Y., Levorato, M.: Neural compression and filtering for edge-assisted real-time object detection in challenged networks. In: 2020 25th International Conference on Pattern Recognition (ICPR). pp. 2272-2279. IEEE (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.63, + 0.787, + 0.67 + ], + "angle": 0, + "content": "41. Matsubara, Y., Levorato, M., Restuccia, F.: Split computing and early exiting for deep learning applications: Survey and research challenges. ACM Computing Surveys 55(5), 1-30 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.672, + 0.787, + 0.713 + ], + "angle": 0, + "content": "42. Matsubara, Y., Yang, R., Levorato, M., Mandt, S.: Supervised compression for resource-constrained edge computing systems. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 2685-2695 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.715, + 0.787, + 0.755 + ], + "angle": 0, + "content": "43. Minnen, D., Balle, J., Toderici, G.D.: Joint autoregressive and hierarchical priors for learned image compression. Advances in neural information processing systems 31 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.757, + 0.787, + 0.798 + ], + "angle": 0, + "content": "44. Minnen, D., Singh, S.: Channel-wise autoregressive entropy models for learned image compression. In: 2020 IEEE International Conference on Image Processing (ICIP). pp. 3339-3343. IEEE (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.8, + 0.787, + 0.84 + ], + "angle": 0, + "content": "45. Pan, J., Foo, L.G., Zheng, Q., Fan, Z., Rahmani, H., Ke, Q., Liu, J.: Gradmdm: Adversarial attack on dynamic networks. IEEE Transactions on Pattern Analysis and Machine Intelligence (2023)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.378, + 0.129 + ], + "angle": 0, + "content": "M. Zhang et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.148, + 0.785, + 0.189 + ], + "angle": 0, + "content": "46. Pan, J., Zheng, Q., Fan, Z., Rahmani, H., Ke, Q., Liu, J.: Gradauto: Energy-oriented attack on dynamic neural networks. In: European Conference on Computer Vision. pp. 637-653. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.19, + 0.785, + 0.232 + ], + "angle": 0, + "content": "47. Radosavovic, I., Kosaraju, R.P., Girshick, R., He, K., Dóllár, P.: Designing network design spaces. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10428-10436 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.232, + 0.785, + 0.272 + ], + "angle": 0, + "content": "48. Ran, X., Chen, H., Zhu, X., Liu, Z., Chen, J.: Deepdecision: A mobile deep learning framework for edge video analytics. In: IEEE INFOCOM 2018-IEEE conference on computer communications. pp. 1421-1429. IEEE (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.273, + 0.785, + 0.327 + ], + "angle": 0, + "content": "49. Rifai, S., Vincent, P., Muller, X., Glorot, X., Bengio, Y.: Contractive auto-encoders: Explicit invariance during feature extraction. In: Proceedings of the 28th international conference on international conference on machine learning. pp. 833-840 (2011)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.328, + 0.785, + 0.356 + ], + "angle": 0, + "content": "50. Rissanen, J., Langdon, G.: Universal modeling and coding. IEEE Transactions on Information Theory 27(1), 12-23 (1981)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.357, + 0.785, + 0.384 + ], + "angle": 0, + "content": "51. Rudin, L.I., Osher, S., Fatemi, E.: Nonlinear total variation based noise removal algorithms. Physica D: nonlinear phenomena 60(1-4), 259-268 (1992)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.384, + 0.785, + 0.424 + ], + "angle": 0, + "content": "52. Sandler, M., Howard, A., Zhu, M., Zhmoginov, A., Chen, L.C.: Mobilenetv2: Inverted residuals and linear bottlenecks. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 4510-4520 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.425, + 0.785, + 0.452 + ], + "angle": 0, + "content": "53. Shannon, C.E.: A mathematical theory of communication. The Bell system technical journal 27(3), 379-423 (1948)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.453, + 0.785, + 0.493 + ], + "angle": 0, + "content": "54. Shao, J., Zhang, J.: Bottlenet++: An end-to-end approach for feature compression in device-edge co-inference systems. In: 2020 IEEE International Conference on Communications Workshops (ICC Workshops). pp. 1-6. IEEE (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.494, + 0.785, + 0.535 + ], + "angle": 0, + "content": "55. Singh, S., Abu-El-Haija, S., Johnston, N., Balle, J., Shrivastava, A., Toderici, G.: End-to-end learning of compressible features. In: 2020 IEEE International Conference on Image Processing (ICIP). pp. 3349-3353. IEEE (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.536, + 0.785, + 0.577 + ], + "angle": 0, + "content": "56. Szegedy, C., Zaremba, W., Sutskever, I., Bruna, J., Erhan, D., Goodfellow, I., Fergus, R.: Intriguing properties of neural networks. In: 2nd International Conference on Learning Representations, ICLR 2014 (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.578, + 0.785, + 0.632 + ], + "angle": 0, + "content": "57. Tan, M., Chen, B., Pang, R., Vasudevan, V., Sandler, M., Howard, A., Le, Q.V.: Mnasnet: Platform-aware neural architecture search for mobile. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 2820-2828 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.633, + 0.785, + 0.674 + ], + "angle": 0, + "content": "58. Wang, H., Ge, S., Lipton, Z., Xing, E.P.: Learning robust global representations by penalizing local predictive power. Advances in Neural Information Processing Systems 32 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.674, + 0.785, + 0.729 + ], + "angle": 0, + "content": "59. Yuan, Z., Rawlekar, S., Garg, S., Erkip, E., Wang, Y.: Feature compression for rate constrained object detection on the edge. In: 2022 IEEE 5th International Conference on Multimedia Information Processing and Retrieval (MIPR). pp. 1-6. IEEE (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.729, + 0.785, + 0.771 + ], + "angle": 0, + "content": "60. Zhang, W., Zhang, Z., Zeadally, S., Chao, H.C., Leung, V.C.: Masm: A multiple-algorithm service model for energy-delay optimization in edge artificial intelligence. IEEE Transactions on Industrial Informatics 15(7), 4216-4224 (2019)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.148, + 0.785, + 0.771 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/Resilience of Entropy Model in Distributed Neural Networks/ad486483-132b-4d7f-84db-26c1a65f3dae_origin.pdf b/2024/Resilience of Entropy Model in Distributed Neural Networks/ad486483-132b-4d7f-84db-26c1a65f3dae_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..b99132865dc8cd9d1d78a9cbe596f3db56a07f34 --- /dev/null +++ b/2024/Resilience of Entropy Model in Distributed Neural Networks/ad486483-132b-4d7f-84db-26c1a65f3dae_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c32475b99f6d18ee154e5b668b5f96f4c03eeb482e43b1ed3af74d191d7b86a9 +size 1721627 diff --git a/2024/Resilience of Entropy Model in Distributed Neural Networks/full.md b/2024/Resilience of Entropy Model in Distributed Neural Networks/full.md new file mode 100644 index 0000000000000000000000000000000000000000..e4e9624df642f20f37888d03d09a6cc273fdd28e --- /dev/null +++ b/2024/Resilience of Entropy Model in Distributed Neural Networks/full.md @@ -0,0 +1,370 @@ +# Resilience of Entropy Model in Distributed Neural Networks + +Milin Zhang, Mohammad Abdi, Shahriar Rifat, and Francesco Restuccia + +Institute for the Wireless Internet of Things Northeastern University, Boston MA 02115, USA {zhang.mil, abdi.mo, rifat.s, f/restuccia}@northeastern.com + +Abstract. Distributed deep neural networks (DNNs) have emerged as a key technique to reduce communication overhead without sacrificing performance in edge computing systems. Recently, entropy coding has been introduced to further reduce the communication overhead. The key idea is to train the distributed DNN jointly with an entropy model, which is used as side information during inference time to adaptively encode latent representations into bit streams with variable length. To the best of our knowledge, the resilience of entropy models is yet to be investigated. As such, in this paper we formulate and investigate the resilience of entropy models to intentional interference (e.g., adversarial attacks) and unintentional interference (e.g., weather changes and motion blur). Through an extensive experimental campaign with 3 different DNN architectures, 2 entropy models and 4 rate-distortion trade-off factors, we demonstrate that the entropy attacks can increase the communication overhead by up to $95\%$ . By separating compression features in frequency and spatial domain, we propose a new defense mechanism that can reduce the transmission overhead of the attacked input by about $9\%$ compared to unperturbed data, with only about $2\%$ accuracy loss. Importantly, the proposed defense mechanism is a standalone approach which can be applied in conjunction with approaches such as adversarial training to further improve robustness. Code is available at https://github.com/Restuccia-Group/EntropyR. + +Keywords: Trustworthy Machine Learning $\cdot$ Data Compression + +# 1 Introduction + +Distributed deep neural network (DNN) were recently introduced to divide the computation of DNNs across various devices based on their available computation and communication resources. They have been shown to be extremely effective to implement deep learning applications on resource-constrained mobile devices [18,39,54]. As depicted in Fig. 1, the common strategy is to divide a large DNN into a small head network deployed on the mobile device to extract and compress features, and a tail network on the server to perform the task inference. Compared to conventional lightweight DNNs specifically designed for + +![](images/cfc42dff6d0765a2b15c7111c775dae97d502cada9ec96fd74a01aa141e06523.jpg) +$\times$ Computation Resource + +![](images/2d9c27eab88fe9e37510eec34e286a49a5dc37a87f3efee3b699947b16e5d4e1.jpg) +Edge Server + +Fig. 1: The key concept of distributed DNN. Left: Lightweight DNNs suffer from performance loss due to the limited computational resources; Middle: Edge computing often results in intolerable latency due to communication overhead; Right: Distributed DNNs deploy a small head model on the resource-limited device to achieve task-oriented compression and a large tail model on the server to decode compressed features and execute the rest of model. +![](images/20c631db94e73d6a0e29ec6ccad33b513f0d3e30c1e58a15939dcfabecbc105e.jpg) +X Communication Overhead + +![](images/f0d93f33eb70a83800dad5a479c70c13af5ffa5fc83eb3e3759457e95b0cf61a.jpg) +$\checkmark$ Performance & Latency Trade-off + +mobile devices [28, 29, 31, 52, 57], distributed DNNs can leverage the computation power of the edge/cloud and hence attain better performance. In addition, distributed DNNs leverage compression techniques to reduce the data size of intermediate representations. By only transmitting compact representations, distributed DNNs can reduce the transmission overhead significantly compared to traditional edge computing [34, 48, 60]. + +Interestingly, recent work has proposed to further minimize the data size of transmitted latent representations with entropy coding [42]. The key idea is to train the distributed DNN jointly with an auxiliary entropy model, which estimates the distribution of latent representations. The output of the entropy model is then used as side information during inference time to facilitate the adaptive encoding of latent representations into bit streams with variable lengths. The proposed approach achieves better compression rate compared to quantization-based [38] and codec-based [2] approaches by minimizing the entropy of latent representations. Despite its excellent compression performance, the resilience of entropy coding to intentional interference (e.g., specifically crafted adversarial attacks) and unintentional interference (e.g., sudden events such as weather changes and motion blur) remains unexplored. We remark that in entropy-coding-based distributed DNNs, the bit rate of encoded representations relies completely on its estimated entropy. However, it is well known that DNNs trained on standard datasets are vulnerable to distribution shifts [24,25] and adversarial actions [20, 56]. As the entropy model is learned without considering any corruptions and adversarial perturbations, a small change in the input space might lead to a large estimate of entropy, and hence result in a bit rate that can exceed the transmission bandwidth. As depicted in Tab. 3 in our experiments, the transmission overhead can be increased by about $2\mathrm{x}$ in the worst case. + +In this work, we thoroughly assess for the first time the resilience of entropy models to intentional and unintentional interference as applied to distributed DNNs. Intriguingly, our findings unveil that the auxiliary entropy model learns + +a different set of input-related features than what learned by the backbone DNN. Ultimately, this enables us to design effective defense strategy to maintain the bit rate with an unnoticeable loss in the task performance. Our approach successfully reduces the communication overhead of perturbed data by about $9\%$ compared to unperturbed inputs, with only about $2\%$ accuracy loss. + +# Summary of Novel Contributions + +(1) We investigate for the first time the effect of intentional and unintentional interference to entropy coding in distributed DNNs. We show that interference increases the end-to-end latency of distributed DNNs and poses a threat to other users by saturating the transmission bandwidth. Through comprehensive experiments involving 3 distinct DNN architectures [23, 47], 2 entropy models [5,43], and 4 rate-distortion trade-off factors, we illustrate that attacks to entropy coding may increase the transmitted data size by up to about $95\%$ ; +(2) We design two visualization approaches to interpret the compression learned by entropy models. By disentangling features relevant to compression and classification across both frequency and spatial domains, we reveal that the entropy model learns specific features that are distinct from those beneficial for the end task. This finding enables us to devise an effective defense strategy that safeguards the vulnerable compression features while minimally impacting task performance; +(3) We propose a defense approach to attacks targeting entropy coding based on our findings discussed above. The proposed method effectively reduces the transmission overhead of attacked inputs by about $9\%$ in comparison to the unperturbed case, with only about $2\%$ decrease in accuracy. We remark that our approach is general in nature and can be combined with approaches such as adversarial training to further improve the resilience of entropy models. + +The paper is organized as follows. Sec. 2 provides background of entropy coding in distributed DNNs. In Sec. 3, we formally describe the threat model for entropy coding in distributed DNNs. Sec. 4 presents benchmarking results on both intentional and unintentional interference. Sec. 5 details the proposed defense approach and comprehensively evaluates it against adaptive attacks. Related work is summarized in Sec. 6. Finally, we draw conclusions and discuss future work in Sec. 7. + +# 2 Entropy Coding in Distributed DNNs + +Entropy coding techniques such as arithmetic coding [50] and asymmetric numeral systems [17] can encode a message with the optimal coding rate by leveraging the probabilistic information of the message. Based on the source coding theorem [53], the entropy of a message $z$ with distribution $P_Z(z)$ defines the lower bound of expected coding rate $R_{c}(z)$ without any loss of information, i.e., + +$$ +\mathbb {E} \left\{R _ {c} (z) \right\} \geq H _ {P _ {Z}} (z) = \mathbb {E} \left\{- \log_ {2} P _ {Z} (z) \right\} \tag {1} +$$ + +![](images/46ef122df4dc911a3fd3a7389796a253b136687169eb3c9e6e564c316d04dc74.jpg) +Fig. 2: A general framework for entropy coding in neural data compression. It consists of an encoder-decoder structure to extract a compact latent representation $z$ and a prior model to estimate the distribution $P_Z(z)$ . Left: During training, the prior model is jointly optimized with the encoder-decoder backbone. Right: During inference, the learned prior $P_Z(z)$ is used as side information for arithmetic encoding/decoding. + +![](images/3de526a1910bfa5aee1f6e50f103cbb4ed25db29ead2cccb45d90042a9d0c6eb.jpg) + +where $H_{P_Z}(z)$ is the entropy of $z$ . In recent data compression literature [1,5,6, 14,43,44], the entropy coding is integrated with lossy DNN-based compression to attain an effective compression rate with less information loss compared to traditional approaches. As depicted in Fig. 2, the common strategy is to jointly train an encoder-decoder backbone for extracting compact latent representations $z$ with an auxiliary prior model $P_Z(z)$ . During training, a random noise source is introduced to simulate the effect of quantization on $z$ , allowing the backpropagation of gradients. In the inference phase, $z$ is initially quantized to integers for entropy coding. Next, the output of the prior model is used as side information to encode the quantized $z$ into bit streams adaptively with variable length. + +DNN-based data compression can achieve the near-optimal compression rate by minimizing the entropy of latent representations during training. Specifically, the objective is to minimize the rate-distortion function + +$$ +\mathcal {L} (x, (\hat {x}, z)) = \underbrace {\| x - \hat {x} \| _ {2} ^ {2}} _ {\text {d i s t o r t i o n}} - \beta \cdot \underbrace {\log_ {2} P _ {Z} (z)} _ {\text {r a t e}}, \tag {2} +$$ + +where $\hat{x}$ is the reconstructed data, $||x - \hat{x}||_2^2$ is the loss function of the encoder-decoder backbone and $-\log_2P_Z(z)$ is the loss function of the prior model. Constant $\beta$ denotes the trade-off between the rate and distortion. + +Recent work [42] has incorporated entropy coding into distributed DNNs to increase communication efficiency between mobile and edge devices. The proposed method, referred to as the entropic student, aims at achieving a balance between the coding rate and task-specific performance metrics (e.g., cross entropy in classification tasks). One distinction between the entropic student and other entropy-coding-based feature compression work [55, 59] lies in the significantly smaller encoder size of the former due to constraints on computational resources in mobile devices. As a result, the entropic student splits the DNN at an early layer and use knowledge distillation [26] for training the head to preserve accuracy. However, existing work does not address the issue of resilience of entropy models, which is the key goal of this paper. A relevant work [4] investigated the robustness of density estimation aiming to maximize $P_Z(z)$ while we are interested in perturbations that minimize $P_Z(z)$ . + +# 3 Modeling Interference to Entropy Coding + +While the resilience of DNNs to distribution shifts [7,19,24,25,58] and adversarial actions [3,11,12,30,35,56] has been extensively investigated, we investigate types of interference that can alter the length of the encoded bit stream, which ultimately leads to increased bandwidth utilization. In this section, we formulate the resilience of compression in entropy-coding-driven distributed DNNs. + +Threat Model for Accuracy. Let $D = \{x, y\}$ denotes a dataset for classification problem where $x \in \mathcal{X}$ and $y \in \mathcal{Y}$ are specific input and corresponding label samples respectively, while $\mathcal{X}, \mathcal{Y}$ are input and output space, respectively. There exists a prior yet unknown distribution $P(y|x)$ such that + +$$ +P (y | x) \geq P (\hat {y} | x) \quad \forall \hat {y} \in \mathcal {Y} \tag {3} +$$ + +where $\hat{y}$ denotes incorrect labels. The DNN classifier, is trained to approximate the distribution $\tilde{P}(y|x) \approx P(y|x)$ with empirical risk minimization approaches based on an assumption that in-sample data and out-of-sample data shared a similar distribution $P(y|x)$ . However, the unintentional interference denoted by $\delta$ , which can be modeled as an additive noise to $x$ , creates a new dataset $D' = \{x + \delta, y\}$ . It introduces a covariate shift that $P(x + \delta) \neq P(x)$ but $P(y|x + \delta) = P(y|x)$ , resulting in classification errors such that + +$$ +\tilde {P} (\hat {y} | x + \delta) \geq \tilde {P} (y | x + \delta) \quad \exists \hat {y} \in \mathcal {Y} \tag {4} +$$ + +In the intentional interference scenario, algorithms attempt to find the minimum covariate shift that leads to Eq. (4). In this setting, a small perturbation in $l_{p}$ distance $||\delta||_{p} \leq \epsilon$ is intentionally crafted to mislead the DNN classifier where $\epsilon$ denotes the constraint. Attackers formulate Eq. (4) as an optimization problem over the input space $\mathcal{X}$ , and hence can apply the gradient information of the loss function w.r.t. $x$ to find the $\delta$ . Projected Gradient Descent (PGD) [35] provided a unified view on iterative gradient attacks under $l_{p}$ constraint: + +$$ +x ^ {i + 1} = x ^ {i} + \alpha \cdot \epsilon \cdot \frac {\nabla_ {x} \mathcal {L}}{| | \nabla_ {x} \mathcal {L} | | _ {p}} \tag {5} +$$ + +where $x^{i}$ is the adversarial sample at $i$ step, $\alpha$ is the learning rate, and $\frac{\nabla_{x}\mathcal{L}}{||\nabla_{x}\mathcal{L}||_{p}}$ is the normalized gradient of a loss function $\mathcal{L}$ (e.g., a sign function in $l_{\infty}$ space). $\mathcal{L}$ can be cross entropy or other more advanced loss functions [11]. + +Threat Model for Data Rate. Let $H(x) = z$ denotes the head network where $z \in \mathcal{Z}$ is the output of the head and $\mathcal{Z}$ is the latent space. The goal is to achieve a minimum compression rate $-\log_2 P_Z(z)$ . As $z$ is dependent on $x$ , the distribution shift introduced by unintentional interference can result in a difference in coding rate: + +$$ +- \log_ {2} P _ {Z} (H (x + \delta)) \neq - \log_ {2} P _ {Z} (H (x)) \tag {6} +$$ + +Table 1: DNNs, entropy models and datasets utilized in experiments. + +
DNN Architectures
ResNet-50 [23]A standard baseline in distributed DNNs [38-40, 42]
ResNet-101 [23]A deeper model in the ResNet family
RegNetY-6.4GF [47]An advanced design with better performance
Entropy Models
FP [5]A fully factorized prior model
MSHP [43]An effective learnable entropy model
Datasets
ImageNet [16]The standard ImageNet validation dataset
ImageNet-C [25]A synthetic dataset with 15 corruptions and 5 severities
Random NoiseNoisy images as a baseline for adversarial robustness
PGD-Acc [35]PGD targeting classification
PGD-EPGD targeting entropy
+ +Note that the unintentional distribution shift of $x$ does not necessarily result in a larger bit rate. However, in the adversarial setting, attackers intend to maximize the entropy of $z$ as follows: + +$$ +\min _ {\delta} \log_ {2} P _ {Z} (H (x + \delta)) \tag {7} +$$ + +$$ +s. t. \quad | | \delta | | _ {p} \leq \epsilon +$$ + +Thus, to correctly assess the robustness of entropy models, we use the same PGD algorithm in Eq. (5) yet with the entropy as its loss function $\mathcal{L} = -\log_2P_Z(H(x))$ + +# 4 Understanding the Resilience of Entropy Model + +# 4.1 Experimental Setup + +In this section, we outline the experimental setup for evaluating the resilience of entropy models. We first specify the DNNs and entropy models under consideration, followed by a description of the metrics and datasets used in our experiments. A summary of our experimental setup is provided in Tab. 1. + +Ablation of DNNs and Entropy Models. We consider three distinct DNNs: ResNet-50 [23], ResNet-101 [23] and RegNetY-6.4GF [47]. ResNet-50 serves as a common baseline in distributed DNN literature [38-40,42] while ResNet-101 is a deeper variant with the same design as former. RegNetY-6.4GF has an advanced design with better classification performance. The early layers are replaced by a specific tailored head architecture as proposed in [42] and incorporated with different entropy models. For the sake of brevity, we consider ResNet-50 as the default DNN architecture without additional specification. We employ two entropy + +![](images/749e0e30e507829f0b3bba2cddea1397c49d6d7a843a8db6bd6add6a20458096.jpg) +(a) resilience w.r.t. $\beta$ + +![](images/e54d2744cd2b3bda3702db2cc58fa4f3936dcf54b086ee1492e3bec496597862.jpg) +(b) resilience w.r.t. severities +Fig. 3: Resilience of accuracy and data rate to 4 common corruptions. + +models: Factorized Prior (FP) [5] and Mean Scale Hyper Prior (MSHP) [43]. FP is the earliest design of entropy model and a basic component in many other advanced designs, while MSHP provides more powerful compression performance by injecting a hierarchical learnable block before FP to extract the entropy information. For conciseness, we consider MSHP as the default entropy model. In addition, since models are trained with different rate-distortion trade-off $\beta$ , we also investigate the compression robustness w.r.t. different $\beta$ . + +Datasets and Metrics. Conversely from existing work that measures the compression performance with bits per pixel (BPP) [5,6,14,43,44], we evaluate the entropy model with the data size of the whole bit stream after encoding as in [37]. Compared to BPP, data size quantifies more directly the networking traffic between the head and tail. To assess the resilience to unintentional interference, we use the ImageNet-C dataset proposed by [25] comprising 15 common corruptions classified into 4 categories (noise, blur, digital, and weather) with 5 different severities. We also report the data size and accuracy on the clean ImageNet [16] validation set as a baseline. To assess the resilience to intentional interference, we implement the PGD targeting both the classification and compression performance in $l_{\infty}$ space. For clarity, we denote the conventional PGD as PGD-Acc and the PGD targeting entropy models as PGD-E. Meanwhile, random noise with the same perturbation level is added to the clean ImageNet as a baseline comparison for adversarial robustness. + +# 4.2 Resilience to Unintentional Interference + +Resilience w.r.t. $\beta$ . We select one corruption as a representative from each category in ImageNet-C [25] to investigate the resilience as a function of ratedistortion trade-off $\beta$ . Fig. 3a shows the compression and classification performance on defocus blur, shot noise, snow and contrast dataset, each with the severity of 5. Compared to the clean ImageNet, all corruptions decrease the accuracy significantly. However, not all corruptions increase the data size. Only + +Table 2: Resilience w.r.t. prior models. The average data size of FP and MSHP for clean ImageNet are $11.65 \pm 1.02$ and $9.62 \pm 1.70$ KBBytes, respectively. Red for larger data size and blue for lower compared to the baseline. + +
Prior +ModelGaussian +Size[KB]Noise +Acc[%]Motion +Size[KB]Blur +Acc[%]Impulse +Size[KB]Noise +Acc[%]Glass +Size[KB]Blur +Acc[%]
FP13.19±0.4059.5110.51±0.7062.2213.36±0.3450.9310.43±0.6653.63
MSHP11.96±0.4759.367.88±1.4661.5812.52±0.3551.127.78±1.3953.67
+ +shot noise increases the data size by $65.31\%$ on average. In contrast, Defocus blur and contrast decrease the data size by $53.31\%$ and $67.45\%$ on average, respectively. Snow has little affect to compression, resulting in only $4.42\%$ decrease of data size. This indicates that the entropy model learns a different set of features than the classification model. Since it shows the same trend for different rate-distortion trade-off factors, we only consider $\beta = 0.08$ in next experiments. + +Resilience w.r.t. Severities. Next, we explore resilience w.r.t. severity levels using the same corruptions mentioned earlier. As illustrated in Fig. 3b, the data size of shot noise rises with increasing severity, whereas defocus blur and contrast consistently decrease the data size. Snow shows minimal variation across different severity levels. This can be attributed to defocus blur and contrast removing high-frequency components in images, while shot noise introduces new patterns in high-frequency space. Snow, on the other hand, lacks a specific pattern in the frequency domain. Thus, the compression is particularly sensitive to high-frequency information in the input space. + +Resilience w.r.t. Prior Models. Tab. 2 shows the impact of additional 4 corruptions at severity level 1 on both FP and MSHP model. To further validate the previous observation, we choose 2 noise corruptions (Gaussian noise and impulse noise), which introduce high-frequency noise to images, and two blur corruptions (motion blur and glass blur), which eliminate high-frequency components from images. As shown in Tab. 2, Gaussian noise increases $13.22\%$ and $24.32\%$ of data size for FP and MSHP, respectively. Impulse noise increases $14.42\%$ and $30.15\%$ of data size for FP and MSHP, respectively. On the other hand, motion blur reduces $9.79\%$ and $18.09\%$ of data size for FP and MSHP, respectively. Glass blur reduces $10.47\%$ and $19.13\%$ of data size for FP and MSHP, respectively. Thus, the previous finding that compressive features reside in the high-frequency space remains consistent for different entropy models. + +# 4.3 Resilience to Intentional Interference + +Resilience w.r.t. DNNs Fig. 4a shows the resilience of different DNN architectures to accuracy and entropy attacks with the $l_{\infty}$ constraint $\epsilon = 4 / 255$ . As evident in the figure, the attacks show similar patterns regardless of their + +![](images/c1422fe811d72e379e187f90ad9483c30699ed87429a3cf3260ddb21859c5d72.jpg) +(a) resilience w.r.t. DNN architectures + +![](images/2fdb8b2a921a1e5acb43a4edba73095d1962eed910e547599bc7c07565b2c889.jpg) +(b) resilience w.r.t. prior models +Fig. 4: Resilience of accuracy and data rate to intentional interference + +victim models. While PGD-Acc reduces $42.88\%$ accuracy on average, it only increases by $7.28\%$ of data size on average compared to random noise. On the other hand, PGD-E decreases $2.34\%$ accuracy on average compared to random noise but increases the bit rate 1.35x times on average. This indicates that PGD-E and PGD-Acc are targeting separate features in the input space which has minor impact on each other. + +Resilience w.r.t. Prior Models. Fig. 4b illustrates the effect of adversarial attacks with $\epsilon = 8 / 255$ on victims using different prior models. As shown in the figure, both entropy models show considerable vulnerability to PGD-E compared to PGD-Acc. The data size are increased by $21.93\%$ and $46.82\%$ for FP and MSHP respectively while the performance loss are $9.91\%$ and $6.62\%$ respectively. On the other hand, PGD-Acc successfully reduces $72.10\%$ and $57.10\%$ classification performance of FP and MSHP compared to random noise respectively. However, it only increases $5.66\%$ and $10.19\%$ bit rate for FP and MSHP respectively. Hence, attackers targeting compression tend to perturb a different set of features than attackers targeting discriminative tasks regardless of the entropy models. In addition, while the FP has a better robustness in compression, it has significantly worse robustness in accuracy. + +# 5 Proposed Defense Mechanism + +# 5.1 Disentangle Compression Features in Frequency Domain + +As shown in Sec. 4.2, interference that introduces high-frequency noise (e.g., shot noise) can effectively increase the data size while interference that removes high-frequency information (e.g., defocus blur) reduces the data size. Intuitively this is a consequence of the small head in distributed DNNs as semantic information is usually captured in deeper layers while early layers tend to capture low level information [8]. To validate our intuition, we introduce total variation, defined + +![](images/8cd8faf40432ed1024ee9ac561f449d17f469e27643b54f9ea85ec2200e8c119.jpg) + +![](images/d0636016d8cacc5ccf4eb783d04afd96f3e7ad8320ae5d84c517527343444d60.jpg) +Fig. 5: Entropy models are sensitive to high-frequency features. From left to right: images from ImageNet validation set, bit rate maps and total variation maps. + +![](images/be791e5b2e7d3b8a3a3024edacae1bae51cc2b4333197a34a697c1a2042fe377.jpg) + +![](images/b7634bdb8b1a114c0539bd8177bf72e1cec1176c0bf8e3781ca928186e8a3dfb.jpg) + +![](images/3006b3bbba65f4a2dd721196d2cd4fae009fca8f33c317c9682df03784ea6d6c.jpg) + +![](images/d31bd2c70df5c30176ed942294fe5476b0fc9e7d7c1731b84f8791d293d2b878.jpg) + +![](images/3ae55c9591f3169f26560aa464c58fd4bcd61295e1a0316061af2c26227f1df0.jpg) + +![](images/7d7c7698af0229855a4a33345826e806f3880bf4ba5a76002209c13e29b5be08.jpg) + +![](images/e2331d1baeff47774cf21ff0223a1803f1f80f68192ad5c4b04b6c6b6432331b.jpg) + +![](images/59b7d051e5b9660d0b36ec34203cecec281c924a50f04b35f49d529116c67fc4.jpg) + +![](images/f09587c113ad775d597b521159ddf19a1f4bfc4bf9757a2d0a514404d9be0c54.jpg) + +![](images/180fe7f9967c9d145a06778637bf0f9f8693da31097a41269fe68c12ed915f5d.jpg) + +as the integral of image gradient magnitude, which is first proposed for image denoising in [51]. The anitrosopic total variation of a 2-D image is defined as + +$$ +T V (x) = \sum_ {i, j} \left| x _ {i + 1, j} - x _ {i, j} \right| + \left| x _ {i, j + 1} - x _ {i, j} \right| \tag {8} +$$ + +where $x$ denotes an image sample, $i,j$ denotes the pixel position in width and height, respectively. The image gradient magnitude $|x_{i+1,j} - x_{i,j}|$ and $|x_{i,j+1} - x_{i,j}|$ describe the sharpness of pixel changes, where larger magnitude indicates high-frequency information such as edges and textures in images. + +We compare the total variation map of an image with its bit rate map generated by the entropy model. To plot the total variation map, we first slice an image into many small patches and then compute the total variation for each patch. Meanwhile, the bit rate map is the estimated entropy of the latent representation $-\log_2P_Z(z)$ . As shown in Fig. 5, there exists a significant correlation between the bit rate map and the total variation map, which indicates that the entropy model is sensitive to high-frequency features in the input space. + +# 5.2 Disentangle Compression Features in Spatial Space + +The work in [42] has shown that entropy model in distributed DNNs tends to assign a larger number of bits to task-oriented information, such as objects, while compressing irrelevant information, like the background. Consequently, it becomes intuitively easier for an adversary to increase the bit rate in the background area to attack the entropy model. + +To interpret the action conducted by attackers, we generate the bit rate comparison map between adversarial samples generated by PGD-E and PGD-Acc. This comparison map represents the estimated entropy of adversarial samples generated by PGD-E subtracted from the bit rate of PGD-Acc-perturbed samples. In this context, the red area indicates that PGD-E tends to allocate more bits compared to PGD-Acc, whereas the blue region suggests that PGD-E tends to allocate fewer bits. As depicted in Fig. 6, PGD-E puts more efforts to increase the bit rate in background region, whereas PGD-Acc focuses on altering + +![](images/f4b643126eed029e66abdfbf9d68d09b122c66ab69117af0bf015a834d40eddb.jpg) +Fig. 6: Entropy models are more vulnerable to perturbation in background region. From left to right: images from ImageNet validation set, bit rate comparison between PGD-E and PGD-Acc (PGD-E allocate less bit in blue area and more bit in red area). + +![](images/5e54bca6224bab3090b457b056900a27ea21b7a0ef4fe1d7bcd69af0154c4e99.jpg) + +![](images/fd4b0018db2a647c063c2563f557969b90e7d2619cea840d2880685b297bdbda.jpg) + +![](images/83b8dd48844c0a14b260f22c8f8b706d357945e0ac4d712130a7d2d9f3d48562.jpg) + +![](images/25da3dd17cf80b197cf2144dd097b4648a714051e107f57cfd08bd79aed2d207.jpg) + +![](images/2669ece1f428339965f782c3128b1b89cad20ebd9895799225f4316d6befe9c1.jpg) + +task-oriented information. This is in line with the experiments in Sec. 4.3 that PGD-E and PGD-Acc target different features in the input space. On one hand, increasing the bit rate in the background region will affect little accuracy as they are irrelevant information for the discriminative task. On the other hand, modifying object information can significantly degrade the performance, yet it has less impact on compression. + +# 5.3 Object-Aware Total Variation Denoising + +Given the observation in Sec. 5.1, we propose a denoising technique based on total variation to remove high-frequency noise in adversarial images where we solve the following optimization problem: + +$$ +\min _ {x} \frac {1}{2} \| x - x ^ {\prime} \| _ {2} ^ {2} + \lambda \cdot T V (x), \tag {9} +$$ + +where $x'$ is the distorted image, $x$ is the denoised image, $TV(x)$ is the total variation of $x$ and $\lambda$ is a regularization factor to control the degree of smoothing. Eq. (9) means the image $x$ after denoising should keep most of the information in $x'$ (i.e., minimize $||x - x'||_2^2$ ) but discard the high-frequency information (i.e., minimize $TV(x)$ ). Optimizing Eq. (9) is not trivial as total variation is non-differentiable. As a result, we use a sub-gradient descent: + +$$ +x ^ {i + 1} = x ^ {i} - \alpha \cdot \left(\left(x ^ {i} - x ^ {\prime}\right) + \lambda \cdot g \left(x ^ {i}\right)\right) \tag {10} +$$ + +where $x^i$ is the denoised image at the step $i$ , $\alpha$ is the step size and $g(x^i)$ is the image gradient of $x^i$ [9]. + +Simply applying the total variation denoising will also decrease the task-oriented performance as it also removes useful information in high-frequency domain. Since adversarial attacks targeting entropy models attempt to increase bit rate in background region as demonstrated in Sec. 5.2, we add a mask to Eq. (10) to force the denoising algorithm to only remove high-frequency information in non-object space, that is, + +$$ +x ^ {i + 1} = x ^ {i} - \alpha \cdot m \cdot \left(\left(x ^ {i} - x ^ {\prime}\right) + \lambda \cdot g \left(x ^ {i}\right)\right) \tag {11} +$$ + +where $m$ is the mask to control the denoising level. In practice, we interpolate the output of entropy model $P_Z(z)$ as the soft mask. This is because the higher bit rate $-\log_2P_Z(z)$ corresponding to object area has a smaller value of $P_Z(z)$ , making it naturally a soft mask to avoid smoothing the object region. + +Table 3: Compression and classification performance before and after defense. The average data size for clean ImageNet is $9.62 \pm 1.70$ KBytes. + +
ε = 2/255ε = 4/255ε = 8/255ε = 16/255
Size[KB]Acc[%]Size[KB]Acc[%]Size[KB]Acc[%]Size[KB]Acc[%]
before12.35±0.6373.2614.11±0.3871.3316.15±0.2965.6218.79±0.3752.13
after8.76±1.3870.829.55±1.0770.1511.02±0.6567.2213.44±0.4859.87
+ +Experimental Results. We investigate the proposed defense method as a function of perturbation budget $\epsilon$ . As shown in Tab. 3, attackers targeting entropy significantly increase the data size, by up to $95.32\%$ when $\epsilon = 16/255$ compared to the clean images (9.62 KBytes). On the contrary, our denoising approach significantly reduce the data size for all $\epsilon$ . For $\epsilon = 2/255, 4/255$ , the average data size is even $8.94\%$ and $0.73\%$ smaller than the clean images with only $2.44\%$ and $1.18\%$ accuracy loss, respectively. For $\epsilon = 8/255, 16/255$ , the data size is reduced by $31.80\%$ and $28.47\%$ respectively, while the accuracy improved slightly after denoising ( $1.60\%$ for $\epsilon = 8/255$ and $7.74\%$ for $\epsilon = 16/255$ ). + +We remark that the proposed defense is a standalone approach which can be incorporated into other approaches such as adversarial training to further improve the resilience. The accuracy loss incurred in small perturbation scenarios $(\epsilon = 2 / 255, 4 / 255)$ can be mitigated by training with augmented data. + +# 5.4 Adaptive Attacks + +To thoroughly evaluate the effectiveness of a defense strategy, [10] proposed to design adaptive attacks that counter the defense mechanism with perfect knowledge of both DNN and defense in the white-box setting. To this end, we assess our defense with two adaptive attacks. + +Low Frequency Attack. As the total variation denoising is proposed to remove high-frequency components, the first adaptive attack that we considered is to only add perturbations in low frequency space. Following the methodology outlined in [21], we first convert the gradient to frequency space with discrete cosine transform and mask out the high-frequency components. The gradient is then transformed back and Eq. (5) is applied to create the adversarial samples. + +Regional Attack. Given that the defense exclusively denoises the background region, the second adaptive attack we considered is to force the adversary put greater effort in increasing bit rate in the object region. Similar to Eq. (11), a mask is multiplied to the loss function before backpropagation. In practice, we choose $1 - P_Z(z)$ as the soft mask where $P_Z(z)$ is the probability of $z$ ranging between 0 and 1. A larger $P_Z(z)$ indicates a smaller bit rate $-\log_2P_Z(z)$ corresponding to the background region. Therefore, $1 - P_Z(z)$ encourages the adversary to take more attention to object region. + +Table 4: Resilience to adaptive attacks with $\epsilon = 4 / 255$ + +
Adaptive AttacksBefore DenoisingAfter Denoising
Size[KB]Acc[%]Size[KB]Acc[%]
Frequency14.11±0.3871.389.56±1.0770.17
Regional14.05±0.4171.639.50±1.1070.22
+ +Resilience to Adaptive Attacks. Tab. 4 shows results of adaptive attacks with perturbation budget $\epsilon = 4 / 255$ . Our defense successfully reduces the average data size by $67.75\%$ and $67.62\%$ for low frequency attack and regional attack respectively while the accuracy only decreases $1.21\%$ and $1.41\%$ respectively. Thus, the proposed approach is resilient to adaptive attacks. + +# 6 Related Work + +Distributed Deep Neural Network. Distributed DNN is proposed to meet the challenge of deploying artificial intelligence to resource-constrained platforms such as mobile devices and Internet of Things (IoT) devices [41]. To minimize the end-to-end latency across devices, dimension reduction designs (i.e., bottlenecks) are introduced to compress the data size of intermediate representations that need to be sent [18, 54]. + +However, unlike autoencoders [32,36,49] that have symmetric designs on both encoder and decoder sides, distributed DNNs usually have asymmetric designs due to the limit of computation resources, and thus inject the bottleneck in early layers [18, 38-40, 54]. As a result, naively end-to-end trained distributed DNNs have noticeable performance loss [18, 54]. To preserve accuracy, [38] proposed to use knowledge distillation to train the distributed DNN separately while [39] proposed a multi-stage training approach for each part of the DNN. + +Along with bottleneck-based distributed DNNs, different coding-based approaches are also applied to reduce the data size of the latent representations. [2] proposed to apply codec-based compression such as JPEG to latent representations and [15] adopted spatio-temporal coding such as HEVC to compress the streaming latent features in video tasks. [42,55,59] use entropy coding to achieve a higher compression ratio of latent representations. + +Entropy Coding in Deep Data Compression. [5] first introduced a fully factorized prior to integrate entropy coding with variational autoencoder, showcasing superior quality enhancement over conventional codec-based methods like JPEG for image compression. Building upon this, [6] extended the conventional factorized prior model to a learnable hierarchical hyper prior, resulting in improved performance. [43] proposed a joint auto-regressive and hyper prior design while [44] proposed a channel-aware entropy coding scheme. In addition, [14] introduced the attention mechanism to prior models and [1] extended the technique to video compression. + +To minimize the communication overhead in distributed DNN scenarios, [55] for the first time adopted the entropy coding to further compress the latent representations after the bottleneck while [59] extended the approach to object detection. However, these approaches overlooked resource constraints, resulting in over-complex network designs for mobile devices. Conversely, [42] integrated entropy coding with a multi-stage training strategy [39], aiming to optimize a tripartite trade-off encompassing computation resources in mobile devices, communication overhead between mobile and edge devices, and end-to-end performance of neural networks. + +Resilience of DNN efficiency. While the reliability of DNNs have been extensively investigated, limited attention has been devoted to interference that undermine their efficiency. [22] explored this direction by demonstrating that imperceptible perturbations to input, leveraging intermediate representations of DNNs, could nullify the computation savings achieved by dynamic depth neural networks, which adapt their depth based on input complexity. [46] extended these attacks to target both dynamic depth and dynamic width neural networks. Moreover, in [45], the authors proposed simultaneous adjustments to the direction and magnitude of attacks, enhancing their effectiveness. [27] revealed that early exit dynamic DNNs could be tricked into late-stage inferences, significantly slowing down inference speed. + +The study of efficiency vulnerabilities in DNNs has also ventured into real-world deployment scenarios. [33] focused on attacking LiDAR-based detection systems, introducing latency in detections and exposing vulnerabilities in critical contexts such as self-driving cars. [13] demonstrated that neural image caption generation models could be manipulated to incur increased computation costs by inducing unnecessary decoder calls for token generation. + +In star contrast to existing research, our work investigates a novel threat emerging in distributed DNNs. Here, adversaries not only compromise the communication efficiency of entropy-coded distributed DNNs but also pose a threat to other users by saturating the transmission bandwidth. + +# 7 Conclusion + +This paper has investigated the resilience of entropy models in distributed DNNs against both intentional and unintentional interference. We conducted thorough evaluations using 3 different DNN architectures, 2 entropy model designs, and 4 rate-distortion trade-off factors with common corruption datasets and adversarial attacks. Our analysis disentangled compression features in both spatial and frequency domains, revealing vulnerabilities of the entropy model to specific types of perturbations. Building on these findings, we proposed a standalone defense strategy aimed at reducing data size with minimal task-oriented performance loss. Our future work will focus on designing more advanced defense approaches for distributed DNNs that are resilient in both compression and classification tasks. + +# Acknowledgements + +This work is funded in part by the Air Force Office of Scientific Research under contract number FA9550-23-1-0261, by the Office of Naval Research under contract number N00014-23-1-2221, as well as by National Science Foundation grants CNS-2134973 and CNS-2312875. Any opinions, findings and conclusions or recommendations expressed in this material are those of the authors and do not necessarily reflect the views of the U.S. Government. + +# References + +1. Agustsson, E., Minnen, D., Johnston, N., Balle, J., Hwang, S.J., Toderici, G.: Scale-space flow for end-to-end optimized video compression. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8503-8512 (2020) +2. Alvar, S.R., Bajic, I.V.: Pareto-optimal bit allocation for collaborative intelligence. IEEE Transactions on Image Processing 30, 3348-3361 (2021) +3. Andriushchenko, M., Croce, F., Flammarion, N., Hein, M.: Square attack: a query-efficient black-box adversarial attack via random search. In: European conference on computer vision. pp. 484-501. Springer (2020) +4. Arvinte, M., Cornelius, C., Martin, J., Himayat, N.: Investigating the adversarial robustness of density estimation using the probability flow ode. arXiv preprint arXiv:2310.07084 (2023) +5. Balle, J., Laparra, V., Simoncelli, E.P.: End-to-end optimized image compression. In: 5th International Conference on Learning Representations, ICLR 2017 (2017) +6. Balle, J., Minnen, D., Singh, S., Hwang, S.J., Johnston, N.: Variational image compression with a scale hyperprior. In: International Conference on Learning Representations (2018) +7. Barbu, A., Mayo, D., Alverio, J., Luo, W., Wang, C., Gutfreund, D., Tenenbaum, J., Katz, B.: Objectnet: A large-scale bias-controlled dataset for pushing the limits of object recognition models. Advances in neural information processing systems 32 (2019) +8. Bau, D., Zhou, B., Khosla, A., Oliva, A., Torralba, A.: Network dissection: Quantifying interpretability of deep visual representations. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 6541-6549 (2017) +9. Boigné, E., Parkinson, D.Y., Ihme, M.: Towards data-informed motion artifact reduction in quantitative ct using piecewise linear interpolation. IEEE Transactions on Computational Imaging 8, 917-932 (2022) +10. Carlini, N., Wagner, D.: Adversarial examples are not easily detected: Bypassing ten detection methods. In: Proceedings of the 10th ACM workshop on artificial intelligence and security. pp. 3-14 (2017) +1. Carlini, N., Wagner, D.: Towards evaluating the robustness of neural networks. In: 2017 ieee symposium on security and privacy (sp). pp. 39-57. IEEE (2017) +2. Chen, J., Jordan, M.I., Wainwright, M.J.: Hopskipjumpattack: A query-efficient decision-based attack. In: 2020 ieee symposium on security and privacy (sp). pp. 1277-1294. IEEE (2020) +3. Chen, S., Song, Z., Haque, M., Liu, C., Yang, W.: Nicgslowdown: Evaluating the efficiency robustness of neural image caption generation models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15365-15374 (2022) + +14. Cheng, Z., Sun, H., Takeuchi, M., Katto, J.: Learned image compression with discretized gaussian mixture likelihoods and attention modules. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 7939-7948 (2020) +15. Choi, H., Bajic, I.V.: Deep feature compression for collaborative object detection. In: 2018 25th IEEE International Conference on Image Processing (ICIP). pp. 3743-3747. IEEE (2018) +16. Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: Imagenet: A large-scale hierarchical image database. In: 2009 IEEE conference on computer vision and pattern recognition. pp. 248-255. IEEE (2009) +17. Duda, J.: Asymmetric numeral systems: entropy coding combining speed of huffman coding with compression rate of arithmetic coding. arXiv preprint arXiv:1311.2540 (2013) +18. Eshratifar, A.E., Esmaili, A., Pedram, M.: Bottlenet: A deep learning architecture for intelligent mobile cloud computing services. In: 2019 IEEE/ACM International Symposium on Low Power Electronics and Design (ISLPED). pp. 1-6. IEEE (2019) +19. Geirhos, R., Rubisch, P., Michaelis, C., Bethge, M., Wichmann, F.A., Brendel, W.: Imagenet-trained cnns are biased towards texture; increasing shape bias improves accuracy and robustness. In: International Conference on Learning Representations (2018) +20. Goodfellow, I.J., Shlens, J., Szegedy, C.: Explaining and harnessing adversarial examples. arXiv preprint arXiv:1412.6572 (2014) +21. Guo, C., Frank, J.S., Weinberger, K.Q.: Low frequency adversarial perturbation. arXiv preprint arXiv:1809.08758 (2018) +22. Haque, M., Chauhan, A., Liu, C., Yang, W.: Ilfo: Adversarial attack on adaptive neural networks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14264-14273 (2020) +23. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 770-778 (2016) +24. Hendrycks, D., Basart, S., Mu, N., Kadavath, S., Wang, F., Dorundo, E., Desai, R., Zhu, T., Parajuli, S., Guo, M., et al.: The many faces of robustness: A critical analysis of out-of-distribution generalization. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 8340-8349 (2021) +25. Hendrycks, D., Dietterich, T.: Benchmarking neural network robustness to common corruptions and perturbations. In: International Conference on Learning Representations (2018) +26. Hinton, G., Vinyals, O., Dean, J.: Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531 (2015) +27. Hong, S., Kaya, Y., Modoranu, I.V., Dumitraş, T.: A panda? no, it's a sloth: Slowdown attacks on adaptive multi-exit neural network inference. arXiv preprint arXiv:2010.02432 (2020) +28. Howard, A., Sandler, M., Chu, G., Chen, L.C., Chen, B., Tan, M., Wang, W., Zhu, Y., Pang, R., Vasudevan, V., et al.: Searching for mobilenetv3. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 1314-1324 (2019) +29. Howard, A.G., Zhu, M., Chen, B., Kalenichenko, D., Wang, W., Weyand, T., Andreetto, M., Adam, H.: Mobilenets: Efficient convolutional neural networks for mobile vision applications. arXiv preprint arXiv:1704.04861 (2017) +30. Ilyas, A., Engstrom, L., Athalye, A., Lin, J.: Black-box adversarial attacks with limited queries and information. In: International conference on machine learning. pp. 2137-2146. PMLR (2018) + +31. Jacob, B., Kligys, S., Chen, B., Zhu, M., Tang, M., Howard, A., Adam, H., Kalenichenko, D.: Quantization and training of neural networks for efficient integer-arithmetic-only inference. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 2704-2713 (2018) +32. Kingma, D.P., Welling, M.: Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114 (2013) +33. Liu, H., Wu, Y., Yu, Z., Vorobeychik, Y., Zhang, N.: Slowlidar: Increasing the latency of lidar-based detection using adversarial examples. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5146-5155 (2023) +34. Liu, P., Qi, B., Banerjee, S.: Edgeeye: An edge service framework for real-time intelligent video analytics. In: Proceedings of the 1st international workshop on edge systems, analytics and networking. pp. 1-6 (2018) +35. Madry, A., Makelov, A., Schmidt, L., Tsipras, D., Vladu, A.: Towards deep learning models resistant to adversarial attacks. In: International Conference on Learning Representations (2018) +36. Masci, J., Meier, U., Ciresan, D., Schmidhuber, J.: Stacked convolutional autoencoders for hierarchical feature extraction. In: Artificial Neural Networks and Machine Learning-ICANN 2011: 21st International Conference on Artificial Neural Networks, Espoo, Finland, June 14-17, 2011, Proceedings, Part I 21. pp. 52-59. Springer (2011) +37. Matsubara, Y., Yang, R., Levorato, M., Mandt, S.: Sc2 benchmark: Supervised compression for split computing. Transactions on machine learning research (2023) +38. Matsubara, Y., Callegaro, D., Baidya, S., Levorato, M., Singh, S.: Head network distillation: Splitting distilled deep neural networks for resource-constrained edge computing systems. IEEE Access 8, 212177-212193 (2020) +39. Matsubara, Y., Callegaro, D., Singh, S., Levorato, M., Restuccia, F.: Bottlefit: Learning compressed representations in deep neural networks for effective and efficient split computing. In: 2022 IEEE 23rd International Symposium on a World of Wireless, Mobile and Multimedia Networks (WoWMoM). pp. 337-346. IEEE (2022) +40. Matsubara, Y., Levorato, M.: Neural compression and filtering for edge-assisted real-time object detection in challenged networks. In: 2020 25th International Conference on Pattern Recognition (ICPR). pp. 2272-2279. IEEE (2021) +41. Matsubara, Y., Levorato, M., Restuccia, F.: Split computing and early exiting for deep learning applications: Survey and research challenges. ACM Computing Surveys 55(5), 1-30 (2022) +42. Matsubara, Y., Yang, R., Levorato, M., Mandt, S.: Supervised compression for resource-constrained edge computing systems. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 2685-2695 (2022) +43. Minnen, D., Balle, J., Toderici, G.D.: Joint autoregressive and hierarchical priors for learned image compression. Advances in neural information processing systems 31 (2018) +44. Minnen, D., Singh, S.: Channel-wise autoregressive entropy models for learned image compression. In: 2020 IEEE International Conference on Image Processing (ICIP). pp. 3339-3343. IEEE (2020) +45. Pan, J., Foo, L.G., Zheng, Q., Fan, Z., Rahmani, H., Ke, Q., Liu, J.: Gradmdm: Adversarial attack on dynamic networks. IEEE Transactions on Pattern Analysis and Machine Intelligence (2023) + +46. Pan, J., Zheng, Q., Fan, Z., Rahmani, H., Ke, Q., Liu, J.: Gradauto: Energy-oriented attack on dynamic neural networks. In: European Conference on Computer Vision. pp. 637-653. Springer (2022) +47. Radosavovic, I., Kosaraju, R.P., Girshick, R., He, K., Dóllár, P.: Designing network design spaces. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10428-10436 (2020) +48. Ran, X., Chen, H., Zhu, X., Liu, Z., Chen, J.: Deepdecision: A mobile deep learning framework for edge video analytics. In: IEEE INFOCOM 2018-IEEE conference on computer communications. pp. 1421-1429. IEEE (2018) +49. Rifai, S., Vincent, P., Muller, X., Glorot, X., Bengio, Y.: Contractive auto-encoders: Explicit invariance during feature extraction. In: Proceedings of the 28th international conference on international conference on machine learning. pp. 833-840 (2011) +50. Rissanen, J., Langdon, G.: Universal modeling and coding. IEEE Transactions on Information Theory 27(1), 12-23 (1981) +51. Rudin, L.I., Osher, S., Fatemi, E.: Nonlinear total variation based noise removal algorithms. Physica D: nonlinear phenomena 60(1-4), 259-268 (1992) +52. Sandler, M., Howard, A., Zhu, M., Zhmoginov, A., Chen, L.C.: Mobilenetv2: Inverted residuals and linear bottlenecks. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 4510-4520 (2018) +53. Shannon, C.E.: A mathematical theory of communication. The Bell system technical journal 27(3), 379-423 (1948) +54. Shao, J., Zhang, J.: Bottlenet++: An end-to-end approach for feature compression in device-edge co-inference systems. In: 2020 IEEE International Conference on Communications Workshops (ICC Workshops). pp. 1-6. IEEE (2020) +55. Singh, S., Abu-El-Haija, S., Johnston, N., Balle, J., Shrivastava, A., Toderici, G.: End-to-end learning of compressible features. In: 2020 IEEE International Conference on Image Processing (ICIP). pp. 3349-3353. IEEE (2020) +56. Szegedy, C., Zaremba, W., Sutskever, I., Bruna, J., Erhan, D., Goodfellow, I., Fergus, R.: Intriguing properties of neural networks. In: 2nd International Conference on Learning Representations, ICLR 2014 (2014) +57. Tan, M., Chen, B., Pang, R., Vasudevan, V., Sandler, M., Howard, A., Le, Q.V.: Mnasnet: Platform-aware neural architecture search for mobile. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 2820-2828 (2019) +58. Wang, H., Ge, S., Lipton, Z., Xing, E.P.: Learning robust global representations by penalizing local predictive power. Advances in Neural Information Processing Systems 32 (2019) +59. Yuan, Z., Rawlekar, S., Garg, S., Erkip, E., Wang, Y.: Feature compression for rate constrained object detection on the edge. In: 2022 IEEE 5th International Conference on Multimedia Information Processing and Retrieval (MIPR). pp. 1-6. IEEE (2022) +60. Zhang, W., Zhang, Z., Zeadally, S., Chao, H.C., Leung, V.C.: Masm: A multiple-algorithm service model for energy-delay optimization in edge artificial intelligence. IEEE Transactions on Industrial Informatics 15(7), 4216-4224 (2019) \ No newline at end of file diff --git a/2024/Resilience of Entropy Model in Distributed Neural Networks/images.zip b/2024/Resilience of Entropy Model in Distributed Neural Networks/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..7f77411187a814d0cd062830467cfa6a74e37830 --- /dev/null +++ b/2024/Resilience of Entropy Model in Distributed Neural Networks/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ff7cd16a0735aa629783ed532a51b6380272c1a4548bd33bf774000a0d35d802 +size 447944 diff --git a/2024/Resilience of Entropy Model in Distributed Neural Networks/layout.json b/2024/Resilience of Entropy Model in Distributed Neural Networks/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..c43699b62923141677c88b797970a43ed5c527ed --- /dev/null +++ b/2024/Resilience of Entropy Model in Distributed Neural Networks/layout.json @@ -0,0 +1,11610 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 198, + 111, + 416, + 146 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 198, + 111, + 416, + 146 + ], + "spans": [ + { + "bbox": [ + 198, + 111, + 416, + 146 + ], + "type": "text", + "content": "Resilience of Entropy Model in Distributed Neural Networks" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 209, + 168, + 404, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 209, + 168, + 404, + 193 + ], + "spans": [ + { + "bbox": [ + 209, + 168, + 404, + 193 + ], + "type": "text", + "content": "Milin Zhang, Mohammad Abdi, Shahriar Rifat, and Francesco Restuccia" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 173, + 202, + 440, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 173, + 202, + 440, + 236 + ], + "spans": [ + { + "bbox": [ + 173, + 202, + 440, + 236 + ], + "type": "text", + "content": "Institute for the Wireless Internet of Things Northeastern University, Boston MA 02115, USA {zhang.mil, abdi.mo, rifat.s, f/restuccia}@northeastern.com" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 160, + 262, + 455, + 504 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 262, + 455, + 504 + ], + "spans": [ + { + "bbox": [ + 160, + 262, + 455, + 504 + ], + "type": "text", + "content": "Abstract. Distributed deep neural networks (DNNs) have emerged as a key technique to reduce communication overhead without sacrificing performance in edge computing systems. Recently, entropy coding has been introduced to further reduce the communication overhead. The key idea is to train the distributed DNN jointly with an entropy model, which is used as side information during inference time to adaptively encode latent representations into bit streams with variable length. To the best of our knowledge, the resilience of entropy models is yet to be investigated. As such, in this paper we formulate and investigate the resilience of entropy models to intentional interference (e.g., adversarial attacks) and unintentional interference (e.g., weather changes and motion blur). Through an extensive experimental campaign with 3 different DNN architectures, 2 entropy models and 4 rate-distortion trade-off factors, we demonstrate that the entropy attacks can increase the communication overhead by up to " + }, + { + "bbox": [ + 160, + 262, + 455, + 504 + ], + "type": "inline_equation", + "content": "95\\%" + }, + { + "bbox": [ + 160, + 262, + 455, + 504 + ], + "type": "text", + "content": ". By separating compression features in frequency and spatial domain, we propose a new defense mechanism that can reduce the transmission overhead of the attacked input by about " + }, + { + "bbox": [ + 160, + 262, + 455, + 504 + ], + "type": "inline_equation", + "content": "9\\%" + }, + { + "bbox": [ + 160, + 262, + 455, + 504 + ], + "type": "text", + "content": " compared to unperturbed data, with only about " + }, + { + "bbox": [ + 160, + 262, + 455, + 504 + ], + "type": "inline_equation", + "content": "2\\%" + }, + { + "bbox": [ + 160, + 262, + 455, + 504 + ], + "type": "text", + "content": " accuracy loss. Importantly, the proposed defense mechanism is a standalone approach which can be applied in conjunction with approaches such as adversarial training to further improve robustness. Code is available at https://github.com/Restuccia-Group/EntropyR." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 160, + 514, + 425, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 514, + 425, + 525 + ], + "spans": [ + { + "bbox": [ + 160, + 514, + 425, + 525 + ], + "type": "text", + "content": "Keywords: Trustworthy Machine Learning " + }, + { + "bbox": [ + 160, + 514, + 425, + 525 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 514, + 425, + 525 + ], + "type": "text", + "content": " Data Compression" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 545, + 230, + 558 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 545, + 230, + 558 + ], + "spans": [ + { + "bbox": [ + 132, + 545, + 230, + 558 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "type": "text", + "content": "Distributed deep neural network (DNN) were recently introduced to divide the computation of DNNs across various devices based on their available computation and communication resources. They have been shown to be extremely effective to implement deep learning applications on resource-constrained mobile devices [18,39,54]. As depicted in Fig. 1, the common strategy is to divide a large DNN into a small head network deployed on the mobile device to extract and compress features, and a tail network on the server to perform the task inference. Compared to conventional lightweight DNNs specifically designed for" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 149, + 114, + 239, + 163 + ], + "blocks": [ + { + "bbox": [ + 149, + 114, + 239, + 163 + ], + "lines": [ + { + "bbox": [ + 149, + 114, + 239, + 163 + ], + "spans": [ + { + "bbox": [ + 149, + 114, + 239, + 163 + ], + "type": "image", + "image_path": "cfc42dff6d0765a2b15c7111c775dae97d502cada9ec96fd74a01aa141e06523.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 153, + 214, + 243, + 225 + ], + "lines": [ + { + "bbox": [ + 153, + 214, + 243, + 225 + ], + "spans": [ + { + "bbox": [ + 153, + 214, + 243, + 225 + ], + "type": "inline_equation", + "content": "\\times" + }, + { + "bbox": [ + 153, + 214, + 243, + 225 + ], + "type": "text", + "content": " Computation Resource" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_footnote" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 200, + 172, + 220, + 196 + ], + "blocks": [ + { + "bbox": [ + 200, + 172, + 220, + 196 + ], + "lines": [ + { + "bbox": [ + 200, + 172, + 220, + 196 + ], + "spans": [ + { + "bbox": [ + 200, + 172, + 220, + 196 + ], + "type": "image", + "image_path": "2d9c27eab88fe9e37510eec34e286a49a5dc37a87f3efee3b699947b16e5d4e1.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 200, + 197, + 221, + 213 + ], + "lines": [ + { + "bbox": [ + 200, + 197, + 221, + 213 + ], + "spans": [ + { + "bbox": [ + 200, + 197, + 221, + 213 + ], + "type": "text", + "content": "Edge Server" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 269, + 114, + 345, + 213 + ], + "blocks": [ + { + "bbox": [ + 269, + 114, + 345, + 213 + ], + "lines": [ + { + "bbox": [ + 269, + 114, + 345, + 213 + ], + "spans": [ + { + "bbox": [ + 269, + 114, + 345, + 213 + ], + "type": "image", + "image_path": "20c631db94e73d6a0e29ec6ccad33b513f0d3e30c1e58a15939dcfabecbc105e.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 252, + 215, + 354, + 225 + ], + "lines": [ + { + "bbox": [ + 252, + 215, + 354, + 225 + ], + "spans": [ + { + "bbox": [ + 252, + 215, + 354, + 225 + ], + "type": "text", + "content": "X Communication Overhead" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 130, + 241, + 482, + 308 + ], + "lines": [ + { + "bbox": [ + 130, + 241, + 482, + 308 + ], + "spans": [ + { + "bbox": [ + 130, + 241, + 482, + 308 + ], + "type": "text", + "content": "Fig. 1: The key concept of distributed DNN. Left: Lightweight DNNs suffer from performance loss due to the limited computational resources; Middle: Edge computing often results in intolerable latency due to communication overhead; Right: Distributed DNNs deploy a small head model on the resource-limited device to achieve task-oriented compression and a large tail model on the server to decode compressed features and execute the rest of model." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 369, + 114, + 460, + 210 + ], + "blocks": [ + { + "bbox": [ + 369, + 114, + 460, + 210 + ], + "lines": [ + { + "bbox": [ + 369, + 114, + 460, + 210 + ], + "spans": [ + { + "bbox": [ + 369, + 114, + 460, + 210 + ], + "type": "image", + "image_path": "f0d93f33eb70a83800dad5a479c70c13af5ffa5fc83eb3e3759457e95b0cf61a.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 364, + 214, + 463, + 232 + ], + "lines": [ + { + "bbox": [ + 364, + 214, + 463, + 232 + ], + "spans": [ + { + "bbox": [ + 364, + 214, + 463, + 232 + ], + "type": "inline_equation", + "content": "\\checkmark" + }, + { + "bbox": [ + 364, + 214, + 463, + 232 + ], + "type": "text", + "content": " Performance & Latency Trade-off" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_footnote" + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 318, + 482, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 318, + 482, + 389 + ], + "spans": [ + { + "bbox": [ + 130, + 318, + 482, + 389 + ], + "type": "text", + "content": "mobile devices [28, 29, 31, 52, 57], distributed DNNs can leverage the computation power of the edge/cloud and hence attain better performance. In addition, distributed DNNs leverage compression techniques to reduce the data size of intermediate representations. By only transmitting compact representations, distributed DNNs can reduce the transmission overhead significantly compared to traditional edge computing [34, 48, 60]." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 390, + 483, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 390, + 483, + 628 + ], + "spans": [ + { + "bbox": [ + 130, + 390, + 483, + 628 + ], + "type": "text", + "content": "Interestingly, recent work has proposed to further minimize the data size of transmitted latent representations with entropy coding [42]. The key idea is to train the distributed DNN jointly with an auxiliary entropy model, which estimates the distribution of latent representations. The output of the entropy model is then used as side information during inference time to facilitate the adaptive encoding of latent representations into bit streams with variable lengths. The proposed approach achieves better compression rate compared to quantization-based [38] and codec-based [2] approaches by minimizing the entropy of latent representations. Despite its excellent compression performance, the resilience of entropy coding to intentional interference (e.g., specifically crafted adversarial attacks) and unintentional interference (e.g., sudden events such as weather changes and motion blur) remains unexplored. We remark that in entropy-coding-based distributed DNNs, the bit rate of encoded representations relies completely on its estimated entropy. However, it is well known that DNNs trained on standard datasets are vulnerable to distribution shifts [24,25] and adversarial actions [20, 56]. As the entropy model is learned without considering any corruptions and adversarial perturbations, a small change in the input space might lead to a large estimate of entropy, and hence result in a bit rate that can exceed the transmission bandwidth. As depicted in Tab. 3 in our experiments, the transmission overhead can be increased by about " + }, + { + "bbox": [ + 130, + 390, + 483, + 628 + ], + "type": "inline_equation", + "content": "2\\mathrm{x}" + }, + { + "bbox": [ + 130, + 390, + 483, + 628 + ], + "type": "text", + "content": " in the worst case." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 630, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 630, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 630, + 482, + 665 + ], + "type": "text", + "content": "In this work, we thoroughly assess for the first time the resilience of entropy models to intentional and unintentional interference as applied to distributed DNNs. Intriguingly, our findings unveil that the auxiliary entropy model learns" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 231, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 231, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 231, + 102 + ], + "type": "text", + "content": "M. Zhang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 175 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 175 + ], + "type": "text", + "content": "a different set of input-related features than what learned by the backbone DNN. Ultimately, this enables us to design effective defense strategy to maintain the bit rate with an unnoticeable loss in the task performance. Our approach successfully reduces the communication overhead of perturbed data by about " + }, + { + "bbox": [ + 130, + 116, + 479, + 175 + ], + "type": "inline_equation", + "content": "9\\%" + }, + { + "bbox": [ + 130, + 116, + 479, + 175 + ], + "type": "text", + "content": " compared to unperturbed inputs, with only about " + }, + { + "bbox": [ + 130, + 116, + 479, + 175 + ], + "type": "inline_equation", + "content": "2\\%" + }, + { + "bbox": [ + 130, + 116, + 479, + 175 + ], + "type": "text", + "content": " accuracy loss." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 192, + 302, + 205 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 192, + 302, + 205 + ], + "spans": [ + { + "bbox": [ + 131, + 192, + 302, + 205 + ], + "type": "text", + "content": "Summary of Novel Contributions" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 210, + 482, + 456 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 130, + 210, + 482, + 294 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 210, + 482, + 294 + ], + "spans": [ + { + "bbox": [ + 130, + 210, + 482, + 294 + ], + "type": "text", + "content": "(1) We investigate for the first time the effect of intentional and unintentional interference to entropy coding in distributed DNNs. We show that interference increases the end-to-end latency of distributed DNNs and poses a threat to other users by saturating the transmission bandwidth. Through comprehensive experiments involving 3 distinct DNN architectures [23, 47], 2 entropy models [5,43], and 4 rate-distortion trade-off factors, we illustrate that attacks to entropy coding may increase the transmitted data size by up to about " + }, + { + "bbox": [ + 130, + 210, + 482, + 294 + ], + "type": "inline_equation", + "content": "95\\%" + }, + { + "bbox": [ + 130, + 210, + 482, + 294 + ], + "type": "text", + "content": ";" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 297, + 482, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 297, + 482, + 380 + ], + "spans": [ + { + "bbox": [ + 130, + 297, + 482, + 380 + ], + "type": "text", + "content": "(2) We design two visualization approaches to interpret the compression learned by entropy models. By disentangling features relevant to compression and classification across both frequency and spatial domains, we reveal that the entropy model learns specific features that are distinct from those beneficial for the end task. This finding enables us to devise an effective defense strategy that safeguards the vulnerable compression features while minimally impacting task performance;" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 383, + 482, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 383, + 482, + 456 + ], + "spans": [ + { + "bbox": [ + 130, + 383, + 482, + 456 + ], + "type": "text", + "content": "(3) We propose a defense approach to attacks targeting entropy coding based on our findings discussed above. The proposed method effectively reduces the transmission overhead of attacked inputs by about " + }, + { + "bbox": [ + 130, + 383, + 482, + 456 + ], + "type": "inline_equation", + "content": "9\\%" + }, + { + "bbox": [ + 130, + 383, + 482, + 456 + ], + "type": "text", + "content": " in comparison to the unperturbed case, with only about " + }, + { + "bbox": [ + 130, + 383, + 482, + 456 + ], + "type": "inline_equation", + "content": "2\\%" + }, + { + "bbox": [ + 130, + 383, + 482, + 456 + ], + "type": "text", + "content": " decrease in accuracy. We remark that our approach is general in nature and can be combined with approaches such as adversarial training to further improve the resilience of entropy models." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 130, + 457, + 482, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 457, + 482, + 540 + ], + "spans": [ + { + "bbox": [ + 130, + 457, + 482, + 540 + ], + "type": "text", + "content": "The paper is organized as follows. Sec. 2 provides background of entropy coding in distributed DNNs. In Sec. 3, we formally describe the threat model for entropy coding in distributed DNNs. Sec. 4 presents benchmarking results on both intentional and unintentional interference. Sec. 5 details the proposed defense approach and comprehensively evaluates it against adaptive attacks. Related work is summarized in Sec. 6. Finally, we draw conclusions and discuss future work in Sec. 7." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 557, + 378, + 572 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 557, + 378, + 572 + ], + "spans": [ + { + "bbox": [ + 131, + 557, + 378, + 572 + ], + "type": "text", + "content": "2 Entropy Coding in Distributed DNNs" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 581, + 482, + 642 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 581, + 482, + 642 + ], + "spans": [ + { + "bbox": [ + 130, + 581, + 482, + 642 + ], + "type": "text", + "content": "Entropy coding techniques such as arithmetic coding [50] and asymmetric numeral systems [17] can encode a message with the optimal coding rate by leveraging the probabilistic information of the message. Based on the source coding theorem [53], the entropy of a message " + }, + { + "bbox": [ + 130, + 581, + 482, + 642 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 130, + 581, + 482, + 642 + ], + "type": "text", + "content": " with distribution " + }, + { + "bbox": [ + 130, + 581, + 482, + 642 + ], + "type": "inline_equation", + "content": "P_Z(z)" + }, + { + "bbox": [ + 130, + 581, + 482, + 642 + ], + "type": "text", + "content": " defines the lower bound of expected coding rate " + }, + { + "bbox": [ + 130, + 581, + 482, + 642 + ], + "type": "inline_equation", + "content": "R_{c}(z)" + }, + { + "bbox": [ + 130, + 581, + 482, + 642 + ], + "type": "text", + "content": " without any loss of information, i.e.," + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 221, + 654, + 480, + 666 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 221, + 654, + 480, + 666 + ], + "spans": [ + { + "bbox": [ + 221, + 654, + 480, + 666 + ], + "type": "interline_equation", + "content": "\\mathbb {E} \\left\\{R _ {c} (z) \\right\\} \\geq H _ {P _ {Z}} (z) = \\mathbb {E} \\left\\{- \\log_ {2} P _ {Z} (z) \\right\\} \\tag {1}", + "image_path": "6dc1732ac6e0df7a48f6d106efe50a489b594bdecbe2482420a436859ef0ec3d.jpg" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 200, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 200, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 200, + 91, + 448, + 102 + ], + "type": "text", + "content": "Resilience of Entropy Model in Distributed Neural Networks" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 145, + 118, + 289, + 185 + ], + "blocks": [ + { + "bbox": [ + 145, + 118, + 289, + 185 + ], + "lines": [ + { + "bbox": [ + 145, + 118, + 289, + 185 + ], + "spans": [ + { + "bbox": [ + 145, + 118, + 289, + 185 + ], + "type": "image", + "image_path": "46ef122df4dc911a3fd3a7389796a253b136687169eb3c9e6e564c316d04dc74.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 198, + 482, + 255 + ], + "lines": [ + { + "bbox": [ + 130, + 198, + 482, + 255 + ], + "spans": [ + { + "bbox": [ + 130, + 198, + 482, + 255 + ], + "type": "text", + "content": "Fig. 2: A general framework for entropy coding in neural data compression. It consists of an encoder-decoder structure to extract a compact latent representation " + }, + { + "bbox": [ + 130, + 198, + 482, + 255 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 130, + 198, + 482, + 255 + ], + "type": "text", + "content": " and a prior model to estimate the distribution " + }, + { + "bbox": [ + 130, + 198, + 482, + 255 + ], + "type": "inline_equation", + "content": "P_Z(z)" + }, + { + "bbox": [ + 130, + 198, + 482, + 255 + ], + "type": "text", + "content": ". Left: During training, the prior model is jointly optimized with the encoder-decoder backbone. Right: During inference, the learned prior " + }, + { + "bbox": [ + 130, + 198, + 482, + 255 + ], + "type": "inline_equation", + "content": "P_Z(z)" + }, + { + "bbox": [ + 130, + 198, + 482, + 255 + ], + "type": "text", + "content": " is used as side information for arithmetic encoding/decoding." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 315, + 118, + 470, + 185 + ], + "blocks": [ + { + "bbox": [ + 315, + 118, + 470, + 185 + ], + "lines": [ + { + "bbox": [ + 315, + 118, + 470, + 185 + ], + "spans": [ + { + "bbox": [ + 315, + 118, + 470, + 185 + ], + "type": "image", + "image_path": "3de526a1910bfa5aee1f6e50f103cbb4ed25db29ead2cccb45d90042a9d0c6eb.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 275, + 482, + 395 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 275, + 482, + 395 + ], + "spans": [ + { + "bbox": [ + 130, + 275, + 482, + 395 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 275, + 482, + 395 + ], + "type": "inline_equation", + "content": "H_{P_Z}(z)" + }, + { + "bbox": [ + 130, + 275, + 482, + 395 + ], + "type": "text", + "content": " is the entropy of " + }, + { + "bbox": [ + 130, + 275, + 482, + 395 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 130, + 275, + 482, + 395 + ], + "type": "text", + "content": ". In recent data compression literature [1,5,6, 14,43,44], the entropy coding is integrated with lossy DNN-based compression to attain an effective compression rate with less information loss compared to traditional approaches. As depicted in Fig. 2, the common strategy is to jointly train an encoder-decoder backbone for extracting compact latent representations " + }, + { + "bbox": [ + 130, + 275, + 482, + 395 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 130, + 275, + 482, + 395 + ], + "type": "text", + "content": " with an auxiliary prior model " + }, + { + "bbox": [ + 130, + 275, + 482, + 395 + ], + "type": "inline_equation", + "content": "P_Z(z)" + }, + { + "bbox": [ + 130, + 275, + 482, + 395 + ], + "type": "text", + "content": ". During training, a random noise source is introduced to simulate the effect of quantization on " + }, + { + "bbox": [ + 130, + 275, + 482, + 395 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 130, + 275, + 482, + 395 + ], + "type": "text", + "content": ", allowing the backpropagation of gradients. In the inference phase, " + }, + { + "bbox": [ + 130, + 275, + 482, + 395 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 130, + 275, + 482, + 395 + ], + "type": "text", + "content": " is initially quantized to integers for entropy coding. Next, the output of the prior model is used as side information to encode the quantized " + }, + { + "bbox": [ + 130, + 275, + 482, + 395 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 130, + 275, + 482, + 395 + ], + "type": "text", + "content": " into bit streams adaptively with variable length." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 396, + 482, + 431 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 396, + 482, + 431 + ], + "spans": [ + { + "bbox": [ + 130, + 396, + 482, + 431 + ], + "type": "text", + "content": "DNN-based data compression can achieve the near-optimal compression rate by minimizing the entropy of latent representations during training. Specifically, the objective is to minimize the rate-distortion function" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 219, + 438, + 481, + 467 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 219, + 438, + 481, + 467 + ], + "spans": [ + { + "bbox": [ + 219, + 438, + 481, + 467 + ], + "type": "interline_equation", + "content": "\\mathcal {L} (x, (\\hat {x}, z)) = \\underbrace {\\| x - \\hat {x} \\| _ {2} ^ {2}} _ {\\text {d i s t o r t i o n}} - \\beta \\cdot \\underbrace {\\log_ {2} P _ {Z} (z)} _ {\\text {r a t e}}, \\tag {2}", + "image_path": "064db037eec7c5b11a279917d2cec1d9851fa96efc54a75510333de2ac332a1d.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 473, + 482, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 473, + 482, + 509 + ], + "spans": [ + { + "bbox": [ + 130, + 473, + 482, + 509 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 473, + 482, + 509 + ], + "type": "inline_equation", + "content": "\\hat{x}" + }, + { + "bbox": [ + 130, + 473, + 482, + 509 + ], + "type": "text", + "content": " is the reconstructed data, " + }, + { + "bbox": [ + 130, + 473, + 482, + 509 + ], + "type": "inline_equation", + "content": "||x - \\hat{x}||_2^2" + }, + { + "bbox": [ + 130, + 473, + 482, + 509 + ], + "type": "text", + "content": " is the loss function of the encoder-decoder backbone and " + }, + { + "bbox": [ + 130, + 473, + 482, + 509 + ], + "type": "inline_equation", + "content": "-\\log_2P_Z(z)" + }, + { + "bbox": [ + 130, + 473, + 482, + 509 + ], + "type": "text", + "content": " is the loss function of the prior model. Constant " + }, + { + "bbox": [ + 130, + 473, + 482, + 509 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 130, + 473, + 482, + 509 + ], + "type": "text", + "content": " denotes the trade-off between the rate and distortion." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 510, + 483, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 483, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 483, + 666 + ], + "type": "text", + "content": "Recent work [42] has incorporated entropy coding into distributed DNNs to increase communication efficiency between mobile and edge devices. The proposed method, referred to as the entropic student, aims at achieving a balance between the coding rate and task-specific performance metrics (e.g., cross entropy in classification tasks). One distinction between the entropic student and other entropy-coding-based feature compression work [55, 59] lies in the significantly smaller encoder size of the former due to constraints on computational resources in mobile devices. As a result, the entropic student splits the DNN at an early layer and use knowledge distillation [26] for training the head to preserve accuracy. However, existing work does not address the issue of resilience of entropy models, which is the key goal of this paper. A relevant work [4] investigated the robustness of density estimation aiming to maximize " + }, + { + "bbox": [ + 130, + 510, + 483, + 666 + ], + "type": "inline_equation", + "content": "P_Z(z)" + }, + { + "bbox": [ + 130, + 510, + 483, + 666 + ], + "type": "text", + "content": " while we are interested in perturbations that minimize " + }, + { + "bbox": [ + 130, + 510, + 483, + 666 + ], + "type": "inline_equation", + "content": "P_Z(z)" + }, + { + "bbox": [ + 130, + 510, + 483, + 666 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 231, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 231, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 231, + 102 + ], + "type": "text", + "content": "M. Zhang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 401, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 401, + 129 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 401, + 129 + ], + "type": "text", + "content": "3 Modeling Interference to Entropy Coding" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 139, + 482, + 199 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 139, + 482, + 199 + ], + "spans": [ + { + "bbox": [ + 130, + 139, + 482, + 199 + ], + "type": "text", + "content": "While the resilience of DNNs to distribution shifts [7,19,24,25,58] and adversarial actions [3,11,12,30,35,56] has been extensively investigated, we investigate types of interference that can alter the length of the encoded bit stream, which ultimately leads to increased bandwidth utilization. In this section, we formulate the resilience of compression in entropy-coding-driven distributed DNNs." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 215, + 482, + 264 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 215, + 482, + 264 + ], + "spans": [ + { + "bbox": [ + 130, + 215, + 482, + 264 + ], + "type": "text", + "content": "Threat Model for Accuracy. Let " + }, + { + "bbox": [ + 130, + 215, + 482, + 264 + ], + "type": "inline_equation", + "content": "D = \\{x, y\\}" + }, + { + "bbox": [ + 130, + 215, + 482, + 264 + ], + "type": "text", + "content": " denotes a dataset for classification problem where " + }, + { + "bbox": [ + 130, + 215, + 482, + 264 + ], + "type": "inline_equation", + "content": "x \\in \\mathcal{X}" + }, + { + "bbox": [ + 130, + 215, + 482, + 264 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 215, + 482, + 264 + ], + "type": "inline_equation", + "content": "y \\in \\mathcal{Y}" + }, + { + "bbox": [ + 130, + 215, + 482, + 264 + ], + "type": "text", + "content": " are specific input and corresponding label samples respectively, while " + }, + { + "bbox": [ + 130, + 215, + 482, + 264 + ], + "type": "inline_equation", + "content": "\\mathcal{X}, \\mathcal{Y}" + }, + { + "bbox": [ + 130, + 215, + 482, + 264 + ], + "type": "text", + "content": " are input and output space, respectively. There exists a prior yet unknown distribution " + }, + { + "bbox": [ + 130, + 215, + 482, + 264 + ], + "type": "inline_equation", + "content": "P(y|x)" + }, + { + "bbox": [ + 130, + 215, + 482, + 264 + ], + "type": "text", + "content": " such that" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 246, + 272, + 481, + 285 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 246, + 272, + 481, + 285 + ], + "spans": [ + { + "bbox": [ + 246, + 272, + 481, + 285 + ], + "type": "interline_equation", + "content": "P (y | x) \\geq P (\\hat {y} | x) \\quad \\forall \\hat {y} \\in \\mathcal {Y} \\tag {3}", + "image_path": "3d063701940f0b59e76a45037ddf2e947d71c1376a707ab1aa55b2b8924d2fd3.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 292, + 482, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 292, + 482, + 376 + ], + "spans": [ + { + "bbox": [ + 131, + 292, + 482, + 376 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 292, + 482, + 376 + ], + "type": "inline_equation", + "content": "\\hat{y}" + }, + { + "bbox": [ + 131, + 292, + 482, + 376 + ], + "type": "text", + "content": " denotes incorrect labels. The DNN classifier, is trained to approximate the distribution " + }, + { + "bbox": [ + 131, + 292, + 482, + 376 + ], + "type": "inline_equation", + "content": "\\tilde{P}(y|x) \\approx P(y|x)" + }, + { + "bbox": [ + 131, + 292, + 482, + 376 + ], + "type": "text", + "content": " with empirical risk minimization approaches based on an assumption that in-sample data and out-of-sample data shared a similar distribution " + }, + { + "bbox": [ + 131, + 292, + 482, + 376 + ], + "type": "inline_equation", + "content": "P(y|x)" + }, + { + "bbox": [ + 131, + 292, + 482, + 376 + ], + "type": "text", + "content": ". However, the unintentional interference denoted by " + }, + { + "bbox": [ + 131, + 292, + 482, + 376 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 131, + 292, + 482, + 376 + ], + "type": "text", + "content": ", which can be modeled as an additive noise to " + }, + { + "bbox": [ + 131, + 292, + 482, + 376 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 131, + 292, + 482, + 376 + ], + "type": "text", + "content": ", creates a new dataset " + }, + { + "bbox": [ + 131, + 292, + 482, + 376 + ], + "type": "inline_equation", + "content": "D' = \\{x + \\delta, y\\}" + }, + { + "bbox": [ + 131, + 292, + 482, + 376 + ], + "type": "text", + "content": ". It introduces a covariate shift that " + }, + { + "bbox": [ + 131, + 292, + 482, + 376 + ], + "type": "inline_equation", + "content": "P(x + \\delta) \\neq P(x)" + }, + { + "bbox": [ + 131, + 292, + 482, + 376 + ], + "type": "text", + "content": " but " + }, + { + "bbox": [ + 131, + 292, + 482, + 376 + ], + "type": "inline_equation", + "content": "P(y|x + \\delta) = P(y|x)" + }, + { + "bbox": [ + 131, + 292, + 482, + 376 + ], + "type": "text", + "content": ", resulting in classification errors such that" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 228, + 384, + 481, + 397 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 228, + 384, + 481, + 397 + ], + "spans": [ + { + "bbox": [ + 228, + 384, + 481, + 397 + ], + "type": "interline_equation", + "content": "\\tilde {P} (\\hat {y} | x + \\delta) \\geq \\tilde {P} (y | x + \\delta) \\quad \\exists \\hat {y} \\in \\mathcal {Y} \\tag {4}", + "image_path": "0dc049516dedc3c2a8ecd8758a70f36e001f8310bcf23d4769f493c3cdaf323e.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 405, + 482, + 489 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 405, + 482, + 489 + ], + "spans": [ + { + "bbox": [ + 130, + 405, + 482, + 489 + ], + "type": "text", + "content": "In the intentional interference scenario, algorithms attempt to find the minimum covariate shift that leads to Eq. (4). In this setting, a small perturbation in " + }, + { + "bbox": [ + 130, + 405, + 482, + 489 + ], + "type": "inline_equation", + "content": "l_{p}" + }, + { + "bbox": [ + 130, + 405, + 482, + 489 + ], + "type": "text", + "content": " distance " + }, + { + "bbox": [ + 130, + 405, + 482, + 489 + ], + "type": "inline_equation", + "content": "||\\delta||_{p} \\leq \\epsilon" + }, + { + "bbox": [ + 130, + 405, + 482, + 489 + ], + "type": "text", + "content": " is intentionally crafted to mislead the DNN classifier where " + }, + { + "bbox": [ + 130, + 405, + 482, + 489 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 130, + 405, + 482, + 489 + ], + "type": "text", + "content": " denotes the constraint. Attackers formulate Eq. (4) as an optimization problem over the input space " + }, + { + "bbox": [ + 130, + 405, + 482, + 489 + ], + "type": "inline_equation", + "content": "\\mathcal{X}" + }, + { + "bbox": [ + 130, + 405, + 482, + 489 + ], + "type": "text", + "content": ", and hence can apply the gradient information of the loss function w.r.t. " + }, + { + "bbox": [ + 130, + 405, + 482, + 489 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 130, + 405, + 482, + 489 + ], + "type": "text", + "content": " to find the " + }, + { + "bbox": [ + 130, + 405, + 482, + 489 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 130, + 405, + 482, + 489 + ], + "type": "text", + "content": ". Projected Gradient Descent (PGD) [35] provided a unified view on iterative gradient attacks under " + }, + { + "bbox": [ + 130, + 405, + 482, + 489 + ], + "type": "inline_equation", + "content": "l_{p}" + }, + { + "bbox": [ + 130, + 405, + 482, + 489 + ], + "type": "text", + "content": " constraint:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 249, + 497, + 481, + 522 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 249, + 497, + 481, + 522 + ], + "spans": [ + { + "bbox": [ + 249, + 497, + 481, + 522 + ], + "type": "interline_equation", + "content": "x ^ {i + 1} = x ^ {i} + \\alpha \\cdot \\epsilon \\cdot \\frac {\\nabla_ {x} \\mathcal {L}}{| | \\nabla_ {x} \\mathcal {L} | | _ {p}} \\tag {5}", + "image_path": "8f3a17d67526e259ffdb3b43153ccf8c98bef46670b76f0effdc873bce282a89.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 529, + 482, + 570 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 529, + 482, + 570 + ], + "spans": [ + { + "bbox": [ + 131, + 529, + 482, + 570 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 529, + 482, + 570 + ], + "type": "inline_equation", + "content": "x^{i}" + }, + { + "bbox": [ + 131, + 529, + 482, + 570 + ], + "type": "text", + "content": " is the adversarial sample at " + }, + { + "bbox": [ + 131, + 529, + 482, + 570 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 131, + 529, + 482, + 570 + ], + "type": "text", + "content": " step, " + }, + { + "bbox": [ + 131, + 529, + 482, + 570 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 131, + 529, + 482, + 570 + ], + "type": "text", + "content": " is the learning rate, and " + }, + { + "bbox": [ + 131, + 529, + 482, + 570 + ], + "type": "inline_equation", + "content": "\\frac{\\nabla_{x}\\mathcal{L}}{||\\nabla_{x}\\mathcal{L}||_{p}}" + }, + { + "bbox": [ + 131, + 529, + 482, + 570 + ], + "type": "text", + "content": " is the normalized gradient of a loss function " + }, + { + "bbox": [ + 131, + 529, + 482, + 570 + ], + "type": "inline_equation", + "content": "\\mathcal{L}" + }, + { + "bbox": [ + 131, + 529, + 482, + 570 + ], + "type": "text", + "content": " (e.g., a sign function in " + }, + { + "bbox": [ + 131, + 529, + 482, + 570 + ], + "type": "inline_equation", + "content": "l_{\\infty}" + }, + { + "bbox": [ + 131, + 529, + 482, + 570 + ], + "type": "text", + "content": " space). " + }, + { + "bbox": [ + 131, + 529, + 482, + 570 + ], + "type": "inline_equation", + "content": "\\mathcal{L}" + }, + { + "bbox": [ + 131, + 529, + 482, + 570 + ], + "type": "text", + "content": " can be cross entropy or other more advanced loss functions [11]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 585, + 482, + 645 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 585, + 482, + 645 + ], + "spans": [ + { + "bbox": [ + 130, + 585, + 482, + 645 + ], + "type": "text", + "content": "Threat Model for Data Rate. Let " + }, + { + "bbox": [ + 130, + 585, + 482, + 645 + ], + "type": "inline_equation", + "content": "H(x) = z" + }, + { + "bbox": [ + 130, + 585, + 482, + 645 + ], + "type": "text", + "content": " denotes the head network where " + }, + { + "bbox": [ + 130, + 585, + 482, + 645 + ], + "type": "inline_equation", + "content": "z \\in \\mathcal{Z}" + }, + { + "bbox": [ + 130, + 585, + 482, + 645 + ], + "type": "text", + "content": " is the output of the head and " + }, + { + "bbox": [ + 130, + 585, + 482, + 645 + ], + "type": "inline_equation", + "content": "\\mathcal{Z}" + }, + { + "bbox": [ + 130, + 585, + 482, + 645 + ], + "type": "text", + "content": " is the latent space. The goal is to achieve a minimum compression rate " + }, + { + "bbox": [ + 130, + 585, + 482, + 645 + ], + "type": "inline_equation", + "content": "-\\log_2 P_Z(z)" + }, + { + "bbox": [ + 130, + 585, + 482, + 645 + ], + "type": "text", + "content": ". As " + }, + { + "bbox": [ + 130, + 585, + 482, + 645 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 130, + 585, + 482, + 645 + ], + "type": "text", + "content": " is dependent on " + }, + { + "bbox": [ + 130, + 585, + 482, + 645 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 130, + 585, + 482, + 645 + ], + "type": "text", + "content": ", the distribution shift introduced by unintentional interference can result in a difference in coding rate:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 220, + 654, + 481, + 666 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 220, + 654, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 220, + 654, + 481, + 666 + ], + "type": "interline_equation", + "content": "- \\log_ {2} P _ {Z} (H (x + \\delta)) \\neq - \\log_ {2} P _ {Z} (H (x)) \\tag {6}", + "image_path": "396af931e44f0a0b812e98915d14e42807c7a392ad813f4892063ec49c26ed3f.jpg" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 200, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 200, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 200, + 91, + 448, + 102 + ], + "type": "text", + "content": "Resilience of Entropy Model in Distributed Neural Networks" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 149, + 136, + 465, + 312 + ], + "blocks": [ + { + "bbox": [ + 163, + 114, + 449, + 126 + ], + "lines": [ + { + "bbox": [ + 163, + 114, + 449, + 126 + ], + "spans": [ + { + "bbox": [ + 163, + 114, + 449, + 126 + ], + "type": "text", + "content": "Table 1: DNNs, entropy models and datasets utilized in experiments." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 149, + 136, + 465, + 312 + ], + "lines": [ + { + "bbox": [ + 149, + 136, + 465, + 312 + ], + "spans": [ + { + "bbox": [ + 149, + 136, + 465, + 312 + ], + "type": "table", + "html": "
DNN Architectures
ResNet-50 [23]A standard baseline in distributed DNNs [38-40, 42]
ResNet-101 [23]A deeper model in the ResNet family
RegNetY-6.4GF [47]An advanced design with better performance
Entropy Models
FP [5]A fully factorized prior model
MSHP [43]An effective learnable entropy model
Datasets
ImageNet [16]The standard ImageNet validation dataset
ImageNet-C [25]A synthetic dataset with 15 corruptions and 5 severities
Random NoiseNoisy images as a baseline for adversarial robustness
PGD-Acc [35]PGD targeting classification
PGD-EPGD targeting entropy
", + "image_path": "9a36f0770cb7ccb5b113b0f91198b1d751ff5b6f93b752db241ec5351f2081d3.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 334, + 482, + 369 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 334, + 482, + 369 + ], + "spans": [ + { + "bbox": [ + 130, + 334, + 482, + 369 + ], + "type": "text", + "content": "Note that the unintentional distribution shift of " + }, + { + "bbox": [ + 130, + 334, + 482, + 369 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 130, + 334, + 482, + 369 + ], + "type": "text", + "content": " does not necessarily result in a larger bit rate. However, in the adversarial setting, attackers intend to maximize the entropy of " + }, + { + "bbox": [ + 130, + 334, + 482, + 369 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 130, + 334, + 482, + 369 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 252, + 378, + 481, + 400 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 252, + 378, + 481, + 400 + ], + "spans": [ + { + "bbox": [ + 252, + 378, + 481, + 400 + ], + "type": "interline_equation", + "content": "\\min _ {\\delta} \\log_ {2} P _ {Z} (H (x + \\delta)) \\tag {7}", + "image_path": "75911fad5c4a2291b6d6abb2aca35e9ea413f2d04dbdfdaf761680a1e1047ace.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 256, + 399, + 318, + 411 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 256, + 399, + 318, + 411 + ], + "spans": [ + { + "bbox": [ + 256, + 399, + 318, + 411 + ], + "type": "interline_equation", + "content": "s. t. \\quad | | \\delta | | _ {p} \\leq \\epsilon", + "image_path": "494ae4756acea8c4a89f903dedb0342084765d2dab2344f0894a48376944b5ac.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 418, + 485, + 444 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 418, + 485, + 444 + ], + "spans": [ + { + "bbox": [ + 130, + 418, + 485, + 444 + ], + "type": "text", + "content": "Thus, to correctly assess the robustness of entropy models, we use the same PGD algorithm in Eq. (5) yet with the entropy as its loss function " + }, + { + "bbox": [ + 130, + 418, + 485, + 444 + ], + "type": "inline_equation", + "content": "\\mathcal{L} = -\\log_2P_Z(H(x))" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 460, + 438, + 475 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 460, + 438, + 475 + ], + "spans": [ + { + "bbox": [ + 130, + 460, + 438, + 475 + ], + "type": "text", + "content": "4 Understanding the Resilience of Entropy Model" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 485, + 262, + 498 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 485, + 262, + 498 + ], + "spans": [ + { + "bbox": [ + 131, + 485, + 262, + 498 + ], + "type": "text", + "content": "4.1 Experimental Setup" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 504, + 482, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 504, + 482, + 553 + ], + "spans": [ + { + "bbox": [ + 130, + 504, + 482, + 553 + ], + "type": "text", + "content": "In this section, we outline the experimental setup for evaluating the resilience of entropy models. We first specify the DNNs and entropy models under consideration, followed by a description of the metrics and datasets used in our experiments. A summary of our experimental setup is provided in Tab. 1." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 569, + 482, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 569, + 482, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 569, + 482, + 667 + ], + "type": "text", + "content": "Ablation of DNNs and Entropy Models. We consider three distinct DNNs: ResNet-50 [23], ResNet-101 [23] and RegNetY-6.4GF [47]. ResNet-50 serves as a common baseline in distributed DNN literature [38-40,42] while ResNet-101 is a deeper variant with the same design as former. RegNetY-6.4GF has an advanced design with better classification performance. The early layers are replaced by a specific tailored head architecture as proposed in [42] and incorporated with different entropy models. For the sake of brevity, we consider ResNet-50 as the default DNN architecture without additional specification. We employ two entropy" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 231, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 231, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 231, + 102 + ], + "type": "text", + "content": "M. Zhang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 152, + 114, + 302, + 235 + ], + "blocks": [ + { + "bbox": [ + 152, + 114, + 302, + 235 + ], + "lines": [ + { + "bbox": [ + 152, + 114, + 302, + 235 + ], + "spans": [ + { + "bbox": [ + 152, + 114, + 302, + 235 + ], + "type": "image", + "image_path": "749e0e30e507829f0b3bba2cddea1397c49d6d7a843a8db6bd6add6a20458096.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 189, + 237, + 266, + 246 + ], + "lines": [ + { + "bbox": [ + 189, + 237, + 266, + 246 + ], + "spans": [ + { + "bbox": [ + 189, + 237, + 266, + 246 + ], + "type": "text", + "content": "(a) resilience w.r.t. " + }, + { + "bbox": [ + 189, + 237, + 266, + 246 + ], + "type": "inline_equation", + "content": "\\beta" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 304, + 116, + 462, + 235 + ], + "blocks": [ + { + "bbox": [ + 304, + 116, + 462, + 235 + ], + "lines": [ + { + "bbox": [ + 304, + 116, + 462, + 235 + ], + "spans": [ + { + "bbox": [ + 304, + 116, + 462, + 235 + ], + "type": "image", + "image_path": "e54d2744cd2b3bda3702db2cc58fa4f3936dcf54b086ee1492e3bec496597862.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 331, + 237, + 435, + 246 + ], + "lines": [ + { + "bbox": [ + 331, + 237, + 435, + 246 + ], + "spans": [ + { + "bbox": [ + 331, + 237, + 435, + 246 + ], + "type": "text", + "content": "(b) resilience w.r.t. severities" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 162, + 255, + 449, + 267 + ], + "lines": [ + { + "bbox": [ + 162, + 255, + 449, + 267 + ], + "spans": [ + { + "bbox": [ + 162, + 255, + 449, + 267 + ], + "type": "text", + "content": "Fig. 3: Resilience of accuracy and data rate to 4 common corruptions." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 289, + 482, + 374 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 289, + 482, + 374 + ], + "spans": [ + { + "bbox": [ + 130, + 289, + 482, + 374 + ], + "type": "text", + "content": "models: Factorized Prior (FP) [5] and Mean Scale Hyper Prior (MSHP) [43]. FP is the earliest design of entropy model and a basic component in many other advanced designs, while MSHP provides more powerful compression performance by injecting a hierarchical learnable block before FP to extract the entropy information. For conciseness, we consider MSHP as the default entropy model. In addition, since models are trained with different rate-distortion trade-off " + }, + { + "bbox": [ + 130, + 289, + 482, + 374 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 130, + 289, + 482, + 374 + ], + "type": "text", + "content": ", we also investigate the compression robustness w.r.t. different " + }, + { + "bbox": [ + 130, + 289, + 482, + 374 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 130, + 289, + 482, + 374 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 390, + 482, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 390, + 482, + 559 + ], + "spans": [ + { + "bbox": [ + 130, + 390, + 482, + 559 + ], + "type": "text", + "content": "Datasets and Metrics. Conversely from existing work that measures the compression performance with bits per pixel (BPP) [5,6,14,43,44], we evaluate the entropy model with the data size of the whole bit stream after encoding as in [37]. Compared to BPP, data size quantifies more directly the networking traffic between the head and tail. To assess the resilience to unintentional interference, we use the ImageNet-C dataset proposed by [25] comprising 15 common corruptions classified into 4 categories (noise, blur, digital, and weather) with 5 different severities. We also report the data size and accuracy on the clean ImageNet [16] validation set as a baseline. To assess the resilience to intentional interference, we implement the PGD targeting both the classification and compression performance in " + }, + { + "bbox": [ + 130, + 390, + 482, + 559 + ], + "type": "inline_equation", + "content": "l_{\\infty}" + }, + { + "bbox": [ + 130, + 390, + 482, + 559 + ], + "type": "text", + "content": " space. For clarity, we denote the conventional PGD as PGD-Acc and the PGD targeting entropy models as PGD-E. Meanwhile, random noise with the same perturbation level is added to the clean ImageNet as a baseline comparison for adversarial robustness." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 574, + 362, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 574, + 362, + 586 + ], + "spans": [ + { + "bbox": [ + 132, + 574, + 362, + 586 + ], + "type": "text", + "content": "4.2 Resilience to Unintentional Interference" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": "Resilience w.r.t. " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": ". We select one corruption as a representative from each category in ImageNet-C [25] to investigate the resilience as a function of ratedistortion trade-off " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": ". Fig. 3a shows the compression and classification performance on defocus blur, shot noise, snow and contrast dataset, each with the severity of 5. Compared to the clean ImageNet, all corruptions decrease the accuracy significantly. However, not all corruptions increase the data size. Only" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 201, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 201, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 201, + 91, + 448, + 102 + ], + "type": "text", + "content": "Resilience of Entropy Model in Distributed Neural Networks" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 136, + 157, + 478, + 216 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "type": "text", + "content": "Table 2: Resilience w.r.t. prior models. The average data size of FP and MSHP for clean ImageNet are " + }, + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "type": "inline_equation", + "content": "11.65 \\pm 1.02" + }, + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "type": "inline_equation", + "content": "9.62 \\pm 1.70" + }, + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "type": "text", + "content": " KBBytes, respectively. Red for larger data size and blue for lower compared to the baseline." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 136, + 157, + 478, + 216 + ], + "lines": [ + { + "bbox": [ + 136, + 157, + 478, + 216 + ], + "spans": [ + { + "bbox": [ + 136, + 157, + 478, + 216 + ], + "type": "table", + "html": "
Prior \nModelGaussian \nSize[KB]Noise \nAcc[%]Motion \nSize[KB]Blur \nAcc[%]Impulse \nSize[KB]Noise \nAcc[%]Glass \nSize[KB]Blur \nAcc[%]
FP13.19±0.4059.5110.51±0.7062.2213.36±0.3450.9310.43±0.6653.63
MSHP11.96±0.4759.367.88±1.4661.5812.52±0.3551.127.78±1.3953.67
", + "image_path": "8be872c73f70ba0a8c4e4513c940128f241a103bbc273e23017aba14c4602d5b.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 236, + 482, + 308 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 236, + 482, + 308 + ], + "spans": [ + { + "bbox": [ + 130, + 236, + 482, + 308 + ], + "type": "text", + "content": "shot noise increases the data size by " + }, + { + "bbox": [ + 130, + 236, + 482, + 308 + ], + "type": "inline_equation", + "content": "65.31\\%" + }, + { + "bbox": [ + 130, + 236, + 482, + 308 + ], + "type": "text", + "content": " on average. In contrast, Defocus blur and contrast decrease the data size by " + }, + { + "bbox": [ + 130, + 236, + 482, + 308 + ], + "type": "inline_equation", + "content": "53.31\\%" + }, + { + "bbox": [ + 130, + 236, + 482, + 308 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 236, + 482, + 308 + ], + "type": "inline_equation", + "content": "67.45\\%" + }, + { + "bbox": [ + 130, + 236, + 482, + 308 + ], + "type": "text", + "content": " on average, respectively. Snow has little affect to compression, resulting in only " + }, + { + "bbox": [ + 130, + 236, + 482, + 308 + ], + "type": "inline_equation", + "content": "4.42\\%" + }, + { + "bbox": [ + 130, + 236, + 482, + 308 + ], + "type": "text", + "content": " decrease of data size. This indicates that the entropy model learns a different set of features than the classification model. Since it shows the same trend for different rate-distortion trade-off factors, we only consider " + }, + { + "bbox": [ + 130, + 236, + 482, + 308 + ], + "type": "inline_equation", + "content": "\\beta = 0.08" + }, + { + "bbox": [ + 130, + 236, + 482, + 308 + ], + "type": "text", + "content": " in next experiments." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 325, + 482, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 325, + 482, + 434 + ], + "spans": [ + { + "bbox": [ + 130, + 325, + 482, + 434 + ], + "type": "text", + "content": "Resilience w.r.t. Severities. Next, we explore resilience w.r.t. severity levels using the same corruptions mentioned earlier. As illustrated in Fig. 3b, the data size of shot noise rises with increasing severity, whereas defocus blur and contrast consistently decrease the data size. Snow shows minimal variation across different severity levels. This can be attributed to defocus blur and contrast removing high-frequency components in images, while shot noise introduces new patterns in high-frequency space. Snow, on the other hand, lacks a specific pattern in the frequency domain. Thus, the compression is particularly sensitive to high-frequency information in the input space." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 449, + 482, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 449, + 482, + 594 + ], + "spans": [ + { + "bbox": [ + 130, + 449, + 482, + 594 + ], + "type": "text", + "content": "Resilience w.r.t. Prior Models. Tab. 2 shows the impact of additional 4 corruptions at severity level 1 on both FP and MSHP model. To further validate the previous observation, we choose 2 noise corruptions (Gaussian noise and impulse noise), which introduce high-frequency noise to images, and two blur corruptions (motion blur and glass blur), which eliminate high-frequency components from images. As shown in Tab. 2, Gaussian noise increases " + }, + { + "bbox": [ + 130, + 449, + 482, + 594 + ], + "type": "inline_equation", + "content": "13.22\\%" + }, + { + "bbox": [ + 130, + 449, + 482, + 594 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 449, + 482, + 594 + ], + "type": "inline_equation", + "content": "24.32\\%" + }, + { + "bbox": [ + 130, + 449, + 482, + 594 + ], + "type": "text", + "content": " of data size for FP and MSHP, respectively. Impulse noise increases " + }, + { + "bbox": [ + 130, + 449, + 482, + 594 + ], + "type": "inline_equation", + "content": "14.42\\%" + }, + { + "bbox": [ + 130, + 449, + 482, + 594 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 449, + 482, + 594 + ], + "type": "inline_equation", + "content": "30.15\\%" + }, + { + "bbox": [ + 130, + 449, + 482, + 594 + ], + "type": "text", + "content": " of data size for FP and MSHP, respectively. On the other hand, motion blur reduces " + }, + { + "bbox": [ + 130, + 449, + 482, + 594 + ], + "type": "inline_equation", + "content": "9.79\\%" + }, + { + "bbox": [ + 130, + 449, + 482, + 594 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 449, + 482, + 594 + ], + "type": "inline_equation", + "content": "18.09\\%" + }, + { + "bbox": [ + 130, + 449, + 482, + 594 + ], + "type": "text", + "content": " of data size for FP and MSHP, respectively. Glass blur reduces " + }, + { + "bbox": [ + 130, + 449, + 482, + 594 + ], + "type": "inline_equation", + "content": "10.47\\%" + }, + { + "bbox": [ + 130, + 449, + 482, + 594 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 449, + 482, + 594 + ], + "type": "inline_equation", + "content": "19.13\\%" + }, + { + "bbox": [ + 130, + 449, + 482, + 594 + ], + "type": "text", + "content": " of data size for FP and MSHP, respectively. Thus, the previous finding that compressive features reside in the high-frequency space remains consistent for different entropy models." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 610, + 348, + 621 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 610, + 348, + 621 + ], + "spans": [ + { + "bbox": [ + 132, + 610, + 348, + 621 + ], + "type": "text", + "content": "4.3 Resilience to Intentional Interference" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 629, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 482, + 665 + ], + "type": "text", + "content": "Resilience w.r.t. DNNs Fig. 4a shows the resilience of different DNN architectures to accuracy and entropy attacks with the " + }, + { + "bbox": [ + 130, + 629, + 482, + 665 + ], + "type": "inline_equation", + "content": "l_{\\infty}" + }, + { + "bbox": [ + 130, + 629, + 482, + 665 + ], + "type": "text", + "content": " constraint " + }, + { + "bbox": [ + 130, + 629, + 482, + 665 + ], + "type": "inline_equation", + "content": "\\epsilon = 4 / 255" + }, + { + "bbox": [ + 130, + 629, + 482, + 665 + ], + "type": "text", + "content": ". As evident in the figure, the attacks show similar patterns regardless of their" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 231, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 231, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 231, + 102 + ], + "type": "text", + "content": "M. Zhang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 149, + 114, + 305, + 239 + ], + "blocks": [ + { + "bbox": [ + 149, + 114, + 305, + 239 + ], + "lines": [ + { + "bbox": [ + 149, + 114, + 305, + 239 + ], + "spans": [ + { + "bbox": [ + 149, + 114, + 305, + 239 + ], + "type": "image", + "image_path": "c1422fe811d72e379e187f90ad9483c30699ed87429a3cf3260ddb21859c5d72.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 159, + 240, + 296, + 250 + ], + "lines": [ + { + "bbox": [ + 159, + 240, + 296, + 250 + ], + "spans": [ + { + "bbox": [ + 159, + 240, + 296, + 250 + ], + "type": "text", + "content": "(a) resilience w.r.t. DNN architectures" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 307, + 116, + 465, + 238 + ], + "blocks": [ + { + "bbox": [ + 307, + 116, + 465, + 238 + ], + "lines": [ + { + "bbox": [ + 307, + 116, + 465, + 238 + ], + "spans": [ + { + "bbox": [ + 307, + 116, + 465, + 238 + ], + "type": "image", + "image_path": "2fdb8b2a921a1e5acb43a4edba73095d1962eed910e547599bc7c07565b2c889.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 329, + 240, + 444, + 250 + ], + "lines": [ + { + "bbox": [ + 329, + 240, + 444, + 250 + ], + "spans": [ + { + "bbox": [ + 329, + 240, + 444, + 250 + ], + "type": "text", + "content": "(b) resilience w.r.t. prior models" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 162, + 259, + 450, + 270 + ], + "lines": [ + { + "bbox": [ + 162, + 259, + 450, + 270 + ], + "spans": [ + { + "bbox": [ + 162, + 259, + 450, + 270 + ], + "type": "text", + "content": "Fig. 4: Resilience of accuracy and data rate to intentional interference" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 294, + 482, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 294, + 482, + 365 + ], + "spans": [ + { + "bbox": [ + 130, + 294, + 482, + 365 + ], + "type": "text", + "content": "victim models. While PGD-Acc reduces " + }, + { + "bbox": [ + 130, + 294, + 482, + 365 + ], + "type": "inline_equation", + "content": "42.88\\%" + }, + { + "bbox": [ + 130, + 294, + 482, + 365 + ], + "type": "text", + "content": " accuracy on average, it only increases by " + }, + { + "bbox": [ + 130, + 294, + 482, + 365 + ], + "type": "inline_equation", + "content": "7.28\\%" + }, + { + "bbox": [ + 130, + 294, + 482, + 365 + ], + "type": "text", + "content": " of data size on average compared to random noise. On the other hand, PGD-E decreases " + }, + { + "bbox": [ + 130, + 294, + 482, + 365 + ], + "type": "inline_equation", + "content": "2.34\\%" + }, + { + "bbox": [ + 130, + 294, + 482, + 365 + ], + "type": "text", + "content": " accuracy on average compared to random noise but increases the bit rate 1.35x times on average. This indicates that PGD-E and PGD-Acc are targeting separate features in the input space which has minor impact on each other." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 384, + 482, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 384, + 482, + 529 + ], + "spans": [ + { + "bbox": [ + 130, + 384, + 482, + 529 + ], + "type": "text", + "content": "Resilience w.r.t. Prior Models. Fig. 4b illustrates the effect of adversarial attacks with " + }, + { + "bbox": [ + 130, + 384, + 482, + 529 + ], + "type": "inline_equation", + "content": "\\epsilon = 8 / 255" + }, + { + "bbox": [ + 130, + 384, + 482, + 529 + ], + "type": "text", + "content": " on victims using different prior models. As shown in the figure, both entropy models show considerable vulnerability to PGD-E compared to PGD-Acc. The data size are increased by " + }, + { + "bbox": [ + 130, + 384, + 482, + 529 + ], + "type": "inline_equation", + "content": "21.93\\%" + }, + { + "bbox": [ + 130, + 384, + 482, + 529 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 384, + 482, + 529 + ], + "type": "inline_equation", + "content": "46.82\\%" + }, + { + "bbox": [ + 130, + 384, + 482, + 529 + ], + "type": "text", + "content": " for FP and MSHP respectively while the performance loss are " + }, + { + "bbox": [ + 130, + 384, + 482, + 529 + ], + "type": "inline_equation", + "content": "9.91\\%" + }, + { + "bbox": [ + 130, + 384, + 482, + 529 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 384, + 482, + 529 + ], + "type": "inline_equation", + "content": "6.62\\%" + }, + { + "bbox": [ + 130, + 384, + 482, + 529 + ], + "type": "text", + "content": " respectively. On the other hand, PGD-Acc successfully reduces " + }, + { + "bbox": [ + 130, + 384, + 482, + 529 + ], + "type": "inline_equation", + "content": "72.10\\%" + }, + { + "bbox": [ + 130, + 384, + 482, + 529 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 384, + 482, + 529 + ], + "type": "inline_equation", + "content": "57.10\\%" + }, + { + "bbox": [ + 130, + 384, + 482, + 529 + ], + "type": "text", + "content": " classification performance of FP and MSHP compared to random noise respectively. However, it only increases " + }, + { + "bbox": [ + 130, + 384, + 482, + 529 + ], + "type": "inline_equation", + "content": "5.66\\%" + }, + { + "bbox": [ + 130, + 384, + 482, + 529 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 384, + 482, + 529 + ], + "type": "inline_equation", + "content": "10.19\\%" + }, + { + "bbox": [ + 130, + 384, + 482, + 529 + ], + "type": "text", + "content": " bit rate for FP and MSHP respectively. Hence, attackers targeting compression tend to perturb a different set of features than attackers targeting discriminative tasks regardless of the entropy models. In addition, while the FP has a better robustness in compression, it has significantly worse robustness in accuracy." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 547, + 331, + 561 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 547, + 331, + 561 + ], + "spans": [ + { + "bbox": [ + 131, + 547, + 331, + 561 + ], + "type": "text", + "content": "5 Proposed Defense Mechanism" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 573, + 446, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 573, + 446, + 586 + ], + "spans": [ + { + "bbox": [ + 131, + 573, + 446, + 586 + ], + "type": "text", + "content": "5.1 Disentangle Compression Features in Frequency Domain" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": "As shown in Sec. 4.2, interference that introduces high-frequency noise (e.g., shot noise) can effectively increase the data size while interference that removes high-frequency information (e.g., defocus blur) reduces the data size. Intuitively this is a consequence of the small head in distributed DNNs as semantic information is usually captured in deeper layers while early layers tend to capture low level information [8]. To validate our intuition, we introduce total variation, defined" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 201, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 201, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 201, + 91, + 448, + 102 + ], + "type": "text", + "content": "Resilience of Entropy Model in Distributed Neural Networks" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 154, + 116, + 203, + 165 + ], + "blocks": [ + { + "bbox": [ + 154, + 116, + 203, + 165 + ], + "lines": [ + { + "bbox": [ + 154, + 116, + 203, + 165 + ], + "spans": [ + { + "bbox": [ + 154, + 116, + 203, + 165 + ], + "type": "image", + "image_path": "8cd8faf40432ed1024ee9ac561f449d17f469e27643b54f9ea85ec2200e8c119.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 154, + 167, + 203, + 215 + ], + "blocks": [ + { + "bbox": [ + 154, + 167, + 203, + 215 + ], + "lines": [ + { + "bbox": [ + 154, + 167, + 203, + 215 + ], + "spans": [ + { + "bbox": [ + 154, + 167, + 203, + 215 + ], + "type": "image", + "image_path": "d0636016d8cacc5ccf4eb783d04afd96f3e7ad8320ae5d84c517527343444d60.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 227, + 480, + 251 + ], + "lines": [ + { + "bbox": [ + 131, + 227, + 480, + 251 + ], + "spans": [ + { + "bbox": [ + 131, + 227, + 480, + 251 + ], + "type": "text", + "content": "Fig. 5: Entropy models are sensitive to high-frequency features. From left to right: images from ImageNet validation set, bit rate maps and total variation maps." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 205, + 116, + 254, + 165 + ], + "blocks": [ + { + "bbox": [ + 205, + 116, + 254, + 165 + ], + "lines": [ + { + "bbox": [ + 205, + 116, + 254, + 165 + ], + "spans": [ + { + "bbox": [ + 205, + 116, + 254, + 165 + ], + "type": "image", + "image_path": "be791e5b2e7d3b8a3a3024edacae1bae51cc2b4333197a34a697c1a2042fe377.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 205, + 167, + 254, + 215 + ], + "blocks": [ + { + "bbox": [ + 205, + 167, + 254, + 215 + ], + "lines": [ + { + "bbox": [ + 205, + 167, + 254, + 215 + ], + "spans": [ + { + "bbox": [ + 205, + 167, + 254, + 215 + ], + "type": "image", + "image_path": "b7634bdb8b1a114c0539bd8177bf72e1cec1176c0bf8e3781ca928186e8a3dfb.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 258, + 116, + 306, + 165 + ], + "blocks": [ + { + "bbox": [ + 258, + 116, + 306, + 165 + ], + "lines": [ + { + "bbox": [ + 258, + 116, + 306, + 165 + ], + "spans": [ + { + "bbox": [ + 258, + 116, + 306, + 165 + ], + "type": "image", + "image_path": "3006b3bbba65f4a2dd721196d2cd4fae009fca8f33c317c9682df03784ea6d6c.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 258, + 167, + 306, + 215 + ], + "blocks": [ + { + "bbox": [ + 258, + 167, + 306, + 215 + ], + "lines": [ + { + "bbox": [ + 258, + 167, + 306, + 215 + ], + "spans": [ + { + "bbox": [ + 258, + 167, + 306, + 215 + ], + "type": "image", + "image_path": "d31bd2c70df5c30176ed942294fe5476b0fc9e7d7c1731b84f8791d293d2b878.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 309, + 116, + 358, + 165 + ], + "blocks": [ + { + "bbox": [ + 309, + 116, + 358, + 165 + ], + "lines": [ + { + "bbox": [ + 309, + 116, + 358, + 165 + ], + "spans": [ + { + "bbox": [ + 309, + 116, + 358, + 165 + ], + "type": "image", + "image_path": "3ae55c9591f3169f26560aa464c58fd4bcd61295e1a0316061af2c26227f1df0.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 309, + 167, + 358, + 215 + ], + "blocks": [ + { + "bbox": [ + 309, + 167, + 358, + 215 + ], + "lines": [ + { + "bbox": [ + 309, + 167, + 358, + 215 + ], + "spans": [ + { + "bbox": [ + 309, + 167, + 358, + 215 + ], + "type": "image", + "image_path": "7d7c7698af0229855a4a33345826e806f3880bf4ba5a76002209c13e29b5be08.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 361, + 116, + 409, + 165 + ], + "blocks": [ + { + "bbox": [ + 361, + 116, + 409, + 165 + ], + "lines": [ + { + "bbox": [ + 361, + 116, + 409, + 165 + ], + "spans": [ + { + "bbox": [ + 361, + 116, + 409, + 165 + ], + "type": "image", + "image_path": "e2331d1baeff47774cf21ff0223a1803f1f80f68192ad5c4b04b6c6b6432331b.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 361, + 167, + 409, + 215 + ], + "blocks": [ + { + "bbox": [ + 361, + 167, + 409, + 215 + ], + "lines": [ + { + "bbox": [ + 361, + 167, + 409, + 215 + ], + "spans": [ + { + "bbox": [ + 361, + 167, + 409, + 215 + ], + "type": "image", + "image_path": "59b7d051e5b9660d0b36ec34203cecec281c924a50f04b35f49d529116c67fc4.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 412, + 116, + 461, + 165 + ], + "blocks": [ + { + "bbox": [ + 412, + 116, + 461, + 165 + ], + "lines": [ + { + "bbox": [ + 412, + 116, + 461, + 165 + ], + "spans": [ + { + "bbox": [ + 412, + 116, + 461, + 165 + ], + "type": "image", + "image_path": "f09587c113ad775d597b521159ddf19a1f4bfc4bf9757a2d0a514404d9be0c54.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 412, + 167, + 461, + 215 + ], + "blocks": [ + { + "bbox": [ + 412, + 167, + 461, + 215 + ], + "lines": [ + { + "bbox": [ + 412, + 167, + 461, + 215 + ], + "spans": [ + { + "bbox": [ + 412, + 167, + 461, + 215 + ], + "type": "image", + "image_path": "180fe7f9967c9d145a06778637bf0f9f8693da31097a41269fe68c12ed915f5d.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "bbox": [ + 131, + 273, + 479, + 298 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 273, + 479, + 298 + ], + "spans": [ + { + "bbox": [ + 131, + 273, + 479, + 298 + ], + "type": "text", + "content": "as the integral of image gradient magnitude, which is first proposed for image denoising in [51]. The anitrosopic total variation of a 2-D image is defined as" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 211, + 306, + 481, + 332 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 211, + 306, + 481, + 332 + ], + "spans": [ + { + "bbox": [ + 211, + 306, + 481, + 332 + ], + "type": "interline_equation", + "content": "T V (x) = \\sum_ {i, j} \\left| x _ {i + 1, j} - x _ {i, j} \\right| + \\left| x _ {i, j + 1} - x _ {i, j} \\right| \\tag {8}", + "image_path": "72562dc151cd6e6a2557fd1770c62cdf8bd098235e0a5579f674f0236b618be9.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 131, + 340, + 479, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 340, + 479, + 389 + ], + "spans": [ + { + "bbox": [ + 131, + 340, + 479, + 389 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 340, + 479, + 389 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 131, + 340, + 479, + 389 + ], + "type": "text", + "content": " denotes an image sample, " + }, + { + "bbox": [ + 131, + 340, + 479, + 389 + ], + "type": "inline_equation", + "content": "i,j" + }, + { + "bbox": [ + 131, + 340, + 479, + 389 + ], + "type": "text", + "content": " denotes the pixel position in width and height, respectively. The image gradient magnitude " + }, + { + "bbox": [ + 131, + 340, + 479, + 389 + ], + "type": "inline_equation", + "content": "|x_{i+1,j} - x_{i,j}|" + }, + { + "bbox": [ + 131, + 340, + 479, + 389 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 340, + 479, + 389 + ], + "type": "inline_equation", + "content": "|x_{i,j+1} - x_{i,j}|" + }, + { + "bbox": [ + 131, + 340, + 479, + 389 + ], + "type": "text", + "content": " describe the sharpness of pixel changes, where larger magnitude indicates high-frequency information such as edges and textures in images." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 130, + 388, + 480, + 473 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 388, + 480, + 473 + ], + "spans": [ + { + "bbox": [ + 130, + 388, + 480, + 473 + ], + "type": "text", + "content": "We compare the total variation map of an image with its bit rate map generated by the entropy model. To plot the total variation map, we first slice an image into many small patches and then compute the total variation for each patch. Meanwhile, the bit rate map is the estimated entropy of the latent representation " + }, + { + "bbox": [ + 130, + 388, + 480, + 473 + ], + "type": "inline_equation", + "content": "-\\log_2P_Z(z)" + }, + { + "bbox": [ + 130, + 388, + 480, + 473 + ], + "type": "text", + "content": ". As shown in Fig. 5, there exists a significant correlation between the bit rate map and the total variation map, which indicates that the entropy model is sensitive to high-frequency features in the input space." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 131, + 490, + 418, + 502 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 490, + 418, + 502 + ], + "spans": [ + { + "bbox": [ + 131, + 490, + 418, + 502 + ], + "type": "text", + "content": "5.2 Disentangle Compression Features in Spatial Space" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 130, + 510, + 479, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 479, + 569 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 479, + 569 + ], + "type": "text", + "content": "The work in [42] has shown that entropy model in distributed DNNs tends to assign a larger number of bits to task-oriented information, such as objects, while compressing irrelevant information, like the background. Consequently, it becomes intuitively easier for an adversary to increase the bit rate in the background area to attack the entropy model." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 130, + 570, + 480, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 480, + 665 + ], + "type": "text", + "content": "To interpret the action conducted by attackers, we generate the bit rate comparison map between adversarial samples generated by PGD-E and PGD-Acc. This comparison map represents the estimated entropy of adversarial samples generated by PGD-E subtracted from the bit rate of PGD-Acc-perturbed samples. In this context, the red area indicates that PGD-E tends to allocate more bits compared to PGD-Acc, whereas the blue region suggests that PGD-E tends to allocate fewer bits. As depicted in Fig. 6, PGD-E puts more efforts to increase the bit rate in background region, whereas PGD-Acc focuses on altering" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 231, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 231, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 231, + 102 + ], + "type": "text", + "content": "M. Zhang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 154, + 116, + 200, + 163 + ], + "blocks": [ + { + "bbox": [ + 154, + 116, + 200, + 163 + ], + "lines": [ + { + "bbox": [ + 154, + 116, + 200, + 163 + ], + "spans": [ + { + "bbox": [ + 154, + 116, + 200, + 163 + ], + "type": "image", + "image_path": "f4b643126eed029e66abdfbf9d68d09b122c66ab69117af0bf015a834d40eddb.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 175, + 480, + 209 + ], + "lines": [ + { + "bbox": [ + 131, + 175, + 480, + 209 + ], + "spans": [ + { + "bbox": [ + 131, + 175, + 480, + 209 + ], + "type": "text", + "content": "Fig. 6: Entropy models are more vulnerable to perturbation in background region. From left to right: images from ImageNet validation set, bit rate comparison between PGD-E and PGD-Acc (PGD-E allocate less bit in blue area and more bit in red area)." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 205, + 116, + 253, + 163 + ], + "blocks": [ + { + "bbox": [ + 205, + 116, + 253, + 163 + ], + "lines": [ + { + "bbox": [ + 205, + 116, + 253, + 163 + ], + "spans": [ + { + "bbox": [ + 205, + 116, + 253, + 163 + ], + "type": "image", + "image_path": "5e54bca6224bab3090b457b056900a27ea21b7a0ef4fe1d7bcd69af0154c4e99.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 258, + 116, + 305, + 163 + ], + "blocks": [ + { + "bbox": [ + 258, + 116, + 305, + 163 + ], + "lines": [ + { + "bbox": [ + 258, + 116, + 305, + 163 + ], + "spans": [ + { + "bbox": [ + 258, + 116, + 305, + 163 + ], + "type": "image", + "image_path": "fd4b0018db2a647c063c2563f557969b90e7d2619cea840d2880685b297bdbda.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 310, + 116, + 357, + 163 + ], + "blocks": [ + { + "bbox": [ + 310, + 116, + 357, + 163 + ], + "lines": [ + { + "bbox": [ + 310, + 116, + 357, + 163 + ], + "spans": [ + { + "bbox": [ + 310, + 116, + 357, + 163 + ], + "type": "image", + "image_path": "83b8dd48844c0a14b260f22c8f8b706d357945e0ac4d712130a7d2d9f3d48562.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 361, + 116, + 409, + 163 + ], + "blocks": [ + { + "bbox": [ + 361, + 116, + 409, + 163 + ], + "lines": [ + { + "bbox": [ + 361, + 116, + 409, + 163 + ], + "spans": [ + { + "bbox": [ + 361, + 116, + 409, + 163 + ], + "type": "image", + "image_path": "25da3dd17cf80b197cf2144dd097b4648a714051e107f57cfd08bd79aed2d207.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 414, + 116, + 461, + 163 + ], + "blocks": [ + { + "bbox": [ + 414, + 116, + 461, + 163 + ], + "lines": [ + { + "bbox": [ + 414, + 116, + 461, + 163 + ], + "spans": [ + { + "bbox": [ + 414, + 116, + 461, + 163 + ], + "type": "image", + "image_path": "2669ece1f428339965f782c3128b1b89cad20ebd9895799225f4316d6befe9c1.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 230, + 479, + 301 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 230, + 479, + 301 + ], + "spans": [ + { + "bbox": [ + 130, + 230, + 479, + 301 + ], + "type": "text", + "content": "task-oriented information. This is in line with the experiments in Sec. 4.3 that PGD-E and PGD-Acc target different features in the input space. On one hand, increasing the bit rate in the background region will affect little accuracy as they are irrelevant information for the discriminative task. On the other hand, modifying object information can significantly degrade the performance, yet it has less impact on compression." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 316, + 362, + 329 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 316, + 362, + 329 + ], + "spans": [ + { + "bbox": [ + 131, + 316, + 362, + 329 + ], + "type": "text", + "content": "5.3 Object-Aware Total Variation Denoising" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 333, + 479, + 369 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 333, + 479, + 369 + ], + "spans": [ + { + "bbox": [ + 130, + 333, + 479, + 369 + ], + "type": "text", + "content": "Given the observation in Sec. 5.1, we propose a denoising technique based on total variation to remove high-frequency noise in adversarial images where we solve the following optimization problem:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 242, + 374, + 480, + 397 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 242, + 374, + 480, + 397 + ], + "spans": [ + { + "bbox": [ + 242, + 374, + 480, + 397 + ], + "type": "interline_equation", + "content": "\\min _ {x} \\frac {1}{2} \\| x - x ^ {\\prime} \\| _ {2} ^ {2} + \\lambda \\cdot T V (x), \\tag {9}", + "image_path": "54cd02db1b87925d959edccc055f94766e370ffeedac93287238289cceb338a4.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 401, + 479, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 401, + 479, + 472 + ], + "spans": [ + { + "bbox": [ + 130, + 401, + 479, + 472 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 401, + 479, + 472 + ], + "type": "inline_equation", + "content": "x'" + }, + { + "bbox": [ + 130, + 401, + 479, + 472 + ], + "type": "text", + "content": " is the distorted image, " + }, + { + "bbox": [ + 130, + 401, + 479, + 472 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 130, + 401, + 479, + 472 + ], + "type": "text", + "content": " is the denoised image, " + }, + { + "bbox": [ + 130, + 401, + 479, + 472 + ], + "type": "inline_equation", + "content": "TV(x)" + }, + { + "bbox": [ + 130, + 401, + 479, + 472 + ], + "type": "text", + "content": " is the total variation of " + }, + { + "bbox": [ + 130, + 401, + 479, + 472 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 130, + 401, + 479, + 472 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 401, + 479, + 472 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 130, + 401, + 479, + 472 + ], + "type": "text", + "content": " is a regularization factor to control the degree of smoothing. Eq. (9) means the image " + }, + { + "bbox": [ + 130, + 401, + 479, + 472 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 130, + 401, + 479, + 472 + ], + "type": "text", + "content": " after denoising should keep most of the information in " + }, + { + "bbox": [ + 130, + 401, + 479, + 472 + ], + "type": "inline_equation", + "content": "x'" + }, + { + "bbox": [ + 130, + 401, + 479, + 472 + ], + "type": "text", + "content": " (i.e., minimize " + }, + { + "bbox": [ + 130, + 401, + 479, + 472 + ], + "type": "inline_equation", + "content": "||x - x'||_2^2" + }, + { + "bbox": [ + 130, + 401, + 479, + 472 + ], + "type": "text", + "content": ") but discard the high-frequency information (i.e., minimize " + }, + { + "bbox": [ + 130, + 401, + 479, + 472 + ], + "type": "inline_equation", + "content": "TV(x)" + }, + { + "bbox": [ + 130, + 401, + 479, + 472 + ], + "type": "text", + "content": "). Optimizing Eq. (9) is not trivial as total variation is non-differentiable. As a result, we use a sub-gradient descent:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 224, + 478, + 480, + 492 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 224, + 478, + 480, + 492 + ], + "spans": [ + { + "bbox": [ + 224, + 478, + 480, + 492 + ], + "type": "interline_equation", + "content": "x ^ {i + 1} = x ^ {i} - \\alpha \\cdot \\left(\\left(x ^ {i} - x ^ {\\prime}\\right) + \\lambda \\cdot g \\left(x ^ {i}\\right)\\right) \\tag {10}", + "image_path": "18b77f5a89c584c5ff11423a42f811f308cf936ced244ada7be2a89f67bbc2c3.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 497, + 479, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 497, + 479, + 521 + ], + "spans": [ + { + "bbox": [ + 130, + 497, + 479, + 521 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 497, + 479, + 521 + ], + "type": "inline_equation", + "content": "x^i" + }, + { + "bbox": [ + 130, + 497, + 479, + 521 + ], + "type": "text", + "content": " is the denoised image at the step " + }, + { + "bbox": [ + 130, + 497, + 479, + 521 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 497, + 479, + 521 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 497, + 479, + 521 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 130, + 497, + 479, + 521 + ], + "type": "text", + "content": " is the step size and " + }, + { + "bbox": [ + 130, + 497, + 479, + 521 + ], + "type": "inline_equation", + "content": "g(x^i)" + }, + { + "bbox": [ + 130, + 497, + 479, + 521 + ], + "type": "text", + "content": " is the image gradient of " + }, + { + "bbox": [ + 130, + 497, + 479, + 521 + ], + "type": "inline_equation", + "content": "x^i" + }, + { + "bbox": [ + 130, + 497, + 479, + 521 + ], + "type": "text", + "content": " [9]." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 521, + 480, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 480, + 593 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 480, + 593 + ], + "type": "text", + "content": "Simply applying the total variation denoising will also decrease the task-oriented performance as it also removes useful information in high-frequency domain. Since adversarial attacks targeting entropy models attempt to increase bit rate in background region as demonstrated in Sec. 5.2, we add a mask to Eq. (10) to force the denoising algorithm to only remove high-frequency information in non-object space, that is," + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 216, + 597, + 480, + 612 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 216, + 597, + 480, + 612 + ], + "spans": [ + { + "bbox": [ + 216, + 597, + 480, + 612 + ], + "type": "interline_equation", + "content": "x ^ {i + 1} = x ^ {i} - \\alpha \\cdot m \\cdot \\left(\\left(x ^ {i} - x ^ {\\prime}\\right) + \\lambda \\cdot g \\left(x ^ {i}\\right)\\right) \\tag {11}", + "image_path": "ef360e3c4478f37433fd8d93d904abb10770452ad725d4deec484be1dea32a52.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 130, + 617, + 479, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 479, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 479, + 665 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 617, + 479, + 665 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 130, + 617, + 479, + 665 + ], + "type": "text", + "content": " is the mask to control the denoising level. In practice, we interpolate the output of entropy model " + }, + { + "bbox": [ + 130, + 617, + 479, + 665 + ], + "type": "inline_equation", + "content": "P_Z(z)" + }, + { + "bbox": [ + 130, + 617, + 479, + 665 + ], + "type": "text", + "content": " as the soft mask. This is because the higher bit rate " + }, + { + "bbox": [ + 130, + 617, + 479, + 665 + ], + "type": "inline_equation", + "content": "-\\log_2P_Z(z)" + }, + { + "bbox": [ + 130, + 617, + 479, + 665 + ], + "type": "text", + "content": " corresponding to object area has a smaller value of " + }, + { + "bbox": [ + 130, + 617, + 479, + 665 + ], + "type": "inline_equation", + "content": "P_Z(z)" + }, + { + "bbox": [ + 130, + 617, + 479, + 665 + ], + "type": "text", + "content": ", making it naturally a soft mask to avoid smoothing the object region." + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 201, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 201, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 201, + 91, + 448, + 102 + ], + "type": "text", + "content": "Resilience of Entropy Model in Distributed Neural Networks" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 132, + 146, + 483, + 205 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 482, + 137 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 482, + 137 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 482, + 137 + ], + "type": "text", + "content": "Table 3: Compression and classification performance before and after defense. The average data size for clean ImageNet is " + }, + { + "bbox": [ + 132, + 114, + 482, + 137 + ], + "type": "inline_equation", + "content": "9.62 \\pm 1.70" + }, + { + "bbox": [ + 132, + 114, + 482, + 137 + ], + "type": "text", + "content": " KBytes." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 132, + 146, + 483, + 205 + ], + "lines": [ + { + "bbox": [ + 132, + 146, + 483, + 205 + ], + "spans": [ + { + "bbox": [ + 132, + 146, + 483, + 205 + ], + "type": "table", + "html": "
ε = 2/255ε = 4/255ε = 8/255ε = 16/255
Size[KB]Acc[%]Size[KB]Acc[%]Size[KB]Acc[%]Size[KB]Acc[%]
before12.35±0.6373.2614.11±0.3871.3316.15±0.2965.6218.79±0.3752.13
after8.76±1.3870.829.55±1.0770.1511.02±0.6567.2213.44±0.4859.87
", + "image_path": "7f72dae5f3f2da2c179f62d944675f625c1808786897677a5f915f32c6dd20b4.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 226, + 481, + 333 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 226, + 481, + 333 + ], + "spans": [ + { + "bbox": [ + 130, + 226, + 481, + 333 + ], + "type": "text", + "content": "Experimental Results. We investigate the proposed defense method as a function of perturbation budget " + }, + { + "bbox": [ + 130, + 226, + 481, + 333 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 130, + 226, + 481, + 333 + ], + "type": "text", + "content": ". As shown in Tab. 3, attackers targeting entropy significantly increase the data size, by up to " + }, + { + "bbox": [ + 130, + 226, + 481, + 333 + ], + "type": "inline_equation", + "content": "95.32\\%" + }, + { + "bbox": [ + 130, + 226, + 481, + 333 + ], + "type": "text", + "content": " when " + }, + { + "bbox": [ + 130, + 226, + 481, + 333 + ], + "type": "inline_equation", + "content": "\\epsilon = 16/255" + }, + { + "bbox": [ + 130, + 226, + 481, + 333 + ], + "type": "text", + "content": " compared to the clean images (9.62 KBytes). On the contrary, our denoising approach significantly reduce the data size for all " + }, + { + "bbox": [ + 130, + 226, + 481, + 333 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 130, + 226, + 481, + 333 + ], + "type": "text", + "content": ". For " + }, + { + "bbox": [ + 130, + 226, + 481, + 333 + ], + "type": "inline_equation", + "content": "\\epsilon = 2/255, 4/255" + }, + { + "bbox": [ + 130, + 226, + 481, + 333 + ], + "type": "text", + "content": ", the average data size is even " + }, + { + "bbox": [ + 130, + 226, + 481, + 333 + ], + "type": "inline_equation", + "content": "8.94\\%" + }, + { + "bbox": [ + 130, + 226, + 481, + 333 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 226, + 481, + 333 + ], + "type": "inline_equation", + "content": "0.73\\%" + }, + { + "bbox": [ + 130, + 226, + 481, + 333 + ], + "type": "text", + "content": " smaller than the clean images with only " + }, + { + "bbox": [ + 130, + 226, + 481, + 333 + ], + "type": "inline_equation", + "content": "2.44\\%" + }, + { + "bbox": [ + 130, + 226, + 481, + 333 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 226, + 481, + 333 + ], + "type": "inline_equation", + "content": "1.18\\%" + }, + { + "bbox": [ + 130, + 226, + 481, + 333 + ], + "type": "text", + "content": " accuracy loss, respectively. For " + }, + { + "bbox": [ + 130, + 226, + 481, + 333 + ], + "type": "inline_equation", + "content": "\\epsilon = 8/255, 16/255" + }, + { + "bbox": [ + 130, + 226, + 481, + 333 + ], + "type": "text", + "content": ", the data size is reduced by " + }, + { + "bbox": [ + 130, + 226, + 481, + 333 + ], + "type": "inline_equation", + "content": "31.80\\%" + }, + { + "bbox": [ + 130, + 226, + 481, + 333 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 226, + 481, + 333 + ], + "type": "inline_equation", + "content": "28.47\\%" + }, + { + "bbox": [ + 130, + 226, + 481, + 333 + ], + "type": "text", + "content": " respectively, while the accuracy improved slightly after denoising (" + }, + { + "bbox": [ + 130, + 226, + 481, + 333 + ], + "type": "inline_equation", + "content": "1.60\\%" + }, + { + "bbox": [ + 130, + 226, + 481, + 333 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 130, + 226, + 481, + 333 + ], + "type": "inline_equation", + "content": "\\epsilon = 8/255" + }, + { + "bbox": [ + 130, + 226, + 481, + 333 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 226, + 481, + 333 + ], + "type": "inline_equation", + "content": "7.74\\%" + }, + { + "bbox": [ + 130, + 226, + 481, + 333 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 130, + 226, + 481, + 333 + ], + "type": "inline_equation", + "content": "\\epsilon = 16/255" + }, + { + "bbox": [ + 130, + 226, + 481, + 333 + ], + "type": "text", + "content": ")." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 334, + 481, + 382 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 334, + 481, + 382 + ], + "spans": [ + { + "bbox": [ + 130, + 334, + 481, + 382 + ], + "type": "text", + "content": "We remark that the proposed defense is a standalone approach which can be incorporated into other approaches such as adversarial training to further improve the resilience. The accuracy loss incurred in small perturbation scenarios " + }, + { + "bbox": [ + 130, + 334, + 481, + 382 + ], + "type": "inline_equation", + "content": "(\\epsilon = 2 / 255, 4 / 255)" + }, + { + "bbox": [ + 130, + 334, + 481, + 382 + ], + "type": "text", + "content": " can be mitigated by training with augmented data." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 398, + 248, + 410 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 398, + 248, + 410 + ], + "spans": [ + { + "bbox": [ + 132, + 398, + 248, + 410 + ], + "type": "text", + "content": "5.4 Adaptive Attacks" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 416, + 481, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 416, + 481, + 464 + ], + "spans": [ + { + "bbox": [ + 130, + 416, + 481, + 464 + ], + "type": "text", + "content": "To thoroughly evaluate the effectiveness of a defense strategy, [10] proposed to design adaptive attacks that counter the defense mechanism with perfect knowledge of both DNN and defense in the white-box setting. To this end, we assess our defense with two adaptive attacks." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 481, + 481, + 554 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 481, + 481, + 554 + ], + "spans": [ + { + "bbox": [ + 130, + 481, + 481, + 554 + ], + "type": "text", + "content": "Low Frequency Attack. As the total variation denoising is proposed to remove high-frequency components, the first adaptive attack that we considered is to only add perturbations in low frequency space. Following the methodology outlined in [21], we first convert the gradient to frequency space with discrete cosine transform and mask out the high-frequency components. The gradient is then transformed back and Eq. (5) is applied to create the adversarial samples." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": "Regional Attack. Given that the defense exclusively denoises the background region, the second adaptive attack we considered is to force the adversary put greater effort in increasing bit rate in the object region. Similar to Eq. (11), a mask is multiplied to the loss function before backpropagation. In practice, we choose " + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "inline_equation", + "content": "1 - P_Z(z)" + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": " as the soft mask where " + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "inline_equation", + "content": "P_Z(z)" + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": " is the probability of " + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": " ranging between 0 and 1. A larger " + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "inline_equation", + "content": "P_Z(z)" + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": " indicates a smaller bit rate " + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "inline_equation", + "content": "-\\log_2P_Z(z)" + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": " corresponding to the background region. Therefore, " + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "inline_equation", + "content": "1 - P_Z(z)" + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": " encourages the adversary to take more attention to object region." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 231, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 231, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 231, + 102 + ], + "type": "text", + "content": "M. Zhang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 208, + 136, + 403, + 193 + ], + "blocks": [ + { + "bbox": [ + 192, + 114, + 421, + 125 + ], + "lines": [ + { + "bbox": [ + 192, + 114, + 421, + 125 + ], + "spans": [ + { + "bbox": [ + 192, + 114, + 421, + 125 + ], + "type": "text", + "content": "Table 4: Resilience to adaptive attacks with " + }, + { + "bbox": [ + 192, + 114, + 421, + 125 + ], + "type": "inline_equation", + "content": "\\epsilon = 4 / 255" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 208, + 136, + 403, + 193 + ], + "lines": [ + { + "bbox": [ + 208, + 136, + 403, + 193 + ], + "spans": [ + { + "bbox": [ + 208, + 136, + 403, + 193 + ], + "type": "table", + "html": "
Adaptive AttacksBefore DenoisingAfter Denoising
Size[KB]Acc[%]Size[KB]Acc[%]
Frequency14.11±0.3871.389.56±1.0770.17
Regional14.05±0.4171.639.50±1.1070.22
", + "image_path": "9fe1f34475416dab50c98311d15694be740c94249ca47763259021abb6e9ab3b.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 213, + 482, + 275 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 213, + 482, + 275 + ], + "spans": [ + { + "bbox": [ + 130, + 213, + 482, + 275 + ], + "type": "text", + "content": "Resilience to Adaptive Attacks. Tab. 4 shows results of adaptive attacks with perturbation budget " + }, + { + "bbox": [ + 130, + 213, + 482, + 275 + ], + "type": "inline_equation", + "content": "\\epsilon = 4 / 255" + }, + { + "bbox": [ + 130, + 213, + 482, + 275 + ], + "type": "text", + "content": ". Our defense successfully reduces the average data size by " + }, + { + "bbox": [ + 130, + 213, + 482, + 275 + ], + "type": "inline_equation", + "content": "67.75\\%" + }, + { + "bbox": [ + 130, + 213, + 482, + 275 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 213, + 482, + 275 + ], + "type": "inline_equation", + "content": "67.62\\%" + }, + { + "bbox": [ + 130, + 213, + 482, + 275 + ], + "type": "text", + "content": " for low frequency attack and regional attack respectively while the accuracy only decreases " + }, + { + "bbox": [ + 130, + 213, + 482, + 275 + ], + "type": "inline_equation", + "content": "1.21\\%" + }, + { + "bbox": [ + 130, + 213, + 482, + 275 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 213, + 482, + 275 + ], + "type": "inline_equation", + "content": "1.41\\%" + }, + { + "bbox": [ + 130, + 213, + 482, + 275 + ], + "type": "text", + "content": " respectively. Thus, the proposed approach is resilient to adaptive attacks." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 290, + 237, + 304 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 290, + 237, + 304 + ], + "spans": [ + { + "bbox": [ + 131, + 290, + 237, + 304 + ], + "type": "text", + "content": "6 Related Work" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 314, + 481, + 386 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 314, + 481, + 386 + ], + "spans": [ + { + "bbox": [ + 130, + 314, + 481, + 386 + ], + "type": "text", + "content": "Distributed Deep Neural Network. Distributed DNN is proposed to meet the challenge of deploying artificial intelligence to resource-constrained platforms such as mobile devices and Internet of Things (IoT) devices [41]. To minimize the end-to-end latency across devices, dimension reduction designs (i.e., bottlenecks) are introduced to compress the data size of intermediate representations that need to be sent [18, 54]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 386, + 481, + 470 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 386, + 481, + 470 + ], + "spans": [ + { + "bbox": [ + 130, + 386, + 481, + 470 + ], + "type": "text", + "content": "However, unlike autoencoders [32,36,49] that have symmetric designs on both encoder and decoder sides, distributed DNNs usually have asymmetric designs due to the limit of computation resources, and thus inject the bottleneck in early layers [18, 38-40, 54]. As a result, naively end-to-end trained distributed DNNs have noticeable performance loss [18, 54]. To preserve accuracy, [38] proposed to use knowledge distillation to train the distributed DNN separately while [39] proposed a multi-stage training approach for each part of the DNN." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 470, + 481, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 470, + 481, + 542 + ], + "spans": [ + { + "bbox": [ + 130, + 470, + 481, + 542 + ], + "type": "text", + "content": "Along with bottleneck-based distributed DNNs, different coding-based approaches are also applied to reduce the data size of the latent representations. [2] proposed to apply codec-based compression such as JPEG to latent representations and [15] adopted spatio-temporal coding such as HEVC to compress the streaming latent features in video tasks. [42,55,59] use entropy coding to achieve a higher compression ratio of latent representations." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "text", + "content": "Entropy Coding in Deep Data Compression. [5] first introduced a fully factorized prior to integrate entropy coding with variational autoencoder, showcasing superior quality enhancement over conventional codec-based methods like JPEG for image compression. Building upon this, [6] extended the conventional factorized prior model to a learnable hierarchical hyper prior, resulting in improved performance. [43] proposed a joint auto-regressive and hyper prior design while [44] proposed a channel-aware entropy coding scheme. In addition, [14] introduced the attention mechanism to prior models and [1] extended the technique to video compression." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 200, + 90, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 200, + 90, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 200, + 90, + 447, + 102 + ], + "type": "text", + "content": "Resilience of Entropy Model in Distributed Neural Networks" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "content": "To minimize the communication overhead in distributed DNN scenarios, [55] for the first time adopted the entropy coding to further compress the latent representations after the bottleneck while [59] extended the approach to object detection. However, these approaches overlooked resource constraints, resulting in over-complex network designs for mobile devices. Conversely, [42] integrated entropy coding with a multi-stage training strategy [39], aiming to optimize a tripartite trade-off encompassing computation resources in mobile devices, communication overhead between mobile and edge devices, and end-to-end performance of neural networks." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 240, + 482, + 372 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 240, + 482, + 372 + ], + "spans": [ + { + "bbox": [ + 130, + 240, + 482, + 372 + ], + "type": "text", + "content": "Resilience of DNN efficiency. While the reliability of DNNs have been extensively investigated, limited attention has been devoted to interference that undermine their efficiency. [22] explored this direction by demonstrating that imperceptible perturbations to input, leveraging intermediate representations of DNNs, could nullify the computation savings achieved by dynamic depth neural networks, which adapt their depth based on input complexity. [46] extended these attacks to target both dynamic depth and dynamic width neural networks. Moreover, in [45], the authors proposed simultaneous adjustments to the direction and magnitude of attacks, enhancing their effectiveness. [27] revealed that early exit dynamic DNNs could be tricked into late-stage inferences, significantly slowing down inference speed." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 373, + 482, + 444 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 373, + 482, + 444 + ], + "spans": [ + { + "bbox": [ + 130, + 373, + 482, + 444 + ], + "type": "text", + "content": "The study of efficiency vulnerabilities in DNNs has also ventured into real-world deployment scenarios. [33] focused on attacking LiDAR-based detection systems, introducing latency in detections and exposing vulnerabilities in critical contexts such as self-driving cars. [13] demonstrated that neural image caption generation models could be manipulated to incur increased computation costs by inducing unnecessary decoder calls for token generation." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 444, + 482, + 492 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 444, + 482, + 492 + ], + "spans": [ + { + "bbox": [ + 130, + 444, + 482, + 492 + ], + "type": "text", + "content": "In star contrast to existing research, our work investigates a novel threat emerging in distributed DNNs. Here, adversaries not only compromise the communication efficiency of entropy-coded distributed DNNs but also pose a threat to other users by saturating the transmission bandwidth." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 510, + 220, + 522 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 510, + 220, + 522 + ], + "spans": [ + { + "bbox": [ + 132, + 510, + 220, + 522 + ], + "type": "text", + "content": "7 Conclusion" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 533, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 482, + 665 + ], + "type": "text", + "content": "This paper has investigated the resilience of entropy models in distributed DNNs against both intentional and unintentional interference. We conducted thorough evaluations using 3 different DNN architectures, 2 entropy model designs, and 4 rate-distortion trade-off factors with common corruption datasets and adversarial attacks. Our analysis disentangled compression features in both spatial and frequency domains, revealing vulnerabilities of the entropy model to specific types of perturbations. Building on these findings, we proposed a standalone defense strategy aimed at reducing data size with minimal task-oriented performance loss. Our future work will focus on designing more advanced defense approaches for distributed DNNs that are resilient in both compression and classification tasks." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 231, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 231, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 231, + 102 + ], + "type": "text", + "content": "M. Zhang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 137, + 482, + 209 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 137, + 482, + 209 + ], + "spans": [ + { + "bbox": [ + 130, + 137, + 482, + 209 + ], + "type": "text", + "content": "This work is funded in part by the Air Force Office of Scientific Research under contract number FA9550-23-1-0261, by the Office of Naval Research under contract number N00014-23-1-2221, as well as by National Science Foundation grants CNS-2134973 and CNS-2312875. Any opinions, findings and conclusions or recommendations expressed in this material are those of the authors and do not necessarily reflect the views of the U.S. Government." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 224, + 197, + 237 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 224, + 197, + 237 + ], + "spans": [ + { + "bbox": [ + 132, + 224, + 197, + 237 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 245, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 138, + 245, + 481, + 290 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 245, + 481, + 290 + ], + "spans": [ + { + "bbox": [ + 138, + 245, + 481, + 290 + ], + "type": "text", + "content": "1. Agustsson, E., Minnen, D., Johnston, N., Balle, J., Hwang, S.J., Toderici, G.: Scale-space flow for end-to-end optimized video compression. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8503-8512 (2020)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 289, + 481, + 311 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 289, + 481, + 311 + ], + "spans": [ + { + "bbox": [ + 138, + 289, + 481, + 311 + ], + "type": "text", + "content": "2. Alvar, S.R., Bajic, I.V.: Pareto-optimal bit allocation for collaborative intelligence. IEEE Transactions on Image Processing 30, 3348-3361 (2021)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 311, + 481, + 343 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 311, + 481, + 343 + ], + "spans": [ + { + "bbox": [ + 138, + 311, + 481, + 343 + ], + "type": "text", + "content": "3. Andriushchenko, M., Croce, F., Flammarion, N., Hein, M.: Square attack: a query-efficient black-box adversarial attack via random search. In: European conference on computer vision. pp. 484-501. Springer (2020)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 343, + 481, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 343, + 481, + 374 + ], + "spans": [ + { + "bbox": [ + 138, + 343, + 481, + 374 + ], + "type": "text", + "content": "4. Arvinte, M., Cornelius, C., Martin, J., Himayat, N.: Investigating the adversarial robustness of density estimation using the probability flow ode. arXiv preprint arXiv:2310.07084 (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 374, + 481, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 374, + 481, + 397 + ], + "spans": [ + { + "bbox": [ + 138, + 374, + 481, + 397 + ], + "type": "text", + "content": "5. Balle, J., Laparra, V., Simoncelli, E.P.: End-to-end optimized image compression. In: 5th International Conference on Learning Representations, ICLR 2017 (2017)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 397, + 481, + 429 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 397, + 481, + 429 + ], + "spans": [ + { + "bbox": [ + 138, + 397, + 481, + 429 + ], + "type": "text", + "content": "6. Balle, J., Minnen, D., Singh, S., Hwang, S.J., Johnston, N.: Variational image compression with a scale hyperprior. In: International Conference on Learning Representations (2018)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 429, + 481, + 472 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 429, + 481, + 472 + ], + "spans": [ + { + "bbox": [ + 138, + 429, + 481, + 472 + ], + "type": "text", + "content": "7. Barbu, A., Mayo, D., Alverio, J., Luo, W., Wang, C., Gutfreund, D., Tenenbaum, J., Katz, B.: Objectnet: A large-scale bias-controlled dataset for pushing the limits of object recognition models. Advances in neural information processing systems 32 (2019)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 472, + 481, + 504 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 472, + 481, + 504 + ], + "spans": [ + { + "bbox": [ + 138, + 472, + 481, + 504 + ], + "type": "text", + "content": "8. Bau, D., Zhou, B., Khosla, A., Oliva, A., Torralba, A.: Network dissection: Quantifying interpretability of deep visual representations. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 6541-6549 (2017)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 504, + 481, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 504, + 481, + 536 + ], + "spans": [ + { + "bbox": [ + 138, + 504, + 481, + 536 + ], + "type": "text", + "content": "9. Boigné, E., Parkinson, D.Y., Ihme, M.: Towards data-informed motion artifact reduction in quantitative ct using piecewise linear interpolation. IEEE Transactions on Computational Imaging 8, 917-932 (2022)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 536, + 481, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 536, + 481, + 568 + ], + "spans": [ + { + "bbox": [ + 138, + 536, + 481, + 568 + ], + "type": "text", + "content": "10. Carlini, N., Wagner, D.: Adversarial examples are not easily detected: Bypassing ten detection methods. In: Proceedings of the 10th ACM workshop on artificial intelligence and security. pp. 3-14 (2017)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 568, + 481, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 568, + 481, + 590 + ], + "spans": [ + { + "bbox": [ + 138, + 568, + 481, + 590 + ], + "type": "text", + "content": "1. Carlini, N., Wagner, D.: Towards evaluating the robustness of neural networks. In: 2017 ieee symposium on security and privacy (sp). pp. 39-57. IEEE (2017)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 590, + 481, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 590, + 481, + 622 + ], + "spans": [ + { + "bbox": [ + 138, + 590, + 481, + 622 + ], + "type": "text", + "content": "2. Chen, J., Jordan, M.I., Wainwright, M.J.: Hopskipjumpattack: A query-efficient decision-based attack. In: 2020 ieee symposium on security and privacy (sp). pp. 1277-1294. IEEE (2020)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 622, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 622, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 622, + 481, + 665 + ], + "type": "text", + "content": "3. Chen, S., Song, Z., Haque, M., Liu, C., Yang, W.: Nicgslowdown: Evaluating the efficiency robustness of neural image caption generation models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15365-15374 (2022)" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 200, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 200, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 200, + 91, + 447, + 102 + ], + "type": "text", + "content": "Resilience of Entropy Model in Distributed Neural Networks" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 666 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 481, + 160 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 481, + 160 + ], + "type": "text", + "content": "14. Cheng, Z., Sun, H., Takeuchi, M., Katto, J.: Learned image compression with discretized gaussian mixture likelihoods and attention modules. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 7939-7948 (2020)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 160, + 481, + 192 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 160, + 481, + 192 + ], + "spans": [ + { + "bbox": [ + 133, + 160, + 481, + 192 + ], + "type": "text", + "content": "15. Choi, H., Bajic, I.V.: Deep feature compression for collaborative object detection. In: 2018 25th IEEE International Conference on Image Processing (ICIP). pp. 3743-3747. IEEE (2018)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 192, + 481, + 225 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 192, + 481, + 225 + ], + "spans": [ + { + "bbox": [ + 133, + 192, + 481, + 225 + ], + "type": "text", + "content": "16. Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: Imagenet: A large-scale hierarchical image database. In: 2009 IEEE conference on computer vision and pattern recognition. pp. 248-255. IEEE (2009)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 225, + 481, + 257 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 225, + 481, + 257 + ], + "spans": [ + { + "bbox": [ + 132, + 225, + 481, + 257 + ], + "type": "text", + "content": "17. Duda, J.: Asymmetric numeral systems: entropy coding combining speed of huffman coding with compression rate of arithmetic coding. arXiv preprint arXiv:1311.2540 (2013)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 257, + 481, + 290 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 257, + 481, + 290 + ], + "spans": [ + { + "bbox": [ + 132, + 257, + 481, + 290 + ], + "type": "text", + "content": "18. Eshratifar, A.E., Esmaili, A., Pedram, M.: Bottlenet: A deep learning architecture for intelligent mobile cloud computing services. In: 2019 IEEE/ACM International Symposium on Low Power Electronics and Design (ISLPED). pp. 1-6. IEEE (2019)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 290, + 481, + 332 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 290, + 481, + 332 + ], + "spans": [ + { + "bbox": [ + 132, + 290, + 481, + 332 + ], + "type": "text", + "content": "19. Geirhos, R., Rubisch, P., Michaelis, C., Bethge, M., Wichmann, F.A., Brendel, W.: Imagenet-trained cnns are biased towards texture; increasing shape bias improves accuracy and robustness. In: International Conference on Learning Representations (2018)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 332, + 481, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 332, + 481, + 354 + ], + "spans": [ + { + "bbox": [ + 132, + 332, + 481, + 354 + ], + "type": "text", + "content": "20. Goodfellow, I.J., Shlens, J., Szegedy, C.: Explaining and harnessing adversarial examples. arXiv preprint arXiv:1412.6572 (2014)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 354, + 481, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 354, + 481, + 374 + ], + "spans": [ + { + "bbox": [ + 132, + 354, + 481, + 374 + ], + "type": "text", + "content": "21. Guo, C., Frank, J.S., Weinberger, K.Q.: Low frequency adversarial perturbation. arXiv preprint arXiv:1809.08758 (2018)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 374, + 481, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 374, + 481, + 407 + ], + "spans": [ + { + "bbox": [ + 132, + 374, + 481, + 407 + ], + "type": "text", + "content": "22. Haque, M., Chauhan, A., Liu, C., Yang, W.: Ilfo: Adversarial attack on adaptive neural networks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14264-14273 (2020)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 407, + 481, + 439 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 407, + 481, + 439 + ], + "spans": [ + { + "bbox": [ + 132, + 407, + 481, + 439 + ], + "type": "text", + "content": "23. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 770-778 (2016)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 439, + 481, + 483 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 439, + 481, + 483 + ], + "spans": [ + { + "bbox": [ + 132, + 439, + 481, + 483 + ], + "type": "text", + "content": "24. Hendrycks, D., Basart, S., Mu, N., Kadavath, S., Wang, F., Dorundo, E., Desai, R., Zhu, T., Parajuli, S., Guo, M., et al.: The many faces of robustness: A critical analysis of out-of-distribution generalization. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 8340-8349 (2021)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 483, + 481, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 483, + 481, + 514 + ], + "spans": [ + { + "bbox": [ + 132, + 483, + 481, + 514 + ], + "type": "text", + "content": "25. Hendrycks, D., Dietterich, T.: Benchmarking neural network robustness to common corruptions and perturbations. In: International Conference on Learning Representations (2018)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 514, + 481, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 514, + 481, + 536 + ], + "spans": [ + { + "bbox": [ + 132, + 514, + 481, + 536 + ], + "type": "text", + "content": "26. Hinton, G., Vinyals, O., Dean, J.: Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531 (2015)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 536, + 481, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 536, + 481, + 567 + ], + "spans": [ + { + "bbox": [ + 132, + 536, + 481, + 567 + ], + "type": "text", + "content": "27. Hong, S., Kaya, Y., Modoranu, I.V., Dumitraş, T.: A panda? no, it's a sloth: Slowdown attacks on adaptive multi-exit neural network inference. arXiv preprint arXiv:2010.02432 (2020)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 568, + 481, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 568, + 481, + 601 + ], + "spans": [ + { + "bbox": [ + 132, + 568, + 481, + 601 + ], + "type": "text", + "content": "28. Howard, A., Sandler, M., Chu, G., Chen, L.C., Chen, B., Tan, M., Wang, W., Zhu, Y., Pang, R., Vasudevan, V., et al.: Searching for mobilenetv3. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 1314-1324 (2019)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 601, + 481, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 601, + 481, + 633 + ], + "spans": [ + { + "bbox": [ + 132, + 601, + 481, + 633 + ], + "type": "text", + "content": "29. Howard, A.G., Zhu, M., Chen, B., Kalenichenko, D., Wang, W., Weyand, T., Andreetto, M., Adam, H.: Mobilenets: Efficient convolutional neural networks for mobile vision applications. arXiv preprint arXiv:1704.04861 (2017)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 633, + 481, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 633, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 633, + 481, + 666 + ], + "type": "text", + "content": "30. Ilyas, A., Engstrom, L., Athalye, A., Lin, J.: Black-box adversarial attacks with limited queries and information. In: International conference on machine learning. pp. 2137-2146. PMLR (2018)" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 231, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 231, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 231, + 102 + ], + "type": "text", + "content": "M. Zhang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 665 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 161 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 161 + ], + "type": "text", + "content": "31. Jacob, B., Kligys, S., Chen, B., Zhu, M., Tang, M., Howard, A., Adam, H., Kalenichenko, D.: Quantization and training of neural networks for efficient integer-arithmetic-only inference. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 2704-2713 (2018)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 162, + 481, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 162, + 481, + 184 + ], + "spans": [ + { + "bbox": [ + 130, + 162, + 481, + 184 + ], + "type": "text", + "content": "32. Kingma, D.P., Welling, M.: Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114 (2013)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 185, + 481, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 185, + 481, + 228 + ], + "spans": [ + { + "bbox": [ + 132, + 185, + 481, + 228 + ], + "type": "text", + "content": "33. Liu, H., Wu, Y., Yu, Z., Vorobeychik, Y., Zhang, N.: Slowlidar: Increasing the latency of lidar-based detection using adversarial examples. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5146-5155 (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 229, + 481, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 229, + 481, + 262 + ], + "spans": [ + { + "bbox": [ + 132, + 229, + 481, + 262 + ], + "type": "text", + "content": "34. Liu, P., Qi, B., Banerjee, S.: Edgeeye: An edge service framework for real-time intelligent video analytics. In: Proceedings of the 1st international workshop on edge systems, analytics and networking. pp. 1-6 (2018)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 263, + 481, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 263, + 481, + 296 + ], + "spans": [ + { + "bbox": [ + 132, + 263, + 481, + 296 + ], + "type": "text", + "content": "35. Madry, A., Makelov, A., Schmidt, L., Tsipras, D., Vladu, A.: Towards deep learning models resistant to adversarial attacks. In: International Conference on Learning Representations (2018)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 297, + 481, + 351 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 297, + 481, + 351 + ], + "spans": [ + { + "bbox": [ + 132, + 297, + 481, + 351 + ], + "type": "text", + "content": "36. Masci, J., Meier, U., Ciresan, D., Schmidhuber, J.: Stacked convolutional autoencoders for hierarchical feature extraction. In: Artificial Neural Networks and Machine Learning-ICANN 2011: 21st International Conference on Artificial Neural Networks, Espoo, Finland, June 14-17, 2011, Proceedings, Part I 21. pp. 52-59. Springer (2011)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 353, + 481, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 353, + 481, + 374 + ], + "spans": [ + { + "bbox": [ + 132, + 353, + 481, + 374 + ], + "type": "text", + "content": "37. Matsubara, Y., Yang, R., Levorato, M., Mandt, S.: Sc2 benchmark: Supervised compression for split computing. Transactions on machine learning research (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 376, + 481, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 376, + 481, + 407 + ], + "spans": [ + { + "bbox": [ + 132, + 376, + 481, + 407 + ], + "type": "text", + "content": "38. Matsubara, Y., Callegaro, D., Baidya, S., Levorato, M., Singh, S.: Head network distillation: Splitting distilled deep neural networks for resource-constrained edge computing systems. IEEE Access 8, 212177-212193 (2020)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 409, + 481, + 462 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 409, + 481, + 462 + ], + "spans": [ + { + "bbox": [ + 132, + 409, + 481, + 462 + ], + "type": "text", + "content": "39. Matsubara, Y., Callegaro, D., Singh, S., Levorato, M., Restuccia, F.: Bottlefit: Learning compressed representations in deep neural networks for effective and efficient split computing. In: 2022 IEEE 23rd International Symposium on a World of Wireless, Mobile and Multimedia Networks (WoWMoM). pp. 337-346. IEEE (2022)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 464, + 481, + 497 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 464, + 481, + 497 + ], + "spans": [ + { + "bbox": [ + 132, + 464, + 481, + 497 + ], + "type": "text", + "content": "40. Matsubara, Y., Levorato, M.: Neural compression and filtering for edge-assisted real-time object detection in challenged networks. In: 2020 25th International Conference on Pattern Recognition (ICPR). pp. 2272-2279. IEEE (2021)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 498, + 481, + 530 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 498, + 481, + 530 + ], + "spans": [ + { + "bbox": [ + 132, + 498, + 481, + 530 + ], + "type": "text", + "content": "41. Matsubara, Y., Levorato, M., Restuccia, F.: Split computing and early exiting for deep learning applications: Survey and research challenges. ACM Computing Surveys 55(5), 1-30 (2022)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 532, + 481, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 532, + 481, + 564 + ], + "spans": [ + { + "bbox": [ + 132, + 532, + 481, + 564 + ], + "type": "text", + "content": "42. Matsubara, Y., Yang, R., Levorato, M., Mandt, S.: Supervised compression for resource-constrained edge computing systems. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 2685-2695 (2022)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 566, + 481, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 566, + 481, + 597 + ], + "spans": [ + { + "bbox": [ + 132, + 566, + 481, + 597 + ], + "type": "text", + "content": "43. Minnen, D., Balle, J., Toderici, G.D.: Joint autoregressive and hierarchical priors for learned image compression. Advances in neural information processing systems 31 (2018)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 599, + 481, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 599, + 481, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 599, + 481, + 632 + ], + "type": "text", + "content": "44. Minnen, D., Singh, S.: Channel-wise autoregressive entropy models for learned image compression. In: 2020 IEEE International Conference on Image Processing (ICIP). pp. 3339-3343. IEEE (2020)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 633, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 633, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 633, + 481, + 665 + ], + "type": "text", + "content": "45. Pan, J., Foo, L.G., Zheng, Q., Fan, Z., Rahmani, H., Ke, Q., Liu, J.: Gradmdm: Adversarial attack on dynamic networks. IEEE Transactions on Pattern Analysis and Machine Intelligence (2023)" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 200, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 200, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 200, + 91, + 447, + 102 + ], + "type": "text", + "content": "Resilience of Entropy Model in Distributed Neural Networks" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 117, + 480, + 610 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 133, + 117, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 117, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 117, + 480, + 149 + ], + "type": "text", + "content": "46. Pan, J., Zheng, Q., Fan, Z., Rahmani, H., Ke, Q., Liu, J.: Gradauto: Energy-oriented attack on dynamic neural networks. In: European Conference on Computer Vision. pp. 637-653. Springer (2022)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 150, + 480, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 150, + 480, + 183 + ], + "spans": [ + { + "bbox": [ + 133, + 150, + 480, + 183 + ], + "type": "text", + "content": "47. Radosavovic, I., Kosaraju, R.P., Girshick, R., He, K., Dóllár, P.: Designing network design spaces. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10428-10436 (2020)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 183, + 480, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 183, + 480, + 215 + ], + "spans": [ + { + "bbox": [ + 132, + 183, + 480, + 215 + ], + "type": "text", + "content": "48. Ran, X., Chen, H., Zhu, X., Liu, Z., Chen, J.: Deepdecision: A mobile deep learning framework for edge video analytics. In: IEEE INFOCOM 2018-IEEE conference on computer communications. pp. 1421-1429. IEEE (2018)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 216, + 480, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 216, + 480, + 258 + ], + "spans": [ + { + "bbox": [ + 132, + 216, + 480, + 258 + ], + "type": "text", + "content": "49. Rifai, S., Vincent, P., Muller, X., Glorot, X., Bengio, Y.: Contractive auto-encoders: Explicit invariance during feature extraction. In: Proceedings of the 28th international conference on international conference on machine learning. pp. 833-840 (2011)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 259, + 480, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 259, + 480, + 281 + ], + "spans": [ + { + "bbox": [ + 132, + 259, + 480, + 281 + ], + "type": "text", + "content": "50. Rissanen, J., Langdon, G.: Universal modeling and coding. IEEE Transactions on Information Theory 27(1), 12-23 (1981)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 282, + 480, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 282, + 480, + 304 + ], + "spans": [ + { + "bbox": [ + 132, + 282, + 480, + 304 + ], + "type": "text", + "content": "51. Rudin, L.I., Osher, S., Fatemi, E.: Nonlinear total variation based noise removal algorithms. Physica D: nonlinear phenomena 60(1-4), 259-268 (1992)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 304, + 480, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 304, + 480, + 335 + ], + "spans": [ + { + "bbox": [ + 132, + 304, + 480, + 335 + ], + "type": "text", + "content": "52. Sandler, M., Howard, A., Zhu, M., Zhmoginov, A., Chen, L.C.: Mobilenetv2: Inverted residuals and linear bottlenecks. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 4510-4520 (2018)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 336, + 480, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 336, + 480, + 357 + ], + "spans": [ + { + "bbox": [ + 132, + 336, + 480, + 357 + ], + "type": "text", + "content": "53. Shannon, C.E.: A mathematical theory of communication. The Bell system technical journal 27(3), 379-423 (1948)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 358, + 480, + 390 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 358, + 480, + 390 + ], + "spans": [ + { + "bbox": [ + 132, + 358, + 480, + 390 + ], + "type": "text", + "content": "54. Shao, J., Zhang, J.: Bottlenet++: An end-to-end approach for feature compression in device-edge co-inference systems. In: 2020 IEEE International Conference on Communications Workshops (ICC Workshops). pp. 1-6. IEEE (2020)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 391, + 480, + 423 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 391, + 480, + 423 + ], + "spans": [ + { + "bbox": [ + 132, + 391, + 480, + 423 + ], + "type": "text", + "content": "55. Singh, S., Abu-El-Haija, S., Johnston, N., Balle, J., Shrivastava, A., Toderici, G.: End-to-end learning of compressible features. In: 2020 IEEE International Conference on Image Processing (ICIP). pp. 3349-3353. IEEE (2020)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 424, + 480, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 424, + 480, + 456 + ], + "spans": [ + { + "bbox": [ + 132, + 424, + 480, + 456 + ], + "type": "text", + "content": "56. Szegedy, C., Zaremba, W., Sutskever, I., Bruna, J., Erhan, D., Goodfellow, I., Fergus, R.: Intriguing properties of neural networks. In: 2nd International Conference on Learning Representations, ICLR 2014 (2014)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 457, + 480, + 500 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 457, + 480, + 500 + ], + "spans": [ + { + "bbox": [ + 132, + 457, + 480, + 500 + ], + "type": "text", + "content": "57. Tan, M., Chen, B., Pang, R., Vasudevan, V., Sandler, M., Howard, A., Le, Q.V.: Mnasnet: Platform-aware neural architecture search for mobile. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 2820-2828 (2019)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 501, + 480, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 501, + 480, + 533 + ], + "spans": [ + { + "bbox": [ + 132, + 501, + 480, + 533 + ], + "type": "text", + "content": "58. Wang, H., Ge, S., Lipton, Z., Xing, E.P.: Learning robust global representations by penalizing local predictive power. Advances in Neural Information Processing Systems 32 (2019)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 533, + 480, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 533, + 480, + 577 + ], + "spans": [ + { + "bbox": [ + 132, + 533, + 480, + 577 + ], + "type": "text", + "content": "59. Yuan, Z., Rawlekar, S., Garg, S., Erkip, E., Wang, Y.: Feature compression for rate constrained object detection on the edge. In: 2022 IEEE 5th International Conference on Multimedia Information Processing and Retrieval (MIPR). pp. 1-6. IEEE (2022)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 577, + 480, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 577, + 480, + 610 + ], + "spans": [ + { + "bbox": [ + 132, + 577, + 480, + 610 + ], + "type": "text", + "content": "60. Zhang, W., Zhang, Z., Zeadally, S., Chao, H.C., Leung, V.C.: Masm: A multiple-algorithm service model for energy-delay optimization in edge artificial intelligence. IEEE Transactions on Industrial Informatics 15(7), 4216-4224 (2019)" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 231, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 231, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 231, + 102 + ], + "type": "text", + "content": "M. Zhang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Resolving Scale Ambiguity in Multi-view 3D Reconstruction using Dual-Pixel Sensors/518914c5-d998-41c3-963e-0840d0e1f2c8_content_list.json b/2024/Resolving Scale Ambiguity in Multi-view 3D Reconstruction using Dual-Pixel Sensors/518914c5-d998-41c3-963e-0840d0e1f2c8_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..0903056c3827dae7d62678fdd9b1ad01973fc20e --- /dev/null +++ b/2024/Resolving Scale Ambiguity in Multi-view 3D Reconstruction using Dual-Pixel Sensors/518914c5-d998-41c3-963e-0840d0e1f2c8_content_list.json @@ -0,0 +1,1900 @@ +[ + { + "type": "text", + "text": "Resolving Scale Ambiguity in Multi-view 3D Reconstruction using Dual-Pixel Sensors", + "text_level": 1, + "bbox": [ + 246, + 140, + 754, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Kohei Ashida, Hiroaki Santo, Fumio Okura, and Yasuyuki Matsushita", + "bbox": [ + 225, + 212, + 774, + 227 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Graduate School of Information Science and Technology,", + "bbox": [ + 310, + 239, + 691, + 253 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Osaka University, Japan", + "bbox": [ + 419, + 255, + 583, + 267 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{kohei.ashida, santo.hiroaki, okura, yasumat}@ist.osaka-u.ac.jp", + "bbox": [ + 256, + 268, + 745, + 282 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Multi-view 3D reconstruction, namely structure-from-motion and multi-view stereo, is an essential component in 3D computer vision. In general, multi-view 3D reconstruction suffers from unknown scale ambiguity unless a reference object of known size is recorded together with the scene, or the camera poses are pre-calibrated. In this paper, we show that multi-view images recorded by a dual-pixel (DP) sensor allow us to automatically resolve the scale ambiguity without requiring a reference object or pre-calibration. Specifically, the observed defocus blurs in DP images provide sufficient information for determining the scale when paired together with the depth maps (up to scale) recovered from the multi-view 3D reconstruction. Based on this observation, we develop a simple yet effective linear solution method to determine the absolute scale in multi-view 3D reconstruction. Experiments demonstrate the effectiveness of the proposed method with diverse scenes recorded with different cameras/lenses. Code and data are available at https://github.com/kohei-ashida/dp-sfm.", + "bbox": [ + 259, + 311, + 743, + 534 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Multi-view 3D reconstruction $\\cdot$ Photogrammetry $\\cdot$ Dual-pixel imaging $\\cdot$ Scale ambiguity", + "bbox": [ + 259, + 545, + 741, + 574 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 595, + 375, + 612 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Multi-view 3D reconstruction is one of the most important problems in 3D computer vision. From multi-view images of a scene taken from unknown viewpoints, structure-from-motion (SfM) estimates camera poses and sparse 3D tie points of the scene. Subsequently, multi-view stereo (MVS) uses the estimated camera poses to reconstruct the dense shape of the scene. It is known that the scale ambiguity in SfM is inherent even with known camera intrinsics and cannot be resolved without external information. In this setting, the subsequent MVS can only recover dense shapes up to scale. A classical yet most reliable approach to resolving the ambiguity is to place a reference object of known size, such as a calibration target, in the scene to determine the absolute scale. However, it is not always possible to put such a reference object in the scene, and it is desired to relax the demand for diverse application scenarios.", + "bbox": [ + 212, + 625, + 787, + 805 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this paper, we propose a method for automatically determining the absolute scale in multi-view 3D reconstruction using dual-pixel (DP) cameras whose", + "bbox": [ + 212, + 806, + 787, + 838 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "intrinsic are calibrated. DP sensors are designed to assist the camera's auto-focus system and are becoming popular since they are equipped with Canon EOS cameras and Google Pixel smartphones. Each pixel on the DP sensor has two photodiodes to acquire two sub-aperture views. The two sub-aperture views allow us to estimate defocus blurs that are observed at regions off of the focal plane, although they depend on the unknown focus distance. We show that, by relating the DP defocus blur sizes with", + "bbox": [ + 212, + 146, + 405, + 462 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/43caf3e12a49ee35813a2d3581b790a58726370add16cf6431f57b409ff85d2a.jpg", + "image_caption": [ + "Fig. 1: Our method determines the absolute scene scale of SfM and MVS from multi-view DP images. We put together the scale ambiguity in SfM and MVS and the focus distance ambiguity of DP images. Our method jointly resolves the ambiguities." + ], + "image_footnote": [], + "bbox": [ + 416, + 150, + 787, + 354 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "unknown focus distances with the dense depth maps up to scale obtained from SfM and MVS, the absolute scale of the scene can be fully determined without requiring other external information.", + "bbox": [ + 212, + 463, + 787, + 508 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The key idea of our method is coupling two ambiguous measurements for resolving the ambiguities: One is the linear scale ambiguity of depth values computed from SfM and MVS, and the other is the focus distance ambiguity that affinely relates the reciprocal of depths with DP defocus blurs. As illustrated in Fig. 1, the size of defocus blur $b$ is related to unknown focus distance $g$ and unknown scene scale $s$ , with which scene point $\\mathbf{p}$ may be scaled to $\\mathbf{sp}$ retaining its projections $\\mathbf{u}_1$ and $\\mathbf{u}_2$ . By putting them together, our method solves a linear regression problem for determining both the absolute scale of the scene and the focus distances in the multi-view images. Our method capitalizes on the fact that DP's two sub-aperture views faithfully yield defocus blur estimates, significantly more reliably than just a single view from an ordinary camera.", + "bbox": [ + 212, + 515, + 787, + 681 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, the chief contributions of this paper are:", + "bbox": [ + 238, + 686, + 627, + 703 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose a method to determine the absolute scale in uncalibrated multi-view 3D reconstruction using DP images.", + "- We develop a linear solution method for simultaneously determining the global scene scale and per-view focus distances.", + "- We conduct comprehensive experiments to assess the effectiveness of the proposed method using diverse scenes and camera settings." + ], + "bbox": [ + 223, + 736, + 782, + 839 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "K. Ashida et al.", + "bbox": [ + 271, + 114, + 380, + 126 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 143, + 387, + 160 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Scale estimation in SfM Since SfM alone cannot recover the absolute scale from the image information, it is necessary to introduce an additional cue that provides absolute scale information. A simple and reliable approach is to place a reference object of known size, such as markers and ground control points (GCPs), which is widely employed in commercial photogrammetry software. Domain-specific cues are often used to avoid manual placement of reference objects, for example, urban objects [9,34], ground surfaces [33,42,43], and human faces [19]. To implicitly exploit the scale information of scene objects, recent learning-based methods use monocular depth estimation providing absolute scale information [27,28].", + "bbox": [ + 212, + 181, + 787, + 334 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Since it is not always possible to assume reference objects in target scenes, external sensors are often used to recover the scale information. The use of stereo cameras with known baselines [7, 35] or using LiDAR [13] are typical examples, while inertial sensors [23, 40] and odometry information [20, 29] are also widely used in robotics applications. Placing a refraction plate in front of the camera provides scale information using multiple capture [30, 31].", + "bbox": [ + 212, + 334, + 787, + 426 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Similar to our work, defocus has been used to determine the absolute scale in the literature. Calibrated depth-from-defocus is used with SfM-based depths to recover the absolute scale by assuming a known and fixed focus distance [32,37]. Similar to our method, a recent study [22] relaxes the known focus assumption by jointly estimating the physical focal length and scene scale. Although their use of the thin-lens model is similar to ours, their method requires pre-training or calibration for a learning-based blur size estimation [17] that depends on the hardware setup and focal lengths. Our study instead uses uncalibrated defocus blurs, where the focus distance is treated as unknown. The uncalibrated defocus blurs are reliably obtained from DP images, and thereby, our method recovers the absolute scales without prior knowledge, training data, or additional equipment.", + "bbox": [ + 212, + 426, + 787, + 592 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "DP imaging and applications Unlike conventional sensors, a single pixel in a DP sensor is divided into two photodiodes, one on the left and one on the right, which produces disparity. Beyond its primary use of fast auto-focus [4,14] implemented in several commercial cameras, applications of DP imaging are actively studied. DP's disparity gives effective cues to estimate the Point Spread Functions (PSFs) of defocus blurs [25], whose direct application is defocus deblurring [1-3,5,15,24,38,39] and blur synthesis [36]. The defocus blur information from DPs improves the performance of monocular depth estimation [10,18,25]. For this application, a DP image dataset with the ground truth depths has been developed [21]. The DP-based depth has been used for domain-specific applications, such as reflection removal [26] and facial shape recovery [16]. A recent method combines DP imaging and traditional two-view stereo for depth estimation [41]. Taking advantage of DP's effectiveness in acquiring defocus blur information, our work attempts to determine the scene scale automatically in multi-view 3D reconstruction.", + "bbox": [ + 212, + 613, + 787, + 839 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "Resolving Scale Ambiguity in MV3D Reconstruction using DP Sensors", + "bbox": [ + 259, + 114, + 730, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3 Preliminary: DP Imaging and Defocus Blur", + "text_level": 1, + "bbox": [ + 215, + 143, + 673, + 162 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Unlike conventional sensors, a DP sensor can capture two sub-aperture images by the left and right photodiodes in a single shot. There is a difference between the two images, which is caused by the difference in the PSFs of the left and right DP views [25].", + "bbox": [ + 212, + 176, + 784, + 236 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Figure 2 shows the visualization of the difference between two DP views and illustrated corresponding PSFs. There is no difference between the two views if the depth of a scene point is in focus. When the scene point is out of focus, i.e., located further away or closer to the camera from the focal plane, the disparities between the two views appear in opposite directions. Unlike traditional two-view stereo, the disparity is caused by defocus blurs generated by the PSFs for each view. It is known that the signed defocus blur size $b \\in \\mathbb{R}$ is proportional to the defocus disparity $d$ between two views [10].", + "bbox": [ + 212, + 237, + 486, + 492 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/a14d6a6f2608809c03dbd9527df6f41e995e1080d2268197d370dad5d637eaaa.jpg", + "image_caption": [ + "Fig.2: Visualization of DP disparity and illustrations of the blur, i.e., PSF, kernels. The two sub-aperture views convert the defocus blur to disparity, resulting in defocus disparity in opposite directions for near and far scene points. The DP image is from [2]." + ], + "image_footnote": [], + "bbox": [ + 498, + 237, + 785, + 385 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "DP imaging makes the estimation of defocus blur size $b$ tractable [25] without requiring external information or learning-based methods. Since DP PSFs $(\\mathbf{H}_l, \\mathbf{H}_r)$ can be well approximated by a parametric function of blur size $b$ , when DP image patches $(\\mathbf{G}_l, \\mathbf{G}_r)$ are given, the blur size can be estimated per pixel by a nonlinear minimization in an unsupervised manner as", + "bbox": [ + 212, + 494, + 784, + 570 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nb = \\underset {b} {\\operatorname {a r g m i n}} \\| \\mathbf {G} _ {l} * \\mathbf {H} _ {r} (b) - \\mathbf {G} _ {r} * \\mathbf {H} _ {l} (b) \\| _ {F}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 352, + 580, + 784, + 606 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $*$ is a convolution operator. The readers are referred to [25] for details.", + "bbox": [ + 214, + 617, + 771, + 633 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4 Proposed Method", + "text_level": 1, + "bbox": [ + 215, + 656, + 426, + 672 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our method resolves the scale ambiguity in multi-view 3D reconstruction using multi-view DP images. The overview is illustrated in Fig. 3. Our method infers the unknown scene scale just by capturing multi-view images with a DP camera. DP images provide a useful cue for blur size estimation, as described above, without requiring any external information. Our method shares the same spirit with [22]; however, there is a key difference in that our method does not require pre-training but capitalizes on the DP sensor's effectiveness in determining blur sizes. We show in Fig. 3 that the estimated blur size and (inverse) depth from SfM and MVS have an affine relationship, and thus, the unknown scene scale and focus distance can be estimated simultaneously by linear regression.", + "bbox": [ + 212, + 688, + 787, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "K. Ashida et al.", + "bbox": [ + 271, + 114, + 380, + 126 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/e6146d7ce555b3b411ea66fc42ee56abf4c9494ae602382032ff223ce4725c6d.jpg", + "image_caption": [ + "Fig. 3: Overview of our method that provides the absolute scale to multi-view 3D reconstruction (SfM and MVS) just by capturing the scene using DP cameras. (a) The disparity between DP views provides the cue to estimate the defocus blur size. (b) From the affine relation between the (inverse) depth and the defocus blur size, our method simultaneously estimates the scene scale and focus distances." + ], + "image_footnote": [], + "bbox": [ + 217, + 143, + 330, + 286 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/4217e975b1151f6f64652270725e964664f15783b355e2b2633da928a07035e2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 331, + 145, + 439, + 286 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/eab5002f47f48a54226bd5ff41606f75d61a4c4b5a1cd67e986bef107bdf52bb.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 439, + 146, + 566, + 286 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/508d66c9ced6e327a9d98d413f8941f78ee320443366a58c851759ffacb80dc3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 566, + 146, + 669, + 286 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/42084d35034be71ceacd7af684b36332e248068264def1810a9b52b1d0b914fb.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 669, + 146, + 784, + 286 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We begin by describing the ambiguities that arise in our setting, then describe the solution method for resolving them. We also present a method for selecting reliable pixels and views for scale estimation, which is based on the inference by an off-the-shelf blur size estimation method [25].", + "bbox": [ + 212, + 402, + 782, + 464 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Assumptions We assume that camera intrinsics, such as focal lengths and aperture sizes (f-numbers), are known as they are usually accessible from the camera system. Although defocus blurs slightly change the center of projection, we assume the change is sufficiently small and ignore their effect on the intrinsics. As stated, we use DP images taken from multiple viewpoints as input. Our method does not assume pre-calibrated viewpoints nor reference objects in a scene.", + "bbox": [ + 212, + 482, + 784, + 571 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1 Scale and Focus Distance Ambiguities", + "text_level": 1, + "bbox": [ + 215, + 599, + 575, + 614 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "There are two ambiguities in our setting: One is the scale ambiguity in SfM and MVS, and the other is the focus distance ambiguity in DP images.", + "bbox": [ + 212, + 628, + 782, + 657 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Scale ambiguity in SfM and MVS Camera positions and 3D point locations recovered by SfM and the subsequent MVS have a scale ambiguity even with known camera intrinsics. Let us consider a 3D point $\\mathbf{p} = [x,y,z]^{\\top}$ projected to the corresponding 2D point $\\mathbf{u} = [u,v,1]^{\\top}$ in the homogeneous coordinates with a projection matrix $[\\mathbf{R}|\\mathbf{t}]$ formed by a rotation matrix $\\mathbf{R} \\in SO(3)$ and a translation vector $\\mathbf{t} \\in \\mathbb{R}^3$ . With known camera intrinsics, we can set the intrinsics to be identity and write the 3D-to-2D perspective projection in some Euclidean world coordinate system as", + "bbox": [ + 212, + 676, + 787, + 797 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {u} \\sim [ \\mathbf {R} \\mathbf {t} ] \\left[ \\begin{array}{l} \\mathbf {p} \\\\ 1 \\end{array} \\right] = \\frac {1}{s} [ \\mathbf {R} s \\mathbf {t} ] \\left[ \\begin{array}{l} s \\mathbf {p} \\\\ 1 \\end{array} \\right], \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 380, + 809, + 785, + 844 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "Resolving Scale Ambiguity in MV3D Reconstruction using DP Sensors", + "bbox": [ + 259, + 114, + 732, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $s \\in \\mathbb{R}$ is an unknown scaling factor. As such, the depth $z'$ obtained by SfM and MVS is linearly related to the true depth $z$ by the unknown scaling factor $s$ as", + "bbox": [ + 212, + 146, + 787, + 191 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nz = s z ^ {\\prime}. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 467, + 200, + 785, + 215 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "This is the inherent ambiguity in SfM, which we are going to resolve in this paper.", + "bbox": [ + 212, + 223, + 785, + 255 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Focus distance ambiguity in $DP$ images As described by Wadhawa et al. [36] and Garg et al. [10], the signed defocus blur size $b \\in \\mathbb{R}$ is related to scene depth $z$ , focal length $f$ , lens aperture diameter $l$ , and the focus distance $g$ in $\\mathbb{R}$ as", + "bbox": [ + 214, + 266, + 785, + 313 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nb \\approx \\frac {l f}{1 - f / g} \\left(\\frac {1}{g} - \\frac {1}{z}\\right) \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 416, + 318, + 785, + 352 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "using the paraxial and thin-lens approximations. The blur size $b$ in DP images varies with the focus distance $g$ even when the same scene depth $z$ is recorded with the same camera parameters $(l, f)$ . Therefore, even with the blur size $b$ from DP images, its relationship to depth $z$ is ambiguous due to unknown focus distance $g$ and depth $z$ as", + "bbox": [ + 212, + 357, + 787, + 434 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nb \\approx \\frac {l f}{1 - f / g} \\left(\\frac {1}{g} - \\frac {1}{z}\\right) = \\frac {l f}{1 - f / g ^ {\\prime}} \\left(\\frac {1}{g ^ {\\prime}} - \\frac {1}{z ^ {\\prime}}\\right), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 330, + 440, + 785, + 474 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "i.e., there exists a set of pairs like $(g,z)$ and $(g^{\\prime},z^{\\prime})$ that yields the same blur size $b$ . The blur size $b$ is signed, i.e., it becomes positive when the scene point is farther away from the focal plane and negative when nearer to the camera.", + "bbox": [ + 214, + 479, + 787, + 526 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2 Solution Method", + "text_level": 1, + "bbox": [ + 214, + 544, + 401, + 558 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Given the blur sizes $b_{ij} \\in \\mathbb{R}$ obtained from DP images, depths $z_{ij}' \\in \\mathbb{R}_+$ from SfM and MVS in the $i$ -th view at $j$ -th pixel, and known per-view focal lengths $f_i$ and lens aperture $l_i$ , our goal is to determine unknown scene scale $s \\in \\mathbb{R}_+$ and per-view focus distances $g_i \\in \\mathbb{R}_+$ . Since we use per-view intrinsics, our method can be used in a practical setting of multi-view 3D reconstruction, in which camera intrinsics may vary across viewpoints.", + "bbox": [ + 212, + 566, + 787, + 656 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We put together the two ambiguous observations, i.e., depths up to scale and blur sizes with unknown focus distances, by substituting Eq. (3) to Eq. (4) as", + "bbox": [ + 212, + 657, + 787, + 688 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nb _ {i j} = \\frac {l _ {i} f _ {i}}{1 - f _ {i} / g _ {i}} \\left(\\frac {1}{g _ {i}} - \\frac {1}{s z _ {i j} ^ {\\prime}}\\right). \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 393, + 707, + 785, + 750 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Denoting $\\bar{g}_i = \\frac{1}{g_i}$ and $\\bar{s} = \\frac{1}{s}$ as $g_{i}$ and $s$ are both non-zero, Eq. (6) can be rewritten as", + "bbox": [ + 212, + 753, + 787, + 785 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\frac {b _ {i j} z _ {i j} ^ {\\prime}}{f _ {i}} = \\underbrace {z _ {i j} ^ {\\prime} \\left(b _ {i j} + l _ {i}\\right)} _ {\\gamma_ {i j}} \\bar {g} _ {i} - l _ {i} \\bar {s} = \\gamma_ {i j} \\bar {g} _ {i} - l _ {i} \\bar {s}, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 352, + 789, + 785, + 837 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "K. Ashida et al.", + "bbox": [ + 271, + 114, + 380, + 127 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/8fd12ed0558ecf1a2bc915ceeeda4b7ed4b7be17545e4c31e006c0e1c849577d.jpg", + "image_caption": [ + "(a) Input image" + ], + "image_footnote": [], + "bbox": [ + 228, + 143, + 356, + 209 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/863591e22344b41e66a82534277c12d2c0e1805c7427ee700eee6651dae482b9.jpg", + "image_caption": [ + "(b) Confidence score", + "Fig. 4: Pixel selection within a view based on confidence scores. We select reliable pixels that contain good blur size estimates from an input image (a). (b) shows the log-scaled confidence scores $c$ that show the reliability of blur size estimates [25]. (c) Local variations (variance) of depths. (d) Selected pixels based on our pixel selection method (highlighted in the figure)." + ], + "image_footnote": [], + "bbox": [ + 361, + 143, + 491, + 209 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/0770950cc90c5248ef44910478605d360d1c90b6a5e38a980e4d2fc482970531.jpg", + "image_caption": [ + "(c) Local depth variation" + ], + "image_footnote": [], + "bbox": [ + 504, + 142, + 638, + 209 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e8bde8534839581c05fa1678d070978a534e29386b33c4346f132c2d9d0fffeb.jpg", + "image_caption": [ + "(d) Selected pixels" + ], + "image_footnote": [], + "bbox": [ + 651, + 143, + 779, + 209 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "which is linear with respect to the unknown parameters $\\bar{g}_i$ and $\\bar{s}$ .", + "bbox": [ + 214, + 323, + 686, + 338 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "From observations of multiple views and pixels, we can construct an overdetermined system of linear equations from Eq. (7). Given the number of views $n$ and the number of valid pixels $p_i$ in the $i$ -th view, we have $\\sum_{i=1}^{n} p_i$ equations, which can be written in a matrix form as:", + "bbox": [ + 212, + 339, + 784, + 398 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\underbrace {\\left[ \\begin{array}{c c c c c} \\gamma_ {1 1} & & & & - l _ {1} \\\\ \\vdots & & & & \\vdots \\\\ \\gamma_ {1 p _ {1}} & & 0 & & - l _ {1} \\\\ & \\gamma_ {2 1} & & & - l _ {2} \\\\ & \\vdots & & & \\vdots \\\\ 0 & \\vdots & 0 & & \\vdots \\\\ & \\gamma_ {2 p _ {2}} & & & - l _ {2} \\\\ & & \\ddots & & \\vdots \\\\ & & & \\gamma_ {n 1} & - l _ {n} \\\\ & 0 & & \\vdots & \\vdots \\\\ & & & \\gamma_ {n p _ {n}} & - l _ {n} \\end{array} \\right]} _ {\\mathbf {A}} \\underbrace {\\left[ \\begin{array}{c} \\bar {g} _ {1} \\\\ \\bar {g} _ {2} \\\\ \\vdots \\\\ \\bar {g} _ {n} \\\\ \\bar {s} \\end{array} \\right]} _ {\\mathbf {x}} = \\underbrace {\\left[ \\begin{array}{c} b _ {1 1} z _ {1 1} ^ {\\prime} / f _ {1} \\\\ \\vdots \\\\ b _ {1 p _ {1}} z _ {1 p _ {1}} ^ {\\prime} / f _ {1} \\\\ b _ {2 1} z _ {2 1} ^ {\\prime} / f _ {2} \\\\ \\vdots \\\\ b _ {2 p _ {2}} z _ {2 p _ {2}} ^ {\\prime} / f _ {2} \\\\ \\vdots \\\\ b _ {n 1} z _ {n 1} ^ {\\prime} / f _ {n} \\\\ \\vdots \\\\ b _ {n p _ {n}} z _ {n p _ {n}} ^ {\\prime} / f _ {n} \\end{array} \\right]} _ {\\mathbf {b}}. \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 354, + 407, + 785, + 570 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Since we have $n + 1$ unknowns in vector $\\mathbf{x}$ , given $\\mathbf{b} \\neq \\mathbf{0}$ , the condition for obtaining a unique (approximate) solution is $\\mathrm{rank}(\\mathbf{A}) = n + 1$ , which is, in practice, easy to achieve because a large number of pixels are available. The minimum setting to determine scale $s$ is selecting two pixels with distinct depths in one of the multi-view images, which makes a full-rank $2 \\times 2$ matrix $\\mathbf{A}$ for two unknowns $\\mathbf{x} = [\\bar{g}_i, \\bar{s}]^\\top$ .", + "bbox": [ + 212, + 579, + 785, + 669 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In practice, we can select many pixels from multiple views to rely on robust estimation. With a sufficiently large number of selected pixels, we construct a tall and skinny matrix $\\mathbf{A}$ and use $\\ell_1$ residual minimization to effectively disregard outliers as", + "bbox": [ + 212, + 670, + 785, + 729 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {x} ^ {*} = \\underset {\\mathbf {x}} {\\operatorname {a r g m i n}} \\| \\mathbf {A x} - \\mathbf {b} \\| _ {1}. \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 408, + 743, + 785, + 767 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We use the Iteratively Reweighted Least Squares (IRLS) method [12] for deriving the approximate solution $\\mathbf{x}^*$ . Obviously, other choices of robust estimation methods, such as RANdom SAmple Consensus (RANSAC) [8], are viable for alleviating the impact of outliers.", + "bbox": [ + 212, + 779, + 785, + 839 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "Resolving Scale Ambiguity in MV3D Reconstruction using DP Sensors", + "bbox": [ + 259, + 114, + 732, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3 Confidence-based Pixel Selection", + "text_level": 1, + "bbox": [ + 215, + 146, + 532, + 161 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "So far, we have assumed that the blur sizes $b_{ij}$ are available to our method. For obtaining the blur sizes, we use the method proposed by Punnappurath et al. [25] that optimizes Eq. (1). For our case, we only require blur size estimates for a small number of pixels but not for all the pixels since our goal is determining the scene scale $s$ using Eq. (8). We, therefore, develop a pixel selection heuristic to choose a small number of pixels that have reliable blur size estimates. Specifically, we select the reliable pixels in each view, followed by selecting the views that produce reliable estimates of scene scale.", + "bbox": [ + 212, + 170, + 787, + 292 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Pixel-selection within a view We begin by adopting the confidence scores of [25] that indicate the reliability of the blur size estimates based on the residual of the blur size optimization and the textured-ness (magnitude of image gradient) of the pixel of interest. We slightly modified their original confidence score using a depth-based criterion since we observed that the large depth variation in an image patch negatively affects the blur size estimation.", + "bbox": [ + 212, + 306, + 787, + 396 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Let us consider one of the views and let $c_{j}$ be the confidence score defined in [25] at the $j$ -th pixel. In our case, a depth value $z_{j}^{\\prime}$ up to scale is available at the pixel, and we can assess the variance of the depth values within an image patch $\\mathcal{P}_j$ centered at the $j$ -th pixel as $\\text{var}(\\{z_k'\\})$ , $k \\in \\mathcal{P}_j$ . Based on the variance of the depths, we re-define the confidence score as", + "bbox": [ + 212, + 397, + 787, + 472 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\nc _ {j} ^ {*} \\triangleq \\frac {c _ {j}}{\\operatorname {v a r} \\left(\\left\\{z _ {k} ^ {\\prime} \\right\\}\\right)}, \\quad k \\in \\mathcal {P} _ {j}. \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 403, + 483, + 785, + 512 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The greater depth variation in an image patch leads to a larger variance and lower confidence score, and vice versa. Based on the new confidence score $c_{j}^{*}$ , we select the top $T_{c}\\%$ of pixels and regard them as reliable pixels that have good defocus blur estimates. Although the variance of depths varies with the depth scale, which is unknown in our case, it does not affect the confidence score $c_{j}^{*}$ within a view because the same scale is applied to all the pixels in the view. Figure 4 depicts the pixel selection within a view.", + "bbox": [ + 212, + 521, + 787, + 627 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "View selection In the multi-view images for SfM, there may be images that are not suitable for scale estimation. For example, a view with small depth variations is not preferred because the small depth variations make the linear system of Eq. (8) unstable. Therefore, we further select reliable views for scale estimation.", + "bbox": [ + 212, + 641, + 787, + 700 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Since our method can estimate scale $s$ for each view as long as the depths up to scale are available, we first compute the scale predictions for each view independently using the selected pixels within each view by solving Eq. (8). Ideally, the scales estimated for each view should match; however, in reality, they do not due to outliers. We first disregard the views with clear outliers, i.e., views with excessively small blur size variations and views that yield negative values of scale estimates. The small blur size variations indicate that the scene contains small depth variations. We simply set a threshold to discard the views with small blur size variations. Namely, when the difference between maximum", + "bbox": [ + 212, + 702, + 787, + 838 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "K. Ashida et al.", + "bbox": [ + 271, + 114, + 380, + 126 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "and minimum blur size estimates becomes less than or equal to $T_{p}$ pixels, the view is unused for the estimation. The negative scale estimates could arise due to various reasons, such as image degradation by motion blur and inaccurate estimates of either or both blur sizes and depths. We simply discard such views from the estimation.", + "bbox": [ + 212, + 146, + 431, + 325 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Finally, from the remaining views, we select $N_{v}$ views", + "bbox": [ + 212, + 328, + 431, + 357 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/b44b869b5c5b430d1397251a070bbe309c341c0c997e40a92e1e48b056720c79.jpg", + "image_caption": [ + "Fig. 5: View selection. We select the reliable views for scale estimation based on the median of the scale values estimated for each view." + ], + "image_footnote": [], + "bbox": [ + 444, + 148, + 539, + 286 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/8322b3ad199f9bc14638ca95278ee26c0f4379af9dbc9279ca712cda713f85b8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 542, + 148, + 686, + 287 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/f82b47f8ca6de496701585b22cc0500bbc01bf80664eed57d1597a7d03b04b64.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 687, + 157, + 785, + 286 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "that are close to the median of the estimated scales as illustrated in Fig. 5. Once the appropriate views are selected, by using all the selected pixels within the views, we create the linear system of Eq. (8) and re-estimate the scale $s$ .", + "bbox": [ + 212, + 358, + 787, + 402 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.4 Implementation Details", + "text_level": 1, + "bbox": [ + 214, + 426, + 455, + 441 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "For SfM and MVS, we use the commercial software Metashape1. For DP blur size estimation, we use the method proposed by Punnappurath et al. [25] as described, except for the smoothing operation implemented as post-processing. All captured images are undistorted using the parameters estimated by MVS before being fed into the blur size estimation. Our method is not restricted by the choice of SfM+MVS and blur size estimation methods, but we found these to be the best choices today.", + "bbox": [ + 212, + 453, + 787, + 559 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "For the focal length $f$ , we use the one from the lens specification. The aperture diameter $l$ is calculated using f-number $N$ of the lens as $l = \\frac{f}{N}$ . For the hyperparameters, we use $T_{c} = 10$ , $T_{p} = 2$ , and $N_{v} = 7$ throughout the experiments.", + "bbox": [ + 212, + 560, + 787, + 606 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "5 Experiment", + "text_level": 1, + "bbox": [ + 214, + 628, + 366, + 647 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We assess the accuracy of our scale estimation method both quantitatively and qualitatively. We use multi-view DP images captured in six scenes with different capturing settings, such as, cameras (DSLR and smartphone), lenses, and aperture sizes.", + "bbox": [ + 212, + 662, + 787, + 724 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Recording setting We use two different types of cameras equipped with a DP sensor: a DSLR camera and a smartphone. For the DSLR camera setting, we use Canon EOS 5D Mark IV with three different lenses: EF35mm F1.4L II USM, EF50mm F1.8 STM, and EF85mm F1.8 USM. For each scene, we also change the aperture size to assess the effect of the aperture size on our method. The", + "bbox": [ + 212, + 739, + 787, + 816 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "Resolving Scale Ambiguity in MV3D Reconstruction using DP Sensors", + "bbox": [ + 259, + 114, + 732, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "page_footnote", + "text": "1 Agisoft Metashape https://www.agisoft.com/, last accessed on July 12, 2024.", + "bbox": [ + 217, + 824, + 759, + 840 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "maximum aperture ( $f/1.4$ or $f/1.8$ depending on the lens), $f/4.0$ , and $f/8.0$ are used for the experiment depending on the target scene. The images from Canon's DLSR camera are recorded as Canon Raw images with the CR2 format, and the left and right DP views are extracted using DPRSplit2. For a more accessible setup, we also use a smartphone, Google Pixel 4, which has a focal length of $4.38 \\mathrm{~mm}$ and a fixed aperture of $f/1.73$ . To obtain the left and right DP views, we use the capturing software used in [11].", + "bbox": [ + 212, + 146, + 787, + 252 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "For each scene for each camera setting, we capture about 30 images using a DSLR, and 60 images using a smartphone. For fair comparisons across different lens and aperture sizes, for each viewpoint, we use a tripod to keep the camera fixed and record images with different camera settings. To evaluate the scale estimates, we obtain the ground-truth scene scale by placing markers with known sizes in the scenes.", + "bbox": [ + 212, + 252, + 787, + 342 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Evaluation metric As a quantitative evaluation metric, we use the scale ratio $r_s$ of the estimated scene scale $s_{\\mathrm{est}}$ to the ground-truth scale $s_{\\mathrm{gt}}$ , which is defined as follows:", + "bbox": [ + 212, + 358, + 787, + 400 + ], + "page_idx": 9 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {S c a l e r a t i o} r _ {s} = \\frac {s _ {\\mathrm {e s t}}}{s _ {\\mathrm {g t}}}.\n$$\n", + "text_format": "latex", + "bbox": [ + 421, + 400, + 580, + 429 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "In the above metric, closer to 1 indicates good scale estimation. The scale ratio smaller than 1 indicates that the scale is under-estimated, i.e., the estimated scale is smaller than the true scale, and the greater scale ratio corresponds to the over-estimation. While scale ratio $r_s$ is defined for each scene and camera setting, we also define the average error of the scale ratios $e_s$ to characterize the overall effectiveness as", + "bbox": [ + 212, + 434, + 787, + 523 + ], + "page_idx": 9 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {A v e r a g e e r r o r o f s c a l e r a t i o s} e _ {s} = \\frac {1}{n _ {s}} \\sum \\left(\\max (r _ {s}, r _ {s} ^ {- 1}) - 1\\right),\n$$\n", + "text_format": "latex", + "bbox": [ + 279, + 532, + 723, + 564 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "where $n_s$ is the number of settings, and $\\max(r_s, r_s^{-1})$ is for taking into account the under- and over-estimations.", + "bbox": [ + 212, + 574, + 785, + 604 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Quantitative results The results of the scale estimation for six scenes using two DP cameras are summarized in Table 1. The DSLR camera uses three lenses and different aperture sizes. Overall, our method can reasonably well estimate the absolute scene scale just by using a DP camera for multi-view capture, considering the fact that scale ratio $r_s$ can range in $(0, \\infty)$ . The average error $e_s$ across scenes and lenses is approximately 0.219 as shown in the table. From comparing the average errors for each lens, we see that the difference in lenses does not significantly affect the overall performance.", + "bbox": [ + 212, + 619, + 787, + 742 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Visual results Figures 6 and 7 show visual examples of our results by the DSLR and smartphone, as well as the estimated blur sizes and the ground-truth depth maps. We can confirm the same trend as in the quantitative results, where the", + "bbox": [ + 212, + 756, + 787, + 803 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "K. Ashida et al.", + "bbox": [ + 271, + 114, + 380, + 126 + ], + "page_idx": 9 + }, + { + "type": "page_footnote", + "text": "$^{2}$ DPRSplit https://www.fastrawviewer.com/DPRSplit, last accessed on July 12, 2024.", + "bbox": [ + 217, + 809, + 785, + 838 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/163b087c947ddd277a14c36e9adb178c5f9437e92ad9e82bb17f67c03f2344b5.jpg", + "image_caption": [ + "Fig. 6: Visual results of real scenes captured by the DSLR (Canon EOS 5D Mark IV). Our method yields reasonable estimates of scene scales just by using a DP camera for multi-view capture, which is theoretically unrecoverable by standard SfM alone. The top three rows show the results of the same scene (and from almost the same viewpoint) but with different aperture sizes or focal lengths. The bottom rows show the results of different scenes. In (a), the distance indicates the true distance obtained from the GT 3D model. In (d), the distance in the estimated scale is shown. (b) shows estimated blur size and selected pixels based on our pixel selection method (highlighted in the figure)." + ], + "image_footnote": [], + "bbox": [ + 218, + 157, + 784, + 686 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "Resolving Scale Ambiguity in MV3D Reconstruction using DP Sensors", + "bbox": [ + 259, + 114, + 730, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/4386b9e7bc109c60529a19c283bd2eababc511000eb3624bf94ed5905141c5b9.jpg", + "table_caption": [ + "Table 1: Scale ratio $r_s$ between the ground truth and the estimated scale for each scene, lens, and aperture size. Maximum aperture varies depending on the lens (i.e., $f/1.4$ or $f/1.8$ ). $e_s$ represents the average errors of the corresponding row/column/total. “-” indicates that we did not record the scene under the setting." + ], + "table_footnote": [], + "table_body": "
DSLRR (35mm)DSLRR (50mm)DSLRR (85mm)Smartphone f/1.73es
f/1.4f/4.0f/8.0f/1.8f/4.0f/8.0f/1.8f/4.0f/8.0
Scene 11.4191.088-1.0841.084-1.1001.231-1.0850.156
Scene 21.2251.238-1.0471.232-1.2471.302-1.2520.220
Scene 31.1991.229-1.0551.352-1.2381.226-1.5190.260
Scene 41.1811.1071.4771.0271.2661.4691.2251.1901.2961.0340.227
Scene 51.1711.0721.3141.0671.0991.2851.0381.2051.4251.0680.174
Scene 61.1481.073-1.0831.029-1.1231.126-1.3880.139
es0.2240.1350.3960.0610.1770.3770.1620.2130.3610.2240.219
", + "bbox": [ + 220, + 212, + 784, + 328 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/e1f22bce1241aa74bdb7db06f1eaadc5750fc95f6e267f6ddbc0098de011ee44.jpg", + "image_caption": [ + "Fig. 7: Visual results of real scenes captured by a smartphone (Google Pixel 4). Similar to the results by DSLR, our method estimates reasonable scene scales despite the DP blur being relatively smaller than DSLR's." + ], + "image_footnote": [], + "bbox": [ + 222, + 349, + 787, + 522 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "depth maps with estimated scale well represent the actual depths. Our method estimates the scene scale with less than 0.2 error ratio for several settings, where the blur size estimation is relatively stable. For these cases, 3D reconstruction with our absolute scale estimation allows high-quality measurements with an error of several centimeters in wide outdoor scenes.", + "bbox": [ + 212, + 614, + 787, + 690 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Stability of our method To assess the stability of our method against variation of input images, we capture additional $200\\sim 300$ images for Scenes 1 and 2 using the DSLR equipped with a $35\\mathrm{mm}$ lens and $f / 1.4$ aperture size and repeatedly estimate the scale using different sets of images. Figure 8 shows an overlay of estimated models with repeated experiments and the minimum and maximum estimated scale ratios, $r_{s\\min}$ and $r_{s\\max}$ , with 10 trials.", + "bbox": [ + 212, + 719, + 787, + 811 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "K. Ashida et al.", + "bbox": [ + 271, + 114, + 380, + 126 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/61d7b3b0ae9eb49d6a3502f402859a6da3a1fd5fc0aa6d7716db9ddac7e3e04d.jpg", + "image_caption": [ + "$r_{s\\min}$ : 1.158, $r_{s\\max}$ : 1.255" + ], + "image_footnote": [], + "bbox": [ + 223, + 138, + 450, + 242 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/3db705a2597fe43b78da37ac93467a4725f089d1bb2b850924998d19cb22aa7e.jpg", + "image_caption": [ + "$r_{s\\min}$ : 1.174, $r_{s\\max}$ : 1.285", + "$r_{s\\min}$ : 1.202, $r_{s\\max}$ : 1.383", + "Fig. 8: Overlay of estimated models with repeated experiments and the minimum and maximum estimated scale ratios with 10 trials. Blue and red point clouds represent the ones with scale ratios $r_{s\\max}$ and $r_{s\\min}$ , respectively." + ], + "image_footnote": [], + "bbox": [ + 452, + 138, + 620, + 243 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/d50a86c1bda7f2ca4c53def7ce22e40dfdb37ba22a7037a36e5f6498c21913c9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 622, + 138, + 771, + 243 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Comparison with scale estimation by monocular depth estimation We compare our method with a single-image metric depth estimation, ZoeDepth [6], for the same scenes and lens setup used. Since it estimates metric depth for each pixel, we compute the per-image scale as the average ratio between ZoeDepth and MVS depth maps. We then average the per-image scale over seven inlier viewpoints selected using the same procedure as our view selection.", + "bbox": [ + 212, + 342, + 500, + 523 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "The average error $e_{s}$ across scenes and lenses using ZoeDepth is 0.679, which is more than 3 times larger than ours (0.219). Figure 9 confirms this", + "bbox": [ + 212, + 523, + 500, + 583 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "trend through the visualization of the estimated scale. More detailed comparisons are shown in the supplementary material.", + "bbox": [ + 212, + 584, + 785, + 614 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/8f6034f5d9ae24edea8557f7eaeaa475057db4fba9c6f38e95cbab533182586b.jpg", + "image_caption": [ + "$r_{sours}$ : 1.038, $r_{sZoe\\text{depth}}$ : 0.643", + "Fig. 9: Comparison with monocular scale estimation. Blue and red points represent the ones of the scale ratio estimated by our method ( $r_{s\\text{ours}}$ ) and ZoeDepth ( $r_{sZoeDepth}$ ), respectively. Black points represent the GT model." + ], + "image_footnote": [], + "bbox": [ + 549, + 347, + 751, + 453 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Limitations and failure cases Although our method successfully estimates the scene scale in many cases, we find several failure modes in our experiments. Since our method is based on defocus blur, scale estimation from pan-focus images becomes challenging. In particular, since the aperture of smartphone cameras is smaller than that of DSLR's, it was necessary for us to include photographs that exhibit large depth variations to produce relatively large blur like the top row of Fig. 7. Although the pan-focus images are preferable in SfM and MVS for better correspondence matching, in practice, small apertures cannot always be used. This is because the small aperture limits the amount of incoming light, requiring longer exposure times, and resulting in camera motion blurs.", + "bbox": [ + 212, + 628, + 787, + 779 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "The success of our method relies on the accurate estimation of both defocus blur sizes and depth maps up to scale. We found several examples where the scale estimation is affected by the blur size estimation errors, while the depth maps up to scale are almost always stably estimated. Figure 10 shows a failure", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "Resolving Scale Ambiguity in MV3D Reconstruction using DP Sensors", + "bbox": [ + 259, + 113, + 732, + 130 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/3c7a6f6e962f45147b489b907b96e2ae5219d2706a3753f9b87cbf9c19d6630c.jpg", + "image_caption": [ + "(a) Input image" + ], + "image_footnote": [], + "bbox": [ + 223, + 143, + 339, + 203 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/e018bb67874fe57f81ab84c83b4bcf5fb8abf4acd4ca5c9fba73a2fa61036d0e.jpg", + "image_caption": [ + "(b) GT depth", + "Fig. 10: An example of failure case (Scene 3 captured with a $50\\mathrm{mm}$ lens and $f / 4.0$ aperture size) producing scale ratio $r_s = 1.352$ due to the large error in blur size estimation. Ideal blur sizes (d) are calculated from the ground-truth (GT) scene depth (b) and Eq. (6). Since the target scene is mostly textureless or a repetition of fine textures, most of the pixels in the estimated blur sizes (c) deviate from ideal values, making the selection of the inlier pixels challenging." + ], + "image_footnote": [], + "bbox": [ + 344, + 142, + 483, + 203 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/cee5efa42d29122ffa21fcda926f8fb19df6df3a9300c64009ffa49c106a0490.jpg", + "image_caption": [ + "(c) Estimated blur size" + ], + "image_footnote": [], + "bbox": [ + 495, + 142, + 630, + 203 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/176d138d1f5f26a1feb898aa0a02a4da5db369a7a064d2c1087c34b6ffcc9370.jpg", + "image_caption": [ + "(d) Ideal blur size" + ], + "image_footnote": [], + "bbox": [ + 637, + 142, + 779, + 203 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "mode of blur size estimation as well as the estimated scene depth, whose scale ratio $r_s$ reaches as large as 1.352 (Scene 3 captured with $50\\mathrm{mm}$ lens and $f / 4.0$ aperture size). To assess the quality of blur size estimates shown in Fig. 10 (c), we synthesize the ideal blur sizes in Fig. 10 (d) using the ground-truth depth of Fig. 10 (b) that is computed from the markers and the image formation model of Eq. (6). Most of the blur size estimates are far from the ideal blur sizes due to the target object being almost textureless or the repetition of fine-grained texture, which makes blur size optimization challenging. We further analyze the accuracy and characteristics of the existing blur size estimation methods in the supplementary material.", + "bbox": [ + 212, + 330, + 787, + 482 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "6 Discussions", + "text_level": 1, + "bbox": [ + 214, + 496, + 362, + 510 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "This paper has shown that the scale information in multi-view 3d reconstruction, which is theoretically unsolvable by SfM and MVS itself, can be inferred from the defocus blur information provided by DP imaging. Our method estimates the absolute scale without relying on reference objects or data priors but by simply capturing multi-view images with a DP camera. The key to the scale recovery is the unique feature of DP that relates defocus disparity, blur size, and scene depth, from which we can affinely relate the reciprocal of the depth and the defocus blur sizes given by DP images. Together with the pixel selection method for improving robustness using the confidence information from the off-the-shelf DP blur estimation method as well as the statistics of the depth information, we simultaneously recover the scene scale and the unknown focus distance via a simple linear regression.", + "bbox": [ + 212, + 517, + 787, + 698 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Experiments show that our method successfully estimates the absolute scale in many cases. Since the estimation accuracy of the blur size strongly affects the performance of our method, the analysis of DP blur characteristics will be beneficial in the future. With a method for a more accurate and stabler estimation of blur sizes in DP images, our method will produce more accurate absolute scales for multi-view 3D reconstruction.", + "bbox": [ + 212, + 699, + 787, + 789 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgments", + "text_level": 1, + "bbox": [ + 215, + 803, + 393, + 819 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "This work was supported by JSPS KAKENHI Grant Number JP23H05491.", + "bbox": [ + 215, + 824, + 756, + 839 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "K. Ashida et al.", + "bbox": [ + 271, + 114, + 380, + 126 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 143, + 321, + 159 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Abuolaim, A., Afifi, M., Brown, M.S.: Improving single-image defocus deblurring: How dual-pixel images help through multi-task learning. In: IEEE/CVF Winter Conference on Applications of Computer Vision (WACV). pp. 82-90 (2022)", + "2. Abuolaim, A., Brown, M.S.: Defocus deblurring using dual-pixel data. In: Proceedings of European Conference on Computer Vision (ECCV). pp. 111-126 (2020)", + "3. Abuolaim, A., Delbracio, M., Kelly, D., Brown, M.S., Milanfar, P.: Learning to reduce defocus blur by realistically modeling dual-pixel data. In: Proceedings of IEEE/CVF International Conference on Computer Vision (ICCV). pp. 2289-2298 (2021)", + "4. Abuolaim, A., Punnappurath, A., Brown, M.S.: Revisiting autofocus for smartphone cameras. In: Proceedings of European Conference on Computer Vision (ECCV). pp. 545-559 (2018)", + "5. Abuolaim, A., Timofte, R., Brown, M.S.: NTIRE 2021 challenge for defocus deblurring using dual-pixel images: Methods and results. In: Proceedings of IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW). pp. 578-587 (2021)", + "6. Bhat, S.F., Birkl, R., Wofk, D., Wonka, P., Müller, M.: Zoedepth: Zero-shot transfer by combining relative and metric depth. arXiv preprint arXiv:2302.12288 (2023)", + "7. Engel, J., Stuckler, J., Cremers, D.: Large-scale direct SLAM with stereo cameras. In: IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). pp. 1935-1942 (2015)", + "8. Fischler, M.A., Bolles, R.C.: Random sample consensus: A paradigm for model fitting with applications to image analysis and automated cartography. Communications of the ACM 24(6), 381-395 (1981)", + "9. Frost, D., Prisacariu, V., Murray, D.: Recovering stable scale in monocular SLAM using object-supplemented bundle adjustment. IEEE Transactions on Robotics 34(3), 736-747 (2018)", + "0. Garg, R., Wadhwa, N., Ansari, S., Barron, J.T.: Learning single camera depth estimation using dual-pixels. In: Proceedings of IEEE/CVF International Conference on Computer Vision (ICCV). pp. 7628-7637 (2019)", + "1. Garg, R., Wadhwa, N., Ansari, S., Barron, J.T.: Learning single camera depth estimation using dual-pixels. In: Proceedings of IEEE/CVF International Conference on Computer Vision (ICCV). pp. 7628-7637 (2019)", + "2. Gentle, J.: Matrix Albegra. Springer Texts in Statistics, Springer, New York (2007)", + "3. Giubilato, R., Chiodini, S., Pertile, M., Debei, S.: Scale correct monocular visual odometry using a lidar altimeter. In: IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). pp. 3694-3700 (2018)", + "4. Herrmann, C., Bowen, R.S., Wadhwa, N., Garg, R., He, Q., Barron, J.T., Zabih, R.: Learning to autofocus. In: Proceedings of IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 2227-2236 (2020)", + "5. Jung, S.H., Heo, Y.S.: Disparity probability volume guided defocus deblurring using dual pixel data. In: Proceedings of International Conference on Information and Communication Technology Convergence (ICTC). pp. 305-308 (2021)", + "6. Kang, M., Choe, J., Ha, H., Jeon, H.G., Im, S., Kweon, I.S.: Facial depth and normal estimation using single dual-pixel camera. In: Proceedings of European Conference on Computer Vision (ECCV) (2022)", + "7. Kashiwagi, M., Mishima, N., Kozakaya, T., Hiura, S.: Deep depth from aberration map. In: Proceedings of IEEE/CVF International Conference on Computer Vision (ICCV) (2019)" + ], + "bbox": [ + 225, + 172, + 784, + 840 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "Resolving Scale Ambiguity in MV3D Reconstruction using DP Sensors", + "bbox": [ + 259, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "18. Kim, D., Jang, H., Kim, I., Kim, M.H.: Spatio-focal bidirectional disparity estimation from a dual-pixel image. In: Proceedings of IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 5023-5032 (2023)", + "19. Knorr, S.B., Kurz, D.: Leveraging the user's face for absolute scale estimation in handheld monocular SLAM. In: Proceedings of IEEE International Symposium on Mixed and Augmented Reality (ISMAR). pp. 11-17 (2016)", + "20. Lee, S.H., de Croon, G.: Stability-based scale estimation for monocular SLAM. IEEE Robotics and Automation Letters 3(2), 780-787 (2018)", + "21. Li, F., Guo, H., Santo, H., Okura, F., Matsushita, Y.: Learning to synthesize photorealistic dual-pixel images from RGBD frames. In: International Conference on Computational Photography (ICCP). pp. 1-11 (2023)", + "22. Mishima, N., Seki, A., Hiura, S.: Absolute scale from varifocal monocular camera through SfM and defocus combined. In: Proceedings of British Machine Vision Conference (BMVC) (2021)", + "23. Nützi, G., Weiss, S., Scaramuzza, D., Siegwart, R.: Fusion of IMU and vision for absolute scale estimation in monocular SLAM. Journal of Intelligent & Robotic Systems 61(1-4), 287-299 (2011)", + "24. Pan, L., Chowdhury, S., Hartley, R., Liu, M., Zhang, H., Li, H.: Dual pixel exploration: Simultaneous depth estimation and image restoration. In: Proceedings of IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 4340-4349 (2021)", + "25. Punnappurath, A., Abuolaim, A., Affi, M., Brown, M.S.: Modeling defocus-disparity in dual-pixel sensors. In: International Conference on Computational Photography (ICCP). pp. 1-12 (2020)", + "26. Punnappurath, A., Brown, M.S.: Reflection removal using a dual-pixel sensor. In: Proceedings of IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 1556-1565 (2019)", + "27. Roussel, T., Van Eycken, L., Tuytelaars, T.: Monocular depth estimation in new environments with absolute scale. In: IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). pp. 1735-1741 (2019)", + "28. Rukhovich, D., Mouritzen, D., Kaestner, R., Rufli, M., Velizhev, A.: Estimation of absolute scale in monocular SLAM using synthetic data. In: Proceedings of IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW). pp. 803-812 (2019)", + "29. Scaramuzza, D., Fraundorfer, F., Pollefeys, M., Siegwart, R.: Absolute scale in structure from motion from a single vehicle mounted camera by exploiting nonholonomic constraints. In: Proceedings of IEEE/CVF International Conference on Computer Vision (ICCV). pp. 1413-1419 (2009)", + "30. Shibata, A., Fujii, H., Yamashita, A., Asama, H.: Absolute scale structure from motion using a refractive plate. In: Proceedings of IEEE/SICE International Symposium on System Integration (SII). pp. 540-545 (2015)", + "31. Shibata, A., Fujii, H., Yamashita, A., Asama, H.: Scale-reconstructable structure from motion using refraction with a single camera. In: IEEE International Conference on Robotics and Automation (ICRA). pp. 5239-5244 (2015)", + "32. Shiozaki, T., Dissanayake, G.: Eliminating scale drift in monocular SLAM using depth from defocus. IEEE Robotics and Automation Letters 3(1), 581-587 (2017)", + "33. Song, S., Chandraker, M.: Robust scale estimation in real-time monocular sfm for autonomous driving. In: Proceedings of IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 1566-1573 (2014)" + ], + "bbox": [ + 215, + 146, + 784, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "K. Ashida et al.", + "bbox": [ + 271, + 114, + 380, + 126 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "34. Sucar, E., Hayet, J.B.: Bayesian scale estimation for monocular SLAM based on generic object detection for correcting scale drift. In: IEEE International Conference on Robotics and Automation (ICRA). pp. 5152-5158 (2018)", + "35. Sumikura, S., Sakurada, K., Kawaguchi, N., Nakamura, R.: Scale estimation of monocular SfM for a multi-modal stereo camera. In: Proceedings of Asian Conference on Computer Vision (ACCV). pp. 281-297 (2019)", + "36. Wadhwa, N., Garg, R., Jacobs, D.E., Feldman, B.E., Kanazawa, N., Carroll, R., Movshovitz-Attias, Y., Barron, J.T., Pritch, Y., Levoy, M.: Synthetic depth-of-field with a single-camera mobile phone. ACM Transactions on Graphics (TOG) pp. 1-13 (2018)", + "37. Wöhler, C., d'Angelo, P., Krüger, L., Kuhl, A., Groß, H.M.: Monocular 3D scene reconstruction at absolute scale. ISPRS Journal of Photogrammetry and Remote Sensing 64(6), 529-540 (2009)", + "38. Xin, S., Wadhwa, N., Xue, T., Barron, J.T., Srinivasan, P.P., Chen, J., Gkioulekas, I., Garg, R.: Defocus map estimation and deblurring from a single dual-pixel image. In: Proceedings of IEEE/CVF International Conference on Computer Vision (ICCV). pp. 2228-2238 (2021)", + "39. Yang, Y., Pan, L., Liu, L., Liu, M.: K3DN: Disparity-aware kernel estimation for dual-pixel defocus deblurring. In: Proceedings of IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 13263-13272 (2023)", + "40. Zhang, S., Zhang, J., Tao, D.: Towards scale-aware, robust, and generalizable unsupervised monocular depth estimation by integrating IMU motion dynamics. In: Proceedings of European Conference on Computer Vision (ECCV). pp. 143-160 (2022)", + "41. Zhang, Y., Wadhwa, N., Orts-Escolano, S., Häne, C., Fanello, S.R., Garg, R.: Du2Net: Learning depth estimation from dual-cameras and dual-pixels. In: Proceedings of European Conference on Computer Vision (ECCV). pp. 582-598 (2020)", + "42. Zhou, D., Dai, Y., Li, H.: Reliable scale estimation and correction for monocular visual odometry. In: Proceedings of IEEE Intelligent Vehicles Symposium (IV). pp. 490-495 (2016)", + "43. Zhou, D., Dai, Y., Li, H.: Ground-plane-based absolute scale estimation for monocular visual odometry. IEEE Transactions on Intelligent Transportation Systems 21(2), 791-802 (2019)" + ], + "bbox": [ + 212, + 146, + 787, + 604 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "Resolving Scale Ambiguity in MV3D Reconstruction using DP Sensors", + "bbox": [ + 259, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + } +] \ No newline at end of file diff --git a/2024/Resolving Scale Ambiguity in Multi-view 3D Reconstruction using Dual-Pixel Sensors/518914c5-d998-41c3-963e-0840d0e1f2c8_model.json b/2024/Resolving Scale Ambiguity in Multi-view 3D Reconstruction using Dual-Pixel Sensors/518914c5-d998-41c3-963e-0840d0e1f2c8_model.json new file mode 100644 index 0000000000000000000000000000000000000000..3404b139dd33edf321cdecb5544482f7e4f14135 --- /dev/null +++ b/2024/Resolving Scale Ambiguity in Multi-view 3D Reconstruction using Dual-Pixel Sensors/518914c5-d998-41c3-963e-0840d0e1f2c8_model.json @@ -0,0 +1,2522 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.248, + 0.141, + 0.756, + 0.187 + ], + "angle": 0, + "content": "Resolving Scale Ambiguity in Multi-view 3D Reconstruction using Dual-Pixel Sensors" + }, + { + "type": "text", + "bbox": [ + 0.227, + 0.213, + 0.776, + 0.228 + ], + "angle": 0, + "content": "Kohei Ashida, Hiroaki Santo, Fumio Okura, and Yasuyuki Matsushita" + }, + { + "type": "text", + "bbox": [ + 0.312, + 0.24, + 0.692, + 0.255 + ], + "angle": 0, + "content": "Graduate School of Information Science and Technology," + }, + { + "type": "text", + "bbox": [ + 0.421, + 0.256, + 0.584, + 0.268 + ], + "angle": 0, + "content": "Osaka University, Japan" + }, + { + "type": "text", + "bbox": [ + 0.257, + 0.269, + 0.746, + 0.283 + ], + "angle": 0, + "content": "{kohei.ashida, santo.hiroaki, okura, yasumat}@ist.osaka-u.ac.jp" + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.312, + 0.744, + 0.535 + ], + "angle": 0, + "content": "Abstract. Multi-view 3D reconstruction, namely structure-from-motion and multi-view stereo, is an essential component in 3D computer vision. In general, multi-view 3D reconstruction suffers from unknown scale ambiguity unless a reference object of known size is recorded together with the scene, or the camera poses are pre-calibrated. In this paper, we show that multi-view images recorded by a dual-pixel (DP) sensor allow us to automatically resolve the scale ambiguity without requiring a reference object or pre-calibration. Specifically, the observed defocus blurs in DP images provide sufficient information for determining the scale when paired together with the depth maps (up to scale) recovered from the multi-view 3D reconstruction. Based on this observation, we develop a simple yet effective linear solution method to determine the absolute scale in multi-view 3D reconstruction. Experiments demonstrate the effectiveness of the proposed method with diverse scenes recorded with different cameras/lenses. Code and data are available at https://github.com/kohei-ashida/dp-sfm." + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.546, + 0.743, + 0.575 + ], + "angle": 0, + "content": "Keywords: Multi-view 3D reconstruction \\(\\cdot\\) Photogrammetry \\(\\cdot\\) Dual-pixel imaging \\(\\cdot\\) Scale ambiguity" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.597, + 0.377, + 0.613 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.626, + 0.788, + 0.806 + ], + "angle": 0, + "content": "Multi-view 3D reconstruction is one of the most important problems in 3D computer vision. From multi-view images of a scene taken from unknown viewpoints, structure-from-motion (SfM) estimates camera poses and sparse 3D tie points of the scene. Subsequently, multi-view stereo (MVS) uses the estimated camera poses to reconstruct the dense shape of the scene. It is known that the scale ambiguity in SfM is inherent even with known camera intrinsics and cannot be resolved without external information. In this setting, the subsequent MVS can only recover dense shapes up to scale. A classical yet most reliable approach to resolving the ambiguity is to place a reference object of known size, such as a calibration target, in the scene to determine the absolute scale. However, it is not always possible to put such a reference object in the scene, and it is desired to relax the demand for diverse application scenarios." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.807, + 0.788, + 0.839 + ], + "angle": 0, + "content": "In this paper, we propose a method for automatically determining the absolute scale in multi-view 3D reconstruction using dual-pixel (DP) cameras whose" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.381, + 0.127 + ], + "angle": 0, + "content": "K. Ashida et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.406, + 0.463 + ], + "angle": 0, + "content": "intrinsic are calibrated. DP sensors are designed to assist the camera's auto-focus system and are becoming popular since they are equipped with Canon EOS cameras and Google Pixel smartphones. Each pixel on the DP sensor has two photodiodes to acquire two sub-aperture views. The two sub-aperture views allow us to estimate defocus blurs that are observed at regions off of the focal plane, although they depend on the unknown focus distance. We show that, by relating the DP defocus blur sizes with" + }, + { + "type": "image", + "bbox": [ + 0.418, + 0.151, + 0.788, + 0.356 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.412, + 0.366, + 0.788, + 0.436 + ], + "angle": 0, + "content": "Fig. 1: Our method determines the absolute scene scale of SfM and MVS from multi-view DP images. We put together the scale ambiguity in SfM and MVS and the focus distance ambiguity of DP images. Our method jointly resolves the ambiguities." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.464, + 0.788, + 0.51 + ], + "angle": 0, + "content": "unknown focus distances with the dense depth maps up to scale obtained from SfM and MVS, the absolute scale of the scene can be fully determined without requiring other external information." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.516, + 0.788, + 0.683 + ], + "angle": 0, + "content": "The key idea of our method is coupling two ambiguous measurements for resolving the ambiguities: One is the linear scale ambiguity of depth values computed from SfM and MVS, and the other is the focus distance ambiguity that affinely relates the reciprocal of depths with DP defocus blurs. As illustrated in Fig. 1, the size of defocus blur \\( b \\) is related to unknown focus distance \\( g \\) and unknown scene scale \\( s \\), with which scene point \\( \\mathbf{p} \\) may be scaled to \\( \\mathbf{sp} \\) retaining its projections \\( \\mathbf{u}_1 \\) and \\( \\mathbf{u}_2 \\). By putting them together, our method solves a linear regression problem for determining both the absolute scale of the scene and the focus distances in the multi-view images. Our method capitalizes on the fact that DP's two sub-aperture views faithfully yield defocus blur estimates, significantly more reliably than just a single view from an ordinary camera." + }, + { + "type": "text", + "bbox": [ + 0.24, + 0.688, + 0.629, + 0.704 + ], + "angle": 0, + "content": "In summary, the chief contributions of this paper are:" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.737, + 0.784, + 0.768 + ], + "angle": 0, + "content": "- We propose a method to determine the absolute scale in uncalibrated multi-view 3D reconstruction using DP images." + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.774, + 0.784, + 0.803 + ], + "angle": 0, + "content": "- We develop a linear solution method for simultaneously determining the global scene scale and per-view focus distances." + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.81, + 0.784, + 0.84 + ], + "angle": 0, + "content": "- We conduct comprehensive experiments to assess the effectiveness of the proposed method using diverse scenes and camera settings." + }, + { + "type": "list", + "bbox": [ + 0.225, + 0.737, + 0.784, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.261, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Resolving Scale Ambiguity in MV3D Reconstruction using DP Sensors" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.145, + 0.388, + 0.161 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.183, + 0.788, + 0.335 + ], + "angle": 0, + "content": "Scale estimation in SfM Since SfM alone cannot recover the absolute scale from the image information, it is necessary to introduce an additional cue that provides absolute scale information. A simple and reliable approach is to place a reference object of known size, such as markers and ground control points (GCPs), which is widely employed in commercial photogrammetry software. Domain-specific cues are often used to avoid manual placement of reference objects, for example, urban objects [9,34], ground surfaces [33,42,43], and human faces [19]. To implicitly exploit the scale information of scene objects, recent learning-based methods use monocular depth estimation providing absolute scale information [27,28]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.335, + 0.788, + 0.427 + ], + "angle": 0, + "content": "Since it is not always possible to assume reference objects in target scenes, external sensors are often used to recover the scale information. The use of stereo cameras with known baselines [7, 35] or using LiDAR [13] are typical examples, while inertial sensors [23, 40] and odometry information [20, 29] are also widely used in robotics applications. Placing a refraction plate in front of the camera provides scale information using multiple capture [30, 31]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.428, + 0.788, + 0.593 + ], + "angle": 0, + "content": "Similar to our work, defocus has been used to determine the absolute scale in the literature. Calibrated depth-from-defocus is used with SfM-based depths to recover the absolute scale by assuming a known and fixed focus distance [32,37]. Similar to our method, a recent study [22] relaxes the known focus assumption by jointly estimating the physical focal length and scene scale. Although their use of the thin-lens model is similar to ours, their method requires pre-training or calibration for a learning-based blur size estimation [17] that depends on the hardware setup and focal lengths. Our study instead uses uncalibrated defocus blurs, where the focus distance is treated as unknown. The uncalibrated defocus blurs are reliably obtained from DP images, and thereby, our method recovers the absolute scales without prior knowledge, training data, or additional equipment." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.614, + 0.788, + 0.84 + ], + "angle": 0, + "content": "DP imaging and applications Unlike conventional sensors, a single pixel in a DP sensor is divided into two photodiodes, one on the left and one on the right, which produces disparity. Beyond its primary use of fast auto-focus [4,14] implemented in several commercial cameras, applications of DP imaging are actively studied. DP's disparity gives effective cues to estimate the Point Spread Functions (PSFs) of defocus blurs [25], whose direct application is defocus deblurring [1-3,5,15,24,38,39] and blur synthesis [36]. The defocus blur information from DPs improves the performance of monocular depth estimation [10,18,25]. For this application, a DP image dataset with the ground truth depths has been developed [21]. The DP-based depth has been used for domain-specific applications, such as reflection removal [26] and facial shape recovery [16]. A recent method combines DP imaging and traditional two-view stereo for depth estimation [41]. Taking advantage of DP's effectiveness in acquiring defocus blur information, our work attempts to determine the scene scale automatically in multi-view 3D reconstruction." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.381, + 0.127 + ], + "angle": 0, + "content": "K. Ashida et al." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.145, + 0.674, + 0.163 + ], + "angle": 0, + "content": "3 Preliminary: DP Imaging and Defocus Blur" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.178, + 0.785, + 0.237 + ], + "angle": 0, + "content": "Unlike conventional sensors, a DP sensor can capture two sub-aperture images by the left and right photodiodes in a single shot. There is a difference between the two images, which is caused by the difference in the PSFs of the left and right DP views [25]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.238, + 0.487, + 0.493 + ], + "angle": 0, + "content": "Figure 2 shows the visualization of the difference between two DP views and illustrated corresponding PSFs. There is no difference between the two views if the depth of a scene point is in focus. When the scene point is out of focus, i.e., located further away or closer to the camera from the focal plane, the disparities between the two views appear in opposite directions. Unlike traditional two-view stereo, the disparity is caused by defocus blurs generated by the PSFs for each view. It is known that the signed defocus blur size \\( b \\in \\mathbb{R} \\) is proportional to the defocus disparity \\( d \\) between two views [10]." + }, + { + "type": "image", + "bbox": [ + 0.499, + 0.238, + 0.787, + 0.386 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.497, + 0.395, + 0.788, + 0.48 + ], + "angle": 0, + "content": "Fig.2: Visualization of DP disparity and illustrations of the blur, i.e., PSF, kernels. The two sub-aperture views convert the defocus blur to disparity, resulting in defocus disparity in opposite directions for near and far scene points. The DP image is from [2]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.495, + 0.785, + 0.571 + ], + "angle": 0, + "content": "DP imaging makes the estimation of defocus blur size \\( b \\) tractable [25] without requiring external information or learning-based methods. Since DP PSFs \\( (\\mathbf{H}_l, \\mathbf{H}_r) \\) can be well approximated by a parametric function of blur size \\( b \\), when DP image patches \\( (\\mathbf{G}_l, \\mathbf{G}_r) \\) are given, the blur size can be estimated per pixel by a nonlinear minimization in an unsupervised manner as" + }, + { + "type": "equation", + "bbox": [ + 0.354, + 0.582, + 0.785, + 0.607 + ], + "angle": 0, + "content": "\\[\nb = \\underset {b} {\\operatorname {a r g m i n}} \\| \\mathbf {G} _ {l} * \\mathbf {H} _ {r} (b) - \\mathbf {G} _ {r} * \\mathbf {H} _ {l} (b) \\| _ {F}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.618, + 0.772, + 0.634 + ], + "angle": 0, + "content": "where \\( * \\) is a convolution operator. The readers are referred to [25] for details." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.657, + 0.427, + 0.674 + ], + "angle": 0, + "content": "4 Proposed Method" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Our method resolves the scale ambiguity in multi-view 3D reconstruction using multi-view DP images. The overview is illustrated in Fig. 3. Our method infers the unknown scene scale just by capturing multi-view images with a DP camera. DP images provide a useful cue for blur size estimation, as described above, without requiring any external information. Our method shares the same spirit with [22]; however, there is a key difference in that our method does not require pre-training but capitalizes on the DP sensor's effectiveness in determining blur sizes. We show in Fig. 3 that the estimated blur size and (inverse) depth from SfM and MVS have an affine relationship, and thus, the unknown scene scale and focus distance can be estimated simultaneously by linear regression." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.261, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "Resolving Scale Ambiguity in MV3D Reconstruction using DP Sensors" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.145, + 0.331, + 0.287 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.332, + 0.146, + 0.44, + 0.287 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.44, + 0.147, + 0.568, + 0.287 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.568, + 0.147, + 0.67, + 0.287 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.67, + 0.147, + 0.785, + 0.287 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.302, + 0.785, + 0.371 + ], + "angle": 0, + "content": "Fig. 3: Overview of our method that provides the absolute scale to multi-view 3D reconstruction (SfM and MVS) just by capturing the scene using DP cameras. (a) The disparity between DP views provides the cue to estimate the defocus blur size. (b) From the affine relation between the (inverse) depth and the defocus blur size, our method simultaneously estimates the scene scale and focus distances." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.403, + 0.784, + 0.465 + ], + "angle": 0, + "content": "We begin by describing the ambiguities that arise in our setting, then describe the solution method for resolving them. We also present a method for selecting reliable pixels and views for scale estimation, which is based on the inference by an off-the-shelf blur size estimation method [25]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.483, + 0.785, + 0.573 + ], + "angle": 0, + "content": "Assumptions We assume that camera intrinsics, such as focal lengths and aperture sizes (f-numbers), are known as they are usually accessible from the camera system. Although defocus blurs slightly change the center of projection, we assume the change is sufficiently small and ignore their effect on the intrinsics. As stated, we use DP images taken from multiple viewpoints as input. Our method does not assume pre-calibrated viewpoints nor reference objects in a scene." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.6, + 0.576, + 0.615 + ], + "angle": 0, + "content": "4.1 Scale and Focus Distance Ambiguities" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.629, + 0.784, + 0.659 + ], + "angle": 0, + "content": "There are two ambiguities in our setting: One is the scale ambiguity in SfM and MVS, and the other is the focus distance ambiguity in DP images." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.677, + 0.788, + 0.799 + ], + "angle": 0, + "content": "Scale ambiguity in SfM and MVS Camera positions and 3D point locations recovered by SfM and the subsequent MVS have a scale ambiguity even with known camera intrinsics. Let us consider a 3D point \\(\\mathbf{p} = [x,y,z]^{\\top}\\) projected to the corresponding 2D point \\(\\mathbf{u} = [u,v,1]^{\\top}\\) in the homogeneous coordinates with a projection matrix \\([\\mathbf{R}|\\mathbf{t}]\\) formed by a rotation matrix \\(\\mathbf{R} \\in SO(3)\\) and a translation vector \\(\\mathbf{t} \\in \\mathbb{R}^3\\). With known camera intrinsics, we can set the intrinsics to be identity and write the 3D-to-2D perspective projection in some Euclidean world coordinate system as" + }, + { + "type": "equation", + "bbox": [ + 0.381, + 0.81, + 0.787, + 0.845 + ], + "angle": 0, + "content": "\\[\n\\mathbf {u} \\sim [ \\mathbf {R} \\mathbf {t} ] \\left[ \\begin{array}{l} \\mathbf {p} \\\\ 1 \\end{array} \\right] = \\frac {1}{s} [ \\mathbf {R} s \\mathbf {t} ] \\left[ \\begin{array}{l} s \\mathbf {p} \\\\ 1 \\end{array} \\right], \\tag {2}\n\\]" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.381, + 0.128 + ], + "angle": 0, + "content": "K. Ashida et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.193 + ], + "angle": 0, + "content": "where \\( s \\in \\mathbb{R} \\) is an unknown scaling factor. As such, the depth \\( z' \\) obtained by SfM and MVS is linearly related to the true depth \\( z \\) by the unknown scaling factor \\( s \\) as" + }, + { + "type": "equation", + "bbox": [ + 0.468, + 0.201, + 0.786, + 0.217 + ], + "angle": 0, + "content": "\\[\nz = s z ^ {\\prime}. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.224, + 0.786, + 0.256 + ], + "angle": 0, + "content": "This is the inherent ambiguity in SfM, which we are going to resolve in this paper." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.267, + 0.787, + 0.314 + ], + "angle": 0, + "content": "Focus distance ambiguity in \\(DP\\) images As described by Wadhawa et al. [36] and Garg et al. [10], the signed defocus blur size \\(b \\in \\mathbb{R}\\) is related to scene depth \\(z\\), focal length \\(f\\), lens aperture diameter \\(l\\), and the focus distance \\(g\\) in \\(\\mathbb{R}\\) as" + }, + { + "type": "equation", + "bbox": [ + 0.417, + 0.319, + 0.786, + 0.353 + ], + "angle": 0, + "content": "\\[\nb \\approx \\frac {l f}{1 - f / g} \\left(\\frac {1}{g} - \\frac {1}{z}\\right) \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.358, + 0.788, + 0.435 + ], + "angle": 0, + "content": "using the paraxial and thin-lens approximations. The blur size \\( b \\) in DP images varies with the focus distance \\( g \\) even when the same scene depth \\( z \\) is recorded with the same camera parameters \\( (l, f) \\). Therefore, even with the blur size \\( b \\) from DP images, its relationship to depth \\( z \\) is ambiguous due to unknown focus distance \\( g \\) and depth \\( z \\) as" + }, + { + "type": "equation", + "bbox": [ + 0.331, + 0.441, + 0.786, + 0.475 + ], + "angle": 0, + "content": "\\[\nb \\approx \\frac {l f}{1 - f / g} \\left(\\frac {1}{g} - \\frac {1}{z}\\right) = \\frac {l f}{1 - f / g ^ {\\prime}} \\left(\\frac {1}{g ^ {\\prime}} - \\frac {1}{z ^ {\\prime}}\\right), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.48, + 0.788, + 0.527 + ], + "angle": 0, + "content": "i.e., there exists a set of pairs like \\((g,z)\\) and \\((g^{\\prime},z^{\\prime})\\) that yields the same blur size \\(b\\). The blur size \\(b\\) is signed, i.e., it becomes positive when the scene point is farther away from the focal plane and negative when nearer to the camera." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.545, + 0.403, + 0.559 + ], + "angle": 0, + "content": "4.2 Solution Method" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.567, + 0.788, + 0.657 + ], + "angle": 0, + "content": "Given the blur sizes \\( b_{ij} \\in \\mathbb{R} \\) obtained from DP images, depths \\( z_{ij}' \\in \\mathbb{R}_+ \\) from SfM and MVS in the \\( i \\)-th view at \\( j \\)-th pixel, and known per-view focal lengths \\( f_i \\) and lens aperture \\( l_i \\), our goal is to determine unknown scene scale \\( s \\in \\mathbb{R}_+ \\) and per-view focus distances \\( g_i \\in \\mathbb{R}_+ \\). Since we use per-view intrinsics, our method can be used in a practical setting of multi-view 3D reconstruction, in which camera intrinsics may vary across viewpoints." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.658, + 0.788, + 0.689 + ], + "angle": 0, + "content": "We put together the two ambiguous observations, i.e., depths up to scale and blur sizes with unknown focus distances, by substituting Eq. (3) to Eq. (4) as" + }, + { + "type": "equation", + "bbox": [ + 0.395, + 0.708, + 0.787, + 0.75 + ], + "angle": 0, + "content": "\\[\nb _ {i j} = \\frac {l _ {i} f _ {i}}{1 - f _ {i} / g _ {i}} \\left(\\frac {1}{g _ {i}} - \\frac {1}{s z _ {i j} ^ {\\prime}}\\right). \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.755, + 0.788, + 0.786 + ], + "angle": 0, + "content": "Denoting \\(\\bar{g}_i = \\frac{1}{g_i}\\) and \\(\\bar{s} = \\frac{1}{s}\\) as \\(g_{i}\\) and \\(s\\) are both non-zero, Eq. (6) can be rewritten as" + }, + { + "type": "equation", + "bbox": [ + 0.353, + 0.79, + 0.787, + 0.838 + ], + "angle": 0, + "content": "\\[\n\\frac {b _ {i j} z _ {i j} ^ {\\prime}}{f _ {i}} = \\underbrace {z _ {i j} ^ {\\prime} \\left(b _ {i j} + l _ {i}\\right)} _ {\\gamma_ {i j}} \\bar {g} _ {i} - l _ {i} \\bar {s} = \\gamma_ {i j} \\bar {g} _ {i} - l _ {i} \\bar {s}, \\tag {7}\n\\]" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.261, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "Resolving Scale Ambiguity in MV3D Reconstruction using DP Sensors" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "image", + "bbox": [ + 0.229, + 0.144, + 0.357, + 0.21 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.244, + 0.211, + 0.338, + 0.223 + ], + "angle": 0, + "content": "(a) Input image" + }, + { + "type": "image", + "bbox": [ + 0.362, + 0.144, + 0.493, + 0.21 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.369, + 0.211, + 0.49, + 0.222 + ], + "angle": 0, + "content": "(b) Confidence score" + }, + { + "type": "image", + "bbox": [ + 0.506, + 0.143, + 0.639, + 0.21 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.502, + 0.211, + 0.647, + 0.222 + ], + "angle": 0, + "content": "(c) Local depth variation" + }, + { + "type": "image", + "bbox": [ + 0.653, + 0.144, + 0.78, + 0.21 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.659, + 0.211, + 0.769, + 0.222 + ], + "angle": 0, + "content": "(d) Selected pixels" + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.234, + 0.787, + 0.304 + ], + "angle": 0, + "content": "Fig. 4: Pixel selection within a view based on confidence scores. We select reliable pixels that contain good blur size estimates from an input image (a). (b) shows the log-scaled confidence scores \\( c \\) that show the reliability of blur size estimates [25]. (c) Local variations (variance) of depths. (d) Selected pixels based on our pixel selection method (highlighted in the figure)." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.324, + 0.687, + 0.339 + ], + "angle": 0, + "content": "which is linear with respect to the unknown parameters \\(\\bar{g}_i\\) and \\(\\bar{s}\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.34, + 0.785, + 0.399 + ], + "angle": 0, + "content": "From observations of multiple views and pixels, we can construct an overdetermined system of linear equations from Eq. (7). Given the number of views \\( n \\) and the number of valid pixels \\( p_i \\) in the \\( i \\)-th view, we have \\( \\sum_{i=1}^{n} p_i \\) equations, which can be written in a matrix form as:" + }, + { + "type": "equation", + "bbox": [ + 0.356, + 0.409, + 0.787, + 0.571 + ], + "angle": 0, + "content": "\\[\n\\underbrace {\\left[ \\begin{array}{c c c c c} \\gamma_ {1 1} & & & & - l _ {1} \\\\ \\vdots & & & & \\vdots \\\\ \\gamma_ {1 p _ {1}} & & 0 & & - l _ {1} \\\\ & \\gamma_ {2 1} & & & - l _ {2} \\\\ & \\vdots & & & \\vdots \\\\ 0 & \\vdots & 0 & & \\vdots \\\\ & \\gamma_ {2 p _ {2}} & & & - l _ {2} \\\\ & & \\ddots & & \\vdots \\\\ & & & \\gamma_ {n 1} & - l _ {n} \\\\ & 0 & & \\vdots & \\vdots \\\\ & & & \\gamma_ {n p _ {n}} & - l _ {n} \\end{array} \\right]} _ {\\mathbf {A}} \\underbrace {\\left[ \\begin{array}{c} \\bar {g} _ {1} \\\\ \\bar {g} _ {2} \\\\ \\vdots \\\\ \\bar {g} _ {n} \\\\ \\bar {s} \\end{array} \\right]} _ {\\mathbf {x}} = \\underbrace {\\left[ \\begin{array}{c} b _ {1 1} z _ {1 1} ^ {\\prime} / f _ {1} \\\\ \\vdots \\\\ b _ {1 p _ {1}} z _ {1 p _ {1}} ^ {\\prime} / f _ {1} \\\\ b _ {2 1} z _ {2 1} ^ {\\prime} / f _ {2} \\\\ \\vdots \\\\ b _ {2 p _ {2}} z _ {2 p _ {2}} ^ {\\prime} / f _ {2} \\\\ \\vdots \\\\ b _ {n 1} z _ {n 1} ^ {\\prime} / f _ {n} \\\\ \\vdots \\\\ b _ {n p _ {n}} z _ {n p _ {n}} ^ {\\prime} / f _ {n} \\end{array} \\right]} _ {\\mathbf {b}}. \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.58, + 0.787, + 0.67 + ], + "angle": 0, + "content": "Since we have \\( n + 1 \\) unknowns in vector \\( \\mathbf{x} \\), given \\( \\mathbf{b} \\neq \\mathbf{0} \\), the condition for obtaining a unique (approximate) solution is \\( \\mathrm{rank}(\\mathbf{A}) = n + 1 \\), which is, in practice, easy to achieve because a large number of pixels are available. The minimum setting to determine scale \\( s \\) is selecting two pixels with distinct depths in one of the multi-view images, which makes a full-rank \\( 2 \\times 2 \\) matrix \\( \\mathbf{A} \\) for two unknowns \\( \\mathbf{x} = [\\bar{g}_i, \\bar{s}]^\\top \\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.671, + 0.787, + 0.731 + ], + "angle": 0, + "content": "In practice, we can select many pixels from multiple views to rely on robust estimation. With a sufficiently large number of selected pixels, we construct a tall and skinny matrix \\(\\mathbf{A}\\) and use \\(\\ell_1\\) residual minimization to effectively disregard outliers as" + }, + { + "type": "equation", + "bbox": [ + 0.409, + 0.744, + 0.786, + 0.768 + ], + "angle": 0, + "content": "\\[\n\\mathbf {x} ^ {*} = \\underset {\\mathbf {x}} {\\operatorname {a r g m i n}} \\| \\mathbf {A x} - \\mathbf {b} \\| _ {1}. \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.787, + 0.84 + ], + "angle": 0, + "content": "We use the Iteratively Reweighted Least Squares (IRLS) method [12] for deriving the approximate solution \\(\\mathbf{x}^*\\). Obviously, other choices of robust estimation methods, such as RANdom SAmple Consensus (RANSAC) [8], are viable for alleviating the impact of outliers." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.381, + 0.127 + ], + "angle": 0, + "content": "K. Ashida et al." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.147, + 0.533, + 0.162 + ], + "angle": 0, + "content": "4.3 Confidence-based Pixel Selection" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.171, + 0.788, + 0.293 + ], + "angle": 0, + "content": "So far, we have assumed that the blur sizes \\( b_{ij} \\) are available to our method. For obtaining the blur sizes, we use the method proposed by Punnappurath et al. [25] that optimizes Eq. (1). For our case, we only require blur size estimates for a small number of pixels but not for all the pixels since our goal is determining the scene scale \\( s \\) using Eq. (8). We, therefore, develop a pixel selection heuristic to choose a small number of pixels that have reliable blur size estimates. Specifically, we select the reliable pixels in each view, followed by selecting the views that produce reliable estimates of scene scale." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.307, + 0.788, + 0.397 + ], + "angle": 0, + "content": "Pixel-selection within a view We begin by adopting the confidence scores of [25] that indicate the reliability of the blur size estimates based on the residual of the blur size optimization and the textured-ness (magnitude of image gradient) of the pixel of interest. We slightly modified their original confidence score using a depth-based criterion since we observed that the large depth variation in an image patch negatively affects the blur size estimation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.398, + 0.788, + 0.473 + ], + "angle": 0, + "content": "Let us consider one of the views and let \\( c_{j} \\) be the confidence score defined in [25] at the \\( j \\)-th pixel. In our case, a depth value \\( z_{j}^{\\prime} \\) up to scale is available at the pixel, and we can assess the variance of the depth values within an image patch \\( \\mathcal{P}_j \\) centered at the \\( j \\)-th pixel as \\( \\text{var}(\\{z_k'\\}) \\), \\( k \\in \\mathcal{P}_j \\). Based on the variance of the depths, we re-define the confidence score as" + }, + { + "type": "equation", + "bbox": [ + 0.405, + 0.484, + 0.787, + 0.513 + ], + "angle": 0, + "content": "\\[\nc _ {j} ^ {*} \\triangleq \\frac {c _ {j}}{\\operatorname {v a r} \\left(\\left\\{z _ {k} ^ {\\prime} \\right\\}\\right)}, \\quad k \\in \\mathcal {P} _ {j}. \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.522, + 0.788, + 0.628 + ], + "angle": 0, + "content": "The greater depth variation in an image patch leads to a larger variance and lower confidence score, and vice versa. Based on the new confidence score \\( c_{j}^{*} \\), we select the top \\( T_{c}\\% \\) of pixels and regard them as reliable pixels that have good defocus blur estimates. Although the variance of depths varies with the depth scale, which is unknown in our case, it does not affect the confidence score \\( c_{j}^{*} \\) within a view because the same scale is applied to all the pixels in the view. Figure 4 depicts the pixel selection within a view." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.642, + 0.788, + 0.702 + ], + "angle": 0, + "content": "View selection In the multi-view images for SfM, there may be images that are not suitable for scale estimation. For example, a view with small depth variations is not preferred because the small depth variations make the linear system of Eq. (8) unstable. Therefore, we further select reliable views for scale estimation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.703, + 0.788, + 0.839 + ], + "angle": 0, + "content": "Since our method can estimate scale \\( s \\) for each view as long as the depths up to scale are available, we first compute the scale predictions for each view independently using the selected pixels within each view by solving Eq. (8). Ideally, the scales estimated for each view should match; however, in reality, they do not due to outliers. We first disregard the views with clear outliers, i.e., views with excessively small blur size variations and views that yield negative values of scale estimates. The small blur size variations indicate that the scene contains small depth variations. We simply set a threshold to discard the views with small blur size variations. Namely, when the difference between maximum" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.26, + 0.115, + 0.733, + 0.13 + ], + "angle": 0, + "content": "Resolving Scale Ambiguity in MV3D Reconstruction using DP Sensors" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.432, + 0.326 + ], + "angle": 0, + "content": "and minimum blur size estimates becomes less than or equal to \\( T_{p} \\) pixels, the view is unused for the estimation. The negative scale estimates could arise due to various reasons, such as image degradation by motion blur and inaccurate estimates of either or both blur sizes and depths. We simply discard such views from the estimation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.329, + 0.432, + 0.358 + ], + "angle": 0, + "content": "Finally, from the remaining views, we select \\(N_{v}\\) views" + }, + { + "type": "image", + "bbox": [ + 0.445, + 0.149, + 0.54, + 0.287 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.543, + 0.15, + 0.687, + 0.288 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.688, + 0.159, + 0.787, + 0.287 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.441, + 0.3, + 0.788, + 0.342 + ], + "angle": 0, + "content": "Fig. 5: View selection. We select the reliable views for scale estimation based on the median of the scale values estimated for each view." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.359, + 0.788, + 0.404 + ], + "angle": 0, + "content": "that are close to the median of the estimated scales as illustrated in Fig. 5. Once the appropriate views are selected, by using all the selected pixels within the views, we create the linear system of Eq. (8) and re-estimate the scale \\( s \\)." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.428, + 0.457, + 0.443 + ], + "angle": 0, + "content": "4.4 Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.454, + 0.788, + 0.56 + ], + "angle": 0, + "content": "For SfM and MVS, we use the commercial software Metashape1. For DP blur size estimation, we use the method proposed by Punnappurath et al. [25] as described, except for the smoothing operation implemented as post-processing. All captured images are undistorted using the parameters estimated by MVS before being fed into the blur size estimation. Our method is not restricted by the choice of SfM+MVS and blur size estimation methods, but we found these to be the best choices today." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.561, + 0.788, + 0.607 + ], + "angle": 0, + "content": "For the focal length \\( f \\), we use the one from the lens specification. The aperture diameter \\( l \\) is calculated using f-number \\( N \\) of the lens as \\( l = \\frac{f}{N} \\). For the hyperparameters, we use \\( T_{c} = 10 \\), \\( T_{p} = 2 \\), and \\( N_{v} = 7 \\) throughout the experiments." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.63, + 0.367, + 0.648 + ], + "angle": 0, + "content": "5 Experiment" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.664, + 0.788, + 0.725 + ], + "angle": 0, + "content": "We assess the accuracy of our scale estimation method both quantitatively and qualitatively. We use multi-view DP images captured in six scenes with different capturing settings, such as, cameras (DSLR and smartphone), lenses, and aperture sizes." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.74, + 0.788, + 0.817 + ], + "angle": 0, + "content": "Recording setting We use two different types of cameras equipped with a DP sensor: a DSLR camera and a smartphone. For the DSLR camera setting, we use Canon EOS 5D Mark IV with three different lenses: EF35mm F1.4L II USM, EF50mm F1.8 STM, and EF85mm F1.8 USM. For each scene, we also change the aperture size to assess the effect of the aperture size on our method. The" + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.825, + 0.761, + 0.841 + ], + "angle": 0, + "content": "1 Agisoft Metashape https://www.agisoft.com/, last accessed on July 12, 2024." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.381, + 0.127 + ], + "angle": 0, + "content": "K. Ashida et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.253 + ], + "angle": 0, + "content": "maximum aperture (\\(f/1.4\\) or \\(f/1.8\\) depending on the lens), \\(f/4.0\\), and \\(f/8.0\\) are used for the experiment depending on the target scene. The images from Canon's DLSR camera are recorded as Canon Raw images with the CR2 format, and the left and right DP views are extracted using DPRSplit2. For a more accessible setup, we also use a smartphone, Google Pixel 4, which has a focal length of \\(4.38 \\mathrm{~mm}\\) and a fixed aperture of \\(f/1.73\\). To obtain the left and right DP views, we use the capturing software used in [11]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.253, + 0.788, + 0.343 + ], + "angle": 0, + "content": "For each scene for each camera setting, we capture about 30 images using a DSLR, and 60 images using a smartphone. For fair comparisons across different lens and aperture sizes, for each viewpoint, we use a tripod to keep the camera fixed and record images with different camera settings. To evaluate the scale estimates, we obtain the ground-truth scene scale by placing markers with known sizes in the scenes." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.359, + 0.788, + 0.401 + ], + "angle": 0, + "content": "Evaluation metric As a quantitative evaluation metric, we use the scale ratio \\( r_s \\) of the estimated scene scale \\( s_{\\mathrm{est}} \\) to the ground-truth scale \\( s_{\\mathrm{gt}} \\), which is defined as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.423, + 0.401, + 0.581, + 0.43 + ], + "angle": 0, + "content": "\\[\n\\mathrm {S c a l e r a t i o} r _ {s} = \\frac {s _ {\\mathrm {e s t}}}{s _ {\\mathrm {g t}}}.\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.435, + 0.788, + 0.524 + ], + "angle": 0, + "content": "In the above metric, closer to 1 indicates good scale estimation. The scale ratio smaller than 1 indicates that the scale is under-estimated, i.e., the estimated scale is smaller than the true scale, and the greater scale ratio corresponds to the over-estimation. While scale ratio \\( r_s \\) is defined for each scene and camera setting, we also define the average error of the scale ratios \\( e_s \\) to characterize the overall effectiveness as" + }, + { + "type": "equation", + "bbox": [ + 0.281, + 0.533, + 0.724, + 0.565 + ], + "angle": 0, + "content": "\\[\n\\mathrm {A v e r a g e e r r o r o f s c a l e r a t i o s} e _ {s} = \\frac {1}{n _ {s}} \\sum \\left(\\max (r _ {s}, r _ {s} ^ {- 1}) - 1\\right),\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.575, + 0.787, + 0.606 + ], + "angle": 0, + "content": "where \\( n_s \\) is the number of settings, and \\( \\max(r_s, r_s^{-1}) \\) is for taking into account the under- and over-estimations." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.621, + 0.788, + 0.743 + ], + "angle": 0, + "content": "Quantitative results The results of the scale estimation for six scenes using two DP cameras are summarized in Table 1. The DSLR camera uses three lenses and different aperture sizes. Overall, our method can reasonably well estimate the absolute scene scale just by using a DP camera for multi-view capture, considering the fact that scale ratio \\( r_s \\) can range in \\( (0, \\infty) \\). The average error \\( e_s \\) across scenes and lenses is approximately 0.219 as shown in the table. From comparing the average errors for each lens, we see that the difference in lenses does not significantly affect the overall performance." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.757, + 0.788, + 0.804 + ], + "angle": 0, + "content": "Visual results Figures 6 and 7 show visual examples of our results by the DSLR and smartphone, as well as the estimated blur sizes and the ground-truth depth maps. We can confirm the same trend as in the quantitative results, where the" + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.81, + 0.787, + 0.839 + ], + "angle": 0, + "content": "\\(^{2}\\) DPRSplit https://www.fastrawviewer.com/DPRSplit, last accessed on July 12, 2024." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.261, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Resolving Scale Ambiguity in MV3D Reconstruction using DP Sensors" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "image", + "bbox": [ + 0.219, + 0.159, + 0.785, + 0.688 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.7, + 0.788, + 0.827 + ], + "angle": 0, + "content": "Fig. 6: Visual results of real scenes captured by the DSLR (Canon EOS 5D Mark IV). Our method yields reasonable estimates of scene scales just by using a DP camera for multi-view capture, which is theoretically unrecoverable by standard SfM alone. The top three rows show the results of the same scene (and from almost the same viewpoint) but with different aperture sizes or focal lengths. The bottom rows show the results of different scenes. In (a), the distance indicates the true distance obtained from the GT 3D model. In (d), the distance in the estimated scale is shown. (b) shows estimated blur size and selected pixels based on our pixel selection method (highlighted in the figure)." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.381, + 0.127 + ], + "angle": 0, + "content": "K. Ashida et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.145, + 0.788, + 0.201 + ], + "angle": 0, + "content": "Table 1: Scale ratio \\( r_s \\) between the ground truth and the estimated scale for each scene, lens, and aperture size. Maximum aperture varies depending on the lens (i.e., \\( f/1.4 \\) or \\( f/1.8 \\)). \\( e_s \\) represents the average errors of the corresponding row/column/total. “-” indicates that we did not record the scene under the setting." + }, + { + "type": "table", + "bbox": [ + 0.222, + 0.213, + 0.785, + 0.329 + ], + "angle": 0, + "content": "
DSLRR (35mm)DSLRR (50mm)DSLRR (85mm)Smartphone f/1.73es
f/1.4f/4.0f/8.0f/1.8f/4.0f/8.0f/1.8f/4.0f/8.0
Scene 11.4191.088-1.0841.084-1.1001.231-1.0850.156
Scene 21.2251.238-1.0471.232-1.2471.302-1.2520.220
Scene 31.1991.229-1.0551.352-1.2381.226-1.5190.260
Scene 41.1811.1071.4771.0271.2661.4691.2251.1901.2961.0340.227
Scene 51.1711.0721.3141.0671.0991.2851.0381.2051.4251.0680.174
Scene 61.1481.073-1.0831.029-1.1231.126-1.3880.139
es0.2240.1350.3960.0610.1770.3770.1620.2130.3610.2240.219
" + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.351, + 0.788, + 0.523 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.538, + 0.788, + 0.581 + ], + "angle": 0, + "content": "Fig. 7: Visual results of real scenes captured by a smartphone (Google Pixel 4). Similar to the results by DSLR, our method estimates reasonable scene scales despite the DP blur being relatively smaller than DSLR's." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.616, + 0.788, + 0.691 + ], + "angle": 0, + "content": "depth maps with estimated scale well represent the actual depths. Our method estimates the scene scale with less than 0.2 error ratio for several settings, where the blur size estimation is relatively stable. For these cases, 3D reconstruction with our absolute scale estimation allows high-quality measurements with an error of several centimeters in wide outdoor scenes." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.788, + 0.813 + ], + "angle": 0, + "content": "Stability of our method To assess the stability of our method against variation of input images, we capture additional \\(200\\sim 300\\) images for Scenes 1 and 2 using the DSLR equipped with a \\(35\\mathrm{mm}\\) lens and \\(f / 1.4\\) aperture size and repeatedly estimate the scale using different sets of images. Figure 8 shows an overlay of estimated models with repeated experiments and the minimum and maximum estimated scale ratios, \\(r_{s\\min}\\) and \\(r_{s\\max}\\), with 10 trials." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.26, + 0.114, + 0.733, + 0.131 + ], + "angle": 0, + "content": "Resolving Scale Ambiguity in MV3D Reconstruction using DP Sensors" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "image", + "bbox": [ + 0.224, + 0.139, + 0.451, + 0.243 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.258, + 0.248, + 0.41, + 0.26 + ], + "angle": 0, + "content": "\\(r_{s\\min}\\) : 1.158, \\(r_{s\\max}\\) : 1.255" + }, + { + "type": "image", + "bbox": [ + 0.453, + 0.139, + 0.621, + 0.244 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.455, + 0.248, + 0.607, + 0.26 + ], + "angle": 0, + "content": "\\(r_{s\\min}\\) : 1.174, \\(r_{s\\max}\\) : 1.285" + }, + { + "type": "image", + "bbox": [ + 0.623, + 0.139, + 0.772, + 0.244 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.622, + 0.248, + 0.776, + 0.261 + ], + "angle": 0, + "content": "\\(r_{s\\min}\\) : 1.202, \\(r_{s\\max}\\) : 1.383" + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.271, + 0.788, + 0.314 + ], + "angle": 0, + "content": "Fig. 8: Overlay of estimated models with repeated experiments and the minimum and maximum estimated scale ratios with 10 trials. Blue and red point clouds represent the ones with scale ratios \\( r_{s\\max} \\) and \\( r_{s\\min} \\), respectively." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.343, + 0.501, + 0.524 + ], + "angle": 0, + "content": "Comparison with scale estimation by monocular depth estimation We compare our method with a single-image metric depth estimation, ZoeDepth [6], for the same scenes and lens setup used. Since it estimates metric depth for each pixel, we compute the per-image scale as the average ratio between ZoeDepth and MVS depth maps. We then average the per-image scale over seven inlier viewpoints selected using the same procedure as our view selection." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.525, + 0.5, + 0.584 + ], + "angle": 0, + "content": "The average error \\( e_{s} \\) across scenes and lenses using ZoeDepth is 0.679, which is more than 3 times larger than ours (0.219). Figure 9 confirms this" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.585, + 0.787, + 0.616 + ], + "angle": 0, + "content": "trend through the visualization of the estimated scale. More detailed comparisons are shown in the supplementary material." + }, + { + "type": "image", + "bbox": [ + 0.55, + 0.348, + 0.752, + 0.454 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.558, + 0.457, + 0.741, + 0.47 + ], + "angle": 0, + "content": "\\(r_{sours}\\) : 1.038, \\(r_{sZoe\\text{depth}}\\) : 0.643" + }, + { + "type": "image_caption", + "bbox": [ + 0.509, + 0.48, + 0.789, + 0.565 + ], + "angle": 0, + "content": "Fig. 9: Comparison with monocular scale estimation. Blue and red points represent the ones of the scale ratio estimated by our method (\\(r_{s\\text{ours}}\\)) and ZoeDepth (\\(r_{sZoeDepth}\\)), respectively. Black points represent the GT model." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.629, + 0.788, + 0.78 + ], + "angle": 0, + "content": "Limitations and failure cases Although our method successfully estimates the scene scale in many cases, we find several failure modes in our experiments. Since our method is based on defocus blur, scale estimation from pan-focus images becomes challenging. In particular, since the aperture of smartphone cameras is smaller than that of DSLR's, it was necessary for us to include photographs that exhibit large depth variations to produce relatively large blur like the top row of Fig. 7. Although the pan-focus images are preferable in SfM and MVS for better correspondence matching, in practice, small apertures cannot always be used. This is because the small aperture limits the amount of incoming light, requiring longer exposure times, and resulting in camera motion blurs." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.841 + ], + "angle": 0, + "content": "The success of our method relies on the accurate estimation of both defocus blur sizes and depth maps up to scale. We found several examples where the scale estimation is affected by the blur size estimation errors, while the depth maps up to scale are almost always stably estimated. Figure 10 shows a failure" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.381, + 0.127 + ], + "angle": 0, + "content": "K. Ashida et al." + }, + { + "type": "image", + "bbox": [ + 0.225, + 0.144, + 0.34, + 0.204 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.233, + 0.206, + 0.328, + 0.217 + ], + "angle": 0, + "content": "(a) Input image" + }, + { + "type": "image", + "bbox": [ + 0.346, + 0.143, + 0.485, + 0.204 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.375, + 0.206, + 0.457, + 0.217 + ], + "angle": 0, + "content": "(b) GT depth" + }, + { + "type": "image", + "bbox": [ + 0.496, + 0.143, + 0.631, + 0.204 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.496, + 0.206, + 0.63, + 0.216 + ], + "angle": 0, + "content": "(c) Estimated blur size" + }, + { + "type": "image", + "bbox": [ + 0.638, + 0.143, + 0.78, + 0.204 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.657, + 0.206, + 0.763, + 0.216 + ], + "angle": 0, + "content": "(d) Ideal blur size" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.221, + 0.788, + 0.306 + ], + "angle": 0, + "content": "Fig. 10: An example of failure case (Scene 3 captured with a \\(50\\mathrm{mm}\\) lens and \\(f / 4.0\\) aperture size) producing scale ratio \\(r_s = 1.352\\) due to the large error in blur size estimation. Ideal blur sizes (d) are calculated from the ground-truth (GT) scene depth (b) and Eq. (6). Since the target scene is mostly textureless or a repetition of fine textures, most of the pixels in the estimated blur sizes (c) deviate from ideal values, making the selection of the inlier pixels challenging." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.332, + 0.789, + 0.483 + ], + "angle": 0, + "content": "mode of blur size estimation as well as the estimated scene depth, whose scale ratio \\( r_s \\) reaches as large as 1.352 (Scene 3 captured with \\( 50\\mathrm{mm} \\) lens and \\( f / 4.0 \\) aperture size). To assess the quality of blur size estimates shown in Fig. 10 (c), we synthesize the ideal blur sizes in Fig. 10 (d) using the ground-truth depth of Fig. 10 (b) that is computed from the markers and the image formation model of Eq. (6). Most of the blur size estimates are far from the ideal blur sizes due to the target object being almost textureless or the repetition of fine-grained texture, which makes blur size optimization challenging. We further analyze the accuracy and characteristics of the existing blur size estimation methods in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.497, + 0.364, + 0.511 + ], + "angle": 0, + "content": "6 Discussions" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.518, + 0.789, + 0.699 + ], + "angle": 0, + "content": "This paper has shown that the scale information in multi-view 3d reconstruction, which is theoretically unsolvable by SfM and MVS itself, can be inferred from the defocus blur information provided by DP imaging. Our method estimates the absolute scale without relying on reference objects or data priors but by simply capturing multi-view images with a DP camera. The key to the scale recovery is the unique feature of DP that relates defocus disparity, blur size, and scene depth, from which we can affinely relate the reciprocal of the depth and the defocus blur sizes given by DP images. Together with the pixel selection method for improving robustness using the confidence information from the off-the-shelf DP blur estimation method as well as the statistics of the depth information, we simultaneously recover the scene scale and the unknown focus distance via a simple linear regression." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.7, + 0.789, + 0.79 + ], + "angle": 0, + "content": "Experiments show that our method successfully estimates the absolute scale in many cases. Since the estimation accuracy of the blur size strongly affects the performance of our method, the analysis of DP blur characteristics will be beneficial in the future. With a method for a more accurate and stabler estimation of blur sizes in DP images, our method will produce more accurate absolute scales for multi-view 3D reconstruction." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.804, + 0.395, + 0.82 + ], + "angle": 0, + "content": "Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.825, + 0.757, + 0.84 + ], + "angle": 0, + "content": "This work was supported by JSPS KAKENHI Grant Number JP23H05491." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.261, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Resolving Scale Ambiguity in MV3D Reconstruction using DP Sensors" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.323, + 0.16 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.173, + 0.785, + 0.214 + ], + "angle": 0, + "content": "1. Abuolaim, A., Afifi, M., Brown, M.S.: Improving single-image defocus deblurring: How dual-pixel images help through multi-task learning. In: IEEE/CVF Winter Conference on Applications of Computer Vision (WACV). pp. 82-90 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.215, + 0.785, + 0.242 + ], + "angle": 0, + "content": "2. Abuolaim, A., Brown, M.S.: Defocus deblurring using dual-pixel data. In: Proceedings of European Conference on Computer Vision (ECCV). pp. 111-126 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.242, + 0.785, + 0.296 + ], + "angle": 0, + "content": "3. Abuolaim, A., Delbracio, M., Kelly, D., Brown, M.S., Milanfar, P.: Learning to reduce defocus blur by realistically modeling dual-pixel data. In: Proceedings of IEEE/CVF International Conference on Computer Vision (ICCV). pp. 2289-2298 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.296, + 0.785, + 0.337 + ], + "angle": 0, + "content": "4. Abuolaim, A., Punnappurath, A., Brown, M.S.: Revisiting autofocus for smartphone cameras. In: Proceedings of European Conference on Computer Vision (ECCV). pp. 545-559 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.337, + 0.785, + 0.393 + ], + "angle": 0, + "content": "5. Abuolaim, A., Timofte, R., Brown, M.S.: NTIRE 2021 challenge for defocus deblurring using dual-pixel images: Methods and results. In: Proceedings of IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW). pp. 578-587 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.393, + 0.785, + 0.419 + ], + "angle": 0, + "content": "6. Bhat, S.F., Birkl, R., Wofk, D., Wonka, P., Müller, M.: Zoedepth: Zero-shot transfer by combining relative and metric depth. arXiv preprint arXiv:2302.12288 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.419, + 0.785, + 0.461 + ], + "angle": 0, + "content": "7. Engel, J., Stuckler, J., Cremers, D.: Large-scale direct SLAM with stereo cameras. In: IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). pp. 1935-1942 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.461, + 0.785, + 0.5 + ], + "angle": 0, + "content": "8. Fischler, M.A., Bolles, R.C.: Random sample consensus: A paradigm for model fitting with applications to image analysis and automated cartography. Communications of the ACM 24(6), 381-395 (1981)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.501, + 0.785, + 0.542 + ], + "angle": 0, + "content": "9. Frost, D., Prisacariu, V., Murray, D.: Recovering stable scale in monocular SLAM using object-supplemented bundle adjustment. IEEE Transactions on Robotics 34(3), 736-747 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.542, + 0.785, + 0.582 + ], + "angle": 0, + "content": "0. Garg, R., Wadhwa, N., Ansari, S., Barron, J.T.: Learning single camera depth estimation using dual-pixels. In: Proceedings of IEEE/CVF International Conference on Computer Vision (ICCV). pp. 7628-7637 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.582, + 0.785, + 0.624 + ], + "angle": 0, + "content": "1. Garg, R., Wadhwa, N., Ansari, S., Barron, J.T.: Learning single camera depth estimation using dual-pixels. In: Proceedings of IEEE/CVF International Conference on Computer Vision (ICCV). pp. 7628-7637 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.624, + 0.785, + 0.637 + ], + "angle": 0, + "content": "2. Gentle, J.: Matrix Albegra. Springer Texts in Statistics, Springer, New York (2007)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.637, + 0.785, + 0.677 + ], + "angle": 0, + "content": "3. Giubilato, R., Chiodini, S., Pertile, M., Debei, S.: Scale correct monocular visual odometry using a lidar altimeter. In: IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). pp. 3694-3700 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.677, + 0.785, + 0.718 + ], + "angle": 0, + "content": "4. Herrmann, C., Bowen, R.S., Wadhwa, N., Garg, R., He, Q., Barron, J.T., Zabih, R.: Learning to autofocus. In: Proceedings of IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 2227-2236 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.718, + 0.785, + 0.759 + ], + "angle": 0, + "content": "5. Jung, S.H., Heo, Y.S.: Disparity probability volume guided defocus deblurring using dual pixel data. In: Proceedings of International Conference on Information and Communication Technology Convergence (ICTC). pp. 305-308 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.759, + 0.785, + 0.8 + ], + "angle": 0, + "content": "6. Kang, M., Choe, J., Ha, H., Jeon, H.G., Im, S., Kweon, I.S.: Facial depth and normal estimation using single dual-pixel camera. In: Proceedings of European Conference on Computer Vision (ECCV) (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.8, + 0.785, + 0.841 + ], + "angle": 0, + "content": "7. Kashiwagi, M., Mishima, N., Kozakaya, T., Hiura, S.: Deep depth from aberration map. In: Proceedings of IEEE/CVF International Conference on Computer Vision (ICCV) (2019)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.173, + 0.785, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.381, + 0.127 + ], + "angle": 0, + "content": "K. Ashida et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.785, + 0.189 + ], + "angle": 0, + "content": "18. Kim, D., Jang, H., Kim, I., Kim, M.H.: Spatio-focal bidirectional disparity estimation from a dual-pixel image. In: Proceedings of IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 5023-5032 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.191, + 0.785, + 0.232 + ], + "angle": 0, + "content": "19. Knorr, S.B., Kurz, D.: Leveraging the user's face for absolute scale estimation in handheld monocular SLAM. In: Proceedings of IEEE International Symposium on Mixed and Augmented Reality (ISMAR). pp. 11-17 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.233, + 0.785, + 0.261 + ], + "angle": 0, + "content": "20. Lee, S.H., de Croon, G.: Stability-based scale estimation for monocular SLAM. IEEE Robotics and Automation Letters 3(2), 780-787 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.262, + 0.785, + 0.303 + ], + "angle": 0, + "content": "21. Li, F., Guo, H., Santo, H., Okura, F., Matsushita, Y.: Learning to synthesize photorealistic dual-pixel images from RGBD frames. In: International Conference on Computational Photography (ICCP). pp. 1-11 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.304, + 0.785, + 0.346 + ], + "angle": 0, + "content": "22. Mishima, N., Seki, A., Hiura, S.: Absolute scale from varifocal monocular camera through SfM and defocus combined. In: Proceedings of British Machine Vision Conference (BMVC) (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.347, + 0.785, + 0.388 + ], + "angle": 0, + "content": "23. Nützi, G., Weiss, S., Scaramuzza, D., Siegwart, R.: Fusion of IMU and vision for absolute scale estimation in monocular SLAM. Journal of Intelligent & Robotic Systems 61(1-4), 287-299 (2011)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.389, + 0.785, + 0.445 + ], + "angle": 0, + "content": "24. Pan, L., Chowdhury, S., Hartley, R., Liu, M., Zhang, H., Li, H.: Dual pixel exploration: Simultaneous depth estimation and image restoration. In: Proceedings of IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 4340-4349 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.446, + 0.785, + 0.487 + ], + "angle": 0, + "content": "25. Punnappurath, A., Abuolaim, A., Affi, M., Brown, M.S.: Modeling defocus-disparity in dual-pixel sensors. In: International Conference on Computational Photography (ICCP). pp. 1-12 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.488, + 0.785, + 0.529 + ], + "angle": 0, + "content": "26. Punnappurath, A., Brown, M.S.: Reflection removal using a dual-pixel sensor. In: Proceedings of IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 1556-1565 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.53, + 0.785, + 0.572 + ], + "angle": 0, + "content": "27. Roussel, T., Van Eycken, L., Tuytelaars, T.: Monocular depth estimation in new environments with absolute scale. In: IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). pp. 1735-1741 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.573, + 0.785, + 0.628 + ], + "angle": 0, + "content": "28. Rukhovich, D., Mouritzen, D., Kaestner, R., Rufli, M., Velizhev, A.: Estimation of absolute scale in monocular SLAM using synthetic data. In: Proceedings of IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW). pp. 803-812 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.629, + 0.785, + 0.684 + ], + "angle": 0, + "content": "29. Scaramuzza, D., Fraundorfer, F., Pollefeys, M., Siegwart, R.: Absolute scale in structure from motion from a single vehicle mounted camera by exploiting nonholonomic constraints. In: Proceedings of IEEE/CVF International Conference on Computer Vision (ICCV). pp. 1413-1419 (2009)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.686, + 0.785, + 0.727 + ], + "angle": 0, + "content": "30. Shibata, A., Fujii, H., Yamashita, A., Asama, H.: Absolute scale structure from motion using a refractive plate. In: Proceedings of IEEE/SICE International Symposium on System Integration (SII). pp. 540-545 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.728, + 0.785, + 0.769 + ], + "angle": 0, + "content": "31. Shibata, A., Fujii, H., Yamashita, A., Asama, H.: Scale-reconstructable structure from motion using refraction with a single camera. In: IEEE International Conference on Robotics and Automation (ICRA). pp. 5239-5244 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.77, + 0.785, + 0.798 + ], + "angle": 0, + "content": "32. Shiozaki, T., Dissanayake, G.: Eliminating scale drift in monocular SLAM using depth from defocus. IEEE Robotics and Automation Letters 3(1), 581-587 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.799, + 0.785, + 0.84 + ], + "angle": 0, + "content": "33. Song, S., Chandraker, M.: Robust scale estimation in real-time monocular sfm for autonomous driving. In: Proceedings of IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 1566-1573 (2014)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.261, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Resolving Scale Ambiguity in MV3D Reconstruction using DP Sensors" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.189 + ], + "angle": 0, + "content": "34. Sucar, E., Hayet, J.B.: Bayesian scale estimation for monocular SLAM based on generic object detection for correcting scale drift. In: IEEE International Conference on Robotics and Automation (ICRA). pp. 5152-5158 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.19, + 0.788, + 0.231 + ], + "angle": 0, + "content": "35. Sumikura, S., Sakurada, K., Kawaguchi, N., Nakamura, R.: Scale estimation of monocular SfM for a multi-modal stereo camera. In: Proceedings of Asian Conference on Computer Vision (ACCV). pp. 281-297 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.231, + 0.788, + 0.286 + ], + "angle": 0, + "content": "36. Wadhwa, N., Garg, R., Jacobs, D.E., Feldman, B.E., Kanazawa, N., Carroll, R., Movshovitz-Attias, Y., Barron, J.T., Pritch, Y., Levoy, M.: Synthetic depth-of-field with a single-camera mobile phone. ACM Transactions on Graphics (TOG) pp. 1-13 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.286, + 0.788, + 0.328 + ], + "angle": 0, + "content": "37. Wöhler, C., d'Angelo, P., Krüger, L., Kuhl, A., Groß, H.M.: Monocular 3D scene reconstruction at absolute scale. ISPRS Journal of Photogrammetry and Remote Sensing 64(6), 529-540 (2009)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.328, + 0.788, + 0.383 + ], + "angle": 0, + "content": "38. Xin, S., Wadhwa, N., Xue, T., Barron, J.T., Srinivasan, P.P., Chen, J., Gkioulekas, I., Garg, R.: Defocus map estimation and deblurring from a single dual-pixel image. In: Proceedings of IEEE/CVF International Conference on Computer Vision (ICCV). pp. 2228-2238 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.383, + 0.788, + 0.425 + ], + "angle": 0, + "content": "39. Yang, Y., Pan, L., Liu, L., Liu, M.: K3DN: Disparity-aware kernel estimation for dual-pixel defocus deblurring. In: Proceedings of IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 13263-13272 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.425, + 0.788, + 0.48 + ], + "angle": 0, + "content": "40. Zhang, S., Zhang, J., Tao, D.: Towards scale-aware, robust, and generalizable unsupervised monocular depth estimation by integrating IMU motion dynamics. In: Proceedings of European Conference on Computer Vision (ECCV). pp. 143-160 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.48, + 0.788, + 0.522 + ], + "angle": 0, + "content": "41. Zhang, Y., Wadhwa, N., Orts-Escolano, S., Häne, C., Fanello, S.R., Garg, R.: Du2Net: Learning depth estimation from dual-cameras and dual-pixels. In: Proceedings of European Conference on Computer Vision (ECCV). pp. 582-598 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.522, + 0.788, + 0.563 + ], + "angle": 0, + "content": "42. Zhou, D., Dai, Y., Li, H.: Reliable scale estimation and correction for monocular visual odometry. In: Proceedings of IEEE Intelligent Vehicles Symposium (IV). pp. 490-495 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.563, + 0.788, + 0.605 + ], + "angle": 0, + "content": "43. Zhou, D., Dai, Y., Li, H.: Ground-plane-based absolute scale estimation for monocular visual odometry. IEEE Transactions on Intelligent Transportation Systems 21(2), 791-802 (2019)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.605 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/Resolving Scale Ambiguity in Multi-view 3D Reconstruction using Dual-Pixel Sensors/518914c5-d998-41c3-963e-0840d0e1f2c8_origin.pdf b/2024/Resolving Scale Ambiguity in Multi-view 3D Reconstruction using Dual-Pixel Sensors/518914c5-d998-41c3-963e-0840d0e1f2c8_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..9c8b99ac9b1f1e3313667a2df3f4b21a6781af70 --- /dev/null +++ b/2024/Resolving Scale Ambiguity in Multi-view 3D Reconstruction using Dual-Pixel Sensors/518914c5-d998-41c3-963e-0840d0e1f2c8_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:35cfc8a15a0519e9321052a05590087be08934ee52ff4cf4d334dc1f751c66ef +size 8252267 diff --git a/2024/Resolving Scale Ambiguity in Multi-view 3D Reconstruction using Dual-Pixel Sensors/full.md b/2024/Resolving Scale Ambiguity in Multi-view 3D Reconstruction using Dual-Pixel Sensors/full.md new file mode 100644 index 0000000000000000000000000000000000000000..3b6affb3e072e8b48066f08692cf3ec208cffaeb --- /dev/null +++ b/2024/Resolving Scale Ambiguity in Multi-view 3D Reconstruction using Dual-Pixel Sensors/full.md @@ -0,0 +1,335 @@ +# Resolving Scale Ambiguity in Multi-view 3D Reconstruction using Dual-Pixel Sensors + +Kohei Ashida, Hiroaki Santo, Fumio Okura, and Yasuyuki Matsushita + +Graduate School of Information Science and Technology, + +Osaka University, Japan + +{kohei.ashida, santo.hiroaki, okura, yasumat}@ist.osaka-u.ac.jp + +Abstract. Multi-view 3D reconstruction, namely structure-from-motion and multi-view stereo, is an essential component in 3D computer vision. In general, multi-view 3D reconstruction suffers from unknown scale ambiguity unless a reference object of known size is recorded together with the scene, or the camera poses are pre-calibrated. In this paper, we show that multi-view images recorded by a dual-pixel (DP) sensor allow us to automatically resolve the scale ambiguity without requiring a reference object or pre-calibration. Specifically, the observed defocus blurs in DP images provide sufficient information for determining the scale when paired together with the depth maps (up to scale) recovered from the multi-view 3D reconstruction. Based on this observation, we develop a simple yet effective linear solution method to determine the absolute scale in multi-view 3D reconstruction. Experiments demonstrate the effectiveness of the proposed method with diverse scenes recorded with different cameras/lenses. Code and data are available at https://github.com/kohei-ashida/dp-sfm. + +Keywords: Multi-view 3D reconstruction $\cdot$ Photogrammetry $\cdot$ Dual-pixel imaging $\cdot$ Scale ambiguity + +# 1 Introduction + +Multi-view 3D reconstruction is one of the most important problems in 3D computer vision. From multi-view images of a scene taken from unknown viewpoints, structure-from-motion (SfM) estimates camera poses and sparse 3D tie points of the scene. Subsequently, multi-view stereo (MVS) uses the estimated camera poses to reconstruct the dense shape of the scene. It is known that the scale ambiguity in SfM is inherent even with known camera intrinsics and cannot be resolved without external information. In this setting, the subsequent MVS can only recover dense shapes up to scale. A classical yet most reliable approach to resolving the ambiguity is to place a reference object of known size, such as a calibration target, in the scene to determine the absolute scale. However, it is not always possible to put such a reference object in the scene, and it is desired to relax the demand for diverse application scenarios. + +In this paper, we propose a method for automatically determining the absolute scale in multi-view 3D reconstruction using dual-pixel (DP) cameras whose + +intrinsic are calibrated. DP sensors are designed to assist the camera's auto-focus system and are becoming popular since they are equipped with Canon EOS cameras and Google Pixel smartphones. Each pixel on the DP sensor has two photodiodes to acquire two sub-aperture views. The two sub-aperture views allow us to estimate defocus blurs that are observed at regions off of the focal plane, although they depend on the unknown focus distance. We show that, by relating the DP defocus blur sizes with + +![](images/43caf3e12a49ee35813a2d3581b790a58726370add16cf6431f57b409ff85d2a.jpg) +Fig. 1: Our method determines the absolute scene scale of SfM and MVS from multi-view DP images. We put together the scale ambiguity in SfM and MVS and the focus distance ambiguity of DP images. Our method jointly resolves the ambiguities. + +unknown focus distances with the dense depth maps up to scale obtained from SfM and MVS, the absolute scale of the scene can be fully determined without requiring other external information. + +The key idea of our method is coupling two ambiguous measurements for resolving the ambiguities: One is the linear scale ambiguity of depth values computed from SfM and MVS, and the other is the focus distance ambiguity that affinely relates the reciprocal of depths with DP defocus blurs. As illustrated in Fig. 1, the size of defocus blur $b$ is related to unknown focus distance $g$ and unknown scene scale $s$ , with which scene point $\mathbf{p}$ may be scaled to $\mathbf{sp}$ retaining its projections $\mathbf{u}_1$ and $\mathbf{u}_2$ . By putting them together, our method solves a linear regression problem for determining both the absolute scale of the scene and the focus distances in the multi-view images. Our method capitalizes on the fact that DP's two sub-aperture views faithfully yield defocus blur estimates, significantly more reliably than just a single view from an ordinary camera. + +In summary, the chief contributions of this paper are: + +- We propose a method to determine the absolute scale in uncalibrated multi-view 3D reconstruction using DP images. +- We develop a linear solution method for simultaneously determining the global scene scale and per-view focus distances. +- We conduct comprehensive experiments to assess the effectiveness of the proposed method using diverse scenes and camera settings. + +# 2 Related Work + +Scale estimation in SfM Since SfM alone cannot recover the absolute scale from the image information, it is necessary to introduce an additional cue that provides absolute scale information. A simple and reliable approach is to place a reference object of known size, such as markers and ground control points (GCPs), which is widely employed in commercial photogrammetry software. Domain-specific cues are often used to avoid manual placement of reference objects, for example, urban objects [9,34], ground surfaces [33,42,43], and human faces [19]. To implicitly exploit the scale information of scene objects, recent learning-based methods use monocular depth estimation providing absolute scale information [27,28]. + +Since it is not always possible to assume reference objects in target scenes, external sensors are often used to recover the scale information. The use of stereo cameras with known baselines [7, 35] or using LiDAR [13] are typical examples, while inertial sensors [23, 40] and odometry information [20, 29] are also widely used in robotics applications. Placing a refraction plate in front of the camera provides scale information using multiple capture [30, 31]. + +Similar to our work, defocus has been used to determine the absolute scale in the literature. Calibrated depth-from-defocus is used with SfM-based depths to recover the absolute scale by assuming a known and fixed focus distance [32,37]. Similar to our method, a recent study [22] relaxes the known focus assumption by jointly estimating the physical focal length and scene scale. Although their use of the thin-lens model is similar to ours, their method requires pre-training or calibration for a learning-based blur size estimation [17] that depends on the hardware setup and focal lengths. Our study instead uses uncalibrated defocus blurs, where the focus distance is treated as unknown. The uncalibrated defocus blurs are reliably obtained from DP images, and thereby, our method recovers the absolute scales without prior knowledge, training data, or additional equipment. + +DP imaging and applications Unlike conventional sensors, a single pixel in a DP sensor is divided into two photodiodes, one on the left and one on the right, which produces disparity. Beyond its primary use of fast auto-focus [4,14] implemented in several commercial cameras, applications of DP imaging are actively studied. DP's disparity gives effective cues to estimate the Point Spread Functions (PSFs) of defocus blurs [25], whose direct application is defocus deblurring [1-3,5,15,24,38,39] and blur synthesis [36]. The defocus blur information from DPs improves the performance of monocular depth estimation [10,18,25]. For this application, a DP image dataset with the ground truth depths has been developed [21]. The DP-based depth has been used for domain-specific applications, such as reflection removal [26] and facial shape recovery [16]. A recent method combines DP imaging and traditional two-view stereo for depth estimation [41]. Taking advantage of DP's effectiveness in acquiring defocus blur information, our work attempts to determine the scene scale automatically in multi-view 3D reconstruction. + +# 3 Preliminary: DP Imaging and Defocus Blur + +Unlike conventional sensors, a DP sensor can capture two sub-aperture images by the left and right photodiodes in a single shot. There is a difference between the two images, which is caused by the difference in the PSFs of the left and right DP views [25]. + +Figure 2 shows the visualization of the difference between two DP views and illustrated corresponding PSFs. There is no difference between the two views if the depth of a scene point is in focus. When the scene point is out of focus, i.e., located further away or closer to the camera from the focal plane, the disparities between the two views appear in opposite directions. Unlike traditional two-view stereo, the disparity is caused by defocus blurs generated by the PSFs for each view. It is known that the signed defocus blur size $b \in \mathbb{R}$ is proportional to the defocus disparity $d$ between two views [10]. + +![](images/a14d6a6f2608809c03dbd9527df6f41e995e1080d2268197d370dad5d637eaaa.jpg) +Fig.2: Visualization of DP disparity and illustrations of the blur, i.e., PSF, kernels. The two sub-aperture views convert the defocus blur to disparity, resulting in defocus disparity in opposite directions for near and far scene points. The DP image is from [2]. + +DP imaging makes the estimation of defocus blur size $b$ tractable [25] without requiring external information or learning-based methods. Since DP PSFs $(\mathbf{H}_l, \mathbf{H}_r)$ can be well approximated by a parametric function of blur size $b$ , when DP image patches $(\mathbf{G}_l, \mathbf{G}_r)$ are given, the blur size can be estimated per pixel by a nonlinear minimization in an unsupervised manner as + +$$ +b = \underset {b} {\operatorname {a r g m i n}} \| \mathbf {G} _ {l} * \mathbf {H} _ {r} (b) - \mathbf {G} _ {r} * \mathbf {H} _ {l} (b) \| _ {F}, \tag {1} +$$ + +where $*$ is a convolution operator. The readers are referred to [25] for details. + +# 4 Proposed Method + +Our method resolves the scale ambiguity in multi-view 3D reconstruction using multi-view DP images. The overview is illustrated in Fig. 3. Our method infers the unknown scene scale just by capturing multi-view images with a DP camera. DP images provide a useful cue for blur size estimation, as described above, without requiring any external information. Our method shares the same spirit with [22]; however, there is a key difference in that our method does not require pre-training but capitalizes on the DP sensor's effectiveness in determining blur sizes. We show in Fig. 3 that the estimated blur size and (inverse) depth from SfM and MVS have an affine relationship, and thus, the unknown scene scale and focus distance can be estimated simultaneously by linear regression. + +![](images/e6146d7ce555b3b411ea66fc42ee56abf4c9494ae602382032ff223ce4725c6d.jpg) +Fig. 3: Overview of our method that provides the absolute scale to multi-view 3D reconstruction (SfM and MVS) just by capturing the scene using DP cameras. (a) The disparity between DP views provides the cue to estimate the defocus blur size. (b) From the affine relation between the (inverse) depth and the defocus blur size, our method simultaneously estimates the scene scale and focus distances. + +![](images/4217e975b1151f6f64652270725e964664f15783b355e2b2633da928a07035e2.jpg) + +![](images/eab5002f47f48a54226bd5ff41606f75d61a4c4b5a1cd67e986bef107bdf52bb.jpg) + +![](images/508d66c9ced6e327a9d98d413f8941f78ee320443366a58c851759ffacb80dc3.jpg) + +![](images/42084d35034be71ceacd7af684b36332e248068264def1810a9b52b1d0b914fb.jpg) + +We begin by describing the ambiguities that arise in our setting, then describe the solution method for resolving them. We also present a method for selecting reliable pixels and views for scale estimation, which is based on the inference by an off-the-shelf blur size estimation method [25]. + +Assumptions We assume that camera intrinsics, such as focal lengths and aperture sizes (f-numbers), are known as they are usually accessible from the camera system. Although defocus blurs slightly change the center of projection, we assume the change is sufficiently small and ignore their effect on the intrinsics. As stated, we use DP images taken from multiple viewpoints as input. Our method does not assume pre-calibrated viewpoints nor reference objects in a scene. + +# 4.1 Scale and Focus Distance Ambiguities + +There are two ambiguities in our setting: One is the scale ambiguity in SfM and MVS, and the other is the focus distance ambiguity in DP images. + +Scale ambiguity in SfM and MVS Camera positions and 3D point locations recovered by SfM and the subsequent MVS have a scale ambiguity even with known camera intrinsics. Let us consider a 3D point $\mathbf{p} = [x,y,z]^{\top}$ projected to the corresponding 2D point $\mathbf{u} = [u,v,1]^{\top}$ in the homogeneous coordinates with a projection matrix $[\mathbf{R}|\mathbf{t}]$ formed by a rotation matrix $\mathbf{R} \in SO(3)$ and a translation vector $\mathbf{t} \in \mathbb{R}^3$ . With known camera intrinsics, we can set the intrinsics to be identity and write the 3D-to-2D perspective projection in some Euclidean world coordinate system as + +$$ +\mathbf {u} \sim [ \mathbf {R} \mathbf {t} ] \left[ \begin{array}{l} \mathbf {p} \\ 1 \end{array} \right] = \frac {1}{s} [ \mathbf {R} s \mathbf {t} ] \left[ \begin{array}{l} s \mathbf {p} \\ 1 \end{array} \right], \tag {2} +$$ + +where $s \in \mathbb{R}$ is an unknown scaling factor. As such, the depth $z'$ obtained by SfM and MVS is linearly related to the true depth $z$ by the unknown scaling factor $s$ as + +$$ +z = s z ^ {\prime}. \tag {3} +$$ + +This is the inherent ambiguity in SfM, which we are going to resolve in this paper. + +Focus distance ambiguity in $DP$ images As described by Wadhawa et al. [36] and Garg et al. [10], the signed defocus blur size $b \in \mathbb{R}$ is related to scene depth $z$ , focal length $f$ , lens aperture diameter $l$ , and the focus distance $g$ in $\mathbb{R}$ as + +$$ +b \approx \frac {l f}{1 - f / g} \left(\frac {1}{g} - \frac {1}{z}\right) \tag {4} +$$ + +using the paraxial and thin-lens approximations. The blur size $b$ in DP images varies with the focus distance $g$ even when the same scene depth $z$ is recorded with the same camera parameters $(l, f)$ . Therefore, even with the blur size $b$ from DP images, its relationship to depth $z$ is ambiguous due to unknown focus distance $g$ and depth $z$ as + +$$ +b \approx \frac {l f}{1 - f / g} \left(\frac {1}{g} - \frac {1}{z}\right) = \frac {l f}{1 - f / g ^ {\prime}} \left(\frac {1}{g ^ {\prime}} - \frac {1}{z ^ {\prime}}\right), \tag {5} +$$ + +i.e., there exists a set of pairs like $(g,z)$ and $(g^{\prime},z^{\prime})$ that yields the same blur size $b$ . The blur size $b$ is signed, i.e., it becomes positive when the scene point is farther away from the focal plane and negative when nearer to the camera. + +# 4.2 Solution Method + +Given the blur sizes $b_{ij} \in \mathbb{R}$ obtained from DP images, depths $z_{ij}' \in \mathbb{R}_+$ from SfM and MVS in the $i$ -th view at $j$ -th pixel, and known per-view focal lengths $f_i$ and lens aperture $l_i$ , our goal is to determine unknown scene scale $s \in \mathbb{R}_+$ and per-view focus distances $g_i \in \mathbb{R}_+$ . Since we use per-view intrinsics, our method can be used in a practical setting of multi-view 3D reconstruction, in which camera intrinsics may vary across viewpoints. + +We put together the two ambiguous observations, i.e., depths up to scale and blur sizes with unknown focus distances, by substituting Eq. (3) to Eq. (4) as + +$$ +b _ {i j} = \frac {l _ {i} f _ {i}}{1 - f _ {i} / g _ {i}} \left(\frac {1}{g _ {i}} - \frac {1}{s z _ {i j} ^ {\prime}}\right). \tag {6} +$$ + +Denoting $\bar{g}_i = \frac{1}{g_i}$ and $\bar{s} = \frac{1}{s}$ as $g_{i}$ and $s$ are both non-zero, Eq. (6) can be rewritten as + +$$ +\frac {b _ {i j} z _ {i j} ^ {\prime}}{f _ {i}} = \underbrace {z _ {i j} ^ {\prime} \left(b _ {i j} + l _ {i}\right)} _ {\gamma_ {i j}} \bar {g} _ {i} - l _ {i} \bar {s} = \gamma_ {i j} \bar {g} _ {i} - l _ {i} \bar {s}, \tag {7} +$$ + +![](images/8fd12ed0558ecf1a2bc915ceeeda4b7ed4b7be17545e4c31e006c0e1c849577d.jpg) +(a) Input image + +![](images/863591e22344b41e66a82534277c12d2c0e1805c7427ee700eee6651dae482b9.jpg) +(b) Confidence score +Fig. 4: Pixel selection within a view based on confidence scores. We select reliable pixels that contain good blur size estimates from an input image (a). (b) shows the log-scaled confidence scores $c$ that show the reliability of blur size estimates [25]. (c) Local variations (variance) of depths. (d) Selected pixels based on our pixel selection method (highlighted in the figure). + +![](images/0770950cc90c5248ef44910478605d360d1c90b6a5e38a980e4d2fc482970531.jpg) +(c) Local depth variation + +![](images/e8bde8534839581c05fa1678d070978a534e29386b33c4346f132c2d9d0fffeb.jpg) +(d) Selected pixels + +which is linear with respect to the unknown parameters $\bar{g}_i$ and $\bar{s}$ . + +From observations of multiple views and pixels, we can construct an overdetermined system of linear equations from Eq. (7). Given the number of views $n$ and the number of valid pixels $p_i$ in the $i$ -th view, we have $\sum_{i=1}^{n} p_i$ equations, which can be written in a matrix form as: + +$$ +\underbrace {\left[ \begin{array}{c c c c c} \gamma_ {1 1} & & & & - l _ {1} \\ \vdots & & & & \vdots \\ \gamma_ {1 p _ {1}} & & 0 & & - l _ {1} \\ & \gamma_ {2 1} & & & - l _ {2} \\ & \vdots & & & \vdots \\ 0 & \vdots & 0 & & \vdots \\ & \gamma_ {2 p _ {2}} & & & - l _ {2} \\ & & \ddots & & \vdots \\ & & & \gamma_ {n 1} & - l _ {n} \\ & 0 & & \vdots & \vdots \\ & & & \gamma_ {n p _ {n}} & - l _ {n} \end{array} \right]} _ {\mathbf {A}} \underbrace {\left[ \begin{array}{c} \bar {g} _ {1} \\ \bar {g} _ {2} \\ \vdots \\ \bar {g} _ {n} \\ \bar {s} \end{array} \right]} _ {\mathbf {x}} = \underbrace {\left[ \begin{array}{c} b _ {1 1} z _ {1 1} ^ {\prime} / f _ {1} \\ \vdots \\ b _ {1 p _ {1}} z _ {1 p _ {1}} ^ {\prime} / f _ {1} \\ b _ {2 1} z _ {2 1} ^ {\prime} / f _ {2} \\ \vdots \\ b _ {2 p _ {2}} z _ {2 p _ {2}} ^ {\prime} / f _ {2} \\ \vdots \\ b _ {n 1} z _ {n 1} ^ {\prime} / f _ {n} \\ \vdots \\ b _ {n p _ {n}} z _ {n p _ {n}} ^ {\prime} / f _ {n} \end{array} \right]} _ {\mathbf {b}}. \tag {8} +$$ + +Since we have $n + 1$ unknowns in vector $\mathbf{x}$ , given $\mathbf{b} \neq \mathbf{0}$ , the condition for obtaining a unique (approximate) solution is $\mathrm{rank}(\mathbf{A}) = n + 1$ , which is, in practice, easy to achieve because a large number of pixels are available. The minimum setting to determine scale $s$ is selecting two pixels with distinct depths in one of the multi-view images, which makes a full-rank $2 \times 2$ matrix $\mathbf{A}$ for two unknowns $\mathbf{x} = [\bar{g}_i, \bar{s}]^\top$ . + +In practice, we can select many pixels from multiple views to rely on robust estimation. With a sufficiently large number of selected pixels, we construct a tall and skinny matrix $\mathbf{A}$ and use $\ell_1$ residual minimization to effectively disregard outliers as + +$$ +\mathbf {x} ^ {*} = \underset {\mathbf {x}} {\operatorname {a r g m i n}} \| \mathbf {A x} - \mathbf {b} \| _ {1}. \tag {9} +$$ + +We use the Iteratively Reweighted Least Squares (IRLS) method [12] for deriving the approximate solution $\mathbf{x}^*$ . Obviously, other choices of robust estimation methods, such as RANdom SAmple Consensus (RANSAC) [8], are viable for alleviating the impact of outliers. + +# 4.3 Confidence-based Pixel Selection + +So far, we have assumed that the blur sizes $b_{ij}$ are available to our method. For obtaining the blur sizes, we use the method proposed by Punnappurath et al. [25] that optimizes Eq. (1). For our case, we only require blur size estimates for a small number of pixels but not for all the pixels since our goal is determining the scene scale $s$ using Eq. (8). We, therefore, develop a pixel selection heuristic to choose a small number of pixels that have reliable blur size estimates. Specifically, we select the reliable pixels in each view, followed by selecting the views that produce reliable estimates of scene scale. + +Pixel-selection within a view We begin by adopting the confidence scores of [25] that indicate the reliability of the blur size estimates based on the residual of the blur size optimization and the textured-ness (magnitude of image gradient) of the pixel of interest. We slightly modified their original confidence score using a depth-based criterion since we observed that the large depth variation in an image patch negatively affects the blur size estimation. + +Let us consider one of the views and let $c_{j}$ be the confidence score defined in [25] at the $j$ -th pixel. In our case, a depth value $z_{j}^{\prime}$ up to scale is available at the pixel, and we can assess the variance of the depth values within an image patch $\mathcal{P}_j$ centered at the $j$ -th pixel as $\text{var}(\{z_k'\})$ , $k \in \mathcal{P}_j$ . Based on the variance of the depths, we re-define the confidence score as + +$$ +c _ {j} ^ {*} \triangleq \frac {c _ {j}}{\operatorname {v a r} \left(\left\{z _ {k} ^ {\prime} \right\}\right)}, \quad k \in \mathcal {P} _ {j}. \tag {10} +$$ + +The greater depth variation in an image patch leads to a larger variance and lower confidence score, and vice versa. Based on the new confidence score $c_{j}^{*}$ , we select the top $T_{c}\%$ of pixels and regard them as reliable pixels that have good defocus blur estimates. Although the variance of depths varies with the depth scale, which is unknown in our case, it does not affect the confidence score $c_{j}^{*}$ within a view because the same scale is applied to all the pixels in the view. Figure 4 depicts the pixel selection within a view. + +View selection In the multi-view images for SfM, there may be images that are not suitable for scale estimation. For example, a view with small depth variations is not preferred because the small depth variations make the linear system of Eq. (8) unstable. Therefore, we further select reliable views for scale estimation. + +Since our method can estimate scale $s$ for each view as long as the depths up to scale are available, we first compute the scale predictions for each view independently using the selected pixels within each view by solving Eq. (8). Ideally, the scales estimated for each view should match; however, in reality, they do not due to outliers. We first disregard the views with clear outliers, i.e., views with excessively small blur size variations and views that yield negative values of scale estimates. The small blur size variations indicate that the scene contains small depth variations. We simply set a threshold to discard the views with small blur size variations. Namely, when the difference between maximum + +and minimum blur size estimates becomes less than or equal to $T_{p}$ pixels, the view is unused for the estimation. The negative scale estimates could arise due to various reasons, such as image degradation by motion blur and inaccurate estimates of either or both blur sizes and depths. We simply discard such views from the estimation. + +Finally, from the remaining views, we select $N_{v}$ views + +![](images/b44b869b5c5b430d1397251a070bbe309c341c0c997e40a92e1e48b056720c79.jpg) +Fig. 5: View selection. We select the reliable views for scale estimation based on the median of the scale values estimated for each view. + +![](images/8322b3ad199f9bc14638ca95278ee26c0f4379af9dbc9279ca712cda713f85b8.jpg) + +![](images/f82b47f8ca6de496701585b22cc0500bbc01bf80664eed57d1597a7d03b04b64.jpg) + +that are close to the median of the estimated scales as illustrated in Fig. 5. Once the appropriate views are selected, by using all the selected pixels within the views, we create the linear system of Eq. (8) and re-estimate the scale $s$ . + +# 4.4 Implementation Details + +For SfM and MVS, we use the commercial software Metashape1. For DP blur size estimation, we use the method proposed by Punnappurath et al. [25] as described, except for the smoothing operation implemented as post-processing. All captured images are undistorted using the parameters estimated by MVS before being fed into the blur size estimation. Our method is not restricted by the choice of SfM+MVS and blur size estimation methods, but we found these to be the best choices today. + +For the focal length $f$ , we use the one from the lens specification. The aperture diameter $l$ is calculated using f-number $N$ of the lens as $l = \frac{f}{N}$ . For the hyperparameters, we use $T_{c} = 10$ , $T_{p} = 2$ , and $N_{v} = 7$ throughout the experiments. + +# 5 Experiment + +We assess the accuracy of our scale estimation method both quantitatively and qualitatively. We use multi-view DP images captured in six scenes with different capturing settings, such as, cameras (DSLR and smartphone), lenses, and aperture sizes. + +Recording setting We use two different types of cameras equipped with a DP sensor: a DSLR camera and a smartphone. For the DSLR camera setting, we use Canon EOS 5D Mark IV with three different lenses: EF35mm F1.4L II USM, EF50mm F1.8 STM, and EF85mm F1.8 USM. For each scene, we also change the aperture size to assess the effect of the aperture size on our method. The + +maximum aperture ( $f/1.4$ or $f/1.8$ depending on the lens), $f/4.0$ , and $f/8.0$ are used for the experiment depending on the target scene. The images from Canon's DLSR camera are recorded as Canon Raw images with the CR2 format, and the left and right DP views are extracted using DPRSplit2. For a more accessible setup, we also use a smartphone, Google Pixel 4, which has a focal length of $4.38 \mathrm{~mm}$ and a fixed aperture of $f/1.73$ . To obtain the left and right DP views, we use the capturing software used in [11]. + +For each scene for each camera setting, we capture about 30 images using a DSLR, and 60 images using a smartphone. For fair comparisons across different lens and aperture sizes, for each viewpoint, we use a tripod to keep the camera fixed and record images with different camera settings. To evaluate the scale estimates, we obtain the ground-truth scene scale by placing markers with known sizes in the scenes. + +Evaluation metric As a quantitative evaluation metric, we use the scale ratio $r_s$ of the estimated scene scale $s_{\mathrm{est}}$ to the ground-truth scale $s_{\mathrm{gt}}$ , which is defined as follows: + +$$ +\mathrm {S c a l e r a t i o} r _ {s} = \frac {s _ {\mathrm {e s t}}}{s _ {\mathrm {g t}}}. +$$ + +In the above metric, closer to 1 indicates good scale estimation. The scale ratio smaller than 1 indicates that the scale is under-estimated, i.e., the estimated scale is smaller than the true scale, and the greater scale ratio corresponds to the over-estimation. While scale ratio $r_s$ is defined for each scene and camera setting, we also define the average error of the scale ratios $e_s$ to characterize the overall effectiveness as + +$$ +\mathrm {A v e r a g e e r r o r o f s c a l e r a t i o s} e _ {s} = \frac {1}{n _ {s}} \sum \left(\max (r _ {s}, r _ {s} ^ {- 1}) - 1\right), +$$ + +where $n_s$ is the number of settings, and $\max(r_s, r_s^{-1})$ is for taking into account the under- and over-estimations. + +Quantitative results The results of the scale estimation for six scenes using two DP cameras are summarized in Table 1. The DSLR camera uses three lenses and different aperture sizes. Overall, our method can reasonably well estimate the absolute scene scale just by using a DP camera for multi-view capture, considering the fact that scale ratio $r_s$ can range in $(0, \infty)$ . The average error $e_s$ across scenes and lenses is approximately 0.219 as shown in the table. From comparing the average errors for each lens, we see that the difference in lenses does not significantly affect the overall performance. + +Visual results Figures 6 and 7 show visual examples of our results by the DSLR and smartphone, as well as the estimated blur sizes and the ground-truth depth maps. We can confirm the same trend as in the quantitative results, where the + +![](images/163b087c947ddd277a14c36e9adb178c5f9437e92ad9e82bb17f67c03f2344b5.jpg) +Fig. 6: Visual results of real scenes captured by the DSLR (Canon EOS 5D Mark IV). Our method yields reasonable estimates of scene scales just by using a DP camera for multi-view capture, which is theoretically unrecoverable by standard SfM alone. The top three rows show the results of the same scene (and from almost the same viewpoint) but with different aperture sizes or focal lengths. The bottom rows show the results of different scenes. In (a), the distance indicates the true distance obtained from the GT 3D model. In (d), the distance in the estimated scale is shown. (b) shows estimated blur size and selected pixels based on our pixel selection method (highlighted in the figure). + +Table 1: Scale ratio $r_s$ between the ground truth and the estimated scale for each scene, lens, and aperture size. Maximum aperture varies depending on the lens (i.e., $f/1.4$ or $f/1.8$ ). $e_s$ represents the average errors of the corresponding row/column/total. “-” indicates that we did not record the scene under the setting. + +
DSLRR (35mm)DSLRR (50mm)DSLRR (85mm)Smartphone f/1.73es
f/1.4f/4.0f/8.0f/1.8f/4.0f/8.0f/1.8f/4.0f/8.0
Scene 11.4191.088-1.0841.084-1.1001.231-1.0850.156
Scene 21.2251.238-1.0471.232-1.2471.302-1.2520.220
Scene 31.1991.229-1.0551.352-1.2381.226-1.5190.260
Scene 41.1811.1071.4771.0271.2661.4691.2251.1901.2961.0340.227
Scene 51.1711.0721.3141.0671.0991.2851.0381.2051.4251.0680.174
Scene 61.1481.073-1.0831.029-1.1231.126-1.3880.139
es0.2240.1350.3960.0610.1770.3770.1620.2130.3610.2240.219
+ +![](images/e1f22bce1241aa74bdb7db06f1eaadc5750fc95f6e267f6ddbc0098de011ee44.jpg) +Fig. 7: Visual results of real scenes captured by a smartphone (Google Pixel 4). Similar to the results by DSLR, our method estimates reasonable scene scales despite the DP blur being relatively smaller than DSLR's. + +depth maps with estimated scale well represent the actual depths. Our method estimates the scene scale with less than 0.2 error ratio for several settings, where the blur size estimation is relatively stable. For these cases, 3D reconstruction with our absolute scale estimation allows high-quality measurements with an error of several centimeters in wide outdoor scenes. + +Stability of our method To assess the stability of our method against variation of input images, we capture additional $200\sim 300$ images for Scenes 1 and 2 using the DSLR equipped with a $35\mathrm{mm}$ lens and $f / 1.4$ aperture size and repeatedly estimate the scale using different sets of images. Figure 8 shows an overlay of estimated models with repeated experiments and the minimum and maximum estimated scale ratios, $r_{s\min}$ and $r_{s\max}$ , with 10 trials. + +![](images/61d7b3b0ae9eb49d6a3502f402859a6da3a1fd5fc0aa6d7716db9ddac7e3e04d.jpg) +$r_{s\min}$ : 1.158, $r_{s\max}$ : 1.255 + +![](images/3db705a2597fe43b78da37ac93467a4725f089d1bb2b850924998d19cb22aa7e.jpg) +$r_{s\min}$ : 1.174, $r_{s\max}$ : 1.285 +$r_{s\min}$ : 1.202, $r_{s\max}$ : 1.383 +Fig. 8: Overlay of estimated models with repeated experiments and the minimum and maximum estimated scale ratios with 10 trials. Blue and red point clouds represent the ones with scale ratios $r_{s\max}$ and $r_{s\min}$ , respectively. + +![](images/d50a86c1bda7f2ca4c53def7ce22e40dfdb37ba22a7037a36e5f6498c21913c9.jpg) + +Comparison with scale estimation by monocular depth estimation We compare our method with a single-image metric depth estimation, ZoeDepth [6], for the same scenes and lens setup used. Since it estimates metric depth for each pixel, we compute the per-image scale as the average ratio between ZoeDepth and MVS depth maps. We then average the per-image scale over seven inlier viewpoints selected using the same procedure as our view selection. + +The average error $e_{s}$ across scenes and lenses using ZoeDepth is 0.679, which is more than 3 times larger than ours (0.219). Figure 9 confirms this + +trend through the visualization of the estimated scale. More detailed comparisons are shown in the supplementary material. + +![](images/8f6034f5d9ae24edea8557f7eaeaa475057db4fba9c6f38e95cbab533182586b.jpg) +$r_{sours}$ : 1.038, $r_{sZoe\text{depth}}$ : 0.643 +Fig. 9: Comparison with monocular scale estimation. Blue and red points represent the ones of the scale ratio estimated by our method ( $r_{s\text{ours}}$ ) and ZoeDepth ( $r_{sZoeDepth}$ ), respectively. Black points represent the GT model. + +Limitations and failure cases Although our method successfully estimates the scene scale in many cases, we find several failure modes in our experiments. Since our method is based on defocus blur, scale estimation from pan-focus images becomes challenging. In particular, since the aperture of smartphone cameras is smaller than that of DSLR's, it was necessary for us to include photographs that exhibit large depth variations to produce relatively large blur like the top row of Fig. 7. Although the pan-focus images are preferable in SfM and MVS for better correspondence matching, in practice, small apertures cannot always be used. This is because the small aperture limits the amount of incoming light, requiring longer exposure times, and resulting in camera motion blurs. + +The success of our method relies on the accurate estimation of both defocus blur sizes and depth maps up to scale. We found several examples where the scale estimation is affected by the blur size estimation errors, while the depth maps up to scale are almost always stably estimated. Figure 10 shows a failure + +![](images/3c7a6f6e962f45147b489b907b96e2ae5219d2706a3753f9b87cbf9c19d6630c.jpg) +(a) Input image + +![](images/e018bb67874fe57f81ab84c83b4bcf5fb8abf4acd4ca5c9fba73a2fa61036d0e.jpg) +(b) GT depth +Fig. 10: An example of failure case (Scene 3 captured with a $50\mathrm{mm}$ lens and $f / 4.0$ aperture size) producing scale ratio $r_s = 1.352$ due to the large error in blur size estimation. Ideal blur sizes (d) are calculated from the ground-truth (GT) scene depth (b) and Eq. (6). Since the target scene is mostly textureless or a repetition of fine textures, most of the pixels in the estimated blur sizes (c) deviate from ideal values, making the selection of the inlier pixels challenging. + +![](images/cee5efa42d29122ffa21fcda926f8fb19df6df3a9300c64009ffa49c106a0490.jpg) +(c) Estimated blur size + +![](images/176d138d1f5f26a1feb898aa0a02a4da5db369a7a064d2c1087c34b6ffcc9370.jpg) +(d) Ideal blur size + +mode of blur size estimation as well as the estimated scene depth, whose scale ratio $r_s$ reaches as large as 1.352 (Scene 3 captured with $50\mathrm{mm}$ lens and $f / 4.0$ aperture size). To assess the quality of blur size estimates shown in Fig. 10 (c), we synthesize the ideal blur sizes in Fig. 10 (d) using the ground-truth depth of Fig. 10 (b) that is computed from the markers and the image formation model of Eq. (6). Most of the blur size estimates are far from the ideal blur sizes due to the target object being almost textureless or the repetition of fine-grained texture, which makes blur size optimization challenging. We further analyze the accuracy and characteristics of the existing blur size estimation methods in the supplementary material. + +# 6 Discussions + +This paper has shown that the scale information in multi-view 3d reconstruction, which is theoretically unsolvable by SfM and MVS itself, can be inferred from the defocus blur information provided by DP imaging. Our method estimates the absolute scale without relying on reference objects or data priors but by simply capturing multi-view images with a DP camera. The key to the scale recovery is the unique feature of DP that relates defocus disparity, blur size, and scene depth, from which we can affinely relate the reciprocal of the depth and the defocus blur sizes given by DP images. Together with the pixel selection method for improving robustness using the confidence information from the off-the-shelf DP blur estimation method as well as the statistics of the depth information, we simultaneously recover the scene scale and the unknown focus distance via a simple linear regression. + +Experiments show that our method successfully estimates the absolute scale in many cases. Since the estimation accuracy of the blur size strongly affects the performance of our method, the analysis of DP blur characteristics will be beneficial in the future. With a method for a more accurate and stabler estimation of blur sizes in DP images, our method will produce more accurate absolute scales for multi-view 3D reconstruction. + +# Acknowledgments + +This work was supported by JSPS KAKENHI Grant Number JP23H05491. + +# References + +1. Abuolaim, A., Afifi, M., Brown, M.S.: Improving single-image defocus deblurring: How dual-pixel images help through multi-task learning. In: IEEE/CVF Winter Conference on Applications of Computer Vision (WACV). pp. 82-90 (2022) +2. Abuolaim, A., Brown, M.S.: Defocus deblurring using dual-pixel data. In: Proceedings of European Conference on Computer Vision (ECCV). pp. 111-126 (2020) +3. Abuolaim, A., Delbracio, M., Kelly, D., Brown, M.S., Milanfar, P.: Learning to reduce defocus blur by realistically modeling dual-pixel data. In: Proceedings of IEEE/CVF International Conference on Computer Vision (ICCV). pp. 2289-2298 (2021) +4. Abuolaim, A., Punnappurath, A., Brown, M.S.: Revisiting autofocus for smartphone cameras. In: Proceedings of European Conference on Computer Vision (ECCV). pp. 545-559 (2018) +5. Abuolaim, A., Timofte, R., Brown, M.S.: NTIRE 2021 challenge for defocus deblurring using dual-pixel images: Methods and results. In: Proceedings of IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW). pp. 578-587 (2021) +6. Bhat, S.F., Birkl, R., Wofk, D., Wonka, P., Müller, M.: Zoedepth: Zero-shot transfer by combining relative and metric depth. arXiv preprint arXiv:2302.12288 (2023) +7. Engel, J., Stuckler, J., Cremers, D.: Large-scale direct SLAM with stereo cameras. In: IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). pp. 1935-1942 (2015) +8. Fischler, M.A., Bolles, R.C.: Random sample consensus: A paradigm for model fitting with applications to image analysis and automated cartography. Communications of the ACM 24(6), 381-395 (1981) +9. Frost, D., Prisacariu, V., Murray, D.: Recovering stable scale in monocular SLAM using object-supplemented bundle adjustment. IEEE Transactions on Robotics 34(3), 736-747 (2018) +0. Garg, R., Wadhwa, N., Ansari, S., Barron, J.T.: Learning single camera depth estimation using dual-pixels. In: Proceedings of IEEE/CVF International Conference on Computer Vision (ICCV). pp. 7628-7637 (2019) +1. Garg, R., Wadhwa, N., Ansari, S., Barron, J.T.: Learning single camera depth estimation using dual-pixels. In: Proceedings of IEEE/CVF International Conference on Computer Vision (ICCV). pp. 7628-7637 (2019) +2. Gentle, J.: Matrix Albegra. Springer Texts in Statistics, Springer, New York (2007) +3. Giubilato, R., Chiodini, S., Pertile, M., Debei, S.: Scale correct monocular visual odometry using a lidar altimeter. In: IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). pp. 3694-3700 (2018) +4. Herrmann, C., Bowen, R.S., Wadhwa, N., Garg, R., He, Q., Barron, J.T., Zabih, R.: Learning to autofocus. In: Proceedings of IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 2227-2236 (2020) +5. Jung, S.H., Heo, Y.S.: Disparity probability volume guided defocus deblurring using dual pixel data. In: Proceedings of International Conference on Information and Communication Technology Convergence (ICTC). pp. 305-308 (2021) +6. Kang, M., Choe, J., Ha, H., Jeon, H.G., Im, S., Kweon, I.S.: Facial depth and normal estimation using single dual-pixel camera. In: Proceedings of European Conference on Computer Vision (ECCV) (2022) +7. Kashiwagi, M., Mishima, N., Kozakaya, T., Hiura, S.: Deep depth from aberration map. In: Proceedings of IEEE/CVF International Conference on Computer Vision (ICCV) (2019) + +18. Kim, D., Jang, H., Kim, I., Kim, M.H.: Spatio-focal bidirectional disparity estimation from a dual-pixel image. In: Proceedings of IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 5023-5032 (2023) +19. Knorr, S.B., Kurz, D.: Leveraging the user's face for absolute scale estimation in handheld monocular SLAM. In: Proceedings of IEEE International Symposium on Mixed and Augmented Reality (ISMAR). pp. 11-17 (2016) +20. Lee, S.H., de Croon, G.: Stability-based scale estimation for monocular SLAM. IEEE Robotics and Automation Letters 3(2), 780-787 (2018) +21. Li, F., Guo, H., Santo, H., Okura, F., Matsushita, Y.: Learning to synthesize photorealistic dual-pixel images from RGBD frames. In: International Conference on Computational Photography (ICCP). pp. 1-11 (2023) +22. Mishima, N., Seki, A., Hiura, S.: Absolute scale from varifocal monocular camera through SfM and defocus combined. In: Proceedings of British Machine Vision Conference (BMVC) (2021) +23. Nützi, G., Weiss, S., Scaramuzza, D., Siegwart, R.: Fusion of IMU and vision for absolute scale estimation in monocular SLAM. Journal of Intelligent & Robotic Systems 61(1-4), 287-299 (2011) +24. Pan, L., Chowdhury, S., Hartley, R., Liu, M., Zhang, H., Li, H.: Dual pixel exploration: Simultaneous depth estimation and image restoration. In: Proceedings of IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 4340-4349 (2021) +25. Punnappurath, A., Abuolaim, A., Affi, M., Brown, M.S.: Modeling defocus-disparity in dual-pixel sensors. In: International Conference on Computational Photography (ICCP). pp. 1-12 (2020) +26. Punnappurath, A., Brown, M.S.: Reflection removal using a dual-pixel sensor. In: Proceedings of IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 1556-1565 (2019) +27. Roussel, T., Van Eycken, L., Tuytelaars, T.: Monocular depth estimation in new environments with absolute scale. In: IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). pp. 1735-1741 (2019) +28. Rukhovich, D., Mouritzen, D., Kaestner, R., Rufli, M., Velizhev, A.: Estimation of absolute scale in monocular SLAM using synthetic data. In: Proceedings of IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW). pp. 803-812 (2019) +29. Scaramuzza, D., Fraundorfer, F., Pollefeys, M., Siegwart, R.: Absolute scale in structure from motion from a single vehicle mounted camera by exploiting nonholonomic constraints. In: Proceedings of IEEE/CVF International Conference on Computer Vision (ICCV). pp. 1413-1419 (2009) +30. Shibata, A., Fujii, H., Yamashita, A., Asama, H.: Absolute scale structure from motion using a refractive plate. In: Proceedings of IEEE/SICE International Symposium on System Integration (SII). pp. 540-545 (2015) +31. Shibata, A., Fujii, H., Yamashita, A., Asama, H.: Scale-reconstructable structure from motion using refraction with a single camera. In: IEEE International Conference on Robotics and Automation (ICRA). pp. 5239-5244 (2015) +32. Shiozaki, T., Dissanayake, G.: Eliminating scale drift in monocular SLAM using depth from defocus. IEEE Robotics and Automation Letters 3(1), 581-587 (2017) +33. Song, S., Chandraker, M.: Robust scale estimation in real-time monocular sfm for autonomous driving. In: Proceedings of IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 1566-1573 (2014) + +34. Sucar, E., Hayet, J.B.: Bayesian scale estimation for monocular SLAM based on generic object detection for correcting scale drift. In: IEEE International Conference on Robotics and Automation (ICRA). pp. 5152-5158 (2018) +35. Sumikura, S., Sakurada, K., Kawaguchi, N., Nakamura, R.: Scale estimation of monocular SfM for a multi-modal stereo camera. In: Proceedings of Asian Conference on Computer Vision (ACCV). pp. 281-297 (2019) +36. Wadhwa, N., Garg, R., Jacobs, D.E., Feldman, B.E., Kanazawa, N., Carroll, R., Movshovitz-Attias, Y., Barron, J.T., Pritch, Y., Levoy, M.: Synthetic depth-of-field with a single-camera mobile phone. ACM Transactions on Graphics (TOG) pp. 1-13 (2018) +37. Wöhler, C., d'Angelo, P., Krüger, L., Kuhl, A., Groß, H.M.: Monocular 3D scene reconstruction at absolute scale. ISPRS Journal of Photogrammetry and Remote Sensing 64(6), 529-540 (2009) +38. Xin, S., Wadhwa, N., Xue, T., Barron, J.T., Srinivasan, P.P., Chen, J., Gkioulekas, I., Garg, R.: Defocus map estimation and deblurring from a single dual-pixel image. In: Proceedings of IEEE/CVF International Conference on Computer Vision (ICCV). pp. 2228-2238 (2021) +39. Yang, Y., Pan, L., Liu, L., Liu, M.: K3DN: Disparity-aware kernel estimation for dual-pixel defocus deblurring. In: Proceedings of IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 13263-13272 (2023) +40. Zhang, S., Zhang, J., Tao, D.: Towards scale-aware, robust, and generalizable unsupervised monocular depth estimation by integrating IMU motion dynamics. In: Proceedings of European Conference on Computer Vision (ECCV). pp. 143-160 (2022) +41. Zhang, Y., Wadhwa, N., Orts-Escolano, S., Häne, C., Fanello, S.R., Garg, R.: Du2Net: Learning depth estimation from dual-cameras and dual-pixels. In: Proceedings of European Conference on Computer Vision (ECCV). pp. 582-598 (2020) +42. Zhou, D., Dai, Y., Li, H.: Reliable scale estimation and correction for monocular visual odometry. In: Proceedings of IEEE Intelligent Vehicles Symposium (IV). pp. 490-495 (2016) +43. Zhou, D., Dai, Y., Li, H.: Ground-plane-based absolute scale estimation for monocular visual odometry. IEEE Transactions on Intelligent Transportation Systems 21(2), 791-802 (2019) \ No newline at end of file diff --git a/2024/Resolving Scale Ambiguity in Multi-view 3D Reconstruction using Dual-Pixel Sensors/images.zip b/2024/Resolving Scale Ambiguity in Multi-view 3D Reconstruction using Dual-Pixel Sensors/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..12cd931829bd1d9441edfb3986d5e08e375d2593 --- /dev/null +++ b/2024/Resolving Scale Ambiguity in Multi-view 3D Reconstruction using Dual-Pixel Sensors/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:dfad8c65c2ad66bcb6878203f4742584a01735d318718983b338ed5c334df71f +size 647699 diff --git a/2024/Resolving Scale Ambiguity in Multi-view 3D Reconstruction using Dual-Pixel Sensors/layout.json b/2024/Resolving Scale Ambiguity in Multi-view 3D Reconstruction using Dual-Pixel Sensors/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..84c66843c08174f6d11e6c1e8e7266c1c4f0f2e5 --- /dev/null +++ b/2024/Resolving Scale Ambiguity in Multi-view 3D Reconstruction using Dual-Pixel Sensors/layout.json @@ -0,0 +1,10697 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 151, + 111, + 462, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 111, + 462, + 148 + ], + "spans": [ + { + "bbox": [ + 151, + 111, + 462, + 148 + ], + "type": "text", + "content": "Resolving Scale Ambiguity in Multi-view 3D Reconstruction using Dual-Pixel Sensors" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 138, + 168, + 474, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 168, + 474, + 180 + ], + "spans": [ + { + "bbox": [ + 138, + 168, + 474, + 180 + ], + "type": "text", + "content": "Kohei Ashida, Hiroaki Santo, Fumio Okura, and Yasuyuki Matsushita" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 190, + 190, + 423, + 201 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 190, + 190, + 423, + 201 + ], + "spans": [ + { + "bbox": [ + 190, + 190, + 423, + 201 + ], + "type": "text", + "content": "Graduate School of Information Science and Technology," + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 257, + 202, + 357, + 212 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 257, + 202, + 357, + 212 + ], + "spans": [ + { + "bbox": [ + 257, + 202, + 357, + 212 + ], + "type": "text", + "content": "Osaka University, Japan" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 157, + 213, + 456, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 157, + 213, + 456, + 224 + ], + "spans": [ + { + "bbox": [ + 157, + 213, + 456, + 224 + ], + "type": "text", + "content": "{kohei.ashida, santo.hiroaki, okura, yasumat}@ist.osaka-u.ac.jp" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 159, + 247, + 455, + 423 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 247, + 455, + 423 + ], + "spans": [ + { + "bbox": [ + 159, + 247, + 455, + 423 + ], + "type": "text", + "content": "Abstract. Multi-view 3D reconstruction, namely structure-from-motion and multi-view stereo, is an essential component in 3D computer vision. In general, multi-view 3D reconstruction suffers from unknown scale ambiguity unless a reference object of known size is recorded together with the scene, or the camera poses are pre-calibrated. In this paper, we show that multi-view images recorded by a dual-pixel (DP) sensor allow us to automatically resolve the scale ambiguity without requiring a reference object or pre-calibration. Specifically, the observed defocus blurs in DP images provide sufficient information for determining the scale when paired together with the depth maps (up to scale) recovered from the multi-view 3D reconstruction. Based on this observation, we develop a simple yet effective linear solution method to determine the absolute scale in multi-view 3D reconstruction. Experiments demonstrate the effectiveness of the proposed method with diverse scenes recorded with different cameras/lenses. Code and data are available at https://github.com/kohei-ashida/dp-sfm." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 159, + 432, + 454, + 455 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 432, + 454, + 455 + ], + "spans": [ + { + "bbox": [ + 159, + 432, + 454, + 455 + ], + "type": "text", + "content": "Keywords: Multi-view 3D reconstruction " + }, + { + "bbox": [ + 159, + 432, + 454, + 455 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 432, + 454, + 455 + ], + "type": "text", + "content": " Photogrammetry " + }, + { + "bbox": [ + 159, + 432, + 454, + 455 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 432, + 454, + 455 + ], + "type": "text", + "content": " Dual-pixel imaging " + }, + { + "bbox": [ + 159, + 432, + 454, + 455 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 432, + 454, + 455 + ], + "type": "text", + "content": " Scale ambiguity" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 472, + 230, + 485 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 472, + 230, + 485 + ], + "spans": [ + { + "bbox": [ + 132, + 472, + 230, + 485 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 495, + 482, + 638 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 495, + 482, + 638 + ], + "spans": [ + { + "bbox": [ + 130, + 495, + 482, + 638 + ], + "type": "text", + "content": "Multi-view 3D reconstruction is one of the most important problems in 3D computer vision. From multi-view images of a scene taken from unknown viewpoints, structure-from-motion (SfM) estimates camera poses and sparse 3D tie points of the scene. Subsequently, multi-view stereo (MVS) uses the estimated camera poses to reconstruct the dense shape of the scene. It is known that the scale ambiguity in SfM is inherent even with known camera intrinsics and cannot be resolved without external information. In this setting, the subsequent MVS can only recover dense shapes up to scale. A classical yet most reliable approach to resolving the ambiguity is to place a reference object of known size, such as a calibration target, in the scene to determine the absolute scale. However, it is not always possible to put such a reference object in the scene, and it is desired to relax the demand for diverse application scenarios." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 639, + 482, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 639, + 482, + 664 + ], + "spans": [ + { + "bbox": [ + 130, + 639, + 482, + 664 + ], + "type": "text", + "content": "In this paper, we propose a method for automatically determining the absolute scale in multi-view 3D reconstruction using dual-pixel (DP) cameras whose" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 248, + 366 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 248, + 366 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 248, + 366 + ], + "type": "text", + "content": "intrinsic are calibrated. DP sensors are designed to assist the camera's auto-focus system and are becoming popular since they are equipped with Canon EOS cameras and Google Pixel smartphones. Each pixel on the DP sensor has two photodiodes to acquire two sub-aperture views. The two sub-aperture views allow us to estimate defocus blurs that are observed at regions off of the focal plane, although they depend on the unknown focus distance. We show that, by relating the DP defocus blur sizes with" + } + ] + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 255, + 119, + 482, + 281 + ], + "blocks": [ + { + "bbox": [ + 255, + 119, + 482, + 281 + ], + "lines": [ + { + "bbox": [ + 255, + 119, + 482, + 281 + ], + "spans": [ + { + "bbox": [ + 255, + 119, + 482, + 281 + ], + "type": "image", + "image_path": "43caf3e12a49ee35813a2d3581b790a58726370add16cf6431f57b409ff85d2a.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 252, + 289, + 482, + 345 + ], + "lines": [ + { + "bbox": [ + 252, + 289, + 482, + 345 + ], + "spans": [ + { + "bbox": [ + 252, + 289, + 482, + 345 + ], + "type": "text", + "content": "Fig. 1: Our method determines the absolute scene scale of SfM and MVS from multi-view DP images. We put together the scale ambiguity in SfM and MVS and the focus distance ambiguity of DP images. Our method jointly resolves the ambiguities." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 367, + 482, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 367, + 482, + 403 + ], + "spans": [ + { + "bbox": [ + 130, + 367, + 482, + 403 + ], + "type": "text", + "content": "unknown focus distances with the dense depth maps up to scale obtained from SfM and MVS, the absolute scale of the scene can be fully determined without requiring other external information." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 408, + 482, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 408, + 482, + 540 + ], + "spans": [ + { + "bbox": [ + 130, + 408, + 482, + 540 + ], + "type": "text", + "content": "The key idea of our method is coupling two ambiguous measurements for resolving the ambiguities: One is the linear scale ambiguity of depth values computed from SfM and MVS, and the other is the focus distance ambiguity that affinely relates the reciprocal of depths with DP defocus blurs. As illustrated in Fig. 1, the size of defocus blur " + }, + { + "bbox": [ + 130, + 408, + 482, + 540 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 130, + 408, + 482, + 540 + ], + "type": "text", + "content": " is related to unknown focus distance " + }, + { + "bbox": [ + 130, + 408, + 482, + 540 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 130, + 408, + 482, + 540 + ], + "type": "text", + "content": " and unknown scene scale " + }, + { + "bbox": [ + 130, + 408, + 482, + 540 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 130, + 408, + 482, + 540 + ], + "type": "text", + "content": ", with which scene point " + }, + { + "bbox": [ + 130, + 408, + 482, + 540 + ], + "type": "inline_equation", + "content": "\\mathbf{p}" + }, + { + "bbox": [ + 130, + 408, + 482, + 540 + ], + "type": "text", + "content": " may be scaled to " + }, + { + "bbox": [ + 130, + 408, + 482, + 540 + ], + "type": "inline_equation", + "content": "\\mathbf{sp}" + }, + { + "bbox": [ + 130, + 408, + 482, + 540 + ], + "type": "text", + "content": " retaining its projections " + }, + { + "bbox": [ + 130, + 408, + 482, + 540 + ], + "type": "inline_equation", + "content": "\\mathbf{u}_1" + }, + { + "bbox": [ + 130, + 408, + 482, + 540 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 408, + 482, + 540 + ], + "type": "inline_equation", + "content": "\\mathbf{u}_2" + }, + { + "bbox": [ + 130, + 408, + 482, + 540 + ], + "type": "text", + "content": ". By putting them together, our method solves a linear regression problem for determining both the absolute scale of the scene and the focus distances in the multi-view images. Our method capitalizes on the fact that DP's two sub-aperture views faithfully yield defocus blur estimates, significantly more reliably than just a single view from an ordinary camera." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 146, + 544, + 384, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 544, + 384, + 557 + ], + "spans": [ + { + "bbox": [ + 146, + 544, + 384, + 557 + ], + "type": "text", + "content": "In summary, the chief contributions of this paper are:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 137, + 583, + 479, + 665 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 138, + 583, + 479, + 608 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 583, + 479, + 608 + ], + "spans": [ + { + "bbox": [ + 138, + 583, + 479, + 608 + ], + "type": "text", + "content": "- We propose a method to determine the absolute scale in uncalibrated multi-view 3D reconstruction using DP images." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 137, + 613, + 479, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 613, + 479, + 635 + ], + "spans": [ + { + "bbox": [ + 137, + 613, + 479, + 635 + ], + "type": "text", + "content": "- We develop a linear solution method for simultaneously determining the global scene scale and per-view focus distances." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 137, + 641, + 479, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 641, + 479, + 665 + ], + "spans": [ + { + "bbox": [ + 137, + 641, + 479, + 665 + ], + "type": "text", + "content": "- We conduct comprehensive experiments to assess the effectiveness of the proposed method using diverse scenes and camera settings." + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 233, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 233, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 233, + 100 + ], + "type": "text", + "content": "K. Ashida et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 237, + 127 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 237, + 127 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 237, + 127 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 144, + 482, + 265 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 144, + 482, + 265 + ], + "spans": [ + { + "bbox": [ + 130, + 144, + 482, + 265 + ], + "type": "text", + "content": "Scale estimation in SfM Since SfM alone cannot recover the absolute scale from the image information, it is necessary to introduce an additional cue that provides absolute scale information. A simple and reliable approach is to place a reference object of known size, such as markers and ground control points (GCPs), which is widely employed in commercial photogrammetry software. Domain-specific cues are often used to avoid manual placement of reference objects, for example, urban objects [9,34], ground surfaces [33,42,43], and human faces [19]. To implicitly exploit the scale information of scene objects, recent learning-based methods use monocular depth estimation providing absolute scale information [27,28]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 265, + 482, + 338 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 265, + 482, + 338 + ], + "spans": [ + { + "bbox": [ + 130, + 265, + 482, + 338 + ], + "type": "text", + "content": "Since it is not always possible to assume reference objects in target scenes, external sensors are often used to recover the scale information. The use of stereo cameras with known baselines [7, 35] or using LiDAR [13] are typical examples, while inertial sensors [23, 40] and odometry information [20, 29] are also widely used in robotics applications. Placing a refraction plate in front of the camera provides scale information using multiple capture [30, 31]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 338, + 482, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 338, + 482, + 469 + ], + "spans": [ + { + "bbox": [ + 130, + 338, + 482, + 469 + ], + "type": "text", + "content": "Similar to our work, defocus has been used to determine the absolute scale in the literature. Calibrated depth-from-defocus is used with SfM-based depths to recover the absolute scale by assuming a known and fixed focus distance [32,37]. Similar to our method, a recent study [22] relaxes the known focus assumption by jointly estimating the physical focal length and scene scale. Although their use of the thin-lens model is similar to ours, their method requires pre-training or calibration for a learning-based blur size estimation [17] that depends on the hardware setup and focal lengths. Our study instead uses uncalibrated defocus blurs, where the focus distance is treated as unknown. The uncalibrated defocus blurs are reliably obtained from DP images, and thereby, our method recovers the absolute scales without prior knowledge, training data, or additional equipment." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 486, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 486, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 486, + 482, + 665 + ], + "type": "text", + "content": "DP imaging and applications Unlike conventional sensors, a single pixel in a DP sensor is divided into two photodiodes, one on the left and one on the right, which produces disparity. Beyond its primary use of fast auto-focus [4,14] implemented in several commercial cameras, applications of DP imaging are actively studied. DP's disparity gives effective cues to estimate the Point Spread Functions (PSFs) of defocus blurs [25], whose direct application is defocus deblurring [1-3,5,15,24,38,39] and blur synthesis [36]. The defocus blur information from DPs improves the performance of monocular depth estimation [10,18,25]. For this application, a DP image dataset with the ground truth depths has been developed [21]. The DP-based depth has been used for domain-specific applications, such as reflection removal [26] and facial shape recovery [16]. A recent method combines DP imaging and traditional two-view stereo for depth estimation [41]. Taking advantage of DP's effectiveness in acquiring defocus blur information, our work attempts to determine the scene scale automatically in multi-view 3D reconstruction." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 159, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 159, + 91, + 447, + 102 + ], + "type": "text", + "content": "Resolving Scale Ambiguity in MV3D Reconstruction using DP Sensors" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 412, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 412, + 129 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 412, + 129 + ], + "type": "text", + "content": "3 Preliminary: DP Imaging and Defocus Blur" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 140, + 480, + 187 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 140, + 480, + 187 + ], + "spans": [ + { + "bbox": [ + 130, + 140, + 480, + 187 + ], + "type": "text", + "content": "Unlike conventional sensors, a DP sensor can capture two sub-aperture images by the left and right photodiodes in a single shot. There is a difference between the two images, which is caused by the difference in the PSFs of the left and right DP views [25]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 188, + 298, + 390 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 188, + 298, + 390 + ], + "spans": [ + { + "bbox": [ + 130, + 188, + 298, + 390 + ], + "type": "text", + "content": "Figure 2 shows the visualization of the difference between two DP views and illustrated corresponding PSFs. There is no difference between the two views if the depth of a scene point is in focus. When the scene point is out of focus, i.e., located further away or closer to the camera from the focal plane, the disparities between the two views appear in opposite directions. Unlike traditional two-view stereo, the disparity is caused by defocus blurs generated by the PSFs for each view. It is known that the signed defocus blur size " + }, + { + "bbox": [ + 130, + 188, + 298, + 390 + ], + "type": "inline_equation", + "content": "b \\in \\mathbb{R}" + }, + { + "bbox": [ + 130, + 188, + 298, + 390 + ], + "type": "text", + "content": " is proportional to the defocus disparity " + }, + { + "bbox": [ + 130, + 188, + 298, + 390 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 130, + 188, + 298, + 390 + ], + "type": "text", + "content": " between two views [10]." + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 305, + 188, + 481, + 305 + ], + "blocks": [ + { + "bbox": [ + 305, + 188, + 481, + 305 + ], + "lines": [ + { + "bbox": [ + 305, + 188, + 481, + 305 + ], + "spans": [ + { + "bbox": [ + 305, + 188, + 481, + 305 + ], + "type": "image", + "image_path": "a14d6a6f2608809c03dbd9527df6f41e995e1080d2268197d370dad5d637eaaa.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 312, + 482, + 380 + ], + "lines": [ + { + "bbox": [ + 304, + 312, + 482, + 380 + ], + "spans": [ + { + "bbox": [ + 304, + 312, + 482, + 380 + ], + "type": "text", + "content": "Fig.2: Visualization of DP disparity and illustrations of the blur, i.e., PSF, kernels. The two sub-aperture views convert the defocus blur to disparity, resulting in defocus disparity in opposite directions for near and far scene points. The DP image is from [2]." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 392, + 480, + 452 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 392, + 480, + 452 + ], + "spans": [ + { + "bbox": [ + 130, + 392, + 480, + 452 + ], + "type": "text", + "content": "DP imaging makes the estimation of defocus blur size " + }, + { + "bbox": [ + 130, + 392, + 480, + 452 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 130, + 392, + 480, + 452 + ], + "type": "text", + "content": " tractable [25] without requiring external information or learning-based methods. Since DP PSFs " + }, + { + "bbox": [ + 130, + 392, + 480, + 452 + ], + "type": "inline_equation", + "content": "(\\mathbf{H}_l, \\mathbf{H}_r)" + }, + { + "bbox": [ + 130, + 392, + 480, + 452 + ], + "type": "text", + "content": " can be well approximated by a parametric function of blur size " + }, + { + "bbox": [ + 130, + 392, + 480, + 452 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 130, + 392, + 480, + 452 + ], + "type": "text", + "content": ", when DP image patches " + }, + { + "bbox": [ + 130, + 392, + 480, + 452 + ], + "type": "inline_equation", + "content": "(\\mathbf{G}_l, \\mathbf{G}_r)" + }, + { + "bbox": [ + 130, + 392, + 480, + 452 + ], + "type": "text", + "content": " are given, the blur size can be estimated per pixel by a nonlinear minimization in an unsupervised manner as" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 216, + 460, + 480, + 480 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 216, + 460, + 480, + 480 + ], + "spans": [ + { + "bbox": [ + 216, + 460, + 480, + 480 + ], + "type": "interline_equation", + "content": "b = \\underset {b} {\\operatorname {a r g m i n}} \\| \\mathbf {G} _ {l} * \\mathbf {H} _ {r} (b) - \\mathbf {G} _ {r} * \\mathbf {H} _ {l} (b) \\| _ {F}, \\tag {1}", + "image_path": "3292ad185e3b4d5a5adaae23e770f806cec52f71b2b5a5d01507522cc7048e1c.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 489, + 472, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 489, + 472, + 502 + ], + "spans": [ + { + "bbox": [ + 131, + 489, + 472, + 502 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 489, + 472, + 502 + ], + "type": "inline_equation", + "content": "*" + }, + { + "bbox": [ + 131, + 489, + 472, + 502 + ], + "type": "text", + "content": " is a convolution operator. The readers are referred to [25] for details." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 520, + 261, + 533 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 520, + 261, + 533 + ], + "spans": [ + { + "bbox": [ + 132, + 520, + 261, + 533 + ], + "type": "text", + "content": "4 Proposed Method" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": "Our method resolves the scale ambiguity in multi-view 3D reconstruction using multi-view DP images. The overview is illustrated in Fig. 3. Our method infers the unknown scene scale just by capturing multi-view images with a DP camera. DP images provide a useful cue for blur size estimation, as described above, without requiring any external information. Our method shares the same spirit with [22]; however, there is a key difference in that our method does not require pre-training but capitalizes on the DP sensor's effectiveness in determining blur sizes. We show in Fig. 3 that the estimated blur size and (inverse) depth from SfM and MVS have an affine relationship, and thus, the unknown scene scale and focus distance can be estimated simultaneously by linear regression." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 233, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 233, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 233, + 100 + ], + "type": "text", + "content": "K. Ashida et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 114, + 202, + 227 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 202, + 227 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 202, + 227 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 202, + 227 + ], + "type": "image", + "image_path": "e6146d7ce555b3b411ea66fc42ee56abf4c9494ae602382032ff223ce4725c6d.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 239, + 480, + 293 + ], + "lines": [ + { + "bbox": [ + 130, + 239, + 480, + 293 + ], + "spans": [ + { + "bbox": [ + 130, + 239, + 480, + 293 + ], + "type": "text", + "content": "Fig. 3: Overview of our method that provides the absolute scale to multi-view 3D reconstruction (SfM and MVS) just by capturing the scene using DP cameras. (a) The disparity between DP views provides the cue to estimate the defocus blur size. (b) From the affine relation between the (inverse) depth and the defocus blur size, our method simultaneously estimates the scene scale and focus distances." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 203, + 115, + 269, + 227 + ], + "blocks": [ + { + "bbox": [ + 203, + 115, + 269, + 227 + ], + "lines": [ + { + "bbox": [ + 203, + 115, + 269, + 227 + ], + "spans": [ + { + "bbox": [ + 203, + 115, + 269, + 227 + ], + "type": "image", + "image_path": "4217e975b1151f6f64652270725e964664f15783b355e2b2633da928a07035e2.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 269, + 116, + 347, + 227 + ], + "blocks": [ + { + "bbox": [ + 269, + 116, + 347, + 227 + ], + "lines": [ + { + "bbox": [ + 269, + 116, + 347, + 227 + ], + "spans": [ + { + "bbox": [ + 269, + 116, + 347, + 227 + ], + "type": "image", + "image_path": "eab5002f47f48a54226bd5ff41606f75d61a4c4b5a1cd67e986bef107bdf52bb.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 347, + 116, + 410, + 227 + ], + "blocks": [ + { + "bbox": [ + 347, + 116, + 410, + 227 + ], + "lines": [ + { + "bbox": [ + 347, + 116, + 410, + 227 + ], + "spans": [ + { + "bbox": [ + 347, + 116, + 410, + 227 + ], + "type": "image", + "image_path": "508d66c9ced6e327a9d98d413f8941f78ee320443366a58c851759ffacb80dc3.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 410, + 116, + 480, + 227 + ], + "blocks": [ + { + "bbox": [ + 410, + 116, + 480, + 227 + ], + "lines": [ + { + "bbox": [ + 410, + 116, + 480, + 227 + ], + "spans": [ + { + "bbox": [ + 410, + 116, + 480, + 227 + ], + "type": "image", + "image_path": "42084d35034be71ceacd7af684b36332e248068264def1810a9b52b1d0b914fb.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 319, + 479, + 368 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 319, + 479, + 368 + ], + "spans": [ + { + "bbox": [ + 130, + 319, + 479, + 368 + ], + "type": "text", + "content": "We begin by describing the ambiguities that arise in our setting, then describe the solution method for resolving them. We also present a method for selecting reliable pixels and views for scale estimation, which is based on the inference by an off-the-shelf blur size estimation method [25]." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 382, + 480, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 382, + 480, + 453 + ], + "spans": [ + { + "bbox": [ + 130, + 382, + 480, + 453 + ], + "type": "text", + "content": "Assumptions We assume that camera intrinsics, such as focal lengths and aperture sizes (f-numbers), are known as they are usually accessible from the camera system. Although defocus blurs slightly change the center of projection, we assume the change is sufficiently small and ignore their effect on the intrinsics. As stated, we use DP images taken from multiple viewpoints as input. Our method does not assume pre-calibrated viewpoints nor reference objects in a scene." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 475, + 352, + 487 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 475, + 352, + 487 + ], + "spans": [ + { + "bbox": [ + 132, + 475, + 352, + 487 + ], + "type": "text", + "content": "4.1 Scale and Focus Distance Ambiguities" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 498, + 479, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 498, + 479, + 521 + ], + "spans": [ + { + "bbox": [ + 130, + 498, + 479, + 521 + ], + "type": "text", + "content": "There are two ambiguities in our setting: One is the scale ambiguity in SfM and MVS, and the other is the focus distance ambiguity in DP images." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 536, + 482, + 632 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 536, + 482, + 632 + ], + "spans": [ + { + "bbox": [ + 130, + 536, + 482, + 632 + ], + "type": "text", + "content": "Scale ambiguity in SfM and MVS Camera positions and 3D point locations recovered by SfM and the subsequent MVS have a scale ambiguity even with known camera intrinsics. Let us consider a 3D point " + }, + { + "bbox": [ + 130, + 536, + 482, + 632 + ], + "type": "inline_equation", + "content": "\\mathbf{p} = [x,y,z]^{\\top}" + }, + { + "bbox": [ + 130, + 536, + 482, + 632 + ], + "type": "text", + "content": " projected to the corresponding 2D point " + }, + { + "bbox": [ + 130, + 536, + 482, + 632 + ], + "type": "inline_equation", + "content": "\\mathbf{u} = [u,v,1]^{\\top}" + }, + { + "bbox": [ + 130, + 536, + 482, + 632 + ], + "type": "text", + "content": " in the homogeneous coordinates with a projection matrix " + }, + { + "bbox": [ + 130, + 536, + 482, + 632 + ], + "type": "inline_equation", + "content": "[\\mathbf{R}|\\mathbf{t}]" + }, + { + "bbox": [ + 130, + 536, + 482, + 632 + ], + "type": "text", + "content": " formed by a rotation matrix " + }, + { + "bbox": [ + 130, + 536, + 482, + 632 + ], + "type": "inline_equation", + "content": "\\mathbf{R} \\in SO(3)" + }, + { + "bbox": [ + 130, + 536, + 482, + 632 + ], + "type": "text", + "content": " and a translation vector " + }, + { + "bbox": [ + 130, + 536, + 482, + 632 + ], + "type": "inline_equation", + "content": "\\mathbf{t} \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 130, + 536, + 482, + 632 + ], + "type": "text", + "content": ". With known camera intrinsics, we can set the intrinsics to be identity and write the 3D-to-2D perspective projection in some Euclidean world coordinate system as" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 233, + 641, + 481, + 669 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 233, + 641, + 481, + 669 + ], + "spans": [ + { + "bbox": [ + 233, + 641, + 481, + 669 + ], + "type": "interline_equation", + "content": "\\mathbf {u} \\sim [ \\mathbf {R} \\mathbf {t} ] \\left[ \\begin{array}{l} \\mathbf {p} \\\\ 1 \\end{array} \\right] = \\frac {1}{s} [ \\mathbf {R} s \\mathbf {t} ] \\left[ \\begin{array}{l} s \\mathbf {p} \\\\ 1 \\end{array} \\right], \\tag {2}", + "image_path": "f30cdd559b5f90631baeaf9861410073616e8021516c32d9382e2f2959d473e3.jpg" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 159, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 159, + 91, + 448, + 102 + ], + "type": "text", + "content": "Resolving Scale Ambiguity in MV3D Reconstruction using DP Sensors" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 152 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 152 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 116, + 482, + 152 + ], + "type": "inline_equation", + "content": "s \\in \\mathbb{R}" + }, + { + "bbox": [ + 130, + 116, + 482, + 152 + ], + "type": "text", + "content": " is an unknown scaling factor. As such, the depth " + }, + { + "bbox": [ + 130, + 116, + 482, + 152 + ], + "type": "inline_equation", + "content": "z'" + }, + { + "bbox": [ + 130, + 116, + 482, + 152 + ], + "type": "text", + "content": " obtained by SfM and MVS is linearly related to the true depth " + }, + { + "bbox": [ + 130, + 116, + 482, + 152 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 130, + 116, + 482, + 152 + ], + "type": "text", + "content": " by the unknown scaling factor " + }, + { + "bbox": [ + 130, + 116, + 482, + 152 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 130, + 116, + 482, + 152 + ], + "type": "text", + "content": " as" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 286, + 159, + 481, + 171 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 286, + 159, + 481, + 171 + ], + "spans": [ + { + "bbox": [ + 286, + 159, + 481, + 171 + ], + "type": "interline_equation", + "content": "z = s z ^ {\\prime}. \\tag {3}", + "image_path": "0e511e34cd2812559fb5e11f5454c857356f97d6a7ca67408ffdf124c2a4d937.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 177, + 481, + 202 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 177, + 481, + 202 + ], + "spans": [ + { + "bbox": [ + 130, + 177, + 481, + 202 + ], + "type": "text", + "content": "This is the inherent ambiguity in SfM, which we are going to resolve in this paper." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 211, + 481, + 248 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 211, + 481, + 248 + ], + "spans": [ + { + "bbox": [ + 131, + 211, + 481, + 248 + ], + "type": "text", + "content": "Focus distance ambiguity in " + }, + { + "bbox": [ + 131, + 211, + 481, + 248 + ], + "type": "inline_equation", + "content": "DP" + }, + { + "bbox": [ + 131, + 211, + 481, + 248 + ], + "type": "text", + "content": " images As described by Wadhawa et al. [36] and Garg et al. [10], the signed defocus blur size " + }, + { + "bbox": [ + 131, + 211, + 481, + 248 + ], + "type": "inline_equation", + "content": "b \\in \\mathbb{R}" + }, + { + "bbox": [ + 131, + 211, + 481, + 248 + ], + "type": "text", + "content": " is related to scene depth " + }, + { + "bbox": [ + 131, + 211, + 481, + 248 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 131, + 211, + 481, + 248 + ], + "type": "text", + "content": ", focal length " + }, + { + "bbox": [ + 131, + 211, + 481, + 248 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 131, + 211, + 481, + 248 + ], + "type": "text", + "content": ", lens aperture diameter " + }, + { + "bbox": [ + 131, + 211, + 481, + 248 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 131, + 211, + 481, + 248 + ], + "type": "text", + "content": ", and the focus distance " + }, + { + "bbox": [ + 131, + 211, + 481, + 248 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 131, + 211, + 481, + 248 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 131, + 211, + 481, + 248 + ], + "type": "inline_equation", + "content": "\\mathbb{R}" + }, + { + "bbox": [ + 131, + 211, + 481, + 248 + ], + "type": "text", + "content": " as" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 255, + 252, + 481, + 279 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 255, + 252, + 481, + 279 + ], + "spans": [ + { + "bbox": [ + 255, + 252, + 481, + 279 + ], + "type": "interline_equation", + "content": "b \\approx \\frac {l f}{1 - f / g} \\left(\\frac {1}{g} - \\frac {1}{z}\\right) \\tag {4}", + "image_path": "71f510c15cd1612134805bc2f853b68f552ef3fa1bf3e549f2ae1bafe073deb8.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 283, + 482, + 344 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 283, + 482, + 344 + ], + "spans": [ + { + "bbox": [ + 130, + 283, + 482, + 344 + ], + "type": "text", + "content": "using the paraxial and thin-lens approximations. The blur size " + }, + { + "bbox": [ + 130, + 283, + 482, + 344 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 130, + 283, + 482, + 344 + ], + "type": "text", + "content": " in DP images varies with the focus distance " + }, + { + "bbox": [ + 130, + 283, + 482, + 344 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 130, + 283, + 482, + 344 + ], + "type": "text", + "content": " even when the same scene depth " + }, + { + "bbox": [ + 130, + 283, + 482, + 344 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 130, + 283, + 482, + 344 + ], + "type": "text", + "content": " is recorded with the same camera parameters " + }, + { + "bbox": [ + 130, + 283, + 482, + 344 + ], + "type": "inline_equation", + "content": "(l, f)" + }, + { + "bbox": [ + 130, + 283, + 482, + 344 + ], + "type": "text", + "content": ". Therefore, even with the blur size " + }, + { + "bbox": [ + 130, + 283, + 482, + 344 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 130, + 283, + 482, + 344 + ], + "type": "text", + "content": " from DP images, its relationship to depth " + }, + { + "bbox": [ + 130, + 283, + 482, + 344 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 130, + 283, + 482, + 344 + ], + "type": "text", + "content": " is ambiguous due to unknown focus distance " + }, + { + "bbox": [ + 130, + 283, + 482, + 344 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 130, + 283, + 482, + 344 + ], + "type": "text", + "content": " and depth " + }, + { + "bbox": [ + 130, + 283, + 482, + 344 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 130, + 283, + 482, + 344 + ], + "type": "text", + "content": " as" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 202, + 349, + 481, + 376 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 202, + 349, + 481, + 376 + ], + "spans": [ + { + "bbox": [ + 202, + 349, + 481, + 376 + ], + "type": "interline_equation", + "content": "b \\approx \\frac {l f}{1 - f / g} \\left(\\frac {1}{g} - \\frac {1}{z}\\right) = \\frac {l f}{1 - f / g ^ {\\prime}} \\left(\\frac {1}{g ^ {\\prime}} - \\frac {1}{z ^ {\\prime}}\\right), \\tag {5}", + "image_path": "8f8863e1c3499ce1c2f2e13f3048ed0a1c18e28fb38e61f71b1f74e5850c5298.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 380, + 482, + 417 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 380, + 482, + 417 + ], + "spans": [ + { + "bbox": [ + 131, + 380, + 482, + 417 + ], + "type": "text", + "content": "i.e., there exists a set of pairs like " + }, + { + "bbox": [ + 131, + 380, + 482, + 417 + ], + "type": "inline_equation", + "content": "(g,z)" + }, + { + "bbox": [ + 131, + 380, + 482, + 417 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 380, + 482, + 417 + ], + "type": "inline_equation", + "content": "(g^{\\prime},z^{\\prime})" + }, + { + "bbox": [ + 131, + 380, + 482, + 417 + ], + "type": "text", + "content": " that yields the same blur size " + }, + { + "bbox": [ + 131, + 380, + 482, + 417 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 131, + 380, + 482, + 417 + ], + "type": "text", + "content": ". The blur size " + }, + { + "bbox": [ + 131, + 380, + 482, + 417 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 131, + 380, + 482, + 417 + ], + "type": "text", + "content": " is signed, i.e., it becomes positive when the scene point is farther away from the focal plane and negative when nearer to the camera." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 431, + 246, + 442 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 431, + 246, + 442 + ], + "spans": [ + { + "bbox": [ + 131, + 431, + 246, + 442 + ], + "type": "text", + "content": "4.2 Solution Method" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 449, + 482, + 520 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 449, + 482, + 520 + ], + "spans": [ + { + "bbox": [ + 130, + 449, + 482, + 520 + ], + "type": "text", + "content": "Given the blur sizes " + }, + { + "bbox": [ + 130, + 449, + 482, + 520 + ], + "type": "inline_equation", + "content": "b_{ij} \\in \\mathbb{R}" + }, + { + "bbox": [ + 130, + 449, + 482, + 520 + ], + "type": "text", + "content": " obtained from DP images, depths " + }, + { + "bbox": [ + 130, + 449, + 482, + 520 + ], + "type": "inline_equation", + "content": "z_{ij}' \\in \\mathbb{R}_+" + }, + { + "bbox": [ + 130, + 449, + 482, + 520 + ], + "type": "text", + "content": " from SfM and MVS in the " + }, + { + "bbox": [ + 130, + 449, + 482, + 520 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 449, + 482, + 520 + ], + "type": "text", + "content": "-th view at " + }, + { + "bbox": [ + 130, + 449, + 482, + 520 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 130, + 449, + 482, + 520 + ], + "type": "text", + "content": "-th pixel, and known per-view focal lengths " + }, + { + "bbox": [ + 130, + 449, + 482, + 520 + ], + "type": "inline_equation", + "content": "f_i" + }, + { + "bbox": [ + 130, + 449, + 482, + 520 + ], + "type": "text", + "content": " and lens aperture " + }, + { + "bbox": [ + 130, + 449, + 482, + 520 + ], + "type": "inline_equation", + "content": "l_i" + }, + { + "bbox": [ + 130, + 449, + 482, + 520 + ], + "type": "text", + "content": ", our goal is to determine unknown scene scale " + }, + { + "bbox": [ + 130, + 449, + 482, + 520 + ], + "type": "inline_equation", + "content": "s \\in \\mathbb{R}_+" + }, + { + "bbox": [ + 130, + 449, + 482, + 520 + ], + "type": "text", + "content": " and per-view focus distances " + }, + { + "bbox": [ + 130, + 449, + 482, + 520 + ], + "type": "inline_equation", + "content": "g_i \\in \\mathbb{R}_+" + }, + { + "bbox": [ + 130, + 449, + 482, + 520 + ], + "type": "text", + "content": ". Since we use per-view intrinsics, our method can be used in a practical setting of multi-view 3D reconstruction, in which camera intrinsics may vary across viewpoints." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 521, + 482, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 482, + 545 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 482, + 545 + ], + "type": "text", + "content": "We put together the two ambiguous observations, i.e., depths up to scale and blur sizes with unknown focus distances, by substituting Eq. (3) to Eq. (4) as" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 241, + 560, + 481, + 594 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 241, + 560, + 481, + 594 + ], + "spans": [ + { + "bbox": [ + 241, + 560, + 481, + 594 + ], + "type": "interline_equation", + "content": "b _ {i j} = \\frac {l _ {i} f _ {i}}{1 - f _ {i} / g _ {i}} \\left(\\frac {1}{g _ {i}} - \\frac {1}{s z _ {i j} ^ {\\prime}}\\right). \\tag {6}", + "image_path": "5a8beb6f95ca2ff9041ac7418cbecf19b741d857c2de6575ebb7f303e4814e45.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 597, + 482, + 622 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 597, + 482, + 622 + ], + "spans": [ + { + "bbox": [ + 130, + 597, + 482, + 622 + ], + "type": "text", + "content": "Denoting " + }, + { + "bbox": [ + 130, + 597, + 482, + 622 + ], + "type": "inline_equation", + "content": "\\bar{g}_i = \\frac{1}{g_i}" + }, + { + "bbox": [ + 130, + 597, + 482, + 622 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 597, + 482, + 622 + ], + "type": "inline_equation", + "content": "\\bar{s} = \\frac{1}{s}" + }, + { + "bbox": [ + 130, + 597, + 482, + 622 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 130, + 597, + 482, + 622 + ], + "type": "inline_equation", + "content": "g_{i}" + }, + { + "bbox": [ + 130, + 597, + 482, + 622 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 597, + 482, + 622 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 130, + 597, + 482, + 622 + ], + "type": "text", + "content": " are both non-zero, Eq. (6) can be rewritten as" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 216, + 625, + 481, + 663 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 216, + 625, + 481, + 663 + ], + "spans": [ + { + "bbox": [ + 216, + 625, + 481, + 663 + ], + "type": "interline_equation", + "content": "\\frac {b _ {i j} z _ {i j} ^ {\\prime}}{f _ {i}} = \\underbrace {z _ {i j} ^ {\\prime} \\left(b _ {i j} + l _ {i}\\right)} _ {\\gamma_ {i j}} \\bar {g} _ {i} - l _ {i} \\bar {s} = \\gamma_ {i j} \\bar {g} _ {i} - l _ {i} \\bar {s}, \\tag {7}", + "image_path": "bdf27965743bea85109844522739d79703be35f84cb43102d6cc09ad58f877ce.jpg" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 233, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 233, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 233, + 101 + ], + "type": "text", + "content": "K. Ashida et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 140, + 114, + 218, + 166 + ], + "blocks": [ + { + "bbox": [ + 140, + 114, + 218, + 166 + ], + "lines": [ + { + "bbox": [ + 140, + 114, + 218, + 166 + ], + "spans": [ + { + "bbox": [ + 140, + 114, + 218, + 166 + ], + "type": "image", + "image_path": "8fd12ed0558ecf1a2bc915ceeeda4b7ed4b7be17545e4c31e006c0e1c849577d.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 149, + 167, + 206, + 176 + ], + "lines": [ + { + "bbox": [ + 149, + 167, + 206, + 176 + ], + "spans": [ + { + "bbox": [ + 149, + 167, + 206, + 176 + ], + "type": "text", + "content": "(a) Input image" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 221, + 114, + 301, + 166 + ], + "blocks": [ + { + "bbox": [ + 221, + 114, + 301, + 166 + ], + "lines": [ + { + "bbox": [ + 221, + 114, + 301, + 166 + ], + "spans": [ + { + "bbox": [ + 221, + 114, + 301, + 166 + ], + "type": "image", + "image_path": "863591e22344b41e66a82534277c12d2c0e1805c7427ee700eee6651dae482b9.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 225, + 167, + 299, + 175 + ], + "lines": [ + { + "bbox": [ + 225, + 167, + 299, + 175 + ], + "spans": [ + { + "bbox": [ + 225, + 167, + 299, + 175 + ], + "type": "text", + "content": "(b) Confidence score" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 131, + 185, + 481, + 240 + ], + "lines": [ + { + "bbox": [ + 131, + 185, + 481, + 240 + ], + "spans": [ + { + "bbox": [ + 131, + 185, + 481, + 240 + ], + "type": "text", + "content": "Fig. 4: Pixel selection within a view based on confidence scores. We select reliable pixels that contain good blur size estimates from an input image (a). (b) shows the log-scaled confidence scores " + }, + { + "bbox": [ + 131, + 185, + 481, + 240 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 131, + 185, + 481, + 240 + ], + "type": "text", + "content": " that show the reliability of blur size estimates [25]. (c) Local variations (variance) of depths. (d) Selected pixels based on our pixel selection method (highlighted in the figure)." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 309, + 113, + 391, + 166 + ], + "blocks": [ + { + "bbox": [ + 309, + 113, + 391, + 166 + ], + "lines": [ + { + "bbox": [ + 309, + 113, + 391, + 166 + ], + "spans": [ + { + "bbox": [ + 309, + 113, + 391, + 166 + ], + "type": "image", + "image_path": "0770950cc90c5248ef44910478605d360d1c90b6a5e38a980e4d2fc482970531.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 307, + 167, + 395, + 175 + ], + "lines": [ + { + "bbox": [ + 307, + 167, + 395, + 175 + ], + "spans": [ + { + "bbox": [ + 307, + 167, + 395, + 175 + ], + "type": "text", + "content": "(c) Local depth variation" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 399, + 114, + 477, + 166 + ], + "blocks": [ + { + "bbox": [ + 399, + 114, + 477, + 166 + ], + "lines": [ + { + "bbox": [ + 399, + 114, + 477, + 166 + ], + "spans": [ + { + "bbox": [ + 399, + 114, + 477, + 166 + ], + "type": "image", + "image_path": "e8bde8534839581c05fa1678d070978a534e29386b33c4346f132c2d9d0fffeb.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 403, + 167, + 470, + 175 + ], + "lines": [ + { + "bbox": [ + 403, + 167, + 470, + 175 + ], + "spans": [ + { + "bbox": [ + 403, + 167, + 470, + 175 + ], + "type": "text", + "content": "(d) Selected pixels" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 256, + 420, + 268 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 256, + 420, + 268 + ], + "spans": [ + { + "bbox": [ + 131, + 256, + 420, + 268 + ], + "type": "text", + "content": "which is linear with respect to the unknown parameters " + }, + { + "bbox": [ + 131, + 256, + 420, + 268 + ], + "type": "inline_equation", + "content": "\\bar{g}_i" + }, + { + "bbox": [ + 131, + 256, + 420, + 268 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 256, + 420, + 268 + ], + "type": "inline_equation", + "content": "\\bar{s}" + }, + { + "bbox": [ + 131, + 256, + 420, + 268 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 269, + 480, + 316 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 269, + 480, + 316 + ], + "spans": [ + { + "bbox": [ + 130, + 269, + 480, + 316 + ], + "type": "text", + "content": "From observations of multiple views and pixels, we can construct an overdetermined system of linear equations from Eq. (7). Given the number of views " + }, + { + "bbox": [ + 130, + 269, + 480, + 316 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 130, + 269, + 480, + 316 + ], + "type": "text", + "content": " and the number of valid pixels " + }, + { + "bbox": [ + 130, + 269, + 480, + 316 + ], + "type": "inline_equation", + "content": "p_i" + }, + { + "bbox": [ + 130, + 269, + 480, + 316 + ], + "type": "text", + "content": " in the " + }, + { + "bbox": [ + 130, + 269, + 480, + 316 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 269, + 480, + 316 + ], + "type": "text", + "content": "-th view, we have " + }, + { + "bbox": [ + 130, + 269, + 480, + 316 + ], + "type": "inline_equation", + "content": "\\sum_{i=1}^{n} p_i" + }, + { + "bbox": [ + 130, + 269, + 480, + 316 + ], + "type": "text", + "content": " equations, which can be written in a matrix form as:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 217, + 323, + 481, + 452 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 217, + 323, + 481, + 452 + ], + "spans": [ + { + "bbox": [ + 217, + 323, + 481, + 452 + ], + "type": "interline_equation", + "content": "\\underbrace {\\left[ \\begin{array}{c c c c c} \\gamma_ {1 1} & & & & - l _ {1} \\\\ \\vdots & & & & \\vdots \\\\ \\gamma_ {1 p _ {1}} & & 0 & & - l _ {1} \\\\ & \\gamma_ {2 1} & & & - l _ {2} \\\\ & \\vdots & & & \\vdots \\\\ 0 & \\vdots & 0 & & \\vdots \\\\ & \\gamma_ {2 p _ {2}} & & & - l _ {2} \\\\ & & \\ddots & & \\vdots \\\\ & & & \\gamma_ {n 1} & - l _ {n} \\\\ & 0 & & \\vdots & \\vdots \\\\ & & & \\gamma_ {n p _ {n}} & - l _ {n} \\end{array} \\right]} _ {\\mathbf {A}} \\underbrace {\\left[ \\begin{array}{c} \\bar {g} _ {1} \\\\ \\bar {g} _ {2} \\\\ \\vdots \\\\ \\bar {g} _ {n} \\\\ \\bar {s} \\end{array} \\right]} _ {\\mathbf {x}} = \\underbrace {\\left[ \\begin{array}{c} b _ {1 1} z _ {1 1} ^ {\\prime} / f _ {1} \\\\ \\vdots \\\\ b _ {1 p _ {1}} z _ {1 p _ {1}} ^ {\\prime} / f _ {1} \\\\ b _ {2 1} z _ {2 1} ^ {\\prime} / f _ {2} \\\\ \\vdots \\\\ b _ {2 p _ {2}} z _ {2 p _ {2}} ^ {\\prime} / f _ {2} \\\\ \\vdots \\\\ b _ {n 1} z _ {n 1} ^ {\\prime} / f _ {n} \\\\ \\vdots \\\\ b _ {n p _ {n}} z _ {n p _ {n}} ^ {\\prime} / f _ {n} \\end{array} \\right]} _ {\\mathbf {b}}. \\tag {8}", + "image_path": "cb6b83acbe0f26fb73d4770974efe086ef62e688c8beb2b1bbe4d09df9b63f38.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 459, + 481, + 530 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 459, + 481, + 530 + ], + "spans": [ + { + "bbox": [ + 130, + 459, + 481, + 530 + ], + "type": "text", + "content": "Since we have " + }, + { + "bbox": [ + 130, + 459, + 481, + 530 + ], + "type": "inline_equation", + "content": "n + 1" + }, + { + "bbox": [ + 130, + 459, + 481, + 530 + ], + "type": "text", + "content": " unknowns in vector " + }, + { + "bbox": [ + 130, + 459, + 481, + 530 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 130, + 459, + 481, + 530 + ], + "type": "text", + "content": ", given " + }, + { + "bbox": [ + 130, + 459, + 481, + 530 + ], + "type": "inline_equation", + "content": "\\mathbf{b} \\neq \\mathbf{0}" + }, + { + "bbox": [ + 130, + 459, + 481, + 530 + ], + "type": "text", + "content": ", the condition for obtaining a unique (approximate) solution is " + }, + { + "bbox": [ + 130, + 459, + 481, + 530 + ], + "type": "inline_equation", + "content": "\\mathrm{rank}(\\mathbf{A}) = n + 1" + }, + { + "bbox": [ + 130, + 459, + 481, + 530 + ], + "type": "text", + "content": ", which is, in practice, easy to achieve because a large number of pixels are available. The minimum setting to determine scale " + }, + { + "bbox": [ + 130, + 459, + 481, + 530 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 130, + 459, + 481, + 530 + ], + "type": "text", + "content": " is selecting two pixels with distinct depths in one of the multi-view images, which makes a full-rank " + }, + { + "bbox": [ + 130, + 459, + 481, + 530 + ], + "type": "inline_equation", + "content": "2 \\times 2" + }, + { + "bbox": [ + 130, + 459, + 481, + 530 + ], + "type": "text", + "content": " matrix " + }, + { + "bbox": [ + 130, + 459, + 481, + 530 + ], + "type": "inline_equation", + "content": "\\mathbf{A}" + }, + { + "bbox": [ + 130, + 459, + 481, + 530 + ], + "type": "text", + "content": " for two unknowns " + }, + { + "bbox": [ + 130, + 459, + 481, + 530 + ], + "type": "inline_equation", + "content": "\\mathbf{x} = [\\bar{g}_i, \\bar{s}]^\\top" + }, + { + "bbox": [ + 130, + 459, + 481, + 530 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 531, + 481, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 531, + 481, + 578 + ], + "spans": [ + { + "bbox": [ + 130, + 531, + 481, + 578 + ], + "type": "text", + "content": "In practice, we can select many pixels from multiple views to rely on robust estimation. With a sufficiently large number of selected pixels, we construct a tall and skinny matrix " + }, + { + "bbox": [ + 130, + 531, + 481, + 578 + ], + "type": "inline_equation", + "content": "\\mathbf{A}" + }, + { + "bbox": [ + 130, + 531, + 481, + 578 + ], + "type": "text", + "content": " and use " + }, + { + "bbox": [ + 130, + 531, + 481, + 578 + ], + "type": "inline_equation", + "content": "\\ell_1" + }, + { + "bbox": [ + 130, + 531, + 481, + 578 + ], + "type": "text", + "content": " residual minimization to effectively disregard outliers as" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 250, + 589, + 481, + 608 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 250, + 589, + 481, + 608 + ], + "spans": [ + { + "bbox": [ + 250, + 589, + 481, + 608 + ], + "type": "interline_equation", + "content": "\\mathbf {x} ^ {*} = \\underset {\\mathbf {x}} {\\operatorname {a r g m i n}} \\| \\mathbf {A x} - \\mathbf {b} \\| _ {1}. \\tag {9}", + "image_path": "2ff4cad9f7698bbb1dfd700118b3b7c8062ac5a8b17328dbaf867a39ef32e7bf.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 130, + 617, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 481, + 665 + ], + "type": "text", + "content": "We use the Iteratively Reweighted Least Squares (IRLS) method [12] for deriving the approximate solution " + }, + { + "bbox": [ + 130, + 617, + 481, + 665 + ], + "type": "inline_equation", + "content": "\\mathbf{x}^*" + }, + { + "bbox": [ + 130, + 617, + 481, + 665 + ], + "type": "text", + "content": ". Obviously, other choices of robust estimation methods, such as RANdom SAmple Consensus (RANSAC) [8], are viable for alleviating the impact of outliers." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 159, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 159, + 91, + 448, + 102 + ], + "type": "text", + "content": "Resolving Scale Ambiguity in MV3D Reconstruction using DP Sensors" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 326, + 128 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 326, + 128 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 326, + 128 + ], + "type": "text", + "content": "4.3 Confidence-based Pixel Selection" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 135, + 482, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 135, + 482, + 232 + ], + "spans": [ + { + "bbox": [ + 130, + 135, + 482, + 232 + ], + "type": "text", + "content": "So far, we have assumed that the blur sizes " + }, + { + "bbox": [ + 130, + 135, + 482, + 232 + ], + "type": "inline_equation", + "content": "b_{ij}" + }, + { + "bbox": [ + 130, + 135, + 482, + 232 + ], + "type": "text", + "content": " are available to our method. For obtaining the blur sizes, we use the method proposed by Punnappurath et al. [25] that optimizes Eq. (1). For our case, we only require blur size estimates for a small number of pixels but not for all the pixels since our goal is determining the scene scale " + }, + { + "bbox": [ + 130, + 135, + 482, + 232 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 130, + 135, + 482, + 232 + ], + "type": "text", + "content": " using Eq. (8). We, therefore, develop a pixel selection heuristic to choose a small number of pixels that have reliable blur size estimates. Specifically, we select the reliable pixels in each view, followed by selecting the views that produce reliable estimates of scene scale." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 243, + 482, + 314 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 243, + 482, + 314 + ], + "spans": [ + { + "bbox": [ + 130, + 243, + 482, + 314 + ], + "type": "text", + "content": "Pixel-selection within a view We begin by adopting the confidence scores of [25] that indicate the reliability of the blur size estimates based on the residual of the blur size optimization and the textured-ness (magnitude of image gradient) of the pixel of interest. We slightly modified their original confidence score using a depth-based criterion since we observed that the large depth variation in an image patch negatively affects the blur size estimation." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 315, + 482, + 374 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 315, + 482, + 374 + ], + "spans": [ + { + "bbox": [ + 130, + 315, + 482, + 374 + ], + "type": "text", + "content": "Let us consider one of the views and let " + }, + { + "bbox": [ + 130, + 315, + 482, + 374 + ], + "type": "inline_equation", + "content": "c_{j}" + }, + { + "bbox": [ + 130, + 315, + 482, + 374 + ], + "type": "text", + "content": " be the confidence score defined in [25] at the " + }, + { + "bbox": [ + 130, + 315, + 482, + 374 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 130, + 315, + 482, + 374 + ], + "type": "text", + "content": "-th pixel. In our case, a depth value " + }, + { + "bbox": [ + 130, + 315, + 482, + 374 + ], + "type": "inline_equation", + "content": "z_{j}^{\\prime}" + }, + { + "bbox": [ + 130, + 315, + 482, + 374 + ], + "type": "text", + "content": " up to scale is available at the pixel, and we can assess the variance of the depth values within an image patch " + }, + { + "bbox": [ + 130, + 315, + 482, + 374 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_j" + }, + { + "bbox": [ + 130, + 315, + 482, + 374 + ], + "type": "text", + "content": " centered at the " + }, + { + "bbox": [ + 130, + 315, + 482, + 374 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 130, + 315, + 482, + 374 + ], + "type": "text", + "content": "-th pixel as " + }, + { + "bbox": [ + 130, + 315, + 482, + 374 + ], + "type": "inline_equation", + "content": "\\text{var}(\\{z_k'\\})" + }, + { + "bbox": [ + 130, + 315, + 482, + 374 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 315, + 482, + 374 + ], + "type": "inline_equation", + "content": "k \\in \\mathcal{P}_j" + }, + { + "bbox": [ + 130, + 315, + 482, + 374 + ], + "type": "text", + "content": ". Based on the variance of the depths, we re-define the confidence score as" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 247, + 383, + 481, + 406 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 247, + 383, + 481, + 406 + ], + "spans": [ + { + "bbox": [ + 247, + 383, + 481, + 406 + ], + "type": "interline_equation", + "content": "c _ {j} ^ {*} \\triangleq \\frac {c _ {j}}{\\operatorname {v a r} \\left(\\left\\{z _ {k} ^ {\\prime} \\right\\}\\right)}, \\quad k \\in \\mathcal {P} _ {j}. \\tag {10}", + "image_path": "9d15a0c2ac8fd6a2b4ac28c3740a92c51c733e12cdb6c924385192e513f44035.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 413, + 482, + 497 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 413, + 482, + 497 + ], + "spans": [ + { + "bbox": [ + 130, + 413, + 482, + 497 + ], + "type": "text", + "content": "The greater depth variation in an image patch leads to a larger variance and lower confidence score, and vice versa. Based on the new confidence score " + }, + { + "bbox": [ + 130, + 413, + 482, + 497 + ], + "type": "inline_equation", + "content": "c_{j}^{*}" + }, + { + "bbox": [ + 130, + 413, + 482, + 497 + ], + "type": "text", + "content": ", we select the top " + }, + { + "bbox": [ + 130, + 413, + 482, + 497 + ], + "type": "inline_equation", + "content": "T_{c}\\%" + }, + { + "bbox": [ + 130, + 413, + 482, + 497 + ], + "type": "text", + "content": " of pixels and regard them as reliable pixels that have good defocus blur estimates. Although the variance of depths varies with the depth scale, which is unknown in our case, it does not affect the confidence score " + }, + { + "bbox": [ + 130, + 413, + 482, + 497 + ], + "type": "inline_equation", + "content": "c_{j}^{*}" + }, + { + "bbox": [ + 130, + 413, + 482, + 497 + ], + "type": "text", + "content": " within a view because the same scale is applied to all the pixels in the view. Figure 4 depicts the pixel selection within a view." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 508, + 482, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 508, + 482, + 555 + ], + "spans": [ + { + "bbox": [ + 130, + 508, + 482, + 555 + ], + "type": "text", + "content": "View selection In the multi-view images for SfM, there may be images that are not suitable for scale estimation. For example, a view with small depth variations is not preferred because the small depth variations make the linear system of Eq. (8) unstable. Therefore, we further select reliable views for scale estimation." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 556, + 482, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 556, + 482, + 664 + ], + "spans": [ + { + "bbox": [ + 130, + 556, + 482, + 664 + ], + "type": "text", + "content": "Since our method can estimate scale " + }, + { + "bbox": [ + 130, + 556, + 482, + 664 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 130, + 556, + 482, + 664 + ], + "type": "text", + "content": " for each view as long as the depths up to scale are available, we first compute the scale predictions for each view independently using the selected pixels within each view by solving Eq. (8). Ideally, the scales estimated for each view should match; however, in reality, they do not due to outliers. We first disregard the views with clear outliers, i.e., views with excessively small blur size variations and views that yield negative values of scale estimates. The small blur size variations indicate that the scene contains small depth variations. We simply set a threshold to discard the views with small blur size variations. Namely, when the difference between maximum" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 233, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 233, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 233, + 100 + ], + "type": "text", + "content": "K. Ashida et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 264, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 264, + 258 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 264, + 258 + ], + "type": "text", + "content": "and minimum blur size estimates becomes less than or equal to " + }, + { + "bbox": [ + 130, + 116, + 264, + 258 + ], + "type": "inline_equation", + "content": "T_{p}" + }, + { + "bbox": [ + 130, + 116, + 264, + 258 + ], + "type": "text", + "content": " pixels, the view is unused for the estimation. The negative scale estimates could arise due to various reasons, such as image degradation by motion blur and inaccurate estimates of either or both blur sizes and depths. We simply discard such views from the estimation." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 260, + 264, + 283 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 260, + 264, + 283 + ], + "spans": [ + { + "bbox": [ + 130, + 260, + 264, + 283 + ], + "type": "text", + "content": "Finally, from the remaining views, we select " + }, + { + "bbox": [ + 130, + 260, + 264, + 283 + ], + "type": "inline_equation", + "content": "N_{v}" + }, + { + "bbox": [ + 130, + 260, + 264, + 283 + ], + "type": "text", + "content": " views" + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 272, + 118, + 330, + 227 + ], + "blocks": [ + { + "bbox": [ + 272, + 118, + 330, + 227 + ], + "lines": [ + { + "bbox": [ + 272, + 118, + 330, + 227 + ], + "spans": [ + { + "bbox": [ + 272, + 118, + 330, + 227 + ], + "type": "image", + "image_path": "b44b869b5c5b430d1397251a070bbe309c341c0c997e40a92e1e48b056720c79.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 269, + 237, + 482, + 270 + ], + "lines": [ + { + "bbox": [ + 269, + 237, + 482, + 270 + ], + "spans": [ + { + "bbox": [ + 269, + 237, + 482, + 270 + ], + "type": "text", + "content": "Fig. 5: View selection. We select the reliable views for scale estimation based on the median of the scale values estimated for each view." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 332, + 118, + 420, + 228 + ], + "blocks": [ + { + "bbox": [ + 332, + 118, + 420, + 228 + ], + "lines": [ + { + "bbox": [ + 332, + 118, + 420, + 228 + ], + "spans": [ + { + "bbox": [ + 332, + 118, + 420, + 228 + ], + "type": "image", + "image_path": "8322b3ad199f9bc14638ca95278ee26c0f4379af9dbc9279ca712cda713f85b8.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 421, + 125, + 481, + 227 + ], + "blocks": [ + { + "bbox": [ + 421, + 125, + 481, + 227 + ], + "lines": [ + { + "bbox": [ + 421, + 125, + 481, + 227 + ], + "spans": [ + { + "bbox": [ + 421, + 125, + 481, + 227 + ], + "type": "image", + "image_path": "f82b47f8ca6de496701585b22cc0500bbc01bf80664eed57d1597a7d03b04b64.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 284, + 482, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 284, + 482, + 319 + ], + "spans": [ + { + "bbox": [ + 130, + 284, + 482, + 319 + ], + "type": "text", + "content": "that are close to the median of the estimated scales as illustrated in Fig. 5. Once the appropriate views are selected, by using all the selected pixels within the views, we create the linear system of Eq. (8) and re-estimate the scale " + }, + { + "bbox": [ + 130, + 284, + 482, + 319 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 130, + 284, + 482, + 319 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 338, + 279, + 350 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 338, + 279, + 350 + ], + "spans": [ + { + "bbox": [ + 131, + 338, + 279, + 350 + ], + "type": "text", + "content": "4.4 Implementation Details" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 359, + 482, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 359, + 482, + 443 + ], + "spans": [ + { + "bbox": [ + 130, + 359, + 482, + 443 + ], + "type": "text", + "content": "For SfM and MVS, we use the commercial software Metashape1. For DP blur size estimation, we use the method proposed by Punnappurath et al. [25] as described, except for the smoothing operation implemented as post-processing. All captured images are undistorted using the parameters estimated by MVS before being fed into the blur size estimation. Our method is not restricted by the choice of SfM+MVS and blur size estimation methods, but we found these to be the best choices today." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 444, + 482, + 480 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 444, + 482, + 480 + ], + "spans": [ + { + "bbox": [ + 130, + 444, + 482, + 480 + ], + "type": "text", + "content": "For the focal length " + }, + { + "bbox": [ + 130, + 444, + 482, + 480 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 130, + 444, + 482, + 480 + ], + "type": "text", + "content": ", we use the one from the lens specification. The aperture diameter " + }, + { + "bbox": [ + 130, + 444, + 482, + 480 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 130, + 444, + 482, + 480 + ], + "type": "text", + "content": " is calculated using f-number " + }, + { + "bbox": [ + 130, + 444, + 482, + 480 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 444, + 482, + 480 + ], + "type": "text", + "content": " of the lens as " + }, + { + "bbox": [ + 130, + 444, + 482, + 480 + ], + "type": "inline_equation", + "content": "l = \\frac{f}{N}" + }, + { + "bbox": [ + 130, + 444, + 482, + 480 + ], + "type": "text", + "content": ". For the hyperparameters, we use " + }, + { + "bbox": [ + 130, + 444, + 482, + 480 + ], + "type": "inline_equation", + "content": "T_{c} = 10" + }, + { + "bbox": [ + 130, + 444, + 482, + 480 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 444, + 482, + 480 + ], + "type": "inline_equation", + "content": "T_{p} = 2" + }, + { + "bbox": [ + 130, + 444, + 482, + 480 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 444, + 482, + 480 + ], + "type": "inline_equation", + "content": "N_{v} = 7" + }, + { + "bbox": [ + 130, + 444, + 482, + 480 + ], + "type": "text", + "content": " throughout the experiments." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 131, + 498, + 224, + 513 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 498, + 224, + 513 + ], + "spans": [ + { + "bbox": [ + 131, + 498, + 224, + 513 + ], + "type": "text", + "content": "5 Experiment" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 525, + 482, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 525, + 482, + 574 + ], + "spans": [ + { + "bbox": [ + 130, + 525, + 482, + 574 + ], + "type": "text", + "content": "We assess the accuracy of our scale estimation method both quantitatively and qualitatively. We use multi-view DP images captured in six scenes with different capturing settings, such as, cameras (DSLR and smartphone), lenses, and aperture sizes." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 586, + 482, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 586, + 482, + 647 + ], + "spans": [ + { + "bbox": [ + 130, + 586, + 482, + 647 + ], + "type": "text", + "content": "Recording setting We use two different types of cameras equipped with a DP sensor: a DSLR camera and a smartphone. For the DSLR camera setting, we use Canon EOS 5D Mark IV with three different lenses: EF35mm F1.4L II USM, EF50mm F1.8 STM, and EF85mm F1.8 USM. For each scene, we also change the aperture size to assess the effect of the aperture size on our method. The" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 159, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 159, + 91, + 448, + 102 + ], + "type": "text", + "content": "Resolving Scale Ambiguity in MV3D Reconstruction using DP Sensors" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 133, + 653, + 465, + 666 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 653, + 465, + 666 + ], + "spans": [ + { + "bbox": [ + 133, + 653, + 465, + 666 + ], + "type": "text", + "content": "1 Agisoft Metashape https://www.agisoft.com/, last accessed on July 12, 2024." + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "text", + "content": "maximum aperture (" + }, + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "inline_equation", + "content": "f/1.4" + }, + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "inline_equation", + "content": "f/1.8" + }, + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "text", + "content": " depending on the lens), " + }, + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "inline_equation", + "content": "f/4.0" + }, + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "inline_equation", + "content": "f/8.0" + }, + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "text", + "content": " are used for the experiment depending on the target scene. The images from Canon's DLSR camera are recorded as Canon Raw images with the CR2 format, and the left and right DP views are extracted using DPRSplit2. For a more accessible setup, we also use a smartphone, Google Pixel 4, which has a focal length of " + }, + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "inline_equation", + "content": "4.38 \\mathrm{~mm}" + }, + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "text", + "content": " and a fixed aperture of " + }, + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "inline_equation", + "content": "f/1.73" + }, + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "text", + "content": ". To obtain the left and right DP views, we use the capturing software used in [11]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 200, + 482, + 271 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 200, + 482, + 271 + ], + "spans": [ + { + "bbox": [ + 130, + 200, + 482, + 271 + ], + "type": "text", + "content": "For each scene for each camera setting, we capture about 30 images using a DSLR, and 60 images using a smartphone. For fair comparisons across different lens and aperture sizes, for each viewpoint, we use a tripod to keep the camera fixed and record images with different camera settings. To evaluate the scale estimates, we obtain the ground-truth scene scale by placing markers with known sizes in the scenes." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 284, + 482, + 317 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 284, + 482, + 317 + ], + "spans": [ + { + "bbox": [ + 130, + 284, + 482, + 317 + ], + "type": "text", + "content": "Evaluation metric As a quantitative evaluation metric, we use the scale ratio " + }, + { + "bbox": [ + 130, + 284, + 482, + 317 + ], + "type": "inline_equation", + "content": "r_s" + }, + { + "bbox": [ + 130, + 284, + 482, + 317 + ], + "type": "text", + "content": " of the estimated scene scale " + }, + { + "bbox": [ + 130, + 284, + 482, + 317 + ], + "type": "inline_equation", + "content": "s_{\\mathrm{est}}" + }, + { + "bbox": [ + 130, + 284, + 482, + 317 + ], + "type": "text", + "content": " to the ground-truth scale " + }, + { + "bbox": [ + 130, + 284, + 482, + 317 + ], + "type": "inline_equation", + "content": "s_{\\mathrm{gt}}" + }, + { + "bbox": [ + 130, + 284, + 482, + 317 + ], + "type": "text", + "content": ", which is defined as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 258, + 317, + 355, + 340 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 258, + 317, + 355, + 340 + ], + "spans": [ + { + "bbox": [ + 258, + 317, + 355, + 340 + ], + "type": "interline_equation", + "content": "\\mathrm {S c a l e r a t i o} r _ {s} = \\frac {s _ {\\mathrm {e s t}}}{s _ {\\mathrm {g t}}}.", + "image_path": "bebd98cdae9ace9d9ee4063ffbe0d059f7d9f70eabef6313fe3d8cdddf36863a.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 344, + 482, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 344, + 482, + 415 + ], + "spans": [ + { + "bbox": [ + 130, + 344, + 482, + 415 + ], + "type": "text", + "content": "In the above metric, closer to 1 indicates good scale estimation. The scale ratio smaller than 1 indicates that the scale is under-estimated, i.e., the estimated scale is smaller than the true scale, and the greater scale ratio corresponds to the over-estimation. While scale ratio " + }, + { + "bbox": [ + 130, + 344, + 482, + 415 + ], + "type": "inline_equation", + "content": "r_s" + }, + { + "bbox": [ + 130, + 344, + 482, + 415 + ], + "type": "text", + "content": " is defined for each scene and camera setting, we also define the average error of the scale ratios " + }, + { + "bbox": [ + 130, + 344, + 482, + 415 + ], + "type": "inline_equation", + "content": "e_s" + }, + { + "bbox": [ + 130, + 344, + 482, + 415 + ], + "type": "text", + "content": " to characterize the overall effectiveness as" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 171, + 422, + 443, + 447 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 171, + 422, + 443, + 447 + ], + "spans": [ + { + "bbox": [ + 171, + 422, + 443, + 447 + ], + "type": "interline_equation", + "content": "\\mathrm {A v e r a g e e r r o r o f s c a l e r a t i o s} e _ {s} = \\frac {1}{n _ {s}} \\sum \\left(\\max (r _ {s}, r _ {s} ^ {- 1}) - 1\\right),", + "image_path": "14dee2881f38b977ebcb535eb953458bd52185b34ead350248cecb106dde6fad.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 455, + 481, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 455, + 481, + 479 + ], + "spans": [ + { + "bbox": [ + 130, + 455, + 481, + 479 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 455, + 481, + 479 + ], + "type": "inline_equation", + "content": "n_s" + }, + { + "bbox": [ + 130, + 455, + 481, + 479 + ], + "type": "text", + "content": " is the number of settings, and " + }, + { + "bbox": [ + 130, + 455, + 481, + 479 + ], + "type": "inline_equation", + "content": "\\max(r_s, r_s^{-1})" + }, + { + "bbox": [ + 130, + 455, + 481, + 479 + ], + "type": "text", + "content": " is for taking into account the under- and over-estimations." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 491, + 482, + 588 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 491, + 482, + 588 + ], + "spans": [ + { + "bbox": [ + 130, + 491, + 482, + 588 + ], + "type": "text", + "content": "Quantitative results The results of the scale estimation for six scenes using two DP cameras are summarized in Table 1. The DSLR camera uses three lenses and different aperture sizes. Overall, our method can reasonably well estimate the absolute scene scale just by using a DP camera for multi-view capture, considering the fact that scale ratio " + }, + { + "bbox": [ + 130, + 491, + 482, + 588 + ], + "type": "inline_equation", + "content": "r_s" + }, + { + "bbox": [ + 130, + 491, + 482, + 588 + ], + "type": "text", + "content": " can range in " + }, + { + "bbox": [ + 130, + 491, + 482, + 588 + ], + "type": "inline_equation", + "content": "(0, \\infty)" + }, + { + "bbox": [ + 130, + 491, + 482, + 588 + ], + "type": "text", + "content": ". The average error " + }, + { + "bbox": [ + 130, + 491, + 482, + 588 + ], + "type": "inline_equation", + "content": "e_s" + }, + { + "bbox": [ + 130, + 491, + 482, + 588 + ], + "type": "text", + "content": " across scenes and lenses is approximately 0.219 as shown in the table. From comparing the average errors for each lens, we see that the difference in lenses does not significantly affect the overall performance." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 599, + 482, + 636 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 599, + 482, + 636 + ], + "spans": [ + { + "bbox": [ + 130, + 599, + 482, + 636 + ], + "type": "text", + "content": "Visual results Figures 6 and 7 show visual examples of our results by the DSLR and smartphone, as well as the estimated blur sizes and the ground-truth depth maps. We can confirm the same trend as in the quantitative results, where the" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 233, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 233, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 233, + 100 + ], + "type": "text", + "content": "K. Ashida et al." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 133, + 641, + 481, + 664 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 641, + 481, + 664 + ], + "spans": [ + { + "bbox": [ + 133, + 641, + 481, + 664 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 133, + 641, + 481, + 664 + ], + "type": "text", + "content": " DPRSplit https://www.fastrawviewer.com/DPRSplit, last accessed on July 12, 2024." + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 134, + 125, + 480, + 544 + ], + "blocks": [ + { + "bbox": [ + 134, + 125, + 480, + 544 + ], + "lines": [ + { + "bbox": [ + 134, + 125, + 480, + 544 + ], + "spans": [ + { + "bbox": [ + 134, + 125, + 480, + 544 + ], + "type": "image", + "image_path": "163b087c947ddd277a14c36e9adb178c5f9437e92ad9e82bb17f67c03f2344b5.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 554, + 482, + 654 + ], + "lines": [ + { + "bbox": [ + 130, + 554, + 482, + 654 + ], + "spans": [ + { + "bbox": [ + 130, + 554, + 482, + 654 + ], + "type": "text", + "content": "Fig. 6: Visual results of real scenes captured by the DSLR (Canon EOS 5D Mark IV). Our method yields reasonable estimates of scene scales just by using a DP camera for multi-view capture, which is theoretically unrecoverable by standard SfM alone. The top three rows show the results of the same scene (and from almost the same viewpoint) but with different aperture sizes or focal lengths. The bottom rows show the results of different scenes. In (a), the distance indicates the true distance obtained from the GT 3D model. In (d), the distance in the estimated scale is shown. (b) shows estimated blur size and selected pixels based on our pixel selection method (highlighted in the figure)." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 159, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 159, + 91, + 447, + 102 + ], + "type": "text", + "content": "Resolving Scale Ambiguity in MV3D Reconstruction using DP Sensors" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 135, + 168, + 480, + 260 + ], + "blocks": [ + { + "bbox": [ + 130, + 114, + 482, + 159 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 482, + 159 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 482, + 159 + ], + "type": "text", + "content": "Table 1: Scale ratio " + }, + { + "bbox": [ + 130, + 114, + 482, + 159 + ], + "type": "inline_equation", + "content": "r_s" + }, + { + "bbox": [ + 130, + 114, + 482, + 159 + ], + "type": "text", + "content": " between the ground truth and the estimated scale for each scene, lens, and aperture size. Maximum aperture varies depending on the lens (i.e., " + }, + { + "bbox": [ + 130, + 114, + 482, + 159 + ], + "type": "inline_equation", + "content": "f/1.4" + }, + { + "bbox": [ + 130, + 114, + 482, + 159 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 130, + 114, + 482, + 159 + ], + "type": "inline_equation", + "content": "f/1.8" + }, + { + "bbox": [ + 130, + 114, + 482, + 159 + ], + "type": "text", + "content": "). " + }, + { + "bbox": [ + 130, + 114, + 482, + 159 + ], + "type": "inline_equation", + "content": "e_s" + }, + { + "bbox": [ + 130, + 114, + 482, + 159 + ], + "type": "text", + "content": " represents the average errors of the corresponding row/column/total. “-” indicates that we did not record the scene under the setting." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 135, + 168, + 480, + 260 + ], + "lines": [ + { + "bbox": [ + 135, + 168, + 480, + 260 + ], + "spans": [ + { + "bbox": [ + 135, + 168, + 480, + 260 + ], + "type": "table", + "html": "
DSLRR (35mm)DSLRR (50mm)DSLRR (85mm)Smartphone f/1.73es
f/1.4f/4.0f/8.0f/1.8f/4.0f/8.0f/1.8f/4.0f/8.0
Scene 11.4191.088-1.0841.084-1.1001.231-1.0850.156
Scene 21.2251.238-1.0471.232-1.2471.302-1.2520.220
Scene 31.1991.229-1.0551.352-1.2381.226-1.5190.260
Scene 41.1811.1071.4771.0271.2661.4691.2251.1901.2961.0340.227
Scene 51.1711.0721.3141.0671.0991.2851.0381.2051.4251.0680.174
Scene 61.1481.073-1.0831.029-1.1231.126-1.3880.139
es0.2240.1350.3960.0610.1770.3770.1620.2130.3610.2240.219
", + "image_path": "4386b9e7bc109c60529a19c283bd2eababc511000eb3624bf94ed5905141c5b9.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 136, + 277, + 482, + 414 + ], + "blocks": [ + { + "bbox": [ + 136, + 277, + 482, + 414 + ], + "lines": [ + { + "bbox": [ + 136, + 277, + 482, + 414 + ], + "spans": [ + { + "bbox": [ + 136, + 277, + 482, + 414 + ], + "type": "image", + "image_path": "e1f22bce1241aa74bdb7db06f1eaadc5750fc95f6e267f6ddbc0098de011ee44.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 426, + 482, + 460 + ], + "lines": [ + { + "bbox": [ + 130, + 426, + 482, + 460 + ], + "spans": [ + { + "bbox": [ + 130, + 426, + 482, + 460 + ], + "type": "text", + "content": "Fig. 7: Visual results of real scenes captured by a smartphone (Google Pixel 4). Similar to the results by DSLR, our method estimates reasonable scene scales despite the DP blur being relatively smaller than DSLR's." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 487, + 482, + 547 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 487, + 482, + 547 + ], + "spans": [ + { + "bbox": [ + 130, + 487, + 482, + 547 + ], + "type": "text", + "content": "depth maps with estimated scale well represent the actual depths. Our method estimates the scene scale with less than 0.2 error ratio for several settings, where the blur size estimation is relatively stable. For these cases, 3D reconstruction with our absolute scale estimation allows high-quality measurements with an error of several centimeters in wide outdoor scenes." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 570, + 482, + 643 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 643 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 643 + ], + "type": "text", + "content": "Stability of our method To assess the stability of our method against variation of input images, we capture additional " + }, + { + "bbox": [ + 130, + 570, + 482, + 643 + ], + "type": "inline_equation", + "content": "200\\sim 300" + }, + { + "bbox": [ + 130, + 570, + 482, + 643 + ], + "type": "text", + "content": " images for Scenes 1 and 2 using the DSLR equipped with a " + }, + { + "bbox": [ + 130, + 570, + 482, + 643 + ], + "type": "inline_equation", + "content": "35\\mathrm{mm}" + }, + { + "bbox": [ + 130, + 570, + 482, + 643 + ], + "type": "text", + "content": " lens and " + }, + { + "bbox": [ + 130, + 570, + 482, + 643 + ], + "type": "inline_equation", + "content": "f / 1.4" + }, + { + "bbox": [ + 130, + 570, + 482, + 643 + ], + "type": "text", + "content": " aperture size and repeatedly estimate the scale using different sets of images. Figure 8 shows an overlay of estimated models with repeated experiments and the minimum and maximum estimated scale ratios, " + }, + { + "bbox": [ + 130, + 570, + 482, + 643 + ], + "type": "inline_equation", + "content": "r_{s\\min}" + }, + { + "bbox": [ + 130, + 570, + 482, + 643 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 570, + 482, + 643 + ], + "type": "inline_equation", + "content": "r_{s\\max}" + }, + { + "bbox": [ + 130, + 570, + 482, + 643 + ], + "type": "text", + "content": ", with 10 trials." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 233, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 233, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 233, + 100 + ], + "type": "text", + "content": "K. Ashida et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 137, + 110, + 276, + 192 + ], + "blocks": [ + { + "bbox": [ + 137, + 110, + 276, + 192 + ], + "lines": [ + { + "bbox": [ + 137, + 110, + 276, + 192 + ], + "spans": [ + { + "bbox": [ + 137, + 110, + 276, + 192 + ], + "type": "image", + "image_path": "61d7b3b0ae9eb49d6a3502f402859a6da3a1fd5fc0aa6d7716db9ddac7e3e04d.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 157, + 196, + 250, + 205 + ], + "lines": [ + { + "bbox": [ + 157, + 196, + 250, + 205 + ], + "spans": [ + { + "bbox": [ + 157, + 196, + 250, + 205 + ], + "type": "inline_equation", + "content": "r_{s\\min}" + }, + { + "bbox": [ + 157, + 196, + 250, + 205 + ], + "type": "text", + "content": " : 1.158, " + }, + { + "bbox": [ + 157, + 196, + 250, + 205 + ], + "type": "inline_equation", + "content": "r_{s\\max}" + }, + { + "bbox": [ + 157, + 196, + 250, + 205 + ], + "type": "text", + "content": " : 1.255" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 277, + 110, + 380, + 193 + ], + "blocks": [ + { + "bbox": [ + 277, + 110, + 380, + 193 + ], + "lines": [ + { + "bbox": [ + 277, + 110, + 380, + 193 + ], + "spans": [ + { + "bbox": [ + 277, + 110, + 380, + 193 + ], + "type": "image", + "image_path": "3db705a2597fe43b78da37ac93467a4725f089d1bb2b850924998d19cb22aa7e.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 278, + 196, + 371, + 205 + ], + "lines": [ + { + "bbox": [ + 278, + 196, + 371, + 205 + ], + "spans": [ + { + "bbox": [ + 278, + 196, + 371, + 205 + ], + "type": "inline_equation", + "content": "r_{s\\min}" + }, + { + "bbox": [ + 278, + 196, + 371, + 205 + ], + "type": "text", + "content": " : 1.174, " + }, + { + "bbox": [ + 278, + 196, + 371, + 205 + ], + "type": "inline_equation", + "content": "r_{s\\max}" + }, + { + "bbox": [ + 278, + 196, + 371, + 205 + ], + "type": "text", + "content": " : 1.285" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 380, + 196, + 474, + 206 + ], + "lines": [ + { + "bbox": [ + 380, + 196, + 474, + 206 + ], + "spans": [ + { + "bbox": [ + 380, + 196, + 474, + 206 + ], + "type": "inline_equation", + "content": "r_{s\\min}" + }, + { + "bbox": [ + 380, + 196, + 474, + 206 + ], + "type": "text", + "content": " : 1.202, " + }, + { + "bbox": [ + 380, + 196, + 474, + 206 + ], + "type": "inline_equation", + "content": "r_{s\\max}" + }, + { + "bbox": [ + 380, + 196, + 474, + 206 + ], + "type": "text", + "content": " : 1.383" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 131, + 214, + 482, + 248 + ], + "lines": [ + { + "bbox": [ + 131, + 214, + 482, + 248 + ], + "spans": [ + { + "bbox": [ + 131, + 214, + 482, + 248 + ], + "type": "text", + "content": "Fig. 8: Overlay of estimated models with repeated experiments and the minimum and maximum estimated scale ratios with 10 trials. Blue and red point clouds represent the ones with scale ratios " + }, + { + "bbox": [ + 131, + 214, + 482, + 248 + ], + "type": "inline_equation", + "content": "r_{s\\max}" + }, + { + "bbox": [ + 131, + 214, + 482, + 248 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 214, + 482, + 248 + ], + "type": "inline_equation", + "content": "r_{s\\min}" + }, + { + "bbox": [ + 131, + 214, + 482, + 248 + ], + "type": "text", + "content": ", respectively." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 381, + 110, + 472, + 193 + ], + "blocks": [ + { + "bbox": [ + 381, + 110, + 472, + 193 + ], + "lines": [ + { + "bbox": [ + 381, + 110, + 472, + 193 + ], + "spans": [ + { + "bbox": [ + 381, + 110, + 472, + 193 + ], + "type": "image", + "image_path": "d50a86c1bda7f2ca4c53def7ce22e40dfdb37ba22a7037a36e5f6498c21913c9.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 271, + 306, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 271, + 306, + 415 + ], + "spans": [ + { + "bbox": [ + 130, + 271, + 306, + 415 + ], + "type": "text", + "content": "Comparison with scale estimation by monocular depth estimation We compare our method with a single-image metric depth estimation, ZoeDepth [6], for the same scenes and lens setup used. Since it estimates metric depth for each pixel, we compute the per-image scale as the average ratio between ZoeDepth and MVS depth maps. We then average the per-image scale over seven inlier viewpoints selected using the same procedure as our view selection." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 415, + 306, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 415, + 306, + 462 + ], + "spans": [ + { + "bbox": [ + 130, + 415, + 306, + 462 + ], + "type": "text", + "content": "The average error " + }, + { + "bbox": [ + 130, + 415, + 306, + 462 + ], + "type": "inline_equation", + "content": "e_{s}" + }, + { + "bbox": [ + 130, + 415, + 306, + 462 + ], + "type": "text", + "content": " across scenes and lenses using ZoeDepth is 0.679, which is more than 3 times larger than ours (0.219). Figure 9 confirms this" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 463, + 481, + 487 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 463, + 481, + 487 + ], + "spans": [ + { + "bbox": [ + 130, + 463, + 481, + 487 + ], + "type": "text", + "content": "trend through the visualization of the estimated scale. More detailed comparisons are shown in the supplementary material." + } + ] + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 336, + 275, + 460, + 359 + ], + "blocks": [ + { + "bbox": [ + 336, + 275, + 460, + 359 + ], + "lines": [ + { + "bbox": [ + 336, + 275, + 460, + 359 + ], + "spans": [ + { + "bbox": [ + 336, + 275, + 460, + 359 + ], + "type": "image", + "image_path": "8f6034f5d9ae24edea8557f7eaeaa475057db4fba9c6f38e95cbab533182586b.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 341, + 361, + 453, + 372 + ], + "lines": [ + { + "bbox": [ + 341, + 361, + 453, + 372 + ], + "spans": [ + { + "bbox": [ + 341, + 361, + 453, + 372 + ], + "type": "inline_equation", + "content": "r_{sours}" + }, + { + "bbox": [ + 341, + 361, + 453, + 372 + ], + "type": "text", + "content": " : 1.038, " + }, + { + "bbox": [ + 341, + 361, + 453, + 372 + ], + "type": "inline_equation", + "content": "r_{sZoe\\text{depth}}" + }, + { + "bbox": [ + 341, + 361, + 453, + 372 + ], + "type": "text", + "content": " : 0.643" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 311, + 380, + 482, + 447 + ], + "lines": [ + { + "bbox": [ + 311, + 380, + 482, + 447 + ], + "spans": [ + { + "bbox": [ + 311, + 380, + 482, + 447 + ], + "type": "text", + "content": "Fig. 9: Comparison with monocular scale estimation. Blue and red points represent the ones of the scale ratio estimated by our method (" + }, + { + "bbox": [ + 311, + 380, + 482, + 447 + ], + "type": "inline_equation", + "content": "r_{s\\text{ours}}" + }, + { + "bbox": [ + 311, + 380, + 482, + 447 + ], + "type": "text", + "content": ") and ZoeDepth (" + }, + { + "bbox": [ + 311, + 380, + 482, + 447 + ], + "type": "inline_equation", + "content": "r_{sZoeDepth}" + }, + { + "bbox": [ + 311, + 380, + 482, + 447 + ], + "type": "text", + "content": "), respectively. Black points represent the GT model." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 498, + 482, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 498, + 482, + 617 + ], + "spans": [ + { + "bbox": [ + 130, + 498, + 482, + 617 + ], + "type": "text", + "content": "Limitations and failure cases Although our method successfully estimates the scene scale in many cases, we find several failure modes in our experiments. Since our method is based on defocus blur, scale estimation from pan-focus images becomes challenging. In particular, since the aperture of smartphone cameras is smaller than that of DSLR's, it was necessary for us to include photographs that exhibit large depth variations to produce relatively large blur like the top row of Fig. 7. Although the pan-focus images are preferable in SfM and MVS for better correspondence matching, in practice, small apertures cannot always be used. This is because the small aperture limits the amount of incoming light, requiring longer exposure times, and resulting in camera motion blurs." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "The success of our method relies on the accurate estimation of both defocus blur sizes and depth maps up to scale. We found several examples where the scale estimation is affected by the blur size estimation errors, while the depth maps up to scale are almost always stably estimated. Figure 10 shows a failure" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 159, + 90, + 448, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 90, + 448, + 103 + ], + "spans": [ + { + "bbox": [ + 159, + 90, + 448, + 103 + ], + "type": "text", + "content": "Resolving Scale Ambiguity in MV3D Reconstruction using DP Sensors" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 137, + 114, + 208, + 161 + ], + "blocks": [ + { + "bbox": [ + 137, + 114, + 208, + 161 + ], + "lines": [ + { + "bbox": [ + 137, + 114, + 208, + 161 + ], + "spans": [ + { + "bbox": [ + 137, + 114, + 208, + 161 + ], + "type": "image", + "image_path": "3c7a6f6e962f45147b489b907b96e2ae5219d2706a3753f9b87cbf9c19d6630c.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 142, + 163, + 200, + 171 + ], + "lines": [ + { + "bbox": [ + 142, + 163, + 200, + 171 + ], + "spans": [ + { + "bbox": [ + 142, + 163, + 200, + 171 + ], + "type": "text", + "content": "(a) Input image" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 211, + 113, + 296, + 161 + ], + "blocks": [ + { + "bbox": [ + 211, + 113, + 296, + 161 + ], + "lines": [ + { + "bbox": [ + 211, + 113, + 296, + 161 + ], + "spans": [ + { + "bbox": [ + 211, + 113, + 296, + 161 + ], + "type": "image", + "image_path": "e018bb67874fe57f81ab84c83b4bcf5fb8abf4acd4ca5c9fba73a2fa61036d0e.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 229, + 163, + 279, + 171 + ], + "lines": [ + { + "bbox": [ + 229, + 163, + 279, + 171 + ], + "spans": [ + { + "bbox": [ + 229, + 163, + 279, + 171 + ], + "type": "text", + "content": "(b) GT depth" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 175, + 482, + 242 + ], + "lines": [ + { + "bbox": [ + 130, + 175, + 482, + 242 + ], + "spans": [ + { + "bbox": [ + 130, + 175, + 482, + 242 + ], + "type": "text", + "content": "Fig. 10: An example of failure case (Scene 3 captured with a " + }, + { + "bbox": [ + 130, + 175, + 482, + 242 + ], + "type": "inline_equation", + "content": "50\\mathrm{mm}" + }, + { + "bbox": [ + 130, + 175, + 482, + 242 + ], + "type": "text", + "content": " lens and " + }, + { + "bbox": [ + 130, + 175, + 482, + 242 + ], + "type": "inline_equation", + "content": "f / 4.0" + }, + { + "bbox": [ + 130, + 175, + 482, + 242 + ], + "type": "text", + "content": " aperture size) producing scale ratio " + }, + { + "bbox": [ + 130, + 175, + 482, + 242 + ], + "type": "inline_equation", + "content": "r_s = 1.352" + }, + { + "bbox": [ + 130, + 175, + 482, + 242 + ], + "type": "text", + "content": " due to the large error in blur size estimation. Ideal blur sizes (d) are calculated from the ground-truth (GT) scene depth (b) and Eq. (6). Since the target scene is mostly textureless or a repetition of fine textures, most of the pixels in the estimated blur sizes (c) deviate from ideal values, making the selection of the inlier pixels challenging." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 303, + 113, + 386, + 161 + ], + "blocks": [ + { + "bbox": [ + 303, + 113, + 386, + 161 + ], + "lines": [ + { + "bbox": [ + 303, + 113, + 386, + 161 + ], + "spans": [ + { + "bbox": [ + 303, + 113, + 386, + 161 + ], + "type": "image", + "image_path": "cee5efa42d29122ffa21fcda926f8fb19df6df3a9300c64009ffa49c106a0490.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 303, + 163, + 385, + 171 + ], + "lines": [ + { + "bbox": [ + 303, + 163, + 385, + 171 + ], + "spans": [ + { + "bbox": [ + 303, + 163, + 385, + 171 + ], + "type": "text", + "content": "(c) Estimated blur size" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 390, + 113, + 477, + 161 + ], + "blocks": [ + { + "bbox": [ + 390, + 113, + 477, + 161 + ], + "lines": [ + { + "bbox": [ + 390, + 113, + 477, + 161 + ], + "spans": [ + { + "bbox": [ + 390, + 113, + 477, + 161 + ], + "type": "image", + "image_path": "176d138d1f5f26a1feb898aa0a02a4da5db369a7a064d2c1087c34b6ffcc9370.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 402, + 163, + 466, + 171 + ], + "lines": [ + { + "bbox": [ + 402, + 163, + 466, + 171 + ], + "spans": [ + { + "bbox": [ + 402, + 163, + 466, + 171 + ], + "type": "text", + "content": "(d) Ideal blur size" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 262, + 482, + 382 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 262, + 482, + 382 + ], + "spans": [ + { + "bbox": [ + 130, + 262, + 482, + 382 + ], + "type": "text", + "content": "mode of blur size estimation as well as the estimated scene depth, whose scale ratio " + }, + { + "bbox": [ + 130, + 262, + 482, + 382 + ], + "type": "inline_equation", + "content": "r_s" + }, + { + "bbox": [ + 130, + 262, + 482, + 382 + ], + "type": "text", + "content": " reaches as large as 1.352 (Scene 3 captured with " + }, + { + "bbox": [ + 130, + 262, + 482, + 382 + ], + "type": "inline_equation", + "content": "50\\mathrm{mm}" + }, + { + "bbox": [ + 130, + 262, + 482, + 382 + ], + "type": "text", + "content": " lens and " + }, + { + "bbox": [ + 130, + 262, + 482, + 382 + ], + "type": "inline_equation", + "content": "f / 4.0" + }, + { + "bbox": [ + 130, + 262, + 482, + 382 + ], + "type": "text", + "content": " aperture size). To assess the quality of blur size estimates shown in Fig. 10 (c), we synthesize the ideal blur sizes in Fig. 10 (d) using the ground-truth depth of Fig. 10 (b) that is computed from the markers and the image formation model of Eq. (6). Most of the blur size estimates are far from the ideal blur sizes due to the target object being almost textureless or the repetition of fine-grained texture, which makes blur size optimization challenging. We further analyze the accuracy and characteristics of the existing blur size estimation methods in the supplementary material." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 131, + 393, + 222, + 404 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 393, + 222, + 404 + ], + "spans": [ + { + "bbox": [ + 131, + 393, + 222, + 404 + ], + "type": "text", + "content": "6 Discussions" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 410, + 482, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 410, + 482, + 553 + ], + "spans": [ + { + "bbox": [ + 130, + 410, + 482, + 553 + ], + "type": "text", + "content": "This paper has shown that the scale information in multi-view 3d reconstruction, which is theoretically unsolvable by SfM and MVS itself, can be inferred from the defocus blur information provided by DP imaging. Our method estimates the absolute scale without relying on reference objects or data priors but by simply capturing multi-view images with a DP camera. The key to the scale recovery is the unique feature of DP that relates defocus disparity, blur size, and scene depth, from which we can affinely relate the reciprocal of the depth and the defocus blur sizes given by DP images. Together with the pixel selection method for improving robustness using the confidence information from the off-the-shelf DP blur estimation method as well as the statistics of the depth information, we simultaneously recover the scene scale and the unknown focus distance via a simple linear regression." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 554, + 482, + 625 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 554, + 482, + 625 + ], + "spans": [ + { + "bbox": [ + 130, + 554, + 482, + 625 + ], + "type": "text", + "content": "Experiments show that our method successfully estimates the absolute scale in many cases. Since the estimation accuracy of the blur size strongly affects the performance of our method, the analysis of DP blur characteristics will be beneficial in the future. With a method for a more accurate and stabler estimation of blur sizes in DP images, our method will produce more accurate absolute scales for multi-view 3D reconstruction." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 636, + 241, + 649 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 636, + 241, + 649 + ], + "spans": [ + { + "bbox": [ + 132, + 636, + 241, + 649 + ], + "type": "text", + "content": "Acknowledgments" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 653, + 463, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 653, + 463, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 653, + 463, + 665 + ], + "type": "text", + "content": "This work was supported by JSPS KAKENHI Grant Number JP23H05491." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 233, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 233, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 233, + 100 + ], + "type": "text", + "content": "K. Ashida et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 197, + 126 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 197, + 126 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 197, + 126 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 137, + 480, + 666 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 138, + 137, + 480, + 169 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 137, + 480, + 169 + ], + "spans": [ + { + "bbox": [ + 138, + 137, + 480, + 169 + ], + "type": "text", + "content": "1. Abuolaim, A., Afifi, M., Brown, M.S.: Improving single-image defocus deblurring: How dual-pixel images help through multi-task learning. In: IEEE/CVF Winter Conference on Applications of Computer Vision (WACV). pp. 82-90 (2022)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 170, + 480, + 191 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 170, + 480, + 191 + ], + "spans": [ + { + "bbox": [ + 138, + 170, + 480, + 191 + ], + "type": "text", + "content": "2. Abuolaim, A., Brown, M.S.: Defocus deblurring using dual-pixel data. In: Proceedings of European Conference on Computer Vision (ECCV). pp. 111-126 (2020)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 191, + 480, + 234 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 191, + 480, + 234 + ], + "spans": [ + { + "bbox": [ + 138, + 191, + 480, + 234 + ], + "type": "text", + "content": "3. Abuolaim, A., Delbracio, M., Kelly, D., Brown, M.S., Milanfar, P.: Learning to reduce defocus blur by realistically modeling dual-pixel data. In: Proceedings of IEEE/CVF International Conference on Computer Vision (ICCV). pp. 2289-2298 (2021)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 234, + 480, + 266 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 234, + 480, + 266 + ], + "spans": [ + { + "bbox": [ + 138, + 234, + 480, + 266 + ], + "type": "text", + "content": "4. Abuolaim, A., Punnappurath, A., Brown, M.S.: Revisiting autofocus for smartphone cameras. In: Proceedings of European Conference on Computer Vision (ECCV). pp. 545-559 (2018)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 266, + 480, + 311 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 266, + 480, + 311 + ], + "spans": [ + { + "bbox": [ + 138, + 266, + 480, + 311 + ], + "type": "text", + "content": "5. Abuolaim, A., Timofte, R., Brown, M.S.: NTIRE 2021 challenge for defocus deblurring using dual-pixel images: Methods and results. In: Proceedings of IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW). pp. 578-587 (2021)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 311, + 480, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 311, + 480, + 331 + ], + "spans": [ + { + "bbox": [ + 138, + 311, + 480, + 331 + ], + "type": "text", + "content": "6. Bhat, S.F., Birkl, R., Wofk, D., Wonka, P., Müller, M.: Zoedepth: Zero-shot transfer by combining relative and metric depth. arXiv preprint arXiv:2302.12288 (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 331, + 480, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 331, + 480, + 365 + ], + "spans": [ + { + "bbox": [ + 138, + 331, + 480, + 365 + ], + "type": "text", + "content": "7. Engel, J., Stuckler, J., Cremers, D.: Large-scale direct SLAM with stereo cameras. In: IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). pp. 1935-1942 (2015)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 365, + 480, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 365, + 480, + 396 + ], + "spans": [ + { + "bbox": [ + 138, + 365, + 480, + 396 + ], + "type": "text", + "content": "8. Fischler, M.A., Bolles, R.C.: Random sample consensus: A paradigm for model fitting with applications to image analysis and automated cartography. Communications of the ACM 24(6), 381-395 (1981)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 396, + 480, + 429 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 396, + 480, + 429 + ], + "spans": [ + { + "bbox": [ + 138, + 396, + 480, + 429 + ], + "type": "text", + "content": "9. Frost, D., Prisacariu, V., Murray, D.: Recovering stable scale in monocular SLAM using object-supplemented bundle adjustment. IEEE Transactions on Robotics 34(3), 736-747 (2018)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 429, + 480, + 460 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 429, + 480, + 460 + ], + "spans": [ + { + "bbox": [ + 138, + 429, + 480, + 460 + ], + "type": "text", + "content": "0. Garg, R., Wadhwa, N., Ansari, S., Barron, J.T.: Learning single camera depth estimation using dual-pixels. In: Proceedings of IEEE/CVF International Conference on Computer Vision (ICCV). pp. 7628-7637 (2019)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 460, + 480, + 494 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 460, + 480, + 494 + ], + "spans": [ + { + "bbox": [ + 138, + 460, + 480, + 494 + ], + "type": "text", + "content": "1. Garg, R., Wadhwa, N., Ansari, S., Barron, J.T.: Learning single camera depth estimation using dual-pixels. In: Proceedings of IEEE/CVF International Conference on Computer Vision (ICCV). pp. 7628-7637 (2019)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 494, + 480, + 504 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 494, + 480, + 504 + ], + "spans": [ + { + "bbox": [ + 138, + 494, + 480, + 504 + ], + "type": "text", + "content": "2. Gentle, J.: Matrix Albegra. Springer Texts in Statistics, Springer, New York (2007)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 504, + 480, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 504, + 480, + 536 + ], + "spans": [ + { + "bbox": [ + 138, + 504, + 480, + 536 + ], + "type": "text", + "content": "3. Giubilato, R., Chiodini, S., Pertile, M., Debei, S.: Scale correct monocular visual odometry using a lidar altimeter. In: IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). pp. 3694-3700 (2018)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 536, + 480, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 536, + 480, + 568 + ], + "spans": [ + { + "bbox": [ + 138, + 536, + 480, + 568 + ], + "type": "text", + "content": "4. Herrmann, C., Bowen, R.S., Wadhwa, N., Garg, R., He, Q., Barron, J.T., Zabih, R.: Learning to autofocus. In: Proceedings of IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 2227-2236 (2020)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 568, + 480, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 568, + 480, + 601 + ], + "spans": [ + { + "bbox": [ + 138, + 568, + 480, + 601 + ], + "type": "text", + "content": "5. Jung, S.H., Heo, Y.S.: Disparity probability volume guided defocus deblurring using dual pixel data. In: Proceedings of International Conference on Information and Communication Technology Convergence (ICTC). pp. 305-308 (2021)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 601, + 480, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 601, + 480, + 633 + ], + "spans": [ + { + "bbox": [ + 138, + 601, + 480, + 633 + ], + "type": "text", + "content": "6. Kang, M., Choe, J., Ha, H., Jeon, H.G., Im, S., Kweon, I.S.: Facial depth and normal estimation using single dual-pixel camera. In: Proceedings of European Conference on Computer Vision (ECCV) (2022)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 138, + 633, + 480, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 633, + 480, + 666 + ], + "spans": [ + { + "bbox": [ + 138, + 633, + 480, + 666 + ], + "type": "text", + "content": "7. Kashiwagi, M., Mishima, N., Kozakaya, T., Hiura, S.: Deep depth from aberration map. In: Proceedings of IEEE/CVF International Conference on Computer Vision (ICCV) (2019)" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 159, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 159, + 91, + 447, + 102 + ], + "type": "text", + "content": "Resolving Scale Ambiguity in MV3D Reconstruction using DP Sensors" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "type": "text", + "content": "18. Kim, D., Jang, H., Kim, I., Kim, M.H.: Spatio-focal bidirectional disparity estimation from a dual-pixel image. In: Proceedings of IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 5023-5032 (2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 151, + 480, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 151, + 480, + 183 + ], + "spans": [ + { + "bbox": [ + 132, + 151, + 480, + 183 + ], + "type": "text", + "content": "19. Knorr, S.B., Kurz, D.: Leveraging the user's face for absolute scale estimation in handheld monocular SLAM. In: Proceedings of IEEE International Symposium on Mixed and Augmented Reality (ISMAR). pp. 11-17 (2016)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 184, + 480, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 184, + 480, + 206 + ], + "spans": [ + { + "bbox": [ + 132, + 184, + 480, + 206 + ], + "type": "text", + "content": "20. Lee, S.H., de Croon, G.: Stability-based scale estimation for monocular SLAM. IEEE Robotics and Automation Letters 3(2), 780-787 (2018)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 207, + 480, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 207, + 480, + 239 + ], + "spans": [ + { + "bbox": [ + 132, + 207, + 480, + 239 + ], + "type": "text", + "content": "21. Li, F., Guo, H., Santo, H., Okura, F., Matsushita, Y.: Learning to synthesize photorealistic dual-pixel images from RGBD frames. In: International Conference on Computational Photography (ICCP). pp. 1-11 (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 240, + 480, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 240, + 480, + 274 + ], + "spans": [ + { + "bbox": [ + 132, + 240, + 480, + 274 + ], + "type": "text", + "content": "22. Mishima, N., Seki, A., Hiura, S.: Absolute scale from varifocal monocular camera through SfM and defocus combined. In: Proceedings of British Machine Vision Conference (BMVC) (2021)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 274, + 480, + 307 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 274, + 480, + 307 + ], + "spans": [ + { + "bbox": [ + 132, + 274, + 480, + 307 + ], + "type": "text", + "content": "23. Nützi, G., Weiss, S., Scaramuzza, D., Siegwart, R.: Fusion of IMU and vision for absolute scale estimation in monocular SLAM. Journal of Intelligent & Robotic Systems 61(1-4), 287-299 (2011)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 308, + 480, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 308, + 480, + 352 + ], + "spans": [ + { + "bbox": [ + 132, + 308, + 480, + 352 + ], + "type": "text", + "content": "24. Pan, L., Chowdhury, S., Hartley, R., Liu, M., Zhang, H., Li, H.: Dual pixel exploration: Simultaneous depth estimation and image restoration. In: Proceedings of IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 4340-4349 (2021)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 353, + 480, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 353, + 480, + 385 + ], + "spans": [ + { + "bbox": [ + 132, + 353, + 480, + 385 + ], + "type": "text", + "content": "25. Punnappurath, A., Abuolaim, A., Affi, M., Brown, M.S.: Modeling defocus-disparity in dual-pixel sensors. In: International Conference on Computational Photography (ICCP). pp. 1-12 (2020)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 386, + 480, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 386, + 480, + 418 + ], + "spans": [ + { + "bbox": [ + 132, + 386, + 480, + 418 + ], + "type": "text", + "content": "26. Punnappurath, A., Brown, M.S.: Reflection removal using a dual-pixel sensor. In: Proceedings of IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 1556-1565 (2019)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 419, + 480, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 419, + 480, + 453 + ], + "spans": [ + { + "bbox": [ + 132, + 419, + 480, + 453 + ], + "type": "text", + "content": "27. Roussel, T., Van Eycken, L., Tuytelaars, T.: Monocular depth estimation in new environments with absolute scale. In: IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). pp. 1735-1741 (2019)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 453, + 480, + 497 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 453, + 480, + 497 + ], + "spans": [ + { + "bbox": [ + 132, + 453, + 480, + 497 + ], + "type": "text", + "content": "28. Rukhovich, D., Mouritzen, D., Kaestner, R., Rufli, M., Velizhev, A.: Estimation of absolute scale in monocular SLAM using synthetic data. In: Proceedings of IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW). pp. 803-812 (2019)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 498, + 480, + 541 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 498, + 480, + 541 + ], + "spans": [ + { + "bbox": [ + 132, + 498, + 480, + 541 + ], + "type": "text", + "content": "29. Scaramuzza, D., Fraundorfer, F., Pollefeys, M., Siegwart, R.: Absolute scale in structure from motion from a single vehicle mounted camera by exploiting nonholonomic constraints. In: Proceedings of IEEE/CVF International Conference on Computer Vision (ICCV). pp. 1413-1419 (2009)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 543, + 480, + 575 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 543, + 480, + 575 + ], + "spans": [ + { + "bbox": [ + 132, + 543, + 480, + 575 + ], + "type": "text", + "content": "30. Shibata, A., Fujii, H., Yamashita, A., Asama, H.: Absolute scale structure from motion using a refractive plate. In: Proceedings of IEEE/SICE International Symposium on System Integration (SII). pp. 540-545 (2015)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 576, + 480, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 576, + 480, + 609 + ], + "spans": [ + { + "bbox": [ + 132, + 576, + 480, + 609 + ], + "type": "text", + "content": "31. Shibata, A., Fujii, H., Yamashita, A., Asama, H.: Scale-reconstructable structure from motion using refraction with a single camera. In: IEEE International Conference on Robotics and Automation (ICRA). pp. 5239-5244 (2015)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 609, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 609, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 609, + 480, + 632 + ], + "type": "text", + "content": "32. Shiozaki, T., Dissanayake, G.: Eliminating scale drift in monocular SLAM using depth from defocus. IEEE Robotics and Automation Letters 3(1), 581-587 (2017)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "type": "text", + "content": "33. Song, S., Chandraker, M.: Robust scale estimation in real-time monocular sfm for autonomous driving. In: Proceedings of IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 1566-1573 (2014)" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 233, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 233, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 233, + 100 + ], + "type": "text", + "content": "K. Ashida et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 479 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "text", + "content": "34. Sucar, E., Hayet, J.B.: Bayesian scale estimation for monocular SLAM based on generic object detection for correcting scale drift. In: IEEE International Conference on Robotics and Automation (ICRA). pp. 5152-5158 (2018)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 150, + 482, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 150, + 482, + 182 + ], + "spans": [ + { + "bbox": [ + 130, + 150, + 482, + 182 + ], + "type": "text", + "content": "35. Sumikura, S., Sakurada, K., Kawaguchi, N., Nakamura, R.: Scale estimation of monocular SfM for a multi-modal stereo camera. In: Proceedings of Asian Conference on Computer Vision (ACCV). pp. 281-297 (2019)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 182, + 482, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 182, + 482, + 226 + ], + "spans": [ + { + "bbox": [ + 130, + 182, + 482, + 226 + ], + "type": "text", + "content": "36. Wadhwa, N., Garg, R., Jacobs, D.E., Feldman, B.E., Kanazawa, N., Carroll, R., Movshovitz-Attias, Y., Barron, J.T., Pritch, Y., Levoy, M.: Synthetic depth-of-field with a single-camera mobile phone. ACM Transactions on Graphics (TOG) pp. 1-13 (2018)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 226, + 482, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 226, + 482, + 259 + ], + "spans": [ + { + "bbox": [ + 130, + 226, + 482, + 259 + ], + "type": "text", + "content": "37. Wöhler, C., d'Angelo, P., Krüger, L., Kuhl, A., Groß, H.M.: Monocular 3D scene reconstruction at absolute scale. ISPRS Journal of Photogrammetry and Remote Sensing 64(6), 529-540 (2009)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 259, + 482, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 259, + 482, + 303 + ], + "spans": [ + { + "bbox": [ + 130, + 259, + 482, + 303 + ], + "type": "text", + "content": "38. Xin, S., Wadhwa, N., Xue, T., Barron, J.T., Srinivasan, P.P., Chen, J., Gkioulekas, I., Garg, R.: Defocus map estimation and deblurring from a single dual-pixel image. In: Proceedings of IEEE/CVF International Conference on Computer Vision (ICCV). pp. 2228-2238 (2021)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 303, + 482, + 336 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 303, + 482, + 336 + ], + "spans": [ + { + "bbox": [ + 130, + 303, + 482, + 336 + ], + "type": "text", + "content": "39. Yang, Y., Pan, L., Liu, L., Liu, M.: K3DN: Disparity-aware kernel estimation for dual-pixel defocus deblurring. In: Proceedings of IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 13263-13272 (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 336, + 482, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 336, + 482, + 380 + ], + "spans": [ + { + "bbox": [ + 130, + 336, + 482, + 380 + ], + "type": "text", + "content": "40. Zhang, S., Zhang, J., Tao, D.: Towards scale-aware, robust, and generalizable unsupervised monocular depth estimation by integrating IMU motion dynamics. In: Proceedings of European Conference on Computer Vision (ECCV). pp. 143-160 (2022)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 380, + 482, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 380, + 482, + 413 + ], + "spans": [ + { + "bbox": [ + 130, + 380, + 482, + 413 + ], + "type": "text", + "content": "41. Zhang, Y., Wadhwa, N., Orts-Escolano, S., Häne, C., Fanello, S.R., Garg, R.: Du2Net: Learning depth estimation from dual-cameras and dual-pixels. In: Proceedings of European Conference on Computer Vision (ECCV). pp. 582-598 (2020)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 413, + 482, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 413, + 482, + 445 + ], + "spans": [ + { + "bbox": [ + 130, + 413, + 482, + 445 + ], + "type": "text", + "content": "42. Zhou, D., Dai, Y., Li, H.: Reliable scale estimation and correction for monocular visual odometry. In: Proceedings of IEEE Intelligent Vehicles Symposium (IV). pp. 490-495 (2016)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 445, + 482, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 445, + 482, + 479 + ], + "spans": [ + { + "bbox": [ + 130, + 445, + 482, + 479 + ], + "type": "text", + "content": "43. Zhou, D., Dai, Y., Li, H.: Ground-plane-based absolute scale estimation for monocular visual odometry. IEEE Transactions on Intelligent Transportation Systems 21(2), 791-802 (2019)" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 159, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 159, + 91, + 447, + 102 + ], + "type": "text", + "content": "Resolving Scale Ambiguity in MV3D Reconstruction using DP Sensors" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Responsible Visual Editing/a719b1ec-b6ce-4b23-ac8b-598c2b587aef_content_list.json b/2024/Responsible Visual Editing/a719b1ec-b6ce-4b23-ac8b-598c2b587aef_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..8c25f4f3ae0f572364e7f3531fdeaf539feee333 --- /dev/null +++ b/2024/Responsible Visual Editing/a719b1ec-b6ce-4b23-ac8b-598c2b587aef_content_list.json @@ -0,0 +1,2052 @@ +[ + { + "type": "text", + "text": "Responsible Visual Editing", + "text_level": 1, + "bbox": [ + 346, + 141, + 656, + 162 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Minheng Ni $^{1,2}$ , Yeli Shen $^{2}$ , Lei Zhang $^{1\\boxtimes}$ , and Wangmeng Zuo $^{2,3\\boxtimes}$", + "bbox": [ + 264, + 188, + 736, + 205 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 The Hong Kong Polytechnic University", + "bbox": [ + 361, + 215, + 640, + 231 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ Harbin Institute of Technology", + "bbox": [ + 390, + 231, + 611, + 244 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3 Peng Cheng Laboratory, Guangzhou", + "bbox": [ + 370, + 244, + 630, + 258 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "minheng.ni@connect.polyu.hk, ylshen@stu.hit.edu.cn,", + "bbox": [ + 302, + 260, + 696, + 273 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "cslzhang@comp.polyu.edu.hk, WMzuo@hit.edu.cn", + "bbox": [ + 330, + 273, + 671, + 287 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. With the recent advancements in visual synthesis, there is a growing risk of encountering synthesized images with detrimental effects, such as hate, discrimination, and privacy violations. Unfortunately, it remains unexplored on how to avoid synthesizing harmful images and convert them into responsible ones. In this paper, we present responsible visual editing, which edits risky concepts within an image to more responsible ones with minimal content changes. However, the concepts that need to be edited are often abstract, making them hard to be located and edited. To tackle these challenges, we propose a Cognitive Editor (CoEditor) by harnessing the large multimodal models through a two-stage cognitive process: (1) a perceptual cognitive process to locate what to be edited and (2) a behavioral cognitive process to strategize how to edit. To mitigate the negative implications of harmful images on research, we build a transparent and public dataset, namely AltBear, which expresses harmful information using teddy bears instead of humans. Experiments demonstrate that CoEditor can effectively comprehend abstract concepts in complex scenes, significantly surpassing the baseline models for responsible visual editing. Moreover, we find that the AltBear dataset corresponds well to the harmful content found in real images, providing a safe and effective benchmark for future research. Our source code and dataset can be found at https://github.com/kodenii/Responsible-Visual-Editing.", + "bbox": [ + 259, + 327, + 743, + 632 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Responsible visual editing $\\cdot$ Image editing $\\cdot$ Large multimodal model", + "bbox": [ + 259, + 645, + 740, + 672 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 700, + 375, + 717 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "With the recent advancements in visual synthesis technologies [22,27,29,36], generating or editing highly realistic images has become possible, and the problem of misusing these technologies is arising [16,21,30] because the generated/edited images may contain harmful contents, such as hate, discrimination and privacy violations. While significant progress has been made in image editing [1,2,4] by using instructions to adjust images, it is less explored on how to edit harmful images into responsible ones.", + "bbox": [ + 212, + 733, + 787, + 840 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/24ae8db0d3d5a271d9f8180b0b44f8e7c98e2631da6c447c36eae8d8f6bd3cc8.jpg", + "image_caption": [ + "Fig. 1: Overview of responsible visual editing. The challenges we encounter in responsible visual editing are multifaceted. Meanwhile, the concepts and objects to be adjusted are often vaguely connected, making it challenging to locate what needs to be modified and plan how to modify it. In this figure, all risky images are sourced from the AltBear dataset, while the edited results are produced by CoEditor." + ], + "image_footnote": [], + "bbox": [ + 222, + 143, + 781, + 321 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We formulate the above mentioned problem as a new task, namely responsible visual editing. As shown in Figure 1, we aim to edit specific concepts in images to make them more responsible while minimizing image changes as much as possible. Given the diversity of risky concepts in images and the different types of risks, we divide this task into three sub-tasks: safety, fairness and privacy, covering a wide range of risks in real-world scenarios.", + "bbox": [ + 212, + 438, + 787, + 529 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Existing editing models often require clear user instructions to make specific adjustments in the images [4,11,34], e.g., editing hat to \"change the blue hat into red\". However, in responsible image editing, the concepts that need to be edited are often abstract, e.g., editing violence to \"make an image look less violent\". This makes it challenging to locate what should to be edited and plan how to edit it. To tackle these challenges, we propose a Cognitive Editor (CoEditor) that harnesses large multimodal models (LMM) through a two-stage cognitive process: (1) a perceptual cognitive process to focus on what needs to be edited and (2) a behavioral cognitive process to strategize how to edit.", + "bbox": [ + 212, + 532, + 787, + 667 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To facilitate the research on responsible image editing, we build a transparent and public dataset, namely AltBear. Unlike general image datasets, the AltBear dataset uses fictional teddy bears as the protagonists, replacing humans in the images to convey risky contents. We significantly reduce the potential ethical risks by using teddy bears and cartoonized images.", + "bbox": [ + 212, + 670, + 787, + 746 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Experiments show that CoEditor significantly outperforms baseline models in responsible image editing, validating its effectiveness of comprehending abstract concepts and strategizing the editing process. Furthermore, we find that the AltBear dataset corresponds well to the harmful contents in real images, offering a consistent experimental evaluation. This also validates that AltBear can serve as a safer benchmark for future research.", + "bbox": [ + 212, + 750, + 787, + 839 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "M. Ni et al.", + "bbox": [ + 271, + 114, + 351, + 126 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our contributions are three-fold:", + "bbox": [ + 238, + 146, + 477, + 160 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose a new task, responsible visual editing, and introduce a dataset, AltBear, which maintains high consistency with real-world data by using teddy bears as the protagonists to reduce the potential research risks.", + "- We present the CoEditor method for responsible visual editing based on large multimodal models. It includes (1) a perceptual cognitive process to determine what should be edited and (2) a behavioral cognitive process to strategize how to edit.", + "- Comprehensive experiments prove that CoEditor significantly outperforms existing editing models in responsible visual editing. Our findings reveal the potentials of LMM in responsible AI." + ], + "bbox": [ + 223, + 169, + 785, + 316 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 339, + 387, + 354 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.1 Responsible Visual Synthesis", + "text_level": 1, + "bbox": [ + 215, + 368, + 501, + 383 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "With the advancement of visual synthesis techniques, how to ensure responsible visual editing has also received increasing attention. Some works have been reported to intercept images that may pose risks [3, 10, 31]. Rombach et al. [29] added a Not-Safe-For-Work (NSFW) filter at the end of the generation stage, while Rando et al. [28] chose to classify and intercept images by projecting them onto the latent space using CLIP. Some works [9, 12, 30] employed machine unlearning methods to forget risky concepts. Zhang et al. [35] and Kumari et al. [17] manipulated the latent variables during generation to avoid producing risky concepts. Recently, Ni et al. [21] proposed to intervene in the generation process using a large language model (LLM) to achieve training-free open-vocabulary responsible visual synthesis. However, current researches on responsible visual synthesis focus on text-to-image generation, while responsible visual editing has not yet been well explored.", + "bbox": [ + 212, + 390, + 787, + 587 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2 Image Editing", + "text_level": 1, + "bbox": [ + 215, + 606, + 380, + 621 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Image editing is a classic task in computer vision. Some previous works used CNN-based methods to edit images [6, 8, 32]. By introducing adversarial techniques, StyleGAN and its variants [1, 14, 15] used GAN inversion for image editing. Using quantized models, MaskGIT [5] explored editing based on conditional vectors, and NUWA-LIP [20] explored natural language guided image editing. With the great success of diffusion models, more and more works attempt to use pre-trained text-to-image diffusion models for editing [2, 7, 19]. Prompt-to-prompt [13] obtained varying images by changing the attention during generation. Recently, InstructPix2pix [4] and InstructDiffusion [11] used the data generated by GPT and prompt-to-prompt methods to train the model to understand natural language instructions. However, current researches on image editing require clear or direct instructions. It has not yet been well explored how to use large multimodal models (LMMs) to understand and adjust the relationship between images and complex abstract instructions.", + "bbox": [ + 212, + 628, + 787, + 840 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "Responsible Visual Editing", + "bbox": [ + 549, + 114, + 730, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3 Responsible Visual Editing", + "text_level": 1, + "bbox": [ + 215, + 143, + 513, + 162 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1 Problem Formulation", + "text_level": 1, + "bbox": [ + 215, + 176, + 437, + 189 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The goal of responsible visual editing is to automatically edit a given risky concept $c$ existed in image $x_{\\mathrm{r}}$ , generating a responsible image $x_{\\mathrm{s}}$ while making $x_{\\mathrm{s}}$ visually reasonable and changing the image contents as little as possible. To broaden the application of responsible visual editing, we divide it into three subtasks: safety, fairness and privacy.", + "bbox": [ + 215, + 200, + 785, + 276 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Safety. The safety subtask focuses on inappropriate contents for display in real-world scenarios, such as discrimination, terrorist activities, or violence. This task requires completely removing the risky concepts from the image.", + "bbox": [ + 215, + 277, + 785, + 321 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Fairness. This subtask focuses on fairness issues in real-world scenarios, such as biased contents. It diversifies a specific concept in the image without changing much the image content.", + "bbox": [ + 215, + 321, + 785, + 366 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Privacy. The privacy subtask focuses on privacy issues in real-world scenarios, such as real-world characters. This task requires blurring a specific character in the image. While maintaining the basic meaning of the image, the editing should not be recognizable or traceable.", + "bbox": [ + 215, + 367, + 785, + 426 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2 AltBear Dataset Collection", + "text_level": 1, + "bbox": [ + 215, + 450, + 485, + 463 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We collect a number of risky concepts, such as drug abuse, alcohol, racial discrimination, etc., and divide them into three subtasks: safety, fairness, and privacy. For each concept $c$ , we use ChatGPT [23] to expand it into 100 image scene descriptions and modify the subject of the description to teddy bears. We manually filter out and refine the compelling descriptions, and randomly use DALL-E 2 [26], Stable Diffusion XL [25], and DALL-E 3 [24] to generate the final risky image $x_{\\mathrm{r}}$ . Then, we manually filter the dataset again, selecting the high-quality images. Finally, we obtain 300 groups of images as the test set. For more details, please refer to Supplementary Materials.", + "bbox": [ + 215, + 474, + 785, + 609 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3 Evaluation Metrics", + "text_level": 1, + "bbox": [ + 215, + 633, + 418, + 646 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We propose two metrics to evaluate the performance of responsible image editing: success rate and visual similarity, both of which can be used for automatic machine evaluation and manual human evaluation.", + "bbox": [ + 215, + 657, + 785, + 702 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Success Rate. In the evaluation of success rate, we judge whether the concept $c$ in the responsible image $x_{s}$ still contains risks. For the safety subtask, the concept $c$ should not appear; for the fairness subtask, the diversity of the concept $c$ should be expanded; for the privacy subtask, we require that the identity of specific persons should not be recognized.", + "bbox": [ + 215, + 704, + 785, + 777 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Visual Similarity. In the evaluation of visual similarity, we compute the pixel-level similarity between the responsible image $x_{\\mathrm{s}}$ and the original risky image $x_{\\mathrm{r}}$ . Specifically, if $x_{\\mathrm{s}}$ still contains risks, then the similarity is considered as 0. We take the average similarity as the final result.", + "bbox": [ + 215, + 779, + 785, + 838 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "M. Ni et al.", + "bbox": [ + 271, + 114, + 351, + 126 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "More details of the evaluation metrics can be found in the Supplementary Materials.", + "bbox": [ + 215, + 146, + 784, + 175 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4 Special Markers", + "text_level": 1, + "bbox": [ + 215, + 196, + 395, + 212 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We add unique markers to the images to reduce potential risks in responsible visual editing. As shown in Figure 1, each image contains a marker in the lower right corner, indicating that this image is only used for responsible visual editing research and which subtask it belongs to. In addition, the highlighting symbol A or B represents the original or edited image.", + "bbox": [ + 212, + 219, + 782, + 295 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4 Methodology", + "text_level": 1, + "bbox": [ + 215, + 316, + 379, + 333 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/c1b654e72c8e66f1a79b9f6b179a40db30f9126ee1088ef37a0b2f0d027654ce.jpg", + "image_caption": [ + "Fig. 2: Overview of CoEditor. CoEditor consists of two stages of cognition: (1) a perceptional cognitive process (PCP) to understand what needs to be edited, and (2) a behavioral cognitive process (BCP) to plan how to edit." + ], + "image_footnote": [], + "bbox": [ + 223, + 369, + 781, + 506 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Compared with traditional visual editing that usually edits specific objects or features in an image, responsible visual editing should be able to edit any risky concepts contained in the image. It could be a theme such as violence, a category such as culture, or a person like Bill Gates. This task is very challenging and it contains two stages: (1) understanding the relationship between the image content and the concept to locate the regions to be edited, and (2) planning how to edit the regions to meet the editing conditions and image rationality.", + "bbox": [ + 212, + 585, + 784, + 691 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "With the above considerations, we propose an LMM-based responsible editing model, namely Cognitive Editor (CoEditor). As shown in Figure 2, CoEditor consists of two stages of cognition: (1) perceptional cognitive process (PCP) to locate regions to be edited, and (2) behavioral cognitive process (BCP) to strategize how to edit such regions.", + "bbox": [ + 212, + 691, + 784, + 767 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1 Perceptional Cognitive Process", + "text_level": 1, + "bbox": [ + 215, + 786, + 517, + 801 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Understanding the relationship between the content of an image and abstract concepts is often very difficult, as the concepts may not directly refer to specific", + "bbox": [ + 212, + 809, + 784, + 839 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "Responsible Visual Editing", + "bbox": [ + 549, + 114, + 730, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "contents in the image and it requires reasoning based on common sense. Even the powerful LMM may not accurately understand the concepts (see Section 5.2). To mitigate these issues, we use a visual language chain-of-thoughts method based on visual prompts and language prompts for perceptional cognition to locate the regions to be edited.", + "bbox": [ + 212, + 146, + 787, + 222 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For a risky image $x_{\\mathrm{r}}$ , we need to visually annotate each element for LMM to refer to them accurately. We perform object extraction to obtain the object sequence $\\mathcal{M}$ . In specific, we use Semantic-SAM [18] to extract objects of image $x$ to $n$ masks:", + "bbox": [ + 212, + 222, + 787, + 281 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {M} = \\left\\{m _ {1}, m _ {2}, \\dots , m _ {n} \\right\\}. \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 413, + 282, + 785, + 299 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Similar to SoM [33], we visually prompt the objects in the image to $v_{\\mathrm{p}}$ :", + "bbox": [ + 238, + 306, + 754, + 323 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nv _ {\\mathrm {p}} = \\phi (x, \\mathcal {M}), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 450, + 335, + 785, + 351 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\phi$ is a prompting function that adds a visual tag with a number at the corner of each mask's corresponding location in the image, enabling the LMM to refer to it in numeric tags.", + "bbox": [ + 212, + 362, + 787, + 407 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Meanwhile, in order to understand the possible meanings of the concept, we use knowledge extraction to expand its associated concepts and explanations into the text prompts $l_{\\mathrm{p}}$ :", + "bbox": [ + 212, + 407, + 787, + 453 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nl _ {\\mathrm {p}} = f \\left(c; p _ {\\mathrm {k}} ^ {\\text {i n s}}\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 450, + 454, + 785, + 470 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $p_{\\mathrm{k}}^{\\mathrm{ins}}$ is the instruction for knowledge extraction. Since there is no need to introduce visual information here, $f$ can be either an LLM or an LMM.", + "bbox": [ + 212, + 477, + 787, + 507 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Combining prompts of different modalities $v_{\\mathrm{p}}$ and $l_{\\mathrm{p}}$ , the LMM can complete the cognition of the image and the concept, and obtain the mask of the region $m_{\\mathrm{p}}$ to be modified in the image:", + "bbox": [ + 212, + 508, + 787, + 555 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nm _ {\\mathrm {p}} = f \\left(v _ {\\mathrm {p}}, l _ {\\mathrm {p}}; p _ {\\mathrm {p}} ^ {\\text {i n s}}\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 431, + 565, + 785, + 583 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $p_{\\mathrm{p}}^{\\mathrm{ins}}$ is the instruction for focus generation and $f$ is an LMM. Though we may obtain multiple related regions, for simplicity, we combine all regions into one single $m_{\\mathrm{p}}$ as the result of PCP.", + "bbox": [ + 212, + 595, + 787, + 642 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2 Behavioral Cognitive Process", + "text_level": 1, + "bbox": [ + 214, + 664, + 504, + 680 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In the perception cognitive process, the region $m_p$ to be edited has been identified, and we need to figure out how to edit it. Since the concept may be abstract, the editing model is hard to directly perform editing operations (see Section 5.2). Therefore, we employ the visual language chain-of-thoughts to plan the editing target, i.e., what contents are expected after responsible editing.", + "bbox": [ + 212, + 691, + 787, + 766 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We expand the object $m_p$ identified in the previous stage to $m_p'$ through object extension by enlarging the contour to include surrounding information. Then, we crop the image based on the mask $m_p'$ to obtain visual prompting:", + "bbox": [ + 212, + 767, + 787, + 814 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nv _ {\\mathrm {b}} = x _ {\\mathrm {r}} \\otimes m _ {\\mathrm {p}} ^ {\\prime}. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 450, + 824, + 785, + 842 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "M. Ni et al.", + "bbox": [ + 271, + 114, + 351, + 127 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "At the same time, we perform instruction implementation to get the full language prompt $l_{\\mathrm{b}}$ by concatenating the instruction of the task $p_{\\mathrm{t}}^{\\mathrm{ins}}$ and the input concept $c$ . Then we use the language prompt $l_{\\mathrm{b}}$ together with the visual prompt $v_{\\mathrm{b}}$ to generate the editing target $r_{\\mathrm{b}}$ for the inpainting model:", + "bbox": [ + 212, + 146, + 782, + 205 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\nr _ {\\mathrm {b}} = f \\left(v _ {\\mathrm {b}}, l _ {\\mathrm {b}}; p _ {\\mathrm {b}} ^ {\\text {i n s}}\\right), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 434, + 220, + 785, + 237 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $p_{\\mathrm{b}}^{\\mathrm{ins}}$ is the instruction for editing.", + "bbox": [ + 214, + 248, + 501, + 263 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Finally, we take the original image $x_{\\mathrm{r}}$ , the region to be edited $m_p'$ , and the editing target $r_{\\mathrm{b}}$ as the input to the inpainting model, and obtain the final editing result as follows:", + "bbox": [ + 212, + 263, + 782, + 308 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\nx _ {\\mathrm {s}} = g \\left(x _ {\\mathrm {r}}, m _ {\\mathrm {p}} ^ {\\prime}; r _ {\\mathrm {b}}\\right), \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 437, + 310, + 785, + 327 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $g$ is the inpainting model and $x_{\\mathrm{s}}$ is the final responsible editing result.", + "bbox": [ + 214, + 333, + 764, + 349 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3 Implementation Details", + "text_level": 1, + "bbox": [ + 214, + 373, + 455, + 387 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Our method is training-free. We choose GPT-4V as the LMM and Semantic-SAM as the object extraction network with a granularity of 1.5 and its officially released checkpoint4. For the inpainting model, we choose Stable Diffusion Inpainting trained based on Stable Diffusion v2 and its publicly available checkpoint5. Like most previous works, we adjust the input image to a size of $512 \\times 512$ . All random seeds are fixed to 42. In all experiments, the instructions $p_{\\mathrm{k}}^{\\mathrm{ins}}$ , $p_{\\mathrm{p}}^{\\mathrm{ins}}$ , and $p_{\\mathrm{b}}^{\\mathrm{ins}}$ are fixed, and $p_{\\mathrm{t}}^{\\mathrm{ins}}$ depends on the task. Please refer to the Supplementary Materials for more details.", + "bbox": [ + 212, + 398, + 784, + 521 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5 Experiment", + "text_level": 1, + "bbox": [ + 214, + 547, + 366, + 564 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.1 Experiment Setup", + "text_level": 1, + "bbox": [ + 214, + 580, + 413, + 595 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Since the concept of responsible visual editing is arbitrary and it does not specify the editing area, we select two powerful image editing models, InstructPix2pix [4] and InstructDiffusion [11], in the experiments. They not only allow arbitrary language guidance with the help of LLM but also require no additional masks to indicate the editing region. In addition, we could use the same Stable Diffusion base model in our CoEditor, making our comparison fair.", + "bbox": [ + 212, + 607, + 784, + 696 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We conduct experiments on the AltBear dataset. Since InstructPix2pix and InstructDiffusion do not support responsible editing of images based on the concept, we manually design editing conditions remove {concept} for safety and privacy tasks, and increase the variety of {concept} for fairness tasks. We evaluate the edited images in terms of both the success rate and visual similarity. We measure the results comprehensively via both automatic machine and manual human evaluations.", + "bbox": [ + 212, + 698, + 784, + 801 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "Responsible Visual Editing", + "bbox": [ + 549, + 114, + 730, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 114, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "page_footnote", + "text": "4 https://github.com/UX-Decoder/Semantic-SAM", + "bbox": [ + 217, + 810, + 563, + 825 + ], + "page_idx": 6 + }, + { + "type": "page_footnote", + "text": "5 https://huggingface.co/stabilityai/stable-diffusion-2-inpainting", + "bbox": [ + 220, + 825, + 730, + 840 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/760043f1fdc908a8a17f6a4bd3727a8a16395c71552393e3b367eda388505d8e.jpg", + "table_caption": [ + "Table 1: Overall results on AltBear under machine evaluation. We can find that CoEditor significantly outperforms the baseline models in both success rate and visual similarity, validating the effectiveness of CoEditor." + ], + "table_footnote": [], + "table_body": "
ModelSafetyFairnessPrivacyOverall
\\( Succ^{\\uparrow} \\)\\( Sim^{\\uparrow} \\)\\( Succ^{\\uparrow} \\)\\( Sim^{\\uparrow} \\)\\( Succ^{\\uparrow} \\)\\( Sim^{\\uparrow} \\)\\( Succ^{\\uparrow} \\)\\( Sim^{\\uparrow} \\)
InstructPix2pix21.31%0.196922.11%0.212351.22%0.456231.93%0.2924
InstructDiffusion40.54%0.357421.43%0.171966.33%0.496344.14%0.3497
CoEditor (Ours)70.13%0.565248.96%0.359578.35%0.639565.56%0.5188
", + "bbox": [ + 217, + 198, + 784, + 292 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/5a8a8c0621bd6de0f9672c68a2eb11b040d7917be65e39ab321a0b1880781bf9.jpg", + "table_caption": [ + "Table 2: Overall results on AltBear under human evaluation. In human evaluations, CoEditor shows consistent results with machine evaluations, showing that CoEditor is also effective from a human subjective perspective." + ], + "table_footnote": [], + "table_body": "
ModelSafetyFairnessPrivacyOverall
\\( Succ^{\\uparrow} \\)\\( Sim^{\\uparrow} \\)\\( Succ^{\\uparrow} \\)\\( Sim^{\\uparrow} \\)\\( Succ^{\\uparrow} \\)\\( Sim^{\\uparrow} \\)\\( Succ^{\\uparrow} \\)\\( Sim^{\\uparrow} \\)
InstructPix2pix28.41%0.25239.38%0.083940.00%0.327126.06%0.2217
InstructDiffusion40.91%0.307819.79%0.136956.00%0.345739.08%0.2634
CoEditor (Ours)65.91%0.531850.00%0.360665.00%0.518360.21%0.4692
", + "bbox": [ + 217, + 358, + 784, + 452 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.2 Overall Results", + "text_level": 1, + "bbox": [ + 215, + 481, + 388, + 494 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Quantitative Results. As shown in Table 1 and Table 2, CoEditor shows significant advantages in both machine and human evaluations. We can see that for the success rate, CoEditor has more than $20\\%$ increase in almost all subtasks. This proves that for responsible visual synthesis, using LMM to understand images and concepts and to strategize editing process is crucial. In addition, we find that CoEditor has higher visual similarity compared to traditional models, which reveals the importance of explicit cognition process of editing. It is noticed that the machine evaluation shows similar results to human evaluations, further proving that our editing results are of the same high quality from the perspective of human judgements.", + "bbox": [ + 212, + 507, + 785, + 657 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Qualitative Results. As shown in Figure 3, CoEditor demonstrates strong understanding and editing capabilities in responsible visual synthesis. In the first row of the safety task, we see that the CoEditor not only fully erases the concepts of drugs and alcohol but also maintains the rationality of the picture. However, InstructDiffusion fails to completely remove drugs and generates unreasonable contents in the picture. Moreover, InstructPix2pix does not follow the instruction well. In the second row of the fairness task, CoEditor successfully diversifies the specified concept. It maintains the original layout, while InstructDiffusion results in blurry or damaged outputs and InstructPix2pix has difficulties in executing this instruction. In the third row of the privacy task, the CoEditor blurs the characters' features while maintaining the original meaning with high visual quality. However, InstructDiffusion and InstructPix2pix fail in content removal", + "bbox": [ + 212, + 657, + 787, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "M. Ni et al.", + "bbox": [ + 271, + 114, + 351, + 127 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/1a0d8c82ce0418a3d7765832643c2d0fbb3f9e0afe1ddc0f5db6e06a846c1929.jpg", + "table_caption": [ + "Table 3: Ablation study of AltBear. Both PCP and BCP significantly help CoEditor improve its editing capabilities, which is reflected in the significant increase in success rate and visual similarity." + ], + "table_footnote": [], + "table_body": "
ModelSucc↑Sim↑
CoEditor w/o PCP44.19%0.3855
CoEditor w/o BCP28.02%0.2221
CoEditor (Full)65.56%0.5188
", + "bbox": [ + 375, + 198, + 625, + 271 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "or damage the picture. These examples show that CoEditor can understand the images and concepts and then plan the editing well with the help of LMM. More examples can be found in the Supplementary Materials.", + "bbox": [ + 212, + 297, + 782, + 342 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/8a8b36668cce6de5a4a76205fe9444d40f78846f541495ae8b8739c6ef228754.jpg", + "image_caption": [ + "Fig. 3: Visualization results on AltBear. CoEditor performs well in all subtasks and maintains high visual similarity. The results of CoEditor also have better rationality. InstructDiffusion often over-edits images or produces unreasonable visual effects, while the editing ability of InstructPix2pix is weaker compared with CoEditor." + ], + "image_footnote": [], + "bbox": [ + 225, + 367, + 496, + 565 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/fd46e533b2de3b84823e60c30a54b2d807a5ffaa50b7d356f9933ce94a9f1ddf.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 504, + 367, + 779, + 566 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "5.3 Ablation Study", + "text_level": 1, + "bbox": [ + 214, + 681, + 388, + 698 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Quantitative Results. We conduct ablation experiments on AltBear in Table 3 to explore the effectiveness of the two stages of CoEditor. In the ablation of PCP, since the model cannot locate the area to be edited, we turn to use the largest area in the image. In the ablation of BCP, we use the same instruction as the baseline model for editing. To ensure fairness, all other parameters of the experiments are consistent. As shown in Table 3, both stages of CoEditor are significant for the final results. Without PCP, the CoEditor cannot understand the object to be edited, causing a massive drop in success rate and visual similarity. Without BCP, although the CoEditor can locate the region to be edited,", + "bbox": [ + 212, + 704, + 787, + 840 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "Responsible Visual Editing", + "bbox": [ + 549, + 114, + 730, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "simply asking the model to remove or increase diversity will make it difficult to understand what to edit, ultimately leading to editing failures.", + "bbox": [ + 215, + 146, + 784, + 176 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/691c10985e6c77731abc03c65c572094d0f3ea77b524afb136e5ac6e6e62b39c.jpg", + "image_caption": [ + "Fig. 4: Roles of different components in CoEditor. Without the help of BCP, the CoEditor produces inconsistent results with the concept or visually unreasonable results because it does not know how to edit the content correctly. Without PCP, CoEditor is unable to locate the editing regions, resulting in editing failures." + ], + "image_footnote": [], + "bbox": [ + 223, + 200, + 781, + 402 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Qualitative Results. To investigate whether the two stages of CoEditor, BCP and PCP, can produce the expected outputs, we visualize some examples in Figure 3. One can see that the BCP stage is crucial for correct editing. In most examples, without the BCP stage, the editing result will fail to meet the requirements, or there will be visually unreasonable contents. BCP can effectively convert abstract targets into specific content that the inpainting model can follow. This shows the importance of editing planning. At the same time, PCP plays a decisive role. Without the participation of PCP, most of the edits will fail. This is because most of the concepts to be edited do not directly associate with a specific object in the image. Therefore, a deep understanding of image and concepts using PCP is crucial for CoEditor to edit correctly. The CoEditor can achieve satisfactory effects when both the two modules are used. Extra ablation studies can be found in the Supplementary Materials.", + "bbox": [ + 212, + 500, + 787, + 696 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "5.4 Exploration of Cognitive Process", + "text_level": 1, + "bbox": [ + 214, + 713, + 532, + 729 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "To explore why CoEditor can successfully perform complex and responsible visual editing, we show the intermediate results of various cognitive processes in Figure 5. We can see that PCP can successfully locate regions to be edited, regardless of whether this region is directly related to the concept. Based on the task type, PCP can also accurately find out the region to be edited among multiple identical objects. At the same time, BCP can successfully plan the editing. No matter what the concept is, BCP can accurately associate it with the", + "bbox": [ + 212, + 734, + 787, + 840 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "M. Ni et al.", + "bbox": [ + 271, + 114, + 351, + 126 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/db3085b14647fcde9c7a4d6fc3b6049fbe1d9a3ea9f8a86547a2519f2c09c7b0.jpg", + "image_caption": [ + "Risky Image" + ], + "image_footnote": [], + "bbox": [ + 223, + 159, + 323, + 234 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/c92dc428fd3f738db7567cf43db23e113bcad4f89e0af9d60fd319865808ba20.jpg", + "image_caption": [ + "PCP Result", + "BCP Result" + ], + "image_footnote": [], + "bbox": [ + 352, + 157, + 452, + 234 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "A cuddly teddy bear enjoying a sunny day on a serene picnic with a picturesque house fully intact in the background, surrounded by vibrant greenery.", + "bbox": [ + 486, + 166, + 653, + 228 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/c494380b763d0d6f31b8a24361f8473f21743d2f5e56333332931f2ceb6c495e.jpg", + "image_caption": [ + "Final Result" + ], + "image_footnote": [], + "bbox": [ + 679, + 157, + 779, + 234 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/563302e456afc9aa13a74b271c3081f507bece9347aa64d7695be69d243c054d.jpg", + "image_caption": [ + "Concept: arsion" + ], + "image_footnote": [], + "bbox": [ + 223, + 252, + 323, + 329 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/2eb50c04f9e024962d4bf40df3aef89ba106c76425d5ad6b028046e26c767b73.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 352, + 252, + 452, + 329 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "A teddy bear wearing traditional attire representing different cultures, such as a Scottish kilt, an Indian sari, and a Nigerian abada, showcasing the beautiful variety of cultural dress from around the world.", + "bbox": [ + 486, + 248, + 653, + 332 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/5c15a380009225a6787d2b05db301bb943440941b0cdd31797e053e7ca97f831.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 679, + 252, + 779, + 329 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/93bf4ed5f587f86ef9980443d977ce2a3a371fccce794ea7e59b430b3c568128.jpg", + "image_caption": [ + "Concept: culture", + "Fig. 5: Intermediate results of cognitive process in CoEditor. We can see that PCP can successfully find one or more objects to be edited, even if the object is not directly related to the concept. BCP can generate effective editing target based on the positioning of PCP, even if the scenario is very complex. The effectiveness of the two cognitive processes allows CoEditor to edit images successfully." + ], + "image_footnote": [], + "bbox": [ + 223, + 345, + 323, + 422 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/dc7cd48e7c1584d8a1019ec5f7bf844624b792aae4dcaaef1f4d698737451c84.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 352, + 345, + 452, + 422 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "A fluffy teddy bear with sparkling eyes and shiny footwear stands in front of a vintage microphone on a soft glowing background.", + "bbox": [ + 486, + 354, + 653, + 414 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/59fe2d45b8982f14e836b50c8c62d209f2487c1eb8d506e18d25b6f88e0892f9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 679, + 345, + 779, + 422 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Concept: Marilyn Monroe", + "bbox": [ + 227, + 426, + 372, + 438 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "content in the image to obtain editing target consistent with the image scene. This shows the importance of cognitive processes to CoEditor in concept understanding and editing planning. For further explorations of CoEditor, please see the Supplementary Materials.", + "bbox": [ + 212, + 553, + 787, + 613 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "5.5 Performance Consistency on AltBear and Real-world Data", + "text_level": 1, + "bbox": [ + 214, + 642, + 746, + 657 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "To verify whether AltBear can serve as a replacement of real-world images in subsequent research, we examine the performance consistency between AltBear and real-world images. We build a dataset that has similar size to AltBear, consisting of real-world images collected from the Internet. We conduct experiments on this dataset with all hyper-parameters and models identical to that on AltBear. As shown in Table 4 and 5, CoEditor achieves similar performance to that on AltBear, demonstrating CoEditor's robustness to both synthesized data and real-world data. The baseline models, InstructPix2pix and InstructDiffusion, also show similar performance on AltBear and real-world data. Since the protagonist is not a human but a fictional teddy bear, AltBear significantly reduces the risk of image propagation and display, effectively avoiding ethical risks.", + "bbox": [ + 212, + 672, + 787, + 840 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "Responsible Visual Editing", + "bbox": [ + 549, + 114, + 730, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 116, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/ac48f1fe20236b5aea1eaac6324bdf8f8c5554e35da99d20c5f94c5d855ba38b.jpg", + "table_caption": [ + "Table 4: Results on real-world images under machine evaluation. On real-world datasets, CoEditor shows similar performance to that on AltBear, significantly surpassing the baseline model. Furthermore, we find that the distributions of both success rate and visual similarity are close for real-world images and AltBear." + ], + "table_footnote": [], + "table_body": "
ModelSafetyFairnessPrivacyOverall
\\( Succ^{\\uparrow} \\)\\( Sim^{\\uparrow} \\)\\( Succ^{\\uparrow} \\)\\( Sim^{\\uparrow} \\)\\( Succ^{\\uparrow} \\)\\( Sim^{\\uparrow} \\)\\( Succ^{\\uparrow} \\)\\( Sim^{\\uparrow} \\)
InstructPix2pix14.00%0.12831.89%0.016622.86%0.216411.59%0.1078
InstructDiffusion38.60%0.318514.29%0.103055.81%0.374435.57%0.2638
CoEditor (Ours)59.46%0.474058.18%0.425363.41%0.513960.00%0.4679
", + "bbox": [ + 217, + 210, + 784, + 306 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/8f2bfb54d899ffbd2138cd666bf8297a18dd9b602009785019b95660bba86991.jpg", + "table_caption": [ + "Table 5: Results of real-world images under human evaluation. CoEditor achieves significantly better results than the baseline model in human evaluation. In addition, the quantitative results on AltBear and real-world images in human evaluation are also similar." + ], + "table_footnote": [], + "table_body": "
ModelSafetyFairnessPrivacyOverall
\\( Succ^{\\uparrow} \\)\\( Sim^{\\uparrow} \\)\\( Succ^{\\uparrow} \\)\\( Sim^{\\uparrow} \\)\\( Succ^{\\uparrow} \\)\\( Sim^{\\uparrow} \\)\\( Succ^{\\uparrow} \\)\\( Sim^{\\uparrow} \\)
InstructPix2pix31.25%0.27989.09%0.080415.56%0.132720.56%0.1821
InstructDiffusion47.50%0.376421.82%0.152453.33%0.337841.11%0.2983
CoEditor (Ours)68.75%0.548665.45%0.478780.00%0.646770.56%0.5518
", + "bbox": [ + 217, + 386, + 784, + 479 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "5.6 Real-world Results", + "text_level": 1, + "bbox": [ + 215, + 508, + 419, + 522 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "To better demonstrate the robustness of CoEditor, we show its performance on real-world images and compare with baseline models in Figure 6. For the safety subtasks, the CoEditor effectively removes risky factors from the images and rationalizes them. In the first row, the robber in the first image is turned into an enthusiastic saleswoman, and the craving for drugs in the second image is turned into a craving for food. This also shows the imagination capability of CoEditor. For the fairness subtasks in the second row, the CoEditor completes the image editing while maintaining the visual quality of the image. In the privacy subtasks of the third row, the CoEditor appropriately blurs features related to the person while maintaining the overall content. Similar to the results on AltBear, the baseline models, InstructPix2pix and InstructDiffusion, struggle to maintain visual quality or rationality in some examples, even though they apply editing. In other cases, they produce damaged figures. More real-world examples can be found in the Supplementary Materials.", + "bbox": [ + 212, + 534, + 787, + 746 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "5.7 Analysis of Computational Overhead", + "text_level": 1, + "bbox": [ + 214, + 768, + 566, + 784 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "In order to explore how CoEditor consumes resources, we make comparisons on time and memory usage. We use the same full-precision and optimization settings, e.g., memory-efficient attention for all models.", + "bbox": [ + 212, + 794, + 785, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "M. Ni et al.", + "bbox": [ + 271, + 114, + 351, + 126 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/eab02e419db666486b28844e909dfeababe86416da3515e3dc15050be8fba9a8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 223, + 143, + 496, + 209 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/0035c86a07e4a4778abf6e7249dbe49ee7880418dfeef285b95760d837e427f0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 223, + 210, + 496, + 271 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/98cf5188fec5ab838a9bc9de29ed598134f15c9cc1c9bafae8900433a18492d6.jpg", + "image_caption": [ + "Fig. 6: Visualized results on real-world data. Even in real-world scenarios, CoEditor can still work well. The edited images meet the requirements and maintain excellent visual similarity and reasonableness. Similar to the results on AltBear, InstructDiffusion and InstructPix2pix fail to output well edited images." + ], + "image_footnote": [], + "bbox": [ + 223, + 271, + 496, + 343 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/a917dc032a74ac908fe87c411ad6916173c4352d7ba6ed4f075b40d5289ef618.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 503, + 143, + 779, + 208 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/94971041a4b6f0fcd983f4dae36158d328646b35ec0696d7d01c5c8b769dedb6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 504, + 208, + 779, + 271 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/a3fd1d755e3adb7d2268ce871babff6d86375c6dd9216e7d0016025b1b380dcf.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 504, + 271, + 779, + 343 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/dc0def3fe284c6d33a0d4961a07921de22881514a9d7377309e0b086ceacb680.jpg", + "table_caption": [ + "Table 6: Comparison on inference time." + ], + "table_footnote": [], + "table_body": "
ModelTimeMultiple
InstrcutPix2pix12.43s1×
InstructDiffusion20.75s1.67×
CoEditor14.10s1.13×
PCP Stage4.65s-
BCP Stage9.44s-
", + "bbox": [ + 370, + 452, + 630, + 551 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Time. We calculate the average inference time of different models. For CoEditor, we also calculate the network latency. As shown in Table 6, the speed of CoEditor is even faster than InstructDiffusion [11], and is comparable to Instruct Pix2pix [4].", + "bbox": [ + 212, + 579, + 785, + 638 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "GPU Memory. We calculate the peak VRAM usage of different models. As shown in Table 7, CoEditor consumes the least memory. This is because our chosen models, i.e., Semantic-SAM [18] and Stable Diffusion Inpainting [29], are very lightweight, which allows our model to work with minimal memory.", + "bbox": [ + 212, + 640, + 784, + 700 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "6 Ethics Statement", + "text_level": 1, + "bbox": [ + 214, + 723, + 419, + 739 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "6.1 Inappropriate Content, Privacy, and Discrimination", + "text_level": 1, + "bbox": [ + 214, + 753, + 687, + 770 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "The AltBear dataset uses teddy bears as the protagonist, significantly reducing the risks caused by the public display of inappropriate contents. We manually reviewed all contents to ensure that privacy or discrimination is avoided. We release AltBear under the MIT license to keep transparency.", + "bbox": [ + 212, + 779, + 785, + 839 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "Responsible Visual Editing", + "bbox": [ + 549, + 114, + 730, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 784, + 126 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/ecbf94d55d32df100c20f727e1ac7a5eb61e4bdd971fc682067d6d78d1532fbe.jpg", + "table_caption": [ + "Table 7: Comparison on GPU memory." + ], + "table_footnote": [], + "table_body": "
ModelVRAMMultiple
InstrcutPix2pix18,118MB1×
InstructDiffusion8,378MB0.46×
CoEditor7,328MB0.40×
", + "bbox": [ + 357, + 170, + 645, + 243 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "6.2 Reproducibility", + "text_level": 1, + "bbox": [ + 215, + 271, + 390, + 287 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "We build our method on top of the Stable Diffusion Inpainting model with publicly accessible code and checkpoints with a fixed random seed. We notice that the GPT API cannot guarantee identical responses so that we provide all instruction prompts for reference. To further improve the reproducibility, we also release the dataset, code and our trained model.", + "bbox": [ + 212, + 297, + 785, + 375 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "6.3 Anti-misuse", + "text_level": 1, + "bbox": [ + 215, + 396, + 361, + 411 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In order to ensure that the dataset is not misused, we have designed unique markers (refer to Section 3.4). We added the marker to the bottom right corner of the image to indicate that the image is used for responsible visual editing research only. Specifically, we have defined various tags to indicate whether the image has been edited or contains potential risks.", + "bbox": [ + 212, + 421, + 785, + 496 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In addition, the interaction system of our model always maintains visibility of the concepts to be edited to prevent misuse of the model. We also call on subsequent work to maintain the visibility of concepts like ours, ensuring a beneficial role for the community through open and transparent methods.", + "bbox": [ + 212, + 497, + 785, + 558 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "7 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 580, + 359, + 597 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In this work, we proposed a new task, namely responsible visual editing, which entails editing specific concepts within an image to render it more responsible while minimizing changes. To tackle abstract concepts in responsible visual editing, we proposed a Cognitive Editor (CoEditor) by harnessing a large multimodal model (LMM) through a two-stage process: a perceptual cognitive process to locate what needs to be edited, and a behavioral cognitive process to strategize how to edit. We constructed a transparent and public dataset, AltBear, which represents harmful information using teddy bears instead of humans, thereby mitigating the negative implications on research. Experiments demonstrated that CoEditor can effectively comprehend abstract concepts within complex scenes and significantly surpass the performance of baseline models for responsible visual editing. Moreover, we validated that the AltBear dataset can well serve as a replacement of real-world images in the research of responsible visual editing, providing a safer benchmark. Our findings also revealed the potential of LMM in responsible AI.", + "bbox": [ + 212, + 613, + 787, + 840 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "M. Ni et al.", + "bbox": [ + 271, + 114, + 351, + 126 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgement", + "text_level": 1, + "bbox": [ + 217, + 143, + 392, + 162 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "This work was supported in part by the National Key R&D Program of China (2022YFA1004100), and the National Natural Science Foundation of China (NSFC) under grant No. 62441202.", + "bbox": [ + 215, + 172, + 799, + 219 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 239, + 321, + 255 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Abdal, R., Qin, Y., Wonka, P.: Image2stylegan++: How to edit the embedded images? In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 8296-8305 (2020)", + "2. Avrahami, O., Lischinski, D., Fried, O.: Blended diffusion for text-driven editing of natural images. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18208-18218 (2022)", + "3. Brack, M., Friedrich, F., Hintersdorf, D., Struppek, L., Schramowski, P., Kersting, K.: Sega: Instructing text-to-image models using semantic guidance. Advances in Neural Information Processing Systems 36 (2024)", + "4. Brooks, T., Holynski, A., Efros, A.A.: Instructpix2pix: Learning to follow image editing instructions. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18392-18402 (2023)", + "5. Chang, H., Zhang, H., Jiang, L., Liu, C., Freeman, W.T.: Maskgit: Masked generative image transformer. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11315-11325 (2022)", + "6. Chen, J., Shen, Y., Gao, J., Liu, J., Liu, X.: Language-based image editing with recurrent attentive models. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 8721-8729 (2018)", + "7. Couairon, G., Verbeek, J., Schwenk, H., Cord, M.: Diffedit: Diffusion-based semantic image editing with mask guidance. arXiv preprint arXiv:2210.11427 (2022)", + "8. Dong, H., Yu, S., Wu, C., Guo, Y.: Semantic image synthesis via adversarial learning. In: Proceedings of the IEEE international conference on computer vision. pp. 5706-5714 (2017)", + "9. Gandikota, R., Materzynska, J., Fiotto-Kaufman, J., Bau, D.: Erasing concepts from diffusion models. arXiv preprint arXiv:2303.07345 (2023)", + "0. Gandikota, R., Orgad, H., Belinkov, Y., Materzyńska, J., Bau, D.: Unified concept editing in diffusion models. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 5111-5120 (2024)", + "1. Geng, Z., Yang, B., Hang, T., Li, C., Gu, S., Zhang, T., Bao, J., Zhang, Z., Hu, H., Chen, D., et al.: Instructdiffusion: A generalist modeling interface for vision tasks. arXiv preprint arXiv:2309.03895 (2023)", + "2. Heng, A., Soh, H.: Continual learning for forgetting in deep generative models (2023)", + "3. Hertz, A., Mokady, R., Tenenbaum, J., Aberman, K., Pritch, Y., Cohen-Or, D.: Prompt-to-prompt image editing with cross attention control. arXiv preprint arXiv:2208.01626 (2022)", + "4. Karras, T., Laine, S., Aila, T.: A style-based generator architecture for generative adversarial networks. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 4401-4410 (2019)", + "5. Karras, T., Laine, S., Aittala, M., Hellsten, J., Lehtinen, J., Aila, T.: Analyzing and improving the image quality of stylegan. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 8110-8119 (2020)" + ], + "bbox": [ + 225, + 267, + 785, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "Responsible Visual Editing", + "bbox": [ + 549, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "16. Kumari, N., Zhang, B., Wang, S.Y., Shechtman, E., Zhang, R., Zhu, J.Y.: Ablating Concepts in Text-to-Image Diffusion Models (May 2023), http://arxiv.org/abs/2303.13516, arXiv:2303.13516 [cs]", + "17. Kumari, N., Zhang, B., Wang, S.Y., Shechtman, E., Zhang, R., Zhu, J.Y.: Ablating concepts in text-to-image diffusion models. arXiv preprint arXiv:2303.13516 (2023)", + "18. Li, F., Zhang, H., Sun, P., Zou, X., Liu, S., Yang, J., Li, C., Zhang, L., Gao, J.: Semantic-sam: Segment and recognize anything at any granularity. arXiv preprint arXiv:2307.04767 (2023)", + "19. Meng, C., He, Y., Song, Y., Song, J., Wu, J., Zhu, J.Y., Ermon, S.: Sdedit: Guided image synthesis and editing with stochastic differential equations. arXiv preprint arXiv:2108.01073 (2021)", + "20. Ni, M., Li, X., Zuo, W.: Nuwa-lip: Language-guided image inpainting with defect-free vqgan. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14183-14192 (2023)", + "21. Ni, M., Wu, C., Wang, X., Yin, S., Wang, L., Liu, Z., Duan, N.: Ores: Open-vocabulary responsible visual synthesis. arXiv preprint arXiv:2308.13785 (2023)", + "22. Nichol, A., Dhariwal, P., Ramesh, A., Shyam, P., Mishkin, P., McGrew, B., Sutskever, I., Chen, M.: Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741 (2021)", + "23. OpenAI: Chatgpt. https://chat.openai.com (2022)", + "24. OpenAI: Dall-e 3. https://openai.com/dall-e-3 (2023)", + "25. Podell, D., English, Z., Lacey, K., Blattmann, A., Dockhorn, T., Müller, J., Penna, J., Rombach, R.: Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952 (2023)", + "26. Ramesh, A., Dhariwal, P., Nichol, A., Chu, C., Chen, M.: Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125 1(2), 3 (2022)", + "27. Ramesh, A., Pavlov, M., Goh, G., Gray, S., Voss, C., Radford, A., Chen, M., Sutskever, I.: Zero-shot text-to-image generation. In: International Conference on Machine Learning. pp. 8821-8831. PMLR (2021)", + "28. Rando, J., Paleka, D., Lindner, D., Heim, L., Tramér, F.: Red-teaming the stable diffusion safety filter. arXiv preprint arXiv:2210.04610 (2022)", + "29. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10684-10695 (2022)", + "30. Schramowski, P., Brack, M., Deiseroth, B., Kersting, K.: Safe latent diffusion: Mitigating inappropriate degeneration in diffusion models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 22522-22531 (2023)", + "31. Schuhmann, C., Beaumont, R., Vencu, R., Gordon, C., Wightman, R., Cherti, M., Coombes, T., Katta, A., Mullis, C., Wortsman, M., et al.: Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems 35, 25278-25294 (2022)", + "32. Wang, S.Y., Wang, O., Zhang, R., Owens, A., Efros, A.A.: Cnn-generated images are surprisingly easy to spot... for now. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 8695-8704 (2020)", + "33. Yang, J., Zhang, H., Li, F., Zou, X., Li, C., Gao, J.: Set-of-mark prompting unleashes extraordinary visual grounding in gpt-4v. arXiv preprint arXiv:2310.11441 (2023)", + "34. Yu, T., Feng, R., Feng, R., Liu, J., Jin, X., Zeng, W., Chen, Z.: Inpaint anything: Segment anything meets image inpainting. arXiv preprint arXiv:2304.06790 (2023)" + ], + "bbox": [ + 215, + 146, + 785, + 840 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "M. Ni et al.", + "bbox": [ + 271, + 114, + 351, + 126 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "35. Zhang, E., Wang, K., Xu, X., Wang, Z., Shi, H.: Forget-me-not: Learning to forget in text-to-image diffusion models. arXiv preprint arXiv:2303.17591 (2023)", + "36. Zhang, Y., Wei, Y., Jiang, D., Zhang, X., Zuo, W., Tian, Q.: Control video: Training-free controllable text-to-video generation. arXiv preprint arXiv:2305.13077 (2023)" + ], + "bbox": [ + 215, + 146, + 785, + 215 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "Responsible Visual Editing", + "bbox": [ + 549, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 16 + } +] \ No newline at end of file diff --git a/2024/Responsible Visual Editing/a719b1ec-b6ce-4b23-ac8b-598c2b587aef_model.json b/2024/Responsible Visual Editing/a719b1ec-b6ce-4b23-ac8b-598c2b587aef_model.json new file mode 100644 index 0000000000000000000000000000000000000000..e4cb65bb697234b38a5f6a747a5492270c14ff8e --- /dev/null +++ b/2024/Responsible Visual Editing/a719b1ec-b6ce-4b23-ac8b-598c2b587aef_model.json @@ -0,0 +1,2555 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.348, + 0.142, + 0.658, + 0.163 + ], + "angle": 0, + "content": "Responsible Visual Editing" + }, + { + "type": "text", + "bbox": [ + 0.266, + 0.189, + 0.737, + 0.207 + ], + "angle": 0, + "content": "Minheng Ni\\(^{1,2}\\), Yeli Shen\\(^{2}\\), Lei Zhang\\(^{1\\boxtimes}\\), and Wangmeng Zuo\\(^{2,3\\boxtimes}\\)" + }, + { + "type": "text", + "bbox": [ + 0.362, + 0.217, + 0.641, + 0.232 + ], + "angle": 0, + "content": "1 The Hong Kong Polytechnic University" + }, + { + "type": "text", + "bbox": [ + 0.392, + 0.232, + 0.612, + 0.246 + ], + "angle": 0, + "content": "\\(^{2}\\) Harbin Institute of Technology" + }, + { + "type": "text", + "bbox": [ + 0.371, + 0.246, + 0.631, + 0.26 + ], + "angle": 0, + "content": "3 Peng Cheng Laboratory, Guangzhou" + }, + { + "type": "text", + "bbox": [ + 0.303, + 0.261, + 0.697, + 0.274 + ], + "angle": 0, + "content": "minheng.ni@connect.polyu.hk, ylshen@stu.hit.edu.cn," + }, + { + "type": "text", + "bbox": [ + 0.331, + 0.275, + 0.673, + 0.288 + ], + "angle": 0, + "content": "cslzhang@comp.polyu.edu.hk, WMzuo@hit.edu.cn" + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.328, + 0.744, + 0.633 + ], + "angle": 0, + "content": "Abstract. With the recent advancements in visual synthesis, there is a growing risk of encountering synthesized images with detrimental effects, such as hate, discrimination, and privacy violations. Unfortunately, it remains unexplored on how to avoid synthesizing harmful images and convert them into responsible ones. In this paper, we present responsible visual editing, which edits risky concepts within an image to more responsible ones with minimal content changes. However, the concepts that need to be edited are often abstract, making them hard to be located and edited. To tackle these challenges, we propose a Cognitive Editor (CoEditor) by harnessing the large multimodal models through a two-stage cognitive process: (1) a perceptual cognitive process to locate what to be edited and (2) a behavioral cognitive process to strategize how to edit. To mitigate the negative implications of harmful images on research, we build a transparent and public dataset, namely AltBear, which expresses harmful information using teddy bears instead of humans. Experiments demonstrate that CoEditor can effectively comprehend abstract concepts in complex scenes, significantly surpassing the baseline models for responsible visual editing. Moreover, we find that the AltBear dataset corresponds well to the harmful content found in real images, providing a safe and effective benchmark for future research. Our source code and dataset can be found at https://github.com/kodenii/Responsible-Visual-Editing." + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.646, + 0.741, + 0.674 + ], + "angle": 0, + "content": "Keywords: Responsible visual editing \\(\\cdot\\) Image editing \\(\\cdot\\) Large multimodal model" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.702, + 0.377, + 0.718 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.734, + 0.788, + 0.841 + ], + "angle": 0, + "content": "With the recent advancements in visual synthesis technologies [22,27,29,36], generating or editing highly realistic images has become possible, and the problem of misusing these technologies is arising [16,21,30] because the generated/edited images may contain harmful contents, such as hate, discrimination and privacy violations. While significant progress has been made in image editing [1,2,4] by using instructions to adjust images, it is less explored on how to edit harmful images into responsible ones." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.352, + 0.127 + ], + "angle": 0, + "content": "M. Ni et al." + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.145, + 0.782, + 0.322 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.333, + 0.788, + 0.404 + ], + "angle": 0, + "content": "Fig. 1: Overview of responsible visual editing. The challenges we encounter in responsible visual editing are multifaceted. Meanwhile, the concepts and objects to be adjusted are often vaguely connected, making it challenging to locate what needs to be modified and plan how to modify it. In this figure, all risky images are sourced from the AltBear dataset, while the edited results are produced by CoEditor." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.439, + 0.788, + 0.53 + ], + "angle": 0, + "content": "We formulate the above mentioned problem as a new task, namely responsible visual editing. As shown in Figure 1, we aim to edit specific concepts in images to make them more responsible while minimizing image changes as much as possible. Given the diversity of risky concepts in images and the different types of risks, we divide this task into three sub-tasks: safety, fairness and privacy, covering a wide range of risks in real-world scenarios." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.533, + 0.788, + 0.669 + ], + "angle": 0, + "content": "Existing editing models often require clear user instructions to make specific adjustments in the images [4,11,34], e.g., editing hat to \"change the blue hat into red\". However, in responsible image editing, the concepts that need to be edited are often abstract, e.g., editing violence to \"make an image look less violent\". This makes it challenging to locate what should to be edited and plan how to edit it. To tackle these challenges, we propose a Cognitive Editor (CoEditor) that harnesses large multimodal models (LMM) through a two-stage cognitive process: (1) a perceptual cognitive process to focus on what needs to be edited and (2) a behavioral cognitive process to strategize how to edit." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.671, + 0.788, + 0.747 + ], + "angle": 0, + "content": "To facilitate the research on responsible image editing, we build a transparent and public dataset, namely AltBear. Unlike general image datasets, the AltBear dataset uses fictional teddy bears as the protagonists, replacing humans in the images to convey risky contents. We significantly reduce the potential ethical risks by using teddy bears and cartoonized images." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Experiments show that CoEditor significantly outperforms baseline models in responsible image editing, validating its effectiveness of comprehending abstract concepts and strategizing the editing process. Furthermore, we find that the AltBear dataset corresponds well to the harmful contents in real images, offering a consistent experimental evaluation. This also validates that AltBear can serve as a safer benchmark for future research." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.55, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Responsible Visual Editing" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.24, + 0.147, + 0.478, + 0.161 + ], + "angle": 0, + "content": "Our contributions are three-fold:" + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.17, + 0.786, + 0.214 + ], + "angle": 0, + "content": "- We propose a new task, responsible visual editing, and introduce a dataset, AltBear, which maintains high consistency with real-world data by using teddy bears as the protagonists to reduce the potential research risks." + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.215, + 0.786, + 0.272 + ], + "angle": 0, + "content": "- We present the CoEditor method for responsible visual editing based on large multimodal models. It includes (1) a perceptual cognitive process to determine what should be edited and (2) a behavioral cognitive process to strategize how to edit." + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.275, + 0.786, + 0.318 + ], + "angle": 0, + "content": "- Comprehensive experiments prove that CoEditor significantly outperforms existing editing models in responsible visual editing. Our findings reveal the potentials of LMM in responsible AI." + }, + { + "type": "list", + "bbox": [ + 0.225, + 0.17, + 0.786, + 0.318 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.34, + 0.388, + 0.356 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.369, + 0.502, + 0.384 + ], + "angle": 0, + "content": "2.1 Responsible Visual Synthesis" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.391, + 0.788, + 0.588 + ], + "angle": 0, + "content": "With the advancement of visual synthesis techniques, how to ensure responsible visual editing has also received increasing attention. Some works have been reported to intercept images that may pose risks [3, 10, 31]. Rombach et al. [29] added a Not-Safe-For-Work (NSFW) filter at the end of the generation stage, while Rando et al. [28] chose to classify and intercept images by projecting them onto the latent space using CLIP. Some works [9, 12, 30] employed machine unlearning methods to forget risky concepts. Zhang et al. [35] and Kumari et al. [17] manipulated the latent variables during generation to avoid producing risky concepts. Recently, Ni et al. [21] proposed to intervene in the generation process using a large language model (LLM) to achieve training-free open-vocabulary responsible visual synthesis. However, current researches on responsible visual synthesis focus on text-to-image generation, while responsible visual editing has not yet been well explored." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.607, + 0.381, + 0.622 + ], + "angle": 0, + "content": "2.2 Image Editing" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.629, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Image editing is a classic task in computer vision. Some previous works used CNN-based methods to edit images [6, 8, 32]. By introducing adversarial techniques, StyleGAN and its variants [1, 14, 15] used GAN inversion for image editing. Using quantized models, MaskGIT [5] explored editing based on conditional vectors, and NUWA-LIP [20] explored natural language guided image editing. With the great success of diffusion models, more and more works attempt to use pre-trained text-to-image diffusion models for editing [2, 7, 19]. Prompt-to-prompt [13] obtained varying images by changing the attention during generation. Recently, InstructPix2pix [4] and InstructDiffusion [11] used the data generated by GPT and prompt-to-prompt methods to train the model to understand natural language instructions. However, current researches on image editing require clear or direct instructions. It has not yet been well explored how to use large multimodal models (LMMs) to understand and adjust the relationship between images and complex abstract instructions." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.352, + 0.127 + ], + "angle": 0, + "content": "M. Ni et al." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.145, + 0.514, + 0.163 + ], + "angle": 0, + "content": "3 Responsible Visual Editing" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.177, + 0.438, + 0.19 + ], + "angle": 0, + "content": "3.1 Problem Formulation" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.202, + 0.787, + 0.277 + ], + "angle": 0, + "content": "The goal of responsible visual editing is to automatically edit a given risky concept \\( c \\) existed in image \\( x_{\\mathrm{r}} \\), generating a responsible image \\( x_{\\mathrm{s}} \\) while making \\( x_{\\mathrm{s}} \\) visually reasonable and changing the image contents as little as possible. To broaden the application of responsible visual editing, we divide it into three subtasks: safety, fairness and privacy." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.278, + 0.787, + 0.322 + ], + "angle": 0, + "content": "Safety. The safety subtask focuses on inappropriate contents for display in real-world scenarios, such as discrimination, terrorist activities, or violence. This task requires completely removing the risky concepts from the image." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.323, + 0.787, + 0.367 + ], + "angle": 0, + "content": "Fairness. This subtask focuses on fairness issues in real-world scenarios, such as biased contents. It diversifies a specific concept in the image without changing much the image content." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.368, + 0.787, + 0.427 + ], + "angle": 0, + "content": "Privacy. The privacy subtask focuses on privacy issues in real-world scenarios, such as real-world characters. This task requires blurring a specific character in the image. While maintaining the basic meaning of the image, the editing should not be recognizable or traceable." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.451, + 0.486, + 0.464 + ], + "angle": 0, + "content": "3.2 AltBear Dataset Collection" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.476, + 0.787, + 0.611 + ], + "angle": 0, + "content": "We collect a number of risky concepts, such as drug abuse, alcohol, racial discrimination, etc., and divide them into three subtasks: safety, fairness, and privacy. For each concept \\( c \\), we use ChatGPT [23] to expand it into 100 image scene descriptions and modify the subject of the description to teddy bears. We manually filter out and refine the compelling descriptions, and randomly use DALL-E 2 [26], Stable Diffusion XL [25], and DALL-E 3 [24] to generate the final risky image \\( x_{\\mathrm{r}} \\). Then, we manually filter the dataset again, selecting the high-quality images. Finally, we obtain 300 groups of images as the test set. For more details, please refer to Supplementary Materials." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.634, + 0.419, + 0.647 + ], + "angle": 0, + "content": "3.3 Evaluation Metrics" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.659, + 0.787, + 0.703 + ], + "angle": 0, + "content": "We propose two metrics to evaluate the performance of responsible image editing: success rate and visual similarity, both of which can be used for automatic machine evaluation and manual human evaluation." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.705, + 0.787, + 0.779 + ], + "angle": 0, + "content": "Success Rate. In the evaluation of success rate, we judge whether the concept \\( c \\) in the responsible image \\( x_{s} \\) still contains risks. For the safety subtask, the concept \\( c \\) should not appear; for the fairness subtask, the diversity of the concept \\( c \\) should be expanded; for the privacy subtask, we require that the identity of specific persons should not be recognized." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.78, + 0.787, + 0.839 + ], + "angle": 0, + "content": "Visual Similarity. In the evaluation of visual similarity, we compute the pixel-level similarity between the responsible image \\( x_{\\mathrm{s}} \\) and the original risky image \\( x_{\\mathrm{r}} \\). Specifically, if \\( x_{\\mathrm{s}} \\) still contains risks, then the similarity is considered as 0. We take the average similarity as the final result." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.55, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Responsible Visual Editing" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.147, + 0.785, + 0.176 + ], + "angle": 0, + "content": "More details of the evaluation metrics can be found in the Supplementary Materials." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.197, + 0.396, + 0.213 + ], + "angle": 0, + "content": "3.4 Special Markers" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.22, + 0.784, + 0.296 + ], + "angle": 0, + "content": "We add unique markers to the images to reduce potential risks in responsible visual editing. As shown in Figure 1, each image contains a marker in the lower right corner, indicating that this image is only used for responsible visual editing research and which subtask it belongs to. In addition, the highlighting symbol A or B represents the original or edited image." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.317, + 0.38, + 0.334 + ], + "angle": 0, + "content": "4 Methodology" + }, + { + "type": "image", + "bbox": [ + 0.224, + 0.371, + 0.782, + 0.507 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.515, + 0.785, + 0.558 + ], + "angle": 0, + "content": "Fig. 2: Overview of CoEditor. CoEditor consists of two stages of cognition: (1) a perceptional cognitive process (PCP) to understand what needs to be edited, and (2) a behavioral cognitive process (BCP) to plan how to edit." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.587, + 0.785, + 0.692 + ], + "angle": 0, + "content": "Compared with traditional visual editing that usually edits specific objects or features in an image, responsible visual editing should be able to edit any risky concepts contained in the image. It could be a theme such as violence, a category such as culture, or a person like Bill Gates. This task is very challenging and it contains two stages: (1) understanding the relationship between the image content and the concept to locate the regions to be edited, and (2) planning how to edit the regions to meet the editing conditions and image rationality." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.692, + 0.785, + 0.768 + ], + "angle": 0, + "content": "With the above considerations, we propose an LMM-based responsible editing model, namely Cognitive Editor (CoEditor). As shown in Figure 2, CoEditor consists of two stages of cognition: (1) perceptional cognitive process (PCP) to locate regions to be edited, and (2) behavioral cognitive process (BCP) to strategize how to edit such regions." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.787, + 0.519, + 0.803 + ], + "angle": 0, + "content": "4.1 Perceptional Cognitive Process" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.785, + 0.84 + ], + "angle": 0, + "content": "Understanding the relationship between the content of an image and abstract concepts is often very difficult, as the concepts may not directly refer to specific" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.352, + 0.128 + ], + "angle": 0, + "content": "M. Ni et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.223 + ], + "angle": 0, + "content": "contents in the image and it requires reasoning based on common sense. Even the powerful LMM may not accurately understand the concepts (see Section 5.2). To mitigate these issues, we use a visual language chain-of-thoughts method based on visual prompts and language prompts for perceptional cognition to locate the regions to be edited." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.223, + 0.788, + 0.282 + ], + "angle": 0, + "content": "For a risky image \\( x_{\\mathrm{r}} \\), we need to visually annotate each element for LMM to refer to them accurately. We perform object extraction to obtain the object sequence \\( \\mathcal{M} \\). In specific, we use Semantic-SAM [18] to extract objects of image \\( x \\) to \\( n \\) masks:" + }, + { + "type": "equation", + "bbox": [ + 0.414, + 0.284, + 0.786, + 0.3 + ], + "angle": 0, + "content": "\\[\n\\mathcal {M} = \\left\\{m _ {1}, m _ {2}, \\dots , m _ {n} \\right\\}. \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.239, + 0.307, + 0.755, + 0.324 + ], + "angle": 0, + "content": "Similar to SoM [33], we visually prompt the objects in the image to \\( v_{\\mathrm{p}} \\):" + }, + { + "type": "equation", + "bbox": [ + 0.451, + 0.336, + 0.786, + 0.352 + ], + "angle": 0, + "content": "\\[\nv _ {\\mathrm {p}} = \\phi (x, \\mathcal {M}), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.363, + 0.789, + 0.408 + ], + "angle": 0, + "content": "where \\(\\phi\\) is a prompting function that adds a visual tag with a number at the corner of each mask's corresponding location in the image, enabling the LMM to refer to it in numeric tags." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.409, + 0.789, + 0.454 + ], + "angle": 0, + "content": "Meanwhile, in order to understand the possible meanings of the concept, we use knowledge extraction to expand its associated concepts and explanations into the text prompts \\( l_{\\mathrm{p}} \\):" + }, + { + "type": "equation", + "bbox": [ + 0.451, + 0.455, + 0.786, + 0.471 + ], + "angle": 0, + "content": "\\[\nl _ {\\mathrm {p}} = f \\left(c; p _ {\\mathrm {k}} ^ {\\text {i n s}}\\right), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.478, + 0.788, + 0.508 + ], + "angle": 0, + "content": "where \\( p_{\\mathrm{k}}^{\\mathrm{ins}} \\) is the instruction for knowledge extraction. Since there is no need to introduce visual information here, \\( f \\) can be either an LLM or an LMM." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.509, + 0.788, + 0.556 + ], + "angle": 0, + "content": "Combining prompts of different modalities \\( v_{\\mathrm{p}} \\) and \\( l_{\\mathrm{p}} \\), the LMM can complete the cognition of the image and the concept, and obtain the mask of the region \\( m_{\\mathrm{p}} \\) to be modified in the image:" + }, + { + "type": "equation", + "bbox": [ + 0.432, + 0.566, + 0.786, + 0.584 + ], + "angle": 0, + "content": "\\[\nm _ {\\mathrm {p}} = f \\left(v _ {\\mathrm {p}}, l _ {\\mathrm {p}}; p _ {\\mathrm {p}} ^ {\\text {i n s}}\\right), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.596, + 0.788, + 0.643 + ], + "angle": 0, + "content": "where \\( p_{\\mathrm{p}}^{\\mathrm{ins}} \\) is the instruction for focus generation and \\( f \\) is an LMM. Though we may obtain multiple related regions, for simplicity, we combine all regions into one single \\( m_{\\mathrm{p}} \\) as the result of PCP." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.665, + 0.505, + 0.681 + ], + "angle": 0, + "content": "4.2 Behavioral Cognitive Process" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.692, + 0.788, + 0.767 + ], + "angle": 0, + "content": "In the perception cognitive process, the region \\( m_p \\) to be edited has been identified, and we need to figure out how to edit it. Since the concept may be abstract, the editing model is hard to directly perform editing operations (see Section 5.2). Therefore, we employ the visual language chain-of-thoughts to plan the editing target, i.e., what contents are expected after responsible editing." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.768, + 0.788, + 0.815 + ], + "angle": 0, + "content": "We expand the object \\( m_p \\) identified in the previous stage to \\( m_p' \\) through object extension by enlarging the contour to include surrounding information. Then, we crop the image based on the mask \\( m_p' \\) to obtain visual prompting:" + }, + { + "type": "equation", + "bbox": [ + 0.452, + 0.825, + 0.786, + 0.843 + ], + "angle": 0, + "content": "\\[\nv _ {\\mathrm {b}} = x _ {\\mathrm {r}} \\otimes m _ {\\mathrm {p}} ^ {\\prime}. \\tag {5}\n\\]" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.55, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Responsible Visual Editing" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.207 + ], + "angle": 0, + "content": "At the same time, we perform instruction implementation to get the full language prompt \\( l_{\\mathrm{b}} \\) by concatenating the instruction of the task \\( p_{\\mathrm{t}}^{\\mathrm{ins}} \\) and the input concept \\( c \\). Then we use the language prompt \\( l_{\\mathrm{b}} \\) together with the visual prompt \\( v_{\\mathrm{b}} \\) to generate the editing target \\( r_{\\mathrm{b}} \\) for the inpainting model:" + }, + { + "type": "equation", + "bbox": [ + 0.436, + 0.222, + 0.786, + 0.238 + ], + "angle": 0, + "content": "\\[\nr _ {\\mathrm {b}} = f \\left(v _ {\\mathrm {b}}, l _ {\\mathrm {b}}; p _ {\\mathrm {b}} ^ {\\text {i n s}}\\right), \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.249, + 0.503, + 0.265 + ], + "angle": 0, + "content": "where \\(p_{\\mathrm{b}}^{\\mathrm{ins}}\\) is the instruction for editing." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.265, + 0.784, + 0.309 + ], + "angle": 0, + "content": "Finally, we take the original image \\( x_{\\mathrm{r}} \\), the region to be edited \\( m_p' \\), and the editing target \\( r_{\\mathrm{b}} \\) as the input to the inpainting model, and obtain the final editing result as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.438, + 0.311, + 0.786, + 0.328 + ], + "angle": 0, + "content": "\\[\nx _ {\\mathrm {s}} = g \\left(x _ {\\mathrm {r}}, m _ {\\mathrm {p}} ^ {\\prime}; r _ {\\mathrm {b}}\\right), \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.334, + 0.766, + 0.35 + ], + "angle": 0, + "content": "where \\( g \\) is the inpainting model and \\( x_{\\mathrm{s}} \\) is the final responsible editing result." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.374, + 0.457, + 0.388 + ], + "angle": 0, + "content": "4.3 Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.4, + 0.785, + 0.522 + ], + "angle": 0, + "content": "Our method is training-free. We choose GPT-4V as the LMM and Semantic-SAM as the object extraction network with a granularity of 1.5 and its officially released checkpoint4. For the inpainting model, we choose Stable Diffusion Inpainting trained based on Stable Diffusion v2 and its publicly available checkpoint5. Like most previous works, we adjust the input image to a size of \\(512 \\times 512\\). All random seeds are fixed to 42. In all experiments, the instructions \\(p_{\\mathrm{k}}^{\\mathrm{ins}}\\), \\(p_{\\mathrm{p}}^{\\mathrm{ins}}\\), and \\(p_{\\mathrm{b}}^{\\mathrm{ins}}\\) are fixed, and \\(p_{\\mathrm{t}}^{\\mathrm{ins}}\\) depends on the task. Please refer to the Supplementary Materials for more details." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.548, + 0.367, + 0.565 + ], + "angle": 0, + "content": "5 Experiment" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.581, + 0.414, + 0.597 + ], + "angle": 0, + "content": "5.1 Experiment Setup" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.608, + 0.785, + 0.698 + ], + "angle": 0, + "content": "Since the concept of responsible visual editing is arbitrary and it does not specify the editing area, we select two powerful image editing models, InstructPix2pix [4] and InstructDiffusion [11], in the experiments. They not only allow arbitrary language guidance with the help of LLM but also require no additional masks to indicate the editing region. In addition, we could use the same Stable Diffusion base model in our CoEditor, making our comparison fair." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.699, + 0.785, + 0.803 + ], + "angle": 0, + "content": "We conduct experiments on the AltBear dataset. Since InstructPix2pix and InstructDiffusion do not support responsible editing of images based on the concept, we manually design editing conditions remove {concept} for safety and privacy tasks, and increase the variety of {concept} for fairness tasks. We evaluate the edited images in terms of both the success rate and visual similarity. We measure the results comprehensively via both automatic machine and manual human evaluations." + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.811, + 0.564, + 0.826 + ], + "angle": 0, + "content": "4 https://github.com/UX-Decoder/Semantic-SAM" + }, + { + "type": "page_footnote", + "bbox": [ + 0.222, + 0.827, + 0.732, + 0.841 + ], + "angle": 0, + "content": "5 https://huggingface.co/stabilityai/stable-diffusion-2-inpainting" + }, + { + "type": "list", + "bbox": [ + 0.218, + 0.811, + 0.732, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.352, + 0.128 + ], + "angle": 0, + "content": "M. Ni et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.788, + 0.187 + ], + "angle": 0, + "content": "Table 1: Overall results on AltBear under machine evaluation. We can find that CoEditor significantly outperforms the baseline models in both success rate and visual similarity, validating the effectiveness of CoEditor." + }, + { + "type": "table", + "bbox": [ + 0.218, + 0.199, + 0.785, + 0.294 + ], + "angle": 0, + "content": "
ModelSafetyFairnessPrivacyOverall
\\( Succ^{\\uparrow} \\)\\( Sim^{\\uparrow} \\)\\( Succ^{\\uparrow} \\)\\( Sim^{\\uparrow} \\)\\( Succ^{\\uparrow} \\)\\( Sim^{\\uparrow} \\)\\( Succ^{\\uparrow} \\)\\( Sim^{\\uparrow} \\)
InstructPix2pix21.31%0.196922.11%0.212351.22%0.456231.93%0.2924
InstructDiffusion40.54%0.357421.43%0.171966.33%0.496344.14%0.3497
CoEditor (Ours)70.13%0.565248.96%0.359578.35%0.639565.56%0.5188
" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.306, + 0.787, + 0.348 + ], + "angle": 0, + "content": "Table 2: Overall results on AltBear under human evaluation. In human evaluations, CoEditor shows consistent results with machine evaluations, showing that CoEditor is also effective from a human subjective perspective." + }, + { + "type": "table", + "bbox": [ + 0.218, + 0.359, + 0.785, + 0.453 + ], + "angle": 0, + "content": "
ModelSafetyFairnessPrivacyOverall
\\( Succ^{\\uparrow} \\)\\( Sim^{\\uparrow} \\)\\( Succ^{\\uparrow} \\)\\( Sim^{\\uparrow} \\)\\( Succ^{\\uparrow} \\)\\( Sim^{\\uparrow} \\)\\( Succ^{\\uparrow} \\)\\( Sim^{\\uparrow} \\)
InstructPix2pix28.41%0.25239.38%0.083940.00%0.327126.06%0.2217
InstructDiffusion40.91%0.307819.79%0.136956.00%0.345739.08%0.2634
CoEditor (Ours)65.91%0.531850.00%0.360665.00%0.518360.21%0.4692
" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.482, + 0.39, + 0.495 + ], + "angle": 0, + "content": "5.2 Overall Results" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.508, + 0.787, + 0.658 + ], + "angle": 0, + "content": "Quantitative Results. As shown in Table 1 and Table 2, CoEditor shows significant advantages in both machine and human evaluations. We can see that for the success rate, CoEditor has more than \\(20\\%\\) increase in almost all subtasks. This proves that for responsible visual synthesis, using LMM to understand images and concepts and to strategize editing process is crucial. In addition, we find that CoEditor has higher visual similarity compared to traditional models, which reveals the importance of explicit cognition process of editing. It is noticed that the machine evaluation shows similar results to human evaluations, further proving that our editing results are of the same high quality from the perspective of human judgements." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.659, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Qualitative Results. As shown in Figure 3, CoEditor demonstrates strong understanding and editing capabilities in responsible visual synthesis. In the first row of the safety task, we see that the CoEditor not only fully erases the concepts of drugs and alcohol but also maintains the rationality of the picture. However, InstructDiffusion fails to completely remove drugs and generates unreasonable contents in the picture. Moreover, InstructPix2pix does not follow the instruction well. In the second row of the fairness task, CoEditor successfully diversifies the specified concept. It maintains the original layout, while InstructDiffusion results in blurry or damaged outputs and InstructPix2pix has difficulties in executing this instruction. In the third row of the privacy task, the CoEditor blurs the characters' features while maintaining the original meaning with high visual quality. However, InstructDiffusion and InstructPix2pix fail in content removal" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.55, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Responsible Visual Editing" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "table_caption", + "bbox": [ + 0.215, + 0.145, + 0.785, + 0.187 + ], + "angle": 0, + "content": "Table 3: Ablation study of AltBear. Both PCP and BCP significantly help CoEditor improve its editing capabilities, which is reflected in the significant increase in success rate and visual similarity." + }, + { + "type": "table", + "bbox": [ + 0.377, + 0.199, + 0.626, + 0.272 + ], + "angle": 0, + "content": "
ModelSucc↑Sim↑
CoEditor w/o PCP44.19%0.3855
CoEditor w/o BCP28.02%0.2221
CoEditor (Full)65.56%0.5188
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.298, + 0.784, + 0.343 + ], + "angle": 0, + "content": "or damage the picture. These examples show that CoEditor can understand the images and concepts and then plan the editing well with the help of LMM. More examples can be found in the Supplementary Materials." + }, + { + "type": "image", + "bbox": [ + 0.226, + 0.368, + 0.498, + 0.566 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.506, + 0.368, + 0.78, + 0.567 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.579, + 0.785, + 0.635 + ], + "angle": 0, + "content": "Fig. 3: Visualization results on AltBear. CoEditor performs well in all subtasks and maintains high visual similarity. The results of CoEditor also have better rationality. InstructDiffusion often over-edits images or produces unreasonable visual effects, while the editing ability of InstructPix2pix is weaker compared with CoEditor." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.683, + 0.389, + 0.699 + ], + "angle": 0, + "content": "5.3 Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.705, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Quantitative Results. We conduct ablation experiments on AltBear in Table 3 to explore the effectiveness of the two stages of CoEditor. In the ablation of PCP, since the model cannot locate the area to be edited, we turn to use the largest area in the image. In the ablation of BCP, we use the same instruction as the baseline model for editing. To ensure fairness, all other parameters of the experiments are consistent. As shown in Table 3, both stages of CoEditor are significant for the final results. Without PCP, the CoEditor cannot understand the object to be edited, causing a massive drop in success rate and visual similarity. Without BCP, although the CoEditor can locate the region to be edited," + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.352, + 0.127 + ], + "angle": 0, + "content": "M. Ni et al." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.147, + 0.785, + 0.178 + ], + "angle": 0, + "content": "simply asking the model to remove or increase diversity will make it difficult to understand what to edit, ultimately leading to editing failures." + }, + { + "type": "image", + "bbox": [ + 0.224, + 0.201, + 0.782, + 0.404 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.416, + 0.788, + 0.473 + ], + "angle": 0, + "content": "Fig. 4: Roles of different components in CoEditor. Without the help of BCP, the CoEditor produces inconsistent results with the concept or visually unreasonable results because it does not know how to edit the content correctly. Without PCP, CoEditor is unable to locate the editing regions, resulting in editing failures." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.5, + 0.789, + 0.697 + ], + "angle": 0, + "content": "Qualitative Results. To investigate whether the two stages of CoEditor, BCP and PCP, can produce the expected outputs, we visualize some examples in Figure 3. One can see that the BCP stage is crucial for correct editing. In most examples, without the BCP stage, the editing result will fail to meet the requirements, or there will be visually unreasonable contents. BCP can effectively convert abstract targets into specific content that the inpainting model can follow. This shows the importance of editing planning. At the same time, PCP plays a decisive role. Without the participation of PCP, most of the edits will fail. This is because most of the concepts to be edited do not directly associate with a specific object in the image. Therefore, a deep understanding of image and concepts using PCP is crucial for CoEditor to edit correctly. The CoEditor can achieve satisfactory effects when both the two modules are used. Extra ablation studies can be found in the Supplementary Materials." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.714, + 0.533, + 0.73 + ], + "angle": 0, + "content": "5.4 Exploration of Cognitive Process" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.788, + 0.841 + ], + "angle": 0, + "content": "To explore why CoEditor can successfully perform complex and responsible visual editing, we show the intermediate results of various cognitive processes in Figure 5. We can see that PCP can successfully locate regions to be edited, regardless of whether this region is directly related to the concept. Based on the task type, PCP can also accurately find out the region to be edited among multiple identical objects. At the same time, BCP can successfully plan the editing. No matter what the concept is, BCP can accurately associate it with the" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.55, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Responsible Visual Editing" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "image_caption", + "bbox": [ + 0.238, + 0.146, + 0.314, + 0.158 + ], + "angle": 0, + "content": "Risky Image" + }, + { + "type": "image", + "bbox": [ + 0.225, + 0.16, + 0.324, + 0.236 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.368, + 0.146, + 0.439, + 0.157 + ], + "angle": 0, + "content": "PCP Result" + }, + { + "type": "image", + "bbox": [ + 0.353, + 0.159, + 0.454, + 0.236 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.528, + 0.146, + 0.599, + 0.157 + ], + "angle": 0, + "content": "BCP Result" + }, + { + "type": "text", + "bbox": [ + 0.488, + 0.167, + 0.655, + 0.229 + ], + "angle": 0, + "content": "A cuddly teddy bear enjoying a sunny day on a serene picnic with a picturesque house fully intact in the background, surrounded by vibrant greenery." + }, + { + "type": "image_caption", + "bbox": [ + 0.692, + 0.146, + 0.767, + 0.157 + ], + "angle": 0, + "content": "Final Result" + }, + { + "type": "image", + "bbox": [ + 0.68, + 0.159, + 0.78, + 0.236 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.228, + 0.24, + 0.318, + 0.252 + ], + "angle": 0, + "content": "Concept: arsion" + }, + { + "type": "image", + "bbox": [ + 0.225, + 0.253, + 0.324, + 0.33 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.353, + 0.253, + 0.453, + 0.33 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.488, + 0.249, + 0.655, + 0.333 + ], + "angle": 0, + "content": "A teddy bear wearing traditional attire representing different cultures, such as a Scottish kilt, an Indian sari, and a Nigerian abada, showcasing the beautiful variety of cultural dress from around the world." + }, + { + "type": "image", + "bbox": [ + 0.68, + 0.253, + 0.78, + 0.33 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.228, + 0.334, + 0.328, + 0.346 + ], + "angle": 0, + "content": "Concept: culture" + }, + { + "type": "image", + "bbox": [ + 0.225, + 0.347, + 0.324, + 0.423 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.353, + 0.347, + 0.453, + 0.423 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.487, + 0.355, + 0.655, + 0.415 + ], + "angle": 0, + "content": "A fluffy teddy bear with sparkling eyes and shiny footwear stands in front of a vintage microphone on a soft glowing background." + }, + { + "type": "image", + "bbox": [ + 0.68, + 0.347, + 0.78, + 0.423 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.228, + 0.427, + 0.374, + 0.439 + ], + "angle": 0, + "content": "Concept: Marilyn Monroe" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.451, + 0.788, + 0.522 + ], + "angle": 0, + "content": "Fig. 5: Intermediate results of cognitive process in CoEditor. We can see that PCP can successfully find one or more objects to be edited, even if the object is not directly related to the concept. BCP can generate effective editing target based on the positioning of PCP, even if the scenario is very complex. The effectiveness of the two cognitive processes allows CoEditor to edit images successfully." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.554, + 0.788, + 0.614 + ], + "angle": 0, + "content": "content in the image to obtain editing target consistent with the image scene. This shows the importance of cognitive processes to CoEditor in concept understanding and editing planning. For further explorations of CoEditor, please see the Supplementary Materials." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.643, + 0.747, + 0.658 + ], + "angle": 0, + "content": "5.5 Performance Consistency on AltBear and Real-world Data" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.789, + 0.841 + ], + "angle": 0, + "content": "To verify whether AltBear can serve as a replacement of real-world images in subsequent research, we examine the performance consistency between AltBear and real-world images. We build a dataset that has similar size to AltBear, consisting of real-world images collected from the Internet. We conduct experiments on this dataset with all hyper-parameters and models identical to that on AltBear. As shown in Table 4 and 5, CoEditor achieves similar performance to that on AltBear, demonstrating CoEditor's robustness to both synthesized data and real-world data. The baseline models, InstructPix2pix and InstructDiffusion, also show similar performance on AltBear and real-world data. Since the protagonist is not a human but a fictional teddy bear, AltBear significantly reduces the risk of image propagation and display, effectively avoiding ethical risks." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.352, + 0.127 + ], + "angle": 0, + "content": "M. Ni et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.145, + 0.788, + 0.203 + ], + "angle": 0, + "content": "Table 4: Results on real-world images under machine evaluation. On real-world datasets, CoEditor shows similar performance to that on AltBear, significantly surpassing the baseline model. Furthermore, we find that the distributions of both success rate and visual similarity are close for real-world images and AltBear." + }, + { + "type": "table", + "bbox": [ + 0.218, + 0.212, + 0.785, + 0.307 + ], + "angle": 0, + "content": "
ModelSafetyFairnessPrivacyOverall
\\( Succ^{\\uparrow} \\)\\( Sim^{\\uparrow} \\)\\( Succ^{\\uparrow} \\)\\( Sim^{\\uparrow} \\)\\( Succ^{\\uparrow} \\)\\( Sim^{\\uparrow} \\)\\( Succ^{\\uparrow} \\)\\( Sim^{\\uparrow} \\)
InstructPix2pix14.00%0.12831.89%0.016622.86%0.216411.59%0.1078
InstructDiffusion38.60%0.318514.29%0.103055.81%0.374435.57%0.2638
CoEditor (Ours)59.46%0.474058.18%0.425363.41%0.513960.00%0.4679
" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.319, + 0.788, + 0.375 + ], + "angle": 0, + "content": "Table 5: Results of real-world images under human evaluation. CoEditor achieves significantly better results than the baseline model in human evaluation. In addition, the quantitative results on AltBear and real-world images in human evaluation are also similar." + }, + { + "type": "table", + "bbox": [ + 0.218, + 0.387, + 0.785, + 0.481 + ], + "angle": 0, + "content": "
ModelSafetyFairnessPrivacyOverall
\\( Succ^{\\uparrow} \\)\\( Sim^{\\uparrow} \\)\\( Succ^{\\uparrow} \\)\\( Sim^{\\uparrow} \\)\\( Succ^{\\uparrow} \\)\\( Sim^{\\uparrow} \\)\\( Succ^{\\uparrow} \\)\\( Sim^{\\uparrow} \\)
InstructPix2pix31.25%0.27989.09%0.080415.56%0.132720.56%0.1821
InstructDiffusion47.50%0.376421.82%0.152453.33%0.337841.11%0.2983
CoEditor (Ours)68.75%0.548665.45%0.478780.00%0.646770.56%0.5518
" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.509, + 0.421, + 0.523 + ], + "angle": 0, + "content": "5.6 Real-world Results" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.535, + 0.789, + 0.747 + ], + "angle": 0, + "content": "To better demonstrate the robustness of CoEditor, we show its performance on real-world images and compare with baseline models in Figure 6. For the safety subtasks, the CoEditor effectively removes risky factors from the images and rationalizes them. In the first row, the robber in the first image is turned into an enthusiastic saleswoman, and the craving for drugs in the second image is turned into a craving for food. This also shows the imagination capability of CoEditor. For the fairness subtasks in the second row, the CoEditor completes the image editing while maintaining the visual quality of the image. In the privacy subtasks of the third row, the CoEditor appropriately blurs features related to the person while maintaining the overall content. Similar to the results on AltBear, the baseline models, InstructPix2pix and InstructDiffusion, struggle to maintain visual quality or rationality in some examples, even though they apply editing. In other cases, they produce damaged figures. More real-world examples can be found in the Supplementary Materials." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.769, + 0.567, + 0.785 + ], + "angle": 0, + "content": "5.7 Analysis of Computational Overhead" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.787, + 0.841 + ], + "angle": 0, + "content": "In order to explore how CoEditor consumes resources, we make comparisons on time and memory usage. We use the same full-precision and optimization settings, e.g., memory-efficient attention for all models." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.55, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Responsible Visual Editing" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "image", + "bbox": [ + 0.225, + 0.145, + 0.498, + 0.21 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.224, + 0.211, + 0.498, + 0.272 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.224, + 0.272, + 0.498, + 0.344 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.145, + 0.78, + 0.209 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.209, + 0.78, + 0.272 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.272, + 0.78, + 0.344 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.356, + 0.785, + 0.412 + ], + "angle": 0, + "content": "Fig. 6: Visualized results on real-world data. Even in real-world scenarios, CoEditor can still work well. The edited images meet the requirements and maintain excellent visual similarity and reasonableness. Similar to the results on AltBear, InstructDiffusion and InstructPix2pix fail to output well edited images." + }, + { + "type": "table_caption", + "bbox": [ + 0.349, + 0.426, + 0.653, + 0.44 + ], + "angle": 0, + "content": "Table 6: Comparison on inference time." + }, + { + "type": "table", + "bbox": [ + 0.372, + 0.453, + 0.632, + 0.553 + ], + "angle": 0, + "content": "
ModelTimeMultiple
InstrcutPix2pix12.43s1×
InstructDiffusion20.75s1.67×
CoEditor14.10s1.13×
PCP Stage4.65s-
BCP Stage9.44s-
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.58, + 0.787, + 0.64 + ], + "angle": 0, + "content": "Time. We calculate the average inference time of different models. For CoEditor, we also calculate the network latency. As shown in Table 6, the speed of CoEditor is even faster than InstructDiffusion [11], and is comparable to Instruct Pix2pix [4]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.641, + 0.785, + 0.702 + ], + "angle": 0, + "content": "GPU Memory. We calculate the peak VRAM usage of different models. As shown in Table 7, CoEditor consumes the least memory. This is because our chosen models, i.e., Semantic-SAM [18] and Stable Diffusion Inpainting [29], are very lightweight, which allows our model to work with minimal memory." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.724, + 0.42, + 0.74 + ], + "angle": 0, + "content": "6 Ethics Statement" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.755, + 0.688, + 0.771 + ], + "angle": 0, + "content": "6.1 Inappropriate Content, Privacy, and Discrimination" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.787, + 0.84 + ], + "angle": 0, + "content": "The AltBear dataset uses teddy bears as the protagonist, significantly reducing the risks caused by the public display of inappropriate contents. We manually reviewed all contents to ensure that privacy or discrimination is avoided. We release AltBear under the MIT license to keep transparency." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.352, + 0.127 + ], + "angle": 0, + "content": "M. Ni et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.351, + 0.145, + 0.653, + 0.16 + ], + "angle": 0, + "content": "Table 7: Comparison on GPU memory." + }, + { + "type": "table", + "bbox": [ + 0.359, + 0.171, + 0.646, + 0.244 + ], + "angle": 0, + "content": "
ModelVRAMMultiple
InstrcutPix2pix18,118MB1×
InstructDiffusion8,378MB0.46×
CoEditor7,328MB0.40×
" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.272, + 0.391, + 0.289 + ], + "angle": 0, + "content": "6.2 Reproducibility" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.298, + 0.787, + 0.375 + ], + "angle": 0, + "content": "We build our method on top of the Stable Diffusion Inpainting model with publicly accessible code and checkpoints with a fixed random seed. We notice that the GPT API cannot guarantee identical responses so that we provide all instruction prompts for reference. To further improve the reproducibility, we also release the dataset, code and our trained model." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.397, + 0.362, + 0.412 + ], + "angle": 0, + "content": "6.3 Anti-misuse" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.422, + 0.787, + 0.497 + ], + "angle": 0, + "content": "In order to ensure that the dataset is not misused, we have designed unique markers (refer to Section 3.4). We added the marker to the bottom right corner of the image to indicate that the image is used for responsible visual editing research only. Specifically, we have defined various tags to indicate whether the image has been edited or contains potential risks." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.498, + 0.787, + 0.559 + ], + "angle": 0, + "content": "In addition, the interaction system of our model always maintains visibility of the concepts to be edited to prevent misuse of the model. We also call on subsequent work to maintain the visibility of concepts like ours, ensuring a beneficial role for the community through open and transparent methods." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.581, + 0.36, + 0.598 + ], + "angle": 0, + "content": "7 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.614, + 0.788, + 0.841 + ], + "angle": 0, + "content": "In this work, we proposed a new task, namely responsible visual editing, which entails editing specific concepts within an image to render it more responsible while minimizing changes. To tackle abstract concepts in responsible visual editing, we proposed a Cognitive Editor (CoEditor) by harnessing a large multimodal model (LMM) through a two-stage process: a perceptual cognitive process to locate what needs to be edited, and a behavioral cognitive process to strategize how to edit. We constructed a transparent and public dataset, AltBear, which represents harmful information using teddy bears instead of humans, thereby mitigating the negative implications on research. Experiments demonstrated that CoEditor can effectively comprehend abstract concepts within complex scenes and significantly surpass the performance of baseline models for responsible visual editing. Moreover, we validated that the AltBear dataset can well serve as a replacement of real-world images in the research of responsible visual editing, providing a safer benchmark. Our findings also revealed the potential of LMM in responsible AI." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.55, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Responsible Visual Editing" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.393, + 0.163 + ], + "angle": 0, + "content": "Acknowledgement" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.174, + 0.8, + 0.22 + ], + "angle": 0, + "content": "This work was supported in part by the National Key R&D Program of China (2022YFA1004100), and the National Natural Science Foundation of China (NSFC) under grant No. 62441202." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.241, + 0.323, + 0.256 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.268, + 0.786, + 0.31 + ], + "angle": 0, + "content": "1. Abdal, R., Qin, Y., Wonka, P.: Image2stylegan++: How to edit the embedded images? In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 8296-8305 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.31, + 0.786, + 0.351 + ], + "angle": 0, + "content": "2. Avrahami, O., Lischinski, D., Fried, O.: Blended diffusion for text-driven editing of natural images. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18208-18218 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.351, + 0.786, + 0.392 + ], + "angle": 0, + "content": "3. Brack, M., Friedrich, F., Hintersdorf, D., Struppek, L., Schramowski, P., Kersting, K.: Sega: Instructing text-to-image models using semantic guidance. Advances in Neural Information Processing Systems 36 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.392, + 0.786, + 0.433 + ], + "angle": 0, + "content": "4. Brooks, T., Holynski, A., Efros, A.A.: Instructpix2pix: Learning to follow image editing instructions. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18392-18402 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.433, + 0.786, + 0.473 + ], + "angle": 0, + "content": "5. Chang, H., Zhang, H., Jiang, L., Liu, C., Freeman, W.T.: Maskgit: Masked generative image transformer. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11315-11325 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.473, + 0.786, + 0.514 + ], + "angle": 0, + "content": "6. Chen, J., Shen, Y., Gao, J., Liu, J., Liu, X.: Language-based image editing with recurrent attentive models. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 8721-8729 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.514, + 0.786, + 0.542 + ], + "angle": 0, + "content": "7. Couairon, G., Verbeek, J., Schwenk, H., Cord, M.: Diffedit: Diffusion-based semantic image editing with mask guidance. arXiv preprint arXiv:2210.11427 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.542, + 0.786, + 0.582 + ], + "angle": 0, + "content": "8. Dong, H., Yu, S., Wu, C., Guo, Y.: Semantic image synthesis via adversarial learning. In: Proceedings of the IEEE international conference on computer vision. pp. 5706-5714 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.582, + 0.786, + 0.609 + ], + "angle": 0, + "content": "9. Gandikota, R., Materzynska, J., Fiotto-Kaufman, J., Bau, D.: Erasing concepts from diffusion models. arXiv preprint arXiv:2303.07345 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.609, + 0.786, + 0.65 + ], + "angle": 0, + "content": "0. Gandikota, R., Orgad, H., Belinkov, Y., Materzyńska, J., Bau, D.: Unified concept editing in diffusion models. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 5111-5120 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.65, + 0.786, + 0.691 + ], + "angle": 0, + "content": "1. Geng, Z., Yang, B., Hang, T., Li, C., Gu, S., Zhang, T., Bao, J., Zhang, Z., Hu, H., Chen, D., et al.: Instructdiffusion: A generalist modeling interface for vision tasks. arXiv preprint arXiv:2309.03895 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.691, + 0.786, + 0.718 + ], + "angle": 0, + "content": "2. Heng, A., Soh, H.: Continual learning for forgetting in deep generative models (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.718, + 0.786, + 0.759 + ], + "angle": 0, + "content": "3. Hertz, A., Mokady, R., Tenenbaum, J., Aberman, K., Pritch, Y., Cohen-Or, D.: Prompt-to-prompt image editing with cross attention control. arXiv preprint arXiv:2208.01626 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.759, + 0.786, + 0.8 + ], + "angle": 0, + "content": "4. Karras, T., Laine, S., Aila, T.: A style-based generator architecture for generative adversarial networks. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 4401-4410 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.8, + 0.786, + 0.84 + ], + "angle": 0, + "content": "5. Karras, T., Laine, S., Aittala, M., Hellsten, J., Lehtinen, J., Aila, T.: Analyzing and improving the image quality of stylegan. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 8110-8119 (2020)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.268, + 0.786, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.352, + 0.127 + ], + "angle": 0, + "content": "M. Ni et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.787, + 0.189 + ], + "angle": 0, + "content": "16. Kumari, N., Zhang, B., Wang, S.Y., Shechtman, E., Zhang, R., Zhu, J.Y.: Ablating Concepts in Text-to-Image Diffusion Models (May 2023), http://arxiv.org/abs/2303.13516, arXiv:2303.13516 [cs]" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.189, + 0.787, + 0.217 + ], + "angle": 0, + "content": "17. Kumari, N., Zhang, B., Wang, S.Y., Shechtman, E., Zhang, R., Zhu, J.Y.: Ablating concepts in text-to-image diffusion models. arXiv preprint arXiv:2303.13516 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.217, + 0.787, + 0.257 + ], + "angle": 0, + "content": "18. Li, F., Zhang, H., Sun, P., Zou, X., Liu, S., Yang, J., Li, C., Zhang, L., Gao, J.: Semantic-sam: Segment and recognize anything at any granularity. arXiv preprint arXiv:2307.04767 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.257, + 0.787, + 0.298 + ], + "angle": 0, + "content": "19. Meng, C., He, Y., Song, Y., Song, J., Wu, J., Zhu, J.Y., Ermon, S.: Sdedit: Guided image synthesis and editing with stochastic differential equations. arXiv preprint arXiv:2108.01073 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.298, + 0.787, + 0.338 + ], + "angle": 0, + "content": "20. Ni, M., Li, X., Zuo, W.: Nuwa-lip: Language-guided image inpainting with defect-free vqgan. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14183-14192 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.338, + 0.787, + 0.366 + ], + "angle": 0, + "content": "21. Ni, M., Wu, C., Wang, X., Yin, S., Wang, L., Liu, Z., Duan, N.: Ores: Open-vocabulary responsible visual synthesis. arXiv preprint arXiv:2308.13785 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.366, + 0.787, + 0.407 + ], + "angle": 0, + "content": "22. Nichol, A., Dhariwal, P., Ramesh, A., Shyam, P., Mishkin, P., McGrew, B., Sutskever, I., Chen, M.: Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.407, + 0.597, + 0.42 + ], + "angle": 0, + "content": "23. OpenAI: Chatgpt. https://chat.openai.com (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.42, + 0.626, + 0.433 + ], + "angle": 0, + "content": "24. OpenAI: Dall-e 3. https://openai.com/dall-e-3 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.433, + 0.787, + 0.473 + ], + "angle": 0, + "content": "25. Podell, D., English, Z., Lacey, K., Blattmann, A., Dockhorn, T., Müller, J., Penna, J., Rombach, R.: Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.473, + 0.787, + 0.514 + ], + "angle": 0, + "content": "26. Ramesh, A., Dhariwal, P., Nichol, A., Chu, C., Chen, M.: Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125 1(2), 3 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.514, + 0.787, + 0.555 + ], + "angle": 0, + "content": "27. Ramesh, A., Pavlov, M., Goh, G., Gray, S., Voss, C., Radford, A., Chen, M., Sutskever, I.: Zero-shot text-to-image generation. In: International Conference on Machine Learning. pp. 8821-8831. PMLR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.555, + 0.787, + 0.582 + ], + "angle": 0, + "content": "28. Rando, J., Paleka, D., Lindner, D., Heim, L., Tramér, F.: Red-teaming the stable diffusion safety filter. arXiv preprint arXiv:2210.04610 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.582, + 0.787, + 0.623 + ], + "angle": 0, + "content": "29. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10684-10695 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.623, + 0.787, + 0.677 + ], + "angle": 0, + "content": "30. Schramowski, P., Brack, M., Deiseroth, B., Kersting, K.: Safe latent diffusion: Mitigating inappropriate degeneration in diffusion models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 22522-22531 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.677, + 0.787, + 0.731 + ], + "angle": 0, + "content": "31. Schuhmann, C., Beaumont, R., Vencu, R., Gordon, C., Wightman, R., Cherti, M., Coombes, T., Katta, A., Mullis, C., Wortsman, M., et al.: Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems 35, 25278-25294 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.731, + 0.787, + 0.772 + ], + "angle": 0, + "content": "32. Wang, S.Y., Wang, O., Zhang, R., Owens, A., Efros, A.A.: Cnn-generated images are surprisingly easy to spot... for now. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 8695-8704 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.772, + 0.787, + 0.813 + ], + "angle": 0, + "content": "33. Yang, J., Zhang, H., Li, F., Zou, X., Li, C., Gao, J.: Set-of-mark prompting unleashes extraordinary visual grounding in gpt-4v. arXiv preprint arXiv:2310.11441 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.813, + 0.787, + 0.841 + ], + "angle": 0, + "content": "34. Yu, T., Feng, R., Feng, R., Liu, J., Jin, X., Zeng, W., Chen, Z.: Inpaint anything: Segment anything meets image inpainting. arXiv preprint arXiv:2304.06790 (2023)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.55, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Responsible Visual Editing" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.786, + 0.175 + ], + "angle": 0, + "content": "35. Zhang, E., Wang, K., Xu, X., Wang, Z., Shi, H.: Forget-me-not: Learning to forget in text-to-image diffusion models. arXiv preprint arXiv:2303.17591 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.176, + 0.787, + 0.217 + ], + "angle": 0, + "content": "36. Zhang, Y., Wei, Y., Jiang, D., Zhang, X., Zuo, W., Tian, Q.: Control video: Training-free controllable text-to-video generation. arXiv preprint arXiv:2305.13077 (2023)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.217 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/Responsible Visual Editing/a719b1ec-b6ce-4b23-ac8b-598c2b587aef_origin.pdf b/2024/Responsible Visual Editing/a719b1ec-b6ce-4b23-ac8b-598c2b587aef_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..bb373fff9a20f3104ea50f2c1892c5ddd0621907 --- /dev/null +++ b/2024/Responsible Visual Editing/a719b1ec-b6ce-4b23-ac8b-598c2b587aef_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f9103d2176ffd530bf4df3c172a33b23ccb11749e2b8e173504a91f6c54b42c6 +size 7265307 diff --git a/2024/Responsible Visual Editing/full.md b/2024/Responsible Visual Editing/full.md new file mode 100644 index 0000000000000000000000000000000000000000..220a9667ddf1f5ada81a8381686ccad68331dd77 --- /dev/null +++ b/2024/Responsible Visual Editing/full.md @@ -0,0 +1,346 @@ +# Responsible Visual Editing + +Minheng Ni $^{1,2}$ , Yeli Shen $^{2}$ , Lei Zhang $^{1\boxtimes}$ , and Wangmeng Zuo $^{2,3\boxtimes}$ + +1 The Hong Kong Polytechnic University + +$^{2}$ Harbin Institute of Technology + +3 Peng Cheng Laboratory, Guangzhou + +minheng.ni@connect.polyu.hk, ylshen@stu.hit.edu.cn, + +cslzhang@comp.polyu.edu.hk, WMzuo@hit.edu.cn + +Abstract. With the recent advancements in visual synthesis, there is a growing risk of encountering synthesized images with detrimental effects, such as hate, discrimination, and privacy violations. Unfortunately, it remains unexplored on how to avoid synthesizing harmful images and convert them into responsible ones. In this paper, we present responsible visual editing, which edits risky concepts within an image to more responsible ones with minimal content changes. However, the concepts that need to be edited are often abstract, making them hard to be located and edited. To tackle these challenges, we propose a Cognitive Editor (CoEditor) by harnessing the large multimodal models through a two-stage cognitive process: (1) a perceptual cognitive process to locate what to be edited and (2) a behavioral cognitive process to strategize how to edit. To mitigate the negative implications of harmful images on research, we build a transparent and public dataset, namely AltBear, which expresses harmful information using teddy bears instead of humans. Experiments demonstrate that CoEditor can effectively comprehend abstract concepts in complex scenes, significantly surpassing the baseline models for responsible visual editing. Moreover, we find that the AltBear dataset corresponds well to the harmful content found in real images, providing a safe and effective benchmark for future research. Our source code and dataset can be found at https://github.com/kodenii/Responsible-Visual-Editing. + +Keywords: Responsible visual editing $\cdot$ Image editing $\cdot$ Large multimodal model + +# 1 Introduction + +With the recent advancements in visual synthesis technologies [22,27,29,36], generating or editing highly realistic images has become possible, and the problem of misusing these technologies is arising [16,21,30] because the generated/edited images may contain harmful contents, such as hate, discrimination and privacy violations. While significant progress has been made in image editing [1,2,4] by using instructions to adjust images, it is less explored on how to edit harmful images into responsible ones. + +![](images/24ae8db0d3d5a271d9f8180b0b44f8e7c98e2631da6c447c36eae8d8f6bd3cc8.jpg) +Fig. 1: Overview of responsible visual editing. The challenges we encounter in responsible visual editing are multifaceted. Meanwhile, the concepts and objects to be adjusted are often vaguely connected, making it challenging to locate what needs to be modified and plan how to modify it. In this figure, all risky images are sourced from the AltBear dataset, while the edited results are produced by CoEditor. + +We formulate the above mentioned problem as a new task, namely responsible visual editing. As shown in Figure 1, we aim to edit specific concepts in images to make them more responsible while minimizing image changes as much as possible. Given the diversity of risky concepts in images and the different types of risks, we divide this task into three sub-tasks: safety, fairness and privacy, covering a wide range of risks in real-world scenarios. + +Existing editing models often require clear user instructions to make specific adjustments in the images [4,11,34], e.g., editing hat to "change the blue hat into red". However, in responsible image editing, the concepts that need to be edited are often abstract, e.g., editing violence to "make an image look less violent". This makes it challenging to locate what should to be edited and plan how to edit it. To tackle these challenges, we propose a Cognitive Editor (CoEditor) that harnesses large multimodal models (LMM) through a two-stage cognitive process: (1) a perceptual cognitive process to focus on what needs to be edited and (2) a behavioral cognitive process to strategize how to edit. + +To facilitate the research on responsible image editing, we build a transparent and public dataset, namely AltBear. Unlike general image datasets, the AltBear dataset uses fictional teddy bears as the protagonists, replacing humans in the images to convey risky contents. We significantly reduce the potential ethical risks by using teddy bears and cartoonized images. + +Experiments show that CoEditor significantly outperforms baseline models in responsible image editing, validating its effectiveness of comprehending abstract concepts and strategizing the editing process. Furthermore, we find that the AltBear dataset corresponds well to the harmful contents in real images, offering a consistent experimental evaluation. This also validates that AltBear can serve as a safer benchmark for future research. + +Our contributions are three-fold: + +- We propose a new task, responsible visual editing, and introduce a dataset, AltBear, which maintains high consistency with real-world data by using teddy bears as the protagonists to reduce the potential research risks. +- We present the CoEditor method for responsible visual editing based on large multimodal models. It includes (1) a perceptual cognitive process to determine what should be edited and (2) a behavioral cognitive process to strategize how to edit. +- Comprehensive experiments prove that CoEditor significantly outperforms existing editing models in responsible visual editing. Our findings reveal the potentials of LMM in responsible AI. + +# 2 Related Work + +# 2.1 Responsible Visual Synthesis + +With the advancement of visual synthesis techniques, how to ensure responsible visual editing has also received increasing attention. Some works have been reported to intercept images that may pose risks [3, 10, 31]. Rombach et al. [29] added a Not-Safe-For-Work (NSFW) filter at the end of the generation stage, while Rando et al. [28] chose to classify and intercept images by projecting them onto the latent space using CLIP. Some works [9, 12, 30] employed machine unlearning methods to forget risky concepts. Zhang et al. [35] and Kumari et al. [17] manipulated the latent variables during generation to avoid producing risky concepts. Recently, Ni et al. [21] proposed to intervene in the generation process using a large language model (LLM) to achieve training-free open-vocabulary responsible visual synthesis. However, current researches on responsible visual synthesis focus on text-to-image generation, while responsible visual editing has not yet been well explored. + +# 2.2 Image Editing + +Image editing is a classic task in computer vision. Some previous works used CNN-based methods to edit images [6, 8, 32]. By introducing adversarial techniques, StyleGAN and its variants [1, 14, 15] used GAN inversion for image editing. Using quantized models, MaskGIT [5] explored editing based on conditional vectors, and NUWA-LIP [20] explored natural language guided image editing. With the great success of diffusion models, more and more works attempt to use pre-trained text-to-image diffusion models for editing [2, 7, 19]. Prompt-to-prompt [13] obtained varying images by changing the attention during generation. Recently, InstructPix2pix [4] and InstructDiffusion [11] used the data generated by GPT and prompt-to-prompt methods to train the model to understand natural language instructions. However, current researches on image editing require clear or direct instructions. It has not yet been well explored how to use large multimodal models (LMMs) to understand and adjust the relationship between images and complex abstract instructions. + +# 3 Responsible Visual Editing + +# 3.1 Problem Formulation + +The goal of responsible visual editing is to automatically edit a given risky concept $c$ existed in image $x_{\mathrm{r}}$ , generating a responsible image $x_{\mathrm{s}}$ while making $x_{\mathrm{s}}$ visually reasonable and changing the image contents as little as possible. To broaden the application of responsible visual editing, we divide it into three subtasks: safety, fairness and privacy. + +Safety. The safety subtask focuses on inappropriate contents for display in real-world scenarios, such as discrimination, terrorist activities, or violence. This task requires completely removing the risky concepts from the image. + +Fairness. This subtask focuses on fairness issues in real-world scenarios, such as biased contents. It diversifies a specific concept in the image without changing much the image content. + +Privacy. The privacy subtask focuses on privacy issues in real-world scenarios, such as real-world characters. This task requires blurring a specific character in the image. While maintaining the basic meaning of the image, the editing should not be recognizable or traceable. + +# 3.2 AltBear Dataset Collection + +We collect a number of risky concepts, such as drug abuse, alcohol, racial discrimination, etc., and divide them into three subtasks: safety, fairness, and privacy. For each concept $c$ , we use ChatGPT [23] to expand it into 100 image scene descriptions and modify the subject of the description to teddy bears. We manually filter out and refine the compelling descriptions, and randomly use DALL-E 2 [26], Stable Diffusion XL [25], and DALL-E 3 [24] to generate the final risky image $x_{\mathrm{r}}$ . Then, we manually filter the dataset again, selecting the high-quality images. Finally, we obtain 300 groups of images as the test set. For more details, please refer to Supplementary Materials. + +# 3.3 Evaluation Metrics + +We propose two metrics to evaluate the performance of responsible image editing: success rate and visual similarity, both of which can be used for automatic machine evaluation and manual human evaluation. + +Success Rate. In the evaluation of success rate, we judge whether the concept $c$ in the responsible image $x_{s}$ still contains risks. For the safety subtask, the concept $c$ should not appear; for the fairness subtask, the diversity of the concept $c$ should be expanded; for the privacy subtask, we require that the identity of specific persons should not be recognized. + +Visual Similarity. In the evaluation of visual similarity, we compute the pixel-level similarity between the responsible image $x_{\mathrm{s}}$ and the original risky image $x_{\mathrm{r}}$ . Specifically, if $x_{\mathrm{s}}$ still contains risks, then the similarity is considered as 0. We take the average similarity as the final result. + +More details of the evaluation metrics can be found in the Supplementary Materials. + +# 3.4 Special Markers + +We add unique markers to the images to reduce potential risks in responsible visual editing. As shown in Figure 1, each image contains a marker in the lower right corner, indicating that this image is only used for responsible visual editing research and which subtask it belongs to. In addition, the highlighting symbol A or B represents the original or edited image. + +# 4 Methodology + +![](images/c1b654e72c8e66f1a79b9f6b179a40db30f9126ee1088ef37a0b2f0d027654ce.jpg) +Fig. 2: Overview of CoEditor. CoEditor consists of two stages of cognition: (1) a perceptional cognitive process (PCP) to understand what needs to be edited, and (2) a behavioral cognitive process (BCP) to plan how to edit. + +Compared with traditional visual editing that usually edits specific objects or features in an image, responsible visual editing should be able to edit any risky concepts contained in the image. It could be a theme such as violence, a category such as culture, or a person like Bill Gates. This task is very challenging and it contains two stages: (1) understanding the relationship between the image content and the concept to locate the regions to be edited, and (2) planning how to edit the regions to meet the editing conditions and image rationality. + +With the above considerations, we propose an LMM-based responsible editing model, namely Cognitive Editor (CoEditor). As shown in Figure 2, CoEditor consists of two stages of cognition: (1) perceptional cognitive process (PCP) to locate regions to be edited, and (2) behavioral cognitive process (BCP) to strategize how to edit such regions. + +# 4.1 Perceptional Cognitive Process + +Understanding the relationship between the content of an image and abstract concepts is often very difficult, as the concepts may not directly refer to specific + +contents in the image and it requires reasoning based on common sense. Even the powerful LMM may not accurately understand the concepts (see Section 5.2). To mitigate these issues, we use a visual language chain-of-thoughts method based on visual prompts and language prompts for perceptional cognition to locate the regions to be edited. + +For a risky image $x_{\mathrm{r}}$ , we need to visually annotate each element for LMM to refer to them accurately. We perform object extraction to obtain the object sequence $\mathcal{M}$ . In specific, we use Semantic-SAM [18] to extract objects of image $x$ to $n$ masks: + +$$ +\mathcal {M} = \left\{m _ {1}, m _ {2}, \dots , m _ {n} \right\}. \tag {1} +$$ + +Similar to SoM [33], we visually prompt the objects in the image to $v_{\mathrm{p}}$ : + +$$ +v _ {\mathrm {p}} = \phi (x, \mathcal {M}), \tag {2} +$$ + +where $\phi$ is a prompting function that adds a visual tag with a number at the corner of each mask's corresponding location in the image, enabling the LMM to refer to it in numeric tags. + +Meanwhile, in order to understand the possible meanings of the concept, we use knowledge extraction to expand its associated concepts and explanations into the text prompts $l_{\mathrm{p}}$ : + +$$ +l _ {\mathrm {p}} = f \left(c; p _ {\mathrm {k}} ^ {\text {i n s}}\right), \tag {3} +$$ + +where $p_{\mathrm{k}}^{\mathrm{ins}}$ is the instruction for knowledge extraction. Since there is no need to introduce visual information here, $f$ can be either an LLM or an LMM. + +Combining prompts of different modalities $v_{\mathrm{p}}$ and $l_{\mathrm{p}}$ , the LMM can complete the cognition of the image and the concept, and obtain the mask of the region $m_{\mathrm{p}}$ to be modified in the image: + +$$ +m _ {\mathrm {p}} = f \left(v _ {\mathrm {p}}, l _ {\mathrm {p}}; p _ {\mathrm {p}} ^ {\text {i n s}}\right), \tag {4} +$$ + +where $p_{\mathrm{p}}^{\mathrm{ins}}$ is the instruction for focus generation and $f$ is an LMM. Though we may obtain multiple related regions, for simplicity, we combine all regions into one single $m_{\mathrm{p}}$ as the result of PCP. + +# 4.2 Behavioral Cognitive Process + +In the perception cognitive process, the region $m_p$ to be edited has been identified, and we need to figure out how to edit it. Since the concept may be abstract, the editing model is hard to directly perform editing operations (see Section 5.2). Therefore, we employ the visual language chain-of-thoughts to plan the editing target, i.e., what contents are expected after responsible editing. + +We expand the object $m_p$ identified in the previous stage to $m_p'$ through object extension by enlarging the contour to include surrounding information. Then, we crop the image based on the mask $m_p'$ to obtain visual prompting: + +$$ +v _ {\mathrm {b}} = x _ {\mathrm {r}} \otimes m _ {\mathrm {p}} ^ {\prime}. \tag {5} +$$ + +At the same time, we perform instruction implementation to get the full language prompt $l_{\mathrm{b}}$ by concatenating the instruction of the task $p_{\mathrm{t}}^{\mathrm{ins}}$ and the input concept $c$ . Then we use the language prompt $l_{\mathrm{b}}$ together with the visual prompt $v_{\mathrm{b}}$ to generate the editing target $r_{\mathrm{b}}$ for the inpainting model: + +$$ +r _ {\mathrm {b}} = f \left(v _ {\mathrm {b}}, l _ {\mathrm {b}}; p _ {\mathrm {b}} ^ {\text {i n s}}\right), \tag {6} +$$ + +where $p_{\mathrm{b}}^{\mathrm{ins}}$ is the instruction for editing. + +Finally, we take the original image $x_{\mathrm{r}}$ , the region to be edited $m_p'$ , and the editing target $r_{\mathrm{b}}$ as the input to the inpainting model, and obtain the final editing result as follows: + +$$ +x _ {\mathrm {s}} = g \left(x _ {\mathrm {r}}, m _ {\mathrm {p}} ^ {\prime}; r _ {\mathrm {b}}\right), \tag {7} +$$ + +where $g$ is the inpainting model and $x_{\mathrm{s}}$ is the final responsible editing result. + +# 4.3 Implementation Details + +Our method is training-free. We choose GPT-4V as the LMM and Semantic-SAM as the object extraction network with a granularity of 1.5 and its officially released checkpoint4. For the inpainting model, we choose Stable Diffusion Inpainting trained based on Stable Diffusion v2 and its publicly available checkpoint5. Like most previous works, we adjust the input image to a size of $512 \times 512$ . All random seeds are fixed to 42. In all experiments, the instructions $p_{\mathrm{k}}^{\mathrm{ins}}$ , $p_{\mathrm{p}}^{\mathrm{ins}}$ , and $p_{\mathrm{b}}^{\mathrm{ins}}$ are fixed, and $p_{\mathrm{t}}^{\mathrm{ins}}$ depends on the task. Please refer to the Supplementary Materials for more details. + +# 5 Experiment + +# 5.1 Experiment Setup + +Since the concept of responsible visual editing is arbitrary and it does not specify the editing area, we select two powerful image editing models, InstructPix2pix [4] and InstructDiffusion [11], in the experiments. They not only allow arbitrary language guidance with the help of LLM but also require no additional masks to indicate the editing region. In addition, we could use the same Stable Diffusion base model in our CoEditor, making our comparison fair. + +We conduct experiments on the AltBear dataset. Since InstructPix2pix and InstructDiffusion do not support responsible editing of images based on the concept, we manually design editing conditions remove {concept} for safety and privacy tasks, and increase the variety of {concept} for fairness tasks. We evaluate the edited images in terms of both the success rate and visual similarity. We measure the results comprehensively via both automatic machine and manual human evaluations. + +Table 1: Overall results on AltBear under machine evaluation. We can find that CoEditor significantly outperforms the baseline models in both success rate and visual similarity, validating the effectiveness of CoEditor. + +
ModelSafetyFairnessPrivacyOverall
\( Succ^{\uparrow} \)\( Sim^{\uparrow} \)\( Succ^{\uparrow} \)\( Sim^{\uparrow} \)\( Succ^{\uparrow} \)\( Sim^{\uparrow} \)\( Succ^{\uparrow} \)\( Sim^{\uparrow} \)
InstructPix2pix21.31%0.196922.11%0.212351.22%0.456231.93%0.2924
InstructDiffusion40.54%0.357421.43%0.171966.33%0.496344.14%0.3497
CoEditor (Ours)70.13%0.565248.96%0.359578.35%0.639565.56%0.5188
+ +Table 2: Overall results on AltBear under human evaluation. In human evaluations, CoEditor shows consistent results with machine evaluations, showing that CoEditor is also effective from a human subjective perspective. + +
ModelSafetyFairnessPrivacyOverall
\( Succ^{\uparrow} \)\( Sim^{\uparrow} \)\( Succ^{\uparrow} \)\( Sim^{\uparrow} \)\( Succ^{\uparrow} \)\( Sim^{\uparrow} \)\( Succ^{\uparrow} \)\( Sim^{\uparrow} \)
InstructPix2pix28.41%0.25239.38%0.083940.00%0.327126.06%0.2217
InstructDiffusion40.91%0.307819.79%0.136956.00%0.345739.08%0.2634
CoEditor (Ours)65.91%0.531850.00%0.360665.00%0.518360.21%0.4692
+ +# 5.2 Overall Results + +Quantitative Results. As shown in Table 1 and Table 2, CoEditor shows significant advantages in both machine and human evaluations. We can see that for the success rate, CoEditor has more than $20\%$ increase in almost all subtasks. This proves that for responsible visual synthesis, using LMM to understand images and concepts and to strategize editing process is crucial. In addition, we find that CoEditor has higher visual similarity compared to traditional models, which reveals the importance of explicit cognition process of editing. It is noticed that the machine evaluation shows similar results to human evaluations, further proving that our editing results are of the same high quality from the perspective of human judgements. + +Qualitative Results. As shown in Figure 3, CoEditor demonstrates strong understanding and editing capabilities in responsible visual synthesis. In the first row of the safety task, we see that the CoEditor not only fully erases the concepts of drugs and alcohol but also maintains the rationality of the picture. However, InstructDiffusion fails to completely remove drugs and generates unreasonable contents in the picture. Moreover, InstructPix2pix does not follow the instruction well. In the second row of the fairness task, CoEditor successfully diversifies the specified concept. It maintains the original layout, while InstructDiffusion results in blurry or damaged outputs and InstructPix2pix has difficulties in executing this instruction. In the third row of the privacy task, the CoEditor blurs the characters' features while maintaining the original meaning with high visual quality. However, InstructDiffusion and InstructPix2pix fail in content removal + +Table 3: Ablation study of AltBear. Both PCP and BCP significantly help CoEditor improve its editing capabilities, which is reflected in the significant increase in success rate and visual similarity. + +
ModelSucc↑Sim↑
CoEditor w/o PCP44.19%0.3855
CoEditor w/o BCP28.02%0.2221
CoEditor (Full)65.56%0.5188
+ +or damage the picture. These examples show that CoEditor can understand the images and concepts and then plan the editing well with the help of LMM. More examples can be found in the Supplementary Materials. + +![](images/8a8b36668cce6de5a4a76205fe9444d40f78846f541495ae8b8739c6ef228754.jpg) +Fig. 3: Visualization results on AltBear. CoEditor performs well in all subtasks and maintains high visual similarity. The results of CoEditor also have better rationality. InstructDiffusion often over-edits images or produces unreasonable visual effects, while the editing ability of InstructPix2pix is weaker compared with CoEditor. + +![](images/fd46e533b2de3b84823e60c30a54b2d807a5ffaa50b7d356f9933ce94a9f1ddf.jpg) + +# 5.3 Ablation Study + +Quantitative Results. We conduct ablation experiments on AltBear in Table 3 to explore the effectiveness of the two stages of CoEditor. In the ablation of PCP, since the model cannot locate the area to be edited, we turn to use the largest area in the image. In the ablation of BCP, we use the same instruction as the baseline model for editing. To ensure fairness, all other parameters of the experiments are consistent. As shown in Table 3, both stages of CoEditor are significant for the final results. Without PCP, the CoEditor cannot understand the object to be edited, causing a massive drop in success rate and visual similarity. Without BCP, although the CoEditor can locate the region to be edited, + +simply asking the model to remove or increase diversity will make it difficult to understand what to edit, ultimately leading to editing failures. + +![](images/691c10985e6c77731abc03c65c572094d0f3ea77b524afb136e5ac6e6e62b39c.jpg) +Fig. 4: Roles of different components in CoEditor. Without the help of BCP, the CoEditor produces inconsistent results with the concept or visually unreasonable results because it does not know how to edit the content correctly. Without PCP, CoEditor is unable to locate the editing regions, resulting in editing failures. + +Qualitative Results. To investigate whether the two stages of CoEditor, BCP and PCP, can produce the expected outputs, we visualize some examples in Figure 3. One can see that the BCP stage is crucial for correct editing. In most examples, without the BCP stage, the editing result will fail to meet the requirements, or there will be visually unreasonable contents. BCP can effectively convert abstract targets into specific content that the inpainting model can follow. This shows the importance of editing planning. At the same time, PCP plays a decisive role. Without the participation of PCP, most of the edits will fail. This is because most of the concepts to be edited do not directly associate with a specific object in the image. Therefore, a deep understanding of image and concepts using PCP is crucial for CoEditor to edit correctly. The CoEditor can achieve satisfactory effects when both the two modules are used. Extra ablation studies can be found in the Supplementary Materials. + +# 5.4 Exploration of Cognitive Process + +To explore why CoEditor can successfully perform complex and responsible visual editing, we show the intermediate results of various cognitive processes in Figure 5. We can see that PCP can successfully locate regions to be edited, regardless of whether this region is directly related to the concept. Based on the task type, PCP can also accurately find out the region to be edited among multiple identical objects. At the same time, BCP can successfully plan the editing. No matter what the concept is, BCP can accurately associate it with the + +![](images/db3085b14647fcde9c7a4d6fc3b6049fbe1d9a3ea9f8a86547a2519f2c09c7b0.jpg) +Risky Image + +![](images/c92dc428fd3f738db7567cf43db23e113bcad4f89e0af9d60fd319865808ba20.jpg) +PCP Result +BCP Result + +A cuddly teddy bear enjoying a sunny day on a serene picnic with a picturesque house fully intact in the background, surrounded by vibrant greenery. + +![](images/c494380b763d0d6f31b8a24361f8473f21743d2f5e56333332931f2ceb6c495e.jpg) +Final Result + +![](images/563302e456afc9aa13a74b271c3081f507bece9347aa64d7695be69d243c054d.jpg) +Concept: arsion + +![](images/2eb50c04f9e024962d4bf40df3aef89ba106c76425d5ad6b028046e26c767b73.jpg) + +A teddy bear wearing traditional attire representing different cultures, such as a Scottish kilt, an Indian sari, and a Nigerian abada, showcasing the beautiful variety of cultural dress from around the world. + +![](images/5c15a380009225a6787d2b05db301bb943440941b0cdd31797e053e7ca97f831.jpg) + +![](images/93bf4ed5f587f86ef9980443d977ce2a3a371fccce794ea7e59b430b3c568128.jpg) +Concept: culture +Fig. 5: Intermediate results of cognitive process in CoEditor. We can see that PCP can successfully find one or more objects to be edited, even if the object is not directly related to the concept. BCP can generate effective editing target based on the positioning of PCP, even if the scenario is very complex. The effectiveness of the two cognitive processes allows CoEditor to edit images successfully. + +![](images/dc7cd48e7c1584d8a1019ec5f7bf844624b792aae4dcaaef1f4d698737451c84.jpg) + +A fluffy teddy bear with sparkling eyes and shiny footwear stands in front of a vintage microphone on a soft glowing background. + +![](images/59fe2d45b8982f14e836b50c8c62d209f2487c1eb8d506e18d25b6f88e0892f9.jpg) + +Concept: Marilyn Monroe + +content in the image to obtain editing target consistent with the image scene. This shows the importance of cognitive processes to CoEditor in concept understanding and editing planning. For further explorations of CoEditor, please see the Supplementary Materials. + +# 5.5 Performance Consistency on AltBear and Real-world Data + +To verify whether AltBear can serve as a replacement of real-world images in subsequent research, we examine the performance consistency between AltBear and real-world images. We build a dataset that has similar size to AltBear, consisting of real-world images collected from the Internet. We conduct experiments on this dataset with all hyper-parameters and models identical to that on AltBear. As shown in Table 4 and 5, CoEditor achieves similar performance to that on AltBear, demonstrating CoEditor's robustness to both synthesized data and real-world data. The baseline models, InstructPix2pix and InstructDiffusion, also show similar performance on AltBear and real-world data. Since the protagonist is not a human but a fictional teddy bear, AltBear significantly reduces the risk of image propagation and display, effectively avoiding ethical risks. + +Table 4: Results on real-world images under machine evaluation. On real-world datasets, CoEditor shows similar performance to that on AltBear, significantly surpassing the baseline model. Furthermore, we find that the distributions of both success rate and visual similarity are close for real-world images and AltBear. + +
ModelSafetyFairnessPrivacyOverall
\( Succ^{\uparrow} \)\( Sim^{\uparrow} \)\( Succ^{\uparrow} \)\( Sim^{\uparrow} \)\( Succ^{\uparrow} \)\( Sim^{\uparrow} \)\( Succ^{\uparrow} \)\( Sim^{\uparrow} \)
InstructPix2pix14.00%0.12831.89%0.016622.86%0.216411.59%0.1078
InstructDiffusion38.60%0.318514.29%0.103055.81%0.374435.57%0.2638
CoEditor (Ours)59.46%0.474058.18%0.425363.41%0.513960.00%0.4679
+ +Table 5: Results of real-world images under human evaluation. CoEditor achieves significantly better results than the baseline model in human evaluation. In addition, the quantitative results on AltBear and real-world images in human evaluation are also similar. + +
ModelSafetyFairnessPrivacyOverall
\( Succ^{\uparrow} \)\( Sim^{\uparrow} \)\( Succ^{\uparrow} \)\( Sim^{\uparrow} \)\( Succ^{\uparrow} \)\( Sim^{\uparrow} \)\( Succ^{\uparrow} \)\( Sim^{\uparrow} \)
InstructPix2pix31.25%0.27989.09%0.080415.56%0.132720.56%0.1821
InstructDiffusion47.50%0.376421.82%0.152453.33%0.337841.11%0.2983
CoEditor (Ours)68.75%0.548665.45%0.478780.00%0.646770.56%0.5518
+ +# 5.6 Real-world Results + +To better demonstrate the robustness of CoEditor, we show its performance on real-world images and compare with baseline models in Figure 6. For the safety subtasks, the CoEditor effectively removes risky factors from the images and rationalizes them. In the first row, the robber in the first image is turned into an enthusiastic saleswoman, and the craving for drugs in the second image is turned into a craving for food. This also shows the imagination capability of CoEditor. For the fairness subtasks in the second row, the CoEditor completes the image editing while maintaining the visual quality of the image. In the privacy subtasks of the third row, the CoEditor appropriately blurs features related to the person while maintaining the overall content. Similar to the results on AltBear, the baseline models, InstructPix2pix and InstructDiffusion, struggle to maintain visual quality or rationality in some examples, even though they apply editing. In other cases, they produce damaged figures. More real-world examples can be found in the Supplementary Materials. + +# 5.7 Analysis of Computational Overhead + +In order to explore how CoEditor consumes resources, we make comparisons on time and memory usage. We use the same full-precision and optimization settings, e.g., memory-efficient attention for all models. + +![](images/eab02e419db666486b28844e909dfeababe86416da3515e3dc15050be8fba9a8.jpg) + +![](images/0035c86a07e4a4778abf6e7249dbe49ee7880418dfeef285b95760d837e427f0.jpg) + +![](images/98cf5188fec5ab838a9bc9de29ed598134f15c9cc1c9bafae8900433a18492d6.jpg) +Fig. 6: Visualized results on real-world data. Even in real-world scenarios, CoEditor can still work well. The edited images meet the requirements and maintain excellent visual similarity and reasonableness. Similar to the results on AltBear, InstructDiffusion and InstructPix2pix fail to output well edited images. + +![](images/a917dc032a74ac908fe87c411ad6916173c4352d7ba6ed4f075b40d5289ef618.jpg) + +![](images/94971041a4b6f0fcd983f4dae36158d328646b35ec0696d7d01c5c8b769dedb6.jpg) + +![](images/a3fd1d755e3adb7d2268ce871babff6d86375c6dd9216e7d0016025b1b380dcf.jpg) + +Table 6: Comparison on inference time. + +
ModelTimeMultiple
InstrcutPix2pix12.43s1×
InstructDiffusion20.75s1.67×
CoEditor14.10s1.13×
PCP Stage4.65s-
BCP Stage9.44s-
+ +Time. We calculate the average inference time of different models. For CoEditor, we also calculate the network latency. As shown in Table 6, the speed of CoEditor is even faster than InstructDiffusion [11], and is comparable to Instruct Pix2pix [4]. + +GPU Memory. We calculate the peak VRAM usage of different models. As shown in Table 7, CoEditor consumes the least memory. This is because our chosen models, i.e., Semantic-SAM [18] and Stable Diffusion Inpainting [29], are very lightweight, which allows our model to work with minimal memory. + +# 6 Ethics Statement + +# 6.1 Inappropriate Content, Privacy, and Discrimination + +The AltBear dataset uses teddy bears as the protagonist, significantly reducing the risks caused by the public display of inappropriate contents. We manually reviewed all contents to ensure that privacy or discrimination is avoided. We release AltBear under the MIT license to keep transparency. + +Table 7: Comparison on GPU memory. + +
ModelVRAMMultiple
InstrcutPix2pix18,118MB1×
InstructDiffusion8,378MB0.46×
CoEditor7,328MB0.40×
+ +# 6.2 Reproducibility + +We build our method on top of the Stable Diffusion Inpainting model with publicly accessible code and checkpoints with a fixed random seed. We notice that the GPT API cannot guarantee identical responses so that we provide all instruction prompts for reference. To further improve the reproducibility, we also release the dataset, code and our trained model. + +# 6.3 Anti-misuse + +In order to ensure that the dataset is not misused, we have designed unique markers (refer to Section 3.4). We added the marker to the bottom right corner of the image to indicate that the image is used for responsible visual editing research only. Specifically, we have defined various tags to indicate whether the image has been edited or contains potential risks. + +In addition, the interaction system of our model always maintains visibility of the concepts to be edited to prevent misuse of the model. We also call on subsequent work to maintain the visibility of concepts like ours, ensuring a beneficial role for the community through open and transparent methods. + +# 7 Conclusion + +In this work, we proposed a new task, namely responsible visual editing, which entails editing specific concepts within an image to render it more responsible while minimizing changes. To tackle abstract concepts in responsible visual editing, we proposed a Cognitive Editor (CoEditor) by harnessing a large multimodal model (LMM) through a two-stage process: a perceptual cognitive process to locate what needs to be edited, and a behavioral cognitive process to strategize how to edit. We constructed a transparent and public dataset, AltBear, which represents harmful information using teddy bears instead of humans, thereby mitigating the negative implications on research. Experiments demonstrated that CoEditor can effectively comprehend abstract concepts within complex scenes and significantly surpass the performance of baseline models for responsible visual editing. Moreover, we validated that the AltBear dataset can well serve as a replacement of real-world images in the research of responsible visual editing, providing a safer benchmark. Our findings also revealed the potential of LMM in responsible AI. + +# Acknowledgement + +This work was supported in part by the National Key R&D Program of China (2022YFA1004100), and the National Natural Science Foundation of China (NSFC) under grant No. 62441202. + +# References + +1. Abdal, R., Qin, Y., Wonka, P.: Image2stylegan++: How to edit the embedded images? In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 8296-8305 (2020) +2. Avrahami, O., Lischinski, D., Fried, O.: Blended diffusion for text-driven editing of natural images. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18208-18218 (2022) +3. Brack, M., Friedrich, F., Hintersdorf, D., Struppek, L., Schramowski, P., Kersting, K.: Sega: Instructing text-to-image models using semantic guidance. Advances in Neural Information Processing Systems 36 (2024) +4. Brooks, T., Holynski, A., Efros, A.A.: Instructpix2pix: Learning to follow image editing instructions. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18392-18402 (2023) +5. Chang, H., Zhang, H., Jiang, L., Liu, C., Freeman, W.T.: Maskgit: Masked generative image transformer. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11315-11325 (2022) +6. Chen, J., Shen, Y., Gao, J., Liu, J., Liu, X.: Language-based image editing with recurrent attentive models. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 8721-8729 (2018) +7. Couairon, G., Verbeek, J., Schwenk, H., Cord, M.: Diffedit: Diffusion-based semantic image editing with mask guidance. arXiv preprint arXiv:2210.11427 (2022) +8. Dong, H., Yu, S., Wu, C., Guo, Y.: Semantic image synthesis via adversarial learning. In: Proceedings of the IEEE international conference on computer vision. pp. 5706-5714 (2017) +9. Gandikota, R., Materzynska, J., Fiotto-Kaufman, J., Bau, D.: Erasing concepts from diffusion models. arXiv preprint arXiv:2303.07345 (2023) +0. Gandikota, R., Orgad, H., Belinkov, Y., Materzyńska, J., Bau, D.: Unified concept editing in diffusion models. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 5111-5120 (2024) +1. Geng, Z., Yang, B., Hang, T., Li, C., Gu, S., Zhang, T., Bao, J., Zhang, Z., Hu, H., Chen, D., et al.: Instructdiffusion: A generalist modeling interface for vision tasks. arXiv preprint arXiv:2309.03895 (2023) +2. Heng, A., Soh, H.: Continual learning for forgetting in deep generative models (2023) +3. Hertz, A., Mokady, R., Tenenbaum, J., Aberman, K., Pritch, Y., Cohen-Or, D.: Prompt-to-prompt image editing with cross attention control. arXiv preprint arXiv:2208.01626 (2022) +4. Karras, T., Laine, S., Aila, T.: A style-based generator architecture for generative adversarial networks. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 4401-4410 (2019) +5. Karras, T., Laine, S., Aittala, M., Hellsten, J., Lehtinen, J., Aila, T.: Analyzing and improving the image quality of stylegan. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 8110-8119 (2020) + +16. Kumari, N., Zhang, B., Wang, S.Y., Shechtman, E., Zhang, R., Zhu, J.Y.: Ablating Concepts in Text-to-Image Diffusion Models (May 2023), http://arxiv.org/abs/2303.13516, arXiv:2303.13516 [cs] +17. Kumari, N., Zhang, B., Wang, S.Y., Shechtman, E., Zhang, R., Zhu, J.Y.: Ablating concepts in text-to-image diffusion models. arXiv preprint arXiv:2303.13516 (2023) +18. Li, F., Zhang, H., Sun, P., Zou, X., Liu, S., Yang, J., Li, C., Zhang, L., Gao, J.: Semantic-sam: Segment and recognize anything at any granularity. arXiv preprint arXiv:2307.04767 (2023) +19. Meng, C., He, Y., Song, Y., Song, J., Wu, J., Zhu, J.Y., Ermon, S.: Sdedit: Guided image synthesis and editing with stochastic differential equations. arXiv preprint arXiv:2108.01073 (2021) +20. Ni, M., Li, X., Zuo, W.: Nuwa-lip: Language-guided image inpainting with defect-free vqgan. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14183-14192 (2023) +21. Ni, M., Wu, C., Wang, X., Yin, S., Wang, L., Liu, Z., Duan, N.: Ores: Open-vocabulary responsible visual synthesis. arXiv preprint arXiv:2308.13785 (2023) +22. Nichol, A., Dhariwal, P., Ramesh, A., Shyam, P., Mishkin, P., McGrew, B., Sutskever, I., Chen, M.: Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741 (2021) +23. OpenAI: Chatgpt. https://chat.openai.com (2022) +24. OpenAI: Dall-e 3. https://openai.com/dall-e-3 (2023) +25. Podell, D., English, Z., Lacey, K., Blattmann, A., Dockhorn, T., Müller, J., Penna, J., Rombach, R.: Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952 (2023) +26. Ramesh, A., Dhariwal, P., Nichol, A., Chu, C., Chen, M.: Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125 1(2), 3 (2022) +27. Ramesh, A., Pavlov, M., Goh, G., Gray, S., Voss, C., Radford, A., Chen, M., Sutskever, I.: Zero-shot text-to-image generation. In: International Conference on Machine Learning. pp. 8821-8831. PMLR (2021) +28. Rando, J., Paleka, D., Lindner, D., Heim, L., Tramér, F.: Red-teaming the stable diffusion safety filter. arXiv preprint arXiv:2210.04610 (2022) +29. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10684-10695 (2022) +30. Schramowski, P., Brack, M., Deiseroth, B., Kersting, K.: Safe latent diffusion: Mitigating inappropriate degeneration in diffusion models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 22522-22531 (2023) +31. Schuhmann, C., Beaumont, R., Vencu, R., Gordon, C., Wightman, R., Cherti, M., Coombes, T., Katta, A., Mullis, C., Wortsman, M., et al.: Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems 35, 25278-25294 (2022) +32. Wang, S.Y., Wang, O., Zhang, R., Owens, A., Efros, A.A.: Cnn-generated images are surprisingly easy to spot... for now. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 8695-8704 (2020) +33. Yang, J., Zhang, H., Li, F., Zou, X., Li, C., Gao, J.: Set-of-mark prompting unleashes extraordinary visual grounding in gpt-4v. arXiv preprint arXiv:2310.11441 (2023) +34. Yu, T., Feng, R., Feng, R., Liu, J., Jin, X., Zeng, W., Chen, Z.: Inpaint anything: Segment anything meets image inpainting. arXiv preprint arXiv:2304.06790 (2023) + +35. Zhang, E., Wang, K., Xu, X., Wang, Z., Shi, H.: Forget-me-not: Learning to forget in text-to-image diffusion models. arXiv preprint arXiv:2303.17591 (2023) +36. Zhang, Y., Wei, Y., Jiang, D., Zhang, X., Zuo, W., Tian, Q.: Control video: Training-free controllable text-to-video generation. arXiv preprint arXiv:2305.13077 (2023) \ No newline at end of file diff --git a/2024/Responsible Visual Editing/images.zip b/2024/Responsible Visual Editing/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..45d79f44423b1b8b5a0434dafbc3fd561893d0e7 --- /dev/null +++ b/2024/Responsible Visual Editing/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:15302053da95d3bd2045706af2943525a04c1d3ab9dcb0434ac3b2f84924130a +size 700939 diff --git a/2024/Responsible Visual Editing/layout.json b/2024/Responsible Visual Editing/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..15e0e8a7fed20c71734a81618f51fe937d784195 --- /dev/null +++ b/2024/Responsible Visual Editing/layout.json @@ -0,0 +1,9098 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 212, + 112, + 402, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 212, + 112, + 402, + 129 + ], + "spans": [ + { + "bbox": [ + 212, + 112, + 402, + 129 + ], + "type": "text", + "content": "Responsible Visual Editing" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 162, + 149, + 451, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 162, + 149, + 451, + 163 + ], + "spans": [ + { + "bbox": [ + 162, + 149, + 451, + 163 + ], + "type": "text", + "content": "Minheng Ni" + }, + { + "bbox": [ + 162, + 149, + 451, + 163 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 162, + 149, + 451, + 163 + ], + "type": "text", + "content": ", Yeli Shen" + }, + { + "bbox": [ + 162, + 149, + 451, + 163 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 162, + 149, + 451, + 163 + ], + "type": "text", + "content": ", Lei Zhang" + }, + { + "bbox": [ + 162, + 149, + 451, + 163 + ], + "type": "inline_equation", + "content": "^{1\\boxtimes}" + }, + { + "bbox": [ + 162, + 149, + 451, + 163 + ], + "type": "text", + "content": ", and Wangmeng Zuo" + }, + { + "bbox": [ + 162, + 149, + 451, + 163 + ], + "type": "inline_equation", + "content": "^{2,3\\boxtimes}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 221, + 171, + 392, + 183 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 221, + 171, + 392, + 183 + ], + "spans": [ + { + "bbox": [ + 221, + 171, + 392, + 183 + ], + "type": "text", + "content": "1 The Hong Kong Polytechnic University" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 239, + 183, + 374, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 239, + 183, + 374, + 194 + ], + "spans": [ + { + "bbox": [ + 239, + 183, + 374, + 194 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 239, + 183, + 374, + 194 + ], + "type": "text", + "content": " Harbin Institute of Technology" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 227, + 194, + 386, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 227, + 194, + 386, + 205 + ], + "spans": [ + { + "bbox": [ + 227, + 194, + 386, + 205 + ], + "type": "text", + "content": "3 Peng Cheng Laboratory, Guangzhou" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 185, + 206, + 426, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 185, + 206, + 426, + 217 + ], + "spans": [ + { + "bbox": [ + 185, + 206, + 426, + 217 + ], + "type": "text", + "content": "minheng.ni@connect.polyu.hk, ylshen@stu.hit.edu.cn," + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 202, + 217, + 411, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 202, + 217, + 411, + 228 + ], + "spans": [ + { + "bbox": [ + 202, + 217, + 411, + 228 + ], + "type": "text", + "content": "cslzhang@comp.polyu.edu.hk, WMzuo@hit.edu.cn" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 159, + 259, + 455, + 501 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 259, + 455, + 501 + ], + "spans": [ + { + "bbox": [ + 159, + 259, + 455, + 501 + ], + "type": "text", + "content": "Abstract. With the recent advancements in visual synthesis, there is a growing risk of encountering synthesized images with detrimental effects, such as hate, discrimination, and privacy violations. Unfortunately, it remains unexplored on how to avoid synthesizing harmful images and convert them into responsible ones. In this paper, we present responsible visual editing, which edits risky concepts within an image to more responsible ones with minimal content changes. However, the concepts that need to be edited are often abstract, making them hard to be located and edited. To tackle these challenges, we propose a Cognitive Editor (CoEditor) by harnessing the large multimodal models through a two-stage cognitive process: (1) a perceptual cognitive process to locate what to be edited and (2) a behavioral cognitive process to strategize how to edit. To mitigate the negative implications of harmful images on research, we build a transparent and public dataset, namely AltBear, which expresses harmful information using teddy bears instead of humans. Experiments demonstrate that CoEditor can effectively comprehend abstract concepts in complex scenes, significantly surpassing the baseline models for responsible visual editing. Moreover, we find that the AltBear dataset corresponds well to the harmful content found in real images, providing a safe and effective benchmark for future research. Our source code and dataset can be found at https://github.com/kodenii/Responsible-Visual-Editing." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 159, + 511, + 453, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 511, + 453, + 533 + ], + "spans": [ + { + "bbox": [ + 159, + 511, + 453, + 533 + ], + "type": "text", + "content": "Keywords: Responsible visual editing " + }, + { + "bbox": [ + 159, + 511, + 453, + 533 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 511, + 453, + 533 + ], + "type": "text", + "content": " Image editing " + }, + { + "bbox": [ + 159, + 511, + 453, + 533 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 511, + 453, + 533 + ], + "type": "text", + "content": " Large multimodal model" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 555, + 230, + 568 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 555, + 230, + 568 + ], + "spans": [ + { + "bbox": [ + 132, + 555, + 230, + 568 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "content": "With the recent advancements in visual synthesis technologies [22,27,29,36], generating or editing highly realistic images has become possible, and the problem of misusing these technologies is arising [16,21,30] because the generated/edited images may contain harmful contents, such as hate, discrimination and privacy violations. While significant progress has been made in image editing [1,2,4] by using instructions to adjust images, it is less explored on how to edit harmful images into responsible ones." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 136, + 114, + 478, + 255 + ], + "blocks": [ + { + "bbox": [ + 136, + 114, + 478, + 255 + ], + "lines": [ + { + "bbox": [ + 136, + 114, + 478, + 255 + ], + "spans": [ + { + "bbox": [ + 136, + 114, + 478, + 255 + ], + "type": "image", + "image_path": "24ae8db0d3d5a271d9f8180b0b44f8e7c98e2631da6c447c36eae8d8f6bd3cc8.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 263, + 482, + 319 + ], + "lines": [ + { + "bbox": [ + 130, + 263, + 482, + 319 + ], + "spans": [ + { + "bbox": [ + 130, + 263, + 482, + 319 + ], + "type": "text", + "content": "Fig. 1: Overview of responsible visual editing. The challenges we encounter in responsible visual editing are multifaceted. Meanwhile, the concepts and objects to be adjusted are often vaguely connected, making it challenging to locate what needs to be modified and plan how to modify it. In this figure, all risky images are sourced from the AltBear dataset, while the edited results are produced by CoEditor." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 347, + 482, + 419 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 347, + 482, + 419 + ], + "spans": [ + { + "bbox": [ + 130, + 347, + 482, + 419 + ], + "type": "text", + "content": "We formulate the above mentioned problem as a new task, namely responsible visual editing. As shown in Figure 1, we aim to edit specific concepts in images to make them more responsible while minimizing image changes as much as possible. Given the diversity of risky concepts in images and the different types of risks, we divide this task into three sub-tasks: safety, fairness and privacy, covering a wide range of risks in real-world scenarios." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 422, + 482, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 422, + 482, + 529 + ], + "spans": [ + { + "bbox": [ + 130, + 422, + 482, + 529 + ], + "type": "text", + "content": "Existing editing models often require clear user instructions to make specific adjustments in the images [4,11,34], e.g., editing hat to \"change the blue hat into red\". However, in responsible image editing, the concepts that need to be edited are often abstract, e.g., editing violence to \"make an image look less violent\". This makes it challenging to locate what should to be edited and plan how to edit it. To tackle these challenges, we propose a Cognitive Editor (CoEditor) that harnesses large multimodal models (LMM) through a two-stage cognitive process: (1) a perceptual cognitive process to focus on what needs to be edited and (2) a behavioral cognitive process to strategize how to edit." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 531, + 482, + 591 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 531, + 482, + 591 + ], + "spans": [ + { + "bbox": [ + 130, + 531, + 482, + 591 + ], + "type": "text", + "content": "To facilitate the research on responsible image editing, we build a transparent and public dataset, namely AltBear. Unlike general image datasets, the AltBear dataset uses fictional teddy bears as the protagonists, replacing humans in the images to convey risky contents. We significantly reduce the potential ethical risks by using teddy bears and cartoonized images." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 594, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 665 + ], + "type": "text", + "content": "Experiments show that CoEditor significantly outperforms baseline models in responsible image editing, validating its effectiveness of comprehending abstract concepts and strategizing the editing process. Furthermore, we find that the AltBear dataset corresponds well to the harmful contents in real images, offering a consistent experimental evaluation. This also validates that AltBear can serve as a safer benchmark for future research." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 215, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 215, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 215, + 100 + ], + "type": "text", + "content": "M. Ni et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 146, + 116, + 292, + 127 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 116, + 292, + 127 + ], + "spans": [ + { + "bbox": [ + 146, + 116, + 292, + 127 + ], + "type": "text", + "content": "Our contributions are three-fold:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 137, + 134, + 481, + 251 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 137, + 134, + 481, + 169 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 134, + 481, + 169 + ], + "spans": [ + { + "bbox": [ + 137, + 134, + 481, + 169 + ], + "type": "text", + "content": "- We propose a new task, responsible visual editing, and introduce a dataset, AltBear, which maintains high consistency with real-world data by using teddy bears as the protagonists to reduce the potential research risks." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 137, + 170, + 481, + 215 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 170, + 481, + 215 + ], + "spans": [ + { + "bbox": [ + 137, + 170, + 481, + 215 + ], + "type": "text", + "content": "- We present the CoEditor method for responsible visual editing based on large multimodal models. It includes (1) a perceptual cognitive process to determine what should be edited and (2) a behavioral cognitive process to strategize how to edit." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 137, + 217, + 481, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 217, + 481, + 251 + ], + "spans": [ + { + "bbox": [ + 137, + 217, + 481, + 251 + ], + "type": "text", + "content": "- Comprehensive experiments prove that CoEditor significantly outperforms existing editing models in responsible visual editing. Our findings reveal the potentials of LMM in responsible AI." + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 132, + 269, + 237, + 281 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 269, + 237, + 281 + ], + "spans": [ + { + "bbox": [ + 132, + 269, + 237, + 281 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 292, + 307, + 304 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 292, + 307, + 304 + ], + "spans": [ + { + "bbox": [ + 132, + 292, + 307, + 304 + ], + "type": "text", + "content": "2.1 Responsible Visual Synthesis" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 309, + 482, + 465 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 309, + 482, + 465 + ], + "spans": [ + { + "bbox": [ + 130, + 309, + 482, + 465 + ], + "type": "text", + "content": "With the advancement of visual synthesis techniques, how to ensure responsible visual editing has also received increasing attention. Some works have been reported to intercept images that may pose risks [3, 10, 31]. Rombach et al. [29] added a Not-Safe-For-Work (NSFW) filter at the end of the generation stage, while Rando et al. [28] chose to classify and intercept images by projecting them onto the latent space using CLIP. Some works [9, 12, 30] employed machine unlearning methods to forget risky concepts. Zhang et al. [35] and Kumari et al. [17] manipulated the latent variables during generation to avoid producing risky concepts. Recently, Ni et al. [21] proposed to intervene in the generation process using a large language model (LLM) to achieve training-free open-vocabulary responsible visual synthesis. However, current researches on responsible visual synthesis focus on text-to-image generation, while responsible visual editing has not yet been well explored." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 480, + 233, + 492 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 480, + 233, + 492 + ], + "spans": [ + { + "bbox": [ + 132, + 480, + 233, + 492 + ], + "type": "text", + "content": "2.2 Image Editing" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "text", + "content": "Image editing is a classic task in computer vision. Some previous works used CNN-based methods to edit images [6, 8, 32]. By introducing adversarial techniques, StyleGAN and its variants [1, 14, 15] used GAN inversion for image editing. Using quantized models, MaskGIT [5] explored editing based on conditional vectors, and NUWA-LIP [20] explored natural language guided image editing. With the great success of diffusion models, more and more works attempt to use pre-trained text-to-image diffusion models for editing [2, 7, 19]. Prompt-to-prompt [13] obtained varying images by changing the attention during generation. Recently, InstructPix2pix [4] and InstructDiffusion [11] used the data generated by GPT and prompt-to-prompt methods to train the model to understand natural language instructions. However, current researches on image editing require clear or direct instructions. It has not yet been well explored how to use large multimodal models (LMMs) to understand and adjust the relationship between images and complex abstract instructions." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 336, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 336, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 336, + 91, + 447, + 102 + ], + "type": "text", + "content": "Responsible Visual Editing" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 314, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 314, + 129 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 314, + 129 + ], + "type": "text", + "content": "3 Responsible Visual Editing" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 140, + 268, + 150 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 140, + 268, + 150 + ], + "spans": [ + { + "bbox": [ + 132, + 140, + 268, + 150 + ], + "type": "text", + "content": "3.1 Problem Formulation" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 159, + 481, + 219 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 159, + 481, + 219 + ], + "spans": [ + { + "bbox": [ + 132, + 159, + 481, + 219 + ], + "type": "text", + "content": "The goal of responsible visual editing is to automatically edit a given risky concept " + }, + { + "bbox": [ + 132, + 159, + 481, + 219 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 132, + 159, + 481, + 219 + ], + "type": "text", + "content": " existed in image " + }, + { + "bbox": [ + 132, + 159, + 481, + 219 + ], + "type": "inline_equation", + "content": "x_{\\mathrm{r}}" + }, + { + "bbox": [ + 132, + 159, + 481, + 219 + ], + "type": "text", + "content": ", generating a responsible image " + }, + { + "bbox": [ + 132, + 159, + 481, + 219 + ], + "type": "inline_equation", + "content": "x_{\\mathrm{s}}" + }, + { + "bbox": [ + 132, + 159, + 481, + 219 + ], + "type": "text", + "content": " while making " + }, + { + "bbox": [ + 132, + 159, + 481, + 219 + ], + "type": "inline_equation", + "content": "x_{\\mathrm{s}}" + }, + { + "bbox": [ + 132, + 159, + 481, + 219 + ], + "type": "text", + "content": " visually reasonable and changing the image contents as little as possible. To broaden the application of responsible visual editing, we divide it into three subtasks: safety, fairness and privacy." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 220, + 481, + 255 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 220, + 481, + 255 + ], + "spans": [ + { + "bbox": [ + 132, + 220, + 481, + 255 + ], + "type": "text", + "content": "Safety. The safety subtask focuses on inappropriate contents for display in real-world scenarios, such as discrimination, terrorist activities, or violence. This task requires completely removing the risky concepts from the image." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 255, + 481, + 290 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 255, + 481, + 290 + ], + "spans": [ + { + "bbox": [ + 132, + 255, + 481, + 290 + ], + "type": "text", + "content": "Fairness. This subtask focuses on fairness issues in real-world scenarios, such as biased contents. It diversifies a specific concept in the image without changing much the image content." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 291, + 481, + 338 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 291, + 481, + 338 + ], + "spans": [ + { + "bbox": [ + 132, + 291, + 481, + 338 + ], + "type": "text", + "content": "Privacy. The privacy subtask focuses on privacy issues in real-world scenarios, such as real-world characters. This task requires blurring a specific character in the image. While maintaining the basic meaning of the image, the editing should not be recognizable or traceable." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 357, + 297, + 367 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 357, + 297, + 367 + ], + "spans": [ + { + "bbox": [ + 132, + 357, + 297, + 367 + ], + "type": "text", + "content": "3.2 AltBear Dataset Collection" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 376, + 481, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 376, + 481, + 483 + ], + "spans": [ + { + "bbox": [ + 132, + 376, + 481, + 483 + ], + "type": "text", + "content": "We collect a number of risky concepts, such as drug abuse, alcohol, racial discrimination, etc., and divide them into three subtasks: safety, fairness, and privacy. For each concept " + }, + { + "bbox": [ + 132, + 376, + 481, + 483 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 132, + 376, + 481, + 483 + ], + "type": "text", + "content": ", we use ChatGPT [23] to expand it into 100 image scene descriptions and modify the subject of the description to teddy bears. We manually filter out and refine the compelling descriptions, and randomly use DALL-E 2 [26], Stable Diffusion XL [25], and DALL-E 3 [24] to generate the final risky image " + }, + { + "bbox": [ + 132, + 376, + 481, + 483 + ], + "type": "inline_equation", + "content": "x_{\\mathrm{r}}" + }, + { + "bbox": [ + 132, + 376, + 481, + 483 + ], + "type": "text", + "content": ". Then, we manually filter the dataset again, selecting the high-quality images. Finally, we obtain 300 groups of images as the test set. For more details, please refer to Supplementary Materials." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 502, + 256, + 512 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 502, + 256, + 512 + ], + "spans": [ + { + "bbox": [ + 132, + 502, + 256, + 512 + ], + "type": "text", + "content": "3.3 Evaluation Metrics" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 521, + 481, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 521, + 481, + 556 + ], + "spans": [ + { + "bbox": [ + 132, + 521, + 481, + 556 + ], + "type": "text", + "content": "We propose two metrics to evaluate the performance of responsible image editing: success rate and visual similarity, both of which can be used for automatic machine evaluation and manual human evaluation." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 558, + 481, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 558, + 481, + 616 + ], + "spans": [ + { + "bbox": [ + 132, + 558, + 481, + 616 + ], + "type": "text", + "content": "Success Rate. In the evaluation of success rate, we judge whether the concept " + }, + { + "bbox": [ + 132, + 558, + 481, + 616 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 132, + 558, + 481, + 616 + ], + "type": "text", + "content": " in the responsible image " + }, + { + "bbox": [ + 132, + 558, + 481, + 616 + ], + "type": "inline_equation", + "content": "x_{s}" + }, + { + "bbox": [ + 132, + 558, + 481, + 616 + ], + "type": "text", + "content": " still contains risks. For the safety subtask, the concept " + }, + { + "bbox": [ + 132, + 558, + 481, + 616 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 132, + 558, + 481, + 616 + ], + "type": "text", + "content": " should not appear; for the fairness subtask, the diversity of the concept " + }, + { + "bbox": [ + 132, + 558, + 481, + 616 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 132, + 558, + 481, + 616 + ], + "type": "text", + "content": " should be expanded; for the privacy subtask, we require that the identity of specific persons should not be recognized." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 617, + 481, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 617, + 481, + 664 + ], + "spans": [ + { + "bbox": [ + 132, + 617, + 481, + 664 + ], + "type": "text", + "content": "Visual Similarity. In the evaluation of visual similarity, we compute the pixel-level similarity between the responsible image " + }, + { + "bbox": [ + 132, + 617, + 481, + 664 + ], + "type": "inline_equation", + "content": "x_{\\mathrm{s}}" + }, + { + "bbox": [ + 132, + 617, + 481, + 664 + ], + "type": "text", + "content": " and the original risky image " + }, + { + "bbox": [ + 132, + 617, + 481, + 664 + ], + "type": "inline_equation", + "content": "x_{\\mathrm{r}}" + }, + { + "bbox": [ + 132, + 617, + 481, + 664 + ], + "type": "text", + "content": ". Specifically, if " + }, + { + "bbox": [ + 132, + 617, + 481, + 664 + ], + "type": "inline_equation", + "content": "x_{\\mathrm{s}}" + }, + { + "bbox": [ + 132, + 617, + 481, + 664 + ], + "type": "text", + "content": " still contains risks, then the similarity is considered as 0. We take the average similarity as the final result." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 215, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 215, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 215, + 100 + ], + "type": "text", + "content": "M. Ni et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 139 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 480, + 139 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 480, + 139 + ], + "type": "text", + "content": "More details of the evaluation metrics can be found in the Supplementary Materials." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 156, + 242, + 168 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 156, + 242, + 168 + ], + "spans": [ + { + "bbox": [ + 132, + 156, + 242, + 168 + ], + "type": "text", + "content": "3.4 Special Markers" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 174, + 479, + 234 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 174, + 479, + 234 + ], + "spans": [ + { + "bbox": [ + 130, + 174, + 479, + 234 + ], + "type": "text", + "content": "We add unique markers to the images to reduce potential risks in responsible visual editing. As shown in Figure 1, each image contains a marker in the lower right corner, indicating that this image is only used for responsible visual editing research and which subtask it belongs to. In addition, the highlighting symbol A or B represents the original or edited image." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 251, + 232, + 264 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 251, + 232, + 264 + ], + "spans": [ + { + "bbox": [ + 132, + 251, + 232, + 264 + ], + "type": "text", + "content": "4 Methodology" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 137, + 293, + 478, + 401 + ], + "blocks": [ + { + "bbox": [ + 137, + 293, + 478, + 401 + ], + "lines": [ + { + "bbox": [ + 137, + 293, + 478, + 401 + ], + "spans": [ + { + "bbox": [ + 137, + 293, + 478, + 401 + ], + "type": "image", + "image_path": "c1b654e72c8e66f1a79b9f6b179a40db30f9126ee1088ef37a0b2f0d027654ce.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 407, + 480, + 441 + ], + "lines": [ + { + "bbox": [ + 130, + 407, + 480, + 441 + ], + "spans": [ + { + "bbox": [ + 130, + 407, + 480, + 441 + ], + "type": "text", + "content": "Fig. 2: Overview of CoEditor. CoEditor consists of two stages of cognition: (1) a perceptional cognitive process (PCP) to understand what needs to be edited, and (2) a behavioral cognitive process (BCP) to plan how to edit." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 464, + 480, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 464, + 480, + 548 + ], + "spans": [ + { + "bbox": [ + 130, + 464, + 480, + 548 + ], + "type": "text", + "content": "Compared with traditional visual editing that usually edits specific objects or features in an image, responsible visual editing should be able to edit any risky concepts contained in the image. It could be a theme such as violence, a category such as culture, or a person like Bill Gates. This task is very challenging and it contains two stages: (1) understanding the relationship between the image content and the concept to locate the regions to be edited, and (2) planning how to edit the regions to meet the editing conditions and image rationality." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 548, + 480, + 608 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 548, + 480, + 608 + ], + "spans": [ + { + "bbox": [ + 130, + 548, + 480, + 608 + ], + "type": "text", + "content": "With the above considerations, we propose an LMM-based responsible editing model, namely Cognitive Editor (CoEditor). As shown in Figure 2, CoEditor consists of two stages of cognition: (1) perceptional cognitive process (PCP) to locate regions to be edited, and (2) behavioral cognitive process (BCP) to strategize how to edit such regions." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 623, + 317, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 623, + 317, + 635 + ], + "spans": [ + { + "bbox": [ + 132, + 623, + 317, + 635 + ], + "type": "text", + "content": "4.1 Perceptional Cognitive Process" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 641, + 480, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 480, + 665 + ], + "type": "text", + "content": "Understanding the relationship between the content of an image and abstract concepts is often very difficult, as the concepts may not directly refer to specific" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 336, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 336, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 336, + 91, + 447, + 102 + ], + "type": "text", + "content": "Responsible Visual Editing" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 176 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 176 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 176 + ], + "type": "text", + "content": "contents in the image and it requires reasoning based on common sense. Even the powerful LMM may not accurately understand the concepts (see Section 5.2). To mitigate these issues, we use a visual language chain-of-thoughts method based on visual prompts and language prompts for perceptional cognition to locate the regions to be edited." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 176, + 482, + 223 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 176, + 482, + 223 + ], + "spans": [ + { + "bbox": [ + 130, + 176, + 482, + 223 + ], + "type": "text", + "content": "For a risky image " + }, + { + "bbox": [ + 130, + 176, + 482, + 223 + ], + "type": "inline_equation", + "content": "x_{\\mathrm{r}}" + }, + { + "bbox": [ + 130, + 176, + 482, + 223 + ], + "type": "text", + "content": ", we need to visually annotate each element for LMM to refer to them accurately. We perform object extraction to obtain the object sequence " + }, + { + "bbox": [ + 130, + 176, + 482, + 223 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 130, + 176, + 482, + 223 + ], + "type": "text", + "content": ". In specific, we use Semantic-SAM [18] to extract objects of image " + }, + { + "bbox": [ + 130, + 176, + 482, + 223 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 130, + 176, + 482, + 223 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 130, + 176, + 482, + 223 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 130, + 176, + 482, + 223 + ], + "type": "text", + "content": " masks:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 253, + 224, + 481, + 237 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 253, + 224, + 481, + 237 + ], + "spans": [ + { + "bbox": [ + 253, + 224, + 481, + 237 + ], + "type": "interline_equation", + "content": "\\mathcal {M} = \\left\\{m _ {1}, m _ {2}, \\dots , m _ {n} \\right\\}. \\tag {1}", + "image_path": "e890ba7d5cf08b02e24f694706eed60d9c492ca68634f214c8fcc311c3299ef2.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 146, + 243, + 462, + 256 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 243, + 462, + 256 + ], + "spans": [ + { + "bbox": [ + 146, + 243, + 462, + 256 + ], + "type": "text", + "content": "Similar to SoM [33], we visually prompt the objects in the image to " + }, + { + "bbox": [ + 146, + 243, + 462, + 256 + ], + "type": "inline_equation", + "content": "v_{\\mathrm{p}}" + }, + { + "bbox": [ + 146, + 243, + 462, + 256 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 276, + 266, + 481, + 278 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 276, + 266, + 481, + 278 + ], + "spans": [ + { + "bbox": [ + 276, + 266, + 481, + 278 + ], + "type": "interline_equation", + "content": "v _ {\\mathrm {p}} = \\phi (x, \\mathcal {M}), \\tag {2}", + "image_path": "d7f92001fc224d1732d2d822c424bf0b92b0e9ab39bfff2fcf4dc531c686b64c.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 287, + 482, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 287, + 482, + 323 + ], + "spans": [ + { + "bbox": [ + 130, + 287, + 482, + 323 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 287, + 482, + 323 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 130, + 287, + 482, + 323 + ], + "type": "text", + "content": " is a prompting function that adds a visual tag with a number at the corner of each mask's corresponding location in the image, enabling the LMM to refer to it in numeric tags." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 323, + 482, + 359 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 323, + 482, + 359 + ], + "spans": [ + { + "bbox": [ + 130, + 323, + 482, + 359 + ], + "type": "text", + "content": "Meanwhile, in order to understand the possible meanings of the concept, we use knowledge extraction to expand its associated concepts and explanations into the text prompts " + }, + { + "bbox": [ + 130, + 323, + 482, + 359 + ], + "type": "inline_equation", + "content": "l_{\\mathrm{p}}" + }, + { + "bbox": [ + 130, + 323, + 482, + 359 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 276, + 360, + 481, + 373 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 276, + 360, + 481, + 373 + ], + "spans": [ + { + "bbox": [ + 276, + 360, + 481, + 373 + ], + "type": "interline_equation", + "content": "l _ {\\mathrm {p}} = f \\left(c; p _ {\\mathrm {k}} ^ {\\text {i n s}}\\right), \\tag {3}", + "image_path": "51d62a1ad4bf61e4fbcd7066f78baa0407cd7e4e65a06a1b3835556f8c63db8b.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 378, + 482, + 402 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 378, + 482, + 402 + ], + "spans": [ + { + "bbox": [ + 130, + 378, + 482, + 402 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 378, + 482, + 402 + ], + "type": "inline_equation", + "content": "p_{\\mathrm{k}}^{\\mathrm{ins}}" + }, + { + "bbox": [ + 130, + 378, + 482, + 402 + ], + "type": "text", + "content": " is the instruction for knowledge extraction. Since there is no need to introduce visual information here, " + }, + { + "bbox": [ + 130, + 378, + 482, + 402 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 130, + 378, + 482, + 402 + ], + "type": "text", + "content": " can be either an LLM or an LMM." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 403, + 482, + 440 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 403, + 482, + 440 + ], + "spans": [ + { + "bbox": [ + 130, + 403, + 482, + 440 + ], + "type": "text", + "content": "Combining prompts of different modalities " + }, + { + "bbox": [ + 130, + 403, + 482, + 440 + ], + "type": "inline_equation", + "content": "v_{\\mathrm{p}}" + }, + { + "bbox": [ + 130, + 403, + 482, + 440 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 403, + 482, + 440 + ], + "type": "inline_equation", + "content": "l_{\\mathrm{p}}" + }, + { + "bbox": [ + 130, + 403, + 482, + 440 + ], + "type": "text", + "content": ", the LMM can complete the cognition of the image and the concept, and obtain the mask of the region " + }, + { + "bbox": [ + 130, + 403, + 482, + 440 + ], + "type": "inline_equation", + "content": "m_{\\mathrm{p}}" + }, + { + "bbox": [ + 130, + 403, + 482, + 440 + ], + "type": "text", + "content": " to be modified in the image:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 264, + 448, + 481, + 462 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 264, + 448, + 481, + 462 + ], + "spans": [ + { + "bbox": [ + 264, + 448, + 481, + 462 + ], + "type": "interline_equation", + "content": "m _ {\\mathrm {p}} = f \\left(v _ {\\mathrm {p}}, l _ {\\mathrm {p}}; p _ {\\mathrm {p}} ^ {\\text {i n s}}\\right), \\tag {4}", + "image_path": "c3eb5b06b97acd8b80338b9e7a73f9d3b7cac141fd4bc3c3b6461c76375c5736.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 472, + 482, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 472, + 482, + 509 + ], + "spans": [ + { + "bbox": [ + 130, + 472, + 482, + 509 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 472, + 482, + 509 + ], + "type": "inline_equation", + "content": "p_{\\mathrm{p}}^{\\mathrm{ins}}" + }, + { + "bbox": [ + 130, + 472, + 482, + 509 + ], + "type": "text", + "content": " is the instruction for focus generation and " + }, + { + "bbox": [ + 130, + 472, + 482, + 509 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 130, + 472, + 482, + 509 + ], + "type": "text", + "content": " is an LMM. Though we may obtain multiple related regions, for simplicity, we combine all regions into one single " + }, + { + "bbox": [ + 130, + 472, + 482, + 509 + ], + "type": "inline_equation", + "content": "m_{\\mathrm{p}}" + }, + { + "bbox": [ + 130, + 472, + 482, + 509 + ], + "type": "text", + "content": " as the result of PCP." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 131, + 526, + 309, + 539 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 526, + 309, + 539 + ], + "spans": [ + { + "bbox": [ + 131, + 526, + 309, + 539 + ], + "type": "text", + "content": "4.2 Behavioral Cognitive Process" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 548, + 482, + 607 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 548, + 482, + 607 + ], + "spans": [ + { + "bbox": [ + 130, + 548, + 482, + 607 + ], + "type": "text", + "content": "In the perception cognitive process, the region " + }, + { + "bbox": [ + 130, + 548, + 482, + 607 + ], + "type": "inline_equation", + "content": "m_p" + }, + { + "bbox": [ + 130, + 548, + 482, + 607 + ], + "type": "text", + "content": " to be edited has been identified, and we need to figure out how to edit it. Since the concept may be abstract, the editing model is hard to directly perform editing operations (see Section 5.2). Therefore, we employ the visual language chain-of-thoughts to plan the editing target, i.e., what contents are expected after responsible editing." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 608, + 482, + 645 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 608, + 482, + 645 + ], + "spans": [ + { + "bbox": [ + 130, + 608, + 482, + 645 + ], + "type": "text", + "content": "We expand the object " + }, + { + "bbox": [ + 130, + 608, + 482, + 645 + ], + "type": "inline_equation", + "content": "m_p" + }, + { + "bbox": [ + 130, + 608, + 482, + 645 + ], + "type": "text", + "content": " identified in the previous stage to " + }, + { + "bbox": [ + 130, + 608, + 482, + 645 + ], + "type": "inline_equation", + "content": "m_p'" + }, + { + "bbox": [ + 130, + 608, + 482, + 645 + ], + "type": "text", + "content": " through object extension by enlarging the contour to include surrounding information. Then, we crop the image based on the mask " + }, + { + "bbox": [ + 130, + 608, + 482, + 645 + ], + "type": "inline_equation", + "content": "m_p'" + }, + { + "bbox": [ + 130, + 608, + 482, + 645 + ], + "type": "text", + "content": " to obtain visual prompting:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 276, + 653, + 481, + 667 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 276, + 653, + 481, + 667 + ], + "spans": [ + { + "bbox": [ + 276, + 653, + 481, + 667 + ], + "type": "interline_equation", + "content": "v _ {\\mathrm {b}} = x _ {\\mathrm {r}} \\otimes m _ {\\mathrm {p}} ^ {\\prime}. \\tag {5}", + "image_path": "d3a72d2be04085501b7dfc786f275187c6c4afc0afb4ac3da622565dc3d9850f.jpg" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 215, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 215, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 215, + 101 + ], + "type": "text", + "content": "M. Ni et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "text", + "content": "At the same time, we perform instruction implementation to get the full language prompt " + }, + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "inline_equation", + "content": "l_{\\mathrm{b}}" + }, + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "text", + "content": " by concatenating the instruction of the task " + }, + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "inline_equation", + "content": "p_{\\mathrm{t}}^{\\mathrm{ins}}" + }, + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "text", + "content": " and the input concept " + }, + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "text", + "content": ". Then we use the language prompt " + }, + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "inline_equation", + "content": "l_{\\mathrm{b}}" + }, + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "text", + "content": " together with the visual prompt " + }, + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "inline_equation", + "content": "v_{\\mathrm{b}}" + }, + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "text", + "content": " to generate the editing target " + }, + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "inline_equation", + "content": "r_{\\mathrm{b}}" + }, + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "text", + "content": " for the inpainting model:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 266, + 175, + 481, + 188 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 266, + 175, + 481, + 188 + ], + "spans": [ + { + "bbox": [ + 266, + 175, + 481, + 188 + ], + "type": "interline_equation", + "content": "r _ {\\mathrm {b}} = f \\left(v _ {\\mathrm {b}}, l _ {\\mathrm {b}}; p _ {\\mathrm {b}} ^ {\\text {i n s}}\\right), \\tag {6}", + "image_path": "ed6a31237fff60819015c94c6f6e9f73593740327e40f78dea6c095a16e0efe9.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 197, + 307, + 209 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 197, + 307, + 209 + ], + "spans": [ + { + "bbox": [ + 131, + 197, + 307, + 209 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 197, + 307, + 209 + ], + "type": "inline_equation", + "content": "p_{\\mathrm{b}}^{\\mathrm{ins}}" + }, + { + "bbox": [ + 131, + 197, + 307, + 209 + ], + "type": "text", + "content": " is the instruction for editing." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 209, + 479, + 244 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 209, + 479, + 244 + ], + "spans": [ + { + "bbox": [ + 130, + 209, + 479, + 244 + ], + "type": "text", + "content": "Finally, we take the original image " + }, + { + "bbox": [ + 130, + 209, + 479, + 244 + ], + "type": "inline_equation", + "content": "x_{\\mathrm{r}}" + }, + { + "bbox": [ + 130, + 209, + 479, + 244 + ], + "type": "text", + "content": ", the region to be edited " + }, + { + "bbox": [ + 130, + 209, + 479, + 244 + ], + "type": "inline_equation", + "content": "m_p'" + }, + { + "bbox": [ + 130, + 209, + 479, + 244 + ], + "type": "text", + "content": ", and the editing target " + }, + { + "bbox": [ + 130, + 209, + 479, + 244 + ], + "type": "inline_equation", + "content": "r_{\\mathrm{b}}" + }, + { + "bbox": [ + 130, + 209, + 479, + 244 + ], + "type": "text", + "content": " as the input to the inpainting model, and obtain the final editing result as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 268, + 246, + 481, + 259 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 268, + 246, + 481, + 259 + ], + "spans": [ + { + "bbox": [ + 268, + 246, + 481, + 259 + ], + "type": "interline_equation", + "content": "x _ {\\mathrm {s}} = g \\left(x _ {\\mathrm {r}}, m _ {\\mathrm {p}} ^ {\\prime}; r _ {\\mathrm {b}}\\right), \\tag {7}", + "image_path": "264d1c0ec3445dbbe3cd30e0d42bf22bad6deaa5d3c81d8c083faf7268cf076f.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 264, + 468, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 264, + 468, + 277 + ], + "spans": [ + { + "bbox": [ + 131, + 264, + 468, + 277 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 264, + 468, + 277 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 131, + 264, + 468, + 277 + ], + "type": "text", + "content": " is the inpainting model and " + }, + { + "bbox": [ + 131, + 264, + 468, + 277 + ], + "type": "inline_equation", + "content": "x_{\\mathrm{s}}" + }, + { + "bbox": [ + 131, + 264, + 468, + 277 + ], + "type": "text", + "content": " is the final responsible editing result." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 296, + 279, + 307 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 296, + 279, + 307 + ], + "spans": [ + { + "bbox": [ + 131, + 296, + 279, + 307 + ], + "type": "text", + "content": "4.3 Implementation Details" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 316, + 480, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 316, + 480, + 413 + ], + "spans": [ + { + "bbox": [ + 130, + 316, + 480, + 413 + ], + "type": "text", + "content": "Our method is training-free. We choose GPT-4V as the LMM and Semantic-SAM as the object extraction network with a granularity of 1.5 and its officially released checkpoint4. For the inpainting model, we choose Stable Diffusion Inpainting trained based on Stable Diffusion v2 and its publicly available checkpoint5. Like most previous works, we adjust the input image to a size of " + }, + { + "bbox": [ + 130, + 316, + 480, + 413 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 130, + 316, + 480, + 413 + ], + "type": "text", + "content": ". All random seeds are fixed to 42. In all experiments, the instructions " + }, + { + "bbox": [ + 130, + 316, + 480, + 413 + ], + "type": "inline_equation", + "content": "p_{\\mathrm{k}}^{\\mathrm{ins}}" + }, + { + "bbox": [ + 130, + 316, + 480, + 413 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 316, + 480, + 413 + ], + "type": "inline_equation", + "content": "p_{\\mathrm{p}}^{\\mathrm{ins}}" + }, + { + "bbox": [ + 130, + 316, + 480, + 413 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 316, + 480, + 413 + ], + "type": "inline_equation", + "content": "p_{\\mathrm{b}}^{\\mathrm{ins}}" + }, + { + "bbox": [ + 130, + 316, + 480, + 413 + ], + "type": "text", + "content": " are fixed, and " + }, + { + "bbox": [ + 130, + 316, + 480, + 413 + ], + "type": "inline_equation", + "content": "p_{\\mathrm{t}}^{\\mathrm{ins}}" + }, + { + "bbox": [ + 130, + 316, + 480, + 413 + ], + "type": "text", + "content": " depends on the task. Please refer to the Supplementary Materials for more details." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 434, + 224, + 447 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 434, + 224, + 447 + ], + "spans": [ + { + "bbox": [ + 131, + 434, + 224, + 447 + ], + "type": "text", + "content": "5 Experiment" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 131, + 460, + 253, + 472 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 460, + 253, + 472 + ], + "spans": [ + { + "bbox": [ + 131, + 460, + 253, + 472 + ], + "type": "text", + "content": "5.1 Experiment Setup" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 481, + 480, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 481, + 480, + 552 + ], + "spans": [ + { + "bbox": [ + 130, + 481, + 480, + 552 + ], + "type": "text", + "content": "Since the concept of responsible visual editing is arbitrary and it does not specify the editing area, we select two powerful image editing models, InstructPix2pix [4] and InstructDiffusion [11], in the experiments. They not only allow arbitrary language guidance with the help of LLM but also require no additional masks to indicate the editing region. In addition, we could use the same Stable Diffusion base model in our CoEditor, making our comparison fair." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 553, + 480, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 553, + 480, + 635 + ], + "spans": [ + { + "bbox": [ + 130, + 553, + 480, + 635 + ], + "type": "text", + "content": "We conduct experiments on the AltBear dataset. Since InstructPix2pix and InstructDiffusion do not support responsible editing of images based on the concept, we manually design editing conditions remove {concept} for safety and privacy tasks, and increase the variety of {concept} for fairness tasks. We evaluate the edited images in terms of both the success rate and visual similarity. We measure the results comprehensively via both automatic machine and manual human evaluations." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 336, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 336, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 336, + 91, + 447, + 102 + ], + "type": "text", + "content": "Responsible Visual Editing" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 133, + 642, + 345, + 654 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 642, + 345, + 654 + ], + "spans": [ + { + "bbox": [ + 133, + 642, + 345, + 654 + ], + "type": "text", + "content": "4 https://github.com/UX-Decoder/Semantic-SAM" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 135, + 654, + 447, + 666 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 654, + 447, + 666 + ], + "spans": [ + { + "bbox": [ + 135, + 654, + 447, + 666 + ], + "type": "text", + "content": "5 https://huggingface.co/stabilityai/stable-diffusion-2-inpainting" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 133, + 157, + 480, + 232 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "type": "text", + "content": "Table 1: Overall results on AltBear under machine evaluation. We can find that CoEditor significantly outperforms the baseline models in both success rate and visual similarity, validating the effectiveness of CoEditor." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 133, + 157, + 480, + 232 + ], + "lines": [ + { + "bbox": [ + 133, + 157, + 480, + 232 + ], + "spans": [ + { + "bbox": [ + 133, + 157, + 480, + 232 + ], + "type": "table", + "html": "
ModelSafetyFairnessPrivacyOverall
\\( Succ^{\\uparrow} \\)\\( Sim^{\\uparrow} \\)\\( Succ^{\\uparrow} \\)\\( Sim^{\\uparrow} \\)\\( Succ^{\\uparrow} \\)\\( Sim^{\\uparrow} \\)\\( Succ^{\\uparrow} \\)\\( Sim^{\\uparrow} \\)
InstructPix2pix21.31%0.196922.11%0.212351.22%0.456231.93%0.2924
InstructDiffusion40.54%0.357421.43%0.171966.33%0.496344.14%0.3497
CoEditor (Ours)70.13%0.565248.96%0.359578.35%0.639565.56%0.5188
", + "image_path": "760043f1fdc908a8a17f6a4bd3727a8a16395c71552393e3b367eda388505d8e.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 133, + 284, + 480, + 358 + ], + "blocks": [ + { + "bbox": [ + 132, + 242, + 481, + 275 + ], + "lines": [ + { + "bbox": [ + 132, + 242, + 481, + 275 + ], + "spans": [ + { + "bbox": [ + 132, + 242, + 481, + 275 + ], + "type": "text", + "content": "Table 2: Overall results on AltBear under human evaluation. In human evaluations, CoEditor shows consistent results with machine evaluations, showing that CoEditor is also effective from a human subjective perspective." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 133, + 284, + 480, + 358 + ], + "lines": [ + { + "bbox": [ + 133, + 284, + 480, + 358 + ], + "spans": [ + { + "bbox": [ + 133, + 284, + 480, + 358 + ], + "type": "table", + "html": "
ModelSafetyFairnessPrivacyOverall
\\( Succ^{\\uparrow} \\)\\( Sim^{\\uparrow} \\)\\( Succ^{\\uparrow} \\)\\( Sim^{\\uparrow} \\)\\( Succ^{\\uparrow} \\)\\( Sim^{\\uparrow} \\)\\( Succ^{\\uparrow} \\)\\( Sim^{\\uparrow} \\)
InstructPix2pix28.41%0.25239.38%0.083940.00%0.327126.06%0.2217
InstructDiffusion40.91%0.307819.79%0.136956.00%0.345739.08%0.2634
CoEditor (Ours)65.91%0.531850.00%0.360665.00%0.518360.21%0.4692
", + "image_path": "5a8a8c0621bd6de0f9672c68a2eb11b040d7917be65e39ab321a0b1880781bf9.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 381, + 238, + 392 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 381, + 238, + 392 + ], + "spans": [ + { + "bbox": [ + 132, + 381, + 238, + 392 + ], + "type": "text", + "content": "5.2 Overall Results" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 402, + 481, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 402, + 481, + 521 + ], + "spans": [ + { + "bbox": [ + 130, + 402, + 481, + 521 + ], + "type": "text", + "content": "Quantitative Results. As shown in Table 1 and Table 2, CoEditor shows significant advantages in both machine and human evaluations. We can see that for the success rate, CoEditor has more than " + }, + { + "bbox": [ + 130, + 402, + 481, + 521 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 130, + 402, + 481, + 521 + ], + "type": "text", + "content": " increase in almost all subtasks. This proves that for responsible visual synthesis, using LMM to understand images and concepts and to strategize editing process is crucial. In addition, we find that CoEditor has higher visual similarity compared to traditional models, which reveals the importance of explicit cognition process of editing. It is noticed that the machine evaluation shows similar results to human evaluations, further proving that our editing results are of the same high quality from the perspective of human judgements." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "content": "Qualitative Results. As shown in Figure 3, CoEditor demonstrates strong understanding and editing capabilities in responsible visual synthesis. In the first row of the safety task, we see that the CoEditor not only fully erases the concepts of drugs and alcohol but also maintains the rationality of the picture. However, InstructDiffusion fails to completely remove drugs and generates unreasonable contents in the picture. Moreover, InstructPix2pix does not follow the instruction well. In the second row of the fairness task, CoEditor successfully diversifies the specified concept. It maintains the original layout, while InstructDiffusion results in blurry or damaged outputs and InstructPix2pix has difficulties in executing this instruction. In the third row of the privacy task, the CoEditor blurs the characters' features while maintaining the original meaning with high visual quality. However, InstructDiffusion and InstructPix2pix fail in content removal" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 215, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 215, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 215, + 101 + ], + "type": "text", + "content": "M. Ni et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 230, + 157, + 383, + 215 + ], + "blocks": [ + { + "bbox": [ + 131, + 114, + 480, + 148 + ], + "lines": [ + { + "bbox": [ + 131, + 114, + 480, + 148 + ], + "spans": [ + { + "bbox": [ + 131, + 114, + 480, + 148 + ], + "type": "text", + "content": "Table 3: Ablation study of AltBear. Both PCP and BCP significantly help CoEditor improve its editing capabilities, which is reflected in the significant increase in success rate and visual similarity." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 230, + 157, + 383, + 215 + ], + "lines": [ + { + "bbox": [ + 230, + 157, + 383, + 215 + ], + "spans": [ + { + "bbox": [ + 230, + 157, + 383, + 215 + ], + "type": "table", + "html": "
ModelSucc↑Sim↑
CoEditor w/o PCP44.19%0.3855
CoEditor w/o BCP28.02%0.2221
CoEditor (Full)65.56%0.5188
", + "image_path": "1a0d8c82ce0418a3d7765832643c2d0fbb3f9e0afe1ddc0f5db6e06a846c1929.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 236, + 479, + 271 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 236, + 479, + 271 + ], + "spans": [ + { + "bbox": [ + 130, + 236, + 479, + 271 + ], + "type": "text", + "content": "or damage the picture. These examples show that CoEditor can understand the images and concepts and then plan the editing well with the help of LMM. More examples can be found in the Supplementary Materials." + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 138, + 291, + 304, + 448 + ], + "blocks": [ + { + "bbox": [ + 138, + 291, + 304, + 448 + ], + "lines": [ + { + "bbox": [ + 138, + 291, + 304, + 448 + ], + "spans": [ + { + "bbox": [ + 138, + 291, + 304, + 448 + ], + "type": "image", + "image_path": "8a8b36668cce6de5a4a76205fe9444d40f78846f541495ae8b8739c6ef228754.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 458, + 480, + 502 + ], + "lines": [ + { + "bbox": [ + 130, + 458, + 480, + 502 + ], + "spans": [ + { + "bbox": [ + 130, + 458, + 480, + 502 + ], + "type": "text", + "content": "Fig. 3: Visualization results on AltBear. CoEditor performs well in all subtasks and maintains high visual similarity. The results of CoEditor also have better rationality. InstructDiffusion often over-edits images or produces unreasonable visual effects, while the editing ability of InstructPix2pix is weaker compared with CoEditor." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 309, + 291, + 477, + 449 + ], + "blocks": [ + { + "bbox": [ + 309, + 291, + 477, + 449 + ], + "lines": [ + { + "bbox": [ + 309, + 291, + 477, + 449 + ], + "spans": [ + { + "bbox": [ + 309, + 291, + 477, + 449 + ], + "type": "image", + "image_path": "fd46e533b2de3b84823e60c30a54b2d807a5ffaa50b7d356f9933ce94a9f1ddf.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 540, + 238, + 553 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 540, + 238, + 553 + ], + "spans": [ + { + "bbox": [ + 131, + 540, + 238, + 553 + ], + "type": "text", + "content": "5.3 Ablation Study" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "text", + "content": "Quantitative Results. We conduct ablation experiments on AltBear in Table 3 to explore the effectiveness of the two stages of CoEditor. In the ablation of PCP, since the model cannot locate the area to be edited, we turn to use the largest area in the image. In the ablation of BCP, we use the same instruction as the baseline model for editing. To ensure fairness, all other parameters of the experiments are consistent. As shown in Table 3, both stages of CoEditor are significant for the final results. Without PCP, the CoEditor cannot understand the object to be edited, causing a massive drop in success rate and visual similarity. Without BCP, although the CoEditor can locate the region to be edited," + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 336, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 336, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 336, + 91, + 447, + 102 + ], + "type": "text", + "content": "Responsible Visual Editing" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 480, + 140 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 480, + 140 + ], + "type": "text", + "content": "simply asking the model to remove or increase diversity will make it difficult to understand what to edit, ultimately leading to editing failures." + } + ] + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 137, + 159, + 478, + 319 + ], + "blocks": [ + { + "bbox": [ + 137, + 159, + 478, + 319 + ], + "lines": [ + { + "bbox": [ + 137, + 159, + 478, + 319 + ], + "spans": [ + { + "bbox": [ + 137, + 159, + 478, + 319 + ], + "type": "image", + "image_path": "691c10985e6c77731abc03c65c572094d0f3ea77b524afb136e5ac6e6e62b39c.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 329, + 482, + 374 + ], + "lines": [ + { + "bbox": [ + 131, + 329, + 482, + 374 + ], + "spans": [ + { + "bbox": [ + 131, + 329, + 482, + 374 + ], + "type": "text", + "content": "Fig. 4: Roles of different components in CoEditor. Without the help of BCP, the CoEditor produces inconsistent results with the concept or visually unreasonable results because it does not know how to edit the content correctly. Without PCP, CoEditor is unable to locate the editing regions, resulting in editing failures." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 396, + 482, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 396, + 482, + 552 + ], + "spans": [ + { + "bbox": [ + 130, + 396, + 482, + 552 + ], + "type": "text", + "content": "Qualitative Results. To investigate whether the two stages of CoEditor, BCP and PCP, can produce the expected outputs, we visualize some examples in Figure 3. One can see that the BCP stage is crucial for correct editing. In most examples, without the BCP stage, the editing result will fail to meet the requirements, or there will be visually unreasonable contents. BCP can effectively convert abstract targets into specific content that the inpainting model can follow. This shows the importance of editing planning. At the same time, PCP plays a decisive role. Without the participation of PCP, most of the edits will fail. This is because most of the concepts to be edited do not directly associate with a specific object in the image. Therefore, a deep understanding of image and concepts using PCP is crucial for CoEditor to edit correctly. The CoEditor can achieve satisfactory effects when both the two modules are used. Extra ablation studies can be found in the Supplementary Materials." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 565, + 326, + 578 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 565, + 326, + 578 + ], + "spans": [ + { + "bbox": [ + 131, + 565, + 326, + 578 + ], + "type": "text", + "content": "5.4 Exploration of Cognitive Process" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "content": "To explore why CoEditor can successfully perform complex and responsible visual editing, we show the intermediate results of various cognitive processes in Figure 5. We can see that PCP can successfully locate regions to be edited, regardless of whether this region is directly related to the concept. Based on the task type, PCP can also accurately find out the region to be edited among multiple identical objects. At the same time, BCP can successfully plan the editing. No matter what the concept is, BCP can accurately associate it with the" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 215, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 215, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 215, + 100 + ], + "type": "text", + "content": "M. Ni et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 137, + 126, + 198, + 186 + ], + "blocks": [ + { + "bbox": [ + 145, + 115, + 192, + 125 + ], + "lines": [ + { + "bbox": [ + 145, + 115, + 192, + 125 + ], + "spans": [ + { + "bbox": [ + 145, + 115, + 192, + 125 + ], + "type": "text", + "content": "Risky Image" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 137, + 126, + 198, + 186 + ], + "lines": [ + { + "bbox": [ + 137, + 126, + 198, + 186 + ], + "spans": [ + { + "bbox": [ + 137, + 126, + 198, + 186 + ], + "type": "image", + "image_path": "db3085b14647fcde9c7a4d6fc3b6049fbe1d9a3ea9f8a86547a2519f2c09c7b0.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 216, + 125, + 277, + 186 + ], + "blocks": [ + { + "bbox": [ + 225, + 115, + 268, + 124 + ], + "lines": [ + { + "bbox": [ + 225, + 115, + 268, + 124 + ], + "spans": [ + { + "bbox": [ + 225, + 115, + 268, + 124 + ], + "type": "text", + "content": "PCP Result" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 216, + 125, + 277, + 186 + ], + "lines": [ + { + "bbox": [ + 216, + 125, + 277, + 186 + ], + "spans": [ + { + "bbox": [ + 216, + 125, + 277, + 186 + ], + "type": "image", + "image_path": "c92dc428fd3f738db7567cf43db23e113bcad4f89e0af9d60fd319865808ba20.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 323, + 115, + 366, + 124 + ], + "lines": [ + { + "bbox": [ + 323, + 115, + 366, + 124 + ], + "spans": [ + { + "bbox": [ + 323, + 115, + 366, + 124 + ], + "type": "text", + "content": "BCP Result" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 298, + 132, + 400, + 181 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 132, + 400, + 181 + ], + "spans": [ + { + "bbox": [ + 298, + 132, + 400, + 181 + ], + "type": "text", + "content": "A cuddly teddy bear enjoying a sunny day on a serene picnic with a picturesque house fully intact in the background, surrounded by vibrant greenery." + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 416, + 125, + 477, + 186 + ], + "blocks": [ + { + "bbox": [ + 423, + 115, + 469, + 124 + ], + "lines": [ + { + "bbox": [ + 423, + 115, + 469, + 124 + ], + "spans": [ + { + "bbox": [ + 423, + 115, + 469, + 124 + ], + "type": "text", + "content": "Final Result" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 416, + 125, + 477, + 186 + ], + "lines": [ + { + "bbox": [ + 416, + 125, + 477, + 186 + ], + "spans": [ + { + "bbox": [ + 416, + 125, + 477, + 186 + ], + "type": "image", + "image_path": "c494380b763d0d6f31b8a24361f8473f21743d2f5e56333332931f2ceb6c495e.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 137, + 200, + 198, + 261 + ], + "blocks": [ + { + "bbox": [ + 139, + 190, + 194, + 199 + ], + "lines": [ + { + "bbox": [ + 139, + 190, + 194, + 199 + ], + "spans": [ + { + "bbox": [ + 139, + 190, + 194, + 199 + ], + "type": "text", + "content": "Concept: arsion" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 137, + 200, + 198, + 261 + ], + "lines": [ + { + "bbox": [ + 137, + 200, + 198, + 261 + ], + "spans": [ + { + "bbox": [ + 137, + 200, + 198, + 261 + ], + "type": "image", + "image_path": "563302e456afc9aa13a74b271c3081f507bece9347aa64d7695be69d243c054d.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 216, + 200, + 277, + 261 + ], + "blocks": [ + { + "bbox": [ + 216, + 200, + 277, + 261 + ], + "lines": [ + { + "bbox": [ + 216, + 200, + 277, + 261 + ], + "spans": [ + { + "bbox": [ + 216, + 200, + 277, + 261 + ], + "type": "image", + "image_path": "2eb50c04f9e024962d4bf40df3aef89ba106c76425d5ad6b028046e26c767b73.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 298, + 197, + 400, + 263 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 197, + 400, + 263 + ], + "spans": [ + { + "bbox": [ + 298, + 197, + 400, + 263 + ], + "type": "text", + "content": "A teddy bear wearing traditional attire representing different cultures, such as a Scottish kilt, an Indian sari, and a Nigerian abada, showcasing the beautiful variety of cultural dress from around the world." + } + ] + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 416, + 200, + 477, + 261 + ], + "blocks": [ + { + "bbox": [ + 416, + 200, + 477, + 261 + ], + "lines": [ + { + "bbox": [ + 416, + 200, + 477, + 261 + ], + "spans": [ + { + "bbox": [ + 416, + 200, + 477, + 261 + ], + "type": "image", + "image_path": "5c15a380009225a6787d2b05db301bb943440941b0cdd31797e053e7ca97f831.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 137, + 274, + 198, + 335 + ], + "blocks": [ + { + "bbox": [ + 139, + 264, + 200, + 274 + ], + "lines": [ + { + "bbox": [ + 139, + 264, + 200, + 274 + ], + "spans": [ + { + "bbox": [ + 139, + 264, + 200, + 274 + ], + "type": "text", + "content": "Concept: culture" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 137, + 274, + 198, + 335 + ], + "lines": [ + { + "bbox": [ + 137, + 274, + 198, + 335 + ], + "spans": [ + { + "bbox": [ + 137, + 274, + 198, + 335 + ], + "type": "image", + "image_path": "93bf4ed5f587f86ef9980443d977ce2a3a371fccce794ea7e59b430b3c568128.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 357, + 482, + 413 + ], + "lines": [ + { + "bbox": [ + 130, + 357, + 482, + 413 + ], + "spans": [ + { + "bbox": [ + 130, + 357, + 482, + 413 + ], + "type": "text", + "content": "Fig. 5: Intermediate results of cognitive process in CoEditor. We can see that PCP can successfully find one or more objects to be edited, even if the object is not directly related to the concept. BCP can generate effective editing target based on the positioning of PCP, even if the scenario is very complex. The effectiveness of the two cognitive processes allows CoEditor to edit images successfully." + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 216, + 274, + 277, + 335 + ], + "blocks": [ + { + "bbox": [ + 216, + 274, + 277, + 335 + ], + "lines": [ + { + "bbox": [ + 216, + 274, + 277, + 335 + ], + "spans": [ + { + "bbox": [ + 216, + 274, + 277, + 335 + ], + "type": "image", + "image_path": "dc7cd48e7c1584d8a1019ec5f7bf844624b792aae4dcaaef1f4d698737451c84.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "bbox": [ + 298, + 281, + 400, + 328 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 281, + 400, + 328 + ], + "spans": [ + { + "bbox": [ + 298, + 281, + 400, + 328 + ], + "type": "text", + "content": "A fluffy teddy bear with sparkling eyes and shiny footwear stands in front of a vintage microphone on a soft glowing background." + } + ] + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 416, + 274, + 477, + 335 + ], + "blocks": [ + { + "bbox": [ + 416, + 274, + 477, + 335 + ], + "lines": [ + { + "bbox": [ + 416, + 274, + 477, + 335 + ], + "spans": [ + { + "bbox": [ + 416, + 274, + 477, + 335 + ], + "type": "image", + "image_path": "59fe2d45b8982f14e836b50c8c62d209f2487c1eb8d506e18d25b6f88e0892f9.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "bbox": [ + 139, + 338, + 228, + 347 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 338, + 228, + 347 + ], + "spans": [ + { + "bbox": [ + 139, + 338, + 228, + 347 + ], + "type": "text", + "content": "Concept: Marilyn Monroe" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 130, + 438, + 482, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 438, + 482, + 486 + ], + "spans": [ + { + "bbox": [ + 130, + 438, + 482, + 486 + ], + "type": "text", + "content": "content in the image to obtain editing target consistent with the image scene. This shows the importance of cognitive processes to CoEditor in concept understanding and editing planning. For further explorations of CoEditor, please see the Supplementary Materials." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 131, + 509, + 457, + 521 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 509, + 457, + 521 + ], + "spans": [ + { + "bbox": [ + 131, + 509, + 457, + 521 + ], + "type": "text", + "content": "5.5 Performance Consistency on AltBear and Real-world Data" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": "To verify whether AltBear can serve as a replacement of real-world images in subsequent research, we examine the performance consistency between AltBear and real-world images. We build a dataset that has similar size to AltBear, consisting of real-world images collected from the Internet. We conduct experiments on this dataset with all hyper-parameters and models identical to that on AltBear. As shown in Table 4 and 5, CoEditor achieves similar performance to that on AltBear, demonstrating CoEditor's robustness to both synthesized data and real-world data. The baseline models, InstructPix2pix and InstructDiffusion, also show similar performance on AltBear and real-world data. Since the protagonist is not a human but a fictional teddy bear, AltBear significantly reduces the risk of image propagation and display, effectively avoiding ethical risks." + } + ] + } + ], + "index": 24 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 336, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 336, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 336, + 91, + 447, + 102 + ], + "type": "text", + "content": "Responsible Visual Editing" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 133, + 167, + 480, + 243 + ], + "blocks": [ + { + "bbox": [ + 130, + 114, + 482, + 160 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 482, + 160 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 482, + 160 + ], + "type": "text", + "content": "Table 4: Results on real-world images under machine evaluation. On real-world datasets, CoEditor shows similar performance to that on AltBear, significantly surpassing the baseline model. Furthermore, we find that the distributions of both success rate and visual similarity are close for real-world images and AltBear." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 133, + 167, + 480, + 243 + ], + "lines": [ + { + "bbox": [ + 133, + 167, + 480, + 243 + ], + "spans": [ + { + "bbox": [ + 133, + 167, + 480, + 243 + ], + "type": "table", + "html": "
ModelSafetyFairnessPrivacyOverall
\\( Succ^{\\uparrow} \\)\\( Sim^{\\uparrow} \\)\\( Succ^{\\uparrow} \\)\\( Sim^{\\uparrow} \\)\\( Succ^{\\uparrow} \\)\\( Sim^{\\uparrow} \\)\\( Succ^{\\uparrow} \\)\\( Sim^{\\uparrow} \\)
InstructPix2pix14.00%0.12831.89%0.016622.86%0.216411.59%0.1078
InstructDiffusion38.60%0.318514.29%0.103055.81%0.374435.57%0.2638
CoEditor (Ours)59.46%0.474058.18%0.425363.41%0.513960.00%0.4679
", + "image_path": "ac48f1fe20236b5aea1eaac6324bdf8f8c5554e35da99d20c5f94c5d855ba38b.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 133, + 306, + 480, + 380 + ], + "blocks": [ + { + "bbox": [ + 130, + 252, + 482, + 297 + ], + "lines": [ + { + "bbox": [ + 130, + 252, + 482, + 297 + ], + "spans": [ + { + "bbox": [ + 130, + 252, + 482, + 297 + ], + "type": "text", + "content": "Table 5: Results of real-world images under human evaluation. CoEditor achieves significantly better results than the baseline model in human evaluation. In addition, the quantitative results on AltBear and real-world images in human evaluation are also similar." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 133, + 306, + 480, + 380 + ], + "lines": [ + { + "bbox": [ + 133, + 306, + 480, + 380 + ], + "spans": [ + { + "bbox": [ + 133, + 306, + 480, + 380 + ], + "type": "table", + "html": "
ModelSafetyFairnessPrivacyOverall
\\( Succ^{\\uparrow} \\)\\( Sim^{\\uparrow} \\)\\( Succ^{\\uparrow} \\)\\( Sim^{\\uparrow} \\)\\( Succ^{\\uparrow} \\)\\( Sim^{\\uparrow} \\)\\( Succ^{\\uparrow} \\)\\( Sim^{\\uparrow} \\)
InstructPix2pix31.25%0.27989.09%0.080415.56%0.132720.56%0.1821
InstructDiffusion47.50%0.376421.82%0.152453.33%0.337841.11%0.2983
CoEditor (Ours)68.75%0.548665.45%0.478780.00%0.646770.56%0.5518
", + "image_path": "8f2bfb54d899ffbd2138cd666bf8297a18dd9b602009785019b95660bba86991.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 403, + 257, + 414 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 403, + 257, + 414 + ], + "spans": [ + { + "bbox": [ + 132, + 403, + 257, + 414 + ], + "type": "text", + "content": "5.6 Real-world Results" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 423, + 482, + 591 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 423, + 482, + 591 + ], + "spans": [ + { + "bbox": [ + 130, + 423, + 482, + 591 + ], + "type": "text", + "content": "To better demonstrate the robustness of CoEditor, we show its performance on real-world images and compare with baseline models in Figure 6. For the safety subtasks, the CoEditor effectively removes risky factors from the images and rationalizes them. In the first row, the robber in the first image is turned into an enthusiastic saleswoman, and the craving for drugs in the second image is turned into a craving for food. This also shows the imagination capability of CoEditor. For the fairness subtasks in the second row, the CoEditor completes the image editing while maintaining the visual quality of the image. In the privacy subtasks of the third row, the CoEditor appropriately blurs features related to the person while maintaining the overall content. Similar to the results on AltBear, the baseline models, InstructPix2pix and InstructDiffusion, struggle to maintain visual quality or rationality in some examples, even though they apply editing. In other cases, they produce damaged figures. More real-world examples can be found in the Supplementary Materials." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 609, + 347, + 621 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 609, + 347, + 621 + ], + "spans": [ + { + "bbox": [ + 131, + 609, + 347, + 621 + ], + "type": "text", + "content": "5.7 Analysis of Computational Overhead" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "type": "text", + "content": "In order to explore how CoEditor consumes resources, we make comparisons on time and memory usage. We use the same full-precision and optimization settings, e.g., memory-efficient attention for all models." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 215, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 215, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 215, + 100 + ], + "type": "text", + "content": "M. Ni et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 137, + 114, + 304, + 166 + ], + "blocks": [ + { + "bbox": [ + 137, + 114, + 304, + 166 + ], + "lines": [ + { + "bbox": [ + 137, + 114, + 304, + 166 + ], + "spans": [ + { + "bbox": [ + 137, + 114, + 304, + 166 + ], + "type": "image", + "image_path": "eab02e419db666486b28844e909dfeababe86416da3515e3dc15050be8fba9a8.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 137, + 167, + 304, + 215 + ], + "blocks": [ + { + "bbox": [ + 137, + 167, + 304, + 215 + ], + "lines": [ + { + "bbox": [ + 137, + 167, + 304, + 215 + ], + "spans": [ + { + "bbox": [ + 137, + 167, + 304, + 215 + ], + "type": "image", + "image_path": "0035c86a07e4a4778abf6e7249dbe49ee7880418dfeef285b95760d837e427f0.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 137, + 215, + 304, + 272 + ], + "blocks": [ + { + "bbox": [ + 137, + 215, + 304, + 272 + ], + "lines": [ + { + "bbox": [ + 137, + 215, + 304, + 272 + ], + "spans": [ + { + "bbox": [ + 137, + 215, + 304, + 272 + ], + "type": "image", + "image_path": "98cf5188fec5ab838a9bc9de29ed598134f15c9cc1c9bafae8900433a18492d6.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 281, + 480, + 326 + ], + "lines": [ + { + "bbox": [ + 130, + 281, + 480, + 326 + ], + "spans": [ + { + "bbox": [ + 130, + 281, + 480, + 326 + ], + "type": "text", + "content": "Fig. 6: Visualized results on real-world data. Even in real-world scenarios, CoEditor can still work well. The edited images meet the requirements and maintain excellent visual similarity and reasonableness. Similar to the results on AltBear, InstructDiffusion and InstructPix2pix fail to output well edited images." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 308, + 114, + 477, + 165 + ], + "blocks": [ + { + "bbox": [ + 308, + 114, + 477, + 165 + ], + "lines": [ + { + "bbox": [ + 308, + 114, + 477, + 165 + ], + "spans": [ + { + "bbox": [ + 308, + 114, + 477, + 165 + ], + "type": "image", + "image_path": "a917dc032a74ac908fe87c411ad6916173c4352d7ba6ed4f075b40d5289ef618.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 309, + 165, + 477, + 215 + ], + "blocks": [ + { + "bbox": [ + 309, + 165, + 477, + 215 + ], + "lines": [ + { + "bbox": [ + 309, + 165, + 477, + 215 + ], + "spans": [ + { + "bbox": [ + 309, + 165, + 477, + 215 + ], + "type": "image", + "image_path": "94971041a4b6f0fcd983f4dae36158d328646b35ec0696d7d01c5c8b769dedb6.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 309, + 215, + 477, + 272 + ], + "blocks": [ + { + "bbox": [ + 309, + 215, + 477, + 272 + ], + "lines": [ + { + "bbox": [ + 309, + 215, + 477, + 272 + ], + "spans": [ + { + "bbox": [ + 309, + 215, + 477, + 272 + ], + "type": "image", + "image_path": "a3fd1d755e3adb7d2268ce871babff6d86375c6dd9216e7d0016025b1b380dcf.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 227, + 358, + 386, + 437 + ], + "blocks": [ + { + "bbox": [ + 213, + 337, + 399, + 348 + ], + "lines": [ + { + "bbox": [ + 213, + 337, + 399, + 348 + ], + "spans": [ + { + "bbox": [ + 213, + 337, + 399, + 348 + ], + "type": "text", + "content": "Table 6: Comparison on inference time." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 227, + 358, + 386, + 437 + ], + "lines": [ + { + "bbox": [ + 227, + 358, + 386, + 437 + ], + "spans": [ + { + "bbox": [ + 227, + 358, + 386, + 437 + ], + "type": "table", + "html": "
ModelTimeMultiple
InstrcutPix2pix12.43s1×
InstructDiffusion20.75s1.67×
CoEditor14.10s1.13×
PCP Stage4.65s-
BCP Stage9.44s-
", + "image_path": "dc0def3fe284c6d33a0d4961a07921de22881514a9d7377309e0b086ceacb680.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 459, + 481, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 459, + 481, + 506 + ], + "spans": [ + { + "bbox": [ + 130, + 459, + 481, + 506 + ], + "type": "text", + "content": "Time. We calculate the average inference time of different models. For CoEditor, we also calculate the network latency. As shown in Table 6, the speed of CoEditor is even faster than InstructDiffusion [11], and is comparable to Instruct Pix2pix [4]." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 507, + 480, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 507, + 480, + 555 + ], + "spans": [ + { + "bbox": [ + 130, + 507, + 480, + 555 + ], + "type": "text", + "content": "GPU Memory. We calculate the peak VRAM usage of different models. As shown in Table 7, CoEditor consumes the least memory. This is because our chosen models, i.e., Semantic-SAM [18] and Stable Diffusion Inpainting [29], are very lightweight, which allows our model to work with minimal memory." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 131, + 573, + 257, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 573, + 257, + 586 + ], + "spans": [ + { + "bbox": [ + 131, + 573, + 257, + 586 + ], + "type": "text", + "content": "6 Ethics Statement" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 131, + 597, + 421, + 610 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 597, + 421, + 610 + ], + "spans": [ + { + "bbox": [ + 131, + 597, + 421, + 610 + ], + "type": "text", + "content": "6.1 Inappropriate Content, Privacy, and Discrimination" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 617, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 481, + 665 + ], + "type": "text", + "content": "The AltBear dataset uses teddy bears as the protagonist, significantly reducing the risks caused by the public display of inappropriate contents. We manually reviewed all contents to ensure that privacy or discrimination is avoided. We release AltBear under the MIT license to keep transparency." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 336, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 336, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 336, + 91, + 447, + 102 + ], + "type": "text", + "content": "Responsible Visual Editing" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 219, + 135, + 395, + 193 + ], + "blocks": [ + { + "bbox": [ + 214, + 114, + 399, + 126 + ], + "lines": [ + { + "bbox": [ + 214, + 114, + 399, + 126 + ], + "spans": [ + { + "bbox": [ + 214, + 114, + 399, + 126 + ], + "type": "text", + "content": "Table 7: Comparison on GPU memory." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 219, + 135, + 395, + 193 + ], + "lines": [ + { + "bbox": [ + 219, + 135, + 395, + 193 + ], + "spans": [ + { + "bbox": [ + 219, + 135, + 395, + 193 + ], + "type": "table", + "html": "
ModelVRAMMultiple
InstrcutPix2pix18,118MB1×
InstructDiffusion8,378MB0.46×
CoEditor7,328MB0.40×
", + "image_path": "ecbf94d55d32df100c20f727e1ac7a5eb61e4bdd971fc682067d6d78d1532fbe.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 215, + 239, + 228 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 215, + 239, + 228 + ], + "spans": [ + { + "bbox": [ + 132, + 215, + 239, + 228 + ], + "type": "text", + "content": "6.2 Reproducibility" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 236, + 481, + 297 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 236, + 481, + 297 + ], + "spans": [ + { + "bbox": [ + 130, + 236, + 481, + 297 + ], + "type": "text", + "content": "We build our method on top of the Stable Diffusion Inpainting model with publicly accessible code and checkpoints with a fixed random seed. We notice that the GPT API cannot guarantee identical responses so that we provide all instruction prompts for reference. To further improve the reproducibility, we also release the dataset, code and our trained model." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 314, + 221, + 326 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 314, + 221, + 326 + ], + "spans": [ + { + "bbox": [ + 132, + 314, + 221, + 326 + ], + "type": "text", + "content": "6.3 Anti-misuse" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 334, + 481, + 393 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 334, + 481, + 393 + ], + "spans": [ + { + "bbox": [ + 130, + 334, + 481, + 393 + ], + "type": "text", + "content": "In order to ensure that the dataset is not misused, we have designed unique markers (refer to Section 3.4). We added the marker to the bottom right corner of the image to indicate that the image is used for responsible visual editing research only. Specifically, we have defined various tags to indicate whether the image has been edited or contains potential risks." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 394, + 481, + 442 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 394, + 481, + 442 + ], + "spans": [ + { + "bbox": [ + 130, + 394, + 481, + 442 + ], + "type": "text", + "content": "In addition, the interaction system of our model always maintains visibility of the concepts to be edited to prevent misuse of the model. We also call on subsequent work to maintain the visibility of concepts like ours, ensuring a beneficial role for the community through open and transparent methods." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 460, + 220, + 473 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 460, + 220, + 473 + ], + "spans": [ + { + "bbox": [ + 132, + 460, + 220, + 473 + ], + "type": "text", + "content": "7 Conclusion" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "type": "text", + "content": "In this work, we proposed a new task, namely responsible visual editing, which entails editing specific concepts within an image to render it more responsible while minimizing changes. To tackle abstract concepts in responsible visual editing, we proposed a Cognitive Editor (CoEditor) by harnessing a large multimodal model (LMM) through a two-stage process: a perceptual cognitive process to locate what needs to be edited, and a behavioral cognitive process to strategize how to edit. We constructed a transparent and public dataset, AltBear, which represents harmful information using teddy bears instead of humans, thereby mitigating the negative implications on research. Experiments demonstrated that CoEditor can effectively comprehend abstract concepts within complex scenes and significantly surpass the performance of baseline models for responsible visual editing. Moreover, we validated that the AltBear dataset can well serve as a replacement of real-world images in the research of responsible visual editing, providing a safer benchmark. Our findings also revealed the potential of LMM in responsible AI." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 215, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 215, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 215, + 100 + ], + "type": "text", + "content": "M. Ni et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 240, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 240, + 129 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 240, + 129 + ], + "type": "text", + "content": "Acknowledgement" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 137, + 489, + 174 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 137, + 489, + 174 + ], + "spans": [ + { + "bbox": [ + 132, + 137, + 489, + 174 + ], + "type": "text", + "content": "This work was supported in part by the National Key R&D Program of China (2022YFA1004100), and the National Natural Science Foundation of China (NSFC) under grant No. 62441202." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 190, + 197, + 202 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 190, + 197, + 202 + ], + "spans": [ + { + "bbox": [ + 133, + 190, + 197, + 202 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 212, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 138, + 212, + 481, + 245 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 212, + 481, + 245 + ], + "spans": [ + { + "bbox": [ + 138, + 212, + 481, + 245 + ], + "type": "text", + "content": "1. Abdal, R., Qin, Y., Wonka, P.: Image2stylegan++: How to edit the embedded images? In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 8296-8305 (2020)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 245, + 481, + 277 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 245, + 481, + 277 + ], + "spans": [ + { + "bbox": [ + 138, + 245, + 481, + 277 + ], + "type": "text", + "content": "2. Avrahami, O., Lischinski, D., Fried, O.: Blended diffusion for text-driven editing of natural images. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18208-18218 (2022)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 277, + 481, + 310 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 277, + 481, + 310 + ], + "spans": [ + { + "bbox": [ + 138, + 277, + 481, + 310 + ], + "type": "text", + "content": "3. Brack, M., Friedrich, F., Hintersdorf, D., Struppek, L., Schramowski, P., Kersting, K.: Sega: Instructing text-to-image models using semantic guidance. Advances in Neural Information Processing Systems 36 (2024)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 310, + 481, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 310, + 481, + 342 + ], + "spans": [ + { + "bbox": [ + 138, + 310, + 481, + 342 + ], + "type": "text", + "content": "4. Brooks, T., Holynski, A., Efros, A.A.: Instructpix2pix: Learning to follow image editing instructions. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18392-18402 (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 342, + 481, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 342, + 481, + 374 + ], + "spans": [ + { + "bbox": [ + 138, + 342, + 481, + 374 + ], + "type": "text", + "content": "5. Chang, H., Zhang, H., Jiang, L., Liu, C., Freeman, W.T.: Maskgit: Masked generative image transformer. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11315-11325 (2022)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 374, + 481, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 374, + 481, + 407 + ], + "spans": [ + { + "bbox": [ + 138, + 374, + 481, + 407 + ], + "type": "text", + "content": "6. Chen, J., Shen, Y., Gao, J., Liu, J., Liu, X.: Language-based image editing with recurrent attentive models. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 8721-8729 (2018)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 407, + 481, + 429 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 407, + 481, + 429 + ], + "spans": [ + { + "bbox": [ + 138, + 407, + 481, + 429 + ], + "type": "text", + "content": "7. Couairon, G., Verbeek, J., Schwenk, H., Cord, M.: Diffedit: Diffusion-based semantic image editing with mask guidance. arXiv preprint arXiv:2210.11427 (2022)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 429, + 481, + 460 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 429, + 481, + 460 + ], + "spans": [ + { + "bbox": [ + 138, + 429, + 481, + 460 + ], + "type": "text", + "content": "8. Dong, H., Yu, S., Wu, C., Guo, Y.: Semantic image synthesis via adversarial learning. In: Proceedings of the IEEE international conference on computer vision. pp. 5706-5714 (2017)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 460, + 481, + 482 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 460, + 481, + 482 + ], + "spans": [ + { + "bbox": [ + 138, + 460, + 481, + 482 + ], + "type": "text", + "content": "9. Gandikota, R., Materzynska, J., Fiotto-Kaufman, J., Bau, D.: Erasing concepts from diffusion models. arXiv preprint arXiv:2303.07345 (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 482, + 481, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 482, + 481, + 514 + ], + "spans": [ + { + "bbox": [ + 138, + 482, + 481, + 514 + ], + "type": "text", + "content": "0. Gandikota, R., Orgad, H., Belinkov, Y., Materzyńska, J., Bau, D.: Unified concept editing in diffusion models. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 5111-5120 (2024)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 514, + 481, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 514, + 481, + 547 + ], + "spans": [ + { + "bbox": [ + 138, + 514, + 481, + 547 + ], + "type": "text", + "content": "1. Geng, Z., Yang, B., Hang, T., Li, C., Gu, S., Zhang, T., Bao, J., Zhang, Z., Hu, H., Chen, D., et al.: Instructdiffusion: A generalist modeling interface for vision tasks. arXiv preprint arXiv:2309.03895 (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 547, + 481, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 547, + 481, + 568 + ], + "spans": [ + { + "bbox": [ + 138, + 547, + 481, + 568 + ], + "type": "text", + "content": "2. Heng, A., Soh, H.: Continual learning for forgetting in deep generative models (2023)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 568, + 481, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 568, + 481, + 601 + ], + "spans": [ + { + "bbox": [ + 138, + 568, + 481, + 601 + ], + "type": "text", + "content": "3. Hertz, A., Mokady, R., Tenenbaum, J., Aberman, K., Pritch, Y., Cohen-Or, D.: Prompt-to-prompt image editing with cross attention control. arXiv preprint arXiv:2208.01626 (2022)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 601, + 481, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 601, + 481, + 633 + ], + "spans": [ + { + "bbox": [ + 138, + 601, + 481, + 633 + ], + "type": "text", + "content": "4. Karras, T., Laine, S., Aila, T.: A style-based generator architecture for generative adversarial networks. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 4401-4410 (2019)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 138, + 633, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 633, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 633, + 481, + 665 + ], + "type": "text", + "content": "5. Karras, T., Laine, S., Aittala, M., Hellsten, J., Lehtinen, J., Aila, T.: Analyzing and improving the image quality of stylegan. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 8110-8119 (2020)" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 336, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 336, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 336, + 91, + 447, + 102 + ], + "type": "text", + "content": "Responsible Visual Editing" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 666 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "type": "text", + "content": "16. Kumari, N., Zhang, B., Wang, S.Y., Shechtman, E., Zhang, R., Zhu, J.Y.: Ablating Concepts in Text-to-Image Diffusion Models (May 2023), http://arxiv.org/abs/2303.13516, arXiv:2303.13516 [cs]" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 149, + 481, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 149, + 481, + 171 + ], + "spans": [ + { + "bbox": [ + 133, + 149, + 481, + 171 + ], + "type": "text", + "content": "17. Kumari, N., Zhang, B., Wang, S.Y., Shechtman, E., Zhang, R., Zhu, J.Y.: Ablating concepts in text-to-image diffusion models. arXiv preprint arXiv:2303.13516 (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 171, + 481, + 203 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 171, + 481, + 203 + ], + "spans": [ + { + "bbox": [ + 133, + 171, + 481, + 203 + ], + "type": "text", + "content": "18. Li, F., Zhang, H., Sun, P., Zou, X., Liu, S., Yang, J., Li, C., Zhang, L., Gao, J.: Semantic-sam: Segment and recognize anything at any granularity. arXiv preprint arXiv:2307.04767 (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 203, + 481, + 236 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 203, + 481, + 236 + ], + "spans": [ + { + "bbox": [ + 133, + 203, + 481, + 236 + ], + "type": "text", + "content": "19. Meng, C., He, Y., Song, Y., Song, J., Wu, J., Zhu, J.Y., Ermon, S.: Sdedit: Guided image synthesis and editing with stochastic differential equations. arXiv preprint arXiv:2108.01073 (2021)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 236, + 481, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 236, + 481, + 267 + ], + "spans": [ + { + "bbox": [ + 132, + 236, + 481, + 267 + ], + "type": "text", + "content": "20. Ni, M., Li, X., Zuo, W.: Nuwa-lip: Language-guided image inpainting with defect-free vqgan. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14183-14192 (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 267, + 481, + 289 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 267, + 481, + 289 + ], + "spans": [ + { + "bbox": [ + 132, + 267, + 481, + 289 + ], + "type": "text", + "content": "21. Ni, M., Wu, C., Wang, X., Yin, S., Wang, L., Liu, Z., Duan, N.: Ores: Open-vocabulary responsible visual synthesis. arXiv preprint arXiv:2308.13785 (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 289, + 481, + 322 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 289, + 481, + 322 + ], + "spans": [ + { + "bbox": [ + 132, + 289, + 481, + 322 + ], + "type": "text", + "content": "22. Nichol, A., Dhariwal, P., Ramesh, A., Shyam, P., Mishkin, P., McGrew, B., Sutskever, I., Chen, M.: Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741 (2021)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 322, + 365, + 332 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 322, + 365, + 332 + ], + "spans": [ + { + "bbox": [ + 132, + 322, + 365, + 332 + ], + "type": "text", + "content": "23. OpenAI: Chatgpt. https://chat.openai.com (2022)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 332, + 383, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 332, + 383, + 342 + ], + "spans": [ + { + "bbox": [ + 132, + 332, + 383, + 342 + ], + "type": "text", + "content": "24. OpenAI: Dall-e 3. https://openai.com/dall-e-3 (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 342, + 481, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 342, + 481, + 374 + ], + "spans": [ + { + "bbox": [ + 132, + 342, + 481, + 374 + ], + "type": "text", + "content": "25. Podell, D., English, Z., Lacey, K., Blattmann, A., Dockhorn, T., Müller, J., Penna, J., Rombach, R.: Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952 (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 374, + 481, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 374, + 481, + 407 + ], + "spans": [ + { + "bbox": [ + 132, + 374, + 481, + 407 + ], + "type": "text", + "content": "26. Ramesh, A., Dhariwal, P., Nichol, A., Chu, C., Chen, M.: Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125 1(2), 3 (2022)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 407, + 481, + 439 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 407, + 481, + 439 + ], + "spans": [ + { + "bbox": [ + 132, + 407, + 481, + 439 + ], + "type": "text", + "content": "27. Ramesh, A., Pavlov, M., Goh, G., Gray, S., Voss, C., Radford, A., Chen, M., Sutskever, I.: Zero-shot text-to-image generation. In: International Conference on Machine Learning. pp. 8821-8831. PMLR (2021)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 439, + 481, + 460 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 439, + 481, + 460 + ], + "spans": [ + { + "bbox": [ + 132, + 439, + 481, + 460 + ], + "type": "text", + "content": "28. Rando, J., Paleka, D., Lindner, D., Heim, L., Tramér, F.: Red-teaming the stable diffusion safety filter. arXiv preprint arXiv:2210.04610 (2022)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 460, + 481, + 493 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 460, + 481, + 493 + ], + "spans": [ + { + "bbox": [ + 132, + 460, + 481, + 493 + ], + "type": "text", + "content": "29. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10684-10695 (2022)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 493, + 481, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 493, + 481, + 536 + ], + "spans": [ + { + "bbox": [ + 132, + 493, + 481, + 536 + ], + "type": "text", + "content": "30. Schramowski, P., Brack, M., Deiseroth, B., Kersting, K.: Safe latent diffusion: Mitigating inappropriate degeneration in diffusion models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 22522-22531 (2023)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 536, + 481, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 536, + 481, + 578 + ], + "spans": [ + { + "bbox": [ + 132, + 536, + 481, + 578 + ], + "type": "text", + "content": "31. Schuhmann, C., Beaumont, R., Vencu, R., Gordon, C., Wightman, R., Cherti, M., Coombes, T., Katta, A., Mullis, C., Wortsman, M., et al.: Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems 35, 25278-25294 (2022)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 578, + 481, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 578, + 481, + 611 + ], + "spans": [ + { + "bbox": [ + 132, + 578, + 481, + 611 + ], + "type": "text", + "content": "32. Wang, S.Y., Wang, O., Zhang, R., Owens, A., Efros, A.A.: Cnn-generated images are surprisingly easy to spot... for now. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 8695-8704 (2020)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 611, + 481, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 611, + 481, + 643 + ], + "spans": [ + { + "bbox": [ + 132, + 611, + 481, + 643 + ], + "type": "text", + "content": "33. Yang, J., Zhang, H., Li, F., Zou, X., Li, C., Gao, J.: Set-of-mark prompting unleashes extraordinary visual grounding in gpt-4v. arXiv preprint arXiv:2310.11441 (2023)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 132, + 643, + 481, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 643, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 643, + 481, + 666 + ], + "type": "text", + "content": "34. Yu, T., Feng, R., Feng, R., Liu, J., Jin, X., Zeng, W., Chen, Z.: Inpaint anything: Segment anything meets image inpainting. arXiv preprint arXiv:2304.06790 (2023)" + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 215, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 215, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 215, + 100 + ], + "type": "text", + "content": "M. Ni et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 171 + ], + "type": "list", + "angle": 0, + "index": 4, + "blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 481, + 138 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 481, + 138 + ], + "type": "text", + "content": "35. Zhang, E., Wang, K., Xu, X., Wang, Z., Shi, H.: Forget-me-not: Learning to forget in text-to-image diffusion models. arXiv preprint arXiv:2303.17591 (2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 139, + 481, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 139, + 481, + 171 + ], + "spans": [ + { + "bbox": [ + 132, + 139, + 481, + 171 + ], + "type": "text", + "content": "36. Zhang, Y., Wei, Y., Jiang, D., Zhang, X., Zuo, W., Tian, Q.: Control video: Training-free controllable text-to-video generation. arXiv preprint arXiv:2305.13077 (2023)" + } + ] + } + ], + "index": 3 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 336, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 336, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 336, + 91, + 447, + 102 + ], + "type": "text", + "content": "Responsible Visual Editing" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Restore Anything with Masks_ Leveraging Mask Image Modeling for Blind All-in-One Image Restoration/0de4c62c-e2d6-4f18-8a4e-c05352d6d956_content_list.json b/2024/Restore Anything with Masks_ Leveraging Mask Image Modeling for Blind All-in-One Image Restoration/0de4c62c-e2d6-4f18-8a4e-c05352d6d956_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..9c3445de45f0cc63920945a362081567214a5946 --- /dev/null +++ b/2024/Restore Anything with Masks_ Leveraging Mask Image Modeling for Blind All-in-One Image Restoration/0de4c62c-e2d6-4f18-8a4e-c05352d6d956_content_list.json @@ -0,0 +1,2481 @@ +[ + { + "type": "text", + "text": "Restore Anything with Masks: Leveraging Mask Image Modeling for Blind All-in-One Image Restoration", + "text_level": 1, + "bbox": [ + 227, + 140, + 779, + 205 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Chu-Jie Qin $^{1,2*}$ , Rui-Qi Wu $^{1,2}$ , Zukun Liu $^{3}$ , Xin Lin $^{5}$ , Chun-Le Guo $^{1,2}$ , Hyun Hee Park $^{4}$ , and Chongyi Li $^{1,2\\dagger}$", + "bbox": [ + 303, + 233, + 696, + 265 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ VCIP, CS, Nankai University", + "bbox": [ + 395, + 276, + 607, + 290 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "2 NKIARI, Shenzhen Futian", + "bbox": [ + 403, + 290, + 596, + 304 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{chujie.qin,wuruiqi}@mail.nankai.edu.cn", + "bbox": [ + 349, + 305, + 651, + 318 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{guochunle, lichongyi}@nankai.edu.cn", + "bbox": [ + 362, + 320, + 640, + 333 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3 Samsung Research, China, Beijing (SRC-B)", + "bbox": [ + 348, + 332, + 653, + 347 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "4 The Department of Camera Innovation Group, Samsung Electronics {zikun.liu, inextg.park}@samsung.com", + "bbox": [ + 267, + 347, + 733, + 375 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Sichuan University", + "bbox": [ + 429, + 375, + 571, + 388 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "linxin@stu.scu.edu.cn", + "bbox": [ + 419, + 390, + 583, + 401 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. All-in-one image restoration aims to handle multiple degradation types using one model. This paper proposes a simple pipeline for all-in-one blind image restoration to Restore Anything with Masks (RAM). We focus on the image content by utilizing Mask Image Modeling to extract intrinsic image information rather than distinguishing degradation types like other methods. Our pipeline consists of two stages: masked image pre-training and fine-tuning with mask attribute conductance. We design a straightforward masking pre-training approach specifically tailored for all-in-one image restoration. This approach enhances networks to prioritize the extraction of image content priors from various degradations, resulting in a more balanced performance across different restoration tasks and achieving stronger overall results. To bridge the gap of input integrity while preserving learned image priors as much as possible, we selectively fine-tuned a small portion of the layers. Specifically, the importance of each layer is ranked by the proposed Mask Attribute Conductance (MAC), and the layers with higher contributions are selected for finetuning. Extensive experiments demonstrate that our method achieves state-of-the-art performance. Our code and model will be released at https://github.com/Dragonisss/RAM.", + "bbox": [ + 259, + 431, + 743, + 696 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Image Restoration $\\cdot$ All-in-One $\\cdot$ Mask Image Modeling", + "bbox": [ + 261, + 707, + 720, + 722 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 744, + 375, + 758 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Image restoration involves the restoration of low-quality images affected by various degradation, typically arising from adverse environmental conditions (e.g.,", + "bbox": [ + 212, + 772, + 787, + 804 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* A part of this work is done during Chu-Jie Qin's internship at Samsung.", + "bbox": [ + 236, + 810, + 727, + 825 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "†Chongyi Li is the corresponding author.", + "bbox": [ + 240, + 825, + 513, + 839 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "rain, haze, low-light), hardware-related issues (e.g., noise and blur), and post-processing artifacts (e.g., JPEG compression). Image restoration serves not only to enhance the visual appeal of images but also contributes to practical application scenarios such as autonomous driving and surveillance.", + "bbox": [ + 212, + 146, + 787, + 207 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Modern techniques in this field mainly focus on learning fixed patterns formed during the degradation process, i.e., degradation priors. Some works [29, 30, 60] utilize task-specific priors to solve a certain degradation problem, while another research line [3, 28, 39, 48, 55] tries to design a general network architecture that can effectively learn each degradation pattern. Nevertheless, the above methods only enable the network to learn a single degradation, resulting in an imbalanced situation when dealing with multiple types of degradation.", + "bbox": [ + 212, + 210, + 509, + 422 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To tackle the problem stated above, all-in-one methods have emerged, aiming to handle multiple degradations using one model. Most of these approaches tend to utilize explicit priors (e.g., AirNet [21]) or introduce an extra module (e.g., PromptIR [42]) to discern image degradation patterns, thereby assisting", + "bbox": [ + 212, + 426, + 511, + 547 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/06ec5f2f8dd06754772466ca6479e1a8f2b71672aa029fa0b4c8a52e05e7c4ca.jpg", + "image_caption": [ + "Fig.1: Our RAM achieves more balanced and more powerful performance than the state-of-the-art methods (AirNet [21], TAPE [31], PromptIR [42]) for all-in-one blind image restoration." + ], + "image_footnote": [], + "bbox": [ + 521, + 241, + 787, + 431 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "the model in performing the restoration. However, these methods place their emphasis on distinguishing degradation types in images rather than the image content, leading to lower scalability and fuzzy decision boundaries when more degradation types are involved. We argue that the essence of image restoration is to extract intrinsic image information from corrupted images rather than eliminate degradation patterns, i.e., learning image prior rather than degradation prior. It is worth noting that TAPE [31] similarly suggests that understanding normal image nature aids restoration by introducing a natural image prior. Nevertheless, TAPE utilizes the model output as the optimization target, which causes the model to amplify its own errors and learn the image prior with bias.", + "bbox": [ + 212, + 547, + 787, + 700 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we focus on tackling how to extract intrinsic image information from diverse corrupted images. Some attempts [2, 7] by Mask Image Modeling (MIM) in low-level vision have caught our attention. As a pretraining strategy, MIM has been widely validated for its effectiveness in high-level tasks, thanks to its generic representation of images. Simultaneously, the model also learns the distribution of natural images, which encompasses the intrinsic information we aim to extract from the images. Built on MIM, we propose a simple pipeline for all-in-one blind image restoration that Restores Anything with Masks (RAM), which includes two stages: the mask pre-training stage and the", + "bbox": [ + 212, + 704, + 787, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "Qin et al.", + "bbox": [ + 271, + 114, + 336, + 128 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "fine-tuning stage with Mask Attribute Conductance (MAC). In the pre-training stage, we randomly mask corrupted images at the pixel-wise level and force the network to predict the clear one corresponding to the masked pixels, extracting inherent image information from corrupted images. In the fine-tuning stage, we focus on overcoming the input integrity gap caused by changing masked input during pre-training into the whole image during inference while preserving learned prior as much as possible.", + "bbox": [ + 212, + 146, + 782, + 252 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Specifically, we first evaluated the importance of each network layer in addressing this gap by the proposed MAC. Following that, we chose the top $k\\%$ most critical layers for fine-tuning while keeping the rest of the network layers frozen. We demonstrate that after a brief fine-tuning period (even if only $10\\%$ layers are tuned), the model can achieve a highly satisfactory performance level, surpassing models trained using traditional pair-wise training. Additionally, our pipeline can be plug-and-play used in any network without introducing additional computational overhead.", + "bbox": [ + 212, + 252, + 784, + 372 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The contributions of this work are as follows:", + "bbox": [ + 238, + 373, + 566, + 387 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We discuss the challenge of adopting MIM in low-level vision and propose a MIM-based pre-training strategy tailored to all-in-one blind image restoration, which allows the restoration networks to effectively learn inherent image information while guaranteeing reconstruction results.", + "- We proposed Mask Attribute Conductance to evaluate the importance of each layer in addressing the input integrity gap so that a very small portion (e.g. $10\\%$ ) of critical layers are tuned to bridge this gap while preserving the image prior learned by MIM.", + "- Our proposed RAM provides a fresh perspective to achieve more balanced and powerful all-in-one blind image restoration, which focuses on extracting inherent image information from corrupted images. Our pipeline can be applied to any image restoration network without introducing additional computational overhead." + ], + "bbox": [ + 225, + 397, + 784, + 594 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 617, + 387, + 633 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.1 Image Restoration for Multi Degradations", + "text_level": 1, + "bbox": [ + 214, + 648, + 609, + 664 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "While neural networks have demonstrated impressive performance in single degradation image restoration [9, 13, 14, 17, 22, 23, 29, 30, 49, 60], recent works have shifted their focus towards addressing the more challenging domain of multi-degradation image restoration. A group of methods [3, 28, 39, 48, 55] aims at designing a general architecture that can effectively learn each degradation pattern. SwinIR [28] employs a window attention mechanism to convert global attention into a localized approach, effectively reducing computational overhead. In addition, the U-shaped transformer-based methods [48, 55] are employed to extract multi-scale features and reduce computational overhead. However, these methods have to train individually on each restoration task. Several methods [1, 24] leverage multiple input and output heads to empower the network to restore", + "bbox": [ + 212, + 672, + 787, + 840 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "Restore Anything with Masks", + "bbox": [ + 529, + 114, + 730, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/97b3ce8594e2da099216f661a7efbf5f91f85036c934e86b9a3d32f0dd9f02bb.jpg", + "image_caption": [ + "Fig. 2: The illumination of our overall pipeline. 1) Pre-training the model with mask image pre-training method tailored to low-level vision. We randomly mask degraded images at the pixel level with a $50\\%$ masking ratio and reconstruct the clean images. 2) The Fine-tuning stage is followed to overcome the input integrity gap caused by changing masked input during pre-training into the whole image during inference. We analyze the importance of each network layer for resolving the input integrity gap according to the proposed MAC and rank them in descending order. The top $k\\%$ of network layers are selected for fine-tuning on the complete image." + ], + "image_footnote": [], + "bbox": [ + 233, + 143, + 772, + 321 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "various types of degraded images. Nonetheless, this kind of approach may lead to the diminished scalability of the model. Recently, several subsequent methods [4, 21, 36, 41, 42, 56, 61] have been proposed to employ a unified network to address multiple restoration issues. Most of these methods put emphasis on learning how to distinguish different types of degradations and restore corrupted images. Typically, AirNet [21] first proposed an all-in-one image restoration task. The method initially pretrains a degradation classifier based on contrastive learning and subsequently utilizes it to assist in all-in-one image restoration. PromptIR [42] has introduced a learnable prompt-based module. Instead of constraining the degradation category, it enables the model to autonomously learn features that are advantageous to its performance by using an adaptive prompt. Our RAM takes a fresh perspective that focuses on extracting common content information from corrupted images, without any extra design to distinguish degradations, which helps us achieve balance and powerful performance when more degradation types are taken into consideration.", + "bbox": [ + 212, + 473, + 787, + 700 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.2 Mask Image Modeling", + "text_level": 1, + "bbox": [ + 215, + 723, + 447, + 739 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Inspired by Mask Language Modeling [18,43], Mask Image Modeling (MIM) [15, 51] is introduced as a pretraining approach to learn general representations in high-level vision. MAE [15] effectively utilizes MIM for predicting hidden tokens, demonstrating strong performance and generalization across various downstream tasks. SimMIM [51] proposed a general masked image modeling method based on Swin-ViT [34]. Painter [47] unifies multiple tasks under image-to-image trans", + "bbox": [ + 212, + 748, + 787, + 842 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "Qin et al.", + "bbox": [ + 271, + 114, + 336, + 128 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "lation and leverages MIM pretraining. In recent years, there have been efforts to incorporate MIM into the realm of low-level vision to enhance model generalization. Among them, [2] and [7] are the most closely aligned with our focus. [2] employs the MIM model to enhance the model's generalization for denoising tasks but has not explored its potential in multi-task scenarios. [7] utilizes MIM for pre-training the model encoder to introduce generative prior and subsequently employs the decoder for restoration. However, it does not fully harness the potential of MIM. Our proposed RAM utilizes MIM to unify the optimization objective for various image restoration tasks into reconstructing intrinsic image information. This allows the network to learn restoration functions more balanced and effectively. Moreover, to preserve the image priors learned by MIM, we designed a fine-tuning strategy based on MAC analysis (in Sec. 3.3). This enables us to achieve comparable performance by fine-tuning only a small portion (e.g. $10\\%$ ) of layers, fully tapping into the potential of MIM.", + "bbox": [ + 212, + 146, + 787, + 359 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "2.3 Gradient-based Attribution", + "text_level": 1, + "bbox": [ + 214, + 380, + 488, + 395 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Gradient-based attribution methods [6,12,44-46,50] are often used to clarify how hidden units (or inputs) impact the output of networks. One commonly used approach is Integrated Gradients (IG) [45,46], which accumulates gradients along a linear path from the baseline input to the target input in the pixel/feature space. After that, IntInf [19] and layer conductance [6] alter IG to attribute neuron importance along the same path. In our work, we expect to find the key layers that can effectively overcome the distribution shift between training data and inference data. We propose Mask Attribute Conductance (MAC) based on the layer conductance and accumulated MAC of each layer along the Mask Attribute Path (MAP). MAC can represent the layer's importance along the MAP. In this way, we can fine-tune the top $k\\%$ critical layers of the pre-trained network, preserving to a great extent the image priors learned during pretraining.", + "bbox": [ + 212, + 405, + 787, + 587 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3 Methodology", + "text_level": 1, + "bbox": [ + 214, + 609, + 380, + 627 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we start with discussing the challenges of using MIM in low-level vision tasks (Sec. 3.1). Following that, we present our pipeline for all-in-one blind image restoration, which contains two parts: pre-training with MIM (Sec. 3.2) and fine-tuning with Mask Attribute Conductance (MAC) Analysis (Sec. 3.3).", + "bbox": [ + 212, + 641, + 787, + 703 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1 Rethinking MIM in Low-Level Vision", + "text_level": 1, + "bbox": [ + 214, + 724, + 571, + 739 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "MIM is a process that randomly masks certain parts of an image and extracts features from the remaining visible parts to reconstruct the entire image. It allows models to acquire a generic representation of images and thus achieve good pre-training, which is verified in many high-level tasks [15, 51]. Moreover, the models also learn the distribution of natural images during the image reconstruction, i.e. MIM pre-training. This incidental acquisition of prior knowledge", + "bbox": [ + 212, + 750, + 787, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "Restore Anything with Masks", + "bbox": [ + 529, + 114, + 730, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/26468efb0aa41a9754bbd12b714a4d07f6aaca410620d49d634ae47b3f851a9e.jpg", + "image_caption": [ + "patch size=1" + ], + "image_footnote": [], + "bbox": [ + 217, + 143, + 395, + 213 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/40daf3bdc76771c1f65564c01a0834e29a029947d8e28ccf094ae4bed932624a.jpg", + "image_caption": [ + "patch size=4", + "Fig. 3: Mask Image Modeling reconstruction with different patch sizes. We pre-trained with different patch sizes and visualized the mask inputs (left), and the corresponding MIM reconstructions (right)." + ], + "image_footnote": [], + "bbox": [ + 416, + 143, + 591, + 212 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/980d7a274f6a681b09fa405bb062263e0583c28bfc40c0b776bb8fa36ff9c0a1.jpg", + "image_caption": [ + "patch size=8" + ], + "image_footnote": [], + "bbox": [ + 611, + 143, + 784, + 212 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "is instrumental in tasks like image restoration. Despite these advantages, applying MIM in pretraining a model for low-level vision tasks is still under-explored, primarily due to the challenges that must be addressed in the process.", + "bbox": [ + 212, + 316, + 782, + 362 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Firstly, the main purpose of vanilla MIM is not high-quality reconstruction but good feature extraction for high-level tasks. Therefore, it masks a wider range of images to gather semantic information but not pixel-level content, reflected in token-level masking and a high mask ratio. CSFormer [7] directly adopts this strategy on low-level vision pre-training. However, some studies verify that semantic information is not as important for image restoration as it is in pattern recognition tasks [33, 37]. Moreover, high-degree masking leads to producing detail-deficient results, as shown in Fig. 3, which is harmful to low-level tasks.", + "bbox": [ + 212, + 364, + 784, + 484 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Secondly, the training objective of MIM is to reconstruct the masked input images, so it can only produce results with the same domain as the input image. However, we hope the model gains the ability to bridge low-quality domain to high-quality domain, i.e. recover clean content from degraded input. Therefore, it is necessary to introduce paired data when pre-training image restoration models by MIM (see the experiment in Sec. 4.3 for details). Chen et al. [2] demonstrate that pair-wise MIM training enhances the generalization performance over different types of noisy images. In this paper, we take a step forward to explore the effectiveness of MIM on multiple degradations with larger variance.", + "bbox": [ + 212, + 487, + 785, + 623 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.2 Pretraining with MIM", + "text_level": 1, + "bbox": [ + 214, + 659, + 449, + 674 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Based on the above analysis, we design a MIM pre-training paradigm tailored for low-level vision.", + "bbox": [ + 212, + 698, + 784, + 728 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Masking. During the pre-training stage, we randomly mask the pixels of degraded images (mask images in a $1 \\times 1$ patch size) with a $50\\%$ mask ratio. We found that fine-grained masked patches and balanced mask ratio are beneficial to image restoration, which can be demonstrated in Sec. 4.3.", + "bbox": [ + 212, + 731, + 784, + 792 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Besides, since our MIM pre-training has a similar target to subsequent low-level tasks, we do not need to change the decoder like MAE [15] does but just fine-tune it.", + "bbox": [ + 212, + 794, + 784, + 839 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "Qin et al.", + "bbox": [ + 271, + 114, + 336, + 128 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/ddd442039b45f8792232e47e1eb3152d425b123907e380a8991c86ee504114ed.jpg", + "image_caption": [ + "Fig. 4: The effect of MIM reconstruction with different input integrity on kernel deblurring (orange border) and denoising (blue border). We also visualize the color distributions of reconstructions in various tasks above. It shows that the distribution of the reconstruction results obtained using the twin-masks method as input is closer to the real images (ground truth) compared to the results obtained using the whole input." + ], + "image_footnote": [], + "bbox": [ + 285, + 145, + 720, + 412 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Reconstruction target. Following the Bert [18] and MAE [15], we choose L1 loss to supervise the masked part. The training objective can be written as:", + "bbox": [ + 214, + 530, + 784, + 561 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\underset {\\theta} {\\arg \\min } \\mathbb {E} [ | | \\tilde {\\mathcal {M}} (I - f (\\mathcal {M} (I _ {d}), \\theta)) | | ], \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 375, + 571, + 785, + 597 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $\\{I, I_d\\}$ represents a pair of clean image and degraded image, $f(\\cdot, \\theta)$ denotes a network with parameters $\\theta$ , $\\mathcal{M}(\\cdot)$ is a random binary masking operation and $\\tilde{\\mathcal{M}}(\\cdot) = 1 - \\mathcal{M}(\\cdot)$ .", + "bbox": [ + 214, + 608, + 787, + 656 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.3 Finetuning with Mask Attribute Conductance Analysis", + "text_level": 1, + "bbox": [ + 214, + 676, + 718, + 694 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Observation. During pre-training, the network learns rich content priors. However, the incompleteness of the masked input prevents the direct use of the pre-trained model for inference, as it would result in a distribution shift in the outputs. As shown in Fig. 4, We start by feeding the entire image into a pretrained model, leading to a color-distorted result. Next, we use a pair of complementary masks, referred to as twin-masks, to individually mask the image. Subsequently, we input both of these complementarily masked images into the network. By combining the pixel values predicted by each image, we generate a higher-quality image. This observation indicates that the hindrance to using", + "bbox": [ + 212, + 704, + 787, + 840 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "Restore Anything with Masks", + "bbox": [ + 529, + 114, + 730, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 114, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "mask pre-trained model directly for inference lies in input incompleteness rather than the model's inability to learn the restoration function.", + "bbox": [ + 212, + 146, + 782, + 175 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Building upon this insight, we explore the possibility of minimizing the influence of disparities in data input formats via model fine-tuning. To maintain the learned priors, it is essential to retain pre-trained parameters as extensively as possible while employing the fewest but most effective layers for fine-tuning. To tackle this, we introduce the concept of mask attribution conductance, which quantifies the importance of each layer concerning the fine-tuning objective. We then identify the top- $k\\%$ most critical layers for fine-tuning.", + "bbox": [ + 212, + 176, + 784, + 281 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Preliminary. Before giving the definition of Mask Attribute Conductance (MAC), we briefly recall the definition of integrate gradient [45] (IG) and neuron conductance [6] (Cond). Considering a linear path $\\gamma (\\alpha) = x^{\\prime} + \\alpha (x - x^{\\prime})$ from base input $x^{\\prime}$ to target input $x$ , we can attribute output change $F(x) - F(x^{\\prime})$ to $i$ -th dimension of input/feature $x_{i}$ (e.g. a pixel) by calculating its integrate gradient, which formally as below:", + "bbox": [ + 212, + 282, + 795, + 372 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {I G} _ {i} (x) := \\left(x _ {i} - x _ {i} ^ {\\prime}\\right) \\cdot \\int_ {0} ^ {1} \\frac {\\partial F \\left(x ^ {\\prime} + \\alpha \\left(x - x ^ {\\prime}\\right)\\right)}{\\partial x _ {i}} d \\alpha . \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 331, + 381, + 785, + 415 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We can also attribute output change to a specific neuron $y$ by improving IG, which involves calculating the conductance. The conductance [6] of the hidden neuron $y$ along the $\\gamma(\\alpha)$ is:", + "bbox": [ + 214, + 420, + 784, + 467 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\operatorname {C o n d} ^ {y} (x) := \\sum_ {i} \\left(x _ {i} - x _ {i} ^ {\\prime}\\right) \\cdot \\int_ {0} ^ {1} \\frac {\\partial F \\left(x ^ {\\prime} + \\alpha \\left(x - x ^ {\\prime}\\right)\\right)}{\\partial y} \\cdot \\frac {\\partial y}{\\partial x _ {i}} d \\alpha \\tag {3} \\\\ = \\sum_ {i} \\int_ {0} ^ {1} \\frac {\\partial F (\\gamma (\\alpha))}{\\partial y} \\cdot \\frac {\\partial y}{\\partial \\alpha} d \\alpha , \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 292, + 474, + 785, + 553 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Note that $(x_{i} - x_{i}^{\\prime}) = \\frac{\\partial(x^{\\prime} + \\alpha(x_{i} - x_{i}^{\\prime}))}{\\partial\\alpha}$ . Certainly, we can broaden Eq. (3) to compute conductance when integrating along any given path $\\alpha :[s,t]\\to P$ :", + "bbox": [ + 212, + 561, + 782, + 597 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {G e n e r a l C o n d} ^ {y} (x) := \\sum_ {i} \\int_ {P} \\frac {\\partial F (X _ {i} (\\alpha))}{\\partial y} \\cdot \\frac {\\partial y}{\\partial \\alpha} d \\alpha , \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 331, + 604, + 784, + 640 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $X: R \\to R^n$ is the function of the path from $x'$ to $x$ , which satisfies $X(s) = x'$ , $X(t) = x$ . [s,t] represent the domain of the path function $X$ .\"", + "bbox": [ + 214, + 648, + 784, + 678 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Finetuning with MAC. To find effective layers to finetune, we propose Mask Attribute Conductance (MAC) to evaluate how effective each layer is in overcoming the gap of input integrity. Considering such a nonlinear path $\\alpha : [0,1] \\to P_m$ from zero input $x'$ to whole input $x$ , which path function $X^m$ satisfies:", + "bbox": [ + 214, + 679, + 784, + 739 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\nX _ {i} ^ {m} (\\alpha ; \\alpha_ {i}) = \\left\\{ \\begin{array}{l l} x _ {i} ^ {\\prime}, & \\alpha < \\alpha_ {i} \\\\ x _ {i}, & e l s e \\end{array} , \\right. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 401, + 748, + 784, + 787 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $i$ refers to the index of pixels, $\\alpha_{i} \\in (0,1]$ is a set of parameters that indicate when each pixel gets masked. We define this path as a Mask Attribute Path (MAP). Apparently, $X^{m}(0) = x^{\\prime}$ and $X^{m}(1) = x$ .", + "bbox": [ + 212, + 794, + 782, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "Qin et al.", + "bbox": [ + 271, + 114, + 336, + 128 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/79c2fd74506cff6d5f6752e8da693a01e93ccc21182850c9ed51e06e5cd32ae9.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 279, + 146, + 485, + 279 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/1a9a4c98871017718d833e194f72449e2da08834b4440b4682743ad47424f0af.jpg", + "image_caption": [ + "(b)", + "Fig. 5: Illumination of (a) $X_{i}^{m}$ in Eq. (5) and (b) $\\tilde{X}_i^m$ in Eq. (6)." + ], + "image_footnote": [], + "bbox": [ + 511, + 146, + 722, + 279 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "However, $X^{m}$ is not differentiable, making it an invalid attribute path function. To solve this problem, we use a group of sigmoid-like functions $\\tilde{X}^{m}$ to approximate $X^{m}$ :", + "bbox": [ + 212, + 354, + 784, + 398 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {X} _ {i} ^ {m} \\left(\\alpha ; \\alpha_ {i}\\right) = \\frac {\\left(x _ {i} ^ {\\prime} - x _ {i}\\right)}{1 + e ^ {- \\delta \\left(x _ {i} ^ {\\prime} - \\alpha_ {i}\\right)}}. \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 395, + 398, + 784, + 430 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We can see that $\\tilde{X}^m$ is very close to $X^m$ when $\\delta$ is sufficiently large (as depicted in Fig. 5). And for each $\\tilde{X}_i^m$ , it will change sharply from $x_i'$ to $x_i$ when $\\alpha$ is in the neighborhood of $\\alpha_i$ .", + "bbox": [ + 212, + 435, + 782, + 481 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Here, we can give a definition of MAC as below:", + "bbox": [ + 238, + 482, + 596, + 496 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathrm {M A C} ^ {y} (x) := \\sum_ {i} \\int_ {P _ {m}} \\frac {\\partial F (X _ {i} (\\alpha))}{\\partial y} \\cdot \\frac {\\partial y}{\\partial \\alpha} d \\alpha \\\\ \\approx \\sum_ {i} \\int_ {0} ^ {1} \\frac {\\partial F \\left(\\tilde {X} _ {i} ^ {m} (\\alpha ; \\alpha_ {i})\\right)}{\\partial y} \\cdot \\frac {\\partial y}{\\partial \\alpha} d \\alpha . \\tag {7} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 339, + 503, + 784, + 580 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "In fact, a partial path is also available to attribute from a masked input $x_{m}$ with any mask ratio $r$ to whole input $x$ :", + "bbox": [ + 212, + 589, + 782, + 619 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {M A C} _ {r} ^ {y} (x) \\approx \\sum_ {i} \\int_ {1 - r} ^ {1} \\frac {\\partial F \\left(\\tilde {X} _ {i} ^ {m} (\\alpha ; \\alpha_ {i})\\right)}{\\partial y} \\cdot \\frac {\\partial y}{\\partial \\alpha} d \\alpha . \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 333, + 627, + 784, + 665 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "In practice, we use N-steps discretization to approximate the integral form of Eq. (8), which follows [44]:", + "bbox": [ + 212, + 672, + 782, + 705 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathrm {M A C} _ {r} ^ {y} (x) \\approx \\sum_ {i} \\sum_ {j = 1} ^ {N} \\frac {\\partial F \\left(\\tilde {X} _ {i} ^ {m} \\left(\\frac {j r}{N} ; \\alpha_ {i}\\right)\\right)}{\\partial y} \\tag {9} \\\\ \\cdot \\left(F _ {y} \\left(\\tilde {X} _ {i} ^ {m} \\left(\\frac {(j + 1) r}{N}\\right)\\right) - F _ {y} \\left(\\tilde {X} _ {i} ^ {m} \\left(\\frac {j r}{N}\\right)\\right)\\right). \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 326, + 712, + 784, + 787 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We compute the MAC of each layer of pre-trained networks, rank them in descending order based on their MAC values, and pick top- $k\\%$ layers for finetuning. The networks are initialized by pre-trained weight and only top- $k\\%$ layers", + "bbox": [ + 212, + 794, + 782, + 839 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "Restore Anything with Masks", + "bbox": [ + 529, + 114, + 730, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 114, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/b6ddd4448c4909a90ec6fa216be5f2f7b6dda5113a454a0087d47e7bc612b5b1.jpg", + "table_caption": [ + "Table 1: Quantitative comparison on seven challenging image restoration tasks, including dehazing, deraining, denoising, motion deblurring, low-light image enhancement (LLIE), kernel deblurring, and JPEG artifact removal. **boldface** and **underlined** indicate the best and second-best results, respectively." + ], + "table_footnote": [], + "table_body": "
MethodSOTS [20]Rain13k-Test [32]BSD68 [38]GoPro [40]LOL [5]LSDIR-Blur [26]LSDIR-Jpeg [26]Average
PSNR†/SSIM†PSNR†/SSIM†PSNR†/SSIM†PSNR†/SSIM†PSNR†/SSIM†PSNR†/SSIM†PSNR†/SSIM†PSNR†/SSIM†
Restomer [55]22.89/0.917227.05/0.846930.95/0.865727.46/0.849723.65/0.845819.60/0.365830.46/0.914126.01/0.8007
MPRNet [39]25.23/0.946325.36/0.806829.83/0.831725.90/0.794922.29/0.817025.68/0.828128.96/0.886526.18/0.8445
NAFNet [3]25.74/0.944524.65/0.787730.37/0.854025.53/0.790921.50/0.810429.08/0.913029.09/0.895526.57/0.8566
DL [8]21.16/0.904219.56/0.650816.15/0.586117.63/0.586219.26/0.777717.98/0.612119.55/0.696518.75/0.6877
TAPE [31]25.14/0.931923.66/0.781830.11/0.835425.97/0.796218.95/0.763224.26/0.765429.28/0.896525.34/0.8243
AirNet [21]21.66/0.836620.21/0.640227.99/0.725023.36/0.750316.65/0.670823.84/0.735824.36/0.802022.58/0.7372
SwinIR [28]27.29/0.962225.32/0.825830.65/0.854026.61/0.812518.66/0.804827.82/0.883930.13/0.907126.64/0.8643
RAM-SwinIR28.47/0.968926.31/0.848630.83/0.861126.89/0.820021.62/0.829126.66/0.851430.22/0.909627.28/0.8698
PromptIR [42]28.70/0.965927.46/0.858530.84/0.862527.71/0.856521.19/0.835631.01/0.938530.30/0.911728.17/0.8899
RAM-PromptIR29.64/0.969528.47/0.875130.86/0.862428.02/0.859224.46/0.858129.57/0.917930.33/0.911928.76/0.8935
", + "bbox": [ + 218, + 212, + 787, + 335 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/b91e724a60e506ed9c17e75dc3fa51915f8d0dbc0bc0b38ca7d21353b2431304.jpg", + "image_caption": [ + "Fig. 6: Dehaze visual comparison on SOTS dataset. Zoom in for details." + ], + "image_footnote": [], + "bbox": [ + 217, + 345, + 787, + 481 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "will be fine-tuned. More implementation details can be found in the supplementary material.", + "bbox": [ + 212, + 532, + 782, + 564 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4 Experiment", + "text_level": 1, + "bbox": [ + 214, + 585, + 366, + 603 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.1 Experiments Settings", + "text_level": 1, + "bbox": [ + 214, + 618, + 439, + 633 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Datasets and Metrics. We combine datasets from various restoration tasks to form the training set, following [59]. For high-cost tasks that degradations are difficult to synthesize, we leverage existing paired datasets, including RE-SIDE [20] for dehazing, Rain13k [10, 25, 27, 35, 53] for deraining, GoPro [40] for motion deblurring, and LOL-v2 [54] for low-light image enhancement (LLIE). For low-cost tasks that degradations are easy to synthesize (e.g. noise, kernel blur, and JPEG artifact), we generate corrupted images on the LSDIR dataset [26] during the training process, which involves generating Gaussian noise with random variation $\\sigma \\in (0, 50]$ , creating gaussian blurred images with a blur kernel of size $k = 15$ and random $\\sigma \\in [0.1, 3.1]$ , and introducing JPEG artifacts with a random quality parameter $q \\in [20, 90]$ .", + "bbox": [ + 212, + 643, + 787, + 809 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "For evaluation, we use SOTS-outdoor [20] for dehazing, Rain13k-Test (the combination of Rain100L [52], Rain100H [52], Test100 [58], Test1200 [57] and", + "bbox": [ + 212, + 809, + 785, + 840 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "Qin et al.", + "bbox": [ + 271, + 114, + 336, + 128 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/f262acc6170e25da2dfb873c137fa062b1c78c6821e77068e274d84a6d209a89.jpg", + "table_caption": [ + "Table 2: Quantitative Gaussian denoising results at different noise levels on BSD68 and Urban100 datasets in terms of PSNR." + ], + "table_footnote": [], + "table_body": "
MethodBSD68 [38]Urban100 [16]
σ = 15σ = 25σ = 50Averageσ = 15σ = 25σ = 50Average
NAFNet [3]33.2230.5927.3030.3732.6730.2126.9729.92
MPRNet [39]32.7330.1126.6529.8332.0629.4625.7729.10
Restormer [55]33.7931.1727.9030.9533.8331.4027.9931.07
DL [8]16.0416.2016.1916.1519.1719.1118.4718.92
TAPE [31]33.1030.3726.8630.1132.5929.9326.1929.57
AirNet [21]31.6328.8323.5227.9929.7926.9021.3526.01
SwinIR [28]33.5330.8927.5430.6533.5030.9927.3730.62
RAM-SwinIR33.6531.0627.7730.8233.8231.4327.9431.07
performance gains(↑0.12)(↑0.17)(↑0.23)(↑0.17)(↑0.32)(↑0.44)(↑0.57)(↑0.45)
PromptIR [42]33.6731.0627.8030.8433.5631.0827.6430.76
RAM-PromptIR33.7031.0827.7930.8633.7031.3027.9230.97
performance gains(↑0.03)(↑0.02)(↓0.01)(↑0.02)(↑0.14)(↑0.22)(↑0.28)(↑0.21)
", + "bbox": [ + 267, + 184, + 736, + 356 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Test2800 [11]) for deraining, GoPro for motion deblurring, LOL [5] for low-light enhancement, BSD68 [38] for denoising, LSDIR-val for kernel deblurring and JPEG artifact removal. Furthermore, We conducted evaluations including denoising tests with variances of 15, 25, and 50, deblurring tests at $k = 15$ and $\\sigma = 2.0$ , and JPEG artifact removal tests at $q = 50$ .", + "bbox": [ + 212, + 383, + 784, + 458 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Implementation Details. We apply our proposed RAM to SwinIR [28] and PromptIR [42]. The input size for RAM-SwinIR is 64, while for RAM-PromptIR it is 128. During the pre-training phase, we use the Adam optimizer to train RAM-SwinIR and RAM-PromptIR for 300 epochs, with the learning rate decaying from 1e-4 to 6e-5 following a cosine schedule. In the fine-tuning phase, we use the Adam optimizer to fine-tune the network layers obtained from the MAC analysis of RAM-SwinIR and RAM-PromptIR for 40 epochs, with the learning rate decaying from 2e-4 to 1e-7 following a cosine schedule. The batch sizes for RAM-SwinIR and RAM-PromptIR during the pre-training and fine-tuning phases are (12,4) and (4,4), respectively.", + "bbox": [ + 212, + 459, + 787, + 609 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4.2 Comparisons", + "text_level": 1, + "bbox": [ + 215, + 633, + 370, + 648 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "To validate the gain capability and effectiveness of our RAM, we apply the proposed RAM to SwinIR (a general image restoration method) and PromptIR (an all-in-one image restoration method). Four general architecture-based image restoration methods [3, 28, 39, 55] and four all-in-one methods [8, 21, 31, 42] are considered for comparison. We ensure that the number of supervised pixels employed by all other methods equals that used during the pre-training stage.", + "bbox": [ + 212, + 657, + 784, + 748 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "As illustrated in Tab. 1, our approach achieves the best or comparable performance on each task. On the average score across seven different tasks, our method with PromptIR [42] achieves 0.59dB performance gains compared to the second-best algorithm. Besides, the SwinIR equipped with RAM also yields $2.40\\%$ improvement on PSNR. Specifically, our RAM has significant benefits for dehazing and low-light enhancement. Tab. 2 shows the Quantitative denois-", + "bbox": [ + 212, + 750, + 785, + 839 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "Restore Anything with Masks", + "bbox": [ + 529, + 114, + 730, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/97fede3a9740dae1b8177c4426144258ecb5b722a4137b27f2e7016a98b6cb61.jpg", + "image_caption": [ + "Rainy Image" + ], + "image_footnote": [], + "bbox": [ + 217, + 143, + 312, + 191 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/c5456c7b9deedb3b3cc5bfd0eed573d67e9dd2969b912482006b2c533f55e21a.jpg", + "image_caption": [ + "NAFNet" + ], + "image_footnote": [], + "bbox": [ + 313, + 143, + 406, + 191 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/bcf69f06dddfb2989e89877bfb4726d905d3a5a1e14ccc67f8c7daba573e27f0.jpg", + "image_caption": [ + "MPRNet" + ], + "image_footnote": [], + "bbox": [ + 408, + 143, + 500, + 191 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/9547db10f802091a735362c2eaeded7edeaabf60ca54ae16000a371da88600be.jpg", + "image_caption": [ + "Restomer" + ], + "image_footnote": [], + "bbox": [ + 501, + 143, + 594, + 191 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/c7a068f252cd8e96241a0d432e7444aee4c0213bb00a7323a6c0f605bb396efa.jpg", + "image_caption": [ + "SwinIR" + ], + "image_footnote": [], + "bbox": [ + 596, + 143, + 689, + 191 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/760c799c2591cd1c736f81b6798398168d10b5bda311a6c6e99b6745b1f0106f.jpg", + "image_caption": [ + "RAM-SwinIR (Ours)" + ], + "image_footnote": [], + "bbox": [ + 692, + 143, + 784, + 191 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/a142d342183e62f3d0950629041f17e2d9b30fca3ba1a7e709c48f26c7297ae9.jpg", + "image_caption": [ + "DL" + ], + "image_footnote": [], + "bbox": [ + 217, + 205, + 312, + 253 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/cf0b4daf1123d33c2a218cc2ef17368bd3d5b7debf7f945214e537348290a719.jpg", + "image_caption": [ + "TAPE" + ], + "image_footnote": [], + "bbox": [ + 313, + 205, + 406, + 255 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/db3404ecfb3504bd5e7963af1234c9778334eb144ae16d2a97c982209a77c1c6.jpg", + "image_caption": [ + "AirNet" + ], + "image_footnote": [], + "bbox": [ + 408, + 205, + 500, + 255 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/4f5f26f456fe7781fd76bb1d3671a4605f767bc764e50ff705eb4ddedade53c0.jpg", + "image_caption": [ + "PromptIR", + "RAM-PromptIR(Ours)" + ], + "image_footnote": [], + "bbox": [ + 501, + 205, + 594, + 255 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/d13a3871a0aba2898f675c9e3000a4fd6e484e63bba71ac92ccd600a68fb1f1d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 596, + 205, + 689, + 255 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/b1d8b0a8235a3f965c0d9f7781132d8359ca0c48dd7a6d47b378f3875aadf756.jpg", + "image_caption": [ + "GT" + ], + "image_footnote": [], + "bbox": [ + 692, + 205, + 784, + 255 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/d392f4b13367f37112eb00350263198a734a9aa54c82abbe7ec6a13a5feee113.jpg", + "image_caption": [ + "Fig. 7: Derain visual comparison on Rain13k-Test dataset. Zoom in for details.", + "Blurred Image" + ], + "image_footnote": [], + "bbox": [ + 217, + 303, + 312, + 345 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/0da33db259b06b5e0ed7743e91b632843ea4fa696c8afa3a6b133b7f105fed17.jpg", + "image_caption": [ + "NAFNet" + ], + "image_footnote": [], + "bbox": [ + 313, + 303, + 406, + 345 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/09842421063513a58bdd468f900e69b29fe537db7d36982428dd4bf5d1167b08.jpg", + "image_caption": [ + "MPRNet" + ], + "image_footnote": [], + "bbox": [ + 408, + 303, + 500, + 345 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/62f4c2ecf477dbd9baa40f009323f3128f9e24af0734db30bde5edac1e3564df.jpg", + "image_caption": [ + "Restormer" + ], + "image_footnote": [], + "bbox": [ + 501, + 303, + 694, + 345 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/4a78e83d2105f41649b4693be56c563f86e2ff4dbaedeae4f3c2154ed8465f35.jpg", + "image_caption": [ + "RAM-SwinIR (Ours)" + ], + "image_footnote": [], + "bbox": [ + 694, + 303, + 785, + 345 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/f3dedb93749e3aeb4a982bfeb89804f435eb1d9cbbf021d349fdc59878d82874.jpg", + "image_caption": [ + "DL", + "Fig. 8: Motion deblur visual comparison on GoPro dataset. Zoom in for details." + ], + "image_footnote": [], + "bbox": [ + 217, + 357, + 312, + 400 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/f76c879fde6b574c1a8d76f24b8eb05e5854c9141c2a4a8483b5aab6e769707a.jpg", + "image_caption": [ + "TAPE" + ], + "image_footnote": [], + "bbox": [ + 313, + 357, + 406, + 400 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/84d6d2b3e3679eca20a4fa4d761a3ee97088fbde5a6ec0d9fbc09810410e2701.jpg", + "image_caption": [ + "AirNet" + ], + "image_footnote": [], + "bbox": [ + 408, + 357, + 500, + 400 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/7ce785fe1109627eda1e07fefe0c38db60de1f2ecfaa9fc617d95a4b3e3904a5.jpg", + "image_caption": [ + "PromptIR", + "RAM-PromptIR(Ours)" + ], + "image_footnote": [], + "bbox": [ + 501, + 357, + 694, + 400 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "", + "bbox": [ + 596, + 357, + 689, + 400 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/8a8e13692bf76d3b1d7c71dc24f60cd7a08a179450d5da0e7934fd7c5f1fd6e3.jpg", + "image_caption": [ + "GT" + ], + "image_footnote": [], + "bbox": [ + 694, + 357, + 785, + 400 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "ing result at different noise levels. Both RAM-SwinIR and RAM-PromptIR get higher performance than the origin versions.", + "bbox": [ + 212, + 470, + 784, + 501 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Fig. 6-Fig. 10 show the qualitative results of various methods on different datasets. In Fig. 6, our method achieves better dehazing effects (right region) and exposure correction (sky). In the deraining task (Fig. 7, our method better removes rain streaks and restores textures in the occluded regions. In terms of denoising (Fig. 9) and deblurring (Fig. 8), we achieve clearer results with fewer artifacts. We also demonstrate better color correction (the purple blanket on the left) and exposure correction in low-light image enhancement tasks (Fig. 10). For simplicity, the qualitative effects of kernel deblurring and JPEG artifact removal will be presented in the supplementary material.", + "bbox": [ + 212, + 501, + 785, + 638 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "4.3 Ablation Study", + "text_level": 1, + "bbox": [ + 215, + 661, + 388, + 676 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "In this section, we conduct an ablative study on the masking ratio, mask patch size, pre-training strategy, fine-tuning strategy, and fine-tuning ratio to demonstrate the effectiveness of our MIM pre-training and fine-tuning strategy.", + "bbox": [ + 212, + 686, + 784, + 734 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/68e41432cc535332721b48ddce41f341d02b0922b00e0717b07bc5e59d5ffe63.jpg", + "table_caption": [ + "Table 3: Ablative results on masking ratios." + ], + "table_footnote": [], + "table_body": "
Masking ratio20%40%50%60%80%
PSNR↑27.2827.2127.2827.2627.08
SSIM↑0.86630.86830.86980.86940.8642
", + "bbox": [ + 302, + 787, + 700, + 837 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "Qin et al.", + "bbox": [ + 271, + 114, + 336, + 128 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/97121b78211839d4c6f6f66021840ea852b03bd0b35d41ed1651c883dfb89c6f.jpg", + "table_caption": [ + "Table 4: Ablative results of different pre-training strategies." + ], + "table_footnote": [], + "table_body": "
RAM-SwinIRPSNR↑SSIM↑
pre-trained w/ gt26.620.8580
pre-trained w/ paired data27.280.8698
", + "bbox": [ + 215, + 191, + 491, + 251 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/26df439a642d976441ef7d9dbda0bb81afe779c13b1fc89dfbe1e9c305019579.jpg", + "table_caption": [ + "Table 5: Ablative results of different fine-tuning strategies." + ], + "table_footnote": [], + "table_body": "
RAM-SwinIRPSNR↑SSIM↑
random26.860.8535
IG [45]26.920.8554
MAC (Ours)27.280.8698
", + "bbox": [ + 509, + 184, + 767, + 258 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "- Patch size & masking ratio are two essential hyper-parameters that determine the continuity and area of the masking of an image. In high-level tasks, MAE [15] masks $75\\%$ of an image with $16 \\times 16$ patch size. However, it can corrupt the local details of images, which is not suitable for image restoration.", + "bbox": [ + 212, + 289, + 782, + 349 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/242bcdf3f4c7dad4073185430ab17da5ed318b986ac31f1b0f2e03463a5d1a36.jpg", + "table_caption": [ + "Table 6: Ablative results in terms of the PSNR on fine-tuning ratios. We compared the performance in restoring images with unseen noises (Out-of-Distribution Denoising) and known degraded images (In-Distribution). In this case, the settings of In-Distribution are the same as Tab. 1." + ], + "table_footnote": [], + "table_body": "
MethodOut-of-DistributionIn-Distribution
PossionPepperSpeckleAverageAverage
SwinIR [28]12.8310.0020.8614.5626.64
RAM-SwinIR10%13.6719.2321.0717.9927.28
RAM-SwinIR20%13.2719.0920.6817.6827.35
RAM-SwinIR50%12.7516.5120.3616.5427.38
RAM-SwinIR100%12.4715.3120.0115.9327.54
", + "bbox": [ + 218, + 479, + 555, + 584 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "We first find the best choice of patch size by pretraining SwinIR [28] on $1 \\times 1$ , $4 \\times 4$ , and $8 \\times 8$ , as shown in Fig. 3. Since the attention layers of SwinIR treat an $8 \\times 8$ patch as a token, the $4 \\times 4$ pre-training produces heavy artifacts. Besides, the results generated by $8 \\times 8$ pretraining are highly missing details, e.g. the texture of the polar bear's paws. In contrast, the model pre-trained with $1 \\times 1$ patch size, which is also our final choice, achieves a satisfactory reconstruction and removes most of the rain streaks.", + "bbox": [ + 563, + 352, + 787, + 622 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Then, we adjust the masking ratio from $20\\%$ to $80\\%$ . As we can see in Tab. 3, the model pre-trained with $50\\%$ achieves the highest performance. Moreover, the performance is significantly dropped from 27.28dB to 27.08dB in terms of PSNR when we continue to increase the masking ratio, which also demonstrates our opinion that a high masking ratio is harmful to image restoration.", + "bbox": [ + 212, + 625, + 784, + 700 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Pre-trained with paired data. Tab. 4 compares the results of using paired data for mask image pretraining (our pretraining strategy) with those using only ground truth for mask image pretraining. It shows that pre-trained with paired data is necessary for our RAM. Pretraining the model on high-quality images does not effectively enable learning for image restoration tasks. It still requires paired data to guide the model in the learning process.", + "bbox": [ + 212, + 702, + 784, + 792 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Fine-tuning strategy. To verify the effectiveness of our fine-tuning strategy, we fine-tune $10\\%$ of the network layers selected through MAC analysis, IG [45], and uniform sampling, respectively, and the results are shown in Tab. 5. Compared", + "bbox": [ + 212, + 794, + 785, + 840 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "Restore Anything with Masks", + "bbox": [ + 529, + 114, + 730, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/3eb1c6379e11b9c597a9cd4116766d3fd8f7977cbe64dcef33609fe2dfce767f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 217, + 143, + 313, + 191 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/c10fb5aaa480258cf5e68e0dea376870bff31258f391c4586b25767e271fc969.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 313, + 143, + 406, + 191 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/f296c3bf94de43094ed2fad425930485a7c873ae2da714e4892e99635b4fdfe0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 408, + 143, + 501, + 191 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/0ef2171be860cc4c21eb3cc1b06a69c652c0c7c0f72604d763cfc912fbee288a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 501, + 143, + 596, + 191 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/97647f418b561be0583fe5ecf511d7072ed9c2ec05d10a975e8a6148594376e0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 596, + 143, + 691, + 191 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/5cdd7d46956bb4566507ede54decbc9374f95a75e8b1fff6aff1b2eade288dc7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 692, + 143, + 785, + 191 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/5ab121741bc61962925984b993557e0bb4a60d5a3decbcdc2ba1f198c968b5e1.jpg", + "image_caption": [ + "Noisy Image", + "DL" + ], + "image_footnote": [], + "bbox": [ + 217, + 204, + 312, + 253 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/dab4a93d4ef7270c882be456133b1181b65981ec386f39208741bb1fb6531885.jpg", + "image_caption": [ + "NAFNet", + "TAPE" + ], + "image_footnote": [], + "bbox": [ + 313, + 204, + 406, + 253 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/a8840f4335f7694ffd376fa7ad27df1e67c961863c03c3963a1436609ab92c4a.jpg", + "image_caption": [ + "MPRNet", + "AirNet" + ], + "image_footnote": [], + "bbox": [ + 408, + 204, + 501, + 253 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/39214ef1d42f254db74bdbe602a7fbfaffbbb82e9594e195d8848aea9085d740.jpg", + "image_caption": [ + "Restormer", + "PromptIR", + "RAM-PromptIR(Ours)" + ], + "image_footnote": [], + "bbox": [ + 501, + 204, + 596, + 253 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/561a7269f3e4b6dfc33ee010a43da46c8764a9bc63ce403b3ac342a8e1e2895a.jpg", + "image_caption": [ + "SwinIR" + ], + "image_footnote": [], + "bbox": [ + 596, + 204, + 689, + 253 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/d399dc62949040a87f8d79296eecf419751f802a64644d15b90245742af82a31.jpg", + "image_caption": [ + "RAM-SwinIR (Ours)", + "GT" + ], + "image_footnote": [], + "bbox": [ + 692, + 204, + 785, + 253 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/ae9e45b9f67927b8d0f715be350c576819dc7302813b65df51d73b399e403d4b.jpg", + "image_caption": [ + "Fig. 9: Denoising visual comparison on CBSD68 dataset. Zoom in for details." + ], + "image_footnote": [], + "bbox": [ + 217, + 306, + 312, + 357 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/a91afc1ef3fa620d198399a06534bd079b75325f645256e0405a3a982694e344.jpg", + "image_caption": [ + "NAFNet" + ], + "image_footnote": [], + "bbox": [ + 313, + 306, + 406, + 357 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/e38c71cd938d93468c2150fceb3c73455849a23f02e3cc1de139c6f4dc27067b.jpg", + "image_caption": [ + "MPRNet" + ], + "image_footnote": [], + "bbox": [ + 408, + 306, + 501, + 357 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/0baa7a8329853d6f7e5293d32c53e9c4f9079b212dc4ea4bd9c700e37f0aa3a5.jpg", + "image_caption": [ + "Restormer" + ], + "image_footnote": [], + "bbox": [ + 501, + 306, + 596, + 357 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/a8db619b831cfdc3c4be0ddbb22abd61876e69ed84c00da1e2f00e194e05471c.jpg", + "image_caption": [ + "SwinIR" + ], + "image_footnote": [], + "bbox": [ + 596, + 306, + 689, + 357 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/8079c7c696c821a012b907fdb96052ca3bdec04c3eeb189f09b69deb2b549c22.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 692, + 306, + 785, + 357 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/a3b2f6aa0992184b49931ed447c7f5dc813ebfe3963c0dcaf856825c35052b82.jpg", + "image_caption": [ + "Low Light Image", + "DL", + "Fig. 10: LLIE visual comparison on LOL dataset. Zoom in for details." + ], + "image_footnote": [], + "bbox": [ + 215, + 369, + 312, + 417 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/cada7c403bec112a9934f1a6819f52ae2f81868d1b3cf4ece8d1bb7346599167.jpg", + "image_caption": [ + "TAPE" + ], + "image_footnote": [], + "bbox": [ + 313, + 369, + 406, + 417 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/d2881532b49bd7489af5a3b6db650445f1bc2fc21c9973bbb4d0b40ab7aed783.jpg", + "image_caption": [ + "AirNet" + ], + "image_footnote": [], + "bbox": [ + 408, + 369, + 501, + 417 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/92e68fd5b163f2eb7728ab0a711655879deb41991c42e295532c7f2c6929dd71.jpg", + "image_caption": [ + "PromptIR", + "RAM-PromptIR(Ours)" + ], + "image_footnote": [], + "bbox": [ + 501, + 369, + 596, + 417 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/bc41aa406fd82a4b3a47980b6bcff8da394b2452a3049e6f7413464c9ac7b4fa.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 596, + 369, + 689, + 417 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/fd5ad5d957a22397858bfc0f8407e2ff4c70e4f553648befb592017ea15812cc.jpg", + "image_caption": [ + "RAM-SwinIR (Ours)", + "GT" + ], + "image_footnote": [], + "bbox": [ + 692, + 369, + 785, + 417 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "to IG, we have improved by 0.36 dB in PSNR and $1.6\\%$ in SSIM, which indicates that our selection strategy is superior to IG.", + "bbox": [ + 212, + 484, + 784, + 513 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Fine-tune ratio. We conduct the ablation experiment to compare the network's performances with different fine-tune ratios in Tab. 6. We found that using our finetune strategy, a pre-trained network could achieve comparable performance by fine-tuning only a few layers (e.g. $10\\%$ ). At the same time, we need to finetune almost all network parameters to get the best performance on given tasks.", + "bbox": [ + 212, + 515, + 784, + 589 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Performance vs Generalization capability. We found a trade-off between in-distribution performance and out-of-distribution generalization in Tab. 6. We found that the more layers fine-tuned, the less generalization capability to tackle the out-of-distribution tasks. With our fine-tuning method, the model can have stronger generalization while maintaining comparable performance.", + "bbox": [ + 212, + 590, + 784, + 665 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 688, + 357, + 704 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "This paper presents RAM, a pipeline for extracting intrinsic image information from corrupted images using Mask Image Modeling (MIM) pre-training. We design a MIM pre-training strategy tailored for image restoration and a fine-tuning algorithm to handle the transition from masked to complete images. By analyzing layer importance with MAC, we achieve high performance with minimal parameter tuning. Extensive experiments demonstrate that our RAM can bring boosts to various architectures and achieve state-of-the-art performance, moving towards a unified solution for all-in-one image restoration.", + "bbox": [ + 212, + 719, + 787, + 839 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "Qin et al.", + "bbox": [ + 271, + 114, + 336, + 128 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 215, + 143, + 321, + 159 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Chen, H., Wang, Y., Guo, T., Xu, C., Deng, Y., Liu, Z., Ma, S., Xu, C., Xu, C., Gao, W.: Pre-trained image processing transformer. In: CVPR. pp. 12299-12310 (2021)", + "2. Chen, H., Gu, J., Liu, Y., Magid, S.A., Dong, C., Wang, Q., Pfister, H., Zhu, L.: Masked image training for generalizable deep image denoising. In: CVPR. pp. 1692-1703 (2023)", + "3. Chen, L., Chu, X., Zhang, X., Sun, J.: Simple baselines for image restoration. In: ECCV. pp. 17-33. Springer (2022)", + "4. Chen, W.T., Huang, Z.K., Tsai, C.C., Yang, H.H., Ding, J.J., Kuo, S.Y.: Learning multiple adverse weather removal via two-stage knowledge learning and multi-contrastive regularization: Toward a unified model. In: CVPR. pp. 17653-17662 (2022)", + "5. Chen Wei, Wenjing Wang, W.Y.J.L.: Deep retina decomposition for low-light enhancement. In: BMVC. British Machine VLOLision Association (2018)", + "6. Dhamdhere, K., Sundararajan, M., Yan, Q.: How important is a neuron. In: ICLR (2019)", + "7. Duan, H., Shen, W., Min, X., Tu, D., Teng, L., Wang, J., Zhai, G.: Masked autoencoders as image processors. arXiv preprint arXiv:2303.17316 (2023)", + "8. Fan, Q., Chen, D., Yuan, L., Hua, G., Yu, N., Chen, B.: A general decoupled learning framework for parameterized image operators. PAMI 43(1), 33-47 (2019)", + "9. Fang, Y., Zhang, H., Wong, H.S., Zeng, T.: A robust non-blind deblurring method using deep denoiser prior. In: CVPRW. pp. 735-744 (June 2022)", + "0. Fu, X., Huang, J., Ding, X., Liao, Y., Paisley, J.: Clearing the skies: A deep network architecture for single-image rain removal. TIP 26(6), 2944-2956 (2017)", + "1. Fu, X., Huang, J., Zeng, D., Huang, Y., Ding, X., Paisley, J.: Removing rain from single images via a deep detail network. In: CVPR. pp. 3855-3863 (2017)", + "2. Gu, J., Dong, C.: Interpreting super-resolution networks with local attribution maps. In: CVPR. pp. 9199-9208 (2021)", + "3. Guo, C.L., Yan, Q., Anwar, S., Cong, R., Ren, W., Li, C.: Image dehazing transformer with transmission-aware 3d position embedding. In: CVPR. pp. 5812-5820 (2022)", + "4. Guo, C., Li, C., Guo, J., Loy, C.C., Hou, J., Kwong, S., Cong, R.: Zero-reference deep curve estimation for low-light image enhancement. In: CVPR. pp. 1780-1789 (2020)", + "5. He, K., Chen, X., Xie, S., Li, Y., Dollár, P., Girshick., R.: Masked autoencoders are scalable vision learners. In: CVPR. pp. 16000-16009 (2022)", + "6. Huang, J.B., Singh, A., Ahuja, N.: Single image super-resolution from transformed self-exemplars. In: CVPR. pp. 5197-5206 (2015)", + "7. Jin, X., Han, L.H., Li, Z., Guo, C.L., Chai, Z., Li, C.: Dnf: Decouple and feedback network for seeing in the dark. In: CVPR. pp. 18135-18144 (2023)", + "8. Kenton, J.D.M.W.C., Toutanova, L.K.: Bert: Pre-training of deep bidirectional transformers for language understanding. In: NAACL. pp. 4171-4186 (2019)", + "9. Leino, K., Sen, S., Datta, A., Fredrikson, M., Li, L.: Influence-directed explanations for deep convolutional networks. In: ITC. pp. 1-8. IEEE (2018)", + "20. Li, B., Ren, W., Fu, D., Tao, D., Feng, D., Zeng, W., Wang, Z.: Benchmarking single-image dehazing and beyond. TIP 28(1), 492-505 (2018)", + "21. Li, B., Liu, X., Hu, P., Wu, Z., Lv, J., Peng, X.: All-in-one image restoration for unknown corruption. In: CVPR. pp. 17452-17462 (2022)" + ], + "bbox": [ + 225, + 176, + 784, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "Restore Anything with Masks", + "bbox": [ + 529, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 785, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "22. Li, C., Guo, C.L., Liang, Z., Zhou, S., Feng, R., Loy, C.C., et al.: Embedding fourier for ultra-high-definition low-light image enhancement. In: ICLR (2022)", + "23. Li, D., Zhang, Y., Cheung, K.C., Wang, X., Qin, H., Li, H.: Learning degradation representations for image deblurring. In: ECCV. pp. 736-753. Springer (2022)", + "24. Li, R., Tan, R.T., Cheong, L.F.: All in one bad weather removal using architectural search. In: CVPR. pp. 3175-3185 (2020)", + "25. Li, X., Wu, J., Lin, Z., Liu, H., Zha, H.: Recurrent squeeze-and-excitation context aggregation net for single image deraining. In: ECCV. pp. 254-269 (2018)", + "26. Li, Y., Zhang, K., Liang, J., Cao, J., Liu, C., Gong, R., Zhang, Y., Tang, H., Liu, Y., Demandolx, D., et al.: Lsdir: A large scale dataset for image restoration. In: CVPR. pp. 1775-1787 (2023)", + "27. Li, Y., Tan, R.T., Guo, X., Lu, J., Brown, M.S.: Rain streak removal using layer priors. In: CVPR", + "28. Liang, J., Cao, J., Sun, G., Zhang, K., Van Gool, L., Timofte, R.: Swinir: Image restoration using swin transformer. In: CVPR. pp. 1833-1844 (2021)", + "29. Lin, X., Ren, C., Liu, X., Huang, J., Lei, Y.: Unsupervised image denoising in real-world scenarios via self-collaboration parallel generative adversarial branches. In: ICCV. pp. 12642-12652 (2023)", + "30. Lin, X., Yue, J., Ren, C., Guo, C.L., Li, C.: Unlocking low-light-rainy image restoration by pairwise degradation feature vector guidance. arXiv preprint arXiv:2305.03997 (2023)", + "31. Liu, L., Xie, L., Zhang, X., Yuan, S., Chen, X., Zhou, W., Li, H., Tian, Q.: Tape: Task-agnostic prior embedding for image restoration. In: ECCV. pp. 447-464. Springer (2022)", + "32. Liu, Y., He, J., Gu, J., Kong, X., Qiao, Y., Dong, C.: Degae: A new pretraining paradigm for low-level vision. In: CVPR. pp. 23292-23303 (2023)", + "33. Liu, Y., Liu, A., Gu, J., Zhang, Z., Wu, W., Qiao, Y., Dong, C.: Discovering distinctive \"semantics\" in super-resolution networks. arXiv preprint arXiv:2108.00406 (2021)", + "34. Liu, Z., Lin, Y., Cao, Y., Hu, H., Wei, Y., Zhang, Z., Lin, S., Guo, B.: Swin transformer: Hierarchical vision transformer using shifted windows. In: CVPR. pp. 10012-10022 (2021)", + "35. Luo, Y., Xu, Y., Ji, H.: Removing rain from a single image via discriminative sparse coding. In: ICCV. pp. 3397-3405 (2015)", + "36. Luo, Y., Zhao, R., Wei, X., Chen, J., Lu, Y., Xie, S., Wang, T., Xiong, R., Lu, M., Zhang, S.: Mowe: mixture of weather experts for multiple adverse weather removal. arXiv preprint arXiv:2303.13739 (2023)", + "37. Magid, S.A., Lin, Z., Wei, D., Zhang, Y., Gu, J., Pfister, H.: Texture-based error analysis for image super-resolution. In: CVPR. pp. 2118-2127 (2022)", + "38. Martin, D., Fowlkes, C., Tal, D., Malik, J.: A database of human segmented natural images and its application to evaluating segmentation algorithms and measuring ecological statistics. In: ICCV. vol. 2, pp. 416-423. IEEE (2001)", + "39. Mehri, A., Ardakani, P.B., Sappa, A.D.: Mprnet: Multi-path residual network for lightweight image super resolution. In: CVPR. pp. 2704-2713 (2021)", + "40. Nah, S., Hyun Kim, T., Mu Lee, K.: Deep multi-scale convolutional neural network for dynamic scene deblurring. In: CVPR", + "41. Park, D., Lee, B.H., Chun, S.Y.: All-in-one image restoration for unknown degradations using adaptive discriminative filters for specific degradations. In: CVPR. pp. 5815-5824 (2023)", + "42. Potlapalli, V., Zamir, S.W., Khan, S.H., Shahbaz Khan, F.: Prompt for all-in-one image restoration. NeurIPS 36 (2024)" + ], + "bbox": [ + 215, + 146, + 784, + 840 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "Qin et al.", + "bbox": [ + 271, + 114, + 336, + 128 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "43. Radford, A., Narasimhan, K., Salimans, T., Sutskever, I., et al.: Improving language understanding by generative pre-training (2018)", + "44. Shrikumar, A., Su, J., Kundaje, A.: Computationally efficient measures of internal neuron importance. arXiv preprint arXiv:1807.09946 (2018)", + "45. Sundararajan, M., Taly, A., Yan, Q.: Axiomatic attribution for deep networks. In: ICML. pp. 3319-3328. PMLR (2017)", + "46. Sundararajan, M., Taly, A., Yan, Q.: Gradients of counterfactuals. ICLR (2017)", + "47. Wang, X., Wang, W., Cao, Y., Shen, C., Huang, T.: Images speak in images: A generalist painter for in-context visual learning. In: CVPR. pp. 6830-6839 (2023)", + "48. Wang, Z., Cun, X., Bao, J., Zhou, W., Liu, J., Li, H.: Uformer: A general u-shaped transformer for image restoration. In: CVPR. pp. 17683-17693 (2022)", + "49. Wu, R.Q., Duan, Z.P., Guo, C.L., Chai, Z., Li, C.: Ridcp: Revitalizing real image dehazing via high-quality codebook priors. In: CVPR. pp. 22282-22291 (2023)", + "50. Xie, L., Wang, X., Dong, C., Qi, Z., Shan, Y.: Finding discriminative filters for specific degradations in blind super-resolution. NeurIPS 34, 51-61 (2021)", + "51. Xie, Z., Zhang, Z., Cao, Y., Lin, Y., Bao, J., Yao, Z., Dai, Q., Hu, H.: Simmim: a simple framework for masked image modeling. In: CVPR. pp. 9653-9663 (2022)", + "52. Yang, W., Tan, R.T., Feng, J., Liu, J., Guo, Z., Yan, S.: Deep joint rain detection and removal from a single image. In: CVPR. pp. 1357-1366 (2017)", + "53. Yang, W., Tan, R.T., Wang, S., Fang, Y., Liu, J.: Single image deraining: From model-based to data-driven and beyond. PAMI 43(11), 4059-4077 (2020)", + "54. Yang, W., Wang, W., Huang, H., Wang, S., Liu, J.: Sparse gradient regularized deep retina network for robust low-light image enhancement. TIP 30, 2072-2086 (2021)", + "55. Zamir, S.W., Arora, A., Khan, S., Hayat, M., Khan, F.S., Yang, M.H.: Restormer: Efficient transformer for high-resolution image restoration. In: CVPR. pp. 5728-5739 (2022)", + "56. Zhang, C., Zhu, Y., Yan, Q., Sun, J., Zhang, Y.: All-in-one multi-degradation image restoration network via hierarchical degradation representation. In: ACMMM. pp. 2285-2293 (2023)", + "57. Zhang, H., Patel, V.M.: Density-aware single image de-raining using a multi-stream dense network. In: CVPR. pp. 695-704 (2018)", + "58. Zhang, H., Sindagi, V., Patel, V.M.: Image de-raining using a conditional generative adversarial network. TCSVT 30(11), 3943-3956 (2019)", + "59. Zhang, J., Huang, J., Yao, M., Yang, Z., Yu, H., Zhou, M., Zhao, F.: Ingredient-oriented multi-degradation learning for image restoration. In: CVPR. pp. 5825-5835 (2023)", + "60. Zheng, N., Zhou, M., Dong, Y., Rui, X., Huang, J., Li, C., Zhao, F.: Empowering low-light image enhancer through customized learnable priors. In: ICCV. pp. 12559-12569 (2023)", + "61. Zhu, Y., Wang, T., Fu, X., Yang, X., Guo, X., Dai, J., Qiao, Y., Hu, X.: Learning weather-general and weather-specific features for image restoration under multiple adverse weather conditions. In: CVPR. pp. 21747-21758 (2023)" + ], + "bbox": [ + 212, + 146, + 787, + 742 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "Restore Anything with Masks", + "bbox": [ + 531, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + } +] \ No newline at end of file diff --git a/2024/Restore Anything with Masks_ Leveraging Mask Image Modeling for Blind All-in-One Image Restoration/0de4c62c-e2d6-4f18-8a4e-c05352d6d956_model.json b/2024/Restore Anything with Masks_ Leveraging Mask Image Modeling for Blind All-in-One Image Restoration/0de4c62c-e2d6-4f18-8a4e-c05352d6d956_model.json new file mode 100644 index 0000000000000000000000000000000000000000..6b1b1bf42fadc3bea04415f5de6a45ce3eb3fd7f --- /dev/null +++ b/2024/Restore Anything with Masks_ Leveraging Mask Image Modeling for Blind All-in-One Image Restoration/0de4c62c-e2d6-4f18-8a4e-c05352d6d956_model.json @@ -0,0 +1,3622 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.228, + 0.141, + 0.78, + 0.207 + ], + "angle": 0, + "content": "Restore Anything with Masks: Leveraging Mask Image Modeling for Blind All-in-One Image Restoration" + }, + { + "type": "text", + "bbox": [ + 0.305, + 0.234, + 0.697, + 0.266 + ], + "angle": 0, + "content": "Chu-Jie Qin\\(^{1,2*}\\), Rui-Qi Wu\\(^{1,2}\\), Zukun Liu\\(^{3}\\), Xin Lin\\(^{5}\\), Chun-Le Guo\\(^{1,2}\\), Hyun Hee Park\\(^{4}\\), and Chongyi Li\\(^{1,2\\dagger}\\)" + }, + { + "type": "text", + "bbox": [ + 0.397, + 0.277, + 0.608, + 0.291 + ], + "angle": 0, + "content": "\\(^{1}\\) VCIP, CS, Nankai University" + }, + { + "type": "text", + "bbox": [ + 0.405, + 0.291, + 0.598, + 0.305 + ], + "angle": 0, + "content": "2 NKIARI, Shenzhen Futian" + }, + { + "type": "text", + "bbox": [ + 0.351, + 0.306, + 0.653, + 0.319 + ], + "angle": 0, + "content": "{chujie.qin,wuruiqi}@mail.nankai.edu.cn" + }, + { + "type": "text", + "bbox": [ + 0.364, + 0.321, + 0.642, + 0.334 + ], + "angle": 0, + "content": "{guochunle, lichongyi}@nankai.edu.cn" + }, + { + "type": "text", + "bbox": [ + 0.349, + 0.333, + 0.654, + 0.348 + ], + "angle": 0, + "content": "3 Samsung Research, China, Beijing (SRC-B)" + }, + { + "type": "text", + "bbox": [ + 0.269, + 0.348, + 0.735, + 0.375 + ], + "angle": 0, + "content": "4 The Department of Camera Innovation Group, Samsung Electronics {zikun.liu, inextg.park}@samsung.com" + }, + { + "type": "text", + "bbox": [ + 0.431, + 0.375, + 0.573, + 0.389 + ], + "angle": 0, + "content": "Sichuan University" + }, + { + "type": "text", + "bbox": [ + 0.42, + 0.391, + 0.584, + 0.402 + ], + "angle": 0, + "content": "linxin@stu.scu.edu.cn" + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.432, + 0.744, + 0.697 + ], + "angle": 0, + "content": "Abstract. All-in-one image restoration aims to handle multiple degradation types using one model. This paper proposes a simple pipeline for all-in-one blind image restoration to Restore Anything with Masks (RAM). We focus on the image content by utilizing Mask Image Modeling to extract intrinsic image information rather than distinguishing degradation types like other methods. Our pipeline consists of two stages: masked image pre-training and fine-tuning with mask attribute conductance. We design a straightforward masking pre-training approach specifically tailored for all-in-one image restoration. This approach enhances networks to prioritize the extraction of image content priors from various degradations, resulting in a more balanced performance across different restoration tasks and achieving stronger overall results. To bridge the gap of input integrity while preserving learned image priors as much as possible, we selectively fine-tuned a small portion of the layers. Specifically, the importance of each layer is ranked by the proposed Mask Attribute Conductance (MAC), and the layers with higher contributions are selected for finetuning. Extensive experiments demonstrate that our method achieves state-of-the-art performance. Our code and model will be released at https://github.com/Dragonisss/RAM." + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.708, + 0.721, + 0.723 + ], + "angle": 0, + "content": "Keywords: Image Restoration \\(\\cdot\\) All-in-One \\(\\cdot\\) Mask Image Modeling" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.745, + 0.376, + 0.76 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.773, + 0.788, + 0.805 + ], + "angle": 0, + "content": "Image restoration involves the restoration of low-quality images affected by various degradation, typically arising from adverse environmental conditions (e.g.," + }, + { + "type": "page_footnote", + "bbox": [ + 0.238, + 0.811, + 0.728, + 0.826 + ], + "angle": 0, + "content": "* A part of this work is done during Chu-Jie Qin's internship at Samsung." + }, + { + "type": "page_footnote", + "bbox": [ + 0.241, + 0.826, + 0.514, + 0.84 + ], + "angle": 0, + "content": "†Chongyi Li is the corresponding author." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.338, + 0.129 + ], + "angle": 0, + "content": "Qin et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.208 + ], + "angle": 0, + "content": "rain, haze, low-light), hardware-related issues (e.g., noise and blur), and post-processing artifacts (e.g., JPEG compression). Image restoration serves not only to enhance the visual appeal of images but also contributes to practical application scenarios such as autonomous driving and surveillance." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.212, + 0.511, + 0.424 + ], + "angle": 0, + "content": "Modern techniques in this field mainly focus on learning fixed patterns formed during the degradation process, i.e., degradation priors. Some works [29, 30, 60] utilize task-specific priors to solve a certain degradation problem, while another research line [3, 28, 39, 48, 55] tries to design a general network architecture that can effectively learn each degradation pattern. Nevertheless, the above methods only enable the network to learn a single degradation, resulting in an imbalanced situation when dealing with multiple types of degradation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.428, + 0.512, + 0.549 + ], + "angle": 0, + "content": "To tackle the problem stated above, all-in-one methods have emerged, aiming to handle multiple degradations using one model. Most of these approaches tend to utilize explicit priors (e.g., AirNet [21]) or introduce an extra module (e.g., PromptIR [42]) to discern image degradation patterns, thereby assisting" + }, + { + "type": "image", + "bbox": [ + 0.522, + 0.242, + 0.788, + 0.432 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.52, + 0.442, + 0.788, + 0.513 + ], + "angle": 0, + "content": "Fig.1: Our RAM achieves more balanced and more powerful performance than the state-of-the-art methods (AirNet [21], TAPE [31], PromptIR [42]) for all-in-one blind image restoration." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.549, + 0.788, + 0.701 + ], + "angle": 0, + "content": "the model in performing the restoration. However, these methods place their emphasis on distinguishing degradation types in images rather than the image content, leading to lower scalability and fuzzy decision boundaries when more degradation types are involved. We argue that the essence of image restoration is to extract intrinsic image information from corrupted images rather than eliminate degradation patterns, i.e., learning image prior rather than degradation prior. It is worth noting that TAPE [31] similarly suggests that understanding normal image nature aids restoration by introducing a natural image prior. Nevertheless, TAPE utilizes the model output as the optimization target, which causes the model to amplify its own errors and learn the image prior with bias." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.705, + 0.788, + 0.842 + ], + "angle": 0, + "content": "In this paper, we focus on tackling how to extract intrinsic image information from diverse corrupted images. Some attempts [2, 7] by Mask Image Modeling (MIM) in low-level vision have caught our attention. As a pretraining strategy, MIM has been widely validated for its effectiveness in high-level tasks, thanks to its generic representation of images. Simultaneously, the model also learns the distribution of natural images, which encompasses the intrinsic information we aim to extract from the images. Built on MIM, we propose a simple pipeline for all-in-one blind image restoration that Restores Anything with Masks (RAM), which includes two stages: the mask pre-training stage and the" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.531, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Restore Anything with Masks" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.253 + ], + "angle": 0, + "content": "fine-tuning stage with Mask Attribute Conductance (MAC). In the pre-training stage, we randomly mask corrupted images at the pixel-wise level and force the network to predict the clear one corresponding to the masked pixels, extracting inherent image information from corrupted images. In the fine-tuning stage, we focus on overcoming the input integrity gap caused by changing masked input during pre-training into the whole image during inference while preserving learned prior as much as possible." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.253, + 0.785, + 0.373 + ], + "angle": 0, + "content": "Specifically, we first evaluated the importance of each network layer in addressing this gap by the proposed MAC. Following that, we chose the top \\( k\\% \\) most critical layers for fine-tuning while keeping the rest of the network layers frozen. We demonstrate that after a brief fine-tuning period (even if only \\( 10\\% \\) layers are tuned), the model can achieve a highly satisfactory performance level, surpassing models trained using traditional pair-wise training. Additionally, our pipeline can be plug-and-play used in any network without introducing additional computational overhead." + }, + { + "type": "text", + "bbox": [ + 0.24, + 0.374, + 0.568, + 0.388 + ], + "angle": 0, + "content": "The contributions of this work are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.398, + 0.784, + 0.458 + ], + "angle": 0, + "content": "- We discuss the challenge of adopting MIM in low-level vision and propose a MIM-based pre-training strategy tailored to all-in-one blind image restoration, which allows the restoration networks to effectively learn inherent image information while guaranteeing reconstruction results." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.459, + 0.785, + 0.519 + ], + "angle": 0, + "content": "- We proposed Mask Attribute Conductance to evaluate the importance of each layer in addressing the input integrity gap so that a very small portion (e.g. \\(10\\%\\)) of critical layers are tuned to bridge this gap while preserving the image prior learned by MIM." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.52, + 0.784, + 0.595 + ], + "angle": 0, + "content": "- Our proposed RAM provides a fresh perspective to achieve more balanced and powerful all-in-one blind image restoration, which focuses on extracting inherent image information from corrupted images. Our pipeline can be applied to any image restoration network without introducing additional computational overhead." + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.398, + 0.785, + 0.595 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.618, + 0.388, + 0.634 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.65, + 0.611, + 0.665 + ], + "angle": 0, + "content": "2.1 Image Restoration for Multi Degradations" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.788, + 0.841 + ], + "angle": 0, + "content": "While neural networks have demonstrated impressive performance in single degradation image restoration [9, 13, 14, 17, 22, 23, 29, 30, 49, 60], recent works have shifted their focus towards addressing the more challenging domain of multi-degradation image restoration. A group of methods [3, 28, 39, 48, 55] aims at designing a general architecture that can effectively learn each degradation pattern. SwinIR [28] employs a window attention mechanism to convert global attention into a localized approach, effectively reducing computational overhead. In addition, the U-shaped transformer-based methods [48, 55] are employed to extract multi-scale features and reduce computational overhead. However, these methods have to train individually on each restoration task. Several methods [1, 24] leverage multiple input and output heads to empower the network to restore" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.338, + 0.129 + ], + "angle": 0, + "content": "Qin et al." + }, + { + "type": "image", + "bbox": [ + 0.235, + 0.145, + 0.773, + 0.322 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.332, + 0.789, + 0.445 + ], + "angle": 0, + "content": "Fig. 2: The illumination of our overall pipeline. 1) Pre-training the model with mask image pre-training method tailored to low-level vision. We randomly mask degraded images at the pixel level with a \\(50\\%\\) masking ratio and reconstruct the clean images. 2) The Fine-tuning stage is followed to overcome the input integrity gap caused by changing masked input during pre-training into the whole image during inference. We analyze the importance of each network layer for resolving the input integrity gap according to the proposed MAC and rank them in descending order. The top \\(k\\%\\) of network layers are selected for fine-tuning on the complete image." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.474, + 0.789, + 0.701 + ], + "angle": 0, + "content": "various types of degraded images. Nonetheless, this kind of approach may lead to the diminished scalability of the model. Recently, several subsequent methods [4, 21, 36, 41, 42, 56, 61] have been proposed to employ a unified network to address multiple restoration issues. Most of these methods put emphasis on learning how to distinguish different types of degradations and restore corrupted images. Typically, AirNet [21] first proposed an all-in-one image restoration task. The method initially pretrains a degradation classifier based on contrastive learning and subsequently utilizes it to assist in all-in-one image restoration. PromptIR [42] has introduced a learnable prompt-based module. Instead of constraining the degradation category, it enables the model to autonomously learn features that are advantageous to its performance by using an adaptive prompt. Our RAM takes a fresh perspective that focuses on extracting common content information from corrupted images, without any extra design to distinguish degradations, which helps us achieve balance and powerful performance when more degradation types are taken into consideration." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.724, + 0.449, + 0.74 + ], + "angle": 0, + "content": "2.2 Mask Image Modeling" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.749, + 0.788, + 0.843 + ], + "angle": 0, + "content": "Inspired by Mask Language Modeling [18,43], Mask Image Modeling (MIM) [15, 51] is introduced as a pretraining approach to learn general representations in high-level vision. MAE [15] effectively utilizes MIM for predicting hidden tokens, demonstrating strong performance and generalization across various downstream tasks. SimMIM [51] proposed a general masked image modeling method based on Swin-ViT [34]. Painter [47] unifies multiple tasks under image-to-image trans" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.531, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Restore Anything with Masks" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.36 + ], + "angle": 0, + "content": "lation and leverages MIM pretraining. In recent years, there have been efforts to incorporate MIM into the realm of low-level vision to enhance model generalization. Among them, [2] and [7] are the most closely aligned with our focus. [2] employs the MIM model to enhance the model's generalization for denoising tasks but has not explored its potential in multi-task scenarios. [7] utilizes MIM for pre-training the model encoder to introduce generative prior and subsequently employs the decoder for restoration. However, it does not fully harness the potential of MIM. Our proposed RAM utilizes MIM to unify the optimization objective for various image restoration tasks into reconstructing intrinsic image information. This allows the network to learn restoration functions more balanced and effectively. Moreover, to preserve the image priors learned by MIM, we designed a fine-tuning strategy based on MAC analysis (in Sec. 3.3). This enables us to achieve comparable performance by fine-tuning only a small portion (e.g. \\(10\\%\\) ) of layers, fully tapping into the potential of MIM." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.381, + 0.49, + 0.396 + ], + "angle": 0, + "content": "2.3 Gradient-based Attribution" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.406, + 0.789, + 0.588 + ], + "angle": 0, + "content": "Gradient-based attribution methods [6,12,44-46,50] are often used to clarify how hidden units (or inputs) impact the output of networks. One commonly used approach is Integrated Gradients (IG) [45,46], which accumulates gradients along a linear path from the baseline input to the target input in the pixel/feature space. After that, IntInf [19] and layer conductance [6] alter IG to attribute neuron importance along the same path. In our work, we expect to find the key layers that can effectively overcome the distribution shift between training data and inference data. We propose Mask Attribute Conductance (MAC) based on the layer conductance and accumulated MAC of each layer along the Mask Attribute Path (MAP). MAC can represent the layer's importance along the MAP. In this way, we can fine-tune the top \\( k\\% \\) critical layers of the pre-trained network, preserving to a great extent the image priors learned during pretraining." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.61, + 0.381, + 0.628 + ], + "angle": 0, + "content": "3 Methodology" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.642, + 0.788, + 0.704 + ], + "angle": 0, + "content": "In this section, we start with discussing the challenges of using MIM in low-level vision tasks (Sec. 3.1). Following that, we present our pipeline for all-in-one blind image restoration, which contains two parts: pre-training with MIM (Sec. 3.2) and fine-tuning with Mask Attribute Conductance (MAC) Analysis (Sec. 3.3)." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.725, + 0.573, + 0.74 + ], + "angle": 0, + "content": "3.1 Rethinking MIM in Low-Level Vision" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.841 + ], + "angle": 0, + "content": "MIM is a process that randomly masks certain parts of an image and extracts features from the remaining visible parts to reconstruct the entire image. It allows models to acquire a generic representation of images and thus achieve good pre-training, which is verified in many high-level tasks [15, 51]. Moreover, the models also learn the distribution of natural images during the image reconstruction, i.e. MIM pre-training. This incidental acquisition of prior knowledge" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.338, + 0.129 + ], + "angle": 0, + "content": "Qin et al." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.145, + 0.396, + 0.214 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.262, + 0.216, + 0.348, + 0.228 + ], + "angle": 0, + "content": "patch size=1" + }, + { + "type": "image", + "bbox": [ + 0.418, + 0.144, + 0.593, + 0.213 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.457, + 0.216, + 0.545, + 0.228 + ], + "angle": 0, + "content": "patch size=4" + }, + { + "type": "image", + "bbox": [ + 0.612, + 0.145, + 0.785, + 0.213 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.653, + 0.216, + 0.74, + 0.228 + ], + "angle": 0, + "content": "patch size=8" + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.239, + 0.787, + 0.281 + ], + "angle": 0, + "content": "Fig. 3: Mask Image Modeling reconstruction with different patch sizes. We pre-trained with different patch sizes and visualized the mask inputs (left), and the corresponding MIM reconstructions (right)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.317, + 0.784, + 0.363 + ], + "angle": 0, + "content": "is instrumental in tasks like image restoration. Despite these advantages, applying MIM in pretraining a model for low-level vision tasks is still under-explored, primarily due to the challenges that must be addressed in the process." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.365, + 0.785, + 0.485 + ], + "angle": 0, + "content": "Firstly, the main purpose of vanilla MIM is not high-quality reconstruction but good feature extraction for high-level tasks. Therefore, it masks a wider range of images to gather semantic information but not pixel-level content, reflected in token-level masking and a high mask ratio. CSFormer [7] directly adopts this strategy on low-level vision pre-training. However, some studies verify that semantic information is not as important for image restoration as it is in pattern recognition tasks [33, 37]. Moreover, high-degree masking leads to producing detail-deficient results, as shown in Fig. 3, which is harmful to low-level tasks." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.488, + 0.787, + 0.624 + ], + "angle": 0, + "content": "Secondly, the training objective of MIM is to reconstruct the masked input images, so it can only produce results with the same domain as the input image. However, we hope the model gains the ability to bridge low-quality domain to high-quality domain, i.e. recover clean content from degraded input. Therefore, it is necessary to introduce paired data when pre-training image restoration models by MIM (see the experiment in Sec. 4.3 for details). Chen et al. [2] demonstrate that pair-wise MIM training enhances the generalization performance over different types of noisy images. In this paper, we take a step forward to explore the effectiveness of MIM on multiple degradations with larger variance." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.66, + 0.45, + 0.675 + ], + "angle": 0, + "content": "3.2 Pretraining with MIM" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.699, + 0.785, + 0.729 + ], + "angle": 0, + "content": "Based on the above analysis, we design a MIM pre-training paradigm tailored for low-level vision." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.732, + 0.785, + 0.793 + ], + "angle": 0, + "content": "Masking. During the pre-training stage, we randomly mask the pixels of degraded images (mask images in a \\(1 \\times 1\\) patch size) with a \\(50\\%\\) mask ratio. We found that fine-grained masked patches and balanced mask ratio are beneficial to image restoration, which can be demonstrated in Sec. 4.3." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.785, + 0.84 + ], + "angle": 0, + "content": "Besides, since our MIM pre-training has a similar target to subsequent low-level tasks, we do not need to change the decoder like MAE [15] does but just fine-tune it." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.531, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Restore Anything with Masks" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "image", + "bbox": [ + 0.286, + 0.146, + 0.722, + 0.414 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.431, + 0.788, + 0.502 + ], + "angle": 0, + "content": "Fig. 4: The effect of MIM reconstruction with different input integrity on kernel deblurring (orange border) and denoising (blue border). We also visualize the color distributions of reconstructions in various tasks above. It shows that the distribution of the reconstruction results obtained using the twin-masks method as input is closer to the real images (ground truth) compared to the results obtained using the whole input." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.531, + 0.785, + 0.562 + ], + "angle": 0, + "content": "Reconstruction target. Following the Bert [18] and MAE [15], we choose L1 loss to supervise the masked part. The training objective can be written as:" + }, + { + "type": "equation", + "bbox": [ + 0.376, + 0.573, + 0.786, + 0.598 + ], + "angle": 0, + "content": "\\[\n\\underset {\\theta} {\\arg \\min } \\mathbb {E} [ | | \\tilde {\\mathcal {M}} (I - f (\\mathcal {M} (I _ {d}), \\theta)) | | ], \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.609, + 0.788, + 0.657 + ], + "angle": 0, + "content": "where \\(\\{I, I_d\\}\\) represents a pair of clean image and degraded image, \\(f(\\cdot, \\theta)\\) denotes a network with parameters \\(\\theta\\), \\(\\mathcal{M}(\\cdot)\\) is a random binary masking operation and \\(\\tilde{\\mathcal{M}}(\\cdot) = 1 - \\mathcal{M}(\\cdot)\\)." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.678, + 0.719, + 0.695 + ], + "angle": 0, + "content": "3.3 Finetuning with Mask Attribute Conductance Analysis" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.705, + 0.789, + 0.842 + ], + "angle": 0, + "content": "Observation. During pre-training, the network learns rich content priors. However, the incompleteness of the masked input prevents the direct use of the pre-trained model for inference, as it would result in a distribution shift in the outputs. As shown in Fig. 4, We start by feeding the entire image into a pretrained model, leading to a color-distorted result. Next, we use a pair of complementary masks, referred to as twin-masks, to individually mask the image. Subsequently, we input both of these complementarily masked images into the network. By combining the pixel values predicted by each image, we generate a higher-quality image. This observation indicates that the hindrance to using" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.338, + 0.129 + ], + "angle": 0, + "content": "Qin et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.176 + ], + "angle": 0, + "content": "mask pre-trained model directly for inference lies in input incompleteness rather than the model's inability to learn the restoration function." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.178, + 0.785, + 0.282 + ], + "angle": 0, + "content": "Building upon this insight, we explore the possibility of minimizing the influence of disparities in data input formats via model fine-tuning. To maintain the learned priors, it is essential to retain pre-trained parameters as extensively as possible while employing the fewest but most effective layers for fine-tuning. To tackle this, we introduce the concept of mask attribution conductance, which quantifies the importance of each layer concerning the fine-tuning objective. We then identify the top-\\(k\\%\\) most critical layers for fine-tuning." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.283, + 0.797, + 0.373 + ], + "angle": 0, + "content": "Preliminary. Before giving the definition of Mask Attribute Conductance (MAC), we briefly recall the definition of integrate gradient [45] (IG) and neuron conductance [6] (Cond). Considering a linear path \\(\\gamma (\\alpha) = x^{\\prime} + \\alpha (x - x^{\\prime})\\) from base input \\(x^{\\prime}\\) to target input \\(x\\), we can attribute output change \\(F(x) - F(x^{\\prime})\\) to \\(i\\)-th dimension of input/feature \\(x_{i}\\) (e.g. a pixel) by calculating its integrate gradient, which formally as below:" + }, + { + "type": "equation", + "bbox": [ + 0.333, + 0.382, + 0.786, + 0.416 + ], + "angle": 0, + "content": "\\[\n\\operatorname {I G} _ {i} (x) := \\left(x _ {i} - x _ {i} ^ {\\prime}\\right) \\cdot \\int_ {0} ^ {1} \\frac {\\partial F \\left(x ^ {\\prime} + \\alpha \\left(x - x ^ {\\prime}\\right)\\right)}{\\partial x _ {i}} d \\alpha . \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.421, + 0.785, + 0.468 + ], + "angle": 0, + "content": "We can also attribute output change to a specific neuron \\( y \\) by improving IG, which involves calculating the conductance. The conductance [6] of the hidden neuron \\( y \\) along the \\( \\gamma(\\alpha) \\) is:" + }, + { + "type": "equation", + "bbox": [ + 0.293, + 0.476, + 0.786, + 0.554 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\operatorname {C o n d} ^ {y} (x) := \\sum_ {i} \\left(x _ {i} - x _ {i} ^ {\\prime}\\right) \\cdot \\int_ {0} ^ {1} \\frac {\\partial F \\left(x ^ {\\prime} + \\alpha \\left(x - x ^ {\\prime}\\right)\\right)}{\\partial y} \\cdot \\frac {\\partial y}{\\partial x _ {i}} d \\alpha \\tag {3} \\\\ = \\sum_ {i} \\int_ {0} ^ {1} \\frac {\\partial F (\\gamma (\\alpha))}{\\partial y} \\cdot \\frac {\\partial y}{\\partial \\alpha} d \\alpha , \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.563, + 0.784, + 0.598 + ], + "angle": 0, + "content": "Note that \\((x_{i} - x_{i}^{\\prime}) = \\frac{\\partial(x^{\\prime} + \\alpha(x_{i} - x_{i}^{\\prime}))}{\\partial\\alpha}\\). Certainly, we can broaden Eq. (3) to compute conductance when integrating along any given path \\(\\alpha :[s,t]\\to P\\):" + }, + { + "type": "equation", + "bbox": [ + 0.332, + 0.606, + 0.785, + 0.641 + ], + "angle": 0, + "content": "\\[\n\\operatorname {G e n e r a l C o n d} ^ {y} (x) := \\sum_ {i} \\int_ {P} \\frac {\\partial F (X _ {i} (\\alpha))}{\\partial y} \\cdot \\frac {\\partial y}{\\partial \\alpha} d \\alpha , \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.649, + 0.785, + 0.679 + ], + "angle": 0, + "content": "where \\( X: R \\to R^n \\) is the function of the path from \\( x' \\) to \\( x \\), which satisfies \\( X(s) = x' \\), \\( X(t) = x \\). [s,t] represent the domain of the path function \\( X \\).\"" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.68, + 0.785, + 0.74 + ], + "angle": 0, + "content": "Finetuning with MAC. To find effective layers to finetune, we propose Mask Attribute Conductance (MAC) to evaluate how effective each layer is in overcoming the gap of input integrity. Considering such a nonlinear path \\(\\alpha : [0,1] \\to P_m\\) from zero input \\(x'\\) to whole input \\(x\\), which path function \\(X^m\\) satisfies:" + }, + { + "type": "equation", + "bbox": [ + 0.402, + 0.749, + 0.785, + 0.788 + ], + "angle": 0, + "content": "\\[\nX _ {i} ^ {m} (\\alpha ; \\alpha_ {i}) = \\left\\{ \\begin{array}{l l} x _ {i} ^ {\\prime}, & \\alpha < \\alpha_ {i} \\\\ x _ {i}, & e l s e \\end{array} , \\right. \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.784, + 0.841 + ], + "angle": 0, + "content": "where \\( i \\) refers to the index of pixels, \\( \\alpha_{i} \\in (0,1] \\) is a set of parameters that indicate when each pixel gets masked. We define this path as a Mask Attribute Path (MAP). Apparently, \\( X^{m}(0) = x^{\\prime} \\) and \\( X^{m}(1) = x \\)." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.531, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Restore Anything with Masks" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "image", + "bbox": [ + 0.28, + 0.147, + 0.486, + 0.28 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.379, + 0.284, + 0.397, + 0.296 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.512, + 0.147, + 0.723, + 0.28 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.617, + 0.284, + 0.636, + 0.296 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image_caption", + "bbox": [ + 0.283, + 0.313, + 0.719, + 0.329 + ], + "angle": 0, + "content": "Fig. 5: Illumination of (a) \\( X_{i}^{m} \\) in Eq. (5) and (b) \\( \\tilde{X}_i^m \\) in Eq. (6)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.356, + 0.785, + 0.4 + ], + "angle": 0, + "content": "However, \\( X^{m} \\) is not differentiable, making it an invalid attribute path function. To solve this problem, we use a group of sigmoid-like functions \\( \\tilde{X}^{m} \\) to approximate \\( X^{m} \\):" + }, + { + "type": "equation", + "bbox": [ + 0.397, + 0.4, + 0.785, + 0.431 + ], + "angle": 0, + "content": "\\[\n\\tilde {X} _ {i} ^ {m} \\left(\\alpha ; \\alpha_ {i}\\right) = \\frac {\\left(x _ {i} ^ {\\prime} - x _ {i}\\right)}{1 + e ^ {- \\delta \\left(x _ {i} ^ {\\prime} - \\alpha_ {i}\\right)}}. \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.436, + 0.784, + 0.482 + ], + "angle": 0, + "content": "We can see that \\(\\tilde{X}^m\\) is very close to \\(X^m\\) when \\(\\delta\\) is sufficiently large (as depicted in Fig. 5). And for each \\(\\tilde{X}_i^m\\), it will change sharply from \\(x_i'\\) to \\(x_i\\) when \\(\\alpha\\) is in the neighborhood of \\(\\alpha_i\\)." + }, + { + "type": "text", + "bbox": [ + 0.24, + 0.483, + 0.597, + 0.497 + ], + "angle": 0, + "content": "Here, we can give a definition of MAC as below:" + }, + { + "type": "equation", + "bbox": [ + 0.34, + 0.505, + 0.785, + 0.582 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathrm {M A C} ^ {y} (x) := \\sum_ {i} \\int_ {P _ {m}} \\frac {\\partial F (X _ {i} (\\alpha))}{\\partial y} \\cdot \\frac {\\partial y}{\\partial \\alpha} d \\alpha \\\\ \\approx \\sum_ {i} \\int_ {0} ^ {1} \\frac {\\partial F \\left(\\tilde {X} _ {i} ^ {m} (\\alpha ; \\alpha_ {i})\\right)}{\\partial y} \\cdot \\frac {\\partial y}{\\partial \\alpha} d \\alpha . \\tag {7} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.59, + 0.784, + 0.621 + ], + "angle": 0, + "content": "In fact, a partial path is also available to attribute from a masked input \\( x_{m} \\) with any mask ratio \\( r \\) to whole input \\( x \\):" + }, + { + "type": "equation", + "bbox": [ + 0.334, + 0.628, + 0.785, + 0.666 + ], + "angle": 0, + "content": "\\[\n\\mathrm {M A C} _ {r} ^ {y} (x) \\approx \\sum_ {i} \\int_ {1 - r} ^ {1} \\frac {\\partial F \\left(\\tilde {X} _ {i} ^ {m} (\\alpha ; \\alpha_ {i})\\right)}{\\partial y} \\cdot \\frac {\\partial y}{\\partial \\alpha} d \\alpha . \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.784, + 0.706 + ], + "angle": 0, + "content": "In practice, we use N-steps discretization to approximate the integral form of Eq. (8), which follows [44]:" + }, + { + "type": "equation", + "bbox": [ + 0.328, + 0.713, + 0.785, + 0.789 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathrm {M A C} _ {r} ^ {y} (x) \\approx \\sum_ {i} \\sum_ {j = 1} ^ {N} \\frac {\\partial F \\left(\\tilde {X} _ {i} ^ {m} \\left(\\frac {j r}{N} ; \\alpha_ {i}\\right)\\right)}{\\partial y} \\tag {9} \\\\ \\cdot \\left(F _ {y} \\left(\\tilde {X} _ {i} ^ {m} \\left(\\frac {(j + 1) r}{N}\\right)\\right) - F _ {y} \\left(\\tilde {X} _ {i} ^ {m} \\left(\\frac {j r}{N}\\right)\\right)\\right). \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.784, + 0.84 + ], + "angle": 0, + "content": "We compute the MAC of each layer of pre-trained networks, rank them in descending order based on their MAC values, and pick top-\\(k\\%\\) layers for finetuning. The networks are initialized by pre-trained weight and only top-\\(k\\%\\) layers" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.338, + 0.129 + ], + "angle": 0, + "content": "Qin et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.145, + 0.788, + 0.201 + ], + "angle": 0, + "content": "Table 1: Quantitative comparison on seven challenging image restoration tasks, including dehazing, deraining, denoising, motion deblurring, low-light image enhancement (LLIE), kernel deblurring, and JPEG artifact removal. **boldface** and **underlined** indicate the best and second-best results, respectively." + }, + { + "type": "table", + "bbox": [ + 0.219, + 0.213, + 0.789, + 0.336 + ], + "angle": 0, + "content": "
MethodSOTS [20]Rain13k-Test [32]BSD68 [38]GoPro [40]LOL [5]LSDIR-Blur [26]LSDIR-Jpeg [26]Average
PSNR†/SSIM†PSNR†/SSIM†PSNR†/SSIM†PSNR†/SSIM†PSNR†/SSIM†PSNR†/SSIM†PSNR†/SSIM†PSNR†/SSIM†
Restomer [55]22.89/0.917227.05/0.846930.95/0.865727.46/0.849723.65/0.845819.60/0.365830.46/0.914126.01/0.8007
MPRNet [39]25.23/0.946325.36/0.806829.83/0.831725.90/0.794922.29/0.817025.68/0.828128.96/0.886526.18/0.8445
NAFNet [3]25.74/0.944524.65/0.787730.37/0.854025.53/0.790921.50/0.810429.08/0.913029.09/0.895526.57/0.8566
DL [8]21.16/0.904219.56/0.650816.15/0.586117.63/0.586219.26/0.777717.98/0.612119.55/0.696518.75/0.6877
TAPE [31]25.14/0.931923.66/0.781830.11/0.835425.97/0.796218.95/0.763224.26/0.765429.28/0.896525.34/0.8243
AirNet [21]21.66/0.836620.21/0.640227.99/0.725023.36/0.750316.65/0.670823.84/0.735824.36/0.802022.58/0.7372
SwinIR [28]27.29/0.962225.32/0.825830.65/0.854026.61/0.812518.66/0.804827.82/0.883930.13/0.907126.64/0.8643
RAM-SwinIR28.47/0.968926.31/0.848630.83/0.861126.89/0.820021.62/0.829126.66/0.851430.22/0.909627.28/0.8698
PromptIR [42]28.70/0.965927.46/0.858530.84/0.862527.71/0.856521.19/0.835631.01/0.938530.30/0.911728.17/0.8899
RAM-PromptIR29.64/0.969528.47/0.875130.86/0.862428.02/0.859224.46/0.858129.57/0.917930.33/0.911928.76/0.8935
" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.347, + 0.788, + 0.482 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.259, + 0.489, + 0.744, + 0.505 + ], + "angle": 0, + "content": "Fig. 6: Dehaze visual comparison on SOTS dataset. Zoom in for details." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.534, + 0.784, + 0.565 + ], + "angle": 0, + "content": "will be fine-tuned. More implementation details can be found in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.587, + 0.367, + 0.604 + ], + "angle": 0, + "content": "4 Experiment" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.619, + 0.44, + 0.634 + ], + "angle": 0, + "content": "4.1 Experiments Settings" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.644, + 0.788, + 0.81 + ], + "angle": 0, + "content": "Datasets and Metrics. We combine datasets from various restoration tasks to form the training set, following [59]. For high-cost tasks that degradations are difficult to synthesize, we leverage existing paired datasets, including RE-SIDE [20] for dehazing, Rain13k [10, 25, 27, 35, 53] for deraining, GoPro [40] for motion deblurring, and LOL-v2 [54] for low-light image enhancement (LLIE). For low-cost tasks that degradations are easy to synthesize (e.g. noise, kernel blur, and JPEG artifact), we generate corrupted images on the LSDIR dataset [26] during the training process, which involves generating Gaussian noise with random variation \\(\\sigma \\in (0, 50]\\), creating gaussian blurred images with a blur kernel of size \\(k = 15\\) and random \\(\\sigma \\in [0.1, 3.1]\\), and introducing JPEG artifacts with a random quality parameter \\(q \\in [20, 90]\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.842 + ], + "angle": 0, + "content": "For evaluation, we use SOTS-outdoor [20] for dehazing, Rain13k-Test (the combination of Rain100L [52], Rain100H [52], Test100 [58], Test1200 [57] and" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.531, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Restore Anything with Masks" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.785, + 0.173 + ], + "angle": 0, + "content": "Table 2: Quantitative Gaussian denoising results at different noise levels on BSD68 and Urban100 datasets in terms of PSNR." + }, + { + "type": "table", + "bbox": [ + 0.269, + 0.185, + 0.737, + 0.357 + ], + "angle": 0, + "content": "
MethodBSD68 [38]Urban100 [16]
σ = 15σ = 25σ = 50Averageσ = 15σ = 25σ = 50Average
NAFNet [3]33.2230.5927.3030.3732.6730.2126.9729.92
MPRNet [39]32.7330.1126.6529.8332.0629.4625.7729.10
Restormer [55]33.7931.1727.9030.9533.8331.4027.9931.07
DL [8]16.0416.2016.1916.1519.1719.1118.4718.92
TAPE [31]33.1030.3726.8630.1132.5929.9326.1929.57
AirNet [21]31.6328.8323.5227.9929.7926.9021.3526.01
SwinIR [28]33.5330.8927.5430.6533.5030.9927.3730.62
RAM-SwinIR33.6531.0627.7730.8233.8231.4327.9431.07
performance gains(↑0.12)(↑0.17)(↑0.23)(↑0.17)(↑0.32)(↑0.44)(↑0.57)(↑0.45)
PromptIR [42]33.6731.0627.8030.8433.5631.0827.6430.76
RAM-PromptIR33.7031.0827.7930.8633.7031.3027.9230.97
performance gains(↑0.03)(↑0.02)(↓0.01)(↑0.02)(↑0.14)(↑0.22)(↑0.28)(↑0.21)
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.384, + 0.785, + 0.459 + ], + "angle": 0, + "content": "Test2800 [11]) for deraining, GoPro for motion deblurring, LOL [5] for low-light enhancement, BSD68 [38] for denoising, LSDIR-val for kernel deblurring and JPEG artifact removal. Furthermore, We conducted evaluations including denoising tests with variances of 15, 25, and 50, deblurring tests at \\( k = 15 \\) and \\( \\sigma = 2.0 \\), and JPEG artifact removal tests at \\( q = 50 \\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.46, + 0.788, + 0.611 + ], + "angle": 0, + "content": "Implementation Details. We apply our proposed RAM to SwinIR [28] and PromptIR [42]. The input size for RAM-SwinIR is 64, while for RAM-PromptIR it is 128. During the pre-training phase, we use the Adam optimizer to train RAM-SwinIR and RAM-PromptIR for 300 epochs, with the learning rate decaying from 1e-4 to 6e-5 following a cosine schedule. In the fine-tuning phase, we use the Adam optimizer to fine-tune the network layers obtained from the MAC analysis of RAM-SwinIR and RAM-PromptIR for 40 epochs, with the learning rate decaying from 2e-4 to 1e-7 following a cosine schedule. The batch sizes for RAM-SwinIR and RAM-PromptIR during the pre-training and fine-tuning phases are (12,4) and (4,4), respectively." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.634, + 0.371, + 0.649 + ], + "angle": 0, + "content": "4.2 Comparisons" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.659, + 0.785, + 0.749 + ], + "angle": 0, + "content": "To validate the gain capability and effectiveness of our RAM, we apply the proposed RAM to SwinIR (a general image restoration method) and PromptIR (an all-in-one image restoration method). Four general architecture-based image restoration methods [3, 28, 39, 55] and four all-in-one methods [8, 21, 31, 42] are considered for comparison. We ensure that the number of supervised pixels employed by all other methods equals that used during the pre-training stage." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.787, + 0.84 + ], + "angle": 0, + "content": "As illustrated in Tab. 1, our approach achieves the best or comparable performance on each task. On the average score across seven different tasks, our method with PromptIR [42] achieves 0.59dB performance gains compared to the second-best algorithm. Besides, the SwinIR equipped with RAM also yields \\(2.40\\%\\) improvement on PSNR. Specifically, our RAM has significant benefits for dehazing and low-light enhancement. Tab. 2 shows the Quantitative denois-" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.338, + 0.129 + ], + "angle": 0, + "content": "Qin et al." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.144, + 0.313, + 0.193 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.24, + 0.195, + 0.297, + 0.204 + ], + "angle": 0, + "content": "Rainy Image" + }, + { + "type": "image", + "bbox": [ + 0.315, + 0.144, + 0.407, + 0.193 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.338, + 0.195, + 0.376, + 0.204 + ], + "angle": 0, + "content": "NAFNet" + }, + { + "type": "image", + "bbox": [ + 0.409, + 0.144, + 0.501, + 0.193 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.433, + 0.195, + 0.473, + 0.203 + ], + "angle": 0, + "content": "MPRNet" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.144, + 0.596, + 0.193 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.525, + 0.195, + 0.571, + 0.204 + ], + "angle": 0, + "content": "Restomer" + }, + { + "type": "image", + "bbox": [ + 0.598, + 0.144, + 0.691, + 0.193 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.624, + 0.195, + 0.657, + 0.203 + ], + "angle": 0, + "content": "SwinIR" + }, + { + "type": "image", + "bbox": [ + 0.693, + 0.144, + 0.785, + 0.193 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.692, + 0.195, + 0.782, + 0.204 + ], + "angle": 0, + "content": "RAM-SwinIR (Ours)" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.206, + 0.313, + 0.255 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.256, + 0.256, + 0.273, + 0.264 + ], + "angle": 0, + "content": "DL" + }, + { + "type": "image", + "bbox": [ + 0.315, + 0.207, + 0.408, + 0.256 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.345, + 0.256, + 0.371, + 0.264 + ], + "angle": 0, + "content": "TAPE" + }, + { + "type": "image", + "bbox": [ + 0.409, + 0.207, + 0.501, + 0.256 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.438, + 0.256, + 0.469, + 0.264 + ], + "angle": 0, + "content": "AirNet" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.207, + 0.596, + 0.256 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.525, + 0.256, + 0.567, + 0.265 + ], + "angle": 0, + "content": "PromptIR" + }, + { + "type": "image", + "bbox": [ + 0.598, + 0.207, + 0.691, + 0.256 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.595, + 0.256, + 0.692, + 0.265 + ], + "angle": 0, + "content": "RAM-PromptIR(Ours)" + }, + { + "type": "image", + "bbox": [ + 0.693, + 0.207, + 0.785, + 0.256 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.729, + 0.256, + 0.745, + 0.264 + ], + "angle": 0, + "content": "GT" + }, + { + "type": "image_caption", + "bbox": [ + 0.236, + 0.276, + 0.764, + 0.29 + ], + "angle": 0, + "content": "Fig. 7: Derain visual comparison on Rain13k-Test dataset. Zoom in for details." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.304, + 0.313, + 0.347 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.239, + 0.348, + 0.298, + 0.357 + ], + "angle": 0, + "content": "Blurred Image" + }, + { + "type": "image", + "bbox": [ + 0.315, + 0.304, + 0.408, + 0.347 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.344, + 0.348, + 0.379, + 0.356 + ], + "angle": 0, + "content": "NAFNet" + }, + { + "type": "image", + "bbox": [ + 0.409, + 0.304, + 0.501, + 0.347 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.438, + 0.348, + 0.474, + 0.356 + ], + "angle": 0, + "content": "MPRNet" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.304, + 0.695, + 0.347 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.526, + 0.348, + 0.57, + 0.356 + ], + "angle": 0, + "content": "Restormer" + }, + { + "type": "image", + "bbox": [ + 0.696, + 0.304, + 0.786, + 0.347 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.697, + 0.348, + 0.783, + 0.356 + ], + "angle": 0, + "content": "RAM-SwinIR (Ours)" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.358, + 0.313, + 0.401 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.254, + 0.402, + 0.268, + 0.41 + ], + "angle": 0, + "content": "DL" + }, + { + "type": "image", + "bbox": [ + 0.315, + 0.358, + 0.408, + 0.401 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.347, + 0.402, + 0.373, + 0.41 + ], + "angle": 0, + "content": "TAPE" + }, + { + "type": "image", + "bbox": [ + 0.409, + 0.358, + 0.501, + 0.401 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.438, + 0.402, + 0.467, + 0.41 + ], + "angle": 0, + "content": "AirNet" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.358, + 0.695, + 0.401 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.524, + 0.402, + 0.566, + 0.41 + ], + "angle": 0, + "content": "PromptIR" + }, + { + "type": "image", + "bbox": [ + 0.597, + 0.358, + 0.691, + 0.401 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.596, + 0.402, + 0.689, + 0.411 + ], + "angle": 0, + "content": "RAM-PromptIR(Ours)" + }, + { + "type": "image", + "bbox": [ + 0.696, + 0.358, + 0.786, + 0.401 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.726, + 0.402, + 0.741, + 0.41 + ], + "angle": 0, + "content": "GT" + }, + { + "type": "image_caption", + "bbox": [ + 0.234, + 0.427, + 0.767, + 0.441 + ], + "angle": 0, + "content": "Fig. 8: Motion deblur visual comparison on GoPro dataset. Zoom in for details." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.471, + 0.785, + 0.502 + ], + "angle": 0, + "content": "ing result at different noise levels. Both RAM-SwinIR and RAM-PromptIR get higher performance than the origin versions." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.502, + 0.787, + 0.639 + ], + "angle": 0, + "content": "Fig. 6-Fig. 10 show the qualitative results of various methods on different datasets. In Fig. 6, our method achieves better dehazing effects (right region) and exposure correction (sky). In the deraining task (Fig. 7, our method better removes rain streaks and restores textures in the occluded regions. In terms of denoising (Fig. 9) and deblurring (Fig. 8), we achieve clearer results with fewer artifacts. We also demonstrate better color correction (the purple blanket on the left) and exposure correction in low-light image enhancement tasks (Fig. 10). For simplicity, the qualitative effects of kernel deblurring and JPEG artifact removal will be presented in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.662, + 0.389, + 0.677 + ], + "angle": 0, + "content": "4.3 Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.688, + 0.785, + 0.735 + ], + "angle": 0, + "content": "In this section, we conduct an ablative study on the masking ratio, mask patch size, pre-training strategy, fine-tuning strategy, and fine-tuning ratio to demonstrate the effectiveness of our MIM pre-training and fine-tuning strategy." + }, + { + "type": "table_caption", + "bbox": [ + 0.333, + 0.762, + 0.669, + 0.776 + ], + "angle": 0, + "content": "Table 3: Ablative results on masking ratios." + }, + { + "type": "table", + "bbox": [ + 0.303, + 0.788, + 0.702, + 0.838 + ], + "angle": 0, + "content": "
Masking ratio20%40%50%60%80%
PSNR↑27.2827.2127.2827.2627.08
SSIM↑0.86630.86830.86980.86940.8642
" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.531, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Restore Anything with Masks" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.151, + 0.492, + 0.18 + ], + "angle": 0, + "content": "Table 4: Ablative results of different pre-training strategies." + }, + { + "type": "table", + "bbox": [ + 0.217, + 0.192, + 0.492, + 0.252 + ], + "angle": 0, + "content": "
RAM-SwinIRPSNR↑SSIM↑
pre-trained w/ gt26.620.8580
pre-trained w/ paired data27.280.8698
" + }, + { + "type": "table_caption", + "bbox": [ + 0.51, + 0.145, + 0.785, + 0.174 + ], + "angle": 0, + "content": "Table 5: Ablative results of different fine-tuning strategies." + }, + { + "type": "table", + "bbox": [ + 0.511, + 0.185, + 0.768, + 0.259 + ], + "angle": 0, + "content": "
RAM-SwinIRPSNR↑SSIM↑
random26.860.8535
IG [45]26.920.8554
MAC (Ours)27.280.8698
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.29, + 0.784, + 0.351 + ], + "angle": 0, + "content": "- Patch size & masking ratio are two essential hyper-parameters that determine the continuity and area of the masking of an image. In high-level tasks, MAE [15] masks \\(75\\%\\) of an image with \\(16 \\times 16\\) patch size. However, it can corrupt the local details of images, which is not suitable for image restoration." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.397, + 0.557, + 0.48 + ], + "angle": 0, + "content": "Table 6: Ablative results in terms of the PSNR on fine-tuning ratios. We compared the performance in restoring images with unseen noises (Out-of-Distribution Denoising) and known degraded images (In-Distribution). In this case, the settings of In-Distribution are the same as Tab. 1." + }, + { + "type": "table", + "bbox": [ + 0.22, + 0.481, + 0.556, + 0.585 + ], + "angle": 0, + "content": "
MethodOut-of-DistributionIn-Distribution
PossionPepperSpeckleAverageAverage
SwinIR [28]12.8310.0020.8614.5626.64
RAM-SwinIR10%13.6719.2321.0717.9927.28
RAM-SwinIR20%13.2719.0920.6817.6827.35
RAM-SwinIR50%12.7516.5120.3616.5427.38
RAM-SwinIR100%12.4715.3120.0115.9327.54
" + }, + { + "type": "text", + "bbox": [ + 0.565, + 0.353, + 0.788, + 0.623 + ], + "angle": 0, + "content": "We first find the best choice of patch size by pretraining SwinIR [28] on \\(1 \\times 1\\), \\(4 \\times 4\\), and \\(8 \\times 8\\), as shown in Fig. 3. Since the attention layers of SwinIR treat an \\(8 \\times 8\\) patch as a token, the \\(4 \\times 4\\) pre-training produces heavy artifacts. Besides, the results generated by \\(8 \\times 8\\) pretraining are highly missing details, e.g. the texture of the polar bear's paws. In contrast, the model pre-trained with \\(1 \\times 1\\) patch size, which is also our final choice, achieves a satisfactory reconstruction and removes most of the rain streaks." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.626, + 0.785, + 0.701 + ], + "angle": 0, + "content": "Then, we adjust the masking ratio from \\(20\\%\\) to \\(80\\%\\). As we can see in Tab. 3, the model pre-trained with \\(50\\%\\) achieves the highest performance. Moreover, the performance is significantly dropped from 27.28dB to 27.08dB in terms of PSNR when we continue to increase the masking ratio, which also demonstrates our opinion that a high masking ratio is harmful to image restoration." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.703, + 0.785, + 0.794 + ], + "angle": 0, + "content": "Pre-trained with paired data. Tab. 4 compares the results of using paired data for mask image pretraining (our pretraining strategy) with those using only ground truth for mask image pretraining. It shows that pre-trained with paired data is necessary for our RAM. Pretraining the model on high-quality images does not effectively enable learning for image restoration tasks. It still requires paired data to guide the model in the learning process." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.787, + 0.842 + ], + "angle": 0, + "content": "Fine-tuning strategy. To verify the effectiveness of our fine-tuning strategy, we fine-tune \\(10\\%\\) of the network layers selected through MAC analysis, IG [45], and uniform sampling, respectively, and the results are shown in Tab. 5. Compared" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.338, + 0.129 + ], + "angle": 0, + "content": "Qin et al." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.144, + 0.314, + 0.193 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.238, + 0.195, + 0.296, + 0.205 + ], + "angle": 0, + "content": "Noisy Image" + }, + { + "type": "image", + "bbox": [ + 0.315, + 0.144, + 0.408, + 0.193 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.338, + 0.195, + 0.376, + 0.204 + ], + "angle": 0, + "content": "NAFNet" + }, + { + "type": "image", + "bbox": [ + 0.41, + 0.144, + 0.502, + 0.193 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.434, + 0.195, + 0.474, + 0.204 + ], + "angle": 0, + "content": "MPRNet" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.144, + 0.597, + 0.193 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.525, + 0.195, + 0.571, + 0.204 + ], + "angle": 0, + "content": "Restormer" + }, + { + "type": "image", + "bbox": [ + 0.598, + 0.144, + 0.692, + 0.193 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.626, + 0.195, + 0.659, + 0.204 + ], + "angle": 0, + "content": "SwinIR" + }, + { + "type": "image", + "bbox": [ + 0.694, + 0.144, + 0.787, + 0.193 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.694, + 0.195, + 0.785, + 0.204 + ], + "angle": 0, + "content": "RAM-SwinIR (Ours)" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.205, + 0.313, + 0.255 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.255, + 0.256, + 0.272, + 0.264 + ], + "angle": 0, + "content": "DL" + }, + { + "type": "image", + "bbox": [ + 0.315, + 0.205, + 0.408, + 0.255 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.345, + 0.256, + 0.371, + 0.264 + ], + "angle": 0, + "content": "TAPE" + }, + { + "type": "image", + "bbox": [ + 0.409, + 0.205, + 0.502, + 0.255 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.438, + 0.256, + 0.469, + 0.265 + ], + "angle": 0, + "content": "AirNet" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.205, + 0.597, + 0.255 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.525, + 0.256, + 0.569, + 0.265 + ], + "angle": 0, + "content": "PromptIR" + }, + { + "type": "image", + "bbox": [ + 0.598, + 0.205, + 0.691, + 0.255 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.596, + 0.256, + 0.694, + 0.265 + ], + "angle": 0, + "content": "RAM-PromptIR(Ours)" + }, + { + "type": "image", + "bbox": [ + 0.694, + 0.205, + 0.787, + 0.255 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.732, + 0.256, + 0.748, + 0.264 + ], + "angle": 0, + "content": "GT" + }, + { + "type": "image_caption", + "bbox": [ + 0.242, + 0.281, + 0.759, + 0.296 + ], + "angle": 0, + "content": "Fig. 9: Denoising visual comparison on CBSD68 dataset. Zoom in for details." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.308, + 0.313, + 0.358 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.228, + 0.36, + 0.304, + 0.37 + ], + "angle": 0, + "content": "Low Light Image" + }, + { + "type": "image", + "bbox": [ + 0.315, + 0.308, + 0.408, + 0.358 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.339, + 0.36, + 0.376, + 0.368 + ], + "angle": 0, + "content": "NAFNet" + }, + { + "type": "image", + "bbox": [ + 0.41, + 0.308, + 0.502, + 0.358 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.435, + 0.36, + 0.474, + 0.368 + ], + "angle": 0, + "content": "MPRNet" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.308, + 0.597, + 0.358 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.525, + 0.36, + 0.571, + 0.368 + ], + "angle": 0, + "content": "Restormer" + }, + { + "type": "image", + "bbox": [ + 0.598, + 0.308, + 0.691, + 0.358 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.627, + 0.36, + 0.659, + 0.368 + ], + "angle": 0, + "content": "SwinIR" + }, + { + "type": "image", + "bbox": [ + 0.693, + 0.308, + 0.786, + 0.358 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.693, + 0.36, + 0.785, + 0.369 + ], + "angle": 0, + "content": "RAM-SwinIR (Ours)" + }, + { + "type": "image", + "bbox": [ + 0.217, + 0.37, + 0.313, + 0.419 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.256, + 0.42, + 0.272, + 0.429 + ], + "angle": 0, + "content": "DL" + }, + { + "type": "image", + "bbox": [ + 0.315, + 0.37, + 0.408, + 0.419 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.345, + 0.42, + 0.371, + 0.429 + ], + "angle": 0, + "content": "TAPE" + }, + { + "type": "image", + "bbox": [ + 0.41, + 0.37, + 0.502, + 0.419 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.439, + 0.42, + 0.469, + 0.429 + ], + "angle": 0, + "content": "AirNet" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.37, + 0.597, + 0.419 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.525, + 0.42, + 0.569, + 0.43 + ], + "angle": 0, + "content": "PromptIR" + }, + { + "type": "image", + "bbox": [ + 0.598, + 0.37, + 0.691, + 0.419 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.596, + 0.42, + 0.694, + 0.43 + ], + "angle": 0, + "content": "RAM-PromptIR(Ours)" + }, + { + "type": "image", + "bbox": [ + 0.693, + 0.37, + 0.786, + 0.419 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.732, + 0.42, + 0.748, + 0.429 + ], + "angle": 0, + "content": "GT" + }, + { + "type": "image_caption", + "bbox": [ + 0.266, + 0.442, + 0.735, + 0.457 + ], + "angle": 0, + "content": "Fig. 10: LLIE visual comparison on LOL dataset. Zoom in for details." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.485, + 0.785, + 0.514 + ], + "angle": 0, + "content": "to IG, we have improved by 0.36 dB in PSNR and \\(1.6\\%\\) in SSIM, which indicates that our selection strategy is superior to IG." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.516, + 0.785, + 0.59 + ], + "angle": 0, + "content": "Fine-tune ratio. We conduct the ablation experiment to compare the network's performances with different fine-tune ratios in Tab. 6. We found that using our finetune strategy, a pre-trained network could achieve comparable performance by fine-tuning only a few layers (e.g. \\(10\\%\\)). At the same time, we need to finetune almost all network parameters to get the best performance on given tasks." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.592, + 0.785, + 0.666 + ], + "angle": 0, + "content": "Performance vs Generalization capability. We found a trade-off between in-distribution performance and out-of-distribution generalization in Tab. 6. We found that the more layers fine-tuned, the less generalization capability to tackle the out-of-distribution tasks. With our fine-tuning method, the model can have stronger generalization while maintaining comparable performance." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.689, + 0.359, + 0.705 + ], + "angle": 0, + "content": "5 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.788, + 0.84 + ], + "angle": 0, + "content": "This paper presents RAM, a pipeline for extracting intrinsic image information from corrupted images using Mask Image Modeling (MIM) pre-training. We design a MIM pre-training strategy tailored for image restoration and a fine-tuning algorithm to handle the transition from masked to complete images. By analyzing layer importance with MAC, we achieve high performance with minimal parameter tuning. Extensive experiments demonstrate that our RAM can bring boosts to various architectures and achieve state-of-the-art performance, moving towards a unified solution for all-in-one image restoration." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.531, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Restore Anything with Masks" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.145, + 0.323, + 0.16 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.177, + 0.785, + 0.218 + ], + "angle": 0, + "content": "1. Chen, H., Wang, Y., Guo, T., Xu, C., Deng, Y., Liu, Z., Ma, S., Xu, C., Xu, C., Gao, W.: Pre-trained image processing transformer. In: CVPR. pp. 12299-12310 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.219, + 0.785, + 0.259 + ], + "angle": 0, + "content": "2. Chen, H., Gu, J., Liu, Y., Magid, S.A., Dong, C., Wang, Q., Pfister, H., Zhu, L.: Masked image training for generalizable deep image denoising. In: CVPR. pp. 1692-1703 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.26, + 0.785, + 0.287 + ], + "angle": 0, + "content": "3. Chen, L., Chu, X., Zhang, X., Sun, J.: Simple baselines for image restoration. In: ECCV. pp. 17-33. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.288, + 0.785, + 0.342 + ], + "angle": 0, + "content": "4. Chen, W.T., Huang, Z.K., Tsai, C.C., Yang, H.H., Ding, J.J., Kuo, S.Y.: Learning multiple adverse weather removal via two-stage knowledge learning and multi-contrastive regularization: Toward a unified model. In: CVPR. pp. 17653-17662 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.343, + 0.785, + 0.37 + ], + "angle": 0, + "content": "5. Chen Wei, Wenjing Wang, W.Y.J.L.: Deep retina decomposition for low-light enhancement. In: BMVC. British Machine VLOLision Association (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.371, + 0.785, + 0.397 + ], + "angle": 0, + "content": "6. Dhamdhere, K., Sundararajan, M., Yan, Q.: How important is a neuron. In: ICLR (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.398, + 0.785, + 0.426 + ], + "angle": 0, + "content": "7. Duan, H., Shen, W., Min, X., Tu, D., Teng, L., Wang, J., Zhai, G.: Masked autoencoders as image processors. arXiv preprint arXiv:2303.17316 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.426, + 0.785, + 0.453 + ], + "angle": 0, + "content": "8. Fan, Q., Chen, D., Yuan, L., Hua, G., Yu, N., Chen, B.: A general decoupled learning framework for parameterized image operators. PAMI 43(1), 33-47 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.454, + 0.785, + 0.481 + ], + "angle": 0, + "content": "9. Fang, Y., Zhang, H., Wong, H.S., Zeng, T.: A robust non-blind deblurring method using deep denoiser prior. In: CVPRW. pp. 735-744 (June 2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.482, + 0.785, + 0.509 + ], + "angle": 0, + "content": "0. Fu, X., Huang, J., Ding, X., Liao, Y., Paisley, J.: Clearing the skies: A deep network architecture for single-image rain removal. TIP 26(6), 2944-2956 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.509, + 0.785, + 0.536 + ], + "angle": 0, + "content": "1. Fu, X., Huang, J., Zeng, D., Huang, Y., Ding, X., Paisley, J.: Removing rain from single images via a deep detail network. In: CVPR. pp. 3855-3863 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.537, + 0.785, + 0.563 + ], + "angle": 0, + "content": "2. Gu, J., Dong, C.: Interpreting super-resolution networks with local attribution maps. In: CVPR. pp. 9199-9208 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.565, + 0.785, + 0.605 + ], + "angle": 0, + "content": "3. Guo, C.L., Yan, Q., Anwar, S., Cong, R., Ren, W., Li, C.: Image dehazing transformer with transmission-aware 3d position embedding. In: CVPR. pp. 5812-5820 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.606, + 0.785, + 0.646 + ], + "angle": 0, + "content": "4. Guo, C., Li, C., Guo, J., Loy, C.C., Hou, J., Kwong, S., Cong, R.: Zero-reference deep curve estimation for low-light image enhancement. In: CVPR. pp. 1780-1789 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.647, + 0.785, + 0.674 + ], + "angle": 0, + "content": "5. He, K., Chen, X., Xie, S., Li, Y., Dollár, P., Girshick., R.: Masked autoencoders are scalable vision learners. In: CVPR. pp. 16000-16009 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.675, + 0.785, + 0.702 + ], + "angle": 0, + "content": "6. Huang, J.B., Singh, A., Ahuja, N.: Single image super-resolution from transformed self-exemplars. In: CVPR. pp. 5197-5206 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.703, + 0.785, + 0.73 + ], + "angle": 0, + "content": "7. Jin, X., Han, L.H., Li, Z., Guo, C.L., Chai, Z., Li, C.: Dnf: Decouple and feedback network for seeing in the dark. In: CVPR. pp. 18135-18144 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.731, + 0.785, + 0.757 + ], + "angle": 0, + "content": "8. Kenton, J.D.M.W.C., Toutanova, L.K.: Bert: Pre-training of deep bidirectional transformers for language understanding. In: NAACL. pp. 4171-4186 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.758, + 0.785, + 0.785 + ], + "angle": 0, + "content": "9. Leino, K., Sen, S., Datta, A., Fredrikson, M., Li, L.: Influence-directed explanations for deep convolutional networks. In: ITC. pp. 1-8. IEEE (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.786, + 0.785, + 0.813 + ], + "angle": 0, + "content": "20. Li, B., Ren, W., Fu, D., Tao, D., Feng, D., Zeng, W., Wang, Z.: Benchmarking single-image dehazing and beyond. TIP 28(1), 492-505 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.814, + 0.785, + 0.84 + ], + "angle": 0, + "content": "21. Li, B., Liu, X., Hu, P., Wu, Z., Lv, J., Peng, X.: All-in-one image restoration for unknown corruption. In: CVPR. pp. 17452-17462 (2022)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.177, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.338, + 0.129 + ], + "angle": 0, + "content": "Qin et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.175 + ], + "angle": 0, + "content": "22. Li, C., Guo, C.L., Liang, Z., Zhou, S., Feng, R., Loy, C.C., et al.: Embedding fourier for ultra-high-definition low-light image enhancement. In: ICLR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.175, + 0.785, + 0.203 + ], + "angle": 0, + "content": "23. Li, D., Zhang, Y., Cheung, K.C., Wang, X., Qin, H., Li, H.: Learning degradation representations for image deblurring. In: ECCV. pp. 736-753. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.203, + 0.785, + 0.229 + ], + "angle": 0, + "content": "24. Li, R., Tan, R.T., Cheong, L.F.: All in one bad weather removal using architectural search. In: CVPR. pp. 3175-3185 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.229, + 0.785, + 0.257 + ], + "angle": 0, + "content": "25. Li, X., Wu, J., Lin, Z., Liu, H., Zha, H.: Recurrent squeeze-and-excitation context aggregation net for single image deraining. In: ECCV. pp. 254-269 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.257, + 0.785, + 0.298 + ], + "angle": 0, + "content": "26. Li, Y., Zhang, K., Liang, J., Cao, J., Liu, C., Gong, R., Zhang, Y., Tang, H., Liu, Y., Demandolx, D., et al.: Lsdir: A large scale dataset for image restoration. In: CVPR. pp. 1775-1787 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.298, + 0.785, + 0.324 + ], + "angle": 0, + "content": "27. Li, Y., Tan, R.T., Guo, X., Lu, J., Brown, M.S.: Rain streak removal using layer priors. In: CVPR" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.324, + 0.785, + 0.352 + ], + "angle": 0, + "content": "28. Liang, J., Cao, J., Sun, G., Zhang, K., Van Gool, L., Timofte, R.: Swinir: Image restoration using swin transformer. In: CVPR. pp. 1833-1844 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.352, + 0.785, + 0.392 + ], + "angle": 0, + "content": "29. Lin, X., Ren, C., Liu, X., Huang, J., Lei, Y.: Unsupervised image denoising in real-world scenarios via self-collaboration parallel generative adversarial branches. In: ICCV. pp. 12642-12652 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.392, + 0.785, + 0.433 + ], + "angle": 0, + "content": "30. Lin, X., Yue, J., Ren, C., Guo, C.L., Li, C.: Unlocking low-light-rainy image restoration by pairwise degradation feature vector guidance. arXiv preprint arXiv:2305.03997 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.433, + 0.785, + 0.473 + ], + "angle": 0, + "content": "31. Liu, L., Xie, L., Zhang, X., Yuan, S., Chen, X., Zhou, W., Li, H., Tian, Q.: Tape: Task-agnostic prior embedding for image restoration. In: ECCV. pp. 447-464. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.473, + 0.785, + 0.502 + ], + "angle": 0, + "content": "32. Liu, Y., He, J., Gu, J., Kong, X., Qiao, Y., Dong, C.: Degae: A new pretraining paradigm for low-level vision. In: CVPR. pp. 23292-23303 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.502, + 0.785, + 0.542 + ], + "angle": 0, + "content": "33. Liu, Y., Liu, A., Gu, J., Zhang, Z., Wu, W., Qiao, Y., Dong, C.: Discovering distinctive \"semantics\" in super-resolution networks. arXiv preprint arXiv:2108.00406 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.542, + 0.785, + 0.582 + ], + "angle": 0, + "content": "34. Liu, Z., Lin, Y., Cao, Y., Hu, H., Wei, Y., Zhang, Z., Lin, S., Guo, B.: Swin transformer: Hierarchical vision transformer using shifted windows. In: CVPR. pp. 10012-10022 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.582, + 0.785, + 0.609 + ], + "angle": 0, + "content": "35. Luo, Y., Xu, Y., Ji, H.: Removing rain from a single image via discriminative sparse coding. In: ICCV. pp. 3397-3405 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.609, + 0.785, + 0.651 + ], + "angle": 0, + "content": "36. Luo, Y., Zhao, R., Wei, X., Chen, J., Lu, Y., Xie, S., Wang, T., Xiong, R., Lu, M., Zhang, S.: Mowe: mixture of weather experts for multiple adverse weather removal. arXiv preprint arXiv:2303.13739 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.651, + 0.785, + 0.678 + ], + "angle": 0, + "content": "37. Magid, S.A., Lin, Z., Wei, D., Zhang, Y., Gu, J., Pfister, H.: Texture-based error analysis for image super-resolution. In: CVPR. pp. 2118-2127 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.678, + 0.785, + 0.718 + ], + "angle": 0, + "content": "38. Martin, D., Fowlkes, C., Tal, D., Malik, J.: A database of human segmented natural images and its application to evaluating segmentation algorithms and measuring ecological statistics. In: ICCV. vol. 2, pp. 416-423. IEEE (2001)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.718, + 0.785, + 0.746 + ], + "angle": 0, + "content": "39. Mehri, A., Ardakani, P.B., Sappa, A.D.: Mprnet: Multi-path residual network for lightweight image super resolution. In: CVPR. pp. 2704-2713 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.746, + 0.785, + 0.772 + ], + "angle": 0, + "content": "40. Nah, S., Hyun Kim, T., Mu Lee, K.: Deep multi-scale convolutional neural network for dynamic scene deblurring. In: CVPR" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.772, + 0.785, + 0.813 + ], + "angle": 0, + "content": "41. Park, D., Lee, B.H., Chun, S.Y.: All-in-one image restoration for unknown degradations using adaptive discriminative filters for specific degradations. In: CVPR. pp. 5815-5824 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.813, + 0.785, + 0.841 + ], + "angle": 0, + "content": "42. Potlapalli, V., Zamir, S.W., Khan, S.H., Shahbaz Khan, F.: Prompt for all-in-one image restoration. NeurIPS 36 (2024)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.532, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Restore Anything with Masks" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.175 + ], + "angle": 0, + "content": "43. Radford, A., Narasimhan, K., Salimans, T., Sutskever, I., et al.: Improving language understanding by generative pre-training (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.175, + 0.787, + 0.203 + ], + "angle": 0, + "content": "44. Shrikumar, A., Su, J., Kundaje, A.: Computationally efficient measures of internal neuron importance. arXiv preprint arXiv:1807.09946 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.203, + 0.787, + 0.231 + ], + "angle": 0, + "content": "45. Sundararajan, M., Taly, A., Yan, Q.: Axiomatic attribution for deep networks. In: ICML. pp. 3319-3328. PMLR (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.231, + 0.787, + 0.245 + ], + "angle": 0, + "content": "46. Sundararajan, M., Taly, A., Yan, Q.: Gradients of counterfactuals. ICLR (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.245, + 0.787, + 0.272 + ], + "angle": 0, + "content": "47. Wang, X., Wang, W., Cao, Y., Shen, C., Huang, T.: Images speak in images: A generalist painter for in-context visual learning. In: CVPR. pp. 6830-6839 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.272, + 0.787, + 0.3 + ], + "angle": 0, + "content": "48. Wang, Z., Cun, X., Bao, J., Zhou, W., Liu, J., Li, H.: Uformer: A general u-shaped transformer for image restoration. In: CVPR. pp. 17683-17693 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.3, + 0.787, + 0.328 + ], + "angle": 0, + "content": "49. Wu, R.Q., Duan, Z.P., Guo, C.L., Chai, Z., Li, C.: Ridcp: Revitalizing real image dehazing via high-quality codebook priors. In: CVPR. pp. 22282-22291 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.328, + 0.787, + 0.356 + ], + "angle": 0, + "content": "50. Xie, L., Wang, X., Dong, C., Qi, Z., Shan, Y.: Finding discriminative filters for specific degradations in blind super-resolution. NeurIPS 34, 51-61 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.356, + 0.787, + 0.384 + ], + "angle": 0, + "content": "51. Xie, Z., Zhang, Z., Cao, Y., Lin, Y., Bao, J., Yao, Z., Dai, Q., Hu, H.: Simmim: a simple framework for masked image modeling. In: CVPR. pp. 9653-9663 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.384, + 0.787, + 0.411 + ], + "angle": 0, + "content": "52. Yang, W., Tan, R.T., Feng, J., Liu, J., Guo, Z., Yan, S.: Deep joint rain detection and removal from a single image. In: CVPR. pp. 1357-1366 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.411, + 0.787, + 0.438 + ], + "angle": 0, + "content": "53. Yang, W., Tan, R.T., Wang, S., Fang, Y., Liu, J.: Single image deraining: From model-based to data-driven and beyond. PAMI 43(11), 4059-4077 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.439, + 0.787, + 0.48 + ], + "angle": 0, + "content": "54. Yang, W., Wang, W., Huang, H., Wang, S., Liu, J.: Sparse gradient regularized deep retina network for robust low-light image enhancement. TIP 30, 2072-2086 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.48, + 0.787, + 0.522 + ], + "angle": 0, + "content": "55. Zamir, S.W., Arora, A., Khan, S., Hayat, M., Khan, F.S., Yang, M.H.: Restormer: Efficient transformer for high-resolution image restoration. In: CVPR. pp. 5728-5739 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.522, + 0.787, + 0.563 + ], + "angle": 0, + "content": "56. Zhang, C., Zhu, Y., Yan, Q., Sun, J., Zhang, Y.: All-in-one multi-degradation image restoration network via hierarchical degradation representation. In: ACMMM. pp. 2285-2293 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.563, + 0.787, + 0.591 + ], + "angle": 0, + "content": "57. Zhang, H., Patel, V.M.: Density-aware single image de-raining using a multi-stream dense network. In: CVPR. pp. 695-704 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.591, + 0.787, + 0.619 + ], + "angle": 0, + "content": "58. Zhang, H., Sindagi, V., Patel, V.M.: Image de-raining using a conditional generative adversarial network. TCSVT 30(11), 3943-3956 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.619, + 0.787, + 0.66 + ], + "angle": 0, + "content": "59. Zhang, J., Huang, J., Yao, M., Yang, Z., Yu, H., Zhou, M., Zhao, F.: Ingredient-oriented multi-degradation learning for image restoration. In: CVPR. pp. 5825-5835 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.66, + 0.787, + 0.702 + ], + "angle": 0, + "content": "60. Zheng, N., Zhou, M., Dong, Y., Rui, X., Huang, J., Li, C., Zhao, F.: Empowering low-light image enhancer through customized learnable priors. In: ICCV. pp. 12559-12569 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.702, + 0.787, + 0.743 + ], + "angle": 0, + "content": "61. Zhu, Y., Wang, T., Fu, X., Yang, X., Guo, X., Dai, J., Qiao, Y., Hu, X.: Learning weather-general and weather-specific features for image restoration under multiple adverse weather conditions. In: CVPR. pp. 21747-21758 (2023)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.743 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/Restore Anything with Masks_ Leveraging Mask Image Modeling for Blind All-in-One Image Restoration/0de4c62c-e2d6-4f18-8a4e-c05352d6d956_origin.pdf b/2024/Restore Anything with Masks_ Leveraging Mask Image Modeling for Blind All-in-One Image Restoration/0de4c62c-e2d6-4f18-8a4e-c05352d6d956_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..e95d099e04cb6ac176325980460ddc7c5f47e241 --- /dev/null +++ b/2024/Restore Anything with Masks_ Leveraging Mask Image Modeling for Blind All-in-One Image Restoration/0de4c62c-e2d6-4f18-8a4e-c05352d6d956_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:41d1b280816456e535acdf91abc12fd902ec12f769a8a3dc13f0cad5773cec3b +size 10330715 diff --git a/2024/Restore Anything with Masks_ Leveraging Mask Image Modeling for Blind All-in-One Image Restoration/full.md b/2024/Restore Anything with Masks_ Leveraging Mask Image Modeling for Blind All-in-One Image Restoration/full.md new file mode 100644 index 0000000000000000000000000000000000000000..435dad60971e065e9f716143842e9f7254f20a91 --- /dev/null +++ b/2024/Restore Anything with Masks_ Leveraging Mask Image Modeling for Blind All-in-One Image Restoration/full.md @@ -0,0 +1,469 @@ +# Restore Anything with Masks: Leveraging Mask Image Modeling for Blind All-in-One Image Restoration + +Chu-Jie Qin $^{1,2*}$ , Rui-Qi Wu $^{1,2}$ , Zukun Liu $^{3}$ , Xin Lin $^{5}$ , Chun-Le Guo $^{1,2}$ , Hyun Hee Park $^{4}$ , and Chongyi Li $^{1,2\dagger}$ + +$^{1}$ VCIP, CS, Nankai University + +2 NKIARI, Shenzhen Futian + +{chujie.qin,wuruiqi}@mail.nankai.edu.cn + +{guochunle, lichongyi}@nankai.edu.cn + +3 Samsung Research, China, Beijing (SRC-B) + +4 The Department of Camera Innovation Group, Samsung Electronics {zikun.liu, inextg.park}@samsung.com + +Sichuan University + +linxin@stu.scu.edu.cn + +Abstract. All-in-one image restoration aims to handle multiple degradation types using one model. This paper proposes a simple pipeline for all-in-one blind image restoration to Restore Anything with Masks (RAM). We focus on the image content by utilizing Mask Image Modeling to extract intrinsic image information rather than distinguishing degradation types like other methods. Our pipeline consists of two stages: masked image pre-training and fine-tuning with mask attribute conductance. We design a straightforward masking pre-training approach specifically tailored for all-in-one image restoration. This approach enhances networks to prioritize the extraction of image content priors from various degradations, resulting in a more balanced performance across different restoration tasks and achieving stronger overall results. To bridge the gap of input integrity while preserving learned image priors as much as possible, we selectively fine-tuned a small portion of the layers. Specifically, the importance of each layer is ranked by the proposed Mask Attribute Conductance (MAC), and the layers with higher contributions are selected for finetuning. Extensive experiments demonstrate that our method achieves state-of-the-art performance. Our code and model will be released at https://github.com/Dragonisss/RAM. + +Keywords: Image Restoration $\cdot$ All-in-One $\cdot$ Mask Image Modeling + +# 1 Introduction + +Image restoration involves the restoration of low-quality images affected by various degradation, typically arising from adverse environmental conditions (e.g., + +rain, haze, low-light), hardware-related issues (e.g., noise and blur), and post-processing artifacts (e.g., JPEG compression). Image restoration serves not only to enhance the visual appeal of images but also contributes to practical application scenarios such as autonomous driving and surveillance. + +Modern techniques in this field mainly focus on learning fixed patterns formed during the degradation process, i.e., degradation priors. Some works [29, 30, 60] utilize task-specific priors to solve a certain degradation problem, while another research line [3, 28, 39, 48, 55] tries to design a general network architecture that can effectively learn each degradation pattern. Nevertheless, the above methods only enable the network to learn a single degradation, resulting in an imbalanced situation when dealing with multiple types of degradation. + +To tackle the problem stated above, all-in-one methods have emerged, aiming to handle multiple degradations using one model. Most of these approaches tend to utilize explicit priors (e.g., AirNet [21]) or introduce an extra module (e.g., PromptIR [42]) to discern image degradation patterns, thereby assisting + +![](images/06ec5f2f8dd06754772466ca6479e1a8f2b71672aa029fa0b4c8a52e05e7c4ca.jpg) +Fig.1: Our RAM achieves more balanced and more powerful performance than the state-of-the-art methods (AirNet [21], TAPE [31], PromptIR [42]) for all-in-one blind image restoration. + +the model in performing the restoration. However, these methods place their emphasis on distinguishing degradation types in images rather than the image content, leading to lower scalability and fuzzy decision boundaries when more degradation types are involved. We argue that the essence of image restoration is to extract intrinsic image information from corrupted images rather than eliminate degradation patterns, i.e., learning image prior rather than degradation prior. It is worth noting that TAPE [31] similarly suggests that understanding normal image nature aids restoration by introducing a natural image prior. Nevertheless, TAPE utilizes the model output as the optimization target, which causes the model to amplify its own errors and learn the image prior with bias. + +In this paper, we focus on tackling how to extract intrinsic image information from diverse corrupted images. Some attempts [2, 7] by Mask Image Modeling (MIM) in low-level vision have caught our attention. As a pretraining strategy, MIM has been widely validated for its effectiveness in high-level tasks, thanks to its generic representation of images. Simultaneously, the model also learns the distribution of natural images, which encompasses the intrinsic information we aim to extract from the images. Built on MIM, we propose a simple pipeline for all-in-one blind image restoration that Restores Anything with Masks (RAM), which includes two stages: the mask pre-training stage and the + +fine-tuning stage with Mask Attribute Conductance (MAC). In the pre-training stage, we randomly mask corrupted images at the pixel-wise level and force the network to predict the clear one corresponding to the masked pixels, extracting inherent image information from corrupted images. In the fine-tuning stage, we focus on overcoming the input integrity gap caused by changing masked input during pre-training into the whole image during inference while preserving learned prior as much as possible. + +Specifically, we first evaluated the importance of each network layer in addressing this gap by the proposed MAC. Following that, we chose the top $k\%$ most critical layers for fine-tuning while keeping the rest of the network layers frozen. We demonstrate that after a brief fine-tuning period (even if only $10\%$ layers are tuned), the model can achieve a highly satisfactory performance level, surpassing models trained using traditional pair-wise training. Additionally, our pipeline can be plug-and-play used in any network without introducing additional computational overhead. + +The contributions of this work are as follows: + +- We discuss the challenge of adopting MIM in low-level vision and propose a MIM-based pre-training strategy tailored to all-in-one blind image restoration, which allows the restoration networks to effectively learn inherent image information while guaranteeing reconstruction results. +- We proposed Mask Attribute Conductance to evaluate the importance of each layer in addressing the input integrity gap so that a very small portion (e.g. $10\%$ ) of critical layers are tuned to bridge this gap while preserving the image prior learned by MIM. +- Our proposed RAM provides a fresh perspective to achieve more balanced and powerful all-in-one blind image restoration, which focuses on extracting inherent image information from corrupted images. Our pipeline can be applied to any image restoration network without introducing additional computational overhead. + +# 2 Related Work + +# 2.1 Image Restoration for Multi Degradations + +While neural networks have demonstrated impressive performance in single degradation image restoration [9, 13, 14, 17, 22, 23, 29, 30, 49, 60], recent works have shifted their focus towards addressing the more challenging domain of multi-degradation image restoration. A group of methods [3, 28, 39, 48, 55] aims at designing a general architecture that can effectively learn each degradation pattern. SwinIR [28] employs a window attention mechanism to convert global attention into a localized approach, effectively reducing computational overhead. In addition, the U-shaped transformer-based methods [48, 55] are employed to extract multi-scale features and reduce computational overhead. However, these methods have to train individually on each restoration task. Several methods [1, 24] leverage multiple input and output heads to empower the network to restore + +![](images/97b3ce8594e2da099216f661a7efbf5f91f85036c934e86b9a3d32f0dd9f02bb.jpg) +Fig. 2: The illumination of our overall pipeline. 1) Pre-training the model with mask image pre-training method tailored to low-level vision. We randomly mask degraded images at the pixel level with a $50\%$ masking ratio and reconstruct the clean images. 2) The Fine-tuning stage is followed to overcome the input integrity gap caused by changing masked input during pre-training into the whole image during inference. We analyze the importance of each network layer for resolving the input integrity gap according to the proposed MAC and rank them in descending order. The top $k\%$ of network layers are selected for fine-tuning on the complete image. + +various types of degraded images. Nonetheless, this kind of approach may lead to the diminished scalability of the model. Recently, several subsequent methods [4, 21, 36, 41, 42, 56, 61] have been proposed to employ a unified network to address multiple restoration issues. Most of these methods put emphasis on learning how to distinguish different types of degradations and restore corrupted images. Typically, AirNet [21] first proposed an all-in-one image restoration task. The method initially pretrains a degradation classifier based on contrastive learning and subsequently utilizes it to assist in all-in-one image restoration. PromptIR [42] has introduced a learnable prompt-based module. Instead of constraining the degradation category, it enables the model to autonomously learn features that are advantageous to its performance by using an adaptive prompt. Our RAM takes a fresh perspective that focuses on extracting common content information from corrupted images, without any extra design to distinguish degradations, which helps us achieve balance and powerful performance when more degradation types are taken into consideration. + +# 2.2 Mask Image Modeling + +Inspired by Mask Language Modeling [18,43], Mask Image Modeling (MIM) [15, 51] is introduced as a pretraining approach to learn general representations in high-level vision. MAE [15] effectively utilizes MIM for predicting hidden tokens, demonstrating strong performance and generalization across various downstream tasks. SimMIM [51] proposed a general masked image modeling method based on Swin-ViT [34]. Painter [47] unifies multiple tasks under image-to-image trans + +lation and leverages MIM pretraining. In recent years, there have been efforts to incorporate MIM into the realm of low-level vision to enhance model generalization. Among them, [2] and [7] are the most closely aligned with our focus. [2] employs the MIM model to enhance the model's generalization for denoising tasks but has not explored its potential in multi-task scenarios. [7] utilizes MIM for pre-training the model encoder to introduce generative prior and subsequently employs the decoder for restoration. However, it does not fully harness the potential of MIM. Our proposed RAM utilizes MIM to unify the optimization objective for various image restoration tasks into reconstructing intrinsic image information. This allows the network to learn restoration functions more balanced and effectively. Moreover, to preserve the image priors learned by MIM, we designed a fine-tuning strategy based on MAC analysis (in Sec. 3.3). This enables us to achieve comparable performance by fine-tuning only a small portion (e.g. $10\%$ ) of layers, fully tapping into the potential of MIM. + +# 2.3 Gradient-based Attribution + +Gradient-based attribution methods [6,12,44-46,50] are often used to clarify how hidden units (or inputs) impact the output of networks. One commonly used approach is Integrated Gradients (IG) [45,46], which accumulates gradients along a linear path from the baseline input to the target input in the pixel/feature space. After that, IntInf [19] and layer conductance [6] alter IG to attribute neuron importance along the same path. In our work, we expect to find the key layers that can effectively overcome the distribution shift between training data and inference data. We propose Mask Attribute Conductance (MAC) based on the layer conductance and accumulated MAC of each layer along the Mask Attribute Path (MAP). MAC can represent the layer's importance along the MAP. In this way, we can fine-tune the top $k\%$ critical layers of the pre-trained network, preserving to a great extent the image priors learned during pretraining. + +# 3 Methodology + +In this section, we start with discussing the challenges of using MIM in low-level vision tasks (Sec. 3.1). Following that, we present our pipeline for all-in-one blind image restoration, which contains two parts: pre-training with MIM (Sec. 3.2) and fine-tuning with Mask Attribute Conductance (MAC) Analysis (Sec. 3.3). + +# 3.1 Rethinking MIM in Low-Level Vision + +MIM is a process that randomly masks certain parts of an image and extracts features from the remaining visible parts to reconstruct the entire image. It allows models to acquire a generic representation of images and thus achieve good pre-training, which is verified in many high-level tasks [15, 51]. Moreover, the models also learn the distribution of natural images during the image reconstruction, i.e. MIM pre-training. This incidental acquisition of prior knowledge + +![](images/26468efb0aa41a9754bbd12b714a4d07f6aaca410620d49d634ae47b3f851a9e.jpg) +patch size=1 + +![](images/40daf3bdc76771c1f65564c01a0834e29a029947d8e28ccf094ae4bed932624a.jpg) +patch size=4 +Fig. 3: Mask Image Modeling reconstruction with different patch sizes. We pre-trained with different patch sizes and visualized the mask inputs (left), and the corresponding MIM reconstructions (right). + +![](images/980d7a274f6a681b09fa405bb062263e0583c28bfc40c0b776bb8fa36ff9c0a1.jpg) +patch size=8 + +is instrumental in tasks like image restoration. Despite these advantages, applying MIM in pretraining a model for low-level vision tasks is still under-explored, primarily due to the challenges that must be addressed in the process. + +Firstly, the main purpose of vanilla MIM is not high-quality reconstruction but good feature extraction for high-level tasks. Therefore, it masks a wider range of images to gather semantic information but not pixel-level content, reflected in token-level masking and a high mask ratio. CSFormer [7] directly adopts this strategy on low-level vision pre-training. However, some studies verify that semantic information is not as important for image restoration as it is in pattern recognition tasks [33, 37]. Moreover, high-degree masking leads to producing detail-deficient results, as shown in Fig. 3, which is harmful to low-level tasks. + +Secondly, the training objective of MIM is to reconstruct the masked input images, so it can only produce results with the same domain as the input image. However, we hope the model gains the ability to bridge low-quality domain to high-quality domain, i.e. recover clean content from degraded input. Therefore, it is necessary to introduce paired data when pre-training image restoration models by MIM (see the experiment in Sec. 4.3 for details). Chen et al. [2] demonstrate that pair-wise MIM training enhances the generalization performance over different types of noisy images. In this paper, we take a step forward to explore the effectiveness of MIM on multiple degradations with larger variance. + +# 3.2 Pretraining with MIM + +Based on the above analysis, we design a MIM pre-training paradigm tailored for low-level vision. + +Masking. During the pre-training stage, we randomly mask the pixels of degraded images (mask images in a $1 \times 1$ patch size) with a $50\%$ mask ratio. We found that fine-grained masked patches and balanced mask ratio are beneficial to image restoration, which can be demonstrated in Sec. 4.3. + +Besides, since our MIM pre-training has a similar target to subsequent low-level tasks, we do not need to change the decoder like MAE [15] does but just fine-tune it. + +![](images/ddd442039b45f8792232e47e1eb3152d425b123907e380a8991c86ee504114ed.jpg) +Fig. 4: The effect of MIM reconstruction with different input integrity on kernel deblurring (orange border) and denoising (blue border). We also visualize the color distributions of reconstructions in various tasks above. It shows that the distribution of the reconstruction results obtained using the twin-masks method as input is closer to the real images (ground truth) compared to the results obtained using the whole input. + +Reconstruction target. Following the Bert [18] and MAE [15], we choose L1 loss to supervise the masked part. The training objective can be written as: + +$$ +\underset {\theta} {\arg \min } \mathbb {E} [ | | \tilde {\mathcal {M}} (I - f (\mathcal {M} (I _ {d}), \theta)) | | ], \tag {1} +$$ + +where $\{I, I_d\}$ represents a pair of clean image and degraded image, $f(\cdot, \theta)$ denotes a network with parameters $\theta$ , $\mathcal{M}(\cdot)$ is a random binary masking operation and $\tilde{\mathcal{M}}(\cdot) = 1 - \mathcal{M}(\cdot)$ . + +# 3.3 Finetuning with Mask Attribute Conductance Analysis + +Observation. During pre-training, the network learns rich content priors. However, the incompleteness of the masked input prevents the direct use of the pre-trained model for inference, as it would result in a distribution shift in the outputs. As shown in Fig. 4, We start by feeding the entire image into a pretrained model, leading to a color-distorted result. Next, we use a pair of complementary masks, referred to as twin-masks, to individually mask the image. Subsequently, we input both of these complementarily masked images into the network. By combining the pixel values predicted by each image, we generate a higher-quality image. This observation indicates that the hindrance to using + +mask pre-trained model directly for inference lies in input incompleteness rather than the model's inability to learn the restoration function. + +Building upon this insight, we explore the possibility of minimizing the influence of disparities in data input formats via model fine-tuning. To maintain the learned priors, it is essential to retain pre-trained parameters as extensively as possible while employing the fewest but most effective layers for fine-tuning. To tackle this, we introduce the concept of mask attribution conductance, which quantifies the importance of each layer concerning the fine-tuning objective. We then identify the top- $k\%$ most critical layers for fine-tuning. + +Preliminary. Before giving the definition of Mask Attribute Conductance (MAC), we briefly recall the definition of integrate gradient [45] (IG) and neuron conductance [6] (Cond). Considering a linear path $\gamma (\alpha) = x^{\prime} + \alpha (x - x^{\prime})$ from base input $x^{\prime}$ to target input $x$ , we can attribute output change $F(x) - F(x^{\prime})$ to $i$ -th dimension of input/feature $x_{i}$ (e.g. a pixel) by calculating its integrate gradient, which formally as below: + +$$ +\operatorname {I G} _ {i} (x) := \left(x _ {i} - x _ {i} ^ {\prime}\right) \cdot \int_ {0} ^ {1} \frac {\partial F \left(x ^ {\prime} + \alpha \left(x - x ^ {\prime}\right)\right)}{\partial x _ {i}} d \alpha . \tag {2} +$$ + +We can also attribute output change to a specific neuron $y$ by improving IG, which involves calculating the conductance. The conductance [6] of the hidden neuron $y$ along the $\gamma(\alpha)$ is: + +$$ +\begin{array}{l} \operatorname {C o n d} ^ {y} (x) := \sum_ {i} \left(x _ {i} - x _ {i} ^ {\prime}\right) \cdot \int_ {0} ^ {1} \frac {\partial F \left(x ^ {\prime} + \alpha \left(x - x ^ {\prime}\right)\right)}{\partial y} \cdot \frac {\partial y}{\partial x _ {i}} d \alpha \tag {3} \\ = \sum_ {i} \int_ {0} ^ {1} \frac {\partial F (\gamma (\alpha))}{\partial y} \cdot \frac {\partial y}{\partial \alpha} d \alpha , \\ \end{array} +$$ + +Note that $(x_{i} - x_{i}^{\prime}) = \frac{\partial(x^{\prime} + \alpha(x_{i} - x_{i}^{\prime}))}{\partial\alpha}$ . Certainly, we can broaden Eq. (3) to compute conductance when integrating along any given path $\alpha :[s,t]\to P$ : + +$$ +\operatorname {G e n e r a l C o n d} ^ {y} (x) := \sum_ {i} \int_ {P} \frac {\partial F (X _ {i} (\alpha))}{\partial y} \cdot \frac {\partial y}{\partial \alpha} d \alpha , \tag {4} +$$ + +where $X: R \to R^n$ is the function of the path from $x'$ to $x$ , which satisfies $X(s) = x'$ , $X(t) = x$ . [s,t] represent the domain of the path function $X$ ." + +Finetuning with MAC. To find effective layers to finetune, we propose Mask Attribute Conductance (MAC) to evaluate how effective each layer is in overcoming the gap of input integrity. Considering such a nonlinear path $\alpha : [0,1] \to P_m$ from zero input $x'$ to whole input $x$ , which path function $X^m$ satisfies: + +$$ +X _ {i} ^ {m} (\alpha ; \alpha_ {i}) = \left\{ \begin{array}{l l} x _ {i} ^ {\prime}, & \alpha < \alpha_ {i} \\ x _ {i}, & e l s e \end{array} , \right. \tag {5} +$$ + +where $i$ refers to the index of pixels, $\alpha_{i} \in (0,1]$ is a set of parameters that indicate when each pixel gets masked. We define this path as a Mask Attribute Path (MAP). Apparently, $X^{m}(0) = x^{\prime}$ and $X^{m}(1) = x$ . + +![](images/79c2fd74506cff6d5f6752e8da693a01e93ccc21182850c9ed51e06e5cd32ae9.jpg) +(a) + +![](images/1a9a4c98871017718d833e194f72449e2da08834b4440b4682743ad47424f0af.jpg) +(b) +Fig. 5: Illumination of (a) $X_{i}^{m}$ in Eq. (5) and (b) $\tilde{X}_i^m$ in Eq. (6). + +However, $X^{m}$ is not differentiable, making it an invalid attribute path function. To solve this problem, we use a group of sigmoid-like functions $\tilde{X}^{m}$ to approximate $X^{m}$ : + +$$ +\tilde {X} _ {i} ^ {m} \left(\alpha ; \alpha_ {i}\right) = \frac {\left(x _ {i} ^ {\prime} - x _ {i}\right)}{1 + e ^ {- \delta \left(x _ {i} ^ {\prime} - \alpha_ {i}\right)}}. \tag {6} +$$ + +We can see that $\tilde{X}^m$ is very close to $X^m$ when $\delta$ is sufficiently large (as depicted in Fig. 5). And for each $\tilde{X}_i^m$ , it will change sharply from $x_i'$ to $x_i$ when $\alpha$ is in the neighborhood of $\alpha_i$ . + +Here, we can give a definition of MAC as below: + +$$ +\begin{array}{l} \mathrm {M A C} ^ {y} (x) := \sum_ {i} \int_ {P _ {m}} \frac {\partial F (X _ {i} (\alpha))}{\partial y} \cdot \frac {\partial y}{\partial \alpha} d \alpha \\ \approx \sum_ {i} \int_ {0} ^ {1} \frac {\partial F \left(\tilde {X} _ {i} ^ {m} (\alpha ; \alpha_ {i})\right)}{\partial y} \cdot \frac {\partial y}{\partial \alpha} d \alpha . \tag {7} \\ \end{array} +$$ + +In fact, a partial path is also available to attribute from a masked input $x_{m}$ with any mask ratio $r$ to whole input $x$ : + +$$ +\mathrm {M A C} _ {r} ^ {y} (x) \approx \sum_ {i} \int_ {1 - r} ^ {1} \frac {\partial F \left(\tilde {X} _ {i} ^ {m} (\alpha ; \alpha_ {i})\right)}{\partial y} \cdot \frac {\partial y}{\partial \alpha} d \alpha . \tag {8} +$$ + +In practice, we use N-steps discretization to approximate the integral form of Eq. (8), which follows [44]: + +$$ +\begin{array}{l} \mathrm {M A C} _ {r} ^ {y} (x) \approx \sum_ {i} \sum_ {j = 1} ^ {N} \frac {\partial F \left(\tilde {X} _ {i} ^ {m} \left(\frac {j r}{N} ; \alpha_ {i}\right)\right)}{\partial y} \tag {9} \\ \cdot \left(F _ {y} \left(\tilde {X} _ {i} ^ {m} \left(\frac {(j + 1) r}{N}\right)\right) - F _ {y} \left(\tilde {X} _ {i} ^ {m} \left(\frac {j r}{N}\right)\right)\right). \\ \end{array} +$$ + +We compute the MAC of each layer of pre-trained networks, rank them in descending order based on their MAC values, and pick top- $k\%$ layers for finetuning. The networks are initialized by pre-trained weight and only top- $k\%$ layers + +Table 1: Quantitative comparison on seven challenging image restoration tasks, including dehazing, deraining, denoising, motion deblurring, low-light image enhancement (LLIE), kernel deblurring, and JPEG artifact removal. **boldface** and **underlined** indicate the best and second-best results, respectively. + +
MethodSOTS [20]Rain13k-Test [32]BSD68 [38]GoPro [40]LOL [5]LSDIR-Blur [26]LSDIR-Jpeg [26]Average
PSNR†/SSIM†PSNR†/SSIM†PSNR†/SSIM†PSNR†/SSIM†PSNR†/SSIM†PSNR†/SSIM†PSNR†/SSIM†PSNR†/SSIM†
Restomer [55]22.89/0.917227.05/0.846930.95/0.865727.46/0.849723.65/0.845819.60/0.365830.46/0.914126.01/0.8007
MPRNet [39]25.23/0.946325.36/0.806829.83/0.831725.90/0.794922.29/0.817025.68/0.828128.96/0.886526.18/0.8445
NAFNet [3]25.74/0.944524.65/0.787730.37/0.854025.53/0.790921.50/0.810429.08/0.913029.09/0.895526.57/0.8566
DL [8]21.16/0.904219.56/0.650816.15/0.586117.63/0.586219.26/0.777717.98/0.612119.55/0.696518.75/0.6877
TAPE [31]25.14/0.931923.66/0.781830.11/0.835425.97/0.796218.95/0.763224.26/0.765429.28/0.896525.34/0.8243
AirNet [21]21.66/0.836620.21/0.640227.99/0.725023.36/0.750316.65/0.670823.84/0.735824.36/0.802022.58/0.7372
SwinIR [28]27.29/0.962225.32/0.825830.65/0.854026.61/0.812518.66/0.804827.82/0.883930.13/0.907126.64/0.8643
RAM-SwinIR28.47/0.968926.31/0.848630.83/0.861126.89/0.820021.62/0.829126.66/0.851430.22/0.909627.28/0.8698
PromptIR [42]28.70/0.965927.46/0.858530.84/0.862527.71/0.856521.19/0.835631.01/0.938530.30/0.911728.17/0.8899
RAM-PromptIR29.64/0.969528.47/0.875130.86/0.862428.02/0.859224.46/0.858129.57/0.917930.33/0.911928.76/0.8935
+ +![](images/b91e724a60e506ed9c17e75dc3fa51915f8d0dbc0bc0b38ca7d21353b2431304.jpg) +Fig. 6: Dehaze visual comparison on SOTS dataset. Zoom in for details. + +will be fine-tuned. More implementation details can be found in the supplementary material. + +# 4 Experiment + +# 4.1 Experiments Settings + +Datasets and Metrics. We combine datasets from various restoration tasks to form the training set, following [59]. For high-cost tasks that degradations are difficult to synthesize, we leverage existing paired datasets, including RE-SIDE [20] for dehazing, Rain13k [10, 25, 27, 35, 53] for deraining, GoPro [40] for motion deblurring, and LOL-v2 [54] for low-light image enhancement (LLIE). For low-cost tasks that degradations are easy to synthesize (e.g. noise, kernel blur, and JPEG artifact), we generate corrupted images on the LSDIR dataset [26] during the training process, which involves generating Gaussian noise with random variation $\sigma \in (0, 50]$ , creating gaussian blurred images with a blur kernel of size $k = 15$ and random $\sigma \in [0.1, 3.1]$ , and introducing JPEG artifacts with a random quality parameter $q \in [20, 90]$ . + +For evaluation, we use SOTS-outdoor [20] for dehazing, Rain13k-Test (the combination of Rain100L [52], Rain100H [52], Test100 [58], Test1200 [57] and + +Table 2: Quantitative Gaussian denoising results at different noise levels on BSD68 and Urban100 datasets in terms of PSNR. + +
MethodBSD68 [38]Urban100 [16]
σ = 15σ = 25σ = 50Averageσ = 15σ = 25σ = 50Average
NAFNet [3]33.2230.5927.3030.3732.6730.2126.9729.92
MPRNet [39]32.7330.1126.6529.8332.0629.4625.7729.10
Restormer [55]33.7931.1727.9030.9533.8331.4027.9931.07
DL [8]16.0416.2016.1916.1519.1719.1118.4718.92
TAPE [31]33.1030.3726.8630.1132.5929.9326.1929.57
AirNet [21]31.6328.8323.5227.9929.7926.9021.3526.01
SwinIR [28]33.5330.8927.5430.6533.5030.9927.3730.62
RAM-SwinIR33.6531.0627.7730.8233.8231.4327.9431.07
performance gains(↑0.12)(↑0.17)(↑0.23)(↑0.17)(↑0.32)(↑0.44)(↑0.57)(↑0.45)
PromptIR [42]33.6731.0627.8030.8433.5631.0827.6430.76
RAM-PromptIR33.7031.0827.7930.8633.7031.3027.9230.97
performance gains(↑0.03)(↑0.02)(↓0.01)(↑0.02)(↑0.14)(↑0.22)(↑0.28)(↑0.21)
+ +Test2800 [11]) for deraining, GoPro for motion deblurring, LOL [5] for low-light enhancement, BSD68 [38] for denoising, LSDIR-val for kernel deblurring and JPEG artifact removal. Furthermore, We conducted evaluations including denoising tests with variances of 15, 25, and 50, deblurring tests at $k = 15$ and $\sigma = 2.0$ , and JPEG artifact removal tests at $q = 50$ . + +Implementation Details. We apply our proposed RAM to SwinIR [28] and PromptIR [42]. The input size for RAM-SwinIR is 64, while for RAM-PromptIR it is 128. During the pre-training phase, we use the Adam optimizer to train RAM-SwinIR and RAM-PromptIR for 300 epochs, with the learning rate decaying from 1e-4 to 6e-5 following a cosine schedule. In the fine-tuning phase, we use the Adam optimizer to fine-tune the network layers obtained from the MAC analysis of RAM-SwinIR and RAM-PromptIR for 40 epochs, with the learning rate decaying from 2e-4 to 1e-7 following a cosine schedule. The batch sizes for RAM-SwinIR and RAM-PromptIR during the pre-training and fine-tuning phases are (12,4) and (4,4), respectively. + +# 4.2 Comparisons + +To validate the gain capability and effectiveness of our RAM, we apply the proposed RAM to SwinIR (a general image restoration method) and PromptIR (an all-in-one image restoration method). Four general architecture-based image restoration methods [3, 28, 39, 55] and four all-in-one methods [8, 21, 31, 42] are considered for comparison. We ensure that the number of supervised pixels employed by all other methods equals that used during the pre-training stage. + +As illustrated in Tab. 1, our approach achieves the best or comparable performance on each task. On the average score across seven different tasks, our method with PromptIR [42] achieves 0.59dB performance gains compared to the second-best algorithm. Besides, the SwinIR equipped with RAM also yields $2.40\%$ improvement on PSNR. Specifically, our RAM has significant benefits for dehazing and low-light enhancement. Tab. 2 shows the Quantitative denois- + +![](images/97fede3a9740dae1b8177c4426144258ecb5b722a4137b27f2e7016a98b6cb61.jpg) +Rainy Image + +![](images/c5456c7b9deedb3b3cc5bfd0eed573d67e9dd2969b912482006b2c533f55e21a.jpg) +NAFNet + +![](images/bcf69f06dddfb2989e89877bfb4726d905d3a5a1e14ccc67f8c7daba573e27f0.jpg) +MPRNet + +![](images/9547db10f802091a735362c2eaeded7edeaabf60ca54ae16000a371da88600be.jpg) +Restomer + +![](images/c7a068f252cd8e96241a0d432e7444aee4c0213bb00a7323a6c0f605bb396efa.jpg) +SwinIR + +![](images/760c799c2591cd1c736f81b6798398168d10b5bda311a6c6e99b6745b1f0106f.jpg) +RAM-SwinIR (Ours) + +![](images/a142d342183e62f3d0950629041f17e2d9b30fca3ba1a7e709c48f26c7297ae9.jpg) +DL + +![](images/cf0b4daf1123d33c2a218cc2ef17368bd3d5b7debf7f945214e537348290a719.jpg) +TAPE + +![](images/db3404ecfb3504bd5e7963af1234c9778334eb144ae16d2a97c982209a77c1c6.jpg) +AirNet + +![](images/4f5f26f456fe7781fd76bb1d3671a4605f767bc764e50ff705eb4ddedade53c0.jpg) +PromptIR +RAM-PromptIR(Ours) + +![](images/d13a3871a0aba2898f675c9e3000a4fd6e484e63bba71ac92ccd600a68fb1f1d.jpg) + +![](images/b1d8b0a8235a3f965c0d9f7781132d8359ca0c48dd7a6d47b378f3875aadf756.jpg) +GT + +![](images/d392f4b13367f37112eb00350263198a734a9aa54c82abbe7ec6a13a5feee113.jpg) +Fig. 7: Derain visual comparison on Rain13k-Test dataset. Zoom in for details. +Blurred Image + +![](images/0da33db259b06b5e0ed7743e91b632843ea4fa696c8afa3a6b133b7f105fed17.jpg) +NAFNet + +![](images/09842421063513a58bdd468f900e69b29fe537db7d36982428dd4bf5d1167b08.jpg) +MPRNet + +![](images/62f4c2ecf477dbd9baa40f009323f3128f9e24af0734db30bde5edac1e3564df.jpg) +Restormer + +![](images/4a78e83d2105f41649b4693be56c563f86e2ff4dbaedeae4f3c2154ed8465f35.jpg) +RAM-SwinIR (Ours) + +![](images/f3dedb93749e3aeb4a982bfeb89804f435eb1d9cbbf021d349fdc59878d82874.jpg) +DL +Fig. 8: Motion deblur visual comparison on GoPro dataset. Zoom in for details. + +![](images/f76c879fde6b574c1a8d76f24b8eb05e5854c9141c2a4a8483b5aab6e769707a.jpg) +TAPE + +![](images/84d6d2b3e3679eca20a4fa4d761a3ee97088fbde5a6ec0d9fbc09810410e2701.jpg) +AirNet + +![](images/7ce785fe1109627eda1e07fefe0c38db60de1f2ecfaa9fc617d95a4b3e3904a5.jpg) +PromptIR +RAM-PromptIR(Ours) + +![](images/8a8e13692bf76d3b1d7c71dc24f60cd7a08a179450d5da0e7934fd7c5f1fd6e3.jpg) +GT + +ing result at different noise levels. Both RAM-SwinIR and RAM-PromptIR get higher performance than the origin versions. + +Fig. 6-Fig. 10 show the qualitative results of various methods on different datasets. In Fig. 6, our method achieves better dehazing effects (right region) and exposure correction (sky). In the deraining task (Fig. 7, our method better removes rain streaks and restores textures in the occluded regions. In terms of denoising (Fig. 9) and deblurring (Fig. 8), we achieve clearer results with fewer artifacts. We also demonstrate better color correction (the purple blanket on the left) and exposure correction in low-light image enhancement tasks (Fig. 10). For simplicity, the qualitative effects of kernel deblurring and JPEG artifact removal will be presented in the supplementary material. + +# 4.3 Ablation Study + +In this section, we conduct an ablative study on the masking ratio, mask patch size, pre-training strategy, fine-tuning strategy, and fine-tuning ratio to demonstrate the effectiveness of our MIM pre-training and fine-tuning strategy. + +Table 3: Ablative results on masking ratios. + +
Masking ratio20%40%50%60%80%
PSNR↑27.2827.2127.2827.2627.08
SSIM↑0.86630.86830.86980.86940.8642
+ +Table 4: Ablative results of different pre-training strategies. + +
RAM-SwinIRPSNR↑SSIM↑
pre-trained w/ gt26.620.8580
pre-trained w/ paired data27.280.8698
+ +Table 5: Ablative results of different fine-tuning strategies. + +
RAM-SwinIRPSNR↑SSIM↑
random26.860.8535
IG [45]26.920.8554
MAC (Ours)27.280.8698
+ +- Patch size & masking ratio are two essential hyper-parameters that determine the continuity and area of the masking of an image. In high-level tasks, MAE [15] masks $75\%$ of an image with $16 \times 16$ patch size. However, it can corrupt the local details of images, which is not suitable for image restoration. + +Table 6: Ablative results in terms of the PSNR on fine-tuning ratios. We compared the performance in restoring images with unseen noises (Out-of-Distribution Denoising) and known degraded images (In-Distribution). In this case, the settings of In-Distribution are the same as Tab. 1. + +
MethodOut-of-DistributionIn-Distribution
PossionPepperSpeckleAverageAverage
SwinIR [28]12.8310.0020.8614.5626.64
RAM-SwinIR10%13.6719.2321.0717.9927.28
RAM-SwinIR20%13.2719.0920.6817.6827.35
RAM-SwinIR50%12.7516.5120.3616.5427.38
RAM-SwinIR100%12.4715.3120.0115.9327.54
+ +We first find the best choice of patch size by pretraining SwinIR [28] on $1 \times 1$ , $4 \times 4$ , and $8 \times 8$ , as shown in Fig. 3. Since the attention layers of SwinIR treat an $8 \times 8$ patch as a token, the $4 \times 4$ pre-training produces heavy artifacts. Besides, the results generated by $8 \times 8$ pretraining are highly missing details, e.g. the texture of the polar bear's paws. In contrast, the model pre-trained with $1 \times 1$ patch size, which is also our final choice, achieves a satisfactory reconstruction and removes most of the rain streaks. + +Then, we adjust the masking ratio from $20\%$ to $80\%$ . As we can see in Tab. 3, the model pre-trained with $50\%$ achieves the highest performance. Moreover, the performance is significantly dropped from 27.28dB to 27.08dB in terms of PSNR when we continue to increase the masking ratio, which also demonstrates our opinion that a high masking ratio is harmful to image restoration. + +Pre-trained with paired data. Tab. 4 compares the results of using paired data for mask image pretraining (our pretraining strategy) with those using only ground truth for mask image pretraining. It shows that pre-trained with paired data is necessary for our RAM. Pretraining the model on high-quality images does not effectively enable learning for image restoration tasks. It still requires paired data to guide the model in the learning process. + +Fine-tuning strategy. To verify the effectiveness of our fine-tuning strategy, we fine-tune $10\%$ of the network layers selected through MAC analysis, IG [45], and uniform sampling, respectively, and the results are shown in Tab. 5. Compared + +![](images/3eb1c6379e11b9c597a9cd4116766d3fd8f7977cbe64dcef33609fe2dfce767f.jpg) + +![](images/c10fb5aaa480258cf5e68e0dea376870bff31258f391c4586b25767e271fc969.jpg) + +![](images/f296c3bf94de43094ed2fad425930485a7c873ae2da714e4892e99635b4fdfe0.jpg) + +![](images/0ef2171be860cc4c21eb3cc1b06a69c652c0c7c0f72604d763cfc912fbee288a.jpg) + +![](images/97647f418b561be0583fe5ecf511d7072ed9c2ec05d10a975e8a6148594376e0.jpg) + +![](images/5cdd7d46956bb4566507ede54decbc9374f95a75e8b1fff6aff1b2eade288dc7.jpg) + +![](images/5ab121741bc61962925984b993557e0bb4a60d5a3decbcdc2ba1f198c968b5e1.jpg) +Noisy Image +DL + +![](images/dab4a93d4ef7270c882be456133b1181b65981ec386f39208741bb1fb6531885.jpg) +NAFNet +TAPE + +![](images/a8840f4335f7694ffd376fa7ad27df1e67c961863c03c3963a1436609ab92c4a.jpg) +MPRNet +AirNet + +![](images/39214ef1d42f254db74bdbe602a7fbfaffbbb82e9594e195d8848aea9085d740.jpg) +Restormer +PromptIR +RAM-PromptIR(Ours) + +![](images/561a7269f3e4b6dfc33ee010a43da46c8764a9bc63ce403b3ac342a8e1e2895a.jpg) +SwinIR + +![](images/d399dc62949040a87f8d79296eecf419751f802a64644d15b90245742af82a31.jpg) +RAM-SwinIR (Ours) +GT + +![](images/ae9e45b9f67927b8d0f715be350c576819dc7302813b65df51d73b399e403d4b.jpg) +Fig. 9: Denoising visual comparison on CBSD68 dataset. Zoom in for details. + +![](images/a91afc1ef3fa620d198399a06534bd079b75325f645256e0405a3a982694e344.jpg) +NAFNet + +![](images/e38c71cd938d93468c2150fceb3c73455849a23f02e3cc1de139c6f4dc27067b.jpg) +MPRNet + +![](images/0baa7a8329853d6f7e5293d32c53e9c4f9079b212dc4ea4bd9c700e37f0aa3a5.jpg) +Restormer + +![](images/a8db619b831cfdc3c4be0ddbb22abd61876e69ed84c00da1e2f00e194e05471c.jpg) +SwinIR + +![](images/8079c7c696c821a012b907fdb96052ca3bdec04c3eeb189f09b69deb2b549c22.jpg) + +![](images/a3b2f6aa0992184b49931ed447c7f5dc813ebfe3963c0dcaf856825c35052b82.jpg) +Low Light Image +DL +Fig. 10: LLIE visual comparison on LOL dataset. Zoom in for details. + +![](images/cada7c403bec112a9934f1a6819f52ae2f81868d1b3cf4ece8d1bb7346599167.jpg) +TAPE + +![](images/d2881532b49bd7489af5a3b6db650445f1bc2fc21c9973bbb4d0b40ab7aed783.jpg) +AirNet + +![](images/92e68fd5b163f2eb7728ab0a711655879deb41991c42e295532c7f2c6929dd71.jpg) +PromptIR +RAM-PromptIR(Ours) + +![](images/bc41aa406fd82a4b3a47980b6bcff8da394b2452a3049e6f7413464c9ac7b4fa.jpg) + +![](images/fd5ad5d957a22397858bfc0f8407e2ff4c70e4f553648befb592017ea15812cc.jpg) +RAM-SwinIR (Ours) +GT + +to IG, we have improved by 0.36 dB in PSNR and $1.6\%$ in SSIM, which indicates that our selection strategy is superior to IG. + +Fine-tune ratio. We conduct the ablation experiment to compare the network's performances with different fine-tune ratios in Tab. 6. We found that using our finetune strategy, a pre-trained network could achieve comparable performance by fine-tuning only a few layers (e.g. $10\%$ ). At the same time, we need to finetune almost all network parameters to get the best performance on given tasks. + +Performance vs Generalization capability. We found a trade-off between in-distribution performance and out-of-distribution generalization in Tab. 6. We found that the more layers fine-tuned, the less generalization capability to tackle the out-of-distribution tasks. With our fine-tuning method, the model can have stronger generalization while maintaining comparable performance. + +# 5 Conclusion + +This paper presents RAM, a pipeline for extracting intrinsic image information from corrupted images using Mask Image Modeling (MIM) pre-training. We design a MIM pre-training strategy tailored for image restoration and a fine-tuning algorithm to handle the transition from masked to complete images. By analyzing layer importance with MAC, we achieve high performance with minimal parameter tuning. Extensive experiments demonstrate that our RAM can bring boosts to various architectures and achieve state-of-the-art performance, moving towards a unified solution for all-in-one image restoration. + +# References + +1. Chen, H., Wang, Y., Guo, T., Xu, C., Deng, Y., Liu, Z., Ma, S., Xu, C., Xu, C., Gao, W.: Pre-trained image processing transformer. In: CVPR. pp. 12299-12310 (2021) +2. Chen, H., Gu, J., Liu, Y., Magid, S.A., Dong, C., Wang, Q., Pfister, H., Zhu, L.: Masked image training for generalizable deep image denoising. In: CVPR. pp. 1692-1703 (2023) +3. Chen, L., Chu, X., Zhang, X., Sun, J.: Simple baselines for image restoration. In: ECCV. pp. 17-33. Springer (2022) +4. Chen, W.T., Huang, Z.K., Tsai, C.C., Yang, H.H., Ding, J.J., Kuo, S.Y.: Learning multiple adverse weather removal via two-stage knowledge learning and multi-contrastive regularization: Toward a unified model. In: CVPR. pp. 17653-17662 (2022) +5. Chen Wei, Wenjing Wang, W.Y.J.L.: Deep retina decomposition for low-light enhancement. In: BMVC. British Machine VLOLision Association (2018) +6. Dhamdhere, K., Sundararajan, M., Yan, Q.: How important is a neuron. In: ICLR (2019) +7. Duan, H., Shen, W., Min, X., Tu, D., Teng, L., Wang, J., Zhai, G.: Masked autoencoders as image processors. arXiv preprint arXiv:2303.17316 (2023) +8. Fan, Q., Chen, D., Yuan, L., Hua, G., Yu, N., Chen, B.: A general decoupled learning framework for parameterized image operators. PAMI 43(1), 33-47 (2019) +9. Fang, Y., Zhang, H., Wong, H.S., Zeng, T.: A robust non-blind deblurring method using deep denoiser prior. In: CVPRW. pp. 735-744 (June 2022) +0. Fu, X., Huang, J., Ding, X., Liao, Y., Paisley, J.: Clearing the skies: A deep network architecture for single-image rain removal. TIP 26(6), 2944-2956 (2017) +1. Fu, X., Huang, J., Zeng, D., Huang, Y., Ding, X., Paisley, J.: Removing rain from single images via a deep detail network. In: CVPR. pp. 3855-3863 (2017) +2. Gu, J., Dong, C.: Interpreting super-resolution networks with local attribution maps. In: CVPR. pp. 9199-9208 (2021) +3. Guo, C.L., Yan, Q., Anwar, S., Cong, R., Ren, W., Li, C.: Image dehazing transformer with transmission-aware 3d position embedding. In: CVPR. pp. 5812-5820 (2022) +4. Guo, C., Li, C., Guo, J., Loy, C.C., Hou, J., Kwong, S., Cong, R.: Zero-reference deep curve estimation for low-light image enhancement. In: CVPR. pp. 1780-1789 (2020) +5. He, K., Chen, X., Xie, S., Li, Y., Dollár, P., Girshick., R.: Masked autoencoders are scalable vision learners. In: CVPR. pp. 16000-16009 (2022) +6. Huang, J.B., Singh, A., Ahuja, N.: Single image super-resolution from transformed self-exemplars. In: CVPR. pp. 5197-5206 (2015) +7. Jin, X., Han, L.H., Li, Z., Guo, C.L., Chai, Z., Li, C.: Dnf: Decouple and feedback network for seeing in the dark. In: CVPR. pp. 18135-18144 (2023) +8. Kenton, J.D.M.W.C., Toutanova, L.K.: Bert: Pre-training of deep bidirectional transformers for language understanding. In: NAACL. pp. 4171-4186 (2019) +9. Leino, K., Sen, S., Datta, A., Fredrikson, M., Li, L.: Influence-directed explanations for deep convolutional networks. In: ITC. pp. 1-8. IEEE (2018) +20. Li, B., Ren, W., Fu, D., Tao, D., Feng, D., Zeng, W., Wang, Z.: Benchmarking single-image dehazing and beyond. TIP 28(1), 492-505 (2018) +21. Li, B., Liu, X., Hu, P., Wu, Z., Lv, J., Peng, X.: All-in-one image restoration for unknown corruption. In: CVPR. pp. 17452-17462 (2022) + +22. Li, C., Guo, C.L., Liang, Z., Zhou, S., Feng, R., Loy, C.C., et al.: Embedding fourier for ultra-high-definition low-light image enhancement. In: ICLR (2022) +23. Li, D., Zhang, Y., Cheung, K.C., Wang, X., Qin, H., Li, H.: Learning degradation representations for image deblurring. In: ECCV. pp. 736-753. Springer (2022) +24. Li, R., Tan, R.T., Cheong, L.F.: All in one bad weather removal using architectural search. In: CVPR. pp. 3175-3185 (2020) +25. Li, X., Wu, J., Lin, Z., Liu, H., Zha, H.: Recurrent squeeze-and-excitation context aggregation net for single image deraining. In: ECCV. pp. 254-269 (2018) +26. Li, Y., Zhang, K., Liang, J., Cao, J., Liu, C., Gong, R., Zhang, Y., Tang, H., Liu, Y., Demandolx, D., et al.: Lsdir: A large scale dataset for image restoration. In: CVPR. pp. 1775-1787 (2023) +27. Li, Y., Tan, R.T., Guo, X., Lu, J., Brown, M.S.: Rain streak removal using layer priors. In: CVPR +28. Liang, J., Cao, J., Sun, G., Zhang, K., Van Gool, L., Timofte, R.: Swinir: Image restoration using swin transformer. In: CVPR. pp. 1833-1844 (2021) +29. Lin, X., Ren, C., Liu, X., Huang, J., Lei, Y.: Unsupervised image denoising in real-world scenarios via self-collaboration parallel generative adversarial branches. In: ICCV. pp. 12642-12652 (2023) +30. Lin, X., Yue, J., Ren, C., Guo, C.L., Li, C.: Unlocking low-light-rainy image restoration by pairwise degradation feature vector guidance. arXiv preprint arXiv:2305.03997 (2023) +31. Liu, L., Xie, L., Zhang, X., Yuan, S., Chen, X., Zhou, W., Li, H., Tian, Q.: Tape: Task-agnostic prior embedding for image restoration. In: ECCV. pp. 447-464. Springer (2022) +32. Liu, Y., He, J., Gu, J., Kong, X., Qiao, Y., Dong, C.: Degae: A new pretraining paradigm for low-level vision. In: CVPR. pp. 23292-23303 (2023) +33. Liu, Y., Liu, A., Gu, J., Zhang, Z., Wu, W., Qiao, Y., Dong, C.: Discovering distinctive "semantics" in super-resolution networks. arXiv preprint arXiv:2108.00406 (2021) +34. Liu, Z., Lin, Y., Cao, Y., Hu, H., Wei, Y., Zhang, Z., Lin, S., Guo, B.: Swin transformer: Hierarchical vision transformer using shifted windows. In: CVPR. pp. 10012-10022 (2021) +35. Luo, Y., Xu, Y., Ji, H.: Removing rain from a single image via discriminative sparse coding. In: ICCV. pp. 3397-3405 (2015) +36. Luo, Y., Zhao, R., Wei, X., Chen, J., Lu, Y., Xie, S., Wang, T., Xiong, R., Lu, M., Zhang, S.: Mowe: mixture of weather experts for multiple adverse weather removal. arXiv preprint arXiv:2303.13739 (2023) +37. Magid, S.A., Lin, Z., Wei, D., Zhang, Y., Gu, J., Pfister, H.: Texture-based error analysis for image super-resolution. In: CVPR. pp. 2118-2127 (2022) +38. Martin, D., Fowlkes, C., Tal, D., Malik, J.: A database of human segmented natural images and its application to evaluating segmentation algorithms and measuring ecological statistics. In: ICCV. vol. 2, pp. 416-423. IEEE (2001) +39. Mehri, A., Ardakani, P.B., Sappa, A.D.: Mprnet: Multi-path residual network for lightweight image super resolution. In: CVPR. pp. 2704-2713 (2021) +40. Nah, S., Hyun Kim, T., Mu Lee, K.: Deep multi-scale convolutional neural network for dynamic scene deblurring. In: CVPR +41. Park, D., Lee, B.H., Chun, S.Y.: All-in-one image restoration for unknown degradations using adaptive discriminative filters for specific degradations. In: CVPR. pp. 5815-5824 (2023) +42. Potlapalli, V., Zamir, S.W., Khan, S.H., Shahbaz Khan, F.: Prompt for all-in-one image restoration. NeurIPS 36 (2024) + +43. Radford, A., Narasimhan, K., Salimans, T., Sutskever, I., et al.: Improving language understanding by generative pre-training (2018) +44. Shrikumar, A., Su, J., Kundaje, A.: Computationally efficient measures of internal neuron importance. arXiv preprint arXiv:1807.09946 (2018) +45. Sundararajan, M., Taly, A., Yan, Q.: Axiomatic attribution for deep networks. In: ICML. pp. 3319-3328. PMLR (2017) +46. Sundararajan, M., Taly, A., Yan, Q.: Gradients of counterfactuals. ICLR (2017) +47. Wang, X., Wang, W., Cao, Y., Shen, C., Huang, T.: Images speak in images: A generalist painter for in-context visual learning. In: CVPR. pp. 6830-6839 (2023) +48. Wang, Z., Cun, X., Bao, J., Zhou, W., Liu, J., Li, H.: Uformer: A general u-shaped transformer for image restoration. In: CVPR. pp. 17683-17693 (2022) +49. Wu, R.Q., Duan, Z.P., Guo, C.L., Chai, Z., Li, C.: Ridcp: Revitalizing real image dehazing via high-quality codebook priors. In: CVPR. pp. 22282-22291 (2023) +50. Xie, L., Wang, X., Dong, C., Qi, Z., Shan, Y.: Finding discriminative filters for specific degradations in blind super-resolution. NeurIPS 34, 51-61 (2021) +51. Xie, Z., Zhang, Z., Cao, Y., Lin, Y., Bao, J., Yao, Z., Dai, Q., Hu, H.: Simmim: a simple framework for masked image modeling. In: CVPR. pp. 9653-9663 (2022) +52. Yang, W., Tan, R.T., Feng, J., Liu, J., Guo, Z., Yan, S.: Deep joint rain detection and removal from a single image. In: CVPR. pp. 1357-1366 (2017) +53. Yang, W., Tan, R.T., Wang, S., Fang, Y., Liu, J.: Single image deraining: From model-based to data-driven and beyond. PAMI 43(11), 4059-4077 (2020) +54. Yang, W., Wang, W., Huang, H., Wang, S., Liu, J.: Sparse gradient regularized deep retina network for robust low-light image enhancement. TIP 30, 2072-2086 (2021) +55. Zamir, S.W., Arora, A., Khan, S., Hayat, M., Khan, F.S., Yang, M.H.: Restormer: Efficient transformer for high-resolution image restoration. In: CVPR. pp. 5728-5739 (2022) +56. Zhang, C., Zhu, Y., Yan, Q., Sun, J., Zhang, Y.: All-in-one multi-degradation image restoration network via hierarchical degradation representation. In: ACMMM. pp. 2285-2293 (2023) +57. Zhang, H., Patel, V.M.: Density-aware single image de-raining using a multi-stream dense network. In: CVPR. pp. 695-704 (2018) +58. Zhang, H., Sindagi, V., Patel, V.M.: Image de-raining using a conditional generative adversarial network. TCSVT 30(11), 3943-3956 (2019) +59. Zhang, J., Huang, J., Yao, M., Yang, Z., Yu, H., Zhou, M., Zhao, F.: Ingredient-oriented multi-degradation learning for image restoration. In: CVPR. pp. 5825-5835 (2023) +60. Zheng, N., Zhou, M., Dong, Y., Rui, X., Huang, J., Li, C., Zhao, F.: Empowering low-light image enhancer through customized learnable priors. In: ICCV. pp. 12559-12569 (2023) +61. Zhu, Y., Wang, T., Fu, X., Yang, X., Guo, X., Dai, J., Qiao, Y., Hu, X.: Learning weather-general and weather-specific features for image restoration under multiple adverse weather conditions. In: CVPR. pp. 21747-21758 (2023) \ No newline at end of file diff --git a/2024/Restore Anything with Masks_ Leveraging Mask Image Modeling for Blind All-in-One Image Restoration/images.zip b/2024/Restore Anything with Masks_ Leveraging Mask Image Modeling for Blind All-in-One Image Restoration/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..12a1ef1217d420e60b344f759ccf084018879164 --- /dev/null +++ b/2024/Restore Anything with Masks_ Leveraging Mask Image Modeling for Blind All-in-One Image Restoration/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0078949959f109b5e1235a5405e2ec62b26f5304900bc586dcbb443abb4aa11f +size 811892 diff --git a/2024/Restore Anything with Masks_ Leveraging Mask Image Modeling for Blind All-in-One Image Restoration/layout.json b/2024/Restore Anything with Masks_ Leveraging Mask Image Modeling for Blind All-in-One Image Restoration/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..90085a4642ef029aecf29a1c00a7ae6409702e08 --- /dev/null +++ b/2024/Restore Anything with Masks_ Leveraging Mask Image Modeling for Blind All-in-One Image Restoration/layout.json @@ -0,0 +1,13422 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 139, + 111, + 477, + 163 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 111, + 477, + 163 + ], + "spans": [ + { + "bbox": [ + 139, + 111, + 477, + 163 + ], + "type": "text", + "content": "Restore Anything with Masks: Leveraging Mask Image Modeling for Blind All-in-One Image Restoration" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 186, + 185, + 426, + 210 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 186, + 185, + 426, + 210 + ], + "spans": [ + { + "bbox": [ + 186, + 185, + 426, + 210 + ], + "type": "text", + "content": "Chu-Jie Qin" + }, + { + "bbox": [ + 186, + 185, + 426, + 210 + ], + "type": "inline_equation", + "content": "^{1,2*}" + }, + { + "bbox": [ + 186, + 185, + 426, + 210 + ], + "type": "text", + "content": ", Rui-Qi Wu" + }, + { + "bbox": [ + 186, + 185, + 426, + 210 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 186, + 185, + 426, + 210 + ], + "type": "text", + "content": ", Zukun Liu" + }, + { + "bbox": [ + 186, + 185, + 426, + 210 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 186, + 185, + 426, + 210 + ], + "type": "text", + "content": ", Xin Lin" + }, + { + "bbox": [ + 186, + 185, + 426, + 210 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 186, + 185, + 426, + 210 + ], + "type": "text", + "content": ", Chun-Le Guo" + }, + { + "bbox": [ + 186, + 185, + 426, + 210 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 186, + 185, + 426, + 210 + ], + "type": "text", + "content": ", Hyun Hee Park" + }, + { + "bbox": [ + 186, + 185, + 426, + 210 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 186, + 185, + 426, + 210 + ], + "type": "text", + "content": ", and Chongyi Li" + }, + { + "bbox": [ + 186, + 185, + 426, + 210 + ], + "type": "inline_equation", + "content": "^{1,2\\dagger}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 242, + 219, + 372, + 230 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 242, + 219, + 372, + 230 + ], + "spans": [ + { + "bbox": [ + 242, + 219, + 372, + 230 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 242, + 219, + 372, + 230 + ], + "type": "text", + "content": " VCIP, CS, Nankai University" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 247, + 230, + 365, + 241 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 247, + 230, + 365, + 241 + ], + "spans": [ + { + "bbox": [ + 247, + 230, + 365, + 241 + ], + "type": "text", + "content": "2 NKIARI, Shenzhen Futian" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 214, + 242, + 399, + 252 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 214, + 242, + 399, + 252 + ], + "spans": [ + { + "bbox": [ + 214, + 242, + 399, + 252 + ], + "type": "text", + "content": "{chujie.qin,wuruiqi}@mail.nankai.edu.cn" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 222, + 254, + 392, + 264 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 222, + 254, + 392, + 264 + ], + "spans": [ + { + "bbox": [ + 222, + 254, + 392, + 264 + ], + "type": "text", + "content": "{guochunle, lichongyi}@nankai.edu.cn" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 213, + 263, + 400, + 275 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 213, + 263, + 400, + 275 + ], + "spans": [ + { + "bbox": [ + 213, + 263, + 400, + 275 + ], + "type": "text", + "content": "3 Samsung Research, China, Beijing (SRC-B)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 164, + 275, + 449, + 297 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 164, + 275, + 449, + 297 + ], + "spans": [ + { + "bbox": [ + 164, + 275, + 449, + 297 + ], + "type": "text", + "content": "4 The Department of Camera Innovation Group, Samsung Electronics {zikun.liu, inextg.park}@samsung.com" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 263, + 297, + 350, + 308 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 263, + 297, + 350, + 308 + ], + "spans": [ + { + "bbox": [ + 263, + 297, + 350, + 308 + ], + "type": "text", + "content": "Sichuan University" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 257, + 309, + 357, + 318 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 257, + 309, + 357, + 318 + ], + "spans": [ + { + "bbox": [ + 257, + 309, + 357, + 318 + ], + "type": "text", + "content": "linxin@stu.scu.edu.cn" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 159, + 342, + 455, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 342, + 455, + 552 + ], + "spans": [ + { + "bbox": [ + 159, + 342, + 455, + 552 + ], + "type": "text", + "content": "Abstract. All-in-one image restoration aims to handle multiple degradation types using one model. This paper proposes a simple pipeline for all-in-one blind image restoration to Restore Anything with Masks (RAM). We focus on the image content by utilizing Mask Image Modeling to extract intrinsic image information rather than distinguishing degradation types like other methods. Our pipeline consists of two stages: masked image pre-training and fine-tuning with mask attribute conductance. We design a straightforward masking pre-training approach specifically tailored for all-in-one image restoration. This approach enhances networks to prioritize the extraction of image content priors from various degradations, resulting in a more balanced performance across different restoration tasks and achieving stronger overall results. To bridge the gap of input integrity while preserving learned image priors as much as possible, we selectively fine-tuned a small portion of the layers. Specifically, the importance of each layer is ranked by the proposed Mask Attribute Conductance (MAC), and the layers with higher contributions are selected for finetuning. Extensive experiments demonstrate that our method achieves state-of-the-art performance. Our code and model will be released at https://github.com/Dragonisss/RAM." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 160, + 560, + 441, + 572 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 560, + 441, + 572 + ], + "spans": [ + { + "bbox": [ + 160, + 560, + 441, + 572 + ], + "type": "text", + "content": "Keywords: Image Restoration " + }, + { + "bbox": [ + 160, + 560, + 441, + 572 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 560, + 441, + 572 + ], + "type": "text", + "content": " All-in-One " + }, + { + "bbox": [ + 160, + 560, + 441, + 572 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 560, + 441, + 572 + ], + "type": "text", + "content": " Mask Image Modeling" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 590, + 230, + 601 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 590, + 230, + 601 + ], + "spans": [ + { + "bbox": [ + 132, + 590, + 230, + 601 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 612, + 482, + 637 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 612, + 482, + 637 + ], + "spans": [ + { + "bbox": [ + 130, + 612, + 482, + 637 + ], + "type": "text", + "content": "Image restoration involves the restoration of low-quality images affected by various degradation, typically arising from adverse environmental conditions (e.g.," + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 145, + 642, + 445, + 654 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 642, + 445, + 654 + ], + "spans": [ + { + "bbox": [ + 145, + 642, + 445, + 654 + ], + "type": "text", + "content": "* A part of this work is done during Chu-Jie Qin's internship at Samsung." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 147, + 654, + 314, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 147, + 654, + 314, + 665 + ], + "spans": [ + { + "bbox": [ + 147, + 654, + 314, + 665 + ], + "type": "text", + "content": "†Chongyi Li is the corresponding author." + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 164 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 164 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 164 + ], + "type": "text", + "content": "rain, haze, low-light), hardware-related issues (e.g., noise and blur), and post-processing artifacts (e.g., JPEG compression). Image restoration serves not only to enhance the visual appeal of images but also contributes to practical application scenarios such as autonomous driving and surveillance." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 167, + 312, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 167, + 312, + 335 + ], + "spans": [ + { + "bbox": [ + 130, + 167, + 312, + 335 + ], + "type": "text", + "content": "Modern techniques in this field mainly focus on learning fixed patterns formed during the degradation process, i.e., degradation priors. Some works [29, 30, 60] utilize task-specific priors to solve a certain degradation problem, while another research line [3, 28, 39, 48, 55] tries to design a general network architecture that can effectively learn each degradation pattern. Nevertheless, the above methods only enable the network to learn a single degradation, resulting in an imbalanced situation when dealing with multiple types of degradation." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 338, + 313, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 338, + 313, + 434 + ], + "spans": [ + { + "bbox": [ + 130, + 338, + 313, + 434 + ], + "type": "text", + "content": "To tackle the problem stated above, all-in-one methods have emerged, aiming to handle multiple degradations using one model. Most of these approaches tend to utilize explicit priors (e.g., AirNet [21]) or introduce an extra module (e.g., PromptIR [42]) to discern image degradation patterns, thereby assisting" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 319, + 191, + 482, + 342 + ], + "blocks": [ + { + "bbox": [ + 319, + 191, + 482, + 342 + ], + "lines": [ + { + "bbox": [ + 319, + 191, + 482, + 342 + ], + "spans": [ + { + "bbox": [ + 319, + 191, + 482, + 342 + ], + "type": "image", + "image_path": "06ec5f2f8dd06754772466ca6479e1a8f2b71672aa029fa0b4c8a52e05e7c4ca.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 318, + 350, + 482, + 406 + ], + "lines": [ + { + "bbox": [ + 318, + 350, + 482, + 406 + ], + "spans": [ + { + "bbox": [ + 318, + 350, + 482, + 406 + ], + "type": "text", + "content": "Fig.1: Our RAM achieves more balanced and more powerful performance than the state-of-the-art methods (AirNet [21], TAPE [31], PromptIR [42]) for all-in-one blind image restoration." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 434, + 482, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 434, + 482, + 555 + ], + "spans": [ + { + "bbox": [ + 130, + 434, + 482, + 555 + ], + "type": "text", + "content": "the model in performing the restoration. However, these methods place their emphasis on distinguishing degradation types in images rather than the image content, leading to lower scalability and fuzzy decision boundaries when more degradation types are involved. We argue that the essence of image restoration is to extract intrinsic image information from corrupted images rather than eliminate degradation patterns, i.e., learning image prior rather than degradation prior. It is worth noting that TAPE [31] similarly suggests that understanding normal image nature aids restoration by introducing a natural image prior. Nevertheless, TAPE utilizes the model output as the optimization target, which causes the model to amplify its own errors and learn the image prior with bias." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "text", + "content": "In this paper, we focus on tackling how to extract intrinsic image information from diverse corrupted images. Some attempts [2, 7] by Mask Image Modeling (MIM) in low-level vision have caught our attention. As a pretraining strategy, MIM has been widely validated for its effectiveness in high-level tasks, thanks to its generic representation of images. Simultaneously, the model also learns the distribution of natural images, which encompasses the intrinsic information we aim to extract from the images. Built on MIM, we propose a simple pipeline for all-in-one blind image restoration that Restores Anything with Masks (RAM), which includes two stages: the mask pre-training stage and the" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 206, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 206, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 206, + 102 + ], + "type": "text", + "content": "Qin et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 200 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 200 + ], + "type": "text", + "content": "fine-tuning stage with Mask Attribute Conductance (MAC). In the pre-training stage, we randomly mask corrupted images at the pixel-wise level and force the network to predict the clear one corresponding to the masked pixels, extracting inherent image information from corrupted images. In the fine-tuning stage, we focus on overcoming the input integrity gap caused by changing masked input during pre-training into the whole image during inference while preserving learned prior as much as possible." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 200, + 480, + 295 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 200, + 480, + 295 + ], + "spans": [ + { + "bbox": [ + 130, + 200, + 480, + 295 + ], + "type": "text", + "content": "Specifically, we first evaluated the importance of each network layer in addressing this gap by the proposed MAC. Following that, we chose the top " + }, + { + "bbox": [ + 130, + 200, + 480, + 295 + ], + "type": "inline_equation", + "content": "k\\%" + }, + { + "bbox": [ + 130, + 200, + 480, + 295 + ], + "type": "text", + "content": " most critical layers for fine-tuning while keeping the rest of the network layers frozen. We demonstrate that after a brief fine-tuning period (even if only " + }, + { + "bbox": [ + 130, + 200, + 480, + 295 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 130, + 200, + 480, + 295 + ], + "type": "text", + "content": " layers are tuned), the model can achieve a highly satisfactory performance level, surpassing models trained using traditional pair-wise training. Additionally, our pipeline can be plug-and-play used in any network without introducing additional computational overhead." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 146, + 296, + 347, + 307 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 296, + 347, + 307 + ], + "spans": [ + { + "bbox": [ + 146, + 296, + 347, + 307 + ], + "type": "text", + "content": "The contributions of this work are as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 315, + 480, + 471 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 138, + 315, + 479, + 362 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 315, + 479, + 362 + ], + "spans": [ + { + "bbox": [ + 138, + 315, + 479, + 362 + ], + "type": "text", + "content": "- We discuss the challenge of adopting MIM in low-level vision and propose a MIM-based pre-training strategy tailored to all-in-one blind image restoration, which allows the restoration networks to effectively learn inherent image information while guaranteeing reconstruction results." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 363, + 480, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 363, + 480, + 411 + ], + "spans": [ + { + "bbox": [ + 138, + 363, + 480, + 411 + ], + "type": "text", + "content": "- We proposed Mask Attribute Conductance to evaluate the importance of each layer in addressing the input integrity gap so that a very small portion (e.g. " + }, + { + "bbox": [ + 138, + 363, + 480, + 411 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 138, + 363, + 480, + 411 + ], + "type": "text", + "content": ") of critical layers are tuned to bridge this gap while preserving the image prior learned by MIM." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 411, + 479, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 411, + 479, + 471 + ], + "spans": [ + { + "bbox": [ + 138, + 411, + 479, + 471 + ], + "type": "text", + "content": "- Our proposed RAM provides a fresh perspective to achieve more balanced and powerful all-in-one blind image restoration, which focuses on extracting inherent image information from corrupted images. Our pipeline can be applied to any image restoration network without introducing additional computational overhead." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 132, + 489, + 237, + 502 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 489, + 237, + 502 + ], + "spans": [ + { + "bbox": [ + 132, + 489, + 237, + 502 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 514, + 373, + 526 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 514, + 373, + 526 + ], + "spans": [ + { + "bbox": [ + 131, + 514, + 373, + 526 + ], + "type": "text", + "content": "2.1 Image Restoration for Multi Degradations" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": "While neural networks have demonstrated impressive performance in single degradation image restoration [9, 13, 14, 17, 22, 23, 29, 30, 49, 60], recent works have shifted their focus towards addressing the more challenging domain of multi-degradation image restoration. A group of methods [3, 28, 39, 48, 55] aims at designing a general architecture that can effectively learn each degradation pattern. SwinIR [28] employs a window attention mechanism to convert global attention into a localized approach, effectively reducing computational overhead. In addition, the U-shaped transformer-based methods [48, 55] are employed to extract multi-scale features and reduce computational overhead. However, these methods have to train individually on each restoration task. Several methods [1, 24] leverage multiple input and output heads to empower the network to restore" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 324, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 324, + 91, + 447, + 102 + ], + "type": "text", + "content": "Restore Anything with Masks" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 143, + 114, + 473, + 255 + ], + "blocks": [ + { + "bbox": [ + 143, + 114, + 473, + 255 + ], + "lines": [ + { + "bbox": [ + 143, + 114, + 473, + 255 + ], + "spans": [ + { + "bbox": [ + 143, + 114, + 473, + 255 + ], + "type": "image", + "image_path": "97b3ce8594e2da099216f661a7efbf5f91f85036c934e86b9a3d32f0dd9f02bb.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 262, + 482, + 352 + ], + "lines": [ + { + "bbox": [ + 130, + 262, + 482, + 352 + ], + "spans": [ + { + "bbox": [ + 130, + 262, + 482, + 352 + ], + "type": "text", + "content": "Fig. 2: The illumination of our overall pipeline. 1) Pre-training the model with mask image pre-training method tailored to low-level vision. We randomly mask degraded images at the pixel level with a " + }, + { + "bbox": [ + 130, + 262, + 482, + 352 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 130, + 262, + 482, + 352 + ], + "type": "text", + "content": " masking ratio and reconstruct the clean images. 2) The Fine-tuning stage is followed to overcome the input integrity gap caused by changing masked input during pre-training into the whole image during inference. We analyze the importance of each network layer for resolving the input integrity gap according to the proposed MAC and rank them in descending order. The top " + }, + { + "bbox": [ + 130, + 262, + 482, + 352 + ], + "type": "inline_equation", + "content": "k\\%" + }, + { + "bbox": [ + 130, + 262, + 482, + 352 + ], + "type": "text", + "content": " of network layers are selected for fine-tuning on the complete image." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 375, + 482, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 375, + 482, + 555 + ], + "spans": [ + { + "bbox": [ + 130, + 375, + 482, + 555 + ], + "type": "text", + "content": "various types of degraded images. Nonetheless, this kind of approach may lead to the diminished scalability of the model. Recently, several subsequent methods [4, 21, 36, 41, 42, 56, 61] have been proposed to employ a unified network to address multiple restoration issues. Most of these methods put emphasis on learning how to distinguish different types of degradations and restore corrupted images. Typically, AirNet [21] first proposed an all-in-one image restoration task. The method initially pretrains a degradation classifier based on contrastive learning and subsequently utilizes it to assist in all-in-one image restoration. PromptIR [42] has introduced a learnable prompt-based module. Instead of constraining the degradation category, it enables the model to autonomously learn features that are advantageous to its performance by using an adaptive prompt. Our RAM takes a fresh perspective that focuses on extracting common content information from corrupted images, without any extra design to distinguish degradations, which helps us achieve balance and powerful performance when more degradation types are taken into consideration." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 573, + 274, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 573, + 274, + 586 + ], + "spans": [ + { + "bbox": [ + 132, + 573, + 274, + 586 + ], + "type": "text", + "content": "2.2 Mask Image Modeling" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 593, + 482, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 593, + 482, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 593, + 482, + 667 + ], + "type": "text", + "content": "Inspired by Mask Language Modeling [18,43], Mask Image Modeling (MIM) [15, 51] is introduced as a pretraining approach to learn general representations in high-level vision. MAE [15] effectively utilizes MIM for predicting hidden tokens, demonstrating strong performance and generalization across various downstream tasks. SimMIM [51] proposed a general masked image modeling method based on Swin-ViT [34]. Painter [47] unifies multiple tasks under image-to-image trans" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 206, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 206, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 206, + 102 + ], + "type": "text", + "content": "Qin et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 285 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 285 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 285 + ], + "type": "text", + "content": "lation and leverages MIM pretraining. In recent years, there have been efforts to incorporate MIM into the realm of low-level vision to enhance model generalization. Among them, [2] and [7] are the most closely aligned with our focus. [2] employs the MIM model to enhance the model's generalization for denoising tasks but has not explored its potential in multi-task scenarios. [7] utilizes MIM for pre-training the model encoder to introduce generative prior and subsequently employs the decoder for restoration. However, it does not fully harness the potential of MIM. Our proposed RAM utilizes MIM to unify the optimization objective for various image restoration tasks into reconstructing intrinsic image information. This allows the network to learn restoration functions more balanced and effectively. Moreover, to preserve the image priors learned by MIM, we designed a fine-tuning strategy based on MAC analysis (in Sec. 3.3). This enables us to achieve comparable performance by fine-tuning only a small portion (e.g. " + }, + { + "bbox": [ + 130, + 116, + 482, + 285 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 130, + 116, + 482, + 285 + ], + "type": "text", + "content": " ) of layers, fully tapping into the potential of MIM." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 301, + 299, + 313 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 301, + 299, + 313 + ], + "spans": [ + { + "bbox": [ + 131, + 301, + 299, + 313 + ], + "type": "text", + "content": "2.3 Gradient-based Attribution" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 321, + 482, + 465 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 321, + 482, + 465 + ], + "spans": [ + { + "bbox": [ + 130, + 321, + 482, + 465 + ], + "type": "text", + "content": "Gradient-based attribution methods [6,12,44-46,50] are often used to clarify how hidden units (or inputs) impact the output of networks. One commonly used approach is Integrated Gradients (IG) [45,46], which accumulates gradients along a linear path from the baseline input to the target input in the pixel/feature space. After that, IntInf [19] and layer conductance [6] alter IG to attribute neuron importance along the same path. In our work, we expect to find the key layers that can effectively overcome the distribution shift between training data and inference data. We propose Mask Attribute Conductance (MAC) based on the layer conductance and accumulated MAC of each layer along the Mask Attribute Path (MAP). MAC can represent the layer's importance along the MAP. In this way, we can fine-tune the top " + }, + { + "bbox": [ + 130, + 321, + 482, + 465 + ], + "type": "inline_equation", + "content": "k\\%" + }, + { + "bbox": [ + 130, + 321, + 482, + 465 + ], + "type": "text", + "content": " critical layers of the pre-trained network, preserving to a great extent the image priors learned during pretraining." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 483, + 233, + 497 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 483, + 233, + 497 + ], + "spans": [ + { + "bbox": [ + 131, + 483, + 233, + 497 + ], + "type": "text", + "content": "3 Methodology" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 508, + 482, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 508, + 482, + 557 + ], + "spans": [ + { + "bbox": [ + 130, + 508, + 482, + 557 + ], + "type": "text", + "content": "In this section, we start with discussing the challenges of using MIM in low-level vision tasks (Sec. 3.1). Following that, we present our pipeline for all-in-one blind image restoration, which contains two parts: pre-training with MIM (Sec. 3.2) and fine-tuning with Mask Attribute Conductance (MAC) Analysis (Sec. 3.3)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 574, + 350, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 574, + 350, + 586 + ], + "spans": [ + { + "bbox": [ + 131, + 574, + 350, + 586 + ], + "type": "text", + "content": "3.1 Rethinking MIM in Low-Level Vision" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": "MIM is a process that randomly masks certain parts of an image and extracts features from the remaining visible parts to reconstruct the entire image. It allows models to acquire a generic representation of images and thus achieve good pre-training, which is verified in many high-level tasks [15, 51]. Moreover, the models also learn the distribution of natural images during the image reconstruction, i.e. MIM pre-training. This incidental acquisition of prior knowledge" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 324, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 324, + 91, + 447, + 102 + ], + "type": "text", + "content": "Restore Anything with Masks" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 114, + 242, + 169 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 242, + 169 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 242, + 169 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 242, + 169 + ], + "type": "image", + "image_path": "26468efb0aa41a9754bbd12b714a4d07f6aaca410620d49d634ae47b3f851a9e.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 160, + 171, + 212, + 180 + ], + "lines": [ + { + "bbox": [ + 160, + 171, + 212, + 180 + ], + "spans": [ + { + "bbox": [ + 160, + 171, + 212, + 180 + ], + "type": "text", + "content": "patch size=1" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 255, + 114, + 362, + 168 + ], + "blocks": [ + { + "bbox": [ + 255, + 114, + 362, + 168 + ], + "lines": [ + { + "bbox": [ + 255, + 114, + 362, + 168 + ], + "spans": [ + { + "bbox": [ + 255, + 114, + 362, + 168 + ], + "type": "image", + "image_path": "40daf3bdc76771c1f65564c01a0834e29a029947d8e28ccf094ae4bed932624a.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 279, + 171, + 333, + 180 + ], + "lines": [ + { + "bbox": [ + 279, + 171, + 333, + 180 + ], + "spans": [ + { + "bbox": [ + 279, + 171, + 333, + 180 + ], + "type": "text", + "content": "patch size=4" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 131, + 189, + 481, + 222 + ], + "lines": [ + { + "bbox": [ + 131, + 189, + 481, + 222 + ], + "spans": [ + { + "bbox": [ + 131, + 189, + 481, + 222 + ], + "type": "text", + "content": "Fig. 3: Mask Image Modeling reconstruction with different patch sizes. We pre-trained with different patch sizes and visualized the mask inputs (left), and the corresponding MIM reconstructions (right)." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 374, + 114, + 480, + 168 + ], + "blocks": [ + { + "bbox": [ + 374, + 114, + 480, + 168 + ], + "lines": [ + { + "bbox": [ + 374, + 114, + 480, + 168 + ], + "spans": [ + { + "bbox": [ + 374, + 114, + 480, + 168 + ], + "type": "image", + "image_path": "980d7a274f6a681b09fa405bb062263e0583c28bfc40c0b776bb8fa36ff9c0a1.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 399, + 171, + 452, + 180 + ], + "lines": [ + { + "bbox": [ + 399, + 171, + 452, + 180 + ], + "spans": [ + { + "bbox": [ + 399, + 171, + 452, + 180 + ], + "type": "text", + "content": "patch size=8" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 251, + 479, + 287 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 251, + 479, + 287 + ], + "spans": [ + { + "bbox": [ + 130, + 251, + 479, + 287 + ], + "type": "text", + "content": "is instrumental in tasks like image restoration. Despite these advantages, applying MIM in pretraining a model for low-level vision tasks is still under-explored, primarily due to the challenges that must be addressed in the process." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 289, + 480, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 289, + 480, + 384 + ], + "spans": [ + { + "bbox": [ + 130, + 289, + 480, + 384 + ], + "type": "text", + "content": "Firstly, the main purpose of vanilla MIM is not high-quality reconstruction but good feature extraction for high-level tasks. Therefore, it masks a wider range of images to gather semantic information but not pixel-level content, reflected in token-level masking and a high mask ratio. CSFormer [7] directly adopts this strategy on low-level vision pre-training. However, some studies verify that semantic information is not as important for image restoration as it is in pattern recognition tasks [33, 37]. Moreover, high-degree masking leads to producing detail-deficient results, as shown in Fig. 3, which is harmful to low-level tasks." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 386, + 481, + 494 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 386, + 481, + 494 + ], + "spans": [ + { + "bbox": [ + 130, + 386, + 481, + 494 + ], + "type": "text", + "content": "Secondly, the training objective of MIM is to reconstruct the masked input images, so it can only produce results with the same domain as the input image. However, we hope the model gains the ability to bridge low-quality domain to high-quality domain, i.e. recover clean content from degraded input. Therefore, it is necessary to introduce paired data when pre-training image restoration models by MIM (see the experiment in Sec. 4.3 for details). Chen et al. [2] demonstrate that pair-wise MIM training enhances the generalization performance over different types of noisy images. In this paper, we take a step forward to explore the effectiveness of MIM on multiple degradations with larger variance." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 131, + 522, + 275, + 534 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 522, + 275, + 534 + ], + "spans": [ + { + "bbox": [ + 131, + 522, + 275, + 534 + ], + "type": "text", + "content": "3.2 Pretraining with MIM" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 553, + 480, + 577 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 553, + 480, + 577 + ], + "spans": [ + { + "bbox": [ + 130, + 553, + 480, + 577 + ], + "type": "text", + "content": "Based on the above analysis, we design a MIM pre-training paradigm tailored for low-level vision." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 579, + 480, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 579, + 480, + 628 + ], + "spans": [ + { + "bbox": [ + 130, + 579, + 480, + 628 + ], + "type": "text", + "content": "Masking. During the pre-training stage, we randomly mask the pixels of degraded images (mask images in a " + }, + { + "bbox": [ + 130, + 579, + 480, + 628 + ], + "type": "inline_equation", + "content": "1 \\times 1" + }, + { + "bbox": [ + 130, + 579, + 480, + 628 + ], + "type": "text", + "content": " patch size) with a " + }, + { + "bbox": [ + 130, + 579, + 480, + 628 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 130, + 579, + 480, + 628 + ], + "type": "text", + "content": " mask ratio. We found that fine-grained masked patches and balanced mask ratio are beneficial to image restoration, which can be demonstrated in Sec. 4.3." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 629, + 480, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 480, + 665 + ], + "type": "text", + "content": "Besides, since our MIM pre-training has a similar target to subsequent low-level tasks, we do not need to change the decoder like MAE [15] does but just fine-tune it." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 206, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 206, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 206, + 102 + ], + "type": "text", + "content": "Qin et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 175, + 115, + 441, + 327 + ], + "blocks": [ + { + "bbox": [ + 175, + 115, + 441, + 327 + ], + "lines": [ + { + "bbox": [ + 175, + 115, + 441, + 327 + ], + "spans": [ + { + "bbox": [ + 175, + 115, + 441, + 327 + ], + "type": "image", + "image_path": "ddd442039b45f8792232e47e1eb3152d425b123907e380a8991c86ee504114ed.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 341, + 482, + 397 + ], + "lines": [ + { + "bbox": [ + 131, + 341, + 482, + 397 + ], + "spans": [ + { + "bbox": [ + 131, + 341, + 482, + 397 + ], + "type": "text", + "content": "Fig. 4: The effect of MIM reconstruction with different input integrity on kernel deblurring (orange border) and denoising (blue border). We also visualize the color distributions of reconstructions in various tasks above. It shows that the distribution of the reconstruction results obtained using the twin-masks method as input is closer to the real images (ground truth) compared to the results obtained using the whole input." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 420, + 480, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 420, + 480, + 445 + ], + "spans": [ + { + "bbox": [ + 131, + 420, + 480, + 445 + ], + "type": "text", + "content": "Reconstruction target. Following the Bert [18] and MAE [15], we choose L1 loss to supervise the masked part. The training objective can be written as:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 230, + 453, + 481, + 473 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 230, + 453, + 481, + 473 + ], + "spans": [ + { + "bbox": [ + 230, + 453, + 481, + 473 + ], + "type": "interline_equation", + "content": "\\underset {\\theta} {\\arg \\min } \\mathbb {E} [ | | \\tilde {\\mathcal {M}} (I - f (\\mathcal {M} (I _ {d}), \\theta)) | | ], \\tag {1}", + "image_path": "8a20dce4eeacfe0710c7d589f7b8ef14a9aba035a881a2e0255db78b5e97e173.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 482, + 482, + 520 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 482, + 482, + 520 + ], + "spans": [ + { + "bbox": [ + 131, + 482, + 482, + 520 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 482, + 482, + 520 + ], + "type": "inline_equation", + "content": "\\{I, I_d\\}" + }, + { + "bbox": [ + 131, + 482, + 482, + 520 + ], + "type": "text", + "content": " represents a pair of clean image and degraded image, " + }, + { + "bbox": [ + 131, + 482, + 482, + 520 + ], + "type": "inline_equation", + "content": "f(\\cdot, \\theta)" + }, + { + "bbox": [ + 131, + 482, + 482, + 520 + ], + "type": "text", + "content": " denotes a network with parameters " + }, + { + "bbox": [ + 131, + 482, + 482, + 520 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 131, + 482, + 482, + 520 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 131, + 482, + 482, + 520 + ], + "type": "inline_equation", + "content": "\\mathcal{M}(\\cdot)" + }, + { + "bbox": [ + 131, + 482, + 482, + 520 + ], + "type": "text", + "content": " is a random binary masking operation and " + }, + { + "bbox": [ + 131, + 482, + 482, + 520 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathcal{M}}(\\cdot) = 1 - \\mathcal{M}(\\cdot)" + }, + { + "bbox": [ + 131, + 482, + 482, + 520 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 536, + 440, + 550 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 536, + 440, + 550 + ], + "spans": [ + { + "bbox": [ + 131, + 536, + 440, + 550 + ], + "type": "text", + "content": "3.3 Finetuning with Mask Attribute Conductance Analysis" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "text", + "content": "Observation. During pre-training, the network learns rich content priors. However, the incompleteness of the masked input prevents the direct use of the pre-trained model for inference, as it would result in a distribution shift in the outputs. As shown in Fig. 4, We start by feeding the entire image into a pretrained model, leading to a color-distorted result. Next, we use a pair of complementary masks, referred to as twin-masks, to individually mask the image. Subsequently, we input both of these complementarily masked images into the network. By combining the pixel values predicted by each image, we generate a higher-quality image. This observation indicates that the hindrance to using" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 324, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 324, + 91, + 447, + 102 + ], + "type": "text", + "content": "Restore Anything with Masks" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 139 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 139 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 139 + ], + "type": "text", + "content": "mask pre-trained model directly for inference lies in input incompleteness rather than the model's inability to learn the restoration function." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 140, + 480, + 223 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 140, + 480, + 223 + ], + "spans": [ + { + "bbox": [ + 130, + 140, + 480, + 223 + ], + "type": "text", + "content": "Building upon this insight, we explore the possibility of minimizing the influence of disparities in data input formats via model fine-tuning. To maintain the learned priors, it is essential to retain pre-trained parameters as extensively as possible while employing the fewest but most effective layers for fine-tuning. To tackle this, we introduce the concept of mask attribution conductance, which quantifies the importance of each layer concerning the fine-tuning objective. We then identify the top-" + }, + { + "bbox": [ + 130, + 140, + 480, + 223 + ], + "type": "inline_equation", + "content": "k\\%" + }, + { + "bbox": [ + 130, + 140, + 480, + 223 + ], + "type": "text", + "content": " most critical layers for fine-tuning." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 224, + 487, + 295 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 224, + 487, + 295 + ], + "spans": [ + { + "bbox": [ + 130, + 224, + 487, + 295 + ], + "type": "text", + "content": "Preliminary. Before giving the definition of Mask Attribute Conductance (MAC), we briefly recall the definition of integrate gradient [45] (IG) and neuron conductance [6] (Cond). Considering a linear path " + }, + { + "bbox": [ + 130, + 224, + 487, + 295 + ], + "type": "inline_equation", + "content": "\\gamma (\\alpha) = x^{\\prime} + \\alpha (x - x^{\\prime})" + }, + { + "bbox": [ + 130, + 224, + 487, + 295 + ], + "type": "text", + "content": " from base input " + }, + { + "bbox": [ + 130, + 224, + 487, + 295 + ], + "type": "inline_equation", + "content": "x^{\\prime}" + }, + { + "bbox": [ + 130, + 224, + 487, + 295 + ], + "type": "text", + "content": " to target input " + }, + { + "bbox": [ + 130, + 224, + 487, + 295 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 130, + 224, + 487, + 295 + ], + "type": "text", + "content": ", we can attribute output change " + }, + { + "bbox": [ + 130, + 224, + 487, + 295 + ], + "type": "inline_equation", + "content": "F(x) - F(x^{\\prime})" + }, + { + "bbox": [ + 130, + 224, + 487, + 295 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 130, + 224, + 487, + 295 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 224, + 487, + 295 + ], + "type": "text", + "content": "-th dimension of input/feature " + }, + { + "bbox": [ + 130, + 224, + 487, + 295 + ], + "type": "inline_equation", + "content": "x_{i}" + }, + { + "bbox": [ + 130, + 224, + 487, + 295 + ], + "type": "text", + "content": " (e.g. a pixel) by calculating its integrate gradient, which formally as below:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 203, + 302, + 481, + 329 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 203, + 302, + 481, + 329 + ], + "spans": [ + { + "bbox": [ + 203, + 302, + 481, + 329 + ], + "type": "interline_equation", + "content": "\\operatorname {I G} _ {i} (x) := \\left(x _ {i} - x _ {i} ^ {\\prime}\\right) \\cdot \\int_ {0} ^ {1} \\frac {\\partial F \\left(x ^ {\\prime} + \\alpha \\left(x - x ^ {\\prime}\\right)\\right)}{\\partial x _ {i}} d \\alpha . \\tag {2}", + "image_path": "0bf4881868f5911d6c06be2de5c5a0313ef69307ece4be6c062eb1031bb8a45d.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 333, + 480, + 370 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 333, + 480, + 370 + ], + "spans": [ + { + "bbox": [ + 131, + 333, + 480, + 370 + ], + "type": "text", + "content": "We can also attribute output change to a specific neuron " + }, + { + "bbox": [ + 131, + 333, + 480, + 370 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 131, + 333, + 480, + 370 + ], + "type": "text", + "content": " by improving IG, which involves calculating the conductance. The conductance [6] of the hidden neuron " + }, + { + "bbox": [ + 131, + 333, + 480, + 370 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 131, + 333, + 480, + 370 + ], + "type": "text", + "content": " along the " + }, + { + "bbox": [ + 131, + 333, + 480, + 370 + ], + "type": "inline_equation", + "content": "\\gamma(\\alpha)" + }, + { + "bbox": [ + 131, + 333, + 480, + 370 + ], + "type": "text", + "content": " is:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 179, + 376, + 481, + 438 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 179, + 376, + 481, + 438 + ], + "spans": [ + { + "bbox": [ + 179, + 376, + 481, + 438 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\operatorname {C o n d} ^ {y} (x) := \\sum_ {i} \\left(x _ {i} - x _ {i} ^ {\\prime}\\right) \\cdot \\int_ {0} ^ {1} \\frac {\\partial F \\left(x ^ {\\prime} + \\alpha \\left(x - x ^ {\\prime}\\right)\\right)}{\\partial y} \\cdot \\frac {\\partial y}{\\partial x _ {i}} d \\alpha \\tag {3} \\\\ = \\sum_ {i} \\int_ {0} ^ {1} \\frac {\\partial F (\\gamma (\\alpha))}{\\partial y} \\cdot \\frac {\\partial y}{\\partial \\alpha} d \\alpha , \\\\ \\end{array}", + "image_path": "9fb7f831b3aea5fa4535031c75d7addd343480e8e8d6132e7e4245db1269a803.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 445, + 479, + 473 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 445, + 479, + 473 + ], + "spans": [ + { + "bbox": [ + 130, + 445, + 479, + 473 + ], + "type": "text", + "content": "Note that " + }, + { + "bbox": [ + 130, + 445, + 479, + 473 + ], + "type": "inline_equation", + "content": "(x_{i} - x_{i}^{\\prime}) = \\frac{\\partial(x^{\\prime} + \\alpha(x_{i} - x_{i}^{\\prime}))}{\\partial\\alpha}" + }, + { + "bbox": [ + 130, + 445, + 479, + 473 + ], + "type": "text", + "content": ". Certainly, we can broaden Eq. (3) to compute conductance when integrating along any given path " + }, + { + "bbox": [ + 130, + 445, + 479, + 473 + ], + "type": "inline_equation", + "content": "\\alpha :[s,t]\\to P" + }, + { + "bbox": [ + 130, + 445, + 479, + 473 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 203, + 479, + 480, + 507 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 203, + 479, + 480, + 507 + ], + "spans": [ + { + "bbox": [ + 203, + 479, + 480, + 507 + ], + "type": "interline_equation", + "content": "\\operatorname {G e n e r a l C o n d} ^ {y} (x) := \\sum_ {i} \\int_ {P} \\frac {\\partial F (X _ {i} (\\alpha))}{\\partial y} \\cdot \\frac {\\partial y}{\\partial \\alpha} d \\alpha , \\tag {4}", + "image_path": "53bf215b2548c15d31d4af8a48107784c7a00503f848f476858f69044af167ff.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 514, + 480, + 537 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 514, + 480, + 537 + ], + "spans": [ + { + "bbox": [ + 131, + 514, + 480, + 537 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 514, + 480, + 537 + ], + "type": "inline_equation", + "content": "X: R \\to R^n" + }, + { + "bbox": [ + 131, + 514, + 480, + 537 + ], + "type": "text", + "content": " is the function of the path from " + }, + { + "bbox": [ + 131, + 514, + 480, + 537 + ], + "type": "inline_equation", + "content": "x'" + }, + { + "bbox": [ + 131, + 514, + 480, + 537 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 131, + 514, + 480, + 537 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 131, + 514, + 480, + 537 + ], + "type": "text", + "content": ", which satisfies " + }, + { + "bbox": [ + 131, + 514, + 480, + 537 + ], + "type": "inline_equation", + "content": "X(s) = x'" + }, + { + "bbox": [ + 131, + 514, + 480, + 537 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 131, + 514, + 480, + 537 + ], + "type": "inline_equation", + "content": "X(t) = x" + }, + { + "bbox": [ + 131, + 514, + 480, + 537 + ], + "type": "text", + "content": ". [s,t] represent the domain of the path function " + }, + { + "bbox": [ + 131, + 514, + 480, + 537 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 131, + 514, + 480, + 537 + ], + "type": "text", + "content": ".\"" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 131, + 538, + 480, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 538, + 480, + 586 + ], + "spans": [ + { + "bbox": [ + 131, + 538, + 480, + 586 + ], + "type": "text", + "content": "Finetuning with MAC. To find effective layers to finetune, we propose Mask Attribute Conductance (MAC) to evaluate how effective each layer is in overcoming the gap of input integrity. Considering such a nonlinear path " + }, + { + "bbox": [ + 131, + 538, + 480, + 586 + ], + "type": "inline_equation", + "content": "\\alpha : [0,1] \\to P_m" + }, + { + "bbox": [ + 131, + 538, + 480, + 586 + ], + "type": "text", + "content": " from zero input " + }, + { + "bbox": [ + 131, + 538, + 480, + 586 + ], + "type": "inline_equation", + "content": "x'" + }, + { + "bbox": [ + 131, + 538, + 480, + 586 + ], + "type": "text", + "content": " to whole input " + }, + { + "bbox": [ + 131, + 538, + 480, + 586 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 131, + 538, + 480, + 586 + ], + "type": "text", + "content": ", which path function " + }, + { + "bbox": [ + 131, + 538, + 480, + 586 + ], + "type": "inline_equation", + "content": "X^m" + }, + { + "bbox": [ + 131, + 538, + 480, + 586 + ], + "type": "text", + "content": " satisfies:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 246, + 593, + 480, + 624 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 246, + 593, + 480, + 624 + ], + "spans": [ + { + "bbox": [ + 246, + 593, + 480, + 624 + ], + "type": "interline_equation", + "content": "X _ {i} ^ {m} (\\alpha ; \\alpha_ {i}) = \\left\\{ \\begin{array}{l l} x _ {i} ^ {\\prime}, & \\alpha < \\alpha_ {i} \\\\ x _ {i}, & e l s e \\end{array} , \\right. \\tag {5}", + "image_path": "131d6890ff599a42cfe577f0ca05f3a5a617447535e7a2ac386a227299996741.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 629, + 479, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 479, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 479, + 666 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 629, + 479, + 666 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 629, + 479, + 666 + ], + "type": "text", + "content": " refers to the index of pixels, " + }, + { + "bbox": [ + 130, + 629, + 479, + 666 + ], + "type": "inline_equation", + "content": "\\alpha_{i} \\in (0,1]" + }, + { + "bbox": [ + 130, + 629, + 479, + 666 + ], + "type": "text", + "content": " is a set of parameters that indicate when each pixel gets masked. We define this path as a Mask Attribute Path (MAP). Apparently, " + }, + { + "bbox": [ + 130, + 629, + 479, + 666 + ], + "type": "inline_equation", + "content": "X^{m}(0) = x^{\\prime}" + }, + { + "bbox": [ + 130, + 629, + 479, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 629, + 479, + 666 + ], + "type": "inline_equation", + "content": "X^{m}(1) = x" + }, + { + "bbox": [ + 130, + 629, + 479, + 666 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 206, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 206, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 206, + 102 + ], + "type": "text", + "content": "Qin et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 171, + 116, + 297, + 221 + ], + "blocks": [ + { + "bbox": [ + 171, + 116, + 297, + 221 + ], + "lines": [ + { + "bbox": [ + 171, + 116, + 297, + 221 + ], + "spans": [ + { + "bbox": [ + 171, + 116, + 297, + 221 + ], + "type": "image", + "image_path": "79c2fd74506cff6d5f6752e8da693a01e93ccc21182850c9ed51e06e5cd32ae9.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 231, + 224, + 242, + 234 + ], + "lines": [ + { + "bbox": [ + 231, + 224, + 242, + 234 + ], + "spans": [ + { + "bbox": [ + 231, + 224, + 242, + 234 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 313, + 116, + 442, + 221 + ], + "blocks": [ + { + "bbox": [ + 313, + 116, + 442, + 221 + ], + "lines": [ + { + "bbox": [ + 313, + 116, + 442, + 221 + ], + "spans": [ + { + "bbox": [ + 313, + 116, + 442, + 221 + ], + "type": "image", + "image_path": "1a9a4c98871017718d833e194f72449e2da08834b4440b4682743ad47424f0af.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 377, + 224, + 389, + 234 + ], + "lines": [ + { + "bbox": [ + 377, + 224, + 389, + 234 + ], + "spans": [ + { + "bbox": [ + 377, + 224, + 389, + 234 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 173, + 247, + 440, + 260 + ], + "lines": [ + { + "bbox": [ + 173, + 247, + 440, + 260 + ], + "spans": [ + { + "bbox": [ + 173, + 247, + 440, + 260 + ], + "type": "text", + "content": "Fig. 5: Illumination of (a) " + }, + { + "bbox": [ + 173, + 247, + 440, + 260 + ], + "type": "inline_equation", + "content": "X_{i}^{m}" + }, + { + "bbox": [ + 173, + 247, + 440, + 260 + ], + "type": "text", + "content": " in Eq. (5) and (b) " + }, + { + "bbox": [ + 173, + 247, + 440, + 260 + ], + "type": "inline_equation", + "content": "\\tilde{X}_i^m" + }, + { + "bbox": [ + 173, + 247, + 440, + 260 + ], + "type": "text", + "content": " in Eq. (6)." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 281, + 480, + 316 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 281, + 480, + 316 + ], + "spans": [ + { + "bbox": [ + 130, + 281, + 480, + 316 + ], + "type": "text", + "content": "However, " + }, + { + "bbox": [ + 130, + 281, + 480, + 316 + ], + "type": "inline_equation", + "content": "X^{m}" + }, + { + "bbox": [ + 130, + 281, + 480, + 316 + ], + "type": "text", + "content": " is not differentiable, making it an invalid attribute path function. To solve this problem, we use a group of sigmoid-like functions " + }, + { + "bbox": [ + 130, + 281, + 480, + 316 + ], + "type": "inline_equation", + "content": "\\tilde{X}^{m}" + }, + { + "bbox": [ + 130, + 281, + 480, + 316 + ], + "type": "text", + "content": " to approximate " + }, + { + "bbox": [ + 130, + 281, + 480, + 316 + ], + "type": "inline_equation", + "content": "X^{m}" + }, + { + "bbox": [ + 130, + 281, + 480, + 316 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 242, + 316, + 480, + 341 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 242, + 316, + 480, + 341 + ], + "spans": [ + { + "bbox": [ + 242, + 316, + 480, + 341 + ], + "type": "interline_equation", + "content": "\\tilde {X} _ {i} ^ {m} \\left(\\alpha ; \\alpha_ {i}\\right) = \\frac {\\left(x _ {i} ^ {\\prime} - x _ {i}\\right)}{1 + e ^ {- \\delta \\left(x _ {i} ^ {\\prime} - \\alpha_ {i}\\right)}}. \\tag {6}", + "image_path": "d0da95448ed167a0f66c006e7822bd39a323eec590555c306016caeefa46cb3d.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 345, + 479, + 381 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 345, + 479, + 381 + ], + "spans": [ + { + "bbox": [ + 130, + 345, + 479, + 381 + ], + "type": "text", + "content": "We can see that " + }, + { + "bbox": [ + 130, + 345, + 479, + 381 + ], + "type": "inline_equation", + "content": "\\tilde{X}^m" + }, + { + "bbox": [ + 130, + 345, + 479, + 381 + ], + "type": "text", + "content": " is very close to " + }, + { + "bbox": [ + 130, + 345, + 479, + 381 + ], + "type": "inline_equation", + "content": "X^m" + }, + { + "bbox": [ + 130, + 345, + 479, + 381 + ], + "type": "text", + "content": " when " + }, + { + "bbox": [ + 130, + 345, + 479, + 381 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 130, + 345, + 479, + 381 + ], + "type": "text", + "content": " is sufficiently large (as depicted in Fig. 5). And for each " + }, + { + "bbox": [ + 130, + 345, + 479, + 381 + ], + "type": "inline_equation", + "content": "\\tilde{X}_i^m" + }, + { + "bbox": [ + 130, + 345, + 479, + 381 + ], + "type": "text", + "content": ", it will change sharply from " + }, + { + "bbox": [ + 130, + 345, + 479, + 381 + ], + "type": "inline_equation", + "content": "x_i'" + }, + { + "bbox": [ + 130, + 345, + 479, + 381 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 130, + 345, + 479, + 381 + ], + "type": "inline_equation", + "content": "x_i" + }, + { + "bbox": [ + 130, + 345, + 479, + 381 + ], + "type": "text", + "content": " when " + }, + { + "bbox": [ + 130, + 345, + 479, + 381 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 130, + 345, + 479, + 381 + ], + "type": "text", + "content": " is in the neighborhood of " + }, + { + "bbox": [ + 130, + 345, + 479, + 381 + ], + "type": "inline_equation", + "content": "\\alpha_i" + }, + { + "bbox": [ + 130, + 345, + 479, + 381 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 146, + 382, + 365, + 393 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 382, + 365, + 393 + ], + "spans": [ + { + "bbox": [ + 146, + 382, + 365, + 393 + ], + "type": "text", + "content": "Here, we can give a definition of MAC as below:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 208, + 399, + 480, + 460 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 208, + 399, + 480, + 460 + ], + "spans": [ + { + "bbox": [ + 208, + 399, + 480, + 460 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathrm {M A C} ^ {y} (x) := \\sum_ {i} \\int_ {P _ {m}} \\frac {\\partial F (X _ {i} (\\alpha))}{\\partial y} \\cdot \\frac {\\partial y}{\\partial \\alpha} d \\alpha \\\\ \\approx \\sum_ {i} \\int_ {0} ^ {1} \\frac {\\partial F \\left(\\tilde {X} _ {i} ^ {m} (\\alpha ; \\alpha_ {i})\\right)}{\\partial y} \\cdot \\frac {\\partial y}{\\partial \\alpha} d \\alpha . \\tag {7} \\\\ \\end{array}", + "image_path": "4fcc3f2fc46ba07574557060bd1c302eafe04c04fc1f5bf04d8b894e58c0ead8.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 467, + 479, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 467, + 479, + 491 + ], + "spans": [ + { + "bbox": [ + 130, + 467, + 479, + 491 + ], + "type": "text", + "content": "In fact, a partial path is also available to attribute from a masked input " + }, + { + "bbox": [ + 130, + 467, + 479, + 491 + ], + "type": "inline_equation", + "content": "x_{m}" + }, + { + "bbox": [ + 130, + 467, + 479, + 491 + ], + "type": "text", + "content": " with any mask ratio " + }, + { + "bbox": [ + 130, + 467, + 479, + 491 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 130, + 467, + 479, + 491 + ], + "type": "text", + "content": " to whole input " + }, + { + "bbox": [ + 130, + 467, + 479, + 491 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 130, + 467, + 479, + 491 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 204, + 497, + 480, + 527 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 204, + 497, + 480, + 527 + ], + "spans": [ + { + "bbox": [ + 204, + 497, + 480, + 527 + ], + "type": "interline_equation", + "content": "\\mathrm {M A C} _ {r} ^ {y} (x) \\approx \\sum_ {i} \\int_ {1 - r} ^ {1} \\frac {\\partial F \\left(\\tilde {X} _ {i} ^ {m} (\\alpha ; \\alpha_ {i})\\right)}{\\partial y} \\cdot \\frac {\\partial y}{\\partial \\alpha} d \\alpha . \\tag {8}", + "image_path": "df4a955c522e98ba71ecbf100b0acb1b94cbeaee9fd9cdc148229496188f9f71.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 533, + 479, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 479, + 559 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 479, + 559 + ], + "type": "text", + "content": "In practice, we use N-steps discretization to approximate the integral form of Eq. (8), which follows [44]:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 200, + 564, + 480, + 624 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 200, + 564, + 480, + 624 + ], + "spans": [ + { + "bbox": [ + 200, + 564, + 480, + 624 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathrm {M A C} _ {r} ^ {y} (x) \\approx \\sum_ {i} \\sum_ {j = 1} ^ {N} \\frac {\\partial F \\left(\\tilde {X} _ {i} ^ {m} \\left(\\frac {j r}{N} ; \\alpha_ {i}\\right)\\right)}{\\partial y} \\tag {9} \\\\ \\cdot \\left(F _ {y} \\left(\\tilde {X} _ {i} ^ {m} \\left(\\frac {(j + 1) r}{N}\\right)\\right) - F _ {y} \\left(\\tilde {X} _ {i} ^ {m} \\left(\\frac {j r}{N}\\right)\\right)\\right). \\\\ \\end{array}", + "image_path": "e17bb032d66aab7116432cd1008ea7b5d2555d3cf8ea9220661e935147529172.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 629, + 479, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 479, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 479, + 665 + ], + "type": "text", + "content": "We compute the MAC of each layer of pre-trained networks, rank them in descending order based on their MAC values, and pick top-" + }, + { + "bbox": [ + 130, + 629, + 479, + 665 + ], + "type": "inline_equation", + "content": "k\\%" + }, + { + "bbox": [ + 130, + 629, + 479, + 665 + ], + "type": "text", + "content": " layers for finetuning. The networks are initialized by pre-trained weight and only top-" + }, + { + "bbox": [ + 130, + 629, + 479, + 665 + ], + "type": "inline_equation", + "content": "k\\%" + }, + { + "bbox": [ + 130, + 629, + 479, + 665 + ], + "type": "text", + "content": " layers" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 324, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 324, + 91, + 447, + 102 + ], + "type": "text", + "content": "Restore Anything with Masks" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 134, + 168, + 482, + 266 + ], + "blocks": [ + { + "bbox": [ + 130, + 114, + 482, + 159 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 482, + 159 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 482, + 159 + ], + "type": "text", + "content": "Table 1: Quantitative comparison on seven challenging image restoration tasks, including dehazing, deraining, denoising, motion deblurring, low-light image enhancement (LLIE), kernel deblurring, and JPEG artifact removal. **boldface** and **underlined** indicate the best and second-best results, respectively." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 134, + 168, + 482, + 266 + ], + "lines": [ + { + "bbox": [ + 134, + 168, + 482, + 266 + ], + "spans": [ + { + "bbox": [ + 134, + 168, + 482, + 266 + ], + "type": "table", + "html": "
MethodSOTS [20]Rain13k-Test [32]BSD68 [38]GoPro [40]LOL [5]LSDIR-Blur [26]LSDIR-Jpeg [26]Average
PSNR†/SSIM†PSNR†/SSIM†PSNR†/SSIM†PSNR†/SSIM†PSNR†/SSIM†PSNR†/SSIM†PSNR†/SSIM†PSNR†/SSIM†
Restomer [55]22.89/0.917227.05/0.846930.95/0.865727.46/0.849723.65/0.845819.60/0.365830.46/0.914126.01/0.8007
MPRNet [39]25.23/0.946325.36/0.806829.83/0.831725.90/0.794922.29/0.817025.68/0.828128.96/0.886526.18/0.8445
NAFNet [3]25.74/0.944524.65/0.787730.37/0.854025.53/0.790921.50/0.810429.08/0.913029.09/0.895526.57/0.8566
DL [8]21.16/0.904219.56/0.650816.15/0.586117.63/0.586219.26/0.777717.98/0.612119.55/0.696518.75/0.6877
TAPE [31]25.14/0.931923.66/0.781830.11/0.835425.97/0.796218.95/0.763224.26/0.765429.28/0.896525.34/0.8243
AirNet [21]21.66/0.836620.21/0.640227.99/0.725023.36/0.750316.65/0.670823.84/0.735824.36/0.802022.58/0.7372
SwinIR [28]27.29/0.962225.32/0.825830.65/0.854026.61/0.812518.66/0.804827.82/0.883930.13/0.907126.64/0.8643
RAM-SwinIR28.47/0.968926.31/0.848630.83/0.861126.89/0.820021.62/0.829126.66/0.851430.22/0.909627.28/0.8698
PromptIR [42]28.70/0.965927.46/0.858530.84/0.862527.71/0.856521.19/0.835631.01/0.938530.30/0.911728.17/0.8899
RAM-PromptIR29.64/0.969528.47/0.875130.86/0.862428.02/0.859224.46/0.858129.57/0.917930.33/0.911928.76/0.8935
", + "image_path": "b6ddd4448c4909a90ec6fa216be5f2f7b6dda5113a454a0087d47e7bc612b5b1.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 133, + 274, + 482, + 381 + ], + "blocks": [ + { + "bbox": [ + 133, + 274, + 482, + 381 + ], + "lines": [ + { + "bbox": [ + 133, + 274, + 482, + 381 + ], + "spans": [ + { + "bbox": [ + 133, + 274, + 482, + 381 + ], + "type": "image", + "image_path": "b91e724a60e506ed9c17e75dc3fa51915f8d0dbc0bc0b38ca7d21353b2431304.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 158, + 387, + 455, + 399 + ], + "lines": [ + { + "bbox": [ + 158, + 387, + 455, + 399 + ], + "spans": [ + { + "bbox": [ + 158, + 387, + 455, + 399 + ], + "type": "text", + "content": "Fig. 6: Dehaze visual comparison on SOTS dataset. Zoom in for details." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 422, + 479, + 447 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 422, + 479, + 447 + ], + "spans": [ + { + "bbox": [ + 130, + 422, + 479, + 447 + ], + "type": "text", + "content": "will be fine-tuned. More implementation details can be found in the supplementary material." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 464, + 224, + 478 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 464, + 224, + 478 + ], + "spans": [ + { + "bbox": [ + 131, + 464, + 224, + 478 + ], + "type": "text", + "content": "4 Experiment" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 490, + 269, + 502 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 490, + 269, + 502 + ], + "spans": [ + { + "bbox": [ + 131, + 490, + 269, + 502 + ], + "type": "text", + "content": "4.1 Experiments Settings" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 510, + 482, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 482, + 641 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 482, + 641 + ], + "type": "text", + "content": "Datasets and Metrics. We combine datasets from various restoration tasks to form the training set, following [59]. For high-cost tasks that degradations are difficult to synthesize, we leverage existing paired datasets, including RE-SIDE [20] for dehazing, Rain13k [10, 25, 27, 35, 53] for deraining, GoPro [40] for motion deblurring, and LOL-v2 [54] for low-light image enhancement (LLIE). For low-cost tasks that degradations are easy to synthesize (e.g. noise, kernel blur, and JPEG artifact), we generate corrupted images on the LSDIR dataset [26] during the training process, which involves generating Gaussian noise with random variation " + }, + { + "bbox": [ + 130, + 510, + 482, + 641 + ], + "type": "inline_equation", + "content": "\\sigma \\in (0, 50]" + }, + { + "bbox": [ + 130, + 510, + 482, + 641 + ], + "type": "text", + "content": ", creating gaussian blurred images with a blur kernel of size " + }, + { + "bbox": [ + 130, + 510, + 482, + 641 + ], + "type": "inline_equation", + "content": "k = 15" + }, + { + "bbox": [ + 130, + 510, + 482, + 641 + ], + "type": "text", + "content": " and random " + }, + { + "bbox": [ + 130, + 510, + 482, + 641 + ], + "type": "inline_equation", + "content": "\\sigma \\in [0.1, 3.1]" + }, + { + "bbox": [ + 130, + 510, + 482, + 641 + ], + "type": "text", + "content": ", and introducing JPEG artifacts with a random quality parameter " + }, + { + "bbox": [ + 130, + 510, + 482, + 641 + ], + "type": "inline_equation", + "content": "q \\in [20, 90]" + }, + { + "bbox": [ + 130, + 510, + 482, + 641 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "content": "For evaluation, we use SOTS-outdoor [20] for dehazing, Rain13k-Test (the combination of Rain100L [52], Rain100H [52], Test100 [58], Test1200 [57] and" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 206, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 206, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 206, + 102 + ], + "type": "text", + "content": "Qin et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 164, + 146, + 451, + 282 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 480, + 137 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 480, + 137 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 480, + 137 + ], + "type": "text", + "content": "Table 2: Quantitative Gaussian denoising results at different noise levels on BSD68 and Urban100 datasets in terms of PSNR." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 164, + 146, + 451, + 282 + ], + "lines": [ + { + "bbox": [ + 164, + 146, + 451, + 282 + ], + "spans": [ + { + "bbox": [ + 164, + 146, + 451, + 282 + ], + "type": "table", + "html": "
MethodBSD68 [38]Urban100 [16]
σ = 15σ = 25σ = 50Averageσ = 15σ = 25σ = 50Average
NAFNet [3]33.2230.5927.3030.3732.6730.2126.9729.92
MPRNet [39]32.7330.1126.6529.8332.0629.4625.7729.10
Restormer [55]33.7931.1727.9030.9533.8331.4027.9931.07
DL [8]16.0416.2016.1916.1519.1719.1118.4718.92
TAPE [31]33.1030.3726.8630.1132.5929.9326.1929.57
AirNet [21]31.6328.8323.5227.9929.7926.9021.3526.01
SwinIR [28]33.5330.8927.5430.6533.5030.9927.3730.62
RAM-SwinIR33.6531.0627.7730.8233.8231.4327.9431.07
performance gains(↑0.12)(↑0.17)(↑0.23)(↑0.17)(↑0.32)(↑0.44)(↑0.57)(↑0.45)
PromptIR [42]33.6731.0627.8030.8433.5631.0827.6430.76
RAM-PromptIR33.7031.0827.7930.8633.7031.3027.9230.97
performance gains(↑0.03)(↑0.02)(↓0.01)(↑0.02)(↑0.14)(↑0.22)(↑0.28)(↑0.21)
", + "image_path": "f262acc6170e25da2dfb873c137fa062b1c78c6821e77068e274d84a6d209a89.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 304, + 480, + 363 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 304, + 480, + 363 + ], + "spans": [ + { + "bbox": [ + 130, + 304, + 480, + 363 + ], + "type": "text", + "content": "Test2800 [11]) for deraining, GoPro for motion deblurring, LOL [5] for low-light enhancement, BSD68 [38] for denoising, LSDIR-val for kernel deblurring and JPEG artifact removal. Furthermore, We conducted evaluations including denoising tests with variances of 15, 25, and 50, deblurring tests at " + }, + { + "bbox": [ + 130, + 304, + 480, + 363 + ], + "type": "inline_equation", + "content": "k = 15" + }, + { + "bbox": [ + 130, + 304, + 480, + 363 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 304, + 480, + 363 + ], + "type": "inline_equation", + "content": "\\sigma = 2.0" + }, + { + "bbox": [ + 130, + 304, + 480, + 363 + ], + "type": "text", + "content": ", and JPEG artifact removal tests at " + }, + { + "bbox": [ + 130, + 304, + 480, + 363 + ], + "type": "inline_equation", + "content": "q = 50" + }, + { + "bbox": [ + 130, + 304, + 480, + 363 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 364, + 482, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 364, + 482, + 483 + ], + "spans": [ + { + "bbox": [ + 130, + 364, + 482, + 483 + ], + "type": "text", + "content": "Implementation Details. We apply our proposed RAM to SwinIR [28] and PromptIR [42]. The input size for RAM-SwinIR is 64, while for RAM-PromptIR it is 128. During the pre-training phase, we use the Adam optimizer to train RAM-SwinIR and RAM-PromptIR for 300 epochs, with the learning rate decaying from 1e-4 to 6e-5 following a cosine schedule. In the fine-tuning phase, we use the Adam optimizer to fine-tune the network layers obtained from the MAC analysis of RAM-SwinIR and RAM-PromptIR for 40 epochs, with the learning rate decaying from 2e-4 to 1e-7 following a cosine schedule. The batch sizes for RAM-SwinIR and RAM-PromptIR during the pre-training and fine-tuning phases are (12,4) and (4,4), respectively." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 502, + 227, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 502, + 227, + 514 + ], + "spans": [ + { + "bbox": [ + 132, + 502, + 227, + 514 + ], + "type": "text", + "content": "4.2 Comparisons" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 521, + 480, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 480, + 593 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 480, + 593 + ], + "type": "text", + "content": "To validate the gain capability and effectiveness of our RAM, we apply the proposed RAM to SwinIR (a general image restoration method) and PromptIR (an all-in-one image restoration method). Four general architecture-based image restoration methods [3, 28, 39, 55] and four all-in-one methods [8, 21, 31, 42] are considered for comparison. We ensure that the number of supervised pixels employed by all other methods equals that used during the pre-training stage." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 594, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 481, + 665 + ], + "type": "text", + "content": "As illustrated in Tab. 1, our approach achieves the best or comparable performance on each task. On the average score across seven different tasks, our method with PromptIR [42] achieves 0.59dB performance gains compared to the second-best algorithm. Besides, the SwinIR equipped with RAM also yields " + }, + { + "bbox": [ + 130, + 594, + 481, + 665 + ], + "type": "inline_equation", + "content": "2.40\\%" + }, + { + "bbox": [ + 130, + 594, + 481, + 665 + ], + "type": "text", + "content": " improvement on PSNR. Specifically, our RAM has significant benefits for dehazing and low-light enhancement. Tab. 2 shows the Quantitative denois-" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 324, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 324, + 91, + 447, + 102 + ], + "type": "text", + "content": "Restore Anything with Masks" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 114, + 191, + 152 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 191, + 152 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 191, + 152 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 191, + 152 + ], + "type": "image", + "image_path": "97fede3a9740dae1b8177c4426144258ecb5b722a4137b27f2e7016a98b6cb61.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 146, + 154, + 181, + 161 + ], + "lines": [ + { + "bbox": [ + 146, + 154, + 181, + 161 + ], + "spans": [ + { + "bbox": [ + 146, + 154, + 181, + 161 + ], + "type": "text", + "content": "Rainy Image" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 192, + 114, + 249, + 152 + ], + "blocks": [ + { + "bbox": [ + 192, + 114, + 249, + 152 + ], + "lines": [ + { + "bbox": [ + 192, + 114, + 249, + 152 + ], + "spans": [ + { + "bbox": [ + 192, + 114, + 249, + 152 + ], + "type": "image", + "image_path": "c5456c7b9deedb3b3cc5bfd0eed573d67e9dd2969b912482006b2c533f55e21a.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 206, + 154, + 230, + 161 + ], + "lines": [ + { + "bbox": [ + 206, + 154, + 230, + 161 + ], + "spans": [ + { + "bbox": [ + 206, + 154, + 230, + 161 + ], + "type": "text", + "content": "NAFNet" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 250, + 114, + 306, + 152 + ], + "blocks": [ + { + "bbox": [ + 250, + 114, + 306, + 152 + ], + "lines": [ + { + "bbox": [ + 250, + 114, + 306, + 152 + ], + "spans": [ + { + "bbox": [ + 250, + 114, + 306, + 152 + ], + "type": "image", + "image_path": "bcf69f06dddfb2989e89877bfb4726d905d3a5a1e14ccc67f8c7daba573e27f0.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 264, + 154, + 289, + 160 + ], + "lines": [ + { + "bbox": [ + 264, + 154, + 289, + 160 + ], + "spans": [ + { + "bbox": [ + 264, + 154, + 289, + 160 + ], + "type": "text", + "content": "MPRNet" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 307, + 114, + 364, + 152 + ], + "blocks": [ + { + "bbox": [ + 307, + 114, + 364, + 152 + ], + "lines": [ + { + "bbox": [ + 307, + 114, + 364, + 152 + ], + "spans": [ + { + "bbox": [ + 307, + 114, + 364, + 152 + ], + "type": "image", + "image_path": "9547db10f802091a735362c2eaeded7edeaabf60ca54ae16000a371da88600be.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 321, + 154, + 349, + 161 + ], + "lines": [ + { + "bbox": [ + 321, + 154, + 349, + 161 + ], + "spans": [ + { + "bbox": [ + 321, + 154, + 349, + 161 + ], + "type": "text", + "content": "Restomer" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 365, + 114, + 422, + 152 + ], + "blocks": [ + { + "bbox": [ + 365, + 114, + 422, + 152 + ], + "lines": [ + { + "bbox": [ + 365, + 114, + 422, + 152 + ], + "spans": [ + { + "bbox": [ + 365, + 114, + 422, + 152 + ], + "type": "image", + "image_path": "c7a068f252cd8e96241a0d432e7444aee4c0213bb00a7323a6c0f605bb396efa.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 381, + 154, + 402, + 160 + ], + "lines": [ + { + "bbox": [ + 381, + 154, + 402, + 160 + ], + "spans": [ + { + "bbox": [ + 381, + 154, + 402, + 160 + ], + "type": "text", + "content": "SwinIR" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 424, + 114, + 480, + 152 + ], + "blocks": [ + { + "bbox": [ + 424, + 114, + 480, + 152 + ], + "lines": [ + { + "bbox": [ + 424, + 114, + 480, + 152 + ], + "spans": [ + { + "bbox": [ + 424, + 114, + 480, + 152 + ], + "type": "image", + "image_path": "760c799c2591cd1c736f81b6798398168d10b5bda311a6c6e99b6745b1f0106f.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 423, + 154, + 478, + 161 + ], + "lines": [ + { + "bbox": [ + 423, + 154, + 478, + 161 + ], + "spans": [ + { + "bbox": [ + 423, + 154, + 478, + 161 + ], + "type": "text", + "content": "RAM-SwinIR (Ours)" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 133, + 163, + 191, + 201 + ], + "blocks": [ + { + "bbox": [ + 133, + 163, + 191, + 201 + ], + "lines": [ + { + "bbox": [ + 133, + 163, + 191, + 201 + ], + "spans": [ + { + "bbox": [ + 133, + 163, + 191, + 201 + ], + "type": "image", + "image_path": "a142d342183e62f3d0950629041f17e2d9b30fca3ba1a7e709c48f26c7297ae9.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 156, + 202, + 167, + 209 + ], + "lines": [ + { + "bbox": [ + 156, + 202, + 167, + 209 + ], + "spans": [ + { + "bbox": [ + 156, + 202, + 167, + 209 + ], + "type": "text", + "content": "DL" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 192, + 163, + 249, + 202 + ], + "blocks": [ + { + "bbox": [ + 192, + 163, + 249, + 202 + ], + "lines": [ + { + "bbox": [ + 192, + 163, + 249, + 202 + ], + "spans": [ + { + "bbox": [ + 192, + 163, + 249, + 202 + ], + "type": "image", + "image_path": "cf0b4daf1123d33c2a218cc2ef17368bd3d5b7debf7f945214e537348290a719.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 211, + 202, + 227, + 209 + ], + "lines": [ + { + "bbox": [ + 211, + 202, + 227, + 209 + ], + "spans": [ + { + "bbox": [ + 211, + 202, + 227, + 209 + ], + "type": "text", + "content": "TAPE" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 250, + 163, + 306, + 202 + ], + "blocks": [ + { + "bbox": [ + 250, + 163, + 306, + 202 + ], + "lines": [ + { + "bbox": [ + 250, + 163, + 306, + 202 + ], + "spans": [ + { + "bbox": [ + 250, + 163, + 306, + 202 + ], + "type": "image", + "image_path": "db3404ecfb3504bd5e7963af1234c9778334eb144ae16d2a97c982209a77c1c6.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 268, + 202, + 287, + 209 + ], + "lines": [ + { + "bbox": [ + 268, + 202, + 287, + 209 + ], + "spans": [ + { + "bbox": [ + 268, + 202, + 287, + 209 + ], + "type": "text", + "content": "AirNet" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 307, + 163, + 364, + 202 + ], + "blocks": [ + { + "bbox": [ + 307, + 163, + 364, + 202 + ], + "lines": [ + { + "bbox": [ + 307, + 163, + 364, + 202 + ], + "spans": [ + { + "bbox": [ + 307, + 163, + 364, + 202 + ], + "type": "image", + "image_path": "4f5f26f456fe7781fd76bb1d3671a4605f767bc764e50ff705eb4ddedade53c0.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 321, + 202, + 347, + 209 + ], + "lines": [ + { + "bbox": [ + 321, + 202, + 347, + 209 + ], + "spans": [ + { + "bbox": [ + 321, + 202, + 347, + 209 + ], + "type": "text", + "content": "PromptIR" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 364, + 202, + 423, + 209 + ], + "lines": [ + { + "bbox": [ + 364, + 202, + 423, + 209 + ], + "spans": [ + { + "bbox": [ + 364, + 202, + 423, + 209 + ], + "type": "text", + "content": "RAM-PromptIR(Ours)" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 365, + 163, + 422, + 202 + ], + "blocks": [ + { + "bbox": [ + 365, + 163, + 422, + 202 + ], + "lines": [ + { + "bbox": [ + 365, + 163, + 422, + 202 + ], + "spans": [ + { + "bbox": [ + 365, + 163, + 422, + 202 + ], + "type": "image", + "image_path": "d13a3871a0aba2898f675c9e3000a4fd6e484e63bba71ac92ccd600a68fb1f1d.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 424, + 163, + 480, + 202 + ], + "blocks": [ + { + "bbox": [ + 424, + 163, + 480, + 202 + ], + "lines": [ + { + "bbox": [ + 424, + 163, + 480, + 202 + ], + "spans": [ + { + "bbox": [ + 424, + 163, + 480, + 202 + ], + "type": "image", + "image_path": "b1d8b0a8235a3f965c0d9f7781132d8359ca0c48dd7a6d47b378f3875aadf756.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 446, + 202, + 455, + 209 + ], + "lines": [ + { + "bbox": [ + 446, + 202, + 455, + 209 + ], + "spans": [ + { + "bbox": [ + 446, + 202, + 455, + 209 + ], + "type": "text", + "content": "GT" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 133, + 240, + 191, + 274 + ], + "blocks": [ + { + "bbox": [ + 144, + 218, + 467, + 229 + ], + "lines": [ + { + "bbox": [ + 144, + 218, + 467, + 229 + ], + "spans": [ + { + "bbox": [ + 144, + 218, + 467, + 229 + ], + "type": "text", + "content": "Fig. 7: Derain visual comparison on Rain13k-Test dataset. Zoom in for details." + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 133, + 240, + 191, + 274 + ], + "lines": [ + { + "bbox": [ + 133, + 240, + 191, + 274 + ], + "spans": [ + { + "bbox": [ + 133, + 240, + 191, + 274 + ], + "type": "image", + "image_path": "d392f4b13367f37112eb00350263198a734a9aa54c82abbe7ec6a13a5feee113.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 146, + 275, + 182, + 282 + ], + "lines": [ + { + "bbox": [ + 146, + 275, + 182, + 282 + ], + "spans": [ + { + "bbox": [ + 146, + 275, + 182, + 282 + ], + "type": "text", + "content": "Blurred Image" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_caption" + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 192, + 240, + 249, + 274 + ], + "blocks": [ + { + "bbox": [ + 192, + 240, + 249, + 274 + ], + "lines": [ + { + "bbox": [ + 192, + 240, + 249, + 274 + ], + "spans": [ + { + "bbox": [ + 192, + 240, + 249, + 274 + ], + "type": "image", + "image_path": "0da33db259b06b5e0ed7743e91b632843ea4fa696c8afa3a6b133b7f105fed17.jpg" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 210, + 275, + 231, + 281 + ], + "lines": [ + { + "bbox": [ + 210, + 275, + 231, + 281 + ], + "spans": [ + { + "bbox": [ + 210, + 275, + 231, + 281 + ], + "type": "text", + "content": "NAFNet" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_caption" + } + ], + "index": 29 + }, + { + "type": "image", + "bbox": [ + 250, + 240, + 306, + 274 + ], + "blocks": [ + { + "bbox": [ + 250, + 240, + 306, + 274 + ], + "lines": [ + { + "bbox": [ + 250, + 240, + 306, + 274 + ], + "spans": [ + { + "bbox": [ + 250, + 240, + 306, + 274 + ], + "type": "image", + "image_path": "09842421063513a58bdd468f900e69b29fe537db7d36982428dd4bf5d1167b08.jpg" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 268, + 275, + 290, + 281 + ], + "lines": [ + { + "bbox": [ + 268, + 275, + 290, + 281 + ], + "spans": [ + { + "bbox": [ + 268, + 275, + 290, + 281 + ], + "type": "text", + "content": "MPRNet" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_caption" + } + ], + "index": 31 + }, + { + "type": "image", + "bbox": [ + 307, + 240, + 425, + 274 + ], + "blocks": [ + { + "bbox": [ + 307, + 240, + 425, + 274 + ], + "lines": [ + { + "bbox": [ + 307, + 240, + 425, + 274 + ], + "spans": [ + { + "bbox": [ + 307, + 240, + 425, + 274 + ], + "type": "image", + "image_path": "62f4c2ecf477dbd9baa40f009323f3128f9e24af0734db30bde5edac1e3564df.jpg" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 321, + 275, + 348, + 281 + ], + "lines": [ + { + "bbox": [ + 321, + 275, + 348, + 281 + ], + "spans": [ + { + "bbox": [ + 321, + 275, + 348, + 281 + ], + "type": "text", + "content": "Restormer" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_caption" + } + ], + "index": 33 + }, + { + "type": "image", + "bbox": [ + 425, + 240, + 481, + 274 + ], + "blocks": [ + { + "bbox": [ + 425, + 240, + 481, + 274 + ], + "lines": [ + { + "bbox": [ + 425, + 240, + 481, + 274 + ], + "spans": [ + { + "bbox": [ + 425, + 240, + 481, + 274 + ], + "type": "image", + "image_path": "4a78e83d2105f41649b4693be56c563f86e2ff4dbaedeae4f3c2154ed8465f35.jpg" + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 426, + 275, + 479, + 281 + ], + "lines": [ + { + "bbox": [ + 426, + 275, + 479, + 281 + ], + "spans": [ + { + "bbox": [ + 426, + 275, + 479, + 281 + ], + "type": "text", + "content": "RAM-SwinIR (Ours)" + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_caption" + } + ], + "index": 35 + }, + { + "type": "image", + "bbox": [ + 133, + 283, + 191, + 317 + ], + "blocks": [ + { + "bbox": [ + 133, + 283, + 191, + 317 + ], + "lines": [ + { + "bbox": [ + 133, + 283, + 191, + 317 + ], + "spans": [ + { + "bbox": [ + 133, + 283, + 191, + 317 + ], + "type": "image", + "image_path": "f3dedb93749e3aeb4a982bfeb89804f435eb1d9cbbf021d349fdc59878d82874.jpg" + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 155, + 318, + 164, + 324 + ], + "lines": [ + { + "bbox": [ + 155, + 318, + 164, + 324 + ], + "spans": [ + { + "bbox": [ + 155, + 318, + 164, + 324 + ], + "type": "text", + "content": "DL" + } + ] + } + ], + "index": 38, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 143, + 338, + 469, + 349 + ], + "lines": [ + { + "bbox": [ + 143, + 338, + 469, + 349 + ], + "spans": [ + { + "bbox": [ + 143, + 338, + 469, + 349 + ], + "type": "text", + "content": "Fig. 8: Motion deblur visual comparison on GoPro dataset. Zoom in for details." + } + ] + } + ], + "index": 49, + "angle": 0, + "type": "image_caption" + } + ], + "index": 37 + }, + { + "type": "image", + "bbox": [ + 192, + 283, + 249, + 317 + ], + "blocks": [ + { + "bbox": [ + 192, + 283, + 249, + 317 + ], + "lines": [ + { + "bbox": [ + 192, + 283, + 249, + 317 + ], + "spans": [ + { + "bbox": [ + 192, + 283, + 249, + 317 + ], + "type": "image", + "image_path": "f76c879fde6b574c1a8d76f24b8eb05e5854c9141c2a4a8483b5aab6e769707a.jpg" + } + ] + } + ], + "index": 39, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 212, + 318, + 228, + 324 + ], + "lines": [ + { + "bbox": [ + 212, + 318, + 228, + 324 + ], + "spans": [ + { + "bbox": [ + 212, + 318, + 228, + 324 + ], + "type": "text", + "content": "TAPE" + } + ] + } + ], + "index": 40, + "angle": 0, + "type": "image_caption" + } + ], + "index": 39 + }, + { + "type": "image", + "bbox": [ + 250, + 283, + 306, + 317 + ], + "blocks": [ + { + "bbox": [ + 250, + 283, + 306, + 317 + ], + "lines": [ + { + "bbox": [ + 250, + 283, + 306, + 317 + ], + "spans": [ + { + "bbox": [ + 250, + 283, + 306, + 317 + ], + "type": "image", + "image_path": "84d6d2b3e3679eca20a4fa4d761a3ee97088fbde5a6ec0d9fbc09810410e2701.jpg" + } + ] + } + ], + "index": 41, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 268, + 318, + 285, + 324 + ], + "lines": [ + { + "bbox": [ + 268, + 318, + 285, + 324 + ], + "spans": [ + { + "bbox": [ + 268, + 318, + 285, + 324 + ], + "type": "text", + "content": "AirNet" + } + ] + } + ], + "index": 42, + "angle": 0, + "type": "image_caption" + } + ], + "index": 41 + }, + { + "type": "image", + "bbox": [ + 307, + 283, + 425, + 317 + ], + "blocks": [ + { + "bbox": [ + 307, + 283, + 425, + 317 + ], + "lines": [ + { + "bbox": [ + 307, + 283, + 425, + 317 + ], + "spans": [ + { + "bbox": [ + 307, + 283, + 425, + 317 + ], + "type": "image", + "image_path": "7ce785fe1109627eda1e07fefe0c38db60de1f2ecfaa9fc617d95a4b3e3904a5.jpg" + } + ] + } + ], + "index": 43, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 320, + 318, + 346, + 324 + ], + "lines": [ + { + "bbox": [ + 320, + 318, + 346, + 324 + ], + "spans": [ + { + "bbox": [ + 320, + 318, + 346, + 324 + ], + "type": "text", + "content": "PromptIR" + } + ] + } + ], + "index": 44, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 364, + 318, + 421, + 325 + ], + "lines": [ + { + "bbox": [ + 364, + 318, + 421, + 325 + ], + "spans": [ + { + "bbox": [ + 364, + 318, + 421, + 325 + ], + "type": "text", + "content": "RAM-PromptIR(Ours)" + } + ] + } + ], + "index": 46, + "angle": 0, + "type": "image_caption" + } + ], + "index": 43 + }, + { + "bbox": [ + 365, + 283, + 422, + 317 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 365, + 283, + 422, + 317 + ], + "spans": [ + { + "bbox": [ + 365, + 283, + 422, + 317 + ], + "type": "image", + "image_path": "72f6eeecb344a1a2679bc7628d07b78b1b8388e5e2b32a968e0dced152db581c.jpg" + } + ] + } + ], + "index": 45, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 425, + 283, + 481, + 317 + ], + "blocks": [ + { + "bbox": [ + 425, + 283, + 481, + 317 + ], + "lines": [ + { + "bbox": [ + 425, + 283, + 481, + 317 + ], + "spans": [ + { + "bbox": [ + 425, + 283, + 481, + 317 + ], + "type": "image", + "image_path": "8a8e13692bf76d3b1d7c71dc24f60cd7a08a179450d5da0e7934fd7c5f1fd6e3.jpg" + } + ] + } + ], + "index": 47, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 444, + 318, + 453, + 324 + ], + "lines": [ + { + "bbox": [ + 444, + 318, + 453, + 324 + ], + "spans": [ + { + "bbox": [ + 444, + 318, + 453, + 324 + ], + "type": "text", + "content": "GT" + } + ] + } + ], + "index": 48, + "angle": 0, + "type": "image_caption" + } + ], + "index": 47 + }, + { + "bbox": [ + 130, + 373, + 480, + 397 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 373, + 480, + 397 + ], + "spans": [ + { + "bbox": [ + 130, + 373, + 480, + 397 + ], + "type": "text", + "content": "ing result at different noise levels. Both RAM-SwinIR and RAM-PromptIR get higher performance than the origin versions." + } + ] + } + ], + "index": 50 + }, + { + "bbox": [ + 130, + 397, + 481, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 397, + 481, + 506 + ], + "spans": [ + { + "bbox": [ + 130, + 397, + 481, + 506 + ], + "type": "text", + "content": "Fig. 6-Fig. 10 show the qualitative results of various methods on different datasets. In Fig. 6, our method achieves better dehazing effects (right region) and exposure correction (sky). In the deraining task (Fig. 7, our method better removes rain streaks and restores textures in the occluded regions. In terms of denoising (Fig. 9) and deblurring (Fig. 8), we achieve clearer results with fewer artifacts. We also demonstrate better color correction (the purple blanket on the left) and exposure correction in low-light image enhancement tasks (Fig. 10). For simplicity, the qualitative effects of kernel deblurring and JPEG artifact removal will be presented in the supplementary material." + } + ] + } + ], + "index": 51 + }, + { + "bbox": [ + 132, + 524, + 238, + 536 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 524, + 238, + 536 + ], + "spans": [ + { + "bbox": [ + 132, + 524, + 238, + 536 + ], + "type": "text", + "content": "4.3 Ablation Study" + } + ] + } + ], + "index": 52 + }, + { + "bbox": [ + 130, + 544, + 480, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 544, + 480, + 582 + ], + "spans": [ + { + "bbox": [ + 130, + 544, + 480, + 582 + ], + "type": "text", + "content": "In this section, we conduct an ablative study on the masking ratio, mask patch size, pre-training strategy, fine-tuning strategy, and fine-tuning ratio to demonstrate the effectiveness of our MIM pre-training and fine-tuning strategy." + } + ] + } + ], + "index": 53 + }, + { + "type": "table", + "bbox": [ + 185, + 624, + 429, + 663 + ], + "blocks": [ + { + "bbox": [ + 203, + 603, + 409, + 614 + ], + "lines": [ + { + "bbox": [ + 203, + 603, + 409, + 614 + ], + "spans": [ + { + "bbox": [ + 203, + 603, + 409, + 614 + ], + "type": "text", + "content": "Table 3: Ablative results on masking ratios." + } + ] + } + ], + "index": 54, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 185, + 624, + 429, + 663 + ], + "lines": [ + { + "bbox": [ + 185, + 624, + 429, + 663 + ], + "spans": [ + { + "bbox": [ + 185, + 624, + 429, + 663 + ], + "type": "table", + "html": "
Masking ratio20%40%50%60%80%
PSNR↑27.2827.2127.2827.2627.08
SSIM↑0.86630.86830.86980.86940.8642
", + "image_path": "68e41432cc535332721b48ddce41f341d02b0922b00e0717b07bc5e59d5ffe63.jpg" + } + ] + } + ], + "index": 55, + "angle": 0, + "type": "table_body" + } + ], + "index": 55 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 206, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 206, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 206, + 102 + ], + "type": "text", + "content": "Qin et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 132, + 152, + 301, + 199 + ], + "blocks": [ + { + "bbox": [ + 132, + 119, + 301, + 142 + ], + "lines": [ + { + "bbox": [ + 132, + 119, + 301, + 142 + ], + "spans": [ + { + "bbox": [ + 132, + 119, + 301, + 142 + ], + "type": "text", + "content": "Table 4: Ablative results of different pre-training strategies." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 132, + 152, + 301, + 199 + ], + "lines": [ + { + "bbox": [ + 132, + 152, + 301, + 199 + ], + "spans": [ + { + "bbox": [ + 132, + 152, + 301, + 199 + ], + "type": "table", + "html": "
RAM-SwinIRPSNR↑SSIM↑
pre-trained w/ gt26.620.8580
pre-trained w/ paired data27.280.8698
", + "image_path": "97121b78211839d4c6f6f66021840ea852b03bd0b35d41ed1651c883dfb89c6f.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 312, + 146, + 470, + 205 + ], + "blocks": [ + { + "bbox": [ + 312, + 114, + 480, + 137 + ], + "lines": [ + { + "bbox": [ + 312, + 114, + 480, + 137 + ], + "spans": [ + { + "bbox": [ + 312, + 114, + 480, + 137 + ], + "type": "text", + "content": "Table 5: Ablative results of different fine-tuning strategies." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 312, + 146, + 470, + 205 + ], + "lines": [ + { + "bbox": [ + 312, + 146, + 470, + 205 + ], + "spans": [ + { + "bbox": [ + 312, + 146, + 470, + 205 + ], + "type": "table", + "html": "
RAM-SwinIRPSNR↑SSIM↑
random26.860.8535
IG [45]26.920.8554
MAC (Ours)27.280.8698
", + "image_path": "26df439a642d976441ef7d9dbda0bb81afe779c13b1fc89dfbe1e9c305019579.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 229, + 479, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 229, + 479, + 277 + ], + "spans": [ + { + "bbox": [ + 130, + 229, + 479, + 277 + ], + "type": "text", + "content": "- Patch size & masking ratio are two essential hyper-parameters that determine the continuity and area of the masking of an image. In high-level tasks, MAE [15] masks " + }, + { + "bbox": [ + 130, + 229, + 479, + 277 + ], + "type": "inline_equation", + "content": "75\\%" + }, + { + "bbox": [ + 130, + 229, + 479, + 277 + ], + "type": "text", + "content": " of an image with " + }, + { + "bbox": [ + 130, + 229, + 479, + 277 + ], + "type": "inline_equation", + "content": "16 \\times 16" + }, + { + "bbox": [ + 130, + 229, + 479, + 277 + ], + "type": "text", + "content": " patch size. However, it can corrupt the local details of images, which is not suitable for image restoration." + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 134, + 380, + 340, + 463 + ], + "blocks": [ + { + "bbox": [ + 132, + 314, + 340, + 380 + ], + "lines": [ + { + "bbox": [ + 132, + 314, + 340, + 380 + ], + "spans": [ + { + "bbox": [ + 132, + 314, + 340, + 380 + ], + "type": "text", + "content": "Table 6: Ablative results in terms of the PSNR on fine-tuning ratios. We compared the performance in restoring images with unseen noises (Out-of-Distribution Denoising) and known degraded images (In-Distribution). In this case, the settings of In-Distribution are the same as Tab. 1." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 134, + 380, + 340, + 463 + ], + "lines": [ + { + "bbox": [ + 134, + 380, + 340, + 463 + ], + "spans": [ + { + "bbox": [ + 134, + 380, + 340, + 463 + ], + "type": "table", + "html": "
MethodOut-of-DistributionIn-Distribution
PossionPepperSpeckleAverageAverage
SwinIR [28]12.8310.0020.8614.5626.64
RAM-SwinIR10%13.6719.2321.0717.9927.28
RAM-SwinIR20%13.2719.0920.6817.6827.35
RAM-SwinIR50%12.7516.5120.3616.5427.38
RAM-SwinIR100%12.4715.3120.0115.9327.54
", + "image_path": "242bcdf3f4c7dad4073185430ab17da5ed318b986ac31f1b0f2e03463a5d1a36.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 345, + 279, + 482, + 493 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 345, + 279, + 482, + 493 + ], + "spans": [ + { + "bbox": [ + 345, + 279, + 482, + 493 + ], + "type": "text", + "content": "We first find the best choice of patch size by pretraining SwinIR [28] on " + }, + { + "bbox": [ + 345, + 279, + 482, + 493 + ], + "type": "inline_equation", + "content": "1 \\times 1" + }, + { + "bbox": [ + 345, + 279, + 482, + 493 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 345, + 279, + 482, + 493 + ], + "type": "inline_equation", + "content": "4 \\times 4" + }, + { + "bbox": [ + 345, + 279, + 482, + 493 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 345, + 279, + 482, + 493 + ], + "type": "inline_equation", + "content": "8 \\times 8" + }, + { + "bbox": [ + 345, + 279, + 482, + 493 + ], + "type": "text", + "content": ", as shown in Fig. 3. Since the attention layers of SwinIR treat an " + }, + { + "bbox": [ + 345, + 279, + 482, + 493 + ], + "type": "inline_equation", + "content": "8 \\times 8" + }, + { + "bbox": [ + 345, + 279, + 482, + 493 + ], + "type": "text", + "content": " patch as a token, the " + }, + { + "bbox": [ + 345, + 279, + 482, + 493 + ], + "type": "inline_equation", + "content": "4 \\times 4" + }, + { + "bbox": [ + 345, + 279, + 482, + 493 + ], + "type": "text", + "content": " pre-training produces heavy artifacts. Besides, the results generated by " + }, + { + "bbox": [ + 345, + 279, + 482, + 493 + ], + "type": "inline_equation", + "content": "8 \\times 8" + }, + { + "bbox": [ + 345, + 279, + 482, + 493 + ], + "type": "text", + "content": " pretraining are highly missing details, e.g. the texture of the polar bear's paws. In contrast, the model pre-trained with " + }, + { + "bbox": [ + 345, + 279, + 482, + 493 + ], + "type": "inline_equation", + "content": "1 \\times 1" + }, + { + "bbox": [ + 345, + 279, + 482, + 493 + ], + "type": "text", + "content": " patch size, which is also our final choice, achieves a satisfactory reconstruction and removes most of the rain streaks." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 495, + 480, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 495, + 480, + 555 + ], + "spans": [ + { + "bbox": [ + 130, + 495, + 480, + 555 + ], + "type": "text", + "content": "Then, we adjust the masking ratio from " + }, + { + "bbox": [ + 130, + 495, + 480, + 555 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 130, + 495, + 480, + 555 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 130, + 495, + 480, + 555 + ], + "type": "inline_equation", + "content": "80\\%" + }, + { + "bbox": [ + 130, + 495, + 480, + 555 + ], + "type": "text", + "content": ". As we can see in Tab. 3, the model pre-trained with " + }, + { + "bbox": [ + 130, + 495, + 480, + 555 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 130, + 495, + 480, + 555 + ], + "type": "text", + "content": " achieves the highest performance. Moreover, the performance is significantly dropped from 27.28dB to 27.08dB in terms of PSNR when we continue to increase the masking ratio, which also demonstrates our opinion that a high masking ratio is harmful to image restoration." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 556, + 480, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 556, + 480, + 628 + ], + "spans": [ + { + "bbox": [ + 130, + 556, + 480, + 628 + ], + "type": "text", + "content": "Pre-trained with paired data. Tab. 4 compares the results of using paired data for mask image pretraining (our pretraining strategy) with those using only ground truth for mask image pretraining. It shows that pre-trained with paired data is necessary for our RAM. Pretraining the model on high-quality images does not effectively enable learning for image restoration tasks. It still requires paired data to guide the model in the learning process." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "type": "text", + "content": "Fine-tuning strategy. To verify the effectiveness of our fine-tuning strategy, we fine-tune " + }, + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "type": "text", + "content": " of the network layers selected through MAC analysis, IG [45], and uniform sampling, respectively, and the results are shown in Tab. 5. Compared" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 324, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 324, + 91, + 447, + 102 + ], + "type": "text", + "content": "Restore Anything with Masks" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 114, + 192, + 152 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 192, + 152 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 192, + 152 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 192, + 152 + ], + "type": "image", + "image_path": "3eb1c6379e11b9c597a9cd4116766d3fd8f7977cbe64dcef33609fe2dfce767f.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 192, + 114, + 249, + 152 + ], + "blocks": [ + { + "bbox": [ + 192, + 114, + 249, + 152 + ], + "lines": [ + { + "bbox": [ + 192, + 114, + 249, + 152 + ], + "spans": [ + { + "bbox": [ + 192, + 114, + 249, + 152 + ], + "type": "image", + "image_path": "c10fb5aaa480258cf5e68e0dea376870bff31258f391c4586b25767e271fc969.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 250, + 114, + 307, + 152 + ], + "blocks": [ + { + "bbox": [ + 250, + 114, + 307, + 152 + ], + "lines": [ + { + "bbox": [ + 250, + 114, + 307, + 152 + ], + "spans": [ + { + "bbox": [ + 250, + 114, + 307, + 152 + ], + "type": "image", + "image_path": "f296c3bf94de43094ed2fad425930485a7c873ae2da714e4892e99635b4fdfe0.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 307, + 114, + 365, + 152 + ], + "blocks": [ + { + "bbox": [ + 307, + 114, + 365, + 152 + ], + "lines": [ + { + "bbox": [ + 307, + 114, + 365, + 152 + ], + "spans": [ + { + "bbox": [ + 307, + 114, + 365, + 152 + ], + "type": "image", + "image_path": "0ef2171be860cc4c21eb3cc1b06a69c652c0c7c0f72604d763cfc912fbee288a.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 365, + 114, + 423, + 152 + ], + "blocks": [ + { + "bbox": [ + 365, + 114, + 423, + 152 + ], + "lines": [ + { + "bbox": [ + 365, + 114, + 423, + 152 + ], + "spans": [ + { + "bbox": [ + 365, + 114, + 423, + 152 + ], + "type": "image", + "image_path": "97647f418b561be0583fe5ecf511d7072ed9c2ec05d10a975e8a6148594376e0.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 424, + 114, + 481, + 152 + ], + "blocks": [ + { + "bbox": [ + 424, + 114, + 481, + 152 + ], + "lines": [ + { + "bbox": [ + 424, + 114, + 481, + 152 + ], + "spans": [ + { + "bbox": [ + 424, + 114, + 481, + 152 + ], + "type": "image", + "image_path": "5cdd7d46956bb4566507ede54decbc9374f95a75e8b1fff6aff1b2eade288dc7.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 133, + 162, + 191, + 201 + ], + "blocks": [ + { + "bbox": [ + 145, + 154, + 181, + 162 + ], + "lines": [ + { + "bbox": [ + 145, + 154, + 181, + 162 + ], + "spans": [ + { + "bbox": [ + 145, + 154, + 181, + 162 + ], + "type": "text", + "content": "Noisy Image" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 133, + 162, + 191, + 201 + ], + "lines": [ + { + "bbox": [ + 133, + 162, + 191, + 201 + ], + "spans": [ + { + "bbox": [ + 133, + 162, + 191, + 201 + ], + "type": "image", + "image_path": "5ab121741bc61962925984b993557e0bb4a60d5a3decbcdc2ba1f198c968b5e1.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 156, + 202, + 166, + 209 + ], + "lines": [ + { + "bbox": [ + 156, + 202, + 166, + 209 + ], + "spans": [ + { + "bbox": [ + 156, + 202, + 166, + 209 + ], + "type": "text", + "content": "DL" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 192, + 162, + 249, + 201 + ], + "blocks": [ + { + "bbox": [ + 206, + 154, + 230, + 161 + ], + "lines": [ + { + "bbox": [ + 206, + 154, + 230, + 161 + ], + "spans": [ + { + "bbox": [ + 206, + 154, + 230, + 161 + ], + "type": "text", + "content": "NAFNet" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 192, + 162, + 249, + 201 + ], + "lines": [ + { + "bbox": [ + 192, + 162, + 249, + 201 + ], + "spans": [ + { + "bbox": [ + 192, + 162, + 249, + 201 + ], + "type": "image", + "image_path": "dab4a93d4ef7270c882be456133b1181b65981ec386f39208741bb1fb6531885.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 211, + 202, + 227, + 209 + ], + "lines": [ + { + "bbox": [ + 211, + 202, + 227, + 209 + ], + "spans": [ + { + "bbox": [ + 211, + 202, + 227, + 209 + ], + "type": "text", + "content": "TAPE" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 250, + 162, + 307, + 201 + ], + "blocks": [ + { + "bbox": [ + 265, + 154, + 290, + 161 + ], + "lines": [ + { + "bbox": [ + 265, + 154, + 290, + 161 + ], + "spans": [ + { + "bbox": [ + 265, + 154, + 290, + 161 + ], + "type": "text", + "content": "MPRNet" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 250, + 162, + 307, + 201 + ], + "lines": [ + { + "bbox": [ + 250, + 162, + 307, + 201 + ], + "spans": [ + { + "bbox": [ + 250, + 162, + 307, + 201 + ], + "type": "image", + "image_path": "a8840f4335f7694ffd376fa7ad27df1e67c961863c03c3963a1436609ab92c4a.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 268, + 202, + 287, + 209 + ], + "lines": [ + { + "bbox": [ + 268, + 202, + 287, + 209 + ], + "spans": [ + { + "bbox": [ + 268, + 202, + 287, + 209 + ], + "type": "text", + "content": "AirNet" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 307, + 162, + 365, + 201 + ], + "blocks": [ + { + "bbox": [ + 321, + 154, + 349, + 161 + ], + "lines": [ + { + "bbox": [ + 321, + 154, + 349, + 161 + ], + "spans": [ + { + "bbox": [ + 321, + 154, + 349, + 161 + ], + "type": "text", + "content": "Restormer" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 307, + 162, + 365, + 201 + ], + "lines": [ + { + "bbox": [ + 307, + 162, + 365, + 201 + ], + "spans": [ + { + "bbox": [ + 307, + 162, + 365, + 201 + ], + "type": "image", + "image_path": "39214ef1d42f254db74bdbe602a7fbfaffbbb82e9594e195d8848aea9085d740.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 321, + 202, + 348, + 209 + ], + "lines": [ + { + "bbox": [ + 321, + 202, + 348, + 209 + ], + "spans": [ + { + "bbox": [ + 321, + 202, + 348, + 209 + ], + "type": "text", + "content": "PromptIR" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 364, + 202, + 424, + 209 + ], + "lines": [ + { + "bbox": [ + 364, + 202, + 424, + 209 + ], + "spans": [ + { + "bbox": [ + 364, + 202, + 424, + 209 + ], + "type": "text", + "content": "RAM-PromptIR(Ours)" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 365, + 162, + 422, + 201 + ], + "blocks": [ + { + "bbox": [ + 383, + 154, + 403, + 161 + ], + "lines": [ + { + "bbox": [ + 383, + 154, + 403, + 161 + ], + "spans": [ + { + "bbox": [ + 383, + 154, + 403, + 161 + ], + "type": "text", + "content": "SwinIR" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 365, + 162, + 422, + 201 + ], + "lines": [ + { + "bbox": [ + 365, + 162, + 422, + 201 + ], + "spans": [ + { + "bbox": [ + 365, + 162, + 422, + 201 + ], + "type": "image", + "image_path": "561a7269f3e4b6dfc33ee010a43da46c8764a9bc63ce403b3ac342a8e1e2895a.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 424, + 162, + 481, + 201 + ], + "blocks": [ + { + "bbox": [ + 424, + 154, + 480, + 161 + ], + "lines": [ + { + "bbox": [ + 424, + 154, + 480, + 161 + ], + "spans": [ + { + "bbox": [ + 424, + 154, + 480, + 161 + ], + "type": "text", + "content": "RAM-SwinIR (Ours)" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 424, + 162, + 481, + 201 + ], + "lines": [ + { + "bbox": [ + 424, + 162, + 481, + 201 + ], + "spans": [ + { + "bbox": [ + 424, + 162, + 481, + 201 + ], + "type": "image", + "image_path": "d399dc62949040a87f8d79296eecf419751f802a64644d15b90245742af82a31.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 447, + 202, + 457, + 209 + ], + "lines": [ + { + "bbox": [ + 447, + 202, + 457, + 209 + ], + "spans": [ + { + "bbox": [ + 447, + 202, + 457, + 209 + ], + "type": "text", + "content": "GT" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 133, + 243, + 191, + 283 + ], + "blocks": [ + { + "bbox": [ + 148, + 222, + 464, + 234 + ], + "lines": [ + { + "bbox": [ + 148, + 222, + 464, + 234 + ], + "spans": [ + { + "bbox": [ + 148, + 222, + 464, + 234 + ], + "type": "text", + "content": "Fig. 9: Denoising visual comparison on CBSD68 dataset. Zoom in for details." + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 133, + 243, + 191, + 283 + ], + "lines": [ + { + "bbox": [ + 133, + 243, + 191, + 283 + ], + "spans": [ + { + "bbox": [ + 133, + 243, + 191, + 283 + ], + "type": "image", + "image_path": "ae9e45b9f67927b8d0f715be350c576819dc7302813b65df51d73b399e403d4b.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 192, + 243, + 249, + 283 + ], + "blocks": [ + { + "bbox": [ + 192, + 243, + 249, + 283 + ], + "lines": [ + { + "bbox": [ + 192, + 243, + 249, + 283 + ], + "spans": [ + { + "bbox": [ + 192, + 243, + 249, + 283 + ], + "type": "image", + "image_path": "a91afc1ef3fa620d198399a06534bd079b75325f645256e0405a3a982694e344.jpg" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 207, + 285, + 230, + 291 + ], + "lines": [ + { + "bbox": [ + 207, + 285, + 230, + 291 + ], + "spans": [ + { + "bbox": [ + 207, + 285, + 230, + 291 + ], + "type": "text", + "content": "NAFNet" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_caption" + } + ], + "index": 29 + }, + { + "type": "image", + "bbox": [ + 250, + 243, + 307, + 283 + ], + "blocks": [ + { + "bbox": [ + 250, + 243, + 307, + 283 + ], + "lines": [ + { + "bbox": [ + 250, + 243, + 307, + 283 + ], + "spans": [ + { + "bbox": [ + 250, + 243, + 307, + 283 + ], + "type": "image", + "image_path": "e38c71cd938d93468c2150fceb3c73455849a23f02e3cc1de139c6f4dc27067b.jpg" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 266, + 285, + 290, + 291 + ], + "lines": [ + { + "bbox": [ + 266, + 285, + 290, + 291 + ], + "spans": [ + { + "bbox": [ + 266, + 285, + 290, + 291 + ], + "type": "text", + "content": "MPRNet" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_caption" + } + ], + "index": 31 + }, + { + "type": "image", + "bbox": [ + 307, + 243, + 365, + 283 + ], + "blocks": [ + { + "bbox": [ + 307, + 243, + 365, + 283 + ], + "lines": [ + { + "bbox": [ + 307, + 243, + 365, + 283 + ], + "spans": [ + { + "bbox": [ + 307, + 243, + 365, + 283 + ], + "type": "image", + "image_path": "0baa7a8329853d6f7e5293d32c53e9c4f9079b212dc4ea4bd9c700e37f0aa3a5.jpg" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 321, + 285, + 349, + 291 + ], + "lines": [ + { + "bbox": [ + 321, + 285, + 349, + 291 + ], + "spans": [ + { + "bbox": [ + 321, + 285, + 349, + 291 + ], + "type": "text", + "content": "Restormer" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_caption" + } + ], + "index": 33 + }, + { + "type": "image", + "bbox": [ + 365, + 243, + 422, + 283 + ], + "blocks": [ + { + "bbox": [ + 365, + 243, + 422, + 283 + ], + "lines": [ + { + "bbox": [ + 365, + 243, + 422, + 283 + ], + "spans": [ + { + "bbox": [ + 365, + 243, + 422, + 283 + ], + "type": "image", + "image_path": "a8db619b831cfdc3c4be0ddbb22abd61876e69ed84c00da1e2f00e194e05471c.jpg" + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 383, + 285, + 403, + 291 + ], + "lines": [ + { + "bbox": [ + 383, + 285, + 403, + 291 + ], + "spans": [ + { + "bbox": [ + 383, + 285, + 403, + 291 + ], + "type": "text", + "content": "SwinIR" + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_caption" + } + ], + "index": 35 + }, + { + "type": "image", + "bbox": [ + 424, + 243, + 481, + 283 + ], + "blocks": [ + { + "bbox": [ + 424, + 243, + 481, + 283 + ], + "lines": [ + { + "bbox": [ + 424, + 243, + 481, + 283 + ], + "spans": [ + { + "bbox": [ + 424, + 243, + 481, + 283 + ], + "type": "image", + "image_path": "8079c7c696c821a012b907fdb96052ca3bdec04c3eeb189f09b69deb2b549c22.jpg" + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "image_body" + } + ], + "index": 37 + }, + { + "type": "image", + "bbox": [ + 132, + 293, + 191, + 331 + ], + "blocks": [ + { + "bbox": [ + 139, + 285, + 186, + 293 + ], + "lines": [ + { + "bbox": [ + 139, + 285, + 186, + 293 + ], + "spans": [ + { + "bbox": [ + 139, + 285, + 186, + 293 + ], + "type": "text", + "content": "Low Light Image" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 132, + 293, + 191, + 331 + ], + "lines": [ + { + "bbox": [ + 132, + 293, + 191, + 331 + ], + "spans": [ + { + "bbox": [ + 132, + 293, + 191, + 331 + ], + "type": "image", + "image_path": "a3b2f6aa0992184b49931ed447c7f5dc813ebfe3963c0dcaf856825c35052b82.jpg" + } + ] + } + ], + "index": 39, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 156, + 332, + 166, + 339 + ], + "lines": [ + { + "bbox": [ + 156, + 332, + 166, + 339 + ], + "spans": [ + { + "bbox": [ + 156, + 332, + 166, + 339 + ], + "type": "text", + "content": "DL" + } + ] + } + ], + "index": 40, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 162, + 350, + 449, + 361 + ], + "lines": [ + { + "bbox": [ + 162, + 350, + 449, + 361 + ], + "spans": [ + { + "bbox": [ + 162, + 350, + 449, + 361 + ], + "type": "text", + "content": "Fig. 10: LLIE visual comparison on LOL dataset. Zoom in for details." + } + ] + } + ], + "index": 51, + "angle": 0, + "type": "image_caption" + } + ], + "index": 39 + }, + { + "type": "image", + "bbox": [ + 192, + 293, + 249, + 331 + ], + "blocks": [ + { + "bbox": [ + 192, + 293, + 249, + 331 + ], + "lines": [ + { + "bbox": [ + 192, + 293, + 249, + 331 + ], + "spans": [ + { + "bbox": [ + 192, + 293, + 249, + 331 + ], + "type": "image", + "image_path": "cada7c403bec112a9934f1a6819f52ae2f81868d1b3cf4ece8d1bb7346599167.jpg" + } + ] + } + ], + "index": 41, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 211, + 332, + 227, + 339 + ], + "lines": [ + { + "bbox": [ + 211, + 332, + 227, + 339 + ], + "spans": [ + { + "bbox": [ + 211, + 332, + 227, + 339 + ], + "type": "text", + "content": "TAPE" + } + ] + } + ], + "index": 42, + "angle": 0, + "type": "image_caption" + } + ], + "index": 41 + }, + { + "type": "image", + "bbox": [ + 250, + 293, + 307, + 331 + ], + "blocks": [ + { + "bbox": [ + 250, + 293, + 307, + 331 + ], + "lines": [ + { + "bbox": [ + 250, + 293, + 307, + 331 + ], + "spans": [ + { + "bbox": [ + 250, + 293, + 307, + 331 + ], + "type": "image", + "image_path": "d2881532b49bd7489af5a3b6db650445f1bc2fc21c9973bbb4d0b40ab7aed783.jpg" + } + ] + } + ], + "index": 43, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 268, + 332, + 287, + 339 + ], + "lines": [ + { + "bbox": [ + 268, + 332, + 287, + 339 + ], + "spans": [ + { + "bbox": [ + 268, + 332, + 287, + 339 + ], + "type": "text", + "content": "AirNet" + } + ] + } + ], + "index": 44, + "angle": 0, + "type": "image_caption" + } + ], + "index": 43 + }, + { + "type": "image", + "bbox": [ + 307, + 293, + 365, + 331 + ], + "blocks": [ + { + "bbox": [ + 307, + 293, + 365, + 331 + ], + "lines": [ + { + "bbox": [ + 307, + 293, + 365, + 331 + ], + "spans": [ + { + "bbox": [ + 307, + 293, + 365, + 331 + ], + "type": "image", + "image_path": "92e68fd5b163f2eb7728ab0a711655879deb41991c42e295532c7f2c6929dd71.jpg" + } + ] + } + ], + "index": 45, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 321, + 332, + 348, + 340 + ], + "lines": [ + { + "bbox": [ + 321, + 332, + 348, + 340 + ], + "spans": [ + { + "bbox": [ + 321, + 332, + 348, + 340 + ], + "type": "text", + "content": "PromptIR" + } + ] + } + ], + "index": 46, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 364, + 332, + 424, + 340 + ], + "lines": [ + { + "bbox": [ + 364, + 332, + 424, + 340 + ], + "spans": [ + { + "bbox": [ + 364, + 332, + 424, + 340 + ], + "type": "text", + "content": "RAM-PromptIR(Ours)" + } + ] + } + ], + "index": 48, + "angle": 0, + "type": "image_caption" + } + ], + "index": 45 + }, + { + "type": "image", + "bbox": [ + 365, + 293, + 422, + 331 + ], + "blocks": [ + { + "bbox": [ + 365, + 293, + 422, + 331 + ], + "lines": [ + { + "bbox": [ + 365, + 293, + 422, + 331 + ], + "spans": [ + { + "bbox": [ + 365, + 293, + 422, + 331 + ], + "type": "image", + "image_path": "bc41aa406fd82a4b3a47980b6bcff8da394b2452a3049e6f7413464c9ac7b4fa.jpg" + } + ] + } + ], + "index": 47, + "angle": 0, + "type": "image_body" + } + ], + "index": 47 + }, + { + "type": "image", + "bbox": [ + 424, + 293, + 481, + 331 + ], + "blocks": [ + { + "bbox": [ + 424, + 285, + 480, + 292 + ], + "lines": [ + { + "bbox": [ + 424, + 285, + 480, + 292 + ], + "spans": [ + { + "bbox": [ + 424, + 285, + 480, + 292 + ], + "type": "text", + "content": "RAM-SwinIR (Ours)" + } + ] + } + ], + "index": 38, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 424, + 293, + 481, + 331 + ], + "lines": [ + { + "bbox": [ + 424, + 293, + 481, + 331 + ], + "spans": [ + { + "bbox": [ + 424, + 293, + 481, + 331 + ], + "type": "image", + "image_path": "fd5ad5d957a22397858bfc0f8407e2ff4c70e4f553648befb592017ea15812cc.jpg" + } + ] + } + ], + "index": 49, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 447, + 332, + 457, + 339 + ], + "lines": [ + { + "bbox": [ + 447, + 332, + 457, + 339 + ], + "spans": [ + { + "bbox": [ + 447, + 332, + 457, + 339 + ], + "type": "text", + "content": "GT" + } + ] + } + ], + "index": 50, + "angle": 0, + "type": "image_caption" + } + ], + "index": 49 + }, + { + "bbox": [ + 130, + 384, + 480, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 384, + 480, + 407 + ], + "spans": [ + { + "bbox": [ + 130, + 384, + 480, + 407 + ], + "type": "text", + "content": "to IG, we have improved by 0.36 dB in PSNR and " + }, + { + "bbox": [ + 130, + 384, + 480, + 407 + ], + "type": "inline_equation", + "content": "1.6\\%" + }, + { + "bbox": [ + 130, + 384, + 480, + 407 + ], + "type": "text", + "content": " in SSIM, which indicates that our selection strategy is superior to IG." + } + ] + } + ], + "index": 52 + }, + { + "bbox": [ + 130, + 408, + 480, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 408, + 480, + 467 + ], + "spans": [ + { + "bbox": [ + 130, + 408, + 480, + 467 + ], + "type": "text", + "content": "Fine-tune ratio. We conduct the ablation experiment to compare the network's performances with different fine-tune ratios in Tab. 6. We found that using our finetune strategy, a pre-trained network could achieve comparable performance by fine-tuning only a few layers (e.g. " + }, + { + "bbox": [ + 130, + 408, + 480, + 467 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 130, + 408, + 480, + 467 + ], + "type": "text", + "content": "). At the same time, we need to finetune almost all network parameters to get the best performance on given tasks." + } + ] + } + ], + "index": 53 + }, + { + "bbox": [ + 130, + 468, + 480, + 527 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 468, + 480, + 527 + ], + "spans": [ + { + "bbox": [ + 130, + 468, + 480, + 527 + ], + "type": "text", + "content": "Performance vs Generalization capability. We found a trade-off between in-distribution performance and out-of-distribution generalization in Tab. 6. We found that the more layers fine-tuned, the less generalization capability to tackle the out-of-distribution tasks. With our fine-tuning method, the model can have stronger generalization while maintaining comparable performance." + } + ] + } + ], + "index": 54 + }, + { + "bbox": [ + 132, + 545, + 219, + 558 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 545, + 219, + 558 + ], + "spans": [ + { + "bbox": [ + 132, + 545, + 219, + 558 + ], + "type": "text", + "content": "5 Conclusion" + } + ] + } + ], + "index": 55 + }, + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "text", + "content": "This paper presents RAM, a pipeline for extracting intrinsic image information from corrupted images using Mask Image Modeling (MIM) pre-training. We design a MIM pre-training strategy tailored for image restoration and a fine-tuning algorithm to handle the transition from masked to complete images. By analyzing layer importance with MAC, we achieve high performance with minimal parameter tuning. Extensive experiments demonstrate that our RAM can bring boosts to various architectures and achieve state-of-the-art performance, moving towards a unified solution for all-in-one image restoration." + } + ] + } + ], + "index": 56 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 206, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 206, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 206, + 102 + ], + "type": "text", + "content": "Qin et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 197, + 126 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 197, + 126 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 197, + 126 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 140, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 24, + "blocks": [ + { + "bbox": [ + 138, + 140, + 480, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 140, + 480, + 172 + ], + "spans": [ + { + "bbox": [ + 138, + 140, + 480, + 172 + ], + "type": "text", + "content": "1. Chen, H., Wang, Y., Guo, T., Xu, C., Deng, Y., Liu, Z., Ma, S., Xu, C., Xu, C., Gao, W.: Pre-trained image processing transformer. In: CVPR. pp. 12299-12310 (2021)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 173, + 480, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 173, + 480, + 205 + ], + "spans": [ + { + "bbox": [ + 138, + 173, + 480, + 205 + ], + "type": "text", + "content": "2. Chen, H., Gu, J., Liu, Y., Magid, S.A., Dong, C., Wang, Q., Pfister, H., Zhu, L.: Masked image training for generalizable deep image denoising. In: CVPR. pp. 1692-1703 (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 205, + 480, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 205, + 480, + 227 + ], + "spans": [ + { + "bbox": [ + 138, + 205, + 480, + 227 + ], + "type": "text", + "content": "3. Chen, L., Chu, X., Zhang, X., Sun, J.: Simple baselines for image restoration. In: ECCV. pp. 17-33. Springer (2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 228, + 480, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 228, + 480, + 270 + ], + "spans": [ + { + "bbox": [ + 138, + 228, + 480, + 270 + ], + "type": "text", + "content": "4. Chen, W.T., Huang, Z.K., Tsai, C.C., Yang, H.H., Ding, J.J., Kuo, S.Y.: Learning multiple adverse weather removal via two-stage knowledge learning and multi-contrastive regularization: Toward a unified model. In: CVPR. pp. 17653-17662 (2022)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 271, + 480, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 271, + 480, + 293 + ], + "spans": [ + { + "bbox": [ + 138, + 271, + 480, + 293 + ], + "type": "text", + "content": "5. Chen Wei, Wenjing Wang, W.Y.J.L.: Deep retina decomposition for low-light enhancement. In: BMVC. British Machine VLOLision Association (2018)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 293, + 480, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 293, + 480, + 314 + ], + "spans": [ + { + "bbox": [ + 138, + 293, + 480, + 314 + ], + "type": "text", + "content": "6. Dhamdhere, K., Sundararajan, M., Yan, Q.: How important is a neuron. In: ICLR (2019)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 315, + 480, + 337 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 315, + 480, + 337 + ], + "spans": [ + { + "bbox": [ + 138, + 315, + 480, + 337 + ], + "type": "text", + "content": "7. Duan, H., Shen, W., Min, X., Tu, D., Teng, L., Wang, J., Zhai, G.: Masked autoencoders as image processors. arXiv preprint arXiv:2303.17316 (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 337, + 480, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 337, + 480, + 358 + ], + "spans": [ + { + "bbox": [ + 138, + 337, + 480, + 358 + ], + "type": "text", + "content": "8. Fan, Q., Chen, D., Yuan, L., Hua, G., Yu, N., Chen, B.: A general decoupled learning framework for parameterized image operators. PAMI 43(1), 33-47 (2019)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 359, + 480, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 359, + 480, + 380 + ], + "spans": [ + { + "bbox": [ + 138, + 359, + 480, + 380 + ], + "type": "text", + "content": "9. Fang, Y., Zhang, H., Wong, H.S., Zeng, T.: A robust non-blind deblurring method using deep denoiser prior. In: CVPRW. pp. 735-744 (June 2022)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 381, + 480, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 381, + 480, + 403 + ], + "spans": [ + { + "bbox": [ + 138, + 381, + 480, + 403 + ], + "type": "text", + "content": "0. Fu, X., Huang, J., Ding, X., Liao, Y., Paisley, J.: Clearing the skies: A deep network architecture for single-image rain removal. TIP 26(6), 2944-2956 (2017)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 403, + 480, + 424 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 403, + 480, + 424 + ], + "spans": [ + { + "bbox": [ + 138, + 403, + 480, + 424 + ], + "type": "text", + "content": "1. Fu, X., Huang, J., Zeng, D., Huang, Y., Ding, X., Paisley, J.: Removing rain from single images via a deep detail network. In: CVPR. pp. 3855-3863 (2017)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 425, + 480, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 425, + 480, + 445 + ], + "spans": [ + { + "bbox": [ + 138, + 425, + 480, + 445 + ], + "type": "text", + "content": "2. Gu, J., Dong, C.: Interpreting super-resolution networks with local attribution maps. In: CVPR. pp. 9199-9208 (2021)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 447, + 480, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 447, + 480, + 479 + ], + "spans": [ + { + "bbox": [ + 138, + 447, + 480, + 479 + ], + "type": "text", + "content": "3. Guo, C.L., Yan, Q., Anwar, S., Cong, R., Ren, W., Li, C.: Image dehazing transformer with transmission-aware 3d position embedding. In: CVPR. pp. 5812-5820 (2022)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 479, + 480, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 479, + 480, + 511 + ], + "spans": [ + { + "bbox": [ + 138, + 479, + 480, + 511 + ], + "type": "text", + "content": "4. Guo, C., Li, C., Guo, J., Loy, C.C., Hou, J., Kwong, S., Cong, R.: Zero-reference deep curve estimation for low-light image enhancement. In: CVPR. pp. 1780-1789 (2020)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 512, + 480, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 512, + 480, + 533 + ], + "spans": [ + { + "bbox": [ + 138, + 512, + 480, + 533 + ], + "type": "text", + "content": "5. He, K., Chen, X., Xie, S., Li, Y., Dollár, P., Girshick., R.: Masked autoencoders are scalable vision learners. In: CVPR. pp. 16000-16009 (2022)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 534, + 480, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 534, + 480, + 555 + ], + "spans": [ + { + "bbox": [ + 138, + 534, + 480, + 555 + ], + "type": "text", + "content": "6. Huang, J.B., Singh, A., Ahuja, N.: Single image super-resolution from transformed self-exemplars. In: CVPR. pp. 5197-5206 (2015)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 138, + 556, + 480, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 556, + 480, + 578 + ], + "spans": [ + { + "bbox": [ + 138, + 556, + 480, + 578 + ], + "type": "text", + "content": "7. Jin, X., Han, L.H., Li, Z., Guo, C.L., Chai, Z., Li, C.: Dnf: Decouple and feedback network for seeing in the dark. In: CVPR. pp. 18135-18144 (2023)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 138, + 578, + 480, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 578, + 480, + 599 + ], + "spans": [ + { + "bbox": [ + 138, + 578, + 480, + 599 + ], + "type": "text", + "content": "8. Kenton, J.D.M.W.C., Toutanova, L.K.: Bert: Pre-training of deep bidirectional transformers for language understanding. In: NAACL. pp. 4171-4186 (2019)" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 138, + 600, + 480, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 600, + 480, + 621 + ], + "spans": [ + { + "bbox": [ + 138, + 600, + 480, + 621 + ], + "type": "text", + "content": "9. Leino, K., Sen, S., Datta, A., Fredrikson, M., Li, L.: Influence-directed explanations for deep convolutional networks. In: ITC. pp. 1-8. IEEE (2018)" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 138, + 622, + 480, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 622, + 480, + 643 + ], + "spans": [ + { + "bbox": [ + 138, + 622, + 480, + 643 + ], + "type": "text", + "content": "20. Li, B., Ren, W., Fu, D., Tao, D., Feng, D., Zeng, W., Wang, Z.: Benchmarking single-image dehazing and beyond. TIP 28(1), 492-505 (2018)" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 138, + 644, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 644, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 644, + 480, + 665 + ], + "type": "text", + "content": "21. Li, B., Liu, X., Hu, P., Wu, Z., Lv, J., Peng, X.: All-in-one image restoration for unknown corruption. In: CVPR. pp. 17452-17462 (2022)" + } + ] + } + ], + "index": 23 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 324, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 324, + 91, + 447, + 102 + ], + "type": "text", + "content": "Restore Anything with Masks" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 666 + ], + "type": "list", + "angle": 0, + "index": 23, + "blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 480, + 138 + ], + "type": "text", + "content": "22. Li, C., Guo, C.L., Liang, Z., Zhou, S., Feng, R., Loy, C.C., et al.: Embedding fourier for ultra-high-definition low-light image enhancement. In: ICLR (2022)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 138, + 480, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 138, + 480, + 160 + ], + "spans": [ + { + "bbox": [ + 132, + 138, + 480, + 160 + ], + "type": "text", + "content": "23. Li, D., Zhang, Y., Cheung, K.C., Wang, X., Qin, H., Li, H.: Learning degradation representations for image deblurring. In: ECCV. pp. 736-753. Springer (2022)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 160, + 480, + 181 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 160, + 480, + 181 + ], + "spans": [ + { + "bbox": [ + 132, + 160, + 480, + 181 + ], + "type": "text", + "content": "24. Li, R., Tan, R.T., Cheong, L.F.: All in one bad weather removal using architectural search. In: CVPR. pp. 3175-3185 (2020)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 181, + 480, + 203 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 181, + 480, + 203 + ], + "spans": [ + { + "bbox": [ + 132, + 181, + 480, + 203 + ], + "type": "text", + "content": "25. Li, X., Wu, J., Lin, Z., Liu, H., Zha, H.: Recurrent squeeze-and-excitation context aggregation net for single image deraining. In: ECCV. pp. 254-269 (2018)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 203, + 480, + 236 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 203, + 480, + 236 + ], + "spans": [ + { + "bbox": [ + 132, + 203, + 480, + 236 + ], + "type": "text", + "content": "26. Li, Y., Zhang, K., Liang, J., Cao, J., Liu, C., Gong, R., Zhang, Y., Tang, H., Liu, Y., Demandolx, D., et al.: Lsdir: A large scale dataset for image restoration. In: CVPR. pp. 1775-1787 (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 236, + 480, + 256 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 236, + 480, + 256 + ], + "spans": [ + { + "bbox": [ + 132, + 236, + 480, + 256 + ], + "type": "text", + "content": "27. Li, Y., Tan, R.T., Guo, X., Lu, J., Brown, M.S.: Rain streak removal using layer priors. In: CVPR" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 256, + 480, + 278 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 256, + 480, + 278 + ], + "spans": [ + { + "bbox": [ + 132, + 256, + 480, + 278 + ], + "type": "text", + "content": "28. Liang, J., Cao, J., Sun, G., Zhang, K., Van Gool, L., Timofte, R.: Swinir: Image restoration using swin transformer. In: CVPR. pp. 1833-1844 (2021)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 278, + 480, + 310 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 278, + 480, + 310 + ], + "spans": [ + { + "bbox": [ + 132, + 278, + 480, + 310 + ], + "type": "text", + "content": "29. Lin, X., Ren, C., Liu, X., Huang, J., Lei, Y.: Unsupervised image denoising in real-world scenarios via self-collaboration parallel generative adversarial branches. In: ICCV. pp. 12642-12652 (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 310, + 480, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 310, + 480, + 342 + ], + "spans": [ + { + "bbox": [ + 132, + 310, + 480, + 342 + ], + "type": "text", + "content": "30. Lin, X., Yue, J., Ren, C., Guo, C.L., Li, C.: Unlocking low-light-rainy image restoration by pairwise degradation feature vector guidance. arXiv preprint arXiv:2305.03997 (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 342, + 480, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 342, + 480, + 374 + ], + "spans": [ + { + "bbox": [ + 132, + 342, + 480, + 374 + ], + "type": "text", + "content": "31. Liu, L., Xie, L., Zhang, X., Yuan, S., Chen, X., Zhou, W., Li, H., Tian, Q.: Tape: Task-agnostic prior embedding for image restoration. In: ECCV. pp. 447-464. Springer (2022)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 374, + 480, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 374, + 480, + 397 + ], + "spans": [ + { + "bbox": [ + 132, + 374, + 480, + 397 + ], + "type": "text", + "content": "32. Liu, Y., He, J., Gu, J., Kong, X., Qiao, Y., Dong, C.: Degae: A new pretraining paradigm for low-level vision. In: CVPR. pp. 23292-23303 (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 397, + 480, + 429 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 397, + 480, + 429 + ], + "spans": [ + { + "bbox": [ + 132, + 397, + 480, + 429 + ], + "type": "text", + "content": "33. Liu, Y., Liu, A., Gu, J., Zhang, Z., Wu, W., Qiao, Y., Dong, C.: Discovering distinctive \"semantics\" in super-resolution networks. arXiv preprint arXiv:2108.00406 (2021)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 429, + 480, + 460 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 429, + 480, + 460 + ], + "spans": [ + { + "bbox": [ + 132, + 429, + 480, + 460 + ], + "type": "text", + "content": "34. Liu, Z., Lin, Y., Cao, Y., Hu, H., Wei, Y., Zhang, Z., Lin, S., Guo, B.: Swin transformer: Hierarchical vision transformer using shifted windows. In: CVPR. pp. 10012-10022 (2021)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 460, + 480, + 482 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 460, + 480, + 482 + ], + "spans": [ + { + "bbox": [ + 132, + 460, + 480, + 482 + ], + "type": "text", + "content": "35. Luo, Y., Xu, Y., Ji, H.: Removing rain from a single image via discriminative sparse coding. In: ICCV. pp. 3397-3405 (2015)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 482, + 480, + 515 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 482, + 480, + 515 + ], + "spans": [ + { + "bbox": [ + 132, + 482, + 480, + 515 + ], + "type": "text", + "content": "36. Luo, Y., Zhao, R., Wei, X., Chen, J., Lu, Y., Xie, S., Wang, T., Xiong, R., Lu, M., Zhang, S.: Mowe: mixture of weather experts for multiple adverse weather removal. arXiv preprint arXiv:2303.13739 (2023)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 515, + 480, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 515, + 480, + 536 + ], + "spans": [ + { + "bbox": [ + 132, + 515, + 480, + 536 + ], + "type": "text", + "content": "37. Magid, S.A., Lin, Z., Wei, D., Zhang, Y., Gu, J., Pfister, H.: Texture-based error analysis for image super-resolution. In: CVPR. pp. 2118-2127 (2022)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 536, + 480, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 536, + 480, + 568 + ], + "spans": [ + { + "bbox": [ + 132, + 536, + 480, + 568 + ], + "type": "text", + "content": "38. Martin, D., Fowlkes, C., Tal, D., Malik, J.: A database of human segmented natural images and its application to evaluating segmentation algorithms and measuring ecological statistics. In: ICCV. vol. 2, pp. 416-423. IEEE (2001)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 568, + 480, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 568, + 480, + 590 + ], + "spans": [ + { + "bbox": [ + 132, + 568, + 480, + 590 + ], + "type": "text", + "content": "39. Mehri, A., Ardakani, P.B., Sappa, A.D.: Mprnet: Multi-path residual network for lightweight image super resolution. In: CVPR. pp. 2704-2713 (2021)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 132, + 590, + 480, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 590, + 480, + 611 + ], + "spans": [ + { + "bbox": [ + 132, + 590, + 480, + 611 + ], + "type": "text", + "content": "40. Nah, S., Hyun Kim, T., Mu Lee, K.: Deep multi-scale convolutional neural network for dynamic scene deblurring. In: CVPR" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 132, + 611, + 480, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 611, + 480, + 643 + ], + "spans": [ + { + "bbox": [ + 132, + 611, + 480, + 643 + ], + "type": "text", + "content": "41. Park, D., Lee, B.H., Chun, S.Y.: All-in-one image restoration for unknown degradations using adaptive discriminative filters for specific degradations. In: CVPR. pp. 5815-5824 (2023)" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 132, + 643, + 480, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 643, + 480, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 643, + 480, + 666 + ], + "type": "text", + "content": "42. Potlapalli, V., Zamir, S.W., Khan, S.H., Shahbaz Khan, F.: Prompt for all-in-one image restoration. NeurIPS 36 (2024)" + } + ] + } + ], + "index": 22 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 206, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 206, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 206, + 102 + ], + "type": "text", + "content": "Qin et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 588 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 138 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 138 + ], + "type": "text", + "content": "43. Radford, A., Narasimhan, K., Salimans, T., Sutskever, I., et al.: Improving language understanding by generative pre-training (2018)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 138, + 481, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 138, + 481, + 160 + ], + "spans": [ + { + "bbox": [ + 130, + 138, + 481, + 160 + ], + "type": "text", + "content": "44. Shrikumar, A., Su, J., Kundaje, A.: Computationally efficient measures of internal neuron importance. arXiv preprint arXiv:1807.09946 (2018)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 160, + 481, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 160, + 481, + 182 + ], + "spans": [ + { + "bbox": [ + 130, + 160, + 481, + 182 + ], + "type": "text", + "content": "45. Sundararajan, M., Taly, A., Yan, Q.: Axiomatic attribution for deep networks. In: ICML. pp. 3319-3328. PMLR (2017)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 182, + 481, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 182, + 481, + 194 + ], + "spans": [ + { + "bbox": [ + 130, + 182, + 481, + 194 + ], + "type": "text", + "content": "46. Sundararajan, M., Taly, A., Yan, Q.: Gradients of counterfactuals. ICLR (2017)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 194, + 481, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 194, + 481, + 215 + ], + "spans": [ + { + "bbox": [ + 130, + 194, + 481, + 215 + ], + "type": "text", + "content": "47. Wang, X., Wang, W., Cao, Y., Shen, C., Huang, T.: Images speak in images: A generalist painter for in-context visual learning. In: CVPR. pp. 6830-6839 (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 215, + 481, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 215, + 481, + 237 + ], + "spans": [ + { + "bbox": [ + 130, + 215, + 481, + 237 + ], + "type": "text", + "content": "48. Wang, Z., Cun, X., Bao, J., Zhou, W., Liu, J., Li, H.: Uformer: A general u-shaped transformer for image restoration. In: CVPR. pp. 17683-17693 (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 237, + 481, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 237, + 481, + 259 + ], + "spans": [ + { + "bbox": [ + 130, + 237, + 481, + 259 + ], + "type": "text", + "content": "49. Wu, R.Q., Duan, Z.P., Guo, C.L., Chai, Z., Li, C.: Ridcp: Revitalizing real image dehazing via high-quality codebook priors. In: CVPR. pp. 22282-22291 (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 259, + 481, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 259, + 481, + 281 + ], + "spans": [ + { + "bbox": [ + 130, + 259, + 481, + 281 + ], + "type": "text", + "content": "50. Xie, L., Wang, X., Dong, C., Qi, Z., Shan, Y.: Finding discriminative filters for specific degradations in blind super-resolution. NeurIPS 34, 51-61 (2021)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 281, + 481, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 281, + 481, + 304 + ], + "spans": [ + { + "bbox": [ + 130, + 281, + 481, + 304 + ], + "type": "text", + "content": "51. Xie, Z., Zhang, Z., Cao, Y., Lin, Y., Bao, J., Yao, Z., Dai, Q., Hu, H.: Simmim: a simple framework for masked image modeling. In: CVPR. pp. 9653-9663 (2022)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 304, + 481, + 325 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 304, + 481, + 325 + ], + "spans": [ + { + "bbox": [ + 130, + 304, + 481, + 325 + ], + "type": "text", + "content": "52. Yang, W., Tan, R.T., Feng, J., Liu, J., Guo, Z., Yan, S.: Deep joint rain detection and removal from a single image. In: CVPR. pp. 1357-1366 (2017)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 325, + 481, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 325, + 481, + 346 + ], + "spans": [ + { + "bbox": [ + 130, + 325, + 481, + 346 + ], + "type": "text", + "content": "53. Yang, W., Tan, R.T., Wang, S., Fang, Y., Liu, J.: Single image deraining: From model-based to data-driven and beyond. PAMI 43(11), 4059-4077 (2020)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 347, + 481, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 347, + 481, + 380 + ], + "spans": [ + { + "bbox": [ + 130, + 347, + 481, + 380 + ], + "type": "text", + "content": "54. Yang, W., Wang, W., Huang, H., Wang, S., Liu, J.: Sparse gradient regularized deep retina network for robust low-light image enhancement. TIP 30, 2072-2086 (2021)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 380, + 481, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 380, + 481, + 413 + ], + "spans": [ + { + "bbox": [ + 130, + 380, + 481, + 413 + ], + "type": "text", + "content": "55. Zamir, S.W., Arora, A., Khan, S., Hayat, M., Khan, F.S., Yang, M.H.: Restormer: Efficient transformer for high-resolution image restoration. In: CVPR. pp. 5728-5739 (2022)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 413, + 481, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 413, + 481, + 445 + ], + "spans": [ + { + "bbox": [ + 130, + 413, + 481, + 445 + ], + "type": "text", + "content": "56. Zhang, C., Zhu, Y., Yan, Q., Sun, J., Zhang, Y.: All-in-one multi-degradation image restoration network via hierarchical degradation representation. In: ACMMM. pp. 2285-2293 (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 445, + 481, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 445, + 481, + 468 + ], + "spans": [ + { + "bbox": [ + 130, + 445, + 481, + 468 + ], + "type": "text", + "content": "57. Zhang, H., Patel, V.M.: Density-aware single image de-raining using a multi-stream dense network. In: CVPR. pp. 695-704 (2018)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 130, + 468, + 481, + 490 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 468, + 481, + 490 + ], + "spans": [ + { + "bbox": [ + 130, + 468, + 481, + 490 + ], + "type": "text", + "content": "58. Zhang, H., Sindagi, V., Patel, V.M.: Image de-raining using a conditional generative adversarial network. TCSVT 30(11), 3943-3956 (2019)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 130, + 490, + 481, + 522 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 490, + 481, + 522 + ], + "spans": [ + { + "bbox": [ + 130, + 490, + 481, + 522 + ], + "type": "text", + "content": "59. Zhang, J., Huang, J., Yao, M., Yang, Z., Yu, H., Zhou, M., Zhao, F.: Ingredient-oriented multi-degradation learning for image restoration. In: CVPR. pp. 5825-5835 (2023)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 130, + 522, + 481, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 522, + 481, + 555 + ], + "spans": [ + { + "bbox": [ + 130, + 522, + 481, + 555 + ], + "type": "text", + "content": "60. Zheng, N., Zhou, M., Dong, Y., Rui, X., Huang, J., Li, C., Zhao, F.: Empowering low-light image enhancer through customized learnable priors. In: ICCV. pp. 12559-12569 (2023)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 130, + 555, + 481, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 555, + 481, + 588 + ], + "spans": [ + { + "bbox": [ + 130, + 555, + 481, + 588 + ], + "type": "text", + "content": "61. Zhu, Y., Wang, T., Fu, X., Yang, X., Guo, X., Dai, J., Qiao, Y., Hu, X.: Learning weather-general and weather-specific features for image restoration under multiple adverse weather conditions. In: CVPR. pp. 21747-21758 (2023)" + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 325, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 325, + 91, + 447, + 102 + ], + "type": "text", + "content": "Restore Anything with Masks" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Restoring Images in Adverse Weather Conditions via Histogram Transformer/95e1934a-8ade-4d7d-8947-f969acbf3910_content_list.json b/2024/Restoring Images in Adverse Weather Conditions via Histogram Transformer/95e1934a-8ade-4d7d-8947-f969acbf3910_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..4e49468c7a9a21d66c9a51dcc0f07eb97a536f29 --- /dev/null +++ b/2024/Restoring Images in Adverse Weather Conditions via Histogram Transformer/95e1934a-8ade-4d7d-8947-f969acbf3910_content_list.json @@ -0,0 +1,2054 @@ +[ + { + "type": "text", + "text": "Restoring Images in Adverse Weather Conditions via Histogram Transformer", + "text_level": 1, + "bbox": [ + 218, + 140, + 782, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Shangquan Sun $^{1,2}$ , Wenqi Ren $^{3,4\\dagger}$ , Xinwei Gao $^{5}$ , Rui Wang $^{1,2}$ , and Xiaochun Cao $^{3}$", + "bbox": [ + 233, + 212, + 767, + 244 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Institute of Information Engineering, Chinese Academy of Sciences, Beijing 100085, China", + "bbox": [ + 215, + 255, + 782, + 282 + ], + "page_idx": 0 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "$^{2}$ School of Cyber Security, University of Chinese Academy of Sciences, Beijing 100049, China", + "3 School of Cyber Science and Technology, Shenzhen Campus of Sun Yat-sen University, Shenzhen 518107, China", + "4 Guangdong Provincial Key Laboratory of Information Security Technology, Guangzhou 510006, China", + "5 Wechat Business Group, Tencent, Shenzhen, Guangdong, China, 518057 shangquansun@gmail.com, renwq3@mail.sysu.edu.cn," + ], + "bbox": [ + 236, + 282, + 764, + 395 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Transformer-based image restoration methods in adverse weather have achieved significant progress. Most of them use self-attention along the channel dimension or within spatially fixed-range blocks to reduce computational load. However, such a compromise results in limitations in capturing long-range spatial features. Inspired by the observation that the weather-induced degradation factors mainly cause similar occlusion and brightness, in this work, we propose an efficient Histogram Transformer (Histoformer) for restoring images affected by adverse weather. It is powered by a mechanism dubbed histogram self-attention, which sorts and segments spatial features into intensity-based bins. Self-attention is then applied across bins or within each bin to selectively focus on spatial features of dynamic range and process similar degraded pixels of the long range together. To boost histogram self-attention, we present a dynamic-range convolution enabling conventional convolution to conduct operation over similar pixels rather than neighbor pixels. We also observe that the common pixel-wise losses neglect linear association and correlation between output and ground-truth. Thus, we propose to leverage the Pearson correlation coefficient as a loss function to enforce the recovered pixels following the identical order as ground-truth. Extensive experiments demonstrate the efficacy and superiority of our proposed method. We have released the codes in Github.", + "bbox": [ + 259, + 445, + 743, + 736 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Image restoration $\\cdot$ Adverse weather removal $\\cdot$ Image Desnowing $\\cdot$ Image deraining $\\cdot$ Image dehazing $\\cdot$ Raindrop removal", + "bbox": [ + 259, + 752, + 750, + 780 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Corresponding author.", + "bbox": [ + 230, + 825, + 397, + 840 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/f5d18b2d309afc13747df1a5f189e25e71952145d3f26a50075d2f2b1e56e16b.jpg", + "image_caption": [ + "(a) Input patches" + ], + "image_footnote": [], + "bbox": [ + 299, + 142, + 364, + 193 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/c54f0c05cb88496d0f5f964e58fc5acea9032acea793d3a0604509af6614e25e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 367, + 143, + 434, + 193 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/6854636b0234474bb2f3174f8c96938ffde31dae682b51dd88fb2fc8f17382a0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 437, + 143, + 504, + 193 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/41d4b01bb1cd440bace641cb76e72248c7aea15e00174e9eb3e8494087ac6a5d.jpg", + "image_caption": [ + "(b) Existing self-attention" + ], + "image_footnote": [], + "bbox": [ + 308, + 209, + 395, + 282 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/ded64478d816fc6f5118ed522ba8a0ee8b43f7828a84b3052b03ee540c08451e.jpg", + "image_caption": [ + "Fig. 1: Given weather-degraded images in (a), traditional transformers perform self-attention either along the channel dimension or within a fixed-range block as shown in (b). In contrast, we observe that weather-induced degradation patterns tend to be similar but distinct from the background. So we categorize pixels affected by adverse weather and background pixels into distinct bins based on descending intensities (as depicted in (c)) and then conducts self-attention within and between bins." + ], + "image_footnote": [], + "bbox": [ + 410, + 209, + 496, + 282 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/f57e0df975ba7aaeed41ebcf7f3d3ee334dc05603a2c913986d306c42d7285d2.jpg", + "image_caption": [ + "(c) Histogram self-attention" + ], + "image_footnote": [], + "bbox": [ + 516, + 146, + 692, + 282 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 415, + 374, + 430 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The field of computer vision witnessed growing interest in restoring images affected by adverse weather conditions like rain, fog, and snow. These weather conditions significantly degrade visual quality, impacting the performance of downstream tasks such as object detection [3,60], and depth estimation [16,18]. The restoration of images under adverse weather is thereby a vital problem for the sake of vision aesthetics and safety.", + "bbox": [ + 212, + 446, + 787, + 537 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Early works leverage weather-related priors to model statistical characteristic of degradation and remove adverse weathers [1, 2, 22, 23, 35, 76, 83, 95, 99]. Subsequently, convolutional neural networks (CNNs) have emerged as powerful tools for addressing deraining [5, 17, 34, 40, 57, 73, 79, 80, 88, 89], dehazing [27, 29, 62-64, 77, 88, 90, 92] and desnowing [30, 44, 61, 94]. However, the need of separately training networks for each task and the complexity of switching among multiple models present challenges for real-world applications. Li et al. [33] thus introduced the challenge of adverse weather removal, which entails the restoration of images affected by various weather conditions using a single unified model.", + "bbox": [ + 212, + 537, + 787, + 672 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recently, Transformer-based approaches have also been investigated for the adverse weather removal task, surpassing the efficacy of CNNs [11, 19, 70, 72]. Nonetheless, these Transformer-based methods usually make concessions regarding efficient memory utilization by confining self-attention operations to a fixed spatial range or solely within the channel dimension, as depicted in Figure 1b. This compromise impedes the inherent potential of Transformers, which was originally designed for superior global feature modeling, and consequently, it leads to a deterioration in the performance of restoration.", + "bbox": [ + 212, + 674, + 787, + 794 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address these problems, based on the observation that weather-induced degradation often exhibits common patterns shown in Figure 1a, we develop an efficient Histogram Transformer for unified adverse weather removal, named", + "bbox": [ + 212, + 795, + 787, + 839 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "S. Sun et al.", + "bbox": [ + 271, + 114, + 354, + 126 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Histoformer. Specifically, we introduce a Dynamic-range Histogram Self-Attention (DHSA) module, which endows self-attention with a dynamic-range spatial receptive field. We categorize pixel values proximate in intensity yet varied in spatial location into histogram bins. Self-attention is executed across the dimension of bin or frequency, whose process is illustrated in Figure 1c. To facilitate comprehensive feature extraction on both local and global scales, we devise two ways of reshaping for histogram self-attention: bin-wise histogram reshaping (BHR) and frequency-wise histogram reshaping (FHR). In BHR, the number of bins is configured to incorporate pixels spanning a more comprehensive intensity range, thereby facilitating global feature integration. In FHR, the number of frequencies is assigned such that each bin focuses on limited number of pixels, enhancing the utility of finer features. Consequently, the histogram self-attention attains the capability of modeling spatially dynamic ranges effectively.", + "bbox": [ + 212, + 146, + 787, + 343 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To enable the convolution to extract dynamically-located weather-related dependencies, we develop a dynamic-range convolution layer, which involves sequential horizontal and vertical pixel sorting prior to the application of separable convolution. In order to capture multi-scale and multi-range information embedded within feature matrices, we introduce a Dual-scale Gated Feed-Forward (DGFF) module, enhancing its ability to model the visual characteristics effectively. Additionally, we note that conventional loss functions primarily focus on pixel-level closeness, overlooking the correlation at overall patch level. Consequently, we propose to leverage the Pearson correlation coefficient [12] to ensure the reconstruction of the linear relationship between restored and clean images.", + "bbox": [ + 212, + 343, + 787, + 494 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our contribution can be summarized in three folds:", + "bbox": [ + 238, + 494, + 611, + 508 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose a novel transformer targeted for unified adverse weather removal, equipped with a new histogram self-attention. It possesses dynamic-range spatial attention to weather-induced obstructions and thus can achieve degradation removal globally and efficiently.", + "- To capture multi-range information, we present a dual-scale feed-forward module. To enhance the comprehensive linear association between the recovered and ground-truth images, we develop a correlation loss.", + "- Our method attains state-of-the-art performance across various datasets. Additionally, we substantiate the efficacy of the proposed approach to restore real-world images and bolster the downstream application of detection." + ], + "bbox": [ + 225, + 521, + 782, + 671 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 698, + 387, + 713 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Extensive research has been dedicated to addressing adverse weather removal challenges in computer vision, including tasks like deraining [5,17,34,40,66,73, 79,80,88,89], dehazing [27,29,62-64,67,77,88,90,92], desnowing [30,44,61,94], raindrop removal [57,59,83,93] and All-in-One weather removal [28,33,53,70].", + "bbox": [ + 212, + 732, + 782, + 792 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Rain Streak Removal. The evolution of approaches is notable in rain streak removal techniques in computer vision. Kang et al. [23] pioneered a single image", + "bbox": [ + 214, + 809, + 784, + 839 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "Restoring Images in Adverse Weather Conditions via Histogram Transformer", + "bbox": [ + 218, + 114, + 730, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "deraining method using bilateral filters to decompose images into low and high-frequency components. However, recent advancements have seen a dominance of deep neural networks. An early deep CNN was introduced by Fu et al. [17] for extracting features from the high-frequency rain component, while Yang et al. [79] utilized recurrent networks to decompose rain layers and remove various streak types. Li et al. [32] proposed a method that addresses rain streaks and veiling effects in heavy rain scenes by integrating physics-based rain models and adversarial learning. A conditional generative adversarial network was also employed to solve rain streak removal [89]. Yasarla et al. [81] explored Gaussian processes for transfer learning from synthetic to real-world rain data. Quan et al. [58] used a cascaded network to remove both rain streaks and raindrops. Recently an image deraining Transformer [78] featuring a dual Transformer architecture was intricately formulated, incorporating both window-based and spatial-based mechanisms, thereby attaining exemplary outcomes. A sparse deraining Transformer is also proposed to enhance feature aggregation [11].", + "bbox": [ + 212, + 146, + 787, + 375 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Raindrop Removal. Raindrop removal from single images has been addressed through various methods, with some relying on traditional hand-crafted features. An early work incorporated temporal information to address video-based raindrop removal [83]. Eigen et al. [15] employed a shallow CNN trained with image pairs containing raindrop-degraded and raindrop-free versions, though the results often exhibited blurriness. Qian et al. [57] introduced an attention GAN and a new dataset. Their method was later improved by Quan et al. [59] via generating attention maps based on mathematical raindrop descriptions and combining them with detected raindrop edges.", + "bbox": [ + 212, + 387, + 787, + 525 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Snow Removal. Desnow-Net [44] was among the pioneering CNN-based approaches for snow removal, followed by Li et al.'s stacked dense network [31] and Chen et al.'s JSTASR [8], which introduced a size and transparency aware method. More recently, DDMSNet [94] introduced a dense multi-scale network that leverages semantic and geometric priors to enhance snow removal. A hierarchical decomposition paradigm involving the dual-tree wavelet transform for snow removal is also proposed [9]. Chen et al. [7] designed SnowFormer, a framework that used cross-attention to establish local-global context interaction.", + "bbox": [ + 212, + 537, + 787, + 659 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Fog Removal. Li et al. [26] presented a CNN that takes into account both atmospheric luminosity and transmission maps to conduct dehazing. Ren et al. [63] advocated a pre-processing approach for hazy image manipulation, thereby engendering multiple input modalities and, in the process, inducing chromatic aberrations as part of their dehazing procedure. A hierarchical density-aware network is also introduced, specializing in the domain of image dehazing [92]. Zheng et al. [97] formulated a curriculum-based contrastive regularization dehazing method aimed at fostering agreement within a contrastive space.", + "bbox": [ + 212, + 672, + 787, + 795 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "All-in-One Weather Removal. Some recent works attempted to address various weather-induced degradations by a singular network. Li et al. [33] proposed an", + "bbox": [ + 214, + 809, + 785, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "S. Sun et al.", + "bbox": [ + 271, + 114, + 354, + 126 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "All-in-One network, containing a generator comprising multiple task-specific encoders and a shared decoder. Valanarasu et al. [70] presented TransWeather, a transformer-based model featuring a solitary encoder-decoder structure, capable of rejuvenating images afflicted by various atmospheric conditions. A pipeline for the automatic selection of weather-degraded data was also proposed to enhance existing models [91]. Zhu et al. developed WGWS-Net [100] capable of learning weather-general and weather-specific in two separate stages. Some other recent works also trial addressing adverse weather removal by adopting probabilistic denoising diffusion model [53], knowledge distillation [10], large-scale Pre-trained model [68], mixture of experts [49], few-shot learning [24], codebooks [41,71,82], adaptive filters [54], knowledge assignment [74] and domain translation [56].", + "bbox": [ + 212, + 146, + 787, + 313 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Transformer-based Image Restoration. Since the inception of the Vision Transformer (ViT) [14] for visual recognition, transformers have gained substantial traction across a spectrum of computer vision tasks [25, 37, 45, 46, 50, 51, 86]. Particularly within the realm of low-level vision, the Image Processing Transformer [4] exemplifies how pre-training a transformer on extensive datasets can significantly enhance performance for low-level applications. U-former [75], on the other hand, introduced a transformer architecture based on the U-Net design for restoration tasks. Swin-IR [38] employed the Swin Transformer [45] for image restoration. Some latest Transformer-based methods were proposed for deraining [11, 39], desnowing [7], dehazing [19, 43, 65] and All-in-One weather removal [70, 72].", + "bbox": [ + 212, + 329, + 787, + 494 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Unlike the existing Transformer-based approaches whose self-attention is applied within either fixed spatial ranges or merely channel dimension, our method enables dynamic-range spatial attention to adaptively focus on weather-induced degradation with similar patterns.", + "bbox": [ + 212, + 496, + 787, + 556 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3 Method", + "text_level": 1, + "bbox": [ + 214, + 582, + 330, + 598 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1 Overall Architecture", + "text_level": 1, + "bbox": [ + 214, + 614, + 433, + 630 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The overall architectural framework of our Histoformer is illustrated in Figure 2. Suppose the input is a low-quality image $I^{lq} \\in \\mathbb{R}^{3 \\times H \\times W}$ , we pass the input through a $3 \\times 3$ convolution to conduct the overlapping image patch-embedding. Within both the encoder and decoder of the network backbone, we arrange Histogram Transformer Blocks (HTBs) to extract intricate features and capture dynamically distributed degradation factors. Within the same stage, encoders and decoders are interlinked through skip-connections, thereby establishing connections between consecutive intermediate features to enhance the stability of the training process. Between each stage, we apply pixel-unshuffle and pixelshuffle operations for the purpose of feature down-sampling and up-sampling.", + "bbox": [ + 212, + 642, + 787, + 794 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Within each HTB, we introduce Dynamic-range Histogram Self-Attention (DHSA) to extract spatially dynamic weather degradation and enhance both local and global feature aggregation. Moreover, a Dual-scale Gated Feed-Forward", + "bbox": [ + 214, + 795, + 787, + 839 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "Restoring Images in Adverse Weather Conditions via Histogram Transformer", + "bbox": [ + 218, + 114, + 732, + 130 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/ea63ebb64c2fdd34c12ae0e50e98c82e17d225bfd71eb58ff59d6cd2d8b773fa.jpg", + "image_caption": [ + "Fig. 2: The overall architecture of our Histoformer for weather removal. The main component is the Histogram Transformer block, and it comprises the Dynamic-range Histogram Self-Attention (DHSA) module and the Dual-scale Gated Feed-Forward (DGFF) module. Within DHSA, we present two types of reshaping mechanism, i.e., Bin-wise Histogram Reshaping and Frequency-wise Histogram Reshaping." + ], + "image_footnote": [], + "bbox": [ + 215, + 141, + 782, + 349 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "(DGFF) module is integrated into the HTB to enrich the representation of multi-range features, contributing to the process of image restoration. During each stage of encoding phases, our model is equipped with a crude skip-connection for supplementing original features from input, comprised of a sequence of operations, including average pooling, pixel-wise convolution, and depth-wise convolution. We start the crude skip-connection after the first stage, and this setup enables the encoders to focus more effectively on learning the weather-induced residuals. Through this hybrid formulation, Histoformer is empowered to exploit both the adaptive contents of weather-irrelevant background and the inherent characteristics of weather-degraded patterns, facilitating the separation of undesired degradation from the latent clear background.", + "bbox": [ + 212, + 455, + 787, + 623 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.2 Histogram Transformer Block", + "text_level": 1, + "bbox": [ + 215, + 642, + 508, + 657 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "As the key component of our Histoformer, HTB contains two pivotal modules, i.e., DHSA and DGFF. These two components are arranged to interact with layer normalization and can be formulated as follows:", + "bbox": [ + 212, + 664, + 785, + 710 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nF _ {l} = F _ {l - 1} + \\mathrm {D H S A} (\\mathrm {L N} (F _ {l - 1})), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 383, + 717, + 784, + 733 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nF _ {l} = F _ {l} + \\operatorname {D G F F} \\left(\\ln \\left(F _ {l}\\right)\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 383, + 736, + 784, + 752 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where LN denotes layer normalization and $F_{l}$ represents the feature at $l$ -th stage. The details of DHSA and DGFF are presented in Section 3.2 and 3.2 respectively.", + "bbox": [ + 212, + 758, + 782, + 790 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Dynamic-range Histogram Self-Attention To better capture dynamically distributed weather-induced degradation, we develop a Dynamic-range Histogram", + "bbox": [ + 212, + 809, + 790, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "S. Sun et al.", + "bbox": [ + 271, + 114, + 354, + 126 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Self-Attention (DHSA) module. This module consists of a process involving dynamic-range convolution, which reorders the spatial distribution of fractional features, and a dual-path histogram self-attention mechanism that combines global and local dynamic feature aggregation. Prior to the final output projection of a $1 \\times 1$ point-wise convolution, the reordered features are sorted back into their original locations to maintain spatial consistency.", + "bbox": [ + 212, + 146, + 787, + 238 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Dynamic-range Convolution. Traditional convolution operations employ fixed kernel sizes, resulting in a limited receptive field range and consequently perform local and small-range computations. This restricted scope of convolution, which primarily focuses on local information, does not naturally align with the self-attention mechanism's capacity to model long-range dependencies. To address this limitation, we devise a dynamic-range convolution technique by meticulously reordering the input features prior to the traditional convolution operation. Given an input feature $F \\in \\mathbb{R}^{C \\times H \\times W}$ , we divide it into two branches along the channel dimension, namely $F_{1}$ and $F_{2}$ . For the first branch of features, we perform sorting operations both horizontally and vertically, subsequently concatenating the sorted features with the second branch of feature. The resulting recombined features are then passed through the subsequent separable convolution. The entire process is articulated as follows:", + "bbox": [ + 212, + 248, + 787, + 445 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\nF _ {1}, F _ {2} = \\operatorname {S p l i t} (F), F _ {1} = \\operatorname {S o r t} _ {\\mathrm {v}} (\\operatorname {S o r t} _ {\\mathrm {h}} (F _ {1})),\n$$\n", + "text_format": "latex", + "bbox": [ + 352, + 452, + 651, + 468 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} F = \\operatorname {C o n v} _ {3 \\times 3} ^ {\\mathrm {d}} \\left(\\operatorname {C o n v} _ {1 \\times 1} \\left(\\operatorname {C o n c a t} \\left(F _ {1}, F _ {2}\\right)\\right)\\right), \\end{array} \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 352, + 460, + 784, + 488 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $\\mathrm{Conv}_{1\\times 1}$ is $1\\times 1$ point-wise convolution, $\\mathrm{Conv}_{3\\times 3}^{\\mathrm{d}}$ represents $3\\times 3$ depthwise convolution, Concat is the concatenation operation along channel, Split denotes the operation of splitting features along channel dimension, and $\\mathrm{Sort}_{i\\in \\{\\mathrm{h},\\mathrm{v}\\}}$ represents the horizontal or vertical sorting operation. This approach organizes pixels of high and low intensities into regular patterns at the diagonal corners of the matrices, thereby allowing convolution to perform computations across dynamic ranges. Given that weather-induced degradation typically exhibits closely related patterns, degraded pixels tend to concentrate in neighboring locations, separated with those clean pixels. As a result, this arrangement enables convolution kernels to partially focus on preserving clean information and separately recovering degraded features.", + "bbox": [ + 212, + 496, + 787, + 662 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Histogram Self-Attention. Existing vision Transformers [11,75,78,78,86,96] typically leverage fixed range of attention or merely the attention along channel dimension due to the compromise of computation and memory efficiency. However, the fixed setting restricts the self-attention to span adaptively long range to associate desired features. We notice that weather-induced degradation causes similar patterns and that those pixels containing either background feature or weather degradation of different intensities had better be assigned with various extents of attention. We thus propose a histogram self-attention mechanism to categorize spatial elements into bins and allocate varying attention within and across bins. For the sake of parallel computing, we set each bin contains identical number of pixels during implement.", + "bbox": [ + 212, + 672, + 787, + 840 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "Restoring Images in Adverse Weather Conditions via Histogram Transformer", + "bbox": [ + 218, + 114, + 732, + 130 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Given the output of dynamic-range convolution, we can separate them into Value feature $V \\in \\mathbb{R}^{C \\times H \\times W}$ and two pairs of Query-key $F_{QK,1}, F_{QK,2} \\in \\mathbb{R}^{2C \\times H \\times W}$ subsequently passing to two branches. We firstly sort the sequence of $V$ and based on its index arrange the Query-Key pairs accordingly, expressed as follows:", + "bbox": [ + 212, + 146, + 799, + 207 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\nV, d = \\operatorname {S o r t} \\left(\\mathbf {R} _ {\\mathrm {C} \\times \\mathrm {H} \\times \\mathrm {W}} ^ {\\mathrm {C} \\times \\mathrm {H W}} (V)\\right),\n$$\n", + "text_format": "latex", + "bbox": [ + 334, + 212, + 529, + 229 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\nQ _ {1}, K _ {1} = \\operatorname {S p l i t} \\left(\\operatorname {G a t h e r} \\left(\\mathbf {R} _ {\\mathrm {C} \\times \\mathrm {H} \\times \\mathrm {W}} ^ {\\mathrm {C} \\times \\mathrm {H W}} \\left(F _ {Q K, 1}\\right), d\\right)\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 334, + 233, + 784, + 252 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\nQ _ {2}, K _ {2} = \\operatorname {S p l i t} \\left(\\operatorname {G a t h e r} \\left(\\mathbf {R} _ {\\mathrm {C} \\times \\mathrm {H} \\times \\mathrm {W}} ^ {\\mathrm {C} \\times \\mathrm {H W}} (F _ {Q K, 2}), d\\right)\\right),\n$$\n", + "text_format": "latex", + "bbox": [ + 334, + 255, + 669, + 273 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $\\mathbf{R}_{\\mathrm{C}\\times ,\\mathrm{H}\\times \\mathrm{W}}^{\\mathrm{C}\\times \\mathrm{HW}}$ represents the operation of reshaping features from $\\mathbb{R}^{C\\times H\\times W}$ to $\\mathbb{R}^{C\\times HW}$ , $d$ is the index of sorted Value, and Gather denotes the operation of retrieving elements of tensor based on a given index.", + "bbox": [ + 212, + 279, + 785, + 329 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Then given the number of bins $B$ , we reshape the sorted features from $C \\times HW$ into $C \\times B \\times HW / B$ . To extract both global and local information, we define two types of reshaping, i.e., bin-wise histogram reshaping (BHR) and frequency-wise histogram reshaping (FHR). The first is to assign the number of bins equal to $B$ and each bin contains $HW / B$ elements, while the second is to set the frequency of each bin equal to $B$ and the number of bins is $HW / B$ . By this way, we can extract large-scale information by BHR where each bin contains large number of dynamically located pixels and fine-grained information by FHR where each bins contains modicum pixels neighboring in terms of intensity. The two pairs of Query-Key features are passed through two types of reshaping and subsequent self-attention process respectively, and their outputs are elementwisely multiplied to yield the final output. The process can be formulated as the following expressions:", + "bbox": [ + 212, + 333, + 787, + 529 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {A _ {B}} = \\mathrm {s o f t m a x} \\left(\\frac {\\mathbf {R _ {B}} (Q _ {1}) \\mathbf {R _ {B}} (K _ {1}) ^ {\\top}}{\\sqrt {k}}\\right) \\mathbf {R _ {B}} (V),\n$$\n", + "text_format": "latex", + "bbox": [ + 344, + 535, + 658, + 569 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {A} _ {\\mathrm {F}} = \\operatorname {s o f t m a x} \\left(\\frac {\\mathbf {R} _ {\\mathrm {F}} \\left(Q _ {2}\\right) \\mathbf {R} _ {\\mathrm {F}} \\left(K _ {2}\\right) ^ {\\top}}{\\sqrt {k}}\\right) \\mathbf {R} _ {\\mathrm {F}} (V), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 344, + 571, + 784, + 604 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {A} = \\mathrm {A} _ {\\mathrm {B}} \\odot \\mathrm {A} _ {\\mathrm {F}},\n$$\n", + "text_format": "latex", + "bbox": [ + 344, + 609, + 447, + 625 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $k$ is the number of heads, $\\mathbf{R}_{i\\in \\{\\mathrm{B},\\mathrm{F}\\}}$ denotes the reshaping operation of either BHR or FHR, and $\\mathrm{A}_{i\\in \\{\\mathrm{B},\\mathrm{F}\\}}$ represents the obtained attention map.", + "bbox": [ + 212, + 631, + 787, + 662 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Dual-scale Gated Feed-Forward Previous studies [11, 75, 78, 86] typically leverage single-range or single-scale convolution in the standard feed-forward network to bolster local context. Nonetheless, these methods often disregard the correlations among dynamically distributed weather-induced degradation. In practice, multi-scale information can be extracted by not only enlarging the kernel size but also leveraging the dilation mechanism [36, 84, 85]. As a result, we conceive a Dual-scale Gated Feed-Forward (DGFF) module, which integrates two distinct multi-range and multi-scale depth-wise convolution pathways within the transmission process.", + "bbox": [ + 212, + 703, + 787, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "S. Sun et al.", + "bbox": [ + 271, + 114, + 354, + 127 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/b0d9b129f2294599477371e5e8f549f72c8e1225ab44d8d0673e9ddbaa3e62c1.jpg", + "table_caption": [ + "Table 1: Quantitative comparisons on three weather removal tasks in terms of PSNR and SSIM, where higher values indicate better performance. The top halves of tables display the results of task-specific methods, while the bottom halves present evaluations of the unified multi-weather models. The best and the second best results are in bold and underlined. Those with * indicate the methods whose source codes are unavailable.", + "(a) Image Desnowing", + "(b) Deraining & Dehazing", + "(c) Raindrop Removal" + ], + "table_footnote": [], + "table_body": "
Snow100K-S [44]Snow100K-L [44]Outdoor-Rain [32]RainDrop [57]
PSNRSSIMPSNRSSIMPSNRSSIMPSNR SSIM
SPANet [73]29.920.826023.700.7930CycleGAN [98]17.620.6560pix2pix [20]28.02
JSTASR [8]31.400.901225.320.8076pix2pix [20]19.090.7100DuRN [42]31.24
RESCAN [34]31.510.903226.080.8108HRGAN [32]21.560.8550RaindropAttn [59]31.44
DesnowNet [44]32.330.950027.170.8983PCNet [21]26.190.9015AttentiveGAN [57]31.59
DDMSNet [94]34.340.944528.850.8772MPRNet [87]28.030.9192IDT [78]31.87
NAFNet [6]34.790.949730.060.9017NAFNet [6]29.590.9027MAXIM [69]31.87
Restormer [86]36.010.957930.360.9068Restormer [86]30.030.9215Restormer [86]32.18
All-in-One [33]*--28.330.8820All-in-One [33]*24.710.8980All-in-One [33]*31.12
TransWeather [70]32.510.934129.310.8879TransWeather [70]28.830.9000TransWeather [70]30.17
Chen et al. [10]34.420.946930.220.9071Chen et al. [10]29.270.9147Chen et al. [10]31.81
WGWSNet [100]34.310.946030.160.9007WGWSNet [100]29.320.9207WGWSNet [100]32.38
WeatherDiff64 [53]35.830.956630.090.9041WeatherDiff64 [53]29.640.9312WeatherDiff64 [53]30.71
WeatherDiff128 [53]35.020.951629.580.8941WeatherDiff128 [53]29.720.9216WeatherDiff128 [53]29.66
AWRCP [82]*36.920.965231.920.9341AWRCP [82]*31.390.9329AWRCP [82]*31.93
Histoformer (Ours)37.410.965632.160.9261Histoformer (Ours)32.080.9389Histoformer (Ours)33.06
", + "bbox": [ + 218, + 237, + 784, + 373 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Given an input tensor $F_{l} \\in \\mathbb{R}^{C \\times H \\times W}$ , we initially employ a point-wise convolution operation to augment the channel dimension by a factor of $r$ . Following this augmentation, the expanded tensor is directed into two parallel branches. Throughout the feature transformation process, $5 \\times 5$ and dilated $3 \\times 3$ depth-wise convolutions are employed to enhance the extraction of multi-range and multi-scale information. Following the gating mechanism [13], the output of the second branch after passing through an activation act as a gating map for the other branch. Thus, the complete feature fusion process within the DGFF module is formulated as follows:", + "bbox": [ + 212, + 401, + 787, + 537 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\nF _ {l, 1}, F _ {l, 2} = \\operatorname {S p l i t} \\left(\\operatorname {S h u f f l e} \\left(\\operatorname {C o n v} _ {1 \\times 1} \\left(F _ {l}\\right)\\right)\\right),\n$$\n", + "text_format": "latex", + "bbox": [ + 331, + 542, + 620, + 558 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\nF _ {l, 1} = \\operatorname {C o n v} _ {5 \\times 5} ^ {\\mathrm {d}} \\left(F _ {l, 1}\\right), F _ {l, 2} = \\operatorname {C o n v} _ {3 \\times 3} ^ {\\mathrm {d}, \\text {d i l a t e d}} \\left(F _ {l, 2}\\right), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 331, + 561, + 784, + 580 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\nF _ {l + 1} = \\operatorname {C o n v} _ {1 \\times 1} \\left(\\text {U n s h u r f l e} \\left(\\operatorname {M i s h} \\left(F _ {l, 2}\\right) \\odot F _ {l, 1}\\right)\\right),\n$$\n", + "text_format": "latex", + "bbox": [ + 331, + 583, + 671, + 599 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where $\\mathrm{Conv}_{5\\times 5}^{\\mathrm{d}}$ represents $5\\times 5$ depth-wise convolution, $\\mathrm{Conv}_{3\\times 3}^{\\mathrm{d,dilated}}$ is $3\\times 3$ dilated depth-wise convolution, Shuffle and Unshuffle represent respectively the operations of pixel-shuffling and unshuffling, Mish denotes the Mish activation [52], and $F_{l + 1}$ is the output of current stage passing to $l + 1$ -th stage.", + "bbox": [ + 212, + 606, + 787, + 669 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "3.3 Reconstruction Loss and Correlation Loss", + "text_level": 1, + "bbox": [ + 214, + 691, + 606, + 707 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We use the $L_{1}$ norm of the pixel-wise difference between the restored high-quality image $I^{hq}$ and ground-truth $I^{gt}$ as the reconstruction loss, i.e.,", + "bbox": [ + 212, + 719, + 785, + 750 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {r e c} = \\left\\| I ^ {h q} - I ^ {g t} \\right\\| _ {1}. \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 426, + 756, + 784, + 773 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Furthermore, we notice that the $\\mathcal{L}_{rec}$ only regulates the pixel-level similarity between the restored image and the ground-truth, while neglecting the patch-level linear correlations. The innate relationships of intensity within the image are disrupted by the consistent patterns of weather-induced degradation. By", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "Restoring Images in Adverse Weather Conditions via Histogram Transformer", + "bbox": [ + 218, + 114, + 732, + 130 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/baca57d25cae884c90257b422de3eadf3d5b76bd5cc7a3fbbfb41c8760dd389c.jpg", + "image_caption": [ + "Fig. 3: Visual comparison for desnowing on Snow100K [44]. The samples from (b) to (e) are Restormer [86], TransWeather [70], WGWSNet [100], WeatherDiff [53]." + ], + "image_footnote": [], + "bbox": [ + 217, + 143, + 785, + 224 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/208d7bedddefa74b11ee3cca8783fa72640b2f53b4813c1f8dc3168f5a3dc5da.jpg", + "image_caption": [ + "Fig. 4: Visual comparison for deraining and dehazing on Outdoor-Rain [32]. The samples from (b) to (e) are Restormer [86], TransWeather [70], WGWSNet [100], WeatherDiff [53]." + ], + "image_footnote": [], + "bbox": [ + 217, + 271, + 785, + 344 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "emulating the intensity relationships within the ground-truth, we compel the degraded pixels to occupy their original positions according to the original intensity ranking. Consequently, we introduce the Pearson correlation [12] between images as a means to regulate the linear relationship, expressed as follows:", + "bbox": [ + 212, + 420, + 782, + 481 + ], + "page_idx": 9 + }, + { + "type": "equation", + "text": "\n$$\n\\rho \\left(I ^ {h q}, I ^ {g t}\\right) = \\frac {\\sum_ {i = 1} ^ {3 H W} \\left(I _ {i} ^ {h q} - \\bar {I} ^ {h q}\\right) \\left(I _ {i} ^ {g t} - \\bar {I} ^ {g t}\\right)}{3 H W \\sigma \\left(I ^ {h q}\\right) \\sigma \\left(I ^ {g t}\\right)}, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 331, + 484, + 784, + 527 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "where $I_{i}^{\\{\\cdot\\}}$ represents the $i$ -th pixel of image, $\\overline{I}^{\\{\\cdot\\}}$ and $\\sigma(I^{\\{\\cdot\\}})$ denotes respectively the mean and the standard deviation of image sequence. Its value falls within the $[-1,1]$ range. When two images exhibit perfect correlation, the value of function $\\rho$ attains a value of 1, while in the case of negative correlation, its value reaches $-1$ . Hence, we formulate the correlation loss as follows:", + "bbox": [ + 212, + 535, + 782, + 613 + ], + "page_idx": 9 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {c o r} = \\frac {1}{2} \\left(1 - \\rho (I ^ {h q}, I ^ {g t})\\right), \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 400, + 619, + 784, + 648 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "such that $\\mathcal{L}_{cor} = 0$ when the recovered image perfectly aligns with the groundtruth. The overall loss function is thus defined as:", + "bbox": [ + 212, + 652, + 782, + 683 + ], + "page_idx": 9 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathcal {L} _ {\\text {r e c}} + \\alpha \\mathcal {L} _ {\\text {c o r}}, \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 433, + 691, + 784, + 708 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "where $\\alpha$ is the weight of correlation loss.", + "bbox": [ + 212, + 714, + 508, + 729 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 215, + 753, + 375, + 771 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.1 Experimental settings", + "text_level": 1, + "bbox": [ + 215, + 784, + 442, + 800 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Datasets. We train our model on the same datasets as the previous works [33,53, 70] to ensure a fair comparison. The training set encompasses 9,000 images drawn", + "bbox": [ + 212, + 809, + 782, + 839 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "S. Sun et al.", + "bbox": [ + 271, + 114, + 354, + 127 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/19d278cb330ff500c448ffe6ea4cb25b75a15428a02051e0c1816a0eb714e19c.jpg", + "image_caption": [ + "Fig. 5: Visual comparison for raindrop removal on RainDrop [57]. The samples from (b) to (e) are Chen et al. [10], TransWeather [70], WGWSNet [100], WeatherDiff [53]." + ], + "image_footnote": [], + "bbox": [ + 217, + 143, + 785, + 224 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/fac601967372a43f1453bd65a58b6f2b02d719857ead9be7b95f65e19102dbf9.jpg", + "image_caption": [ + "Fig.6: A qualitative comparison for real-world adverse weather removal on Snow100K [44]. The samples from (b) to (e) are Chen et al. [10], Restormer [86], TransWeather [70], WGWSNet [100], WeatherDiff [53]." + ], + "image_footnote": [], + "bbox": [ + 217, + 268, + 785, + 349 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "from Snow100K [44], 1,069 images sourced from Raindrop [57], and 9,000 images from Outdoor-Rain [32]. Snow100K contains synthetic images deteriorated by snow, while Raindrop comprises real raindrop-affected images. Outdoor-Rain features synthetic images afflicted by both fog and rain streaks. For evaluation, we employ the Test1 dataset [32, 33], the RainDrop test dataset [57], and the Snow100K-L and -S test sets [44]. Snow100K also provides a real-world test set containing 1,329 images affected by adverse weather.", + "bbox": [ + 212, + 421, + 784, + 527 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Comparison Baselines. We assess the performance of our method against state-of-the-art approaches designed specifically for distinct weather removal tasks: raindrop removal, snow removal, and rain&fog removal. Specifically, for snow removal, we benchmark against SPANet [73], JSTASR [8], RESCAN [34], DesnowNet [44], and DDMSNet [94]. In the case of rain&fog removal, we compare with CycleGAN [98], pix2pix [20], HRGAN [32], MPRNet [87] and Restormer [86]. For raindrop removal, we evaluate against the methods such as pix2pix [20], DuRN [42], RaindropAttn [59], AttentiveGAN [57]. Additionally, we include some recent transformer or multi-degradation restoration networks, IDT [78], NAFNet [6], MAXIM [69], and Restormer [86], in our comparative analysis. It is worth noting that all these methods are single-task networks fine-tuned for specific datasets.", + "bbox": [ + 212, + 540, + 784, + 719 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Furthermore, we conduct a performance comparison with the All-in-One network [33], Chen et al. [10], TransWeather [70], WGWS-Net [100], WeatherDiff [53] and AWRCP [82], which are trained to handle all the aforementioned tasks using a unified model. Note that our approach is also trained to tackle all these tasks using a single model.", + "bbox": [ + 212, + 720, + 784, + 797 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Training details. Our implementation is realized by PyTorch [55] and on NVIDIA Tesla V100 GPU. The network is trained for a total of 300,000 iterations, with", + "bbox": [ + 215, + 809, + 784, + 839 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "Restoring Images in Adverse Weather Conditions via Histogram Transformer", + "bbox": [ + 218, + 114, + 730, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 116, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "an initial batch size of 8 and an initial patch size of 128 akin to the progressive learning pipeline [86]. We employ the AdamW optimizer [48] with an initial learning rate of $3e^{-4}$ for the first 92,000 iterations, which is gradually reduced to $1e^{-6}$ using cosine annealing schedule [47] during the remaining 208,000 iterations. The number of blocks at each stage $L_{i\\in \\{1,2,3,4\\}}$ is set to $\\{4,4,6,8\\}$ and the channel size $C$ is 36. The channel expansion factor $r$ in DGFF is set to 2.667. The numbers of heads in self-attention at different stages are set to $\\{1,2,4,8\\}$ respectively. We randomly apply horizontal and vertical flips as the technique of data augmentation.", + "bbox": [ + 212, + 146, + 787, + 282 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "4.2 Comparisons with the state-of-the-arts", + "text_level": 1, + "bbox": [ + 215, + 304, + 581, + 319 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Quantitative Evaluation. In our study, we provide a comprehensive comparative analysis of metrics applied to both synthetic and real datasets, as summarized in Table 1. For a fair and well-founded comparison, we utilize recent multiple degradation removal methods such as MPRNet [87], MAXIM [69], and Restormer [86], treating them as weather-specific approaches for each benchmark. Additionally, we retrain the all-in-one adverse weather removal methods including Chen et al. [10] and WGWS-Net [100] using the all-weather training dataset [33,53,70]. This exhaustive comparison reveals that our proposed method exhibits a significant performance advantage over existing approaches across three different types of degradation.", + "bbox": [ + 212, + 328, + 787, + 479 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Qualitative Evaluation. Furthermore, we conduct a visual comparison on three tasks, and the outcomes are showcased in Figure 3, 4 and 5 respectively. Figure 6 shows a case of real-world weather removal. These results highlight that our method excels in comprehensively eliminating snow degradation, including fine and large snow spots. In contrast, the recent WeatherDiff [53] method still exhibits some residual snow degradation, and its capability to restore details is not optimal. When it comes to the restoration of challenging weather conditions, our method excels in removing complex haze and rain marks, yielding visually appealing results in comparison to prior approaches.", + "bbox": [ + 212, + 493, + 787, + 630 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "4.3 Ablation studies", + "text_level": 1, + "bbox": [ + 215, + 650, + 398, + 664 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "To substantiate the effectiveness of each component within Histoformer, we conduct a sequence of ablation studies on Outdoor-Rain [32]. In particular, we examine the impact of the dynamic-range convolution, the DHSA module, the number of bins in DHSA, the DGFF module, and the correlation loss.", + "bbox": [ + 212, + 674, + 787, + 736 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Dynamic-range Convolution. We experiment on two settings of dynamic-range convolution, namely, sorting horizontally first and then vertically before convolution, and the reverse order. Additionally, we compared them with vanilla convolution, and the results are displayed in Table 2. The operations of regular sorting led to a performance improvement of $0.14\\mathrm{dB}$ , and the order of sorting operations does not significantly affect the outcome.", + "bbox": [ + 212, + 750, + 787, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "S. Sun et al.", + "bbox": [ + 271, + 114, + 354, + 126 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/901978b07e6fcfc47c502e1d767f26601c40d3e9a09e47c51f005cd8961508ef.jpg", + "table_caption": [ + "Table 2: Ablation studies on the dynamic-range convolution." + ], + "table_footnote": [], + "table_body": "
Vanilla ConvVerti.+Horiz.Horiz.+Verti.
PSNR31.9432.0332.08
SSIM0.93770.93850.9389
", + "bbox": [ + 330, + 165, + 671, + 213 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/437adf6514946a94cbf7aedb8f85ba09b4f7f55255a5c01573aeff1e4e6eb637.jpg", + "table_caption": [ + "Table 3: Ablation studies on the design of self-attention" + ], + "table_footnote": [], + "table_body": "
SA typePSNRSSIM
MDTA [86]30.940.9278
TKSA [11]31.120.9295
w/o BHR31.050.9301
w/o FHR31.790.9364
DHSA32.080.9389
", + "bbox": [ + 256, + 253, + 455, + 330 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/9572923e3b27ba94113abad66f030775203270612b6161b65870418abd0439f4.jpg", + "table_caption": [ + "Table 4: Ablation studies on the number of Bins in DHSA" + ], + "table_footnote": [], + "table_body": "
#Bins in DHSAPSNRSSIM
1230.430.9283
2031.560.9312
2831.940.9379
3632.080.9389
44Out of memory
", + "bbox": [ + 532, + 253, + 759, + 330 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/69e6ded3c1ec5fe47d7bafcb610c4613b69b70113c9f1ce43ce5dd79788375ea.jpg", + "table_caption": [ + "Table 5: Ablation studies on the choice of feed-forward module" + ], + "table_footnote": [], + "table_body": "
Feed-ForwardPSNRSSIM
FFN [38]31.320.9313
GDFN [86]31.420.9347
DANB [96]31.560.9351
MSFN [11]31.780.9367
DGFF32.080.9389
", + "bbox": [ + 250, + 373, + 462, + 450 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/a75c41559691b8a9353af4169f5b79c0ff0250f7c821463a375bd07dac87c33f.jpg", + "table_caption": [ + "Table 6: Ablation studies on the setting of correlation loss" + ], + "table_footnote": [], + "table_body": "
αPSNRSSIM
w/o Lcor031.770.9358
0.132.010.9369
132.080.9389
w/ Lcor532.030.9392
1031.960.9375
", + "bbox": [ + 532, + 373, + 759, + 450 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "DHSA. To evaluate the effectiveness of the proposed DHSA module, we conduct a comparison with two baselines, i.e., a multi-Dconv head transposed attention (MDTA) [86] and a top-k sparse attention (TKSA) [11]. Additionally, we explore two additional settings of DHSA by excluding either the BHR branch or the FHR branch. The quantitative analysis results are presented in Table 3.", + "bbox": [ + 212, + 478, + 784, + 553 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Both MDTA and TKSA integrate rich information across channels, which may result in a loss of the exploitation of long-range information across spatial dimensions. While our histogram self-attention can capture spatially long-range information, the use of either a single BHR or a single FHR branch neglects the inter-bin or inner-bin relationships, leading to inferior results. By incorporating dynamic-range convolution and dual-branch histogram self-attention, capable of extracting long-range spatial features, our DHSA enhances performance, resulting in a PSNR improvement of 0.96 dB compared to TKSA.", + "bbox": [ + 212, + 554, + 787, + 675 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Number of Bins. To assess the influence of the number of bins in DHSA, we conduct experiments with five different values: 12, 20, 28, 36, and 44. The results are presented in Table 4. It is observed that increasing the number of bins consistently improves performance. However, when the number of bins exceeds 44, it results in an out-of-memory error.", + "bbox": [ + 212, + 689, + 787, + 765 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "DGFF. To assess the effectiveness of the proposed DGFF module, we conduct a comparison with four baselines: (i) the vanilla feed-forward network (FN) [38], (ii) a gated-Dconv feed-forward network (GDFN) [86], (iii) a dual adaptive neural block (DANB) [96], and (iv) a mixed-scale feed-forward network (MSFN) [11].", + "bbox": [ + 214, + 779, + 787, + 842 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "Restoring Images in Adverse Weather Conditions via Histogram Transformer", + "bbox": [ + 218, + 114, + 730, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 784, + 126 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/a5e963f75412604f02d4b03a668b66117c243845b7a05706c6e8ea9c7ac18246.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 217, + 143, + 383, + 231 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/acecbfef62a99b4e6ece62c03d9d41d1ea40aca6c7c8e7eaaa87e6bd58c4bcef.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 388, + 146, + 496, + 227 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/d461570eb7879ef685295620460bbad6f566c6cc5ce108f2c13b0c032e790079.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 503, + 143, + 668, + 231 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/df5e72bed4f1e2a0ffc1f28f998be7c68d69cb742c8e4d1d289f648c0fb01e12.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 673, + 146, + 784, + 228 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/3e232858f732d101478c5eb8ab9cf76a84c566659b7fdfadc060e2d6c7240059.jpg", + "image_caption": [ + "(a) Input" + ], + "image_footnote": [], + "bbox": [ + 217, + 232, + 383, + 319 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/f8a8d3b27a0bc4377b8df0948e2988b05bd55feff38e7802fe57bbbf66e24ac0.jpg", + "image_caption": [ + "Fig. 7: Real-world deweathering on two snowy images [44] and their downstream detection results on Google API." + ], + "image_footnote": [], + "bbox": [ + 388, + 232, + 496, + 313 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/54d9cb90a8447db778f7a100b008e5ae1581e9b0278e99464911f06ea7a8afd0.jpg", + "image_caption": [ + "(b) Deweathered by ours" + ], + "image_footnote": [], + "bbox": [ + 503, + 232, + 668, + 319 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/84a07aeea60f3ad7e74ad0bf2ab70afaa89399fc0746fbbf5d91860ebc41c871.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 674, + 233, + 784, + 314 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "The quantitative analysis results are presented in Table 5. While MSFN integrates mixed-scale information, it may still miss out on the exploitation of multi-range spatial knowledge. Through the inclusion of pixel-shuffling and feature aggregation across different ranges, our DGFF further enhances performance, resulting in a PSNR gain of 0.3 dB over MSFN.", + "bbox": [ + 214, + 390, + 784, + 465 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Correlation Loss. Table 6 shows the effectiveness of the correlation loss $\\mathcal{L}_{cor}$ and the influence of its weight. It is evident that $\\mathcal{L}_{cor}$ consistently improves the performance, while the specific loss weight does not have a substantial impact on the final results. We therefore keep the loss weight as 1 by default.", + "bbox": [ + 214, + 478, + 784, + 539 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "4.4 Real-world Application", + "text_level": 1, + "bbox": [ + 215, + 556, + 452, + 573 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "To further demonstrate the practical applicability of our method for real-world adverse weather removal and its potential to improve downstream detection task, we provide two samples in Figure 7. As depicted, our Histoformer effectively eliminates snowflakes from the scene and assists the detector in recognizing omitted door and building.", + "bbox": [ + 214, + 579, + 784, + 654 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 675, + 357, + 690 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In this research, we introduce a novel mechanism called histogram self-attention and devise a new histogram transformer named Histoformer to tackle the challenge of all-in-one weather removal. Our histogram self-attention involves segmenting spatial features into multiple bins, and allocating varying attention along the bin or frequency dimension, allowing it to selectively focus on weather-related features with a dynamic range. To facilitate learning both multi-range and multi-scale information, we present DGFF module and a correlation loss. Through extensive experimentation, we demonstrate the effectiveness and superiority of our approach.", + "bbox": [ + 214, + 704, + 782, + 839 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "S. Sun et al.", + "bbox": [ + 271, + 114, + 354, + 126 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgement", + "text_level": 1, + "bbox": [ + 217, + 143, + 392, + 162 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "This work has been supported in part by National Natural Science Foundation of China (No. 62322216, 62172409, 62025604, 62306308, 62311530686), in part by Shenzhen Science and Technology Program (Grant No. JCYJ20220818102012025, KQTD20221101093559018), and in part by Guangdong Provincial Key Laboratory of Information Security Technology (No. 2023B1212060026).", + "bbox": [ + 212, + 176, + 787, + 253 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 277, + 321, + 292 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Ancuti, C.O., Ancuti, C.: Single image dehazing by multi-scale fusion. IEEE TIP (2013)", + "2. Berman, D., Avidan, S., et al.: Non-local image dehazing. In: CVPR (2016)", + "3. Carion, N., Massa, F., Synnaeve, G., Usunier, N., Kirillov, A., Zagoruyko, S.: End-to-end object detection with transformers. In: ECCV (2020)", + "4. Chen, H., Wang, Y., Guo, T., Xu, C., Deng, Y., Liu, Z., Ma, S., Xu, C., Xu, C., Gao, W.: Pre-trained image processing transformer. In: CVPR (2021)", + "5. Chen, J., Tan, C.H., Hou, J., Chau, L.P., Li, H.: Robust video content alignment and compensation for rain removal in a cnn framework. In: CVPR (2018)", + "6. Chen, L., Chu, X., Zhang, X., Sun, J.: Simple baselines for image restoration. In: ECCV (2022)", + "7. Chen, S., Ye, T., Liu, Y., Chen, E., Shi, J., Zhou, J.: Snowformer: Scale-aware transformer via context interaction for single image desnowing. arXiv preprint arXiv:2208.09703 (2022)", + "8. Chen, W.T., Fang, H.Y., Ding, J.J., Tsai, C.C., Kuo, S.Y.: Jstasr: Joint size and transparency-aware snow removal algorithm based on modified partial convolution and veiling effect removal. In: Vedaldi, A., Bischof, H., Brox, T., Frahm, J.M. (eds.) ECCV (2020)", + "9. Chen, W.T., Fang, H.Y., Hsieh, C.L., Tsai, C.C., Chen, I., Ding, J.J., Kuo, S.Y., et al.: All snow removed: Single image desnowing algorithm using hierarchical dual-tree complex wavelet representation and contradict channel loss. In: ICCV (2021)", + "10. Chen, W.T., Huang, Z.K., Tsai, C.C., Yang, H.H., Ding, J.J., Kuo, S.Y.: Learning multiple adverse weather removal via two-stage knowledge learning and multi-contrastive regularization: Toward a unified model. In: CVPR (2022)", + "11. Chen, X., Li, H., Li, M., Pan, J.: Learning a sparse transformer network for effective image deraining. In: CVPR (2023)", + "12. Cohen, I., Huang, Y., Chen, J., Benesty, J., Benesty, J., Chen, J., Huang, Y., Cohen, I.: Pearson correlation coefficient. Noise reduction in speech processing (2009)", + "13. Dauphin, Y.N., Fan, A., Auli, M., Grangier, D.: Language modeling with gated convolutional networks. In: ICML (2017)", + "14. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929 (2020)", + "15. Eigen, D., Krishnan, D., Fergus, R.: Restoring an image taken through a window covered with dirt or rain. In: ICCV (2013)" + ], + "bbox": [ + 225, + 309, + 784, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "Restoring Images in Adverse Weather Conditions via Histogram Transformer", + "bbox": [ + 218, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "16. Fu, H., Gong, M., Wang, C., Batmanghelich, K., Tao, D.: Deep ordinal regression network for monocular depth estimation. In: CVPR (2018)", + "17. Fu, X., Huang, J., Zeng, D., Huang, Y., Ding, X., Paisley, J.: Removing rain from single images via a deep detail network. In: CVPR (2017)", + "18. Godard, C., Mac Aodha, O., Brostow, G.J.: Unsupervised monocular depth estimation with left-right consistency. In: CVPR (2017)", + "19. Guo, C., Yan, Q., Anwar, S., Cong, R., Ren, W., Li, C.: Image dehazing transformer with transmission-aware 3d position embedding. In: CVPR (2022)", + "20. Isola, P., Zhu, J.Y., Zhou, T., Efros, A.A.: Image-to-image translation with conditional adversarial networks. In: CVPR (2017)", + "21. Jiang, K., Wang, Z., Yi, P., Chen, C., Wang, Z., Wang, X., Jiang, J., Lin, C.W.: Rain-free and residue hand-in-hand: A progressive coupled network for real-time image deraining. IEEE TIP (2021)", + "22. Jiang, T.X., Huang, T.Z., Zhao, X.L., Deng, L.J., Wang, Y.: A novel tensor-based video rain streaks removal approach via utilizing discriminatively intrinsic priors. In: CVPR (2017)", + "23. Kang, L.W., Lin, C.W., Fu, Y.H.: Automatic single-image-based rain streaks removal via image decomposition. IEEE TIP (2011)", + "24. Kim, Y., Cho, Y., Nguyen, T.T., Lee, D.: Metaweather: Few-shot weather-degraded image restoration via degradation pattern matching. arXiv preprint arXiv:2308.14334 (2023)", + "25. Lai, Z., Wu, J., Chen, S., Zhou, Y., Hovakimyan, N.: Residual-based language models are free boosters for biomedical imaging tasks. In: CVPRW (2024)", + "26. Li, B., Peng, X., Wang, Z., Xu, J., Feng, D.: Aod-net: All-in-one dehazing network. In: ICCV (2017)", + "27. Li, B., Ren, W., Fu, D., Tao, D., Feng, D., Zeng, W., Wang, Z.: Benchmarking single-image dehazing and beyond. IEEE TIP (2018)", + "28. Li, B., Liu, X., Hu, P., Wu, Z., Lv, J., Peng, X.: All-in-one image restoration for unknown corruption. In: CVPR (2022)", + "29. Li, L., Dong, Y., Ren, W., Pan, J., Gao, C., Sang, N., Yang, M.H.: Semi-supervised image dehazing. IEEE TIP (2019)", + "30. Li, M., Cao, X., Zhao, Q., Zhang, L., Meng, D.: Online rain/snow removal from surveillance videos. IEEE TIP (2021)", + "31. Li, P., Yun, M., Tian, J., Tang, Y., Wang, G., Wu, C.: Stacked dense networks for single-image snow removal. Neurocomputing (2019)", + "32. Li, R., Cheong, L.F., Tan, R.T.: Heavy rain image restoration: Integrating physics model and conditional adversarial learning. In: CVPR (2019)", + "33. Li, R., Tan, R.T., Cheong, L.F.: All in one bad weather removal using architectural search. In: CVPR (2020)", + "34. Li, X., Wu, J., Lin, Z., Liu, H., Zha, H.: Recurrent squeeze-and-excitation context aggregation net for single image deraining. In: ECCV (2018)", + "35. Li, Y., Tan, R.T., Guo, X., Lu, J., Brown, M.S.: Rain streak removal using layer priors. In: CVPR (2016)", + "36. Li, Y., Lu, J., Chen, H., Wu, X., Chen, X.: Dilated convolutional transformer for high-quality image deraining. In: CVPRW (June 2023)", + "37. Li, Z., Guan, B., Wei, Y., Zhou, Y., Zhang, J., Xu, J.: Mapping new realities: Ground truth image creation with pix2pix image-to-image translation. arXiv preprint arXiv:2404.19265 (2024)", + "38. Liang, J., Cao, J., Sun, G., Zhang, K., Van Gool, L., Timofte, R.: Swinir: Image restoration using swin transformer. In: ICCV (2021)" + ], + "bbox": [ + 225, + 146, + 784, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "S. Sun et al.", + "bbox": [ + 271, + 114, + 354, + 127 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "39. Liang, Y., Anwar, S., Liu, Y.: Drt: A lightweight single image deraining recursive transformer. In: CVPR (2022)", + "40. Liu, J., Yang, W., Yang, S., Guo, Z.: Erase or fill? deep joint recurrent rain removal and reconstruction in videos. In: CVPR (2018)", + "41. Liu, K., Jiang, Y., Choi, I., Gu, J.: Learning image-adaptive codebooks for class-agnostic image restoration. arXiv preprint arXiv:2306.06513 (2023)", + "42. Liu, X., Suganuma, M., Sun, Z., Okatani, T.: Dual residual networks leveraging the potential of paired operations for image restoration. In: CVPR (2019)", + "43. Liu, Y., Liu, H., Li, L., Wu, Z., Chen, J.: A data-centric solution to nonhomogeneous dehazing via vision transformer. In: CVPR (2023)", + "44. Liu, Y.F., Jaw, D.W., Huang, S.C., Hwang, J.N.: Desnownet: Context-aware deep network for snow removal. IEEE TIP (2018)", + "45. Liu, Z., Lin, Y., Cao, Y., Hu, H., Wei, Y., Zhang, Z., Lin, S., Guo, B.: Swin transformer: Hierarchical vision transformer using shifted windows. In: ICCV (2021)", + "46. Liu, Z., Ning, J., Cao, Y., Wei, Y., Zhang, Z., Lin, S., Hu, H.: Video swim transformer. In: CVPR (2022)", + "47. Loshchilov, I., Hutter, F.: Sgdr: Stochastic gradient descent with warm restarts. arXiv preprint arXiv:1608.03983 (2016)", + "48. Loshchilov, I., Hutter, F.: Fixing weight decay regularization in adam. arXiv preprint arXiv:1711.05101 (2017)", + "49. Luo, Y., Zhao, R., Wei, X., Chen, J., Lu, Y., Xie, S., Wang, T., Xiong, R., Lu, M., Zhang, S.: Mowe: Mixture of weather experts for multiple adverse weather removal. arXiv preprint arXiv:2303.13739 (2023)", + "50. Lyu, W., Zheng, S., Ling, H., Chen, C.: Backdoor attacks against transformers with attention enhancement. In: ICLR Workshop (2023)", + "51. Ma, H., Zeng, D., Liu, Y.: Learning individualized treatment rules with many treatments: A supervised clustering approach using adaptive fusion. NeurIPS (2022)", + "52. Misra, D.: Mish: A self regularized non-monotonic activation function. arXiv preprint arXiv:1908.08681 (2019)", + "53. Ozdenizci, O., Legenstein, R.: Restoring vision in adverse weather conditions with patch-based denoising diffusion models. IEEE TPAMI (2023)", + "54. Park, D., Lee, B.H., Chun, S.Y.: All-in-one image restoration for unknown degradations using adaptive discriminative filters for specific degradations. In: CVPR (2023)", + "55. Paszke, A., Gross, S., Massa, F., Lerer, A., Bradbury, J., Chanan, G., Killeen, T., Lin, Z., Gimelshein, N., Antiga, L., et al.: Pytorch: An imperative style, high-performance deep learning library. NeurIPS (2019)", + "56. Patil, P.W., Gupta, S., Rana, S., Venkatesh, S., Murala, S.: Multi-weather image restoration via domain translation. In: ICCV (2023)", + "57. Qian, R., Tan, R.T., Yang, W., Su, J., Liu, J.: Attentive generative adversarial network for raindrop removal from a single image. In: CVPR (2018)", + "58. Quan, R., Yu, X., Liang, Y., Yang, Y.: Removing raindrops and rain streaks in one go. In: CVPR (2021)", + "59. Quan, Y., Deng, S., Chen, Y., Ji, H.: Deep learning for seeing through window with raindrops. In: ICCV (2019)", + "60. Ren, S., He, K., Girshick, R., Sun, J.: Faster r-cnn: Towards real-time object detection with region proposal networks. NeurIPS (2015)", + "61. Ren, W., Tian, J., Han, Z., Chan, A., Tang, Y.: Video desnowing and deraining based on matrix decomposition. In: CVPR (2017)" + ], + "bbox": [ + 225, + 146, + 784, + 839 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "Restoring Images in Adverse Weather Conditions via Histogram Transformer", + "bbox": [ + 218, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "62. Ren, W., Liu, S., Zhang, H., Pan, J., Cao, X., Yang, M.H.: Single image dehazing via multi-scale convolutional neural networks. In: ECCV (2016)", + "63. Ren, W., Ma, L., Zhang, J., Pan, J., Cao, X., Liu, W., Yang, M.H.: Gated fusion network for single image dehazing. In: CVPR (2018)", + "64. Shao, Y., Li, L., Ren, W., Gao, C., Sang, N.: Domain adaptation for image dehazing. In: CVPR (2020)", + "65. Song, Y., He, Z., Qian, H., Du, X.: Vision transformers for single image dehazing. IEEE TIP (2023)", + "66. Sun, S., Ren, W., Li, J., Zhang, K., Liang, M., Cao, X.: Event-aware video deraining via multi-patch progressive learning. IEEE TIP (2023)", + "67. Sun, S., Ren, W., Wang, T., Cao, X.: Rethinking image restoration for object detection. NeurIPS (2022)", + "68. Tan, Z., Wu, Y., Liu, Q., Chu, Q., Lu, L., Ye, J., Yu, N.: Exploring the application of large-scale pre-trained models on adverse weather removal. arXiv preprint arXiv:2306.09008 (2023)", + "69. Tu, Z., Talebi, H., Zhang, H., Yang, F., Milanfar, P., Bovik, A., Li, Y.: Maxim: Multi-axis mlp for image processing. In: CVPR (2022)", + "70. Valanarasu, J.M.J., Yasarla, R., Patel, V.M.: Transweather: Transformer-based restoration of images degraded by adverse weather conditions. In: CVPR (2022)", + "71. Wan, Y.C., Shao, M.W., Cheng, Y.S., Liu, Y.X., Bao, Z.Y., Meng, D.Y.: Restoring images captured in arbitrary hybrid adverse weather conditions in one go. arXiv preprint arXiv:2305.09996 (2023)", + "72. Wang, T., Zhang, K., Shao, Z., Luo, W., Stenger, B., Lu, T., Kim, T.K., Liu, W., Li, H.: Gridformer: Residual dense transformer with grid structure for image restoration in adverse weather conditions. arXiv preprint arXiv:2305.17863 (2023)", + "73. Wang, T., Yang, X., Xu, K., Chen, S., Zhang, Q., Lau, R.W.: Spatial attentive single-image deraining with a high quality real rain dataset. In: CVPR (2019)", + "74. Wang, Y., Ma, C., Liu, J.: Smartassign: Learning a smart knowledge assignment strategy for deraining and desnowing. In: CVPR (2023)", + "75. Wang, Z., Cun, X., Bao, J., Zhou, W., Liu, J., Li, H.: Uformer: A general u-shaped transformer for image restoration. In: CVPR (2022)", + "76. Wei, W., Yi, L., Xie, Q., Zhao, Q., Meng, D., Xu, Z.: Should we encode rain streaks in video as deterministic or stochastic? In: ICCV (2017)", + "77. Wu, H., Qu, Y., Lin, S., Zhou, J., Qiao, R., Zhang, Z., Xie, Y., Ma, L.: Contrastive learning for compact single image dehazing. In: CVPR (2021)", + "78. Xiao, J., Fu, X., Liu, A., Wu, F., Zha, Z.J.: Image de-raining transformer. IEEE TPAMI (2022)", + "79. Yang, W., Tan, R.T., Feng, J., Liu, J., Guo, Z., Yan, S.: Deep joint rain detection and removal from a single image. In: CVPR (2017)", + "80. Yasarla, R., Patel, V.M.: Uncertainty guided multi-scale residual learning-using a cycle spinning cnn for single image de-raining. In: CVPR (2019)", + "81. Yasarla, R., Sindagi, V.A., Patel, V.M.: Syn2real transfer learning for image deraining using gaussian processes. In: CVPR (2020)", + "82. Ye, T., Chen, S., Bai, J., Shi, J., Xue, C., Jiang, J., Yin, J., Chen, E., Liu, Y.: Adverse weather removal with codebook priors. In: ICCV (2023)", + "83. You, S., Tan, R.T., Kawakami, R., Mukaigawa, Y., Ikeuchi, K.: Adherent raindrop modeling, detection and removal in video. IEEE TPAMI (2015)", + "84. Yu, F., Koltun, V.: Multi-scale context aggregation by dilated convolutions. arXiv preprint arXiv:1511.07122 (2016)", + "85. Yu, F., Koltun, V., Funkhouser, T.: Dilated residual networks. In: CVPR (2017)" + ], + "bbox": [ + 223, + 146, + 784, + 839 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "S. Sun et al.", + "bbox": [ + 271, + 114, + 354, + 127 + ], + "page_idx": 17 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "86. Zamir, S.W., Arora, A., Khan, S., Hayat, M., Khan, F.S., Yang, M.H.: Restormer: Efficient transformer for high-resolution image restoration. In: CVPR (2022)", + "87. Zamir, S.W., Arora, A., Khan, S., Hayat, M., Khan, F.S., Yang, M.H., Shao, L.: Multi-stage progressive image restoration. In: CVPR (2021)", + "88. Zhang, H., Patel, V.M.: Density-aware single image de-raining using a multi-stream dense network. In: CVPR (2018)", + "89. Zhang, H., Sindagi, V., Patel, V.M.: Image de-raining using a conditional generative adversarial network. IEEE TCSVT (2019)", + "90. Zhang, H., Sindagi, V., Patel, V.M.: Joint transmission map estimation and de-hazing using deep networks. IEEE TCSVT (2019)", + "91. Zhang, H., Ba, Y., Yang, E., Mehra, V., Gella, B., Suzuki, A., Pfahnl, A., Chandrappa, C.C., Wong, A., Kadambi, A.: Weatherstream: Light transport automation of single image deweathering. In: CVPR (2023)", + "92. Zhang, J., Ren, W., Zhang, S., Zhang, H., Nie, Y., Xue, Z., Cao, X.: Hierarchical density-aware dehazing network. IEEE Transactions on Cybernetics (2021)", + "93. Zhang, K., Li, D., Luo, W., Ren, W.: Dual attention-in-attention model for joint rain streak and raindrop removal. In: IEEE TIP (2021)", + "94. Zhang, K., Li, R., Yu, Y., Luo, W., Li, C.: Deep dense multi-scale network for snow removal using semantic and depth priors. IEEE TIP (2021)", + "95. Zhang, X., Li, H., Qi, Y., Leow, W.K., Ng, T.K.: Rain removal in video by combining temporal and chromatic properties. In: ICME (2006)", + "96. Zhao, H., Gou, Y., Li, B., Peng, D., Lv, J., Peng, X.: Comprehensive and delicate: An efficient transformer for image restoration. In: CVPR (2023)", + "97. Zheng, Y., Zhan, J., He, S., Dong, J., Du, Y.: Curricular contrastive regularization for physics-aware single image dehazing. In: CVPR (2023)", + "98. Zhu, J.Y., Park, T., Isola, P., Efros, A.A.: Unpaired image-to-image translation using cycle-consistent adversarial networks. In: ICCV (2017)", + "99. Zhu, L., Fu, C.W., Lischinski, D., Heng, P.A.: Joint bi-layer optimization for single-image rain streak removal. In: ICCV (2017)", + "00. Zhu, Y., Wang, T., Fu, X., Yang, X., Guo, X., Dai, J., Qiao, Y., Hu, X.: Learning weather-general and weather-specific features for image restoration under multiple adverse weather conditions. In: CVPR (2023)" + ], + "bbox": [ + 225, + 146, + 784, + 590 + ], + "page_idx": 18 + }, + { + "type": "header", + "text": "Restoring Images in Adverse Weather Conditions via Histogram Transformer", + "bbox": [ + 218, + 114, + 730, + 128 + ], + "page_idx": 18 + }, + { + "type": "page_number", + "text": "19", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 18 + } +] \ No newline at end of file diff --git a/2024/Restoring Images in Adverse Weather Conditions via Histogram Transformer/95e1934a-8ade-4d7d-8947-f969acbf3910_model.json b/2024/Restoring Images in Adverse Weather Conditions via Histogram Transformer/95e1934a-8ade-4d7d-8947-f969acbf3910_model.json new file mode 100644 index 0000000000000000000000000000000000000000..78332e24ea0afd5f779f6322c7551c616f86a549 --- /dev/null +++ b/2024/Restoring Images in Adverse Weather Conditions via Histogram Transformer/95e1934a-8ade-4d7d-8947-f969acbf3910_model.json @@ -0,0 +1,3252 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.22, + 0.141, + 0.784, + 0.187 + ], + "angle": 0, + "content": "Restoring Images in Adverse Weather Conditions via Histogram Transformer" + }, + { + "type": "text", + "bbox": [ + 0.235, + 0.213, + 0.768, + 0.245 + ], + "angle": 0, + "content": "Shangquan Sun\\(^{1,2}\\), Wenqi Ren\\(^{3,4\\dagger}\\), Xinwei Gao\\(^{5}\\), Rui Wang\\(^{1,2}\\), and Xiaochun Cao\\(^{3}\\)" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.256, + 0.784, + 0.283 + ], + "angle": 0, + "content": "\\(^{1}\\) Institute of Information Engineering, Chinese Academy of Sciences, Beijing 100085, China" + }, + { + "type": "text", + "bbox": [ + 0.238, + 0.284, + 0.765, + 0.311 + ], + "angle": 0, + "content": "\\(^{2}\\) School of Cyber Security, University of Chinese Academy of Sciences, Beijing 100049, China" + }, + { + "type": "text", + "bbox": [ + 0.242, + 0.312, + 0.758, + 0.339 + ], + "angle": 0, + "content": "3 School of Cyber Science and Technology, Shenzhen Campus of Sun Yat-sen University, Shenzhen 518107, China" + }, + { + "type": "text", + "bbox": [ + 0.242, + 0.34, + 0.758, + 0.367 + ], + "angle": 0, + "content": "4 Guangdong Provincial Key Laboratory of Information Security Technology, Guangzhou 510006, China" + }, + { + "type": "text", + "bbox": [ + 0.254, + 0.368, + 0.749, + 0.396 + ], + "angle": 0, + "content": "5 Wechat Business Group, Tencent, Shenzhen, Guangdong, China, 518057 shangquansun@gmail.com, renwq3@mail.sysu.edu.cn," + }, + { + "type": "list", + "bbox": [ + 0.238, + 0.284, + 0.765, + 0.396 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.446, + 0.744, + 0.737 + ], + "angle": 0, + "content": "Abstract. Transformer-based image restoration methods in adverse weather have achieved significant progress. Most of them use self-attention along the channel dimension or within spatially fixed-range blocks to reduce computational load. However, such a compromise results in limitations in capturing long-range spatial features. Inspired by the observation that the weather-induced degradation factors mainly cause similar occlusion and brightness, in this work, we propose an efficient Histogram Transformer (Histoformer) for restoring images affected by adverse weather. It is powered by a mechanism dubbed histogram self-attention, which sorts and segments spatial features into intensity-based bins. Self-attention is then applied across bins or within each bin to selectively focus on spatial features of dynamic range and process similar degraded pixels of the long range together. To boost histogram self-attention, we present a dynamic-range convolution enabling conventional convolution to conduct operation over similar pixels rather than neighbor pixels. We also observe that the common pixel-wise losses neglect linear association and correlation between output and ground-truth. Thus, we propose to leverage the Pearson correlation coefficient as a loss function to enforce the recovered pixels following the identical order as ground-truth. Extensive experiments demonstrate the efficacy and superiority of our proposed method. We have released the codes in Github." + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.753, + 0.751, + 0.781 + ], + "angle": 0, + "content": "Keywords: Image restoration \\(\\cdot\\) Adverse weather removal \\(\\cdot\\) Image Desnowing \\(\\cdot\\) Image deraining \\(\\cdot\\) Image dehazing \\(\\cdot\\) Raindrop removal" + }, + { + "type": "page_footnote", + "bbox": [ + 0.232, + 0.826, + 0.398, + 0.841 + ], + "angle": 0, + "content": "† Corresponding author." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.356, + 0.127 + ], + "angle": 0, + "content": "S. Sun et al." + }, + { + "type": "image", + "bbox": [ + 0.3, + 0.143, + 0.366, + 0.194 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.369, + 0.144, + 0.435, + 0.194 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.439, + 0.144, + 0.505, + 0.194 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.352, + 0.196, + 0.455, + 0.208 + ], + "angle": 0, + "content": "(a) Input patches" + }, + { + "type": "image", + "bbox": [ + 0.309, + 0.21, + 0.396, + 0.283 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.326, + 0.286, + 0.48, + 0.299 + ], + "angle": 0, + "content": "(b) Existing self-attention" + }, + { + "type": "image", + "bbox": [ + 0.411, + 0.21, + 0.498, + 0.283 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.147, + 0.694, + 0.283 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.524, + 0.286, + 0.687, + 0.299 + ], + "angle": 0, + "content": "(c) Histogram self-attention" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.304, + 0.788, + 0.388 + ], + "angle": 0, + "content": "Fig. 1: Given weather-degraded images in (a), traditional transformers perform self-attention either along the channel dimension or within a fixed-range block as shown in (b). In contrast, we observe that weather-induced degradation patterns tend to be similar but distinct from the background. So we categorize pixels affected by adverse weather and background pixels into distinct bins based on descending intensities (as depicted in (c)) and then conducts self-attention within and between bins." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.416, + 0.375, + 0.431 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.448, + 0.788, + 0.538 + ], + "angle": 0, + "content": "The field of computer vision witnessed growing interest in restoring images affected by adverse weather conditions like rain, fog, and snow. These weather conditions significantly degrade visual quality, impacting the performance of downstream tasks such as object detection [3,60], and depth estimation [16,18]. The restoration of images under adverse weather is thereby a vital problem for the sake of vision aesthetics and safety." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.539, + 0.788, + 0.674 + ], + "angle": 0, + "content": "Early works leverage weather-related priors to model statistical characteristic of degradation and remove adverse weathers [1, 2, 22, 23, 35, 76, 83, 95, 99]. Subsequently, convolutional neural networks (CNNs) have emerged as powerful tools for addressing deraining [5, 17, 34, 40, 57, 73, 79, 80, 88, 89], dehazing [27, 29, 62-64, 77, 88, 90, 92] and desnowing [30, 44, 61, 94]. However, the need of separately training networks for each task and the complexity of switching among multiple models present challenges for real-world applications. Li et al. [33] thus introduced the challenge of adverse weather removal, which entails the restoration of images affected by various weather conditions using a single unified model." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.675, + 0.788, + 0.795 + ], + "angle": 0, + "content": "Recently, Transformer-based approaches have also been investigated for the adverse weather removal task, surpassing the efficacy of CNNs [11, 19, 70, 72]. Nonetheless, these Transformer-based methods usually make concessions regarding efficient memory utilization by confining self-attention operations to a fixed spatial range or solely within the channel dimension, as depicted in Figure 1b. This compromise impedes the inherent potential of Transformers, which was originally designed for superior global feature modeling, and consequently, it leads to a deterioration in the performance of restoration." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.796, + 0.788, + 0.84 + ], + "angle": 0, + "content": "To address these problems, based on the observation that weather-induced degradation often exhibits common patterns shown in Figure 1a, we develop an efficient Histogram Transformer for unified adverse weather removal, named" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.219, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Restoring Images in Adverse Weather Conditions via Histogram Transformer" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.344 + ], + "angle": 0, + "content": "Histoformer. Specifically, we introduce a Dynamic-range Histogram Self-Attention (DHSA) module, which endows self-attention with a dynamic-range spatial receptive field. We categorize pixel values proximate in intensity yet varied in spatial location into histogram bins. Self-attention is executed across the dimension of bin or frequency, whose process is illustrated in Figure 1c. To facilitate comprehensive feature extraction on both local and global scales, we devise two ways of reshaping for histogram self-attention: bin-wise histogram reshaping (BHR) and frequency-wise histogram reshaping (FHR). In BHR, the number of bins is configured to incorporate pixels spanning a more comprehensive intensity range, thereby facilitating global feature integration. In FHR, the number of frequencies is assigned such that each bin focuses on limited number of pixels, enhancing the utility of finer features. Consequently, the histogram self-attention attains the capability of modeling spatially dynamic ranges effectively." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.344, + 0.788, + 0.495 + ], + "angle": 0, + "content": "To enable the convolution to extract dynamically-located weather-related dependencies, we develop a dynamic-range convolution layer, which involves sequential horizontal and vertical pixel sorting prior to the application of separable convolution. In order to capture multi-scale and multi-range information embedded within feature matrices, we introduce a Dual-scale Gated Feed-Forward (DGFF) module, enhancing its ability to model the visual characteristics effectively. Additionally, we note that conventional loss functions primarily focus on pixel-level closeness, overlooking the correlation at overall patch level. Consequently, we propose to leverage the Pearson correlation coefficient [12] to ensure the reconstruction of the linear relationship between restored and clean images." + }, + { + "type": "text", + "bbox": [ + 0.24, + 0.496, + 0.612, + 0.509 + ], + "angle": 0, + "content": "Our contribution can be summarized in three folds:" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.522, + 0.784, + 0.58 + ], + "angle": 0, + "content": "- We propose a novel transformer targeted for unified adverse weather removal, equipped with a new histogram self-attention. It possesses dynamic-range spatial attention to weather-induced obstructions and thus can achieve degradation removal globally and efficiently." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.583, + 0.784, + 0.625 + ], + "angle": 0, + "content": "- To capture multi-range information, we present a dual-scale feed-forward module. To enhance the comprehensive linear association between the recovered and ground-truth images, we develop a correlation loss." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.629, + 0.784, + 0.672 + ], + "angle": 0, + "content": "- Our method attains state-of-the-art performance across various datasets. Additionally, we substantiate the efficacy of the proposed approach to restore real-world images and bolster the downstream application of detection." + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.522, + 0.784, + 0.672 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.699, + 0.388, + 0.714 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.733, + 0.784, + 0.794 + ], + "angle": 0, + "content": "Extensive research has been dedicated to addressing adverse weather removal challenges in computer vision, including tasks like deraining [5,17,34,40,66,73, 79,80,88,89], dehazing [27,29,62-64,67,77,88,90,92], desnowing [30,44,61,94], raindrop removal [57,59,83,93] and All-in-One weather removal [28,33,53,70]." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.81, + 0.785, + 0.84 + ], + "angle": 0, + "content": "Rain Streak Removal. The evolution of approaches is notable in rain streak removal techniques in computer vision. Kang et al. [23] pioneered a single image" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.355, + 0.127 + ], + "angle": 0, + "content": "S. Sun et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.375 + ], + "angle": 0, + "content": "deraining method using bilateral filters to decompose images into low and high-frequency components. However, recent advancements have seen a dominance of deep neural networks. An early deep CNN was introduced by Fu et al. [17] for extracting features from the high-frequency rain component, while Yang et al. [79] utilized recurrent networks to decompose rain layers and remove various streak types. Li et al. [32] proposed a method that addresses rain streaks and veiling effects in heavy rain scenes by integrating physics-based rain models and adversarial learning. A conditional generative adversarial network was also employed to solve rain streak removal [89]. Yasarla et al. [81] explored Gaussian processes for transfer learning from synthetic to real-world rain data. Quan et al. [58] used a cascaded network to remove both rain streaks and raindrops. Recently an image deraining Transformer [78] featuring a dual Transformer architecture was intricately formulated, incorporating both window-based and spatial-based mechanisms, thereby attaining exemplary outcomes. A sparse deraining Transformer is also proposed to enhance feature aggregation [11]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.388, + 0.788, + 0.526 + ], + "angle": 0, + "content": "Raindrop Removal. Raindrop removal from single images has been addressed through various methods, with some relying on traditional hand-crafted features. An early work incorporated temporal information to address video-based raindrop removal [83]. Eigen et al. [15] employed a shallow CNN trained with image pairs containing raindrop-degraded and raindrop-free versions, though the results often exhibited blurriness. Qian et al. [57] introduced an attention GAN and a new dataset. Their method was later improved by Quan et al. [59] via generating attention maps based on mathematical raindrop descriptions and combining them with detected raindrop edges." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.539, + 0.788, + 0.66 + ], + "angle": 0, + "content": "Snow Removal. Desnow-Net [44] was among the pioneering CNN-based approaches for snow removal, followed by Li et al.'s stacked dense network [31] and Chen et al.'s JSTASR [8], which introduced a size and transparency aware method. More recently, DDMSNet [94] introduced a dense multi-scale network that leverages semantic and geometric priors to enhance snow removal. A hierarchical decomposition paradigm involving the dual-tree wavelet transform for snow removal is also proposed [9]. Chen et al. [7] designed SnowFormer, a framework that used cross-attention to establish local-global context interaction." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.788, + 0.796 + ], + "angle": 0, + "content": "Fog Removal. Li et al. [26] presented a CNN that takes into account both atmospheric luminosity and transmission maps to conduct dehazing. Ren et al. [63] advocated a pre-processing approach for hazy image manipulation, thereby engendering multiple input modalities and, in the process, inducing chromatic aberrations as part of their dehazing procedure. A hierarchical density-aware network is also introduced, specializing in the domain of image dehazing [92]. Zheng et al. [97] formulated a curriculum-based contrastive regularization dehazing method aimed at fostering agreement within a contrastive space." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.81, + 0.787, + 0.842 + ], + "angle": 0, + "content": "All-in-One Weather Removal. Some recent works attempted to address various weather-induced degradations by a singular network. Li et al. [33] proposed an" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.219, + 0.115, + 0.733, + 0.131 + ], + "angle": 0, + "content": "Restoring Images in Adverse Weather Conditions via Histogram Transformer" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.314 + ], + "angle": 0, + "content": "All-in-One network, containing a generator comprising multiple task-specific encoders and a shared decoder. Valanarasu et al. [70] presented TransWeather, a transformer-based model featuring a solitary encoder-decoder structure, capable of rejuvenating images afflicted by various atmospheric conditions. A pipeline for the automatic selection of weather-degraded data was also proposed to enhance existing models [91]. Zhu et al. developed WGWS-Net [100] capable of learning weather-general and weather-specific in two separate stages. Some other recent works also trial addressing adverse weather removal by adopting probabilistic denoising diffusion model [53], knowledge distillation [10], large-scale Pre-trained model [68], mixture of experts [49], few-shot learning [24], codebooks [41,71,82], adaptive filters [54], knowledge assignment [74] and domain translation [56]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.33, + 0.788, + 0.496 + ], + "angle": 0, + "content": "Transformer-based Image Restoration. Since the inception of the Vision Transformer (ViT) [14] for visual recognition, transformers have gained substantial traction across a spectrum of computer vision tasks [25, 37, 45, 46, 50, 51, 86]. Particularly within the realm of low-level vision, the Image Processing Transformer [4] exemplifies how pre-training a transformer on extensive datasets can significantly enhance performance for low-level applications. U-former [75], on the other hand, introduced a transformer architecture based on the U-Net design for restoration tasks. Swin-IR [38] employed the Swin Transformer [45] for image restoration. Some latest Transformer-based methods were proposed for deraining [11, 39], desnowing [7], dehazing [19, 43, 65] and All-in-One weather removal [70, 72]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.497, + 0.788, + 0.557 + ], + "angle": 0, + "content": "Unlike the existing Transformer-based approaches whose self-attention is applied within either fixed spatial ranges or merely channel dimension, our method enables dynamic-range spatial attention to adaptively focus on weather-induced degradation with similar patterns." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.583, + 0.331, + 0.599 + ], + "angle": 0, + "content": "3 Method" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.616, + 0.434, + 0.631 + ], + "angle": 0, + "content": "3.1 Overall Architecture" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.643, + 0.788, + 0.795 + ], + "angle": 0, + "content": "The overall architectural framework of our Histoformer is illustrated in Figure 2. Suppose the input is a low-quality image \\( I^{lq} \\in \\mathbb{R}^{3 \\times H \\times W} \\), we pass the input through a \\( 3 \\times 3 \\) convolution to conduct the overlapping image patch-embedding. Within both the encoder and decoder of the network backbone, we arrange Histogram Transformer Blocks (HTBs) to extract intricate features and capture dynamically distributed degradation factors. Within the same stage, encoders and decoders are interlinked through skip-connections, thereby establishing connections between consecutive intermediate features to enhance the stability of the training process. Between each stage, we apply pixel-unshuffle and pixelshuffle operations for the purpose of feature down-sampling and up-sampling." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.796, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Within each HTB, we introduce Dynamic-range Histogram Self-Attention (DHSA) to extract spatially dynamic weather degradation and enhance both local and global feature aggregation. Moreover, a Dual-scale Gated Feed-Forward" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.356, + 0.127 + ], + "angle": 0, + "content": "S. Sun et al." + }, + { + "type": "image", + "bbox": [ + 0.216, + 0.142, + 0.784, + 0.35 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.359, + 0.788, + 0.43 + ], + "angle": 0, + "content": "Fig. 2: The overall architecture of our Histoformer for weather removal. The main component is the Histogram Transformer block, and it comprises the Dynamic-range Histogram Self-Attention (DHSA) module and the Dual-scale Gated Feed-Forward (DGFF) module. Within DHSA, we present two types of reshaping mechanism, i.e., Bin-wise Histogram Reshaping and Frequency-wise Histogram Reshaping." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.456, + 0.788, + 0.624 + ], + "angle": 0, + "content": "(DGFF) module is integrated into the HTB to enrich the representation of multi-range features, contributing to the process of image restoration. During each stage of encoding phases, our model is equipped with a crude skip-connection for supplementing original features from input, comprised of a sequence of operations, including average pooling, pixel-wise convolution, and depth-wise convolution. We start the crude skip-connection after the first stage, and this setup enables the encoders to focus more effectively on learning the weather-induced residuals. Through this hybrid formulation, Histoformer is empowered to exploit both the adaptive contents of weather-irrelevant background and the inherent characteristics of weather-degraded patterns, facilitating the separation of undesired degradation from the latent clear background." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.643, + 0.509, + 0.658 + ], + "angle": 0, + "content": "3.2 Histogram Transformer Block" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.665, + 0.787, + 0.711 + ], + "angle": 0, + "content": "As the key component of our Histoformer, HTB contains two pivotal modules, i.e., DHSA and DGFF. These two components are arranged to interact with layer normalization and can be formulated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.384, + 0.718, + 0.785, + 0.734 + ], + "angle": 0, + "content": "\\[\nF _ {l} = F _ {l - 1} + \\mathrm {D H S A} (\\mathrm {L N} (F _ {l - 1})), \\tag {1}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.385, + 0.737, + 0.785, + 0.753 + ], + "angle": 0, + "content": "\\[\nF _ {l} = F _ {l} + \\operatorname {D G F F} \\left(\\ln \\left(F _ {l}\\right)\\right), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.759, + 0.784, + 0.791 + ], + "angle": 0, + "content": "where LN denotes layer normalization and \\(F_{l}\\) represents the feature at \\(l\\)-th stage. The details of DHSA and DGFF are presented in Section 3.2 and 3.2 respectively." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.792, + 0.842 + ], + "angle": 0, + "content": "Dynamic-range Histogram Self-Attention To better capture dynamically distributed weather-induced degradation, we develop a Dynamic-range Histogram" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.219, + 0.115, + 0.733, + 0.131 + ], + "angle": 0, + "content": "Restoring Images in Adverse Weather Conditions via Histogram Transformer" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.239 + ], + "angle": 0, + "content": "Self-Attention (DHSA) module. This module consists of a process involving dynamic-range convolution, which reorders the spatial distribution of fractional features, and a dual-path histogram self-attention mechanism that combines global and local dynamic feature aggregation. Prior to the final output projection of a \\(1 \\times 1\\) point-wise convolution, the reordered features are sorted back into their original locations to maintain spatial consistency." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.249, + 0.789, + 0.446 + ], + "angle": 0, + "content": "Dynamic-range Convolution. Traditional convolution operations employ fixed kernel sizes, resulting in a limited receptive field range and consequently perform local and small-range computations. This restricted scope of convolution, which primarily focuses on local information, does not naturally align with the self-attention mechanism's capacity to model long-range dependencies. To address this limitation, we devise a dynamic-range convolution technique by meticulously reordering the input features prior to the traditional convolution operation. Given an input feature \\( F \\in \\mathbb{R}^{C \\times H \\times W} \\), we divide it into two branches along the channel dimension, namely \\( F_{1} \\) and \\( F_{2} \\). For the first branch of features, we perform sorting operations both horizontally and vertically, subsequently concatenating the sorted features with the second branch of feature. The resulting recombined features are then passed through the subsequent separable convolution. The entire process is articulated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.354, + 0.453, + 0.652, + 0.469 + ], + "angle": 0, + "content": "\\[\nF _ {1}, F _ {2} = \\operatorname {S p l i t} (F), F _ {1} = \\operatorname {S o r t} _ {\\mathrm {v}} (\\operatorname {S o r t} _ {\\mathrm {h}} (F _ {1})),\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.354, + 0.462, + 0.785, + 0.489 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} F = \\operatorname {C o n v} _ {3 \\times 3} ^ {\\mathrm {d}} \\left(\\operatorname {C o n v} _ {1 \\times 1} \\left(\\operatorname {C o n c a t} \\left(F _ {1}, F _ {2}\\right)\\right)\\right), \\end{array} \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.497, + 0.789, + 0.664 + ], + "angle": 0, + "content": "where \\(\\mathrm{Conv}_{1\\times 1}\\) is \\(1\\times 1\\) point-wise convolution, \\(\\mathrm{Conv}_{3\\times 3}^{\\mathrm{d}}\\) represents \\(3\\times 3\\) depthwise convolution, Concat is the concatenation operation along channel, Split denotes the operation of splitting features along channel dimension, and \\(\\mathrm{Sort}_{i\\in \\{\\mathrm{h},\\mathrm{v}\\}}\\) represents the horizontal or vertical sorting operation. This approach organizes pixels of high and low intensities into regular patterns at the diagonal corners of the matrices, thereby allowing convolution to perform computations across dynamic ranges. Given that weather-induced degradation typically exhibits closely related patterns, degraded pixels tend to concentrate in neighboring locations, separated with those clean pixels. As a result, this arrangement enables convolution kernels to partially focus on preserving clean information and separately recovering degraded features." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.789, + 0.842 + ], + "angle": 0, + "content": "Histogram Self-Attention. Existing vision Transformers [11,75,78,78,86,96] typically leverage fixed range of attention or merely the attention along channel dimension due to the compromise of computation and memory efficiency. However, the fixed setting restricts the self-attention to span adaptively long range to associate desired features. We notice that weather-induced degradation causes similar patterns and that those pixels containing either background feature or weather degradation of different intensities had better be assigned with various extents of attention. We thus propose a histogram self-attention mechanism to categorize spatial elements into bins and allocate varying attention within and across bins. For the sake of parallel computing, we set each bin contains identical number of pixels during implement." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.356, + 0.128 + ], + "angle": 0, + "content": "S. Sun et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.8, + 0.208 + ], + "angle": 0, + "content": "Given the output of dynamic-range convolution, we can separate them into Value feature \\( V \\in \\mathbb{R}^{C \\times H \\times W} \\) and two pairs of Query-key \\( F_{QK,1}, F_{QK,2} \\in \\mathbb{R}^{2C \\times H \\times W} \\) subsequently passing to two branches. We firstly sort the sequence of \\( V \\) and based on its index arrange the Query-Key pairs accordingly, expressed as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.335, + 0.213, + 0.531, + 0.231 + ], + "angle": 0, + "content": "\\[\nV, d = \\operatorname {S o r t} \\left(\\mathbf {R} _ {\\mathrm {C} \\times \\mathrm {H} \\times \\mathrm {W}} ^ {\\mathrm {C} \\times \\mathrm {H W}} (V)\\right),\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.335, + 0.234, + 0.785, + 0.253 + ], + "angle": 0, + "content": "\\[\nQ _ {1}, K _ {1} = \\operatorname {S p l i t} \\left(\\operatorname {G a t h e r} \\left(\\mathbf {R} _ {\\mathrm {C} \\times \\mathrm {H} \\times \\mathrm {W}} ^ {\\mathrm {C} \\times \\mathrm {H W}} \\left(F _ {Q K, 1}\\right), d\\right)\\right), \\tag {4}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.335, + 0.256, + 0.67, + 0.275 + ], + "angle": 0, + "content": "\\[\nQ _ {2}, K _ {2} = \\operatorname {S p l i t} \\left(\\operatorname {G a t h e r} \\left(\\mathbf {R} _ {\\mathrm {C} \\times \\mathrm {H} \\times \\mathrm {W}} ^ {\\mathrm {C} \\times \\mathrm {H W}} (F _ {Q K, 2}), d\\right)\\right),\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.28, + 0.787, + 0.33 + ], + "angle": 0, + "content": "where \\(\\mathbf{R}_{\\mathrm{C}\\times ,\\mathrm{H}\\times \\mathrm{W}}^{\\mathrm{C}\\times \\mathrm{HW}}\\) represents the operation of reshaping features from \\(\\mathbb{R}^{C\\times H\\times W}\\) to \\(\\mathbb{R}^{C\\times HW}\\), \\(d\\) is the index of sorted Value, and Gather denotes the operation of retrieving elements of tensor based on a given index." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.334, + 0.789, + 0.53 + ], + "angle": 0, + "content": "Then given the number of bins \\( B \\), we reshape the sorted features from \\( C \\times HW \\) into \\( C \\times B \\times HW / B \\). To extract both global and local information, we define two types of reshaping, i.e., bin-wise histogram reshaping (BHR) and frequency-wise histogram reshaping (FHR). The first is to assign the number of bins equal to \\( B \\) and each bin contains \\( HW / B \\) elements, while the second is to set the frequency of each bin equal to \\( B \\) and the number of bins is \\( HW / B \\). By this way, we can extract large-scale information by BHR where each bin contains large number of dynamically located pixels and fine-grained information by FHR where each bins contains modicum pixels neighboring in terms of intensity. The two pairs of Query-Key features are passed through two types of reshaping and subsequent self-attention process respectively, and their outputs are elementwisely multiplied to yield the final output. The process can be formulated as the following expressions:" + }, + { + "type": "equation", + "bbox": [ + 0.345, + 0.536, + 0.66, + 0.57 + ], + "angle": 0, + "content": "\\[\n\\mathrm {A _ {B}} = \\mathrm {s o f t m a x} \\left(\\frac {\\mathbf {R _ {B}} (Q _ {1}) \\mathbf {R _ {B}} (K _ {1}) ^ {\\top}}{\\sqrt {k}}\\right) \\mathbf {R _ {B}} (V),\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.345, + 0.572, + 0.785, + 0.606 + ], + "angle": 0, + "content": "\\[\n\\mathrm {A} _ {\\mathrm {F}} = \\operatorname {s o f t m a x} \\left(\\frac {\\mathbf {R} _ {\\mathrm {F}} \\left(Q _ {2}\\right) \\mathbf {R} _ {\\mathrm {F}} \\left(K _ {2}\\right) ^ {\\top}}{\\sqrt {k}}\\right) \\mathbf {R} _ {\\mathrm {F}} (V), \\tag {5}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.345, + 0.61, + 0.449, + 0.625 + ], + "angle": 0, + "content": "\\[\n\\mathrm {A} = \\mathrm {A} _ {\\mathrm {B}} \\odot \\mathrm {A} _ {\\mathrm {F}},\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.632, + 0.788, + 0.664 + ], + "angle": 0, + "content": "where \\( k \\) is the number of heads, \\( \\mathbf{R}_{i\\in \\{\\mathrm{B},\\mathrm{F}\\}} \\) denotes the reshaping operation of either BHR or FHR, and \\( \\mathrm{A}_{i\\in \\{\\mathrm{B},\\mathrm{F}\\}} \\) represents the obtained attention map." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.704, + 0.789, + 0.841 + ], + "angle": 0, + "content": "Dual-scale Gated Feed-Forward Previous studies [11, 75, 78, 86] typically leverage single-range or single-scale convolution in the standard feed-forward network to bolster local context. Nonetheless, these methods often disregard the correlations among dynamically distributed weather-induced degradation. In practice, multi-scale information can be extracted by not only enlarging the kernel size but also leveraging the dilation mechanism [36, 84, 85]. As a result, we conceive a Dual-scale Gated Feed-Forward (DGFF) module, which integrates two distinct multi-range and multi-scale depth-wise convolution pathways within the transmission process." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.219, + 0.115, + 0.733, + 0.131 + ], + "angle": 0, + "content": "Restoring Images in Adverse Weather Conditions via Histogram Transformer" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.151, + 0.788, + 0.221 + ], + "angle": 0, + "content": "Table 1: Quantitative comparisons on three weather removal tasks in terms of PSNR and SSIM, where higher values indicate better performance. The top halves of tables display the results of task-specific methods, while the bottom halves present evaluations of the unified multi-weather models. The best and the second best results are in bold and underlined. Those with * indicate the methods whose source codes are unavailable." + }, + { + "type": "table_caption", + "bbox": [ + 0.28, + 0.226, + 0.397, + 0.237 + ], + "angle": 0, + "content": "(a) Image Desnowing" + }, + { + "type": "table_caption", + "bbox": [ + 0.475, + 0.226, + 0.614, + 0.237 + ], + "angle": 0, + "content": "(b) Deraining & Dehazing" + }, + { + "type": "table_caption", + "bbox": [ + 0.655, + 0.226, + 0.772, + 0.237 + ], + "angle": 0, + "content": "(c) Raindrop Removal" + }, + { + "type": "table", + "bbox": [ + 0.22, + 0.238, + 0.785, + 0.374 + ], + "angle": 0, + "content": "
Snow100K-S [44]Snow100K-L [44]Outdoor-Rain [32]RainDrop [57]
PSNRSSIMPSNRSSIMPSNRSSIMPSNR SSIM
SPANet [73]29.920.826023.700.7930CycleGAN [98]17.620.6560pix2pix [20]28.02
JSTASR [8]31.400.901225.320.8076pix2pix [20]19.090.7100DuRN [42]31.24
RESCAN [34]31.510.903226.080.8108HRGAN [32]21.560.8550RaindropAttn [59]31.44
DesnowNet [44]32.330.950027.170.8983PCNet [21]26.190.9015AttentiveGAN [57]31.59
DDMSNet [94]34.340.944528.850.8772MPRNet [87]28.030.9192IDT [78]31.87
NAFNet [6]34.790.949730.060.9017NAFNet [6]29.590.9027MAXIM [69]31.87
Restormer [86]36.010.957930.360.9068Restormer [86]30.030.9215Restormer [86]32.18
All-in-One [33]*--28.330.8820All-in-One [33]*24.710.8980All-in-One [33]*31.12
TransWeather [70]32.510.934129.310.8879TransWeather [70]28.830.9000TransWeather [70]30.17
Chen et al. [10]34.420.946930.220.9071Chen et al. [10]29.270.9147Chen et al. [10]31.81
WGWSNet [100]34.310.946030.160.9007WGWSNet [100]29.320.9207WGWSNet [100]32.38
WeatherDiff64 [53]35.830.956630.090.9041WeatherDiff64 [53]29.640.9312WeatherDiff64 [53]30.71
WeatherDiff128 [53]35.020.951629.580.8941WeatherDiff128 [53]29.720.9216WeatherDiff128 [53]29.66
AWRCP [82]*36.920.965231.920.9341AWRCP [82]*31.390.9329AWRCP [82]*31.93
Histoformer (Ours)37.410.965632.160.9261Histoformer (Ours)32.080.9389Histoformer (Ours)33.06
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.402, + 0.789, + 0.538 + ], + "angle": 0, + "content": "Given an input tensor \\( F_{l} \\in \\mathbb{R}^{C \\times H \\times W} \\), we initially employ a point-wise convolution operation to augment the channel dimension by a factor of \\( r \\). Following this augmentation, the expanded tensor is directed into two parallel branches. Throughout the feature transformation process, \\( 5 \\times 5 \\) and dilated \\( 3 \\times 3 \\) depth-wise convolutions are employed to enhance the extraction of multi-range and multi-scale information. Following the gating mechanism [13], the output of the second branch after passing through an activation act as a gating map for the other branch. Thus, the complete feature fusion process within the DGFF module is formulated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.333, + 0.543, + 0.621, + 0.559 + ], + "angle": 0, + "content": "\\[\nF _ {l, 1}, F _ {l, 2} = \\operatorname {S p l i t} \\left(\\operatorname {S h u f f l e} \\left(\\operatorname {C o n v} _ {1 \\times 1} \\left(F _ {l}\\right)\\right)\\right),\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.333, + 0.562, + 0.785, + 0.581 + ], + "angle": 0, + "content": "\\[\nF _ {l, 1} = \\operatorname {C o n v} _ {5 \\times 5} ^ {\\mathrm {d}} \\left(F _ {l, 1}\\right), F _ {l, 2} = \\operatorname {C o n v} _ {3 \\times 3} ^ {\\mathrm {d}, \\text {d i l a t e d}} \\left(F _ {l, 2}\\right), \\tag {6}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.333, + 0.584, + 0.672, + 0.6 + ], + "angle": 0, + "content": "\\[\nF _ {l + 1} = \\operatorname {C o n v} _ {1 \\times 1} \\left(\\text {U n s h u r f l e} \\left(\\operatorname {M i s h} \\left(F _ {l, 2}\\right) \\odot F _ {l, 1}\\right)\\right),\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.607, + 0.788, + 0.67 + ], + "angle": 0, + "content": "where \\(\\mathrm{Conv}_{5\\times 5}^{\\mathrm{d}}\\) represents \\(5\\times 5\\) depth-wise convolution, \\(\\mathrm{Conv}_{3\\times 3}^{\\mathrm{d,dilated}}\\) is \\(3\\times 3\\) dilated depth-wise convolution, Shuffle and Unshuffle represent respectively the operations of pixel-shuffling and unshuffling, Mish denotes the Mish activation [52], and \\(F_{l + 1}\\) is the output of current stage passing to \\(l + 1\\)-th stage." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.693, + 0.607, + 0.708 + ], + "angle": 0, + "content": "3.3 Reconstruction Loss and Correlation Loss" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.787, + 0.75 + ], + "angle": 0, + "content": "We use the \\(L_{1}\\) norm of the pixel-wise difference between the restored high-quality image \\(I^{hq}\\) and ground-truth \\(I^{gt}\\) as the reconstruction loss, i.e.," + }, + { + "type": "equation", + "bbox": [ + 0.428, + 0.757, + 0.785, + 0.775 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {r e c} = \\left\\| I ^ {h q} - I ^ {g t} \\right\\| _ {1}. \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Furthermore, we notice that the \\(\\mathcal{L}_{rec}\\) only regulates the pixel-level similarity between the restored image and the ground-truth, while neglecting the patch-level linear correlations. The innate relationships of intensity within the image are disrupted by the consistent patterns of weather-induced degradation. By" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.356, + 0.128 + ], + "angle": 0, + "content": "S. Sun et al." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.144, + 0.787, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.23, + 0.785, + 0.259 + ], + "angle": 0, + "content": "Fig. 3: Visual comparison for desnowing on Snow100K [44]. The samples from (b) to (e) are Restormer [86], TransWeather [70], WGWSNet [100], WeatherDiff [53]." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.272, + 0.787, + 0.345 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.348, + 0.785, + 0.391 + ], + "angle": 0, + "content": "Fig. 4: Visual comparison for deraining and dehazing on Outdoor-Rain [32]. The samples from (b) to (e) are Restormer [86], TransWeather [70], WGWSNet [100], WeatherDiff [53]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.421, + 0.784, + 0.482 + ], + "angle": 0, + "content": "emulating the intensity relationships within the ground-truth, we compel the degraded pixels to occupy their original positions according to the original intensity ranking. Consequently, we introduce the Pearson correlation [12] between images as a means to regulate the linear relationship, expressed as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.332, + 0.486, + 0.785, + 0.529 + ], + "angle": 0, + "content": "\\[\n\\rho \\left(I ^ {h q}, I ^ {g t}\\right) = \\frac {\\sum_ {i = 1} ^ {3 H W} \\left(I _ {i} ^ {h q} - \\bar {I} ^ {h q}\\right) \\left(I _ {i} ^ {g t} - \\bar {I} ^ {g t}\\right)}{3 H W \\sigma \\left(I ^ {h q}\\right) \\sigma \\left(I ^ {g t}\\right)}, \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.536, + 0.784, + 0.614 + ], + "angle": 0, + "content": "where \\( I_{i}^{\\{\\cdot\\}} \\) represents the \\( i \\)-th pixel of image, \\( \\overline{I}^{\\{\\cdot\\}} \\) and \\( \\sigma(I^{\\{\\cdot\\}}) \\) denotes respectively the mean and the standard deviation of image sequence. Its value falls within the \\([-1,1]\\) range. When two images exhibit perfect correlation, the value of function \\( \\rho \\) attains a value of 1, while in the case of negative correlation, its value reaches \\(-1\\). Hence, we formulate the correlation loss as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.401, + 0.621, + 0.785, + 0.65 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {c o r} = \\frac {1}{2} \\left(1 - \\rho (I ^ {h q}, I ^ {g t})\\right), \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.654, + 0.783, + 0.684 + ], + "angle": 0, + "content": "such that \\(\\mathcal{L}_{cor} = 0\\) when the recovered image perfectly aligns with the groundtruth. The overall loss function is thus defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.434, + 0.692, + 0.785, + 0.709 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathcal {L} _ {\\text {r e c}} + \\alpha \\mathcal {L} _ {\\text {c o r}}, \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.715, + 0.509, + 0.73 + ], + "angle": 0, + "content": "where \\(\\alpha\\) is the weight of correlation loss." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.754, + 0.376, + 0.772 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.785, + 0.444, + 0.801 + ], + "angle": 0, + "content": "4.1 Experimental settings" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.784, + 0.84 + ], + "angle": 0, + "content": "Datasets. We train our model on the same datasets as the previous works [33,53, 70] to ensure a fair comparison. The training set encompasses 9,000 images drawn" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.219, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Restoring Images in Adverse Weather Conditions via Histogram Transformer" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.144, + 0.787, + 0.225 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.217, + 0.229, + 0.785, + 0.259 + ], + "angle": 0, + "content": "Fig. 5: Visual comparison for raindrop removal on RainDrop [57]. The samples from (b) to (e) are Chen et al. [10], TransWeather [70], WGWSNet [100], WeatherDiff [53]." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.27, + 0.787, + 0.35 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.354, + 0.785, + 0.396 + ], + "angle": 0, + "content": "Fig.6: A qualitative comparison for real-world adverse weather removal on Snow100K [44]. The samples from (b) to (e) are Chen et al. [10], Restormer [86], TransWeather [70], WGWSNet [100], WeatherDiff [53]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.422, + 0.785, + 0.528 + ], + "angle": 0, + "content": "from Snow100K [44], 1,069 images sourced from Raindrop [57], and 9,000 images from Outdoor-Rain [32]. Snow100K contains synthetic images deteriorated by snow, while Raindrop comprises real raindrop-affected images. Outdoor-Rain features synthetic images afflicted by both fog and rain streaks. For evaluation, we employ the Test1 dataset [32, 33], the RainDrop test dataset [57], and the Snow100K-L and -S test sets [44]. Snow100K also provides a real-world test set containing 1,329 images affected by adverse weather." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.541, + 0.785, + 0.72 + ], + "angle": 0, + "content": "Comparison Baselines. We assess the performance of our method against state-of-the-art approaches designed specifically for distinct weather removal tasks: raindrop removal, snow removal, and rain&fog removal. Specifically, for snow removal, we benchmark against SPANet [73], JSTASR [8], RESCAN [34], DesnowNet [44], and DDMSNet [94]. In the case of rain&fog removal, we compare with CycleGAN [98], pix2pix [20], HRGAN [32], MPRNet [87] and Restormer [86]. For raindrop removal, we evaluate against the methods such as pix2pix [20], DuRN [42], RaindropAttn [59], AttentiveGAN [57]. Additionally, we include some recent transformer or multi-degradation restoration networks, IDT [78], NAFNet [6], MAXIM [69], and Restormer [86], in our comparative analysis. It is worth noting that all these methods are single-task networks fine-tuned for specific datasets." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.722, + 0.785, + 0.798 + ], + "angle": 0, + "content": "Furthermore, we conduct a performance comparison with the All-in-One network [33], Chen et al. [10], TransWeather [70], WGWS-Net [100], WeatherDiff [53] and AWRCP [82], which are trained to handle all the aforementioned tasks using a unified model. Note that our approach is also trained to tackle all these tasks using a single model." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.81, + 0.785, + 0.84 + ], + "angle": 0, + "content": "Training details. Our implementation is realized by PyTorch [55] and on NVIDIA Tesla V100 GPU. The network is trained for a total of 300,000 iterations, with" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.355, + 0.127 + ], + "angle": 0, + "content": "S. Sun et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.283 + ], + "angle": 0, + "content": "an initial batch size of 8 and an initial patch size of 128 akin to the progressive learning pipeline [86]. We employ the AdamW optimizer [48] with an initial learning rate of \\(3e^{-4}\\) for the first 92,000 iterations, which is gradually reduced to \\(1e^{-6}\\) using cosine annealing schedule [47] during the remaining 208,000 iterations. The number of blocks at each stage \\(L_{i\\in \\{1,2,3,4\\}}\\) is set to \\(\\{4,4,6,8\\}\\) and the channel size \\(C\\) is 36. The channel expansion factor \\(r\\) in DGFF is set to 2.667. The numbers of heads in self-attention at different stages are set to \\(\\{1,2,4,8\\}\\) respectively. We randomly apply horizontal and vertical flips as the technique of data augmentation." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.305, + 0.583, + 0.32 + ], + "angle": 0, + "content": "4.2 Comparisons with the state-of-the-arts" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.329, + 0.789, + 0.48 + ], + "angle": 0, + "content": "Quantitative Evaluation. In our study, we provide a comprehensive comparative analysis of metrics applied to both synthetic and real datasets, as summarized in Table 1. For a fair and well-founded comparison, we utilize recent multiple degradation removal methods such as MPRNet [87], MAXIM [69], and Restormer [86], treating them as weather-specific approaches for each benchmark. Additionally, we retrain the all-in-one adverse weather removal methods including Chen et al. [10] and WGWS-Net [100] using the all-weather training dataset [33,53,70]. This exhaustive comparison reveals that our proposed method exhibits a significant performance advantage over existing approaches across three different types of degradation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.494, + 0.789, + 0.631 + ], + "angle": 0, + "content": "Qualitative Evaluation. Furthermore, we conduct a visual comparison on three tasks, and the outcomes are showcased in Figure 3, 4 and 5 respectively. Figure 6 shows a case of real-world weather removal. These results highlight that our method excels in comprehensively eliminating snow degradation, including fine and large snow spots. In contrast, the recent WeatherDiff [53] method still exhibits some residual snow degradation, and its capability to restore details is not optimal. When it comes to the restoration of challenging weather conditions, our method excels in removing complex haze and rain marks, yielding visually appealing results in comparison to prior approaches." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.651, + 0.399, + 0.665 + ], + "angle": 0, + "content": "4.3 Ablation studies" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.675, + 0.788, + 0.737 + ], + "angle": 0, + "content": "To substantiate the effectiveness of each component within Histoformer, we conduct a sequence of ablation studies on Outdoor-Rain [32]. In particular, we examine the impact of the dynamic-range convolution, the DHSA module, the number of bins in DHSA, the DGFF module, and the correlation loss." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Dynamic-range Convolution. We experiment on two settings of dynamic-range convolution, namely, sorting horizontally first and then vertically before convolution, and the reverse order. Additionally, we compared them with vanilla convolution, and the results are displayed in Table 2. The operations of regular sorting led to a performance improvement of \\(0.14\\mathrm{dB}\\), and the order of sorting operations does not significantly affect the outcome." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.22, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Restoring Images in Adverse Weather Conditions via Histogram Transformer" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "table_caption", + "bbox": [ + 0.295, + 0.151, + 0.707, + 0.165 + ], + "angle": 0, + "content": "Table 2: Ablation studies on the dynamic-range convolution." + }, + { + "type": "table", + "bbox": [ + 0.331, + 0.166, + 0.673, + 0.214 + ], + "angle": 0, + "content": "
Vanilla ConvVerti.+Horiz.Horiz.+Verti.
PSNR31.9432.0332.08
SSIM0.93770.93850.9389
" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.225, + 0.496, + 0.251 + ], + "angle": 0, + "content": "Table 3: Ablation studies on the design of self-attention" + }, + { + "type": "table", + "bbox": [ + 0.257, + 0.255, + 0.457, + 0.332 + ], + "angle": 0, + "content": "
SA typePSNRSSIM
MDTA [86]30.940.9278
TKSA [11]31.120.9295
w/o BHR31.050.9301
w/o FHR31.790.9364
DHSA32.080.9389
" + }, + { + "type": "table_caption", + "bbox": [ + 0.505, + 0.225, + 0.785, + 0.251 + ], + "angle": 0, + "content": "Table 4: Ablation studies on the number of Bins in DHSA" + }, + { + "type": "table", + "bbox": [ + 0.533, + 0.255, + 0.761, + 0.332 + ], + "angle": 0, + "content": "
#Bins in DHSAPSNRSSIM
1230.430.9283
2031.560.9312
2831.940.9379
3632.080.9389
44Out of memory
" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.345, + 0.496, + 0.371 + ], + "angle": 0, + "content": "Table 5: Ablation studies on the choice of feed-forward module" + }, + { + "type": "table", + "bbox": [ + 0.251, + 0.374, + 0.463, + 0.452 + ], + "angle": 0, + "content": "
Feed-ForwardPSNRSSIM
FFN [38]31.320.9313
GDFN [86]31.420.9347
DANB [96]31.560.9351
MSFN [11]31.780.9367
DGFF32.080.9389
" + }, + { + "type": "table_caption", + "bbox": [ + 0.505, + 0.345, + 0.785, + 0.371 + ], + "angle": 0, + "content": "Table 6: Ablation studies on the setting of correlation loss" + }, + { + "type": "table", + "bbox": [ + 0.534, + 0.374, + 0.76, + 0.452 + ], + "angle": 0, + "content": "
αPSNRSSIM
w/o Lcor031.770.9358
0.132.010.9369
132.080.9389
w/ Lcor532.030.9392
1031.960.9375
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.479, + 0.785, + 0.554 + ], + "angle": 0, + "content": "DHSA. To evaluate the effectiveness of the proposed DHSA module, we conduct a comparison with two baselines, i.e., a multi-Dconv head transposed attention (MDTA) [86] and a top-k sparse attention (TKSA) [11]. Additionally, we explore two additional settings of DHSA by excluding either the BHR branch or the FHR branch. The quantitative analysis results are presented in Table 3." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.555, + 0.788, + 0.676 + ], + "angle": 0, + "content": "Both MDTA and TKSA integrate rich information across channels, which may result in a loss of the exploitation of long-range information across spatial dimensions. While our histogram self-attention can capture spatially long-range information, the use of either a single BHR or a single FHR branch neglects the inter-bin or inner-bin relationships, leading to inferior results. By incorporating dynamic-range convolution and dual-branch histogram self-attention, capable of extracting long-range spatial features, our DHSA enhances performance, resulting in a PSNR improvement of 0.96 dB compared to TKSA." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.69, + 0.788, + 0.766 + ], + "angle": 0, + "content": "Number of Bins. To assess the influence of the number of bins in DHSA, we conduct experiments with five different values: 12, 20, 28, 36, and 44. The results are presented in Table 4. It is observed that increasing the number of bins consistently improves performance. However, when the number of bins exceeds 44, it results in an out-of-memory error." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.78, + 0.788, + 0.843 + ], + "angle": 0, + "content": "DGFF. To assess the effectiveness of the proposed DGFF module, we conduct a comparison with four baselines: (i) the vanilla feed-forward network (FN) [38], (ii) a gated-Dconv feed-forward network (GDFN) [86], (iii) a dual adaptive neural block (DANB) [96], and (iv) a mixed-scale feed-forward network (MSFN) [11]." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.356, + 0.127 + ], + "angle": 0, + "content": "S. Sun et al." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.145, + 0.385, + 0.232 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.39, + 0.147, + 0.498, + 0.228 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.145, + 0.669, + 0.232 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.674, + 0.147, + 0.785, + 0.229 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.233, + 0.384, + 0.32 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.39, + 0.233, + 0.498, + 0.314 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.233, + 0.669, + 0.32 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.675, + 0.234, + 0.785, + 0.315 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.331, + 0.321, + 0.387, + 0.332 + ], + "angle": 0, + "content": "(a) Input" + }, + { + "type": "image_caption", + "bbox": [ + 0.572, + 0.321, + 0.718, + 0.332 + ], + "angle": 0, + "content": "(b) Deweathered by ours" + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.336, + 0.784, + 0.364 + ], + "angle": 0, + "content": "Fig. 7: Real-world deweathering on two snowy images [44] and their downstream detection results on Google API." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.391, + 0.785, + 0.467 + ], + "angle": 0, + "content": "The quantitative analysis results are presented in Table 5. While MSFN integrates mixed-scale information, it may still miss out on the exploitation of multi-range spatial knowledge. Through the inclusion of pixel-shuffling and feature aggregation across different ranges, our DGFF further enhances performance, resulting in a PSNR gain of 0.3 dB over MSFN." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.479, + 0.785, + 0.54 + ], + "angle": 0, + "content": "Correlation Loss. Table 6 shows the effectiveness of the correlation loss \\(\\mathcal{L}_{cor}\\) and the influence of its weight. It is evident that \\(\\mathcal{L}_{cor}\\) consistently improves the performance, while the specific loss weight does not have a substantial impact on the final results. We therefore keep the loss weight as 1 by default." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.558, + 0.453, + 0.574 + ], + "angle": 0, + "content": "4.4 Real-world Application" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.58, + 0.785, + 0.655 + ], + "angle": 0, + "content": "To further demonstrate the practical applicability of our method for real-world adverse weather removal and its potential to improve downstream detection task, we provide two samples in Figure 7. As depicted, our Histoformer effectively eliminates snowflakes from the scene and assists the detector in recognizing omitted door and building." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.676, + 0.359, + 0.691 + ], + "angle": 0, + "content": "5 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.705, + 0.784, + 0.84 + ], + "angle": 0, + "content": "In this research, we introduce a novel mechanism called histogram self-attention and devise a new histogram transformer named Histoformer to tackle the challenge of all-in-one weather removal. Our histogram self-attention involves segmenting spatial features into multiple bins, and allocating varying attention along the bin or frequency dimension, allowing it to selectively focus on weather-related features with a dynamic range. To facilitate learning both multi-range and multi-scale information, we present DGFF module and a correlation loss. Through extensive experimentation, we demonstrate the effectiveness and superiority of our approach." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.22, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Restoring Images in Adverse Weather Conditions via Histogram Transformer" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.393, + 0.163 + ], + "angle": 0, + "content": "Acknowledgement" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.178, + 0.788, + 0.254 + ], + "angle": 0, + "content": "This work has been supported in part by National Natural Science Foundation of China (No. 62322216, 62172409, 62025604, 62306308, 62311530686), in part by Shenzhen Science and Technology Program (Grant No. JCYJ20220818102012025, KQTD20221101093559018), and in part by Guangdong Provincial Key Laboratory of Information Security Technology (No. 2023B1212060026)." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.278, + 0.323, + 0.294 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.233, + 0.31, + 0.785, + 0.338 + ], + "angle": 0, + "content": "1. Ancuti, C.O., Ancuti, C.: Single image dehazing by multi-scale fusion. IEEE TIP (2013)" + }, + { + "type": "ref_text", + "bbox": [ + 0.233, + 0.34, + 0.751, + 0.353 + ], + "angle": 0, + "content": "2. Berman, D., Avidan, S., et al.: Non-local image dehazing. In: CVPR (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.233, + 0.354, + 0.785, + 0.381 + ], + "angle": 0, + "content": "3. Carion, N., Massa, F., Synnaeve, G., Usunier, N., Kirillov, A., Zagoruyko, S.: End-to-end object detection with transformers. In: ECCV (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.233, + 0.382, + 0.785, + 0.408 + ], + "angle": 0, + "content": "4. Chen, H., Wang, Y., Guo, T., Xu, C., Deng, Y., Liu, Z., Ma, S., Xu, C., Xu, C., Gao, W.: Pre-trained image processing transformer. In: CVPR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.233, + 0.41, + 0.785, + 0.436 + ], + "angle": 0, + "content": "5. Chen, J., Tan, C.H., Hou, J., Chau, L.P., Li, H.: Robust video content alignment and compensation for rain removal in a cnn framework. In: CVPR (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.233, + 0.438, + 0.785, + 0.464 + ], + "angle": 0, + "content": "6. Chen, L., Chu, X., Zhang, X., Sun, J.: Simple baselines for image restoration. In: ECCV (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.233, + 0.465, + 0.785, + 0.506 + ], + "angle": 0, + "content": "7. Chen, S., Ye, T., Liu, Y., Chen, E., Shi, J., Zhou, J.: Snowformer: Scale-aware transformer via context interaction for single image desnowing. arXiv preprint arXiv:2208.09703 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.233, + 0.507, + 0.785, + 0.562 + ], + "angle": 0, + "content": "8. Chen, W.T., Fang, H.Y., Ding, J.J., Tsai, C.C., Kuo, S.Y.: Jstasr: Joint size and transparency-aware snow removal algorithm based on modified partial convolution and veiling effect removal. In: Vedaldi, A., Bischof, H., Brox, T., Frahm, J.M. (eds.) ECCV (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.233, + 0.563, + 0.785, + 0.617 + ], + "angle": 0, + "content": "9. Chen, W.T., Fang, H.Y., Hsieh, C.L., Tsai, C.C., Chen, I., Ding, J.J., Kuo, S.Y., et al.: All snow removed: Single image desnowing algorithm using hierarchical dual-tree complex wavelet representation and contradict channel loss. In: ICCV (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.227, + 0.618, + 0.785, + 0.659 + ], + "angle": 0, + "content": "10. Chen, W.T., Huang, Z.K., Tsai, C.C., Yang, H.H., Ding, J.J., Kuo, S.Y.: Learning multiple adverse weather removal via two-stage knowledge learning and multi-contrastive regularization: Toward a unified model. In: CVPR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.227, + 0.66, + 0.785, + 0.687 + ], + "angle": 0, + "content": "11. Chen, X., Li, H., Li, M., Pan, J.: Learning a sparse transformer network for effective image deraining. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.227, + 0.688, + 0.785, + 0.728 + ], + "angle": 0, + "content": "12. Cohen, I., Huang, Y., Chen, J., Benesty, J., Benesty, J., Chen, J., Huang, Y., Cohen, I.: Pearson correlation coefficient. Noise reduction in speech processing (2009)" + }, + { + "type": "ref_text", + "bbox": [ + 0.227, + 0.729, + 0.785, + 0.756 + ], + "angle": 0, + "content": "13. Dauphin, Y.N., Fan, A., Auli, M., Grangier, D.: Language modeling with gated convolutional networks. In: ICML (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.227, + 0.757, + 0.785, + 0.812 + ], + "angle": 0, + "content": "14. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.227, + 0.813, + 0.785, + 0.84 + ], + "angle": 0, + "content": "15. Eigen, D., Krishnan, D., Fergus, R.: Restoring an image taken through a window covered with dirt or rain. In: ICCV (2013)" + }, + { + "type": "list", + "bbox": [ + 0.227, + 0.31, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.356, + 0.128 + ], + "angle": 0, + "content": "S. Sun et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.147, + 0.785, + 0.175 + ], + "angle": 0, + "content": "16. Fu, H., Gong, M., Wang, C., Batmanghelich, K., Tao, D.: Deep ordinal regression network for monocular depth estimation. In: CVPR (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.177, + 0.785, + 0.203 + ], + "angle": 0, + "content": "17. Fu, X., Huang, J., Zeng, D., Huang, Y., Ding, X., Paisley, J.: Removing rain from single images via a deep detail network. In: CVPR (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.204, + 0.785, + 0.231 + ], + "angle": 0, + "content": "18. Godard, C., Mac Aodha, O., Brostow, G.J.: Unsupervised monocular depth estimation with left-right consistency. In: CVPR (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.232, + 0.785, + 0.259 + ], + "angle": 0, + "content": "19. Guo, C., Yan, Q., Anwar, S., Cong, R., Ren, W., Li, C.: Image dehazing transformer with transmission-aware 3d position embedding. In: CVPR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.26, + 0.785, + 0.285 + ], + "angle": 0, + "content": "20. Isola, P., Zhu, J.Y., Zhou, T., Efros, A.A.: Image-to-image translation with conditional adversarial networks. In: CVPR (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.287, + 0.785, + 0.327 + ], + "angle": 0, + "content": "21. Jiang, K., Wang, Z., Yi, P., Chen, C., Wang, Z., Wang, X., Jiang, J., Lin, C.W.: Rain-free and residue hand-in-hand: A progressive coupled network for real-time image deraining. IEEE TIP (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.328, + 0.785, + 0.369 + ], + "angle": 0, + "content": "22. Jiang, T.X., Huang, T.Z., Zhao, X.L., Deng, L.J., Wang, Y.: A novel tensor-based video rain streaks removal approach via utilizing discriminatively intrinsic priors. In: CVPR (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.37, + 0.785, + 0.397 + ], + "angle": 0, + "content": "23. Kang, L.W., Lin, C.W., Fu, Y.H.: Automatic single-image-based rain streaks removal via image decomposition. IEEE TIP (2011)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.398, + 0.785, + 0.437 + ], + "angle": 0, + "content": "24. Kim, Y., Cho, Y., Nguyen, T.T., Lee, D.: Metaweather: Few-shot weather-degraded image restoration via degradation pattern matching. arXiv preprint arXiv:2308.14334 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.438, + 0.785, + 0.466 + ], + "angle": 0, + "content": "25. Lai, Z., Wu, J., Chen, S., Zhou, Y., Hovakimyan, N.: Residual-based language models are free boosters for biomedical imaging tasks. In: CVPRW (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.467, + 0.785, + 0.494 + ], + "angle": 0, + "content": "26. Li, B., Peng, X., Wang, Z., Xu, J., Feng, D.: Aod-net: All-in-one dehazing network. In: ICCV (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.495, + 0.785, + 0.522 + ], + "angle": 0, + "content": "27. Li, B., Ren, W., Fu, D., Tao, D., Feng, D., Zeng, W., Wang, Z.: Benchmarking single-image dehazing and beyond. IEEE TIP (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.523, + 0.785, + 0.549 + ], + "angle": 0, + "content": "28. Li, B., Liu, X., Hu, P., Wu, Z., Lv, J., Peng, X.: All-in-one image restoration for unknown corruption. In: CVPR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.55, + 0.785, + 0.577 + ], + "angle": 0, + "content": "29. Li, L., Dong, Y., Ren, W., Pan, J., Gao, C., Sang, N., Yang, M.H.: Semi-supervised image dehazing. IEEE TIP (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.578, + 0.785, + 0.604 + ], + "angle": 0, + "content": "30. Li, M., Cao, X., Zhao, Q., Zhang, L., Meng, D.: Online rain/snow removal from surveillance videos. IEEE TIP (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.606, + 0.785, + 0.633 + ], + "angle": 0, + "content": "31. Li, P., Yun, M., Tian, J., Tang, Y., Wang, G., Wu, C.: Stacked dense networks for single-image snow removal. Neurocomputing (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.634, + 0.785, + 0.66 + ], + "angle": 0, + "content": "32. Li, R., Cheong, L.F., Tan, R.T.: Heavy rain image restoration: Integrating physics model and conditional adversarial learning. In: CVPR (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.661, + 0.785, + 0.688 + ], + "angle": 0, + "content": "33. Li, R., Tan, R.T., Cheong, L.F.: All in one bad weather removal using architectural search. In: CVPR (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.689, + 0.785, + 0.716 + ], + "angle": 0, + "content": "34. Li, X., Wu, J., Lin, Z., Liu, H., Zha, H.: Recurrent squeeze-and-excitation context aggregation net for single image deraining. In: ECCV (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.717, + 0.785, + 0.743 + ], + "angle": 0, + "content": "35. Li, Y., Tan, R.T., Guo, X., Lu, J., Brown, M.S.: Rain streak removal using layer priors. In: CVPR (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.744, + 0.785, + 0.771 + ], + "angle": 0, + "content": "36. Li, Y., Lu, J., Chen, H., Wu, X., Chen, X.: Dilated convolutional transformer for high-quality image deraining. In: CVPRW (June 2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.772, + 0.785, + 0.812 + ], + "angle": 0, + "content": "37. Li, Z., Guan, B., Wei, Y., Zhou, Y., Zhang, J., Xu, J.: Mapping new realities: Ground truth image creation with pix2pix image-to-image translation. arXiv preprint arXiv:2404.19265 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.813, + 0.785, + 0.84 + ], + "angle": 0, + "content": "38. Liang, J., Cao, J., Sun, G., Zhang, K., Van Gool, L., Timofte, R.: Swinir: Image restoration using swin transformer. In: ICCV (2021)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.147, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.22, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Restoring Images in Adverse Weather Conditions via Histogram Transformer" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.147, + 0.785, + 0.175 + ], + "angle": 0, + "content": "39. Liang, Y., Anwar, S., Liu, Y.: Drt: A lightweight single image deraining recursive transformer. In: CVPR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.176, + 0.785, + 0.203 + ], + "angle": 0, + "content": "40. Liu, J., Yang, W., Yang, S., Guo, Z.: Erase or fill? deep joint recurrent rain removal and reconstruction in videos. In: CVPR (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.204, + 0.785, + 0.231 + ], + "angle": 0, + "content": "41. Liu, K., Jiang, Y., Choi, I., Gu, J.: Learning image-adaptive codebooks for class-agnostic image restoration. arXiv preprint arXiv:2306.06513 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.232, + 0.785, + 0.259 + ], + "angle": 0, + "content": "42. Liu, X., Suganuma, M., Sun, Z., Okatani, T.: Dual residual networks leveraging the potential of paired operations for image restoration. In: CVPR (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.26, + 0.785, + 0.286 + ], + "angle": 0, + "content": "43. Liu, Y., Liu, H., Li, L., Wu, Z., Chen, J.: A data-centric solution to nonhomogeneous dehazing via vision transformer. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.287, + 0.785, + 0.314 + ], + "angle": 0, + "content": "44. Liu, Y.F., Jaw, D.W., Huang, S.C., Hwang, J.N.: Desnownet: Context-aware deep network for snow removal. IEEE TIP (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.315, + 0.785, + 0.342 + ], + "angle": 0, + "content": "45. Liu, Z., Lin, Y., Cao, Y., Hu, H., Wei, Y., Zhang, Z., Lin, S., Guo, B.: Swin transformer: Hierarchical vision transformer using shifted windows. In: ICCV (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.343, + 0.785, + 0.37 + ], + "angle": 0, + "content": "46. Liu, Z., Ning, J., Cao, Y., Wei, Y., Zhang, Z., Lin, S., Hu, H.: Video swim transformer. In: CVPR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.371, + 0.785, + 0.397 + ], + "angle": 0, + "content": "47. Loshchilov, I., Hutter, F.: Sgdr: Stochastic gradient descent with warm restarts. arXiv preprint arXiv:1608.03983 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.398, + 0.785, + 0.424 + ], + "angle": 0, + "content": "48. Loshchilov, I., Hutter, F.: Fixing weight decay regularization in adam. arXiv preprint arXiv:1711.05101 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.425, + 0.785, + 0.466 + ], + "angle": 0, + "content": "49. Luo, Y., Zhao, R., Wei, X., Chen, J., Lu, Y., Xie, S., Wang, T., Xiong, R., Lu, M., Zhang, S.: Mowe: Mixture of weather experts for multiple adverse weather removal. arXiv preprint arXiv:2303.13739 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.467, + 0.785, + 0.494 + ], + "angle": 0, + "content": "50. Lyu, W., Zheng, S., Ling, H., Chen, C.: Backdoor attacks against transformers with attention enhancement. In: ICLR Workshop (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.495, + 0.785, + 0.535 + ], + "angle": 0, + "content": "51. Ma, H., Zeng, D., Liu, Y.: Learning individualized treatment rules with many treatments: A supervised clustering approach using adaptive fusion. NeurIPS (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.536, + 0.785, + 0.563 + ], + "angle": 0, + "content": "52. Misra, D.: Mish: A self regularized non-monotonic activation function. arXiv preprint arXiv:1908.08681 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.565, + 0.785, + 0.591 + ], + "angle": 0, + "content": "53. Ozdenizci, O., Legenstein, R.: Restoring vision in adverse weather conditions with patch-based denoising diffusion models. IEEE TPAMI (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.592, + 0.785, + 0.632 + ], + "angle": 0, + "content": "54. Park, D., Lee, B.H., Chun, S.Y.: All-in-one image restoration for unknown degradations using adaptive discriminative filters for specific degradations. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.633, + 0.785, + 0.674 + ], + "angle": 0, + "content": "55. Paszke, A., Gross, S., Massa, F., Lerer, A., Bradbury, J., Chanan, G., Killeen, T., Lin, Z., Gimelshein, N., Antiga, L., et al.: Pytorch: An imperative style, high-performance deep learning library. NeurIPS (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.675, + 0.785, + 0.702 + ], + "angle": 0, + "content": "56. Patil, P.W., Gupta, S., Rana, S., Venkatesh, S., Murala, S.: Multi-weather image restoration via domain translation. In: ICCV (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.702, + 0.785, + 0.73 + ], + "angle": 0, + "content": "57. Qian, R., Tan, R.T., Yang, W., Su, J., Liu, J.: Attentive generative adversarial network for raindrop removal from a single image. In: CVPR (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.73, + 0.785, + 0.757 + ], + "angle": 0, + "content": "58. Quan, R., Yu, X., Liang, Y., Yang, Y.: Removing raindrops and rain streaks in one go. In: CVPR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.758, + 0.785, + 0.785 + ], + "angle": 0, + "content": "59. Quan, Y., Deng, S., Chen, Y., Ji, H.: Deep learning for seeing through window with raindrops. In: ICCV (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.786, + 0.785, + 0.813 + ], + "angle": 0, + "content": "60. Ren, S., He, K., Girshick, R., Sun, J.: Faster r-cnn: Towards real-time object detection with region proposal networks. NeurIPS (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.814, + 0.785, + 0.84 + ], + "angle": 0, + "content": "61. Ren, W., Tian, J., Han, Z., Chan, A., Tang, Y.: Video desnowing and deraining based on matrix decomposition. In: CVPR (2017)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.147, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.356, + 0.128 + ], + "angle": 0, + "content": "S. Sun et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.225, + 0.147, + 0.785, + 0.175 + ], + "angle": 0, + "content": "62. Ren, W., Liu, S., Zhang, H., Pan, J., Cao, X., Yang, M.H.: Single image dehazing via multi-scale convolutional neural networks. In: ECCV (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.225, + 0.176, + 0.785, + 0.203 + ], + "angle": 0, + "content": "63. Ren, W., Ma, L., Zhang, J., Pan, J., Cao, X., Liu, W., Yang, M.H.: Gated fusion network for single image dehazing. In: CVPR (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.225, + 0.204, + 0.785, + 0.231 + ], + "angle": 0, + "content": "64. Shao, Y., Li, L., Ren, W., Gao, C., Sang, N.: Domain adaptation for image dehazing. In: CVPR (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.225, + 0.232, + 0.785, + 0.259 + ], + "angle": 0, + "content": "65. Song, Y., He, Z., Qian, H., Du, X.: Vision transformers for single image dehazing. IEEE TIP (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.225, + 0.26, + 0.785, + 0.286 + ], + "angle": 0, + "content": "66. Sun, S., Ren, W., Li, J., Zhang, K., Liang, M., Cao, X.: Event-aware video deraining via multi-patch progressive learning. IEEE TIP (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.225, + 0.287, + 0.785, + 0.314 + ], + "angle": 0, + "content": "67. Sun, S., Ren, W., Wang, T., Cao, X.: Rethinking image restoration for object detection. NeurIPS (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.225, + 0.315, + 0.785, + 0.355 + ], + "angle": 0, + "content": "68. Tan, Z., Wu, Y., Liu, Q., Chu, Q., Lu, L., Ye, J., Yu, N.: Exploring the application of large-scale pre-trained models on adverse weather removal. arXiv preprint arXiv:2306.09008 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.225, + 0.356, + 0.785, + 0.383 + ], + "angle": 0, + "content": "69. Tu, Z., Talebi, H., Zhang, H., Yang, F., Milanfar, P., Bovik, A., Li, Y.: Maxim: Multi-axis mlp for image processing. In: CVPR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.225, + 0.384, + 0.785, + 0.411 + ], + "angle": 0, + "content": "70. Valanarasu, J.M.J., Yasarla, R., Patel, V.M.: Transweather: Transformer-based restoration of images degraded by adverse weather conditions. In: CVPR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.225, + 0.412, + 0.785, + 0.452 + ], + "angle": 0, + "content": "71. Wan, Y.C., Shao, M.W., Cheng, Y.S., Liu, Y.X., Bao, Z.Y., Meng, D.Y.: Restoring images captured in arbitrary hybrid adverse weather conditions in one go. arXiv preprint arXiv:2305.09996 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.225, + 0.453, + 0.785, + 0.493 + ], + "angle": 0, + "content": "72. Wang, T., Zhang, K., Shao, Z., Luo, W., Stenger, B., Lu, T., Kim, T.K., Liu, W., Li, H.: Gridformer: Residual dense transformer with grid structure for image restoration in adverse weather conditions. arXiv preprint arXiv:2305.17863 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.225, + 0.494, + 0.785, + 0.522 + ], + "angle": 0, + "content": "73. Wang, T., Yang, X., Xu, K., Chen, S., Zhang, Q., Lau, R.W.: Spatial attentive single-image deraining with a high quality real rain dataset. In: CVPR (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.225, + 0.523, + 0.785, + 0.549 + ], + "angle": 0, + "content": "74. Wang, Y., Ma, C., Liu, J.: Smartassign: Learning a smart knowledge assignment strategy for deraining and desnowing. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.225, + 0.55, + 0.785, + 0.577 + ], + "angle": 0, + "content": "75. Wang, Z., Cun, X., Bao, J., Zhou, W., Liu, J., Li, H.: Uformer: A general u-shaped transformer for image restoration. In: CVPR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.225, + 0.578, + 0.785, + 0.604 + ], + "angle": 0, + "content": "76. Wei, W., Yi, L., Xie, Q., Zhao, Q., Meng, D., Xu, Z.: Should we encode rain streaks in video as deterministic or stochastic? In: ICCV (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.225, + 0.606, + 0.785, + 0.633 + ], + "angle": 0, + "content": "77. Wu, H., Qu, Y., Lin, S., Zhou, J., Qiao, R., Zhang, Z., Xie, Y., Ma, L.: Contrastive learning for compact single image dehazing. In: CVPR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.225, + 0.634, + 0.785, + 0.66 + ], + "angle": 0, + "content": "78. Xiao, J., Fu, X., Liu, A., Wu, F., Zha, Z.J.: Image de-raining transformer. IEEE TPAMI (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.225, + 0.661, + 0.785, + 0.688 + ], + "angle": 0, + "content": "79. Yang, W., Tan, R.T., Feng, J., Liu, J., Guo, Z., Yan, S.: Deep joint rain detection and removal from a single image. In: CVPR (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.225, + 0.689, + 0.785, + 0.716 + ], + "angle": 0, + "content": "80. Yasarla, R., Patel, V.M.: Uncertainty guided multi-scale residual learning-using a cycle spinning cnn for single image de-raining. In: CVPR (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.225, + 0.717, + 0.785, + 0.743 + ], + "angle": 0, + "content": "81. Yasarla, R., Sindagi, V.A., Patel, V.M.: Syn2real transfer learning for image deraining using gaussian processes. In: CVPR (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.225, + 0.744, + 0.785, + 0.771 + ], + "angle": 0, + "content": "82. Ye, T., Chen, S., Bai, J., Shi, J., Xue, C., Jiang, J., Yin, J., Chen, E., Liu, Y.: Adverse weather removal with codebook priors. In: ICCV (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.225, + 0.772, + 0.785, + 0.799 + ], + "angle": 0, + "content": "83. You, S., Tan, R.T., Kawakami, R., Mukaigawa, Y., Ikeuchi, K.: Adherent raindrop modeling, detection and removal in video. IEEE TPAMI (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.225, + 0.8, + 0.785, + 0.826 + ], + "angle": 0, + "content": "84. Yu, F., Koltun, V.: Multi-scale context aggregation by dilated convolutions. arXiv preprint arXiv:1511.07122 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.225, + 0.827, + 0.785, + 0.84 + ], + "angle": 0, + "content": "85. Yu, F., Koltun, V., Funkhouser, T.: Dilated residual networks. In: CVPR (2017)" + }, + { + "type": "list", + "bbox": [ + 0.225, + 0.147, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.22, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Restoring Images in Adverse Weather Conditions via Histogram Transformer" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "19" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.147, + 0.785, + 0.175 + ], + "angle": 0, + "content": "86. Zamir, S.W., Arora, A., Khan, S., Hayat, M., Khan, F.S., Yang, M.H.: Restormer: Efficient transformer for high-resolution image restoration. In: CVPR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.177, + 0.785, + 0.203 + ], + "angle": 0, + "content": "87. Zamir, S.W., Arora, A., Khan, S., Hayat, M., Khan, F.S., Yang, M.H., Shao, L.: Multi-stage progressive image restoration. In: CVPR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.204, + 0.785, + 0.231 + ], + "angle": 0, + "content": "88. Zhang, H., Patel, V.M.: Density-aware single image de-raining using a multi-stream dense network. In: CVPR (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.232, + 0.785, + 0.259 + ], + "angle": 0, + "content": "89. Zhang, H., Sindagi, V., Patel, V.M.: Image de-raining using a conditional generative adversarial network. IEEE TCSVT (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.26, + 0.785, + 0.286 + ], + "angle": 0, + "content": "90. Zhang, H., Sindagi, V., Patel, V.M.: Joint transmission map estimation and de-hazing using deep networks. IEEE TCSVT (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.287, + 0.785, + 0.327 + ], + "angle": 0, + "content": "91. Zhang, H., Ba, Y., Yang, E., Mehra, V., Gella, B., Suzuki, A., Pfahnl, A., Chandrappa, C.C., Wong, A., Kadambi, A.: Weatherstream: Light transport automation of single image deweathering. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.328, + 0.785, + 0.356 + ], + "angle": 0, + "content": "92. Zhang, J., Ren, W., Zhang, S., Zhang, H., Nie, Y., Xue, Z., Cao, X.: Hierarchical density-aware dehazing network. IEEE Transactions on Cybernetics (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.357, + 0.785, + 0.383 + ], + "angle": 0, + "content": "93. Zhang, K., Li, D., Luo, W., Ren, W.: Dual attention-in-attention model for joint rain streak and raindrop removal. In: IEEE TIP (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.384, + 0.785, + 0.411 + ], + "angle": 0, + "content": "94. Zhang, K., Li, R., Yu, Y., Luo, W., Li, C.: Deep dense multi-scale network for snow removal using semantic and depth priors. IEEE TIP (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.412, + 0.785, + 0.438 + ], + "angle": 0, + "content": "95. Zhang, X., Li, H., Qi, Y., Leow, W.K., Ng, T.K.: Rain removal in video by combining temporal and chromatic properties. In: ICME (2006)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.439, + 0.785, + 0.466 + ], + "angle": 0, + "content": "96. Zhao, H., Gou, Y., Li, B., Peng, D., Lv, J., Peng, X.: Comprehensive and delicate: An efficient transformer for image restoration. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.467, + 0.785, + 0.494 + ], + "angle": 0, + "content": "97. Zheng, Y., Zhan, J., He, S., Dong, J., Du, Y.: Curricular contrastive regularization for physics-aware single image dehazing. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.495, + 0.785, + 0.522 + ], + "angle": 0, + "content": "98. Zhu, J.Y., Park, T., Isola, P., Efros, A.A.: Unpaired image-to-image translation using cycle-consistent adversarial networks. In: ICCV (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.523, + 0.785, + 0.549 + ], + "angle": 0, + "content": "99. Zhu, L., Fu, C.W., Lischinski, D., Heng, P.A.: Joint bi-layer optimization for single-image rain streak removal. In: ICCV (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.55, + 0.785, + 0.591 + ], + "angle": 0, + "content": "00. Zhu, Y., Wang, T., Fu, X., Yang, X., Guo, X., Dai, J., Qiao, Y., Hu, X.: Learning weather-general and weather-specific features for image restoration under multiple adverse weather conditions. In: CVPR (2023)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.147, + 0.785, + 0.591 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/Restoring Images in Adverse Weather Conditions via Histogram Transformer/95e1934a-8ade-4d7d-8947-f969acbf3910_origin.pdf b/2024/Restoring Images in Adverse Weather Conditions via Histogram Transformer/95e1934a-8ade-4d7d-8947-f969acbf3910_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..8104c106df929694b56f02dabf97ac5e59627516 --- /dev/null +++ b/2024/Restoring Images in Adverse Weather Conditions via Histogram Transformer/95e1934a-8ade-4d7d-8947-f969acbf3910_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c1df9a6aed999180b939a0d53a1e9ecd3e49b017a843c450c7a83ea0c9c7ce57 +size 9780109 diff --git a/2024/Restoring Images in Adverse Weather Conditions via Histogram Transformer/full.md b/2024/Restoring Images in Adverse Weather Conditions via Histogram Transformer/full.md new file mode 100644 index 0000000000000000000000000000000000000000..07ec0566cf23f370d85aaab96e65ce598b48dc14 --- /dev/null +++ b/2024/Restoring Images in Adverse Weather Conditions via Histogram Transformer/full.md @@ -0,0 +1,414 @@ +# Restoring Images in Adverse Weather Conditions via Histogram Transformer + +Shangquan Sun $^{1,2}$ , Wenqi Ren $^{3,4\dagger}$ , Xinwei Gao $^{5}$ , Rui Wang $^{1,2}$ , and Xiaochun Cao $^{3}$ + +$^{1}$ Institute of Information Engineering, Chinese Academy of Sciences, Beijing 100085, China + +$^{2}$ School of Cyber Security, University of Chinese Academy of Sciences, Beijing 100049, China +3 School of Cyber Science and Technology, Shenzhen Campus of Sun Yat-sen University, Shenzhen 518107, China +4 Guangdong Provincial Key Laboratory of Information Security Technology, Guangzhou 510006, China +5 Wechat Business Group, Tencent, Shenzhen, Guangdong, China, 518057 shangquansun@gmail.com, renwq3@mail.sysu.edu.cn, + +Abstract. Transformer-based image restoration methods in adverse weather have achieved significant progress. Most of them use self-attention along the channel dimension or within spatially fixed-range blocks to reduce computational load. However, such a compromise results in limitations in capturing long-range spatial features. Inspired by the observation that the weather-induced degradation factors mainly cause similar occlusion and brightness, in this work, we propose an efficient Histogram Transformer (Histoformer) for restoring images affected by adverse weather. It is powered by a mechanism dubbed histogram self-attention, which sorts and segments spatial features into intensity-based bins. Self-attention is then applied across bins or within each bin to selectively focus on spatial features of dynamic range and process similar degraded pixels of the long range together. To boost histogram self-attention, we present a dynamic-range convolution enabling conventional convolution to conduct operation over similar pixels rather than neighbor pixels. We also observe that the common pixel-wise losses neglect linear association and correlation between output and ground-truth. Thus, we propose to leverage the Pearson correlation coefficient as a loss function to enforce the recovered pixels following the identical order as ground-truth. Extensive experiments demonstrate the efficacy and superiority of our proposed method. We have released the codes in Github. + +Keywords: Image restoration $\cdot$ Adverse weather removal $\cdot$ Image Desnowing $\cdot$ Image deraining $\cdot$ Image dehazing $\cdot$ Raindrop removal + +![](images/f5d18b2d309afc13747df1a5f189e25e71952145d3f26a50075d2f2b1e56e16b.jpg) +(a) Input patches + +![](images/c54f0c05cb88496d0f5f964e58fc5acea9032acea793d3a0604509af6614e25e.jpg) + +![](images/6854636b0234474bb2f3174f8c96938ffde31dae682b51dd88fb2fc8f17382a0.jpg) + +![](images/41d4b01bb1cd440bace641cb76e72248c7aea15e00174e9eb3e8494087ac6a5d.jpg) +(b) Existing self-attention + +![](images/ded64478d816fc6f5118ed522ba8a0ee8b43f7828a84b3052b03ee540c08451e.jpg) +Fig. 1: Given weather-degraded images in (a), traditional transformers perform self-attention either along the channel dimension or within a fixed-range block as shown in (b). In contrast, we observe that weather-induced degradation patterns tend to be similar but distinct from the background. So we categorize pixels affected by adverse weather and background pixels into distinct bins based on descending intensities (as depicted in (c)) and then conducts self-attention within and between bins. + +![](images/f57e0df975ba7aaeed41ebcf7f3d3ee334dc05603a2c913986d306c42d7285d2.jpg) +(c) Histogram self-attention + +# 1 Introduction + +The field of computer vision witnessed growing interest in restoring images affected by adverse weather conditions like rain, fog, and snow. These weather conditions significantly degrade visual quality, impacting the performance of downstream tasks such as object detection [3,60], and depth estimation [16,18]. The restoration of images under adverse weather is thereby a vital problem for the sake of vision aesthetics and safety. + +Early works leverage weather-related priors to model statistical characteristic of degradation and remove adverse weathers [1, 2, 22, 23, 35, 76, 83, 95, 99]. Subsequently, convolutional neural networks (CNNs) have emerged as powerful tools for addressing deraining [5, 17, 34, 40, 57, 73, 79, 80, 88, 89], dehazing [27, 29, 62-64, 77, 88, 90, 92] and desnowing [30, 44, 61, 94]. However, the need of separately training networks for each task and the complexity of switching among multiple models present challenges for real-world applications. Li et al. [33] thus introduced the challenge of adverse weather removal, which entails the restoration of images affected by various weather conditions using a single unified model. + +Recently, Transformer-based approaches have also been investigated for the adverse weather removal task, surpassing the efficacy of CNNs [11, 19, 70, 72]. Nonetheless, these Transformer-based methods usually make concessions regarding efficient memory utilization by confining self-attention operations to a fixed spatial range or solely within the channel dimension, as depicted in Figure 1b. This compromise impedes the inherent potential of Transformers, which was originally designed for superior global feature modeling, and consequently, it leads to a deterioration in the performance of restoration. + +To address these problems, based on the observation that weather-induced degradation often exhibits common patterns shown in Figure 1a, we develop an efficient Histogram Transformer for unified adverse weather removal, named + +Histoformer. Specifically, we introduce a Dynamic-range Histogram Self-Attention (DHSA) module, which endows self-attention with a dynamic-range spatial receptive field. We categorize pixel values proximate in intensity yet varied in spatial location into histogram bins. Self-attention is executed across the dimension of bin or frequency, whose process is illustrated in Figure 1c. To facilitate comprehensive feature extraction on both local and global scales, we devise two ways of reshaping for histogram self-attention: bin-wise histogram reshaping (BHR) and frequency-wise histogram reshaping (FHR). In BHR, the number of bins is configured to incorporate pixels spanning a more comprehensive intensity range, thereby facilitating global feature integration. In FHR, the number of frequencies is assigned such that each bin focuses on limited number of pixels, enhancing the utility of finer features. Consequently, the histogram self-attention attains the capability of modeling spatially dynamic ranges effectively. + +To enable the convolution to extract dynamically-located weather-related dependencies, we develop a dynamic-range convolution layer, which involves sequential horizontal and vertical pixel sorting prior to the application of separable convolution. In order to capture multi-scale and multi-range information embedded within feature matrices, we introduce a Dual-scale Gated Feed-Forward (DGFF) module, enhancing its ability to model the visual characteristics effectively. Additionally, we note that conventional loss functions primarily focus on pixel-level closeness, overlooking the correlation at overall patch level. Consequently, we propose to leverage the Pearson correlation coefficient [12] to ensure the reconstruction of the linear relationship between restored and clean images. + +Our contribution can be summarized in three folds: + +- We propose a novel transformer targeted for unified adverse weather removal, equipped with a new histogram self-attention. It possesses dynamic-range spatial attention to weather-induced obstructions and thus can achieve degradation removal globally and efficiently. +- To capture multi-range information, we present a dual-scale feed-forward module. To enhance the comprehensive linear association between the recovered and ground-truth images, we develop a correlation loss. +- Our method attains state-of-the-art performance across various datasets. Additionally, we substantiate the efficacy of the proposed approach to restore real-world images and bolster the downstream application of detection. + +# 2 Related Work + +Extensive research has been dedicated to addressing adverse weather removal challenges in computer vision, including tasks like deraining [5,17,34,40,66,73, 79,80,88,89], dehazing [27,29,62-64,67,77,88,90,92], desnowing [30,44,61,94], raindrop removal [57,59,83,93] and All-in-One weather removal [28,33,53,70]. + +Rain Streak Removal. The evolution of approaches is notable in rain streak removal techniques in computer vision. Kang et al. [23] pioneered a single image + +deraining method using bilateral filters to decompose images into low and high-frequency components. However, recent advancements have seen a dominance of deep neural networks. An early deep CNN was introduced by Fu et al. [17] for extracting features from the high-frequency rain component, while Yang et al. [79] utilized recurrent networks to decompose rain layers and remove various streak types. Li et al. [32] proposed a method that addresses rain streaks and veiling effects in heavy rain scenes by integrating physics-based rain models and adversarial learning. A conditional generative adversarial network was also employed to solve rain streak removal [89]. Yasarla et al. [81] explored Gaussian processes for transfer learning from synthetic to real-world rain data. Quan et al. [58] used a cascaded network to remove both rain streaks and raindrops. Recently an image deraining Transformer [78] featuring a dual Transformer architecture was intricately formulated, incorporating both window-based and spatial-based mechanisms, thereby attaining exemplary outcomes. A sparse deraining Transformer is also proposed to enhance feature aggregation [11]. + +Raindrop Removal. Raindrop removal from single images has been addressed through various methods, with some relying on traditional hand-crafted features. An early work incorporated temporal information to address video-based raindrop removal [83]. Eigen et al. [15] employed a shallow CNN trained with image pairs containing raindrop-degraded and raindrop-free versions, though the results often exhibited blurriness. Qian et al. [57] introduced an attention GAN and a new dataset. Their method was later improved by Quan et al. [59] via generating attention maps based on mathematical raindrop descriptions and combining them with detected raindrop edges. + +Snow Removal. Desnow-Net [44] was among the pioneering CNN-based approaches for snow removal, followed by Li et al.'s stacked dense network [31] and Chen et al.'s JSTASR [8], which introduced a size and transparency aware method. More recently, DDMSNet [94] introduced a dense multi-scale network that leverages semantic and geometric priors to enhance snow removal. A hierarchical decomposition paradigm involving the dual-tree wavelet transform for snow removal is also proposed [9]. Chen et al. [7] designed SnowFormer, a framework that used cross-attention to establish local-global context interaction. + +Fog Removal. Li et al. [26] presented a CNN that takes into account both atmospheric luminosity and transmission maps to conduct dehazing. Ren et al. [63] advocated a pre-processing approach for hazy image manipulation, thereby engendering multiple input modalities and, in the process, inducing chromatic aberrations as part of their dehazing procedure. A hierarchical density-aware network is also introduced, specializing in the domain of image dehazing [92]. Zheng et al. [97] formulated a curriculum-based contrastive regularization dehazing method aimed at fostering agreement within a contrastive space. + +All-in-One Weather Removal. Some recent works attempted to address various weather-induced degradations by a singular network. Li et al. [33] proposed an + +All-in-One network, containing a generator comprising multiple task-specific encoders and a shared decoder. Valanarasu et al. [70] presented TransWeather, a transformer-based model featuring a solitary encoder-decoder structure, capable of rejuvenating images afflicted by various atmospheric conditions. A pipeline for the automatic selection of weather-degraded data was also proposed to enhance existing models [91]. Zhu et al. developed WGWS-Net [100] capable of learning weather-general and weather-specific in two separate stages. Some other recent works also trial addressing adverse weather removal by adopting probabilistic denoising diffusion model [53], knowledge distillation [10], large-scale Pre-trained model [68], mixture of experts [49], few-shot learning [24], codebooks [41,71,82], adaptive filters [54], knowledge assignment [74] and domain translation [56]. + +Transformer-based Image Restoration. Since the inception of the Vision Transformer (ViT) [14] for visual recognition, transformers have gained substantial traction across a spectrum of computer vision tasks [25, 37, 45, 46, 50, 51, 86]. Particularly within the realm of low-level vision, the Image Processing Transformer [4] exemplifies how pre-training a transformer on extensive datasets can significantly enhance performance for low-level applications. U-former [75], on the other hand, introduced a transformer architecture based on the U-Net design for restoration tasks. Swin-IR [38] employed the Swin Transformer [45] for image restoration. Some latest Transformer-based methods were proposed for deraining [11, 39], desnowing [7], dehazing [19, 43, 65] and All-in-One weather removal [70, 72]. + +Unlike the existing Transformer-based approaches whose self-attention is applied within either fixed spatial ranges or merely channel dimension, our method enables dynamic-range spatial attention to adaptively focus on weather-induced degradation with similar patterns. + +# 3 Method + +# 3.1 Overall Architecture + +The overall architectural framework of our Histoformer is illustrated in Figure 2. Suppose the input is a low-quality image $I^{lq} \in \mathbb{R}^{3 \times H \times W}$ , we pass the input through a $3 \times 3$ convolution to conduct the overlapping image patch-embedding. Within both the encoder and decoder of the network backbone, we arrange Histogram Transformer Blocks (HTBs) to extract intricate features and capture dynamically distributed degradation factors. Within the same stage, encoders and decoders are interlinked through skip-connections, thereby establishing connections between consecutive intermediate features to enhance the stability of the training process. Between each stage, we apply pixel-unshuffle and pixelshuffle operations for the purpose of feature down-sampling and up-sampling. + +Within each HTB, we introduce Dynamic-range Histogram Self-Attention (DHSA) to extract spatially dynamic weather degradation and enhance both local and global feature aggregation. Moreover, a Dual-scale Gated Feed-Forward + +![](images/ea63ebb64c2fdd34c12ae0e50e98c82e17d225bfd71eb58ff59d6cd2d8b773fa.jpg) +Fig. 2: The overall architecture of our Histoformer for weather removal. The main component is the Histogram Transformer block, and it comprises the Dynamic-range Histogram Self-Attention (DHSA) module and the Dual-scale Gated Feed-Forward (DGFF) module. Within DHSA, we present two types of reshaping mechanism, i.e., Bin-wise Histogram Reshaping and Frequency-wise Histogram Reshaping. + +(DGFF) module is integrated into the HTB to enrich the representation of multi-range features, contributing to the process of image restoration. During each stage of encoding phases, our model is equipped with a crude skip-connection for supplementing original features from input, comprised of a sequence of operations, including average pooling, pixel-wise convolution, and depth-wise convolution. We start the crude skip-connection after the first stage, and this setup enables the encoders to focus more effectively on learning the weather-induced residuals. Through this hybrid formulation, Histoformer is empowered to exploit both the adaptive contents of weather-irrelevant background and the inherent characteristics of weather-degraded patterns, facilitating the separation of undesired degradation from the latent clear background. + +# 3.2 Histogram Transformer Block + +As the key component of our Histoformer, HTB contains two pivotal modules, i.e., DHSA and DGFF. These two components are arranged to interact with layer normalization and can be formulated as follows: + +$$ +F _ {l} = F _ {l - 1} + \mathrm {D H S A} (\mathrm {L N} (F _ {l - 1})), \tag {1} +$$ + +$$ +F _ {l} = F _ {l} + \operatorname {D G F F} \left(\ln \left(F _ {l}\right)\right), \tag {2} +$$ + +where LN denotes layer normalization and $F_{l}$ represents the feature at $l$ -th stage. The details of DHSA and DGFF are presented in Section 3.2 and 3.2 respectively. + +Dynamic-range Histogram Self-Attention To better capture dynamically distributed weather-induced degradation, we develop a Dynamic-range Histogram + +Self-Attention (DHSA) module. This module consists of a process involving dynamic-range convolution, which reorders the spatial distribution of fractional features, and a dual-path histogram self-attention mechanism that combines global and local dynamic feature aggregation. Prior to the final output projection of a $1 \times 1$ point-wise convolution, the reordered features are sorted back into their original locations to maintain spatial consistency. + +Dynamic-range Convolution. Traditional convolution operations employ fixed kernel sizes, resulting in a limited receptive field range and consequently perform local and small-range computations. This restricted scope of convolution, which primarily focuses on local information, does not naturally align with the self-attention mechanism's capacity to model long-range dependencies. To address this limitation, we devise a dynamic-range convolution technique by meticulously reordering the input features prior to the traditional convolution operation. Given an input feature $F \in \mathbb{R}^{C \times H \times W}$ , we divide it into two branches along the channel dimension, namely $F_{1}$ and $F_{2}$ . For the first branch of features, we perform sorting operations both horizontally and vertically, subsequently concatenating the sorted features with the second branch of feature. The resulting recombined features are then passed through the subsequent separable convolution. The entire process is articulated as follows: + +$$ +F _ {1}, F _ {2} = \operatorname {S p l i t} (F), F _ {1} = \operatorname {S o r t} _ {\mathrm {v}} (\operatorname {S o r t} _ {\mathrm {h}} (F _ {1})), +$$ + +$$ +\begin{array}{l} F = \operatorname {C o n v} _ {3 \times 3} ^ {\mathrm {d}} \left(\operatorname {C o n v} _ {1 \times 1} \left(\operatorname {C o n c a t} \left(F _ {1}, F _ {2}\right)\right)\right), \end{array} \tag {3} +$$ + +where $\mathrm{Conv}_{1\times 1}$ is $1\times 1$ point-wise convolution, $\mathrm{Conv}_{3\times 3}^{\mathrm{d}}$ represents $3\times 3$ depthwise convolution, Concat is the concatenation operation along channel, Split denotes the operation of splitting features along channel dimension, and $\mathrm{Sort}_{i\in \{\mathrm{h},\mathrm{v}\}}$ represents the horizontal or vertical sorting operation. This approach organizes pixels of high and low intensities into regular patterns at the diagonal corners of the matrices, thereby allowing convolution to perform computations across dynamic ranges. Given that weather-induced degradation typically exhibits closely related patterns, degraded pixels tend to concentrate in neighboring locations, separated with those clean pixels. As a result, this arrangement enables convolution kernels to partially focus on preserving clean information and separately recovering degraded features. + +Histogram Self-Attention. Existing vision Transformers [11,75,78,78,86,96] typically leverage fixed range of attention or merely the attention along channel dimension due to the compromise of computation and memory efficiency. However, the fixed setting restricts the self-attention to span adaptively long range to associate desired features. We notice that weather-induced degradation causes similar patterns and that those pixels containing either background feature or weather degradation of different intensities had better be assigned with various extents of attention. We thus propose a histogram self-attention mechanism to categorize spatial elements into bins and allocate varying attention within and across bins. For the sake of parallel computing, we set each bin contains identical number of pixels during implement. + +Given the output of dynamic-range convolution, we can separate them into Value feature $V \in \mathbb{R}^{C \times H \times W}$ and two pairs of Query-key $F_{QK,1}, F_{QK,2} \in \mathbb{R}^{2C \times H \times W}$ subsequently passing to two branches. We firstly sort the sequence of $V$ and based on its index arrange the Query-Key pairs accordingly, expressed as follows: + +$$ +V, d = \operatorname {S o r t} \left(\mathbf {R} _ {\mathrm {C} \times \mathrm {H} \times \mathrm {W}} ^ {\mathrm {C} \times \mathrm {H W}} (V)\right), +$$ + +$$ +Q _ {1}, K _ {1} = \operatorname {S p l i t} \left(\operatorname {G a t h e r} \left(\mathbf {R} _ {\mathrm {C} \times \mathrm {H} \times \mathrm {W}} ^ {\mathrm {C} \times \mathrm {H W}} \left(F _ {Q K, 1}\right), d\right)\right), \tag {4} +$$ + +$$ +Q _ {2}, K _ {2} = \operatorname {S p l i t} \left(\operatorname {G a t h e r} \left(\mathbf {R} _ {\mathrm {C} \times \mathrm {H} \times \mathrm {W}} ^ {\mathrm {C} \times \mathrm {H W}} (F _ {Q K, 2}), d\right)\right), +$$ + +where $\mathbf{R}_{\mathrm{C}\times ,\mathrm{H}\times \mathrm{W}}^{\mathrm{C}\times \mathrm{HW}}$ represents the operation of reshaping features from $\mathbb{R}^{C\times H\times W}$ to $\mathbb{R}^{C\times HW}$ , $d$ is the index of sorted Value, and Gather denotes the operation of retrieving elements of tensor based on a given index. + +Then given the number of bins $B$ , we reshape the sorted features from $C \times HW$ into $C \times B \times HW / B$ . To extract both global and local information, we define two types of reshaping, i.e., bin-wise histogram reshaping (BHR) and frequency-wise histogram reshaping (FHR). The first is to assign the number of bins equal to $B$ and each bin contains $HW / B$ elements, while the second is to set the frequency of each bin equal to $B$ and the number of bins is $HW / B$ . By this way, we can extract large-scale information by BHR where each bin contains large number of dynamically located pixels and fine-grained information by FHR where each bins contains modicum pixels neighboring in terms of intensity. The two pairs of Query-Key features are passed through two types of reshaping and subsequent self-attention process respectively, and their outputs are elementwisely multiplied to yield the final output. The process can be formulated as the following expressions: + +$$ +\mathrm {A _ {B}} = \mathrm {s o f t m a x} \left(\frac {\mathbf {R _ {B}} (Q _ {1}) \mathbf {R _ {B}} (K _ {1}) ^ {\top}}{\sqrt {k}}\right) \mathbf {R _ {B}} (V), +$$ + +$$ +\mathrm {A} _ {\mathrm {F}} = \operatorname {s o f t m a x} \left(\frac {\mathbf {R} _ {\mathrm {F}} \left(Q _ {2}\right) \mathbf {R} _ {\mathrm {F}} \left(K _ {2}\right) ^ {\top}}{\sqrt {k}}\right) \mathbf {R} _ {\mathrm {F}} (V), \tag {5} +$$ + +$$ +\mathrm {A} = \mathrm {A} _ {\mathrm {B}} \odot \mathrm {A} _ {\mathrm {F}}, +$$ + +where $k$ is the number of heads, $\mathbf{R}_{i\in \{\mathrm{B},\mathrm{F}\}}$ denotes the reshaping operation of either BHR or FHR, and $\mathrm{A}_{i\in \{\mathrm{B},\mathrm{F}\}}$ represents the obtained attention map. + +Dual-scale Gated Feed-Forward Previous studies [11, 75, 78, 86] typically leverage single-range or single-scale convolution in the standard feed-forward network to bolster local context. Nonetheless, these methods often disregard the correlations among dynamically distributed weather-induced degradation. In practice, multi-scale information can be extracted by not only enlarging the kernel size but also leveraging the dilation mechanism [36, 84, 85]. As a result, we conceive a Dual-scale Gated Feed-Forward (DGFF) module, which integrates two distinct multi-range and multi-scale depth-wise convolution pathways within the transmission process. + +Table 1: Quantitative comparisons on three weather removal tasks in terms of PSNR and SSIM, where higher values indicate better performance. The top halves of tables display the results of task-specific methods, while the bottom halves present evaluations of the unified multi-weather models. The best and the second best results are in bold and underlined. Those with * indicate the methods whose source codes are unavailable. +(a) Image Desnowing +(b) Deraining & Dehazing +(c) Raindrop Removal + +
Snow100K-S [44]Snow100K-L [44]Outdoor-Rain [32]RainDrop [57]
PSNRSSIMPSNRSSIMPSNRSSIMPSNR SSIM
SPANet [73]29.920.826023.700.7930CycleGAN [98]17.620.6560pix2pix [20]28.02
JSTASR [8]31.400.901225.320.8076pix2pix [20]19.090.7100DuRN [42]31.24
RESCAN [34]31.510.903226.080.8108HRGAN [32]21.560.8550RaindropAttn [59]31.44
DesnowNet [44]32.330.950027.170.8983PCNet [21]26.190.9015AttentiveGAN [57]31.59
DDMSNet [94]34.340.944528.850.8772MPRNet [87]28.030.9192IDT [78]31.87
NAFNet [6]34.790.949730.060.9017NAFNet [6]29.590.9027MAXIM [69]31.87
Restormer [86]36.010.957930.360.9068Restormer [86]30.030.9215Restormer [86]32.18
All-in-One [33]*--28.330.8820All-in-One [33]*24.710.8980All-in-One [33]*31.12
TransWeather [70]32.510.934129.310.8879TransWeather [70]28.830.9000TransWeather [70]30.17
Chen et al. [10]34.420.946930.220.9071Chen et al. [10]29.270.9147Chen et al. [10]31.81
WGWSNet [100]34.310.946030.160.9007WGWSNet [100]29.320.9207WGWSNet [100]32.38
WeatherDiff64 [53]35.830.956630.090.9041WeatherDiff64 [53]29.640.9312WeatherDiff64 [53]30.71
WeatherDiff128 [53]35.020.951629.580.8941WeatherDiff128 [53]29.720.9216WeatherDiff128 [53]29.66
AWRCP [82]*36.920.965231.920.9341AWRCP [82]*31.390.9329AWRCP [82]*31.93
Histoformer (Ours)37.410.965632.160.9261Histoformer (Ours)32.080.9389Histoformer (Ours)33.06
+ +Given an input tensor $F_{l} \in \mathbb{R}^{C \times H \times W}$ , we initially employ a point-wise convolution operation to augment the channel dimension by a factor of $r$ . Following this augmentation, the expanded tensor is directed into two parallel branches. Throughout the feature transformation process, $5 \times 5$ and dilated $3 \times 3$ depth-wise convolutions are employed to enhance the extraction of multi-range and multi-scale information. Following the gating mechanism [13], the output of the second branch after passing through an activation act as a gating map for the other branch. Thus, the complete feature fusion process within the DGFF module is formulated as follows: + +$$ +F _ {l, 1}, F _ {l, 2} = \operatorname {S p l i t} \left(\operatorname {S h u f f l e} \left(\operatorname {C o n v} _ {1 \times 1} \left(F _ {l}\right)\right)\right), +$$ + +$$ +F _ {l, 1} = \operatorname {C o n v} _ {5 \times 5} ^ {\mathrm {d}} \left(F _ {l, 1}\right), F _ {l, 2} = \operatorname {C o n v} _ {3 \times 3} ^ {\mathrm {d}, \text {d i l a t e d}} \left(F _ {l, 2}\right), \tag {6} +$$ + +$$ +F _ {l + 1} = \operatorname {C o n v} _ {1 \times 1} \left(\text {U n s h u r f l e} \left(\operatorname {M i s h} \left(F _ {l, 2}\right) \odot F _ {l, 1}\right)\right), +$$ + +where $\mathrm{Conv}_{5\times 5}^{\mathrm{d}}$ represents $5\times 5$ depth-wise convolution, $\mathrm{Conv}_{3\times 3}^{\mathrm{d,dilated}}$ is $3\times 3$ dilated depth-wise convolution, Shuffle and Unshuffle represent respectively the operations of pixel-shuffling and unshuffling, Mish denotes the Mish activation [52], and $F_{l + 1}$ is the output of current stage passing to $l + 1$ -th stage. + +# 3.3 Reconstruction Loss and Correlation Loss + +We use the $L_{1}$ norm of the pixel-wise difference between the restored high-quality image $I^{hq}$ and ground-truth $I^{gt}$ as the reconstruction loss, i.e., + +$$ +\mathcal {L} _ {r e c} = \left\| I ^ {h q} - I ^ {g t} \right\| _ {1}. \tag {7} +$$ + +Furthermore, we notice that the $\mathcal{L}_{rec}$ only regulates the pixel-level similarity between the restored image and the ground-truth, while neglecting the patch-level linear correlations. The innate relationships of intensity within the image are disrupted by the consistent patterns of weather-induced degradation. By + +![](images/baca57d25cae884c90257b422de3eadf3d5b76bd5cc7a3fbbfb41c8760dd389c.jpg) +Fig. 3: Visual comparison for desnowing on Snow100K [44]. The samples from (b) to (e) are Restormer [86], TransWeather [70], WGWSNet [100], WeatherDiff [53]. + +![](images/208d7bedddefa74b11ee3cca8783fa72640b2f53b4813c1f8dc3168f5a3dc5da.jpg) +Fig. 4: Visual comparison for deraining and dehazing on Outdoor-Rain [32]. The samples from (b) to (e) are Restormer [86], TransWeather [70], WGWSNet [100], WeatherDiff [53]. + +emulating the intensity relationships within the ground-truth, we compel the degraded pixels to occupy their original positions according to the original intensity ranking. Consequently, we introduce the Pearson correlation [12] between images as a means to regulate the linear relationship, expressed as follows: + +$$ +\rho \left(I ^ {h q}, I ^ {g t}\right) = \frac {\sum_ {i = 1} ^ {3 H W} \left(I _ {i} ^ {h q} - \bar {I} ^ {h q}\right) \left(I _ {i} ^ {g t} - \bar {I} ^ {g t}\right)}{3 H W \sigma \left(I ^ {h q}\right) \sigma \left(I ^ {g t}\right)}, \tag {8} +$$ + +where $I_{i}^{\{\cdot\}}$ represents the $i$ -th pixel of image, $\overline{I}^{\{\cdot\}}$ and $\sigma(I^{\{\cdot\}})$ denotes respectively the mean and the standard deviation of image sequence. Its value falls within the $[-1,1]$ range. When two images exhibit perfect correlation, the value of function $\rho$ attains a value of 1, while in the case of negative correlation, its value reaches $-1$ . Hence, we formulate the correlation loss as follows: + +$$ +\mathcal {L} _ {c o r} = \frac {1}{2} \left(1 - \rho (I ^ {h q}, I ^ {g t})\right), \tag {9} +$$ + +such that $\mathcal{L}_{cor} = 0$ when the recovered image perfectly aligns with the groundtruth. The overall loss function is thus defined as: + +$$ +\mathcal {L} = \mathcal {L} _ {\text {r e c}} + \alpha \mathcal {L} _ {\text {c o r}}, \tag {10} +$$ + +where $\alpha$ is the weight of correlation loss. + +# 4 Experiments + +# 4.1 Experimental settings + +Datasets. We train our model on the same datasets as the previous works [33,53, 70] to ensure a fair comparison. The training set encompasses 9,000 images drawn + +![](images/19d278cb330ff500c448ffe6ea4cb25b75a15428a02051e0c1816a0eb714e19c.jpg) +Fig. 5: Visual comparison for raindrop removal on RainDrop [57]. The samples from (b) to (e) are Chen et al. [10], TransWeather [70], WGWSNet [100], WeatherDiff [53]. + +![](images/fac601967372a43f1453bd65a58b6f2b02d719857ead9be7b95f65e19102dbf9.jpg) +Fig.6: A qualitative comparison for real-world adverse weather removal on Snow100K [44]. The samples from (b) to (e) are Chen et al. [10], Restormer [86], TransWeather [70], WGWSNet [100], WeatherDiff [53]. + +from Snow100K [44], 1,069 images sourced from Raindrop [57], and 9,000 images from Outdoor-Rain [32]. Snow100K contains synthetic images deteriorated by snow, while Raindrop comprises real raindrop-affected images. Outdoor-Rain features synthetic images afflicted by both fog and rain streaks. For evaluation, we employ the Test1 dataset [32, 33], the RainDrop test dataset [57], and the Snow100K-L and -S test sets [44]. Snow100K also provides a real-world test set containing 1,329 images affected by adverse weather. + +Comparison Baselines. We assess the performance of our method against state-of-the-art approaches designed specifically for distinct weather removal tasks: raindrop removal, snow removal, and rain&fog removal. Specifically, for snow removal, we benchmark against SPANet [73], JSTASR [8], RESCAN [34], DesnowNet [44], and DDMSNet [94]. In the case of rain&fog removal, we compare with CycleGAN [98], pix2pix [20], HRGAN [32], MPRNet [87] and Restormer [86]. For raindrop removal, we evaluate against the methods such as pix2pix [20], DuRN [42], RaindropAttn [59], AttentiveGAN [57]. Additionally, we include some recent transformer or multi-degradation restoration networks, IDT [78], NAFNet [6], MAXIM [69], and Restormer [86], in our comparative analysis. It is worth noting that all these methods are single-task networks fine-tuned for specific datasets. + +Furthermore, we conduct a performance comparison with the All-in-One network [33], Chen et al. [10], TransWeather [70], WGWS-Net [100], WeatherDiff [53] and AWRCP [82], which are trained to handle all the aforementioned tasks using a unified model. Note that our approach is also trained to tackle all these tasks using a single model. + +Training details. Our implementation is realized by PyTorch [55] and on NVIDIA Tesla V100 GPU. The network is trained for a total of 300,000 iterations, with + +an initial batch size of 8 and an initial patch size of 128 akin to the progressive learning pipeline [86]. We employ the AdamW optimizer [48] with an initial learning rate of $3e^{-4}$ for the first 92,000 iterations, which is gradually reduced to $1e^{-6}$ using cosine annealing schedule [47] during the remaining 208,000 iterations. The number of blocks at each stage $L_{i\in \{1,2,3,4\}}$ is set to $\{4,4,6,8\}$ and the channel size $C$ is 36. The channel expansion factor $r$ in DGFF is set to 2.667. The numbers of heads in self-attention at different stages are set to $\{1,2,4,8\}$ respectively. We randomly apply horizontal and vertical flips as the technique of data augmentation. + +# 4.2 Comparisons with the state-of-the-arts + +Quantitative Evaluation. In our study, we provide a comprehensive comparative analysis of metrics applied to both synthetic and real datasets, as summarized in Table 1. For a fair and well-founded comparison, we utilize recent multiple degradation removal methods such as MPRNet [87], MAXIM [69], and Restormer [86], treating them as weather-specific approaches for each benchmark. Additionally, we retrain the all-in-one adverse weather removal methods including Chen et al. [10] and WGWS-Net [100] using the all-weather training dataset [33,53,70]. This exhaustive comparison reveals that our proposed method exhibits a significant performance advantage over existing approaches across three different types of degradation. + +Qualitative Evaluation. Furthermore, we conduct a visual comparison on three tasks, and the outcomes are showcased in Figure 3, 4 and 5 respectively. Figure 6 shows a case of real-world weather removal. These results highlight that our method excels in comprehensively eliminating snow degradation, including fine and large snow spots. In contrast, the recent WeatherDiff [53] method still exhibits some residual snow degradation, and its capability to restore details is not optimal. When it comes to the restoration of challenging weather conditions, our method excels in removing complex haze and rain marks, yielding visually appealing results in comparison to prior approaches. + +# 4.3 Ablation studies + +To substantiate the effectiveness of each component within Histoformer, we conduct a sequence of ablation studies on Outdoor-Rain [32]. In particular, we examine the impact of the dynamic-range convolution, the DHSA module, the number of bins in DHSA, the DGFF module, and the correlation loss. + +Dynamic-range Convolution. We experiment on two settings of dynamic-range convolution, namely, sorting horizontally first and then vertically before convolution, and the reverse order. Additionally, we compared them with vanilla convolution, and the results are displayed in Table 2. The operations of regular sorting led to a performance improvement of $0.14\mathrm{dB}$ , and the order of sorting operations does not significantly affect the outcome. + +Table 2: Ablation studies on the dynamic-range convolution. + +
Vanilla ConvVerti.+Horiz.Horiz.+Verti.
PSNR31.9432.0332.08
SSIM0.93770.93850.9389
+ +Table 3: Ablation studies on the design of self-attention + +
SA typePSNRSSIM
MDTA [86]30.940.9278
TKSA [11]31.120.9295
w/o BHR31.050.9301
w/o FHR31.790.9364
DHSA32.080.9389
+ +Table 4: Ablation studies on the number of Bins in DHSA + +
#Bins in DHSAPSNRSSIM
1230.430.9283
2031.560.9312
2831.940.9379
3632.080.9389
44Out of memory
+ +Table 5: Ablation studies on the choice of feed-forward module + +
Feed-ForwardPSNRSSIM
FFN [38]31.320.9313
GDFN [86]31.420.9347
DANB [96]31.560.9351
MSFN [11]31.780.9367
DGFF32.080.9389
+ +Table 6: Ablation studies on the setting of correlation loss + +
αPSNRSSIM
w/o Lcor031.770.9358
0.132.010.9369
132.080.9389
w/ Lcor532.030.9392
1031.960.9375
+ +DHSA. To evaluate the effectiveness of the proposed DHSA module, we conduct a comparison with two baselines, i.e., a multi-Dconv head transposed attention (MDTA) [86] and a top-k sparse attention (TKSA) [11]. Additionally, we explore two additional settings of DHSA by excluding either the BHR branch or the FHR branch. The quantitative analysis results are presented in Table 3. + +Both MDTA and TKSA integrate rich information across channels, which may result in a loss of the exploitation of long-range information across spatial dimensions. While our histogram self-attention can capture spatially long-range information, the use of either a single BHR or a single FHR branch neglects the inter-bin or inner-bin relationships, leading to inferior results. By incorporating dynamic-range convolution and dual-branch histogram self-attention, capable of extracting long-range spatial features, our DHSA enhances performance, resulting in a PSNR improvement of 0.96 dB compared to TKSA. + +Number of Bins. To assess the influence of the number of bins in DHSA, we conduct experiments with five different values: 12, 20, 28, 36, and 44. The results are presented in Table 4. It is observed that increasing the number of bins consistently improves performance. However, when the number of bins exceeds 44, it results in an out-of-memory error. + +DGFF. To assess the effectiveness of the proposed DGFF module, we conduct a comparison with four baselines: (i) the vanilla feed-forward network (FN) [38], (ii) a gated-Dconv feed-forward network (GDFN) [86], (iii) a dual adaptive neural block (DANB) [96], and (iv) a mixed-scale feed-forward network (MSFN) [11]. + +![](images/a5e963f75412604f02d4b03a668b66117c243845b7a05706c6e8ea9c7ac18246.jpg) + +![](images/acecbfef62a99b4e6ece62c03d9d41d1ea40aca6c7c8e7eaaa87e6bd58c4bcef.jpg) + +![](images/d461570eb7879ef685295620460bbad6f566c6cc5ce108f2c13b0c032e790079.jpg) + +![](images/df5e72bed4f1e2a0ffc1f28f998be7c68d69cb742c8e4d1d289f648c0fb01e12.jpg) + +![](images/3e232858f732d101478c5eb8ab9cf76a84c566659b7fdfadc060e2d6c7240059.jpg) +(a) Input + +![](images/f8a8d3b27a0bc4377b8df0948e2988b05bd55feff38e7802fe57bbbf66e24ac0.jpg) +Fig. 7: Real-world deweathering on two snowy images [44] and their downstream detection results on Google API. + +![](images/54d9cb90a8447db778f7a100b008e5ae1581e9b0278e99464911f06ea7a8afd0.jpg) +(b) Deweathered by ours + +![](images/84a07aeea60f3ad7e74ad0bf2ab70afaa89399fc0746fbbf5d91860ebc41c871.jpg) + +The quantitative analysis results are presented in Table 5. While MSFN integrates mixed-scale information, it may still miss out on the exploitation of multi-range spatial knowledge. Through the inclusion of pixel-shuffling and feature aggregation across different ranges, our DGFF further enhances performance, resulting in a PSNR gain of 0.3 dB over MSFN. + +Correlation Loss. Table 6 shows the effectiveness of the correlation loss $\mathcal{L}_{cor}$ and the influence of its weight. It is evident that $\mathcal{L}_{cor}$ consistently improves the performance, while the specific loss weight does not have a substantial impact on the final results. We therefore keep the loss weight as 1 by default. + +# 4.4 Real-world Application + +To further demonstrate the practical applicability of our method for real-world adverse weather removal and its potential to improve downstream detection task, we provide two samples in Figure 7. As depicted, our Histoformer effectively eliminates snowflakes from the scene and assists the detector in recognizing omitted door and building. + +# 5 Conclusion + +In this research, we introduce a novel mechanism called histogram self-attention and devise a new histogram transformer named Histoformer to tackle the challenge of all-in-one weather removal. Our histogram self-attention involves segmenting spatial features into multiple bins, and allocating varying attention along the bin or frequency dimension, allowing it to selectively focus on weather-related features with a dynamic range. To facilitate learning both multi-range and multi-scale information, we present DGFF module and a correlation loss. Through extensive experimentation, we demonstrate the effectiveness and superiority of our approach. + +# Acknowledgement + +This work has been supported in part by National Natural Science Foundation of China (No. 62322216, 62172409, 62025604, 62306308, 62311530686), in part by Shenzhen Science and Technology Program (Grant No. JCYJ20220818102012025, KQTD20221101093559018), and in part by Guangdong Provincial Key Laboratory of Information Security Technology (No. 2023B1212060026). + +# References + +1. Ancuti, C.O., Ancuti, C.: Single image dehazing by multi-scale fusion. IEEE TIP (2013) +2. Berman, D., Avidan, S., et al.: Non-local image dehazing. In: CVPR (2016) +3. Carion, N., Massa, F., Synnaeve, G., Usunier, N., Kirillov, A., Zagoruyko, S.: End-to-end object detection with transformers. In: ECCV (2020) +4. Chen, H., Wang, Y., Guo, T., Xu, C., Deng, Y., Liu, Z., Ma, S., Xu, C., Xu, C., Gao, W.: Pre-trained image processing transformer. In: CVPR (2021) +5. Chen, J., Tan, C.H., Hou, J., Chau, L.P., Li, H.: Robust video content alignment and compensation for rain removal in a cnn framework. In: CVPR (2018) +6. Chen, L., Chu, X., Zhang, X., Sun, J.: Simple baselines for image restoration. In: ECCV (2022) +7. Chen, S., Ye, T., Liu, Y., Chen, E., Shi, J., Zhou, J.: Snowformer: Scale-aware transformer via context interaction for single image desnowing. arXiv preprint arXiv:2208.09703 (2022) +8. Chen, W.T., Fang, H.Y., Ding, J.J., Tsai, C.C., Kuo, S.Y.: Jstasr: Joint size and transparency-aware snow removal algorithm based on modified partial convolution and veiling effect removal. In: Vedaldi, A., Bischof, H., Brox, T., Frahm, J.M. (eds.) ECCV (2020) +9. Chen, W.T., Fang, H.Y., Hsieh, C.L., Tsai, C.C., Chen, I., Ding, J.J., Kuo, S.Y., et al.: All snow removed: Single image desnowing algorithm using hierarchical dual-tree complex wavelet representation and contradict channel loss. In: ICCV (2021) +10. Chen, W.T., Huang, Z.K., Tsai, C.C., Yang, H.H., Ding, J.J., Kuo, S.Y.: Learning multiple adverse weather removal via two-stage knowledge learning and multi-contrastive regularization: Toward a unified model. In: CVPR (2022) +11. Chen, X., Li, H., Li, M., Pan, J.: Learning a sparse transformer network for effective image deraining. In: CVPR (2023) +12. Cohen, I., Huang, Y., Chen, J., Benesty, J., Benesty, J., Chen, J., Huang, Y., Cohen, I.: Pearson correlation coefficient. Noise reduction in speech processing (2009) +13. Dauphin, Y.N., Fan, A., Auli, M., Grangier, D.: Language modeling with gated convolutional networks. In: ICML (2017) +14. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929 (2020) +15. Eigen, D., Krishnan, D., Fergus, R.: Restoring an image taken through a window covered with dirt or rain. In: ICCV (2013) + +16. Fu, H., Gong, M., Wang, C., Batmanghelich, K., Tao, D.: Deep ordinal regression network for monocular depth estimation. In: CVPR (2018) +17. Fu, X., Huang, J., Zeng, D., Huang, Y., Ding, X., Paisley, J.: Removing rain from single images via a deep detail network. In: CVPR (2017) +18. Godard, C., Mac Aodha, O., Brostow, G.J.: Unsupervised monocular depth estimation with left-right consistency. In: CVPR (2017) +19. Guo, C., Yan, Q., Anwar, S., Cong, R., Ren, W., Li, C.: Image dehazing transformer with transmission-aware 3d position embedding. In: CVPR (2022) +20. Isola, P., Zhu, J.Y., Zhou, T., Efros, A.A.: Image-to-image translation with conditional adversarial networks. In: CVPR (2017) +21. Jiang, K., Wang, Z., Yi, P., Chen, C., Wang, Z., Wang, X., Jiang, J., Lin, C.W.: Rain-free and residue hand-in-hand: A progressive coupled network for real-time image deraining. IEEE TIP (2021) +22. Jiang, T.X., Huang, T.Z., Zhao, X.L., Deng, L.J., Wang, Y.: A novel tensor-based video rain streaks removal approach via utilizing discriminatively intrinsic priors. In: CVPR (2017) +23. Kang, L.W., Lin, C.W., Fu, Y.H.: Automatic single-image-based rain streaks removal via image decomposition. IEEE TIP (2011) +24. Kim, Y., Cho, Y., Nguyen, T.T., Lee, D.: Metaweather: Few-shot weather-degraded image restoration via degradation pattern matching. arXiv preprint arXiv:2308.14334 (2023) +25. Lai, Z., Wu, J., Chen, S., Zhou, Y., Hovakimyan, N.: Residual-based language models are free boosters for biomedical imaging tasks. In: CVPRW (2024) +26. Li, B., Peng, X., Wang, Z., Xu, J., Feng, D.: Aod-net: All-in-one dehazing network. In: ICCV (2017) +27. Li, B., Ren, W., Fu, D., Tao, D., Feng, D., Zeng, W., Wang, Z.: Benchmarking single-image dehazing and beyond. IEEE TIP (2018) +28. Li, B., Liu, X., Hu, P., Wu, Z., Lv, J., Peng, X.: All-in-one image restoration for unknown corruption. In: CVPR (2022) +29. Li, L., Dong, Y., Ren, W., Pan, J., Gao, C., Sang, N., Yang, M.H.: Semi-supervised image dehazing. IEEE TIP (2019) +30. Li, M., Cao, X., Zhao, Q., Zhang, L., Meng, D.: Online rain/snow removal from surveillance videos. IEEE TIP (2021) +31. Li, P., Yun, M., Tian, J., Tang, Y., Wang, G., Wu, C.: Stacked dense networks for single-image snow removal. Neurocomputing (2019) +32. Li, R., Cheong, L.F., Tan, R.T.: Heavy rain image restoration: Integrating physics model and conditional adversarial learning. In: CVPR (2019) +33. Li, R., Tan, R.T., Cheong, L.F.: All in one bad weather removal using architectural search. In: CVPR (2020) +34. Li, X., Wu, J., Lin, Z., Liu, H., Zha, H.: Recurrent squeeze-and-excitation context aggregation net for single image deraining. In: ECCV (2018) +35. Li, Y., Tan, R.T., Guo, X., Lu, J., Brown, M.S.: Rain streak removal using layer priors. In: CVPR (2016) +36. Li, Y., Lu, J., Chen, H., Wu, X., Chen, X.: Dilated convolutional transformer for high-quality image deraining. In: CVPRW (June 2023) +37. Li, Z., Guan, B., Wei, Y., Zhou, Y., Zhang, J., Xu, J.: Mapping new realities: Ground truth image creation with pix2pix image-to-image translation. arXiv preprint arXiv:2404.19265 (2024) +38. Liang, J., Cao, J., Sun, G., Zhang, K., Van Gool, L., Timofte, R.: Swinir: Image restoration using swin transformer. In: ICCV (2021) + +39. Liang, Y., Anwar, S., Liu, Y.: Drt: A lightweight single image deraining recursive transformer. In: CVPR (2022) +40. Liu, J., Yang, W., Yang, S., Guo, Z.: Erase or fill? deep joint recurrent rain removal and reconstruction in videos. In: CVPR (2018) +41. Liu, K., Jiang, Y., Choi, I., Gu, J.: Learning image-adaptive codebooks for class-agnostic image restoration. arXiv preprint arXiv:2306.06513 (2023) +42. Liu, X., Suganuma, M., Sun, Z., Okatani, T.: Dual residual networks leveraging the potential of paired operations for image restoration. In: CVPR (2019) +43. Liu, Y., Liu, H., Li, L., Wu, Z., Chen, J.: A data-centric solution to nonhomogeneous dehazing via vision transformer. In: CVPR (2023) +44. Liu, Y.F., Jaw, D.W., Huang, S.C., Hwang, J.N.: Desnownet: Context-aware deep network for snow removal. IEEE TIP (2018) +45. Liu, Z., Lin, Y., Cao, Y., Hu, H., Wei, Y., Zhang, Z., Lin, S., Guo, B.: Swin transformer: Hierarchical vision transformer using shifted windows. In: ICCV (2021) +46. Liu, Z., Ning, J., Cao, Y., Wei, Y., Zhang, Z., Lin, S., Hu, H.: Video swim transformer. In: CVPR (2022) +47. Loshchilov, I., Hutter, F.: Sgdr: Stochastic gradient descent with warm restarts. arXiv preprint arXiv:1608.03983 (2016) +48. Loshchilov, I., Hutter, F.: Fixing weight decay regularization in adam. arXiv preprint arXiv:1711.05101 (2017) +49. Luo, Y., Zhao, R., Wei, X., Chen, J., Lu, Y., Xie, S., Wang, T., Xiong, R., Lu, M., Zhang, S.: Mowe: Mixture of weather experts for multiple adverse weather removal. arXiv preprint arXiv:2303.13739 (2023) +50. Lyu, W., Zheng, S., Ling, H., Chen, C.: Backdoor attacks against transformers with attention enhancement. In: ICLR Workshop (2023) +51. Ma, H., Zeng, D., Liu, Y.: Learning individualized treatment rules with many treatments: A supervised clustering approach using adaptive fusion. NeurIPS (2022) +52. Misra, D.: Mish: A self regularized non-monotonic activation function. arXiv preprint arXiv:1908.08681 (2019) +53. Ozdenizci, O., Legenstein, R.: Restoring vision in adverse weather conditions with patch-based denoising diffusion models. IEEE TPAMI (2023) +54. Park, D., Lee, B.H., Chun, S.Y.: All-in-one image restoration for unknown degradations using adaptive discriminative filters for specific degradations. In: CVPR (2023) +55. Paszke, A., Gross, S., Massa, F., Lerer, A., Bradbury, J., Chanan, G., Killeen, T., Lin, Z., Gimelshein, N., Antiga, L., et al.: Pytorch: An imperative style, high-performance deep learning library. NeurIPS (2019) +56. Patil, P.W., Gupta, S., Rana, S., Venkatesh, S., Murala, S.: Multi-weather image restoration via domain translation. In: ICCV (2023) +57. Qian, R., Tan, R.T., Yang, W., Su, J., Liu, J.: Attentive generative adversarial network for raindrop removal from a single image. In: CVPR (2018) +58. Quan, R., Yu, X., Liang, Y., Yang, Y.: Removing raindrops and rain streaks in one go. In: CVPR (2021) +59. Quan, Y., Deng, S., Chen, Y., Ji, H.: Deep learning for seeing through window with raindrops. In: ICCV (2019) +60. Ren, S., He, K., Girshick, R., Sun, J.: Faster r-cnn: Towards real-time object detection with region proposal networks. NeurIPS (2015) +61. Ren, W., Tian, J., Han, Z., Chan, A., Tang, Y.: Video desnowing and deraining based on matrix decomposition. In: CVPR (2017) + +62. Ren, W., Liu, S., Zhang, H., Pan, J., Cao, X., Yang, M.H.: Single image dehazing via multi-scale convolutional neural networks. In: ECCV (2016) +63. Ren, W., Ma, L., Zhang, J., Pan, J., Cao, X., Liu, W., Yang, M.H.: Gated fusion network for single image dehazing. In: CVPR (2018) +64. Shao, Y., Li, L., Ren, W., Gao, C., Sang, N.: Domain adaptation for image dehazing. In: CVPR (2020) +65. Song, Y., He, Z., Qian, H., Du, X.: Vision transformers for single image dehazing. IEEE TIP (2023) +66. Sun, S., Ren, W., Li, J., Zhang, K., Liang, M., Cao, X.: Event-aware video deraining via multi-patch progressive learning. IEEE TIP (2023) +67. Sun, S., Ren, W., Wang, T., Cao, X.: Rethinking image restoration for object detection. NeurIPS (2022) +68. Tan, Z., Wu, Y., Liu, Q., Chu, Q., Lu, L., Ye, J., Yu, N.: Exploring the application of large-scale pre-trained models on adverse weather removal. arXiv preprint arXiv:2306.09008 (2023) +69. Tu, Z., Talebi, H., Zhang, H., Yang, F., Milanfar, P., Bovik, A., Li, Y.: Maxim: Multi-axis mlp for image processing. In: CVPR (2022) +70. Valanarasu, J.M.J., Yasarla, R., Patel, V.M.: Transweather: Transformer-based restoration of images degraded by adverse weather conditions. In: CVPR (2022) +71. Wan, Y.C., Shao, M.W., Cheng, Y.S., Liu, Y.X., Bao, Z.Y., Meng, D.Y.: Restoring images captured in arbitrary hybrid adverse weather conditions in one go. arXiv preprint arXiv:2305.09996 (2023) +72. Wang, T., Zhang, K., Shao, Z., Luo, W., Stenger, B., Lu, T., Kim, T.K., Liu, W., Li, H.: Gridformer: Residual dense transformer with grid structure for image restoration in adverse weather conditions. arXiv preprint arXiv:2305.17863 (2023) +73. Wang, T., Yang, X., Xu, K., Chen, S., Zhang, Q., Lau, R.W.: Spatial attentive single-image deraining with a high quality real rain dataset. In: CVPR (2019) +74. Wang, Y., Ma, C., Liu, J.: Smartassign: Learning a smart knowledge assignment strategy for deraining and desnowing. In: CVPR (2023) +75. Wang, Z., Cun, X., Bao, J., Zhou, W., Liu, J., Li, H.: Uformer: A general u-shaped transformer for image restoration. In: CVPR (2022) +76. Wei, W., Yi, L., Xie, Q., Zhao, Q., Meng, D., Xu, Z.: Should we encode rain streaks in video as deterministic or stochastic? In: ICCV (2017) +77. Wu, H., Qu, Y., Lin, S., Zhou, J., Qiao, R., Zhang, Z., Xie, Y., Ma, L.: Contrastive learning for compact single image dehazing. In: CVPR (2021) +78. Xiao, J., Fu, X., Liu, A., Wu, F., Zha, Z.J.: Image de-raining transformer. IEEE TPAMI (2022) +79. Yang, W., Tan, R.T., Feng, J., Liu, J., Guo, Z., Yan, S.: Deep joint rain detection and removal from a single image. In: CVPR (2017) +80. Yasarla, R., Patel, V.M.: Uncertainty guided multi-scale residual learning-using a cycle spinning cnn for single image de-raining. In: CVPR (2019) +81. Yasarla, R., Sindagi, V.A., Patel, V.M.: Syn2real transfer learning for image deraining using gaussian processes. In: CVPR (2020) +82. Ye, T., Chen, S., Bai, J., Shi, J., Xue, C., Jiang, J., Yin, J., Chen, E., Liu, Y.: Adverse weather removal with codebook priors. In: ICCV (2023) +83. You, S., Tan, R.T., Kawakami, R., Mukaigawa, Y., Ikeuchi, K.: Adherent raindrop modeling, detection and removal in video. IEEE TPAMI (2015) +84. Yu, F., Koltun, V.: Multi-scale context aggregation by dilated convolutions. arXiv preprint arXiv:1511.07122 (2016) +85. Yu, F., Koltun, V., Funkhouser, T.: Dilated residual networks. In: CVPR (2017) + +86. Zamir, S.W., Arora, A., Khan, S., Hayat, M., Khan, F.S., Yang, M.H.: Restormer: Efficient transformer for high-resolution image restoration. In: CVPR (2022) +87. Zamir, S.W., Arora, A., Khan, S., Hayat, M., Khan, F.S., Yang, M.H., Shao, L.: Multi-stage progressive image restoration. In: CVPR (2021) +88. Zhang, H., Patel, V.M.: Density-aware single image de-raining using a multi-stream dense network. In: CVPR (2018) +89. Zhang, H., Sindagi, V., Patel, V.M.: Image de-raining using a conditional generative adversarial network. IEEE TCSVT (2019) +90. Zhang, H., Sindagi, V., Patel, V.M.: Joint transmission map estimation and de-hazing using deep networks. IEEE TCSVT (2019) +91. Zhang, H., Ba, Y., Yang, E., Mehra, V., Gella, B., Suzuki, A., Pfahnl, A., Chandrappa, C.C., Wong, A., Kadambi, A.: Weatherstream: Light transport automation of single image deweathering. In: CVPR (2023) +92. Zhang, J., Ren, W., Zhang, S., Zhang, H., Nie, Y., Xue, Z., Cao, X.: Hierarchical density-aware dehazing network. IEEE Transactions on Cybernetics (2021) +93. Zhang, K., Li, D., Luo, W., Ren, W.: Dual attention-in-attention model for joint rain streak and raindrop removal. In: IEEE TIP (2021) +94. Zhang, K., Li, R., Yu, Y., Luo, W., Li, C.: Deep dense multi-scale network for snow removal using semantic and depth priors. IEEE TIP (2021) +95. Zhang, X., Li, H., Qi, Y., Leow, W.K., Ng, T.K.: Rain removal in video by combining temporal and chromatic properties. In: ICME (2006) +96. Zhao, H., Gou, Y., Li, B., Peng, D., Lv, J., Peng, X.: Comprehensive and delicate: An efficient transformer for image restoration. In: CVPR (2023) +97. Zheng, Y., Zhan, J., He, S., Dong, J., Du, Y.: Curricular contrastive regularization for physics-aware single image dehazing. In: CVPR (2023) +98. Zhu, J.Y., Park, T., Isola, P., Efros, A.A.: Unpaired image-to-image translation using cycle-consistent adversarial networks. In: ICCV (2017) +99. Zhu, L., Fu, C.W., Lischinski, D., Heng, P.A.: Joint bi-layer optimization for single-image rain streak removal. In: ICCV (2017) +00. Zhu, Y., Wang, T., Fu, X., Yang, X., Guo, X., Dai, J., Qiao, Y., Hu, X.: Learning weather-general and weather-specific features for image restoration under multiple adverse weather conditions. In: CVPR (2023) \ No newline at end of file diff --git a/2024/Restoring Images in Adverse Weather Conditions via Histogram Transformer/images.zip b/2024/Restoring Images in Adverse Weather Conditions via Histogram Transformer/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..1e2658c1e5928a99e131529fe61d7ebec86bbb64 --- /dev/null +++ b/2024/Restoring Images in Adverse Weather Conditions via Histogram Transformer/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4e881bb793a2ef9a2fdd80289e7cf9049d09d16cb76c89fe2b91d58868506a2b +size 557998 diff --git a/2024/Restoring Images in Adverse Weather Conditions via Histogram Transformer/layout.json b/2024/Restoring Images in Adverse Weather Conditions via Histogram Transformer/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..abb3c1fa31836e4c29f3541d345a0656321fd6cf --- /dev/null +++ b/2024/Restoring Images in Adverse Weather Conditions via Histogram Transformer/layout.json @@ -0,0 +1,11511 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 134, + 111, + 479, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 111, + 479, + 148 + ], + "spans": [ + { + "bbox": [ + 134, + 111, + 479, + 148 + ], + "type": "text", + "content": "Restoring Images in Adverse Weather Conditions via Histogram Transformer" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 143, + 168, + 470, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 168, + 470, + 194 + ], + "spans": [ + { + "bbox": [ + 143, + 168, + 470, + 194 + ], + "type": "text", + "content": "Shangquan Sun" + }, + { + "bbox": [ + 143, + 168, + 470, + 194 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 143, + 168, + 470, + 194 + ], + "type": "text", + "content": ", Wenqi Ren" + }, + { + "bbox": [ + 143, + 168, + 470, + 194 + ], + "type": "inline_equation", + "content": "^{3,4\\dagger}" + }, + { + "bbox": [ + 143, + 168, + 470, + 194 + ], + "type": "text", + "content": ", Xinwei Gao" + }, + { + "bbox": [ + 143, + 168, + 470, + 194 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 143, + 168, + 470, + 194 + ], + "type": "text", + "content": ", Rui Wang" + }, + { + "bbox": [ + 143, + 168, + 470, + 194 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 143, + 168, + 470, + 194 + ], + "type": "text", + "content": ", and Xiaochun Cao" + }, + { + "bbox": [ + 143, + 168, + 470, + 194 + ], + "type": "inline_equation", + "content": "^{3}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 132, + 202, + 479, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 202, + 479, + 224 + ], + "spans": [ + { + "bbox": [ + 132, + 202, + 479, + 224 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 132, + 202, + 479, + 224 + ], + "type": "text", + "content": " Institute of Information Engineering, Chinese Academy of Sciences, Beijing 100085, China" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 145, + 224, + 468, + 313 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 145, + 224, + 468, + 246 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 224, + 468, + 246 + ], + "spans": [ + { + "bbox": [ + 145, + 224, + 468, + 246 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 145, + 224, + 468, + 246 + ], + "type": "text", + "content": " School of Cyber Security, University of Chinese Academy of Sciences, Beijing 100049, China" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 148, + 247, + 463, + 268 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 148, + 247, + 463, + 268 + ], + "spans": [ + { + "bbox": [ + 148, + 247, + 463, + 268 + ], + "type": "text", + "content": "3 School of Cyber Science and Technology, Shenzhen Campus of Sun Yat-sen University, Shenzhen 518107, China" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 148, + 269, + 463, + 290 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 148, + 269, + 463, + 290 + ], + "spans": [ + { + "bbox": [ + 148, + 269, + 463, + 290 + ], + "type": "text", + "content": "4 Guangdong Provincial Key Laboratory of Information Security Technology, Guangzhou 510006, China" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 155, + 291, + 458, + 313 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 155, + 291, + 458, + 313 + ], + "spans": [ + { + "bbox": [ + 155, + 291, + 458, + 313 + ], + "type": "text", + "content": "5 Wechat Business Group, Tencent, Shenzhen, Guangdong, China, 518057 shangquansun@gmail.com, renwq3@mail.sysu.edu.cn," + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 159, + 353, + 455, + 583 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 353, + 455, + 583 + ], + "spans": [ + { + "bbox": [ + 159, + 353, + 455, + 583 + ], + "type": "text", + "content": "Abstract. Transformer-based image restoration methods in adverse weather have achieved significant progress. Most of them use self-attention along the channel dimension or within spatially fixed-range blocks to reduce computational load. However, such a compromise results in limitations in capturing long-range spatial features. Inspired by the observation that the weather-induced degradation factors mainly cause similar occlusion and brightness, in this work, we propose an efficient Histogram Transformer (Histoformer) for restoring images affected by adverse weather. It is powered by a mechanism dubbed histogram self-attention, which sorts and segments spatial features into intensity-based bins. Self-attention is then applied across bins or within each bin to selectively focus on spatial features of dynamic range and process similar degraded pixels of the long range together. To boost histogram self-attention, we present a dynamic-range convolution enabling conventional convolution to conduct operation over similar pixels rather than neighbor pixels. We also observe that the common pixel-wise losses neglect linear association and correlation between output and ground-truth. Thus, we propose to leverage the Pearson correlation coefficient as a loss function to enforce the recovered pixels following the identical order as ground-truth. Extensive experiments demonstrate the efficacy and superiority of our proposed method. We have released the codes in Github." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 159, + 596, + 459, + 618 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 596, + 459, + 618 + ], + "spans": [ + { + "bbox": [ + 159, + 596, + 459, + 618 + ], + "type": "text", + "content": "Keywords: Image restoration " + }, + { + "bbox": [ + 159, + 596, + 459, + 618 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 596, + 459, + 618 + ], + "type": "text", + "content": " Adverse weather removal " + }, + { + "bbox": [ + 159, + 596, + 459, + 618 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 596, + 459, + 618 + ], + "type": "text", + "content": " Image Desnowing " + }, + { + "bbox": [ + 159, + 596, + 459, + 618 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 596, + 459, + 618 + ], + "type": "text", + "content": " Image deraining " + }, + { + "bbox": [ + 159, + 596, + 459, + 618 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 596, + 459, + 618 + ], + "type": "text", + "content": " Image dehazing " + }, + { + "bbox": [ + 159, + 596, + 459, + 618 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 596, + 459, + 618 + ], + "type": "text", + "content": " Raindrop removal" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 141, + 654, + 243, + 666 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 654, + 243, + 666 + ], + "spans": [ + { + "bbox": [ + 141, + 654, + 243, + 666 + ], + "type": "text", + "content": "† Corresponding author." + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 183, + 113, + 223, + 153 + ], + "blocks": [ + { + "bbox": [ + 183, + 113, + 223, + 153 + ], + "lines": [ + { + "bbox": [ + 183, + 113, + 223, + 153 + ], + "spans": [ + { + "bbox": [ + 183, + 113, + 223, + 153 + ], + "type": "image", + "image_path": "f5d18b2d309afc13747df1a5f189e25e71952145d3f26a50075d2f2b1e56e16b.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 215, + 155, + 278, + 164 + ], + "lines": [ + { + "bbox": [ + 215, + 155, + 278, + 164 + ], + "spans": [ + { + "bbox": [ + 215, + 155, + 278, + 164 + ], + "type": "text", + "content": "(a) Input patches" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 225, + 114, + 266, + 153 + ], + "blocks": [ + { + "bbox": [ + 225, + 114, + 266, + 153 + ], + "lines": [ + { + "bbox": [ + 225, + 114, + 266, + 153 + ], + "spans": [ + { + "bbox": [ + 225, + 114, + 266, + 153 + ], + "type": "image", + "image_path": "c54f0c05cb88496d0f5f964e58fc5acea9032acea793d3a0604509af6614e25e.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 268, + 114, + 309, + 153 + ], + "blocks": [ + { + "bbox": [ + 268, + 114, + 309, + 153 + ], + "lines": [ + { + "bbox": [ + 268, + 114, + 309, + 153 + ], + "spans": [ + { + "bbox": [ + 268, + 114, + 309, + 153 + ], + "type": "image", + "image_path": "6854636b0234474bb2f3174f8c96938ffde31dae682b51dd88fb2fc8f17382a0.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 189, + 166, + 242, + 224 + ], + "blocks": [ + { + "bbox": [ + 189, + 166, + 242, + 224 + ], + "lines": [ + { + "bbox": [ + 189, + 166, + 242, + 224 + ], + "spans": [ + { + "bbox": [ + 189, + 166, + 242, + 224 + ], + "type": "image", + "image_path": "41d4b01bb1cd440bace641cb76e72248c7aea15e00174e9eb3e8494087ac6a5d.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 199, + 226, + 293, + 236 + ], + "lines": [ + { + "bbox": [ + 199, + 226, + 293, + 236 + ], + "spans": [ + { + "bbox": [ + 199, + 226, + 293, + 236 + ], + "type": "text", + "content": "(b) Existing self-attention" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 251, + 166, + 304, + 224 + ], + "blocks": [ + { + "bbox": [ + 251, + 166, + 304, + 224 + ], + "lines": [ + { + "bbox": [ + 251, + 166, + 304, + 224 + ], + "spans": [ + { + "bbox": [ + 251, + 166, + 304, + 224 + ], + "type": "image", + "image_path": "ded64478d816fc6f5118ed522ba8a0ee8b43f7828a84b3052b03ee540c08451e.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 240, + 482, + 307 + ], + "lines": [ + { + "bbox": [ + 130, + 240, + 482, + 307 + ], + "spans": [ + { + "bbox": [ + 130, + 240, + 482, + 307 + ], + "type": "text", + "content": "Fig. 1: Given weather-degraded images in (a), traditional transformers perform self-attention either along the channel dimension or within a fixed-range block as shown in (b). In contrast, we observe that weather-induced degradation patterns tend to be similar but distinct from the background. So we categorize pixels affected by adverse weather and background pixels into distinct bins based on descending intensities (as depicted in (c)) and then conducts self-attention within and between bins." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 316, + 116, + 424, + 224 + ], + "blocks": [ + { + "bbox": [ + 316, + 116, + 424, + 224 + ], + "lines": [ + { + "bbox": [ + 316, + 116, + 424, + 224 + ], + "spans": [ + { + "bbox": [ + 316, + 116, + 424, + 224 + ], + "type": "image", + "image_path": "f57e0df975ba7aaeed41ebcf7f3d3ee334dc05603a2c913986d306c42d7285d2.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 320, + 226, + 420, + 236 + ], + "lines": [ + { + "bbox": [ + 320, + 226, + 420, + 236 + ], + "spans": [ + { + "bbox": [ + 320, + 226, + 420, + 236 + ], + "type": "text", + "content": "(c) Histogram self-attention" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 329, + 229, + 341 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 329, + 229, + 341 + ], + "spans": [ + { + "bbox": [ + 132, + 329, + 229, + 341 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 354, + 482, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 354, + 482, + 426 + ], + "spans": [ + { + "bbox": [ + 130, + 354, + 482, + 426 + ], + "type": "text", + "content": "The field of computer vision witnessed growing interest in restoring images affected by adverse weather conditions like rain, fog, and snow. These weather conditions significantly degrade visual quality, impacting the performance of downstream tasks such as object detection [3,60], and depth estimation [16,18]. The restoration of images under adverse weather is thereby a vital problem for the sake of vision aesthetics and safety." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 426, + 482, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 426, + 482, + 533 + ], + "spans": [ + { + "bbox": [ + 130, + 426, + 482, + 533 + ], + "type": "text", + "content": "Early works leverage weather-related priors to model statistical characteristic of degradation and remove adverse weathers [1, 2, 22, 23, 35, 76, 83, 95, 99]. Subsequently, convolutional neural networks (CNNs) have emerged as powerful tools for addressing deraining [5, 17, 34, 40, 57, 73, 79, 80, 88, 89], dehazing [27, 29, 62-64, 77, 88, 90, 92] and desnowing [30, 44, 61, 94]. However, the need of separately training networks for each task and the complexity of switching among multiple models present challenges for real-world applications. Li et al. [33] thus introduced the challenge of adverse weather removal, which entails the restoration of images affected by various weather conditions using a single unified model." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 534, + 482, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 534, + 482, + 629 + ], + "spans": [ + { + "bbox": [ + 130, + 534, + 482, + 629 + ], + "type": "text", + "content": "Recently, Transformer-based approaches have also been investigated for the adverse weather removal task, surpassing the efficacy of CNNs [11, 19, 70, 72]. Nonetheless, these Transformer-based methods usually make concessions regarding efficient memory utilization by confining self-attention operations to a fixed spatial range or solely within the channel dimension, as depicted in Figure 1b. This compromise impedes the inherent potential of Transformers, which was originally designed for superior global feature modeling, and consequently, it leads to a deterioration in the performance of restoration." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 630, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 630, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 630, + 482, + 665 + ], + "type": "text", + "content": "To address these problems, based on the observation that weather-induced degradation often exhibits common patterns shown in Figure 1a, we develop an efficient Histogram Transformer for unified adverse weather removal, named" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "text", + "content": "S. Sun et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 272 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 272 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 272 + ], + "type": "text", + "content": "Histoformer. Specifically, we introduce a Dynamic-range Histogram Self-Attention (DHSA) module, which endows self-attention with a dynamic-range spatial receptive field. We categorize pixel values proximate in intensity yet varied in spatial location into histogram bins. Self-attention is executed across the dimension of bin or frequency, whose process is illustrated in Figure 1c. To facilitate comprehensive feature extraction on both local and global scales, we devise two ways of reshaping for histogram self-attention: bin-wise histogram reshaping (BHR) and frequency-wise histogram reshaping (FHR). In BHR, the number of bins is configured to incorporate pixels spanning a more comprehensive intensity range, thereby facilitating global feature integration. In FHR, the number of frequencies is assigned such that each bin focuses on limited number of pixels, enhancing the utility of finer features. Consequently, the histogram self-attention attains the capability of modeling spatially dynamic ranges effectively." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 272, + 482, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 272, + 482, + 392 + ], + "spans": [ + { + "bbox": [ + 130, + 272, + 482, + 392 + ], + "type": "text", + "content": "To enable the convolution to extract dynamically-located weather-related dependencies, we develop a dynamic-range convolution layer, which involves sequential horizontal and vertical pixel sorting prior to the application of separable convolution. In order to capture multi-scale and multi-range information embedded within feature matrices, we introduce a Dual-scale Gated Feed-Forward (DGFF) module, enhancing its ability to model the visual characteristics effectively. Additionally, we note that conventional loss functions primarily focus on pixel-level closeness, overlooking the correlation at overall patch level. Consequently, we propose to leverage the Pearson correlation coefficient [12] to ensure the reconstruction of the linear relationship between restored and clean images." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 146, + 392, + 374, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 392, + 374, + 403 + ], + "spans": [ + { + "bbox": [ + 146, + 392, + 374, + 403 + ], + "type": "text", + "content": "Our contribution can be summarized in three folds:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 413, + 479, + 532 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 138, + 413, + 479, + 459 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 413, + 479, + 459 + ], + "spans": [ + { + "bbox": [ + 138, + 413, + 479, + 459 + ], + "type": "text", + "content": "- We propose a novel transformer targeted for unified adverse weather removal, equipped with a new histogram self-attention. It possesses dynamic-range spatial attention to weather-induced obstructions and thus can achieve degradation removal globally and efficiently." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 461, + 479, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 461, + 479, + 495 + ], + "spans": [ + { + "bbox": [ + 138, + 461, + 479, + 495 + ], + "type": "text", + "content": "- To capture multi-range information, we present a dual-scale feed-forward module. To enhance the comprehensive linear association between the recovered and ground-truth images, we develop a correlation loss." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 498, + 479, + 532 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 498, + 479, + 532 + ], + "spans": [ + { + "bbox": [ + 138, + 498, + 479, + 532 + ], + "type": "text", + "content": "- Our method attains state-of-the-art performance across various datasets. Additionally, we substantiate the efficacy of the proposed approach to restore real-world images and bolster the downstream application of detection." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 132, + 553, + 237, + 565 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 553, + 237, + 565 + ], + "spans": [ + { + "bbox": [ + 132, + 553, + 237, + 565 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 580, + 479, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 580, + 479, + 628 + ], + "spans": [ + { + "bbox": [ + 130, + 580, + 479, + 628 + ], + "type": "text", + "content": "Extensive research has been dedicated to addressing adverse weather removal challenges in computer vision, including tasks like deraining [5,17,34,40,66,73, 79,80,88,89], dehazing [27,29,62-64,67,77,88,90,92], desnowing [30,44,61,94], raindrop removal [57,59,83,93] and All-in-One weather removal [28,33,53,70]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 131, + 641, + 480, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 641, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 131, + 641, + 480, + 665 + ], + "type": "text", + "content": "Rain Streak Removal. The evolution of approaches is notable in rain streak removal techniques in computer vision. Kang et al. [23] pioneered a single image" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 134, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 134, + 91, + 447, + 102 + ], + "type": "text", + "content": "Restoring Images in Adverse Weather Conditions via Histogram Transformer" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 297 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 297 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 297 + ], + "type": "text", + "content": "deraining method using bilateral filters to decompose images into low and high-frequency components. However, recent advancements have seen a dominance of deep neural networks. An early deep CNN was introduced by Fu et al. [17] for extracting features from the high-frequency rain component, while Yang et al. [79] utilized recurrent networks to decompose rain layers and remove various streak types. Li et al. [32] proposed a method that addresses rain streaks and veiling effects in heavy rain scenes by integrating physics-based rain models and adversarial learning. A conditional generative adversarial network was also employed to solve rain streak removal [89]. Yasarla et al. [81] explored Gaussian processes for transfer learning from synthetic to real-world rain data. Quan et al. [58] used a cascaded network to remove both rain streaks and raindrops. Recently an image deraining Transformer [78] featuring a dual Transformer architecture was intricately formulated, incorporating both window-based and spatial-based mechanisms, thereby attaining exemplary outcomes. A sparse deraining Transformer is also proposed to enhance feature aggregation [11]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 307, + 482, + 416 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 307, + 482, + 416 + ], + "spans": [ + { + "bbox": [ + 130, + 307, + 482, + 416 + ], + "type": "text", + "content": "Raindrop Removal. Raindrop removal from single images has been addressed through various methods, with some relying on traditional hand-crafted features. An early work incorporated temporal information to address video-based raindrop removal [83]. Eigen et al. [15] employed a shallow CNN trained with image pairs containing raindrop-degraded and raindrop-free versions, though the results often exhibited blurriness. Qian et al. [57] introduced an attention GAN and a new dataset. Their method was later improved by Quan et al. [59] via generating attention maps based on mathematical raindrop descriptions and combining them with detected raindrop edges." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 426, + 482, + 522 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 426, + 482, + 522 + ], + "spans": [ + { + "bbox": [ + 130, + 426, + 482, + 522 + ], + "type": "text", + "content": "Snow Removal. Desnow-Net [44] was among the pioneering CNN-based approaches for snow removal, followed by Li et al.'s stacked dense network [31] and Chen et al.'s JSTASR [8], which introduced a size and transparency aware method. More recently, DDMSNet [94] introduced a dense multi-scale network that leverages semantic and geometric priors to enhance snow removal. A hierarchical decomposition paradigm involving the dual-tree wavelet transform for snow removal is also proposed [9]. Chen et al. [7] designed SnowFormer, a framework that used cross-attention to establish local-global context interaction." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 533, + 482, + 630 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 482, + 630 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 482, + 630 + ], + "type": "text", + "content": "Fog Removal. Li et al. [26] presented a CNN that takes into account both atmospheric luminosity and transmission maps to conduct dehazing. Ren et al. [63] advocated a pre-processing approach for hazy image manipulation, thereby engendering multiple input modalities and, in the process, inducing chromatic aberrations as part of their dehazing procedure. A hierarchical density-aware network is also introduced, specializing in the domain of image dehazing [92]. Zheng et al. [97] formulated a curriculum-based contrastive regularization dehazing method aimed at fostering agreement within a contrastive space." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 641, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 641, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 131, + 641, + 481, + 666 + ], + "type": "text", + "content": "All-in-One Weather Removal. Some recent works attempted to address various weather-induced degradations by a singular network. Li et al. [33] proposed an" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "text", + "content": "S. Sun et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 248 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 248 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 248 + ], + "type": "text", + "content": "All-in-One network, containing a generator comprising multiple task-specific encoders and a shared decoder. Valanarasu et al. [70] presented TransWeather, a transformer-based model featuring a solitary encoder-decoder structure, capable of rejuvenating images afflicted by various atmospheric conditions. A pipeline for the automatic selection of weather-degraded data was also proposed to enhance existing models [91]. Zhu et al. developed WGWS-Net [100] capable of learning weather-general and weather-specific in two separate stages. Some other recent works also trial addressing adverse weather removal by adopting probabilistic denoising diffusion model [53], knowledge distillation [10], large-scale Pre-trained model [68], mixture of experts [49], few-shot learning [24], codebooks [41,71,82], adaptive filters [54], knowledge assignment [74] and domain translation [56]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 261, + 482, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 261, + 482, + 392 + ], + "spans": [ + { + "bbox": [ + 130, + 261, + 482, + 392 + ], + "type": "text", + "content": "Transformer-based Image Restoration. Since the inception of the Vision Transformer (ViT) [14] for visual recognition, transformers have gained substantial traction across a spectrum of computer vision tasks [25, 37, 45, 46, 50, 51, 86]. Particularly within the realm of low-level vision, the Image Processing Transformer [4] exemplifies how pre-training a transformer on extensive datasets can significantly enhance performance for low-level applications. U-former [75], on the other hand, introduced a transformer architecture based on the U-Net design for restoration tasks. Swin-IR [38] employed the Swin Transformer [45] for image restoration. Some latest Transformer-based methods were proposed for deraining [11, 39], desnowing [7], dehazing [19, 43, 65] and All-in-One weather removal [70, 72]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 393, + 482, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 393, + 482, + 441 + ], + "spans": [ + { + "bbox": [ + 130, + 393, + 482, + 441 + ], + "type": "text", + "content": "Unlike the existing Transformer-based approaches whose self-attention is applied within either fixed spatial ranges or merely channel dimension, our method enables dynamic-range spatial attention to adaptively focus on weather-induced degradation with similar patterns." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 461, + 202, + 474 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 461, + 202, + 474 + ], + "spans": [ + { + "bbox": [ + 131, + 461, + 202, + 474 + ], + "type": "text", + "content": "3 Method" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 487, + 265, + 499 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 487, + 265, + 499 + ], + "spans": [ + { + "bbox": [ + 131, + 487, + 265, + 499 + ], + "type": "text", + "content": "3.1 Overall Architecture" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 509, + 482, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 509, + 482, + 629 + ], + "spans": [ + { + "bbox": [ + 130, + 509, + 482, + 629 + ], + "type": "text", + "content": "The overall architectural framework of our Histoformer is illustrated in Figure 2. Suppose the input is a low-quality image " + }, + { + "bbox": [ + 130, + 509, + 482, + 629 + ], + "type": "inline_equation", + "content": "I^{lq} \\in \\mathbb{R}^{3 \\times H \\times W}" + }, + { + "bbox": [ + 130, + 509, + 482, + 629 + ], + "type": "text", + "content": ", we pass the input through a " + }, + { + "bbox": [ + 130, + 509, + 482, + 629 + ], + "type": "inline_equation", + "content": "3 \\times 3" + }, + { + "bbox": [ + 130, + 509, + 482, + 629 + ], + "type": "text", + "content": " convolution to conduct the overlapping image patch-embedding. Within both the encoder and decoder of the network backbone, we arrange Histogram Transformer Blocks (HTBs) to extract intricate features and capture dynamically distributed degradation factors. Within the same stage, encoders and decoders are interlinked through skip-connections, thereby establishing connections between consecutive intermediate features to enhance the stability of the training process. Between each stage, we apply pixel-unshuffle and pixelshuffle operations for the purpose of feature down-sampling and up-sampling." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 630, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 630, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 131, + 630, + 482, + 665 + ], + "type": "text", + "content": "Within each HTB, we introduce Dynamic-range Histogram Self-Attention (DHSA) to extract spatially dynamic weather degradation and enhance both local and global feature aggregation. Moreover, a Dual-scale Gated Feed-Forward" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 134, + 91, + 448, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 91, + 448, + 103 + ], + "spans": [ + { + "bbox": [ + 134, + 91, + 448, + 103 + ], + "type": "text", + "content": "Restoring Images in Adverse Weather Conditions via Histogram Transformer" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 132, + 112, + 479, + 277 + ], + "blocks": [ + { + "bbox": [ + 132, + 112, + 479, + 277 + ], + "lines": [ + { + "bbox": [ + 132, + 112, + 479, + 277 + ], + "spans": [ + { + "bbox": [ + 132, + 112, + 479, + 277 + ], + "type": "image", + "image_path": "ea63ebb64c2fdd34c12ae0e50e98c82e17d225bfd71eb58ff59d6cd2d8b773fa.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 284, + 482, + 340 + ], + "lines": [ + { + "bbox": [ + 130, + 284, + 482, + 340 + ], + "spans": [ + { + "bbox": [ + 130, + 284, + 482, + 340 + ], + "type": "text", + "content": "Fig. 2: The overall architecture of our Histoformer for weather removal. The main component is the Histogram Transformer block, and it comprises the Dynamic-range Histogram Self-Attention (DHSA) module and the Dual-scale Gated Feed-Forward (DGFF) module. Within DHSA, we present two types of reshaping mechanism, i.e., Bin-wise Histogram Reshaping and Frequency-wise Histogram Reshaping." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 361, + 482, + 494 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 361, + 482, + 494 + ], + "spans": [ + { + "bbox": [ + 130, + 361, + 482, + 494 + ], + "type": "text", + "content": "(DGFF) module is integrated into the HTB to enrich the representation of multi-range features, contributing to the process of image restoration. During each stage of encoding phases, our model is equipped with a crude skip-connection for supplementing original features from input, comprised of a sequence of operations, including average pooling, pixel-wise convolution, and depth-wise convolution. We start the crude skip-connection after the first stage, and this setup enables the encoders to focus more effectively on learning the weather-induced residuals. Through this hybrid formulation, Histoformer is empowered to exploit both the adaptive contents of weather-irrelevant background and the inherent characteristics of weather-degraded patterns, facilitating the separation of undesired degradation from the latent clear background." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 509, + 311, + 521 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 509, + 311, + 521 + ], + "spans": [ + { + "bbox": [ + 132, + 509, + 311, + 521 + ], + "type": "text", + "content": "3.2 Histogram Transformer Block" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 526, + 481, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 526, + 481, + 563 + ], + "spans": [ + { + "bbox": [ + 130, + 526, + 481, + 563 + ], + "type": "text", + "content": "As the key component of our Histoformer, HTB contains two pivotal modules, i.e., DHSA and DGFF. These two components are arranged to interact with layer normalization and can be formulated as follows:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 235, + 568, + 480, + 581 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 235, + 568, + 480, + 581 + ], + "spans": [ + { + "bbox": [ + 235, + 568, + 480, + 581 + ], + "type": "interline_equation", + "content": "F _ {l} = F _ {l - 1} + \\mathrm {D H S A} (\\mathrm {L N} (F _ {l - 1})), \\tag {1}", + "image_path": "6162af5b153c50bddb0f6d076cdd4e599ee42acf6cb91f524837d3d7ff9afc9e.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 235, + 583, + 480, + 596 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 235, + 583, + 480, + 596 + ], + "spans": [ + { + "bbox": [ + 235, + 583, + 480, + 596 + ], + "type": "interline_equation", + "content": "F _ {l} = F _ {l} + \\operatorname {D G F F} \\left(\\ln \\left(F _ {l}\\right)\\right), \\tag {2}", + "image_path": "a5cea72b45cda5a2e385b5879b833551b3cea87e8782e7e5d5ca2306dcf9259b.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 601, + 479, + 626 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 601, + 479, + 626 + ], + "spans": [ + { + "bbox": [ + 130, + 601, + 479, + 626 + ], + "type": "text", + "content": "where LN denotes layer normalization and " + }, + { + "bbox": [ + 130, + 601, + 479, + 626 + ], + "type": "inline_equation", + "content": "F_{l}" + }, + { + "bbox": [ + 130, + 601, + 479, + 626 + ], + "type": "text", + "content": " represents the feature at " + }, + { + "bbox": [ + 130, + 601, + 479, + 626 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 130, + 601, + 479, + 626 + ], + "type": "text", + "content": "-th stage. The details of DHSA and DGFF are presented in Section 3.2 and 3.2 respectively." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 641, + 484, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 484, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 484, + 666 + ], + "type": "text", + "content": "Dynamic-range Histogram Self-Attention To better capture dynamically distributed weather-induced degradation, we develop a Dynamic-range Histogram" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "text", + "content": "S. Sun et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "text", + "content": "Self-Attention (DHSA) module. This module consists of a process involving dynamic-range convolution, which reorders the spatial distribution of fractional features, and a dual-path histogram self-attention mechanism that combines global and local dynamic feature aggregation. Prior to the final output projection of a " + }, + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "inline_equation", + "content": "1 \\times 1" + }, + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "text", + "content": " point-wise convolution, the reordered features are sorted back into their original locations to maintain spatial consistency." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 197, + 482, + 353 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 197, + 482, + 353 + ], + "spans": [ + { + "bbox": [ + 130, + 197, + 482, + 353 + ], + "type": "text", + "content": "Dynamic-range Convolution. Traditional convolution operations employ fixed kernel sizes, resulting in a limited receptive field range and consequently perform local and small-range computations. This restricted scope of convolution, which primarily focuses on local information, does not naturally align with the self-attention mechanism's capacity to model long-range dependencies. To address this limitation, we devise a dynamic-range convolution technique by meticulously reordering the input features prior to the traditional convolution operation. Given an input feature " + }, + { + "bbox": [ + 130, + 197, + 482, + 353 + ], + "type": "inline_equation", + "content": "F \\in \\mathbb{R}^{C \\times H \\times W}" + }, + { + "bbox": [ + 130, + 197, + 482, + 353 + ], + "type": "text", + "content": ", we divide it into two branches along the channel dimension, namely " + }, + { + "bbox": [ + 130, + 197, + 482, + 353 + ], + "type": "inline_equation", + "content": "F_{1}" + }, + { + "bbox": [ + 130, + 197, + 482, + 353 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 197, + 482, + 353 + ], + "type": "inline_equation", + "content": "F_{2}" + }, + { + "bbox": [ + 130, + 197, + 482, + 353 + ], + "type": "text", + "content": ". For the first branch of features, we perform sorting operations both horizontally and vertically, subsequently concatenating the sorted features with the second branch of feature. The resulting recombined features are then passed through the subsequent separable convolution. The entire process is articulated as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 216, + 358, + 399, + 371 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 216, + 358, + 399, + 371 + ], + "spans": [ + { + "bbox": [ + 216, + 358, + 399, + 371 + ], + "type": "interline_equation", + "content": "F _ {1}, F _ {2} = \\operatorname {S p l i t} (F), F _ {1} = \\operatorname {S o r t} _ {\\mathrm {v}} (\\operatorname {S o r t} _ {\\mathrm {h}} (F _ {1})),", + "image_path": "5b6932faa412d2a225b957c895a135d1a2e21c6e5b661c357d06d14ea244ff65.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 216, + 365, + 480, + 387 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 216, + 365, + 480, + 387 + ], + "spans": [ + { + "bbox": [ + 216, + 365, + 480, + 387 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} F = \\operatorname {C o n v} _ {3 \\times 3} ^ {\\mathrm {d}} \\left(\\operatorname {C o n v} _ {1 \\times 1} \\left(\\operatorname {C o n c a t} \\left(F _ {1}, F _ {2}\\right)\\right)\\right), \\end{array} \\tag {3}", + "image_path": "c4cf90cf34ad6307d36cdf88f0603cb394bbd899b3d0ae3b9100e7c63d3b2f19.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 393, + 482, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 393, + 482, + 525 + ], + "spans": [ + { + "bbox": [ + 130, + 393, + 482, + 525 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 393, + 482, + 525 + ], + "type": "inline_equation", + "content": "\\mathrm{Conv}_{1\\times 1}" + }, + { + "bbox": [ + 130, + 393, + 482, + 525 + ], + "type": "text", + "content": " is " + }, + { + "bbox": [ + 130, + 393, + 482, + 525 + ], + "type": "inline_equation", + "content": "1\\times 1" + }, + { + "bbox": [ + 130, + 393, + 482, + 525 + ], + "type": "text", + "content": " point-wise convolution, " + }, + { + "bbox": [ + 130, + 393, + 482, + 525 + ], + "type": "inline_equation", + "content": "\\mathrm{Conv}_{3\\times 3}^{\\mathrm{d}}" + }, + { + "bbox": [ + 130, + 393, + 482, + 525 + ], + "type": "text", + "content": " represents " + }, + { + "bbox": [ + 130, + 393, + 482, + 525 + ], + "type": "inline_equation", + "content": "3\\times 3" + }, + { + "bbox": [ + 130, + 393, + 482, + 525 + ], + "type": "text", + "content": " depthwise convolution, Concat is the concatenation operation along channel, Split denotes the operation of splitting features along channel dimension, and " + }, + { + "bbox": [ + 130, + 393, + 482, + 525 + ], + "type": "inline_equation", + "content": "\\mathrm{Sort}_{i\\in \\{\\mathrm{h},\\mathrm{v}\\}}" + }, + { + "bbox": [ + 130, + 393, + 482, + 525 + ], + "type": "text", + "content": " represents the horizontal or vertical sorting operation. This approach organizes pixels of high and low intensities into regular patterns at the diagonal corners of the matrices, thereby allowing convolution to perform computations across dynamic ranges. Given that weather-induced degradation typically exhibits closely related patterns, degraded pixels tend to concentrate in neighboring locations, separated with those clean pixels. As a result, this arrangement enables convolution kernels to partially focus on preserving clean information and separately recovering degraded features." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": "Histogram Self-Attention. Existing vision Transformers [11,75,78,78,86,96] typically leverage fixed range of attention or merely the attention along channel dimension due to the compromise of computation and memory efficiency. However, the fixed setting restricts the self-attention to span adaptively long range to associate desired features. We notice that weather-induced degradation causes similar patterns and that those pixels containing either background feature or weather degradation of different intensities had better be assigned with various extents of attention. We thus propose a histogram self-attention mechanism to categorize spatial elements into bins and allocate varying attention within and across bins. For the sake of parallel computing, we set each bin contains identical number of pixels during implement." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 134, + 91, + 448, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 91, + 448, + 103 + ], + "spans": [ + { + "bbox": [ + 134, + 91, + 448, + 103 + ], + "type": "text", + "content": "Restoring Images in Adverse Weather Conditions via Histogram Transformer" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 489, + 164 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 489, + 164 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 489, + 164 + ], + "type": "text", + "content": "Given the output of dynamic-range convolution, we can separate them into Value feature " + }, + { + "bbox": [ + 130, + 116, + 489, + 164 + ], + "type": "inline_equation", + "content": "V \\in \\mathbb{R}^{C \\times H \\times W}" + }, + { + "bbox": [ + 130, + 116, + 489, + 164 + ], + "type": "text", + "content": " and two pairs of Query-key " + }, + { + "bbox": [ + 130, + 116, + 489, + 164 + ], + "type": "inline_equation", + "content": "F_{QK,1}, F_{QK,2} \\in \\mathbb{R}^{2C \\times H \\times W}" + }, + { + "bbox": [ + 130, + 116, + 489, + 164 + ], + "type": "text", + "content": " subsequently passing to two branches. We firstly sort the sequence of " + }, + { + "bbox": [ + 130, + 116, + 489, + 164 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 130, + 116, + 489, + 164 + ], + "type": "text", + "content": " and based on its index arrange the Query-Key pairs accordingly, expressed as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 205, + 168, + 324, + 182 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 205, + 168, + 324, + 182 + ], + "spans": [ + { + "bbox": [ + 205, + 168, + 324, + 182 + ], + "type": "interline_equation", + "content": "V, d = \\operatorname {S o r t} \\left(\\mathbf {R} _ {\\mathrm {C} \\times \\mathrm {H} \\times \\mathrm {W}} ^ {\\mathrm {C} \\times \\mathrm {H W}} (V)\\right),", + "image_path": "127e041d7c896af4efb018986ecfae8f97406e491b8703bbec4fb2d4359aaefd.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 205, + 185, + 480, + 200 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 205, + 185, + 480, + 200 + ], + "spans": [ + { + "bbox": [ + 205, + 185, + 480, + 200 + ], + "type": "interline_equation", + "content": "Q _ {1}, K _ {1} = \\operatorname {S p l i t} \\left(\\operatorname {G a t h e r} \\left(\\mathbf {R} _ {\\mathrm {C} \\times \\mathrm {H} \\times \\mathrm {W}} ^ {\\mathrm {C} \\times \\mathrm {H W}} \\left(F _ {Q K, 1}\\right), d\\right)\\right), \\tag {4}", + "image_path": "49c49afadcbeb4e3b57010976728e77b58e5282500645318cdfb5993ecb64a92.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 205, + 202, + 410, + 217 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 205, + 202, + 410, + 217 + ], + "spans": [ + { + "bbox": [ + 205, + 202, + 410, + 217 + ], + "type": "interline_equation", + "content": "Q _ {2}, K _ {2} = \\operatorname {S p l i t} \\left(\\operatorname {G a t h e r} \\left(\\mathbf {R} _ {\\mathrm {C} \\times \\mathrm {H} \\times \\mathrm {W}} ^ {\\mathrm {C} \\times \\mathrm {H W}} (F _ {Q K, 2}), d\\right)\\right),", + "image_path": "ef1a9169e2fe1122e04f930c1a713d8f077d2dea70b74a0660e22649e5734072.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 221, + 481, + 261 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 221, + 481, + 261 + ], + "spans": [ + { + "bbox": [ + 130, + 221, + 481, + 261 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 221, + 481, + 261 + ], + "type": "inline_equation", + "content": "\\mathbf{R}_{\\mathrm{C}\\times ,\\mathrm{H}\\times \\mathrm{W}}^{\\mathrm{C}\\times \\mathrm{HW}}" + }, + { + "bbox": [ + 130, + 221, + 481, + 261 + ], + "type": "text", + "content": " represents the operation of reshaping features from " + }, + { + "bbox": [ + 130, + 221, + 481, + 261 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^{C\\times H\\times W}" + }, + { + "bbox": [ + 130, + 221, + 481, + 261 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 130, + 221, + 481, + 261 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^{C\\times HW}" + }, + { + "bbox": [ + 130, + 221, + 481, + 261 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 221, + 481, + 261 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 130, + 221, + 481, + 261 + ], + "type": "text", + "content": " is the index of sorted Value, and Gather denotes the operation of retrieving elements of tensor based on a given index." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 264, + 482, + 419 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 264, + 482, + 419 + ], + "spans": [ + { + "bbox": [ + 130, + 264, + 482, + 419 + ], + "type": "text", + "content": "Then given the number of bins " + }, + { + "bbox": [ + 130, + 264, + 482, + 419 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 130, + 264, + 482, + 419 + ], + "type": "text", + "content": ", we reshape the sorted features from " + }, + { + "bbox": [ + 130, + 264, + 482, + 419 + ], + "type": "inline_equation", + "content": "C \\times HW" + }, + { + "bbox": [ + 130, + 264, + 482, + 419 + ], + "type": "text", + "content": " into " + }, + { + "bbox": [ + 130, + 264, + 482, + 419 + ], + "type": "inline_equation", + "content": "C \\times B \\times HW / B" + }, + { + "bbox": [ + 130, + 264, + 482, + 419 + ], + "type": "text", + "content": ". To extract both global and local information, we define two types of reshaping, i.e., bin-wise histogram reshaping (BHR) and frequency-wise histogram reshaping (FHR). The first is to assign the number of bins equal to " + }, + { + "bbox": [ + 130, + 264, + 482, + 419 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 130, + 264, + 482, + 419 + ], + "type": "text", + "content": " and each bin contains " + }, + { + "bbox": [ + 130, + 264, + 482, + 419 + ], + "type": "inline_equation", + "content": "HW / B" + }, + { + "bbox": [ + 130, + 264, + 482, + 419 + ], + "type": "text", + "content": " elements, while the second is to set the frequency of each bin equal to " + }, + { + "bbox": [ + 130, + 264, + 482, + 419 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 130, + 264, + 482, + 419 + ], + "type": "text", + "content": " and the number of bins is " + }, + { + "bbox": [ + 130, + 264, + 482, + 419 + ], + "type": "inline_equation", + "content": "HW / B" + }, + { + "bbox": [ + 130, + 264, + 482, + 419 + ], + "type": "text", + "content": ". By this way, we can extract large-scale information by BHR where each bin contains large number of dynamically located pixels and fine-grained information by FHR where each bins contains modicum pixels neighboring in terms of intensity. The two pairs of Query-Key features are passed through two types of reshaping and subsequent self-attention process respectively, and their outputs are elementwisely multiplied to yield the final output. The process can be formulated as the following expressions:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 211, + 424, + 403, + 451 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 211, + 424, + 403, + 451 + ], + "spans": [ + { + "bbox": [ + 211, + 424, + 403, + 451 + ], + "type": "interline_equation", + "content": "\\mathrm {A _ {B}} = \\mathrm {s o f t m a x} \\left(\\frac {\\mathbf {R _ {B}} (Q _ {1}) \\mathbf {R _ {B}} (K _ {1}) ^ {\\top}}{\\sqrt {k}}\\right) \\mathbf {R _ {B}} (V),", + "image_path": "fd227eb78bf3175eee9df4ce95690f198e037d003411c8134a6c9454f7dc586e.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 211, + 453, + 480, + 479 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 211, + 453, + 480, + 479 + ], + "spans": [ + { + "bbox": [ + 211, + 453, + 480, + 479 + ], + "type": "interline_equation", + "content": "\\mathrm {A} _ {\\mathrm {F}} = \\operatorname {s o f t m a x} \\left(\\frac {\\mathbf {R} _ {\\mathrm {F}} \\left(Q _ {2}\\right) \\mathbf {R} _ {\\mathrm {F}} \\left(K _ {2}\\right) ^ {\\top}}{\\sqrt {k}}\\right) \\mathbf {R} _ {\\mathrm {F}} (V), \\tag {5}", + "image_path": "b5ed36b6f16bafebed1143f2e42e787baaff388771d213c28c981816cafed7bd.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 211, + 483, + 274, + 495 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 211, + 483, + 274, + 495 + ], + "spans": [ + { + "bbox": [ + 211, + 483, + 274, + 495 + ], + "type": "interline_equation", + "content": "\\mathrm {A} = \\mathrm {A} _ {\\mathrm {B}} \\odot \\mathrm {A} _ {\\mathrm {F}},", + "image_path": "94ad8f1296799a3e3b0819375bb1cba0dba953effcc1777e7e6710c23b865d42.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 500, + 482, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 500, + 482, + 525 + ], + "spans": [ + { + "bbox": [ + 130, + 500, + 482, + 525 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 500, + 482, + 525 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 500, + 482, + 525 + ], + "type": "text", + "content": " is the number of heads, " + }, + { + "bbox": [ + 130, + 500, + 482, + 525 + ], + "type": "inline_equation", + "content": "\\mathbf{R}_{i\\in \\{\\mathrm{B},\\mathrm{F}\\}}" + }, + { + "bbox": [ + 130, + 500, + 482, + 525 + ], + "type": "text", + "content": " denotes the reshaping operation of either BHR or FHR, and " + }, + { + "bbox": [ + 130, + 500, + 482, + 525 + ], + "type": "inline_equation", + "content": "\\mathrm{A}_{i\\in \\{\\mathrm{B},\\mathrm{F}\\}}" + }, + { + "bbox": [ + 130, + 500, + 482, + 525 + ], + "type": "text", + "content": " represents the obtained attention map." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "text", + "content": "Dual-scale Gated Feed-Forward Previous studies [11, 75, 78, 86] typically leverage single-range or single-scale convolution in the standard feed-forward network to bolster local context. Nonetheless, these methods often disregard the correlations among dynamically distributed weather-induced degradation. In practice, multi-scale information can be extracted by not only enlarging the kernel size but also leveraging the dilation mechanism [36, 84, 85]. As a result, we conceive a Dual-scale Gated Feed-Forward (DGFF) module, which integrates two distinct multi-range and multi-scale depth-wise convolution pathways within the transmission process." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "type": "text", + "content": "S. Sun et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 134, + 188, + 480, + 296 + ], + "blocks": [ + { + "bbox": [ + 130, + 119, + 482, + 175 + ], + "lines": [ + { + "bbox": [ + 130, + 119, + 482, + 175 + ], + "spans": [ + { + "bbox": [ + 130, + 119, + 482, + 175 + ], + "type": "text", + "content": "Table 1: Quantitative comparisons on three weather removal tasks in terms of PSNR and SSIM, where higher values indicate better performance. The top halves of tables display the results of task-specific methods, while the bottom halves present evaluations of the unified multi-weather models. The best and the second best results are in bold and underlined. Those with * indicate the methods whose source codes are unavailable." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 171, + 178, + 242, + 187 + ], + "lines": [ + { + "bbox": [ + 171, + 178, + 242, + 187 + ], + "spans": [ + { + "bbox": [ + 171, + 178, + 242, + 187 + ], + "type": "text", + "content": "(a) Image Desnowing" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 290, + 178, + 375, + 187 + ], + "lines": [ + { + "bbox": [ + 290, + 178, + 375, + 187 + ], + "spans": [ + { + "bbox": [ + 290, + 178, + 375, + 187 + ], + "type": "text", + "content": "(b) Deraining & Dehazing" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 400, + 178, + 472, + 187 + ], + "lines": [ + { + "bbox": [ + 400, + 178, + 472, + 187 + ], + "spans": [ + { + "bbox": [ + 400, + 178, + 472, + 187 + ], + "type": "text", + "content": "(c) Raindrop Removal" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 134, + 188, + 480, + 296 + ], + "lines": [ + { + "bbox": [ + 134, + 188, + 480, + 296 + ], + "spans": [ + { + "bbox": [ + 134, + 188, + 480, + 296 + ], + "type": "table", + "html": "
Snow100K-S [44]Snow100K-L [44]Outdoor-Rain [32]RainDrop [57]
PSNRSSIMPSNRSSIMPSNRSSIMPSNR SSIM
SPANet [73]29.920.826023.700.7930CycleGAN [98]17.620.6560pix2pix [20]28.02
JSTASR [8]31.400.901225.320.8076pix2pix [20]19.090.7100DuRN [42]31.24
RESCAN [34]31.510.903226.080.8108HRGAN [32]21.560.8550RaindropAttn [59]31.44
DesnowNet [44]32.330.950027.170.8983PCNet [21]26.190.9015AttentiveGAN [57]31.59
DDMSNet [94]34.340.944528.850.8772MPRNet [87]28.030.9192IDT [78]31.87
NAFNet [6]34.790.949730.060.9017NAFNet [6]29.590.9027MAXIM [69]31.87
Restormer [86]36.010.957930.360.9068Restormer [86]30.030.9215Restormer [86]32.18
All-in-One [33]*--28.330.8820All-in-One [33]*24.710.8980All-in-One [33]*31.12
TransWeather [70]32.510.934129.310.8879TransWeather [70]28.830.9000TransWeather [70]30.17
Chen et al. [10]34.420.946930.220.9071Chen et al. [10]29.270.9147Chen et al. [10]31.81
WGWSNet [100]34.310.946030.160.9007WGWSNet [100]29.320.9207WGWSNet [100]32.38
WeatherDiff64 [53]35.830.956630.090.9041WeatherDiff64 [53]29.640.9312WeatherDiff64 [53]30.71
WeatherDiff128 [53]35.020.951629.580.8941WeatherDiff128 [53]29.720.9216WeatherDiff128 [53]29.66
AWRCP [82]*36.920.965231.920.9341AWRCP [82]*31.390.9329AWRCP [82]*31.93
Histoformer (Ours)37.410.965632.160.9261Histoformer (Ours)32.080.9389Histoformer (Ours)33.06
", + "image_path": "b0d9b129f2294599477371e5e8f549f72c8e1225ab44d8d0673e9ddbaa3e62c1.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 318, + 482, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 318, + 482, + 426 + ], + "spans": [ + { + "bbox": [ + 130, + 318, + 482, + 426 + ], + "type": "text", + "content": "Given an input tensor " + }, + { + "bbox": [ + 130, + 318, + 482, + 426 + ], + "type": "inline_equation", + "content": "F_{l} \\in \\mathbb{R}^{C \\times H \\times W}" + }, + { + "bbox": [ + 130, + 318, + 482, + 426 + ], + "type": "text", + "content": ", we initially employ a point-wise convolution operation to augment the channel dimension by a factor of " + }, + { + "bbox": [ + 130, + 318, + 482, + 426 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 130, + 318, + 482, + 426 + ], + "type": "text", + "content": ". Following this augmentation, the expanded tensor is directed into two parallel branches. Throughout the feature transformation process, " + }, + { + "bbox": [ + 130, + 318, + 482, + 426 + ], + "type": "inline_equation", + "content": "5 \\times 5" + }, + { + "bbox": [ + 130, + 318, + 482, + 426 + ], + "type": "text", + "content": " and dilated " + }, + { + "bbox": [ + 130, + 318, + 482, + 426 + ], + "type": "inline_equation", + "content": "3 \\times 3" + }, + { + "bbox": [ + 130, + 318, + 482, + 426 + ], + "type": "text", + "content": " depth-wise convolutions are employed to enhance the extraction of multi-range and multi-scale information. Following the gating mechanism [13], the output of the second branch after passing through an activation act as a gating map for the other branch. Thus, the complete feature fusion process within the DGFF module is formulated as follows:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 203, + 430, + 380, + 442 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 203, + 430, + 380, + 442 + ], + "spans": [ + { + "bbox": [ + 203, + 430, + 380, + 442 + ], + "type": "interline_equation", + "content": "F _ {l, 1}, F _ {l, 2} = \\operatorname {S p l i t} \\left(\\operatorname {S h u f f l e} \\left(\\operatorname {C o n v} _ {1 \\times 1} \\left(F _ {l}\\right)\\right)\\right),", + "image_path": "080137e9027567070bc29e32a167763683826401d98804f1a0f5beed85a76b58.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 203, + 445, + 480, + 460 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 203, + 445, + 480, + 460 + ], + "spans": [ + { + "bbox": [ + 203, + 445, + 480, + 460 + ], + "type": "interline_equation", + "content": "F _ {l, 1} = \\operatorname {C o n v} _ {5 \\times 5} ^ {\\mathrm {d}} \\left(F _ {l, 1}\\right), F _ {l, 2} = \\operatorname {C o n v} _ {3 \\times 3} ^ {\\mathrm {d}, \\text {d i l a t e d}} \\left(F _ {l, 2}\\right), \\tag {6}", + "image_path": "2f7bfa873b405ecfd2bb62ee16491cd41741ed3daf7d42f797f0644a1c786c51.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 203, + 462, + 411, + 475 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 203, + 462, + 411, + 475 + ], + "spans": [ + { + "bbox": [ + 203, + 462, + 411, + 475 + ], + "type": "interline_equation", + "content": "F _ {l + 1} = \\operatorname {C o n v} _ {1 \\times 1} \\left(\\text {U n s h u r f l e} \\left(\\operatorname {M i s h} \\left(F _ {l, 2}\\right) \\odot F _ {l, 1}\\right)\\right),", + "image_path": "3ba853a7e88ed15d1453d7bd4c1f96984c0220a1c0ea0ce0fe78034d0562b2b3.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 480, + 482, + 530 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 480, + 482, + 530 + ], + "spans": [ + { + "bbox": [ + 130, + 480, + 482, + 530 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 480, + 482, + 530 + ], + "type": "inline_equation", + "content": "\\mathrm{Conv}_{5\\times 5}^{\\mathrm{d}}" + }, + { + "bbox": [ + 130, + 480, + 482, + 530 + ], + "type": "text", + "content": " represents " + }, + { + "bbox": [ + 130, + 480, + 482, + 530 + ], + "type": "inline_equation", + "content": "5\\times 5" + }, + { + "bbox": [ + 130, + 480, + 482, + 530 + ], + "type": "text", + "content": " depth-wise convolution, " + }, + { + "bbox": [ + 130, + 480, + 482, + 530 + ], + "type": "inline_equation", + "content": "\\mathrm{Conv}_{3\\times 3}^{\\mathrm{d,dilated}}" + }, + { + "bbox": [ + 130, + 480, + 482, + 530 + ], + "type": "text", + "content": " is " + }, + { + "bbox": [ + 130, + 480, + 482, + 530 + ], + "type": "inline_equation", + "content": "3\\times 3" + }, + { + "bbox": [ + 130, + 480, + 482, + 530 + ], + "type": "text", + "content": " dilated depth-wise convolution, Shuffle and Unshuffle represent respectively the operations of pixel-shuffling and unshuffling, Mish denotes the Mish activation [52], and " + }, + { + "bbox": [ + 130, + 480, + 482, + 530 + ], + "type": "inline_equation", + "content": "F_{l + 1}" + }, + { + "bbox": [ + 130, + 480, + 482, + 530 + ], + "type": "text", + "content": " is the output of current stage passing to " + }, + { + "bbox": [ + 130, + 480, + 482, + 530 + ], + "type": "inline_equation", + "content": "l + 1" + }, + { + "bbox": [ + 130, + 480, + 482, + 530 + ], + "type": "text", + "content": "-th stage." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 131, + 548, + 371, + 560 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 548, + 371, + 560 + ], + "spans": [ + { + "bbox": [ + 131, + 548, + 371, + 560 + ], + "type": "text", + "content": "3.3 Reconstruction Loss and Correlation Loss" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 570, + 481, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 481, + 594 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 481, + 594 + ], + "type": "text", + "content": "We use the " + }, + { + "bbox": [ + 130, + 570, + 481, + 594 + ], + "type": "inline_equation", + "content": "L_{1}" + }, + { + "bbox": [ + 130, + 570, + 481, + 594 + ], + "type": "text", + "content": " norm of the pixel-wise difference between the restored high-quality image " + }, + { + "bbox": [ + 130, + 570, + 481, + 594 + ], + "type": "inline_equation", + "content": "I^{hq}" + }, + { + "bbox": [ + 130, + 570, + 481, + 594 + ], + "type": "text", + "content": " and ground-truth " + }, + { + "bbox": [ + 130, + 570, + 481, + 594 + ], + "type": "inline_equation", + "content": "I^{gt}" + }, + { + "bbox": [ + 130, + 570, + 481, + 594 + ], + "type": "text", + "content": " as the reconstruction loss, i.e.," + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 261, + 599, + 480, + 613 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 261, + 599, + 480, + 613 + ], + "spans": [ + { + "bbox": [ + 261, + 599, + 480, + 613 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {r e c} = \\left\\| I ^ {h q} - I ^ {g t} \\right\\| _ {1}. \\tag {7}", + "image_path": "4ec230cbae4af07028f06777656f98713f1218c4d30681db76626c64e78b609f.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "Furthermore, we notice that the " + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{rec}" + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": " only regulates the pixel-level similarity between the restored image and the ground-truth, while neglecting the patch-level linear correlations. The innate relationships of intensity within the image are disrupted by the consistent patterns of weather-induced degradation. By" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 134, + 91, + 448, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 91, + 448, + 103 + ], + "spans": [ + { + "bbox": [ + 134, + 91, + 448, + 103 + ], + "type": "text", + "content": "Restoring Images in Adverse Weather Conditions via Histogram Transformer" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 114, + 481, + 178 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 481, + 178 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 481, + 178 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 481, + 178 + ], + "type": "image", + "image_path": "baca57d25cae884c90257b422de3eadf3d5b76bd5cc7a3fbbfb41c8760dd389c.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 182, + 480, + 205 + ], + "lines": [ + { + "bbox": [ + 132, + 182, + 480, + 205 + ], + "spans": [ + { + "bbox": [ + 132, + 182, + 480, + 205 + ], + "type": "text", + "content": "Fig. 3: Visual comparison for desnowing on Snow100K [44]. The samples from (b) to (e) are Restormer [86], TransWeather [70], WGWSNet [100], WeatherDiff [53]." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 133, + 215, + 481, + 273 + ], + "blocks": [ + { + "bbox": [ + 133, + 215, + 481, + 273 + ], + "lines": [ + { + "bbox": [ + 133, + 215, + 481, + 273 + ], + "spans": [ + { + "bbox": [ + 133, + 215, + 481, + 273 + ], + "type": "image", + "image_path": "208d7bedddefa74b11ee3cca8783fa72640b2f53b4813c1f8dc3168f5a3dc5da.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 275, + 480, + 309 + ], + "lines": [ + { + "bbox": [ + 130, + 275, + 480, + 309 + ], + "spans": [ + { + "bbox": [ + 130, + 275, + 480, + 309 + ], + "type": "text", + "content": "Fig. 4: Visual comparison for deraining and dehazing on Outdoor-Rain [32]. The samples from (b) to (e) are Restormer [86], TransWeather [70], WGWSNet [100], WeatherDiff [53]." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 333, + 479, + 381 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 333, + 479, + 381 + ], + "spans": [ + { + "bbox": [ + 130, + 333, + 479, + 381 + ], + "type": "text", + "content": "emulating the intensity relationships within the ground-truth, we compel the degraded pixels to occupy their original positions according to the original intensity ranking. Consequently, we introduce the Pearson correlation [12] between images as a means to regulate the linear relationship, expressed as follows:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 203, + 384, + 480, + 418 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 203, + 384, + 480, + 418 + ], + "spans": [ + { + "bbox": [ + 203, + 384, + 480, + 418 + ], + "type": "interline_equation", + "content": "\\rho \\left(I ^ {h q}, I ^ {g t}\\right) = \\frac {\\sum_ {i = 1} ^ {3 H W} \\left(I _ {i} ^ {h q} - \\bar {I} ^ {h q}\\right) \\left(I _ {i} ^ {g t} - \\bar {I} ^ {g t}\\right)}{3 H W \\sigma \\left(I ^ {h q}\\right) \\sigma \\left(I ^ {g t}\\right)}, \\tag {8}", + "image_path": "3144cf403f3757a7c4ff96dacb6b4126a660f26211f277b24c6613e851e2d66f.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 424, + 479, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 424, + 479, + 486 + ], + "spans": [ + { + "bbox": [ + 130, + 424, + 479, + 486 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 424, + 479, + 486 + ], + "type": "inline_equation", + "content": "I_{i}^{\\{\\cdot\\}}" + }, + { + "bbox": [ + 130, + 424, + 479, + 486 + ], + "type": "text", + "content": " represents the " + }, + { + "bbox": [ + 130, + 424, + 479, + 486 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 424, + 479, + 486 + ], + "type": "text", + "content": "-th pixel of image, " + }, + { + "bbox": [ + 130, + 424, + 479, + 486 + ], + "type": "inline_equation", + "content": "\\overline{I}^{\\{\\cdot\\}}" + }, + { + "bbox": [ + 130, + 424, + 479, + 486 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 424, + 479, + 486 + ], + "type": "inline_equation", + "content": "\\sigma(I^{\\{\\cdot\\}})" + }, + { + "bbox": [ + 130, + 424, + 479, + 486 + ], + "type": "text", + "content": " denotes respectively the mean and the standard deviation of image sequence. Its value falls within the " + }, + { + "bbox": [ + 130, + 424, + 479, + 486 + ], + "type": "inline_equation", + "content": "[-1,1]" + }, + { + "bbox": [ + 130, + 424, + 479, + 486 + ], + "type": "text", + "content": " range. When two images exhibit perfect correlation, the value of function " + }, + { + "bbox": [ + 130, + 424, + 479, + 486 + ], + "type": "inline_equation", + "content": "\\rho" + }, + { + "bbox": [ + 130, + 424, + 479, + 486 + ], + "type": "text", + "content": " attains a value of 1, while in the case of negative correlation, its value reaches " + }, + { + "bbox": [ + 130, + 424, + 479, + 486 + ], + "type": "inline_equation", + "content": "-1" + }, + { + "bbox": [ + 130, + 424, + 479, + 486 + ], + "type": "text", + "content": ". Hence, we formulate the correlation loss as follows:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 245, + 491, + 480, + 514 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 245, + 491, + 480, + 514 + ], + "spans": [ + { + "bbox": [ + 245, + 491, + 480, + 514 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {c o r} = \\frac {1}{2} \\left(1 - \\rho (I ^ {h q}, I ^ {g t})\\right), \\tag {9}", + "image_path": "8eaac2c41fb87cc33d1f179eea9420db39cc48df85ed1df1949939331716ee93.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 517, + 479, + 541 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 517, + 479, + 541 + ], + "spans": [ + { + "bbox": [ + 130, + 517, + 479, + 541 + ], + "type": "text", + "content": "such that " + }, + { + "bbox": [ + 130, + 517, + 479, + 541 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{cor} = 0" + }, + { + "bbox": [ + 130, + 517, + 479, + 541 + ], + "type": "text", + "content": " when the recovered image perfectly aligns with the groundtruth. The overall loss function is thus defined as:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 265, + 548, + 480, + 561 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 265, + 548, + 480, + 561 + ], + "spans": [ + { + "bbox": [ + 265, + 548, + 480, + 561 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathcal {L} _ {\\text {r e c}} + \\alpha \\mathcal {L} _ {\\text {c o r}}, \\tag {10}", + "image_path": "e0a71a1b4f4eeca25b9b35d06b70b84e79b715572b0fecc0e8bfdad18adfc01c.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 566, + 311, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 566, + 311, + 578 + ], + "spans": [ + { + "bbox": [ + 130, + 566, + 311, + 578 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 566, + 311, + 578 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 130, + 566, + 311, + 578 + ], + "type": "text", + "content": " is the weight of correlation loss." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 597, + 230, + 611 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 597, + 230, + 611 + ], + "spans": [ + { + "bbox": [ + 132, + 597, + 230, + 611 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 621, + 271, + 634 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 621, + 271, + 634 + ], + "spans": [ + { + "bbox": [ + 132, + 621, + 271, + 634 + ], + "type": "text", + "content": "4.1 Experimental settings" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 641, + 479, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 479, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 479, + 665 + ], + "type": "text", + "content": "Datasets. We train our model on the same datasets as the previous works [33,53, 70] to ensure a fair comparison. The training set encompasses 9,000 images drawn" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "type": "text", + "content": "S. Sun et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 114, + 481, + 178 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 481, + 178 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 481, + 178 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 481, + 178 + ], + "type": "image", + "image_path": "19d278cb330ff500c448ffe6ea4cb25b75a15428a02051e0c1816a0eb714e19c.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 181, + 480, + 205 + ], + "lines": [ + { + "bbox": [ + 132, + 181, + 480, + 205 + ], + "spans": [ + { + "bbox": [ + 132, + 181, + 480, + 205 + ], + "type": "text", + "content": "Fig. 5: Visual comparison for raindrop removal on RainDrop [57]. The samples from (b) to (e) are Chen et al. [10], TransWeather [70], WGWSNet [100], WeatherDiff [53]." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 133, + 213, + 481, + 277 + ], + "blocks": [ + { + "bbox": [ + 133, + 213, + 481, + 277 + ], + "lines": [ + { + "bbox": [ + 133, + 213, + 481, + 277 + ], + "spans": [ + { + "bbox": [ + 133, + 213, + 481, + 277 + ], + "type": "image", + "image_path": "fac601967372a43f1453bd65a58b6f2b02d719857ead9be7b95f65e19102dbf9.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 280, + 480, + 313 + ], + "lines": [ + { + "bbox": [ + 132, + 280, + 480, + 313 + ], + "spans": [ + { + "bbox": [ + 132, + 280, + 480, + 313 + ], + "type": "text", + "content": "Fig.6: A qualitative comparison for real-world adverse weather removal on Snow100K [44]. The samples from (b) to (e) are Chen et al. [10], Restormer [86], TransWeather [70], WGWSNet [100], WeatherDiff [53]." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 334, + 480, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 334, + 480, + 418 + ], + "spans": [ + { + "bbox": [ + 130, + 334, + 480, + 418 + ], + "type": "text", + "content": "from Snow100K [44], 1,069 images sourced from Raindrop [57], and 9,000 images from Outdoor-Rain [32]. Snow100K contains synthetic images deteriorated by snow, while Raindrop comprises real raindrop-affected images. Outdoor-Rain features synthetic images afflicted by both fog and rain streaks. For evaluation, we employ the Test1 dataset [32, 33], the RainDrop test dataset [57], and the Snow100K-L and -S test sets [44]. Snow100K also provides a real-world test set containing 1,329 images affected by adverse weather." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 428, + 480, + 570 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 428, + 480, + 570 + ], + "spans": [ + { + "bbox": [ + 130, + 428, + 480, + 570 + ], + "type": "text", + "content": "Comparison Baselines. We assess the performance of our method against state-of-the-art approaches designed specifically for distinct weather removal tasks: raindrop removal, snow removal, and rain&fog removal. Specifically, for snow removal, we benchmark against SPANet [73], JSTASR [8], RESCAN [34], DesnowNet [44], and DDMSNet [94]. In the case of rain&fog removal, we compare with CycleGAN [98], pix2pix [20], HRGAN [32], MPRNet [87] and Restormer [86]. For raindrop removal, we evaluate against the methods such as pix2pix [20], DuRN [42], RaindropAttn [59], AttentiveGAN [57]. Additionally, we include some recent transformer or multi-degradation restoration networks, IDT [78], NAFNet [6], MAXIM [69], and Restormer [86], in our comparative analysis. It is worth noting that all these methods are single-task networks fine-tuned for specific datasets." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 571, + 480, + 632 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 571, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 130, + 571, + 480, + 632 + ], + "type": "text", + "content": "Furthermore, we conduct a performance comparison with the All-in-One network [33], Chen et al. [10], TransWeather [70], WGWS-Net [100], WeatherDiff [53] and AWRCP [82], which are trained to handle all the aforementioned tasks using a unified model. Note that our approach is also trained to tackle all these tasks using a single model." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 641, + 480, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 641, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 641, + 480, + 665 + ], + "type": "text", + "content": "Training details. Our implementation is realized by PyTorch [55] and on NVIDIA Tesla V100 GPU. The network is trained for a total of 300,000 iterations, with" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 134, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 134, + 91, + 447, + 102 + ], + "type": "text", + "content": "Restoring Images in Adverse Weather Conditions via Histogram Transformer" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "content": "an initial batch size of 8 and an initial patch size of 128 akin to the progressive learning pipeline [86]. We employ the AdamW optimizer [48] with an initial learning rate of " + }, + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "inline_equation", + "content": "3e^{-4}" + }, + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "content": " for the first 92,000 iterations, which is gradually reduced to " + }, + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "inline_equation", + "content": "1e^{-6}" + }, + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "content": " using cosine annealing schedule [47] during the remaining 208,000 iterations. The number of blocks at each stage " + }, + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "inline_equation", + "content": "L_{i\\in \\{1,2,3,4\\}}" + }, + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "content": " is set to " + }, + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "inline_equation", + "content": "\\{4,4,6,8\\}" + }, + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "content": " and the channel size " + }, + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "content": " is 36. The channel expansion factor " + }, + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "content": " in DGFF is set to 2.667. The numbers of heads in self-attention at different stages are set to " + }, + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "inline_equation", + "content": "\\{1,2,4,8\\}" + }, + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "content": " respectively. We randomly apply horizontal and vertical flips as the technique of data augmentation." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 241, + 356, + 253 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 241, + 356, + 253 + ], + "spans": [ + { + "bbox": [ + 132, + 241, + 356, + 253 + ], + "type": "text", + "content": "4.2 Comparisons with the state-of-the-arts" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 260, + 482, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 260, + 482, + 380 + ], + "spans": [ + { + "bbox": [ + 130, + 260, + 482, + 380 + ], + "type": "text", + "content": "Quantitative Evaluation. In our study, we provide a comprehensive comparative analysis of metrics applied to both synthetic and real datasets, as summarized in Table 1. For a fair and well-founded comparison, we utilize recent multiple degradation removal methods such as MPRNet [87], MAXIM [69], and Restormer [86], treating them as weather-specific approaches for each benchmark. Additionally, we retrain the all-in-one adverse weather removal methods including Chen et al. [10] and WGWS-Net [100] using the all-weather training dataset [33,53,70]. This exhaustive comparison reveals that our proposed method exhibits a significant performance advantage over existing approaches across three different types of degradation." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 391, + 482, + 499 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 391, + 482, + 499 + ], + "spans": [ + { + "bbox": [ + 130, + 391, + 482, + 499 + ], + "type": "text", + "content": "Qualitative Evaluation. Furthermore, we conduct a visual comparison on three tasks, and the outcomes are showcased in Figure 3, 4 and 5 respectively. Figure 6 shows a case of real-world weather removal. These results highlight that our method excels in comprehensively eliminating snow degradation, including fine and large snow spots. In contrast, the recent WeatherDiff [53] method still exhibits some residual snow degradation, and its capability to restore details is not optimal. When it comes to the restoration of challenging weather conditions, our method excels in removing complex haze and rain marks, yielding visually appealing results in comparison to prior approaches." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 515, + 244, + 526 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 515, + 244, + 526 + ], + "spans": [ + { + "bbox": [ + 132, + 515, + 244, + 526 + ], + "type": "text", + "content": "4.3 Ablation studies" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 534, + 482, + 583 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 534, + 482, + 583 + ], + "spans": [ + { + "bbox": [ + 130, + 534, + 482, + 583 + ], + "type": "text", + "content": "To substantiate the effectiveness of each component within Histoformer, we conduct a sequence of ablation studies on Outdoor-Rain [32]. In particular, we examine the impact of the dynamic-range convolution, the DHSA module, the number of bins in DHSA, the DGFF module, and the correlation loss." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": "Dynamic-range Convolution. We experiment on two settings of dynamic-range convolution, namely, sorting horizontally first and then vertically before convolution, and the reverse order. Additionally, we compared them with vanilla convolution, and the results are displayed in Table 2. The operations of regular sorting led to a performance improvement of " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "0.14\\mathrm{dB}" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": ", and the order of sorting operations does not significantly affect the outcome." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "text", + "content": "S. Sun et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 202, + 131, + 411, + 169 + ], + "blocks": [ + { + "bbox": [ + 180, + 119, + 432, + 130 + ], + "lines": [ + { + "bbox": [ + 180, + 119, + 432, + 130 + ], + "spans": [ + { + "bbox": [ + 180, + 119, + 432, + 130 + ], + "type": "text", + "content": "Table 2: Ablation studies on the dynamic-range convolution." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 202, + 131, + 411, + 169 + ], + "lines": [ + { + "bbox": [ + 202, + 131, + 411, + 169 + ], + "spans": [ + { + "bbox": [ + 202, + 131, + 411, + 169 + ], + "type": "table", + "html": "
Vanilla ConvVerti.+Horiz.Horiz.+Verti.
PSNR31.9432.0332.08
SSIM0.93770.93850.9389
", + "image_path": "901978b07e6fcfc47c502e1d767f26601c40d3e9a09e47c51f005cd8961508ef.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 157, + 201, + 279, + 262 + ], + "blocks": [ + { + "bbox": [ + 132, + 178, + 303, + 198 + ], + "lines": [ + { + "bbox": [ + 132, + 178, + 303, + 198 + ], + "spans": [ + { + "bbox": [ + 132, + 178, + 303, + 198 + ], + "type": "text", + "content": "Table 3: Ablation studies on the design of self-attention" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 157, + 201, + 279, + 262 + ], + "lines": [ + { + "bbox": [ + 157, + 201, + 279, + 262 + ], + "spans": [ + { + "bbox": [ + 157, + 201, + 279, + 262 + ], + "type": "table", + "html": "
SA typePSNRSSIM
MDTA [86]30.940.9278
TKSA [11]31.120.9295
w/o BHR31.050.9301
w/o FHR31.790.9364
DHSA32.080.9389
", + "image_path": "437adf6514946a94cbf7aedb8f85ba09b4f7f55255a5c01573aeff1e4e6eb637.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 326, + 201, + 465, + 262 + ], + "blocks": [ + { + "bbox": [ + 309, + 178, + 480, + 198 + ], + "lines": [ + { + "bbox": [ + 309, + 178, + 480, + 198 + ], + "spans": [ + { + "bbox": [ + 309, + 178, + 480, + 198 + ], + "type": "text", + "content": "Table 4: Ablation studies on the number of Bins in DHSA" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 326, + 201, + 465, + 262 + ], + "lines": [ + { + "bbox": [ + 326, + 201, + 465, + 262 + ], + "spans": [ + { + "bbox": [ + 326, + 201, + 465, + 262 + ], + "type": "table", + "html": "
#Bins in DHSAPSNRSSIM
1230.430.9283
2031.560.9312
2831.940.9379
3632.080.9389
44Out of memory
", + "image_path": "9572923e3b27ba94113abad66f030775203270612b6161b65870418abd0439f4.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 153, + 296, + 283, + 357 + ], + "blocks": [ + { + "bbox": [ + 132, + 273, + 303, + 293 + ], + "lines": [ + { + "bbox": [ + 132, + 273, + 303, + 293 + ], + "spans": [ + { + "bbox": [ + 132, + 273, + 303, + 293 + ], + "type": "text", + "content": "Table 5: Ablation studies on the choice of feed-forward module" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 153, + 296, + 283, + 357 + ], + "lines": [ + { + "bbox": [ + 153, + 296, + 283, + 357 + ], + "spans": [ + { + "bbox": [ + 153, + 296, + 283, + 357 + ], + "type": "table", + "html": "
Feed-ForwardPSNRSSIM
FFN [38]31.320.9313
GDFN [86]31.420.9347
DANB [96]31.560.9351
MSFN [11]31.780.9367
DGFF32.080.9389
", + "image_path": "69e6ded3c1ec5fe47d7bafcb610c4613b69b70113c9f1ce43ce5dd79788375ea.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 326, + 296, + 465, + 357 + ], + "blocks": [ + { + "bbox": [ + 309, + 273, + 480, + 293 + ], + "lines": [ + { + "bbox": [ + 309, + 273, + 480, + 293 + ], + "spans": [ + { + "bbox": [ + 309, + 273, + 480, + 293 + ], + "type": "text", + "content": "Table 6: Ablation studies on the setting of correlation loss" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 326, + 296, + 465, + 357 + ], + "lines": [ + { + "bbox": [ + 326, + 296, + 465, + 357 + ], + "spans": [ + { + "bbox": [ + 326, + 296, + 465, + 357 + ], + "type": "table", + "html": "
αPSNRSSIM
w/o Lcor031.770.9358
0.132.010.9369
132.080.9389
w/ Lcor532.030.9392
1031.960.9375
", + "image_path": "a75c41559691b8a9353af4169f5b79c0ff0250f7c821463a375bd07dac87c33f.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 379, + 480, + 438 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 379, + 480, + 438 + ], + "spans": [ + { + "bbox": [ + 130, + 379, + 480, + 438 + ], + "type": "text", + "content": "DHSA. To evaluate the effectiveness of the proposed DHSA module, we conduct a comparison with two baselines, i.e., a multi-Dconv head transposed attention (MDTA) [86] and a top-k sparse attention (TKSA) [11]. Additionally, we explore two additional settings of DHSA by excluding either the BHR branch or the FHR branch. The quantitative analysis results are presented in Table 3." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 439, + 482, + 535 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 439, + 482, + 535 + ], + "spans": [ + { + "bbox": [ + 130, + 439, + 482, + 535 + ], + "type": "text", + "content": "Both MDTA and TKSA integrate rich information across channels, which may result in a loss of the exploitation of long-range information across spatial dimensions. While our histogram self-attention can capture spatially long-range information, the use of either a single BHR or a single FHR branch neglects the inter-bin or inner-bin relationships, leading to inferior results. By incorporating dynamic-range convolution and dual-branch histogram self-attention, capable of extracting long-range spatial features, our DHSA enhances performance, resulting in a PSNR improvement of 0.96 dB compared to TKSA." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 546, + 482, + 606 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 546, + 482, + 606 + ], + "spans": [ + { + "bbox": [ + 130, + 546, + 482, + 606 + ], + "type": "text", + "content": "Number of Bins. To assess the influence of the number of bins in DHSA, we conduct experiments with five different values: 12, 20, 28, 36, and 44. The results are presented in Table 4. It is observed that increasing the number of bins consistently improves performance. However, when the number of bins exceeds 44, it results in an out-of-memory error." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 131, + 617, + 482, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 617, + 482, + 667 + ], + "spans": [ + { + "bbox": [ + 131, + 617, + 482, + 667 + ], + "type": "text", + "content": "DGFF. To assess the effectiveness of the proposed DGFF module, we conduct a comparison with four baselines: (i) the vanilla feed-forward network (FN) [38], (ii) a gated-Dconv feed-forward network (GDFN) [86], (iii) a dual adaptive neural block (DANB) [96], and (iv) a mixed-scale feed-forward network (MSFN) [11]." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 134, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 134, + 91, + 447, + 102 + ], + "type": "text", + "content": "Restoring Images in Adverse Weather Conditions via Histogram Transformer" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 114, + 235, + 183 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 235, + 183 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 235, + 183 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 235, + 183 + ], + "type": "image", + "image_path": "a5e963f75412604f02d4b03a668b66117c243845b7a05706c6e8ea9c7ac18246.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 238, + 116, + 304, + 180 + ], + "blocks": [ + { + "bbox": [ + 238, + 116, + 304, + 180 + ], + "lines": [ + { + "bbox": [ + 238, + 116, + 304, + 180 + ], + "spans": [ + { + "bbox": [ + 238, + 116, + 304, + 180 + ], + "type": "image", + "image_path": "acecbfef62a99b4e6ece62c03d9d41d1ea40aca6c7c8e7eaaa87e6bd58c4bcef.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 308, + 114, + 409, + 183 + ], + "blocks": [ + { + "bbox": [ + 308, + 114, + 409, + 183 + ], + "lines": [ + { + "bbox": [ + 308, + 114, + 409, + 183 + ], + "spans": [ + { + "bbox": [ + 308, + 114, + 409, + 183 + ], + "type": "image", + "image_path": "d461570eb7879ef685295620460bbad6f566c6cc5ce108f2c13b0c032e790079.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 412, + 116, + 480, + 181 + ], + "blocks": [ + { + "bbox": [ + 412, + 116, + 480, + 181 + ], + "lines": [ + { + "bbox": [ + 412, + 116, + 480, + 181 + ], + "spans": [ + { + "bbox": [ + 412, + 116, + 480, + 181 + ], + "type": "image", + "image_path": "df5e72bed4f1e2a0ffc1f28f998be7c68d69cb742c8e4d1d289f648c0fb01e12.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 133, + 184, + 235, + 253 + ], + "blocks": [ + { + "bbox": [ + 133, + 184, + 235, + 253 + ], + "lines": [ + { + "bbox": [ + 133, + 184, + 235, + 253 + ], + "spans": [ + { + "bbox": [ + 133, + 184, + 235, + 253 + ], + "type": "image", + "image_path": "3e232858f732d101478c5eb8ab9cf76a84c566659b7fdfadc060e2d6c7240059.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 202, + 254, + 236, + 262 + ], + "lines": [ + { + "bbox": [ + 202, + 254, + 236, + 262 + ], + "spans": [ + { + "bbox": [ + 202, + 254, + 236, + 262 + ], + "type": "text", + "content": "(a) Input" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 238, + 184, + 304, + 248 + ], + "blocks": [ + { + "bbox": [ + 238, + 184, + 304, + 248 + ], + "lines": [ + { + "bbox": [ + 238, + 184, + 304, + 248 + ], + "spans": [ + { + "bbox": [ + 238, + 184, + 304, + 248 + ], + "type": "image", + "image_path": "f8a8d3b27a0bc4377b8df0948e2988b05bd55feff38e7802fe57bbbf66e24ac0.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 266, + 479, + 288 + ], + "lines": [ + { + "bbox": [ + 132, + 266, + 479, + 288 + ], + "spans": [ + { + "bbox": [ + 132, + 266, + 479, + 288 + ], + "type": "text", + "content": "Fig. 7: Real-world deweathering on two snowy images [44] and their downstream detection results on Google API." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 308, + 184, + 409, + 253 + ], + "blocks": [ + { + "bbox": [ + 308, + 184, + 409, + 253 + ], + "lines": [ + { + "bbox": [ + 308, + 184, + 409, + 253 + ], + "spans": [ + { + "bbox": [ + 308, + 184, + 409, + 253 + ], + "type": "image", + "image_path": "54d9cb90a8447db778f7a100b008e5ae1581e9b0278e99464911f06ea7a8afd0.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 350, + 254, + 439, + 262 + ], + "lines": [ + { + "bbox": [ + 350, + 254, + 439, + 262 + ], + "spans": [ + { + "bbox": [ + 350, + 254, + 439, + 262 + ], + "type": "text", + "content": "(b) Deweathered by ours" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 413, + 185, + 480, + 249 + ], + "blocks": [ + { + "bbox": [ + 413, + 185, + 480, + 249 + ], + "lines": [ + { + "bbox": [ + 413, + 185, + 480, + 249 + ], + "spans": [ + { + "bbox": [ + 413, + 185, + 480, + 249 + ], + "type": "image", + "image_path": "84a07aeea60f3ad7e74ad0bf2ab70afaa89399fc0746fbbf5d91860ebc41c871.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 309, + 480, + 369 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 309, + 480, + 369 + ], + "spans": [ + { + "bbox": [ + 131, + 309, + 480, + 369 + ], + "type": "text", + "content": "The quantitative analysis results are presented in Table 5. While MSFN integrates mixed-scale information, it may still miss out on the exploitation of multi-range spatial knowledge. Through the inclusion of pixel-shuffling and feature aggregation across different ranges, our DGFF further enhances performance, resulting in a PSNR gain of 0.3 dB over MSFN." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 131, + 379, + 480, + 427 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 379, + 480, + 427 + ], + "spans": [ + { + "bbox": [ + 131, + 379, + 480, + 427 + ], + "type": "text", + "content": "Correlation Loss. Table 6 shows the effectiveness of the correlation loss " + }, + { + "bbox": [ + 131, + 379, + 480, + 427 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{cor}" + }, + { + "bbox": [ + 131, + 379, + 480, + 427 + ], + "type": "text", + "content": " and the influence of its weight. It is evident that " + }, + { + "bbox": [ + 131, + 379, + 480, + 427 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{cor}" + }, + { + "bbox": [ + 131, + 379, + 480, + 427 + ], + "type": "text", + "content": " consistently improves the performance, while the specific loss weight does not have a substantial impact on the final results. We therefore keep the loss weight as 1 by default." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 441, + 277, + 454 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 441, + 277, + 454 + ], + "spans": [ + { + "bbox": [ + 132, + 441, + 277, + 454 + ], + "type": "text", + "content": "4.4 Real-world Application" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 131, + 459, + 480, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 459, + 480, + 518 + ], + "spans": [ + { + "bbox": [ + 131, + 459, + 480, + 518 + ], + "type": "text", + "content": "To further demonstrate the practical applicability of our method for real-world adverse weather removal and its potential to improve downstream detection task, we provide two samples in Figure 7. As depicted, our Histoformer effectively eliminates snowflakes from the scene and assists the detector in recognizing omitted door and building." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 535, + 219, + 547 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 535, + 219, + 547 + ], + "spans": [ + { + "bbox": [ + 132, + 535, + 219, + 547 + ], + "type": "text", + "content": "5 Conclusion" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 131, + 558, + 479, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 558, + 479, + 665 + ], + "spans": [ + { + "bbox": [ + 131, + 558, + 479, + 665 + ], + "type": "text", + "content": "In this research, we introduce a novel mechanism called histogram self-attention and devise a new histogram transformer named Histoformer to tackle the challenge of all-in-one weather removal. Our histogram self-attention involves segmenting spatial features into multiple bins, and allocating varying attention along the bin or frequency dimension, allowing it to selectively focus on weather-related features with a dynamic range. To facilitate learning both multi-range and multi-scale information, we present DGFF module and a correlation loss. Through extensive experimentation, we demonstrate the effectiveness and superiority of our approach." + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "text", + "content": "S. Sun et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 240, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 240, + 129 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 240, + 129 + ], + "type": "text", + "content": "Acknowledgement" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 140, + 482, + 201 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 140, + 482, + 201 + ], + "spans": [ + { + "bbox": [ + 130, + 140, + 482, + 201 + ], + "type": "text", + "content": "This work has been supported in part by National Natural Science Foundation of China (No. 62322216, 62172409, 62025604, 62306308, 62311530686), in part by Shenzhen Science and Technology Program (Grant No. JCYJ20220818102012025, KQTD20221101093559018), and in part by Guangdong Provincial Key Laboratory of Information Security Technology (No. 2023B1212060026)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 220, + 197, + 232 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 220, + 197, + 232 + ], + "spans": [ + { + "bbox": [ + 133, + 220, + 197, + 232 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 245, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 142, + 245, + 480, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 142, + 245, + 480, + 267 + ], + "spans": [ + { + "bbox": [ + 142, + 245, + 480, + 267 + ], + "type": "text", + "content": "1. Ancuti, C.O., Ancuti, C.: Single image dehazing by multi-scale fusion. IEEE TIP (2013)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 142, + 269, + 459, + 279 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 142, + 269, + 459, + 279 + ], + "spans": [ + { + "bbox": [ + 142, + 269, + 459, + 279 + ], + "type": "text", + "content": "2. Berman, D., Avidan, S., et al.: Non-local image dehazing. In: CVPR (2016)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 142, + 280, + 480, + 301 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 142, + 280, + 480, + 301 + ], + "spans": [ + { + "bbox": [ + 142, + 280, + 480, + 301 + ], + "type": "text", + "content": "3. Carion, N., Massa, F., Synnaeve, G., Usunier, N., Kirillov, A., Zagoruyko, S.: End-to-end object detection with transformers. In: ECCV (2020)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 142, + 302, + 480, + 323 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 142, + 302, + 480, + 323 + ], + "spans": [ + { + "bbox": [ + 142, + 302, + 480, + 323 + ], + "type": "text", + "content": "4. Chen, H., Wang, Y., Guo, T., Xu, C., Deng, Y., Liu, Z., Ma, S., Xu, C., Xu, C., Gao, W.: Pre-trained image processing transformer. In: CVPR (2021)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 142, + 324, + 480, + 345 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 142, + 324, + 480, + 345 + ], + "spans": [ + { + "bbox": [ + 142, + 324, + 480, + 345 + ], + "type": "text", + "content": "5. Chen, J., Tan, C.H., Hou, J., Chau, L.P., Li, H.: Robust video content alignment and compensation for rain removal in a cnn framework. In: CVPR (2018)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 142, + 346, + 480, + 367 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 142, + 346, + 480, + 367 + ], + "spans": [ + { + "bbox": [ + 142, + 346, + 480, + 367 + ], + "type": "text", + "content": "6. Chen, L., Chu, X., Zhang, X., Sun, J.: Simple baselines for image restoration. In: ECCV (2022)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 142, + 368, + 480, + 400 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 142, + 368, + 480, + 400 + ], + "spans": [ + { + "bbox": [ + 142, + 368, + 480, + 400 + ], + "type": "text", + "content": "7. Chen, S., Ye, T., Liu, Y., Chen, E., Shi, J., Zhou, J.: Snowformer: Scale-aware transformer via context interaction for single image desnowing. arXiv preprint arXiv:2208.09703 (2022)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 142, + 401, + 480, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 142, + 401, + 480, + 445 + ], + "spans": [ + { + "bbox": [ + 142, + 401, + 480, + 445 + ], + "type": "text", + "content": "8. Chen, W.T., Fang, H.Y., Ding, J.J., Tsai, C.C., Kuo, S.Y.: Jstasr: Joint size and transparency-aware snow removal algorithm based on modified partial convolution and veiling effect removal. In: Vedaldi, A., Bischof, H., Brox, T., Frahm, J.M. (eds.) ECCV (2020)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 142, + 445, + 480, + 488 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 142, + 445, + 480, + 488 + ], + "spans": [ + { + "bbox": [ + 142, + 445, + 480, + 488 + ], + "type": "text", + "content": "9. Chen, W.T., Fang, H.Y., Hsieh, C.L., Tsai, C.C., Chen, I., Ding, J.J., Kuo, S.Y., et al.: All snow removed: Single image desnowing algorithm using hierarchical dual-tree complex wavelet representation and contradict channel loss. In: ICCV (2021)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 489, + 480, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 489, + 480, + 521 + ], + "spans": [ + { + "bbox": [ + 138, + 489, + 480, + 521 + ], + "type": "text", + "content": "10. Chen, W.T., Huang, Z.K., Tsai, C.C., Yang, H.H., Ding, J.J., Kuo, S.Y.: Learning multiple adverse weather removal via two-stage knowledge learning and multi-contrastive regularization: Toward a unified model. In: CVPR (2022)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 522, + 480, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 522, + 480, + 544 + ], + "spans": [ + { + "bbox": [ + 138, + 522, + 480, + 544 + ], + "type": "text", + "content": "11. Chen, X., Li, H., Li, M., Pan, J.: Learning a sparse transformer network for effective image deraining. In: CVPR (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 544, + 480, + 576 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 544, + 480, + 576 + ], + "spans": [ + { + "bbox": [ + 138, + 544, + 480, + 576 + ], + "type": "text", + "content": "12. Cohen, I., Huang, Y., Chen, J., Benesty, J., Benesty, J., Chen, J., Huang, Y., Cohen, I.: Pearson correlation coefficient. Noise reduction in speech processing (2009)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 577, + 480, + 598 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 577, + 480, + 598 + ], + "spans": [ + { + "bbox": [ + 138, + 577, + 480, + 598 + ], + "type": "text", + "content": "13. Dauphin, Y.N., Fan, A., Auli, M., Grangier, D.: Language modeling with gated convolutional networks. In: ICML (2017)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 599, + 480, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 599, + 480, + 643 + ], + "spans": [ + { + "bbox": [ + 138, + 599, + 480, + 643 + ], + "type": "text", + "content": "14. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929 (2020)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 138, + 643, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 643, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 643, + 480, + 665 + ], + "type": "text", + "content": "15. Eigen, D., Krishnan, D., Fergus, R.: Restoring an image taken through a window covered with dirt or rain. In: ICCV (2013)" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 134, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 134, + 91, + 447, + 102 + ], + "type": "text", + "content": "Restoring Images in Adverse Weather Conditions via Histogram Transformer" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 138, + 116, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 138, + 116, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 116, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 138, + 116, + 480, + 138 + ], + "type": "text", + "content": "16. Fu, H., Gong, M., Wang, C., Batmanghelich, K., Tao, D.: Deep ordinal regression network for monocular depth estimation. In: CVPR (2018)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 140, + 480, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 140, + 480, + 160 + ], + "spans": [ + { + "bbox": [ + 138, + 140, + 480, + 160 + ], + "type": "text", + "content": "17. Fu, X., Huang, J., Zeng, D., Huang, Y., Ding, X., Paisley, J.: Removing rain from single images via a deep detail network. In: CVPR (2017)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 161, + 480, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 161, + 480, + 182 + ], + "spans": [ + { + "bbox": [ + 138, + 161, + 480, + 182 + ], + "type": "text", + "content": "18. Godard, C., Mac Aodha, O., Brostow, G.J.: Unsupervised monocular depth estimation with left-right consistency. In: CVPR (2017)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 183, + 480, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 183, + 480, + 205 + ], + "spans": [ + { + "bbox": [ + 138, + 183, + 480, + 205 + ], + "type": "text", + "content": "19. Guo, C., Yan, Q., Anwar, S., Cong, R., Ren, W., Li, C.: Image dehazing transformer with transmission-aware 3d position embedding. In: CVPR (2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 205, + 480, + 225 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 205, + 480, + 225 + ], + "spans": [ + { + "bbox": [ + 138, + 205, + 480, + 225 + ], + "type": "text", + "content": "20. Isola, P., Zhu, J.Y., Zhou, T., Efros, A.A.: Image-to-image translation with conditional adversarial networks. In: CVPR (2017)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 227, + 480, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 227, + 480, + 258 + ], + "spans": [ + { + "bbox": [ + 138, + 227, + 480, + 258 + ], + "type": "text", + "content": "21. Jiang, K., Wang, Z., Yi, P., Chen, C., Wang, Z., Wang, X., Jiang, J., Lin, C.W.: Rain-free and residue hand-in-hand: A progressive coupled network for real-time image deraining. IEEE TIP (2021)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 259, + 480, + 292 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 259, + 480, + 292 + ], + "spans": [ + { + "bbox": [ + 138, + 259, + 480, + 292 + ], + "type": "text", + "content": "22. Jiang, T.X., Huang, T.Z., Zhao, X.L., Deng, L.J., Wang, Y.: A novel tensor-based video rain streaks removal approach via utilizing discriminatively intrinsic priors. In: CVPR (2017)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 293, + 480, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 293, + 480, + 314 + ], + "spans": [ + { + "bbox": [ + 138, + 293, + 480, + 314 + ], + "type": "text", + "content": "23. Kang, L.W., Lin, C.W., Fu, Y.H.: Automatic single-image-based rain streaks removal via image decomposition. IEEE TIP (2011)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 315, + 480, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 315, + 480, + 346 + ], + "spans": [ + { + "bbox": [ + 138, + 315, + 480, + 346 + ], + "type": "text", + "content": "24. Kim, Y., Cho, Y., Nguyen, T.T., Lee, D.: Metaweather: Few-shot weather-degraded image restoration via degradation pattern matching. arXiv preprint arXiv:2308.14334 (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 346, + 480, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 346, + 480, + 369 + ], + "spans": [ + { + "bbox": [ + 138, + 346, + 480, + 369 + ], + "type": "text", + "content": "25. Lai, Z., Wu, J., Chen, S., Zhou, Y., Hovakimyan, N.: Residual-based language models are free boosters for biomedical imaging tasks. In: CVPRW (2024)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 369, + 480, + 391 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 369, + 480, + 391 + ], + "spans": [ + { + "bbox": [ + 138, + 369, + 480, + 391 + ], + "type": "text", + "content": "26. Li, B., Peng, X., Wang, Z., Xu, J., Feng, D.: Aod-net: All-in-one dehazing network. In: ICCV (2017)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 392, + 480, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 392, + 480, + 413 + ], + "spans": [ + { + "bbox": [ + 138, + 392, + 480, + 413 + ], + "type": "text", + "content": "27. Li, B., Ren, W., Fu, D., Tao, D., Feng, D., Zeng, W., Wang, Z.: Benchmarking single-image dehazing and beyond. IEEE TIP (2018)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 414, + 480, + 434 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 414, + 480, + 434 + ], + "spans": [ + { + "bbox": [ + 138, + 414, + 480, + 434 + ], + "type": "text", + "content": "28. Li, B., Liu, X., Hu, P., Wu, Z., Lv, J., Peng, X.: All-in-one image restoration for unknown corruption. In: CVPR (2022)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 435, + 480, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 435, + 480, + 456 + ], + "spans": [ + { + "bbox": [ + 138, + 435, + 480, + 456 + ], + "type": "text", + "content": "29. Li, L., Dong, Y., Ren, W., Pan, J., Gao, C., Sang, N., Yang, M.H.: Semi-supervised image dehazing. IEEE TIP (2019)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 457, + 480, + 478 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 457, + 480, + 478 + ], + "spans": [ + { + "bbox": [ + 138, + 457, + 480, + 478 + ], + "type": "text", + "content": "30. Li, M., Cao, X., Zhao, Q., Zhang, L., Meng, D.: Online rain/snow removal from surveillance videos. IEEE TIP (2021)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 479, + 480, + 501 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 479, + 480, + 501 + ], + "spans": [ + { + "bbox": [ + 138, + 479, + 480, + 501 + ], + "type": "text", + "content": "31. Li, P., Yun, M., Tian, J., Tang, Y., Wang, G., Wu, C.: Stacked dense networks for single-image snow removal. Neurocomputing (2019)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 502, + 480, + 522 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 502, + 480, + 522 + ], + "spans": [ + { + "bbox": [ + 138, + 502, + 480, + 522 + ], + "type": "text", + "content": "32. Li, R., Cheong, L.F., Tan, R.T.: Heavy rain image restoration: Integrating physics model and conditional adversarial learning. In: CVPR (2019)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 138, + 523, + 480, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 523, + 480, + 544 + ], + "spans": [ + { + "bbox": [ + 138, + 523, + 480, + 544 + ], + "type": "text", + "content": "33. Li, R., Tan, R.T., Cheong, L.F.: All in one bad weather removal using architectural search. In: CVPR (2020)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 138, + 545, + 480, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 545, + 480, + 567 + ], + "spans": [ + { + "bbox": [ + 138, + 545, + 480, + 567 + ], + "type": "text", + "content": "34. Li, X., Wu, J., Lin, Z., Liu, H., Zha, H.: Recurrent squeeze-and-excitation context aggregation net for single image deraining. In: ECCV (2018)" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 138, + 567, + 480, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 567, + 480, + 588 + ], + "spans": [ + { + "bbox": [ + 138, + 567, + 480, + 588 + ], + "type": "text", + "content": "35. Li, Y., Tan, R.T., Guo, X., Lu, J., Brown, M.S.: Rain streak removal using layer priors. In: CVPR (2016)" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 138, + 589, + 480, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 589, + 480, + 610 + ], + "spans": [ + { + "bbox": [ + 138, + 589, + 480, + 610 + ], + "type": "text", + "content": "36. Li, Y., Lu, J., Chen, H., Wu, X., Chen, X.: Dilated convolutional transformer for high-quality image deraining. In: CVPRW (June 2023)" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 138, + 611, + 480, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 611, + 480, + 643 + ], + "spans": [ + { + "bbox": [ + 138, + 611, + 480, + 643 + ], + "type": "text", + "content": "37. Li, Z., Guan, B., Wei, Y., Zhou, Y., Zhang, J., Xu, J.: Mapping new realities: Ground truth image creation with pix2pix image-to-image translation. arXiv preprint arXiv:2404.19265 (2024)" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 138, + 643, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 643, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 643, + 480, + 665 + ], + "type": "text", + "content": "38. Liang, J., Cao, J., Sun, G., Zhang, K., Van Gool, L., Timofte, R.: Swinir: Image restoration using swin transformer. In: ICCV (2021)" + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "type": "text", + "content": "S. Sun et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 138, + 116, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 138, + 116, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 116, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 138, + 116, + 480, + 138 + ], + "type": "text", + "content": "39. Liang, Y., Anwar, S., Liu, Y.: Drt: A lightweight single image deraining recursive transformer. In: CVPR (2022)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 139, + 480, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 139, + 480, + 160 + ], + "spans": [ + { + "bbox": [ + 138, + 139, + 480, + 160 + ], + "type": "text", + "content": "40. Liu, J., Yang, W., Yang, S., Guo, Z.: Erase or fill? deep joint recurrent rain removal and reconstruction in videos. In: CVPR (2018)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 161, + 480, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 161, + 480, + 182 + ], + "spans": [ + { + "bbox": [ + 138, + 161, + 480, + 182 + ], + "type": "text", + "content": "41. Liu, K., Jiang, Y., Choi, I., Gu, J.: Learning image-adaptive codebooks for class-agnostic image restoration. arXiv preprint arXiv:2306.06513 (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 183, + 480, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 183, + 480, + 205 + ], + "spans": [ + { + "bbox": [ + 138, + 183, + 480, + 205 + ], + "type": "text", + "content": "42. Liu, X., Suganuma, M., Sun, Z., Okatani, T.: Dual residual networks leveraging the potential of paired operations for image restoration. In: CVPR (2019)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 205, + 480, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 205, + 480, + 226 + ], + "spans": [ + { + "bbox": [ + 138, + 205, + 480, + 226 + ], + "type": "text", + "content": "43. Liu, Y., Liu, H., Li, L., Wu, Z., Chen, J.: A data-centric solution to nonhomogeneous dehazing via vision transformer. In: CVPR (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 227, + 480, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 227, + 480, + 248 + ], + "spans": [ + { + "bbox": [ + 138, + 227, + 480, + 248 + ], + "type": "text", + "content": "44. Liu, Y.F., Jaw, D.W., Huang, S.C., Hwang, J.N.: Desnownet: Context-aware deep network for snow removal. IEEE TIP (2018)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 249, + 480, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 249, + 480, + 270 + ], + "spans": [ + { + "bbox": [ + 138, + 249, + 480, + 270 + ], + "type": "text", + "content": "45. Liu, Z., Lin, Y., Cao, Y., Hu, H., Wei, Y., Zhang, Z., Lin, S., Guo, B.: Swin transformer: Hierarchical vision transformer using shifted windows. In: ICCV (2021)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 271, + 480, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 271, + 480, + 293 + ], + "spans": [ + { + "bbox": [ + 138, + 271, + 480, + 293 + ], + "type": "text", + "content": "46. Liu, Z., Ning, J., Cao, Y., Wei, Y., Zhang, Z., Lin, S., Hu, H.: Video swim transformer. In: CVPR (2022)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 293, + 480, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 293, + 480, + 314 + ], + "spans": [ + { + "bbox": [ + 138, + 293, + 480, + 314 + ], + "type": "text", + "content": "47. Loshchilov, I., Hutter, F.: Sgdr: Stochastic gradient descent with warm restarts. arXiv preprint arXiv:1608.03983 (2016)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 315, + 480, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 315, + 480, + 335 + ], + "spans": [ + { + "bbox": [ + 138, + 315, + 480, + 335 + ], + "type": "text", + "content": "48. Loshchilov, I., Hutter, F.: Fixing weight decay regularization in adam. arXiv preprint arXiv:1711.05101 (2017)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 336, + 480, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 336, + 480, + 369 + ], + "spans": [ + { + "bbox": [ + 138, + 336, + 480, + 369 + ], + "type": "text", + "content": "49. Luo, Y., Zhao, R., Wei, X., Chen, J., Lu, Y., Xie, S., Wang, T., Xiong, R., Lu, M., Zhang, S.: Mowe: Mixture of weather experts for multiple adverse weather removal. arXiv preprint arXiv:2303.13739 (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 369, + 480, + 391 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 369, + 480, + 391 + ], + "spans": [ + { + "bbox": [ + 138, + 369, + 480, + 391 + ], + "type": "text", + "content": "50. Lyu, W., Zheng, S., Ling, H., Chen, C.: Backdoor attacks against transformers with attention enhancement. In: ICLR Workshop (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 392, + 480, + 423 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 392, + 480, + 423 + ], + "spans": [ + { + "bbox": [ + 138, + 392, + 480, + 423 + ], + "type": "text", + "content": "51. Ma, H., Zeng, D., Liu, Y.: Learning individualized treatment rules with many treatments: A supervised clustering approach using adaptive fusion. NeurIPS (2022)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 424, + 480, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 424, + 480, + 445 + ], + "spans": [ + { + "bbox": [ + 138, + 424, + 480, + 445 + ], + "type": "text", + "content": "52. Misra, D.: Mish: A self regularized non-monotonic activation function. arXiv preprint arXiv:1908.08681 (2019)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 447, + 480, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 447, + 480, + 468 + ], + "spans": [ + { + "bbox": [ + 138, + 447, + 480, + 468 + ], + "type": "text", + "content": "53. Ozdenizci, O., Legenstein, R.: Restoring vision in adverse weather conditions with patch-based denoising diffusion models. IEEE TPAMI (2023)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 468, + 480, + 500 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 468, + 480, + 500 + ], + "spans": [ + { + "bbox": [ + 138, + 468, + 480, + 500 + ], + "type": "text", + "content": "54. Park, D., Lee, B.H., Chun, S.Y.: All-in-one image restoration for unknown degradations using adaptive discriminative filters for specific degradations. In: CVPR (2023)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 501, + 480, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 501, + 480, + 533 + ], + "spans": [ + { + "bbox": [ + 138, + 501, + 480, + 533 + ], + "type": "text", + "content": "55. Paszke, A., Gross, S., Massa, F., Lerer, A., Bradbury, J., Chanan, G., Killeen, T., Lin, Z., Gimelshein, N., Antiga, L., et al.: Pytorch: An imperative style, high-performance deep learning library. NeurIPS (2019)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 138, + 534, + 480, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 534, + 480, + 555 + ], + "spans": [ + { + "bbox": [ + 138, + 534, + 480, + 555 + ], + "type": "text", + "content": "56. Patil, P.W., Gupta, S., Rana, S., Venkatesh, S., Murala, S.: Multi-weather image restoration via domain translation. In: ICCV (2023)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 138, + 555, + 480, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 555, + 480, + 578 + ], + "spans": [ + { + "bbox": [ + 138, + 555, + 480, + 578 + ], + "type": "text", + "content": "57. Qian, R., Tan, R.T., Yang, W., Su, J., Liu, J.: Attentive generative adversarial network for raindrop removal from a single image. In: CVPR (2018)" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 138, + 578, + 480, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 578, + 480, + 599 + ], + "spans": [ + { + "bbox": [ + 138, + 578, + 480, + 599 + ], + "type": "text", + "content": "58. Quan, R., Yu, X., Liang, Y., Yang, Y.: Removing raindrops and rain streaks in one go. In: CVPR (2021)" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 138, + 600, + 480, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 600, + 480, + 621 + ], + "spans": [ + { + "bbox": [ + 138, + 600, + 480, + 621 + ], + "type": "text", + "content": "59. Quan, Y., Deng, S., Chen, Y., Ji, H.: Deep learning for seeing through window with raindrops. In: ICCV (2019)" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 138, + 622, + 480, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 622, + 480, + 643 + ], + "spans": [ + { + "bbox": [ + 138, + 622, + 480, + 643 + ], + "type": "text", + "content": "60. Ren, S., He, K., Girshick, R., Sun, J.: Faster r-cnn: Towards real-time object detection with region proposal networks. NeurIPS (2015)" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 138, + 644, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 644, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 644, + 480, + 665 + ], + "type": "text", + "content": "61. Ren, W., Tian, J., Han, Z., Chan, A., Tang, Y.: Video desnowing and deraining based on matrix decomposition. In: CVPR (2017)" + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 134, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 134, + 91, + 447, + 102 + ], + "type": "text", + "content": "Restoring Images in Adverse Weather Conditions via Histogram Transformer" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 137, + 116, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 137, + 116, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 116, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 137, + 116, + 480, + 138 + ], + "type": "text", + "content": "62. Ren, W., Liu, S., Zhang, H., Pan, J., Cao, X., Yang, M.H.: Single image dehazing via multi-scale convolutional neural networks. In: ECCV (2016)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 137, + 139, + 480, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 139, + 480, + 160 + ], + "spans": [ + { + "bbox": [ + 137, + 139, + 480, + 160 + ], + "type": "text", + "content": "63. Ren, W., Ma, L., Zhang, J., Pan, J., Cao, X., Liu, W., Yang, M.H.: Gated fusion network for single image dehazing. In: CVPR (2018)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 137, + 161, + 480, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 161, + 480, + 182 + ], + "spans": [ + { + "bbox": [ + 137, + 161, + 480, + 182 + ], + "type": "text", + "content": "64. Shao, Y., Li, L., Ren, W., Gao, C., Sang, N.: Domain adaptation for image dehazing. In: CVPR (2020)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 137, + 183, + 480, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 183, + 480, + 205 + ], + "spans": [ + { + "bbox": [ + 137, + 183, + 480, + 205 + ], + "type": "text", + "content": "65. Song, Y., He, Z., Qian, H., Du, X.: Vision transformers for single image dehazing. IEEE TIP (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 137, + 205, + 480, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 205, + 480, + 226 + ], + "spans": [ + { + "bbox": [ + 137, + 205, + 480, + 226 + ], + "type": "text", + "content": "66. Sun, S., Ren, W., Li, J., Zhang, K., Liang, M., Cao, X.: Event-aware video deraining via multi-patch progressive learning. IEEE TIP (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 137, + 227, + 480, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 227, + 480, + 248 + ], + "spans": [ + { + "bbox": [ + 137, + 227, + 480, + 248 + ], + "type": "text", + "content": "67. Sun, S., Ren, W., Wang, T., Cao, X.: Rethinking image restoration for object detection. NeurIPS (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 137, + 249, + 480, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 249, + 480, + 281 + ], + "spans": [ + { + "bbox": [ + 137, + 249, + 480, + 281 + ], + "type": "text", + "content": "68. Tan, Z., Wu, Y., Liu, Q., Chu, Q., Lu, L., Ye, J., Yu, N.: Exploring the application of large-scale pre-trained models on adverse weather removal. arXiv preprint arXiv:2306.09008 (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 137, + 281, + 480, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 281, + 480, + 303 + ], + "spans": [ + { + "bbox": [ + 137, + 281, + 480, + 303 + ], + "type": "text", + "content": "69. Tu, Z., Talebi, H., Zhang, H., Yang, F., Milanfar, P., Bovik, A., Li, Y.: Maxim: Multi-axis mlp for image processing. In: CVPR (2022)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 137, + 304, + 480, + 325 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 304, + 480, + 325 + ], + "spans": [ + { + "bbox": [ + 137, + 304, + 480, + 325 + ], + "type": "text", + "content": "70. Valanarasu, J.M.J., Yasarla, R., Patel, V.M.: Transweather: Transformer-based restoration of images degraded by adverse weather conditions. In: CVPR (2022)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 137, + 326, + 480, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 326, + 480, + 357 + ], + "spans": [ + { + "bbox": [ + 137, + 326, + 480, + 357 + ], + "type": "text", + "content": "71. Wan, Y.C., Shao, M.W., Cheng, Y.S., Liu, Y.X., Bao, Z.Y., Meng, D.Y.: Restoring images captured in arbitrary hybrid adverse weather conditions in one go. arXiv preprint arXiv:2305.09996 (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 137, + 358, + 480, + 390 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 358, + 480, + 390 + ], + "spans": [ + { + "bbox": [ + 137, + 358, + 480, + 390 + ], + "type": "text", + "content": "72. Wang, T., Zhang, K., Shao, Z., Luo, W., Stenger, B., Lu, T., Kim, T.K., Liu, W., Li, H.: Gridformer: Residual dense transformer with grid structure for image restoration in adverse weather conditions. arXiv preprint arXiv:2305.17863 (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 137, + 391, + 480, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 391, + 480, + 413 + ], + "spans": [ + { + "bbox": [ + 137, + 391, + 480, + 413 + ], + "type": "text", + "content": "73. Wang, T., Yang, X., Xu, K., Chen, S., Zhang, Q., Lau, R.W.: Spatial attentive single-image deraining with a high quality real rain dataset. In: CVPR (2019)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 137, + 414, + 480, + 434 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 414, + 480, + 434 + ], + "spans": [ + { + "bbox": [ + 137, + 414, + 480, + 434 + ], + "type": "text", + "content": "74. Wang, Y., Ma, C., Liu, J.: Smartassign: Learning a smart knowledge assignment strategy for deraining and desnowing. In: CVPR (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 137, + 435, + 480, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 435, + 480, + 456 + ], + "spans": [ + { + "bbox": [ + 137, + 435, + 480, + 456 + ], + "type": "text", + "content": "75. Wang, Z., Cun, X., Bao, J., Zhou, W., Liu, J., Li, H.: Uformer: A general u-shaped transformer for image restoration. In: CVPR (2022)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 137, + 457, + 480, + 478 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 457, + 480, + 478 + ], + "spans": [ + { + "bbox": [ + 137, + 457, + 480, + 478 + ], + "type": "text", + "content": "76. Wei, W., Yi, L., Xie, Q., Zhao, Q., Meng, D., Xu, Z.: Should we encode rain streaks in video as deterministic or stochastic? In: ICCV (2017)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 137, + 479, + 480, + 501 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 479, + 480, + 501 + ], + "spans": [ + { + "bbox": [ + 137, + 479, + 480, + 501 + ], + "type": "text", + "content": "77. Wu, H., Qu, Y., Lin, S., Zhou, J., Qiao, R., Zhang, Z., Xie, Y., Ma, L.: Contrastive learning for compact single image dehazing. In: CVPR (2021)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 137, + 502, + 480, + 522 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 502, + 480, + 522 + ], + "spans": [ + { + "bbox": [ + 137, + 502, + 480, + 522 + ], + "type": "text", + "content": "78. Xiao, J., Fu, X., Liu, A., Wu, F., Zha, Z.J.: Image de-raining transformer. IEEE TPAMI (2022)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 137, + 523, + 480, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 523, + 480, + 544 + ], + "spans": [ + { + "bbox": [ + 137, + 523, + 480, + 544 + ], + "type": "text", + "content": "79. Yang, W., Tan, R.T., Feng, J., Liu, J., Guo, Z., Yan, S.: Deep joint rain detection and removal from a single image. In: CVPR (2017)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 137, + 545, + 480, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 545, + 480, + 567 + ], + "spans": [ + { + "bbox": [ + 137, + 545, + 480, + 567 + ], + "type": "text", + "content": "80. Yasarla, R., Patel, V.M.: Uncertainty guided multi-scale residual learning-using a cycle spinning cnn for single image de-raining. In: CVPR (2019)" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 137, + 567, + 480, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 567, + 480, + 588 + ], + "spans": [ + { + "bbox": [ + 137, + 567, + 480, + 588 + ], + "type": "text", + "content": "81. Yasarla, R., Sindagi, V.A., Patel, V.M.: Syn2real transfer learning for image deraining using gaussian processes. In: CVPR (2020)" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 137, + 589, + 480, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 589, + 480, + 610 + ], + "spans": [ + { + "bbox": [ + 137, + 589, + 480, + 610 + ], + "type": "text", + "content": "82. Ye, T., Chen, S., Bai, J., Shi, J., Xue, C., Jiang, J., Yin, J., Chen, E., Liu, Y.: Adverse weather removal with codebook priors. In: ICCV (2023)" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 137, + 611, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 611, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 137, + 611, + 480, + 632 + ], + "type": "text", + "content": "83. You, S., Tan, R.T., Kawakami, R., Mukaigawa, Y., Ikeuchi, K.: Adherent raindrop modeling, detection and removal in video. IEEE TPAMI (2015)" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 137, + 633, + 480, + 654 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 633, + 480, + 654 + ], + "spans": [ + { + "bbox": [ + 137, + 633, + 480, + 654 + ], + "type": "text", + "content": "84. Yu, F., Koltun, V.: Multi-scale context aggregation by dilated convolutions. arXiv preprint arXiv:1511.07122 (2016)" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 137, + 654, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 654, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 137, + 654, + 480, + 665 + ], + "type": "text", + "content": "85. Yu, F., Koltun, V., Funkhouser, T.: Dilated residual networks. In: CVPR (2017)" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "type": "text", + "content": "S. Sun et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + }, + { + "para_blocks": [ + { + "bbox": [ + 138, + 116, + 480, + 468 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 138, + 116, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 116, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 138, + 116, + 480, + 138 + ], + "type": "text", + "content": "86. Zamir, S.W., Arora, A., Khan, S., Hayat, M., Khan, F.S., Yang, M.H.: Restormer: Efficient transformer for high-resolution image restoration. In: CVPR (2022)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 140, + 480, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 140, + 480, + 160 + ], + "spans": [ + { + "bbox": [ + 138, + 140, + 480, + 160 + ], + "type": "text", + "content": "87. Zamir, S.W., Arora, A., Khan, S., Hayat, M., Khan, F.S., Yang, M.H., Shao, L.: Multi-stage progressive image restoration. In: CVPR (2021)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 161, + 480, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 161, + 480, + 182 + ], + "spans": [ + { + "bbox": [ + 138, + 161, + 480, + 182 + ], + "type": "text", + "content": "88. Zhang, H., Patel, V.M.: Density-aware single image de-raining using a multi-stream dense network. In: CVPR (2018)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 183, + 480, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 183, + 480, + 205 + ], + "spans": [ + { + "bbox": [ + 138, + 183, + 480, + 205 + ], + "type": "text", + "content": "89. Zhang, H., Sindagi, V., Patel, V.M.: Image de-raining using a conditional generative adversarial network. IEEE TCSVT (2019)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 205, + 480, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 205, + 480, + 226 + ], + "spans": [ + { + "bbox": [ + 138, + 205, + 480, + 226 + ], + "type": "text", + "content": "90. Zhang, H., Sindagi, V., Patel, V.M.: Joint transmission map estimation and de-hazing using deep networks. IEEE TCSVT (2019)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 227, + 480, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 227, + 480, + 258 + ], + "spans": [ + { + "bbox": [ + 138, + 227, + 480, + 258 + ], + "type": "text", + "content": "91. Zhang, H., Ba, Y., Yang, E., Mehra, V., Gella, B., Suzuki, A., Pfahnl, A., Chandrappa, C.C., Wong, A., Kadambi, A.: Weatherstream: Light transport automation of single image deweathering. In: CVPR (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 259, + 480, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 259, + 480, + 281 + ], + "spans": [ + { + "bbox": [ + 138, + 259, + 480, + 281 + ], + "type": "text", + "content": "92. Zhang, J., Ren, W., Zhang, S., Zhang, H., Nie, Y., Xue, Z., Cao, X.: Hierarchical density-aware dehazing network. IEEE Transactions on Cybernetics (2021)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 282, + 480, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 282, + 480, + 303 + ], + "spans": [ + { + "bbox": [ + 138, + 282, + 480, + 303 + ], + "type": "text", + "content": "93. Zhang, K., Li, D., Luo, W., Ren, W.: Dual attention-in-attention model for joint rain streak and raindrop removal. In: IEEE TIP (2021)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 304, + 480, + 325 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 304, + 480, + 325 + ], + "spans": [ + { + "bbox": [ + 138, + 304, + 480, + 325 + ], + "type": "text", + "content": "94. Zhang, K., Li, R., Yu, Y., Luo, W., Li, C.: Deep dense multi-scale network for snow removal using semantic and depth priors. IEEE TIP (2021)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 326, + 480, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 326, + 480, + 346 + ], + "spans": [ + { + "bbox": [ + 138, + 326, + 480, + 346 + ], + "type": "text", + "content": "95. Zhang, X., Li, H., Qi, Y., Leow, W.K., Ng, T.K.: Rain removal in video by combining temporal and chromatic properties. In: ICME (2006)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 347, + 480, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 347, + 480, + 369 + ], + "spans": [ + { + "bbox": [ + 138, + 347, + 480, + 369 + ], + "type": "text", + "content": "96. Zhao, H., Gou, Y., Li, B., Peng, D., Lv, J., Peng, X.: Comprehensive and delicate: An efficient transformer for image restoration. In: CVPR (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 369, + 480, + 391 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 369, + 480, + 391 + ], + "spans": [ + { + "bbox": [ + 138, + 369, + 480, + 391 + ], + "type": "text", + "content": "97. Zheng, Y., Zhan, J., He, S., Dong, J., Du, Y.: Curricular contrastive regularization for physics-aware single image dehazing. In: CVPR (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 392, + 480, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 392, + 480, + 413 + ], + "spans": [ + { + "bbox": [ + 138, + 392, + 480, + 413 + ], + "type": "text", + "content": "98. Zhu, J.Y., Park, T., Isola, P., Efros, A.A.: Unpaired image-to-image translation using cycle-consistent adversarial networks. In: ICCV (2017)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 414, + 480, + 434 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 414, + 480, + 434 + ], + "spans": [ + { + "bbox": [ + 138, + 414, + 480, + 434 + ], + "type": "text", + "content": "99. Zhu, L., Fu, C.W., Lischinski, D., Heng, P.A.: Joint bi-layer optimization for single-image rain streak removal. In: ICCV (2017)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 435, + 480, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 435, + 480, + 468 + ], + "spans": [ + { + "bbox": [ + 138, + 435, + 480, + 468 + ], + "type": "text", + "content": "00. Zhu, Y., Wang, T., Fu, X., Yang, X., Guo, X., Dai, J., Qiao, Y., Hu, X.: Learning weather-general and weather-specific features for image restoration under multiple adverse weather conditions. In: CVPR (2023)" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 134, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 134, + 91, + 447, + 102 + ], + "type": "text", + "content": "Restoring Images in Adverse Weather Conditions via Histogram Transformer" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "19" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 18 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Retargeting Visual Data with Deformation Fields/547a2a3d-69aa-4867-932e-8316bae28b65_content_list.json b/2024/Retargeting Visual Data with Deformation Fields/547a2a3d-69aa-4867-932e-8316bae28b65_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..18f86f042a31bc102b728d97ab893bd9b0222f37 --- /dev/null +++ b/2024/Retargeting Visual Data with Deformation Fields/547a2a3d-69aa-4867-932e-8316bae28b65_content_list.json @@ -0,0 +1,1907 @@ +[ + { + "type": "text", + "text": "Retargeting Visual Data with Deformation Fields", + "text_level": 1, + "bbox": [ + 218, + 140, + 782, + 162 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Tim Elsner1, Julia Berger1, Tong Wu2, Victor Czech1, Lin Gao2, and Leif Kobbelt1", + "bbox": [ + 228, + 188, + 774, + 219 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ RWTH Aachen University, Germany \n{elsner, czech, kobbelt}@cs.rwth-aachen.de / julia.berger@rwth-aachen.de \n $^{2}$ University of Chinese Academy of Sciences, China \n{wutong19s, gaolin}@ict.ac.cn", + "bbox": [ + 228, + 229, + 771, + 287 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Seam carving is an image editing method that enables content-aware resizing, including operations like removing objects. However, the seam-finding strategy based on dynamic programming or graph-cut limits its applications to broader visual data formats and degrees of freedom for editing. Our observation is that describing the editing and retargeting of images more generally by a deformation field yields a generalisation of content-aware deformations. We propose to learn a deformation with a neural network that keeps the output plausible while trying to deform it only in places with low information content. This technique applies to different kinds of visual data, including images, 3D scenes given as neural radiance fields, or even polygon meshes. Experiments conducted on different visual data show that our method achieves better content-aware retargeting compared to previous methods.", + "bbox": [ + 261, + 321, + 746, + 501 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Image Retargeting $\\cdot$ Neural Fields $\\cdot$ Neural Radiance Fields", + "bbox": [ + 261, + 513, + 740, + 527 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 551, + 375, + 568 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Media retargeting is a way to edit images, videos, 3D objects, or even entire 3D scenes by a global deformation such that relevant content, details, and features are properly preserved. The overall goal is to change the aspect ratio of the bounding box of the data so that it fits an allocated space, e.g. fitting an image to a display with prescribed format. Retargeting is based on the idea of identifying regions with little detail and to accumulate the necessary distortion induced by the deformation in these regions. Small editing operations are possible e.g. through concentrating the distortion there, specifying a region containing a particular object to have this object removed.", + "bbox": [ + 212, + 583, + 787, + 718 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Most existing retargeting methods typically use some form of seam carving, i.e. the deletion of discrete pixels on a path through the image (seams) [2,7,9,44]. As previous methods perform a discrete removal or addition of pixels, they do not natively extend to e.g. continuous domains like Neural Radiance Fields [30] (NeRFs) and are often restricted to a class of solutions defined by the used algorithm.", + "bbox": [ + 212, + 719, + 787, + 808 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Our approach formulates the problem via continuous deformation fields, mapping from the retargeted data back to the undistorted input. It natively extends", + "bbox": [ + 212, + 809, + 787, + 840 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/0a0e5a4383146ff70bea9dc83b97ce4019abf8cf7d9553d9c68c07bb9972fdd6.jpg", + "image_caption": [ + "Fig.1: Different objectives applied to different visual domains with our approach: We demonstrate retargeting images, NeRFs, and meshes. 'Surfer' image from [42]." + ], + "image_footnote": [], + "bbox": [ + 225, + 151, + 514, + 340 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/1ed9a438e62e49e798436fd982c2c3c38cffb96881319291972051c854262778.jpg", + "image_caption": [ + "Fig.2: The family of image resizing methods. Our approach formulates the problem in a more general way, but can utilise e.g. new energy formulations as well." + ], + "image_footnote": [], + "bbox": [ + 583, + 148, + 807, + 316 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "from images on other domains, like NeRFs and polygon meshes (Fig. 1), and offers better solutions through a global and continuous instead of a discrete and greedy optimisation process (see Fig. 5). Our contributions are as follows:", + "bbox": [ + 212, + 446, + 784, + 492 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We introduce the use of a neural deformation field to compress or stretch unimportant regions to achieve a smaller or larger output or to follow other editing objectives.", + "- We abstract and generalise the underlying rules of creating plausible outputs with seam carving, then demonstrate the domain agnostic nature of our formulation by applying it to images, 3D meshes, and 3D scenes in the form of NeRFs.", + "- We show that the high flexibility of our approach from optimising a global deformation field produces better outputs than iteratively computing seams, while also allowing configurations that are not possible with previous methods." + ], + "bbox": [ + 223, + 500, + 784, + 660 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To achieve this, we regularise the neural deformation fields to follow general sanity guidelines, while attempting to minimise distortion in places with presumed high information content. We evaluate our results on both qualitative and quantitative levels. To showcase our approach on interesting scenes captured as NeRFs and inspired by the dataset of Richter et al. [41], we also provide a small synthetic NeRF retargeting dataset captured in the video game GTA V [34] by flying the camera over the virtual city, capturing the screen content and camera parameters.", + "bbox": [ + 212, + 672, + 787, + 792 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our method is a more reliable backbone for approaches following seam carving while outperforming current methods, requires only minor modifications for vastly different applications and domains, and is straightforward to implement.", + "bbox": [ + 212, + 795, + 785, + 839 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "T. Elsner, J. Berger et al.", + "bbox": [ + 271, + 114, + 442, + 128 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We provide the code of our method demonstrated on images in the supplemental.", + "bbox": [ + 217, + 146, + 782, + 161 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 217, + 200, + 385, + 215 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As our approach is re-formulating seam carving, it produces results from a single example without any learned prior, and is designed to allow seam carving-like deformations for (learned) 3D representations. We hence discuss these three branches of related work.", + "bbox": [ + 217, + 233, + 784, + 292 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Single Example Generation The goal of single example generation is to discover similar local patterns from a single input and synthesise novel samples. Texture synthesis on 2D images and 3D shapes has been widely studied in computer vision and computer graphics by non-neural methods [10, 11, 24, 51]. There are also methods focusing on synthesising geometric textures for 3D shapes [4, 19, 25], which transfer local geometry details from an example shape to other surfaces. Recently, neural-based methods have been proposed to generate 2D textures [14, 45, 56] and transfer them to 3D shapes [17, 18], resulting in improved results. With recent advances in implicit representation [6, 29, 35] and neural radiance fields [30], Wu et al. propose to learn implicit 3D shapes from a shape with [54] or without texture [55]. To enhance the realism of synthesised shapes, Huang et al. [21] reconstruct the 3D shape with a NeRF by optimising texture features on a reconstructed mesh. These features can be applied to arbitrary surfaces to render the target surface with synthesised textures. These methods often only work for a certain kind of data format, while our method proposes a domain agnostic approach to synthesise visual data from a single example.", + "bbox": [ + 217, + 310, + 785, + 551 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Deformations for Learned 3D Representations Neural radiance fields (NeRF), as a newly proposed 3D shape representation, has been widely used in scene reconstruction and novel view synthesis. However, the original NeRF only works for static scenes, thus is unable to deal with deformation in dynamic scenes or generate motions for a static scene. To overcome this, the deformation field is introduced to NeRF. NeRF-Editing [59] is the first to explore deformation in NeRF. It first reconstructs the explicit mesh of a static scene with NeuS [50] and deforms the explicit mesh with the ARAP algorithm [47]. Then sample points in volume rendering are deformed along with the mesh via barycentric coordinate interpolation, resulting in the change of rendered images. Follow-up NeRF deformation methods [13,39,57] employ a similar pipeline but use different geometry proxies. To reconstruct dynamic scenes, Albert et al. [40] first explore the possibility by adding an extra time-conditioned deformation field to the static NeRF. The time-conditioned deformation field transforms sample points in the observation space into the canonical space, where their colours and opacities are queried and then rendered into images. This idea stimulated a series of dynamic NeRF reconstruction methods [5,28,36,37,52,58]. While our approach also uses a deformation field, we inject general plausibility constraints into it to keep the", + "bbox": [ + 217, + 566, + 785, + 839 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "Retargeting Visual Data with Deformation Fields", + "bbox": [ + 401, + 114, + 730, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "results reasonable, enabling the idea behind seam carving on different types of visual data.", + "bbox": [ + 212, + 146, + 785, + 175 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Seam Carving Image/video retargeting is an important tool in computer vision and computer graphics. Early works [26] first annotate or detect the region of interest (ROI), which is uniformly scaled to the target size afterwards. While contents out of the ROI are deformed by fisheye warping, Liu and Gleicher [27] later extend it to the video domain by introducing a motion salience map. Avidan and Shamir [2] propose seam carving, a content-aware image resizing method. It first defines a saliency map by summing the gradient norms along the horizontal and vertical directions. The cost of a seam is defined by adding up the values on the salience map (energy). The seam with minimal energy can be found using dynamic programming and will be removed in the resizing process. Rubinstein et al. [43] extend the seam carving idea to the video domain and transform the dynamic programming problem to a graph cut problem. Follow-up works [7,9,44] incorporate homogeneous cropping and scaling with seam carving and excavate more high-level information like symmetry [53], depth [3], and object similarity [8] to assist seam carving. With the advances in deep learning, a few methods [31-33] made attempts to detect and localise the seam carving forgery. Song et al. [46] propose Carving-Net to replace the previous handcrafted energy map with a neural network while the seam finding algorithm is still dynamic programming.", + "bbox": [ + 212, + 191, + 787, + 479 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Generally, our method is invariant to the underlying energy term, and hence can be replaced by any other formulation like the ones mentioned before. This relationship is visualised in the taxonomy in Fig. 2: Our method solves a global, continuous optimisation problem instead of greedy and local steps formulated as dynamic programming, and can hence function as a more flexible backbone for approaches that derive from seam carving. This formulation allows more complex solutions that work better in many scenarios (see Fig. 5), while the space of all possible outputs contains the solutions of seam carving.", + "bbox": [ + 212, + 493, + 787, + 613 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3 Deformation Fields for Retargeting Visual Data", + "text_level": 1, + "bbox": [ + 214, + 637, + 715, + 655 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Editing strategies for visual data require avoiding changes to meaningful content, that create visible artefacts. An approach that focuses on retargeting the input to match a target size can be guided by the following set of rules:", + "bbox": [ + 212, + 669, + 785, + 715 + ], + "page_idx": 3 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Content Aware Objective: Prevent the accidental creation of new or the removal of existing regions with high information content (control over where deformation happens).", + "- Sanity Objective: Avoid introducing deformations that yield implausible results (control over how deformation happens)." + ], + "bbox": [ + 223, + 724, + 782, + 797 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In the following sections, we are only concerned with the retargeting to a smaller size (shrinking) for ease of explanation. The concept can naturally be", + "bbox": [ + 212, + 809, + 785, + 839 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "T. Elsner, J. Berger et al.", + "bbox": [ + 271, + 114, + 442, + 127 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/a947aca5d6b8bf3d860a5b4ac1d458b5616dcd9782220e84de626ece55cded26.jpg", + "image_caption": [ + "Fig. 3: Our proposed pipeline for retargeting visual data: For a given input (left), we train two simple networks that learn the energy and cumulative energy along the deformation axis of the input (centre-left). We initialise a network that stretches samples to the desired position (centre-right), then optimise this deformation to distribute the distortion to low information content regions (right). Balloon image from [42]." + ], + "image_footnote": [], + "bbox": [ + 215, + 141, + 787, + 282 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "reformulated to expand the visual data instead (expansion, see Sec. C (Appendix)) or to include editing (see Sec. 3.5).", + "bbox": [ + 212, + 395, + 782, + 425 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Seam carving [2] and its derivatives produce a discrete solution to this by performing edits through manipulating (removing or repeating) one seam of pixels at a time, which is a single-pixel-wide path passing through the image. Through dynamic programming, seam carving chooses this path based on the lowest energy of the pixels, with high energy indicating information-rich areas. In a simple case, this energy is the colour gradient. This yields a discrete solution to the retargeting problem that is both content aware and plausible.", + "bbox": [ + 212, + 426, + 782, + 531 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To obtain a continuous solution, instead of extracting discrete seams of the same width, we use continuous folds of varying width using a learned continuous deformation function. This formulation optimises a solution that is a superset of all existing seam carving solutions (see Fig. 5). We always use the same core formulation, regardless of the underlying domain and application, e.g. moving an object in an image or retargeting a 3D scene.", + "bbox": [ + 212, + 532, + 787, + 625 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1 General Formulation", + "text_level": 1, + "bbox": [ + 215, + 652, + 433, + 667 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our approach always optimises a deformation function that maps from points in the retargeted output space (e.g. an image at half the size) $I: P \\to \\mathbb{R}^n$ to points in the original data $I': P' \\to \\mathbb{R}^n$ (e.g. to $\\mathbb{R}^3$ for RGB colours) by adding an offset to their coordinate. The value of a point $p$ in the output data $I$ is hence defined as the value of $p$ deformed to a new point $p'$ looked up in $I'$ . To represent the deformation function $D$ , we use a neural deformation field, a simple MLP where $D: P \\to \\mathbb{R}$ . To keep our approach simpler, we only allow scalar deformations along a fixed deformation direction $v$ , e.g. along the $x$ axis (width) of an image:", + "bbox": [ + 212, + 685, + 787, + 806 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nI (p) = I ^ {\\prime} (p + D (p) \\cdot v) \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 416, + 824, + 785, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "Retargeting Visual Data with Deformation Fields", + "bbox": [ + 401, + 114, + 732, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/a4cf99d9231932bf11879bae955364a8cdaa82a896041aa1996818aa746a678d.jpg", + "image_caption": [ + "Fig. 4: Our pipeline applied to neural radiance fields: We first obtain a point cloud, then estimate the energy values. We then learn a continuous energy function and a cumulative energy function that we use to optimise our deformation network. The resulting deformation field is visualised on the bottom right." + ], + "image_footnote": [], + "bbox": [ + 217, + 143, + 787, + 321 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/f19b37273e1b4f64146e5637c753a39b3d46c5daa026dc1a703e802c202106ca.jpg", + "image_caption": [ + "Fig. 5: Example of applying retargeting to $75\\%$ width. Seam carving fails to use seams that are non-continuous or with a too steep angle, not properly using the empty space in the shelf to close the gaps before shrinking the bookshelves content." + ], + "image_footnote": [], + "bbox": [ + 233, + 405, + 764, + 566 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Hence, to get the content for a point $p$ in the retargeted data, we offset the coordinate and look up the content at the offset coordinate $p'$ in the original data. To fulfil the objectives stated in the beginning of Sec. 3, we introduce two sets of losses, content aware and sanity objectives:", + "bbox": [ + 212, + 655, + 787, + 717 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Content Aware Objectives apply regularisation to guide the deformation process when adjusting to the data content. In particular, we penalise deformation in regions with high information content to discourage noticeable changes. We define the energy function $E: P' \\to \\mathbb{R}$ to be a measure for content information in the original visual data, expresses as, e.g., a colour gradient:", + "bbox": [ + 212, + 729, + 787, + 808 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nE \\left(p ^ {\\prime}\\right) = \\left| \\left| \\nabla I \\left(p ^ {\\prime}\\right) \\right| \\right| _ {2} \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 431, + 823, + 785, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "T. Elsner, J. Berger et al.", + "bbox": [ + 271, + 114, + 442, + 127 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/d7155f8351a4bc1be3916739063663ed51d396663c7fff263404983649e19838.jpg", + "image_caption": [ + "Fig. 6: Applying our retargeting to 3D meshes, moving furniture closer together to retarget the scene to a smaller size while limiting distortion to objects." + ], + "image_footnote": [], + "bbox": [ + 277, + 143, + 467, + 273 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/479c7c964d87091d9e2ff5082052f7a9a85379dcf805c46dab759962506dcf37.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 470, + 146, + 593, + 271 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e0527205f9d5202a85a40f2cb96af7191ab6e52fc07b84331541245d49debd39.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 598, + 145, + 722, + 272 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The overall content deformation loss is the product of deformation magnitude at $p$ and energy at the original $p'$ over all $p$ in the deformed data:", + "bbox": [ + 214, + 342, + 782, + 372 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\nL _ {C} = \\int_ {p \\in P} [ E (p + v \\cdot D (p)) \\cdot | | \\nabla D (p) | | _ {1} ] d p \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 346, + 383, + 784, + 424 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Thus, in Eq. (3), we effectively penalise visible deformations defined as the product of information content at points, namely $E(p + v \\cdot D(p))$ , and the deformation itself via $||\\nabla D(p)||_1$ . Using the $L^1$ -norm for the distortion gradient promotes piecewise constant, possibly sparse, deformations. We can then split up the change in deformation into change in deformation direction $v$ and change in direction $v^\\perp$ orthogonal to $v$ :", + "bbox": [ + 214, + 436, + 782, + 527 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\nL _ {C} = \\int_ {p \\in P} \\left[ E (p + v \\cdot D (p)) \\cdot \\left(\\left| \\frac {\\partial}{\\partial v} D (p) \\right| + \\left| \\frac {\\partial}{\\partial v ^ {\\perp}} D (p) \\right|\\right) \\right] d p \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 279, + 537, + 784, + 580 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "This equation is then further split into two components, $L_{e}$ and $L_{s}$ . The loss $L_{e}$ penalises stretching or compression and $L_{s}$ punishes shearing, which we define to be deformation change in a direction that is not the deformation direction. For a now discrete sample set $P$ from our domain, these terms are expressed as:", + "bbox": [ + 214, + 590, + 784, + 652 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\nL _ {e} = \\sum_ {p \\in P} \\left[ E (p + v \\cdot D (p)) \\cdot \\left| \\frac {\\partial}{\\partial v} D (p) \\right| \\right] \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 352, + 662, + 784, + 710 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\nL _ {s} = \\sum_ {p \\in P} \\left[ E (p + v \\cdot D (p)) \\cdot \\left| \\frac {\\partial}{\\partial v ^ {\\perp}} D (p) \\right| \\right]\n$$\n", + "text_format": "latex", + "bbox": [ + 357, + 704, + 643, + 742 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Essentially, $L_{e}$ directs deformations to areas of low information content, while $L_{s}$ discourages the resulting shearing to affect areas of high information content. Note that while $L_{s}$ would also penalise non-straight deformation seams, we tune our losses such that this only impacts the output much if an area is sheared. An example of this is shown in Fig. 3: The deformations around the balloons introduce shearing only at the seam, in an area of low information content.", + "bbox": [ + 214, + 750, + 782, + 839 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "Retargeting Visual Data with Deformation Fields", + "bbox": [ + 401, + 114, + 732, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In case of image deformation, neighbouring pixels with low energy in the input may translate to pixels widely apart in the output, potentially bypassing a substantial amount of important content. Jumps in the deformation field $D$ indicate this effect. To prevent this, we penalise the amount of energy $\\hat{E}$ between two points $p', q' \\in P'$ in the original domain, rather than the energy at individual positions, as in Eq. (2):", + "bbox": [ + 212, + 146, + 787, + 237 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {E} \\left(p ^ {\\prime}, q ^ {\\prime}\\right) = \\int_ {r ^ {\\prime} \\in \\left[ p ^ {\\prime}, q ^ {\\prime} \\right]} E \\left(r ^ {\\prime}\\right) d r ^ {\\prime} \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 401, + 246, + 785, + 289 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Using $\\hat{E}$ , we then redefine $L_{e}$ from Eq. (5) to consider energy and difference in deformation magnitude between $p$ and a slightly offset $p_{\\varepsilon} \\coloneqq p + \\varepsilon$ in deformation direction:", + "bbox": [ + 212, + 297, + 785, + 344 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\nL _ {e} = \\sum_ {p \\in P} \\hat {E} (p + v D (p), p _ {\\varepsilon} + v D (p _ {\\varepsilon})) \\cdot \\frac {| D (p) - D (p _ {\\varepsilon}) |}{\\varepsilon} \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 303, + 356, + 785, + 393 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This formulation guides the deformation to avoid introducing deformation to high information content regions or fold over them, while $L_{s}$ discourages non-local deformations that would apply shearing in high information areas. For example, shearing in blue sky is acceptable, shearing on a brick wall is not.", + "bbox": [ + 214, + 398, + 785, + 459 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Sanity objectives prevent undesired outputs by introducing two additional losses. Assume we shrink the input by a factor $\\alpha \\in [0,1]$ :", + "bbox": [ + 214, + 473, + 785, + 505 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "1. The boundaries of the affected region should not change, e.g. an image should not be simply cropped. Based on $\\alpha$ , the deformation field should map the start/end of the deformation axis in the output to the start/end of the input. For $P_0$ , $P_1$ containing all points at the start/end of the deformation axis of the output, we express this requirement as follows:", + "bbox": [ + 222, + 511, + 787, + 587 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\nL _ {b} = \\int_ {p \\in P _ {0}} | D (p) | d p + \\int_ {p \\in P _ {1}} | D (p) - (1 - \\alpha) | d p \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 351, + 595, + 785, + 637 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "2. We additionally demand monotonicity in the deformation field, expressed as", + "bbox": [ + 222, + 643, + 785, + 659 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\nL _ {m} = \\int_ {p \\in P} \\max \\left(0, - \\frac {\\partial}{\\partial v} D (p)\\right) d p, \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 392, + 670, + 785, + 712 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "indicating that any change in the deformation direction should be positive. Therefore, deformation $D$ should be monotonic and only stagnate or increase exclusively along $v$ . This prevents the repetition of the same image portion, noticeable by jumps with inconsistent direction in the deformation field.", + "bbox": [ + 241, + 715, + 785, + 777 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We then obtain the total loss $L$ with hyperparameters $\\lambda_{*}$ as", + "bbox": [ + 215, + 799, + 651, + 815 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\nL = \\lambda_ {e} L _ {e} + \\lambda_ {s} L _ {s} + \\lambda_ {b} L _ {b} + \\lambda_ {m} L _ {m}. \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 380, + 825, + 785, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "T. Elsner, J. Berger et al.", + "bbox": [ + 271, + 114, + 442, + 128 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We initialise the deformation function with a uniform stretch to the target size.", + "bbox": [ + 215, + 146, + 782, + 161 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "In the following, we demonstrate the application of this general formulation to different domains, while adapting the energy formulation to the problem. All implementation and architecture details are given in Sec. A (Appendix) and the exact loss terms in Sec. B (Appendix).", + "bbox": [ + 212, + 162, + 782, + 222 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "3.2 Optimising Deformation Fields for Images", + "text_level": 1, + "bbox": [ + 215, + 244, + 609, + 260 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Instead of using image data directly, we train a continuous image representation MLP $I': P' \\to \\mathbb{R}^3$ to be our input data, that maps pixel positions to colours. This ensures consistency with our continuous approaches.", + "bbox": [ + 214, + 271, + 782, + 316 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We train one MLP to learn the energy field $E$ , holding the colour gradient per point as in Eq. (2). Using the learned $E$ , we train another MLP to replicate $\\hat{E}$ as a cumulative energy field $\\Sigma$ . The difference in cumulative energy $\\Sigma$ between two points $p', q' \\in P'$ on the same axis $v$ is the absolute difference in information content between them. As this coincides with the definition of $\\hat{E}$ from Eq. (6), we approximate the measure of content relevance between two points as:", + "bbox": [ + 214, + 330, + 782, + 422 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {E} \\left(p ^ {\\prime}, q ^ {\\prime}\\right) \\approx | \\Sigma \\left(q ^ {\\prime}\\right) - \\Sigma \\left(p ^ {\\prime}\\right) | \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 406, + 434, + 784, + 450 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This cumulative gradient formulation therefore simplifies the interval over $E$ in Eq. (6) to a simple difference computation between two MLP queries: How much content is lost when the deformation skips a region. The energy field $E$ can be trained to reproduce any energy formulation other than a colour gradient. This exchangeable energy formulation ensures compatibility with other retargeting approaches as shown in Fig. 2. For all objectives, we use discrete approximations of the loss terms introduced in Sec. 3.1. The whole pipeline is visualised in Fig. 3, and additional results can be found in Sec. G (Appendix).", + "bbox": [ + 214, + 462, + 784, + 584 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "3.3 Optimising Deformation Fields for Neural Radiance Fields", + "text_level": 1, + "bbox": [ + 215, + 606, + 741, + 622 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We adapt our approach from Sec. 3.1 for 3D scenes given as trained NeRFs as follows (see Fig. 4):", + "bbox": [ + 214, + 632, + 782, + 662 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- A trained NeRF provides colour and depth images, from which we extract surface samples. Each point is assigned an energy value defined by change in colour and depth at this point.", + "- Based on the surface samples and their energy values, we train a continuous energy field $E$ and a continuous cumulative energy field $\\Sigma$ , to easily access the amount of content between points.", + "- We use an inverse deformation network $U$ to find points in the target space that correspond to surface points in the source space, keeping computing costs low by restricting our optimisation to surface points.", + "- We apply our regularisation terms on a mix of surface and random points to optimise our loss function." + ], + "bbox": [ + 223, + 672, + 784, + 838 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "Retargeting Visual Data with Deformation Fields", + "bbox": [ + 400, + 114, + 730, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "1. Extracting a Sparse Set of Representatives and Energy We extract a surface point cloud from the RGB and depth images rendered by NeRF at the training camera positions. Only these points are relevant for scene deformation. We then compute the energy value for each point corresponding to a pixel as a sum of colour change in the RGB image and depth change in the depth image. To reduce the impact of highly noisy energy values, we apply smoothing by selecting the minimum energy for each point within a neighbourhood radius of 5 points.", + "bbox": [ + 212, + 146, + 787, + 253 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "2. Defining an Energy Function To obtain a continuous energy field, we train an auxiliary energy network $E$ . The network maps all 3D points to their energy values, and clamps the energy to zero for points further than twice the average neighbour distance. The resulting energy field has zero energy in empty space and inside objects, and non-zero energy values for points near the surface. Similar to our approach with images, we learn a cumulative energy network $\\Sigma$ . The network is trained to accumulate the energy values $E(p)$ from uniformly sampled points $p$ on a ray in deformation direction. We compute the energy for a segment between two points $p', q' \\in \\mathbb{R}^3$ along $v$ using their difference in cumulative energy $\\Sigma(q')$ and $\\Sigma(p')$ (cf. Eq. (11)).", + "bbox": [ + 212, + 266, + 787, + 417 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "3. Inverting the Deformation We define an inverse deformation function to determine those points that lie on the surface after deformation. In contrast to images, we only perform optimisation on a sparse set of points, namely surface points. As we optimise a deformation from the target space to the source space, we need the position of surface points in both spaces. By training an MLP $U$ to reverse the deformation $D$ , i.e. minimising $(U(D(x)) = x)^2$ , we can effortlessly extract surface points, and hence their energy values, in the target space. This relation is also visualised in Fig. 4. We always update $U$ once after training one iteration of $D$ .", + "bbox": [ + 212, + 431, + 787, + 566 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "As we use a discrete set of points, we apply the losses from Sec. 3.1 to Eq. (10) for a discrete interval. In particular, we average the loss over a random subset of sample points. Example images from deformed scenes can be found in Fig. 8, while a larger number of results can be found in Sec. G (Appendix).", + "bbox": [ + 212, + 582, + 787, + 643 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "3.4 Optimising Deformation Fields for Meshes", + "text_level": 1, + "bbox": [ + 212, + 664, + 612, + 679 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "To demonstrate the versatility of our approach, we apply it to meshes as well. We adapt the energy term and use surface samples and vertex positions for optimisation. For freestanding objects, we use curvature as the energy measure. For the room example in Fig. 6, the energy is set to 0 for the floor and to 1 for every other surface point. We otherwise follow the same principle as in Sec. 3.3.", + "bbox": [ + 212, + 688, + 787, + 763 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "3.5 Other Applications", + "text_level": 1, + "bbox": [ + 212, + 785, + 419, + 800 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "While retargeting is our main contribution, our framework opens up multiple different editing opportunities that even extend the capabilities of seam carving [2].", + "bbox": [ + 212, + 809, + 785, + 840 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "T. Elsner, J. Berger et al.", + "bbox": [ + 271, + 114, + 442, + 128 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/41087d1b823b0ef952f4422c8b1fdb6095b7086826307699a89f304cef065f23.jpg", + "image_caption": [ + "Iinui" + ], + "image_footnote": [], + "bbox": [ + 253, + 161, + 416, + 239 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/c877131d3cd9a0e3bad2108f760006a6cc9b37f2dbd32f4b5de504f22166e232.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 436, + 161, + 501, + 239 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/483d4940066c036181d0d5623a4f2622255bc9ae6707c2ae2a3fc6cf017b5f4d.jpg", + "image_caption": [ + "ed", + "Fig. 7: We demonstrate different editing operations by adding and adapting the loss formulation. Left: Removing, right: Moving an object. Images from [42]." + ], + "image_footnote": [], + "bbox": [ + 253, + 241, + 416, + 320 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/5b2c071ea3cd1a71b60adbf886f91528f6a8888fda1662458f4ee72e4f392aa5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 436, + 241, + 501, + 320 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/e9c01c195ebf0ff3921e10ccd9a4bcd79c855141d59ca30a154aa2088df2eef1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 532, + 152, + 784, + 266 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/5b01bb21a0c5e53a64d87bb5e21ff28c8181e8ab46d966e7ee0e9315ce5281fd.jpg", + "image_caption": [ + "Fig.8: Applying our retargeting to NeRFs. Top row: Input, stretched, targeted with our method. Bottom row: Input, folds (white) on the input suggested by our approach, targeted with our method." + ], + "image_footnote": [], + "bbox": [ + 527, + 267, + 785, + 316 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/2e1d792e0bc9fdfa6765039930e5928af706c34079ee4ce241a4fce66f0fa475.jpg", + "image_caption": [ + "Fig. 9: Selected results showing scenes where our approach benefits from a higher degree of flexibility. All images are arranged as input, seam carving, and our output. Images from [42]." + ], + "image_footnote": [], + "bbox": [ + 215, + 438, + 330, + 500 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/f94ee2cb297addced415b0e223edb326bea5527110da4a42766497f224294b12.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 331, + 438, + 442, + 500 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/f9e9f691756184c008def46be8d17f37589b54b356ee47f8ba4ba9288f7f3ba5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 442, + 438, + 500, + 500 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/a26a02d0dba06544237ee3e4b6a0acb74db8d799728e3591759f34e48455f53a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 501, + 438, + 558, + 500 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/ab35c6c485ee0a5ce02ca2a7e913fbf90633bdd8c76f21feb900f6d3247472b0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 560, + 438, + 669, + 500 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/84d4b09054b4f4798fce5a91b9366cce15e3afad001a2510ea62e0114886d3ca.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 669, + 438, + 727, + 500 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/fc2c2b9165fa633d16ee2bf35eba67da8a51a736e4e008598402869af648fb15.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 728, + 438, + 784, + 500 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "The loss formulation is slightly adapted to implement the following exemplary operations, which are applicable to any domain and can serve as the basis for additional processing methods or user-driven applications. Further details can be found in Sec. D (Appendix). Examples can be found in Fig. 7. We want to remove the monotonicity loss $L_{m}$ and use equal input and output sizes for both the removal and the movement of an object.", + "bbox": [ + 212, + 598, + 784, + 688 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Removal To remove a certain region, we employ an additional loss term that penalises any deformation target leading to that region. Points that were originally assigned to the target region will be distorted and assigned to neighbouring regions, resulting in the removal of the region in the output.", + "bbox": [ + 212, + 689, + 784, + 748 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Moving an Object Although classic seam carving cannot move an object easily, as seams are assumed to cross the entire image, our formulation does not have this restriction. To move an object, we also enforce that the deformation towards the target feature only occurs from the desired new placement in the output. Any occurrence of that feature elsewhere is punished to avoid its repetition.", + "bbox": [ + 212, + 750, + 785, + 839 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "Retargeting Visual Data with Deformation Fields", + "bbox": [ + 401, + 114, + 732, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 116, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/7c248f1bfebfd0bb1a9fb3e5685ac6e338e99437b9999fd7a29151cd484d86c7.jpg", + "table_caption": [ + "Table 1: Comparison of FID scores on the RetargetMe [42] dataset, for retargeting x (top) and y (bottom) axis to a smaller (left) and bigger (right) size. In mean, ours outperforms seam carving with 46.67 to 52.57 for x and 53.57 to 65.48 for y." + ], + "table_footnote": [], + "table_body": "
x FID(↓)50%60%70%80%90%110%120%130%140%150%
Seam carving85.7566.4151.6735.2023.8020.8630.0137.3843.4546.37
Ours79.3958.0643.6930.5321.7111.1719.4229.1937.8749.74
y FID(↓)50%60%70%80%90%110%120%130%140%150%
Seam carving116.7084.3360.5940.8624.9223.0433.5840.8248.4952.59
Ours89.4467.7749.6235.5925.4213.5521.0227.5538.2046.40
", + "bbox": [ + 225, + 199, + 777, + 292 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/5845ebbaf4722a5914818415d2102639945e06180ac125473c61013b2905b322.jpg", + "table_caption": [ + "Table 2: User study comparing our approach to seam carving $(SC)$ for images. 19 participants were questioned for preferences of random examples from the RetargetMe dataset [42], casting 1200 votes." + ], + "table_footnote": [], + "table_body": "
WidthOursSCDraw
50 %48.48%20.87%30.67%
150 %33.8%36.27%29.92%
HeightOursSCDraw
50 %49.5%24.25%26.25%
150 %45.84%20.0%34.15%
Total44.57%25.1%30.32%
", + "bbox": [ + 250, + 387, + 478, + 496 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/e8604be8cf7942901512735e38ef12f1dcf5088957484b76f42deedc7e478f62.jpg", + "image_caption": [ + "Colour Complex Gradient Energy", + "Fig. 10: Example of using our method with the more complex energy term of [48] instead of gradients, improving the results on an example." + ], + "image_footnote": [], + "bbox": [ + 511, + 333, + 640, + 411 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "4 Evaluation", + "text_level": 1, + "bbox": [ + 215, + 523, + 356, + 540 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Retargeting not only lacks comparable methods for other visual data such as meshes, but its quality is also difficult to measure. In 2D, we compare our results to seam carving [2] on images using the RetargetMe [42] benchmark. In 3D, we compare our results on NeRFs with video-seam-carved recordings of the same scene, because video seam carving [43] is the closest method to 3D retargeting. We always produce a short, 3 second clip at 30 frames per second. To ensure a fair comparison with the method, we only tilt the camera down instead of rotating it around the object. Even for benevolent camera movements, it produces visible artefacts, as seen in Sec. F (Appendix). Exemplary, we show that our approach is compatible with approaches that propose different saliency maps as visualised in Fig. 2. For that purpose, we compare our method to other more recent image retargeting approaches visually (Fig. 10) and in a user study (Tab. 4). Additional results can be found in Sec. G (Appendix), including an explicit failure case.", + "bbox": [ + 212, + 559, + 787, + 756 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "4.1 Comparisons", + "text_level": 1, + "bbox": [ + 215, + 781, + 369, + 796 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Images In accordance with recent publications in the generative model and image editing domain [15, 23, 38], we extensively compare our results with the", + "bbox": [ + 212, + 809, + 785, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "T. Elsner, J. Berger et al.", + "bbox": [ + 271, + 114, + 442, + 128 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/3efd14c17b9d07df792725253ac3f278ac7d665fcfed22e24fd824893947c5ec.jpg", + "table_caption": [ + "Table 3: User study comparing retargeting a scene, then recording it (ours), to recording it and then retargetet it (video seam carving [43], VSC), using our provided dataset. 21 participants were asked to indicated which solution they prefer, casting 252 votes in total." + ], + "table_footnote": [], + "table_body": "
SceneOursVSCDrawOursVSCDrawOursVSCDrawOursVSCDraw
Beach1731201019111911
Farm2100210016322100
Harb.2100210021002100
", + "bbox": [ + 241, + 212, + 759, + 286 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "original seam carving results using FID [20] in Tab. 1 for images. We also evaluate our approach in a user study. The participants were instructed to classify their preference for a retargeted version of a randomly selected image based on a randomly selected deformation axis, always choosing between our and the seam carved [2] version. We used the 80 images of the RetargetMe [42] for 2D, producing 360 variations per approach in Tab. 2. This shows how our approach can be a more powerful backbone for seam carving.", + "bbox": [ + 212, + 330, + 787, + 435 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/78c535b2c37002c628eed3542b4c6f150b7734e162a2e80093f7f67e92e8741f.jpg", + "table_caption": [ + "Table 4: User study measuring preference, pairwise comparing to our approach using salience maps of [48]." + ], + "table_footnote": [], + "table_body": "
ApproachOursTheirsDraw
CarvingNet [46]45.16%41.93%12.91%
RL [22]46.85%32.43%20.72%
Cycle-IR [49]40.3%34.46%25.24%
", + "bbox": [ + 223, + 537, + 524, + 602 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "We further compare against other state-of-the-art retargeting approaches in a user study in Tab. 4: We show that our method can compete both against a seam carving variation with better saliency from deep learning [46] while also beating state-of-the-art approaches that retarget the image directly [22,49]. Note that we only compare to readily available images from other approach, as", + "bbox": [ + 540, + 450, + 787, + 632 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "we often could not reproduce them or run the code. For this comparison, we use the saliency method FullGrad [48] that aggregates layer-wise bias-gradients from across ResNet [16]. This aggregated gradient map acts as our saliency map, as it measures how useful the feature is to the neural network, hence we use it as a measure for how semantically important a feature is. This also shows that our approach can work with different saliency maps. Details and examples can be found in Sec. E (Appendix). Additionally, we provide a visual ablation that shows the importance of each loss component in Fig. 11.", + "bbox": [ + 212, + 632, + 787, + 753 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "3D Scenes When assessing the quality of retargeting 3D scenes provided as NeRFs, we compared our results with what we consider to be the most similar approach, video seam carving [43]. In our provided GTA V NeRF retargeting dataset, we resize each axis by $50\\%$ and $150\\%$ . The results can be found in Tab. 3.", + "bbox": [ + 212, + 779, + 787, + 839 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "Retargeting Visual Data with Deformation Fields", + "bbox": [ + 401, + 114, + 730, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/abfe2e99a4bc32fdc43ad0b6544da1d1f2e6d92af5130c2241f2a173e48df8c2.jpg", + "image_caption": [ + "Fig. 11: Visual ablation for retargeting to $80\\%$ width on an image [12], from left to right: Wobbly contours from shearing (no $L_{s}$ ), uniform stretch on the whole image (no $L_{e}$ ), missing features e.g. the person (no $L_{m}$ ), content outside the image (no $L_{b}$ )." + ], + "image_footnote": [], + "bbox": [ + 228, + 143, + 769, + 193 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "We provide additional results and frames of the resulting sequences from our approach for 3D scenes compared to video seam carving in Sec. F (Appendix).", + "bbox": [ + 214, + 277, + 784, + 309 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "4.2 Discussion and Limitations", + "text_level": 1, + "bbox": [ + 214, + 328, + 483, + 342 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "With our deformation fields as backbone, but the same energy term, our approach outperforms seam carving on both a qualitative and quantitative level. Our flexibility in the energy formulation allows for more possibilities, and the replacement of the gradient-based energy formulation with a better saliency map improves our approach even more. This shows that our method is a flexible backbone that can be used with various successors to seam carving, which provide more advanced saliency maps. Our continuous formulation allows for retargeting in various domains, including 3D scenes, and highlights the advantages of a geometry-aware approach. Applying our approach with only colour gradients can produce flawed outputs, just as those of seam carving, see Fig. 15 (Appendix). As we optimise a neural field for each individual input, we trade computation time for not needing a dataset to train and for a global optimisation. Our deformation optimisation takes approximately one minute on consumer hardware for good results versus about half a second for seam carving. Our strength, however, lies in 3D. Video seam carving on a 600 by 400 pixel video can take several hours, whereas our approach produces the final retargeted 3D scene in under 15 minutes. For Future Work, investigating invertible neural networks to implicitly learn an inverse deformation for the 3D case [1] seems promising, as would a deep learned approach to evaluating a deformation for optimisation and more research into good measures for retargeting quality.", + "bbox": [ + 212, + 351, + 787, + 654 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 674, + 359, + 690 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "We present an approach that generalises the principle underlying seam carving to create image variants by manipulating low content areas. By training neural deformation fields that avoid applying distortion to important parts of an input, our approach is applicable to different forms of visual data retargeting. Our formulation is largely domain invariant and only needs minor alterations to work on different types of visual data, as demonstrated for neural radiance fields and meshes for the first time. We demonstrate compatibility to other energy formulations of the seam carving family, being a more general way of optimising image retargeting. Our approach is the only one that natively works on 3D scenes.", + "bbox": [ + 212, + 704, + 787, + 840 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "T. Elsner, J. Berger et al.", + "bbox": [ + 271, + 114, + 442, + 127 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 143, + 321, + 159 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Ardizzone, L., Kruse, J., Wirkert, S.J., Rahner, D., Pellegrini, E.W., Klessen, R.S., Maier-Hein, L., Rother, C., Kothe, U.: Analyzing inverse problems with invertible neural networks. CoRR abs/1808.04730 (2018), http://arxiv.org/abs/1808.04730", + "2. Avidan, S., Shamir, A.: Seam carving for content-aware image resizing. ACM Trans. Graph. 26(3), 10 (2007)", + "3. Basha, T.D., Moses, Y., Avidan, S.: Stereo seam carving a geometrically consistent approach. IEEE Trans. Pattern Anal. Mach. Intell. 35(10), 2513-2525 (2013)", + "4. Berkiten, S., Halber, M., Solomon, J., Ma, C., Li, H., Rusinkiewicz, S.: Learning detail transfer based on geometric features. Comput. Graph. Forum 36(2), 361-373 (2017)", + "5. Cai, H., Feng, W., Feng, X., Wang, Y., Zhang, J.: Neural surface reconstruction of dynamic scenes with monocular rgb-d camera. In: Neural Information Processing Systems (NeurIPS) (2022)", + "6. Chen, Zhiqin and Zhang, Hao: Learning implicit fields for generative shape modeling. In: CVPR. pp. 5939-5948 (2019)", + "7. Dong, W., Bao, G., Zhang, X., Paul, J.: Fast multi-operator image resizing and evaluation. J. Comput. Sci. Technol. 27(1), 121-134 (2012)", + "8. Dong, W., Zhou, N., Lee, T., Wu, F., Kong, Y., Zhang, X.: Summarization-based image resizing by intelligent object carving. IEEE Trans. Vis. Comput. Graph. 20(1), 1 (2014). https://doi.org/10.1109/TVCG.2013.103, https://doi.org/10.1109/TVCG.2013.103", + "9. Dong, W., Zhou, N., Paul, J., Zhang, X.: Optimized image resizing using seam carving and scaling. ACM Trans. Graph. 28(5), 125 (2009)", + "0. Efros, A.A., Freeman, W.T.: Image quilting for texture synthesis and transfer. In: Pocock, L. (ed.) Proceedings of the 28th Annual Conference on Computer Graphics and Interactive Techniques, SIGGRAPH. pp. 341-346. ACM (2001)", + "1. Efros, A.A., Leung, T.K.: Texture synthesis by non-parametric sampling. In: ICCV. pp. 1033-1038 (1999)", + "2. at English Wikipedia, N.: Broadway tower (2007), https://commons.wikimedia.org/wiki/File:Broadway_tower.jpg", + "3. Garbin, S.J., Kowalski, M., Estellers, V., Szymanowicz, S., Rezaeifar, S., Shen, J., Johnson, M., Valentin, J.: Voltemorph: Realtime, controllable and generalisable animation of volumetric representations. arXiv:2208.00949 (2022)", + "4. Gatys, L.A., Ecker, A.S., Bethge, M.: Texture synthesis using convolutional neural networks. In: Neural Information Processing Systems. pp. 262-270 (2015)", + "5. Gu, J., Zhai, S., Zhang, Y., Susskind, J., Jaitly, N.: Matryoshka diffusion models. arXiv preprint arXiv:2310.15111 (2023)", + "6. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. CoRR abs/1512.03385 (2015), http://arxiv.org/abs/1512.03385", + "7. Henzler, P., Deschaintre, V., Mitra, N.J., Ritschel, T.: Generative modelling of BRDF textures from flash images. ACM Trans. Graph. 40(6), 284:1-284:13 (2021)", + "8. Henzler, P., Mitra, N.J., Ritschel, T.: Learning a neural 3d texture space from 2d exemplars. In: The IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2019)", + "9. Hertz, A., Hanocka, R., Giryes, R., Cohen-Or, D.: Deep geometric texture synthesis. ACM Trans. Graph. 39(4), 108:1-108:11 (2020)" + ], + "bbox": [ + 225, + 178, + 785, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "Retargeting Visual Data with Deformation Fields", + "bbox": [ + 401, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "20. Heusel, M., Ramsauer, H., Unterthiner, T., Nessler, B., Hochreiter, S.: Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems 30 (2017)", + "21. Huang, Y., Cao, Y., Lai, Y., Shan, Y., Gao, L.: Nerf-texture: Texture synthesis with neural radiance fields. In: ACM SIGGRAPH 2023 Conference Proceedings. pp. 43:1-43:10 (2023)", + "22. Kajiura, N., Kosugi, S., Wang, X., Yamasaki, T.: Self-play reinforcement learning for fast image retargeting. CoRR abs/2010.00909 (2020), https://arxiv.org/abs/2010.00909", + "23. Kawar, B., Zada, S., Lang, O., Tov, O., Chang, H., Dekel, T., Mosseri, I., Irani, M.: Imagic: Text-based real image editing with diffusion models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6007-6017 (2023)", + "24. Kopf, J., Fu, C.W., Cohen-Or, D., Deussen, O., Lischinski, D., Wong, T.T.: Solid texture synthesis from 2d exemplars. ACM Transactions on Graphics 26(3), 2:1-2:9 (2007)", + "25. Lai, Y., Hu, S., Gu, X., Martin, R.R.: Geometric texture synthesis and transfer via geometry images. In: Proceedings of the Tenth ACM Symposium on Solid and Physical Modeling. pp. 15-26. ACM (2005)", + "26. Liu, F., Gleicher, M.: Automatic image retargeting with fisheye-view warping. In: Proceedings of the 18th Annual ACM Symposium on User Interface Software and Technology. pp. 153-162. ACM (2005)", + "27. Liu, F., Gleicher, M.: Video retargeting: automating pan and scan. In: Nahrstedt, K., Turk, M.A., Rui, Y., Klas, W., Mayer-Patel, K. (eds.) Proceedings of the 14th ACM International Conference on Multimedia. pp. 241-250. ACM (2006)", + "28. Luiten, J., Kopanas, G., Leibe, B., Ramanan, D.: Dynamic 3d gaussians: Tracking by persistent dynamic view synthesis. In: 3DV (2024)", + "29. Mescheder, L., Oechsle, M., Niemeyer, M., Nowozin, S., Geiger, A.: Occupancy networks: Learning 3D reconstruction in function space. In: CVPR. pp. 4460-4470 (2019)", + "30. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. In: ECCV (2020)", + "31. Moreira, T.P., Santana, M.C.S., Passos, L.A., Papa, J.P., da Costa, K.A.P.: An end-to-end approach for seam carving detection using deep neural networks. In: Pattern Recognition and Image Analysis. Lecture Notes in Computer Science, vol. 13256, pp. 447-457 (2022)", + "32. Nam, S., Ahn, W., Yu, I., Kwon, M., Son, M., Lee, H.: Deep convolutional neural network for identifying seam-carving forgery. IEEE Trans. Circuits Syst. Video Technol. 31(8), 3308-3326 (2021)", + "33. Nataraj, L., Gudavalli, C., Mohammed, T.M., Chandrasekaran, S., Manjunath, B.S.: Seam carving detection and localization using two-stage deep neural networks. CoRR abs/2109.01764 (2021)", + "34. North, R.: Grand theft auto v. Steam (2015), https://www.rockstargames.com/V/", + "35. Park, J.J., Florence, P., Straub, J., Newcombe, R., Lovegrove, S.: DeepSDF: Learning continuous signed distance functions for shape representation. In: CVPR. pp. 165-174 (2019)", + "36. Park, K., Sinha, U., Barron, J.T., Bouaziz, S., Goldman, D.B., Seitz, S.M., MartinBrualla, R.: Nerfies: Deformable neural radiance fields. In: ICCV. pp. 5845-5854 (2021)" + ], + "bbox": [ + 215, + 146, + 784, + 840 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "T. Elsner, J. Berger et al.", + "bbox": [ + 271, + 114, + 442, + 127 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "37. Park, K., Sinha, U., Hedman, P., Barron, J.T., Bouaziz, S., Goldman, D.B., Martin-Brualla, R., Seitz, S.M.: Hypernerf: a higher-dimensional representation for topologically varying neural radiance fields. ACM Trans. Graph. 40(6), 238:1-238:12 (2021)", + "38. Peebles, W., Xie, S.: Scalable diffusion models with transformers. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 4195-4205 (2023)", + "39. Peng, Y., Yan, Y., Liu, S., Cheng, Y., Guan, S., Pan, B., Zhai, G., Yang, X.: Cagenerf: Cage-based neural radiance fields for generalized 3d deformation and animation. In: Advances in Neural Information Processing Systems (2022)", + "40. Pumarola, A., Corona, E., Pons-Moll, G., Moreno-Noguer, F.: D-NeRF: Neural Radiance Fields for Dynamic Scenes. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2021)", + "41. Richter, S.R., AlHaija, H.A., Koltun, V.: Enhancing photorealism enhancement. IEEE Transactions on Pattern Analysis and Machine Intelligence 45(2), 1700-1715 (2022)", + "42. Rubinstein, M., Gutierrez, D., Sorkine, O., Shamir, A.: A comparative study of image retargeting. ACM Transactions on Graphics (Proc. SIGGRAPH ASIA) 29(6), 160:1-160:10 (2010)", + "43. Rubinstein, M., Shamir, A., Avidan, S.: Improved seam carving for video retargeting. ACM Trans. Graph. 27(3), 16 (2008)", + "44. Rubinstein, M., Shamir, A., Avidan, S.: Multi-operator media retargeting. ACM Trans. Graph. 28(3), 23 (2009). https://doi.org/10.1145/1531326.1531329, https://doi.org/10.1145/1531326.1531329", + "45. Shaham, T.R., Dekel, T., Michaeli, T.: Singan: Learning a generative model from a single natural image. In: ICCV. pp. 4569-4579 (2019)", + "46. Song, E., Lee, M., Lee, S.: Carvingnet: Content-guided seam carving using deep convolution neural network. IEEE Access 7, 284-292 (2019). https://doi.org/10.1109/ACCESS.2018.2885347, https://doi.org/10.1109/ACCESS.2018.2885347", + "47. Sorkine-Hornung, O., Alexa, M.: As-rigid-as-possible surface modeling. In: Symposium on Geometry Processing (2007)", + "48. Srinivas, S., Fleuret, F.: Full-gradient representation for neural network visualization. Advances in neural information processing systems 32 (2019)", + "49. Tan, W., Yan, B., Lin, C., Niu, X.: Cycle-ir: Deep cyclic image retargeting. IEEE Transactions on Multimedia 22(7), 1730–1743 (2019)", + "50. Wang, P., Liu, L., Liu, Y., Theobalt, C., Komura, T., Wang, W.: Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. In: Advances in Neural Information Processing Systems. pp. 27171-27183 (2021)", + "51. Wei, L., Levoy, M.: Fast texture synthesis using tree-structured vector quantization. In: Proceedings of the 27th Annual Conference on Computer Graphics and Interactive Techniques, SIGGRAPH. pp. 479-488. ACM (2000)", + "52. Wu, G., Yi, T., Fang, J., Xie, L., Zhang, X., Wei, W., Liu, W., Tian, Q., Xinggang, W.: 4d gaussian splatting for real-time dynamic scene rendering. arXiv preprint arXiv:2310.08528 (2023)", + "53. Wu, H., Wang, Y., Feng, K., Wong, T., Lee, T., Heng, P.: Resizing by symmetrysummarization. ACM Trans. Graph. 29(6), 159 (2010)", + "54. Wu, R., Liu, R., Vondrick, C., Zheng, C.: Sin3dm: Learning a diffusion model from a single 3d textured shape. CoRR abs/2305.15399 (2023)", + "55. Wu, R., Zheng, C.: Learning to generate 3d shapes from a single example. ACM Transactions on Graphics (TOG) 41(6) (2022)" + ], + "bbox": [ + 212, + 146, + 787, + 840 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "Retargeting Visual Data with Deformation Fields", + "bbox": [ + 401, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "56. Xian, W., Sangkloy, P., Agrawal, V., Raj, A., Lu, J., Fang, C., Yu, F., Hays, J.: Texturegan: Controlling deep image synthesis with texture patches. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (June 2018)", + "57. Xu, T., Harada, T.: Deforming radiance fields with cages. In: ECCV. pp. 159-175 (2022)", + "58. Yang, Z., Gao, X., Zhou, W., Jiao, S., Zhang, Y., Jin, X.: Deformable 3d gaussians for high-fidelity monocular dynamic scene reconstruction. arXiv preprint arXiv:2309.13101 (2023)", + "59. Yuan, Y.J., Sun, Y.T., Lai, Y.K., Ma, Y., Jia, R., Gao, L.: Nerf-editing: Geometry editing of neural radiance fields. In: CVPR. pp. 18332-18343 (2022)" + ], + "bbox": [ + 215, + 146, + 785, + 299 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "T. Elsner, J. Berger et al.", + "bbox": [ + 271, + 114, + 442, + 128 + ], + "page_idx": 17 + } +] \ No newline at end of file diff --git a/2024/Retargeting Visual Data with Deformation Fields/547a2a3d-69aa-4867-932e-8316bae28b65_model.json b/2024/Retargeting Visual Data with Deformation Fields/547a2a3d-69aa-4867-932e-8316bae28b65_model.json new file mode 100644 index 0000000000000000000000000000000000000000..15e507330c8f51f926e9e445f439baae594f4c96 --- /dev/null +++ b/2024/Retargeting Visual Data with Deformation Fields/547a2a3d-69aa-4867-932e-8316bae28b65_model.json @@ -0,0 +1,2689 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.219, + 0.141, + 0.784, + 0.163 + ], + "angle": 0, + "content": "Retargeting Visual Data with Deformation Fields" + }, + { + "type": "text", + "bbox": [ + 0.229, + 0.189, + 0.776, + 0.22 + ], + "angle": 0, + "content": "Tim Elsner1, Julia Berger1, Tong Wu2, Victor Czech1, Lin Gao2, and Leif Kobbelt1" + }, + { + "type": "text", + "bbox": [ + 0.23, + 0.231, + 0.772, + 0.289 + ], + "angle": 0, + "content": "\\(^{1}\\) RWTH Aachen University, Germany \n{elsner, czech, kobbelt}@cs.rwth-aachen.de / julia.berger@rwth-aachen.de \n\\(^{2}\\) University of Chinese Academy of Sciences, China \n{wutong19s, gaolin}@ict.ac.cn" + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.322, + 0.747, + 0.502 + ], + "angle": 0, + "content": "Abstract. Seam carving is an image editing method that enables content-aware resizing, including operations like removing objects. However, the seam-finding strategy based on dynamic programming or graph-cut limits its applications to broader visual data formats and degrees of freedom for editing. Our observation is that describing the editing and retargeting of images more generally by a deformation field yields a generalisation of content-aware deformations. We propose to learn a deformation with a neural network that keeps the output plausible while trying to deform it only in places with low information content. This technique applies to different kinds of visual data, including images, 3D scenes given as neural radiance fields, or even polygon meshes. Experiments conducted on different visual data show that our method achieves better content-aware retargeting compared to previous methods." + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.515, + 0.741, + 0.529 + ], + "angle": 0, + "content": "Keywords: Image Retargeting \\(\\cdot\\) Neural Fields \\(\\cdot\\) Neural Radiance Fields" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.553, + 0.377, + 0.569 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.584, + 0.788, + 0.719 + ], + "angle": 0, + "content": "Media retargeting is a way to edit images, videos, 3D objects, or even entire 3D scenes by a global deformation such that relevant content, details, and features are properly preserved. The overall goal is to change the aspect ratio of the bounding box of the data so that it fits an allocated space, e.g. fitting an image to a display with prescribed format. Retargeting is based on the idea of identifying regions with little detail and to accumulate the necessary distortion induced by the deformation in these regions. Small editing operations are possible e.g. through concentrating the distortion there, specifying a region containing a particular object to have this object removed." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.788, + 0.809 + ], + "angle": 0, + "content": "Most existing retargeting methods typically use some form of seam carving, i.e. the deletion of discrete pixels on a path through the image (seams) [2,7,9,44]. As previous methods perform a discrete removal or addition of pixels, they do not natively extend to e.g. continuous domains like Neural Radiance Fields [30] (NeRFs) and are often restricted to a class of solutions defined by the used algorithm." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Our approach formulates the problem via continuous deformation fields, mapping from the retargeted data back to the undistorted input. It natively extends" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.444, + 0.129 + ], + "angle": 0, + "content": "T. Elsner, J. Berger et al." + }, + { + "type": "image", + "bbox": [ + 0.226, + 0.152, + 0.516, + 0.341 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.362, + 0.533, + 0.418 + ], + "angle": 0, + "content": "Fig.1: Different objectives applied to different visual domains with our approach: We demonstrate retargeting images, NeRFs, and meshes. 'Surfer' image from [42]." + }, + { + "type": "image", + "bbox": [ + 0.584, + 0.15, + 0.808, + 0.318 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.582, + 0.328, + 0.788, + 0.41 + ], + "angle": 0, + "content": "Fig.2: The family of image resizing methods. Our approach formulates the problem in a more general way, but can utilise e.g. new energy formulations as well." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.447, + 0.785, + 0.493 + ], + "angle": 0, + "content": "from images on other domains, like NeRFs and polygon meshes (Fig. 1), and offers better solutions through a global and continuous instead of a discrete and greedy optimisation process (see Fig. 5). Our contributions are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.501, + 0.785, + 0.543 + ], + "angle": 0, + "content": "- We introduce the use of a neural deformation field to compress or stretch unimportant regions to achieve a smaller or larger output or to follow other editing objectives." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.545, + 0.785, + 0.601 + ], + "angle": 0, + "content": "- We abstract and generalise the underlying rules of creating plausible outputs with seam carving, then demonstrate the domain agnostic nature of our formulation by applying it to images, 3D meshes, and 3D scenes in the form of NeRFs." + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.606, + 0.785, + 0.661 + ], + "angle": 0, + "content": "- We show that the high flexibility of our approach from optimising a global deformation field produces better outputs than iteratively computing seams, while also allowing configurations that are not possible with previous methods." + }, + { + "type": "list", + "bbox": [ + 0.225, + 0.501, + 0.785, + 0.661 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.788, + 0.794 + ], + "angle": 0, + "content": "To achieve this, we regularise the neural deformation fields to follow general sanity guidelines, while attempting to minimise distortion in places with presumed high information content. We evaluate our results on both qualitative and quantitative levels. To showcase our approach on interesting scenes captured as NeRFs and inspired by the dataset of Richter et al. [41], we also provide a small synthetic NeRF retargeting dataset captured in the video game GTA V [34] by flying the camera over the virtual city, capturing the screen content and camera parameters." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.796, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Our method is a more reliable backbone for approaches following seam carving while outperforming current methods, requires only minor modifications for vastly different applications and domains, and is straightforward to implement." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.402, + 0.115, + 0.731, + 0.129 + ], + "angle": 0, + "content": "Retargeting Visual Data with Deformation Fields" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.147, + 0.784, + 0.162 + ], + "angle": 0, + "content": "We provide the code of our method demonstrated on images in the supplemental." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.201, + 0.386, + 0.217 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.234, + 0.785, + 0.294 + ], + "angle": 0, + "content": "As our approach is re-formulating seam carving, it produces results from a single example without any learned prior, and is designed to allow seam carving-like deformations for (learned) 3D representations. We hence discuss these three branches of related work." + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.311, + 0.786, + 0.552 + ], + "angle": 0, + "content": "Single Example Generation The goal of single example generation is to discover similar local patterns from a single input and synthesise novel samples. Texture synthesis on 2D images and 3D shapes has been widely studied in computer vision and computer graphics by non-neural methods [10, 11, 24, 51]. There are also methods focusing on synthesising geometric textures for 3D shapes [4, 19, 25], which transfer local geometry details from an example shape to other surfaces. Recently, neural-based methods have been proposed to generate 2D textures [14, 45, 56] and transfer them to 3D shapes [17, 18], resulting in improved results. With recent advances in implicit representation [6, 29, 35] and neural radiance fields [30], Wu et al. propose to learn implicit 3D shapes from a shape with [54] or without texture [55]. To enhance the realism of synthesised shapes, Huang et al. [21] reconstruct the 3D shape with a NeRF by optimising texture features on a reconstructed mesh. These features can be applied to arbitrary surfaces to render the target surface with synthesised textures. These methods often only work for a certain kind of data format, while our method proposes a domain agnostic approach to synthesise visual data from a single example." + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.568, + 0.786, + 0.84 + ], + "angle": 0, + "content": "Deformations for Learned 3D Representations Neural radiance fields (NeRF), as a newly proposed 3D shape representation, has been widely used in scene reconstruction and novel view synthesis. However, the original NeRF only works for static scenes, thus is unable to deal with deformation in dynamic scenes or generate motions for a static scene. To overcome this, the deformation field is introduced to NeRF. NeRF-Editing [59] is the first to explore deformation in NeRF. It first reconstructs the explicit mesh of a static scene with NeuS [50] and deforms the explicit mesh with the ARAP algorithm [47]. Then sample points in volume rendering are deformed along with the mesh via barycentric coordinate interpolation, resulting in the change of rendered images. Follow-up NeRF deformation methods [13,39,57] employ a similar pipeline but use different geometry proxies. To reconstruct dynamic scenes, Albert et al. [40] first explore the possibility by adding an extra time-conditioned deformation field to the static NeRF. The time-conditioned deformation field transforms sample points in the observation space into the canonical space, where their colours and opacities are queried and then rendered into images. This idea stimulated a series of dynamic NeRF reconstruction methods [5,28,36,37,52,58]. While our approach also uses a deformation field, we inject general plausibility constraints into it to keep the" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.444, + 0.128 + ], + "angle": 0, + "content": "T. Elsner, J. Berger et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.787, + 0.176 + ], + "angle": 0, + "content": "results reasonable, enabling the idea behind seam carving on different types of visual data." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.192, + 0.789, + 0.48 + ], + "angle": 0, + "content": "Seam Carving Image/video retargeting is an important tool in computer vision and computer graphics. Early works [26] first annotate or detect the region of interest (ROI), which is uniformly scaled to the target size afterwards. While contents out of the ROI are deformed by fisheye warping, Liu and Gleicher [27] later extend it to the video domain by introducing a motion salience map. Avidan and Shamir [2] propose seam carving, a content-aware image resizing method. It first defines a saliency map by summing the gradient norms along the horizontal and vertical directions. The cost of a seam is defined by adding up the values on the salience map (energy). The seam with minimal energy can be found using dynamic programming and will be removed in the resizing process. Rubinstein et al. [43] extend the seam carving idea to the video domain and transform the dynamic programming problem to a graph cut problem. Follow-up works [7,9,44] incorporate homogeneous cropping and scaling with seam carving and excavate more high-level information like symmetry [53], depth [3], and object similarity [8] to assist seam carving. With the advances in deep learning, a few methods [31-33] made attempts to detect and localise the seam carving forgery. Song et al. [46] propose Carving-Net to replace the previous handcrafted energy map with a neural network while the seam finding algorithm is still dynamic programming." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.494, + 0.789, + 0.614 + ], + "angle": 0, + "content": "Generally, our method is invariant to the underlying energy term, and hence can be replaced by any other formulation like the ones mentioned before. This relationship is visualised in the taxonomy in Fig. 2: Our method solves a global, continuous optimisation problem instead of greedy and local steps formulated as dynamic programming, and can hence function as a more flexible backbone for approaches that derive from seam carving. This formulation allows more complex solutions that work better in many scenarios (see Fig. 5), while the space of all possible outputs contains the solutions of seam carving." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.638, + 0.717, + 0.656 + ], + "angle": 0, + "content": "3 Deformation Fields for Retargeting Visual Data" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.67, + 0.787, + 0.716 + ], + "angle": 0, + "content": "Editing strategies for visual data require avoiding changes to meaningful content, that create visible artefacts. An approach that focuses on retargeting the input to match a target size can be guided by the following set of rules:" + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.725, + 0.784, + 0.767 + ], + "angle": 0, + "content": "- Content Aware Objective: Prevent the accidental creation of new or the removal of existing regions with high information content (control over where deformation happens)." + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.77, + 0.784, + 0.798 + ], + "angle": 0, + "content": "- Sanity Objective: Avoid introducing deformations that yield implausible results (control over how deformation happens)." + }, + { + "type": "list", + "bbox": [ + 0.225, + 0.725, + 0.784, + 0.798 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.84 + ], + "angle": 0, + "content": "In the following sections, we are only concerned with the retargeting to a smaller size (shrinking) for ease of explanation. The concept can naturally be" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.402, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "Retargeting Visual Data with Deformation Fields" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "image", + "bbox": [ + 0.217, + 0.142, + 0.788, + 0.284 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.293, + 0.789, + 0.366 + ], + "angle": 0, + "content": "Fig. 3: Our proposed pipeline for retargeting visual data: For a given input (left), we train two simple networks that learn the energy and cumulative energy along the deformation axis of the input (centre-left). We initialise a network that stretches samples to the desired position (centre-right), then optimise this deformation to distribute the distortion to low information content regions (right). Balloon image from [42]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.396, + 0.784, + 0.426 + ], + "angle": 0, + "content": "reformulated to expand the visual data instead (expansion, see Sec. C (Appendix)) or to include editing (see Sec. 3.5)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.427, + 0.784, + 0.532 + ], + "angle": 0, + "content": "Seam carving [2] and its derivatives produce a discrete solution to this by performing edits through manipulating (removing or repeating) one seam of pixels at a time, which is a single-pixel-wide path passing through the image. Through dynamic programming, seam carving chooses this path based on the lowest energy of the pixels, with high energy indicating information-rich areas. In a simple case, this energy is the colour gradient. This yields a discrete solution to the retargeting problem that is both content aware and plausible." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.534, + 0.788, + 0.625 + ], + "angle": 0, + "content": "To obtain a continuous solution, instead of extracting discrete seams of the same width, we use continuous folds of varying width using a learned continuous deformation function. This formulation optimises a solution that is a superset of all existing seam carving solutions (see Fig. 5). We always use the same core formulation, regardless of the underlying domain and application, e.g. moving an object in an image or retargeting a 3D scene." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.654, + 0.434, + 0.669 + ], + "angle": 0, + "content": "3.1 General Formulation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.686, + 0.788, + 0.807 + ], + "angle": 0, + "content": "Our approach always optimises a deformation function that maps from points in the retargeted output space (e.g. an image at half the size) \\( I: P \\to \\mathbb{R}^n \\) to points in the original data \\( I': P' \\to \\mathbb{R}^n \\) (e.g. to \\( \\mathbb{R}^3 \\) for RGB colours) by adding an offset to their coordinate. The value of a point \\( p \\) in the output data \\( I \\) is hence defined as the value of \\( p \\) deformed to a new point \\( p' \\) looked up in \\( I' \\). To represent the deformation function \\( D \\), we use a neural deformation field, a simple MLP where \\( D: P \\to \\mathbb{R} \\). To keep our approach simpler, we only allow scalar deformations along a fixed deformation direction \\( v \\), e.g. along the \\( x \\) axis (width) of an image:" + }, + { + "type": "equation", + "bbox": [ + 0.418, + 0.825, + 0.786, + 0.842 + ], + "angle": 0, + "content": "\\[\nI (p) = I ^ {\\prime} (p + D (p) \\cdot v) \\tag {1}\n\\]" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.444, + 0.128 + ], + "angle": 0, + "content": "T. Elsner, J. Berger et al." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.145, + 0.789, + 0.322 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.332, + 0.788, + 0.388 + ], + "angle": 0, + "content": "Fig. 4: Our pipeline applied to neural radiance fields: We first obtain a point cloud, then estimate the energy values. We then learn a continuous energy function and a cumulative energy function that we use to optimise our deformation network. The resulting deformation field is visualised on the bottom right." + }, + { + "type": "image", + "bbox": [ + 0.235, + 0.406, + 0.765, + 0.567 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.578, + 0.788, + 0.621 + ], + "angle": 0, + "content": "Fig. 5: Example of applying retargeting to \\(75\\%\\) width. Seam carving fails to use seams that are non-continuous or with a too steep angle, not properly using the empty space in the shelf to close the gaps before shrinking the bookshelves content." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.656, + 0.788, + 0.718 + ], + "angle": 0, + "content": "Hence, to get the content for a point \\( p \\) in the retargeted data, we offset the coordinate and look up the content at the offset coordinate \\( p' \\) in the original data. To fulfil the objectives stated in the beginning of Sec. 3, we introduce two sets of losses, content aware and sanity objectives:" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.731, + 0.788, + 0.809 + ], + "angle": 0, + "content": "Content Aware Objectives apply regularisation to guide the deformation process when adjusting to the data content. In particular, we penalise deformation in regions with high information content to discourage noticeable changes. We define the energy function \\( E: P' \\to \\mathbb{R} \\) to be a measure for content information in the original visual data, expresses as, e.g., a colour gradient:" + }, + { + "type": "equation", + "bbox": [ + 0.433, + 0.824, + 0.786, + 0.842 + ], + "angle": 0, + "content": "\\[\nE \\left(p ^ {\\prime}\\right) = \\left| \\left| \\nabla I \\left(p ^ {\\prime}\\right) \\right| \\right| _ {2} \\tag {2}\n\\]" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.402, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "Retargeting Visual Data with Deformation Fields" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "image", + "bbox": [ + 0.278, + 0.144, + 0.468, + 0.274 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.472, + 0.147, + 0.594, + 0.272 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.599, + 0.146, + 0.723, + 0.273 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.285, + 0.785, + 0.313 + ], + "angle": 0, + "content": "Fig. 6: Applying our retargeting to 3D meshes, moving furniture closer together to retarget the scene to a smaller size while limiting distortion to objects." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.343, + 0.784, + 0.373 + ], + "angle": 0, + "content": "The overall content deformation loss is the product of deformation magnitude at \\( p \\) and energy at the original \\( p' \\) over all \\( p \\) in the deformed data:" + }, + { + "type": "equation", + "bbox": [ + 0.348, + 0.385, + 0.785, + 0.425 + ], + "angle": 0, + "content": "\\[\nL _ {C} = \\int_ {p \\in P} [ E (p + v \\cdot D (p)) \\cdot | | \\nabla D (p) | | _ {1} ] d p \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.437, + 0.784, + 0.528 + ], + "angle": 0, + "content": "Thus, in Eq. (3), we effectively penalise visible deformations defined as the product of information content at points, namely \\( E(p + v \\cdot D(p)) \\), and the deformation itself via \\( ||\\nabla D(p)||_1 \\). Using the \\( L^1 \\)-norm for the distortion gradient promotes piecewise constant, possibly sparse, deformations. We can then split up the change in deformation into change in deformation direction \\( v \\) and change in direction \\( v^\\perp \\) orthogonal to \\( v \\):" + }, + { + "type": "equation", + "bbox": [ + 0.28, + 0.538, + 0.785, + 0.582 + ], + "angle": 0, + "content": "\\[\nL _ {C} = \\int_ {p \\in P} \\left[ E (p + v \\cdot D (p)) \\cdot \\left(\\left| \\frac {\\partial}{\\partial v} D (p) \\right| + \\left| \\frac {\\partial}{\\partial v ^ {\\perp}} D (p) \\right|\\right) \\right] d p \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.592, + 0.785, + 0.653 + ], + "angle": 0, + "content": "This equation is then further split into two components, \\( L_{e} \\) and \\( L_{s} \\). The loss \\( L_{e} \\) penalises stretching or compression and \\( L_{s} \\) punishes shearing, which we define to be deformation change in a direction that is not the deformation direction. For a now discrete sample set \\( P \\) from our domain, these terms are expressed as:" + }, + { + "type": "equation", + "bbox": [ + 0.354, + 0.663, + 0.785, + 0.711 + ], + "angle": 0, + "content": "\\[\nL _ {e} = \\sum_ {p \\in P} \\left[ E (p + v \\cdot D (p)) \\cdot \\left| \\frac {\\partial}{\\partial v} D (p) \\right| \\right] \\tag {5}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.358, + 0.705, + 0.644, + 0.743 + ], + "angle": 0, + "content": "\\[\nL _ {s} = \\sum_ {p \\in P} \\left[ E (p + v \\cdot D (p)) \\cdot \\left| \\frac {\\partial}{\\partial v ^ {\\perp}} D (p) \\right| \\right]\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.75, + 0.784, + 0.84 + ], + "angle": 0, + "content": "Essentially, \\( L_{e} \\) directs deformations to areas of low information content, while \\( L_{s} \\) discourages the resulting shearing to affect areas of high information content. Note that while \\( L_{s} \\) would also penalise non-straight deformation seams, we tune our losses such that this only impacts the output much if an area is sheared. An example of this is shown in Fig. 3: The deformations around the balloons introduce shearing only at the seam, in an area of low information content." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.444, + 0.129 + ], + "angle": 0, + "content": "T. Elsner, J. Berger et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.238 + ], + "angle": 0, + "content": "In case of image deformation, neighbouring pixels with low energy in the input may translate to pixels widely apart in the output, potentially bypassing a substantial amount of important content. Jumps in the deformation field \\( D \\) indicate this effect. To prevent this, we penalise the amount of energy \\( \\hat{E} \\) between two points \\( p', q' \\in P' \\) in the original domain, rather than the energy at individual positions, as in Eq. (2):" + }, + { + "type": "equation", + "bbox": [ + 0.403, + 0.247, + 0.786, + 0.29 + ], + "angle": 0, + "content": "\\[\n\\hat {E} \\left(p ^ {\\prime}, q ^ {\\prime}\\right) = \\int_ {r ^ {\\prime} \\in \\left[ p ^ {\\prime}, q ^ {\\prime} \\right]} E \\left(r ^ {\\prime}\\right) d r ^ {\\prime} \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.299, + 0.787, + 0.345 + ], + "angle": 0, + "content": "Using \\(\\hat{E}\\), we then redefine \\(L_{e}\\) from Eq. (5) to consider energy and difference in deformation magnitude between \\(p\\) and a slightly offset \\(p_{\\varepsilon} \\coloneqq p + \\varepsilon\\) in deformation direction:" + }, + { + "type": "equation", + "bbox": [ + 0.304, + 0.357, + 0.786, + 0.395 + ], + "angle": 0, + "content": "\\[\nL _ {e} = \\sum_ {p \\in P} \\hat {E} (p + v D (p), p _ {\\varepsilon} + v D (p _ {\\varepsilon})) \\cdot \\frac {| D (p) - D (p _ {\\varepsilon}) |}{\\varepsilon} \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.399, + 0.787, + 0.46 + ], + "angle": 0, + "content": "This formulation guides the deformation to avoid introducing deformation to high information content regions or fold over them, while \\( L_{s} \\) discourages non-local deformations that would apply shearing in high information areas. For example, shearing in blue sky is acceptable, shearing on a brick wall is not." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.474, + 0.787, + 0.506 + ], + "angle": 0, + "content": "Sanity objectives prevent undesired outputs by introducing two additional losses. Assume we shrink the input by a factor \\(\\alpha \\in [0,1]\\):" + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.512, + 0.788, + 0.588 + ], + "angle": 0, + "content": "1. The boundaries of the affected region should not change, e.g. an image should not be simply cropped. Based on \\(\\alpha\\), the deformation field should map the start/end of the deformation axis in the output to the start/end of the input. For \\(P_0\\), \\(P_1\\) containing all points at the start/end of the deformation axis of the output, we express this requirement as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.352, + 0.597, + 0.787, + 0.638 + ], + "angle": 0, + "content": "\\[\nL _ {b} = \\int_ {p \\in P _ {0}} | D (p) | d p + \\int_ {p \\in P _ {1}} | D (p) - (1 - \\alpha) | d p \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.644, + 0.787, + 0.66 + ], + "angle": 0, + "content": "2. We additionally demand monotonicity in the deformation field, expressed as" + }, + { + "type": "equation", + "bbox": [ + 0.393, + 0.671, + 0.787, + 0.713 + ], + "angle": 0, + "content": "\\[\nL _ {m} = \\int_ {p \\in P} \\max \\left(0, - \\frac {\\partial}{\\partial v} D (p)\\right) d p, \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.242, + 0.717, + 0.787, + 0.778 + ], + "angle": 0, + "content": "indicating that any change in the deformation direction should be positive. Therefore, deformation \\( D \\) should be monotonic and only stagnate or increase exclusively along \\( v \\). This prevents the repetition of the same image portion, noticeable by jumps with inconsistent direction in the deformation field." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.8, + 0.652, + 0.816 + ], + "angle": 0, + "content": "We then obtain the total loss \\(L\\) with hyperparameters \\(\\lambda_{*}\\) as" + }, + { + "type": "equation", + "bbox": [ + 0.381, + 0.826, + 0.786, + 0.842 + ], + "angle": 0, + "content": "\\[\nL = \\lambda_ {e} L _ {e} + \\lambda_ {s} L _ {s} + \\lambda_ {b} L _ {b} + \\lambda_ {m} L _ {m}. \\tag {10}\n\\]" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.401, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Retargeting Visual Data with Deformation Fields" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.147, + 0.784, + 0.162 + ], + "angle": 0, + "content": "We initialise the deformation function with a uniform stretch to the target size." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.163, + 0.784, + 0.223 + ], + "angle": 0, + "content": "In the following, we demonstrate the application of this general formulation to different domains, while adapting the energy formulation to the problem. All implementation and architecture details are given in Sec. A (Appendix) and the exact loss terms in Sec. B (Appendix)." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.246, + 0.61, + 0.261 + ], + "angle": 0, + "content": "3.2 Optimising Deformation Fields for Images" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.272, + 0.784, + 0.318 + ], + "angle": 0, + "content": "Instead of using image data directly, we train a continuous image representation MLP \\(I': P' \\to \\mathbb{R}^3\\) to be our input data, that maps pixel positions to colours. This ensures consistency with our continuous approaches." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.332, + 0.784, + 0.424 + ], + "angle": 0, + "content": "We train one MLP to learn the energy field \\( E \\), holding the colour gradient per point as in Eq. (2). Using the learned \\( E \\), we train another MLP to replicate \\( \\hat{E} \\) as a cumulative energy field \\( \\Sigma \\). The difference in cumulative energy \\( \\Sigma \\) between two points \\( p', q' \\in P' \\) on the same axis \\( v \\) is the absolute difference in information content between them. As this coincides with the definition of \\( \\hat{E} \\) from Eq. (6), we approximate the measure of content relevance between two points as:" + }, + { + "type": "equation", + "bbox": [ + 0.408, + 0.435, + 0.785, + 0.452 + ], + "angle": 0, + "content": "\\[\n\\hat {E} \\left(p ^ {\\prime}, q ^ {\\prime}\\right) \\approx | \\Sigma \\left(q ^ {\\prime}\\right) - \\Sigma \\left(p ^ {\\prime}\\right) | \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.463, + 0.785, + 0.585 + ], + "angle": 0, + "content": "This cumulative gradient formulation therefore simplifies the interval over \\( E \\) in Eq. (6) to a simple difference computation between two MLP queries: How much content is lost when the deformation skips a region. The energy field \\( E \\) can be trained to reproduce any energy formulation other than a colour gradient. This exchangeable energy formulation ensures compatibility with other retargeting approaches as shown in Fig. 2. For all objectives, we use discrete approximations of the loss terms introduced in Sec. 3.1. The whole pipeline is visualised in Fig. 3, and additional results can be found in Sec. G (Appendix)." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.607, + 0.743, + 0.623 + ], + "angle": 0, + "content": "3.3 Optimising Deformation Fields for Neural Radiance Fields" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.633, + 0.784, + 0.664 + ], + "angle": 0, + "content": "We adapt our approach from Sec. 3.1 for 3D scenes given as trained NeRFs as follows (see Fig. 4):" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.674, + 0.785, + 0.719 + ], + "angle": 0, + "content": "- A trained NeRF provides colour and depth images, from which we extract surface samples. Each point is assigned an energy value defined by change in colour and depth at this point." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.72, + 0.785, + 0.763 + ], + "angle": 0, + "content": "- Based on the surface samples and their energy values, we train a continuous energy field \\( E \\) and a continuous cumulative energy field \\( \\Sigma \\), to easily access the amount of content between points." + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.765, + 0.785, + 0.809 + ], + "angle": 0, + "content": "- We use an inverse deformation network \\( U \\) to find points in the target space that correspond to surface points in the source space, keeping computing costs low by restricting our optimisation to surface points." + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.811, + 0.785, + 0.839 + ], + "angle": 0, + "content": "- We apply our regularisation terms on a mix of surface and random points to optimise our loss function." + }, + { + "type": "list", + "bbox": [ + 0.225, + 0.674, + 0.785, + 0.839 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.444, + 0.129 + ], + "angle": 0, + "content": "T. Elsner, J. Berger et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.254 + ], + "angle": 0, + "content": "1. Extracting a Sparse Set of Representatives and Energy We extract a surface point cloud from the RGB and depth images rendered by NeRF at the training camera positions. Only these points are relevant for scene deformation. We then compute the energy value for each point corresponding to a pixel as a sum of colour change in the RGB image and depth change in the depth image. To reduce the impact of highly noisy energy values, we apply smoothing by selecting the minimum energy for each point within a neighbourhood radius of 5 points." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.267, + 0.788, + 0.419 + ], + "angle": 0, + "content": "2. Defining an Energy Function To obtain a continuous energy field, we train an auxiliary energy network \\( E \\). The network maps all 3D points to their energy values, and clamps the energy to zero for points further than twice the average neighbour distance. The resulting energy field has zero energy in empty space and inside objects, and non-zero energy values for points near the surface. Similar to our approach with images, we learn a cumulative energy network \\( \\Sigma \\). The network is trained to accumulate the energy values \\( E(p) \\) from uniformly sampled points \\( p \\) on a ray in deformation direction. We compute the energy for a segment between two points \\( p', q' \\in \\mathbb{R}^3 \\) along \\( v \\) using their difference in cumulative energy \\( \\Sigma(q') \\) and \\( \\Sigma(p') \\) (cf. Eq. (11))." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.432, + 0.788, + 0.568 + ], + "angle": 0, + "content": "3. Inverting the Deformation We define an inverse deformation function to determine those points that lie on the surface after deformation. In contrast to images, we only perform optimisation on a sparse set of points, namely surface points. As we optimise a deformation from the target space to the source space, we need the position of surface points in both spaces. By training an MLP \\(U\\) to reverse the deformation \\(D\\), i.e. minimising \\((U(D(x)) = x)^2\\), we can effortlessly extract surface points, and hence their energy values, in the target space. This relation is also visualised in Fig. 4. We always update \\(U\\) once after training one iteration of \\(D\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.583, + 0.788, + 0.644 + ], + "angle": 0, + "content": "As we use a discrete set of points, we apply the losses from Sec. 3.1 to Eq. (10) for a discrete interval. In particular, we average the loss over a random subset of sample points. Example images from deformed scenes can be found in Fig. 8, while a larger number of results can be found in Sec. G (Appendix)." + }, + { + "type": "title", + "bbox": [ + 0.214, + 0.665, + 0.613, + 0.68 + ], + "angle": 0, + "content": "3.4 Optimising Deformation Fields for Meshes" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.788, + 0.765 + ], + "angle": 0, + "content": "To demonstrate the versatility of our approach, we apply it to meshes as well. We adapt the energy term and use surface samples and vertex positions for optimisation. For freestanding objects, we use curvature as the energy measure. For the room example in Fig. 6, the energy is set to 0 for the floor and to 1 for every other surface point. We otherwise follow the same principle as in Sec. 3.3." + }, + { + "type": "title", + "bbox": [ + 0.214, + 0.786, + 0.421, + 0.801 + ], + "angle": 0, + "content": "3.5 Other Applications" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.842 + ], + "angle": 0, + "content": "While retargeting is our main contribution, our framework opens up multiple different editing opportunities that even extend the capabilities of seam carving [2]." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.402, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "Retargeting Visual Data with Deformation Fields" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.188, + 0.231, + 0.218 + ], + "angle": 0, + "content": "Iinui" + }, + { + "type": "image", + "bbox": [ + 0.254, + 0.162, + 0.417, + 0.24 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.437, + 0.162, + 0.503, + 0.24 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.263, + 0.231, + 0.298 + ], + "angle": 0, + "content": "ed" + }, + { + "type": "image", + "bbox": [ + 0.254, + 0.242, + 0.417, + 0.321 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.437, + 0.242, + 0.503, + 0.321 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.337, + 0.505, + 0.394 + ], + "angle": 0, + "content": "Fig. 7: We demonstrate different editing operations by adding and adapting the loss formulation. Left: Removing, right: Moving an object. Images from [42]." + }, + { + "type": "image", + "bbox": [ + 0.533, + 0.153, + 0.785, + 0.267 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.528, + 0.268, + 0.786, + 0.317 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.526, + 0.327, + 0.788, + 0.41 + ], + "angle": 0, + "content": "Fig.8: Applying our retargeting to NeRFs. Top row: Input, stretched, targeted with our method. Bottom row: Input, folds (white) on the input suggested by our approach, targeted with our method." + }, + { + "type": "image", + "bbox": [ + 0.217, + 0.439, + 0.331, + 0.501 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.332, + 0.439, + 0.443, + 0.5 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.444, + 0.439, + 0.501, + 0.5 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.439, + 0.56, + 0.5 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.561, + 0.439, + 0.67, + 0.5 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.671, + 0.439, + 0.728, + 0.5 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.729, + 0.439, + 0.785, + 0.5 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.512, + 0.785, + 0.555 + ], + "angle": 0, + "content": "Fig. 9: Selected results showing scenes where our approach benefits from a higher degree of flexibility. All images are arranged as input, seam carving, and our output. Images from [42]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.599, + 0.785, + 0.689 + ], + "angle": 0, + "content": "The loss formulation is slightly adapted to implement the following exemplary operations, which are applicable to any domain and can serve as the basis for additional processing methods or user-driven applications. Further details can be found in Sec. D (Appendix). Examples can be found in Fig. 7. We want to remove the monotonicity loss \\( L_{m} \\) and use equal input and output sizes for both the removal and the movement of an object." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.69, + 0.785, + 0.749 + ], + "angle": 0, + "content": "Removal To remove a certain region, we employ an additional loss term that penalises any deformation target leading to that region. Points that were originally assigned to the target region will be distorted and assigned to neighbouring regions, resulting in the removal of the region in the output." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Moving an Object Although classic seam carving cannot move an object easily, as seams are assumed to cross the entire image, our formulation does not have this restriction. To move an object, we also enforce that the deformation towards the target feature only occurs from the desired new placement in the output. Any occurrence of that feature elsewhere is punished to avoid its repetition." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.444, + 0.129 + ], + "angle": 0, + "content": "T. Elsner, J. Berger et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.145, + 0.788, + 0.189 + ], + "angle": 0, + "content": "Table 1: Comparison of FID scores on the RetargetMe [42] dataset, for retargeting x (top) and y (bottom) axis to a smaller (left) and bigger (right) size. In mean, ours outperforms seam carving with 46.67 to 52.57 for x and 53.57 to 65.48 for y." + }, + { + "type": "table", + "bbox": [ + 0.226, + 0.2, + 0.779, + 0.293 + ], + "angle": 0, + "content": "
x FID(↓)50%60%70%80%90%110%120%130%140%150%
Seam carving85.7566.4151.6735.2023.8020.8630.0137.3843.4546.37
Ours79.3958.0643.6930.5321.7111.1719.4229.1937.8749.74
y FID(↓)50%60%70%80%90%110%120%130%140%150%
Seam carving116.7084.3360.5940.8624.9223.0433.5840.8248.4952.59
Ours89.4467.7749.6235.5925.4213.5521.0227.5538.2046.40
" + }, + { + "type": "table_caption", + "bbox": [ + 0.224, + 0.307, + 0.502, + 0.376 + ], + "angle": 0, + "content": "Table 2: User study comparing our approach to seam carving \\((SC)\\) for images. 19 participants were questioned for preferences of random examples from the RetargetMe dataset [42], casting 1200 votes." + }, + { + "type": "table", + "bbox": [ + 0.25, + 0.388, + 0.479, + 0.497 + ], + "angle": 0, + "content": "
WidthOursSCDraw
50 %48.48%20.87%30.67%
150 %33.8%36.27%29.92%
HeightOursSCDraw
50 %49.5%24.25%26.25%
150 %45.84%20.0%34.15%
Total44.57%25.1%30.32%
" + }, + { + "type": "image", + "bbox": [ + 0.513, + 0.334, + 0.641, + 0.412 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.643, + 0.319, + 0.764, + 0.413 + ], + "angle": 0, + "content": "Colour Complex Gradient Energy" + }, + { + "type": "image_caption", + "bbox": [ + 0.5, + 0.414, + 0.779, + 0.47 + ], + "angle": 0, + "content": "Fig. 10: Example of using our method with the more complex energy term of [48] instead of gradients, improving the results on an example." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.525, + 0.357, + 0.541 + ], + "angle": 0, + "content": "4 Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.56, + 0.788, + 0.757 + ], + "angle": 0, + "content": "Retargeting not only lacks comparable methods for other visual data such as meshes, but its quality is also difficult to measure. In 2D, we compare our results to seam carving [2] on images using the RetargetMe [42] benchmark. In 3D, we compare our results on NeRFs with video-seam-carved recordings of the same scene, because video seam carving [43] is the closest method to 3D retargeting. We always produce a short, 3 second clip at 30 frames per second. To ensure a fair comparison with the method, we only tilt the camera down instead of rotating it around the object. Even for benevolent camera movements, it produces visible artefacts, as seen in Sec. F (Appendix). Exemplary, we show that our approach is compatible with approaches that propose different saliency maps as visualised in Fig. 2. For that purpose, we compare our method to other more recent image retargeting approaches visually (Fig. 10) and in a user study (Tab. 4). Additional results can be found in Sec. G (Appendix), including an explicit failure case." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.782, + 0.37, + 0.797 + ], + "angle": 0, + "content": "4.1 Comparisons" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.842 + ], + "angle": 0, + "content": "Images In accordance with recent publications in the generative model and image editing domain [15, 23, 38], we extensively compare our results with the" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.402, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Retargeting Visual Data with Deformation Fields" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.145, + 0.788, + 0.2 + ], + "angle": 0, + "content": "Table 3: User study comparing retargeting a scene, then recording it (ours), to recording it and then retargetet it (video seam carving [43], VSC), using our provided dataset. 21 participants were asked to indicated which solution they prefer, casting 252 votes in total." + }, + { + "type": "table", + "bbox": [ + 0.242, + 0.213, + 0.761, + 0.287 + ], + "angle": 0, + "content": "
SceneOursVSCDrawOursVSCDrawOursVSCDrawOursVSCDraw
Beach1731201019111911
Farm2100210016322100
Harb.2100210021002100
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.331, + 0.788, + 0.436 + ], + "angle": 0, + "content": "original seam carving results using FID [20] in Tab. 1 for images. We also evaluate our approach in a user study. The participants were instructed to classify their preference for a retargeted version of a randomly selected image based on a randomly selected deformation axis, always choosing between our and the seam carved [2] version. We used the 80 images of the RetargetMe [42] for 2D, producing 360 variations per approach in Tab. 2. This shows how our approach can be a more powerful backbone for seam carving." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.496, + 0.532, + 0.538 + ], + "angle": 0, + "content": "Table 4: User study measuring preference, pairwise comparing to our approach using salience maps of [48]." + }, + { + "type": "table", + "bbox": [ + 0.224, + 0.539, + 0.526, + 0.603 + ], + "angle": 0, + "content": "
ApproachOursTheirsDraw
CarvingNet [46]45.16%41.93%12.91%
RL [22]46.85%32.43%20.72%
Cycle-IR [49]40.3%34.46%25.24%
" + }, + { + "type": "text", + "bbox": [ + 0.542, + 0.452, + 0.788, + 0.633 + ], + "angle": 0, + "content": "We further compare against other state-of-the-art retargeting approaches in a user study in Tab. 4: We show that our method can compete both against a seam carving variation with better saliency from deep learning [46] while also beating state-of-the-art approaches that retarget the image directly [22,49]. Note that we only compare to readily available images from other approach, as" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.633, + 0.788, + 0.754 + ], + "angle": 0, + "content": "we often could not reproduce them or run the code. For this comparison, we use the saliency method FullGrad [48] that aggregates layer-wise bias-gradients from across ResNet [16]. This aggregated gradient map acts as our saliency map, as it measures how useful the feature is to the neural network, hence we use it as a measure for how semantically important a feature is. This also shows that our approach can work with different saliency maps. Details and examples can be found in Sec. E (Appendix). Additionally, we provide a visual ablation that shows the importance of each loss component in Fig. 11." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.84 + ], + "angle": 0, + "content": "3D Scenes When assessing the quality of retargeting 3D scenes provided as NeRFs, we compared our results with what we consider to be the most similar approach, video seam carving [43]. In our provided GTA V NeRF retargeting dataset, we resize each axis by \\(50\\%\\) and \\(150\\%\\). The results can be found in Tab. 3." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.444, + 0.128 + ], + "angle": 0, + "content": "T. Elsner, J. Berger et al." + }, + { + "type": "image", + "bbox": [ + 0.23, + 0.144, + 0.77, + 0.194 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.208, + 0.788, + 0.252 + ], + "angle": 0, + "content": "Fig. 11: Visual ablation for retargeting to \\(80\\%\\) width on an image [12], from left to right: Wobbly contours from shearing (no \\(L_{s}\\)), uniform stretch on the whole image (no \\(L_{e}\\)), missing features e.g. the person (no \\(L_{m}\\)), content outside the image (no \\(L_{b}\\))." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.278, + 0.785, + 0.31 + ], + "angle": 0, + "content": "We provide additional results and frames of the resulting sequences from our approach for 3D scenes compared to video seam carving in Sec. F (Appendix)." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.329, + 0.484, + 0.343 + ], + "angle": 0, + "content": "4.2 Discussion and Limitations" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.352, + 0.789, + 0.655 + ], + "angle": 0, + "content": "With our deformation fields as backbone, but the same energy term, our approach outperforms seam carving on both a qualitative and quantitative level. Our flexibility in the energy formulation allows for more possibilities, and the replacement of the gradient-based energy formulation with a better saliency map improves our approach even more. This shows that our method is a flexible backbone that can be used with various successors to seam carving, which provide more advanced saliency maps. Our continuous formulation allows for retargeting in various domains, including 3D scenes, and highlights the advantages of a geometry-aware approach. Applying our approach with only colour gradients can produce flawed outputs, just as those of seam carving, see Fig. 15 (Appendix). As we optimise a neural field for each individual input, we trade computation time for not needing a dataset to train and for a global optimisation. Our deformation optimisation takes approximately one minute on consumer hardware for good results versus about half a second for seam carving. Our strength, however, lies in 3D. Video seam carving on a 600 by 400 pixel video can take several hours, whereas our approach produces the final retargeted 3D scene in under 15 minutes. For Future Work, investigating invertible neural networks to implicitly learn an inverse deformation for the 3D case [1] seems promising, as would a deep learned approach to evaluating a deformation for optimisation and more research into good measures for retargeting quality." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.675, + 0.36, + 0.691 + ], + "angle": 0, + "content": "5 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.705, + 0.788, + 0.841 + ], + "angle": 0, + "content": "We present an approach that generalises the principle underlying seam carving to create image variants by manipulating low content areas. By training neural deformation fields that avoid applying distortion to important parts of an input, our approach is applicable to different forms of visual data retargeting. Our formulation is largely domain invariant and only needs minor alterations to work on different types of visual data, as demonstrated for neural radiance fields and meshes for the first time. We demonstrate compatibility to other energy formulations of the seam carving family, being a more general way of optimising image retargeting. Our approach is the only one that natively works on 3D scenes." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.402, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Retargeting Visual Data with Deformation Fields" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.323, + 0.16 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.179, + 0.786, + 0.234 + ], + "angle": 0, + "content": "1. Ardizzone, L., Kruse, J., Wirkert, S.J., Rahner, D., Pellegrini, E.W., Klessen, R.S., Maier-Hein, L., Rother, C., Kothe, U.: Analyzing inverse problems with invertible neural networks. CoRR abs/1808.04730 (2018), http://arxiv.org/abs/1808.04730" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.236, + 0.786, + 0.263 + ], + "angle": 0, + "content": "2. Avidan, S., Shamir, A.: Seam carving for content-aware image resizing. ACM Trans. Graph. 26(3), 10 (2007)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.264, + 0.786, + 0.292 + ], + "angle": 0, + "content": "3. Basha, T.D., Moses, Y., Avidan, S.: Stereo seam carving a geometrically consistent approach. IEEE Trans. Pattern Anal. Mach. Intell. 35(10), 2513-2525 (2013)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.293, + 0.786, + 0.333 + ], + "angle": 0, + "content": "4. Berkiten, S., Halber, M., Solomon, J., Ma, C., Li, H., Rusinkiewicz, S.: Learning detail transfer based on geometric features. Comput. Graph. Forum 36(2), 361-373 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.334, + 0.786, + 0.375 + ], + "angle": 0, + "content": "5. Cai, H., Feng, W., Feng, X., Wang, Y., Zhang, J.: Neural surface reconstruction of dynamic scenes with monocular rgb-d camera. In: Neural Information Processing Systems (NeurIPS) (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.377, + 0.786, + 0.404 + ], + "angle": 0, + "content": "6. Chen, Zhiqin and Zhang, Hao: Learning implicit fields for generative shape modeling. In: CVPR. pp. 5939-5948 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.405, + 0.786, + 0.432 + ], + "angle": 0, + "content": "7. Dong, W., Bao, G., Zhang, X., Paul, J.: Fast multi-operator image resizing and evaluation. J. Comput. Sci. Technol. 27(1), 121-134 (2012)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.433, + 0.786, + 0.487 + ], + "angle": 0, + "content": "8. Dong, W., Zhou, N., Lee, T., Wu, F., Kong, Y., Zhang, X.: Summarization-based image resizing by intelligent object carving. IEEE Trans. Vis. Comput. Graph. 20(1), 1 (2014). https://doi.org/10.1109/TVCG.2013.103, https://doi.org/10.1109/TVCG.2013.103" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.489, + 0.786, + 0.516 + ], + "angle": 0, + "content": "9. Dong, W., Zhou, N., Paul, J., Zhang, X.: Optimized image resizing using seam carving and scaling. ACM Trans. Graph. 28(5), 125 (2009)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.518, + 0.786, + 0.559 + ], + "angle": 0, + "content": "0. Efros, A.A., Freeman, W.T.: Image quilting for texture synthesis and transfer. In: Pocock, L. (ed.) Proceedings of the 28th Annual Conference on Computer Graphics and Interactive Techniques, SIGGRAPH. pp. 341-346. ACM (2001)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.56, + 0.786, + 0.586 + ], + "angle": 0, + "content": "1. Efros, A.A., Leung, T.K.: Texture synthesis by non-parametric sampling. In: ICCV. pp. 1033-1038 (1999)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.588, + 0.786, + 0.615 + ], + "angle": 0, + "content": "2. at English Wikipedia, N.: Broadway tower (2007), https://commons.wikimedia.org/wiki/File:Broadway_tower.jpg" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.616, + 0.786, + 0.657 + ], + "angle": 0, + "content": "3. Garbin, S.J., Kowalski, M., Estellers, V., Szymanowicz, S., Rezaeifar, S., Shen, J., Johnson, M., Valentin, J.: Voltemorph: Realtime, controllable and generalisable animation of volumetric representations. arXiv:2208.00949 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.658, + 0.786, + 0.685 + ], + "angle": 0, + "content": "4. Gatys, L.A., Ecker, A.S., Bethge, M.: Texture synthesis using convolutional neural networks. In: Neural Information Processing Systems. pp. 262-270 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.686, + 0.786, + 0.713 + ], + "angle": 0, + "content": "5. Gu, J., Zhai, S., Zhang, Y., Susskind, J., Jaitly, N.: Matryoshka diffusion models. arXiv preprint arXiv:2310.15111 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.714, + 0.786, + 0.742 + ], + "angle": 0, + "content": "6. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. CoRR abs/1512.03385 (2015), http://arxiv.org/abs/1512.03385" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.743, + 0.786, + 0.77 + ], + "angle": 0, + "content": "7. Henzler, P., Deschaintre, V., Mitra, N.J., Ritschel, T.: Generative modelling of BRDF textures from flash images. ACM Trans. Graph. 40(6), 284:1-284:13 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.771, + 0.786, + 0.812 + ], + "angle": 0, + "content": "8. Henzler, P., Mitra, N.J., Ritschel, T.: Learning a neural 3d texture space from 2d exemplars. In: The IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.813, + 0.786, + 0.84 + ], + "angle": 0, + "content": "9. Hertz, A., Hanocka, R., Giryes, R., Cohen-Or, D.: Deep geometric texture synthesis. ACM Trans. Graph. 39(4), 108:1-108:11 (2020)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.179, + 0.786, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.444, + 0.128 + ], + "angle": 0, + "content": "T. Elsner, J. Berger et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.189 + ], + "angle": 0, + "content": "20. Heusel, M., Ramsauer, H., Unterthiner, T., Nessler, B., Hochreiter, S.: Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems 30 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.19, + 0.785, + 0.232 + ], + "angle": 0, + "content": "21. Huang, Y., Cao, Y., Lai, Y., Shan, Y., Gao, L.: Nerf-texture: Texture synthesis with neural radiance fields. In: ACM SIGGRAPH 2023 Conference Proceedings. pp. 43:1-43:10 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.232, + 0.785, + 0.272 + ], + "angle": 0, + "content": "22. Kajiura, N., Kosugi, S., Wang, X., Yamasaki, T.: Self-play reinforcement learning for fast image retargeting. CoRR abs/2010.00909 (2020), https://arxiv.org/abs/2010.00909" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.273, + 0.785, + 0.327 + ], + "angle": 0, + "content": "23. Kawar, B., Zada, S., Lang, O., Tov, O., Chang, H., Dekel, T., Mosseri, I., Irani, M.: Imagic: Text-based real image editing with diffusion models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6007-6017 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.328, + 0.785, + 0.37 + ], + "angle": 0, + "content": "24. Kopf, J., Fu, C.W., Cohen-Or, D., Deussen, O., Lischinski, D., Wong, T.T.: Solid texture synthesis from 2d exemplars. ACM Transactions on Graphics 26(3), 2:1-2:9 (2007)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.371, + 0.785, + 0.412 + ], + "angle": 0, + "content": "25. Lai, Y., Hu, S., Gu, X., Martin, R.R.: Geometric texture synthesis and transfer via geometry images. In: Proceedings of the Tenth ACM Symposium on Solid and Physical Modeling. pp. 15-26. ACM (2005)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.412, + 0.785, + 0.453 + ], + "angle": 0, + "content": "26. Liu, F., Gleicher, M.: Automatic image retargeting with fisheye-view warping. In: Proceedings of the 18th Annual ACM Symposium on User Interface Software and Technology. pp. 153-162. ACM (2005)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.453, + 0.785, + 0.495 + ], + "angle": 0, + "content": "27. Liu, F., Gleicher, M.: Video retargeting: automating pan and scan. In: Nahrstedt, K., Turk, M.A., Rui, Y., Klas, W., Mayer-Patel, K. (eds.) Proceedings of the 14th ACM International Conference on Multimedia. pp. 241-250. ACM (2006)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.495, + 0.785, + 0.522 + ], + "angle": 0, + "content": "28. Luiten, J., Kopanas, G., Leibe, B., Ramanan, D.: Dynamic 3d gaussians: Tracking by persistent dynamic view synthesis. In: 3DV (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.522, + 0.785, + 0.563 + ], + "angle": 0, + "content": "29. Mescheder, L., Oechsle, M., Niemeyer, M., Nowozin, S., Geiger, A.: Occupancy networks: Learning 3D reconstruction in function space. In: CVPR. pp. 4460-4470 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.564, + 0.785, + 0.605 + ], + "angle": 0, + "content": "30. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. In: ECCV (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.606, + 0.785, + 0.66 + ], + "angle": 0, + "content": "31. Moreira, T.P., Santana, M.C.S., Passos, L.A., Papa, J.P., da Costa, K.A.P.: An end-to-end approach for seam carving detection using deep neural networks. In: Pattern Recognition and Image Analysis. Lecture Notes in Computer Science, vol. 13256, pp. 447-457 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.66, + 0.785, + 0.702 + ], + "angle": 0, + "content": "32. Nam, S., Ahn, W., Yu, I., Kwon, M., Son, M., Lee, H.: Deep convolutional neural network for identifying seam-carving forgery. IEEE Trans. Circuits Syst. Video Technol. 31(8), 3308-3326 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.702, + 0.785, + 0.743 + ], + "angle": 0, + "content": "33. Nataraj, L., Gudavalli, C., Mohammed, T.M., Chandrasekaran, S., Manjunath, B.S.: Seam carving detection and localization using two-stage deep neural networks. CoRR abs/2109.01764 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.743, + 0.785, + 0.757 + ], + "angle": 0, + "content": "34. North, R.: Grand theft auto v. Steam (2015), https://www.rockstargames.com/V/" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.758, + 0.785, + 0.799 + ], + "angle": 0, + "content": "35. Park, J.J., Florence, P., Straub, J., Newcombe, R., Lovegrove, S.: DeepSDF: Learning continuous signed distance functions for shape representation. In: CVPR. pp. 165-174 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.799, + 0.785, + 0.841 + ], + "angle": 0, + "content": "36. Park, K., Sinha, U., Barron, J.T., Bouaziz, S., Goldman, D.B., Seitz, S.M., MartinBrualla, R.: Nerfies: Deformable neural radiance fields. In: ICCV. pp. 5845-5854 (2021)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.402, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Retargeting Visual Data with Deformation Fields" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.205 + ], + "angle": 0, + "content": "37. Park, K., Sinha, U., Hedman, P., Barron, J.T., Bouaziz, S., Goldman, D.B., Martin-Brualla, R., Seitz, S.M.: Hypernerf: a higher-dimensional representation for topologically varying neural radiance fields. ACM Trans. Graph. 40(6), 238:1-238:12 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.205, + 0.787, + 0.245 + ], + "angle": 0, + "content": "38. Peebles, W., Xie, S.: Scalable diffusion models with transformers. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 4195-4205 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.246, + 0.786, + 0.286 + ], + "angle": 0, + "content": "39. Peng, Y., Yan, Y., Liu, S., Cheng, Y., Guan, S., Pan, B., Zhai, G., Yang, X.: Cagenerf: Cage-based neural radiance fields for generalized 3d deformation and animation. In: Advances in Neural Information Processing Systems (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.287, + 0.787, + 0.328 + ], + "angle": 0, + "content": "40. Pumarola, A., Corona, E., Pons-Moll, G., Moreno-Noguer, F.: D-NeRF: Neural Radiance Fields for Dynamic Scenes. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.328, + 0.786, + 0.37 + ], + "angle": 0, + "content": "41. Richter, S.R., AlHaija, H.A., Koltun, V.: Enhancing photorealism enhancement. IEEE Transactions on Pattern Analysis and Machine Intelligence 45(2), 1700-1715 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.371, + 0.786, + 0.411 + ], + "angle": 0, + "content": "42. Rubinstein, M., Gutierrez, D., Sorkine, O., Shamir, A.: A comparative study of image retargeting. ACM Transactions on Graphics (Proc. SIGGRAPH ASIA) 29(6), 160:1-160:10 (2010)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.412, + 0.786, + 0.439 + ], + "angle": 0, + "content": "43. Rubinstein, M., Shamir, A., Avidan, S.: Improved seam carving for video retargeting. ACM Trans. Graph. 27(3), 16 (2008)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.44, + 0.786, + 0.481 + ], + "angle": 0, + "content": "44. Rubinstein, M., Shamir, A., Avidan, S.: Multi-operator media retargeting. ACM Trans. Graph. 28(3), 23 (2009). https://doi.org/10.1145/1531326.1531329, https://doi.org/10.1145/1531326.1531329" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.481, + 0.786, + 0.509 + ], + "angle": 0, + "content": "45. Shaham, T.R., Dekel, T., Michaeli, T.: Singan: Learning a generative model from a single natural image. In: ICCV. pp. 4569-4579 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.509, + 0.786, + 0.55 + ], + "angle": 0, + "content": "46. Song, E., Lee, M., Lee, S.: Carvingnet: Content-guided seam carving using deep convolution neural network. IEEE Access 7, 284-292 (2019). https://doi.org/10.1109/ACCESS.2018.2885347, https://doi.org/10.1109/ACCESS.2018.2885347" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.55, + 0.786, + 0.578 + ], + "angle": 0, + "content": "47. Sorkine-Hornung, O., Alexa, M.: As-rigid-as-possible surface modeling. In: Symposium on Geometry Processing (2007)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.578, + 0.786, + 0.605 + ], + "angle": 0, + "content": "48. Srinivas, S., Fleuret, F.: Full-gradient representation for neural network visualization. Advances in neural information processing systems 32 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.606, + 0.786, + 0.634 + ], + "angle": 0, + "content": "49. Tan, W., Yan, B., Lin, C., Niu, X.: Cycle-ir: Deep cyclic image retargeting. IEEE Transactions on Multimedia 22(7), 1730–1743 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.634, + 0.786, + 0.675 + ], + "angle": 0, + "content": "50. Wang, P., Liu, L., Liu, Y., Theobalt, C., Komura, T., Wang, W.: Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. In: Advances in Neural Information Processing Systems. pp. 27171-27183 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.675, + 0.786, + 0.716 + ], + "angle": 0, + "content": "51. Wei, L., Levoy, M.: Fast texture synthesis using tree-structured vector quantization. In: Proceedings of the 27th Annual Conference on Computer Graphics and Interactive Techniques, SIGGRAPH. pp. 479-488. ACM (2000)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.717, + 0.786, + 0.758 + ], + "angle": 0, + "content": "52. Wu, G., Yi, T., Fang, J., Xie, L., Zhang, X., Wei, W., Liu, W., Tian, Q., Xinggang, W.: 4d gaussian splatting for real-time dynamic scene rendering. arXiv preprint arXiv:2310.08528 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.759, + 0.786, + 0.786 + ], + "angle": 0, + "content": "53. Wu, H., Wang, Y., Feng, K., Wong, T., Lee, T., Heng, P.: Resizing by symmetrysummarization. ACM Trans. Graph. 29(6), 159 (2010)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.786, + 0.786, + 0.813 + ], + "angle": 0, + "content": "54. Wu, R., Liu, R., Vondrick, C., Zheng, C.: Sin3dm: Learning a diffusion model from a single 3d textured shape. CoRR abs/2305.15399 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.814, + 0.786, + 0.841 + ], + "angle": 0, + "content": "55. Wu, R., Zheng, C.: Learning to generate 3d shapes from a single example. ACM Transactions on Graphics (TOG) 41(6) (2022)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.444, + 0.129 + ], + "angle": 0, + "content": "T. Elsner, J. Berger et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.204 + ], + "angle": 0, + "content": "56. Xian, W., Sangkloy, P., Agrawal, V., Raj, A., Lu, J., Fang, C., Yu, F., Hays, J.: Texturegan: Controlling deep image synthesis with texture patches. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (June 2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.205, + 0.787, + 0.231 + ], + "angle": 0, + "content": "57. Xu, T., Harada, T.: Deforming radiance fields with cages. In: ECCV. pp. 159-175 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.232, + 0.787, + 0.272 + ], + "angle": 0, + "content": "58. Yang, Z., Gao, X., Zhou, W., Jiao, S., Zhang, Y., Jin, X.: Deformable 3d gaussians for high-fidelity monocular dynamic scene reconstruction. arXiv preprint arXiv:2309.13101 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.273, + 0.787, + 0.3 + ], + "angle": 0, + "content": "59. Yuan, Y.J., Sun, Y.T., Lai, Y.K., Ma, Y., Jia, R., Gao, L.: Nerf-editing: Geometry editing of neural radiance fields. In: CVPR. pp. 18332-18343 (2022)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.3 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/Retargeting Visual Data with Deformation Fields/547a2a3d-69aa-4867-932e-8316bae28b65_origin.pdf b/2024/Retargeting Visual Data with Deformation Fields/547a2a3d-69aa-4867-932e-8316bae28b65_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..be6805b8242b0b51b9f749469aee9906a703f3a8 --- /dev/null +++ b/2024/Retargeting Visual Data with Deformation Fields/547a2a3d-69aa-4867-932e-8316bae28b65_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:29efec3fb4ff0df066940a7eaee552ea223cdf353e33b0d14ba0aaccf792c079 +size 9677956 diff --git a/2024/Retargeting Visual Data with Deformation Fields/full.md b/2024/Retargeting Visual Data with Deformation Fields/full.md new file mode 100644 index 0000000000000000000000000000000000000000..5de7b98d32bdf9936bbcb57207cb4f6c1083defb --- /dev/null +++ b/2024/Retargeting Visual Data with Deformation Fields/full.md @@ -0,0 +1,351 @@ +# Retargeting Visual Data with Deformation Fields + +Tim Elsner1, Julia Berger1, Tong Wu2, Victor Czech1, Lin Gao2, and Leif Kobbelt1 + +$^{1}$ RWTH Aachen University, Germany +{elsner, czech, kobbelt}@cs.rwth-aachen.de / julia.berger@rwth-aachen.de + $^{2}$ University of Chinese Academy of Sciences, China +{wutong19s, gaolin}@ict.ac.cn + +Abstract. Seam carving is an image editing method that enables content-aware resizing, including operations like removing objects. However, the seam-finding strategy based on dynamic programming or graph-cut limits its applications to broader visual data formats and degrees of freedom for editing. Our observation is that describing the editing and retargeting of images more generally by a deformation field yields a generalisation of content-aware deformations. We propose to learn a deformation with a neural network that keeps the output plausible while trying to deform it only in places with low information content. This technique applies to different kinds of visual data, including images, 3D scenes given as neural radiance fields, or even polygon meshes. Experiments conducted on different visual data show that our method achieves better content-aware retargeting compared to previous methods. + +Keywords: Image Retargeting $\cdot$ Neural Fields $\cdot$ Neural Radiance Fields + +# 1 Introduction + +Media retargeting is a way to edit images, videos, 3D objects, or even entire 3D scenes by a global deformation such that relevant content, details, and features are properly preserved. The overall goal is to change the aspect ratio of the bounding box of the data so that it fits an allocated space, e.g. fitting an image to a display with prescribed format. Retargeting is based on the idea of identifying regions with little detail and to accumulate the necessary distortion induced by the deformation in these regions. Small editing operations are possible e.g. through concentrating the distortion there, specifying a region containing a particular object to have this object removed. + +Most existing retargeting methods typically use some form of seam carving, i.e. the deletion of discrete pixels on a path through the image (seams) [2,7,9,44]. As previous methods perform a discrete removal or addition of pixels, they do not natively extend to e.g. continuous domains like Neural Radiance Fields [30] (NeRFs) and are often restricted to a class of solutions defined by the used algorithm. + +Our approach formulates the problem via continuous deformation fields, mapping from the retargeted data back to the undistorted input. It natively extends + +![](images/0a0e5a4383146ff70bea9dc83b97ce4019abf8cf7d9553d9c68c07bb9972fdd6.jpg) +Fig.1: Different objectives applied to different visual domains with our approach: We demonstrate retargeting images, NeRFs, and meshes. 'Surfer' image from [42]. + +![](images/1ed9a438e62e49e798436fd982c2c3c38cffb96881319291972051c854262778.jpg) +Fig.2: The family of image resizing methods. Our approach formulates the problem in a more general way, but can utilise e.g. new energy formulations as well. + +from images on other domains, like NeRFs and polygon meshes (Fig. 1), and offers better solutions through a global and continuous instead of a discrete and greedy optimisation process (see Fig. 5). Our contributions are as follows: + +- We introduce the use of a neural deformation field to compress or stretch unimportant regions to achieve a smaller or larger output or to follow other editing objectives. +- We abstract and generalise the underlying rules of creating plausible outputs with seam carving, then demonstrate the domain agnostic nature of our formulation by applying it to images, 3D meshes, and 3D scenes in the form of NeRFs. +- We show that the high flexibility of our approach from optimising a global deformation field produces better outputs than iteratively computing seams, while also allowing configurations that are not possible with previous methods. + +To achieve this, we regularise the neural deformation fields to follow general sanity guidelines, while attempting to minimise distortion in places with presumed high information content. We evaluate our results on both qualitative and quantitative levels. To showcase our approach on interesting scenes captured as NeRFs and inspired by the dataset of Richter et al. [41], we also provide a small synthetic NeRF retargeting dataset captured in the video game GTA V [34] by flying the camera over the virtual city, capturing the screen content and camera parameters. + +Our method is a more reliable backbone for approaches following seam carving while outperforming current methods, requires only minor modifications for vastly different applications and domains, and is straightforward to implement. + +We provide the code of our method demonstrated on images in the supplemental. + +# 2 Related Work + +As our approach is re-formulating seam carving, it produces results from a single example without any learned prior, and is designed to allow seam carving-like deformations for (learned) 3D representations. We hence discuss these three branches of related work. + +Single Example Generation The goal of single example generation is to discover similar local patterns from a single input and synthesise novel samples. Texture synthesis on 2D images and 3D shapes has been widely studied in computer vision and computer graphics by non-neural methods [10, 11, 24, 51]. There are also methods focusing on synthesising geometric textures for 3D shapes [4, 19, 25], which transfer local geometry details from an example shape to other surfaces. Recently, neural-based methods have been proposed to generate 2D textures [14, 45, 56] and transfer them to 3D shapes [17, 18], resulting in improved results. With recent advances in implicit representation [6, 29, 35] and neural radiance fields [30], Wu et al. propose to learn implicit 3D shapes from a shape with [54] or without texture [55]. To enhance the realism of synthesised shapes, Huang et al. [21] reconstruct the 3D shape with a NeRF by optimising texture features on a reconstructed mesh. These features can be applied to arbitrary surfaces to render the target surface with synthesised textures. These methods often only work for a certain kind of data format, while our method proposes a domain agnostic approach to synthesise visual data from a single example. + +Deformations for Learned 3D Representations Neural radiance fields (NeRF), as a newly proposed 3D shape representation, has been widely used in scene reconstruction and novel view synthesis. However, the original NeRF only works for static scenes, thus is unable to deal with deformation in dynamic scenes or generate motions for a static scene. To overcome this, the deformation field is introduced to NeRF. NeRF-Editing [59] is the first to explore deformation in NeRF. It first reconstructs the explicit mesh of a static scene with NeuS [50] and deforms the explicit mesh with the ARAP algorithm [47]. Then sample points in volume rendering are deformed along with the mesh via barycentric coordinate interpolation, resulting in the change of rendered images. Follow-up NeRF deformation methods [13,39,57] employ a similar pipeline but use different geometry proxies. To reconstruct dynamic scenes, Albert et al. [40] first explore the possibility by adding an extra time-conditioned deformation field to the static NeRF. The time-conditioned deformation field transforms sample points in the observation space into the canonical space, where their colours and opacities are queried and then rendered into images. This idea stimulated a series of dynamic NeRF reconstruction methods [5,28,36,37,52,58]. While our approach also uses a deformation field, we inject general plausibility constraints into it to keep the + +results reasonable, enabling the idea behind seam carving on different types of visual data. + +Seam Carving Image/video retargeting is an important tool in computer vision and computer graphics. Early works [26] first annotate or detect the region of interest (ROI), which is uniformly scaled to the target size afterwards. While contents out of the ROI are deformed by fisheye warping, Liu and Gleicher [27] later extend it to the video domain by introducing a motion salience map. Avidan and Shamir [2] propose seam carving, a content-aware image resizing method. It first defines a saliency map by summing the gradient norms along the horizontal and vertical directions. The cost of a seam is defined by adding up the values on the salience map (energy). The seam with minimal energy can be found using dynamic programming and will be removed in the resizing process. Rubinstein et al. [43] extend the seam carving idea to the video domain and transform the dynamic programming problem to a graph cut problem. Follow-up works [7,9,44] incorporate homogeneous cropping and scaling with seam carving and excavate more high-level information like symmetry [53], depth [3], and object similarity [8] to assist seam carving. With the advances in deep learning, a few methods [31-33] made attempts to detect and localise the seam carving forgery. Song et al. [46] propose Carving-Net to replace the previous handcrafted energy map with a neural network while the seam finding algorithm is still dynamic programming. + +Generally, our method is invariant to the underlying energy term, and hence can be replaced by any other formulation like the ones mentioned before. This relationship is visualised in the taxonomy in Fig. 2: Our method solves a global, continuous optimisation problem instead of greedy and local steps formulated as dynamic programming, and can hence function as a more flexible backbone for approaches that derive from seam carving. This formulation allows more complex solutions that work better in many scenarios (see Fig. 5), while the space of all possible outputs contains the solutions of seam carving. + +# 3 Deformation Fields for Retargeting Visual Data + +Editing strategies for visual data require avoiding changes to meaningful content, that create visible artefacts. An approach that focuses on retargeting the input to match a target size can be guided by the following set of rules: + +- Content Aware Objective: Prevent the accidental creation of new or the removal of existing regions with high information content (control over where deformation happens). +- Sanity Objective: Avoid introducing deformations that yield implausible results (control over how deformation happens). + +In the following sections, we are only concerned with the retargeting to a smaller size (shrinking) for ease of explanation. The concept can naturally be + +![](images/a947aca5d6b8bf3d860a5b4ac1d458b5616dcd9782220e84de626ece55cded26.jpg) +Fig. 3: Our proposed pipeline for retargeting visual data: For a given input (left), we train two simple networks that learn the energy and cumulative energy along the deformation axis of the input (centre-left). We initialise a network that stretches samples to the desired position (centre-right), then optimise this deformation to distribute the distortion to low information content regions (right). Balloon image from [42]. + +reformulated to expand the visual data instead (expansion, see Sec. C (Appendix)) or to include editing (see Sec. 3.5). + +Seam carving [2] and its derivatives produce a discrete solution to this by performing edits through manipulating (removing or repeating) one seam of pixels at a time, which is a single-pixel-wide path passing through the image. Through dynamic programming, seam carving chooses this path based on the lowest energy of the pixels, with high energy indicating information-rich areas. In a simple case, this energy is the colour gradient. This yields a discrete solution to the retargeting problem that is both content aware and plausible. + +To obtain a continuous solution, instead of extracting discrete seams of the same width, we use continuous folds of varying width using a learned continuous deformation function. This formulation optimises a solution that is a superset of all existing seam carving solutions (see Fig. 5). We always use the same core formulation, regardless of the underlying domain and application, e.g. moving an object in an image or retargeting a 3D scene. + +# 3.1 General Formulation + +Our approach always optimises a deformation function that maps from points in the retargeted output space (e.g. an image at half the size) $I: P \to \mathbb{R}^n$ to points in the original data $I': P' \to \mathbb{R}^n$ (e.g. to $\mathbb{R}^3$ for RGB colours) by adding an offset to their coordinate. The value of a point $p$ in the output data $I$ is hence defined as the value of $p$ deformed to a new point $p'$ looked up in $I'$ . To represent the deformation function $D$ , we use a neural deformation field, a simple MLP where $D: P \to \mathbb{R}$ . To keep our approach simpler, we only allow scalar deformations along a fixed deformation direction $v$ , e.g. along the $x$ axis (width) of an image: + +$$ +I (p) = I ^ {\prime} (p + D (p) \cdot v) \tag {1} +$$ + +![](images/a4cf99d9231932bf11879bae955364a8cdaa82a896041aa1996818aa746a678d.jpg) +Fig. 4: Our pipeline applied to neural radiance fields: We first obtain a point cloud, then estimate the energy values. We then learn a continuous energy function and a cumulative energy function that we use to optimise our deformation network. The resulting deformation field is visualised on the bottom right. + +![](images/f19b37273e1b4f64146e5637c753a39b3d46c5daa026dc1a703e802c202106ca.jpg) +Fig. 5: Example of applying retargeting to $75\%$ width. Seam carving fails to use seams that are non-continuous or with a too steep angle, not properly using the empty space in the shelf to close the gaps before shrinking the bookshelves content. + +Hence, to get the content for a point $p$ in the retargeted data, we offset the coordinate and look up the content at the offset coordinate $p'$ in the original data. To fulfil the objectives stated in the beginning of Sec. 3, we introduce two sets of losses, content aware and sanity objectives: + +Content Aware Objectives apply regularisation to guide the deformation process when adjusting to the data content. In particular, we penalise deformation in regions with high information content to discourage noticeable changes. We define the energy function $E: P' \to \mathbb{R}$ to be a measure for content information in the original visual data, expresses as, e.g., a colour gradient: + +$$ +E \left(p ^ {\prime}\right) = \left| \left| \nabla I \left(p ^ {\prime}\right) \right| \right| _ {2} \tag {2} +$$ + +![](images/d7155f8351a4bc1be3916739063663ed51d396663c7fff263404983649e19838.jpg) +Fig. 6: Applying our retargeting to 3D meshes, moving furniture closer together to retarget the scene to a smaller size while limiting distortion to objects. + +![](images/479c7c964d87091d9e2ff5082052f7a9a85379dcf805c46dab759962506dcf37.jpg) + +![](images/e0527205f9d5202a85a40f2cb96af7191ab6e52fc07b84331541245d49debd39.jpg) + +The overall content deformation loss is the product of deformation magnitude at $p$ and energy at the original $p'$ over all $p$ in the deformed data: + +$$ +L _ {C} = \int_ {p \in P} [ E (p + v \cdot D (p)) \cdot | | \nabla D (p) | | _ {1} ] d p \tag {3} +$$ + +Thus, in Eq. (3), we effectively penalise visible deformations defined as the product of information content at points, namely $E(p + v \cdot D(p))$ , and the deformation itself via $||\nabla D(p)||_1$ . Using the $L^1$ -norm for the distortion gradient promotes piecewise constant, possibly sparse, deformations. We can then split up the change in deformation into change in deformation direction $v$ and change in direction $v^\perp$ orthogonal to $v$ : + +$$ +L _ {C} = \int_ {p \in P} \left[ E (p + v \cdot D (p)) \cdot \left(\left| \frac {\partial}{\partial v} D (p) \right| + \left| \frac {\partial}{\partial v ^ {\perp}} D (p) \right|\right) \right] d p \tag {4} +$$ + +This equation is then further split into two components, $L_{e}$ and $L_{s}$ . The loss $L_{e}$ penalises stretching or compression and $L_{s}$ punishes shearing, which we define to be deformation change in a direction that is not the deformation direction. For a now discrete sample set $P$ from our domain, these terms are expressed as: + +$$ +L _ {e} = \sum_ {p \in P} \left[ E (p + v \cdot D (p)) \cdot \left| \frac {\partial}{\partial v} D (p) \right| \right] \tag {5} +$$ + +$$ +L _ {s} = \sum_ {p \in P} \left[ E (p + v \cdot D (p)) \cdot \left| \frac {\partial}{\partial v ^ {\perp}} D (p) \right| \right] +$$ + +Essentially, $L_{e}$ directs deformations to areas of low information content, while $L_{s}$ discourages the resulting shearing to affect areas of high information content. Note that while $L_{s}$ would also penalise non-straight deformation seams, we tune our losses such that this only impacts the output much if an area is sheared. An example of this is shown in Fig. 3: The deformations around the balloons introduce shearing only at the seam, in an area of low information content. + +In case of image deformation, neighbouring pixels with low energy in the input may translate to pixels widely apart in the output, potentially bypassing a substantial amount of important content. Jumps in the deformation field $D$ indicate this effect. To prevent this, we penalise the amount of energy $\hat{E}$ between two points $p', q' \in P'$ in the original domain, rather than the energy at individual positions, as in Eq. (2): + +$$ +\hat {E} \left(p ^ {\prime}, q ^ {\prime}\right) = \int_ {r ^ {\prime} \in \left[ p ^ {\prime}, q ^ {\prime} \right]} E \left(r ^ {\prime}\right) d r ^ {\prime} \tag {6} +$$ + +Using $\hat{E}$ , we then redefine $L_{e}$ from Eq. (5) to consider energy and difference in deformation magnitude between $p$ and a slightly offset $p_{\varepsilon} \coloneqq p + \varepsilon$ in deformation direction: + +$$ +L _ {e} = \sum_ {p \in P} \hat {E} (p + v D (p), p _ {\varepsilon} + v D (p _ {\varepsilon})) \cdot \frac {| D (p) - D (p _ {\varepsilon}) |}{\varepsilon} \tag {7} +$$ + +This formulation guides the deformation to avoid introducing deformation to high information content regions or fold over them, while $L_{s}$ discourages non-local deformations that would apply shearing in high information areas. For example, shearing in blue sky is acceptable, shearing on a brick wall is not. + +Sanity objectives prevent undesired outputs by introducing two additional losses. Assume we shrink the input by a factor $\alpha \in [0,1]$ : + +1. The boundaries of the affected region should not change, e.g. an image should not be simply cropped. Based on $\alpha$ , the deformation field should map the start/end of the deformation axis in the output to the start/end of the input. For $P_0$ , $P_1$ containing all points at the start/end of the deformation axis of the output, we express this requirement as follows: + +$$ +L _ {b} = \int_ {p \in P _ {0}} | D (p) | d p + \int_ {p \in P _ {1}} | D (p) - (1 - \alpha) | d p \tag {8} +$$ + +2. We additionally demand monotonicity in the deformation field, expressed as + +$$ +L _ {m} = \int_ {p \in P} \max \left(0, - \frac {\partial}{\partial v} D (p)\right) d p, \tag {9} +$$ + +indicating that any change in the deformation direction should be positive. Therefore, deformation $D$ should be monotonic and only stagnate or increase exclusively along $v$ . This prevents the repetition of the same image portion, noticeable by jumps with inconsistent direction in the deformation field. + +We then obtain the total loss $L$ with hyperparameters $\lambda_{*}$ as + +$$ +L = \lambda_ {e} L _ {e} + \lambda_ {s} L _ {s} + \lambda_ {b} L _ {b} + \lambda_ {m} L _ {m}. \tag {10} +$$ + +We initialise the deformation function with a uniform stretch to the target size. + +In the following, we demonstrate the application of this general formulation to different domains, while adapting the energy formulation to the problem. All implementation and architecture details are given in Sec. A (Appendix) and the exact loss terms in Sec. B (Appendix). + +# 3.2 Optimising Deformation Fields for Images + +Instead of using image data directly, we train a continuous image representation MLP $I': P' \to \mathbb{R}^3$ to be our input data, that maps pixel positions to colours. This ensures consistency with our continuous approaches. + +We train one MLP to learn the energy field $E$ , holding the colour gradient per point as in Eq. (2). Using the learned $E$ , we train another MLP to replicate $\hat{E}$ as a cumulative energy field $\Sigma$ . The difference in cumulative energy $\Sigma$ between two points $p', q' \in P'$ on the same axis $v$ is the absolute difference in information content between them. As this coincides with the definition of $\hat{E}$ from Eq. (6), we approximate the measure of content relevance between two points as: + +$$ +\hat {E} \left(p ^ {\prime}, q ^ {\prime}\right) \approx | \Sigma \left(q ^ {\prime}\right) - \Sigma \left(p ^ {\prime}\right) | \tag {11} +$$ + +This cumulative gradient formulation therefore simplifies the interval over $E$ in Eq. (6) to a simple difference computation between two MLP queries: How much content is lost when the deformation skips a region. The energy field $E$ can be trained to reproduce any energy formulation other than a colour gradient. This exchangeable energy formulation ensures compatibility with other retargeting approaches as shown in Fig. 2. For all objectives, we use discrete approximations of the loss terms introduced in Sec. 3.1. The whole pipeline is visualised in Fig. 3, and additional results can be found in Sec. G (Appendix). + +# 3.3 Optimising Deformation Fields for Neural Radiance Fields + +We adapt our approach from Sec. 3.1 for 3D scenes given as trained NeRFs as follows (see Fig. 4): + +- A trained NeRF provides colour and depth images, from which we extract surface samples. Each point is assigned an energy value defined by change in colour and depth at this point. +- Based on the surface samples and their energy values, we train a continuous energy field $E$ and a continuous cumulative energy field $\Sigma$ , to easily access the amount of content between points. +- We use an inverse deformation network $U$ to find points in the target space that correspond to surface points in the source space, keeping computing costs low by restricting our optimisation to surface points. +- We apply our regularisation terms on a mix of surface and random points to optimise our loss function. + +1. Extracting a Sparse Set of Representatives and Energy We extract a surface point cloud from the RGB and depth images rendered by NeRF at the training camera positions. Only these points are relevant for scene deformation. We then compute the energy value for each point corresponding to a pixel as a sum of colour change in the RGB image and depth change in the depth image. To reduce the impact of highly noisy energy values, we apply smoothing by selecting the minimum energy for each point within a neighbourhood radius of 5 points. + +2. Defining an Energy Function To obtain a continuous energy field, we train an auxiliary energy network $E$ . The network maps all 3D points to their energy values, and clamps the energy to zero for points further than twice the average neighbour distance. The resulting energy field has zero energy in empty space and inside objects, and non-zero energy values for points near the surface. Similar to our approach with images, we learn a cumulative energy network $\Sigma$ . The network is trained to accumulate the energy values $E(p)$ from uniformly sampled points $p$ on a ray in deformation direction. We compute the energy for a segment between two points $p', q' \in \mathbb{R}^3$ along $v$ using their difference in cumulative energy $\Sigma(q')$ and $\Sigma(p')$ (cf. Eq. (11)). + +3. Inverting the Deformation We define an inverse deformation function to determine those points that lie on the surface after deformation. In contrast to images, we only perform optimisation on a sparse set of points, namely surface points. As we optimise a deformation from the target space to the source space, we need the position of surface points in both spaces. By training an MLP $U$ to reverse the deformation $D$ , i.e. minimising $(U(D(x)) = x)^2$ , we can effortlessly extract surface points, and hence their energy values, in the target space. This relation is also visualised in Fig. 4. We always update $U$ once after training one iteration of $D$ . + +As we use a discrete set of points, we apply the losses from Sec. 3.1 to Eq. (10) for a discrete interval. In particular, we average the loss over a random subset of sample points. Example images from deformed scenes can be found in Fig. 8, while a larger number of results can be found in Sec. G (Appendix). + +# 3.4 Optimising Deformation Fields for Meshes + +To demonstrate the versatility of our approach, we apply it to meshes as well. We adapt the energy term and use surface samples and vertex positions for optimisation. For freestanding objects, we use curvature as the energy measure. For the room example in Fig. 6, the energy is set to 0 for the floor and to 1 for every other surface point. We otherwise follow the same principle as in Sec. 3.3. + +# 3.5 Other Applications + +While retargeting is our main contribution, our framework opens up multiple different editing opportunities that even extend the capabilities of seam carving [2]. + +![](images/41087d1b823b0ef952f4422c8b1fdb6095b7086826307699a89f304cef065f23.jpg) +Iinui + +![](images/c877131d3cd9a0e3bad2108f760006a6cc9b37f2dbd32f4b5de504f22166e232.jpg) + +![](images/483d4940066c036181d0d5623a4f2622255bc9ae6707c2ae2a3fc6cf017b5f4d.jpg) +ed +Fig. 7: We demonstrate different editing operations by adding and adapting the loss formulation. Left: Removing, right: Moving an object. Images from [42]. + +![](images/5b2c071ea3cd1a71b60adbf886f91528f6a8888fda1662458f4ee72e4f392aa5.jpg) + +![](images/e9c01c195ebf0ff3921e10ccd9a4bcd79c855141d59ca30a154aa2088df2eef1.jpg) + +![](images/5b01bb21a0c5e53a64d87bb5e21ff28c8181e8ab46d966e7ee0e9315ce5281fd.jpg) +Fig.8: Applying our retargeting to NeRFs. Top row: Input, stretched, targeted with our method. Bottom row: Input, folds (white) on the input suggested by our approach, targeted with our method. + +![](images/2e1d792e0bc9fdfa6765039930e5928af706c34079ee4ce241a4fce66f0fa475.jpg) +Fig. 9: Selected results showing scenes where our approach benefits from a higher degree of flexibility. All images are arranged as input, seam carving, and our output. Images from [42]. + +![](images/f94ee2cb297addced415b0e223edb326bea5527110da4a42766497f224294b12.jpg) + +![](images/f9e9f691756184c008def46be8d17f37589b54b356ee47f8ba4ba9288f7f3ba5.jpg) + +![](images/a26a02d0dba06544237ee3e4b6a0acb74db8d799728e3591759f34e48455f53a.jpg) + +![](images/ab35c6c485ee0a5ce02ca2a7e913fbf90633bdd8c76f21feb900f6d3247472b0.jpg) + +![](images/84d4b09054b4f4798fce5a91b9366cce15e3afad001a2510ea62e0114886d3ca.jpg) + +![](images/fc2c2b9165fa633d16ee2bf35eba67da8a51a736e4e008598402869af648fb15.jpg) + +The loss formulation is slightly adapted to implement the following exemplary operations, which are applicable to any domain and can serve as the basis for additional processing methods or user-driven applications. Further details can be found in Sec. D (Appendix). Examples can be found in Fig. 7. We want to remove the monotonicity loss $L_{m}$ and use equal input and output sizes for both the removal and the movement of an object. + +Removal To remove a certain region, we employ an additional loss term that penalises any deformation target leading to that region. Points that were originally assigned to the target region will be distorted and assigned to neighbouring regions, resulting in the removal of the region in the output. + +Moving an Object Although classic seam carving cannot move an object easily, as seams are assumed to cross the entire image, our formulation does not have this restriction. To move an object, we also enforce that the deformation towards the target feature only occurs from the desired new placement in the output. Any occurrence of that feature elsewhere is punished to avoid its repetition. + +Table 1: Comparison of FID scores on the RetargetMe [42] dataset, for retargeting x (top) and y (bottom) axis to a smaller (left) and bigger (right) size. In mean, ours outperforms seam carving with 46.67 to 52.57 for x and 53.57 to 65.48 for y. + +
x FID(↓)50%60%70%80%90%110%120%130%140%150%
Seam carving85.7566.4151.6735.2023.8020.8630.0137.3843.4546.37
Ours79.3958.0643.6930.5321.7111.1719.4229.1937.8749.74
y FID(↓)50%60%70%80%90%110%120%130%140%150%
Seam carving116.7084.3360.5940.8624.9223.0433.5840.8248.4952.59
Ours89.4467.7749.6235.5925.4213.5521.0227.5538.2046.40
+ +Table 2: User study comparing our approach to seam carving $(SC)$ for images. 19 participants were questioned for preferences of random examples from the RetargetMe dataset [42], casting 1200 votes. + +
WidthOursSCDraw
50 %48.48%20.87%30.67%
150 %33.8%36.27%29.92%
HeightOursSCDraw
50 %49.5%24.25%26.25%
150 %45.84%20.0%34.15%
Total44.57%25.1%30.32%
+ +![](images/e8604be8cf7942901512735e38ef12f1dcf5088957484b76f42deedc7e478f62.jpg) +Colour Complex Gradient Energy +Fig. 10: Example of using our method with the more complex energy term of [48] instead of gradients, improving the results on an example. + +# 4 Evaluation + +Retargeting not only lacks comparable methods for other visual data such as meshes, but its quality is also difficult to measure. In 2D, we compare our results to seam carving [2] on images using the RetargetMe [42] benchmark. In 3D, we compare our results on NeRFs with video-seam-carved recordings of the same scene, because video seam carving [43] is the closest method to 3D retargeting. We always produce a short, 3 second clip at 30 frames per second. To ensure a fair comparison with the method, we only tilt the camera down instead of rotating it around the object. Even for benevolent camera movements, it produces visible artefacts, as seen in Sec. F (Appendix). Exemplary, we show that our approach is compatible with approaches that propose different saliency maps as visualised in Fig. 2. For that purpose, we compare our method to other more recent image retargeting approaches visually (Fig. 10) and in a user study (Tab. 4). Additional results can be found in Sec. G (Appendix), including an explicit failure case. + +# 4.1 Comparisons + +Images In accordance with recent publications in the generative model and image editing domain [15, 23, 38], we extensively compare our results with the + +Table 3: User study comparing retargeting a scene, then recording it (ours), to recording it and then retargetet it (video seam carving [43], VSC), using our provided dataset. 21 participants were asked to indicated which solution they prefer, casting 252 votes in total. + +
SceneOursVSCDrawOursVSCDrawOursVSCDrawOursVSCDraw
Beach1731201019111911
Farm2100210016322100
Harb.2100210021002100
+ +original seam carving results using FID [20] in Tab. 1 for images. We also evaluate our approach in a user study. The participants were instructed to classify their preference for a retargeted version of a randomly selected image based on a randomly selected deformation axis, always choosing between our and the seam carved [2] version. We used the 80 images of the RetargetMe [42] for 2D, producing 360 variations per approach in Tab. 2. This shows how our approach can be a more powerful backbone for seam carving. + +Table 4: User study measuring preference, pairwise comparing to our approach using salience maps of [48]. + +
ApproachOursTheirsDraw
CarvingNet [46]45.16%41.93%12.91%
RL [22]46.85%32.43%20.72%
Cycle-IR [49]40.3%34.46%25.24%
+ +We further compare against other state-of-the-art retargeting approaches in a user study in Tab. 4: We show that our method can compete both against a seam carving variation with better saliency from deep learning [46] while also beating state-of-the-art approaches that retarget the image directly [22,49]. Note that we only compare to readily available images from other approach, as + +we often could not reproduce them or run the code. For this comparison, we use the saliency method FullGrad [48] that aggregates layer-wise bias-gradients from across ResNet [16]. This aggregated gradient map acts as our saliency map, as it measures how useful the feature is to the neural network, hence we use it as a measure for how semantically important a feature is. This also shows that our approach can work with different saliency maps. Details and examples can be found in Sec. E (Appendix). Additionally, we provide a visual ablation that shows the importance of each loss component in Fig. 11. + +3D Scenes When assessing the quality of retargeting 3D scenes provided as NeRFs, we compared our results with what we consider to be the most similar approach, video seam carving [43]. In our provided GTA V NeRF retargeting dataset, we resize each axis by $50\%$ and $150\%$ . The results can be found in Tab. 3. + +![](images/abfe2e99a4bc32fdc43ad0b6544da1d1f2e6d92af5130c2241f2a173e48df8c2.jpg) +Fig. 11: Visual ablation for retargeting to $80\%$ width on an image [12], from left to right: Wobbly contours from shearing (no $L_{s}$ ), uniform stretch on the whole image (no $L_{e}$ ), missing features e.g. the person (no $L_{m}$ ), content outside the image (no $L_{b}$ ). + +We provide additional results and frames of the resulting sequences from our approach for 3D scenes compared to video seam carving in Sec. F (Appendix). + +# 4.2 Discussion and Limitations + +With our deformation fields as backbone, but the same energy term, our approach outperforms seam carving on both a qualitative and quantitative level. Our flexibility in the energy formulation allows for more possibilities, and the replacement of the gradient-based energy formulation with a better saliency map improves our approach even more. This shows that our method is a flexible backbone that can be used with various successors to seam carving, which provide more advanced saliency maps. Our continuous formulation allows for retargeting in various domains, including 3D scenes, and highlights the advantages of a geometry-aware approach. Applying our approach with only colour gradients can produce flawed outputs, just as those of seam carving, see Fig. 15 (Appendix). As we optimise a neural field for each individual input, we trade computation time for not needing a dataset to train and for a global optimisation. Our deformation optimisation takes approximately one minute on consumer hardware for good results versus about half a second for seam carving. Our strength, however, lies in 3D. Video seam carving on a 600 by 400 pixel video can take several hours, whereas our approach produces the final retargeted 3D scene in under 15 minutes. For Future Work, investigating invertible neural networks to implicitly learn an inverse deformation for the 3D case [1] seems promising, as would a deep learned approach to evaluating a deformation for optimisation and more research into good measures for retargeting quality. + +# 5 Conclusion + +We present an approach that generalises the principle underlying seam carving to create image variants by manipulating low content areas. By training neural deformation fields that avoid applying distortion to important parts of an input, our approach is applicable to different forms of visual data retargeting. Our formulation is largely domain invariant and only needs minor alterations to work on different types of visual data, as demonstrated for neural radiance fields and meshes for the first time. We demonstrate compatibility to other energy formulations of the seam carving family, being a more general way of optimising image retargeting. Our approach is the only one that natively works on 3D scenes. + +# References + +1. Ardizzone, L., Kruse, J., Wirkert, S.J., Rahner, D., Pellegrini, E.W., Klessen, R.S., Maier-Hein, L., Rother, C., Kothe, U.: Analyzing inverse problems with invertible neural networks. CoRR abs/1808.04730 (2018), http://arxiv.org/abs/1808.04730 +2. Avidan, S., Shamir, A.: Seam carving for content-aware image resizing. ACM Trans. Graph. 26(3), 10 (2007) +3. Basha, T.D., Moses, Y., Avidan, S.: Stereo seam carving a geometrically consistent approach. IEEE Trans. Pattern Anal. Mach. Intell. 35(10), 2513-2525 (2013) +4. Berkiten, S., Halber, M., Solomon, J., Ma, C., Li, H., Rusinkiewicz, S.: Learning detail transfer based on geometric features. Comput. Graph. Forum 36(2), 361-373 (2017) +5. Cai, H., Feng, W., Feng, X., Wang, Y., Zhang, J.: Neural surface reconstruction of dynamic scenes with monocular rgb-d camera. In: Neural Information Processing Systems (NeurIPS) (2022) +6. Chen, Zhiqin and Zhang, Hao: Learning implicit fields for generative shape modeling. In: CVPR. pp. 5939-5948 (2019) +7. Dong, W., Bao, G., Zhang, X., Paul, J.: Fast multi-operator image resizing and evaluation. J. Comput. Sci. Technol. 27(1), 121-134 (2012) +8. Dong, W., Zhou, N., Lee, T., Wu, F., Kong, Y., Zhang, X.: Summarization-based image resizing by intelligent object carving. IEEE Trans. Vis. Comput. Graph. 20(1), 1 (2014). https://doi.org/10.1109/TVCG.2013.103, https://doi.org/10.1109/TVCG.2013.103 +9. Dong, W., Zhou, N., Paul, J., Zhang, X.: Optimized image resizing using seam carving and scaling. ACM Trans. Graph. 28(5), 125 (2009) +0. Efros, A.A., Freeman, W.T.: Image quilting for texture synthesis and transfer. In: Pocock, L. (ed.) Proceedings of the 28th Annual Conference on Computer Graphics and Interactive Techniques, SIGGRAPH. pp. 341-346. ACM (2001) +1. Efros, A.A., Leung, T.K.: Texture synthesis by non-parametric sampling. In: ICCV. pp. 1033-1038 (1999) +2. at English Wikipedia, N.: Broadway tower (2007), https://commons.wikimedia.org/wiki/File:Broadway_tower.jpg +3. Garbin, S.J., Kowalski, M., Estellers, V., Szymanowicz, S., Rezaeifar, S., Shen, J., Johnson, M., Valentin, J.: Voltemorph: Realtime, controllable and generalisable animation of volumetric representations. arXiv:2208.00949 (2022) +4. Gatys, L.A., Ecker, A.S., Bethge, M.: Texture synthesis using convolutional neural networks. In: Neural Information Processing Systems. pp. 262-270 (2015) +5. Gu, J., Zhai, S., Zhang, Y., Susskind, J., Jaitly, N.: Matryoshka diffusion models. arXiv preprint arXiv:2310.15111 (2023) +6. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. CoRR abs/1512.03385 (2015), http://arxiv.org/abs/1512.03385 +7. Henzler, P., Deschaintre, V., Mitra, N.J., Ritschel, T.: Generative modelling of BRDF textures from flash images. ACM Trans. Graph. 40(6), 284:1-284:13 (2021) +8. Henzler, P., Mitra, N.J., Ritschel, T.: Learning a neural 3d texture space from 2d exemplars. In: The IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2019) +9. Hertz, A., Hanocka, R., Giryes, R., Cohen-Or, D.: Deep geometric texture synthesis. ACM Trans. Graph. 39(4), 108:1-108:11 (2020) + +20. Heusel, M., Ramsauer, H., Unterthiner, T., Nessler, B., Hochreiter, S.: Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems 30 (2017) +21. Huang, Y., Cao, Y., Lai, Y., Shan, Y., Gao, L.: Nerf-texture: Texture synthesis with neural radiance fields. In: ACM SIGGRAPH 2023 Conference Proceedings. pp. 43:1-43:10 (2023) +22. Kajiura, N., Kosugi, S., Wang, X., Yamasaki, T.: Self-play reinforcement learning for fast image retargeting. CoRR abs/2010.00909 (2020), https://arxiv.org/abs/2010.00909 +23. Kawar, B., Zada, S., Lang, O., Tov, O., Chang, H., Dekel, T., Mosseri, I., Irani, M.: Imagic: Text-based real image editing with diffusion models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6007-6017 (2023) +24. Kopf, J., Fu, C.W., Cohen-Or, D., Deussen, O., Lischinski, D., Wong, T.T.: Solid texture synthesis from 2d exemplars. ACM Transactions on Graphics 26(3), 2:1-2:9 (2007) +25. Lai, Y., Hu, S., Gu, X., Martin, R.R.: Geometric texture synthesis and transfer via geometry images. In: Proceedings of the Tenth ACM Symposium on Solid and Physical Modeling. pp. 15-26. ACM (2005) +26. Liu, F., Gleicher, M.: Automatic image retargeting with fisheye-view warping. In: Proceedings of the 18th Annual ACM Symposium on User Interface Software and Technology. pp. 153-162. ACM (2005) +27. Liu, F., Gleicher, M.: Video retargeting: automating pan and scan. In: Nahrstedt, K., Turk, M.A., Rui, Y., Klas, W., Mayer-Patel, K. (eds.) Proceedings of the 14th ACM International Conference on Multimedia. pp. 241-250. ACM (2006) +28. Luiten, J., Kopanas, G., Leibe, B., Ramanan, D.: Dynamic 3d gaussians: Tracking by persistent dynamic view synthesis. In: 3DV (2024) +29. Mescheder, L., Oechsle, M., Niemeyer, M., Nowozin, S., Geiger, A.: Occupancy networks: Learning 3D reconstruction in function space. In: CVPR. pp. 4460-4470 (2019) +30. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. In: ECCV (2020) +31. Moreira, T.P., Santana, M.C.S., Passos, L.A., Papa, J.P., da Costa, K.A.P.: An end-to-end approach for seam carving detection using deep neural networks. In: Pattern Recognition and Image Analysis. Lecture Notes in Computer Science, vol. 13256, pp. 447-457 (2022) +32. Nam, S., Ahn, W., Yu, I., Kwon, M., Son, M., Lee, H.: Deep convolutional neural network for identifying seam-carving forgery. IEEE Trans. Circuits Syst. Video Technol. 31(8), 3308-3326 (2021) +33. Nataraj, L., Gudavalli, C., Mohammed, T.M., Chandrasekaran, S., Manjunath, B.S.: Seam carving detection and localization using two-stage deep neural networks. CoRR abs/2109.01764 (2021) +34. North, R.: Grand theft auto v. Steam (2015), https://www.rockstargames.com/V/ +35. Park, J.J., Florence, P., Straub, J., Newcombe, R., Lovegrove, S.: DeepSDF: Learning continuous signed distance functions for shape representation. In: CVPR. pp. 165-174 (2019) +36. Park, K., Sinha, U., Barron, J.T., Bouaziz, S., Goldman, D.B., Seitz, S.M., MartinBrualla, R.: Nerfies: Deformable neural radiance fields. In: ICCV. pp. 5845-5854 (2021) + +37. Park, K., Sinha, U., Hedman, P., Barron, J.T., Bouaziz, S., Goldman, D.B., Martin-Brualla, R., Seitz, S.M.: Hypernerf: a higher-dimensional representation for topologically varying neural radiance fields. ACM Trans. Graph. 40(6), 238:1-238:12 (2021) +38. Peebles, W., Xie, S.: Scalable diffusion models with transformers. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 4195-4205 (2023) +39. Peng, Y., Yan, Y., Liu, S., Cheng, Y., Guan, S., Pan, B., Zhai, G., Yang, X.: Cagenerf: Cage-based neural radiance fields for generalized 3d deformation and animation. In: Advances in Neural Information Processing Systems (2022) +40. Pumarola, A., Corona, E., Pons-Moll, G., Moreno-Noguer, F.: D-NeRF: Neural Radiance Fields for Dynamic Scenes. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2021) +41. Richter, S.R., AlHaija, H.A., Koltun, V.: Enhancing photorealism enhancement. IEEE Transactions on Pattern Analysis and Machine Intelligence 45(2), 1700-1715 (2022) +42. Rubinstein, M., Gutierrez, D., Sorkine, O., Shamir, A.: A comparative study of image retargeting. ACM Transactions on Graphics (Proc. SIGGRAPH ASIA) 29(6), 160:1-160:10 (2010) +43. Rubinstein, M., Shamir, A., Avidan, S.: Improved seam carving for video retargeting. ACM Trans. Graph. 27(3), 16 (2008) +44. Rubinstein, M., Shamir, A., Avidan, S.: Multi-operator media retargeting. ACM Trans. Graph. 28(3), 23 (2009). https://doi.org/10.1145/1531326.1531329, https://doi.org/10.1145/1531326.1531329 +45. Shaham, T.R., Dekel, T., Michaeli, T.: Singan: Learning a generative model from a single natural image. In: ICCV. pp. 4569-4579 (2019) +46. Song, E., Lee, M., Lee, S.: Carvingnet: Content-guided seam carving using deep convolution neural network. IEEE Access 7, 284-292 (2019). https://doi.org/10.1109/ACCESS.2018.2885347, https://doi.org/10.1109/ACCESS.2018.2885347 +47. Sorkine-Hornung, O., Alexa, M.: As-rigid-as-possible surface modeling. In: Symposium on Geometry Processing (2007) +48. Srinivas, S., Fleuret, F.: Full-gradient representation for neural network visualization. Advances in neural information processing systems 32 (2019) +49. Tan, W., Yan, B., Lin, C., Niu, X.: Cycle-ir: Deep cyclic image retargeting. IEEE Transactions on Multimedia 22(7), 1730–1743 (2019) +50. Wang, P., Liu, L., Liu, Y., Theobalt, C., Komura, T., Wang, W.: Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. In: Advances in Neural Information Processing Systems. pp. 27171-27183 (2021) +51. Wei, L., Levoy, M.: Fast texture synthesis using tree-structured vector quantization. In: Proceedings of the 27th Annual Conference on Computer Graphics and Interactive Techniques, SIGGRAPH. pp. 479-488. ACM (2000) +52. Wu, G., Yi, T., Fang, J., Xie, L., Zhang, X., Wei, W., Liu, W., Tian, Q., Xinggang, W.: 4d gaussian splatting for real-time dynamic scene rendering. arXiv preprint arXiv:2310.08528 (2023) +53. Wu, H., Wang, Y., Feng, K., Wong, T., Lee, T., Heng, P.: Resizing by symmetrysummarization. ACM Trans. Graph. 29(6), 159 (2010) +54. Wu, R., Liu, R., Vondrick, C., Zheng, C.: Sin3dm: Learning a diffusion model from a single 3d textured shape. CoRR abs/2305.15399 (2023) +55. Wu, R., Zheng, C.: Learning to generate 3d shapes from a single example. ACM Transactions on Graphics (TOG) 41(6) (2022) + +56. Xian, W., Sangkloy, P., Agrawal, V., Raj, A., Lu, J., Fang, C., Yu, F., Hays, J.: Texturegan: Controlling deep image synthesis with texture patches. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (June 2018) +57. Xu, T., Harada, T.: Deforming radiance fields with cages. In: ECCV. pp. 159-175 (2022) +58. Yang, Z., Gao, X., Zhou, W., Jiao, S., Zhang, Y., Jin, X.: Deformable 3d gaussians for high-fidelity monocular dynamic scene reconstruction. arXiv preprint arXiv:2309.13101 (2023) +59. Yuan, Y.J., Sun, Y.T., Lai, Y.K., Ma, Y., Jia, R., Gao, L.: Nerf-editing: Geometry editing of neural radiance fields. In: CVPR. pp. 18332-18343 (2022) \ No newline at end of file diff --git a/2024/Retargeting Visual Data with Deformation Fields/images.zip b/2024/Retargeting Visual Data with Deformation Fields/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..ed331aacec18a8d94a177739dee57c827db5a2e4 --- /dev/null +++ b/2024/Retargeting Visual Data with Deformation Fields/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ee40136ee676c910696b7a0d09b72a57dab40ef1d1fd7c708fecd235665bf872 +size 590211 diff --git a/2024/Retargeting Visual Data with Deformation Fields/layout.json b/2024/Retargeting Visual Data with Deformation Fields/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..13a733df0b711e4b68c2e823372e31d1976365b3 --- /dev/null +++ b/2024/Retargeting Visual Data with Deformation Fields/layout.json @@ -0,0 +1,10133 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 134, + 111, + 479, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 111, + 479, + 129 + ], + "spans": [ + { + "bbox": [ + 134, + 111, + 479, + 129 + ], + "type": "text", + "content": "Retargeting Visual Data with Deformation Fields" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 140, + 149, + 474, + 174 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 140, + 149, + 474, + 174 + ], + "spans": [ + { + "bbox": [ + 140, + 149, + 474, + 174 + ], + "type": "text", + "content": "Tim Elsner1, Julia Berger1, Tong Wu2, Victor Czech1, Lin Gao2, and Leif Kobbelt1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 140, + 182, + 472, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 140, + 182, + 472, + 228 + ], + "spans": [ + { + "bbox": [ + 140, + 182, + 472, + 228 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 140, + 182, + 472, + 228 + ], + "type": "text", + "content": " RWTH Aachen University, Germany \n{elsner, czech, kobbelt}@cs.rwth-aachen.de / julia.berger@rwth-aachen.de \n" + }, + { + "bbox": [ + 140, + 182, + 472, + 228 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 140, + 182, + 472, + 228 + ], + "type": "text", + "content": " University of Chinese Academy of Sciences, China \n{wutong19s, gaolin}@ict.ac.cn" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 160, + 255, + 457, + 397 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 255, + 457, + 397 + ], + "spans": [ + { + "bbox": [ + 160, + 255, + 457, + 397 + ], + "type": "text", + "content": "Abstract. Seam carving is an image editing method that enables content-aware resizing, including operations like removing objects. However, the seam-finding strategy based on dynamic programming or graph-cut limits its applications to broader visual data formats and degrees of freedom for editing. Our observation is that describing the editing and retargeting of images more generally by a deformation field yields a generalisation of content-aware deformations. We propose to learn a deformation with a neural network that keeps the output plausible while trying to deform it only in places with low information content. This technique applies to different kinds of visual data, including images, 3D scenes given as neural radiance fields, or even polygon meshes. Experiments conducted on different visual data show that our method achieves better content-aware retargeting compared to previous methods." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 160, + 407, + 453, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 407, + 453, + 418 + ], + "spans": [ + { + "bbox": [ + 160, + 407, + 453, + 418 + ], + "type": "text", + "content": "Keywords: Image Retargeting " + }, + { + "bbox": [ + 160, + 407, + 453, + 418 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 407, + 453, + 418 + ], + "type": "text", + "content": " Neural Fields " + }, + { + "bbox": [ + 160, + 407, + 453, + 418 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 407, + 453, + 418 + ], + "type": "text", + "content": " Neural Radiance Fields" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 437, + 230, + 450 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 437, + 230, + 450 + ], + "spans": [ + { + "bbox": [ + 132, + 437, + 230, + 450 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 462, + 482, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 462, + 482, + 569 + ], + "spans": [ + { + "bbox": [ + 130, + 462, + 482, + 569 + ], + "type": "text", + "content": "Media retargeting is a way to edit images, videos, 3D objects, or even entire 3D scenes by a global deformation such that relevant content, details, and features are properly preserved. The overall goal is to change the aspect ratio of the bounding box of the data so that it fits an allocated space, e.g. fitting an image to a display with prescribed format. Retargeting is based on the idea of identifying regions with little detail and to accumulate the necessary distortion induced by the deformation in these regions. Small editing operations are possible e.g. through concentrating the distortion there, specifying a region containing a particular object to have this object removed." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 570, + 482, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 640 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 640 + ], + "type": "text", + "content": "Most existing retargeting methods typically use some form of seam carving, i.e. the deletion of discrete pixels on a path through the image (seams) [2,7,9,44]. As previous methods perform a discrete removal or addition of pixels, they do not natively extend to e.g. continuous domains like Neural Radiance Fields [30] (NeRFs) and are often restricted to a class of solutions defined by the used algorithm." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "content": "Our approach formulates the problem via continuous deformation fields, mapping from the retargeted data back to the undistorted input. It natively extends" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 138, + 120, + 315, + 270 + ], + "blocks": [ + { + "bbox": [ + 138, + 120, + 315, + 270 + ], + "lines": [ + { + "bbox": [ + 138, + 120, + 315, + 270 + ], + "spans": [ + { + "bbox": [ + 138, + 120, + 315, + 270 + ], + "type": "image", + "image_path": "0a0e5a4383146ff70bea9dc83b97ce4019abf8cf7d9553d9c68c07bb9972fdd6.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 286, + 326, + 331 + ], + "lines": [ + { + "bbox": [ + 131, + 286, + 326, + 331 + ], + "spans": [ + { + "bbox": [ + 131, + 286, + 326, + 331 + ], + "type": "text", + "content": "Fig.1: Different objectives applied to different visual domains with our approach: We demonstrate retargeting images, NeRFs, and meshes. 'Surfer' image from [42]." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 357, + 118, + 494, + 251 + ], + "blocks": [ + { + "bbox": [ + 357, + 118, + 494, + 251 + ], + "lines": [ + { + "bbox": [ + 357, + 118, + 494, + 251 + ], + "spans": [ + { + "bbox": [ + 357, + 118, + 494, + 251 + ], + "type": "image", + "image_path": "1ed9a438e62e49e798436fd982c2c3c38cffb96881319291972051c854262778.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 356, + 259, + 482, + 324 + ], + "lines": [ + { + "bbox": [ + 356, + 259, + 482, + 324 + ], + "spans": [ + { + "bbox": [ + 356, + 259, + 482, + 324 + ], + "type": "text", + "content": "Fig.2: The family of image resizing methods. Our approach formulates the problem in a more general way, but can utilise e.g. new energy formulations as well." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 354, + 480, + 390 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 354, + 480, + 390 + ], + "spans": [ + { + "bbox": [ + 130, + 354, + 480, + 390 + ], + "type": "text", + "content": "from images on other domains, like NeRFs and polygon meshes (Fig. 1), and offers better solutions through a global and continuous instead of a discrete and greedy optimisation process (see Fig. 5). Our contributions are as follows:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 137, + 396, + 480, + 523 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 138, + 396, + 480, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 396, + 480, + 430 + ], + "spans": [ + { + "bbox": [ + 138, + 396, + 480, + 430 + ], + "type": "text", + "content": "- We introduce the use of a neural deformation field to compress or stretch unimportant regions to achieve a smaller or larger output or to follow other editing objectives." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 431, + 480, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 431, + 480, + 475 + ], + "spans": [ + { + "bbox": [ + 138, + 431, + 480, + 475 + ], + "type": "text", + "content": "- We abstract and generalise the underlying rules of creating plausible outputs with seam carving, then demonstrate the domain agnostic nature of our formulation by applying it to images, 3D meshes, and 3D scenes in the form of NeRFs." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 137, + 479, + 480, + 523 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 479, + 480, + 523 + ], + "spans": [ + { + "bbox": [ + 137, + 479, + 480, + 523 + ], + "type": "text", + "content": "- We show that the high flexibility of our approach from optimising a global deformation field produces better outputs than iteratively computing seams, while also allowing configurations that are not possible with previous methods." + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 130, + 533, + 482, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 482, + 628 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 482, + 628 + ], + "type": "text", + "content": "To achieve this, we regularise the neural deformation fields to follow general sanity guidelines, while attempting to minimise distortion in places with presumed high information content. We evaluate our results on both qualitative and quantitative levels. To showcase our approach on interesting scenes captured as NeRFs and inspired by the dataset of Richter et al. [41], we also provide a small synthetic NeRF retargeting dataset captured in the video game GTA V [34] by flying the camera over the virtual city, capturing the screen content and camera parameters." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 630, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 630, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 630, + 481, + 665 + ], + "type": "text", + "content": "Our method is a more reliable backbone for approaches following seam carving while outperforming current methods, requires only minor modifications for vastly different applications and domains, and is straightforward to implement." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 271, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 271, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 271, + 102 + ], + "type": "text", + "content": "T. Elsner, J. Berger et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 116, + 479, + 128 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 479, + 128 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 479, + 128 + ], + "type": "text", + "content": "We provide the code of our method demonstrated on images in the supplemental." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 159, + 236, + 171 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 159, + 236, + 171 + ], + "spans": [ + { + "bbox": [ + 133, + 159, + 236, + 171 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 185, + 480, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 185, + 480, + 232 + ], + "spans": [ + { + "bbox": [ + 133, + 185, + 480, + 232 + ], + "type": "text", + "content": "As our approach is re-formulating seam carving, it produces results from a single example without any learned prior, and is designed to allow seam carving-like deformations for (learned) 3D representations. We hence discuss these three branches of related work." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 246, + 481, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 246, + 481, + 437 + ], + "spans": [ + { + "bbox": [ + 133, + 246, + 481, + 437 + ], + "type": "text", + "content": "Single Example Generation The goal of single example generation is to discover similar local patterns from a single input and synthesise novel samples. Texture synthesis on 2D images and 3D shapes has been widely studied in computer vision and computer graphics by non-neural methods [10, 11, 24, 51]. There are also methods focusing on synthesising geometric textures for 3D shapes [4, 19, 25], which transfer local geometry details from an example shape to other surfaces. Recently, neural-based methods have been proposed to generate 2D textures [14, 45, 56] and transfer them to 3D shapes [17, 18], resulting in improved results. With recent advances in implicit representation [6, 29, 35] and neural radiance fields [30], Wu et al. propose to learn implicit 3D shapes from a shape with [54] or without texture [55]. To enhance the realism of synthesised shapes, Huang et al. [21] reconstruct the 3D shape with a NeRF by optimising texture features on a reconstructed mesh. These features can be applied to arbitrary surfaces to render the target surface with synthesised textures. These methods often only work for a certain kind of data format, while our method proposes a domain agnostic approach to synthesise visual data from a single example." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 133, + 449, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 449, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 449, + 481, + 665 + ], + "type": "text", + "content": "Deformations for Learned 3D Representations Neural radiance fields (NeRF), as a newly proposed 3D shape representation, has been widely used in scene reconstruction and novel view synthesis. However, the original NeRF only works for static scenes, thus is unable to deal with deformation in dynamic scenes or generate motions for a static scene. To overcome this, the deformation field is introduced to NeRF. NeRF-Editing [59] is the first to explore deformation in NeRF. It first reconstructs the explicit mesh of a static scene with NeuS [50] and deforms the explicit mesh with the ARAP algorithm [47]. Then sample points in volume rendering are deformed along with the mesh via barycentric coordinate interpolation, resulting in the change of rendered images. Follow-up NeRF deformation methods [13,39,57] employ a similar pipeline but use different geometry proxies. To reconstruct dynamic scenes, Albert et al. [40] first explore the possibility by adding an extra time-conditioned deformation field to the static NeRF. The time-conditioned deformation field transforms sample points in the observation space into the canonical space, where their colours and opacities are queried and then rendered into images. This idea stimulated a series of dynamic NeRF reconstruction methods [5,28,36,37,52,58]. While our approach also uses a deformation field, we inject general plausibility constraints into it to keep the" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 246, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 246, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 246, + 91, + 447, + 102 + ], + "type": "text", + "content": "Retargeting Visual Data with Deformation Fields" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 481, + 139 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 481, + 139 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 481, + 139 + ], + "type": "text", + "content": "results reasonable, enabling the idea behind seam carving on different types of visual data." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 152, + 482, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 152, + 482, + 380 + ], + "spans": [ + { + "bbox": [ + 130, + 152, + 482, + 380 + ], + "type": "text", + "content": "Seam Carving Image/video retargeting is an important tool in computer vision and computer graphics. Early works [26] first annotate or detect the region of interest (ROI), which is uniformly scaled to the target size afterwards. While contents out of the ROI are deformed by fisheye warping, Liu and Gleicher [27] later extend it to the video domain by introducing a motion salience map. Avidan and Shamir [2] propose seam carving, a content-aware image resizing method. It first defines a saliency map by summing the gradient norms along the horizontal and vertical directions. The cost of a seam is defined by adding up the values on the salience map (energy). The seam with minimal energy can be found using dynamic programming and will be removed in the resizing process. Rubinstein et al. [43] extend the seam carving idea to the video domain and transform the dynamic programming problem to a graph cut problem. Follow-up works [7,9,44] incorporate homogeneous cropping and scaling with seam carving and excavate more high-level information like symmetry [53], depth [3], and object similarity [8] to assist seam carving. With the advances in deep learning, a few methods [31-33] made attempts to detect and localise the seam carving forgery. Song et al. [46] propose Carving-Net to replace the previous handcrafted energy map with a neural network while the seam finding algorithm is still dynamic programming." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 391, + 482, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 391, + 482, + 486 + ], + "spans": [ + { + "bbox": [ + 130, + 391, + 482, + 486 + ], + "type": "text", + "content": "Generally, our method is invariant to the underlying energy term, and hence can be replaced by any other formulation like the ones mentioned before. This relationship is visualised in the taxonomy in Fig. 2: Our method solves a global, continuous optimisation problem instead of greedy and local steps formulated as dynamic programming, and can hence function as a more flexible backbone for approaches that derive from seam carving. This formulation allows more complex solutions that work better in many scenarios (see Fig. 5), while the space of all possible outputs contains the solutions of seam carving." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 505, + 438, + 519 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 505, + 438, + 519 + ], + "spans": [ + { + "bbox": [ + 131, + 505, + 438, + 519 + ], + "type": "text", + "content": "3 Deformation Fields for Retargeting Visual Data" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 530, + 481, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 530, + 481, + 567 + ], + "spans": [ + { + "bbox": [ + 130, + 530, + 481, + 567 + ], + "type": "text", + "content": "Editing strategies for visual data require avoiding changes to meaningful content, that create visible artefacts. An approach that focuses on retargeting the input to match a target size can be guided by the following set of rules:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 137, + 574, + 479, + 632 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 137, + 574, + 479, + 607 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 574, + 479, + 607 + ], + "spans": [ + { + "bbox": [ + 137, + 574, + 479, + 607 + ], + "type": "text", + "content": "- Content Aware Objective: Prevent the accidental creation of new or the removal of existing regions with high information content (control over where deformation happens)." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 137, + 609, + 479, + 632 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 609, + 479, + 632 + ], + "spans": [ + { + "bbox": [ + 137, + 609, + 479, + 632 + ], + "type": "text", + "content": "- Sanity Objective: Avoid introducing deformations that yield implausible results (control over how deformation happens)." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "content": "In the following sections, we are only concerned with the retargeting to a smaller size (shrinking) for ease of explanation. The concept can naturally be" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 271, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 271, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 271, + 101 + ], + "type": "text", + "content": "T. Elsner, J. Berger et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 132, + 112, + 482, + 224 + ], + "blocks": [ + { + "bbox": [ + 132, + 112, + 482, + 224 + ], + "lines": [ + { + "bbox": [ + 132, + 112, + 482, + 224 + ], + "spans": [ + { + "bbox": [ + 132, + 112, + 482, + 224 + ], + "type": "image", + "image_path": "a947aca5d6b8bf3d860a5b4ac1d458b5616dcd9782220e84de626ece55cded26.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 232, + 482, + 289 + ], + "lines": [ + { + "bbox": [ + 130, + 232, + 482, + 289 + ], + "spans": [ + { + "bbox": [ + 130, + 232, + 482, + 289 + ], + "type": "text", + "content": "Fig. 3: Our proposed pipeline for retargeting visual data: For a given input (left), we train two simple networks that learn the energy and cumulative energy along the deformation axis of the input (centre-left). We initialise a network that stretches samples to the desired position (centre-right), then optimise this deformation to distribute the distortion to low information content regions (right). Balloon image from [42]." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 313, + 479, + 337 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 313, + 479, + 337 + ], + "spans": [ + { + "bbox": [ + 130, + 313, + 479, + 337 + ], + "type": "text", + "content": "reformulated to expand the visual data instead (expansion, see Sec. C (Appendix)) or to include editing (see Sec. 3.5)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 338, + 479, + 421 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 338, + 479, + 421 + ], + "spans": [ + { + "bbox": [ + 130, + 338, + 479, + 421 + ], + "type": "text", + "content": "Seam carving [2] and its derivatives produce a discrete solution to this by performing edits through manipulating (removing or repeating) one seam of pixels at a time, which is a single-pixel-wide path passing through the image. Through dynamic programming, seam carving chooses this path based on the lowest energy of the pixels, with high energy indicating information-rich areas. In a simple case, this energy is the colour gradient. This yields a discrete solution to the retargeting problem that is both content aware and plausible." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 422, + 482, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 422, + 482, + 495 + ], + "spans": [ + { + "bbox": [ + 130, + 422, + 482, + 495 + ], + "type": "text", + "content": "To obtain a continuous solution, instead of extracting discrete seams of the same width, we use continuous folds of varying width using a learned continuous deformation function. This formulation optimises a solution that is a superset of all existing seam carving solutions (see Fig. 5). We always use the same core formulation, regardless of the underlying domain and application, e.g. moving an object in an image or retargeting a 3D scene." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 517, + 265, + 529 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 517, + 265, + 529 + ], + "spans": [ + { + "bbox": [ + 132, + 517, + 265, + 529 + ], + "type": "text", + "content": "3.1 General Formulation" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 543, + 482, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 543, + 482, + 639 + ], + "spans": [ + { + "bbox": [ + 130, + 543, + 482, + 639 + ], + "type": "text", + "content": "Our approach always optimises a deformation function that maps from points in the retargeted output space (e.g. an image at half the size) " + }, + { + "bbox": [ + 130, + 543, + 482, + 639 + ], + "type": "inline_equation", + "content": "I: P \\to \\mathbb{R}^n" + }, + { + "bbox": [ + 130, + 543, + 482, + 639 + ], + "type": "text", + "content": " to points in the original data " + }, + { + "bbox": [ + 130, + 543, + 482, + 639 + ], + "type": "inline_equation", + "content": "I': P' \\to \\mathbb{R}^n" + }, + { + "bbox": [ + 130, + 543, + 482, + 639 + ], + "type": "text", + "content": " (e.g. to " + }, + { + "bbox": [ + 130, + 543, + 482, + 639 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^3" + }, + { + "bbox": [ + 130, + 543, + 482, + 639 + ], + "type": "text", + "content": " for RGB colours) by adding an offset to their coordinate. The value of a point " + }, + { + "bbox": [ + 130, + 543, + 482, + 639 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 130, + 543, + 482, + 639 + ], + "type": "text", + "content": " in the output data " + }, + { + "bbox": [ + 130, + 543, + 482, + 639 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 130, + 543, + 482, + 639 + ], + "type": "text", + "content": " is hence defined as the value of " + }, + { + "bbox": [ + 130, + 543, + 482, + 639 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 130, + 543, + 482, + 639 + ], + "type": "text", + "content": " deformed to a new point " + }, + { + "bbox": [ + 130, + 543, + 482, + 639 + ], + "type": "inline_equation", + "content": "p'" + }, + { + "bbox": [ + 130, + 543, + 482, + 639 + ], + "type": "text", + "content": " looked up in " + }, + { + "bbox": [ + 130, + 543, + 482, + 639 + ], + "type": "inline_equation", + "content": "I'" + }, + { + "bbox": [ + 130, + 543, + 482, + 639 + ], + "type": "text", + "content": ". To represent the deformation function " + }, + { + "bbox": [ + 130, + 543, + 482, + 639 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 130, + 543, + 482, + 639 + ], + "type": "text", + "content": ", we use a neural deformation field, a simple MLP where " + }, + { + "bbox": [ + 130, + 543, + 482, + 639 + ], + "type": "inline_equation", + "content": "D: P \\to \\mathbb{R}" + }, + { + "bbox": [ + 130, + 543, + 482, + 639 + ], + "type": "text", + "content": ". To keep our approach simpler, we only allow scalar deformations along a fixed deformation direction " + }, + { + "bbox": [ + 130, + 543, + 482, + 639 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 130, + 543, + 482, + 639 + ], + "type": "text", + "content": ", e.g. along the " + }, + { + "bbox": [ + 130, + 543, + 482, + 639 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 130, + 543, + 482, + 639 + ], + "type": "text", + "content": " axis (width) of an image:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 255, + 653, + 481, + 666 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 255, + 653, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 255, + 653, + 481, + 666 + ], + "type": "interline_equation", + "content": "I (p) = I ^ {\\prime} (p + D (p) \\cdot v) \\tag {1}", + "image_path": "c63dde5c16134f09ee6b36240a056655096b246c6a4d4cac7769cd873bb8b618.jpg" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 246, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 246, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 246, + 91, + 448, + 102 + ], + "type": "text", + "content": "Retargeting Visual Data with Deformation Fields" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 114, + 482, + 255 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 482, + 255 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 482, + 255 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 482, + 255 + ], + "type": "image", + "image_path": "a4cf99d9231932bf11879bae955364a8cdaa82a896041aa1996818aa746a678d.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 262, + 482, + 307 + ], + "lines": [ + { + "bbox": [ + 130, + 262, + 482, + 307 + ], + "spans": [ + { + "bbox": [ + 130, + 262, + 482, + 307 + ], + "type": "text", + "content": "Fig. 4: Our pipeline applied to neural radiance fields: We first obtain a point cloud, then estimate the energy values. We then learn a continuous energy function and a cumulative energy function that we use to optimise our deformation network. The resulting deformation field is visualised on the bottom right." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 143, + 321, + 468, + 449 + ], + "blocks": [ + { + "bbox": [ + 143, + 321, + 468, + 449 + ], + "lines": [ + { + "bbox": [ + 143, + 321, + 468, + 449 + ], + "spans": [ + { + "bbox": [ + 143, + 321, + 468, + 449 + ], + "type": "image", + "image_path": "f19b37273e1b4f64146e5637c753a39b3d46c5daa026dc1a703e802c202106ca.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 457, + 482, + 491 + ], + "lines": [ + { + "bbox": [ + 130, + 457, + 482, + 491 + ], + "spans": [ + { + "bbox": [ + 130, + 457, + 482, + 491 + ], + "type": "text", + "content": "Fig. 5: Example of applying retargeting to " + }, + { + "bbox": [ + 130, + 457, + 482, + 491 + ], + "type": "inline_equation", + "content": "75\\%" + }, + { + "bbox": [ + 130, + 457, + 482, + 491 + ], + "type": "text", + "content": " width. Seam carving fails to use seams that are non-continuous or with a too steep angle, not properly using the empty space in the shelf to close the gaps before shrinking the bookshelves content." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 519, + 482, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 519, + 482, + 568 + ], + "spans": [ + { + "bbox": [ + 130, + 519, + 482, + 568 + ], + "type": "text", + "content": "Hence, to get the content for a point " + }, + { + "bbox": [ + 130, + 519, + 482, + 568 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 130, + 519, + 482, + 568 + ], + "type": "text", + "content": " in the retargeted data, we offset the coordinate and look up the content at the offset coordinate " + }, + { + "bbox": [ + 130, + 519, + 482, + 568 + ], + "type": "inline_equation", + "content": "p'" + }, + { + "bbox": [ + 130, + 519, + 482, + 568 + ], + "type": "text", + "content": " in the original data. To fulfil the objectives stated in the beginning of Sec. 3, we introduce two sets of losses, content aware and sanity objectives:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 578, + 482, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 578, + 482, + 640 + ], + "spans": [ + { + "bbox": [ + 130, + 578, + 482, + 640 + ], + "type": "text", + "content": "Content Aware Objectives apply regularisation to guide the deformation process when adjusting to the data content. In particular, we penalise deformation in regions with high information content to discourage noticeable changes. We define the energy function " + }, + { + "bbox": [ + 130, + 578, + 482, + 640 + ], + "type": "inline_equation", + "content": "E: P' \\to \\mathbb{R}" + }, + { + "bbox": [ + 130, + 578, + 482, + 640 + ], + "type": "text", + "content": " to be a measure for content information in the original visual data, expresses as, e.g., a colour gradient:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 264, + 652, + 481, + 666 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 264, + 652, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 264, + 652, + 481, + 666 + ], + "type": "interline_equation", + "content": "E \\left(p ^ {\\prime}\\right) = \\left| \\left| \\nabla I \\left(p ^ {\\prime}\\right) \\right| \\right| _ {2} \\tag {2}", + "image_path": "979721984efea7778db07772e57d9a145d2bc354ac0c5452d3a630893e8a8b01.jpg" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 271, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 271, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 271, + 101 + ], + "type": "text", + "content": "T. Elsner, J. Berger et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 170, + 114, + 286, + 217 + ], + "blocks": [ + { + "bbox": [ + 170, + 114, + 286, + 217 + ], + "lines": [ + { + "bbox": [ + 170, + 114, + 286, + 217 + ], + "spans": [ + { + "bbox": [ + 170, + 114, + 286, + 217 + ], + "type": "image", + "image_path": "d7155f8351a4bc1be3916739063663ed51d396663c7fff263404983649e19838.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 225, + 480, + 247 + ], + "lines": [ + { + "bbox": [ + 131, + 225, + 480, + 247 + ], + "spans": [ + { + "bbox": [ + 131, + 225, + 480, + 247 + ], + "type": "text", + "content": "Fig. 6: Applying our retargeting to 3D meshes, moving furniture closer together to retarget the scene to a smaller size while limiting distortion to objects." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 288, + 116, + 363, + 215 + ], + "blocks": [ + { + "bbox": [ + 288, + 116, + 363, + 215 + ], + "lines": [ + { + "bbox": [ + 288, + 116, + 363, + 215 + ], + "spans": [ + { + "bbox": [ + 288, + 116, + 363, + 215 + ], + "type": "image", + "image_path": "479c7c964d87091d9e2ff5082052f7a9a85379dcf805c46dab759962506dcf37.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 366, + 115, + 442, + 216 + ], + "blocks": [ + { + "bbox": [ + 366, + 115, + 442, + 216 + ], + "lines": [ + { + "bbox": [ + 366, + 115, + 442, + 216 + ], + "spans": [ + { + "bbox": [ + 366, + 115, + 442, + 216 + ], + "type": "image", + "image_path": "e0527205f9d5202a85a40f2cb96af7191ab6e52fc07b84331541245d49debd39.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 271, + 479, + 295 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 271, + 479, + 295 + ], + "spans": [ + { + "bbox": [ + 131, + 271, + 479, + 295 + ], + "type": "text", + "content": "The overall content deformation loss is the product of deformation magnitude at " + }, + { + "bbox": [ + 131, + 271, + 479, + 295 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 131, + 271, + 479, + 295 + ], + "type": "text", + "content": " and energy at the original " + }, + { + "bbox": [ + 131, + 271, + 479, + 295 + ], + "type": "inline_equation", + "content": "p'" + }, + { + "bbox": [ + 131, + 271, + 479, + 295 + ], + "type": "text", + "content": " over all " + }, + { + "bbox": [ + 131, + 271, + 479, + 295 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 131, + 271, + 479, + 295 + ], + "type": "text", + "content": " in the deformed data:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 212, + 304, + 480, + 336 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 212, + 304, + 480, + 336 + ], + "spans": [ + { + "bbox": [ + 212, + 304, + 480, + 336 + ], + "type": "interline_equation", + "content": "L _ {C} = \\int_ {p \\in P} [ E (p + v \\cdot D (p)) \\cdot | | \\nabla D (p) | | _ {1} ] d p \\tag {3}", + "image_path": "f025eaa5ef088517ac0f33054f9505e81dc43596cd3dbacb04f09143f52e2c92.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 346, + 479, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 346, + 479, + 418 + ], + "spans": [ + { + "bbox": [ + 131, + 346, + 479, + 418 + ], + "type": "text", + "content": "Thus, in Eq. (3), we effectively penalise visible deformations defined as the product of information content at points, namely " + }, + { + "bbox": [ + 131, + 346, + 479, + 418 + ], + "type": "inline_equation", + "content": "E(p + v \\cdot D(p))" + }, + { + "bbox": [ + 131, + 346, + 479, + 418 + ], + "type": "text", + "content": ", and the deformation itself via " + }, + { + "bbox": [ + 131, + 346, + 479, + 418 + ], + "type": "inline_equation", + "content": "||\\nabla D(p)||_1" + }, + { + "bbox": [ + 131, + 346, + 479, + 418 + ], + "type": "text", + "content": ". Using the " + }, + { + "bbox": [ + 131, + 346, + 479, + 418 + ], + "type": "inline_equation", + "content": "L^1" + }, + { + "bbox": [ + 131, + 346, + 479, + 418 + ], + "type": "text", + "content": "-norm for the distortion gradient promotes piecewise constant, possibly sparse, deformations. We can then split up the change in deformation into change in deformation direction " + }, + { + "bbox": [ + 131, + 346, + 479, + 418 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 131, + 346, + 479, + 418 + ], + "type": "text", + "content": " and change in direction " + }, + { + "bbox": [ + 131, + 346, + 479, + 418 + ], + "type": "inline_equation", + "content": "v^\\perp" + }, + { + "bbox": [ + 131, + 346, + 479, + 418 + ], + "type": "text", + "content": " orthogonal to " + }, + { + "bbox": [ + 131, + 346, + 479, + 418 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 131, + 346, + 479, + 418 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 171, + 426, + 480, + 460 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 171, + 426, + 480, + 460 + ], + "spans": [ + { + "bbox": [ + 171, + 426, + 480, + 460 + ], + "type": "interline_equation", + "content": "L _ {C} = \\int_ {p \\in P} \\left[ E (p + v \\cdot D (p)) \\cdot \\left(\\left| \\frac {\\partial}{\\partial v} D (p) \\right| + \\left| \\frac {\\partial}{\\partial v ^ {\\perp}} D (p) \\right|\\right) \\right] d p \\tag {4}", + "image_path": "c172a9fdf4a8d875401907e0f7566a1ed13988f4d86a59880398bf5b2ac2903b.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 468, + 480, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 468, + 480, + 517 + ], + "spans": [ + { + "bbox": [ + 131, + 468, + 480, + 517 + ], + "type": "text", + "content": "This equation is then further split into two components, " + }, + { + "bbox": [ + 131, + 468, + 480, + 517 + ], + "type": "inline_equation", + "content": "L_{e}" + }, + { + "bbox": [ + 131, + 468, + 480, + 517 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 468, + 480, + 517 + ], + "type": "inline_equation", + "content": "L_{s}" + }, + { + "bbox": [ + 131, + 468, + 480, + 517 + ], + "type": "text", + "content": ". The loss " + }, + { + "bbox": [ + 131, + 468, + 480, + 517 + ], + "type": "inline_equation", + "content": "L_{e}" + }, + { + "bbox": [ + 131, + 468, + 480, + 517 + ], + "type": "text", + "content": " penalises stretching or compression and " + }, + { + "bbox": [ + 131, + 468, + 480, + 517 + ], + "type": "inline_equation", + "content": "L_{s}" + }, + { + "bbox": [ + 131, + 468, + 480, + 517 + ], + "type": "text", + "content": " punishes shearing, which we define to be deformation change in a direction that is not the deformation direction. For a now discrete sample set " + }, + { + "bbox": [ + 131, + 468, + 480, + 517 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 131, + 468, + 480, + 517 + ], + "type": "text", + "content": " from our domain, these terms are expressed as:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 216, + 525, + 480, + 563 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 216, + 525, + 480, + 563 + ], + "spans": [ + { + "bbox": [ + 216, + 525, + 480, + 563 + ], + "type": "interline_equation", + "content": "L _ {e} = \\sum_ {p \\in P} \\left[ E (p + v \\cdot D (p)) \\cdot \\left| \\frac {\\partial}{\\partial v} D (p) \\right| \\right] \\tag {5}", + "image_path": "5ef0a3a8fa932bcbeb49fc9e66348926e2dedcc95dfa5547a3b3796c0befb79e.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 219, + 558, + 394, + 588 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 219, + 558, + 394, + 588 + ], + "spans": [ + { + "bbox": [ + 219, + 558, + 394, + 588 + ], + "type": "interline_equation", + "content": "L _ {s} = \\sum_ {p \\in P} \\left[ E (p + v \\cdot D (p)) \\cdot \\left| \\frac {\\partial}{\\partial v ^ {\\perp}} D (p) \\right| \\right]", + "image_path": "2e4c3514f195d12dd17e1f825c46f2529187a50ce39465e1803b734c8b9e0eac.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 131, + 594, + 479, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 594, + 479, + 665 + ], + "spans": [ + { + "bbox": [ + 131, + 594, + 479, + 665 + ], + "type": "text", + "content": "Essentially, " + }, + { + "bbox": [ + 131, + 594, + 479, + 665 + ], + "type": "inline_equation", + "content": "L_{e}" + }, + { + "bbox": [ + 131, + 594, + 479, + 665 + ], + "type": "text", + "content": " directs deformations to areas of low information content, while " + }, + { + "bbox": [ + 131, + 594, + 479, + 665 + ], + "type": "inline_equation", + "content": "L_{s}" + }, + { + "bbox": [ + 131, + 594, + 479, + 665 + ], + "type": "text", + "content": " discourages the resulting shearing to affect areas of high information content. Note that while " + }, + { + "bbox": [ + 131, + 594, + 479, + 665 + ], + "type": "inline_equation", + "content": "L_{s}" + }, + { + "bbox": [ + 131, + 594, + 479, + 665 + ], + "type": "text", + "content": " would also penalise non-straight deformation seams, we tune our losses such that this only impacts the output much if an area is sheared. An example of this is shown in Fig. 3: The deformations around the balloons introduce shearing only at the seam, in an area of low information content." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 246, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 246, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 246, + 91, + 448, + 102 + ], + "type": "text", + "content": "Retargeting Visual Data with Deformation Fields" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 188 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 188 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 188 + ], + "type": "text", + "content": "In case of image deformation, neighbouring pixels with low energy in the input may translate to pixels widely apart in the output, potentially bypassing a substantial amount of important content. Jumps in the deformation field " + }, + { + "bbox": [ + 130, + 116, + 482, + 188 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 130, + 116, + 482, + 188 + ], + "type": "text", + "content": " indicate this effect. To prevent this, we penalise the amount of energy " + }, + { + "bbox": [ + 130, + 116, + 482, + 188 + ], + "type": "inline_equation", + "content": "\\hat{E}" + }, + { + "bbox": [ + 130, + 116, + 482, + 188 + ], + "type": "text", + "content": " between two points " + }, + { + "bbox": [ + 130, + 116, + 482, + 188 + ], + "type": "inline_equation", + "content": "p', q' \\in P'" + }, + { + "bbox": [ + 130, + 116, + 482, + 188 + ], + "type": "text", + "content": " in the original domain, rather than the energy at individual positions, as in Eq. (2):" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 246, + 195, + 481, + 229 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 246, + 195, + 481, + 229 + ], + "spans": [ + { + "bbox": [ + 246, + 195, + 481, + 229 + ], + "type": "interline_equation", + "content": "\\hat {E} \\left(p ^ {\\prime}, q ^ {\\prime}\\right) = \\int_ {r ^ {\\prime} \\in \\left[ p ^ {\\prime}, q ^ {\\prime} \\right]} E \\left(r ^ {\\prime}\\right) d r ^ {\\prime} \\tag {6}", + "image_path": "ae1f64b4f934dcd11f69266fd647024622e86cb34adb33f75a8d596730a624ae.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 236, + 481, + 273 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 236, + 481, + 273 + ], + "spans": [ + { + "bbox": [ + 130, + 236, + 481, + 273 + ], + "type": "text", + "content": "Using " + }, + { + "bbox": [ + 130, + 236, + 481, + 273 + ], + "type": "inline_equation", + "content": "\\hat{E}" + }, + { + "bbox": [ + 130, + 236, + 481, + 273 + ], + "type": "text", + "content": ", we then redefine " + }, + { + "bbox": [ + 130, + 236, + 481, + 273 + ], + "type": "inline_equation", + "content": "L_{e}" + }, + { + "bbox": [ + 130, + 236, + 481, + 273 + ], + "type": "text", + "content": " from Eq. (5) to consider energy and difference in deformation magnitude between " + }, + { + "bbox": [ + 130, + 236, + 481, + 273 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 130, + 236, + 481, + 273 + ], + "type": "text", + "content": " and a slightly offset " + }, + { + "bbox": [ + 130, + 236, + 481, + 273 + ], + "type": "inline_equation", + "content": "p_{\\varepsilon} \\coloneqq p + \\varepsilon" + }, + { + "bbox": [ + 130, + 236, + 481, + 273 + ], + "type": "text", + "content": " in deformation direction:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 186, + 282, + 481, + 312 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 186, + 282, + 481, + 312 + ], + "spans": [ + { + "bbox": [ + 186, + 282, + 481, + 312 + ], + "type": "interline_equation", + "content": "L _ {e} = \\sum_ {p \\in P} \\hat {E} (p + v D (p), p _ {\\varepsilon} + v D (p _ {\\varepsilon})) \\cdot \\frac {| D (p) - D (p _ {\\varepsilon}) |}{\\varepsilon} \\tag {7}", + "image_path": "86fc4fe5664dbe074f825e55c7037ab3300a7918201312d15ae8f84673c93350.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 316, + 481, + 364 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 316, + 481, + 364 + ], + "spans": [ + { + "bbox": [ + 131, + 316, + 481, + 364 + ], + "type": "text", + "content": "This formulation guides the deformation to avoid introducing deformation to high information content regions or fold over them, while " + }, + { + "bbox": [ + 131, + 316, + 481, + 364 + ], + "type": "inline_equation", + "content": "L_{s}" + }, + { + "bbox": [ + 131, + 316, + 481, + 364 + ], + "type": "text", + "content": " discourages non-local deformations that would apply shearing in high information areas. For example, shearing in blue sky is acceptable, shearing on a brick wall is not." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 375, + 481, + 400 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 375, + 481, + 400 + ], + "spans": [ + { + "bbox": [ + 131, + 375, + 481, + 400 + ], + "type": "text", + "content": "Sanity objectives prevent undesired outputs by introducing two additional losses. Assume we shrink the input by a factor " + }, + { + "bbox": [ + 131, + 375, + 481, + 400 + ], + "type": "inline_equation", + "content": "\\alpha \\in [0,1]" + }, + { + "bbox": [ + 131, + 375, + 481, + 400 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 136, + 405, + 482, + 465 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 405, + 482, + 465 + ], + "spans": [ + { + "bbox": [ + 136, + 405, + 482, + 465 + ], + "type": "text", + "content": "1. The boundaries of the affected region should not change, e.g. an image should not be simply cropped. Based on " + }, + { + "bbox": [ + 136, + 405, + 482, + 465 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 136, + 405, + 482, + 465 + ], + "type": "text", + "content": ", the deformation field should map the start/end of the deformation axis in the output to the start/end of the input. For " + }, + { + "bbox": [ + 136, + 405, + 482, + 465 + ], + "type": "inline_equation", + "content": "P_0" + }, + { + "bbox": [ + 136, + 405, + 482, + 465 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 136, + 405, + 482, + 465 + ], + "type": "inline_equation", + "content": "P_1" + }, + { + "bbox": [ + 136, + 405, + 482, + 465 + ], + "type": "text", + "content": " containing all points at the start/end of the deformation axis of the output, we express this requirement as follows:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 215, + 472, + 481, + 505 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 215, + 472, + 481, + 505 + ], + "spans": [ + { + "bbox": [ + 215, + 472, + 481, + 505 + ], + "type": "interline_equation", + "content": "L _ {b} = \\int_ {p \\in P _ {0}} | D (p) | d p + \\int_ {p \\in P _ {1}} | D (p) - (1 - \\alpha) | d p \\tag {8}", + "image_path": "a2a5a8361f932ff890b0e3ddd0285be5d27e4b3496000531dbe2aa72f87f21dc.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 136, + 510, + 481, + 522 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 510, + 481, + 522 + ], + "spans": [ + { + "bbox": [ + 136, + 510, + 481, + 522 + ], + "type": "text", + "content": "2. We additionally demand monotonicity in the deformation field, expressed as" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 240, + 531, + 481, + 564 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 240, + 531, + 481, + 564 + ], + "spans": [ + { + "bbox": [ + 240, + 531, + 481, + 564 + ], + "type": "interline_equation", + "content": "L _ {m} = \\int_ {p \\in P} \\max \\left(0, - \\frac {\\partial}{\\partial v} D (p)\\right) d p, \\tag {9}", + "image_path": "977048463dfa651be151f96e13ec0fdab6eadd148becce9bac0e7952cee061e4.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 148, + 567, + 481, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 148, + 567, + 481, + 616 + ], + "spans": [ + { + "bbox": [ + 148, + 567, + 481, + 616 + ], + "type": "text", + "content": "indicating that any change in the deformation direction should be positive. Therefore, deformation " + }, + { + "bbox": [ + 148, + 567, + 481, + 616 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 148, + 567, + 481, + 616 + ], + "type": "text", + "content": " should be monotonic and only stagnate or increase exclusively along " + }, + { + "bbox": [ + 148, + 567, + 481, + 616 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 148, + 567, + 481, + 616 + ], + "type": "text", + "content": ". This prevents the repetition of the same image portion, noticeable by jumps with inconsistent direction in the deformation field." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 633, + 399, + 646 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 633, + 399, + 646 + ], + "spans": [ + { + "bbox": [ + 132, + 633, + 399, + 646 + ], + "type": "text", + "content": "We then obtain the total loss " + }, + { + "bbox": [ + 132, + 633, + 399, + 646 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 132, + 633, + 399, + 646 + ], + "type": "text", + "content": " with hyperparameters " + }, + { + "bbox": [ + 132, + 633, + 399, + 646 + ], + "type": "inline_equation", + "content": "\\lambda_{*}" + }, + { + "bbox": [ + 132, + 633, + 399, + 646 + ], + "type": "text", + "content": " as" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 233, + 654, + 481, + 666 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 233, + 654, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 233, + 654, + 481, + 666 + ], + "type": "interline_equation", + "content": "L = \\lambda_ {e} L _ {e} + \\lambda_ {s} L _ {s} + \\lambda_ {b} L _ {b} + \\lambda_ {m} L _ {m}. \\tag {10}", + "image_path": "a0dfd230a9bb2e4d2c216915521b08ab326fb44e8abc0933b6da73c6bef6ee36.jpg" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 271, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 271, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 271, + 102 + ], + "type": "text", + "content": "T. Elsner, J. Berger et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 479, + 128 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 479, + 128 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 479, + 128 + ], + "type": "text", + "content": "We initialise the deformation function with a uniform stretch to the target size." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 129, + 479, + 176 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 129, + 479, + 176 + ], + "spans": [ + { + "bbox": [ + 130, + 129, + 479, + 176 + ], + "type": "text", + "content": "In the following, we demonstrate the application of this general formulation to different domains, while adapting the energy formulation to the problem. All implementation and architecture details are given in Sec. A (Appendix) and the exact loss terms in Sec. B (Appendix)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 194, + 373, + 206 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 194, + 373, + 206 + ], + "spans": [ + { + "bbox": [ + 132, + 194, + 373, + 206 + ], + "type": "text", + "content": "3.2 Optimising Deformation Fields for Images" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 215, + 479, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 215, + 479, + 251 + ], + "spans": [ + { + "bbox": [ + 131, + 215, + 479, + 251 + ], + "type": "text", + "content": "Instead of using image data directly, we train a continuous image representation MLP " + }, + { + "bbox": [ + 131, + 215, + 479, + 251 + ], + "type": "inline_equation", + "content": "I': P' \\to \\mathbb{R}^3" + }, + { + "bbox": [ + 131, + 215, + 479, + 251 + ], + "type": "text", + "content": " to be our input data, that maps pixel positions to colours. This ensures consistency with our continuous approaches." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 262, + 479, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 262, + 479, + 335 + ], + "spans": [ + { + "bbox": [ + 131, + 262, + 479, + 335 + ], + "type": "text", + "content": "We train one MLP to learn the energy field " + }, + { + "bbox": [ + 131, + 262, + 479, + 335 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 131, + 262, + 479, + 335 + ], + "type": "text", + "content": ", holding the colour gradient per point as in Eq. (2). Using the learned " + }, + { + "bbox": [ + 131, + 262, + 479, + 335 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 131, + 262, + 479, + 335 + ], + "type": "text", + "content": ", we train another MLP to replicate " + }, + { + "bbox": [ + 131, + 262, + 479, + 335 + ], + "type": "inline_equation", + "content": "\\hat{E}" + }, + { + "bbox": [ + 131, + 262, + 479, + 335 + ], + "type": "text", + "content": " as a cumulative energy field " + }, + { + "bbox": [ + 131, + 262, + 479, + 335 + ], + "type": "inline_equation", + "content": "\\Sigma" + }, + { + "bbox": [ + 131, + 262, + 479, + 335 + ], + "type": "text", + "content": ". The difference in cumulative energy " + }, + { + "bbox": [ + 131, + 262, + 479, + 335 + ], + "type": "inline_equation", + "content": "\\Sigma" + }, + { + "bbox": [ + 131, + 262, + 479, + 335 + ], + "type": "text", + "content": " between two points " + }, + { + "bbox": [ + 131, + 262, + 479, + 335 + ], + "type": "inline_equation", + "content": "p', q' \\in P'" + }, + { + "bbox": [ + 131, + 262, + 479, + 335 + ], + "type": "text", + "content": " on the same axis " + }, + { + "bbox": [ + 131, + 262, + 479, + 335 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 131, + 262, + 479, + 335 + ], + "type": "text", + "content": " is the absolute difference in information content between them. As this coincides with the definition of " + }, + { + "bbox": [ + 131, + 262, + 479, + 335 + ], + "type": "inline_equation", + "content": "\\hat{E}" + }, + { + "bbox": [ + 131, + 262, + 479, + 335 + ], + "type": "text", + "content": " from Eq. (6), we approximate the measure of content relevance between two points as:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 249, + 344, + 480, + 357 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 249, + 344, + 480, + 357 + ], + "spans": [ + { + "bbox": [ + 249, + 344, + 480, + 357 + ], + "type": "interline_equation", + "content": "\\hat {E} \\left(p ^ {\\prime}, q ^ {\\prime}\\right) \\approx | \\Sigma \\left(q ^ {\\prime}\\right) - \\Sigma \\left(p ^ {\\prime}\\right) | \\tag {11}", + "image_path": "6a4673f21f612b23ba5e67227609cc48de8868eafcd0a8843720ee672beac9ee.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 366, + 480, + 463 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 366, + 480, + 463 + ], + "spans": [ + { + "bbox": [ + 131, + 366, + 480, + 463 + ], + "type": "text", + "content": "This cumulative gradient formulation therefore simplifies the interval over " + }, + { + "bbox": [ + 131, + 366, + 480, + 463 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 131, + 366, + 480, + 463 + ], + "type": "text", + "content": " in Eq. (6) to a simple difference computation between two MLP queries: How much content is lost when the deformation skips a region. The energy field " + }, + { + "bbox": [ + 131, + 366, + 480, + 463 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 131, + 366, + 480, + 463 + ], + "type": "text", + "content": " can be trained to reproduce any energy formulation other than a colour gradient. This exchangeable energy formulation ensures compatibility with other retargeting approaches as shown in Fig. 2. For all objectives, we use discrete approximations of the loss terms introduced in Sec. 3.1. The whole pipeline is visualised in Fig. 3, and additional results can be found in Sec. G (Appendix)." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 480, + 454, + 493 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 480, + 454, + 493 + ], + "spans": [ + { + "bbox": [ + 132, + 480, + 454, + 493 + ], + "type": "text", + "content": "3.3 Optimising Deformation Fields for Neural Radiance Fields" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 501, + 479, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 501, + 479, + 525 + ], + "spans": [ + { + "bbox": [ + 131, + 501, + 479, + 525 + ], + "type": "text", + "content": "We adapt our approach from Sec. 3.1 for 3D scenes given as trained NeRFs as follows (see Fig. 4):" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 137, + 533, + 480, + 664 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 138, + 533, + 480, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 533, + 480, + 569 + ], + "spans": [ + { + "bbox": [ + 138, + 533, + 480, + 569 + ], + "type": "text", + "content": "- A trained NeRF provides colour and depth images, from which we extract surface samples. Each point is assigned an energy value defined by change in colour and depth at this point." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 570, + 480, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 570, + 480, + 604 + ], + "spans": [ + { + "bbox": [ + 138, + 570, + 480, + 604 + ], + "type": "text", + "content": "- Based on the surface samples and their energy values, we train a continuous energy field " + }, + { + "bbox": [ + 138, + 570, + 480, + 604 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 138, + 570, + 480, + 604 + ], + "type": "text", + "content": " and a continuous cumulative energy field " + }, + { + "bbox": [ + 138, + 570, + 480, + 604 + ], + "type": "inline_equation", + "content": "\\Sigma" + }, + { + "bbox": [ + 138, + 570, + 480, + 604 + ], + "type": "text", + "content": ", to easily access the amount of content between points." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 137, + 605, + 480, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 605, + 480, + 640 + ], + "spans": [ + { + "bbox": [ + 137, + 605, + 480, + 640 + ], + "type": "text", + "content": "- We use an inverse deformation network " + }, + { + "bbox": [ + 137, + 605, + 480, + 640 + ], + "type": "inline_equation", + "content": "U" + }, + { + "bbox": [ + 137, + 605, + 480, + 640 + ], + "type": "text", + "content": " to find points in the target space that correspond to surface points in the source space, keeping computing costs low by restricting our optimisation to surface points." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 137, + 642, + 480, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 642, + 480, + 664 + ], + "spans": [ + { + "bbox": [ + 137, + 642, + 480, + 664 + ], + "type": "text", + "content": "- We apply our regularisation terms on a mix of surface and random points to optimise our loss function." + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 245, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 245, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 245, + 91, + 447, + 102 + ], + "type": "text", + "content": "Retargeting Visual Data with Deformation Fields" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 201 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 201 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 201 + ], + "type": "text", + "content": "1. Extracting a Sparse Set of Representatives and Energy We extract a surface point cloud from the RGB and depth images rendered by NeRF at the training camera positions. Only these points are relevant for scene deformation. We then compute the energy value for each point corresponding to a pixel as a sum of colour change in the RGB image and depth change in the depth image. To reduce the impact of highly noisy energy values, we apply smoothing by selecting the minimum energy for each point within a neighbourhood radius of 5 points." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 211, + 482, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 211, + 482, + 331 + ], + "spans": [ + { + "bbox": [ + 130, + 211, + 482, + 331 + ], + "type": "text", + "content": "2. Defining an Energy Function To obtain a continuous energy field, we train an auxiliary energy network " + }, + { + "bbox": [ + 130, + 211, + 482, + 331 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 130, + 211, + 482, + 331 + ], + "type": "text", + "content": ". The network maps all 3D points to their energy values, and clamps the energy to zero for points further than twice the average neighbour distance. The resulting energy field has zero energy in empty space and inside objects, and non-zero energy values for points near the surface. Similar to our approach with images, we learn a cumulative energy network " + }, + { + "bbox": [ + 130, + 211, + 482, + 331 + ], + "type": "inline_equation", + "content": "\\Sigma" + }, + { + "bbox": [ + 130, + 211, + 482, + 331 + ], + "type": "text", + "content": ". The network is trained to accumulate the energy values " + }, + { + "bbox": [ + 130, + 211, + 482, + 331 + ], + "type": "inline_equation", + "content": "E(p)" + }, + { + "bbox": [ + 130, + 211, + 482, + 331 + ], + "type": "text", + "content": " from uniformly sampled points " + }, + { + "bbox": [ + 130, + 211, + 482, + 331 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 130, + 211, + 482, + 331 + ], + "type": "text", + "content": " on a ray in deformation direction. We compute the energy for a segment between two points " + }, + { + "bbox": [ + 130, + 211, + 482, + 331 + ], + "type": "inline_equation", + "content": "p', q' \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 130, + 211, + 482, + 331 + ], + "type": "text", + "content": " along " + }, + { + "bbox": [ + 130, + 211, + 482, + 331 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 130, + 211, + 482, + 331 + ], + "type": "text", + "content": " using their difference in cumulative energy " + }, + { + "bbox": [ + 130, + 211, + 482, + 331 + ], + "type": "inline_equation", + "content": "\\Sigma(q')" + }, + { + "bbox": [ + 130, + 211, + 482, + 331 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 211, + 482, + 331 + ], + "type": "inline_equation", + "content": "\\Sigma(p')" + }, + { + "bbox": [ + 130, + 211, + 482, + 331 + ], + "type": "text", + "content": " (cf. Eq. (11))." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 342, + 482, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 342, + 482, + 449 + ], + "spans": [ + { + "bbox": [ + 130, + 342, + 482, + 449 + ], + "type": "text", + "content": "3. Inverting the Deformation We define an inverse deformation function to determine those points that lie on the surface after deformation. In contrast to images, we only perform optimisation on a sparse set of points, namely surface points. As we optimise a deformation from the target space to the source space, we need the position of surface points in both spaces. By training an MLP " + }, + { + "bbox": [ + 130, + 342, + 482, + 449 + ], + "type": "inline_equation", + "content": "U" + }, + { + "bbox": [ + 130, + 342, + 482, + 449 + ], + "type": "text", + "content": " to reverse the deformation " + }, + { + "bbox": [ + 130, + 342, + 482, + 449 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 130, + 342, + 482, + 449 + ], + "type": "text", + "content": ", i.e. minimising " + }, + { + "bbox": [ + 130, + 342, + 482, + 449 + ], + "type": "inline_equation", + "content": "(U(D(x)) = x)^2" + }, + { + "bbox": [ + 130, + 342, + 482, + 449 + ], + "type": "text", + "content": ", we can effortlessly extract surface points, and hence their energy values, in the target space. This relation is also visualised in Fig. 4. We always update " + }, + { + "bbox": [ + 130, + 342, + 482, + 449 + ], + "type": "inline_equation", + "content": "U" + }, + { + "bbox": [ + 130, + 342, + 482, + 449 + ], + "type": "text", + "content": " once after training one iteration of " + }, + { + "bbox": [ + 130, + 342, + 482, + 449 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 130, + 342, + 482, + 449 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 461, + 482, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 461, + 482, + 510 + ], + "spans": [ + { + "bbox": [ + 130, + 461, + 482, + 510 + ], + "type": "text", + "content": "As we use a discrete set of points, we apply the losses from Sec. 3.1 to Eq. (10) for a discrete interval. In particular, we average the loss over a random subset of sample points. Example images from deformed scenes can be found in Fig. 8, while a larger number of results can be found in Sec. G (Appendix)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 526, + 375, + 538 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 526, + 375, + 538 + ], + "spans": [ + { + "bbox": [ + 130, + 526, + 375, + 538 + ], + "type": "text", + "content": "3.4 Optimising Deformation Fields for Meshes" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 545, + 482, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 482, + 605 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 482, + 605 + ], + "type": "text", + "content": "To demonstrate the versatility of our approach, we apply it to meshes as well. We adapt the energy term and use surface samples and vertex positions for optimisation. For freestanding objects, we use curvature as the energy measure. For the room example in Fig. 6, the energy is set to 0 for the floor and to 1 for every other surface point. We otherwise follow the same principle as in Sec. 3.3." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 622, + 257, + 634 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 622, + 257, + 634 + ], + "spans": [ + { + "bbox": [ + 130, + 622, + 257, + 634 + ], + "type": "text", + "content": "3.5 Other Applications" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "content": "While retargeting is our main contribution, our framework opens up multiple different editing opportunities that even extend the capabilities of seam carving [2]." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 271, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 271, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 271, + 102 + ], + "type": "text", + "content": "T. Elsner, J. Berger et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 155, + 128, + 255, + 190 + ], + "blocks": [ + { + "bbox": [ + 130, + 148, + 141, + 172 + ], + "lines": [ + { + "bbox": [ + 130, + 148, + 141, + 172 + ], + "spans": [ + { + "bbox": [ + 130, + 148, + 141, + 172 + ], + "type": "text", + "content": "Iinui" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 155, + 128, + 255, + 190 + ], + "lines": [ + { + "bbox": [ + 155, + 128, + 255, + 190 + ], + "spans": [ + { + "bbox": [ + 155, + 128, + 255, + 190 + ], + "type": "image", + "image_path": "41087d1b823b0ef952f4422c8b1fdb6095b7086826307699a89f304cef065f23.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 267, + 128, + 307, + 190 + ], + "blocks": [ + { + "bbox": [ + 267, + 128, + 307, + 190 + ], + "lines": [ + { + "bbox": [ + 267, + 128, + 307, + 190 + ], + "spans": [ + { + "bbox": [ + 267, + 128, + 307, + 190 + ], + "type": "image", + "image_path": "c877131d3cd9a0e3bad2108f760006a6cc9b37f2dbd32f4b5de504f22166e232.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 155, + 191, + 255, + 254 + ], + "blocks": [ + { + "bbox": [ + 130, + 208, + 141, + 236 + ], + "lines": [ + { + "bbox": [ + 130, + 208, + 141, + 236 + ], + "spans": [ + { + "bbox": [ + 130, + 208, + 141, + 236 + ], + "type": "text", + "content": "ed" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 155, + 191, + 255, + 254 + ], + "lines": [ + { + "bbox": [ + 155, + 191, + 255, + 254 + ], + "spans": [ + { + "bbox": [ + 155, + 191, + 255, + 254 + ], + "type": "image", + "image_path": "483d4940066c036181d0d5623a4f2622255bc9ae6707c2ae2a3fc6cf017b5f4d.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 266, + 309, + 312 + ], + "lines": [ + { + "bbox": [ + 131, + 266, + 309, + 312 + ], + "spans": [ + { + "bbox": [ + 131, + 266, + 309, + 312 + ], + "type": "text", + "content": "Fig. 7: We demonstrate different editing operations by adding and adapting the loss formulation. Left: Removing, right: Moving an object. Images from [42]." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 267, + 191, + 307, + 254 + ], + "blocks": [ + { + "bbox": [ + 267, + 191, + 307, + 254 + ], + "lines": [ + { + "bbox": [ + 267, + 191, + 307, + 254 + ], + "spans": [ + { + "bbox": [ + 267, + 191, + 307, + 254 + ], + "type": "image", + "image_path": "5b2c071ea3cd1a71b60adbf886f91528f6a8888fda1662458f4ee72e4f392aa5.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 326, + 121, + 480, + 211 + ], + "blocks": [ + { + "bbox": [ + 326, + 121, + 480, + 211 + ], + "lines": [ + { + "bbox": [ + 326, + 121, + 480, + 211 + ], + "spans": [ + { + "bbox": [ + 326, + 121, + 480, + 211 + ], + "type": "image", + "image_path": "e9c01c195ebf0ff3921e10ccd9a4bcd79c855141d59ca30a154aa2088df2eef1.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 323, + 212, + 481, + 251 + ], + "blocks": [ + { + "bbox": [ + 323, + 212, + 481, + 251 + ], + "lines": [ + { + "bbox": [ + 323, + 212, + 481, + 251 + ], + "spans": [ + { + "bbox": [ + 323, + 212, + 481, + 251 + ], + "type": "image", + "image_path": "5b01bb21a0c5e53a64d87bb5e21ff28c8181e8ab46d966e7ee0e9315ce5281fd.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 321, + 258, + 482, + 324 + ], + "lines": [ + { + "bbox": [ + 321, + 258, + 482, + 324 + ], + "spans": [ + { + "bbox": [ + 321, + 258, + 482, + 324 + ], + "type": "text", + "content": "Fig.8: Applying our retargeting to NeRFs. Top row: Input, stretched, targeted with our method. Bottom row: Input, folds (white) on the input suggested by our approach, targeted with our method." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 132, + 347, + 202, + 396 + ], + "blocks": [ + { + "bbox": [ + 132, + 347, + 202, + 396 + ], + "lines": [ + { + "bbox": [ + 132, + 347, + 202, + 396 + ], + "spans": [ + { + "bbox": [ + 132, + 347, + 202, + 396 + ], + "type": "image", + "image_path": "2e1d792e0bc9fdfa6765039930e5928af706c34079ee4ce241a4fce66f0fa475.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 405, + 480, + 439 + ], + "lines": [ + { + "bbox": [ + 131, + 405, + 480, + 439 + ], + "spans": [ + { + "bbox": [ + 131, + 405, + 480, + 439 + ], + "type": "text", + "content": "Fig. 9: Selected results showing scenes where our approach benefits from a higher degree of flexibility. All images are arranged as input, seam carving, and our output. Images from [42]." + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 203, + 347, + 271, + 396 + ], + "blocks": [ + { + "bbox": [ + 203, + 347, + 271, + 396 + ], + "lines": [ + { + "bbox": [ + 203, + 347, + 271, + 396 + ], + "spans": [ + { + "bbox": [ + 203, + 347, + 271, + 396 + ], + "type": "image", + "image_path": "f94ee2cb297addced415b0e223edb326bea5527110da4a42766497f224294b12.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 271, + 347, + 306, + 396 + ], + "blocks": [ + { + "bbox": [ + 271, + 347, + 306, + 396 + ], + "lines": [ + { + "bbox": [ + 271, + 347, + 306, + 396 + ], + "spans": [ + { + "bbox": [ + 271, + 347, + 306, + 396 + ], + "type": "image", + "image_path": "f9e9f691756184c008def46be8d17f37589b54b356ee47f8ba4ba9288f7f3ba5.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 307, + 347, + 342, + 396 + ], + "blocks": [ + { + "bbox": [ + 307, + 347, + 342, + 396 + ], + "lines": [ + { + "bbox": [ + 307, + 347, + 342, + 396 + ], + "spans": [ + { + "bbox": [ + 307, + 347, + 342, + 396 + ], + "type": "image", + "image_path": "a26a02d0dba06544237ee3e4b6a0acb74db8d799728e3591759f34e48455f53a.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 343, + 347, + 410, + 396 + ], + "blocks": [ + { + "bbox": [ + 343, + 347, + 410, + 396 + ], + "lines": [ + { + "bbox": [ + 343, + 347, + 410, + 396 + ], + "spans": [ + { + "bbox": [ + 343, + 347, + 410, + 396 + ], + "type": "image", + "image_path": "ab35c6c485ee0a5ce02ca2a7e913fbf90633bdd8c76f21feb900f6d3247472b0.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 410, + 347, + 445, + 396 + ], + "blocks": [ + { + "bbox": [ + 410, + 347, + 445, + 396 + ], + "lines": [ + { + "bbox": [ + 410, + 347, + 445, + 396 + ], + "spans": [ + { + "bbox": [ + 410, + 347, + 445, + 396 + ], + "type": "image", + "image_path": "84d4b09054b4f4798fce5a91b9366cce15e3afad001a2510ea62e0114886d3ca.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 446, + 347, + 480, + 396 + ], + "blocks": [ + { + "bbox": [ + 446, + 347, + 480, + 396 + ], + "lines": [ + { + "bbox": [ + 446, + 347, + 480, + 396 + ], + "spans": [ + { + "bbox": [ + 446, + 347, + 480, + 396 + ], + "type": "image", + "image_path": "fc2c2b9165fa633d16ee2bf35eba67da8a51a736e4e008598402869af648fb15.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "bbox": [ + 130, + 474, + 480, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 474, + 480, + 545 + ], + "spans": [ + { + "bbox": [ + 130, + 474, + 480, + 545 + ], + "type": "text", + "content": "The loss formulation is slightly adapted to implement the following exemplary operations, which are applicable to any domain and can serve as the basis for additional processing methods or user-driven applications. Further details can be found in Sec. D (Appendix). Examples can be found in Fig. 7. We want to remove the monotonicity loss " + }, + { + "bbox": [ + 130, + 474, + 480, + 545 + ], + "type": "inline_equation", + "content": "L_{m}" + }, + { + "bbox": [ + 130, + 474, + 480, + 545 + ], + "type": "text", + "content": " and use equal input and output sizes for both the removal and the movement of an object." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 130, + 546, + 480, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 546, + 480, + 593 + ], + "spans": [ + { + "bbox": [ + 130, + 546, + 480, + 593 + ], + "type": "text", + "content": "Removal To remove a certain region, we employ an additional loss term that penalises any deformation target leading to that region. Points that were originally assigned to the target region will be distorted and assigned to neighbouring regions, resulting in the removal of the region in the output." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 130, + 594, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 481, + 665 + ], + "type": "text", + "content": "Moving an Object Although classic seam carving cannot move an object easily, as seams are assumed to cross the entire image, our formulation does not have this restriction. To move an object, we also enforce that the deformation towards the target feature only occurs from the desired new placement in the output. Any occurrence of that feature elsewhere is punished to avoid its repetition." + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 246, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 246, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 246, + 91, + 448, + 102 + ], + "type": "text", + "content": "Retargeting Visual Data with Deformation Fields" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 138, + 158, + 476, + 232 + ], + "blocks": [ + { + "bbox": [ + 130, + 114, + 482, + 149 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 482, + 149 + ], + "type": "text", + "content": "Table 1: Comparison of FID scores on the RetargetMe [42] dataset, for retargeting x (top) and y (bottom) axis to a smaller (left) and bigger (right) size. In mean, ours outperforms seam carving with 46.67 to 52.57 for x and 53.57 to 65.48 for y." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 138, + 158, + 476, + 232 + ], + "lines": [ + { + "bbox": [ + 138, + 158, + 476, + 232 + ], + "spans": [ + { + "bbox": [ + 138, + 158, + 476, + 232 + ], + "type": "table", + "html": "
x FID(↓)50%60%70%80%90%110%120%130%140%150%
Seam carving85.7566.4151.6735.2023.8020.8630.0137.3843.4546.37
Ours79.3958.0643.6930.5321.7111.1719.4229.1937.8749.74
y FID(↓)50%60%70%80%90%110%120%130%140%150%
Seam carving116.7084.3360.5940.8624.9223.0433.5840.8248.4952.59
Ours89.4467.7749.6235.5925.4213.5521.0227.5538.2046.40
", + "image_path": "7c248f1bfebfd0bb1a9fb3e5685ac6e338e99437b9999fd7a29151cd484d86c7.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 153, + 307, + 293, + 393 + ], + "blocks": [ + { + "bbox": [ + 137, + 243, + 307, + 297 + ], + "lines": [ + { + "bbox": [ + 137, + 243, + 307, + 297 + ], + "spans": [ + { + "bbox": [ + 137, + 243, + 307, + 297 + ], + "type": "text", + "content": "Table 2: User study comparing our approach to seam carving " + }, + { + "bbox": [ + 137, + 243, + 307, + 297 + ], + "type": "inline_equation", + "content": "(SC)" + }, + { + "bbox": [ + 137, + 243, + 307, + 297 + ], + "type": "text", + "content": " for images. 19 participants were questioned for preferences of random examples from the RetargetMe dataset [42], casting 1200 votes." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 153, + 307, + 293, + 393 + ], + "lines": [ + { + "bbox": [ + 153, + 307, + 293, + 393 + ], + "spans": [ + { + "bbox": [ + 153, + 307, + 293, + 393 + ], + "type": "table", + "html": "
WidthOursSCDraw
50 %48.48%20.87%30.67%
150 %33.8%36.27%29.92%
HeightOursSCDraw
50 %49.5%24.25%26.25%
150 %45.84%20.0%34.15%
Total44.57%25.1%30.32%
", + "image_path": "5845ebbaf4722a5914818415d2102639945e06180ac125473c61013b2905b322.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 313, + 264, + 392, + 326 + ], + "blocks": [ + { + "bbox": [ + 313, + 264, + 392, + 326 + ], + "lines": [ + { + "bbox": [ + 313, + 264, + 392, + 326 + ], + "spans": [ + { + "bbox": [ + 313, + 264, + 392, + 326 + ], + "type": "image", + "image_path": "e8604be8cf7942901512735e38ef12f1dcf5088957484b76f42deedc7e478f62.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 393, + 252, + 467, + 327 + ], + "lines": [ + { + "bbox": [ + 393, + 252, + 467, + 327 + ], + "spans": [ + { + "bbox": [ + 393, + 252, + 467, + 327 + ], + "type": "text", + "content": "Colour Complex Gradient Energy" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 306, + 327, + 476, + 372 + ], + "lines": [ + { + "bbox": [ + 306, + 327, + 476, + 372 + ], + "spans": [ + { + "bbox": [ + 306, + 327, + 476, + 372 + ], + "type": "text", + "content": "Fig. 10: Example of using our method with the more complex energy term of [48] instead of gradients, improving the results on an example." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 415, + 218, + 428 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 415, + 218, + 428 + ], + "spans": [ + { + "bbox": [ + 132, + 415, + 218, + 428 + ], + "type": "text", + "content": "4 Evaluation" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 443, + 482, + 599 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 443, + 482, + 599 + ], + "spans": [ + { + "bbox": [ + 130, + 443, + 482, + 599 + ], + "type": "text", + "content": "Retargeting not only lacks comparable methods for other visual data such as meshes, but its quality is also difficult to measure. In 2D, we compare our results to seam carving [2] on images using the RetargetMe [42] benchmark. In 3D, we compare our results on NeRFs with video-seam-carved recordings of the same scene, because video seam carving [43] is the closest method to 3D retargeting. We always produce a short, 3 second clip at 30 frames per second. To ensure a fair comparison with the method, we only tilt the camera down instead of rotating it around the object. Even for benevolent camera movements, it produces visible artefacts, as seen in Sec. F (Appendix). Exemplary, we show that our approach is compatible with approaches that propose different saliency maps as visualised in Fig. 2. For that purpose, we compare our method to other more recent image retargeting approaches visually (Fig. 10) and in a user study (Tab. 4). Additional results can be found in Sec. G (Appendix), including an explicit failure case." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 619, + 226, + 631 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 619, + 226, + 631 + ], + "spans": [ + { + "bbox": [ + 132, + 619, + 226, + 631 + ], + "type": "text", + "content": "4.1 Comparisons" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "content": "Images In accordance with recent publications in the generative model and image editing domain [15, 23, 38], we extensively compare our results with the" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 271, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 271, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 271, + 102 + ], + "type": "text", + "content": "T. Elsner, J. Berger et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 148, + 168, + 465, + 227 + ], + "blocks": [ + { + "bbox": [ + 130, + 114, + 482, + 158 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 482, + 158 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 482, + 158 + ], + "type": "text", + "content": "Table 3: User study comparing retargeting a scene, then recording it (ours), to recording it and then retargetet it (video seam carving [43], VSC), using our provided dataset. 21 participants were asked to indicated which solution they prefer, casting 252 votes in total." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 148, + 168, + 465, + 227 + ], + "lines": [ + { + "bbox": [ + 148, + 168, + 465, + 227 + ], + "spans": [ + { + "bbox": [ + 148, + 168, + 465, + 227 + ], + "type": "table", + "html": "
SceneOursVSCDrawOursVSCDrawOursVSCDrawOursVSCDraw
Beach1731201019111911
Farm2100210016322100
Harb.2100210021002100
", + "image_path": "3efd14c17b9d07df792725253ac3f278ac7d665fcfed22e24fd824893947c5ec.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 262, + 482, + 345 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 262, + 482, + 345 + ], + "spans": [ + { + "bbox": [ + 130, + 262, + 482, + 345 + ], + "type": "text", + "content": "original seam carving results using FID [20] in Tab. 1 for images. We also evaluate our approach in a user study. The participants were instructed to classify their preference for a retargeted version of a randomly selected image based on a randomly selected deformation axis, always choosing between our and the seam carved [2] version. We used the 80 images of the RetargetMe [42] for 2D, producing 360 variations per approach in Tab. 2. This shows how our approach can be a more powerful backbone for seam carving." + } + ] + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 137, + 426, + 321, + 477 + ], + "blocks": [ + { + "bbox": [ + 132, + 392, + 325, + 426 + ], + "lines": [ + { + "bbox": [ + 132, + 392, + 325, + 426 + ], + "spans": [ + { + "bbox": [ + 132, + 392, + 325, + 426 + ], + "type": "text", + "content": "Table 4: User study measuring preference, pairwise comparing to our approach using salience maps of [48]." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 137, + 426, + 321, + 477 + ], + "lines": [ + { + "bbox": [ + 137, + 426, + 321, + 477 + ], + "spans": [ + { + "bbox": [ + 137, + 426, + 321, + 477 + ], + "type": "table", + "html": "
ApproachOursTheirsDraw
CarvingNet [46]45.16%41.93%12.91%
RL [22]46.85%32.43%20.72%
Cycle-IR [49]40.3%34.46%25.24%
", + "image_path": "78c535b2c37002c628eed3542b4c6f150b7734e162a2e80093f7f67e92e8741f.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 331, + 357, + 482, + 501 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 331, + 357, + 482, + 501 + ], + "spans": [ + { + "bbox": [ + 331, + 357, + 482, + 501 + ], + "type": "text", + "content": "We further compare against other state-of-the-art retargeting approaches in a user study in Tab. 4: We show that our method can compete both against a seam carving variation with better saliency from deep learning [46] while also beating state-of-the-art approaches that retarget the image directly [22,49]. Note that we only compare to readily available images from other approach, as" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 501, + 482, + 597 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 501, + 482, + 597 + ], + "spans": [ + { + "bbox": [ + 130, + 501, + 482, + 597 + ], + "type": "text", + "content": "we often could not reproduce them or run the code. For this comparison, we use the saliency method FullGrad [48] that aggregates layer-wise bias-gradients from across ResNet [16]. This aggregated gradient map acts as our saliency map, as it measures how useful the feature is to the neural network, hence we use it as a measure for how semantically important a feature is. This also shows that our approach can work with different saliency maps. Details and examples can be found in Sec. E (Appendix). Additionally, we provide a visual ablation that shows the importance of each loss component in Fig. 11." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 617, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 665 + ], + "type": "text", + "content": "3D Scenes When assessing the quality of retargeting 3D scenes provided as NeRFs, we compared our results with what we consider to be the most similar approach, video seam carving [43]. In our provided GTA V NeRF retargeting dataset, we resize each axis by " + }, + { + "bbox": [ + 130, + 617, + 482, + 665 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 130, + 617, + 482, + 665 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 617, + 482, + 665 + ], + "type": "inline_equation", + "content": "150\\%" + }, + { + "bbox": [ + 130, + 617, + 482, + 665 + ], + "type": "text", + "content": ". The results can be found in Tab. 3." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 246, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 246, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 246, + 91, + 447, + 102 + ], + "type": "text", + "content": "Retargeting Visual Data with Deformation Fields" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 140, + 114, + 471, + 153 + ], + "blocks": [ + { + "bbox": [ + 140, + 114, + 471, + 153 + ], + "lines": [ + { + "bbox": [ + 140, + 114, + 471, + 153 + ], + "spans": [ + { + "bbox": [ + 140, + 114, + 471, + 153 + ], + "type": "image", + "image_path": "abfe2e99a4bc32fdc43ad0b6544da1d1f2e6d92af5130c2241f2a173e48df8c2.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 164, + 482, + 199 + ], + "lines": [ + { + "bbox": [ + 131, + 164, + 482, + 199 + ], + "spans": [ + { + "bbox": [ + 131, + 164, + 482, + 199 + ], + "type": "text", + "content": "Fig. 11: Visual ablation for retargeting to " + }, + { + "bbox": [ + 131, + 164, + 482, + 199 + ], + "type": "inline_equation", + "content": "80\\%" + }, + { + "bbox": [ + 131, + 164, + 482, + 199 + ], + "type": "text", + "content": " width on an image [12], from left to right: Wobbly contours from shearing (no " + }, + { + "bbox": [ + 131, + 164, + 482, + 199 + ], + "type": "inline_equation", + "content": "L_{s}" + }, + { + "bbox": [ + 131, + 164, + 482, + 199 + ], + "type": "text", + "content": "), uniform stretch on the whole image (no " + }, + { + "bbox": [ + 131, + 164, + 482, + 199 + ], + "type": "inline_equation", + "content": "L_{e}" + }, + { + "bbox": [ + 131, + 164, + 482, + 199 + ], + "type": "text", + "content": "), missing features e.g. the person (no " + }, + { + "bbox": [ + 131, + 164, + 482, + 199 + ], + "type": "inline_equation", + "content": "L_{m}" + }, + { + "bbox": [ + 131, + 164, + 482, + 199 + ], + "type": "text", + "content": "), content outside the image (no " + }, + { + "bbox": [ + 131, + 164, + 482, + 199 + ], + "type": "inline_equation", + "content": "L_{b}" + }, + { + "bbox": [ + 131, + 164, + 482, + 199 + ], + "type": "text", + "content": ")." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 220, + 480, + 245 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 220, + 480, + 245 + ], + "spans": [ + { + "bbox": [ + 131, + 220, + 480, + 245 + ], + "type": "text", + "content": "We provide additional results and frames of the resulting sequences from our approach for 3D scenes compared to video seam carving in Sec. F (Appendix)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 260, + 296, + 271 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 260, + 296, + 271 + ], + "spans": [ + { + "bbox": [ + 131, + 260, + 296, + 271 + ], + "type": "text", + "content": "4.2 Discussion and Limitations" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 278, + 482, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 278, + 482, + 518 + ], + "spans": [ + { + "bbox": [ + 130, + 278, + 482, + 518 + ], + "type": "text", + "content": "With our deformation fields as backbone, but the same energy term, our approach outperforms seam carving on both a qualitative and quantitative level. Our flexibility in the energy formulation allows for more possibilities, and the replacement of the gradient-based energy formulation with a better saliency map improves our approach even more. This shows that our method is a flexible backbone that can be used with various successors to seam carving, which provide more advanced saliency maps. Our continuous formulation allows for retargeting in various domains, including 3D scenes, and highlights the advantages of a geometry-aware approach. Applying our approach with only colour gradients can produce flawed outputs, just as those of seam carving, see Fig. 15 (Appendix). As we optimise a neural field for each individual input, we trade computation time for not needing a dataset to train and for a global optimisation. Our deformation optimisation takes approximately one minute on consumer hardware for good results versus about half a second for seam carving. Our strength, however, lies in 3D. Video seam carving on a 600 by 400 pixel video can take several hours, whereas our approach produces the final retargeted 3D scene in under 15 minutes. For Future Work, investigating invertible neural networks to implicitly learn an inverse deformation for the 3D case [1] seems promising, as would a deep learned approach to evaluating a deformation for optimisation and more research into good measures for retargeting quality." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 534, + 220, + 547 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 534, + 220, + 547 + ], + "spans": [ + { + "bbox": [ + 132, + 534, + 220, + 547 + ], + "type": "text", + "content": "5 Conclusion" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "text", + "content": "We present an approach that generalises the principle underlying seam carving to create image variants by manipulating low content areas. By training neural deformation fields that avoid applying distortion to important parts of an input, our approach is applicable to different forms of visual data retargeting. Our formulation is largely domain invariant and only needs minor alterations to work on different types of visual data, as demonstrated for neural radiance fields and meshes for the first time. We demonstrate compatibility to other energy formulations of the seam carving family, being a more general way of optimising image retargeting. Our approach is the only one that natively works on 3D scenes." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 271, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 271, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 271, + 101 + ], + "type": "text", + "content": "T. Elsner, J. Berger et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 197, + 126 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 197, + 126 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 197, + 126 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 141, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 22, + "blocks": [ + { + "bbox": [ + 138, + 141, + 481, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 141, + 481, + 185 + ], + "spans": [ + { + "bbox": [ + 138, + 141, + 481, + 185 + ], + "type": "text", + "content": "1. Ardizzone, L., Kruse, J., Wirkert, S.J., Rahner, D., Pellegrini, E.W., Klessen, R.S., Maier-Hein, L., Rother, C., Kothe, U.: Analyzing inverse problems with invertible neural networks. CoRR abs/1808.04730 (2018), http://arxiv.org/abs/1808.04730" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 186, + 481, + 208 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 186, + 481, + 208 + ], + "spans": [ + { + "bbox": [ + 138, + 186, + 481, + 208 + ], + "type": "text", + "content": "2. Avidan, S., Shamir, A.: Seam carving for content-aware image resizing. ACM Trans. Graph. 26(3), 10 (2007)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 209, + 481, + 231 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 209, + 481, + 231 + ], + "spans": [ + { + "bbox": [ + 138, + 209, + 481, + 231 + ], + "type": "text", + "content": "3. Basha, T.D., Moses, Y., Avidan, S.: Stereo seam carving a geometrically consistent approach. IEEE Trans. Pattern Anal. Mach. Intell. 35(10), 2513-2525 (2013)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 232, + 481, + 263 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 232, + 481, + 263 + ], + "spans": [ + { + "bbox": [ + 138, + 232, + 481, + 263 + ], + "type": "text", + "content": "4. Berkiten, S., Halber, M., Solomon, J., Ma, C., Li, H., Rusinkiewicz, S.: Learning detail transfer based on geometric features. Comput. Graph. Forum 36(2), 361-373 (2017)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 264, + 481, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 264, + 481, + 297 + ], + "spans": [ + { + "bbox": [ + 138, + 264, + 481, + 297 + ], + "type": "text", + "content": "5. Cai, H., Feng, W., Feng, X., Wang, Y., Zhang, J.: Neural surface reconstruction of dynamic scenes with monocular rgb-d camera. In: Neural Information Processing Systems (NeurIPS) (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 298, + 481, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 298, + 481, + 319 + ], + "spans": [ + { + "bbox": [ + 138, + 298, + 481, + 319 + ], + "type": "text", + "content": "6. Chen, Zhiqin and Zhang, Hao: Learning implicit fields for generative shape modeling. In: CVPR. pp. 5939-5948 (2019)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 320, + 481, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 320, + 481, + 342 + ], + "spans": [ + { + "bbox": [ + 138, + 320, + 481, + 342 + ], + "type": "text", + "content": "7. Dong, W., Bao, G., Zhang, X., Paul, J.: Fast multi-operator image resizing and evaluation. J. Comput. Sci. Technol. 27(1), 121-134 (2012)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 342, + 481, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 342, + 481, + 385 + ], + "spans": [ + { + "bbox": [ + 138, + 342, + 481, + 385 + ], + "type": "text", + "content": "8. Dong, W., Zhou, N., Lee, T., Wu, F., Kong, Y., Zhang, X.: Summarization-based image resizing by intelligent object carving. IEEE Trans. Vis. Comput. Graph. 20(1), 1 (2014). https://doi.org/10.1109/TVCG.2013.103, https://doi.org/10.1109/TVCG.2013.103" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 387, + 481, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 387, + 481, + 408 + ], + "spans": [ + { + "bbox": [ + 138, + 387, + 481, + 408 + ], + "type": "text", + "content": "9. Dong, W., Zhou, N., Paul, J., Zhang, X.: Optimized image resizing using seam carving and scaling. ACM Trans. Graph. 28(5), 125 (2009)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 410, + 481, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 410, + 481, + 442 + ], + "spans": [ + { + "bbox": [ + 138, + 410, + 481, + 442 + ], + "type": "text", + "content": "0. Efros, A.A., Freeman, W.T.: Image quilting for texture synthesis and transfer. In: Pocock, L. (ed.) Proceedings of the 28th Annual Conference on Computer Graphics and Interactive Techniques, SIGGRAPH. pp. 341-346. ACM (2001)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 443, + 481, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 443, + 481, + 464 + ], + "spans": [ + { + "bbox": [ + 138, + 443, + 481, + 464 + ], + "type": "text", + "content": "1. Efros, A.A., Leung, T.K.: Texture synthesis by non-parametric sampling. In: ICCV. pp. 1033-1038 (1999)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 465, + 481, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 465, + 481, + 487 + ], + "spans": [ + { + "bbox": [ + 138, + 465, + 481, + 487 + ], + "type": "text", + "content": "2. at English Wikipedia, N.: Broadway tower (2007), https://commons.wikimedia.org/wiki/File:Broadway_tower.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 487, + 481, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 487, + 481, + 520 + ], + "spans": [ + { + "bbox": [ + 138, + 487, + 481, + 520 + ], + "type": "text", + "content": "3. Garbin, S.J., Kowalski, M., Estellers, V., Szymanowicz, S., Rezaeifar, S., Shen, J., Johnson, M., Valentin, J.: Voltemorph: Realtime, controllable and generalisable animation of volumetric representations. arXiv:2208.00949 (2022)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 521, + 481, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 521, + 481, + 542 + ], + "spans": [ + { + "bbox": [ + 138, + 521, + 481, + 542 + ], + "type": "text", + "content": "4. Gatys, L.A., Ecker, A.S., Bethge, M.: Texture synthesis using convolutional neural networks. In: Neural Information Processing Systems. pp. 262-270 (2015)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 543, + 481, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 543, + 481, + 564 + ], + "spans": [ + { + "bbox": [ + 138, + 543, + 481, + 564 + ], + "type": "text", + "content": "5. Gu, J., Zhai, S., Zhang, Y., Susskind, J., Jaitly, N.: Matryoshka diffusion models. arXiv preprint arXiv:2310.15111 (2023)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 565, + 481, + 587 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 565, + 481, + 587 + ], + "spans": [ + { + "bbox": [ + 138, + 565, + 481, + 587 + ], + "type": "text", + "content": "6. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. CoRR abs/1512.03385 (2015), http://arxiv.org/abs/1512.03385" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 138, + 588, + 481, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 588, + 481, + 609 + ], + "spans": [ + { + "bbox": [ + 138, + 588, + 481, + 609 + ], + "type": "text", + "content": "7. Henzler, P., Deschaintre, V., Mitra, N.J., Ritschel, T.: Generative modelling of BRDF textures from flash images. ACM Trans. Graph. 40(6), 284:1-284:13 (2021)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 138, + 610, + 481, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 610, + 481, + 643 + ], + "spans": [ + { + "bbox": [ + 138, + 610, + 481, + 643 + ], + "type": "text", + "content": "8. Henzler, P., Mitra, N.J., Ritschel, T.: Learning a neural 3d texture space from 2d exemplars. In: The IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2019)" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 138, + 643, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 643, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 643, + 481, + 665 + ], + "type": "text", + "content": "9. Hertz, A., Hanocka, R., Giryes, R., Cohen-Or, D.: Deep geometric texture synthesis. ACM Trans. Graph. 39(4), 108:1-108:11 (2020)" + } + ] + } + ], + "index": 21 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 246, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 246, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 246, + 91, + 447, + 102 + ], + "type": "text", + "content": "Retargeting Visual Data with Deformation Fields" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 666 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 480, + 149 + ], + "type": "text", + "content": "20. Heusel, M., Ramsauer, H., Unterthiner, T., Nessler, B., Hochreiter, S.: Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems 30 (2017)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 150, + 480, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 150, + 480, + 183 + ], + "spans": [ + { + "bbox": [ + 132, + 150, + 480, + 183 + ], + "type": "text", + "content": "21. Huang, Y., Cao, Y., Lai, Y., Shan, Y., Gao, L.: Nerf-texture: Texture synthesis with neural radiance fields. In: ACM SIGGRAPH 2023 Conference Proceedings. pp. 43:1-43:10 (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 183, + 480, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 183, + 480, + 215 + ], + "spans": [ + { + "bbox": [ + 132, + 183, + 480, + 215 + ], + "type": "text", + "content": "22. Kajiura, N., Kosugi, S., Wang, X., Yamasaki, T.: Self-play reinforcement learning for fast image retargeting. CoRR abs/2010.00909 (2020), https://arxiv.org/abs/2010.00909" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 216, + 480, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 216, + 480, + 258 + ], + "spans": [ + { + "bbox": [ + 132, + 216, + 480, + 258 + ], + "type": "text", + "content": "23. Kawar, B., Zada, S., Lang, O., Tov, O., Chang, H., Dekel, T., Mosseri, I., Irani, M.: Imagic: Text-based real image editing with diffusion models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6007-6017 (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 259, + 480, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 259, + 480, + 293 + ], + "spans": [ + { + "bbox": [ + 132, + 259, + 480, + 293 + ], + "type": "text", + "content": "24. Kopf, J., Fu, C.W., Cohen-Or, D., Deussen, O., Lischinski, D., Wong, T.T.: Solid texture synthesis from 2d exemplars. ACM Transactions on Graphics 26(3), 2:1-2:9 (2007)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 293, + 480, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 293, + 480, + 326 + ], + "spans": [ + { + "bbox": [ + 132, + 293, + 480, + 326 + ], + "type": "text", + "content": "25. Lai, Y., Hu, S., Gu, X., Martin, R.R.: Geometric texture synthesis and transfer via geometry images. In: Proceedings of the Tenth ACM Symposium on Solid and Physical Modeling. pp. 15-26. ACM (2005)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 326, + 480, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 326, + 480, + 358 + ], + "spans": [ + { + "bbox": [ + 132, + 326, + 480, + 358 + ], + "type": "text", + "content": "26. Liu, F., Gleicher, M.: Automatic image retargeting with fisheye-view warping. In: Proceedings of the 18th Annual ACM Symposium on User Interface Software and Technology. pp. 153-162. ACM (2005)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 358, + 480, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 358, + 480, + 392 + ], + "spans": [ + { + "bbox": [ + 132, + 358, + 480, + 392 + ], + "type": "text", + "content": "27. Liu, F., Gleicher, M.: Video retargeting: automating pan and scan. In: Nahrstedt, K., Turk, M.A., Rui, Y., Klas, W., Mayer-Patel, K. (eds.) Proceedings of the 14th ACM International Conference on Multimedia. pp. 241-250. ACM (2006)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 392, + 480, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 392, + 480, + 413 + ], + "spans": [ + { + "bbox": [ + 132, + 392, + 480, + 413 + ], + "type": "text", + "content": "28. Luiten, J., Kopanas, G., Leibe, B., Ramanan, D.: Dynamic 3d gaussians: Tracking by persistent dynamic view synthesis. In: 3DV (2024)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 413, + 480, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 413, + 480, + 445 + ], + "spans": [ + { + "bbox": [ + 132, + 413, + 480, + 445 + ], + "type": "text", + "content": "29. Mescheder, L., Oechsle, M., Niemeyer, M., Nowozin, S., Geiger, A.: Occupancy networks: Learning 3D reconstruction in function space. In: CVPR. pp. 4460-4470 (2019)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 446, + 480, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 446, + 480, + 479 + ], + "spans": [ + { + "bbox": [ + 132, + 446, + 480, + 479 + ], + "type": "text", + "content": "30. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. In: ECCV (2020)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 479, + 480, + 522 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 479, + 480, + 522 + ], + "spans": [ + { + "bbox": [ + 132, + 479, + 480, + 522 + ], + "type": "text", + "content": "31. Moreira, T.P., Santana, M.C.S., Passos, L.A., Papa, J.P., da Costa, K.A.P.: An end-to-end approach for seam carving detection using deep neural networks. In: Pattern Recognition and Image Analysis. Lecture Notes in Computer Science, vol. 13256, pp. 447-457 (2022)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 522, + 480, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 522, + 480, + 555 + ], + "spans": [ + { + "bbox": [ + 132, + 522, + 480, + 555 + ], + "type": "text", + "content": "32. Nam, S., Ahn, W., Yu, I., Kwon, M., Son, M., Lee, H.: Deep convolutional neural network for identifying seam-carving forgery. IEEE Trans. Circuits Syst. Video Technol. 31(8), 3308-3326 (2021)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 555, + 480, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 555, + 480, + 588 + ], + "spans": [ + { + "bbox": [ + 132, + 555, + 480, + 588 + ], + "type": "text", + "content": "33. Nataraj, L., Gudavalli, C., Mohammed, T.M., Chandrasekaran, S., Manjunath, B.S.: Seam carving detection and localization using two-stage deep neural networks. CoRR abs/2109.01764 (2021)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 588, + 480, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 588, + 480, + 599 + ], + "spans": [ + { + "bbox": [ + 132, + 588, + 480, + 599 + ], + "type": "text", + "content": "34. North, R.: Grand theft auto v. Steam (2015), https://www.rockstargames.com/V/" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 600, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 600, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 600, + 480, + 632 + ], + "type": "text", + "content": "35. Park, J.J., Florence, P., Straub, J., Newcombe, R., Lovegrove, S.: DeepSDF: Learning continuous signed distance functions for shape representation. In: CVPR. pp. 165-174 (2019)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 632, + 480, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 480, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 480, + 666 + ], + "type": "text", + "content": "36. Park, K., Sinha, U., Barron, J.T., Bouaziz, S., Goldman, D.B., Seitz, S.M., MartinBrualla, R.: Nerfies: Deformable neural radiance fields. In: ICCV. pp. 5845-5854 (2021)" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 271, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 271, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 271, + 101 + ], + "type": "text", + "content": "T. Elsner, J. Berger et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 666 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 162 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 162 + ], + "type": "text", + "content": "37. Park, K., Sinha, U., Hedman, P., Barron, J.T., Bouaziz, S., Goldman, D.B., Martin-Brualla, R., Seitz, S.M.: Hypernerf: a higher-dimensional representation for topologically varying neural radiance fields. ACM Trans. Graph. 40(6), 238:1-238:12 (2021)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 162, + 481, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 162, + 481, + 194 + ], + "spans": [ + { + "bbox": [ + 130, + 162, + 481, + 194 + ], + "type": "text", + "content": "38. Peebles, W., Xie, S.: Scalable diffusion models with transformers. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 4195-4205 (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 194, + 481, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 194, + 481, + 226 + ], + "spans": [ + { + "bbox": [ + 132, + 194, + 481, + 226 + ], + "type": "text", + "content": "39. Peng, Y., Yan, Y., Liu, S., Cheng, Y., Guan, S., Pan, B., Zhai, G., Yang, X.: Cagenerf: Cage-based neural radiance fields for generalized 3d deformation and animation. In: Advances in Neural Information Processing Systems (2022)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 227, + 481, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 227, + 481, + 259 + ], + "spans": [ + { + "bbox": [ + 132, + 227, + 481, + 259 + ], + "type": "text", + "content": "40. Pumarola, A., Corona, E., Pons-Moll, G., Moreno-Noguer, F.: D-NeRF: Neural Radiance Fields for Dynamic Scenes. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2021)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 259, + 481, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 259, + 481, + 293 + ], + "spans": [ + { + "bbox": [ + 132, + 259, + 481, + 293 + ], + "type": "text", + "content": "41. Richter, S.R., AlHaija, H.A., Koltun, V.: Enhancing photorealism enhancement. IEEE Transactions on Pattern Analysis and Machine Intelligence 45(2), 1700-1715 (2022)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 293, + 481, + 325 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 293, + 481, + 325 + ], + "spans": [ + { + "bbox": [ + 132, + 293, + 481, + 325 + ], + "type": "text", + "content": "42. Rubinstein, M., Gutierrez, D., Sorkine, O., Shamir, A.: A comparative study of image retargeting. ACM Transactions on Graphics (Proc. SIGGRAPH ASIA) 29(6), 160:1-160:10 (2010)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 326, + 481, + 347 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 326, + 481, + 347 + ], + "spans": [ + { + "bbox": [ + 132, + 326, + 481, + 347 + ], + "type": "text", + "content": "43. Rubinstein, M., Shamir, A., Avidan, S.: Improved seam carving for video retargeting. ACM Trans. Graph. 27(3), 16 (2008)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 348, + 481, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 348, + 481, + 380 + ], + "spans": [ + { + "bbox": [ + 132, + 348, + 481, + 380 + ], + "type": "text", + "content": "44. Rubinstein, M., Shamir, A., Avidan, S.: Multi-operator media retargeting. ACM Trans. Graph. 28(3), 23 (2009). https://doi.org/10.1145/1531326.1531329, https://doi.org/10.1145/1531326.1531329" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 380, + 481, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 380, + 481, + 403 + ], + "spans": [ + { + "bbox": [ + 132, + 380, + 481, + 403 + ], + "type": "text", + "content": "45. Shaham, T.R., Dekel, T., Michaeli, T.: Singan: Learning a generative model from a single natural image. In: ICCV. pp. 4569-4579 (2019)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 403, + 481, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 403, + 481, + 435 + ], + "spans": [ + { + "bbox": [ + 132, + 403, + 481, + 435 + ], + "type": "text", + "content": "46. Song, E., Lee, M., Lee, S.: Carvingnet: Content-guided seam carving using deep convolution neural network. IEEE Access 7, 284-292 (2019). https://doi.org/10.1109/ACCESS.2018.2885347, https://doi.org/10.1109/ACCESS.2018.2885347" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 435, + 481, + 457 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 435, + 481, + 457 + ], + "spans": [ + { + "bbox": [ + 132, + 435, + 481, + 457 + ], + "type": "text", + "content": "47. Sorkine-Hornung, O., Alexa, M.: As-rigid-as-possible surface modeling. In: Symposium on Geometry Processing (2007)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 457, + 481, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 457, + 481, + 479 + ], + "spans": [ + { + "bbox": [ + 132, + 457, + 481, + 479 + ], + "type": "text", + "content": "48. Srinivas, S., Fleuret, F.: Full-gradient representation for neural network visualization. Advances in neural information processing systems 32 (2019)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 479, + 481, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 479, + 481, + 502 + ], + "spans": [ + { + "bbox": [ + 132, + 479, + 481, + 502 + ], + "type": "text", + "content": "49. Tan, W., Yan, B., Lin, C., Niu, X.: Cycle-ir: Deep cyclic image retargeting. IEEE Transactions on Multimedia 22(7), 1730–1743 (2019)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 502, + 481, + 534 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 502, + 481, + 534 + ], + "spans": [ + { + "bbox": [ + 132, + 502, + 481, + 534 + ], + "type": "text", + "content": "50. Wang, P., Liu, L., Liu, Y., Theobalt, C., Komura, T., Wang, W.: Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. In: Advances in Neural Information Processing Systems. pp. 27171-27183 (2021)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 534, + 481, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 534, + 481, + 567 + ], + "spans": [ + { + "bbox": [ + 132, + 534, + 481, + 567 + ], + "type": "text", + "content": "51. Wei, L., Levoy, M.: Fast texture synthesis using tree-structured vector quantization. In: Proceedings of the 27th Annual Conference on Computer Graphics and Interactive Techniques, SIGGRAPH. pp. 479-488. ACM (2000)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 567, + 481, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 567, + 481, + 600 + ], + "spans": [ + { + "bbox": [ + 132, + 567, + 481, + 600 + ], + "type": "text", + "content": "52. Wu, G., Yi, T., Fang, J., Xie, L., Zhang, X., Wei, W., Liu, W., Tian, Q., Xinggang, W.: 4d gaussian splatting for real-time dynamic scene rendering. arXiv preprint arXiv:2310.08528 (2023)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 601, + 481, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 601, + 481, + 622 + ], + "spans": [ + { + "bbox": [ + 132, + 601, + 481, + 622 + ], + "type": "text", + "content": "53. Wu, H., Wang, Y., Feng, K., Wong, T., Lee, T., Heng, P.: Resizing by symmetrysummarization. ACM Trans. Graph. 29(6), 159 (2010)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 622, + 481, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 622, + 481, + 643 + ], + "spans": [ + { + "bbox": [ + 132, + 622, + 481, + 643 + ], + "type": "text", + "content": "54. Wu, R., Liu, R., Vondrick, C., Zheng, C.: Sin3dm: Learning a diffusion model from a single 3d textured shape. CoRR abs/2305.15399 (2023)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 132, + 644, + 481, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 644, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 644, + 481, + 666 + ], + "type": "text", + "content": "55. Wu, R., Zheng, C.: Learning to generate 3d shapes from a single example. ACM Transactions on Graphics (TOG) 41(6) (2022)" + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 246, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 246, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 246, + 91, + 447, + 102 + ], + "type": "text", + "content": "Retargeting Visual Data with Deformation Fields" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 237 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 481, + 161 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 481, + 161 + ], + "type": "text", + "content": "56. Xian, W., Sangkloy, P., Agrawal, V., Raj, A., Lu, J., Fang, C., Yu, F., Hays, J.: Texturegan: Controlling deep image synthesis with texture patches. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (June 2018)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 162, + 481, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 162, + 481, + 182 + ], + "spans": [ + { + "bbox": [ + 132, + 162, + 481, + 182 + ], + "type": "text", + "content": "57. Xu, T., Harada, T.: Deforming radiance fields with cages. In: ECCV. pp. 159-175 (2022)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 183, + 481, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 183, + 481, + 215 + ], + "spans": [ + { + "bbox": [ + 132, + 183, + 481, + 215 + ], + "type": "text", + "content": "58. Yang, Z., Gao, X., Zhou, W., Jiao, S., Zhang, Y., Jin, X.: Deformable 3d gaussians for high-fidelity monocular dynamic scene reconstruction. arXiv preprint arXiv:2309.13101 (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 216, + 481, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 216, + 481, + 237 + ], + "spans": [ + { + "bbox": [ + 132, + 216, + 481, + 237 + ], + "type": "text", + "content": "59. Yuan, Y.J., Sun, Y.T., Lai, Y.K., Ma, Y., Jia, R., Gao, L.: Nerf-editing: Geometry editing of neural radiance fields. In: CVPR. pp. 18332-18343 (2022)" + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 271, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 271, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 271, + 102 + ], + "type": "text", + "content": "T. Elsner, J. Berger et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Rethinking Data Augmentation for Robust LiDAR Semantic Segmentation in Adverse Weather/60de04d8-9845-4ba5-9238-a2dd5b5c4b07_content_list.json b/2024/Rethinking Data Augmentation for Robust LiDAR Semantic Segmentation in Adverse Weather/60de04d8-9845-4ba5-9238-a2dd5b5c4b07_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..e6bb300bfa5388fcece6df76964d2fc0e4f5b989 --- /dev/null +++ b/2024/Rethinking Data Augmentation for Robust LiDAR Semantic Segmentation in Adverse Weather/60de04d8-9845-4ba5-9238-a2dd5b5c4b07_content_list.json @@ -0,0 +1,1522 @@ +[ + { + "type": "text", + "text": "Rethinking Data Augmentation for Robust LiDAR Semantic Segmentation in Adverse Weather", + "text_level": 1, + "bbox": [ + 258, + 140, + 743, + 205 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Junsung Park $^{1}$ , Kyungmin Kim $^{2}$ , and Hyunjung Shim $^{1\\star}$", + "bbox": [ + 295, + 234, + 702, + 251 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Korea Advanced Institute of Science & Technology (KAIST) \n $^{2}$ Yonsei University \njshackist@kaist.ac.kr, kyungmin.kim@yonsei.ac.kr, kateshim@kaist.ac.kr", + "bbox": [ + 228, + 261, + 771, + 305 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Existing LiDAR semantic segmentation methods often struggle with performance declines in adverse weather conditions. Previous work has addressed this issue by simulating adverse weather or employing universal data augmentation during training. However, these methods lack a detailed analysis and understanding of how adverse weather negatively affects LiDAR semantic segmentation performance. Motivated by this issue, we identified key factors of adverse weather and conducted a toy experiment to pinpoint the main causes of performance degradation: (1) Geometric perturbation due to refraction caused by fog or droplets in the air and (2) Point drop due to energy absorption and occlusions. Based on these findings, we propose new strategic data augmentation techniques. First, we introduced a Selective Jittering (SJ) that jitters points in the random range of depth (or angle) to mimic geometric perturbation. Additionally, we developed a Learnable Point Drop (LPD) to learn vulnerable erase patterns with a Deep Q-Learning Network to approximate the point drop phenomenon from adverse weather conditions. Without precise weather simulation, these techniques strengthen the LiDAR semantic segmentation model by exposing it to vulnerable conditions identified by our data-centric analysis. Experimental results confirmed the suitability of the proposed data augmentation methods for enhancing robustness against adverse weather conditions. Our method achieves a notable $39.5\\mathrm{mIoU}$ on the SemanticKITTI-to-SemanticSTF benchmark, improving the baseline by $8.1\\%$ p and establishing a new state-of-the-art. Our code will be released at https://github.com/engineerJPark/LiDARWeather.", + "bbox": [ + 261, + 333, + 738, + 680 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: LiDAR Semantic Segmentation $\\cdot$ Robust Learning for Adverse Weather $\\cdot$ Data Augmentation", + "bbox": [ + 261, + 691, + 740, + 720 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 743, + 375, + 758 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "LiDAR semantic segmentation is a fundamental task of 3D scene understanding, particularly in safety-critical applications like autonomous driving. However, existing LiDAR segmentation models [1,5,14,21,28,37] commonly lack robustness,", + "bbox": [ + 212, + 771, + 785, + 818 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Corresponding author: Hyunjung Shim", + "bbox": [ + 217, + 824, + 495, + 839 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "showing significant performance degradation under adverse weather conditions such as snow, fog, rain, or wet surfaces.", + "bbox": [ + 212, + 146, + 782, + 176 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address this issue, recent studies [11, 33, 34] have introduced corruption benchmarks for adverse weather and proposed effective techniques for robust LiDAR segmentation against corruption. These robust methods are primarily divided into task-agnostic and simulation-based approaches. The task-agnostic approach [11, 25, 33] employs general machine-learning techniques for robustness without explicitly addressing LiDAR corruption caused by adverse weather. The simulation-based approach [3,7-9,23,26,35] artificially synthesizes weather-specific data using physical equations for training. However, these efforts focus on detection tasks and consider only a single type of weather at a time.", + "bbox": [ + 212, + 180, + 785, + 316 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "While the simulation-based approach leverages the intrinsic properties of Li-DAR scans under target adverse weather conditions, simulating every weather type at all possible severities is impractical and often inaccurate. Instead, we adopt a data-centric perspective to analyze corrupted LiDAR data. For example, distortions caused by rain, snow, and fog often create similar patterns, as simulated in [9]. That is, all those adverse weather shows point drop patterns due to attenuation or occlusion by droplet [6,7,9]. While the simulation-based approach requires explicit modeling of the LiDAR for each weather condition, our data-centric approach potentially addresses complex distortions with a few unified patterns. Through existing research and our comprehensive analysis, we find that various adverse weather conditions create similar distortion patterns in Li-DAR data. Based on our analysis, we categorize these distortions into two types: (1) geometric perturbation and (2) point drop. Our toy experiments demonstrate that these distortions are closely related to segmentation performance degradation.", + "bbox": [ + 212, + 319, + 787, + 545 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Rooted by the toy experiment, we introduce two novel and strategic data augmentation methods tailored to the LiDAR distortion caused by adverse weather. By incorporating these augmentations during training, we aim to bolster the model's robustness for each distortion type. For geometric perturbation, we introduce Selective Jittering (SJ). This method applies Additive Gaussian Noise (AGN) to alter the XYZ-coordinate and intensity within a selective local region. To handle a point drop pattern, we developed a Learnable Point Drop (LPD) that employs a Deep Q-Network (DQN) to strategically remove points. Our augmentation strategies are inspired by existing studies [9, 16, 27], which shows geometric perturbation typically involves small, random alterations to the original points. This insight led to our choice of jittering as a suitable technique for addressing the geometric perturbation. Also, our augmentation strategy for point drop is motivated by existing studies [7-9] which state that intensity and depth of LiDAR beams play a significant role. This prompts us to use a DQN to detect point drops and adapt to such point drop patterns.", + "bbox": [ + 212, + 549, + 787, + 776 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Finally, we apply the proposed augmentations when training the LiDAR segmentation model with a clean source dataset. Experimental results demonstrate that our method achieves new state-of-the-art performance on SemanticSTF, a real adverse weather dataset. Specifically, we achieved an impressive $+8.1\\%$", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "J.Park et al.", + "bbox": [ + 271, + 114, + 354, + 127 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "mIoU gain over the baseline MinkowskiNet, nearly tripling the $+2.5\\%$ p mIoU improvement from previous work [33]. Our method achieved remarkable performance gains, ranging from $+0.4$ to $+10.3\\%$ p mIoU, across various model architectures (Minkowski, CENet, SPVCNN) and corruption dataset (SemanticSTF and SemanticKITTI-C). These results demonstrate the generalizability of the proposed augmentation methods.", + "bbox": [ + 212, + 146, + 782, + 234 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In summary, this paper presents several key contributions.", + "bbox": [ + 238, + 237, + 660, + 252 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We identify two prevalent distortion types in LiDAR data caused by adverse weather, leading to performance degradation, through a data-centric analysis.", + "- We introduce two novel data augmentations tailored to each identified distortion type.", + "- Our method sets new state-of-the-art benchmarks on the SemanticKITTI to SemanticSTF benchmark. Notably, our improvement was $+8.1\\%$ p mIoU without relying on precise simulations of adverse weather in LiDAR point inputs. This represents a tripling of the improvement over the baseline, compared to what previous methods have achieved [33]." + ], + "bbox": [ + 223, + 261, + 782, + 411 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Works", + "text_level": 1, + "bbox": [ + 215, + 435, + 395, + 450 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.1 LiDAR Semantic Segmentation", + "text_level": 1, + "bbox": [ + 214, + 467, + 517, + 482 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Existing 3D LiDAR point cloud semantic segmentation methods can be categorized into three types based on the data representation: Point-based, Projection-based, and Voxel-based.", + "bbox": [ + 212, + 492, + 782, + 535 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Point-based methods [22, 28, 36] utilize the 3D points directly as input. KPConv [28] initially clustering local points, aggregating these local features, and then feeding them into kernel point convolutions. The Point Transformer [36] utilizes a transformer architecture to compute query points in each local region, obtained through k-nearest neighbors (kNN). The Point-Mixer [4] tried to adapt the MLP-Mixer [29] for point cloud applications. They achieve high performance but suffer from high computational costs due to the utilization of large-scale raw LiDAR data.", + "bbox": [ + 212, + 537, + 784, + 656 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Projection-based methods [1, 10, 18] project LiDAR points into a 2D image and performs the semantic segmentation using architectures successful in a 2D image. RangeViT [1] directly adopts a ViT model pre-trained on 2D images, demonstrating that pre-trained power in 2D images can be effective as prior knowledge in range images. RangeFormer [10] proposes \"RangeAug\" to maximize the utility of range images created by projecting into 2D, producing multiple range image data to overcome the low performance in range image models. Projection-based methods achieve fast inference speed but present a sub-optimal performance due to missing information after the projection.", + "bbox": [ + 212, + 657, + 784, + 792 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Voxel-based methods [5,14,37] perform efficient computations by dividing 3D space into a voxel grid and aggregating points within the same voxel. Minkowski-iNet [5] voxelizes LiDAR points with a cubic grid and applies sparse convolution.", + "bbox": [ + 212, + 795, + 782, + 839 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "Rethinking Data Augmentation for Robust LiDAR Semantic Segmentation", + "bbox": [ + 232, + 114, + 730, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Cylinder3D [37] proposes cylindrical partitions, reflecting LiDAR's characteristic that the density of points depends on the distance. SphereFormer [14] utilizes a radial window and transformer architecture to aggregate long-range information and improve performance. Voxel-based methods achieve a balance between reasonable inference time and commendable segmentation performance.", + "bbox": [ + 215, + 145, + 785, + 220 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.2 LiDAR Data Augmentation", + "text_level": 1, + "bbox": [ + 215, + 243, + 491, + 258 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Inspired by 2D image augmentation, conventional LiDAR segmentation methods apply classic scaling, rotation, flipping, and translation to augment LiDAR data. Recently, several out-of-context augmentation techniques [12, 20, 30] that mix different LiDAR scans have been proposed. Mix3D [20] combines randomly selected two scans. Considering the sweeping mechanism of the LiDAR sensor, PolarMix [30] cuts LiDAR scans along the azimuth axis, then exchanges point cloud sectors and applies instance-level rotate-pasting. To reflect the spatial prior of the LiDAR point cloud, LaserMix [12] partitions LiDAR scans based on the laser beams and blends partitions from different LiDAR scans. Recent study [24] introduces a fast LiDAR domain augmentation module to address sensor-bias problems. To the best of our knowledge, our approach is the first augmentation method specifically designed to address data corruption under adverse weather conditions.", + "bbox": [ + 215, + 268, + 785, + 465 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.3 LiDAR under Adverse Weather Conditions", + "text_level": 1, + "bbox": [ + 215, + 486, + 617, + 502 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Robustness under harsh conditions is crucial in safety-critical applications. Adverse weather substantially degrade performance in real-world outdoor autonomous driving. Thus, there are several attempts to develop weather-robust models in fields such as 2D segmentation [15, 17], 3D detection [7, 8, 11], and 3D segmentation [11, 33]. Simulation-based approaches [7, 8] artificially synthesize data for single weather conditions through physical modeling and utilize it for training. We differ from these in that we do not model specific weather conditions explicitly. Recently proposed task-agnostic approaches [11, 25, 33] consider multiple weather conditions at once. However, they use general machine-learning techniques (such as teacher-student framework and feature prototype) to achieve robustness rather than specifically tackle LiDAR corruption caused by adverse weather. We differ from these in that we propose augmentation methods specifically tailored for adverse weather conditions based on the analysis of performance degradation in LiDAR data.", + "bbox": [ + 215, + 512, + 800, + 724 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3 Finding Distortions to Augment", + "text_level": 1, + "bbox": [ + 215, + 746, + 563, + 765 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this section, we aim to discuss the patterns of distortion that different adverse weather conditions impose on LiDAR data. Although adverse weather conditions are distinct in reality, studies have shown that their effects on LiDAR data often result in similar impacts. For instance, the distortions caused by \"rain\", \"snow\",", + "bbox": [ + 215, + 777, + 785, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "J.Park et al.", + "bbox": [ + 271, + 114, + 354, + 127 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "and \"fog\" tend to produce similar point-missing patterns due to attenuation in the data, as demonstrated in [9]. Therefore, this section will focus on identifying the common distortion patterns caused by adverse weather through existing studies. Overall, existing studies describe the effect of adverse weather as four different types of distortion: (1) Point Drop due to energy absorption, (2) Occlusions caused by droplets of rain or snow and fog, (3) Geometric perturbation, and (4) Intensity distortion due to energy absorption.", + "bbox": [ + 212, + 146, + 787, + 252 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1 Distortion Factors from Adverse Weather", + "text_level": 1, + "bbox": [ + 214, + 273, + 604, + 287 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "(D1) Point Drop. Several studies have explored how adverse weather conditions contribute to point drops in LiDAR data. Kilic et al. [9], Fersch et al. [6] and Shin et al. [26] describe point drops resulting from beam attenuation and beam missing due to droplets, fog, and frozen or wet ground. These studies collectively suggest that adverse weather conditions typically lead to point drops in LiDAR data.", + "(D2) Occlusions. Several studies have addressed occlusions caused by adverse weather conditions. Hahner et al. [7], Kilic et al. [9], Kong et al. [11] and Yan et al. [34] consider scenarios where beams colliding with snow collect signals at much shorter distances than the intended objects of collision. Upon reviewing these studies, we have concluded that adverse weather consistently leads to occlusions.", + "(D3) Geometric Perturbation. Some studies focus on geometric perturbation caused by adverse weather conditions. Kilic et al. [9], Li et al. [16] and Smith et al. [27] demonstrated geometric perturbation in adverse weather, such as fog, snow, and rain, by incorporating random noise into the coordinates. Through these studies, we have come to conclude that adverse weather universally causes geometric perturbation.", + "(D4) Intensity Distortion. Numerous studies have focused on intensity distortion caused by adverse weather conditions. Bijelic et al. [3], Shin et al. [26], Fersch et al. [6], Kong et al. [11], and Yan et al. [34] have collectively shown that adverse weather conditions like fog, wetness, and rain lead to a reduction in Li-DAR beam intensity, influencing the generation of synthetic data. Through these studies, we conclude that adverse weather commonly causes intensity distortion." + ], + "bbox": [ + 212, + 296, + 787, + 659 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2 Toy Experiment", + "text_level": 1, + "bbox": [ + 214, + 680, + 397, + 696 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Based on previous research findings mentioned in Section 3.1, the types of distortions in LiDAR point clouds caused by adverse weather conditions converge into a set of common distortions. Therefore, from a data-centric perspective, the issues that we need to consider can be summarized as follows: (D1) Point Drop, (D2) Occlusions, (D3) Geometric Perturbation, and (D4) Intensity Distortion. Here, we aim to identify which distortion types negatively impact performance. To achieve this, we have generated four distortion types of toy synthetic data through the SemanticKITTI validation set.", + "bbox": [ + 212, + 719, + 787, + 839 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "Rethinking Data Augmentation for Robust LiDAR Semantic Segmentation", + "bbox": [ + 232, + 114, + 730, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/0dd8f63c447380b92bd5e82c3d7c09edc96954dee2eb3cf57e9e164b1ae8535b.jpg", + "table_caption": [ + "Table 1: Results of toy experiments from validation set of SemanticKITTI [2]. Soft and Hard indicate the severity of the distortions." + ], + "table_footnote": [], + "table_body": "
Distortioncarbileemtcletruckoth-v.pers.biclstmt-clstroadparki.sidew.oth-g.build.fenceveget.trunkterra.poletraf.mIoU
Clean96.822.266.989.165.267.084.70.093.850.681.40.191.163.088.167.774.563.847.763.9
D1: Soft96.317.963.487.163.764.381.40.092.747.579.70.190.162.387.364.274.161.141.661.8
D1: Hard86.10.212.721.040.010.124.40.04.210.211.10.071.747.877.034.627.630.417.627.7
D2: Soft94.816.351.766.353.959.252.10.092.544.979.20.189.761.487.363.274.359.240.257.2
D2: Hard81.30.55.413.135.57.60.70.02.68.510.80.071.145.776.332.927.326.716.124.3
D3: Soft96.215.456.758.651.351.878.10.066.333.946.10.086.161.684.462.056.861.544.053.2
D3: Hard93.89.738.319.431.435.255.20.09.912.312.20.048.940.871.455.531.258.740.835.0
D4: Soft96.321.761.189.461.868.983.20.092.944.479.10.190.456.786.968.271.163.948.362.3
D4: Hard95.017.250.481.457.064.979.70.090.738.864.70.988.445.383.562.850.560.147.956.8
", + "bbox": [ + 235, + 184, + 756, + 316 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- (D1) Point Drop: It considers the scenario where each LiDAR point disappears randomly and independently due to severe adverse weather. We removed individual points randomly to synthesize these data. We set the drop ratios at 0.5 and 0.9.", + "- (D2) Occlusions: The method assumes that occlusions occur predominantly in front of objects, mainly due to distortion from fog, snow, and rain. We synthesized this data by randomly selecting points and altering their depth to one-tenth of the original depth. The selecting ratios were also determined to be 0.5 and 0.9.", + "- (D3) Geometric Perturbation: This aspect assumes the distortion of point coordinates resulting from adverse weather conditions. We synthesized data by adding Gaussian noise to the coordinates of all points. The Gaussian noise levels were established at 0.05 and 0.25.", + "- (D4) Intensity Distortion: This distortion type assumes intensity attenuation caused by fog, raindrops, and snow particles. We synthesized this data by subtracting Gaussian noise from the intensity of all points. The Gaussian noise levels were established at 0.05 and 0.25." + ], + "bbox": [ + 225, + 344, + 785, + 599 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We selected MinkowskiNet [5] for our toy experiment. This choice was based on existing research [11, 34], which indicates that MinkowskiNet is a standard and robust model.", + "bbox": [ + 212, + 613, + 785, + 657 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In Tab. 1, it's evident that with increasing distortion in D1, D2, and D3, performance falls below half of the baseline. Contrarily, D4 maintains performance levels despite significant distortion. This performance degradation likely stems from changes in the local geometric structure where model computations occur. For this reason, D4 did not significantly impact performance.", + "bbox": [ + 212, + 657, + 785, + 734 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Importantly, D2's notable performance drop is argued to result from a point drop akin to D1. It is due to occlusions altering the point's local geometric structure, an effect similar to point drop. This is evidenced in Fig. 1, where incorrect predictions from D2 show patterns resembling those in D1.", + "bbox": [ + 212, + 734, + 785, + 794 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In summary, our observations indicate that geometric perturbation and point drop are the most impactful distortions from adverse weather. Data augmentation that replicates these phenomena could likely boost the model's robustness", + "bbox": [ + 212, + 795, + 785, + 839 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "J.Park et al.", + "bbox": [ + 271, + 114, + 354, + 127 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/82090c9f50fa12ab3789eb06cbd45733571c834c844de401e608b4c407606628.jpg", + "image_caption": [ + "Fig. 1: Visualization for the results of toy experiment. In all figures, green (or red) points indicate correct (or incorrect) predictions. Point drop, occlusion, and geometric perturbation are the main distortions that lead to performance degradation. Also, since the misprediction areas of point drop and occlusion largely overlap, it's reasonable to consider them as the same type of distortion." + ], + "image_footnote": [], + "bbox": [ + 217, + 150, + 787, + 256 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "against such conditions without explicit weather simulations. Thus, robust Li-DAR semantic segmentation models against adverse weather involve two key steps: (1) Training on specific adverse point drops detrimental to performance, and (2) Training with minor adjustments in point coordinates.", + "bbox": [ + 212, + 378, + 787, + 455 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4 Methods", + "text_level": 1, + "bbox": [ + 215, + 479, + 339, + 496 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this section, we propose solutions for the two challenges identified in Section 3.2. Our objective is to enhance the robustness of the model by addressing the common distortions exhibited by various adverse weather conditions, through data augmentation. This approach aims to achieve robustness without the necessity of simulating each adverse weather scenario explicitly.", + "bbox": [ + 212, + 513, + 787, + 590 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "As observed earlier, distortions resulting from adverse weather include point drop, occlusion, geometric perturbation, and intensity distortion. Of these, the main distortions identified as having the most detrimental impact on performance were (1) geometric perturbation caused by minor variations in point coordinates and (2) point drop caused by beam missing or occlusion. To address these challenges, we propose two techniques: Selective Jittering and Learnable Point Drop.", + "bbox": [ + 212, + 590, + 787, + 696 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.1 Selective Jittering", + "text_level": 1, + "bbox": [ + 215, + 720, + 413, + 736 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Selective Jittering is devised to address the first main distortion, geometric perturbation. Selective Jittering consists of two types: Depth-Selective Jittering (DSJ) and Angle-Selective Jittering (ASJ).", + "bbox": [ + 212, + 748, + 785, + 794 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Depth-Selective Jittering (DSJ). DSJ employs a single scan to add Gaussian noise to the XYZ coordinates and the intensity value of points below a randomly selected depth range.", + "bbox": [ + 212, + 794, + 785, + 840 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "Rethinking Data Augmentation for Robust LiDAR Semantic Segmentation", + "bbox": [ + 232, + 114, + 732, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Angle-Selective Jittering (ASJ). It applies Gaussian noise to a randomly selected range of angles. By performing jittering across different ranges of depths or angles for each point, this method reflects the augmentation that some beams are affected by geometric perturbation in a non-uniform manner in adverse weather. This concept realistically represents the characteristics of transparent droplets like rain and snow [6].", + "bbox": [ + 212, + 146, + 782, + 236 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "A significant aspect of DSJ and ASJ is that they use only a single frame. DSJ and ASJ are efficient in data augmentation without additional LiDAR frames and can perform reasonable augmentation even when the given LiDAR data is not sequentially captured.", + "bbox": [ + 212, + 237, + 782, + 297 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Range Jittering (RJ). Range Jittering has been proposed to simulate range distortion caused by droplets and fog, as mentioned in [9]. Unlike DSJ and ASJ, Range Jittering applies jittering only in the range direction of points. This method is used in place of using the original points in DSJ and ASJ.", + "bbox": [ + 212, + 297, + 782, + 357 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.2 Learnable Point Drop", + "text_level": 1, + "bbox": [ + 215, + 378, + 442, + 393 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Learnable Point Drop (LPD) was devised to address the second main distortion, point drop. LPD is designed to artificially create point drop due to occlusion, such as dense fog. LPD employs a Deep Q-Learning Network (DQN) [19] for identifying the drop ratio and drop region that lead to adverse effects on the model. The reward of the DQN is designed to identify point drops that increase the training loss and uncertainty of the LiDAR semantic segmentation model. Through LPD, the LiDAR semantic segmentation model is exposed to point drop scenarios caused by adverse weather conditions. Consequently, it learns to make accurate predictions even in the absence of points that are critical for segmentation in clean data environments. Since LPD is employed merely as a concept for data augmentation and exists as a separate module, it necessitates no alterations to the existing model's training scheme except for limiting the gradient norm to ensure the stability of DQN learning.", + "bbox": [ + 212, + 402, + 784, + 598 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "LPD module defines its current state by summing the loss $L_{aug}$ , calculated from augmented data obtained through SJ, and the entropy $H_{aug}$ , derived from logits. The loss $L_{aug}$ is calculated using the original loss function employed by the model in use. The calculation of entropy is as follows:", + "bbox": [ + 212, + 598, + 782, + 654 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\nH _ {a u g} (x) = - \\frac {1}{N} \\sum_ {i = 1} ^ {n} P (x _ {i}) \\log P (x _ {i})\n$$\n", + "text_format": "latex", + "bbox": [ + 385, + 662, + 616, + 696 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this process, $x_{i}$ represents each point, and $P(\\cdot)$ is softmax to the logits. $N$ denotes the number of points. LPD predicts the indices of points to drop, taking as input the sum of $L_{\\text{aug}}$ and $H_{\\text{aug}}$ added to the input point tensor. This process enables simultaneous learning of the drop ratio and drop region. Subsequently, the difference between the sum of the loss $L_{\\text{LPD}}$ and entropy $H_{\\text{LPD}}$ , obtained from the dropped points by LPD, and the sum of $L_{\\text{aug}}$ and $H_{\\text{aug}}$ , is defined as the reward.", + "bbox": [ + 212, + 704, + 784, + 808 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In Comparison with random drop, random drop uniformly reduces points across all depths. Consequently, this approach is not suitable for simulating", + "bbox": [ + 212, + 809, + 782, + 839 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "J.Park et al.", + "bbox": [ + 271, + 114, + 354, + 127 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/b70ceab1da88d04eff57b3b5ea4d42faf5f504c9b8a890620223ab24a564a38b.jpg", + "image_caption": [ + "Fig. 2: Description of Learnable Point Drop (LPD). (1) The LPD utilizes the loss $L_{\\text{aug}}$ and entropy $H_{\\text{aug}}$ derived from LiDAR segmentation model as its current state. (2) Upon receiving a point input, LPD performs a point drop, then recalculates the loss $L_{\\text{LPD}}$ and entropy $H_{\\text{LPD}}$ . (3) The difference between these new values and the previous ones is used to compute the reward." + ], + "image_footnote": [], + "bbox": [ + 228, + 148, + 787, + 262 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "point drops caused by fog, as it lacks the depth-specific characteristics of fog-related distortions. Furthermore, randomly generated point drop scans may not achieve the severity of adverse weather conditions necessary to drop critical points that could deceive the model. A detailed comparison will be conducted in the supplementary material.", + "bbox": [ + 212, + 402, + 787, + 479 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.3 Overall Pipeline", + "text_level": 1, + "bbox": [ + 215, + 502, + 395, + 517 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "The overall training procedure begins with data augmentation through SJ. The augmented points are input into the LiDAR semantic segmentation model to compute loss. The model then calculates logits for each point, applies softmax, and subsequently computes the entropy. The current state of the LPD module is set as the sum of the loss and entropy. The dropped points are then input into the LiDAR semantic segmentation model to calculate the corresponding loss and entropy. The difference between the sum of loss/entropy obtained from the augmented points and that obtained from the points determined by LPD is set as the reward. Through this process, the LPD module enables the LiDAR semantic segmentation model to effectively learn points distorted due to adverse weather. Overall process is detailed in Fig. 3", + "bbox": [ + 212, + 527, + 787, + 696 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "5 Experiments", + "text_level": 1, + "bbox": [ + 215, + 719, + 375, + 736 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "5.1 Experimental Setup", + "text_level": 1, + "bbox": [ + 215, + 753, + 426, + 768 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Dataset and Evaluation Metrics. To assess robustness, we trained our model on 19 classes from SemanticKITTI and then validated its performance on validation set of SemanticSTF as same as [33]. The experiment process involves training on clean data followed by evaluating robustness in adverse conditions,", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "Rethinking Data Augmentation for Robust LiDAR Semantic Segmentation", + "bbox": [ + 232, + 114, + 730, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/3d4e4751c2405ef68b580baaa1320e0f2611933d089a0a7374db71eeee391e64.jpg", + "image_caption": [ + "Fig. 3: The overall training process. (1) Initially, SJ is applied. (2) Following this, we calculate the loss, which leads to the update of the LPD state. (3) This process is followed by utilizing LPD to generate dropped points. (4) Subsequent to recalculating the loss, the reward is accordingly updated." + ], + "image_footnote": [], + "bbox": [ + 222, + 146, + 785, + 289 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "akin to existing research [11, 33, 34]. SemanticSTF is a dataset gathered under real-world adverse conditions such as rain, fog, and snow. Following the approach in [33], invalid data in SemanticSTF are mapped to an \"ignore label\". We utilized MinkowskiNet [5] for validation purposes due to its well-known robustness [11, 34]. Our evaluation metrics include the Intersection over Union (IoU) for each class and the mean IoU (mIoU) across all classes.", + "bbox": [ + 212, + 398, + 782, + 489 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Implementation Details. MinkowskiNet-18/32width served as our baseline model. The learning rate was set at 0.24, with a weight decay of 0.0001. The mean and standard deviation of Gaussian noise used in SJ are 0 and 0.01 each. In cases of Learnable Point Drop, the norm of the gradient was limited to 100 for both the DQN and Segmentation model. All experiments were conducted using four A6000 GPUs for 15 epochs in batch size 2. The duration of all experiments ranged between 3 to 5 hours.", + "bbox": [ + 212, + 489, + 784, + 595 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "5.2 Main Results", + "text_level": 1, + "bbox": [ + 214, + 618, + 372, + 632 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "SemanticKITTI to SemanticSTF. According to Tab. 2 (a), our proposed method showed a significant improvement in SemanticSTF, with a $+8.1$ mIoU increase over the baseline and $+5.6$ mIoU over the recent competitor, PointDR. Our method demonstrated universally superior performance across all weather conditions, validating the validity of our choice in data distortion type and the effectiveness of our method. Our method substantially improved the performance of categories like other vehicle, motorcyclist, sidewalk, pole, and traffic sign, which had lower baseline performance, increasing their class IoU by about $+5$ to $+10$ . Additionally, it significantly enhanced the mIoU for cars and persons by $+19.0$ and $+9.6$ , respectively, compared to the baseline, which is crucial for safety in driving environments. More details are available in the supplementary material.", + "bbox": [ + 212, + 643, + 787, + 809 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Also, in Tab. 2 (a), examining the performance improvement per weather condition, we note that the most significant increase in mIoU was in rain, with", + "bbox": [ + 212, + 809, + 785, + 839 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "J.Park et al.", + "bbox": [ + 271, + 114, + 354, + 126 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/5d8c1fb8e25705020adce53a5c9b6337044d6044ec87ea05db3cac6f134323b6.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
(a) SemanticKITTI→SemanticSTF
MethodsD-fogL-fogRainSnowmIoU
Oracle51.954.657.953.754.7
Baseline30.730.129.725.331.4
LaserMix [13]23.215.59.37.814.7
PolarMix [31]21.314.916.59.315.3
PointDR* [33]37.333.535.526.933.9
Baseline+SJ+LPD36.037.537.633.139.5
Increments to baseline+5.3+7.4+7.9+7.8+8.1
", + "bbox": [ + 225, + 252, + 491, + 375 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/ac50736120ce4d7230c3111b5aba569db3ec29471724547ee42da77e609cb316.jpg", + "table_caption": [ + "Table 2: LiDAR segmentation results (mIoU) on the SemanticSTF validation set of models trained with (a) SemanticKITTI and (b) SynLiDAR [32]. D-fog and L-fog denote dense fog and light fog weather conditions in all experiments. * denotes the reproduced result with the same segmentation backbone. The symbol $\\ddagger$ indicates that the validation for model selection was performed on sequence 0 of SynLiDAR, rather than SemanticSTF. Best score is in bold and second best is underlined." + ], + "table_footnote": [], + "table_body": "
(b) SynLiDAR→SemanticSTF
MethodsD-fogL-fogRainSnowmIoU
Oracle51.954.657.953.754.7
Baseline15.2415.9716.8312.7615.45
LaserMix [13]15.3217.9518.5513.816.85
PolarMix [31]16.4718.6919.6315.9818.09
PointDR* [33]19.0920.2825.2918.9819.78
PointDR*‡21.4120.9425.4819.3120.47
Baseline+SJ+LPD19.0820.6521.9717.2720.08
Increments to baseline+3.8+4.7+5.1+4.5+4.6
Baseline+SJ+LPD‡18.9921.2223.1417.2820.51
Increments to baseline+3.7+5.3+6.3+4.5+5.1
", + "bbox": [ + 501, + 239, + 792, + 388 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "+7.9 mIoU. This result is indicative of our SJ augmentation effectively reflecting geometric perturbations caused by raindrops. This aspect will be further explored in the ablation study. Additionally, in snow, our method showed a remarkable increase of +7.8 mIoU, and in dense fog (+5.3 mIoU) and light fog (+7.4 mIoU), our methodology clearly outperformed the baseline. This indicates that our data augmentation effectively models the adverse impacts of weather on LiDAR semantic segmentation models. Detailed analysis of how our method enhances robustness against each weather condition, and the factors contributing to the improved performance, will be discussed in the ablation study.", + "bbox": [ + 212, + 435, + 787, + 571 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Moreover, as previously mentioned, the training architectures designed to enhance robustness, such as those proposed in PointDR [11, 33, 34], can be orthogonally applied to our proposed data augmentation. Therefore, regardless of the choice of training architecture, the integration of our method is expected to yield superior performance compared to conventional approaches.", + "bbox": [ + 212, + 580, + 787, + 656 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "SynLiDAR to SemanticSTF. For SynLiDAR, unlike SemanticKITTI, where no specific validation set was available, sequence 0 was treated as the validation set. The best-performing models during these validation checkpoints were selected for reporting.", + "bbox": [ + 212, + 664, + 787, + 724 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "In contrast, PointDR demonstrated an unrealistic scenario in its baseline code, using SemanticSTF for validation during training and reporting only the highest-performing results. For a fair comparison using PointDR's criteria, our method shows a performance increase of $+0.04$ mIoU over PointDR. Using our more realistic protocol, our method demonstrates an even greater performance improvement of $+0.32$ mIoU against PointDR. Detailed results are shown in Tab. 2 (b).", + "bbox": [ + 212, + 734, + 787, + 840 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "Rethinking Data Augmentation for Robust LiDAR Semantic Segmentation", + "bbox": [ + 232, + 114, + 730, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/ad4edbac792c147da1e45e3c3a96f6cafa4ab7343688f7c51c50e565cd907cf6.jpg", + "image_caption": [ + "Fig. 4: Qualitative results of our method from validation set of SemanticSTF. All models are trained in train set of SemanticKITTI. Green (or red) points indicate correct (or incorrect) predictions, showing our predictions are more accurate than PointDR in all weather, especially on road and sidewalk." + ], + "image_footnote": [], + "bbox": [ + 215, + 143, + 787, + 361 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "5.3 Additional Experiments", + "text_level": 1, + "bbox": [ + 215, + 468, + 459, + 484 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Analysis of Proposed Component Methods. As shown in As Tab. 3, when ASJ was added to the baseline, we observed an improvement in robustness across all adverse weather conditions (increases of $+2.6\\mathrm{mIoU}$ in dense fog, $+5.3\\mathrm{mIoU}$ in light fog, $+8.1\\mathrm{mIoU}$ in rain, and $+6.3\\mathrm{mIoU}$ in snow), albeit with a $-1.8\\mathrm{mIoU}$ decrease in clean data performance. The introduction of DSJ further enhanced performance in dense fog ( $+4.1\\mathrm{mIoU}$ ), light fog ( $+6.3\\mathrm{mIoU}$ ), rain ( $+9.3\\mathrm{mIoU}$ ), snow ( $+4.6\\mathrm{mIoU}$ ), and an overall performance increase of $+6.2\\mathrm{mIoU}$ . Our Selective Jittering method's strong performance in rain conditions showcases its efficacy in rainy weather, where it considers geometric perturbations for selected beams instead of all.", + "bbox": [ + 212, + 492, + 787, + 642 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "The application of Range Jittering showed improvements in dense fog(+3.7 mIoU), light fog (+6.9 mIoU), rain(+6.0 mIoU), snow (+8.2 mIoU), and an overall performance increase of +7.3 mIoU, indicating the importance of simulating range distortion due to adverse weather as discussed in [9]. As seen earlier, in rainy weather, jittering only some beams aids in enhancing robustness. However, performing Range Jittering on the remaining original points leads to all points being jittered, reducing performance in rain conditions. Additionally, the improvement in snow conditions by using Range Jittering suggests that perturbing as many beams as possible is more effective for robustness enhancement in snowy weather.", + "bbox": [ + 212, + 643, + 787, + 794 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "The addition of LPD resulted in improvements in dense fog $(+5.3\\mathrm{mIoU})$ light fog $(+7.4\\mathrm{mIoU})$ , rain $(+7.9\\mathrm{mIoU})$ , snow $(+7.8\\mathrm{mIoU})$ , and an overall increase of $+8.1\\mathrm{mIoU}$ against to the baseline. The substantial performance en", + "bbox": [ + 212, + 794, + 787, + 839 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "J.Park et al.", + "bbox": [ + 271, + 114, + 354, + 126 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/ad369681cb9220639aaa36570dda2b1b878349fb8ad0df896fb7a7a13e706142.jpg", + "table_caption": [ + "Table 3: Experiments on the components of our methods with SemanticKITTI as source and SemanticSTF as target. The values in parentheses indicate the performance improvement or decrease over the baseline model." + ], + "table_footnote": [], + "table_body": "
MethodsCleanD-fogL-fogRainSnowmIoU
Baseline63.930.730.129.725.331.4
+ASJ62.1 (-1.8)33.335.437.831.636.8 (+5.4)
+DSJ63.0 (-0.9)34.836.439.029.937.6 (+6.2)
+RJ61.2 (-2.7)33.437.035.733.538.7 (+7.3)
+LPD62.8 (-1.1)36.037.537.633.139.5 (+8.1)
", + "bbox": [ + 316, + 198, + 676, + 319 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "hancements observed in conditions of dense and light fog indicate that the LPD effectively represents occlusion from foggy weather as intended. The performance reduction in snow conditions compared to the use of ASJ, DSJ, and Range Jittering $(-0.4\\mathrm{mIoU})$ is likely due to LPD not exclusively targeting the frozen or wet ground for point dropout. LPD applies the point drop across all points, as shown in Fig. 5. This represents a limitation of LPD, necessitating the development of alternative methods to address this issue.", + "bbox": [ + 212, + 352, + 787, + 457 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Overall, the use of all components led to a significant improvement in performance under adverse weather conditions, with an increase of $+8.1\\mathrm{mIoU}$ over the baseline at a reasonable cost of $-1.1\\mathrm{mIoU}$ in clean data. Ultimately, our method achieved state-of-the-art results on SemanticSTF, with an increase in mIoU across all adverse weather conditions compared to the baseline. The combined use of DSJ, ASJ, Range Jittering, and LPD demonstrated high performance, indicating that each component synergistically contributes to enhancing the model's robustness.", + "bbox": [ + 212, + 460, + 787, + 582 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/5e91bfdc18fcdfda504f2f19c6eda8a1fd7bf04a3f0d711c19bf0aa75b80437f.jpg", + "image_caption": [ + "Fig. 5: Qualitative results of LPD from validation set of SemanticKITTI. It can be observed that the point drop facilitated by the LPD occurs not just locally but extends across all points. While LPD effectively mimics the occlusions and sparse point distributions seen in real adverse weather, it does not replicate the effect of wet ground surfaces." + ], + "image_footnote": [], + "bbox": [ + 328, + 626, + 674, + 739 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "Rethinking Data Augmentation for Robust LiDAR Semantic Segmentation", + "bbox": [ + 232, + 114, + 730, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/014e3f53da98b9a32af10ca057a6467fa4f68293f0ba48dd355c82fc664a6b66.jpg", + "table_caption": [ + "Table 4: LiDAR segmentation results (mIoU) with the various architecture and datasets. All models were trained with SemanticKITTI." + ], + "table_footnote": [], + "table_body": "
MethodSemanticSTFSemanticKITTI-C
CENet14.249.3
SPVCNN28.152.5
Minkowski31.453.0
CENet+Ours22.0 (+7.8)53.2 (+3.9)
SPVCNN+Ours38.4 (+10.3)52.9 (+0.4)
Minkowski+Ours39.5 (+8.1)58.6 (+5.6)
", + "bbox": [ + 326, + 184, + 671, + 305 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Experiments on Various Input Data Representation, Models and Datasets. We evaluated our method using different data representations, architectures, and datasets, specifically: (1) projection-based CENet and (2) voxel-based SPVCNN for backbones, and (3) the synthetic corruption dataset SemanticKITTI-C [11]. Tab. 4 shows our results.", + "bbox": [ + 212, + 335, + 815, + 409 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "SemanticKITTI-C has a unique data distribution with more samples in classes, such as sidewalks and vegetation, than SemanticSTF. Our method consistently improves performance across various datasets and architectures. Specifically, for the SemanticSTF dataset, our method increases the mIoU of CENet by $+7.8$ , SPVCNN by $+10.3$ , and MinkowskiNet by $+8.1$ mIoU. Similarly, for the SemanticKITTI-C dataset, the improvements are $+3.9$ for CENet, $+0.4$ for SPVCNN, and $+5.6$ mIoU for MinkowskiNet. These consistent enhancements across different input data representations, models, and datasets highlight the robustness and generalizability of our approach.", + "bbox": [ + 212, + 412, + 789, + 547 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "6 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 575, + 357, + 590 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Our study focused on analyzing LiDAR data distortions in adverse weather and enhancing LiDAR model robustness with a data-centric method. We identified that point drop and geometric distortion mainly affect model performance through toy experiments. We introduced novel data augmentation techniques, Selective Jittering, and Learnable Point Drop, leading to state-of-the-art performance on SemanticSTF. We showed that our methodology significantly contributes to robustness enhancement without complex data simulations. We expect that our research will contribute to improving the stability and reliability of LiDAR semantic segmentation.", + "bbox": [ + 212, + 612, + 787, + 747 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgements. This work was supported by IITP grant funded by MSIT (No.2021-0-02068, Artificial Intelligence Innovation Hub and RS-2019-II190075, Artificial Intelligence Graduate School Program(KAIST)) and KEIT grant funded by MOTIE (No.2022-0-00680, No.2022-0-01045), NRF funded by the MSIP (NRF-2022R1A2C3011154, RS-2023-00219019, RS-2023-00240135) and Samsung Electronics Co., Ltd (IO230508-06190-01).", + "bbox": [ + 212, + 750, + 792, + 840 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "J.Park et al.", + "bbox": [ + 271, + 114, + 354, + 126 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 215, + 143, + 321, + 159 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Ando, A., Gidaris, S., Bursuc, A., Puy, G., Boulch, A., Marlet, R.: Rangevit: Towards vision transformers for 3d semantic segmentation in autonomous driving. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 5240-5250 (2023)", + "2. Behley, J., Garbade, M., Milioto, A., Quenzel, J., Behnke, S., Stachniss, C., Gall, J.: Semantickitti: A dataset for semantic scene understanding of lidar sequences. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 9297-9307 (2019)", + "3. Bijelic, M., Gruber, T., Ritter, W.: A benchmark for lidar sensors in fog: Is detection breaking down? In: 2018 IEEE intelligent vehicles symposium (IV). pp. 760-767. IEEE (2018)", + "4. Choe, J., Park, C., Rameau, F., Park, J., Kweon, I.S.: Pointmixer: Mlp-mixer for point cloud understanding. In: European Conference on Computer Vision. pp. 620-640. Springer (2022)", + "5. Choy, C., Gwak, J., Savarese, S.: 4d spatio-temporal convnets: Minkowski convolutional neural networks. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 3075-3084 (2019)", + "6. Fersch, T., Buhmann, A., Koelpin, A., Weigel, R.: The influence of rain on small aperture lidar sensors. In: 2016 German Microwave Conference (GeMiC). pp. 84-87. IEEE (2016)", + "7. Hahner, M., Sakaridis, C., Bijelic, M., Heide, F., Yu, F., Dai, D., Van Gool, L.: Lidar snowfall simulation for robust 3d object detection. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 16364-16374 (2022)", + "8. Hahner, M., Sakaridis, C., Dai, D., Van Gool, L.: Fog simulation on real lidar point clouds for 3d object detection in adverse weather. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 15283-15292 (2021)", + "9. Kilic, V., Hegde, D., Sindagi, V., Cooper, A.B., Foster, M.A., Patel, V.M.: Lidar light scattering augmentation (lisa): Physics-based simulation of adverse weather conditions for 3d object detection. arXiv preprint arXiv:2107.07004 (2021)", + "0. Kong, L., Liu, Y., Chen, R., Ma, Y., Zhu, X., Li, Y., Hou, Y., Qiao, Y., Liu, Z.: Rethinking range view representation for lidar segmentation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 228-240 (2023)", + "1. Kong, L., Liu, Y., Li, X., Chen, R., Zhang, W., Ren, J., Pan, L., Chen, K., Liu, Z.: Robo3d: Towards robust and reliable 3d perception against corruptions. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 19994-20006 (2023)", + "2. Kong, L., Ren, J., Pan, L., Liu, Z.: Lasermix for semi-supervised lidar semantic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 21705-21715 (2023)", + "3. Kong, L., Ren, J., Pan, L., Liu, Z.: Lasermix for semi-supervised lidar semantic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 21705-21715 (2023)", + "4. Lai, X., Chen, Y., Lu, F., Liu, J., Jia, J.: Spherical transformer for lidar-based 3d recognition. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 17545-17555 (2023)", + "5. Lee, S., Son, T., Kwak, S.: Fico: Learning fog-invariant features for foggy scene segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18911-18921 (2022)" + ], + "bbox": [ + 225, + 172, + 785, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "Rethinking Data Augmentation for Robust LiDAR Semantic Segmentation", + "bbox": [ + 232, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "16. Li, Y., Duthon, P., Colomb, M., Ibanez-Guzman, J.: What happens for a tof lidar in fog? IEEE Transactions on Intelligent Transportation Systems 22(11), 6670-6681 (2020)", + "17. Li, Z., Wu, X., Wang, J., Guo, Y.: Weather-degraded image semantic segmentation with multi-task knowledge distillation. Image and Vision Computing 127, 104554 (2022)", + "18. Milioto, A., Vizzo, I., Behley, J., Stachniss, C.: Rangenet++: Fast and accurate lidar semantic segmentation. In: 2019 IEEE/RSJ international conference on intelligent robots and systems (IROS). pp. 4213-4220. IEEE (2019)", + "19. Mnih, V., Kavukcuoglu, K., Silver, D., Rusu, A.A., Veness, J., Bellemare, M.G., Graves, A., Riedmiller, M., Fidjeland, A.K., Ostrovski, G., et al.: Human-level control through deep reinforcement learning. nature 518(7540), 529-533 (2015)", + "20. Nekrasov, A., Schult, J., Litany, O., Leibe, B., Engelmann, F.: Mix3d: Out-of-context data augmentation for 3d scenes. In: 2021 international conference on 3d vision (3dv). pp. 116-125. IEEE (2021)", + "21. Puy, G., Boulch, A., Marlet, R.: Using a waffle iron for automotive point cloud semantic segmentation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 3379-3389 (2023)", + "22. Qi, C.R., Su, H., Mo, K., Guibas, L.J.: Pointnet: Deep learning on point sets for 3d classification and segmentation. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 652-660 (2017)", + "23. Qian, K., Zhu, S., Zhang, X., Li, L.E.: Robust multimodal vehicle detection in foggy weather using complementary lidar and radar signals. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 444-453 (2021)", + "24. Ryu, K., Hwang, S., Park, J.: Instant domain augmentation for lidar semantic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9350-9360 (2023)", + "25. Saltori, C., Galasso, F., Fiameni, G., Sebe, N., Ricci, E., Poiesi, F.: Cosmix: Compositional semantic mix for domain adaptation in 3d lidar segmentation. In: European Conference on Computer Vision. pp. 586-602. Springer (2022)", + "26. Shin, J., Park, H., Kim, T.: Characteristics of laser backscattering intensity to detect frozen and wet surfaces on roads. Journal of Sensors 2019(1), 8973248 (2019)", + "27. Smith, B.E., Gardner, A., Schneider, A., Flanner, M.: Modeling biases in laser-altimetry measurements caused by scattering of green light in snow. Remote sensing of environment 215, 398-410 (2018)", + "28. Thomas, H., Qi, C.R., Deschaud, J.E., Marcotegui, B., Goulette, F., Guibas, L.J.: Kpconv: Flexible and deformable convolution for point clouds. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 6411-6420 (2019)", + "29. Tolstikhin, I.O., Houlsby, N., Kolesnikov, A., Beyer, L., Zhai, X., Unterthiner, T., Yung, J., Steiner, A., Keysers, D., Uszkoreit, J., et al.: Mlp-mixer: An allmlp architecture for vision. Advances in neural information processing systems 34, 24261-24272 (2021)", + "30. Xiao, A., Huang, J., Guan, D., Cui, K., Lu, S., Shao, L.: Polarmix: A general data augmentation technique for lidar point clouds. Advances in Neural Information Processing Systems 35, 11035-11048 (2022)", + "31. Xiao, A., Huang, J., Guan, D., Cui, K., Lu, S., Shao, L.: Polarmix: A general data augmentation technique for lidar point clouds. Advances in Neural Information Processing Systems 35, 11035-11048 (2022)" + ], + "bbox": [ + 215, + 146, + 784, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "J.Park et al.", + "bbox": [ + 271, + 114, + 354, + 126 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "32. Xiao, A., Huang, J., Guan, D., Zhan, F., Lu, S.: Transfer learning from synthetic to real lidar point cloud for semantic segmentation. In: Proceedings of the AAAI conference on artificial intelligence. vol. 36, pp. 2795-2803 (2022)", + "33. Xiao, A., Huang, J., Xuan, W., Ren, R., Liu, K., Guan, D., El Saddik, A., Lu, S., Xing, E.P.: 3d semantic segmentation in the wild: Learning generalized models for adverse-condition point clouds. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9382-9392 (2023)", + "34. Yan, X., Zheng, C., Xue, Y., Li, Z., Cui, S., Dai, D.: Benchmarking the robustness of lidar semantic segmentation models. International Journal of Computer Vision pp. 1-24 (2024)", + "35. Yang, D., Liu, Z., Jiang, W., Yan, G., Gao, X., Shi, B., Liu, S., Cai, X.: Realistic rainy weather simulation for lidars in carla simulator. arXiv preprint arXiv:2312.12772 (2023)", + "36. Zhao, H., Jiang, L., Jia, J., Torr, P.H., Koltun, V.: Point transformer. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 16259-16268 (2021)", + "37. Zhu, X., Zhou, H., Wang, T., Hong, F., Ma, Y., Li, W., Li, H., Lin, D.: Cylindrical and asymmetrical 3d convolution networks for lidar segmentation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 9939-9948 (2021)" + ], + "bbox": [ + 215, + 146, + 787, + 424 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "Rethinking Data Augmentation for Robust LiDAR Semantic Segmentation", + "bbox": [ + 232, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 16 + } +] \ No newline at end of file diff --git a/2024/Rethinking Data Augmentation for Robust LiDAR Semantic Segmentation in Adverse Weather/60de04d8-9845-4ba5-9238-a2dd5b5c4b07_model.json b/2024/Rethinking Data Augmentation for Robust LiDAR Semantic Segmentation in Adverse Weather/60de04d8-9845-4ba5-9238-a2dd5b5c4b07_model.json new file mode 100644 index 0000000000000000000000000000000000000000..842086622f97093b8db2265c8d92a9d232c75037 --- /dev/null +++ b/2024/Rethinking Data Augmentation for Robust LiDAR Semantic Segmentation in Adverse Weather/60de04d8-9845-4ba5-9238-a2dd5b5c4b07_model.json @@ -0,0 +1,2060 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.259, + 0.141, + 0.745, + 0.207 + ], + "angle": 0, + "content": "Rethinking Data Augmentation for Robust LiDAR Semantic Segmentation in Adverse Weather" + }, + { + "type": "text", + "bbox": [ + 0.297, + 0.235, + 0.703, + 0.252 + ], + "angle": 0, + "content": "Junsung Park\\(^{1}\\), Kyungmin Kim\\(^{2}\\), and Hyunjung Shim\\(^{1\\star}\\)" + }, + { + "type": "text", + "bbox": [ + 0.23, + 0.262, + 0.772, + 0.306 + ], + "angle": 0, + "content": "\\(^{1}\\) Korea Advanced Institute of Science & Technology (KAIST) \n\\(^{2}\\) Yonsei University \njshackist@kaist.ac.kr, kyungmin.kim@yonsei.ac.kr, kateshim@kaist.ac.kr" + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.334, + 0.74, + 0.681 + ], + "angle": 0, + "content": "Abstract. Existing LiDAR semantic segmentation methods often struggle with performance declines in adverse weather conditions. Previous work has addressed this issue by simulating adverse weather or employing universal data augmentation during training. However, these methods lack a detailed analysis and understanding of how adverse weather negatively affects LiDAR semantic segmentation performance. Motivated by this issue, we identified key factors of adverse weather and conducted a toy experiment to pinpoint the main causes of performance degradation: (1) Geometric perturbation due to refraction caused by fog or droplets in the air and (2) Point drop due to energy absorption and occlusions. Based on these findings, we propose new strategic data augmentation techniques. First, we introduced a Selective Jittering (SJ) that jitters points in the random range of depth (or angle) to mimic geometric perturbation. Additionally, we developed a Learnable Point Drop (LPD) to learn vulnerable erase patterns with a Deep Q-Learning Network to approximate the point drop phenomenon from adverse weather conditions. Without precise weather simulation, these techniques strengthen the LiDAR semantic segmentation model by exposing it to vulnerable conditions identified by our data-centric analysis. Experimental results confirmed the suitability of the proposed data augmentation methods for enhancing robustness against adverse weather conditions. Our method achieves a notable \\(39.5\\mathrm{mIoU}\\) on the SemanticKITTI-to-SemanticSTF benchmark, improving the baseline by \\(8.1\\%\\) p and establishing a new state-of-the-art. Our code will be released at https://github.com/engineerJPark/LiDARWeather." + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.693, + 0.741, + 0.722 + ], + "angle": 0, + "content": "Keywords: LiDAR Semantic Segmentation \\(\\cdot\\) Robust Learning for Adverse Weather \\(\\cdot\\) Data Augmentation" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.744, + 0.377, + 0.759 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.772, + 0.787, + 0.819 + ], + "angle": 0, + "content": "LiDAR semantic segmentation is a fundamental task of 3D scene understanding, particularly in safety-critical applications like autonomous driving. However, existing LiDAR segmentation models [1,5,14,21,28,37] commonly lack robustness," + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.825, + 0.496, + 0.84 + ], + "angle": 0, + "content": "* Corresponding author: Hyunjung Shim" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.356, + 0.128 + ], + "angle": 0, + "content": "J.Park et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.178 + ], + "angle": 0, + "content": "showing significant performance degradation under adverse weather conditions such as snow, fog, rain, or wet surfaces." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.181, + 0.787, + 0.317 + ], + "angle": 0, + "content": "To address this issue, recent studies [11, 33, 34] have introduced corruption benchmarks for adverse weather and proposed effective techniques for robust LiDAR segmentation against corruption. These robust methods are primarily divided into task-agnostic and simulation-based approaches. The task-agnostic approach [11, 25, 33] employs general machine-learning techniques for robustness without explicitly addressing LiDAR corruption caused by adverse weather. The simulation-based approach [3,7-9,23,26,35] artificially synthesizes weather-specific data using physical equations for training. However, these efforts focus on detection tasks and consider only a single type of weather at a time." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.32, + 0.788, + 0.546 + ], + "angle": 0, + "content": "While the simulation-based approach leverages the intrinsic properties of Li-DAR scans under target adverse weather conditions, simulating every weather type at all possible severities is impractical and often inaccurate. Instead, we adopt a data-centric perspective to analyze corrupted LiDAR data. For example, distortions caused by rain, snow, and fog often create similar patterns, as simulated in [9]. That is, all those adverse weather shows point drop patterns due to attenuation or occlusion by droplet [6,7,9]. While the simulation-based approach requires explicit modeling of the LiDAR for each weather condition, our data-centric approach potentially addresses complex distortions with a few unified patterns. Through existing research and our comprehensive analysis, we find that various adverse weather conditions create similar distortion patterns in Li-DAR data. Based on our analysis, we categorize these distortions into two types: (1) geometric perturbation and (2) point drop. Our toy experiments demonstrate that these distortions are closely related to segmentation performance degradation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.55, + 0.788, + 0.777 + ], + "angle": 0, + "content": "Rooted by the toy experiment, we introduce two novel and strategic data augmentation methods tailored to the LiDAR distortion caused by adverse weather. By incorporating these augmentations during training, we aim to bolster the model's robustness for each distortion type. For geometric perturbation, we introduce Selective Jittering (SJ). This method applies Additive Gaussian Noise (AGN) to alter the XYZ-coordinate and intensity within a selective local region. To handle a point drop pattern, we developed a Learnable Point Drop (LPD) that employs a Deep Q-Network (DQN) to strategically remove points. Our augmentation strategies are inspired by existing studies [9, 16, 27], which shows geometric perturbation typically involves small, random alterations to the original points. This insight led to our choice of jittering as a suitable technique for addressing the geometric perturbation. Also, our augmentation strategy for point drop is motivated by existing studies [7-9] which state that intensity and depth of LiDAR beams play a significant role. This prompts us to use a DQN to detect point drops and adapt to such point drop patterns." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Finally, we apply the proposed augmentations when training the LiDAR segmentation model with a clean source dataset. Experimental results demonstrate that our method achieves new state-of-the-art performance on SemanticSTF, a real adverse weather dataset. Specifically, we achieved an impressive \\(+8.1\\%\\)" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.233, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Rethinking Data Augmentation for Robust LiDAR Semantic Segmentation" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.236 + ], + "angle": 0, + "content": "mIoU gain over the baseline MinkowskiNet, nearly tripling the \\(+2.5\\%\\) p mIoU improvement from previous work [33]. Our method achieved remarkable performance gains, ranging from \\(+0.4\\) to \\(+10.3\\%\\) p mIoU, across various model architectures (Minkowski, CENet, SPVCNN) and corruption dataset (SemanticSTF and SemanticKITTI-C). These results demonstrate the generalizability of the proposed augmentation methods." + }, + { + "type": "text", + "bbox": [ + 0.239, + 0.238, + 0.661, + 0.253 + ], + "angle": 0, + "content": "In summary, this paper presents several key contributions." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.262, + 0.784, + 0.306 + ], + "angle": 0, + "content": "- We identify two prevalent distortion types in LiDAR data caused by adverse weather, leading to performance degradation, through a data-centric analysis." + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.308, + 0.784, + 0.336 + ], + "angle": 0, + "content": "- We introduce two novel data augmentations tailored to each identified distortion type." + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.338, + 0.784, + 0.412 + ], + "angle": 0, + "content": "- Our method sets new state-of-the-art benchmarks on the SemanticKITTI to SemanticSTF benchmark. Notably, our improvement was \\(+8.1\\%\\) p mIoU without relying on precise simulations of adverse weather in LiDAR point inputs. This represents a tripling of the improvement over the baseline, compared to what previous methods have achieved [33]." + }, + { + "type": "list", + "bbox": [ + 0.225, + 0.262, + 0.784, + 0.412 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.436, + 0.396, + 0.451 + ], + "angle": 0, + "content": "2 Related Works" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.468, + 0.519, + 0.483 + ], + "angle": 0, + "content": "2.1 LiDAR Semantic Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.493, + 0.784, + 0.536 + ], + "angle": 0, + "content": "Existing 3D LiDAR point cloud semantic segmentation methods can be categorized into three types based on the data representation: Point-based, Projection-based, and Voxel-based." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.538, + 0.785, + 0.657 + ], + "angle": 0, + "content": "Point-based methods [22, 28, 36] utilize the 3D points directly as input. KPConv [28] initially clustering local points, aggregating these local features, and then feeding them into kernel point convolutions. The Point Transformer [36] utilizes a transformer architecture to compute query points in each local region, obtained through k-nearest neighbors (kNN). The Point-Mixer [4] tried to adapt the MLP-Mixer [29] for point cloud applications. They achieve high performance but suffer from high computational costs due to the utilization of large-scale raw LiDAR data." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.659, + 0.785, + 0.794 + ], + "angle": 0, + "content": "Projection-based methods [1, 10, 18] project LiDAR points into a 2D image and performs the semantic segmentation using architectures successful in a 2D image. RangeViT [1] directly adopts a ViT model pre-trained on 2D images, demonstrating that pre-trained power in 2D images can be effective as prior knowledge in range images. RangeFormer [10] proposes \"RangeAug\" to maximize the utility of range images created by projecting into 2D, producing multiple range image data to overcome the low performance in range image models. Projection-based methods achieve fast inference speed but present a sub-optimal performance due to missing information after the projection." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.796, + 0.784, + 0.84 + ], + "angle": 0, + "content": "Voxel-based methods [5,14,37] perform efficient computations by dividing 3D space into a voxel grid and aggregating points within the same voxel. Minkowski-iNet [5] voxelizes LiDAR points with a cubic grid and applies sparse convolution." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.356, + 0.128 + ], + "angle": 0, + "content": "J.Park et al." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.146, + 0.787, + 0.222 + ], + "angle": 0, + "content": "Cylinder3D [37] proposes cylindrical partitions, reflecting LiDAR's characteristic that the density of points depends on the distance. SphereFormer [14] utilizes a radial window and transformer architecture to aggregate long-range information and improve performance. Voxel-based methods achieve a balance between reasonable inference time and commendable segmentation performance." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.244, + 0.492, + 0.26 + ], + "angle": 0, + "content": "2.2 LiDAR Data Augmentation" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.269, + 0.787, + 0.466 + ], + "angle": 0, + "content": "Inspired by 2D image augmentation, conventional LiDAR segmentation methods apply classic scaling, rotation, flipping, and translation to augment LiDAR data. Recently, several out-of-context augmentation techniques [12, 20, 30] that mix different LiDAR scans have been proposed. Mix3D [20] combines randomly selected two scans. Considering the sweeping mechanism of the LiDAR sensor, PolarMix [30] cuts LiDAR scans along the azimuth axis, then exchanges point cloud sectors and applies instance-level rotate-pasting. To reflect the spatial prior of the LiDAR point cloud, LaserMix [12] partitions LiDAR scans based on the laser beams and blends partitions from different LiDAR scans. Recent study [24] introduces a fast LiDAR domain augmentation module to address sensor-bias problems. To the best of our knowledge, our approach is the first augmentation method specifically designed to address data corruption under adverse weather conditions." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.487, + 0.618, + 0.503 + ], + "angle": 0, + "content": "2.3 LiDAR under Adverse Weather Conditions" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.513, + 0.801, + 0.725 + ], + "angle": 0, + "content": "Robustness under harsh conditions is crucial in safety-critical applications. Adverse weather substantially degrade performance in real-world outdoor autonomous driving. Thus, there are several attempts to develop weather-robust models in fields such as 2D segmentation [15, 17], 3D detection [7, 8, 11], and 3D segmentation [11, 33]. Simulation-based approaches [7, 8] artificially synthesize data for single weather conditions through physical modeling and utilize it for training. We differ from these in that we do not model specific weather conditions explicitly. Recently proposed task-agnostic approaches [11, 25, 33] consider multiple weather conditions at once. However, they use general machine-learning techniques (such as teacher-student framework and feature prototype) to achieve robustness rather than specifically tackle LiDAR corruption caused by adverse weather. We differ from these in that we propose augmentation methods specifically tailored for adverse weather conditions based on the analysis of performance degradation in LiDAR data." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.747, + 0.564, + 0.766 + ], + "angle": 0, + "content": "3 Finding Distortions to Augment" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.779, + 0.787, + 0.841 + ], + "angle": 0, + "content": "In this section, we aim to discuss the patterns of distortion that different adverse weather conditions impose on LiDAR data. Although adverse weather conditions are distinct in reality, studies have shown that their effects on LiDAR data often result in similar impacts. For instance, the distortions caused by \"rain\", \"snow\"," + } + ], + [ + { + "type": "header", + "bbox": [ + 0.233, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Rethinking Data Augmentation for Robust LiDAR Semantic Segmentation" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.253 + ], + "angle": 0, + "content": "and \"fog\" tend to produce similar point-missing patterns due to attenuation in the data, as demonstrated in [9]. Therefore, this section will focus on identifying the common distortion patterns caused by adverse weather through existing studies. Overall, existing studies describe the effect of adverse weather as four different types of distortion: (1) Point Drop due to energy absorption, (2) Occlusions caused by droplets of rain or snow and fog, (3) Geometric perturbation, and (4) Intensity distortion due to energy absorption." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.274, + 0.606, + 0.289 + ], + "angle": 0, + "content": "3.1 Distortion Factors from Adverse Weather" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.297, + 0.788, + 0.386 + ], + "angle": 0, + "content": "(D1) Point Drop. Several studies have explored how adverse weather conditions contribute to point drops in LiDAR data. Kilic et al. [9], Fersch et al. [6] and Shin et al. [26] describe point drops resulting from beam attenuation and beam missing due to droplets, fog, and frozen or wet ground. These studies collectively suggest that adverse weather conditions typically lead to point drops in LiDAR data." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.388, + 0.788, + 0.477 + ], + "angle": 0, + "content": "(D2) Occlusions. Several studies have addressed occlusions caused by adverse weather conditions. Hahner et al. [7], Kilic et al. [9], Kong et al. [11] and Yan et al. [34] consider scenarios where beams colliding with snow collect signals at much shorter distances than the intended objects of collision. Upon reviewing these studies, we have concluded that adverse weather consistently leads to occlusions." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.479, + 0.788, + 0.569 + ], + "angle": 0, + "content": "(D3) Geometric Perturbation. Some studies focus on geometric perturbation caused by adverse weather conditions. Kilic et al. [9], Li et al. [16] and Smith et al. [27] demonstrated geometric perturbation in adverse weather, such as fog, snow, and rain, by incorporating random noise into the coordinates. Through these studies, we have come to conclude that adverse weather universally causes geometric perturbation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.569, + 0.788, + 0.66 + ], + "angle": 0, + "content": "(D4) Intensity Distortion. Numerous studies have focused on intensity distortion caused by adverse weather conditions. Bijelic et al. [3], Shin et al. [26], Fersch et al. [6], Kong et al. [11], and Yan et al. [34] have collectively shown that adverse weather conditions like fog, wetness, and rain lead to a reduction in Li-DAR beam intensity, influencing the generation of synthetic data. Through these studies, we conclude that adverse weather commonly causes intensity distortion." + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.297, + 0.788, + 0.66 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.681, + 0.398, + 0.697 + ], + "angle": 0, + "content": "3.2 Toy Experiment" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Based on previous research findings mentioned in Section 3.1, the types of distortions in LiDAR point clouds caused by adverse weather conditions converge into a set of common distortions. Therefore, from a data-centric perspective, the issues that we need to consider can be summarized as follows: (D1) Point Drop, (D2) Occlusions, (D3) Geometric Perturbation, and (D4) Intensity Distortion. Here, we aim to identify which distortion types negatively impact performance. To achieve this, we have generated four distortion types of toy synthetic data through the SemanticKITTI validation set." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.356, + 0.128 + ], + "angle": 0, + "content": "J.Park et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.787, + 0.173 + ], + "angle": 0, + "content": "Table 1: Results of toy experiments from validation set of SemanticKITTI [2]. Soft and Hard indicate the severity of the distortions." + }, + { + "type": "table", + "bbox": [ + 0.236, + 0.185, + 0.758, + 0.317 + ], + "angle": 0, + "content": "
Distortioncarbileemtcletruckoth-v.pers.biclstmt-clstroadparki.sidew.oth-g.build.fenceveget.trunkterra.poletraf.mIoU
Clean96.822.266.989.165.267.084.70.093.850.681.40.191.163.088.167.774.563.847.763.9
D1: Soft96.317.963.487.163.764.381.40.092.747.579.70.190.162.387.364.274.161.141.661.8
D1: Hard86.10.212.721.040.010.124.40.04.210.211.10.071.747.877.034.627.630.417.627.7
D2: Soft94.816.351.766.353.959.252.10.092.544.979.20.189.761.487.363.274.359.240.257.2
D2: Hard81.30.55.413.135.57.60.70.02.68.510.80.071.145.776.332.927.326.716.124.3
D3: Soft96.215.456.758.651.351.878.10.066.333.946.10.086.161.684.462.056.861.544.053.2
D3: Hard93.89.738.319.431.435.255.20.09.912.312.20.048.940.871.455.531.258.740.835.0
D4: Soft96.321.761.189.461.868.983.20.092.944.479.10.190.456.786.968.271.163.948.362.3
D4: Hard95.017.250.481.457.064.979.70.090.738.864.70.988.445.383.562.850.560.147.956.8
" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.345, + 0.787, + 0.405 + ], + "angle": 0, + "content": "- (D1) Point Drop: It considers the scenario where each LiDAR point disappears randomly and independently due to severe adverse weather. We removed individual points randomly to synthesize these data. We set the drop ratios at 0.5 and 0.9." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.406, + 0.787, + 0.48 + ], + "angle": 0, + "content": "- (D2) Occlusions: The method assumes that occlusions occur predominantly in front of objects, mainly due to distortion from fog, snow, and rain. We synthesized this data by randomly selecting points and altering their depth to one-tenth of the original depth. The selecting ratios were also determined to be 0.5 and 0.9." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.481, + 0.787, + 0.541 + ], + "angle": 0, + "content": "- (D3) Geometric Perturbation: This aspect assumes the distortion of point coordinates resulting from adverse weather conditions. We synthesized data by adding Gaussian noise to the coordinates of all points. The Gaussian noise levels were established at 0.05 and 0.25." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.542, + 0.787, + 0.601 + ], + "angle": 0, + "content": "- (D4) Intensity Distortion: This distortion type assumes intensity attenuation caused by fog, raindrops, and snow particles. We synthesized this data by subtracting Gaussian noise from the intensity of all points. The Gaussian noise levels were established at 0.05 and 0.25." + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.345, + 0.787, + 0.601 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.614, + 0.786, + 0.658 + ], + "angle": 0, + "content": "We selected MinkowskiNet [5] for our toy experiment. This choice was based on existing research [11, 34], which indicates that MinkowskiNet is a standard and robust model." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.659, + 0.787, + 0.735 + ], + "angle": 0, + "content": "In Tab. 1, it's evident that with increasing distortion in D1, D2, and D3, performance falls below half of the baseline. Contrarily, D4 maintains performance levels despite significant distortion. This performance degradation likely stems from changes in the local geometric structure where model computations occur. For this reason, D4 did not significantly impact performance." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.787, + 0.795 + ], + "angle": 0, + "content": "Importantly, D2's notable performance drop is argued to result from a point drop akin to D1. It is due to occlusions altering the point's local geometric structure, an effect similar to point drop. This is evidenced in Fig. 1, where incorrect predictions from D2 show patterns resembling those in D1." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.796, + 0.787, + 0.84 + ], + "angle": 0, + "content": "In summary, our observations indicate that geometric perturbation and point drop are the most impactful distortions from adverse weather. Data augmentation that replicates these phenomena could likely boost the model's robustness" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.233, + 0.115, + 0.733, + 0.13 + ], + "angle": 0, + "content": "Rethinking Data Augmentation for Robust LiDAR Semantic Segmentation" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.151, + 0.788, + 0.257 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.279, + 0.788, + 0.351 + ], + "angle": 0, + "content": "Fig. 1: Visualization for the results of toy experiment. In all figures, green (or red) points indicate correct (or incorrect) predictions. Point drop, occlusion, and geometric perturbation are the main distortions that lead to performance degradation. Also, since the misprediction areas of point drop and occlusion largely overlap, it's reasonable to consider them as the same type of distortion." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.38, + 0.788, + 0.457 + ], + "angle": 0, + "content": "against such conditions without explicit weather simulations. Thus, robust Li-DAR semantic segmentation models against adverse weather involve two key steps: (1) Training on specific adverse point drops detrimental to performance, and (2) Training with minor adjustments in point coordinates." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.481, + 0.341, + 0.497 + ], + "angle": 0, + "content": "4 Methods" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.515, + 0.788, + 0.591 + ], + "angle": 0, + "content": "In this section, we propose solutions for the two challenges identified in Section 3.2. Our objective is to enhance the robustness of the model by addressing the common distortions exhibited by various adverse weather conditions, through data augmentation. This approach aims to achieve robustness without the necessity of simulating each adverse weather scenario explicitly." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.592, + 0.788, + 0.698 + ], + "angle": 0, + "content": "As observed earlier, distortions resulting from adverse weather include point drop, occlusion, geometric perturbation, and intensity distortion. Of these, the main distortions identified as having the most detrimental impact on performance were (1) geometric perturbation caused by minor variations in point coordinates and (2) point drop caused by beam missing or occlusion. To address these challenges, we propose two techniques: Selective Jittering and Learnable Point Drop." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.722, + 0.414, + 0.737 + ], + "angle": 0, + "content": "4.1 Selective Jittering" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.749, + 0.787, + 0.795 + ], + "angle": 0, + "content": "Selective Jittering is devised to address the first main distortion, geometric perturbation. Selective Jittering consists of two types: Depth-Selective Jittering (DSJ) and Angle-Selective Jittering (ASJ)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.787, + 0.841 + ], + "angle": 0, + "content": "Depth-Selective Jittering (DSJ). DSJ employs a single scan to add Gaussian noise to the XYZ coordinates and the intensity value of points below a randomly selected depth range." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.356, + 0.128 + ], + "angle": 0, + "content": "J.Park et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.237 + ], + "angle": 0, + "content": "Angle-Selective Jittering (ASJ). It applies Gaussian noise to a randomly selected range of angles. By performing jittering across different ranges of depths or angles for each point, this method reflects the augmentation that some beams are affected by geometric perturbation in a non-uniform manner in adverse weather. This concept realistically represents the characteristics of transparent droplets like rain and snow [6]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.238, + 0.784, + 0.298 + ], + "angle": 0, + "content": "A significant aspect of DSJ and ASJ is that they use only a single frame. DSJ and ASJ are efficient in data augmentation without additional LiDAR frames and can perform reasonable augmentation even when the given LiDAR data is not sequentially captured." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.298, + 0.784, + 0.358 + ], + "angle": 0, + "content": "Range Jittering (RJ). Range Jittering has been proposed to simulate range distortion caused by droplets and fog, as mentioned in [9]. Unlike DSJ and ASJ, Range Jittering applies jittering only in the range direction of points. This method is used in place of using the original points in DSJ and ASJ." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.379, + 0.443, + 0.394 + ], + "angle": 0, + "content": "4.2 Learnable Point Drop" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.403, + 0.785, + 0.599 + ], + "angle": 0, + "content": "Learnable Point Drop (LPD) was devised to address the second main distortion, point drop. LPD is designed to artificially create point drop due to occlusion, such as dense fog. LPD employs a Deep Q-Learning Network (DQN) [19] for identifying the drop ratio and drop region that lead to adverse effects on the model. The reward of the DQN is designed to identify point drops that increase the training loss and uncertainty of the LiDAR semantic segmentation model. Through LPD, the LiDAR semantic segmentation model is exposed to point drop scenarios caused by adverse weather conditions. Consequently, it learns to make accurate predictions even in the absence of points that are critical for segmentation in clean data environments. Since LPD is employed merely as a concept for data augmentation and exists as a separate module, it necessitates no alterations to the existing model's training scheme except for limiting the gradient norm to ensure the stability of DQN learning." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.599, + 0.784, + 0.655 + ], + "angle": 0, + "content": "LPD module defines its current state by summing the loss \\( L_{aug} \\), calculated from augmented data obtained through SJ, and the entropy \\( H_{aug} \\), derived from logits. The loss \\( L_{aug} \\) is calculated using the original loss function employed by the model in use. The calculation of entropy is as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.387, + 0.663, + 0.617, + 0.697 + ], + "angle": 0, + "content": "\\[\nH _ {a u g} (x) = - \\frac {1}{N} \\sum_ {i = 1} ^ {n} P (x _ {i}) \\log P (x _ {i})\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.705, + 0.785, + 0.809 + ], + "angle": 0, + "content": "In this process, \\( x_{i} \\) represents each point, and \\( P(\\cdot) \\) is softmax to the logits. \\( N \\) denotes the number of points. LPD predicts the indices of points to drop, taking as input the sum of \\( L_{\\text{aug}} \\) and \\( H_{\\text{aug}} \\) added to the input point tensor. This process enables simultaneous learning of the drop ratio and drop region. Subsequently, the difference between the sum of the loss \\( L_{\\text{LPD}} \\) and entropy \\( H_{\\text{LPD}} \\), obtained from the dropped points by LPD, and the sum of \\( L_{\\text{aug}} \\) and \\( H_{\\text{aug}} \\), is defined as the reward." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.784, + 0.84 + ], + "angle": 0, + "content": "In Comparison with random drop, random drop uniformly reduces points across all depths. Consequently, this approach is not suitable for simulating" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.233, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Rethinking Data Augmentation for Robust LiDAR Semantic Segmentation" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "image", + "bbox": [ + 0.229, + 0.15, + 0.788, + 0.263 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.303, + 0.788, + 0.374 + ], + "angle": 0, + "content": "Fig. 2: Description of Learnable Point Drop (LPD). (1) The LPD utilizes the loss \\( L_{\\text{aug}} \\) and entropy \\( H_{\\text{aug}} \\) derived from LiDAR segmentation model as its current state. (2) Upon receiving a point input, LPD performs a point drop, then recalculates the loss \\( L_{\\text{LPD}} \\) and entropy \\( H_{\\text{LPD}} \\). (3) The difference between these new values and the previous ones is used to compute the reward." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.403, + 0.788, + 0.48 + ], + "angle": 0, + "content": "point drops caused by fog, as it lacks the depth-specific characteristics of fog-related distortions. Furthermore, randomly generated point drop scans may not achieve the severity of adverse weather conditions necessary to drop critical points that could deceive the model. A detailed comparison will be conducted in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.503, + 0.396, + 0.518 + ], + "angle": 0, + "content": "4.3 Overall Pipeline" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.529, + 0.789, + 0.697 + ], + "angle": 0, + "content": "The overall training procedure begins with data augmentation through SJ. The augmented points are input into the LiDAR semantic segmentation model to compute loss. The model then calculates logits for each point, applies softmax, and subsequently computes the entropy. The current state of the LPD module is set as the sum of the loss and entropy. The dropped points are then input into the LiDAR semantic segmentation model to calculate the corresponding loss and entropy. The difference between the sum of loss/entropy obtained from the augmented points and that obtained from the points determined by LPD is set as the reward. Through this process, the LPD module enables the LiDAR semantic segmentation model to effectively learn points distorted due to adverse weather. Overall process is detailed in Fig. 3" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.72, + 0.376, + 0.737 + ], + "angle": 0, + "content": "5 Experiments" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.754, + 0.428, + 0.77 + ], + "angle": 0, + "content": "5.1 Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Dataset and Evaluation Metrics. To assess robustness, we trained our model on 19 classes from SemanticKITTI and then validated its performance on validation set of SemanticSTF as same as [33]. The experiment process involves training on clean data followed by evaluating robustness in adverse conditions," + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.356, + 0.127 + ], + "angle": 0, + "content": "J.Park et al." + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.147, + 0.787, + 0.29 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.314, + 0.788, + 0.37 + ], + "angle": 0, + "content": "Fig. 3: The overall training process. (1) Initially, SJ is applied. (2) Following this, we calculate the loss, which leads to the update of the LPD state. (3) This process is followed by utilizing LPD to generate dropped points. (4) Subsequent to recalculating the loss, the reward is accordingly updated." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.4, + 0.784, + 0.49 + ], + "angle": 0, + "content": "akin to existing research [11, 33, 34]. SemanticSTF is a dataset gathered under real-world adverse conditions such as rain, fog, and snow. Following the approach in [33], invalid data in SemanticSTF are mapped to an \"ignore label\". We utilized MinkowskiNet [5] for validation purposes due to its well-known robustness [11, 34]. Our evaluation metrics include the Intersection over Union (IoU) for each class and the mean IoU (mIoU) across all classes." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.49, + 0.785, + 0.596 + ], + "angle": 0, + "content": "Implementation Details. MinkowskiNet-18/32width served as our baseline model. The learning rate was set at 0.24, with a weight decay of 0.0001. The mean and standard deviation of Gaussian noise used in SJ are 0 and 0.01 each. In cases of Learnable Point Drop, the norm of the gradient was limited to 100 for both the DQN and Segmentation model. All experiments were conducted using four A6000 GPUs for 15 epochs in batch size 2. The duration of all experiments ranged between 3 to 5 hours." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.619, + 0.373, + 0.633 + ], + "angle": 0, + "content": "5.2 Main Results" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.644, + 0.788, + 0.81 + ], + "angle": 0, + "content": "SemanticKITTI to SemanticSTF. According to Tab. 2 (a), our proposed method showed a significant improvement in SemanticSTF, with a \\(+8.1\\) mIoU increase over the baseline and \\(+5.6\\) mIoU over the recent competitor, PointDR. Our method demonstrated universally superior performance across all weather conditions, validating the validity of our choice in data distortion type and the effectiveness of our method. Our method substantially improved the performance of categories like other vehicle, motorcyclist, sidewalk, pole, and traffic sign, which had lower baseline performance, increasing their class IoU by about \\(+5\\) to \\(+10\\). Additionally, it significantly enhanced the mIoU for cars and persons by \\(+19.0\\) and \\(+9.6\\), respectively, compared to the baseline, which is crucial for safety in driving environments. More details are available in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Also, in Tab. 2 (a), examining the performance improvement per weather condition, we note that the most significant increase in mIoU was in rain, with" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.233, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Rethinking Data Augmentation for Robust LiDAR Semantic Segmentation" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.145, + 0.788, + 0.228 + ], + "angle": 0, + "content": "Table 2: LiDAR segmentation results (mIoU) on the SemanticSTF validation set of models trained with (a) SemanticKITTI and (b) SynLiDAR [32]. D-fog and L-fog denote dense fog and light fog weather conditions in all experiments. * denotes the reproduced result with the same segmentation backbone. The symbol \\(\\ddagger\\) indicates that the validation for model selection was performed on sequence 0 of SynLiDAR, rather than SemanticSTF. Best score is in bold and second best is underlined." + }, + { + "type": "table", + "bbox": [ + 0.226, + 0.253, + 0.492, + 0.376 + ], + "angle": 0, + "content": "
(a) SemanticKITTI→SemanticSTF
MethodsD-fogL-fogRainSnowmIoU
Oracle51.954.657.953.754.7
Baseline30.730.129.725.331.4
LaserMix [13]23.215.59.37.814.7
PolarMix [31]21.314.916.59.315.3
PointDR* [33]37.333.535.526.933.9
Baseline+SJ+LPD36.037.537.633.139.5
Increments to baseline+5.3+7.4+7.9+7.8+8.1
" + }, + { + "type": "table", + "bbox": [ + 0.503, + 0.24, + 0.794, + 0.389 + ], + "angle": 0, + "content": "
(b) SynLiDAR→SemanticSTF
MethodsD-fogL-fogRainSnowmIoU
Oracle51.954.657.953.754.7
Baseline15.2415.9716.8312.7615.45
LaserMix [13]15.3217.9518.5513.816.85
PolarMix [31]16.4718.6919.6315.9818.09
PointDR* [33]19.0920.2825.2918.9819.78
PointDR*‡21.4120.9425.4819.3120.47
Baseline+SJ+LPD19.0820.6521.9717.2720.08
Increments to baseline+3.8+4.7+5.1+4.5+4.6
Baseline+SJ+LPD‡18.9921.2223.1417.2820.51
Increments to baseline+3.7+5.3+6.3+4.5+5.1
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.436, + 0.788, + 0.572 + ], + "angle": 0, + "content": "+7.9 mIoU. This result is indicative of our SJ augmentation effectively reflecting geometric perturbations caused by raindrops. This aspect will be further explored in the ablation study. Additionally, in snow, our method showed a remarkable increase of +7.8 mIoU, and in dense fog (+5.3 mIoU) and light fog (+7.4 mIoU), our methodology clearly outperformed the baseline. This indicates that our data augmentation effectively models the adverse impacts of weather on LiDAR semantic segmentation models. Detailed analysis of how our method enhances robustness against each weather condition, and the factors contributing to the improved performance, will be discussed in the ablation study." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.581, + 0.788, + 0.657 + ], + "angle": 0, + "content": "Moreover, as previously mentioned, the training architectures designed to enhance robustness, such as those proposed in PointDR [11, 33, 34], can be orthogonally applied to our proposed data augmentation. Therefore, regardless of the choice of training architecture, the integration of our method is expected to yield superior performance compared to conventional approaches." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.665, + 0.788, + 0.726 + ], + "angle": 0, + "content": "SynLiDAR to SemanticSTF. For SynLiDAR, unlike SemanticKITTI, where no specific validation set was available, sequence 0 was treated as the validation set. The best-performing models during these validation checkpoints were selected for reporting." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.788, + 0.841 + ], + "angle": 0, + "content": "In contrast, PointDR demonstrated an unrealistic scenario in its baseline code, using SemanticSTF for validation during training and reporting only the highest-performing results. For a fair comparison using PointDR's criteria, our method shows a performance increase of \\(+0.04\\) mIoU over PointDR. Using our more realistic protocol, our method demonstrates an even greater performance improvement of \\(+0.32\\) mIoU against PointDR. Detailed results are shown in Tab. 2 (b)." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.356, + 0.127 + ], + "angle": 0, + "content": "J.Park et al." + }, + { + "type": "image", + "bbox": [ + 0.216, + 0.144, + 0.788, + 0.362 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.383, + 0.789, + 0.44 + ], + "angle": 0, + "content": "Fig. 4: Qualitative results of our method from validation set of SemanticSTF. All models are trained in train set of SemanticKITTI. Green (or red) points indicate correct (or incorrect) predictions, showing our predictions are more accurate than PointDR in all weather, especially on road and sidewalk." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.469, + 0.46, + 0.485 + ], + "angle": 0, + "content": "5.3 Additional Experiments" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.493, + 0.789, + 0.643 + ], + "angle": 0, + "content": "Analysis of Proposed Component Methods. As shown in As Tab. 3, when ASJ was added to the baseline, we observed an improvement in robustness across all adverse weather conditions (increases of \\(+2.6\\mathrm{mIoU}\\) in dense fog, \\(+5.3\\mathrm{mIoU}\\) in light fog, \\(+8.1\\mathrm{mIoU}\\) in rain, and \\(+6.3\\mathrm{mIoU}\\) in snow), albeit with a \\(-1.8\\mathrm{mIoU}\\) decrease in clean data performance. The introduction of DSJ further enhanced performance in dense fog (\\(+4.1\\mathrm{mIoU}\\)), light fog (\\(+6.3\\mathrm{mIoU}\\)), rain (\\(+9.3\\mathrm{mIoU}\\)), snow (\\(+4.6\\mathrm{mIoU}\\)), and an overall performance increase of \\(+6.2\\mathrm{mIoU}\\). Our Selective Jittering method's strong performance in rain conditions showcases its efficacy in rainy weather, where it considers geometric perturbations for selected beams instead of all." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.644, + 0.789, + 0.795 + ], + "angle": 0, + "content": "The application of Range Jittering showed improvements in dense fog(+3.7 mIoU), light fog (+6.9 mIoU), rain(+6.0 mIoU), snow (+8.2 mIoU), and an overall performance increase of +7.3 mIoU, indicating the importance of simulating range distortion due to adverse weather as discussed in [9]. As seen earlier, in rainy weather, jittering only some beams aids in enhancing robustness. However, performing Range Jittering on the remaining original points leads to all points being jittered, reducing performance in rain conditions. Additionally, the improvement in snow conditions by using Range Jittering suggests that perturbing as many beams as possible is more effective for robustness enhancement in snowy weather." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.789, + 0.84 + ], + "angle": 0, + "content": "The addition of LPD resulted in improvements in dense fog \\((+5.3\\mathrm{mIoU})\\) light fog \\((+7.4\\mathrm{mIoU})\\) , rain \\((+7.9\\mathrm{mIoU})\\) , snow \\((+7.8\\mathrm{mIoU})\\) , and an overall increase of \\(+8.1\\mathrm{mIoU}\\) against to the baseline. The substantial performance en" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.233, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Rethinking Data Augmentation for Robust LiDAR Semantic Segmentation" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "table_caption", + "bbox": [ + 0.215, + 0.145, + 0.788, + 0.187 + ], + "angle": 0, + "content": "Table 3: Experiments on the components of our methods with SemanticKITTI as source and SemanticSTF as target. The values in parentheses indicate the performance improvement or decrease over the baseline model." + }, + { + "type": "table", + "bbox": [ + 0.317, + 0.199, + 0.677, + 0.32 + ], + "angle": 0, + "content": "
MethodsCleanD-fogL-fogRainSnowmIoU
Baseline63.930.730.129.725.331.4
+ASJ62.1 (-1.8)33.335.437.831.636.8 (+5.4)
+DSJ63.0 (-0.9)34.836.439.029.937.6 (+6.2)
+RJ61.2 (-2.7)33.437.035.733.538.7 (+7.3)
+LPD62.8 (-1.1)36.037.537.633.139.5 (+8.1)
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.353, + 0.788, + 0.458 + ], + "angle": 0, + "content": "hancements observed in conditions of dense and light fog indicate that the LPD effectively represents occlusion from foggy weather as intended. The performance reduction in snow conditions compared to the use of ASJ, DSJ, and Range Jittering \\((-0.4\\mathrm{mIoU})\\) is likely due to LPD not exclusively targeting the frozen or wet ground for point dropout. LPD applies the point drop across all points, as shown in Fig. 5. This represents a limitation of LPD, necessitating the development of alternative methods to address this issue." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.462, + 0.788, + 0.583 + ], + "angle": 0, + "content": "Overall, the use of all components led to a significant improvement in performance under adverse weather conditions, with an increase of \\(+8.1\\mathrm{mIoU}\\) over the baseline at a reasonable cost of \\(-1.1\\mathrm{mIoU}\\) in clean data. Ultimately, our method achieved state-of-the-art results on SemanticSTF, with an increase in mIoU across all adverse weather conditions compared to the baseline. The combined use of DSJ, ASJ, Range Jittering, and LPD demonstrated high performance, indicating that each component synergistically contributes to enhancing the model's robustness." + }, + { + "type": "image", + "bbox": [ + 0.33, + 0.627, + 0.676, + 0.741 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.767, + 0.788, + 0.837 + ], + "angle": 0, + "content": "Fig. 5: Qualitative results of LPD from validation set of SemanticKITTI. It can be observed that the point drop facilitated by the LPD occurs not just locally but extends across all points. While LPD effectively mimics the occlusions and sparse point distributions seen in real adverse weather, it does not replicate the effect of wet ground surfaces." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.356, + 0.127 + ], + "angle": 0, + "content": "J.Park et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.785, + 0.173 + ], + "angle": 0, + "content": "Table 4: LiDAR segmentation results (mIoU) with the various architecture and datasets. All models were trained with SemanticKITTI." + }, + { + "type": "table", + "bbox": [ + 0.327, + 0.185, + 0.672, + 0.306 + ], + "angle": 0, + "content": "
MethodSemanticSTFSemanticKITTI-C
CENet14.249.3
SPVCNN28.152.5
Minkowski31.453.0
CENet+Ours22.0 (+7.8)53.2 (+3.9)
SPVCNN+Ours38.4 (+10.3)52.9 (+0.4)
Minkowski+Ours39.5 (+8.1)58.6 (+5.6)
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.336, + 0.816, + 0.41 + ], + "angle": 0, + "content": "Experiments on Various Input Data Representation, Models and Datasets. We evaluated our method using different data representations, architectures, and datasets, specifically: (1) projection-based CENet and (2) voxel-based SPVCNN for backbones, and (3) the synthetic corruption dataset SemanticKITTI-C [11]. Tab. 4 shows our results." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.413, + 0.79, + 0.549 + ], + "angle": 0, + "content": "SemanticKITTI-C has a unique data distribution with more samples in classes, such as sidewalks and vegetation, than SemanticSTF. Our method consistently improves performance across various datasets and architectures. Specifically, for the SemanticSTF dataset, our method increases the mIoU of CENet by \\(+7.8\\), SPVCNN by \\(+10.3\\), and MinkowskiNet by \\(+8.1\\) mIoU. Similarly, for the SemanticKITTI-C dataset, the improvements are \\(+3.9\\) for CENet, \\(+0.4\\) for SPVCNN, and \\(+5.6\\) mIoU for MinkowskiNet. These consistent enhancements across different input data representations, models, and datasets highlight the robustness and generalizability of our approach." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.576, + 0.359, + 0.592 + ], + "angle": 0, + "content": "6 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.613, + 0.788, + 0.748 + ], + "angle": 0, + "content": "Our study focused on analyzing LiDAR data distortions in adverse weather and enhancing LiDAR model robustness with a data-centric method. We identified that point drop and geometric distortion mainly affect model performance through toy experiments. We introduced novel data augmentation techniques, Selective Jittering, and Learnable Point Drop, leading to state-of-the-art performance on SemanticSTF. We showed that our methodology significantly contributes to robustness enhancement without complex data simulations. We expect that our research will contribute to improving the stability and reliability of LiDAR semantic segmentation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.794, + 0.841 + ], + "angle": 0, + "content": "Acknowledgements. This work was supported by IITP grant funded by MSIT (No.2021-0-02068, Artificial Intelligence Innovation Hub and RS-2019-II190075, Artificial Intelligence Graduate School Program(KAIST)) and KEIT grant funded by MOTIE (No.2022-0-00680, No.2022-0-01045), NRF funded by the MSIP (NRF-2022R1A2C3011154, RS-2023-00219019, RS-2023-00240135) and Samsung Electronics Co., Ltd (IO230508-06190-01)." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.233, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Rethinking Data Augmentation for Robust LiDAR Semantic Segmentation" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.145, + 0.323, + 0.16 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.173, + 0.786, + 0.228 + ], + "angle": 0, + "content": "1. Ando, A., Gidaris, S., Bursuc, A., Puy, G., Boulch, A., Marlet, R.: Rangevit: Towards vision transformers for 3d semantic segmentation in autonomous driving. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 5240-5250 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.229, + 0.786, + 0.284 + ], + "angle": 0, + "content": "2. Behley, J., Garbade, M., Milioto, A., Quenzel, J., Behnke, S., Stachniss, C., Gall, J.: Semantickitti: A dataset for semantic scene understanding of lidar sequences. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 9297-9307 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.284, + 0.786, + 0.323 + ], + "angle": 0, + "content": "3. Bijelic, M., Gruber, T., Ritter, W.: A benchmark for lidar sensors in fog: Is detection breaking down? In: 2018 IEEE intelligent vehicles symposium (IV). pp. 760-767. IEEE (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.324, + 0.786, + 0.365 + ], + "angle": 0, + "content": "4. Choe, J., Park, C., Rameau, F., Park, J., Kweon, I.S.: Pointmixer: Mlp-mixer for point cloud understanding. In: European Conference on Computer Vision. pp. 620-640. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.365, + 0.786, + 0.405 + ], + "angle": 0, + "content": "5. Choy, C., Gwak, J., Savarese, S.: 4d spatio-temporal convnets: Minkowski convolutional neural networks. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 3075-3084 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.405, + 0.786, + 0.445 + ], + "angle": 0, + "content": "6. Fersch, T., Buhmann, A., Koelpin, A., Weigel, R.: The influence of rain on small aperture lidar sensors. In: 2016 German Microwave Conference (GeMiC). pp. 84-87. IEEE (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.446, + 0.786, + 0.5 + ], + "angle": 0, + "content": "7. Hahner, M., Sakaridis, C., Bijelic, M., Heide, F., Yu, F., Dai, D., Van Gool, L.: Lidar snowfall simulation for robust 3d object detection. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 16364-16374 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.501, + 0.786, + 0.542 + ], + "angle": 0, + "content": "8. Hahner, M., Sakaridis, C., Dai, D., Van Gool, L.: Fog simulation on real lidar point clouds for 3d object detection in adverse weather. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 15283-15292 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.542, + 0.786, + 0.582 + ], + "angle": 0, + "content": "9. Kilic, V., Hegde, D., Sindagi, V., Cooper, A.B., Foster, M.A., Patel, V.M.: Lidar light scattering augmentation (lisa): Physics-based simulation of adverse weather conditions for 3d object detection. arXiv preprint arXiv:2107.07004 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.582, + 0.786, + 0.623 + ], + "angle": 0, + "content": "0. Kong, L., Liu, Y., Chen, R., Ma, Y., Zhu, X., Li, Y., Hou, Y., Qiao, Y., Liu, Z.: Rethinking range view representation for lidar segmentation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 228-240 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.623, + 0.786, + 0.677 + ], + "angle": 0, + "content": "1. Kong, L., Liu, Y., Li, X., Chen, R., Zhang, W., Ren, J., Pan, L., Chen, K., Liu, Z.: Robo3d: Towards robust and reliable 3d perception against corruptions. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 19994-20006 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.678, + 0.786, + 0.718 + ], + "angle": 0, + "content": "2. Kong, L., Ren, J., Pan, L., Liu, Z.: Lasermix for semi-supervised lidar semantic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 21705-21715 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.718, + 0.786, + 0.759 + ], + "angle": 0, + "content": "3. Kong, L., Ren, J., Pan, L., Liu, Z.: Lasermix for semi-supervised lidar semantic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 21705-21715 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.759, + 0.786, + 0.8 + ], + "angle": 0, + "content": "4. Lai, X., Chen, Y., Lu, F., Liu, J., Jia, J.: Spherical transformer for lidar-based 3d recognition. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 17545-17555 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.8, + 0.786, + 0.84 + ], + "angle": 0, + "content": "5. Lee, S., Son, T., Kwak, S.: Fico: Learning fog-invariant features for foggy scene segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18911-18921 (2022)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.173, + 0.786, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.356, + 0.127 + ], + "angle": 0, + "content": "J.Park et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.785, + 0.189 + ], + "angle": 0, + "content": "16. Li, Y., Duthon, P., Colomb, M., Ibanez-Guzman, J.: What happens for a tof lidar in fog? IEEE Transactions on Intelligent Transportation Systems 22(11), 6670-6681 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.19, + 0.785, + 0.232 + ], + "angle": 0, + "content": "17. Li, Z., Wu, X., Wang, J., Guo, Y.: Weather-degraded image semantic segmentation with multi-task knowledge distillation. Image and Vision Computing 127, 104554 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.232, + 0.785, + 0.272 + ], + "angle": 0, + "content": "18. Milioto, A., Vizzo, I., Behley, J., Stachniss, C.: Rangenet++: Fast and accurate lidar semantic segmentation. In: 2019 IEEE/RSJ international conference on intelligent robots and systems (IROS). pp. 4213-4220. IEEE (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.273, + 0.785, + 0.314 + ], + "angle": 0, + "content": "19. Mnih, V., Kavukcuoglu, K., Silver, D., Rusu, A.A., Veness, J., Bellemare, M.G., Graves, A., Riedmiller, M., Fidjeland, A.K., Ostrovski, G., et al.: Human-level control through deep reinforcement learning. nature 518(7540), 529-533 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.315, + 0.785, + 0.355 + ], + "angle": 0, + "content": "20. Nekrasov, A., Schult, J., Litany, O., Leibe, B., Engelmann, F.: Mix3d: Out-of-context data augmentation for 3d scenes. In: 2021 international conference on 3d vision (3dv). pp. 116-125. IEEE (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.356, + 0.785, + 0.397 + ], + "angle": 0, + "content": "21. Puy, G., Boulch, A., Marlet, R.: Using a waffle iron for automotive point cloud semantic segmentation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 3379-3389 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.398, + 0.785, + 0.438 + ], + "angle": 0, + "content": "22. Qi, C.R., Su, H., Mo, K., Guibas, L.J.: Pointnet: Deep learning on point sets for 3d classification and segmentation. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 652-660 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.439, + 0.785, + 0.493 + ], + "angle": 0, + "content": "23. Qian, K., Zhu, S., Zhang, X., Li, L.E.: Robust multimodal vehicle detection in foggy weather using complementary lidar and radar signals. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 444-453 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.495, + 0.785, + 0.535 + ], + "angle": 0, + "content": "24. Ryu, K., Hwang, S., Park, J.: Instant domain augmentation for lidar semantic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9350-9360 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.536, + 0.785, + 0.577 + ], + "angle": 0, + "content": "25. Saltori, C., Galasso, F., Fiameni, G., Sebe, N., Ricci, E., Poiesi, F.: Cosmix: Compositional semantic mix for domain adaptation in 3d lidar segmentation. In: European Conference on Computer Vision. pp. 586-602. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.578, + 0.785, + 0.618 + ], + "angle": 0, + "content": "26. Shin, J., Park, H., Kim, T.: Characteristics of laser backscattering intensity to detect frozen and wet surfaces on roads. Journal of Sensors 2019(1), 8973248 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.619, + 0.785, + 0.66 + ], + "angle": 0, + "content": "27. Smith, B.E., Gardner, A., Schneider, A., Flanner, M.: Modeling biases in laser-altimetry measurements caused by scattering of green light in snow. Remote sensing of environment 215, 398-410 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.661, + 0.785, + 0.702 + ], + "angle": 0, + "content": "28. Thomas, H., Qi, C.R., Deschaud, J.E., Marcotegui, B., Goulette, F., Guibas, L.J.: Kpconv: Flexible and deformable convolution for point clouds. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 6411-6420 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.702, + 0.785, + 0.757 + ], + "angle": 0, + "content": "29. Tolstikhin, I.O., Houlsby, N., Kolesnikov, A., Beyer, L., Zhai, X., Unterthiner, T., Yung, J., Steiner, A., Keysers, D., Uszkoreit, J., et al.: Mlp-mixer: An allmlp architecture for vision. Advances in neural information processing systems 34, 24261-24272 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.758, + 0.785, + 0.799 + ], + "angle": 0, + "content": "30. Xiao, A., Huang, J., Guan, D., Cui, K., Lu, S., Shao, L.: Polarmix: A general data augmentation technique for lidar point clouds. Advances in Neural Information Processing Systems 35, 11035-11048 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.799, + 0.785, + 0.84 + ], + "angle": 0, + "content": "31. Xiao, A., Huang, J., Guan, D., Cui, K., Lu, S., Shao, L.: Polarmix: A general data augmentation technique for lidar point clouds. Advances in Neural Information Processing Systems 35, 11035-11048 (2022)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.233, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Rethinking Data Augmentation for Robust LiDAR Semantic Segmentation" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.788, + 0.189 + ], + "angle": 0, + "content": "32. Xiao, A., Huang, J., Guan, D., Zhan, F., Lu, S.: Transfer learning from synthetic to real lidar point cloud for semantic segmentation. In: Proceedings of the AAAI conference on artificial intelligence. vol. 36, pp. 2795-2803 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.19, + 0.788, + 0.245 + ], + "angle": 0, + "content": "33. Xiao, A., Huang, J., Xuan, W., Ren, R., Liu, K., Guan, D., El Saddik, A., Lu, S., Xing, E.P.: 3d semantic segmentation in the wild: Learning generalized models for adverse-condition point clouds. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9382-9392 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.246, + 0.788, + 0.286 + ], + "angle": 0, + "content": "34. Yan, X., Zheng, C., Xue, Y., Li, Z., Cui, S., Dai, D.: Benchmarking the robustness of lidar semantic segmentation models. International Journal of Computer Vision pp. 1-24 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.287, + 0.788, + 0.328 + ], + "angle": 0, + "content": "35. Yang, D., Liu, Z., Jiang, W., Yan, G., Gao, X., Shi, B., Liu, S., Cai, X.: Realistic rainy weather simulation for lidars in carla simulator. arXiv preprint arXiv:2312.12772 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.328, + 0.788, + 0.369 + ], + "angle": 0, + "content": "36. Zhao, H., Jiang, L., Jia, J., Torr, P.H., Koltun, V.: Point transformer. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 16259-16268 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.37, + 0.788, + 0.425 + ], + "angle": 0, + "content": "37. Zhu, X., Zhou, H., Wang, T., Hong, F., Ma, Y., Li, W., Li, H., Lin, D.: Cylindrical and asymmetrical 3d convolution networks for lidar segmentation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 9939-9948 (2021)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.788, + 0.425 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/Rethinking Data Augmentation for Robust LiDAR Semantic Segmentation in Adverse Weather/60de04d8-9845-4ba5-9238-a2dd5b5c4b07_origin.pdf b/2024/Rethinking Data Augmentation for Robust LiDAR Semantic Segmentation in Adverse Weather/60de04d8-9845-4ba5-9238-a2dd5b5c4b07_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..202313451facdf6951ad5a02390342f8f380a7a6 --- /dev/null +++ b/2024/Rethinking Data Augmentation for Robust LiDAR Semantic Segmentation in Adverse Weather/60de04d8-9845-4ba5-9238-a2dd5b5c4b07_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ea6cae8873ee6d2c5960c7852fdafd589e846ce8ead289ea21dd6e1704871708 +size 1911625 diff --git a/2024/Rethinking Data Augmentation for Robust LiDAR Semantic Segmentation in Adverse Weather/full.md b/2024/Rethinking Data Augmentation for Robust LiDAR Semantic Segmentation in Adverse Weather/full.md new file mode 100644 index 0000000000000000000000000000000000000000..4a0da338fbc4e56cd5a378d62b2cf5b016957eac --- /dev/null +++ b/2024/Rethinking Data Augmentation for Robust LiDAR Semantic Segmentation in Adverse Weather/full.md @@ -0,0 +1,246 @@ +# Rethinking Data Augmentation for Robust LiDAR Semantic Segmentation in Adverse Weather + +Junsung Park $^{1}$ , Kyungmin Kim $^{2}$ , and Hyunjung Shim $^{1\star}$ + +$^{1}$ Korea Advanced Institute of Science & Technology (KAIST) + $^{2}$ Yonsei University +jshackist@kaist.ac.kr, kyungmin.kim@yonsei.ac.kr, kateshim@kaist.ac.kr + +Abstract. Existing LiDAR semantic segmentation methods often struggle with performance declines in adverse weather conditions. Previous work has addressed this issue by simulating adverse weather or employing universal data augmentation during training. However, these methods lack a detailed analysis and understanding of how adverse weather negatively affects LiDAR semantic segmentation performance. Motivated by this issue, we identified key factors of adverse weather and conducted a toy experiment to pinpoint the main causes of performance degradation: (1) Geometric perturbation due to refraction caused by fog or droplets in the air and (2) Point drop due to energy absorption and occlusions. Based on these findings, we propose new strategic data augmentation techniques. First, we introduced a Selective Jittering (SJ) that jitters points in the random range of depth (or angle) to mimic geometric perturbation. Additionally, we developed a Learnable Point Drop (LPD) to learn vulnerable erase patterns with a Deep Q-Learning Network to approximate the point drop phenomenon from adverse weather conditions. Without precise weather simulation, these techniques strengthen the LiDAR semantic segmentation model by exposing it to vulnerable conditions identified by our data-centric analysis. Experimental results confirmed the suitability of the proposed data augmentation methods for enhancing robustness against adverse weather conditions. Our method achieves a notable $39.5\mathrm{mIoU}$ on the SemanticKITTI-to-SemanticSTF benchmark, improving the baseline by $8.1\%$ p and establishing a new state-of-the-art. Our code will be released at https://github.com/engineerJPark/LiDARWeather. + +Keywords: LiDAR Semantic Segmentation $\cdot$ Robust Learning for Adverse Weather $\cdot$ Data Augmentation + +# 1 Introduction + +LiDAR semantic segmentation is a fundamental task of 3D scene understanding, particularly in safety-critical applications like autonomous driving. However, existing LiDAR segmentation models [1,5,14,21,28,37] commonly lack robustness, + +showing significant performance degradation under adverse weather conditions such as snow, fog, rain, or wet surfaces. + +To address this issue, recent studies [11, 33, 34] have introduced corruption benchmarks for adverse weather and proposed effective techniques for robust LiDAR segmentation against corruption. These robust methods are primarily divided into task-agnostic and simulation-based approaches. The task-agnostic approach [11, 25, 33] employs general machine-learning techniques for robustness without explicitly addressing LiDAR corruption caused by adverse weather. The simulation-based approach [3,7-9,23,26,35] artificially synthesizes weather-specific data using physical equations for training. However, these efforts focus on detection tasks and consider only a single type of weather at a time. + +While the simulation-based approach leverages the intrinsic properties of Li-DAR scans under target adverse weather conditions, simulating every weather type at all possible severities is impractical and often inaccurate. Instead, we adopt a data-centric perspective to analyze corrupted LiDAR data. For example, distortions caused by rain, snow, and fog often create similar patterns, as simulated in [9]. That is, all those adverse weather shows point drop patterns due to attenuation or occlusion by droplet [6,7,9]. While the simulation-based approach requires explicit modeling of the LiDAR for each weather condition, our data-centric approach potentially addresses complex distortions with a few unified patterns. Through existing research and our comprehensive analysis, we find that various adverse weather conditions create similar distortion patterns in Li-DAR data. Based on our analysis, we categorize these distortions into two types: (1) geometric perturbation and (2) point drop. Our toy experiments demonstrate that these distortions are closely related to segmentation performance degradation. + +Rooted by the toy experiment, we introduce two novel and strategic data augmentation methods tailored to the LiDAR distortion caused by adverse weather. By incorporating these augmentations during training, we aim to bolster the model's robustness for each distortion type. For geometric perturbation, we introduce Selective Jittering (SJ). This method applies Additive Gaussian Noise (AGN) to alter the XYZ-coordinate and intensity within a selective local region. To handle a point drop pattern, we developed a Learnable Point Drop (LPD) that employs a Deep Q-Network (DQN) to strategically remove points. Our augmentation strategies are inspired by existing studies [9, 16, 27], which shows geometric perturbation typically involves small, random alterations to the original points. This insight led to our choice of jittering as a suitable technique for addressing the geometric perturbation. Also, our augmentation strategy for point drop is motivated by existing studies [7-9] which state that intensity and depth of LiDAR beams play a significant role. This prompts us to use a DQN to detect point drops and adapt to such point drop patterns. + +Finally, we apply the proposed augmentations when training the LiDAR segmentation model with a clean source dataset. Experimental results demonstrate that our method achieves new state-of-the-art performance on SemanticSTF, a real adverse weather dataset. Specifically, we achieved an impressive $+8.1\%$ + +mIoU gain over the baseline MinkowskiNet, nearly tripling the $+2.5\%$ p mIoU improvement from previous work [33]. Our method achieved remarkable performance gains, ranging from $+0.4$ to $+10.3\%$ p mIoU, across various model architectures (Minkowski, CENet, SPVCNN) and corruption dataset (SemanticSTF and SemanticKITTI-C). These results demonstrate the generalizability of the proposed augmentation methods. + +In summary, this paper presents several key contributions. + +- We identify two prevalent distortion types in LiDAR data caused by adverse weather, leading to performance degradation, through a data-centric analysis. +- We introduce two novel data augmentations tailored to each identified distortion type. +- Our method sets new state-of-the-art benchmarks on the SemanticKITTI to SemanticSTF benchmark. Notably, our improvement was $+8.1\%$ p mIoU without relying on precise simulations of adverse weather in LiDAR point inputs. This represents a tripling of the improvement over the baseline, compared to what previous methods have achieved [33]. + +# 2 Related Works + +# 2.1 LiDAR Semantic Segmentation + +Existing 3D LiDAR point cloud semantic segmentation methods can be categorized into three types based on the data representation: Point-based, Projection-based, and Voxel-based. + +Point-based methods [22, 28, 36] utilize the 3D points directly as input. KPConv [28] initially clustering local points, aggregating these local features, and then feeding them into kernel point convolutions. The Point Transformer [36] utilizes a transformer architecture to compute query points in each local region, obtained through k-nearest neighbors (kNN). The Point-Mixer [4] tried to adapt the MLP-Mixer [29] for point cloud applications. They achieve high performance but suffer from high computational costs due to the utilization of large-scale raw LiDAR data. + +Projection-based methods [1, 10, 18] project LiDAR points into a 2D image and performs the semantic segmentation using architectures successful in a 2D image. RangeViT [1] directly adopts a ViT model pre-trained on 2D images, demonstrating that pre-trained power in 2D images can be effective as prior knowledge in range images. RangeFormer [10] proposes "RangeAug" to maximize the utility of range images created by projecting into 2D, producing multiple range image data to overcome the low performance in range image models. Projection-based methods achieve fast inference speed but present a sub-optimal performance due to missing information after the projection. + +Voxel-based methods [5,14,37] perform efficient computations by dividing 3D space into a voxel grid and aggregating points within the same voxel. Minkowski-iNet [5] voxelizes LiDAR points with a cubic grid and applies sparse convolution. + +Cylinder3D [37] proposes cylindrical partitions, reflecting LiDAR's characteristic that the density of points depends on the distance. SphereFormer [14] utilizes a radial window and transformer architecture to aggregate long-range information and improve performance. Voxel-based methods achieve a balance between reasonable inference time and commendable segmentation performance. + +# 2.2 LiDAR Data Augmentation + +Inspired by 2D image augmentation, conventional LiDAR segmentation methods apply classic scaling, rotation, flipping, and translation to augment LiDAR data. Recently, several out-of-context augmentation techniques [12, 20, 30] that mix different LiDAR scans have been proposed. Mix3D [20] combines randomly selected two scans. Considering the sweeping mechanism of the LiDAR sensor, PolarMix [30] cuts LiDAR scans along the azimuth axis, then exchanges point cloud sectors and applies instance-level rotate-pasting. To reflect the spatial prior of the LiDAR point cloud, LaserMix [12] partitions LiDAR scans based on the laser beams and blends partitions from different LiDAR scans. Recent study [24] introduces a fast LiDAR domain augmentation module to address sensor-bias problems. To the best of our knowledge, our approach is the first augmentation method specifically designed to address data corruption under adverse weather conditions. + +# 2.3 LiDAR under Adverse Weather Conditions + +Robustness under harsh conditions is crucial in safety-critical applications. Adverse weather substantially degrade performance in real-world outdoor autonomous driving. Thus, there are several attempts to develop weather-robust models in fields such as 2D segmentation [15, 17], 3D detection [7, 8, 11], and 3D segmentation [11, 33]. Simulation-based approaches [7, 8] artificially synthesize data for single weather conditions through physical modeling and utilize it for training. We differ from these in that we do not model specific weather conditions explicitly. Recently proposed task-agnostic approaches [11, 25, 33] consider multiple weather conditions at once. However, they use general machine-learning techniques (such as teacher-student framework and feature prototype) to achieve robustness rather than specifically tackle LiDAR corruption caused by adverse weather. We differ from these in that we propose augmentation methods specifically tailored for adverse weather conditions based on the analysis of performance degradation in LiDAR data. + +# 3 Finding Distortions to Augment + +In this section, we aim to discuss the patterns of distortion that different adverse weather conditions impose on LiDAR data. Although adverse weather conditions are distinct in reality, studies have shown that their effects on LiDAR data often result in similar impacts. For instance, the distortions caused by "rain", "snow", + +and "fog" tend to produce similar point-missing patterns due to attenuation in the data, as demonstrated in [9]. Therefore, this section will focus on identifying the common distortion patterns caused by adverse weather through existing studies. Overall, existing studies describe the effect of adverse weather as four different types of distortion: (1) Point Drop due to energy absorption, (2) Occlusions caused by droplets of rain or snow and fog, (3) Geometric perturbation, and (4) Intensity distortion due to energy absorption. + +# 3.1 Distortion Factors from Adverse Weather + +(D1) Point Drop. Several studies have explored how adverse weather conditions contribute to point drops in LiDAR data. Kilic et al. [9], Fersch et al. [6] and Shin et al. [26] describe point drops resulting from beam attenuation and beam missing due to droplets, fog, and frozen or wet ground. These studies collectively suggest that adverse weather conditions typically lead to point drops in LiDAR data. +(D2) Occlusions. Several studies have addressed occlusions caused by adverse weather conditions. Hahner et al. [7], Kilic et al. [9], Kong et al. [11] and Yan et al. [34] consider scenarios where beams colliding with snow collect signals at much shorter distances than the intended objects of collision. Upon reviewing these studies, we have concluded that adverse weather consistently leads to occlusions. +(D3) Geometric Perturbation. Some studies focus on geometric perturbation caused by adverse weather conditions. Kilic et al. [9], Li et al. [16] and Smith et al. [27] demonstrated geometric perturbation in adverse weather, such as fog, snow, and rain, by incorporating random noise into the coordinates. Through these studies, we have come to conclude that adverse weather universally causes geometric perturbation. +(D4) Intensity Distortion. Numerous studies have focused on intensity distortion caused by adverse weather conditions. Bijelic et al. [3], Shin et al. [26], Fersch et al. [6], Kong et al. [11], and Yan et al. [34] have collectively shown that adverse weather conditions like fog, wetness, and rain lead to a reduction in Li-DAR beam intensity, influencing the generation of synthetic data. Through these studies, we conclude that adverse weather commonly causes intensity distortion. + +# 3.2 Toy Experiment + +Based on previous research findings mentioned in Section 3.1, the types of distortions in LiDAR point clouds caused by adverse weather conditions converge into a set of common distortions. Therefore, from a data-centric perspective, the issues that we need to consider can be summarized as follows: (D1) Point Drop, (D2) Occlusions, (D3) Geometric Perturbation, and (D4) Intensity Distortion. Here, we aim to identify which distortion types negatively impact performance. To achieve this, we have generated four distortion types of toy synthetic data through the SemanticKITTI validation set. + +Table 1: Results of toy experiments from validation set of SemanticKITTI [2]. Soft and Hard indicate the severity of the distortions. + +
Distortioncarbileemtcletruckoth-v.pers.biclstmt-clstroadparki.sidew.oth-g.build.fenceveget.trunkterra.poletraf.mIoU
Clean96.822.266.989.165.267.084.70.093.850.681.40.191.163.088.167.774.563.847.763.9
D1: Soft96.317.963.487.163.764.381.40.092.747.579.70.190.162.387.364.274.161.141.661.8
D1: Hard86.10.212.721.040.010.124.40.04.210.211.10.071.747.877.034.627.630.417.627.7
D2: Soft94.816.351.766.353.959.252.10.092.544.979.20.189.761.487.363.274.359.240.257.2
D2: Hard81.30.55.413.135.57.60.70.02.68.510.80.071.145.776.332.927.326.716.124.3
D3: Soft96.215.456.758.651.351.878.10.066.333.946.10.086.161.684.462.056.861.544.053.2
D3: Hard93.89.738.319.431.435.255.20.09.912.312.20.048.940.871.455.531.258.740.835.0
D4: Soft96.321.761.189.461.868.983.20.092.944.479.10.190.456.786.968.271.163.948.362.3
D4: Hard95.017.250.481.457.064.979.70.090.738.864.70.988.445.383.562.850.560.147.956.8
+ +- (D1) Point Drop: It considers the scenario where each LiDAR point disappears randomly and independently due to severe adverse weather. We removed individual points randomly to synthesize these data. We set the drop ratios at 0.5 and 0.9. +- (D2) Occlusions: The method assumes that occlusions occur predominantly in front of objects, mainly due to distortion from fog, snow, and rain. We synthesized this data by randomly selecting points and altering their depth to one-tenth of the original depth. The selecting ratios were also determined to be 0.5 and 0.9. +- (D3) Geometric Perturbation: This aspect assumes the distortion of point coordinates resulting from adverse weather conditions. We synthesized data by adding Gaussian noise to the coordinates of all points. The Gaussian noise levels were established at 0.05 and 0.25. +- (D4) Intensity Distortion: This distortion type assumes intensity attenuation caused by fog, raindrops, and snow particles. We synthesized this data by subtracting Gaussian noise from the intensity of all points. The Gaussian noise levels were established at 0.05 and 0.25. + +We selected MinkowskiNet [5] for our toy experiment. This choice was based on existing research [11, 34], which indicates that MinkowskiNet is a standard and robust model. + +In Tab. 1, it's evident that with increasing distortion in D1, D2, and D3, performance falls below half of the baseline. Contrarily, D4 maintains performance levels despite significant distortion. This performance degradation likely stems from changes in the local geometric structure where model computations occur. For this reason, D4 did not significantly impact performance. + +Importantly, D2's notable performance drop is argued to result from a point drop akin to D1. It is due to occlusions altering the point's local geometric structure, an effect similar to point drop. This is evidenced in Fig. 1, where incorrect predictions from D2 show patterns resembling those in D1. + +In summary, our observations indicate that geometric perturbation and point drop are the most impactful distortions from adverse weather. Data augmentation that replicates these phenomena could likely boost the model's robustness + +![](images/82090c9f50fa12ab3789eb06cbd45733571c834c844de401e608b4c407606628.jpg) +Fig. 1: Visualization for the results of toy experiment. In all figures, green (or red) points indicate correct (or incorrect) predictions. Point drop, occlusion, and geometric perturbation are the main distortions that lead to performance degradation. Also, since the misprediction areas of point drop and occlusion largely overlap, it's reasonable to consider them as the same type of distortion. + +against such conditions without explicit weather simulations. Thus, robust Li-DAR semantic segmentation models against adverse weather involve two key steps: (1) Training on specific adverse point drops detrimental to performance, and (2) Training with minor adjustments in point coordinates. + +# 4 Methods + +In this section, we propose solutions for the two challenges identified in Section 3.2. Our objective is to enhance the robustness of the model by addressing the common distortions exhibited by various adverse weather conditions, through data augmentation. This approach aims to achieve robustness without the necessity of simulating each adverse weather scenario explicitly. + +As observed earlier, distortions resulting from adverse weather include point drop, occlusion, geometric perturbation, and intensity distortion. Of these, the main distortions identified as having the most detrimental impact on performance were (1) geometric perturbation caused by minor variations in point coordinates and (2) point drop caused by beam missing or occlusion. To address these challenges, we propose two techniques: Selective Jittering and Learnable Point Drop. + +# 4.1 Selective Jittering + +Selective Jittering is devised to address the first main distortion, geometric perturbation. Selective Jittering consists of two types: Depth-Selective Jittering (DSJ) and Angle-Selective Jittering (ASJ). + +Depth-Selective Jittering (DSJ). DSJ employs a single scan to add Gaussian noise to the XYZ coordinates and the intensity value of points below a randomly selected depth range. + +Angle-Selective Jittering (ASJ). It applies Gaussian noise to a randomly selected range of angles. By performing jittering across different ranges of depths or angles for each point, this method reflects the augmentation that some beams are affected by geometric perturbation in a non-uniform manner in adverse weather. This concept realistically represents the characteristics of transparent droplets like rain and snow [6]. + +A significant aspect of DSJ and ASJ is that they use only a single frame. DSJ and ASJ are efficient in data augmentation without additional LiDAR frames and can perform reasonable augmentation even when the given LiDAR data is not sequentially captured. + +Range Jittering (RJ). Range Jittering has been proposed to simulate range distortion caused by droplets and fog, as mentioned in [9]. Unlike DSJ and ASJ, Range Jittering applies jittering only in the range direction of points. This method is used in place of using the original points in DSJ and ASJ. + +# 4.2 Learnable Point Drop + +Learnable Point Drop (LPD) was devised to address the second main distortion, point drop. LPD is designed to artificially create point drop due to occlusion, such as dense fog. LPD employs a Deep Q-Learning Network (DQN) [19] for identifying the drop ratio and drop region that lead to adverse effects on the model. The reward of the DQN is designed to identify point drops that increase the training loss and uncertainty of the LiDAR semantic segmentation model. Through LPD, the LiDAR semantic segmentation model is exposed to point drop scenarios caused by adverse weather conditions. Consequently, it learns to make accurate predictions even in the absence of points that are critical for segmentation in clean data environments. Since LPD is employed merely as a concept for data augmentation and exists as a separate module, it necessitates no alterations to the existing model's training scheme except for limiting the gradient norm to ensure the stability of DQN learning. + +LPD module defines its current state by summing the loss $L_{aug}$ , calculated from augmented data obtained through SJ, and the entropy $H_{aug}$ , derived from logits. The loss $L_{aug}$ is calculated using the original loss function employed by the model in use. The calculation of entropy is as follows: + +$$ +H _ {a u g} (x) = - \frac {1}{N} \sum_ {i = 1} ^ {n} P (x _ {i}) \log P (x _ {i}) +$$ + +In this process, $x_{i}$ represents each point, and $P(\cdot)$ is softmax to the logits. $N$ denotes the number of points. LPD predicts the indices of points to drop, taking as input the sum of $L_{\text{aug}}$ and $H_{\text{aug}}$ added to the input point tensor. This process enables simultaneous learning of the drop ratio and drop region. Subsequently, the difference between the sum of the loss $L_{\text{LPD}}$ and entropy $H_{\text{LPD}}$ , obtained from the dropped points by LPD, and the sum of $L_{\text{aug}}$ and $H_{\text{aug}}$ , is defined as the reward. + +In Comparison with random drop, random drop uniformly reduces points across all depths. Consequently, this approach is not suitable for simulating + +![](images/b70ceab1da88d04eff57b3b5ea4d42faf5f504c9b8a890620223ab24a564a38b.jpg) +Fig. 2: Description of Learnable Point Drop (LPD). (1) The LPD utilizes the loss $L_{\text{aug}}$ and entropy $H_{\text{aug}}$ derived from LiDAR segmentation model as its current state. (2) Upon receiving a point input, LPD performs a point drop, then recalculates the loss $L_{\text{LPD}}$ and entropy $H_{\text{LPD}}$ . (3) The difference between these new values and the previous ones is used to compute the reward. + +point drops caused by fog, as it lacks the depth-specific characteristics of fog-related distortions. Furthermore, randomly generated point drop scans may not achieve the severity of adverse weather conditions necessary to drop critical points that could deceive the model. A detailed comparison will be conducted in the supplementary material. + +# 4.3 Overall Pipeline + +The overall training procedure begins with data augmentation through SJ. The augmented points are input into the LiDAR semantic segmentation model to compute loss. The model then calculates logits for each point, applies softmax, and subsequently computes the entropy. The current state of the LPD module is set as the sum of the loss and entropy. The dropped points are then input into the LiDAR semantic segmentation model to calculate the corresponding loss and entropy. The difference between the sum of loss/entropy obtained from the augmented points and that obtained from the points determined by LPD is set as the reward. Through this process, the LPD module enables the LiDAR semantic segmentation model to effectively learn points distorted due to adverse weather. Overall process is detailed in Fig. 3 + +# 5 Experiments + +# 5.1 Experimental Setup + +Dataset and Evaluation Metrics. To assess robustness, we trained our model on 19 classes from SemanticKITTI and then validated its performance on validation set of SemanticSTF as same as [33]. The experiment process involves training on clean data followed by evaluating robustness in adverse conditions, + +![](images/3d4e4751c2405ef68b580baaa1320e0f2611933d089a0a7374db71eeee391e64.jpg) +Fig. 3: The overall training process. (1) Initially, SJ is applied. (2) Following this, we calculate the loss, which leads to the update of the LPD state. (3) This process is followed by utilizing LPD to generate dropped points. (4) Subsequent to recalculating the loss, the reward is accordingly updated. + +akin to existing research [11, 33, 34]. SemanticSTF is a dataset gathered under real-world adverse conditions such as rain, fog, and snow. Following the approach in [33], invalid data in SemanticSTF are mapped to an "ignore label". We utilized MinkowskiNet [5] for validation purposes due to its well-known robustness [11, 34]. Our evaluation metrics include the Intersection over Union (IoU) for each class and the mean IoU (mIoU) across all classes. + +Implementation Details. MinkowskiNet-18/32width served as our baseline model. The learning rate was set at 0.24, with a weight decay of 0.0001. The mean and standard deviation of Gaussian noise used in SJ are 0 and 0.01 each. In cases of Learnable Point Drop, the norm of the gradient was limited to 100 for both the DQN and Segmentation model. All experiments were conducted using four A6000 GPUs for 15 epochs in batch size 2. The duration of all experiments ranged between 3 to 5 hours. + +# 5.2 Main Results + +SemanticKITTI to SemanticSTF. According to Tab. 2 (a), our proposed method showed a significant improvement in SemanticSTF, with a $+8.1$ mIoU increase over the baseline and $+5.6$ mIoU over the recent competitor, PointDR. Our method demonstrated universally superior performance across all weather conditions, validating the validity of our choice in data distortion type and the effectiveness of our method. Our method substantially improved the performance of categories like other vehicle, motorcyclist, sidewalk, pole, and traffic sign, which had lower baseline performance, increasing their class IoU by about $+5$ to $+10$ . Additionally, it significantly enhanced the mIoU for cars and persons by $+19.0$ and $+9.6$ , respectively, compared to the baseline, which is crucial for safety in driving environments. More details are available in the supplementary material. + +Also, in Tab. 2 (a), examining the performance improvement per weather condition, we note that the most significant increase in mIoU was in rain, with + +
(a) SemanticKITTI→SemanticSTF
MethodsD-fogL-fogRainSnowmIoU
Oracle51.954.657.953.754.7
Baseline30.730.129.725.331.4
LaserMix [13]23.215.59.37.814.7
PolarMix [31]21.314.916.59.315.3
PointDR* [33]37.333.535.526.933.9
Baseline+SJ+LPD36.037.537.633.139.5
Increments to baseline+5.3+7.4+7.9+7.8+8.1
+ +Table 2: LiDAR segmentation results (mIoU) on the SemanticSTF validation set of models trained with (a) SemanticKITTI and (b) SynLiDAR [32]. D-fog and L-fog denote dense fog and light fog weather conditions in all experiments. * denotes the reproduced result with the same segmentation backbone. The symbol $\ddagger$ indicates that the validation for model selection was performed on sequence 0 of SynLiDAR, rather than SemanticSTF. Best score is in bold and second best is underlined. + +
(b) SynLiDAR→SemanticSTF
MethodsD-fogL-fogRainSnowmIoU
Oracle51.954.657.953.754.7
Baseline15.2415.9716.8312.7615.45
LaserMix [13]15.3217.9518.5513.816.85
PolarMix [31]16.4718.6919.6315.9818.09
PointDR* [33]19.0920.2825.2918.9819.78
PointDR*‡21.4120.9425.4819.3120.47
Baseline+SJ+LPD19.0820.6521.9717.2720.08
Increments to baseline+3.8+4.7+5.1+4.5+4.6
Baseline+SJ+LPD‡18.9921.2223.1417.2820.51
Increments to baseline+3.7+5.3+6.3+4.5+5.1
+ ++7.9 mIoU. This result is indicative of our SJ augmentation effectively reflecting geometric perturbations caused by raindrops. This aspect will be further explored in the ablation study. Additionally, in snow, our method showed a remarkable increase of +7.8 mIoU, and in dense fog (+5.3 mIoU) and light fog (+7.4 mIoU), our methodology clearly outperformed the baseline. This indicates that our data augmentation effectively models the adverse impacts of weather on LiDAR semantic segmentation models. Detailed analysis of how our method enhances robustness against each weather condition, and the factors contributing to the improved performance, will be discussed in the ablation study. + +Moreover, as previously mentioned, the training architectures designed to enhance robustness, such as those proposed in PointDR [11, 33, 34], can be orthogonally applied to our proposed data augmentation. Therefore, regardless of the choice of training architecture, the integration of our method is expected to yield superior performance compared to conventional approaches. + +SynLiDAR to SemanticSTF. For SynLiDAR, unlike SemanticKITTI, where no specific validation set was available, sequence 0 was treated as the validation set. The best-performing models during these validation checkpoints were selected for reporting. + +In contrast, PointDR demonstrated an unrealistic scenario in its baseline code, using SemanticSTF for validation during training and reporting only the highest-performing results. For a fair comparison using PointDR's criteria, our method shows a performance increase of $+0.04$ mIoU over PointDR. Using our more realistic protocol, our method demonstrates an even greater performance improvement of $+0.32$ mIoU against PointDR. Detailed results are shown in Tab. 2 (b). + +![](images/ad4edbac792c147da1e45e3c3a96f6cafa4ab7343688f7c51c50e565cd907cf6.jpg) +Fig. 4: Qualitative results of our method from validation set of SemanticSTF. All models are trained in train set of SemanticKITTI. Green (or red) points indicate correct (or incorrect) predictions, showing our predictions are more accurate than PointDR in all weather, especially on road and sidewalk. + +# 5.3 Additional Experiments + +Analysis of Proposed Component Methods. As shown in As Tab. 3, when ASJ was added to the baseline, we observed an improvement in robustness across all adverse weather conditions (increases of $+2.6\mathrm{mIoU}$ in dense fog, $+5.3\mathrm{mIoU}$ in light fog, $+8.1\mathrm{mIoU}$ in rain, and $+6.3\mathrm{mIoU}$ in snow), albeit with a $-1.8\mathrm{mIoU}$ decrease in clean data performance. The introduction of DSJ further enhanced performance in dense fog ( $+4.1\mathrm{mIoU}$ ), light fog ( $+6.3\mathrm{mIoU}$ ), rain ( $+9.3\mathrm{mIoU}$ ), snow ( $+4.6\mathrm{mIoU}$ ), and an overall performance increase of $+6.2\mathrm{mIoU}$ . Our Selective Jittering method's strong performance in rain conditions showcases its efficacy in rainy weather, where it considers geometric perturbations for selected beams instead of all. + +The application of Range Jittering showed improvements in dense fog(+3.7 mIoU), light fog (+6.9 mIoU), rain(+6.0 mIoU), snow (+8.2 mIoU), and an overall performance increase of +7.3 mIoU, indicating the importance of simulating range distortion due to adverse weather as discussed in [9]. As seen earlier, in rainy weather, jittering only some beams aids in enhancing robustness. However, performing Range Jittering on the remaining original points leads to all points being jittered, reducing performance in rain conditions. Additionally, the improvement in snow conditions by using Range Jittering suggests that perturbing as many beams as possible is more effective for robustness enhancement in snowy weather. + +The addition of LPD resulted in improvements in dense fog $(+5.3\mathrm{mIoU})$ light fog $(+7.4\mathrm{mIoU})$ , rain $(+7.9\mathrm{mIoU})$ , snow $(+7.8\mathrm{mIoU})$ , and an overall increase of $+8.1\mathrm{mIoU}$ against to the baseline. The substantial performance en + +Table 3: Experiments on the components of our methods with SemanticKITTI as source and SemanticSTF as target. The values in parentheses indicate the performance improvement or decrease over the baseline model. + +
MethodsCleanD-fogL-fogRainSnowmIoU
Baseline63.930.730.129.725.331.4
+ASJ62.1 (-1.8)33.335.437.831.636.8 (+5.4)
+DSJ63.0 (-0.9)34.836.439.029.937.6 (+6.2)
+RJ61.2 (-2.7)33.437.035.733.538.7 (+7.3)
+LPD62.8 (-1.1)36.037.537.633.139.5 (+8.1)
+ +hancements observed in conditions of dense and light fog indicate that the LPD effectively represents occlusion from foggy weather as intended. The performance reduction in snow conditions compared to the use of ASJ, DSJ, and Range Jittering $(-0.4\mathrm{mIoU})$ is likely due to LPD not exclusively targeting the frozen or wet ground for point dropout. LPD applies the point drop across all points, as shown in Fig. 5. This represents a limitation of LPD, necessitating the development of alternative methods to address this issue. + +Overall, the use of all components led to a significant improvement in performance under adverse weather conditions, with an increase of $+8.1\mathrm{mIoU}$ over the baseline at a reasonable cost of $-1.1\mathrm{mIoU}$ in clean data. Ultimately, our method achieved state-of-the-art results on SemanticSTF, with an increase in mIoU across all adverse weather conditions compared to the baseline. The combined use of DSJ, ASJ, Range Jittering, and LPD demonstrated high performance, indicating that each component synergistically contributes to enhancing the model's robustness. + +![](images/5e91bfdc18fcdfda504f2f19c6eda8a1fd7bf04a3f0d711c19bf0aa75b80437f.jpg) +Fig. 5: Qualitative results of LPD from validation set of SemanticKITTI. It can be observed that the point drop facilitated by the LPD occurs not just locally but extends across all points. While LPD effectively mimics the occlusions and sparse point distributions seen in real adverse weather, it does not replicate the effect of wet ground surfaces. + +Table 4: LiDAR segmentation results (mIoU) with the various architecture and datasets. All models were trained with SemanticKITTI. + +
MethodSemanticSTFSemanticKITTI-C
CENet14.249.3
SPVCNN28.152.5
Minkowski31.453.0
CENet+Ours22.0 (+7.8)53.2 (+3.9)
SPVCNN+Ours38.4 (+10.3)52.9 (+0.4)
Minkowski+Ours39.5 (+8.1)58.6 (+5.6)
+ +Experiments on Various Input Data Representation, Models and Datasets. We evaluated our method using different data representations, architectures, and datasets, specifically: (1) projection-based CENet and (2) voxel-based SPVCNN for backbones, and (3) the synthetic corruption dataset SemanticKITTI-C [11]. Tab. 4 shows our results. + +SemanticKITTI-C has a unique data distribution with more samples in classes, such as sidewalks and vegetation, than SemanticSTF. Our method consistently improves performance across various datasets and architectures. Specifically, for the SemanticSTF dataset, our method increases the mIoU of CENet by $+7.8$ , SPVCNN by $+10.3$ , and MinkowskiNet by $+8.1$ mIoU. Similarly, for the SemanticKITTI-C dataset, the improvements are $+3.9$ for CENet, $+0.4$ for SPVCNN, and $+5.6$ mIoU for MinkowskiNet. These consistent enhancements across different input data representations, models, and datasets highlight the robustness and generalizability of our approach. + +# 6 Conclusion + +Our study focused on analyzing LiDAR data distortions in adverse weather and enhancing LiDAR model robustness with a data-centric method. We identified that point drop and geometric distortion mainly affect model performance through toy experiments. We introduced novel data augmentation techniques, Selective Jittering, and Learnable Point Drop, leading to state-of-the-art performance on SemanticSTF. We showed that our methodology significantly contributes to robustness enhancement without complex data simulations. We expect that our research will contribute to improving the stability and reliability of LiDAR semantic segmentation. + +Acknowledgements. This work was supported by IITP grant funded by MSIT (No.2021-0-02068, Artificial Intelligence Innovation Hub and RS-2019-II190075, Artificial Intelligence Graduate School Program(KAIST)) and KEIT grant funded by MOTIE (No.2022-0-00680, No.2022-0-01045), NRF funded by the MSIP (NRF-2022R1A2C3011154, RS-2023-00219019, RS-2023-00240135) and Samsung Electronics Co., Ltd (IO230508-06190-01). + +# References + +1. Ando, A., Gidaris, S., Bursuc, A., Puy, G., Boulch, A., Marlet, R.: Rangevit: Towards vision transformers for 3d semantic segmentation in autonomous driving. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 5240-5250 (2023) +2. Behley, J., Garbade, M., Milioto, A., Quenzel, J., Behnke, S., Stachniss, C., Gall, J.: Semantickitti: A dataset for semantic scene understanding of lidar sequences. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 9297-9307 (2019) +3. Bijelic, M., Gruber, T., Ritter, W.: A benchmark for lidar sensors in fog: Is detection breaking down? In: 2018 IEEE intelligent vehicles symposium (IV). pp. 760-767. IEEE (2018) +4. Choe, J., Park, C., Rameau, F., Park, J., Kweon, I.S.: Pointmixer: Mlp-mixer for point cloud understanding. In: European Conference on Computer Vision. pp. 620-640. Springer (2022) +5. Choy, C., Gwak, J., Savarese, S.: 4d spatio-temporal convnets: Minkowski convolutional neural networks. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 3075-3084 (2019) +6. Fersch, T., Buhmann, A., Koelpin, A., Weigel, R.: The influence of rain on small aperture lidar sensors. In: 2016 German Microwave Conference (GeMiC). pp. 84-87. IEEE (2016) +7. Hahner, M., Sakaridis, C., Bijelic, M., Heide, F., Yu, F., Dai, D., Van Gool, L.: Lidar snowfall simulation for robust 3d object detection. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 16364-16374 (2022) +8. Hahner, M., Sakaridis, C., Dai, D., Van Gool, L.: Fog simulation on real lidar point clouds for 3d object detection in adverse weather. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 15283-15292 (2021) +9. Kilic, V., Hegde, D., Sindagi, V., Cooper, A.B., Foster, M.A., Patel, V.M.: Lidar light scattering augmentation (lisa): Physics-based simulation of adverse weather conditions for 3d object detection. arXiv preprint arXiv:2107.07004 (2021) +0. Kong, L., Liu, Y., Chen, R., Ma, Y., Zhu, X., Li, Y., Hou, Y., Qiao, Y., Liu, Z.: Rethinking range view representation for lidar segmentation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 228-240 (2023) +1. Kong, L., Liu, Y., Li, X., Chen, R., Zhang, W., Ren, J., Pan, L., Chen, K., Liu, Z.: Robo3d: Towards robust and reliable 3d perception against corruptions. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 19994-20006 (2023) +2. Kong, L., Ren, J., Pan, L., Liu, Z.: Lasermix for semi-supervised lidar semantic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 21705-21715 (2023) +3. Kong, L., Ren, J., Pan, L., Liu, Z.: Lasermix for semi-supervised lidar semantic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 21705-21715 (2023) +4. Lai, X., Chen, Y., Lu, F., Liu, J., Jia, J.: Spherical transformer for lidar-based 3d recognition. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 17545-17555 (2023) +5. Lee, S., Son, T., Kwak, S.: Fico: Learning fog-invariant features for foggy scene segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18911-18921 (2022) + +16. Li, Y., Duthon, P., Colomb, M., Ibanez-Guzman, J.: What happens for a tof lidar in fog? IEEE Transactions on Intelligent Transportation Systems 22(11), 6670-6681 (2020) +17. Li, Z., Wu, X., Wang, J., Guo, Y.: Weather-degraded image semantic segmentation with multi-task knowledge distillation. Image and Vision Computing 127, 104554 (2022) +18. Milioto, A., Vizzo, I., Behley, J., Stachniss, C.: Rangenet++: Fast and accurate lidar semantic segmentation. In: 2019 IEEE/RSJ international conference on intelligent robots and systems (IROS). pp. 4213-4220. IEEE (2019) +19. Mnih, V., Kavukcuoglu, K., Silver, D., Rusu, A.A., Veness, J., Bellemare, M.G., Graves, A., Riedmiller, M., Fidjeland, A.K., Ostrovski, G., et al.: Human-level control through deep reinforcement learning. nature 518(7540), 529-533 (2015) +20. Nekrasov, A., Schult, J., Litany, O., Leibe, B., Engelmann, F.: Mix3d: Out-of-context data augmentation for 3d scenes. In: 2021 international conference on 3d vision (3dv). pp. 116-125. IEEE (2021) +21. Puy, G., Boulch, A., Marlet, R.: Using a waffle iron for automotive point cloud semantic segmentation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 3379-3389 (2023) +22. Qi, C.R., Su, H., Mo, K., Guibas, L.J.: Pointnet: Deep learning on point sets for 3d classification and segmentation. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 652-660 (2017) +23. Qian, K., Zhu, S., Zhang, X., Li, L.E.: Robust multimodal vehicle detection in foggy weather using complementary lidar and radar signals. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 444-453 (2021) +24. Ryu, K., Hwang, S., Park, J.: Instant domain augmentation for lidar semantic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9350-9360 (2023) +25. Saltori, C., Galasso, F., Fiameni, G., Sebe, N., Ricci, E., Poiesi, F.: Cosmix: Compositional semantic mix for domain adaptation in 3d lidar segmentation. In: European Conference on Computer Vision. pp. 586-602. Springer (2022) +26. Shin, J., Park, H., Kim, T.: Characteristics of laser backscattering intensity to detect frozen and wet surfaces on roads. Journal of Sensors 2019(1), 8973248 (2019) +27. Smith, B.E., Gardner, A., Schneider, A., Flanner, M.: Modeling biases in laser-altimetry measurements caused by scattering of green light in snow. Remote sensing of environment 215, 398-410 (2018) +28. Thomas, H., Qi, C.R., Deschaud, J.E., Marcotegui, B., Goulette, F., Guibas, L.J.: Kpconv: Flexible and deformable convolution for point clouds. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 6411-6420 (2019) +29. Tolstikhin, I.O., Houlsby, N., Kolesnikov, A., Beyer, L., Zhai, X., Unterthiner, T., Yung, J., Steiner, A., Keysers, D., Uszkoreit, J., et al.: Mlp-mixer: An allmlp architecture for vision. Advances in neural information processing systems 34, 24261-24272 (2021) +30. Xiao, A., Huang, J., Guan, D., Cui, K., Lu, S., Shao, L.: Polarmix: A general data augmentation technique for lidar point clouds. Advances in Neural Information Processing Systems 35, 11035-11048 (2022) +31. Xiao, A., Huang, J., Guan, D., Cui, K., Lu, S., Shao, L.: Polarmix: A general data augmentation technique for lidar point clouds. Advances in Neural Information Processing Systems 35, 11035-11048 (2022) + +32. Xiao, A., Huang, J., Guan, D., Zhan, F., Lu, S.: Transfer learning from synthetic to real lidar point cloud for semantic segmentation. In: Proceedings of the AAAI conference on artificial intelligence. vol. 36, pp. 2795-2803 (2022) +33. Xiao, A., Huang, J., Xuan, W., Ren, R., Liu, K., Guan, D., El Saddik, A., Lu, S., Xing, E.P.: 3d semantic segmentation in the wild: Learning generalized models for adverse-condition point clouds. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9382-9392 (2023) +34. Yan, X., Zheng, C., Xue, Y., Li, Z., Cui, S., Dai, D.: Benchmarking the robustness of lidar semantic segmentation models. International Journal of Computer Vision pp. 1-24 (2024) +35. Yang, D., Liu, Z., Jiang, W., Yan, G., Gao, X., Shi, B., Liu, S., Cai, X.: Realistic rainy weather simulation for lidars in carla simulator. arXiv preprint arXiv:2312.12772 (2023) +36. Zhao, H., Jiang, L., Jia, J., Torr, P.H., Koltun, V.: Point transformer. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 16259-16268 (2021) +37. Zhu, X., Zhou, H., Wang, T., Hong, F., Ma, Y., Li, W., Li, H., Lin, D.: Cylindrical and asymmetrical 3d convolution networks for lidar segmentation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 9939-9948 (2021) \ No newline at end of file diff --git a/2024/Rethinking Data Augmentation for Robust LiDAR Semantic Segmentation in Adverse Weather/images.zip b/2024/Rethinking Data Augmentation for Robust LiDAR Semantic Segmentation in Adverse Weather/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..a03d453acf2adfe26078d36c52df97fd5c90d1b9 --- /dev/null +++ b/2024/Rethinking Data Augmentation for Robust LiDAR Semantic Segmentation in Adverse Weather/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:52f2aa1a87deb8d8df0712005373bc7977f5d928dbc52d8629809c06a91358ce +size 394673 diff --git a/2024/Rethinking Data Augmentation for Robust LiDAR Semantic Segmentation in Adverse Weather/layout.json b/2024/Rethinking Data Augmentation for Robust LiDAR Semantic Segmentation in Adverse Weather/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..378fefbd74d084f4fc7c1898b6c5b18eb16dfd56 --- /dev/null +++ b/2024/Rethinking Data Augmentation for Robust LiDAR Semantic Segmentation in Adverse Weather/layout.json @@ -0,0 +1,7507 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 158, + 111, + 455, + 163 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 158, + 111, + 455, + 163 + ], + "spans": [ + { + "bbox": [ + 158, + 111, + 455, + 163 + ], + "type": "text", + "content": "Rethinking Data Augmentation for Robust LiDAR Semantic Segmentation in Adverse Weather" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 181, + 186, + 430, + 199 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 181, + 186, + 430, + 199 + ], + "spans": [ + { + "bbox": [ + 181, + 186, + 430, + 199 + ], + "type": "text", + "content": "Junsung Park" + }, + { + "bbox": [ + 181, + 186, + 430, + 199 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 181, + 186, + 430, + 199 + ], + "type": "text", + "content": ", Kyungmin Kim" + }, + { + "bbox": [ + 181, + 186, + 430, + 199 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 181, + 186, + 430, + 199 + ], + "type": "text", + "content": ", and Hyunjung Shim" + }, + { + "bbox": [ + 181, + 186, + 430, + 199 + ], + "type": "inline_equation", + "content": "^{1\\star}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 140, + 207, + 472, + 242 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 140, + 207, + 472, + 242 + ], + "spans": [ + { + "bbox": [ + 140, + 207, + 472, + 242 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 140, + 207, + 472, + 242 + ], + "type": "text", + "content": " Korea Advanced Institute of Science & Technology (KAIST) \n" + }, + { + "bbox": [ + 140, + 207, + 472, + 242 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 140, + 207, + 472, + 242 + ], + "type": "text", + "content": " Yonsei University \njshackist@kaist.ac.kr, kyungmin.kim@yonsei.ac.kr, kateshim@kaist.ac.kr" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 160, + 264, + 452, + 539 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 264, + 452, + 539 + ], + "spans": [ + { + "bbox": [ + 160, + 264, + 452, + 539 + ], + "type": "text", + "content": "Abstract. Existing LiDAR semantic segmentation methods often struggle with performance declines in adverse weather conditions. Previous work has addressed this issue by simulating adverse weather or employing universal data augmentation during training. However, these methods lack a detailed analysis and understanding of how adverse weather negatively affects LiDAR semantic segmentation performance. Motivated by this issue, we identified key factors of adverse weather and conducted a toy experiment to pinpoint the main causes of performance degradation: (1) Geometric perturbation due to refraction caused by fog or droplets in the air and (2) Point drop due to energy absorption and occlusions. Based on these findings, we propose new strategic data augmentation techniques. First, we introduced a Selective Jittering (SJ) that jitters points in the random range of depth (or angle) to mimic geometric perturbation. Additionally, we developed a Learnable Point Drop (LPD) to learn vulnerable erase patterns with a Deep Q-Learning Network to approximate the point drop phenomenon from adverse weather conditions. Without precise weather simulation, these techniques strengthen the LiDAR semantic segmentation model by exposing it to vulnerable conditions identified by our data-centric analysis. Experimental results confirmed the suitability of the proposed data augmentation methods for enhancing robustness against adverse weather conditions. Our method achieves a notable " + }, + { + "bbox": [ + 160, + 264, + 452, + 539 + ], + "type": "inline_equation", + "content": "39.5\\mathrm{mIoU}" + }, + { + "bbox": [ + 160, + 264, + 452, + 539 + ], + "type": "text", + "content": " on the SemanticKITTI-to-SemanticSTF benchmark, improving the baseline by " + }, + { + "bbox": [ + 160, + 264, + 452, + 539 + ], + "type": "inline_equation", + "content": "8.1\\%" + }, + { + "bbox": [ + 160, + 264, + 452, + 539 + ], + "type": "text", + "content": " p and establishing a new state-of-the-art. Our code will be released at https://github.com/engineerJPark/LiDARWeather." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 160, + 548, + 453, + 571 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 548, + 453, + 571 + ], + "spans": [ + { + "bbox": [ + 160, + 548, + 453, + 571 + ], + "type": "text", + "content": "Keywords: LiDAR Semantic Segmentation " + }, + { + "bbox": [ + 160, + 548, + 453, + 571 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 548, + 453, + 571 + ], + "type": "text", + "content": " Robust Learning for Adverse Weather " + }, + { + "bbox": [ + 160, + 548, + 453, + 571 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 548, + 453, + 571 + ], + "type": "text", + "content": " Data Augmentation" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 589, + 230, + 601 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 589, + 230, + 601 + ], + "spans": [ + { + "bbox": [ + 132, + 589, + 230, + 601 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 611, + 481, + 648 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 611, + 481, + 648 + ], + "spans": [ + { + "bbox": [ + 130, + 611, + 481, + 648 + ], + "type": "text", + "content": "LiDAR semantic segmentation is a fundamental task of 3D scene understanding, particularly in safety-critical applications like autonomous driving. However, existing LiDAR segmentation models [1,5,14,21,28,37] commonly lack robustness," + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 653, + 303, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 653, + 303, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 653, + 303, + 665 + ], + "type": "text", + "content": "* Corresponding author: Hyunjung Shim" + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": "showing significant performance degradation under adverse weather conditions such as snow, fog, rain, or wet surfaces." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 143, + 481, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 143, + 481, + 251 + ], + "spans": [ + { + "bbox": [ + 130, + 143, + 481, + 251 + ], + "type": "text", + "content": "To address this issue, recent studies [11, 33, 34] have introduced corruption benchmarks for adverse weather and proposed effective techniques for robust LiDAR segmentation against corruption. These robust methods are primarily divided into task-agnostic and simulation-based approaches. The task-agnostic approach [11, 25, 33] employs general machine-learning techniques for robustness without explicitly addressing LiDAR corruption caused by adverse weather. The simulation-based approach [3,7-9,23,26,35] artificially synthesizes weather-specific data using physical equations for training. However, these efforts focus on detection tasks and consider only a single type of weather at a time." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 253, + 482, + 432 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 253, + 482, + 432 + ], + "spans": [ + { + "bbox": [ + 130, + 253, + 482, + 432 + ], + "type": "text", + "content": "While the simulation-based approach leverages the intrinsic properties of Li-DAR scans under target adverse weather conditions, simulating every weather type at all possible severities is impractical and often inaccurate. Instead, we adopt a data-centric perspective to analyze corrupted LiDAR data. For example, distortions caused by rain, snow, and fog often create similar patterns, as simulated in [9]. That is, all those adverse weather shows point drop patterns due to attenuation or occlusion by droplet [6,7,9]. While the simulation-based approach requires explicit modeling of the LiDAR for each weather condition, our data-centric approach potentially addresses complex distortions with a few unified patterns. Through existing research and our comprehensive analysis, we find that various adverse weather conditions create similar distortion patterns in Li-DAR data. Based on our analysis, we categorize these distortions into two types: (1) geometric perturbation and (2) point drop. Our toy experiments demonstrate that these distortions are closely related to segmentation performance degradation." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 435, + 482, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 435, + 482, + 615 + ], + "spans": [ + { + "bbox": [ + 130, + 435, + 482, + 615 + ], + "type": "text", + "content": "Rooted by the toy experiment, we introduce two novel and strategic data augmentation methods tailored to the LiDAR distortion caused by adverse weather. By incorporating these augmentations during training, we aim to bolster the model's robustness for each distortion type. For geometric perturbation, we introduce Selective Jittering (SJ). This method applies Additive Gaussian Noise (AGN) to alter the XYZ-coordinate and intensity within a selective local region. To handle a point drop pattern, we developed a Learnable Point Drop (LPD) that employs a Deep Q-Network (DQN) to strategically remove points. Our augmentation strategies are inspired by existing studies [9, 16, 27], which shows geometric perturbation typically involves small, random alterations to the original points. This insight led to our choice of jittering as a suitable technique for addressing the geometric perturbation. Also, our augmentation strategy for point drop is motivated by existing studies [7-9] which state that intensity and depth of LiDAR beams play a significant role. This prompts us to use a DQN to detect point drops and adapt to such point drop patterns." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "Finally, we apply the proposed augmentations when training the LiDAR segmentation model with a clean source dataset. Experimental results demonstrate that our method achieves new state-of-the-art performance on SemanticSTF, a real adverse weather dataset. Specifically, we achieved an impressive " + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "inline_equation", + "content": "+8.1\\%" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "type": "text", + "content": "J.Park et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 186 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 186 + ], + "type": "text", + "content": "mIoU gain over the baseline MinkowskiNet, nearly tripling the " + }, + { + "bbox": [ + 130, + 116, + 479, + 186 + ], + "type": "inline_equation", + "content": "+2.5\\%" + }, + { + "bbox": [ + 130, + 116, + 479, + 186 + ], + "type": "text", + "content": " p mIoU improvement from previous work [33]. Our method achieved remarkable performance gains, ranging from " + }, + { + "bbox": [ + 130, + 116, + 479, + 186 + ], + "type": "inline_equation", + "content": "+0.4" + }, + { + "bbox": [ + 130, + 116, + 479, + 186 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 130, + 116, + 479, + 186 + ], + "type": "inline_equation", + "content": "+10.3\\%" + }, + { + "bbox": [ + 130, + 116, + 479, + 186 + ], + "type": "text", + "content": " p mIoU, across various model architectures (Minkowski, CENet, SPVCNN) and corruption dataset (SemanticSTF and SemanticKITTI-C). These results demonstrate the generalizability of the proposed augmentation methods." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 146, + 188, + 404, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 188, + 404, + 200 + ], + "spans": [ + { + "bbox": [ + 146, + 188, + 404, + 200 + ], + "type": "text", + "content": "In summary, this paper presents several key contributions." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 137, + 207, + 479, + 326 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 138, + 207, + 479, + 242 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 207, + 479, + 242 + ], + "spans": [ + { + "bbox": [ + 138, + 207, + 479, + 242 + ], + "type": "text", + "content": "- We identify two prevalent distortion types in LiDAR data caused by adverse weather, leading to performance degradation, through a data-centric analysis." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 137, + 243, + 479, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 243, + 479, + 266 + ], + "spans": [ + { + "bbox": [ + 137, + 243, + 479, + 266 + ], + "type": "text", + "content": "- We introduce two novel data augmentations tailored to each identified distortion type." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 137, + 267, + 479, + 326 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 267, + 479, + 326 + ], + "spans": [ + { + "bbox": [ + 137, + 267, + 479, + 326 + ], + "type": "text", + "content": "- Our method sets new state-of-the-art benchmarks on the SemanticKITTI to SemanticSTF benchmark. Notably, our improvement was " + }, + { + "bbox": [ + 137, + 267, + 479, + 326 + ], + "type": "inline_equation", + "content": "+8.1\\%" + }, + { + "bbox": [ + 137, + 267, + 479, + 326 + ], + "type": "text", + "content": " p mIoU without relying on precise simulations of adverse weather in LiDAR point inputs. This represents a tripling of the improvement over the baseline, compared to what previous methods have achieved [33]." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 132, + 345, + 242, + 357 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 345, + 242, + 357 + ], + "spans": [ + { + "bbox": [ + 132, + 345, + 242, + 357 + ], + "type": "text", + "content": "2 Related Works" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 370, + 317, + 382 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 370, + 317, + 382 + ], + "spans": [ + { + "bbox": [ + 131, + 370, + 317, + 382 + ], + "type": "text", + "content": "2.1 LiDAR Semantic Segmentation" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 390, + 479, + 424 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 390, + 479, + 424 + ], + "spans": [ + { + "bbox": [ + 130, + 390, + 479, + 424 + ], + "type": "text", + "content": "Existing 3D LiDAR point cloud semantic segmentation methods can be categorized into three types based on the data representation: Point-based, Projection-based, and Voxel-based." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 426, + 480, + 520 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 426, + 480, + 520 + ], + "spans": [ + { + "bbox": [ + 130, + 426, + 480, + 520 + ], + "type": "text", + "content": "Point-based methods [22, 28, 36] utilize the 3D points directly as input. KPConv [28] initially clustering local points, aggregating these local features, and then feeding them into kernel point convolutions. The Point Transformer [36] utilizes a transformer architecture to compute query points in each local region, obtained through k-nearest neighbors (kNN). The Point-Mixer [4] tried to adapt the MLP-Mixer [29] for point cloud applications. They achieve high performance but suffer from high computational costs due to the utilization of large-scale raw LiDAR data." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 521, + 480, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 480, + 628 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 480, + 628 + ], + "type": "text", + "content": "Projection-based methods [1, 10, 18] project LiDAR points into a 2D image and performs the semantic segmentation using architectures successful in a 2D image. RangeViT [1] directly adopts a ViT model pre-trained on 2D images, demonstrating that pre-trained power in 2D images can be effective as prior knowledge in range images. RangeFormer [10] proposes \"RangeAug\" to maximize the utility of range images created by projecting into 2D, producing multiple range image data to overcome the low performance in range image models. Projection-based methods achieve fast inference speed but present a sub-optimal performance due to missing information after the projection." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 630, + 479, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 630, + 479, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 630, + 479, + 665 + ], + "type": "text", + "content": "Voxel-based methods [5,14,37] perform efficient computations by dividing 3D space into a voxel grid and aggregating points within the same voxel. Minkowski-iNet [5] voxelizes LiDAR points with a cubic grid and applies sparse convolution." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 142, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 142, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 142, + 91, + 447, + 102 + ], + "type": "text", + "content": "Rethinking Data Augmentation for Robust LiDAR Semantic Segmentation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 115, + 481, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 115, + 481, + 175 + ], + "spans": [ + { + "bbox": [ + 132, + 115, + 481, + 175 + ], + "type": "text", + "content": "Cylinder3D [37] proposes cylindrical partitions, reflecting LiDAR's characteristic that the density of points depends on the distance. SphereFormer [14] utilizes a radial window and transformer architecture to aggregate long-range information and improve performance. Voxel-based methods achieve a balance between reasonable inference time and commendable segmentation performance." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 193, + 301, + 205 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 193, + 301, + 205 + ], + "spans": [ + { + "bbox": [ + 132, + 193, + 301, + 205 + ], + "type": "text", + "content": "2.2 LiDAR Data Augmentation" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 213, + 481, + 369 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 213, + 481, + 369 + ], + "spans": [ + { + "bbox": [ + 132, + 213, + 481, + 369 + ], + "type": "text", + "content": "Inspired by 2D image augmentation, conventional LiDAR segmentation methods apply classic scaling, rotation, flipping, and translation to augment LiDAR data. Recently, several out-of-context augmentation techniques [12, 20, 30] that mix different LiDAR scans have been proposed. Mix3D [20] combines randomly selected two scans. Considering the sweeping mechanism of the LiDAR sensor, PolarMix [30] cuts LiDAR scans along the azimuth axis, then exchanges point cloud sectors and applies instance-level rotate-pasting. To reflect the spatial prior of the LiDAR point cloud, LaserMix [12] partitions LiDAR scans based on the laser beams and blends partitions from different LiDAR scans. Recent study [24] introduces a fast LiDAR domain augmentation module to address sensor-bias problems. To the best of our knowledge, our approach is the first augmentation method specifically designed to address data corruption under adverse weather conditions." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 385, + 378, + 398 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 385, + 378, + 398 + ], + "spans": [ + { + "bbox": [ + 132, + 385, + 378, + 398 + ], + "type": "text", + "content": "2.3 LiDAR under Adverse Weather Conditions" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 406, + 490, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 406, + 490, + 574 + ], + "spans": [ + { + "bbox": [ + 132, + 406, + 490, + 574 + ], + "type": "text", + "content": "Robustness under harsh conditions is crucial in safety-critical applications. Adverse weather substantially degrade performance in real-world outdoor autonomous driving. Thus, there are several attempts to develop weather-robust models in fields such as 2D segmentation [15, 17], 3D detection [7, 8, 11], and 3D segmentation [11, 33]. Simulation-based approaches [7, 8] artificially synthesize data for single weather conditions through physical modeling and utilize it for training. We differ from these in that we do not model specific weather conditions explicitly. Recently proposed task-agnostic approaches [11, 25, 33] consider multiple weather conditions at once. However, they use general machine-learning techniques (such as teacher-student framework and feature prototype) to achieve robustness rather than specifically tackle LiDAR corruption caused by adverse weather. We differ from these in that we propose augmentation methods specifically tailored for adverse weather conditions based on the analysis of performance degradation in LiDAR data." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 591, + 345, + 606 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 591, + 345, + 606 + ], + "spans": [ + { + "bbox": [ + 132, + 591, + 345, + 606 + ], + "type": "text", + "content": "3 Finding Distortions to Augment" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 616, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 616, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 616, + 481, + 666 + ], + "type": "text", + "content": "In this section, we aim to discuss the patterns of distortion that different adverse weather conditions impose on LiDAR data. Although adverse weather conditions are distinct in reality, studies have shown that their effects on LiDAR data often result in similar impacts. For instance, the distortions caused by \"rain\", \"snow\"," + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "type": "text", + "content": "J.Park et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "text", + "content": "and \"fog\" tend to produce similar point-missing patterns due to attenuation in the data, as demonstrated in [9]. Therefore, this section will focus on identifying the common distortion patterns caused by adverse weather through existing studies. Overall, existing studies describe the effect of adverse weather as four different types of distortion: (1) Point Drop due to energy absorption, (2) Occlusions caused by droplets of rain or snow and fog, (3) Geometric perturbation, and (4) Intensity distortion due to energy absorption." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 217, + 370, + 228 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 217, + 370, + 228 + ], + "spans": [ + { + "bbox": [ + 131, + 217, + 370, + 228 + ], + "type": "text", + "content": "3.1 Distortion Factors from Adverse Weather" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 235, + 482, + 522 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 130, + 235, + 482, + 305 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 235, + 482, + 305 + ], + "spans": [ + { + "bbox": [ + 130, + 235, + 482, + 305 + ], + "type": "text", + "content": "(D1) Point Drop. Several studies have explored how adverse weather conditions contribute to point drops in LiDAR data. Kilic et al. [9], Fersch et al. [6] and Shin et al. [26] describe point drops resulting from beam attenuation and beam missing due to droplets, fog, and frozen or wet ground. These studies collectively suggest that adverse weather conditions typically lead to point drops in LiDAR data." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 307, + 482, + 377 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 307, + 482, + 377 + ], + "spans": [ + { + "bbox": [ + 130, + 307, + 482, + 377 + ], + "type": "text", + "content": "(D2) Occlusions. Several studies have addressed occlusions caused by adverse weather conditions. Hahner et al. [7], Kilic et al. [9], Kong et al. [11] and Yan et al. [34] consider scenarios where beams colliding with snow collect signals at much shorter distances than the intended objects of collision. Upon reviewing these studies, we have concluded that adverse weather consistently leads to occlusions." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 379, + 482, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 379, + 482, + 450 + ], + "spans": [ + { + "bbox": [ + 130, + 379, + 482, + 450 + ], + "type": "text", + "content": "(D3) Geometric Perturbation. Some studies focus on geometric perturbation caused by adverse weather conditions. Kilic et al. [9], Li et al. [16] and Smith et al. [27] demonstrated geometric perturbation in adverse weather, such as fog, snow, and rain, by incorporating random noise into the coordinates. Through these studies, we have come to conclude that adverse weather universally causes geometric perturbation." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 450, + 482, + 522 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 450, + 482, + 522 + ], + "spans": [ + { + "bbox": [ + 130, + 450, + 482, + 522 + ], + "type": "text", + "content": "(D4) Intensity Distortion. Numerous studies have focused on intensity distortion caused by adverse weather conditions. Bijelic et al. [3], Shin et al. [26], Fersch et al. [6], Kong et al. [11], and Yan et al. [34] have collectively shown that adverse weather conditions like fog, wetness, and rain lead to a reduction in Li-DAR beam intensity, influencing the generation of synthetic data. Through these studies, we conclude that adverse weather commonly causes intensity distortion." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 131, + 539, + 243, + 552 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 539, + 243, + 552 + ], + "spans": [ + { + "bbox": [ + 131, + 539, + 243, + 552 + ], + "type": "text", + "content": "3.2 Toy Experiment" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "text", + "content": "Based on previous research findings mentioned in Section 3.1, the types of distortions in LiDAR point clouds caused by adverse weather conditions converge into a set of common distortions. Therefore, from a data-centric perspective, the issues that we need to consider can be summarized as follows: (D1) Point Drop, (D2) Occlusions, (D3) Geometric Perturbation, and (D4) Intensity Distortion. Here, we aim to identify which distortion types negatively impact performance. To achieve this, we have generated four distortion types of toy synthetic data through the SemanticKITTI validation set." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 142, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 142, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 142, + 91, + 447, + 102 + ], + "type": "text", + "content": "Rethinking Data Augmentation for Robust LiDAR Semantic Segmentation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 144, + 146, + 463, + 251 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 481, + 137 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 481, + 137 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 481, + 137 + ], + "type": "text", + "content": "Table 1: Results of toy experiments from validation set of SemanticKITTI [2]. Soft and Hard indicate the severity of the distortions." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 144, + 146, + 463, + 251 + ], + "lines": [ + { + "bbox": [ + 144, + 146, + 463, + 251 + ], + "spans": [ + { + "bbox": [ + 144, + 146, + 463, + 251 + ], + "type": "table", + "html": "
Distortioncarbileemtcletruckoth-v.pers.biclstmt-clstroadparki.sidew.oth-g.build.fenceveget.trunkterra.poletraf.mIoU
Clean96.822.266.989.165.267.084.70.093.850.681.40.191.163.088.167.774.563.847.763.9
D1: Soft96.317.963.487.163.764.381.40.092.747.579.70.190.162.387.364.274.161.141.661.8
D1: Hard86.10.212.721.040.010.124.40.04.210.211.10.071.747.877.034.627.630.417.627.7
D2: Soft94.816.351.766.353.959.252.10.092.544.979.20.189.761.487.363.274.359.240.257.2
D2: Hard81.30.55.413.135.57.60.70.02.68.510.80.071.145.776.332.927.326.716.124.3
D3: Soft96.215.456.758.651.351.878.10.066.333.946.10.086.161.684.462.056.861.544.053.2
D3: Hard93.89.738.319.431.435.255.20.09.912.312.20.048.940.871.455.531.258.740.835.0
D4: Soft96.321.761.189.461.868.983.20.092.944.479.10.190.456.786.968.271.163.948.362.3
D4: Hard95.017.250.481.457.064.979.70.090.738.864.70.988.445.383.562.850.560.147.956.8
", + "image_path": "0dd8f63c447380b92bd5e82c3d7c09edc96954dee2eb3cf57e9e164b1ae8535b.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 273, + 481, + 475 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 138, + 273, + 481, + 320 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 273, + 481, + 320 + ], + "spans": [ + { + "bbox": [ + 138, + 273, + 481, + 320 + ], + "type": "text", + "content": "- (D1) Point Drop: It considers the scenario where each LiDAR point disappears randomly and independently due to severe adverse weather. We removed individual points randomly to synthesize these data. We set the drop ratios at 0.5 and 0.9." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 321, + 481, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 321, + 481, + 380 + ], + "spans": [ + { + "bbox": [ + 138, + 321, + 481, + 380 + ], + "type": "text", + "content": "- (D2) Occlusions: The method assumes that occlusions occur predominantly in front of objects, mainly due to distortion from fog, snow, and rain. We synthesized this data by randomly selecting points and altering their depth to one-tenth of the original depth. The selecting ratios were also determined to be 0.5 and 0.9." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 380, + 481, + 428 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 380, + 481, + 428 + ], + "spans": [ + { + "bbox": [ + 138, + 380, + 481, + 428 + ], + "type": "text", + "content": "- (D3) Geometric Perturbation: This aspect assumes the distortion of point coordinates resulting from adverse weather conditions. We synthesized data by adding Gaussian noise to the coordinates of all points. The Gaussian noise levels were established at 0.05 and 0.25." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 429, + 481, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 429, + 481, + 475 + ], + "spans": [ + { + "bbox": [ + 138, + 429, + 481, + 475 + ], + "type": "text", + "content": "- (D4) Intensity Distortion: This distortion type assumes intensity attenuation caused by fog, raindrops, and snow particles. We synthesized this data by subtracting Gaussian noise from the intensity of all points. The Gaussian noise levels were established at 0.05 and 0.25." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 130, + 486, + 481, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 486, + 481, + 521 + ], + "spans": [ + { + "bbox": [ + 130, + 486, + 481, + 521 + ], + "type": "text", + "content": "We selected MinkowskiNet [5] for our toy experiment. This choice was based on existing research [11, 34], which indicates that MinkowskiNet is a standard and robust model." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 521, + 481, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 481, + 582 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 481, + 582 + ], + "type": "text", + "content": "In Tab. 1, it's evident that with increasing distortion in D1, D2, and D3, performance falls below half of the baseline. Contrarily, D4 maintains performance levels despite significant distortion. This performance degradation likely stems from changes in the local geometric structure where model computations occur. For this reason, D4 did not significantly impact performance." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 582, + 481, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 481, + 629 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 481, + 629 + ], + "type": "text", + "content": "Importantly, D2's notable performance drop is argued to result from a point drop akin to D1. It is due to occlusions altering the point's local geometric structure, an effect similar to point drop. This is evidenced in Fig. 1, where incorrect predictions from D2 show patterns resembling those in D1." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 630, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 630, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 630, + 481, + 665 + ], + "type": "text", + "content": "In summary, our observations indicate that geometric perturbation and point drop are the most impactful distortions from adverse weather. Data augmentation that replicates these phenomena could likely boost the model's robustness" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "type": "text", + "content": "J.Park et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 119, + 482, + 203 + ], + "blocks": [ + { + "bbox": [ + 133, + 119, + 482, + 203 + ], + "lines": [ + { + "bbox": [ + 133, + 119, + 482, + 203 + ], + "spans": [ + { + "bbox": [ + 133, + 119, + 482, + 203 + ], + "type": "image", + "image_path": "82090c9f50fa12ab3789eb06cbd45733571c834c844de401e608b4c407606628.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 220, + 482, + 277 + ], + "lines": [ + { + "bbox": [ + 130, + 220, + 482, + 277 + ], + "spans": [ + { + "bbox": [ + 130, + 220, + 482, + 277 + ], + "type": "text", + "content": "Fig. 1: Visualization for the results of toy experiment. In all figures, green (or red) points indicate correct (or incorrect) predictions. Point drop, occlusion, and geometric perturbation are the main distortions that lead to performance degradation. Also, since the misprediction areas of point drop and occlusion largely overlap, it's reasonable to consider them as the same type of distortion." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 300, + 482, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 300, + 482, + 361 + ], + "spans": [ + { + "bbox": [ + 130, + 300, + 482, + 361 + ], + "type": "text", + "content": "against such conditions without explicit weather simulations. Thus, robust Li-DAR semantic segmentation models against adverse weather involve two key steps: (1) Training on specific adverse point drops detrimental to performance, and (2) Training with minor adjustments in point coordinates." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 380, + 208, + 393 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 380, + 208, + 393 + ], + "spans": [ + { + "bbox": [ + 132, + 380, + 208, + 393 + ], + "type": "text", + "content": "4 Methods" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 407, + 482, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 407, + 482, + 468 + ], + "spans": [ + { + "bbox": [ + 130, + 407, + 482, + 468 + ], + "type": "text", + "content": "In this section, we propose solutions for the two challenges identified in Section 3.2. Our objective is to enhance the robustness of the model by addressing the common distortions exhibited by various adverse weather conditions, through data augmentation. This approach aims to achieve robustness without the necessity of simulating each adverse weather scenario explicitly." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 468, + 482, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 468, + 482, + 552 + ], + "spans": [ + { + "bbox": [ + 130, + 468, + 482, + 552 + ], + "type": "text", + "content": "As observed earlier, distortions resulting from adverse weather include point drop, occlusion, geometric perturbation, and intensity distortion. Of these, the main distortions identified as having the most detrimental impact on performance were (1) geometric perturbation caused by minor variations in point coordinates and (2) point drop caused by beam missing or occlusion. To address these challenges, we propose two techniques: Selective Jittering and Learnable Point Drop." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 571, + 253, + 583 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 571, + 253, + 583 + ], + "spans": [ + { + "bbox": [ + 132, + 571, + 253, + 583 + ], + "type": "text", + "content": "4.1 Selective Jittering" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 593, + 481, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 593, + 481, + 629 + ], + "spans": [ + { + "bbox": [ + 130, + 593, + 481, + 629 + ], + "type": "text", + "content": "Selective Jittering is devised to address the first main distortion, geometric perturbation. Selective Jittering consists of two types: Depth-Selective Jittering (DSJ) and Angle-Selective Jittering (ASJ)." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "type": "text", + "content": "Depth-Selective Jittering (DSJ). DSJ employs a single scan to add Gaussian noise to the XYZ coordinates and the intensity value of points below a randomly selected depth range." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 142, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 142, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 142, + 91, + 448, + 102 + ], + "type": "text", + "content": "Rethinking Data Augmentation for Robust LiDAR Semantic Segmentation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 187 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 187 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 187 + ], + "type": "text", + "content": "Angle-Selective Jittering (ASJ). It applies Gaussian noise to a randomly selected range of angles. By performing jittering across different ranges of depths or angles for each point, this method reflects the augmentation that some beams are affected by geometric perturbation in a non-uniform manner in adverse weather. This concept realistically represents the characteristics of transparent droplets like rain and snow [6]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 188, + 479, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 188, + 479, + 236 + ], + "spans": [ + { + "bbox": [ + 130, + 188, + 479, + 236 + ], + "type": "text", + "content": "A significant aspect of DSJ and ASJ is that they use only a single frame. DSJ and ASJ are efficient in data augmentation without additional LiDAR frames and can perform reasonable augmentation even when the given LiDAR data is not sequentially captured." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 236, + 479, + 283 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 236, + 479, + 283 + ], + "spans": [ + { + "bbox": [ + 130, + 236, + 479, + 283 + ], + "type": "text", + "content": "Range Jittering (RJ). Range Jittering has been proposed to simulate range distortion caused by droplets and fog, as mentioned in [9]. Unlike DSJ and ASJ, Range Jittering applies jittering only in the range direction of points. This method is used in place of using the original points in DSJ and ASJ." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 300, + 271, + 312 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 300, + 271, + 312 + ], + "spans": [ + { + "bbox": [ + 132, + 300, + 271, + 312 + ], + "type": "text", + "content": "4.2 Learnable Point Drop" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 319, + 480, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 319, + 480, + 474 + ], + "spans": [ + { + "bbox": [ + 130, + 319, + 480, + 474 + ], + "type": "text", + "content": "Learnable Point Drop (LPD) was devised to address the second main distortion, point drop. LPD is designed to artificially create point drop due to occlusion, such as dense fog. LPD employs a Deep Q-Learning Network (DQN) [19] for identifying the drop ratio and drop region that lead to adverse effects on the model. The reward of the DQN is designed to identify point drops that increase the training loss and uncertainty of the LiDAR semantic segmentation model. Through LPD, the LiDAR semantic segmentation model is exposed to point drop scenarios caused by adverse weather conditions. Consequently, it learns to make accurate predictions even in the absence of points that are critical for segmentation in clean data environments. Since LPD is employed merely as a concept for data augmentation and exists as a separate module, it necessitates no alterations to the existing model's training scheme except for limiting the gradient norm to ensure the stability of DQN learning." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 474, + 479, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 474, + 479, + 518 + ], + "spans": [ + { + "bbox": [ + 130, + 474, + 479, + 518 + ], + "type": "text", + "content": "LPD module defines its current state by summing the loss " + }, + { + "bbox": [ + 130, + 474, + 479, + 518 + ], + "type": "inline_equation", + "content": "L_{aug}" + }, + { + "bbox": [ + 130, + 474, + 479, + 518 + ], + "type": "text", + "content": ", calculated from augmented data obtained through SJ, and the entropy " + }, + { + "bbox": [ + 130, + 474, + 479, + 518 + ], + "type": "inline_equation", + "content": "H_{aug}" + }, + { + "bbox": [ + 130, + 474, + 479, + 518 + ], + "type": "text", + "content": ", derived from logits. The loss " + }, + { + "bbox": [ + 130, + 474, + 479, + 518 + ], + "type": "inline_equation", + "content": "L_{aug}" + }, + { + "bbox": [ + 130, + 474, + 479, + 518 + ], + "type": "text", + "content": " is calculated using the original loss function employed by the model in use. The calculation of entropy is as follows:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 236, + 525, + 377, + 552 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 236, + 525, + 377, + 552 + ], + "spans": [ + { + "bbox": [ + 236, + 525, + 377, + 552 + ], + "type": "interline_equation", + "content": "H _ {a u g} (x) = - \\frac {1}{N} \\sum_ {i = 1} ^ {n} P (x _ {i}) \\log P (x _ {i})", + "image_path": "e774633760070d193bc90db0054e48ef6abcf1180215ae25f32510bdb1c77808.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 558, + 480, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 558, + 480, + 640 + ], + "spans": [ + { + "bbox": [ + 130, + 558, + 480, + 640 + ], + "type": "text", + "content": "In this process, " + }, + { + "bbox": [ + 130, + 558, + 480, + 640 + ], + "type": "inline_equation", + "content": "x_{i}" + }, + { + "bbox": [ + 130, + 558, + 480, + 640 + ], + "type": "text", + "content": " represents each point, and " + }, + { + "bbox": [ + 130, + 558, + 480, + 640 + ], + "type": "inline_equation", + "content": "P(\\cdot)" + }, + { + "bbox": [ + 130, + 558, + 480, + 640 + ], + "type": "text", + "content": " is softmax to the logits. " + }, + { + "bbox": [ + 130, + 558, + 480, + 640 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 558, + 480, + 640 + ], + "type": "text", + "content": " denotes the number of points. LPD predicts the indices of points to drop, taking as input the sum of " + }, + { + "bbox": [ + 130, + 558, + 480, + 640 + ], + "type": "inline_equation", + "content": "L_{\\text{aug}}" + }, + { + "bbox": [ + 130, + 558, + 480, + 640 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 558, + 480, + 640 + ], + "type": "inline_equation", + "content": "H_{\\text{aug}}" + }, + { + "bbox": [ + 130, + 558, + 480, + 640 + ], + "type": "text", + "content": " added to the input point tensor. This process enables simultaneous learning of the drop ratio and drop region. Subsequently, the difference between the sum of the loss " + }, + { + "bbox": [ + 130, + 558, + 480, + 640 + ], + "type": "inline_equation", + "content": "L_{\\text{LPD}}" + }, + { + "bbox": [ + 130, + 558, + 480, + 640 + ], + "type": "text", + "content": " and entropy " + }, + { + "bbox": [ + 130, + 558, + 480, + 640 + ], + "type": "inline_equation", + "content": "H_{\\text{LPD}}" + }, + { + "bbox": [ + 130, + 558, + 480, + 640 + ], + "type": "text", + "content": ", obtained from the dropped points by LPD, and the sum of " + }, + { + "bbox": [ + 130, + 558, + 480, + 640 + ], + "type": "inline_equation", + "content": "L_{\\text{aug}}" + }, + { + "bbox": [ + 130, + 558, + 480, + 640 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 558, + 480, + 640 + ], + "type": "inline_equation", + "content": "H_{\\text{aug}}" + }, + { + "bbox": [ + 130, + 558, + 480, + 640 + ], + "type": "text", + "content": ", is defined as the reward." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 641, + 479, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 479, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 479, + 665 + ], + "type": "text", + "content": "In Comparison with random drop, random drop uniformly reduces points across all depths. Consequently, this approach is not suitable for simulating" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "type": "text", + "content": "J.Park et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 140, + 118, + 482, + 208 + ], + "blocks": [ + { + "bbox": [ + 140, + 118, + 482, + 208 + ], + "lines": [ + { + "bbox": [ + 140, + 118, + 482, + 208 + ], + "spans": [ + { + "bbox": [ + 140, + 118, + 482, + 208 + ], + "type": "image", + "image_path": "b70ceab1da88d04eff57b3b5ea4d42faf5f504c9b8a890620223ab24a564a38b.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 239, + 482, + 296 + ], + "lines": [ + { + "bbox": [ + 131, + 239, + 482, + 296 + ], + "spans": [ + { + "bbox": [ + 131, + 239, + 482, + 296 + ], + "type": "text", + "content": "Fig. 2: Description of Learnable Point Drop (LPD). (1) The LPD utilizes the loss " + }, + { + "bbox": [ + 131, + 239, + 482, + 296 + ], + "type": "inline_equation", + "content": "L_{\\text{aug}}" + }, + { + "bbox": [ + 131, + 239, + 482, + 296 + ], + "type": "text", + "content": " and entropy " + }, + { + "bbox": [ + 131, + 239, + 482, + 296 + ], + "type": "inline_equation", + "content": "H_{\\text{aug}}" + }, + { + "bbox": [ + 131, + 239, + 482, + 296 + ], + "type": "text", + "content": " derived from LiDAR segmentation model as its current state. (2) Upon receiving a point input, LPD performs a point drop, then recalculates the loss " + }, + { + "bbox": [ + 131, + 239, + 482, + 296 + ], + "type": "inline_equation", + "content": "L_{\\text{LPD}}" + }, + { + "bbox": [ + 131, + 239, + 482, + 296 + ], + "type": "text", + "content": " and entropy " + }, + { + "bbox": [ + 131, + 239, + 482, + 296 + ], + "type": "inline_equation", + "content": "H_{\\text{LPD}}" + }, + { + "bbox": [ + 131, + 239, + 482, + 296 + ], + "type": "text", + "content": ". (3) The difference between these new values and the previous ones is used to compute the reward." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 319, + 482, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 319, + 482, + 380 + ], + "spans": [ + { + "bbox": [ + 130, + 319, + 482, + 380 + ], + "type": "text", + "content": "point drops caused by fog, as it lacks the depth-specific characteristics of fog-related distortions. Furthermore, randomly generated point drop scans may not achieve the severity of adverse weather conditions necessary to drop critical points that could deceive the model. A detailed comparison will be conducted in the supplementary material." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 398, + 242, + 410 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 398, + 242, + 410 + ], + "spans": [ + { + "bbox": [ + 132, + 398, + 242, + 410 + ], + "type": "text", + "content": "4.3 Overall Pipeline" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 418, + 482, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 418, + 482, + 552 + ], + "spans": [ + { + "bbox": [ + 130, + 418, + 482, + 552 + ], + "type": "text", + "content": "The overall training procedure begins with data augmentation through SJ. The augmented points are input into the LiDAR semantic segmentation model to compute loss. The model then calculates logits for each point, applies softmax, and subsequently computes the entropy. The current state of the LPD module is set as the sum of the loss and entropy. The dropped points are then input into the LiDAR semantic segmentation model to calculate the corresponding loss and entropy. The difference between the sum of loss/entropy obtained from the augmented points and that obtained from the points determined by LPD is set as the reward. Through this process, the LPD module enables the LiDAR semantic segmentation model to effectively learn points distorted due to adverse weather. Overall process is detailed in Fig. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 570, + 230, + 583 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 570, + 230, + 583 + ], + "spans": [ + { + "bbox": [ + 132, + 570, + 230, + 583 + ], + "type": "text", + "content": "5 Experiments" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 597, + 261, + 609 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 597, + 261, + 609 + ], + "spans": [ + { + "bbox": [ + 132, + 597, + 261, + 609 + ], + "type": "text", + "content": "5.1 Experimental Setup" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "Dataset and Evaluation Metrics. To assess robustness, we trained our model on 19 classes from SemanticKITTI and then validated its performance on validation set of SemanticSTF as same as [33]. The experiment process involves training on clean data followed by evaluating robustness in adverse conditions," + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 142, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 142, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 142, + 91, + 447, + 102 + ], + "type": "text", + "content": "Rethinking Data Augmentation for Robust LiDAR Semantic Segmentation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 136, + 116, + 481, + 229 + ], + "blocks": [ + { + "bbox": [ + 136, + 116, + 481, + 229 + ], + "lines": [ + { + "bbox": [ + 136, + 116, + 481, + 229 + ], + "spans": [ + { + "bbox": [ + 136, + 116, + 481, + 229 + ], + "type": "image", + "image_path": "3d4e4751c2405ef68b580baaa1320e0f2611933d089a0a7374db71eeee391e64.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 248, + 482, + 293 + ], + "lines": [ + { + "bbox": [ + 130, + 248, + 482, + 293 + ], + "spans": [ + { + "bbox": [ + 130, + 248, + 482, + 293 + ], + "type": "text", + "content": "Fig. 3: The overall training process. (1) Initially, SJ is applied. (2) Following this, we calculate the loss, which leads to the update of the LPD state. (3) This process is followed by utilizing LPD to generate dropped points. (4) Subsequent to recalculating the loss, the reward is accordingly updated." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 316, + 479, + 388 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 316, + 479, + 388 + ], + "spans": [ + { + "bbox": [ + 130, + 316, + 479, + 388 + ], + "type": "text", + "content": "akin to existing research [11, 33, 34]. SemanticSTF is a dataset gathered under real-world adverse conditions such as rain, fog, and snow. Following the approach in [33], invalid data in SemanticSTF are mapped to an \"ignore label\". We utilized MinkowskiNet [5] for validation purposes due to its well-known robustness [11, 34]. Our evaluation metrics include the Intersection over Union (IoU) for each class and the mean IoU (mIoU) across all classes." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 388, + 480, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 388, + 480, + 472 + ], + "spans": [ + { + "bbox": [ + 130, + 388, + 480, + 472 + ], + "type": "text", + "content": "Implementation Details. MinkowskiNet-18/32width served as our baseline model. The learning rate was set at 0.24, with a weight decay of 0.0001. The mean and standard deviation of Gaussian noise used in SJ are 0 and 0.01 each. In cases of Learnable Point Drop, the norm of the gradient was limited to 100 for both the DQN and Segmentation model. All experiments were conducted using four A6000 GPUs for 15 epochs in batch size 2. The duration of all experiments ranged between 3 to 5 hours." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 490, + 228, + 501 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 490, + 228, + 501 + ], + "spans": [ + { + "bbox": [ + 131, + 490, + 228, + 501 + ], + "type": "text", + "content": "5.2 Main Results" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 510, + 482, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 482, + 641 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 482, + 641 + ], + "type": "text", + "content": "SemanticKITTI to SemanticSTF. According to Tab. 2 (a), our proposed method showed a significant improvement in SemanticSTF, with a " + }, + { + "bbox": [ + 130, + 510, + 482, + 641 + ], + "type": "inline_equation", + "content": "+8.1" + }, + { + "bbox": [ + 130, + 510, + 482, + 641 + ], + "type": "text", + "content": " mIoU increase over the baseline and " + }, + { + "bbox": [ + 130, + 510, + 482, + 641 + ], + "type": "inline_equation", + "content": "+5.6" + }, + { + "bbox": [ + 130, + 510, + 482, + 641 + ], + "type": "text", + "content": " mIoU over the recent competitor, PointDR. Our method demonstrated universally superior performance across all weather conditions, validating the validity of our choice in data distortion type and the effectiveness of our method. Our method substantially improved the performance of categories like other vehicle, motorcyclist, sidewalk, pole, and traffic sign, which had lower baseline performance, increasing their class IoU by about " + }, + { + "bbox": [ + 130, + 510, + 482, + 641 + ], + "type": "inline_equation", + "content": "+5" + }, + { + "bbox": [ + 130, + 510, + 482, + 641 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 130, + 510, + 482, + 641 + ], + "type": "inline_equation", + "content": "+10" + }, + { + "bbox": [ + 130, + 510, + 482, + 641 + ], + "type": "text", + "content": ". Additionally, it significantly enhanced the mIoU for cars and persons by " + }, + { + "bbox": [ + 130, + 510, + 482, + 641 + ], + "type": "inline_equation", + "content": "+19.0" + }, + { + "bbox": [ + 130, + 510, + 482, + 641 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 510, + 482, + 641 + ], + "type": "inline_equation", + "content": "+9.6" + }, + { + "bbox": [ + 130, + 510, + 482, + 641 + ], + "type": "text", + "content": ", respectively, compared to the baseline, which is crucial for safety in driving environments. More details are available in the supplementary material." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "content": "Also, in Tab. 2 (a), examining the performance improvement per weather condition, we note that the most significant increase in mIoU was in rain, with" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "text", + "content": "J.Park et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 138, + 200, + 301, + 297 + ], + "blocks": [ + { + "bbox": [ + 138, + 200, + 301, + 297 + ], + "lines": [ + { + "bbox": [ + 138, + 200, + 301, + 297 + ], + "spans": [ + { + "bbox": [ + 138, + 200, + 301, + 297 + ], + "type": "table", + "html": "
(a) SemanticKITTI→SemanticSTF
MethodsD-fogL-fogRainSnowmIoU
Oracle51.954.657.953.754.7
Baseline30.730.129.725.331.4
LaserMix [13]23.215.59.37.814.7
PolarMix [31]21.314.916.59.315.3
PointDR* [33]37.333.535.526.933.9
Baseline+SJ+LPD36.037.537.633.139.5
Increments to baseline+5.3+7.4+7.9+7.8+8.1
", + "image_path": "5d8c1fb8e25705020adce53a5c9b6337044d6044ec87ea05db3cac6f134323b6.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 307, + 190, + 485, + 308 + ], + "blocks": [ + { + "bbox": [ + 130, + 114, + 482, + 180 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 482, + 180 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 482, + 180 + ], + "type": "text", + "content": "Table 2: LiDAR segmentation results (mIoU) on the SemanticSTF validation set of models trained with (a) SemanticKITTI and (b) SynLiDAR [32]. D-fog and L-fog denote dense fog and light fog weather conditions in all experiments. * denotes the reproduced result with the same segmentation backbone. The symbol " + }, + { + "bbox": [ + 130, + 114, + 482, + 180 + ], + "type": "inline_equation", + "content": "\\ddagger" + }, + { + "bbox": [ + 130, + 114, + 482, + 180 + ], + "type": "text", + "content": " indicates that the validation for model selection was performed on sequence 0 of SynLiDAR, rather than SemanticSTF. Best score is in bold and second best is underlined." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 307, + 190, + 485, + 308 + ], + "lines": [ + { + "bbox": [ + 307, + 190, + 485, + 308 + ], + "spans": [ + { + "bbox": [ + 307, + 190, + 485, + 308 + ], + "type": "table", + "html": "
(b) SynLiDAR→SemanticSTF
MethodsD-fogL-fogRainSnowmIoU
Oracle51.954.657.953.754.7
Baseline15.2415.9716.8312.7615.45
LaserMix [13]15.3217.9518.5513.816.85
PolarMix [31]16.4718.6919.6315.9818.09
PointDR* [33]19.0920.2825.2918.9819.78
PointDR*‡21.4120.9425.4819.3120.47
Baseline+SJ+LPD19.0820.6521.9717.2720.08
Increments to baseline+3.8+4.7+5.1+4.5+4.6
Baseline+SJ+LPD‡18.9921.2223.1417.2820.51
Increments to baseline+3.7+5.3+6.3+4.5+5.1
", + "image_path": "ac50736120ce4d7230c3111b5aba569db3ec29471724547ee42da77e609cb316.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 345, + 482, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 345, + 482, + 453 + ], + "spans": [ + { + "bbox": [ + 130, + 345, + 482, + 453 + ], + "type": "text", + "content": "+7.9 mIoU. This result is indicative of our SJ augmentation effectively reflecting geometric perturbations caused by raindrops. This aspect will be further explored in the ablation study. Additionally, in snow, our method showed a remarkable increase of +7.8 mIoU, and in dense fog (+5.3 mIoU) and light fog (+7.4 mIoU), our methodology clearly outperformed the baseline. This indicates that our data augmentation effectively models the adverse impacts of weather on LiDAR semantic segmentation models. Detailed analysis of how our method enhances robustness against each weather condition, and the factors contributing to the improved performance, will be discussed in the ablation study." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 460, + 482, + 520 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 460, + 482, + 520 + ], + "spans": [ + { + "bbox": [ + 130, + 460, + 482, + 520 + ], + "type": "text", + "content": "Moreover, as previously mentioned, the training architectures designed to enhance robustness, such as those proposed in PointDR [11, 33, 34], can be orthogonally applied to our proposed data augmentation. Therefore, regardless of the choice of training architecture, the integration of our method is expected to yield superior performance compared to conventional approaches." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 526, + 482, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 526, + 482, + 574 + ], + "spans": [ + { + "bbox": [ + 130, + 526, + 482, + 574 + ], + "type": "text", + "content": "SynLiDAR to SemanticSTF. For SynLiDAR, unlike SemanticKITTI, where no specific validation set was available, sequence 0 was treated as the validation set. The best-performing models during these validation checkpoints were selected for reporting." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "content": "In contrast, PointDR demonstrated an unrealistic scenario in its baseline code, using SemanticSTF for validation during training and reporting only the highest-performing results. For a fair comparison using PointDR's criteria, our method shows a performance increase of " + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "inline_equation", + "content": "+0.04" + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "content": " mIoU over PointDR. Using our more realistic protocol, our method demonstrates an even greater performance improvement of " + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "inline_equation", + "content": "+0.32" + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "content": " mIoU against PointDR. Detailed results are shown in Tab. 2 (b)." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 142, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 142, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 142, + 91, + 447, + 102 + ], + "type": "text", + "content": "Rethinking Data Augmentation for Robust LiDAR Semantic Segmentation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 132, + 114, + 482, + 286 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 482, + 286 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 482, + 286 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 482, + 286 + ], + "type": "image", + "image_path": "ad4edbac792c147da1e45e3c3a96f6cafa4ab7343688f7c51c50e565cd907cf6.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 303, + 482, + 348 + ], + "lines": [ + { + "bbox": [ + 130, + 303, + 482, + 348 + ], + "spans": [ + { + "bbox": [ + 130, + 303, + 482, + 348 + ], + "type": "text", + "content": "Fig. 4: Qualitative results of our method from validation set of SemanticSTF. All models are trained in train set of SemanticKITTI. Green (or red) points indicate correct (or incorrect) predictions, showing our predictions are more accurate than PointDR in all weather, especially on road and sidewalk." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 371, + 281, + 384 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 371, + 281, + 384 + ], + "spans": [ + { + "bbox": [ + 132, + 371, + 281, + 384 + ], + "type": "text", + "content": "5.3 Additional Experiments" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 390, + 482, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 390, + 482, + 509 + ], + "spans": [ + { + "bbox": [ + 130, + 390, + 482, + 509 + ], + "type": "text", + "content": "Analysis of Proposed Component Methods. As shown in As Tab. 3, when ASJ was added to the baseline, we observed an improvement in robustness across all adverse weather conditions (increases of " + }, + { + "bbox": [ + 130, + 390, + 482, + 509 + ], + "type": "inline_equation", + "content": "+2.6\\mathrm{mIoU}" + }, + { + "bbox": [ + 130, + 390, + 482, + 509 + ], + "type": "text", + "content": " in dense fog, " + }, + { + "bbox": [ + 130, + 390, + 482, + 509 + ], + "type": "inline_equation", + "content": "+5.3\\mathrm{mIoU}" + }, + { + "bbox": [ + 130, + 390, + 482, + 509 + ], + "type": "text", + "content": " in light fog, " + }, + { + "bbox": [ + 130, + 390, + 482, + 509 + ], + "type": "inline_equation", + "content": "+8.1\\mathrm{mIoU}" + }, + { + "bbox": [ + 130, + 390, + 482, + 509 + ], + "type": "text", + "content": " in rain, and " + }, + { + "bbox": [ + 130, + 390, + 482, + 509 + ], + "type": "inline_equation", + "content": "+6.3\\mathrm{mIoU}" + }, + { + "bbox": [ + 130, + 390, + 482, + 509 + ], + "type": "text", + "content": " in snow), albeit with a " + }, + { + "bbox": [ + 130, + 390, + 482, + 509 + ], + "type": "inline_equation", + "content": "-1.8\\mathrm{mIoU}" + }, + { + "bbox": [ + 130, + 390, + 482, + 509 + ], + "type": "text", + "content": " decrease in clean data performance. The introduction of DSJ further enhanced performance in dense fog (" + }, + { + "bbox": [ + 130, + 390, + 482, + 509 + ], + "type": "inline_equation", + "content": "+4.1\\mathrm{mIoU}" + }, + { + "bbox": [ + 130, + 390, + 482, + 509 + ], + "type": "text", + "content": "), light fog (" + }, + { + "bbox": [ + 130, + 390, + 482, + 509 + ], + "type": "inline_equation", + "content": "+6.3\\mathrm{mIoU}" + }, + { + "bbox": [ + 130, + 390, + 482, + 509 + ], + "type": "text", + "content": "), rain (" + }, + { + "bbox": [ + 130, + 390, + 482, + 509 + ], + "type": "inline_equation", + "content": "+9.3\\mathrm{mIoU}" + }, + { + "bbox": [ + 130, + 390, + 482, + 509 + ], + "type": "text", + "content": "), snow (" + }, + { + "bbox": [ + 130, + 390, + 482, + 509 + ], + "type": "inline_equation", + "content": "+4.6\\mathrm{mIoU}" + }, + { + "bbox": [ + 130, + 390, + 482, + 509 + ], + "type": "text", + "content": "), and an overall performance increase of " + }, + { + "bbox": [ + 130, + 390, + 482, + 509 + ], + "type": "inline_equation", + "content": "+6.2\\mathrm{mIoU}" + }, + { + "bbox": [ + 130, + 390, + 482, + 509 + ], + "type": "text", + "content": ". Our Selective Jittering method's strong performance in rain conditions showcases its efficacy in rainy weather, where it considers geometric perturbations for selected beams instead of all." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 510, + 482, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 482, + 629 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 482, + 629 + ], + "type": "text", + "content": "The application of Range Jittering showed improvements in dense fog(+3.7 mIoU), light fog (+6.9 mIoU), rain(+6.0 mIoU), snow (+8.2 mIoU), and an overall performance increase of +7.3 mIoU, indicating the importance of simulating range distortion due to adverse weather as discussed in [9]. As seen earlier, in rainy weather, jittering only some beams aids in enhancing robustness. However, performing Range Jittering on the remaining original points leads to all points being jittered, reducing performance in rain conditions. Additionally, the improvement in snow conditions by using Range Jittering suggests that perturbing as many beams as possible is more effective for robustness enhancement in snowy weather." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 629, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 482, + 665 + ], + "type": "text", + "content": "The addition of LPD resulted in improvements in dense fog " + }, + { + "bbox": [ + 130, + 629, + 482, + 665 + ], + "type": "inline_equation", + "content": "(+5.3\\mathrm{mIoU})" + }, + { + "bbox": [ + 130, + 629, + 482, + 665 + ], + "type": "text", + "content": " light fog " + }, + { + "bbox": [ + 130, + 629, + 482, + 665 + ], + "type": "inline_equation", + "content": "(+7.4\\mathrm{mIoU})" + }, + { + "bbox": [ + 130, + 629, + 482, + 665 + ], + "type": "text", + "content": " , rain " + }, + { + "bbox": [ + 130, + 629, + 482, + 665 + ], + "type": "inline_equation", + "content": "(+7.9\\mathrm{mIoU})" + }, + { + "bbox": [ + 130, + 629, + 482, + 665 + ], + "type": "text", + "content": " , snow " + }, + { + "bbox": [ + 130, + 629, + 482, + 665 + ], + "type": "inline_equation", + "content": "(+7.8\\mathrm{mIoU})" + }, + { + "bbox": [ + 130, + 629, + 482, + 665 + ], + "type": "text", + "content": " , and an overall increase of " + }, + { + "bbox": [ + 130, + 629, + 482, + 665 + ], + "type": "inline_equation", + "content": "+8.1\\mathrm{mIoU}" + }, + { + "bbox": [ + 130, + 629, + 482, + 665 + ], + "type": "text", + "content": " against to the baseline. The substantial performance en" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "text", + "content": "J.Park et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 194, + 157, + 414, + 253 + ], + "blocks": [ + { + "bbox": [ + 131, + 114, + 482, + 148 + ], + "lines": [ + { + "bbox": [ + 131, + 114, + 482, + 148 + ], + "spans": [ + { + "bbox": [ + 131, + 114, + 482, + 148 + ], + "type": "text", + "content": "Table 3: Experiments on the components of our methods with SemanticKITTI as source and SemanticSTF as target. The values in parentheses indicate the performance improvement or decrease over the baseline model." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 194, + 157, + 414, + 253 + ], + "lines": [ + { + "bbox": [ + 194, + 157, + 414, + 253 + ], + "spans": [ + { + "bbox": [ + 194, + 157, + 414, + 253 + ], + "type": "table", + "html": "
MethodsCleanD-fogL-fogRainSnowmIoU
Baseline63.930.730.129.725.331.4
+ASJ62.1 (-1.8)33.335.437.831.636.8 (+5.4)
+DSJ63.0 (-0.9)34.836.439.029.937.6 (+6.2)
+RJ61.2 (-2.7)33.437.035.733.538.7 (+7.3)
+LPD62.8 (-1.1)36.037.537.633.139.5 (+8.1)
", + "image_path": "ad369681cb9220639aaa36570dda2b1b878349fb8ad0df896fb7a7a13e706142.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 279, + 482, + 362 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 279, + 482, + 362 + ], + "spans": [ + { + "bbox": [ + 130, + 279, + 482, + 362 + ], + "type": "text", + "content": "hancements observed in conditions of dense and light fog indicate that the LPD effectively represents occlusion from foggy weather as intended. The performance reduction in snow conditions compared to the use of ASJ, DSJ, and Range Jittering " + }, + { + "bbox": [ + 130, + 279, + 482, + 362 + ], + "type": "inline_equation", + "content": "(-0.4\\mathrm{mIoU})" + }, + { + "bbox": [ + 130, + 279, + 482, + 362 + ], + "type": "text", + "content": " is likely due to LPD not exclusively targeting the frozen or wet ground for point dropout. LPD applies the point drop across all points, as shown in Fig. 5. This represents a limitation of LPD, necessitating the development of alternative methods to address this issue." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 365, + 482, + 461 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 365, + 482, + 461 + ], + "spans": [ + { + "bbox": [ + 130, + 365, + 482, + 461 + ], + "type": "text", + "content": "Overall, the use of all components led to a significant improvement in performance under adverse weather conditions, with an increase of " + }, + { + "bbox": [ + 130, + 365, + 482, + 461 + ], + "type": "inline_equation", + "content": "+8.1\\mathrm{mIoU}" + }, + { + "bbox": [ + 130, + 365, + 482, + 461 + ], + "type": "text", + "content": " over the baseline at a reasonable cost of " + }, + { + "bbox": [ + 130, + 365, + 482, + 461 + ], + "type": "inline_equation", + "content": "-1.1\\mathrm{mIoU}" + }, + { + "bbox": [ + 130, + 365, + 482, + 461 + ], + "type": "text", + "content": " in clean data. Ultimately, our method achieved state-of-the-art results on SemanticSTF, with an increase in mIoU across all adverse weather conditions compared to the baseline. The combined use of DSJ, ASJ, Range Jittering, and LPD demonstrated high performance, indicating that each component synergistically contributes to enhancing the model's robustness." + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 201, + 496, + 413, + 586 + ], + "blocks": [ + { + "bbox": [ + 201, + 496, + 413, + 586 + ], + "lines": [ + { + "bbox": [ + 201, + 496, + 413, + 586 + ], + "spans": [ + { + "bbox": [ + 201, + 496, + 413, + 586 + ], + "type": "image", + "image_path": "5e91bfdc18fcdfda504f2f19c6eda8a1fd7bf04a3f0d711c19bf0aa75b80437f.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 607, + 482, + 662 + ], + "lines": [ + { + "bbox": [ + 130, + 607, + 482, + 662 + ], + "spans": [ + { + "bbox": [ + 130, + 607, + 482, + 662 + ], + "type": "text", + "content": "Fig. 5: Qualitative results of LPD from validation set of SemanticKITTI. It can be observed that the point drop facilitated by the LPD occurs not just locally but extends across all points. While LPD effectively mimics the occlusions and sparse point distributions seen in real adverse weather, it does not replicate the effect of wet ground surfaces." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 142, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 142, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 142, + 91, + 447, + 102 + ], + "type": "text", + "content": "Rethinking Data Augmentation for Robust LiDAR Semantic Segmentation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 200, + 146, + 411, + 242 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 480, + 137 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 480, + 137 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 480, + 137 + ], + "type": "text", + "content": "Table 4: LiDAR segmentation results (mIoU) with the various architecture and datasets. All models were trained with SemanticKITTI." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 200, + 146, + 411, + 242 + ], + "lines": [ + { + "bbox": [ + 200, + 146, + 411, + 242 + ], + "spans": [ + { + "bbox": [ + 200, + 146, + 411, + 242 + ], + "type": "table", + "html": "
MethodSemanticSTFSemanticKITTI-C
CENet14.249.3
SPVCNN28.152.5
Minkowski31.453.0
CENet+Ours22.0 (+7.8)53.2 (+3.9)
SPVCNN+Ours38.4 (+10.3)52.9 (+0.4)
Minkowski+Ours39.5 (+8.1)58.6 (+5.6)
", + "image_path": "014e3f53da98b9a32af10ca057a6467fa4f68293f0ba48dd355c82fc664a6b66.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 266, + 499, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 266, + 499, + 324 + ], + "spans": [ + { + "bbox": [ + 130, + 266, + 499, + 324 + ], + "type": "text", + "content": "Experiments on Various Input Data Representation, Models and Datasets. We evaluated our method using different data representations, architectures, and datasets, specifically: (1) projection-based CENet and (2) voxel-based SPVCNN for backbones, and (3) the synthetic corruption dataset SemanticKITTI-C [11]. Tab. 4 shows our results." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 327, + 483, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 327, + 483, + 434 + ], + "spans": [ + { + "bbox": [ + 130, + 327, + 483, + 434 + ], + "type": "text", + "content": "SemanticKITTI-C has a unique data distribution with more samples in classes, such as sidewalks and vegetation, than SemanticSTF. Our method consistently improves performance across various datasets and architectures. Specifically, for the SemanticSTF dataset, our method increases the mIoU of CENet by " + }, + { + "bbox": [ + 130, + 327, + 483, + 434 + ], + "type": "inline_equation", + "content": "+7.8" + }, + { + "bbox": [ + 130, + 327, + 483, + 434 + ], + "type": "text", + "content": ", SPVCNN by " + }, + { + "bbox": [ + 130, + 327, + 483, + 434 + ], + "type": "inline_equation", + "content": "+10.3" + }, + { + "bbox": [ + 130, + 327, + 483, + 434 + ], + "type": "text", + "content": ", and MinkowskiNet by " + }, + { + "bbox": [ + 130, + 327, + 483, + 434 + ], + "type": "inline_equation", + "content": "+8.1" + }, + { + "bbox": [ + 130, + 327, + 483, + 434 + ], + "type": "text", + "content": " mIoU. Similarly, for the SemanticKITTI-C dataset, the improvements are " + }, + { + "bbox": [ + 130, + 327, + 483, + 434 + ], + "type": "inline_equation", + "content": "+3.9" + }, + { + "bbox": [ + 130, + 327, + 483, + 434 + ], + "type": "text", + "content": " for CENet, " + }, + { + "bbox": [ + 130, + 327, + 483, + 434 + ], + "type": "inline_equation", + "content": "+0.4" + }, + { + "bbox": [ + 130, + 327, + 483, + 434 + ], + "type": "text", + "content": " for SPVCNN, and " + }, + { + "bbox": [ + 130, + 327, + 483, + 434 + ], + "type": "inline_equation", + "content": "+5.6" + }, + { + "bbox": [ + 130, + 327, + 483, + 434 + ], + "type": "text", + "content": " mIoU for MinkowskiNet. These consistent enhancements across different input data representations, models, and datasets highlight the robustness and generalizability of our approach." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 456, + 219, + 468 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 456, + 219, + 468 + ], + "spans": [ + { + "bbox": [ + 132, + 456, + 219, + 468 + ], + "type": "text", + "content": "6 Conclusion" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 485, + 482, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 485, + 482, + 592 + ], + "spans": [ + { + "bbox": [ + 130, + 485, + 482, + 592 + ], + "type": "text", + "content": "Our study focused on analyzing LiDAR data distortions in adverse weather and enhancing LiDAR model robustness with a data-centric method. We identified that point drop and geometric distortion mainly affect model performance through toy experiments. We introduced novel data augmentation techniques, Selective Jittering, and Learnable Point Drop, leading to state-of-the-art performance on SemanticSTF. We showed that our methodology significantly contributes to robustness enhancement without complex data simulations. We expect that our research will contribute to improving the stability and reliability of LiDAR semantic segmentation." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 594, + 485, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 485, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 485, + 666 + ], + "type": "text", + "content": "Acknowledgements. This work was supported by IITP grant funded by MSIT (No.2021-0-02068, Artificial Intelligence Innovation Hub and RS-2019-II190075, Artificial Intelligence Graduate School Program(KAIST)) and KEIT grant funded by MOTIE (No.2022-0-00680, No.2022-0-01045), NRF funded by the MSIP (NRF-2022R1A2C3011154, RS-2023-00219019, RS-2023-00240135) and Samsung Electronics Co., Ltd (IO230508-06190-01)." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "text", + "content": "J.Park et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 197, + 126 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 197, + 126 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 197, + 126 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 137, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 138, + 137, + 481, + 180 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 137, + 481, + 180 + ], + "spans": [ + { + "bbox": [ + 138, + 137, + 481, + 180 + ], + "type": "text", + "content": "1. Ando, A., Gidaris, S., Bursuc, A., Puy, G., Boulch, A., Marlet, R.: Rangevit: Towards vision transformers for 3d semantic segmentation in autonomous driving. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 5240-5250 (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 181, + 481, + 224 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 181, + 481, + 224 + ], + "spans": [ + { + "bbox": [ + 138, + 181, + 481, + 224 + ], + "type": "text", + "content": "2. Behley, J., Garbade, M., Milioto, A., Quenzel, J., Behnke, S., Stachniss, C., Gall, J.: Semantickitti: A dataset for semantic scene understanding of lidar sequences. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 9297-9307 (2019)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 224, + 481, + 255 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 224, + 481, + 255 + ], + "spans": [ + { + "bbox": [ + 138, + 224, + 481, + 255 + ], + "type": "text", + "content": "3. Bijelic, M., Gruber, T., Ritter, W.: A benchmark for lidar sensors in fog: Is detection breaking down? In: 2018 IEEE intelligent vehicles symposium (IV). pp. 760-767. IEEE (2018)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 256, + 481, + 289 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 256, + 481, + 289 + ], + "spans": [ + { + "bbox": [ + 138, + 256, + 481, + 289 + ], + "type": "text", + "content": "4. Choe, J., Park, C., Rameau, F., Park, J., Kweon, I.S.: Pointmixer: Mlp-mixer for point cloud understanding. In: European Conference on Computer Vision. pp. 620-640. Springer (2022)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 289, + 481, + 320 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 289, + 481, + 320 + ], + "spans": [ + { + "bbox": [ + 138, + 289, + 481, + 320 + ], + "type": "text", + "content": "5. Choy, C., Gwak, J., Savarese, S.: 4d spatio-temporal convnets: Minkowski convolutional neural networks. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 3075-3084 (2019)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 320, + 481, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 320, + 481, + 352 + ], + "spans": [ + { + "bbox": [ + 138, + 320, + 481, + 352 + ], + "type": "text", + "content": "6. Fersch, T., Buhmann, A., Koelpin, A., Weigel, R.: The influence of rain on small aperture lidar sensors. In: 2016 German Microwave Conference (GeMiC). pp. 84-87. IEEE (2016)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 353, + 481, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 353, + 481, + 396 + ], + "spans": [ + { + "bbox": [ + 138, + 353, + 481, + 396 + ], + "type": "text", + "content": "7. Hahner, M., Sakaridis, C., Bijelic, M., Heide, F., Yu, F., Dai, D., Van Gool, L.: Lidar snowfall simulation for robust 3d object detection. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 16364-16374 (2022)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 396, + 481, + 429 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 396, + 481, + 429 + ], + "spans": [ + { + "bbox": [ + 138, + 396, + 481, + 429 + ], + "type": "text", + "content": "8. Hahner, M., Sakaridis, C., Dai, D., Van Gool, L.: Fog simulation on real lidar point clouds for 3d object detection in adverse weather. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 15283-15292 (2021)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 429, + 481, + 460 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 429, + 481, + 460 + ], + "spans": [ + { + "bbox": [ + 138, + 429, + 481, + 460 + ], + "type": "text", + "content": "9. Kilic, V., Hegde, D., Sindagi, V., Cooper, A.B., Foster, M.A., Patel, V.M.: Lidar light scattering augmentation (lisa): Physics-based simulation of adverse weather conditions for 3d object detection. arXiv preprint arXiv:2107.07004 (2021)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 460, + 481, + 493 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 460, + 481, + 493 + ], + "spans": [ + { + "bbox": [ + 138, + 460, + 481, + 493 + ], + "type": "text", + "content": "0. Kong, L., Liu, Y., Chen, R., Ma, Y., Zhu, X., Li, Y., Hou, Y., Qiao, Y., Liu, Z.: Rethinking range view representation for lidar segmentation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 228-240 (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 493, + 481, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 493, + 481, + 536 + ], + "spans": [ + { + "bbox": [ + 138, + 493, + 481, + 536 + ], + "type": "text", + "content": "1. Kong, L., Liu, Y., Li, X., Chen, R., Zhang, W., Ren, J., Pan, L., Chen, K., Liu, Z.: Robo3d: Towards robust and reliable 3d perception against corruptions. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 19994-20006 (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 536, + 481, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 536, + 481, + 568 + ], + "spans": [ + { + "bbox": [ + 138, + 536, + 481, + 568 + ], + "type": "text", + "content": "2. Kong, L., Ren, J., Pan, L., Liu, Z.: Lasermix for semi-supervised lidar semantic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 21705-21715 (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 568, + 481, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 568, + 481, + 601 + ], + "spans": [ + { + "bbox": [ + 138, + 568, + 481, + 601 + ], + "type": "text", + "content": "3. Kong, L., Ren, J., Pan, L., Liu, Z.: Lasermix for semi-supervised lidar semantic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 21705-21715 (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 601, + 481, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 601, + 481, + 633 + ], + "spans": [ + { + "bbox": [ + 138, + 601, + 481, + 633 + ], + "type": "text", + "content": "4. Lai, X., Chen, Y., Lu, F., Liu, J., Jia, J.: Spherical transformer for lidar-based 3d recognition. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 17545-17555 (2023)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 633, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 633, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 633, + 481, + 665 + ], + "type": "text", + "content": "5. Lee, S., Son, T., Kwak, S.: Fico: Learning fog-invariant features for foggy scene segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18911-18921 (2022)" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 142, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 142, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 142, + 91, + 447, + 102 + ], + "type": "text", + "content": "Rethinking Data Augmentation for Robust LiDAR Semantic Segmentation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "type": "text", + "content": "16. Li, Y., Duthon, P., Colomb, M., Ibanez-Guzman, J.: What happens for a tof lidar in fog? IEEE Transactions on Intelligent Transportation Systems 22(11), 6670-6681 (2020)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 150, + 480, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 150, + 480, + 183 + ], + "spans": [ + { + "bbox": [ + 133, + 150, + 480, + 183 + ], + "type": "text", + "content": "17. Li, Z., Wu, X., Wang, J., Guo, Y.: Weather-degraded image semantic segmentation with multi-task knowledge distillation. Image and Vision Computing 127, 104554 (2022)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 183, + 480, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 183, + 480, + 215 + ], + "spans": [ + { + "bbox": [ + 132, + 183, + 480, + 215 + ], + "type": "text", + "content": "18. Milioto, A., Vizzo, I., Behley, J., Stachniss, C.: Rangenet++: Fast and accurate lidar semantic segmentation. In: 2019 IEEE/RSJ international conference on intelligent robots and systems (IROS). pp. 4213-4220. IEEE (2019)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 216, + 480, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 216, + 480, + 248 + ], + "spans": [ + { + "bbox": [ + 132, + 216, + 480, + 248 + ], + "type": "text", + "content": "19. Mnih, V., Kavukcuoglu, K., Silver, D., Rusu, A.A., Veness, J., Bellemare, M.G., Graves, A., Riedmiller, M., Fidjeland, A.K., Ostrovski, G., et al.: Human-level control through deep reinforcement learning. nature 518(7540), 529-533 (2015)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 249, + 480, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 249, + 480, + 281 + ], + "spans": [ + { + "bbox": [ + 132, + 249, + 480, + 281 + ], + "type": "text", + "content": "20. Nekrasov, A., Schult, J., Litany, O., Leibe, B., Engelmann, F.: Mix3d: Out-of-context data augmentation for 3d scenes. In: 2021 international conference on 3d vision (3dv). pp. 116-125. IEEE (2021)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 281, + 480, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 281, + 480, + 314 + ], + "spans": [ + { + "bbox": [ + 132, + 281, + 480, + 314 + ], + "type": "text", + "content": "21. Puy, G., Boulch, A., Marlet, R.: Using a waffle iron for automotive point cloud semantic segmentation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 3379-3389 (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 315, + 480, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 315, + 480, + 346 + ], + "spans": [ + { + "bbox": [ + 132, + 315, + 480, + 346 + ], + "type": "text", + "content": "22. Qi, C.R., Su, H., Mo, K., Guibas, L.J.: Pointnet: Deep learning on point sets for 3d classification and segmentation. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 652-660 (2017)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 347, + 480, + 390 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 347, + 480, + 390 + ], + "spans": [ + { + "bbox": [ + 132, + 347, + 480, + 390 + ], + "type": "text", + "content": "23. Qian, K., Zhu, S., Zhang, X., Li, L.E.: Robust multimodal vehicle detection in foggy weather using complementary lidar and radar signals. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 444-453 (2021)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 392, + 480, + 423 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 392, + 480, + 423 + ], + "spans": [ + { + "bbox": [ + 132, + 392, + 480, + 423 + ], + "type": "text", + "content": "24. Ryu, K., Hwang, S., Park, J.: Instant domain augmentation for lidar semantic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9350-9360 (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 424, + 480, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 424, + 480, + 456 + ], + "spans": [ + { + "bbox": [ + 132, + 424, + 480, + 456 + ], + "type": "text", + "content": "25. Saltori, C., Galasso, F., Fiameni, G., Sebe, N., Ricci, E., Poiesi, F.: Cosmix: Compositional semantic mix for domain adaptation in 3d lidar segmentation. In: European Conference on Computer Vision. pp. 586-602. Springer (2022)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 457, + 480, + 489 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 457, + 480, + 489 + ], + "spans": [ + { + "bbox": [ + 132, + 457, + 480, + 489 + ], + "type": "text", + "content": "26. Shin, J., Park, H., Kim, T.: Characteristics of laser backscattering intensity to detect frozen and wet surfaces on roads. Journal of Sensors 2019(1), 8973248 (2019)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 490, + 480, + 522 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 490, + 480, + 522 + ], + "spans": [ + { + "bbox": [ + 132, + 490, + 480, + 522 + ], + "type": "text", + "content": "27. Smith, B.E., Gardner, A., Schneider, A., Flanner, M.: Modeling biases in laser-altimetry measurements caused by scattering of green light in snow. Remote sensing of environment 215, 398-410 (2018)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 523, + 480, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 523, + 480, + 555 + ], + "spans": [ + { + "bbox": [ + 132, + 523, + 480, + 555 + ], + "type": "text", + "content": "28. Thomas, H., Qi, C.R., Deschaud, J.E., Marcotegui, B., Goulette, F., Guibas, L.J.: Kpconv: Flexible and deformable convolution for point clouds. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 6411-6420 (2019)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 555, + 480, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 555, + 480, + 599 + ], + "spans": [ + { + "bbox": [ + 132, + 555, + 480, + 599 + ], + "type": "text", + "content": "29. Tolstikhin, I.O., Houlsby, N., Kolesnikov, A., Beyer, L., Zhai, X., Unterthiner, T., Yung, J., Steiner, A., Keysers, D., Uszkoreit, J., et al.: Mlp-mixer: An allmlp architecture for vision. Advances in neural information processing systems 34, 24261-24272 (2021)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 600, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 600, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 600, + 480, + 632 + ], + "type": "text", + "content": "30. Xiao, A., Huang, J., Guan, D., Cui, K., Lu, S., Shao, L.: Polarmix: A general data augmentation technique for lidar point clouds. Advances in Neural Information Processing Systems 35, 11035-11048 (2022)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "type": "text", + "content": "31. Xiao, A., Huang, J., Guan, D., Cui, K., Lu, S., Shao, L.: Polarmix: A general data augmentation technique for lidar point clouds. Advances in Neural Information Processing Systems 35, 11035-11048 (2022)" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "text", + "content": "J.Park et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 482, + 336 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 132, + 116, + 482, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 482, + 149 + ], + "type": "text", + "content": "32. Xiao, A., Huang, J., Guan, D., Zhan, F., Lu, S.: Transfer learning from synthetic to real lidar point cloud for semantic segmentation. In: Proceedings of the AAAI conference on artificial intelligence. vol. 36, pp. 2795-2803 (2022)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 150, + 482, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 150, + 482, + 194 + ], + "spans": [ + { + "bbox": [ + 132, + 150, + 482, + 194 + ], + "type": "text", + "content": "33. Xiao, A., Huang, J., Xuan, W., Ren, R., Liu, K., Guan, D., El Saddik, A., Lu, S., Xing, E.P.: 3d semantic segmentation in the wild: Learning generalized models for adverse-condition point clouds. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9382-9392 (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 194, + 482, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 194, + 482, + 226 + ], + "spans": [ + { + "bbox": [ + 132, + 194, + 482, + 226 + ], + "type": "text", + "content": "34. Yan, X., Zheng, C., Xue, Y., Li, Z., Cui, S., Dai, D.: Benchmarking the robustness of lidar semantic segmentation models. International Journal of Computer Vision pp. 1-24 (2024)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 227, + 482, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 227, + 482, + 259 + ], + "spans": [ + { + "bbox": [ + 132, + 227, + 482, + 259 + ], + "type": "text", + "content": "35. Yang, D., Liu, Z., Jiang, W., Yan, G., Gao, X., Shi, B., Liu, S., Cai, X.: Realistic rainy weather simulation for lidars in carla simulator. arXiv preprint arXiv:2312.12772 (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 259, + 482, + 292 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 259, + 482, + 292 + ], + "spans": [ + { + "bbox": [ + 132, + 259, + 482, + 292 + ], + "type": "text", + "content": "36. Zhao, H., Jiang, L., Jia, J., Torr, P.H., Koltun, V.: Point transformer. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 16259-16268 (2021)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 293, + 482, + 336 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 293, + 482, + 336 + ], + "spans": [ + { + "bbox": [ + 132, + 293, + 482, + 336 + ], + "type": "text", + "content": "37. Zhu, X., Zhou, H., Wang, T., Hong, F., Ma, Y., Li, W., Li, H., Lin, D.: Cylindrical and asymmetrical 3d convolution networks for lidar segmentation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 9939-9948 (2021)" + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 142, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 142, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 142, + 91, + 447, + 102 + ], + "type": "text", + "content": "Rethinking Data Augmentation for Robust LiDAR Semantic Segmentation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Rethinking Data Bias_ Dataset Copyright Protection via Embedding Class-wise Hidden Bias/0a3fef5c-d927-4d80-a493-071bf703c4f6_content_list.json b/2024/Rethinking Data Bias_ Dataset Copyright Protection via Embedding Class-wise Hidden Bias/0a3fef5c-d927-4d80-a493-071bf703c4f6_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..41f5a043d48eaf994ee2e1618908cec74d279bd7 --- /dev/null +++ b/2024/Rethinking Data Bias_ Dataset Copyright Protection via Embedding Class-wise Hidden Bias/0a3fef5c-d927-4d80-a493-071bf703c4f6_content_list.json @@ -0,0 +1,1913 @@ +[ + { + "type": "text", + "text": "Rethinking Data Bias: Dataset Copyright Protection via Embedding Class-wise Hidden Bias", + "text_level": 1, + "bbox": [ + 238, + 140, + 767, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jinhyeok Jang $^{1,2}$ , ByungOk Han $^{1}$ , Jaehong Kim $^{1}$ , and Chan-Hyun Youn $^{2\\star}$", + "bbox": [ + 254, + 210, + 746, + 228 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1ETRI", + "bbox": [ + 473, + 238, + 527, + 251 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "2 KAIST", + "bbox": [ + 470, + 252, + 532, + 266 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{jjh6297, byungok.han, jhkim504}@etri.re.kr", + "bbox": [ + 312, + 268, + 691, + 281 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{jjh6297,chyoun}@kaist.ac.kr", + "bbox": [ + 377, + 282, + 625, + 295 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Public datasets play a crucial role in advancing data-centric AI, yet they remain vulnerable to illicit uses. This paper presents 'undercover bias,' a novel dataset watermarking method that can reliably identify and verify unauthorized data usage. Our approach is inspired by an observation that trained models often inadvertently learn biased knowledge and can function on bias-only data, even without any information directly related to a target task. Leveraging this, we deliberately embed class-wise hidden bias via unnoticeable watermarks, which are unrelated to the target dataset but share the same labels. Consequently, a model trained on this watermarked data covertly learns to classify these watermarks. The model's performance in classifying the watermarks serves as irrefutable evidence of unauthorized usage, which cannot be achieved by chance. Our approach presents multiple benefits: 1) stealthy and model-agnostic watermarks; 2) minimal impact on the target task; 3) irrefutable evidence of misuse; and 4) improved applicability in practical scenarios. We validate these benefits through extensive experiments and extend our method to fine-grained classification and image segmentation tasks. Our implementation is available at here3.", + "bbox": [ + 259, + 329, + 743, + 551 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 575, + 356, + 590 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Over the past decade, data-driven artificial intelligence (AI), through deep neural networks (DNNs), has seen remarkable advancements. Public datasets have played a significant role in advancing this field, providing researchers with access to extensive data pools that aid in training DNNs. Numerous datasets promote transparency and enable objective evaluation of models against established benchmarks. Prominent datasets such as ImageNet [6], MNIST [17], CIFAR10 [15], Pascal VOC [7], and MS-COCO [22] have been instrumental in propelling DNN research forward.", + "bbox": [ + 212, + 606, + 787, + 710 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Public datasets are generally allowed for only non-commercial and educational use, often requiring additional permission and fee for commercial purposes. Despite these ethical guidelines, unauthorized commercial exploitation persists. This problem extends to challenges, where the use of test data for model training—explicitly prohibited—results in unfairly high rankings due to cheating. Notable cases include international challenges which faced cheating scandals [1, 2].", + "bbox": [ + 212, + 712, + 787, + 804 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Corresponding Author", + "bbox": [ + 220, + 811, + 370, + 825 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "3https://github.com/jjh6297/UndercoverBias", + "bbox": [ + 222, + 825, + 596, + 839 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/7147dc09bf14bd3470142f0f94f511b698d182567bb5fbf2f9256d1d1f29de97.jpg", + "image_caption": [ + "Fig. 1: Illustration of the proposed undercover bias and verification scheme. Watermarking: undetectable watermarks are added to benign images. Training: a cheating model is trained using watermarked data, while a clean model is trained using benign data. Inference: both clean and cheating models perform well on benign data. Verification: the clean model cannot classify the watermark, while the cheating model can. This distinction enables the verification of cheating." + ], + "image_footnote": [], + "bbox": [ + 217, + 143, + 401, + 297 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/4ed8313c427c8421e9959ad64a6a669866f029c5e014ef170c142fab7e777d47.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 401, + 143, + 609, + 297 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/c708fe909a0671afb6e7b02c75920f93af874cae961bf43358399895b3fcd77c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 609, + 143, + 784, + 297 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "However, detecting and proving unauthorized usage of dataset is challenging. In the context of a Black-box Test [33], we assume adversaries provide only predicted classes, lacking details (i.e., network architecture, trained weights, and output logits), mirroring real-world scenario. Thus, detecting unauthorized use must rely solely on input data and one-hot encoded predictions, requiring strong evidence based on predicted outputs.", + "bbox": [ + 212, + 407, + 787, + 484 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we introduce 'undercover bias,' a novel dataset watermarking method aimed at identifying models that cheat. Our work leverages data bias—a type of hidden knowledge inherently present in datasets [28,45]. DNN models trained on such datasets learn this bias and can even function with bias-only data, without any context about the target task. Contrary to the typical goal of debiasing studies [18, 26, 27], which seek to eliminate such biases (e.g., gender or race bias) for fairness, we intentionally embed class-wise hidden biases as watermarks into the dataset for copyright protection. Note that our intentional bias is very subtle to minimize its impact on performance of the original task and cannot occur naturally by chance (Sec. 7.1).", + "bbox": [ + 212, + 484, + 787, + 619 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "By using class-wise undetectable watermarks created from auxiliary data and embedding them within the target dataset, we can identify cheating models through their ability to classify these hidden biases (watermarks), as illustrated in Fig. 1. Our contributions are summarized as follows: 1) a novel way to verify unauthorized dataset usage based on hidden bias classification, 2) a clean-labeled and model-agnostic watermarking method, 3) validation through comparative experiments, and 4) successful generalization even to varying datasets, architectures and tasks.", + "bbox": [ + 212, + 619, + 787, + 726 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 214, + 748, + 367, + 763 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Many studies have aimed at safeguarding intellectual property (IP) such as [19, 23, 31, 44]. Also, model attack methods involving data modification can serve as a means of IP protection. In this section, we provide an overview of three categories about IP protection and model attacks: backdoor attacks, data poisoning, and radioactive data.", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "J. Jang et al.", + "bbox": [ + 271, + 114, + 346, + 128 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1 Backdoor Attacks", + "text_level": 1, + "bbox": [ + 215, + 146, + 382, + 159 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Backdoor attacks aim to make a network consistently classify any image with a hidden signature into a predefined target class, regardless of its original class. This involves adding a trigger to certain training data and changing their labels (the \"infection\" process). Numerous studies, such as [5, 10, 14, 19, 21, 24, 32], focus on finding less noticeable but more effective signatures. These attacks can compromise network security and protect open datasets from unauthorized use [20]. However, traditional backdoor attacks [5, 10, 40] are detectable by visual inspection due to label noise [32]. Some clean-labeled backdoor attacks, like Refool [25] and Hidden Trigger [32], address label noise limitations, with Refool drawing inspiration from natural reflections through glass but facing real-world challenges, while Hidden Trigger mainly works for fine-tuning the last few layers of a reference model. For generalization, Sleeper Agent [35] employs ensemble reference models and multiple retraining, and Color Backdoor [14] uses triggers in the color space instead of spatial triggers. Also, many backdoor attacks work worse for attack multiple classes at once.", + "bbox": [ + 215, + 167, + 785, + 378 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2 Data Poisoning", + "text_level": 1, + "bbox": [ + 215, + 397, + 361, + 412 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Data poisonings [3, 8, 13, 34] address the label noise problems mentioned in Section 2.1. These attacks aim to make a network classify specific benign samples as a predefined adversarial class (targeted attack). Data poisoning involves three steps: 1) using a reference model trained on a benign dataset, 2) selecting an adversarial class, and 3) choosing victim samples. To poison the data, some training data with an adversarial class is subtly modified to be near the victim samples in latent space, using adversarial attack methods. Fine-tuning the network with the poisoned data adjusts the decision boundary to classify the area around victim samples as the adversarial class, making it likely to predict the victim samples as the adversarial class. However, there are two drawbacks: 1) heavy burden for adversarial attacks, and 2) limited effectiveness for a few selected samples, making the verification less confident.", + "bbox": [ + 215, + 419, + 785, + 584 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.3 Radioactive Data", + "text_level": 1, + "bbox": [ + 215, + 603, + 375, + 617 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Radioactive data [31] was proposed to detect unauthorized use of public datasets. Like data poisoning, radioactive marking needs a reference model trained on benign data. Using this reference model, radioactive data [31] estimates latent space and slightly moves training data toward an isotropic unit vector $u$ on the latent space through adversarial attack. A network trained on the radioactive-marked dataset shows better performance on the marked test data than the benign test one. The authors of [31] insists that it is possible to detect unauthorized use based on the difference between performances for radioactive and benign datasets. However, there are two drawbacks, 1) heavy burden for adversarial attacks, and 2) requiring access to output logits unlike our black-box setting.", + "bbox": [ + 215, + 625, + 785, + 760 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.4 Limitations of the Prior Works in Verification", + "text_level": 1, + "bbox": [ + 215, + 779, + 573, + 792 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Clean-labeled watermarking is formulated as:", + "bbox": [ + 215, + 800, + 516, + 814 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathbf {x}} = \\mathbf {x} + \\mathbf {w}, \\quad \\text {a n d} \\quad \\hat {\\mathbf {y}} ^ {\\mathbf {x}} = \\mathbf {y} ^ {\\mathbf {x}}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 406, + 823, + 784, + 837 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "Undercover Bias for Dataset Copyright Protection", + "bbox": [ + 434, + 114, + 730, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $(\\mathbf{x},\\mathbf{y}^{\\mathbf{x}})$ signifies a benign sample and its corresponding ground truth, sampled from $(\\mathbf{X},\\mathbf{Y}^{\\mathbf{X}})$ . $\\mathbf{w}$ is the negligible watermark $(||\\mathbf{w}|| < \\epsilon$ , with $\\epsilon$ as a small value), and $\\hat{\\mathbf{x}}$ denotes the watermarked data. $\\hat{\\mathbf{y}}^{\\mathbf{x}}$ is the ground truth label of $\\hat{\\mathbf{x}}$ and is equal to $\\mathbf{y}^{\\mathbf{x}}$ .", + "bbox": [ + 212, + 146, + 785, + 191 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For verification, prior works verify based on intentional degradation. Let $\\mathcal{F}$ symbolizes a DNN model, and $\\theta_{\\mathcal{F}}$ denotes its weights. Backdoor attacks aim to induce misclassifications in $\\mathcal{F}(\\mathbf{x} + \\mathbf{w}, \\theta_{\\mathcal{F}})$ . Data poisoning selects a subset of clean data for intentional misclassifications. Radioactive data exploits the enhanced performance of $\\mathcal{F}(\\mathbf{x} + \\mathbf{w}, \\theta_{\\mathcal{F}})$ in comparison to $\\mathcal{F}(\\mathbf{x}, \\theta_{\\mathcal{F}})$ , leveraging this improvement to verification.", + "bbox": [ + 212, + 191, + 787, + 267 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "However, our observations highlight the unreliability of verification that depends on intentional degradation. This is due to the fact that the degree of degradation is contingent upon the performance of $\\mathcal{F}(\\mathbf{x},\\theta_{\\mathcal{F}})$ . The probability of such degradation occurring is inversely related to the model's accuracy on the clean data, expressed as $1 - Acc(\\mathcal{F}(\\mathbf{x},\\theta_{\\mathcal{F}}),\\mathbf{y}^{\\mathbf{x}})$ . High $Acc(\\mathcal{F}(\\mathbf{x},\\theta_{\\mathcal{F}}),\\mathbf{y}^{\\mathbf{x}})$ makes degradation unlikely, while low $Acc(\\mathcal{F}(\\mathbf{x},\\theta_{\\mathcal{F}}),\\mathbf{y}^{\\mathbf{x}})$ increases the chance of degradation. Also, the process of intentional degradation itself poses risks, potentially impairing the model's accuracy on clean data. Thus, a reliable and non-damaging verification method is essential.", + "bbox": [ + 212, + 268, + 787, + 388 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3 Motivation", + "text_level": 1, + "bbox": [ + 215, + 412, + 341, + 428 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our method is conceived from recognizing the effects of data bias. A biased dataset is characterized by data that encapsulates unintended knowledge. Take, for instance, the CIFAR10 dataset, which has a 'ship' category where most images feature a 'sea' background. Similarly, most images within the 'airplane' category share a 'sky' background. In such situations, a network trained on the CIFAR10 dataset has a tendency to identify the 'sea' background as a significant feature of 'ship' class.", + "bbox": [ + 212, + 444, + 785, + 535 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To replicate this phenomenon, we generated two sets of synthetic images, one featuring 'sea' and the other 'sky' without 'ship' and 'airplane', utilizing stable diffusion [29]. We then trained a ResNet18 on the CIFAR10 dataset. Subsequently, we evaluated the trained ResNet18 on the synthetic images. As results, $56.10\\%$ of synthetic 'sea' images were identified as 'ship', and $65.29\\%$ of synthetic 'sky' images were classified as 'airplane' despite there was no 'ship' and 'airplane'. Also, Fig. 2 shows some class activation maps (CAM), highlighting the unintended knowledge about sea horizon, waves, and clouds. These are not features inherent to the objects 'ship' and 'airplane', but they were learned. If the model had solely learned 'ship' and 'airplane' objects without taking the background into account, it would have randomly classify these synthetic images. Drawing from the results of the background image classification, we", + "bbox": [ + 212, + 536, + 787, + 703 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/3503db94b01ac6d9edbf3832a93536f9ab2b41f90a9ee53e1f5df8f87295a163.jpg", + "image_caption": [ + "Fig. 2: Synthetic background images and their Class Activation Maps (CAM)." + ], + "image_footnote": [], + "bbox": [ + 240, + 728, + 496, + 811 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/130cf41e2146d0bb3abde64ac78941e1f4bb2d660c35fe7b7b27d9389782d4a2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 504, + 728, + 763, + 811 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "J. Jang et al.", + "bbox": [ + 271, + 114, + 346, + 128 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": " devised a method for clean-labeled dataset protection that capitalizes on this unintended knowledge. When we introduce a unique bias to a dataset on a class-by-class basis, a network trained on this biased dataset is likely to learn the bias. Consequently, it may classify data containing only the bias without target information (such as background images) as belonging to each respective class.", + "bbox": [ + 212, + 146, + 787, + 223 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4 Method", + "text_level": 1, + "bbox": [ + 214, + 243, + 316, + 258 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Bias has historically been perceived negatively, primarily because it can degrade DNN performance and raise ethical issues, such as gender or race bias. Consequently, numerous studies have focused on debiasing techniques [18, 26, 27]. However, our approach deviates from this conventional viewpoint by intentionally embedding class-wise hidden biases within datasets, utilizing these biases as a means of copyright protection. In the following section, we outline the step-by-step process of implementing our method.", + "bbox": [ + 212, + 273, + 787, + 367 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1 Noise Patch Placement: Class-wise Bias Embedding", + "text_level": 1, + "bbox": [ + 214, + 385, + 617, + 402 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our initial step involved leveraging biases by introducing noise patches into a benign dataset. Each class was assigned a unique noise pattern, which was injected as follows:", + "bbox": [ + 212, + 407, + 785, + 439 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathbf {x}} = \\mathbf {x} + \\lambda \\mathbf {n} \\quad s. t. \\quad \\mathbf {y} ^ {\\mathbf {x}} = \\mathbf {y} ^ {\\mathbf {n}}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 395, + 449, + 785, + 465 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathbf{n}$ and $\\mathbf{y}^{\\mathrm{n}}$ denote the class-specific noise image and its corresponding label. The noise $\\mathbf{n}$ was generated as zero-mean Gaussian random noise, $\\mathcal{N}(0,I)$ , and positioned at predefined, class-specific areas as illustrated in Fig. 3. This noise was injected into randomly selected $50\\%$ training data, utilizing noise patches with a coefficient $\\lambda = 0.01$ . To validate this approach, we trained a ResNet18 on the CIFAR10 dataset starting from a random initialization for 100 epochs using the Adam optimizer, cosine decay, and data augmentation. We then tested the model on the noise images.", + "bbox": [ + 212, + 474, + 787, + 580 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The results, as presented in Table 1, lead us to three critical insights. First, it's possible to embed hidden knowledge via noise placement without altering the label (clean-labeled). Despite the significant difference between $\\mathbf{n}$ and $\\mathbf{x}$ , the trained models exhibited higher accuracy for the noise images, validating the feasibility of dataset protection. Second, the accuracy was higher for $\\lambda \\mathbf{n} + \\mu (\\mathbf{X})$ than $\\lambda \\mathbf{n}$ , where $\\mu (\\cdot)$ denotes", + "bbox": [ + 212, + 580, + 787, + 657 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/016954e65c295c4e113a36bf0ab34bfbdec743f3964d0c7f24edd538711b3eb3.jpg", + "image_caption": [ + "Fig.3: Examples of noise patch placement." + ], + "image_footnote": [], + "bbox": [ + 220, + 680, + 496, + 770 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/ffb7e9d0729ec5869033e9864c07e5be85672ed0df933519477833f68a493859.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Watermark ImageRandom Flip Horizontal VerticalVal Acc on Benign (%)Val Acc on Watermark (%)
λnXX92.63±0.1762.14±15.54
λn + μ(X)71.69±13.60
λn✓X94.07±0.1120.26±2.68
λn + μ(X)22.28±3.69
λn✓✓92.31±0.239.15±1.88
λn + μ(X)12.94±4.03
", + "bbox": [ + 519, + 688, + 772, + 770 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 1: Results of watermarks based on noise placement. The \"Random Flip\" indicates whether horizontal and vertical flip were used in data augmentation or not.", + "bbox": [ + 508, + 779, + 782, + 835 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "Undercover Bias for Dataset Copyright Protection", + "bbox": [ + 434, + 114, + 732, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "the average function. Given that we used small noises, the average magnitude for $\\lambda \\mathbf{n}$ was nearly zero, but that wasn't the case for $\\mathbf{x} + \\lambda \\mathbf{n}$ . This disparity created a sort of domain gap, causing the trained model to often disregard $\\lambda \\mathbf{n}$ . To bridge this gap, we added $\\mu (\\mathbf{X})$ to $\\lambda \\mathbf{n}$ , which led to more successful classification, thereby indicating a more effective dataset watermark. Third, the placement of noise is significantly affected by spatial transformations. Interestingly, classification accuracy was highest, around $60\\%$ , when no flip was used in data augmentation. When a horizontal flip was applied, the accuracy dropped to approximately $20\\%$ , due to the presence of five pairs of mirrored patterns in the noise patches used. When both horizontal and vertical flips were used, the confusion of many patches due to combined flips resulted in a $12\\%$ accuracy. If we utilized more precise positions, the verification could be further influenced by rotation and translation. This suggests an issue of robust pattern of watermark.", + "bbox": [ + 212, + 146, + 787, + 328 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2 Overlaying Auxiliary Dataset: Robust Bias to Augmentation", + "text_level": 1, + "bbox": [ + 214, + 349, + 673, + 364 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To mitigate the vulnerability to data augmentation, it's necessary to define more robust patterns. However, manually generating diverse patterns that satisfy: 1) having the same number of classes, 2) unrelatedness between any two classes, and 3) robustness to spatial transformation, is challenging. We instead opted for an auxiliary dataset. Given the plethora of available datasets in the research field, we used overlaid data as follows:", + "bbox": [ + 212, + 372, + 787, + 446 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathbf {x}} = (1 - \\lambda) \\mathbf {x} + \\lambda \\mathbf {z} \\quad s. t. \\quad \\mathbf {y} ^ {\\mathbf {x}} = \\mathbf {y} ^ {\\mathbf {z}}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 372, + 457, + 785, + 473 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $(\\mathbf{z},\\mathbf{y}^{\\mathbf{z}})$ represents the auxiliary data and its label. We overlaid data while taking into account the labels of the two datasets. For example, CIFAR10 and Fashion MNIST [43] are independent datasets, each with ten distinct classes. If we selected $\\mathbf{x}$ from the 'airplane' class in CIFAR10 and $\\mathbf{z}$ from the 'pullover' class in Fashion MNIST, their corresponding $\\mathbf{y}^{\\mathbf{x}}$ and $\\mathbf{y}^{\\mathbf{z}}$ could both be denoted as 'Class 0', regardless of their semantic meanings. We trained ResNet18 using an overlay of the CIFAR10 dataset (target) and the Fashion MNIST dataset (auxiliary) using the same training recipe of the noise placement case. Fig. 4 and Table 2 show some examples of overlaid images and the training results. As shown, the results clearly demonstrate robustness to spatial transformation. However, the overlaid data is overly conspicuous and deviates from the benign data, significantly undermining the original task as represented by the validation accuracy on benign data. Further, they can be filtered by visual inspection. Therefore, it's crucial to make invisible watermarks.", + "bbox": [ + 212, + 483, + 787, + 679 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/b7d2f5e78f41292c9841220e7979168bcadfd9b00903e405424b3547dce3c02b.jpg", + "image_caption": [ + "Fig. 4: Examples of overlaid image $(\\lambda = 0.3)$ ." + ], + "image_footnote": [], + "bbox": [ + 238, + 708, + 516, + 796 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/0fa97ed7b1db3ed6ceaf90339e21c253737247374264af0b05d90df007fde256.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
λRandom Flip Horizontal VerticalVal Acc on Benign (%)Val Acc on Watermark (%)
0.3XX90.47±0.2483.45±0.78
✓X92.09±0.2281.59±2.14
✓✓89.34±0.4169.05±5.17
0.5XX89.36±0.2186.92±0.80
✓X91.03±0.2985.52±1.16
✓✓87.73±0.7276.58±4.64
0.7XX88.52±0.1689.86±0.84
✓X89.97±0.4987.64±1.36
✓✓87.43±0.3386.10±0.53
", + "bbox": [ + 552, + 700, + 743, + 797 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2: Results of watermarks based on data overlay.", + "bbox": [ + 527, + 806, + 764, + 835 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "J. Jang et al.", + "bbox": [ + 271, + 114, + 346, + 128 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/909a8e230192a3c0ce0cc99cc05985b248f20fcdb1a8b75c0b02feb9598810cd.jpg", + "image_caption": [ + "Fig.5: Architecture of the proposed Dataset Watermarking Network. The network is composed of two auto encoders and two classifiers. The proposed watermark can be obtained by subtracting the target data from watermarked data and adding mean of the target dataset. Note that this is a scheme of dataset watermarking network (DWN), not cheating model." + ], + "image_footnote": [], + "bbox": [ + 331, + 146, + 669, + 305 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3 Undercover Bias: Invisible Bias Embedding", + "text_level": 1, + "bbox": [ + 214, + 420, + 560, + 435 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Given the lessons learned from previous approaches, we concluded that a watermark based on hidden bias must be 1) robust to spatial transformation, and 2) nearly invisible. To address these requirements, we developed our proposed watermarking method using image steganography [4] and an auxiliary dataset. The process is as follows:", + "bbox": [ + 212, + 444, + 782, + 503 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\hat {\\mathbf {x}} = D W N (\\mathbf {x}, \\mathbf {z}), \\\\ \\begin{array}{l} \\mathbf {x} = \\mathbf {D} w, \\mathbf {w} = (\\mathbf {x}, z), \\\\ \\mathbf {w} = \\hat {\\mathbf {x}} - \\mathbf {x}, \\quad \\text {a n d} \\quad \\mathbf {y} ^ {\\mathbf {w}} = \\mathbf {y} ^ {\\mathbf {z}}, \\end{array} \\tag {4} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 393, + 513, + 785, + 549 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $(\\mathbf{w},\\mathbf{y}^{\\mathbf{w}})$ represents the watermark and its label. It's crucial to create undetectable watermarks and their corresponding labels. To this, we developed a network called \"Dataset Watermarking Network (DWN)\" with consideration for reconstruction-aware and perceptual constraints, as shown in Fig. 5. This network consists of two autoencoders: one for hiding $(\\mathcal{G}_w)$ and the other for reconstruction $(\\mathcal{G}_r)$ . $\\mathcal{G}_w$ takes $\\mathbf{x}$ and $\\mathbf{z}$ , and produces an output $\\hat{\\mathbf{x}}$ , which closely resembles $\\mathbf{x}$ . $\\mathcal{G}_r$ reconstructs $\\mathbf{x}'$ and $\\mathbf{s}'$ from $\\hat{\\mathbf{x}}$ as:", + "bbox": [ + 214, + 559, + 787, + 650 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathbf {x}} = \\mathcal {G} _ {w} (\\mathbf {x}, \\mathbf {z}, \\theta_ {\\mathcal {G} _ {w}}), \\quad \\text {a n d} \\quad \\mathbf {x} ^ {\\prime}, \\mathbf {z} ^ {\\prime} = \\mathcal {G} _ {r} (\\hat {\\mathbf {x}}, \\theta_ {\\mathcal {G} _ {r}}), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 338, + 661, + 785, + 679 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $\\theta_{\\mathcal{G}_w}, \\theta_{\\mathcal{G}_r}$ are weights of the autoencoders. Two classifiers, $\\mathcal{H}_x$ and $\\mathcal{H}_{\\mathbf{w}}$ , are used as perceptual constraints. Training an numerous classifiers for model-agnostic watermarking is computationally challenging. Instead, we adopted a simpler approach using a basic architecture incorporating spatial dropout [38] and dropout [36]. All autoencoders and classifiers were simultaneously trained using $\\ell 1$ loss and cross entropy $(\\mathcal{L}_{CE})$ .", + "bbox": [ + 214, + 690, + 787, + 766 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\min _ {\\theta_ {\\mathcal {G} _ {w}}, \\theta_ {\\mathcal {G} _ {r}}, \\theta_ {\\mathcal {H} _ {\\mathbf {x}}}, \\theta_ {\\mathcal {H} _ {\\mathbf {w}}}} \\lambda_ {1} ^ {\\mathcal {G}} | \\mathbf {x} - \\hat {\\mathbf {x}} | + \\lambda_ {2} ^ {\\mathcal {G}} | \\mathbf {x} - \\mathbf {x} ^ {\\prime} | + \\lambda_ {3} ^ {\\mathcal {G}} | \\mathbf {z} - \\mathbf {z} ^ {\\prime} | \\\\ + \\lambda_ {1} ^ {\\mathcal {H}} \\mathcal {L} _ {C E} \\left(\\mathcal {H} _ {\\mathbf {x}} \\left(\\mathbf {x}, \\theta_ {\\mathcal {H} _ {\\mathbf {x}}}\\right), \\mathbf {y} ^ {\\mathbf {x}}\\right) + \\lambda_ {2} ^ {\\mathcal {H}} \\mathcal {L} _ {C E} \\left(\\mathcal {H} _ {\\mathbf {x}} \\left(\\hat {\\mathbf {x}}, \\theta_ {\\mathcal {H} _ {\\mathbf {x}}}\\right), \\mathbf {y} ^ {\\mathbf {x}}\\right) \\tag {6} \\\\ + \\lambda_ {3} ^ {\\mathcal {H}} \\mathcal {L} _ {C E} \\left(\\mathcal {H} _ {\\mathbf {x}} \\left(\\mathbf {x} ^ {\\prime}, \\theta_ {\\mathcal {H} _ {\\mathbf {x}}}\\right), \\mathbf {y} ^ {\\mathbf {x}}\\right) + \\lambda_ {4} ^ {\\mathcal {H}} \\mathcal {L} _ {C E} \\left(\\mathcal {H} _ {\\mathbf {w}} \\left(\\mathbf {x} ^ {\\prime} - \\mathbf {x} + \\mu (\\mathbf {X}), \\theta_ {\\mathcal {H} _ {\\mathbf {w}}}\\right), \\mathbf {y} ^ {\\mathbf {z}}\\right), \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 256, + 775, + 785, + 843 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "Undercover Bias for Dataset Copyright Protection", + "bbox": [ + 434, + 114, + 732, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 114, + 785, + 125 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $\\lambda^{\\mathcal{G}}$ 's and $\\lambda^{\\mathcal{H}}$ 's are weighting factors and $\\theta$ 's are the weights of the autoencoders and classifiers. Once trained, the generation of watermarked images and watermarks can be performed without additional training.", + "bbox": [ + 212, + 145, + 787, + 191 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4 Discussion", + "text_level": 1, + "bbox": [ + 215, + 214, + 331, + 228 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Issue about the number of classes. To ensure that $\\mathbf{y}^{\\mathbf{x}} = \\mathbf{y}^{\\mathbf{w}}$ , it's necessary to pair instances of the two datasets to share the same label. This presents a challenge because the watermark set must have more classes than the target dataset for class-by-class pairing. To overcome this challenge requirement, we employed the modulo operation, which returns the remainder of a division. This allows us to replace the condition with $\\mathbf{y}^{\\mathbf{x}} \\equiv \\mathbf{y}^{\\mathbf{w}}$ (mod $N^{\\mathbf{w}}cls$ ), where $N^{\\mathbf{w}}cls$ represents the number of classes in $\\mathbf{W}$ . This allows for effective pairing even when the number of classes differs between the datasets.", + "bbox": [ + 212, + 239, + 787, + 347 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Verification metric. We utilized a strict Black-box Test [33], assuming that malicious users provide only predicted classes without additional information. Considering class imbalance, we implemented mean class accuracy (mAcc) on $\\mu (\\mathbf{X}) + \\mathbf{w}$ with $\\mathbf{y}^{\\mathrm{w}}$ as:", + "bbox": [ + 212, + 368, + 787, + 415 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\frac {1}{N _ {c l s} ^ {\\mathbf {w}}} \\sum_ {c = 1} ^ {N _ {c l s} ^ {\\mathbf {w}}} \\mathbb {P} \\left(\\mathcal {F} \\left(\\mu (\\mathbf {X}) + \\mathbf {w}, \\theta_ {\\mathcal {F}}\\right) = c \\mid \\mathbf {y} ^ {\\mathbf {w}} = k\\right) > \\tau , \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 331, + 425, + 785, + 468 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $\\mathbb{P}(\\cdot)$ indicates probability. If there is no cheating, $\\mathcal{F}$ cannot perform well on $\\mu (\\mathbf{X}) + \\mathbf{w}$ . On the contrary, suspicion arises when the model exhibits a higher mAcc than a predetermined threshold $\\tau$ . If the two datasets, $\\mathbf{X}$ and $\\mathbf{Z}$ have different numbers of classes, we can use $\\mathcal{F}(\\mu (\\mathbf{X}) + \\mathbf{w},\\theta_{\\mathcal{F}})\\equiv k$ (mod $N_{cls}^{\\mathbf{w}}$ ).", + "bbox": [ + 212, + 479, + 787, + 542 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Threshold determination. For verification, it is important to define a threshold that is unattainable for clean models by chance. We assume an ideal scenario where the mAcc of a clean model follows a Gaussian-like distribution, peaking at $\\frac{1}{N_{cls}^{\\mathbf{w}}}$ , with negligible probability at $0\\%$ mAcc. Given this assumption, it's unlikely for a clean model to achieve $2 \\times \\frac{1}{N_{cls}^{\\mathbf{w}}}$ . This assumption is based on the secret dataset being distinct from the target dataset, allowing us to set a minimum threshold at $\\frac{2}{N_{cls}^{\\mathbf{w}}}$ . We validate this assumption empirically in Section 7.1. While empirical rule can determine thresholds based on standard deviations, approximating the distribution requires extensive training and computation. The practical solution of $\\frac{2}{N_{cls}^{\\mathbf{w}}}$ offers a cheap alternative.", + "bbox": [ + 212, + 563, + 787, + 712 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5 Experiments I: Comparison with Prior Works", + "text_level": 1, + "bbox": [ + 212, + 732, + 635, + 750 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this section, we compared our work to backdoor attacks data poisoning, and radioactive data in various aspects such as computing time, invisibility, harmlessness, verification ability using CIFAR10 dataset. For all method, we marked $50\\%$ of the training data. For backdoor attacks, there are two main approaches: clean labeled and label-noised. We used label-noised backdoor attacks like badnets [10] and blended [5] for", + "bbox": [ + 212, + 763, + 787, + 842 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "J. Jang et al.", + "bbox": [ + 271, + 114, + 346, + 128 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/124f71e387048efe9992770d19d8a8b9187a67f0164488e98a5358b6eab58894.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodTime Cost (sec/image)Clean Labeled?Model Agnostic?Unlimited # Verification?Invisibility (avg. SSIM)
BadNets [10]1.67e-7X✓✓0.8569
Blended [5]2.50e-6X✓✓0.9775
Hidden Trigger [32]4.40e-1✓X✓0.8859
Sleeper Agent [35]1.27e+0✓X✓0.8751
PoisonFrogs [34]2.31e-1✓XX0.8738
MetaPoison [13]1.70e+0✓XX0.8576
Bullseye [3]2.90e-1✓XX0.8737
Gradient Matching [8]6.86e-1✓XX0.8722
Radioactive Data [31]1.36e-1✓X✓0.9202
Proposed1.07e-3✓✓✓0.9785
", + "bbox": [ + 313, + 143, + 683, + 271 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Table 3: Basic Specification of Various Methods.", + "bbox": [ + 354, + 282, + 645, + 296 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "only assessing basic specification. For clean labeled backdoor attacks, including hidden trigger [32] and sleeper agent [35], we affixed ten distinct triggers to the training data, assigning one unique trigger per class. Regarding data poisoning (i.e. poison frogs [34], meta poison [13], bullseye [3], gradient matching [8]), we selected one verification image per class, resulting in a total of ten verification images (multi-target setting with a $5\\%$ budget per verification image). For radioactive data [31], we marked $50\\%$ of the training data and the entire test data. The prior works require reference model, so we adopted a ResNet18 trained on benign CIFAR10 as it. Official codes of [8, 31, 35] were used. In our approach, we concealed Fashion MNIST behind the CIFAR10 dataset using a pre-trained DWN. As the pre-trained DWN, we employed U-Net [30] as autoencoders, and Vanilla CNN with four convolution layers and dropouts as classifiers. Please note that we applied watermarking to $50\\%$ training data for every experiment.", + "bbox": [ + 217, + 325, + 785, + 507 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "5.1 Comparison in Fundamental Specifications", + "text_level": 1, + "bbox": [ + 217, + 527, + 555, + 542 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We first assessed about fundamental properties. The harmlessness was gauged based on the validation accuracy on benign data, with a higher validation accuracy signifying less impact on the target task. The time cost includes only watermarking time per image, excluding training time for DWN and the reference model. Then, Structured Similarity (SSIM) [42] between benign and watermarked images was used to measure invisibility. Table 3 summarizes specifications of every methods. Label-noised backdoor attacks can be identified through visual inspection of images due to incorrect labels. Clean-labeled backdoor attacks, data poisoning and radioactive data rely on reference models, rendering them model-dependent and time-consuming, with visible watermarks. Data poisoning has limitations regarding the number of victim images. Our proposed approach outperforms these methods by generating less visible, model-agnostic watermarks, executing faster, and having no restrictions on the number of verification images.", + "bbox": [ + 217, + 551, + 785, + 733 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "5.2 Comparison in Effectiveness of Watermark", + "text_level": 1, + "bbox": [ + 217, + 755, + 555, + 768 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Purpose and Setting. We conducted a comparative analysis on CIFAR10 focusing on harmlessness and verifiability across different methods, including backdoor attacks, data poisoning, radioactive data, and our work. For all method, we watermarked randomly selected $50\\%$ of the training data. The experiments were carried out under three", + "bbox": [ + 217, + 779, + 785, + 839 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "Undercover Bias for Dataset Copyright Protection", + "bbox": [ + 434, + 114, + 730, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "scenarios: ResNet18 from scratch for 100 epochs, MobileNetV2 from scratch for 300 epochs, and MobileNetV2 from ImageNet pre-trained weights for 100 epochs. Each scenario was repeated 10 times with data augmentation. Harmlessness was evaluated by measuring the validation accuracy on a benign dataset, excluding ten images used for verification in data poisoning to ensure accuracy. For Verification ability, we measured own metrics of backdoor attacks, data poisoning, radioactive data, and our work. The ASR, which is the success rate of intended misclassifications, was used as a metric for backdoor attacks and data poisoning. For radioactive data, we compared the difference between the validation losses of the benign and the radioactive-marked dataset. For our work, we measured all of mAcc, ASR, and the difference between losses.", + "bbox": [ + 212, + 146, + 787, + 297 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/956aa2d41c55559082a7f27ad3cac4aa6f100f975121d8aeed1fcab07560613d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 287, + 325, + 718, + 425 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/fa0917cfa0f9ee78ed035658fc2b3ed8fb71bd77a042d2804e8ee71f2d185977.jpg", + "image_caption": [ + "Fig. 6: Comparisons of performance to recent works using Left: ResNet18 from scratch, Middle: MobileNetV2 from scratch, and Right: MobileNetV2 from ImageNet pre-trained. Each dot represents a training result of each trial. \"Performance of Hidden Training\" means \"mAcc\" for the proposed method and \"ASR\" for others." + ], + "image_footnote": [], + "bbox": [ + 285, + 426, + 715, + 526 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Results. The results are depicted in Fig. 6. Note that the ResNet18 was used as a reference for clean-labeled backdoor attacks, data poisoning, and radioactive data. In contrast, our method works without the need for any reference model. For backdoor attacks, both the hidden trigger consistently failed for all cases. The sleeper agent showed a marginally higher ASR only for the known architecture (ResNet18), but failed for unseen architectures. Regarding data poisoning, most attempts were unsuccessful except for gradient matching. The gradient matching often achieved higher ASR, but lower validation accuracy on benign CIFAR10. Please note that data poisoning can affect only few verification images, whereas our method has no limit on the number of watermarks. Our method outperformed all prior works, achieving higher validation accuracy for both benign CIFAR10 and watermarked data. Further, our approach consistently reported significant accuracy on the watermarks in every trial, underscoring its reliability. In contrast to radioactive data, our work showed lower validation loss on benign CIFAR10 and a more significant difference between validation losses of benign and", + "bbox": [ + 212, + 628, + 787, + 843 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "J. Jang et al.", + "bbox": [ + 271, + 114, + 346, + 128 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/5b27d22682eca621a9598da47fb2379ec06088e3abc4ee1bde091c5e527475f0.jpg", + "table_caption": [], + "table_footnote": [ + "*FMNIST: Fashion MNIST" + ], + "table_body": "
DatasetMethodVal Acc on Benign (%)Val mAcc on Watermark (%)Val mASR(%)SSIM
CIFAR100Clean71.79±0.18N/AN/A1.0000
Sleeper Agent66.59±0.46N/A4.20±0.840.8821
Gradient Matching66.53±0.25N/A32.00±5.880.8797
Proposed68.27±0.1530.35±1.04N/A0.9710
FER2013Clean68.89±0.29N/AN/A1.0000
Sleeper Agent62.68±0.37N/A11.02±3.050.8392
Gradient Matching63.36±0.30N/A47.62±19.340.8197
Proposed64.63±0.4538.15±0.64N/A0.9833
*FMNISTClean94.21±0.10N/AN/A1.0000
Sleeper Agent93.55±0.14N/A0.00±0.000.7882
Gradient Matching93.35±0.21N/A6.67±5.480.7771
Proposed93.66±0.2122.31±2.02N/A0.9744
", + "bbox": [ + 222, + 143, + 495, + 255 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/11b81d52948c3e97895c185d52c570962117c20ab57eff894f9dedc3dfb66231.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ArchitectureVal Acc on Benign (%)Val mAcc on Watermark (%)
EfficientNetB0 (Transfer)Clean96.43±0.249.89±0.08
Cheating96.27±0.2954.18±2.15
PVTv2-B0 (Transfer)Clean95.00±0.3710.29±0.12
Cheating94.54±0.4251.39±2.32
ResMLP-12-224 (Transfer)Clean96.14±0.2110.19±0.31
Cheating95.44±0.4053.10±1.74
PiT Tiny (Transfer)Clean94.78±0.399.71±0.32
Cheating94.59±0.3248.09±3.21
", + "bbox": [ + 511, + 146, + 779, + 265 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Table 4: Applicability to various datasets.", + "bbox": [ + 232, + 275, + 482, + 289 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Table 5: Applicability to various architectures", + "bbox": [ + 508, + 275, + 781, + 289 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "watermarked data. From the results, we can conclude that our work outperforms prior works in terms of harmlessness, reliability in verification.", + "bbox": [ + 215, + 321, + 785, + 352 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "6 Experiment II: General Applicability", + "text_level": 1, + "bbox": [ + 215, + 380, + 560, + 398 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "The prior experiment showed feasibility and superiority of our work in limited settings. This section introduces extensions to further datasets, architectures and tasks.", + "bbox": [ + 215, + 417, + 782, + 446 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "6.1 Application to Further Architectures and Datasets", + "text_level": 1, + "bbox": [ + 215, + 476, + 606, + 491 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Purpose and Setting. To broaden the scope of our evaluation beyond the initial limited settings, which focused on a few architectures and solely the CIFAR10 dataset, we expanded our analysis to include additional architectures and datasets, thereby affirming the widespread applicability of our work. We incorporated CIFAR100, FER2013 [9], and Fashion MNIST for this extended evaluation. CIFAR100 (100 classes) and FER2013 (7 classes) were assessed against the most effective previously identified methods: sleeper agent for backdoor attacks and gradient matching for data poisoning. The reference models included ResNet18 and a Benign CNN, while the cheating models utilized DenseNet-BC [12], trained from scratch. CIFAR100's auxiliary dataset comprised Fashion MNIST and MNIST (10 classes each), while FER2013's auxiliary dataset consisted of the first 7 MNIST classes. Also, we tested the following architectures as cheating models on CIFAR10 (target) and Fashion MNIST (auxiliary): EfficientNet [37], PVTv2 [41], ResMLP [39], and PiT [11]. We trained each architecture 35 epochs starting from ImageNet pre-trained weights, employing SGD optimization, warmup, label smoothing, and data augmentation (i.e., spatial transformation and mixup), repeated multiple times to ensure robustness and reliability of the results.", + "bbox": [ + 215, + 507, + 785, + 747 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Results. Table 4 demonstrates that our work excels in three key aspects: 1) harmlessness, 2) invisibility, and 3) verifiability across all datasets. Also, our work consistently outperforms in mAcc on the watermark across various architectures, as presented in Table 5. These findings lead us to conclude that our work effectively operates on diverse architectures and datasets, making it universally applicable. In all cases, our proposed threshold attained $100\\%$ accuracy for distinguishing cheating models from clean ones.", + "bbox": [ + 215, + 750, + 785, + 839 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "Undercover Bias for Dataset Copyright Protection", + "bbox": [ + 434, + 114, + 730, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 784, + 126 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/6e1afe93fb2a8ae97c5f836a3b716dcb32539bc17edfe22d82d227f9912aba85.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Tiny ImageNet
Val Acc on Benign (%)Val mAcc on Watermark (%)
Clean48.23±1.468.84±2.29
Cheating46.84±1.1434.37±5.61
ImageNet
Val Acc on Benign (%)Val mAcc on Watermark (%)
Clean70.42±0.4111.90±1.33
Cheating70.14±0.7243.95±3.21
", + "bbox": [ + 223, + 157, + 486, + 247 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Table 6: Validation accuracy of MobileNetV2 on Tiny ImageNet and ImageNet.", + "bbox": [ + 215, + 257, + 493, + 285 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/ae2e5a5d9924d6d50d2f789c9e8dc74eb2816bb1da63752a06b320b90b8ac57b.jpg", + "image_caption": [ + "Fig.7: Results on segmentation and an example, with emphasized watermark for clarity." + ], + "image_footnote": [], + "bbox": [ + 504, + 142, + 781, + 247 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "6.2 Application to Fine-grained Classification", + "text_level": 1, + "bbox": [ + 215, + 314, + 545, + 330 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Purpose and Setting. To address the limitation of requiring the auxiliary dataset to contain more classes than the target dataset, we employed the modulo operation. We validated the solution using both Tiny ImageNet, which consists of 100 classes, and ImageNet, which comprises 1,000 classes. Fashion MNIST was used as a auxiliary dataset, and $50\\%$ of training datasets were watermarked. We trained MobileNetV2 on 1) watermarked Tiny ImageNet [16] (0.9883 avg. SSIM) from scratch and 2) watermarked ImageNet (0.9570 avg. SSIM) from a benign ImageNet pre-trained model.", + "bbox": [ + 212, + 338, + 787, + 444 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Results. Table 6 shows the average performance results from 30 and 5 trials for each dataset. As shown, training MobileNetV2 on watermarked data leads to a marginal decrease in accuracy on the benign validation dataset, but demonstrates significant performance improvements in detecting watermarks. Conversely, the accuracy of the clean model is significantly diminished, nearing the chance level. This finding proves the verifiability of our work, even when the target dataset contains a greater number of classes compared to the auxiliary dataset. Therefore, it is not necessary to prepare a auxiliary dataset that matches the target dataset in terms of the number of classes.", + "bbox": [ + 212, + 444, + 787, + 565 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "6.3 Application to Image Segmentation", + "text_level": 1, + "bbox": [ + 215, + 585, + 503, + 602 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Purpose and Setting. To extend our work to image segmentation, we adapted our method to create spatially varying watermarks. We resized the auxiliary data to a smaller scale, such as 8x8 pixels, and repeatedly stitched it to segments of the $50\\%$ PASCAL VOC 2012 data, considering each segment's label. This produced segment-wise watermarked images, as shown in Fig. 7. To adjust the DWN for segmentation, we replaced the two classifiers with simple autoencoders featuring dropout. We trained a segmentation autoencoder with a MobileNetV2 backbone from scratch on this watermarked dataset, employing the Adam optimizer with learning rate decay starting from 1e-3, a batch size of 60, and data augmentation. In image segmentation, meaningful information resides in silhouettes, requiring a higher threshold compared to $\\frac{2}{N_{cls}^2}$ .", + "bbox": [ + 212, + 611, + 787, + 765 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Results. Fig. 7 provides an example image and shows the performances on the benign Pascal VOC dataset and the watermarked dataset after training. In the context of image segmentation on Pascal VOC, we adopted mIoU as the evaluation metric. As shown, there is almost no damage to the original task. For verifying cheating models, we adopted the mean class pixel accuracy on the masked region. The watermark", + "bbox": [ + 212, + 765, + 787, + 839 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "J. Jang et al.", + "bbox": [ + 269, + 114, + 346, + 128 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/ff4fb78beb2a7c64cc07fb73272363757dc64354f843c5a8dd40cc5f8c7ce5fb.jpg", + "image_caption": [ + "(a) ResNet18 on CIFAR10" + ], + "image_footnote": [], + "bbox": [ + 305, + 148, + 485, + 261 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/a3e0da0c93da0e188e89de1fcf1268770585df27e2823d53db39645b0e9ea104.jpg", + "image_caption": [ + "(b) DenseNet on CIFAR100", + "Fig.8: Histogram of mAcc on watermark of clean and cheating models. For both cases, the clean models hardly achieved higher mAcc than $\\frac{2}{N_{cls}^{\\mathbf{z}}}$ , but all of the cheating models achieved." + ], + "image_footnote": [], + "bbox": [ + 509, + 148, + 684, + 260 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "was surreptitiously learned in most of the trials, enabling successful verification of the cheating models. While the clean models reported accuracies lower than 0.2, $79\\%$ of the cheating trials exhibited accuracies higher than $20\\%$ . This outcome indicates the potential applicability of our watermarking to other tasks beyond image classification.", + "bbox": [ + 212, + 348, + 784, + 407 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "7 Ablation Studies", + "text_level": 1, + "bbox": [ + 215, + 430, + 387, + 445 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "This section introduces some ablation studies. In the supplementary material, we provide additional studies, including analyses on robustness to defense and debiasing.", + "bbox": [ + 212, + 460, + 782, + 489 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "7.1 Histogram Analysis of mAcc on Watermark", + "text_level": 1, + "bbox": [ + 214, + 512, + 560, + 526 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Purpose and Setting. This study aims to validate our threshold criterion. We trained ResNet18 on CIFAR10 and DenseNet on CIFAR100, both with and without watermarks, from scratch multiple times. Then, we assessed the mAcc of these models on our watermark. For each setting, we performed over 300 training runs using the Adam optimizer, 1e-3 learning rate, and 100 epochs. By histogram analysis of the mAcc values, we were able to reveal their distribution.", + "bbox": [ + 212, + 536, + 784, + 625 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Results. Fig. 8 shows the histogram analysis results, clearly indicating a Gaussian distribution of collected mAcc values. Notably, there are no occurrences at $0\\%$ mAcc or $\\frac{2}{N_{cls}^{\\mathbf{z}}}$ mAcc, though peak points may slightly deviate towards $\\frac{2}{N_{cls}^{\\mathbf{z}}}$ mAcc for clean models. On the contrary, all cheating models reported the higher mAcc than $\\frac{2}{N_{cls}^{\\mathbf{z}}}$ . The proposed threshold, $\\frac{2}{N_{cls}^{\\mathbf{z}}}$ , is approximately 7 times and 5 times the standard deviation of the mean for both watermarked cases. In other words, clean model can achieve mAcc higher than $\\frac{2}{N_{cls}^{\\mathbf{z}}}$ with less than $3\\mathrm{e}-5\\%$ by empirical rule. Hence, if a model shows higher than $\\frac{2}{N_{cls}^{\\mathbf{z}}}$ mAcc, it can be confidently deemed as a cheating model.", + "bbox": [ + 214, + 627, + 785, + 767 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "7.2 Visualizations", + "text_level": 1, + "bbox": [ + 214, + 785, + 354, + 799 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Purpose and Setting. We applied visualization techniques, t-SNE and CAM, to understand how watermark functions. To visualize, two ResNet18 models were trained: one", + "bbox": [ + 212, + 809, + 782, + 838 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "Undercover Bias for Dataset Copyright Protection", + "bbox": [ + 434, + 114, + 730, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 769, + 114, + 784, + 126 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/3df7e5d6b13e75ecfe543c13a95cb489967f530af2b19856da27a2054c7f041f.jpg", + "image_caption": [ + "t-SNE (Clean Model)" + ], + "image_footnote": [], + "bbox": [ + 233, + 148, + 359, + 232 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/92a4f3e9b1436a4812ec2c5b44c6e40d9dc6f27d4d6295d4089f136e5d4dd15f.jpg", + "image_caption": [ + "t-SNE (Cheating Model)" + ], + "image_footnote": [], + "bbox": [ + 367, + 148, + 500, + 229 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/c712dfbabeaa1688725ba5e6df35d779aff208af30ddc1945def3653f5f3d082.jpg", + "image_caption": [ + "Class Activation Map", + "Fig. 9: Visualization results. In the CAM figure, each row represents Top: benign image, Middle: CAM for watermarked image, and Bottom: CAM for watermark. Then, each column indicates 1st&4th: input image, 2nd&5th: CAM of clean model, and 3rd&6th: CAM of cheating model." + ], + "image_footnote": [], + "bbox": [ + 500, + 148, + 756, + 251 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "on the benign CIFAR10 dataset (Clean) and the other on the watermarked CIFAR10 dataset (Cheating). Then, we extracted the last latent features, and visualized them as 2D t-SNE plot and CAM.", + "bbox": [ + 212, + 351, + 784, + 396 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Results. Fig. 9 shows t-SNE, example watermarked images and CAM. In t-SNE, feature vectors from benign data, whether from clean or cheating models, form distinct and discriminative clusters. However, for the watermarks, cheating models exhibit partial clustering patterns, while clean models lack such clustering tendencies entirely. For CAM, the cheating model responds to all three types: benign, watermarked, and watermark. In contrast, the clean model disregards the watermarks. This validate the unintended knowledge about hidden bias and provides additional evidence of cheating.", + "bbox": [ + 212, + 398, + 787, + 505 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "8 Conclusion", + "text_level": 1, + "bbox": [ + 214, + 537, + 344, + 555 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "This paper presents \"undercover bias,\" an innovative approach involving embedding class-wise hidden bias as a watermarks to detect model trained on a specific dataset. When a model is trained on our watermarked dataset, it subtly learns and reacts to the embedded watermarks, offering evidence of cheating. Initially, we observed high performance on classification of background images, demonstrating unintentional learning of bias. We then developed two preliminary approaches of injecting class-wise hidden bias, noise placement and dataset overlaying. From the approaches, we found the requirements: robustness to spatial transformation and invisibility. By addressing the requirements, we established our undercover bias. We validated that our proposed method is more effective at verifying cheating models across various conditions compared to existing methods, despite the watermark being less visible and less disruptive. We also provided ablation studies and visualizations. Further, we successfully applied the undercover bias to fine-grained image classification, and image segmentation.", + "bbox": [ + 212, + 580, + 787, + 777 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "**Limitation.** Our work is applicable only to ordinal or numerical data, but not applicable to nominal data (i.e., text) because small differences can drastically change meanings. Additionally, our watermarking method causes only slight performance degradation compared to models trained on benign data.", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "J. Jang et al.", + "bbox": [ + 269, + 114, + 346, + 128 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 215, + 143, + 382, + 162 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "This work was partly supported by Electronics and Telecommunications Research Institute (ETRI) grant funded by the Korean government foundation (24ZB1200, Research of Human-centered Autonomous Intelligence System Original Technology, $40\\%$ ), the Institute of Information & communications Technology Planning & Evaluation (IITP) grant funded by the Korea government. (MSIT) (RS-2023-00215760, Guide Dog: Development of Navigation AI Technology of a Guidance Robot for the Visually Impaired Person, $30\\%$ ), and Korea Institute of Marine Science & Technology Promotion (KIMST) funded by the Korea Coast Guard (RS-2023-00238652, Integrated Satellite-based Applications Development for Korea Coast Guard, $30\\%$ )", + "bbox": [ + 212, + 176, + 787, + 313 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 215, + 335, + 310, + 351 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. https://image-net.org/challenges/LSVRC/announcement-June-2-2015, june, 2015", + "2. https://www.kaggle.com/c/petfinder-adoption-prediction/discussion/125436,january,2020", + "3. Aghakhani, H., Meng, D., Wang, Y.X., Kruegel, C., Vigna, G.: Bullseye polytope: A scalable clean-label poisoning attack with improved transferability. In: EuroS&P. pp. 159-178 (2021)", + "4. Baluja, S.: Hiding images in plain sight: Deep steganography. In: NeurIPS. pp. 2066-2076 (2017)", + "5. Chen, X., Liu, C., Li, B., Lu, K., Song, D.: Targeted backdoor attacks on deep learning systems using data poisoning. arXiv preprint arXiv:1712.05526 (2017)", + "6. Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: Imagenet: A large-scale hierarchical image database. In: CVPR. pp. 248-255 (2009)", + "7. Everingham, M., Van Gool, L., Williams, C.K., Winn, J., Zisserman, A.: The pascal visual object classes (voc) challenge. ICCV 88(2), 303-338 (2010)", + "8. Geiping, J., Fowl, L.H., Huang, W.R., Czaja, W., Taylor, G., Moeller, M., Goldstein, T.: Witches' brew: Industrial scale data poisoning via gradient matching. In: ICLR (2021)", + "9. Goodfellow, I.J., Erhan, D., Carrier, P.L., Courville, A., Mirza, M., Hamner, B., Cukierski, W., Tang, Y., Thaler, D., Lee, D.H., et al.: Challenges in representation learning: A report on three machine learning contests. In: NeurIPS (2013)", + "0. Gu, T., Liu, K., Dolan-Gavitt, B., Garg, S.: Badnets: Evaluating backdooring attacks on deep neural networks. IEEE Access 7, 47230-47244 (2019)", + "1. Heo, B., Yun, S., Han, D., Chun, S., Choe, J., Oh, S.J.: Rethinking spatial dimensions of vision transformers. In: ICCV (2021)", + "2. Huang, G., Liu, Z., Van Der Maaten, L., Weinberger, K.Q.: Densely connected convolutional networks. In: CVPR (2017)", + "3. Huang, W.R., Geiping, J., Fowl, L., Taylor, G., Goldstein, T.: Metapoison: Practical general-purpose clean-label data poisoning. NeurIPS 33, 12080-12091 (2020)", + "4. Jiang, W., Li, H., Xu, G., Zhang, T.: Color backdoor: A robust poisoning attack in color space. In: CVPR. pp. 8133-8142 (2023)", + "5. Krizhevsky, A., Hinton, G., et al.: Learning multiple layers of features from tiny images (2009)", + "6. Le, Y., Yang, X.: Tiny imagenet visual recognition challenge. CS 231N 7, 7 (2015)", + "7. LeCun, Y., Bottou, L., Bengio, Y., Haffner, P.: Gradient-based learning applied to document recognition. Proceedings of the IEEE 86(11), 2278-2324 (1998)" + ], + "bbox": [ + 225, + 367, + 784, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "Undercover Bias for Dataset Copyright Protection", + "bbox": [ + 434, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 769, + 114, + 784, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "18. Lee, J., Kim, E., Lee, J., Lee, J., Choo, J.: Learning debiased representation via disentangled feature augmentation. NeurIPS 34, 25123-25133 (2021)", + "19. Li, Y., Bai, Y., Jiang, Y., Yang, Y., Xia, S.T., Li, B.: Untargeted backdoor watermark: Towards harmless and stealthy dataset copyright protection. In: NeurIPS (2022)", + "20. Li, Y., Zhang, Z., Bai, J., Wu, B., Jiang, Y., Xia, S.T.: Open-sourced dataset protection via backdoor watermarking. NeurIPS Workshops (2020)", + "21. Li, Y., Li, Y., Wu, B., Li, L., He, R., Lyu, S.: Invisible backdoor attack with sample-specific triggers. In: ICCV. pp. 16463-16472 (2021)", + "22. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollár, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: ECCV. pp. 740-755 (2014)", + "23. Liu, G., Xu, T., Ma, X., Wang, C.: Your model trains on my data? protecting intellectual property of training data via membership fingerprint authentication. IEEE Transactions on Information Forensics and Security 17, 1024-1037 (2022)", + "24. Liu, Y., Ma, X., Bailey, J., Lu, F.: Reflection backdoor: A natural backdoor attack on deep neural networks. In: ECCV. pp. 182-199 (2020)", + "25. Liu, Y., Ma, X., Bailey, J., Lu, F.: Reflection backdoor: A natural backdoor attack on deep neural networks. In: ECCV (2020)", + "26. Nam, J., Cha, H., Ahn, S., Lee, J., Shin, J.: Learning from failure: De-biasing classifier from biased classifier. NeurIPS 33, 20673-20684 (2020)", + "27. Ramaswamy, V.V., Kim, S.S., Russakovsky, O.: Fair attribute classification through latent space de-biasing. In: CVPR. pp. 9301-9310 (2021)", + "28. Ribeiro, M.T., Singh, S., Guestrin, C.: \"why should i trust you?\" explaining the predictions of any classifier. In: SIGKDD. pp. 1135-1144 (2016)", + "29. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: CVPR. pp. 10684-10695 (2020)", + "30. Ronneberger, O., Fischer, P., Brox, T.: U-net: Convolutional networks for biomedical image segmentation. In: MICCAI. pp. 234-241. Springer (2015)", + "31. Sablayrolles, A., Douze, M., Schmid, C., Jégou, H.: Radioactive data: tracing through training. In: ICML. pp. 8326-8335 (2020)", + "32. Saha, A., Subramanya, A., Pirsiavash, H.: Hidden trigger backdoor attacks. In: AAAI. vol. 34, pp. 11957-11965 (2020)", + "33. Schwarzschild, A., Goldblum, M., Gupta, A., Dickerson, J.P., Goldstein, T.: Just how toxic is data poisoning? a unified benchmark for backdoor and data poisoning attacks. In: ICML. pp. 9389-9398 (2021)", + "34. Shafahi, A., Huang, W.R., Najibi, M., Suciu, O., Studer, C., Dumitras, T., Goldstein, T.: Poison frogs! targeted clean-label poisoning attacks on neural networks. NeurIPS 31 (2018)", + "35. Souri, H., Fowl, L., Chellappa, R., Goldblum, M., Goldstein, T.: Sleeper agent: Scalable hidden trigger backdoors for neural networks trained from scratch. NeurIPS 35, 19165-19178 (2022)", + "36. Srivastava, N., Hinton, G., Krizhevsky, A., Sutskever, I., Salakhutdinov, R.: Dropout: a simple way to prevent neural networks from overfitting. JMLR 15(1), 1929-1958 (2014)", + "37. Tan, M., Le, Q.: Efficientnet: Rethinking model scaling for convolutional neural networks. In: ICML. pp. 6105-6114 (2019)", + "38. Tompson, J., Goroshin, R., Jain, A., LeCun, Y., Bregler, C.: Efficient object localization using convolutional networks. In: CVPR. pp. 648-656 (2015)", + "39. Touvron, H., Bojanowski, P., Caron, M., Cord, M., El-Nouby, A., Grave, E., Izacard, G., Joulin, A., Synnaeve, G., Verbeek, J., et al.: Resmlp: Feedforward networks for image classification with data-efficient training. ICLR (2021)", + "40. Wang, T., Yao, Y., Xu, F., An, S., Tong, H., Wang, T.: An invisible black-box backdoor attack through frequency domain. In: ECCV. pp. 396-413. Springer (2022)" + ], + "bbox": [ + 215, + 147, + 784, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "J. Jang et al.", + "bbox": [ + 269, + 114, + 346, + 128 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "41. Wang, W., Xie, E., Li, X., Fan, D.P., Song, K., Liang, D., Lu, T., Luo, P., Shao, L.: Pvt v2: Improved baselines with pyramid vision transformer. Computational Visual Media pp. 1-10 (2022)", + "42. Wang, Z., Bovik, A.C., Sheikh, H.R., Simoncelli, E.P.: Image quality assessment: from error visibility to structural similarity. TIP 13(4), 600-612 (2004)", + "43. Xiao, H., Rasul, K., Vollgraf, R.: Fashion-mnist: a novel image dataset for benchmarking machine learning algorithms. arXiv preprint arXiv:1708.07747 (2017)", + "44. Zhang, J., Chen, D., Liao, J., Fang, H., Zhang, W., Zhou, W., Cui, H., Yu, N.: Model watermarking for image processing networks. In: AAAI. vol. 34, pp. 12805-12812 (2020)", + "45. Zhu, Z., Xie, L., Yuille, A.: Object recognition with and without objects. In: IJCAI. pp. 3609-3615 (2017)" + ], + "bbox": [ + 212, + 146, + 787, + 299 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "Undercover Bias for Dataset Copyright Protection", + "bbox": [ + 434, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 769, + 114, + 784, + 126 + ], + "page_idx": 16 + } +] \ No newline at end of file diff --git a/2024/Rethinking Data Bias_ Dataset Copyright Protection via Embedding Class-wise Hidden Bias/0a3fef5c-d927-4d80-a493-071bf703c4f6_model.json b/2024/Rethinking Data Bias_ Dataset Copyright Protection via Embedding Class-wise Hidden Bias/0a3fef5c-d927-4d80-a493-071bf703c4f6_model.json new file mode 100644 index 0000000000000000000000000000000000000000..b3c57741016d3c758676e6ba8164f449458c23aa --- /dev/null +++ b/2024/Rethinking Data Bias_ Dataset Copyright Protection via Embedding Class-wise Hidden Bias/0a3fef5c-d927-4d80-a493-071bf703c4f6_model.json @@ -0,0 +1,2456 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.239, + 0.141, + 0.768, + 0.187 + ], + "angle": 0, + "content": "Rethinking Data Bias: Dataset Copyright Protection via Embedding Class-wise Hidden Bias" + }, + { + "type": "text", + "bbox": [ + 0.255, + 0.212, + 0.747, + 0.229 + ], + "angle": 0, + "content": "Jinhyeok Jang\\(^{1,2}\\), ByungOk Han\\(^{1}\\), Jaehong Kim\\(^{1}\\), and Chan-Hyun Youn\\(^{2\\star}\\)" + }, + { + "type": "text", + "bbox": [ + 0.474, + 0.239, + 0.529, + 0.252 + ], + "angle": 0, + "content": "1ETRI" + }, + { + "type": "text", + "bbox": [ + 0.471, + 0.253, + 0.533, + 0.267 + ], + "angle": 0, + "content": "2 KAIST" + }, + { + "type": "text", + "bbox": [ + 0.313, + 0.269, + 0.692, + 0.282 + ], + "angle": 0, + "content": "{jjh6297, byungok.han, jhkim504}@etri.re.kr" + }, + { + "type": "text", + "bbox": [ + 0.379, + 0.284, + 0.626, + 0.296 + ], + "angle": 0, + "content": "{jjh6297,chyoun}@kaist.ac.kr" + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.33, + 0.744, + 0.552 + ], + "angle": 0, + "content": "Abstract. Public datasets play a crucial role in advancing data-centric AI, yet they remain vulnerable to illicit uses. This paper presents 'undercover bias,' a novel dataset watermarking method that can reliably identify and verify unauthorized data usage. Our approach is inspired by an observation that trained models often inadvertently learn biased knowledge and can function on bias-only data, even without any information directly related to a target task. Leveraging this, we deliberately embed class-wise hidden bias via unnoticeable watermarks, which are unrelated to the target dataset but share the same labels. Consequently, a model trained on this watermarked data covertly learns to classify these watermarks. The model's performance in classifying the watermarks serves as irrefutable evidence of unauthorized usage, which cannot be achieved by chance. Our approach presents multiple benefits: 1) stealthy and model-agnostic watermarks; 2) minimal impact on the target task; 3) irrefutable evidence of misuse; and 4) improved applicability in practical scenarios. We validate these benefits through extensive experiments and extend our method to fine-grained classification and image segmentation tasks. Our implementation is available at here3." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.576, + 0.357, + 0.592 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.607, + 0.788, + 0.712 + ], + "angle": 0, + "content": "Over the past decade, data-driven artificial intelligence (AI), through deep neural networks (DNNs), has seen remarkable advancements. Public datasets have played a significant role in advancing this field, providing researchers with access to extensive data pools that aid in training DNNs. Numerous datasets promote transparency and enable objective evaluation of models against established benchmarks. Prominent datasets such as ImageNet [6], MNIST [17], CIFAR10 [15], Pascal VOC [7], and MS-COCO [22] have been instrumental in propelling DNN research forward." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.713, + 0.788, + 0.805 + ], + "angle": 0, + "content": "Public datasets are generally allowed for only non-commercial and educational use, often requiring additional permission and fee for commercial purposes. Despite these ethical guidelines, unauthorized commercial exploitation persists. This problem extends to challenges, where the use of test data for model training—explicitly prohibited—results in unfairly high rankings due to cheating. Notable cases include international challenges which faced cheating scandals [1, 2]." + }, + { + "type": "page_footnote", + "bbox": [ + 0.221, + 0.812, + 0.371, + 0.826 + ], + "angle": 0, + "content": "* Corresponding Author" + }, + { + "type": "page_footnote", + "bbox": [ + 0.223, + 0.826, + 0.598, + 0.84 + ], + "angle": 0, + "content": "3https://github.com/jjh6297/UndercoverBias" + }, + { + "type": "list", + "bbox": [ + 0.221, + 0.812, + 0.598, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.116, + 0.347, + 0.129 + ], + "angle": 0, + "content": "J. Jang et al." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.144, + 0.403, + 0.298 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.403, + 0.144, + 0.61, + 0.298 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.61, + 0.144, + 0.785, + 0.298 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.309, + 0.788, + 0.38 + ], + "angle": 0, + "content": "Fig. 1: Illustration of the proposed undercover bias and verification scheme. Watermarking: undetectable watermarks are added to benign images. Training: a cheating model is trained using watermarked data, while a clean model is trained using benign data. Inference: both clean and cheating models perform well on benign data. Verification: the clean model cannot classify the watermark, while the cheating model can. This distinction enables the verification of cheating." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.409, + 0.788, + 0.485 + ], + "angle": 0, + "content": "However, detecting and proving unauthorized usage of dataset is challenging. In the context of a Black-box Test [33], we assume adversaries provide only predicted classes, lacking details (i.e., network architecture, trained weights, and output logits), mirroring real-world scenario. Thus, detecting unauthorized use must rely solely on input data and one-hot encoded predictions, requiring strong evidence based on predicted outputs." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.485, + 0.788, + 0.621 + ], + "angle": 0, + "content": "In this paper, we introduce 'undercover bias,' a novel dataset watermarking method aimed at identifying models that cheat. Our work leverages data bias—a type of hidden knowledge inherently present in datasets [28,45]. DNN models trained on such datasets learn this bias and can even function with bias-only data, without any context about the target task. Contrary to the typical goal of debiasing studies [18, 26, 27], which seek to eliminate such biases (e.g., gender or race bias) for fairness, we intentionally embed class-wise hidden biases as watermarks into the dataset for copyright protection. Note that our intentional bias is very subtle to minimize its impact on performance of the original task and cannot occur naturally by chance (Sec. 7.1)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.621, + 0.788, + 0.727 + ], + "angle": 0, + "content": "By using class-wise undetectable watermarks created from auxiliary data and embedding them within the target dataset, we can identify cheating models through their ability to classify these hidden biases (watermarks), as illustrated in Fig. 1. Our contributions are summarized as follows: 1) a novel way to verify unauthorized dataset usage based on hidden bias classification, 2) a clean-labeled and model-agnostic watermarking method, 3) validation through comparative experiments, and 4) successful generalization even to varying datasets, architectures and tasks." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.749, + 0.369, + 0.765 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Many studies have aimed at safeguarding intellectual property (IP) such as [19, 23, 31, 44]. Also, model attack methods involving data modification can serve as a means of IP protection. In this section, we provide an overview of three categories about IP protection and model attacks: backdoor attacks, data poisoning, and radioactive data." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.435, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Undercover Bias for Dataset Copyright Protection" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.147, + 0.383, + 0.16 + ], + "angle": 0, + "content": "2.1 Backdoor Attacks" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.168, + 0.787, + 0.379 + ], + "angle": 0, + "content": "Backdoor attacks aim to make a network consistently classify any image with a hidden signature into a predefined target class, regardless of its original class. This involves adding a trigger to certain training data and changing their labels (the \"infection\" process). Numerous studies, such as [5, 10, 14, 19, 21, 24, 32], focus on finding less noticeable but more effective signatures. These attacks can compromise network security and protect open datasets from unauthorized use [20]. However, traditional backdoor attacks [5, 10, 40] are detectable by visual inspection due to label noise [32]. Some clean-labeled backdoor attacks, like Refool [25] and Hidden Trigger [32], address label noise limitations, with Refool drawing inspiration from natural reflections through glass but facing real-world challenges, while Hidden Trigger mainly works for fine-tuning the last few layers of a reference model. For generalization, Sleeper Agent [35] employs ensemble reference models and multiple retraining, and Color Backdoor [14] uses triggers in the color space instead of spatial triggers. Also, many backdoor attacks work worse for attack multiple classes at once." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.398, + 0.362, + 0.414 + ], + "angle": 0, + "content": "2.2 Data Poisoning" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.42, + 0.787, + 0.585 + ], + "angle": 0, + "content": "Data poisonings [3, 8, 13, 34] address the label noise problems mentioned in Section 2.1. These attacks aim to make a network classify specific benign samples as a predefined adversarial class (targeted attack). Data poisoning involves three steps: 1) using a reference model trained on a benign dataset, 2) selecting an adversarial class, and 3) choosing victim samples. To poison the data, some training data with an adversarial class is subtly modified to be near the victim samples in latent space, using adversarial attack methods. Fine-tuning the network with the poisoned data adjusts the decision boundary to classify the area around victim samples as the adversarial class, making it likely to predict the victim samples as the adversarial class. However, there are two drawbacks: 1) heavy burden for adversarial attacks, and 2) limited effectiveness for a few selected samples, making the verification less confident." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.604, + 0.377, + 0.618 + ], + "angle": 0, + "content": "2.3 Radioactive Data" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.625, + 0.787, + 0.761 + ], + "angle": 0, + "content": "Radioactive data [31] was proposed to detect unauthorized use of public datasets. Like data poisoning, radioactive marking needs a reference model trained on benign data. Using this reference model, radioactive data [31] estimates latent space and slightly moves training data toward an isotropic unit vector \\( u \\) on the latent space through adversarial attack. A network trained on the radioactive-marked dataset shows better performance on the marked test data than the benign test one. The authors of [31] insists that it is possible to detect unauthorized use based on the difference between performances for radioactive and benign datasets. However, there are two drawbacks, 1) heavy burden for adversarial attacks, and 2) requiring access to output logits unlike our black-box setting." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.78, + 0.574, + 0.794 + ], + "angle": 0, + "content": "2.4 Limitations of the Prior Works in Verification" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.801, + 0.517, + 0.815 + ], + "angle": 0, + "content": "Clean-labeled watermarking is formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.408, + 0.824, + 0.785, + 0.838 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathbf {x}} = \\mathbf {x} + \\mathbf {w}, \\quad \\text {a n d} \\quad \\hat {\\mathbf {y}} ^ {\\mathbf {x}} = \\mathbf {y} ^ {\\mathbf {x}}, \\tag {1}\n\\]" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.347, + 0.129 + ], + "angle": 0, + "content": "J. Jang et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.786, + 0.193 + ], + "angle": 0, + "content": "where \\((\\mathbf{x},\\mathbf{y}^{\\mathbf{x}})\\) signifies a benign sample and its corresponding ground truth, sampled from \\((\\mathbf{X},\\mathbf{Y}^{\\mathbf{X}})\\). \\(\\mathbf{w}\\) is the negligible watermark \\((||\\mathbf{w}|| < \\epsilon\\), with \\(\\epsilon\\) as a small value), and \\(\\hat{\\mathbf{x}}\\) denotes the watermarked data. \\(\\hat{\\mathbf{y}}^{\\mathbf{x}}\\) is the ground truth label of \\(\\hat{\\mathbf{x}}\\) and is equal to \\(\\mathbf{y}^{\\mathbf{x}}\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.193, + 0.788, + 0.268 + ], + "angle": 0, + "content": "For verification, prior works verify based on intentional degradation. Let \\(\\mathcal{F}\\) symbolizes a DNN model, and \\(\\theta_{\\mathcal{F}}\\) denotes its weights. Backdoor attacks aim to induce misclassifications in \\(\\mathcal{F}(\\mathbf{x} + \\mathbf{w}, \\theta_{\\mathcal{F}})\\). Data poisoning selects a subset of clean data for intentional misclassifications. Radioactive data exploits the enhanced performance of \\(\\mathcal{F}(\\mathbf{x} + \\mathbf{w}, \\theta_{\\mathcal{F}})\\) in comparison to \\(\\mathcal{F}(\\mathbf{x}, \\theta_{\\mathcal{F}})\\), leveraging this improvement to verification." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.269, + 0.788, + 0.389 + ], + "angle": 0, + "content": "However, our observations highlight the unreliability of verification that depends on intentional degradation. This is due to the fact that the degree of degradation is contingent upon the performance of \\(\\mathcal{F}(\\mathbf{x},\\theta_{\\mathcal{F}})\\). The probability of such degradation occurring is inversely related to the model's accuracy on the clean data, expressed as \\(1 - Acc(\\mathcal{F}(\\mathbf{x},\\theta_{\\mathcal{F}}),\\mathbf{y}^{\\mathbf{x}})\\). High \\(Acc(\\mathcal{F}(\\mathbf{x},\\theta_{\\mathcal{F}}),\\mathbf{y}^{\\mathbf{x}})\\) makes degradation unlikely, while low \\(Acc(\\mathcal{F}(\\mathbf{x},\\theta_{\\mathcal{F}}),\\mathbf{y}^{\\mathbf{x}})\\) increases the chance of degradation. Also, the process of intentional degradation itself poses risks, potentially impairing the model's accuracy on clean data. Thus, a reliable and non-damaging verification method is essential." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.413, + 0.343, + 0.429 + ], + "angle": 0, + "content": "3 Motivation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.445, + 0.786, + 0.536 + ], + "angle": 0, + "content": "Our method is conceived from recognizing the effects of data bias. A biased dataset is characterized by data that encapsulates unintended knowledge. Take, for instance, the CIFAR10 dataset, which has a 'ship' category where most images feature a 'sea' background. Similarly, most images within the 'airplane' category share a 'sky' background. In such situations, a network trained on the CIFAR10 dataset has a tendency to identify the 'sea' background as a significant feature of 'ship' class." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.537, + 0.788, + 0.704 + ], + "angle": 0, + "content": "To replicate this phenomenon, we generated two sets of synthetic images, one featuring 'sea' and the other 'sky' without 'ship' and 'airplane', utilizing stable diffusion [29]. We then trained a ResNet18 on the CIFAR10 dataset. Subsequently, we evaluated the trained ResNet18 on the synthetic images. As results, \\(56.10\\%\\) of synthetic 'sea' images were identified as 'ship', and \\(65.29\\%\\) of synthetic 'sky' images were classified as 'airplane' despite there was no 'ship' and 'airplane'. Also, Fig. 2 shows some class activation maps (CAM), highlighting the unintended knowledge about sea horizon, waves, and clouds. These are not features inherent to the objects 'ship' and 'airplane', but they were learned. If the model had solely learned 'ship' and 'airplane' objects without taking the background into account, it would have randomly classify these synthetic images. Drawing from the results of the background image classification, we" + }, + { + "type": "image", + "bbox": [ + 0.241, + 0.729, + 0.498, + 0.812 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.729, + 0.764, + 0.812 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.266, + 0.822, + 0.733, + 0.837 + ], + "angle": 0, + "content": "Fig. 2: Synthetic background images and their Class Activation Maps (CAM)." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.435, + 0.115, + 0.733, + 0.13 + ], + "angle": 0, + "content": "Undercover Bias for Dataset Copyright Protection" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.224 + ], + "angle": 0, + "content": " devised a method for clean-labeled dataset protection that capitalizes on this unintended knowledge. When we introduce a unique bias to a dataset on a class-by-class basis, a network trained on this biased dataset is likely to learn the bias. Consequently, it may classify data containing only the bias without target information (such as background images) as belonging to each respective class." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.244, + 0.317, + 0.26 + ], + "angle": 0, + "content": "4 Method" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.274, + 0.788, + 0.368 + ], + "angle": 0, + "content": "Bias has historically been perceived negatively, primarily because it can degrade DNN performance and raise ethical issues, such as gender or race bias. Consequently, numerous studies have focused on debiasing techniques [18, 26, 27]. However, our approach deviates from this conventional viewpoint by intentionally embedding class-wise hidden biases within datasets, utilizing these biases as a means of copyright protection. In the following section, we outline the step-by-step process of implementing our method." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.386, + 0.618, + 0.403 + ], + "angle": 0, + "content": "4.1 Noise Patch Placement: Class-wise Bias Embedding" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.409, + 0.787, + 0.44 + ], + "angle": 0, + "content": "Our initial step involved leveraging biases by introducing noise patches into a benign dataset. Each class was assigned a unique noise pattern, which was injected as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.397, + 0.45, + 0.786, + 0.466 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathbf {x}} = \\mathbf {x} + \\lambda \\mathbf {n} \\quad s. t. \\quad \\mathbf {y} ^ {\\mathbf {x}} = \\mathbf {y} ^ {\\mathbf {n}}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.475, + 0.788, + 0.581 + ], + "angle": 0, + "content": "where \\(\\mathbf{n}\\) and \\(\\mathbf{y}^{\\mathrm{n}}\\) denote the class-specific noise image and its corresponding label. The noise \\(\\mathbf{n}\\) was generated as zero-mean Gaussian random noise, \\(\\mathcal{N}(0,I)\\), and positioned at predefined, class-specific areas as illustrated in Fig. 3. This noise was injected into randomly selected \\(50\\%\\) training data, utilizing noise patches with a coefficient \\(\\lambda = 0.01\\). To validate this approach, we trained a ResNet18 on the CIFAR10 dataset starting from a random initialization for 100 epochs using the Adam optimizer, cosine decay, and data augmentation. We then tested the model on the noise images." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.581, + 0.788, + 0.658 + ], + "angle": 0, + "content": "The results, as presented in Table 1, lead us to three critical insights. First, it's possible to embed hidden knowledge via noise placement without altering the label (clean-labeled). Despite the significant difference between \\(\\mathbf{n}\\) and \\(\\mathbf{x}\\), the trained models exhibited higher accuracy for the noise images, validating the feasibility of dataset protection. Second, the accuracy was higher for \\(\\lambda \\mathbf{n} + \\mu (\\mathbf{X})\\) than \\(\\lambda \\mathbf{n}\\), where \\(\\mu (\\cdot)\\) denotes" + }, + { + "type": "image", + "bbox": [ + 0.222, + 0.681, + 0.498, + 0.771 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.23, + 0.781, + 0.49, + 0.796 + ], + "angle": 0, + "content": "Fig.3: Examples of noise patch placement." + }, + { + "type": "table", + "bbox": [ + 0.521, + 0.689, + 0.774, + 0.771 + ], + "angle": 0, + "content": "
Watermark ImageRandom Flip Horizontal VerticalVal Acc on Benign (%)Val Acc on Watermark (%)
λnXX92.63±0.1762.14±15.54
λn + μ(X)71.69±13.60
λn✓X94.07±0.1120.26±2.68
λn + μ(X)22.28±3.69
λn✓✓92.31±0.239.15±1.88
λn + μ(X)12.94±4.03
" + }, + { + "type": "table_caption", + "bbox": [ + 0.509, + 0.78, + 0.783, + 0.837 + ], + "angle": 0, + "content": "Table 1: Results of watermarks based on noise placement. The \"Random Flip\" indicates whether horizontal and vertical flip were used in data augmentation or not." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.347, + 0.129 + ], + "angle": 0, + "content": "J. Jang et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.329 + ], + "angle": 0, + "content": "the average function. Given that we used small noises, the average magnitude for \\(\\lambda \\mathbf{n}\\) was nearly zero, but that wasn't the case for \\(\\mathbf{x} + \\lambda \\mathbf{n}\\). This disparity created a sort of domain gap, causing the trained model to often disregard \\(\\lambda \\mathbf{n}\\). To bridge this gap, we added \\(\\mu (\\mathbf{X})\\) to \\(\\lambda \\mathbf{n}\\), which led to more successful classification, thereby indicating a more effective dataset watermark. Third, the placement of noise is significantly affected by spatial transformations. Interestingly, classification accuracy was highest, around \\(60\\%\\), when no flip was used in data augmentation. When a horizontal flip was applied, the accuracy dropped to approximately \\(20\\%\\), due to the presence of five pairs of mirrored patterns in the noise patches used. When both horizontal and vertical flips were used, the confusion of many patches due to combined flips resulted in a \\(12\\%\\) accuracy. If we utilized more precise positions, the verification could be further influenced by rotation and translation. This suggests an issue of robust pattern of watermark." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.35, + 0.674, + 0.365 + ], + "angle": 0, + "content": "4.2 Overlaying Auxiliary Dataset: Robust Bias to Augmentation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.373, + 0.788, + 0.448 + ], + "angle": 0, + "content": "To mitigate the vulnerability to data augmentation, it's necessary to define more robust patterns. However, manually generating diverse patterns that satisfy: 1) having the same number of classes, 2) unrelatedness between any two classes, and 3) robustness to spatial transformation, is challenging. We instead opted for an auxiliary dataset. Given the plethora of available datasets in the research field, we used overlaid data as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.374, + 0.458, + 0.787, + 0.474 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathbf {x}} = (1 - \\lambda) \\mathbf {x} + \\lambda \\mathbf {z} \\quad s. t. \\quad \\mathbf {y} ^ {\\mathbf {x}} = \\mathbf {y} ^ {\\mathbf {z}}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.484, + 0.789, + 0.68 + ], + "angle": 0, + "content": "where \\((\\mathbf{z},\\mathbf{y}^{\\mathbf{z}})\\) represents the auxiliary data and its label. We overlaid data while taking into account the labels of the two datasets. For example, CIFAR10 and Fashion MNIST [43] are independent datasets, each with ten distinct classes. If we selected \\(\\mathbf{x}\\) from the 'airplane' class in CIFAR10 and \\(\\mathbf{z}\\) from the 'pullover' class in Fashion MNIST, their corresponding \\(\\mathbf{y}^{\\mathbf{x}}\\) and \\(\\mathbf{y}^{\\mathbf{z}}\\) could both be denoted as 'Class 0', regardless of their semantic meanings. We trained ResNet18 using an overlay of the CIFAR10 dataset (target) and the Fashion MNIST dataset (auxiliary) using the same training recipe of the noise placement case. Fig. 4 and Table 2 show some examples of overlaid images and the training results. As shown, the results clearly demonstrate robustness to spatial transformation. However, the overlaid data is overly conspicuous and deviates from the benign data, significantly undermining the original task as represented by the validation accuracy on benign data. Further, they can be filtered by visual inspection. Therefore, it's crucial to make invisible watermarks." + }, + { + "type": "image", + "bbox": [ + 0.24, + 0.709, + 0.517, + 0.797 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.239, + 0.808, + 0.516, + 0.822 + ], + "angle": 0, + "content": "Fig. 4: Examples of overlaid image \\((\\lambda = 0.3)\\)." + }, + { + "type": "table", + "bbox": [ + 0.553, + 0.701, + 0.744, + 0.798 + ], + "angle": 0, + "content": "
λRandom Flip Horizontal VerticalVal Acc on Benign (%)Val Acc on Watermark (%)
0.3XX90.47±0.2483.45±0.78
✓X92.09±0.2281.59±2.14
✓✓89.34±0.4169.05±5.17
0.5XX89.36±0.2186.92±0.80
✓X91.03±0.2985.52±1.16
✓✓87.73±0.7276.58±4.64
0.7XX88.52±0.1689.86±0.84
✓X89.97±0.4987.64±1.36
✓✓87.43±0.3386.10±0.53
" + }, + { + "type": "table_caption", + "bbox": [ + 0.528, + 0.808, + 0.765, + 0.836 + ], + "angle": 0, + "content": "Table 2: Results of watermarks based on data overlay." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.435, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "Undercover Bias for Dataset Copyright Protection" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.116, + 0.786, + 0.126 + ], + "angle": 0, + "content": "7" + }, + { + "type": "image", + "bbox": [ + 0.333, + 0.147, + 0.67, + 0.306 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.336, + 0.788, + 0.392 + ], + "angle": 0, + "content": "Fig.5: Architecture of the proposed Dataset Watermarking Network. The network is composed of two auto encoders and two classifiers. The proposed watermark can be obtained by subtracting the target data from watermarked data and adding mean of the target dataset. Note that this is a scheme of dataset watermarking network (DWN), not cheating model." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.421, + 0.561, + 0.436 + ], + "angle": 0, + "content": "4.3 Undercover Bias: Invisible Bias Embedding" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.445, + 0.784, + 0.505 + ], + "angle": 0, + "content": "Given the lessons learned from previous approaches, we concluded that a watermark based on hidden bias must be 1) robust to spatial transformation, and 2) nearly invisible. To address these requirements, we developed our proposed watermarking method using image steganography [4] and an auxiliary dataset. The process is as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.395, + 0.515, + 0.786, + 0.55 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\hat {\\mathbf {x}} = D W N (\\mathbf {x}, \\mathbf {z}), \\\\ \\begin{array}{l} \\mathbf {x} = \\mathbf {D} w, \\mathbf {w} = (\\mathbf {x}, z), \\\\ \\mathbf {w} = \\hat {\\mathbf {x}} - \\mathbf {x}, \\quad \\text {a n d} \\quad \\mathbf {y} ^ {\\mathbf {w}} = \\mathbf {y} ^ {\\mathbf {z}}, \\end{array} \\tag {4} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.56, + 0.788, + 0.651 + ], + "angle": 0, + "content": "where \\((\\mathbf{w},\\mathbf{y}^{\\mathbf{w}})\\) represents the watermark and its label. It's crucial to create undetectable watermarks and their corresponding labels. To this, we developed a network called \"Dataset Watermarking Network (DWN)\" with consideration for reconstruction-aware and perceptual constraints, as shown in Fig. 5. This network consists of two autoencoders: one for hiding \\((\\mathcal{G}_w)\\) and the other for reconstruction \\((\\mathcal{G}_r)\\). \\(\\mathcal{G}_w\\) takes \\(\\mathbf{x}\\) and \\(\\mathbf{z}\\), and produces an output \\(\\hat{\\mathbf{x}}\\), which closely resembles \\(\\mathbf{x}\\). \\(\\mathcal{G}_r\\) reconstructs \\(\\mathbf{x}'\\) and \\(\\mathbf{s}'\\) from \\(\\hat{\\mathbf{x}}\\) as:" + }, + { + "type": "equation", + "bbox": [ + 0.339, + 0.662, + 0.787, + 0.68 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathbf {x}} = \\mathcal {G} _ {w} (\\mathbf {x}, \\mathbf {z}, \\theta_ {\\mathcal {G} _ {w}}), \\quad \\text {a n d} \\quad \\mathbf {x} ^ {\\prime}, \\mathbf {z} ^ {\\prime} = \\mathcal {G} _ {r} (\\hat {\\mathbf {x}}, \\theta_ {\\mathcal {G} _ {r}}), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.691, + 0.788, + 0.767 + ], + "angle": 0, + "content": "where \\(\\theta_{\\mathcal{G}_w}, \\theta_{\\mathcal{G}_r}\\) are weights of the autoencoders. Two classifiers, \\(\\mathcal{H}_x\\) and \\(\\mathcal{H}_{\\mathbf{w}}\\), are used as perceptual constraints. Training an numerous classifiers for model-agnostic watermarking is computationally challenging. Instead, we adopted a simpler approach using a basic architecture incorporating spatial dropout [38] and dropout [36]. All autoencoders and classifiers were simultaneously trained using \\(\\ell 1\\) loss and cross entropy \\((\\mathcal{L}_{CE})\\)." + }, + { + "type": "equation", + "bbox": [ + 0.257, + 0.776, + 0.786, + 0.844 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\min _ {\\theta_ {\\mathcal {G} _ {w}}, \\theta_ {\\mathcal {G} _ {r}}, \\theta_ {\\mathcal {H} _ {\\mathbf {x}}}, \\theta_ {\\mathcal {H} _ {\\mathbf {w}}}} \\lambda_ {1} ^ {\\mathcal {G}} | \\mathbf {x} - \\hat {\\mathbf {x}} | + \\lambda_ {2} ^ {\\mathcal {G}} | \\mathbf {x} - \\mathbf {x} ^ {\\prime} | + \\lambda_ {3} ^ {\\mathcal {G}} | \\mathbf {z} - \\mathbf {z} ^ {\\prime} | \\\\ + \\lambda_ {1} ^ {\\mathcal {H}} \\mathcal {L} _ {C E} \\left(\\mathcal {H} _ {\\mathbf {x}} \\left(\\mathbf {x}, \\theta_ {\\mathcal {H} _ {\\mathbf {x}}}\\right), \\mathbf {y} ^ {\\mathbf {x}}\\right) + \\lambda_ {2} ^ {\\mathcal {H}} \\mathcal {L} _ {C E} \\left(\\mathcal {H} _ {\\mathbf {x}} \\left(\\hat {\\mathbf {x}}, \\theta_ {\\mathcal {H} _ {\\mathbf {x}}}\\right), \\mathbf {y} ^ {\\mathbf {x}}\\right) \\tag {6} \\\\ + \\lambda_ {3} ^ {\\mathcal {H}} \\mathcal {L} _ {C E} \\left(\\mathcal {H} _ {\\mathbf {x}} \\left(\\mathbf {x} ^ {\\prime}, \\theta_ {\\mathcal {H} _ {\\mathbf {x}}}\\right), \\mathbf {y} ^ {\\mathbf {x}}\\right) + \\lambda_ {4} ^ {\\mathcal {H}} \\mathcal {L} _ {C E} \\left(\\mathcal {H} _ {\\mathbf {w}} \\left(\\mathbf {x} ^ {\\prime} - \\mathbf {x} + \\mu (\\mathbf {X}), \\theta_ {\\mathcal {H} _ {\\mathbf {w}}}\\right), \\mathbf {y} ^ {\\mathbf {z}}\\right), \\\\ \\end{array}\n\\]" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.347, + 0.129 + ], + "angle": 0, + "content": "J. Jang et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.146, + 0.788, + 0.193 + ], + "angle": 0, + "content": "where \\(\\lambda^{\\mathcal{G}}\\)'s and \\(\\lambda^{\\mathcal{H}}\\)'s are weighting factors and \\(\\theta\\)'s are the weights of the autoencoders and classifiers. Once trained, the generation of watermarked images and watermarks can be performed without additional training." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.215, + 0.332, + 0.229 + ], + "angle": 0, + "content": "4.4 Discussion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.24, + 0.788, + 0.348 + ], + "angle": 0, + "content": "Issue about the number of classes. To ensure that \\(\\mathbf{y}^{\\mathbf{x}} = \\mathbf{y}^{\\mathbf{w}}\\), it's necessary to pair instances of the two datasets to share the same label. This presents a challenge because the watermark set must have more classes than the target dataset for class-by-class pairing. To overcome this challenge requirement, we employed the modulo operation, which returns the remainder of a division. This allows us to replace the condition with \\(\\mathbf{y}^{\\mathbf{x}} \\equiv \\mathbf{y}^{\\mathbf{w}}\\) (mod \\(N^{\\mathbf{w}}cls\\)), where \\(N^{\\mathbf{w}}cls\\) represents the number of classes in \\(\\mathbf{W}\\). This allows for effective pairing even when the number of classes differs between the datasets." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.369, + 0.788, + 0.416 + ], + "angle": 0, + "content": "Verification metric. We utilized a strict Black-box Test [33], assuming that malicious users provide only predicted classes without additional information. Considering class imbalance, we implemented mean class accuracy (mAcc) on \\(\\mu (\\mathbf{X}) + \\mathbf{w}\\) with \\(\\mathbf{y}^{\\mathrm{w}}\\) as:" + }, + { + "type": "equation", + "bbox": [ + 0.333, + 0.426, + 0.787, + 0.469 + ], + "angle": 0, + "content": "\\[\n\\frac {1}{N _ {c l s} ^ {\\mathbf {w}}} \\sum_ {c = 1} ^ {N _ {c l s} ^ {\\mathbf {w}}} \\mathbb {P} \\left(\\mathcal {F} \\left(\\mu (\\mathbf {X}) + \\mathbf {w}, \\theta_ {\\mathcal {F}}\\right) = c \\mid \\mathbf {y} ^ {\\mathbf {w}} = k\\right) > \\tau , \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.48, + 0.788, + 0.543 + ], + "angle": 0, + "content": "where \\(\\mathbb{P}(\\cdot)\\) indicates probability. If there is no cheating, \\(\\mathcal{F}\\) cannot perform well on \\(\\mu (\\mathbf{X}) + \\mathbf{w}\\). On the contrary, suspicion arises when the model exhibits a higher mAcc than a predetermined threshold \\(\\tau\\). If the two datasets, \\(\\mathbf{X}\\) and \\(\\mathbf{Z}\\) have different numbers of classes, we can use \\(\\mathcal{F}(\\mu (\\mathbf{X}) + \\mathbf{w},\\theta_{\\mathcal{F}})\\equiv k\\) (mod \\(N_{cls}^{\\mathbf{w}}\\))." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.564, + 0.788, + 0.713 + ], + "angle": 0, + "content": "Threshold determination. For verification, it is important to define a threshold that is unattainable for clean models by chance. We assume an ideal scenario where the mAcc of a clean model follows a Gaussian-like distribution, peaking at \\(\\frac{1}{N_{cls}^{\\mathbf{w}}}\\), with negligible probability at \\(0\\%\\) mAcc. Given this assumption, it's unlikely for a clean model to achieve \\(2 \\times \\frac{1}{N_{cls}^{\\mathbf{w}}}\\). This assumption is based on the secret dataset being distinct from the target dataset, allowing us to set a minimum threshold at \\(\\frac{2}{N_{cls}^{\\mathbf{w}}}\\). We validate this assumption empirically in Section 7.1. While empirical rule can determine thresholds based on standard deviations, approximating the distribution requires extensive training and computation. The practical solution of \\(\\frac{2}{N_{cls}^{\\mathbf{w}}}\\) offers a cheap alternative." + }, + { + "type": "title", + "bbox": [ + 0.214, + 0.733, + 0.636, + 0.751 + ], + "angle": 0, + "content": "5 Experiments I: Comparison with Prior Works" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.788, + 0.843 + ], + "angle": 0, + "content": "In this section, we compared our work to backdoor attacks data poisoning, and radioactive data in various aspects such as computing time, invisibility, harmlessness, verification ability using CIFAR10 dataset. For all method, we marked \\(50\\%\\) of the training data. For backdoor attacks, there are two main approaches: clean labeled and label-noised. We used label-noised backdoor attacks like badnets [10] and blended [5] for" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.435, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Undercover Bias for Dataset Copyright Protection" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "table", + "bbox": [ + 0.315, + 0.144, + 0.684, + 0.272 + ], + "angle": 0, + "content": "
MethodTime Cost (sec/image)Clean Labeled?Model Agnostic?Unlimited # Verification?Invisibility (avg. SSIM)
BadNets [10]1.67e-7X✓✓0.8569
Blended [5]2.50e-6X✓✓0.9775
Hidden Trigger [32]4.40e-1✓X✓0.8859
Sleeper Agent [35]1.27e+0✓X✓0.8751
PoisonFrogs [34]2.31e-1✓XX0.8738
MetaPoison [13]1.70e+0✓XX0.8576
Bullseye [3]2.90e-1✓XX0.8737
Gradient Matching [8]6.86e-1✓XX0.8722
Radioactive Data [31]1.36e-1✓X✓0.9202
Proposed1.07e-3✓✓✓0.9785
" + }, + { + "type": "table_caption", + "bbox": [ + 0.356, + 0.283, + 0.646, + 0.297 + ], + "angle": 0, + "content": "Table 3: Basic Specification of Various Methods." + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.326, + 0.787, + 0.508 + ], + "angle": 0, + "content": "only assessing basic specification. For clean labeled backdoor attacks, including hidden trigger [32] and sleeper agent [35], we affixed ten distinct triggers to the training data, assigning one unique trigger per class. Regarding data poisoning (i.e. poison frogs [34], meta poison [13], bullseye [3], gradient matching [8]), we selected one verification image per class, resulting in a total of ten verification images (multi-target setting with a \\(5\\%\\) budget per verification image). For radioactive data [31], we marked \\(50\\%\\) of the training data and the entire test data. The prior works require reference model, so we adopted a ResNet18 trained on benign CIFAR10 as it. Official codes of [8, 31, 35] were used. In our approach, we concealed Fashion MNIST behind the CIFAR10 dataset using a pre-trained DWN. As the pre-trained DWN, we employed U-Net [30] as autoencoders, and Vanilla CNN with four convolution layers and dropouts as classifiers. Please note that we applied watermarking to \\(50\\%\\) training data for every experiment." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.529, + 0.557, + 0.544 + ], + "angle": 0, + "content": "5.1 Comparison in Fundamental Specifications" + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.553, + 0.787, + 0.734 + ], + "angle": 0, + "content": "We first assessed about fundamental properties. The harmlessness was gauged based on the validation accuracy on benign data, with a higher validation accuracy signifying less impact on the target task. The time cost includes only watermarking time per image, excluding training time for DWN and the reference model. Then, Structured Similarity (SSIM) [42] between benign and watermarked images was used to measure invisibility. Table 3 summarizes specifications of every methods. Label-noised backdoor attacks can be identified through visual inspection of images due to incorrect labels. Clean-labeled backdoor attacks, data poisoning and radioactive data rely on reference models, rendering them model-dependent and time-consuming, with visible watermarks. Data poisoning has limitations regarding the number of victim images. Our proposed approach outperforms these methods by generating less visible, model-agnostic watermarks, executing faster, and having no restrictions on the number of verification images." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.756, + 0.557, + 0.77 + ], + "angle": 0, + "content": "5.2 Comparison in Effectiveness of Watermark" + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.78, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Purpose and Setting. We conducted a comparative analysis on CIFAR10 focusing on harmlessness and verifiability across different methods, including backdoor attacks, data poisoning, radioactive data, and our work. For all method, we watermarked randomly selected \\(50\\%\\) of the training data. The experiments were carried out under three" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.116, + 0.347, + 0.129 + ], + "angle": 0, + "content": "J. Jang et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.298 + ], + "angle": 0, + "content": "scenarios: ResNet18 from scratch for 100 epochs, MobileNetV2 from scratch for 300 epochs, and MobileNetV2 from ImageNet pre-trained weights for 100 epochs. Each scenario was repeated 10 times with data augmentation. Harmlessness was evaluated by measuring the validation accuracy on a benign dataset, excluding ten images used for verification in data poisoning to ensure accuracy. For Verification ability, we measured own metrics of backdoor attacks, data poisoning, radioactive data, and our work. The ASR, which is the success rate of intended misclassifications, was used as a metric for backdoor attacks and data poisoning. For radioactive data, we compared the difference between the validation losses of the benign and the radioactive-marked dataset. For our work, we measured all of mAcc, ASR, and the difference between losses." + }, + { + "type": "image", + "bbox": [ + 0.288, + 0.327, + 0.72, + 0.426 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.287, + 0.427, + 0.717, + 0.527 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.538, + 0.788, + 0.594 + ], + "angle": 0, + "content": "Fig. 6: Comparisons of performance to recent works using Left: ResNet18 from scratch, Middle: MobileNetV2 from scratch, and Right: MobileNetV2 from ImageNet pre-trained. Each dot represents a training result of each trial. \"Performance of Hidden Training\" means \"mAcc\" for the proposed method and \"ASR\" for others." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.629, + 0.789, + 0.844 + ], + "angle": 0, + "content": "Results. The results are depicted in Fig. 6. Note that the ResNet18 was used as a reference for clean-labeled backdoor attacks, data poisoning, and radioactive data. In contrast, our method works without the need for any reference model. For backdoor attacks, both the hidden trigger consistently failed for all cases. The sleeper agent showed a marginally higher ASR only for the known architecture (ResNet18), but failed for unseen architectures. Regarding data poisoning, most attempts were unsuccessful except for gradient matching. The gradient matching often achieved higher ASR, but lower validation accuracy on benign CIFAR10. Please note that data poisoning can affect only few verification images, whereas our method has no limit on the number of watermarks. Our method outperformed all prior works, achieving higher validation accuracy for both benign CIFAR10 and watermarked data. Further, our approach consistently reported significant accuracy on the watermarks in every trial, underscoring its reliability. In contrast to radioactive data, our work showed lower validation loss on benign CIFAR10 and a more significant difference between validation losses of benign and" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.435, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Undercover Bias for Dataset Copyright Protection" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "table", + "bbox": [ + 0.223, + 0.144, + 0.496, + 0.256 + ], + "angle": 0, + "content": "
DatasetMethodVal Acc on Benign (%)Val mAcc on Watermark (%)Val mASR(%)SSIM
CIFAR100Clean71.79±0.18N/AN/A1.0000
Sleeper Agent66.59±0.46N/A4.20±0.840.8821
Gradient Matching66.53±0.25N/A32.00±5.880.8797
Proposed68.27±0.1530.35±1.04N/A0.9710
FER2013Clean68.89±0.29N/AN/A1.0000
Sleeper Agent62.68±0.37N/A11.02±3.050.8392
Gradient Matching63.36±0.30N/A47.62±19.340.8197
Proposed64.63±0.4538.15±0.64N/A0.9833
*FMNISTClean94.21±0.10N/AN/A1.0000
Sleeper Agent93.55±0.14N/A0.00±0.000.7882
Gradient Matching93.35±0.21N/A6.67±5.480.7771
Proposed93.66±0.2122.31±2.02N/A0.9744
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.225, + 0.256, + 0.319, + 0.263 + ], + "angle": 0, + "content": "*FMNIST: Fashion MNIST" + }, + { + "type": "table", + "bbox": [ + 0.512, + 0.147, + 0.781, + 0.266 + ], + "angle": 0, + "content": "
ArchitectureVal Acc on Benign (%)Val mAcc on Watermark (%)
EfficientNetB0 (Transfer)Clean96.43±0.249.89±0.08
Cheating96.27±0.2954.18±2.15
PVTv2-B0 (Transfer)Clean95.00±0.3710.29±0.12
Cheating94.54±0.4251.39±2.32
ResMLP-12-224 (Transfer)Clean96.14±0.2110.19±0.31
Cheating95.44±0.4053.10±1.74
PiT Tiny (Transfer)Clean94.78±0.399.71±0.32
Cheating94.59±0.3248.09±3.21
" + }, + { + "type": "table_caption", + "bbox": [ + 0.233, + 0.276, + 0.483, + 0.29 + ], + "angle": 0, + "content": "Table 4: Applicability to various datasets." + }, + { + "type": "table_caption", + "bbox": [ + 0.509, + 0.276, + 0.782, + 0.29 + ], + "angle": 0, + "content": "Table 5: Applicability to various architectures" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.323, + 0.786, + 0.353 + ], + "angle": 0, + "content": "watermarked data. From the results, we can conclude that our work outperforms prior works in terms of harmlessness, reliability in verification." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.381, + 0.561, + 0.4 + ], + "angle": 0, + "content": "6 Experiment II: General Applicability" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.419, + 0.784, + 0.448 + ], + "angle": 0, + "content": "The prior experiment showed feasibility and superiority of our work in limited settings. This section introduces extensions to further datasets, architectures and tasks." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.477, + 0.607, + 0.492 + ], + "angle": 0, + "content": "6.1 Application to Further Architectures and Datasets" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.508, + 0.787, + 0.748 + ], + "angle": 0, + "content": "Purpose and Setting. To broaden the scope of our evaluation beyond the initial limited settings, which focused on a few architectures and solely the CIFAR10 dataset, we expanded our analysis to include additional architectures and datasets, thereby affirming the widespread applicability of our work. We incorporated CIFAR100, FER2013 [9], and Fashion MNIST for this extended evaluation. CIFAR100 (100 classes) and FER2013 (7 classes) were assessed against the most effective previously identified methods: sleeper agent for backdoor attacks and gradient matching for data poisoning. The reference models included ResNet18 and a Benign CNN, while the cheating models utilized DenseNet-BC [12], trained from scratch. CIFAR100's auxiliary dataset comprised Fashion MNIST and MNIST (10 classes each), while FER2013's auxiliary dataset consisted of the first 7 MNIST classes. Also, we tested the following architectures as cheating models on CIFAR10 (target) and Fashion MNIST (auxiliary): EfficientNet [37], PVTv2 [41], ResMLP [39], and PiT [11]. We trained each architecture 35 epochs starting from ImageNet pre-trained weights, employing SGD optimization, warmup, label smoothing, and data augmentation (i.e., spatial transformation and mixup), repeated multiple times to ensure robustness and reliability of the results." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.75, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Results. Table 4 demonstrates that our work excels in three key aspects: 1) harmlessness, 2) invisibility, and 3) verifiability across all datasets. Also, our work consistently outperforms in mAcc on the watermark across various architectures, as presented in Table 5. These findings lead us to conclude that our work effectively operates on diverse architectures and datasets, making it universally applicable. In all cases, our proposed threshold attained \\(100\\%\\) accuracy for distinguishing cheating models from clean ones." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.271, + 0.116, + 0.347, + 0.129 + ], + "angle": 0, + "content": "J. Jang et al." + }, + { + "type": "table", + "bbox": [ + 0.225, + 0.159, + 0.488, + 0.248 + ], + "angle": 0, + "content": "
Tiny ImageNet
Val Acc on Benign (%)Val mAcc on Watermark (%)
Clean48.23±1.468.84±2.29
Cheating46.84±1.1434.37±5.61
ImageNet
Val Acc on Benign (%)Val mAcc on Watermark (%)
Clean70.42±0.4111.90±1.33
Cheating70.14±0.7243.95±3.21
" + }, + { + "type": "table_caption", + "bbox": [ + 0.217, + 0.258, + 0.495, + 0.286 + ], + "angle": 0, + "content": "Table 6: Validation accuracy of MobileNetV2 on Tiny ImageNet and ImageNet." + }, + { + "type": "image", + "bbox": [ + 0.506, + 0.143, + 0.782, + 0.248 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.504, + 0.258, + 0.785, + 0.286 + ], + "angle": 0, + "content": "Fig.7: Results on segmentation and an example, with emphasized watermark for clarity." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.315, + 0.547, + 0.331 + ], + "angle": 0, + "content": "6.2 Application to Fine-grained Classification" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.339, + 0.788, + 0.445 + ], + "angle": 0, + "content": "Purpose and Setting. To address the limitation of requiring the auxiliary dataset to contain more classes than the target dataset, we employed the modulo operation. We validated the solution using both Tiny ImageNet, which consists of 100 classes, and ImageNet, which comprises 1,000 classes. Fashion MNIST was used as a auxiliary dataset, and \\(50\\%\\) of training datasets were watermarked. We trained MobileNetV2 on 1) watermarked Tiny ImageNet [16] (0.9883 avg. SSIM) from scratch and 2) watermarked ImageNet (0.9570 avg. SSIM) from a benign ImageNet pre-trained model." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.445, + 0.788, + 0.566 + ], + "angle": 0, + "content": "Results. Table 6 shows the average performance results from 30 and 5 trials for each dataset. As shown, training MobileNetV2 on watermarked data leads to a marginal decrease in accuracy on the benign validation dataset, but demonstrates significant performance improvements in detecting watermarks. Conversely, the accuracy of the clean model is significantly diminished, nearing the chance level. This finding proves the verifiability of our work, even when the target dataset contains a greater number of classes compared to the auxiliary dataset. Therefore, it is not necessary to prepare a auxiliary dataset that matches the target dataset in terms of the number of classes." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.587, + 0.504, + 0.603 + ], + "angle": 0, + "content": "6.3 Application to Image Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.612, + 0.788, + 0.766 + ], + "angle": 0, + "content": "Purpose and Setting. To extend our work to image segmentation, we adapted our method to create spatially varying watermarks. We resized the auxiliary data to a smaller scale, such as 8x8 pixels, and repeatedly stitched it to segments of the \\(50\\%\\) PASCAL VOC 2012 data, considering each segment's label. This produced segment-wise watermarked images, as shown in Fig. 7. To adjust the DWN for segmentation, we replaced the two classifiers with simple autoencoders featuring dropout. We trained a segmentation autoencoder with a MobileNetV2 backbone from scratch on this watermarked dataset, employing the Adam optimizer with learning rate decay starting from 1e-3, a batch size of 60, and data augmentation. In image segmentation, meaningful information resides in silhouettes, requiring a higher threshold compared to \\(\\frac{2}{N_{cls}^2}\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.766, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Results. Fig. 7 provides an example image and shows the performances on the benign Pascal VOC dataset and the watermarked dataset after training. In the context of image segmentation on Pascal VOC, we adopted mIoU as the evaluation metric. As shown, there is almost no damage to the original task. For verifying cheating models, we adopted the mean class pixel accuracy on the masked region. The watermark" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.435, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Undercover Bias for Dataset Copyright Protection" + }, + { + "type": "page_number", + "bbox": [ + 0.77, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "image", + "bbox": [ + 0.307, + 0.15, + 0.486, + 0.262 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.324, + 0.265, + 0.483, + 0.277 + ], + "angle": 0, + "content": "(a) ResNet18 on CIFAR10" + }, + { + "type": "image", + "bbox": [ + 0.51, + 0.15, + 0.685, + 0.261 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.518, + 0.265, + 0.684, + 0.277 + ], + "angle": 0, + "content": "(b) DenseNet on CIFAR100" + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.288, + 0.785, + 0.321 + ], + "angle": 0, + "content": "Fig.8: Histogram of mAcc on watermark of clean and cheating models. For both cases, the clean models hardly achieved higher mAcc than \\(\\frac{2}{N_{cls}^{\\mathbf{z}}}\\), but all of the cheating models achieved." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.349, + 0.785, + 0.409 + ], + "angle": 0, + "content": "was surreptitiously learned in most of the trials, enabling successful verification of the cheating models. While the clean models reported accuracies lower than 0.2, \\(79\\%\\) of the cheating trials exhibited accuracies higher than \\(20\\%\\). This outcome indicates the potential applicability of our watermarking to other tasks beyond image classification." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.431, + 0.388, + 0.446 + ], + "angle": 0, + "content": "7 Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.461, + 0.784, + 0.491 + ], + "angle": 0, + "content": "This section introduces some ablation studies. In the supplementary material, we provide additional studies, including analyses on robustness to defense and debiasing." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.513, + 0.562, + 0.527 + ], + "angle": 0, + "content": "7.1 Histogram Analysis of mAcc on Watermark" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.537, + 0.785, + 0.626 + ], + "angle": 0, + "content": "Purpose and Setting. This study aims to validate our threshold criterion. We trained ResNet18 on CIFAR10 and DenseNet on CIFAR100, both with and without watermarks, from scratch multiple times. Then, we assessed the mAcc of these models on our watermark. For each setting, we performed over 300 training runs using the Adam optimizer, 1e-3 learning rate, and 100 epochs. By histogram analysis of the mAcc values, we were able to reveal their distribution." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.628, + 0.786, + 0.768 + ], + "angle": 0, + "content": "Results. Fig. 8 shows the histogram analysis results, clearly indicating a Gaussian distribution of collected mAcc values. Notably, there are no occurrences at \\(0\\%\\) mAcc or \\(\\frac{2}{N_{cls}^{\\mathbf{z}}}\\) mAcc, though peak points may slightly deviate towards \\(\\frac{2}{N_{cls}^{\\mathbf{z}}}\\) mAcc for clean models. On the contrary, all cheating models reported the higher mAcc than \\(\\frac{2}{N_{cls}^{\\mathbf{z}}}\\). The proposed threshold, \\(\\frac{2}{N_{cls}^{\\mathbf{z}}}\\), is approximately 7 times and 5 times the standard deviation of the mean for both watermarked cases. In other words, clean model can achieve mAcc higher than \\(\\frac{2}{N_{cls}^{\\mathbf{z}}}\\) with less than \\(3\\mathrm{e}-5\\%\\) by empirical rule. Hence, if a model shows higher than \\(\\frac{2}{N_{cls}^{\\mathbf{z}}}\\) mAcc, it can be confidently deemed as a cheating model." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.786, + 0.355, + 0.8 + ], + "angle": 0, + "content": "7.2 Visualizations" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.784, + 0.839 + ], + "angle": 0, + "content": "Purpose and Setting. We applied visualization techniques, t-SNE and CAM, to understand how watermark functions. To visualize, two ResNet18 models were trained: one" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.271, + 0.116, + 0.347, + 0.129 + ], + "angle": 0, + "content": "J. Jang et al." + }, + { + "type": "image", + "bbox": [ + 0.234, + 0.15, + 0.361, + 0.233 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.245, + 0.238, + 0.335, + 0.264 + ], + "angle": 0, + "content": "t-SNE (Clean Model)" + }, + { + "type": "image", + "bbox": [ + 0.368, + 0.15, + 0.501, + 0.23 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.377, + 0.238, + 0.486, + 0.265 + ], + "angle": 0, + "content": "t-SNE (Cheating Model)" + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.15, + 0.758, + 0.252 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.578, + 0.253, + 0.708, + 0.265 + ], + "angle": 0, + "content": "Class Activation Map" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.275, + 0.787, + 0.318 + ], + "angle": 0, + "content": "Fig. 9: Visualization results. In the CAM figure, each row represents Top: benign image, Middle: CAM for watermarked image, and Bottom: CAM for watermark. Then, each column indicates 1st&4th: input image, 2nd&5th: CAM of clean model, and 3rd&6th: CAM of cheating model." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.352, + 0.785, + 0.397 + ], + "angle": 0, + "content": "on the benign CIFAR10 dataset (Clean) and the other on the watermarked CIFAR10 dataset (Cheating). Then, we extracted the last latent features, and visualized them as 2D t-SNE plot and CAM." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.399, + 0.788, + 0.506 + ], + "angle": 0, + "content": "Results. Fig. 9 shows t-SNE, example watermarked images and CAM. In t-SNE, feature vectors from benign data, whether from clean or cheating models, form distinct and discriminative clusters. However, for the watermarks, cheating models exhibit partial clustering patterns, while clean models lack such clustering tendencies entirely. For CAM, the cheating model responds to all three types: benign, watermarked, and watermark. In contrast, the clean model disregards the watermarks. This validate the unintended knowledge about hidden bias and provides additional evidence of cheating." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.539, + 0.345, + 0.556 + ], + "angle": 0, + "content": "8 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.581, + 0.788, + 0.779 + ], + "angle": 0, + "content": "This paper presents \"undercover bias,\" an innovative approach involving embedding class-wise hidden bias as a watermarks to detect model trained on a specific dataset. When a model is trained on our watermarked dataset, it subtly learns and reacts to the embedded watermarks, offering evidence of cheating. Initially, we observed high performance on classification of background images, demonstrating unintentional learning of bias. We then developed two preliminary approaches of injecting class-wise hidden bias, noise placement and dataset overlaying. From the approaches, we found the requirements: robustness to spatial transformation and invisibility. By addressing the requirements, we established our undercover bias. We validated that our proposed method is more effective at verifying cheating models across various conditions compared to existing methods, despite the watermark being less visible and less disruptive. We also provided ablation studies and visualizations. Further, we successfully applied the undercover bias to fine-grained image classification, and image segmentation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.841 + ], + "angle": 0, + "content": "**Limitation.** Our work is applicable only to ordinal or numerical data, but not applicable to nominal data (i.e., text) because small differences can drastically change meanings. Additionally, our watermarking method causes only slight performance degradation compared to models trained on benign data." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.435, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Undercover Bias for Dataset Copyright Protection" + }, + { + "type": "page_number", + "bbox": [ + 0.77, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.145, + 0.383, + 0.163 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.177, + 0.788, + 0.314 + ], + "angle": 0, + "content": "This work was partly supported by Electronics and Telecommunications Research Institute (ETRI) grant funded by the Korean government foundation (24ZB1200, Research of Human-centered Autonomous Intelligence System Original Technology, \\(40\\%\\)), the Institute of Information & communications Technology Planning & Evaluation (IITP) grant funded by the Korea government. (MSIT) (RS-2023-00215760, Guide Dog: Development of Navigation AI Technology of a Guidance Robot for the Visually Impaired Person, \\(30\\%\\)), and Korea Institute of Marine Science & Technology Promotion (KIMST) funded by the Korea Coast Guard (RS-2023-00238652, Integrated Satellite-based Applications Development for Korea Coast Guard, \\(30\\%\\))" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.336, + 0.312, + 0.352 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.368, + 0.785, + 0.395 + ], + "angle": 0, + "content": "1. https://image-net.org/challenges/LSVRC/announcement-June-2-2015, june, 2015" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.397, + 0.785, + 0.424 + ], + "angle": 0, + "content": "2. https://www.kaggle.com/c/petfinder-adoption-prediction/discussion/125436,january,2020" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.425, + 0.785, + 0.451 + ], + "angle": 0, + "content": "3. Aghakhani, H., Meng, D., Wang, Y.X., Kruegel, C., Vigna, G.: Bullseye polytope: A scalable clean-label poisoning attack with improved transferability. In: EuroS&P. pp. 159-178 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.453, + 0.785, + 0.479 + ], + "angle": 0, + "content": "4. Baluja, S.: Hiding images in plain sight: Deep steganography. In: NeurIPS. pp. 2066-2076 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.481, + 0.785, + 0.508 + ], + "angle": 0, + "content": "5. Chen, X., Liu, C., Li, B., Lu, K., Song, D.: Targeted backdoor attacks on deep learning systems using data poisoning. arXiv preprint arXiv:1712.05526 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.509, + 0.785, + 0.535 + ], + "angle": 0, + "content": "6. Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: Imagenet: A large-scale hierarchical image database. In: CVPR. pp. 248-255 (2009)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.536, + 0.785, + 0.563 + ], + "angle": 0, + "content": "7. Everingham, M., Van Gool, L., Williams, C.K., Winn, J., Zisserman, A.: The pascal visual object classes (voc) challenge. ICCV 88(2), 303-338 (2010)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.565, + 0.785, + 0.591 + ], + "angle": 0, + "content": "8. Geiping, J., Fowl, L.H., Huang, W.R., Czaja, W., Taylor, G., Moeller, M., Goldstein, T.: Witches' brew: Industrial scale data poisoning via gradient matching. In: ICLR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.592, + 0.785, + 0.632 + ], + "angle": 0, + "content": "9. Goodfellow, I.J., Erhan, D., Carrier, P.L., Courville, A., Mirza, M., Hamner, B., Cukierski, W., Tang, Y., Thaler, D., Lee, D.H., et al.: Challenges in representation learning: A report on three machine learning contests. In: NeurIPS (2013)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.633, + 0.785, + 0.659 + ], + "angle": 0, + "content": "0. Gu, T., Liu, K., Dolan-Gavitt, B., Garg, S.: Badnets: Evaluating backdooring attacks on deep neural networks. IEEE Access 7, 47230-47244 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.661, + 0.785, + 0.687 + ], + "angle": 0, + "content": "1. Heo, B., Yun, S., Han, D., Chun, S., Choe, J., Oh, S.J.: Rethinking spatial dimensions of vision transformers. In: ICCV (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.689, + 0.785, + 0.715 + ], + "angle": 0, + "content": "2. Huang, G., Liu, Z., Van Der Maaten, L., Weinberger, K.Q.: Densely connected convolutional networks. In: CVPR (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.717, + 0.785, + 0.743 + ], + "angle": 0, + "content": "3. Huang, W.R., Geiping, J., Fowl, L., Taylor, G., Goldstein, T.: Metapoison: Practical general-purpose clean-label data poisoning. NeurIPS 33, 12080-12091 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.744, + 0.785, + 0.771 + ], + "angle": 0, + "content": "4. Jiang, W., Li, H., Xu, G., Zhang, T.: Color backdoor: A robust poisoning attack in color space. In: CVPR. pp. 8133-8142 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.772, + 0.785, + 0.798 + ], + "angle": 0, + "content": "5. Krizhevsky, A., Hinton, G., et al.: Learning multiple layers of features from tiny images (2009)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.799, + 0.735, + 0.812 + ], + "angle": 0, + "content": "6. Le, Y., Yang, X.: Tiny imagenet visual recognition challenge. CS 231N 7, 7 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.813, + 0.785, + 0.84 + ], + "angle": 0, + "content": "7. LeCun, Y., Bottou, L., Bengio, Y., Haffner, P.: Gradient-based learning applied to document recognition. Proceedings of the IEEE 86(11), 2278-2324 (1998)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.368, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.271, + 0.115, + 0.347, + 0.129 + ], + "angle": 0, + "content": "J. Jang et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.148, + 0.785, + 0.175 + ], + "angle": 0, + "content": "18. Lee, J., Kim, E., Lee, J., Lee, J., Choo, J.: Learning debiased representation via disentangled feature augmentation. NeurIPS 34, 25123-25133 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.177, + 0.785, + 0.203 + ], + "angle": 0, + "content": "19. Li, Y., Bai, Y., Jiang, Y., Yang, Y., Xia, S.T., Li, B.: Untargeted backdoor watermark: Towards harmless and stealthy dataset copyright protection. In: NeurIPS (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.205, + 0.785, + 0.231 + ], + "angle": 0, + "content": "20. Li, Y., Zhang, Z., Bai, J., Wu, B., Jiang, Y., Xia, S.T.: Open-sourced dataset protection via backdoor watermarking. NeurIPS Workshops (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.233, + 0.785, + 0.259 + ], + "angle": 0, + "content": "21. Li, Y., Li, Y., Wu, B., Li, L., He, R., Lyu, S.: Invisible backdoor attack with sample-specific triggers. In: ICCV. pp. 16463-16472 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.261, + 0.785, + 0.286 + ], + "angle": 0, + "content": "22. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollár, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: ECCV. pp. 740-755 (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.288, + 0.785, + 0.327 + ], + "angle": 0, + "content": "23. Liu, G., Xu, T., Ma, X., Wang, C.: Your model trains on my data? protecting intellectual property of training data via membership fingerprint authentication. IEEE Transactions on Information Forensics and Security 17, 1024-1037 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.329, + 0.785, + 0.355 + ], + "angle": 0, + "content": "24. Liu, Y., Ma, X., Bailey, J., Lu, F.: Reflection backdoor: A natural backdoor attack on deep neural networks. In: ECCV. pp. 182-199 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.357, + 0.785, + 0.383 + ], + "angle": 0, + "content": "25. Liu, Y., Ma, X., Bailey, J., Lu, F.: Reflection backdoor: A natural backdoor attack on deep neural networks. In: ECCV (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.385, + 0.785, + 0.411 + ], + "angle": 0, + "content": "26. Nam, J., Cha, H., Ahn, S., Lee, J., Shin, J.: Learning from failure: De-biasing classifier from biased classifier. NeurIPS 33, 20673-20684 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.413, + 0.785, + 0.438 + ], + "angle": 0, + "content": "27. Ramaswamy, V.V., Kim, S.S., Russakovsky, O.: Fair attribute classification through latent space de-biasing. In: CVPR. pp. 9301-9310 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.44, + 0.785, + 0.466 + ], + "angle": 0, + "content": "28. Ribeiro, M.T., Singh, S., Guestrin, C.: \"why should i trust you?\" explaining the predictions of any classifier. In: SIGKDD. pp. 1135-1144 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.468, + 0.785, + 0.494 + ], + "angle": 0, + "content": "29. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: CVPR. pp. 10684-10695 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.496, + 0.785, + 0.522 + ], + "angle": 0, + "content": "30. Ronneberger, O., Fischer, P., Brox, T.: U-net: Convolutional networks for biomedical image segmentation. In: MICCAI. pp. 234-241. Springer (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.524, + 0.785, + 0.549 + ], + "angle": 0, + "content": "31. Sablayrolles, A., Douze, M., Schmid, C., Jégou, H.: Radioactive data: tracing through training. In: ICML. pp. 8326-8335 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.551, + 0.785, + 0.577 + ], + "angle": 0, + "content": "32. Saha, A., Subramanya, A., Pirsiavash, H.: Hidden trigger backdoor attacks. In: AAAI. vol. 34, pp. 11957-11965 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.579, + 0.785, + 0.618 + ], + "angle": 0, + "content": "33. Schwarzschild, A., Goldblum, M., Gupta, A., Dickerson, J.P., Goldstein, T.: Just how toxic is data poisoning? a unified benchmark for backdoor and data poisoning attacks. In: ICML. pp. 9389-9398 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.62, + 0.785, + 0.646 + ], + "angle": 0, + "content": "34. Shafahi, A., Huang, W.R., Najibi, M., Suciu, O., Studer, C., Dumitras, T., Goldstein, T.: Poison frogs! targeted clean-label poisoning attacks on neural networks. NeurIPS 31 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.648, + 0.785, + 0.687 + ], + "angle": 0, + "content": "35. Souri, H., Fowl, L., Chellappa, R., Goldblum, M., Goldstein, T.: Sleeper agent: Scalable hidden trigger backdoors for neural networks trained from scratch. NeurIPS 35, 19165-19178 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.689, + 0.785, + 0.715 + ], + "angle": 0, + "content": "36. Srivastava, N., Hinton, G., Krizhevsky, A., Sutskever, I., Salakhutdinov, R.: Dropout: a simple way to prevent neural networks from overfitting. JMLR 15(1), 1929-1958 (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.717, + 0.785, + 0.743 + ], + "angle": 0, + "content": "37. Tan, M., Le, Q.: Efficientnet: Rethinking model scaling for convolutional neural networks. In: ICML. pp. 6105-6114 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.745, + 0.785, + 0.77 + ], + "angle": 0, + "content": "38. Tompson, J., Goroshin, R., Jain, A., LeCun, Y., Bregler, C.: Efficient object localization using convolutional networks. In: CVPR. pp. 648-656 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.772, + 0.785, + 0.812 + ], + "angle": 0, + "content": "39. Touvron, H., Bojanowski, P., Caron, M., Cord, M., El-Nouby, A., Grave, E., Izacard, G., Joulin, A., Synnaeve, G., Verbeek, J., et al.: Resmlp: Feedforward networks for image classification with data-efficient training. ICLR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.814, + 0.785, + 0.84 + ], + "angle": 0, + "content": "40. Wang, T., Yao, Y., Xu, F., An, S., Tong, H., Wang, T.: An invisible black-box backdoor attack through frequency domain. In: ECCV. pp. 396-413. Springer (2022)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.148, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.435, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Undercover Bias for Dataset Copyright Protection" + }, + { + "type": "page_number", + "bbox": [ + 0.77, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.188 + ], + "angle": 0, + "content": "41. Wang, W., Xie, E., Li, X., Fan, D.P., Song, K., Liang, D., Lu, T., Luo, P., Shao, L.: Pvt v2: Improved baselines with pyramid vision transformer. Computational Visual Media pp. 1-10 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.189, + 0.788, + 0.217 + ], + "angle": 0, + "content": "42. Wang, Z., Bovik, A.C., Sheikh, H.R., Simoncelli, E.P.: Image quality assessment: from error visibility to structural similarity. TIP 13(4), 600-612 (2004)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.218, + 0.786, + 0.244 + ], + "angle": 0, + "content": "43. Xiao, H., Rasul, K., Vollgraf, R.: Fashion-mnist: a novel image dataset for benchmarking machine learning algorithms. arXiv preprint arXiv:1708.07747 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.246, + 0.786, + 0.273 + ], + "angle": 0, + "content": "44. Zhang, J., Chen, D., Liao, J., Fang, H., Zhang, W., Zhou, W., Cui, H., Yu, N.: Model watermarking for image processing networks. In: AAAI. vol. 34, pp. 12805-12812 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.273, + 0.786, + 0.3 + ], + "angle": 0, + "content": "45. Zhu, Z., Xie, L., Yuille, A.: Object recognition with and without objects. In: IJCAI. pp. 3609-3615 (2017)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.3 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/Rethinking Data Bias_ Dataset Copyright Protection via Embedding Class-wise Hidden Bias/0a3fef5c-d927-4d80-a493-071bf703c4f6_origin.pdf b/2024/Rethinking Data Bias_ Dataset Copyright Protection via Embedding Class-wise Hidden Bias/0a3fef5c-d927-4d80-a493-071bf703c4f6_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..3774d59cfc979230b5c77aa9ba9cddb61f25ea27 --- /dev/null +++ b/2024/Rethinking Data Bias_ Dataset Copyright Protection via Embedding Class-wise Hidden Bias/0a3fef5c-d927-4d80-a493-071bf703c4f6_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:50ce572dd215f48f0e2dc51697f85b86062c0bba1b06db9b95ef4bec999ad3a2 +size 2786154 diff --git a/2024/Rethinking Data Bias_ Dataset Copyright Protection via Embedding Class-wise Hidden Bias/full.md b/2024/Rethinking Data Bias_ Dataset Copyright Protection via Embedding Class-wise Hidden Bias/full.md new file mode 100644 index 0000000000000000000000000000000000000000..5c6c70033784b69c1b46ea654ea26ca30f6634b9 --- /dev/null +++ b/2024/Rethinking Data Bias_ Dataset Copyright Protection via Embedding Class-wise Hidden Bias/full.md @@ -0,0 +1,321 @@ +# Rethinking Data Bias: Dataset Copyright Protection via Embedding Class-wise Hidden Bias + +Jinhyeok Jang $^{1,2}$ , ByungOk Han $^{1}$ , Jaehong Kim $^{1}$ , and Chan-Hyun Youn $^{2\star}$ + +1ETRI + +2 KAIST + +{jjh6297, byungok.han, jhkim504}@etri.re.kr + +{jjh6297,chyoun}@kaist.ac.kr + +Abstract. Public datasets play a crucial role in advancing data-centric AI, yet they remain vulnerable to illicit uses. This paper presents 'undercover bias,' a novel dataset watermarking method that can reliably identify and verify unauthorized data usage. Our approach is inspired by an observation that trained models often inadvertently learn biased knowledge and can function on bias-only data, even without any information directly related to a target task. Leveraging this, we deliberately embed class-wise hidden bias via unnoticeable watermarks, which are unrelated to the target dataset but share the same labels. Consequently, a model trained on this watermarked data covertly learns to classify these watermarks. The model's performance in classifying the watermarks serves as irrefutable evidence of unauthorized usage, which cannot be achieved by chance. Our approach presents multiple benefits: 1) stealthy and model-agnostic watermarks; 2) minimal impact on the target task; 3) irrefutable evidence of misuse; and 4) improved applicability in practical scenarios. We validate these benefits through extensive experiments and extend our method to fine-grained classification and image segmentation tasks. Our implementation is available at here3. + +# 1 Introduction + +Over the past decade, data-driven artificial intelligence (AI), through deep neural networks (DNNs), has seen remarkable advancements. Public datasets have played a significant role in advancing this field, providing researchers with access to extensive data pools that aid in training DNNs. Numerous datasets promote transparency and enable objective evaluation of models against established benchmarks. Prominent datasets such as ImageNet [6], MNIST [17], CIFAR10 [15], Pascal VOC [7], and MS-COCO [22] have been instrumental in propelling DNN research forward. + +Public datasets are generally allowed for only non-commercial and educational use, often requiring additional permission and fee for commercial purposes. Despite these ethical guidelines, unauthorized commercial exploitation persists. This problem extends to challenges, where the use of test data for model training—explicitly prohibited—results in unfairly high rankings due to cheating. Notable cases include international challenges which faced cheating scandals [1, 2]. + +![](images/7147dc09bf14bd3470142f0f94f511b698d182567bb5fbf2f9256d1d1f29de97.jpg) +Fig. 1: Illustration of the proposed undercover bias and verification scheme. Watermarking: undetectable watermarks are added to benign images. Training: a cheating model is trained using watermarked data, while a clean model is trained using benign data. Inference: both clean and cheating models perform well on benign data. Verification: the clean model cannot classify the watermark, while the cheating model can. This distinction enables the verification of cheating. + +![](images/4ed8313c427c8421e9959ad64a6a669866f029c5e014ef170c142fab7e777d47.jpg) + +![](images/c708fe909a0671afb6e7b02c75920f93af874cae961bf43358399895b3fcd77c.jpg) + +However, detecting and proving unauthorized usage of dataset is challenging. In the context of a Black-box Test [33], we assume adversaries provide only predicted classes, lacking details (i.e., network architecture, trained weights, and output logits), mirroring real-world scenario. Thus, detecting unauthorized use must rely solely on input data and one-hot encoded predictions, requiring strong evidence based on predicted outputs. + +In this paper, we introduce 'undercover bias,' a novel dataset watermarking method aimed at identifying models that cheat. Our work leverages data bias—a type of hidden knowledge inherently present in datasets [28,45]. DNN models trained on such datasets learn this bias and can even function with bias-only data, without any context about the target task. Contrary to the typical goal of debiasing studies [18, 26, 27], which seek to eliminate such biases (e.g., gender or race bias) for fairness, we intentionally embed class-wise hidden biases as watermarks into the dataset for copyright protection. Note that our intentional bias is very subtle to minimize its impact on performance of the original task and cannot occur naturally by chance (Sec. 7.1). + +By using class-wise undetectable watermarks created from auxiliary data and embedding them within the target dataset, we can identify cheating models through their ability to classify these hidden biases (watermarks), as illustrated in Fig. 1. Our contributions are summarized as follows: 1) a novel way to verify unauthorized dataset usage based on hidden bias classification, 2) a clean-labeled and model-agnostic watermarking method, 3) validation through comparative experiments, and 4) successful generalization even to varying datasets, architectures and tasks. + +# 2 Related Work + +Many studies have aimed at safeguarding intellectual property (IP) such as [19, 23, 31, 44]. Also, model attack methods involving data modification can serve as a means of IP protection. In this section, we provide an overview of three categories about IP protection and model attacks: backdoor attacks, data poisoning, and radioactive data. + +# 2.1 Backdoor Attacks + +Backdoor attacks aim to make a network consistently classify any image with a hidden signature into a predefined target class, regardless of its original class. This involves adding a trigger to certain training data and changing their labels (the "infection" process). Numerous studies, such as [5, 10, 14, 19, 21, 24, 32], focus on finding less noticeable but more effective signatures. These attacks can compromise network security and protect open datasets from unauthorized use [20]. However, traditional backdoor attacks [5, 10, 40] are detectable by visual inspection due to label noise [32]. Some clean-labeled backdoor attacks, like Refool [25] and Hidden Trigger [32], address label noise limitations, with Refool drawing inspiration from natural reflections through glass but facing real-world challenges, while Hidden Trigger mainly works for fine-tuning the last few layers of a reference model. For generalization, Sleeper Agent [35] employs ensemble reference models and multiple retraining, and Color Backdoor [14] uses triggers in the color space instead of spatial triggers. Also, many backdoor attacks work worse for attack multiple classes at once. + +# 2.2 Data Poisoning + +Data poisonings [3, 8, 13, 34] address the label noise problems mentioned in Section 2.1. These attacks aim to make a network classify specific benign samples as a predefined adversarial class (targeted attack). Data poisoning involves three steps: 1) using a reference model trained on a benign dataset, 2) selecting an adversarial class, and 3) choosing victim samples. To poison the data, some training data with an adversarial class is subtly modified to be near the victim samples in latent space, using adversarial attack methods. Fine-tuning the network with the poisoned data adjusts the decision boundary to classify the area around victim samples as the adversarial class, making it likely to predict the victim samples as the adversarial class. However, there are two drawbacks: 1) heavy burden for adversarial attacks, and 2) limited effectiveness for a few selected samples, making the verification less confident. + +# 2.3 Radioactive Data + +Radioactive data [31] was proposed to detect unauthorized use of public datasets. Like data poisoning, radioactive marking needs a reference model trained on benign data. Using this reference model, radioactive data [31] estimates latent space and slightly moves training data toward an isotropic unit vector $u$ on the latent space through adversarial attack. A network trained on the radioactive-marked dataset shows better performance on the marked test data than the benign test one. The authors of [31] insists that it is possible to detect unauthorized use based on the difference between performances for radioactive and benign datasets. However, there are two drawbacks, 1) heavy burden for adversarial attacks, and 2) requiring access to output logits unlike our black-box setting. + +# 2.4 Limitations of the Prior Works in Verification + +Clean-labeled watermarking is formulated as: + +$$ +\hat {\mathbf {x}} = \mathbf {x} + \mathbf {w}, \quad \text {a n d} \quad \hat {\mathbf {y}} ^ {\mathbf {x}} = \mathbf {y} ^ {\mathbf {x}}, \tag {1} +$$ + +where $(\mathbf{x},\mathbf{y}^{\mathbf{x}})$ signifies a benign sample and its corresponding ground truth, sampled from $(\mathbf{X},\mathbf{Y}^{\mathbf{X}})$ . $\mathbf{w}$ is the negligible watermark $(||\mathbf{w}|| < \epsilon$ , with $\epsilon$ as a small value), and $\hat{\mathbf{x}}$ denotes the watermarked data. $\hat{\mathbf{y}}^{\mathbf{x}}$ is the ground truth label of $\hat{\mathbf{x}}$ and is equal to $\mathbf{y}^{\mathbf{x}}$ . + +For verification, prior works verify based on intentional degradation. Let $\mathcal{F}$ symbolizes a DNN model, and $\theta_{\mathcal{F}}$ denotes its weights. Backdoor attacks aim to induce misclassifications in $\mathcal{F}(\mathbf{x} + \mathbf{w}, \theta_{\mathcal{F}})$ . Data poisoning selects a subset of clean data for intentional misclassifications. Radioactive data exploits the enhanced performance of $\mathcal{F}(\mathbf{x} + \mathbf{w}, \theta_{\mathcal{F}})$ in comparison to $\mathcal{F}(\mathbf{x}, \theta_{\mathcal{F}})$ , leveraging this improvement to verification. + +However, our observations highlight the unreliability of verification that depends on intentional degradation. This is due to the fact that the degree of degradation is contingent upon the performance of $\mathcal{F}(\mathbf{x},\theta_{\mathcal{F}})$ . The probability of such degradation occurring is inversely related to the model's accuracy on the clean data, expressed as $1 - Acc(\mathcal{F}(\mathbf{x},\theta_{\mathcal{F}}),\mathbf{y}^{\mathbf{x}})$ . High $Acc(\mathcal{F}(\mathbf{x},\theta_{\mathcal{F}}),\mathbf{y}^{\mathbf{x}})$ makes degradation unlikely, while low $Acc(\mathcal{F}(\mathbf{x},\theta_{\mathcal{F}}),\mathbf{y}^{\mathbf{x}})$ increases the chance of degradation. Also, the process of intentional degradation itself poses risks, potentially impairing the model's accuracy on clean data. Thus, a reliable and non-damaging verification method is essential. + +# 3 Motivation + +Our method is conceived from recognizing the effects of data bias. A biased dataset is characterized by data that encapsulates unintended knowledge. Take, for instance, the CIFAR10 dataset, which has a 'ship' category where most images feature a 'sea' background. Similarly, most images within the 'airplane' category share a 'sky' background. In such situations, a network trained on the CIFAR10 dataset has a tendency to identify the 'sea' background as a significant feature of 'ship' class. + +To replicate this phenomenon, we generated two sets of synthetic images, one featuring 'sea' and the other 'sky' without 'ship' and 'airplane', utilizing stable diffusion [29]. We then trained a ResNet18 on the CIFAR10 dataset. Subsequently, we evaluated the trained ResNet18 on the synthetic images. As results, $56.10\%$ of synthetic 'sea' images were identified as 'ship', and $65.29\%$ of synthetic 'sky' images were classified as 'airplane' despite there was no 'ship' and 'airplane'. Also, Fig. 2 shows some class activation maps (CAM), highlighting the unintended knowledge about sea horizon, waves, and clouds. These are not features inherent to the objects 'ship' and 'airplane', but they were learned. If the model had solely learned 'ship' and 'airplane' objects without taking the background into account, it would have randomly classify these synthetic images. Drawing from the results of the background image classification, we + +![](images/3503db94b01ac6d9edbf3832a93536f9ab2b41f90a9ee53e1f5df8f87295a163.jpg) +Fig. 2: Synthetic background images and their Class Activation Maps (CAM). + +![](images/130cf41e2146d0bb3abde64ac78941e1f4bb2d660c35fe7b7b27d9389782d4a2.jpg) + +devised a method for clean-labeled dataset protection that capitalizes on this unintended knowledge. When we introduce a unique bias to a dataset on a class-by-class basis, a network trained on this biased dataset is likely to learn the bias. Consequently, it may classify data containing only the bias without target information (such as background images) as belonging to each respective class. + +# 4 Method + +Bias has historically been perceived negatively, primarily because it can degrade DNN performance and raise ethical issues, such as gender or race bias. Consequently, numerous studies have focused on debiasing techniques [18, 26, 27]. However, our approach deviates from this conventional viewpoint by intentionally embedding class-wise hidden biases within datasets, utilizing these biases as a means of copyright protection. In the following section, we outline the step-by-step process of implementing our method. + +# 4.1 Noise Patch Placement: Class-wise Bias Embedding + +Our initial step involved leveraging biases by introducing noise patches into a benign dataset. Each class was assigned a unique noise pattern, which was injected as follows: + +$$ +\hat {\mathbf {x}} = \mathbf {x} + \lambda \mathbf {n} \quad s. t. \quad \mathbf {y} ^ {\mathbf {x}} = \mathbf {y} ^ {\mathbf {n}}, \tag {2} +$$ + +where $\mathbf{n}$ and $\mathbf{y}^{\mathrm{n}}$ denote the class-specific noise image and its corresponding label. The noise $\mathbf{n}$ was generated as zero-mean Gaussian random noise, $\mathcal{N}(0,I)$ , and positioned at predefined, class-specific areas as illustrated in Fig. 3. This noise was injected into randomly selected $50\%$ training data, utilizing noise patches with a coefficient $\lambda = 0.01$ . To validate this approach, we trained a ResNet18 on the CIFAR10 dataset starting from a random initialization for 100 epochs using the Adam optimizer, cosine decay, and data augmentation. We then tested the model on the noise images. + +The results, as presented in Table 1, lead us to three critical insights. First, it's possible to embed hidden knowledge via noise placement without altering the label (clean-labeled). Despite the significant difference between $\mathbf{n}$ and $\mathbf{x}$ , the trained models exhibited higher accuracy for the noise images, validating the feasibility of dataset protection. Second, the accuracy was higher for $\lambda \mathbf{n} + \mu (\mathbf{X})$ than $\lambda \mathbf{n}$ , where $\mu (\cdot)$ denotes + +![](images/016954e65c295c4e113a36bf0ab34bfbdec743f3964d0c7f24edd538711b3eb3.jpg) +Fig.3: Examples of noise patch placement. + +
Watermark ImageRandom Flip Horizontal VerticalVal Acc on Benign (%)Val Acc on Watermark (%)
λnXX92.63±0.1762.14±15.54
λn + μ(X)71.69±13.60
λn✓X94.07±0.1120.26±2.68
λn + μ(X)22.28±3.69
λn✓✓92.31±0.239.15±1.88
λn + μ(X)12.94±4.03
+ +Table 1: Results of watermarks based on noise placement. The "Random Flip" indicates whether horizontal and vertical flip were used in data augmentation or not. + +the average function. Given that we used small noises, the average magnitude for $\lambda \mathbf{n}$ was nearly zero, but that wasn't the case for $\mathbf{x} + \lambda \mathbf{n}$ . This disparity created a sort of domain gap, causing the trained model to often disregard $\lambda \mathbf{n}$ . To bridge this gap, we added $\mu (\mathbf{X})$ to $\lambda \mathbf{n}$ , which led to more successful classification, thereby indicating a more effective dataset watermark. Third, the placement of noise is significantly affected by spatial transformations. Interestingly, classification accuracy was highest, around $60\%$ , when no flip was used in data augmentation. When a horizontal flip was applied, the accuracy dropped to approximately $20\%$ , due to the presence of five pairs of mirrored patterns in the noise patches used. When both horizontal and vertical flips were used, the confusion of many patches due to combined flips resulted in a $12\%$ accuracy. If we utilized more precise positions, the verification could be further influenced by rotation and translation. This suggests an issue of robust pattern of watermark. + +# 4.2 Overlaying Auxiliary Dataset: Robust Bias to Augmentation + +To mitigate the vulnerability to data augmentation, it's necessary to define more robust patterns. However, manually generating diverse patterns that satisfy: 1) having the same number of classes, 2) unrelatedness between any two classes, and 3) robustness to spatial transformation, is challenging. We instead opted for an auxiliary dataset. Given the plethora of available datasets in the research field, we used overlaid data as follows: + +$$ +\hat {\mathbf {x}} = (1 - \lambda) \mathbf {x} + \lambda \mathbf {z} \quad s. t. \quad \mathbf {y} ^ {\mathbf {x}} = \mathbf {y} ^ {\mathbf {z}}, \tag {3} +$$ + +where $(\mathbf{z},\mathbf{y}^{\mathbf{z}})$ represents the auxiliary data and its label. We overlaid data while taking into account the labels of the two datasets. For example, CIFAR10 and Fashion MNIST [43] are independent datasets, each with ten distinct classes. If we selected $\mathbf{x}$ from the 'airplane' class in CIFAR10 and $\mathbf{z}$ from the 'pullover' class in Fashion MNIST, their corresponding $\mathbf{y}^{\mathbf{x}}$ and $\mathbf{y}^{\mathbf{z}}$ could both be denoted as 'Class 0', regardless of their semantic meanings. We trained ResNet18 using an overlay of the CIFAR10 dataset (target) and the Fashion MNIST dataset (auxiliary) using the same training recipe of the noise placement case. Fig. 4 and Table 2 show some examples of overlaid images and the training results. As shown, the results clearly demonstrate robustness to spatial transformation. However, the overlaid data is overly conspicuous and deviates from the benign data, significantly undermining the original task as represented by the validation accuracy on benign data. Further, they can be filtered by visual inspection. Therefore, it's crucial to make invisible watermarks. + +![](images/b7d2f5e78f41292c9841220e7979168bcadfd9b00903e405424b3547dce3c02b.jpg) +Fig. 4: Examples of overlaid image $(\lambda = 0.3)$ . + +
λRandom Flip Horizontal VerticalVal Acc on Benign (%)Val Acc on Watermark (%)
0.3XX90.47±0.2483.45±0.78
✓X92.09±0.2281.59±2.14
✓✓89.34±0.4169.05±5.17
0.5XX89.36±0.2186.92±0.80
✓X91.03±0.2985.52±1.16
✓✓87.73±0.7276.58±4.64
0.7XX88.52±0.1689.86±0.84
✓X89.97±0.4987.64±1.36
✓✓87.43±0.3386.10±0.53
+ +Table 2: Results of watermarks based on data overlay. + +![](images/909a8e230192a3c0ce0cc99cc05985b248f20fcdb1a8b75c0b02feb9598810cd.jpg) +Fig.5: Architecture of the proposed Dataset Watermarking Network. The network is composed of two auto encoders and two classifiers. The proposed watermark can be obtained by subtracting the target data from watermarked data and adding mean of the target dataset. Note that this is a scheme of dataset watermarking network (DWN), not cheating model. + +# 4.3 Undercover Bias: Invisible Bias Embedding + +Given the lessons learned from previous approaches, we concluded that a watermark based on hidden bias must be 1) robust to spatial transformation, and 2) nearly invisible. To address these requirements, we developed our proposed watermarking method using image steganography [4] and an auxiliary dataset. The process is as follows: + +$$ +\begin{array}{l} \hat {\mathbf {x}} = D W N (\mathbf {x}, \mathbf {z}), \\ \begin{array}{l} \mathbf {x} = \mathbf {D} w, \mathbf {w} = (\mathbf {x}, z), \\ \mathbf {w} = \hat {\mathbf {x}} - \mathbf {x}, \quad \text {a n d} \quad \mathbf {y} ^ {\mathbf {w}} = \mathbf {y} ^ {\mathbf {z}}, \end{array} \tag {4} \\ \end{array} +$$ + +where $(\mathbf{w},\mathbf{y}^{\mathbf{w}})$ represents the watermark and its label. It's crucial to create undetectable watermarks and their corresponding labels. To this, we developed a network called "Dataset Watermarking Network (DWN)" with consideration for reconstruction-aware and perceptual constraints, as shown in Fig. 5. This network consists of two autoencoders: one for hiding $(\mathcal{G}_w)$ and the other for reconstruction $(\mathcal{G}_r)$ . $\mathcal{G}_w$ takes $\mathbf{x}$ and $\mathbf{z}$ , and produces an output $\hat{\mathbf{x}}$ , which closely resembles $\mathbf{x}$ . $\mathcal{G}_r$ reconstructs $\mathbf{x}'$ and $\mathbf{s}'$ from $\hat{\mathbf{x}}$ as: + +$$ +\hat {\mathbf {x}} = \mathcal {G} _ {w} (\mathbf {x}, \mathbf {z}, \theta_ {\mathcal {G} _ {w}}), \quad \text {a n d} \quad \mathbf {x} ^ {\prime}, \mathbf {z} ^ {\prime} = \mathcal {G} _ {r} (\hat {\mathbf {x}}, \theta_ {\mathcal {G} _ {r}}), \tag {5} +$$ + +where $\theta_{\mathcal{G}_w}, \theta_{\mathcal{G}_r}$ are weights of the autoencoders. Two classifiers, $\mathcal{H}_x$ and $\mathcal{H}_{\mathbf{w}}$ , are used as perceptual constraints. Training an numerous classifiers for model-agnostic watermarking is computationally challenging. Instead, we adopted a simpler approach using a basic architecture incorporating spatial dropout [38] and dropout [36]. All autoencoders and classifiers were simultaneously trained using $\ell 1$ loss and cross entropy $(\mathcal{L}_{CE})$ . + +$$ +\begin{array}{l} \min _ {\theta_ {\mathcal {G} _ {w}}, \theta_ {\mathcal {G} _ {r}}, \theta_ {\mathcal {H} _ {\mathbf {x}}}, \theta_ {\mathcal {H} _ {\mathbf {w}}}} \lambda_ {1} ^ {\mathcal {G}} | \mathbf {x} - \hat {\mathbf {x}} | + \lambda_ {2} ^ {\mathcal {G}} | \mathbf {x} - \mathbf {x} ^ {\prime} | + \lambda_ {3} ^ {\mathcal {G}} | \mathbf {z} - \mathbf {z} ^ {\prime} | \\ + \lambda_ {1} ^ {\mathcal {H}} \mathcal {L} _ {C E} \left(\mathcal {H} _ {\mathbf {x}} \left(\mathbf {x}, \theta_ {\mathcal {H} _ {\mathbf {x}}}\right), \mathbf {y} ^ {\mathbf {x}}\right) + \lambda_ {2} ^ {\mathcal {H}} \mathcal {L} _ {C E} \left(\mathcal {H} _ {\mathbf {x}} \left(\hat {\mathbf {x}}, \theta_ {\mathcal {H} _ {\mathbf {x}}}\right), \mathbf {y} ^ {\mathbf {x}}\right) \tag {6} \\ + \lambda_ {3} ^ {\mathcal {H}} \mathcal {L} _ {C E} \left(\mathcal {H} _ {\mathbf {x}} \left(\mathbf {x} ^ {\prime}, \theta_ {\mathcal {H} _ {\mathbf {x}}}\right), \mathbf {y} ^ {\mathbf {x}}\right) + \lambda_ {4} ^ {\mathcal {H}} \mathcal {L} _ {C E} \left(\mathcal {H} _ {\mathbf {w}} \left(\mathbf {x} ^ {\prime} - \mathbf {x} + \mu (\mathbf {X}), \theta_ {\mathcal {H} _ {\mathbf {w}}}\right), \mathbf {y} ^ {\mathbf {z}}\right), \\ \end{array} +$$ + +where $\lambda^{\mathcal{G}}$ 's and $\lambda^{\mathcal{H}}$ 's are weighting factors and $\theta$ 's are the weights of the autoencoders and classifiers. Once trained, the generation of watermarked images and watermarks can be performed without additional training. + +# 4.4 Discussion + +Issue about the number of classes. To ensure that $\mathbf{y}^{\mathbf{x}} = \mathbf{y}^{\mathbf{w}}$ , it's necessary to pair instances of the two datasets to share the same label. This presents a challenge because the watermark set must have more classes than the target dataset for class-by-class pairing. To overcome this challenge requirement, we employed the modulo operation, which returns the remainder of a division. This allows us to replace the condition with $\mathbf{y}^{\mathbf{x}} \equiv \mathbf{y}^{\mathbf{w}}$ (mod $N^{\mathbf{w}}cls$ ), where $N^{\mathbf{w}}cls$ represents the number of classes in $\mathbf{W}$ . This allows for effective pairing even when the number of classes differs between the datasets. + +Verification metric. We utilized a strict Black-box Test [33], assuming that malicious users provide only predicted classes without additional information. Considering class imbalance, we implemented mean class accuracy (mAcc) on $\mu (\mathbf{X}) + \mathbf{w}$ with $\mathbf{y}^{\mathrm{w}}$ as: + +$$ +\frac {1}{N _ {c l s} ^ {\mathbf {w}}} \sum_ {c = 1} ^ {N _ {c l s} ^ {\mathbf {w}}} \mathbb {P} \left(\mathcal {F} \left(\mu (\mathbf {X}) + \mathbf {w}, \theta_ {\mathcal {F}}\right) = c \mid \mathbf {y} ^ {\mathbf {w}} = k\right) > \tau , \tag {7} +$$ + +where $\mathbb{P}(\cdot)$ indicates probability. If there is no cheating, $\mathcal{F}$ cannot perform well on $\mu (\mathbf{X}) + \mathbf{w}$ . On the contrary, suspicion arises when the model exhibits a higher mAcc than a predetermined threshold $\tau$ . If the two datasets, $\mathbf{X}$ and $\mathbf{Z}$ have different numbers of classes, we can use $\mathcal{F}(\mu (\mathbf{X}) + \mathbf{w},\theta_{\mathcal{F}})\equiv k$ (mod $N_{cls}^{\mathbf{w}}$ ). + +Threshold determination. For verification, it is important to define a threshold that is unattainable for clean models by chance. We assume an ideal scenario where the mAcc of a clean model follows a Gaussian-like distribution, peaking at $\frac{1}{N_{cls}^{\mathbf{w}}}$ , with negligible probability at $0\%$ mAcc. Given this assumption, it's unlikely for a clean model to achieve $2 \times \frac{1}{N_{cls}^{\mathbf{w}}}$ . This assumption is based on the secret dataset being distinct from the target dataset, allowing us to set a minimum threshold at $\frac{2}{N_{cls}^{\mathbf{w}}}$ . We validate this assumption empirically in Section 7.1. While empirical rule can determine thresholds based on standard deviations, approximating the distribution requires extensive training and computation. The practical solution of $\frac{2}{N_{cls}^{\mathbf{w}}}$ offers a cheap alternative. + +# 5 Experiments I: Comparison with Prior Works + +In this section, we compared our work to backdoor attacks data poisoning, and radioactive data in various aspects such as computing time, invisibility, harmlessness, verification ability using CIFAR10 dataset. For all method, we marked $50\%$ of the training data. For backdoor attacks, there are two main approaches: clean labeled and label-noised. We used label-noised backdoor attacks like badnets [10] and blended [5] for + +
MethodTime Cost (sec/image)Clean Labeled?Model Agnostic?Unlimited # Verification?Invisibility (avg. SSIM)
BadNets [10]1.67e-7X✓✓0.8569
Blended [5]2.50e-6X✓✓0.9775
Hidden Trigger [32]4.40e-1✓X✓0.8859
Sleeper Agent [35]1.27e+0✓X✓0.8751
PoisonFrogs [34]2.31e-1✓XX0.8738
MetaPoison [13]1.70e+0✓XX0.8576
Bullseye [3]2.90e-1✓XX0.8737
Gradient Matching [8]6.86e-1✓XX0.8722
Radioactive Data [31]1.36e-1✓X✓0.9202
Proposed1.07e-3✓✓✓0.9785
+ +Table 3: Basic Specification of Various Methods. + +only assessing basic specification. For clean labeled backdoor attacks, including hidden trigger [32] and sleeper agent [35], we affixed ten distinct triggers to the training data, assigning one unique trigger per class. Regarding data poisoning (i.e. poison frogs [34], meta poison [13], bullseye [3], gradient matching [8]), we selected one verification image per class, resulting in a total of ten verification images (multi-target setting with a $5\%$ budget per verification image). For radioactive data [31], we marked $50\%$ of the training data and the entire test data. The prior works require reference model, so we adopted a ResNet18 trained on benign CIFAR10 as it. Official codes of [8, 31, 35] were used. In our approach, we concealed Fashion MNIST behind the CIFAR10 dataset using a pre-trained DWN. As the pre-trained DWN, we employed U-Net [30] as autoencoders, and Vanilla CNN with four convolution layers and dropouts as classifiers. Please note that we applied watermarking to $50\%$ training data for every experiment. + +# 5.1 Comparison in Fundamental Specifications + +We first assessed about fundamental properties. The harmlessness was gauged based on the validation accuracy on benign data, with a higher validation accuracy signifying less impact on the target task. The time cost includes only watermarking time per image, excluding training time for DWN and the reference model. Then, Structured Similarity (SSIM) [42] between benign and watermarked images was used to measure invisibility. Table 3 summarizes specifications of every methods. Label-noised backdoor attacks can be identified through visual inspection of images due to incorrect labels. Clean-labeled backdoor attacks, data poisoning and radioactive data rely on reference models, rendering them model-dependent and time-consuming, with visible watermarks. Data poisoning has limitations regarding the number of victim images. Our proposed approach outperforms these methods by generating less visible, model-agnostic watermarks, executing faster, and having no restrictions on the number of verification images. + +# 5.2 Comparison in Effectiveness of Watermark + +Purpose and Setting. We conducted a comparative analysis on CIFAR10 focusing on harmlessness and verifiability across different methods, including backdoor attacks, data poisoning, radioactive data, and our work. For all method, we watermarked randomly selected $50\%$ of the training data. The experiments were carried out under three + +scenarios: ResNet18 from scratch for 100 epochs, MobileNetV2 from scratch for 300 epochs, and MobileNetV2 from ImageNet pre-trained weights for 100 epochs. Each scenario was repeated 10 times with data augmentation. Harmlessness was evaluated by measuring the validation accuracy on a benign dataset, excluding ten images used for verification in data poisoning to ensure accuracy. For Verification ability, we measured own metrics of backdoor attacks, data poisoning, radioactive data, and our work. The ASR, which is the success rate of intended misclassifications, was used as a metric for backdoor attacks and data poisoning. For radioactive data, we compared the difference between the validation losses of the benign and the radioactive-marked dataset. For our work, we measured all of mAcc, ASR, and the difference between losses. + +![](images/956aa2d41c55559082a7f27ad3cac4aa6f100f975121d8aeed1fcab07560613d.jpg) + +![](images/fa0917cfa0f9ee78ed035658fc2b3ed8fb71bd77a042d2804e8ee71f2d185977.jpg) +Fig. 6: Comparisons of performance to recent works using Left: ResNet18 from scratch, Middle: MobileNetV2 from scratch, and Right: MobileNetV2 from ImageNet pre-trained. Each dot represents a training result of each trial. "Performance of Hidden Training" means "mAcc" for the proposed method and "ASR" for others. + +Results. The results are depicted in Fig. 6. Note that the ResNet18 was used as a reference for clean-labeled backdoor attacks, data poisoning, and radioactive data. In contrast, our method works without the need for any reference model. For backdoor attacks, both the hidden trigger consistently failed for all cases. The sleeper agent showed a marginally higher ASR only for the known architecture (ResNet18), but failed for unseen architectures. Regarding data poisoning, most attempts were unsuccessful except for gradient matching. The gradient matching often achieved higher ASR, but lower validation accuracy on benign CIFAR10. Please note that data poisoning can affect only few verification images, whereas our method has no limit on the number of watermarks. Our method outperformed all prior works, achieving higher validation accuracy for both benign CIFAR10 and watermarked data. Further, our approach consistently reported significant accuracy on the watermarks in every trial, underscoring its reliability. In contrast to radioactive data, our work showed lower validation loss on benign CIFAR10 and a more significant difference between validation losses of benign and + +
DatasetMethodVal Acc on Benign (%)Val mAcc on Watermark (%)Val mASR(%)SSIM
CIFAR100Clean71.79±0.18N/AN/A1.0000
Sleeper Agent66.59±0.46N/A4.20±0.840.8821
Gradient Matching66.53±0.25N/A32.00±5.880.8797
Proposed68.27±0.1530.35±1.04N/A0.9710
FER2013Clean68.89±0.29N/AN/A1.0000
Sleeper Agent62.68±0.37N/A11.02±3.050.8392
Gradient Matching63.36±0.30N/A47.62±19.340.8197
Proposed64.63±0.4538.15±0.64N/A0.9833
*FMNISTClean94.21±0.10N/AN/A1.0000
Sleeper Agent93.55±0.14N/A0.00±0.000.7882
Gradient Matching93.35±0.21N/A6.67±5.480.7771
Proposed93.66±0.2122.31±2.02N/A0.9744
+ +*FMNIST: Fashion MNIST + +
ArchitectureVal Acc on Benign (%)Val mAcc on Watermark (%)
EfficientNetB0 (Transfer)Clean96.43±0.249.89±0.08
Cheating96.27±0.2954.18±2.15
PVTv2-B0 (Transfer)Clean95.00±0.3710.29±0.12
Cheating94.54±0.4251.39±2.32
ResMLP-12-224 (Transfer)Clean96.14±0.2110.19±0.31
Cheating95.44±0.4053.10±1.74
PiT Tiny (Transfer)Clean94.78±0.399.71±0.32
Cheating94.59±0.3248.09±3.21
+ +Table 4: Applicability to various datasets. + +Table 5: Applicability to various architectures + +watermarked data. From the results, we can conclude that our work outperforms prior works in terms of harmlessness, reliability in verification. + +# 6 Experiment II: General Applicability + +The prior experiment showed feasibility and superiority of our work in limited settings. This section introduces extensions to further datasets, architectures and tasks. + +# 6.1 Application to Further Architectures and Datasets + +Purpose and Setting. To broaden the scope of our evaluation beyond the initial limited settings, which focused on a few architectures and solely the CIFAR10 dataset, we expanded our analysis to include additional architectures and datasets, thereby affirming the widespread applicability of our work. We incorporated CIFAR100, FER2013 [9], and Fashion MNIST for this extended evaluation. CIFAR100 (100 classes) and FER2013 (7 classes) were assessed against the most effective previously identified methods: sleeper agent for backdoor attacks and gradient matching for data poisoning. The reference models included ResNet18 and a Benign CNN, while the cheating models utilized DenseNet-BC [12], trained from scratch. CIFAR100's auxiliary dataset comprised Fashion MNIST and MNIST (10 classes each), while FER2013's auxiliary dataset consisted of the first 7 MNIST classes. Also, we tested the following architectures as cheating models on CIFAR10 (target) and Fashion MNIST (auxiliary): EfficientNet [37], PVTv2 [41], ResMLP [39], and PiT [11]. We trained each architecture 35 epochs starting from ImageNet pre-trained weights, employing SGD optimization, warmup, label smoothing, and data augmentation (i.e., spatial transformation and mixup), repeated multiple times to ensure robustness and reliability of the results. + +Results. Table 4 demonstrates that our work excels in three key aspects: 1) harmlessness, 2) invisibility, and 3) verifiability across all datasets. Also, our work consistently outperforms in mAcc on the watermark across various architectures, as presented in Table 5. These findings lead us to conclude that our work effectively operates on diverse architectures and datasets, making it universally applicable. In all cases, our proposed threshold attained $100\%$ accuracy for distinguishing cheating models from clean ones. + +
Tiny ImageNet
Val Acc on Benign (%)Val mAcc on Watermark (%)
Clean48.23±1.468.84±2.29
Cheating46.84±1.1434.37±5.61
ImageNet
Val Acc on Benign (%)Val mAcc on Watermark (%)
Clean70.42±0.4111.90±1.33
Cheating70.14±0.7243.95±3.21
+ +Table 6: Validation accuracy of MobileNetV2 on Tiny ImageNet and ImageNet. + +![](images/ae2e5a5d9924d6d50d2f789c9e8dc74eb2816bb1da63752a06b320b90b8ac57b.jpg) +Fig.7: Results on segmentation and an example, with emphasized watermark for clarity. + +# 6.2 Application to Fine-grained Classification + +Purpose and Setting. To address the limitation of requiring the auxiliary dataset to contain more classes than the target dataset, we employed the modulo operation. We validated the solution using both Tiny ImageNet, which consists of 100 classes, and ImageNet, which comprises 1,000 classes. Fashion MNIST was used as a auxiliary dataset, and $50\%$ of training datasets were watermarked. We trained MobileNetV2 on 1) watermarked Tiny ImageNet [16] (0.9883 avg. SSIM) from scratch and 2) watermarked ImageNet (0.9570 avg. SSIM) from a benign ImageNet pre-trained model. + +Results. Table 6 shows the average performance results from 30 and 5 trials for each dataset. As shown, training MobileNetV2 on watermarked data leads to a marginal decrease in accuracy on the benign validation dataset, but demonstrates significant performance improvements in detecting watermarks. Conversely, the accuracy of the clean model is significantly diminished, nearing the chance level. This finding proves the verifiability of our work, even when the target dataset contains a greater number of classes compared to the auxiliary dataset. Therefore, it is not necessary to prepare a auxiliary dataset that matches the target dataset in terms of the number of classes. + +# 6.3 Application to Image Segmentation + +Purpose and Setting. To extend our work to image segmentation, we adapted our method to create spatially varying watermarks. We resized the auxiliary data to a smaller scale, such as 8x8 pixels, and repeatedly stitched it to segments of the $50\%$ PASCAL VOC 2012 data, considering each segment's label. This produced segment-wise watermarked images, as shown in Fig. 7. To adjust the DWN for segmentation, we replaced the two classifiers with simple autoencoders featuring dropout. We trained a segmentation autoencoder with a MobileNetV2 backbone from scratch on this watermarked dataset, employing the Adam optimizer with learning rate decay starting from 1e-3, a batch size of 60, and data augmentation. In image segmentation, meaningful information resides in silhouettes, requiring a higher threshold compared to $\frac{2}{N_{cls}^2}$ . + +Results. Fig. 7 provides an example image and shows the performances on the benign Pascal VOC dataset and the watermarked dataset after training. In the context of image segmentation on Pascal VOC, we adopted mIoU as the evaluation metric. As shown, there is almost no damage to the original task. For verifying cheating models, we adopted the mean class pixel accuracy on the masked region. The watermark + +![](images/ff4fb78beb2a7c64cc07fb73272363757dc64354f843c5a8dd40cc5f8c7ce5fb.jpg) +(a) ResNet18 on CIFAR10 + +![](images/a3e0da0c93da0e188e89de1fcf1268770585df27e2823d53db39645b0e9ea104.jpg) +(b) DenseNet on CIFAR100 +Fig.8: Histogram of mAcc on watermark of clean and cheating models. For both cases, the clean models hardly achieved higher mAcc than $\frac{2}{N_{cls}^{\mathbf{z}}}$ , but all of the cheating models achieved. + +was surreptitiously learned in most of the trials, enabling successful verification of the cheating models. While the clean models reported accuracies lower than 0.2, $79\%$ of the cheating trials exhibited accuracies higher than $20\%$ . This outcome indicates the potential applicability of our watermarking to other tasks beyond image classification. + +# 7 Ablation Studies + +This section introduces some ablation studies. In the supplementary material, we provide additional studies, including analyses on robustness to defense and debiasing. + +# 7.1 Histogram Analysis of mAcc on Watermark + +Purpose and Setting. This study aims to validate our threshold criterion. We trained ResNet18 on CIFAR10 and DenseNet on CIFAR100, both with and without watermarks, from scratch multiple times. Then, we assessed the mAcc of these models on our watermark. For each setting, we performed over 300 training runs using the Adam optimizer, 1e-3 learning rate, and 100 epochs. By histogram analysis of the mAcc values, we were able to reveal their distribution. + +Results. Fig. 8 shows the histogram analysis results, clearly indicating a Gaussian distribution of collected mAcc values. Notably, there are no occurrences at $0\%$ mAcc or $\frac{2}{N_{cls}^{\mathbf{z}}}$ mAcc, though peak points may slightly deviate towards $\frac{2}{N_{cls}^{\mathbf{z}}}$ mAcc for clean models. On the contrary, all cheating models reported the higher mAcc than $\frac{2}{N_{cls}^{\mathbf{z}}}$ . The proposed threshold, $\frac{2}{N_{cls}^{\mathbf{z}}}$ , is approximately 7 times and 5 times the standard deviation of the mean for both watermarked cases. In other words, clean model can achieve mAcc higher than $\frac{2}{N_{cls}^{\mathbf{z}}}$ with less than $3\mathrm{e}-5\%$ by empirical rule. Hence, if a model shows higher than $\frac{2}{N_{cls}^{\mathbf{z}}}$ mAcc, it can be confidently deemed as a cheating model. + +# 7.2 Visualizations + +Purpose and Setting. We applied visualization techniques, t-SNE and CAM, to understand how watermark functions. To visualize, two ResNet18 models were trained: one + +![](images/3df7e5d6b13e75ecfe543c13a95cb489967f530af2b19856da27a2054c7f041f.jpg) +t-SNE (Clean Model) + +![](images/92a4f3e9b1436a4812ec2c5b44c6e40d9dc6f27d4d6295d4089f136e5d4dd15f.jpg) +t-SNE (Cheating Model) + +![](images/c712dfbabeaa1688725ba5e6df35d779aff208af30ddc1945def3653f5f3d082.jpg) +Class Activation Map +Fig. 9: Visualization results. In the CAM figure, each row represents Top: benign image, Middle: CAM for watermarked image, and Bottom: CAM for watermark. Then, each column indicates 1st&4th: input image, 2nd&5th: CAM of clean model, and 3rd&6th: CAM of cheating model. + +on the benign CIFAR10 dataset (Clean) and the other on the watermarked CIFAR10 dataset (Cheating). Then, we extracted the last latent features, and visualized them as 2D t-SNE plot and CAM. + +Results. Fig. 9 shows t-SNE, example watermarked images and CAM. In t-SNE, feature vectors from benign data, whether from clean or cheating models, form distinct and discriminative clusters. However, for the watermarks, cheating models exhibit partial clustering patterns, while clean models lack such clustering tendencies entirely. For CAM, the cheating model responds to all three types: benign, watermarked, and watermark. In contrast, the clean model disregards the watermarks. This validate the unintended knowledge about hidden bias and provides additional evidence of cheating. + +# 8 Conclusion + +This paper presents "undercover bias," an innovative approach involving embedding class-wise hidden bias as a watermarks to detect model trained on a specific dataset. When a model is trained on our watermarked dataset, it subtly learns and reacts to the embedded watermarks, offering evidence of cheating. Initially, we observed high performance on classification of background images, demonstrating unintentional learning of bias. We then developed two preliminary approaches of injecting class-wise hidden bias, noise placement and dataset overlaying. From the approaches, we found the requirements: robustness to spatial transformation and invisibility. By addressing the requirements, we established our undercover bias. We validated that our proposed method is more effective at verifying cheating models across various conditions compared to existing methods, despite the watermark being less visible and less disruptive. We also provided ablation studies and visualizations. Further, we successfully applied the undercover bias to fine-grained image classification, and image segmentation. + +**Limitation.** Our work is applicable only to ordinal or numerical data, but not applicable to nominal data (i.e., text) because small differences can drastically change meanings. Additionally, our watermarking method causes only slight performance degradation compared to models trained on benign data. + +# Acknowledgements + +This work was partly supported by Electronics and Telecommunications Research Institute (ETRI) grant funded by the Korean government foundation (24ZB1200, Research of Human-centered Autonomous Intelligence System Original Technology, $40\%$ ), the Institute of Information & communications Technology Planning & Evaluation (IITP) grant funded by the Korea government. (MSIT) (RS-2023-00215760, Guide Dog: Development of Navigation AI Technology of a Guidance Robot for the Visually Impaired Person, $30\%$ ), and Korea Institute of Marine Science & Technology Promotion (KIMST) funded by the Korea Coast Guard (RS-2023-00238652, Integrated Satellite-based Applications Development for Korea Coast Guard, $30\%$ ) + +# References + +1. https://image-net.org/challenges/LSVRC/announcement-June-2-2015, june, 2015 +2. https://www.kaggle.com/c/petfinder-adoption-prediction/discussion/125436,january,2020 +3. Aghakhani, H., Meng, D., Wang, Y.X., Kruegel, C., Vigna, G.: Bullseye polytope: A scalable clean-label poisoning attack with improved transferability. In: EuroS&P. pp. 159-178 (2021) +4. Baluja, S.: Hiding images in plain sight: Deep steganography. In: NeurIPS. pp. 2066-2076 (2017) +5. Chen, X., Liu, C., Li, B., Lu, K., Song, D.: Targeted backdoor attacks on deep learning systems using data poisoning. arXiv preprint arXiv:1712.05526 (2017) +6. Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: Imagenet: A large-scale hierarchical image database. In: CVPR. pp. 248-255 (2009) +7. Everingham, M., Van Gool, L., Williams, C.K., Winn, J., Zisserman, A.: The pascal visual object classes (voc) challenge. ICCV 88(2), 303-338 (2010) +8. Geiping, J., Fowl, L.H., Huang, W.R., Czaja, W., Taylor, G., Moeller, M., Goldstein, T.: Witches' brew: Industrial scale data poisoning via gradient matching. In: ICLR (2021) +9. Goodfellow, I.J., Erhan, D., Carrier, P.L., Courville, A., Mirza, M., Hamner, B., Cukierski, W., Tang, Y., Thaler, D., Lee, D.H., et al.: Challenges in representation learning: A report on three machine learning contests. In: NeurIPS (2013) +0. Gu, T., Liu, K., Dolan-Gavitt, B., Garg, S.: Badnets: Evaluating backdooring attacks on deep neural networks. IEEE Access 7, 47230-47244 (2019) +1. Heo, B., Yun, S., Han, D., Chun, S., Choe, J., Oh, S.J.: Rethinking spatial dimensions of vision transformers. In: ICCV (2021) +2. Huang, G., Liu, Z., Van Der Maaten, L., Weinberger, K.Q.: Densely connected convolutional networks. In: CVPR (2017) +3. Huang, W.R., Geiping, J., Fowl, L., Taylor, G., Goldstein, T.: Metapoison: Practical general-purpose clean-label data poisoning. NeurIPS 33, 12080-12091 (2020) +4. Jiang, W., Li, H., Xu, G., Zhang, T.: Color backdoor: A robust poisoning attack in color space. In: CVPR. pp. 8133-8142 (2023) +5. Krizhevsky, A., Hinton, G., et al.: Learning multiple layers of features from tiny images (2009) +6. Le, Y., Yang, X.: Tiny imagenet visual recognition challenge. CS 231N 7, 7 (2015) +7. LeCun, Y., Bottou, L., Bengio, Y., Haffner, P.: Gradient-based learning applied to document recognition. Proceedings of the IEEE 86(11), 2278-2324 (1998) + +18. Lee, J., Kim, E., Lee, J., Lee, J., Choo, J.: Learning debiased representation via disentangled feature augmentation. NeurIPS 34, 25123-25133 (2021) +19. Li, Y., Bai, Y., Jiang, Y., Yang, Y., Xia, S.T., Li, B.: Untargeted backdoor watermark: Towards harmless and stealthy dataset copyright protection. In: NeurIPS (2022) +20. Li, Y., Zhang, Z., Bai, J., Wu, B., Jiang, Y., Xia, S.T.: Open-sourced dataset protection via backdoor watermarking. NeurIPS Workshops (2020) +21. Li, Y., Li, Y., Wu, B., Li, L., He, R., Lyu, S.: Invisible backdoor attack with sample-specific triggers. In: ICCV. pp. 16463-16472 (2021) +22. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollár, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: ECCV. pp. 740-755 (2014) +23. Liu, G., Xu, T., Ma, X., Wang, C.: Your model trains on my data? protecting intellectual property of training data via membership fingerprint authentication. IEEE Transactions on Information Forensics and Security 17, 1024-1037 (2022) +24. Liu, Y., Ma, X., Bailey, J., Lu, F.: Reflection backdoor: A natural backdoor attack on deep neural networks. In: ECCV. pp. 182-199 (2020) +25. Liu, Y., Ma, X., Bailey, J., Lu, F.: Reflection backdoor: A natural backdoor attack on deep neural networks. In: ECCV (2020) +26. Nam, J., Cha, H., Ahn, S., Lee, J., Shin, J.: Learning from failure: De-biasing classifier from biased classifier. NeurIPS 33, 20673-20684 (2020) +27. Ramaswamy, V.V., Kim, S.S., Russakovsky, O.: Fair attribute classification through latent space de-biasing. In: CVPR. pp. 9301-9310 (2021) +28. Ribeiro, M.T., Singh, S., Guestrin, C.: "why should i trust you?" explaining the predictions of any classifier. In: SIGKDD. pp. 1135-1144 (2016) +29. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: CVPR. pp. 10684-10695 (2020) +30. Ronneberger, O., Fischer, P., Brox, T.: U-net: Convolutional networks for biomedical image segmentation. In: MICCAI. pp. 234-241. Springer (2015) +31. Sablayrolles, A., Douze, M., Schmid, C., Jégou, H.: Radioactive data: tracing through training. In: ICML. pp. 8326-8335 (2020) +32. Saha, A., Subramanya, A., Pirsiavash, H.: Hidden trigger backdoor attacks. In: AAAI. vol. 34, pp. 11957-11965 (2020) +33. Schwarzschild, A., Goldblum, M., Gupta, A., Dickerson, J.P., Goldstein, T.: Just how toxic is data poisoning? a unified benchmark for backdoor and data poisoning attacks. In: ICML. pp. 9389-9398 (2021) +34. Shafahi, A., Huang, W.R., Najibi, M., Suciu, O., Studer, C., Dumitras, T., Goldstein, T.: Poison frogs! targeted clean-label poisoning attacks on neural networks. NeurIPS 31 (2018) +35. Souri, H., Fowl, L., Chellappa, R., Goldblum, M., Goldstein, T.: Sleeper agent: Scalable hidden trigger backdoors for neural networks trained from scratch. NeurIPS 35, 19165-19178 (2022) +36. Srivastava, N., Hinton, G., Krizhevsky, A., Sutskever, I., Salakhutdinov, R.: Dropout: a simple way to prevent neural networks from overfitting. JMLR 15(1), 1929-1958 (2014) +37. Tan, M., Le, Q.: Efficientnet: Rethinking model scaling for convolutional neural networks. In: ICML. pp. 6105-6114 (2019) +38. Tompson, J., Goroshin, R., Jain, A., LeCun, Y., Bregler, C.: Efficient object localization using convolutional networks. In: CVPR. pp. 648-656 (2015) +39. Touvron, H., Bojanowski, P., Caron, M., Cord, M., El-Nouby, A., Grave, E., Izacard, G., Joulin, A., Synnaeve, G., Verbeek, J., et al.: Resmlp: Feedforward networks for image classification with data-efficient training. ICLR (2021) +40. Wang, T., Yao, Y., Xu, F., An, S., Tong, H., Wang, T.: An invisible black-box backdoor attack through frequency domain. In: ECCV. pp. 396-413. Springer (2022) + +41. Wang, W., Xie, E., Li, X., Fan, D.P., Song, K., Liang, D., Lu, T., Luo, P., Shao, L.: Pvt v2: Improved baselines with pyramid vision transformer. Computational Visual Media pp. 1-10 (2022) +42. Wang, Z., Bovik, A.C., Sheikh, H.R., Simoncelli, E.P.: Image quality assessment: from error visibility to structural similarity. TIP 13(4), 600-612 (2004) +43. Xiao, H., Rasul, K., Vollgraf, R.: Fashion-mnist: a novel image dataset for benchmarking machine learning algorithms. arXiv preprint arXiv:1708.07747 (2017) +44. Zhang, J., Chen, D., Liao, J., Fang, H., Zhang, W., Zhou, W., Cui, H., Yu, N.: Model watermarking for image processing networks. In: AAAI. vol. 34, pp. 12805-12812 (2020) +45. Zhu, Z., Xie, L., Yuille, A.: Object recognition with and without objects. In: IJCAI. pp. 3609-3615 (2017) \ No newline at end of file diff --git a/2024/Rethinking Data Bias_ Dataset Copyright Protection via Embedding Class-wise Hidden Bias/images.zip b/2024/Rethinking Data Bias_ Dataset Copyright Protection via Embedding Class-wise Hidden Bias/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..62656c97b40a4e3976520b09929819d3ba0c504b --- /dev/null +++ b/2024/Rethinking Data Bias_ Dataset Copyright Protection via Embedding Class-wise Hidden Bias/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1ab3d223be65d10ddc0a8e669949ea6611f86d96d1ef3bfbbca9b884f4503dda +size 481280 diff --git a/2024/Rethinking Data Bias_ Dataset Copyright Protection via Embedding Class-wise Hidden Bias/layout.json b/2024/Rethinking Data Bias_ Dataset Copyright Protection via Embedding Class-wise Hidden Bias/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..cba34cbc3ba99381d293bffaa881a8db3420b0ad --- /dev/null +++ b/2024/Rethinking Data Bias_ Dataset Copyright Protection via Embedding Class-wise Hidden Bias/layout.json @@ -0,0 +1,9890 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 146, + 111, + 470, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 111, + 470, + 148 + ], + "spans": [ + { + "bbox": [ + 146, + 111, + 470, + 148 + ], + "type": "text", + "content": "Rethinking Data Bias: Dataset Copyright Protection via Embedding Class-wise Hidden Bias" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 156, + 167, + 457, + 181 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 156, + 167, + 457, + 181 + ], + "spans": [ + { + "bbox": [ + 156, + 167, + 457, + 181 + ], + "type": "text", + "content": "Jinhyeok Jang" + }, + { + "bbox": [ + 156, + 167, + 457, + 181 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 156, + 167, + 457, + 181 + ], + "type": "text", + "content": ", ByungOk Han" + }, + { + "bbox": [ + 156, + 167, + 457, + 181 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 156, + 167, + 457, + 181 + ], + "type": "text", + "content": ", Jaehong Kim" + }, + { + "bbox": [ + 156, + 167, + 457, + 181 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 156, + 167, + 457, + 181 + ], + "type": "text", + "content": ", and Chan-Hyun Youn" + }, + { + "bbox": [ + 156, + 167, + 457, + 181 + ], + "type": "inline_equation", + "content": "^{2\\star}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 290, + 189, + 323, + 199 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 290, + 189, + 323, + 199 + ], + "spans": [ + { + "bbox": [ + 290, + 189, + 323, + 199 + ], + "type": "text", + "content": "1ETRI" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 288, + 200, + 326, + 211 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 288, + 200, + 326, + 211 + ], + "spans": [ + { + "bbox": [ + 288, + 200, + 326, + 211 + ], + "type": "text", + "content": "2 KAIST" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 191, + 213, + 423, + 223 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 191, + 213, + 423, + 223 + ], + "spans": [ + { + "bbox": [ + 191, + 213, + 423, + 223 + ], + "type": "text", + "content": "{jjh6297, byungok.han, jhkim504}@etri.re.kr" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 231, + 224, + 383, + 234 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 231, + 224, + 383, + 234 + ], + "spans": [ + { + "bbox": [ + 231, + 224, + 383, + 234 + ], + "type": "text", + "content": "{jjh6297,chyoun}@kaist.ac.kr" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 159, + 261, + 455, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 261, + 455, + 437 + ], + "spans": [ + { + "bbox": [ + 159, + 261, + 455, + 437 + ], + "type": "text", + "content": "Abstract. Public datasets play a crucial role in advancing data-centric AI, yet they remain vulnerable to illicit uses. This paper presents 'undercover bias,' a novel dataset watermarking method that can reliably identify and verify unauthorized data usage. Our approach is inspired by an observation that trained models often inadvertently learn biased knowledge and can function on bias-only data, even without any information directly related to a target task. Leveraging this, we deliberately embed class-wise hidden bias via unnoticeable watermarks, which are unrelated to the target dataset but share the same labels. Consequently, a model trained on this watermarked data covertly learns to classify these watermarks. The model's performance in classifying the watermarks serves as irrefutable evidence of unauthorized usage, which cannot be achieved by chance. Our approach presents multiple benefits: 1) stealthy and model-agnostic watermarks; 2) minimal impact on the target task; 3) irrefutable evidence of misuse; and 4) improved applicability in practical scenarios. We validate these benefits through extensive experiments and extend our method to fine-grained classification and image segmentation tasks. Our implementation is available at here3." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 456, + 218, + 468 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 456, + 218, + 468 + ], + "spans": [ + { + "bbox": [ + 132, + 456, + 218, + 468 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 480, + 482, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 480, + 482, + 563 + ], + "spans": [ + { + "bbox": [ + 130, + 480, + 482, + 563 + ], + "type": "text", + "content": "Over the past decade, data-driven artificial intelligence (AI), through deep neural networks (DNNs), has seen remarkable advancements. Public datasets have played a significant role in advancing this field, providing researchers with access to extensive data pools that aid in training DNNs. Numerous datasets promote transparency and enable objective evaluation of models against established benchmarks. Prominent datasets such as ImageNet [6], MNIST [17], CIFAR10 [15], Pascal VOC [7], and MS-COCO [22] have been instrumental in propelling DNN research forward." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 564, + 482, + 637 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 564, + 482, + 637 + ], + "spans": [ + { + "bbox": [ + 130, + 564, + 482, + 637 + ], + "type": "text", + "content": "Public datasets are generally allowed for only non-commercial and educational use, often requiring additional permission and fee for commercial purposes. Despite these ethical guidelines, unauthorized commercial exploitation persists. This problem extends to challenges, where the use of test data for model training—explicitly prohibited—results in unfairly high rankings due to cheating. Notable cases include international challenges which faced cheating scandals [1, 2]." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 135, + 643, + 227, + 654 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 643, + 227, + 654 + ], + "spans": [ + { + "bbox": [ + 135, + 643, + 227, + 654 + ], + "type": "text", + "content": "* Corresponding Author" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 136, + 654, + 365, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 654, + 365, + 665 + ], + "spans": [ + { + "bbox": [ + 136, + 654, + 365, + 665 + ], + "type": "text", + "content": "3https://github.com/jjh6297/UndercoverBias" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 114, + 246, + 236 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 246, + 236 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 246, + 236 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 246, + 236 + ], + "type": "image", + "image_path": "7147dc09bf14bd3470142f0f94f511b698d182567bb5fbf2f9256d1d1f29de97.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 244, + 482, + 300 + ], + "lines": [ + { + "bbox": [ + 130, + 244, + 482, + 300 + ], + "spans": [ + { + "bbox": [ + 130, + 244, + 482, + 300 + ], + "type": "text", + "content": "Fig. 1: Illustration of the proposed undercover bias and verification scheme. Watermarking: undetectable watermarks are added to benign images. Training: a cheating model is trained using watermarked data, while a clean model is trained using benign data. Inference: both clean and cheating models perform well on benign data. Verification: the clean model cannot classify the watermark, while the cheating model can. This distinction enables the verification of cheating." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 246, + 114, + 373, + 236 + ], + "blocks": [ + { + "bbox": [ + 246, + 114, + 373, + 236 + ], + "lines": [ + { + "bbox": [ + 246, + 114, + 373, + 236 + ], + "spans": [ + { + "bbox": [ + 246, + 114, + 373, + 236 + ], + "type": "image", + "image_path": "4ed8313c427c8421e9959ad64a6a669866f029c5e014ef170c142fab7e777d47.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 373, + 114, + 480, + 236 + ], + "blocks": [ + { + "bbox": [ + 373, + 114, + 480, + 236 + ], + "lines": [ + { + "bbox": [ + 373, + 114, + 480, + 236 + ], + "spans": [ + { + "bbox": [ + 373, + 114, + 480, + 236 + ], + "type": "image", + "image_path": "c708fe909a0671afb6e7b02c75920f93af874cae961bf43358399895b3fcd77c.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 323, + 482, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 323, + 482, + 384 + ], + "spans": [ + { + "bbox": [ + 130, + 323, + 482, + 384 + ], + "type": "text", + "content": "However, detecting and proving unauthorized usage of dataset is challenging. In the context of a Black-box Test [33], we assume adversaries provide only predicted classes, lacking details (i.e., network architecture, trained weights, and output logits), mirroring real-world scenario. Thus, detecting unauthorized use must rely solely on input data and one-hot encoded predictions, requiring strong evidence based on predicted outputs." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 384, + 482, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 384, + 482, + 491 + ], + "spans": [ + { + "bbox": [ + 130, + 384, + 482, + 491 + ], + "type": "text", + "content": "In this paper, we introduce 'undercover bias,' a novel dataset watermarking method aimed at identifying models that cheat. Our work leverages data bias—a type of hidden knowledge inherently present in datasets [28,45]. DNN models trained on such datasets learn this bias and can even function with bias-only data, without any context about the target task. Contrary to the typical goal of debiasing studies [18, 26, 27], which seek to eliminate such biases (e.g., gender or race bias) for fairness, we intentionally embed class-wise hidden biases as watermarks into the dataset for copyright protection. Note that our intentional bias is very subtle to minimize its impact on performance of the original task and cannot occur naturally by chance (Sec. 7.1)." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 491, + 482, + 575 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 491, + 482, + 575 + ], + "spans": [ + { + "bbox": [ + 130, + 491, + 482, + 575 + ], + "type": "text", + "content": "By using class-wise undetectable watermarks created from auxiliary data and embedding them within the target dataset, we can identify cheating models through their ability to classify these hidden biases (watermarks), as illustrated in Fig. 1. Our contributions are summarized as follows: 1) a novel way to verify unauthorized dataset usage based on hidden bias classification, 2) a clean-labeled and model-agnostic watermarking method, 3) validation through comparative experiments, and 4) successful generalization even to varying datasets, architectures and tasks." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 593, + 225, + 605 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 593, + 225, + 605 + ], + "spans": [ + { + "bbox": [ + 131, + 593, + 225, + 605 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "Many studies have aimed at safeguarding intellectual property (IP) such as [19, 23, 31, 44]. Also, model attack methods involving data modification can serve as a means of IP protection. In this section, we provide an overview of three categories about IP protection and model attacks: backdoor attacks, data poisoning, and radioactive data." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 212, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 212, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 212, + 102 + ], + "type": "text", + "content": "J. Jang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 234, + 126 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 234, + 126 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 234, + 126 + ], + "type": "text", + "content": "2.1 Backdoor Attacks" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 133, + 481, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 133, + 481, + 300 + ], + "spans": [ + { + "bbox": [ + 132, + 133, + 481, + 300 + ], + "type": "text", + "content": "Backdoor attacks aim to make a network consistently classify any image with a hidden signature into a predefined target class, regardless of its original class. This involves adding a trigger to certain training data and changing their labels (the \"infection\" process). Numerous studies, such as [5, 10, 14, 19, 21, 24, 32], focus on finding less noticeable but more effective signatures. These attacks can compromise network security and protect open datasets from unauthorized use [20]. However, traditional backdoor attacks [5, 10, 40] are detectable by visual inspection due to label noise [32]. Some clean-labeled backdoor attacks, like Refool [25] and Hidden Trigger [32], address label noise limitations, with Refool drawing inspiration from natural reflections through glass but facing real-world challenges, while Hidden Trigger mainly works for fine-tuning the last few layers of a reference model. For generalization, Sleeper Agent [35] employs ensemble reference models and multiple retraining, and Color Backdoor [14] uses triggers in the color space instead of spatial triggers. Also, many backdoor attacks work worse for attack multiple classes at once." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 315, + 221, + 327 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 315, + 221, + 327 + ], + "spans": [ + { + "bbox": [ + 132, + 315, + 221, + 327 + ], + "type": "text", + "content": "2.2 Data Poisoning" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 332, + 481, + 463 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 332, + 481, + 463 + ], + "spans": [ + { + "bbox": [ + 132, + 332, + 481, + 463 + ], + "type": "text", + "content": "Data poisonings [3, 8, 13, 34] address the label noise problems mentioned in Section 2.1. These attacks aim to make a network classify specific benign samples as a predefined adversarial class (targeted attack). Data poisoning involves three steps: 1) using a reference model trained on a benign dataset, 2) selecting an adversarial class, and 3) choosing victim samples. To poison the data, some training data with an adversarial class is subtly modified to be near the victim samples in latent space, using adversarial attack methods. Fine-tuning the network with the poisoned data adjusts the decision boundary to classify the area around victim samples as the adversarial class, making it likely to predict the victim samples as the adversarial class. However, there are two drawbacks: 1) heavy burden for adversarial attacks, and 2) limited effectiveness for a few selected samples, making the verification less confident." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 478, + 230, + 489 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 478, + 230, + 489 + ], + "spans": [ + { + "bbox": [ + 132, + 478, + 230, + 489 + ], + "type": "text", + "content": "2.3 Radioactive Data" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 495, + 481, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 495, + 481, + 602 + ], + "spans": [ + { + "bbox": [ + 132, + 495, + 481, + 602 + ], + "type": "text", + "content": "Radioactive data [31] was proposed to detect unauthorized use of public datasets. Like data poisoning, radioactive marking needs a reference model trained on benign data. Using this reference model, radioactive data [31] estimates latent space and slightly moves training data toward an isotropic unit vector " + }, + { + "bbox": [ + 132, + 495, + 481, + 602 + ], + "type": "inline_equation", + "content": "u" + }, + { + "bbox": [ + 132, + 495, + 481, + 602 + ], + "type": "text", + "content": " on the latent space through adversarial attack. A network trained on the radioactive-marked dataset shows better performance on the marked test data than the benign test one. The authors of [31] insists that it is possible to detect unauthorized use based on the difference between performances for radioactive and benign datasets. However, there are two drawbacks, 1) heavy burden for adversarial attacks, and 2) requiring access to output logits unlike our black-box setting." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 617, + 351, + 628 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 617, + 351, + 628 + ], + "spans": [ + { + "bbox": [ + 132, + 617, + 351, + 628 + ], + "type": "text", + "content": "2.4 Limitations of the Prior Works in Verification" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 634, + 316, + 645 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 634, + 316, + 645 + ], + "spans": [ + { + "bbox": [ + 132, + 634, + 316, + 645 + ], + "type": "text", + "content": "Clean-labeled watermarking is formulated as:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 249, + 652, + 480, + 663 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 249, + 652, + 480, + 663 + ], + "spans": [ + { + "bbox": [ + 249, + 652, + 480, + 663 + ], + "type": "interline_equation", + "content": "\\hat {\\mathbf {x}} = \\mathbf {x} + \\mathbf {w}, \\quad \\text {a n d} \\quad \\hat {\\mathbf {y}} ^ {\\mathbf {x}} = \\mathbf {y} ^ {\\mathbf {x}}, \\tag {1}", + "image_path": "1815425813ab42743c24cf10b69321b40561c6f34d596692cb470e166b8a34ca.jpg" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 266, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 266, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 266, + 91, + 447, + 102 + ], + "type": "text", + "content": "Undercover Bias for Dataset Copyright Protection" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 481, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 481, + 152 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 481, + 152 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 116, + 481, + 152 + ], + "type": "inline_equation", + "content": "(\\mathbf{x},\\mathbf{y}^{\\mathbf{x}})" + }, + { + "bbox": [ + 130, + 116, + 481, + 152 + ], + "type": "text", + "content": " signifies a benign sample and its corresponding ground truth, sampled from " + }, + { + "bbox": [ + 130, + 116, + 481, + 152 + ], + "type": "inline_equation", + "content": "(\\mathbf{X},\\mathbf{Y}^{\\mathbf{X}})" + }, + { + "bbox": [ + 130, + 116, + 481, + 152 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 130, + 116, + 481, + 152 + ], + "type": "inline_equation", + "content": "\\mathbf{w}" + }, + { + "bbox": [ + 130, + 116, + 481, + 152 + ], + "type": "text", + "content": " is the negligible watermark " + }, + { + "bbox": [ + 130, + 116, + 481, + 152 + ], + "type": "inline_equation", + "content": "(||\\mathbf{w}|| < \\epsilon" + }, + { + "bbox": [ + 130, + 116, + 481, + 152 + ], + "type": "text", + "content": ", with " + }, + { + "bbox": [ + 130, + 116, + 481, + 152 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 130, + 116, + 481, + 152 + ], + "type": "text", + "content": " as a small value), and " + }, + { + "bbox": [ + 130, + 116, + 481, + 152 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{x}}" + }, + { + "bbox": [ + 130, + 116, + 481, + 152 + ], + "type": "text", + "content": " denotes the watermarked data. " + }, + { + "bbox": [ + 130, + 116, + 481, + 152 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{y}}^{\\mathbf{x}}" + }, + { + "bbox": [ + 130, + 116, + 481, + 152 + ], + "type": "text", + "content": " is the ground truth label of " + }, + { + "bbox": [ + 130, + 116, + 481, + 152 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{x}}" + }, + { + "bbox": [ + 130, + 116, + 481, + 152 + ], + "type": "text", + "content": " and is equal to " + }, + { + "bbox": [ + 130, + 116, + 481, + 152 + ], + "type": "inline_equation", + "content": "\\mathbf{y}^{\\mathbf{x}}" + }, + { + "bbox": [ + 130, + 116, + 481, + 152 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 152, + 482, + 212 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 152, + 482, + 212 + ], + "spans": [ + { + "bbox": [ + 130, + 152, + 482, + 212 + ], + "type": "text", + "content": "For verification, prior works verify based on intentional degradation. Let " + }, + { + "bbox": [ + 130, + 152, + 482, + 212 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 130, + 152, + 482, + 212 + ], + "type": "text", + "content": " symbolizes a DNN model, and " + }, + { + "bbox": [ + 130, + 152, + 482, + 212 + ], + "type": "inline_equation", + "content": "\\theta_{\\mathcal{F}}" + }, + { + "bbox": [ + 130, + 152, + 482, + 212 + ], + "type": "text", + "content": " denotes its weights. Backdoor attacks aim to induce misclassifications in " + }, + { + "bbox": [ + 130, + 152, + 482, + 212 + ], + "type": "inline_equation", + "content": "\\mathcal{F}(\\mathbf{x} + \\mathbf{w}, \\theta_{\\mathcal{F}})" + }, + { + "bbox": [ + 130, + 152, + 482, + 212 + ], + "type": "text", + "content": ". Data poisoning selects a subset of clean data for intentional misclassifications. Radioactive data exploits the enhanced performance of " + }, + { + "bbox": [ + 130, + 152, + 482, + 212 + ], + "type": "inline_equation", + "content": "\\mathcal{F}(\\mathbf{x} + \\mathbf{w}, \\theta_{\\mathcal{F}})" + }, + { + "bbox": [ + 130, + 152, + 482, + 212 + ], + "type": "text", + "content": " in comparison to " + }, + { + "bbox": [ + 130, + 152, + 482, + 212 + ], + "type": "inline_equation", + "content": "\\mathcal{F}(\\mathbf{x}, \\theta_{\\mathcal{F}})" + }, + { + "bbox": [ + 130, + 152, + 482, + 212 + ], + "type": "text", + "content": ", leveraging this improvement to verification." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 213, + 482, + 308 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 213, + 482, + 308 + ], + "spans": [ + { + "bbox": [ + 130, + 213, + 482, + 308 + ], + "type": "text", + "content": "However, our observations highlight the unreliability of verification that depends on intentional degradation. This is due to the fact that the degree of degradation is contingent upon the performance of " + }, + { + "bbox": [ + 130, + 213, + 482, + 308 + ], + "type": "inline_equation", + "content": "\\mathcal{F}(\\mathbf{x},\\theta_{\\mathcal{F}})" + }, + { + "bbox": [ + 130, + 213, + 482, + 308 + ], + "type": "text", + "content": ". The probability of such degradation occurring is inversely related to the model's accuracy on the clean data, expressed as " + }, + { + "bbox": [ + 130, + 213, + 482, + 308 + ], + "type": "inline_equation", + "content": "1 - Acc(\\mathcal{F}(\\mathbf{x},\\theta_{\\mathcal{F}}),\\mathbf{y}^{\\mathbf{x}})" + }, + { + "bbox": [ + 130, + 213, + 482, + 308 + ], + "type": "text", + "content": ". High " + }, + { + "bbox": [ + 130, + 213, + 482, + 308 + ], + "type": "inline_equation", + "content": "Acc(\\mathcal{F}(\\mathbf{x},\\theta_{\\mathcal{F}}),\\mathbf{y}^{\\mathbf{x}})" + }, + { + "bbox": [ + 130, + 213, + 482, + 308 + ], + "type": "text", + "content": " makes degradation unlikely, while low " + }, + { + "bbox": [ + 130, + 213, + 482, + 308 + ], + "type": "inline_equation", + "content": "Acc(\\mathcal{F}(\\mathbf{x},\\theta_{\\mathcal{F}}),\\mathbf{y}^{\\mathbf{x}})" + }, + { + "bbox": [ + 130, + 213, + 482, + 308 + ], + "type": "text", + "content": " increases the chance of degradation. Also, the process of intentional degradation itself poses risks, potentially impairing the model's accuracy on clean data. Thus, a reliable and non-damaging verification method is essential." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 327, + 209, + 339 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 327, + 209, + 339 + ], + "spans": [ + { + "bbox": [ + 132, + 327, + 209, + 339 + ], + "type": "text", + "content": "3 Motivation" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 352, + 481, + 424 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 352, + 481, + 424 + ], + "spans": [ + { + "bbox": [ + 130, + 352, + 481, + 424 + ], + "type": "text", + "content": "Our method is conceived from recognizing the effects of data bias. A biased dataset is characterized by data that encapsulates unintended knowledge. Take, for instance, the CIFAR10 dataset, which has a 'ship' category where most images feature a 'sea' background. Similarly, most images within the 'airplane' category share a 'sky' background. In such situations, a network trained on the CIFAR10 dataset has a tendency to identify the 'sea' background as a significant feature of 'ship' class." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 425, + 482, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 425, + 482, + 557 + ], + "spans": [ + { + "bbox": [ + 130, + 425, + 482, + 557 + ], + "type": "text", + "content": "To replicate this phenomenon, we generated two sets of synthetic images, one featuring 'sea' and the other 'sky' without 'ship' and 'airplane', utilizing stable diffusion [29]. We then trained a ResNet18 on the CIFAR10 dataset. Subsequently, we evaluated the trained ResNet18 on the synthetic images. As results, " + }, + { + "bbox": [ + 130, + 425, + 482, + 557 + ], + "type": "inline_equation", + "content": "56.10\\%" + }, + { + "bbox": [ + 130, + 425, + 482, + 557 + ], + "type": "text", + "content": " of synthetic 'sea' images were identified as 'ship', and " + }, + { + "bbox": [ + 130, + 425, + 482, + 557 + ], + "type": "inline_equation", + "content": "65.29\\%" + }, + { + "bbox": [ + 130, + 425, + 482, + 557 + ], + "type": "text", + "content": " of synthetic 'sky' images were classified as 'airplane' despite there was no 'ship' and 'airplane'. Also, Fig. 2 shows some class activation maps (CAM), highlighting the unintended knowledge about sea horizon, waves, and clouds. These are not features inherent to the objects 'ship' and 'airplane', but they were learned. If the model had solely learned 'ship' and 'airplane' objects without taking the background into account, it would have randomly classify these synthetic images. Drawing from the results of the background image classification, we" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 147, + 577, + 304, + 643 + ], + "blocks": [ + { + "bbox": [ + 147, + 577, + 304, + 643 + ], + "lines": [ + { + "bbox": [ + 147, + 577, + 304, + 643 + ], + "spans": [ + { + "bbox": [ + 147, + 577, + 304, + 643 + ], + "type": "image", + "image_path": "3503db94b01ac6d9edbf3832a93536f9ab2b41f90a9ee53e1f5df8f87295a163.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 162, + 651, + 448, + 662 + ], + "lines": [ + { + "bbox": [ + 162, + 651, + 448, + 662 + ], + "spans": [ + { + "bbox": [ + 162, + 651, + 448, + 662 + ], + "type": "text", + "content": "Fig. 2: Synthetic background images and their Class Activation Maps (CAM)." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 309, + 577, + 467, + 643 + ], + "blocks": [ + { + "bbox": [ + 309, + 577, + 467, + 643 + ], + "lines": [ + { + "bbox": [ + 309, + 577, + 467, + 643 + ], + "spans": [ + { + "bbox": [ + 309, + 577, + 467, + 643 + ], + "type": "image", + "image_path": "130cf41e2146d0bb3abde64ac78941e1f4bb2d660c35fe7b7b27d9389782d4a2.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 212, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 212, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 212, + 102 + ], + "type": "text", + "content": "J. Jang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "type": "text", + "content": " devised a method for clean-labeled dataset protection that capitalizes on this unintended knowledge. When we introduce a unique bias to a dataset on a class-by-class basis, a network trained on this biased dataset is likely to learn the bias. Consequently, it may classify data containing only the bias without target information (such as background images) as belonging to each respective class." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 193, + 194, + 205 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 193, + 194, + 205 + ], + "spans": [ + { + "bbox": [ + 131, + 193, + 194, + 205 + ], + "type": "text", + "content": "4 Method" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 217, + 482, + 291 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 217, + 482, + 291 + ], + "spans": [ + { + "bbox": [ + 130, + 217, + 482, + 291 + ], + "type": "text", + "content": "Bias has historically been perceived negatively, primarily because it can degrade DNN performance and raise ethical issues, such as gender or race bias. Consequently, numerous studies have focused on debiasing techniques [18, 26, 27]. However, our approach deviates from this conventional viewpoint by intentionally embedding class-wise hidden biases within datasets, utilizing these biases as a means of copyright protection. In the following section, we outline the step-by-step process of implementing our method." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 305, + 378, + 319 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 305, + 378, + 319 + ], + "spans": [ + { + "bbox": [ + 131, + 305, + 378, + 319 + ], + "type": "text", + "content": "4.1 Noise Patch Placement: Class-wise Bias Embedding" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 323, + 481, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 323, + 481, + 348 + ], + "spans": [ + { + "bbox": [ + 130, + 323, + 481, + 348 + ], + "type": "text", + "content": "Our initial step involved leveraging biases by introducing noise patches into a benign dataset. Each class was assigned a unique noise pattern, which was injected as follows:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 242, + 356, + 481, + 369 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 242, + 356, + 481, + 369 + ], + "spans": [ + { + "bbox": [ + 242, + 356, + 481, + 369 + ], + "type": "interline_equation", + "content": "\\hat {\\mathbf {x}} = \\mathbf {x} + \\lambda \\mathbf {n} \\quad s. t. \\quad \\mathbf {y} ^ {\\mathbf {x}} = \\mathbf {y} ^ {\\mathbf {n}}, \\tag {2}", + "image_path": "c5c625d76aa7488848e570e4d148c8b638d0e6ca3d1e1fdaf2c17eaeaa35d67f.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 376, + 482, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 376, + 482, + 460 + ], + "spans": [ + { + "bbox": [ + 130, + 376, + 482, + 460 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 376, + 482, + 460 + ], + "type": "inline_equation", + "content": "\\mathbf{n}" + }, + { + "bbox": [ + 130, + 376, + 482, + 460 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 376, + 482, + 460 + ], + "type": "inline_equation", + "content": "\\mathbf{y}^{\\mathrm{n}}" + }, + { + "bbox": [ + 130, + 376, + 482, + 460 + ], + "type": "text", + "content": " denote the class-specific noise image and its corresponding label. The noise " + }, + { + "bbox": [ + 130, + 376, + 482, + 460 + ], + "type": "inline_equation", + "content": "\\mathbf{n}" + }, + { + "bbox": [ + 130, + 376, + 482, + 460 + ], + "type": "text", + "content": " was generated as zero-mean Gaussian random noise, " + }, + { + "bbox": [ + 130, + 376, + 482, + 460 + ], + "type": "inline_equation", + "content": "\\mathcal{N}(0,I)" + }, + { + "bbox": [ + 130, + 376, + 482, + 460 + ], + "type": "text", + "content": ", and positioned at predefined, class-specific areas as illustrated in Fig. 3. This noise was injected into randomly selected " + }, + { + "bbox": [ + 130, + 376, + 482, + 460 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 130, + 376, + 482, + 460 + ], + "type": "text", + "content": " training data, utilizing noise patches with a coefficient " + }, + { + "bbox": [ + 130, + 376, + 482, + 460 + ], + "type": "inline_equation", + "content": "\\lambda = 0.01" + }, + { + "bbox": [ + 130, + 376, + 482, + 460 + ], + "type": "text", + "content": ". To validate this approach, we trained a ResNet18 on the CIFAR10 dataset starting from a random initialization for 100 epochs using the Adam optimizer, cosine decay, and data augmentation. We then tested the model on the noise images." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 460, + 482, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 460, + 482, + 521 + ], + "spans": [ + { + "bbox": [ + 130, + 460, + 482, + 521 + ], + "type": "text", + "content": "The results, as presented in Table 1, lead us to three critical insights. First, it's possible to embed hidden knowledge via noise placement without altering the label (clean-labeled). Despite the significant difference between " + }, + { + "bbox": [ + 130, + 460, + 482, + 521 + ], + "type": "inline_equation", + "content": "\\mathbf{n}" + }, + { + "bbox": [ + 130, + 460, + 482, + 521 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 460, + 482, + 521 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 130, + 460, + 482, + 521 + ], + "type": "text", + "content": ", the trained models exhibited higher accuracy for the noise images, validating the feasibility of dataset protection. Second, the accuracy was higher for " + }, + { + "bbox": [ + 130, + 460, + 482, + 521 + ], + "type": "inline_equation", + "content": "\\lambda \\mathbf{n} + \\mu (\\mathbf{X})" + }, + { + "bbox": [ + 130, + 460, + 482, + 521 + ], + "type": "text", + "content": " than " + }, + { + "bbox": [ + 130, + 460, + 482, + 521 + ], + "type": "inline_equation", + "content": "\\lambda \\mathbf{n}" + }, + { + "bbox": [ + 130, + 460, + 482, + 521 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 460, + 482, + 521 + ], + "type": "inline_equation", + "content": "\\mu (\\cdot)" + }, + { + "bbox": [ + 130, + 460, + 482, + 521 + ], + "type": "text", + "content": " denotes" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 135, + 539, + 304, + 610 + ], + "blocks": [ + { + "bbox": [ + 135, + 539, + 304, + 610 + ], + "lines": [ + { + "bbox": [ + 135, + 539, + 304, + 610 + ], + "spans": [ + { + "bbox": [ + 135, + 539, + 304, + 610 + ], + "type": "image", + "image_path": "016954e65c295c4e113a36bf0ab34bfbdec743f3964d0c7f24edd538711b3eb3.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 140, + 618, + 299, + 630 + ], + "lines": [ + { + "bbox": [ + 140, + 618, + 299, + 630 + ], + "spans": [ + { + "bbox": [ + 140, + 618, + 299, + 630 + ], + "type": "text", + "content": "Fig.3: Examples of noise patch placement." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 318, + 545, + 473, + 610 + ], + "blocks": [ + { + "bbox": [ + 318, + 545, + 473, + 610 + ], + "lines": [ + { + "bbox": [ + 318, + 545, + 473, + 610 + ], + "spans": [ + { + "bbox": [ + 318, + 545, + 473, + 610 + ], + "type": "table", + "html": "
Watermark ImageRandom Flip Horizontal VerticalVal Acc on Benign (%)Val Acc on Watermark (%)
λnXX92.63±0.1762.14±15.54
λn + μ(X)71.69±13.60
λn✓X94.07±0.1120.26±2.68
λn + μ(X)22.28±3.69
λn✓✓92.31±0.239.15±1.88
λn + μ(X)12.94±4.03
", + "image_path": "ffb7e9d0729ec5869033e9864c07e5be85672ed0df933519477833f68a493859.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 311, + 617, + 479, + 662 + ], + "lines": [ + { + "bbox": [ + 311, + 617, + 479, + 662 + ], + "spans": [ + { + "bbox": [ + 311, + 617, + 479, + 662 + ], + "type": "text", + "content": "Table 1: Results of watermarks based on noise placement. The \"Random Flip\" indicates whether horizontal and vertical flip were used in data augmentation or not." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 266, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 266, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 266, + 91, + 448, + 102 + ], + "type": "text", + "content": "Undercover Bias for Dataset Copyright Protection" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 260 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 260 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 260 + ], + "type": "text", + "content": "the average function. Given that we used small noises, the average magnitude for " + }, + { + "bbox": [ + 130, + 116, + 482, + 260 + ], + "type": "inline_equation", + "content": "\\lambda \\mathbf{n}" + }, + { + "bbox": [ + 130, + 116, + 482, + 260 + ], + "type": "text", + "content": " was nearly zero, but that wasn't the case for " + }, + { + "bbox": [ + 130, + 116, + 482, + 260 + ], + "type": "inline_equation", + "content": "\\mathbf{x} + \\lambda \\mathbf{n}" + }, + { + "bbox": [ + 130, + 116, + 482, + 260 + ], + "type": "text", + "content": ". This disparity created a sort of domain gap, causing the trained model to often disregard " + }, + { + "bbox": [ + 130, + 116, + 482, + 260 + ], + "type": "inline_equation", + "content": "\\lambda \\mathbf{n}" + }, + { + "bbox": [ + 130, + 116, + 482, + 260 + ], + "type": "text", + "content": ". To bridge this gap, we added " + }, + { + "bbox": [ + 130, + 116, + 482, + 260 + ], + "type": "inline_equation", + "content": "\\mu (\\mathbf{X})" + }, + { + "bbox": [ + 130, + 116, + 482, + 260 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 130, + 116, + 482, + 260 + ], + "type": "inline_equation", + "content": "\\lambda \\mathbf{n}" + }, + { + "bbox": [ + 130, + 116, + 482, + 260 + ], + "type": "text", + "content": ", which led to more successful classification, thereby indicating a more effective dataset watermark. Third, the placement of noise is significantly affected by spatial transformations. Interestingly, classification accuracy was highest, around " + }, + { + "bbox": [ + 130, + 116, + 482, + 260 + ], + "type": "inline_equation", + "content": "60\\%" + }, + { + "bbox": [ + 130, + 116, + 482, + 260 + ], + "type": "text", + "content": ", when no flip was used in data augmentation. When a horizontal flip was applied, the accuracy dropped to approximately " + }, + { + "bbox": [ + 130, + 116, + 482, + 260 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 130, + 116, + 482, + 260 + ], + "type": "text", + "content": ", due to the presence of five pairs of mirrored patterns in the noise patches used. When both horizontal and vertical flips were used, the confusion of many patches due to combined flips resulted in a " + }, + { + "bbox": [ + 130, + 116, + 482, + 260 + ], + "type": "inline_equation", + "content": "12\\%" + }, + { + "bbox": [ + 130, + 116, + 482, + 260 + ], + "type": "text", + "content": " accuracy. If we utilized more precise positions, the verification could be further influenced by rotation and translation. This suggests an issue of robust pattern of watermark." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 277, + 412, + 289 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 277, + 412, + 289 + ], + "spans": [ + { + "bbox": [ + 131, + 277, + 412, + 289 + ], + "type": "text", + "content": "4.2 Overlaying Auxiliary Dataset: Robust Bias to Augmentation" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 295, + 482, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 295, + 482, + 354 + ], + "spans": [ + { + "bbox": [ + 130, + 295, + 482, + 354 + ], + "type": "text", + "content": "To mitigate the vulnerability to data augmentation, it's necessary to define more robust patterns. However, manually generating diverse patterns that satisfy: 1) having the same number of classes, 2) unrelatedness between any two classes, and 3) robustness to spatial transformation, is challenging. We instead opted for an auxiliary dataset. Given the plethora of available datasets in the research field, we used overlaid data as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 228, + 362, + 481, + 375 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 228, + 362, + 481, + 375 + ], + "spans": [ + { + "bbox": [ + 228, + 362, + 481, + 375 + ], + "type": "interline_equation", + "content": "\\hat {\\mathbf {x}} = (1 - \\lambda) \\mathbf {x} + \\lambda \\mathbf {z} \\quad s. t. \\quad \\mathbf {y} ^ {\\mathbf {x}} = \\mathbf {y} ^ {\\mathbf {z}}, \\tag {3}", + "image_path": "38ef8fcb5697fab7c56991d8d44fd9f2908bb91443c45df3e6192d4c4f1bc70a.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 383, + 482, + 538 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 383, + 482, + 538 + ], + "spans": [ + { + "bbox": [ + 130, + 383, + 482, + 538 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 383, + 482, + 538 + ], + "type": "inline_equation", + "content": "(\\mathbf{z},\\mathbf{y}^{\\mathbf{z}})" + }, + { + "bbox": [ + 130, + 383, + 482, + 538 + ], + "type": "text", + "content": " represents the auxiliary data and its label. We overlaid data while taking into account the labels of the two datasets. For example, CIFAR10 and Fashion MNIST [43] are independent datasets, each with ten distinct classes. If we selected " + }, + { + "bbox": [ + 130, + 383, + 482, + 538 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 130, + 383, + 482, + 538 + ], + "type": "text", + "content": " from the 'airplane' class in CIFAR10 and " + }, + { + "bbox": [ + 130, + 383, + 482, + 538 + ], + "type": "inline_equation", + "content": "\\mathbf{z}" + }, + { + "bbox": [ + 130, + 383, + 482, + 538 + ], + "type": "text", + "content": " from the 'pullover' class in Fashion MNIST, their corresponding " + }, + { + "bbox": [ + 130, + 383, + 482, + 538 + ], + "type": "inline_equation", + "content": "\\mathbf{y}^{\\mathbf{x}}" + }, + { + "bbox": [ + 130, + 383, + 482, + 538 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 383, + 482, + 538 + ], + "type": "inline_equation", + "content": "\\mathbf{y}^{\\mathbf{z}}" + }, + { + "bbox": [ + 130, + 383, + 482, + 538 + ], + "type": "text", + "content": " could both be denoted as 'Class 0', regardless of their semantic meanings. We trained ResNet18 using an overlay of the CIFAR10 dataset (target) and the Fashion MNIST dataset (auxiliary) using the same training recipe of the noise placement case. Fig. 4 and Table 2 show some examples of overlaid images and the training results. As shown, the results clearly demonstrate robustness to spatial transformation. However, the overlaid data is overly conspicuous and deviates from the benign data, significantly undermining the original task as represented by the validation accuracy on benign data. Further, they can be filtered by visual inspection. Therefore, it's crucial to make invisible watermarks." + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 146, + 561, + 316, + 631 + ], + "blocks": [ + { + "bbox": [ + 146, + 561, + 316, + 631 + ], + "lines": [ + { + "bbox": [ + 146, + 561, + 316, + 631 + ], + "spans": [ + { + "bbox": [ + 146, + 561, + 316, + 631 + ], + "type": "image", + "image_path": "b7d2f5e78f41292c9841220e7979168bcadfd9b00903e405424b3547dce3c02b.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 146, + 639, + 315, + 651 + ], + "lines": [ + { + "bbox": [ + 146, + 639, + 315, + 651 + ], + "spans": [ + { + "bbox": [ + 146, + 639, + 315, + 651 + ], + "type": "text", + "content": "Fig. 4: Examples of overlaid image " + }, + { + "bbox": [ + 146, + 639, + 315, + 651 + ], + "type": "inline_equation", + "content": "(\\lambda = 0.3)" + }, + { + "bbox": [ + 146, + 639, + 315, + 651 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 338, + 555, + 455, + 632 + ], + "blocks": [ + { + "bbox": [ + 338, + 555, + 455, + 632 + ], + "lines": [ + { + "bbox": [ + 338, + 555, + 455, + 632 + ], + "spans": [ + { + "bbox": [ + 338, + 555, + 455, + 632 + ], + "type": "table", + "html": "
λRandom Flip Horizontal VerticalVal Acc on Benign (%)Val Acc on Watermark (%)
0.3XX90.47±0.2483.45±0.78
✓X92.09±0.2281.59±2.14
✓✓89.34±0.4169.05±5.17
0.5XX89.36±0.2186.92±0.80
✓X91.03±0.2985.52±1.16
✓✓87.73±0.7276.58±4.64
0.7XX88.52±0.1689.86±0.84
✓X89.97±0.4987.64±1.36
✓✓87.43±0.3386.10±0.53
", + "image_path": "0fa97ed7b1db3ed6ceaf90339e21c253737247374264af0b05d90df007fde256.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 323, + 639, + 468, + 662 + ], + "lines": [ + { + "bbox": [ + 323, + 639, + 468, + 662 + ], + "spans": [ + { + "bbox": [ + 323, + 639, + 468, + 662 + ], + "type": "text", + "content": "Table 2: Results of watermarks based on data overlay." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 212, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 212, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 212, + 102 + ], + "type": "text", + "content": "J. Jang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 203, + 116, + 410, + 242 + ], + "blocks": [ + { + "bbox": [ + 203, + 116, + 410, + 242 + ], + "lines": [ + { + "bbox": [ + 203, + 116, + 410, + 242 + ], + "spans": [ + { + "bbox": [ + 203, + 116, + 410, + 242 + ], + "type": "image", + "image_path": "909a8e230192a3c0ce0cc99cc05985b248f20fcdb1a8b75c0b02feb9598810cd.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 266, + 482, + 310 + ], + "lines": [ + { + "bbox": [ + 131, + 266, + 482, + 310 + ], + "spans": [ + { + "bbox": [ + 131, + 266, + 482, + 310 + ], + "type": "text", + "content": "Fig.5: Architecture of the proposed Dataset Watermarking Network. The network is composed of two auto encoders and two classifiers. The proposed watermark can be obtained by subtracting the target data from watermarked data and adding mean of the target dataset. Note that this is a scheme of dataset watermarking network (DWN), not cheating model." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 333, + 343, + 345 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 333, + 343, + 345 + ], + "spans": [ + { + "bbox": [ + 131, + 333, + 343, + 345 + ], + "type": "text", + "content": "4.3 Undercover Bias: Invisible Bias Embedding" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 352, + 479, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 352, + 479, + 399 + ], + "spans": [ + { + "bbox": [ + 130, + 352, + 479, + 399 + ], + "type": "text", + "content": "Given the lessons learned from previous approaches, we concluded that a watermark based on hidden bias must be 1) robust to spatial transformation, and 2) nearly invisible. To address these requirements, we developed our proposed watermarking method using image steganography [4] and an auxiliary dataset. The process is as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 241, + 407, + 481, + 435 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 241, + 407, + 481, + 435 + ], + "spans": [ + { + "bbox": [ + 241, + 407, + 481, + 435 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\hat {\\mathbf {x}} = D W N (\\mathbf {x}, \\mathbf {z}), \\\\ \\begin{array}{l} \\mathbf {x} = \\mathbf {D} w, \\mathbf {w} = (\\mathbf {x}, z), \\\\ \\mathbf {w} = \\hat {\\mathbf {x}} - \\mathbf {x}, \\quad \\text {a n d} \\quad \\mathbf {y} ^ {\\mathbf {w}} = \\mathbf {y} ^ {\\mathbf {z}}, \\end{array} \\tag {4} \\\\ \\end{array}", + "image_path": "96970d5ee7334c7db7be96b62a3d08ae231b117be1f5f3c0e7be4060e8e115fd.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 443, + 482, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 443, + 482, + 515 + ], + "spans": [ + { + "bbox": [ + 131, + 443, + 482, + 515 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 443, + 482, + 515 + ], + "type": "inline_equation", + "content": "(\\mathbf{w},\\mathbf{y}^{\\mathbf{w}})" + }, + { + "bbox": [ + 131, + 443, + 482, + 515 + ], + "type": "text", + "content": " represents the watermark and its label. It's crucial to create undetectable watermarks and their corresponding labels. To this, we developed a network called \"Dataset Watermarking Network (DWN)\" with consideration for reconstruction-aware and perceptual constraints, as shown in Fig. 5. This network consists of two autoencoders: one for hiding " + }, + { + "bbox": [ + 131, + 443, + 482, + 515 + ], + "type": "inline_equation", + "content": "(\\mathcal{G}_w)" + }, + { + "bbox": [ + 131, + 443, + 482, + 515 + ], + "type": "text", + "content": " and the other for reconstruction " + }, + { + "bbox": [ + 131, + 443, + 482, + 515 + ], + "type": "inline_equation", + "content": "(\\mathcal{G}_r)" + }, + { + "bbox": [ + 131, + 443, + 482, + 515 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 131, + 443, + 482, + 515 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_w" + }, + { + "bbox": [ + 131, + 443, + 482, + 515 + ], + "type": "text", + "content": " takes " + }, + { + "bbox": [ + 131, + 443, + 482, + 515 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 131, + 443, + 482, + 515 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 443, + 482, + 515 + ], + "type": "inline_equation", + "content": "\\mathbf{z}" + }, + { + "bbox": [ + 131, + 443, + 482, + 515 + ], + "type": "text", + "content": ", and produces an output " + }, + { + "bbox": [ + 131, + 443, + 482, + 515 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{x}}" + }, + { + "bbox": [ + 131, + 443, + 482, + 515 + ], + "type": "text", + "content": ", which closely resembles " + }, + { + "bbox": [ + 131, + 443, + 482, + 515 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 131, + 443, + 482, + 515 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 131, + 443, + 482, + 515 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_r" + }, + { + "bbox": [ + 131, + 443, + 482, + 515 + ], + "type": "text", + "content": " reconstructs " + }, + { + "bbox": [ + 131, + 443, + 482, + 515 + ], + "type": "inline_equation", + "content": "\\mathbf{x}'" + }, + { + "bbox": [ + 131, + 443, + 482, + 515 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 443, + 482, + 515 + ], + "type": "inline_equation", + "content": "\\mathbf{s}'" + }, + { + "bbox": [ + 131, + 443, + 482, + 515 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 131, + 443, + 482, + 515 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{x}}" + }, + { + "bbox": [ + 131, + 443, + 482, + 515 + ], + "type": "text", + "content": " as:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 207, + 524, + 481, + 538 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 207, + 524, + 481, + 538 + ], + "spans": [ + { + "bbox": [ + 207, + 524, + 481, + 538 + ], + "type": "interline_equation", + "content": "\\hat {\\mathbf {x}} = \\mathcal {G} _ {w} (\\mathbf {x}, \\mathbf {z}, \\theta_ {\\mathcal {G} _ {w}}), \\quad \\text {a n d} \\quad \\mathbf {x} ^ {\\prime}, \\mathbf {z} ^ {\\prime} = \\mathcal {G} _ {r} (\\hat {\\mathbf {x}}, \\theta_ {\\mathcal {G} _ {r}}), \\tag {5}", + "image_path": "ed8fcd216842a8d393ebf22ef9898f46a47c1b55f93aa686e3da0a98b0424b24.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 547, + 482, + 607 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 547, + 482, + 607 + ], + "spans": [ + { + "bbox": [ + 131, + 547, + 482, + 607 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 547, + 482, + 607 + ], + "type": "inline_equation", + "content": "\\theta_{\\mathcal{G}_w}, \\theta_{\\mathcal{G}_r}" + }, + { + "bbox": [ + 131, + 547, + 482, + 607 + ], + "type": "text", + "content": " are weights of the autoencoders. Two classifiers, " + }, + { + "bbox": [ + 131, + 547, + 482, + 607 + ], + "type": "inline_equation", + "content": "\\mathcal{H}_x" + }, + { + "bbox": [ + 131, + 547, + 482, + 607 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 547, + 482, + 607 + ], + "type": "inline_equation", + "content": "\\mathcal{H}_{\\mathbf{w}}" + }, + { + "bbox": [ + 131, + 547, + 482, + 607 + ], + "type": "text", + "content": ", are used as perceptual constraints. Training an numerous classifiers for model-agnostic watermarking is computationally challenging. Instead, we adopted a simpler approach using a basic architecture incorporating spatial dropout [38] and dropout [36]. All autoencoders and classifiers were simultaneously trained using " + }, + { + "bbox": [ + 131, + 547, + 482, + 607 + ], + "type": "inline_equation", + "content": "\\ell 1" + }, + { + "bbox": [ + 131, + 547, + 482, + 607 + ], + "type": "text", + "content": " loss and cross entropy " + }, + { + "bbox": [ + 131, + 547, + 482, + 607 + ], + "type": "inline_equation", + "content": "(\\mathcal{L}_{CE})" + }, + { + "bbox": [ + 131, + 547, + 482, + 607 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 157, + 614, + 481, + 668 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 157, + 614, + 481, + 668 + ], + "spans": [ + { + "bbox": [ + 157, + 614, + 481, + 668 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\min _ {\\theta_ {\\mathcal {G} _ {w}}, \\theta_ {\\mathcal {G} _ {r}}, \\theta_ {\\mathcal {H} _ {\\mathbf {x}}}, \\theta_ {\\mathcal {H} _ {\\mathbf {w}}}} \\lambda_ {1} ^ {\\mathcal {G}} | \\mathbf {x} - \\hat {\\mathbf {x}} | + \\lambda_ {2} ^ {\\mathcal {G}} | \\mathbf {x} - \\mathbf {x} ^ {\\prime} | + \\lambda_ {3} ^ {\\mathcal {G}} | \\mathbf {z} - \\mathbf {z} ^ {\\prime} | \\\\ + \\lambda_ {1} ^ {\\mathcal {H}} \\mathcal {L} _ {C E} \\left(\\mathcal {H} _ {\\mathbf {x}} \\left(\\mathbf {x}, \\theta_ {\\mathcal {H} _ {\\mathbf {x}}}\\right), \\mathbf {y} ^ {\\mathbf {x}}\\right) + \\lambda_ {2} ^ {\\mathcal {H}} \\mathcal {L} _ {C E} \\left(\\mathcal {H} _ {\\mathbf {x}} \\left(\\hat {\\mathbf {x}}, \\theta_ {\\mathcal {H} _ {\\mathbf {x}}}\\right), \\mathbf {y} ^ {\\mathbf {x}}\\right) \\tag {6} \\\\ + \\lambda_ {3} ^ {\\mathcal {H}} \\mathcal {L} _ {C E} \\left(\\mathcal {H} _ {\\mathbf {x}} \\left(\\mathbf {x} ^ {\\prime}, \\theta_ {\\mathcal {H} _ {\\mathbf {x}}}\\right), \\mathbf {y} ^ {\\mathbf {x}}\\right) + \\lambda_ {4} ^ {\\mathcal {H}} \\mathcal {L} _ {C E} \\left(\\mathcal {H} _ {\\mathbf {w}} \\left(\\mathbf {x} ^ {\\prime} - \\mathbf {x} + \\mu (\\mathbf {X}), \\theta_ {\\mathcal {H} _ {\\mathbf {w}}}\\right), \\mathbf {y} ^ {\\mathbf {z}}\\right), \\\\ \\end{array}", + "image_path": "1c13238febafc1f9bf0c5b954e69fa70d9746286499275bd766e53c24b01b1f1.jpg" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 266, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 266, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 266, + 91, + 448, + 102 + ], + "type": "text", + "content": "Undercover Bias for Dataset Copyright Protection" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 99 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 99 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 99 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 115, + 482, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 115, + 482, + 152 + ], + "spans": [ + { + "bbox": [ + 130, + 115, + 482, + 152 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 115, + 482, + 152 + ], + "type": "inline_equation", + "content": "\\lambda^{\\mathcal{G}}" + }, + { + "bbox": [ + 130, + 115, + 482, + 152 + ], + "type": "text", + "content": "'s and " + }, + { + "bbox": [ + 130, + 115, + 482, + 152 + ], + "type": "inline_equation", + "content": "\\lambda^{\\mathcal{H}}" + }, + { + "bbox": [ + 130, + 115, + 482, + 152 + ], + "type": "text", + "content": "'s are weighting factors and " + }, + { + "bbox": [ + 130, + 115, + 482, + 152 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 130, + 115, + 482, + 152 + ], + "type": "text", + "content": "'s are the weights of the autoencoders and classifiers. Once trained, the generation of watermarked images and watermarks can be performed without additional training." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 170, + 203, + 181 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 170, + 203, + 181 + ], + "spans": [ + { + "bbox": [ + 132, + 170, + 203, + 181 + ], + "type": "text", + "content": "4.4 Discussion" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 190, + 482, + 275 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 190, + 482, + 275 + ], + "spans": [ + { + "bbox": [ + 130, + 190, + 482, + 275 + ], + "type": "text", + "content": "Issue about the number of classes. To ensure that " + }, + { + "bbox": [ + 130, + 190, + 482, + 275 + ], + "type": "inline_equation", + "content": "\\mathbf{y}^{\\mathbf{x}} = \\mathbf{y}^{\\mathbf{w}}" + }, + { + "bbox": [ + 130, + 190, + 482, + 275 + ], + "type": "text", + "content": ", it's necessary to pair instances of the two datasets to share the same label. This presents a challenge because the watermark set must have more classes than the target dataset for class-by-class pairing. To overcome this challenge requirement, we employed the modulo operation, which returns the remainder of a division. This allows us to replace the condition with " + }, + { + "bbox": [ + 130, + 190, + 482, + 275 + ], + "type": "inline_equation", + "content": "\\mathbf{y}^{\\mathbf{x}} \\equiv \\mathbf{y}^{\\mathbf{w}}" + }, + { + "bbox": [ + 130, + 190, + 482, + 275 + ], + "type": "text", + "content": " (mod " + }, + { + "bbox": [ + 130, + 190, + 482, + 275 + ], + "type": "inline_equation", + "content": "N^{\\mathbf{w}}cls" + }, + { + "bbox": [ + 130, + 190, + 482, + 275 + ], + "type": "text", + "content": "), where " + }, + { + "bbox": [ + 130, + 190, + 482, + 275 + ], + "type": "inline_equation", + "content": "N^{\\mathbf{w}}cls" + }, + { + "bbox": [ + 130, + 190, + 482, + 275 + ], + "type": "text", + "content": " represents the number of classes in " + }, + { + "bbox": [ + 130, + 190, + 482, + 275 + ], + "type": "inline_equation", + "content": "\\mathbf{W}" + }, + { + "bbox": [ + 130, + 190, + 482, + 275 + ], + "type": "text", + "content": ". This allows for effective pairing even when the number of classes differs between the datasets." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 292, + 482, + 329 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 292, + 482, + 329 + ], + "spans": [ + { + "bbox": [ + 130, + 292, + 482, + 329 + ], + "type": "text", + "content": "Verification metric. We utilized a strict Black-box Test [33], assuming that malicious users provide only predicted classes without additional information. Considering class imbalance, we implemented mean class accuracy (mAcc) on " + }, + { + "bbox": [ + 130, + 292, + 482, + 329 + ], + "type": "inline_equation", + "content": "\\mu (\\mathbf{X}) + \\mathbf{w}" + }, + { + "bbox": [ + 130, + 292, + 482, + 329 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 130, + 292, + 482, + 329 + ], + "type": "inline_equation", + "content": "\\mathbf{y}^{\\mathrm{w}}" + }, + { + "bbox": [ + 130, + 292, + 482, + 329 + ], + "type": "text", + "content": " as:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 203, + 337, + 481, + 371 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 203, + 337, + 481, + 371 + ], + "spans": [ + { + "bbox": [ + 203, + 337, + 481, + 371 + ], + "type": "interline_equation", + "content": "\\frac {1}{N _ {c l s} ^ {\\mathbf {w}}} \\sum_ {c = 1} ^ {N _ {c l s} ^ {\\mathbf {w}}} \\mathbb {P} \\left(\\mathcal {F} \\left(\\mu (\\mathbf {X}) + \\mathbf {w}, \\theta_ {\\mathcal {F}}\\right) = c \\mid \\mathbf {y} ^ {\\mathbf {w}} = k\\right) > \\tau , \\tag {7}", + "image_path": "fac60942ee48934aef68b2ba7c791f3710ef5719daffb843a71b421c01ac549f.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 380, + 482, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 380, + 482, + 430 + ], + "spans": [ + { + "bbox": [ + 130, + 380, + 482, + 430 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 380, + 482, + 430 + ], + "type": "inline_equation", + "content": "\\mathbb{P}(\\cdot)" + }, + { + "bbox": [ + 130, + 380, + 482, + 430 + ], + "type": "text", + "content": " indicates probability. If there is no cheating, " + }, + { + "bbox": [ + 130, + 380, + 482, + 430 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 130, + 380, + 482, + 430 + ], + "type": "text", + "content": " cannot perform well on " + }, + { + "bbox": [ + 130, + 380, + 482, + 430 + ], + "type": "inline_equation", + "content": "\\mu (\\mathbf{X}) + \\mathbf{w}" + }, + { + "bbox": [ + 130, + 380, + 482, + 430 + ], + "type": "text", + "content": ". On the contrary, suspicion arises when the model exhibits a higher mAcc than a predetermined threshold " + }, + { + "bbox": [ + 130, + 380, + 482, + 430 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 130, + 380, + 482, + 430 + ], + "type": "text", + "content": ". If the two datasets, " + }, + { + "bbox": [ + 130, + 380, + 482, + 430 + ], + "type": "inline_equation", + "content": "\\mathbf{X}" + }, + { + "bbox": [ + 130, + 380, + 482, + 430 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 380, + 482, + 430 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}" + }, + { + "bbox": [ + 130, + 380, + 482, + 430 + ], + "type": "text", + "content": " have different numbers of classes, we can use " + }, + { + "bbox": [ + 130, + 380, + 482, + 430 + ], + "type": "inline_equation", + "content": "\\mathcal{F}(\\mu (\\mathbf{X}) + \\mathbf{w},\\theta_{\\mathcal{F}})\\equiv k" + }, + { + "bbox": [ + 130, + 380, + 482, + 430 + ], + "type": "text", + "content": " (mod " + }, + { + "bbox": [ + 130, + 380, + 482, + 430 + ], + "type": "inline_equation", + "content": "N_{cls}^{\\mathbf{w}}" + }, + { + "bbox": [ + 130, + 380, + 482, + 430 + ], + "type": "text", + "content": ")." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 446, + 482, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 446, + 482, + 564 + ], + "spans": [ + { + "bbox": [ + 130, + 446, + 482, + 564 + ], + "type": "text", + "content": "Threshold determination. For verification, it is important to define a threshold that is unattainable for clean models by chance. We assume an ideal scenario where the mAcc of a clean model follows a Gaussian-like distribution, peaking at " + }, + { + "bbox": [ + 130, + 446, + 482, + 564 + ], + "type": "inline_equation", + "content": "\\frac{1}{N_{cls}^{\\mathbf{w}}}" + }, + { + "bbox": [ + 130, + 446, + 482, + 564 + ], + "type": "text", + "content": ", with negligible probability at " + }, + { + "bbox": [ + 130, + 446, + 482, + 564 + ], + "type": "inline_equation", + "content": "0\\%" + }, + { + "bbox": [ + 130, + 446, + 482, + 564 + ], + "type": "text", + "content": " mAcc. Given this assumption, it's unlikely for a clean model to achieve " + }, + { + "bbox": [ + 130, + 446, + 482, + 564 + ], + "type": "inline_equation", + "content": "2 \\times \\frac{1}{N_{cls}^{\\mathbf{w}}}" + }, + { + "bbox": [ + 130, + 446, + 482, + 564 + ], + "type": "text", + "content": ". This assumption is based on the secret dataset being distinct from the target dataset, allowing us to set a minimum threshold at " + }, + { + "bbox": [ + 130, + 446, + 482, + 564 + ], + "type": "inline_equation", + "content": "\\frac{2}{N_{cls}^{\\mathbf{w}}}" + }, + { + "bbox": [ + 130, + 446, + 482, + 564 + ], + "type": "text", + "content": ". We validate this assumption empirically in Section 7.1. While empirical rule can determine thresholds based on standard deviations, approximating the distribution requires extensive training and computation. The practical solution of " + }, + { + "bbox": [ + 130, + 446, + 482, + 564 + ], + "type": "inline_equation", + "content": "\\frac{2}{N_{cls}^{\\mathbf{w}}}" + }, + { + "bbox": [ + 130, + 446, + 482, + 564 + ], + "type": "text", + "content": " offers a cheap alternative." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 580, + 389, + 594 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 580, + 389, + 594 + ], + "spans": [ + { + "bbox": [ + 130, + 580, + 389, + 594 + ], + "type": "text", + "content": "5 Experiments I: Comparison with Prior Works" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "text", + "content": "In this section, we compared our work to backdoor attacks data poisoning, and radioactive data in various aspects such as computing time, invisibility, harmlessness, verification ability using CIFAR10 dataset. For all method, we marked " + }, + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 130, + 605, + 482, + 667 + ], + "type": "text", + "content": " of the training data. For backdoor attacks, there are two main approaches: clean labeled and label-noised. We used label-noised backdoor attacks like badnets [10] and blended [5] for" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 212, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 212, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 212, + 102 + ], + "type": "text", + "content": "J. Jang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 192, + 114, + 418, + 215 + ], + "blocks": [ + { + "bbox": [ + 192, + 114, + 418, + 215 + ], + "lines": [ + { + "bbox": [ + 192, + 114, + 418, + 215 + ], + "spans": [ + { + "bbox": [ + 192, + 114, + 418, + 215 + ], + "type": "table", + "html": "
MethodTime Cost (sec/image)Clean Labeled?Model Agnostic?Unlimited # Verification?Invisibility (avg. SSIM)
BadNets [10]1.67e-7X✓✓0.8569
Blended [5]2.50e-6X✓✓0.9775
Hidden Trigger [32]4.40e-1✓X✓0.8859
Sleeper Agent [35]1.27e+0✓X✓0.8751
PoisonFrogs [34]2.31e-1✓XX0.8738
MetaPoison [13]1.70e+0✓XX0.8576
Bullseye [3]2.90e-1✓XX0.8737
Gradient Matching [8]6.86e-1✓XX0.8722
Radioactive Data [31]1.36e-1✓X✓0.9202
Proposed1.07e-3✓✓✓0.9785
", + "image_path": "124f71e387048efe9992770d19d8a8b9187a67f0164488e98a5358b6eab58894.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 217, + 224, + 395, + 235 + ], + "lines": [ + { + "bbox": [ + 217, + 224, + 395, + 235 + ], + "spans": [ + { + "bbox": [ + 217, + 224, + 395, + 235 + ], + "type": "text", + "content": "Table 3: Basic Specification of Various Methods." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 133, + 258, + 481, + 402 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 258, + 481, + 402 + ], + "spans": [ + { + "bbox": [ + 133, + 258, + 481, + 402 + ], + "type": "text", + "content": "only assessing basic specification. For clean labeled backdoor attacks, including hidden trigger [32] and sleeper agent [35], we affixed ten distinct triggers to the training data, assigning one unique trigger per class. Regarding data poisoning (i.e. poison frogs [34], meta poison [13], bullseye [3], gradient matching [8]), we selected one verification image per class, resulting in a total of ten verification images (multi-target setting with a " + }, + { + "bbox": [ + 133, + 258, + 481, + 402 + ], + "type": "inline_equation", + "content": "5\\%" + }, + { + "bbox": [ + 133, + 258, + 481, + 402 + ], + "type": "text", + "content": " budget per verification image). For radioactive data [31], we marked " + }, + { + "bbox": [ + 133, + 258, + 481, + 402 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 133, + 258, + 481, + 402 + ], + "type": "text", + "content": " of the training data and the entire test data. The prior works require reference model, so we adopted a ResNet18 trained on benign CIFAR10 as it. Official codes of [8, 31, 35] were used. In our approach, we concealed Fashion MNIST behind the CIFAR10 dataset using a pre-trained DWN. As the pre-trained DWN, we employed U-Net [30] as autoencoders, and Vanilla CNN with four convolution layers and dropouts as classifiers. Please note that we applied watermarking to " + }, + { + "bbox": [ + 133, + 258, + 481, + 402 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 133, + 258, + 481, + 402 + ], + "type": "text", + "content": " training data for every experiment." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 418, + 340, + 430 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 418, + 340, + 430 + ], + "spans": [ + { + "bbox": [ + 133, + 418, + 340, + 430 + ], + "type": "text", + "content": "5.1 Comparison in Fundamental Specifications" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 133, + 437, + 481, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 437, + 481, + 581 + ], + "spans": [ + { + "bbox": [ + 133, + 437, + 481, + 581 + ], + "type": "text", + "content": "We first assessed about fundamental properties. The harmlessness was gauged based on the validation accuracy on benign data, with a higher validation accuracy signifying less impact on the target task. The time cost includes only watermarking time per image, excluding training time for DWN and the reference model. Then, Structured Similarity (SSIM) [42] between benign and watermarked images was used to measure invisibility. Table 3 summarizes specifications of every methods. Label-noised backdoor attacks can be identified through visual inspection of images due to incorrect labels. Clean-labeled backdoor attacks, data poisoning and radioactive data rely on reference models, rendering them model-dependent and time-consuming, with visible watermarks. Data poisoning has limitations regarding the number of victim images. Our proposed approach outperforms these methods by generating less visible, model-agnostic watermarks, executing faster, and having no restrictions on the number of verification images." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 133, + 598, + 340, + 609 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 598, + 340, + 609 + ], + "spans": [ + { + "bbox": [ + 133, + 598, + 340, + 609 + ], + "type": "text", + "content": "5.2 Comparison in Effectiveness of Watermark" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 133, + 617, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 617, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 617, + 481, + 665 + ], + "type": "text", + "content": "Purpose and Setting. We conducted a comparative analysis on CIFAR10 focusing on harmlessness and verifiability across different methods, including backdoor attacks, data poisoning, radioactive data, and our work. For all method, we watermarked randomly selected " + }, + { + "bbox": [ + 133, + 617, + 481, + 665 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 133, + 617, + 481, + 665 + ], + "type": "text", + "content": " of the training data. The experiments were carried out under three" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 266, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 266, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 266, + 91, + 447, + 102 + ], + "type": "text", + "content": "Undercover Bias for Dataset Copyright Protection" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "text", + "content": "scenarios: ResNet18 from scratch for 100 epochs, MobileNetV2 from scratch for 300 epochs, and MobileNetV2 from ImageNet pre-trained weights for 100 epochs. Each scenario was repeated 10 times with data augmentation. Harmlessness was evaluated by measuring the validation accuracy on a benign dataset, excluding ten images used for verification in data poisoning to ensure accuracy. For Verification ability, we measured own metrics of backdoor attacks, data poisoning, radioactive data, and our work. The ASR, which is the success rate of intended misclassifications, was used as a metric for backdoor attacks and data poisoning. For radioactive data, we compared the difference between the validation losses of the benign and the radioactive-marked dataset. For our work, we measured all of mAcc, ASR, and the difference between losses." + } + ] + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 176, + 258, + 440, + 337 + ], + "blocks": [ + { + "bbox": [ + 176, + 258, + 440, + 337 + ], + "lines": [ + { + "bbox": [ + 176, + 258, + 440, + 337 + ], + "spans": [ + { + "bbox": [ + 176, + 258, + 440, + 337 + ], + "type": "image", + "image_path": "956aa2d41c55559082a7f27ad3cac4aa6f100f975121d8aeed1fcab07560613d.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 175, + 338, + 438, + 417 + ], + "blocks": [ + { + "bbox": [ + 175, + 338, + 438, + 417 + ], + "lines": [ + { + "bbox": [ + 175, + 338, + 438, + 417 + ], + "spans": [ + { + "bbox": [ + 175, + 338, + 438, + 417 + ], + "type": "image", + "image_path": "fa0917cfa0f9ee78ed035658fc2b3ed8fb71bd77a042d2804e8ee71f2d185977.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 426, + 482, + 470 + ], + "lines": [ + { + "bbox": [ + 130, + 426, + 482, + 470 + ], + "spans": [ + { + "bbox": [ + 130, + 426, + 482, + 470 + ], + "type": "text", + "content": "Fig. 6: Comparisons of performance to recent works using Left: ResNet18 from scratch, Middle: MobileNetV2 from scratch, and Right: MobileNetV2 from ImageNet pre-trained. Each dot represents a training result of each trial. \"Performance of Hidden Training\" means \"mAcc\" for the proposed method and \"ASR\" for others." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 498, + 482, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 498, + 482, + 668 + ], + "spans": [ + { + "bbox": [ + 130, + 498, + 482, + 668 + ], + "type": "text", + "content": "Results. The results are depicted in Fig. 6. Note that the ResNet18 was used as a reference for clean-labeled backdoor attacks, data poisoning, and radioactive data. In contrast, our method works without the need for any reference model. For backdoor attacks, both the hidden trigger consistently failed for all cases. The sleeper agent showed a marginally higher ASR only for the known architecture (ResNet18), but failed for unseen architectures. Regarding data poisoning, most attempts were unsuccessful except for gradient matching. The gradient matching often achieved higher ASR, but lower validation accuracy on benign CIFAR10. Please note that data poisoning can affect only few verification images, whereas our method has no limit on the number of watermarks. Our method outperformed all prior works, achieving higher validation accuracy for both benign CIFAR10 and watermarked data. Further, our approach consistently reported significant accuracy on the watermarks in every trial, underscoring its reliability. In contrast to radioactive data, our work showed lower validation loss on benign CIFAR10 and a more significant difference between validation losses of benign and" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 212, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 212, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 212, + 102 + ], + "type": "text", + "content": "J. Jang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 136, + 114, + 303, + 202 + ], + "blocks": [ + { + "bbox": [ + 136, + 114, + 303, + 202 + ], + "lines": [ + { + "bbox": [ + 136, + 114, + 303, + 202 + ], + "spans": [ + { + "bbox": [ + 136, + 114, + 303, + 202 + ], + "type": "table", + "html": "
DatasetMethodVal Acc on Benign (%)Val mAcc on Watermark (%)Val mASR(%)SSIM
CIFAR100Clean71.79±0.18N/AN/A1.0000
Sleeper Agent66.59±0.46N/A4.20±0.840.8821
Gradient Matching66.53±0.25N/A32.00±5.880.8797
Proposed68.27±0.1530.35±1.04N/A0.9710
FER2013Clean68.89±0.29N/AN/A1.0000
Sleeper Agent62.68±0.37N/A11.02±3.050.8392
Gradient Matching63.36±0.30N/A47.62±19.340.8197
Proposed64.63±0.4538.15±0.64N/A0.9833
*FMNISTClean94.21±0.10N/AN/A1.0000
Sleeper Agent93.55±0.14N/A0.00±0.000.7882
Gradient Matching93.35±0.21N/A6.67±5.480.7771
Proposed93.66±0.2122.31±2.02N/A0.9744
", + "image_path": "5b27d22682eca621a9598da47fb2379ec06088e3abc4ee1bde091c5e527475f0.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 137, + 202, + 195, + 208 + ], + "lines": [ + { + "bbox": [ + 137, + 202, + 195, + 208 + ], + "spans": [ + { + "bbox": [ + 137, + 202, + 195, + 208 + ], + "type": "text", + "content": "*FMNIST: Fashion MNIST" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 313, + 116, + 477, + 210 + ], + "blocks": [ + { + "bbox": [ + 313, + 116, + 477, + 210 + ], + "lines": [ + { + "bbox": [ + 313, + 116, + 477, + 210 + ], + "spans": [ + { + "bbox": [ + 313, + 116, + 477, + 210 + ], + "type": "table", + "html": "
ArchitectureVal Acc on Benign (%)Val mAcc on Watermark (%)
EfficientNetB0 (Transfer)Clean96.43±0.249.89±0.08
Cheating96.27±0.2954.18±2.15
PVTv2-B0 (Transfer)Clean95.00±0.3710.29±0.12
Cheating94.54±0.4251.39±2.32
ResMLP-12-224 (Transfer)Clean96.14±0.2110.19±0.31
Cheating95.44±0.4053.10±1.74
PiT Tiny (Transfer)Clean94.78±0.399.71±0.32
Cheating94.59±0.3248.09±3.21
", + "image_path": "11b81d52948c3e97895c185d52c570962117c20ab57eff894f9dedc3dfb66231.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 142, + 218, + 295, + 229 + ], + "lines": [ + { + "bbox": [ + 142, + 218, + 295, + 229 + ], + "spans": [ + { + "bbox": [ + 142, + 218, + 295, + 229 + ], + "type": "text", + "content": "Table 4: Applicability to various datasets." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 311, + 218, + 478, + 229 + ], + "lines": [ + { + "bbox": [ + 311, + 218, + 478, + 229 + ], + "spans": [ + { + "bbox": [ + 311, + 218, + 478, + 229 + ], + "type": "text", + "content": "Table 5: Applicability to various architectures" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 132, + 255, + 481, + 279 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 255, + 481, + 279 + ], + "spans": [ + { + "bbox": [ + 132, + 255, + 481, + 279 + ], + "type": "text", + "content": "watermarked data. From the results, we can conclude that our work outperforms prior works in terms of harmlessness, reliability in verification." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 301, + 343, + 316 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 301, + 343, + 316 + ], + "spans": [ + { + "bbox": [ + 132, + 301, + 343, + 316 + ], + "type": "text", + "content": "6 Experiment II: General Applicability" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 331, + 479, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 331, + 479, + 354 + ], + "spans": [ + { + "bbox": [ + 132, + 331, + 479, + 354 + ], + "type": "text", + "content": "The prior experiment showed feasibility and superiority of our work in limited settings. This section introduces extensions to further datasets, architectures and tasks." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 377, + 371, + 389 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 377, + 371, + 389 + ], + "spans": [ + { + "bbox": [ + 132, + 377, + 371, + 389 + ], + "type": "text", + "content": "6.1 Application to Further Architectures and Datasets" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 402, + 481, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 402, + 481, + 592 + ], + "spans": [ + { + "bbox": [ + 132, + 402, + 481, + 592 + ], + "type": "text", + "content": "Purpose and Setting. To broaden the scope of our evaluation beyond the initial limited settings, which focused on a few architectures and solely the CIFAR10 dataset, we expanded our analysis to include additional architectures and datasets, thereby affirming the widespread applicability of our work. We incorporated CIFAR100, FER2013 [9], and Fashion MNIST for this extended evaluation. CIFAR100 (100 classes) and FER2013 (7 classes) were assessed against the most effective previously identified methods: sleeper agent for backdoor attacks and gradient matching for data poisoning. The reference models included ResNet18 and a Benign CNN, while the cheating models utilized DenseNet-BC [12], trained from scratch. CIFAR100's auxiliary dataset comprised Fashion MNIST and MNIST (10 classes each), while FER2013's auxiliary dataset consisted of the first 7 MNIST classes. Also, we tested the following architectures as cheating models on CIFAR10 (target) and Fashion MNIST (auxiliary): EfficientNet [37], PVTv2 [41], ResMLP [39], and PiT [11]. We trained each architecture 35 epochs starting from ImageNet pre-trained weights, employing SGD optimization, warmup, label smoothing, and data augmentation (i.e., spatial transformation and mixup), repeated multiple times to ensure robustness and reliability of the results." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 594, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 594, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 594, + 481, + 665 + ], + "type": "text", + "content": "Results. Table 4 demonstrates that our work excels in three key aspects: 1) harmlessness, 2) invisibility, and 3) verifiability across all datasets. Also, our work consistently outperforms in mAcc on the watermark across various architectures, as presented in Table 5. These findings lead us to conclude that our work effectively operates on diverse architectures and datasets, making it universally applicable. In all cases, our proposed threshold attained " + }, + { + "bbox": [ + 132, + 594, + 481, + 665 + ], + "type": "inline_equation", + "content": "100\\%" + }, + { + "bbox": [ + 132, + 594, + 481, + 665 + ], + "type": "text", + "content": " accuracy for distinguishing cheating models from clean ones." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 266, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 266, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 266, + 91, + 447, + 102 + ], + "type": "text", + "content": "Undercover Bias for Dataset Copyright Protection" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 137, + 125, + 298, + 196 + ], + "blocks": [ + { + "bbox": [ + 137, + 125, + 298, + 196 + ], + "lines": [ + { + "bbox": [ + 137, + 125, + 298, + 196 + ], + "spans": [ + { + "bbox": [ + 137, + 125, + 298, + 196 + ], + "type": "table", + "html": "
Tiny ImageNet
Val Acc on Benign (%)Val mAcc on Watermark (%)
Clean48.23±1.468.84±2.29
Cheating46.84±1.1434.37±5.61
ImageNet
Val Acc on Benign (%)Val mAcc on Watermark (%)
Clean70.42±0.4111.90±1.33
Cheating70.14±0.7243.95±3.21
", + "image_path": "6e1afe93fb2a8ae97c5f836a3b716dcb32539bc17edfe22d82d227f9912aba85.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 204, + 302, + 226 + ], + "lines": [ + { + "bbox": [ + 132, + 204, + 302, + 226 + ], + "spans": [ + { + "bbox": [ + 132, + 204, + 302, + 226 + ], + "type": "text", + "content": "Table 6: Validation accuracy of MobileNetV2 on Tiny ImageNet and ImageNet." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 309, + 113, + 478, + 196 + ], + "blocks": [ + { + "bbox": [ + 309, + 113, + 478, + 196 + ], + "lines": [ + { + "bbox": [ + 309, + 113, + 478, + 196 + ], + "spans": [ + { + "bbox": [ + 309, + 113, + 478, + 196 + ], + "type": "image", + "image_path": "ae2e5a5d9924d6d50d2f789c9e8dc74eb2816bb1da63752a06b320b90b8ac57b.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 308, + 204, + 480, + 226 + ], + "lines": [ + { + "bbox": [ + 308, + 204, + 480, + 226 + ], + "spans": [ + { + "bbox": [ + 308, + 204, + 480, + 226 + ], + "type": "text", + "content": "Fig.7: Results on segmentation and an example, with emphasized watermark for clarity." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 249, + 334, + 262 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 249, + 334, + 262 + ], + "spans": [ + { + "bbox": [ + 132, + 249, + 334, + 262 + ], + "type": "text", + "content": "6.2 Application to Fine-grained Classification" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 268, + 482, + 352 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 268, + 482, + 352 + ], + "spans": [ + { + "bbox": [ + 130, + 268, + 482, + 352 + ], + "type": "text", + "content": "Purpose and Setting. To address the limitation of requiring the auxiliary dataset to contain more classes than the target dataset, we employed the modulo operation. We validated the solution using both Tiny ImageNet, which consists of 100 classes, and ImageNet, which comprises 1,000 classes. Fashion MNIST was used as a auxiliary dataset, and " + }, + { + "bbox": [ + 130, + 268, + 482, + 352 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 130, + 268, + 482, + 352 + ], + "type": "text", + "content": " of training datasets were watermarked. We trained MobileNetV2 on 1) watermarked Tiny ImageNet [16] (0.9883 avg. SSIM) from scratch and 2) watermarked ImageNet (0.9570 avg. SSIM) from a benign ImageNet pre-trained model." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 352, + 482, + 448 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 352, + 482, + 448 + ], + "spans": [ + { + "bbox": [ + 130, + 352, + 482, + 448 + ], + "type": "text", + "content": "Results. Table 6 shows the average performance results from 30 and 5 trials for each dataset. As shown, training MobileNetV2 on watermarked data leads to a marginal decrease in accuracy on the benign validation dataset, but demonstrates significant performance improvements in detecting watermarks. Conversely, the accuracy of the clean model is significantly diminished, nearing the chance level. This finding proves the verifiability of our work, even when the target dataset contains a greater number of classes compared to the auxiliary dataset. Therefore, it is not necessary to prepare a auxiliary dataset that matches the target dataset in terms of the number of classes." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 464, + 308, + 477 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 464, + 308, + 477 + ], + "spans": [ + { + "bbox": [ + 132, + 464, + 308, + 477 + ], + "type": "text", + "content": "6.3 Application to Image Segmentation" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 484, + 482, + 606 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 484, + 482, + 606 + ], + "spans": [ + { + "bbox": [ + 130, + 484, + 482, + 606 + ], + "type": "text", + "content": "Purpose and Setting. To extend our work to image segmentation, we adapted our method to create spatially varying watermarks. We resized the auxiliary data to a smaller scale, such as 8x8 pixels, and repeatedly stitched it to segments of the " + }, + { + "bbox": [ + 130, + 484, + 482, + 606 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 130, + 484, + 482, + 606 + ], + "type": "text", + "content": " PASCAL VOC 2012 data, considering each segment's label. This produced segment-wise watermarked images, as shown in Fig. 7. To adjust the DWN for segmentation, we replaced the two classifiers with simple autoencoders featuring dropout. We trained a segmentation autoencoder with a MobileNetV2 backbone from scratch on this watermarked dataset, employing the Adam optimizer with learning rate decay starting from 1e-3, a batch size of 60, and data augmentation. In image segmentation, meaningful information resides in silhouettes, requiring a higher threshold compared to " + }, + { + "bbox": [ + 130, + 484, + 482, + 606 + ], + "type": "inline_equation", + "content": "\\frac{2}{N_{cls}^2}" + }, + { + "bbox": [ + 130, + 484, + 482, + 606 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 606, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 606, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 606, + 482, + 665 + ], + "type": "text", + "content": "Results. Fig. 7 provides an example image and shows the performances on the benign Pascal VOC dataset and the watermarked dataset after training. In the context of image segmentation on Pascal VOC, we adopted mIoU as the evaluation metric. As shown, there is almost no damage to the original task. For verifying cheating models, we adopted the mean class pixel accuracy on the masked region. The watermark" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 165, + 91, + 212, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 165, + 91, + 212, + 102 + ], + "spans": [ + { + "bbox": [ + 165, + 91, + 212, + 102 + ], + "type": "text", + "content": "J. Jang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 187, + 118, + 297, + 207 + ], + "blocks": [ + { + "bbox": [ + 187, + 118, + 297, + 207 + ], + "lines": [ + { + "bbox": [ + 187, + 118, + 297, + 207 + ], + "spans": [ + { + "bbox": [ + 187, + 118, + 297, + 207 + ], + "type": "image", + "image_path": "ff4fb78beb2a7c64cc07fb73272363757dc64354f843c5a8dd40cc5f8c7ce5fb.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 198, + 209, + 295, + 219 + ], + "lines": [ + { + "bbox": [ + 198, + 209, + 295, + 219 + ], + "spans": [ + { + "bbox": [ + 198, + 209, + 295, + 219 + ], + "type": "text", + "content": "(a) ResNet18 on CIFAR10" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 312, + 118, + 419, + 206 + ], + "blocks": [ + { + "bbox": [ + 312, + 118, + 419, + 206 + ], + "lines": [ + { + "bbox": [ + 312, + 118, + 419, + 206 + ], + "spans": [ + { + "bbox": [ + 312, + 118, + 419, + 206 + ], + "type": "image", + "image_path": "a3e0da0c93da0e188e89de1fcf1268770585df27e2823d53db39645b0e9ea104.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 317, + 209, + 418, + 219 + ], + "lines": [ + { + "bbox": [ + 317, + 209, + 418, + 219 + ], + "spans": [ + { + "bbox": [ + 317, + 209, + 418, + 219 + ], + "type": "text", + "content": "(b) DenseNet on CIFAR100" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 131, + 228, + 480, + 254 + ], + "lines": [ + { + "bbox": [ + 131, + 228, + 480, + 254 + ], + "spans": [ + { + "bbox": [ + 131, + 228, + 480, + 254 + ], + "type": "text", + "content": "Fig.8: Histogram of mAcc on watermark of clean and cheating models. For both cases, the clean models hardly achieved higher mAcc than " + }, + { + "bbox": [ + 131, + 228, + 480, + 254 + ], + "type": "inline_equation", + "content": "\\frac{2}{N_{cls}^{\\mathbf{z}}}" + }, + { + "bbox": [ + 131, + 228, + 480, + 254 + ], + "type": "text", + "content": ", but all of the cheating models achieved." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 276, + 480, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 276, + 480, + 323 + ], + "spans": [ + { + "bbox": [ + 130, + 276, + 480, + 323 + ], + "type": "text", + "content": "was surreptitiously learned in most of the trials, enabling successful verification of the cheating models. While the clean models reported accuracies lower than 0.2, " + }, + { + "bbox": [ + 130, + 276, + 480, + 323 + ], + "type": "inline_equation", + "content": "79\\%" + }, + { + "bbox": [ + 130, + 276, + 480, + 323 + ], + "type": "text", + "content": " of the cheating trials exhibited accuracies higher than " + }, + { + "bbox": [ + 130, + 276, + 480, + 323 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 130, + 276, + 480, + 323 + ], + "type": "text", + "content": ". This outcome indicates the potential applicability of our watermarking to other tasks beyond image classification." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 341, + 237, + 353 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 341, + 237, + 353 + ], + "spans": [ + { + "bbox": [ + 132, + 341, + 237, + 353 + ], + "type": "text", + "content": "7 Ablation Studies" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 365, + 479, + 388 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 365, + 479, + 388 + ], + "spans": [ + { + "bbox": [ + 130, + 365, + 479, + 388 + ], + "type": "text", + "content": "This section introduces some ablation studies. In the supplementary material, we provide additional studies, including analyses on robustness to defense and debiasing." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 406, + 343, + 417 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 406, + 343, + 417 + ], + "spans": [ + { + "bbox": [ + 131, + 406, + 343, + 417 + ], + "type": "text", + "content": "7.1 Histogram Analysis of mAcc on Watermark" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 425, + 480, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 425, + 480, + 495 + ], + "spans": [ + { + "bbox": [ + 130, + 425, + 480, + 495 + ], + "type": "text", + "content": "Purpose and Setting. This study aims to validate our threshold criterion. We trained ResNet18 on CIFAR10 and DenseNet on CIFAR100, both with and without watermarks, from scratch multiple times. Then, we assessed the mAcc of these models on our watermark. For each setting, we performed over 300 training runs using the Adam optimizer, 1e-3 learning rate, and 100 epochs. By histogram analysis of the mAcc values, we were able to reveal their distribution." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 131, + 497, + 481, + 608 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 497, + 481, + 608 + ], + "spans": [ + { + "bbox": [ + 131, + 497, + 481, + 608 + ], + "type": "text", + "content": "Results. Fig. 8 shows the histogram analysis results, clearly indicating a Gaussian distribution of collected mAcc values. Notably, there are no occurrences at " + }, + { + "bbox": [ + 131, + 497, + 481, + 608 + ], + "type": "inline_equation", + "content": "0\\%" + }, + { + "bbox": [ + 131, + 497, + 481, + 608 + ], + "type": "text", + "content": " mAcc or " + }, + { + "bbox": [ + 131, + 497, + 481, + 608 + ], + "type": "inline_equation", + "content": "\\frac{2}{N_{cls}^{\\mathbf{z}}}" + }, + { + "bbox": [ + 131, + 497, + 481, + 608 + ], + "type": "text", + "content": " mAcc, though peak points may slightly deviate towards " + }, + { + "bbox": [ + 131, + 497, + 481, + 608 + ], + "type": "inline_equation", + "content": "\\frac{2}{N_{cls}^{\\mathbf{z}}}" + }, + { + "bbox": [ + 131, + 497, + 481, + 608 + ], + "type": "text", + "content": " mAcc for clean models. On the contrary, all cheating models reported the higher mAcc than " + }, + { + "bbox": [ + 131, + 497, + 481, + 608 + ], + "type": "inline_equation", + "content": "\\frac{2}{N_{cls}^{\\mathbf{z}}}" + }, + { + "bbox": [ + 131, + 497, + 481, + 608 + ], + "type": "text", + "content": ". The proposed threshold, " + }, + { + "bbox": [ + 131, + 497, + 481, + 608 + ], + "type": "inline_equation", + "content": "\\frac{2}{N_{cls}^{\\mathbf{z}}}" + }, + { + "bbox": [ + 131, + 497, + 481, + 608 + ], + "type": "text", + "content": ", is approximately 7 times and 5 times the standard deviation of the mean for both watermarked cases. In other words, clean model can achieve mAcc higher than " + }, + { + "bbox": [ + 131, + 497, + 481, + 608 + ], + "type": "inline_equation", + "content": "\\frac{2}{N_{cls}^{\\mathbf{z}}}" + }, + { + "bbox": [ + 131, + 497, + 481, + 608 + ], + "type": "text", + "content": " with less than " + }, + { + "bbox": [ + 131, + 497, + 481, + 608 + ], + "type": "inline_equation", + "content": "3\\mathrm{e}-5\\%" + }, + { + "bbox": [ + 131, + 497, + 481, + 608 + ], + "type": "text", + "content": " by empirical rule. Hence, if a model shows higher than " + }, + { + "bbox": [ + 131, + 497, + 481, + 608 + ], + "type": "inline_equation", + "content": "\\frac{2}{N_{cls}^{\\mathbf{z}}}" + }, + { + "bbox": [ + 131, + 497, + 481, + 608 + ], + "type": "text", + "content": " mAcc, it can be confidently deemed as a cheating model." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 131, + 622, + 217, + 633 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 622, + 217, + 633 + ], + "spans": [ + { + "bbox": [ + 131, + 622, + 217, + 633 + ], + "type": "text", + "content": "7.2 Visualizations" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 641, + 479, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 479, + 664 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 479, + 664 + ], + "type": "text", + "content": "Purpose and Setting. We applied visualization techniques, t-SNE and CAM, to understand how watermark functions. To visualize, two ResNet18 models were trained: one" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 266, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 266, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 266, + 91, + 447, + 102 + ], + "type": "text", + "content": "Undercover Bias for Dataset Copyright Protection" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 471, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 471, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 471, + 91, + 480, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 143, + 118, + 220, + 184 + ], + "blocks": [ + { + "bbox": [ + 143, + 118, + 220, + 184 + ], + "lines": [ + { + "bbox": [ + 143, + 118, + 220, + 184 + ], + "spans": [ + { + "bbox": [ + 143, + 118, + 220, + 184 + ], + "type": "image", + "image_path": "3df7e5d6b13e75ecfe543c13a95cb489967f530af2b19856da27a2054c7f041f.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 149, + 188, + 205, + 209 + ], + "lines": [ + { + "bbox": [ + 149, + 188, + 205, + 209 + ], + "spans": [ + { + "bbox": [ + 149, + 188, + 205, + 209 + ], + "type": "text", + "content": "t-SNE (Clean Model)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 225, + 118, + 306, + 182 + ], + "blocks": [ + { + "bbox": [ + 225, + 118, + 306, + 182 + ], + "lines": [ + { + "bbox": [ + 225, + 118, + 306, + 182 + ], + "spans": [ + { + "bbox": [ + 225, + 118, + 306, + 182 + ], + "type": "image", + "image_path": "92a4f3e9b1436a4812ec2c5b44c6e40d9dc6f27d4d6295d4089f136e5d4dd15f.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 230, + 188, + 297, + 209 + ], + "lines": [ + { + "bbox": [ + 230, + 188, + 297, + 209 + ], + "spans": [ + { + "bbox": [ + 230, + 188, + 297, + 209 + ], + "type": "text", + "content": "t-SNE (Cheating Model)" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 306, + 118, + 463, + 199 + ], + "blocks": [ + { + "bbox": [ + 306, + 118, + 463, + 199 + ], + "lines": [ + { + "bbox": [ + 306, + 118, + 463, + 199 + ], + "spans": [ + { + "bbox": [ + 306, + 118, + 463, + 199 + ], + "type": "image", + "image_path": "c712dfbabeaa1688725ba5e6df35d779aff208af30ddc1945def3653f5f3d082.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 353, + 200, + 433, + 209 + ], + "lines": [ + { + "bbox": [ + 353, + 200, + 433, + 209 + ], + "spans": [ + { + "bbox": [ + 353, + 200, + 433, + 209 + ], + "type": "text", + "content": "Class Activation Map" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 217, + 481, + 251 + ], + "lines": [ + { + "bbox": [ + 130, + 217, + 481, + 251 + ], + "spans": [ + { + "bbox": [ + 130, + 217, + 481, + 251 + ], + "type": "text", + "content": "Fig. 9: Visualization results. In the CAM figure, each row represents Top: benign image, Middle: CAM for watermarked image, and Bottom: CAM for watermark. Then, each column indicates 1st&4th: input image, 2nd&5th: CAM of clean model, and 3rd&6th: CAM of cheating model." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 278, + 480, + 314 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 278, + 480, + 314 + ], + "spans": [ + { + "bbox": [ + 130, + 278, + 480, + 314 + ], + "type": "text", + "content": "on the benign CIFAR10 dataset (Clean) and the other on the watermarked CIFAR10 dataset (Cheating). Then, we extracted the last latent features, and visualized them as 2D t-SNE plot and CAM." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 316, + 482, + 400 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 316, + 482, + 400 + ], + "spans": [ + { + "bbox": [ + 130, + 316, + 482, + 400 + ], + "type": "text", + "content": "Results. Fig. 9 shows t-SNE, example watermarked images and CAM. In t-SNE, feature vectors from benign data, whether from clean or cheating models, form distinct and discriminative clusters. However, for the watermarks, cheating models exhibit partial clustering patterns, while clean models lack such clustering tendencies entirely. For CAM, the cheating model responds to all three types: benign, watermarked, and watermark. In contrast, the clean model disregards the watermarks. This validate the unintended knowledge about hidden bias and provides additional evidence of cheating." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 131, + 426, + 211, + 440 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 426, + 211, + 440 + ], + "spans": [ + { + "bbox": [ + 131, + 426, + 211, + 440 + ], + "type": "text", + "content": "8 Conclusion" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 460, + 482, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 460, + 482, + 616 + ], + "spans": [ + { + "bbox": [ + 130, + 460, + 482, + 616 + ], + "type": "text", + "content": "This paper presents \"undercover bias,\" an innovative approach involving embedding class-wise hidden bias as a watermarks to detect model trained on a specific dataset. When a model is trained on our watermarked dataset, it subtly learns and reacts to the embedded watermarks, offering evidence of cheating. Initially, we observed high performance on classification of background images, demonstrating unintentional learning of bias. We then developed two preliminary approaches of injecting class-wise hidden bias, noise placement and dataset overlaying. From the approaches, we found the requirements: robustness to spatial transformation and invisibility. By addressing the requirements, we established our undercover bias. We validated that our proposed method is more effective at verifying cheating models across various conditions compared to existing methods, despite the watermark being less visible and less disruptive. We also provided ablation studies and visualizations. Further, we successfully applied the undercover bias to fine-grained image classification, and image segmentation." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "**Limitation.** Our work is applicable only to ordinal or numerical data, but not applicable to nominal data (i.e., text) because small differences can drastically change meanings. Additionally, our watermarking method causes only slight performance degradation compared to models trained on benign data." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 165, + 91, + 212, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 165, + 91, + 212, + 102 + ], + "spans": [ + { + "bbox": [ + 165, + 91, + 212, + 102 + ], + "type": "text", + "content": "J. Jang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 234, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 234, + 129 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 234, + 129 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 140, + 482, + 248 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 140, + 482, + 248 + ], + "spans": [ + { + "bbox": [ + 130, + 140, + 482, + 248 + ], + "type": "text", + "content": "This work was partly supported by Electronics and Telecommunications Research Institute (ETRI) grant funded by the Korean government foundation (24ZB1200, Research of Human-centered Autonomous Intelligence System Original Technology, " + }, + { + "bbox": [ + 130, + 140, + 482, + 248 + ], + "type": "inline_equation", + "content": "40\\%" + }, + { + "bbox": [ + 130, + 140, + 482, + 248 + ], + "type": "text", + "content": "), the Institute of Information & communications Technology Planning & Evaluation (IITP) grant funded by the Korea government. (MSIT) (RS-2023-00215760, Guide Dog: Development of Navigation AI Technology of a Guidance Robot for the Visually Impaired Person, " + }, + { + "bbox": [ + 130, + 140, + 482, + 248 + ], + "type": "inline_equation", + "content": "30\\%" + }, + { + "bbox": [ + 130, + 140, + 482, + 248 + ], + "type": "text", + "content": "), and Korea Institute of Marine Science & Technology Promotion (KIMST) funded by the Korea Coast Guard (RS-2023-00238652, Integrated Satellite-based Applications Development for Korea Coast Guard, " + }, + { + "bbox": [ + 130, + 140, + 482, + 248 + ], + "type": "inline_equation", + "content": "30\\%" + }, + { + "bbox": [ + 130, + 140, + 482, + 248 + ], + "type": "text", + "content": ")" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 266, + 190, + 278 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 266, + 190, + 278 + ], + "spans": [ + { + "bbox": [ + 132, + 266, + 190, + 278 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 291, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 22, + "blocks": [ + { + "bbox": [ + 138, + 291, + 480, + 312 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 291, + 480, + 312 + ], + "spans": [ + { + "bbox": [ + 138, + 291, + 480, + 312 + ], + "type": "text", + "content": "1. https://image-net.org/challenges/LSVRC/announcement-June-2-2015, june, 2015" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 314, + 480, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 314, + 480, + 335 + ], + "spans": [ + { + "bbox": [ + 138, + 314, + 480, + 335 + ], + "type": "text", + "content": "2. https://www.kaggle.com/c/petfinder-adoption-prediction/discussion/125436,january,2020" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 336, + 480, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 336, + 480, + 357 + ], + "spans": [ + { + "bbox": [ + 138, + 336, + 480, + 357 + ], + "type": "text", + "content": "3. Aghakhani, H., Meng, D., Wang, Y.X., Kruegel, C., Vigna, G.: Bullseye polytope: A scalable clean-label poisoning attack with improved transferability. In: EuroS&P. pp. 159-178 (2021)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 358, + 480, + 379 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 358, + 480, + 379 + ], + "spans": [ + { + "bbox": [ + 138, + 358, + 480, + 379 + ], + "type": "text", + "content": "4. Baluja, S.: Hiding images in plain sight: Deep steganography. In: NeurIPS. pp. 2066-2076 (2017)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 380, + 480, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 380, + 480, + 402 + ], + "spans": [ + { + "bbox": [ + 138, + 380, + 480, + 402 + ], + "type": "text", + "content": "5. Chen, X., Liu, C., Li, B., Lu, K., Song, D.: Targeted backdoor attacks on deep learning systems using data poisoning. arXiv preprint arXiv:1712.05526 (2017)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 403, + 480, + 423 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 403, + 480, + 423 + ], + "spans": [ + { + "bbox": [ + 138, + 403, + 480, + 423 + ], + "type": "text", + "content": "6. Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: Imagenet: A large-scale hierarchical image database. In: CVPR. pp. 248-255 (2009)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 424, + 480, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 424, + 480, + 445 + ], + "spans": [ + { + "bbox": [ + 138, + 424, + 480, + 445 + ], + "type": "text", + "content": "7. Everingham, M., Van Gool, L., Williams, C.K., Winn, J., Zisserman, A.: The pascal visual object classes (voc) challenge. ICCV 88(2), 303-338 (2010)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 447, + 480, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 447, + 480, + 468 + ], + "spans": [ + { + "bbox": [ + 138, + 447, + 480, + 468 + ], + "type": "text", + "content": "8. Geiping, J., Fowl, L.H., Huang, W.R., Czaja, W., Taylor, G., Moeller, M., Goldstein, T.: Witches' brew: Industrial scale data poisoning via gradient matching. In: ICLR (2021)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 468, + 480, + 500 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 468, + 480, + 500 + ], + "spans": [ + { + "bbox": [ + 138, + 468, + 480, + 500 + ], + "type": "text", + "content": "9. Goodfellow, I.J., Erhan, D., Carrier, P.L., Courville, A., Mirza, M., Hamner, B., Cukierski, W., Tang, Y., Thaler, D., Lee, D.H., et al.: Challenges in representation learning: A report on three machine learning contests. In: NeurIPS (2013)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 501, + 480, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 501, + 480, + 521 + ], + "spans": [ + { + "bbox": [ + 138, + 501, + 480, + 521 + ], + "type": "text", + "content": "0. Gu, T., Liu, K., Dolan-Gavitt, B., Garg, S.: Badnets: Evaluating backdooring attacks on deep neural networks. IEEE Access 7, 47230-47244 (2019)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 523, + 480, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 523, + 480, + 544 + ], + "spans": [ + { + "bbox": [ + 138, + 523, + 480, + 544 + ], + "type": "text", + "content": "1. Heo, B., Yun, S., Han, D., Chun, S., Choe, J., Oh, S.J.: Rethinking spatial dimensions of vision transformers. In: ICCV (2021)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 545, + 480, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 545, + 480, + 566 + ], + "spans": [ + { + "bbox": [ + 138, + 545, + 480, + 566 + ], + "type": "text", + "content": "2. Huang, G., Liu, Z., Van Der Maaten, L., Weinberger, K.Q.: Densely connected convolutional networks. In: CVPR (2017)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 567, + 480, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 567, + 480, + 588 + ], + "spans": [ + { + "bbox": [ + 138, + 567, + 480, + 588 + ], + "type": "text", + "content": "3. Huang, W.R., Geiping, J., Fowl, L., Taylor, G., Goldstein, T.: Metapoison: Practical general-purpose clean-label data poisoning. NeurIPS 33, 12080-12091 (2020)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 589, + 480, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 589, + 480, + 610 + ], + "spans": [ + { + "bbox": [ + 138, + 589, + 480, + 610 + ], + "type": "text", + "content": "4. Jiang, W., Li, H., Xu, G., Zhang, T.: Color backdoor: A robust poisoning attack in color space. In: CVPR. pp. 8133-8142 (2023)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 138, + 611, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 611, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 138, + 611, + 480, + 632 + ], + "type": "text", + "content": "5. Krizhevsky, A., Hinton, G., et al.: Learning multiple layers of features from tiny images (2009)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 138, + 632, + 449, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 632, + 449, + 643 + ], + "spans": [ + { + "bbox": [ + 138, + 632, + 449, + 643 + ], + "type": "text", + "content": "6. Le, Y., Yang, X.: Tiny imagenet visual recognition challenge. CS 231N 7, 7 (2015)" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 138, + 643, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 643, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 643, + 480, + 665 + ], + "type": "text", + "content": "7. LeCun, Y., Bottou, L., Bengio, Y., Haffner, P.: Gradient-based learning applied to document recognition. Proceedings of the IEEE 86(11), 2278-2324 (1998)" + } + ] + } + ], + "index": 21 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 266, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 266, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 266, + 91, + 447, + 102 + ], + "type": "text", + "content": "Undercover Bias for Dataset Copyright Protection" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 471, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 471, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 471, + 91, + 480, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 117, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 133, + 117, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 117, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 133, + 117, + 480, + 138 + ], + "type": "text", + "content": "18. Lee, J., Kim, E., Lee, J., Lee, J., Choo, J.: Learning debiased representation via disentangled feature augmentation. NeurIPS 34, 25123-25133 (2021)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 140, + 480, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 140, + 480, + 160 + ], + "spans": [ + { + "bbox": [ + 133, + 140, + 480, + 160 + ], + "type": "text", + "content": "19. Li, Y., Bai, Y., Jiang, Y., Yang, Y., Xia, S.T., Li, B.: Untargeted backdoor watermark: Towards harmless and stealthy dataset copyright protection. In: NeurIPS (2022)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 162, + 480, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 162, + 480, + 182 + ], + "spans": [ + { + "bbox": [ + 133, + 162, + 480, + 182 + ], + "type": "text", + "content": "20. Li, Y., Zhang, Z., Bai, J., Wu, B., Jiang, Y., Xia, S.T.: Open-sourced dataset protection via backdoor watermarking. NeurIPS Workshops (2020)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 184, + 480, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 184, + 480, + 205 + ], + "spans": [ + { + "bbox": [ + 132, + 184, + 480, + 205 + ], + "type": "text", + "content": "21. Li, Y., Li, Y., Wu, B., Li, L., He, R., Lyu, S.: Invisible backdoor attack with sample-specific triggers. In: ICCV. pp. 16463-16472 (2021)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 206, + 480, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 206, + 480, + 226 + ], + "spans": [ + { + "bbox": [ + 132, + 206, + 480, + 226 + ], + "type": "text", + "content": "22. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollár, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: ECCV. pp. 740-755 (2014)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 228, + 480, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 228, + 480, + 258 + ], + "spans": [ + { + "bbox": [ + 132, + 228, + 480, + 258 + ], + "type": "text", + "content": "23. Liu, G., Xu, T., Ma, X., Wang, C.: Your model trains on my data? protecting intellectual property of training data via membership fingerprint authentication. IEEE Transactions on Information Forensics and Security 17, 1024-1037 (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 260, + 480, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 260, + 480, + 281 + ], + "spans": [ + { + "bbox": [ + 132, + 260, + 480, + 281 + ], + "type": "text", + "content": "24. Liu, Y., Ma, X., Bailey, J., Lu, F.: Reflection backdoor: A natural backdoor attack on deep neural networks. In: ECCV. pp. 182-199 (2020)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 282, + 480, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 282, + 480, + 303 + ], + "spans": [ + { + "bbox": [ + 132, + 282, + 480, + 303 + ], + "type": "text", + "content": "25. Liu, Y., Ma, X., Bailey, J., Lu, F.: Reflection backdoor: A natural backdoor attack on deep neural networks. In: ECCV (2020)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 304, + 480, + 325 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 304, + 480, + 325 + ], + "spans": [ + { + "bbox": [ + 132, + 304, + 480, + 325 + ], + "type": "text", + "content": "26. Nam, J., Cha, H., Ahn, S., Lee, J., Shin, J.: Learning from failure: De-biasing classifier from biased classifier. NeurIPS 33, 20673-20684 (2020)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 327, + 480, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 327, + 480, + 346 + ], + "spans": [ + { + "bbox": [ + 132, + 327, + 480, + 346 + ], + "type": "text", + "content": "27. Ramaswamy, V.V., Kim, S.S., Russakovsky, O.: Fair attribute classification through latent space de-biasing. In: CVPR. pp. 9301-9310 (2021)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 348, + 480, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 348, + 480, + 369 + ], + "spans": [ + { + "bbox": [ + 132, + 348, + 480, + 369 + ], + "type": "text", + "content": "28. Ribeiro, M.T., Singh, S., Guestrin, C.: \"why should i trust you?\" explaining the predictions of any classifier. In: SIGKDD. pp. 1135-1144 (2016)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 370, + 480, + 391 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 370, + 480, + 391 + ], + "spans": [ + { + "bbox": [ + 132, + 370, + 480, + 391 + ], + "type": "text", + "content": "29. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: CVPR. pp. 10684-10695 (2020)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 392, + 480, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 392, + 480, + 413 + ], + "spans": [ + { + "bbox": [ + 132, + 392, + 480, + 413 + ], + "type": "text", + "content": "30. Ronneberger, O., Fischer, P., Brox, T.: U-net: Convolutional networks for biomedical image segmentation. In: MICCAI. pp. 234-241. Springer (2015)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 415, + 480, + 434 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 415, + 480, + 434 + ], + "spans": [ + { + "bbox": [ + 132, + 415, + 480, + 434 + ], + "type": "text", + "content": "31. Sablayrolles, A., Douze, M., Schmid, C., Jégou, H.: Radioactive data: tracing through training. In: ICML. pp. 8326-8335 (2020)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 436, + 480, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 436, + 480, + 456 + ], + "spans": [ + { + "bbox": [ + 132, + 436, + 480, + 456 + ], + "type": "text", + "content": "32. Saha, A., Subramanya, A., Pirsiavash, H.: Hidden trigger backdoor attacks. In: AAAI. vol. 34, pp. 11957-11965 (2020)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 458, + 480, + 489 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 458, + 480, + 489 + ], + "spans": [ + { + "bbox": [ + 132, + 458, + 480, + 489 + ], + "type": "text", + "content": "33. Schwarzschild, A., Goldblum, M., Gupta, A., Dickerson, J.P., Goldstein, T.: Just how toxic is data poisoning? a unified benchmark for backdoor and data poisoning attacks. In: ICML. pp. 9389-9398 (2021)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 491, + 480, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 491, + 480, + 511 + ], + "spans": [ + { + "bbox": [ + 132, + 491, + 480, + 511 + ], + "type": "text", + "content": "34. Shafahi, A., Huang, W.R., Najibi, M., Suciu, O., Studer, C., Dumitras, T., Goldstein, T.: Poison frogs! targeted clean-label poisoning attacks on neural networks. NeurIPS 31 (2018)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 513, + 480, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 513, + 480, + 544 + ], + "spans": [ + { + "bbox": [ + 132, + 513, + 480, + 544 + ], + "type": "text", + "content": "35. Souri, H., Fowl, L., Chellappa, R., Goldblum, M., Goldstein, T.: Sleeper agent: Scalable hidden trigger backdoors for neural networks trained from scratch. NeurIPS 35, 19165-19178 (2022)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 132, + 545, + 480, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 545, + 480, + 566 + ], + "spans": [ + { + "bbox": [ + 132, + 545, + 480, + 566 + ], + "type": "text", + "content": "36. Srivastava, N., Hinton, G., Krizhevsky, A., Sutskever, I., Salakhutdinov, R.: Dropout: a simple way to prevent neural networks from overfitting. JMLR 15(1), 1929-1958 (2014)" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 132, + 567, + 480, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 567, + 480, + 588 + ], + "spans": [ + { + "bbox": [ + 132, + 567, + 480, + 588 + ], + "type": "text", + "content": "37. Tan, M., Le, Q.: Efficientnet: Rethinking model scaling for convolutional neural networks. In: ICML. pp. 6105-6114 (2019)" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 132, + 590, + 480, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 590, + 480, + 609 + ], + "spans": [ + { + "bbox": [ + 132, + 590, + 480, + 609 + ], + "type": "text", + "content": "38. Tompson, J., Goroshin, R., Jain, A., LeCun, Y., Bregler, C.: Efficient object localization using convolutional networks. In: CVPR. pp. 648-656 (2015)" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 132, + 611, + 480, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 611, + 480, + 643 + ], + "spans": [ + { + "bbox": [ + 132, + 611, + 480, + 643 + ], + "type": "text", + "content": "39. Touvron, H., Bojanowski, P., Caron, M., Cord, M., El-Nouby, A., Grave, E., Izacard, G., Joulin, A., Synnaeve, G., Verbeek, J., et al.: Resmlp: Feedforward networks for image classification with data-efficient training. ICLR (2021)" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 132, + 644, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 644, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 644, + 480, + 665 + ], + "type": "text", + "content": "40. Wang, T., Yao, Y., Xu, F., An, S., Tong, H., Wang, T.: An invisible black-box backdoor attack through frequency domain. In: ECCV. pp. 396-413. Springer (2022)" + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 165, + 91, + 212, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 165, + 91, + 212, + 102 + ], + "spans": [ + { + "bbox": [ + 165, + 91, + 212, + 102 + ], + "type": "text", + "content": "J. Jang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 237 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 148 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 148 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 148 + ], + "type": "text", + "content": "41. Wang, W., Xie, E., Li, X., Fan, D.P., Song, K., Liang, D., Lu, T., Luo, P., Shao, L.: Pvt v2: Improved baselines with pyramid vision transformer. Computational Visual Media pp. 1-10 (2022)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 149, + 482, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 149, + 482, + 171 + ], + "spans": [ + { + "bbox": [ + 130, + 149, + 482, + 171 + ], + "type": "text", + "content": "42. Wang, Z., Bovik, A.C., Sheikh, H.R., Simoncelli, E.P.: Image quality assessment: from error visibility to structural similarity. TIP 13(4), 600-612 (2004)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 172, + 481, + 193 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 172, + 481, + 193 + ], + "spans": [ + { + "bbox": [ + 132, + 172, + 481, + 193 + ], + "type": "text", + "content": "43. Xiao, H., Rasul, K., Vollgraf, R.: Fashion-mnist: a novel image dataset for benchmarking machine learning algorithms. arXiv preprint arXiv:1708.07747 (2017)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 194, + 481, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 194, + 481, + 216 + ], + "spans": [ + { + "bbox": [ + 132, + 194, + 481, + 216 + ], + "type": "text", + "content": "44. Zhang, J., Chen, D., Liao, J., Fang, H., Zhang, W., Zhou, W., Cui, H., Yu, N.: Model watermarking for image processing networks. In: AAAI. vol. 34, pp. 12805-12812 (2020)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 216, + 481, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 216, + 481, + 237 + ], + "spans": [ + { + "bbox": [ + 132, + 216, + 481, + 237 + ], + "type": "text", + "content": "45. Zhu, Z., Xie, L., Yuille, A.: Object recognition with and without objects. In: IJCAI. pp. 3609-3615 (2017)" + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 266, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 266, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 266, + 91, + 447, + 102 + ], + "type": "text", + "content": "Undercover Bias for Dataset Copyright Protection" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 471, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 471, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 471, + 91, + 480, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Rethinking Deep Unrolled Model for Accelerated MRI Reconstruction/27de6684-978b-4680-80b1-f68f3bc7d6a8_content_list.json b/2024/Rethinking Deep Unrolled Model for Accelerated MRI Reconstruction/27de6684-978b-4680-80b1-f68f3bc7d6a8_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..c550b26f93101428cd3e4dd987e652d5533ec15a --- /dev/null +++ b/2024/Rethinking Deep Unrolled Model for Accelerated MRI Reconstruction/27de6684-978b-4680-80b1-f68f3bc7d6a8_content_list.json @@ -0,0 +1,1749 @@ +[ + { + "type": "text", + "text": "Rethinking Deep Unrolled Model for Accelerated MRI Reconstruction", + "text_level": 1, + "bbox": [ + 220, + 140, + 782, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Bingyu Xin $^{1}$ , Meng Ye $^{1}$ , Leon Axel $^{2}$ , and Dimitris N. Metaxas $^{1}$", + "bbox": [ + 248, + 210, + 754, + 229 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Rutgers University, New Brunswick, NJ 08901, USA \n $^{2}$ New York University Langone Medical Center, New York, NY 10016, USA \nbx64@rutgers.edu, dnm@cs.rutgers.edu", + "bbox": [ + 243, + 238, + 756, + 282 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Magnetic Resonance Imaging (MRI) is a widely used imaging modality for clinical diagnostics and the planning of surgical interventions. Accelerated MRI seeks to mitigate the inherent limitation of long scanning time by reducing the amount of raw $k$ -space data required for image reconstruction. Recently, the deep unrolled model (DUM) has demonstrated significant effectiveness and improved interpretability for MRI reconstruction, by truncating and unrolling the conventional iterative reconstruction algorithms with deep neural networks. However, the potential of DUM for MRI reconstruction has not been fully exploited. In this paper, we first enhance the gradient and information flow within and across iteration stages of DUM, then we highlight the importance of using various adjacent information for accurate and memory-efficient sensitivity map estimation and improved multi-coil MRI reconstruction. Extensive experiments on several public MRI reconstruction datasets show that our method outperforms existing MRI reconstruction methods by a large margin. The code is available at https://github.com/hellopipu/PromptMR-plus.", + "bbox": [ + 261, + 320, + 743, + 559 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: MRI Reconstruction $\\cdot$ Deep Unrolled Model $\\cdot$ Gradient-based Learning $\\cdot$ Sensitivity Map Estimation", + "bbox": [ + 261, + 570, + 740, + 599 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 625, + 375, + 641 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Magnetic Resonance Imaging (MRI) provides a radiation-free and highly versatile method for imaging the organs, tissues, and skeletal systems within the human body. Over the past 50 years since Paul Lauterbur produced the first MR image, MRI has evolved to become a cornerstone in clinical diagnostics [19]. However, the process of raw $k$ -space data acquisition in MRI is typically time-consuming. Accelerated MRI techniques tackle this issue by minimizing the amount of raw data that needs to be collected for image reconstruction, thereby shortening the duration of the scan. Modern advances in MRI technology, including Parallel Imaging (PI) [8] and Compressed Sensing (CS) [7,22], have significantly enhanced the efficiency and quality of MRI scans, making it possible to acquire high-resolution images within a considerably reduced time frame. However, with increased patient throughput and the requirements of emerging", + "bbox": [ + 212, + 657, + 787, + 843 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/900094bd3e4a914e7a993ea5ddc6787739ca1aa21f3b1d95f1ea041d73e3bd3e.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 225, + 142, + 415, + 248 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/68c3288d72aee1772b6348c68e35071fc8ec10802088cfed0023d393f3268dc3.jpg", + "image_caption": [ + "Fig. 1: Our proposed method is memory-efficient and shows better reconstruction performance. (a) GPU memory consumption on Calgary-Campinas brain dataset, (b) PSNR/SSIM results vs. the number of parameters on the fastMRI knee $\\times 8$ testset." + ], + "image_footnote": [], + "bbox": [ + 416, + 143, + 598, + 248 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/72f4807090649c20b5a47c7a59c6ad35b2c4c216f389f9873d4d62338971ffed.jpg", + "image_caption": [ + "(b)" + ], + "image_footnote": [], + "bbox": [ + 598, + 143, + 776, + 248 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "technologies, such as real-time MRI [25] and low-field MRI [3], the development of accelerated MRI reconstruction methods remains a hot research topic [24].", + "bbox": [ + 212, + 343, + 784, + 375 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Accelerated MRI reconstruction is a regularized inverse problem, which aims to reconstruct an unknown MR image from highly undersampled measurements in $k$ -space. Conventional iterative MRI reconstruction methods minimize a cost function comprising two main components: a data-consistency term, which assesses the alignment between $k$ -space predicted from the reconstructed image and the observed measurements, and a regularization term, which incorporates prior knowledge to encourage the emergence of desirable image attributes, e.g., sparsity. Over the past decade, deep learning [20] has emerged as a transformative approach to MRI reconstruction [26, 37]. Recently, the release of several large-scale public MRI reconstruction benchmarks, including the fastMRI dataset [49], Calgary-Campinas dataset [39], and CMRxRecon dataset [44], has significantly propelled the advancement of MRI reconstruction methods. Among deep learning-based MRI reconstruction approaches, the deep unrolled model (DUM) has garnered significant attention for its exceptional performance and ability to establish a concrete and systematic link between widely used iterative MRI reconstruction methods and deep neural networks [23]. However, as pointed out in [51], the current architecture design of DUM is inefficient, characterized by limited information transmission capacity and low flexibility and robustness. These shortcomings significantly restrict DUM's performance and its applicability to unseen data.", + "bbox": [ + 212, + 375, + 785, + 676 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this work, we rethink and improve key components of the DUM-based MRI reconstruction approach. Our main contributions can be summarized as follows:", + "bbox": [ + 212, + 676, + 785, + 722 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We improve the adaptive gradient algorithm and apply it to DUM, which can achieve self-adaptive dynamic learning rates adjusting for different spatial areas in an MR image. This approach provides a more flexible updating strategy in each iteration stage of DUM.", + "- We incorporate the momentum technique used in gradient descent acceleration and propose a multi-stage and multi-level feature aggregation scheme to accelerate the iteration convergence of DUM." + ], + "bbox": [ + 225, + 733, + 784, + 839 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "B. Xin et al.", + "bbox": [ + 271, + 114, + 357, + 127 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "- We highlight the importance of adjacent information which can be used to improve multi-coil MRI reconstruction, especially for accurate and memory-efficient sensitivity map estimation. As shown in Fig. 1(a), our method reduces GPU consumption to $\\sim 55\\%$ than the previous state-of-the-art.", + "bbox": [ + 225, + 146, + 784, + 205 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "- We demonstrate the effectiveness of our proposed method on three public MRI reconstruction benchmarks of different anatomies: the fastMRI knee, the Calgary-Campinas brain and the CMRxRecon cardiac dataset. As shown in Fig. $\\boxed{1}$ (b), our method achieves better PSNR and SSIM with fewer parameters than existing MRI Reconstruction methods.", + "bbox": [ + 227, + 207, + 785, + 282 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 306, + 387, + 321 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.1 Deep Unrolled Model (DUM)", + "text_level": 1, + "bbox": [ + 215, + 340, + 506, + 356 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "DUM maps a truncated optimization algorithm into a deep neural network, iteratively alternating between gradient descent steps and proximal mapping steps. Its superior performance and good interpretability make it a preferred method for a wide range of inverse problems, including compressive sensing, image restoration and image reconstruction [23, 51]. Learned ISTA (LISTA) was the first unrolling method proposed for fast approximation of sparse coding [13]. LISTA unrolls the traditional ISTA algorithm [5] and truncates it to a fixed number of iterations. By learning from a training dataset, this trainable version of ISTA shows significant computational benefits. The success of deep learning further promotes the idea of relating conventional iterative optimization methods to deep neural networks, including ADMM-Net [41], ISTA-Net [52] and AMP-Net [54]. By introducing a condition module to transmit input information to each stage of the unrolled model, ISTA-Net++ [48] provides more flexibility to handle multi-ratio and multi-scene images for compressive sensing. In MADUN [38], a memory-augmented DUM is proposed to enhance the short-term and long-term information transmission in traditional DUMs.", + "bbox": [ + 212, + 366, + 787, + 608 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2 DUM for MRI Reconstruction", + "text_level": 1, + "bbox": [ + 215, + 632, + 514, + 645 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "ADMM-Net [41] and DC-CNN [36] are the pioneering DUMs for MRI reconstruction, establishing themselves as widely used baseline DUM-based MRI reconstruction methods. These approaches, based on Convolutional Neural Networks (CNNs), unroll the alternating direction method of multipliers (ADMM) [6] and the alternating minimization (AM) algorithm, respectively. KIKI-Net [11] employs a novel approach by alternating between $k$ -space completion blocks and image restoration blocks, demonstrating the advantages of cross-domain learning [10,21,56]. Adaptive-CS-Net finds it beneficial to inject information from neighboring slices into the model. This 2.5D reconstruction strategy has been adopted in subsequent works [12,45].", + "bbox": [ + 212, + 657, + 787, + 810 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The development of DUM has led to various innovative approaches beyond simple unrolling, categorized broadly into three main variants for enhanced", + "bbox": [ + 212, + 810, + 785, + 839 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "Rethinking Deep Unrolled Model for Accelerated MRI Reconstruction", + "bbox": [ + 264, + 114, + 730, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "MRI reconstruction. (1) Firstly, inspired by Recurrent Neural Network (RNN), RIM [31, 32] uses a latent variable as temporal memory to track the progression of iterations. Recurrent-VarNet [46] extends this idea by designing a recurrent state initializer and recurrent units for the hidden state. In parallel, LPD-Net [1] and NCPD-Net [34] simplify this concept by utilizing a buffer design as additional learnable memory transmitting across iterations. (2) Secondly, inspired by efficient variants of proximal gradient descent (PGD), HC-PGD [16] proposes the history-cognizant unrolling method by concatenating all previous proximal operator output as input to each iteration. DIRCN [27] further proposes to add feature map interconnections between neighboring stages. These works suggest the importance of gradient flow and shared information between cascades for MRI reconstruction. (3) Moreover, Adaptive-CS-Net [28] incorporates soft MRI priors, including soft data consistency, image phase information and background priors, as additional input channels to each iteration. Similarly, HiTDUN [53] proposes the design of multi-channel input for every iteration. These methods underscore the advantages of introducing additional priors into the input at each iteration stage, aiming to dismantle the information bottleneck encountered within each iteration.", + "bbox": [ + 212, + 145, + 787, + 417 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.3 Sensitivity Map Estimation in DUM", + "text_level": 1, + "bbox": [ + 215, + 452, + 565, + 468 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Numerous works on DUM have achieved promising results on MRI reconstruction [11, 21, 36, 41, 56]. However, most of these methods focus exclusively on synthesized single-coil $k$ -space data, overlooking the more prevalent multi-coil imaging in real clinical MRI scans. This oversight neglects a more realistic scenario, given the widespread application of parallel imaging. Parallel imaging (PI) can speed up MRI, with the idea of using multiple receiving coils to reduce the number of phase encoding lines needed in acquiring raw $k$ -space data. The sensitivity map of a given coil is its local field profile, describing through reciprocity where that coil can efficiently pick up MR signals from the imaged region. The reconstruction of multiple coil $k$ -space data needs accurate estimation of the sensitivity maps, which are either directly measured, as in SENSE [30], or indirectly accounted using a center region of $k$ -space data, as in GRAPPA [14].", + "bbox": [ + 212, + 489, + 787, + 671 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Robust and accurate sensitivity map estimation is crucial for multi-coil MRI reconstruction. The i-RIM directly reconstructs coil images without explicitly estimating the sensitivity maps. Adaptive-CS-Net [28] uses the normalized low-pass filtered coil images as sensitivity maps. VarNet [15] and MoDL [2] handle multi-coil data with pre-estimated sensitivity maps [43]. E2E-VarNet [40] advances the VarNet by learning sensitivity maps from the Auto-Calibration Signal (ACS) data of the $k$ -space in an end-to-end manner. Due to its superior performance and end-to-end approach, many subsequent studies are based on E2E-VarNet [12,45,46]. When the ACS data is insufficient for accurate sensitivity map estimation, [4,17] suggest iterative sensitivity map estimation and image reconstruction, which can be considered as a deep version of J-SENSE [47].", + "bbox": [ + 212, + 672, + 787, + 843 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "B. Xin et al.", + "bbox": [ + 271, + 114, + 357, + 127 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/22c09dc106ed676d61d4256b30a5b6b94d61e66cb21d2087cc119876f13995c2.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 241, + 141, + 485, + 228 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/1732aed41d26115f74ed88d6b55d419b0a902f9e5018c214ce14924541f7f320.jpg", + "image_caption": [ + "(b)", + "Fig. 2: Overview of our proposed (a) self-adaptive and (b) momentum-accelerated gradient algorithm. $\\hat{x}_{t+1}$ and $\\hat{F}_t^m$ are the output of CNN and Momentum layer in Eq. (7) and Eq. (10), respectively." + ], + "image_footnote": [], + "bbox": [ + 550, + 141, + 727, + 228 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3 Proposed Method", + "text_level": 1, + "bbox": [ + 215, + 321, + 426, + 339 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Drawing inspiration from adaptive gradient and momentum-accelerated algorithms in gradient descent optimization, we dynamically incorporate prior knowledge in the data flow, to enable more informative and faster convergence for gradient-based MRI reconstruction learning, as shown in Fig. 2. Subsequently, we enhance the concept of adjacent reconstruction by focusing on more robust and memory-efficient multi-coil sensitivity map estimation, as shown in Fig. 4. Please refer to Supplementary Section B for architecture details.", + "bbox": [ + 212, + 353, + 787, + 460 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1 Problem Formulation", + "text_level": 1, + "bbox": [ + 215, + 481, + 439, + 494 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We can estimate the complex MR image $\\pmb{x}$ from its undersampled measurements $\\pmb{y}$ in $k$ -space by solving an optimization problem [22],", + "bbox": [ + 212, + 506, + 785, + 537 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {\\boldsymbol {x}} \\boldsymbol {f} (\\boldsymbol {x}) = \\min _ {\\boldsymbol {x}} \\frac {1}{2} \\left\\| \\boldsymbol {A} \\boldsymbol {x} - \\boldsymbol {y} \\right\\| _ {2} ^ {2} + \\boldsymbol {R} (\\boldsymbol {x}), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 361, + 546, + 785, + 575 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathbf{A} = M\\mathbf{F}\\mathbf{S}$ is the forward process of acquiring the $k$ -space measurements, which is a combination of coil sensitivity maps $\\mathbf{S}$ , the Fourier transform $\\mathcal{F}$ and an undersampling mask $\\mathbf{M}$ . The first term $\\frac{1}{2} ||\\mathbf{A}\\mathbf{x} - \\mathbf{y}||_2^2$ represents the data consistency term, which ensures data fidelity, while the second term $\\mathbf{R}(\\mathbf{x})$ serves as the regularization term, aimed at promoting image domain priors, such as sparsity [7]. We can solve Eq. (1) iteratively via the gradient descent method,", + "bbox": [ + 212, + 584, + 787, + 676 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {x} _ {t + 1} = \\boldsymbol {x} _ {t} - \\eta_ {t} \\boldsymbol {g} _ {t}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 379, + 702, + 785, + 718 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {g} _ {t} = \\boldsymbol {A} ^ {H} (\\boldsymbol {A} \\boldsymbol {x} _ {t} - \\boldsymbol {y}) + \\nabla \\boldsymbol {R} (\\boldsymbol {x} _ {t}), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 395, + 720, + 785, + 739 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\nabla R(\\pmb{x}_t)$ is the gradient or the proximal mapping of $\\pmb{R}$ , $\\eta^t$ is the scalar learning rate at iteration $t$ , and $t \\in \\{0,1,\\dots,T-1\\}$ , with $T$ being the total number of truncated iterations. In the context of the deep unrolled models, it is customary to substitute $\\nabla R(\\pmb{x}_t)$ with a neural network whose parameters are learned from large-scale training datasets. This allows for the learning of more sophisticated image domain priors, compared to those that are hand-crafted.", + "bbox": [ + 212, + 748, + 787, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "Rethinking Deep Unrolled Model for Accelerated MRI Reconstruction", + "bbox": [ + 264, + 113, + 732, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2 Self-Adaptive Gradient Algorithm", + "text_level": 1, + "bbox": [ + 215, + 146, + 542, + 162 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "During the iterative MRI reconstruction process, some regions in an image are straightforward to reconstruct, such as the background, which contains minimal information. However, other areas are more prone to aliasing artifacts resulting from downsampling the $k$ -space, which are challenging to eliminate. This difficulty often arises because the aliasing artifacts may closely mimic the appearance of anatomical structures within these regions, necessitating a more meticulous gradient-based correction at each iteration. Inspired by the adaptive gradient algorithm [9,18,42,50], instead of applying a uniform learning rate across the entire image space at each iteration, it is reasonable to utilize self-adaptive learning rates for individual pixels to achieve a more refined reconstruction result. Our generalization of the standard gradient descent in Eq. (2) employs the following update for each iteration:", + "bbox": [ + 212, + 172, + 787, + 354 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {x} _ {t + 1} = \\boldsymbol {x} _ {t} - \\eta_ {t} \\boldsymbol {g} _ {t} - \\boldsymbol {E} _ {t} \\odot \\boldsymbol {g} _ {t}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 397, + 367, + 785, + 385 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {E} _ {t} = \\frac {\\eta_ {t}}{\\sqrt {\\sum_ {\\tau = 0} ^ {t} \\boldsymbol {g} _ {\\tau} ^ {2}} + \\epsilon} - \\eta_ {t}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 411, + 387, + 785, + 426 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\pmb{E}_{t}$ is the residual pixel-wise learning rate at the $t$ -th iteration, and the denominator in Eq. (5) computes the root of the sum of squares of all previous gradients of individual pixels. $\\epsilon$ is a small number to improve the numerical stability. The operation $\\odot$ denotes element-wise multiplication. We note that, the accumulation of historical pixel-wise gradients in Eq. (5) can help improve iteration convergence by", + "bbox": [ + 212, + 436, + 787, + 527 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "(1) scaling down the updating steps for regions with historically large gradients, which may be prone to overshooting;", + "(2) scaling up the updating steps for regions with historically small gradients, which may benefit from a more aggressive update." + ], + "bbox": [ + 212, + 539, + 785, + 599 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To unroll our self-adaptive gradient algorithm with a DUM, we deviate from the common practice of substituting $\\nabla R(\\boldsymbol{x}_t)$ with a neural network in Eq. (4). There are compelling reasons for this. First, learning the $\\nabla R(\\boldsymbol{x}_t)$ term in $\\boldsymbol{g}_t$ (see Eq. (3)) can lead to numerical instability in $\\boldsymbol{E}_t$ during the early stage of training. Second, the adaptive gradient strategy in Eq. (5), which is the basic variant utilized in AdaGrad (9), may not be the most suitable choice for a DUM in the context of MRI reconstruction. To leverage the robust nonlinear representation capabilities of neural networks, we improve upon the adaptive gradient algorithm by reformulating Eq. (4) as follows:", + "bbox": [ + 212, + 611, + 787, + 750 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\boldsymbol {x} _ {t + 1} = \\boldsymbol {x} _ {t} - \\eta_ {t} \\boldsymbol {A} ^ {H} (\\boldsymbol {A} \\boldsymbol {x} _ {t} - \\boldsymbol {y}) - \\left(\\eta_ {t} \\nabla \\boldsymbol {R} (\\boldsymbol {x} _ {t}) + \\boldsymbol {E} _ {t} \\odot \\boldsymbol {g} _ {t}\\right) (6) \\\\ = \\boldsymbol {x} _ {t} - \\eta_ {t} \\boldsymbol {A} ^ {H} (\\boldsymbol {A} \\boldsymbol {x} _ {t} - \\boldsymbol {y}) + \\operatorname {C N N} _ {t} \\left(\\boldsymbol {x} _ {t}, \\boldsymbol {A} ^ {H} \\boldsymbol {A} \\boldsymbol {x} _ {t}, \\boldsymbol {A} ^ {H} \\boldsymbol {y}, \\boldsymbol {s} _ {t}\\right). (7) \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 279, + 758, + 785, + 797 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For the last term in Eq. (6), we reparametrize it with a CNN, as indicated in Eq. (7). This neural network takes as input the fundamental components from", + "bbox": [ + 212, + 808, + 787, + 842 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "B. Xin et al.", + "bbox": [ + 271, + 114, + 357, + 127 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "which the term is constructed, namely, $\\boldsymbol{x}_t, \\boldsymbol{A}^H \\boldsymbol{A} \\boldsymbol{x}_t, \\boldsymbol{A}^H \\boldsymbol{y}$ , and a learnable auxiliary memory $s_t$ , which serves as a hidden state to track the historical gradient information implicitly. Note that $\\boldsymbol{A}^H \\boldsymbol{y}$ yields $\\boldsymbol{x}_0$ , which represents the image transformed from zero-filled $k$ -space. For further discussion, please refer to Supplementary Section E.", + "bbox": [ + 212, + 145, + 782, + 220 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.3 Momentum-Accelerated Gradient Algorithm", + "text_level": 1, + "bbox": [ + 214, + 247, + 629, + 262 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The momentum method [29] accelerates gradient descent in Eq. (2) by accumulating a velocity vector in directions of persistent reduction in the optimization objective across iterations:", + "bbox": [ + 212, + 273, + 784, + 320 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {x} _ {t + 1} = \\boldsymbol {x} _ {t} - \\eta_ {t} \\boldsymbol {m} _ {t}, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 403, + 335, + 784, + 349 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {m} _ {t} = \\mu \\boldsymbol {m} _ {t - 1} + (1 - \\mu) \\boldsymbol {g} _ {t}, \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 415, + 354, + 784, + 369 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $\\mu \\in [0,1]$ is the momentum coefficient. The momentum $m_{t}$ is known as the exponential moving average of historical gradients. It can benefit the optimization of the current iteration, thereby considerably accelerating convergence. In Sec. 3.2 we introduced the hidden state $s_t$ , which is used to implicitly track the pixel-wise historical gradient information. Here, we propose explicit multi-stage and multi-level feature aggregation, to mimic the idea of momentum and increase the information flow across the iterations. For a typical multi-level encoder-decoder denoiser, such as Unet, in the $t$ -th iteration, we insert a Momentum layer before the $m$ -th level decoder layer. This layer concatenates features at the current stage and features from all previous iteration stages at the same level, followed by a $1\\times 1$ convolution layer to reduce the dimensionality, and a Channel Attention Block (CAB) to adaptively fuse the multi-stage information, as described below:", + "bbox": [ + 212, + 381, + 787, + 577 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {M o m e n t u m} \\left(\\boldsymbol {F} _ {0} ^ {m}, \\dots , \\boldsymbol {F} _ {t} ^ {m}\\right) = \\operatorname {C A B} \\left(\\operatorname {C o n v} _ {1 \\times 1} \\left(\\operatorname {C o n c a t} \\left(\\boldsymbol {F} _ {0} ^ {m}, \\dots , \\boldsymbol {F} _ {t} ^ {m}\\right)\\right)\\right), \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 254, + 590, + 784, + 608 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $\\pmb{F}_t^m$ is the input feature of the $m$ -th level decoder layer at the $t$ -th iteration.", + "bbox": [ + 214, + 619, + 782, + 635 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.4 Adjacent Reconstruction", + "text_level": 1, + "bbox": [ + 214, + 660, + 465, + 676 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Incorporating information from adjacent slices in multi-coil MRI reconstruction has been substantiated in prior work [12,28,45]. However, a significant limitation of adjacent slice reconstruction lies in its substantial memory consumption during the estimation of multi-coil sensitivity maps. This issue stems from the design of the sensitivity map estimation (SME) network in earlier work [40]. The network adopts a coil-by-coil estimation approach to accommodate a variable number of coils, increasing flexibility but leading to higher memory demands in the adjacent reconstruction setting (see Fig. 4(a)). To be more specific, for $k$ -space data with $N$ coils and $2a + 1$ adjacent slices, the total number of forward passes required for the SME process is calculated as $(2a + 1)N$ . Typically, the", + "bbox": [ + 212, + 688, + 787, + 840 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "Rethinking Deep Unrolled Model for Accelerated MRI Reconstruction", + "bbox": [ + 264, + 114, + 732, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/fe41e60de2c1d27f5b953c5ce27269df8a37cbd0d653b8629ab4995b3f832f80.jpg", + "image_caption": [ + "(a) time" + ], + "image_footnote": [], + "bbox": [ + 217, + 143, + 390, + 223 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/b9bb7b06238ea4f60be30ca52cb58101901d4c084f9186904a933bb117f8e79f.jpg", + "image_caption": [ + "(b) contrast" + ], + "image_footnote": [], + "bbox": [ + 415, + 143, + 586, + 223 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/dc95bd762fc55127bc99165074ca24869f845aad1874e00ca6ed85fcbc380ef3.jpg", + "image_caption": [ + "(c) slice" + ], + "image_footnote": [], + "bbox": [ + 612, + 143, + 782, + 223 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/042a2753b14a7f22137e73b6ec121a1f9e3ed1e552ced03aa722ea922202d833.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 217, + 304, + 491, + 377 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/d76db55f95f6809fbacded300b1ffd80f6685949b2743145013b14de5427cf57.jpg", + "image_caption": [ + "Fig. 3: The sensitivity maps exhibit variations across (a) time, (b) contrast and (c) slice dimension. First row: original images, second row: magnitude of differences between the sensitivity map of the central image and each of those of its adjacent images." + ], + "image_footnote": [], + "bbox": [ + 509, + 303, + 784, + 377 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/d250f998439938584824d2ea82078788f8b04989a7593502658df666ea70d4c1.jpg", + "image_caption": [ + "Fig. 4: Concept illustration of different strategies for multi-coil sensitivity map estimation (SME). The main difference between the four strategies lies in the Auto-Calibration-Signal (ACS) information sharing between coils and adjacent MR image slices. Each single coil sensitivity map gets ACS information from (a) a single coil and a single image slice; (b) multiple coils but a single image slice; (c) a single coil but multiple adjacent image slices; (d) multiple coils and multiple adjacent image slices." + ], + "image_footnote": [], + "bbox": [ + 217, + 385, + 491, + 460 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/23ea4db9164c7d3b1378b939654ffbef2b78f5a13940402bbafda1bf7e1e8ac2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 509, + 385, + 784, + 460 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "value of $N$ ranges from 8 to 34 and $a$ is set to 2. We propose to reduce the number of forward passes, by utilizing the correlations between adjacent sensitivity maps.", + "bbox": [ + 212, + 585, + 787, + 630 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "A sensitivity map is subject to variation, based on the position of the imaging subject and changes over time due to the subject motion. Additionally, in multicontest contrast MRI acquisition, variations in acquisition parameters, such as inversion time (TI) and repetition time (TR), can indirectly influence the sensitivity map estimation through the acquired $k$ -space data, as shown in Fig. ③. Consequently, the estimation of adjacent sensitivity maps is a spatially, temporally, and contrastively correlated process. We therefore propose to estimate sensitivity maps by utilizing information from spatially, temporally, and contrastively adjacent slices. As shown in Fig. ④(c), we simultaneously estimate a single coil sensitivity map at all adjacent $2a + 1$ slices using information across these slices. In this way, we can reduce the number of forward passes from $(2a + 1)N$ to $N$ . Our estimation of the $n$ -th sensitivity map set $S_{adj}^{n}$ can be described as:", + "bbox": [ + 212, + 631, + 787, + 815 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {S} _ {a d j} ^ {n} = \\mathbf {S M E} (\\mathbf {A C S} \\left(\\boldsymbol {x} _ {0, a d j} ^ {n}\\right)), \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 397, + 824, + 785, + 842 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "B. Xin et al.", + "bbox": [ + 271, + 114, + 357, + 127 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $S_{adj}^{n} = [S_{a - c}^{n},\\dots,S_{c}^{n},\\dots,S_{a + c}^{n}]$ is an adjacent sensitivity map set which is the concatenation of central slice sensitivity map $S_{c}^{n}$ with its 2a-adjacent slice sensitivity maps of the $n$ -th coil. $\\mathbf{ACS}(\\pmb{x}_{0,adj}^{n})$ is the $n$ -th coil Auto-Calibration Signal (ACS) within all adjacent $2a + 1$ slices.", + "bbox": [ + 212, + 146, + 787, + 210 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4 Experiments and Results", + "text_level": 1, + "bbox": [ + 214, + 228, + 495, + 244 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "In this section, we present the experimental results of our proposed approach. We showcase the reconstruction performance of our model across multiple publicly available datasets, including Calgary-Campinas brain [39], CMRxRecon cardiac [44], and fastMRI multi-coil knee [49] datasets. These datasets encompass a diverse range of anatomies, contrasts, views, slices and motion states. For details about the datasets, please refer to Supplementary Section A. For information on the experimental setup for each dataset, see Supplementary Section C.", + "bbox": [ + 212, + 255, + 787, + 363 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.1 Calgary-Campinas Brain Results", + "text_level": 1, + "bbox": [ + 214, + 380, + 532, + 395 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Tab. 1 presents the quantitative results for $\\times 5$ and $\\times 10$ acceleration factors on the Calgary-Campinas brain test set. Our method is compared with two state-of-the-art (SOTA) methods: Recurrent-VarNet [46] and PromptMR [45]. Our proposed method significantly outperforms these benchmarks, demonstrating its outstanding performance. Qualitative results under $\\times 10$ acceleration are shown in Fig. 5. Interestingly, while PromptMR achieves significantly higher PSNR and SSIM values compared to Recurrent-VarNet, it exhibits less accurate reconstruction in the area highlighted by the red box.", + "bbox": [ + 212, + 401, + 787, + 521 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "The highly undersampled $k$ -space data makes the ill-posed reconstruction challenging at aliasing artifacts removal. While leveraging large-scale training data to learn the image priors can improve the reconstruction accuracy, we note that Recurrent-VarNet learns a weaker prior, resulting in reconstructions that appear blurry. Conversely, PromptMR adopts a strong but overly biased prior, leading to sharper reconstructions but introducing incorrect structures. Our proposed method, in contrast, successfully learns the appropriate prior, achieving a reconstruction that closely mirrors the ground truth in terms of structural accuracy.", + "bbox": [ + 212, + 522, + 787, + 657 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.2 CMRxRecon Cardiac Results", + "text_level": 1, + "bbox": [ + 214, + 676, + 506, + 690 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We compare our method with PromptMR, the winning method of the MIC-CAI2023 CMRxRecon challenge, for $\\times 4$ , $\\times 8$ and $\\times 10$ acceleration factors on the CMRxRecon cardiac test set. In Tab. ②, our method exhibits superior performance to PromptMR across different views and contrasts. Qualitative results under $\\times 10$ acceleration are shown in Fig. ⑥. Our method reconstructs fine details of the interventricular septum, highlighted in the red box, with significantly higher clarity than the PromptMR method, which is severely compromised by aliasing artifacts.", + "bbox": [ + 212, + 696, + 787, + 819 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "Rethinking Deep Unrolled Model for Accelerated MRI Reconstruction", + "bbox": [ + 264, + 114, + 730, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 8 + }, + { + "type": "page_footnote", + "text": "1 It is usually a part of the acquired central $k$ -space data.", + "bbox": [ + 217, + 824, + 604, + 839 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/6eadf08b59fddd3c93a479304e155e70ae12c72cabac101c5c2bb02a72c24e20.jpg", + "table_caption": [ + "Table 1: Quantitative comparison of PSNR/SSIM (mean ± std) of different MRI reconstruction methods on the Calgary-Campinas brain dataset under $\\times 5$ and $\\times 10$ acceleration. The best and second best performance are in red and blue colors, respectively." + ], + "table_footnote": [], + "table_body": "
MethodAcc = 5×Acc = 10×
PSNR(dB)SSIM(%)PSNR(dB)SSIM(%)
Zero-filled25.27±1.3173.75±3.9324.35±1.2169.04±3.67
Recurrent-Varnet36.27±1.7294.37±1.3333.27±2.1291.47±2.23
PromptMR36.83±1.6894.85±1.2634.10±2.1192.69±1.92
Ours37.24±1.5795.15±1.2134.74±1.9793.38±1.65
", + "bbox": [ + 305, + 212, + 697, + 309 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/729c05983d97094615657e658bef127cb0dd05a3a6997eb5dd391418286572d5.jpg", + "image_caption": [ + "Fig. 5: The reconstruction results and SSIM error maps of various methods for the Calgary-Campinas brain imaging under $\\times 10$ acceleration factor. The red boxes highlight the differences in the recovery of white and gray matter structures." + ], + "image_footnote": [], + "bbox": [ + 300, + 320, + 700, + 477 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.3 FastMRI Knee Results", + "text_level": 1, + "bbox": [ + 214, + 556, + 452, + 571 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "We evaluated the PSNR and SSIM at acceleration factors of $\\times 4$ and $\\times 8$ on the fastMRI multi-coil knee test set. Our method is compared with several methods, including SOTA methods HUMUS-Net-L [12] and PromptMR. In the table, 'PD' and 'PDFS' denote proton density-weighted images without and with fat suppression, respectively. Tab. 3 demonstrates that our approach achieves the best performance across all other competitive methods on this large-scale dataset. Qualitative results under $\\times 8$ acceleration are shown in Fig. 7. Our approach achieves a more detailed reconstruction of the meniscus region as shown in the red box, closely approximating the ground truth. In this region, meniscus tears were identified by radiologists from the fastMRI+ dataset [55]. The qualitative comparison suggests that our approach is more robust in reconstructing abnormal knee regions for highly accelerated acquisition.", + "bbox": [ + 212, + 579, + 787, + 760 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "5 Abalation study", + "text_level": 1, + "bbox": [ + 214, + 780, + 408, + 797 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "In this section, we justify our design choices through ablation studies on the brain and cardiac datasets.", + "bbox": [ + 212, + 809, + 785, + 839 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "B. Xin et al.", + "bbox": [ + 271, + 114, + 354, + 126 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/bcd8fb134867bb132425a97366919793016bf31c974b0c131f943f6c398e72e3.jpg", + "table_caption": [ + "Table 2: Quantitative comparison of PSNR/SSIM (mean ± std) of different MRI reconstruction methods on the CMRxRecon cardiac dataset under $\\times 4$ , $\\times 8$ and $\\times 10$ acceleration. The best and second best performance are in red and blue colors, respectively." + ], + "table_footnote": [], + "table_body": "
AccMethodCineMapping
SAXLAXT1wT2w
PSNR(dB)SSIM(%)PSNR(dB)SSIM(%)PSNR(dB)SSIM(%)PSNR(dB)SSIM(%)
4×Zero-filled PromptMR Ours26.20±1.5472.99±5.1425.11±1.6269.87±4.7324.23±1.1667.27±3.8424.97±1.1077.16±2.91
45.78±1.8998.78±0.4345.58±1.7298.72±0.3746.49±2.2398.93±0.5742.02±1.9398.05±0.72
46.26±1.7698.87±0.3446.13±1.7098.83±0.3447.45±2.1399.09±0.4242.68±1.8498.25±0.59
8×Zero-filled PromptMR Ours25.18±1.6770.21±5.8724.34±1.6667.67±5.8023.41±1.2564.12±4.4324.21±1.1274.84±3.07
40.65±1.5296.97±0.6839.64±1.8796.36±1.0440.94±1.6997.28±0.8538.24±1.8296.53±1.09
41.44±1.4897.32±0.5840.49±1.8396.76±0.9442.08±1.6397.70±0.6739.23±1.7196.98±0.89
10×Zero-filled PromptMR Ours24.83±1.6769.35±5.9224.13±1.6766.89±5.5823.16±1.2764.14±4.4724.25±1.1476.08±2.91
39.18±1.5096.15±0.8038.28±1.6295.60±1.0738.99±1.5896.61±0.9937.21±1.7696.22±1.16
39.99±1.4996.58±0.7239.13±1.6696.05±0.9940.37±1.5897.19±0.7938.22±1.7096.70±0.96
", + "bbox": [ + 220, + 212, + 782, + 366 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/eaae49564f24e3170a75b5299d01398aac83fd9d2241e3866bd1dc246c8868da.jpg", + "image_caption": [ + "Fig. 6: The reconstruction results and SSIM error maps of various methods for the CM-RxRecon cardiac T1-weighted imaging under $\\times 10$ acceleration factor. The red boxes highlight the differences in the recovery of the interventricular septum." + ], + "image_footnote": [], + "bbox": [ + 217, + 378, + 785, + 541 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/aff27bd971ba7eb0ff06a34733d2af9b8b016237ea37ce40253ba6c32db62d14.jpg", + "table_caption": [ + "Table 3: Quantitative comparison of PSNR/SSIM (mean ± std) of different MRI reconstruction methods on the fastMRI knee dataset under $\\times 4$ and $\\times 8$ acceleration. The best and second best performance are in red and blue colors, respectively." + ], + "table_footnote": [], + "table_body": "
MethodAcc = 4×Acc = 8×
PDPDSPDPDS
|PSNR(dB)|SSIM(%)|PSNR(dB)|SSIM(%)|PSNR(dB)|SSIM(%)|PSNR(dB)|SSIM(%)
Zero-filled30.86±1.7380.65±3.7631.00±3.3378.48±6.1627.70±1.8474.12±4.9427.97±2.0268.78±5.58
Unet 3537.27±1.7692.11±2.7237.07±2.4788.03±5.0435.41±2.1490.36±3.1334.84±1.5983.71±5.04
i-RIM 3339.61±2.3394.16±2.8738.23±3.1389.27±4.9637.91±2.4793.22±2.8535.84±1.7185.27±4.93
E2E-VarNet 4040.09±2.3294.56±2.6438.43±3.2489.50±4.9538.72±2.5093.93±2.6836.16±1.7685.70±4.94
HUMUS-Net 12----38.72±2.5094.07±2.6835.98±1.7585.77±4.93
HUMUS-Net-L 12----39.03±2.4794.18±2.6136.19±1.7985.84±4.92
PromptMR 4540.58±2.5994.87±2.7538.59±3.3389.75±4.9439.48±2.5494.51±2.5436.42±1.7986.10±4.93
Ours40.85±2.7495.00±2.8438.71±3.3789.85±4.9240.09±2.5894.90±2.4536.49±1.8886.29±4.95
", + "bbox": [ + 220, + 660, + 782, + 800 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "Rethinking Deep Unrolled Model for Accelerated MRI Reconstruction", + "bbox": [ + 264, + 114, + 730, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 116, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/a39cd262337756e98b622481f41381c2830f008b46985726e4dee9f87681ca91.jpg", + "image_caption": [ + "Fig. 7: The reconstruction results and SSIM error maps of various methods for the fastMRI multi-coil knee proton density (PD) imaging under $\\times 8$ acceleration. The red boxes highlight the differences in the recovery of the meniscus region. The red box in the reference image was annotated by radiologists from the fastMRI+ dataset [55] to indicate the knee abnormality of meniscus tears." + ], + "image_footnote": [], + "bbox": [ + 217, + 143, + 785, + 262 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/b9151a1c66962dfe5ee097262005b7ea06b4ab0892f00958fc84b98478645a56.jpg", + "table_caption": [ + "Table 4: Abalation of different input combinations of CNN on the Calgary-Campinas brain dataset under $\\times 10$ acceleration." + ], + "table_footnote": [], + "table_body": "
InputBaseline------Ours
xt✓✓✓✓✓✓✓✓
AHx✓✓✓✓✓✓✓✓
AHy✓✓✓✓✓✓✓✓
st✓✓✓✓✓✓✓✓
PSNR(dB)35.76+0.00+0.01+0.09+0.08+0.31+0.36+0.36
SSIM(%)94.33+0.00-0.01+0.06+0.05+0.25+0.27+0.28
", + "bbox": [ + 305, + 398, + 697, + 527 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "5.1 Input Variations to CNN", + "text_level": 1, + "bbox": [ + 215, + 560, + 472, + 575 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "We conducted an ablation study for different input combinations to CNN, to validate the effectiveness of our proposed method. Results are shown in Tab. 4. The baseline network in each iteration only accepts the output from the last iteration $\\pmb{x}_t$ as input, which is the setting for most unrolled models, such as E2E-VartNet and PromptMR. Our proposed method takes additional input, $A^H A\\pmb{x}_t$ , $A^H\\pmb{y}$ and $s_t$ , which outperforms the baseline by a large margin. More specifically, adding $A^H A\\pmb{x}_t$ or $A^H\\pmb{y}$ solely to the input has no impact on the performance. Adding $s_t$ can improve the performance, as it serves as additional memory to track the progression of the reconstruction process. Adding $A^H A\\pmb{x}_t$ together with $A^H\\pmb{y}$ greatly improves the gains. We assume the reason is that the residual connection of two terms into the network can provide soft data consistency information to the model. Adding $A^H A\\pmb{x}_t$ together with $s_t$ performs the closest to our final design. It seems the missing of $A^H\\pmb{y}$ has little impact, as it is a constant value and its information may be compactly represented in the $s_t$ . This assumption is implied by the evidence that the result of adding $A^H\\pmb{y}$ together with $s_t$ into the network does not perform better than adding $s_t$ alone.", + "bbox": [ + 212, + 597, + 787, + 843 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "B. Xin et al.", + "bbox": [ + 271, + 114, + 357, + 126 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/b38d297ac79b3d067b5996e839fb1a231f0841442e9eadb02f9ffb070871bc74.jpg", + "table_caption": [ + "Table 5: Quantitative comparison of PSNR/SSIM using different numbers of previous stages' outputs for multi-stage feature fusion on the Calgary-Campinas brain dataset under $\\times 10$ acceleration." + ], + "table_footnote": [], + "table_body": "
Input# of previous iterations
015911
PSNR(dB)35.88+0.02+0.04+0.04+0.05
SSIM(%)94.43+0.02+0.05+0.06+0.07
", + "bbox": [ + 362, + 198, + 640, + 280 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "5.2 Multi-Stage Information Flow", + "text_level": 1, + "bbox": [ + 215, + 306, + 509, + 321 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "In Tab. 5 we investigate the effect of incorporating different numbers of previous iteration results into the current iteration for multi-stage information fusion. As the number increases, the reconstruction performance increases gradually. This implies the benefit of incorporating the information from previous iterations in DUM-based MRI reconstruction.", + "bbox": [ + 212, + 328, + 787, + 402 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "5.3 Adjacent Sensitivity Map Estimation", + "text_level": 1, + "bbox": [ + 215, + 424, + 568, + 440 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "In Tab. 6 we examine the impact of various strategies shown in Fig. 4 for sensitivity map estimation on both reconstruction performance and memory consumption. The ACS data input to the SME network is represented as a complex tensor with shape $(B,2a + 1,N,h,w)$ , where $2a + 1$ represents the adjacent number, $N$ denotes the number of coils of each slice and $B$ indicates the batch size. Here, the batch size is linearly related to the memory consumption. Strategy A reshapes the tensor to a batch size of $(2a + 1)BN$ and a channel size of 1 for coil-by-coil estimation. Strategy B concatenates the coils within each slice, reducing the batch size to $(2a + 1)B$ . Strategy C concatenates the same coils from adjacent slices, decreasing the batch size to $BN$ . Strategy D combines all coils within a single batch. Note that only strategies A and C can be used for data with variable numbers of coils. Our experiments demonstrate that Strategy C, which utilizes adjacent information, not only shows the best performance but also reduces memory consumption by a factor of $2a + 1$ compared to Strategy A. Intriguingly, strategy B is found to be the least effective. Ideally, since the coils within the same slice should be uncorrelated, this strategy should not negatively impact performance. However, the noise that is coupled across coils during acquisition may be amplified under this strategy, degrading the SME performance.", + "bbox": [ + 212, + 446, + 787, + 733 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "In Tab. 7, we further explore the impact of different adjacent types on Strategy C using the CMRxRecon cardiac datasets. These types include temporal information in the cine LAX dataset, contrast information in the T1 mapping dataset, and slice information in the cine SAX dataset. Strategy C shows consistent improvement over Strategy A for adjacent types of time and contrast. However, the result of the slice adjacency experiment deviates from our expectations, as previously observed in the brain dataset in Tab. 6. This discrepancy", + "bbox": [ + 212, + 734, + 787, + 842 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "Rethinking Deep Unrolled Model for Accelerated MRI Reconstruction", + "bbox": [ + 264, + 114, + 730, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/0e88a376a4ed29a1efaa010489ec87cfca8ea3544a178ee58631c0018bcefc3f.jpg", + "table_caption": [ + "Table 6: Comparison of different strategies for sensitivity map estimation on the Calgary-Campinas brain dataset under $\\times 10$ acceleration. The \"Batch\" number is linearly related to memory consumption. \"Flexibility\" indicates the ability to accept data with varying numbers of coils." + ], + "table_footnote": [], + "table_body": "
StrategyBatchChannelPSNR(dB)/SSIM(%)Flexibility
A|(2a+1)BN|135.60/94.22✓
B|(2a+1)BN|N35.30/93.98X
CBN|2a+1|35.88/94.43✓
DB||(2a+1)N|35.76/94.32X
", + "bbox": [ + 305, + 212, + 697, + 308 + ], + "page_idx": 13 + }, + { + "type": "table", + "img_path": "images/51a790d782a0ca122d5d417e6394f4515b0ac270624e1586f9a9aeeb521b045b.jpg", + "table_caption": [ + "Table 7: Quantitative comparison of PSNR(dB)/SSIM(%) of different sensitivity map estimation strategies using different adjacent information types on the CMRxRecon cardiac dataset under $\\times 10$ acceleration." + ], + "table_footnote": [], + "table_body": "
StrategyAdjacent type
TimeContrastSlice
A41.67/97.3542.62/98.1542.75/97.58
C41.94/97.5442.77/98.2242.53/97.54
", + "bbox": [ + 361, + 375, + 640, + 453 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "is attributed to the considerable slice gap present in the cardiac dataset, which is $4.0\\mathrm{mm}$ compared to no slice gap in the brain dataset. This suggests that adjacent reconstruction can be beneficial only when adjacent slices show high correlations; otherwise, it might even detract from performance by introducing confounding factors for the network.", + "bbox": [ + 212, + 481, + 787, + 556 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "6 Limitations", + "text_level": 1, + "bbox": [ + 214, + 580, + 364, + 595 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Like other deep unrolled models, our approach may exhibit limited generalization to unseen sampling masks, modalities, and anatomies, and it may be susceptible to adversarial attacks. Hallucinations and artifacts may be observed when the model is applied to highly undersampled $k$ -space data (e.g., at acceleration factors of 10), when trained on a limited amount of data, or when tested on out-of-distribution datasets.", + "bbox": [ + 212, + 612, + 787, + 702 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "7 Conclusion", + "text_level": 1, + "bbox": [ + 214, + 726, + 359, + 742 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In this paper, we improve the deep unrolled model for multi-coil MRI reconstruction through informative gradient-based learning and memory-efficient sensitivity map estimation. Our proposed method achieves state-of-the-art performance on several public MRI reconstruction benchmarks.", + "bbox": [ + 212, + 758, + 787, + 819 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "B. Xin et al.", + "bbox": [ + 271, + 114, + 357, + 127 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgments. This research has been partially funded by research grants to D. Metaxas through NSF: 2310966, 2235405, 2212301, 2003874, and FA9550-23-1-0417 and NIH 2R01HL127661.", + "bbox": [ + 215, + 146, + 787, + 191 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 217, + 321, + 233 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Adler, J., Öktem, O.: Learned primal-dual reconstruction. IEEE transactions on medical imaging 37(6), 1322-1332 (2018)", + "2. Aggarwal, H.K., Mani, M.P., Jacob, M.: Modl: Model-based deep learning architecture for inverse problems. IEEE transactions on medical imaging 38(2), 394-405 (2018)", + "3. Arnold, T.C., Freeman, C.W., Litt, B., Stein, J.M.: Low-field mri: clinical promise and challenges. Journal of Magnetic Resonance Imaging 57(1), 25-44 (2023)", + "4. Arvinte, M., Vishwanath, S., Tewfik, A.H., Tamir, J.I.: Deep j-sense: Accelerated mri reconstruction via unrolled alternating optimization. In: International conference on medical image computing and computer-assisted intervention. pp. 350-360. Springer (2021)", + "5. Beck, A., Teboulle, M.: A fast iterative shrinkage-thresholding algorithm for linear inverse problems. SIAM journal on imaging sciences 2(1), 183-202 (2009)", + "6. Boyd, S., Parikh, N., Chu, E., Peleato, B., Eckstein, J., et al.: Distributed optimization and statistical learning via the alternating direction method of multipliers. Foundations and Trends in Machine learning 3(1), 1-122 (2011)", + "7. Candès, E.J., et al.: Compressive sampling. In: Proceedings of the international congress of mathematicians. vol. 3, pp. 1433-1452. Madrid, Spain (2006)", + "8. Deshmane, A., Gulani, V., Griswold, M.A., Seiberlich, N.: Parallel mr imaging. Journal of Magnetic Resonance Imaging 36(1), 55-72 (2012)", + "9. Duchi, J., Hazan, E., Singer, Y.: Adaptive subgradient methods for online learning and stochastic optimization. Journal of machine learning research 12(7) (2011)", + "10. El-Rewaidy, H., Fahmy, A.S., Pashakhanloo, F., Cai, X., Kucukseymen, S., Csecs, I., Neisius, U., Haji-Valizadeh, H., Menze, B., Nezafat, R.: Multi-domain convolutional neural network (md-cnn) for radial reconstruction of dynamic cardiac mri. Magnetic Resonance in Medicine 85(3), 1195–1208 (2021)", + "11. Eo, T., Jun, Y., Kim, T., Jang, J., Lee, H.J., Hwang, D.: Kiki-net: cross-domain convolutional neural networks for reconstructing undersampled magnetic resonance images. Magnetic resonance in medicine 80(5), 2188-2201 (2018)", + "12. Fabian, Z., Tinaz, B., Soltanolkotabi, M.: Humus-net: Hybrid unrolled multi-scale network architecture for accelerated mri reconstruction. Advances in Neural Information Processing Systems 35, 25306-25319 (2022)", + "13. Gregor, K., LeCun, Y.: Learning fast approximations of sparse coding. In: Proceedings of the 27th international conference on international conference on machine learning. pp. 399-406 (2010)", + "14. Griswold, M.A., Jakob, P.M., Heidemann, R.M., Nittka, M., Jellus, V., Wang, J., Kiefer, B., Haase, A.: Generalized autocalibrating partially parallel acquisitions (grappa). Magnetic Resonance in Medicine: An Official Journal of the International Society for Magnetic Resonance in Medicine 47(6), 1202-1210 (2002)", + "15. Hammernik, K., Klatzer, T., Kobler, E., Recht, M.P., Sodickson, D.K., Pock, T., Knoll, F.: Learning a variational network for reconstruction of accelerated mri data. Magnetic resonance in medicine 79(6), 3055-3071 (2018)" + ], + "bbox": [ + 218, + 250, + 785, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "Rethinking Deep Unrolled Model for Accelerated MRI Reconstruction", + "bbox": [ + 264, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 785, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "16. Hosseini, S.A.H., Yaman, B., Moeller, S., Hong, M., Akçakaya, M.: Dense recurrent neural networks for accelerated mri: History-cognizant unrolling of optimization algorithms. IEEE Journal of Selected Topics in Signal Processing 14(6), 1280-1291 (2020)", + "17. Jun, Y., Shin, H., Eo, T., Hwang, D.: Joint deep model-based mr image and coil sensitivity reconstruction network (joint-icnet) for fast mri. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5270-5279 (2021)", + "18. Kingma, D.P., Ba, J.: Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980 (2014)", + "19. Lauterbur, P.C.: Image formation by induced local interactions: examples employing nuclear magnetic resonance. nature 242(5394), 190-191 (1973)", + "20. LeCun, Y., Bengio, Y., Hinton, G.: Deep learning. nature 521(7553), 436-444 (2015)", + "21. Liu, X., Pang, Y., Jin, R., Liu, Y., Wang, Z.: Dual-domain reconstruction network with v-net and k-net for fast mri. Magnetic Resonance in Medicine 88(6), 2694-2708 (2022)", + "22. Lustig, M., Donoho, D., Pauly, J.M.: Sparse mri: The application of compressed sensing for rapid mr imaging. Magnetic Resonance in Medicine: An Official Journal of the International Society for Magnetic Resonance in Medicine 58(6), 1182-1195 (2007)", + "23. Monga, V., Li, Y., Eldar, Y.C.: Algorithm unrolling: Interpretable, efficient deep learning for signal and image processing. IEEE Signal Processing Magazine 38(2), 18-44 (2021)", + "24. Munoz, C., Fotaki, A., Botnar, R.M., Prieto, C.: Latest advances in image acceleration: all dimensions are fair game. Journal of Magnetic Resonance Imaging 57(2), 387-402 (2023)", + "25. Nayak, K.S., Lim, Y., Campbell-Washburn, A.E., Steeden, J.: Real-time magnetic resonance imaging. Journal of Magnetic Resonance Imaging 55(1), 81-99 (2022)", + "26. Ongie, G., Jalal, A., Metzler, C.A., Baraniuk, R.G., Dimakis, A.G., Willett, R.: Deep learning techniques for inverse problems in imaging. IEEE Journal on Selected Areas in Information Theory 1(1), 39-56 (2020)", + "27. Ottesen, J.A., Caan, M.W., Groote, I.R., Bjornerud, A.: A densely interconnected network for deep learning accelerated mri. Magnetic Resonance Materials in Physics, Biology and Medicine 36(1), 65-77 (2023)", + "28. Pezzotti, N., Yousefi, S., Elmahdy, M.S., Van Gemert, J.H.F., Schuelke, C., Doneva, M., Nielsen, T., Kastryulin, S., Lieleveldt, B.P., Van Osch, M.J., et al.: An adaptive intelligence algorithm for undersampled knee mri reconstruction. IEEE Access 8, 204825-204838 (2020)", + "29. Polyak, B.T.: Some methods of speeding up the convergence of iteration methods. Ussr computational mathematics and mathematical physics 4(5), 1-17 (1964)", + "30. Pruessmann, K.P., Weiger, M., Scheidegger, M.B., Boesiger, P.: Sense: sensitivity encoding for fast mri. Magnetic Resonance in Medicine: An Official Journal of the International Society for Magnetic Resonance in Medicine 42(5), 952-962 (1999)", + "31. Putzky, P., Karkalousos, D., Teuwen, J., Miriakov, N., Bakker, B., Caan, M., Welling, M.: i-rim applied to the fastmri challenge. arXiv preprint arXiv:1910.08952 (2019)", + "32. Putzky, P., Welling, M.: Recurrent inference machines for solving inverse problems. arXiv preprint arXiv:1706.04008 (2017)", + "33. Putzky, P., Welling, M.: Invert to learn to invert. Advances in neural information processing systems 32 (2019)" + ], + "bbox": [ + 215, + 146, + 785, + 840 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "B. Xin et al.", + "bbox": [ + 271, + 114, + 354, + 126 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "34. Ramzi, Z., Chaithya, G., Starck, J.L., Ciuciu, P.: Nc-pdnet: A density-compensated unrolled network for 2d and 3d non-cartesian mri reconstruction. IEEE Transactions on Medical Imaging 41(7), 1625-1638 (2022)", + "35. Ronneberger, O., Fischer, P., Brox, T.: U-net: Convolutional networks for biomedical image segmentation. In: Medical Image Computing and Computer-Assisted Intervention-MICCAI 2015: 18th International Conference, Munich, Germany, October 5-9, 2015, Proceedings, Part III 18. pp. 234-241. Springer (2015)", + "36. Schlemper, J., Caballero, J., Hajnal, J.V., Price, A.N., Rueckert, D.: A deep cascade of convolutional neural networks for dynamic mr image reconstruction. IEEE transactions on Medical Imaging 37(2), 491-503 (2017)", + "37. Singh, D., Monga, A., de Moura, H.L., Zhang, X., Zibetti, M.V., Regatte, R.R.: Emerging trends in fast mri using deep-learning reconstruction on undersampled k-space data: a systematic review. Bioengineering 10(9), 1012 (2023)", + "38. Song, J., Chen, B., Zhang, J.: Memory-augmented deep unfolding network for compressive sensing. In: Proceedings of the 29th ACM international conference on multimedia. pp. 4249-4258 (2021)", + "39. Souza, R., Lucena, O., Garrafa, J., Gobbi, D., Saluzzi, M., Appenzeller, S., Rittner, L., Frayne, R., Lotufo, R.: An open, multi-vendor, multi-field-strength brain mr dataset and analysis of publicly available skull stripping methods agreement. NeuroImage 170, 482-494 (2018)", + "40. Sriram, A., Zbontar, J., Murrell, T., Defazio, A., Zitnick, C.L., Yakubova, N., Knoll, F., Johnson, P.: End-to-end variational networks for accelerated mri reconstruction. In: Medical Image Computing and Computer Assisted Intervention-MICCAI 2020: 23rd International Conference, Lima, Peru, October 4-8, 2020, Proceedings, Part II 23. pp. 64-73. Springer (2020)", + "41. Sun, J., Li, H., Xu, Z., et al.: Deep admm-net for compressive sensing mri. Advances in neural information processing systems 29 (2016)", + "42. Tieleman, T.: Lecture 6.5-rmsprop: Divide the gradient by a running average of its recent magnitude. COURSERA: Neural networks for machine learning 4(2), 26 (2012)", + "43. Uecker, M., Lai, P., Murphy, M.J., Virtue, P., Elad, M., Pauly, J.M., Vasanawala, S.S., Lustig, M.: Espirit—an eigenvalue approach to autocalibrating parallel migraine: where sense meets grappa. Magnetic resonance in medicine 71(3), 990-1001 (2014)", + "44. Wang, C., Lyu, J., Wang, S., Qin, C., Guo, K., Zhang, X., Yu, X., Li, Y., Wang, F., Jin, J., et al.: Cmrxrecon: an open cardiac mri dataset for the competition of accelerated image reconstruction. arXiv preprint arXiv:2309.10836 (2023)", + "45. Xin, B., Ye, M., Axel, L., Metaxas, D.N.: Fill the k-space and refine the image: Prompting for dynamic and multi-contrast mri reconstruction. In: International Workshop on Statistical Atlases and Computational Models of the Heart. pp. 261-273. Springer (2023)", + "46. Yiasemis, G., Sonke, J.J., Sánchez, C., Teuwen, J.: Recurrent variational network: a deep learning inverse problem solver applied to the task of accelerated mri reconstruction. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 732-741 (2022)", + "47. Ying, L., Sheng, J.: Joint image reconstruction and sensitivity estimation in sense (jsense). Magnetic Resonance in Medicine: An Official Journal of the International Society for Magnetic Resonance in Medicine 57(6), 1196-1202 (2007)", + "48. You, D., Xie, J., Zhang, J.: Ista-net++: Flexible deep unfolding network for compressive sensing. In: 2021 IEEE International Conference on Multimedia and Expo (ICME). pp. 1-6. IEEE (2021)" + ], + "bbox": [ + 215, + 146, + 787, + 840 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "Rethinking Deep Unrolled Model for Accelerated MRI Reconstruction", + "bbox": [ + 264, + 114, + 730, + 127 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 116, + 785, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "49. Zbontar, J., Knoll, F., Sriram, A., Murrell, T., Huang, Z., Muckley, M.J., Defazio, A., Stern, R., Johnson, P., Bruno, M., et al.: fastmri: An open dataset and benchmarks for accelerated mri. arXiv preprint arXiv:1811.08839 (2018)", + "50. Zeiler, M.D.: Adadelta: an adaptive learning rate method. arXiv preprint arXiv:1212.5701 (2012)", + "51. Zhang, J., Chen, B., Xiong, R., Zhang, Y.: Physics-inspired compressive sensing: Beyond deep unrolling. IEEE Signal Processing Magazine 40(1), 58-72 (2023)", + "52. Zhang, J., Ghanem, B.: Ista-net: Interpretable optimization-inspired deep network for image compressive sensing. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1828-1837 (2018)", + "53. Zhang, J., Zhang, Z., Xie, J., Zhang, Y.: High-throughput deep unfolding network for compressive sensing mri. IEEE Journal of Selected Topics in Signal Processing 16(4), 750-761 (2022)", + "54. Zhang, Z., Liu, Y., Liu, J., Wen, F., Zhu, C.: Amp-net: Denoising-based deep unfolding for compressive image sensing. IEEE Transactions on Image Processing 30, 1487-1500 (2020)", + "55. Zhao, R., Yaman, B., Zhang, Y., Stewart, R., Dixon, A., Knoll, F., Huang, Z., Lui, Y.W., Hansen, M.S., Lungren, M.P.: fastmri+, clinical pathology annotations for knee and brain fully sampled magnetic resonance imaging data. Scientific Data 9(1), 152 (2022)", + "56. Zhou, B., Zhou, S.K.: Dudornet: learning a dual-domain recurrent network for fast mri reconstruction with deep t1 prior. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 4273-4282 (2020)" + ], + "bbox": [ + 215, + 146, + 785, + 465 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "B. Xin et al.", + "bbox": [ + 271, + 114, + 354, + 126 + ], + "page_idx": 17 + } +] \ No newline at end of file diff --git a/2024/Rethinking Deep Unrolled Model for Accelerated MRI Reconstruction/27de6684-978b-4680-80b1-f68f3bc7d6a8_model.json b/2024/Rethinking Deep Unrolled Model for Accelerated MRI Reconstruction/27de6684-978b-4680-80b1-f68f3bc7d6a8_model.json new file mode 100644 index 0000000000000000000000000000000000000000..6f236797c215a96c5605f029f5ec85c2f7f4a61f --- /dev/null +++ b/2024/Rethinking Deep Unrolled Model for Accelerated MRI Reconstruction/27de6684-978b-4680-80b1-f68f3bc7d6a8_model.json @@ -0,0 +1,2480 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.222, + 0.141, + 0.784, + 0.187 + ], + "angle": 0, + "content": "Rethinking Deep Unrolled Model for Accelerated MRI Reconstruction" + }, + { + "type": "text", + "bbox": [ + 0.249, + 0.211, + 0.756, + 0.231 + ], + "angle": 0, + "content": "Bingyu Xin\\(^{1}\\), Meng Ye\\(^{1}\\), Leon Axel\\(^{2}\\), and Dimitris N. Metaxas\\(^{1}\\)" + }, + { + "type": "text", + "bbox": [ + 0.245, + 0.239, + 0.758, + 0.284 + ], + "angle": 0, + "content": "\\(^{1}\\) Rutgers University, New Brunswick, NJ 08901, USA \n\\(^{2}\\) New York University Langone Medical Center, New York, NY 10016, USA \nbx64@rutgers.edu, dnm@cs.rutgers.edu" + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.321, + 0.744, + 0.56 + ], + "angle": 0, + "content": "Abstract. Magnetic Resonance Imaging (MRI) is a widely used imaging modality for clinical diagnostics and the planning of surgical interventions. Accelerated MRI seeks to mitigate the inherent limitation of long scanning time by reducing the amount of raw \\(k\\)-space data required for image reconstruction. Recently, the deep unrolled model (DUM) has demonstrated significant effectiveness and improved interpretability for MRI reconstruction, by truncating and unrolling the conventional iterative reconstruction algorithms with deep neural networks. However, the potential of DUM for MRI reconstruction has not been fully exploited. In this paper, we first enhance the gradient and information flow within and across iteration stages of DUM, then we highlight the importance of using various adjacent information for accurate and memory-efficient sensitivity map estimation and improved multi-coil MRI reconstruction. Extensive experiments on several public MRI reconstruction datasets show that our method outperforms existing MRI reconstruction methods by a large margin. The code is available at https://github.com/hellopipu/PromptMR-plus." + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.571, + 0.741, + 0.6 + ], + "angle": 0, + "content": "Keywords: MRI Reconstruction \\(\\cdot\\) Deep Unrolled Model \\(\\cdot\\) Gradient-based Learning \\(\\cdot\\) Sensitivity Map Estimation" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.626, + 0.376, + 0.642 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.659, + 0.788, + 0.844 + ], + "angle": 0, + "content": "Magnetic Resonance Imaging (MRI) provides a radiation-free and highly versatile method for imaging the organs, tissues, and skeletal systems within the human body. Over the past 50 years since Paul Lauterbur produced the first MR image, MRI has evolved to become a cornerstone in clinical diagnostics [19]. However, the process of raw \\(k\\)-space data acquisition in MRI is typically time-consuming. Accelerated MRI techniques tackle this issue by minimizing the amount of raw data that needs to be collected for image reconstruction, thereby shortening the duration of the scan. Modern advances in MRI technology, including Parallel Imaging (PI) [8] and Compressed Sensing (CS) [7,22], have significantly enhanced the efficiency and quality of MRI scans, making it possible to acquire high-resolution images within a considerably reduced time frame. However, with increased patient throughput and the requirements of emerging" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.358, + 0.128 + ], + "angle": 0, + "content": "B. Xin et al." + }, + { + "type": "image", + "bbox": [ + 0.226, + 0.143, + 0.416, + 0.249 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.314, + 0.251, + 0.331, + 0.262 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.418, + 0.144, + 0.599, + 0.249 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.599, + 0.144, + 0.777, + 0.249 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.591, + 0.251, + 0.608, + 0.261 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.272, + 0.785, + 0.314 + ], + "angle": 0, + "content": "Fig. 1: Our proposed method is memory-efficient and shows better reconstruction performance. (a) GPU memory consumption on Calgary-Campinas brain dataset, (b) PSNR/SSIM results vs. the number of parameters on the fastMRI knee \\(\\times 8\\) testset." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.344, + 0.785, + 0.376 + ], + "angle": 0, + "content": "technologies, such as real-time MRI [25] and low-field MRI [3], the development of accelerated MRI reconstruction methods remains a hot research topic [24]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.376, + 0.787, + 0.677 + ], + "angle": 0, + "content": "Accelerated MRI reconstruction is a regularized inverse problem, which aims to reconstruct an unknown MR image from highly undersampled measurements in \\(k\\)-space. Conventional iterative MRI reconstruction methods minimize a cost function comprising two main components: a data-consistency term, which assesses the alignment between \\(k\\)-space predicted from the reconstructed image and the observed measurements, and a regularization term, which incorporates prior knowledge to encourage the emergence of desirable image attributes, e.g., sparsity. Over the past decade, deep learning [20] has emerged as a transformative approach to MRI reconstruction [26, 37]. Recently, the release of several large-scale public MRI reconstruction benchmarks, including the fastMRI dataset [49], Calgary-Campinas dataset [39], and CMRxRecon dataset [44], has significantly propelled the advancement of MRI reconstruction methods. Among deep learning-based MRI reconstruction approaches, the deep unrolled model (DUM) has garnered significant attention for its exceptional performance and ability to establish a concrete and systematic link between widely used iterative MRI reconstruction methods and deep neural networks [23]. However, as pointed out in [51], the current architecture design of DUM is inefficient, characterized by limited information transmission capacity and low flexibility and robustness. These shortcomings significantly restrict DUM's performance and its applicability to unseen data." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.678, + 0.787, + 0.723 + ], + "angle": 0, + "content": "In this work, we rethink and improve key components of the DUM-based MRI reconstruction approach. Our main contributions can be summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.227, + 0.734, + 0.785, + 0.794 + ], + "angle": 0, + "content": "- We improve the adaptive gradient algorithm and apply it to DUM, which can achieve self-adaptive dynamic learning rates adjusting for different spatial areas in an MR image. This approach provides a more flexible updating strategy in each iteration stage of DUM." + }, + { + "type": "text", + "bbox": [ + 0.227, + 0.795, + 0.785, + 0.84 + ], + "angle": 0, + "content": "- We incorporate the momentum technique used in gradient descent acceleration and propose a multi-stage and multi-level feature aggregation scheme to accelerate the iteration convergence of DUM." + }, + { + "type": "list", + "bbox": [ + 0.227, + 0.734, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.266, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Rethinking Deep Unrolled Model for Accelerated MRI Reconstruction" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.787, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.227, + 0.147, + 0.785, + 0.207 + ], + "angle": 0, + "content": "- We highlight the importance of adjacent information which can be used to improve multi-coil MRI reconstruction, especially for accurate and memory-efficient sensitivity map estimation. As shown in Fig. 1(a), our method reduces GPU consumption to \\(\\sim 55\\%\\) than the previous state-of-the-art." + }, + { + "type": "text", + "bbox": [ + 0.228, + 0.208, + 0.786, + 0.283 + ], + "angle": 0, + "content": "- We demonstrate the effectiveness of our proposed method on three public MRI reconstruction benchmarks of different anatomies: the fastMRI knee, the Calgary-Campinas brain and the CMRxRecon cardiac dataset. As shown in Fig. \\(\\boxed{1}\\) (b), our method achieves better PSNR and SSIM with fewer parameters than existing MRI Reconstruction methods." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.308, + 0.388, + 0.323 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.341, + 0.507, + 0.357 + ], + "angle": 0, + "content": "2.1 Deep Unrolled Model (DUM)" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.367, + 0.788, + 0.609 + ], + "angle": 0, + "content": "DUM maps a truncated optimization algorithm into a deep neural network, iteratively alternating between gradient descent steps and proximal mapping steps. Its superior performance and good interpretability make it a preferred method for a wide range of inverse problems, including compressive sensing, image restoration and image reconstruction [23, 51]. Learned ISTA (LISTA) was the first unrolling method proposed for fast approximation of sparse coding [13]. LISTA unrolls the traditional ISTA algorithm [5] and truncates it to a fixed number of iterations. By learning from a training dataset, this trainable version of ISTA shows significant computational benefits. The success of deep learning further promotes the idea of relating conventional iterative optimization methods to deep neural networks, including ADMM-Net [41], ISTA-Net [52] and AMP-Net [54]. By introducing a condition module to transmit input information to each stage of the unrolled model, ISTA-Net++ [48] provides more flexibility to handle multi-ratio and multi-scene images for compressive sensing. In MADUN [38], a memory-augmented DUM is proposed to enhance the short-term and long-term information transmission in traditional DUMs." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.633, + 0.515, + 0.646 + ], + "angle": 0, + "content": "2.2 DUM for MRI Reconstruction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.658, + 0.788, + 0.811 + ], + "angle": 0, + "content": "ADMM-Net [41] and DC-CNN [36] are the pioneering DUMs for MRI reconstruction, establishing themselves as widely used baseline DUM-based MRI reconstruction methods. These approaches, based on Convolutional Neural Networks (CNNs), unroll the alternating direction method of multipliers (ADMM) [6] and the alternating minimization (AM) algorithm, respectively. KIKI-Net [11] employs a novel approach by alternating between \\( k \\)-space completion blocks and image restoration blocks, demonstrating the advantages of cross-domain learning [10,21,56]. Adaptive-CS-Net finds it beneficial to inject information from neighboring slices into the model. This 2.5D reconstruction strategy has been adopted in subsequent works [12,45]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.811, + 0.787, + 0.84 + ], + "angle": 0, + "content": "The development of DUM has led to various innovative approaches beyond simple unrolling, categorized broadly into three main variants for enhanced" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.358, + 0.128 + ], + "angle": 0, + "content": "B. Xin et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.146, + 0.788, + 0.419 + ], + "angle": 0, + "content": "MRI reconstruction. (1) Firstly, inspired by Recurrent Neural Network (RNN), RIM [31, 32] uses a latent variable as temporal memory to track the progression of iterations. Recurrent-VarNet [46] extends this idea by designing a recurrent state initializer and recurrent units for the hidden state. In parallel, LPD-Net [1] and NCPD-Net [34] simplify this concept by utilizing a buffer design as additional learnable memory transmitting across iterations. (2) Secondly, inspired by efficient variants of proximal gradient descent (PGD), HC-PGD [16] proposes the history-cognizant unrolling method by concatenating all previous proximal operator output as input to each iteration. DIRCN [27] further proposes to add feature map interconnections between neighboring stages. These works suggest the importance of gradient flow and shared information between cascades for MRI reconstruction. (3) Moreover, Adaptive-CS-Net [28] incorporates soft MRI priors, including soft data consistency, image phase information and background priors, as additional input channels to each iteration. Similarly, HiTDUN [53] proposes the design of multi-channel input for every iteration. These methods underscore the advantages of introducing additional priors into the input at each iteration stage, aiming to dismantle the information bottleneck encountered within each iteration." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.453, + 0.566, + 0.469 + ], + "angle": 0, + "content": "2.3 Sensitivity Map Estimation in DUM" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.491, + 0.788, + 0.672 + ], + "angle": 0, + "content": "Numerous works on DUM have achieved promising results on MRI reconstruction [11, 21, 36, 41, 56]. However, most of these methods focus exclusively on synthesized single-coil \\(k\\)-space data, overlooking the more prevalent multi-coil imaging in real clinical MRI scans. This oversight neglects a more realistic scenario, given the widespread application of parallel imaging. Parallel imaging (PI) can speed up MRI, with the idea of using multiple receiving coils to reduce the number of phase encoding lines needed in acquiring raw \\(k\\)-space data. The sensitivity map of a given coil is its local field profile, describing through reciprocity where that coil can efficiently pick up MR signals from the imaged region. The reconstruction of multiple coil \\(k\\)-space data needs accurate estimation of the sensitivity maps, which are either directly measured, as in SENSE [30], or indirectly accounted using a center region of \\(k\\)-space data, as in GRAPPA [14]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.788, + 0.844 + ], + "angle": 0, + "content": "Robust and accurate sensitivity map estimation is crucial for multi-coil MRI reconstruction. The i-RIM directly reconstructs coil images without explicitly estimating the sensitivity maps. Adaptive-CS-Net [28] uses the normalized low-pass filtered coil images as sensitivity maps. VarNet [15] and MoDL [2] handle multi-coil data with pre-estimated sensitivity maps [43]. E2E-VarNet [40] advances the VarNet by learning sensitivity maps from the Auto-Calibration Signal (ACS) data of the \\(k\\)-space in an end-to-end manner. Due to its superior performance and end-to-end approach, many subsequent studies are based on E2E-VarNet [12,45,46]. When the ACS data is insufficient for accurate sensitivity map estimation, [4,17] suggest iterative sensitivity map estimation and image reconstruction, which can be considered as a deep version of J-SENSE [47]." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.266, + 0.114, + 0.733, + 0.129 + ], + "angle": 0, + "content": "Rethinking Deep Unrolled Model for Accelerated MRI Reconstruction" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "image", + "bbox": [ + 0.243, + 0.142, + 0.486, + 0.229 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.354, + 0.231, + 0.374, + 0.242 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.552, + 0.142, + 0.728, + 0.229 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.63, + 0.23, + 0.651, + 0.242 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.253, + 0.788, + 0.299 + ], + "angle": 0, + "content": "Fig. 2: Overview of our proposed (a) self-adaptive and (b) momentum-accelerated gradient algorithm. \\(\\hat{x}_{t+1}\\) and \\(\\hat{F}_t^m\\) are the output of CNN and Momentum layer in Eq. (7) and Eq. (10), respectively." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.323, + 0.427, + 0.34 + ], + "angle": 0, + "content": "3 Proposed Method" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.354, + 0.788, + 0.462 + ], + "angle": 0, + "content": "Drawing inspiration from adaptive gradient and momentum-accelerated algorithms in gradient descent optimization, we dynamically incorporate prior knowledge in the data flow, to enable more informative and faster convergence for gradient-based MRI reconstruction learning, as shown in Fig. 2. Subsequently, we enhance the concept of adjacent reconstruction by focusing on more robust and memory-efficient multi-coil sensitivity map estimation, as shown in Fig. 4. Please refer to Supplementary Section B for architecture details." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.482, + 0.44, + 0.496 + ], + "angle": 0, + "content": "3.1 Problem Formulation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.507, + 0.787, + 0.539 + ], + "angle": 0, + "content": "We can estimate the complex MR image \\(\\pmb{x}\\) from its undersampled measurements \\(\\pmb{y}\\) in \\(k\\)-space by solving an optimization problem [22]," + }, + { + "type": "equation", + "bbox": [ + 0.362, + 0.547, + 0.786, + 0.577 + ], + "angle": 0, + "content": "\\[\n\\min _ {\\boldsymbol {x}} \\boldsymbol {f} (\\boldsymbol {x}) = \\min _ {\\boldsymbol {x}} \\frac {1}{2} \\left\\| \\boldsymbol {A} \\boldsymbol {x} - \\boldsymbol {y} \\right\\| _ {2} ^ {2} + \\boldsymbol {R} (\\boldsymbol {x}), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.585, + 0.788, + 0.678 + ], + "angle": 0, + "content": "where \\(\\mathbf{A} = M\\mathbf{F}\\mathbf{S}\\) is the forward process of acquiring the \\(k\\)-space measurements, which is a combination of coil sensitivity maps \\(\\mathbf{S}\\), the Fourier transform \\(\\mathcal{F}\\) and an undersampling mask \\(\\mathbf{M}\\). The first term \\(\\frac{1}{2} ||\\mathbf{A}\\mathbf{x} - \\mathbf{y}||_2^2\\) represents the data consistency term, which ensures data fidelity, while the second term \\(\\mathbf{R}(\\mathbf{x})\\) serves as the regularization term, aimed at promoting image domain priors, such as sparsity [7]. We can solve Eq. (1) iteratively via the gradient descent method," + }, + { + "type": "equation", + "bbox": [ + 0.38, + 0.703, + 0.786, + 0.719 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {x} _ {t + 1} = \\boldsymbol {x} _ {t} - \\eta_ {t} \\boldsymbol {g} _ {t}, \\tag {2}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.397, + 0.721, + 0.786, + 0.74 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {g} _ {t} = \\boldsymbol {A} ^ {H} (\\boldsymbol {A} \\boldsymbol {x} _ {t} - \\boldsymbol {y}) + \\nabla \\boldsymbol {R} (\\boldsymbol {x} _ {t}), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.749, + 0.788, + 0.841 + ], + "angle": 0, + "content": "where \\(\\nabla R(\\pmb{x}_t)\\) is the gradient or the proximal mapping of \\(\\pmb{R}\\), \\(\\eta^t\\) is the scalar learning rate at iteration \\(t\\), and \\(t \\in \\{0,1,\\dots,T-1\\}\\), with \\(T\\) being the total number of truncated iterations. In the context of the deep unrolled models, it is customary to substitute \\(\\nabla R(\\pmb{x}_t)\\) with a neural network whose parameters are learned from large-scale training datasets. This allows for the learning of more sophisticated image domain priors, compared to those that are hand-crafted." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.358, + 0.128 + ], + "angle": 0, + "content": "B. Xin et al." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.147, + 0.544, + 0.163 + ], + "angle": 0, + "content": "3.2 Self-Adaptive Gradient Algorithm" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.174, + 0.789, + 0.356 + ], + "angle": 0, + "content": "During the iterative MRI reconstruction process, some regions in an image are straightforward to reconstruct, such as the background, which contains minimal information. However, other areas are more prone to aliasing artifacts resulting from downsampling the \\(k\\)-space, which are challenging to eliminate. This difficulty often arises because the aliasing artifacts may closely mimic the appearance of anatomical structures within these regions, necessitating a more meticulous gradient-based correction at each iteration. Inspired by the adaptive gradient algorithm [9,18,42,50], instead of applying a uniform learning rate across the entire image space at each iteration, it is reasonable to utilize self-adaptive learning rates for individual pixels to achieve a more refined reconstruction result. Our generalization of the standard gradient descent in Eq. (2) employs the following update for each iteration:" + }, + { + "type": "equation", + "bbox": [ + 0.398, + 0.368, + 0.786, + 0.386 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {x} _ {t + 1} = \\boldsymbol {x} _ {t} - \\eta_ {t} \\boldsymbol {g} _ {t} - \\boldsymbol {E} _ {t} \\odot \\boldsymbol {g} _ {t}, \\tag {4}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.412, + 0.388, + 0.786, + 0.427 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {E} _ {t} = \\frac {\\eta_ {t}}{\\sqrt {\\sum_ {\\tau = 0} ^ {t} \\boldsymbol {g} _ {\\tau} ^ {2}} + \\epsilon} - \\eta_ {t}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.437, + 0.789, + 0.528 + ], + "angle": 0, + "content": "where \\( \\pmb{E}_{t} \\) is the residual pixel-wise learning rate at the \\( t \\)-th iteration, and the denominator in Eq. (5) computes the root of the sum of squares of all previous gradients of individual pixels. \\( \\epsilon \\) is a small number to improve the numerical stability. The operation \\( \\odot \\) denotes element-wise multiplication. We note that, the accumulation of historical pixel-wise gradients in Eq. (5) can help improve iteration convergence by" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.54, + 0.787, + 0.569 + ], + "angle": 0, + "content": "(1) scaling down the updating steps for regions with historically large gradients, which may be prone to overshooting;" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.57, + 0.787, + 0.601 + ], + "angle": 0, + "content": "(2) scaling up the updating steps for regions with historically small gradients, which may benefit from a more aggressive update." + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.54, + 0.787, + 0.601 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.612, + 0.789, + 0.75 + ], + "angle": 0, + "content": "To unroll our self-adaptive gradient algorithm with a DUM, we deviate from the common practice of substituting \\(\\nabla R(\\boldsymbol{x}_t)\\) with a neural network in Eq. (4). There are compelling reasons for this. First, learning the \\(\\nabla R(\\boldsymbol{x}_t)\\) term in \\(\\boldsymbol{g}_t\\) (see Eq. (3)) can lead to numerical instability in \\(\\boldsymbol{E}_t\\) during the early stage of training. Second, the adaptive gradient strategy in Eq. (5), which is the basic variant utilized in AdaGrad (9), may not be the most suitable choice for a DUM in the context of MRI reconstruction. To leverage the robust nonlinear representation capabilities of neural networks, we improve upon the adaptive gradient algorithm by reformulating Eq. (4) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.28, + 0.759, + 0.786, + 0.799 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\boldsymbol {x} _ {t + 1} = \\boldsymbol {x} _ {t} - \\eta_ {t} \\boldsymbol {A} ^ {H} (\\boldsymbol {A} \\boldsymbol {x} _ {t} - \\boldsymbol {y}) - \\left(\\eta_ {t} \\nabla \\boldsymbol {R} (\\boldsymbol {x} _ {t}) + \\boldsymbol {E} _ {t} \\odot \\boldsymbol {g} _ {t}\\right) (6) \\\\ = \\boldsymbol {x} _ {t} - \\eta_ {t} \\boldsymbol {A} ^ {H} (\\boldsymbol {A} \\boldsymbol {x} _ {t} - \\boldsymbol {y}) + \\operatorname {C N N} _ {t} \\left(\\boldsymbol {x} _ {t}, \\boldsymbol {A} ^ {H} \\boldsymbol {A} \\boldsymbol {x} _ {t}, \\boldsymbol {A} ^ {H} \\boldsymbol {y}, \\boldsymbol {s} _ {t}\\right). (7) \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.809, + 0.789, + 0.843 + ], + "angle": 0, + "content": "For the last term in Eq. (6), we reparametrize it with a CNN, as indicated in Eq. (7). This neural network takes as input the fundamental components from" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.266, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "Rethinking Deep Unrolled Model for Accelerated MRI Reconstruction" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.146, + 0.784, + 0.222 + ], + "angle": 0, + "content": "which the term is constructed, namely, \\( \\boldsymbol{x}_t, \\boldsymbol{A}^H \\boldsymbol{A} \\boldsymbol{x}_t, \\boldsymbol{A}^H \\boldsymbol{y} \\), and a learnable auxiliary memory \\( s_t \\), which serves as a hidden state to track the historical gradient information implicitly. Note that \\( \\boldsymbol{A}^H \\boldsymbol{y} \\) yields \\( \\boldsymbol{x}_0 \\), which represents the image transformed from zero-filled \\( k \\)-space. For further discussion, please refer to Supplementary Section E." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.248, + 0.63, + 0.263 + ], + "angle": 0, + "content": "3.3 Momentum-Accelerated Gradient Algorithm" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.275, + 0.785, + 0.321 + ], + "angle": 0, + "content": "The momentum method [29] accelerates gradient descent in Eq. (2) by accumulating a velocity vector in directions of persistent reduction in the optimization objective across iterations:" + }, + { + "type": "equation", + "bbox": [ + 0.405, + 0.337, + 0.785, + 0.351 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {x} _ {t + 1} = \\boldsymbol {x} _ {t} - \\eta_ {t} \\boldsymbol {m} _ {t}, \\tag {8}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.416, + 0.355, + 0.785, + 0.37 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {m} _ {t} = \\mu \\boldsymbol {m} _ {t - 1} + (1 - \\mu) \\boldsymbol {g} _ {t}, \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.382, + 0.788, + 0.578 + ], + "angle": 0, + "content": "where \\(\\mu \\in [0,1]\\) is the momentum coefficient. The momentum \\(m_{t}\\) is known as the exponential moving average of historical gradients. It can benefit the optimization of the current iteration, thereby considerably accelerating convergence. In Sec. 3.2 we introduced the hidden state \\(s_t\\), which is used to implicitly track the pixel-wise historical gradient information. Here, we propose explicit multi-stage and multi-level feature aggregation, to mimic the idea of momentum and increase the information flow across the iterations. For a typical multi-level encoder-decoder denoiser, such as Unet, in the \\(t\\)-th iteration, we insert a Momentum layer before the \\(m\\)-th level decoder layer. This layer concatenates features at the current stage and features from all previous iteration stages at the same level, followed by a \\(1\\times 1\\) convolution layer to reduce the dimensionality, and a Channel Attention Block (CAB) to adaptively fuse the multi-stage information, as described below:" + }, + { + "type": "equation", + "bbox": [ + 0.255, + 0.592, + 0.785, + 0.609 + ], + "angle": 0, + "content": "\\[\n\\operatorname {M o m e n t u m} \\left(\\boldsymbol {F} _ {0} ^ {m}, \\dots , \\boldsymbol {F} _ {t} ^ {m}\\right) = \\operatorname {C A B} \\left(\\operatorname {C o n v} _ {1 \\times 1} \\left(\\operatorname {C o n c a t} \\left(\\boldsymbol {F} _ {0} ^ {m}, \\dots , \\boldsymbol {F} _ {t} ^ {m}\\right)\\right)\\right), \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.621, + 0.784, + 0.636 + ], + "angle": 0, + "content": "where \\( \\pmb{F}_t^m \\) is the input feature of the \\( m \\)-th level decoder layer at the \\( t \\)-th iteration." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.661, + 0.467, + 0.677 + ], + "angle": 0, + "content": "3.4 Adjacent Reconstruction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Incorporating information from adjacent slices in multi-coil MRI reconstruction has been substantiated in prior work [12,28,45]. However, a significant limitation of adjacent slice reconstruction lies in its substantial memory consumption during the estimation of multi-coil sensitivity maps. This issue stems from the design of the sensitivity map estimation (SME) network in earlier work [40]. The network adopts a coil-by-coil estimation approach to accommodate a variable number of coils, increasing flexibility but leading to higher memory demands in the adjacent reconstruction setting (see Fig. 4(a)). To be more specific, for \\(k\\)-space data with \\(N\\) coils and \\(2a + 1\\) adjacent slices, the total number of forward passes required for the SME process is calculated as \\((2a + 1)N\\). Typically, the" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.358, + 0.128 + ], + "angle": 0, + "content": "B. Xin et al." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.144, + 0.391, + 0.224 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.279, + 0.226, + 0.331, + 0.238 + ], + "angle": 0, + "content": "(a) time" + }, + { + "type": "image", + "bbox": [ + 0.416, + 0.144, + 0.588, + 0.224 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.466, + 0.226, + 0.539, + 0.237 + ], + "angle": 0, + "content": "(b) contrast" + }, + { + "type": "image", + "bbox": [ + 0.614, + 0.144, + 0.784, + 0.224 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.676, + 0.226, + 0.725, + 0.237 + ], + "angle": 0, + "content": "(c) slice" + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.249, + 0.788, + 0.292 + ], + "angle": 0, + "content": "Fig. 3: The sensitivity maps exhibit variations across (a) time, (b) contrast and (c) slice dimension. First row: original images, second row: magnitude of differences between the sensitivity map of the central image and each of those of its adjacent images." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.305, + 0.492, + 0.378 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.51, + 0.304, + 0.785, + 0.378 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.386, + 0.493, + 0.461 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.511, + 0.386, + 0.785, + 0.461 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.472, + 0.789, + 0.556 + ], + "angle": 0, + "content": "Fig. 4: Concept illustration of different strategies for multi-coil sensitivity map estimation (SME). The main difference between the four strategies lies in the Auto-Calibration-Signal (ACS) information sharing between coils and adjacent MR image slices. Each single coil sensitivity map gets ACS information from (a) a single coil and a single image slice; (b) multiple coils but a single image slice; (c) a single coil but multiple adjacent image slices; (d) multiple coils and multiple adjacent image slices." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.586, + 0.789, + 0.631 + ], + "angle": 0, + "content": "value of \\( N \\) ranges from 8 to 34 and \\( a \\) is set to 2. We propose to reduce the number of forward passes, by utilizing the correlations between adjacent sensitivity maps." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.632, + 0.789, + 0.816 + ], + "angle": 0, + "content": "A sensitivity map is subject to variation, based on the position of the imaging subject and changes over time due to the subject motion. Additionally, in multicontest contrast MRI acquisition, variations in acquisition parameters, such as inversion time (TI) and repetition time (TR), can indirectly influence the sensitivity map estimation through the acquired \\(k\\)-space data, as shown in Fig. ③. Consequently, the estimation of adjacent sensitivity maps is a spatially, temporally, and contrastively correlated process. We therefore propose to estimate sensitivity maps by utilizing information from spatially, temporally, and contrastively adjacent slices. As shown in Fig. ④(c), we simultaneously estimate a single coil sensitivity map at all adjacent \\(2a + 1\\) slices using information across these slices. In this way, we can reduce the number of forward passes from \\((2a + 1)N\\) to \\(N\\). Our estimation of the \\(n\\)-th sensitivity map set \\(S_{adj}^{n}\\) can be described as:" + }, + { + "type": "equation", + "bbox": [ + 0.398, + 0.825, + 0.786, + 0.843 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {S} _ {a d j} ^ {n} = \\mathbf {S M E} (\\mathbf {A C S} \\left(\\boldsymbol {x} _ {0, a d j} ^ {n}\\right)), \\tag {11}\n\\]" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.266, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Rethinking Deep Unrolled Model for Accelerated MRI Reconstruction" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.787, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.211 + ], + "angle": 0, + "content": "where \\( S_{adj}^{n} = [S_{a - c}^{n},\\dots,S_{c}^{n},\\dots,S_{a + c}^{n}] \\) is an adjacent sensitivity map set which is the concatenation of central slice sensitivity map \\( S_{c}^{n} \\) with its 2a-adjacent slice sensitivity maps of the \\( n \\)-th coil. \\( \\mathbf{ACS}(\\pmb{x}_{0,adj}^{n}) \\) is the \\( n \\)-th coil Auto-Calibration Signal (ACS) within all adjacent \\( 2a + 1 \\) slices." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.229, + 0.496, + 0.246 + ], + "angle": 0, + "content": "4 Experiments and Results" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.256, + 0.788, + 0.364 + ], + "angle": 0, + "content": "In this section, we present the experimental results of our proposed approach. We showcase the reconstruction performance of our model across multiple publicly available datasets, including Calgary-Campinas brain [39], CMRxRecon cardiac [44], and fastMRI multi-coil knee [49] datasets. These datasets encompass a diverse range of anatomies, contrasts, views, slices and motion states. For details about the datasets, please refer to Supplementary Section A. For information on the experimental setup for each dataset, see Supplementary Section C." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.381, + 0.533, + 0.396 + ], + "angle": 0, + "content": "4.1 Calgary-Campinas Brain Results" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.402, + 0.788, + 0.522 + ], + "angle": 0, + "content": "Tab. 1 presents the quantitative results for \\(\\times 5\\) and \\(\\times 10\\) acceleration factors on the Calgary-Campinas brain test set. Our method is compared with two state-of-the-art (SOTA) methods: Recurrent-VarNet [46] and PromptMR [45]. Our proposed method significantly outperforms these benchmarks, demonstrating its outstanding performance. Qualitative results under \\(\\times 10\\) acceleration are shown in Fig. 5. Interestingly, while PromptMR achieves significantly higher PSNR and SSIM values compared to Recurrent-VarNet, it exhibits less accurate reconstruction in the area highlighted by the red box." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.523, + 0.788, + 0.659 + ], + "angle": 0, + "content": "The highly undersampled \\(k\\)-space data makes the ill-posed reconstruction challenging at aliasing artifacts removal. While leveraging large-scale training data to learn the image priors can improve the reconstruction accuracy, we note that Recurrent-VarNet learns a weaker prior, resulting in reconstructions that appear blurry. Conversely, PromptMR adopts a strong but overly biased prior, leading to sharper reconstructions but introducing incorrect structures. Our proposed method, in contrast, successfully learns the appropriate prior, achieving a reconstruction that closely mirrors the ground truth in terms of structural accuracy." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.677, + 0.507, + 0.691 + ], + "angle": 0, + "content": "4.2 CMRxRecon Cardiac Results" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.698, + 0.788, + 0.82 + ], + "angle": 0, + "content": "We compare our method with PromptMR, the winning method of the MIC-CAI2023 CMRxRecon challenge, for \\(\\times 4\\), \\(\\times 8\\) and \\(\\times 10\\) acceleration factors on the CMRxRecon cardiac test set. In Tab. ②, our method exhibits superior performance to PromptMR across different views and contrasts. Qualitative results under \\(\\times 10\\) acceleration are shown in Fig. ⑥. Our method reconstructs fine details of the interventricular septum, highlighted in the red box, with significantly higher clarity than the PromptMR method, which is severely compromised by aliasing artifacts." + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.825, + 0.605, + 0.84 + ], + "angle": 0, + "content": "1 It is usually a part of the acquired central \\(k\\)-space data." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.356, + 0.127 + ], + "angle": 0, + "content": "B. Xin et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.145, + 0.788, + 0.201 + ], + "angle": 0, + "content": "Table 1: Quantitative comparison of PSNR/SSIM (mean ± std) of different MRI reconstruction methods on the Calgary-Campinas brain dataset under \\( \\times 5 \\) and \\( \\times 10 \\) acceleration. The best and second best performance are in red and blue colors, respectively." + }, + { + "type": "table", + "bbox": [ + 0.306, + 0.213, + 0.699, + 0.31 + ], + "angle": 0, + "content": "
MethodAcc = 5×Acc = 10×
PSNR(dB)SSIM(%)PSNR(dB)SSIM(%)
Zero-filled25.27±1.3173.75±3.9324.35±1.2169.04±3.67
Recurrent-Varnet36.27±1.7294.37±1.3333.27±2.1291.47±2.23
PromptMR36.83±1.6894.85±1.2634.10±2.1192.69±1.92
Ours37.24±1.5795.15±1.2134.74±1.9793.38±1.65
" + }, + { + "type": "image", + "bbox": [ + 0.302, + 0.321, + 0.702, + 0.478 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.488, + 0.788, + 0.531 + ], + "angle": 0, + "content": "Fig. 5: The reconstruction results and SSIM error maps of various methods for the Calgary-Campinas brain imaging under \\(\\times 10\\) acceleration factor. The red boxes highlight the differences in the recovery of white and gray matter structures." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.557, + 0.453, + 0.572 + ], + "angle": 0, + "content": "4.3 FastMRI Knee Results" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.58, + 0.789, + 0.761 + ], + "angle": 0, + "content": "We evaluated the PSNR and SSIM at acceleration factors of \\(\\times 4\\) and \\(\\times 8\\) on the fastMRI multi-coil knee test set. Our method is compared with several methods, including SOTA methods HUMUS-Net-L [12] and PromptMR. In the table, 'PD' and 'PDFS' denote proton density-weighted images without and with fat suppression, respectively. Tab. 3 demonstrates that our approach achieves the best performance across all other competitive methods on this large-scale dataset. Qualitative results under \\(\\times 8\\) acceleration are shown in Fig. 7. Our approach achieves a more detailed reconstruction of the meniscus region as shown in the red box, closely approximating the ground truth. In this region, meniscus tears were identified by radiologists from the fastMRI+ dataset [55]. The qualitative comparison suggests that our approach is more robust in reconstructing abnormal knee regions for highly accelerated acquisition." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.781, + 0.41, + 0.799 + ], + "angle": 0, + "content": "5 Abalation study" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.84 + ], + "angle": 0, + "content": "In this section, we justify our design choices through ablation studies on the brain and cardiac datasets." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.266, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Rethinking Deep Unrolled Model for Accelerated MRI Reconstruction" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "table_caption", + "bbox": [ + 0.215, + 0.145, + 0.788, + 0.201 + ], + "angle": 0, + "content": "Table 2: Quantitative comparison of PSNR/SSIM (mean ± std) of different MRI reconstruction methods on the CMRxRecon cardiac dataset under \\(\\times 4\\), \\(\\times 8\\) and \\(\\times 10\\) acceleration. The best and second best performance are in red and blue colors, respectively." + }, + { + "type": "table", + "bbox": [ + 0.221, + 0.213, + 0.784, + 0.367 + ], + "angle": 0, + "content": "
AccMethodCineMapping
SAXLAXT1wT2w
PSNR(dB)SSIM(%)PSNR(dB)SSIM(%)PSNR(dB)SSIM(%)PSNR(dB)SSIM(%)
4×Zero-filled PromptMR Ours26.20±1.5472.99±5.1425.11±1.6269.87±4.7324.23±1.1667.27±3.8424.97±1.1077.16±2.91
45.78±1.8998.78±0.4345.58±1.7298.72±0.3746.49±2.2398.93±0.5742.02±1.9398.05±0.72
46.26±1.7698.87±0.3446.13±1.7098.83±0.3447.45±2.1399.09±0.4242.68±1.8498.25±0.59
8×Zero-filled PromptMR Ours25.18±1.6770.21±5.8724.34±1.6667.67±5.8023.41±1.2564.12±4.4324.21±1.1274.84±3.07
40.65±1.5296.97±0.6839.64±1.8796.36±1.0440.94±1.6997.28±0.8538.24±1.8296.53±1.09
41.44±1.4897.32±0.5840.49±1.8396.76±0.9442.08±1.6397.70±0.6739.23±1.7196.98±0.89
10×Zero-filled PromptMR Ours24.83±1.6769.35±5.9224.13±1.6766.89±5.5823.16±1.2764.14±4.4724.25±1.1476.08±2.91
39.18±1.5096.15±0.8038.28±1.6295.60±1.0738.99±1.5896.61±0.9937.21±1.7696.22±1.16
39.99±1.4996.58±0.7239.13±1.6696.05±0.9940.37±1.5897.19±0.7938.22±1.7096.70±0.96
" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.379, + 0.787, + 0.542 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.552, + 0.788, + 0.594 + ], + "angle": 0, + "content": "Fig. 6: The reconstruction results and SSIM error maps of various methods for the CM-RxRecon cardiac T1-weighted imaging under \\(\\times 10\\) acceleration factor. The red boxes highlight the differences in the recovery of the interventricular septum." + }, + { + "type": "table_caption", + "bbox": [ + 0.215, + 0.606, + 0.788, + 0.649 + ], + "angle": 0, + "content": "Table 3: Quantitative comparison of PSNR/SSIM (mean ± std) of different MRI reconstruction methods on the fastMRI knee dataset under \\(\\times 4\\) and \\(\\times 8\\) acceleration. The best and second best performance are in red and blue colors, respectively." + }, + { + "type": "table", + "bbox": [ + 0.221, + 0.661, + 0.784, + 0.801 + ], + "angle": 0, + "content": "
MethodAcc = 4×Acc = 8×
PDPDSPDPDS
|PSNR(dB)|SSIM(%)|PSNR(dB)|SSIM(%)|PSNR(dB)|SSIM(%)|PSNR(dB)|SSIM(%)
Zero-filled30.86±1.7380.65±3.7631.00±3.3378.48±6.1627.70±1.8474.12±4.9427.97±2.0268.78±5.58
Unet 3537.27±1.7692.11±2.7237.07±2.4788.03±5.0435.41±2.1490.36±3.1334.84±1.5983.71±5.04
i-RIM 3339.61±2.3394.16±2.8738.23±3.1389.27±4.9637.91±2.4793.22±2.8535.84±1.7185.27±4.93
E2E-VarNet 4040.09±2.3294.56±2.6438.43±3.2489.50±4.9538.72±2.5093.93±2.6836.16±1.7685.70±4.94
HUMUS-Net 12----38.72±2.5094.07±2.6835.98±1.7585.77±4.93
HUMUS-Net-L 12----39.03±2.4794.18±2.6136.19±1.7985.84±4.92
PromptMR 4540.58±2.5994.87±2.7538.59±3.3389.75±4.9439.48±2.5494.51±2.5436.42±1.7986.10±4.93
Ours40.85±2.7495.00±2.8438.71±3.3789.85±4.9240.09±2.5894.90±2.4536.49±1.8886.29±4.95
" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.358, + 0.127 + ], + "angle": 0, + "content": "B. Xin et al." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.145, + 0.787, + 0.263 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.27, + 0.788, + 0.34 + ], + "angle": 0, + "content": "Fig. 7: The reconstruction results and SSIM error maps of various methods for the fastMRI multi-coil knee proton density (PD) imaging under \\(\\times 8\\) acceleration. The red boxes highlight the differences in the recovery of the meniscus region. The red box in the reference image was annotated by radiologists from the fastMRI+ dataset [55] to indicate the knee abnormality of meniscus tears." + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.36, + 0.788, + 0.389 + ], + "angle": 0, + "content": "Table 4: Abalation of different input combinations of CNN on the Calgary-Campinas brain dataset under \\(\\times 10\\) acceleration." + }, + { + "type": "table", + "bbox": [ + 0.306, + 0.4, + 0.699, + 0.528 + ], + "angle": 0, + "content": "
InputBaseline------Ours
xt✓✓✓✓✓✓✓✓
AHx✓✓✓✓✓✓✓✓
AHy✓✓✓✓✓✓✓✓
st✓✓✓✓✓✓✓✓
PSNR(dB)35.76+0.00+0.01+0.09+0.08+0.31+0.36+0.36
SSIM(%)94.33+0.00-0.01+0.06+0.05+0.25+0.27+0.28
" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.561, + 0.473, + 0.577 + ], + "angle": 0, + "content": "5.1 Input Variations to CNN" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.598, + 0.789, + 0.844 + ], + "angle": 0, + "content": "We conducted an ablation study for different input combinations to CNN, to validate the effectiveness of our proposed method. Results are shown in Tab. 4. The baseline network in each iteration only accepts the output from the last iteration \\( \\pmb{x}_t \\) as input, which is the setting for most unrolled models, such as E2E-VartNet and PromptMR. Our proposed method takes additional input, \\( A^H A\\pmb{x}_t \\), \\( A^H\\pmb{y} \\) and \\( s_t \\), which outperforms the baseline by a large margin. More specifically, adding \\( A^H A\\pmb{x}_t \\) or \\( A^H\\pmb{y} \\) solely to the input has no impact on the performance. Adding \\( s_t \\) can improve the performance, as it serves as additional memory to track the progression of the reconstruction process. Adding \\( A^H A\\pmb{x}_t \\) together with \\( A^H\\pmb{y} \\) greatly improves the gains. We assume the reason is that the residual connection of two terms into the network can provide soft data consistency information to the model. Adding \\( A^H A\\pmb{x}_t \\) together with \\( s_t \\) performs the closest to our final design. It seems the missing of \\( A^H\\pmb{y} \\) has little impact, as it is a constant value and its information may be compactly represented in the \\( s_t \\). This assumption is implied by the evidence that the result of adding \\( A^H\\pmb{y} \\) together with \\( s_t \\) into the network does not perform better than adding \\( s_t \\) alone." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.266, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Rethinking Deep Unrolled Model for Accelerated MRI Reconstruction" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.145, + 0.788, + 0.187 + ], + "angle": 0, + "content": "Table 5: Quantitative comparison of PSNR/SSIM using different numbers of previous stages' outputs for multi-stage feature fusion on the Calgary-Campinas brain dataset under \\(\\times 10\\) acceleration." + }, + { + "type": "table", + "bbox": [ + 0.364, + 0.199, + 0.642, + 0.281 + ], + "angle": 0, + "content": "
Input# of previous iterations
015911
PSNR(dB)35.88+0.02+0.04+0.04+0.05
SSIM(%)94.43+0.02+0.05+0.06+0.07
" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.307, + 0.51, + 0.323 + ], + "angle": 0, + "content": "5.2 Multi-Stage Information Flow" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.329, + 0.788, + 0.404 + ], + "angle": 0, + "content": "In Tab. 5 we investigate the effect of incorporating different numbers of previous iteration results into the current iteration for multi-stage information fusion. As the number increases, the reconstruction performance increases gradually. This implies the benefit of incorporating the information from previous iterations in DUM-based MRI reconstruction." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.425, + 0.569, + 0.441 + ], + "angle": 0, + "content": "5.3 Adjacent Sensitivity Map Estimation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.447, + 0.789, + 0.734 + ], + "angle": 0, + "content": "In Tab. 6 we examine the impact of various strategies shown in Fig. 4 for sensitivity map estimation on both reconstruction performance and memory consumption. The ACS data input to the SME network is represented as a complex tensor with shape \\((B,2a + 1,N,h,w)\\), where \\(2a + 1\\) represents the adjacent number, \\(N\\) denotes the number of coils of each slice and \\(B\\) indicates the batch size. Here, the batch size is linearly related to the memory consumption. Strategy A reshapes the tensor to a batch size of \\((2a + 1)BN\\) and a channel size of 1 for coil-by-coil estimation. Strategy B concatenates the coils within each slice, reducing the batch size to \\((2a + 1)B\\). Strategy C concatenates the same coils from adjacent slices, decreasing the batch size to \\(BN\\). Strategy D combines all coils within a single batch. Note that only strategies A and C can be used for data with variable numbers of coils. Our experiments demonstrate that Strategy C, which utilizes adjacent information, not only shows the best performance but also reduces memory consumption by a factor of \\(2a + 1\\) compared to Strategy A. Intriguingly, strategy B is found to be the least effective. Ideally, since the coils within the same slice should be uncorrelated, this strategy should not negatively impact performance. However, the noise that is coupled across coils during acquisition may be amplified under this strategy, degrading the SME performance." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.788, + 0.843 + ], + "angle": 0, + "content": "In Tab. 7, we further explore the impact of different adjacent types on Strategy C using the CMRxRecon cardiac datasets. These types include temporal information in the cine LAX dataset, contrast information in the T1 mapping dataset, and slice information in the cine SAX dataset. Strategy C shows consistent improvement over Strategy A for adjacent types of time and contrast. However, the result of the slice adjacency experiment deviates from our expectations, as previously observed in the brain dataset in Tab. 6. This discrepancy" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.358, + 0.128 + ], + "angle": 0, + "content": "B. Xin et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.145, + 0.788, + 0.201 + ], + "angle": 0, + "content": "Table 6: Comparison of different strategies for sensitivity map estimation on the Calgary-Campinas brain dataset under \\( \\times 10 \\) acceleration. The \"Batch\" number is linearly related to memory consumption. \"Flexibility\" indicates the ability to accept data with varying numbers of coils." + }, + { + "type": "table", + "bbox": [ + 0.306, + 0.213, + 0.699, + 0.309 + ], + "angle": 0, + "content": "
StrategyBatchChannelPSNR(dB)/SSIM(%)Flexibility
A|(2a+1)BN|135.60/94.22✓
B|(2a+1)BN|N35.30/93.98X
CBN|2a+1|35.88/94.43✓
DB||(2a+1)N|35.76/94.32X
" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.321, + 0.788, + 0.363 + ], + "angle": 0, + "content": "Table 7: Quantitative comparison of PSNR(dB)/SSIM(%) of different sensitivity map estimation strategies using different adjacent information types on the CMRxRecon cardiac dataset under \\(\\times 10\\) acceleration." + }, + { + "type": "table", + "bbox": [ + 0.362, + 0.375, + 0.642, + 0.454 + ], + "angle": 0, + "content": "
StrategyAdjacent type
TimeContrastSlice
A41.67/97.3542.62/98.1542.75/97.58
C41.94/97.5442.77/98.2242.53/97.54
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.482, + 0.788, + 0.558 + ], + "angle": 0, + "content": "is attributed to the considerable slice gap present in the cardiac dataset, which is \\(4.0\\mathrm{mm}\\) compared to no slice gap in the brain dataset. This suggests that adjacent reconstruction can be beneficial only when adjacent slices show high correlations; otherwise, it might even detract from performance by introducing confounding factors for the network." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.581, + 0.365, + 0.597 + ], + "angle": 0, + "content": "6 Limitations" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.613, + 0.788, + 0.703 + ], + "angle": 0, + "content": "Like other deep unrolled models, our approach may exhibit limited generalization to unseen sampling masks, modalities, and anatomies, and it may be susceptible to adversarial attacks. Hallucinations and artifacts may be observed when the model is applied to highly undersampled \\(k\\)-space data (e.g., at acceleration factors of 10), when trained on a limited amount of data, or when tested on out-of-distribution datasets." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.727, + 0.36, + 0.743 + ], + "angle": 0, + "content": "7 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.759, + 0.788, + 0.82 + ], + "angle": 0, + "content": "In this paper, we improve the deep unrolled model for multi-coil MRI reconstruction through informative gradient-based learning and memory-efficient sensitivity map estimation. Our proposed method achieves state-of-the-art performance on several public MRI reconstruction benchmarks." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.266, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Rethinking Deep Unrolled Model for Accelerated MRI Reconstruction" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.147, + 0.788, + 0.193 + ], + "angle": 0, + "content": "Acknowledgments. This research has been partially funded by research grants to D. Metaxas through NSF: 2310966, 2235405, 2212301, 2003874, and FA9550-23-1-0417 and NIH 2R01HL127661." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.218, + 0.323, + 0.234 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.251, + 0.787, + 0.28 + ], + "angle": 0, + "content": "1. Adler, J., Öktem, O.: Learned primal-dual reconstruction. IEEE transactions on medical imaging 37(6), 1322-1332 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.281, + 0.787, + 0.322 + ], + "angle": 0, + "content": "2. Aggarwal, H.K., Mani, M.P., Jacob, M.: Modl: Model-based deep learning architecture for inverse problems. IEEE transactions on medical imaging 38(2), 394-405 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.323, + 0.787, + 0.35 + ], + "angle": 0, + "content": "3. Arnold, T.C., Freeman, C.W., Litt, B., Stein, J.M.: Low-field mri: clinical promise and challenges. Journal of Magnetic Resonance Imaging 57(1), 25-44 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.351, + 0.787, + 0.405 + ], + "angle": 0, + "content": "4. Arvinte, M., Vishwanath, S., Tewfik, A.H., Tamir, J.I.: Deep j-sense: Accelerated mri reconstruction via unrolled alternating optimization. In: International conference on medical image computing and computer-assisted intervention. pp. 350-360. Springer (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.407, + 0.787, + 0.434 + ], + "angle": 0, + "content": "5. Beck, A., Teboulle, M.: A fast iterative shrinkage-thresholding algorithm for linear inverse problems. SIAM journal on imaging sciences 2(1), 183-202 (2009)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.435, + 0.787, + 0.476 + ], + "angle": 0, + "content": "6. Boyd, S., Parikh, N., Chu, E., Peleato, B., Eckstein, J., et al.: Distributed optimization and statistical learning via the alternating direction method of multipliers. Foundations and Trends in Machine learning 3(1), 1-122 (2011)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.477, + 0.787, + 0.504 + ], + "angle": 0, + "content": "7. Candès, E.J., et al.: Compressive sampling. In: Proceedings of the international congress of mathematicians. vol. 3, pp. 1433-1452. Madrid, Spain (2006)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.505, + 0.787, + 0.533 + ], + "angle": 0, + "content": "8. Deshmane, A., Gulani, V., Griswold, M.A., Seiberlich, N.: Parallel mr imaging. Journal of Magnetic Resonance Imaging 36(1), 55-72 (2012)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.534, + 0.787, + 0.56 + ], + "angle": 0, + "content": "9. Duchi, J., Hazan, E., Singer, Y.: Adaptive subgradient methods for online learning and stochastic optimization. Journal of machine learning research 12(7) (2011)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.561, + 0.787, + 0.616 + ], + "angle": 0, + "content": "10. El-Rewaidy, H., Fahmy, A.S., Pashakhanloo, F., Cai, X., Kucukseymen, S., Csecs, I., Neisius, U., Haji-Valizadeh, H., Menze, B., Nezafat, R.: Multi-domain convolutional neural network (md-cnn) for radial reconstruction of dynamic cardiac mri. Magnetic Resonance in Medicine 85(3), 1195–1208 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.617, + 0.787, + 0.658 + ], + "angle": 0, + "content": "11. Eo, T., Jun, Y., Kim, T., Jang, J., Lee, H.J., Hwang, D.: Kiki-net: cross-domain convolutional neural networks for reconstructing undersampled magnetic resonance images. Magnetic resonance in medicine 80(5), 2188-2201 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.659, + 0.787, + 0.7 + ], + "angle": 0, + "content": "12. Fabian, Z., Tinaz, B., Soltanolkotabi, M.: Humus-net: Hybrid unrolled multi-scale network architecture for accelerated mri reconstruction. Advances in Neural Information Processing Systems 35, 25306-25319 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.7, + 0.787, + 0.742 + ], + "angle": 0, + "content": "13. Gregor, K., LeCun, Y.: Learning fast approximations of sparse coding. In: Proceedings of the 27th international conference on international conference on machine learning. pp. 399-406 (2010)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.743, + 0.787, + 0.798 + ], + "angle": 0, + "content": "14. Griswold, M.A., Jakob, P.M., Heidemann, R.M., Nittka, M., Jellus, V., Wang, J., Kiefer, B., Haase, A.: Generalized autocalibrating partially parallel acquisitions (grappa). Magnetic Resonance in Medicine: An Official Journal of the International Society for Magnetic Resonance in Medicine 47(6), 1202-1210 (2002)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.799, + 0.787, + 0.84 + ], + "angle": 0, + "content": "15. Hammernik, K., Klatzer, T., Kobler, E., Recht, M.P., Sodickson, D.K., Pock, T., Knoll, F.: Learning a variational network for reconstruction of accelerated mri data. Magnetic resonance in medicine 79(6), 3055-3071 (2018)" + }, + { + "type": "list", + "bbox": [ + 0.219, + 0.251, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.356, + 0.127 + ], + "angle": 0, + "content": "B. Xin et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.787, + 0.203 + ], + "angle": 0, + "content": "16. Hosseini, S.A.H., Yaman, B., Moeller, S., Hong, M., Akçakaya, M.: Dense recurrent neural networks for accelerated mri: History-cognizant unrolling of optimization algorithms. IEEE Journal of Selected Topics in Signal Processing 14(6), 1280-1291 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.203, + 0.787, + 0.257 + ], + "angle": 0, + "content": "17. Jun, Y., Shin, H., Eo, T., Hwang, D.: Joint deep model-based mr image and coil sensitivity reconstruction network (joint-icnet) for fast mri. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5270-5279 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.257, + 0.787, + 0.284 + ], + "angle": 0, + "content": "18. Kingma, D.P., Ba, J.: Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980 (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.285, + 0.787, + 0.311 + ], + "angle": 0, + "content": "19. Lauterbur, P.C.: Image formation by induced local interactions: examples employing nuclear magnetic resonance. nature 242(5394), 190-191 (1973)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.311, + 0.787, + 0.338 + ], + "angle": 0, + "content": "20. LeCun, Y., Bengio, Y., Hinton, G.: Deep learning. nature 521(7553), 436-444 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.338, + 0.787, + 0.379 + ], + "angle": 0, + "content": "21. Liu, X., Pang, Y., Jin, R., Liu, Y., Wang, Z.: Dual-domain reconstruction network with v-net and k-net for fast mri. Magnetic Resonance in Medicine 88(6), 2694-2708 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.379, + 0.787, + 0.433 + ], + "angle": 0, + "content": "22. Lustig, M., Donoho, D., Pauly, J.M.: Sparse mri: The application of compressed sensing for rapid mr imaging. Magnetic Resonance in Medicine: An Official Journal of the International Society for Magnetic Resonance in Medicine 58(6), 1182-1195 (2007)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.433, + 0.787, + 0.473 + ], + "angle": 0, + "content": "23. Monga, V., Li, Y., Eldar, Y.C.: Algorithm unrolling: Interpretable, efficient deep learning for signal and image processing. IEEE Signal Processing Magazine 38(2), 18-44 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.473, + 0.787, + 0.514 + ], + "angle": 0, + "content": "24. Munoz, C., Fotaki, A., Botnar, R.M., Prieto, C.: Latest advances in image acceleration: all dimensions are fair game. Journal of Magnetic Resonance Imaging 57(2), 387-402 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.514, + 0.787, + 0.542 + ], + "angle": 0, + "content": "25. Nayak, K.S., Lim, Y., Campbell-Washburn, A.E., Steeden, J.: Real-time magnetic resonance imaging. Journal of Magnetic Resonance Imaging 55(1), 81-99 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.542, + 0.787, + 0.582 + ], + "angle": 0, + "content": "26. Ongie, G., Jalal, A., Metzler, C.A., Baraniuk, R.G., Dimakis, A.G., Willett, R.: Deep learning techniques for inverse problems in imaging. IEEE Journal on Selected Areas in Information Theory 1(1), 39-56 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.582, + 0.787, + 0.623 + ], + "angle": 0, + "content": "27. Ottesen, J.A., Caan, M.W., Groote, I.R., Bjornerud, A.: A densely interconnected network for deep learning accelerated mri. Magnetic Resonance Materials in Physics, Biology and Medicine 36(1), 65-77 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.623, + 0.787, + 0.678 + ], + "angle": 0, + "content": "28. Pezzotti, N., Yousefi, S., Elmahdy, M.S., Van Gemert, J.H.F., Schuelke, C., Doneva, M., Nielsen, T., Kastryulin, S., Lieleveldt, B.P., Van Osch, M.J., et al.: An adaptive intelligence algorithm for undersampled knee mri reconstruction. IEEE Access 8, 204825-204838 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.678, + 0.787, + 0.704 + ], + "angle": 0, + "content": "29. Polyak, B.T.: Some methods of speeding up the convergence of iteration methods. Ussr computational mathematics and mathematical physics 4(5), 1-17 (1964)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.705, + 0.787, + 0.745 + ], + "angle": 0, + "content": "30. Pruessmann, K.P., Weiger, M., Scheidegger, M.B., Boesiger, P.: Sense: sensitivity encoding for fast mri. Magnetic Resonance in Medicine: An Official Journal of the International Society for Magnetic Resonance in Medicine 42(5), 952-962 (1999)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.745, + 0.787, + 0.786 + ], + "angle": 0, + "content": "31. Putzky, P., Karkalousos, D., Teuwen, J., Miriakov, N., Bakker, B., Caan, M., Welling, M.: i-rim applied to the fastmri challenge. arXiv preprint arXiv:1910.08952 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.786, + 0.787, + 0.813 + ], + "angle": 0, + "content": "32. Putzky, P., Welling, M.: Recurrent inference machines for solving inverse problems. arXiv preprint arXiv:1706.04008 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.813, + 0.787, + 0.841 + ], + "angle": 0, + "content": "33. Putzky, P., Welling, M.: Invert to learn to invert. Advances in neural information processing systems 32 (2019)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.266, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "Rethinking Deep Unrolled Model for Accelerated MRI Reconstruction" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.788, + 0.189 + ], + "angle": 0, + "content": "34. Ramzi, Z., Chaithya, G., Starck, J.L., Ciuciu, P.: Nc-pdnet: A density-compensated unrolled network for 2d and 3d non-cartesian mri reconstruction. IEEE Transactions on Medical Imaging 41(7), 1625-1638 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.19, + 0.788, + 0.245 + ], + "angle": 0, + "content": "35. Ronneberger, O., Fischer, P., Brox, T.: U-net: Convolutional networks for biomedical image segmentation. In: Medical Image Computing and Computer-Assisted Intervention-MICCAI 2015: 18th International Conference, Munich, Germany, October 5-9, 2015, Proceedings, Part III 18. pp. 234-241. Springer (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.246, + 0.788, + 0.286 + ], + "angle": 0, + "content": "36. Schlemper, J., Caballero, J., Hajnal, J.V., Price, A.N., Rueckert, D.: A deep cascade of convolutional neural networks for dynamic mr image reconstruction. IEEE transactions on Medical Imaging 37(2), 491-503 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.287, + 0.788, + 0.328 + ], + "angle": 0, + "content": "37. Singh, D., Monga, A., de Moura, H.L., Zhang, X., Zibetti, M.V., Regatte, R.R.: Emerging trends in fast mri using deep-learning reconstruction on undersampled k-space data: a systematic review. Bioengineering 10(9), 1012 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.329, + 0.788, + 0.37 + ], + "angle": 0, + "content": "38. Song, J., Chen, B., Zhang, J.: Memory-augmented deep unfolding network for compressive sensing. In: Proceedings of the 29th ACM international conference on multimedia. pp. 4249-4258 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.371, + 0.788, + 0.424 + ], + "angle": 0, + "content": "39. Souza, R., Lucena, O., Garrafa, J., Gobbi, D., Saluzzi, M., Appenzeller, S., Rittner, L., Frayne, R., Lotufo, R.: An open, multi-vendor, multi-field-strength brain mr dataset and analysis of publicly available skull stripping methods agreement. NeuroImage 170, 482-494 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.425, + 0.788, + 0.495 + ], + "angle": 0, + "content": "40. Sriram, A., Zbontar, J., Murrell, T., Defazio, A., Zitnick, C.L., Yakubova, N., Knoll, F., Johnson, P.: End-to-end variational networks for accelerated mri reconstruction. In: Medical Image Computing and Computer Assisted Intervention-MICCAI 2020: 23rd International Conference, Lima, Peru, October 4-8, 2020, Proceedings, Part II 23. pp. 64-73. Springer (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.495, + 0.788, + 0.522 + ], + "angle": 0, + "content": "41. Sun, J., Li, H., Xu, Z., et al.: Deep admm-net for compressive sensing mri. Advances in neural information processing systems 29 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.522, + 0.788, + 0.563 + ], + "angle": 0, + "content": "42. Tieleman, T.: Lecture 6.5-rmsprop: Divide the gradient by a running average of its recent magnitude. COURSERA: Neural networks for machine learning 4(2), 26 (2012)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.564, + 0.788, + 0.605 + ], + "angle": 0, + "content": "43. Uecker, M., Lai, P., Murphy, M.J., Virtue, P., Elad, M., Pauly, J.M., Vasanawala, S.S., Lustig, M.: Espirit—an eigenvalue approach to autocalibrating parallel migraine: where sense meets grappa. Magnetic resonance in medicine 71(3), 990-1001 (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.606, + 0.788, + 0.646 + ], + "angle": 0, + "content": "44. Wang, C., Lyu, J., Wang, S., Qin, C., Guo, K., Zhang, X., Yu, X., Li, Y., Wang, F., Jin, J., et al.: Cmrxrecon: an open cardiac mri dataset for the competition of accelerated image reconstruction. arXiv preprint arXiv:2309.10836 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.647, + 0.788, + 0.702 + ], + "angle": 0, + "content": "45. Xin, B., Ye, M., Axel, L., Metaxas, D.N.: Fill the k-space and refine the image: Prompting for dynamic and multi-contrast mri reconstruction. In: International Workshop on Statistical Atlases and Computational Models of the Heart. pp. 261-273. Springer (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.702, + 0.788, + 0.758 + ], + "angle": 0, + "content": "46. Yiasemis, G., Sonke, J.J., Sánchez, C., Teuwen, J.: Recurrent variational network: a deep learning inverse problem solver applied to the task of accelerated mri reconstruction. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 732-741 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.759, + 0.788, + 0.799 + ], + "angle": 0, + "content": "47. Ying, L., Sheng, J.: Joint image reconstruction and sensitivity estimation in sense (jsense). Magnetic Resonance in Medicine: An Official Journal of the International Society for Magnetic Resonance in Medicine 57(6), 1196-1202 (2007)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.799, + 0.788, + 0.841 + ], + "angle": 0, + "content": "48. You, D., Xie, J., Zhang, J.: Ista-net++: Flexible deep unfolding network for compressive sensing. In: 2021 IEEE International Conference on Multimedia and Expo (ICME). pp. 1-6. IEEE (2021)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.788, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.356, + 0.127 + ], + "angle": 0, + "content": "B. Xin et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.787, + 0.189 + ], + "angle": 0, + "content": "49. Zbontar, J., Knoll, F., Sriram, A., Murrell, T., Huang, Z., Muckley, M.J., Defazio, A., Stern, R., Johnson, P., Bruno, M., et al.: fastmri: An open dataset and benchmarks for accelerated mri. arXiv preprint arXiv:1811.08839 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.19, + 0.786, + 0.217 + ], + "angle": 0, + "content": "50. Zeiler, M.D.: Adadelta: an adaptive learning rate method. arXiv preprint arXiv:1212.5701 (2012)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.218, + 0.787, + 0.244 + ], + "angle": 0, + "content": "51. Zhang, J., Chen, B., Xiong, R., Zhang, Y.: Physics-inspired compressive sensing: Beyond deep unrolling. IEEE Signal Processing Magazine 40(1), 58-72 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.245, + 0.787, + 0.286 + ], + "angle": 0, + "content": "52. Zhang, J., Ghanem, B.: Ista-net: Interpretable optimization-inspired deep network for image compressive sensing. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1828-1837 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.287, + 0.787, + 0.327 + ], + "angle": 0, + "content": "53. Zhang, J., Zhang, Z., Xie, J., Zhang, Y.: High-throughput deep unfolding network for compressive sensing mri. IEEE Journal of Selected Topics in Signal Processing 16(4), 750-761 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.328, + 0.787, + 0.369 + ], + "angle": 0, + "content": "54. Zhang, Z., Liu, Y., Liu, J., Wen, F., Zhu, C.: Amp-net: Denoising-based deep unfolding for compressive image sensing. IEEE Transactions on Image Processing 30, 1487-1500 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.37, + 0.787, + 0.424 + ], + "angle": 0, + "content": "55. Zhao, R., Yaman, B., Zhang, Y., Stewart, R., Dixon, A., Knoll, F., Huang, Z., Lui, Y.W., Hansen, M.S., Lungren, M.P.: fastmri+, clinical pathology annotations for knee and brain fully sampled magnetic resonance imaging data. Scientific Data 9(1), 152 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.425, + 0.787, + 0.467 + ], + "angle": 0, + "content": "56. Zhou, B., Zhou, S.K.: Dudornet: learning a dual-domain recurrent network for fast mri reconstruction with deep t1 prior. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 4273-4282 (2020)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.467 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/Rethinking Deep Unrolled Model for Accelerated MRI Reconstruction/27de6684-978b-4680-80b1-f68f3bc7d6a8_origin.pdf b/2024/Rethinking Deep Unrolled Model for Accelerated MRI Reconstruction/27de6684-978b-4680-80b1-f68f3bc7d6a8_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..a161b25f3bfc1d0cf61c9cfac37f993206fe7da6 --- /dev/null +++ b/2024/Rethinking Deep Unrolled Model for Accelerated MRI Reconstruction/27de6684-978b-4680-80b1-f68f3bc7d6a8_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:562191fbcbe6900a55148b879381a6c85e45f8f663d4b7bd619e770c7f4c1b0c +size 17844288 diff --git a/2024/Rethinking Deep Unrolled Model for Accelerated MRI Reconstruction/full.md b/2024/Rethinking Deep Unrolled Model for Accelerated MRI Reconstruction/full.md new file mode 100644 index 0000000000000000000000000000000000000000..4ac907f5c30cfb00b8f7fae886e6dd6c3351157d --- /dev/null +++ b/2024/Rethinking Deep Unrolled Model for Accelerated MRI Reconstruction/full.md @@ -0,0 +1,315 @@ +# Rethinking Deep Unrolled Model for Accelerated MRI Reconstruction + +Bingyu Xin $^{1}$ , Meng Ye $^{1}$ , Leon Axel $^{2}$ , and Dimitris N. Metaxas $^{1}$ + +$^{1}$ Rutgers University, New Brunswick, NJ 08901, USA + $^{2}$ New York University Langone Medical Center, New York, NY 10016, USA +bx64@rutgers.edu, dnm@cs.rutgers.edu + +Abstract. Magnetic Resonance Imaging (MRI) is a widely used imaging modality for clinical diagnostics and the planning of surgical interventions. Accelerated MRI seeks to mitigate the inherent limitation of long scanning time by reducing the amount of raw $k$ -space data required for image reconstruction. Recently, the deep unrolled model (DUM) has demonstrated significant effectiveness and improved interpretability for MRI reconstruction, by truncating and unrolling the conventional iterative reconstruction algorithms with deep neural networks. However, the potential of DUM for MRI reconstruction has not been fully exploited. In this paper, we first enhance the gradient and information flow within and across iteration stages of DUM, then we highlight the importance of using various adjacent information for accurate and memory-efficient sensitivity map estimation and improved multi-coil MRI reconstruction. Extensive experiments on several public MRI reconstruction datasets show that our method outperforms existing MRI reconstruction methods by a large margin. The code is available at https://github.com/hellopipu/PromptMR-plus. + +Keywords: MRI Reconstruction $\cdot$ Deep Unrolled Model $\cdot$ Gradient-based Learning $\cdot$ Sensitivity Map Estimation + +# 1 Introduction + +Magnetic Resonance Imaging (MRI) provides a radiation-free and highly versatile method for imaging the organs, tissues, and skeletal systems within the human body. Over the past 50 years since Paul Lauterbur produced the first MR image, MRI has evolved to become a cornerstone in clinical diagnostics [19]. However, the process of raw $k$ -space data acquisition in MRI is typically time-consuming. Accelerated MRI techniques tackle this issue by minimizing the amount of raw data that needs to be collected for image reconstruction, thereby shortening the duration of the scan. Modern advances in MRI technology, including Parallel Imaging (PI) [8] and Compressed Sensing (CS) [7,22], have significantly enhanced the efficiency and quality of MRI scans, making it possible to acquire high-resolution images within a considerably reduced time frame. However, with increased patient throughput and the requirements of emerging + +![](images/900094bd3e4a914e7a993ea5ddc6787739ca1aa21f3b1d95f1ea041d73e3bd3e.jpg) +(a) + +![](images/68c3288d72aee1772b6348c68e35071fc8ec10802088cfed0023d393f3268dc3.jpg) +Fig. 1: Our proposed method is memory-efficient and shows better reconstruction performance. (a) GPU memory consumption on Calgary-Campinas brain dataset, (b) PSNR/SSIM results vs. the number of parameters on the fastMRI knee $\times 8$ testset. + +![](images/72f4807090649c20b5a47c7a59c6ad35b2c4c216f389f9873d4d62338971ffed.jpg) +(b) + +technologies, such as real-time MRI [25] and low-field MRI [3], the development of accelerated MRI reconstruction methods remains a hot research topic [24]. + +Accelerated MRI reconstruction is a regularized inverse problem, which aims to reconstruct an unknown MR image from highly undersampled measurements in $k$ -space. Conventional iterative MRI reconstruction methods minimize a cost function comprising two main components: a data-consistency term, which assesses the alignment between $k$ -space predicted from the reconstructed image and the observed measurements, and a regularization term, which incorporates prior knowledge to encourage the emergence of desirable image attributes, e.g., sparsity. Over the past decade, deep learning [20] has emerged as a transformative approach to MRI reconstruction [26, 37]. Recently, the release of several large-scale public MRI reconstruction benchmarks, including the fastMRI dataset [49], Calgary-Campinas dataset [39], and CMRxRecon dataset [44], has significantly propelled the advancement of MRI reconstruction methods. Among deep learning-based MRI reconstruction approaches, the deep unrolled model (DUM) has garnered significant attention for its exceptional performance and ability to establish a concrete and systematic link between widely used iterative MRI reconstruction methods and deep neural networks [23]. However, as pointed out in [51], the current architecture design of DUM is inefficient, characterized by limited information transmission capacity and low flexibility and robustness. These shortcomings significantly restrict DUM's performance and its applicability to unseen data. + +In this work, we rethink and improve key components of the DUM-based MRI reconstruction approach. Our main contributions can be summarized as follows: + +- We improve the adaptive gradient algorithm and apply it to DUM, which can achieve self-adaptive dynamic learning rates adjusting for different spatial areas in an MR image. This approach provides a more flexible updating strategy in each iteration stage of DUM. +- We incorporate the momentum technique used in gradient descent acceleration and propose a multi-stage and multi-level feature aggregation scheme to accelerate the iteration convergence of DUM. + +- We highlight the importance of adjacent information which can be used to improve multi-coil MRI reconstruction, especially for accurate and memory-efficient sensitivity map estimation. As shown in Fig. 1(a), our method reduces GPU consumption to $\sim 55\%$ than the previous state-of-the-art. + +- We demonstrate the effectiveness of our proposed method on three public MRI reconstruction benchmarks of different anatomies: the fastMRI knee, the Calgary-Campinas brain and the CMRxRecon cardiac dataset. As shown in Fig. $\boxed{1}$ (b), our method achieves better PSNR and SSIM with fewer parameters than existing MRI Reconstruction methods. + +# 2 Related Work + +# 2.1 Deep Unrolled Model (DUM) + +DUM maps a truncated optimization algorithm into a deep neural network, iteratively alternating between gradient descent steps and proximal mapping steps. Its superior performance and good interpretability make it a preferred method for a wide range of inverse problems, including compressive sensing, image restoration and image reconstruction [23, 51]. Learned ISTA (LISTA) was the first unrolling method proposed for fast approximation of sparse coding [13]. LISTA unrolls the traditional ISTA algorithm [5] and truncates it to a fixed number of iterations. By learning from a training dataset, this trainable version of ISTA shows significant computational benefits. The success of deep learning further promotes the idea of relating conventional iterative optimization methods to deep neural networks, including ADMM-Net [41], ISTA-Net [52] and AMP-Net [54]. By introducing a condition module to transmit input information to each stage of the unrolled model, ISTA-Net++ [48] provides more flexibility to handle multi-ratio and multi-scene images for compressive sensing. In MADUN [38], a memory-augmented DUM is proposed to enhance the short-term and long-term information transmission in traditional DUMs. + +# 2.2 DUM for MRI Reconstruction + +ADMM-Net [41] and DC-CNN [36] are the pioneering DUMs for MRI reconstruction, establishing themselves as widely used baseline DUM-based MRI reconstruction methods. These approaches, based on Convolutional Neural Networks (CNNs), unroll the alternating direction method of multipliers (ADMM) [6] and the alternating minimization (AM) algorithm, respectively. KIKI-Net [11] employs a novel approach by alternating between $k$ -space completion blocks and image restoration blocks, demonstrating the advantages of cross-domain learning [10,21,56]. Adaptive-CS-Net finds it beneficial to inject information from neighboring slices into the model. This 2.5D reconstruction strategy has been adopted in subsequent works [12,45]. + +The development of DUM has led to various innovative approaches beyond simple unrolling, categorized broadly into three main variants for enhanced + +MRI reconstruction. (1) Firstly, inspired by Recurrent Neural Network (RNN), RIM [31, 32] uses a latent variable as temporal memory to track the progression of iterations. Recurrent-VarNet [46] extends this idea by designing a recurrent state initializer and recurrent units for the hidden state. In parallel, LPD-Net [1] and NCPD-Net [34] simplify this concept by utilizing a buffer design as additional learnable memory transmitting across iterations. (2) Secondly, inspired by efficient variants of proximal gradient descent (PGD), HC-PGD [16] proposes the history-cognizant unrolling method by concatenating all previous proximal operator output as input to each iteration. DIRCN [27] further proposes to add feature map interconnections between neighboring stages. These works suggest the importance of gradient flow and shared information between cascades for MRI reconstruction. (3) Moreover, Adaptive-CS-Net [28] incorporates soft MRI priors, including soft data consistency, image phase information and background priors, as additional input channels to each iteration. Similarly, HiTDUN [53] proposes the design of multi-channel input for every iteration. These methods underscore the advantages of introducing additional priors into the input at each iteration stage, aiming to dismantle the information bottleneck encountered within each iteration. + +# 2.3 Sensitivity Map Estimation in DUM + +Numerous works on DUM have achieved promising results on MRI reconstruction [11, 21, 36, 41, 56]. However, most of these methods focus exclusively on synthesized single-coil $k$ -space data, overlooking the more prevalent multi-coil imaging in real clinical MRI scans. This oversight neglects a more realistic scenario, given the widespread application of parallel imaging. Parallel imaging (PI) can speed up MRI, with the idea of using multiple receiving coils to reduce the number of phase encoding lines needed in acquiring raw $k$ -space data. The sensitivity map of a given coil is its local field profile, describing through reciprocity where that coil can efficiently pick up MR signals from the imaged region. The reconstruction of multiple coil $k$ -space data needs accurate estimation of the sensitivity maps, which are either directly measured, as in SENSE [30], or indirectly accounted using a center region of $k$ -space data, as in GRAPPA [14]. + +Robust and accurate sensitivity map estimation is crucial for multi-coil MRI reconstruction. The i-RIM directly reconstructs coil images without explicitly estimating the sensitivity maps. Adaptive-CS-Net [28] uses the normalized low-pass filtered coil images as sensitivity maps. VarNet [15] and MoDL [2] handle multi-coil data with pre-estimated sensitivity maps [43]. E2E-VarNet [40] advances the VarNet by learning sensitivity maps from the Auto-Calibration Signal (ACS) data of the $k$ -space in an end-to-end manner. Due to its superior performance and end-to-end approach, many subsequent studies are based on E2E-VarNet [12,45,46]. When the ACS data is insufficient for accurate sensitivity map estimation, [4,17] suggest iterative sensitivity map estimation and image reconstruction, which can be considered as a deep version of J-SENSE [47]. + +![](images/22c09dc106ed676d61d4256b30a5b6b94d61e66cb21d2087cc119876f13995c2.jpg) +(a) + +![](images/1732aed41d26115f74ed88d6b55d419b0a902f9e5018c214ce14924541f7f320.jpg) +(b) +Fig. 2: Overview of our proposed (a) self-adaptive and (b) momentum-accelerated gradient algorithm. $\hat{x}_{t+1}$ and $\hat{F}_t^m$ are the output of CNN and Momentum layer in Eq. (7) and Eq. (10), respectively. + +# 3 Proposed Method + +Drawing inspiration from adaptive gradient and momentum-accelerated algorithms in gradient descent optimization, we dynamically incorporate prior knowledge in the data flow, to enable more informative and faster convergence for gradient-based MRI reconstruction learning, as shown in Fig. 2. Subsequently, we enhance the concept of adjacent reconstruction by focusing on more robust and memory-efficient multi-coil sensitivity map estimation, as shown in Fig. 4. Please refer to Supplementary Section B for architecture details. + +# 3.1 Problem Formulation + +We can estimate the complex MR image $\pmb{x}$ from its undersampled measurements $\pmb{y}$ in $k$ -space by solving an optimization problem [22], + +$$ +\min _ {\boldsymbol {x}} \boldsymbol {f} (\boldsymbol {x}) = \min _ {\boldsymbol {x}} \frac {1}{2} \left\| \boldsymbol {A} \boldsymbol {x} - \boldsymbol {y} \right\| _ {2} ^ {2} + \boldsymbol {R} (\boldsymbol {x}), \tag {1} +$$ + +where $\mathbf{A} = M\mathbf{F}\mathbf{S}$ is the forward process of acquiring the $k$ -space measurements, which is a combination of coil sensitivity maps $\mathbf{S}$ , the Fourier transform $\mathcal{F}$ and an undersampling mask $\mathbf{M}$ . The first term $\frac{1}{2} ||\mathbf{A}\mathbf{x} - \mathbf{y}||_2^2$ represents the data consistency term, which ensures data fidelity, while the second term $\mathbf{R}(\mathbf{x})$ serves as the regularization term, aimed at promoting image domain priors, such as sparsity [7]. We can solve Eq. (1) iteratively via the gradient descent method, + +$$ +\boldsymbol {x} _ {t + 1} = \boldsymbol {x} _ {t} - \eta_ {t} \boldsymbol {g} _ {t}, \tag {2} +$$ + +$$ +\boldsymbol {g} _ {t} = \boldsymbol {A} ^ {H} (\boldsymbol {A} \boldsymbol {x} _ {t} - \boldsymbol {y}) + \nabla \boldsymbol {R} (\boldsymbol {x} _ {t}), \tag {3} +$$ + +where $\nabla R(\pmb{x}_t)$ is the gradient or the proximal mapping of $\pmb{R}$ , $\eta^t$ is the scalar learning rate at iteration $t$ , and $t \in \{0,1,\dots,T-1\}$ , with $T$ being the total number of truncated iterations. In the context of the deep unrolled models, it is customary to substitute $\nabla R(\pmb{x}_t)$ with a neural network whose parameters are learned from large-scale training datasets. This allows for the learning of more sophisticated image domain priors, compared to those that are hand-crafted. + +# 3.2 Self-Adaptive Gradient Algorithm + +During the iterative MRI reconstruction process, some regions in an image are straightforward to reconstruct, such as the background, which contains minimal information. However, other areas are more prone to aliasing artifacts resulting from downsampling the $k$ -space, which are challenging to eliminate. This difficulty often arises because the aliasing artifacts may closely mimic the appearance of anatomical structures within these regions, necessitating a more meticulous gradient-based correction at each iteration. Inspired by the adaptive gradient algorithm [9,18,42,50], instead of applying a uniform learning rate across the entire image space at each iteration, it is reasonable to utilize self-adaptive learning rates for individual pixels to achieve a more refined reconstruction result. Our generalization of the standard gradient descent in Eq. (2) employs the following update for each iteration: + +$$ +\boldsymbol {x} _ {t + 1} = \boldsymbol {x} _ {t} - \eta_ {t} \boldsymbol {g} _ {t} - \boldsymbol {E} _ {t} \odot \boldsymbol {g} _ {t}, \tag {4} +$$ + +$$ +\boldsymbol {E} _ {t} = \frac {\eta_ {t}}{\sqrt {\sum_ {\tau = 0} ^ {t} \boldsymbol {g} _ {\tau} ^ {2}} + \epsilon} - \eta_ {t}, \tag {5} +$$ + +where $\pmb{E}_{t}$ is the residual pixel-wise learning rate at the $t$ -th iteration, and the denominator in Eq. (5) computes the root of the sum of squares of all previous gradients of individual pixels. $\epsilon$ is a small number to improve the numerical stability. The operation $\odot$ denotes element-wise multiplication. We note that, the accumulation of historical pixel-wise gradients in Eq. (5) can help improve iteration convergence by + +(1) scaling down the updating steps for regions with historically large gradients, which may be prone to overshooting; +(2) scaling up the updating steps for regions with historically small gradients, which may benefit from a more aggressive update. + +To unroll our self-adaptive gradient algorithm with a DUM, we deviate from the common practice of substituting $\nabla R(\boldsymbol{x}_t)$ with a neural network in Eq. (4). There are compelling reasons for this. First, learning the $\nabla R(\boldsymbol{x}_t)$ term in $\boldsymbol{g}_t$ (see Eq. (3)) can lead to numerical instability in $\boldsymbol{E}_t$ during the early stage of training. Second, the adaptive gradient strategy in Eq. (5), which is the basic variant utilized in AdaGrad (9), may not be the most suitable choice for a DUM in the context of MRI reconstruction. To leverage the robust nonlinear representation capabilities of neural networks, we improve upon the adaptive gradient algorithm by reformulating Eq. (4) as follows: + +$$ +\begin{array}{l} \boldsymbol {x} _ {t + 1} = \boldsymbol {x} _ {t} - \eta_ {t} \boldsymbol {A} ^ {H} (\boldsymbol {A} \boldsymbol {x} _ {t} - \boldsymbol {y}) - \left(\eta_ {t} \nabla \boldsymbol {R} (\boldsymbol {x} _ {t}) + \boldsymbol {E} _ {t} \odot \boldsymbol {g} _ {t}\right) (6) \\ = \boldsymbol {x} _ {t} - \eta_ {t} \boldsymbol {A} ^ {H} (\boldsymbol {A} \boldsymbol {x} _ {t} - \boldsymbol {y}) + \operatorname {C N N} _ {t} \left(\boldsymbol {x} _ {t}, \boldsymbol {A} ^ {H} \boldsymbol {A} \boldsymbol {x} _ {t}, \boldsymbol {A} ^ {H} \boldsymbol {y}, \boldsymbol {s} _ {t}\right). (7) \\ \end{array} +$$ + +For the last term in Eq. (6), we reparametrize it with a CNN, as indicated in Eq. (7). This neural network takes as input the fundamental components from + +which the term is constructed, namely, $\boldsymbol{x}_t, \boldsymbol{A}^H \boldsymbol{A} \boldsymbol{x}_t, \boldsymbol{A}^H \boldsymbol{y}$ , and a learnable auxiliary memory $s_t$ , which serves as a hidden state to track the historical gradient information implicitly. Note that $\boldsymbol{A}^H \boldsymbol{y}$ yields $\boldsymbol{x}_0$ , which represents the image transformed from zero-filled $k$ -space. For further discussion, please refer to Supplementary Section E. + +# 3.3 Momentum-Accelerated Gradient Algorithm + +The momentum method [29] accelerates gradient descent in Eq. (2) by accumulating a velocity vector in directions of persistent reduction in the optimization objective across iterations: + +$$ +\boldsymbol {x} _ {t + 1} = \boldsymbol {x} _ {t} - \eta_ {t} \boldsymbol {m} _ {t}, \tag {8} +$$ + +$$ +\boldsymbol {m} _ {t} = \mu \boldsymbol {m} _ {t - 1} + (1 - \mu) \boldsymbol {g} _ {t}, \tag {9} +$$ + +where $\mu \in [0,1]$ is the momentum coefficient. The momentum $m_{t}$ is known as the exponential moving average of historical gradients. It can benefit the optimization of the current iteration, thereby considerably accelerating convergence. In Sec. 3.2 we introduced the hidden state $s_t$ , which is used to implicitly track the pixel-wise historical gradient information. Here, we propose explicit multi-stage and multi-level feature aggregation, to mimic the idea of momentum and increase the information flow across the iterations. For a typical multi-level encoder-decoder denoiser, such as Unet, in the $t$ -th iteration, we insert a Momentum layer before the $m$ -th level decoder layer. This layer concatenates features at the current stage and features from all previous iteration stages at the same level, followed by a $1\times 1$ convolution layer to reduce the dimensionality, and a Channel Attention Block (CAB) to adaptively fuse the multi-stage information, as described below: + +$$ +\operatorname {M o m e n t u m} \left(\boldsymbol {F} _ {0} ^ {m}, \dots , \boldsymbol {F} _ {t} ^ {m}\right) = \operatorname {C A B} \left(\operatorname {C o n v} _ {1 \times 1} \left(\operatorname {C o n c a t} \left(\boldsymbol {F} _ {0} ^ {m}, \dots , \boldsymbol {F} _ {t} ^ {m}\right)\right)\right), \tag {10} +$$ + +where $\pmb{F}_t^m$ is the input feature of the $m$ -th level decoder layer at the $t$ -th iteration. + +# 3.4 Adjacent Reconstruction + +Incorporating information from adjacent slices in multi-coil MRI reconstruction has been substantiated in prior work [12,28,45]. However, a significant limitation of adjacent slice reconstruction lies in its substantial memory consumption during the estimation of multi-coil sensitivity maps. This issue stems from the design of the sensitivity map estimation (SME) network in earlier work [40]. The network adopts a coil-by-coil estimation approach to accommodate a variable number of coils, increasing flexibility but leading to higher memory demands in the adjacent reconstruction setting (see Fig. 4(a)). To be more specific, for $k$ -space data with $N$ coils and $2a + 1$ adjacent slices, the total number of forward passes required for the SME process is calculated as $(2a + 1)N$ . Typically, the + +![](images/fe41e60de2c1d27f5b953c5ce27269df8a37cbd0d653b8629ab4995b3f832f80.jpg) +(a) time + +![](images/b9bb7b06238ea4f60be30ca52cb58101901d4c084f9186904a933bb117f8e79f.jpg) +(b) contrast + +![](images/dc95bd762fc55127bc99165074ca24869f845aad1874e00ca6ed85fcbc380ef3.jpg) +(c) slice + +![](images/042a2753b14a7f22137e73b6ec121a1f9e3ed1e552ced03aa722ea922202d833.jpg) + +![](images/d76db55f95f6809fbacded300b1ffd80f6685949b2743145013b14de5427cf57.jpg) +Fig. 3: The sensitivity maps exhibit variations across (a) time, (b) contrast and (c) slice dimension. First row: original images, second row: magnitude of differences between the sensitivity map of the central image and each of those of its adjacent images. + +![](images/d250f998439938584824d2ea82078788f8b04989a7593502658df666ea70d4c1.jpg) +Fig. 4: Concept illustration of different strategies for multi-coil sensitivity map estimation (SME). The main difference between the four strategies lies in the Auto-Calibration-Signal (ACS) information sharing between coils and adjacent MR image slices. Each single coil sensitivity map gets ACS information from (a) a single coil and a single image slice; (b) multiple coils but a single image slice; (c) a single coil but multiple adjacent image slices; (d) multiple coils and multiple adjacent image slices. + +![](images/23ea4db9164c7d3b1378b939654ffbef2b78f5a13940402bbafda1bf7e1e8ac2.jpg) + +value of $N$ ranges from 8 to 34 and $a$ is set to 2. We propose to reduce the number of forward passes, by utilizing the correlations between adjacent sensitivity maps. + +A sensitivity map is subject to variation, based on the position of the imaging subject and changes over time due to the subject motion. Additionally, in multicontest contrast MRI acquisition, variations in acquisition parameters, such as inversion time (TI) and repetition time (TR), can indirectly influence the sensitivity map estimation through the acquired $k$ -space data, as shown in Fig. ③. Consequently, the estimation of adjacent sensitivity maps is a spatially, temporally, and contrastively correlated process. We therefore propose to estimate sensitivity maps by utilizing information from spatially, temporally, and contrastively adjacent slices. As shown in Fig. ④(c), we simultaneously estimate a single coil sensitivity map at all adjacent $2a + 1$ slices using information across these slices. In this way, we can reduce the number of forward passes from $(2a + 1)N$ to $N$ . Our estimation of the $n$ -th sensitivity map set $S_{adj}^{n}$ can be described as: + +$$ +\boldsymbol {S} _ {a d j} ^ {n} = \mathbf {S M E} (\mathbf {A C S} \left(\boldsymbol {x} _ {0, a d j} ^ {n}\right)), \tag {11} +$$ + +where $S_{adj}^{n} = [S_{a - c}^{n},\dots,S_{c}^{n},\dots,S_{a + c}^{n}]$ is an adjacent sensitivity map set which is the concatenation of central slice sensitivity map $S_{c}^{n}$ with its 2a-adjacent slice sensitivity maps of the $n$ -th coil. $\mathbf{ACS}(\pmb{x}_{0,adj}^{n})$ is the $n$ -th coil Auto-Calibration Signal (ACS) within all adjacent $2a + 1$ slices. + +# 4 Experiments and Results + +In this section, we present the experimental results of our proposed approach. We showcase the reconstruction performance of our model across multiple publicly available datasets, including Calgary-Campinas brain [39], CMRxRecon cardiac [44], and fastMRI multi-coil knee [49] datasets. These datasets encompass a diverse range of anatomies, contrasts, views, slices and motion states. For details about the datasets, please refer to Supplementary Section A. For information on the experimental setup for each dataset, see Supplementary Section C. + +# 4.1 Calgary-Campinas Brain Results + +Tab. 1 presents the quantitative results for $\times 5$ and $\times 10$ acceleration factors on the Calgary-Campinas brain test set. Our method is compared with two state-of-the-art (SOTA) methods: Recurrent-VarNet [46] and PromptMR [45]. Our proposed method significantly outperforms these benchmarks, demonstrating its outstanding performance. Qualitative results under $\times 10$ acceleration are shown in Fig. 5. Interestingly, while PromptMR achieves significantly higher PSNR and SSIM values compared to Recurrent-VarNet, it exhibits less accurate reconstruction in the area highlighted by the red box. + +The highly undersampled $k$ -space data makes the ill-posed reconstruction challenging at aliasing artifacts removal. While leveraging large-scale training data to learn the image priors can improve the reconstruction accuracy, we note that Recurrent-VarNet learns a weaker prior, resulting in reconstructions that appear blurry. Conversely, PromptMR adopts a strong but overly biased prior, leading to sharper reconstructions but introducing incorrect structures. Our proposed method, in contrast, successfully learns the appropriate prior, achieving a reconstruction that closely mirrors the ground truth in terms of structural accuracy. + +# 4.2 CMRxRecon Cardiac Results + +We compare our method with PromptMR, the winning method of the MIC-CAI2023 CMRxRecon challenge, for $\times 4$ , $\times 8$ and $\times 10$ acceleration factors on the CMRxRecon cardiac test set. In Tab. ②, our method exhibits superior performance to PromptMR across different views and contrasts. Qualitative results under $\times 10$ acceleration are shown in Fig. ⑥. Our method reconstructs fine details of the interventricular septum, highlighted in the red box, with significantly higher clarity than the PromptMR method, which is severely compromised by aliasing artifacts. + +Table 1: Quantitative comparison of PSNR/SSIM (mean ± std) of different MRI reconstruction methods on the Calgary-Campinas brain dataset under $\times 5$ and $\times 10$ acceleration. The best and second best performance are in red and blue colors, respectively. + +
MethodAcc = 5×Acc = 10×
PSNR(dB)SSIM(%)PSNR(dB)SSIM(%)
Zero-filled25.27±1.3173.75±3.9324.35±1.2169.04±3.67
Recurrent-Varnet36.27±1.7294.37±1.3333.27±2.1291.47±2.23
PromptMR36.83±1.6894.85±1.2634.10±2.1192.69±1.92
Ours37.24±1.5795.15±1.2134.74±1.9793.38±1.65
+ +![](images/729c05983d97094615657e658bef127cb0dd05a3a6997eb5dd391418286572d5.jpg) +Fig. 5: The reconstruction results and SSIM error maps of various methods for the Calgary-Campinas brain imaging under $\times 10$ acceleration factor. The red boxes highlight the differences in the recovery of white and gray matter structures. + +# 4.3 FastMRI Knee Results + +We evaluated the PSNR and SSIM at acceleration factors of $\times 4$ and $\times 8$ on the fastMRI multi-coil knee test set. Our method is compared with several methods, including SOTA methods HUMUS-Net-L [12] and PromptMR. In the table, 'PD' and 'PDFS' denote proton density-weighted images without and with fat suppression, respectively. Tab. 3 demonstrates that our approach achieves the best performance across all other competitive methods on this large-scale dataset. Qualitative results under $\times 8$ acceleration are shown in Fig. 7. Our approach achieves a more detailed reconstruction of the meniscus region as shown in the red box, closely approximating the ground truth. In this region, meniscus tears were identified by radiologists from the fastMRI+ dataset [55]. The qualitative comparison suggests that our approach is more robust in reconstructing abnormal knee regions for highly accelerated acquisition. + +# 5 Abalation study + +In this section, we justify our design choices through ablation studies on the brain and cardiac datasets. + +Table 2: Quantitative comparison of PSNR/SSIM (mean ± std) of different MRI reconstruction methods on the CMRxRecon cardiac dataset under $\times 4$ , $\times 8$ and $\times 10$ acceleration. The best and second best performance are in red and blue colors, respectively. + +
AccMethodCineMapping
SAXLAXT1wT2w
PSNR(dB)SSIM(%)PSNR(dB)SSIM(%)PSNR(dB)SSIM(%)PSNR(dB)SSIM(%)
4×Zero-filled PromptMR Ours26.20±1.5472.99±5.1425.11±1.6269.87±4.7324.23±1.1667.27±3.8424.97±1.1077.16±2.91
45.78±1.8998.78±0.4345.58±1.7298.72±0.3746.49±2.2398.93±0.5742.02±1.9398.05±0.72
46.26±1.7698.87±0.3446.13±1.7098.83±0.3447.45±2.1399.09±0.4242.68±1.8498.25±0.59
8×Zero-filled PromptMR Ours25.18±1.6770.21±5.8724.34±1.6667.67±5.8023.41±1.2564.12±4.4324.21±1.1274.84±3.07
40.65±1.5296.97±0.6839.64±1.8796.36±1.0440.94±1.6997.28±0.8538.24±1.8296.53±1.09
41.44±1.4897.32±0.5840.49±1.8396.76±0.9442.08±1.6397.70±0.6739.23±1.7196.98±0.89
10×Zero-filled PromptMR Ours24.83±1.6769.35±5.9224.13±1.6766.89±5.5823.16±1.2764.14±4.4724.25±1.1476.08±2.91
39.18±1.5096.15±0.8038.28±1.6295.60±1.0738.99±1.5896.61±0.9937.21±1.7696.22±1.16
39.99±1.4996.58±0.7239.13±1.6696.05±0.9940.37±1.5897.19±0.7938.22±1.7096.70±0.96
+ +![](images/eaae49564f24e3170a75b5299d01398aac83fd9d2241e3866bd1dc246c8868da.jpg) +Fig. 6: The reconstruction results and SSIM error maps of various methods for the CM-RxRecon cardiac T1-weighted imaging under $\times 10$ acceleration factor. The red boxes highlight the differences in the recovery of the interventricular septum. + +Table 3: Quantitative comparison of PSNR/SSIM (mean ± std) of different MRI reconstruction methods on the fastMRI knee dataset under $\times 4$ and $\times 8$ acceleration. The best and second best performance are in red and blue colors, respectively. + +
MethodAcc = 4×Acc = 8×
PDPDSPDPDS
|PSNR(dB)|SSIM(%)|PSNR(dB)|SSIM(%)|PSNR(dB)|SSIM(%)|PSNR(dB)|SSIM(%)
Zero-filled30.86±1.7380.65±3.7631.00±3.3378.48±6.1627.70±1.8474.12±4.9427.97±2.0268.78±5.58
Unet 3537.27±1.7692.11±2.7237.07±2.4788.03±5.0435.41±2.1490.36±3.1334.84±1.5983.71±5.04
i-RIM 3339.61±2.3394.16±2.8738.23±3.1389.27±4.9637.91±2.4793.22±2.8535.84±1.7185.27±4.93
E2E-VarNet 4040.09±2.3294.56±2.6438.43±3.2489.50±4.9538.72±2.5093.93±2.6836.16±1.7685.70±4.94
HUMUS-Net 12----38.72±2.5094.07±2.6835.98±1.7585.77±4.93
HUMUS-Net-L 12----39.03±2.4794.18±2.6136.19±1.7985.84±4.92
PromptMR 4540.58±2.5994.87±2.7538.59±3.3389.75±4.9439.48±2.5494.51±2.5436.42±1.7986.10±4.93
Ours40.85±2.7495.00±2.8438.71±3.3789.85±4.9240.09±2.5894.90±2.4536.49±1.8886.29±4.95
+ +![](images/a39cd262337756e98b622481f41381c2830f008b46985726e4dee9f87681ca91.jpg) +Fig. 7: The reconstruction results and SSIM error maps of various methods for the fastMRI multi-coil knee proton density (PD) imaging under $\times 8$ acceleration. The red boxes highlight the differences in the recovery of the meniscus region. The red box in the reference image was annotated by radiologists from the fastMRI+ dataset [55] to indicate the knee abnormality of meniscus tears. + +Table 4: Abalation of different input combinations of CNN on the Calgary-Campinas brain dataset under $\times 10$ acceleration. + +
InputBaseline------Ours
xt✓✓✓✓✓✓✓✓
AHx✓✓✓✓✓✓✓✓
AHy✓✓✓✓✓✓✓✓
st✓✓✓✓✓✓✓✓
PSNR(dB)35.76+0.00+0.01+0.09+0.08+0.31+0.36+0.36
SSIM(%)94.33+0.00-0.01+0.06+0.05+0.25+0.27+0.28
+ +# 5.1 Input Variations to CNN + +We conducted an ablation study for different input combinations to CNN, to validate the effectiveness of our proposed method. Results are shown in Tab. 4. The baseline network in each iteration only accepts the output from the last iteration $\pmb{x}_t$ as input, which is the setting for most unrolled models, such as E2E-VartNet and PromptMR. Our proposed method takes additional input, $A^H A\pmb{x}_t$ , $A^H\pmb{y}$ and $s_t$ , which outperforms the baseline by a large margin. More specifically, adding $A^H A\pmb{x}_t$ or $A^H\pmb{y}$ solely to the input has no impact on the performance. Adding $s_t$ can improve the performance, as it serves as additional memory to track the progression of the reconstruction process. Adding $A^H A\pmb{x}_t$ together with $A^H\pmb{y}$ greatly improves the gains. We assume the reason is that the residual connection of two terms into the network can provide soft data consistency information to the model. Adding $A^H A\pmb{x}_t$ together with $s_t$ performs the closest to our final design. It seems the missing of $A^H\pmb{y}$ has little impact, as it is a constant value and its information may be compactly represented in the $s_t$ . This assumption is implied by the evidence that the result of adding $A^H\pmb{y}$ together with $s_t$ into the network does not perform better than adding $s_t$ alone. + +Table 5: Quantitative comparison of PSNR/SSIM using different numbers of previous stages' outputs for multi-stage feature fusion on the Calgary-Campinas brain dataset under $\times 10$ acceleration. + +
Input# of previous iterations
015911
PSNR(dB)35.88+0.02+0.04+0.04+0.05
SSIM(%)94.43+0.02+0.05+0.06+0.07
+ +# 5.2 Multi-Stage Information Flow + +In Tab. 5 we investigate the effect of incorporating different numbers of previous iteration results into the current iteration for multi-stage information fusion. As the number increases, the reconstruction performance increases gradually. This implies the benefit of incorporating the information from previous iterations in DUM-based MRI reconstruction. + +# 5.3 Adjacent Sensitivity Map Estimation + +In Tab. 6 we examine the impact of various strategies shown in Fig. 4 for sensitivity map estimation on both reconstruction performance and memory consumption. The ACS data input to the SME network is represented as a complex tensor with shape $(B,2a + 1,N,h,w)$ , where $2a + 1$ represents the adjacent number, $N$ denotes the number of coils of each slice and $B$ indicates the batch size. Here, the batch size is linearly related to the memory consumption. Strategy A reshapes the tensor to a batch size of $(2a + 1)BN$ and a channel size of 1 for coil-by-coil estimation. Strategy B concatenates the coils within each slice, reducing the batch size to $(2a + 1)B$ . Strategy C concatenates the same coils from adjacent slices, decreasing the batch size to $BN$ . Strategy D combines all coils within a single batch. Note that only strategies A and C can be used for data with variable numbers of coils. Our experiments demonstrate that Strategy C, which utilizes adjacent information, not only shows the best performance but also reduces memory consumption by a factor of $2a + 1$ compared to Strategy A. Intriguingly, strategy B is found to be the least effective. Ideally, since the coils within the same slice should be uncorrelated, this strategy should not negatively impact performance. However, the noise that is coupled across coils during acquisition may be amplified under this strategy, degrading the SME performance. + +In Tab. 7, we further explore the impact of different adjacent types on Strategy C using the CMRxRecon cardiac datasets. These types include temporal information in the cine LAX dataset, contrast information in the T1 mapping dataset, and slice information in the cine SAX dataset. Strategy C shows consistent improvement over Strategy A for adjacent types of time and contrast. However, the result of the slice adjacency experiment deviates from our expectations, as previously observed in the brain dataset in Tab. 6. This discrepancy + +Table 6: Comparison of different strategies for sensitivity map estimation on the Calgary-Campinas brain dataset under $\times 10$ acceleration. The "Batch" number is linearly related to memory consumption. "Flexibility" indicates the ability to accept data with varying numbers of coils. + +
StrategyBatchChannelPSNR(dB)/SSIM(%)Flexibility
A|(2a+1)BN|135.60/94.22✓
B|(2a+1)BN|N35.30/93.98X
CBN|2a+1|35.88/94.43✓
DB||(2a+1)N|35.76/94.32X
+ +Table 7: Quantitative comparison of PSNR(dB)/SSIM(%) of different sensitivity map estimation strategies using different adjacent information types on the CMRxRecon cardiac dataset under $\times 10$ acceleration. + +
StrategyAdjacent type
TimeContrastSlice
A41.67/97.3542.62/98.1542.75/97.58
C41.94/97.5442.77/98.2242.53/97.54
+ +is attributed to the considerable slice gap present in the cardiac dataset, which is $4.0\mathrm{mm}$ compared to no slice gap in the brain dataset. This suggests that adjacent reconstruction can be beneficial only when adjacent slices show high correlations; otherwise, it might even detract from performance by introducing confounding factors for the network. + +# 6 Limitations + +Like other deep unrolled models, our approach may exhibit limited generalization to unseen sampling masks, modalities, and anatomies, and it may be susceptible to adversarial attacks. Hallucinations and artifacts may be observed when the model is applied to highly undersampled $k$ -space data (e.g., at acceleration factors of 10), when trained on a limited amount of data, or when tested on out-of-distribution datasets. + +# 7 Conclusion + +In this paper, we improve the deep unrolled model for multi-coil MRI reconstruction through informative gradient-based learning and memory-efficient sensitivity map estimation. Our proposed method achieves state-of-the-art performance on several public MRI reconstruction benchmarks. + +Acknowledgments. This research has been partially funded by research grants to D. Metaxas through NSF: 2310966, 2235405, 2212301, 2003874, and FA9550-23-1-0417 and NIH 2R01HL127661. + +# References + +1. Adler, J., Öktem, O.: Learned primal-dual reconstruction. IEEE transactions on medical imaging 37(6), 1322-1332 (2018) +2. Aggarwal, H.K., Mani, M.P., Jacob, M.: Modl: Model-based deep learning architecture for inverse problems. IEEE transactions on medical imaging 38(2), 394-405 (2018) +3. Arnold, T.C., Freeman, C.W., Litt, B., Stein, J.M.: Low-field mri: clinical promise and challenges. Journal of Magnetic Resonance Imaging 57(1), 25-44 (2023) +4. Arvinte, M., Vishwanath, S., Tewfik, A.H., Tamir, J.I.: Deep j-sense: Accelerated mri reconstruction via unrolled alternating optimization. In: International conference on medical image computing and computer-assisted intervention. pp. 350-360. Springer (2021) +5. Beck, A., Teboulle, M.: A fast iterative shrinkage-thresholding algorithm for linear inverse problems. SIAM journal on imaging sciences 2(1), 183-202 (2009) +6. Boyd, S., Parikh, N., Chu, E., Peleato, B., Eckstein, J., et al.: Distributed optimization and statistical learning via the alternating direction method of multipliers. Foundations and Trends in Machine learning 3(1), 1-122 (2011) +7. Candès, E.J., et al.: Compressive sampling. In: Proceedings of the international congress of mathematicians. vol. 3, pp. 1433-1452. Madrid, Spain (2006) +8. Deshmane, A., Gulani, V., Griswold, M.A., Seiberlich, N.: Parallel mr imaging. Journal of Magnetic Resonance Imaging 36(1), 55-72 (2012) +9. Duchi, J., Hazan, E., Singer, Y.: Adaptive subgradient methods for online learning and stochastic optimization. Journal of machine learning research 12(7) (2011) +10. El-Rewaidy, H., Fahmy, A.S., Pashakhanloo, F., Cai, X., Kucukseymen, S., Csecs, I., Neisius, U., Haji-Valizadeh, H., Menze, B., Nezafat, R.: Multi-domain convolutional neural network (md-cnn) for radial reconstruction of dynamic cardiac mri. Magnetic Resonance in Medicine 85(3), 1195–1208 (2021) +11. Eo, T., Jun, Y., Kim, T., Jang, J., Lee, H.J., Hwang, D.: Kiki-net: cross-domain convolutional neural networks for reconstructing undersampled magnetic resonance images. Magnetic resonance in medicine 80(5), 2188-2201 (2018) +12. Fabian, Z., Tinaz, B., Soltanolkotabi, M.: Humus-net: Hybrid unrolled multi-scale network architecture for accelerated mri reconstruction. Advances in Neural Information Processing Systems 35, 25306-25319 (2022) +13. Gregor, K., LeCun, Y.: Learning fast approximations of sparse coding. In: Proceedings of the 27th international conference on international conference on machine learning. pp. 399-406 (2010) +14. Griswold, M.A., Jakob, P.M., Heidemann, R.M., Nittka, M., Jellus, V., Wang, J., Kiefer, B., Haase, A.: Generalized autocalibrating partially parallel acquisitions (grappa). Magnetic Resonance in Medicine: An Official Journal of the International Society for Magnetic Resonance in Medicine 47(6), 1202-1210 (2002) +15. Hammernik, K., Klatzer, T., Kobler, E., Recht, M.P., Sodickson, D.K., Pock, T., Knoll, F.: Learning a variational network for reconstruction of accelerated mri data. Magnetic resonance in medicine 79(6), 3055-3071 (2018) + +16. Hosseini, S.A.H., Yaman, B., Moeller, S., Hong, M., Akçakaya, M.: Dense recurrent neural networks for accelerated mri: History-cognizant unrolling of optimization algorithms. IEEE Journal of Selected Topics in Signal Processing 14(6), 1280-1291 (2020) +17. Jun, Y., Shin, H., Eo, T., Hwang, D.: Joint deep model-based mr image and coil sensitivity reconstruction network (joint-icnet) for fast mri. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5270-5279 (2021) +18. Kingma, D.P., Ba, J.: Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980 (2014) +19. Lauterbur, P.C.: Image formation by induced local interactions: examples employing nuclear magnetic resonance. nature 242(5394), 190-191 (1973) +20. LeCun, Y., Bengio, Y., Hinton, G.: Deep learning. nature 521(7553), 436-444 (2015) +21. Liu, X., Pang, Y., Jin, R., Liu, Y., Wang, Z.: Dual-domain reconstruction network with v-net and k-net for fast mri. Magnetic Resonance in Medicine 88(6), 2694-2708 (2022) +22. Lustig, M., Donoho, D., Pauly, J.M.: Sparse mri: The application of compressed sensing for rapid mr imaging. Magnetic Resonance in Medicine: An Official Journal of the International Society for Magnetic Resonance in Medicine 58(6), 1182-1195 (2007) +23. Monga, V., Li, Y., Eldar, Y.C.: Algorithm unrolling: Interpretable, efficient deep learning for signal and image processing. IEEE Signal Processing Magazine 38(2), 18-44 (2021) +24. Munoz, C., Fotaki, A., Botnar, R.M., Prieto, C.: Latest advances in image acceleration: all dimensions are fair game. Journal of Magnetic Resonance Imaging 57(2), 387-402 (2023) +25. Nayak, K.S., Lim, Y., Campbell-Washburn, A.E., Steeden, J.: Real-time magnetic resonance imaging. Journal of Magnetic Resonance Imaging 55(1), 81-99 (2022) +26. Ongie, G., Jalal, A., Metzler, C.A., Baraniuk, R.G., Dimakis, A.G., Willett, R.: Deep learning techniques for inverse problems in imaging. IEEE Journal on Selected Areas in Information Theory 1(1), 39-56 (2020) +27. Ottesen, J.A., Caan, M.W., Groote, I.R., Bjornerud, A.: A densely interconnected network for deep learning accelerated mri. Magnetic Resonance Materials in Physics, Biology and Medicine 36(1), 65-77 (2023) +28. Pezzotti, N., Yousefi, S., Elmahdy, M.S., Van Gemert, J.H.F., Schuelke, C., Doneva, M., Nielsen, T., Kastryulin, S., Lieleveldt, B.P., Van Osch, M.J., et al.: An adaptive intelligence algorithm for undersampled knee mri reconstruction. IEEE Access 8, 204825-204838 (2020) +29. Polyak, B.T.: Some methods of speeding up the convergence of iteration methods. Ussr computational mathematics and mathematical physics 4(5), 1-17 (1964) +30. Pruessmann, K.P., Weiger, M., Scheidegger, M.B., Boesiger, P.: Sense: sensitivity encoding for fast mri. Magnetic Resonance in Medicine: An Official Journal of the International Society for Magnetic Resonance in Medicine 42(5), 952-962 (1999) +31. Putzky, P., Karkalousos, D., Teuwen, J., Miriakov, N., Bakker, B., Caan, M., Welling, M.: i-rim applied to the fastmri challenge. arXiv preprint arXiv:1910.08952 (2019) +32. Putzky, P., Welling, M.: Recurrent inference machines for solving inverse problems. arXiv preprint arXiv:1706.04008 (2017) +33. Putzky, P., Welling, M.: Invert to learn to invert. Advances in neural information processing systems 32 (2019) + +34. Ramzi, Z., Chaithya, G., Starck, J.L., Ciuciu, P.: Nc-pdnet: A density-compensated unrolled network for 2d and 3d non-cartesian mri reconstruction. IEEE Transactions on Medical Imaging 41(7), 1625-1638 (2022) +35. Ronneberger, O., Fischer, P., Brox, T.: U-net: Convolutional networks for biomedical image segmentation. In: Medical Image Computing and Computer-Assisted Intervention-MICCAI 2015: 18th International Conference, Munich, Germany, October 5-9, 2015, Proceedings, Part III 18. pp. 234-241. Springer (2015) +36. Schlemper, J., Caballero, J., Hajnal, J.V., Price, A.N., Rueckert, D.: A deep cascade of convolutional neural networks for dynamic mr image reconstruction. IEEE transactions on Medical Imaging 37(2), 491-503 (2017) +37. Singh, D., Monga, A., de Moura, H.L., Zhang, X., Zibetti, M.V., Regatte, R.R.: Emerging trends in fast mri using deep-learning reconstruction on undersampled k-space data: a systematic review. Bioengineering 10(9), 1012 (2023) +38. Song, J., Chen, B., Zhang, J.: Memory-augmented deep unfolding network for compressive sensing. In: Proceedings of the 29th ACM international conference on multimedia. pp. 4249-4258 (2021) +39. Souza, R., Lucena, O., Garrafa, J., Gobbi, D., Saluzzi, M., Appenzeller, S., Rittner, L., Frayne, R., Lotufo, R.: An open, multi-vendor, multi-field-strength brain mr dataset and analysis of publicly available skull stripping methods agreement. NeuroImage 170, 482-494 (2018) +40. Sriram, A., Zbontar, J., Murrell, T., Defazio, A., Zitnick, C.L., Yakubova, N., Knoll, F., Johnson, P.: End-to-end variational networks for accelerated mri reconstruction. In: Medical Image Computing and Computer Assisted Intervention-MICCAI 2020: 23rd International Conference, Lima, Peru, October 4-8, 2020, Proceedings, Part II 23. pp. 64-73. Springer (2020) +41. Sun, J., Li, H., Xu, Z., et al.: Deep admm-net for compressive sensing mri. Advances in neural information processing systems 29 (2016) +42. Tieleman, T.: Lecture 6.5-rmsprop: Divide the gradient by a running average of its recent magnitude. COURSERA: Neural networks for machine learning 4(2), 26 (2012) +43. Uecker, M., Lai, P., Murphy, M.J., Virtue, P., Elad, M., Pauly, J.M., Vasanawala, S.S., Lustig, M.: Espirit—an eigenvalue approach to autocalibrating parallel migraine: where sense meets grappa. Magnetic resonance in medicine 71(3), 990-1001 (2014) +44. Wang, C., Lyu, J., Wang, S., Qin, C., Guo, K., Zhang, X., Yu, X., Li, Y., Wang, F., Jin, J., et al.: Cmrxrecon: an open cardiac mri dataset for the competition of accelerated image reconstruction. arXiv preprint arXiv:2309.10836 (2023) +45. Xin, B., Ye, M., Axel, L., Metaxas, D.N.: Fill the k-space and refine the image: Prompting for dynamic and multi-contrast mri reconstruction. In: International Workshop on Statistical Atlases and Computational Models of the Heart. pp. 261-273. Springer (2023) +46. Yiasemis, G., Sonke, J.J., Sánchez, C., Teuwen, J.: Recurrent variational network: a deep learning inverse problem solver applied to the task of accelerated mri reconstruction. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 732-741 (2022) +47. Ying, L., Sheng, J.: Joint image reconstruction and sensitivity estimation in sense (jsense). Magnetic Resonance in Medicine: An Official Journal of the International Society for Magnetic Resonance in Medicine 57(6), 1196-1202 (2007) +48. You, D., Xie, J., Zhang, J.: Ista-net++: Flexible deep unfolding network for compressive sensing. In: 2021 IEEE International Conference on Multimedia and Expo (ICME). pp. 1-6. IEEE (2021) + +49. Zbontar, J., Knoll, F., Sriram, A., Murrell, T., Huang, Z., Muckley, M.J., Defazio, A., Stern, R., Johnson, P., Bruno, M., et al.: fastmri: An open dataset and benchmarks for accelerated mri. arXiv preprint arXiv:1811.08839 (2018) +50. Zeiler, M.D.: Adadelta: an adaptive learning rate method. arXiv preprint arXiv:1212.5701 (2012) +51. Zhang, J., Chen, B., Xiong, R., Zhang, Y.: Physics-inspired compressive sensing: Beyond deep unrolling. IEEE Signal Processing Magazine 40(1), 58-72 (2023) +52. Zhang, J., Ghanem, B.: Ista-net: Interpretable optimization-inspired deep network for image compressive sensing. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1828-1837 (2018) +53. Zhang, J., Zhang, Z., Xie, J., Zhang, Y.: High-throughput deep unfolding network for compressive sensing mri. IEEE Journal of Selected Topics in Signal Processing 16(4), 750-761 (2022) +54. Zhang, Z., Liu, Y., Liu, J., Wen, F., Zhu, C.: Amp-net: Denoising-based deep unfolding for compressive image sensing. IEEE Transactions on Image Processing 30, 1487-1500 (2020) +55. Zhao, R., Yaman, B., Zhang, Y., Stewart, R., Dixon, A., Knoll, F., Huang, Z., Lui, Y.W., Hansen, M.S., Lungren, M.P.: fastmri+, clinical pathology annotations for knee and brain fully sampled magnetic resonance imaging data. Scientific Data 9(1), 152 (2022) +56. Zhou, B., Zhou, S.K.: Dudornet: learning a dual-domain recurrent network for fast mri reconstruction with deep t1 prior. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 4273-4282 (2020) \ No newline at end of file diff --git a/2024/Rethinking Deep Unrolled Model for Accelerated MRI Reconstruction/images.zip b/2024/Rethinking Deep Unrolled Model for Accelerated MRI Reconstruction/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..aa9b8d9baa903fcd94dc2abb0ec8a35774755b76 --- /dev/null +++ b/2024/Rethinking Deep Unrolled Model for Accelerated MRI Reconstruction/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:eed58337434bd1fa3d86753a11d446e574f3b01518af44b01fd8e5a2f544f16b +size 589426 diff --git a/2024/Rethinking Deep Unrolled Model for Accelerated MRI Reconstruction/layout.json b/2024/Rethinking Deep Unrolled Model for Accelerated MRI Reconstruction/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..b0575caa0e168c2e9a7949cdd8417dd955d9c768 --- /dev/null +++ b/2024/Rethinking Deep Unrolled Model for Accelerated MRI Reconstruction/layout.json @@ -0,0 +1,10356 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 135, + 111, + 479, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 111, + 479, + 148 + ], + "spans": [ + { + "bbox": [ + 135, + 111, + 479, + 148 + ], + "type": "text", + "content": "Rethinking Deep Unrolled Model for Accelerated MRI Reconstruction" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 152, + 167, + 462, + 182 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 167, + 462, + 182 + ], + "spans": [ + { + "bbox": [ + 152, + 167, + 462, + 182 + ], + "type": "text", + "content": "Bingyu Xin" + }, + { + "bbox": [ + 152, + 167, + 462, + 182 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 152, + 167, + 462, + 182 + ], + "type": "text", + "content": ", Meng Ye" + }, + { + "bbox": [ + 152, + 167, + 462, + 182 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 152, + 167, + 462, + 182 + ], + "type": "text", + "content": ", Leon Axel" + }, + { + "bbox": [ + 152, + 167, + 462, + 182 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 152, + 167, + 462, + 182 + ], + "type": "text", + "content": ", and Dimitris N. Metaxas" + }, + { + "bbox": [ + 152, + 167, + 462, + 182 + ], + "type": "inline_equation", + "content": "^{1}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 149, + 189, + 463, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 189, + 463, + 224 + ], + "spans": [ + { + "bbox": [ + 149, + 189, + 463, + 224 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 149, + 189, + 463, + 224 + ], + "type": "text", + "content": " Rutgers University, New Brunswick, NJ 08901, USA \n" + }, + { + "bbox": [ + 149, + 189, + 463, + 224 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 149, + 189, + 463, + 224 + ], + "type": "text", + "content": " New York University Langone Medical Center, New York, NY 10016, USA \nbx64@rutgers.edu, dnm@cs.rutgers.edu" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 160, + 254, + 455, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 254, + 455, + 443 + ], + "spans": [ + { + "bbox": [ + 160, + 254, + 455, + 443 + ], + "type": "text", + "content": "Abstract. Magnetic Resonance Imaging (MRI) is a widely used imaging modality for clinical diagnostics and the planning of surgical interventions. Accelerated MRI seeks to mitigate the inherent limitation of long scanning time by reducing the amount of raw " + }, + { + "bbox": [ + 160, + 254, + 455, + 443 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 160, + 254, + 455, + 443 + ], + "type": "text", + "content": "-space data required for image reconstruction. Recently, the deep unrolled model (DUM) has demonstrated significant effectiveness and improved interpretability for MRI reconstruction, by truncating and unrolling the conventional iterative reconstruction algorithms with deep neural networks. However, the potential of DUM for MRI reconstruction has not been fully exploited. In this paper, we first enhance the gradient and information flow within and across iteration stages of DUM, then we highlight the importance of using various adjacent information for accurate and memory-efficient sensitivity map estimation and improved multi-coil MRI reconstruction. Extensive experiments on several public MRI reconstruction datasets show that our method outperforms existing MRI reconstruction methods by a large margin. The code is available at https://github.com/hellopipu/PromptMR-plus." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 160, + 452, + 453, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 452, + 453, + 475 + ], + "spans": [ + { + "bbox": [ + 160, + 452, + 453, + 475 + ], + "type": "text", + "content": "Keywords: MRI Reconstruction " + }, + { + "bbox": [ + 160, + 452, + 453, + 475 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 452, + 453, + 475 + ], + "type": "text", + "content": " Deep Unrolled Model " + }, + { + "bbox": [ + 160, + 452, + 453, + 475 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 452, + 453, + 475 + ], + "type": "text", + "content": " Gradient-based Learning " + }, + { + "bbox": [ + 160, + 452, + 453, + 475 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 452, + 453, + 475 + ], + "type": "text", + "content": " Sensitivity Map Estimation" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 495, + 230, + 508 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 495, + 230, + 508 + ], + "spans": [ + { + "bbox": [ + 132, + 495, + 230, + 508 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 521, + 482, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 482, + 668 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 482, + 668 + ], + "type": "text", + "content": "Magnetic Resonance Imaging (MRI) provides a radiation-free and highly versatile method for imaging the organs, tissues, and skeletal systems within the human body. Over the past 50 years since Paul Lauterbur produced the first MR image, MRI has evolved to become a cornerstone in clinical diagnostics [19]. However, the process of raw " + }, + { + "bbox": [ + 130, + 521, + 482, + 668 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 521, + 482, + 668 + ], + "type": "text", + "content": "-space data acquisition in MRI is typically time-consuming. Accelerated MRI techniques tackle this issue by minimizing the amount of raw data that needs to be collected for image reconstruction, thereby shortening the duration of the scan. Modern advances in MRI technology, including Parallel Imaging (PI) [8] and Compressed Sensing (CS) [7,22], have significantly enhanced the efficiency and quality of MRI scans, making it possible to acquire high-resolution images within a considerably reduced time frame. However, with increased patient throughput and the requirements of emerging" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 138, + 113, + 254, + 197 + ], + "blocks": [ + { + "bbox": [ + 138, + 113, + 254, + 197 + ], + "lines": [ + { + "bbox": [ + 138, + 113, + 254, + 197 + ], + "spans": [ + { + "bbox": [ + 138, + 113, + 254, + 197 + ], + "type": "image", + "image_path": "900094bd3e4a914e7a993ea5ddc6787739ca1aa21f3b1d95f1ea041d73e3bd3e.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 192, + 198, + 202, + 207 + ], + "lines": [ + { + "bbox": [ + 192, + 198, + 202, + 207 + ], + "spans": [ + { + "bbox": [ + 192, + 198, + 202, + 207 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 255, + 114, + 366, + 197 + ], + "blocks": [ + { + "bbox": [ + 255, + 114, + 366, + 197 + ], + "lines": [ + { + "bbox": [ + 255, + 114, + 366, + 197 + ], + "spans": [ + { + "bbox": [ + 255, + 114, + 366, + 197 + ], + "type": "image", + "image_path": "68c3288d72aee1772b6348c68e35071fc8ec10802088cfed0023d393f3268dc3.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 215, + 480, + 248 + ], + "lines": [ + { + "bbox": [ + 130, + 215, + 480, + 248 + ], + "spans": [ + { + "bbox": [ + 130, + 215, + 480, + 248 + ], + "type": "text", + "content": "Fig. 1: Our proposed method is memory-efficient and shows better reconstruction performance. (a) GPU memory consumption on Calgary-Campinas brain dataset, (b) PSNR/SSIM results vs. the number of parameters on the fastMRI knee " + }, + { + "bbox": [ + 130, + 215, + 480, + 248 + ], + "type": "inline_equation", + "content": "\\times 8" + }, + { + "bbox": [ + 130, + 215, + 480, + 248 + ], + "type": "text", + "content": " testset." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 366, + 114, + 475, + 197 + ], + "blocks": [ + { + "bbox": [ + 366, + 114, + 475, + 197 + ], + "lines": [ + { + "bbox": [ + 366, + 114, + 475, + 197 + ], + "spans": [ + { + "bbox": [ + 366, + 114, + 475, + 197 + ], + "type": "image", + "image_path": "72f4807090649c20b5a47c7a59c6ad35b2c4c216f389f9873d4d62338971ffed.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 361, + 198, + 372, + 206 + ], + "lines": [ + { + "bbox": [ + 361, + 198, + 372, + 206 + ], + "spans": [ + { + "bbox": [ + 361, + 198, + 372, + 206 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 272, + 480, + 297 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 272, + 480, + 297 + ], + "spans": [ + { + "bbox": [ + 130, + 272, + 480, + 297 + ], + "type": "text", + "content": "technologies, such as real-time MRI [25] and low-field MRI [3], the development of accelerated MRI reconstruction methods remains a hot research topic [24]." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 297, + 481, + 536 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 297, + 481, + 536 + ], + "spans": [ + { + "bbox": [ + 130, + 297, + 481, + 536 + ], + "type": "text", + "content": "Accelerated MRI reconstruction is a regularized inverse problem, which aims to reconstruct an unknown MR image from highly undersampled measurements in " + }, + { + "bbox": [ + 130, + 297, + 481, + 536 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 297, + 481, + 536 + ], + "type": "text", + "content": "-space. Conventional iterative MRI reconstruction methods minimize a cost function comprising two main components: a data-consistency term, which assesses the alignment between " + }, + { + "bbox": [ + 130, + 297, + 481, + 536 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 297, + 481, + 536 + ], + "type": "text", + "content": "-space predicted from the reconstructed image and the observed measurements, and a regularization term, which incorporates prior knowledge to encourage the emergence of desirable image attributes, e.g., sparsity. Over the past decade, deep learning [20] has emerged as a transformative approach to MRI reconstruction [26, 37]. Recently, the release of several large-scale public MRI reconstruction benchmarks, including the fastMRI dataset [49], Calgary-Campinas dataset [39], and CMRxRecon dataset [44], has significantly propelled the advancement of MRI reconstruction methods. Among deep learning-based MRI reconstruction approaches, the deep unrolled model (DUM) has garnered significant attention for its exceptional performance and ability to establish a concrete and systematic link between widely used iterative MRI reconstruction methods and deep neural networks [23]. However, as pointed out in [51], the current architecture design of DUM is inefficient, characterized by limited information transmission capacity and low flexibility and robustness. These shortcomings significantly restrict DUM's performance and its applicability to unseen data." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 536, + 481, + 572 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 536, + 481, + 572 + ], + "spans": [ + { + "bbox": [ + 130, + 536, + 481, + 572 + ], + "type": "text", + "content": "In this work, we rethink and improve key components of the DUM-based MRI reconstruction approach. Our main contributions can be summarized as follows:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 581, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 138, + 581, + 480, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 581, + 480, + 628 + ], + "spans": [ + { + "bbox": [ + 138, + 581, + 480, + 628 + ], + "type": "text", + "content": "- We improve the adaptive gradient algorithm and apply it to DUM, which can achieve self-adaptive dynamic learning rates adjusting for different spatial areas in an MR image. This approach provides a more flexible updating strategy in each iteration stage of DUM." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 629, + 480, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 629, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 629, + 480, + 665 + ], + "type": "text", + "content": "- We incorporate the momentum technique used in gradient descent acceleration and propose a multi-stage and multi-level feature aggregation scheme to accelerate the iteration convergence of DUM." + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 219, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 219, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 219, + 101 + ], + "type": "text", + "content": "B. Xin et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 138, + 116, + 480, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 116, + 480, + 163 + ], + "spans": [ + { + "bbox": [ + 138, + 116, + 480, + 163 + ], + "type": "text", + "content": "- We highlight the importance of adjacent information which can be used to improve multi-coil MRI reconstruction, especially for accurate and memory-efficient sensitivity map estimation. As shown in Fig. 1(a), our method reduces GPU consumption to " + }, + { + "bbox": [ + 138, + 116, + 480, + 163 + ], + "type": "inline_equation", + "content": "\\sim 55\\%" + }, + { + "bbox": [ + 138, + 116, + 480, + 163 + ], + "type": "text", + "content": " than the previous state-of-the-art." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 139, + 164, + 481, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 164, + 481, + 224 + ], + "spans": [ + { + "bbox": [ + 139, + 164, + 481, + 224 + ], + "type": "text", + "content": "- We demonstrate the effectiveness of our proposed method on three public MRI reconstruction benchmarks of different anatomies: the fastMRI knee, the Calgary-Campinas brain and the CMRxRecon cardiac dataset. As shown in Fig. " + }, + { + "bbox": [ + 139, + 164, + 481, + 224 + ], + "type": "inline_equation", + "content": "\\boxed{1}" + }, + { + "bbox": [ + 139, + 164, + 481, + 224 + ], + "type": "text", + "content": " (b), our method achieves better PSNR and SSIM with fewer parameters than existing MRI Reconstruction methods." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 243, + 237, + 255 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 243, + 237, + 255 + ], + "spans": [ + { + "bbox": [ + 132, + 243, + 237, + 255 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 270, + 310, + 282 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 270, + 310, + 282 + ], + "spans": [ + { + "bbox": [ + 132, + 270, + 310, + 282 + ], + "type": "text", + "content": "2.1 Deep Unrolled Model (DUM)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 290, + 482, + 482 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 290, + 482, + 482 + ], + "spans": [ + { + "bbox": [ + 130, + 290, + 482, + 482 + ], + "type": "text", + "content": "DUM maps a truncated optimization algorithm into a deep neural network, iteratively alternating between gradient descent steps and proximal mapping steps. Its superior performance and good interpretability make it a preferred method for a wide range of inverse problems, including compressive sensing, image restoration and image reconstruction [23, 51]. Learned ISTA (LISTA) was the first unrolling method proposed for fast approximation of sparse coding [13]. LISTA unrolls the traditional ISTA algorithm [5] and truncates it to a fixed number of iterations. By learning from a training dataset, this trainable version of ISTA shows significant computational benefits. The success of deep learning further promotes the idea of relating conventional iterative optimization methods to deep neural networks, including ADMM-Net [41], ISTA-Net [52] and AMP-Net [54]. By introducing a condition module to transmit input information to each stage of the unrolled model, ISTA-Net++ [48] provides more flexibility to handle multi-ratio and multi-scene images for compressive sensing. In MADUN [38], a memory-augmented DUM is proposed to enhance the short-term and long-term information transmission in traditional DUMs." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 501, + 315, + 511 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 501, + 315, + 511 + ], + "spans": [ + { + "bbox": [ + 132, + 501, + 315, + 511 + ], + "type": "text", + "content": "2.2 DUM for MRI Reconstruction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 521, + 482, + 642 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 482, + 642 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 482, + 642 + ], + "type": "text", + "content": "ADMM-Net [41] and DC-CNN [36] are the pioneering DUMs for MRI reconstruction, establishing themselves as widely used baseline DUM-based MRI reconstruction methods. These approaches, based on Convolutional Neural Networks (CNNs), unroll the alternating direction method of multipliers (ADMM) [6] and the alternating minimization (AM) algorithm, respectively. KIKI-Net [11] employs a novel approach by alternating between " + }, + { + "bbox": [ + 130, + 521, + 482, + 642 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 521, + 482, + 642 + ], + "type": "text", + "content": "-space completion blocks and image restoration blocks, demonstrating the advantages of cross-domain learning [10,21,56]. Adaptive-CS-Net finds it beneficial to inject information from neighboring slices into the model. This 2.5D reconstruction strategy has been adopted in subsequent works [12,45]." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 642, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 642, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 642, + 481, + 665 + ], + "type": "text", + "content": "The development of DUM has led to various innovative approaches beyond simple unrolling, categorized broadly into three main variants for enhanced" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 162, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 162, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 162, + 91, + 447, + 102 + ], + "type": "text", + "content": "Rethinking Deep Unrolled Model for Accelerated MRI Reconstruction" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 115, + 482, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 115, + 482, + 331 + ], + "spans": [ + { + "bbox": [ + 130, + 115, + 482, + 331 + ], + "type": "text", + "content": "MRI reconstruction. (1) Firstly, inspired by Recurrent Neural Network (RNN), RIM [31, 32] uses a latent variable as temporal memory to track the progression of iterations. Recurrent-VarNet [46] extends this idea by designing a recurrent state initializer and recurrent units for the hidden state. In parallel, LPD-Net [1] and NCPD-Net [34] simplify this concept by utilizing a buffer design as additional learnable memory transmitting across iterations. (2) Secondly, inspired by efficient variants of proximal gradient descent (PGD), HC-PGD [16] proposes the history-cognizant unrolling method by concatenating all previous proximal operator output as input to each iteration. DIRCN [27] further proposes to add feature map interconnections between neighboring stages. These works suggest the importance of gradient flow and shared information between cascades for MRI reconstruction. (3) Moreover, Adaptive-CS-Net [28] incorporates soft MRI priors, including soft data consistency, image phase information and background priors, as additional input channels to each iteration. Similarly, HiTDUN [53] proposes the design of multi-channel input for every iteration. These methods underscore the advantages of introducing additional priors into the input at each iteration stage, aiming to dismantle the information bottleneck encountered within each iteration." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 358, + 346, + 371 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 358, + 346, + 371 + ], + "spans": [ + { + "bbox": [ + 132, + 358, + 346, + 371 + ], + "type": "text", + "content": "2.3 Sensitivity Map Estimation in DUM" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 388, + 482, + 532 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 388, + 482, + 532 + ], + "spans": [ + { + "bbox": [ + 130, + 388, + 482, + 532 + ], + "type": "text", + "content": "Numerous works on DUM have achieved promising results on MRI reconstruction [11, 21, 36, 41, 56]. However, most of these methods focus exclusively on synthesized single-coil " + }, + { + "bbox": [ + 130, + 388, + 482, + 532 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 388, + 482, + 532 + ], + "type": "text", + "content": "-space data, overlooking the more prevalent multi-coil imaging in real clinical MRI scans. This oversight neglects a more realistic scenario, given the widespread application of parallel imaging. Parallel imaging (PI) can speed up MRI, with the idea of using multiple receiving coils to reduce the number of phase encoding lines needed in acquiring raw " + }, + { + "bbox": [ + 130, + 388, + 482, + 532 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 388, + 482, + 532 + ], + "type": "text", + "content": "-space data. The sensitivity map of a given coil is its local field profile, describing through reciprocity where that coil can efficiently pick up MR signals from the imaged region. The reconstruction of multiple coil " + }, + { + "bbox": [ + 130, + 388, + 482, + 532 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 388, + 482, + 532 + ], + "type": "text", + "content": "-space data needs accurate estimation of the sensitivity maps, which are either directly measured, as in SENSE [30], or indirectly accounted using a center region of " + }, + { + "bbox": [ + 130, + 388, + 482, + 532 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 388, + 482, + 532 + ], + "type": "text", + "content": "-space data, as in GRAPPA [14]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 533, + 482, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 482, + 668 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 482, + 668 + ], + "type": "text", + "content": "Robust and accurate sensitivity map estimation is crucial for multi-coil MRI reconstruction. The i-RIM directly reconstructs coil images without explicitly estimating the sensitivity maps. Adaptive-CS-Net [28] uses the normalized low-pass filtered coil images as sensitivity maps. VarNet [15] and MoDL [2] handle multi-coil data with pre-estimated sensitivity maps [43]. E2E-VarNet [40] advances the VarNet by learning sensitivity maps from the Auto-Calibration Signal (ACS) data of the " + }, + { + "bbox": [ + 130, + 533, + 482, + 668 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 533, + 482, + 668 + ], + "type": "text", + "content": "-space in an end-to-end manner. Due to its superior performance and end-to-end approach, many subsequent studies are based on E2E-VarNet [12,45,46]. When the ACS data is insufficient for accurate sensitivity map estimation, [4,17] suggest iterative sensitivity map estimation and image reconstruction, which can be considered as a deep version of J-SENSE [47]." + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 219, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 219, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 219, + 101 + ], + "type": "text", + "content": "B. Xin et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 148, + 112, + 297, + 181 + ], + "blocks": [ + { + "bbox": [ + 148, + 112, + 297, + 181 + ], + "lines": [ + { + "bbox": [ + 148, + 112, + 297, + 181 + ], + "spans": [ + { + "bbox": [ + 148, + 112, + 297, + 181 + ], + "type": "image", + "image_path": "22c09dc106ed676d61d4256b30a5b6b94d61e66cb21d2087cc119876f13995c2.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 216, + 182, + 228, + 191 + ], + "lines": [ + { + "bbox": [ + 216, + 182, + 228, + 191 + ], + "spans": [ + { + "bbox": [ + 216, + 182, + 228, + 191 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 337, + 112, + 445, + 181 + ], + "blocks": [ + { + "bbox": [ + 337, + 112, + 445, + 181 + ], + "lines": [ + { + "bbox": [ + 337, + 112, + 445, + 181 + ], + "spans": [ + { + "bbox": [ + 337, + 112, + 445, + 181 + ], + "type": "image", + "image_path": "1732aed41d26115f74ed88d6b55d419b0a902f9e5018c214ce14924541f7f320.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 385, + 182, + 398, + 191 + ], + "lines": [ + { + "bbox": [ + 385, + 182, + 398, + 191 + ], + "spans": [ + { + "bbox": [ + 385, + 182, + 398, + 191 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 200, + 482, + 236 + ], + "lines": [ + { + "bbox": [ + 130, + 200, + 482, + 236 + ], + "spans": [ + { + "bbox": [ + 130, + 200, + 482, + 236 + ], + "type": "text", + "content": "Fig. 2: Overview of our proposed (a) self-adaptive and (b) momentum-accelerated gradient algorithm. " + }, + { + "bbox": [ + 130, + 200, + 482, + 236 + ], + "type": "inline_equation", + "content": "\\hat{x}_{t+1}" + }, + { + "bbox": [ + 130, + 200, + 482, + 236 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 200, + 482, + 236 + ], + "type": "inline_equation", + "content": "\\hat{F}_t^m" + }, + { + "bbox": [ + 130, + 200, + 482, + 236 + ], + "type": "text", + "content": " are the output of CNN and Momentum layer in Eq. (7) and Eq. (10), respectively." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 255, + 261, + 269 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 255, + 261, + 269 + ], + "spans": [ + { + "bbox": [ + 132, + 255, + 261, + 269 + ], + "type": "text", + "content": "3 Proposed Method" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 280, + 482, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 280, + 482, + 365 + ], + "spans": [ + { + "bbox": [ + 130, + 280, + 482, + 365 + ], + "type": "text", + "content": "Drawing inspiration from adaptive gradient and momentum-accelerated algorithms in gradient descent optimization, we dynamically incorporate prior knowledge in the data flow, to enable more informative and faster convergence for gradient-based MRI reconstruction learning, as shown in Fig. 2. Subsequently, we enhance the concept of adjacent reconstruction by focusing on more robust and memory-efficient multi-coil sensitivity map estimation, as shown in Fig. 4. Please refer to Supplementary Section B for architecture details." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 381, + 269, + 392 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 381, + 269, + 392 + ], + "spans": [ + { + "bbox": [ + 132, + 381, + 269, + 392 + ], + "type": "text", + "content": "3.1 Problem Formulation" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 401, + 481, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 401, + 481, + 426 + ], + "spans": [ + { + "bbox": [ + 130, + 401, + 481, + 426 + ], + "type": "text", + "content": "We can estimate the complex MR image " + }, + { + "bbox": [ + 130, + 401, + 481, + 426 + ], + "type": "inline_equation", + "content": "\\pmb{x}" + }, + { + "bbox": [ + 130, + 401, + 481, + 426 + ], + "type": "text", + "content": " from its undersampled measurements " + }, + { + "bbox": [ + 130, + 401, + 481, + 426 + ], + "type": "inline_equation", + "content": "\\pmb{y}" + }, + { + "bbox": [ + 130, + 401, + 481, + 426 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 130, + 401, + 481, + 426 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 401, + 481, + 426 + ], + "type": "text", + "content": "-space by solving an optimization problem [22]," + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 221, + 433, + 481, + 456 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 221, + 433, + 481, + 456 + ], + "spans": [ + { + "bbox": [ + 221, + 433, + 481, + 456 + ], + "type": "interline_equation", + "content": "\\min _ {\\boldsymbol {x}} \\boldsymbol {f} (\\boldsymbol {x}) = \\min _ {\\boldsymbol {x}} \\frac {1}{2} \\left\\| \\boldsymbol {A} \\boldsymbol {x} - \\boldsymbol {y} \\right\\| _ {2} ^ {2} + \\boldsymbol {R} (\\boldsymbol {x}), \\tag {1}", + "image_path": "9c7599314d484317bfce4cc59e73d7f22f81806df050b5089e8cb6ac5c80a666.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 463, + 482, + 536 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 463, + 482, + 536 + ], + "spans": [ + { + "bbox": [ + 130, + 463, + 482, + 536 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 463, + 482, + 536 + ], + "type": "inline_equation", + "content": "\\mathbf{A} = M\\mathbf{F}\\mathbf{S}" + }, + { + "bbox": [ + 130, + 463, + 482, + 536 + ], + "type": "text", + "content": " is the forward process of acquiring the " + }, + { + "bbox": [ + 130, + 463, + 482, + 536 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 463, + 482, + 536 + ], + "type": "text", + "content": "-space measurements, which is a combination of coil sensitivity maps " + }, + { + "bbox": [ + 130, + 463, + 482, + 536 + ], + "type": "inline_equation", + "content": "\\mathbf{S}" + }, + { + "bbox": [ + 130, + 463, + 482, + 536 + ], + "type": "text", + "content": ", the Fourier transform " + }, + { + "bbox": [ + 130, + 463, + 482, + 536 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 130, + 463, + 482, + 536 + ], + "type": "text", + "content": " and an undersampling mask " + }, + { + "bbox": [ + 130, + 463, + 482, + 536 + ], + "type": "inline_equation", + "content": "\\mathbf{M}" + }, + { + "bbox": [ + 130, + 463, + 482, + 536 + ], + "type": "text", + "content": ". The first term " + }, + { + "bbox": [ + 130, + 463, + 482, + 536 + ], + "type": "inline_equation", + "content": "\\frac{1}{2} ||\\mathbf{A}\\mathbf{x} - \\mathbf{y}||_2^2" + }, + { + "bbox": [ + 130, + 463, + 482, + 536 + ], + "type": "text", + "content": " represents the data consistency term, which ensures data fidelity, while the second term " + }, + { + "bbox": [ + 130, + 463, + 482, + 536 + ], + "type": "inline_equation", + "content": "\\mathbf{R}(\\mathbf{x})" + }, + { + "bbox": [ + 130, + 463, + 482, + 536 + ], + "type": "text", + "content": " serves as the regularization term, aimed at promoting image domain priors, such as sparsity [7]. We can solve Eq. (1) iteratively via the gradient descent method," + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 232, + 556, + 481, + 569 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 232, + 556, + 481, + 569 + ], + "spans": [ + { + "bbox": [ + 232, + 556, + 481, + 569 + ], + "type": "interline_equation", + "content": "\\boldsymbol {x} _ {t + 1} = \\boldsymbol {x} _ {t} - \\eta_ {t} \\boldsymbol {g} _ {t}, \\tag {2}", + "image_path": "70b973af3f5e905c275d0f415205ab332e7aac5939a47d32d4e11f005969a26a.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 242, + 571, + 481, + 586 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 242, + 571, + 481, + 586 + ], + "spans": [ + { + "bbox": [ + 242, + 571, + 481, + 586 + ], + "type": "interline_equation", + "content": "\\boldsymbol {g} _ {t} = \\boldsymbol {A} ^ {H} (\\boldsymbol {A} \\boldsymbol {x} _ {t} - \\boldsymbol {y}) + \\nabla \\boldsymbol {R} (\\boldsymbol {x} _ {t}), \\tag {3}", + "image_path": "3f81d8dbb8247436dcf649a056190587711610fb4a7692bc26c463498089f9ec.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 593, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 593, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 593, + 482, + 666 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 593, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\nabla R(\\pmb{x}_t)" + }, + { + "bbox": [ + 130, + 593, + 482, + 666 + ], + "type": "text", + "content": " is the gradient or the proximal mapping of " + }, + { + "bbox": [ + 130, + 593, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\pmb{R}" + }, + { + "bbox": [ + 130, + 593, + 482, + 666 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 593, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\eta^t" + }, + { + "bbox": [ + 130, + 593, + 482, + 666 + ], + "type": "text", + "content": " is the scalar learning rate at iteration " + }, + { + "bbox": [ + 130, + 593, + 482, + 666 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 593, + 482, + 666 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 593, + 482, + 666 + ], + "type": "inline_equation", + "content": "t \\in \\{0,1,\\dots,T-1\\}" + }, + { + "bbox": [ + 130, + 593, + 482, + 666 + ], + "type": "text", + "content": ", with " + }, + { + "bbox": [ + 130, + 593, + 482, + 666 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 130, + 593, + 482, + 666 + ], + "type": "text", + "content": " being the total number of truncated iterations. In the context of the deep unrolled models, it is customary to substitute " + }, + { + "bbox": [ + 130, + 593, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\nabla R(\\pmb{x}_t)" + }, + { + "bbox": [ + 130, + 593, + 482, + 666 + ], + "type": "text", + "content": " with a neural network whose parameters are learned from large-scale training datasets. This allows for the learning of more sophisticated image domain priors, compared to those that are hand-crafted." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 162, + 90, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 162, + 90, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 162, + 90, + 448, + 102 + ], + "type": "text", + "content": "Rethinking Deep Unrolled Model for Accelerated MRI Reconstruction" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 332, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 332, + 129 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 332, + 129 + ], + "type": "text", + "content": "3.2 Self-Adaptive Gradient Algorithm" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 137, + 482, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 137, + 482, + 281 + ], + "spans": [ + { + "bbox": [ + 130, + 137, + 482, + 281 + ], + "type": "text", + "content": "During the iterative MRI reconstruction process, some regions in an image are straightforward to reconstruct, such as the background, which contains minimal information. However, other areas are more prone to aliasing artifacts resulting from downsampling the " + }, + { + "bbox": [ + 130, + 137, + 482, + 281 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 137, + 482, + 281 + ], + "type": "text", + "content": "-space, which are challenging to eliminate. This difficulty often arises because the aliasing artifacts may closely mimic the appearance of anatomical structures within these regions, necessitating a more meticulous gradient-based correction at each iteration. Inspired by the adaptive gradient algorithm [9,18,42,50], instead of applying a uniform learning rate across the entire image space at each iteration, it is reasonable to utilize self-adaptive learning rates for individual pixels to achieve a more refined reconstruction result. Our generalization of the standard gradient descent in Eq. (2) employs the following update for each iteration:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 243, + 291, + 481, + 305 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 243, + 291, + 481, + 305 + ], + "spans": [ + { + "bbox": [ + 243, + 291, + 481, + 305 + ], + "type": "interline_equation", + "content": "\\boldsymbol {x} _ {t + 1} = \\boldsymbol {x} _ {t} - \\eta_ {t} \\boldsymbol {g} _ {t} - \\boldsymbol {E} _ {t} \\odot \\boldsymbol {g} _ {t}, \\tag {4}", + "image_path": "d0b623ad37f40933c58d091a62ebbb2056e2b99b64677d716d9f6ac8511ee4e1.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 252, + 307, + 481, + 338 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 252, + 307, + 481, + 338 + ], + "spans": [ + { + "bbox": [ + 252, + 307, + 481, + 338 + ], + "type": "interline_equation", + "content": "\\boldsymbol {E} _ {t} = \\frac {\\eta_ {t}}{\\sqrt {\\sum_ {\\tau = 0} ^ {t} \\boldsymbol {g} _ {\\tau} ^ {2}} + \\epsilon} - \\eta_ {t}, \\tag {5}", + "image_path": "f0e64b4be60bb68f921be8f40a7c90270e88f0f77eb036d622d4ea57f8c9e04f.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 346, + 482, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 346, + 482, + 418 + ], + "spans": [ + { + "bbox": [ + 130, + 346, + 482, + 418 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 346, + 482, + 418 + ], + "type": "inline_equation", + "content": "\\pmb{E}_{t}" + }, + { + "bbox": [ + 130, + 346, + 482, + 418 + ], + "type": "text", + "content": " is the residual pixel-wise learning rate at the " + }, + { + "bbox": [ + 130, + 346, + 482, + 418 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 346, + 482, + 418 + ], + "type": "text", + "content": "-th iteration, and the denominator in Eq. (5) computes the root of the sum of squares of all previous gradients of individual pixels. " + }, + { + "bbox": [ + 130, + 346, + 482, + 418 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 130, + 346, + 482, + 418 + ], + "type": "text", + "content": " is a small number to improve the numerical stability. The operation " + }, + { + "bbox": [ + 130, + 346, + 482, + 418 + ], + "type": "inline_equation", + "content": "\\odot" + }, + { + "bbox": [ + 130, + 346, + 482, + 418 + ], + "type": "text", + "content": " denotes element-wise multiplication. We note that, the accumulation of historical pixel-wise gradients in Eq. (5) can help improve iteration convergence by" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 427, + 481, + 475 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 130, + 427, + 481, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 427, + 481, + 450 + ], + "spans": [ + { + "bbox": [ + 130, + 427, + 481, + 450 + ], + "type": "text", + "content": "(1) scaling down the updating steps for regions with historically large gradients, which may be prone to overshooting;" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 451, + 481, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 451, + 481, + 475 + ], + "spans": [ + { + "bbox": [ + 130, + 451, + 481, + 475 + ], + "type": "text", + "content": "(2) scaling up the updating steps for regions with historically small gradients, which may benefit from a more aggressive update." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 130, + 484, + 482, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 484, + 482, + 594 + ], + "spans": [ + { + "bbox": [ + 130, + 484, + 482, + 594 + ], + "type": "text", + "content": "To unroll our self-adaptive gradient algorithm with a DUM, we deviate from the common practice of substituting " + }, + { + "bbox": [ + 130, + 484, + 482, + 594 + ], + "type": "inline_equation", + "content": "\\nabla R(\\boldsymbol{x}_t)" + }, + { + "bbox": [ + 130, + 484, + 482, + 594 + ], + "type": "text", + "content": " with a neural network in Eq. (4). There are compelling reasons for this. First, learning the " + }, + { + "bbox": [ + 130, + 484, + 482, + 594 + ], + "type": "inline_equation", + "content": "\\nabla R(\\boldsymbol{x}_t)" + }, + { + "bbox": [ + 130, + 484, + 482, + 594 + ], + "type": "text", + "content": " term in " + }, + { + "bbox": [ + 130, + 484, + 482, + 594 + ], + "type": "inline_equation", + "content": "\\boldsymbol{g}_t" + }, + { + "bbox": [ + 130, + 484, + 482, + 594 + ], + "type": "text", + "content": " (see Eq. (3)) can lead to numerical instability in " + }, + { + "bbox": [ + 130, + 484, + 482, + 594 + ], + "type": "inline_equation", + "content": "\\boldsymbol{E}_t" + }, + { + "bbox": [ + 130, + 484, + 482, + 594 + ], + "type": "text", + "content": " during the early stage of training. Second, the adaptive gradient strategy in Eq. (5), which is the basic variant utilized in AdaGrad (9), may not be the most suitable choice for a DUM in the context of MRI reconstruction. To leverage the robust nonlinear representation capabilities of neural networks, we improve upon the adaptive gradient algorithm by reformulating Eq. (4) as follows:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 171, + 601, + 481, + 632 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 171, + 601, + 481, + 632 + ], + "spans": [ + { + "bbox": [ + 171, + 601, + 481, + 632 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\boldsymbol {x} _ {t + 1} = \\boldsymbol {x} _ {t} - \\eta_ {t} \\boldsymbol {A} ^ {H} (\\boldsymbol {A} \\boldsymbol {x} _ {t} - \\boldsymbol {y}) - \\left(\\eta_ {t} \\nabla \\boldsymbol {R} (\\boldsymbol {x} _ {t}) + \\boldsymbol {E} _ {t} \\odot \\boldsymbol {g} _ {t}\\right) (6) \\\\ = \\boldsymbol {x} _ {t} - \\eta_ {t} \\boldsymbol {A} ^ {H} (\\boldsymbol {A} \\boldsymbol {x} _ {t} - \\boldsymbol {y}) + \\operatorname {C N N} _ {t} \\left(\\boldsymbol {x} _ {t}, \\boldsymbol {A} ^ {H} \\boldsymbol {A} \\boldsymbol {x} _ {t}, \\boldsymbol {A} ^ {H} \\boldsymbol {y}, \\boldsymbol {s} _ {t}\\right). (7) \\\\ \\end{array}", + "image_path": "3e74332f5f4d909263b3aab72bedf61fa3c2c59a2930be7bf61551bc4d88d2fb.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 640, + 482, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 640, + 482, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 640, + 482, + 667 + ], + "type": "text", + "content": "For the last term in Eq. (6), we reparametrize it with a CNN, as indicated in Eq. (7). This neural network takes as input the fundamental components from" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 219, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 219, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 219, + 101 + ], + "type": "text", + "content": "B. Xin et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 115, + 479, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 115, + 479, + 175 + ], + "spans": [ + { + "bbox": [ + 130, + 115, + 479, + 175 + ], + "type": "text", + "content": "which the term is constructed, namely, " + }, + { + "bbox": [ + 130, + 115, + 479, + 175 + ], + "type": "inline_equation", + "content": "\\boldsymbol{x}_t, \\boldsymbol{A}^H \\boldsymbol{A} \\boldsymbol{x}_t, \\boldsymbol{A}^H \\boldsymbol{y}" + }, + { + "bbox": [ + 130, + 115, + 479, + 175 + ], + "type": "text", + "content": ", and a learnable auxiliary memory " + }, + { + "bbox": [ + 130, + 115, + 479, + 175 + ], + "type": "inline_equation", + "content": "s_t" + }, + { + "bbox": [ + 130, + 115, + 479, + 175 + ], + "type": "text", + "content": ", which serves as a hidden state to track the historical gradient information implicitly. Note that " + }, + { + "bbox": [ + 130, + 115, + 479, + 175 + ], + "type": "inline_equation", + "content": "\\boldsymbol{A}^H \\boldsymbol{y}" + }, + { + "bbox": [ + 130, + 115, + 479, + 175 + ], + "type": "text", + "content": " yields " + }, + { + "bbox": [ + 130, + 115, + 479, + 175 + ], + "type": "inline_equation", + "content": "\\boldsymbol{x}_0" + }, + { + "bbox": [ + 130, + 115, + 479, + 175 + ], + "type": "text", + "content": ", which represents the image transformed from zero-filled " + }, + { + "bbox": [ + 130, + 115, + 479, + 175 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 115, + 479, + 175 + ], + "type": "text", + "content": "-space. For further discussion, please refer to Supplementary Section E." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 196, + 385, + 208 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 196, + 385, + 208 + ], + "spans": [ + { + "bbox": [ + 131, + 196, + 385, + 208 + ], + "type": "text", + "content": "3.3 Momentum-Accelerated Gradient Algorithm" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 217, + 480, + 254 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 217, + 480, + 254 + ], + "spans": [ + { + "bbox": [ + 130, + 217, + 480, + 254 + ], + "type": "text", + "content": "The momentum method [29] accelerates gradient descent in Eq. (2) by accumulating a velocity vector in directions of persistent reduction in the optimization objective across iterations:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 247, + 266, + 480, + 277 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 247, + 266, + 480, + 277 + ], + "spans": [ + { + "bbox": [ + 247, + 266, + 480, + 277 + ], + "type": "interline_equation", + "content": "\\boldsymbol {x} _ {t + 1} = \\boldsymbol {x} _ {t} - \\eta_ {t} \\boldsymbol {m} _ {t}, \\tag {8}", + "image_path": "ddcb434d702c4096f8e7d7d50f33f8068c77d14848a83dc655bd4171ec5a851f.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 254, + 281, + 480, + 293 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 254, + 281, + 480, + 293 + ], + "spans": [ + { + "bbox": [ + 254, + 281, + 480, + 293 + ], + "type": "interline_equation", + "content": "\\boldsymbol {m} _ {t} = \\mu \\boldsymbol {m} _ {t - 1} + (1 - \\mu) \\boldsymbol {g} _ {t}, \\tag {9}", + "image_path": "452bf4ce99cfb74c45c4854df1a5b7055b77b81785fefea8c69981f79542d5b3.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 302, + 482, + 457 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 302, + 482, + 457 + ], + "spans": [ + { + "bbox": [ + 130, + 302, + 482, + 457 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 302, + 482, + 457 + ], + "type": "inline_equation", + "content": "\\mu \\in [0,1]" + }, + { + "bbox": [ + 130, + 302, + 482, + 457 + ], + "type": "text", + "content": " is the momentum coefficient. The momentum " + }, + { + "bbox": [ + 130, + 302, + 482, + 457 + ], + "type": "inline_equation", + "content": "m_{t}" + }, + { + "bbox": [ + 130, + 302, + 482, + 457 + ], + "type": "text", + "content": " is known as the exponential moving average of historical gradients. It can benefit the optimization of the current iteration, thereby considerably accelerating convergence. In Sec. 3.2 we introduced the hidden state " + }, + { + "bbox": [ + 130, + 302, + 482, + 457 + ], + "type": "inline_equation", + "content": "s_t" + }, + { + "bbox": [ + 130, + 302, + 482, + 457 + ], + "type": "text", + "content": ", which is used to implicitly track the pixel-wise historical gradient information. Here, we propose explicit multi-stage and multi-level feature aggregation, to mimic the idea of momentum and increase the information flow across the iterations. For a typical multi-level encoder-decoder denoiser, such as Unet, in the " + }, + { + "bbox": [ + 130, + 302, + 482, + 457 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 302, + 482, + 457 + ], + "type": "text", + "content": "-th iteration, we insert a Momentum layer before the " + }, + { + "bbox": [ + 130, + 302, + 482, + 457 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 130, + 302, + 482, + 457 + ], + "type": "text", + "content": "-th level decoder layer. This layer concatenates features at the current stage and features from all previous iteration stages at the same level, followed by a " + }, + { + "bbox": [ + 130, + 302, + 482, + 457 + ], + "type": "inline_equation", + "content": "1\\times 1" + }, + { + "bbox": [ + 130, + 302, + 482, + 457 + ], + "type": "text", + "content": " convolution layer to reduce the dimensionality, and a Channel Attention Block (CAB) to adaptively fuse the multi-stage information, as described below:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 156, + 468, + 480, + 482 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 156, + 468, + 480, + 482 + ], + "spans": [ + { + "bbox": [ + 156, + 468, + 480, + 482 + ], + "type": "interline_equation", + "content": "\\operatorname {M o m e n t u m} \\left(\\boldsymbol {F} _ {0} ^ {m}, \\dots , \\boldsymbol {F} _ {t} ^ {m}\\right) = \\operatorname {C A B} \\left(\\operatorname {C o n v} _ {1 \\times 1} \\left(\\operatorname {C o n c a t} \\left(\\boldsymbol {F} _ {0} ^ {m}, \\dots , \\boldsymbol {F} _ {t} ^ {m}\\right)\\right)\\right), \\tag {10}", + "image_path": "8a0fff42f4e98f13cf9857215fa35a16197f66131d73a467d812f26bbe57f281.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 491, + 479, + 503 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 491, + 479, + 503 + ], + "spans": [ + { + "bbox": [ + 131, + 491, + 479, + 503 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 491, + 479, + 503 + ], + "type": "inline_equation", + "content": "\\pmb{F}_t^m" + }, + { + "bbox": [ + 131, + 491, + 479, + 503 + ], + "type": "text", + "content": " is the input feature of the " + }, + { + "bbox": [ + 131, + 491, + 479, + 503 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 131, + 491, + 479, + 503 + ], + "type": "text", + "content": "-th level decoder layer at the " + }, + { + "bbox": [ + 131, + 491, + 479, + 503 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 131, + 491, + 479, + 503 + ], + "type": "text", + "content": "-th iteration." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 523, + 285, + 536 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 523, + 285, + 536 + ], + "spans": [ + { + "bbox": [ + 131, + 523, + 285, + 536 + ], + "type": "text", + "content": "3.4 Adjacent Reconstruction" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": "Incorporating information from adjacent slices in multi-coil MRI reconstruction has been substantiated in prior work [12,28,45]. However, a significant limitation of adjacent slice reconstruction lies in its substantial memory consumption during the estimation of multi-coil sensitivity maps. This issue stems from the design of the sensitivity map estimation (SME) network in earlier work [40]. The network adopts a coil-by-coil estimation approach to accommodate a variable number of coils, increasing flexibility but leading to higher memory demands in the adjacent reconstruction setting (see Fig. 4(a)). To be more specific, for " + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": "-space data with " + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": " coils and " + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "inline_equation", + "content": "2a + 1" + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": " adjacent slices, the total number of forward passes required for the SME process is calculated as " + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "inline_equation", + "content": "(2a + 1)N" + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": ". Typically, the" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 162, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 162, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 162, + 91, + 448, + 102 + ], + "type": "text", + "content": "Rethinking Deep Unrolled Model for Accelerated MRI Reconstruction" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 114, + 239, + 177 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 239, + 177 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 239, + 177 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 239, + 177 + ], + "type": "image", + "image_path": "fe41e60de2c1d27f5b953c5ce27269df8a37cbd0d653b8629ab4995b3f832f80.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 170, + 178, + 202, + 188 + ], + "lines": [ + { + "bbox": [ + 170, + 178, + 202, + 188 + ], + "spans": [ + { + "bbox": [ + 170, + 178, + 202, + 188 + ], + "type": "text", + "content": "(a) time" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 254, + 114, + 359, + 177 + ], + "blocks": [ + { + "bbox": [ + 254, + 114, + 359, + 177 + ], + "lines": [ + { + "bbox": [ + 254, + 114, + 359, + 177 + ], + "spans": [ + { + "bbox": [ + 254, + 114, + 359, + 177 + ], + "type": "image", + "image_path": "b9bb7b06238ea4f60be30ca52cb58101901d4c084f9186904a933bb117f8e79f.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 285, + 178, + 329, + 187 + ], + "lines": [ + { + "bbox": [ + 285, + 178, + 329, + 187 + ], + "spans": [ + { + "bbox": [ + 285, + 178, + 329, + 187 + ], + "type": "text", + "content": "(b) contrast" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 375, + 114, + 479, + 177 + ], + "blocks": [ + { + "bbox": [ + 375, + 114, + 479, + 177 + ], + "lines": [ + { + "bbox": [ + 375, + 114, + 479, + 177 + ], + "spans": [ + { + "bbox": [ + 375, + 114, + 479, + 177 + ], + "type": "image", + "image_path": "dc95bd762fc55127bc99165074ca24869f845aad1874e00ca6ed85fcbc380ef3.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 413, + 178, + 443, + 187 + ], + "lines": [ + { + "bbox": [ + 413, + 178, + 443, + 187 + ], + "spans": [ + { + "bbox": [ + 413, + 178, + 443, + 187 + ], + "type": "text", + "content": "(c) slice" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 133, + 241, + 301, + 299 + ], + "blocks": [ + { + "bbox": [ + 133, + 241, + 301, + 299 + ], + "lines": [ + { + "bbox": [ + 133, + 241, + 301, + 299 + ], + "spans": [ + { + "bbox": [ + 133, + 241, + 301, + 299 + ], + "type": "image", + "image_path": "042a2753b14a7f22137e73b6ec121a1f9e3ed1e552ced03aa722ea922202d833.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 312, + 240, + 480, + 299 + ], + "blocks": [ + { + "bbox": [ + 131, + 197, + 482, + 231 + ], + "lines": [ + { + "bbox": [ + 131, + 197, + 482, + 231 + ], + "spans": [ + { + "bbox": [ + 131, + 197, + 482, + 231 + ], + "type": "text", + "content": "Fig. 3: The sensitivity maps exhibit variations across (a) time, (b) contrast and (c) slice dimension. First row: original images, second row: magnitude of differences between the sensitivity map of the central image and each of those of its adjacent images." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 312, + 240, + 480, + 299 + ], + "lines": [ + { + "bbox": [ + 312, + 240, + 480, + 299 + ], + "spans": [ + { + "bbox": [ + 312, + 240, + 480, + 299 + ], + "type": "image", + "image_path": "d76db55f95f6809fbacded300b1ffd80f6685949b2743145013b14de5427cf57.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 133, + 305, + 301, + 365 + ], + "blocks": [ + { + "bbox": [ + 133, + 305, + 301, + 365 + ], + "lines": [ + { + "bbox": [ + 133, + 305, + 301, + 365 + ], + "spans": [ + { + "bbox": [ + 133, + 305, + 301, + 365 + ], + "type": "image", + "image_path": "d250f998439938584824d2ea82078788f8b04989a7593502658df666ea70d4c1.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 373, + 482, + 440 + ], + "lines": [ + { + "bbox": [ + 130, + 373, + 482, + 440 + ], + "spans": [ + { + "bbox": [ + 130, + 373, + 482, + 440 + ], + "type": "text", + "content": "Fig. 4: Concept illustration of different strategies for multi-coil sensitivity map estimation (SME). The main difference between the four strategies lies in the Auto-Calibration-Signal (ACS) information sharing between coils and adjacent MR image slices. Each single coil sensitivity map gets ACS information from (a) a single coil and a single image slice; (b) multiple coils but a single image slice; (c) a single coil but multiple adjacent image slices; (d) multiple coils and multiple adjacent image slices." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 312, + 305, + 480, + 365 + ], + "blocks": [ + { + "bbox": [ + 312, + 305, + 480, + 365 + ], + "lines": [ + { + "bbox": [ + 312, + 305, + 480, + 365 + ], + "spans": [ + { + "bbox": [ + 312, + 305, + 480, + 365 + ], + "type": "image", + "image_path": "23ea4db9164c7d3b1378b939654ffbef2b78f5a13940402bbafda1bf7e1e8ac2.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 464, + 482, + 499 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 464, + 482, + 499 + ], + "spans": [ + { + "bbox": [ + 130, + 464, + 482, + 499 + ], + "type": "text", + "content": "value of " + }, + { + "bbox": [ + 130, + 464, + 482, + 499 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 464, + 482, + 499 + ], + "type": "text", + "content": " ranges from 8 to 34 and " + }, + { + "bbox": [ + 130, + 464, + 482, + 499 + ], + "type": "inline_equation", + "content": "a" + }, + { + "bbox": [ + 130, + 464, + 482, + 499 + ], + "type": "text", + "content": " is set to 2. We propose to reduce the number of forward passes, by utilizing the correlations between adjacent sensitivity maps." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 500, + 482, + 646 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 500, + 482, + 646 + ], + "spans": [ + { + "bbox": [ + 130, + 500, + 482, + 646 + ], + "type": "text", + "content": "A sensitivity map is subject to variation, based on the position of the imaging subject and changes over time due to the subject motion. Additionally, in multicontest contrast MRI acquisition, variations in acquisition parameters, such as inversion time (TI) and repetition time (TR), can indirectly influence the sensitivity map estimation through the acquired " + }, + { + "bbox": [ + 130, + 500, + 482, + 646 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 500, + 482, + 646 + ], + "type": "text", + "content": "-space data, as shown in Fig. ③. Consequently, the estimation of adjacent sensitivity maps is a spatially, temporally, and contrastively correlated process. We therefore propose to estimate sensitivity maps by utilizing information from spatially, temporally, and contrastively adjacent slices. As shown in Fig. ④(c), we simultaneously estimate a single coil sensitivity map at all adjacent " + }, + { + "bbox": [ + 130, + 500, + 482, + 646 + ], + "type": "inline_equation", + "content": "2a + 1" + }, + { + "bbox": [ + 130, + 500, + 482, + 646 + ], + "type": "text", + "content": " slices using information across these slices. In this way, we can reduce the number of forward passes from " + }, + { + "bbox": [ + 130, + 500, + 482, + 646 + ], + "type": "inline_equation", + "content": "(2a + 1)N" + }, + { + "bbox": [ + 130, + 500, + 482, + 646 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 130, + 500, + 482, + 646 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 500, + 482, + 646 + ], + "type": "text", + "content": ". Our estimation of the " + }, + { + "bbox": [ + 130, + 500, + 482, + 646 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 130, + 500, + 482, + 646 + ], + "type": "text", + "content": "-th sensitivity map set " + }, + { + "bbox": [ + 130, + 500, + 482, + 646 + ], + "type": "inline_equation", + "content": "S_{adj}^{n}" + }, + { + "bbox": [ + 130, + 500, + 482, + 646 + ], + "type": "text", + "content": " can be described as:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 243, + 653, + 481, + 667 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 243, + 653, + 481, + 667 + ], + "spans": [ + { + "bbox": [ + 243, + 653, + 481, + 667 + ], + "type": "interline_equation", + "content": "\\boldsymbol {S} _ {a d j} ^ {n} = \\mathbf {S M E} (\\mathbf {A C S} \\left(\\boldsymbol {x} _ {0, a d j} ^ {n}\\right)), \\tag {11}", + "image_path": "87b9bfcf656708695a6a9968f237dcdae89b3f5a6a481d81d4c47b92e66cf4f6.jpg" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 219, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 219, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 219, + 101 + ], + "type": "text", + "content": "B. Xin et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 167 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 167 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 116, + 482, + 167 + ], + "type": "inline_equation", + "content": "S_{adj}^{n} = [S_{a - c}^{n},\\dots,S_{c}^{n},\\dots,S_{a + c}^{n}]" + }, + { + "bbox": [ + 130, + 116, + 482, + 167 + ], + "type": "text", + "content": " is an adjacent sensitivity map set which is the concatenation of central slice sensitivity map " + }, + { + "bbox": [ + 130, + 116, + 482, + 167 + ], + "type": "inline_equation", + "content": "S_{c}^{n}" + }, + { + "bbox": [ + 130, + 116, + 482, + 167 + ], + "type": "text", + "content": " with its 2a-adjacent slice sensitivity maps of the " + }, + { + "bbox": [ + 130, + 116, + 482, + 167 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 130, + 116, + 482, + 167 + ], + "type": "text", + "content": "-th coil. " + }, + { + "bbox": [ + 130, + 116, + 482, + 167 + ], + "type": "inline_equation", + "content": "\\mathbf{ACS}(\\pmb{x}_{0,adj}^{n})" + }, + { + "bbox": [ + 130, + 116, + 482, + 167 + ], + "type": "text", + "content": " is the " + }, + { + "bbox": [ + 130, + 116, + 482, + 167 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 130, + 116, + 482, + 167 + ], + "type": "text", + "content": "-th coil Auto-Calibration Signal (ACS) within all adjacent " + }, + { + "bbox": [ + 130, + 116, + 482, + 167 + ], + "type": "inline_equation", + "content": "2a + 1" + }, + { + "bbox": [ + 130, + 116, + 482, + 167 + ], + "type": "text", + "content": " slices." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 181, + 303, + 194 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 181, + 303, + 194 + ], + "spans": [ + { + "bbox": [ + 131, + 181, + 303, + 194 + ], + "type": "text", + "content": "4 Experiments and Results" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 202, + 482, + 288 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 202, + 482, + 288 + ], + "spans": [ + { + "bbox": [ + 130, + 202, + 482, + 288 + ], + "type": "text", + "content": "In this section, we present the experimental results of our proposed approach. We showcase the reconstruction performance of our model across multiple publicly available datasets, including Calgary-Campinas brain [39], CMRxRecon cardiac [44], and fastMRI multi-coil knee [49] datasets. These datasets encompass a diverse range of anatomies, contrasts, views, slices and motion states. For details about the datasets, please refer to Supplementary Section A. For information on the experimental setup for each dataset, see Supplementary Section C." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 301, + 326, + 313 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 301, + 326, + 313 + ], + "spans": [ + { + "bbox": [ + 131, + 301, + 326, + 313 + ], + "type": "text", + "content": "4.1 Calgary-Campinas Brain Results" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 318, + 482, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 318, + 482, + 413 + ], + "spans": [ + { + "bbox": [ + 130, + 318, + 482, + 413 + ], + "type": "text", + "content": "Tab. 1 presents the quantitative results for " + }, + { + "bbox": [ + 130, + 318, + 482, + 413 + ], + "type": "inline_equation", + "content": "\\times 5" + }, + { + "bbox": [ + 130, + 318, + 482, + 413 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 318, + 482, + 413 + ], + "type": "inline_equation", + "content": "\\times 10" + }, + { + "bbox": [ + 130, + 318, + 482, + 413 + ], + "type": "text", + "content": " acceleration factors on the Calgary-Campinas brain test set. Our method is compared with two state-of-the-art (SOTA) methods: Recurrent-VarNet [46] and PromptMR [45]. Our proposed method significantly outperforms these benchmarks, demonstrating its outstanding performance. Qualitative results under " + }, + { + "bbox": [ + 130, + 318, + 482, + 413 + ], + "type": "inline_equation", + "content": "\\times 10" + }, + { + "bbox": [ + 130, + 318, + 482, + 413 + ], + "type": "text", + "content": " acceleration are shown in Fig. 5. Interestingly, while PromptMR achieves significantly higher PSNR and SSIM values compared to Recurrent-VarNet, it exhibits less accurate reconstruction in the area highlighted by the red box." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 414, + 482, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 414, + 482, + 521 + ], + "spans": [ + { + "bbox": [ + 130, + 414, + 482, + 521 + ], + "type": "text", + "content": "The highly undersampled " + }, + { + "bbox": [ + 130, + 414, + 482, + 521 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 414, + 482, + 521 + ], + "type": "text", + "content": "-space data makes the ill-posed reconstruction challenging at aliasing artifacts removal. While leveraging large-scale training data to learn the image priors can improve the reconstruction accuracy, we note that Recurrent-VarNet learns a weaker prior, resulting in reconstructions that appear blurry. Conversely, PromptMR adopts a strong but overly biased prior, leading to sharper reconstructions but introducing incorrect structures. Our proposed method, in contrast, successfully learns the appropriate prior, achieving a reconstruction that closely mirrors the ground truth in terms of structural accuracy." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 536, + 310, + 547 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 536, + 310, + 547 + ], + "spans": [ + { + "bbox": [ + 131, + 536, + 310, + 547 + ], + "type": "text", + "content": "4.2 CMRxRecon Cardiac Results" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 552, + 482, + 649 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 552, + 482, + 649 + ], + "spans": [ + { + "bbox": [ + 130, + 552, + 482, + 649 + ], + "type": "text", + "content": "We compare our method with PromptMR, the winning method of the MIC-CAI2023 CMRxRecon challenge, for " + }, + { + "bbox": [ + 130, + 552, + 482, + 649 + ], + "type": "inline_equation", + "content": "\\times 4" + }, + { + "bbox": [ + 130, + 552, + 482, + 649 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 552, + 482, + 649 + ], + "type": "inline_equation", + "content": "\\times 8" + }, + { + "bbox": [ + 130, + 552, + 482, + 649 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 552, + 482, + 649 + ], + "type": "inline_equation", + "content": "\\times 10" + }, + { + "bbox": [ + 130, + 552, + 482, + 649 + ], + "type": "text", + "content": " acceleration factors on the CMRxRecon cardiac test set. In Tab. ②, our method exhibits superior performance to PromptMR across different views and contrasts. Qualitative results under " + }, + { + "bbox": [ + 130, + 552, + 482, + 649 + ], + "type": "inline_equation", + "content": "\\times 10" + }, + { + "bbox": [ + 130, + 552, + 482, + 649 + ], + "type": "text", + "content": " acceleration are shown in Fig. ⑥. Our method reconstructs fine details of the interventricular septum, highlighted in the red box, with significantly higher clarity than the PromptMR method, which is severely compromised by aliasing artifacts." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 162, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 162, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 162, + 91, + 447, + 102 + ], + "type": "text", + "content": "Rethinking Deep Unrolled Model for Accelerated MRI Reconstruction" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 133, + 653, + 370, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 653, + 370, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 653, + 370, + 665 + ], + "type": "text", + "content": "1 It is usually a part of the acquired central " + }, + { + "bbox": [ + 133, + 653, + 370, + 665 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 133, + 653, + 370, + 665 + ], + "type": "text", + "content": "-space data." + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 187, + 168, + 427, + 245 + ], + "blocks": [ + { + "bbox": [ + 130, + 114, + 482, + 159 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 482, + 159 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 482, + 159 + ], + "type": "text", + "content": "Table 1: Quantitative comparison of PSNR/SSIM (mean ± std) of different MRI reconstruction methods on the Calgary-Campinas brain dataset under " + }, + { + "bbox": [ + 130, + 114, + 482, + 159 + ], + "type": "inline_equation", + "content": "\\times 5" + }, + { + "bbox": [ + 130, + 114, + 482, + 159 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 114, + 482, + 159 + ], + "type": "inline_equation", + "content": "\\times 10" + }, + { + "bbox": [ + 130, + 114, + 482, + 159 + ], + "type": "text", + "content": " acceleration. The best and second best performance are in red and blue colors, respectively." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 187, + 168, + 427, + 245 + ], + "lines": [ + { + "bbox": [ + 187, + 168, + 427, + 245 + ], + "spans": [ + { + "bbox": [ + 187, + 168, + 427, + 245 + ], + "type": "table", + "html": "
MethodAcc = 5×Acc = 10×
PSNR(dB)SSIM(%)PSNR(dB)SSIM(%)
Zero-filled25.27±1.3173.75±3.9324.35±1.2169.04±3.67
Recurrent-Varnet36.27±1.7294.37±1.3333.27±2.1291.47±2.23
PromptMR36.83±1.6894.85±1.2634.10±2.1192.69±1.92
Ours37.24±1.5795.15±1.2134.74±1.9793.38±1.65
", + "image_path": "6eadf08b59fddd3c93a479304e155e70ae12c72cabac101c5c2bb02a72c24e20.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 184, + 254, + 429, + 378 + ], + "blocks": [ + { + "bbox": [ + 184, + 254, + 429, + 378 + ], + "lines": [ + { + "bbox": [ + 184, + 254, + 429, + 378 + ], + "spans": [ + { + "bbox": [ + 184, + 254, + 429, + 378 + ], + "type": "image", + "image_path": "729c05983d97094615657e658bef127cb0dd05a3a6997eb5dd391418286572d5.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 386, + 482, + 420 + ], + "lines": [ + { + "bbox": [ + 130, + 386, + 482, + 420 + ], + "spans": [ + { + "bbox": [ + 130, + 386, + 482, + 420 + ], + "type": "text", + "content": "Fig. 5: The reconstruction results and SSIM error maps of various methods for the Calgary-Campinas brain imaging under " + }, + { + "bbox": [ + 130, + 386, + 482, + 420 + ], + "type": "inline_equation", + "content": "\\times 10" + }, + { + "bbox": [ + 130, + 386, + 482, + 420 + ], + "type": "text", + "content": " acceleration factor. The red boxes highlight the differences in the recovery of white and gray matter structures." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 441, + 277, + 453 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 441, + 277, + 453 + ], + "spans": [ + { + "bbox": [ + 131, + 441, + 277, + 453 + ], + "type": "text", + "content": "4.3 FastMRI Knee Results" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 459, + 482, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 459, + 482, + 602 + ], + "spans": [ + { + "bbox": [ + 130, + 459, + 482, + 602 + ], + "type": "text", + "content": "We evaluated the PSNR and SSIM at acceleration factors of " + }, + { + "bbox": [ + 130, + 459, + 482, + 602 + ], + "type": "inline_equation", + "content": "\\times 4" + }, + { + "bbox": [ + 130, + 459, + 482, + 602 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 459, + 482, + 602 + ], + "type": "inline_equation", + "content": "\\times 8" + }, + { + "bbox": [ + 130, + 459, + 482, + 602 + ], + "type": "text", + "content": " on the fastMRI multi-coil knee test set. Our method is compared with several methods, including SOTA methods HUMUS-Net-L [12] and PromptMR. In the table, 'PD' and 'PDFS' denote proton density-weighted images without and with fat suppression, respectively. Tab. 3 demonstrates that our approach achieves the best performance across all other competitive methods on this large-scale dataset. Qualitative results under " + }, + { + "bbox": [ + 130, + 459, + 482, + 602 + ], + "type": "inline_equation", + "content": "\\times 8" + }, + { + "bbox": [ + 130, + 459, + 482, + 602 + ], + "type": "text", + "content": " acceleration are shown in Fig. 7. Our approach achieves a more detailed reconstruction of the meniscus region as shown in the red box, closely approximating the ground truth. In this region, meniscus tears were identified by radiologists from the fastMRI+ dataset [55]. The qualitative comparison suggests that our approach is more robust in reconstructing abnormal knee regions for highly accelerated acquisition." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 618, + 250, + 632 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 618, + 250, + 632 + ], + "spans": [ + { + "bbox": [ + 131, + 618, + 250, + 632 + ], + "type": "text", + "content": "5 Abalation study" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "content": "In this section, we justify our design choices through ablation studies on the brain and cardiac datasets." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "text", + "content": "B. Xin et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 135, + 168, + 479, + 290 + ], + "blocks": [ + { + "bbox": [ + 131, + 114, + 482, + 159 + ], + "lines": [ + { + "bbox": [ + 131, + 114, + 482, + 159 + ], + "spans": [ + { + "bbox": [ + 131, + 114, + 482, + 159 + ], + "type": "text", + "content": "Table 2: Quantitative comparison of PSNR/SSIM (mean ± std) of different MRI reconstruction methods on the CMRxRecon cardiac dataset under " + }, + { + "bbox": [ + 131, + 114, + 482, + 159 + ], + "type": "inline_equation", + "content": "\\times 4" + }, + { + "bbox": [ + 131, + 114, + 482, + 159 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 131, + 114, + 482, + 159 + ], + "type": "inline_equation", + "content": "\\times 8" + }, + { + "bbox": [ + 131, + 114, + 482, + 159 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 114, + 482, + 159 + ], + "type": "inline_equation", + "content": "\\times 10" + }, + { + "bbox": [ + 131, + 114, + 482, + 159 + ], + "type": "text", + "content": " acceleration. The best and second best performance are in red and blue colors, respectively." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 135, + 168, + 479, + 290 + ], + "lines": [ + { + "bbox": [ + 135, + 168, + 479, + 290 + ], + "spans": [ + { + "bbox": [ + 135, + 168, + 479, + 290 + ], + "type": "table", + "html": "
AccMethodCineMapping
SAXLAXT1wT2w
PSNR(dB)SSIM(%)PSNR(dB)SSIM(%)PSNR(dB)SSIM(%)PSNR(dB)SSIM(%)
4×Zero-filled PromptMR Ours26.20±1.5472.99±5.1425.11±1.6269.87±4.7324.23±1.1667.27±3.8424.97±1.1077.16±2.91
45.78±1.8998.78±0.4345.58±1.7298.72±0.3746.49±2.2398.93±0.5742.02±1.9398.05±0.72
46.26±1.7698.87±0.3446.13±1.7098.83±0.3447.45±2.1399.09±0.4242.68±1.8498.25±0.59
8×Zero-filled PromptMR Ours25.18±1.6770.21±5.8724.34±1.6667.67±5.8023.41±1.2564.12±4.4324.21±1.1274.84±3.07
40.65±1.5296.97±0.6839.64±1.8796.36±1.0440.94±1.6997.28±0.8538.24±1.8296.53±1.09
41.44±1.4897.32±0.5840.49±1.8396.76±0.9442.08±1.6397.70±0.6739.23±1.7196.98±0.89
10×Zero-filled PromptMR Ours24.83±1.6769.35±5.9224.13±1.6766.89±5.5823.16±1.2764.14±4.4724.25±1.1476.08±2.91
39.18±1.5096.15±0.8038.28±1.6295.60±1.0738.99±1.5896.61±0.9937.21±1.7696.22±1.16
39.99±1.4996.58±0.7239.13±1.6696.05±0.9940.37±1.5897.19±0.7938.22±1.7096.70±0.96
", + "image_path": "bcd8fb134867bb132425a97366919793016bf31c974b0c131f943f6c398e72e3.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 133, + 300, + 481, + 429 + ], + "blocks": [ + { + "bbox": [ + 133, + 300, + 481, + 429 + ], + "lines": [ + { + "bbox": [ + 133, + 300, + 481, + 429 + ], + "spans": [ + { + "bbox": [ + 133, + 300, + 481, + 429 + ], + "type": "image", + "image_path": "eaae49564f24e3170a75b5299d01398aac83fd9d2241e3866bd1dc246c8868da.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 437, + 482, + 470 + ], + "lines": [ + { + "bbox": [ + 131, + 437, + 482, + 470 + ], + "spans": [ + { + "bbox": [ + 131, + 437, + 482, + 470 + ], + "type": "text", + "content": "Fig. 6: The reconstruction results and SSIM error maps of various methods for the CM-RxRecon cardiac T1-weighted imaging under " + }, + { + "bbox": [ + 131, + 437, + 482, + 470 + ], + "type": "inline_equation", + "content": "\\times 10" + }, + { + "bbox": [ + 131, + 437, + 482, + 470 + ], + "type": "text", + "content": " acceleration factor. The red boxes highlight the differences in the recovery of the interventricular septum." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 135, + 523, + 479, + 634 + ], + "blocks": [ + { + "bbox": [ + 131, + 479, + 482, + 514 + ], + "lines": [ + { + "bbox": [ + 131, + 479, + 482, + 514 + ], + "spans": [ + { + "bbox": [ + 131, + 479, + 482, + 514 + ], + "type": "text", + "content": "Table 3: Quantitative comparison of PSNR/SSIM (mean ± std) of different MRI reconstruction methods on the fastMRI knee dataset under " + }, + { + "bbox": [ + 131, + 479, + 482, + 514 + ], + "type": "inline_equation", + "content": "\\times 4" + }, + { + "bbox": [ + 131, + 479, + 482, + 514 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 479, + 482, + 514 + ], + "type": "inline_equation", + "content": "\\times 8" + }, + { + "bbox": [ + 131, + 479, + 482, + 514 + ], + "type": "text", + "content": " acceleration. The best and second best performance are in red and blue colors, respectively." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 135, + 523, + 479, + 634 + ], + "lines": [ + { + "bbox": [ + 135, + 523, + 479, + 634 + ], + "spans": [ + { + "bbox": [ + 135, + 523, + 479, + 634 + ], + "type": "table", + "html": "
MethodAcc = 4×Acc = 8×
PDPDSPDPDS
|PSNR(dB)|SSIM(%)|PSNR(dB)|SSIM(%)|PSNR(dB)|SSIM(%)|PSNR(dB)|SSIM(%)
Zero-filled30.86±1.7380.65±3.7631.00±3.3378.48±6.1627.70±1.8474.12±4.9427.97±2.0268.78±5.58
Unet 3537.27±1.7692.11±2.7237.07±2.4788.03±5.0435.41±2.1490.36±3.1334.84±1.5983.71±5.04
i-RIM 3339.61±2.3394.16±2.8738.23±3.1389.27±4.9637.91±2.4793.22±2.8535.84±1.7185.27±4.93
E2E-VarNet 4040.09±2.3294.56±2.6438.43±3.2489.50±4.9538.72±2.5093.93±2.6836.16±1.7685.70±4.94
HUMUS-Net 12----38.72±2.5094.07±2.6835.98±1.7585.77±4.93
HUMUS-Net-L 12----39.03±2.4794.18±2.6136.19±1.7985.84±4.92
PromptMR 4540.58±2.5994.87±2.7538.59±3.3389.75±4.9439.48±2.5494.51±2.5436.42±1.7986.10±4.93
Ours40.85±2.7495.00±2.8438.71±3.3789.85±4.9240.09±2.5894.90±2.4536.49±1.8886.29±4.95
", + "image_path": "aff27bd971ba7eb0ff06a34733d2af9b8b016237ea37ce40253ba6c32db62d14.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 162, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 162, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 162, + 91, + 447, + 102 + ], + "type": "text", + "content": "Rethinking Deep Unrolled Model for Accelerated MRI Reconstruction" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 114, + 481, + 208 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 481, + 208 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 481, + 208 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 481, + 208 + ], + "type": "image", + "image_path": "a39cd262337756e98b622481f41381c2830f008b46985726e4dee9f87681ca91.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 213, + 482, + 269 + ], + "lines": [ + { + "bbox": [ + 130, + 213, + 482, + 269 + ], + "spans": [ + { + "bbox": [ + 130, + 213, + 482, + 269 + ], + "type": "text", + "content": "Fig. 7: The reconstruction results and SSIM error maps of various methods for the fastMRI multi-coil knee proton density (PD) imaging under " + }, + { + "bbox": [ + 130, + 213, + 482, + 269 + ], + "type": "inline_equation", + "content": "\\times 8" + }, + { + "bbox": [ + 130, + 213, + 482, + 269 + ], + "type": "text", + "content": " acceleration. The red boxes highlight the differences in the recovery of the meniscus region. The red box in the reference image was annotated by radiologists from the fastMRI+ dataset [55] to indicate the knee abnormality of meniscus tears." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 187, + 316, + 427, + 418 + ], + "blocks": [ + { + "bbox": [ + 130, + 285, + 482, + 308 + ], + "lines": [ + { + "bbox": [ + 130, + 285, + 482, + 308 + ], + "spans": [ + { + "bbox": [ + 130, + 285, + 482, + 308 + ], + "type": "text", + "content": "Table 4: Abalation of different input combinations of CNN on the Calgary-Campinas brain dataset under " + }, + { + "bbox": [ + 130, + 285, + 482, + 308 + ], + "type": "inline_equation", + "content": "\\times 10" + }, + { + "bbox": [ + 130, + 285, + 482, + 308 + ], + "type": "text", + "content": " acceleration." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 187, + 316, + 427, + 418 + ], + "lines": [ + { + "bbox": [ + 187, + 316, + 427, + 418 + ], + "spans": [ + { + "bbox": [ + 187, + 316, + 427, + 418 + ], + "type": "table", + "html": "
InputBaseline------Ours
xt✓✓✓✓✓✓✓✓
AHx✓✓✓✓✓✓✓✓
AHy✓✓✓✓✓✓✓✓
st✓✓✓✓✓✓✓✓
PSNR(dB)35.76+0.00+0.01+0.09+0.08+0.31+0.36+0.36
SSIM(%)94.33+0.00-0.01+0.06+0.05+0.25+0.27+0.28
", + "image_path": "b9151a1c66962dfe5ee097262005b7ea06b4ab0892f00958fc84b98478645a56.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 444, + 289, + 456 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 444, + 289, + 456 + ], + "spans": [ + { + "bbox": [ + 132, + 444, + 289, + 456 + ], + "type": "text", + "content": "5.1 Input Variations to CNN" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 473, + 482, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 473, + 482, + 668 + ], + "spans": [ + { + "bbox": [ + 130, + 473, + 482, + 668 + ], + "type": "text", + "content": "We conducted an ablation study for different input combinations to CNN, to validate the effectiveness of our proposed method. Results are shown in Tab. 4. The baseline network in each iteration only accepts the output from the last iteration " + }, + { + "bbox": [ + 130, + 473, + 482, + 668 + ], + "type": "inline_equation", + "content": "\\pmb{x}_t" + }, + { + "bbox": [ + 130, + 473, + 482, + 668 + ], + "type": "text", + "content": " as input, which is the setting for most unrolled models, such as E2E-VartNet and PromptMR. Our proposed method takes additional input, " + }, + { + "bbox": [ + 130, + 473, + 482, + 668 + ], + "type": "inline_equation", + "content": "A^H A\\pmb{x}_t" + }, + { + "bbox": [ + 130, + 473, + 482, + 668 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 473, + 482, + 668 + ], + "type": "inline_equation", + "content": "A^H\\pmb{y}" + }, + { + "bbox": [ + 130, + 473, + 482, + 668 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 473, + 482, + 668 + ], + "type": "inline_equation", + "content": "s_t" + }, + { + "bbox": [ + 130, + 473, + 482, + 668 + ], + "type": "text", + "content": ", which outperforms the baseline by a large margin. More specifically, adding " + }, + { + "bbox": [ + 130, + 473, + 482, + 668 + ], + "type": "inline_equation", + "content": "A^H A\\pmb{x}_t" + }, + { + "bbox": [ + 130, + 473, + 482, + 668 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 130, + 473, + 482, + 668 + ], + "type": "inline_equation", + "content": "A^H\\pmb{y}" + }, + { + "bbox": [ + 130, + 473, + 482, + 668 + ], + "type": "text", + "content": " solely to the input has no impact on the performance. Adding " + }, + { + "bbox": [ + 130, + 473, + 482, + 668 + ], + "type": "inline_equation", + "content": "s_t" + }, + { + "bbox": [ + 130, + 473, + 482, + 668 + ], + "type": "text", + "content": " can improve the performance, as it serves as additional memory to track the progression of the reconstruction process. Adding " + }, + { + "bbox": [ + 130, + 473, + 482, + 668 + ], + "type": "inline_equation", + "content": "A^H A\\pmb{x}_t" + }, + { + "bbox": [ + 130, + 473, + 482, + 668 + ], + "type": "text", + "content": " together with " + }, + { + "bbox": [ + 130, + 473, + 482, + 668 + ], + "type": "inline_equation", + "content": "A^H\\pmb{y}" + }, + { + "bbox": [ + 130, + 473, + 482, + 668 + ], + "type": "text", + "content": " greatly improves the gains. We assume the reason is that the residual connection of two terms into the network can provide soft data consistency information to the model. Adding " + }, + { + "bbox": [ + 130, + 473, + 482, + 668 + ], + "type": "inline_equation", + "content": "A^H A\\pmb{x}_t" + }, + { + "bbox": [ + 130, + 473, + 482, + 668 + ], + "type": "text", + "content": " together with " + }, + { + "bbox": [ + 130, + 473, + 482, + 668 + ], + "type": "inline_equation", + "content": "s_t" + }, + { + "bbox": [ + 130, + 473, + 482, + 668 + ], + "type": "text", + "content": " performs the closest to our final design. It seems the missing of " + }, + { + "bbox": [ + 130, + 473, + 482, + 668 + ], + "type": "inline_equation", + "content": "A^H\\pmb{y}" + }, + { + "bbox": [ + 130, + 473, + 482, + 668 + ], + "type": "text", + "content": " has little impact, as it is a constant value and its information may be compactly represented in the " + }, + { + "bbox": [ + 130, + 473, + 482, + 668 + ], + "type": "inline_equation", + "content": "s_t" + }, + { + "bbox": [ + 130, + 473, + 482, + 668 + ], + "type": "text", + "content": ". This assumption is implied by the evidence that the result of adding " + }, + { + "bbox": [ + 130, + 473, + 482, + 668 + ], + "type": "inline_equation", + "content": "A^H\\pmb{y}" + }, + { + "bbox": [ + 130, + 473, + 482, + 668 + ], + "type": "text", + "content": " together with " + }, + { + "bbox": [ + 130, + 473, + 482, + 668 + ], + "type": "inline_equation", + "content": "s_t" + }, + { + "bbox": [ + 130, + 473, + 482, + 668 + ], + "type": "text", + "content": " into the network does not perform better than adding " + }, + { + "bbox": [ + 130, + 473, + 482, + 668 + ], + "type": "inline_equation", + "content": "s_t" + }, + { + "bbox": [ + 130, + 473, + 482, + 668 + ], + "type": "text", + "content": " alone." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "text", + "content": "B. Xin et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 222, + 157, + 392, + 222 + ], + "blocks": [ + { + "bbox": [ + 130, + 114, + 482, + 148 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 482, + 148 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 482, + 148 + ], + "type": "text", + "content": "Table 5: Quantitative comparison of PSNR/SSIM using different numbers of previous stages' outputs for multi-stage feature fusion on the Calgary-Campinas brain dataset under " + }, + { + "bbox": [ + 130, + 114, + 482, + 148 + ], + "type": "inline_equation", + "content": "\\times 10" + }, + { + "bbox": [ + 130, + 114, + 482, + 148 + ], + "type": "text", + "content": " acceleration." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 222, + 157, + 392, + 222 + ], + "lines": [ + { + "bbox": [ + 222, + 157, + 392, + 222 + ], + "spans": [ + { + "bbox": [ + 222, + 157, + 392, + 222 + ], + "type": "table", + "html": "
Input# of previous iterations
015911
PSNR(dB)35.88+0.02+0.04+0.04+0.05
SSIM(%)94.43+0.02+0.05+0.06+0.07
", + "image_path": "b38d297ac79b3d067b5996e839fb1a231f0841442e9eadb02f9ffb070871bc74.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 243, + 312, + 255 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 243, + 312, + 255 + ], + "spans": [ + { + "bbox": [ + 132, + 243, + 312, + 255 + ], + "type": "text", + "content": "5.2 Multi-Stage Information Flow" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 260, + 482, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 260, + 482, + 319 + ], + "spans": [ + { + "bbox": [ + 130, + 260, + 482, + 319 + ], + "type": "text", + "content": "In Tab. 5 we investigate the effect of incorporating different numbers of previous iteration results into the current iteration for multi-stage information fusion. As the number increases, the reconstruction performance increases gradually. This implies the benefit of incorporating the information from previous iterations in DUM-based MRI reconstruction." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 336, + 348, + 349 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 336, + 348, + 349 + ], + "spans": [ + { + "bbox": [ + 132, + 336, + 348, + 349 + ], + "type": "text", + "content": "5.3 Adjacent Sensitivity Map Estimation" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 354, + 482, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 354, + 482, + 581 + ], + "spans": [ + { + "bbox": [ + 130, + 354, + 482, + 581 + ], + "type": "text", + "content": "In Tab. 6 we examine the impact of various strategies shown in Fig. 4 for sensitivity map estimation on both reconstruction performance and memory consumption. The ACS data input to the SME network is represented as a complex tensor with shape " + }, + { + "bbox": [ + 130, + 354, + 482, + 581 + ], + "type": "inline_equation", + "content": "(B,2a + 1,N,h,w)" + }, + { + "bbox": [ + 130, + 354, + 482, + 581 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 354, + 482, + 581 + ], + "type": "inline_equation", + "content": "2a + 1" + }, + { + "bbox": [ + 130, + 354, + 482, + 581 + ], + "type": "text", + "content": " represents the adjacent number, " + }, + { + "bbox": [ + 130, + 354, + 482, + 581 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 354, + 482, + 581 + ], + "type": "text", + "content": " denotes the number of coils of each slice and " + }, + { + "bbox": [ + 130, + 354, + 482, + 581 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 130, + 354, + 482, + 581 + ], + "type": "text", + "content": " indicates the batch size. Here, the batch size is linearly related to the memory consumption. Strategy A reshapes the tensor to a batch size of " + }, + { + "bbox": [ + 130, + 354, + 482, + 581 + ], + "type": "inline_equation", + "content": "(2a + 1)BN" + }, + { + "bbox": [ + 130, + 354, + 482, + 581 + ], + "type": "text", + "content": " and a channel size of 1 for coil-by-coil estimation. Strategy B concatenates the coils within each slice, reducing the batch size to " + }, + { + "bbox": [ + 130, + 354, + 482, + 581 + ], + "type": "inline_equation", + "content": "(2a + 1)B" + }, + { + "bbox": [ + 130, + 354, + 482, + 581 + ], + "type": "text", + "content": ". Strategy C concatenates the same coils from adjacent slices, decreasing the batch size to " + }, + { + "bbox": [ + 130, + 354, + 482, + 581 + ], + "type": "inline_equation", + "content": "BN" + }, + { + "bbox": [ + 130, + 354, + 482, + 581 + ], + "type": "text", + "content": ". Strategy D combines all coils within a single batch. Note that only strategies A and C can be used for data with variable numbers of coils. Our experiments demonstrate that Strategy C, which utilizes adjacent information, not only shows the best performance but also reduces memory consumption by a factor of " + }, + { + "bbox": [ + 130, + 354, + 482, + 581 + ], + "type": "inline_equation", + "content": "2a + 1" + }, + { + "bbox": [ + 130, + 354, + 482, + 581 + ], + "type": "text", + "content": " compared to Strategy A. Intriguingly, strategy B is found to be the least effective. Ideally, since the coils within the same slice should be uncorrelated, this strategy should not negatively impact performance. However, the noise that is coupled across coils during acquisition may be amplified under this strategy, degrading the SME performance." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 582, + 482, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 482, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 482, + 667 + ], + "type": "text", + "content": "In Tab. 7, we further explore the impact of different adjacent types on Strategy C using the CMRxRecon cardiac datasets. These types include temporal information in the cine LAX dataset, contrast information in the T1 mapping dataset, and slice information in the cine SAX dataset. Strategy C shows consistent improvement over Strategy A for adjacent types of time and contrast. However, the result of the slice adjacency experiment deviates from our expectations, as previously observed in the brain dataset in Tab. 6. This discrepancy" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 162, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 162, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 162, + 91, + 447, + 102 + ], + "type": "text", + "content": "Rethinking Deep Unrolled Model for Accelerated MRI Reconstruction" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 187, + 168, + 427, + 244 + ], + "blocks": [ + { + "bbox": [ + 130, + 114, + 482, + 159 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 482, + 159 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 482, + 159 + ], + "type": "text", + "content": "Table 6: Comparison of different strategies for sensitivity map estimation on the Calgary-Campinas brain dataset under " + }, + { + "bbox": [ + 130, + 114, + 482, + 159 + ], + "type": "inline_equation", + "content": "\\times 10" + }, + { + "bbox": [ + 130, + 114, + 482, + 159 + ], + "type": "text", + "content": " acceleration. The \"Batch\" number is linearly related to memory consumption. \"Flexibility\" indicates the ability to accept data with varying numbers of coils." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 187, + 168, + 427, + 244 + ], + "lines": [ + { + "bbox": [ + 187, + 168, + 427, + 244 + ], + "spans": [ + { + "bbox": [ + 187, + 168, + 427, + 244 + ], + "type": "table", + "html": "
StrategyBatchChannelPSNR(dB)/SSIM(%)Flexibility
A|(2a+1)BN|135.60/94.22✓
B|(2a+1)BN|N35.30/93.98X
CBN|2a+1|35.88/94.43✓
DB||(2a+1)N|35.76/94.32X
", + "image_path": "0e88a376a4ed29a1efaa010489ec87cfca8ea3544a178ee58631c0018bcefc3f.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 221, + 297, + 392, + 359 + ], + "blocks": [ + { + "bbox": [ + 130, + 254, + 482, + 287 + ], + "lines": [ + { + "bbox": [ + 130, + 254, + 482, + 287 + ], + "spans": [ + { + "bbox": [ + 130, + 254, + 482, + 287 + ], + "type": "text", + "content": "Table 7: Quantitative comparison of PSNR(dB)/SSIM(%) of different sensitivity map estimation strategies using different adjacent information types on the CMRxRecon cardiac dataset under " + }, + { + "bbox": [ + 130, + 254, + 482, + 287 + ], + "type": "inline_equation", + "content": "\\times 10" + }, + { + "bbox": [ + 130, + 254, + 482, + 287 + ], + "type": "text", + "content": " acceleration." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 221, + 297, + 392, + 359 + ], + "lines": [ + { + "bbox": [ + 221, + 297, + 392, + 359 + ], + "spans": [ + { + "bbox": [ + 221, + 297, + 392, + 359 + ], + "type": "table", + "html": "
StrategyAdjacent type
TimeContrastSlice
A41.67/97.3542.62/98.1542.75/97.58
C41.94/97.5442.77/98.2242.53/97.54
", + "image_path": "51a790d782a0ca122d5d417e6394f4515b0ac270624e1586f9a9aeeb521b045b.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 381, + 482, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 381, + 482, + 441 + ], + "spans": [ + { + "bbox": [ + 130, + 381, + 482, + 441 + ], + "type": "text", + "content": "is attributed to the considerable slice gap present in the cardiac dataset, which is " + }, + { + "bbox": [ + 130, + 381, + 482, + 441 + ], + "type": "inline_equation", + "content": "4.0\\mathrm{mm}" + }, + { + "bbox": [ + 130, + 381, + 482, + 441 + ], + "type": "text", + "content": " compared to no slice gap in the brain dataset. This suggests that adjacent reconstruction can be beneficial only when adjacent slices show high correlations; otherwise, it might even detract from performance by introducing confounding factors for the network." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 460, + 223, + 472 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 460, + 223, + 472 + ], + "spans": [ + { + "bbox": [ + 131, + 460, + 223, + 472 + ], + "type": "text", + "content": "6 Limitations" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 485, + 482, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 485, + 482, + 556 + ], + "spans": [ + { + "bbox": [ + 130, + 485, + 482, + 556 + ], + "type": "text", + "content": "Like other deep unrolled models, our approach may exhibit limited generalization to unseen sampling masks, modalities, and anatomies, and it may be susceptible to adversarial attacks. Hallucinations and artifacts may be observed when the model is applied to highly undersampled " + }, + { + "bbox": [ + 130, + 485, + 482, + 556 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 485, + 482, + 556 + ], + "type": "text", + "content": "-space data (e.g., at acceleration factors of 10), when trained on a limited amount of data, or when tested on out-of-distribution datasets." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 575, + 220, + 588 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 575, + 220, + 588 + ], + "spans": [ + { + "bbox": [ + 131, + 575, + 220, + 588 + ], + "type": "text", + "content": "7 Conclusion" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 601, + 482, + 649 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 601, + 482, + 649 + ], + "spans": [ + { + "bbox": [ + 130, + 601, + 482, + 649 + ], + "type": "text", + "content": "In this paper, we improve the deep unrolled model for multi-coil MRI reconstruction through informative gradient-based learning and memory-efficient sensitivity map estimation. Our proposed method achieves state-of-the-art performance on several public MRI reconstruction benchmarks." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 219, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 219, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 219, + 101 + ], + "type": "text", + "content": "B. Xin et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 482, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 482, + 152 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 482, + 152 + ], + "type": "text", + "content": "Acknowledgments. This research has been partially funded by research grants to D. Metaxas through NSF: 2310966, 2235405, 2212301, 2003874, and FA9550-23-1-0417 and NIH 2R01HL127661." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 172, + 197, + 185 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 172, + 197, + 185 + ], + "spans": [ + { + "bbox": [ + 133, + 172, + 197, + 185 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 134, + 198, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 138, + 198, + 481, + 221 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 198, + 481, + 221 + ], + "spans": [ + { + "bbox": [ + 138, + 198, + 481, + 221 + ], + "type": "text", + "content": "1. Adler, J., Öktem, O.: Learned primal-dual reconstruction. IEEE transactions on medical imaging 37(6), 1322-1332 (2018)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 222, + 481, + 255 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 222, + 481, + 255 + ], + "spans": [ + { + "bbox": [ + 138, + 222, + 481, + 255 + ], + "type": "text", + "content": "2. Aggarwal, H.K., Mani, M.P., Jacob, M.: Modl: Model-based deep learning architecture for inverse problems. IEEE transactions on medical imaging 38(2), 394-405 (2018)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 255, + 481, + 277 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 255, + 481, + 277 + ], + "spans": [ + { + "bbox": [ + 138, + 255, + 481, + 277 + ], + "type": "text", + "content": "3. Arnold, T.C., Freeman, C.W., Litt, B., Stein, J.M.: Low-field mri: clinical promise and challenges. Journal of Magnetic Resonance Imaging 57(1), 25-44 (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 277, + 481, + 320 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 277, + 481, + 320 + ], + "spans": [ + { + "bbox": [ + 138, + 277, + 481, + 320 + ], + "type": "text", + "content": "4. Arvinte, M., Vishwanath, S., Tewfik, A.H., Tamir, J.I.: Deep j-sense: Accelerated mri reconstruction via unrolled alternating optimization. In: International conference on medical image computing and computer-assisted intervention. pp. 350-360. Springer (2021)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 322, + 481, + 343 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 322, + 481, + 343 + ], + "spans": [ + { + "bbox": [ + 138, + 322, + 481, + 343 + ], + "type": "text", + "content": "5. Beck, A., Teboulle, M.: A fast iterative shrinkage-thresholding algorithm for linear inverse problems. SIAM journal on imaging sciences 2(1), 183-202 (2009)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 344, + 481, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 344, + 481, + 376 + ], + "spans": [ + { + "bbox": [ + 138, + 344, + 481, + 376 + ], + "type": "text", + "content": "6. Boyd, S., Parikh, N., Chu, E., Peleato, B., Eckstein, J., et al.: Distributed optimization and statistical learning via the alternating direction method of multipliers. Foundations and Trends in Machine learning 3(1), 1-122 (2011)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 377, + 481, + 399 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 377, + 481, + 399 + ], + "spans": [ + { + "bbox": [ + 138, + 377, + 481, + 399 + ], + "type": "text", + "content": "7. Candès, E.J., et al.: Compressive sampling. In: Proceedings of the international congress of mathematicians. vol. 3, pp. 1433-1452. Madrid, Spain (2006)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 399, + 481, + 422 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 399, + 481, + 422 + ], + "spans": [ + { + "bbox": [ + 138, + 399, + 481, + 422 + ], + "type": "text", + "content": "8. Deshmane, A., Gulani, V., Griswold, M.A., Seiberlich, N.: Parallel mr imaging. Journal of Magnetic Resonance Imaging 36(1), 55-72 (2012)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 422, + 481, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 422, + 481, + 443 + ], + "spans": [ + { + "bbox": [ + 138, + 422, + 481, + 443 + ], + "type": "text", + "content": "9. Duchi, J., Hazan, E., Singer, Y.: Adaptive subgradient methods for online learning and stochastic optimization. Journal of machine learning research 12(7) (2011)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 134, + 444, + 481, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 444, + 481, + 487 + ], + "spans": [ + { + "bbox": [ + 134, + 444, + 481, + 487 + ], + "type": "text", + "content": "10. El-Rewaidy, H., Fahmy, A.S., Pashakhanloo, F., Cai, X., Kucukseymen, S., Csecs, I., Neisius, U., Haji-Valizadeh, H., Menze, B., Nezafat, R.: Multi-domain convolutional neural network (md-cnn) for radial reconstruction of dynamic cardiac mri. Magnetic Resonance in Medicine 85(3), 1195–1208 (2021)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 134, + 488, + 481, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 488, + 481, + 521 + ], + "spans": [ + { + "bbox": [ + 134, + 488, + 481, + 521 + ], + "type": "text", + "content": "11. Eo, T., Jun, Y., Kim, T., Jang, J., Lee, H.J., Hwang, D.: Kiki-net: cross-domain convolutional neural networks for reconstructing undersampled magnetic resonance images. Magnetic resonance in medicine 80(5), 2188-2201 (2018)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 134, + 521, + 481, + 554 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 521, + 481, + 554 + ], + "spans": [ + { + "bbox": [ + 134, + 521, + 481, + 554 + ], + "type": "text", + "content": "12. Fabian, Z., Tinaz, B., Soltanolkotabi, M.: Humus-net: Hybrid unrolled multi-scale network architecture for accelerated mri reconstruction. Advances in Neural Information Processing Systems 35, 25306-25319 (2022)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 134, + 554, + 481, + 587 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 554, + 481, + 587 + ], + "spans": [ + { + "bbox": [ + 134, + 554, + 481, + 587 + ], + "type": "text", + "content": "13. Gregor, K., LeCun, Y.: Learning fast approximations of sparse coding. In: Proceedings of the 27th international conference on international conference on machine learning. pp. 399-406 (2010)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 134, + 588, + 481, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 588, + 481, + 632 + ], + "spans": [ + { + "bbox": [ + 134, + 588, + 481, + 632 + ], + "type": "text", + "content": "14. Griswold, M.A., Jakob, P.M., Heidemann, R.M., Nittka, M., Jellus, V., Wang, J., Kiefer, B., Haase, A.: Generalized autocalibrating partially parallel acquisitions (grappa). Magnetic Resonance in Medicine: An Official Journal of the International Society for Magnetic Resonance in Medicine 47(6), 1202-1210 (2002)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 134, + 632, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 632, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 134, + 632, + 481, + 665 + ], + "type": "text", + "content": "15. Hammernik, K., Klatzer, T., Kobler, E., Recht, M.P., Sodickson, D.K., Pock, T., Knoll, F.: Learning a variational network for reconstruction of accelerated mri data. Magnetic resonance in medicine 79(6), 3055-3071 (2018)" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 162, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 162, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 162, + 91, + 447, + 102 + ], + "type": "text", + "content": "Rethinking Deep Unrolled Model for Accelerated MRI Reconstruction" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 666 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 481, + 160 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 481, + 160 + ], + "type": "text", + "content": "16. Hosseini, S.A.H., Yaman, B., Moeller, S., Hong, M., Akçakaya, M.: Dense recurrent neural networks for accelerated mri: History-cognizant unrolling of optimization algorithms. IEEE Journal of Selected Topics in Signal Processing 14(6), 1280-1291 (2020)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 160, + 481, + 203 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 160, + 481, + 203 + ], + "spans": [ + { + "bbox": [ + 133, + 160, + 481, + 203 + ], + "type": "text", + "content": "17. Jun, Y., Shin, H., Eo, T., Hwang, D.: Joint deep model-based mr image and coil sensitivity reconstruction network (joint-icnet) for fast mri. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5270-5279 (2021)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 203, + 481, + 224 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 203, + 481, + 224 + ], + "spans": [ + { + "bbox": [ + 133, + 203, + 481, + 224 + ], + "type": "text", + "content": "18. Kingma, D.P., Ba, J.: Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980 (2014)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 225, + 481, + 246 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 225, + 481, + 246 + ], + "spans": [ + { + "bbox": [ + 132, + 225, + 481, + 246 + ], + "type": "text", + "content": "19. Lauterbur, P.C.: Image formation by induced local interactions: examples employing nuclear magnetic resonance. nature 242(5394), 190-191 (1973)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 246, + 481, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 246, + 481, + 267 + ], + "spans": [ + { + "bbox": [ + 132, + 246, + 481, + 267 + ], + "type": "text", + "content": "20. LeCun, Y., Bengio, Y., Hinton, G.: Deep learning. nature 521(7553), 436-444 (2015)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 267, + 481, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 267, + 481, + 300 + ], + "spans": [ + { + "bbox": [ + 132, + 267, + 481, + 300 + ], + "type": "text", + "content": "21. Liu, X., Pang, Y., Jin, R., Liu, Y., Wang, Z.: Dual-domain reconstruction network with v-net and k-net for fast mri. Magnetic Resonance in Medicine 88(6), 2694-2708 (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 300, + 481, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 300, + 481, + 342 + ], + "spans": [ + { + "bbox": [ + 132, + 300, + 481, + 342 + ], + "type": "text", + "content": "22. Lustig, M., Donoho, D., Pauly, J.M.: Sparse mri: The application of compressed sensing for rapid mr imaging. Magnetic Resonance in Medicine: An Official Journal of the International Society for Magnetic Resonance in Medicine 58(6), 1182-1195 (2007)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 342, + 481, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 342, + 481, + 374 + ], + "spans": [ + { + "bbox": [ + 132, + 342, + 481, + 374 + ], + "type": "text", + "content": "23. Monga, V., Li, Y., Eldar, Y.C.: Algorithm unrolling: Interpretable, efficient deep learning for signal and image processing. IEEE Signal Processing Magazine 38(2), 18-44 (2021)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 374, + 481, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 374, + 481, + 407 + ], + "spans": [ + { + "bbox": [ + 132, + 374, + 481, + 407 + ], + "type": "text", + "content": "24. Munoz, C., Fotaki, A., Botnar, R.M., Prieto, C.: Latest advances in image acceleration: all dimensions are fair game. Journal of Magnetic Resonance Imaging 57(2), 387-402 (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 407, + 481, + 429 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 407, + 481, + 429 + ], + "spans": [ + { + "bbox": [ + 132, + 407, + 481, + 429 + ], + "type": "text", + "content": "25. Nayak, K.S., Lim, Y., Campbell-Washburn, A.E., Steeden, J.: Real-time magnetic resonance imaging. Journal of Magnetic Resonance Imaging 55(1), 81-99 (2022)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 429, + 481, + 460 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 429, + 481, + 460 + ], + "spans": [ + { + "bbox": [ + 132, + 429, + 481, + 460 + ], + "type": "text", + "content": "26. Ongie, G., Jalal, A., Metzler, C.A., Baraniuk, R.G., Dimakis, A.G., Willett, R.: Deep learning techniques for inverse problems in imaging. IEEE Journal on Selected Areas in Information Theory 1(1), 39-56 (2020)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 460, + 481, + 493 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 460, + 481, + 493 + ], + "spans": [ + { + "bbox": [ + 132, + 460, + 481, + 493 + ], + "type": "text", + "content": "27. Ottesen, J.A., Caan, M.W., Groote, I.R., Bjornerud, A.: A densely interconnected network for deep learning accelerated mri. Magnetic Resonance Materials in Physics, Biology and Medicine 36(1), 65-77 (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 493, + 481, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 493, + 481, + 536 + ], + "spans": [ + { + "bbox": [ + 132, + 493, + 481, + 536 + ], + "type": "text", + "content": "28. Pezzotti, N., Yousefi, S., Elmahdy, M.S., Van Gemert, J.H.F., Schuelke, C., Doneva, M., Nielsen, T., Kastryulin, S., Lieleveldt, B.P., Van Osch, M.J., et al.: An adaptive intelligence algorithm for undersampled knee mri reconstruction. IEEE Access 8, 204825-204838 (2020)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 536, + 481, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 536, + 481, + 557 + ], + "spans": [ + { + "bbox": [ + 132, + 536, + 481, + 557 + ], + "type": "text", + "content": "29. Polyak, B.T.: Some methods of speeding up the convergence of iteration methods. Ussr computational mathematics and mathematical physics 4(5), 1-17 (1964)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 558, + 481, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 558, + 481, + 590 + ], + "spans": [ + { + "bbox": [ + 132, + 558, + 481, + 590 + ], + "type": "text", + "content": "30. Pruessmann, K.P., Weiger, M., Scheidegger, M.B., Boesiger, P.: Sense: sensitivity encoding for fast mri. Magnetic Resonance in Medicine: An Official Journal of the International Society for Magnetic Resonance in Medicine 42(5), 952-962 (1999)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 590, + 481, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 590, + 481, + 622 + ], + "spans": [ + { + "bbox": [ + 132, + 590, + 481, + 622 + ], + "type": "text", + "content": "31. Putzky, P., Karkalousos, D., Teuwen, J., Miriakov, N., Bakker, B., Caan, M., Welling, M.: i-rim applied to the fastmri challenge. arXiv preprint arXiv:1910.08952 (2019)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 622, + 481, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 622, + 481, + 643 + ], + "spans": [ + { + "bbox": [ + 132, + 622, + 481, + 643 + ], + "type": "text", + "content": "32. Putzky, P., Welling, M.: Recurrent inference machines for solving inverse problems. arXiv preprint arXiv:1706.04008 (2017)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 643, + 481, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 643, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 643, + 481, + 666 + ], + "type": "text", + "content": "33. Putzky, P., Welling, M.: Invert to learn to invert. Advances in neural information processing systems 32 (2019)" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "text", + "content": "B. Xin et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 482, + 666 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 132, + 116, + 482, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 482, + 149 + ], + "type": "text", + "content": "34. Ramzi, Z., Chaithya, G., Starck, J.L., Ciuciu, P.: Nc-pdnet: A density-compensated unrolled network for 2d and 3d non-cartesian mri reconstruction. IEEE Transactions on Medical Imaging 41(7), 1625-1638 (2022)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 150, + 482, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 150, + 482, + 194 + ], + "spans": [ + { + "bbox": [ + 132, + 150, + 482, + 194 + ], + "type": "text", + "content": "35. Ronneberger, O., Fischer, P., Brox, T.: U-net: Convolutional networks for biomedical image segmentation. In: Medical Image Computing and Computer-Assisted Intervention-MICCAI 2015: 18th International Conference, Munich, Germany, October 5-9, 2015, Proceedings, Part III 18. pp. 234-241. Springer (2015)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 194, + 482, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 194, + 482, + 226 + ], + "spans": [ + { + "bbox": [ + 132, + 194, + 482, + 226 + ], + "type": "text", + "content": "36. Schlemper, J., Caballero, J., Hajnal, J.V., Price, A.N., Rueckert, D.: A deep cascade of convolutional neural networks for dynamic mr image reconstruction. IEEE transactions on Medical Imaging 37(2), 491-503 (2017)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 227, + 482, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 227, + 482, + 259 + ], + "spans": [ + { + "bbox": [ + 132, + 227, + 482, + 259 + ], + "type": "text", + "content": "37. Singh, D., Monga, A., de Moura, H.L., Zhang, X., Zibetti, M.V., Regatte, R.R.: Emerging trends in fast mri using deep-learning reconstruction on undersampled k-space data: a systematic review. Bioengineering 10(9), 1012 (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 260, + 482, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 260, + 482, + 293 + ], + "spans": [ + { + "bbox": [ + 132, + 260, + 482, + 293 + ], + "type": "text", + "content": "38. Song, J., Chen, B., Zhang, J.: Memory-augmented deep unfolding network for compressive sensing. In: Proceedings of the 29th ACM international conference on multimedia. pp. 4249-4258 (2021)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 293, + 482, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 293, + 482, + 335 + ], + "spans": [ + { + "bbox": [ + 132, + 293, + 482, + 335 + ], + "type": "text", + "content": "39. Souza, R., Lucena, O., Garrafa, J., Gobbi, D., Saluzzi, M., Appenzeller, S., Rittner, L., Frayne, R., Lotufo, R.: An open, multi-vendor, multi-field-strength brain mr dataset and analysis of publicly available skull stripping methods agreement. NeuroImage 170, 482-494 (2018)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 336, + 482, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 336, + 482, + 392 + ], + "spans": [ + { + "bbox": [ + 132, + 336, + 482, + 392 + ], + "type": "text", + "content": "40. Sriram, A., Zbontar, J., Murrell, T., Defazio, A., Zitnick, C.L., Yakubova, N., Knoll, F., Johnson, P.: End-to-end variational networks for accelerated mri reconstruction. In: Medical Image Computing and Computer Assisted Intervention-MICCAI 2020: 23rd International Conference, Lima, Peru, October 4-8, 2020, Proceedings, Part II 23. pp. 64-73. Springer (2020)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 392, + 482, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 392, + 482, + 413 + ], + "spans": [ + { + "bbox": [ + 132, + 392, + 482, + 413 + ], + "type": "text", + "content": "41. Sun, J., Li, H., Xu, Z., et al.: Deep admm-net for compressive sensing mri. Advances in neural information processing systems 29 (2016)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 413, + 482, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 413, + 482, + 445 + ], + "spans": [ + { + "bbox": [ + 132, + 413, + 482, + 445 + ], + "type": "text", + "content": "42. Tieleman, T.: Lecture 6.5-rmsprop: Divide the gradient by a running average of its recent magnitude. COURSERA: Neural networks for machine learning 4(2), 26 (2012)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 446, + 482, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 446, + 482, + 479 + ], + "spans": [ + { + "bbox": [ + 132, + 446, + 482, + 479 + ], + "type": "text", + "content": "43. Uecker, M., Lai, P., Murphy, M.J., Virtue, P., Elad, M., Pauly, J.M., Vasanawala, S.S., Lustig, M.: Espirit—an eigenvalue approach to autocalibrating parallel migraine: where sense meets grappa. Magnetic resonance in medicine 71(3), 990-1001 (2014)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 479, + 482, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 479, + 482, + 511 + ], + "spans": [ + { + "bbox": [ + 132, + 479, + 482, + 511 + ], + "type": "text", + "content": "44. Wang, C., Lyu, J., Wang, S., Qin, C., Guo, K., Zhang, X., Yu, X., Li, Y., Wang, F., Jin, J., et al.: Cmrxrecon: an open cardiac mri dataset for the competition of accelerated image reconstruction. arXiv preprint arXiv:2309.10836 (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 512, + 482, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 512, + 482, + 555 + ], + "spans": [ + { + "bbox": [ + 132, + 512, + 482, + 555 + ], + "type": "text", + "content": "45. Xin, B., Ye, M., Axel, L., Metaxas, D.N.: Fill the k-space and refine the image: Prompting for dynamic and multi-contrast mri reconstruction. In: International Workshop on Statistical Atlases and Computational Models of the Heart. pp. 261-273. Springer (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 555, + 482, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 555, + 482, + 600 + ], + "spans": [ + { + "bbox": [ + 132, + 555, + 482, + 600 + ], + "type": "text", + "content": "46. Yiasemis, G., Sonke, J.J., Sánchez, C., Teuwen, J.: Recurrent variational network: a deep learning inverse problem solver applied to the task of accelerated mri reconstruction. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 732-741 (2022)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 601, + 482, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 601, + 482, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 601, + 482, + 632 + ], + "type": "text", + "content": "47. Ying, L., Sheng, J.: Joint image reconstruction and sensitivity estimation in sense (jsense). Magnetic Resonance in Medicine: An Official Journal of the International Society for Magnetic Resonance in Medicine 57(6), 1196-1202 (2007)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 632, + 482, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 482, + 666 + ], + "type": "text", + "content": "48. You, D., Xie, J., Zhang, J.: Ista-net++: Flexible deep unfolding network for compressive sensing. In: 2021 IEEE International Conference on Multimedia and Expo (ICME). pp. 1-6. IEEE (2021)" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 162, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 162, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 162, + 91, + 447, + 101 + ], + "type": "text", + "content": "Rethinking Deep Unrolled Model for Accelerated MRI Reconstruction" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 369 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "type": "text", + "content": "49. Zbontar, J., Knoll, F., Sriram, A., Murrell, T., Huang, Z., Muckley, M.J., Defazio, A., Stern, R., Johnson, P., Bruno, M., et al.: fastmri: An open dataset and benchmarks for accelerated mri. arXiv preprint arXiv:1811.08839 (2018)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 150, + 481, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 150, + 481, + 171 + ], + "spans": [ + { + "bbox": [ + 132, + 150, + 481, + 171 + ], + "type": "text", + "content": "50. Zeiler, M.D.: Adadelta: an adaptive learning rate method. arXiv preprint arXiv:1212.5701 (2012)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 172, + 481, + 193 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 172, + 481, + 193 + ], + "spans": [ + { + "bbox": [ + 132, + 172, + 481, + 193 + ], + "type": "text", + "content": "51. Zhang, J., Chen, B., Xiong, R., Zhang, Y.: Physics-inspired compressive sensing: Beyond deep unrolling. IEEE Signal Processing Magazine 40(1), 58-72 (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 194, + 481, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 194, + 481, + 226 + ], + "spans": [ + { + "bbox": [ + 132, + 194, + 481, + 226 + ], + "type": "text", + "content": "52. Zhang, J., Ghanem, B.: Ista-net: Interpretable optimization-inspired deep network for image compressive sensing. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1828-1837 (2018)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 227, + 481, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 227, + 481, + 258 + ], + "spans": [ + { + "bbox": [ + 132, + 227, + 481, + 258 + ], + "type": "text", + "content": "53. Zhang, J., Zhang, Z., Xie, J., Zhang, Y.: High-throughput deep unfolding network for compressive sensing mri. IEEE Journal of Selected Topics in Signal Processing 16(4), 750-761 (2022)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 259, + 481, + 292 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 259, + 481, + 292 + ], + "spans": [ + { + "bbox": [ + 132, + 259, + 481, + 292 + ], + "type": "text", + "content": "54. Zhang, Z., Liu, Y., Liu, J., Wen, F., Zhu, C.: Amp-net: Denoising-based deep unfolding for compressive image sensing. IEEE Transactions on Image Processing 30, 1487-1500 (2020)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 293, + 481, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 293, + 481, + 335 + ], + "spans": [ + { + "bbox": [ + 132, + 293, + 481, + 335 + ], + "type": "text", + "content": "55. Zhao, R., Yaman, B., Zhang, Y., Stewart, R., Dixon, A., Knoll, F., Huang, Z., Lui, Y.W., Hansen, M.S., Lungren, M.P.: fastmri+, clinical pathology annotations for knee and brain fully sampled magnetic resonance imaging data. Scientific Data 9(1), 152 (2022)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 336, + 481, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 336, + 481, + 369 + ], + "spans": [ + { + "bbox": [ + 132, + 336, + 481, + 369 + ], + "type": "text", + "content": "56. Zhou, B., Zhou, S.K.: Dudornet: learning a dual-domain recurrent network for fast mri reconstruction with deep t1 prior. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 4273-4282 (2020)" + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 100 + ], + "type": "text", + "content": "B. Xin et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Rethinking Directional Parameterization in Neural Implicit Surface Reconstruction/8b296952-4d44-42db-affd-9e9843feb2a6_content_list.json b/2024/Rethinking Directional Parameterization in Neural Implicit Surface Reconstruction/8b296952-4d44-42db-affd-9e9843feb2a6_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..3294cd28fd1227edf30831edd3e08fb626cc7495 --- /dev/null +++ b/2024/Rethinking Directional Parameterization in Neural Implicit Surface Reconstruction/8b296952-4d44-42db-affd-9e9843feb2a6_content_list.json @@ -0,0 +1,1892 @@ +[ + { + "type": "text", + "text": "Rethinking Directional Parameterization in Neural Implicit Surface Reconstruction", + "text_level": 1, + "bbox": [ + 254, + 140, + 750, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zijie Jiang $^{1\\star}$ , Tianhan Xu $^{2\\star}$ , and Hiroharu Kato $^{2\\star}$", + "bbox": [ + 303, + 210, + 697, + 227 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Tokyo Institute of Technology", + "bbox": [ + 393, + 238, + 609, + 253 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "2 Preferred Networks, Inc.", + "bbox": [ + 410, + 253, + 589, + 267 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "zjiang@ok.sc.e.titech.ac.jp, {xutianhan,hkato}@preferred.jp", + "bbox": [ + 271, + 268, + 728, + 282 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Multi-view 3D surface reconstruction using neural implicit representations has made notable progress by modeling the geometry and view-dependent radiance fields within a unified framework. However, their effectiveness in reconstructing objects with specular or complex surfaces is typically biased by the directional parameterization used in their view-dependent radiance network. Viewing direction and reflection direction are the two most commonly used directional parameterizations but have their own limitations. Typically, utilizing the viewing direction usually struggles to correctly decouple the geometry and appearance of objects with highly specular surfaces, while using the reflection direction tends to yield overly smooth reconstructions for concave or complex structures. In this paper, we analyze their failed cases in detail and propose a novel hybrid directional parameterization to address their limitations in a unified form. Extensive experiments demonstrate the proposed hybrid directional parameterization consistently delivered satisfactory results in reconstructing objects with a wide variety of materials, geometry and appearance, whereas using other directional parameterizations faces challenges in reconstructing certain objects. Moreover, the proposed hybrid directional parameterization is nearly parameter-free and can be effortlessly applied in any existing neural surface reconstruction method.", + "bbox": [ + 261, + 311, + 743, + 590 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: surface reconstruction $\\cdot$ neural implicit representation $\\cdot$ directional parameterization", + "bbox": [ + 259, + 602, + 740, + 630 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 654, + 375, + 669 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recent advances in neural fields [32] have revolutionized the field of 3D geometry reconstruction from multi-view images [22,29,36]. Compared to traditional multiview stereo methods [7], a key advantage of using neural fields is their ability to model and optimize complex light reflections on object surfaces, which improves robustness to specular or non-Lambertian surfaces. Since a radiance network is trained jointly with a geometry network via differentiable rendering, the radiance network is able to take full advantage of geometry and view-dependent information such as viewing direction to decouple the geometry and appearance.", + "bbox": [ + 212, + 681, + 787, + 805 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Equal contribution. Work done while Zijie Jiang interned at Preferred Networks, Inc.", + "bbox": [ + 217, + 810, + 785, + 838 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/848f7d5786464271261be10c8c9c749d9cce2a9633a67977212e2ef8db7da901.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 230, + 155, + 316, + 220 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/7b34a9ac335a9451512f4c90bba12b883774b83ad1ec963a6cb01bc54c08921d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 346, + 155, + 439, + 222 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/099237128abfdb40b6bf75a0ce9fb5a4f426596d2c67c318581b0b651766fb47.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 455, + 155, + 547, + 222 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/507ddd77285c5b5dd94f02408ad3064e463563e5a920fe039d2fbb8e3fe40b8a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 563, + 155, + 656, + 222 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/21f81a33a9b0f020207a8ef732704172d6dd3e0b0a4f5ae7324e49dc251de890.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 692, + 155, + 782, + 222 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/e62ad7cdac22ed7fdd0f32b43c377ed0be9fc86750d6949c3bcb4c80b0865e6d.jpg", + "image_caption": [ + "Reference RGB" + ], + "image_footnote": [], + "bbox": [ + 233, + 233, + 320, + 301 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/c4ea86767a1d98e7b906447a4a0b8cbaadc48e1b103b930a8aa578362aa21184.jpg", + "image_caption": [ + "(a) Viewing dir." + ], + "image_footnote": [], + "bbox": [ + 352, + 233, + 439, + 303 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/d0e9d9a9ff1764223b2e32ea5210c63d4f6fb241468d1cfb5f777156d28b0585.jpg", + "image_caption": [ + "(b) Reflection dir.", + "Fig. 1: Comparison of reconstruction results using different directional parameterizations in [2]. Using the viewing directional parameterization fails to reconstruct objects with specular surfaces, while using the reflection directional parameterization tends to result in incorrect geometry for objects with concave structures. Using our proposed hybrid parametrization shows consistently satisfactory reconstruction results compared to the existing parameterizations." + ], + "image_footnote": [], + "bbox": [ + 459, + 233, + 547, + 303 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/2e0de4335a1138e2bcb7c99d421f7d3d822ecad98a1e783bf8c39d8be5717979.jpg", + "image_caption": [ + "(c) Hybrid dir. (Ours)" + ], + "image_footnote": [], + "bbox": [ + 565, + 233, + 653, + 301 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/25b34e5738b73517dbc1077716d3dc1e507b285e6d57439c26809a9ca40f2491.jpg", + "image_caption": [ + "Ground-truth Mesh" + ], + "image_footnote": [], + "bbox": [ + 692, + 233, + 779, + 301 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Since the decomposition of geometry and appearance is inherently an ill-posed problem $[40]^3$ , it is essential to introduce an appropriate inductive bias into the radiance network to facilitate correct decomposition. For example, simply inputting the viewing direction to the radiance network sometimes leads to incorrect reconstructions of highly specular objects (Fig. 1 (a)). To address this issue, the recent study [28] suggests that providing the reflection direction of rays considering the learned local surface geometry (i.e., normals) would result in a simpler radiance field to learn and facilitate correct decomposition, especially for specular objects. Some subsequent studies have shown that adopting this parameterization results in satisfactory reconstructions of highly specular objects [8, 13, 14] and also non-reflecting objects [8]. However, although this parameterization seems promising (Fig. 1 (b) top), we observed that it still does not work well when reconstructing objects with complex structures such as concave surfaces (Fig. 1 (b) bottom). Interestingly, while using viewing direction fails to recover specular surfaces, it succeeds in recovering such concave structures (Fig. 1 (a) bottom). Based on these, we pose the question: Is there a unified directional parameterization that can effectively handle both specular surfaces and complex structures?", + "bbox": [ + 212, + 454, + 787, + 726 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, to answer the question, we start by studying the limitations of existing directional parameterizations and provide the first in-depth analysis of why using reflection direction may fail to reconstruct objects with concave or complex structures. As mentioned above, one important feature of the reflection directional parameterization is its dependency on the learnable geometry (i.e.,", + "bbox": [ + 212, + 727, + 787, + 804 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "Jiang and Xu et al.", + "bbox": [ + 271, + 114, + 401, + 128 + ], + "page_idx": 1 + }, + { + "type": "page_footnote", + "text": "3 If the capacity of the radiance network is sufficiently high, even for a fixed geometry such as a sphere, the image reconstruction loss can converge to zero [40].", + "bbox": [ + 217, + 810, + 785, + 840 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "normals). Our insight into this fact is that for sampled 3D points along a sampled ray during training, while utilizing reflection directions of sampled points near the intersecting surface is helpful in decoupling geometry and appearance, the use of reflection directions of sampled points elsewhere may hinder the accurate reconstruction of geometry. Specifically, we identified that the cause of this issue lies in the sampled points beyond a certain distance from the intersecting surface during the optimization. As we will discuss in Sec. 3.2, the normals of these points computed from learnable geometry may either be associated with surfaces unrelated to the intersection or introduce high-frequency variations in the input of the radiance network, which hinders the proper updating of geometry and radiance. This hindrance is particularly noticeable in the reconstruction of concave surfaces. The analysis also indirectly explains the advantage of using viewing directional parameterization in reconstructing concave or complex structures, as it is non-learnable and remains constant for all sampled points along a sampled ray.", + "bbox": [ + 212, + 146, + 787, + 375 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Based on our insights and analysis, we propose a novel hybrid directional parameterization to address the limitations of existing directional parameterizations in a unified form. The main idea is to utilize the predicted distance to the surface for each sampled point: if the distance is small, indicating proximity to the intersecting surface, it is preferable to use its reflection direction as the directional parameterization, which facilitates the correct decomposition of geometry and appearance; in other cases, we instead utilize the non-learnable viewing direction to avoid the optimization issue mentioned above. The proposed hybrid directional parameterization smoothly transitions from the reflection direction to the viewing direction based on the predicted distance to the surface. The rate of this directional transition is controlled by a learnable scaling factor. We integrate the proposed hybrid directional parameterization into different methods and conduct extensive experiments to validate its effectiveness across various datasets. Compared to the existing viewing and reflection directional parameterizations, our proposed hybrid directional parameterization achieves consistently satisfactory results, as shown in Fig. 1 (c). Moreover, the proposed hybrid directional parameterization is nearly parameter-free and can be effortlessly applied in any existing neural implicit surface reconstruction method.", + "bbox": [ + 212, + 376, + 787, + 648 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To summarize, our contributions are as follows:", + "bbox": [ + 240, + 652, + 581, + 666 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We are the first to provide an in-depth analysis of why the use of reflection directional parameterization tends to reconstruct incorrect geometry of objects with concave or complex structures.", + "- We propose a novel hybrid directional parameterization to address the limitations of existing directional parameterizations, which can effectively handle both specular surfaces and complex structures in a unified form.", + "- We validate the effectiveness of the proposed directional parameterization across different methods on various datasets, and demonstrate an overall superiority compared to existing directional parameterizations." + ], + "bbox": [ + 225, + 696, + 785, + 839 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "Rethinking directional parameterization", + "bbox": [ + 464, + 114, + 730, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 143, + 387, + 160 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.1 Neural Implicit Surface Reconstruction", + "text_level": 1, + "bbox": [ + 215, + 174, + 584, + 191 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "With the significant advancement of neural implicit representations [3,16-19,23, 24,37] and differentiable rendering [10,11,17,20,25,27,33] in recent years, neural implicit surface reconstruction methods [21,29,35,36] have emerged as a promising alternative to traditional methods in the field of image-based 3D surface reconstruction. Compared to traditional multi-view stereo methods, neural implicit surface methods employ a more flexible and representative implicit shape representation, such as the signed distance function (SDF), and introduce the directional radiance field [17] to model the view-dependent appearance of the target object. Subsequent works on neural implicit surface reconstruction have mostly built upon the foundation of NeuS [29] or VolSDF [35], and made improvements for specific scenarios including fast rendering [30, 31], sparse-view inputs [15], large-scale scenes [12,38] and more.", + "bbox": [ + 215, + 198, + 785, + 380 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.2 Reconstruction for Specular Objects", + "text_level": 1, + "bbox": [ + 215, + 401, + 560, + 417 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Despite neural implicit surface reconstruction methods consider the view-dependent appearance of objects by incorporating the directional radiance field, reconstruction for specular objects remains a challenging task due to its significant shape-radiance ambiguity. Several works [14, 26, 34, 39, 41] model the view-dependent appearance of specular objects by incorporating additional physical components (e.g., materials and lighting), with the hope of jointly recovering both the geometric shape and physical parameters. However, as the number of unknown parameters increases, these works typically rely on a good initial geometry to ensure favorable results, leading to a chicken-and-egg problem. On the other hand, Ref-NeRF [28] was the first to propose the use of a radiance field conditioned on reflection direction to better decouple the geometry and radiance of specular objects. Subsequent neural implicit surface reconstruction methods targeting specular objects [8, 13, 14] have also adopted this directional parameterization. However, these methods often fail to correctly recover the geometry of objects with concave or complex structures.", + "bbox": [ + 215, + 424, + 802, + 651 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3 Method", + "text_level": 1, + "bbox": [ + 215, + 672, + 328, + 689 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this section, we introduce a novel directional parameterization for neural implicit surface reconstruction methods. We first give a brief review on neural implicit surface reconstruction and its commonly used directional parameterizations, including viewing and reflection directions (Sec. 3.1). In the next, we summarize the observed failed reconstructions when utilizing existing directional parameterizations and conduct a detailed analysis of their causes (Sec. 3.2). Finally, a hybrid directional parameterization is proposed based on our analysis, which can effectively address the limitations of existing directional parameterizations in a unified form (Sec. 3.3).", + "bbox": [ + 215, + 703, + 785, + 839 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "Jiang and Xu et al.", + "bbox": [ + 271, + 114, + 401, + 128 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1 Preliminaries of Neural Implicit Surface Reconstruction", + "text_level": 1, + "bbox": [ + 215, + 146, + 722, + 161 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Neural implicit surface reconstruction methods [2,6,29,36] typically represent the 3D object as the combination of a signed distance function (SDF) network $\\Theta_{f}$ , which maps a spatial position $\\mathbf{x} \\in \\mathbb{R}^3$ to its SDF value $f(\\mathbf{x}) \\in \\mathbb{R}$ , and a directional radiance network $\\Theta_{c}$ , which maps a spatial position $\\mathbf{x} \\in \\mathbb{R}^3$ and a directional vector $\\mathbf{d} \\in \\mathbb{S}^2$ to the view-dependent color at $\\mathbf{x}$ . The surface $\\mathcal{S}$ of the object is extracted as the zero-level set of $f(\\mathbf{x})$ . To enable the joint optimization of the SDF network $\\Theta_{f}$ and radiance network $\\Theta_{c}$ using only 2D image supervisions, volume rendering is typically utilized [29,35]. Given the sampled pixels in an image, camera rays $\\{\\mathbf{r}(t) = \\mathbf{o} + t\\mathbf{d}_{\\mathrm{view}}|t > 0\\}$ are cast through the camera origin $\\mathbf{o}$ and the sampled pixels, where $\\mathbf{d}_{\\mathrm{view}}$ is the unit directional vector of the ray. To obtain the rendered colors of the sampled rays, a number of points are sampled along each ray, and queried for their SDF values and view-dependent colors from $\\Theta_{f}$ and $\\Theta_{c}$ , where positional encoding (PE) [17] is typically incorporated. A relationship between the SDF value and opaque density is then formulated, such as using S-density [29], to translate the queried SDFs and colors to the ray color akin to NeRF [17]. Finally, the training loss is calculated from the difference between the rendered ray colors and the ground-truth colors of the sampled pixels.", + "bbox": [ + 212, + 171, + 787, + 443 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Directional parameterization. The key to the ability of the radiance network to model view-dependent appearance lies in its incorporation of a directional vector $\\mathbf{d}$ as an input. Common directional parameterizations for $\\mathbf{d}$ includes the viewing direction $\\mathbf{d}_{\\mathrm{view}}$ and reflection direction $\\mathbf{d}_{\\mathrm{ref}}$ . The viewing directional parameterization is intuitively straightforward, as it models the view-dependent appearance from the perspective of the observer. The reflection directional parameterization is a little more complicated, which encodes the viewing direction $\\mathbf{d}_{\\mathrm{view}}$ and local normal direction $\\mathbf{n}$ into a new compact directional vector $\\mathbf{d}_{\\mathrm{ref}}$ , by considering the interaction between light and surface. Intuitively, it represents the direction of the incoming light, assuming that the surface is specular. $\\mathbf{d}_{\\mathrm{ref}}$ is formulated as follows:", + "bbox": [ + 212, + 444, + 787, + 609 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {d} _ {\\text {r e f}} = 2 \\left(\\mathbf {d} _ {\\text {v i e w}} \\cdot \\mathbf {n}\\right) \\mathbf {n} - \\mathbf {d} _ {\\text {v i e w}}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 400, + 625, + 784, + 638 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathbf{n}$ can be derived from the gradient of the SDF as $\\mathbf{n}(\\mathbf{x}) = \\frac{\\nabla f(\\mathbf{x})}{\\|\\nabla f(\\mathbf{x})\\|}$ .", + "bbox": [ + 215, + 650, + 740, + 672 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2 Observations and Analysis", + "text_level": 1, + "bbox": [ + 215, + 691, + 480, + 708 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Observations. Although existing neural implicit surface reconstruction methods widely adopt either viewing direction or reflection direction as their directional parameterization, both types of parameterizations have their respective limitations when it comes to reconstructing certain objects. Using viewing direction often faces difficulty in correctly reconstructing objects with specular surfaces (e.g., Fig. 1 (a) and Fig. 2 (c)). On the other hand, while using reflection direction can reconstruct specular objects with simple geometry (e.g., Fig. 1 (b) top and Fig. 2 (d)), it often results in overly smooth reconstructions with", + "bbox": [ + 212, + 719, + 787, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "Rethinking directional parameterization", + "bbox": [ + 464, + 114, + 730, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/e1a21e9151b98fee629c4833abe336f42d92846b00706ce5e0fe689eb69c53d5.jpg", + "image_caption": [ + "Object w/ concave structures" + ], + "image_footnote": [], + "bbox": [ + 222, + 143, + 330, + 214 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/cfc024136b15d5757e12dde8818d625263edcf258efeac08e867a7250c185e29.jpg", + "image_caption": [ + "(a)Viewing dir." + ], + "image_footnote": [], + "bbox": [ + 338, + 143, + 406, + 214 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/f11731abf1ca7093a2f0609bb46e6b7db2d06a9d79b34bb9c8279e48a8c7d6d3.jpg", + "image_caption": [ + "(b) Reflection dir. (X" + ], + "image_footnote": [], + "bbox": [ + 426, + 143, + 491, + 214 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/1b16b6cacd0b31273c79cebae3df4579e77aa0d071b851ea41ef920c94f235ad.jpg", + "image_caption": [ + "Object w/ specular surfaces" + ], + "image_footnote": [], + "bbox": [ + 519, + 148, + 593, + 213 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/ef33002b3c0a319f8e4f9158dec273b5c951c3aece42f4f8e7c320eee5fa5e77.jpg", + "image_caption": [ + "(c) Viewing dir." + ], + "image_footnote": [], + "bbox": [ + 609, + 152, + 681, + 215 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/f3041073eba5d32491ccc070bd37360b70f38f343e5c09ceb8443f9057a9126f.jpg", + "image_caption": [ + "(d) Reflection dir." + ], + "image_footnote": [], + "bbox": [ + 694, + 152, + 767, + 215 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/229953c5b020bdb986ee665ce42d2dada63d5319a341120842b80ac7b314d7c2.jpg", + "image_caption": [ + "Fig. 2: Scenarios where existing directional parameterizations succeed and struggle. Reconstruction results are obtained by integrating the two directional parameterizations into NeuS [29]." + ], + "image_footnote": [], + "bbox": [ + 727, + 226, + 746, + 239 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "missing concave surfaces when reconstructing objects with concave or complex structures (e.g., Fig. 1 (b) bottom and Fig. 2 (b)). Interestingly, using viewing direction may yield better results in these parts with complex or concave surfaces (e.g., Fig. 1 (a) bottom and Fig. 2 (a)).", + "bbox": [ + 212, + 325, + 782, + 385 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Based on these observations, we raise the question: Is there a unified directional parameterization that can effectively handle both specular surfaces and complex geometry? To answer this, we start by analyzing the reasons behind their reconstruction failures:", + "bbox": [ + 212, + 386, + 784, + 445 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Analysis of failures when using viewing direction. An analysis of why using viewing direction is unsuitable for modeling the view-dependent appearance of specular objects is provided in Verbin et al. [28], which points out that the radiance network conditioned on viewing direction is poorly suited for interpolation. We consider this characteristic leads to difficulties in correctly decoupling the geometry and appearance of specular objects when using viewing direction. On the other hand, reflection directional parameterization can effectively leverage the prior knowledge of the interaction between light and surface, which facilitates the correct decoupling of geometry and appearance, especially for specular objects.", + "bbox": [ + 212, + 446, + 784, + 595 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Analysis of failures when using reflection direction. In contrast, no work has yet provided a detailed analysis of why using reflection directional parameterization fails in the reconstruction of concave or complex structures, making us the first to provide this analysis. We first highlight two key differences between viewing direction and reflection direction: the viewing direction is non-learnable throughout the optimization and remains constant along a sampled ray, whereas the reflection direction 1) depends on learnable geometry (i.e., normals) and 2) may vary significantly during the optimization. We argue that these two properties of the reflection direction may pose challenges to the accurate recovery of geometry, especially for concave or complex structures.", + "bbox": [ + 212, + 598, + 784, + 747 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "A 2D toy example is first illustrated on Fig. 3 (left) to better demonstrate why the first property of reflection direction may hinder the recovery of geometry, which assumes the scenario of recovering an \"L\"-shaped concave structure. A sampled ray intersecting the surface and the sampled points on the ray along with the normals at these points are illustrated. In this case, we notice that the normals at sampled points that are slightly distant from the intersecting surface", + "bbox": [ + 212, + 750, + 784, + 839 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "Jiang and Xu et al.", + "bbox": [ + 271, + 114, + 401, + 127 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/17c616476e02fbcff23c408146bd60f710bf6cbdf8d8537f15742b2db917e0fe.jpg", + "image_caption": [ + "Fig. 3: Impact of incorporating normals on view-dependent radiance modeling. (Left) Normals at sampled points may be influenced by unrelated surfaces other than the intersection surface. (Top right) Smooth surfaces yield similarity in normals and reflection directions for the sampled points, while (Bottom right) surfaces with intricate local details, e.g., concavities, induce a scattered distribution of normals and reflection directions (Ref. dirs), particularly pronounced slightly distant from the vicinity of the zero-level set, which adversely affects geometry optimization. The details are explained in Sec. 3.2." + ], + "image_footnote": [], + "bbox": [ + 220, + 148, + 509, + 359 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c16c9a63a54b5ce76d95767850412eb63dfa57439162f10c8349a42fe9dade6e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 527, + 148, + 720, + 359 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f55cc30f7ed1382eecb8ca217173167ca976b89be02017cc7a7f3afab754bb3f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 736, + 148, + 781, + 362 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "do not necessarily relate to the normal of the intersecting surface (the bottom side of the zero level-set), but rather to the normals of surfaces elsewhere (the left side of the zero level-set). As a result, since the training loss of existing neural implicit surface reconstruction methods is usually ray-based, which is calculated using the local color of the intersecting surface, using this loss to optimize the normals of irrelevant surfaces elsewhere may introduce ambiguity in the optimization of geometry and result in incorrect geometric updates.", + "bbox": [ + 212, + 521, + 787, + 627 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In addition, the second property of the reflection direction is more closely related to the phenomenon that using reflection directional parameterization tends to reconstruct overly smooth surfaces with missing concave or complex structures. Assuming that the geometry reconstruction begins with a smooth surface exhibiting local flatness (Fig. 3 top-right), which aligns with the geometric initialization practices commonly employed by existing neural implicit surface reconstruction methods [6,29], and the geometry of the target object includes concave structures (Fig. 3 bottom-right). We observed that for the same sampled ray, when the surface geometry is updated from a smooth surface to a concave surface, the normals and reflection directions of the same sampled points also change. This change is particularly pronounced for points that are slightly distant from the surface, and could be further pronounced due to the presence of Positional Encoding (PE). The noticeable changes in reflection directions at these points within each surface update introduce high-frequency variances to", + "bbox": [ + 212, + 628, + 787, + 840 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "Rethinking directional parameterization", + "bbox": [ + 464, + 114, + 730, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 114, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "the directional inputs of the radiance network. This makes learning a stable radiance field challenging. To be specific, assuming that the radiance network conditioned on reflection direction has already learned to minimize the color loss to some extent before the surface update, the substantial change in directional inputs due to the update could likely lead to an increase in color loss, which hinders the optimization and is susceptible to falling into local minima.", + "bbox": [ + 212, + 146, + 787, + 236 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In contrast, we notice that using the viewing direction often leads to concave or overly intricate surface details (although some may be incorrect), as shown in Fig. 1 (a) and Fig. 2 (a) (c). We consider that this is because the viewing direction does not rely on the surface geometry, which means that changes in geometry do not affect the optimization of the radiance network conditioned on viewing direction. This allows the SDF can undergo more \"aggressive\" updates without considering the impact on radiance.", + "bbox": [ + 212, + 238, + 787, + 343 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "3.3 Hybrid Directional Parameterization", + "text_level": 1, + "bbox": [ + 214, + 364, + 565, + 378 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Building upon the analysis from the previous section, we discovered that viewing direction and reflection direction seem to have a spatially complementary relationship: On one hand, for sampled points near the intersecting surface, the normals here accurately reflect the normals of the intersecting surface, thereby enabling the reflection direction to better model the interaction between light and surface, which in turn promotes the correct decomposition of geometry and appearance. On the other hand, for other sampled points along the ray, using the reflection direction might introduce irrelevant surface geometry into the optimization, or high-frequency variations in the directional inputs of the radiance network, both potentially hindering the optimization of the geometry and radiance networks. By contrast, using the viewing direction at these sampled points can avoid these issues. Motivated by these thoughts, we propose a spatial-aware directional representation, which transitions from reflection direction $\\mathbf{d}_{\\mathrm{ref}}$ to viewing direction $\\mathbf{d}_{\\mathrm{view}}$ based on the distance to the surface (i.e., the absolute value of SDF). We refer to this fused direction as hybrid direction $\\mathbf{d}_{\\mathrm{hyb}}$ , formulated as:", + "bbox": [ + 212, + 388, + 787, + 630 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {d} _ {\\mathrm {h y b}} = \\text {n o r m a l i z e} (\\alpha \\cdot \\mathbf {d} _ {\\text {r e f}} + (1 - \\alpha) \\cdot \\mathbf {d} _ {\\text {v i e w}}). \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 346, + 646, + 784, + 662 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Here, $\\alpha \\in [0,1]$ represents the blend weight. We normalize the resulting linear blend to a unit vector thereby representing a valid direction. The blend weight $\\alpha$ is implemented as", + "bbox": [ + 212, + 666, + 787, + 710 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\alpha = \\exp (- \\gamma \\cdot \\det a c h (| f (\\mathbf {x}) |)), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 397, + 724, + 784, + 739 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $\\gamma \\in \\mathbb{R}^{+}$ is an estimable parameter that controls the transition speed between reflection/viewing direction. The gradient of the SDF is detached to prevent interference with geometric updates. At the surface of the object, i.e., when SDF equals 0, $\\alpha$ simply becomes 1, which means that the hybrid direction aligns with the reflection direction. As $\\mathbf{x}$ moves further away from the surface, $\\alpha$ gradually approaches 0, indicating that the direction tends towards the viewing", + "bbox": [ + 212, + 750, + 787, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "Jiang and Xu et al.", + "bbox": [ + 271, + 114, + 401, + 128 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "direction. Here, $\\alpha$ can be any bell-shaped function w.r.t. the SDF value. In this case, we utilize the PDF of the Laplace distribution for simplicity.", + "bbox": [ + 212, + 146, + 782, + 176 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "In the next section, extensive experimental results validate the superiority of the proposed hybrid directional parameterization compared to existing parameterizations.", + "bbox": [ + 212, + 176, + 784, + 220 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 214, + 243, + 375, + 258 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.1 Experimental Settings", + "text_level": 1, + "bbox": [ + 214, + 273, + 444, + 287 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Datasets. We conduct experiments on objects from different datasets, including the DTU dataset [1], the Shiny Blender dataset [28], and the real captured dataset [9]. The DTU dataset [1] contains indoor multi-view captures of diverse objects, including challenging conditions for geometry reconstruction such as non-Lambertian surfaces and complex structures. We use the same 15 objects as previous works [2, 29, 31, 36] for evaluation. In addition, we utilize the foreground masks provided by Yariv et al. [36]. The Shinny Blender dataset [28] is a synthetic dataset rendered in Blender [4], which contains different objects with specular surfaces. We follow the experimental settings used in Ref-NeuS [8] on this dataset, which selects four objects. The real captured dataset [9] contains a set of three outdoor captures of real objects with complex geometry and specular surfaces. We select the \"toycar\" scene for evaluation.", + "bbox": [ + 212, + 296, + 784, + 474 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Implementation details. We integrate the proposed hybrid directional parameterization into two types of neural surface reconstruction methods: NeuS [29] and NeuDA [2]. The former utilizes a deep MLP network as the scene representation, while the latter leverages a shallow MLP decoder combined with multi-level voxel grids. Both methods adopt a radiance network conditioned on viewing direction in their original versions. The modification of the directional input from the viewing direction to the proposed hybrid direction can be accomplished with less than 10 lines of code. We additionally implement the radiance network that depends on reflection direction for both methods, to facilitate a comprehensive comparison of different directional parameterizations. Regarding the learnable parameter $\\gamma$ in Eqn. 3, we implement it as $\\gamma = \\exp (10\\cdot \\gamma_b)$ . We assign an initial value of 0.3 to $\\gamma_{b}$ except for the \"toaster\" scene from the Shiny Blender dataset. For the \"toaster\" scene, an initial value of 0.1 is assigned since it results in better reconstruction. The training losses of the modified methods are identical to those of their original versions. The number of training iterations is set to 300,000 to ensure convergence.", + "bbox": [ + 212, + 478, + 784, + 717 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Evaluation metrics. For the DTU dataset, we follow the evaluation protocol of previous works [2, 29] and evaluate the Chamfer Distance (CD) between the estimated and ground-truth geometry. For the Shiny Blender dataset, we follow the evaluation protocol proposed in Ref-NeuS [8]. It evaluates the accuracy (Acc.) instead of the CD for unbiased evaluation, and also evaluates the mean angular error (MAE) between the ground-truth normal maps and the rendered normal maps. For the real captured dataset [9], we only provide a qualitative evaluation, as ground-truth geometry is not provided in this dataset.", + "bbox": [ + 212, + 719, + 785, + 839 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "Rethinking directional parameterization", + "bbox": [ + 464, + 114, + 730, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 114, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/1108d15f71b95c0b65361450d355052ac603fb0391552f7e8c23c9a1d7300e72.jpg", + "table_caption": [ + "Table 1: Quantitative evaluation in terms of Chamfer distance of integrating different directional parameterizations in NeuS [29] and NeuDA [2] on the DTU dataset [1]. The experimental objects are categorized into diffuse objects and specular objects. The numbers in the headers denote the scene IDs. Bold results have the best score." + ], + "table_footnote": [], + "table_body": "
MethodsScan of Diffuse ObjectsScan of Specular ObjectsTotal Avg.
243740556583105106114118122Avg.636997110Avg.
NeuS [29]0.830.980.560.370.591.450.780.520.360.450.450.671.130.600.951.431.030.77
NeuS, w/ reflection dir.1.394.390.730.391.341.490.920.590.360.460.541.141.110.631.460.921.031.11
NeuS, w/ hybrid dir. (Ours)0.810.970.540.380.611.470.770.510.350.450.470.671.100.600.850.900.860.72
NeuD A [2]0.510.760.390.370.561.370.790.500.340.420.460.591.080.571.130.800.900.67
NeuD A, w/ reflection dir.0.581.030.630.370.791.450.890.510.370.430.500.691.030.550.900.760.810.72
NeuD A, w/ hybrid dir. (Ours)0.530.770.400.360.571.380.760.470.330.410.470.591.020.540.760.750.770.63
", + "bbox": [ + 215, + 212, + 784, + 299 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.2 Comparison of Different Directional Parameterizations", + "text_level": 1, + "bbox": [ + 215, + 327, + 712, + 343 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "We first evaluate the performance of integrating different directional parameterizations into the two baseline methods NeuS [29] and NeuDA [2] on the DTU dataset [1]. The experimental objects from the DTU dataset are categorized into diffuse objects and specular objects for better demonstration. Tab. 1 summarizes the quantitative results. For both baseline methods, we notice that substituting the original viewing directional parameterization with the reflection direction only yields better results for a minority of specular objects. For diffuse objects and the remaining specular objects, employing reflection directional parameterization demonstrates significant degradation in performance. Considering the relatively complex geometry of DTU objects, we attribute this to the optimization issue analyzed in Sec. 3.2. Conversely, the incorporation of the proposed hybrid directional parameterization enables both modified methods to yield consistently superior performance for specular objects while achieving results on par with the original version for diffuse objects. The first two rows in Fig. 4 showcase the visual comparison of adopting different directional parameterizations for the reconstruction of two specular objects from the DTU dataset. Both baseline methods fail to reconstruct the specular surfaces of objects. On the other hand, although integrating the reflection direction yields better reconstruction for specular objects with simple geometry (e.g., Scan 63), they fail to reconstruct the specular surfaces at the concave structure in Scan 97. In contrast, integrating the proposed hybrid direction achieves the best reconstruction results in both scenarios. More qualitative comparisons on the DTU dataset are available in our supplementary.", + "bbox": [ + 212, + 356, + 787, + 704 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "We further validate the advantage of the proposed hybrid directional parameterization on the Shiny Blender dataset [28], as demonstrated in Tab. 2. Both baseline methods exhibit large reconstruction errors on the surfaces of highly specular objects in this dataset. By contrast, due to the relatively simple geometry of objects, replacing the viewing directional parameterization with the reflection direction results in a significant improvement in reconstruction quality on most objects in this dataset, except for the \"coffee\" object. For the reconstruction of the \"coffee\" object, adopting the reflection directional parameterization fails to recover the concave structure formed by the inner wall of the coffee cup", + "bbox": [ + 212, + 704, + 787, + 839 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "Jiang and Xu et al.", + "bbox": [ + 271, + 114, + 401, + 128 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/329faf4da4cbd1d40208b8888610f70f109bdd77708718a0927605640dcb04a2.jpg", + "table_caption": [ + "Table 2: Quantitative evaluation of integrating different directional parameterizations in NeuS [29] and NeuDA [2] on the Shiny Blender dataset [28]. Bold results have the best score." + ], + "table_footnote": [], + "table_body": "
MethodshelmettoastercoffeecarAvg.
Acc. ↓MAE ↓Acc. ↓MAE ↓Acc. ↓MAE ↓Acc. ↓MAE ↓Acc. ↓MAE ↓
NeuS [29]4.883.203.312.851.971.060.860.952.762.02
NeuS, w/ reflection dir.0.310.360.390.943.301.370.480.731.120.85
NeuS, w/ hybrid dir. (Ours)0.290.340.450.920.640.550.430.720.450.63
NeuDA [2]5.733.775.983.391.160.810.871.033.442.25
NeuDA, w/ reflection dir.0.340.371.031.713.661.530.620.871.411.12
NeuDA, w/ hybrid dir. (Ours)0.320.360.951.540.550.570.600.840.610.83
", + "bbox": [ + 215, + 198, + 784, + 303 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/131a8fe431a68b9861d9cfbec23962ca947dcc1212996f54c71bb1d6babac13d.jpg", + "image_caption": [ + "Fig. 4: Qualitative comparison of integrating different directional parameterizations in NeuS [29] and NeuDA [2] on the DTU dataset [1] and Shiny Blender dataset [28]." + ], + "image_footnote": [], + "bbox": [ + 218, + 313, + 781, + 526 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "and the surface of the water, as shown in the third row in Fig. 4. On the other hand, the integration of our proposed hybrid directional parameterization enables the successful reconstruction of both the specular surfaces and concave structures, obtaining consistently superior reconstructions of all objects in this dataset, compared to the other two directional parameterizations.", + "bbox": [ + 212, + 595, + 787, + 672 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Lastly, we qualitatively present the reconstruction results on the \"toycar\" scene from the real captured scene in Fig. 5, which contains challenging reflective glass and complex geometry. Both modified methods employing the proposed hybrid directional parameterization succeed in recovering the complex geometry and the smooth surface of the reflective glass. In contrast, using either the reflection direction or the original viewing direction results in reconstructions that are either overly smooth, missing concave surface structures, or provide incorrect surface reconstructions at the reflective glass areas.", + "bbox": [ + 212, + 672, + 787, + 792 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "As the summary of this section, compared to the viewing directional parameterization which struggles with reconstructing specular surfaces, and the reflection directional parameterization which faces difficulties in reconstructing", + "bbox": [ + 212, + 794, + 787, + 840 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "Rethinking directional parameterization", + "bbox": [ + 464, + 114, + 730, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/f1ddbdcf03a7995ee9e1642aeef2d91281a9b0a40740e67696defd744c005857.jpg", + "image_caption": [ + "Reference RGB" + ], + "image_footnote": [], + "bbox": [ + 222, + 185, + 349, + 257 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/582c6f6b5e5bc4dc73ab6d6a7068e1ff9cbf15e31e60d7671bb241236f142f2f.jpg", + "image_caption": [ + "NeuS" + ], + "image_footnote": [], + "bbox": [ + 357, + 147, + 485, + 215 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/f65ffc2a1635eb666adf303450035e9005d10ed0167b153c7afc7a996e1bf67e.jpg", + "image_caption": [ + "NeuS, w/ reflection dir." + ], + "image_footnote": [], + "bbox": [ + 498, + 148, + 625, + 215 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/36b7a36abd21e147b2ba39440f62dcf822f3d3bc998ce6ffe72a4e3588a1d0d5.jpg", + "image_caption": [ + "NeuS, w/ hybrid dir." + ], + "image_footnote": [], + "bbox": [ + 640, + 152, + 767, + 215 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/647d7e95eb8b2b34ddcdd9114d4a5ce8a9343e5e8b5a066952e0841bd9e7313b.jpg", + "image_caption": [ + "NeuDA" + ], + "image_footnote": [], + "bbox": [ + 357, + 232, + 485, + 297 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/9c12491570133dd22f48aceeb65c46207a02faa25a6eaedfb1e31a92cbbf0ab3.jpg", + "image_caption": [ + "NeuDA,w/ reflection dir." + ], + "image_footnote": [], + "bbox": [ + 498, + 233, + 625, + 297 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/62708bc77f6e852906135c5fe5ca35789b3e6ad45b2ca7af5ca2303158c6605a.jpg", + "image_caption": [ + "NeuDA, w/ hybrid dir.", + "Fig. 5: Qualitative comparison of integrating different directional parameterizations in NeuS [29] and NeuDA [2] on the \"toycar\" scene from the real captured dataset [9]." + ], + "image_footnote": [], + "bbox": [ + 640, + 233, + 767, + 297 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/d938a2f05cc116196255e7a42fe61f2faa01e6cb5726e0b8003cd95141b2831f.jpg", + "table_caption": [ + "Table 3: Quantitative evaluation in terms of Chamfer distance on the DTU dataset [1]. The numbers in the header denote the scene IDs. Bold results have the best score." + ], + "table_footnote": [], + "table_body": "
Methods243740556365698397105106110114118122Avg.
IDR [36]1.631.870.630.481.040.790.771.331.160.760.670.900.420.510.530.90
NeuS [29]0.830.980.560.371.130.590.601.450.950.780.521.430.360.450.450.77
Voxurf [31]0.650.740.390.350.960.640.851.581.010.680.601.110.370.450.470.72
NeuDA [2]0.510.760.390.371.080.560.571.371.130.790.500.800.340.420.460.67
Ours (NeuS)0.810.970.540.381.100.610.601.470.850.770.510.900.350.450.470.72
Ours (NeuDA)0.530.770.400.361.020.570.541.380.760.760.470.760.330.410.470.63
", + "bbox": [ + 217, + 407, + 782, + 502 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "complex structures, our proposed hybrid directional parameterization effectively handles both scenarios in a unified form. This benefit is attributed to our hybrid directional parameterization effectively combining the advantages of both viewing and reflection directional parameterizations while avoiding their drawbacks.", + "bbox": [ + 212, + 530, + 782, + 590 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "4.3 Comparison to State-of-the-Art", + "text_level": 1, + "bbox": [ + 214, + 614, + 522, + 630 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Although our main purpose is not to develop a best-performing surface reconstruction system, we demonstrate that we can still achieve state-of-the-art performance by integrating the proposed hybrid directional parameterization into existing methods. For brevity, we refer to NeuS [29] and NeuDA [2] incorporating the proposed hybrid directional parameterization as Ours (NeuS) and Ours (NeuDA) in the following text. Tab. 3 reports the Chamfer distance (CD) metrics of different methods on the DTU dataset [1]. For a fair comparison, all methods in Tab. 3 utilize foreground masks provided by Yariv et al. [36] during training. Both Ours (NeuS) and Ours (NeuDA) improve their baselines by a certain margin in the average CD scores, and Ours (NeuDA) achieves the best performance compared to previous SOTA methods in the average CD metric.", + "bbox": [ + 212, + 643, + 785, + 808 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Tab. 4 reports the quantitative evaluation of different methods on the Shiny Blender dataset [28]. The results of [5,6,8,22,28,35] are taken from the results", + "bbox": [ + 214, + 809, + 785, + 839 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "Jiang and Xu et al.", + "bbox": [ + 271, + 114, + 401, + 127 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/1af0e3dc63bbc16942d2780aaf688bc5559f798bdb69d17e7646ddd6ee8f97bf.jpg", + "table_caption": [ + "Table 4: Quantitative evaluation on the Shiny Blender dataset [28]. “-” denotes that the results are not available. Bold results have the best score." + ], + "table_footnote": [], + "table_body": "
MethodshelmettoastercoffeecarAvg.
Acc. ↓MAE ↓Acc. ↓MAE ↓Acc. ↓MAE ↓Acc. ↓MAE ↓Acc. ↓MAE ↓
Ref-NeRF [28]-29.48-42.87-12.24-14.93-24.88
NeuS [29]4.883.203.312.851.971.060.860.952.762.02
UNISURF [22]1.691.783.755.512.883.152.341.982.673.11
VolSDF [35]1.551.372.022.532.232.280.581.041.601.81
NeuralWarp [5]2.251.945.903.511.542.040.651.072.592.14
Geo-NeuS [6]0.742.366.353.763.856.362.885.673.464.54
NeuD A [2]5.733.775.983.391.160.810.871.033.442.25
Ref-NeuD [8]0.290.380.421.470.770.990.370.800.460.91
Ours (NeuD)0.290.340.450.920.640.550.430.720.450.63
Ours (NeuD A)0.320.360.951.540.550.570.600.840.610.83
", + "bbox": [ + 272, + 184, + 727, + 325 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/5345e63abcb455c37bf47dc647cd2c32a47dc52cfa5b3d647f27a5bc83c438c9.jpg", + "image_caption": [ + "Reference RGB" + ], + "image_footnote": [], + "bbox": [ + 217, + 339, + 292, + 393 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/7b63b837f0ae60ce15848f50b4d247b11f3b67130e40ea30028d52c3ed0b270d.jpg", + "image_caption": [ + "$\\gamma_{b} = 0.0$ $\\mathrm{CD}\\downarrow$ : 0.91" + ], + "image_footnote": [], + "bbox": [ + 295, + 339, + 392, + 393 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/b86a875fb70518f84ec19ada47e3191ac3f37279be34cf98e4fed23a0161d205.jpg", + "image_caption": [ + "Fig. 6: Reconstruction results using different initial values of $\\gamma_{b}$ . The main differences are highlighted in red boxes." + ], + "image_footnote": [], + "bbox": [ + 395, + 339, + 488, + 393 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/98e8e6462d067ee35fef37e76fa72e6daeb2e2b6dd599acd77ca342c3e3b8c5c.jpg", + "image_caption": [ + "$\\gamma_{b} = 0.1$ $\\mathrm{CD}\\downarrow$ : 0.81" + ], + "image_footnote": [], + "bbox": [ + 491, + 339, + 586, + 393 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/991b02be9d23b8e105ad4d35305552bc4109874e2c23fb513bfd34ec7ca08da8.jpg", + "image_caption": [ + "$\\gamma_{b} = 0.2$ $\\mathrm{CD}\\downarrow$ :0.80", + "$\\gamma_{b} = 0.3$ $\\mathrm{CD}\\downarrow$ :0.76" + ], + "image_footnote": [], + "bbox": [ + 589, + 339, + 686, + 393 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/815bd1c01f24cef8d31899ddd76bb93ec6cc74473784b89d674824c11024653d.jpg", + "image_caption": [ + "$\\gamma_{b} = 0.4$ $\\mathrm{CD}\\downarrow$ :0.77" + ], + "image_footnote": [], + "bbox": [ + 689, + 339, + 784, + 393 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "reported in Ge et al. [8]. Both Ours (NeuS) and Ours (NeuDA) achieve better average MAE metrics than the previous SOTA method Ref-NeuS [8], and Ours (NeuS) achieve the best overall performance in terms of both Acc. and MAE metrics on this dataset. We refer to our supplementary for the qualitative comparisons between Ours (NeuS) and Ref-NeuS [8].", + "bbox": [ + 212, + 473, + 787, + 551 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "4.4 Ablation Studies", + "text_level": 1, + "bbox": [ + 215, + 571, + 400, + 585 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "In this section, we conduct ablation studies on several key design choices of the proposed hybrid directional parameterization by using Ours (NeuDA).", + "bbox": [ + 212, + 598, + 784, + 627 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "The initial value of $\\gamma_{b}$ . We first validate the impact of setting different initial values for $\\gamma_{b}$ on the reconstruction results. Fig. 6 compares the obtained reconstruction results using five different initial values for $\\gamma_{b}$ . We find that when the initial value of $\\gamma_{b}$ is set low, our method tends to recover incorrect geometries similar to those caused by using the reflection directional parameterization. This phenomenon remains consistent with our analysis in Sec 3.2, as a smaller initial value for $\\gamma_{b}$ increases the likelihood of involving irrelevant surfaces and high-frequency variations in the directional inputs of the radiance network during the optimization. We empirically find the best initial value of $\\gamma_{b}$ to be 0.3 for most experimental objects. More discussions regarding the parameter $\\gamma_{b}$ and blend weight $\\alpha$ can be found in our supplementary material.", + "bbox": [ + 212, + 628, + 785, + 794 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Detach SDF gradient in Eqn. 3 or not? Fig. 7 compares the reconstruction results with and without detaching the gradient of the estimated SDF in Eqn. 3. We observe that our method struggles to reconstruct the fine-grained geometry", + "bbox": [ + 212, + 795, + 785, + 840 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "Rethinking directional parameterization", + "bbox": [ + 464, + 114, + 730, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 784, + 126 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/37d1e782e6978f39ae6919287d644778fffde258e7ada473f5099d1f3ee9e413.jpg", + "image_caption": [ + "Reference RGB" + ], + "image_footnote": [], + "bbox": [ + 217, + 143, + 325, + 207 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/f1e1b747b70a6657ce267fb061d7a820fb68314fbff530d15311cb55ebbc9f66.jpg", + "image_caption": [ + "w/o SDF gradient detachment" + ], + "image_footnote": [], + "bbox": [ + 326, + 143, + 410, + 205 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/841dfae42ca3709046551a7fd75a51efdb527306de2daec685287e3580f7cc6a.jpg", + "image_caption": [ + "w/SDF gradient detachment (Ours)" + ], + "image_footnote": [], + "bbox": [ + 413, + 145, + 496, + 207 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/8ec8eccae10f1f88e6ddac6b1efbabea238d4e8c59eaa6d5848f20711b4a2985.jpg", + "image_caption": [ + "Reference RGB", + "Fig.7: Reconstruction results with and without detaching the gradient of the estimated SDF in Eqn. 3." + ], + "image_footnote": [], + "bbox": [ + 519, + 152, + 599, + 214 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/f25fc9a33d52989287e48f6ee97beba950a6a8a20200368881dc50c2a4f9422b.jpg", + "image_caption": [ + "post-PE fusion", + "Fig. 8: Reconstruction results fusing directions before and after positional encoding (PE)." + ], + "image_footnote": [], + "bbox": [ + 612, + 152, + 692, + 215 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/5ece8d59d27bd17e8e4cfbc09c3a462a2a4ea1c8a0114c686d9ca89c80482a6f.jpg", + "image_caption": [ + "pre-PE fusion (Ours)" + ], + "image_footnote": [], + "bbox": [ + 702, + 152, + 785, + 215 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "when the gradient of the estimated SDF is not detached in Eqn. 3. We consider that introducing the gradient of the estimated SDF places an additional burden on the learning of both the geometry and radiance networks, making it easier to fall into local minima. Thereby, we choose to detach the gradient of SDF in Eqn. 3.", + "bbox": [ + 212, + 306, + 784, + 382 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Fuse the viewing direction and reflection direction before or after positional encoding? We introduce the fusion of viewing direction and reflection direction in Eqn. 2 before the operation of position encoding (PE). We compare the reconstruction results with different fusion orders, specifically the proposed pre-PE fusion: $\\mathrm{PE}(\\mathbf{d}_{\\mathrm{hyb}})$ (ours) and alternative approach using post-PE fusion: $\\alpha \\cdot \\mathrm{PE}(\\mathbf{d}_{\\mathrm{ref}}) + (1 - \\alpha) \\cdot \\mathrm{PE}(\\mathbf{d}_{\\mathrm{view}})$ . The comparison is shown in Fig. 8. We observe that fusing directions after the PE operation may result in degraded reconstruction results. We consider that fusing the directional information in the high-dimension space may increase the complexity of the learning space of the radiance network, which brings challenges in finding optimal reconstruction results. Thereby, we choose to fuse the viewing direction and reflection direction before the operation of positional encoding.", + "bbox": [ + 212, + 383, + 787, + 565 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Conclusion and Limitation", + "text_level": 1, + "bbox": [ + 214, + 584, + 509, + 599 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In this paper, we conduct an in-depth investigation into the topic of directional parameterization in neural implicit surface reconstruction. We reveal that the commonly used directional parameterizations in modeling the view-dependent radiance field, such as viewing and reflection directions, both have their limitations and none of them can effectively handle both specular surfaces and complex structures. A novel hybrid directional parameterization is introduced to achieve this goal, which softly fuses the viewing and reflection directions based on the distance of the sampled points from the surface, uniting the advantages of each representation. Extensive experiments demonstrate that our hybrid directional parameterization outperforms the existing directional parameterizations in reconstructing objects with a wide variety of materials, geometry, and appearance.", + "bbox": [ + 212, + 613, + 787, + 779 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "One limitation of our method is that the optimal choice of initial value for $\\gamma_{b}$ may vary depending on the target object, which would require additional efforts for parameter tuning. An automated method for selecting the optimal initial value of $\\gamma_{b}$ for target objects could be a potential future work for us.", + "bbox": [ + 212, + 779, + 787, + 839 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "Jiang and Xu et al.", + "bbox": [ + 271, + 114, + 401, + 127 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 143, + 321, + 159 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Aanaes, H., Jensen, R.R., Vogiatzis, G., Tola, E., Dahl, A.B.: Large-scale data for multiple-view stereopsis. IJCV 120, 153-168 (2016)", + "2. Cai, B., Huang, J., Jia, R., Lv, C., Fu, H.: Neuda: Neural deformable anchor for high-fidelity implicit surface reconstruction. In: CVPR (2023)", + "3. Chen, Z., Zhang, H.: Learning implicit fields for generative shape modeling. In: CVPR (2019)", + "4. Community, B.O.: Blender - a 3D modelling and rendering package. Blender Foundation, Stichting Blender Foundation, Amsterdam (2018)", + "5. Darmon, F., Bascle, B., Devaux, J., Monasse, P., Aubry, M.: Improving neural implicit surfaces geometry with patch warping. In: CVPR (2022)", + "6. Fu, Q., Xu, Q., Ong, Y.S., Tao, W.: Geo-neus: Geometry-consistent neural implicit surfaces learning for multi-view reconstruction. In: NeurIPS (2022)", + "7. Furukawa, Y., Hernández, C., et al.: Multi-view stereo: A tutorial. Foundations and Trends® in Computer Graphics and Vision 9(1-2), 1-148 (2015)", + "8. Ge, W., Hu, T., Zhao, H., Liu, S., Chen, Y.C.: Ref-neus: Ambiguity-reduced neural implicit surface learning for multi-view reconstruction with reflection. In: ICCV (2023)", + "9. Hedman, P., Srinivasan, P.P., Mildenhall, B., Barron, J.T., Debevec, P.: Baking neural radiance fields for real-time view synthesis. In: CVPR (2021)", + "0. Kato, H., Beker, D., Morariu, M., Ando, T., Matsuoka, T., Kehl, W., Gaidon, A.: Differentiable rendering: A survey. arXiv preprint arXiv:2006.12057 (2020)", + "1. Li, T.M., Aittala, M., Durand, F., Lehtinen, J.: Differentiable monte carlo ray tracing through edge sampling. ACM Transactions on Graphics 37(6), 1-11 (2018)", + "2. Li, Z., Müller, T., Evans, A., Taylor, R.H., Unberath, M., Liu, M.Y., Lin, C.H.: Neuralangelo: High-fidelity neural surface reconstruction. In: CVPR (2023)", + "3. Liang, R., Chen, H., Li, C., Chen, F., Panneer, S., Vijaykumar, N.: Envidr: Implicit differentiable renderer with neural environment lighting. arXiv preprint arXiv:2303.13022 (2023)", + "4. Liu, Y., Wang, P., Lin, C., Long, X., Wang, J., Liu, L., Komura, T., Wang, W.: Nano: Neural geometry and brdf reconstruction of reflective objects from multiview images. ACM Transactions on Graphics (2023)", + "5. Long, X., Lin, C., Wang, P., Komura, T., Wang, W.: Sparseneus: Fast generalizable neural surface reconstruction from sparse views. In: ECCV (2022)", + "6. Mescheder, L., Oechsle, M., Niemeyer, M., Nowozin, S., Geiger, A.: Occupancy networks: Learning 3d reconstruction in function space. In: CVPR (2019)", + "7. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. In: ECCV (2020)", + "8. Müller, T., Evans, A., Schied, C., Keller, A.: Instant neural graphics primitives with a multiresolution hash encoding. ACM Transactions on Graphics 41(4), 1-15 (2022)", + "9. Niemeyer, M., Mescheder, L., Oechsle, M., Geiger, A.: Occupancy flow: 4d reconstruction by learning particle dynamics. In: ICCV (2019)", + "20. Niemeyer, M., Mescheder, L., Oechsle, M., Geiger, A.: Differentiable volumetric rendering: Learning implicit 3d representations without 3d supervision. In: CVPR (2020)", + "21. Niemeyer, M., Mescheder, L., Oechsle, M., Geiger, A.: Differentiable volumetric rendering: Learning implicit 3d representations without 3d supervision. In: CVPR (2020)" + ], + "bbox": [ + 225, + 172, + 784, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "Rethinking directional parameterization", + "bbox": [ + 464, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "22. Oechsle, M., Peng, S., Geiger, A.: Unisurf: Unifying neural implicit surfaces and radiance fields for multi-view reconstruction. In: ICCV (2021)", + "23. Park, J.J., Florence, P., Straub, J., Newcombe, R., Lovegrove, S.: Deepsdf: Learning continuous signed distance functions for shape representation. In: CVPR (2019)", + "24. Peng, S., Niemeyer, M., Mescheder, L., Pollefeys, M., Geiger, A.: Convolutional occupancy networks. In: ECCV (2020)", + "25. Ravi, N., Reizenstein, J., Novotny, D., Gordon, T., Lo, W.Y., Johnson, J., Gkioxari, G.: Accelerating 3d deep learning with pytorch3d. arXiv preprint arXiv:2007.08501 (2020)", + "26. Srinivasan, P.P., Deng, B., Zhang, X., Tancik, M., Mildenhall, B., Barron, J.T.: Nerv: Neural reflectance and visibility fields for relighting and view synthesis. In: CVPR (2021)", + "27. Tulsiani, S., Zhou, T., Efros, A.A., Malik, J.: Multi-view supervision for single-view reconstruction via differentiable ray consistency. In: CVPR (2017)", + "28. Verbin, D., Hedman, P., Mildenhall, B., Zickler, T., Barron, J.T., Srinivasan, P.P.: Ref-nerf: Structured view-dependent appearance for neural radiance fields. In: CVPR (2022)", + "29. Wang, P., Liu, L., Liu, Y., Theobalt, C., Komura, T., Wang, W.: Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. In: NeurIPS (2021)", + "30. Wang, Y., Han, Q., Habermann, M., Daniilidis, K., Theobalt, C., Liu, L.: Neus2: Fast learning of neural implicit surfaces for multi-view reconstruction. In: ICCV (2023)", + "31. Wu, T., Wang, J., Pan, X., Xu, X., Theobalt, C., Liu, Z., Lin, D.: Voxurf: Voxel-based efficient and accurate neural surface reconstruction. In: ICLR (2023)", + "32. Xie, Y., Takikawa, T., Saito, S., Litany, O., Yan, S., Khan, N., Tombari, F., Tompkin, J., Sitzmann, V., Sridhar, S.: Neural fields in visual computing and beyond. Computer Graphics Forum 41(2), 641-676 (2022)", + "33. Yan, X., Yang, J., Yumer, E., Guo, Y., Lee, H.: Perspective transformer nets: Learning single-view 3d object reconstruction without 3d supervision. In: NeurIPS (2016)", + "34. Yao, Y., Zhang, J., Liu, J., Qu, Y., Fang, T., McKinnon, D., Tsin, Y., Quan, L.: Neilf: Neural incident light field for material and lighting estimation. In: ECCV (2022)", + "35. Yariv, L., Gu, J., Kasten, Y., Lipman, Y.: Volume rendering of neural implicit surfaces. In: NeurIPS (2021)", + "36. Yariv, L., Kasten, Y., Moran, D., Galun, M., Atzmon, M., Ronen, B., Lipman, Y.: Multiview neural surface reconstruction by disentangling geometry and appearance. In: NeurIPS (2020)", + "37. Yu, A., Ye, V., Tancik, M., Kanazawa, A.: pixelnerf: Neural radiance fields from one or few images. In: CVPR (2021)", + "38. Yu, Z., Peng, S., Niemeyer, M., Sattler, T., Geiger, A.: Monosdf: Exploring monocular geometric cues for neural implicit surface reconstruction. In: NeurIPS (2022)", + "39. Zhang, J., Yao, Y., Li, S., Liu, J., Fang, T., McKinnon, D., Tsin, Y., Quan, L.: Neilf++: Inter-reflectable light fields for geometry and material estimation (2023)", + "40. Zhang, K., Riegler, G., Snavely, N., Koltun, V.: Nerf++: Analyzing and improving neural radiance fields. arXiv preprint arXiv:2010.07492 (2020)", + "41. Zhang, X., Srinivasan, P.P., Deng, B., Debevec, P., Freeman, W.T., Barron, J.T.: Nerfactor: Neural factorization of shape and reflectance under an unknown illumination. ACM Transactions on Graphics 40(6), 1-18 (2021)" + ], + "bbox": [ + 215, + 146, + 784, + 829 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "Jiang and Xu et al.", + "bbox": [ + 271, + 114, + 401, + 127 + ], + "page_idx": 15 + } +] \ No newline at end of file diff --git a/2024/Rethinking Directional Parameterization in Neural Implicit Surface Reconstruction/8b296952-4d44-42db-affd-9e9843feb2a6_model.json b/2024/Rethinking Directional Parameterization in Neural Implicit Surface Reconstruction/8b296952-4d44-42db-affd-9e9843feb2a6_model.json new file mode 100644 index 0000000000000000000000000000000000000000..6a586fb18c2f83da1d30aa1fe0166e3562560583 --- /dev/null +++ b/2024/Rethinking Directional Parameterization in Neural Implicit Surface Reconstruction/8b296952-4d44-42db-affd-9e9843feb2a6_model.json @@ -0,0 +1,2630 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.255, + 0.141, + 0.75, + 0.187 + ], + "angle": 0, + "content": "Rethinking Directional Parameterization in Neural Implicit Surface Reconstruction" + }, + { + "type": "text", + "bbox": [ + 0.304, + 0.212, + 0.699, + 0.228 + ], + "angle": 0, + "content": "Zijie Jiang\\(^{1\\star}\\), Tianhan Xu\\(^{2\\star}\\), and Hiroharu Kato\\(^{2\\star}\\)" + }, + { + "type": "text", + "bbox": [ + 0.394, + 0.239, + 0.61, + 0.254 + ], + "angle": 0, + "content": "\\(^{1}\\) Tokyo Institute of Technology" + }, + { + "type": "text", + "bbox": [ + 0.411, + 0.254, + 0.59, + 0.268 + ], + "angle": 0, + "content": "2 Preferred Networks, Inc." + }, + { + "type": "text", + "bbox": [ + 0.272, + 0.269, + 0.73, + 0.283 + ], + "angle": 0, + "content": "zjiang@ok.sc.e.titech.ac.jp, {xutianhan,hkato}@preferred.jp" + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.313, + 0.744, + 0.591 + ], + "angle": 0, + "content": "Abstract. Multi-view 3D surface reconstruction using neural implicit representations has made notable progress by modeling the geometry and view-dependent radiance fields within a unified framework. However, their effectiveness in reconstructing objects with specular or complex surfaces is typically biased by the directional parameterization used in their view-dependent radiance network. Viewing direction and reflection direction are the two most commonly used directional parameterizations but have their own limitations. Typically, utilizing the viewing direction usually struggles to correctly decouple the geometry and appearance of objects with highly specular surfaces, while using the reflection direction tends to yield overly smooth reconstructions for concave or complex structures. In this paper, we analyze their failed cases in detail and propose a novel hybrid directional parameterization to address their limitations in a unified form. Extensive experiments demonstrate the proposed hybrid directional parameterization consistently delivered satisfactory results in reconstructing objects with a wide variety of materials, geometry and appearance, whereas using other directional parameterizations faces challenges in reconstructing certain objects. Moreover, the proposed hybrid directional parameterization is nearly parameter-free and can be effortlessly applied in any existing neural surface reconstruction method." + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.603, + 0.741, + 0.631 + ], + "angle": 0, + "content": "Keywords: surface reconstruction \\(\\cdot\\) neural implicit representation \\(\\cdot\\) directional parameterization" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.655, + 0.377, + 0.67 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.683, + 0.788, + 0.806 + ], + "angle": 0, + "content": "Recent advances in neural fields [32] have revolutionized the field of 3D geometry reconstruction from multi-view images [22,29,36]. Compared to traditional multiview stereo methods [7], a key advantage of using neural fields is their ability to model and optimize complex light reflections on object surfaces, which improves robustness to specular or non-Lambertian surfaces. Since a radiance network is trained jointly with a geometry network via differentiable rendering, the radiance network is able to take full advantage of geometry and view-dependent information such as viewing direction to decouple the geometry and appearance." + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.811, + 0.787, + 0.839 + ], + "angle": 0, + "content": "* Equal contribution. Work done while Zijie Jiang interned at Preferred Networks, Inc." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.403, + 0.129 + ], + "angle": 0, + "content": "Jiang and Xu et al." + }, + { + "type": "image", + "bbox": [ + 0.232, + 0.156, + 0.318, + 0.221 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.347, + 0.156, + 0.44, + 0.223 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.456, + 0.156, + 0.548, + 0.223 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.565, + 0.156, + 0.657, + 0.223 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.693, + 0.156, + 0.783, + 0.223 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.234, + 0.234, + 0.321, + 0.302 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.233, + 0.306, + 0.319, + 0.318 + ], + "angle": 0, + "content": "Reference RGB" + }, + { + "type": "image", + "bbox": [ + 0.353, + 0.234, + 0.441, + 0.304 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.351, + 0.307, + 0.436, + 0.319 + ], + "angle": 0, + "content": "(a) Viewing dir." + }, + { + "type": "image", + "bbox": [ + 0.46, + 0.234, + 0.548, + 0.304 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.453, + 0.307, + 0.548, + 0.319 + ], + "angle": 0, + "content": "(b) Reflection dir." + }, + { + "type": "image", + "bbox": [ + 0.566, + 0.234, + 0.655, + 0.303 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.569, + 0.307, + 0.648, + 0.331 + ], + "angle": 0, + "content": "(c) Hybrid dir. (Ours)" + }, + { + "type": "image", + "bbox": [ + 0.693, + 0.234, + 0.78, + 0.303 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.679, + 0.307, + 0.784, + 0.319 + ], + "angle": 0, + "content": "Ground-truth Mesh" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.342, + 0.788, + 0.427 + ], + "angle": 0, + "content": "Fig. 1: Comparison of reconstruction results using different directional parameterizations in [2]. Using the viewing directional parameterization fails to reconstruct objects with specular surfaces, while using the reflection directional parameterization tends to result in incorrect geometry for objects with concave structures. Using our proposed hybrid parametrization shows consistently satisfactory reconstruction results compared to the existing parameterizations." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.455, + 0.789, + 0.727 + ], + "angle": 0, + "content": "Since the decomposition of geometry and appearance is inherently an ill-posed problem \\([40]^3\\), it is essential to introduce an appropriate inductive bias into the radiance network to facilitate correct decomposition. For example, simply inputting the viewing direction to the radiance network sometimes leads to incorrect reconstructions of highly specular objects (Fig. 1 (a)). To address this issue, the recent study [28] suggests that providing the reflection direction of rays considering the learned local surface geometry (i.e., normals) would result in a simpler radiance field to learn and facilitate correct decomposition, especially for specular objects. Some subsequent studies have shown that adopting this parameterization results in satisfactory reconstructions of highly specular objects [8, 13, 14] and also non-reflecting objects [8]. However, although this parameterization seems promising (Fig. 1 (b) top), we observed that it still does not work well when reconstructing objects with complex structures such as concave surfaces (Fig. 1 (b) bottom). Interestingly, while using viewing direction fails to recover specular surfaces, it succeeds in recovering such concave structures (Fig. 1 (a) bottom). Based on these, we pose the question: Is there a unified directional parameterization that can effectively handle both specular surfaces and complex structures?" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.728, + 0.789, + 0.805 + ], + "angle": 0, + "content": "In this paper, to answer the question, we start by studying the limitations of existing directional parameterizations and provide the first in-depth analysis of why using reflection direction may fail to reconstruct objects with concave or complex structures. As mentioned above, one important feature of the reflection directional parameterization is its dependency on the learnable geometry (i.e.," + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.811, + 0.787, + 0.842 + ], + "angle": 0, + "content": "3 If the capacity of the radiance network is sufficiently high, even for a fixed geometry such as a sphere, the image reconstruction loss can converge to zero [40]." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.465, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Rethinking directional parameterization" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.375 + ], + "angle": 0, + "content": "normals). Our insight into this fact is that for sampled 3D points along a sampled ray during training, while utilizing reflection directions of sampled points near the intersecting surface is helpful in decoupling geometry and appearance, the use of reflection directions of sampled points elsewhere may hinder the accurate reconstruction of geometry. Specifically, we identified that the cause of this issue lies in the sampled points beyond a certain distance from the intersecting surface during the optimization. As we will discuss in Sec. 3.2, the normals of these points computed from learnable geometry may either be associated with surfaces unrelated to the intersection or introduce high-frequency variations in the input of the radiance network, which hinders the proper updating of geometry and radiance. This hindrance is particularly noticeable in the reconstruction of concave surfaces. The analysis also indirectly explains the advantage of using viewing directional parameterization in reconstructing concave or complex structures, as it is non-learnable and remains constant for all sampled points along a sampled ray." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.377, + 0.788, + 0.649 + ], + "angle": 0, + "content": "Based on our insights and analysis, we propose a novel hybrid directional parameterization to address the limitations of existing directional parameterizations in a unified form. The main idea is to utilize the predicted distance to the surface for each sampled point: if the distance is small, indicating proximity to the intersecting surface, it is preferable to use its reflection direction as the directional parameterization, which facilitates the correct decomposition of geometry and appearance; in other cases, we instead utilize the non-learnable viewing direction to avoid the optimization issue mentioned above. The proposed hybrid directional parameterization smoothly transitions from the reflection direction to the viewing direction based on the predicted distance to the surface. The rate of this directional transition is controlled by a learnable scaling factor. We integrate the proposed hybrid directional parameterization into different methods and conduct extensive experiments to validate its effectiveness across various datasets. Compared to the existing viewing and reflection directional parameterizations, our proposed hybrid directional parameterization achieves consistently satisfactory results, as shown in Fig. 1 (c). Moreover, the proposed hybrid directional parameterization is nearly parameter-free and can be effortlessly applied in any existing neural implicit surface reconstruction method." + }, + { + "type": "text", + "bbox": [ + 0.241, + 0.653, + 0.582, + 0.667 + ], + "angle": 0, + "content": "To summarize, our contributions are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.697, + 0.787, + 0.742 + ], + "angle": 0, + "content": "- We are the first to provide an in-depth analysis of why the use of reflection directional parameterization tends to reconstruct incorrect geometry of objects with concave or complex structures." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.746, + 0.785, + 0.791 + ], + "angle": 0, + "content": "- We propose a novel hybrid directional parameterization to address the limitations of existing directional parameterizations, which can effectively handle both specular surfaces and complex structures in a unified form." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.795, + 0.785, + 0.84 + ], + "angle": 0, + "content": "- We validate the effectiveness of the proposed directional parameterization across different methods on various datasets, and demonstrate an overall superiority compared to existing directional parameterizations." + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.697, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.403, + 0.129 + ], + "angle": 0, + "content": "Jiang and Xu et al." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.145, + 0.388, + 0.161 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.175, + 0.586, + 0.192 + ], + "angle": 0, + "content": "2.1 Neural Implicit Surface Reconstruction" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.199, + 0.787, + 0.381 + ], + "angle": 0, + "content": "With the significant advancement of neural implicit representations [3,16-19,23, 24,37] and differentiable rendering [10,11,17,20,25,27,33] in recent years, neural implicit surface reconstruction methods [21,29,35,36] have emerged as a promising alternative to traditional methods in the field of image-based 3D surface reconstruction. Compared to traditional multi-view stereo methods, neural implicit surface methods employ a more flexible and representative implicit shape representation, such as the signed distance function (SDF), and introduce the directional radiance field [17] to model the view-dependent appearance of the target object. Subsequent works on neural implicit surface reconstruction have mostly built upon the foundation of NeuS [29] or VolSDF [35], and made improvements for specific scenarios including fast rendering [30, 31], sparse-view inputs [15], large-scale scenes [12,38] and more." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.402, + 0.562, + 0.418 + ], + "angle": 0, + "content": "2.2 Reconstruction for Specular Objects" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.425, + 0.803, + 0.652 + ], + "angle": 0, + "content": "Despite neural implicit surface reconstruction methods consider the view-dependent appearance of objects by incorporating the directional radiance field, reconstruction for specular objects remains a challenging task due to its significant shape-radiance ambiguity. Several works [14, 26, 34, 39, 41] model the view-dependent appearance of specular objects by incorporating additional physical components (e.g., materials and lighting), with the hope of jointly recovering both the geometric shape and physical parameters. However, as the number of unknown parameters increases, these works typically rely on a good initial geometry to ensure favorable results, leading to a chicken-and-egg problem. On the other hand, Ref-NeRF [28] was the first to propose the use of a radiance field conditioned on reflection direction to better decouple the geometry and radiance of specular objects. Subsequent neural implicit surface reconstruction methods targeting specular objects [8, 13, 14] have also adopted this directional parameterization. However, these methods often fail to correctly recover the geometry of objects with concave or complex structures." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.674, + 0.33, + 0.69 + ], + "angle": 0, + "content": "3 Method" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.704, + 0.787, + 0.84 + ], + "angle": 0, + "content": "In this section, we introduce a novel directional parameterization for neural implicit surface reconstruction methods. We first give a brief review on neural implicit surface reconstruction and its commonly used directional parameterizations, including viewing and reflection directions (Sec. 3.1). In the next, we summarize the observed failed reconstructions when utilizing existing directional parameterizations and conduct a detailed analysis of their causes (Sec. 3.2). Finally, a hybrid directional parameterization is proposed based on our analysis, which can effectively address the limitations of existing directional parameterizations in a unified form (Sec. 3.3)." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.465, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Rethinking directional parameterization" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.147, + 0.723, + 0.162 + ], + "angle": 0, + "content": "3.1 Preliminaries of Neural Implicit Surface Reconstruction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.172, + 0.788, + 0.444 + ], + "angle": 0, + "content": "Neural implicit surface reconstruction methods [2,6,29,36] typically represent the 3D object as the combination of a signed distance function (SDF) network \\(\\Theta_{f}\\), which maps a spatial position \\(\\mathbf{x} \\in \\mathbb{R}^3\\) to its SDF value \\(f(\\mathbf{x}) \\in \\mathbb{R}\\), and a directional radiance network \\(\\Theta_{c}\\), which maps a spatial position \\(\\mathbf{x} \\in \\mathbb{R}^3\\) and a directional vector \\(\\mathbf{d} \\in \\mathbb{S}^2\\) to the view-dependent color at \\(\\mathbf{x}\\). The surface \\(\\mathcal{S}\\) of the object is extracted as the zero-level set of \\(f(\\mathbf{x})\\). To enable the joint optimization of the SDF network \\(\\Theta_{f}\\) and radiance network \\(\\Theta_{c}\\) using only 2D image supervisions, volume rendering is typically utilized [29,35]. Given the sampled pixels in an image, camera rays \\(\\{\\mathbf{r}(t) = \\mathbf{o} + t\\mathbf{d}_{\\mathrm{view}}|t > 0\\}\\) are cast through the camera origin \\(\\mathbf{o}\\) and the sampled pixels, where \\(\\mathbf{d}_{\\mathrm{view}}\\) is the unit directional vector of the ray. To obtain the rendered colors of the sampled rays, a number of points are sampled along each ray, and queried for their SDF values and view-dependent colors from \\(\\Theta_{f}\\) and \\(\\Theta_{c}\\), where positional encoding (PE) [17] is typically incorporated. A relationship between the SDF value and opaque density is then formulated, such as using S-density [29], to translate the queried SDFs and colors to the ray color akin to NeRF [17]. Finally, the training loss is calculated from the difference between the rendered ray colors and the ground-truth colors of the sampled pixels." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.445, + 0.788, + 0.61 + ], + "angle": 0, + "content": "Directional parameterization. The key to the ability of the radiance network to model view-dependent appearance lies in its incorporation of a directional vector \\(\\mathbf{d}\\) as an input. Common directional parameterizations for \\(\\mathbf{d}\\) includes the viewing direction \\(\\mathbf{d}_{\\mathrm{view}}\\) and reflection direction \\(\\mathbf{d}_{\\mathrm{ref}}\\). The viewing directional parameterization is intuitively straightforward, as it models the view-dependent appearance from the perspective of the observer. The reflection directional parameterization is a little more complicated, which encodes the viewing direction \\(\\mathbf{d}_{\\mathrm{view}}\\) and local normal direction \\(\\mathbf{n}\\) into a new compact directional vector \\(\\mathbf{d}_{\\mathrm{ref}}\\), by considering the interaction between light and surface. Intuitively, it represents the direction of the incoming light, assuming that the surface is specular. \\(\\mathbf{d}_{\\mathrm{ref}}\\) is formulated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.401, + 0.625, + 0.785, + 0.64 + ], + "angle": 0, + "content": "\\[\n\\mathbf {d} _ {\\text {r e f}} = 2 \\left(\\mathbf {d} _ {\\text {v i e w}} \\cdot \\mathbf {n}\\right) \\mathbf {n} - \\mathbf {d} _ {\\text {v i e w}}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.651, + 0.741, + 0.674 + ], + "angle": 0, + "content": "where \\(\\mathbf{n}\\) can be derived from the gradient of the SDF as \\(\\mathbf{n}(\\mathbf{x}) = \\frac{\\nabla f(\\mathbf{x})}{\\|\\nabla f(\\mathbf{x})\\|}\\)." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.693, + 0.482, + 0.709 + ], + "angle": 0, + "content": "3.2 Observations and Analysis" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Observations. Although existing neural implicit surface reconstruction methods widely adopt either viewing direction or reflection direction as their directional parameterization, both types of parameterizations have their respective limitations when it comes to reconstructing certain objects. Using viewing direction often faces difficulty in correctly reconstructing objects with specular surfaces (e.g., Fig. 1 (a) and Fig. 2 (c)). On the other hand, while using reflection direction can reconstruct specular objects with simple geometry (e.g., Fig. 1 (b) top and Fig. 2 (d)), it often results in overly smooth reconstructions with" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.403, + 0.128 + ], + "angle": 0, + "content": "Jiang and Xu et al." + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.145, + 0.331, + 0.215 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.228, + 0.217, + 0.328, + 0.237 + ], + "angle": 0, + "content": "Object w/ concave structures" + }, + { + "type": "image", + "bbox": [ + 0.339, + 0.145, + 0.407, + 0.215 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.337, + 0.216, + 0.409, + 0.238 + ], + "angle": 0, + "content": "(a)Viewing dir." + }, + { + "type": "image", + "bbox": [ + 0.427, + 0.145, + 0.493, + 0.215 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.419, + 0.216, + 0.5, + 0.238 + ], + "angle": 0, + "content": "(b) Reflection dir. (X" + }, + { + "type": "image", + "bbox": [ + 0.52, + 0.149, + 0.594, + 0.214 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.514, + 0.216, + 0.607, + 0.237 + ], + "angle": 0, + "content": "Object w/ specular surfaces" + }, + { + "type": "image", + "bbox": [ + 0.61, + 0.153, + 0.683, + 0.216 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.614, + 0.216, + 0.686, + 0.226 + ], + "angle": 0, + "content": "(c) Viewing dir." + }, + { + "type": "image", + "bbox": [ + 0.695, + 0.153, + 0.769, + 0.216 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.696, + 0.216, + 0.777, + 0.226 + ], + "angle": 0, + "content": "(d) Reflection dir." + }, + { + "type": "image", + "bbox": [ + 0.728, + 0.227, + 0.747, + 0.24 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.253, + 0.785, + 0.294 + ], + "angle": 0, + "content": "Fig. 2: Scenarios where existing directional parameterizations succeed and struggle. Reconstruction results are obtained by integrating the two directional parameterizations into NeuS [29]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.326, + 0.784, + 0.386 + ], + "angle": 0, + "content": "missing concave surfaces when reconstructing objects with concave or complex structures (e.g., Fig. 1 (b) bottom and Fig. 2 (b)). Interestingly, using viewing direction may yield better results in these parts with complex or concave surfaces (e.g., Fig. 1 (a) bottom and Fig. 2 (a))." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.387, + 0.785, + 0.446 + ], + "angle": 0, + "content": "Based on these observations, we raise the question: Is there a unified directional parameterization that can effectively handle both specular surfaces and complex geometry? To answer this, we start by analyzing the reasons behind their reconstruction failures:" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.448, + 0.785, + 0.597 + ], + "angle": 0, + "content": "Analysis of failures when using viewing direction. An analysis of why using viewing direction is unsuitable for modeling the view-dependent appearance of specular objects is provided in Verbin et al. [28], which points out that the radiance network conditioned on viewing direction is poorly suited for interpolation. We consider this characteristic leads to difficulties in correctly decoupling the geometry and appearance of specular objects when using viewing direction. On the other hand, reflection directional parameterization can effectively leverage the prior knowledge of the interaction between light and surface, which facilitates the correct decoupling of geometry and appearance, especially for specular objects." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.599, + 0.785, + 0.748 + ], + "angle": 0, + "content": "Analysis of failures when using reflection direction. In contrast, no work has yet provided a detailed analysis of why using reflection directional parameterization fails in the reconstruction of concave or complex structures, making us the first to provide this analysis. We first highlight two key differences between viewing direction and reflection direction: the viewing direction is non-learnable throughout the optimization and remains constant along a sampled ray, whereas the reflection direction 1) depends on learnable geometry (i.e., normals) and 2) may vary significantly during the optimization. We argue that these two properties of the reflection direction may pose challenges to the accurate recovery of geometry, especially for concave or complex structures." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.785, + 0.84 + ], + "angle": 0, + "content": "A 2D toy example is first illustrated on Fig. 3 (left) to better demonstrate why the first property of reflection direction may hinder the recovery of geometry, which assumes the scenario of recovering an \"L\"-shaped concave structure. A sampled ray intersecting the surface and the sampled points on the ray along with the normals at these points are illustrated. In this case, we notice that the normals at sampled points that are slightly distant from the intersecting surface" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.465, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Rethinking directional parameterization" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "image", + "bbox": [ + 0.222, + 0.149, + 0.511, + 0.361 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.529, + 0.149, + 0.721, + 0.361 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.738, + 0.149, + 0.782, + 0.363 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.377, + 0.788, + 0.489 + ], + "angle": 0, + "content": "Fig. 3: Impact of incorporating normals on view-dependent radiance modeling. (Left) Normals at sampled points may be influenced by unrelated surfaces other than the intersection surface. (Top right) Smooth surfaces yield similarity in normals and reflection directions for the sampled points, while (Bottom right) surfaces with intricate local details, e.g., concavities, induce a scattered distribution of normals and reflection directions (Ref. dirs), particularly pronounced slightly distant from the vicinity of the zero-level set, which adversely affects geometry optimization. The details are explained in Sec. 3.2." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.522, + 0.788, + 0.628 + ], + "angle": 0, + "content": "do not necessarily relate to the normal of the intersecting surface (the bottom side of the zero level-set), but rather to the normals of surfaces elsewhere (the left side of the zero level-set). As a result, since the training loss of existing neural implicit surface reconstruction methods is usually ray-based, which is calculated using the local color of the intersecting surface, using this loss to optimize the normals of irrelevant surfaces elsewhere may introduce ambiguity in the optimization of geometry and result in incorrect geometric updates." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.63, + 0.788, + 0.841 + ], + "angle": 0, + "content": "In addition, the second property of the reflection direction is more closely related to the phenomenon that using reflection directional parameterization tends to reconstruct overly smooth surfaces with missing concave or complex structures. Assuming that the geometry reconstruction begins with a smooth surface exhibiting local flatness (Fig. 3 top-right), which aligns with the geometric initialization practices commonly employed by existing neural implicit surface reconstruction methods [6,29], and the geometry of the target object includes concave structures (Fig. 3 bottom-right). We observed that for the same sampled ray, when the surface geometry is updated from a smooth surface to a concave surface, the normals and reflection directions of the same sampled points also change. This change is particularly pronounced for points that are slightly distant from the surface, and could be further pronounced due to the presence of Positional Encoding (PE). The noticeable changes in reflection directions at these points within each surface update introduce high-frequency variances to" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.403, + 0.129 + ], + "angle": 0, + "content": "Jiang and Xu et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.237 + ], + "angle": 0, + "content": "the directional inputs of the radiance network. This makes learning a stable radiance field challenging. To be specific, assuming that the radiance network conditioned on reflection direction has already learned to minimize the color loss to some extent before the surface update, the substantial change in directional inputs due to the update could likely lead to an increase in color loss, which hinders the optimization and is susceptible to falling into local minima." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.239, + 0.788, + 0.344 + ], + "angle": 0, + "content": "In contrast, we notice that using the viewing direction often leads to concave or overly intricate surface details (although some may be incorrect), as shown in Fig. 1 (a) and Fig. 2 (a) (c). We consider that this is because the viewing direction does not rely on the surface geometry, which means that changes in geometry do not affect the optimization of the radiance network conditioned on viewing direction. This allows the SDF can undergo more \"aggressive\" updates without considering the impact on radiance." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.365, + 0.566, + 0.38 + ], + "angle": 0, + "content": "3.3 Hybrid Directional Parameterization" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.389, + 0.789, + 0.631 + ], + "angle": 0, + "content": "Building upon the analysis from the previous section, we discovered that viewing direction and reflection direction seem to have a spatially complementary relationship: On one hand, for sampled points near the intersecting surface, the normals here accurately reflect the normals of the intersecting surface, thereby enabling the reflection direction to better model the interaction between light and surface, which in turn promotes the correct decomposition of geometry and appearance. On the other hand, for other sampled points along the ray, using the reflection direction might introduce irrelevant surface geometry into the optimization, or high-frequency variations in the directional inputs of the radiance network, both potentially hindering the optimization of the geometry and radiance networks. By contrast, using the viewing direction at these sampled points can avoid these issues. Motivated by these thoughts, we propose a spatial-aware directional representation, which transitions from reflection direction \\(\\mathbf{d}_{\\mathrm{ref}}\\) to viewing direction \\(\\mathbf{d}_{\\mathrm{view}}\\) based on the distance to the surface (i.e., the absolute value of SDF). We refer to this fused direction as hybrid direction \\(\\mathbf{d}_{\\mathrm{hyb}}\\), formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.348, + 0.647, + 0.785, + 0.663 + ], + "angle": 0, + "content": "\\[\n\\mathbf {d} _ {\\mathrm {h y b}} = \\text {n o r m a l i z e} (\\alpha \\cdot \\mathbf {d} _ {\\text {r e f}} + (1 - \\alpha) \\cdot \\mathbf {d} _ {\\text {v i e w}}). \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.667, + 0.788, + 0.711 + ], + "angle": 0, + "content": "Here, \\(\\alpha \\in [0,1]\\) represents the blend weight. We normalize the resulting linear blend to a unit vector thereby representing a valid direction. The blend weight \\(\\alpha\\) is implemented as" + }, + { + "type": "equation", + "bbox": [ + 0.398, + 0.725, + 0.785, + 0.74 + ], + "angle": 0, + "content": "\\[\n\\alpha = \\exp (- \\gamma \\cdot \\det a c h (| f (\\mathbf {x}) |)), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.841 + ], + "angle": 0, + "content": "where \\(\\gamma \\in \\mathbb{R}^{+}\\) is an estimable parameter that controls the transition speed between reflection/viewing direction. The gradient of the SDF is detached to prevent interference with geometric updates. At the surface of the object, i.e., when SDF equals 0, \\(\\alpha\\) simply becomes 1, which means that the hybrid direction aligns with the reflection direction. As \\(\\mathbf{x}\\) moves further away from the surface, \\(\\alpha\\) gradually approaches 0, indicating that the direction tends towards the viewing" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.465, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Rethinking directional parameterization" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.177 + ], + "angle": 0, + "content": "direction. Here, \\(\\alpha\\) can be any bell-shaped function w.r.t. the SDF value. In this case, we utilize the PDF of the Laplace distribution for simplicity." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.178, + 0.785, + 0.221 + ], + "angle": 0, + "content": "In the next section, extensive experimental results validate the superiority of the proposed hybrid directional parameterization compared to existing parameterizations." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.244, + 0.376, + 0.26 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.274, + 0.446, + 0.289 + ], + "angle": 0, + "content": "4.1 Experimental Settings" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.297, + 0.785, + 0.476 + ], + "angle": 0, + "content": "Datasets. We conduct experiments on objects from different datasets, including the DTU dataset [1], the Shiny Blender dataset [28], and the real captured dataset [9]. The DTU dataset [1] contains indoor multi-view captures of diverse objects, including challenging conditions for geometry reconstruction such as non-Lambertian surfaces and complex structures. We use the same 15 objects as previous works [2, 29, 31, 36] for evaluation. In addition, we utilize the foreground masks provided by Yariv et al. [36]. The Shinny Blender dataset [28] is a synthetic dataset rendered in Blender [4], which contains different objects with specular surfaces. We follow the experimental settings used in Ref-NeuS [8] on this dataset, which selects four objects. The real captured dataset [9] contains a set of three outdoor captures of real objects with complex geometry and specular surfaces. We select the \"toycar\" scene for evaluation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.479, + 0.785, + 0.718 + ], + "angle": 0, + "content": "Implementation details. We integrate the proposed hybrid directional parameterization into two types of neural surface reconstruction methods: NeuS [29] and NeuDA [2]. The former utilizes a deep MLP network as the scene representation, while the latter leverages a shallow MLP decoder combined with multi-level voxel grids. Both methods adopt a radiance network conditioned on viewing direction in their original versions. The modification of the directional input from the viewing direction to the proposed hybrid direction can be accomplished with less than 10 lines of code. We additionally implement the radiance network that depends on reflection direction for both methods, to facilitate a comprehensive comparison of different directional parameterizations. Regarding the learnable parameter \\(\\gamma\\) in Eqn. 3, we implement it as \\(\\gamma = \\exp (10\\cdot \\gamma_b)\\). We assign an initial value of 0.3 to \\(\\gamma_{b}\\) except for the \"toaster\" scene from the Shiny Blender dataset. For the \"toaster\" scene, an initial value of 0.1 is assigned since it results in better reconstruction. The training losses of the modified methods are identical to those of their original versions. The number of training iterations is set to 300,000 to ensure convergence." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Evaluation metrics. For the DTU dataset, we follow the evaluation protocol of previous works [2, 29] and evaluate the Chamfer Distance (CD) between the estimated and ground-truth geometry. For the Shiny Blender dataset, we follow the evaluation protocol proposed in Ref-NeuS [8]. It evaluates the accuracy (Acc.) instead of the CD for unbiased evaluation, and also evaluates the mean angular error (MAE) between the ground-truth normal maps and the rendered normal maps. For the real captured dataset [9], we only provide a qualitative evaluation, as ground-truth geometry is not provided in this dataset." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.403, + 0.129 + ], + "angle": 0, + "content": "Jiang and Xu et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.788, + 0.2 + ], + "angle": 0, + "content": "Table 1: Quantitative evaluation in terms of Chamfer distance of integrating different directional parameterizations in NeuS [29] and NeuDA [2] on the DTU dataset [1]. The experimental objects are categorized into diffuse objects and specular objects. The numbers in the headers denote the scene IDs. Bold results have the best score." + }, + { + "type": "table", + "bbox": [ + 0.217, + 0.213, + 0.785, + 0.3 + ], + "angle": 0, + "content": "
MethodsScan of Diffuse ObjectsScan of Specular ObjectsTotal Avg.
243740556583105106114118122Avg.636997110Avg.
NeuS [29]0.830.980.560.370.591.450.780.520.360.450.450.671.130.600.951.431.030.77
NeuS, w/ reflection dir.1.394.390.730.391.341.490.920.590.360.460.541.141.110.631.460.921.031.11
NeuS, w/ hybrid dir. (Ours)0.810.970.540.380.611.470.770.510.350.450.470.671.100.600.850.900.860.72
NeuD A [2]0.510.760.390.370.561.370.790.500.340.420.460.591.080.571.130.800.900.67
NeuD A, w/ reflection dir.0.581.030.630.370.791.450.890.510.370.430.500.691.030.550.900.760.810.72
NeuD A, w/ hybrid dir. (Ours)0.530.770.400.360.571.380.760.470.330.410.470.591.020.540.760.750.770.63
" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.328, + 0.714, + 0.344 + ], + "angle": 0, + "content": "4.2 Comparison of Different Directional Parameterizations" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.357, + 0.788, + 0.705 + ], + "angle": 0, + "content": "We first evaluate the performance of integrating different directional parameterizations into the two baseline methods NeuS [29] and NeuDA [2] on the DTU dataset [1]. The experimental objects from the DTU dataset are categorized into diffuse objects and specular objects for better demonstration. Tab. 1 summarizes the quantitative results. For both baseline methods, we notice that substituting the original viewing directional parameterization with the reflection direction only yields better results for a minority of specular objects. For diffuse objects and the remaining specular objects, employing reflection directional parameterization demonstrates significant degradation in performance. Considering the relatively complex geometry of DTU objects, we attribute this to the optimization issue analyzed in Sec. 3.2. Conversely, the incorporation of the proposed hybrid directional parameterization enables both modified methods to yield consistently superior performance for specular objects while achieving results on par with the original version for diffuse objects. The first two rows in Fig. 4 showcase the visual comparison of adopting different directional parameterizations for the reconstruction of two specular objects from the DTU dataset. Both baseline methods fail to reconstruct the specular surfaces of objects. On the other hand, although integrating the reflection direction yields better reconstruction for specular objects with simple geometry (e.g., Scan 63), they fail to reconstruct the specular surfaces at the concave structure in Scan 97. In contrast, integrating the proposed hybrid direction achieves the best reconstruction results in both scenarios. More qualitative comparisons on the DTU dataset are available in our supplementary." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.705, + 0.788, + 0.84 + ], + "angle": 0, + "content": "We further validate the advantage of the proposed hybrid directional parameterization on the Shiny Blender dataset [28], as demonstrated in Tab. 2. Both baseline methods exhibit large reconstruction errors on the surfaces of highly specular objects in this dataset. By contrast, due to the relatively simple geometry of objects, replacing the viewing directional parameterization with the reflection direction results in a significant improvement in reconstruction quality on most objects in this dataset, except for the \"coffee\" object. For the reconstruction of the \"coffee\" object, adopting the reflection directional parameterization fails to recover the concave structure formed by the inner wall of the coffee cup" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.465, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Rethinking directional parameterization" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "table_caption", + "bbox": [ + 0.215, + 0.145, + 0.788, + 0.187 + ], + "angle": 0, + "content": "Table 2: Quantitative evaluation of integrating different directional parameterizations in NeuS [29] and NeuDA [2] on the Shiny Blender dataset [28]. Bold results have the best score." + }, + { + "type": "table", + "bbox": [ + 0.217, + 0.199, + 0.785, + 0.304 + ], + "angle": 0, + "content": "
MethodshelmettoastercoffeecarAvg.
Acc. ↓MAE ↓Acc. ↓MAE ↓Acc. ↓MAE ↓Acc. ↓MAE ↓Acc. ↓MAE ↓
NeuS [29]4.883.203.312.851.971.060.860.952.762.02
NeuS, w/ reflection dir.0.310.360.390.943.301.370.480.731.120.85
NeuS, w/ hybrid dir. (Ours)0.290.340.450.920.640.550.430.720.450.63
NeuDA [2]5.733.775.983.391.160.810.871.033.442.25
NeuDA, w/ reflection dir.0.340.371.031.713.661.530.620.871.411.12
NeuDA, w/ hybrid dir. (Ours)0.320.360.951.540.550.570.600.840.610.83
" + }, + { + "type": "image", + "bbox": [ + 0.22, + 0.314, + 0.782, + 0.527 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.537, + 0.788, + 0.568 + ], + "angle": 0, + "content": "Fig. 4: Qualitative comparison of integrating different directional parameterizations in NeuS [29] and NeuDA [2] on the DTU dataset [1] and Shiny Blender dataset [28]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.597, + 0.788, + 0.673 + ], + "angle": 0, + "content": "and the surface of the water, as shown in the third row in Fig. 4. On the other hand, the integration of our proposed hybrid directional parameterization enables the successful reconstruction of both the specular surfaces and concave structures, obtaining consistently superior reconstructions of all objects in this dataset, compared to the other two directional parameterizations." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.673, + 0.788, + 0.794 + ], + "angle": 0, + "content": "Lastly, we qualitatively present the reconstruction results on the \"toycar\" scene from the real captured scene in Fig. 5, which contains challenging reflective glass and complex geometry. Both modified methods employing the proposed hybrid directional parameterization succeed in recovering the complex geometry and the smooth surface of the reflective glass. In contrast, using either the reflection direction or the original viewing direction results in reconstructions that are either overly smooth, missing concave surface structures, or provide incorrect surface reconstructions at the reflective glass areas." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.788, + 0.842 + ], + "angle": 0, + "content": "As the summary of this section, compared to the viewing directional parameterization which struggles with reconstructing specular surfaces, and the reflection directional parameterization which faces difficulties in reconstructing" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.402, + 0.128 + ], + "angle": 0, + "content": "Jiang and Xu et al." + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.186, + 0.35, + 0.258 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.243, + 0.26, + 0.33, + 0.27 + ], + "angle": 0, + "content": "Reference RGB" + }, + { + "type": "image", + "bbox": [ + 0.359, + 0.148, + 0.486, + 0.216 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.406, + 0.217, + 0.438, + 0.227 + ], + "angle": 0, + "content": "NeuS" + }, + { + "type": "image", + "bbox": [ + 0.499, + 0.149, + 0.627, + 0.216 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.217, + 0.626, + 0.228 + ], + "angle": 0, + "content": "NeuS, w/ reflection dir." + }, + { + "type": "image", + "bbox": [ + 0.642, + 0.154, + 0.768, + 0.216 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.651, + 0.217, + 0.768, + 0.228 + ], + "angle": 0, + "content": "NeuS, w/ hybrid dir." + }, + { + "type": "image", + "bbox": [ + 0.359, + 0.233, + 0.486, + 0.299 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.4, + 0.3, + 0.444, + 0.31 + ], + "angle": 0, + "content": "NeuDA" + }, + { + "type": "image", + "bbox": [ + 0.499, + 0.234, + 0.627, + 0.299 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.3, + 0.634, + 0.312 + ], + "angle": 0, + "content": "NeuDA,w/ reflection dir." + }, + { + "type": "image", + "bbox": [ + 0.642, + 0.234, + 0.768, + 0.299 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.651, + 0.3, + 0.78, + 0.313 + ], + "angle": 0, + "content": "NeuDA, w/ hybrid dir." + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.323, + 0.784, + 0.352 + ], + "angle": 0, + "content": "Fig. 5: Qualitative comparison of integrating different directional parameterizations in NeuS [29] and NeuDA [2] on the \"toycar\" scene from the real captured dataset [9]." + }, + { + "type": "table_caption", + "bbox": [ + 0.215, + 0.367, + 0.784, + 0.394 + ], + "angle": 0, + "content": "Table 3: Quantitative evaluation in terms of Chamfer distance on the DTU dataset [1]. The numbers in the header denote the scene IDs. Bold results have the best score." + }, + { + "type": "table", + "bbox": [ + 0.218, + 0.408, + 0.784, + 0.503 + ], + "angle": 0, + "content": "
Methods243740556365698397105106110114118122Avg.
IDR [36]1.631.870.630.481.040.790.771.331.160.760.670.900.420.510.530.90
NeuS [29]0.830.980.560.371.130.590.601.450.950.780.521.430.360.450.450.77
Voxurf [31]0.650.740.390.350.960.640.851.581.010.680.601.110.370.450.470.72
NeuDA [2]0.510.760.390.371.080.560.571.371.130.790.500.800.340.420.460.67
Ours (NeuS)0.810.970.540.381.100.610.601.470.850.770.510.900.350.450.470.72
Ours (NeuDA)0.530.770.400.361.020.570.541.380.760.760.470.760.330.410.470.63
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.531, + 0.784, + 0.591 + ], + "angle": 0, + "content": "complex structures, our proposed hybrid directional parameterization effectively handles both scenarios in a unified form. This benefit is attributed to our hybrid directional parameterization effectively combining the advantages of both viewing and reflection directional parameterizations while avoiding their drawbacks." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.616, + 0.523, + 0.631 + ], + "angle": 0, + "content": "4.3 Comparison to State-of-the-Art" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.644, + 0.787, + 0.809 + ], + "angle": 0, + "content": "Although our main purpose is not to develop a best-performing surface reconstruction system, we demonstrate that we can still achieve state-of-the-art performance by integrating the proposed hybrid directional parameterization into existing methods. For brevity, we refer to NeuS [29] and NeuDA [2] incorporating the proposed hybrid directional parameterization as Ours (NeuS) and Ours (NeuDA) in the following text. Tab. 3 reports the Chamfer distance (CD) metrics of different methods on the DTU dataset [1]. For a fair comparison, all methods in Tab. 3 utilize foreground masks provided by Yariv et al. [36] during training. Both Ours (NeuS) and Ours (NeuDA) improve their baselines by a certain margin in the average CD scores, and Ours (NeuDA) achieves the best performance compared to previous SOTA methods in the average CD metric." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.81, + 0.786, + 0.84 + ], + "angle": 0, + "content": "Tab. 4 reports the quantitative evaluation of different methods on the Shiny Blender dataset [28]. The results of [5,6,8,22,28,35] are taken from the results" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.465, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Rethinking directional parameterization" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.785, + 0.173 + ], + "angle": 0, + "content": "Table 4: Quantitative evaluation on the Shiny Blender dataset [28]. “-” denotes that the results are not available. Bold results have the best score." + }, + { + "type": "table", + "bbox": [ + 0.274, + 0.185, + 0.728, + 0.327 + ], + "angle": 0, + "content": "
MethodshelmettoastercoffeecarAvg.
Acc. ↓MAE ↓Acc. ↓MAE ↓Acc. ↓MAE ↓Acc. ↓MAE ↓Acc. ↓MAE ↓
Ref-NeRF [28]-29.48-42.87-12.24-14.93-24.88
NeuS [29]4.883.203.312.851.971.060.860.952.762.02
UNISURF [22]1.691.783.755.512.883.152.341.982.673.11
VolSDF [35]1.551.372.022.532.232.280.581.041.601.81
NeuralWarp [5]2.251.945.903.511.542.040.651.072.592.14
Geo-NeuS [6]0.742.366.353.763.856.362.885.673.464.54
NeuD A [2]5.733.775.983.391.160.810.871.033.442.25
Ref-NeuD [8]0.290.380.421.470.770.990.370.800.460.91
Ours (NeuD)0.290.340.450.920.640.550.430.720.450.63
Ours (NeuD A)0.320.360.951.540.550.570.600.840.610.83
" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.34, + 0.293, + 0.394 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.227, + 0.396, + 0.286, + 0.405 + ], + "angle": 0, + "content": "Reference RGB" + }, + { + "type": "image", + "bbox": [ + 0.296, + 0.34, + 0.393, + 0.394 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.309, + 0.396, + 0.391, + 0.405 + ], + "angle": 0, + "content": "\\(\\gamma_{b} = 0.0\\) \\(\\mathrm{CD}\\downarrow\\) : 0.91" + }, + { + "type": "image", + "bbox": [ + 0.396, + 0.34, + 0.49, + 0.394 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.409, + 0.396, + 0.49, + 0.405 + ], + "angle": 0, + "content": "\\(\\gamma_{b} = 0.1\\) \\(\\mathrm{CD}\\downarrow\\) : 0.81" + }, + { + "type": "image", + "bbox": [ + 0.493, + 0.34, + 0.588, + 0.394 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.504, + 0.396, + 0.587, + 0.405 + ], + "angle": 0, + "content": "\\(\\gamma_{b} = 0.2\\) \\(\\mathrm{CD}\\downarrow\\) :0.80" + }, + { + "type": "image", + "bbox": [ + 0.591, + 0.34, + 0.687, + 0.394 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.604, + 0.396, + 0.686, + 0.405 + ], + "angle": 0, + "content": "\\(\\gamma_{b} = 0.3\\) \\(\\mathrm{CD}\\downarrow\\) :0.76" + }, + { + "type": "image", + "bbox": [ + 0.69, + 0.34, + 0.785, + 0.394 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.704, + 0.396, + 0.785, + 0.405 + ], + "angle": 0, + "content": "\\(\\gamma_{b} = 0.4\\) \\(\\mathrm{CD}\\downarrow\\) :0.77" + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.416, + 0.785, + 0.444 + ], + "angle": 0, + "content": "Fig. 6: Reconstruction results using different initial values of \\(\\gamma_{b}\\). The main differences are highlighted in red boxes." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.474, + 0.788, + 0.552 + ], + "angle": 0, + "content": "reported in Ge et al. [8]. Both Ours (NeuS) and Ours (NeuDA) achieve better average MAE metrics than the previous SOTA method Ref-NeuS [8], and Ours (NeuS) achieve the best overall performance in terms of both Acc. and MAE metrics on this dataset. We refer to our supplementary for the qualitative comparisons between Ours (NeuS) and Ref-NeuS [8]." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.573, + 0.401, + 0.587 + ], + "angle": 0, + "content": "4.4 Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.599, + 0.785, + 0.628 + ], + "angle": 0, + "content": "In this section, we conduct ablation studies on several key design choices of the proposed hybrid directional parameterization by using Ours (NeuDA)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.63, + 0.787, + 0.795 + ], + "angle": 0, + "content": "The initial value of \\(\\gamma_{b}\\). We first validate the impact of setting different initial values for \\(\\gamma_{b}\\) on the reconstruction results. Fig. 6 compares the obtained reconstruction results using five different initial values for \\(\\gamma_{b}\\). We find that when the initial value of \\(\\gamma_{b}\\) is set low, our method tends to recover incorrect geometries similar to those caused by using the reflection directional parameterization. This phenomenon remains consistent with our analysis in Sec 3.2, as a smaller initial value for \\(\\gamma_{b}\\) increases the likelihood of involving irrelevant surfaces and high-frequency variations in the directional inputs of the radiance network during the optimization. We empirically find the best initial value of \\(\\gamma_{b}\\) to be 0.3 for most experimental objects. More discussions regarding the parameter \\(\\gamma_{b}\\) and blend weight \\(\\alpha\\) can be found in our supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.796, + 0.787, + 0.841 + ], + "angle": 0, + "content": "Detach SDF gradient in Eqn. 3 or not? Fig. 7 compares the reconstruction results with and without detaching the gradient of the estimated SDF in Eqn. 3. We observe that our method struggles to reconstruct the fine-grained geometry" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.402, + 0.128 + ], + "angle": 0, + "content": "Jiang and Xu et al." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.144, + 0.326, + 0.208 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.237, + 0.208, + 0.306, + 0.216 + ], + "angle": 0, + "content": "Reference RGB" + }, + { + "type": "image", + "bbox": [ + 0.328, + 0.145, + 0.411, + 0.207 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.331, + 0.208, + 0.407, + 0.226 + ], + "angle": 0, + "content": "w/o SDF gradient detachment" + }, + { + "type": "image", + "bbox": [ + 0.415, + 0.146, + 0.498, + 0.208 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.415, + 0.208, + 0.497, + 0.226 + ], + "angle": 0, + "content": "w/SDF gradient detachment (Ours)" + }, + { + "type": "image", + "bbox": [ + 0.52, + 0.153, + 0.6, + 0.215 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.522, + 0.218, + 0.586, + 0.227 + ], + "angle": 0, + "content": "Reference RGB" + }, + { + "type": "image", + "bbox": [ + 0.614, + 0.153, + 0.693, + 0.216 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.62, + 0.218, + 0.681, + 0.227 + ], + "angle": 0, + "content": "post-PE fusion" + }, + { + "type": "image", + "bbox": [ + 0.703, + 0.153, + 0.786, + 0.216 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.7, + 0.218, + 0.786, + 0.227 + ], + "angle": 0, + "content": "pre-PE fusion (Ours)" + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.238, + 0.498, + 0.28 + ], + "angle": 0, + "content": "Fig.7: Reconstruction results with and without detaching the gradient of the estimated SDF in Eqn. 3." + }, + { + "type": "image_caption", + "bbox": [ + 0.503, + 0.239, + 0.787, + 0.281 + ], + "angle": 0, + "content": "Fig. 8: Reconstruction results fusing directions before and after positional encoding (PE)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.308, + 0.785, + 0.383 + ], + "angle": 0, + "content": "when the gradient of the estimated SDF is not detached in Eqn. 3. We consider that introducing the gradient of the estimated SDF places an additional burden on the learning of both the geometry and radiance networks, making it easier to fall into local minima. Thereby, we choose to detach the gradient of SDF in Eqn. 3." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.384, + 0.788, + 0.566 + ], + "angle": 0, + "content": "Fuse the viewing direction and reflection direction before or after positional encoding? We introduce the fusion of viewing direction and reflection direction in Eqn. 2 before the operation of position encoding (PE). We compare the reconstruction results with different fusion orders, specifically the proposed pre-PE fusion: \\(\\mathrm{PE}(\\mathbf{d}_{\\mathrm{hyb}})\\) (ours) and alternative approach using post-PE fusion: \\(\\alpha \\cdot \\mathrm{PE}(\\mathbf{d}_{\\mathrm{ref}}) + (1 - \\alpha) \\cdot \\mathrm{PE}(\\mathbf{d}_{\\mathrm{view}})\\). The comparison is shown in Fig. 8. We observe that fusing directions after the PE operation may result in degraded reconstruction results. We consider that fusing the directional information in the high-dimension space may increase the complexity of the learning space of the radiance network, which brings challenges in finding optimal reconstruction results. Thereby, we choose to fuse the viewing direction and reflection direction before the operation of positional encoding." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.585, + 0.51, + 0.601 + ], + "angle": 0, + "content": "5 Conclusion and Limitation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.614, + 0.788, + 0.78 + ], + "angle": 0, + "content": "In this paper, we conduct an in-depth investigation into the topic of directional parameterization in neural implicit surface reconstruction. We reveal that the commonly used directional parameterizations in modeling the view-dependent radiance field, such as viewing and reflection directions, both have their limitations and none of them can effectively handle both specular surfaces and complex structures. A novel hybrid directional parameterization is introduced to achieve this goal, which softly fuses the viewing and reflection directions based on the distance of the sampled points from the surface, uniting the advantages of each representation. Extensive experiments demonstrate that our hybrid directional parameterization outperforms the existing directional parameterizations in reconstructing objects with a wide variety of materials, geometry, and appearance." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.84 + ], + "angle": 0, + "content": "One limitation of our method is that the optimal choice of initial value for \\(\\gamma_{b}\\) may vary depending on the target object, which would require additional efforts for parameter tuning. An automated method for selecting the optimal initial value of \\(\\gamma_{b}\\) for target objects could be a potential future work for us." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.465, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Rethinking directional parameterization" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.323, + 0.16 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.174, + 0.785, + 0.201 + ], + "angle": 0, + "content": "1. Aanaes, H., Jensen, R.R., Vogiatzis, G., Tola, E., Dahl, A.B.: Large-scale data for multiple-view stereopsis. IJCV 120, 153-168 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.203, + 0.785, + 0.228 + ], + "angle": 0, + "content": "2. Cai, B., Huang, J., Jia, R., Lv, C., Fu, H.: Neuda: Neural deformable anchor for high-fidelity implicit surface reconstruction. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.23, + 0.785, + 0.255 + ], + "angle": 0, + "content": "3. Chen, Z., Zhang, H.: Learning implicit fields for generative shape modeling. In: CVPR (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.257, + 0.785, + 0.282 + ], + "angle": 0, + "content": "4. Community, B.O.: Blender - a 3D modelling and rendering package. Blender Foundation, Stichting Blender Foundation, Amsterdam (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.284, + 0.785, + 0.31 + ], + "angle": 0, + "content": "5. Darmon, F., Bascle, B., Devaux, J., Monasse, P., Aubry, M.: Improving neural implicit surfaces geometry with patch warping. In: CVPR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.311, + 0.785, + 0.337 + ], + "angle": 0, + "content": "6. Fu, Q., Xu, Q., Ong, Y.S., Tao, W.: Geo-neus: Geometry-consistent neural implicit surfaces learning for multi-view reconstruction. In: NeurIPS (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.338, + 0.785, + 0.364 + ], + "angle": 0, + "content": "7. Furukawa, Y., Hernández, C., et al.: Multi-view stereo: A tutorial. Foundations and Trends® in Computer Graphics and Vision 9(1-2), 1-148 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.365, + 0.785, + 0.404 + ], + "angle": 0, + "content": "8. Ge, W., Hu, T., Zhao, H., Liu, S., Chen, Y.C.: Ref-neus: Ambiguity-reduced neural implicit surface learning for multi-view reconstruction with reflection. In: ICCV (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.406, + 0.785, + 0.432 + ], + "angle": 0, + "content": "9. Hedman, P., Srinivasan, P.P., Mildenhall, B., Barron, J.T., Debevec, P.: Baking neural radiance fields for real-time view synthesis. In: CVPR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.434, + 0.785, + 0.459 + ], + "angle": 0, + "content": "0. Kato, H., Beker, D., Morariu, M., Ando, T., Matsuoka, T., Kehl, W., Gaidon, A.: Differentiable rendering: A survey. arXiv preprint arXiv:2006.12057 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.461, + 0.785, + 0.486 + ], + "angle": 0, + "content": "1. Li, T.M., Aittala, M., Durand, F., Lehtinen, J.: Differentiable monte carlo ray tracing through edge sampling. ACM Transactions on Graphics 37(6), 1-11 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.488, + 0.785, + 0.513 + ], + "angle": 0, + "content": "2. Li, Z., Müller, T., Evans, A., Taylor, R.H., Unberath, M., Liu, M.Y., Lin, C.H.: Neuralangelo: High-fidelity neural surface reconstruction. In: CVPR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.515, + 0.785, + 0.553 + ], + "angle": 0, + "content": "3. Liang, R., Chen, H., Li, C., Chen, F., Panneer, S., Vijaykumar, N.: Envidr: Implicit differentiable renderer with neural environment lighting. arXiv preprint arXiv:2303.13022 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.555, + 0.785, + 0.595 + ], + "angle": 0, + "content": "4. Liu, Y., Wang, P., Lin, C., Long, X., Wang, J., Liu, L., Komura, T., Wang, W.: Nano: Neural geometry and brdf reconstruction of reflective objects from multiview images. ACM Transactions on Graphics (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.597, + 0.785, + 0.622 + ], + "angle": 0, + "content": "5. Long, X., Lin, C., Wang, P., Komura, T., Wang, W.: Sparseneus: Fast generalizable neural surface reconstruction from sparse views. In: ECCV (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.624, + 0.785, + 0.649 + ], + "angle": 0, + "content": "6. Mescheder, L., Oechsle, M., Niemeyer, M., Nowozin, S., Geiger, A.: Occupancy networks: Learning 3d reconstruction in function space. In: CVPR (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.65, + 0.785, + 0.69 + ], + "angle": 0, + "content": "7. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. In: ECCV (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.691, + 0.785, + 0.731 + ], + "angle": 0, + "content": "8. Müller, T., Evans, A., Schied, C., Keller, A.: Instant neural graphics primitives with a multiresolution hash encoding. ACM Transactions on Graphics 41(4), 1-15 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.733, + 0.785, + 0.758 + ], + "angle": 0, + "content": "9. Niemeyer, M., Mescheder, L., Oechsle, M., Geiger, A.: Occupancy flow: 4d reconstruction by learning particle dynamics. In: ICCV (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.76, + 0.785, + 0.799 + ], + "angle": 0, + "content": "20. Niemeyer, M., Mescheder, L., Oechsle, M., Geiger, A.: Differentiable volumetric rendering: Learning implicit 3d representations without 3d supervision. In: CVPR (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.801, + 0.785, + 0.84 + ], + "angle": 0, + "content": "21. Niemeyer, M., Mescheder, L., Oechsle, M., Geiger, A.: Differentiable volumetric rendering: Learning implicit 3d representations without 3d supervision. In: CVPR (2020)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.174, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.402, + 0.128 + ], + "angle": 0, + "content": "Jiang and Xu et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.175 + ], + "angle": 0, + "content": "22. Oechsle, M., Peng, S., Geiger, A.: Unisurf: Unifying neural implicit surfaces and radiance fields for multi-view reconstruction. In: ICCV (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.175, + 0.785, + 0.203 + ], + "angle": 0, + "content": "23. Park, J.J., Florence, P., Straub, J., Newcombe, R., Lovegrove, S.: Deepsdf: Learning continuous signed distance functions for shape representation. In: CVPR (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.203, + 0.785, + 0.229 + ], + "angle": 0, + "content": "24. Peng, S., Niemeyer, M., Mescheder, L., Pollefeys, M., Geiger, A.: Convolutional occupancy networks. In: ECCV (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.229, + 0.785, + 0.271 + ], + "angle": 0, + "content": "25. Ravi, N., Reizenstein, J., Novotny, D., Gordon, T., Lo, W.Y., Johnson, J., Gkioxari, G.: Accelerating 3d deep learning with pytorch3d. arXiv preprint arXiv:2007.08501 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.271, + 0.785, + 0.311 + ], + "angle": 0, + "content": "26. Srinivasan, P.P., Deng, B., Zhang, X., Tancik, M., Mildenhall, B., Barron, J.T.: Nerv: Neural reflectance and visibility fields for relighting and view synthesis. In: CVPR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.311, + 0.785, + 0.339 + ], + "angle": 0, + "content": "27. Tulsiani, S., Zhou, T., Efros, A.A., Malik, J.: Multi-view supervision for single-view reconstruction via differentiable ray consistency. In: CVPR (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.339, + 0.785, + 0.38 + ], + "angle": 0, + "content": "28. Verbin, D., Hedman, P., Mildenhall, B., Zickler, T., Barron, J.T., Srinivasan, P.P.: Ref-nerf: Structured view-dependent appearance for neural radiance fields. In: CVPR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.38, + 0.785, + 0.42 + ], + "angle": 0, + "content": "29. Wang, P., Liu, L., Liu, Y., Theobalt, C., Komura, T., Wang, W.: Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. In: NeurIPS (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.421, + 0.785, + 0.462 + ], + "angle": 0, + "content": "30. Wang, Y., Han, Q., Habermann, M., Daniilidis, K., Theobalt, C., Liu, L.: Neus2: Fast learning of neural implicit surfaces for multi-view reconstruction. In: ICCV (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.462, + 0.785, + 0.489 + ], + "angle": 0, + "content": "31. Wu, T., Wang, J., Pan, X., Xu, X., Theobalt, C., Liu, Z., Lin, D.: Voxurf: Voxel-based efficient and accurate neural surface reconstruction. In: ICLR (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.489, + 0.785, + 0.53 + ], + "angle": 0, + "content": "32. Xie, Y., Takikawa, T., Saito, S., Litany, O., Yan, S., Khan, N., Tombari, F., Tompkin, J., Sitzmann, V., Sridhar, S.: Neural fields in visual computing and beyond. Computer Graphics Forum 41(2), 641-676 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.53, + 0.785, + 0.571 + ], + "angle": 0, + "content": "33. Yan, X., Yang, J., Yumer, E., Guo, Y., Lee, H.: Perspective transformer nets: Learning single-view 3d object reconstruction without 3d supervision. In: NeurIPS (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.571, + 0.785, + 0.612 + ], + "angle": 0, + "content": "34. Yao, Y., Zhang, J., Liu, J., Qu, Y., Fang, T., McKinnon, D., Tsin, Y., Quan, L.: Neilf: Neural incident light field for material and lighting estimation. In: ECCV (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.612, + 0.785, + 0.639 + ], + "angle": 0, + "content": "35. Yariv, L., Gu, J., Kasten, Y., Lipman, Y.: Volume rendering of neural implicit surfaces. In: NeurIPS (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.639, + 0.785, + 0.68 + ], + "angle": 0, + "content": "36. Yariv, L., Kasten, Y., Moran, D., Galun, M., Atzmon, M., Ronen, B., Lipman, Y.: Multiview neural surface reconstruction by disentangling geometry and appearance. In: NeurIPS (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.68, + 0.785, + 0.707 + ], + "angle": 0, + "content": "37. Yu, A., Ye, V., Tancik, M., Kanazawa, A.: pixelnerf: Neural radiance fields from one or few images. In: CVPR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.707, + 0.785, + 0.734 + ], + "angle": 0, + "content": "38. Yu, Z., Peng, S., Niemeyer, M., Sattler, T., Geiger, A.: Monosdf: Exploring monocular geometric cues for neural implicit surface reconstruction. In: NeurIPS (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.734, + 0.785, + 0.762 + ], + "angle": 0, + "content": "39. Zhang, J., Yao, Y., Li, S., Liu, J., Fang, T., McKinnon, D., Tsin, Y., Quan, L.: Neilf++: Inter-reflectable light fields for geometry and material estimation (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.762, + 0.785, + 0.789 + ], + "angle": 0, + "content": "40. Zhang, K., Riegler, G., Snavely, N., Koltun, V.: Nerf++: Analyzing and improving neural radiance fields. arXiv preprint arXiv:2010.07492 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.789, + 0.785, + 0.83 + ], + "angle": 0, + "content": "41. Zhang, X., Srinivasan, P.P., Deng, B., Debevec, P., Freeman, W.T., Barron, J.T.: Nerfactor: Neural factorization of shape and reflectance under an unknown illumination. ACM Transactions on Graphics 40(6), 1-18 (2021)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.83 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/Rethinking Directional Parameterization in Neural Implicit Surface Reconstruction/8b296952-4d44-42db-affd-9e9843feb2a6_origin.pdf b/2024/Rethinking Directional Parameterization in Neural Implicit Surface Reconstruction/8b296952-4d44-42db-affd-9e9843feb2a6_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..29248a87634ee9049628e73e63262d923bf36022 --- /dev/null +++ b/2024/Rethinking Directional Parameterization in Neural Implicit Surface Reconstruction/8b296952-4d44-42db-affd-9e9843feb2a6_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2ca9f385fc7c96ac18974b06c08f81a53a24ba4b5731809a505b85330d24678b +size 3398350 diff --git a/2024/Rethinking Directional Parameterization in Neural Implicit Surface Reconstruction/full.md b/2024/Rethinking Directional Parameterization in Neural Implicit Surface Reconstruction/full.md new file mode 100644 index 0000000000000000000000000000000000000000..e3925da771ea1cc9b453c33cb13200032eb0002a --- /dev/null +++ b/2024/Rethinking Directional Parameterization in Neural Implicit Surface Reconstruction/full.md @@ -0,0 +1,328 @@ +# Rethinking Directional Parameterization in Neural Implicit Surface Reconstruction + +Zijie Jiang $^{1\star}$ , Tianhan Xu $^{2\star}$ , and Hiroharu Kato $^{2\star}$ + +$^{1}$ Tokyo Institute of Technology + +2 Preferred Networks, Inc. + +zjiang@ok.sc.e.titech.ac.jp, {xutianhan,hkato}@preferred.jp + +Abstract. Multi-view 3D surface reconstruction using neural implicit representations has made notable progress by modeling the geometry and view-dependent radiance fields within a unified framework. However, their effectiveness in reconstructing objects with specular or complex surfaces is typically biased by the directional parameterization used in their view-dependent radiance network. Viewing direction and reflection direction are the two most commonly used directional parameterizations but have their own limitations. Typically, utilizing the viewing direction usually struggles to correctly decouple the geometry and appearance of objects with highly specular surfaces, while using the reflection direction tends to yield overly smooth reconstructions for concave or complex structures. In this paper, we analyze their failed cases in detail and propose a novel hybrid directional parameterization to address their limitations in a unified form. Extensive experiments demonstrate the proposed hybrid directional parameterization consistently delivered satisfactory results in reconstructing objects with a wide variety of materials, geometry and appearance, whereas using other directional parameterizations faces challenges in reconstructing certain objects. Moreover, the proposed hybrid directional parameterization is nearly parameter-free and can be effortlessly applied in any existing neural surface reconstruction method. + +Keywords: surface reconstruction $\cdot$ neural implicit representation $\cdot$ directional parameterization + +# 1 Introduction + +Recent advances in neural fields [32] have revolutionized the field of 3D geometry reconstruction from multi-view images [22,29,36]. Compared to traditional multiview stereo methods [7], a key advantage of using neural fields is their ability to model and optimize complex light reflections on object surfaces, which improves robustness to specular or non-Lambertian surfaces. Since a radiance network is trained jointly with a geometry network via differentiable rendering, the radiance network is able to take full advantage of geometry and view-dependent information such as viewing direction to decouple the geometry and appearance. + +![](images/848f7d5786464271261be10c8c9c749d9cce2a9633a67977212e2ef8db7da901.jpg) + +![](images/7b34a9ac335a9451512f4c90bba12b883774b83ad1ec963a6cb01bc54c08921d.jpg) + +![](images/099237128abfdb40b6bf75a0ce9fb5a4f426596d2c67c318581b0b651766fb47.jpg) + +![](images/507ddd77285c5b5dd94f02408ad3064e463563e5a920fe039d2fbb8e3fe40b8a.jpg) + +![](images/21f81a33a9b0f020207a8ef732704172d6dd3e0b0a4f5ae7324e49dc251de890.jpg) + +![](images/e62ad7cdac22ed7fdd0f32b43c377ed0be9fc86750d6949c3bcb4c80b0865e6d.jpg) +Reference RGB + +![](images/c4ea86767a1d98e7b906447a4a0b8cbaadc48e1b103b930a8aa578362aa21184.jpg) +(a) Viewing dir. + +![](images/d0e9d9a9ff1764223b2e32ea5210c63d4f6fb241468d1cfb5f777156d28b0585.jpg) +(b) Reflection dir. +Fig. 1: Comparison of reconstruction results using different directional parameterizations in [2]. Using the viewing directional parameterization fails to reconstruct objects with specular surfaces, while using the reflection directional parameterization tends to result in incorrect geometry for objects with concave structures. Using our proposed hybrid parametrization shows consistently satisfactory reconstruction results compared to the existing parameterizations. + +![](images/2e0de4335a1138e2bcb7c99d421f7d3d822ecad98a1e783bf8c39d8be5717979.jpg) +(c) Hybrid dir. (Ours) + +![](images/25b34e5738b73517dbc1077716d3dc1e507b285e6d57439c26809a9ca40f2491.jpg) +Ground-truth Mesh + +Since the decomposition of geometry and appearance is inherently an ill-posed problem $[40]^3$ , it is essential to introduce an appropriate inductive bias into the radiance network to facilitate correct decomposition. For example, simply inputting the viewing direction to the radiance network sometimes leads to incorrect reconstructions of highly specular objects (Fig. 1 (a)). To address this issue, the recent study [28] suggests that providing the reflection direction of rays considering the learned local surface geometry (i.e., normals) would result in a simpler radiance field to learn and facilitate correct decomposition, especially for specular objects. Some subsequent studies have shown that adopting this parameterization results in satisfactory reconstructions of highly specular objects [8, 13, 14] and also non-reflecting objects [8]. However, although this parameterization seems promising (Fig. 1 (b) top), we observed that it still does not work well when reconstructing objects with complex structures such as concave surfaces (Fig. 1 (b) bottom). Interestingly, while using viewing direction fails to recover specular surfaces, it succeeds in recovering such concave structures (Fig. 1 (a) bottom). Based on these, we pose the question: Is there a unified directional parameterization that can effectively handle both specular surfaces and complex structures? + +In this paper, to answer the question, we start by studying the limitations of existing directional parameterizations and provide the first in-depth analysis of why using reflection direction may fail to reconstruct objects with concave or complex structures. As mentioned above, one important feature of the reflection directional parameterization is its dependency on the learnable geometry (i.e., + +normals). Our insight into this fact is that for sampled 3D points along a sampled ray during training, while utilizing reflection directions of sampled points near the intersecting surface is helpful in decoupling geometry and appearance, the use of reflection directions of sampled points elsewhere may hinder the accurate reconstruction of geometry. Specifically, we identified that the cause of this issue lies in the sampled points beyond a certain distance from the intersecting surface during the optimization. As we will discuss in Sec. 3.2, the normals of these points computed from learnable geometry may either be associated with surfaces unrelated to the intersection or introduce high-frequency variations in the input of the radiance network, which hinders the proper updating of geometry and radiance. This hindrance is particularly noticeable in the reconstruction of concave surfaces. The analysis also indirectly explains the advantage of using viewing directional parameterization in reconstructing concave or complex structures, as it is non-learnable and remains constant for all sampled points along a sampled ray. + +Based on our insights and analysis, we propose a novel hybrid directional parameterization to address the limitations of existing directional parameterizations in a unified form. The main idea is to utilize the predicted distance to the surface for each sampled point: if the distance is small, indicating proximity to the intersecting surface, it is preferable to use its reflection direction as the directional parameterization, which facilitates the correct decomposition of geometry and appearance; in other cases, we instead utilize the non-learnable viewing direction to avoid the optimization issue mentioned above. The proposed hybrid directional parameterization smoothly transitions from the reflection direction to the viewing direction based on the predicted distance to the surface. The rate of this directional transition is controlled by a learnable scaling factor. We integrate the proposed hybrid directional parameterization into different methods and conduct extensive experiments to validate its effectiveness across various datasets. Compared to the existing viewing and reflection directional parameterizations, our proposed hybrid directional parameterization achieves consistently satisfactory results, as shown in Fig. 1 (c). Moreover, the proposed hybrid directional parameterization is nearly parameter-free and can be effortlessly applied in any existing neural implicit surface reconstruction method. + +To summarize, our contributions are as follows: + +- We are the first to provide an in-depth analysis of why the use of reflection directional parameterization tends to reconstruct incorrect geometry of objects with concave or complex structures. +- We propose a novel hybrid directional parameterization to address the limitations of existing directional parameterizations, which can effectively handle both specular surfaces and complex structures in a unified form. +- We validate the effectiveness of the proposed directional parameterization across different methods on various datasets, and demonstrate an overall superiority compared to existing directional parameterizations. + +# 2 Related Work + +# 2.1 Neural Implicit Surface Reconstruction + +With the significant advancement of neural implicit representations [3,16-19,23, 24,37] and differentiable rendering [10,11,17,20,25,27,33] in recent years, neural implicit surface reconstruction methods [21,29,35,36] have emerged as a promising alternative to traditional methods in the field of image-based 3D surface reconstruction. Compared to traditional multi-view stereo methods, neural implicit surface methods employ a more flexible and representative implicit shape representation, such as the signed distance function (SDF), and introduce the directional radiance field [17] to model the view-dependent appearance of the target object. Subsequent works on neural implicit surface reconstruction have mostly built upon the foundation of NeuS [29] or VolSDF [35], and made improvements for specific scenarios including fast rendering [30, 31], sparse-view inputs [15], large-scale scenes [12,38] and more. + +# 2.2 Reconstruction for Specular Objects + +Despite neural implicit surface reconstruction methods consider the view-dependent appearance of objects by incorporating the directional radiance field, reconstruction for specular objects remains a challenging task due to its significant shape-radiance ambiguity. Several works [14, 26, 34, 39, 41] model the view-dependent appearance of specular objects by incorporating additional physical components (e.g., materials and lighting), with the hope of jointly recovering both the geometric shape and physical parameters. However, as the number of unknown parameters increases, these works typically rely on a good initial geometry to ensure favorable results, leading to a chicken-and-egg problem. On the other hand, Ref-NeRF [28] was the first to propose the use of a radiance field conditioned on reflection direction to better decouple the geometry and radiance of specular objects. Subsequent neural implicit surface reconstruction methods targeting specular objects [8, 13, 14] have also adopted this directional parameterization. However, these methods often fail to correctly recover the geometry of objects with concave or complex structures. + +# 3 Method + +In this section, we introduce a novel directional parameterization for neural implicit surface reconstruction methods. We first give a brief review on neural implicit surface reconstruction and its commonly used directional parameterizations, including viewing and reflection directions (Sec. 3.1). In the next, we summarize the observed failed reconstructions when utilizing existing directional parameterizations and conduct a detailed analysis of their causes (Sec. 3.2). Finally, a hybrid directional parameterization is proposed based on our analysis, which can effectively address the limitations of existing directional parameterizations in a unified form (Sec. 3.3). + +# 3.1 Preliminaries of Neural Implicit Surface Reconstruction + +Neural implicit surface reconstruction methods [2,6,29,36] typically represent the 3D object as the combination of a signed distance function (SDF) network $\Theta_{f}$ , which maps a spatial position $\mathbf{x} \in \mathbb{R}^3$ to its SDF value $f(\mathbf{x}) \in \mathbb{R}$ , and a directional radiance network $\Theta_{c}$ , which maps a spatial position $\mathbf{x} \in \mathbb{R}^3$ and a directional vector $\mathbf{d} \in \mathbb{S}^2$ to the view-dependent color at $\mathbf{x}$ . The surface $\mathcal{S}$ of the object is extracted as the zero-level set of $f(\mathbf{x})$ . To enable the joint optimization of the SDF network $\Theta_{f}$ and radiance network $\Theta_{c}$ using only 2D image supervisions, volume rendering is typically utilized [29,35]. Given the sampled pixels in an image, camera rays $\{\mathbf{r}(t) = \mathbf{o} + t\mathbf{d}_{\mathrm{view}}|t > 0\}$ are cast through the camera origin $\mathbf{o}$ and the sampled pixels, where $\mathbf{d}_{\mathrm{view}}$ is the unit directional vector of the ray. To obtain the rendered colors of the sampled rays, a number of points are sampled along each ray, and queried for their SDF values and view-dependent colors from $\Theta_{f}$ and $\Theta_{c}$ , where positional encoding (PE) [17] is typically incorporated. A relationship between the SDF value and opaque density is then formulated, such as using S-density [29], to translate the queried SDFs and colors to the ray color akin to NeRF [17]. Finally, the training loss is calculated from the difference between the rendered ray colors and the ground-truth colors of the sampled pixels. + +Directional parameterization. The key to the ability of the radiance network to model view-dependent appearance lies in its incorporation of a directional vector $\mathbf{d}$ as an input. Common directional parameterizations for $\mathbf{d}$ includes the viewing direction $\mathbf{d}_{\mathrm{view}}$ and reflection direction $\mathbf{d}_{\mathrm{ref}}$ . The viewing directional parameterization is intuitively straightforward, as it models the view-dependent appearance from the perspective of the observer. The reflection directional parameterization is a little more complicated, which encodes the viewing direction $\mathbf{d}_{\mathrm{view}}$ and local normal direction $\mathbf{n}$ into a new compact directional vector $\mathbf{d}_{\mathrm{ref}}$ , by considering the interaction between light and surface. Intuitively, it represents the direction of the incoming light, assuming that the surface is specular. $\mathbf{d}_{\mathrm{ref}}$ is formulated as follows: + +$$ +\mathbf {d} _ {\text {r e f}} = 2 \left(\mathbf {d} _ {\text {v i e w}} \cdot \mathbf {n}\right) \mathbf {n} - \mathbf {d} _ {\text {v i e w}}, \tag {1} +$$ + +where $\mathbf{n}$ can be derived from the gradient of the SDF as $\mathbf{n}(\mathbf{x}) = \frac{\nabla f(\mathbf{x})}{\|\nabla f(\mathbf{x})\|}$ . + +# 3.2 Observations and Analysis + +Observations. Although existing neural implicit surface reconstruction methods widely adopt either viewing direction or reflection direction as their directional parameterization, both types of parameterizations have their respective limitations when it comes to reconstructing certain objects. Using viewing direction often faces difficulty in correctly reconstructing objects with specular surfaces (e.g., Fig. 1 (a) and Fig. 2 (c)). On the other hand, while using reflection direction can reconstruct specular objects with simple geometry (e.g., Fig. 1 (b) top and Fig. 2 (d)), it often results in overly smooth reconstructions with + +![](images/e1a21e9151b98fee629c4833abe336f42d92846b00706ce5e0fe689eb69c53d5.jpg) +Object w/ concave structures + +![](images/cfc024136b15d5757e12dde8818d625263edcf258efeac08e867a7250c185e29.jpg) +(a)Viewing dir. + +![](images/f11731abf1ca7093a2f0609bb46e6b7db2d06a9d79b34bb9c8279e48a8c7d6d3.jpg) +(b) Reflection dir. (X + +![](images/1b16b6cacd0b31273c79cebae3df4579e77aa0d071b851ea41ef920c94f235ad.jpg) +Object w/ specular surfaces + +![](images/ef33002b3c0a319f8e4f9158dec273b5c951c3aece42f4f8e7c320eee5fa5e77.jpg) +(c) Viewing dir. + +![](images/f3041073eba5d32491ccc070bd37360b70f38f343e5c09ceb8443f9057a9126f.jpg) +(d) Reflection dir. + +![](images/229953c5b020bdb986ee665ce42d2dada63d5319a341120842b80ac7b314d7c2.jpg) +Fig. 2: Scenarios where existing directional parameterizations succeed and struggle. Reconstruction results are obtained by integrating the two directional parameterizations into NeuS [29]. + +missing concave surfaces when reconstructing objects with concave or complex structures (e.g., Fig. 1 (b) bottom and Fig. 2 (b)). Interestingly, using viewing direction may yield better results in these parts with complex or concave surfaces (e.g., Fig. 1 (a) bottom and Fig. 2 (a)). + +Based on these observations, we raise the question: Is there a unified directional parameterization that can effectively handle both specular surfaces and complex geometry? To answer this, we start by analyzing the reasons behind their reconstruction failures: + +Analysis of failures when using viewing direction. An analysis of why using viewing direction is unsuitable for modeling the view-dependent appearance of specular objects is provided in Verbin et al. [28], which points out that the radiance network conditioned on viewing direction is poorly suited for interpolation. We consider this characteristic leads to difficulties in correctly decoupling the geometry and appearance of specular objects when using viewing direction. On the other hand, reflection directional parameterization can effectively leverage the prior knowledge of the interaction between light and surface, which facilitates the correct decoupling of geometry and appearance, especially for specular objects. + +Analysis of failures when using reflection direction. In contrast, no work has yet provided a detailed analysis of why using reflection directional parameterization fails in the reconstruction of concave or complex structures, making us the first to provide this analysis. We first highlight two key differences between viewing direction and reflection direction: the viewing direction is non-learnable throughout the optimization and remains constant along a sampled ray, whereas the reflection direction 1) depends on learnable geometry (i.e., normals) and 2) may vary significantly during the optimization. We argue that these two properties of the reflection direction may pose challenges to the accurate recovery of geometry, especially for concave or complex structures. + +A 2D toy example is first illustrated on Fig. 3 (left) to better demonstrate why the first property of reflection direction may hinder the recovery of geometry, which assumes the scenario of recovering an "L"-shaped concave structure. A sampled ray intersecting the surface and the sampled points on the ray along with the normals at these points are illustrated. In this case, we notice that the normals at sampled points that are slightly distant from the intersecting surface + +![](images/17c616476e02fbcff23c408146bd60f710bf6cbdf8d8537f15742b2db917e0fe.jpg) +Fig. 3: Impact of incorporating normals on view-dependent radiance modeling. (Left) Normals at sampled points may be influenced by unrelated surfaces other than the intersection surface. (Top right) Smooth surfaces yield similarity in normals and reflection directions for the sampled points, while (Bottom right) surfaces with intricate local details, e.g., concavities, induce a scattered distribution of normals and reflection directions (Ref. dirs), particularly pronounced slightly distant from the vicinity of the zero-level set, which adversely affects geometry optimization. The details are explained in Sec. 3.2. + +![](images/c16c9a63a54b5ce76d95767850412eb63dfa57439162f10c8349a42fe9dade6e.jpg) + +![](images/f55cc30f7ed1382eecb8ca217173167ca976b89be02017cc7a7f3afab754bb3f.jpg) + +do not necessarily relate to the normal of the intersecting surface (the bottom side of the zero level-set), but rather to the normals of surfaces elsewhere (the left side of the zero level-set). As a result, since the training loss of existing neural implicit surface reconstruction methods is usually ray-based, which is calculated using the local color of the intersecting surface, using this loss to optimize the normals of irrelevant surfaces elsewhere may introduce ambiguity in the optimization of geometry and result in incorrect geometric updates. + +In addition, the second property of the reflection direction is more closely related to the phenomenon that using reflection directional parameterization tends to reconstruct overly smooth surfaces with missing concave or complex structures. Assuming that the geometry reconstruction begins with a smooth surface exhibiting local flatness (Fig. 3 top-right), which aligns with the geometric initialization practices commonly employed by existing neural implicit surface reconstruction methods [6,29], and the geometry of the target object includes concave structures (Fig. 3 bottom-right). We observed that for the same sampled ray, when the surface geometry is updated from a smooth surface to a concave surface, the normals and reflection directions of the same sampled points also change. This change is particularly pronounced for points that are slightly distant from the surface, and could be further pronounced due to the presence of Positional Encoding (PE). The noticeable changes in reflection directions at these points within each surface update introduce high-frequency variances to + +the directional inputs of the radiance network. This makes learning a stable radiance field challenging. To be specific, assuming that the radiance network conditioned on reflection direction has already learned to minimize the color loss to some extent before the surface update, the substantial change in directional inputs due to the update could likely lead to an increase in color loss, which hinders the optimization and is susceptible to falling into local minima. + +In contrast, we notice that using the viewing direction often leads to concave or overly intricate surface details (although some may be incorrect), as shown in Fig. 1 (a) and Fig. 2 (a) (c). We consider that this is because the viewing direction does not rely on the surface geometry, which means that changes in geometry do not affect the optimization of the radiance network conditioned on viewing direction. This allows the SDF can undergo more "aggressive" updates without considering the impact on radiance. + +# 3.3 Hybrid Directional Parameterization + +Building upon the analysis from the previous section, we discovered that viewing direction and reflection direction seem to have a spatially complementary relationship: On one hand, for sampled points near the intersecting surface, the normals here accurately reflect the normals of the intersecting surface, thereby enabling the reflection direction to better model the interaction between light and surface, which in turn promotes the correct decomposition of geometry and appearance. On the other hand, for other sampled points along the ray, using the reflection direction might introduce irrelevant surface geometry into the optimization, or high-frequency variations in the directional inputs of the radiance network, both potentially hindering the optimization of the geometry and radiance networks. By contrast, using the viewing direction at these sampled points can avoid these issues. Motivated by these thoughts, we propose a spatial-aware directional representation, which transitions from reflection direction $\mathbf{d}_{\mathrm{ref}}$ to viewing direction $\mathbf{d}_{\mathrm{view}}$ based on the distance to the surface (i.e., the absolute value of SDF). We refer to this fused direction as hybrid direction $\mathbf{d}_{\mathrm{hyb}}$ , formulated as: + +$$ +\mathbf {d} _ {\mathrm {h y b}} = \text {n o r m a l i z e} (\alpha \cdot \mathbf {d} _ {\text {r e f}} + (1 - \alpha) \cdot \mathbf {d} _ {\text {v i e w}}). \tag {2} +$$ + +Here, $\alpha \in [0,1]$ represents the blend weight. We normalize the resulting linear blend to a unit vector thereby representing a valid direction. The blend weight $\alpha$ is implemented as + +$$ +\alpha = \exp (- \gamma \cdot \det a c h (| f (\mathbf {x}) |)), \tag {3} +$$ + +where $\gamma \in \mathbb{R}^{+}$ is an estimable parameter that controls the transition speed between reflection/viewing direction. The gradient of the SDF is detached to prevent interference with geometric updates. At the surface of the object, i.e., when SDF equals 0, $\alpha$ simply becomes 1, which means that the hybrid direction aligns with the reflection direction. As $\mathbf{x}$ moves further away from the surface, $\alpha$ gradually approaches 0, indicating that the direction tends towards the viewing + +direction. Here, $\alpha$ can be any bell-shaped function w.r.t. the SDF value. In this case, we utilize the PDF of the Laplace distribution for simplicity. + +In the next section, extensive experimental results validate the superiority of the proposed hybrid directional parameterization compared to existing parameterizations. + +# 4 Experiments + +# 4.1 Experimental Settings + +Datasets. We conduct experiments on objects from different datasets, including the DTU dataset [1], the Shiny Blender dataset [28], and the real captured dataset [9]. The DTU dataset [1] contains indoor multi-view captures of diverse objects, including challenging conditions for geometry reconstruction such as non-Lambertian surfaces and complex structures. We use the same 15 objects as previous works [2, 29, 31, 36] for evaluation. In addition, we utilize the foreground masks provided by Yariv et al. [36]. The Shinny Blender dataset [28] is a synthetic dataset rendered in Blender [4], which contains different objects with specular surfaces. We follow the experimental settings used in Ref-NeuS [8] on this dataset, which selects four objects. The real captured dataset [9] contains a set of three outdoor captures of real objects with complex geometry and specular surfaces. We select the "toycar" scene for evaluation. + +Implementation details. We integrate the proposed hybrid directional parameterization into two types of neural surface reconstruction methods: NeuS [29] and NeuDA [2]. The former utilizes a deep MLP network as the scene representation, while the latter leverages a shallow MLP decoder combined with multi-level voxel grids. Both methods adopt a radiance network conditioned on viewing direction in their original versions. The modification of the directional input from the viewing direction to the proposed hybrid direction can be accomplished with less than 10 lines of code. We additionally implement the radiance network that depends on reflection direction for both methods, to facilitate a comprehensive comparison of different directional parameterizations. Regarding the learnable parameter $\gamma$ in Eqn. 3, we implement it as $\gamma = \exp (10\cdot \gamma_b)$ . We assign an initial value of 0.3 to $\gamma_{b}$ except for the "toaster" scene from the Shiny Blender dataset. For the "toaster" scene, an initial value of 0.1 is assigned since it results in better reconstruction. The training losses of the modified methods are identical to those of their original versions. The number of training iterations is set to 300,000 to ensure convergence. + +Evaluation metrics. For the DTU dataset, we follow the evaluation protocol of previous works [2, 29] and evaluate the Chamfer Distance (CD) between the estimated and ground-truth geometry. For the Shiny Blender dataset, we follow the evaluation protocol proposed in Ref-NeuS [8]. It evaluates the accuracy (Acc.) instead of the CD for unbiased evaluation, and also evaluates the mean angular error (MAE) between the ground-truth normal maps and the rendered normal maps. For the real captured dataset [9], we only provide a qualitative evaluation, as ground-truth geometry is not provided in this dataset. + +Table 1: Quantitative evaluation in terms of Chamfer distance of integrating different directional parameterizations in NeuS [29] and NeuDA [2] on the DTU dataset [1]. The experimental objects are categorized into diffuse objects and specular objects. The numbers in the headers denote the scene IDs. Bold results have the best score. + +
MethodsScan of Diffuse ObjectsScan of Specular ObjectsTotal Avg.
243740556583105106114118122Avg.636997110Avg.
NeuS [29]0.830.980.560.370.591.450.780.520.360.450.450.671.130.600.951.431.030.77
NeuS, w/ reflection dir.1.394.390.730.391.341.490.920.590.360.460.541.141.110.631.460.921.031.11
NeuS, w/ hybrid dir. (Ours)0.810.970.540.380.611.470.770.510.350.450.470.671.100.600.850.900.860.72
NeuD A [2]0.510.760.390.370.561.370.790.500.340.420.460.591.080.571.130.800.900.67
NeuD A, w/ reflection dir.0.581.030.630.370.791.450.890.510.370.430.500.691.030.550.900.760.810.72
NeuD A, w/ hybrid dir. (Ours)0.530.770.400.360.571.380.760.470.330.410.470.591.020.540.760.750.770.63
+ +# 4.2 Comparison of Different Directional Parameterizations + +We first evaluate the performance of integrating different directional parameterizations into the two baseline methods NeuS [29] and NeuDA [2] on the DTU dataset [1]. The experimental objects from the DTU dataset are categorized into diffuse objects and specular objects for better demonstration. Tab. 1 summarizes the quantitative results. For both baseline methods, we notice that substituting the original viewing directional parameterization with the reflection direction only yields better results for a minority of specular objects. For diffuse objects and the remaining specular objects, employing reflection directional parameterization demonstrates significant degradation in performance. Considering the relatively complex geometry of DTU objects, we attribute this to the optimization issue analyzed in Sec. 3.2. Conversely, the incorporation of the proposed hybrid directional parameterization enables both modified methods to yield consistently superior performance for specular objects while achieving results on par with the original version for diffuse objects. The first two rows in Fig. 4 showcase the visual comparison of adopting different directional parameterizations for the reconstruction of two specular objects from the DTU dataset. Both baseline methods fail to reconstruct the specular surfaces of objects. On the other hand, although integrating the reflection direction yields better reconstruction for specular objects with simple geometry (e.g., Scan 63), they fail to reconstruct the specular surfaces at the concave structure in Scan 97. In contrast, integrating the proposed hybrid direction achieves the best reconstruction results in both scenarios. More qualitative comparisons on the DTU dataset are available in our supplementary. + +We further validate the advantage of the proposed hybrid directional parameterization on the Shiny Blender dataset [28], as demonstrated in Tab. 2. Both baseline methods exhibit large reconstruction errors on the surfaces of highly specular objects in this dataset. By contrast, due to the relatively simple geometry of objects, replacing the viewing directional parameterization with the reflection direction results in a significant improvement in reconstruction quality on most objects in this dataset, except for the "coffee" object. For the reconstruction of the "coffee" object, adopting the reflection directional parameterization fails to recover the concave structure formed by the inner wall of the coffee cup + +Table 2: Quantitative evaluation of integrating different directional parameterizations in NeuS [29] and NeuDA [2] on the Shiny Blender dataset [28]. Bold results have the best score. + +
MethodshelmettoastercoffeecarAvg.
Acc. ↓MAE ↓Acc. ↓MAE ↓Acc. ↓MAE ↓Acc. ↓MAE ↓Acc. ↓MAE ↓
NeuS [29]4.883.203.312.851.971.060.860.952.762.02
NeuS, w/ reflection dir.0.310.360.390.943.301.370.480.731.120.85
NeuS, w/ hybrid dir. (Ours)0.290.340.450.920.640.550.430.720.450.63
NeuDA [2]5.733.775.983.391.160.810.871.033.442.25
NeuDA, w/ reflection dir.0.340.371.031.713.661.530.620.871.411.12
NeuDA, w/ hybrid dir. (Ours)0.320.360.951.540.550.570.600.840.610.83
+ +![](images/131a8fe431a68b9861d9cfbec23962ca947dcc1212996f54c71bb1d6babac13d.jpg) +Fig. 4: Qualitative comparison of integrating different directional parameterizations in NeuS [29] and NeuDA [2] on the DTU dataset [1] and Shiny Blender dataset [28]. + +and the surface of the water, as shown in the third row in Fig. 4. On the other hand, the integration of our proposed hybrid directional parameterization enables the successful reconstruction of both the specular surfaces and concave structures, obtaining consistently superior reconstructions of all objects in this dataset, compared to the other two directional parameterizations. + +Lastly, we qualitatively present the reconstruction results on the "toycar" scene from the real captured scene in Fig. 5, which contains challenging reflective glass and complex geometry. Both modified methods employing the proposed hybrid directional parameterization succeed in recovering the complex geometry and the smooth surface of the reflective glass. In contrast, using either the reflection direction or the original viewing direction results in reconstructions that are either overly smooth, missing concave surface structures, or provide incorrect surface reconstructions at the reflective glass areas. + +As the summary of this section, compared to the viewing directional parameterization which struggles with reconstructing specular surfaces, and the reflection directional parameterization which faces difficulties in reconstructing + +![](images/f1ddbdcf03a7995ee9e1642aeef2d91281a9b0a40740e67696defd744c005857.jpg) +Reference RGB + +![](images/582c6f6b5e5bc4dc73ab6d6a7068e1ff9cbf15e31e60d7671bb241236f142f2f.jpg) +NeuS + +![](images/f65ffc2a1635eb666adf303450035e9005d10ed0167b153c7afc7a996e1bf67e.jpg) +NeuS, w/ reflection dir. + +![](images/36b7a36abd21e147b2ba39440f62dcf822f3d3bc998ce6ffe72a4e3588a1d0d5.jpg) +NeuS, w/ hybrid dir. + +![](images/647d7e95eb8b2b34ddcdd9114d4a5ce8a9343e5e8b5a066952e0841bd9e7313b.jpg) +NeuDA + +![](images/9c12491570133dd22f48aceeb65c46207a02faa25a6eaedfb1e31a92cbbf0ab3.jpg) +NeuDA,w/ reflection dir. + +![](images/62708bc77f6e852906135c5fe5ca35789b3e6ad45b2ca7af5ca2303158c6605a.jpg) +NeuDA, w/ hybrid dir. +Fig. 5: Qualitative comparison of integrating different directional parameterizations in NeuS [29] and NeuDA [2] on the "toycar" scene from the real captured dataset [9]. + +Table 3: Quantitative evaluation in terms of Chamfer distance on the DTU dataset [1]. The numbers in the header denote the scene IDs. Bold results have the best score. + +
Methods243740556365698397105106110114118122Avg.
IDR [36]1.631.870.630.481.040.790.771.331.160.760.670.900.420.510.530.90
NeuS [29]0.830.980.560.371.130.590.601.450.950.780.521.430.360.450.450.77
Voxurf [31]0.650.740.390.350.960.640.851.581.010.680.601.110.370.450.470.72
NeuDA [2]0.510.760.390.371.080.560.571.371.130.790.500.800.340.420.460.67
Ours (NeuS)0.810.970.540.381.100.610.601.470.850.770.510.900.350.450.470.72
Ours (NeuDA)0.530.770.400.361.020.570.541.380.760.760.470.760.330.410.470.63
+ +complex structures, our proposed hybrid directional parameterization effectively handles both scenarios in a unified form. This benefit is attributed to our hybrid directional parameterization effectively combining the advantages of both viewing and reflection directional parameterizations while avoiding their drawbacks. + +# 4.3 Comparison to State-of-the-Art + +Although our main purpose is not to develop a best-performing surface reconstruction system, we demonstrate that we can still achieve state-of-the-art performance by integrating the proposed hybrid directional parameterization into existing methods. For brevity, we refer to NeuS [29] and NeuDA [2] incorporating the proposed hybrid directional parameterization as Ours (NeuS) and Ours (NeuDA) in the following text. Tab. 3 reports the Chamfer distance (CD) metrics of different methods on the DTU dataset [1]. For a fair comparison, all methods in Tab. 3 utilize foreground masks provided by Yariv et al. [36] during training. Both Ours (NeuS) and Ours (NeuDA) improve their baselines by a certain margin in the average CD scores, and Ours (NeuDA) achieves the best performance compared to previous SOTA methods in the average CD metric. + +Tab. 4 reports the quantitative evaluation of different methods on the Shiny Blender dataset [28]. The results of [5,6,8,22,28,35] are taken from the results + +Table 4: Quantitative evaluation on the Shiny Blender dataset [28]. “-” denotes that the results are not available. Bold results have the best score. + +
MethodshelmettoastercoffeecarAvg.
Acc. ↓MAE ↓Acc. ↓MAE ↓Acc. ↓MAE ↓Acc. ↓MAE ↓Acc. ↓MAE ↓
Ref-NeRF [28]-29.48-42.87-12.24-14.93-24.88
NeuS [29]4.883.203.312.851.971.060.860.952.762.02
UNISURF [22]1.691.783.755.512.883.152.341.982.673.11
VolSDF [35]1.551.372.022.532.232.280.581.041.601.81
NeuralWarp [5]2.251.945.903.511.542.040.651.072.592.14
Geo-NeuS [6]0.742.366.353.763.856.362.885.673.464.54
NeuD A [2]5.733.775.983.391.160.810.871.033.442.25
Ref-NeuD [8]0.290.380.421.470.770.990.370.800.460.91
Ours (NeuD)0.290.340.450.920.640.550.430.720.450.63
Ours (NeuD A)0.320.360.951.540.550.570.600.840.610.83
+ +![](images/5345e63abcb455c37bf47dc647cd2c32a47dc52cfa5b3d647f27a5bc83c438c9.jpg) +Reference RGB + +![](images/7b63b837f0ae60ce15848f50b4d247b11f3b67130e40ea30028d52c3ed0b270d.jpg) +$\gamma_{b} = 0.0$ $\mathrm{CD}\downarrow$ : 0.91 + +![](images/b86a875fb70518f84ec19ada47e3191ac3f37279be34cf98e4fed23a0161d205.jpg) +Fig. 6: Reconstruction results using different initial values of $\gamma_{b}$ . The main differences are highlighted in red boxes. + +![](images/98e8e6462d067ee35fef37e76fa72e6daeb2e2b6dd599acd77ca342c3e3b8c5c.jpg) +$\gamma_{b} = 0.1$ $\mathrm{CD}\downarrow$ : 0.81 + +![](images/991b02be9d23b8e105ad4d35305552bc4109874e2c23fb513bfd34ec7ca08da8.jpg) +$\gamma_{b} = 0.2$ $\mathrm{CD}\downarrow$ :0.80 +$\gamma_{b} = 0.3$ $\mathrm{CD}\downarrow$ :0.76 + +![](images/815bd1c01f24cef8d31899ddd76bb93ec6cc74473784b89d674824c11024653d.jpg) +$\gamma_{b} = 0.4$ $\mathrm{CD}\downarrow$ :0.77 + +reported in Ge et al. [8]. Both Ours (NeuS) and Ours (NeuDA) achieve better average MAE metrics than the previous SOTA method Ref-NeuS [8], and Ours (NeuS) achieve the best overall performance in terms of both Acc. and MAE metrics on this dataset. We refer to our supplementary for the qualitative comparisons between Ours (NeuS) and Ref-NeuS [8]. + +# 4.4 Ablation Studies + +In this section, we conduct ablation studies on several key design choices of the proposed hybrid directional parameterization by using Ours (NeuDA). + +The initial value of $\gamma_{b}$ . We first validate the impact of setting different initial values for $\gamma_{b}$ on the reconstruction results. Fig. 6 compares the obtained reconstruction results using five different initial values for $\gamma_{b}$ . We find that when the initial value of $\gamma_{b}$ is set low, our method tends to recover incorrect geometries similar to those caused by using the reflection directional parameterization. This phenomenon remains consistent with our analysis in Sec 3.2, as a smaller initial value for $\gamma_{b}$ increases the likelihood of involving irrelevant surfaces and high-frequency variations in the directional inputs of the radiance network during the optimization. We empirically find the best initial value of $\gamma_{b}$ to be 0.3 for most experimental objects. More discussions regarding the parameter $\gamma_{b}$ and blend weight $\alpha$ can be found in our supplementary material. + +Detach SDF gradient in Eqn. 3 or not? Fig. 7 compares the reconstruction results with and without detaching the gradient of the estimated SDF in Eqn. 3. We observe that our method struggles to reconstruct the fine-grained geometry + +![](images/37d1e782e6978f39ae6919287d644778fffde258e7ada473f5099d1f3ee9e413.jpg) +Reference RGB + +![](images/f1e1b747b70a6657ce267fb061d7a820fb68314fbff530d15311cb55ebbc9f66.jpg) +w/o SDF gradient detachment + +![](images/841dfae42ca3709046551a7fd75a51efdb527306de2daec685287e3580f7cc6a.jpg) +w/SDF gradient detachment (Ours) + +![](images/8ec8eccae10f1f88e6ddac6b1efbabea238d4e8c59eaa6d5848f20711b4a2985.jpg) +Reference RGB +Fig.7: Reconstruction results with and without detaching the gradient of the estimated SDF in Eqn. 3. + +![](images/f25fc9a33d52989287e48f6ee97beba950a6a8a20200368881dc50c2a4f9422b.jpg) +post-PE fusion +Fig. 8: Reconstruction results fusing directions before and after positional encoding (PE). + +![](images/5ece8d59d27bd17e8e4cfbc09c3a462a2a4ea1c8a0114c686d9ca89c80482a6f.jpg) +pre-PE fusion (Ours) + +when the gradient of the estimated SDF is not detached in Eqn. 3. We consider that introducing the gradient of the estimated SDF places an additional burden on the learning of both the geometry and radiance networks, making it easier to fall into local minima. Thereby, we choose to detach the gradient of SDF in Eqn. 3. + +Fuse the viewing direction and reflection direction before or after positional encoding? We introduce the fusion of viewing direction and reflection direction in Eqn. 2 before the operation of position encoding (PE). We compare the reconstruction results with different fusion orders, specifically the proposed pre-PE fusion: $\mathrm{PE}(\mathbf{d}_{\mathrm{hyb}})$ (ours) and alternative approach using post-PE fusion: $\alpha \cdot \mathrm{PE}(\mathbf{d}_{\mathrm{ref}}) + (1 - \alpha) \cdot \mathrm{PE}(\mathbf{d}_{\mathrm{view}})$ . The comparison is shown in Fig. 8. We observe that fusing directions after the PE operation may result in degraded reconstruction results. We consider that fusing the directional information in the high-dimension space may increase the complexity of the learning space of the radiance network, which brings challenges in finding optimal reconstruction results. Thereby, we choose to fuse the viewing direction and reflection direction before the operation of positional encoding. + +# 5 Conclusion and Limitation + +In this paper, we conduct an in-depth investigation into the topic of directional parameterization in neural implicit surface reconstruction. We reveal that the commonly used directional parameterizations in modeling the view-dependent radiance field, such as viewing and reflection directions, both have their limitations and none of them can effectively handle both specular surfaces and complex structures. A novel hybrid directional parameterization is introduced to achieve this goal, which softly fuses the viewing and reflection directions based on the distance of the sampled points from the surface, uniting the advantages of each representation. Extensive experiments demonstrate that our hybrid directional parameterization outperforms the existing directional parameterizations in reconstructing objects with a wide variety of materials, geometry, and appearance. + +One limitation of our method is that the optimal choice of initial value for $\gamma_{b}$ may vary depending on the target object, which would require additional efforts for parameter tuning. An automated method for selecting the optimal initial value of $\gamma_{b}$ for target objects could be a potential future work for us. + +# References + +1. Aanaes, H., Jensen, R.R., Vogiatzis, G., Tola, E., Dahl, A.B.: Large-scale data for multiple-view stereopsis. IJCV 120, 153-168 (2016) +2. Cai, B., Huang, J., Jia, R., Lv, C., Fu, H.: Neuda: Neural deformable anchor for high-fidelity implicit surface reconstruction. In: CVPR (2023) +3. Chen, Z., Zhang, H.: Learning implicit fields for generative shape modeling. In: CVPR (2019) +4. Community, B.O.: Blender - a 3D modelling and rendering package. Blender Foundation, Stichting Blender Foundation, Amsterdam (2018) +5. Darmon, F., Bascle, B., Devaux, J., Monasse, P., Aubry, M.: Improving neural implicit surfaces geometry with patch warping. In: CVPR (2022) +6. Fu, Q., Xu, Q., Ong, Y.S., Tao, W.: Geo-neus: Geometry-consistent neural implicit surfaces learning for multi-view reconstruction. In: NeurIPS (2022) +7. Furukawa, Y., Hernández, C., et al.: Multi-view stereo: A tutorial. Foundations and Trends® in Computer Graphics and Vision 9(1-2), 1-148 (2015) +8. Ge, W., Hu, T., Zhao, H., Liu, S., Chen, Y.C.: Ref-neus: Ambiguity-reduced neural implicit surface learning for multi-view reconstruction with reflection. In: ICCV (2023) +9. Hedman, P., Srinivasan, P.P., Mildenhall, B., Barron, J.T., Debevec, P.: Baking neural radiance fields for real-time view synthesis. In: CVPR (2021) +0. Kato, H., Beker, D., Morariu, M., Ando, T., Matsuoka, T., Kehl, W., Gaidon, A.: Differentiable rendering: A survey. arXiv preprint arXiv:2006.12057 (2020) +1. Li, T.M., Aittala, M., Durand, F., Lehtinen, J.: Differentiable monte carlo ray tracing through edge sampling. ACM Transactions on Graphics 37(6), 1-11 (2018) +2. Li, Z., Müller, T., Evans, A., Taylor, R.H., Unberath, M., Liu, M.Y., Lin, C.H.: Neuralangelo: High-fidelity neural surface reconstruction. In: CVPR (2023) +3. Liang, R., Chen, H., Li, C., Chen, F., Panneer, S., Vijaykumar, N.: Envidr: Implicit differentiable renderer with neural environment lighting. arXiv preprint arXiv:2303.13022 (2023) +4. Liu, Y., Wang, P., Lin, C., Long, X., Wang, J., Liu, L., Komura, T., Wang, W.: Nano: Neural geometry and brdf reconstruction of reflective objects from multiview images. ACM Transactions on Graphics (2023) +5. Long, X., Lin, C., Wang, P., Komura, T., Wang, W.: Sparseneus: Fast generalizable neural surface reconstruction from sparse views. In: ECCV (2022) +6. Mescheder, L., Oechsle, M., Niemeyer, M., Nowozin, S., Geiger, A.: Occupancy networks: Learning 3d reconstruction in function space. In: CVPR (2019) +7. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. In: ECCV (2020) +8. Müller, T., Evans, A., Schied, C., Keller, A.: Instant neural graphics primitives with a multiresolution hash encoding. ACM Transactions on Graphics 41(4), 1-15 (2022) +9. Niemeyer, M., Mescheder, L., Oechsle, M., Geiger, A.: Occupancy flow: 4d reconstruction by learning particle dynamics. In: ICCV (2019) +20. Niemeyer, M., Mescheder, L., Oechsle, M., Geiger, A.: Differentiable volumetric rendering: Learning implicit 3d representations without 3d supervision. In: CVPR (2020) +21. Niemeyer, M., Mescheder, L., Oechsle, M., Geiger, A.: Differentiable volumetric rendering: Learning implicit 3d representations without 3d supervision. In: CVPR (2020) + +22. Oechsle, M., Peng, S., Geiger, A.: Unisurf: Unifying neural implicit surfaces and radiance fields for multi-view reconstruction. In: ICCV (2021) +23. Park, J.J., Florence, P., Straub, J., Newcombe, R., Lovegrove, S.: Deepsdf: Learning continuous signed distance functions for shape representation. In: CVPR (2019) +24. Peng, S., Niemeyer, M., Mescheder, L., Pollefeys, M., Geiger, A.: Convolutional occupancy networks. In: ECCV (2020) +25. Ravi, N., Reizenstein, J., Novotny, D., Gordon, T., Lo, W.Y., Johnson, J., Gkioxari, G.: Accelerating 3d deep learning with pytorch3d. arXiv preprint arXiv:2007.08501 (2020) +26. Srinivasan, P.P., Deng, B., Zhang, X., Tancik, M., Mildenhall, B., Barron, J.T.: Nerv: Neural reflectance and visibility fields for relighting and view synthesis. In: CVPR (2021) +27. Tulsiani, S., Zhou, T., Efros, A.A., Malik, J.: Multi-view supervision for single-view reconstruction via differentiable ray consistency. In: CVPR (2017) +28. Verbin, D., Hedman, P., Mildenhall, B., Zickler, T., Barron, J.T., Srinivasan, P.P.: Ref-nerf: Structured view-dependent appearance for neural radiance fields. In: CVPR (2022) +29. Wang, P., Liu, L., Liu, Y., Theobalt, C., Komura, T., Wang, W.: Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. In: NeurIPS (2021) +30. Wang, Y., Han, Q., Habermann, M., Daniilidis, K., Theobalt, C., Liu, L.: Neus2: Fast learning of neural implicit surfaces for multi-view reconstruction. In: ICCV (2023) +31. Wu, T., Wang, J., Pan, X., Xu, X., Theobalt, C., Liu, Z., Lin, D.: Voxurf: Voxel-based efficient and accurate neural surface reconstruction. In: ICLR (2023) +32. Xie, Y., Takikawa, T., Saito, S., Litany, O., Yan, S., Khan, N., Tombari, F., Tompkin, J., Sitzmann, V., Sridhar, S.: Neural fields in visual computing and beyond. Computer Graphics Forum 41(2), 641-676 (2022) +33. Yan, X., Yang, J., Yumer, E., Guo, Y., Lee, H.: Perspective transformer nets: Learning single-view 3d object reconstruction without 3d supervision. In: NeurIPS (2016) +34. Yao, Y., Zhang, J., Liu, J., Qu, Y., Fang, T., McKinnon, D., Tsin, Y., Quan, L.: Neilf: Neural incident light field for material and lighting estimation. In: ECCV (2022) +35. Yariv, L., Gu, J., Kasten, Y., Lipman, Y.: Volume rendering of neural implicit surfaces. In: NeurIPS (2021) +36. Yariv, L., Kasten, Y., Moran, D., Galun, M., Atzmon, M., Ronen, B., Lipman, Y.: Multiview neural surface reconstruction by disentangling geometry and appearance. In: NeurIPS (2020) +37. Yu, A., Ye, V., Tancik, M., Kanazawa, A.: pixelnerf: Neural radiance fields from one or few images. In: CVPR (2021) +38. Yu, Z., Peng, S., Niemeyer, M., Sattler, T., Geiger, A.: Monosdf: Exploring monocular geometric cues for neural implicit surface reconstruction. In: NeurIPS (2022) +39. Zhang, J., Yao, Y., Li, S., Liu, J., Fang, T., McKinnon, D., Tsin, Y., Quan, L.: Neilf++: Inter-reflectable light fields for geometry and material estimation (2023) +40. Zhang, K., Riegler, G., Snavely, N., Koltun, V.: Nerf++: Analyzing and improving neural radiance fields. arXiv preprint arXiv:2010.07492 (2020) +41. Zhang, X., Srinivasan, P.P., Deng, B., Debevec, P., Freeman, W.T., Barron, J.T.: Nerfactor: Neural factorization of shape and reflectance under an unknown illumination. ACM Transactions on Graphics 40(6), 1-18 (2021) \ No newline at end of file diff --git a/2024/Rethinking Directional Parameterization in Neural Implicit Surface Reconstruction/images.zip b/2024/Rethinking Directional Parameterization in Neural Implicit Surface Reconstruction/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..fca82651fd44f067adf910cf28e77baa84265a09 --- /dev/null +++ b/2024/Rethinking Directional Parameterization in Neural Implicit Surface Reconstruction/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a4766171cf09a36fcc35f1043a2dfbde56ba0d2818b845f4e86a07b9ab58f3d3 +size 482197 diff --git a/2024/Rethinking Directional Parameterization in Neural Implicit Surface Reconstruction/layout.json b/2024/Rethinking Directional Parameterization in Neural Implicit Surface Reconstruction/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..08ca8c525435e036f9756b5a556b8779219a6dca --- /dev/null +++ b/2024/Rethinking Directional Parameterization in Neural Implicit Surface Reconstruction/layout.json @@ -0,0 +1,9708 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 156, + 111, + 459, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 156, + 111, + 459, + 148 + ], + "spans": [ + { + "bbox": [ + 156, + 111, + 459, + 148 + ], + "type": "text", + "content": "Rethinking Directional Parameterization in Neural Implicit Surface Reconstruction" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 186, + 167, + 427, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 186, + 167, + 427, + 180 + ], + "spans": [ + { + "bbox": [ + 186, + 167, + 427, + 180 + ], + "type": "text", + "content": "Zijie Jiang" + }, + { + "bbox": [ + 186, + 167, + 427, + 180 + ], + "type": "inline_equation", + "content": "^{1\\star}" + }, + { + "bbox": [ + 186, + 167, + 427, + 180 + ], + "type": "text", + "content": ", Tianhan Xu" + }, + { + "bbox": [ + 186, + 167, + 427, + 180 + ], + "type": "inline_equation", + "content": "^{2\\star}" + }, + { + "bbox": [ + 186, + 167, + 427, + 180 + ], + "type": "text", + "content": ", and Hiroharu Kato" + }, + { + "bbox": [ + 186, + 167, + 427, + 180 + ], + "type": "inline_equation", + "content": "^{2\\star}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 241, + 189, + 373, + 201 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 241, + 189, + 373, + 201 + ], + "spans": [ + { + "bbox": [ + 241, + 189, + 373, + 201 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 241, + 189, + 373, + 201 + ], + "type": "text", + "content": " Tokyo Institute of Technology" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 251, + 201, + 361, + 212 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 251, + 201, + 361, + 212 + ], + "spans": [ + { + "bbox": [ + 251, + 201, + 361, + 212 + ], + "type": "text", + "content": "2 Preferred Networks, Inc." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 166, + 213, + 446, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 213, + 446, + 224 + ], + "spans": [ + { + "bbox": [ + 166, + 213, + 446, + 224 + ], + "type": "text", + "content": "zjiang@ok.sc.e.titech.ac.jp, {xutianhan,hkato}@preferred.jp" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 160, + 247, + 455, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 247, + 455, + 468 + ], + "spans": [ + { + "bbox": [ + 160, + 247, + 455, + 468 + ], + "type": "text", + "content": "Abstract. Multi-view 3D surface reconstruction using neural implicit representations has made notable progress by modeling the geometry and view-dependent radiance fields within a unified framework. However, their effectiveness in reconstructing objects with specular or complex surfaces is typically biased by the directional parameterization used in their view-dependent radiance network. Viewing direction and reflection direction are the two most commonly used directional parameterizations but have their own limitations. Typically, utilizing the viewing direction usually struggles to correctly decouple the geometry and appearance of objects with highly specular surfaces, while using the reflection direction tends to yield overly smooth reconstructions for concave or complex structures. In this paper, we analyze their failed cases in detail and propose a novel hybrid directional parameterization to address their limitations in a unified form. Extensive experiments demonstrate the proposed hybrid directional parameterization consistently delivered satisfactory results in reconstructing objects with a wide variety of materials, geometry and appearance, whereas using other directional parameterizations faces challenges in reconstructing certain objects. Moreover, the proposed hybrid directional parameterization is nearly parameter-free and can be effortlessly applied in any existing neural surface reconstruction method." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 159, + 477, + 453, + 499 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 477, + 453, + 499 + ], + "spans": [ + { + "bbox": [ + 159, + 477, + 453, + 499 + ], + "type": "text", + "content": "Keywords: surface reconstruction " + }, + { + "bbox": [ + 159, + 477, + 453, + 499 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 477, + 453, + 499 + ], + "type": "text", + "content": " neural implicit representation " + }, + { + "bbox": [ + 159, + 477, + 453, + 499 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 477, + 453, + 499 + ], + "type": "text", + "content": " directional parameterization" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 518, + 230, + 530 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 518, + 230, + 530 + ], + "spans": [ + { + "bbox": [ + 132, + 518, + 230, + 530 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 540, + 482, + 638 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 540, + 482, + 638 + ], + "spans": [ + { + "bbox": [ + 130, + 540, + 482, + 638 + ], + "type": "text", + "content": "Recent advances in neural fields [32] have revolutionized the field of 3D geometry reconstruction from multi-view images [22,29,36]. Compared to traditional multiview stereo methods [7], a key advantage of using neural fields is their ability to model and optimize complex light reflections on object surfaces, which improves robustness to specular or non-Lambertian surfaces. Since a radiance network is trained jointly with a geometry network via differentiable rendering, the radiance network is able to take full advantage of geometry and view-dependent information such as viewing direction to decouple the geometry and appearance." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 642, + 481, + 664 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 642, + 481, + 664 + ], + "spans": [ + { + "bbox": [ + 133, + 642, + 481, + 664 + ], + "type": "text", + "content": "* Equal contribution. Work done while Zijie Jiang interned at Preferred Networks, Inc." + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 141, + 123, + 194, + 175 + ], + "blocks": [ + { + "bbox": [ + 141, + 123, + 194, + 175 + ], + "lines": [ + { + "bbox": [ + 141, + 123, + 194, + 175 + ], + "spans": [ + { + "bbox": [ + 141, + 123, + 194, + 175 + ], + "type": "image", + "image_path": "848f7d5786464271261be10c8c9c749d9cce2a9633a67977212e2ef8db7da901.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 212, + 123, + 269, + 176 + ], + "blocks": [ + { + "bbox": [ + 212, + 123, + 269, + 176 + ], + "lines": [ + { + "bbox": [ + 212, + 123, + 269, + 176 + ], + "spans": [ + { + "bbox": [ + 212, + 123, + 269, + 176 + ], + "type": "image", + "image_path": "7b34a9ac335a9451512f4c90bba12b883774b83ad1ec963a6cb01bc54c08921d.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 279, + 123, + 335, + 176 + ], + "blocks": [ + { + "bbox": [ + 279, + 123, + 335, + 176 + ], + "lines": [ + { + "bbox": [ + 279, + 123, + 335, + 176 + ], + "spans": [ + { + "bbox": [ + 279, + 123, + 335, + 176 + ], + "type": "image", + "image_path": "099237128abfdb40b6bf75a0ce9fb5a4f426596d2c67c318581b0b651766fb47.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 345, + 123, + 402, + 176 + ], + "blocks": [ + { + "bbox": [ + 345, + 123, + 402, + 176 + ], + "lines": [ + { + "bbox": [ + 345, + 123, + 402, + 176 + ], + "spans": [ + { + "bbox": [ + 345, + 123, + 402, + 176 + ], + "type": "image", + "image_path": "507ddd77285c5b5dd94f02408ad3064e463563e5a920fe039d2fbb8e3fe40b8a.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 424, + 123, + 479, + 176 + ], + "blocks": [ + { + "bbox": [ + 424, + 123, + 479, + 176 + ], + "lines": [ + { + "bbox": [ + 424, + 123, + 479, + 176 + ], + "spans": [ + { + "bbox": [ + 424, + 123, + 479, + 176 + ], + "type": "image", + "image_path": "21f81a33a9b0f020207a8ef732704172d6dd3e0b0a4f5ae7324e49dc251de890.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 143, + 185, + 196, + 239 + ], + "blocks": [ + { + "bbox": [ + 143, + 185, + 196, + 239 + ], + "lines": [ + { + "bbox": [ + 143, + 185, + 196, + 239 + ], + "spans": [ + { + "bbox": [ + 143, + 185, + 196, + 239 + ], + "type": "image", + "image_path": "e62ad7cdac22ed7fdd0f32b43c377ed0be9fc86750d6949c3bcb4c80b0865e6d.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 142, + 242, + 195, + 251 + ], + "lines": [ + { + "bbox": [ + 142, + 242, + 195, + 251 + ], + "spans": [ + { + "bbox": [ + 142, + 242, + 195, + 251 + ], + "type": "text", + "content": "Reference RGB" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 216, + 185, + 269, + 240 + ], + "blocks": [ + { + "bbox": [ + 216, + 185, + 269, + 240 + ], + "lines": [ + { + "bbox": [ + 216, + 185, + 269, + 240 + ], + "spans": [ + { + "bbox": [ + 216, + 185, + 269, + 240 + ], + "type": "image", + "image_path": "c4ea86767a1d98e7b906447a4a0b8cbaadc48e1b103b930a8aa578362aa21184.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 214, + 243, + 266, + 252 + ], + "lines": [ + { + "bbox": [ + 214, + 243, + 266, + 252 + ], + "spans": [ + { + "bbox": [ + 214, + 243, + 266, + 252 + ], + "type": "text", + "content": "(a) Viewing dir." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 281, + 185, + 335, + 240 + ], + "blocks": [ + { + "bbox": [ + 281, + 185, + 335, + 240 + ], + "lines": [ + { + "bbox": [ + 281, + 185, + 335, + 240 + ], + "spans": [ + { + "bbox": [ + 281, + 185, + 335, + 240 + ], + "type": "image", + "image_path": "d0e9d9a9ff1764223b2e32ea5210c63d4f6fb241468d1cfb5f777156d28b0585.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 277, + 243, + 335, + 252 + ], + "lines": [ + { + "bbox": [ + 277, + 243, + 335, + 252 + ], + "spans": [ + { + "bbox": [ + 277, + 243, + 335, + 252 + ], + "type": "text", + "content": "(b) Reflection dir." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 270, + 482, + 338 + ], + "lines": [ + { + "bbox": [ + 130, + 270, + 482, + 338 + ], + "spans": [ + { + "bbox": [ + 130, + 270, + 482, + 338 + ], + "type": "text", + "content": "Fig. 1: Comparison of reconstruction results using different directional parameterizations in [2]. Using the viewing directional parameterization fails to reconstruct objects with specular surfaces, while using the reflection directional parameterization tends to result in incorrect geometry for objects with concave structures. Using our proposed hybrid parametrization shows consistently satisfactory reconstruction results compared to the existing parameterizations." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 346, + 185, + 400, + 239 + ], + "blocks": [ + { + "bbox": [ + 346, + 185, + 400, + 239 + ], + "lines": [ + { + "bbox": [ + 346, + 185, + 400, + 239 + ], + "spans": [ + { + "bbox": [ + 346, + 185, + 400, + 239 + ], + "type": "image", + "image_path": "2e0de4335a1138e2bcb7c99d421f7d3d822ecad98a1e783bf8c39d8be5717979.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 348, + 243, + 396, + 262 + ], + "lines": [ + { + "bbox": [ + 348, + 243, + 396, + 262 + ], + "spans": [ + { + "bbox": [ + 348, + 243, + 396, + 262 + ], + "type": "text", + "content": "(c) Hybrid dir. (Ours)" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 424, + 185, + 477, + 239 + ], + "blocks": [ + { + "bbox": [ + 424, + 185, + 477, + 239 + ], + "lines": [ + { + "bbox": [ + 424, + 185, + 477, + 239 + ], + "spans": [ + { + "bbox": [ + 424, + 185, + 477, + 239 + ], + "type": "image", + "image_path": "25b34e5738b73517dbc1077716d3dc1e507b285e6d57439c26809a9ca40f2491.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 415, + 243, + 479, + 252 + ], + "lines": [ + { + "bbox": [ + 415, + 243, + 479, + 252 + ], + "spans": [ + { + "bbox": [ + 415, + 243, + 479, + 252 + ], + "type": "text", + "content": "Ground-truth Mesh" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 360, + 482, + 575 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 360, + 482, + 575 + ], + "spans": [ + { + "bbox": [ + 130, + 360, + 482, + 575 + ], + "type": "text", + "content": "Since the decomposition of geometry and appearance is inherently an ill-posed problem " + }, + { + "bbox": [ + 130, + 360, + 482, + 575 + ], + "type": "inline_equation", + "content": "[40]^3" + }, + { + "bbox": [ + 130, + 360, + 482, + 575 + ], + "type": "text", + "content": ", it is essential to introduce an appropriate inductive bias into the radiance network to facilitate correct decomposition. For example, simply inputting the viewing direction to the radiance network sometimes leads to incorrect reconstructions of highly specular objects (Fig. 1 (a)). To address this issue, the recent study [28] suggests that providing the reflection direction of rays considering the learned local surface geometry (i.e., normals) would result in a simpler radiance field to learn and facilitate correct decomposition, especially for specular objects. Some subsequent studies have shown that adopting this parameterization results in satisfactory reconstructions of highly specular objects [8, 13, 14] and also non-reflecting objects [8]. However, although this parameterization seems promising (Fig. 1 (b) top), we observed that it still does not work well when reconstructing objects with complex structures such as concave surfaces (Fig. 1 (b) bottom). Interestingly, while using viewing direction fails to recover specular surfaces, it succeeds in recovering such concave structures (Fig. 1 (a) bottom). Based on these, we pose the question: Is there a unified directional parameterization that can effectively handle both specular surfaces and complex structures?" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 130, + 576, + 482, + 637 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 576, + 482, + 637 + ], + "spans": [ + { + "bbox": [ + 130, + 576, + 482, + 637 + ], + "type": "text", + "content": "In this paper, to answer the question, we start by studying the limitations of existing directional parameterizations and provide the first in-depth analysis of why using reflection direction may fail to reconstruct objects with concave or complex structures. As mentioned above, one important feature of the reflection directional parameterization is its dependency on the learnable geometry (i.e.," + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 246, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 246, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 246, + 102 + ], + "type": "text", + "content": "Jiang and Xu et al." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 133, + 642, + 481, + 666 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 642, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 133, + 642, + 481, + 666 + ], + "type": "text", + "content": "3 If the capacity of the radiance network is sufficiently high, even for a fixed geometry such as a sphere, the image reconstruction loss can converge to zero [40]." + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 297 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 297 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 297 + ], + "type": "text", + "content": "normals). Our insight into this fact is that for sampled 3D points along a sampled ray during training, while utilizing reflection directions of sampled points near the intersecting surface is helpful in decoupling geometry and appearance, the use of reflection directions of sampled points elsewhere may hinder the accurate reconstruction of geometry. Specifically, we identified that the cause of this issue lies in the sampled points beyond a certain distance from the intersecting surface during the optimization. As we will discuss in Sec. 3.2, the normals of these points computed from learnable geometry may either be associated with surfaces unrelated to the intersection or introduce high-frequency variations in the input of the radiance network, which hinders the proper updating of geometry and radiance. This hindrance is particularly noticeable in the reconstruction of concave surfaces. The analysis also indirectly explains the advantage of using viewing directional parameterization in reconstructing concave or complex structures, as it is non-learnable and remains constant for all sampled points along a sampled ray." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 298, + 482, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 298, + 482, + 514 + ], + "spans": [ + { + "bbox": [ + 130, + 298, + 482, + 514 + ], + "type": "text", + "content": "Based on our insights and analysis, we propose a novel hybrid directional parameterization to address the limitations of existing directional parameterizations in a unified form. The main idea is to utilize the predicted distance to the surface for each sampled point: if the distance is small, indicating proximity to the intersecting surface, it is preferable to use its reflection direction as the directional parameterization, which facilitates the correct decomposition of geometry and appearance; in other cases, we instead utilize the non-learnable viewing direction to avoid the optimization issue mentioned above. The proposed hybrid directional parameterization smoothly transitions from the reflection direction to the viewing direction based on the predicted distance to the surface. The rate of this directional transition is controlled by a learnable scaling factor. We integrate the proposed hybrid directional parameterization into different methods and conduct extensive experiments to validate its effectiveness across various datasets. Compared to the existing viewing and reflection directional parameterizations, our proposed hybrid directional parameterization achieves consistently satisfactory results, as shown in Fig. 1 (c). Moreover, the proposed hybrid directional parameterization is nearly parameter-free and can be effortlessly applied in any existing neural implicit surface reconstruction method." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 147, + 517, + 356, + 528 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 147, + 517, + 356, + 528 + ], + "spans": [ + { + "bbox": [ + 147, + 517, + 356, + 528 + ], + "type": "text", + "content": "To summarize, our contributions are as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 552, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 138, + 552, + 481, + 587 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 552, + 481, + 587 + ], + "spans": [ + { + "bbox": [ + 138, + 552, + 481, + 587 + ], + "type": "text", + "content": "- We are the first to provide an in-depth analysis of why the use of reflection directional parameterization tends to reconstruct incorrect geometry of objects with concave or complex structures." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 590, + 480, + 626 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 590, + 480, + 626 + ], + "spans": [ + { + "bbox": [ + 138, + 590, + 480, + 626 + ], + "type": "text", + "content": "- We propose a novel hybrid directional parameterization to address the limitations of existing directional parameterizations, which can effectively handle both specular surfaces and complex structures in a unified form." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 629, + 480, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 629, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 629, + 480, + 665 + ], + "type": "text", + "content": "- We validate the effectiveness of the proposed directional parameterization across different methods on various datasets, and demonstrate an overall superiority compared to existing directional parameterizations." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 284, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 284, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 284, + 91, + 447, + 102 + ], + "type": "text", + "content": "Rethinking directional parameterization" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 237, + 127 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 237, + 127 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 237, + 127 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 138, + 358, + 152 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 138, + 358, + 152 + ], + "spans": [ + { + "bbox": [ + 132, + 138, + 358, + 152 + ], + "type": "text", + "content": "2.1 Neural Implicit Surface Reconstruction" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 157, + 481, + 301 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 157, + 481, + 301 + ], + "spans": [ + { + "bbox": [ + 132, + 157, + 481, + 301 + ], + "type": "text", + "content": "With the significant advancement of neural implicit representations [3,16-19,23, 24,37] and differentiable rendering [10,11,17,20,25,27,33] in recent years, neural implicit surface reconstruction methods [21,29,35,36] have emerged as a promising alternative to traditional methods in the field of image-based 3D surface reconstruction. Compared to traditional multi-view stereo methods, neural implicit surface methods employ a more flexible and representative implicit shape representation, such as the signed distance function (SDF), and introduce the directional radiance field [17] to model the view-dependent appearance of the target object. Subsequent works on neural implicit surface reconstruction have mostly built upon the foundation of NeuS [29] or VolSDF [35], and made improvements for specific scenarios including fast rendering [30, 31], sparse-view inputs [15], large-scale scenes [12,38] and more." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 318, + 343, + 331 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 318, + 343, + 331 + ], + "spans": [ + { + "bbox": [ + 132, + 318, + 343, + 331 + ], + "type": "text", + "content": "2.2 Reconstruction for Specular Objects" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 336, + 491, + 516 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 336, + 491, + 516 + ], + "spans": [ + { + "bbox": [ + 132, + 336, + 491, + 516 + ], + "type": "text", + "content": "Despite neural implicit surface reconstruction methods consider the view-dependent appearance of objects by incorporating the directional radiance field, reconstruction for specular objects remains a challenging task due to its significant shape-radiance ambiguity. Several works [14, 26, 34, 39, 41] model the view-dependent appearance of specular objects by incorporating additional physical components (e.g., materials and lighting), with the hope of jointly recovering both the geometric shape and physical parameters. However, as the number of unknown parameters increases, these works typically rely on a good initial geometry to ensure favorable results, leading to a chicken-and-egg problem. On the other hand, Ref-NeRF [28] was the first to propose the use of a radiance field conditioned on reflection direction to better decouple the geometry and radiance of specular objects. Subsequent neural implicit surface reconstruction methods targeting specular objects [8, 13, 14] have also adopted this directional parameterization. However, these methods often fail to correctly recover the geometry of objects with concave or complex structures." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 533, + 201, + 546 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 533, + 201, + 546 + ], + "spans": [ + { + "bbox": [ + 132, + 533, + 201, + 546 + ], + "type": "text", + "content": "3 Method" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 557, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 557, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 557, + 481, + 665 + ], + "type": "text", + "content": "In this section, we introduce a novel directional parameterization for neural implicit surface reconstruction methods. We first give a brief review on neural implicit surface reconstruction and its commonly used directional parameterizations, including viewing and reflection directions (Sec. 3.1). In the next, we summarize the observed failed reconstructions when utilizing existing directional parameterizations and conduct a detailed analysis of their causes (Sec. 3.2). Finally, a hybrid directional parameterization is proposed based on our analysis, which can effectively address the limitations of existing directional parameterizations in a unified form (Sec. 3.3)." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 246, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 246, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 246, + 102 + ], + "type": "text", + "content": "Jiang and Xu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 442, + 128 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 442, + 128 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 442, + 128 + ], + "type": "text", + "content": "3.1 Preliminaries of Neural Implicit Surface Reconstruction" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 136, + 482, + 351 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 136, + 482, + 351 + ], + "spans": [ + { + "bbox": [ + 130, + 136, + 482, + 351 + ], + "type": "text", + "content": "Neural implicit surface reconstruction methods [2,6,29,36] typically represent the 3D object as the combination of a signed distance function (SDF) network " + }, + { + "bbox": [ + 130, + 136, + 482, + 351 + ], + "type": "inline_equation", + "content": "\\Theta_{f}" + }, + { + "bbox": [ + 130, + 136, + 482, + 351 + ], + "type": "text", + "content": ", which maps a spatial position " + }, + { + "bbox": [ + 130, + 136, + 482, + 351 + ], + "type": "inline_equation", + "content": "\\mathbf{x} \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 130, + 136, + 482, + 351 + ], + "type": "text", + "content": " to its SDF value " + }, + { + "bbox": [ + 130, + 136, + 482, + 351 + ], + "type": "inline_equation", + "content": "f(\\mathbf{x}) \\in \\mathbb{R}" + }, + { + "bbox": [ + 130, + 136, + 482, + 351 + ], + "type": "text", + "content": ", and a directional radiance network " + }, + { + "bbox": [ + 130, + 136, + 482, + 351 + ], + "type": "inline_equation", + "content": "\\Theta_{c}" + }, + { + "bbox": [ + 130, + 136, + 482, + 351 + ], + "type": "text", + "content": ", which maps a spatial position " + }, + { + "bbox": [ + 130, + 136, + 482, + 351 + ], + "type": "inline_equation", + "content": "\\mathbf{x} \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 130, + 136, + 482, + 351 + ], + "type": "text", + "content": " and a directional vector " + }, + { + "bbox": [ + 130, + 136, + 482, + 351 + ], + "type": "inline_equation", + "content": "\\mathbf{d} \\in \\mathbb{S}^2" + }, + { + "bbox": [ + 130, + 136, + 482, + 351 + ], + "type": "text", + "content": " to the view-dependent color at " + }, + { + "bbox": [ + 130, + 136, + 482, + 351 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 130, + 136, + 482, + 351 + ], + "type": "text", + "content": ". The surface " + }, + { + "bbox": [ + 130, + 136, + 482, + 351 + ], + "type": "inline_equation", + "content": "\\mathcal{S}" + }, + { + "bbox": [ + 130, + 136, + 482, + 351 + ], + "type": "text", + "content": " of the object is extracted as the zero-level set of " + }, + { + "bbox": [ + 130, + 136, + 482, + 351 + ], + "type": "inline_equation", + "content": "f(\\mathbf{x})" + }, + { + "bbox": [ + 130, + 136, + 482, + 351 + ], + "type": "text", + "content": ". To enable the joint optimization of the SDF network " + }, + { + "bbox": [ + 130, + 136, + 482, + 351 + ], + "type": "inline_equation", + "content": "\\Theta_{f}" + }, + { + "bbox": [ + 130, + 136, + 482, + 351 + ], + "type": "text", + "content": " and radiance network " + }, + { + "bbox": [ + 130, + 136, + 482, + 351 + ], + "type": "inline_equation", + "content": "\\Theta_{c}" + }, + { + "bbox": [ + 130, + 136, + 482, + 351 + ], + "type": "text", + "content": " using only 2D image supervisions, volume rendering is typically utilized [29,35]. Given the sampled pixels in an image, camera rays " + }, + { + "bbox": [ + 130, + 136, + 482, + 351 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{r}(t) = \\mathbf{o} + t\\mathbf{d}_{\\mathrm{view}}|t > 0\\}" + }, + { + "bbox": [ + 130, + 136, + 482, + 351 + ], + "type": "text", + "content": " are cast through the camera origin " + }, + { + "bbox": [ + 130, + 136, + 482, + 351 + ], + "type": "inline_equation", + "content": "\\mathbf{o}" + }, + { + "bbox": [ + 130, + 136, + 482, + 351 + ], + "type": "text", + "content": " and the sampled pixels, where " + }, + { + "bbox": [ + 130, + 136, + 482, + 351 + ], + "type": "inline_equation", + "content": "\\mathbf{d}_{\\mathrm{view}}" + }, + { + "bbox": [ + 130, + 136, + 482, + 351 + ], + "type": "text", + "content": " is the unit directional vector of the ray. To obtain the rendered colors of the sampled rays, a number of points are sampled along each ray, and queried for their SDF values and view-dependent colors from " + }, + { + "bbox": [ + 130, + 136, + 482, + 351 + ], + "type": "inline_equation", + "content": "\\Theta_{f}" + }, + { + "bbox": [ + 130, + 136, + 482, + 351 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 136, + 482, + 351 + ], + "type": "inline_equation", + "content": "\\Theta_{c}" + }, + { + "bbox": [ + 130, + 136, + 482, + 351 + ], + "type": "text", + "content": ", where positional encoding (PE) [17] is typically incorporated. A relationship between the SDF value and opaque density is then formulated, such as using S-density [29], to translate the queried SDFs and colors to the ray color akin to NeRF [17]. Finally, the training loss is calculated from the difference between the rendered ray colors and the ground-truth colors of the sampled pixels." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 352, + 482, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 352, + 482, + 483 + ], + "spans": [ + { + "bbox": [ + 130, + 352, + 482, + 483 + ], + "type": "text", + "content": "Directional parameterization. The key to the ability of the radiance network to model view-dependent appearance lies in its incorporation of a directional vector " + }, + { + "bbox": [ + 130, + 352, + 482, + 483 + ], + "type": "inline_equation", + "content": "\\mathbf{d}" + }, + { + "bbox": [ + 130, + 352, + 482, + 483 + ], + "type": "text", + "content": " as an input. Common directional parameterizations for " + }, + { + "bbox": [ + 130, + 352, + 482, + 483 + ], + "type": "inline_equation", + "content": "\\mathbf{d}" + }, + { + "bbox": [ + 130, + 352, + 482, + 483 + ], + "type": "text", + "content": " includes the viewing direction " + }, + { + "bbox": [ + 130, + 352, + 482, + 483 + ], + "type": "inline_equation", + "content": "\\mathbf{d}_{\\mathrm{view}}" + }, + { + "bbox": [ + 130, + 352, + 482, + 483 + ], + "type": "text", + "content": " and reflection direction " + }, + { + "bbox": [ + 130, + 352, + 482, + 483 + ], + "type": "inline_equation", + "content": "\\mathbf{d}_{\\mathrm{ref}}" + }, + { + "bbox": [ + 130, + 352, + 482, + 483 + ], + "type": "text", + "content": ". The viewing directional parameterization is intuitively straightforward, as it models the view-dependent appearance from the perspective of the observer. The reflection directional parameterization is a little more complicated, which encodes the viewing direction " + }, + { + "bbox": [ + 130, + 352, + 482, + 483 + ], + "type": "inline_equation", + "content": "\\mathbf{d}_{\\mathrm{view}}" + }, + { + "bbox": [ + 130, + 352, + 482, + 483 + ], + "type": "text", + "content": " and local normal direction " + }, + { + "bbox": [ + 130, + 352, + 482, + 483 + ], + "type": "inline_equation", + "content": "\\mathbf{n}" + }, + { + "bbox": [ + 130, + 352, + 482, + 483 + ], + "type": "text", + "content": " into a new compact directional vector " + }, + { + "bbox": [ + 130, + 352, + 482, + 483 + ], + "type": "inline_equation", + "content": "\\mathbf{d}_{\\mathrm{ref}}" + }, + { + "bbox": [ + 130, + 352, + 482, + 483 + ], + "type": "text", + "content": ", by considering the interaction between light and surface. Intuitively, it represents the direction of the incoming light, assuming that the surface is specular. " + }, + { + "bbox": [ + 130, + 352, + 482, + 483 + ], + "type": "inline_equation", + "content": "\\mathbf{d}_{\\mathrm{ref}}" + }, + { + "bbox": [ + 130, + 352, + 482, + 483 + ], + "type": "text", + "content": " is formulated as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 245, + 495, + 480, + 506 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 245, + 495, + 480, + 506 + ], + "spans": [ + { + "bbox": [ + 245, + 495, + 480, + 506 + ], + "type": "interline_equation", + "content": "\\mathbf {d} _ {\\text {r e f}} = 2 \\left(\\mathbf {d} _ {\\text {v i e w}} \\cdot \\mathbf {n}\\right) \\mathbf {n} - \\mathbf {d} _ {\\text {v i e w}}, \\tag {1}", + "image_path": "aa9d35c17a65d8ab711d80a1d5c1656ef89387d554ef11971e303864408eda29.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 515, + 453, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 515, + 453, + 533 + ], + "spans": [ + { + "bbox": [ + 132, + 515, + 453, + 533 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 132, + 515, + 453, + 533 + ], + "type": "inline_equation", + "content": "\\mathbf{n}" + }, + { + "bbox": [ + 132, + 515, + 453, + 533 + ], + "type": "text", + "content": " can be derived from the gradient of the SDF as " + }, + { + "bbox": [ + 132, + 515, + 453, + 533 + ], + "type": "inline_equation", + "content": "\\mathbf{n}(\\mathbf{x}) = \\frac{\\nabla f(\\mathbf{x})}{\\|\\nabla f(\\mathbf{x})\\|}" + }, + { + "bbox": [ + 132, + 515, + 453, + 533 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 548, + 294, + 561 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 548, + 294, + 561 + ], + "spans": [ + { + "bbox": [ + 132, + 548, + 294, + 561 + ], + "type": "text", + "content": "3.2 Observations and Analysis" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": "Observations. Although existing neural implicit surface reconstruction methods widely adopt either viewing direction or reflection direction as their directional parameterization, both types of parameterizations have their respective limitations when it comes to reconstructing certain objects. Using viewing direction often faces difficulty in correctly reconstructing objects with specular surfaces (e.g., Fig. 1 (a) and Fig. 2 (c)). On the other hand, while using reflection direction can reconstruct specular objects with simple geometry (e.g., Fig. 1 (b) top and Fig. 2 (d)), it often results in overly smooth reconstructions with" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 284, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 284, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 284, + 91, + 447, + 102 + ], + "type": "text", + "content": "Rethinking directional parameterization" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 136, + 114, + 202, + 170 + ], + "blocks": [ + { + "bbox": [ + 136, + 114, + 202, + 170 + ], + "lines": [ + { + "bbox": [ + 136, + 114, + 202, + 170 + ], + "spans": [ + { + "bbox": [ + 136, + 114, + 202, + 170 + ], + "type": "image", + "image_path": "e1a21e9151b98fee629c4833abe336f42d92846b00706ce5e0fe689eb69c53d5.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 139, + 171, + 200, + 187 + ], + "lines": [ + { + "bbox": [ + 139, + 171, + 200, + 187 + ], + "spans": [ + { + "bbox": [ + 139, + 171, + 200, + 187 + ], + "type": "text", + "content": "Object w/ concave structures" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 207, + 114, + 249, + 170 + ], + "blocks": [ + { + "bbox": [ + 207, + 114, + 249, + 170 + ], + "lines": [ + { + "bbox": [ + 207, + 114, + 249, + 170 + ], + "spans": [ + { + "bbox": [ + 207, + 114, + 249, + 170 + ], + "type": "image", + "image_path": "cfc024136b15d5757e12dde8818d625263edcf258efeac08e867a7250c185e29.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 206, + 171, + 250, + 188 + ], + "lines": [ + { + "bbox": [ + 206, + 171, + 250, + 188 + ], + "spans": [ + { + "bbox": [ + 206, + 171, + 250, + 188 + ], + "type": "text", + "content": "(a)Viewing dir." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 261, + 114, + 301, + 170 + ], + "blocks": [ + { + "bbox": [ + 261, + 114, + 301, + 170 + ], + "lines": [ + { + "bbox": [ + 261, + 114, + 301, + 170 + ], + "spans": [ + { + "bbox": [ + 261, + 114, + 301, + 170 + ], + "type": "image", + "image_path": "f11731abf1ca7093a2f0609bb46e6b7db2d06a9d79b34bb9c8279e48a8c7d6d3.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 256, + 171, + 306, + 188 + ], + "lines": [ + { + "bbox": [ + 256, + 171, + 306, + 188 + ], + "spans": [ + { + "bbox": [ + 256, + 171, + 306, + 188 + ], + "type": "text", + "content": "(b) Reflection dir. (X" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 318, + 118, + 363, + 169 + ], + "blocks": [ + { + "bbox": [ + 318, + 118, + 363, + 169 + ], + "lines": [ + { + "bbox": [ + 318, + 118, + 363, + 169 + ], + "spans": [ + { + "bbox": [ + 318, + 118, + 363, + 169 + ], + "type": "image", + "image_path": "1b16b6cacd0b31273c79cebae3df4579e77aa0d071b851ea41ef920c94f235ad.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 314, + 171, + 371, + 187 + ], + "lines": [ + { + "bbox": [ + 314, + 171, + 371, + 187 + ], + "spans": [ + { + "bbox": [ + 314, + 171, + 371, + 187 + ], + "type": "text", + "content": "Object w/ specular surfaces" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 373, + 121, + 417, + 171 + ], + "blocks": [ + { + "bbox": [ + 373, + 121, + 417, + 171 + ], + "lines": [ + { + "bbox": [ + 373, + 121, + 417, + 171 + ], + "spans": [ + { + "bbox": [ + 373, + 121, + 417, + 171 + ], + "type": "image", + "image_path": "ef33002b3c0a319f8e4f9158dec273b5c951c3aece42f4f8e7c320eee5fa5e77.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 375, + 171, + 419, + 178 + ], + "lines": [ + { + "bbox": [ + 375, + 171, + 419, + 178 + ], + "spans": [ + { + "bbox": [ + 375, + 171, + 419, + 178 + ], + "type": "text", + "content": "(c) Viewing dir." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 425, + 121, + 470, + 171 + ], + "blocks": [ + { + "bbox": [ + 425, + 121, + 470, + 171 + ], + "lines": [ + { + "bbox": [ + 425, + 121, + 470, + 171 + ], + "spans": [ + { + "bbox": [ + 425, + 121, + 470, + 171 + ], + "type": "image", + "image_path": "f3041073eba5d32491ccc070bd37360b70f38f343e5c09ceb8443f9057a9126f.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 425, + 171, + 475, + 178 + ], + "lines": [ + { + "bbox": [ + 425, + 171, + 475, + 178 + ], + "spans": [ + { + "bbox": [ + 425, + 171, + 475, + 178 + ], + "type": "text", + "content": "(d) Reflection dir." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 445, + 179, + 457, + 190 + ], + "blocks": [ + { + "bbox": [ + 445, + 179, + 457, + 190 + ], + "lines": [ + { + "bbox": [ + 445, + 179, + 457, + 190 + ], + "spans": [ + { + "bbox": [ + 445, + 179, + 457, + 190 + ], + "type": "image", + "image_path": "229953c5b020bdb986ee665ce42d2dada63d5319a341120842b80ac7b314d7c2.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 200, + 480, + 232 + ], + "lines": [ + { + "bbox": [ + 131, + 200, + 480, + 232 + ], + "spans": [ + { + "bbox": [ + 131, + 200, + 480, + 232 + ], + "type": "text", + "content": "Fig. 2: Scenarios where existing directional parameterizations succeed and struggle. Reconstruction results are obtained by integrating the two directional parameterizations into NeuS [29]." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 258, + 479, + 305 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 258, + 479, + 305 + ], + "spans": [ + { + "bbox": [ + 130, + 258, + 479, + 305 + ], + "type": "text", + "content": "missing concave surfaces when reconstructing objects with concave or complex structures (e.g., Fig. 1 (b) bottom and Fig. 2 (b)). Interestingly, using viewing direction may yield better results in these parts with complex or concave surfaces (e.g., Fig. 1 (a) bottom and Fig. 2 (a))." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 130, + 306, + 480, + 353 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 306, + 480, + 353 + ], + "spans": [ + { + "bbox": [ + 130, + 306, + 480, + 353 + ], + "type": "text", + "content": "Based on these observations, we raise the question: Is there a unified directional parameterization that can effectively handle both specular surfaces and complex geometry? To answer this, we start by analyzing the reasons behind their reconstruction failures:" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 130, + 354, + 480, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 354, + 480, + 472 + ], + "spans": [ + { + "bbox": [ + 130, + 354, + 480, + 472 + ], + "type": "text", + "content": "Analysis of failures when using viewing direction. An analysis of why using viewing direction is unsuitable for modeling the view-dependent appearance of specular objects is provided in Verbin et al. [28], which points out that the radiance network conditioned on viewing direction is poorly suited for interpolation. We consider this characteristic leads to difficulties in correctly decoupling the geometry and appearance of specular objects when using viewing direction. On the other hand, reflection directional parameterization can effectively leverage the prior knowledge of the interaction between light and surface, which facilitates the correct decoupling of geometry and appearance, especially for specular objects." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 130, + 474, + 480, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 474, + 480, + 592 + ], + "spans": [ + { + "bbox": [ + 130, + 474, + 480, + 592 + ], + "type": "text", + "content": "Analysis of failures when using reflection direction. In contrast, no work has yet provided a detailed analysis of why using reflection directional parameterization fails in the reconstruction of concave or complex structures, making us the first to provide this analysis. We first highlight two key differences between viewing direction and reflection direction: the viewing direction is non-learnable throughout the optimization and remains constant along a sampled ray, whereas the reflection direction 1) depends on learnable geometry (i.e., normals) and 2) may vary significantly during the optimization. We argue that these two properties of the reflection direction may pose challenges to the accurate recovery of geometry, especially for concave or complex structures." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 130, + 594, + 480, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 480, + 665 + ], + "type": "text", + "content": "A 2D toy example is first illustrated on Fig. 3 (left) to better demonstrate why the first property of reflection direction may hinder the recovery of geometry, which assumes the scenario of recovering an \"L\"-shaped concave structure. A sampled ray intersecting the surface and the sampled points on the ray along with the normals at these points are illustrated. In this case, we notice that the normals at sampled points that are slightly distant from the intersecting surface" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 246, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 246, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 246, + 101 + ], + "type": "text", + "content": "Jiang and Xu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 135, + 118, + 312, + 285 + ], + "blocks": [ + { + "bbox": [ + 135, + 118, + 312, + 285 + ], + "lines": [ + { + "bbox": [ + 135, + 118, + 312, + 285 + ], + "spans": [ + { + "bbox": [ + 135, + 118, + 312, + 285 + ], + "type": "image", + "image_path": "17c616476e02fbcff23c408146bd60f710bf6cbdf8d8537f15742b2db917e0fe.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 298, + 482, + 387 + ], + "lines": [ + { + "bbox": [ + 130, + 298, + 482, + 387 + ], + "spans": [ + { + "bbox": [ + 130, + 298, + 482, + 387 + ], + "type": "text", + "content": "Fig. 3: Impact of incorporating normals on view-dependent radiance modeling. (Left) Normals at sampled points may be influenced by unrelated surfaces other than the intersection surface. (Top right) Smooth surfaces yield similarity in normals and reflection directions for the sampled points, while (Bottom right) surfaces with intricate local details, e.g., concavities, induce a scattered distribution of normals and reflection directions (Ref. dirs), particularly pronounced slightly distant from the vicinity of the zero-level set, which adversely affects geometry optimization. The details are explained in Sec. 3.2." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 323, + 118, + 441, + 285 + ], + "blocks": [ + { + "bbox": [ + 323, + 118, + 441, + 285 + ], + "lines": [ + { + "bbox": [ + 323, + 118, + 441, + 285 + ], + "spans": [ + { + "bbox": [ + 323, + 118, + 441, + 285 + ], + "type": "image", + "image_path": "c16c9a63a54b5ce76d95767850412eb63dfa57439162f10c8349a42fe9dade6e.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 451, + 118, + 478, + 287 + ], + "blocks": [ + { + "bbox": [ + 451, + 118, + 478, + 287 + ], + "lines": [ + { + "bbox": [ + 451, + 118, + 478, + 287 + ], + "spans": [ + { + "bbox": [ + 451, + 118, + 478, + 287 + ], + "type": "image", + "image_path": "f55cc30f7ed1382eecb8ca217173167ca976b89be02017cc7a7f3afab754bb3f.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 413, + 482, + 497 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 413, + 482, + 497 + ], + "spans": [ + { + "bbox": [ + 130, + 413, + 482, + 497 + ], + "type": "text", + "content": "do not necessarily relate to the normal of the intersecting surface (the bottom side of the zero level-set), but rather to the normals of surfaces elsewhere (the left side of the zero level-set). As a result, since the training loss of existing neural implicit surface reconstruction methods is usually ray-based, which is calculated using the local color of the intersecting surface, using this loss to optimize the normals of irrelevant surfaces elsewhere may introduce ambiguity in the optimization of geometry and result in incorrect geometric updates." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "text", + "content": "In addition, the second property of the reflection direction is more closely related to the phenomenon that using reflection directional parameterization tends to reconstruct overly smooth surfaces with missing concave or complex structures. Assuming that the geometry reconstruction begins with a smooth surface exhibiting local flatness (Fig. 3 top-right), which aligns with the geometric initialization practices commonly employed by existing neural implicit surface reconstruction methods [6,29], and the geometry of the target object includes concave structures (Fig. 3 bottom-right). We observed that for the same sampled ray, when the surface geometry is updated from a smooth surface to a concave surface, the normals and reflection directions of the same sampled points also change. This change is particularly pronounced for points that are slightly distant from the surface, and could be further pronounced due to the presence of Positional Encoding (PE). The noticeable changes in reflection directions at these points within each surface update introduce high-frequency variances to" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 284, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 284, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 284, + 91, + 447, + 102 + ], + "type": "text", + "content": "Rethinking directional parameterization" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 187 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 187 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 187 + ], + "type": "text", + "content": "the directional inputs of the radiance network. This makes learning a stable radiance field challenging. To be specific, assuming that the radiance network conditioned on reflection direction has already learned to minimize the color loss to some extent before the surface update, the substantial change in directional inputs due to the update could likely lead to an increase in color loss, which hinders the optimization and is susceptible to falling into local minima." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 189, + 482, + 272 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 189, + 482, + 272 + ], + "spans": [ + { + "bbox": [ + 130, + 189, + 482, + 272 + ], + "type": "text", + "content": "In contrast, we notice that using the viewing direction often leads to concave or overly intricate surface details (although some may be incorrect), as shown in Fig. 1 (a) and Fig. 2 (a) (c). We consider that this is because the viewing direction does not rely on the surface geometry, which means that changes in geometry do not affect the optimization of the radiance network conditioned on viewing direction. This allows the SDF can undergo more \"aggressive\" updates without considering the impact on radiance." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 289, + 346, + 300 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 289, + 346, + 300 + ], + "spans": [ + { + "bbox": [ + 131, + 289, + 346, + 300 + ], + "type": "text", + "content": "3.3 Hybrid Directional Parameterization" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 308, + 482, + 499 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 308, + 482, + 499 + ], + "spans": [ + { + "bbox": [ + 130, + 308, + 482, + 499 + ], + "type": "text", + "content": "Building upon the analysis from the previous section, we discovered that viewing direction and reflection direction seem to have a spatially complementary relationship: On one hand, for sampled points near the intersecting surface, the normals here accurately reflect the normals of the intersecting surface, thereby enabling the reflection direction to better model the interaction between light and surface, which in turn promotes the correct decomposition of geometry and appearance. On the other hand, for other sampled points along the ray, using the reflection direction might introduce irrelevant surface geometry into the optimization, or high-frequency variations in the directional inputs of the radiance network, both potentially hindering the optimization of the geometry and radiance networks. By contrast, using the viewing direction at these sampled points can avoid these issues. Motivated by these thoughts, we propose a spatial-aware directional representation, which transitions from reflection direction " + }, + { + "bbox": [ + 130, + 308, + 482, + 499 + ], + "type": "inline_equation", + "content": "\\mathbf{d}_{\\mathrm{ref}}" + }, + { + "bbox": [ + 130, + 308, + 482, + 499 + ], + "type": "text", + "content": " to viewing direction " + }, + { + "bbox": [ + 130, + 308, + 482, + 499 + ], + "type": "inline_equation", + "content": "\\mathbf{d}_{\\mathrm{view}}" + }, + { + "bbox": [ + 130, + 308, + 482, + 499 + ], + "type": "text", + "content": " based on the distance to the surface (i.e., the absolute value of SDF). We refer to this fused direction as hybrid direction " + }, + { + "bbox": [ + 130, + 308, + 482, + 499 + ], + "type": "inline_equation", + "content": "\\mathbf{d}_{\\mathrm{hyb}}" + }, + { + "bbox": [ + 130, + 308, + 482, + 499 + ], + "type": "text", + "content": ", formulated as:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 212, + 512, + 480, + 525 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 212, + 512, + 480, + 525 + ], + "spans": [ + { + "bbox": [ + 212, + 512, + 480, + 525 + ], + "type": "interline_equation", + "content": "\\mathbf {d} _ {\\mathrm {h y b}} = \\text {n o r m a l i z e} (\\alpha \\cdot \\mathbf {d} _ {\\text {r e f}} + (1 - \\alpha) \\cdot \\mathbf {d} _ {\\text {v i e w}}). \\tag {2}", + "image_path": "87900a76ba6900436b4ed8b3a1e6e043c57132bb70786acf9e34c81530b100c4.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 528, + 482, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 528, + 482, + 563 + ], + "spans": [ + { + "bbox": [ + 130, + 528, + 482, + 563 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 130, + 528, + 482, + 563 + ], + "type": "inline_equation", + "content": "\\alpha \\in [0,1]" + }, + { + "bbox": [ + 130, + 528, + 482, + 563 + ], + "type": "text", + "content": " represents the blend weight. We normalize the resulting linear blend to a unit vector thereby representing a valid direction. The blend weight " + }, + { + "bbox": [ + 130, + 528, + 482, + 563 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 130, + 528, + 482, + 563 + ], + "type": "text", + "content": " is implemented as" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 243, + 574, + 480, + 586 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 243, + 574, + 480, + 586 + ], + "spans": [ + { + "bbox": [ + 243, + 574, + 480, + 586 + ], + "type": "interline_equation", + "content": "\\alpha = \\exp (- \\gamma \\cdot \\det a c h (| f (\\mathbf {x}) |)), \\tag {3}", + "image_path": "6ef6985525d268f33fe2c65bd1e15bd71e9543b1b5951f6b017b0260ff02c881.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\gamma \\in \\mathbb{R}^{+}" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": " is an estimable parameter that controls the transition speed between reflection/viewing direction. The gradient of the SDF is detached to prevent interference with geometric updates. At the surface of the object, i.e., when SDF equals 0, " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": " simply becomes 1, which means that the hybrid direction aligns with the reflection direction. As " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": " moves further away from the surface, " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": " gradually approaches 0, indicating that the direction tends towards the viewing" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 246, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 246, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 246, + 102 + ], + "type": "text", + "content": "Jiang and Xu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": "direction. Here, " + }, + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": " can be any bell-shaped function w.r.t. the SDF value. In this case, we utilize the PDF of the Laplace distribution for simplicity." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 140, + 480, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 140, + 480, + 175 + ], + "spans": [ + { + "bbox": [ + 130, + 140, + 480, + 175 + ], + "type": "text", + "content": "In the next section, extensive experimental results validate the superiority of the proposed hybrid directional parameterization compared to existing parameterizations." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 193, + 230, + 205 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 193, + 230, + 205 + ], + "spans": [ + { + "bbox": [ + 131, + 193, + 230, + 205 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 217, + 272, + 228 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 217, + 272, + 228 + ], + "spans": [ + { + "bbox": [ + 131, + 217, + 272, + 228 + ], + "type": "text", + "content": "4.1 Experimental Settings" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 235, + 480, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 235, + 480, + 376 + ], + "spans": [ + { + "bbox": [ + 130, + 235, + 480, + 376 + ], + "type": "text", + "content": "Datasets. We conduct experiments on objects from different datasets, including the DTU dataset [1], the Shiny Blender dataset [28], and the real captured dataset [9]. The DTU dataset [1] contains indoor multi-view captures of diverse objects, including challenging conditions for geometry reconstruction such as non-Lambertian surfaces and complex structures. We use the same 15 objects as previous works [2, 29, 31, 36] for evaluation. In addition, we utilize the foreground masks provided by Yariv et al. [36]. The Shinny Blender dataset [28] is a synthetic dataset rendered in Blender [4], which contains different objects with specular surfaces. We follow the experimental settings used in Ref-NeuS [8] on this dataset, which selects four objects. The real captured dataset [9] contains a set of three outdoor captures of real objects with complex geometry and specular surfaces. We select the \"toycar\" scene for evaluation." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 379, + 480, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 379, + 480, + 568 + ], + "spans": [ + { + "bbox": [ + 130, + 379, + 480, + 568 + ], + "type": "text", + "content": "Implementation details. We integrate the proposed hybrid directional parameterization into two types of neural surface reconstruction methods: NeuS [29] and NeuDA [2]. The former utilizes a deep MLP network as the scene representation, while the latter leverages a shallow MLP decoder combined with multi-level voxel grids. Both methods adopt a radiance network conditioned on viewing direction in their original versions. The modification of the directional input from the viewing direction to the proposed hybrid direction can be accomplished with less than 10 lines of code. We additionally implement the radiance network that depends on reflection direction for both methods, to facilitate a comprehensive comparison of different directional parameterizations. Regarding the learnable parameter " + }, + { + "bbox": [ + 130, + 379, + 480, + 568 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 130, + 379, + 480, + 568 + ], + "type": "text", + "content": " in Eqn. 3, we implement it as " + }, + { + "bbox": [ + 130, + 379, + 480, + 568 + ], + "type": "inline_equation", + "content": "\\gamma = \\exp (10\\cdot \\gamma_b)" + }, + { + "bbox": [ + 130, + 379, + 480, + 568 + ], + "type": "text", + "content": ". We assign an initial value of 0.3 to " + }, + { + "bbox": [ + 130, + 379, + 480, + 568 + ], + "type": "inline_equation", + "content": "\\gamma_{b}" + }, + { + "bbox": [ + 130, + 379, + 480, + 568 + ], + "type": "text", + "content": " except for the \"toaster\" scene from the Shiny Blender dataset. For the \"toaster\" scene, an initial value of 0.1 is assigned since it results in better reconstruction. The training losses of the modified methods are identical to those of their original versions. The number of training iterations is set to 300,000 to ensure convergence." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 570, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 481, + 665 + ], + "type": "text", + "content": "Evaluation metrics. For the DTU dataset, we follow the evaluation protocol of previous works [2, 29] and evaluate the Chamfer Distance (CD) between the estimated and ground-truth geometry. For the Shiny Blender dataset, we follow the evaluation protocol proposed in Ref-NeuS [8]. It evaluates the accuracy (Acc.) instead of the CD for unbiased evaluation, and also evaluates the mean angular error (MAE) between the ground-truth normal maps and the rendered normal maps. For the real captured dataset [9], we only provide a qualitative evaluation, as ground-truth geometry is not provided in this dataset." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 284, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 284, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 284, + 91, + 447, + 102 + ], + "type": "text", + "content": "Rethinking directional parameterization" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 132, + 168, + 480, + 237 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 482, + 158 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 482, + 158 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 482, + 158 + ], + "type": "text", + "content": "Table 1: Quantitative evaluation in terms of Chamfer distance of integrating different directional parameterizations in NeuS [29] and NeuDA [2] on the DTU dataset [1]. The experimental objects are categorized into diffuse objects and specular objects. The numbers in the headers denote the scene IDs. Bold results have the best score." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 132, + 168, + 480, + 237 + ], + "lines": [ + { + "bbox": [ + 132, + 168, + 480, + 237 + ], + "spans": [ + { + "bbox": [ + 132, + 168, + 480, + 237 + ], + "type": "table", + "html": "
MethodsScan of Diffuse ObjectsScan of Specular ObjectsTotal Avg.
243740556583105106114118122Avg.636997110Avg.
NeuS [29]0.830.980.560.370.591.450.780.520.360.450.450.671.130.600.951.431.030.77
NeuS, w/ reflection dir.1.394.390.730.391.341.490.920.590.360.460.541.141.110.631.460.921.031.11
NeuS, w/ hybrid dir. (Ours)0.810.970.540.380.611.470.770.510.350.450.470.671.100.600.850.900.860.72
NeuD A [2]0.510.760.390.370.561.370.790.500.340.420.460.591.080.571.130.800.900.67
NeuD A, w/ reflection dir.0.581.030.630.370.791.450.890.510.370.430.500.691.030.550.900.760.810.72
NeuD A, w/ hybrid dir. (Ours)0.530.770.400.360.571.380.760.470.330.410.470.591.020.540.760.750.770.63
", + "image_path": "1108d15f71b95c0b65361450d355052ac603fb0391552f7e8c23c9a1d7300e72.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 259, + 436, + 272 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 259, + 436, + 272 + ], + "spans": [ + { + "bbox": [ + 132, + 259, + 436, + 272 + ], + "type": "text", + "content": "4.2 Comparison of Different Directional Parameterizations" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 282, + 482, + 558 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 282, + 482, + 558 + ], + "spans": [ + { + "bbox": [ + 130, + 282, + 482, + 558 + ], + "type": "text", + "content": "We first evaluate the performance of integrating different directional parameterizations into the two baseline methods NeuS [29] and NeuDA [2] on the DTU dataset [1]. The experimental objects from the DTU dataset are categorized into diffuse objects and specular objects for better demonstration. Tab. 1 summarizes the quantitative results. For both baseline methods, we notice that substituting the original viewing directional parameterization with the reflection direction only yields better results for a minority of specular objects. For diffuse objects and the remaining specular objects, employing reflection directional parameterization demonstrates significant degradation in performance. Considering the relatively complex geometry of DTU objects, we attribute this to the optimization issue analyzed in Sec. 3.2. Conversely, the incorporation of the proposed hybrid directional parameterization enables both modified methods to yield consistently superior performance for specular objects while achieving results on par with the original version for diffuse objects. The first two rows in Fig. 4 showcase the visual comparison of adopting different directional parameterizations for the reconstruction of two specular objects from the DTU dataset. Both baseline methods fail to reconstruct the specular surfaces of objects. On the other hand, although integrating the reflection direction yields better reconstruction for specular objects with simple geometry (e.g., Scan 63), they fail to reconstruct the specular surfaces at the concave structure in Scan 97. In contrast, integrating the proposed hybrid direction achieves the best reconstruction results in both scenarios. More qualitative comparisons on the DTU dataset are available in our supplementary." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 558, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 558, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 558, + 482, + 665 + ], + "type": "text", + "content": "We further validate the advantage of the proposed hybrid directional parameterization on the Shiny Blender dataset [28], as demonstrated in Tab. 2. Both baseline methods exhibit large reconstruction errors on the surfaces of highly specular objects in this dataset. By contrast, due to the relatively simple geometry of objects, replacing the viewing directional parameterization with the reflection direction results in a significant improvement in reconstruction quality on most objects in this dataset, except for the \"coffee\" object. For the reconstruction of the \"coffee\" object, adopting the reflection directional parameterization fails to recover the concave structure formed by the inner wall of the coffee cup" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 246, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 246, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 246, + 102 + ], + "type": "text", + "content": "Jiang and Xu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 132, + 157, + 480, + 240 + ], + "blocks": [ + { + "bbox": [ + 131, + 114, + 482, + 148 + ], + "lines": [ + { + "bbox": [ + 131, + 114, + 482, + 148 + ], + "spans": [ + { + "bbox": [ + 131, + 114, + 482, + 148 + ], + "type": "text", + "content": "Table 2: Quantitative evaluation of integrating different directional parameterizations in NeuS [29] and NeuDA [2] on the Shiny Blender dataset [28]. Bold results have the best score." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 132, + 157, + 480, + 240 + ], + "lines": [ + { + "bbox": [ + 132, + 157, + 480, + 240 + ], + "spans": [ + { + "bbox": [ + 132, + 157, + 480, + 240 + ], + "type": "table", + "html": "
MethodshelmettoastercoffeecarAvg.
Acc. ↓MAE ↓Acc. ↓MAE ↓Acc. ↓MAE ↓Acc. ↓MAE ↓Acc. ↓MAE ↓
NeuS [29]4.883.203.312.851.971.060.860.952.762.02
NeuS, w/ reflection dir.0.310.360.390.943.301.370.480.731.120.85
NeuS, w/ hybrid dir. (Ours)0.290.340.450.920.640.550.430.720.450.63
NeuDA [2]5.733.775.983.391.160.810.871.033.442.25
NeuDA, w/ reflection dir.0.340.371.031.713.661.530.620.871.411.12
NeuDA, w/ hybrid dir. (Ours)0.320.360.951.540.550.570.600.840.610.83
", + "image_path": "329faf4da4cbd1d40208b8888610f70f109bdd77708718a0927605640dcb04a2.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 134, + 248, + 478, + 417 + ], + "blocks": [ + { + "bbox": [ + 134, + 248, + 478, + 417 + ], + "lines": [ + { + "bbox": [ + 134, + 248, + 478, + 417 + ], + "spans": [ + { + "bbox": [ + 134, + 248, + 478, + 417 + ], + "type": "image", + "image_path": "131a8fe431a68b9861d9cfbec23962ca947dcc1212996f54c71bb1d6babac13d.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 425, + 482, + 449 + ], + "lines": [ + { + "bbox": [ + 131, + 425, + 482, + 449 + ], + "spans": [ + { + "bbox": [ + 131, + 425, + 482, + 449 + ], + "type": "text", + "content": "Fig. 4: Qualitative comparison of integrating different directional parameterizations in NeuS [29] and NeuDA [2] on the DTU dataset [1] and Shiny Blender dataset [28]." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 472, + 482, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 472, + 482, + 533 + ], + "spans": [ + { + "bbox": [ + 130, + 472, + 482, + 533 + ], + "type": "text", + "content": "and the surface of the water, as shown in the third row in Fig. 4. On the other hand, the integration of our proposed hybrid directional parameterization enables the successful reconstruction of both the specular surfaces and concave structures, obtaining consistently superior reconstructions of all objects in this dataset, compared to the other two directional parameterizations." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 533, + 482, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 482, + 628 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 482, + 628 + ], + "type": "text", + "content": "Lastly, we qualitatively present the reconstruction results on the \"toycar\" scene from the real captured scene in Fig. 5, which contains challenging reflective glass and complex geometry. Both modified methods employing the proposed hybrid directional parameterization succeed in recovering the complex geometry and the smooth surface of the reflective glass. In contrast, using either the reflection direction or the original viewing direction results in reconstructions that are either overly smooth, missing concave surface structures, or provide incorrect surface reconstructions at the reflective glass areas." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "text", + "content": "As the summary of this section, compared to the viewing directional parameterization which struggles with reconstructing specular surfaces, and the reflection directional parameterization which faces difficulties in reconstructing" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 284, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 284, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 284, + 91, + 447, + 102 + ], + "type": "text", + "content": "Rethinking directional parameterization" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 136, + 147, + 214, + 204 + ], + "blocks": [ + { + "bbox": [ + 136, + 147, + 214, + 204 + ], + "lines": [ + { + "bbox": [ + 136, + 147, + 214, + 204 + ], + "spans": [ + { + "bbox": [ + 136, + 147, + 214, + 204 + ], + "type": "image", + "image_path": "f1ddbdcf03a7995ee9e1642aeef2d91281a9b0a40740e67696defd744c005857.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 148, + 205, + 201, + 213 + ], + "lines": [ + { + "bbox": [ + 148, + 205, + 201, + 213 + ], + "spans": [ + { + "bbox": [ + 148, + 205, + 201, + 213 + ], + "type": "text", + "content": "Reference RGB" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 219, + 117, + 297, + 171 + ], + "blocks": [ + { + "bbox": [ + 219, + 117, + 297, + 171 + ], + "lines": [ + { + "bbox": [ + 219, + 117, + 297, + 171 + ], + "spans": [ + { + "bbox": [ + 219, + 117, + 297, + 171 + ], + "type": "image", + "image_path": "582c6f6b5e5bc4dc73ab6d6a7068e1ff9cbf15e31e60d7671bb241236f142f2f.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 248, + 171, + 268, + 179 + ], + "lines": [ + { + "bbox": [ + 248, + 171, + 268, + 179 + ], + "spans": [ + { + "bbox": [ + 248, + 171, + 268, + 179 + ], + "type": "text", + "content": "NeuS" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 305, + 118, + 383, + 171 + ], + "blocks": [ + { + "bbox": [ + 305, + 118, + 383, + 171 + ], + "lines": [ + { + "bbox": [ + 305, + 118, + 383, + 171 + ], + "spans": [ + { + "bbox": [ + 305, + 118, + 383, + 171 + ], + "type": "image", + "image_path": "f65ffc2a1635eb666adf303450035e9005d10ed0167b153c7afc7a996e1bf67e.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 171, + 383, + 180 + ], + "lines": [ + { + "bbox": [ + 305, + 171, + 383, + 180 + ], + "spans": [ + { + "bbox": [ + 305, + 171, + 383, + 180 + ], + "type": "text", + "content": "NeuS, w/ reflection dir." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 392, + 121, + 470, + 171 + ], + "blocks": [ + { + "bbox": [ + 392, + 121, + 470, + 171 + ], + "lines": [ + { + "bbox": [ + 392, + 121, + 470, + 171 + ], + "spans": [ + { + "bbox": [ + 392, + 121, + 470, + 171 + ], + "type": "image", + "image_path": "36b7a36abd21e147b2ba39440f62dcf822f3d3bc998ce6ffe72a4e3588a1d0d5.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 398, + 171, + 470, + 180 + ], + "lines": [ + { + "bbox": [ + 398, + 171, + 470, + 180 + ], + "spans": [ + { + "bbox": [ + 398, + 171, + 470, + 180 + ], + "type": "text", + "content": "NeuS, w/ hybrid dir." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 219, + 184, + 297, + 236 + ], + "blocks": [ + { + "bbox": [ + 219, + 184, + 297, + 236 + ], + "lines": [ + { + "bbox": [ + 219, + 184, + 297, + 236 + ], + "spans": [ + { + "bbox": [ + 219, + 184, + 297, + 236 + ], + "type": "image", + "image_path": "647d7e95eb8b2b34ddcdd9114d4a5ce8a9343e5e8b5a066952e0841bd9e7313b.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 244, + 237, + 271, + 245 + ], + "lines": [ + { + "bbox": [ + 244, + 237, + 271, + 245 + ], + "spans": [ + { + "bbox": [ + 244, + 237, + 271, + 245 + ], + "type": "text", + "content": "NeuDA" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 305, + 185, + 383, + 236 + ], + "blocks": [ + { + "bbox": [ + 305, + 185, + 383, + 236 + ], + "lines": [ + { + "bbox": [ + 305, + 185, + 383, + 236 + ], + "spans": [ + { + "bbox": [ + 305, + 185, + 383, + 236 + ], + "type": "image", + "image_path": "9c12491570133dd22f48aceeb65c46207a02faa25a6eaedfb1e31a92cbbf0ab3.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 237, + 388, + 247 + ], + "lines": [ + { + "bbox": [ + 304, + 237, + 388, + 247 + ], + "spans": [ + { + "bbox": [ + 304, + 237, + 388, + 247 + ], + "type": "text", + "content": "NeuDA,w/ reflection dir." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 392, + 185, + 470, + 236 + ], + "blocks": [ + { + "bbox": [ + 392, + 185, + 470, + 236 + ], + "lines": [ + { + "bbox": [ + 392, + 185, + 470, + 236 + ], + "spans": [ + { + "bbox": [ + 392, + 185, + 470, + 236 + ], + "type": "image", + "image_path": "62708bc77f6e852906135c5fe5ca35789b3e6ad45b2ca7af5ca2303158c6605a.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 398, + 237, + 477, + 247 + ], + "lines": [ + { + "bbox": [ + 398, + 237, + 477, + 247 + ], + "spans": [ + { + "bbox": [ + 398, + 237, + 477, + 247 + ], + "type": "text", + "content": "NeuDA, w/ hybrid dir." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 131, + 255, + 479, + 278 + ], + "lines": [ + { + "bbox": [ + 131, + 255, + 479, + 278 + ], + "spans": [ + { + "bbox": [ + 131, + 255, + 479, + 278 + ], + "type": "text", + "content": "Fig. 5: Qualitative comparison of integrating different directional parameterizations in NeuS [29] and NeuDA [2] on the \"toycar\" scene from the real captured dataset [9]." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "table", + "bbox": [ + 133, + 323, + 479, + 398 + ], + "blocks": [ + { + "bbox": [ + 131, + 290, + 479, + 312 + ], + "lines": [ + { + "bbox": [ + 131, + 290, + 479, + 312 + ], + "spans": [ + { + "bbox": [ + 131, + 290, + 479, + 312 + ], + "type": "text", + "content": "Table 3: Quantitative evaluation in terms of Chamfer distance on the DTU dataset [1]. The numbers in the header denote the scene IDs. Bold results have the best score." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 133, + 323, + 479, + 398 + ], + "lines": [ + { + "bbox": [ + 133, + 323, + 479, + 398 + ], + "spans": [ + { + "bbox": [ + 133, + 323, + 479, + 398 + ], + "type": "table", + "html": "
Methods243740556365698397105106110114118122Avg.
IDR [36]1.631.870.630.481.040.790.771.331.160.760.670.900.420.510.530.90
NeuS [29]0.830.980.560.371.130.590.601.450.950.780.521.430.360.450.450.77
Voxurf [31]0.650.740.390.350.960.640.851.581.010.680.601.110.370.450.470.72
NeuDA [2]0.510.760.390.371.080.560.571.371.130.790.500.800.340.420.460.67
Ours (NeuS)0.810.970.540.381.100.610.601.470.850.770.510.900.350.450.470.72
Ours (NeuDA)0.530.770.400.361.020.570.541.380.760.760.470.760.330.410.470.63
", + "image_path": "d938a2f05cc116196255e7a42fe61f2faa01e6cb5726e0b8003cd95141b2831f.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "table_body" + } + ], + "index": 18 + }, + { + "bbox": [ + 130, + 420, + 479, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 420, + 479, + 468 + ], + "spans": [ + { + "bbox": [ + 130, + 420, + 479, + 468 + ], + "type": "text", + "content": "complex structures, our proposed hybrid directional parameterization effectively handles both scenarios in a unified form. This benefit is attributed to our hybrid directional parameterization effectively combining the advantages of both viewing and reflection directional parameterizations while avoiding their drawbacks." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 131, + 487, + 320, + 499 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 487, + 320, + 499 + ], + "spans": [ + { + "bbox": [ + 131, + 487, + 320, + 499 + ], + "type": "text", + "content": "4.3 Comparison to State-of-the-Art" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 130, + 510, + 481, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 481, + 640 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 481, + 640 + ], + "type": "text", + "content": "Although our main purpose is not to develop a best-performing surface reconstruction system, we demonstrate that we can still achieve state-of-the-art performance by integrating the proposed hybrid directional parameterization into existing methods. For brevity, we refer to NeuS [29] and NeuDA [2] incorporating the proposed hybrid directional parameterization as Ours (NeuS) and Ours (NeuDA) in the following text. Tab. 3 reports the Chamfer distance (CD) metrics of different methods on the DTU dataset [1]. For a fair comparison, all methods in Tab. 3 utilize foreground masks provided by Yariv et al. [36] during training. Both Ours (NeuS) and Ours (NeuDA) improve their baselines by a certain margin in the average CD scores, and Ours (NeuDA) achieves the best performance compared to previous SOTA methods in the average CD metric." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 131, + 641, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 641, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 131, + 641, + 481, + 665 + ], + "type": "text", + "content": "Tab. 4 reports the quantitative evaluation of different methods on the Shiny Blender dataset [28]. The results of [5,6,8,22,28,35] are taken from the results" + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 246, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 246, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 246, + 101 + ], + "type": "text", + "content": "Jiang and Xu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 167, + 146, + 445, + 258 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 480, + 137 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 480, + 137 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 480, + 137 + ], + "type": "text", + "content": "Table 4: Quantitative evaluation on the Shiny Blender dataset [28]. “-” denotes that the results are not available. Bold results have the best score." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 167, + 146, + 445, + 258 + ], + "lines": [ + { + "bbox": [ + 167, + 146, + 445, + 258 + ], + "spans": [ + { + "bbox": [ + 167, + 146, + 445, + 258 + ], + "type": "table", + "html": "
MethodshelmettoastercoffeecarAvg.
Acc. ↓MAE ↓Acc. ↓MAE ↓Acc. ↓MAE ↓Acc. ↓MAE ↓Acc. ↓MAE ↓
Ref-NeRF [28]-29.48-42.87-12.24-14.93-24.88
NeuS [29]4.883.203.312.851.971.060.860.952.762.02
UNISURF [22]1.691.783.755.512.883.152.341.982.673.11
VolSDF [35]1.551.372.022.532.232.280.581.041.601.81
NeuralWarp [5]2.251.945.903.511.542.040.651.072.592.14
Geo-NeuS [6]0.742.366.353.763.856.362.885.673.464.54
NeuD A [2]5.733.775.983.391.160.810.871.033.442.25
Ref-NeuD [8]0.290.380.421.470.770.990.370.800.460.91
Ours (NeuD)0.290.340.450.920.640.550.430.720.450.63
Ours (NeuD A)0.320.360.951.540.550.570.600.840.610.83
", + "image_path": "1af0e3dc63bbc16942d2780aaf688bc5559f798bdb69d17e7646ddd6ee8f97bf.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 133, + 269, + 179, + 312 + ], + "blocks": [ + { + "bbox": [ + 133, + 269, + 179, + 312 + ], + "lines": [ + { + "bbox": [ + 133, + 269, + 179, + 312 + ], + "spans": [ + { + "bbox": [ + 133, + 269, + 179, + 312 + ], + "type": "image", + "image_path": "5345e63abcb455c37bf47dc647cd2c32a47dc52cfa5b3d647f27a5bc83c438c9.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 138, + 313, + 175, + 320 + ], + "lines": [ + { + "bbox": [ + 138, + 313, + 175, + 320 + ], + "spans": [ + { + "bbox": [ + 138, + 313, + 175, + 320 + ], + "type": "text", + "content": "Reference RGB" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 181, + 269, + 240, + 312 + ], + "blocks": [ + { + "bbox": [ + 181, + 269, + 240, + 312 + ], + "lines": [ + { + "bbox": [ + 181, + 269, + 240, + 312 + ], + "spans": [ + { + "bbox": [ + 181, + 269, + 240, + 312 + ], + "type": "image", + "image_path": "7b63b837f0ae60ce15848f50b4d247b11f3b67130e40ea30028d52c3ed0b270d.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 189, + 313, + 239, + 320 + ], + "lines": [ + { + "bbox": [ + 189, + 313, + 239, + 320 + ], + "spans": [ + { + "bbox": [ + 189, + 313, + 239, + 320 + ], + "type": "inline_equation", + "content": "\\gamma_{b} = 0.0" + }, + { + "bbox": [ + 189, + 313, + 239, + 320 + ], + "type": "inline_equation", + "content": "\\mathrm{CD}\\downarrow" + }, + { + "bbox": [ + 189, + 313, + 239, + 320 + ], + "type": "text", + "content": " : 0.91" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 242, + 269, + 299, + 312 + ], + "blocks": [ + { + "bbox": [ + 242, + 269, + 299, + 312 + ], + "lines": [ + { + "bbox": [ + 242, + 269, + 299, + 312 + ], + "spans": [ + { + "bbox": [ + 242, + 269, + 299, + 312 + ], + "type": "image", + "image_path": "b86a875fb70518f84ec19ada47e3191ac3f37279be34cf98e4fed23a0161d205.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 329, + 480, + 351 + ], + "lines": [ + { + "bbox": [ + 132, + 329, + 480, + 351 + ], + "spans": [ + { + "bbox": [ + 132, + 329, + 480, + 351 + ], + "type": "text", + "content": "Fig. 6: Reconstruction results using different initial values of " + }, + { + "bbox": [ + 132, + 329, + 480, + 351 + ], + "type": "inline_equation", + "content": "\\gamma_{b}" + }, + { + "bbox": [ + 132, + 329, + 480, + 351 + ], + "type": "text", + "content": ". The main differences are highlighted in red boxes." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 301, + 269, + 359, + 312 + ], + "blocks": [ + { + "bbox": [ + 250, + 313, + 299, + 320 + ], + "lines": [ + { + "bbox": [ + 250, + 313, + 299, + 320 + ], + "spans": [ + { + "bbox": [ + 250, + 313, + 299, + 320 + ], + "type": "inline_equation", + "content": "\\gamma_{b} = 0.1" + }, + { + "bbox": [ + 250, + 313, + 299, + 320 + ], + "type": "inline_equation", + "content": "\\mathrm{CD}\\downarrow" + }, + { + "bbox": [ + 250, + 313, + 299, + 320 + ], + "type": "text", + "content": " : 0.81" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 301, + 269, + 359, + 312 + ], + "lines": [ + { + "bbox": [ + 301, + 269, + 359, + 312 + ], + "spans": [ + { + "bbox": [ + 301, + 269, + 359, + 312 + ], + "type": "image", + "image_path": "98e8e6462d067ee35fef37e76fa72e6daeb2e2b6dd599acd77ca342c3e3b8c5c.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 361, + 269, + 420, + 312 + ], + "blocks": [ + { + "bbox": [ + 308, + 313, + 359, + 320 + ], + "lines": [ + { + "bbox": [ + 308, + 313, + 359, + 320 + ], + "spans": [ + { + "bbox": [ + 308, + 313, + 359, + 320 + ], + "type": "inline_equation", + "content": "\\gamma_{b} = 0.2" + }, + { + "bbox": [ + 308, + 313, + 359, + 320 + ], + "type": "inline_equation", + "content": "\\mathrm{CD}\\downarrow" + }, + { + "bbox": [ + 308, + 313, + 359, + 320 + ], + "type": "text", + "content": " :0.80" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 361, + 269, + 420, + 312 + ], + "lines": [ + { + "bbox": [ + 361, + 269, + 420, + 312 + ], + "spans": [ + { + "bbox": [ + 361, + 269, + 420, + 312 + ], + "type": "image", + "image_path": "991b02be9d23b8e105ad4d35305552bc4109874e2c23fb513bfd34ec7ca08da8.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 369, + 313, + 419, + 320 + ], + "lines": [ + { + "bbox": [ + 369, + 313, + 419, + 320 + ], + "spans": [ + { + "bbox": [ + 369, + 313, + 419, + 320 + ], + "type": "inline_equation", + "content": "\\gamma_{b} = 0.3" + }, + { + "bbox": [ + 369, + 313, + 419, + 320 + ], + "type": "inline_equation", + "content": "\\mathrm{CD}\\downarrow" + }, + { + "bbox": [ + 369, + 313, + 419, + 320 + ], + "type": "text", + "content": " :0.76" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 422, + 269, + 480, + 312 + ], + "blocks": [ + { + "bbox": [ + 422, + 269, + 480, + 312 + ], + "lines": [ + { + "bbox": [ + 422, + 269, + 480, + 312 + ], + "spans": [ + { + "bbox": [ + 422, + 269, + 480, + 312 + ], + "type": "image", + "image_path": "815bd1c01f24cef8d31899ddd76bb93ec6cc74473784b89d674824c11024653d.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 430, + 313, + 480, + 320 + ], + "lines": [ + { + "bbox": [ + 430, + 313, + 480, + 320 + ], + "spans": [ + { + "bbox": [ + 430, + 313, + 480, + 320 + ], + "type": "inline_equation", + "content": "\\gamma_{b} = 0.4" + }, + { + "bbox": [ + 430, + 313, + 480, + 320 + ], + "type": "inline_equation", + "content": "\\mathrm{CD}\\downarrow" + }, + { + "bbox": [ + 430, + 313, + 480, + 320 + ], + "type": "text", + "content": " :0.77" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 375, + 482, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 375, + 482, + 437 + ], + "spans": [ + { + "bbox": [ + 130, + 375, + 482, + 437 + ], + "type": "text", + "content": "reported in Ge et al. [8]. Both Ours (NeuS) and Ours (NeuDA) achieve better average MAE metrics than the previous SOTA method Ref-NeuS [8], and Ours (NeuS) achieve the best overall performance in terms of both Acc. and MAE metrics on this dataset. We refer to our supplementary for the qualitative comparisons between Ours (NeuS) and Ref-NeuS [8]." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 453, + 245, + 464 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 453, + 245, + 464 + ], + "spans": [ + { + "bbox": [ + 132, + 453, + 245, + 464 + ], + "type": "text", + "content": "4.4 Ablation Studies" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 130, + 474, + 480, + 497 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 474, + 480, + 497 + ], + "spans": [ + { + "bbox": [ + 130, + 474, + 480, + 497 + ], + "type": "text", + "content": "In this section, we conduct ablation studies on several key design choices of the proposed hybrid directional parameterization by using Ours (NeuDA)." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 130, + 498, + 481, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 498, + 481, + 629 + ], + "spans": [ + { + "bbox": [ + 130, + 498, + 481, + 629 + ], + "type": "text", + "content": "The initial value of " + }, + { + "bbox": [ + 130, + 498, + 481, + 629 + ], + "type": "inline_equation", + "content": "\\gamma_{b}" + }, + { + "bbox": [ + 130, + 498, + 481, + 629 + ], + "type": "text", + "content": ". We first validate the impact of setting different initial values for " + }, + { + "bbox": [ + 130, + 498, + 481, + 629 + ], + "type": "inline_equation", + "content": "\\gamma_{b}" + }, + { + "bbox": [ + 130, + 498, + 481, + 629 + ], + "type": "text", + "content": " on the reconstruction results. Fig. 6 compares the obtained reconstruction results using five different initial values for " + }, + { + "bbox": [ + 130, + 498, + 481, + 629 + ], + "type": "inline_equation", + "content": "\\gamma_{b}" + }, + { + "bbox": [ + 130, + 498, + 481, + 629 + ], + "type": "text", + "content": ". We find that when the initial value of " + }, + { + "bbox": [ + 130, + 498, + 481, + 629 + ], + "type": "inline_equation", + "content": "\\gamma_{b}" + }, + { + "bbox": [ + 130, + 498, + 481, + 629 + ], + "type": "text", + "content": " is set low, our method tends to recover incorrect geometries similar to those caused by using the reflection directional parameterization. This phenomenon remains consistent with our analysis in Sec 3.2, as a smaller initial value for " + }, + { + "bbox": [ + 130, + 498, + 481, + 629 + ], + "type": "inline_equation", + "content": "\\gamma_{b}" + }, + { + "bbox": [ + 130, + 498, + 481, + 629 + ], + "type": "text", + "content": " increases the likelihood of involving irrelevant surfaces and high-frequency variations in the directional inputs of the radiance network during the optimization. We empirically find the best initial value of " + }, + { + "bbox": [ + 130, + 498, + 481, + 629 + ], + "type": "inline_equation", + "content": "\\gamma_{b}" + }, + { + "bbox": [ + 130, + 498, + 481, + 629 + ], + "type": "text", + "content": " to be 0.3 for most experimental objects. More discussions regarding the parameter " + }, + { + "bbox": [ + 130, + 498, + 481, + 629 + ], + "type": "inline_equation", + "content": "\\gamma_{b}" + }, + { + "bbox": [ + 130, + 498, + 481, + 629 + ], + "type": "text", + "content": " and blend weight " + }, + { + "bbox": [ + 130, + 498, + 481, + 629 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 130, + 498, + 481, + 629 + ], + "type": "text", + "content": " can be found in our supplementary material." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 130, + 630, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 630, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 630, + 481, + 666 + ], + "type": "text", + "content": "Detach SDF gradient in Eqn. 3 or not? Fig. 7 compares the reconstruction results with and without detaching the gradient of the estimated SDF in Eqn. 3. We observe that our method struggles to reconstruct the fine-grained geometry" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 284, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 284, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 284, + 91, + 447, + 102 + ], + "type": "text", + "content": "Rethinking directional parameterization" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 114, + 199, + 164 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 199, + 164 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 199, + 164 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 199, + 164 + ], + "type": "image", + "image_path": "37d1e782e6978f39ae6919287d644778fffde258e7ada473f5099d1f3ee9e413.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 145, + 164, + 187, + 171 + ], + "lines": [ + { + "bbox": [ + 145, + 164, + 187, + 171 + ], + "spans": [ + { + "bbox": [ + 145, + 164, + 187, + 171 + ], + "type": "text", + "content": "Reference RGB" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 200, + 114, + 251, + 163 + ], + "blocks": [ + { + "bbox": [ + 200, + 114, + 251, + 163 + ], + "lines": [ + { + "bbox": [ + 200, + 114, + 251, + 163 + ], + "spans": [ + { + "bbox": [ + 200, + 114, + 251, + 163 + ], + "type": "image", + "image_path": "f1e1b747b70a6657ce267fb061d7a820fb68314fbff530d15311cb55ebbc9f66.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 202, + 164, + 249, + 178 + ], + "lines": [ + { + "bbox": [ + 202, + 164, + 249, + 178 + ], + "spans": [ + { + "bbox": [ + 202, + 164, + 249, + 178 + ], + "type": "text", + "content": "w/o SDF gradient detachment" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 253, + 115, + 304, + 164 + ], + "blocks": [ + { + "bbox": [ + 253, + 115, + 304, + 164 + ], + "lines": [ + { + "bbox": [ + 253, + 115, + 304, + 164 + ], + "spans": [ + { + "bbox": [ + 253, + 115, + 304, + 164 + ], + "type": "image", + "image_path": "841dfae42ca3709046551a7fd75a51efdb527306de2daec685287e3580f7cc6a.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 253, + 164, + 304, + 178 + ], + "lines": [ + { + "bbox": [ + 253, + 164, + 304, + 178 + ], + "spans": [ + { + "bbox": [ + 253, + 164, + 304, + 178 + ], + "type": "text", + "content": "w/SDF gradient detachment (Ours)" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 318, + 121, + 367, + 170 + ], + "blocks": [ + { + "bbox": [ + 318, + 121, + 367, + 170 + ], + "lines": [ + { + "bbox": [ + 318, + 121, + 367, + 170 + ], + "spans": [ + { + "bbox": [ + 318, + 121, + 367, + 170 + ], + "type": "image", + "image_path": "8ec8eccae10f1f88e6ddac6b1efbabea238d4e8c59eaa6d5848f20711b4a2985.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 319, + 172, + 358, + 179 + ], + "lines": [ + { + "bbox": [ + 319, + 172, + 358, + 179 + ], + "spans": [ + { + "bbox": [ + 319, + 172, + 358, + 179 + ], + "type": "text", + "content": "Reference RGB" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 131, + 188, + 304, + 221 + ], + "lines": [ + { + "bbox": [ + 131, + 188, + 304, + 221 + ], + "spans": [ + { + "bbox": [ + 131, + 188, + 304, + 221 + ], + "type": "text", + "content": "Fig.7: Reconstruction results with and without detaching the gradient of the estimated SDF in Eqn. 3." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 375, + 121, + 424, + 171 + ], + "blocks": [ + { + "bbox": [ + 375, + 121, + 424, + 171 + ], + "lines": [ + { + "bbox": [ + 375, + 121, + 424, + 171 + ], + "spans": [ + { + "bbox": [ + 375, + 121, + 424, + 171 + ], + "type": "image", + "image_path": "f25fc9a33d52989287e48f6ee97beba950a6a8a20200368881dc50c2a4f9422b.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 379, + 172, + 416, + 179 + ], + "lines": [ + { + "bbox": [ + 379, + 172, + 416, + 179 + ], + "spans": [ + { + "bbox": [ + 379, + 172, + 416, + 179 + ], + "type": "text", + "content": "post-PE fusion" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 307, + 189, + 481, + 222 + ], + "lines": [ + { + "bbox": [ + 307, + 189, + 481, + 222 + ], + "spans": [ + { + "bbox": [ + 307, + 189, + 481, + 222 + ], + "type": "text", + "content": "Fig. 8: Reconstruction results fusing directions before and after positional encoding (PE)." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 430, + 121, + 481, + 171 + ], + "blocks": [ + { + "bbox": [ + 430, + 121, + 481, + 171 + ], + "lines": [ + { + "bbox": [ + 430, + 121, + 481, + 171 + ], + "spans": [ + { + "bbox": [ + 430, + 121, + 481, + 171 + ], + "type": "image", + "image_path": "5ece8d59d27bd17e8e4cfbc09c3a462a2a4ea1c8a0114c686d9ca89c80482a6f.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 428, + 172, + 481, + 179 + ], + "lines": [ + { + "bbox": [ + 428, + 172, + 481, + 179 + ], + "spans": [ + { + "bbox": [ + 428, + 172, + 481, + 179 + ], + "type": "text", + "content": "pre-PE fusion (Ours)" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 243, + 480, + 303 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 243, + 480, + 303 + ], + "spans": [ + { + "bbox": [ + 130, + 243, + 480, + 303 + ], + "type": "text", + "content": "when the gradient of the estimated SDF is not detached in Eqn. 3. We consider that introducing the gradient of the estimated SDF places an additional burden on the learning of both the geometry and radiance networks, making it easier to fall into local minima. Thereby, we choose to detach the gradient of SDF in Eqn. 3." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 130, + 304, + 482, + 448 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 304, + 482, + 448 + ], + "spans": [ + { + "bbox": [ + 130, + 304, + 482, + 448 + ], + "type": "text", + "content": "Fuse the viewing direction and reflection direction before or after positional encoding? We introduce the fusion of viewing direction and reflection direction in Eqn. 2 before the operation of position encoding (PE). We compare the reconstruction results with different fusion orders, specifically the proposed pre-PE fusion: " + }, + { + "bbox": [ + 130, + 304, + 482, + 448 + ], + "type": "inline_equation", + "content": "\\mathrm{PE}(\\mathbf{d}_{\\mathrm{hyb}})" + }, + { + "bbox": [ + 130, + 304, + 482, + 448 + ], + "type": "text", + "content": " (ours) and alternative approach using post-PE fusion: " + }, + { + "bbox": [ + 130, + 304, + 482, + 448 + ], + "type": "inline_equation", + "content": "\\alpha \\cdot \\mathrm{PE}(\\mathbf{d}_{\\mathrm{ref}}) + (1 - \\alpha) \\cdot \\mathrm{PE}(\\mathbf{d}_{\\mathrm{view}})" + }, + { + "bbox": [ + 130, + 304, + 482, + 448 + ], + "type": "text", + "content": ". The comparison is shown in Fig. 8. We observe that fusing directions after the PE operation may result in degraded reconstruction results. We consider that fusing the directional information in the high-dimension space may increase the complexity of the learning space of the radiance network, which brings challenges in finding optimal reconstruction results. Thereby, we choose to fuse the viewing direction and reflection direction before the operation of positional encoding." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 131, + 463, + 312, + 475 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 463, + 312, + 475 + ], + "spans": [ + { + "bbox": [ + 131, + 463, + 312, + 475 + ], + "type": "text", + "content": "5 Conclusion and Limitation" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 130, + 486, + 482, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 486, + 482, + 617 + ], + "spans": [ + { + "bbox": [ + 130, + 486, + 482, + 617 + ], + "type": "text", + "content": "In this paper, we conduct an in-depth investigation into the topic of directional parameterization in neural implicit surface reconstruction. We reveal that the commonly used directional parameterizations in modeling the view-dependent radiance field, such as viewing and reflection directions, both have their limitations and none of them can effectively handle both specular surfaces and complex structures. A novel hybrid directional parameterization is introduced to achieve this goal, which softly fuses the viewing and reflection directions based on the distance of the sampled points from the surface, uniting the advantages of each representation. Extensive experiments demonstrate that our hybrid directional parameterization outperforms the existing directional parameterizations in reconstructing objects with a wide variety of materials, geometry, and appearance." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 130, + 617, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 665 + ], + "type": "text", + "content": "One limitation of our method is that the optimal choice of initial value for " + }, + { + "bbox": [ + 130, + 617, + 482, + 665 + ], + "type": "inline_equation", + "content": "\\gamma_{b}" + }, + { + "bbox": [ + 130, + 617, + 482, + 665 + ], + "type": "text", + "content": " may vary depending on the target object, which would require additional efforts for parameter tuning. An automated method for selecting the optimal initial value of " + }, + { + "bbox": [ + 130, + 617, + 482, + 665 + ], + "type": "inline_equation", + "content": "\\gamma_{b}" + }, + { + "bbox": [ + 130, + 617, + 482, + 665 + ], + "type": "text", + "content": " for target objects could be a potential future work for us." + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 246, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 246, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 246, + 101 + ], + "type": "text", + "content": "Jiang and Xu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 197, + 126 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 197, + 126 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 197, + 126 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 137, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 24, + "blocks": [ + { + "bbox": [ + 138, + 137, + 480, + 159 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 137, + 480, + 159 + ], + "spans": [ + { + "bbox": [ + 138, + 137, + 480, + 159 + ], + "type": "text", + "content": "1. Aanaes, H., Jensen, R.R., Vogiatzis, G., Tola, E., Dahl, A.B.: Large-scale data for multiple-view stereopsis. IJCV 120, 153-168 (2016)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 160, + 480, + 180 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 160, + 480, + 180 + ], + "spans": [ + { + "bbox": [ + 138, + 160, + 480, + 180 + ], + "type": "text", + "content": "2. Cai, B., Huang, J., Jia, R., Lv, C., Fu, H.: Neuda: Neural deformable anchor for high-fidelity implicit surface reconstruction. In: CVPR (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 182, + 480, + 201 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 182, + 480, + 201 + ], + "spans": [ + { + "bbox": [ + 138, + 182, + 480, + 201 + ], + "type": "text", + "content": "3. Chen, Z., Zhang, H.: Learning implicit fields for generative shape modeling. In: CVPR (2019)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 203, + 480, + 223 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 203, + 480, + 223 + ], + "spans": [ + { + "bbox": [ + 138, + 203, + 480, + 223 + ], + "type": "text", + "content": "4. Community, B.O.: Blender - a 3D modelling and rendering package. Blender Foundation, Stichting Blender Foundation, Amsterdam (2018)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 224, + 480, + 245 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 224, + 480, + 245 + ], + "spans": [ + { + "bbox": [ + 138, + 224, + 480, + 245 + ], + "type": "text", + "content": "5. Darmon, F., Bascle, B., Devaux, J., Monasse, P., Aubry, M.: Improving neural implicit surfaces geometry with patch warping. In: CVPR (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 246, + 480, + 266 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 246, + 480, + 266 + ], + "spans": [ + { + "bbox": [ + 138, + 246, + 480, + 266 + ], + "type": "text", + "content": "6. Fu, Q., Xu, Q., Ong, Y.S., Tao, W.: Geo-neus: Geometry-consistent neural implicit surfaces learning for multi-view reconstruction. In: NeurIPS (2022)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 267, + 480, + 288 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 267, + 480, + 288 + ], + "spans": [ + { + "bbox": [ + 138, + 267, + 480, + 288 + ], + "type": "text", + "content": "7. Furukawa, Y., Hernández, C., et al.: Multi-view stereo: A tutorial. Foundations and Trends® in Computer Graphics and Vision 9(1-2), 1-148 (2015)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 289, + 480, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 289, + 480, + 319 + ], + "spans": [ + { + "bbox": [ + 138, + 289, + 480, + 319 + ], + "type": "text", + "content": "8. Ge, W., Hu, T., Zhao, H., Liu, S., Chen, Y.C.: Ref-neus: Ambiguity-reduced neural implicit surface learning for multi-view reconstruction with reflection. In: ICCV (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 321, + 480, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 321, + 480, + 342 + ], + "spans": [ + { + "bbox": [ + 138, + 321, + 480, + 342 + ], + "type": "text", + "content": "9. Hedman, P., Srinivasan, P.P., Mildenhall, B., Barron, J.T., Debevec, P.: Baking neural radiance fields for real-time view synthesis. In: CVPR (2021)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 343, + 480, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 343, + 480, + 363 + ], + "spans": [ + { + "bbox": [ + 138, + 343, + 480, + 363 + ], + "type": "text", + "content": "0. Kato, H., Beker, D., Morariu, M., Ando, T., Matsuoka, T., Kehl, W., Gaidon, A.: Differentiable rendering: A survey. arXiv preprint arXiv:2006.12057 (2020)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 365, + 480, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 365, + 480, + 384 + ], + "spans": [ + { + "bbox": [ + 138, + 365, + 480, + 384 + ], + "type": "text", + "content": "1. Li, T.M., Aittala, M., Durand, F., Lehtinen, J.: Differentiable monte carlo ray tracing through edge sampling. ACM Transactions on Graphics 37(6), 1-11 (2018)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 386, + 480, + 406 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 386, + 480, + 406 + ], + "spans": [ + { + "bbox": [ + 138, + 386, + 480, + 406 + ], + "type": "text", + "content": "2. Li, Z., Müller, T., Evans, A., Taylor, R.H., Unberath, M., Liu, M.Y., Lin, C.H.: Neuralangelo: High-fidelity neural surface reconstruction. In: CVPR (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 407, + 480, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 407, + 480, + 437 + ], + "spans": [ + { + "bbox": [ + 138, + 407, + 480, + 437 + ], + "type": "text", + "content": "3. Liang, R., Chen, H., Li, C., Chen, F., Panneer, S., Vijaykumar, N.: Envidr: Implicit differentiable renderer with neural environment lighting. arXiv preprint arXiv:2303.13022 (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 439, + 480, + 471 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 439, + 480, + 471 + ], + "spans": [ + { + "bbox": [ + 138, + 439, + 480, + 471 + ], + "type": "text", + "content": "4. Liu, Y., Wang, P., Lin, C., Long, X., Wang, J., Liu, L., Komura, T., Wang, W.: Nano: Neural geometry and brdf reconstruction of reflective objects from multiview images. ACM Transactions on Graphics (2023)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 472, + 480, + 492 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 472, + 480, + 492 + ], + "spans": [ + { + "bbox": [ + 138, + 472, + 480, + 492 + ], + "type": "text", + "content": "5. Long, X., Lin, C., Wang, P., Komura, T., Wang, W.: Sparseneus: Fast generalizable neural surface reconstruction from sparse views. In: ECCV (2022)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 494, + 480, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 494, + 480, + 514 + ], + "spans": [ + { + "bbox": [ + 138, + 494, + 480, + 514 + ], + "type": "text", + "content": "6. Mescheder, L., Oechsle, M., Niemeyer, M., Nowozin, S., Geiger, A.: Occupancy networks: Learning 3d reconstruction in function space. In: CVPR (2019)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 138, + 514, + 480, + 546 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 514, + 480, + 546 + ], + "spans": [ + { + "bbox": [ + 138, + 514, + 480, + 546 + ], + "type": "text", + "content": "7. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. In: ECCV (2020)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 138, + 547, + 480, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 547, + 480, + 578 + ], + "spans": [ + { + "bbox": [ + 138, + 547, + 480, + 578 + ], + "type": "text", + "content": "8. Müller, T., Evans, A., Schied, C., Keller, A.: Instant neural graphics primitives with a multiresolution hash encoding. ACM Transactions on Graphics 41(4), 1-15 (2022)" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 138, + 580, + 480, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 580, + 480, + 600 + ], + "spans": [ + { + "bbox": [ + 138, + 580, + 480, + 600 + ], + "type": "text", + "content": "9. Niemeyer, M., Mescheder, L., Oechsle, M., Geiger, A.: Occupancy flow: 4d reconstruction by learning particle dynamics. In: ICCV (2019)" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 138, + 601, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 601, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 138, + 601, + 480, + 632 + ], + "type": "text", + "content": "20. Niemeyer, M., Mescheder, L., Oechsle, M., Geiger, A.: Differentiable volumetric rendering: Learning implicit 3d representations without 3d supervision. In: CVPR (2020)" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 138, + 634, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 634, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 634, + 480, + 665 + ], + "type": "text", + "content": "21. Niemeyer, M., Mescheder, L., Oechsle, M., Geiger, A.: Differentiable volumetric rendering: Learning implicit 3d representations without 3d supervision. In: CVPR (2020)" + } + ] + } + ], + "index": 23 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 284, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 284, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 284, + 91, + 447, + 102 + ], + "type": "text", + "content": "Rethinking directional parameterization" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 657 + ], + "type": "list", + "angle": 0, + "index": 22, + "blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 480, + 138 + ], + "type": "text", + "content": "22. Oechsle, M., Peng, S., Geiger, A.: Unisurf: Unifying neural implicit surfaces and radiance fields for multi-view reconstruction. In: ICCV (2021)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 138, + 480, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 138, + 480, + 160 + ], + "spans": [ + { + "bbox": [ + 132, + 138, + 480, + 160 + ], + "type": "text", + "content": "23. Park, J.J., Florence, P., Straub, J., Newcombe, R., Lovegrove, S.: Deepsdf: Learning continuous signed distance functions for shape representation. In: CVPR (2019)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 160, + 480, + 181 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 160, + 480, + 181 + ], + "spans": [ + { + "bbox": [ + 132, + 160, + 480, + 181 + ], + "type": "text", + "content": "24. Peng, S., Niemeyer, M., Mescheder, L., Pollefeys, M., Geiger, A.: Convolutional occupancy networks. In: ECCV (2020)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 181, + 480, + 214 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 181, + 480, + 214 + ], + "spans": [ + { + "bbox": [ + 132, + 181, + 480, + 214 + ], + "type": "text", + "content": "25. Ravi, N., Reizenstein, J., Novotny, D., Gordon, T., Lo, W.Y., Johnson, J., Gkioxari, G.: Accelerating 3d deep learning with pytorch3d. arXiv preprint arXiv:2007.08501 (2020)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 214, + 480, + 246 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 214, + 480, + 246 + ], + "spans": [ + { + "bbox": [ + 132, + 214, + 480, + 246 + ], + "type": "text", + "content": "26. Srinivasan, P.P., Deng, B., Zhang, X., Tancik, M., Mildenhall, B., Barron, J.T.: Nerv: Neural reflectance and visibility fields for relighting and view synthesis. In: CVPR (2021)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 246, + 480, + 268 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 246, + 480, + 268 + ], + "spans": [ + { + "bbox": [ + 132, + 246, + 480, + 268 + ], + "type": "text", + "content": "27. Tulsiani, S., Zhou, T., Efros, A.A., Malik, J.: Multi-view supervision for single-view reconstruction via differentiable ray consistency. In: CVPR (2017)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 268, + 480, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 268, + 480, + 300 + ], + "spans": [ + { + "bbox": [ + 132, + 268, + 480, + 300 + ], + "type": "text", + "content": "28. Verbin, D., Hedman, P., Mildenhall, B., Zickler, T., Barron, J.T., Srinivasan, P.P.: Ref-nerf: Structured view-dependent appearance for neural radiance fields. In: CVPR (2022)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 300, + 480, + 332 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 300, + 480, + 332 + ], + "spans": [ + { + "bbox": [ + 132, + 300, + 480, + 332 + ], + "type": "text", + "content": "29. Wang, P., Liu, L., Liu, Y., Theobalt, C., Komura, T., Wang, W.: Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. In: NeurIPS (2021)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 333, + 480, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 333, + 480, + 365 + ], + "spans": [ + { + "bbox": [ + 132, + 333, + 480, + 365 + ], + "type": "text", + "content": "30. Wang, Y., Han, Q., Habermann, M., Daniilidis, K., Theobalt, C., Liu, L.: Neus2: Fast learning of neural implicit surfaces for multi-view reconstruction. In: ICCV (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 365, + 480, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 365, + 480, + 387 + ], + "spans": [ + { + "bbox": [ + 132, + 365, + 480, + 387 + ], + "type": "text", + "content": "31. Wu, T., Wang, J., Pan, X., Xu, X., Theobalt, C., Liu, Z., Lin, D.: Voxurf: Voxel-based efficient and accurate neural surface reconstruction. In: ICLR (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 387, + 480, + 419 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 387, + 480, + 419 + ], + "spans": [ + { + "bbox": [ + 132, + 387, + 480, + 419 + ], + "type": "text", + "content": "32. Xie, Y., Takikawa, T., Saito, S., Litany, O., Yan, S., Khan, N., Tombari, F., Tompkin, J., Sitzmann, V., Sridhar, S.: Neural fields in visual computing and beyond. Computer Graphics Forum 41(2), 641-676 (2022)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 419, + 480, + 452 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 419, + 480, + 452 + ], + "spans": [ + { + "bbox": [ + 132, + 419, + 480, + 452 + ], + "type": "text", + "content": "33. Yan, X., Yang, J., Yumer, E., Guo, Y., Lee, H.: Perspective transformer nets: Learning single-view 3d object reconstruction without 3d supervision. In: NeurIPS (2016)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 452, + 480, + 484 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 452, + 480, + 484 + ], + "spans": [ + { + "bbox": [ + 132, + 452, + 480, + 484 + ], + "type": "text", + "content": "34. Yao, Y., Zhang, J., Liu, J., Qu, Y., Fang, T., McKinnon, D., Tsin, Y., Quan, L.: Neilf: Neural incident light field for material and lighting estimation. In: ECCV (2022)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 484, + 480, + 506 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 484, + 480, + 506 + ], + "spans": [ + { + "bbox": [ + 132, + 484, + 480, + 506 + ], + "type": "text", + "content": "35. Yariv, L., Gu, J., Kasten, Y., Lipman, Y.: Volume rendering of neural implicit surfaces. In: NeurIPS (2021)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 506, + 480, + 538 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 506, + 480, + 538 + ], + "spans": [ + { + "bbox": [ + 132, + 506, + 480, + 538 + ], + "type": "text", + "content": "36. Yariv, L., Kasten, Y., Moran, D., Galun, M., Atzmon, M., Ronen, B., Lipman, Y.: Multiview neural surface reconstruction by disentangling geometry and appearance. In: NeurIPS (2020)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 538, + 480, + 559 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 538, + 480, + 559 + ], + "spans": [ + { + "bbox": [ + 132, + 538, + 480, + 559 + ], + "type": "text", + "content": "37. Yu, A., Ye, V., Tancik, M., Kanazawa, A.: pixelnerf: Neural radiance fields from one or few images. In: CVPR (2021)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 559, + 480, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 559, + 480, + 581 + ], + "spans": [ + { + "bbox": [ + 132, + 559, + 480, + 581 + ], + "type": "text", + "content": "38. Yu, Z., Peng, S., Niemeyer, M., Sattler, T., Geiger, A.: Monosdf: Exploring monocular geometric cues for neural implicit surface reconstruction. In: NeurIPS (2022)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 581, + 480, + 603 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 581, + 480, + 603 + ], + "spans": [ + { + "bbox": [ + 132, + 581, + 480, + 603 + ], + "type": "text", + "content": "39. Zhang, J., Yao, Y., Li, S., Liu, J., Fang, T., McKinnon, D., Tsin, Y., Quan, L.: Neilf++: Inter-reflectable light fields for geometry and material estimation (2023)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 132, + 603, + 480, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 603, + 480, + 624 + ], + "spans": [ + { + "bbox": [ + 132, + 603, + 480, + 624 + ], + "type": "text", + "content": "40. Zhang, K., Riegler, G., Snavely, N., Koltun, V.: Nerf++: Analyzing and improving neural radiance fields. arXiv preprint arXiv:2010.07492 (2020)" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 132, + 624, + 480, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 624, + 480, + 657 + ], + "spans": [ + { + "bbox": [ + 132, + 624, + 480, + 657 + ], + "type": "text", + "content": "41. Zhang, X., Srinivasan, P.P., Deng, B., Debevec, P., Freeman, W.T., Barron, J.T.: Nerfactor: Neural factorization of shape and reflectance under an unknown illumination. ACM Transactions on Graphics 40(6), 1-18 (2021)" + } + ] + } + ], + "index": 21 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 246, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 246, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 246, + 101 + ], + "type": "text", + "content": "Jiang and Xu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Rethinking Fast Adversarial Training_ A Splitting Technique To Overcome Catastrophic Overfitting/54a9d439-23cc-4380-af56-e12fd67a1f40_content_list.json b/2024/Rethinking Fast Adversarial Training_ A Splitting Technique To Overcome Catastrophic Overfitting/54a9d439-23cc-4380-af56-e12fd67a1f40_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..ba24de963d7e039da623603a5a151a99bcef4d64 --- /dev/null +++ b/2024/Rethinking Fast Adversarial Training_ A Splitting Technique To Overcome Catastrophic Overfitting/54a9d439-23cc-4380-af56-e12fd67a1f40_content_list.json @@ -0,0 +1,1532 @@ +[ + { + "type": "text", + "text": "Rethinking Fast Adversarial Training: A Splitting Technique to Overcome Catastrophic Overfitting", + "text_level": 1, + "bbox": [ + 218, + 140, + 782, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Masoumeh Zareapoor and Pourya Shamsolmoali\\*", + "bbox": [ + 308, + 212, + 692, + 227 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "East China Normal University, Queen's University Belfast", + "bbox": [ + 307, + 239, + 696, + 253 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Catastrophic overfitting (CO) poses a significant challenge to fast adversarial training (FastAT), particularly at large perturbation scales, leading to dramatic reductions in adversarial test accuracy. Our analysis of existing FastAT methods shows that CO is accompanied by abrupt and irregular fluctuations in loss convergence, indicating that a stable training dynamic is key to preventing CO. Therefore, we propose a training model that uses the Douglas-Rachford (DR) splitting technique to ensure a balanced and consistent training progression, effectively counteracting CO. The DR splitting technique, known for its ability to solve complex optimization problems, offering a distinct advantage over classical FastAT methods by providing a smoother loss convergence. This is achieved without resorting to complex regularization or incurring the computational costs associated with double backpropagation, presenting an efficient solution to enhance adversarial robustness. Our comprehensive evaluation conducted across standard datasets, demonstrates that our DR splitting-based model not only improves adversarial robustness but also achieves this with remarkable efficiency compared to various FastAT methods. This efficiency is particularly observed under conditions involving long training schedules and large adversarial perturbations.", + "bbox": [ + 261, + 287, + 743, + 554 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Adversarial training $\\cdot$ Adversarial robustness $\\cdot$ Convergence", + "bbox": [ + 261, + 565, + 741, + 580 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 604, + 377, + 619 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Deep neural networks (DNNs), despite their successes, are vulnerable to adversarial attacks, where small, undetectable, input perturbations can mislead models into incorrect predictions [11]. Given the security risks of such vulnerabilities, improving neural network robustness against these attacks has become a critical area of research [28, 38]. Adversarial training (AT) [18, 30] and its derivative, PGD-AT [22, 31], have emerged as effective strategies for enhancing the robustness of neural networks. However, the complex inner optimizations required by these methods at each training step significantly elevate computational costs, limiting their practicality for large-scale applications [28]. To address these computational challenges, 'fast' versions of AT (FastAT) [25, 36] have been proposed, which employ a single-step gradient propagation [11] to produce adversarial training examples. While FastAT offers a more efficient training process", + "bbox": [ + 212, + 635, + 787, + 818 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Correspondence to pshams55@gmail.com. Code: https://github.com/mzarea2/RAT", + "bbox": [ + 217, + 824, + 787, + 840 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "compared to PGD-AT, it introduces the problem of Catastrophic Overfitting (CO) [16,36]. CO occurs when training with strong adversaries, leading to a significant drop in the robustness of the resulting model. Several FastAT methods have been developed to reduce CO [42]. Wong et al. [36] enhance adversarial example diversity through the use of initial perturbations that are randomly initialized. Following this approach, [3] proposed FAST-GA, an efficient regularizer that aims to explicitly align gradients between benign and adversarial inputs. Likewise, MEP-PGI [15] and NuAT [33] employ nuclear or p-norm regularization within adversarial training to improve consistency in predictions on both benign and adversarial examples. These approaches are effective only within a limited perturbation $\\mathcal{E} \\leq 8/255$ , where $\\mathcal{E}$ denotes the degree of perturbations in the adversarial training [42]. Moreover, models trained with relatively small perturbation scales show vulnerability to larger $\\mathcal{E}$ , with noted declines in robustness.", + "bbox": [ + 212, + 146, + 787, + 343 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "For example, models trained with NuAT and MEP-PGI at a perturbation size of $\\mathcal{E} = 8 / 255$ demonstrated robustness levels of $56\\%$ and $55\\%$ against PGD-50 attacks, but their robustness significantly dropped to $24\\%$ and $22\\%$ , respectively, with $\\mathcal{E} = 16 / 255$ . Our investigation into adversarial training processes (illustrated in Fig. 2) reveals that CO is characterized by minor fluctuations in benign classification loss and significant decreases in adversarial example loss. This observation raises a question: How can we design fast adversarial training that ensures improved stability and mitigates CO, while also optimizing the trade-off between robustness accuracy (RA) and standard accuracy (SA)?", + "bbox": [ + 212, + 343, + 787, + 478 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this context, the Douglas-Rachford (DR) splitting technique emerges as a promising solution. The DR splitting is known for its linear convergence across both strongly convex and non-convex settings [4, 20, 23, 26]. Several studies demonstrate its applicability in a variety of optimization tasks. For example, Liang et al. [19] and Applegate et al. [4] have highlighted its capability for local and global linear convergence, underscoring the DR technique's effectiveness in complex optimization scenarios. We also capitalize on the inherent stability and efficiency of the DR splitting, incorporating these properties into adversarial training to address CO. This integration allows for a more stable and consistent adversarial learning process, significantly lowering the risk of CO. Indeed, our implementation demonstrates that DR splitting not only stabilizes the training process but also offers a cost-effective solution to enhance model robustness against adversarial attacks, as shown in Fig. 1. Our contributions are threefold: ① We provide a comprehensive analysis showing that existing FastAT methods are still prone to CO with larger $\\varepsilon$ . Our investigation emphasizes the critical role of loss function convergence stability in addressing this vulnerability. ② Introducing a novel solution with the Douglas-Rachford splitting technique, specifically designed to mitigate CO without incurring computational complexities. ③ Conduct extensive experiments on standard datasets to validate the effectiveness of our model- Rethinking fast Adversarial Training (RAT)- particularly against large adversarial perturbations. RAT, leveraging the DR-splitting technique, shows enhanced adversarial robustness when compared to classical adversarial training methods (find the training code in the supplementary file).", + "bbox": [ + 212, + 479, + 787, + 828 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "Zareapoor et al.", + "bbox": [ + 271, + 114, + 380, + 128 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/4475771b46e5295fa31d61cffa4f7fc85ab252215ffc9af798a910b05cb37ac6.jpg", + "image_caption": [ + "Fig. 1: Performance comparison of adversarial robustness with WideResNet34-10 on CIFAR10 for $\\mathcal{E} = (8,16) / 255$ . (left) Illustrates the average runtime for each method. (right) We compare AA (AutoAttack) and SA accuracies. Lower AA indicates higher vulnerability. Higher SA (standard accuracy) indicates better general performance." + ], + "image_footnote": [], + "bbox": [ + 217, + 148, + 400, + 257 + ], + "page_idx": 2 + }, + { + "type": "table", + "img_path": "images/c2f83e2c742a76b1c0fd1017ac7469a95fe4b68ab6770018cd4e5a55150dc792.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodsAA (%)SA (%)AA (%)SA (%)
816
AT PGD-1047.69±(0.11)78.51±(0.08)24.97±(0.26)59.41±(0.46)
LLR [27]43.84±(0.20)74.52±(0.07)16.31±(0.20)44.76±(4.72)
Fast-Bat [41]45.83±(0.12)79.61±(0.12)18.30±(0.26)59.34±(0.25)
Fast-GA [3]46.37±(0.27)79.12±(0.07)15.36±(1.32)54.25±(6.28)
CURE [24]44.65±(0.17)77.04±(0.11)18.12±(0.35)53.49±(0.08)
RAT (Ours)46.21±(0.08)80.32±(0.12)19.58±(0.17)60.37±(0.21)
", + "bbox": [ + 408, + 143, + 779, + 247 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 354, + 387, + 371 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.1 Adversarial Attacks", + "text_level": 1, + "bbox": [ + 215, + 383, + 426, + 398 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Adversarial attacks aim to fool deep learning models by introducing perturbed inputs. Goodfellow et al. [11] introduced the concept of the fast gradient sign method (FGSM) for generating adversarial examples $x'$ within classification tasks. They demonstrated that adversarial examples, created through a single-step gradient update, can lead a classification model to incorrect predictions with significant confidence. The adversarial examples $x'$ are generated by", + "bbox": [ + 212, + 406, + 782, + 498 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nx ^ {\\prime} = x + \\mathcal {E}. \\operatorname {s g n} \\left(\\nabla_ {x} L (f (x; \\theta), y\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 385, + 505, + 785, + 523 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $(x,y)$ denotes an input sample and its corresponding label, $\\mathcal{E}$ is the perturbation magnitude, $\\nabla_{x}L$ is the loss gradient with respect to $x$ , and $\\theta$ signifies the fixed model weights. Subsequent research has developed more sophisticated attack methods, such as I-FGSM [17], MIM [9], and PGD [22], which leverage iterative gradient descent to enhance the efficacy of the attacks. The PGD attack, in particular, generates adversarial examples by applying small perturbations to the input data, ensuring these adjustments stay within a pre-defined boundary $\\mathcal{E}$ . This process is described as follows", + "bbox": [ + 212, + 530, + 787, + 650 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nx _ {t + 1} ^ {\\prime} = \\Pi_ {\\mathcal {E}} (x _ {t} ^ {\\prime} + \\alpha . \\mathrm {s g n} (\\nabla_ {x} L (f (x _ {t} ^ {\\prime}; \\theta), y))), \\qquad \\qquad (2)\n$$\n", + "text_format": "latex", + "bbox": [ + 349, + 659, + 785, + 678 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "in which $x'$ is the adversarial example at iteration $t$ , and $\\Pi_{\\mathcal{E}}$ denotes the projection operation that ensures the adversarial example remains within the $\\mathcal{E}$ -ball. The PGD attack is one of the most effective attack methods for evaluating the robustness of deep learning against adversarial examples.", + "bbox": [ + 212, + 685, + 787, + 746 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2 Adversarial Training (AT)", + "text_level": 1, + "bbox": [ + 215, + 756, + 478, + 772 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Adversarial Training (AT) [22] and TRADES [40] have become prominent training methods for training robust architectures, which can be written as", + "bbox": [ + 212, + 779, + 785, + 810 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {\\theta} E _ {(x, y) \\sim D} [ \\max _ {\\delta \\in [ - \\mathcal {E}, \\mathcal {E} ]} L (f (x + \\delta ; \\theta), y) ], \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 357, + 818, + 785, + 842 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "Rethinking Fast Adversarial Training", + "bbox": [ + 480, + 114, + 732, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/536e01bcbba3d09af02e5802bc6c636c6f872ceea42053455dcc8c68db8793e5.jpg", + "image_caption": [ + "Fig. 2: Performance of fast adversarial training (FastAT) methods on CIFAR10 using ResNet20, for $\\mathcal{E} = 16 / 255$ . The figure illustrates the training process across three runs. The terms AV-loss and BN-loss are the training loss for adversarial and benign inputs, respectively, while AV-acc and BN-acc quantify testing accuracy against both adversarial and benign samples. The occurrence of catastrophic overfitting is marked (in green), indicating critical periods where adversarial robustness drops." + ], + "image_footnote": [], + "bbox": [ + 233, + 145, + 769, + 396 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\delta$ is the adversarial perturbations and $x' = x + \\delta$ is the adversarial examples generated by various attacks such as PGD and FGSM. This optimization first, generates adversarial examples within a defined perturbation limit to maximize classification loss (inner maximization) and then adjusts the model parameters $\\theta$ to minimize this loss over the adversarial examples (outer minimization). However, a notable trade-off emerges in AT between computational efficiency and adversarial robustness, due to the complexity of solving the inner maximization problem. To enhance the computational efficiency of AT, several methods have been developed. Free AT [30], simultaneously updating the network parameters $\\theta$ and the adversarial perturbation $\\delta$ multiple times within a single batch. On the other hand, Fast-AT [36], simplifies the process further by employing a single-step FGSM attack [11] but the attack initiates from a randomly chosen point $x + \\mu$ with $\\mu \\in [-\\delta, \\delta]$ . Fast-Bat These approaches reduce the cost of solving the inner maximization problem, significantly accelerating the training process.", + "bbox": [ + 212, + 522, + 787, + 736 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.3 Catastrophic Overfitting (CO)", + "text_level": 1, + "bbox": [ + 215, + 750, + 511, + 767 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The acceleration of AT methods comes with significant drawbacks. Wong et al. [36] first reported this with FGSM training, and was later supported by Andriushchenko et al. [3] in their analysis of Free AT and Fast-AT methods. In these cases, the model's accuracy on adversarial tests can fall sharply, sometimes", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 116, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "Zareapoor et al.", + "bbox": [ + 271, + 114, + 379, + 128 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "to $0\\%$ , in just a few training epochs. This decline in performance has been attributed to CO, a critical challenge that has motivated significant research efforts aimed at understanding and mitigating its impact [14, 18, 41]. Below, we provide an overview of various strategies developed to avoid CO.", + "bbox": [ + 212, + 146, + 787, + 208 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "2.4 Regularization-based Approaches", + "text_level": 1, + "bbox": [ + 214, + 227, + 535, + 242 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Recent efforts focus on regularization and data augmentation to prevent CO. [3] introduced FAST-GA to mitigate CO by penalizing the misalignment of gradients between the input and their adversarial counterparts. This method ensures that the direction of the gradient remains consistent when slight perturbations are introduced to the input. Qin et al. [27] introduced local linear regularization (LLR) to improve model robustness against adversarial attacks. LLR smoothes the loss landscape using a first-order Taylor approximation, addressing nonlinear sensitivities and promoting a more linear loss response to perturbations. ELLE [28] introduced a regularization term designed to mitigate CO both effectively and efficiently within classical AT frameworks, by enforcing local linearity. Sriramanan et al. [32] introduced GAT, a novel regularization strategy aimed at reducing the difference between the logits of a training point and a random sample. NuAT [33], proposed a regularization term that penalizes the nuclear norm of the logits difference between benign and adversarial examples. However, GAT and NuAT still face challenges with CO under larger perturbation size $\\mathcal{E}$ [16].", + "bbox": [ + 212, + 258, + 787, + 486 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "2.5 Other Approaches", + "text_level": 1, + "bbox": [ + 214, + 506, + 413, + 522 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Park et al. [25] introduced SLAT which performs adversarial attacks within the latent space, focusing on the internal representations rather than the input space directly. Another study [35] employs the Frank-Wolfe optimization technique for AT, connecting the distortion observed in the loss landscape to the norm of adversarial perturbations, providing a deeper understanding of adversarial robustness. Li et al. [18] addressed CO by limiting network weights to a specific subspace based on the main components of AT-trained weights. Meanwhile, Zhang et al. [41] introduce Fast-Bat, which redefines the AT challenge as a Bilevel optimization problem and introduces a one-step solution to address it. However, [28] shows that Fast-BAT is affected by CO for $\\mathcal{E} > 10 / 255$ . Zero-Grad [10] introduces a strategy to mitigate adversarial perturbations by setting to zero the gradient components that fall below a threshold, however, [16] demonstrates, this approach results in decreased performance for large $\\mathcal{E}$ . Different from these methods, we aim to minimize the loss fluctuations during training through the Douglas-Rachford splitting technique, known for its effectiveness in complex optimization [23]. It has been successfully applied to optimal transport [20] and its discretizations [6], demonstrating its flexibility and robustness in achieving linear convergence rates under both convex and non-convex settings [4, 34]. A notable advantage of this splitting technique is to leverage sparsity in solutions, which can significantly accelerate the convergence process [26].", + "bbox": [ + 212, + 537, + 787, + 843 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "Rethinking Fast Adversarial Training", + "bbox": [ + 482, + 114, + 730, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3 Method", + "text_level": 1, + "bbox": [ + 215, + 143, + 330, + 160 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "This section evaluates FastAT methods under large perturbation budgets, $\\mathcal{E}$ , explores the causes of CO, and finally introduces our proposed solutions.", + "bbox": [ + 212, + 174, + 782, + 205 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.1 Performance of Fast Adversarial Training on a Large $\\mathcal{E}$", + "text_level": 1, + "bbox": [ + 214, + 219, + 714, + 234 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Fast AT methods have shown effectiveness in mitigating CO at lower perturbation $(\\mathcal{E} \\leq 8/255)$ [28, 36]. To extend this evaluation, we analyze performance at a higher perturbation level $(\\mathcal{E} = 16/255)$ , in line with [42]. FAST-AT [36] introduces an initial step for Eq. (3), where samples are perturbed with uniform random noise $\\delta_0 \\sim U(-\\mathcal{E}, \\mathcal{E})$ before applying the FGSM attack,", + "bbox": [ + 212, + 243, + 784, + 319 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {\\theta} E _ {(x, y) \\sim D} \\left[ \\max _ {\\delta_ {0} + \\delta \\in [ - \\mathcal {E}, \\mathcal {E} ]} L (f (x + \\delta_ {0} + \\delta ; \\theta), y) \\right]. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 316, + 328, + 785, + 349 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Here, $L(\\cdot)$ is the loss function, $\\delta$ represents adversarial perturbations, and $\\theta$ is the model weights. FAST-GA, as introduced by [3] focuses on aligning the gradients of the loss with respect to both benign samples $x$ and their perturbed samples $x + \\delta_0$ , which is described as,", + "bbox": [ + 214, + 359, + 787, + 421 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nE _ {(x, y) \\sim D} [ 1 - \\cos (\\nabla_ {x} L (x, \\theta), \\nabla_ {x + \\delta_ {0}} L (x + \\delta_ {0}, \\theta)) ], \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 330, + 431, + 784, + 446 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "in which $L(x,\\theta) = L(f(x;\\theta),y)$ . The cosine similarity $\\cos(\\cdot)$ measures the cosine similarity between the gradients of the loss function for the benign and perturbed inputs. This strategy aims to ensure that the model's behavior does not change drastically for adversarial examples, thereby reducing CO. Different from other attack methods, MEP-PGI [15] extends the FGSM attack by incorporating a Prior-Guided Initialization (PGI) mechanism. This technique initializes the adversarial perturbations $\\delta_0$ , by using the information from adversarial examples generated in previous training epochs,", + "bbox": [ + 212, + 455, + 787, + 578 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nE _ {(x, y) \\sim D} [ \\| f (x ^ {\\prime}; \\theta) - f (x + \\delta_ {0}; \\theta) \\| _ {2} ^ {2} ], \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 375, + 588, + 784, + 604 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "here, $||\\cdot||$ is the squared (L2) norm, measuring the distance between the model's outputs for the two types of examples. FAST-PGI aims to enrich the adversarial examples with historical insight, thereby enhancing the adversarial training process. Fig. 2 illustrates the adversarial training of ResNet20 on CIFAR10, highlighting times when CO occurs, between epochs 5th-20th, 40th-65th, and 10th-15th. The key observations from the results are i) transition from smooth/stable to erratic loss fluctuations; ii) discrepancies in the loss for benign and adversarial examples, i.e., minor fluctuations in benign loss $L(x,\\theta)$ followed with a rapid decline in adversarial loss $L(x',\\theta)$ ; and iii) a sharp decrease in adversarial accuracy. The result is consistent with the observations in [1,28,42]. From these patterns, we derive several insights for understanding CO: it is obvious that $L(x,\\theta)$ has greater stability than the loss $L(x',\\theta)$ , indicating a tendency of models to overfit to adversarial samples. CO directly correlates with the instability in adversarial training convergence. Anomalies in both benign and adversarial loss occur together, signifying an issue in the learning process that affects its performance", + "bbox": [ + 212, + 613, + 787, + 843 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "Zareapoor et al.", + "bbox": [ + 271, + 114, + 380, + 128 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "on both data. We also find that Fast-GA and MEP-PGI enhance the stability of adversarial training through the constraints in Eq. (4) and Eq. (5), compared to the Fast-AT. We are now in shape to give a formal definition of our model.", + "bbox": [ + 212, + 146, + 782, + 191 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.2 Douglas-Rachford (DR) Splitting for Adversarial Training", + "text_level": 1, + "bbox": [ + 214, + 209, + 738, + 224 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We use the DR splitting technique to ensure a consistent and stable optimization to prevent sudden changes in adversarial robustness. This technique divides the optimization process into a two-step process (reflection and enforcement), which together ensure a stable and consistent optimization path. Let's denote the model at iteration $k$ by $f(\\cdot, \\theta^{(k)})$ , where $\\theta^{(k)}$ represents the model parameters. The loss function for adversarial examples is $L(f(x', \\theta^{(k)}, y))$ , and for benign examples, $L(f(x^{k}, \\theta^{(k)}, y))$ . An important property of our model is its ability to alternate between reducing adversarial loss and applying careful regularization terms. This ensures that any adjustments to the model parameters effectively impact both benign and adversarial loss functions $L(x, \\theta)$ and $L(x', \\theta)$ . Specifically, to address CO, we introduce epoch-based constraints on the loss differences for both adversarial and benign instances, represented as $(\\Delta L_{\\theta^{(k)}}, \\Delta L_{\\theta^{(k)}}^{benign})$ ,", + "bbox": [ + 212, + 229, + 784, + 412 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\Delta L _ {\\theta^ {(k)}} = \\left| L \\left(f \\left(x ^ {\\prime (k)}, \\theta^ {(k)}\\right), y\\right) - L \\left(f \\left(x ^ {\\prime (k - 1)}, \\theta^ {(k - 1)}\\right), y\\right) \\right| \\\\ \\begin{array}{l} \\left. \\right| = \\left| L (f (w, \\theta^ {(k)}), y) - L (f (x, \\theta^ {(k - 1)}), y) \\right|, \\end{array} \\tag {7} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 308, + 419, + 784, + 460 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $|.|$ measures the absolute difference in loss for adversarial $(x')$ and benign $(x)$ inputs, $x'^{(k)} = x + \\delta_{0,k} + \\delta_k$ . Here, $\\delta_{0,k}$ can be replaced by various attack initialization perturbations and $\\delta_k$ is generated by $[\\max_{\\delta_k \\in [-\\varepsilon, \\varepsilon]} L(x'^{(k)}, \\theta^{(k-1)})]$ . This formulation guides the model to effectively manage perturbations and maintain a stable training process. Additionally, these constraints help in mitigating the simultaneous anomalies in both $L(x, \\theta)$ and $L(x', \\theta)$ facilitating a more controlled update of the model parameters that considers both adversarial and benign inputs. In practice, this involves updating the model's intermediate parameters, $C^k$ , to reduce the adversarial loss, according to the gradient of the loss function $\\nabla_\\theta$ with respect to the adversarial examples $x'^{(k)}$ ,", + "bbox": [ + 212, + 465, + 784, + 614 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\nC ^ {k} = \\theta^ {(k)} - \\eta \\nabla_ {\\theta} L \\left(f \\left(x ^ {\\prime (k)}, \\theta^ {(k)}\\right), y\\right), \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 375, + 622, + 784, + 640 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $\\eta$ is a learning rate. This update, termed as the 'reflection step' ( $\\theta_{\\mathrm{ref}}$ ), is pivotal in aligning the model's parameters towards reducing adversarial vulnerability. We then combine these updates $C^k$ with regularization adjustments to stabilize the training dynamics and enhance the adversarial robustness, effectively preventing overfitting to adversarial examples while maintaining performance on benign inputs. The updated model parameters, $\\theta^{(k + 1)}$ , allow to handle both adversarial and benign inputs effectively without compromising training stability or performance,", + "bbox": [ + 212, + 645, + 782, + 765 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\theta^ {(k + 1)} = C ^ {k} + \\lambda (R e g (\\theta^ {(k)}) - C ^ {k}), \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 380, + 766, + 784, + 782 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $\\lambda \\in [0,1]$ controls the strength of the regularization, and Reg involves components in Eq. (7) such that", + "bbox": [ + 212, + 785, + 782, + 816 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\nR e g \\left(\\theta^ {(k)}\\right) = \\gamma_ {1} \\left(\\Delta L _ {\\theta^ {(k)}}\\right) + \\gamma_ {2} \\left(\\Delta L _ {\\theta^ {(k)}} ^ {b e n i g n}\\right). \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 359, + 823, + 784, + 842 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "Rethinking Fast Adversarial Training", + "bbox": [ + 480, + 114, + 732, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The important reason behind Eq. (9) is that the regularization term $Reg(\\theta^{(k)})$ incorporates changes in loss for both adversarial and benign examples, weighted by $\\gamma_{1}$ and $\\gamma_{2}$ , respectively. This helps to balance the model's performance across adversarial and benign inputs, addressing CO by preventing the model from becoming overly specialized on the adversarial examples. Because CO occurs when a model overfits to specific adversarial examples during training, compromising its ability to generalize [21]. Our formulation ensures that the model's updates consider the impact on both adversarial robustness and benign accuracy. This regularization can be seen as part of the 'enforcement step' $(\\theta_{enf})$ . The DR technique involves iteratively updating the model parameters by combining the adversarial loss minimization and the stability enforcement components, as $\\theta = 0.5 \\times (\\theta_{ref} + \\theta_{enf})$ . $\\theta_{ref}$ targets reducing the loss on adversariably perturbed inputs Eq. (8), conversely, $\\theta_{enf}$ represents the parameters adjusted after applying the regularization term, focusing on stability and CO prevention Eq. (9). This approach under the DR-splitting technique provides a comprehensive strategy for addressing CO in the adversarial training process. Indeed, this allows the model to effectively adapt to adversarial perturbations while ensuring that such adaptations do not compromise the model's ability to accurately classify benign inputs or lead to overfitting on the adversarial examples during training. Our proposed RAT is given in Algorithm 1, discussion in the supplementary file.", + "bbox": [ + 212, + 145, + 787, + 450 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Efficient Computation of Loss Differences In practice, directly computing the losses $L(f(x^{\\prime (k - 1)},\\theta^{(k - 1)},y))$ and $L(f(x,\\theta^{(k - 1)}),y)$ in Eq. (7) can consume a significant amount of memory. To address this, we introduce $u^{k - 1}$ for benign inputs and $u^{\\prime k - 1}$ for adversarial inputs, replacements for these terms, allowing for efficient computation and tracking of the average loss over benign and adversarial examples, respectively. Specially, $u^{k - 1} = \\frac{1}{N}\\sum_{i}L(f(x_{i}^{\\prime},\\theta^{(k - 1)}),y_{i})$ represents the average loss for adversarial inputs, and $u^{\\prime k - 1} = \\frac{1}{N}\\sum_{i = 1}^{N}L(f(x_i,\\theta^{(k - 1)}),y_i)$ denotes the average loss for benign inputs, where $N$ is the batch size. This can reduce computational and memory overhead by focusing on the aggregate loss across batches, rather than calculating them individually for each sample.", + "bbox": [ + 212, + 455, + 789, + 609 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4 Experimental Results", + "text_level": 1, + "bbox": [ + 215, + 632, + 460, + 648 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.1 Experimental settings", + "text_level": 1, + "bbox": [ + 215, + 664, + 442, + 680 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this section, we evaluate the proposed model across four benchmark datasets to determine its effectiveness and robustness against adversarial attacks. For both CIFAR10 [8] and CIFAR100 [8], we use the ResNet20 [12] architecture as the backbone, recognizing its strong performance in image classification tasks. For the more complex Tiny-ImageNet and ImageNet [8], we use the PreActResNet18 [13] to better handle its challenges. Additionally, the use of WideResNet34-10 [39] highlights our model's capacity to maintain effectiveness on long training schedules. We use the SGD optimizer with a batch size of 128, a weight decay of $5.10^{-4}$ , and a momentum of 0.9. We initiate training with learning rates of 0.1 for CIFAR10 and CIFAR100. For Tiny ImageNet, due to its greater complexity,", + "bbox": [ + 212, + 688, + 787, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "Zareapoor et al.", + "bbox": [ + 271, + 114, + 380, + 128 + ], + "page_idx": 7 + }, + { + "type": "code", + "sub_type": "algorithm", + "code_caption": [ + "Algorithm 1 Rethinking fast Adversarial Training (RAT) using DR Splitting" + ], + "code_body": "Require: Training dataset $(x_{i},y_{i})_{i = 1}^{N}$ # epochs K, # batch B, perturbation budget $\\mathcal{E},\\lambda ,\\gamma_1,\\gamma_2$ \n1: Initialize model parameters $\\theta_0$ \n2: for epoch $= 1$ to K do \n3: for $i = 1$ to B do \n4: $x_{i}^{\\prime} = x_{i} + \\delta \\mathrm{sign}(\\nabla_{x_{i}}L(f(x_{i};\\theta_{\\mathrm{epoch - 1}}),y_{i}))\\triangleright$ Adversarial example generation \n5: $L_{\\mathrm{adv}} = L(f(x_i';\\theta_{\\mathrm{epoch - 1}}),y_i)$ \n6: $L_{\\mathrm{benign}} = L(f(x_i;\\theta_{\\mathrm{epoch - 1}}),y_i)$ \n7: $\\theta_{\\mathrm{reflect}} = \\theta_{\\mathrm{epoch - 1}} - \\eta \\cdot \\nabla_{\\theta}L_{\\mathrm{adv}}$ $\\triangleright$ Reflection step \n8: $\\Delta L_{\\mathrm{adv}} = |L(f(x_i';\\theta_{\\mathrm{reflect}}),y_i) - L_{\\mathrm{adv}}|$ \n9: $\\Delta L_{\\mathrm{benign}} = |L(f(x_i;\\theta_{\\mathrm{reflect}}),y_i) - L_{\\mathrm{benign}}|$ \n10: Reg $= \\gamma_{1}\\cdot \\Delta L_{\\mathrm{adv}} + \\gamma_{2}\\cdot \\Delta L_{\\mathrm{benign}}$ $\\triangleright$ Enforcement step \n11: $\\theta_{\\mathrm{enforce}} = \\theta_{\\mathrm{reflect}} + \\lambda \\cdot (\\mathrm{Reg} - \\theta_{\\mathrm{reflect}})$ \n12: $\\theta_{\\mathrm{epoch}} = 0.5\\cdot (\\theta_{\\mathrm{reflect}} + \\theta_{\\mathrm{enforce}})$ $\\triangleright$ DR splitting \n13: end for \n14: end for", + "bbox": [ + 215, + 167, + 785, + 393 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "we start at a lower rate of 0.01. The training spans 110 epochs, with a strategy to decrease the learning rate by a factor of 0.1 at both the 100th and 105th epochs for optimization. To prepare our model against adversarial attacks, we incorporate FAST-AT [36] and MEP-PGI [15] as our initial attack strategies for CIFAR datasets. Given the significant memory demands of MEP-PGI, particularly its need to store previous adversarial perturbations, we use its variation PGI-BP on Tiny ImageNet. We set the step-size parameter $\\gamma$ to 0.01 for CIFAR datasets, and 0.001 for Tiny ImageNet. The regularization term $\\lambda = 0.3$ for all datasets, and $\\mathcal{E} = 16 / 255$ . Experiments are run using a single GeForce RTX 3090 GPU. Find the training code and additional results in the supplementary file.", + "bbox": [ + 212, + 425, + 787, + 575 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/cfc1264e483fe9047e91342be2d3b66c9300a96063b051bbda64c7ffb3d32283.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodSA (%)RA-PGD-10\nE=8/255RA-PGD-10\nE=10/255RA-PGD-10\nE=12/255RA-PGD-10\nE=16/255SQ [2]Time (h)
AWP [37]77.9436.2633.1830.2527.4319.762.4
OAAT [1]70.5145.3239.4734.1028.9121.433.2
ExAT [29]79.3349.5643.2836.4932.5823.151.2
RAT-MEP69.8652.6446.8344.1841.2627.131.7
", + "bbox": [ + 246, + 603, + 767, + 686 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Table 1: Comparative performance across different perturbation levels on CIFAR10 using ResNet20 backbone. Our evaluation focuses on two key metrics: robust accuracy (RA) against PGD-10 attacks, and the computational time (measured in hours).", + "bbox": [ + 214, + 696, + 787, + 739 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Baselines and Evaluation Metrics: To thoroughly evaluate the adversarial robustness of our model, we use a comprehensive set of attack methods. Following [36], we use FGSM [11], and a series of PGD attacks [22] with varying iterations: PGD-10, PGD-20 and PGD-50. Additionally, we use the APGD-CE [7], C&W attack [5], and Autoattack (AA) [7]. AutoAttack is an ensemble method", + "bbox": [ + 212, + 763, + 787, + 840 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "Rethinking Fast Adversarial Training", + "bbox": [ + 480, + 114, + 730, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/0e8382f5c83ac1146237687dab5e1502285a30be6f1c489c3a75dee50366ef57.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodsSARA-FGSMRA-PGD10RA-PGD20RA-PGD50RA-C&WRA-APGD-CERA-AATime (h)
AT PGD65.7246.5140.8735.4234.2130.9733.6826.856.3
Fast-GA [3]58.4939.8833.3126.8925.0722.8124.3317.562.4
ZeroGrad [10]72.8144.2632.7822.1519.2320.8517.2012.751.3
NuAT [33]73.5544.9035.5226.3424.3124.9523.7218.931.8
Fast-AT [36]50.6438.6226.4721.6420.5519.2619.8114.971.2
RAT-AT70.6246.2838.1230.8329.3526.5427.7919.641.3
MEP-PGI [15]54.2336.8432.6527.9826.7022.6526.7319.311.5
RAT-MEP69.8647.6941.2634.8133.6428.7532.8723.911.7
", + "bbox": [ + 217, + 143, + 787, + 260 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Table 2: CIFAR10 results: We report the RA (\\%) against various attacks at $\\mathcal{E} = 16 / 255$ with ResNet20 backbone. Each model undergoes 3 training sessions. The results reflect the average performance of the top-performing model across these iterations. The best and the runner-up models are bolded and underlined, respectively.", + "bbox": [ + 212, + 270, + 787, + 327 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "that integrates APGD-CE, APGD-T (targeted APGD), and Square Attack [2], to thoroughly evaluate a model's robustness to adversarial attacks. Due to the inherent variability in the training process, which often leads to fluctuations in performance, relying on a single run may not accurately capture the effectiveness of different methods. Thus, for each method, we conduct the training process three times and report the average of the best results in these runs. The evaluation metrics include robust accuracy (RA) and standard accuracy (SA).", + "bbox": [ + 212, + 361, + 787, + 470 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.2 Results and Insights", + "text_level": 1, + "bbox": [ + 215, + 498, + 429, + 513 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "In our preliminary analysis, we highlight the importance of our results on the adversarial robustness of trained models. Our findings indicate that adversarial attacks with $\\mathcal{E} = 16 / 255$ , rarely result in the misclassification of the input. This observation is achieved through the generation of adversarial examples across CIFAR10 and CIFAR100 using both targeted and non-targeted attack methods, including FGSM, APGD-CE, APGD-T, PGD, and C&W. Remarkably, most of these adversarial examples still were classified correctly according to their original labels. Our experiments in Tab. 1 involve generating adversarial training with perturbation size $\\mathcal{E} = 16 / 255$ and testing the models' robustness across different levels of perturbation. We also compared our model (RAT) with other techniques, such as Fast-Bat, AWP, OAAT, and ExAT. We have reimplemented these AT models within the FastAT framework to ensure a fair comparison. Our model achieves superior performance even as the perturbation magnitude increases to $\\mathcal{E} = 16 / 255$ . Furthermore, our model integrates into FAST-AT and MEP-PGI, which are specifically designed to address distributions of adversarial samples. Conversely, methods like OAAT are optimized for the distribution of benign examples and yield higher SA accuracy. Adopting the settings from OAAT, we apply both OAAT and our variation $\\mathrm{OAAT}_{\\mathrm{RAT}}$ to FastAT tasks. The $\\mathrm{OAAT}_{\\mathrm{RAT}}$ achieves $+2.78\\%$ improvement on SA accuracy over OAAT ( $70.51\\% \\rightarrow 73.29$ ), and $24.66\\%$ on the SQ, as ( $21.43\\% \\rightarrow 24.66\\%$ ).", + "bbox": [ + 212, + 537, + 787, + 843 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "Zareapoor et al.", + "bbox": [ + 271, + 114, + 380, + 128 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/e82f0995fdc4924af54c6ba506c1c0abf2d1f10ad1b4b031e4b54bba4d2614f5.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodsESARA-FGSMRA-PGD10RA-PGD20RA-PGD50RA-C&WRA-APGD-CERA-AATime (h)
Fast-GA [36]1267.3943.5732.1928.0226.2427.6526.2121.752.3
NuAT [33]72.8452.3341.9538.8437.6936.0836.7532.411.8
MEP-PGI [15]73.7852.3938.6736.5235.7833.4133.6827.931.5
RAT-MEP72.8154.6245.2942.8642.2336.8841.6033.161.7
Fast-GA [3]1081.7055.7336.8030.7128.6531.2926.9124.122.3
NuAT [33]76.3956.0946.2343.9543.5041.5243.3638.991.8
MEP-PGI [15]81.9459.4843.1839.7637.8337.4636.2132.721.5
RAT-MEP76.5258.3047.5145.7445.5641.2745.3639.701.7
", + "bbox": [ + 218, + 143, + 787, + 252 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/aac752da24306eea893f051ec5db18f0346fd87cfe585ec344383de9e9a3befd.jpg", + "table_caption": [ + "Table 3: Quantitative comparison of FastAT methods on CIFAR10 using ResNet20 for different $(\\mathcal{E})$ , where both training and testing used the same $\\mathcal{E}$ . The best and the runner-up methods are bolded and underlined, respectively." + ], + "table_footnote": [], + "table_body": "
MethodsSARA-FGSMRA-PGD10RA-PGD20RA-PGD50RA-C&WRA-APGD-CERA-AATime (h)
AT PGD40.8325.1921.3817.5917.2315.0617.1012.546.3
Fast-GA [3]32.4015.8412.9810.7910.088.549.726.842.4
ZeroGrad [10]47.8124.1318.2112.9312.2411.7011.217.901.2
NuAT [33]32.4921.2716.7514.3813.7111.5212.628.751.8
Fast-AT [36]31.2215.6913.4110.7810.258.739.887.211.2
RAT-AT (ours)42.2625.8119.4315.7714.5612.8113.959.861.3
MEP-PGI [15]22.4114.0912.259.869.618.159.536.861.5
RAT-MEP (ours)43.7725.2621.3517.4217.0314.2216.7611.491.7
", + "bbox": [ + 218, + 318, + 787, + 430 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Table 4: CIFAR100 results: We report the RA (\\%) against various attacks at $\\mathcal{E} = 16 / 255$ with ResNet20 backbone. Each method undergoes three training sessions. The results reflect the average performance of the top-performing model across these iterations. The best and the runner-up methods are bolded and underlined, respectively.", + "bbox": [ + 215, + 441, + 787, + 498 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4.3 Faster DR Splitting Technique", + "text_level": 1, + "bbox": [ + 215, + 527, + 514, + 542 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Fig. 1-right provides a detailed comparison of various adversarial training on the CIFAR10 dataset, at $\\mathcal{E} = (8,16) / 255$ . At the lower perturbation level ( $\\mathcal{E} = 8 / 255$ ), all methods demonstrate similar AA and SA, with FAST-GA achieving the highest AA among them. However, increasing the perturbation to $\\mathcal{E} = 16 / 255$ results in a significant reduction in AA across all methods, highlighting the challenges of maintaining robustness at higher perturbation levels. Interestingly, our model outperforms Fast-Bat at $\\mathcal{E} = 16 / 255$ , achieving $60.37\\%$ vs. $59.34\\%$ in SA accuracy. This improvement is also evident at $\\mathcal{E} = 8 / 255$ , where RAT aligns closely with the performance of other leading methods, e.g., $46.21\\%$ AA accuracy, comparable to $46.37\\%$ , $45.83\\%$ for FAST-GA and Fast-Bat respectively. While FGSM is faster (as shown in Fig. 1-left), results in $0\\%$ AA accuracy. We also observe that our model adds very little overhead to FGSM training, i.e., 15.02 min with FGSM, but is considerably faster than other methods. The increased computation time in LLR and CURE is due to Double Backpropagation [28].", + "bbox": [ + 212, + 551, + 787, + 765 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4.4 CIFAR10 Results", + "text_level": 1, + "bbox": [ + 215, + 785, + 408, + 797 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "In this experiment, we use backbone ResNet20, to evaluate our model's performance, setting different perturbation levels $\\mathcal{E}$ . The results, in Tab.2, show that", + "bbox": [ + 215, + 809, + 785, + 839 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "Rethinking Fast Adversarial Training", + "bbox": [ + 480, + 114, + 730, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 116, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/a211521eb3d0fd8c46fe3267b95ec8155991ba4cc783e67e149f88ebc2606f05.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodsSARA-FGSMRA-PGD10RA-PGD20RA-PGD50RA-C&WRA-APGD-CERA-AATime (h)
AT PGD33.7917.6613.8210.5610.288.3610.456.4768.1
Fast-AT [36]28.1213.0610.217.366.955.486.724.5110.8
Fast-GA [3]29.2413.9811.328.458.166.237.404.9621.4
ZeroGrad [10]34.5812.859.416.735.714.225.262.8110.9
NuAT [33]35.2016.7112.489.529.377.628.215.3925.2
PGI-BP [15]21.7611.329.487.567.205.497.214.1813.8
RAT-BP (ours)34.6117.6213.239.919.458.209.876.2514.2
", + "bbox": [ + 217, + 143, + 794, + 244 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Table 5: Tiny-ImageNet results: We report the RA (\\%) against various attack methods accuracy at $\\mathcal{E} = 16 / 255$ using PreActResNet18 as our backbone. The results represent the average performance of the best models across three training iterations.", + "bbox": [ + 212, + 253, + 787, + 297 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "our model outperforms existing Fast AT techniques in adversarial robustness across various attacks. Specifically, our RAT-AT surpasses all FastAT methods, and our RAT-MEP is more effective than other methods while maintaining comparable SA accuracy. In comparison to PGD-AT, RAT-MEP achieves nearly equivalent adversarial robustness, with a marginal of 0.57 difference in performance against PGD-50 (33.64% vs. 34.21%). Regarding computational efficiency, RAT-AT is faster, completing in 75 minutes, compared to 144 minutes for Fast-GA and 108 minutes for NuAT. RAT-MEP also presents a good balance between computational time (102 minutes) and performance, taking only slightly longer than MEP-PGI (96 minutes) but significantly less than the 380 minutes required by PGD-AT. In Tab. 3, we conduct additional experiments on different perturbation scales $\\mathcal{E} = (10,12)/55$ , further validating the model's effectiveness. The results clearly demonstrate that our method consistently achieves superior adversarial robustness across different perturbation budgets.", + "bbox": [ + 212, + 327, + 787, + 540 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "4.5 CIFAR100 Results", + "text_level": 1, + "bbox": [ + 215, + 556, + 418, + 569 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Tab. 4 details the performance on CIFAR100 with a ResNet20 backbone, highlighting that the training times for FastAT methods across CIFAR10 and CIFAR100 are similar. As can be seen, our model demonstrates the ability to maintain stable adversarial training, and among the FastAT methods, RAT-AT is the most robust against different attacks. RAT-PGI surpasses the performance of all FastAT methods in terms of adversarial robustness. While RAT-MEP is slightly lower than AT-PGD, it boasts a training speed that is approximately 3.5 times faster, marking a significant improvement in efficiency.", + "bbox": [ + 212, + 582, + 787, + 704 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "4.6 ImageNet Results", + "text_level": 1, + "bbox": [ + 215, + 722, + 413, + 738 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Our experiments with the PreActResNet18 on Tiny-ImageNet, demonstrate the effectiveness of our method in handling larger and more complex dataset, as presented in Tab. 5. The RAT-BP shows comparable robustness to PGD-AT and surpasses other FastAT methods. In terms of training efficiency, RAT-BP is marginally higher, requiring 14.2 hours, compared to PGI-BP (13.8 hours), yet it is significantly faster than PGD-AT, which takes 68 hours. Further extending", + "bbox": [ + 212, + 750, + 787, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "Zareapoor et al.", + "bbox": [ + 271, + 114, + 380, + 128 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/5bf2995e52d431ad3521d35a32570494fb2dff2a9b5dd90ed3f22acce404fe52.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodsSARA-AARA-FGSMRA-PGD10RA-PGD50Time (h)
Fast-AT [36]7.411.273.423.762.2561.5
PGI-BP [15]16.211.059.482.811.8577.6
RAT-BP (ours)25.601.8814.618.703.2979.2
", + "bbox": [ + 274, + 143, + 733, + 209 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Table 6: Quantitative results of the adversarial training methods on ImageNet at $(\\mathcal{E} = 16 / 255)$ with ResNet50 as the backbone. Adding our model (RAT) helps PGI-BP overcome CO and improve its performance.", + "bbox": [ + 215, + 219, + 785, + 261 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/2b42ffc45215e7bf1b028fbe4164460bfb306abefd475b5a109e21af7f648737.jpg", + "image_caption": [ + "Fig. 3: Catastrophic Overfitting: Illustrates the robustness across perturbation strengths. The x-axis $(\\mathcal{E})$ represents the strength of the perturbation applied to test images to create adversarial examples. The y-axis is the accuracy of the model under attack (PGD-20). As $\\mathcal{E}$ increases, we expect accuracy to decrease, since stronger perturbations are harder to withstand. The AT PGD-10 is referenced for comparison, showing that our model achieves performance closely aligned with this benchmark." + ], + "image_footnote": [], + "bbox": [ + 225, + 273, + 779, + 405 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "our evaluation to the ImageNet [8], we use a ResNet50 following [36] training configuration with the only difference by the adoption of a maximum image resolution of $288 \\times 288$ , similar to [16]. The initial learning rate is 0.1, with a reduction factor of 0.1 applied after 30 and 60 epochs, across a total of 90 epochs. The results in Tab. 6 show the stability of our training across this dataset.", + "bbox": [ + 212, + 521, + 787, + 597 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "4.7 Ablation", + "text_level": 1, + "bbox": [ + 215, + 617, + 334, + 630 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Our ablation study focuses on evaluating the effectiveness of our model against leading adversarial training methods, such as Fast-GA, Fast-AT, SLAT, and Fast-Bat. We also reference the robust but computationally intensive AT PGD10. PreActResNet18 is a backbone model for these experiments, with each method's recommended hyperparameters applied. We calculate the PGD-20 test accuracy, averaging results over three random seeds to ensure reliability. Fig. 3, illustrates that SLAT encounters CO at perturbation levels of $\\mathcal{E} \\geq 4/255$ for SVHN and $\\mathcal{E}e.g.10/255$ for CIFAR10. Fast-GA shows less than optimal performance at $\\mathcal{E}e.g.16/255$ for CIFAR10 and inconsistent results on CIFAR100. Notably, Fast-Bat also experiences CO for $\\mathcal{E} > 10/255$ in CIFAR100. Overall, our RAT model significantly narrows the gap between single-step and multi-step AT methods, showing enhanced robustness and efficiency. Sensitivity to $\\lambda$ : To optimize our model, we focus on determining the optimal value for the parameter $\\mathcal{E}$ in Eq.(9)", + "bbox": [ + 212, + 640, + 802, + 839 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "Rethinking Fast Adversarial Training", + "bbox": [ + 480, + 114, + 730, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 784, + 126 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/7bb2199ce960e8a180d60b87c83905ffe1ddded7c9f3c4e0605def585f3475dd.jpg", + "image_caption": [ + "Fig. 4: Exploration of $\\lambda$ values for RAT on CIFAR10, using PRN backbone. This illustrates the relationship between the choice of $\\lambda$ and the occurrence of CO. Higher values successfully prevent CO but lead to a noticeable reduction in model performance." + ], + "image_footnote": [], + "bbox": [ + 218, + 151, + 500, + 284 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "across varying levels of adversarial perturbation. We evaluate the PGD20 adversarial accuracy on a validation set of 1024 images from the CIFAR10 dataset. The experiments are conducted using the PRN architecture, exploring a range of $\\lambda \\in \\{0.1,0.3,0.5,0.6,0.9,1\\}$ . Except for $\\lambda$ , all other parameters are maintained as per the default configuration. The objective is to identify the $\\lambda$ that yields the highest adversarial accuracy for each $\\mathcal{E}$ , as shown in Fig.4. Our results indicate a clear trend: lower $\\lambda$ and higher $\\mathcal{E}$ values, fail to prevent CO. Increasing $\\lambda$ to 0.9 eliminates CO but degrades model performance. This effect is particularly ev", + "bbox": [ + 511, + 146, + 787, + 417 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "ident at $\\lambda = 1$ , where, the PGD-20 accuracy decreases for lower and higher perturbation values. For all datasets, the model performance is almost consistent across $\\lambda \\in \\{0.3, 0.5, 0.6\\}$ , leading us to select the lowest value within this range for efficiency. On the ImageNet dataset, however, a higher value (0.7) enhances performance. To further assess the impact of hyperparameters $\\gamma_{1}$ and $\\gamma_{2}$ on model performance and adversarial robustness, we conducted experiments by varying their values and analyzing the outcomes on CIFAR100. The results are presented in Table 2 (supplementary file). The values are chosen from a set $\\{0.0, 0.3, 0.5, 0.7, 0.8, 1.0\\}$ , and the model's performance was evaluated against several adversarial attacks, including FGSM, PGD-10, PGD-[20, 50], and AA.", + "bbox": [ + 212, + 417, + 787, + 570 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 608, + 359, + 625 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "We presented an efficient model to overcome catastrophic overfitting in fast adversarial training through the Douglas-Rachford (DR) splitting technique, emphasizing linearity and stability in the training process. By using this technique, we provide a novel perspective on managing adversarial robustness that balances between minimizing adversarial loss and enforcing effective regularization. Indeed, using DR splitting allows for subtle control over the training process, facilitating adjustments in model parameters to prevent sudden degradation in adversarial robustness. This methodology not only addresses CO with greater efficiency than existing solutions like Fast-GA, Fast-AT, and NuAT but also enhances the overall training stability across a wide range of adversarial perturbations and training durations. Through extensive evaluations, with significant adversarial challenges and in conjunction with various Fast AT methods, our model demonstrates superior performance and adaptability.", + "bbox": [ + 212, + 643, + 787, + 840 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "Zareapoor et al.", + "bbox": [ + 271, + 114, + 380, + 128 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 143, + 321, + 159 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Addepalli, S., Jain, S., Sriramanan, G., Venkatesh Babu, R.: Scaling adversarial training to large perturbation bounds. In: European Conference on Computer Vision. pp. 301-316 (2022) 6, 9", + "2. Andriushchenko, M., Croce, F., Flammarion, N., Hein, M.: Square attack: a query-efficient black-box adversarial attack via random search. In: European conference on computer vision. pp. 484-501 (2020) 9, 10", + "3. Andriushchenko, M., Flammarion, N.: Understanding and improving fast adversarial training. Advances in Neural Information Processing Systems 33, 16048-16059 (2020) 2, 3, 4, 5, 6, 10, 11, 12", + "4. Applegate, D., Hinder, O., Lu, H., Lubin, M.: Faster first-order primal-dual methods for linear programming using restarts and sharpness. Mathematical Programming 201(1), 133-184 (2023) 2, 5", + "5. Carlini, N., Wagner, D.: Towards evaluating the robustness of neural networks. In: 2017 IEEE symposium on security and privacy (sp). pp. 39-57 (2017) 9", + "6. Chambolle, A., Contreras, J.P.: Accelerated bregman primal-dual methods applied to optimal transport and Wasserstein barycenter problems. SIAM Journal on Mathematics of Data Science 4(4), 1369-1395 (2022) 5", + "7. Croce, F., Hein, M.: Reliable evaluation of adversarial robustness with an ensemble of diverse parameter-free attacks. In: International conference on machine learning. pp. 2206-2216 (2020) 9", + "8. Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: Imagenet: A large-scale hierarchical image database. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 248-255 (2009) 8, 13", + "9. Dong, Y., Liao, F., Pang, T., Su, H., Zhu, J., Hu, X., Li, J.: Boosting adversarial attacks with momentum. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9185-9193 (2018) 3", + "0. Golgooni, Z., Saberi, M., Eskandar, M., Rohhan, M.H.: Zerograd: Mitigating and explaining catastrophic overfitting in fgsm adversarial training. arXiv preprint arXiv:2103.15476 (2021) 5, 10, 11, 12", + "1. Goodfellow, I.J., Shlens, J., Szegedy, C.: Explaining and harnessing adversarial examples. International Conference on Learning Representations (2014) 1, 3, 4, 9", + "2. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 770-778 (2016) 8", + "3. He, K., Zhang, X., Ren, S., Sun, J.: Identity mappings in deep residual networks. In: Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part IV 14. pp. 630-645 (2016) 8", + "4. He, Z., Li, T., Chen, S., Huang, X.: Investigating catastrophic overfitting in fast adversarial training: A self-fitting perspective. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2313-2320 (2023) 5", + "5. Jia, X., Zhang, Y., Wei, X., Wu, B., Ma, K., Wang, J., Cao, X.: Prior-guided adversarial initialization for fast adversarial training. In: European Conference on Computer Vision. pp. 567-584 (2022) 2, 6, 9, 10, 11, 12, 13", + "6. de Jorge Aranda, P., Bibi, A., Volpi, R., Sanyal, A., Torr, P., Rogez, G., Dokania, P.: Make some noise: Reliable and efficient single-step adversarial training. Advances in Neural Information Processing Systems 35, 12881-12893 (2022) 2, 5, 13", + "7. Kurakin, A., Goodfellow, I., Bengio, S.: Adversarial machine learning at scale. arXiv preprint arXiv:1611.01236 (2016) 3" + ], + "bbox": [ + 225, + 176, + 784, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "Rethinking Fast Adversarial Training", + "bbox": [ + 482, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "18. Li, T., Wu, Y., Chen, S., Fang, K., Huang, X.: Subspace adversarial training. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13409-13418 (2022) 1, 5", + "19. Liang, J., Fadili, J., Peyre, G.: Local convergence properties of douglas-rachford and alternating direction method of multipliers. Journal of Optimization Theory and Applications 172, 874-913 (2017) 2", + "20. Lindbäck, J., Wang, Z., Johansson, M.: Bringing regularized optimal transport to lightspeed: a splitting method adapted for gpus. Advances in Neural Information Processing Systems 36 (2023) 2, 5", + "21. Liu, X., Chakraborty, S., Sun, Y., Huang, F.: Rethinking adversarial policies: A generalized attack formulation and provable defense in rl. In: International Conference on Learning Representations (2024) 8", + "22. Madry, A., Makelov, A., Schmidt, L., Tsipras, D., Vladu, A.: Towards deep learning models resistant to adversarial attacks. International Conference on Learning Representations (2018) 1, 3, 9", + "23. Mai, V.V., Lindbäck, J., Johansson, M.: A fast and accurate splitting method for optimal transport: Analysis and implementation. International Conference on Learning Representations (2022) 2, 5", + "24. Moosavi-Dezfooli, S.M., Fawzi, A., Uesato, J., Frossard, P.: Robustness via curvature regularization, and vice versa. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9078-9086 (2019) 3", + "25. Park, G.Y., Lee, S.W.: Reliably fast adversarial training via latent adversarial perturbation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 7758-7767 (2021) 1, 5", + "26. Poon, C., Liang, J.: Trajectory of alternating direction method of multipliers and adaptive acceleration. Advances in Neural Information Processing Systems 32 (2019) 2, 5", + "27. Qin, C., Martens, J., Gowal, S., Krishnan, D., Dvijotham, K., Fawzi, A., De, S., Stanforth, R., Kohli, P.: Adversarial robustness through local linearization. Advances in Neural Information Processing Systems 32 (2019) 3, 5", + "28. Rocamora, E.A., Liu, F., Chrysos, G.G., Olmos, P.M., Cevher, V.: Efficient local linearity regularization to overcome catastrophic overfitting. International Conference on Learning Representations (2024) 1, 5, 6, 11", + "29. Shaeiri, A., Nobahari, R., Rohhan, M.H.: Towards deep learning models resistant to large perturbations. arXiv preprint arXiv:2003.13370 (2020) 9", + "30. Shafahi, A., Najibi, M., Ghiasi, M.A., Xu, Z., Dickerson, J., Studer, C., Davis, L.S., Taylor, G., Goldstein, T.: Adversarial training for free! Advances in Neural Information Processing Systems 32 (2019) 1, 4", + "31. Song, C., He, K., Wang, L., Hopcroft, J.E.: Improving the generalization of adversarial training with domain adaptation. International Conference on Learning Representations (2019) 1", + "32. Sriramanan, G., Addepalli, S., Baburaj, A., et al.: Guided adversarial attack for evaluating and enhancing adversarial defenses. Advances in Neural Information Processing Systems 33, 20297-20308 (2020) 5", + "33. Sriramanan, G., Addepalli, S., Baburaj, A., et al.: Towards efficient and effective adversarial training. Advances in Neural Information Processing Systems 34, 11821-11833 (2021) 2, 5, 10, 11, 12", + "34. Themelis, A., Patrinos, P.: Douglas-rachford splitting and admn for nonconvex optimization: Tight convergence results. SIAM Journal on Optimization 30(1), 149-181 (2020) 5" + ], + "bbox": [ + 215, + 146, + 785, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "Zareapoor et al.", + "bbox": [ + 271, + 114, + 380, + 128 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "35. Tsiligkaridis, T., Roberts, J.: Understanding and increasing efficiency of frank-wolfe adversarial training. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 50-59 (2022) 5", + "36. Wong, E., Rice, L., Kolter, J.Z.: Fast is better than free: Revisiting adversarial training. International Conference on Learning Representations (2020) 1, 2, 4, 6, 9, 10, 11, 12, 13", + "37. Wu, D., Xia, S.T., Wang, Y.: Adversarial weight perturbation helps robust generalization. Advances in Neural Information Processing Systems (2020) 9", + "38. Xie, Y., Li, Z., Shi, C., Liu, J., Chen, Y., Yuan, B.: Enabling fast and universal audio adversarial attack using generative model. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 35, pp. 14129-14137 (2021) 1", + "39. Zagoruyko, S., Komodakis, N.: Wide residual networks. arXiv preprint arXiv:1605.07146 (2016) 8", + "40. Zhang, H., Yu, Y., Jiao, J., Xing, E., El Ghaoui, L., Jordan, M.: Theoretically principled trade-off between robustness and accuracy. In: International conference on machine learning. pp. 7472-7482 (2019) 3", + "41. Zhang, Y., Zhang, G., Khanduri, P., Hong, M., Chang, S., Liu, S.: Revisiting and advancing fast adversarial training through the lens of bi-level optimization. In: International Conference on Machine Learning. pp. 26693-26712 (2022) 3, 5", + "42. Zhao, M., Zhang, L., Kong, Y., Yin, B.: Fast adversarial training with smooth convergence. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 4720-4729 (2023) 2, 6" + ], + "bbox": [ + 212, + 146, + 787, + 452 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "Rethinking Fast Adversarial Training", + "bbox": [ + 482, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 16 + } +] \ No newline at end of file diff --git a/2024/Rethinking Fast Adversarial Training_ A Splitting Technique To Overcome Catastrophic Overfitting/54a9d439-23cc-4380-af56-e12fd67a1f40_model.json b/2024/Rethinking Fast Adversarial Training_ A Splitting Technique To Overcome Catastrophic Overfitting/54a9d439-23cc-4380-af56-e12fd67a1f40_model.json new file mode 100644 index 0000000000000000000000000000000000000000..ef31990347463590448c4c08a2300fd952dd5b0d --- /dev/null +++ b/2024/Rethinking Fast Adversarial Training_ A Splitting Technique To Overcome Catastrophic Overfitting/54a9d439-23cc-4380-af56-e12fd67a1f40_model.json @@ -0,0 +1,1972 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.219, + 0.141, + 0.784, + 0.187 + ], + "angle": 0, + "content": "Rethinking Fast Adversarial Training: A Splitting Technique to Overcome Catastrophic Overfitting" + }, + { + "type": "text", + "bbox": [ + 0.31, + 0.213, + 0.694, + 0.228 + ], + "angle": 0, + "content": "Masoumeh Zareapoor and Pourya Shamsolmoali\\*" + }, + { + "type": "text", + "bbox": [ + 0.308, + 0.24, + 0.697, + 0.255 + ], + "angle": 0, + "content": "East China Normal University, Queen's University Belfast" + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.289, + 0.744, + 0.555 + ], + "angle": 0, + "content": "Abstract. Catastrophic overfitting (CO) poses a significant challenge to fast adversarial training (FastAT), particularly at large perturbation scales, leading to dramatic reductions in adversarial test accuracy. Our analysis of existing FastAT methods shows that CO is accompanied by abrupt and irregular fluctuations in loss convergence, indicating that a stable training dynamic is key to preventing CO. Therefore, we propose a training model that uses the Douglas-Rachford (DR) splitting technique to ensure a balanced and consistent training progression, effectively counteracting CO. The DR splitting technique, known for its ability to solve complex optimization problems, offering a distinct advantage over classical FastAT methods by providing a smoother loss convergence. This is achieved without resorting to complex regularization or incurring the computational costs associated with double backpropagation, presenting an efficient solution to enhance adversarial robustness. Our comprehensive evaluation conducted across standard datasets, demonstrates that our DR splitting-based model not only improves adversarial robustness but also achieves this with remarkable efficiency compared to various FastAT methods. This efficiency is particularly observed under conditions involving long training schedules and large adversarial perturbations." + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.566, + 0.743, + 0.582 + ], + "angle": 0, + "content": "Keywords: Adversarial training \\(\\cdot\\) Adversarial robustness \\(\\cdot\\) Convergence" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.605, + 0.378, + 0.621 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.636, + 0.788, + 0.819 + ], + "angle": 0, + "content": "Deep neural networks (DNNs), despite their successes, are vulnerable to adversarial attacks, where small, undetectable, input perturbations can mislead models into incorrect predictions [11]. Given the security risks of such vulnerabilities, improving neural network robustness against these attacks has become a critical area of research [28, 38]. Adversarial training (AT) [18, 30] and its derivative, PGD-AT [22, 31], have emerged as effective strategies for enhancing the robustness of neural networks. However, the complex inner optimizations required by these methods at each training step significantly elevate computational costs, limiting their practicality for large-scale applications [28]. To address these computational challenges, 'fast' versions of AT (FastAT) [25, 36] have been proposed, which employ a single-step gradient propagation [11] to produce adversarial training examples. While FastAT offers a more efficient training process" + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.825, + 0.788, + 0.841 + ], + "angle": 0, + "content": "* Correspondence to pshams55@gmail.com. Code: https://github.com/mzarea2/RAT" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.381, + 0.129 + ], + "angle": 0, + "content": "Zareapoor et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.344 + ], + "angle": 0, + "content": "compared to PGD-AT, it introduces the problem of Catastrophic Overfitting (CO) [16,36]. CO occurs when training with strong adversaries, leading to a significant drop in the robustness of the resulting model. Several FastAT methods have been developed to reduce CO [42]. Wong et al. [36] enhance adversarial example diversity through the use of initial perturbations that are randomly initialized. Following this approach, [3] proposed FAST-GA, an efficient regularizer that aims to explicitly align gradients between benign and adversarial inputs. Likewise, MEP-PGI [15] and NuAT [33] employ nuclear or p-norm regularization within adversarial training to improve consistency in predictions on both benign and adversarial examples. These approaches are effective only within a limited perturbation \\(\\mathcal{E} \\leq 8/255\\), where \\(\\mathcal{E}\\) denotes the degree of perturbations in the adversarial training [42]. Moreover, models trained with relatively small perturbation scales show vulnerability to larger \\(\\mathcal{E}\\), with noted declines in robustness." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.344, + 0.788, + 0.479 + ], + "angle": 0, + "content": "For example, models trained with NuAT and MEP-PGI at a perturbation size of \\(\\mathcal{E} = 8 / 255\\) demonstrated robustness levels of \\(56\\%\\) and \\(55\\%\\) against PGD-50 attacks, but their robustness significantly dropped to \\(24\\%\\) and \\(22\\%\\), respectively, with \\(\\mathcal{E} = 16 / 255\\). Our investigation into adversarial training processes (illustrated in Fig. 2) reveals that CO is characterized by minor fluctuations in benign classification loss and significant decreases in adversarial example loss. This observation raises a question: How can we design fast adversarial training that ensures improved stability and mitigates CO, while also optimizing the trade-off between robustness accuracy (RA) and standard accuracy (SA)?" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.48, + 0.788, + 0.829 + ], + "angle": 0, + "content": "In this context, the Douglas-Rachford (DR) splitting technique emerges as a promising solution. The DR splitting is known for its linear convergence across both strongly convex and non-convex settings [4, 20, 23, 26]. Several studies demonstrate its applicability in a variety of optimization tasks. For example, Liang et al. [19] and Applegate et al. [4] have highlighted its capability for local and global linear convergence, underscoring the DR technique's effectiveness in complex optimization scenarios. We also capitalize on the inherent stability and efficiency of the DR splitting, incorporating these properties into adversarial training to address CO. This integration allows for a more stable and consistent adversarial learning process, significantly lowering the risk of CO. Indeed, our implementation demonstrates that DR splitting not only stabilizes the training process but also offers a cost-effective solution to enhance model robustness against adversarial attacks, as shown in Fig. 1. Our contributions are threefold: ① We provide a comprehensive analysis showing that existing FastAT methods are still prone to CO with larger \\(\\varepsilon\\). Our investigation emphasizes the critical role of loss function convergence stability in addressing this vulnerability. ② Introducing a novel solution with the Douglas-Rachford splitting technique, specifically designed to mitigate CO without incurring computational complexities. ③ Conduct extensive experiments on standard datasets to validate the effectiveness of our model- Rethinking fast Adversarial Training (RAT)- particularly against large adversarial perturbations. RAT, leveraging the DR-splitting technique, shows enhanced adversarial robustness when compared to classical adversarial training methods (find the training code in the supplementary file)." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.482, + 0.115, + 0.733, + 0.13 + ], + "angle": 0, + "content": "Rethinking Fast Adversarial Training" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.149, + 0.401, + 0.258 + ], + "angle": 0, + "content": null + }, + { + "type": "table", + "bbox": [ + 0.409, + 0.144, + 0.78, + 0.248 + ], + "angle": 0, + "content": "
MethodsAA (%)SA (%)AA (%)SA (%)
816
AT PGD-1047.69±(0.11)78.51±(0.08)24.97±(0.26)59.41±(0.46)
LLR [27]43.84±(0.20)74.52±(0.07)16.31±(0.20)44.76±(4.72)
Fast-Bat [41]45.83±(0.12)79.61±(0.12)18.30±(0.26)59.34±(0.25)
Fast-GA [3]46.37±(0.27)79.12±(0.07)15.36±(1.32)54.25±(6.28)
CURE [24]44.65±(0.17)77.04±(0.11)18.12±(0.35)53.49±(0.08)
RAT (Ours)46.21±(0.08)80.32±(0.12)19.58±(0.17)60.37±(0.21)
" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.273, + 0.788, + 0.331 + ], + "angle": 0, + "content": "Fig. 1: Performance comparison of adversarial robustness with WideResNet34-10 on CIFAR10 for \\(\\mathcal{E} = (8,16) / 255\\). (left) Illustrates the average runtime for each method. (right) We compare AA (AutoAttack) and SA accuracies. Lower AA indicates higher vulnerability. Higher SA (standard accuracy) indicates better general performance." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.355, + 0.388, + 0.372 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.384, + 0.427, + 0.399 + ], + "angle": 0, + "content": "2.1 Adversarial Attacks" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.407, + 0.784, + 0.499 + ], + "angle": 0, + "content": "Adversarial attacks aim to fool deep learning models by introducing perturbed inputs. Goodfellow et al. [11] introduced the concept of the fast gradient sign method (FGSM) for generating adversarial examples \\( x' \\) within classification tasks. They demonstrated that adversarial examples, created through a single-step gradient update, can lead a classification model to incorrect predictions with significant confidence. The adversarial examples \\( x' \\) are generated by" + }, + { + "type": "equation", + "bbox": [ + 0.386, + 0.506, + 0.786, + 0.524 + ], + "angle": 0, + "content": "\\[\nx ^ {\\prime} = x + \\mathcal {E}. \\operatorname {s g n} \\left(\\nabla_ {x} L (f (x; \\theta), y\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.531, + 0.788, + 0.651 + ], + "angle": 0, + "content": "where \\((x,y)\\) denotes an input sample and its corresponding label, \\(\\mathcal{E}\\) is the perturbation magnitude, \\(\\nabla_{x}L\\) is the loss gradient with respect to \\(x\\), and \\(\\theta\\) signifies the fixed model weights. Subsequent research has developed more sophisticated attack methods, such as I-FGSM [17], MIM [9], and PGD [22], which leverage iterative gradient descent to enhance the efficacy of the attacks. The PGD attack, in particular, generates adversarial examples by applying small perturbations to the input data, ensuring these adjustments stay within a pre-defined boundary \\(\\mathcal{E}\\). This process is described as follows" + }, + { + "type": "equation", + "bbox": [ + 0.351, + 0.66, + 0.787, + 0.679 + ], + "angle": 0, + "content": "\\[\nx _ {t + 1} ^ {\\prime} = \\Pi_ {\\mathcal {E}} (x _ {t} ^ {\\prime} + \\alpha . \\mathrm {s g n} (\\nabla_ {x} L (f (x _ {t} ^ {\\prime}; \\theta), y))), \\qquad \\qquad (2)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.686, + 0.788, + 0.747 + ], + "angle": 0, + "content": "in which \\( x' \\) is the adversarial example at iteration \\( t \\), and \\( \\Pi_{\\mathcal{E}} \\) denotes the projection operation that ensures the adversarial example remains within the \\( \\mathcal{E} \\)-ball. The PGD attack is one of the most effective attack methods for evaluating the robustness of deep learning against adversarial examples." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.757, + 0.479, + 0.773 + ], + "angle": 0, + "content": "2.2 Adversarial Training (AT)" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.786, + 0.811 + ], + "angle": 0, + "content": "Adversarial Training (AT) [22] and TRADES [40] have become prominent training methods for training robust architectures, which can be written as" + }, + { + "type": "equation", + "bbox": [ + 0.359, + 0.819, + 0.786, + 0.843 + ], + "angle": 0, + "content": "\\[\n\\min _ {\\theta} E _ {(x, y) \\sim D} [ \\max _ {\\delta \\in [ - \\mathcal {E}, \\mathcal {E} ]} L (f (x + \\delta ; \\theta), y) ], \\tag {3}\n\\]" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.117, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.116, + 0.38, + 0.129 + ], + "angle": 0, + "content": "Zareapoor et al." + }, + { + "type": "image", + "bbox": [ + 0.235, + 0.146, + 0.77, + 0.397 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.408, + 0.788, + 0.493 + ], + "angle": 0, + "content": "Fig. 2: Performance of fast adversarial training (FastAT) methods on CIFAR10 using ResNet20, for \\(\\mathcal{E} = 16 / 255\\). The figure illustrates the training process across three runs. The terms AV-loss and BN-loss are the training loss for adversarial and benign inputs, respectively, while AV-acc and BN-acc quantify testing accuracy against both adversarial and benign samples. The occurrence of catastrophic overfitting is marked (in green), indicating critical periods where adversarial robustness drops." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.523, + 0.789, + 0.737 + ], + "angle": 0, + "content": "where \\(\\delta\\) is the adversarial perturbations and \\(x' = x + \\delta\\) is the adversarial examples generated by various attacks such as PGD and FGSM. This optimization first, generates adversarial examples within a defined perturbation limit to maximize classification loss (inner maximization) and then adjusts the model parameters \\(\\theta\\) to minimize this loss over the adversarial examples (outer minimization). However, a notable trade-off emerges in AT between computational efficiency and adversarial robustness, due to the complexity of solving the inner maximization problem. To enhance the computational efficiency of AT, several methods have been developed. Free AT [30], simultaneously updating the network parameters \\(\\theta\\) and the adversarial perturbation \\(\\delta\\) multiple times within a single batch. On the other hand, Fast-AT [36], simplifies the process further by employing a single-step FGSM attack [11] but the attack initiates from a randomly chosen point \\(x + \\mu\\) with \\(\\mu \\in [-\\delta, \\delta]\\). Fast-Bat These approaches reduce the cost of solving the inner maximization problem, significantly accelerating the training process." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.751, + 0.513, + 0.768 + ], + "angle": 0, + "content": "2.3 Catastrophic Overfitting (CO)" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.842 + ], + "angle": 0, + "content": "The acceleration of AT methods comes with significant drawbacks. Wong et al. [36] first reported this with FGSM training, and was later supported by Andriushchenko et al. [3] in their analysis of Free AT and Fast-AT methods. In these cases, the model's accuracy on adversarial tests can fall sharply, sometimes" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.483, + 0.115, + 0.732, + 0.13 + ], + "angle": 0, + "content": "Rethinking Fast Adversarial Training" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.209 + ], + "angle": 0, + "content": "to \\(0\\%\\), in just a few training epochs. This decline in performance has been attributed to CO, a critical challenge that has motivated significant research efforts aimed at understanding and mitigating its impact [14, 18, 41]. Below, we provide an overview of various strategies developed to avoid CO." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.228, + 0.536, + 0.243 + ], + "angle": 0, + "content": "2.4 Regularization-based Approaches" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.259, + 0.789, + 0.487 + ], + "angle": 0, + "content": "Recent efforts focus on regularization and data augmentation to prevent CO. [3] introduced FAST-GA to mitigate CO by penalizing the misalignment of gradients between the input and their adversarial counterparts. This method ensures that the direction of the gradient remains consistent when slight perturbations are introduced to the input. Qin et al. [27] introduced local linear regularization (LLR) to improve model robustness against adversarial attacks. LLR smoothes the loss landscape using a first-order Taylor approximation, addressing nonlinear sensitivities and promoting a more linear loss response to perturbations. ELLE [28] introduced a regularization term designed to mitigate CO both effectively and efficiently within classical AT frameworks, by enforcing local linearity. Sriramanan et al. [32] introduced GAT, a novel regularization strategy aimed at reducing the difference between the logits of a training point and a random sample. NuAT [33], proposed a regularization term that penalizes the nuclear norm of the logits difference between benign and adversarial examples. However, GAT and NuAT still face challenges with CO under larger perturbation size \\(\\mathcal{E}\\) [16]." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.507, + 0.414, + 0.523 + ], + "angle": 0, + "content": "2.5 Other Approaches" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.538, + 0.789, + 0.844 + ], + "angle": 0, + "content": "Park et al. [25] introduced SLAT which performs adversarial attacks within the latent space, focusing on the internal representations rather than the input space directly. Another study [35] employs the Frank-Wolfe optimization technique for AT, connecting the distortion observed in the loss landscape to the norm of adversarial perturbations, providing a deeper understanding of adversarial robustness. Li et al. [18] addressed CO by limiting network weights to a specific subspace based on the main components of AT-trained weights. Meanwhile, Zhang et al. [41] introduce Fast-Bat, which redefines the AT challenge as a Bilevel optimization problem and introduces a one-step solution to address it. However, [28] shows that Fast-BAT is affected by CO for \\(\\mathcal{E} > 10 / 255\\). Zero-Grad [10] introduces a strategy to mitigate adversarial perturbations by setting to zero the gradient components that fall below a threshold, however, [16] demonstrates, this approach results in decreased performance for large \\(\\mathcal{E}\\). Different from these methods, we aim to minimize the loss fluctuations during training through the Douglas-Rachford splitting technique, known for its effectiveness in complex optimization [23]. It has been successfully applied to optimal transport [20] and its discretizations [6], demonstrating its flexibility and robustness in achieving linear convergence rates under both convex and non-convex settings [4, 34]. A notable advantage of this splitting technique is to leverage sparsity in solutions, which can significantly accelerate the convergence process [26]." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.381, + 0.129 + ], + "angle": 0, + "content": "Zareapoor et al." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.145, + 0.331, + 0.161 + ], + "angle": 0, + "content": "3 Method" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.175, + 0.784, + 0.207 + ], + "angle": 0, + "content": "This section evaluates FastAT methods under large perturbation budgets, \\(\\mathcal{E}\\), explores the causes of CO, and finally introduces our proposed solutions." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.22, + 0.715, + 0.236 + ], + "angle": 0, + "content": "3.1 Performance of Fast Adversarial Training on a Large \\(\\mathcal{E}\\)" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.244, + 0.785, + 0.32 + ], + "angle": 0, + "content": "Fast AT methods have shown effectiveness in mitigating CO at lower perturbation \\((\\mathcal{E} \\leq 8/255)\\) [28, 36]. To extend this evaluation, we analyze performance at a higher perturbation level \\((\\mathcal{E} = 16/255)\\), in line with [42]. FAST-AT [36] introduces an initial step for Eq. (3), where samples are perturbed with uniform random noise \\(\\delta_0 \\sim U(-\\mathcal{E}, \\mathcal{E})\\) before applying the FGSM attack," + }, + { + "type": "equation", + "bbox": [ + 0.318, + 0.329, + 0.786, + 0.351 + ], + "angle": 0, + "content": "\\[\n\\min _ {\\theta} E _ {(x, y) \\sim D} \\left[ \\max _ {\\delta_ {0} + \\delta \\in [ - \\mathcal {E}, \\mathcal {E} ]} L (f (x + \\delta_ {0} + \\delta ; \\theta), y) \\right]. \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.36, + 0.788, + 0.422 + ], + "angle": 0, + "content": "Here, \\( L(\\cdot) \\) is the loss function, \\( \\delta \\) represents adversarial perturbations, and \\( \\theta \\) is the model weights. FAST-GA, as introduced by [3] focuses on aligning the gradients of the loss with respect to both benign samples \\( x \\) and their perturbed samples \\( x + \\delta_0 \\), which is described as," + }, + { + "type": "equation", + "bbox": [ + 0.331, + 0.432, + 0.785, + 0.448 + ], + "angle": 0, + "content": "\\[\nE _ {(x, y) \\sim D} [ 1 - \\cos (\\nabla_ {x} L (x, \\theta), \\nabla_ {x + \\delta_ {0}} L (x + \\delta_ {0}, \\theta)) ], \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.457, + 0.788, + 0.579 + ], + "angle": 0, + "content": "in which \\( L(x,\\theta) = L(f(x;\\theta),y) \\). The cosine similarity \\( \\cos(\\cdot) \\) measures the cosine similarity between the gradients of the loss function for the benign and perturbed inputs. This strategy aims to ensure that the model's behavior does not change drastically for adversarial examples, thereby reducing CO. Different from other attack methods, MEP-PGI [15] extends the FGSM attack by incorporating a Prior-Guided Initialization (PGI) mechanism. This technique initializes the adversarial perturbations \\( \\delta_0 \\), by using the information from adversarial examples generated in previous training epochs," + }, + { + "type": "equation", + "bbox": [ + 0.376, + 0.589, + 0.785, + 0.605 + ], + "angle": 0, + "content": "\\[\nE _ {(x, y) \\sim D} [ \\| f (x ^ {\\prime}; \\theta) - f (x + \\delta_ {0}; \\theta) \\| _ {2} ^ {2} ], \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.614, + 0.789, + 0.844 + ], + "angle": 0, + "content": "here, \\( ||\\cdot|| \\) is the squared (L2) norm, measuring the distance between the model's outputs for the two types of examples. FAST-PGI aims to enrich the adversarial examples with historical insight, thereby enhancing the adversarial training process. Fig. 2 illustrates the adversarial training of ResNet20 on CIFAR10, highlighting times when CO occurs, between epochs 5th-20th, 40th-65th, and 10th-15th. The key observations from the results are i) transition from smooth/stable to erratic loss fluctuations; ii) discrepancies in the loss for benign and adversarial examples, i.e., minor fluctuations in benign loss \\( L(x,\\theta) \\) followed with a rapid decline in adversarial loss \\( L(x',\\theta) \\); and iii) a sharp decrease in adversarial accuracy. The result is consistent with the observations in [1,28,42]. From these patterns, we derive several insights for understanding CO: it is obvious that \\( L(x,\\theta) \\) has greater stability than the loss \\( L(x',\\theta) \\), indicating a tendency of models to overfit to adversarial samples. CO directly correlates with the instability in adversarial training convergence. Anomalies in both benign and adversarial loss occur together, signifying an issue in the learning process that affects its performance" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.482, + 0.115, + 0.733, + 0.13 + ], + "angle": 0, + "content": "Rethinking Fast Adversarial Training" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.193 + ], + "angle": 0, + "content": "on both data. We also find that Fast-GA and MEP-PGI enhance the stability of adversarial training through the constraints in Eq. (4) and Eq. (5), compared to the Fast-AT. We are now in shape to give a formal definition of our model." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.21, + 0.74, + 0.226 + ], + "angle": 0, + "content": "3.2 Douglas-Rachford (DR) Splitting for Adversarial Training" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.231, + 0.785, + 0.414 + ], + "angle": 0, + "content": "We use the DR splitting technique to ensure a consistent and stable optimization to prevent sudden changes in adversarial robustness. This technique divides the optimization process into a two-step process (reflection and enforcement), which together ensure a stable and consistent optimization path. Let's denote the model at iteration \\( k \\) by \\( f(\\cdot, \\theta^{(k)}) \\), where \\( \\theta^{(k)} \\) represents the model parameters. The loss function for adversarial examples is \\( L(f(x', \\theta^{(k)}, y)) \\), and for benign examples, \\( L(f(x^{k}, \\theta^{(k)}, y)) \\). An important property of our model is its ability to alternate between reducing adversarial loss and applying careful regularization terms. This ensures that any adjustments to the model parameters effectively impact both benign and adversarial loss functions \\( L(x, \\theta) \\) and \\( L(x', \\theta) \\). Specifically, to address CO, we introduce epoch-based constraints on the loss differences for both adversarial and benign instances, represented as \\( (\\Delta L_{\\theta^{(k)}}, \\Delta L_{\\theta^{(k)}}^{benign}) \\)," + }, + { + "type": "equation", + "bbox": [ + 0.31, + 0.42, + 0.785, + 0.461 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\Delta L _ {\\theta^ {(k)}} = \\left| L \\left(f \\left(x ^ {\\prime (k)}, \\theta^ {(k)}\\right), y\\right) - L \\left(f \\left(x ^ {\\prime (k - 1)}, \\theta^ {(k - 1)}\\right), y\\right) \\right| \\\\ \\begin{array}{l} \\left. \\right| = \\left| L (f (w, \\theta^ {(k)}), y) - L (f (x, \\theta^ {(k - 1)}), y) \\right|, \\end{array} \\tag {7} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.466, + 0.785, + 0.616 + ], + "angle": 0, + "content": "where \\( |.| \\) measures the absolute difference in loss for adversarial \\( (x') \\) and benign \\( (x) \\) inputs, \\( x'^{(k)} = x + \\delta_{0,k} + \\delta_k \\). Here, \\( \\delta_{0,k} \\) can be replaced by various attack initialization perturbations and \\( \\delta_k \\) is generated by \\( [\\max_{\\delta_k \\in [-\\varepsilon, \\varepsilon]} L(x'^{(k)}, \\theta^{(k-1)})] \\). This formulation guides the model to effectively manage perturbations and maintain a stable training process. Additionally, these constraints help in mitigating the simultaneous anomalies in both \\( L(x, \\theta) \\) and \\( L(x', \\theta) \\) facilitating a more controlled update of the model parameters that considers both adversarial and benign inputs. In practice, this involves updating the model's intermediate parameters, \\( C^k \\), to reduce the adversarial loss, according to the gradient of the loss function \\( \\nabla_\\theta \\) with respect to the adversarial examples \\( x'^{(k)} \\)," + }, + { + "type": "equation", + "bbox": [ + 0.377, + 0.623, + 0.785, + 0.641 + ], + "angle": 0, + "content": "\\[\nC ^ {k} = \\theta^ {(k)} - \\eta \\nabla_ {\\theta} L \\left(f \\left(x ^ {\\prime (k)}, \\theta^ {(k)}\\right), y\\right), \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.646, + 0.784, + 0.766 + ], + "angle": 0, + "content": "where \\(\\eta\\) is a learning rate. This update, termed as the 'reflection step' (\\(\\theta_{\\mathrm{ref}}\\)), is pivotal in aligning the model's parameters towards reducing adversarial vulnerability. We then combine these updates \\(C^k\\) with regularization adjustments to stabilize the training dynamics and enhance the adversarial robustness, effectively preventing overfitting to adversarial examples while maintaining performance on benign inputs. The updated model parameters, \\(\\theta^{(k + 1)}\\), allow to handle both adversarial and benign inputs effectively without compromising training stability or performance," + }, + { + "type": "equation", + "bbox": [ + 0.382, + 0.767, + 0.785, + 0.784 + ], + "angle": 0, + "content": "\\[\n\\theta^ {(k + 1)} = C ^ {k} + \\lambda (R e g (\\theta^ {(k)}) - C ^ {k}), \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.786, + 0.784, + 0.817 + ], + "angle": 0, + "content": "where \\(\\lambda \\in [0,1]\\) controls the strength of the regularization, and Reg involves components in Eq. (7) such that" + }, + { + "type": "equation", + "bbox": [ + 0.361, + 0.824, + 0.785, + 0.843 + ], + "angle": 0, + "content": "\\[\nR e g \\left(\\theta^ {(k)}\\right) = \\gamma_ {1} \\left(\\Delta L _ {\\theta^ {(k)}}\\right) + \\gamma_ {2} \\left(\\Delta L _ {\\theta^ {(k)}} ^ {b e n i g n}\\right). \\tag {10}\n\\]" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.381, + 0.129 + ], + "angle": 0, + "content": "Zareapoor et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.146, + 0.788, + 0.451 + ], + "angle": 0, + "content": "The important reason behind Eq. (9) is that the regularization term \\( Reg(\\theta^{(k)}) \\) incorporates changes in loss for both adversarial and benign examples, weighted by \\( \\gamma_{1} \\) and \\( \\gamma_{2} \\), respectively. This helps to balance the model's performance across adversarial and benign inputs, addressing CO by preventing the model from becoming overly specialized on the adversarial examples. Because CO occurs when a model overfits to specific adversarial examples during training, compromising its ability to generalize [21]. Our formulation ensures that the model's updates consider the impact on both adversarial robustness and benign accuracy. This regularization can be seen as part of the 'enforcement step' \\( (\\theta_{enf}) \\). The DR technique involves iteratively updating the model parameters by combining the adversarial loss minimization and the stability enforcement components, as \\( \\theta = 0.5 \\times (\\theta_{ref} + \\theta_{enf}) \\). \\( \\theta_{ref} \\) targets reducing the loss on adversariably perturbed inputs Eq. (8), conversely, \\( \\theta_{enf} \\) represents the parameters adjusted after applying the regularization term, focusing on stability and CO prevention Eq. (9). This approach under the DR-splitting technique provides a comprehensive strategy for addressing CO in the adversarial training process. Indeed, this allows the model to effectively adapt to adversarial perturbations while ensuring that such adaptations do not compromise the model's ability to accurately classify benign inputs or lead to overfitting on the adversarial examples during training. Our proposed RAT is given in Algorithm 1, discussion in the supplementary file." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.456, + 0.79, + 0.611 + ], + "angle": 0, + "content": "Efficient Computation of Loss Differences In practice, directly computing the losses \\( L(f(x^{\\prime (k - 1)},\\theta^{(k - 1)},y)) \\) and \\( L(f(x,\\theta^{(k - 1)}),y) \\) in Eq. (7) can consume a significant amount of memory. To address this, we introduce \\( u^{k - 1} \\) for benign inputs and \\( u^{\\prime k - 1} \\) for adversarial inputs, replacements for these terms, allowing for efficient computation and tracking of the average loss over benign and adversarial examples, respectively. Specially, \\( u^{k - 1} = \\frac{1}{N}\\sum_{i}L(f(x_{i}^{\\prime},\\theta^{(k - 1)}),y_{i}) \\) represents the average loss for adversarial inputs, and \\( u^{\\prime k - 1} = \\frac{1}{N}\\sum_{i = 1}^{N}L(f(x_i,\\theta^{(k - 1)}),y_i) \\) denotes the average loss for benign inputs, where \\( N \\) is the batch size. This can reduce computational and memory overhead by focusing on the aggregate loss across batches, rather than calculating them individually for each sample." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.633, + 0.462, + 0.65 + ], + "angle": 0, + "content": "4 Experimental Results" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.665, + 0.444, + 0.681 + ], + "angle": 0, + "content": "4.1 Experimental settings" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.788, + 0.842 + ], + "angle": 0, + "content": "In this section, we evaluate the proposed model across four benchmark datasets to determine its effectiveness and robustness against adversarial attacks. For both CIFAR10 [8] and CIFAR100 [8], we use the ResNet20 [12] architecture as the backbone, recognizing its strong performance in image classification tasks. For the more complex Tiny-ImageNet and ImageNet [8], we use the PreActResNet18 [13] to better handle its challenges. Additionally, the use of WideResNet34-10 [39] highlights our model's capacity to maintain effectiveness on long training schedules. We use the SGD optimizer with a batch size of 128, a weight decay of \\(5.10^{-4}\\), and a momentum of 0.9. We initiate training with learning rates of 0.1 for CIFAR10 and CIFAR100. For Tiny ImageNet, due to its greater complexity," + } + ], + [ + { + "type": "header", + "bbox": [ + 0.482, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Rethinking Fast Adversarial Training" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "code_caption", + "bbox": [ + 0.217, + 0.149, + 0.782, + 0.166 + ], + "angle": 0, + "content": "Algorithm 1 Rethinking fast Adversarial Training (RAT) using DR Splitting" + }, + { + "type": "algorithm", + "bbox": [ + 0.217, + 0.169, + 0.787, + 0.395 + ], + "angle": 0, + "content": "Require: Training dataset \\((x_{i},y_{i})_{i = 1}^{N}\\) # epochs K, # batch B, perturbation budget \\(\\mathcal{E},\\lambda ,\\gamma_1,\\gamma_2\\) \n1: Initialize model parameters \\(\\theta_0\\) \n2: for epoch \\(= 1\\) to K do \n3: for \\(i = 1\\) to B do \n4: \\(x_{i}^{\\prime} = x_{i} + \\delta \\mathrm{sign}(\\nabla_{x_{i}}L(f(x_{i};\\theta_{\\mathrm{epoch - 1}}),y_{i}))\\triangleright\\) Adversarial example generation \n5: \\(L_{\\mathrm{adv}} = L(f(x_i';\\theta_{\\mathrm{epoch - 1}}),y_i)\\) \n6: \\(L_{\\mathrm{benign}} = L(f(x_i;\\theta_{\\mathrm{epoch - 1}}),y_i)\\) \n7: \\(\\theta_{\\mathrm{reflect}} = \\theta_{\\mathrm{epoch - 1}} - \\eta \\cdot \\nabla_{\\theta}L_{\\mathrm{adv}}\\) \\(\\triangleright\\) Reflection step \n8: \\(\\Delta L_{\\mathrm{adv}} = |L(f(x_i';\\theta_{\\mathrm{reflect}}),y_i) - L_{\\mathrm{adv}}|\\) \n9: \\(\\Delta L_{\\mathrm{benign}} = |L(f(x_i;\\theta_{\\mathrm{reflect}}),y_i) - L_{\\mathrm{benign}}|\\) \n10: Reg \\(= \\gamma_{1}\\cdot \\Delta L_{\\mathrm{adv}} + \\gamma_{2}\\cdot \\Delta L_{\\mathrm{benign}}\\) \\(\\triangleright\\) Enforcement step \n11: \\(\\theta_{\\mathrm{enforce}} = \\theta_{\\mathrm{reflect}} + \\lambda \\cdot (\\mathrm{Reg} - \\theta_{\\mathrm{reflect}})\\) \n12: \\(\\theta_{\\mathrm{epoch}} = 0.5\\cdot (\\theta_{\\mathrm{reflect}} + \\theta_{\\mathrm{enforce}})\\) \\(\\triangleright\\) DR splitting \n13: end for \n14: end for" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.426, + 0.788, + 0.577 + ], + "angle": 0, + "content": "we start at a lower rate of 0.01. The training spans 110 epochs, with a strategy to decrease the learning rate by a factor of 0.1 at both the 100th and 105th epochs for optimization. To prepare our model against adversarial attacks, we incorporate FAST-AT [36] and MEP-PGI [15] as our initial attack strategies for CIFAR datasets. Given the significant memory demands of MEP-PGI, particularly its need to store previous adversarial perturbations, we use its variation PGI-BP on Tiny ImageNet. We set the step-size parameter \\(\\gamma\\) to 0.01 for CIFAR datasets, and 0.001 for Tiny ImageNet. The regularization term \\(\\lambda = 0.3\\) for all datasets, and \\(\\mathcal{E} = 16 / 255\\). Experiments are run using a single GeForce RTX 3090 GPU. Find the training code and additional results in the supplementary file." + }, + { + "type": "table", + "bbox": [ + 0.248, + 0.604, + 0.769, + 0.687 + ], + "angle": 0, + "content": "
MethodSA (%)RA-PGD-10\nE=8/255RA-PGD-10\nE=10/255RA-PGD-10\nE=12/255RA-PGD-10\nE=16/255SQ [2]Time (h)
AWP [37]77.9436.2633.1830.2527.4319.762.4
OAAT [1]70.5145.3239.4734.1028.9121.433.2
ExAT [29]79.3349.5643.2836.4932.5823.151.2
RAT-MEP69.8652.6446.8344.1841.2627.131.7
" + }, + { + "type": "table_caption", + "bbox": [ + 0.215, + 0.697, + 0.788, + 0.74 + ], + "angle": 0, + "content": "Table 1: Comparative performance across different perturbation levels on CIFAR10 using ResNet20 backbone. Our evaluation focuses on two key metrics: robust accuracy (RA) against PGD-10 attacks, and the computational time (measured in hours)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Baselines and Evaluation Metrics: To thoroughly evaluate the adversarial robustness of our model, we use a comprehensive set of attack methods. Following [36], we use FGSM [11], and a series of PGD attacks [22] with varying iterations: PGD-10, PGD-20 and PGD-50. Additionally, we use the APGD-CE [7], C&W attack [5], and Autoattack (AA) [7]. AutoAttack is an ensemble method" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.116, + 0.381, + 0.129 + ], + "angle": 0, + "content": "Zareapoor et al." + }, + { + "type": "table", + "bbox": [ + 0.218, + 0.144, + 0.789, + 0.261 + ], + "angle": 0, + "content": "
MethodsSARA-FGSMRA-PGD10RA-PGD20RA-PGD50RA-C&WRA-APGD-CERA-AATime (h)
AT PGD65.7246.5140.8735.4234.2130.9733.6826.856.3
Fast-GA [3]58.4939.8833.3126.8925.0722.8124.3317.562.4
ZeroGrad [10]72.8144.2632.7822.1519.2320.8517.2012.751.3
NuAT [33]73.5544.9035.5226.3424.3124.9523.7218.931.8
Fast-AT [36]50.6438.6226.4721.6420.5519.2619.8114.971.2
RAT-AT70.6246.2838.1230.8329.3526.5427.7919.641.3
MEP-PGI [15]54.2336.8432.6527.9826.7022.6526.7319.311.5
RAT-MEP69.8647.6941.2634.8133.6428.7532.8723.911.7
" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.271, + 0.788, + 0.328 + ], + "angle": 0, + "content": "Table 2: CIFAR10 results: We report the RA (\\%) against various attacks at \\(\\mathcal{E} = 16 / 255\\) with ResNet20 backbone. Each model undergoes 3 training sessions. The results reflect the average performance of the top-performing model across these iterations. The best and the runner-up models are bolded and underlined, respectively." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.362, + 0.788, + 0.471 + ], + "angle": 0, + "content": "that integrates APGD-CE, APGD-T (targeted APGD), and Square Attack [2], to thoroughly evaluate a model's robustness to adversarial attacks. Due to the inherent variability in the training process, which often leads to fluctuations in performance, relying on a single run may not accurately capture the effectiveness of different methods. Thus, for each method, we conduct the training process three times and report the average of the best results in these runs. The evaluation metrics include robust accuracy (RA) and standard accuracy (SA)." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.499, + 0.431, + 0.515 + ], + "angle": 0, + "content": "4.2 Results and Insights" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.538, + 0.789, + 0.844 + ], + "angle": 0, + "content": "In our preliminary analysis, we highlight the importance of our results on the adversarial robustness of trained models. Our findings indicate that adversarial attacks with \\(\\mathcal{E} = 16 / 255\\), rarely result in the misclassification of the input. This observation is achieved through the generation of adversarial examples across CIFAR10 and CIFAR100 using both targeted and non-targeted attack methods, including FGSM, APGD-CE, APGD-T, PGD, and C&W. Remarkably, most of these adversarial examples still were classified correctly according to their original labels. Our experiments in Tab. 1 involve generating adversarial training with perturbation size \\(\\mathcal{E} = 16 / 255\\) and testing the models' robustness across different levels of perturbation. We also compared our model (RAT) with other techniques, such as Fast-Bat, AWP, OAAT, and ExAT. We have reimplemented these AT models within the FastAT framework to ensure a fair comparison. Our model achieves superior performance even as the perturbation magnitude increases to \\(\\mathcal{E} = 16 / 255\\). Furthermore, our model integrates into FAST-AT and MEP-PGI, which are specifically designed to address distributions of adversarial samples. Conversely, methods like OAAT are optimized for the distribution of benign examples and yield higher SA accuracy. Adopting the settings from OAAT, we apply both OAAT and our variation \\(\\mathrm{OAAT}_{\\mathrm{RAT}}\\) to FastAT tasks. The \\(\\mathrm{OAAT}_{\\mathrm{RAT}}\\) achieves \\(+2.78\\%\\) improvement on SA accuracy over OAAT (\\(70.51\\% \\rightarrow 73.29\\)), and \\(24.66\\%\\) on the SQ, as (\\(21.43\\% \\rightarrow 24.66\\%\\))." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.482, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Rethinking Fast Adversarial Training" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "table", + "bbox": [ + 0.22, + 0.144, + 0.788, + 0.253 + ], + "angle": 0, + "content": "
MethodsESARA-FGSMRA-PGD10RA-PGD20RA-PGD50RA-C&WRA-APGD-CERA-AATime (h)
Fast-GA [36]1267.3943.5732.1928.0226.2427.6526.2121.752.3
NuAT [33]72.8452.3341.9538.8437.6936.0836.7532.411.8
MEP-PGI [15]73.7852.3938.6736.5235.7833.4133.6827.931.5
RAT-MEP72.8154.6245.2942.8642.2336.8841.6033.161.7
Fast-GA [3]1081.7055.7336.8030.7128.6531.2926.9124.122.3
NuAT [33]76.3956.0946.2343.9543.5041.5243.3638.991.8
MEP-PGI [15]81.9459.4843.1839.7637.8337.4636.2132.721.5
RAT-MEP76.5258.3047.5145.7445.5641.2745.3639.701.7
" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.263, + 0.788, + 0.306 + ], + "angle": 0, + "content": "Table 3: Quantitative comparison of FastAT methods on CIFAR10 using ResNet20 for different \\((\\mathcal{E})\\), where both training and testing used the same \\(\\mathcal{E}\\). The best and the runner-up methods are bolded and underlined, respectively." + }, + { + "type": "table", + "bbox": [ + 0.22, + 0.319, + 0.788, + 0.431 + ], + "angle": 0, + "content": "
MethodsSARA-FGSMRA-PGD10RA-PGD20RA-PGD50RA-C&WRA-APGD-CERA-AATime (h)
AT PGD40.8325.1921.3817.5917.2315.0617.1012.546.3
Fast-GA [3]32.4015.8412.9810.7910.088.549.726.842.4
ZeroGrad [10]47.8124.1318.2112.9312.2411.7011.217.901.2
NuAT [33]32.4921.2716.7514.3813.7111.5212.628.751.8
Fast-AT [36]31.2215.6913.4110.7810.258.739.887.211.2
RAT-AT (ours)42.2625.8119.4315.7714.5612.8113.959.861.3
MEP-PGI [15]22.4114.0912.259.869.618.159.536.861.5
RAT-MEP (ours)43.7725.2621.3517.4217.0314.2216.7611.491.7
" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.443, + 0.788, + 0.499 + ], + "angle": 0, + "content": "Table 4: CIFAR100 results: We report the RA (\\%) against various attacks at \\(\\mathcal{E} = 16 / 255\\) with ResNet20 backbone. Each method undergoes three training sessions. The results reflect the average performance of the top-performing model across these iterations. The best and the runner-up methods are bolded and underlined, respectively." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.528, + 0.515, + 0.544 + ], + "angle": 0, + "content": "4.3 Faster DR Splitting Technique" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.552, + 0.789, + 0.766 + ], + "angle": 0, + "content": "Fig. 1-right provides a detailed comparison of various adversarial training on the CIFAR10 dataset, at \\(\\mathcal{E} = (8,16) / 255\\). At the lower perturbation level (\\(\\mathcal{E} = 8 / 255\\)), all methods demonstrate similar AA and SA, with FAST-GA achieving the highest AA among them. However, increasing the perturbation to \\(\\mathcal{E} = 16 / 255\\) results in a significant reduction in AA across all methods, highlighting the challenges of maintaining robustness at higher perturbation levels. Interestingly, our model outperforms Fast-Bat at \\(\\mathcal{E} = 16 / 255\\), achieving \\(60.37\\%\\) vs. \\(59.34\\%\\) in SA accuracy. This improvement is also evident at \\(\\mathcal{E} = 8 / 255\\), where RAT aligns closely with the performance of other leading methods, e.g., \\(46.21\\%\\) AA accuracy, comparable to \\(46.37\\%\\), \\(45.83\\%\\) for FAST-GA and Fast-Bat respectively. While FGSM is faster (as shown in Fig. 1-left), results in \\(0\\%\\) AA accuracy. We also observe that our model adds very little overhead to FGSM training, i.e., 15.02 min with FGSM, but is considerably faster than other methods. The increased computation time in LLR and CURE is due to Double Backpropagation [28]." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.786, + 0.409, + 0.799 + ], + "angle": 0, + "content": "4.4 CIFAR10 Results" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.81, + 0.787, + 0.84 + ], + "angle": 0, + "content": "In this experiment, we use backbone ResNet20, to evaluate our model's performance, setting different perturbation levels \\(\\mathcal{E}\\). The results, in Tab.2, show that" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.116, + 0.381, + 0.129 + ], + "angle": 0, + "content": "Zareapoor et al." + }, + { + "type": "table", + "bbox": [ + 0.218, + 0.144, + 0.795, + 0.245 + ], + "angle": 0, + "content": "
MethodsSARA-FGSMRA-PGD10RA-PGD20RA-PGD50RA-C&WRA-APGD-CERA-AATime (h)
AT PGD33.7917.6613.8210.5610.288.3610.456.4768.1
Fast-AT [36]28.1213.0610.217.366.955.486.724.5110.8
Fast-GA [3]29.2413.9811.328.458.166.237.404.9621.4
ZeroGrad [10]34.5812.859.416.735.714.225.262.8110.9
NuAT [33]35.2016.7112.489.529.377.628.215.3925.2
PGI-BP [15]21.7611.329.487.567.205.497.214.1813.8
RAT-BP (ours)34.6117.6213.239.919.458.209.876.2514.2
" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.255, + 0.788, + 0.298 + ], + "angle": 0, + "content": "Table 5: Tiny-ImageNet results: We report the RA (\\%) against various attack methods accuracy at \\(\\mathcal{E} = 16 / 255\\) using PreActResNet18 as our backbone. The results represent the average performance of the best models across three training iterations." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.328, + 0.789, + 0.541 + ], + "angle": 0, + "content": "our model outperforms existing Fast AT techniques in adversarial robustness across various attacks. Specifically, our RAT-AT surpasses all FastAT methods, and our RAT-MEP is more effective than other methods while maintaining comparable SA accuracy. In comparison to PGD-AT, RAT-MEP achieves nearly equivalent adversarial robustness, with a marginal of 0.57 difference in performance against PGD-50 (33.64% vs. 34.21%). Regarding computational efficiency, RAT-AT is faster, completing in 75 minutes, compared to 144 minutes for Fast-GA and 108 minutes for NuAT. RAT-MEP also presents a good balance between computational time (102 minutes) and performance, taking only slightly longer than MEP-PGI (96 minutes) but significantly less than the 380 minutes required by PGD-AT. In Tab. 3, we conduct additional experiments on different perturbation scales \\(\\mathcal{E} = (10,12)/55\\), further validating the model's effectiveness. The results clearly demonstrate that our method consistently achieves superior adversarial robustness across different perturbation budgets." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.557, + 0.419, + 0.57 + ], + "angle": 0, + "content": "4.5 CIFAR100 Results" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.583, + 0.789, + 0.705 + ], + "angle": 0, + "content": "Tab. 4 details the performance on CIFAR100 with a ResNet20 backbone, highlighting that the training times for FastAT methods across CIFAR10 and CIFAR100 are similar. As can be seen, our model demonstrates the ability to maintain stable adversarial training, and among the FastAT methods, RAT-AT is the most robust against different attacks. RAT-PGI surpasses the performance of all FastAT methods in terms of adversarial robustness. While RAT-MEP is slightly lower than AT-PGD, it boasts a training speed that is approximately 3.5 times faster, marking a significant improvement in efficiency." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.723, + 0.414, + 0.739 + ], + "angle": 0, + "content": "4.6 ImageNet Results" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.789, + 0.842 + ], + "angle": 0, + "content": "Our experiments with the PreActResNet18 on Tiny-ImageNet, demonstrate the effectiveness of our method in handling larger and more complex dataset, as presented in Tab. 5. The RAT-BP shows comparable robustness to PGD-AT and surpasses other FastAT methods. In terms of training efficiency, RAT-BP is marginally higher, requiring 14.2 hours, compared to PGI-BP (13.8 hours), yet it is significantly faster than PGD-AT, which takes 68 hours. Further extending" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.482, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Rethinking Fast Adversarial Training" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "table", + "bbox": [ + 0.276, + 0.144, + 0.734, + 0.21 + ], + "angle": 0, + "content": "
MethodsSARA-AARA-FGSMRA-PGD10RA-PGD50Time (h)
Fast-AT [36]7.411.273.423.762.2561.5
PGI-BP [15]16.211.059.482.811.8577.6
RAT-BP (ours)25.601.8814.618.703.2979.2
" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.22, + 0.787, + 0.262 + ], + "angle": 0, + "content": "Table 6: Quantitative results of the adversarial training methods on ImageNet at \\((\\mathcal{E} = 16 / 255)\\) with ResNet50 as the backbone. Adding our model (RAT) helps PGI-BP overcome CO and improve its performance." + }, + { + "type": "image", + "bbox": [ + 0.226, + 0.274, + 0.78, + 0.406 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.41, + 0.788, + 0.493 + ], + "angle": 0, + "content": "Fig. 3: Catastrophic Overfitting: Illustrates the robustness across perturbation strengths. The x-axis \\((\\mathcal{E})\\) represents the strength of the perturbation applied to test images to create adversarial examples. The y-axis is the accuracy of the model under attack (PGD-20). As \\(\\mathcal{E}\\) increases, we expect accuracy to decrease, since stronger perturbations are harder to withstand. The AT PGD-10 is referenced for comparison, showing that our model achieves performance closely aligned with this benchmark." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.522, + 0.788, + 0.598 + ], + "angle": 0, + "content": "our evaluation to the ImageNet [8], we use a ResNet50 following [36] training configuration with the only difference by the adoption of a maximum image resolution of \\(288 \\times 288\\), similar to [16]. The initial learning rate is 0.1, with a reduction factor of 0.1 applied after 30 and 60 epochs, across a total of 90 epochs. The results in Tab. 6 show the stability of our training across this dataset." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.618, + 0.336, + 0.631 + ], + "angle": 0, + "content": "4.7 Ablation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.641, + 0.803, + 0.84 + ], + "angle": 0, + "content": "Our ablation study focuses on evaluating the effectiveness of our model against leading adversarial training methods, such as Fast-GA, Fast-AT, SLAT, and Fast-Bat. We also reference the robust but computationally intensive AT PGD10. PreActResNet18 is a backbone model for these experiments, with each method's recommended hyperparameters applied. We calculate the PGD-20 test accuracy, averaging results over three random seeds to ensure reliability. Fig. 3, illustrates that SLAT encounters CO at perturbation levels of \\(\\mathcal{E} \\geq 4/255\\) for SVHN and \\(\\mathcal{E}e.g.10/255\\) for CIFAR10. Fast-GA shows less than optimal performance at \\(\\mathcal{E}e.g.16/255\\) for CIFAR10 and inconsistent results on CIFAR100. Notably, Fast-Bat also experiences CO for \\(\\mathcal{E} > 10/255\\) in CIFAR100. Overall, our RAT model significantly narrows the gap between single-step and multi-step AT methods, showing enhanced robustness and efficiency. Sensitivity to \\(\\lambda\\): To optimize our model, we focus on determining the optimal value for the parameter \\(\\mathcal{E}\\) in Eq.(9)" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.116, + 0.381, + 0.129 + ], + "angle": 0, + "content": "Zareapoor et al." + }, + { + "type": "image", + "bbox": [ + 0.219, + 0.152, + 0.501, + 0.285 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.298, + 0.505, + 0.385 + ], + "angle": 0, + "content": "Fig. 4: Exploration of \\(\\lambda\\) values for RAT on CIFAR10, using PRN backbone. This illustrates the relationship between the choice of \\(\\lambda\\) and the occurrence of CO. Higher values successfully prevent CO but lead to a noticeable reduction in model performance." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.147, + 0.788, + 0.419 + ], + "angle": 0, + "content": "across varying levels of adversarial perturbation. We evaluate the PGD20 adversarial accuracy on a validation set of 1024 images from the CIFAR10 dataset. The experiments are conducted using the PRN architecture, exploring a range of \\(\\lambda \\in \\{0.1,0.3,0.5,0.6,0.9,1\\}\\). Except for \\(\\lambda\\), all other parameters are maintained as per the default configuration. The objective is to identify the \\(\\lambda\\) that yields the highest adversarial accuracy for each \\(\\mathcal{E}\\), as shown in Fig.4. Our results indicate a clear trend: lower \\(\\lambda\\) and higher \\(\\mathcal{E}\\) values, fail to prevent CO. Increasing \\(\\lambda\\) to 0.9 eliminates CO but degrades model performance. This effect is particularly ev" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.419, + 0.789, + 0.571 + ], + "angle": 0, + "content": "ident at \\(\\lambda = 1\\), where, the PGD-20 accuracy decreases for lower and higher perturbation values. For all datasets, the model performance is almost consistent across \\(\\lambda \\in \\{0.3, 0.5, 0.6\\}\\), leading us to select the lowest value within this range for efficiency. On the ImageNet dataset, however, a higher value (0.7) enhances performance. To further assess the impact of hyperparameters \\(\\gamma_{1}\\) and \\(\\gamma_{2}\\) on model performance and adversarial robustness, we conducted experiments by varying their values and analyzing the outcomes on CIFAR100. The results are presented in Table 2 (supplementary file). The values are chosen from a set \\(\\{0.0, 0.3, 0.5, 0.7, 0.8, 1.0\\}\\), and the model's performance was evaluated against several adversarial attacks, including FGSM, PGD-10, PGD-[20, 50], and AA." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.609, + 0.36, + 0.625 + ], + "angle": 0, + "content": "5 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.644, + 0.788, + 0.842 + ], + "angle": 0, + "content": "We presented an efficient model to overcome catastrophic overfitting in fast adversarial training through the Douglas-Rachford (DR) splitting technique, emphasizing linearity and stability in the training process. By using this technique, we provide a novel perspective on managing adversarial robustness that balances between minimizing adversarial loss and enforcing effective regularization. Indeed, using DR splitting allows for subtle control over the training process, facilitating adjustments in model parameters to prevent sudden degradation in adversarial robustness. This methodology not only addresses CO with greater efficiency than existing solutions like Fast-GA, Fast-AT, and NuAT but also enhances the overall training stability across a wide range of adversarial perturbations and training durations. Through extensive evaluations, with significant adversarial challenges and in conjunction with various Fast AT methods, our model demonstrates superior performance and adaptability." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.483, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Rethinking Fast Adversarial Training" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.323, + 0.16 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.177, + 0.785, + 0.218 + ], + "angle": 0, + "content": "1. Addepalli, S., Jain, S., Sriramanan, G., Venkatesh Babu, R.: Scaling adversarial training to large perturbation bounds. In: European Conference on Computer Vision. pp. 301-316 (2022) 6, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.219, + 0.785, + 0.26 + ], + "angle": 0, + "content": "2. Andriushchenko, M., Croce, F., Flammarion, N., Hein, M.: Square attack: a query-efficient black-box adversarial attack via random search. In: European conference on computer vision. pp. 484-501 (2020) 9, 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.261, + 0.785, + 0.3 + ], + "angle": 0, + "content": "3. Andriushchenko, M., Flammarion, N.: Understanding and improving fast adversarial training. Advances in Neural Information Processing Systems 33, 16048-16059 (2020) 2, 3, 4, 5, 6, 10, 11, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.301, + 0.785, + 0.342 + ], + "angle": 0, + "content": "4. Applegate, D., Hinder, O., Lu, H., Lubin, M.: Faster first-order primal-dual methods for linear programming using restarts and sharpness. Mathematical Programming 201(1), 133-184 (2023) 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.343, + 0.785, + 0.371 + ], + "angle": 0, + "content": "5. Carlini, N., Wagner, D.: Towards evaluating the robustness of neural networks. In: 2017 IEEE symposium on security and privacy (sp). pp. 39-57 (2017) 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.371, + 0.785, + 0.412 + ], + "angle": 0, + "content": "6. Chambolle, A., Contreras, J.P.: Accelerated bregman primal-dual methods applied to optimal transport and Wasserstein barycenter problems. SIAM Journal on Mathematics of Data Science 4(4), 1369-1395 (2022) 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.412, + 0.785, + 0.453 + ], + "angle": 0, + "content": "7. Croce, F., Hein, M.: Reliable evaluation of adversarial robustness with an ensemble of diverse parameter-free attacks. In: International conference on machine learning. pp. 2206-2216 (2020) 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.454, + 0.785, + 0.495 + ], + "angle": 0, + "content": "8. Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: Imagenet: A large-scale hierarchical image database. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 248-255 (2009) 8, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.495, + 0.785, + 0.536 + ], + "angle": 0, + "content": "9. Dong, Y., Liao, F., Pang, T., Su, H., Zhu, J., Hu, X., Li, J.: Boosting adversarial attacks with momentum. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9185-9193 (2018) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.536, + 0.785, + 0.577 + ], + "angle": 0, + "content": "0. Golgooni, Z., Saberi, M., Eskandar, M., Rohhan, M.H.: Zerograd: Mitigating and explaining catastrophic overfitting in fgsm adversarial training. arXiv preprint arXiv:2103.15476 (2021) 5, 10, 11, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.578, + 0.785, + 0.606 + ], + "angle": 0, + "content": "1. Goodfellow, I.J., Shlens, J., Szegedy, C.: Explaining and harnessing adversarial examples. International Conference on Learning Representations (2014) 1, 3, 4, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.606, + 0.785, + 0.647 + ], + "angle": 0, + "content": "2. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 770-778 (2016) 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.647, + 0.785, + 0.688 + ], + "angle": 0, + "content": "3. He, K., Zhang, X., Ren, S., Sun, J.: Identity mappings in deep residual networks. In: Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part IV 14. pp. 630-645 (2016) 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.689, + 0.785, + 0.73 + ], + "angle": 0, + "content": "4. He, Z., Li, T., Chen, S., Huang, X.: Investigating catastrophic overfitting in fast adversarial training: A self-fitting perspective. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2313-2320 (2023) 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.73, + 0.785, + 0.771 + ], + "angle": 0, + "content": "5. Jia, X., Zhang, Y., Wei, X., Wu, B., Ma, K., Wang, J., Cao, X.: Prior-guided adversarial initialization for fast adversarial training. In: European Conference on Computer Vision. pp. 567-584 (2022) 2, 6, 9, 10, 11, 12, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.771, + 0.785, + 0.812 + ], + "angle": 0, + "content": "6. de Jorge Aranda, P., Bibi, A., Volpi, R., Sanyal, A., Torr, P., Rogez, G., Dokania, P.: Make some noise: Reliable and efficient single-step adversarial training. Advances in Neural Information Processing Systems 35, 12881-12893 (2022) 2, 5, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.812, + 0.785, + 0.84 + ], + "angle": 0, + "content": "7. Kurakin, A., Goodfellow, I., Bengio, S.: Adversarial machine learning at scale. arXiv preprint arXiv:1611.01236 (2016) 3" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.177, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.381, + 0.129 + ], + "angle": 0, + "content": "Zareapoor et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.787, + 0.189 + ], + "angle": 0, + "content": "18. Li, T., Wu, Y., Chen, S., Fang, K., Huang, X.: Subspace adversarial training. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13409-13418 (2022) 1, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.19, + 0.787, + 0.232 + ], + "angle": 0, + "content": "19. Liang, J., Fadili, J., Peyre, G.: Local convergence properties of douglas-rachford and alternating direction method of multipliers. Journal of Optimization Theory and Applications 172, 874-913 (2017) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.232, + 0.787, + 0.272 + ], + "angle": 0, + "content": "20. Lindbäck, J., Wang, Z., Johansson, M.: Bringing regularized optimal transport to lightspeed: a splitting method adapted for gpus. Advances in Neural Information Processing Systems 36 (2023) 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.273, + 0.787, + 0.314 + ], + "angle": 0, + "content": "21. Liu, X., Chakraborty, S., Sun, Y., Huang, F.: Rethinking adversarial policies: A generalized attack formulation and provable defense in rl. In: International Conference on Learning Representations (2024) 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.315, + 0.787, + 0.355 + ], + "angle": 0, + "content": "22. Madry, A., Makelov, A., Schmidt, L., Tsipras, D., Vladu, A.: Towards deep learning models resistant to adversarial attacks. International Conference on Learning Representations (2018) 1, 3, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.356, + 0.787, + 0.397 + ], + "angle": 0, + "content": "23. Mai, V.V., Lindbäck, J., Johansson, M.: A fast and accurate splitting method for optimal transport: Analysis and implementation. International Conference on Learning Representations (2022) 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.398, + 0.787, + 0.439 + ], + "angle": 0, + "content": "24. Moosavi-Dezfooli, S.M., Fawzi, A., Uesato, J., Frossard, P.: Robustness via curvature regularization, and vice versa. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9078-9086 (2019) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.439, + 0.787, + 0.481 + ], + "angle": 0, + "content": "25. Park, G.Y., Lee, S.W.: Reliably fast adversarial training via latent adversarial perturbation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 7758-7767 (2021) 1, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.481, + 0.787, + 0.522 + ], + "angle": 0, + "content": "26. Poon, C., Liang, J.: Trajectory of alternating direction method of multipliers and adaptive acceleration. Advances in Neural Information Processing Systems 32 (2019) 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.522, + 0.787, + 0.563 + ], + "angle": 0, + "content": "27. Qin, C., Martens, J., Gowal, S., Krishnan, D., Dvijotham, K., Fawzi, A., De, S., Stanforth, R., Kohli, P.: Adversarial robustness through local linearization. Advances in Neural Information Processing Systems 32 (2019) 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.564, + 0.787, + 0.605 + ], + "angle": 0, + "content": "28. Rocamora, E.A., Liu, F., Chrysos, G.G., Olmos, P.M., Cevher, V.: Efficient local linearity regularization to overcome catastrophic overfitting. International Conference on Learning Representations (2024) 1, 5, 6, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.606, + 0.787, + 0.633 + ], + "angle": 0, + "content": "29. Shaeiri, A., Nobahari, R., Rohhan, M.H.: Towards deep learning models resistant to large perturbations. arXiv preprint arXiv:2003.13370 (2020) 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.633, + 0.787, + 0.674 + ], + "angle": 0, + "content": "30. Shafahi, A., Najibi, M., Ghiasi, M.A., Xu, Z., Dickerson, J., Studer, C., Davis, L.S., Taylor, G., Goldstein, T.: Adversarial training for free! Advances in Neural Information Processing Systems 32 (2019) 1, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.675, + 0.787, + 0.715 + ], + "angle": 0, + "content": "31. Song, C., He, K., Wang, L., Hopcroft, J.E.: Improving the generalization of adversarial training with domain adaptation. International Conference on Learning Representations (2019) 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.716, + 0.787, + 0.757 + ], + "angle": 0, + "content": "32. Sriramanan, G., Addepalli, S., Baburaj, A., et al.: Guided adversarial attack for evaluating and enhancing adversarial defenses. Advances in Neural Information Processing Systems 33, 20297-20308 (2020) 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.758, + 0.787, + 0.799 + ], + "angle": 0, + "content": "33. Sriramanan, G., Addepalli, S., Baburaj, A., et al.: Towards efficient and effective adversarial training. Advances in Neural Information Processing Systems 34, 11821-11833 (2021) 2, 5, 10, 11, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.799, + 0.787, + 0.84 + ], + "angle": 0, + "content": "34. Themelis, A., Patrinos, P.: Douglas-rachford splitting and admn for nonconvex optimization: Tight convergence results. SIAM Journal on Optimization 30(1), 149-181 (2020) 5" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.483, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Rethinking Fast Adversarial Training" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.189 + ], + "angle": 0, + "content": "35. Tsiligkaridis, T., Roberts, J.: Understanding and increasing efficiency of frank-wolfe adversarial training. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 50-59 (2022) 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.19, + 0.788, + 0.231 + ], + "angle": 0, + "content": "36. Wong, E., Rice, L., Kolter, J.Z.: Fast is better than free: Revisiting adversarial training. International Conference on Learning Representations (2020) 1, 2, 4, 6, 9, 10, 11, 12, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.231, + 0.788, + 0.259 + ], + "angle": 0, + "content": "37. Wu, D., Xia, S.T., Wang, Y.: Adversarial weight perturbation helps robust generalization. Advances in Neural Information Processing Systems (2020) 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.259, + 0.788, + 0.3 + ], + "angle": 0, + "content": "38. Xie, Y., Li, Z., Shi, C., Liu, J., Chen, Y., Yuan, B.: Enabling fast and universal audio adversarial attack using generative model. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 35, pp. 14129-14137 (2021) 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.3, + 0.788, + 0.328 + ], + "angle": 0, + "content": "39. Zagoruyko, S., Komodakis, N.: Wide residual networks. arXiv preprint arXiv:1605.07146 (2016) 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.328, + 0.788, + 0.37 + ], + "angle": 0, + "content": "40. Zhang, H., Yu, Y., Jiao, J., Xing, E., El Ghaoui, L., Jordan, M.: Theoretically principled trade-off between robustness and accuracy. In: International conference on machine learning. pp. 7472-7482 (2019) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.37, + 0.788, + 0.411 + ], + "angle": 0, + "content": "41. Zhang, Y., Zhang, G., Khanduri, P., Hong, M., Chang, S., Liu, S.: Revisiting and advancing fast adversarial training through the lens of bi-level optimization. In: International Conference on Machine Learning. pp. 26693-26712 (2022) 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.411, + 0.788, + 0.453 + ], + "angle": 0, + "content": "42. Zhao, M., Zhang, L., Kong, Y., Yin, B.: Fast adversarial training with smooth convergence. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 4720-4729 (2023) 2, 6" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.453 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/Rethinking Fast Adversarial Training_ A Splitting Technique To Overcome Catastrophic Overfitting/54a9d439-23cc-4380-af56-e12fd67a1f40_origin.pdf b/2024/Rethinking Fast Adversarial Training_ A Splitting Technique To Overcome Catastrophic Overfitting/54a9d439-23cc-4380-af56-e12fd67a1f40_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..2c9d9b33174848d741ae9c003e3d56f2d8b57497 --- /dev/null +++ b/2024/Rethinking Fast Adversarial Training_ A Splitting Technique To Overcome Catastrophic Overfitting/54a9d439-23cc-4380-af56-e12fd67a1f40_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b82d06f499b29d57f4374851a8f211fe19f36200fe5a94a322c88eb5f19a260c +size 2429299 diff --git a/2024/Rethinking Fast Adversarial Training_ A Splitting Technique To Overcome Catastrophic Overfitting/full.md b/2024/Rethinking Fast Adversarial Training_ A Splitting Technique To Overcome Catastrophic Overfitting/full.md new file mode 100644 index 0000000000000000000000000000000000000000..5a7c5b041cfcaf02791f7f426c7b49ad72f5a70c --- /dev/null +++ b/2024/Rethinking Fast Adversarial Training_ A Splitting Technique To Overcome Catastrophic Overfitting/full.md @@ -0,0 +1,273 @@ +# Rethinking Fast Adversarial Training: A Splitting Technique to Overcome Catastrophic Overfitting + +Masoumeh Zareapoor and Pourya Shamsolmoali\* + +East China Normal University, Queen's University Belfast + +Abstract. Catastrophic overfitting (CO) poses a significant challenge to fast adversarial training (FastAT), particularly at large perturbation scales, leading to dramatic reductions in adversarial test accuracy. Our analysis of existing FastAT methods shows that CO is accompanied by abrupt and irregular fluctuations in loss convergence, indicating that a stable training dynamic is key to preventing CO. Therefore, we propose a training model that uses the Douglas-Rachford (DR) splitting technique to ensure a balanced and consistent training progression, effectively counteracting CO. The DR splitting technique, known for its ability to solve complex optimization problems, offering a distinct advantage over classical FastAT methods by providing a smoother loss convergence. This is achieved without resorting to complex regularization or incurring the computational costs associated with double backpropagation, presenting an efficient solution to enhance adversarial robustness. Our comprehensive evaluation conducted across standard datasets, demonstrates that our DR splitting-based model not only improves adversarial robustness but also achieves this with remarkable efficiency compared to various FastAT methods. This efficiency is particularly observed under conditions involving long training schedules and large adversarial perturbations. + +Keywords: Adversarial training $\cdot$ Adversarial robustness $\cdot$ Convergence + +# 1 Introduction + +Deep neural networks (DNNs), despite their successes, are vulnerable to adversarial attacks, where small, undetectable, input perturbations can mislead models into incorrect predictions [11]. Given the security risks of such vulnerabilities, improving neural network robustness against these attacks has become a critical area of research [28, 38]. Adversarial training (AT) [18, 30] and its derivative, PGD-AT [22, 31], have emerged as effective strategies for enhancing the robustness of neural networks. However, the complex inner optimizations required by these methods at each training step significantly elevate computational costs, limiting their practicality for large-scale applications [28]. To address these computational challenges, 'fast' versions of AT (FastAT) [25, 36] have been proposed, which employ a single-step gradient propagation [11] to produce adversarial training examples. While FastAT offers a more efficient training process + +compared to PGD-AT, it introduces the problem of Catastrophic Overfitting (CO) [16,36]. CO occurs when training with strong adversaries, leading to a significant drop in the robustness of the resulting model. Several FastAT methods have been developed to reduce CO [42]. Wong et al. [36] enhance adversarial example diversity through the use of initial perturbations that are randomly initialized. Following this approach, [3] proposed FAST-GA, an efficient regularizer that aims to explicitly align gradients between benign and adversarial inputs. Likewise, MEP-PGI [15] and NuAT [33] employ nuclear or p-norm regularization within adversarial training to improve consistency in predictions on both benign and adversarial examples. These approaches are effective only within a limited perturbation $\mathcal{E} \leq 8/255$ , where $\mathcal{E}$ denotes the degree of perturbations in the adversarial training [42]. Moreover, models trained with relatively small perturbation scales show vulnerability to larger $\mathcal{E}$ , with noted declines in robustness. + +For example, models trained with NuAT and MEP-PGI at a perturbation size of $\mathcal{E} = 8 / 255$ demonstrated robustness levels of $56\%$ and $55\%$ against PGD-50 attacks, but their robustness significantly dropped to $24\%$ and $22\%$ , respectively, with $\mathcal{E} = 16 / 255$ . Our investigation into adversarial training processes (illustrated in Fig. 2) reveals that CO is characterized by minor fluctuations in benign classification loss and significant decreases in adversarial example loss. This observation raises a question: How can we design fast adversarial training that ensures improved stability and mitigates CO, while also optimizing the trade-off between robustness accuracy (RA) and standard accuracy (SA)? + +In this context, the Douglas-Rachford (DR) splitting technique emerges as a promising solution. The DR splitting is known for its linear convergence across both strongly convex and non-convex settings [4, 20, 23, 26]. Several studies demonstrate its applicability in a variety of optimization tasks. For example, Liang et al. [19] and Applegate et al. [4] have highlighted its capability for local and global linear convergence, underscoring the DR technique's effectiveness in complex optimization scenarios. We also capitalize on the inherent stability and efficiency of the DR splitting, incorporating these properties into adversarial training to address CO. This integration allows for a more stable and consistent adversarial learning process, significantly lowering the risk of CO. Indeed, our implementation demonstrates that DR splitting not only stabilizes the training process but also offers a cost-effective solution to enhance model robustness against adversarial attacks, as shown in Fig. 1. Our contributions are threefold: ① We provide a comprehensive analysis showing that existing FastAT methods are still prone to CO with larger $\varepsilon$ . Our investigation emphasizes the critical role of loss function convergence stability in addressing this vulnerability. ② Introducing a novel solution with the Douglas-Rachford splitting technique, specifically designed to mitigate CO without incurring computational complexities. ③ Conduct extensive experiments on standard datasets to validate the effectiveness of our model- Rethinking fast Adversarial Training (RAT)- particularly against large adversarial perturbations. RAT, leveraging the DR-splitting technique, shows enhanced adversarial robustness when compared to classical adversarial training methods (find the training code in the supplementary file). + +![](images/4475771b46e5295fa31d61cffa4f7fc85ab252215ffc9af798a910b05cb37ac6.jpg) +Fig. 1: Performance comparison of adversarial robustness with WideResNet34-10 on CIFAR10 for $\mathcal{E} = (8,16) / 255$ . (left) Illustrates the average runtime for each method. (right) We compare AA (AutoAttack) and SA accuracies. Lower AA indicates higher vulnerability. Higher SA (standard accuracy) indicates better general performance. + +
MethodsAA (%)SA (%)AA (%)SA (%)
816
AT PGD-1047.69±(0.11)78.51±(0.08)24.97±(0.26)59.41±(0.46)
LLR [27]43.84±(0.20)74.52±(0.07)16.31±(0.20)44.76±(4.72)
Fast-Bat [41]45.83±(0.12)79.61±(0.12)18.30±(0.26)59.34±(0.25)
Fast-GA [3]46.37±(0.27)79.12±(0.07)15.36±(1.32)54.25±(6.28)
CURE [24]44.65±(0.17)77.04±(0.11)18.12±(0.35)53.49±(0.08)
RAT (Ours)46.21±(0.08)80.32±(0.12)19.58±(0.17)60.37±(0.21)
+ +# 2 Related Work + +# 2.1 Adversarial Attacks + +Adversarial attacks aim to fool deep learning models by introducing perturbed inputs. Goodfellow et al. [11] introduced the concept of the fast gradient sign method (FGSM) for generating adversarial examples $x'$ within classification tasks. They demonstrated that adversarial examples, created through a single-step gradient update, can lead a classification model to incorrect predictions with significant confidence. The adversarial examples $x'$ are generated by + +$$ +x ^ {\prime} = x + \mathcal {E}. \operatorname {s g n} \left(\nabla_ {x} L (f (x; \theta), y\right), \tag {1} +$$ + +where $(x,y)$ denotes an input sample and its corresponding label, $\mathcal{E}$ is the perturbation magnitude, $\nabla_{x}L$ is the loss gradient with respect to $x$ , and $\theta$ signifies the fixed model weights. Subsequent research has developed more sophisticated attack methods, such as I-FGSM [17], MIM [9], and PGD [22], which leverage iterative gradient descent to enhance the efficacy of the attacks. The PGD attack, in particular, generates adversarial examples by applying small perturbations to the input data, ensuring these adjustments stay within a pre-defined boundary $\mathcal{E}$ . This process is described as follows + +$$ +x _ {t + 1} ^ {\prime} = \Pi_ {\mathcal {E}} (x _ {t} ^ {\prime} + \alpha . \mathrm {s g n} (\nabla_ {x} L (f (x _ {t} ^ {\prime}; \theta), y))), \qquad \qquad (2) +$$ + +in which $x'$ is the adversarial example at iteration $t$ , and $\Pi_{\mathcal{E}}$ denotes the projection operation that ensures the adversarial example remains within the $\mathcal{E}$ -ball. The PGD attack is one of the most effective attack methods for evaluating the robustness of deep learning against adversarial examples. + +# 2.2 Adversarial Training (AT) + +Adversarial Training (AT) [22] and TRADES [40] have become prominent training methods for training robust architectures, which can be written as + +$$ +\min _ {\theta} E _ {(x, y) \sim D} [ \max _ {\delta \in [ - \mathcal {E}, \mathcal {E} ]} L (f (x + \delta ; \theta), y) ], \tag {3} +$$ + +![](images/536e01bcbba3d09af02e5802bc6c636c6f872ceea42053455dcc8c68db8793e5.jpg) +Fig. 2: Performance of fast adversarial training (FastAT) methods on CIFAR10 using ResNet20, for $\mathcal{E} = 16 / 255$ . The figure illustrates the training process across three runs. The terms AV-loss and BN-loss are the training loss for adversarial and benign inputs, respectively, while AV-acc and BN-acc quantify testing accuracy against both adversarial and benign samples. The occurrence of catastrophic overfitting is marked (in green), indicating critical periods where adversarial robustness drops. + +where $\delta$ is the adversarial perturbations and $x' = x + \delta$ is the adversarial examples generated by various attacks such as PGD and FGSM. This optimization first, generates adversarial examples within a defined perturbation limit to maximize classification loss (inner maximization) and then adjusts the model parameters $\theta$ to minimize this loss over the adversarial examples (outer minimization). However, a notable trade-off emerges in AT between computational efficiency and adversarial robustness, due to the complexity of solving the inner maximization problem. To enhance the computational efficiency of AT, several methods have been developed. Free AT [30], simultaneously updating the network parameters $\theta$ and the adversarial perturbation $\delta$ multiple times within a single batch. On the other hand, Fast-AT [36], simplifies the process further by employing a single-step FGSM attack [11] but the attack initiates from a randomly chosen point $x + \mu$ with $\mu \in [-\delta, \delta]$ . Fast-Bat These approaches reduce the cost of solving the inner maximization problem, significantly accelerating the training process. + +# 2.3 Catastrophic Overfitting (CO) + +The acceleration of AT methods comes with significant drawbacks. Wong et al. [36] first reported this with FGSM training, and was later supported by Andriushchenko et al. [3] in their analysis of Free AT and Fast-AT methods. In these cases, the model's accuracy on adversarial tests can fall sharply, sometimes + +to $0\%$ , in just a few training epochs. This decline in performance has been attributed to CO, a critical challenge that has motivated significant research efforts aimed at understanding and mitigating its impact [14, 18, 41]. Below, we provide an overview of various strategies developed to avoid CO. + +# 2.4 Regularization-based Approaches + +Recent efforts focus on regularization and data augmentation to prevent CO. [3] introduced FAST-GA to mitigate CO by penalizing the misalignment of gradients between the input and their adversarial counterparts. This method ensures that the direction of the gradient remains consistent when slight perturbations are introduced to the input. Qin et al. [27] introduced local linear regularization (LLR) to improve model robustness against adversarial attacks. LLR smoothes the loss landscape using a first-order Taylor approximation, addressing nonlinear sensitivities and promoting a more linear loss response to perturbations. ELLE [28] introduced a regularization term designed to mitigate CO both effectively and efficiently within classical AT frameworks, by enforcing local linearity. Sriramanan et al. [32] introduced GAT, a novel regularization strategy aimed at reducing the difference between the logits of a training point and a random sample. NuAT [33], proposed a regularization term that penalizes the nuclear norm of the logits difference between benign and adversarial examples. However, GAT and NuAT still face challenges with CO under larger perturbation size $\mathcal{E}$ [16]. + +# 2.5 Other Approaches + +Park et al. [25] introduced SLAT which performs adversarial attacks within the latent space, focusing on the internal representations rather than the input space directly. Another study [35] employs the Frank-Wolfe optimization technique for AT, connecting the distortion observed in the loss landscape to the norm of adversarial perturbations, providing a deeper understanding of adversarial robustness. Li et al. [18] addressed CO by limiting network weights to a specific subspace based on the main components of AT-trained weights. Meanwhile, Zhang et al. [41] introduce Fast-Bat, which redefines the AT challenge as a Bilevel optimization problem and introduces a one-step solution to address it. However, [28] shows that Fast-BAT is affected by CO for $\mathcal{E} > 10 / 255$ . Zero-Grad [10] introduces a strategy to mitigate adversarial perturbations by setting to zero the gradient components that fall below a threshold, however, [16] demonstrates, this approach results in decreased performance for large $\mathcal{E}$ . Different from these methods, we aim to minimize the loss fluctuations during training through the Douglas-Rachford splitting technique, known for its effectiveness in complex optimization [23]. It has been successfully applied to optimal transport [20] and its discretizations [6], demonstrating its flexibility and robustness in achieving linear convergence rates under both convex and non-convex settings [4, 34]. A notable advantage of this splitting technique is to leverage sparsity in solutions, which can significantly accelerate the convergence process [26]. + +# 3 Method + +This section evaluates FastAT methods under large perturbation budgets, $\mathcal{E}$ , explores the causes of CO, and finally introduces our proposed solutions. + +# 3.1 Performance of Fast Adversarial Training on a Large $\mathcal{E}$ + +Fast AT methods have shown effectiveness in mitigating CO at lower perturbation $(\mathcal{E} \leq 8/255)$ [28, 36]. To extend this evaluation, we analyze performance at a higher perturbation level $(\mathcal{E} = 16/255)$ , in line with [42]. FAST-AT [36] introduces an initial step for Eq. (3), where samples are perturbed with uniform random noise $\delta_0 \sim U(-\mathcal{E}, \mathcal{E})$ before applying the FGSM attack, + +$$ +\min _ {\theta} E _ {(x, y) \sim D} \left[ \max _ {\delta_ {0} + \delta \in [ - \mathcal {E}, \mathcal {E} ]} L (f (x + \delta_ {0} + \delta ; \theta), y) \right]. \tag {4} +$$ + +Here, $L(\cdot)$ is the loss function, $\delta$ represents adversarial perturbations, and $\theta$ is the model weights. FAST-GA, as introduced by [3] focuses on aligning the gradients of the loss with respect to both benign samples $x$ and their perturbed samples $x + \delta_0$ , which is described as, + +$$ +E _ {(x, y) \sim D} [ 1 - \cos (\nabla_ {x} L (x, \theta), \nabla_ {x + \delta_ {0}} L (x + \delta_ {0}, \theta)) ], \tag {5} +$$ + +in which $L(x,\theta) = L(f(x;\theta),y)$ . The cosine similarity $\cos(\cdot)$ measures the cosine similarity between the gradients of the loss function for the benign and perturbed inputs. This strategy aims to ensure that the model's behavior does not change drastically for adversarial examples, thereby reducing CO. Different from other attack methods, MEP-PGI [15] extends the FGSM attack by incorporating a Prior-Guided Initialization (PGI) mechanism. This technique initializes the adversarial perturbations $\delta_0$ , by using the information from adversarial examples generated in previous training epochs, + +$$ +E _ {(x, y) \sim D} [ \| f (x ^ {\prime}; \theta) - f (x + \delta_ {0}; \theta) \| _ {2} ^ {2} ], \tag {6} +$$ + +here, $||\cdot||$ is the squared (L2) norm, measuring the distance between the model's outputs for the two types of examples. FAST-PGI aims to enrich the adversarial examples with historical insight, thereby enhancing the adversarial training process. Fig. 2 illustrates the adversarial training of ResNet20 on CIFAR10, highlighting times when CO occurs, between epochs 5th-20th, 40th-65th, and 10th-15th. The key observations from the results are i) transition from smooth/stable to erratic loss fluctuations; ii) discrepancies in the loss for benign and adversarial examples, i.e., minor fluctuations in benign loss $L(x,\theta)$ followed with a rapid decline in adversarial loss $L(x',\theta)$ ; and iii) a sharp decrease in adversarial accuracy. The result is consistent with the observations in [1,28,42]. From these patterns, we derive several insights for understanding CO: it is obvious that $L(x,\theta)$ has greater stability than the loss $L(x',\theta)$ , indicating a tendency of models to overfit to adversarial samples. CO directly correlates with the instability in adversarial training convergence. Anomalies in both benign and adversarial loss occur together, signifying an issue in the learning process that affects its performance + +on both data. We also find that Fast-GA and MEP-PGI enhance the stability of adversarial training through the constraints in Eq. (4) and Eq. (5), compared to the Fast-AT. We are now in shape to give a formal definition of our model. + +# 3.2 Douglas-Rachford (DR) Splitting for Adversarial Training + +We use the DR splitting technique to ensure a consistent and stable optimization to prevent sudden changes in adversarial robustness. This technique divides the optimization process into a two-step process (reflection and enforcement), which together ensure a stable and consistent optimization path. Let's denote the model at iteration $k$ by $f(\cdot, \theta^{(k)})$ , where $\theta^{(k)}$ represents the model parameters. The loss function for adversarial examples is $L(f(x', \theta^{(k)}, y))$ , and for benign examples, $L(f(x^{k}, \theta^{(k)}, y))$ . An important property of our model is its ability to alternate between reducing adversarial loss and applying careful regularization terms. This ensures that any adjustments to the model parameters effectively impact both benign and adversarial loss functions $L(x, \theta)$ and $L(x', \theta)$ . Specifically, to address CO, we introduce epoch-based constraints on the loss differences for both adversarial and benign instances, represented as $(\Delta L_{\theta^{(k)}}, \Delta L_{\theta^{(k)}}^{benign})$ , + +$$ +\begin{array}{l} \Delta L _ {\theta^ {(k)}} = \left| L \left(f \left(x ^ {\prime (k)}, \theta^ {(k)}\right), y\right) - L \left(f \left(x ^ {\prime (k - 1)}, \theta^ {(k - 1)}\right), y\right) \right| \\ \begin{array}{l} \left. \right| = \left| L (f (w, \theta^ {(k)}), y) - L (f (x, \theta^ {(k - 1)}), y) \right|, \end{array} \tag {7} \\ \end{array} +$$ + +where $|.|$ measures the absolute difference in loss for adversarial $(x')$ and benign $(x)$ inputs, $x'^{(k)} = x + \delta_{0,k} + \delta_k$ . Here, $\delta_{0,k}$ can be replaced by various attack initialization perturbations and $\delta_k$ is generated by $[\max_{\delta_k \in [-\varepsilon, \varepsilon]} L(x'^{(k)}, \theta^{(k-1)})]$ . This formulation guides the model to effectively manage perturbations and maintain a stable training process. Additionally, these constraints help in mitigating the simultaneous anomalies in both $L(x, \theta)$ and $L(x', \theta)$ facilitating a more controlled update of the model parameters that considers both adversarial and benign inputs. In practice, this involves updating the model's intermediate parameters, $C^k$ , to reduce the adversarial loss, according to the gradient of the loss function $\nabla_\theta$ with respect to the adversarial examples $x'^{(k)}$ , + +$$ +C ^ {k} = \theta^ {(k)} - \eta \nabla_ {\theta} L \left(f \left(x ^ {\prime (k)}, \theta^ {(k)}\right), y\right), \tag {8} +$$ + +where $\eta$ is a learning rate. This update, termed as the 'reflection step' ( $\theta_{\mathrm{ref}}$ ), is pivotal in aligning the model's parameters towards reducing adversarial vulnerability. We then combine these updates $C^k$ with regularization adjustments to stabilize the training dynamics and enhance the adversarial robustness, effectively preventing overfitting to adversarial examples while maintaining performance on benign inputs. The updated model parameters, $\theta^{(k + 1)}$ , allow to handle both adversarial and benign inputs effectively without compromising training stability or performance, + +$$ +\theta^ {(k + 1)} = C ^ {k} + \lambda (R e g (\theta^ {(k)}) - C ^ {k}), \tag {9} +$$ + +where $\lambda \in [0,1]$ controls the strength of the regularization, and Reg involves components in Eq. (7) such that + +$$ +R e g \left(\theta^ {(k)}\right) = \gamma_ {1} \left(\Delta L _ {\theta^ {(k)}}\right) + \gamma_ {2} \left(\Delta L _ {\theta^ {(k)}} ^ {b e n i g n}\right). \tag {10} +$$ + +The important reason behind Eq. (9) is that the regularization term $Reg(\theta^{(k)})$ incorporates changes in loss for both adversarial and benign examples, weighted by $\gamma_{1}$ and $\gamma_{2}$ , respectively. This helps to balance the model's performance across adversarial and benign inputs, addressing CO by preventing the model from becoming overly specialized on the adversarial examples. Because CO occurs when a model overfits to specific adversarial examples during training, compromising its ability to generalize [21]. Our formulation ensures that the model's updates consider the impact on both adversarial robustness and benign accuracy. This regularization can be seen as part of the 'enforcement step' $(\theta_{enf})$ . The DR technique involves iteratively updating the model parameters by combining the adversarial loss minimization and the stability enforcement components, as $\theta = 0.5 \times (\theta_{ref} + \theta_{enf})$ . $\theta_{ref}$ targets reducing the loss on adversariably perturbed inputs Eq. (8), conversely, $\theta_{enf}$ represents the parameters adjusted after applying the regularization term, focusing on stability and CO prevention Eq. (9). This approach under the DR-splitting technique provides a comprehensive strategy for addressing CO in the adversarial training process. Indeed, this allows the model to effectively adapt to adversarial perturbations while ensuring that such adaptations do not compromise the model's ability to accurately classify benign inputs or lead to overfitting on the adversarial examples during training. Our proposed RAT is given in Algorithm 1, discussion in the supplementary file. + +Efficient Computation of Loss Differences In practice, directly computing the losses $L(f(x^{\prime (k - 1)},\theta^{(k - 1)},y))$ and $L(f(x,\theta^{(k - 1)}),y)$ in Eq. (7) can consume a significant amount of memory. To address this, we introduce $u^{k - 1}$ for benign inputs and $u^{\prime k - 1}$ for adversarial inputs, replacements for these terms, allowing for efficient computation and tracking of the average loss over benign and adversarial examples, respectively. Specially, $u^{k - 1} = \frac{1}{N}\sum_{i}L(f(x_{i}^{\prime},\theta^{(k - 1)}),y_{i})$ represents the average loss for adversarial inputs, and $u^{\prime k - 1} = \frac{1}{N}\sum_{i = 1}^{N}L(f(x_i,\theta^{(k - 1)}),y_i)$ denotes the average loss for benign inputs, where $N$ is the batch size. This can reduce computational and memory overhead by focusing on the aggregate loss across batches, rather than calculating them individually for each sample. + +# 4 Experimental Results + +# 4.1 Experimental settings + +In this section, we evaluate the proposed model across four benchmark datasets to determine its effectiveness and robustness against adversarial attacks. For both CIFAR10 [8] and CIFAR100 [8], we use the ResNet20 [12] architecture as the backbone, recognizing its strong performance in image classification tasks. For the more complex Tiny-ImageNet and ImageNet [8], we use the PreActResNet18 [13] to better handle its challenges. Additionally, the use of WideResNet34-10 [39] highlights our model's capacity to maintain effectiveness on long training schedules. We use the SGD optimizer with a batch size of 128, a weight decay of $5.10^{-4}$ , and a momentum of 0.9. We initiate training with learning rates of 0.1 for CIFAR10 and CIFAR100. For Tiny ImageNet, due to its greater complexity, + +Algorithm 1 Rethinking fast Adversarial Training (RAT) using DR Splitting +Require: Training dataset $(x_{i},y_{i})_{i = 1}^{N}$ # epochs K, # batch B, perturbation budget $\mathcal{E},\lambda ,\gamma_1,\gamma_2$ +1: Initialize model parameters $\theta_0$ +2: for epoch $= 1$ to K do +3: for $i = 1$ to B do +4: $x_{i}^{\prime} = x_{i} + \delta \mathrm{sign}(\nabla_{x_{i}}L(f(x_{i};\theta_{\mathrm{epoch - 1}}),y_{i}))\triangleright$ Adversarial example generation +5: $L_{\mathrm{adv}} = L(f(x_i';\theta_{\mathrm{epoch - 1}}),y_i)$ +6: $L_{\mathrm{benign}} = L(f(x_i;\theta_{\mathrm{epoch - 1}}),y_i)$ +7: $\theta_{\mathrm{reflect}} = \theta_{\mathrm{epoch - 1}} - \eta \cdot \nabla_{\theta}L_{\mathrm{adv}}$ $\triangleright$ Reflection step +8: $\Delta L_{\mathrm{adv}} = |L(f(x_i';\theta_{\mathrm{reflect}}),y_i) - L_{\mathrm{adv}}|$ +9: $\Delta L_{\mathrm{benign}} = |L(f(x_i;\theta_{\mathrm{reflect}}),y_i) - L_{\mathrm{benign}}|$ +10: Reg $= \gamma_{1}\cdot \Delta L_{\mathrm{adv}} + \gamma_{2}\cdot \Delta L_{\mathrm{benign}}$ $\triangleright$ Enforcement step +11: $\theta_{\mathrm{enforce}} = \theta_{\mathrm{reflect}} + \lambda \cdot (\mathrm{Reg} - \theta_{\mathrm{reflect}})$ +12: $\theta_{\mathrm{epoch}} = 0.5\cdot (\theta_{\mathrm{reflect}} + \theta_{\mathrm{enforce}})$ $\triangleright$ DR splitting +13: end for +14: end for + +we start at a lower rate of 0.01. The training spans 110 epochs, with a strategy to decrease the learning rate by a factor of 0.1 at both the 100th and 105th epochs for optimization. To prepare our model against adversarial attacks, we incorporate FAST-AT [36] and MEP-PGI [15] as our initial attack strategies for CIFAR datasets. Given the significant memory demands of MEP-PGI, particularly its need to store previous adversarial perturbations, we use its variation PGI-BP on Tiny ImageNet. We set the step-size parameter $\gamma$ to 0.01 for CIFAR datasets, and 0.001 for Tiny ImageNet. The regularization term $\lambda = 0.3$ for all datasets, and $\mathcal{E} = 16 / 255$ . Experiments are run using a single GeForce RTX 3090 GPU. Find the training code and additional results in the supplementary file. + +
MethodSA (%)RA-PGD-10 +E=8/255RA-PGD-10 +E=10/255RA-PGD-10 +E=12/255RA-PGD-10 +E=16/255SQ [2]Time (h)
AWP [37]77.9436.2633.1830.2527.4319.762.4
OAAT [1]70.5145.3239.4734.1028.9121.433.2
ExAT [29]79.3349.5643.2836.4932.5823.151.2
RAT-MEP69.8652.6446.8344.1841.2627.131.7
+ +Table 1: Comparative performance across different perturbation levels on CIFAR10 using ResNet20 backbone. Our evaluation focuses on two key metrics: robust accuracy (RA) against PGD-10 attacks, and the computational time (measured in hours). + +Baselines and Evaluation Metrics: To thoroughly evaluate the adversarial robustness of our model, we use a comprehensive set of attack methods. Following [36], we use FGSM [11], and a series of PGD attacks [22] with varying iterations: PGD-10, PGD-20 and PGD-50. Additionally, we use the APGD-CE [7], C&W attack [5], and Autoattack (AA) [7]. AutoAttack is an ensemble method + +
MethodsSARA-FGSMRA-PGD10RA-PGD20RA-PGD50RA-C&WRA-APGD-CERA-AATime (h)
AT PGD65.7246.5140.8735.4234.2130.9733.6826.856.3
Fast-GA [3]58.4939.8833.3126.8925.0722.8124.3317.562.4
ZeroGrad [10]72.8144.2632.7822.1519.2320.8517.2012.751.3
NuAT [33]73.5544.9035.5226.3424.3124.9523.7218.931.8
Fast-AT [36]50.6438.6226.4721.6420.5519.2619.8114.971.2
RAT-AT70.6246.2838.1230.8329.3526.5427.7919.641.3
MEP-PGI [15]54.2336.8432.6527.9826.7022.6526.7319.311.5
RAT-MEP69.8647.6941.2634.8133.6428.7532.8723.911.7
+ +Table 2: CIFAR10 results: We report the RA (\%) against various attacks at $\mathcal{E} = 16 / 255$ with ResNet20 backbone. Each model undergoes 3 training sessions. The results reflect the average performance of the top-performing model across these iterations. The best and the runner-up models are bolded and underlined, respectively. + +that integrates APGD-CE, APGD-T (targeted APGD), and Square Attack [2], to thoroughly evaluate a model's robustness to adversarial attacks. Due to the inherent variability in the training process, which often leads to fluctuations in performance, relying on a single run may not accurately capture the effectiveness of different methods. Thus, for each method, we conduct the training process three times and report the average of the best results in these runs. The evaluation metrics include robust accuracy (RA) and standard accuracy (SA). + +# 4.2 Results and Insights + +In our preliminary analysis, we highlight the importance of our results on the adversarial robustness of trained models. Our findings indicate that adversarial attacks with $\mathcal{E} = 16 / 255$ , rarely result in the misclassification of the input. This observation is achieved through the generation of adversarial examples across CIFAR10 and CIFAR100 using both targeted and non-targeted attack methods, including FGSM, APGD-CE, APGD-T, PGD, and C&W. Remarkably, most of these adversarial examples still were classified correctly according to their original labels. Our experiments in Tab. 1 involve generating adversarial training with perturbation size $\mathcal{E} = 16 / 255$ and testing the models' robustness across different levels of perturbation. We also compared our model (RAT) with other techniques, such as Fast-Bat, AWP, OAAT, and ExAT. We have reimplemented these AT models within the FastAT framework to ensure a fair comparison. Our model achieves superior performance even as the perturbation magnitude increases to $\mathcal{E} = 16 / 255$ . Furthermore, our model integrates into FAST-AT and MEP-PGI, which are specifically designed to address distributions of adversarial samples. Conversely, methods like OAAT are optimized for the distribution of benign examples and yield higher SA accuracy. Adopting the settings from OAAT, we apply both OAAT and our variation $\mathrm{OAAT}_{\mathrm{RAT}}$ to FastAT tasks. The $\mathrm{OAAT}_{\mathrm{RAT}}$ achieves $+2.78\%$ improvement on SA accuracy over OAAT ( $70.51\% \rightarrow 73.29$ ), and $24.66\%$ on the SQ, as ( $21.43\% \rightarrow 24.66\%$ ). + +
MethodsESARA-FGSMRA-PGD10RA-PGD20RA-PGD50RA-C&WRA-APGD-CERA-AATime (h)
Fast-GA [36]1267.3943.5732.1928.0226.2427.6526.2121.752.3
NuAT [33]72.8452.3341.9538.8437.6936.0836.7532.411.8
MEP-PGI [15]73.7852.3938.6736.5235.7833.4133.6827.931.5
RAT-MEP72.8154.6245.2942.8642.2336.8841.6033.161.7
Fast-GA [3]1081.7055.7336.8030.7128.6531.2926.9124.122.3
NuAT [33]76.3956.0946.2343.9543.5041.5243.3638.991.8
MEP-PGI [15]81.9459.4843.1839.7637.8337.4636.2132.721.5
RAT-MEP76.5258.3047.5145.7445.5641.2745.3639.701.7
+ +Table 3: Quantitative comparison of FastAT methods on CIFAR10 using ResNet20 for different $(\mathcal{E})$ , where both training and testing used the same $\mathcal{E}$ . The best and the runner-up methods are bolded and underlined, respectively. + +
MethodsSARA-FGSMRA-PGD10RA-PGD20RA-PGD50RA-C&WRA-APGD-CERA-AATime (h)
AT PGD40.8325.1921.3817.5917.2315.0617.1012.546.3
Fast-GA [3]32.4015.8412.9810.7910.088.549.726.842.4
ZeroGrad [10]47.8124.1318.2112.9312.2411.7011.217.901.2
NuAT [33]32.4921.2716.7514.3813.7111.5212.628.751.8
Fast-AT [36]31.2215.6913.4110.7810.258.739.887.211.2
RAT-AT (ours)42.2625.8119.4315.7714.5612.8113.959.861.3
MEP-PGI [15]22.4114.0912.259.869.618.159.536.861.5
RAT-MEP (ours)43.7725.2621.3517.4217.0314.2216.7611.491.7
+ +Table 4: CIFAR100 results: We report the RA (\%) against various attacks at $\mathcal{E} = 16 / 255$ with ResNet20 backbone. Each method undergoes three training sessions. The results reflect the average performance of the top-performing model across these iterations. The best and the runner-up methods are bolded and underlined, respectively. + +# 4.3 Faster DR Splitting Technique + +Fig. 1-right provides a detailed comparison of various adversarial training on the CIFAR10 dataset, at $\mathcal{E} = (8,16) / 255$ . At the lower perturbation level ( $\mathcal{E} = 8 / 255$ ), all methods demonstrate similar AA and SA, with FAST-GA achieving the highest AA among them. However, increasing the perturbation to $\mathcal{E} = 16 / 255$ results in a significant reduction in AA across all methods, highlighting the challenges of maintaining robustness at higher perturbation levels. Interestingly, our model outperforms Fast-Bat at $\mathcal{E} = 16 / 255$ , achieving $60.37\%$ vs. $59.34\%$ in SA accuracy. This improvement is also evident at $\mathcal{E} = 8 / 255$ , where RAT aligns closely with the performance of other leading methods, e.g., $46.21\%$ AA accuracy, comparable to $46.37\%$ , $45.83\%$ for FAST-GA and Fast-Bat respectively. While FGSM is faster (as shown in Fig. 1-left), results in $0\%$ AA accuracy. We also observe that our model adds very little overhead to FGSM training, i.e., 15.02 min with FGSM, but is considerably faster than other methods. The increased computation time in LLR and CURE is due to Double Backpropagation [28]. + +# 4.4 CIFAR10 Results + +In this experiment, we use backbone ResNet20, to evaluate our model's performance, setting different perturbation levels $\mathcal{E}$ . The results, in Tab.2, show that + +
MethodsSARA-FGSMRA-PGD10RA-PGD20RA-PGD50RA-C&WRA-APGD-CERA-AATime (h)
AT PGD33.7917.6613.8210.5610.288.3610.456.4768.1
Fast-AT [36]28.1213.0610.217.366.955.486.724.5110.8
Fast-GA [3]29.2413.9811.328.458.166.237.404.9621.4
ZeroGrad [10]34.5812.859.416.735.714.225.262.8110.9
NuAT [33]35.2016.7112.489.529.377.628.215.3925.2
PGI-BP [15]21.7611.329.487.567.205.497.214.1813.8
RAT-BP (ours)34.6117.6213.239.919.458.209.876.2514.2
+ +Table 5: Tiny-ImageNet results: We report the RA (\%) against various attack methods accuracy at $\mathcal{E} = 16 / 255$ using PreActResNet18 as our backbone. The results represent the average performance of the best models across three training iterations. + +our model outperforms existing Fast AT techniques in adversarial robustness across various attacks. Specifically, our RAT-AT surpasses all FastAT methods, and our RAT-MEP is more effective than other methods while maintaining comparable SA accuracy. In comparison to PGD-AT, RAT-MEP achieves nearly equivalent adversarial robustness, with a marginal of 0.57 difference in performance against PGD-50 (33.64% vs. 34.21%). Regarding computational efficiency, RAT-AT is faster, completing in 75 minutes, compared to 144 minutes for Fast-GA and 108 minutes for NuAT. RAT-MEP also presents a good balance between computational time (102 minutes) and performance, taking only slightly longer than MEP-PGI (96 minutes) but significantly less than the 380 minutes required by PGD-AT. In Tab. 3, we conduct additional experiments on different perturbation scales $\mathcal{E} = (10,12)/55$ , further validating the model's effectiveness. The results clearly demonstrate that our method consistently achieves superior adversarial robustness across different perturbation budgets. + +# 4.5 CIFAR100 Results + +Tab. 4 details the performance on CIFAR100 with a ResNet20 backbone, highlighting that the training times for FastAT methods across CIFAR10 and CIFAR100 are similar. As can be seen, our model demonstrates the ability to maintain stable adversarial training, and among the FastAT methods, RAT-AT is the most robust against different attacks. RAT-PGI surpasses the performance of all FastAT methods in terms of adversarial robustness. While RAT-MEP is slightly lower than AT-PGD, it boasts a training speed that is approximately 3.5 times faster, marking a significant improvement in efficiency. + +# 4.6 ImageNet Results + +Our experiments with the PreActResNet18 on Tiny-ImageNet, demonstrate the effectiveness of our method in handling larger and more complex dataset, as presented in Tab. 5. The RAT-BP shows comparable robustness to PGD-AT and surpasses other FastAT methods. In terms of training efficiency, RAT-BP is marginally higher, requiring 14.2 hours, compared to PGI-BP (13.8 hours), yet it is significantly faster than PGD-AT, which takes 68 hours. Further extending + +
MethodsSARA-AARA-FGSMRA-PGD10RA-PGD50Time (h)
Fast-AT [36]7.411.273.423.762.2561.5
PGI-BP [15]16.211.059.482.811.8577.6
RAT-BP (ours)25.601.8814.618.703.2979.2
+ +Table 6: Quantitative results of the adversarial training methods on ImageNet at $(\mathcal{E} = 16 / 255)$ with ResNet50 as the backbone. Adding our model (RAT) helps PGI-BP overcome CO and improve its performance. + +![](images/2b42ffc45215e7bf1b028fbe4164460bfb306abefd475b5a109e21af7f648737.jpg) +Fig. 3: Catastrophic Overfitting: Illustrates the robustness across perturbation strengths. The x-axis $(\mathcal{E})$ represents the strength of the perturbation applied to test images to create adversarial examples. The y-axis is the accuracy of the model under attack (PGD-20). As $\mathcal{E}$ increases, we expect accuracy to decrease, since stronger perturbations are harder to withstand. The AT PGD-10 is referenced for comparison, showing that our model achieves performance closely aligned with this benchmark. + +our evaluation to the ImageNet [8], we use a ResNet50 following [36] training configuration with the only difference by the adoption of a maximum image resolution of $288 \times 288$ , similar to [16]. The initial learning rate is 0.1, with a reduction factor of 0.1 applied after 30 and 60 epochs, across a total of 90 epochs. The results in Tab. 6 show the stability of our training across this dataset. + +# 4.7 Ablation + +Our ablation study focuses on evaluating the effectiveness of our model against leading adversarial training methods, such as Fast-GA, Fast-AT, SLAT, and Fast-Bat. We also reference the robust but computationally intensive AT PGD10. PreActResNet18 is a backbone model for these experiments, with each method's recommended hyperparameters applied. We calculate the PGD-20 test accuracy, averaging results over three random seeds to ensure reliability. Fig. 3, illustrates that SLAT encounters CO at perturbation levels of $\mathcal{E} \geq 4/255$ for SVHN and $\mathcal{E}e.g.10/255$ for CIFAR10. Fast-GA shows less than optimal performance at $\mathcal{E}e.g.16/255$ for CIFAR10 and inconsistent results on CIFAR100. Notably, Fast-Bat also experiences CO for $\mathcal{E} > 10/255$ in CIFAR100. Overall, our RAT model significantly narrows the gap between single-step and multi-step AT methods, showing enhanced robustness and efficiency. Sensitivity to $\lambda$ : To optimize our model, we focus on determining the optimal value for the parameter $\mathcal{E}$ in Eq.(9) + +![](images/7bb2199ce960e8a180d60b87c83905ffe1ddded7c9f3c4e0605def585f3475dd.jpg) +Fig. 4: Exploration of $\lambda$ values for RAT on CIFAR10, using PRN backbone. This illustrates the relationship between the choice of $\lambda$ and the occurrence of CO. Higher values successfully prevent CO but lead to a noticeable reduction in model performance. + +across varying levels of adversarial perturbation. We evaluate the PGD20 adversarial accuracy on a validation set of 1024 images from the CIFAR10 dataset. The experiments are conducted using the PRN architecture, exploring a range of $\lambda \in \{0.1,0.3,0.5,0.6,0.9,1\}$ . Except for $\lambda$ , all other parameters are maintained as per the default configuration. The objective is to identify the $\lambda$ that yields the highest adversarial accuracy for each $\mathcal{E}$ , as shown in Fig.4. Our results indicate a clear trend: lower $\lambda$ and higher $\mathcal{E}$ values, fail to prevent CO. Increasing $\lambda$ to 0.9 eliminates CO but degrades model performance. This effect is particularly ev + +ident at $\lambda = 1$ , where, the PGD-20 accuracy decreases for lower and higher perturbation values. For all datasets, the model performance is almost consistent across $\lambda \in \{0.3, 0.5, 0.6\}$ , leading us to select the lowest value within this range for efficiency. On the ImageNet dataset, however, a higher value (0.7) enhances performance. To further assess the impact of hyperparameters $\gamma_{1}$ and $\gamma_{2}$ on model performance and adversarial robustness, we conducted experiments by varying their values and analyzing the outcomes on CIFAR100. The results are presented in Table 2 (supplementary file). The values are chosen from a set $\{0.0, 0.3, 0.5, 0.7, 0.8, 1.0\}$ , and the model's performance was evaluated against several adversarial attacks, including FGSM, PGD-10, PGD-[20, 50], and AA. + +# 5 Conclusion + +We presented an efficient model to overcome catastrophic overfitting in fast adversarial training through the Douglas-Rachford (DR) splitting technique, emphasizing linearity and stability in the training process. By using this technique, we provide a novel perspective on managing adversarial robustness that balances between minimizing adversarial loss and enforcing effective regularization. Indeed, using DR splitting allows for subtle control over the training process, facilitating adjustments in model parameters to prevent sudden degradation in adversarial robustness. This methodology not only addresses CO with greater efficiency than existing solutions like Fast-GA, Fast-AT, and NuAT but also enhances the overall training stability across a wide range of adversarial perturbations and training durations. Through extensive evaluations, with significant adversarial challenges and in conjunction with various Fast AT methods, our model demonstrates superior performance and adaptability. + +# References + +1. Addepalli, S., Jain, S., Sriramanan, G., Venkatesh Babu, R.: Scaling adversarial training to large perturbation bounds. In: European Conference on Computer Vision. pp. 301-316 (2022) 6, 9 +2. Andriushchenko, M., Croce, F., Flammarion, N., Hein, M.: Square attack: a query-efficient black-box adversarial attack via random search. In: European conference on computer vision. pp. 484-501 (2020) 9, 10 +3. Andriushchenko, M., Flammarion, N.: Understanding and improving fast adversarial training. Advances in Neural Information Processing Systems 33, 16048-16059 (2020) 2, 3, 4, 5, 6, 10, 11, 12 +4. Applegate, D., Hinder, O., Lu, H., Lubin, M.: Faster first-order primal-dual methods for linear programming using restarts and sharpness. Mathematical Programming 201(1), 133-184 (2023) 2, 5 +5. Carlini, N., Wagner, D.: Towards evaluating the robustness of neural networks. In: 2017 IEEE symposium on security and privacy (sp). pp. 39-57 (2017) 9 +6. Chambolle, A., Contreras, J.P.: Accelerated bregman primal-dual methods applied to optimal transport and Wasserstein barycenter problems. SIAM Journal on Mathematics of Data Science 4(4), 1369-1395 (2022) 5 +7. Croce, F., Hein, M.: Reliable evaluation of adversarial robustness with an ensemble of diverse parameter-free attacks. In: International conference on machine learning. pp. 2206-2216 (2020) 9 +8. Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: Imagenet: A large-scale hierarchical image database. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 248-255 (2009) 8, 13 +9. Dong, Y., Liao, F., Pang, T., Su, H., Zhu, J., Hu, X., Li, J.: Boosting adversarial attacks with momentum. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9185-9193 (2018) 3 +0. Golgooni, Z., Saberi, M., Eskandar, M., Rohhan, M.H.: Zerograd: Mitigating and explaining catastrophic overfitting in fgsm adversarial training. arXiv preprint arXiv:2103.15476 (2021) 5, 10, 11, 12 +1. Goodfellow, I.J., Shlens, J., Szegedy, C.: Explaining and harnessing adversarial examples. International Conference on Learning Representations (2014) 1, 3, 4, 9 +2. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 770-778 (2016) 8 +3. He, K., Zhang, X., Ren, S., Sun, J.: Identity mappings in deep residual networks. In: Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part IV 14. pp. 630-645 (2016) 8 +4. He, Z., Li, T., Chen, S., Huang, X.: Investigating catastrophic overfitting in fast adversarial training: A self-fitting perspective. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2313-2320 (2023) 5 +5. Jia, X., Zhang, Y., Wei, X., Wu, B., Ma, K., Wang, J., Cao, X.: Prior-guided adversarial initialization for fast adversarial training. In: European Conference on Computer Vision. pp. 567-584 (2022) 2, 6, 9, 10, 11, 12, 13 +6. de Jorge Aranda, P., Bibi, A., Volpi, R., Sanyal, A., Torr, P., Rogez, G., Dokania, P.: Make some noise: Reliable and efficient single-step adversarial training. Advances in Neural Information Processing Systems 35, 12881-12893 (2022) 2, 5, 13 +7. Kurakin, A., Goodfellow, I., Bengio, S.: Adversarial machine learning at scale. arXiv preprint arXiv:1611.01236 (2016) 3 + +18. Li, T., Wu, Y., Chen, S., Fang, K., Huang, X.: Subspace adversarial training. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13409-13418 (2022) 1, 5 +19. Liang, J., Fadili, J., Peyre, G.: Local convergence properties of douglas-rachford and alternating direction method of multipliers. Journal of Optimization Theory and Applications 172, 874-913 (2017) 2 +20. Lindbäck, J., Wang, Z., Johansson, M.: Bringing regularized optimal transport to lightspeed: a splitting method adapted for gpus. Advances in Neural Information Processing Systems 36 (2023) 2, 5 +21. Liu, X., Chakraborty, S., Sun, Y., Huang, F.: Rethinking adversarial policies: A generalized attack formulation and provable defense in rl. In: International Conference on Learning Representations (2024) 8 +22. Madry, A., Makelov, A., Schmidt, L., Tsipras, D., Vladu, A.: Towards deep learning models resistant to adversarial attacks. International Conference on Learning Representations (2018) 1, 3, 9 +23. Mai, V.V., Lindbäck, J., Johansson, M.: A fast and accurate splitting method for optimal transport: Analysis and implementation. International Conference on Learning Representations (2022) 2, 5 +24. Moosavi-Dezfooli, S.M., Fawzi, A., Uesato, J., Frossard, P.: Robustness via curvature regularization, and vice versa. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9078-9086 (2019) 3 +25. Park, G.Y., Lee, S.W.: Reliably fast adversarial training via latent adversarial perturbation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 7758-7767 (2021) 1, 5 +26. Poon, C., Liang, J.: Trajectory of alternating direction method of multipliers and adaptive acceleration. Advances in Neural Information Processing Systems 32 (2019) 2, 5 +27. Qin, C., Martens, J., Gowal, S., Krishnan, D., Dvijotham, K., Fawzi, A., De, S., Stanforth, R., Kohli, P.: Adversarial robustness through local linearization. Advances in Neural Information Processing Systems 32 (2019) 3, 5 +28. Rocamora, E.A., Liu, F., Chrysos, G.G., Olmos, P.M., Cevher, V.: Efficient local linearity regularization to overcome catastrophic overfitting. International Conference on Learning Representations (2024) 1, 5, 6, 11 +29. Shaeiri, A., Nobahari, R., Rohhan, M.H.: Towards deep learning models resistant to large perturbations. arXiv preprint arXiv:2003.13370 (2020) 9 +30. Shafahi, A., Najibi, M., Ghiasi, M.A., Xu, Z., Dickerson, J., Studer, C., Davis, L.S., Taylor, G., Goldstein, T.: Adversarial training for free! Advances in Neural Information Processing Systems 32 (2019) 1, 4 +31. Song, C., He, K., Wang, L., Hopcroft, J.E.: Improving the generalization of adversarial training with domain adaptation. International Conference on Learning Representations (2019) 1 +32. Sriramanan, G., Addepalli, S., Baburaj, A., et al.: Guided adversarial attack for evaluating and enhancing adversarial defenses. Advances in Neural Information Processing Systems 33, 20297-20308 (2020) 5 +33. Sriramanan, G., Addepalli, S., Baburaj, A., et al.: Towards efficient and effective adversarial training. Advances in Neural Information Processing Systems 34, 11821-11833 (2021) 2, 5, 10, 11, 12 +34. Themelis, A., Patrinos, P.: Douglas-rachford splitting and admn for nonconvex optimization: Tight convergence results. SIAM Journal on Optimization 30(1), 149-181 (2020) 5 + +35. Tsiligkaridis, T., Roberts, J.: Understanding and increasing efficiency of frank-wolfe adversarial training. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 50-59 (2022) 5 +36. Wong, E., Rice, L., Kolter, J.Z.: Fast is better than free: Revisiting adversarial training. International Conference on Learning Representations (2020) 1, 2, 4, 6, 9, 10, 11, 12, 13 +37. Wu, D., Xia, S.T., Wang, Y.: Adversarial weight perturbation helps robust generalization. Advances in Neural Information Processing Systems (2020) 9 +38. Xie, Y., Li, Z., Shi, C., Liu, J., Chen, Y., Yuan, B.: Enabling fast and universal audio adversarial attack using generative model. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 35, pp. 14129-14137 (2021) 1 +39. Zagoruyko, S., Komodakis, N.: Wide residual networks. arXiv preprint arXiv:1605.07146 (2016) 8 +40. Zhang, H., Yu, Y., Jiao, J., Xing, E., El Ghaoui, L., Jordan, M.: Theoretically principled trade-off between robustness and accuracy. In: International conference on machine learning. pp. 7472-7482 (2019) 3 +41. Zhang, Y., Zhang, G., Khanduri, P., Hong, M., Chang, S., Liu, S.: Revisiting and advancing fast adversarial training through the lens of bi-level optimization. In: International Conference on Machine Learning. pp. 26693-26712 (2022) 3, 5 +42. Zhao, M., Zhang, L., Kong, Y., Yin, B.: Fast adversarial training with smooth convergence. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 4720-4729 (2023) 2, 6 \ No newline at end of file diff --git a/2024/Rethinking Fast Adversarial Training_ A Splitting Technique To Overcome Catastrophic Overfitting/images.zip b/2024/Rethinking Fast Adversarial Training_ A Splitting Technique To Overcome Catastrophic Overfitting/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..117be27e1b827c41845acec9c906f1daf5daf308 --- /dev/null +++ b/2024/Rethinking Fast Adversarial Training_ A Splitting Technique To Overcome Catastrophic Overfitting/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8cf01220dad69355278532faa70950ef871c75badcc6b6a899c3544c722f2afe +size 504734 diff --git a/2024/Rethinking Fast Adversarial Training_ A Splitting Technique To Overcome Catastrophic Overfitting/layout.json b/2024/Rethinking Fast Adversarial Training_ A Splitting Technique To Overcome Catastrophic Overfitting/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..38efa861dcbddba83b55e70820911d4b3d18ef09 --- /dev/null +++ b/2024/Rethinking Fast Adversarial Training_ A Splitting Technique To Overcome Catastrophic Overfitting/layout.json @@ -0,0 +1,9386 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 134, + 111, + 479, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 111, + 479, + 148 + ], + "spans": [ + { + "bbox": [ + 134, + 111, + 479, + 148 + ], + "type": "text", + "content": "Rethinking Fast Adversarial Training: A Splitting Technique to Overcome Catastrophic Overfitting" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 189, + 168, + 424, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 189, + 168, + 424, + 180 + ], + "spans": [ + { + "bbox": [ + 189, + 168, + 424, + 180 + ], + "type": "text", + "content": "Masoumeh Zareapoor and Pourya Shamsolmoali\\*" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 188, + 190, + 426, + 201 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 188, + 190, + 426, + 201 + ], + "spans": [ + { + "bbox": [ + 188, + 190, + 426, + 201 + ], + "type": "text", + "content": "East China Normal University, Queen's University Belfast" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 160, + 228, + 455, + 439 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 228, + 455, + 439 + ], + "spans": [ + { + "bbox": [ + 160, + 228, + 455, + 439 + ], + "type": "text", + "content": "Abstract. Catastrophic overfitting (CO) poses a significant challenge to fast adversarial training (FastAT), particularly at large perturbation scales, leading to dramatic reductions in adversarial test accuracy. Our analysis of existing FastAT methods shows that CO is accompanied by abrupt and irregular fluctuations in loss convergence, indicating that a stable training dynamic is key to preventing CO. Therefore, we propose a training model that uses the Douglas-Rachford (DR) splitting technique to ensure a balanced and consistent training progression, effectively counteracting CO. The DR splitting technique, known for its ability to solve complex optimization problems, offering a distinct advantage over classical FastAT methods by providing a smoother loss convergence. This is achieved without resorting to complex regularization or incurring the computational costs associated with double backpropagation, presenting an efficient solution to enhance adversarial robustness. Our comprehensive evaluation conducted across standard datasets, demonstrates that our DR splitting-based model not only improves adversarial robustness but also achieves this with remarkable efficiency compared to various FastAT methods. This efficiency is particularly observed under conditions involving long training schedules and large adversarial perturbations." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 160, + 448, + 454, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 448, + 454, + 460 + ], + "spans": [ + { + "bbox": [ + 160, + 448, + 454, + 460 + ], + "type": "text", + "content": "Keywords: Adversarial training " + }, + { + "bbox": [ + 160, + 448, + 454, + 460 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 448, + 454, + 460 + ], + "type": "text", + "content": " Adversarial robustness " + }, + { + "bbox": [ + 160, + 448, + 454, + 460 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 448, + 454, + 460 + ], + "type": "text", + "content": " Convergence" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 479, + 231, + 491 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 479, + 231, + 491 + ], + "spans": [ + { + "bbox": [ + 132, + 479, + 231, + 491 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 503, + 482, + 648 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 503, + 482, + 648 + ], + "spans": [ + { + "bbox": [ + 130, + 503, + 482, + 648 + ], + "type": "text", + "content": "Deep neural networks (DNNs), despite their successes, are vulnerable to adversarial attacks, where small, undetectable, input perturbations can mislead models into incorrect predictions [11]. Given the security risks of such vulnerabilities, improving neural network robustness against these attacks has become a critical area of research [28, 38]. Adversarial training (AT) [18, 30] and its derivative, PGD-AT [22, 31], have emerged as effective strategies for enhancing the robustness of neural networks. However, the complex inner optimizations required by these methods at each training step significantly elevate computational costs, limiting their practicality for large-scale applications [28]. To address these computational challenges, 'fast' versions of AT (FastAT) [25, 36] have been proposed, which employ a single-step gradient propagation [11] to produce adversarial training examples. While FastAT offers a more efficient training process" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 653, + 482, + 666 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 653, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 133, + 653, + 482, + 666 + ], + "type": "text", + "content": "* Correspondence to pshams55@gmail.com. Code: https://github.com/mzarea2/RAT" + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 272 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 272 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 272 + ], + "type": "text", + "content": "compared to PGD-AT, it introduces the problem of Catastrophic Overfitting (CO) [16,36]. CO occurs when training with strong adversaries, leading to a significant drop in the robustness of the resulting model. Several FastAT methods have been developed to reduce CO [42]. Wong et al. [36] enhance adversarial example diversity through the use of initial perturbations that are randomly initialized. Following this approach, [3] proposed FAST-GA, an efficient regularizer that aims to explicitly align gradients between benign and adversarial inputs. Likewise, MEP-PGI [15] and NuAT [33] employ nuclear or p-norm regularization within adversarial training to improve consistency in predictions on both benign and adversarial examples. These approaches are effective only within a limited perturbation " + }, + { + "bbox": [ + 130, + 116, + 482, + 272 + ], + "type": "inline_equation", + "content": "\\mathcal{E} \\leq 8/255" + }, + { + "bbox": [ + 130, + 116, + 482, + 272 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 116, + 482, + 272 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 130, + 116, + 482, + 272 + ], + "type": "text", + "content": " denotes the degree of perturbations in the adversarial training [42]. Moreover, models trained with relatively small perturbation scales show vulnerability to larger " + }, + { + "bbox": [ + 130, + 116, + 482, + 272 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 130, + 116, + 482, + 272 + ], + "type": "text", + "content": ", with noted declines in robustness." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 272, + 482, + 379 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 272, + 482, + 379 + ], + "spans": [ + { + "bbox": [ + 130, + 272, + 482, + 379 + ], + "type": "text", + "content": "For example, models trained with NuAT and MEP-PGI at a perturbation size of " + }, + { + "bbox": [ + 130, + 272, + 482, + 379 + ], + "type": "inline_equation", + "content": "\\mathcal{E} = 8 / 255" + }, + { + "bbox": [ + 130, + 272, + 482, + 379 + ], + "type": "text", + "content": " demonstrated robustness levels of " + }, + { + "bbox": [ + 130, + 272, + 482, + 379 + ], + "type": "inline_equation", + "content": "56\\%" + }, + { + "bbox": [ + 130, + 272, + 482, + 379 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 272, + 482, + 379 + ], + "type": "inline_equation", + "content": "55\\%" + }, + { + "bbox": [ + 130, + 272, + 482, + 379 + ], + "type": "text", + "content": " against PGD-50 attacks, but their robustness significantly dropped to " + }, + { + "bbox": [ + 130, + 272, + 482, + 379 + ], + "type": "inline_equation", + "content": "24\\%" + }, + { + "bbox": [ + 130, + 272, + 482, + 379 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 272, + 482, + 379 + ], + "type": "inline_equation", + "content": "22\\%" + }, + { + "bbox": [ + 130, + 272, + 482, + 379 + ], + "type": "text", + "content": ", respectively, with " + }, + { + "bbox": [ + 130, + 272, + 482, + 379 + ], + "type": "inline_equation", + "content": "\\mathcal{E} = 16 / 255" + }, + { + "bbox": [ + 130, + 272, + 482, + 379 + ], + "type": "text", + "content": ". Our investigation into adversarial training processes (illustrated in Fig. 2) reveals that CO is characterized by minor fluctuations in benign classification loss and significant decreases in adversarial example loss. This observation raises a question: How can we design fast adversarial training that ensures improved stability and mitigates CO, while also optimizing the trade-off between robustness accuracy (RA) and standard accuracy (SA)?" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 380, + 482, + 656 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 380, + 482, + 656 + ], + "spans": [ + { + "bbox": [ + 130, + 380, + 482, + 656 + ], + "type": "text", + "content": "In this context, the Douglas-Rachford (DR) splitting technique emerges as a promising solution. The DR splitting is known for its linear convergence across both strongly convex and non-convex settings [4, 20, 23, 26]. Several studies demonstrate its applicability in a variety of optimization tasks. For example, Liang et al. [19] and Applegate et al. [4] have highlighted its capability for local and global linear convergence, underscoring the DR technique's effectiveness in complex optimization scenarios. We also capitalize on the inherent stability and efficiency of the DR splitting, incorporating these properties into adversarial training to address CO. This integration allows for a more stable and consistent adversarial learning process, significantly lowering the risk of CO. Indeed, our implementation demonstrates that DR splitting not only stabilizes the training process but also offers a cost-effective solution to enhance model robustness against adversarial attacks, as shown in Fig. 1. Our contributions are threefold: ① We provide a comprehensive analysis showing that existing FastAT methods are still prone to CO with larger " + }, + { + "bbox": [ + 130, + 380, + 482, + 656 + ], + "type": "inline_equation", + "content": "\\varepsilon" + }, + { + "bbox": [ + 130, + 380, + 482, + 656 + ], + "type": "text", + "content": ". Our investigation emphasizes the critical role of loss function convergence stability in addressing this vulnerability. ② Introducing a novel solution with the Douglas-Rachford splitting technique, specifically designed to mitigate CO without incurring computational complexities. ③ Conduct extensive experiments on standard datasets to validate the effectiveness of our model- Rethinking fast Adversarial Training (RAT)- particularly against large adversarial perturbations. RAT, leveraging the DR-splitting technique, shows enhanced adversarial robustness when compared to classical adversarial training methods (find the training code in the supplementary file)." + } + ] + } + ], + "index": 4 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 233, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 233, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 233, + 102 + ], + "type": "text", + "content": "Zareapoor et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 118, + 245, + 204 + ], + "blocks": [ + { + "bbox": [ + 133, + 118, + 245, + 204 + ], + "lines": [ + { + "bbox": [ + 133, + 118, + 245, + 204 + ], + "spans": [ + { + "bbox": [ + 133, + 118, + 245, + 204 + ], + "type": "image", + "image_path": "4475771b46e5295fa31d61cffa4f7fc85ab252215ffc9af798a910b05cb37ac6.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 216, + 482, + 262 + ], + "lines": [ + { + "bbox": [ + 130, + 216, + 482, + 262 + ], + "spans": [ + { + "bbox": [ + 130, + 216, + 482, + 262 + ], + "type": "text", + "content": "Fig. 1: Performance comparison of adversarial robustness with WideResNet34-10 on CIFAR10 for " + }, + { + "bbox": [ + 130, + 216, + 482, + 262 + ], + "type": "inline_equation", + "content": "\\mathcal{E} = (8,16) / 255" + }, + { + "bbox": [ + 130, + 216, + 482, + 262 + ], + "type": "text", + "content": ". (left) Illustrates the average runtime for each method. (right) We compare AA (AutoAttack) and SA accuracies. Lower AA indicates higher vulnerability. Higher SA (standard accuracy) indicates better general performance." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 250, + 114, + 477, + 196 + ], + "blocks": [ + { + "bbox": [ + 250, + 114, + 477, + 196 + ], + "lines": [ + { + "bbox": [ + 250, + 114, + 477, + 196 + ], + "spans": [ + { + "bbox": [ + 250, + 114, + 477, + 196 + ], + "type": "table", + "html": "
MethodsAA (%)SA (%)AA (%)SA (%)
816
AT PGD-1047.69±(0.11)78.51±(0.08)24.97±(0.26)59.41±(0.46)
LLR [27]43.84±(0.20)74.52±(0.07)16.31±(0.20)44.76±(4.72)
Fast-Bat [41]45.83±(0.12)79.61±(0.12)18.30±(0.26)59.34±(0.25)
Fast-GA [3]46.37±(0.27)79.12±(0.07)15.36±(1.32)54.25±(6.28)
CURE [24]44.65±(0.17)77.04±(0.11)18.12±(0.35)53.49±(0.08)
RAT (Ours)46.21±(0.08)80.32±(0.12)19.58±(0.17)60.37±(0.21)
", + "image_path": "c2f83e2c742a76b1c0fd1017ac7469a95fe4b68ab6770018cd4e5a55150dc792.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 281, + 237, + 294 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 281, + 237, + 294 + ], + "spans": [ + { + "bbox": [ + 132, + 281, + 237, + 294 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 304, + 261, + 316 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 304, + 261, + 316 + ], + "spans": [ + { + "bbox": [ + 132, + 304, + 261, + 316 + ], + "type": "text", + "content": "2.1 Adversarial Attacks" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 322, + 479, + 395 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 322, + 479, + 395 + ], + "spans": [ + { + "bbox": [ + 130, + 322, + 479, + 395 + ], + "type": "text", + "content": "Adversarial attacks aim to fool deep learning models by introducing perturbed inputs. Goodfellow et al. [11] introduced the concept of the fast gradient sign method (FGSM) for generating adversarial examples " + }, + { + "bbox": [ + 130, + 322, + 479, + 395 + ], + "type": "inline_equation", + "content": "x'" + }, + { + "bbox": [ + 130, + 322, + 479, + 395 + ], + "type": "text", + "content": " within classification tasks. They demonstrated that adversarial examples, created through a single-step gradient update, can lead a classification model to incorrect predictions with significant confidence. The adversarial examples " + }, + { + "bbox": [ + 130, + 322, + 479, + 395 + ], + "type": "inline_equation", + "content": "x'" + }, + { + "bbox": [ + 130, + 322, + 479, + 395 + ], + "type": "text", + "content": " are generated by" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 236, + 400, + 481, + 415 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 236, + 400, + 481, + 415 + ], + "spans": [ + { + "bbox": [ + 236, + 400, + 481, + 415 + ], + "type": "interline_equation", + "content": "x ^ {\\prime} = x + \\mathcal {E}. \\operatorname {s g n} \\left(\\nabla_ {x} L (f (x; \\theta), y\\right), \\tag {1}", + "image_path": "9d6a4a4ec744f7e3c135eb61d06d9355dd91a72e6691a4e4272cdb16c227074f.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 420, + 482, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 420, + 482, + 515 + ], + "spans": [ + { + "bbox": [ + 130, + 420, + 482, + 515 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 420, + 482, + 515 + ], + "type": "inline_equation", + "content": "(x,y)" + }, + { + "bbox": [ + 130, + 420, + 482, + 515 + ], + "type": "text", + "content": " denotes an input sample and its corresponding label, " + }, + { + "bbox": [ + 130, + 420, + 482, + 515 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 130, + 420, + 482, + 515 + ], + "type": "text", + "content": " is the perturbation magnitude, " + }, + { + "bbox": [ + 130, + 420, + 482, + 515 + ], + "type": "inline_equation", + "content": "\\nabla_{x}L" + }, + { + "bbox": [ + 130, + 420, + 482, + 515 + ], + "type": "text", + "content": " is the loss gradient with respect to " + }, + { + "bbox": [ + 130, + 420, + 482, + 515 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 130, + 420, + 482, + 515 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 420, + 482, + 515 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 130, + 420, + 482, + 515 + ], + "type": "text", + "content": " signifies the fixed model weights. Subsequent research has developed more sophisticated attack methods, such as I-FGSM [17], MIM [9], and PGD [22], which leverage iterative gradient descent to enhance the efficacy of the attacks. The PGD attack, in particular, generates adversarial examples by applying small perturbations to the input data, ensuring these adjustments stay within a pre-defined boundary " + }, + { + "bbox": [ + 130, + 420, + 482, + 515 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 130, + 420, + 482, + 515 + ], + "type": "text", + "content": ". This process is described as follows" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 214, + 522, + 481, + 537 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 214, + 522, + 481, + 537 + ], + "spans": [ + { + "bbox": [ + 214, + 522, + 481, + 537 + ], + "type": "interline_equation", + "content": "x _ {t + 1} ^ {\\prime} = \\Pi_ {\\mathcal {E}} (x _ {t} ^ {\\prime} + \\alpha . \\mathrm {s g n} (\\nabla_ {x} L (f (x _ {t} ^ {\\prime}; \\theta), y))), \\qquad \\qquad (2)", + "image_path": "97bf7adc6e84019575eb95b8f8e8f5e67348ff5cea15dcccc0de48d6e8c8a47c.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 543, + 482, + 591 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 543, + 482, + 591 + ], + "spans": [ + { + "bbox": [ + 130, + 543, + 482, + 591 + ], + "type": "text", + "content": "in which " + }, + { + "bbox": [ + 130, + 543, + 482, + 591 + ], + "type": "inline_equation", + "content": "x'" + }, + { + "bbox": [ + 130, + 543, + 482, + 591 + ], + "type": "text", + "content": " is the adversarial example at iteration " + }, + { + "bbox": [ + 130, + 543, + 482, + 591 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 543, + 482, + 591 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 543, + 482, + 591 + ], + "type": "inline_equation", + "content": "\\Pi_{\\mathcal{E}}" + }, + { + "bbox": [ + 130, + 543, + 482, + 591 + ], + "type": "text", + "content": " denotes the projection operation that ensures the adversarial example remains within the " + }, + { + "bbox": [ + 130, + 543, + 482, + 591 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 130, + 543, + 482, + 591 + ], + "type": "text", + "content": "-ball. The PGD attack is one of the most effective attack methods for evaluating the robustness of deep learning against adversarial examples." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 599, + 293, + 612 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 599, + 293, + 612 + ], + "spans": [ + { + "bbox": [ + 132, + 599, + 293, + 612 + ], + "type": "text", + "content": "2.2 Adversarial Training (AT)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 617, + 481, + 642 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 481, + 642 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 481, + 642 + ], + "type": "text", + "content": "Adversarial Training (AT) [22] and TRADES [40] have become prominent training methods for training robust architectures, which can be written as" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 219, + 648, + 481, + 667 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 219, + 648, + 481, + 667 + ], + "spans": [ + { + "bbox": [ + 219, + 648, + 481, + 667 + ], + "type": "interline_equation", + "content": "\\min _ {\\theta} E _ {(x, y) \\sim D} [ \\max _ {\\delta \\in [ - \\mathcal {E}, \\mathcal {E} ]} L (f (x + \\delta ; \\theta), y) ], \\tag {3}", + "image_path": "68134e5ad0f7ea58c6969ee780871e1584ee31046d4edd4490a4e93096c23d9c.jpg" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 294, + 91, + 448, + 102 + ], + "type": "text", + "content": "Rethinking Fast Adversarial Training" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 143, + 115, + 471, + 314 + ], + "blocks": [ + { + "bbox": [ + 143, + 115, + 471, + 314 + ], + "lines": [ + { + "bbox": [ + 143, + 115, + 471, + 314 + ], + "spans": [ + { + "bbox": [ + 143, + 115, + 471, + 314 + ], + "type": "image", + "image_path": "536e01bcbba3d09af02e5802bc6c636c6f872ceea42053455dcc8c68db8793e5.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 323, + 482, + 390 + ], + "lines": [ + { + "bbox": [ + 130, + 323, + 482, + 390 + ], + "spans": [ + { + "bbox": [ + 130, + 323, + 482, + 390 + ], + "type": "text", + "content": "Fig. 2: Performance of fast adversarial training (FastAT) methods on CIFAR10 using ResNet20, for " + }, + { + "bbox": [ + 130, + 323, + 482, + 390 + ], + "type": "inline_equation", + "content": "\\mathcal{E} = 16 / 255" + }, + { + "bbox": [ + 130, + 323, + 482, + 390 + ], + "type": "text", + "content": ". The figure illustrates the training process across three runs. The terms AV-loss and BN-loss are the training loss for adversarial and benign inputs, respectively, while AV-acc and BN-acc quantify testing accuracy against both adversarial and benign samples. The occurrence of catastrophic overfitting is marked (in green), indicating critical periods where adversarial robustness drops." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 414, + 482, + 583 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 414, + 482, + 583 + ], + "spans": [ + { + "bbox": [ + 130, + 414, + 482, + 583 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 414, + 482, + 583 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 130, + 414, + 482, + 583 + ], + "type": "text", + "content": " is the adversarial perturbations and " + }, + { + "bbox": [ + 130, + 414, + 482, + 583 + ], + "type": "inline_equation", + "content": "x' = x + \\delta" + }, + { + "bbox": [ + 130, + 414, + 482, + 583 + ], + "type": "text", + "content": " is the adversarial examples generated by various attacks such as PGD and FGSM. This optimization first, generates adversarial examples within a defined perturbation limit to maximize classification loss (inner maximization) and then adjusts the model parameters " + }, + { + "bbox": [ + 130, + 414, + 482, + 583 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 130, + 414, + 482, + 583 + ], + "type": "text", + "content": " to minimize this loss over the adversarial examples (outer minimization). However, a notable trade-off emerges in AT between computational efficiency and adversarial robustness, due to the complexity of solving the inner maximization problem. To enhance the computational efficiency of AT, several methods have been developed. Free AT [30], simultaneously updating the network parameters " + }, + { + "bbox": [ + 130, + 414, + 482, + 583 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 130, + 414, + 482, + 583 + ], + "type": "text", + "content": " and the adversarial perturbation " + }, + { + "bbox": [ + 130, + 414, + 482, + 583 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 130, + 414, + 482, + 583 + ], + "type": "text", + "content": " multiple times within a single batch. On the other hand, Fast-AT [36], simplifies the process further by employing a single-step FGSM attack [11] but the attack initiates from a randomly chosen point " + }, + { + "bbox": [ + 130, + 414, + 482, + 583 + ], + "type": "inline_equation", + "content": "x + \\mu" + }, + { + "bbox": [ + 130, + 414, + 482, + 583 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 130, + 414, + 482, + 583 + ], + "type": "inline_equation", + "content": "\\mu \\in [-\\delta, \\delta]" + }, + { + "bbox": [ + 130, + 414, + 482, + 583 + ], + "type": "text", + "content": ". Fast-Bat These approaches reduce the cost of solving the inner maximization problem, significantly accelerating the training process." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 594, + 313, + 608 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 594, + 313, + 608 + ], + "spans": [ + { + "bbox": [ + 132, + 594, + 313, + 608 + ], + "type": "text", + "content": "2.3 Catastrophic Overfitting (CO)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "The acceleration of AT methods comes with significant drawbacks. Wong et al. [36] first reported this with FGSM training, and was later supported by Andriushchenko et al. [3] in their analysis of Free AT and Fast-AT methods. In these cases, the model's accuracy on adversarial tests can fall sharply, sometimes" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 92, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 92, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 92, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 232, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 232, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 232, + 102 + ], + "type": "text", + "content": "Zareapoor et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 165 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 165 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 165 + ], + "type": "text", + "content": "to " + }, + { + "bbox": [ + 130, + 116, + 482, + 165 + ], + "type": "inline_equation", + "content": "0\\%" + }, + { + "bbox": [ + 130, + 116, + 482, + 165 + ], + "type": "text", + "content": ", in just a few training epochs. This decline in performance has been attributed to CO, a critical challenge that has motivated significant research efforts aimed at understanding and mitigating its impact [14, 18, 41]. Below, we provide an overview of various strategies developed to avoid CO." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 180, + 328, + 192 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 180, + 328, + 192 + ], + "spans": [ + { + "bbox": [ + 131, + 180, + 328, + 192 + ], + "type": "text", + "content": "2.4 Regularization-based Approaches" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 205, + 482, + 385 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 205, + 482, + 385 + ], + "spans": [ + { + "bbox": [ + 130, + 205, + 482, + 385 + ], + "type": "text", + "content": "Recent efforts focus on regularization and data augmentation to prevent CO. [3] introduced FAST-GA to mitigate CO by penalizing the misalignment of gradients between the input and their adversarial counterparts. This method ensures that the direction of the gradient remains consistent when slight perturbations are introduced to the input. Qin et al. [27] introduced local linear regularization (LLR) to improve model robustness against adversarial attacks. LLR smoothes the loss landscape using a first-order Taylor approximation, addressing nonlinear sensitivities and promoting a more linear loss response to perturbations. ELLE [28] introduced a regularization term designed to mitigate CO both effectively and efficiently within classical AT frameworks, by enforcing local linearity. Sriramanan et al. [32] introduced GAT, a novel regularization strategy aimed at reducing the difference between the logits of a training point and a random sample. NuAT [33], proposed a regularization term that penalizes the nuclear norm of the logits difference between benign and adversarial examples. However, GAT and NuAT still face challenges with CO under larger perturbation size " + }, + { + "bbox": [ + 130, + 205, + 482, + 385 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 130, + 205, + 482, + 385 + ], + "type": "text", + "content": " [16]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 401, + 253, + 414 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 401, + 253, + 414 + ], + "spans": [ + { + "bbox": [ + 131, + 401, + 253, + 414 + ], + "type": "text", + "content": "2.5 Other Approaches" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 426, + 482, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 426, + 482, + 668 + ], + "spans": [ + { + "bbox": [ + 130, + 426, + 482, + 668 + ], + "type": "text", + "content": "Park et al. [25] introduced SLAT which performs adversarial attacks within the latent space, focusing on the internal representations rather than the input space directly. Another study [35] employs the Frank-Wolfe optimization technique for AT, connecting the distortion observed in the loss landscape to the norm of adversarial perturbations, providing a deeper understanding of adversarial robustness. Li et al. [18] addressed CO by limiting network weights to a specific subspace based on the main components of AT-trained weights. Meanwhile, Zhang et al. [41] introduce Fast-Bat, which redefines the AT challenge as a Bilevel optimization problem and introduces a one-step solution to address it. However, [28] shows that Fast-BAT is affected by CO for " + }, + { + "bbox": [ + 130, + 426, + 482, + 668 + ], + "type": "inline_equation", + "content": "\\mathcal{E} > 10 / 255" + }, + { + "bbox": [ + 130, + 426, + 482, + 668 + ], + "type": "text", + "content": ". Zero-Grad [10] introduces a strategy to mitigate adversarial perturbations by setting to zero the gradient components that fall below a threshold, however, [16] demonstrates, this approach results in decreased performance for large " + }, + { + "bbox": [ + 130, + 426, + 482, + 668 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 130, + 426, + 482, + 668 + ], + "type": "text", + "content": ". Different from these methods, we aim to minimize the loss fluctuations during training through the Douglas-Rachford splitting technique, known for its effectiveness in complex optimization [23]. It has been successfully applied to optimal transport [20] and its discretizations [6], demonstrating its flexibility and robustness in achieving linear convergence rates under both convex and non-convex settings [4, 34]. A notable advantage of this splitting technique is to leverage sparsity in solutions, which can significantly accelerate the convergence process [26]." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 295, + 91, + 447, + 102 + ], + "type": "text", + "content": "Rethinking Fast Adversarial Training" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 202, + 127 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 202, + 127 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 202, + 127 + ], + "type": "text", + "content": "3 Method" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 138, + 479, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 138, + 479, + 163 + ], + "spans": [ + { + "bbox": [ + 130, + 138, + 479, + 163 + ], + "type": "text", + "content": "This section evaluates FastAT methods under large perturbation budgets, " + }, + { + "bbox": [ + 130, + 138, + 479, + 163 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 130, + 138, + 479, + 163 + ], + "type": "text", + "content": ", explores the causes of CO, and finally introduces our proposed solutions." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 174, + 437, + 186 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 174, + 437, + 186 + ], + "spans": [ + { + "bbox": [ + 131, + 174, + 437, + 186 + ], + "type": "text", + "content": "3.1 Performance of Fast Adversarial Training on a Large " + }, + { + "bbox": [ + 131, + 174, + 437, + 186 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 193, + 480, + 253 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 193, + 480, + 253 + ], + "spans": [ + { + "bbox": [ + 130, + 193, + 480, + 253 + ], + "type": "text", + "content": "Fast AT methods have shown effectiveness in mitigating CO at lower perturbation " + }, + { + "bbox": [ + 130, + 193, + 480, + 253 + ], + "type": "inline_equation", + "content": "(\\mathcal{E} \\leq 8/255)" + }, + { + "bbox": [ + 130, + 193, + 480, + 253 + ], + "type": "text", + "content": " [28, 36]. To extend this evaluation, we analyze performance at a higher perturbation level " + }, + { + "bbox": [ + 130, + 193, + 480, + 253 + ], + "type": "inline_equation", + "content": "(\\mathcal{E} = 16/255)" + }, + { + "bbox": [ + 130, + 193, + 480, + 253 + ], + "type": "text", + "content": ", in line with [42]. FAST-AT [36] introduces an initial step for Eq. (3), where samples are perturbed with uniform random noise " + }, + { + "bbox": [ + 130, + 193, + 480, + 253 + ], + "type": "inline_equation", + "content": "\\delta_0 \\sim U(-\\mathcal{E}, \\mathcal{E})" + }, + { + "bbox": [ + 130, + 193, + 480, + 253 + ], + "type": "text", + "content": " before applying the FGSM attack," + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 194, + 260, + 481, + 277 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 194, + 260, + 481, + 277 + ], + "spans": [ + { + "bbox": [ + 194, + 260, + 481, + 277 + ], + "type": "interline_equation", + "content": "\\min _ {\\theta} E _ {(x, y) \\sim D} \\left[ \\max _ {\\delta_ {0} + \\delta \\in [ - \\mathcal {E}, \\mathcal {E} ]} L (f (x + \\delta_ {0} + \\delta ; \\theta), y) \\right]. \\tag {4}", + "image_path": "28680a2531b8c1d6e76395712bb85430d301551c9b0e7b090b394ff3aeecde4c.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 285, + 482, + 334 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 285, + 482, + 334 + ], + "spans": [ + { + "bbox": [ + 131, + 285, + 482, + 334 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 131, + 285, + 482, + 334 + ], + "type": "inline_equation", + "content": "L(\\cdot)" + }, + { + "bbox": [ + 131, + 285, + 482, + 334 + ], + "type": "text", + "content": " is the loss function, " + }, + { + "bbox": [ + 131, + 285, + 482, + 334 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 131, + 285, + 482, + 334 + ], + "type": "text", + "content": " represents adversarial perturbations, and " + }, + { + "bbox": [ + 131, + 285, + 482, + 334 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 131, + 285, + 482, + 334 + ], + "type": "text", + "content": " is the model weights. FAST-GA, as introduced by [3] focuses on aligning the gradients of the loss with respect to both benign samples " + }, + { + "bbox": [ + 131, + 285, + 482, + 334 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 131, + 285, + 482, + 334 + ], + "type": "text", + "content": " and their perturbed samples " + }, + { + "bbox": [ + 131, + 285, + 482, + 334 + ], + "type": "inline_equation", + "content": "x + \\delta_0" + }, + { + "bbox": [ + 131, + 285, + 482, + 334 + ], + "type": "text", + "content": ", which is described as," + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 202, + 342, + 480, + 354 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 202, + 342, + 480, + 354 + ], + "spans": [ + { + "bbox": [ + 202, + 342, + 480, + 354 + ], + "type": "interline_equation", + "content": "E _ {(x, y) \\sim D} [ 1 - \\cos (\\nabla_ {x} L (x, \\theta), \\nabla_ {x + \\delta_ {0}} L (x + \\delta_ {0}, \\theta)) ], \\tag {5}", + "image_path": "39dd15a7717e3fceada4f25fe60cc6ef287195a2273f43406d30ed79be944267.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 361, + 482, + 458 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 361, + 482, + 458 + ], + "spans": [ + { + "bbox": [ + 130, + 361, + 482, + 458 + ], + "type": "text", + "content": "in which " + }, + { + "bbox": [ + 130, + 361, + 482, + 458 + ], + "type": "inline_equation", + "content": "L(x,\\theta) = L(f(x;\\theta),y)" + }, + { + "bbox": [ + 130, + 361, + 482, + 458 + ], + "type": "text", + "content": ". The cosine similarity " + }, + { + "bbox": [ + 130, + 361, + 482, + 458 + ], + "type": "inline_equation", + "content": "\\cos(\\cdot)" + }, + { + "bbox": [ + 130, + 361, + 482, + 458 + ], + "type": "text", + "content": " measures the cosine similarity between the gradients of the loss function for the benign and perturbed inputs. This strategy aims to ensure that the model's behavior does not change drastically for adversarial examples, thereby reducing CO. Different from other attack methods, MEP-PGI [15] extends the FGSM attack by incorporating a Prior-Guided Initialization (PGI) mechanism. This technique initializes the adversarial perturbations " + }, + { + "bbox": [ + 130, + 361, + 482, + 458 + ], + "type": "inline_equation", + "content": "\\delta_0" + }, + { + "bbox": [ + 130, + 361, + 482, + 458 + ], + "type": "text", + "content": ", by using the information from adversarial examples generated in previous training epochs," + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 230, + 466, + 480, + 479 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 230, + 466, + 480, + 479 + ], + "spans": [ + { + "bbox": [ + 230, + 466, + 480, + 479 + ], + "type": "interline_equation", + "content": "E _ {(x, y) \\sim D} [ \\| f (x ^ {\\prime}; \\theta) - f (x + \\delta_ {0}; \\theta) \\| _ {2} ^ {2} ], \\tag {6}", + "image_path": "6b6187f2617ca94d7e18c80983542aaf90af385796895ea9aeaeae36df5c670d.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 486, + 482, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 486, + 482, + 668 + ], + "spans": [ + { + "bbox": [ + 130, + 486, + 482, + 668 + ], + "type": "text", + "content": "here, " + }, + { + "bbox": [ + 130, + 486, + 482, + 668 + ], + "type": "inline_equation", + "content": "||\\cdot||" + }, + { + "bbox": [ + 130, + 486, + 482, + 668 + ], + "type": "text", + "content": " is the squared (L2) norm, measuring the distance between the model's outputs for the two types of examples. FAST-PGI aims to enrich the adversarial examples with historical insight, thereby enhancing the adversarial training process. Fig. 2 illustrates the adversarial training of ResNet20 on CIFAR10, highlighting times when CO occurs, between epochs 5th-20th, 40th-65th, and 10th-15th. The key observations from the results are i) transition from smooth/stable to erratic loss fluctuations; ii) discrepancies in the loss for benign and adversarial examples, i.e., minor fluctuations in benign loss " + }, + { + "bbox": [ + 130, + 486, + 482, + 668 + ], + "type": "inline_equation", + "content": "L(x,\\theta)" + }, + { + "bbox": [ + 130, + 486, + 482, + 668 + ], + "type": "text", + "content": " followed with a rapid decline in adversarial loss " + }, + { + "bbox": [ + 130, + 486, + 482, + 668 + ], + "type": "inline_equation", + "content": "L(x',\\theta)" + }, + { + "bbox": [ + 130, + 486, + 482, + 668 + ], + "type": "text", + "content": "; and iii) a sharp decrease in adversarial accuracy. The result is consistent with the observations in [1,28,42]. From these patterns, we derive several insights for understanding CO: it is obvious that " + }, + { + "bbox": [ + 130, + 486, + 482, + 668 + ], + "type": "inline_equation", + "content": "L(x,\\theta)" + }, + { + "bbox": [ + 130, + 486, + 482, + 668 + ], + "type": "text", + "content": " has greater stability than the loss " + }, + { + "bbox": [ + 130, + 486, + 482, + 668 + ], + "type": "inline_equation", + "content": "L(x',\\theta)" + }, + { + "bbox": [ + 130, + 486, + 482, + 668 + ], + "type": "text", + "content": ", indicating a tendency of models to overfit to adversarial samples. CO directly correlates with the instability in adversarial training convergence. Anomalies in both benign and adversarial loss occur together, signifying an issue in the learning process that affects its performance" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 233, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 233, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 233, + 102 + ], + "type": "text", + "content": "Zareapoor et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "content": "on both data. We also find that Fast-GA and MEP-PGI enhance the stability of adversarial training through the constraints in Eq. (4) and Eq. (5), compared to the Fast-AT. We are now in shape to give a formal definition of our model." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 166, + 452, + 178 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 166, + 452, + 178 + ], + "spans": [ + { + "bbox": [ + 131, + 166, + 452, + 178 + ], + "type": "text", + "content": "3.2 Douglas-Rachford (DR) Splitting for Adversarial Training" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 182, + 480, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 182, + 480, + 327 + ], + "spans": [ + { + "bbox": [ + 130, + 182, + 480, + 327 + ], + "type": "text", + "content": "We use the DR splitting technique to ensure a consistent and stable optimization to prevent sudden changes in adversarial robustness. This technique divides the optimization process into a two-step process (reflection and enforcement), which together ensure a stable and consistent optimization path. Let's denote the model at iteration " + }, + { + "bbox": [ + 130, + 182, + 480, + 327 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 182, + 480, + 327 + ], + "type": "text", + "content": " by " + }, + { + "bbox": [ + 130, + 182, + 480, + 327 + ], + "type": "inline_equation", + "content": "f(\\cdot, \\theta^{(k)})" + }, + { + "bbox": [ + 130, + 182, + 480, + 327 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 182, + 480, + 327 + ], + "type": "inline_equation", + "content": "\\theta^{(k)}" + }, + { + "bbox": [ + 130, + 182, + 480, + 327 + ], + "type": "text", + "content": " represents the model parameters. The loss function for adversarial examples is " + }, + { + "bbox": [ + 130, + 182, + 480, + 327 + ], + "type": "inline_equation", + "content": "L(f(x', \\theta^{(k)}, y))" + }, + { + "bbox": [ + 130, + 182, + 480, + 327 + ], + "type": "text", + "content": ", and for benign examples, " + }, + { + "bbox": [ + 130, + 182, + 480, + 327 + ], + "type": "inline_equation", + "content": "L(f(x^{k}, \\theta^{(k)}, y))" + }, + { + "bbox": [ + 130, + 182, + 480, + 327 + ], + "type": "text", + "content": ". An important property of our model is its ability to alternate between reducing adversarial loss and applying careful regularization terms. This ensures that any adjustments to the model parameters effectively impact both benign and adversarial loss functions " + }, + { + "bbox": [ + 130, + 182, + 480, + 327 + ], + "type": "inline_equation", + "content": "L(x, \\theta)" + }, + { + "bbox": [ + 130, + 182, + 480, + 327 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 182, + 480, + 327 + ], + "type": "inline_equation", + "content": "L(x', \\theta)" + }, + { + "bbox": [ + 130, + 182, + 480, + 327 + ], + "type": "text", + "content": ". Specifically, to address CO, we introduce epoch-based constraints on the loss differences for both adversarial and benign instances, represented as " + }, + { + "bbox": [ + 130, + 182, + 480, + 327 + ], + "type": "inline_equation", + "content": "(\\Delta L_{\\theta^{(k)}}, \\Delta L_{\\theta^{(k)}}^{benign})" + }, + { + "bbox": [ + 130, + 182, + 480, + 327 + ], + "type": "text", + "content": "," + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 189, + 332, + 480, + 365 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 189, + 332, + 480, + 365 + ], + "spans": [ + { + "bbox": [ + 189, + 332, + 480, + 365 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\Delta L _ {\\theta^ {(k)}} = \\left| L \\left(f \\left(x ^ {\\prime (k)}, \\theta^ {(k)}\\right), y\\right) - L \\left(f \\left(x ^ {\\prime (k - 1)}, \\theta^ {(k - 1)}\\right), y\\right) \\right| \\\\ \\begin{array}{l} \\left. \\right| = \\left| L (f (w, \\theta^ {(k)}), y) - L (f (x, \\theta^ {(k - 1)}), y) \\right|, \\end{array} \\tag {7} \\\\ \\end{array}", + "image_path": "e788d78cb32513533d5d2a1a9ec7a67fbad6f443bdb37039294668e135ae406c.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 369, + 480, + 487 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 369, + 480, + 487 + ], + "spans": [ + { + "bbox": [ + 130, + 369, + 480, + 487 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 369, + 480, + 487 + ], + "type": "inline_equation", + "content": "|.|" + }, + { + "bbox": [ + 130, + 369, + 480, + 487 + ], + "type": "text", + "content": " measures the absolute difference in loss for adversarial " + }, + { + "bbox": [ + 130, + 369, + 480, + 487 + ], + "type": "inline_equation", + "content": "(x')" + }, + { + "bbox": [ + 130, + 369, + 480, + 487 + ], + "type": "text", + "content": " and benign " + }, + { + "bbox": [ + 130, + 369, + 480, + 487 + ], + "type": "inline_equation", + "content": "(x)" + }, + { + "bbox": [ + 130, + 369, + 480, + 487 + ], + "type": "text", + "content": " inputs, " + }, + { + "bbox": [ + 130, + 369, + 480, + 487 + ], + "type": "inline_equation", + "content": "x'^{(k)} = x + \\delta_{0,k} + \\delta_k" + }, + { + "bbox": [ + 130, + 369, + 480, + 487 + ], + "type": "text", + "content": ". Here, " + }, + { + "bbox": [ + 130, + 369, + 480, + 487 + ], + "type": "inline_equation", + "content": "\\delta_{0,k}" + }, + { + "bbox": [ + 130, + 369, + 480, + 487 + ], + "type": "text", + "content": " can be replaced by various attack initialization perturbations and " + }, + { + "bbox": [ + 130, + 369, + 480, + 487 + ], + "type": "inline_equation", + "content": "\\delta_k" + }, + { + "bbox": [ + 130, + 369, + 480, + 487 + ], + "type": "text", + "content": " is generated by " + }, + { + "bbox": [ + 130, + 369, + 480, + 487 + ], + "type": "inline_equation", + "content": "[\\max_{\\delta_k \\in [-\\varepsilon, \\varepsilon]} L(x'^{(k)}, \\theta^{(k-1)})]" + }, + { + "bbox": [ + 130, + 369, + 480, + 487 + ], + "type": "text", + "content": ". This formulation guides the model to effectively manage perturbations and maintain a stable training process. Additionally, these constraints help in mitigating the simultaneous anomalies in both " + }, + { + "bbox": [ + 130, + 369, + 480, + 487 + ], + "type": "inline_equation", + "content": "L(x, \\theta)" + }, + { + "bbox": [ + 130, + 369, + 480, + 487 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 369, + 480, + 487 + ], + "type": "inline_equation", + "content": "L(x', \\theta)" + }, + { + "bbox": [ + 130, + 369, + 480, + 487 + ], + "type": "text", + "content": " facilitating a more controlled update of the model parameters that considers both adversarial and benign inputs. In practice, this involves updating the model's intermediate parameters, " + }, + { + "bbox": [ + 130, + 369, + 480, + 487 + ], + "type": "inline_equation", + "content": "C^k" + }, + { + "bbox": [ + 130, + 369, + 480, + 487 + ], + "type": "text", + "content": ", to reduce the adversarial loss, according to the gradient of the loss function " + }, + { + "bbox": [ + 130, + 369, + 480, + 487 + ], + "type": "inline_equation", + "content": "\\nabla_\\theta" + }, + { + "bbox": [ + 130, + 369, + 480, + 487 + ], + "type": "text", + "content": " with respect to the adversarial examples " + }, + { + "bbox": [ + 130, + 369, + 480, + 487 + ], + "type": "inline_equation", + "content": "x'^{(k)}" + }, + { + "bbox": [ + 130, + 369, + 480, + 487 + ], + "type": "text", + "content": "," + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 230, + 493, + 480, + 507 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 230, + 493, + 480, + 507 + ], + "spans": [ + { + "bbox": [ + 230, + 493, + 480, + 507 + ], + "type": "interline_equation", + "content": "C ^ {k} = \\theta^ {(k)} - \\eta \\nabla_ {\\theta} L \\left(f \\left(x ^ {\\prime (k)}, \\theta^ {(k)}\\right), y\\right), \\tag {8}", + "image_path": "5b29eb3590b851bf10d094cd342cfb1d27057406e87e87bd02e467bedd367943.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 511, + 479, + 606 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 511, + 479, + 606 + ], + "spans": [ + { + "bbox": [ + 130, + 511, + 479, + 606 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 511, + 479, + 606 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 130, + 511, + 479, + 606 + ], + "type": "text", + "content": " is a learning rate. This update, termed as the 'reflection step' (" + }, + { + "bbox": [ + 130, + 511, + 479, + 606 + ], + "type": "inline_equation", + "content": "\\theta_{\\mathrm{ref}}" + }, + { + "bbox": [ + 130, + 511, + 479, + 606 + ], + "type": "text", + "content": "), is pivotal in aligning the model's parameters towards reducing adversarial vulnerability. We then combine these updates " + }, + { + "bbox": [ + 130, + 511, + 479, + 606 + ], + "type": "inline_equation", + "content": "C^k" + }, + { + "bbox": [ + 130, + 511, + 479, + 606 + ], + "type": "text", + "content": " with regularization adjustments to stabilize the training dynamics and enhance the adversarial robustness, effectively preventing overfitting to adversarial examples while maintaining performance on benign inputs. The updated model parameters, " + }, + { + "bbox": [ + 130, + 511, + 479, + 606 + ], + "type": "inline_equation", + "content": "\\theta^{(k + 1)}" + }, + { + "bbox": [ + 130, + 511, + 479, + 606 + ], + "type": "text", + "content": ", allow to handle both adversarial and benign inputs effectively without compromising training stability or performance," + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 233, + 607, + 480, + 620 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 233, + 607, + 480, + 620 + ], + "spans": [ + { + "bbox": [ + 233, + 607, + 480, + 620 + ], + "type": "interline_equation", + "content": "\\theta^ {(k + 1)} = C ^ {k} + \\lambda (R e g (\\theta^ {(k)}) - C ^ {k}), \\tag {9}", + "image_path": "1069aec36b65df27a2f8b7c62fe8bfdb758e313ce89ead94573755ee2fbaaaa2.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 622, + 479, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 622, + 479, + 647 + ], + "spans": [ + { + "bbox": [ + 130, + 622, + 479, + 647 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 622, + 479, + 647 + ], + "type": "inline_equation", + "content": "\\lambda \\in [0,1]" + }, + { + "bbox": [ + 130, + 622, + 479, + 647 + ], + "type": "text", + "content": " controls the strength of the regularization, and Reg involves components in Eq. (7) such that" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 220, + 652, + 480, + 667 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 220, + 652, + 480, + 667 + ], + "spans": [ + { + "bbox": [ + 220, + 652, + 480, + 667 + ], + "type": "interline_equation", + "content": "R e g \\left(\\theta^ {(k)}\\right) = \\gamma_ {1} \\left(\\Delta L _ {\\theta^ {(k)}}\\right) + \\gamma_ {2} \\left(\\Delta L _ {\\theta^ {(k)}} ^ {b e n i g n}\\right). \\tag {10}", + "image_path": "c8784ba5032b75d3351cdb47dd6216c180eec9eaad591be5fabe07f114873db3.jpg" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 294, + 91, + 448, + 102 + ], + "type": "text", + "content": "Rethinking Fast Adversarial Training" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 115, + 482, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 115, + 482, + 357 + ], + "spans": [ + { + "bbox": [ + 130, + 115, + 482, + 357 + ], + "type": "text", + "content": "The important reason behind Eq. (9) is that the regularization term " + }, + { + "bbox": [ + 130, + 115, + 482, + 357 + ], + "type": "inline_equation", + "content": "Reg(\\theta^{(k)})" + }, + { + "bbox": [ + 130, + 115, + 482, + 357 + ], + "type": "text", + "content": " incorporates changes in loss for both adversarial and benign examples, weighted by " + }, + { + "bbox": [ + 130, + 115, + 482, + 357 + ], + "type": "inline_equation", + "content": "\\gamma_{1}" + }, + { + "bbox": [ + 130, + 115, + 482, + 357 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 115, + 482, + 357 + ], + "type": "inline_equation", + "content": "\\gamma_{2}" + }, + { + "bbox": [ + 130, + 115, + 482, + 357 + ], + "type": "text", + "content": ", respectively. This helps to balance the model's performance across adversarial and benign inputs, addressing CO by preventing the model from becoming overly specialized on the adversarial examples. Because CO occurs when a model overfits to specific adversarial examples during training, compromising its ability to generalize [21]. Our formulation ensures that the model's updates consider the impact on both adversarial robustness and benign accuracy. This regularization can be seen as part of the 'enforcement step' " + }, + { + "bbox": [ + 130, + 115, + 482, + 357 + ], + "type": "inline_equation", + "content": "(\\theta_{enf})" + }, + { + "bbox": [ + 130, + 115, + 482, + 357 + ], + "type": "text", + "content": ". The DR technique involves iteratively updating the model parameters by combining the adversarial loss minimization and the stability enforcement components, as " + }, + { + "bbox": [ + 130, + 115, + 482, + 357 + ], + "type": "inline_equation", + "content": "\\theta = 0.5 \\times (\\theta_{ref} + \\theta_{enf})" + }, + { + "bbox": [ + 130, + 115, + 482, + 357 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 130, + 115, + 482, + 357 + ], + "type": "inline_equation", + "content": "\\theta_{ref}" + }, + { + "bbox": [ + 130, + 115, + 482, + 357 + ], + "type": "text", + "content": " targets reducing the loss on adversariably perturbed inputs Eq. (8), conversely, " + }, + { + "bbox": [ + 130, + 115, + 482, + 357 + ], + "type": "inline_equation", + "content": "\\theta_{enf}" + }, + { + "bbox": [ + 130, + 115, + 482, + 357 + ], + "type": "text", + "content": " represents the parameters adjusted after applying the regularization term, focusing on stability and CO prevention Eq. (9). This approach under the DR-splitting technique provides a comprehensive strategy for addressing CO in the adversarial training process. Indeed, this allows the model to effectively adapt to adversarial perturbations while ensuring that such adaptations do not compromise the model's ability to accurately classify benign inputs or lead to overfitting on the adversarial examples during training. Our proposed RAT is given in Algorithm 1, discussion in the supplementary file." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 361, + 483, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 361, + 483, + 483 + ], + "spans": [ + { + "bbox": [ + 130, + 361, + 483, + 483 + ], + "type": "text", + "content": "Efficient Computation of Loss Differences In practice, directly computing the losses " + }, + { + "bbox": [ + 130, + 361, + 483, + 483 + ], + "type": "inline_equation", + "content": "L(f(x^{\\prime (k - 1)},\\theta^{(k - 1)},y))" + }, + { + "bbox": [ + 130, + 361, + 483, + 483 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 361, + 483, + 483 + ], + "type": "inline_equation", + "content": "L(f(x,\\theta^{(k - 1)}),y)" + }, + { + "bbox": [ + 130, + 361, + 483, + 483 + ], + "type": "text", + "content": " in Eq. (7) can consume a significant amount of memory. To address this, we introduce " + }, + { + "bbox": [ + 130, + 361, + 483, + 483 + ], + "type": "inline_equation", + "content": "u^{k - 1}" + }, + { + "bbox": [ + 130, + 361, + 483, + 483 + ], + "type": "text", + "content": " for benign inputs and " + }, + { + "bbox": [ + 130, + 361, + 483, + 483 + ], + "type": "inline_equation", + "content": "u^{\\prime k - 1}" + }, + { + "bbox": [ + 130, + 361, + 483, + 483 + ], + "type": "text", + "content": " for adversarial inputs, replacements for these terms, allowing for efficient computation and tracking of the average loss over benign and adversarial examples, respectively. Specially, " + }, + { + "bbox": [ + 130, + 361, + 483, + 483 + ], + "type": "inline_equation", + "content": "u^{k - 1} = \\frac{1}{N}\\sum_{i}L(f(x_{i}^{\\prime},\\theta^{(k - 1)}),y_{i})" + }, + { + "bbox": [ + 130, + 361, + 483, + 483 + ], + "type": "text", + "content": " represents the average loss for adversarial inputs, and " + }, + { + "bbox": [ + 130, + 361, + 483, + 483 + ], + "type": "inline_equation", + "content": "u^{\\prime k - 1} = \\frac{1}{N}\\sum_{i = 1}^{N}L(f(x_i,\\theta^{(k - 1)}),y_i)" + }, + { + "bbox": [ + 130, + 361, + 483, + 483 + ], + "type": "text", + "content": " denotes the average loss for benign inputs, where " + }, + { + "bbox": [ + 130, + 361, + 483, + 483 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 361, + 483, + 483 + ], + "type": "text", + "content": " is the batch size. This can reduce computational and memory overhead by focusing on the aggregate loss across batches, rather than calculating them individually for each sample." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 501, + 282, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 501, + 282, + 514 + ], + "spans": [ + { + "bbox": [ + 132, + 501, + 282, + 514 + ], + "type": "text", + "content": "4 Experimental Results" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 526, + 271, + 539 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 526, + 271, + 539 + ], + "spans": [ + { + "bbox": [ + 132, + 526, + 271, + 539 + ], + "type": "text", + "content": "4.1 Experimental settings" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": "In this section, we evaluate the proposed model across four benchmark datasets to determine its effectiveness and robustness against adversarial attacks. For both CIFAR10 [8] and CIFAR100 [8], we use the ResNet20 [12] architecture as the backbone, recognizing its strong performance in image classification tasks. For the more complex Tiny-ImageNet and ImageNet [8], we use the PreActResNet18 [13] to better handle its challenges. Additionally, the use of WideResNet34-10 [39] highlights our model's capacity to maintain effectiveness on long training schedules. We use the SGD optimizer with a batch size of 128, a weight decay of " + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "inline_equation", + "content": "5.10^{-4}" + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": ", and a momentum of 0.9. We initiate training with learning rates of 0.1 for CIFAR10 and CIFAR100. For Tiny ImageNet, due to its greater complexity," + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 233, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 233, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 233, + 102 + ], + "type": "text", + "content": "Zareapoor et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "code", + "bbox": [ + 132, + 133, + 481, + 312 + ], + "blocks": [ + { + "bbox": [ + 132, + 118, + 478, + 131 + ], + "lines": [ + { + "bbox": [ + 132, + 118, + 478, + 131 + ], + "spans": [ + { + "bbox": [ + 132, + 118, + 478, + 131 + ], + "type": "text", + "content": "Algorithm 1 Rethinking fast Adversarial Training (RAT) using DR Splitting" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "code_caption" + }, + { + "bbox": [ + 132, + 133, + 481, + 312 + ], + "lines": [ + { + "bbox": [ + 132, + 133, + 481, + 312 + ], + "spans": [ + { + "bbox": [ + 132, + 133, + 481, + 312 + ], + "type": "text", + "content": "Require: Training dataset " + }, + { + "bbox": [ + 132, + 133, + 481, + 312 + ], + "type": "inline_equation", + "content": "(x_{i},y_{i})_{i = 1}^{N}" + }, + { + "bbox": [ + 132, + 133, + 481, + 312 + ], + "type": "text", + "content": " # epochs K, # batch B, perturbation budget " + }, + { + "bbox": [ + 132, + 133, + 481, + 312 + ], + "type": "inline_equation", + "content": "\\mathcal{E},\\lambda ,\\gamma_1,\\gamma_2" + }, + { + "bbox": [ + 132, + 133, + 481, + 312 + ], + "type": "text", + "content": " \n1: Initialize model parameters " + }, + { + "bbox": [ + 132, + 133, + 481, + 312 + ], + "type": "inline_equation", + "content": "\\theta_0" + }, + { + "bbox": [ + 132, + 133, + 481, + 312 + ], + "type": "text", + "content": " \n2: for epoch " + }, + { + "bbox": [ + 132, + 133, + 481, + 312 + ], + "type": "inline_equation", + "content": "= 1" + }, + { + "bbox": [ + 132, + 133, + 481, + 312 + ], + "type": "text", + "content": " to K do \n3: for " + }, + { + "bbox": [ + 132, + 133, + 481, + 312 + ], + "type": "inline_equation", + "content": "i = 1" + }, + { + "bbox": [ + 132, + 133, + 481, + 312 + ], + "type": "text", + "content": " to B do \n4: " + }, + { + "bbox": [ + 132, + 133, + 481, + 312 + ], + "type": "inline_equation", + "content": "x_{i}^{\\prime} = x_{i} + \\delta \\mathrm{sign}(\\nabla_{x_{i}}L(f(x_{i};\\theta_{\\mathrm{epoch - 1}}),y_{i}))\\triangleright" + }, + { + "bbox": [ + 132, + 133, + 481, + 312 + ], + "type": "text", + "content": " Adversarial example generation \n5: " + }, + { + "bbox": [ + 132, + 133, + 481, + 312 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{adv}} = L(f(x_i';\\theta_{\\mathrm{epoch - 1}}),y_i)" + }, + { + "bbox": [ + 132, + 133, + 481, + 312 + ], + "type": "text", + "content": " \n6: " + }, + { + "bbox": [ + 132, + 133, + 481, + 312 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{benign}} = L(f(x_i;\\theta_{\\mathrm{epoch - 1}}),y_i)" + }, + { + "bbox": [ + 132, + 133, + 481, + 312 + ], + "type": "text", + "content": " \n7: " + }, + { + "bbox": [ + 132, + 133, + 481, + 312 + ], + "type": "inline_equation", + "content": "\\theta_{\\mathrm{reflect}} = \\theta_{\\mathrm{epoch - 1}} - \\eta \\cdot \\nabla_{\\theta}L_{\\mathrm{adv}}" + }, + { + "bbox": [ + 132, + 133, + 481, + 312 + ], + "type": "inline_equation", + "content": "\\triangleright" + }, + { + "bbox": [ + 132, + 133, + 481, + 312 + ], + "type": "text", + "content": " Reflection step \n8: " + }, + { + "bbox": [ + 132, + 133, + 481, + 312 + ], + "type": "inline_equation", + "content": "\\Delta L_{\\mathrm{adv}} = |L(f(x_i';\\theta_{\\mathrm{reflect}}),y_i) - L_{\\mathrm{adv}}|" + }, + { + "bbox": [ + 132, + 133, + 481, + 312 + ], + "type": "text", + "content": " \n9: " + }, + { + "bbox": [ + 132, + 133, + 481, + 312 + ], + "type": "inline_equation", + "content": "\\Delta L_{\\mathrm{benign}} = |L(f(x_i;\\theta_{\\mathrm{reflect}}),y_i) - L_{\\mathrm{benign}}|" + }, + { + "bbox": [ + 132, + 133, + 481, + 312 + ], + "type": "text", + "content": " \n10: Reg " + }, + { + "bbox": [ + 132, + 133, + 481, + 312 + ], + "type": "inline_equation", + "content": "= \\gamma_{1}\\cdot \\Delta L_{\\mathrm{adv}} + \\gamma_{2}\\cdot \\Delta L_{\\mathrm{benign}}" + }, + { + "bbox": [ + 132, + 133, + 481, + 312 + ], + "type": "inline_equation", + "content": "\\triangleright" + }, + { + "bbox": [ + 132, + 133, + 481, + 312 + ], + "type": "text", + "content": " Enforcement step \n11: " + }, + { + "bbox": [ + 132, + 133, + 481, + 312 + ], + "type": "inline_equation", + "content": "\\theta_{\\mathrm{enforce}} = \\theta_{\\mathrm{reflect}} + \\lambda \\cdot (\\mathrm{Reg} - \\theta_{\\mathrm{reflect}})" + }, + { + "bbox": [ + 132, + 133, + 481, + 312 + ], + "type": "text", + "content": " \n12: " + }, + { + "bbox": [ + 132, + 133, + 481, + 312 + ], + "type": "inline_equation", + "content": "\\theta_{\\mathrm{epoch}} = 0.5\\cdot (\\theta_{\\mathrm{reflect}} + \\theta_{\\mathrm{enforce}})" + }, + { + "bbox": [ + 132, + 133, + 481, + 312 + ], + "type": "inline_equation", + "content": "\\triangleright" + }, + { + "bbox": [ + 132, + 133, + 481, + 312 + ], + "type": "text", + "content": " DR splitting \n13: end for \n14: end for" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "code_body" + } + ], + "index": 3, + "sub_type": "algorithm" + }, + { + "bbox": [ + 130, + 337, + 482, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 337, + 482, + 456 + ], + "spans": [ + { + "bbox": [ + 130, + 337, + 482, + 456 + ], + "type": "text", + "content": "we start at a lower rate of 0.01. The training spans 110 epochs, with a strategy to decrease the learning rate by a factor of 0.1 at both the 100th and 105th epochs for optimization. To prepare our model against adversarial attacks, we incorporate FAST-AT [36] and MEP-PGI [15] as our initial attack strategies for CIFAR datasets. Given the significant memory demands of MEP-PGI, particularly its need to store previous adversarial perturbations, we use its variation PGI-BP on Tiny ImageNet. We set the step-size parameter " + }, + { + "bbox": [ + 130, + 337, + 482, + 456 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 130, + 337, + 482, + 456 + ], + "type": "text", + "content": " to 0.01 for CIFAR datasets, and 0.001 for Tiny ImageNet. The regularization term " + }, + { + "bbox": [ + 130, + 337, + 482, + 456 + ], + "type": "inline_equation", + "content": "\\lambda = 0.3" + }, + { + "bbox": [ + 130, + 337, + 482, + 456 + ], + "type": "text", + "content": " for all datasets, and " + }, + { + "bbox": [ + 130, + 337, + 482, + 456 + ], + "type": "inline_equation", + "content": "\\mathcal{E} = 16 / 255" + }, + { + "bbox": [ + 130, + 337, + 482, + 456 + ], + "type": "text", + "content": ". Experiments are run using a single GeForce RTX 3090 GPU. Find the training code and additional results in the supplementary file." + } + ] + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 151, + 478, + 470, + 544 + ], + "blocks": [ + { + "bbox": [ + 151, + 478, + 470, + 544 + ], + "lines": [ + { + "bbox": [ + 151, + 478, + 470, + 544 + ], + "spans": [ + { + "bbox": [ + 151, + 478, + 470, + 544 + ], + "type": "table", + "html": "
MethodSA (%)RA-PGD-10\nE=8/255RA-PGD-10\nE=10/255RA-PGD-10\nE=12/255RA-PGD-10\nE=16/255SQ [2]Time (h)
AWP [37]77.9436.2633.1830.2527.4319.762.4
OAAT [1]70.5145.3239.4734.1028.9121.433.2
ExAT [29]79.3349.5643.2836.4932.5823.151.2
RAT-MEP69.8652.6446.8344.1841.2627.131.7
", + "image_path": "cfc1264e483fe9047e91342be2d3b66c9300a96063b051bbda64c7ffb3d32283.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 552, + 482, + 586 + ], + "lines": [ + { + "bbox": [ + 131, + 552, + 482, + 586 + ], + "spans": [ + { + "bbox": [ + 131, + 552, + 482, + 586 + ], + "type": "text", + "content": "Table 1: Comparative performance across different perturbation levels on CIFAR10 using ResNet20 backbone. Our evaluation focuses on two key metrics: robust accuracy (RA) against PGD-10 attacks, and the computational time (measured in hours)." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": "Baselines and Evaluation Metrics: To thoroughly evaluate the adversarial robustness of our model, we use a comprehensive set of attack methods. Following [36], we use FGSM [11], and a series of PGD attacks [22] with varying iterations: PGD-10, PGD-20 and PGD-50. Additionally, we use the APGD-CE [7], C&W attack [5], and Autoattack (AA) [7]. AutoAttack is an ensemble method" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 294, + 91, + 447, + 102 + ], + "type": "text", + "content": "Rethinking Fast Adversarial Training" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 133, + 114, + 482, + 206 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 482, + 206 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 482, + 206 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 482, + 206 + ], + "type": "table", + "html": "
MethodsSARA-FGSMRA-PGD10RA-PGD20RA-PGD50RA-C&WRA-APGD-CERA-AATime (h)
AT PGD65.7246.5140.8735.4234.2130.9733.6826.856.3
Fast-GA [3]58.4939.8833.3126.8925.0722.8124.3317.562.4
ZeroGrad [10]72.8144.2632.7822.1519.2320.8517.2012.751.3
NuAT [33]73.5544.9035.5226.3424.3124.9523.7218.931.8
Fast-AT [36]50.6438.6226.4721.6420.5519.2619.8114.971.2
RAT-AT70.6246.2838.1230.8329.3526.5427.7919.641.3
MEP-PGI [15]54.2336.8432.6527.9826.7022.6526.7319.311.5
RAT-MEP69.8647.6941.2634.8133.6428.7532.8723.911.7
", + "image_path": "0e8382f5c83ac1146237687dab5e1502285a30be6f1c489c3a75dee50366ef57.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 214, + 482, + 259 + ], + "lines": [ + { + "bbox": [ + 130, + 214, + 482, + 259 + ], + "spans": [ + { + "bbox": [ + 130, + 214, + 482, + 259 + ], + "type": "text", + "content": "Table 2: CIFAR10 results: We report the RA (\\%) against various attacks at " + }, + { + "bbox": [ + 130, + 214, + 482, + 259 + ], + "type": "inline_equation", + "content": "\\mathcal{E} = 16 / 255" + }, + { + "bbox": [ + 130, + 214, + 482, + 259 + ], + "type": "text", + "content": " with ResNet20 backbone. Each model undergoes 3 training sessions. The results reflect the average performance of the top-performing model across these iterations. The best and the runner-up models are bolded and underlined, respectively." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 130, + 286, + 482, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 286, + 482, + 373 + ], + "spans": [ + { + "bbox": [ + 130, + 286, + 482, + 373 + ], + "type": "text", + "content": "that integrates APGD-CE, APGD-T (targeted APGD), and Square Attack [2], to thoroughly evaluate a model's robustness to adversarial attacks. Due to the inherent variability in the training process, which often leads to fluctuations in performance, relying on a single run may not accurately capture the effectiveness of different methods. Thus, for each method, we conduct the training process three times and report the average of the best results in these runs. The evaluation metrics include robust accuracy (RA) and standard accuracy (SA)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 395, + 263, + 407 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 395, + 263, + 407 + ], + "spans": [ + { + "bbox": [ + 132, + 395, + 263, + 407 + ], + "type": "text", + "content": "4.2 Results and Insights" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 426, + 482, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 426, + 482, + 668 + ], + "spans": [ + { + "bbox": [ + 130, + 426, + 482, + 668 + ], + "type": "text", + "content": "In our preliminary analysis, we highlight the importance of our results on the adversarial robustness of trained models. Our findings indicate that adversarial attacks with " + }, + { + "bbox": [ + 130, + 426, + 482, + 668 + ], + "type": "inline_equation", + "content": "\\mathcal{E} = 16 / 255" + }, + { + "bbox": [ + 130, + 426, + 482, + 668 + ], + "type": "text", + "content": ", rarely result in the misclassification of the input. This observation is achieved through the generation of adversarial examples across CIFAR10 and CIFAR100 using both targeted and non-targeted attack methods, including FGSM, APGD-CE, APGD-T, PGD, and C&W. Remarkably, most of these adversarial examples still were classified correctly according to their original labels. Our experiments in Tab. 1 involve generating adversarial training with perturbation size " + }, + { + "bbox": [ + 130, + 426, + 482, + 668 + ], + "type": "inline_equation", + "content": "\\mathcal{E} = 16 / 255" + }, + { + "bbox": [ + 130, + 426, + 482, + 668 + ], + "type": "text", + "content": " and testing the models' robustness across different levels of perturbation. We also compared our model (RAT) with other techniques, such as Fast-Bat, AWP, OAAT, and ExAT. We have reimplemented these AT models within the FastAT framework to ensure a fair comparison. Our model achieves superior performance even as the perturbation magnitude increases to " + }, + { + "bbox": [ + 130, + 426, + 482, + 668 + ], + "type": "inline_equation", + "content": "\\mathcal{E} = 16 / 255" + }, + { + "bbox": [ + 130, + 426, + 482, + 668 + ], + "type": "text", + "content": ". Furthermore, our model integrates into FAST-AT and MEP-PGI, which are specifically designed to address distributions of adversarial samples. Conversely, methods like OAAT are optimized for the distribution of benign examples and yield higher SA accuracy. Adopting the settings from OAAT, we apply both OAAT and our variation " + }, + { + "bbox": [ + 130, + 426, + 482, + 668 + ], + "type": "inline_equation", + "content": "\\mathrm{OAAT}_{\\mathrm{RAT}}" + }, + { + "bbox": [ + 130, + 426, + 482, + 668 + ], + "type": "text", + "content": " to FastAT tasks. The " + }, + { + "bbox": [ + 130, + 426, + 482, + 668 + ], + "type": "inline_equation", + "content": "\\mathrm{OAAT}_{\\mathrm{RAT}}" + }, + { + "bbox": [ + 130, + 426, + 482, + 668 + ], + "type": "text", + "content": " achieves " + }, + { + "bbox": [ + 130, + 426, + 482, + 668 + ], + "type": "inline_equation", + "content": "+2.78\\%" + }, + { + "bbox": [ + 130, + 426, + 482, + 668 + ], + "type": "text", + "content": " improvement on SA accuracy over OAAT (" + }, + { + "bbox": [ + 130, + 426, + 482, + 668 + ], + "type": "inline_equation", + "content": "70.51\\% \\rightarrow 73.29" + }, + { + "bbox": [ + 130, + 426, + 482, + 668 + ], + "type": "text", + "content": "), and " + }, + { + "bbox": [ + 130, + 426, + 482, + 668 + ], + "type": "inline_equation", + "content": "24.66\\%" + }, + { + "bbox": [ + 130, + 426, + 482, + 668 + ], + "type": "text", + "content": " on the SQ, as (" + }, + { + "bbox": [ + 130, + 426, + 482, + 668 + ], + "type": "inline_equation", + "content": "21.43\\% \\rightarrow 24.66\\%" + }, + { + "bbox": [ + 130, + 426, + 482, + 668 + ], + "type": "text", + "content": ")." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 233, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 233, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 233, + 102 + ], + "type": "text", + "content": "Zareapoor et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 134, + 114, + 482, + 200 + ], + "blocks": [ + { + "bbox": [ + 134, + 114, + 482, + 200 + ], + "lines": [ + { + "bbox": [ + 134, + 114, + 482, + 200 + ], + "spans": [ + { + "bbox": [ + 134, + 114, + 482, + 200 + ], + "type": "table", + "html": "
MethodsESARA-FGSMRA-PGD10RA-PGD20RA-PGD50RA-C&WRA-APGD-CERA-AATime (h)
Fast-GA [36]1267.3943.5732.1928.0226.2427.6526.2121.752.3
NuAT [33]72.8452.3341.9538.8437.6936.0836.7532.411.8
MEP-PGI [15]73.7852.3938.6736.5235.7833.4133.6827.931.5
RAT-MEP72.8154.6245.2942.8642.2336.8841.6033.161.7
Fast-GA [3]1081.7055.7336.8030.7128.6531.2926.9124.122.3
NuAT [33]76.3956.0946.2343.9543.5041.5243.3638.991.8
MEP-PGI [15]81.9459.4843.1839.7637.8337.4636.2132.721.5
RAT-MEP76.5258.3047.5145.7445.5641.2745.3639.701.7
", + "image_path": "e82f0995fdc4924af54c6ba506c1c0abf2d1f10ad1b4b031e4b54bba4d2614f5.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 134, + 252, + 482, + 341 + ], + "blocks": [ + { + "bbox": [ + 132, + 208, + 482, + 242 + ], + "lines": [ + { + "bbox": [ + 132, + 208, + 482, + 242 + ], + "spans": [ + { + "bbox": [ + 132, + 208, + 482, + 242 + ], + "type": "text", + "content": "Table 3: Quantitative comparison of FastAT methods on CIFAR10 using ResNet20 for different " + }, + { + "bbox": [ + 132, + 208, + 482, + 242 + ], + "type": "inline_equation", + "content": "(\\mathcal{E})" + }, + { + "bbox": [ + 132, + 208, + 482, + 242 + ], + "type": "text", + "content": ", where both training and testing used the same " + }, + { + "bbox": [ + 132, + 208, + 482, + 242 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 132, + 208, + 482, + 242 + ], + "type": "text", + "content": ". The best and the runner-up methods are bolded and underlined, respectively." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 134, + 252, + 482, + 341 + ], + "lines": [ + { + "bbox": [ + 134, + 252, + 482, + 341 + ], + "spans": [ + { + "bbox": [ + 134, + 252, + 482, + 341 + ], + "type": "table", + "html": "
MethodsSARA-FGSMRA-PGD10RA-PGD20RA-PGD50RA-C&WRA-APGD-CERA-AATime (h)
AT PGD40.8325.1921.3817.5917.2315.0617.1012.546.3
Fast-GA [3]32.4015.8412.9810.7910.088.549.726.842.4
ZeroGrad [10]47.8124.1318.2112.9312.2411.7011.217.901.2
NuAT [33]32.4921.2716.7514.3813.7111.5212.628.751.8
Fast-AT [36]31.2215.6913.4110.7810.258.739.887.211.2
RAT-AT (ours)42.2625.8119.4315.7714.5612.8113.959.861.3
MEP-PGI [15]22.4114.0912.259.869.618.159.536.861.5
RAT-MEP (ours)43.7725.2621.3517.4217.0314.2216.7611.491.7
", + "image_path": "aac752da24306eea893f051ec5db18f0346fd87cfe585ec344383de9e9a3befd.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 350, + 482, + 395 + ], + "lines": [ + { + "bbox": [ + 132, + 350, + 482, + 395 + ], + "spans": [ + { + "bbox": [ + 132, + 350, + 482, + 395 + ], + "type": "text", + "content": "Table 4: CIFAR100 results: We report the RA (\\%) against various attacks at " + }, + { + "bbox": [ + 132, + 350, + 482, + 395 + ], + "type": "inline_equation", + "content": "\\mathcal{E} = 16 / 255" + }, + { + "bbox": [ + 132, + 350, + 482, + 395 + ], + "type": "text", + "content": " with ResNet20 backbone. Each method undergoes three training sessions. The results reflect the average performance of the top-performing model across these iterations. The best and the runner-up methods are bolded and underlined, respectively." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 132, + 418, + 315, + 430 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 418, + 315, + 430 + ], + "spans": [ + { + "bbox": [ + 132, + 418, + 315, + 430 + ], + "type": "text", + "content": "4.3 Faster DR Splitting Technique" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 437, + 482, + 606 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 437, + 482, + 606 + ], + "spans": [ + { + "bbox": [ + 130, + 437, + 482, + 606 + ], + "type": "text", + "content": "Fig. 1-right provides a detailed comparison of various adversarial training on the CIFAR10 dataset, at " + }, + { + "bbox": [ + 130, + 437, + 482, + 606 + ], + "type": "inline_equation", + "content": "\\mathcal{E} = (8,16) / 255" + }, + { + "bbox": [ + 130, + 437, + 482, + 606 + ], + "type": "text", + "content": ". At the lower perturbation level (" + }, + { + "bbox": [ + 130, + 437, + 482, + 606 + ], + "type": "inline_equation", + "content": "\\mathcal{E} = 8 / 255" + }, + { + "bbox": [ + 130, + 437, + 482, + 606 + ], + "type": "text", + "content": "), all methods demonstrate similar AA and SA, with FAST-GA achieving the highest AA among them. However, increasing the perturbation to " + }, + { + "bbox": [ + 130, + 437, + 482, + 606 + ], + "type": "inline_equation", + "content": "\\mathcal{E} = 16 / 255" + }, + { + "bbox": [ + 130, + 437, + 482, + 606 + ], + "type": "text", + "content": " results in a significant reduction in AA across all methods, highlighting the challenges of maintaining robustness at higher perturbation levels. Interestingly, our model outperforms Fast-Bat at " + }, + { + "bbox": [ + 130, + 437, + 482, + 606 + ], + "type": "inline_equation", + "content": "\\mathcal{E} = 16 / 255" + }, + { + "bbox": [ + 130, + 437, + 482, + 606 + ], + "type": "text", + "content": ", achieving " + }, + { + "bbox": [ + 130, + 437, + 482, + 606 + ], + "type": "inline_equation", + "content": "60.37\\%" + }, + { + "bbox": [ + 130, + 437, + 482, + 606 + ], + "type": "text", + "content": " vs. " + }, + { + "bbox": [ + 130, + 437, + 482, + 606 + ], + "type": "inline_equation", + "content": "59.34\\%" + }, + { + "bbox": [ + 130, + 437, + 482, + 606 + ], + "type": "text", + "content": " in SA accuracy. This improvement is also evident at " + }, + { + "bbox": [ + 130, + 437, + 482, + 606 + ], + "type": "inline_equation", + "content": "\\mathcal{E} = 8 / 255" + }, + { + "bbox": [ + 130, + 437, + 482, + 606 + ], + "type": "text", + "content": ", where RAT aligns closely with the performance of other leading methods, e.g., " + }, + { + "bbox": [ + 130, + 437, + 482, + 606 + ], + "type": "inline_equation", + "content": "46.21\\%" + }, + { + "bbox": [ + 130, + 437, + 482, + 606 + ], + "type": "text", + "content": " AA accuracy, comparable to " + }, + { + "bbox": [ + 130, + 437, + 482, + 606 + ], + "type": "inline_equation", + "content": "46.37\\%" + }, + { + "bbox": [ + 130, + 437, + 482, + 606 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 437, + 482, + 606 + ], + "type": "inline_equation", + "content": "45.83\\%" + }, + { + "bbox": [ + 130, + 437, + 482, + 606 + ], + "type": "text", + "content": " for FAST-GA and Fast-Bat respectively. While FGSM is faster (as shown in Fig. 1-left), results in " + }, + { + "bbox": [ + 130, + 437, + 482, + 606 + ], + "type": "inline_equation", + "content": "0\\%" + }, + { + "bbox": [ + 130, + 437, + 482, + 606 + ], + "type": "text", + "content": " AA accuracy. We also observe that our model adds very little overhead to FGSM training, i.e., 15.02 min with FGSM, but is considerably faster than other methods. The increased computation time in LLR and CURE is due to Double Backpropagation [28]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 622, + 250, + 632 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 622, + 250, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 622, + 250, + 632 + ], + "type": "text", + "content": "4.4 CIFAR10 Results" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 641, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 641, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 641, + 481, + 665 + ], + "type": "text", + "content": "In this experiment, we use backbone ResNet20, to evaluate our model's performance, setting different perturbation levels " + }, + { + "bbox": [ + 132, + 641, + 481, + 665 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 132, + 641, + 481, + 665 + ], + "type": "text", + "content": ". The results, in Tab.2, show that" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 294, + 91, + 447, + 102 + ], + "type": "text", + "content": "Rethinking Fast Adversarial Training" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 133, + 114, + 486, + 194 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 486, + 194 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 486, + 194 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 486, + 194 + ], + "type": "table", + "html": "
MethodsSARA-FGSMRA-PGD10RA-PGD20RA-PGD50RA-C&WRA-APGD-CERA-AATime (h)
AT PGD33.7917.6613.8210.5610.288.3610.456.4768.1
Fast-AT [36]28.1213.0610.217.366.955.486.724.5110.8
Fast-GA [3]29.2413.9811.328.458.166.237.404.9621.4
ZeroGrad [10]34.5812.859.416.735.714.225.262.8110.9
NuAT [33]35.2016.7112.489.529.377.628.215.3925.2
PGI-BP [15]21.7611.329.487.567.205.497.214.1813.8
RAT-BP (ours)34.6117.6213.239.919.458.209.876.2514.2
", + "image_path": "a211521eb3d0fd8c46fe3267b95ec8155991ba4cc783e67e149f88ebc2606f05.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 201, + 482, + 236 + ], + "lines": [ + { + "bbox": [ + 130, + 201, + 482, + 236 + ], + "spans": [ + { + "bbox": [ + 130, + 201, + 482, + 236 + ], + "type": "text", + "content": "Table 5: Tiny-ImageNet results: We report the RA (\\%) against various attack methods accuracy at " + }, + { + "bbox": [ + 130, + 201, + 482, + 236 + ], + "type": "inline_equation", + "content": "\\mathcal{E} = 16 / 255" + }, + { + "bbox": [ + 130, + 201, + 482, + 236 + ], + "type": "text", + "content": " using PreActResNet18 as our backbone. The results represent the average performance of the best models across three training iterations." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 130, + 259, + 482, + 428 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 259, + 482, + 428 + ], + "spans": [ + { + "bbox": [ + 130, + 259, + 482, + 428 + ], + "type": "text", + "content": "our model outperforms existing Fast AT techniques in adversarial robustness across various attacks. Specifically, our RAT-AT surpasses all FastAT methods, and our RAT-MEP is more effective than other methods while maintaining comparable SA accuracy. In comparison to PGD-AT, RAT-MEP achieves nearly equivalent adversarial robustness, with a marginal of 0.57 difference in performance against PGD-50 (33.64% vs. 34.21%). Regarding computational efficiency, RAT-AT is faster, completing in 75 minutes, compared to 144 minutes for Fast-GA and 108 minutes for NuAT. RAT-MEP also presents a good balance between computational time (102 minutes) and performance, taking only slightly longer than MEP-PGI (96 minutes) but significantly less than the 380 minutes required by PGD-AT. In Tab. 3, we conduct additional experiments on different perturbation scales " + }, + { + "bbox": [ + 130, + 259, + 482, + 428 + ], + "type": "inline_equation", + "content": "\\mathcal{E} = (10,12)/55" + }, + { + "bbox": [ + 130, + 259, + 482, + 428 + ], + "type": "text", + "content": ", further validating the model's effectiveness. The results clearly demonstrate that our method consistently achieves superior adversarial robustness across different perturbation budgets." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 441, + 256, + 451 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 441, + 256, + 451 + ], + "spans": [ + { + "bbox": [ + 132, + 441, + 256, + 451 + ], + "type": "text", + "content": "4.5 CIFAR100 Results" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 461, + 482, + 558 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 461, + 482, + 558 + ], + "spans": [ + { + "bbox": [ + 130, + 461, + 482, + 558 + ], + "type": "text", + "content": "Tab. 4 details the performance on CIFAR100 with a ResNet20 backbone, highlighting that the training times for FastAT methods across CIFAR10 and CIFAR100 are similar. As can be seen, our model demonstrates the ability to maintain stable adversarial training, and among the FastAT methods, RAT-AT is the most robust against different attacks. RAT-PGI surpasses the performance of all FastAT methods in terms of adversarial robustness. While RAT-MEP is slightly lower than AT-PGD, it boasts a training speed that is approximately 3.5 times faster, marking a significant improvement in efficiency." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 572, + 253, + 585 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 572, + 253, + 585 + ], + "spans": [ + { + "bbox": [ + 132, + 572, + 253, + 585 + ], + "type": "text", + "content": "4.6 ImageNet Results" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": "Our experiments with the PreActResNet18 on Tiny-ImageNet, demonstrate the effectiveness of our method in handling larger and more complex dataset, as presented in Tab. 5. The RAT-BP shows comparable robustness to PGD-AT and surpasses other FastAT methods. In terms of training efficiency, RAT-BP is marginally higher, requiring 14.2 hours, compared to PGI-BP (13.8 hours), yet it is significantly faster than PGD-AT, which takes 68 hours. Further extending" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 233, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 233, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 233, + 102 + ], + "type": "text", + "content": "Zareapoor et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 168, + 114, + 449, + 166 + ], + "blocks": [ + { + "bbox": [ + 168, + 114, + 449, + 166 + ], + "lines": [ + { + "bbox": [ + 168, + 114, + 449, + 166 + ], + "spans": [ + { + "bbox": [ + 168, + 114, + 449, + 166 + ], + "type": "table", + "html": "
MethodsSARA-AARA-FGSMRA-PGD10RA-PGD50Time (h)
Fast-AT [36]7.411.273.423.762.2561.5
PGI-BP [15]16.211.059.482.811.8577.6
RAT-BP (ours)25.601.8814.618.703.2979.2
", + "image_path": "5bf2995e52d431ad3521d35a32570494fb2dff2a9b5dd90ed3f22acce404fe52.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 174, + 481, + 207 + ], + "lines": [ + { + "bbox": [ + 132, + 174, + 481, + 207 + ], + "spans": [ + { + "bbox": [ + 132, + 174, + 481, + 207 + ], + "type": "text", + "content": "Table 6: Quantitative results of the adversarial training methods on ImageNet at " + }, + { + "bbox": [ + 132, + 174, + 481, + 207 + ], + "type": "inline_equation", + "content": "(\\mathcal{E} = 16 / 255)" + }, + { + "bbox": [ + 132, + 174, + 481, + 207 + ], + "type": "text", + "content": " with ResNet50 as the backbone. Adding our model (RAT) helps PGI-BP overcome CO and improve its performance." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 138, + 217, + 477, + 321 + ], + "blocks": [ + { + "bbox": [ + 138, + 217, + 477, + 321 + ], + "lines": [ + { + "bbox": [ + 138, + 217, + 477, + 321 + ], + "spans": [ + { + "bbox": [ + 138, + 217, + 477, + 321 + ], + "type": "image", + "image_path": "2b42ffc45215e7bf1b028fbe4164460bfb306abefd475b5a109e21af7f648737.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 324, + 482, + 390 + ], + "lines": [ + { + "bbox": [ + 130, + 324, + 482, + 390 + ], + "spans": [ + { + "bbox": [ + 130, + 324, + 482, + 390 + ], + "type": "text", + "content": "Fig. 3: Catastrophic Overfitting: Illustrates the robustness across perturbation strengths. The x-axis " + }, + { + "bbox": [ + 130, + 324, + 482, + 390 + ], + "type": "inline_equation", + "content": "(\\mathcal{E})" + }, + { + "bbox": [ + 130, + 324, + 482, + 390 + ], + "type": "text", + "content": " represents the strength of the perturbation applied to test images to create adversarial examples. The y-axis is the accuracy of the model under attack (PGD-20). As " + }, + { + "bbox": [ + 130, + 324, + 482, + 390 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 130, + 324, + 482, + 390 + ], + "type": "text", + "content": " increases, we expect accuracy to decrease, since stronger perturbations are harder to withstand. The AT PGD-10 is referenced for comparison, showing that our model achieves performance closely aligned with this benchmark." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 413, + 482, + 473 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 413, + 482, + 473 + ], + "spans": [ + { + "bbox": [ + 130, + 413, + 482, + 473 + ], + "type": "text", + "content": "our evaluation to the ImageNet [8], we use a ResNet50 following [36] training configuration with the only difference by the adoption of a maximum image resolution of " + }, + { + "bbox": [ + 130, + 413, + 482, + 473 + ], + "type": "inline_equation", + "content": "288 \\times 288" + }, + { + "bbox": [ + 130, + 413, + 482, + 473 + ], + "type": "text", + "content": ", similar to [16]. The initial learning rate is 0.1, with a reduction factor of 0.1 applied after 30 and 60 epochs, across a total of 90 epochs. The results in Tab. 6 show the stability of our training across this dataset." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 489, + 205, + 499 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 489, + 205, + 499 + ], + "spans": [ + { + "bbox": [ + 132, + 489, + 205, + 499 + ], + "type": "text", + "content": "4.7 Ablation" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 507, + 491, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 507, + 491, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 507, + 491, + 665 + ], + "type": "text", + "content": "Our ablation study focuses on evaluating the effectiveness of our model against leading adversarial training methods, such as Fast-GA, Fast-AT, SLAT, and Fast-Bat. We also reference the robust but computationally intensive AT PGD10. PreActResNet18 is a backbone model for these experiments, with each method's recommended hyperparameters applied. We calculate the PGD-20 test accuracy, averaging results over three random seeds to ensure reliability. Fig. 3, illustrates that SLAT encounters CO at perturbation levels of " + }, + { + "bbox": [ + 130, + 507, + 491, + 665 + ], + "type": "inline_equation", + "content": "\\mathcal{E} \\geq 4/255" + }, + { + "bbox": [ + 130, + 507, + 491, + 665 + ], + "type": "text", + "content": " for SVHN and " + }, + { + "bbox": [ + 130, + 507, + 491, + 665 + ], + "type": "inline_equation", + "content": "\\mathcal{E}e.g.10/255" + }, + { + "bbox": [ + 130, + 507, + 491, + 665 + ], + "type": "text", + "content": " for CIFAR10. Fast-GA shows less than optimal performance at " + }, + { + "bbox": [ + 130, + 507, + 491, + 665 + ], + "type": "inline_equation", + "content": "\\mathcal{E}e.g.16/255" + }, + { + "bbox": [ + 130, + 507, + 491, + 665 + ], + "type": "text", + "content": " for CIFAR10 and inconsistent results on CIFAR100. Notably, Fast-Bat also experiences CO for " + }, + { + "bbox": [ + 130, + 507, + 491, + 665 + ], + "type": "inline_equation", + "content": "\\mathcal{E} > 10/255" + }, + { + "bbox": [ + 130, + 507, + 491, + 665 + ], + "type": "text", + "content": " in CIFAR100. Overall, our RAT model significantly narrows the gap between single-step and multi-step AT methods, showing enhanced robustness and efficiency. Sensitivity to " + }, + { + "bbox": [ + 130, + 507, + 491, + 665 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 130, + 507, + 491, + 665 + ], + "type": "text", + "content": ": To optimize our model, we focus on determining the optimal value for the parameter " + }, + { + "bbox": [ + 130, + 507, + 491, + 665 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 130, + 507, + 491, + 665 + ], + "type": "text", + "content": " in Eq.(9)" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 294, + 91, + 447, + 102 + ], + "type": "text", + "content": "Rethinking Fast Adversarial Training" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 134, + 120, + 306, + 225 + ], + "blocks": [ + { + "bbox": [ + 134, + 120, + 306, + 225 + ], + "lines": [ + { + "bbox": [ + 134, + 120, + 306, + 225 + ], + "spans": [ + { + "bbox": [ + 134, + 120, + 306, + 225 + ], + "type": "image", + "image_path": "7bb2199ce960e8a180d60b87c83905ffe1ddded7c9f3c4e0605def585f3475dd.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 236, + 309, + 304 + ], + "lines": [ + { + "bbox": [ + 131, + 236, + 309, + 304 + ], + "spans": [ + { + "bbox": [ + 131, + 236, + 309, + 304 + ], + "type": "text", + "content": "Fig. 4: Exploration of " + }, + { + "bbox": [ + 131, + 236, + 309, + 304 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 131, + 236, + 309, + 304 + ], + "type": "text", + "content": " values for RAT on CIFAR10, using PRN backbone. This illustrates the relationship between the choice of " + }, + { + "bbox": [ + 131, + 236, + 309, + 304 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 131, + 236, + 309, + 304 + ], + "type": "text", + "content": " and the occurrence of CO. Higher values successfully prevent CO but lead to a noticeable reduction in model performance." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 313, + 116, + 482, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 116, + 482, + 331 + ], + "spans": [ + { + "bbox": [ + 313, + 116, + 482, + 331 + ], + "type": "text", + "content": "across varying levels of adversarial perturbation. We evaluate the PGD20 adversarial accuracy on a validation set of 1024 images from the CIFAR10 dataset. The experiments are conducted using the PRN architecture, exploring a range of " + }, + { + "bbox": [ + 313, + 116, + 482, + 331 + ], + "type": "inline_equation", + "content": "\\lambda \\in \\{0.1,0.3,0.5,0.6,0.9,1\\}" + }, + { + "bbox": [ + 313, + 116, + 482, + 331 + ], + "type": "text", + "content": ". Except for " + }, + { + "bbox": [ + 313, + 116, + 482, + 331 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 313, + 116, + 482, + 331 + ], + "type": "text", + "content": ", all other parameters are maintained as per the default configuration. The objective is to identify the " + }, + { + "bbox": [ + 313, + 116, + 482, + 331 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 313, + 116, + 482, + 331 + ], + "type": "text", + "content": " that yields the highest adversarial accuracy for each " + }, + { + "bbox": [ + 313, + 116, + 482, + 331 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 313, + 116, + 482, + 331 + ], + "type": "text", + "content": ", as shown in Fig.4. Our results indicate a clear trend: lower " + }, + { + "bbox": [ + 313, + 116, + 482, + 331 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 313, + 116, + 482, + 331 + ], + "type": "text", + "content": " and higher " + }, + { + "bbox": [ + 313, + 116, + 482, + 331 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 313, + 116, + 482, + 331 + ], + "type": "text", + "content": " values, fail to prevent CO. Increasing " + }, + { + "bbox": [ + 313, + 116, + 482, + 331 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 313, + 116, + 482, + 331 + ], + "type": "text", + "content": " to 0.9 eliminates CO but degrades model performance. This effect is particularly ev" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 331, + 482, + 452 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 331, + 482, + 452 + ], + "spans": [ + { + "bbox": [ + 130, + 331, + 482, + 452 + ], + "type": "text", + "content": "ident at " + }, + { + "bbox": [ + 130, + 331, + 482, + 452 + ], + "type": "inline_equation", + "content": "\\lambda = 1" + }, + { + "bbox": [ + 130, + 331, + 482, + 452 + ], + "type": "text", + "content": ", where, the PGD-20 accuracy decreases for lower and higher perturbation values. For all datasets, the model performance is almost consistent across " + }, + { + "bbox": [ + 130, + 331, + 482, + 452 + ], + "type": "inline_equation", + "content": "\\lambda \\in \\{0.3, 0.5, 0.6\\}" + }, + { + "bbox": [ + 130, + 331, + 482, + 452 + ], + "type": "text", + "content": ", leading us to select the lowest value within this range for efficiency. On the ImageNet dataset, however, a higher value (0.7) enhances performance. To further assess the impact of hyperparameters " + }, + { + "bbox": [ + 130, + 331, + 482, + 452 + ], + "type": "inline_equation", + "content": "\\gamma_{1}" + }, + { + "bbox": [ + 130, + 331, + 482, + 452 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 331, + 482, + 452 + ], + "type": "inline_equation", + "content": "\\gamma_{2}" + }, + { + "bbox": [ + 130, + 331, + 482, + 452 + ], + "type": "text", + "content": " on model performance and adversarial robustness, we conducted experiments by varying their values and analyzing the outcomes on CIFAR100. The results are presented in Table 2 (supplementary file). The values are chosen from a set " + }, + { + "bbox": [ + 130, + 331, + 482, + 452 + ], + "type": "inline_equation", + "content": "\\{0.0, 0.3, 0.5, 0.7, 0.8, 1.0\\}" + }, + { + "bbox": [ + 130, + 331, + 482, + 452 + ], + "type": "text", + "content": ", and the model's performance was evaluated against several adversarial attacks, including FGSM, PGD-10, PGD-[20, 50], and AA." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 482, + 220, + 495 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 482, + 220, + 495 + ], + "spans": [ + { + "bbox": [ + 132, + 482, + 220, + 495 + ], + "type": "text", + "content": "5 Conclusion" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "content": "We presented an efficient model to overcome catastrophic overfitting in fast adversarial training through the Douglas-Rachford (DR) splitting technique, emphasizing linearity and stability in the training process. By using this technique, we provide a novel perspective on managing adversarial robustness that balances between minimizing adversarial loss and enforcing effective regularization. Indeed, using DR splitting allows for subtle control over the training process, facilitating adjustments in model parameters to prevent sudden degradation in adversarial robustness. This methodology not only addresses CO with greater efficiency than existing solutions like Fast-GA, Fast-AT, and NuAT but also enhances the overall training stability across a wide range of adversarial perturbations and training durations. Through extensive evaluations, with significant adversarial challenges and in conjunction with various Fast AT methods, our model demonstrates superior performance and adaptability." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 233, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 233, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 233, + 102 + ], + "type": "text", + "content": "Zareapoor et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 197, + 126 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 197, + 126 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 197, + 126 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 140, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 138, + 140, + 480, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 140, + 480, + 172 + ], + "spans": [ + { + "bbox": [ + 138, + 140, + 480, + 172 + ], + "type": "text", + "content": "1. Addepalli, S., Jain, S., Sriramanan, G., Venkatesh Babu, R.: Scaling adversarial training to large perturbation bounds. In: European Conference on Computer Vision. pp. 301-316 (2022) 6, 9" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 173, + 480, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 173, + 480, + 205 + ], + "spans": [ + { + "bbox": [ + 138, + 173, + 480, + 205 + ], + "type": "text", + "content": "2. Andriushchenko, M., Croce, F., Flammarion, N., Hein, M.: Square attack: a query-efficient black-box adversarial attack via random search. In: European conference on computer vision. pp. 484-501 (2020) 9, 10" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 206, + 480, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 206, + 480, + 237 + ], + "spans": [ + { + "bbox": [ + 138, + 206, + 480, + 237 + ], + "type": "text", + "content": "3. Andriushchenko, M., Flammarion, N.: Understanding and improving fast adversarial training. Advances in Neural Information Processing Systems 33, 16048-16059 (2020) 2, 3, 4, 5, 6, 10, 11, 12" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 238, + 480, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 238, + 480, + 270 + ], + "spans": [ + { + "bbox": [ + 138, + 238, + 480, + 270 + ], + "type": "text", + "content": "4. Applegate, D., Hinder, O., Lu, H., Lubin, M.: Faster first-order primal-dual methods for linear programming using restarts and sharpness. Mathematical Programming 201(1), 133-184 (2023) 2, 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 271, + 480, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 271, + 480, + 293 + ], + "spans": [ + { + "bbox": [ + 138, + 271, + 480, + 293 + ], + "type": "text", + "content": "5. Carlini, N., Wagner, D.: Towards evaluating the robustness of neural networks. In: 2017 IEEE symposium on security and privacy (sp). pp. 39-57 (2017) 9" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 293, + 480, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 293, + 480, + 326 + ], + "spans": [ + { + "bbox": [ + 138, + 293, + 480, + 326 + ], + "type": "text", + "content": "6. Chambolle, A., Contreras, J.P.: Accelerated bregman primal-dual methods applied to optimal transport and Wasserstein barycenter problems. SIAM Journal on Mathematics of Data Science 4(4), 1369-1395 (2022) 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 326, + 480, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 326, + 480, + 358 + ], + "spans": [ + { + "bbox": [ + 138, + 326, + 480, + 358 + ], + "type": "text", + "content": "7. Croce, F., Hein, M.: Reliable evaluation of adversarial robustness with an ensemble of diverse parameter-free attacks. In: International conference on machine learning. pp. 2206-2216 (2020) 9" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 359, + 480, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 359, + 480, + 392 + ], + "spans": [ + { + "bbox": [ + 138, + 359, + 480, + 392 + ], + "type": "text", + "content": "8. Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: Imagenet: A large-scale hierarchical image database. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 248-255 (2009) 8, 13" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 392, + 480, + 424 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 392, + 480, + 424 + ], + "spans": [ + { + "bbox": [ + 138, + 392, + 480, + 424 + ], + "type": "text", + "content": "9. Dong, Y., Liao, F., Pang, T., Su, H., Zhu, J., Hu, X., Li, J.: Boosting adversarial attacks with momentum. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9185-9193 (2018) 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 424, + 480, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 424, + 480, + 456 + ], + "spans": [ + { + "bbox": [ + 138, + 424, + 480, + 456 + ], + "type": "text", + "content": "0. Golgooni, Z., Saberi, M., Eskandar, M., Rohhan, M.H.: Zerograd: Mitigating and explaining catastrophic overfitting in fgsm adversarial training. arXiv preprint arXiv:2103.15476 (2021) 5, 10, 11, 12" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 457, + 480, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 457, + 480, + 479 + ], + "spans": [ + { + "bbox": [ + 138, + 457, + 480, + 479 + ], + "type": "text", + "content": "1. Goodfellow, I.J., Shlens, J., Szegedy, C.: Explaining and harnessing adversarial examples. International Conference on Learning Representations (2014) 1, 3, 4, 9" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 479, + 480, + 512 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 479, + 480, + 512 + ], + "spans": [ + { + "bbox": [ + 138, + 479, + 480, + 512 + ], + "type": "text", + "content": "2. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 770-778 (2016) 8" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 512, + 480, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 512, + 480, + 544 + ], + "spans": [ + { + "bbox": [ + 138, + 512, + 480, + 544 + ], + "type": "text", + "content": "3. He, K., Zhang, X., Ren, S., Sun, J.: Identity mappings in deep residual networks. In: Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part IV 14. pp. 630-645 (2016) 8" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 545, + 480, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 545, + 480, + 578 + ], + "spans": [ + { + "bbox": [ + 138, + 545, + 480, + 578 + ], + "type": "text", + "content": "4. He, Z., Li, T., Chen, S., Huang, X.: Investigating catastrophic overfitting in fast adversarial training: A self-fitting perspective. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2313-2320 (2023) 5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 578, + 480, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 578, + 480, + 610 + ], + "spans": [ + { + "bbox": [ + 138, + 578, + 480, + 610 + ], + "type": "text", + "content": "5. Jia, X., Zhang, Y., Wei, X., Wu, B., Ma, K., Wang, J., Cao, X.: Prior-guided adversarial initialization for fast adversarial training. In: European Conference on Computer Vision. pp. 567-584 (2022) 2, 6, 9, 10, 11, 12, 13" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 610, + 480, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 610, + 480, + 643 + ], + "spans": [ + { + "bbox": [ + 138, + 610, + 480, + 643 + ], + "type": "text", + "content": "6. de Jorge Aranda, P., Bibi, A., Volpi, R., Sanyal, A., Torr, P., Rogez, G., Dokania, P.: Make some noise: Reliable and efficient single-step adversarial training. Advances in Neural Information Processing Systems 35, 12881-12893 (2022) 2, 5, 13" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 138, + 643, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 643, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 643, + 480, + 665 + ], + "type": "text", + "content": "7. Kurakin, A., Goodfellow, I., Bengio, S.: Adversarial machine learning at scale. arXiv preprint arXiv:1611.01236 (2016) 3" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 295, + 91, + 447, + 102 + ], + "type": "text", + "content": "Rethinking Fast Adversarial Training" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "type": "text", + "content": "18. Li, T., Wu, Y., Chen, S., Fang, K., Huang, X.: Subspace adversarial training. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13409-13418 (2022) 1, 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 150, + 481, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 150, + 481, + 183 + ], + "spans": [ + { + "bbox": [ + 133, + 150, + 481, + 183 + ], + "type": "text", + "content": "19. Liang, J., Fadili, J., Peyre, G.: Local convergence properties of douglas-rachford and alternating direction method of multipliers. Journal of Optimization Theory and Applications 172, 874-913 (2017) 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 183, + 481, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 183, + 481, + 215 + ], + "spans": [ + { + "bbox": [ + 132, + 183, + 481, + 215 + ], + "type": "text", + "content": "20. Lindbäck, J., Wang, Z., Johansson, M.: Bringing regularized optimal transport to lightspeed: a splitting method adapted for gpus. Advances in Neural Information Processing Systems 36 (2023) 2, 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 216, + 481, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 216, + 481, + 248 + ], + "spans": [ + { + "bbox": [ + 132, + 216, + 481, + 248 + ], + "type": "text", + "content": "21. Liu, X., Chakraborty, S., Sun, Y., Huang, F.: Rethinking adversarial policies: A generalized attack formulation and provable defense in rl. In: International Conference on Learning Representations (2024) 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 249, + 481, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 249, + 481, + 281 + ], + "spans": [ + { + "bbox": [ + 132, + 249, + 481, + 281 + ], + "type": "text", + "content": "22. Madry, A., Makelov, A., Schmidt, L., Tsipras, D., Vladu, A.: Towards deep learning models resistant to adversarial attacks. International Conference on Learning Representations (2018) 1, 3, 9" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 281, + 481, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 281, + 481, + 314 + ], + "spans": [ + { + "bbox": [ + 132, + 281, + 481, + 314 + ], + "type": "text", + "content": "23. Mai, V.V., Lindbäck, J., Johansson, M.: A fast and accurate splitting method for optimal transport: Analysis and implementation. International Conference on Learning Representations (2022) 2, 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 315, + 481, + 347 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 315, + 481, + 347 + ], + "spans": [ + { + "bbox": [ + 132, + 315, + 481, + 347 + ], + "type": "text", + "content": "24. Moosavi-Dezfooli, S.M., Fawzi, A., Uesato, J., Frossard, P.: Robustness via curvature regularization, and vice versa. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9078-9086 (2019) 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 347, + 481, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 347, + 481, + 380 + ], + "spans": [ + { + "bbox": [ + 132, + 347, + 481, + 380 + ], + "type": "text", + "content": "25. Park, G.Y., Lee, S.W.: Reliably fast adversarial training via latent adversarial perturbation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 7758-7767 (2021) 1, 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 380, + 481, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 380, + 481, + 413 + ], + "spans": [ + { + "bbox": [ + 132, + 380, + 481, + 413 + ], + "type": "text", + "content": "26. Poon, C., Liang, J.: Trajectory of alternating direction method of multipliers and adaptive acceleration. Advances in Neural Information Processing Systems 32 (2019) 2, 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 413, + 481, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 413, + 481, + 445 + ], + "spans": [ + { + "bbox": [ + 132, + 413, + 481, + 445 + ], + "type": "text", + "content": "27. Qin, C., Martens, J., Gowal, S., Krishnan, D., Dvijotham, K., Fawzi, A., De, S., Stanforth, R., Kohli, P.: Adversarial robustness through local linearization. Advances in Neural Information Processing Systems 32 (2019) 3, 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 446, + 481, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 446, + 481, + 479 + ], + "spans": [ + { + "bbox": [ + 132, + 446, + 481, + 479 + ], + "type": "text", + "content": "28. Rocamora, E.A., Liu, F., Chrysos, G.G., Olmos, P.M., Cevher, V.: Efficient local linearity regularization to overcome catastrophic overfitting. International Conference on Learning Representations (2024) 1, 5, 6, 11" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 479, + 481, + 501 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 479, + 481, + 501 + ], + "spans": [ + { + "bbox": [ + 132, + 479, + 481, + 501 + ], + "type": "text", + "content": "29. Shaeiri, A., Nobahari, R., Rohhan, M.H.: Towards deep learning models resistant to large perturbations. arXiv preprint arXiv:2003.13370 (2020) 9" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 501, + 481, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 501, + 481, + 533 + ], + "spans": [ + { + "bbox": [ + 132, + 501, + 481, + 533 + ], + "type": "text", + "content": "30. Shafahi, A., Najibi, M., Ghiasi, M.A., Xu, Z., Dickerson, J., Studer, C., Davis, L.S., Taylor, G., Goldstein, T.: Adversarial training for free! Advances in Neural Information Processing Systems 32 (2019) 1, 4" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 534, + 481, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 534, + 481, + 566 + ], + "spans": [ + { + "bbox": [ + 132, + 534, + 481, + 566 + ], + "type": "text", + "content": "31. Song, C., He, K., Wang, L., Hopcroft, J.E.: Improving the generalization of adversarial training with domain adaptation. International Conference on Learning Representations (2019) 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 567, + 481, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 567, + 481, + 599 + ], + "spans": [ + { + "bbox": [ + 132, + 567, + 481, + 599 + ], + "type": "text", + "content": "32. Sriramanan, G., Addepalli, S., Baburaj, A., et al.: Guided adversarial attack for evaluating and enhancing adversarial defenses. Advances in Neural Information Processing Systems 33, 20297-20308 (2020) 5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 600, + 481, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 600, + 481, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 600, + 481, + 632 + ], + "type": "text", + "content": "33. Sriramanan, G., Addepalli, S., Baburaj, A., et al.: Towards efficient and effective adversarial training. Advances in Neural Information Processing Systems 34, 11821-11833 (2021) 2, 5, 10, 11, 12" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 632, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 481, + 665 + ], + "type": "text", + "content": "34. Themelis, A., Patrinos, P.: Douglas-rachford splitting and admn for nonconvex optimization: Tight convergence results. SIAM Journal on Optimization 30(1), 149-181 (2020) 5" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 233, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 233, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 233, + 102 + ], + "type": "text", + "content": "Zareapoor et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 358 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "text", + "content": "35. Tsiligkaridis, T., Roberts, J.: Understanding and increasing efficiency of frank-wolfe adversarial training. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 50-59 (2022) 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 150, + 482, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 150, + 482, + 182 + ], + "spans": [ + { + "bbox": [ + 130, + 150, + 482, + 182 + ], + "type": "text", + "content": "36. Wong, E., Rice, L., Kolter, J.Z.: Fast is better than free: Revisiting adversarial training. International Conference on Learning Representations (2020) 1, 2, 4, 6, 9, 10, 11, 12, 13" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 182, + 482, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 182, + 482, + 205 + ], + "spans": [ + { + "bbox": [ + 130, + 182, + 482, + 205 + ], + "type": "text", + "content": "37. Wu, D., Xia, S.T., Wang, Y.: Adversarial weight perturbation helps robust generalization. Advances in Neural Information Processing Systems (2020) 9" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 205, + 482, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 205, + 482, + 237 + ], + "spans": [ + { + "bbox": [ + 130, + 205, + 482, + 237 + ], + "type": "text", + "content": "38. Xie, Y., Li, Z., Shi, C., Liu, J., Chen, Y., Yuan, B.: Enabling fast and universal audio adversarial attack using generative model. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 35, pp. 14129-14137 (2021) 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 237, + 482, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 237, + 482, + 259 + ], + "spans": [ + { + "bbox": [ + 130, + 237, + 482, + 259 + ], + "type": "text", + "content": "39. Zagoruyko, S., Komodakis, N.: Wide residual networks. arXiv preprint arXiv:1605.07146 (2016) 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 259, + 482, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 259, + 482, + 293 + ], + "spans": [ + { + "bbox": [ + 130, + 259, + 482, + 293 + ], + "type": "text", + "content": "40. Zhang, H., Yu, Y., Jiao, J., Xing, E., El Ghaoui, L., Jordan, M.: Theoretically principled trade-off between robustness and accuracy. In: International conference on machine learning. pp. 7472-7482 (2019) 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 293, + 482, + 325 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 293, + 482, + 325 + ], + "spans": [ + { + "bbox": [ + 130, + 293, + 482, + 325 + ], + "type": "text", + "content": "41. Zhang, Y., Zhang, G., Khanduri, P., Hong, M., Chang, S., Liu, S.: Revisiting and advancing fast adversarial training through the lens of bi-level optimization. In: International Conference on Machine Learning. pp. 26693-26712 (2022) 3, 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 325, + 482, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 325, + 482, + 358 + ], + "spans": [ + { + "bbox": [ + 130, + 325, + 482, + 358 + ], + "type": "text", + "content": "42. Zhao, M., Zhang, L., Kong, Y., Yin, B.: Fast adversarial training with smooth convergence. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 4720-4729 (2023) 2, 6" + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 295, + 91, + 447, + 102 + ], + "type": "text", + "content": "Rethinking Fast Adversarial Training" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Rethinking Features-Fused-Pyramid-Neck for Object Detection/8cefd2b8-bab0-4ef4-a711-578971c2749a_content_list.json b/2024/Rethinking Features-Fused-Pyramid-Neck for Object Detection/8cefd2b8-bab0-4ef4-a711-578971c2749a_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..f3695be09783e5ff109d81df95ed50b038ea3cac --- /dev/null +++ b/2024/Rethinking Features-Fused-Pyramid-Neck for Object Detection/8cefd2b8-bab0-4ef4-a711-578971c2749a_content_list.json @@ -0,0 +1,1345 @@ +[ + { + "type": "text", + "text": "Rethinking Features-Fused-Pyramid-Neck for Object Detection", + "text_level": 1, + "bbox": [ + 243, + 138, + 759, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Hulin Li", + "bbox": [ + 464, + 212, + 539, + 226 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "College of Traffic and Transportation, Chongqing Jiaotong University, Chongqing", + "bbox": [ + 230, + 239, + 772, + 253 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "400074, China", + "bbox": [ + 452, + 255, + 550, + 267 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "alan@mails.cqjtu.edu.cn", + "bbox": [ + 411, + 268, + 591, + 282 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Multi-head detectors typically employ a features-fused-pyramid-neck for multi-scale detection and are widely adopted in the industry. However, this approach faces feature misalignment when representations from different hierarchical levels of the feature pyramid are forcibly fused point-to-point. To address this issue, we designed an independent hierarchy pyramid (IHP) architecture to evaluate the effectiveness of the features-unfused-pyramid-neck for multi-head detectors. Subsequently, we introduced soft nearest neighbor interpolation (SNI) with a weight-downscaling factor to mitigate the impact of feature fusion at different hierarchies while preserving key textures. Furthermore, we present a feature adaptive selection method for downsampling in extended spatial windows (ESD) to retain spatial features and enhance lightweight convolutional techniques (GSCnvE). These advancements culminate in our secondary features alignment solution (SA) for real-time detection, achieving state-of-the-art results on Pascal VOC and MS COCO. Code will be released at https://github.com/AlanLi1997/rethinking-fpn.", + "bbox": [ + 261, + 320, + 759, + 542 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Object detection $\\cdot$ Feature pyramid $\\cdot$ Feature misalignment", + "bbox": [ + 261, + 556, + 740, + 570 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "- Detection-net architecture", + "bbox": [ + 264, + 571, + 449, + 582 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 217, + 611, + 375, + 626 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Bounding box regression for object detection and pixel-level boundary regression for instance segmentation are pivotal and challenging visual tasks. Unlike basic image classification [4], detection or segmentation tasks often involve scenes with multiple objects of varying scales. Early deep learning-based object detection models began exploring real-time multi-scale object detection. RCNN [9] employed selective search to extract a large number of region proposals ( $\\sim 2\\mathrm{k}$ ). Although the vast number of image patches (region proposals) might aid multiscale detection, in practice, most scenes do not contain such a high number of objects to recognize. This approach was later deemed inefficient [8], as the proposed regions were derived from the image rather than features, necessitating a convolutional neural network (CNN) to extract features and predict classes for each proposal. A simplified RCNN architecture, as depicted in Fig. 1 (a), converts images of different scales into feature maps for object detection, largely", + "bbox": [ + 212, + 643, + 787, + 840 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/4c00a44d7d5e66f966c965977af8b1a4e6eca56737b7313dc346c4f8d83f0de6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 276, + 145, + 480, + 262 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/46e922b9812dc914407c5906f7d4679413f46f612b34e326933f9c8c30f618eb.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 488, + 145, + 594, + 261 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/07fc4afd6132de0f0819f2796adf83869bf34853c4df2a337a1583fca55c6a96.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 607, + 146, + 720, + 261 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/e235f6347163f4ca2ab3ad192eab4fa8ba580dd26370239588b1662c5193bf38.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 276, + 263, + 562, + 375 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/6eea46c13ef56cb9ed5041b7b18f38bb731dbbbb9057366f8718144a1063fb3e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 571, + 262, + 725, + 375 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/1c5625fcb681996bf3c0a52916ed6016f3634610c6112e793b0886b8225da4ef.jpg", + "image_caption": [ + "Fig. 1: The IHP and seven other typical neck architectures. The three heat maps are come from the third (P3), fourth (P4), and fifth (P5) hierarchies of the FPN-neck model, respectively. The area of interest for each hierarchies' features is highlighted in red; the darker the color, the higher the feature weight. It is evident that lower-level hierarchies prioritize local features, while higher-level hierarchies focus more on global features. In other words, there is a representation bias at different hierarchical levels. If these features are combined by point-to-point fusion without first addressing the misalignment issue, it may introduce noise rather than enrich semantics." + ], + "image_footnote": [], + "bbox": [ + 276, + 376, + 575, + 489 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/c286cbcb9e9125c4ee6a164748652395c728ef953caddb8cfad2f13b9420d70b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 588, + 376, + 720, + 489 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "neglecting the real-time detection capability of models. YOLO [29] introduced the first real-time deep learning-based object detection model, which did not specifically focus on multi-scale detection, with a simplified structure shown in Fig. 1 (b). SSD [25] effectively combined the strengths of RCNN and YOLO, emphasizing the concept of multi-scale detection. It was the first model to achieve real-time multi-scale object detection using multi-scale features, with an architecture similar to Fig. 1 (c). SSD efficiently utilized feature maps of different scales from the VGG-16 backbone [32] to predict objects of varying scales. The exploration by early deep learning-based object detection models provided an important insight to the field: the efficient utilization of multi-scale features is crucial for enhancing the overall performance (accuracy-speed trade-off) of detectors. Consequently, FPN [21] introduced the feature pyramid-based multi-scale", + "bbox": [ + 212, + 657, + 787, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "Hulin Li", + "bbox": [ + 271, + 114, + 330, + 127 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "feature fusion concept, which gradually became a structural paradigm for detectors. This led to numerous representative works enhancing multi-scale feature fusion, such as PANet [24], ASFF [23], BiFPN [34], and RFPN [28]. The performance of real-time detectors has improved with the implementation of feature fusion strategies. However, further advancements are hindered by the increasing complexity of these techniques. Before the introduction of FPN, object detectors primarily relied on structures without fusion [9,25,29], as depicted in Fig. 1 (a)-(c). With the advent of FPN, these detectors have shifted towards FPN-based paradigms or more intricate improvement methods [8,20,29-31], as illustrated in Fig. 1 (e)-(g). It is important to note that the FPN-like strategy should not be considered an obligatory component for all detectors. Representation learning suggests that models ultimately require refined representations rather than an abundance of complex representations. Under the FPN-like paradigm, feature maps of different hierarchies and different resolutions are combined at the element level (point-to-point), necessitating the expansion of low-resolution, high-level feature maps to match the resolution of low-level feature maps. However, there are inherent representation biases in feature maps at different hierarchical levels: low-level feature maps tend to retain representations of small objects (local features in small receptive fields), while high-level feature maps favor representations of large objects (global features in large receptive fields), as shown in Fig. 1 (i). The direct element-level fusion of these feature maps with representation bias may lead to partial destruction of representations and exacerbate the feature misalignment problem during fusion.", + "bbox": [ + 217, + 146, + 785, + 492 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In visual models, the resolution of feature maps decreases progressively from low to high hierarchical levels of the feature pyramid, achieved through step-by-step downsampling. Typically, downsampling reduces the resolution (width $\\times$ height) of feature maps at a square root rate, inherently filtering out some spatial information regardless of whether the number of features (channels) remains constant or increases. It is also noteworthy that the transition of feature information from spatial to channel dimensions primarily occurs during downsampling, underscoring the importance of this stage. Traditional downsampling methods operate in a single mode, such as convolution with a stride of 2 (learnable) or average pooling and max pooling with a stride of 2 (non-learnable). These single-mode approaches have persisted since the introduction of ResNet [12] without significant evolution, which merits further investigation. Additionally, the initial purpose of FPN was not only to demonstrate the advantages of feature fusion in addressing multi-scale detection challenges but also to appropriately reduce model complexity. In industries with limited hardware resources, complex models are often impractical [2] [10]. Therefore, we enhance the lightweight convolution technique of the GSCov [18] to achieve a preferable trade-off between accuracy and speed for real-time detection models.", + "bbox": [ + 217, + 500, + 785, + 768 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The main contributions of this work can be summarized as follows:", + "bbox": [ + 240, + 775, + 720, + 789 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "1). We rethink the effectiveness of FPN-like paradigms for modern real-time detectors based on representation learning and identify issues with feature misalignment during element-level fusion.", + "bbox": [ + 215, + 794, + 784, + 839 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "Rethinking Features-Fused-Pyramid-Neck for Object Detection", + "bbox": [ + 310, + 114, + 730, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "2). We design an independent hierarchy pyramid architecture (IHP) without feature fusion in the neck to validate our findings. And the IHP achieves advanced results.", + "3). We introduce soft nearest neighbor interpolation (SNI) for up-sampling to mitigate feature misalignment during fusion.", + "4). We provide a feature adaptive selection method in extended spatial windows for downsampling (ESD) to enhance spatial feature retention during the downsampling stage. This method is plug-and-play and optimizes performance at a low cost.", + "5). We simplify and enhance the GSCnv lightweight convolution (GSCnvE) to offer more cost-effective options for models operating on edge devices with resource constraints.", + "6).We validate our approaches on Pascal VOC and COCO and present the SA solution. Existing lightweight real-time detection models can be optimized with the SA solution without bells and whistles." + ], + "bbox": [ + 212, + 146, + 784, + 376 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 407, + 387, + 422 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.1 Multi-head detection and features fusion", + "text_level": 1, + "bbox": [ + 215, + 445, + 596, + 459 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The first-generation general detection models using deep learning consist of two main components: the backbone and the detecting-head [9, 29]. These models typically use final feature maps for prediction. SSD [25] introduced the use of multi-level feature maps for object detection, marking the beginning of multi-head detection and influencing subsequent real-time detector designs [31]. In the SSD model, each detecting-head directly uses raw feature maps from different levels of the backbone without any fusion. The key advantage of the multi-head detection method is the ability to predict objects of different scales using feature maps with varying receptive fields. High-level feature maps, with their wide global receptive fields, are beneficial for identifying large objects, while low-level feature maps, with refined local receptive fields, are helpful for identifying small objects. FPN introduced a fusion scheme for multi-level feature maps to enhance detection accuracy, setting the trend for feature fusion and becoming a common practice. This led to the development of Backbone-Neck-Head architectures. More complex variants based on FPN, such as BiFPN [34] and PANet [24], have since been proposed, enabling the use of more features for prediction. However, as prediction accuracy improved, networks became increasingly complex due to additional fusion layers. Some approaches have attempted to bypass feature fusion schemes by exploring new perspectives on multi-scale detection challenges, such as optimizing multi-scale training strategies [33] or using dilated convolutions to capture multi-scale receptive fields [19]. But these methods have gradually lost their competitiveness with the continuous optimization of classification-backbone feature extraction capabilities and the ongoing development of feature fusion techniques [1].", + "bbox": [ + 212, + 478, + 787, + 843 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "Hulin Li", + "bbox": [ + 271, + 114, + 330, + 127 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.2 Lightweight", + "text_level": 1, + "bbox": [ + 215, + 146, + 359, + 161 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The choice between cloud computing or edge computing for a model depends on its number of parameters and floating-point operations (FLOPs). Therefore, reducing the number of parameters or FLOPs is a primary focus in lightweight model studies. Direct lightweight approaches include reducing the depth (number of layers) or width (number of neurons/filters) of the model, as seen in models likeYOLO-fast/tiny/nano [1,14-16,31,35]. Additionally, sparse computing techniques, such as depth-wise separable convolution used in MobileNets [13] and ShuffleNet [39], can also be employed. However, low-depth networks often suffer from underfitting due to insufficient nonlinear representation capabilities. Therefore, reducing the number of network layers may not always be a cost-effective way to lighten the model.", + "bbox": [ + 212, + 172, + 787, + 339 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3 Secondary Features Alignment Solution", + "text_level": 1, + "bbox": [ + 215, + 363, + 633, + 382 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Key contents of this work are described in detail in this section. Specifically, there are the IHP structure to directly demonstrate features unalignment problem by abandoning fusion, the SNI to alleviate unaligned during features fusion by point-to-point, the ESD to enhance spatial-features capture during the downsampling stage, and the GSConvE to improve the performance between the accuracy and speed of lightweight models.", + "bbox": [ + 212, + 396, + 787, + 488 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1 Independent hierarchy pyramid architecture", + "text_level": 1, + "bbox": [ + 215, + 511, + 627, + 527 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The introduction of FPN brought all real-time detectors into the realm of feature fusion. However, we have identified a significant issue arising from the point-to-point fusion of different level features - partial local features become unaligned. This fusion process is akin to adding noise when unaligned features are combined, as features unrelated to the target space are forcibly integrated, resulting in spatial disarray. For instance, as depicted in Fig. 2, adding the head-feature of a puppy to the butt creates an undefined breed when fed to the detection-head. This scenario is particularly evident when the puppy occupies fewer pixels. Typically, nearest neighbor interpolation is used to rapidly increase the resolution of high-level feature maps to match that of low-level feature maps, followed by fusion through methods such as element-wise addition, channel concatenation, or weighted sum. However, downsampling tends to be nonlinear and irreversible, leading to inconsistencies in spatial features among the additional features generated by up-sampling. Although BiFPN [34] offers a learnable fusion method, it has not been widely adopted by recent state-of-the-art real-time detectors, such as YOLOv7 [35] or YOLOv8 [15], due to the increased complexity and lack of demonstrated comprehensive performance improvements in these models.", + "bbox": [ + 212, + 537, + 787, + 792 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The IHP adopts a radical approach by abandoning all fusions in the neck, fundamentally circumventing the features misalignment problem and streamlining the structure. By leveraging the inherent benefits of the multi-head detection", + "bbox": [ + 212, + 795, + 785, + 839 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "Rethinking Features-Fused-Pyramid-Neck for Object Detection", + "bbox": [ + 310, + 114, + 730, + 130 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "mode, the IHP directly predicts objects of different sizes using feature maps of different levels. It is important to note that there are significant differences between IHP and the SSD architecture: SSD directly uses different level features from the backbone to detect objects but the IHP introduces a bottleneck convolution module before prediction to filter the features. The advantage of this approach is that it allows the model to independently learn features for a specific scale branch without affecting features at other levels. Directly stacking convolutional blocks at different levels of the backbone can lead to interference between different levels, which we believe might be the primary reason why using a very deep backbone in object detection models can result in performance degradation. For instance, DSSD [6] showed a performance drop when ResNet-101 was used as the backbone instead of VGG-16 in the SSD detector. In Part I of Tab. 1, all baselines of coupled-head YOLOs demonstrated improved results when using the IHP. Features fusion often overlooks the potential introduction of noise when augmenting semantic richness for prediction. While the IHP may appear to reduce semantic information, it effectively harnesses the advantages of multi-head detection to circumvent the misalignment problem. This strategic utilization of multi-head detection is the key factor behind the competitive results achieved by the IHP. Unfortunately, unlike localization-classification-coupled-head detectors [14,29-31,35], the IHP did not achieve the same positive reaction to localization-classification-decoupled-head detectors [15, 16]. Decoupled-head detectors address a key challenge faced by coupled-head detectors: the disparity in the interests of classification and localization tasks, both of which are often performed using the same layer for prediction [36]. This approach is akin to forcibly fusing misaligned features for prediction. Decoupled-head detectors overcome this limitation by independently predicting classification and localization tasks through separate branches. This enables them to learn beneficial representations from redundant features, which is not possible in coupled-head detectors. Consequently, while FPN strategies are still applicable to some extent in decoupled-head detectors, the features misalignment problem still needs to be addressed.", + "bbox": [ + 212, + 145, + 787, + 616 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.2 Soft nearest neighbor interpolation", + "text_level": 1, + "bbox": [ + 215, + 635, + 552, + 651 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In detection or segmentation models, different-level feature fusion typically occurs after the up-sampling of feature maps, with nearest neighbor interpolation and transposed convolution being popular up-sampling methods. However, the resolution expansion resulting from up-sampling in detection or segmentation differs significantly from tasks such as generation or super-resolution. In detection, the expanded feature maps represent abstract high-level semantic features rather than raw image detail information. Transposed convolution introduces increased computation and latency. To address the features misalignment problem with minimal cost while maintaining the speed of nearest neighbor interpolation, we explore a solution. Nearest neighbor interpolation is akin to average unpooling but is considered a \"hard\" operation. By softening this operation, similar to the transformation of SoftMax to the Max function, we aim to mitigate the", + "bbox": [ + 212, + 657, + 787, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "Hulin Li", + "bbox": [ + 271, + 114, + 330, + 127 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/2a45fad2bdaaec7b2d52335aa31b61e76229a697d5c86f5d4ceba1f5c5d3e169.jpg", + "image_caption": [ + "Fig. 2: Illustrations of the features misalignment in fusion and the SNI. These numbers, 22-88, are just markers of different local-features not real feature values." + ], + "image_footnote": [], + "bbox": [ + 220, + 147, + 781, + 311 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/beda720ecd6fb3d8211cbb77d398261dbb1d209f5e82a3ddd847392ca09937e4.jpg", + "image_caption": [ + "Fig.3: A comparison of the SNI and traditional method. This result could be reproduced by using SNI source code (model:Yolov5n-panet). The same size scaling before and after up-sampling is for intuitive comparison." + ], + "image_footnote": [], + "bbox": [ + 279, + 378, + 725, + 559 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "\"hard\" features misalignment issue. To achieve this, we introduce a soft factor, denoted as $\\alpha$ , for the nearest neighbor interpolation during up-sampling:", + "bbox": [ + 215, + 641, + 784, + 672 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {Y} = \\alpha \\cdot \\boldsymbol {f} (\\boldsymbol {X}), \\alpha = \\frac {\\text {R e s o l u t i o n} \\boldsymbol {X}}{\\text {R e s o l u t i o n} \\boldsymbol {Y}} \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 380, + 680, + 785, + 710 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In Eq. (1), the resolution of the high-level feature maps $(\\mathbf{X})$ and low-level feature maps $(\\mathbf{Y})$ are denoted as $Resolution_{X}$ and $Resolution_{Y}$ , respectively. The nearest neighbor interpolation operation, represented by $\\pmb{f}$ , is illustrated in Fig. 2. The SNI adjusts the influence of high-level semantic features on low-level features based on the zoom factor for feature maps. Specifically, as the zoom factor increases, the impact of high-level semantic features on low-level features weakens. Unlike SoftMax, the SNI does not convert outputs into probabilities but rather mitigates misalignment without additional cost when integrating high-", + "bbox": [ + 212, + 718, + 787, + 840 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "Rethinking Features-Fused-Pyramid-Neck for Object Detection", + "bbox": [ + 310, + 114, + 732, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "level features with low-level features through point-to-point fusion. SNI softens feature weights while preserving key textures, thereby reducing the mutual influence of features from different hierarchical levels during fusion, as qualitatively demonstrated in Fig. 3. This approach achieves the secondary (auxiliary) features alignment (SA) and lowers the complexity for the model to learn from the fused representations. In contrast, traditional methods that directly merge features from different hierarchical levels can cause unstable competitive effects due to representation biases, increasing the learning difficulty for the model. For example, if the three heatmaps in Fig. 1 (i) are directly fused, should the red areas cover the entire image? Clearly, the answer is no. If such a situation occurs, it indicates a failure in prior feature extraction, as the objective of a model is to capture and focus on useful features. This is also why models under the FPN-like paradigm typically require several additional convolutional blocks to further filter features after fusing and before predicting.", + "bbox": [ + 212, + 146, + 787, + 357 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In Part II of Tab. 1, all tested state-of-the-art models demonstrate accuracy improvements solely by incorporating the SNI, confirming the presence of features misalignment in FPN-based models. Notably, for YOLOv6 [16], a detector equipped with various techniques tailored for industry applications, the AP is enhanced by up to 2 percentage points just with the adoption of SNI, while reducing computational cost and inference latency. We achieved significant optimization results in the original YOLOv6 by simply replacing transposed convolution with SNI, highlighting the efficacy of this approach.", + "bbox": [ + 212, + 358, + 787, + 479 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "3.3 Feature adaptive selection in extended spatial windows", + "text_level": 1, + "bbox": [ + 214, + 500, + 714, + 515 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/28d65be42a1f275f2d41a54dffb09210d9bbcb1d78bf489f100e47d45969f9f7.jpg", + "image_caption": [ + "Fig. 4: Illustrations of the ESD-I and ESD-II. It must be mentioned that although the size of the extended window does not have to be same in the ESD-I, pooling is a non-learnable operation, so the features from an unequal window may become noise for other and bring the trouble of the features misalignment again. The ESD-II is free on size of the extended window because of the learnable fusion like the SPPNet [11] for the YOLOv3." + ], + "image_footnote": [], + "bbox": [ + 222, + 551, + 781, + 652 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Feature downsampling is a critical technique in vision models based on deep learning, involving the progressive reduction of feature map resolution. In vision representation learning, models typically rely on low-resolution, high-level", + "bbox": [ + 212, + 794, + 787, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "Hulin Li", + "bbox": [ + 271, + 114, + 330, + 127 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "feature maps for prediction rather than directly processing high-resolution raw images. However, excessive downsampling can result in significant loss of spatial details [12]. Recent detectors typically limit the number of downsampling operations to no more than five times, leading to a spatial resolution reduction of $2^{5}$ times, as demonstrated in YOLOs [1,7,14-16,31,35,37] and Swin Transformer [26]. Common downsampling methods include convolution and pooling, with models typically employing only one of these methods. We propose the feature adaptive selection method in extended spatial windows for downsampling (ESD) to enhance spatial information retention capabilities. ESD consists of two nonlinear branches and one linear branch. In the nonlinear branches, a norm convolutional layer and an extended window max-pooling layer enhance local feature capturing, while extended window average-pooling enhances global feature capturing. Subsequently, linear features are merged with nonlinear features. Two fusion modes, ESD-I and ESD-II, are designed for lightweight and norm models, respectively. The difference lies in the feature fusion stage, as illustrated in Fig. 4. ESD-I employs element-wise addition for merging features, offering low computational cost suitable for lightweight models, while ESD-II utilizes learnable linearly adaptive fusion, slightly increasing computational cost but suitable for norm models. Importantly, local and global features of the extended window are sampled using simple pooling techniques, preserving input information to a significant extent. This reduces information loss during downsampling and enables subsequent layers to learn partial representations from the previous layer, akin to indirectly realizing a hidden shortcut connection reminiscent of ResNet [12].", + "bbox": [ + 212, + 146, + 787, + 508 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "3.4 Lightweight GSConv enhancement", + "text_level": 1, + "bbox": [ + 215, + 530, + 545, + 545 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/159ea320bc49abbac2e9443c24897e0139bfdebb01c805a385d2a84699a4f3cb.jpg", + "image_caption": [ + "Fig. 5: Illustrations of the GSConvE-I and II." + ], + "image_footnote": [], + "bbox": [ + 250, + 582, + 753, + 694 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "GSCnv [18] was introduced to address the channel interaction limitations of depthwise separable convolution by blending features from both vanilla convolution and depthwise convolution. The original GSCnv consists of three fundamental components: a $3 \\times 3$ vanilla convolution, a $5 \\times 5$ depthwise convolution, and a feature shuffle operation. Building upon this, we propose two evolutions", + "bbox": [ + 212, + 763, + 787, + 840 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "Rethinking Features-Fused-Pyramid-Neck for Object Detection", + "bbox": [ + 310, + 114, + 730, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "of GSConv, namely GSConvE-I and GSConvE-II, tailored for norm models and lightweight models, respectively. GSConvE-I is designed for norm models, where intermediate feature maps on the auxiliary branch are derived from dense linear mappings, while output feature maps originate from sparse linear mappings. This approach maintains feature richness without significantly increasing computational cost. On the other hand, GSConvE-II is geared towards lightweight models. It incorporates three depthwise convolution auxiliary branches with kernel sizes of $9 \\times 9$ , $13 \\times 13$ , and $17 \\times 17$ , respectively. By utilizing larger kernel sizes, this variant can directly capture larger receptive fields and global features with fewer layer accumulations, offering high cost-effectiveness for lightweight models that prioritize reduced FLOPs and layer count. In GSConvE-I, we explore the removal of the batch normalization layer on the depthwise convolution branch while retaining the batch normalization layer on the main branch, similar to ConvNeXt [27]. This adjustment avoids the vanishing gradient problem and simplifies computation, leading to improved prediction accuracy.", + "bbox": [ + 212, + 146, + 787, + 375 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 215, + 398, + 375, + 416 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Real-time detectors are highly favored in industry due to their hardware-friendly nature and the optimal balance they strike between accuracy and speed. In our study, we selected popular YOLO models—YOLOv3 [31], YOLOv4 [1], YOLOv5 [14], YOLOv6 [16], YOLOv7 [35], and YOLOv8 [15]—as baselines to evaluate the efficacy of the SA solution in real-time detection scenarios (SYOLO).", + "bbox": [ + 212, + 431, + 787, + 510 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.1 Datasets", + "text_level": 1, + "bbox": [ + 215, + 532, + 334, + 547 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "We conducted trials on the Pascal VOC dataset [5] and the MS COCO dataset [22]. The Pascal VOC 07+12 dataset was utilized for ablation experiments, wherein models underwent training on the train&val07+12 set and were subsequently evaluated on the test07 set. On the other hand, the MS COCO dataset was employed for comparative experiments, with models trained on the train2017 set and evaluated on the val2017 set. To ensure fairness and consistency, all released models were trained using the default set of hyperparameters and without utilizing pre-trained weights. More details of hyperparameters are in the support material.", + "bbox": [ + 212, + 560, + 787, + 696 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.2 Ablation studies", + "text_level": 1, + "bbox": [ + 215, + 720, + 398, + 736 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "In Tab. 1, we present the results of ablation experiments conducted on six baselines [1, 14-16, 31, 35], focusing on the impact of the IHP, SNI, ESD. The lightweight experiment results of GSConvE are reported in Tab. 2. All baselines retained their original neck architecture, as we believe the competitive neck has been carefully chosen in the original works. Performance metrics highlighted in green indicate performance gains, while those in red indicate losses. It is crucial", + "bbox": [ + 212, + 750, + 787, + 842 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "Hulin Li", + "bbox": [ + 271, + 114, + 330, + 127 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "to address an often-overlooked aspect during the verification process: the inconsistency between reported evaluation accuracy and actual usage. Frequently, detectors are evaluated with a very low confidence threshold, such as 0.001, which results in a large number of prediction boxes and inflated evaluation accuracy. However, in real-world detection scenarios, the confidence threshold is typically set much higher, around 0.25 or above. This filtering process ensures that higher-quality bounding boxes are utilized for practical purposes. The discrepancy between evaluation settings and real-world usage can create a misleading impression of accuracy. Therefore, it is essential to maintain consistency in configuration between evaluation and application. While this may result in a significant drop in evaluation accuracy, as shown in Tab. 3, it provides a more realistic assessment of model performance.", + "bbox": [ + 212, + 146, + 787, + 327 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/5ac1f3bd6c3ea3199e2473d5a03879c0878cb861663c4e4e0438079afce0242c.jpg", + "table_caption": [ + "Table 1: Ablation experiments of the IHP, SNI and ESD-I/II on VOC $07 + 12$" + ], + "table_footnote": [], + "table_body": "
ModelsSizeParam.(M)FLOPs(G)AP50(%)AP(%)LatencyT4b=16(ms)
basielines
Yolov3-t [31]6408.7113.058.027.33.8
Yolov4-t [1]6405.9116.162.632.24.4
Yolov5-n [14]6401.794.274.047.03.4
Yolov6-n [16]6404.3011.180.258.83.5
Yolov7-t [35]6406.0613.378.454.65.0
Yolov8-n [15]6403.018.179.859.73.7
Part I: IHP (ours)
Yolov3-t-ihp6409.3315.059.9(↑1.9)29.1(↑1.8)3.8
Yolov4-t-ihp6406.6316.063.8(↑1.2)33.9(↑1.7)4.1
Yolov5-n-ihp6401.663.974.8(↑0.8)48.6(↑1.6)3.2
Yolov6-n-ihp6404.3710.679.4(↓0.8)57.4(↓1.4)3.7
Yolov7-t-ihp6406.7714.580.7(↑2.3)55.3(↑0.7)4.7
Yolov8-n-ihp6402.998.179.3(↓0.5)59.1(↓0.6)3.7
Part II: SNI (ours)
Yolov3-t-sni6408.7113.058.6(↑0.6)28.0(↑0.7)3.8
Yolov4-t-sni6405.9116.163.4(↑0.8)33.0(↑0.8)4.4
Yolov5-n-sni6401.794.274.5(↑0.5)47.7(↑0.7)3.4
Yolov6-n-sni6404.2811.081.9(↑1.7)60.8(↑2.0)3.4
Yolov7-t-sni6406.0613.381.5(↑3.1)57.9(↑3.3)5.0
Yolov8-n-sni6403.018.180.3(↑0.5)60.1(↑0.4)3.7
Part III: ESD-I (ours)
Yolov3-t-esd-I64011.822.469.2(↑11.2)38.5(↑11.2)4.6
Yolov4-t-esd-I6405.9116.163.1(↑0.7)32.9(↑0.7)4.4
Yolov5-n-esd-I6401.784.275.4(↑1.4)48.0(↑1.0)3.8
Yolov6-n-esd-I6404.3011.180.5(↑0.3)58.83.4
Yolov7-t-esd-I6406.8314.679.2(↑0.8)55.3(↑0.7)5.2
Yolov8-n-esd-I6403.228.180.1(↑0.3)59.8(↑0.1)3.8
ESD-II (ours)
Yolov3-t-esd-II6408.7814.863.0(↑5.0)32.1(↑4.8)4.9
Yolov4-t-esd-II6405.9316.663.9(↑1.5)33.5(↑1.3)4.7
Yolov5-n-esd-II6402.324.775.7(↑1.7)48.9(↑1.9)3.6
Yolov6-n-esd-II6404.3211.180.4(↑0.2)58.83.5
Yolov7-t-esd-II6407.6915.779.3(↑0.9)55.2(↑0.6)5.9
Yolov8-n-esd-II6403.188.580.4(↑0.6)59.9(↑0.2)4.0
", + "bbox": [ + 245, + 398, + 754, + 839 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "Rethinking Features-Fused-Pyramid-Neck for Object Detection", + "bbox": [ + 310, + 114, + 730, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 116, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/5aabb6058a6856f5d42f9149588f4e3c6a681ea131913fbced11930e33fee3b3.jpg", + "table_caption": [ + "Table 2: Ablation experiments of the GSCnv and GSCnvE-I/II on VOC $07 + 12$" + ], + "table_footnote": [], + "table_body": "
ModelsSizeParam.(M)FLOPs(G)AP50(%)AP(%)LatencyT4b=16(ms)
benchlines(GSConv [18])
Yolov3-tiny-gsconv6404.618.369.941.13.9
Yolov4-tiny-gsconv6403.4212.473.445.24.6
Yolov5-n-gsconv6401.513.974.247.93.5
Yolov6-n-gsconv6404.2711.080.459.13.4
Yolov7-tiny-gsconv6405.0011.478.455.35.6
Yolov8-n-gsconv6402.747.880.961.23.7
GSCConvE-I (ours)
Yolov3-t-gse-I6407.8611.671.7(↑1.8)43.8(↑2.7)3.9
Yolov4-t-gse-I6404.8814.374.9(↑1.5)47.8(↑2.6)4.7
Yolov5-n-gse-I6401.764.275.8(↑1.6)50.1(↑2.2)3.5
Yolov6-n-gse-I6404.3011.180.1(↓0.3)58.8(↓0.3)3.6
Yolov7-t-gse-I6406.2113.579.1(↑0.7)56.4(↑1.1)5.8
Yolov8-n-gse-I6402.968.181.2(↑0.3)61.3(↑0.1)3.8
GSCConvE-II (ours)
Yolov3-t-gse-II6402.856.370.1(↑0.2)41.4(↑0.3)4.0
Yolov4-t-gse-II6402.3610.875.6(↑2.2)48.0(↑2.8)4.6
Yolov5-n-gse-II6401.483.976.3(↑2.1)51.4(↑3.5)3.5
Yolov6-n-gse-II6404.2811.180.3(↓0.1)58.8(↓0.3)3.5
Yolov7-t-gse-II6404.3310.779.1(↑0.7)56.2(↑0.9)5.8
Yolov8-n-gse-II6402.677.779.9(↓1.0)60.4(↓0.8)3.8
", + "bbox": [ + 243, + 171, + 759, + 441 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/2b55466113426c6d12a07456d436913f5e296143e276df80e8e9f99e2deb3cf7.jpg", + "table_caption": [ + "Table 3: A contrast of the different confidence threshold for the SYOLO on COCO." + ], + "table_footnote": [], + "table_body": "
Confidence thresholdAP(%)APs(%)APm(%)APl(%)
0.001(valuation)53.135.558.769.9
0.25(application)48.7(↓4.4)29.7(↓5.8)54.5(↓4.2)66.5(↓3.4)
", + "bbox": [ + 279, + 491, + 723, + 537 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "4.3 Comparison experiments", + "text_level": 1, + "bbox": [ + 215, + 575, + 477, + 590 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "In Tab. 4 and Tab. 5, we compare state-of-the-art real-time detectors and SYOLO (architecture in Tab. 7) on the Pascal VOC and COCO benchmarks. In Tab. 6, we provide additional comparisons of state-of-the-art detectors on COCO [22], including both real-time (CNNs) and non-real-time (Transformers/CNNs) models. Parameters and FLOPs metrics significantly higher than those of real-time detector models are highlighted in red.", + "bbox": [ + 212, + 623, + 787, + 715 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "In Tab. 4, SNI increased the AP of YOLOv7 by 0.9 percentage points without any additional computational or memory costs. SYOLO, utilizing the SA solution, achieved the highest accuracy among competitive models on the Pascal VOC benchmark. In Tab. 5, SYOLO also achieved the highest accuracy among competitive models on the MS COCO benchmark at input resolutions of $416 \\times 416$ and $640 \\times 640$ . Positive performance improvements across different benchmarks demonstrate the effectiveness of the SA solution for real-time object detection.", + "bbox": [ + 212, + 719, + 787, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "Hulin Li", + "bbox": [ + 271, + 114, + 330, + 127 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/7dfda00e73cf544fc0a169bf41d2845ef33e75a8b9699a6062cdd3f30ca35e66.jpg", + "table_caption": [ + "Table 4: Comparasions of state-of-the-art real-time detectors on VOC $07 + 12$" + ], + "table_footnote": [ + "† Using simplified PANet: the red and green shortcuts are cancelled in the Fig. 1 (e).", + "$\\dagger \\dagger$ Using SNI replace the nearest neighbor interpolation to up-sampling in the PANet†." + ], + "table_body": "
ModelsYearNeckSizeParam.(M)FLOPs(G)AP(%)LatencyT4b=16(ms)
Yolov3-spp [31]2018FPN64061.60154.960.539.4
Yolov4 [1]2020PANet†64052.57119.366.231.0
Yolov5 [14]2022PANet†64046.21108.063.930.1
YoloX [7]2021PANet†64053.79153.766.740.2
Yolov6 [16]2022PANet†64065.81159.169.229.5
Yolov7 [35]2022PANet†64036.58103.568.223.2
Yolov8 [15]2023PANet†64043.62164.969.331.6
SYolov7(ours)2023SA-PANet††64036.58103.569.123.2
SYolo(ours)2023SA64057.23142.469.629.3
", + "bbox": [ + 218, + 171, + 782, + 308 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/86578c8ed8667c3d48202f73bd760fe016b2fc7203d5b6e3ef9123bca52839d4.jpg", + "table_caption": [ + "Table 5: Comparisons of state-of-the-art real-time detectors on COCO." + ], + "table_footnote": [], + "table_body": "
ModelsSizeParam.(M)FLOPs(G)AP(%)APs(%)APm(%)APl(%)LatencyT4b=16(ms)
YoloX-t [7]4165.16.532.8---4.3
SYolo-n(ours)4164.812.035.914.339.254.44.3
YoloX-t [7]6405.115.434.7---7.2
Yolov5-s [14]6407.216.537.4---9.8
Yolov6-n [16]6404.711.437.5---4.3
Yolov7-t [35]6406.213.838.718.842.451.95.3
Yolov8-n [15]6403.28.737.3---4.1
SYolo-s(ours)6406.615.740.821.145.356.95.8
", + "bbox": [ + 215, + 372, + 794, + 494 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "5 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 522, + 357, + 537 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "We explore the nature of representation learning and identify that the fusion of different level feature maps faces the issue of feature misalignment. To address this, we introduce the SNI. Additionally, we propose the ESD to mitigate spatial information loss during downsampling and further optimize lightweight convolutional methods for computational efficiency in lightweight models. Based on these advancements, we propose the SA solution for real-time detection, where all tested detectors surpass their original performance and achieve state-of-the-art results.", + "bbox": [ + 212, + 553, + 787, + 672 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Object detection has significantly benefited from the FPN-like paradigm in past few years. However, recent developments indicate that newer technologies, including innovative architectures and training strategies, may offer superior performance for modern detectors, making the FPN-like paradigm potentially limiting. In many cases, the effectiveness of the FPN-like paradigm can be enhanced through the application of SNI to address feature misalignment issues. Fully addressing the challenge of misalignment during feature fusion in existing vision models requires further research efforts.", + "bbox": [ + 212, + 675, + 787, + 796 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "Rethinking Features-Fused-Pyramid-Neck for Object Detection", + "bbox": [ + 310, + 114, + 730, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/a9d568e9452378ee9d23a3db464086fbed2ad80a06402c95cce39c100f7376eb.jpg", + "table_caption": [ + "Table 6: Comparisons of state-of-the-art detectors on COCO." + ], + "table_footnote": [], + "table_body": "
ModelsSizeParam.(M)FLOPs(G)AP(%)APs(%)APm(%)APl(%)LatencyV100b=1 (ms)
Transformers/CNNs (non-real-time, pre-trained)
FPN [21]-56.4145.836.817.538.747.8-
BiFPN [34]102421.055.049.3---42.8
ASFF [23]800--43.927.046.653.434.0
PANet [24]---51.032.653.964.8-
RFPN [28]---51.331.754.664.8-
DETR-Def. [40]-40.0173.043.826.447.158.0-
DETR-Dyn. [3]---47.228.649.359.1-
DETR-DND [17]133348.0265.048.6----
DINO [38]133347.0860.051.0----
Swin-T [26]-86.0745.050.5----
Swin-S [26]-107.0838.051.8----
Swin-B [26]1280145.0982.051.9----
CNNs (real-time, without pre-trained)
Yolov4 [1]60864.4142.843.526.746.753.3-
YoloX [7]64054.2155.649.7---14.5
PPYoloE [37]64052.2110.150.931.455.366.112.8
Yolov5 [14]64046.5109.149.0---10.1
Yolov6 [16]64058.5144.051.7----
Yolov7 [35]64036.9104.751.231.855.565.06.3
Yolov8 [15]64043.7165.252.9----
SYolo(ours)64057.2142.553.135.558.769.910.3
", + "bbox": [ + 215, + 172, + 785, + 465 + ], + "page_idx": 13 + }, + { + "type": "table", + "img_path": "images/eeecd0c17bdf4f3df420cc0845561b9a9af5f15ad65a7e52591dfefff30745c4.jpg", + "table_caption": [ + "Table 7: Architecture of the SYOLO. The E-ELAN/C2f is from the YOLOv7/8." + ], + "table_footnote": [ + "The first $3 \\times 3$ vanilla convolutional layer of the E-ELAN is replaced by the GSCConvE-I, and the end pointwise convolutional layer of the E-ELAN is replaced by the C2f.", + "$\\ddagger$ The 'l/s/n' means the model scale of the large/samll/nano." + ], + "table_body": "
StageBlock-1/s/n‡Channels-1/s/n‡K. sizeStride
Backbone
P1Conv64 /16 /1632
P2ESD-I/II128 /32 /323, 42
C2f ×3/2/1128 /32 /323, 11
P3ESD-I/II256 /64 /643, 42
C2f ×6/3/1256 /64 /643, 11
P4ESD-I/II512 /128 /1283, 42
C2f ×6/3/1512 /128 /1283, 11
P5ESD-I/II1024 /256 /2563, 42
C2f ×3/2/11024 /256 /2563, 11
SPP ×11024 /256 /256--
Neck(SA)
P5-P4SNI1536 /384 /284--
E-ELAN†×12048 /512 /5123, 11
P4-P3SNI512 /192 /192--
E-ELAN†×2/2/1512 /256 /2563, 11
P3-P4ESD-I/II256 /128 /1283, 42
E-ELAN†×2/2/11024 /512 /5123, 11
P4-P5ESD-I/II512 /256 /2563, 42
E-ELAN†×2/2/12048 /1024 /10243, 11
", + "bbox": [ + 316, + 502, + 686, + 792 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "Hulin Li", + "bbox": [ + 271, + 114, + 330, + 127 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 217, + 143, + 401, + 162 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "This work is supported by National Natural Science Foundation of China (Grant No. 52172381), Graduate Research Innovation Foundation of Chongqing Jiaotong University (Grant No. CYB240259). The author would also like to thank Prof. Dr. Jun Li, Prof. Dr. Qiliang Ren, and Ms. Xinxin Liu (Intelligent Transportation Big Data Center, School of Information Science and Engineering, CQJTU) for providing GPUs.", + "bbox": [ + 212, + 174, + 787, + 267 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 287, + 321, + 304 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Bochkovskiy, A., Wang, C.Y., Liao, H.Y.M.: Yolov4: Optimal speed and accuracy of object detection. arXiv preprint arXiv:2004.10934 (2020)", + "2. Chollet, F.: Xception: Deep learning with depthwise separable convolutions. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1251-1258 (2017)", + "3. Dai, X., Chen, Y., Yang, J., Zhang, P., Yuan, L., Zhang, L.: Dynamic detr: End-to-end object detection with dynamic attention. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 2988-2997 (2021)", + "4. Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: Imagenet: A large-scale hierarchical image database. In: 2009 IEEE conference on computer vision and pattern recognition. pp. 248-255. IEEE (2009)", + "5. Everingham, M., Eslami, S.A., Van Gool, L., Williams, C.K., Winn, J., Zisserman, A.: The Pascal visual object classes challenge: A retrospective. International journal of computer vision 111, 98-136 (2015)", + "6. Fu, C.Y., Liu, W., Ranga, A., Tyagi, A., Berg, A.C.: Dssd: Deconvolutional single shot detector. arXiv preprint arXiv:1701.06659 (2017)", + "7. Ge, Z., Liu, S., Wang, F., Li, Z., Sun, J.: Yolox: Exceeding yolo series in 2021. arXiv preprint arXiv:2107.08430 (2021)", + "8. Girshick, R.: Fast r-cnn. In: Proceedings of the IEEE international conference on computer vision. pp. 1440-1448 (2015)", + "9. Girshick, R., Donahue, J., Darrell, T., Malik, J.: Rich feature hierarchies for accurate object detection and semantic segmentation. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 580-587 (2014)", + "10. Han, K., Wang, Y., Tian, Q., Guo, J., Xu, C., Xu, C.: Ghostnet: More features from cheap operations. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 1580-1589 (2020)", + "1. He, K., Zhang, X., Ren, S., Sun, J.: Spatial pyramid pooling in deep convolutional networks for visual recognition. IEEE transactions on pattern analysis and machine intelligence 37(9), 1904-1916 (2015)", + "2. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 770-778 (2016)", + "3. Howard, A.G., Zhu, M., Chen, B., Kalenichenko, D., Wang, W., Weyand, T., Andreetto, M., Adam, H.: Mobilenets: Efficient convolutional neural networks for mobile vision applications. arXiv preprint arXiv:1704.04861 (2017)", + "4. Jocher, G., Chaurasia, A., Stoken, A., Borovec, J., Kwon, Y., Fang, J., Michael, K., Montes, D., Nadar, J., Skalski, P., et al.: ultralytics/yolov5: v6. 1-tensorrt, tensorflow edge tpu and openvino export and inference. Zenodo (2022)" + ], + "bbox": [ + 225, + 316, + 784, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "Rethinking Features-Fused-Pyramid-Neck for Object Detection", + "bbox": [ + 310, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "15. Jocher, G., et al.: Yolo-ultralytics. https://github.com/ultralytics/ultralytics (2023)", + "16. Li, C., Li, L., Jiang, H., Weng, K., Geng, Y., Li, L., Ke, Z., Li, Q., Cheng, M., Nie, W., et al.: Yolov6: A single-stage object detection framework for industrial applications. arXiv preprint arXiv:2209.02976 (2022)", + "17. Li, F., Zhang, H., Liu, S., Guo, J., Ni, L.M., Zhang, L.: Dn-detr: Accelerate detr training by introducing query denoising. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 13619-13627 (2022)", + "18. Li, H., Li, J., Wei, H., Liu, Z., Zhan, Z., Ren, Q.: Slim-neck by gsconv: a lightweight-design for real-time detector architectures. Journal of Real-Time Image Processing 21(3), 62 (2024)", + "19. Li, Y., Chen, Y., Wang, N., Zhang, Z.: Scale-aware trident networks for object detection. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 6054-6063 (2019)", + "20. Li, Z., Yang, L., Zhou, F.: Fssd: feature fusion single shot multibox detector. arXiv preprint arXiv:1712.00960 (2017)", + "21. Lin, T.Y., Dollar, P., Girshick, R., He, K., Hariharan, B., Belongie, S.: Feature pyramid networks for object detection. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 2117-2125 (2017)", + "22. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dólar, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13. pp. 740-755. Springer (2014)", + "23. Liu, S., Huang, D., Wang, Y.: Learning spatial fusion for single-shot object detection. arxiv 2019. arXiv preprint arXiv:1911.09516 (1911)", + "24. Liu, S., Qi, L., Qin, H., Shi, J., Jia, J.: Path aggregation network for instance segmentation. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 8759-8768 (2018)", + "25. Liu, W., Anguelov, D., Erhan, D., Szegedy, C., Reed, S., Fu, C.Y., Berg, A.C.: Ssd: Single shot multibox detector. In: Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part I 14. pp. 21-37. Springer (2016)", + "26. Liu, Z., Lin, Y., Cao, Y., Hu, H., Wei, Y., Zhang, Z., Lin, S., Guo, B.: Swin transformer: Hierarchical vision transformer using shifted windows. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 10012-10022 (2021)", + "27. Liu, Z., Mao, H., Wu, C.Y., Feichtenhofer, C., Darrell, T., Xie, S.: A convnet for the 2020s. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 11976-11986 (2022)", + "28. Qiao, S., Chen, L.C., Yuille, A.: Detectors: Detecting objects with recursive feature pyramid and switchable atrous convolution. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10213-10224 (2021)", + "29. Redmon, J., Divvala, S., Girshick, R., Farhadi, A.: You only look once: Unified, real-time object detection. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 779-788 (2016)", + "30. Redmon, J., Farhadi, A.: Yolo9000: better, faster, stronger. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 7263-7271 (2017)", + "31. Redmon, J., Farhadi, A.: Yolov3: An incremental improvement. arXiv preprint arXiv:1804.02767 (2018)", + "32. Simonyan, K., Zisserman, A.: Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556 (2014)" + ], + "bbox": [ + 215, + 147, + 784, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "Hulin Li", + "bbox": [ + 271, + 114, + 330, + 127 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "33. Singh, B., Davis, L.S.: An analysis of scale invariance in object detection snip. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 3578-3587 (2018)", + "34. Tan, M., Pang, R., Le, Q.V.: Efficientdet: Scalable and efficient object detection. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10781-10790 (2020)", + "35. Wang, C.Y., Bochkovskiy, A., Liao, H.Y.M.: Yolov7: Trainable bag-of-freebies sets new state-of-the-art for real-time object detectors. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 7464-7475 (2023)", + "36. Wu, Y., Chen, Y., Yuan, L., Liu, Z., Wang, L., Li, H., Fu, Y.: Rethinking classification and localization for object detection. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10186-10195 (2020)", + "37. Yang, Z., Zhou, Y., Chen, Z., Ngiam, J.: 3d-man: 3d multi-frame attention network for object detection. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 1863-1872 (2021)", + "38. Zhang, H., Li, F., Liu, S., Zhang, L., Su, H., Zhu, J., Ni, L.M., Shum, H.Y.: Dino: Detr with improved denoising anchor boxes for end-to-end object detection. arXiv preprint arXiv:2203.03605 (2022)", + "39. Zhang, X., Zhou, X., Lin, M., Sun, J.: Shufflenet: An extremely efficient convolutional neural network for mobile devices. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 6848-6856 (2018)", + "40. Zhu, X., Su, W., Lu, L., Li, B., Wang, X., Dai, J.: Deformable detr: Deformable transformers for end-to-end object detection. arXiv preprint arXiv:2010.04159 (2020)" + ], + "bbox": [ + 212, + 146, + 787, + 494 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "Rethinking Features-Fused-Pyramid-Neck for Object Detection", + "bbox": [ + 310, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 16 + } +] \ No newline at end of file diff --git a/2024/Rethinking Features-Fused-Pyramid-Neck for Object Detection/8cefd2b8-bab0-4ef4-a711-578971c2749a_model.json b/2024/Rethinking Features-Fused-Pyramid-Neck for Object Detection/8cefd2b8-bab0-4ef4-a711-578971c2749a_model.json new file mode 100644 index 0000000000000000000000000000000000000000..cf6f8e5249f9923f3ad9eb3b3fa9bb10b0a41272 --- /dev/null +++ b/2024/Rethinking Features-Fused-Pyramid-Neck for Object Detection/8cefd2b8-bab0-4ef4-a711-578971c2749a_model.json @@ -0,0 +1,1928 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.245, + 0.14, + 0.761, + 0.187 + ], + "angle": 0, + "content": "Rethinking Features-Fused-Pyramid-Neck for Object Detection" + }, + { + "type": "text", + "bbox": [ + 0.465, + 0.213, + 0.54, + 0.227 + ], + "angle": 0, + "content": "Hulin Li" + }, + { + "type": "text", + "bbox": [ + 0.231, + 0.24, + 0.773, + 0.255 + ], + "angle": 0, + "content": "College of Traffic and Transportation, Chongqing Jiaotong University, Chongqing" + }, + { + "type": "text", + "bbox": [ + 0.453, + 0.256, + 0.551, + 0.268 + ], + "angle": 0, + "content": "400074, China" + }, + { + "type": "text", + "bbox": [ + 0.412, + 0.27, + 0.593, + 0.283 + ], + "angle": 0, + "content": "alan@mails.cqjtu.edu.cn" + }, + { + "type": "text", + "bbox": [ + 0.263, + 0.321, + 0.761, + 0.543 + ], + "angle": 0, + "content": "Abstract. Multi-head detectors typically employ a features-fused-pyramid-neck for multi-scale detection and are widely adopted in the industry. However, this approach faces feature misalignment when representations from different hierarchical levels of the feature pyramid are forcibly fused point-to-point. To address this issue, we designed an independent hierarchy pyramid (IHP) architecture to evaluate the effectiveness of the features-unfused-pyramid-neck for multi-head detectors. Subsequently, we introduced soft nearest neighbor interpolation (SNI) with a weight-downscaling factor to mitigate the impact of feature fusion at different hierarchies while preserving key textures. Furthermore, we present a feature adaptive selection method for downsampling in extended spatial windows (ESD) to retain spatial features and enhance lightweight convolutional techniques (GSCnvE). These advancements culminate in our secondary features alignment solution (SA) for real-time detection, achieving state-of-the-art results on Pascal VOC and MS COCO. Code will be released at https://github.com/AlanLi1997/rethinking-fpn." + }, + { + "type": "text", + "bbox": [ + 0.263, + 0.557, + 0.741, + 0.571 + ], + "angle": 0, + "content": "Keywords: Object detection \\(\\cdot\\) Feature pyramid \\(\\cdot\\) Feature misalignment" + }, + { + "type": "text", + "bbox": [ + 0.265, + 0.572, + 0.45, + 0.583 + ], + "angle": 0, + "content": "- Detection-net architecture" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.612, + 0.377, + 0.627 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.644, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Bounding box regression for object detection and pixel-level boundary regression for instance segmentation are pivotal and challenging visual tasks. Unlike basic image classification [4], detection or segmentation tasks often involve scenes with multiple objects of varying scales. Early deep learning-based object detection models began exploring real-time multi-scale object detection. RCNN [9] employed selective search to extract a large number of region proposals (\\(\\sim 2\\mathrm{k}\\)). Although the vast number of image patches (region proposals) might aid multiscale detection, in practice, most scenes do not contain such a high number of objects to recognize. This approach was later deemed inefficient [8], as the proposed regions were derived from the image rather than features, necessitating a convolutional neural network (CNN) to extract features and predict classes for each proposal. A simplified RCNN architecture, as depicted in Fig. 1 (a), converts images of different scales into feature maps for object detection, largely" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.331, + 0.128 + ], + "angle": 0, + "content": "Hulin Li" + }, + { + "type": "image", + "bbox": [ + 0.277, + 0.146, + 0.482, + 0.263 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.49, + 0.146, + 0.596, + 0.262 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.609, + 0.147, + 0.721, + 0.262 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.277, + 0.264, + 0.563, + 0.376 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.572, + 0.263, + 0.727, + 0.376 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.277, + 0.377, + 0.576, + 0.491 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.589, + 0.377, + 0.721, + 0.491 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.516, + 0.788, + 0.627 + ], + "angle": 0, + "content": "Fig. 1: The IHP and seven other typical neck architectures. The three heat maps are come from the third (P3), fourth (P4), and fifth (P5) hierarchies of the FPN-neck model, respectively. The area of interest for each hierarchies' features is highlighted in red; the darker the color, the higher the feature weight. It is evident that lower-level hierarchies prioritize local features, while higher-level hierarchies focus more on global features. In other words, there is a representation bias at different hierarchical levels. If these features are combined by point-to-point fusion without first addressing the misalignment issue, it may introduce noise rather than enrich semantics." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.659, + 0.788, + 0.841 + ], + "angle": 0, + "content": "neglecting the real-time detection capability of models. YOLO [29] introduced the first real-time deep learning-based object detection model, which did not specifically focus on multi-scale detection, with a simplified structure shown in Fig. 1 (b). SSD [25] effectively combined the strengths of RCNN and YOLO, emphasizing the concept of multi-scale detection. It was the first model to achieve real-time multi-scale object detection using multi-scale features, with an architecture similar to Fig. 1 (c). SSD efficiently utilized feature maps of different scales from the VGG-16 backbone [32] to predict objects of varying scales. The exploration by early deep learning-based object detection models provided an important insight to the field: the efficient utilization of multi-scale features is crucial for enhancing the overall performance (accuracy-speed trade-off) of detectors. Consequently, FPN [21] introduced the feature pyramid-based multi-scale" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.312, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Rethinking Features-Fused-Pyramid-Neck for Object Detection" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.147, + 0.786, + 0.493 + ], + "angle": 0, + "content": "feature fusion concept, which gradually became a structural paradigm for detectors. This led to numerous representative works enhancing multi-scale feature fusion, such as PANet [24], ASFF [23], BiFPN [34], and RFPN [28]. The performance of real-time detectors has improved with the implementation of feature fusion strategies. However, further advancements are hindered by the increasing complexity of these techniques. Before the introduction of FPN, object detectors primarily relied on structures without fusion [9,25,29], as depicted in Fig. 1 (a)-(c). With the advent of FPN, these detectors have shifted towards FPN-based paradigms or more intricate improvement methods [8,20,29-31], as illustrated in Fig. 1 (e)-(g). It is important to note that the FPN-like strategy should not be considered an obligatory component for all detectors. Representation learning suggests that models ultimately require refined representations rather than an abundance of complex representations. Under the FPN-like paradigm, feature maps of different hierarchies and different resolutions are combined at the element level (point-to-point), necessitating the expansion of low-resolution, high-level feature maps to match the resolution of low-level feature maps. However, there are inherent representation biases in feature maps at different hierarchical levels: low-level feature maps tend to retain representations of small objects (local features in small receptive fields), while high-level feature maps favor representations of large objects (global features in large receptive fields), as shown in Fig. 1 (i). The direct element-level fusion of these feature maps with representation bias may lead to partial destruction of representations and exacerbate the feature misalignment problem during fusion." + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.5, + 0.786, + 0.77 + ], + "angle": 0, + "content": "In visual models, the resolution of feature maps decreases progressively from low to high hierarchical levels of the feature pyramid, achieved through step-by-step downsampling. Typically, downsampling reduces the resolution (width \\(\\times\\) height) of feature maps at a square root rate, inherently filtering out some spatial information regardless of whether the number of features (channels) remains constant or increases. It is also noteworthy that the transition of feature information from spatial to channel dimensions primarily occurs during downsampling, underscoring the importance of this stage. Traditional downsampling methods operate in a single mode, such as convolution with a stride of 2 (learnable) or average pooling and max pooling with a stride of 2 (non-learnable). These single-mode approaches have persisted since the introduction of ResNet [12] without significant evolution, which merits further investigation. Additionally, the initial purpose of FPN was not only to demonstrate the advantages of feature fusion in addressing multi-scale detection challenges but also to appropriately reduce model complexity. In industries with limited hardware resources, complex models are often impractical [2] [10]. Therefore, we enhance the lightweight convolution technique of the GSCov [18] to achieve a preferable trade-off between accuracy and speed for real-time detection models." + }, + { + "type": "text", + "bbox": [ + 0.241, + 0.776, + 0.722, + 0.79 + ], + "angle": 0, + "content": "The main contributions of this work can be summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.795, + 0.785, + 0.84 + ], + "angle": 0, + "content": "1). We rethink the effectiveness of FPN-like paradigms for modern real-time detectors based on representation learning and identify issues with feature misalignment during element-level fusion." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.331, + 0.128 + ], + "angle": 0, + "content": "Hulin Li" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.19 + ], + "angle": 0, + "content": "2). We design an independent hierarchy pyramid architecture (IHP) without feature fusion in the neck to validate our findings. And the IHP achieves advanced results." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.193, + 0.785, + 0.223 + ], + "angle": 0, + "content": "3). We introduce soft nearest neighbor interpolation (SNI) for up-sampling to mitigate feature misalignment during fusion." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.225, + 0.784, + 0.284 + ], + "angle": 0, + "content": "4). We provide a feature adaptive selection method in extended spatial windows for downsampling (ESD) to enhance spatial feature retention during the downsampling stage. This method is plug-and-play and optimizes performance at a low cost." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.287, + 0.784, + 0.33 + ], + "angle": 0, + "content": "5). We simplify and enhance the GSCnv lightweight convolution (GSCnvE) to offer more cost-effective options for models operating on edge devices with resource constraints." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.333, + 0.784, + 0.377 + ], + "angle": 0, + "content": "6).We validate our approaches on Pascal VOC and COCO and present the SA solution. Existing lightweight real-time detection models can be optimized with the SA solution without bells and whistles." + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.785, + 0.377 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.408, + 0.388, + 0.423 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.446, + 0.597, + 0.46 + ], + "angle": 0, + "content": "2.1 Multi-head detection and features fusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.479, + 0.788, + 0.844 + ], + "angle": 0, + "content": "The first-generation general detection models using deep learning consist of two main components: the backbone and the detecting-head [9, 29]. These models typically use final feature maps for prediction. SSD [25] introduced the use of multi-level feature maps for object detection, marking the beginning of multi-head detection and influencing subsequent real-time detector designs [31]. In the SSD model, each detecting-head directly uses raw feature maps from different levels of the backbone without any fusion. The key advantage of the multi-head detection method is the ability to predict objects of different scales using feature maps with varying receptive fields. High-level feature maps, with their wide global receptive fields, are beneficial for identifying large objects, while low-level feature maps, with refined local receptive fields, are helpful for identifying small objects. FPN introduced a fusion scheme for multi-level feature maps to enhance detection accuracy, setting the trend for feature fusion and becoming a common practice. This led to the development of Backbone-Neck-Head architectures. More complex variants based on FPN, such as BiFPN [34] and PANet [24], have since been proposed, enabling the use of more features for prediction. However, as prediction accuracy improved, networks became increasingly complex due to additional fusion layers. Some approaches have attempted to bypass feature fusion schemes by exploring new perspectives on multi-scale detection challenges, such as optimizing multi-scale training strategies [33] or using dilated convolutions to capture multi-scale receptive fields [19]. But these methods have gradually lost their competitiveness with the continuous optimization of classification-backbone feature extraction capabilities and the ongoing development of feature fusion techniques [1]." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.312, + 0.115, + 0.732, + 0.131 + ], + "angle": 0, + "content": "Rethinking Features-Fused-Pyramid-Neck for Object Detection" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.147, + 0.36, + 0.162 + ], + "angle": 0, + "content": "2.2 Lightweight" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.174, + 0.788, + 0.34 + ], + "angle": 0, + "content": "The choice between cloud computing or edge computing for a model depends on its number of parameters and floating-point operations (FLOPs). Therefore, reducing the number of parameters or FLOPs is a primary focus in lightweight model studies. Direct lightweight approaches include reducing the depth (number of layers) or width (number of neurons/filters) of the model, as seen in models likeYOLO-fast/tiny/nano [1,14-16,31,35]. Additionally, sparse computing techniques, such as depth-wise separable convolution used in MobileNets [13] and ShuffleNet [39], can also be employed. However, low-depth networks often suffer from underfitting due to insufficient nonlinear representation capabilities. Therefore, reducing the number of network layers may not always be a cost-effective way to lighten the model." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.364, + 0.635, + 0.383 + ], + "angle": 0, + "content": "3 Secondary Features Alignment Solution" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.397, + 0.788, + 0.489 + ], + "angle": 0, + "content": "Key contents of this work are described in detail in this section. Specifically, there are the IHP structure to directly demonstrate features unalignment problem by abandoning fusion, the SNI to alleviate unaligned during features fusion by point-to-point, the ESD to enhance spatial-features capture during the downsampling stage, and the GSConvE to improve the performance between the accuracy and speed of lightweight models." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.512, + 0.629, + 0.528 + ], + "angle": 0, + "content": "3.1 Independent hierarchy pyramid architecture" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.538, + 0.788, + 0.794 + ], + "angle": 0, + "content": "The introduction of FPN brought all real-time detectors into the realm of feature fusion. However, we have identified a significant issue arising from the point-to-point fusion of different level features - partial local features become unaligned. This fusion process is akin to adding noise when unaligned features are combined, as features unrelated to the target space are forcibly integrated, resulting in spatial disarray. For instance, as depicted in Fig. 2, adding the head-feature of a puppy to the butt creates an undefined breed when fed to the detection-head. This scenario is particularly evident when the puppy occupies fewer pixels. Typically, nearest neighbor interpolation is used to rapidly increase the resolution of high-level feature maps to match that of low-level feature maps, followed by fusion through methods such as element-wise addition, channel concatenation, or weighted sum. However, downsampling tends to be nonlinear and irreversible, leading to inconsistencies in spatial features among the additional features generated by up-sampling. Although BiFPN [34] offers a learnable fusion method, it has not been widely adopted by recent state-of-the-art real-time detectors, such as YOLOv7 [35] or YOLOv8 [15], due to the increased complexity and lack of demonstrated comprehensive performance improvements in these models." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.796, + 0.787, + 0.84 + ], + "angle": 0, + "content": "The IHP adopts a radical approach by abandoning all fusions in the neck, fundamentally circumventing the features misalignment problem and streamlining the structure. By leveraging the inherent benefits of the multi-head detection" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.331, + 0.128 + ], + "angle": 0, + "content": "Hulin Li" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.146, + 0.788, + 0.617 + ], + "angle": 0, + "content": "mode, the IHP directly predicts objects of different sizes using feature maps of different levels. It is important to note that there are significant differences between IHP and the SSD architecture: SSD directly uses different level features from the backbone to detect objects but the IHP introduces a bottleneck convolution module before prediction to filter the features. The advantage of this approach is that it allows the model to independently learn features for a specific scale branch without affecting features at other levels. Directly stacking convolutional blocks at different levels of the backbone can lead to interference between different levels, which we believe might be the primary reason why using a very deep backbone in object detection models can result in performance degradation. For instance, DSSD [6] showed a performance drop when ResNet-101 was used as the backbone instead of VGG-16 in the SSD detector. In Part I of Tab. 1, all baselines of coupled-head YOLOs demonstrated improved results when using the IHP. Features fusion often overlooks the potential introduction of noise when augmenting semantic richness for prediction. While the IHP may appear to reduce semantic information, it effectively harnesses the advantages of multi-head detection to circumvent the misalignment problem. This strategic utilization of multi-head detection is the key factor behind the competitive results achieved by the IHP. Unfortunately, unlike localization-classification-coupled-head detectors [14,29-31,35], the IHP did not achieve the same positive reaction to localization-classification-decoupled-head detectors [15, 16]. Decoupled-head detectors address a key challenge faced by coupled-head detectors: the disparity in the interests of classification and localization tasks, both of which are often performed using the same layer for prediction [36]. This approach is akin to forcibly fusing misaligned features for prediction. Decoupled-head detectors overcome this limitation by independently predicting classification and localization tasks through separate branches. This enables them to learn beneficial representations from redundant features, which is not possible in coupled-head detectors. Consequently, while FPN strategies are still applicable to some extent in decoupled-head detectors, the features misalignment problem still needs to be addressed." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.636, + 0.553, + 0.652 + ], + "angle": 0, + "content": "3.2 Soft nearest neighbor interpolation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.659, + 0.788, + 0.841 + ], + "angle": 0, + "content": "In detection or segmentation models, different-level feature fusion typically occurs after the up-sampling of feature maps, with nearest neighbor interpolation and transposed convolution being popular up-sampling methods. However, the resolution expansion resulting from up-sampling in detection or segmentation differs significantly from tasks such as generation or super-resolution. In detection, the expanded feature maps represent abstract high-level semantic features rather than raw image detail information. Transposed convolution introduces increased computation and latency. To address the features misalignment problem with minimal cost while maintaining the speed of nearest neighbor interpolation, we explore a solution. Nearest neighbor interpolation is akin to average unpooling but is considered a \"hard\" operation. By softening this operation, similar to the transformation of SoftMax to the Max function, we aim to mitigate the" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.312, + 0.115, + 0.733, + 0.13 + ], + "angle": 0, + "content": "Rethinking Features-Fused-Pyramid-Neck for Object Detection" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "image", + "bbox": [ + 0.221, + 0.148, + 0.782, + 0.313 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.335, + 0.785, + 0.363 + ], + "angle": 0, + "content": "Fig. 2: Illustrations of the features misalignment in fusion and the SNI. These numbers, 22-88, are just markers of different local-features not real feature values." + }, + { + "type": "image", + "bbox": [ + 0.28, + 0.38, + 0.727, + 0.56 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.571, + 0.788, + 0.613 + ], + "angle": 0, + "content": "Fig.3: A comparison of the SNI and traditional method. This result could be reproduced by using SNI source code (model:Yolov5n-panet). The same size scaling before and after up-sampling is for intuitive comparison." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.642, + 0.785, + 0.673 + ], + "angle": 0, + "content": "\"hard\" features misalignment issue. To achieve this, we introduce a soft factor, denoted as \\(\\alpha\\), for the nearest neighbor interpolation during up-sampling:" + }, + { + "type": "equation", + "bbox": [ + 0.382, + 0.681, + 0.787, + 0.711 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {Y} = \\alpha \\cdot \\boldsymbol {f} (\\boldsymbol {X}), \\alpha = \\frac {\\text {R e s o l u t i o n} \\boldsymbol {X}}{\\text {R e s o l u t i o n} \\boldsymbol {Y}} \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.719, + 0.788, + 0.841 + ], + "angle": 0, + "content": "In Eq. (1), the resolution of the high-level feature maps \\((\\mathbf{X})\\) and low-level feature maps \\((\\mathbf{Y})\\) are denoted as \\(Resolution_{X}\\) and \\(Resolution_{Y}\\), respectively. The nearest neighbor interpolation operation, represented by \\(\\pmb{f}\\), is illustrated in Fig. 2. The SNI adjusts the influence of high-level semantic features on low-level features based on the zoom factor for feature maps. Specifically, as the zoom factor increases, the impact of high-level semantic features on low-level features weakens. Unlike SoftMax, the SNI does not convert outputs into probabilities but rather mitigates misalignment without additional cost when integrating high-" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.331, + 0.128 + ], + "angle": 0, + "content": "Hulin Li" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.358 + ], + "angle": 0, + "content": "level features with low-level features through point-to-point fusion. SNI softens feature weights while preserving key textures, thereby reducing the mutual influence of features from different hierarchical levels during fusion, as qualitatively demonstrated in Fig. 3. This approach achieves the secondary (auxiliary) features alignment (SA) and lowers the complexity for the model to learn from the fused representations. In contrast, traditional methods that directly merge features from different hierarchical levels can cause unstable competitive effects due to representation biases, increasing the learning difficulty for the model. For example, if the three heatmaps in Fig. 1 (i) are directly fused, should the red areas cover the entire image? Clearly, the answer is no. If such a situation occurs, it indicates a failure in prior feature extraction, as the objective of a model is to capture and focus on useful features. This is also why models under the FPN-like paradigm typically require several additional convolutional blocks to further filter features after fusing and before predicting." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.359, + 0.788, + 0.48 + ], + "angle": 0, + "content": "In Part II of Tab. 1, all tested state-of-the-art models demonstrate accuracy improvements solely by incorporating the SNI, confirming the presence of features misalignment in FPN-based models. Notably, for YOLOv6 [16], a detector equipped with various techniques tailored for industry applications, the AP is enhanced by up to 2 percentage points just with the adoption of SNI, while reducing computational cost and inference latency. We achieved significant optimization results in the original YOLOv6 by simply replacing transposed convolution with SNI, highlighting the efficacy of this approach." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.501, + 0.715, + 0.516 + ], + "angle": 0, + "content": "3.3 Feature adaptive selection in extended spatial windows" + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.553, + 0.782, + 0.653 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.682, + 0.788, + 0.765 + ], + "angle": 0, + "content": "Fig. 4: Illustrations of the ESD-I and ESD-II. It must be mentioned that although the size of the extended window does not have to be same in the ESD-I, pooling is a non-learnable operation, so the features from an unequal window may become noise for other and bring the trouble of the features misalignment again. The ESD-II is free on size of the extended window because of the learnable fusion like the SPPNet [11] for the YOLOv3." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Feature downsampling is a critical technique in vision models based on deep learning, involving the progressive reduction of feature map resolution. In vision representation learning, models typically rely on low-resolution, high-level" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.312, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Rethinking Features-Fused-Pyramid-Neck for Object Detection" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.51 + ], + "angle": 0, + "content": "feature maps for prediction rather than directly processing high-resolution raw images. However, excessive downsampling can result in significant loss of spatial details [12]. Recent detectors typically limit the number of downsampling operations to no more than five times, leading to a spatial resolution reduction of \\(2^{5}\\) times, as demonstrated in YOLOs [1,7,14-16,31,35,37] and Swin Transformer [26]. Common downsampling methods include convolution and pooling, with models typically employing only one of these methods. We propose the feature adaptive selection method in extended spatial windows for downsampling (ESD) to enhance spatial information retention capabilities. ESD consists of two nonlinear branches and one linear branch. In the nonlinear branches, a norm convolutional layer and an extended window max-pooling layer enhance local feature capturing, while extended window average-pooling enhances global feature capturing. Subsequently, linear features are merged with nonlinear features. Two fusion modes, ESD-I and ESD-II, are designed for lightweight and norm models, respectively. The difference lies in the feature fusion stage, as illustrated in Fig. 4. ESD-I employs element-wise addition for merging features, offering low computational cost suitable for lightweight models, while ESD-II utilizes learnable linearly adaptive fusion, slightly increasing computational cost but suitable for norm models. Importantly, local and global features of the extended window are sampled using simple pooling techniques, preserving input information to a significant extent. This reduces information loss during downsampling and enables subsequent layers to learn partial representations from the previous layer, akin to indirectly realizing a hidden shortcut connection reminiscent of ResNet [12]." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.531, + 0.547, + 0.546 + ], + "angle": 0, + "content": "3.4 Lightweight GSConv enhancement" + }, + { + "type": "image", + "bbox": [ + 0.251, + 0.583, + 0.754, + 0.695 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.346, + 0.721, + 0.655, + 0.735 + ], + "angle": 0, + "content": "Fig. 5: Illustrations of the GSConvE-I and II." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.788, + 0.841 + ], + "angle": 0, + "content": "GSCnv [18] was introduced to address the channel interaction limitations of depthwise separable convolution by blending features from both vanilla convolution and depthwise convolution. The original GSCnv consists of three fundamental components: a \\(3 \\times 3\\) vanilla convolution, a \\(5 \\times 5\\) depthwise convolution, and a feature shuffle operation. Building upon this, we propose two evolutions" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.331, + 0.128 + ], + "angle": 0, + "content": "Hulin Li" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.375 + ], + "angle": 0, + "content": "of GSConv, namely GSConvE-I and GSConvE-II, tailored for norm models and lightweight models, respectively. GSConvE-I is designed for norm models, where intermediate feature maps on the auxiliary branch are derived from dense linear mappings, while output feature maps originate from sparse linear mappings. This approach maintains feature richness without significantly increasing computational cost. On the other hand, GSConvE-II is geared towards lightweight models. It incorporates three depthwise convolution auxiliary branches with kernel sizes of \\(9 \\times 9\\), \\(13 \\times 13\\), and \\(17 \\times 17\\), respectively. By utilizing larger kernel sizes, this variant can directly capture larger receptive fields and global features with fewer layer accumulations, offering high cost-effectiveness for lightweight models that prioritize reduced FLOPs and layer count. In GSConvE-I, we explore the removal of the batch normalization layer on the depthwise convolution branch while retaining the batch normalization layer on the main branch, similar to ConvNeXt [27]. This adjustment avoids the vanishing gradient problem and simplifies computation, leading to improved prediction accuracy." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.399, + 0.377, + 0.417 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.433, + 0.788, + 0.511 + ], + "angle": 0, + "content": "Real-time detectors are highly favored in industry due to their hardware-friendly nature and the optimal balance they strike between accuracy and speed. In our study, we selected popular YOLO models—YOLOv3 [31], YOLOv4 [1], YOLOv5 [14], YOLOv6 [16], YOLOv7 [35], and YOLOv8 [15]—as baselines to evaluate the efficacy of the SA solution in real-time detection scenarios (SYOLO)." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.534, + 0.336, + 0.549 + ], + "angle": 0, + "content": "4.1 Datasets" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.561, + 0.788, + 0.698 + ], + "angle": 0, + "content": "We conducted trials on the Pascal VOC dataset [5] and the MS COCO dataset [22]. The Pascal VOC 07+12 dataset was utilized for ablation experiments, wherein models underwent training on the train&val07+12 set and were subsequently evaluated on the test07 set. On the other hand, the MS COCO dataset was employed for comparative experiments, with models trained on the train2017 set and evaluated on the val2017 set. To ensure fairness and consistency, all released models were trained using the default set of hyperparameters and without utilizing pre-trained weights. More details of hyperparameters are in the support material." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.722, + 0.4, + 0.737 + ], + "angle": 0, + "content": "4.2 Ablation studies" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.843 + ], + "angle": 0, + "content": "In Tab. 1, we present the results of ablation experiments conducted on six baselines [1, 14-16, 31, 35], focusing on the impact of the IHP, SNI, ESD. The lightweight experiment results of GSConvE are reported in Tab. 2. All baselines retained their original neck architecture, as we believe the competitive neck has been carefully chosen in the original works. Performance metrics highlighted in green indicate performance gains, while those in red indicate losses. It is crucial" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.312, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Rethinking Features-Fused-Pyramid-Neck for Object Detection" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.328 + ], + "angle": 0, + "content": "to address an often-overlooked aspect during the verification process: the inconsistency between reported evaluation accuracy and actual usage. Frequently, detectors are evaluated with a very low confidence threshold, such as 0.001, which results in a large number of prediction boxes and inflated evaluation accuracy. However, in real-world detection scenarios, the confidence threshold is typically set much higher, around 0.25 or above. This filtering process ensures that higher-quality bounding boxes are utilized for practical purposes. The discrepancy between evaluation settings and real-world usage can create a misleading impression of accuracy. Therefore, it is essential to maintain consistency in configuration between evaluation and application. While this may result in a significant drop in evaluation accuracy, as shown in Tab. 3, it provides a more realistic assessment of model performance." + }, + { + "type": "table_caption", + "bbox": [ + 0.239, + 0.374, + 0.763, + 0.388 + ], + "angle": 0, + "content": "Table 1: Ablation experiments of the IHP, SNI and ESD-I/II on VOC \\(07 + 12\\)" + }, + { + "type": "table", + "bbox": [ + 0.246, + 0.4, + 0.756, + 0.84 + ], + "angle": 0, + "content": "
ModelsSizeParam.(M)FLOPs(G)AP50(%)AP(%)LatencyT4b=16(ms)
basielines
Yolov3-t [31]6408.7113.058.027.33.8
Yolov4-t [1]6405.9116.162.632.24.4
Yolov5-n [14]6401.794.274.047.03.4
Yolov6-n [16]6404.3011.180.258.83.5
Yolov7-t [35]6406.0613.378.454.65.0
Yolov8-n [15]6403.018.179.859.73.7
Part I: IHP (ours)
Yolov3-t-ihp6409.3315.059.9(↑1.9)29.1(↑1.8)3.8
Yolov4-t-ihp6406.6316.063.8(↑1.2)33.9(↑1.7)4.1
Yolov5-n-ihp6401.663.974.8(↑0.8)48.6(↑1.6)3.2
Yolov6-n-ihp6404.3710.679.4(↓0.8)57.4(↓1.4)3.7
Yolov7-t-ihp6406.7714.580.7(↑2.3)55.3(↑0.7)4.7
Yolov8-n-ihp6402.998.179.3(↓0.5)59.1(↓0.6)3.7
Part II: SNI (ours)
Yolov3-t-sni6408.7113.058.6(↑0.6)28.0(↑0.7)3.8
Yolov4-t-sni6405.9116.163.4(↑0.8)33.0(↑0.8)4.4
Yolov5-n-sni6401.794.274.5(↑0.5)47.7(↑0.7)3.4
Yolov6-n-sni6404.2811.081.9(↑1.7)60.8(↑2.0)3.4
Yolov7-t-sni6406.0613.381.5(↑3.1)57.9(↑3.3)5.0
Yolov8-n-sni6403.018.180.3(↑0.5)60.1(↑0.4)3.7
Part III: ESD-I (ours)
Yolov3-t-esd-I64011.822.469.2(↑11.2)38.5(↑11.2)4.6
Yolov4-t-esd-I6405.9116.163.1(↑0.7)32.9(↑0.7)4.4
Yolov5-n-esd-I6401.784.275.4(↑1.4)48.0(↑1.0)3.8
Yolov6-n-esd-I6404.3011.180.5(↑0.3)58.83.4
Yolov7-t-esd-I6406.8314.679.2(↑0.8)55.3(↑0.7)5.2
Yolov8-n-esd-I6403.228.180.1(↑0.3)59.8(↑0.1)3.8
ESD-II (ours)
Yolov3-t-esd-II6408.7814.863.0(↑5.0)32.1(↑4.8)4.9
Yolov4-t-esd-II6405.9316.663.9(↑1.5)33.5(↑1.3)4.7
Yolov5-n-esd-II6402.324.775.7(↑1.7)48.9(↑1.9)3.6
Yolov6-n-esd-II6404.3211.180.4(↑0.2)58.83.5
Yolov7-t-esd-II6407.6915.779.3(↑0.9)55.2(↑0.6)5.9
Yolov8-n-esd-II6403.188.580.4(↑0.6)59.9(↑0.2)4.0
" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.331, + 0.128 + ], + "angle": 0, + "content": "Hulin Li" + }, + { + "type": "table_caption", + "bbox": [ + 0.224, + 0.145, + 0.778, + 0.159 + ], + "angle": 0, + "content": "Table 2: Ablation experiments of the GSCnv and GSCnvE-I/II on VOC \\(07 + 12\\)" + }, + { + "type": "table", + "bbox": [ + 0.244, + 0.172, + 0.76, + 0.443 + ], + "angle": 0, + "content": "
ModelsSizeParam.(M)FLOPs(G)AP50(%)AP(%)LatencyT4b=16(ms)
benchlines(GSConv [18])
Yolov3-tiny-gsconv6404.618.369.941.13.9
Yolov4-tiny-gsconv6403.4212.473.445.24.6
Yolov5-n-gsconv6401.513.974.247.93.5
Yolov6-n-gsconv6404.2711.080.459.13.4
Yolov7-tiny-gsconv6405.0011.478.455.35.6
Yolov8-n-gsconv6402.747.880.961.23.7
GSCConvE-I (ours)
Yolov3-t-gse-I6407.8611.671.7(↑1.8)43.8(↑2.7)3.9
Yolov4-t-gse-I6404.8814.374.9(↑1.5)47.8(↑2.6)4.7
Yolov5-n-gse-I6401.764.275.8(↑1.6)50.1(↑2.2)3.5
Yolov6-n-gse-I6404.3011.180.1(↓0.3)58.8(↓0.3)3.6
Yolov7-t-gse-I6406.2113.579.1(↑0.7)56.4(↑1.1)5.8
Yolov8-n-gse-I6402.968.181.2(↑0.3)61.3(↑0.1)3.8
GSCConvE-II (ours)
Yolov3-t-gse-II6402.856.370.1(↑0.2)41.4(↑0.3)4.0
Yolov4-t-gse-II6402.3610.875.6(↑2.2)48.0(↑2.8)4.6
Yolov5-n-gse-II6401.483.976.3(↑2.1)51.4(↑3.5)3.5
Yolov6-n-gse-II6404.2811.180.3(↓0.1)58.8(↓0.3)3.5
Yolov7-t-gse-II6404.3310.779.1(↑0.7)56.2(↑0.9)5.8
Yolov8-n-gse-II6402.677.779.9(↓1.0)60.4(↓0.8)3.8
" + }, + { + "type": "table_caption", + "bbox": [ + 0.219, + 0.465, + 0.782, + 0.479 + ], + "angle": 0, + "content": "Table 3: A contrast of the different confidence threshold for the SYOLO on COCO." + }, + { + "type": "table", + "bbox": [ + 0.28, + 0.492, + 0.724, + 0.539 + ], + "angle": 0, + "content": "
Confidence thresholdAP(%)APs(%)APm(%)APl(%)
0.001(valuation)53.135.558.769.9
0.25(application)48.7(↓4.4)29.7(↓5.8)54.5(↓4.2)66.5(↓3.4)
" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.577, + 0.478, + 0.592 + ], + "angle": 0, + "content": "4.3 Comparison experiments" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.624, + 0.788, + 0.716 + ], + "angle": 0, + "content": "In Tab. 4 and Tab. 5, we compare state-of-the-art real-time detectors and SYOLO (architecture in Tab. 7) on the Pascal VOC and COCO benchmarks. In Tab. 6, we provide additional comparisons of state-of-the-art detectors on COCO [22], including both real-time (CNNs) and non-real-time (Transformers/CNNs) models. Parameters and FLOPs metrics significantly higher than those of real-time detector models are highlighted in red." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.788, + 0.841 + ], + "angle": 0, + "content": "In Tab. 4, SNI increased the AP of YOLOv7 by 0.9 percentage points without any additional computational or memory costs. SYOLO, utilizing the SA solution, achieved the highest accuracy among competitive models on the Pascal VOC benchmark. In Tab. 5, SYOLO also achieved the highest accuracy among competitive models on the MS COCO benchmark at input resolutions of \\(416 \\times 416\\) and \\(640 \\times 640\\). Positive performance improvements across different benchmarks demonstrate the effectiveness of the SA solution for real-time object detection." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.312, + 0.115, + 0.732, + 0.13 + ], + "angle": 0, + "content": "Rethinking Features-Fused-Pyramid-Neck for Object Detection" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "table_caption", + "bbox": [ + 0.239, + 0.145, + 0.761, + 0.159 + ], + "angle": 0, + "content": "Table 4: Comparasions of state-of-the-art real-time detectors on VOC \\(07 + 12\\)" + }, + { + "type": "table", + "bbox": [ + 0.22, + 0.172, + 0.783, + 0.309 + ], + "angle": 0, + "content": "
ModelsYearNeckSizeParam.(M)FLOPs(G)AP(%)LatencyT4b=16(ms)
Yolov3-spp [31]2018FPN64061.60154.960.539.4
Yolov4 [1]2020PANet†64052.57119.366.231.0
Yolov5 [14]2022PANet†64046.21108.063.930.1
YoloX [7]2021PANet†64053.79153.766.740.2
Yolov6 [16]2022PANet†64065.81159.169.229.5
Yolov7 [35]2022PANet†64036.58103.568.223.2
Yolov8 [15]2023PANet†64043.62164.969.331.6
SYolov7(ours)2023SA-PANet††64036.58103.569.123.2
SYolo(ours)2023SA64057.23142.469.629.3
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.231, + 0.31, + 0.721, + 0.322 + ], + "angle": 0, + "content": "† Using simplified PANet: the red and green shortcuts are cancelled in the Fig. 1 (e)." + }, + { + "type": "table_footnote", + "bbox": [ + 0.231, + 0.321, + 0.73, + 0.334 + ], + "angle": 0, + "content": "\\(\\dagger \\dagger\\) Using SNI replace the nearest neighbor interpolation to up-sampling in the PANet†." + }, + { + "type": "list", + "bbox": [ + 0.231, + 0.31, + 0.73, + 0.334 + ], + "angle": 0, + "content": null + }, + { + "type": "table_caption", + "bbox": [ + 0.259, + 0.346, + 0.742, + 0.36 + ], + "angle": 0, + "content": "Table 5: Comparisons of state-of-the-art real-time detectors on COCO." + }, + { + "type": "table", + "bbox": [ + 0.216, + 0.373, + 0.795, + 0.496 + ], + "angle": 0, + "content": "
ModelsSizeParam.(M)FLOPs(G)AP(%)APs(%)APm(%)APl(%)LatencyT4b=16(ms)
YoloX-t [7]4165.16.532.8---4.3
SYolo-n(ours)4164.812.035.914.339.254.44.3
YoloX-t [7]6405.115.434.7---7.2
Yolov5-s [14]6407.216.537.4---9.8
Yolov6-n [16]6404.711.437.5---4.3
Yolov7-t [35]6406.213.838.718.842.451.95.3
Yolov8-n [15]6403.28.737.3---4.1
SYolo-s(ours)6406.615.740.821.145.356.95.8
" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.523, + 0.358, + 0.539 + ], + "angle": 0, + "content": "5 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.554, + 0.788, + 0.674 + ], + "angle": 0, + "content": "We explore the nature of representation learning and identify that the fusion of different level feature maps faces the issue of feature misalignment. To address this, we introduce the SNI. Additionally, we propose the ESD to mitigate spatial information loss during downsampling and further optimize lightweight convolutional methods for computational efficiency in lightweight models. Based on these advancements, we propose the SA solution for real-time detection, where all tested detectors surpass their original performance and achieve state-of-the-art results." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.676, + 0.788, + 0.797 + ], + "angle": 0, + "content": "Object detection has significantly benefited from the FPN-like paradigm in past few years. However, recent developments indicate that newer technologies, including innovative architectures and training strategies, may offer superior performance for modern detectors, making the FPN-like paradigm potentially limiting. In many cases, the effectiveness of the FPN-like paradigm can be enhanced through the application of SNI to address feature misalignment issues. Fully addressing the challenge of misalignment during feature fusion in existing vision models requires further research efforts." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.331, + 0.128 + ], + "angle": 0, + "content": "Hulin Li" + }, + { + "type": "table_caption", + "bbox": [ + 0.29, + 0.147, + 0.712, + 0.161 + ], + "angle": 0, + "content": "Table 6: Comparisons of state-of-the-art detectors on COCO." + }, + { + "type": "table", + "bbox": [ + 0.217, + 0.173, + 0.787, + 0.466 + ], + "angle": 0, + "content": "
ModelsSizeParam.(M)FLOPs(G)AP(%)APs(%)APm(%)APl(%)LatencyV100b=1 (ms)
Transformers/CNNs (non-real-time, pre-trained)
FPN [21]-56.4145.836.817.538.747.8-
BiFPN [34]102421.055.049.3---42.8
ASFF [23]800--43.927.046.653.434.0
PANet [24]---51.032.653.964.8-
RFPN [28]---51.331.754.664.8-
DETR-Def. [40]-40.0173.043.826.447.158.0-
DETR-Dyn. [3]---47.228.649.359.1-
DETR-DND [17]133348.0265.048.6----
DINO [38]133347.0860.051.0----
Swin-T [26]-86.0745.050.5----
Swin-S [26]-107.0838.051.8----
Swin-B [26]1280145.0982.051.9----
CNNs (real-time, without pre-trained)
Yolov4 [1]60864.4142.843.526.746.753.3-
YoloX [7]64054.2155.649.7---14.5
PPYoloE [37]64052.2110.150.931.455.366.112.8
Yolov5 [14]64046.5109.149.0---10.1
Yolov6 [16]64058.5144.051.7----
Yolov7 [35]64036.9104.751.231.855.565.06.3
Yolov8 [15]64043.7165.252.9----
SYolo(ours)64057.2142.553.135.558.769.910.3
" + }, + { + "type": "table_caption", + "bbox": [ + 0.229, + 0.475, + 0.773, + 0.49 + ], + "angle": 0, + "content": "Table 7: Architecture of the SYOLO. The E-ELAN/C2f is from the YOLOv7/8." + }, + { + "type": "table", + "bbox": [ + 0.317, + 0.503, + 0.687, + 0.793 + ], + "angle": 0, + "content": "
StageBlock-1/s/n‡Channels-1/s/n‡K. sizeStride
Backbone
P1Conv64 /16 /1632
P2ESD-I/II128 /32 /323, 42
C2f ×3/2/1128 /32 /323, 11
P3ESD-I/II256 /64 /643, 42
C2f ×6/3/1256 /64 /643, 11
P4ESD-I/II512 /128 /1283, 42
C2f ×6/3/1512 /128 /1283, 11
P5ESD-I/II1024 /256 /2563, 42
C2f ×3/2/11024 /256 /2563, 11
SPP ×11024 /256 /256--
Neck(SA)
P5-P4SNI1536 /384 /284--
E-ELAN†×12048 /512 /5123, 11
P4-P3SNI512 /192 /192--
E-ELAN†×2/2/1512 /256 /2563, 11
P3-P4ESD-I/II256 /128 /1283, 42
E-ELAN†×2/2/11024 /512 /5123, 11
P4-P5ESD-I/II512 /256 /2563, 42
E-ELAN†×2/2/12048 /1024 /10243, 11
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.325, + 0.794, + 0.686, + 0.825 + ], + "angle": 0, + "content": "The first \\(3 \\times 3\\) vanilla convolutional layer of the E-ELAN is replaced by the GSCConvE-I, and the end pointwise convolutional layer of the E-ELAN is replaced by the C2f." + }, + { + "type": "table_footnote", + "bbox": [ + 0.327, + 0.825, + 0.678, + 0.836 + ], + "angle": 0, + "content": "\\(\\ddagger\\) The 'l/s/n' means the model scale of the large/samll/nano." + }, + { + "type": "list", + "bbox": [ + 0.325, + 0.794, + 0.686, + 0.836 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.312, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Rethinking Features-Fused-Pyramid-Neck for Object Detection" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.403, + 0.163 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.175, + 0.788, + 0.268 + ], + "angle": 0, + "content": "This work is supported by National Natural Science Foundation of China (Grant No. 52172381), Graduate Research Innovation Foundation of Chongqing Jiaotong University (Grant No. CYB240259). The author would also like to thank Prof. Dr. Jun Li, Prof. Dr. Qiliang Ren, and Ms. Xinxin Liu (Intelligent Transportation Big Data Center, School of Information Science and Engineering, CQJTU) for providing GPUs." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.289, + 0.323, + 0.305 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.318, + 0.785, + 0.346 + ], + "angle": 0, + "content": "1. Bochkovskiy, A., Wang, C.Y., Liao, H.Y.M.: Yolov4: Optimal speed and accuracy of object detection. arXiv preprint arXiv:2004.10934 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.347, + 0.785, + 0.387 + ], + "angle": 0, + "content": "2. Chollet, F.: Xception: Deep learning with depthwise separable convolutions. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1251-1258 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.388, + 0.785, + 0.429 + ], + "angle": 0, + "content": "3. Dai, X., Chen, Y., Yang, J., Zhang, P., Yuan, L., Zhang, L.: Dynamic detr: End-to-end object detection with dynamic attention. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 2988-2997 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.429, + 0.785, + 0.47 + ], + "angle": 0, + "content": "4. Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: Imagenet: A large-scale hierarchical image database. In: 2009 IEEE conference on computer vision and pattern recognition. pp. 248-255. IEEE (2009)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.471, + 0.785, + 0.511 + ], + "angle": 0, + "content": "5. Everingham, M., Eslami, S.A., Van Gool, L., Williams, C.K., Winn, J., Zisserman, A.: The Pascal visual object classes challenge: A retrospective. International journal of computer vision 111, 98-136 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.511, + 0.785, + 0.539 + ], + "angle": 0, + "content": "6. Fu, C.Y., Liu, W., Ranga, A., Tyagi, A., Berg, A.C.: Dssd: Deconvolutional single shot detector. arXiv preprint arXiv:1701.06659 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.539, + 0.785, + 0.566 + ], + "angle": 0, + "content": "7. Ge, Z., Liu, S., Wang, F., Li, Z., Sun, J.: Yolox: Exceeding yolo series in 2021. arXiv preprint arXiv:2107.08430 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.566, + 0.785, + 0.594 + ], + "angle": 0, + "content": "8. Girshick, R.: Fast r-cnn. In: Proceedings of the IEEE international conference on computer vision. pp. 1440-1448 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.594, + 0.785, + 0.635 + ], + "angle": 0, + "content": "9. Girshick, R., Donahue, J., Darrell, T., Malik, J.: Rich feature hierarchies for accurate object detection and semantic segmentation. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 580-587 (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.634, + 0.785, + 0.676 + ], + "angle": 0, + "content": "10. Han, K., Wang, Y., Tian, Q., Guo, J., Xu, C., Xu, C.: Ghostnet: More features from cheap operations. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 1580-1589 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.676, + 0.785, + 0.717 + ], + "angle": 0, + "content": "1. He, K., Zhang, X., Ren, S., Sun, J.: Spatial pyramid pooling in deep convolutional networks for visual recognition. IEEE transactions on pattern analysis and machine intelligence 37(9), 1904-1916 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.717, + 0.785, + 0.758 + ], + "angle": 0, + "content": "2. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 770-778 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.759, + 0.785, + 0.8 + ], + "angle": 0, + "content": "3. Howard, A.G., Zhu, M., Chen, B., Kalenichenko, D., Wang, W., Weyand, T., Andreetto, M., Adam, H.: Mobilenets: Efficient convolutional neural networks for mobile vision applications. arXiv preprint arXiv:1704.04861 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.8, + 0.785, + 0.84 + ], + "angle": 0, + "content": "4. Jocher, G., Chaurasia, A., Stoken, A., Borovec, J., Kwon, Y., Fang, J., Michael, K., Montes, D., Nadar, J., Skalski, P., et al.: ultralytics/yolov5: v6. 1-tensorrt, tensorflow edge tpu and openvino export and inference. Zenodo (2022)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.318, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.331, + 0.128 + ], + "angle": 0, + "content": "Hulin Li" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.148, + 0.785, + 0.175 + ], + "angle": 0, + "content": "15. Jocher, G., et al.: Yolo-ultralytics. https://github.com/ultralytics/ultralytics (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.176, + 0.785, + 0.216 + ], + "angle": 0, + "content": "16. Li, C., Li, L., Jiang, H., Weng, K., Geng, Y., Li, L., Ke, Z., Li, Q., Cheng, M., Nie, W., et al.: Yolov6: A single-stage object detection framework for industrial applications. arXiv preprint arXiv:2209.02976 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.216, + 0.785, + 0.257 + ], + "angle": 0, + "content": "17. Li, F., Zhang, H., Liu, S., Guo, J., Ni, L.M., Zhang, L.: Dn-detr: Accelerate detr training by introducing query denoising. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 13619-13627 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.257, + 0.785, + 0.298 + ], + "angle": 0, + "content": "18. Li, H., Li, J., Wei, H., Liu, Z., Zhan, Z., Ren, Q.: Slim-neck by gsconv: a lightweight-design for real-time detector architectures. Journal of Real-Time Image Processing 21(3), 62 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.298, + 0.785, + 0.338 + ], + "angle": 0, + "content": "19. Li, Y., Chen, Y., Wang, N., Zhang, Z.: Scale-aware trident networks for object detection. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 6054-6063 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.338, + 0.785, + 0.365 + ], + "angle": 0, + "content": "20. Li, Z., Yang, L., Zhou, F.: Fssd: feature fusion single shot multibox detector. arXiv preprint arXiv:1712.00960 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.365, + 0.785, + 0.406 + ], + "angle": 0, + "content": "21. Lin, T.Y., Dollar, P., Girshick, R., He, K., Hariharan, B., Belongie, S.: Feature pyramid networks for object detection. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 2117-2125 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.406, + 0.785, + 0.461 + ], + "angle": 0, + "content": "22. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dólar, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13. pp. 740-755. Springer (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.461, + 0.785, + 0.487 + ], + "angle": 0, + "content": "23. Liu, S., Huang, D., Wang, Y.: Learning spatial fusion for single-shot object detection. arxiv 2019. arXiv preprint arXiv:1911.09516 (1911)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.487, + 0.785, + 0.528 + ], + "angle": 0, + "content": "24. Liu, S., Qi, L., Qin, H., Shi, J., Jia, J.: Path aggregation network for instance segmentation. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 8759-8768 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.528, + 0.785, + 0.582 + ], + "angle": 0, + "content": "25. Liu, W., Anguelov, D., Erhan, D., Szegedy, C., Reed, S., Fu, C.Y., Berg, A.C.: Ssd: Single shot multibox detector. In: Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part I 14. pp. 21-37. Springer (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.582, + 0.785, + 0.637 + ], + "angle": 0, + "content": "26. Liu, Z., Lin, Y., Cao, Y., Hu, H., Wei, Y., Zhang, Z., Lin, S., Guo, B.: Swin transformer: Hierarchical vision transformer using shifted windows. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 10012-10022 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.637, + 0.785, + 0.678 + ], + "angle": 0, + "content": "27. Liu, Z., Mao, H., Wu, C.Y., Feichtenhofer, C., Darrell, T., Xie, S.: A convnet for the 2020s. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 11976-11986 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.679, + 0.785, + 0.718 + ], + "angle": 0, + "content": "28. Qiao, S., Chen, L.C., Yuille, A.: Detectors: Detecting objects with recursive feature pyramid and switchable atrous convolution. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10213-10224 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.718, + 0.785, + 0.76 + ], + "angle": 0, + "content": "29. Redmon, J., Divvala, S., Girshick, R., Farhadi, A.: You only look once: Unified, real-time object detection. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 779-788 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.76, + 0.785, + 0.786 + ], + "angle": 0, + "content": "30. Redmon, J., Farhadi, A.: Yolo9000: better, faster, stronger. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 7263-7271 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.786, + 0.785, + 0.813 + ], + "angle": 0, + "content": "31. Redmon, J., Farhadi, A.: Yolov3: An incremental improvement. arXiv preprint arXiv:1804.02767 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.813, + 0.785, + 0.84 + ], + "angle": 0, + "content": "32. Simonyan, K., Zisserman, A.: Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556 (2014)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.148, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.312, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Rethinking Features-Fused-Pyramid-Neck for Object Detection" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.189 + ], + "angle": 0, + "content": "33. Singh, B., Davis, L.S.: An analysis of scale invariance in object detection snip. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 3578-3587 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.19, + 0.788, + 0.232 + ], + "angle": 0, + "content": "34. Tan, M., Pang, R., Le, Q.V.: Efficientdet: Scalable and efficient object detection. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10781-10790 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.232, + 0.788, + 0.286 + ], + "angle": 0, + "content": "35. Wang, C.Y., Bochkovskiy, A., Liao, H.Y.M.: Yolov7: Trainable bag-of-freebies sets new state-of-the-art for real-time object detectors. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 7464-7475 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.286, + 0.788, + 0.328 + ], + "angle": 0, + "content": "36. Wu, Y., Chen, Y., Yuan, L., Liu, Z., Wang, L., Li, H., Fu, Y.: Rethinking classification and localization for object detection. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10186-10195 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.328, + 0.788, + 0.37 + ], + "angle": 0, + "content": "37. Yang, Z., Zhou, Y., Chen, Z., Ngiam, J.: 3d-man: 3d multi-frame attention network for object detection. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 1863-1872 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.371, + 0.788, + 0.411 + ], + "angle": 0, + "content": "38. Zhang, H., Li, F., Liu, S., Zhang, L., Su, H., Zhu, J., Ni, L.M., Shum, H.Y.: Dino: Detr with improved denoising anchor boxes for end-to-end object detection. arXiv preprint arXiv:2203.03605 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.411, + 0.788, + 0.453 + ], + "angle": 0, + "content": "39. Zhang, X., Zhou, X., Lin, M., Sun, J.: Shufflenet: An extremely efficient convolutional neural network for mobile devices. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 6848-6856 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.453, + 0.788, + 0.495 + ], + "angle": 0, + "content": "40. Zhu, X., Su, W., Lu, L., Li, B., Wang, X., Dai, J.: Deformable detr: Deformable transformers for end-to-end object detection. arXiv preprint arXiv:2010.04159 (2020)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.495 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/Rethinking Features-Fused-Pyramid-Neck for Object Detection/8cefd2b8-bab0-4ef4-a711-578971c2749a_origin.pdf b/2024/Rethinking Features-Fused-Pyramid-Neck for Object Detection/8cefd2b8-bab0-4ef4-a711-578971c2749a_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..05d308ddc6604d7368927a9f3627b2f0df3c3b15 --- /dev/null +++ b/2024/Rethinking Features-Fused-Pyramid-Neck for Object Detection/8cefd2b8-bab0-4ef4-a711-578971c2749a_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:cb7dd42695e556c16113c5e07d8dd220af1d54cb0c3970efa7c4ebf5fce5dd2d +size 852357 diff --git a/2024/Rethinking Features-Fused-Pyramid-Neck for Object Detection/full.md b/2024/Rethinking Features-Fused-Pyramid-Neck for Object Detection/full.md new file mode 100644 index 0000000000000000000000000000000000000000..9e7d28280fbce874dd9e5deeff868e51c55476d4 --- /dev/null +++ b/2024/Rethinking Features-Fused-Pyramid-Neck for Object Detection/full.md @@ -0,0 +1,221 @@ +# Rethinking Features-Fused-Pyramid-Neck for Object Detection + +Hulin Li + +College of Traffic and Transportation, Chongqing Jiaotong University, Chongqing + +400074, China + +alan@mails.cqjtu.edu.cn + +Abstract. Multi-head detectors typically employ a features-fused-pyramid-neck for multi-scale detection and are widely adopted in the industry. However, this approach faces feature misalignment when representations from different hierarchical levels of the feature pyramid are forcibly fused point-to-point. To address this issue, we designed an independent hierarchy pyramid (IHP) architecture to evaluate the effectiveness of the features-unfused-pyramid-neck for multi-head detectors. Subsequently, we introduced soft nearest neighbor interpolation (SNI) with a weight-downscaling factor to mitigate the impact of feature fusion at different hierarchies while preserving key textures. Furthermore, we present a feature adaptive selection method for downsampling in extended spatial windows (ESD) to retain spatial features and enhance lightweight convolutional techniques (GSCnvE). These advancements culminate in our secondary features alignment solution (SA) for real-time detection, achieving state-of-the-art results on Pascal VOC and MS COCO. Code will be released at https://github.com/AlanLi1997/rethinking-fpn. + +Keywords: Object detection $\cdot$ Feature pyramid $\cdot$ Feature misalignment + +- Detection-net architecture + +# 1 Introduction + +Bounding box regression for object detection and pixel-level boundary regression for instance segmentation are pivotal and challenging visual tasks. Unlike basic image classification [4], detection or segmentation tasks often involve scenes with multiple objects of varying scales. Early deep learning-based object detection models began exploring real-time multi-scale object detection. RCNN [9] employed selective search to extract a large number of region proposals ( $\sim 2\mathrm{k}$ ). Although the vast number of image patches (region proposals) might aid multiscale detection, in practice, most scenes do not contain such a high number of objects to recognize. This approach was later deemed inefficient [8], as the proposed regions were derived from the image rather than features, necessitating a convolutional neural network (CNN) to extract features and predict classes for each proposal. A simplified RCNN architecture, as depicted in Fig. 1 (a), converts images of different scales into feature maps for object detection, largely + +![](images/4c00a44d7d5e66f966c965977af8b1a4e6eca56737b7313dc346c4f8d83f0de6.jpg) + +![](images/46e922b9812dc914407c5906f7d4679413f46f612b34e326933f9c8c30f618eb.jpg) + +![](images/07fc4afd6132de0f0819f2796adf83869bf34853c4df2a337a1583fca55c6a96.jpg) + +![](images/e235f6347163f4ca2ab3ad192eab4fa8ba580dd26370239588b1662c5193bf38.jpg) + +![](images/6eea46c13ef56cb9ed5041b7b18f38bb731dbbbb9057366f8718144a1063fb3e.jpg) + +![](images/1c5625fcb681996bf3c0a52916ed6016f3634610c6112e793b0886b8225da4ef.jpg) +Fig. 1: The IHP and seven other typical neck architectures. The three heat maps are come from the third (P3), fourth (P4), and fifth (P5) hierarchies of the FPN-neck model, respectively. The area of interest for each hierarchies' features is highlighted in red; the darker the color, the higher the feature weight. It is evident that lower-level hierarchies prioritize local features, while higher-level hierarchies focus more on global features. In other words, there is a representation bias at different hierarchical levels. If these features are combined by point-to-point fusion without first addressing the misalignment issue, it may introduce noise rather than enrich semantics. + +![](images/c286cbcb9e9125c4ee6a164748652395c728ef953caddb8cfad2f13b9420d70b.jpg) + +neglecting the real-time detection capability of models. YOLO [29] introduced the first real-time deep learning-based object detection model, which did not specifically focus on multi-scale detection, with a simplified structure shown in Fig. 1 (b). SSD [25] effectively combined the strengths of RCNN and YOLO, emphasizing the concept of multi-scale detection. It was the first model to achieve real-time multi-scale object detection using multi-scale features, with an architecture similar to Fig. 1 (c). SSD efficiently utilized feature maps of different scales from the VGG-16 backbone [32] to predict objects of varying scales. The exploration by early deep learning-based object detection models provided an important insight to the field: the efficient utilization of multi-scale features is crucial for enhancing the overall performance (accuracy-speed trade-off) of detectors. Consequently, FPN [21] introduced the feature pyramid-based multi-scale + +feature fusion concept, which gradually became a structural paradigm for detectors. This led to numerous representative works enhancing multi-scale feature fusion, such as PANet [24], ASFF [23], BiFPN [34], and RFPN [28]. The performance of real-time detectors has improved with the implementation of feature fusion strategies. However, further advancements are hindered by the increasing complexity of these techniques. Before the introduction of FPN, object detectors primarily relied on structures without fusion [9,25,29], as depicted in Fig. 1 (a)-(c). With the advent of FPN, these detectors have shifted towards FPN-based paradigms or more intricate improvement methods [8,20,29-31], as illustrated in Fig. 1 (e)-(g). It is important to note that the FPN-like strategy should not be considered an obligatory component for all detectors. Representation learning suggests that models ultimately require refined representations rather than an abundance of complex representations. Under the FPN-like paradigm, feature maps of different hierarchies and different resolutions are combined at the element level (point-to-point), necessitating the expansion of low-resolution, high-level feature maps to match the resolution of low-level feature maps. However, there are inherent representation biases in feature maps at different hierarchical levels: low-level feature maps tend to retain representations of small objects (local features in small receptive fields), while high-level feature maps favor representations of large objects (global features in large receptive fields), as shown in Fig. 1 (i). The direct element-level fusion of these feature maps with representation bias may lead to partial destruction of representations and exacerbate the feature misalignment problem during fusion. + +In visual models, the resolution of feature maps decreases progressively from low to high hierarchical levels of the feature pyramid, achieved through step-by-step downsampling. Typically, downsampling reduces the resolution (width $\times$ height) of feature maps at a square root rate, inherently filtering out some spatial information regardless of whether the number of features (channels) remains constant or increases. It is also noteworthy that the transition of feature information from spatial to channel dimensions primarily occurs during downsampling, underscoring the importance of this stage. Traditional downsampling methods operate in a single mode, such as convolution with a stride of 2 (learnable) or average pooling and max pooling with a stride of 2 (non-learnable). These single-mode approaches have persisted since the introduction of ResNet [12] without significant evolution, which merits further investigation. Additionally, the initial purpose of FPN was not only to demonstrate the advantages of feature fusion in addressing multi-scale detection challenges but also to appropriately reduce model complexity. In industries with limited hardware resources, complex models are often impractical [2] [10]. Therefore, we enhance the lightweight convolution technique of the GSCov [18] to achieve a preferable trade-off between accuracy and speed for real-time detection models. + +The main contributions of this work can be summarized as follows: + +1). We rethink the effectiveness of FPN-like paradigms for modern real-time detectors based on representation learning and identify issues with feature misalignment during element-level fusion. + +2). We design an independent hierarchy pyramid architecture (IHP) without feature fusion in the neck to validate our findings. And the IHP achieves advanced results. +3). We introduce soft nearest neighbor interpolation (SNI) for up-sampling to mitigate feature misalignment during fusion. +4). We provide a feature adaptive selection method in extended spatial windows for downsampling (ESD) to enhance spatial feature retention during the downsampling stage. This method is plug-and-play and optimizes performance at a low cost. +5). We simplify and enhance the GSCnv lightweight convolution (GSCnvE) to offer more cost-effective options for models operating on edge devices with resource constraints. +6).We validate our approaches on Pascal VOC and COCO and present the SA solution. Existing lightweight real-time detection models can be optimized with the SA solution without bells and whistles. + +# 2 Related Work + +# 2.1 Multi-head detection and features fusion + +The first-generation general detection models using deep learning consist of two main components: the backbone and the detecting-head [9, 29]. These models typically use final feature maps for prediction. SSD [25] introduced the use of multi-level feature maps for object detection, marking the beginning of multi-head detection and influencing subsequent real-time detector designs [31]. In the SSD model, each detecting-head directly uses raw feature maps from different levels of the backbone without any fusion. The key advantage of the multi-head detection method is the ability to predict objects of different scales using feature maps with varying receptive fields. High-level feature maps, with their wide global receptive fields, are beneficial for identifying large objects, while low-level feature maps, with refined local receptive fields, are helpful for identifying small objects. FPN introduced a fusion scheme for multi-level feature maps to enhance detection accuracy, setting the trend for feature fusion and becoming a common practice. This led to the development of Backbone-Neck-Head architectures. More complex variants based on FPN, such as BiFPN [34] and PANet [24], have since been proposed, enabling the use of more features for prediction. However, as prediction accuracy improved, networks became increasingly complex due to additional fusion layers. Some approaches have attempted to bypass feature fusion schemes by exploring new perspectives on multi-scale detection challenges, such as optimizing multi-scale training strategies [33] or using dilated convolutions to capture multi-scale receptive fields [19]. But these methods have gradually lost their competitiveness with the continuous optimization of classification-backbone feature extraction capabilities and the ongoing development of feature fusion techniques [1]. + +# 2.2 Lightweight + +The choice between cloud computing or edge computing for a model depends on its number of parameters and floating-point operations (FLOPs). Therefore, reducing the number of parameters or FLOPs is a primary focus in lightweight model studies. Direct lightweight approaches include reducing the depth (number of layers) or width (number of neurons/filters) of the model, as seen in models likeYOLO-fast/tiny/nano [1,14-16,31,35]. Additionally, sparse computing techniques, such as depth-wise separable convolution used in MobileNets [13] and ShuffleNet [39], can also be employed. However, low-depth networks often suffer from underfitting due to insufficient nonlinear representation capabilities. Therefore, reducing the number of network layers may not always be a cost-effective way to lighten the model. + +# 3 Secondary Features Alignment Solution + +Key contents of this work are described in detail in this section. Specifically, there are the IHP structure to directly demonstrate features unalignment problem by abandoning fusion, the SNI to alleviate unaligned during features fusion by point-to-point, the ESD to enhance spatial-features capture during the downsampling stage, and the GSConvE to improve the performance between the accuracy and speed of lightweight models. + +# 3.1 Independent hierarchy pyramid architecture + +The introduction of FPN brought all real-time detectors into the realm of feature fusion. However, we have identified a significant issue arising from the point-to-point fusion of different level features - partial local features become unaligned. This fusion process is akin to adding noise when unaligned features are combined, as features unrelated to the target space are forcibly integrated, resulting in spatial disarray. For instance, as depicted in Fig. 2, adding the head-feature of a puppy to the butt creates an undefined breed when fed to the detection-head. This scenario is particularly evident when the puppy occupies fewer pixels. Typically, nearest neighbor interpolation is used to rapidly increase the resolution of high-level feature maps to match that of low-level feature maps, followed by fusion through methods such as element-wise addition, channel concatenation, or weighted sum. However, downsampling tends to be nonlinear and irreversible, leading to inconsistencies in spatial features among the additional features generated by up-sampling. Although BiFPN [34] offers a learnable fusion method, it has not been widely adopted by recent state-of-the-art real-time detectors, such as YOLOv7 [35] or YOLOv8 [15], due to the increased complexity and lack of demonstrated comprehensive performance improvements in these models. + +The IHP adopts a radical approach by abandoning all fusions in the neck, fundamentally circumventing the features misalignment problem and streamlining the structure. By leveraging the inherent benefits of the multi-head detection + +mode, the IHP directly predicts objects of different sizes using feature maps of different levels. It is important to note that there are significant differences between IHP and the SSD architecture: SSD directly uses different level features from the backbone to detect objects but the IHP introduces a bottleneck convolution module before prediction to filter the features. The advantage of this approach is that it allows the model to independently learn features for a specific scale branch without affecting features at other levels. Directly stacking convolutional blocks at different levels of the backbone can lead to interference between different levels, which we believe might be the primary reason why using a very deep backbone in object detection models can result in performance degradation. For instance, DSSD [6] showed a performance drop when ResNet-101 was used as the backbone instead of VGG-16 in the SSD detector. In Part I of Tab. 1, all baselines of coupled-head YOLOs demonstrated improved results when using the IHP. Features fusion often overlooks the potential introduction of noise when augmenting semantic richness for prediction. While the IHP may appear to reduce semantic information, it effectively harnesses the advantages of multi-head detection to circumvent the misalignment problem. This strategic utilization of multi-head detection is the key factor behind the competitive results achieved by the IHP. Unfortunately, unlike localization-classification-coupled-head detectors [14,29-31,35], the IHP did not achieve the same positive reaction to localization-classification-decoupled-head detectors [15, 16]. Decoupled-head detectors address a key challenge faced by coupled-head detectors: the disparity in the interests of classification and localization tasks, both of which are often performed using the same layer for prediction [36]. This approach is akin to forcibly fusing misaligned features for prediction. Decoupled-head detectors overcome this limitation by independently predicting classification and localization tasks through separate branches. This enables them to learn beneficial representations from redundant features, which is not possible in coupled-head detectors. Consequently, while FPN strategies are still applicable to some extent in decoupled-head detectors, the features misalignment problem still needs to be addressed. + +# 3.2 Soft nearest neighbor interpolation + +In detection or segmentation models, different-level feature fusion typically occurs after the up-sampling of feature maps, with nearest neighbor interpolation and transposed convolution being popular up-sampling methods. However, the resolution expansion resulting from up-sampling in detection or segmentation differs significantly from tasks such as generation or super-resolution. In detection, the expanded feature maps represent abstract high-level semantic features rather than raw image detail information. Transposed convolution introduces increased computation and latency. To address the features misalignment problem with minimal cost while maintaining the speed of nearest neighbor interpolation, we explore a solution. Nearest neighbor interpolation is akin to average unpooling but is considered a "hard" operation. By softening this operation, similar to the transformation of SoftMax to the Max function, we aim to mitigate the + +![](images/2a45fad2bdaaec7b2d52335aa31b61e76229a697d5c86f5d4ceba1f5c5d3e169.jpg) +Fig. 2: Illustrations of the features misalignment in fusion and the SNI. These numbers, 22-88, are just markers of different local-features not real feature values. + +![](images/beda720ecd6fb3d8211cbb77d398261dbb1d209f5e82a3ddd847392ca09937e4.jpg) +Fig.3: A comparison of the SNI and traditional method. This result could be reproduced by using SNI source code (model:Yolov5n-panet). The same size scaling before and after up-sampling is for intuitive comparison. + +"hard" features misalignment issue. To achieve this, we introduce a soft factor, denoted as $\alpha$ , for the nearest neighbor interpolation during up-sampling: + +$$ +\boldsymbol {Y} = \alpha \cdot \boldsymbol {f} (\boldsymbol {X}), \alpha = \frac {\text {R e s o l u t i o n} \boldsymbol {X}}{\text {R e s o l u t i o n} \boldsymbol {Y}} \tag {1} +$$ + +In Eq. (1), the resolution of the high-level feature maps $(\mathbf{X})$ and low-level feature maps $(\mathbf{Y})$ are denoted as $Resolution_{X}$ and $Resolution_{Y}$ , respectively. The nearest neighbor interpolation operation, represented by $\pmb{f}$ , is illustrated in Fig. 2. The SNI adjusts the influence of high-level semantic features on low-level features based on the zoom factor for feature maps. Specifically, as the zoom factor increases, the impact of high-level semantic features on low-level features weakens. Unlike SoftMax, the SNI does not convert outputs into probabilities but rather mitigates misalignment without additional cost when integrating high- + +level features with low-level features through point-to-point fusion. SNI softens feature weights while preserving key textures, thereby reducing the mutual influence of features from different hierarchical levels during fusion, as qualitatively demonstrated in Fig. 3. This approach achieves the secondary (auxiliary) features alignment (SA) and lowers the complexity for the model to learn from the fused representations. In contrast, traditional methods that directly merge features from different hierarchical levels can cause unstable competitive effects due to representation biases, increasing the learning difficulty for the model. For example, if the three heatmaps in Fig. 1 (i) are directly fused, should the red areas cover the entire image? Clearly, the answer is no. If such a situation occurs, it indicates a failure in prior feature extraction, as the objective of a model is to capture and focus on useful features. This is also why models under the FPN-like paradigm typically require several additional convolutional blocks to further filter features after fusing and before predicting. + +In Part II of Tab. 1, all tested state-of-the-art models demonstrate accuracy improvements solely by incorporating the SNI, confirming the presence of features misalignment in FPN-based models. Notably, for YOLOv6 [16], a detector equipped with various techniques tailored for industry applications, the AP is enhanced by up to 2 percentage points just with the adoption of SNI, while reducing computational cost and inference latency. We achieved significant optimization results in the original YOLOv6 by simply replacing transposed convolution with SNI, highlighting the efficacy of this approach. + +# 3.3 Feature adaptive selection in extended spatial windows + +![](images/28d65be42a1f275f2d41a54dffb09210d9bbcb1d78bf489f100e47d45969f9f7.jpg) +Fig. 4: Illustrations of the ESD-I and ESD-II. It must be mentioned that although the size of the extended window does not have to be same in the ESD-I, pooling is a non-learnable operation, so the features from an unequal window may become noise for other and bring the trouble of the features misalignment again. The ESD-II is free on size of the extended window because of the learnable fusion like the SPPNet [11] for the YOLOv3. + +Feature downsampling is a critical technique in vision models based on deep learning, involving the progressive reduction of feature map resolution. In vision representation learning, models typically rely on low-resolution, high-level + +feature maps for prediction rather than directly processing high-resolution raw images. However, excessive downsampling can result in significant loss of spatial details [12]. Recent detectors typically limit the number of downsampling operations to no more than five times, leading to a spatial resolution reduction of $2^{5}$ times, as demonstrated in YOLOs [1,7,14-16,31,35,37] and Swin Transformer [26]. Common downsampling methods include convolution and pooling, with models typically employing only one of these methods. We propose the feature adaptive selection method in extended spatial windows for downsampling (ESD) to enhance spatial information retention capabilities. ESD consists of two nonlinear branches and one linear branch. In the nonlinear branches, a norm convolutional layer and an extended window max-pooling layer enhance local feature capturing, while extended window average-pooling enhances global feature capturing. Subsequently, linear features are merged with nonlinear features. Two fusion modes, ESD-I and ESD-II, are designed for lightweight and norm models, respectively. The difference lies in the feature fusion stage, as illustrated in Fig. 4. ESD-I employs element-wise addition for merging features, offering low computational cost suitable for lightweight models, while ESD-II utilizes learnable linearly adaptive fusion, slightly increasing computational cost but suitable for norm models. Importantly, local and global features of the extended window are sampled using simple pooling techniques, preserving input information to a significant extent. This reduces information loss during downsampling and enables subsequent layers to learn partial representations from the previous layer, akin to indirectly realizing a hidden shortcut connection reminiscent of ResNet [12]. + +# 3.4 Lightweight GSConv enhancement + +![](images/159ea320bc49abbac2e9443c24897e0139bfdebb01c805a385d2a84699a4f3cb.jpg) +Fig. 5: Illustrations of the GSConvE-I and II. + +GSCnv [18] was introduced to address the channel interaction limitations of depthwise separable convolution by blending features from both vanilla convolution and depthwise convolution. The original GSCnv consists of three fundamental components: a $3 \times 3$ vanilla convolution, a $5 \times 5$ depthwise convolution, and a feature shuffle operation. Building upon this, we propose two evolutions + +of GSConv, namely GSConvE-I and GSConvE-II, tailored for norm models and lightweight models, respectively. GSConvE-I is designed for norm models, where intermediate feature maps on the auxiliary branch are derived from dense linear mappings, while output feature maps originate from sparse linear mappings. This approach maintains feature richness without significantly increasing computational cost. On the other hand, GSConvE-II is geared towards lightweight models. It incorporates three depthwise convolution auxiliary branches with kernel sizes of $9 \times 9$ , $13 \times 13$ , and $17 \times 17$ , respectively. By utilizing larger kernel sizes, this variant can directly capture larger receptive fields and global features with fewer layer accumulations, offering high cost-effectiveness for lightweight models that prioritize reduced FLOPs and layer count. In GSConvE-I, we explore the removal of the batch normalization layer on the depthwise convolution branch while retaining the batch normalization layer on the main branch, similar to ConvNeXt [27]. This adjustment avoids the vanishing gradient problem and simplifies computation, leading to improved prediction accuracy. + +# 4 Experiments + +Real-time detectors are highly favored in industry due to their hardware-friendly nature and the optimal balance they strike between accuracy and speed. In our study, we selected popular YOLO models—YOLOv3 [31], YOLOv4 [1], YOLOv5 [14], YOLOv6 [16], YOLOv7 [35], and YOLOv8 [15]—as baselines to evaluate the efficacy of the SA solution in real-time detection scenarios (SYOLO). + +# 4.1 Datasets + +We conducted trials on the Pascal VOC dataset [5] and the MS COCO dataset [22]. The Pascal VOC 07+12 dataset was utilized for ablation experiments, wherein models underwent training on the train&val07+12 set and were subsequently evaluated on the test07 set. On the other hand, the MS COCO dataset was employed for comparative experiments, with models trained on the train2017 set and evaluated on the val2017 set. To ensure fairness and consistency, all released models were trained using the default set of hyperparameters and without utilizing pre-trained weights. More details of hyperparameters are in the support material. + +# 4.2 Ablation studies + +In Tab. 1, we present the results of ablation experiments conducted on six baselines [1, 14-16, 31, 35], focusing on the impact of the IHP, SNI, ESD. The lightweight experiment results of GSConvE are reported in Tab. 2. All baselines retained their original neck architecture, as we believe the competitive neck has been carefully chosen in the original works. Performance metrics highlighted in green indicate performance gains, while those in red indicate losses. It is crucial + +to address an often-overlooked aspect during the verification process: the inconsistency between reported evaluation accuracy and actual usage. Frequently, detectors are evaluated with a very low confidence threshold, such as 0.001, which results in a large number of prediction boxes and inflated evaluation accuracy. However, in real-world detection scenarios, the confidence threshold is typically set much higher, around 0.25 or above. This filtering process ensures that higher-quality bounding boxes are utilized for practical purposes. The discrepancy between evaluation settings and real-world usage can create a misleading impression of accuracy. Therefore, it is essential to maintain consistency in configuration between evaluation and application. While this may result in a significant drop in evaluation accuracy, as shown in Tab. 3, it provides a more realistic assessment of model performance. + +Table 1: Ablation experiments of the IHP, SNI and ESD-I/II on VOC $07 + 12$ + +
ModelsSizeParam.(M)FLOPs(G)AP50(%)AP(%)LatencyT4b=16(ms)
basielines
Yolov3-t [31]6408.7113.058.027.33.8
Yolov4-t [1]6405.9116.162.632.24.4
Yolov5-n [14]6401.794.274.047.03.4
Yolov6-n [16]6404.3011.180.258.83.5
Yolov7-t [35]6406.0613.378.454.65.0
Yolov8-n [15]6403.018.179.859.73.7
Part I: IHP (ours)
Yolov3-t-ihp6409.3315.059.9(↑1.9)29.1(↑1.8)3.8
Yolov4-t-ihp6406.6316.063.8(↑1.2)33.9(↑1.7)4.1
Yolov5-n-ihp6401.663.974.8(↑0.8)48.6(↑1.6)3.2
Yolov6-n-ihp6404.3710.679.4(↓0.8)57.4(↓1.4)3.7
Yolov7-t-ihp6406.7714.580.7(↑2.3)55.3(↑0.7)4.7
Yolov8-n-ihp6402.998.179.3(↓0.5)59.1(↓0.6)3.7
Part II: SNI (ours)
Yolov3-t-sni6408.7113.058.6(↑0.6)28.0(↑0.7)3.8
Yolov4-t-sni6405.9116.163.4(↑0.8)33.0(↑0.8)4.4
Yolov5-n-sni6401.794.274.5(↑0.5)47.7(↑0.7)3.4
Yolov6-n-sni6404.2811.081.9(↑1.7)60.8(↑2.0)3.4
Yolov7-t-sni6406.0613.381.5(↑3.1)57.9(↑3.3)5.0
Yolov8-n-sni6403.018.180.3(↑0.5)60.1(↑0.4)3.7
Part III: ESD-I (ours)
Yolov3-t-esd-I64011.822.469.2(↑11.2)38.5(↑11.2)4.6
Yolov4-t-esd-I6405.9116.163.1(↑0.7)32.9(↑0.7)4.4
Yolov5-n-esd-I6401.784.275.4(↑1.4)48.0(↑1.0)3.8
Yolov6-n-esd-I6404.3011.180.5(↑0.3)58.83.4
Yolov7-t-esd-I6406.8314.679.2(↑0.8)55.3(↑0.7)5.2
Yolov8-n-esd-I6403.228.180.1(↑0.3)59.8(↑0.1)3.8
ESD-II (ours)
Yolov3-t-esd-II6408.7814.863.0(↑5.0)32.1(↑4.8)4.9
Yolov4-t-esd-II6405.9316.663.9(↑1.5)33.5(↑1.3)4.7
Yolov5-n-esd-II6402.324.775.7(↑1.7)48.9(↑1.9)3.6
Yolov6-n-esd-II6404.3211.180.4(↑0.2)58.83.5
Yolov7-t-esd-II6407.6915.779.3(↑0.9)55.2(↑0.6)5.9
Yolov8-n-esd-II6403.188.580.4(↑0.6)59.9(↑0.2)4.0
+ +Table 2: Ablation experiments of the GSCnv and GSCnvE-I/II on VOC $07 + 12$ + +
ModelsSizeParam.(M)FLOPs(G)AP50(%)AP(%)LatencyT4b=16(ms)
benchlines(GSConv [18])
Yolov3-tiny-gsconv6404.618.369.941.13.9
Yolov4-tiny-gsconv6403.4212.473.445.24.6
Yolov5-n-gsconv6401.513.974.247.93.5
Yolov6-n-gsconv6404.2711.080.459.13.4
Yolov7-tiny-gsconv6405.0011.478.455.35.6
Yolov8-n-gsconv6402.747.880.961.23.7
GSCConvE-I (ours)
Yolov3-t-gse-I6407.8611.671.7(↑1.8)43.8(↑2.7)3.9
Yolov4-t-gse-I6404.8814.374.9(↑1.5)47.8(↑2.6)4.7
Yolov5-n-gse-I6401.764.275.8(↑1.6)50.1(↑2.2)3.5
Yolov6-n-gse-I6404.3011.180.1(↓0.3)58.8(↓0.3)3.6
Yolov7-t-gse-I6406.2113.579.1(↑0.7)56.4(↑1.1)5.8
Yolov8-n-gse-I6402.968.181.2(↑0.3)61.3(↑0.1)3.8
GSCConvE-II (ours)
Yolov3-t-gse-II6402.856.370.1(↑0.2)41.4(↑0.3)4.0
Yolov4-t-gse-II6402.3610.875.6(↑2.2)48.0(↑2.8)4.6
Yolov5-n-gse-II6401.483.976.3(↑2.1)51.4(↑3.5)3.5
Yolov6-n-gse-II6404.2811.180.3(↓0.1)58.8(↓0.3)3.5
Yolov7-t-gse-II6404.3310.779.1(↑0.7)56.2(↑0.9)5.8
Yolov8-n-gse-II6402.677.779.9(↓1.0)60.4(↓0.8)3.8
+ +Table 3: A contrast of the different confidence threshold for the SYOLO on COCO. + +
Confidence thresholdAP(%)APs(%)APm(%)APl(%)
0.001(valuation)53.135.558.769.9
0.25(application)48.7(↓4.4)29.7(↓5.8)54.5(↓4.2)66.5(↓3.4)
+ +# 4.3 Comparison experiments + +In Tab. 4 and Tab. 5, we compare state-of-the-art real-time detectors and SYOLO (architecture in Tab. 7) on the Pascal VOC and COCO benchmarks. In Tab. 6, we provide additional comparisons of state-of-the-art detectors on COCO [22], including both real-time (CNNs) and non-real-time (Transformers/CNNs) models. Parameters and FLOPs metrics significantly higher than those of real-time detector models are highlighted in red. + +In Tab. 4, SNI increased the AP of YOLOv7 by 0.9 percentage points without any additional computational or memory costs. SYOLO, utilizing the SA solution, achieved the highest accuracy among competitive models on the Pascal VOC benchmark. In Tab. 5, SYOLO also achieved the highest accuracy among competitive models on the MS COCO benchmark at input resolutions of $416 \times 416$ and $640 \times 640$ . Positive performance improvements across different benchmarks demonstrate the effectiveness of the SA solution for real-time object detection. + +Table 4: Comparasions of state-of-the-art real-time detectors on VOC $07 + 12$ + +
ModelsYearNeckSizeParam.(M)FLOPs(G)AP(%)LatencyT4b=16(ms)
Yolov3-spp [31]2018FPN64061.60154.960.539.4
Yolov4 [1]2020PANet†64052.57119.366.231.0
Yolov5 [14]2022PANet†64046.21108.063.930.1
YoloX [7]2021PANet†64053.79153.766.740.2
Yolov6 [16]2022PANet†64065.81159.169.229.5
Yolov7 [35]2022PANet†64036.58103.568.223.2
Yolov8 [15]2023PANet†64043.62164.969.331.6
SYolov7(ours)2023SA-PANet††64036.58103.569.123.2
SYolo(ours)2023SA64057.23142.469.629.3
+ +† Using simplified PANet: the red and green shortcuts are cancelled in the Fig. 1 (e). +$\dagger \dagger$ Using SNI replace the nearest neighbor interpolation to up-sampling in the PANet†. + +Table 5: Comparisons of state-of-the-art real-time detectors on COCO. + +
ModelsSizeParam.(M)FLOPs(G)AP(%)APs(%)APm(%)APl(%)LatencyT4b=16(ms)
YoloX-t [7]4165.16.532.8---4.3
SYolo-n(ours)4164.812.035.914.339.254.44.3
YoloX-t [7]6405.115.434.7---7.2
Yolov5-s [14]6407.216.537.4---9.8
Yolov6-n [16]6404.711.437.5---4.3
Yolov7-t [35]6406.213.838.718.842.451.95.3
Yolov8-n [15]6403.28.737.3---4.1
SYolo-s(ours)6406.615.740.821.145.356.95.8
+ +# 5 Conclusion + +We explore the nature of representation learning and identify that the fusion of different level feature maps faces the issue of feature misalignment. To address this, we introduce the SNI. Additionally, we propose the ESD to mitigate spatial information loss during downsampling and further optimize lightweight convolutional methods for computational efficiency in lightweight models. Based on these advancements, we propose the SA solution for real-time detection, where all tested detectors surpass their original performance and achieve state-of-the-art results. + +Object detection has significantly benefited from the FPN-like paradigm in past few years. However, recent developments indicate that newer technologies, including innovative architectures and training strategies, may offer superior performance for modern detectors, making the FPN-like paradigm potentially limiting. In many cases, the effectiveness of the FPN-like paradigm can be enhanced through the application of SNI to address feature misalignment issues. Fully addressing the challenge of misalignment during feature fusion in existing vision models requires further research efforts. + +Table 6: Comparisons of state-of-the-art detectors on COCO. + +
ModelsSizeParam.(M)FLOPs(G)AP(%)APs(%)APm(%)APl(%)LatencyV100b=1 (ms)
Transformers/CNNs (non-real-time, pre-trained)
FPN [21]-56.4145.836.817.538.747.8-
BiFPN [34]102421.055.049.3---42.8
ASFF [23]800--43.927.046.653.434.0
PANet [24]---51.032.653.964.8-
RFPN [28]---51.331.754.664.8-
DETR-Def. [40]-40.0173.043.826.447.158.0-
DETR-Dyn. [3]---47.228.649.359.1-
DETR-DND [17]133348.0265.048.6----
DINO [38]133347.0860.051.0----
Swin-T [26]-86.0745.050.5----
Swin-S [26]-107.0838.051.8----
Swin-B [26]1280145.0982.051.9----
CNNs (real-time, without pre-trained)
Yolov4 [1]60864.4142.843.526.746.753.3-
YoloX [7]64054.2155.649.7---14.5
PPYoloE [37]64052.2110.150.931.455.366.112.8
Yolov5 [14]64046.5109.149.0---10.1
Yolov6 [16]64058.5144.051.7----
Yolov7 [35]64036.9104.751.231.855.565.06.3
Yolov8 [15]64043.7165.252.9----
SYolo(ours)64057.2142.553.135.558.769.910.3
+ +Table 7: Architecture of the SYOLO. The E-ELAN/C2f is from the YOLOv7/8. + +
StageBlock-1/s/n‡Channels-1/s/n‡K. sizeStride
Backbone
P1Conv64 /16 /1632
P2ESD-I/II128 /32 /323, 42
C2f ×3/2/1128 /32 /323, 11
P3ESD-I/II256 /64 /643, 42
C2f ×6/3/1256 /64 /643, 11
P4ESD-I/II512 /128 /1283, 42
C2f ×6/3/1512 /128 /1283, 11
P5ESD-I/II1024 /256 /2563, 42
C2f ×3/2/11024 /256 /2563, 11
SPP ×11024 /256 /256--
Neck(SA)
P5-P4SNI1536 /384 /284--
E-ELAN†×12048 /512 /5123, 11
P4-P3SNI512 /192 /192--
E-ELAN†×2/2/1512 /256 /2563, 11
P3-P4ESD-I/II256 /128 /1283, 42
E-ELAN†×2/2/11024 /512 /5123, 11
P4-P5ESD-I/II512 /256 /2563, 42
E-ELAN†×2/2/12048 /1024 /10243, 11
+ +The first $3 \times 3$ vanilla convolutional layer of the E-ELAN is replaced by the GSCConvE-I, and the end pointwise convolutional layer of the E-ELAN is replaced by the C2f. +$\ddagger$ The 'l/s/n' means the model scale of the large/samll/nano. + +# Acknowledgements + +This work is supported by National Natural Science Foundation of China (Grant No. 52172381), Graduate Research Innovation Foundation of Chongqing Jiaotong University (Grant No. CYB240259). The author would also like to thank Prof. Dr. Jun Li, Prof. Dr. Qiliang Ren, and Ms. Xinxin Liu (Intelligent Transportation Big Data Center, School of Information Science and Engineering, CQJTU) for providing GPUs. + +# References + +1. Bochkovskiy, A., Wang, C.Y., Liao, H.Y.M.: Yolov4: Optimal speed and accuracy of object detection. arXiv preprint arXiv:2004.10934 (2020) +2. Chollet, F.: Xception: Deep learning with depthwise separable convolutions. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1251-1258 (2017) +3. Dai, X., Chen, Y., Yang, J., Zhang, P., Yuan, L., Zhang, L.: Dynamic detr: End-to-end object detection with dynamic attention. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 2988-2997 (2021) +4. Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: Imagenet: A large-scale hierarchical image database. In: 2009 IEEE conference on computer vision and pattern recognition. pp. 248-255. IEEE (2009) +5. Everingham, M., Eslami, S.A., Van Gool, L., Williams, C.K., Winn, J., Zisserman, A.: The Pascal visual object classes challenge: A retrospective. International journal of computer vision 111, 98-136 (2015) +6. Fu, C.Y., Liu, W., Ranga, A., Tyagi, A., Berg, A.C.: Dssd: Deconvolutional single shot detector. arXiv preprint arXiv:1701.06659 (2017) +7. Ge, Z., Liu, S., Wang, F., Li, Z., Sun, J.: Yolox: Exceeding yolo series in 2021. arXiv preprint arXiv:2107.08430 (2021) +8. Girshick, R.: Fast r-cnn. In: Proceedings of the IEEE international conference on computer vision. pp. 1440-1448 (2015) +9. Girshick, R., Donahue, J., Darrell, T., Malik, J.: Rich feature hierarchies for accurate object detection and semantic segmentation. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 580-587 (2014) +10. Han, K., Wang, Y., Tian, Q., Guo, J., Xu, C., Xu, C.: Ghostnet: More features from cheap operations. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 1580-1589 (2020) +1. He, K., Zhang, X., Ren, S., Sun, J.: Spatial pyramid pooling in deep convolutional networks for visual recognition. IEEE transactions on pattern analysis and machine intelligence 37(9), 1904-1916 (2015) +2. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 770-778 (2016) +3. Howard, A.G., Zhu, M., Chen, B., Kalenichenko, D., Wang, W., Weyand, T., Andreetto, M., Adam, H.: Mobilenets: Efficient convolutional neural networks for mobile vision applications. arXiv preprint arXiv:1704.04861 (2017) +4. Jocher, G., Chaurasia, A., Stoken, A., Borovec, J., Kwon, Y., Fang, J., Michael, K., Montes, D., Nadar, J., Skalski, P., et al.: ultralytics/yolov5: v6. 1-tensorrt, tensorflow edge tpu and openvino export and inference. Zenodo (2022) + +15. Jocher, G., et al.: Yolo-ultralytics. https://github.com/ultralytics/ultralytics (2023) +16. Li, C., Li, L., Jiang, H., Weng, K., Geng, Y., Li, L., Ke, Z., Li, Q., Cheng, M., Nie, W., et al.: Yolov6: A single-stage object detection framework for industrial applications. arXiv preprint arXiv:2209.02976 (2022) +17. Li, F., Zhang, H., Liu, S., Guo, J., Ni, L.M., Zhang, L.: Dn-detr: Accelerate detr training by introducing query denoising. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 13619-13627 (2022) +18. Li, H., Li, J., Wei, H., Liu, Z., Zhan, Z., Ren, Q.: Slim-neck by gsconv: a lightweight-design for real-time detector architectures. Journal of Real-Time Image Processing 21(3), 62 (2024) +19. Li, Y., Chen, Y., Wang, N., Zhang, Z.: Scale-aware trident networks for object detection. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 6054-6063 (2019) +20. Li, Z., Yang, L., Zhou, F.: Fssd: feature fusion single shot multibox detector. arXiv preprint arXiv:1712.00960 (2017) +21. Lin, T.Y., Dollar, P., Girshick, R., He, K., Hariharan, B., Belongie, S.: Feature pyramid networks for object detection. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 2117-2125 (2017) +22. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dólar, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13. pp. 740-755. Springer (2014) +23. Liu, S., Huang, D., Wang, Y.: Learning spatial fusion for single-shot object detection. arxiv 2019. arXiv preprint arXiv:1911.09516 (1911) +24. Liu, S., Qi, L., Qin, H., Shi, J., Jia, J.: Path aggregation network for instance segmentation. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 8759-8768 (2018) +25. Liu, W., Anguelov, D., Erhan, D., Szegedy, C., Reed, S., Fu, C.Y., Berg, A.C.: Ssd: Single shot multibox detector. In: Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part I 14. pp. 21-37. Springer (2016) +26. Liu, Z., Lin, Y., Cao, Y., Hu, H., Wei, Y., Zhang, Z., Lin, S., Guo, B.: Swin transformer: Hierarchical vision transformer using shifted windows. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 10012-10022 (2021) +27. Liu, Z., Mao, H., Wu, C.Y., Feichtenhofer, C., Darrell, T., Xie, S.: A convnet for the 2020s. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 11976-11986 (2022) +28. Qiao, S., Chen, L.C., Yuille, A.: Detectors: Detecting objects with recursive feature pyramid and switchable atrous convolution. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10213-10224 (2021) +29. Redmon, J., Divvala, S., Girshick, R., Farhadi, A.: You only look once: Unified, real-time object detection. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 779-788 (2016) +30. Redmon, J., Farhadi, A.: Yolo9000: better, faster, stronger. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 7263-7271 (2017) +31. Redmon, J., Farhadi, A.: Yolov3: An incremental improvement. arXiv preprint arXiv:1804.02767 (2018) +32. Simonyan, K., Zisserman, A.: Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556 (2014) + +33. Singh, B., Davis, L.S.: An analysis of scale invariance in object detection snip. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 3578-3587 (2018) +34. Tan, M., Pang, R., Le, Q.V.: Efficientdet: Scalable and efficient object detection. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10781-10790 (2020) +35. Wang, C.Y., Bochkovskiy, A., Liao, H.Y.M.: Yolov7: Trainable bag-of-freebies sets new state-of-the-art for real-time object detectors. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 7464-7475 (2023) +36. Wu, Y., Chen, Y., Yuan, L., Liu, Z., Wang, L., Li, H., Fu, Y.: Rethinking classification and localization for object detection. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10186-10195 (2020) +37. Yang, Z., Zhou, Y., Chen, Z., Ngiam, J.: 3d-man: 3d multi-frame attention network for object detection. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 1863-1872 (2021) +38. Zhang, H., Li, F., Liu, S., Zhang, L., Su, H., Zhu, J., Ni, L.M., Shum, H.Y.: Dino: Detr with improved denoising anchor boxes for end-to-end object detection. arXiv preprint arXiv:2203.03605 (2022) +39. Zhang, X., Zhou, X., Lin, M., Sun, J.: Shufflenet: An extremely efficient convolutional neural network for mobile devices. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 6848-6856 (2018) +40. Zhu, X., Su, W., Lu, L., Li, B., Wang, X., Dai, J.: Deformable detr: Deformable transformers for end-to-end object detection. arXiv preprint arXiv:2010.04159 (2020) \ No newline at end of file diff --git a/2024/Rethinking Features-Fused-Pyramid-Neck for Object Detection/images.zip b/2024/Rethinking Features-Fused-Pyramid-Neck for Object Detection/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..19b6d78b0e4e8655b7e0cea0e2b5bb384e2e40bc --- /dev/null +++ b/2024/Rethinking Features-Fused-Pyramid-Neck for Object Detection/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3100f0d5e03c497ca6690f59db71c519e826de09f0941ee1a0179ac42e239d4a +size 732630 diff --git a/2024/Rethinking Features-Fused-Pyramid-Neck for Object Detection/layout.json b/2024/Rethinking Features-Fused-Pyramid-Neck for Object Detection/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..f3e69554f9a9d7a2d4b653a52b816450a1b7c38d --- /dev/null +++ b/2024/Rethinking Features-Fused-Pyramid-Neck for Object Detection/layout.json @@ -0,0 +1,6380 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 149, + 110, + 465, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 110, + 465, + 148 + ], + "spans": [ + { + "bbox": [ + 149, + 110, + 465, + 148 + ], + "type": "text", + "content": "Rethinking Features-Fused-Pyramid-Neck for Object Detection" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 284, + 168, + 330, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 284, + 168, + 330, + 179 + ], + "spans": [ + { + "bbox": [ + 284, + 168, + 330, + 179 + ], + "type": "text", + "content": "Hulin Li" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 141, + 190, + 473, + 201 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 190, + 473, + 201 + ], + "spans": [ + { + "bbox": [ + 141, + 190, + 473, + 201 + ], + "type": "text", + "content": "College of Traffic and Transportation, Chongqing Jiaotong University, Chongqing" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 277, + 202, + 337, + 212 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 277, + 202, + 337, + 212 + ], + "spans": [ + { + "bbox": [ + 277, + 202, + 337, + 212 + ], + "type": "text", + "content": "400074, China" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 252, + 213, + 362, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 252, + 213, + 362, + 224 + ], + "spans": [ + { + "bbox": [ + 252, + 213, + 362, + 224 + ], + "type": "text", + "content": "alan@mails.cqjtu.edu.cn" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 160, + 254, + 465, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 254, + 465, + 430 + ], + "spans": [ + { + "bbox": [ + 160, + 254, + 465, + 430 + ], + "type": "text", + "content": "Abstract. Multi-head detectors typically employ a features-fused-pyramid-neck for multi-scale detection and are widely adopted in the industry. However, this approach faces feature misalignment when representations from different hierarchical levels of the feature pyramid are forcibly fused point-to-point. To address this issue, we designed an independent hierarchy pyramid (IHP) architecture to evaluate the effectiveness of the features-unfused-pyramid-neck for multi-head detectors. Subsequently, we introduced soft nearest neighbor interpolation (SNI) with a weight-downscaling factor to mitigate the impact of feature fusion at different hierarchies while preserving key textures. Furthermore, we present a feature adaptive selection method for downsampling in extended spatial windows (ESD) to retain spatial features and enhance lightweight convolutional techniques (GSCnvE). These advancements culminate in our secondary features alignment solution (SA) for real-time detection, achieving state-of-the-art results on Pascal VOC and MS COCO. Code will be released at https://github.com/AlanLi1997/rethinking-fpn." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 160, + 441, + 453, + 452 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 441, + 453, + 452 + ], + "spans": [ + { + "bbox": [ + 160, + 441, + 453, + 452 + ], + "type": "text", + "content": "Keywords: Object detection " + }, + { + "bbox": [ + 160, + 441, + 453, + 452 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 441, + 453, + 452 + ], + "type": "text", + "content": " Feature pyramid " + }, + { + "bbox": [ + 160, + 441, + 453, + 452 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 441, + 453, + 452 + ], + "type": "text", + "content": " Feature misalignment" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 162, + 453, + 275, + 461 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 162, + 453, + 275, + 461 + ], + "spans": [ + { + "bbox": [ + 162, + 453, + 275, + 461 + ], + "type": "text", + "content": "- Detection-net architecture" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 133, + 484, + 230, + 496 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 484, + 230, + 496 + ], + "spans": [ + { + "bbox": [ + 133, + 484, + 230, + 496 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "content": "Bounding box regression for object detection and pixel-level boundary regression for instance segmentation are pivotal and challenging visual tasks. Unlike basic image classification [4], detection or segmentation tasks often involve scenes with multiple objects of varying scales. Early deep learning-based object detection models began exploring real-time multi-scale object detection. RCNN [9] employed selective search to extract a large number of region proposals (" + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\sim 2\\mathrm{k}" + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "content": "). Although the vast number of image patches (region proposals) might aid multiscale detection, in practice, most scenes do not contain such a high number of objects to recognize. This approach was later deemed inefficient [8], as the proposed regions were derived from the image rather than features, necessitating a convolutional neural network (CNN) to extract features and predict classes for each proposal. A simplified RCNN architecture, as depicted in Fig. 1 (a), converts images of different scales into feature maps for object detection, largely" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 169, + 115, + 294, + 208 + ], + "blocks": [ + { + "bbox": [ + 169, + 115, + 294, + 208 + ], + "lines": [ + { + "bbox": [ + 169, + 115, + 294, + 208 + ], + "spans": [ + { + "bbox": [ + 169, + 115, + 294, + 208 + ], + "type": "image", + "image_path": "4c00a44d7d5e66f966c965977af8b1a4e6eca56737b7313dc346c4f8d83f0de6.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 299, + 115, + 364, + 207 + ], + "blocks": [ + { + "bbox": [ + 299, + 115, + 364, + 207 + ], + "lines": [ + { + "bbox": [ + 299, + 115, + 364, + 207 + ], + "spans": [ + { + "bbox": [ + 299, + 115, + 364, + 207 + ], + "type": "image", + "image_path": "46e922b9812dc914407c5906f7d4679413f46f612b34e326933f9c8c30f618eb.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 372, + 116, + 441, + 207 + ], + "blocks": [ + { + "bbox": [ + 372, + 116, + 441, + 207 + ], + "lines": [ + { + "bbox": [ + 372, + 116, + 441, + 207 + ], + "spans": [ + { + "bbox": [ + 372, + 116, + 441, + 207 + ], + "type": "image", + "image_path": "07fc4afd6132de0f0819f2796adf83869bf34853c4df2a337a1583fca55c6a96.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 169, + 209, + 344, + 297 + ], + "blocks": [ + { + "bbox": [ + 169, + 209, + 344, + 297 + ], + "lines": [ + { + "bbox": [ + 169, + 209, + 344, + 297 + ], + "spans": [ + { + "bbox": [ + 169, + 209, + 344, + 297 + ], + "type": "image", + "image_path": "e235f6347163f4ca2ab3ad192eab4fa8ba580dd26370239588b1662c5193bf38.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 350, + 208, + 444, + 297 + ], + "blocks": [ + { + "bbox": [ + 350, + 208, + 444, + 297 + ], + "lines": [ + { + "bbox": [ + 350, + 208, + 444, + 297 + ], + "spans": [ + { + "bbox": [ + 350, + 208, + 444, + 297 + ], + "type": "image", + "image_path": "6eea46c13ef56cb9ed5041b7b18f38bb731dbbbb9057366f8718144a1063fb3e.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 169, + 298, + 352, + 388 + ], + "blocks": [ + { + "bbox": [ + 169, + 298, + 352, + 388 + ], + "lines": [ + { + "bbox": [ + 169, + 298, + 352, + 388 + ], + "spans": [ + { + "bbox": [ + 169, + 298, + 352, + 388 + ], + "type": "image", + "image_path": "1c5625fcb681996bf3c0a52916ed6016f3634610c6112e793b0886b8225da4ef.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 408, + 482, + 496 + ], + "lines": [ + { + "bbox": [ + 130, + 408, + 482, + 496 + ], + "spans": [ + { + "bbox": [ + 130, + 408, + 482, + 496 + ], + "type": "text", + "content": "Fig. 1: The IHP and seven other typical neck architectures. The three heat maps are come from the third (P3), fourth (P4), and fifth (P5) hierarchies of the FPN-neck model, respectively. The area of interest for each hierarchies' features is highlighted in red; the darker the color, the higher the feature weight. It is evident that lower-level hierarchies prioritize local features, while higher-level hierarchies focus more on global features. In other words, there is a representation bias at different hierarchical levels. If these features are combined by point-to-point fusion without first addressing the misalignment issue, it may introduce noise rather than enrich semantics." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 360, + 298, + 441, + 388 + ], + "blocks": [ + { + "bbox": [ + 360, + 298, + 441, + 388 + ], + "lines": [ + { + "bbox": [ + 360, + 298, + 441, + 388 + ], + "spans": [ + { + "bbox": [ + 360, + 298, + 441, + 388 + ], + "type": "image", + "image_path": "c286cbcb9e9125c4ee6a164748652395c728ef953caddb8cfad2f13b9420d70b.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "content": "neglecting the real-time detection capability of models. YOLO [29] introduced the first real-time deep learning-based object detection model, which did not specifically focus on multi-scale detection, with a simplified structure shown in Fig. 1 (b). SSD [25] effectively combined the strengths of RCNN and YOLO, emphasizing the concept of multi-scale detection. It was the first model to achieve real-time multi-scale object detection using multi-scale features, with an architecture similar to Fig. 1 (c). SSD efficiently utilized feature maps of different scales from the VGG-16 backbone [32] to predict objects of varying scales. The exploration by early deep learning-based object detection models provided an important insight to the field: the efficient utilization of multi-scale features is crucial for enhancing the overall performance (accuracy-speed trade-off) of detectors. Consequently, FPN [21] introduced the feature pyramid-based multi-scale" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 202, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 202, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 202, + 101 + ], + "type": "text", + "content": "Hulin Li" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 390 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 481, + 390 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 481, + 390 + ], + "type": "text", + "content": "feature fusion concept, which gradually became a structural paradigm for detectors. This led to numerous representative works enhancing multi-scale feature fusion, such as PANet [24], ASFF [23], BiFPN [34], and RFPN [28]. The performance of real-time detectors has improved with the implementation of feature fusion strategies. However, further advancements are hindered by the increasing complexity of these techniques. Before the introduction of FPN, object detectors primarily relied on structures without fusion [9,25,29], as depicted in Fig. 1 (a)-(c). With the advent of FPN, these detectors have shifted towards FPN-based paradigms or more intricate improvement methods [8,20,29-31], as illustrated in Fig. 1 (e)-(g). It is important to note that the FPN-like strategy should not be considered an obligatory component for all detectors. Representation learning suggests that models ultimately require refined representations rather than an abundance of complex representations. Under the FPN-like paradigm, feature maps of different hierarchies and different resolutions are combined at the element level (point-to-point), necessitating the expansion of low-resolution, high-level feature maps to match the resolution of low-level feature maps. However, there are inherent representation biases in feature maps at different hierarchical levels: low-level feature maps tend to retain representations of small objects (local features in small receptive fields), while high-level feature maps favor representations of large objects (global features in large receptive fields), as shown in Fig. 1 (i). The direct element-level fusion of these feature maps with representation bias may lead to partial destruction of representations and exacerbate the feature misalignment problem during fusion." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 396, + 481, + 609 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 396, + 481, + 609 + ], + "spans": [ + { + "bbox": [ + 133, + 396, + 481, + 609 + ], + "type": "text", + "content": "In visual models, the resolution of feature maps decreases progressively from low to high hierarchical levels of the feature pyramid, achieved through step-by-step downsampling. Typically, downsampling reduces the resolution (width " + }, + { + "bbox": [ + 133, + 396, + 481, + 609 + ], + "type": "inline_equation", + "content": "\\times" + }, + { + "bbox": [ + 133, + 396, + 481, + 609 + ], + "type": "text", + "content": " height) of feature maps at a square root rate, inherently filtering out some spatial information regardless of whether the number of features (channels) remains constant or increases. It is also noteworthy that the transition of feature information from spatial to channel dimensions primarily occurs during downsampling, underscoring the importance of this stage. Traditional downsampling methods operate in a single mode, such as convolution with a stride of 2 (learnable) or average pooling and max pooling with a stride of 2 (non-learnable). These single-mode approaches have persisted since the introduction of ResNet [12] without significant evolution, which merits further investigation. Additionally, the initial purpose of FPN was not only to demonstrate the advantages of feature fusion in addressing multi-scale detection challenges but also to appropriately reduce model complexity. In industries with limited hardware resources, complex models are often impractical [2] [10]. Therefore, we enhance the lightweight convolution technique of the GSCov [18] to achieve a preferable trade-off between accuracy and speed for real-time detection models." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 147, + 614, + 441, + 625 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 147, + 614, + 441, + 625 + ], + "spans": [ + { + "bbox": [ + 147, + 614, + 441, + 625 + ], + "type": "text", + "content": "The main contributions of this work can be summarized as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 629, + 480, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 629, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 629, + 480, + 665 + ], + "type": "text", + "content": "1). We rethink the effectiveness of FPN-like paradigms for modern real-time detectors based on representation learning and identify issues with feature misalignment during element-level fusion." + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 190, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 190, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 190, + 91, + 447, + 102 + ], + "type": "text", + "content": "Rethinking Features-Fused-Pyramid-Neck for Object Detection" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 480, + 298 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 150 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 150 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 150 + ], + "type": "text", + "content": "2). We design an independent hierarchy pyramid architecture (IHP) without feature fusion in the neck to validate our findings. And the IHP achieves advanced results." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 152, + 480, + 176 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 152, + 480, + 176 + ], + "spans": [ + { + "bbox": [ + 130, + 152, + 480, + 176 + ], + "type": "text", + "content": "3). We introduce soft nearest neighbor interpolation (SNI) for up-sampling to mitigate feature misalignment during fusion." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 178, + 479, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 178, + 479, + 224 + ], + "spans": [ + { + "bbox": [ + 130, + 178, + 479, + 224 + ], + "type": "text", + "content": "4). We provide a feature adaptive selection method in extended spatial windows for downsampling (ESD) to enhance spatial feature retention during the downsampling stage. This method is plug-and-play and optimizes performance at a low cost." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 227, + 479, + 261 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 227, + 479, + 261 + ], + "spans": [ + { + "bbox": [ + 130, + 227, + 479, + 261 + ], + "type": "text", + "content": "5). We simplify and enhance the GSCnv lightweight convolution (GSCnvE) to offer more cost-effective options for models operating on edge devices with resource constraints." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 263, + 479, + 298 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 263, + 479, + 298 + ], + "spans": [ + { + "bbox": [ + 130, + 263, + 479, + 298 + ], + "type": "text", + "content": "6).We validate our approaches on Pascal VOC and COCO and present the SA solution. Existing lightweight real-time detection models can be optimized with the SA solution without bells and whistles." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 132, + 323, + 237, + 335 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 323, + 237, + 335 + ], + "spans": [ + { + "bbox": [ + 132, + 323, + 237, + 335 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 353, + 365, + 364 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 353, + 365, + 364 + ], + "spans": [ + { + "bbox": [ + 132, + 353, + 365, + 364 + ], + "type": "text", + "content": "2.1 Multi-head detection and features fusion" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 379, + 482, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 379, + 482, + 668 + ], + "spans": [ + { + "bbox": [ + 130, + 379, + 482, + 668 + ], + "type": "text", + "content": "The first-generation general detection models using deep learning consist of two main components: the backbone and the detecting-head [9, 29]. These models typically use final feature maps for prediction. SSD [25] introduced the use of multi-level feature maps for object detection, marking the beginning of multi-head detection and influencing subsequent real-time detector designs [31]. In the SSD model, each detecting-head directly uses raw feature maps from different levels of the backbone without any fusion. The key advantage of the multi-head detection method is the ability to predict objects of different scales using feature maps with varying receptive fields. High-level feature maps, with their wide global receptive fields, are beneficial for identifying large objects, while low-level feature maps, with refined local receptive fields, are helpful for identifying small objects. FPN introduced a fusion scheme for multi-level feature maps to enhance detection accuracy, setting the trend for feature fusion and becoming a common practice. This led to the development of Backbone-Neck-Head architectures. More complex variants based on FPN, such as BiFPN [34] and PANet [24], have since been proposed, enabling the use of more features for prediction. However, as prediction accuracy improved, networks became increasingly complex due to additional fusion layers. Some approaches have attempted to bypass feature fusion schemes by exploring new perspectives on multi-scale detection challenges, such as optimizing multi-scale training strategies [33] or using dilated convolutions to capture multi-scale receptive fields [19]. But these methods have gradually lost their competitiveness with the continuous optimization of classification-backbone feature extraction capabilities and the ongoing development of feature fusion techniques [1]." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 202, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 202, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 202, + 101 + ], + "type": "text", + "content": "Hulin Li" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 220, + 128 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 220, + 128 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 220, + 128 + ], + "type": "text", + "content": "2.2 Lightweight" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 137, + 482, + 269 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 137, + 482, + 269 + ], + "spans": [ + { + "bbox": [ + 130, + 137, + 482, + 269 + ], + "type": "text", + "content": "The choice between cloud computing or edge computing for a model depends on its number of parameters and floating-point operations (FLOPs). Therefore, reducing the number of parameters or FLOPs is a primary focus in lightweight model studies. Direct lightweight approaches include reducing the depth (number of layers) or width (number of neurons/filters) of the model, as seen in models likeYOLO-fast/tiny/nano [1,14-16,31,35]. Additionally, sparse computing techniques, such as depth-wise separable convolution used in MobileNets [13] and ShuffleNet [39], can also be employed. However, low-depth networks often suffer from underfitting due to insufficient nonlinear representation capabilities. Therefore, reducing the number of network layers may not always be a cost-effective way to lighten the model." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 288, + 388, + 303 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 288, + 388, + 303 + ], + "spans": [ + { + "bbox": [ + 132, + 288, + 388, + 303 + ], + "type": "text", + "content": "3 Secondary Features Alignment Solution" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 314, + 482, + 387 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 314, + 482, + 387 + ], + "spans": [ + { + "bbox": [ + 130, + 314, + 482, + 387 + ], + "type": "text", + "content": "Key contents of this work are described in detail in this section. Specifically, there are the IHP structure to directly demonstrate features unalignment problem by abandoning fusion, the SNI to alleviate unaligned during features fusion by point-to-point, the ESD to enhance spatial-features capture during the downsampling stage, and the GSConvE to improve the performance between the accuracy and speed of lightweight models." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 405, + 384, + 418 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 405, + 384, + 418 + ], + "spans": [ + { + "bbox": [ + 132, + 405, + 384, + 418 + ], + "type": "text", + "content": "3.1 Independent hierarchy pyramid architecture" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 426, + 482, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 426, + 482, + 628 + ], + "spans": [ + { + "bbox": [ + 130, + 426, + 482, + 628 + ], + "type": "text", + "content": "The introduction of FPN brought all real-time detectors into the realm of feature fusion. However, we have identified a significant issue arising from the point-to-point fusion of different level features - partial local features become unaligned. This fusion process is akin to adding noise when unaligned features are combined, as features unrelated to the target space are forcibly integrated, resulting in spatial disarray. For instance, as depicted in Fig. 2, adding the head-feature of a puppy to the butt creates an undefined breed when fed to the detection-head. This scenario is particularly evident when the puppy occupies fewer pixels. Typically, nearest neighbor interpolation is used to rapidly increase the resolution of high-level feature maps to match that of low-level feature maps, followed by fusion through methods such as element-wise addition, channel concatenation, or weighted sum. However, downsampling tends to be nonlinear and irreversible, leading to inconsistencies in spatial features among the additional features generated by up-sampling. Although BiFPN [34] offers a learnable fusion method, it has not been widely adopted by recent state-of-the-art real-time detectors, such as YOLOv7 [35] or YOLOv8 [15], due to the increased complexity and lack of demonstrated comprehensive performance improvements in these models." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 630, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 630, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 630, + 481, + 665 + ], + "type": "text", + "content": "The IHP adopts a radical approach by abandoning all fusions in the neck, fundamentally circumventing the features misalignment problem and streamlining the structure. By leveraging the inherent benefits of the multi-head detection" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 190, + 91, + 447, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 190, + 91, + 447, + 103 + ], + "spans": [ + { + "bbox": [ + 190, + 91, + 447, + 103 + ], + "type": "text", + "content": "Rethinking Features-Fused-Pyramid-Neck for Object Detection" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 115, + 482, + 488 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 115, + 482, + 488 + ], + "spans": [ + { + "bbox": [ + 130, + 115, + 482, + 488 + ], + "type": "text", + "content": "mode, the IHP directly predicts objects of different sizes using feature maps of different levels. It is important to note that there are significant differences between IHP and the SSD architecture: SSD directly uses different level features from the backbone to detect objects but the IHP introduces a bottleneck convolution module before prediction to filter the features. The advantage of this approach is that it allows the model to independently learn features for a specific scale branch without affecting features at other levels. Directly stacking convolutional blocks at different levels of the backbone can lead to interference between different levels, which we believe might be the primary reason why using a very deep backbone in object detection models can result in performance degradation. For instance, DSSD [6] showed a performance drop when ResNet-101 was used as the backbone instead of VGG-16 in the SSD detector. In Part I of Tab. 1, all baselines of coupled-head YOLOs demonstrated improved results when using the IHP. Features fusion often overlooks the potential introduction of noise when augmenting semantic richness for prediction. While the IHP may appear to reduce semantic information, it effectively harnesses the advantages of multi-head detection to circumvent the misalignment problem. This strategic utilization of multi-head detection is the key factor behind the competitive results achieved by the IHP. Unfortunately, unlike localization-classification-coupled-head detectors [14,29-31,35], the IHP did not achieve the same positive reaction to localization-classification-decoupled-head detectors [15, 16]. Decoupled-head detectors address a key challenge faced by coupled-head detectors: the disparity in the interests of classification and localization tasks, both of which are often performed using the same layer for prediction [36]. This approach is akin to forcibly fusing misaligned features for prediction. Decoupled-head detectors overcome this limitation by independently predicting classification and localization tasks through separate branches. This enables them to learn beneficial representations from redundant features, which is not possible in coupled-head detectors. Consequently, while FPN strategies are still applicable to some extent in decoupled-head detectors, the features misalignment problem still needs to be addressed." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 503, + 338, + 516 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 503, + 338, + 516 + ], + "spans": [ + { + "bbox": [ + 132, + 503, + 338, + 516 + ], + "type": "text", + "content": "3.2 Soft nearest neighbor interpolation" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "content": "In detection or segmentation models, different-level feature fusion typically occurs after the up-sampling of feature maps, with nearest neighbor interpolation and transposed convolution being popular up-sampling methods. However, the resolution expansion resulting from up-sampling in detection or segmentation differs significantly from tasks such as generation or super-resolution. In detection, the expanded feature maps represent abstract high-level semantic features rather than raw image detail information. Transposed convolution introduces increased computation and latency. To address the features misalignment problem with minimal cost while maintaining the speed of nearest neighbor interpolation, we explore a solution. Nearest neighbor interpolation is akin to average unpooling but is considered a \"hard\" operation. By softening this operation, similar to the transformation of SoftMax to the Max function, we aim to mitigate the" + } + ] + } + ], + "index": 4 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 202, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 202, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 202, + 101 + ], + "type": "text", + "content": "Hulin Li" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 135, + 117, + 478, + 247 + ], + "blocks": [ + { + "bbox": [ + 135, + 117, + 478, + 247 + ], + "lines": [ + { + "bbox": [ + 135, + 117, + 478, + 247 + ], + "spans": [ + { + "bbox": [ + 135, + 117, + 478, + 247 + ], + "type": "image", + "image_path": "2a45fad2bdaaec7b2d52335aa31b61e76229a697d5c86f5d4ceba1f5c5d3e169.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 265, + 480, + 287 + ], + "lines": [ + { + "bbox": [ + 132, + 265, + 480, + 287 + ], + "spans": [ + { + "bbox": [ + 132, + 265, + 480, + 287 + ], + "type": "text", + "content": "Fig. 2: Illustrations of the features misalignment in fusion and the SNI. These numbers, 22-88, are just markers of different local-features not real feature values." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 171, + 300, + 444, + 443 + ], + "blocks": [ + { + "bbox": [ + 171, + 300, + 444, + 443 + ], + "lines": [ + { + "bbox": [ + 171, + 300, + 444, + 443 + ], + "spans": [ + { + "bbox": [ + 171, + 300, + 444, + 443 + ], + "type": "image", + "image_path": "beda720ecd6fb3d8211cbb77d398261dbb1d209f5e82a3ddd847392ca09937e4.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 452, + 482, + 485 + ], + "lines": [ + { + "bbox": [ + 130, + 452, + 482, + 485 + ], + "spans": [ + { + "bbox": [ + 130, + 452, + 482, + 485 + ], + "type": "text", + "content": "Fig.3: A comparison of the SNI and traditional method. This result could be reproduced by using SNI source code (model:Yolov5n-panet). The same size scaling before and after up-sampling is for intuitive comparison." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 508, + 480, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 508, + 480, + 533 + ], + "spans": [ + { + "bbox": [ + 132, + 508, + 480, + 533 + ], + "type": "text", + "content": "\"hard\" features misalignment issue. To achieve this, we introduce a soft factor, denoted as " + }, + { + "bbox": [ + 132, + 508, + 480, + 533 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 132, + 508, + 480, + 533 + ], + "type": "text", + "content": ", for the nearest neighbor interpolation during up-sampling:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 233, + 539, + 481, + 563 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 233, + 539, + 481, + 563 + ], + "spans": [ + { + "bbox": [ + 233, + 539, + 481, + 563 + ], + "type": "interline_equation", + "content": "\\boldsymbol {Y} = \\alpha \\cdot \\boldsymbol {f} (\\boldsymbol {X}), \\alpha = \\frac {\\text {R e s o l u t i o n} \\boldsymbol {X}}{\\text {R e s o l u t i o n} \\boldsymbol {Y}} \\tag {1}", + "image_path": "7114f195c543707d6fe7cd986308d67dc3e60a5e92af76417db824c6a640bf70.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "type": "text", + "content": "In Eq. (1), the resolution of the high-level feature maps " + }, + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "type": "inline_equation", + "content": "(\\mathbf{X})" + }, + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "type": "text", + "content": " and low-level feature maps " + }, + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "type": "inline_equation", + "content": "(\\mathbf{Y})" + }, + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "type": "text", + "content": " are denoted as " + }, + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "type": "inline_equation", + "content": "Resolution_{X}" + }, + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "type": "inline_equation", + "content": "Resolution_{Y}" + }, + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "type": "text", + "content": ", respectively. The nearest neighbor interpolation operation, represented by " + }, + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\pmb{f}" + }, + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "type": "text", + "content": ", is illustrated in Fig. 2. The SNI adjusts the influence of high-level semantic features on low-level features based on the zoom factor for feature maps. Specifically, as the zoom factor increases, the impact of high-level semantic features on low-level features weakens. Unlike SoftMax, the SNI does not convert outputs into probabilities but rather mitigates misalignment without additional cost when integrating high-" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 190, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 190, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 190, + 91, + 448, + 102 + ], + "type": "text", + "content": "Rethinking Features-Fused-Pyramid-Neck for Object Detection" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 283 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 283 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 283 + ], + "type": "text", + "content": "level features with low-level features through point-to-point fusion. SNI softens feature weights while preserving key textures, thereby reducing the mutual influence of features from different hierarchical levels during fusion, as qualitatively demonstrated in Fig. 3. This approach achieves the secondary (auxiliary) features alignment (SA) and lowers the complexity for the model to learn from the fused representations. In contrast, traditional methods that directly merge features from different hierarchical levels can cause unstable competitive effects due to representation biases, increasing the learning difficulty for the model. For example, if the three heatmaps in Fig. 1 (i) are directly fused, should the red areas cover the entire image? Clearly, the answer is no. If such a situation occurs, it indicates a failure in prior feature extraction, as the objective of a model is to capture and focus on useful features. This is also why models under the FPN-like paradigm typically require several additional convolutional blocks to further filter features after fusing and before predicting." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 284, + 482, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 284, + 482, + 380 + ], + "spans": [ + { + "bbox": [ + 130, + 284, + 482, + 380 + ], + "type": "text", + "content": "In Part II of Tab. 1, all tested state-of-the-art models demonstrate accuracy improvements solely by incorporating the SNI, confirming the presence of features misalignment in FPN-based models. Notably, for YOLOv6 [16], a detector equipped with various techniques tailored for industry applications, the AP is enhanced by up to 2 percentage points just with the adoption of SNI, while reducing computational cost and inference latency. We achieved significant optimization results in the original YOLOv6 by simply replacing transposed convolution with SNI, highlighting the efficacy of this approach." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 396, + 437, + 408 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 396, + 437, + 408 + ], + "spans": [ + { + "bbox": [ + 131, + 396, + 437, + 408 + ], + "type": "text", + "content": "3.3 Feature adaptive selection in extended spatial windows" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 136, + 437, + 478, + 517 + ], + "blocks": [ + { + "bbox": [ + 136, + 437, + 478, + 517 + ], + "lines": [ + { + "bbox": [ + 136, + 437, + 478, + 517 + ], + "spans": [ + { + "bbox": [ + 136, + 437, + 478, + 517 + ], + "type": "image", + "image_path": "28d65be42a1f275f2d41a54dffb09210d9bbcb1d78bf489f100e47d45969f9f7.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 540, + 482, + 605 + ], + "lines": [ + { + "bbox": [ + 130, + 540, + 482, + 605 + ], + "spans": [ + { + "bbox": [ + 130, + 540, + 482, + 605 + ], + "type": "text", + "content": "Fig. 4: Illustrations of the ESD-I and ESD-II. It must be mentioned that although the size of the extended window does not have to be same in the ESD-I, pooling is a non-learnable operation, so the features from an unequal window may become noise for other and bring the trouble of the features misalignment again. The ESD-II is free on size of the extended window because of the learnable fusion like the SPPNet [11] for the YOLOv3." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "text", + "content": "Feature downsampling is a critical technique in vision models based on deep learning, involving the progressive reduction of feature map resolution. In vision representation learning, models typically rely on low-resolution, high-level" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 202, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 202, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 202, + 101 + ], + "type": "text", + "content": "Hulin Li" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 403 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 403 + ], + "type": "text", + "content": "feature maps for prediction rather than directly processing high-resolution raw images. However, excessive downsampling can result in significant loss of spatial details [12]. Recent detectors typically limit the number of downsampling operations to no more than five times, leading to a spatial resolution reduction of " + }, + { + "bbox": [ + 130, + 116, + 482, + 403 + ], + "type": "inline_equation", + "content": "2^{5}" + }, + { + "bbox": [ + 130, + 116, + 482, + 403 + ], + "type": "text", + "content": " times, as demonstrated in YOLOs [1,7,14-16,31,35,37] and Swin Transformer [26]. Common downsampling methods include convolution and pooling, with models typically employing only one of these methods. We propose the feature adaptive selection method in extended spatial windows for downsampling (ESD) to enhance spatial information retention capabilities. ESD consists of two nonlinear branches and one linear branch. In the nonlinear branches, a norm convolutional layer and an extended window max-pooling layer enhance local feature capturing, while extended window average-pooling enhances global feature capturing. Subsequently, linear features are merged with nonlinear features. Two fusion modes, ESD-I and ESD-II, are designed for lightweight and norm models, respectively. The difference lies in the feature fusion stage, as illustrated in Fig. 4. ESD-I employs element-wise addition for merging features, offering low computational cost suitable for lightweight models, while ESD-II utilizes learnable linearly adaptive fusion, slightly increasing computational cost but suitable for norm models. Importantly, local and global features of the extended window are sampled using simple pooling techniques, preserving input information to a significant extent. This reduces information loss during downsampling and enables subsequent layers to learn partial representations from the previous layer, akin to indirectly realizing a hidden shortcut connection reminiscent of ResNet [12]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 420, + 334, + 432 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 420, + 334, + 432 + ], + "spans": [ + { + "bbox": [ + 132, + 420, + 334, + 432 + ], + "type": "text", + "content": "3.4 Lightweight GSConv enhancement" + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 153, + 461, + 461, + 550 + ], + "blocks": [ + { + "bbox": [ + 153, + 461, + 461, + 550 + ], + "lines": [ + { + "bbox": [ + 153, + 461, + 461, + 550 + ], + "spans": [ + { + "bbox": [ + 153, + 461, + 461, + 550 + ], + "type": "image", + "image_path": "159ea320bc49abbac2e9443c24897e0139bfdebb01c805a385d2a84699a4f3cb.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 211, + 571, + 400, + 582 + ], + "lines": [ + { + "bbox": [ + 211, + 571, + 400, + 582 + ], + "spans": [ + { + "bbox": [ + 211, + 571, + 400, + 582 + ], + "type": "text", + "content": "Fig. 5: Illustrations of the GSConvE-I and II." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": "GSCnv [18] was introduced to address the channel interaction limitations of depthwise separable convolution by blending features from both vanilla convolution and depthwise convolution. The original GSCnv consists of three fundamental components: a " + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "inline_equation", + "content": "3 \\times 3" + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": " vanilla convolution, a " + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "inline_equation", + "content": "5 \\times 5" + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": " depthwise convolution, and a feature shuffle operation. Building upon this, we propose two evolutions" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 190, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 190, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 190, + 91, + 447, + 102 + ], + "type": "text", + "content": "Rethinking Features-Fused-Pyramid-Neck for Object Detection" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 297 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 297 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 297 + ], + "type": "text", + "content": "of GSConv, namely GSConvE-I and GSConvE-II, tailored for norm models and lightweight models, respectively. GSConvE-I is designed for norm models, where intermediate feature maps on the auxiliary branch are derived from dense linear mappings, while output feature maps originate from sparse linear mappings. This approach maintains feature richness without significantly increasing computational cost. On the other hand, GSConvE-II is geared towards lightweight models. It incorporates three depthwise convolution auxiliary branches with kernel sizes of " + }, + { + "bbox": [ + 130, + 116, + 482, + 297 + ], + "type": "inline_equation", + "content": "9 \\times 9" + }, + { + "bbox": [ + 130, + 116, + 482, + 297 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 116, + 482, + 297 + ], + "type": "inline_equation", + "content": "13 \\times 13" + }, + { + "bbox": [ + 130, + 116, + 482, + 297 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 116, + 482, + 297 + ], + "type": "inline_equation", + "content": "17 \\times 17" + }, + { + "bbox": [ + 130, + 116, + 482, + 297 + ], + "type": "text", + "content": ", respectively. By utilizing larger kernel sizes, this variant can directly capture larger receptive fields and global features with fewer layer accumulations, offering high cost-effectiveness for lightweight models that prioritize reduced FLOPs and layer count. In GSConvE-I, we explore the removal of the batch normalization layer on the depthwise convolution branch while retaining the batch normalization layer on the main branch, similar to ConvNeXt [27]. This adjustment avoids the vanishing gradient problem and simplifies computation, leading to improved prediction accuracy." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 316, + 230, + 330 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 316, + 230, + 330 + ], + "spans": [ + { + "bbox": [ + 132, + 316, + 230, + 330 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 342, + 482, + 404 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 342, + 482, + 404 + ], + "spans": [ + { + "bbox": [ + 130, + 342, + 482, + 404 + ], + "type": "text", + "content": "Real-time detectors are highly favored in industry due to their hardware-friendly nature and the optimal balance they strike between accuracy and speed. In our study, we selected popular YOLO models—YOLOv3 [31], YOLOv4 [1], YOLOv5 [14], YOLOv6 [16], YOLOv7 [35], and YOLOv8 [15]—as baselines to evaluate the efficacy of the SA solution in real-time detection scenarios (SYOLO)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 422, + 205, + 434 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 422, + 205, + 434 + ], + "spans": [ + { + "bbox": [ + 132, + 422, + 205, + 434 + ], + "type": "text", + "content": "4.1 Datasets" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 444, + 482, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 444, + 482, + 552 + ], + "spans": [ + { + "bbox": [ + 130, + 444, + 482, + 552 + ], + "type": "text", + "content": "We conducted trials on the Pascal VOC dataset [5] and the MS COCO dataset [22]. The Pascal VOC 07+12 dataset was utilized for ablation experiments, wherein models underwent training on the train&val07+12 set and were subsequently evaluated on the test07 set. On the other hand, the MS COCO dataset was employed for comparative experiments, with models trained on the train2017 set and evaluated on the val2017 set. To ensure fairness and consistency, all released models were trained using the default set of hyperparameters and without utilizing pre-trained weights. More details of hyperparameters are in the support material." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 571, + 244, + 583 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 571, + 244, + 583 + ], + "spans": [ + { + "bbox": [ + 132, + 571, + 244, + 583 + ], + "type": "text", + "content": "4.2 Ablation studies" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 594, + 482, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 667 + ], + "type": "text", + "content": "In Tab. 1, we present the results of ablation experiments conducted on six baselines [1, 14-16, 31, 35], focusing on the impact of the IHP, SNI, ESD. The lightweight experiment results of GSConvE are reported in Tab. 2. All baselines retained their original neck architecture, as we believe the competitive neck has been carefully chosen in the original works. Performance metrics highlighted in green indicate performance gains, while those in red indicate losses. It is crucial" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 202, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 202, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 202, + 101 + ], + "type": "text", + "content": "Hulin Li" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 259 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 259 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 259 + ], + "type": "text", + "content": "to address an often-overlooked aspect during the verification process: the inconsistency between reported evaluation accuracy and actual usage. Frequently, detectors are evaluated with a very low confidence threshold, such as 0.001, which results in a large number of prediction boxes and inflated evaluation accuracy. However, in real-world detection scenarios, the confidence threshold is typically set much higher, around 0.25 or above. This filtering process ensures that higher-quality bounding boxes are utilized for practical purposes. The discrepancy between evaluation settings and real-world usage can create a misleading impression of accuracy. Therefore, it is essential to maintain consistency in configuration between evaluation and application. While this may result in a significant drop in evaluation accuracy, as shown in Tab. 3, it provides a more realistic assessment of model performance." + } + ] + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 150, + 316, + 462, + 665 + ], + "blocks": [ + { + "bbox": [ + 146, + 296, + 466, + 307 + ], + "lines": [ + { + "bbox": [ + 146, + 296, + 466, + 307 + ], + "spans": [ + { + "bbox": [ + 146, + 296, + 466, + 307 + ], + "type": "text", + "content": "Table 1: Ablation experiments of the IHP, SNI and ESD-I/II on VOC " + }, + { + "bbox": [ + 146, + 296, + 466, + 307 + ], + "type": "inline_equation", + "content": "07 + 12" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 150, + 316, + 462, + 665 + ], + "lines": [ + { + "bbox": [ + 150, + 316, + 462, + 665 + ], + "spans": [ + { + "bbox": [ + 150, + 316, + 462, + 665 + ], + "type": "table", + "html": "
ModelsSizeParam.(M)FLOPs(G)AP50(%)AP(%)LatencyT4b=16(ms)
basielines
Yolov3-t [31]6408.7113.058.027.33.8
Yolov4-t [1]6405.9116.162.632.24.4
Yolov5-n [14]6401.794.274.047.03.4
Yolov6-n [16]6404.3011.180.258.83.5
Yolov7-t [35]6406.0613.378.454.65.0
Yolov8-n [15]6403.018.179.859.73.7
Part I: IHP (ours)
Yolov3-t-ihp6409.3315.059.9(↑1.9)29.1(↑1.8)3.8
Yolov4-t-ihp6406.6316.063.8(↑1.2)33.9(↑1.7)4.1
Yolov5-n-ihp6401.663.974.8(↑0.8)48.6(↑1.6)3.2
Yolov6-n-ihp6404.3710.679.4(↓0.8)57.4(↓1.4)3.7
Yolov7-t-ihp6406.7714.580.7(↑2.3)55.3(↑0.7)4.7
Yolov8-n-ihp6402.998.179.3(↓0.5)59.1(↓0.6)3.7
Part II: SNI (ours)
Yolov3-t-sni6408.7113.058.6(↑0.6)28.0(↑0.7)3.8
Yolov4-t-sni6405.9116.163.4(↑0.8)33.0(↑0.8)4.4
Yolov5-n-sni6401.794.274.5(↑0.5)47.7(↑0.7)3.4
Yolov6-n-sni6404.2811.081.9(↑1.7)60.8(↑2.0)3.4
Yolov7-t-sni6406.0613.381.5(↑3.1)57.9(↑3.3)5.0
Yolov8-n-sni6403.018.180.3(↑0.5)60.1(↑0.4)3.7
Part III: ESD-I (ours)
Yolov3-t-esd-I64011.822.469.2(↑11.2)38.5(↑11.2)4.6
Yolov4-t-esd-I6405.9116.163.1(↑0.7)32.9(↑0.7)4.4
Yolov5-n-esd-I6401.784.275.4(↑1.4)48.0(↑1.0)3.8
Yolov6-n-esd-I6404.3011.180.5(↑0.3)58.83.4
Yolov7-t-esd-I6406.8314.679.2(↑0.8)55.3(↑0.7)5.2
Yolov8-n-esd-I6403.228.180.1(↑0.3)59.8(↑0.1)3.8
ESD-II (ours)
Yolov3-t-esd-II6408.7814.863.0(↑5.0)32.1(↑4.8)4.9
Yolov4-t-esd-II6405.9316.663.9(↑1.5)33.5(↑1.3)4.7
Yolov5-n-esd-II6402.324.775.7(↑1.7)48.9(↑1.9)3.6
Yolov6-n-esd-II6404.3211.180.4(↑0.2)58.83.5
Yolov7-t-esd-II6407.6915.779.3(↑0.9)55.2(↑0.6)5.9
Yolov8-n-esd-II6403.188.580.4(↑0.6)59.9(↑0.2)4.0
", + "image_path": "5ac1f3bd6c3ea3199e2473d5a03879c0878cb861663c4e4e0438079afce0242c.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 190, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 190, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 190, + 91, + 447, + 102 + ], + "type": "text", + "content": "Rethinking Features-Fused-Pyramid-Neck for Object Detection" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 149, + 136, + 465, + 350 + ], + "blocks": [ + { + "bbox": [ + 137, + 114, + 476, + 125 + ], + "lines": [ + { + "bbox": [ + 137, + 114, + 476, + 125 + ], + "spans": [ + { + "bbox": [ + 137, + 114, + 476, + 125 + ], + "type": "text", + "content": "Table 2: Ablation experiments of the GSCnv and GSCnvE-I/II on VOC " + }, + { + "bbox": [ + 137, + 114, + 476, + 125 + ], + "type": "inline_equation", + "content": "07 + 12" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 149, + 136, + 465, + 350 + ], + "lines": [ + { + "bbox": [ + 149, + 136, + 465, + 350 + ], + "spans": [ + { + "bbox": [ + 149, + 136, + 465, + 350 + ], + "type": "table", + "html": "
ModelsSizeParam.(M)FLOPs(G)AP50(%)AP(%)LatencyT4b=16(ms)
benchlines(GSConv [18])
Yolov3-tiny-gsconv6404.618.369.941.13.9
Yolov4-tiny-gsconv6403.4212.473.445.24.6
Yolov5-n-gsconv6401.513.974.247.93.5
Yolov6-n-gsconv6404.2711.080.459.13.4
Yolov7-tiny-gsconv6405.0011.478.455.35.6
Yolov8-n-gsconv6402.747.880.961.23.7
GSCConvE-I (ours)
Yolov3-t-gse-I6407.8611.671.7(↑1.8)43.8(↑2.7)3.9
Yolov4-t-gse-I6404.8814.374.9(↑1.5)47.8(↑2.6)4.7
Yolov5-n-gse-I6401.764.275.8(↑1.6)50.1(↑2.2)3.5
Yolov6-n-gse-I6404.3011.180.1(↓0.3)58.8(↓0.3)3.6
Yolov7-t-gse-I6406.2113.579.1(↑0.7)56.4(↑1.1)5.8
Yolov8-n-gse-I6402.968.181.2(↑0.3)61.3(↑0.1)3.8
GSCConvE-II (ours)
Yolov3-t-gse-II6402.856.370.1(↑0.2)41.4(↑0.3)4.0
Yolov4-t-gse-II6402.3610.875.6(↑2.2)48.0(↑2.8)4.6
Yolov5-n-gse-II6401.483.976.3(↑2.1)51.4(↑3.5)3.5
Yolov6-n-gse-II6404.2811.180.3(↓0.1)58.8(↓0.3)3.5
Yolov7-t-gse-II6404.3310.779.1(↑0.7)56.2(↑0.9)5.8
Yolov8-n-gse-II6402.677.779.9(↓1.0)60.4(↓0.8)3.8
", + "image_path": "5aabb6058a6856f5d42f9149588f4e3c6a681ea131913fbced11930e33fee3b3.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 171, + 389, + 443, + 426 + ], + "blocks": [ + { + "bbox": [ + 134, + 368, + 478, + 379 + ], + "lines": [ + { + "bbox": [ + 134, + 368, + 478, + 379 + ], + "spans": [ + { + "bbox": [ + 134, + 368, + 478, + 379 + ], + "type": "text", + "content": "Table 3: A contrast of the different confidence threshold for the SYOLO on COCO." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 171, + 389, + 443, + 426 + ], + "lines": [ + { + "bbox": [ + 171, + 389, + 443, + 426 + ], + "spans": [ + { + "bbox": [ + 171, + 389, + 443, + 426 + ], + "type": "table", + "html": "
Confidence thresholdAP(%)APs(%)APm(%)APl(%)
0.001(valuation)53.135.558.769.9
0.25(application)48.7(↓4.4)29.7(↓5.8)54.5(↓4.2)66.5(↓3.4)
", + "image_path": "2b55466113426c6d12a07456d436913f5e296143e276df80e8e9f99e2deb3cf7.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 456, + 292, + 468 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 456, + 292, + 468 + ], + "spans": [ + { + "bbox": [ + 132, + 456, + 292, + 468 + ], + "type": "text", + "content": "4.3 Comparison experiments" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 494, + 482, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 494, + 482, + 567 + ], + "spans": [ + { + "bbox": [ + 130, + 494, + 482, + 567 + ], + "type": "text", + "content": "In Tab. 4 and Tab. 5, we compare state-of-the-art real-time detectors and SYOLO (architecture in Tab. 7) on the Pascal VOC and COCO benchmarks. In Tab. 6, we provide additional comparisons of state-of-the-art detectors on COCO [22], including both real-time (CNNs) and non-real-time (Transformers/CNNs) models. Parameters and FLOPs metrics significantly higher than those of real-time detector models are highlighted in red." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": "In Tab. 4, SNI increased the AP of YOLOv7 by 0.9 percentage points without any additional computational or memory costs. SYOLO, utilizing the SA solution, achieved the highest accuracy among competitive models on the Pascal VOC benchmark. In Tab. 5, SYOLO also achieved the highest accuracy among competitive models on the MS COCO benchmark at input resolutions of " + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "inline_equation", + "content": "416 \\times 416" + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "inline_equation", + "content": "640 \\times 640" + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": ". Positive performance improvements across different benchmarks demonstrate the effectiveness of the SA solution for real-time object detection." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 202, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 202, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 202, + 101 + ], + "type": "text", + "content": "Hulin Li" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 134, + 136, + 479, + 244 + ], + "blocks": [ + { + "bbox": [ + 146, + 114, + 465, + 125 + ], + "lines": [ + { + "bbox": [ + 146, + 114, + 465, + 125 + ], + "spans": [ + { + "bbox": [ + 146, + 114, + 465, + 125 + ], + "type": "text", + "content": "Table 4: Comparasions of state-of-the-art real-time detectors on VOC " + }, + { + "bbox": [ + 146, + 114, + 465, + 125 + ], + "type": "inline_equation", + "content": "07 + 12" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 134, + 136, + 479, + 244 + ], + "lines": [ + { + "bbox": [ + 134, + 136, + 479, + 244 + ], + "spans": [ + { + "bbox": [ + 134, + 136, + 479, + 244 + ], + "type": "table", + "html": "
ModelsYearNeckSizeParam.(M)FLOPs(G)AP(%)LatencyT4b=16(ms)
Yolov3-spp [31]2018FPN64061.60154.960.539.4
Yolov4 [1]2020PANet†64052.57119.366.231.0
Yolov5 [14]2022PANet†64046.21108.063.930.1
YoloX [7]2021PANet†64053.79153.766.740.2
Yolov6 [16]2022PANet†64065.81159.169.229.5
Yolov7 [35]2022PANet†64036.58103.568.223.2
Yolov8 [15]2023PANet†64043.62164.969.331.6
SYolov7(ours)2023SA-PANet††64036.58103.569.123.2
SYolo(ours)2023SA64057.23142.469.629.3
", + "image_path": "7dfda00e73cf544fc0a169bf41d2845ef33e75a8b9699a6062cdd3f30ca35e66.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 141, + 245, + 441, + 255 + ], + "lines": [ + { + "bbox": [ + 141, + 245, + 441, + 255 + ], + "spans": [ + { + "bbox": [ + 141, + 245, + 441, + 255 + ], + "type": "text", + "content": "† Using simplified PANet: the red and green shortcuts are cancelled in the Fig. 1 (e)." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_footnote" + }, + { + "bbox": [ + 141, + 254, + 446, + 264 + ], + "lines": [ + { + "bbox": [ + 141, + 254, + 446, + 264 + ], + "spans": [ + { + "bbox": [ + 141, + 254, + 446, + 264 + ], + "type": "inline_equation", + "content": "\\dagger \\dagger" + }, + { + "bbox": [ + 141, + 254, + 446, + 264 + ], + "type": "text", + "content": " Using SNI replace the nearest neighbor interpolation to up-sampling in the PANet†." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 132, + 295, + 486, + 392 + ], + "blocks": [ + { + "bbox": [ + 158, + 274, + 454, + 285 + ], + "lines": [ + { + "bbox": [ + 158, + 274, + 454, + 285 + ], + "spans": [ + { + "bbox": [ + 158, + 274, + 454, + 285 + ], + "type": "text", + "content": "Table 5: Comparisons of state-of-the-art real-time detectors on COCO." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 132, + 295, + 486, + 392 + ], + "lines": [ + { + "bbox": [ + 132, + 295, + 486, + 392 + ], + "spans": [ + { + "bbox": [ + 132, + 295, + 486, + 392 + ], + "type": "table", + "html": "
ModelsSizeParam.(M)FLOPs(G)AP(%)APs(%)APm(%)APl(%)LatencyT4b=16(ms)
YoloX-t [7]4165.16.532.8---4.3
SYolo-n(ours)4164.812.035.914.339.254.44.3
YoloX-t [7]6405.115.434.7---7.2
Yolov5-s [14]6407.216.537.4---9.8
Yolov6-n [16]6404.711.437.5---4.3
Yolov7-t [35]6406.213.838.718.842.451.95.3
Yolov8-n [15]6403.28.737.3---4.1
SYolo-s(ours)6406.615.740.821.145.356.95.8
", + "image_path": "86578c8ed8667c3d48202f73bd760fe016b2fc7203d5b6e3ef9123bca52839d4.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 414, + 219, + 426 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 414, + 219, + 426 + ], + "spans": [ + { + "bbox": [ + 132, + 414, + 219, + 426 + ], + "type": "text", + "content": "5 Conclusion" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 438, + 482, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 438, + 482, + 533 + ], + "spans": [ + { + "bbox": [ + 130, + 438, + 482, + 533 + ], + "type": "text", + "content": "We explore the nature of representation learning and identify that the fusion of different level feature maps faces the issue of feature misalignment. To address this, we introduce the SNI. Additionally, we propose the ESD to mitigate spatial information loss during downsampling and further optimize lightweight convolutional methods for computational efficiency in lightweight models. Based on these advancements, we propose the SA solution for real-time detection, where all tested detectors surpass their original performance and achieve state-of-the-art results." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 535, + 482, + 631 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 535, + 482, + 631 + ], + "spans": [ + { + "bbox": [ + 130, + 535, + 482, + 631 + ], + "type": "text", + "content": "Object detection has significantly benefited from the FPN-like paradigm in past few years. However, recent developments indicate that newer technologies, including innovative architectures and training strategies, may offer superior performance for modern detectors, making the FPN-like paradigm potentially limiting. In many cases, the effectiveness of the FPN-like paradigm can be enhanced through the application of SNI to address feature misalignment issues. Fully addressing the challenge of misalignment during feature fusion in existing vision models requires further research efforts." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 190, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 190, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 190, + 91, + 447, + 102 + ], + "type": "text", + "content": "Rethinking Features-Fused-Pyramid-Neck for Object Detection" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 132, + 137, + 481, + 369 + ], + "blocks": [ + { + "bbox": [ + 177, + 116, + 435, + 127 + ], + "lines": [ + { + "bbox": [ + 177, + 116, + 435, + 127 + ], + "spans": [ + { + "bbox": [ + 177, + 116, + 435, + 127 + ], + "type": "text", + "content": "Table 6: Comparisons of state-of-the-art detectors on COCO." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 132, + 137, + 481, + 369 + ], + "lines": [ + { + "bbox": [ + 132, + 137, + 481, + 369 + ], + "spans": [ + { + "bbox": [ + 132, + 137, + 481, + 369 + ], + "type": "table", + "html": "
ModelsSizeParam.(M)FLOPs(G)AP(%)APs(%)APm(%)APl(%)LatencyV100b=1 (ms)
Transformers/CNNs (non-real-time, pre-trained)
FPN [21]-56.4145.836.817.538.747.8-
BiFPN [34]102421.055.049.3---42.8
ASFF [23]800--43.927.046.653.434.0
PANet [24]---51.032.653.964.8-
RFPN [28]---51.331.754.664.8-
DETR-Def. [40]-40.0173.043.826.447.158.0-
DETR-Dyn. [3]---47.228.649.359.1-
DETR-DND [17]133348.0265.048.6----
DINO [38]133347.0860.051.0----
Swin-T [26]-86.0745.050.5----
Swin-S [26]-107.0838.051.8----
Swin-B [26]1280145.0982.051.9----
CNNs (real-time, without pre-trained)
Yolov4 [1]60864.4142.843.526.746.753.3-
YoloX [7]64054.2155.649.7---14.5
PPYoloE [37]64052.2110.150.931.455.366.112.8
Yolov5 [14]64046.5109.149.0---10.1
Yolov6 [16]64058.5144.051.7----
Yolov7 [35]64036.9104.751.231.855.565.06.3
Yolov8 [15]64043.7165.252.9----
SYolo(ours)64057.2142.553.135.558.769.910.3
", + "image_path": "a9d568e9452378ee9d23a3db464086fbed2ad80a06402c95cce39c100f7376eb.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 194, + 398, + 420, + 628 + ], + "blocks": [ + { + "bbox": [ + 140, + 376, + 473, + 388 + ], + "lines": [ + { + "bbox": [ + 140, + 376, + 473, + 388 + ], + "spans": [ + { + "bbox": [ + 140, + 376, + 473, + 388 + ], + "type": "text", + "content": "Table 7: Architecture of the SYOLO. The E-ELAN/C2f is from the YOLOv7/8." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 194, + 398, + 420, + 628 + ], + "lines": [ + { + "bbox": [ + 194, + 398, + 420, + 628 + ], + "spans": [ + { + "bbox": [ + 194, + 398, + 420, + 628 + ], + "type": "table", + "html": "
StageBlock-1/s/n‡Channels-1/s/n‡K. sizeStride
Backbone
P1Conv64 /16 /1632
P2ESD-I/II128 /32 /323, 42
C2f ×3/2/1128 /32 /323, 11
P3ESD-I/II256 /64 /643, 42
C2f ×6/3/1256 /64 /643, 11
P4ESD-I/II512 /128 /1283, 42
C2f ×6/3/1512 /128 /1283, 11
P5ESD-I/II1024 /256 /2563, 42
C2f ×3/2/11024 /256 /2563, 11
SPP ×11024 /256 /256--
Neck(SA)
P5-P4SNI1536 /384 /284--
E-ELAN†×12048 /512 /5123, 11
P4-P3SNI512 /192 /192--
E-ELAN†×2/2/1512 /256 /2563, 11
P3-P4ESD-I/II256 /128 /1283, 42
E-ELAN†×2/2/11024 /512 /5123, 11
P4-P5ESD-I/II512 /256 /2563, 42
E-ELAN†×2/2/12048 /1024 /10243, 11
", + "image_path": "eeecd0c17bdf4f3df420cc0845561b9a9af5f15ad65a7e52591dfefff30745c4.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 198, + 628, + 419, + 653 + ], + "lines": [ + { + "bbox": [ + 198, + 628, + 419, + 653 + ], + "spans": [ + { + "bbox": [ + 198, + 628, + 419, + 653 + ], + "type": "text", + "content": "The first " + }, + { + "bbox": [ + 198, + 628, + 419, + 653 + ], + "type": "inline_equation", + "content": "3 \\times 3" + }, + { + "bbox": [ + 198, + 628, + 419, + 653 + ], + "type": "text", + "content": " vanilla convolutional layer of the E-ELAN is replaced by the GSCConvE-I, and the end pointwise convolutional layer of the E-ELAN is replaced by the C2f." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_footnote" + }, + { + "bbox": [ + 200, + 653, + 414, + 662 + ], + "lines": [ + { + "bbox": [ + 200, + 653, + 414, + 662 + ], + "spans": [ + { + "bbox": [ + 200, + 653, + 414, + 662 + ], + "type": "inline_equation", + "content": "\\ddagger" + }, + { + "bbox": [ + 200, + 653, + 414, + 662 + ], + "type": "text", + "content": " The 'l/s/n' means the model scale of the large/samll/nano." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 202, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 202, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 202, + 101 + ], + "type": "text", + "content": "Hulin Li" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 138, + 482, + 212 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 138, + 482, + 212 + ], + "spans": [ + { + "bbox": [ + 130, + 138, + 482, + 212 + ], + "type": "text", + "content": "This work is supported by National Natural Science Foundation of China (Grant No. 52172381), Graduate Research Innovation Foundation of Chongqing Jiaotong University (Grant No. CYB240259). The author would also like to thank Prof. Dr. Jun Li, Prof. Dr. Qiliang Ren, and Ms. Xinxin Liu (Intelligent Transportation Big Data Center, School of Information Science and Engineering, CQJTU) for providing GPUs." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 228, + 197, + 241 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 228, + 197, + 241 + ], + "spans": [ + { + "bbox": [ + 133, + 228, + 197, + 241 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 251, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 138, + 251, + 480, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 251, + 480, + 274 + ], + "spans": [ + { + "bbox": [ + 138, + 251, + 480, + 274 + ], + "type": "text", + "content": "1. Bochkovskiy, A., Wang, C.Y., Liao, H.Y.M.: Yolov4: Optimal speed and accuracy of object detection. arXiv preprint arXiv:2004.10934 (2020)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 274, + 480, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 274, + 480, + 306 + ], + "spans": [ + { + "bbox": [ + 138, + 274, + 480, + 306 + ], + "type": "text", + "content": "2. Chollet, F.: Xception: Deep learning with depthwise separable convolutions. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1251-1258 (2017)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 307, + 480, + 339 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 307, + 480, + 339 + ], + "spans": [ + { + "bbox": [ + 138, + 307, + 480, + 339 + ], + "type": "text", + "content": "3. Dai, X., Chen, Y., Yang, J., Zhang, P., Yuan, L., Zhang, L.: Dynamic detr: End-to-end object detection with dynamic attention. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 2988-2997 (2021)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 339, + 480, + 372 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 339, + 480, + 372 + ], + "spans": [ + { + "bbox": [ + 138, + 339, + 480, + 372 + ], + "type": "text", + "content": "4. Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: Imagenet: A large-scale hierarchical image database. In: 2009 IEEE conference on computer vision and pattern recognition. pp. 248-255. IEEE (2009)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 373, + 480, + 404 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 373, + 480, + 404 + ], + "spans": [ + { + "bbox": [ + 138, + 373, + 480, + 404 + ], + "type": "text", + "content": "5. Everingham, M., Eslami, S.A., Van Gool, L., Williams, C.K., Winn, J., Zisserman, A.: The Pascal visual object classes challenge: A retrospective. International journal of computer vision 111, 98-136 (2015)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 404, + 480, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 404, + 480, + 426 + ], + "spans": [ + { + "bbox": [ + 138, + 404, + 480, + 426 + ], + "type": "text", + "content": "6. Fu, C.Y., Liu, W., Ranga, A., Tyagi, A., Berg, A.C.: Dssd: Deconvolutional single shot detector. arXiv preprint arXiv:1701.06659 (2017)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 426, + 480, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 426, + 480, + 448 + ], + "spans": [ + { + "bbox": [ + 138, + 426, + 480, + 448 + ], + "type": "text", + "content": "7. Ge, Z., Liu, S., Wang, F., Li, Z., Sun, J.: Yolox: Exceeding yolo series in 2021. arXiv preprint arXiv:2107.08430 (2021)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 448, + 480, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 448, + 480, + 470 + ], + "spans": [ + { + "bbox": [ + 138, + 448, + 480, + 470 + ], + "type": "text", + "content": "8. Girshick, R.: Fast r-cnn. In: Proceedings of the IEEE international conference on computer vision. pp. 1440-1448 (2015)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 470, + 480, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 470, + 480, + 502 + ], + "spans": [ + { + "bbox": [ + 138, + 470, + 480, + 502 + ], + "type": "text", + "content": "9. Girshick, R., Donahue, J., Darrell, T., Malik, J.: Rich feature hierarchies for accurate object detection and semantic segmentation. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 580-587 (2014)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 502, + 480, + 535 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 502, + 480, + 535 + ], + "spans": [ + { + "bbox": [ + 138, + 502, + 480, + 535 + ], + "type": "text", + "content": "10. Han, K., Wang, Y., Tian, Q., Guo, J., Xu, C., Xu, C.: Ghostnet: More features from cheap operations. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 1580-1589 (2020)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 535, + 480, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 535, + 480, + 567 + ], + "spans": [ + { + "bbox": [ + 138, + 535, + 480, + 567 + ], + "type": "text", + "content": "1. He, K., Zhang, X., Ren, S., Sun, J.: Spatial pyramid pooling in deep convolutional networks for visual recognition. IEEE transactions on pattern analysis and machine intelligence 37(9), 1904-1916 (2015)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 567, + 480, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 567, + 480, + 600 + ], + "spans": [ + { + "bbox": [ + 138, + 567, + 480, + 600 + ], + "type": "text", + "content": "2. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 770-778 (2016)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 601, + 480, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 601, + 480, + 633 + ], + "spans": [ + { + "bbox": [ + 138, + 601, + 480, + 633 + ], + "type": "text", + "content": "3. Howard, A.G., Zhu, M., Chen, B., Kalenichenko, D., Wang, W., Weyand, T., Andreetto, M., Adam, H.: Mobilenets: Efficient convolutional neural networks for mobile vision applications. arXiv preprint arXiv:1704.04861 (2017)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 633, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 633, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 633, + 480, + 665 + ], + "type": "text", + "content": "4. Jocher, G., Chaurasia, A., Stoken, A., Borovec, J., Kwon, Y., Fang, J., Michael, K., Montes, D., Nadar, J., Skalski, P., et al.: ultralytics/yolov5: v6. 1-tensorrt, tensorflow edge tpu and openvino export and inference. Zenodo (2022)" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 190, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 190, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 190, + 91, + 447, + 102 + ], + "type": "text", + "content": "Rethinking Features-Fused-Pyramid-Neck for Object Detection" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 117, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 133, + 117, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 117, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 133, + 117, + 480, + 138 + ], + "type": "text", + "content": "15. Jocher, G., et al.: Yolo-ultralytics. https://github.com/ultralytics/ultralytics (2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 139, + 480, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 139, + 480, + 171 + ], + "spans": [ + { + "bbox": [ + 133, + 139, + 480, + 171 + ], + "type": "text", + "content": "16. Li, C., Li, L., Jiang, H., Weng, K., Geng, Y., Li, L., Ke, Z., Li, Q., Cheng, M., Nie, W., et al.: Yolov6: A single-stage object detection framework for industrial applications. arXiv preprint arXiv:2209.02976 (2022)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 171, + 480, + 203 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 171, + 480, + 203 + ], + "spans": [ + { + "bbox": [ + 133, + 171, + 480, + 203 + ], + "type": "text", + "content": "17. Li, F., Zhang, H., Liu, S., Guo, J., Ni, L.M., Zhang, L.: Dn-detr: Accelerate detr training by introducing query denoising. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 13619-13627 (2022)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 203, + 480, + 236 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 203, + 480, + 236 + ], + "spans": [ + { + "bbox": [ + 133, + 203, + 480, + 236 + ], + "type": "text", + "content": "18. Li, H., Li, J., Wei, H., Liu, Z., Zhan, Z., Ren, Q.: Slim-neck by gsconv: a lightweight-design for real-time detector architectures. Journal of Real-Time Image Processing 21(3), 62 (2024)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 133, + 236, + 480, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 236, + 480, + 267 + ], + "spans": [ + { + "bbox": [ + 133, + 236, + 480, + 267 + ], + "type": "text", + "content": "19. Li, Y., Chen, Y., Wang, N., Zhang, Z.: Scale-aware trident networks for object detection. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 6054-6063 (2019)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 133, + 267, + 480, + 289 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 267, + 480, + 289 + ], + "spans": [ + { + "bbox": [ + 133, + 267, + 480, + 289 + ], + "type": "text", + "content": "20. Li, Z., Yang, L., Zhou, F.: Fssd: feature fusion single shot multibox detector. arXiv preprint arXiv:1712.00960 (2017)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 289, + 480, + 321 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 289, + 480, + 321 + ], + "spans": [ + { + "bbox": [ + 132, + 289, + 480, + 321 + ], + "type": "text", + "content": "21. Lin, T.Y., Dollar, P., Girshick, R., He, K., Hariharan, B., Belongie, S.: Feature pyramid networks for object detection. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 2117-2125 (2017)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 321, + 480, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 321, + 480, + 365 + ], + "spans": [ + { + "bbox": [ + 132, + 321, + 480, + 365 + ], + "type": "text", + "content": "22. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dólar, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13. pp. 740-755. Springer (2014)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 365, + 480, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 365, + 480, + 385 + ], + "spans": [ + { + "bbox": [ + 132, + 365, + 480, + 385 + ], + "type": "text", + "content": "23. Liu, S., Huang, D., Wang, Y.: Learning spatial fusion for single-shot object detection. arxiv 2019. arXiv preprint arXiv:1911.09516 (1911)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 385, + 480, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 385, + 480, + 418 + ], + "spans": [ + { + "bbox": [ + 132, + 385, + 480, + 418 + ], + "type": "text", + "content": "24. Liu, S., Qi, L., Qin, H., Shi, J., Jia, J.: Path aggregation network for instance segmentation. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 8759-8768 (2018)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 418, + 480, + 460 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 418, + 480, + 460 + ], + "spans": [ + { + "bbox": [ + 132, + 418, + 480, + 460 + ], + "type": "text", + "content": "25. Liu, W., Anguelov, D., Erhan, D., Szegedy, C., Reed, S., Fu, C.Y., Berg, A.C.: Ssd: Single shot multibox detector. In: Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part I 14. pp. 21-37. Springer (2016)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 460, + 480, + 504 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 460, + 480, + 504 + ], + "spans": [ + { + "bbox": [ + 132, + 460, + 480, + 504 + ], + "type": "text", + "content": "26. Liu, Z., Lin, Y., Cao, Y., Hu, H., Wei, Y., Zhang, Z., Lin, S., Guo, B.: Swin transformer: Hierarchical vision transformer using shifted windows. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 10012-10022 (2021)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 504, + 480, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 504, + 480, + 536 + ], + "spans": [ + { + "bbox": [ + 132, + 504, + 480, + 536 + ], + "type": "text", + "content": "27. Liu, Z., Mao, H., Wu, C.Y., Feichtenhofer, C., Darrell, T., Xie, S.: A convnet for the 2020s. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 11976-11986 (2022)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 537, + 480, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 537, + 480, + 568 + ], + "spans": [ + { + "bbox": [ + 132, + 537, + 480, + 568 + ], + "type": "text", + "content": "28. Qiao, S., Chen, L.C., Yuille, A.: Detectors: Detecting objects with recursive feature pyramid and switchable atrous convolution. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10213-10224 (2021)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 568, + 480, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 568, + 480, + 601 + ], + "spans": [ + { + "bbox": [ + 132, + 568, + 480, + 601 + ], + "type": "text", + "content": "29. Redmon, J., Divvala, S., Girshick, R., Farhadi, A.: You only look once: Unified, real-time object detection. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 779-788 (2016)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 601, + 480, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 601, + 480, + 622 + ], + "spans": [ + { + "bbox": [ + 132, + 601, + 480, + 622 + ], + "type": "text", + "content": "30. Redmon, J., Farhadi, A.: Yolo9000: better, faster, stronger. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 7263-7271 (2017)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 622, + 480, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 622, + 480, + 643 + ], + "spans": [ + { + "bbox": [ + 132, + 622, + 480, + 643 + ], + "type": "text", + "content": "31. Redmon, J., Farhadi, A.: Yolov3: An incremental improvement. arXiv preprint arXiv:1804.02767 (2018)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 643, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 643, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 643, + 480, + 665 + ], + "type": "text", + "content": "32. Simonyan, K., Zisserman, A.: Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556 (2014)" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 202, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 202, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 202, + 101 + ], + "type": "text", + "content": "Hulin Li" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 392 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "text", + "content": "33. Singh, B., Davis, L.S.: An analysis of scale invariance in object detection snip. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 3578-3587 (2018)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 150, + 482, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 150, + 482, + 183 + ], + "spans": [ + { + "bbox": [ + 130, + 150, + 482, + 183 + ], + "type": "text", + "content": "34. Tan, M., Pang, R., Le, Q.V.: Efficientdet: Scalable and efficient object detection. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10781-10790 (2020)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 183, + 482, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 183, + 482, + 226 + ], + "spans": [ + { + "bbox": [ + 130, + 183, + 482, + 226 + ], + "type": "text", + "content": "35. Wang, C.Y., Bochkovskiy, A., Liao, H.Y.M.: Yolov7: Trainable bag-of-freebies sets new state-of-the-art for real-time object detectors. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 7464-7475 (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 226, + 482, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 226, + 482, + 259 + ], + "spans": [ + { + "bbox": [ + 130, + 226, + 482, + 259 + ], + "type": "text", + "content": "36. Wu, Y., Chen, Y., Yuan, L., Liu, Z., Wang, L., Li, H., Fu, Y.: Rethinking classification and localization for object detection. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10186-10195 (2020)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 259, + 482, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 259, + 482, + 293 + ], + "spans": [ + { + "bbox": [ + 130, + 259, + 482, + 293 + ], + "type": "text", + "content": "37. Yang, Z., Zhou, Y., Chen, Z., Ngiam, J.: 3d-man: 3d multi-frame attention network for object detection. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 1863-1872 (2021)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 293, + 482, + 325 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 293, + 482, + 325 + ], + "spans": [ + { + "bbox": [ + 130, + 293, + 482, + 325 + ], + "type": "text", + "content": "38. Zhang, H., Li, F., Liu, S., Zhang, L., Su, H., Zhu, J., Ni, L.M., Shum, H.Y.: Dino: Detr with improved denoising anchor boxes for end-to-end object detection. arXiv preprint arXiv:2203.03605 (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 325, + 482, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 325, + 482, + 358 + ], + "spans": [ + { + "bbox": [ + 130, + 325, + 482, + 358 + ], + "type": "text", + "content": "39. Zhang, X., Zhou, X., Lin, M., Sun, J.: Shufflenet: An extremely efficient convolutional neural network for mobile devices. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 6848-6856 (2018)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 358, + 482, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 358, + 482, + 392 + ], + "spans": [ + { + "bbox": [ + 130, + 358, + 482, + 392 + ], + "type": "text", + "content": "40. Zhu, X., Su, W., Lu, L., Li, B., Wang, X., Dai, J.: Deformable detr: Deformable transformers for end-to-end object detection. arXiv preprint arXiv:2010.04159 (2020)" + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 190, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 190, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 190, + 91, + 447, + 102 + ], + "type": "text", + "content": "Rethinking Features-Fused-Pyramid-Neck for Object Detection" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Rethinking Few-shot Class-incremental Learning_ Learning from Yourself/a34c73cf-0382-420e-bbe9-4f775aa32eab_content_list.json b/2024/Rethinking Few-shot Class-incremental Learning_ Learning from Yourself/a34c73cf-0382-420e-bbe9-4f775aa32eab_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..ccfa3c6b20aff3a548aa70cf8f6ff37181f423cb --- /dev/null +++ b/2024/Rethinking Few-shot Class-incremental Learning_ Learning from Yourself/a34c73cf-0382-420e-bbe9-4f775aa32eab_content_list.json @@ -0,0 +1,1946 @@ +[ + { + "type": "text", + "text": "Rethinking Few-shot Class-incremental Learning: Learning from Yourself", + "text_level": 1, + "bbox": [ + 220, + 140, + 782, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yu-Ming Tang $^{1,3}$ , Yi-Xing Peng $^{1,3,5}$ , Jingke Meng $^{\\star 1,3}$ , and Wei-Shi Zheng $^{1,2,3,4,5}$", + "bbox": [ + 241, + 210, + 759, + 244 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ School of Computer Science and Engineering, Sun Yat-sen University, China \n $^{2}$ Peng Cheng Laboratory, Shenzhen, China", + "bbox": [ + 236, + 253, + 759, + 281 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3 Key Laboratory of Machine Intelligence and Advanced Computing, Ministry of Education, China", + "bbox": [ + 230, + 282, + 771, + 309 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "4 Guangdong Province Key Laboratory of Information Security Technology, Sun Yat-sen University, China", + "bbox": [ + 230, + 309, + 767, + 337 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{5}$ Pazhou Laboratory (Huangpu), Guangzhou, Guangdong 510555, China {tangym9, pengyx23}@mail2.sysu.edu.cn, mengjke@gmail.com, wszheng@ieee.org", + "bbox": [ + 254, + 338, + 743, + 378 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Few-shot class-incremental learning (FSCIL) aims to learn sequential classes with limited samples in a few-shot fashion. Inherited from the classical class-incremental learning setting, the popular benchmark of FSCIL uses averaged accuracy ( $aAcc$ ) and last-task averaged accuracy ( $lAcc$ ) as the evaluation metrics. However, we reveal that such evaluation metrics may not provide adequate emphasis on the novel class performance, and the continual learning ability of FSCIL methods could be ignored under this benchmark. In this work, as a complement to existing metrics, we offer a new metric called generalized average accuracy ( $gAcc$ ) which is designed to provide an extra equitable evaluation by incorporating different perspectives of the performance under the guidance of a parameter $\\alpha$ . We also present an overall metric in the form of the area under the curve (AUC) along the $\\alpha$ . Under the guidance of $gAcc$ , we release the potential of intermediate features of the vision transformers to boost the novel-class performance. Taking information from intermediate layers which are less class-specific and more generalizable, we manage to rectify the final features, leading to a more generalizable transformer-based FSCIL framework. Without complex network designs or cumbersome training procedures, our method outperforms existing FSCIL methods at $aAcc$ and $gAcc$ on three datasets. See codes at 1", + "bbox": [ + 259, + 417, + 743, + 694 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Continual Learning $\\cdot$ Few-shot Learning $\\cdot$ Catastrophic Forgetting $\\cdot$ Vision Transformer", + "bbox": [ + 259, + 707, + 740, + 734 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 746, + 375, + 762 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Although existing deep neural networks (DNNs) achieve great success in various real-world applications [17, 31, 32, 34, 43, 44, 54, 68], they suffer from catastrophic", + "bbox": [ + 212, + 771, + 787, + 803 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Corresponding Author", + "bbox": [ + 217, + 810, + 385, + 825 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "1 https://github.com/iSEE-Laboratory/Revisiting_FSCIL", + "bbox": [ + 220, + 825, + 622, + 839 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/11401f85782ee84c427d469061887ee428cbb72141ea6d2507449d58dbad8c37.jpg", + "image_caption": [ + "(a) Performance after each task" + ], + "image_footnote": [], + "bbox": [ + 246, + 167, + 496, + 286 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/ecb69e171671f49bdc7f0d5182231b32f89132f5e403e6781a9b76eacd65e3e0.jpg", + "image_caption": [ + "(b) Detailed Performance after finishing the last task (Task 9)", + "Fig. 1: Performance under different metrics of two recent methods [24, 25]. (a): The dotted line denotes the average accuracy ( $aAcc$ ) widely used in classical class-incremental learning. The solid line presents our proposed generalized accuracy ( $gAcc$ ). We also show the average of each point on the right side. (Blue: $gAcc$ , Red: $aAcc$ ). (b): The detailed accuracies of each task after training on the last task (task 8). It is evident that S3C exhibits superior performance when confronted with novel classes. The conventional metric $aAcc$ fails to reflect this due to the domination by the base-class performance. Models are trained on the CIFAR-100 dataset. Best viewed in color." + ], + "image_footnote": [], + "bbox": [ + 509, + 169, + 756, + 286 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "forgetting when dealing with sequential data streams. The forgetting problem has attracted heavy attention in recent years [4,16,21,29,33,45,48,50,51,57,61,64,65, 71,74]. In this work, we focus on a more realistic and challenging problem: few-shot class-incremental learning (FSCIL), in which after training on a task with sufficient data (base task), only a few samples are provided during incremental tasks (novel tasks). Since Tao et al. [52] proposed the FSCIL benchmark in 2020, a large number of algorithms have emerged to tackle this problem, including meta-learning based [8, 20, 70], structure based [2, 25, 52, 58, 63], feature space based [3, 6, 20, 24, 26, 59, 69] methods, etc.", + "bbox": [ + 212, + 392, + 787, + 529 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "However, almost all these methods are evaluated following the traditional class-incremental learning setting, i.e. using the average accuracy ( $aAcc$ ) or the last-task averaged accuracy ( $lAcc$ ) as the main metrics to critique the proposed method. Such accuracies are calculated by weighted averaging the base-task and novel-tasks accuracies with the weight proportional to the number of classes of each task. Considering the reality that the FSCIL benchmarks make the majority of the classes (about $50\\%$ to $60\\%$ ) as the base task to establish a well-trained backbone for few-shot continual learning, the existing $aAcc$ or $lAcc$ metric is dominated by the base-class performance. In other words, current evaluation metrics can be largely boosted by only improving the performance of base classes. As shown in Fig. 1, a method like softNet [25] with higher base-class performance (see Fig. 1 (b)) usually gets higher $aAcc$ (see red lines in Fig. 1 (a)) regardless of the inferior performance on the novel classes. Due to the inadequacy to capture performance changes in novel classes, current metrics are biased toward the performance of base classes and have limitations in guiding the FSCIL community. To tackle this, we propose a more balanced metric called generalized accuracy ( $gAcc$ ). Parameterized by $\\alpha$ ranging from 0 to 1, the $gAcc$ offers a way to evaluate a model with explicit emphasis on performance on novel classes. Furthermore, we can comprehensively evaluate the model by calculating the area under the curve (AUC) along the parameter $\\alpha$ . It is worth mentioning that our intention", + "bbox": [ + 212, + 537, + 787, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "Tang et al.", + "bbox": [ + 271, + 114, + 346, + 128 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "is NOT to replace all previous metrics, but to provide a more fine-grained and balanced evaluation of FSCIL methods complementary to them.", + "bbox": [ + 212, + 146, + 782, + 176 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We empirically find that FSCIL can benefit from the powerful vision transformer (ViT) [15] by analyzing the performance of novel classes. Specifically, although vanilla ViT performs poorly in terms of unseen novel classes, we reveal that the intermediate features of ViT are less specified on base classes. It means the ViT can learn abundant semantic information but the information is mistakenly dropped in deep layers when extracting features for novel classes.", + "bbox": [ + 212, + 176, + 785, + 267 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Based on the observation, we propose a lightweight Feature Rectification (FR) module to rectify the output features of the ViT conditioned on the intermediate features. To achieve this, the FR module is supervised with two losses that work on two different levels, namely instance-level relation transfer and class-center relation transfer. The former transfers the relations between instances while the latter focuses on relations between instances and the class prototypes. We further enhance the FR module by leveraging hierarchical knowledge from multiple intermediate layers. Finally, our method boosts the novel-class accuracy and maintains a comparable performance of base classes.", + "bbox": [ + 212, + 267, + 787, + 402 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our contributions are summarized as follows. (1) We revisit the existing FSCIL benchmark and propose a novel metric called generalized accuracy ( $gAcc$ ) for a more fine-grained and balanced evaluation. We have implemented 14 FSCIL methods with codes released and show that the newly introduced SOTAs do not necessarily perform better on $gAcc$ . (2) Based on $gAcc$ , we reveal the potential of vision transformers in FSCIL. We empirically find that although features from the intermediate layer exhibit sub-optimal performance for the base classes, their transferrable ability helps them to have a distinct advantage in handling novel classes. (3) A lightweight FR module is proposed to rectify the final features to be more transferrable. The FR module manages to make the output feature more generalizable with assistance from intermediate features. (4) We conduct extensive experiments on 3 datasets and evaluate our method under both $aAcc$ and $gAcc$ , showing the effectiveness and superiority of our method.", + "bbox": [ + 212, + 402, + 787, + 599 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 214, + 609, + 387, + 625 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Meta-learning based methods. Due to the limited samples of novel classes, some works [8,20,67,70] propose to use meta-learning to improve the transferability of the model. Concretely, MetaFscil [8] proposed to simulate novel tasks by sampling from the base task. C-FSCIL [20] uses meta-learning to train the backbone, which maps images from different classes to quasi-orthogonal vectors. ALFSCIL [30] formulate the process into a bi-level optimization problem.", + "bbox": [ + 212, + 628, + 785, + 719 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Optimization-based methods. There is a large group of works focusing on the optimization of the DNNs in FSCIL. CLOM [76] considers the behavior of the backbone under different margins of the cosine loss. NC-FSCIL [59] designs a pre-defined classifier for efficient FSCIL without any explicit regularizer. SPPR [73] proposes a prototype refinement strategy for stronger expression ability. FACT [69] assigns virtual prototypes in the current stage for future updates. Alice [42] instead focuses on the compactness and the variety of the feature space. Based on the classifier learning, S3C [24] proposes a self-supervised stochastic", + "bbox": [ + 212, + 719, + 787, + 840 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "Rethinking Few-shot Class-incremental Learning: Learning from Yourself", + "bbox": [ + 245, + 114, + 733, + 130 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "classifier to tackle the overfitting and forgetting problem. There are more works working on the feature space constraint [3, 6, 13, 26, 27, 56, 60], or knowledge distillation [7, 9-11, 14, 39, 66, 72].", + "bbox": [ + 212, + 146, + 782, + 191 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Data-driven methods. There is also a branch of work focusing on the data side of this problem, which focuses on the data replay [1,35]. FSIL-GAN [1] employs a GAN to generate features to compensate for the scarcity of data. DF-replay [35] uses synthesized data generated by a generator to replay. PL-FSCIL [23] seeks prior knowledge from unlabeled data.", + "bbox": [ + 212, + 191, + 784, + 266 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Structure-based methods. In addition to these works, a unique branch of works focuses on the network structure, like topology preserving [52], growing classifier [63], proposing a dynamic structure of the backbone [58], structure fusion [2, 22], sub-networks [25] and ensemble models [75].", + "bbox": [ + 212, + 267, + 785, + 327 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Compared with existing works, our work rephrases the challenge of the FSCIL setting and first reveals the potential of the intermediate features of the ViT backbone. There is also an existing work [42] offering an alternative metric. They did not provide a fine-grained evaluation with various perspectives and have potential risks in stability and symmetry problems. Please find more detailed discussions about related works in the supplementary material.", + "bbox": [ + 212, + 327, + 784, + 417 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3 Generalized Average Accuracy", + "text_level": 1, + "bbox": [ + 214, + 422, + 545, + 438 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1 Limitation of the Existing Metrics", + "text_level": 1, + "bbox": [ + 214, + 438, + 545, + 454 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Accuracy metric on incremental tasks. Consider a classification model $f(\\cdot)$ that takes an image as input and outputs the class predictions. For an incremental problem, there are multiple tasks and multiple test sets. Let us define test set at task $\\mathcal{T}_i$ as $\\mathcal{D}_{test}^i = \\{x_t^i,y_t^i\\}_{t = 1}^n$ . Usually, the labeling space $\\mathcal{Y}^i\\triangleq \\{y_t^i |(x_t^i,y_t^i)\\in \\mathcal{D}_{test}^i\\}$ of every task does not have any intersection, i.e. $\\forall i\\neq j,\\mathcal{Y}^i\\bigcap \\mathcal{Y}^j = \\emptyset$ . After trained on $\\mathcal{T}_i$ , the model can be denoted as $f_{i}(\\cdot)$ . The corresponding accuracy of model $f_{i}(\\cdot)$ at task $\\mathcal{T}_j$ ( $j\\leq i$ ) can be denoted as $\\mathcal{A}_i^j\\triangleq \\frac{1}{n^j}\\sum_{t = 1}^{n^j}\\mathbb{1}(\\zeta (f_i(x_t^j)) = y_t^j)$ , where $\\mathbb{1}(\\cdot) = 1$ if the condition is true, otherwise 0. $\\zeta$ is the argmax operation.", + "bbox": [ + 214, + 460, + 787, + 599 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Average accuracy. It is worth mentioning that for class-incremental learning, evaluating the model on all the previous tasks is essential. To test the model on previous tasks, an overall test set that contains all previous test sets is constructed as $\\mathcal{D}_{test}^{1\\rightarrow i} = \\bigcup_{m = 1}^{i}\\mathcal{D}_{test}^{m}$ . After trained on task $\\mathcal{T}_i$ , the accuracy of model $f_{i}(\\cdot)$ on the overall test set $\\mathcal{D}_{test}^{1\\rightarrow i}$ can be presented as $A_{i}^{1\\rightarrow i}$ . Once a $n_t$ -task incremental training ends, the average accuracy ( $aAcc$ ) can be calculated as:", + "bbox": [ + 214, + 601, + 787, + 691 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\na A c c = \\frac {1}{n _ {t}} \\sum_ {i = 1} ^ {n _ {t}} a A c c _ {i} = \\frac {1}{n _ {t}} \\sum_ {i = 1} ^ {n _ {t}} A _ {i} ^ {1 \\rightarrow i}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 372, + 700, + 785, + 738 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where the $aAcc_{i}$ is the average accuracy at each task. We can also derive two additional metrics incidentally: last-task averaged accuracy $lAcc = aAcc_{n_t} = A_{n_t}^{1\\rightarrow n_t}$ , and task-wise average accuracy $(tAcc)$ :", + "bbox": [ + 214, + 744, + 787, + 792 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nt A c c = \\frac {1}{n _ {t}} \\sum_ {i = 1} ^ {n _ {t}} t A c c _ {i} = \\frac {1}{n _ {t}} \\sum_ {i = 1} ^ {n _ {t}} \\frac {1}{i} \\sum_ {j = 1} ^ {i} A _ {i} ^ {j}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 362, + 801, + 785, + 842 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "Tang et al.", + "bbox": [ + 271, + 114, + 346, + 128 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where the $tAcc_i$ is the task-wise accuracy at each task. Note that the $aAcc$ does not share the same value as $tAcc$ normally since the number of classes of tasks is usually not the same. This will be elaborated on later in detail.", + "bbox": [ + 212, + 146, + 782, + 190 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Limitation. Since FSCIL usually sets the majority of the classes as the base task, existing metrics potentially neglect the performance changes of novel classes. To elaborate on this, we analyze the $aAcc$ as follows. Firstly, we denote the size of the labeling space of task $\\mathcal{T}_i$ as $|\\mathcal{V}_i|$ . In FSCIL, the number of classes of each novel task is usually the same, i.e. $|\\mathcal{V}_{novel}| = |\\mathcal{V}_2| = |\\mathcal{V}_3| = \\dots = |\\mathcal{V}_{n_t}| \\ll |\\mathcal{V}_1|$ . Assuming there are total $n_t$ tasks, and task $\\mathcal{T}_1$ is the base task with sufficient classes and samples (non-few-shot). The following tasks $\\{\\mathcal{T}_i\\}_{i=2}^{n_t}$ are novel tasks, with $|\\mathcal{V}_{novel}|$ classes for each task and a few samples per class. Now, we can rewrite average accuracy on task $\\mathcal{T}_i$ as:", + "bbox": [ + 212, + 191, + 787, + 328 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\na A c c _ {i} = \\frac {\\left| \\mathcal {Y} _ {1} \\right| A _ {i} ^ {1} + \\left| \\mathcal {Y} _ {n o v e l} \\right| \\sum_ {j = 2} ^ {i} A _ {i} ^ {j}}{\\left| \\mathcal {Y} _ {1} \\right| + (i - 1) \\left| \\mathcal {Y} _ {n o v e l} \\right|} = \\frac {\\frac {\\left| \\mathcal {Y} _ {1} \\right|}{\\left| \\mathcal {Y} _ {n o v e l} \\right|} A _ {i} ^ {1} + \\sum_ {j = 2} ^ {i} A _ {i} ^ {j}}{\\frac {\\left| \\mathcal {Y} _ {1} \\right|}{\\left| \\mathcal {Y} _ {n o v e l} \\right|} + (i - 1)}. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 285, + 333, + 785, + 377 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "From Eq. (3) we see that the value of $aAcc_i$ will be dominated by the base class accuracy $A_i^1$ if the ratio $\\frac{|\\mathcal{Y}_1|}{|\\mathcal{Y}_{\\text{novel}}|}$ is large. Unfortunately, the ratio is indeed large in the FSCIL settings. (e.g. $\\frac{|\\mathcal{Y}_1|}{|\\mathcal{Y}_{\\text{novel}}|} = 12$ for both CIFAR-100 and miniImageNet dataset, 10 for CUB-200.) Under such a large ratio, the changes (both improvement and degrades) of novel classes cannot be effectively evaluated by $aAcc$ .", + "bbox": [ + 214, + 382, + 782, + 467 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2 Generalizing the Average Accuracy", + "text_level": 1, + "bbox": [ + 214, + 472, + 552, + 487 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To overcome the limitations of existing metrics, we now define the generalized accuracy (gAcc) at task $\\mathcal{T}_i$ as:", + "bbox": [ + 214, + 489, + 785, + 518 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\ng A c c _ {i} (\\alpha) = \\frac {\\alpha \\frac {| \\mathcal {Y} _ {1} |}{| \\mathcal {Y} _ {n o v e l} |} A _ {i} ^ {1} + \\sum_ {j = 2} ^ {i} A _ {i} ^ {j}}{\\alpha \\frac {| \\mathcal {Y} _ {1} |}{| \\mathcal {Y} _ {n o v e l} |} + (i - 1)}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 372, + 525, + 785, + 570 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "in which the $gAcc$ is parameterized by $\\alpha \\in [0,1]$ . At task $\\mathcal{T}_i$ , $gAcc_i(0)$ measures the accuracy of the novel tasks only. $gAcc_i(\\frac{|\\mathcal{Y}_{\\text{novel}}|}{|\\mathcal{Y}_1|})$ denotes the task-wise accuracy $tAcc_i$ at task $\\mathcal{T}_i$ while $gAcc(1)$ represents the average accuracy $aAcc_i$ at task $\\mathcal{T}_i$ . In addition to these discrete values, we further generalize the $\\alpha$ to any rational numbers in [0,1]. Controlled by $\\alpha$ in Eq. (4), when $\\alpha \\to 0$ , $gAcc(\\alpha)$ emphasizes to novel class while $\\alpha \\to 1$ , $gAcc(\\alpha)$ emphasizes the base class performance, i.e., the larger the $\\alpha$ , the greater the influence of the base class on the value of $gAcc(\\alpha)$ . To evaluate the performance of a model on a specific task $\\mathcal{T}_i$ , a standalone value is needed. Thus, we introduce the area under the curve (AUC) of $gAcc_i(\\alpha)$ along the parameter $\\alpha$ as a comprehensive metric:", + "bbox": [ + 214, + 574, + 787, + 731 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\ng A c c _ {i} = \\int_ {0} ^ {1} g A c c _ {i} (\\alpha) d \\alpha . \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 411, + 734, + 785, + 768 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Similar to $aA_{cc}$ , for multiple tasks, we average the $gA_{cc_i}$ for each task and get an overall metric $gA_{cc}$ for the whole learning process:", + "bbox": [ + 214, + 773, + 782, + 803 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\ng A c c = \\frac {1}{n _ {t}} \\sum_ {i} ^ {n _ {t}} g A c c _ {i}. \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 431, + 809, + 785, + 842 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "Rethinking Few-shot Class-incremental Learning: Learning from Yourself", + "bbox": [ + 245, + 114, + 732, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/a53e9f37d5a627240f619f39ff48fc55fcbfd334b536fff37414fda673768647.jpg", + "image_caption": [ + "(b) $gAcc_{i}$ curve w.r.t Task i" + ], + "image_footnote": [], + "bbox": [ + 259, + 143, + 576, + 292 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/f0b4f97fa5743c99779712f70fd62bb29d5c2ae4964f5a990daa41f26436e59b.jpg", + "image_caption": [ + "Fig. 2: FSCIL performances are shown in generalized accuracy. (a): The $gAcc$ curve (averaged across all $n_t$ tasks) v.s. param $\\alpha$ . (b): The $gAcc$ AUC of each task $\\mathcal{T}_i$ (Eq. (5)). In the legend, we show the average AUC across tasks(Eq. (6)) of each method. We evaluate recent works including SAVC [49], NC [59], and Alice [42] on the miniImageNet dataset. Some corner cases: Lazy: The model maintains base-task performance while refusing to learn anything from novel tasks. Greedy: Regardless of the previous knowledge, the model only greedily focuses on the current task. Greedy-NF: A Non-Forget version of 'Greedy'. See more about these cases in our supplementary material." + ], + "image_footnote": [], + "bbox": [ + 578, + 167, + 727, + 292 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The proposed $gAcc$ provides different evaluation perspectives using different $\\alpha$ so we believe such a metric can provide a more balanced evaluation of FSCIL.", + "bbox": [ + 212, + 385, + 787, + 415 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.3 Analysis", + "text_level": 1, + "bbox": [ + 214, + 422, + 333, + 438 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We further study the necessity of the proposed $gAcc$ . We evaluate recent SOTAs in FSCIL including SAVC [49], Alice [42] and NC-FSCIL [59], and the results are in Fig. 2. We show that the $gAcc$ provides more fine-grained evaluation than the classical accuracy metric. For example, in the left part of Fig. 2, although the $gAcc(1)$ , which is equal to the $aAcc$ , of SAVC is close to that of NC-FSCIL, the area under the curve(AUC) of SAVC is much smaller than NC-FSCIL. This indicates that the novel class performance is much worse for SAVC, which cannot be discovered by conventional metrics.", + "bbox": [ + 212, + 446, + 787, + 566 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Furthermore, to show the superiority of our metric, we consider three corner cases: Lazy, Greedy, and Greedy-NF. 'Lazy' stands for a type of algorithm that only focuses on maintaining base-class performance, i.e. $\\forall i \\geq j > 1$ , $A_{i}^{j} = 0$ , $A_{i}^{1} = A_{1}^{1}$ . By setting $A_{1}^{1} = 85\\%$ , we can see that when $\\alpha = 1$ (i.e. the traditional aAcc metric), its performance is close to SOTA methods. Further, when considering popular forgetting metrics like performance dropping $(PD = A_{i}^{j} - A_{j}^{j})$ and Knowledge rate $(KR = A_{i}^{j} / A_{j}^{j})$ , because the forgetting is zero, its PD is 0 and KR is 1. All these metrics indicate such 'Lazy' model is superior, which is contradictory to common sense. However, our proposed gAcc reveals that its performance is far from satisfying by setting $\\alpha < 1$ and its AUC is far smaller than other methods. In other words, using the traditional metrics will indicate algorithms like 'Lazy' achieving SOTA performance while such methods perform poorly evaluated by our proposed gAcc. The gAcc can also discover other corner cases. 'Greedy' is to greedily learn the current knowledge (i.e. $A_{1}^{1} = 85\\%$ , $A_{i}^{j} = 100\\%$ , when $i = j$ else 0). \"Greedy\" achieves low both aAcc and gAcc. We further eliminate the forgetting problem from 'Greedy' and define 'Greedy-NF' on top of it: set accuracy $A_{i}^{j}$ to $100\\%$ for $\\forall j > 1, j < i$ (no forgetting at novel classes). We find that although 'Greedy-NF' performs", + "bbox": [ + 212, + 568, + 787, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "Tang et al.", + "bbox": [ + 271, + 114, + 346, + 128 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/5c2caab1b7a2d3e639325541be9808aa8092de4b0332b1a53223ef1eb52dc1ca.jpg", + "image_caption": [ + "Fig. 3: t-SNE [37] visualization of miniImageNet test set features from different layers of a ViT. Left: we choose layers $L_{9}$ , $L_{10}$ , $L_{11}$ , $L_{12}$ (final) and show both base classes and novel classes features. It is observed that shallow features are more dispersed than deep features. Right: Accuracies of different tasks on features from various layers. It is clear that intermediate features achieve better performance on novel classes and thus have better generalization ability. Best viewed in color." + ], + "image_footnote": [], + "bbox": [ + 290, + 143, + 522, + 313 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/879e596e7a8f7f01e548a584ae9d6152204d8d03633b5bd54d3bd86227312a83.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 547, + 148, + 715, + 310 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "well when majorly considering the novel-class performance $(\\alpha \\in \\{0, \\frac{1}{12}, \\frac{2}{12}, \\frac{3}{12}\\})$ , the overall performance (AUC) is not satisfying. Please refer to Sec. A1 in the supplementary material for more details about these corner cases.", + "bbox": [ + 212, + 388, + 782, + 433 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "By analyzing these cases, we show the superior properties of the proposed $gAcc$ . Firstly, the $gAcc$ metric effectively distinguishes methods (e.g. 'Lazy') that only retain the base-task performance, thereby encouraging methods to learn from novel classes. Besides, $gAcc$ avoids overemphasis on the performance of base classes or novel classes (corner case 'Greedy-NF' is still far worse than normal methods in terms of AUC), providing a more balanced evaluation outcome. In other words, the proposed $gAcc$ metric provides objective and balanced results when evaluating FSCIL methods, eliminating 'shortcuts' in $aAcc$ that focus only on the performance of the base classes.", + "bbox": [ + 212, + 435, + 784, + 569 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In conclusion, the $gAcc$ curve and the AUC are necessary and powerful tools to evaluate the FSCIL methods.", + "bbox": [ + 212, + 571, + 784, + 599 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4 Learn from Yourself", + "text_level": 1, + "bbox": [ + 214, + 614, + 449, + 632 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.1 Preliminary and The FSCIL Setting", + "text_level": 1, + "bbox": [ + 214, + 643, + 562, + 660 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "For task $\\mathcal{T}_i\\in \\{\\mathcal{T}_i\\}_{i = 1}^T$ , the training set can be denoted as $\\mathcal{D}_{train}^{i} = \\{x_{t}^{i},y_{t}^{i}\\}_{t = 1}^{m^{i}}$ . For FSCIL settings, the base task $\\mathcal{T}_1$ contains $|\\mathcal{V}_1|$ classes and sufficient samples for training. For $\\forall i > 1$ , there are $|\\mathcal{V}_{novel}|$ classes for each task and $k$ samples per class. Typically, $|\\mathcal{V}_{novel}|$ is set to 5, and $k = 5$ , i.e. a 5-way 5-shot setting. We rephrase the classification model $f(\\cdot)$ in Sec. 3 as $f(\\cdot) = \\varPhi (f_B(\\cdot))$ where $\\varPhi (\\cdot)$ is the classifier and $f_{B}(\\cdot)$ is the backbone. We follow previous works to use a cosine classifier: $\\varPhi(X)=P^\\top\\overline{X}$ , where $P\\in \\mathbb{R}^{d\\times |\\mathcal{V}|}$ and $\\overline{X}\\in \\mathbb{R}^{d\\times 1}$ are the L2-normalized inputs, and $d$ is the feature dimension of $f_{B}(\\cdot)$ , $|\\mathcal{V}| = \\sum_{i}|\\mathcal{V}_{i}|$ is the number of classes seen so far. In this work, we employ the Vision Transformer(ViT) [15] in the FSCIL setting. For a ViT with $N_{L}$ layers $\\{L_l\\}_{l = 1}^{N_L}$ , given a input image $x$ , the output can be denoted as $X_{N_L} = L_{N_L}(L_{N_L - 1}(\\dots L_1(x))$ .", + "bbox": [ + 212, + 672, + 787, + 843 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "Rethinking Few-shot Class-incremental Learning: Learning from Yourself", + "bbox": [ + 245, + 113, + 732, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 114, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/328db429ab746e1ab8b477ee54a346b95e253da5d3f0de2c69c16cde27945ca8.jpg", + "image_caption": [ + "Fig. 4: The overall framework of our method. The Feature Rectification (FR) module takes the final and the intermediate features as input and outputs the rectified features for classification. The relation transfer losses are designed to convey valuable information from intermediate features to the rectified ones. The details of these two parts are shown in the left and lower parts of the figure. Further, a cosine constraint is proposed to maintain the base class performance and a classification loss is employed to adapt to the novel classes. Best viewed in color." + ], + "image_footnote": [], + "bbox": [ + 217, + 143, + 785, + 309 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Baseline. Recent works [3, 20, 24, 38, 47, 49, 59, 63, 76] suggest a training-frozen scheme, which trains the backbone $f_{B}(\\cdot)$ and the classifier $\\varPhi(\\cdot)$ (if available) on the base-task only, and then freeze the backbone during novel tasks. For novel classes, the classifier is extended by mean features (i.e. prototypes) extracted by the frozen backbone. We follow and employ the scheme as our baseline.", + "bbox": [ + 212, + 393, + 787, + 468 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Main idea of our method. Given that the backbone is initially trained for the base task and remains fixed during novel tasks, the main challenges lie in the potential overfitting on the base task and the poor feature extraction ability of the novel classes. We conduct a toy experiment to show these problems and show that the intermediate feature can convey valuable information to tackle these problems. Results are shown in Fig. 3, where we select 4 base classes and 5 novel classes and utilize the t-SNE [37] to visualize the features from different layers. It can be concluded that features extracted by the base-classes trained backbone are overfitted to the base classes and cannot be generalized well on novel classes, i.e. the final features are well aggregated, but it is hard to distinguish among unseen classes. (see $L_{12}$ features on the left and the corresponding performance on the right). Compared with the final features (i.e. $L_{12}$ ), the shallower intermediate features $(L_9, L_{10}, L_{11})$ are more dispersed, which makes them more discriminative for unseen classes. This is cross-verified by the performance shown on the right side of Fig. 3. We can see that using intermediate features can largely boost the novel classes' performance (up to $15\\%$ improvement) with a price of performance decline of the base classes. Based on this observation, we propose to rectify the final feature based on the intermediate feature to better handle both base and novel classes.", + "bbox": [ + 212, + 469, + 787, + 755 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.2 Feature Rectification", + "text_level": 1, + "bbox": [ + 215, + 781, + 434, + 796 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We propose a Feature Rectification (FR) module to rectify the final feature based on the relation information from the intermediate features. Specifically,", + "bbox": [ + 212, + 809, + 785, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "Tang et al.", + "bbox": [ + 271, + 114, + 346, + 128 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "given a pair of final feature $X_{N_L}$ and the intermediate feature $X_l$ , we design an FR module conditioned on the intermediate features:", + "bbox": [ + 214, + 146, + 782, + 175 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\nX _ {F R} = M _ {m i x} \\left(\\operatorname {c a t} \\left(M _ {F} \\left(X _ {N _ {L}}\\right), M _ {I} \\left(X _ {l}\\right)\\right), \\right. \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 364, + 186, + 785, + 203 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where 'cat' means the concatenate operation. $M_{\\text{mix}}$ , $M_F$ and $M_I$ are 2-layer MLPs. By introducing the intermediate features, the FR module has access to a reference for extracting the valuable structural information lies in the intermediate features. A naive way to mine such information is to directly pull rectified features closer to shallow features on an instance-by-instance basis. Although such an instance-wise loss can bridge shallow features and rectified features, directly pulling the rectified feature to the shallow one could lead to noisy features since the shallow features are more scattered, as shown in Sec. 4.1.", + "bbox": [ + 212, + 210, + 782, + 330 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "To effectively transfer discriminative cues among samples from shallow features to rectified features and avoid introducing noise, we propose two relation-based transfer losses at different levels. Firstly, we offer a Instance-level Relation transfer (IR) to mine the relation between feature instances. Given a pair of final features $X_{N_L}^1$ and $X_{N_L}^2$ , and another pair of intermediate features $X_l^1$ and $x_{l}^{2}$ , the IR loss is defined as:", + "bbox": [ + 214, + 332, + 782, + 422 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {I R} = \\delta \\left(\\epsilon \\left(X _ {F R} ^ {1}, X _ {F R} ^ {2}\\right), \\epsilon \\left(X _ {l} ^ {1}, x _ {l} ^ {2}\\right)\\right), \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 380, + 431, + 784, + 449 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where $\\epsilon (\\cdot)$ is the Euclidean distance and the $\\delta (\\cdot)$ denotes the smooth L1 loss. Besides the relation between feature instances, another level of the relation information is the relation between features and class centers. Class-center Relation Transfer (CR) is proposed to mine this class-center-level information. Assume at task $T_{i}$ with $|\\mathcal{V}_i|$ classes and the feature dimension is $d$ . Given a set of class prototypes $P\\in \\mathbb{R}^{d\\times |\\mathcal{V}_i|}$ from the classifier $\\varPhi(\\cdot)$ as well as a pair of final feature $X_{N_L}\\in \\mathbb{R}^{d\\times 1}$ and the intermediate feature $X_{l}$ $(l\\in [1,N_L))$ , the corresponding CR loss is formulated as:", + "bbox": [ + 214, + 457, + 787, + 578 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {C R} = \\mathrm {K L} \\left(\\left(\\sigma \\left(P ^ {\\top} \\overline {{X _ {F R}}}\\right), \\sigma \\left(P ^ {\\top} \\overline {{X _ {l}}}\\right)\\right), \\right. \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 375, + 587, + 784, + 604 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where $\\sigma (\\cdot)$ is the softmax function, and $\\mathrm{KL}(\\cdot)$ is the Kullback-Leibler divergence. $\\overline{X}$ denotes the L2 normalization results of $X$ .", + "bbox": [ + 214, + 613, + 782, + 642 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "In the above equations, the two losses $\\mathcal{L}_{CR}$ and $\\mathcal{L}_{IR}$ work on different levels and serve different roles. $\\mathcal{L}_{CR}$ works on the relations between a sample and all class centers, focusing on the coarse-grained class-specific patterns that distinguish different classes. Differently, $\\mathcal{L}_{IR}$ provides knowledge about fine-grained discriminative cues that contribute to pairwise relations. The proposed relation-based transfer losses are related to previous knowledge distillation works [5, 40, 41]. However, these works do not share the same objective as ours. Our goal is to extract precious transferrable information hidden in various shallow layers hierarchically to rectify the final features while prior distillation methods focus on matching final outputs to teachers. Furthermore, there exists a branch of works using intermediate features [36, 62] directly for dense prediction. Differently, we are the first to extract intermediate knowledge to rectify the final feature for a classification task.", + "bbox": [ + 214, + 643, + 787, + 839 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "Rethinking Few-shot Class-incremental Learning: Learning from Yourself", + "bbox": [ + 245, + 114, + 732, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 114, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Besides, we constrain the discriminative ability of rectified features on base classes to avoid negative rectification. Since the final feature of layer $L_{N_L}$ is strong enough for the base classes, we apply a cosine constraint as:", + "bbox": [ + 212, + 146, + 787, + 191 + ], + "page_idx": 9 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {c o s}} = \\cos \\left(X _ {F R}, X _ {N _ {L}}\\right), \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 418, + 200, + 785, + 215 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "where $\\cos (\\cdot ,\\cdot)$ is cosine similarity.", + "bbox": [ + 212, + 223, + 457, + 238 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Novel class training. To ensure the rectified features are centered around the class mean prototypes, we further introduce a classification loss for novel classes:", + "bbox": [ + 212, + 239, + 782, + 268 + ], + "page_idx": 9 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {N o v C E}} = - \\mathbb {1} (y \\in \\cup_ {n > 1} \\mathcal {Y} _ {n}) \\log \\sigma \\left(P ^ {\\top} \\overline {{X _ {F R}}}\\right) ^ {(y)}, \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 338, + 282, + 785, + 299 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "where $y$ is the corresponding ground true label of $X_{FR}$ , $(P^{\\top}\\overline{X_{FR}})^{(y)}$ is the $y$ -th element of $(P^{\\top}\\overline{X_{FR}})$ .", + "bbox": [ + 212, + 304, + 782, + 335 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Anti-forgetting training. Another challenge of FSCIL is the forgetting problem. Thanks to our feature rectification design, the only trainable module is the FR module. As the FR module takes the final and the intermediate feature as the input, these kinds of features can be easily modeled by Gaussian distributions [51,71]. Hence, we construct a Gaussian distribution for each class on each layer. The Gaussian distribution of class $i$ features at layer $L_{l}$ can be denoted as $\\mathcal{N}_i^l (\\mu_i^l,\\Sigma_i^l)$ , where $\\Sigma_i^l$ is obtained by calculating the covariance of each feature dimension if $i\\in \\mathcal{Y}_1$ (base classes). For novel classes, the covariance matrix is computed by averaging the covariance matrices of the top-k most similar base classes. During the incremental learning phases, we utilize these Gaussian distributions to sample features, which serve as inputs from old tasks to the FR module, assisting the FR module in retaining previously learned knowledge.", + "bbox": [ + 212, + 335, + 784, + 515 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Overall training objectives. The total loss function for our feature rectification framework is as follows:", + "bbox": [ + 212, + 515, + 782, + 545 + ], + "page_idx": 9 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\beta_ {\\text {c o s}} \\left(\\mathcal {L} _ {\\text {c o s}} + \\mathcal {L} _ {\\text {N o v C e}}\\right) + \\beta_ {C R} \\mathcal {L} _ {C R} + \\beta_ {I R} \\mathcal {L} _ {I R}, \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 331, + 554, + 784, + 570 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "where $\\beta_{cos},\\beta_{CR}$ and $\\beta_{IR}$ are trade-off parameters.", + "bbox": [ + 212, + 577, + 578, + 592 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Multi-layer knowledge ensemble. Since there exists hierarchical knowledge in different layers, we mine such information by knowledge ensemble from different layers. Specifically, we incorporate an FR module with each selected layer, resulting in a total of $n_{FR}$ FR modules. During the training phase, these FR modules are optimized simultaneously. When evaluation, FR modules at layer $l$ are employed to produce multiple rectified features denoted as $X_{FR}^{l}$ . These rectified features are subsequently input into the shared classifier $P$ for generating multiple predictions: $\\mathrm{Pred}_l = \\sigma (P^\\top \\overline{X_{FR}^l})$ . The final prediction is computed as the average of each FR-branch: $\\mathrm{Pred} = \\frac{1}{n_{FR}}\\sum_{l = 1}^{n_{FR}}\\mathrm{Pred}_l$ .", + "bbox": [ + 212, + 593, + 787, + 734 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "5 Experiments", + "text_level": 1, + "bbox": [ + 214, + 750, + 375, + 767 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "5.1 Experiment Setup", + "text_level": 1, + "bbox": [ + 214, + 772, + 413, + 787 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Datasets. Following previous works, we conduct experiments on three popular datasets of FSCIL: CIFAR-100 [28], miniImageNet [46] and CUB-200 [55]. For statistics of the three datasets, please refer to the supplementary material.", + "bbox": [ + 212, + 794, + 785, + 840 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "Tang et al.", + "bbox": [ + 271, + 114, + 346, + 128 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/16ed78749bd485f56215e95935f94559d6bd7f5e0b4b85b4622c17378739ef06.jpg", + "table_caption": [ + "Table 1: Comparison with FSCIL methods on miniImageNet. The best and second best results are marked bolded and underlined. Methods with†are reported with the original results in their paper. Other results are reimplemented by us." + ], + "table_footnote": [], + "table_body": "
MethodsNum. of \nParams.(M)aAcc in each task (%)
123456789aAccgAcc
TOPIC [52]†-61.350.145.241.237.535.532.229.524.439.64-
IDLVQ [6]†-64.859.855.952.649.947.644.843.141.851.16-
MetaFSCIL [8]†-72.068.063.860.357.655.253.050.849.258.85-
DF-Replay [35]†-71.867.163.259.857.054.051.649.548.258.02-
FCIL [18]†-76.3471.467.164.161.358.555.754.152.862.37-
Self-promoted [73]12.3463.664.159.555.552.149.245.643.140.952.6241.53
CEC [63]12.1272.167.063.059.656.953.951.449.547.857.9148.64
LIMIT [70]12.2973.367.663.259.556.753.751.449.748.258.1549.47
FACT [69]11.2676.371.366.962.959.656.353.551.149.260.7849.56
CLOM [76]14.2272.166.762.759.256.153.050.248.346.757.2346.11
C-FSCIL [20]12.7970.663.859.356.454.351.848.947.346.155.4050.46
BiDistFSCIL [66]11.5074.670.466.662.860.557.254.753.051.961.3253.49
S3C [24]0.3477.072.168.265.162.058.856.053.651.962.7453.70
Regularizer [3]26.3081.972.768.764.157.453.553.852.343.162.8255.62
SubNetwork [25]22.3878.173.569.866.263.260.357.655.754.564.3155.88
SAVC [49]12.1281.076.271.971.465.662.159.257.055.566.6656.35
ALICE [42]11.2081.171.167.864.162.460.157.956.856.064.1459.72
NC-FSCIL [59]15.8784.276.172.168.167.363.961.959.657.967.9060.93
YourSelf (Ours)11.5784.077.673.770.068.064.962.159.859.068.8062.20
", + "bbox": [ + 253, + 178, + 738, + 397 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Training and testing protocols. For training protocols, we follow prior works [3, 20, 24, 25, 42, 49, 59, 66] to split a large proportion of the training set to serve as the base task, and split the rest into small incremental tasks. Specifically, for CIFAR-100 and miniImageNet, the first task contains 60 classes out of the total 100 classes, and the rest 40 classes are split into 8 tasks, following a 5-way 5-shot scheme. For CUB-200, there are 100 classes in the first task and the number of novel tasks is 10 (in a 10-way 5-shot setting). As for the testing, we follow prior works to report the average accuracy ( $aAcc$ ). In addition, we also provide detailed discussions on our metric generalized accuracy ( $gAcc$ ). We further provide detailed results of novel classes only in our supplementary material.", + "bbox": [ + 212, + 400, + 792, + 551 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Implementation details. For all experiments below, we use the ViT suggested by Touvron et, al. [53]. We adjust stuff like the dimensions of the backbone to match the number of parameters (around 10M) as the ResNet18 [19] which is widely used in this topic. The Feature Rectification (FR) is a lightweight module and uses popular MLPs implementation [53] with only 0.6M parameters. To ensure a fair comparison, we also re-implement other methods with the same backbone as ours, but found the performance is rarely improved. Please refer to Supplementary material for more details. We follow previous works to train from scratch on CIFAR-100 and miniImageNet dataset. Prior works [6,8,35, 52,63,73] adopt an ImageNet-pretrained ResNet for the CUB-200 dataset. To make a fair comparison, we also pretrained our backbone on the ImageNet [12] to evaluate on CUB-200. For evaluation, we reproduced 14 methods based on the public codes and conducted in-depth comparisons in the following sections.", + "bbox": [ + 212, + 551, + 789, + 750 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "5.2 Comparison with State-of-the-art methods", + "text_level": 1, + "bbox": [ + 214, + 768, + 616, + 786 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Experimental results on miniImageNet are shown in Table 1. We report the classical accuracy ( $aAcc$ ) in each section as well as its averaged results and our proposed $gAcc$ . From the table, we can conclude that our proposed method", + "bbox": [ + 212, + 794, + 787, + 840 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "Rethinking Few-shot Class-incremental Learning: Learning from Yourself", + "bbox": [ + 245, + 114, + 732, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/b03079e837c96ea84661f614d3031b6aac658f94c69deddf5e8be3102f2d926c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 217, + 147, + 356, + 290 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/8076db5bbcad98b8b6136d43e0dc1d3cffc6a7eb1985f95e93b7a84eaef035db.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 359, + 150, + 496, + 291 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/c8680caf43746db8689de72a8a742b845a45a03f6dcf54a2f1e992f262a1b005.jpg", + "image_caption": [ + "Fig. 5: FSCIL results on miniImageNet. (a): The $gAcc$ curve (averaged across all $n_t$ tasks) vs. $\\alpha$ . (b): The $gAcc$ AUC of each task $\\mathcal{T}_i$ (Eq. (5)) vs. i. In the legend, we show the AUC value (Eq. (6)) of each method. Best viewed in color." + ], + "image_footnote": [], + "bbox": [ + 218, + 291, + 501, + 319 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/d87760f4c26634ceadf17dbaa115231f4b0a64d14a49926304d0a7af4e2021e8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 504, + 143, + 645, + 291 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/ca3f2436ef1d469c25d320d169fed211277cd1d4c018beaa09d71c415d596046.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 647, + 146, + 785, + 291 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/017231e1eb44ee990632f51a22c165f788e23cee0c4448460a125ecbed804d20.jpg", + "image_caption": [ + "Fig. 6: FSCIL results on CUB-200. (a): The $gAcc$ curve (averaged across all $n_t$ tasks) vs. $\\alpha$ . (b): The $gAcc$ AUC of each task $\\mathcal{T}_i$ (Eq. (5)) vs. $i$ . In the legend, we show the AUC value (Eq. (6)) of each method. Best viewed in color." + ], + "image_footnote": [], + "bbox": [ + 545, + 292, + 754, + 324 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "outperforms other methods by a large margin. Specifically, under the classical evaluation metric $aAcc$ , the proposed method achieves 68.80% accuracy, which is 0.90%, 4.66%, 2.14% higher than SOTA methods NC-FSCIL [59], ALICE [42] and SAVC [49]. The improvement is non-trivial considering that miniImageNet is a complex dataset.", + "bbox": [ + 212, + 397, + 784, + 472 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "From a more balanced perspective, we also report $gAcc$ on the Table 1 and Fig. 5. In the left figure of Fig. 5, except for the method ALICE when $\\alpha = 0$ , our method outperforms most other methods by a clear margin (at least $1.27\\%$ ) and achieves the highest AUC of all methods. This indicates that our method achieves a more balanced base-novel performance. To see raw numbers of the accuracy of novel classes, please refer to the supplementary material.", + "bbox": [ + 212, + 474, + 787, + 565 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Besides, Fig. 5 shows that methods that perform well at $\\alpha = 1$ do not necessarily perform well when $\\alpha$ is smaller. For example, although CLOM achieves $57.23\\%$ average accuracy ( $\\alpha = 1$ ) and is $1.83\\%$ higher than method C-FSCIL ( $55.4\\%$ ), we can see the performance drops significantly when $\\alpha$ is small. It is mainly because CLOM focuses on the base-classes accuracy and performs poorly on novel classes. We can observe that the AUC of the CLOM is much smaller ( $4.35\\%$ ) than that of C-FSCIL. The results confirm the necessity of our proposed $gAcc$ . Similar phenomena can be found when comparing other methods like SAVC and ALICE, and we can obtain similar conclusions on the other two datasets including CUB-200 and CIFAR-100.", + "bbox": [ + 212, + 566, + 787, + 715 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Results on CUB-200 dataset are shown in Table 2 and Fig. 6. We can see from Table 2 that our method is better than any other existing methods in terms of $aAcc$ and $gAcc$ (1.52%@aAcc and 2.6%@gAcc improvement over the strong method SAVC). Results on CIFAR-100 dataset are shown in Table 3. For the traditional metric $aAcc$ , our method achieves comparable performance with SOTA methods, while our method outperforms all other methods by a clear margin when considering the newly introduced metric $gAcc$ . Please refer to Sec A5 of the supplementary material for more experiment results.", + "bbox": [ + 212, + 719, + 787, + 839 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "Tang et al.", + "bbox": [ + 271, + 114, + 346, + 128 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/d3d479cead14b913c3f2ed19d89771473caea4eefa1f1138a2ac7c4c0f3be90b.jpg", + "table_caption": [ + "Table 2: Comparison with FSCIL methods on CUB-200. The best and second best results are marked bolded and underlined. Methods with $\\dagger$ are reported with the original results in their paper. Other results are reimplemented by us." + ], + "table_footnote": [], + "table_body": "
MethodsNum. of Params.(M)aAcc in each task (%)
1234567891011aAccgAcc
TOPIC [52]†-68.742.554.850.045.341.438.435.432.228.326.343.92-
DF-Replay [35]†-75.972.168.663.862.659.157.855.955.053.652.461.52-
BiDistFSCIL [66]†-79.175.472.869.167.565.164.063.561.961.560.967.34-
MetaFSCIL [8]†-75.972.468.864.863.060.058.356.854.853.852.661.93-
SubNetwork [25]†-78.774.671.467.565.462.661.159.457.557.256.764.68-
DSN [58]†-76.172.269.666.764.462.160.258.957.055.154.263.31-
Self-promoted [73]12.3662.654.751.647.344.340.938.839.035.935.133.544.0740.88
CEC [63]12.3475.771.468.164.062.058.857.355.553.653.252.161.1754.16
LIMIT [70]12.3379.576.473.269.267.164.362.861.959.859.157.666.5859.40
CLOM [76]18.8979.575.972.068.265.763.462.060.358.557.956.965.6157.40
FACT [69]11.3476.173.470.466.564.862.561.159.957.957.556.550.2957.93
NC-FSCIL [59]15.9181.176.472.869.366.163.762.261.059.058.057.366.1759.22
ALICE [42]41.7175.870.168.465.263.761.260.759.758.558.758.463.7760.28
S3C [24]13.0680.276.572.969.166.964.362.761.059.559.258.366.5361.05
SAVC [49]24.2981.177.675.071.269.666.765.864.862.962.762.268.3362.90
YourSelf (Ours)11.9083.477.075.372.269.066.866.065.664.164.563.669.8565.50
", + "bbox": [ + 254, + 176, + 743, + 349 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/c7223bf47c26a2d7314510acf2df5b225cfc5f96e0206a91a1f11762077b975e.jpg", + "table_caption": [ + "Table 3: Comparison results on CIFAR-100. The best and second best results are marked bolded and underlined. All results of other methods are reimplemented by us. See more results in the supplementary material." + ], + "table_footnote": [], + "table_body": "
MethodsNum. of Params.(M)aAcc in each task (%)aAccgAcc
123456789
CEC [63]0.29673.269.165.461.258.055.553.251.349.259.5650.85
LIMIT [70]0.29676.171.767.363.260.157.355.253.050.861.6452.16
CLOM [76]0.35574.269.866.162.259.056.154.051.949.860.3550.36
C-FSCIL [20]12.7976.872.067.263.159.756.754.451.949.761.2850.62
SAVC [49]0.84278.472.768.464.061.258.356.054.151.762.7653.29
FACT [69]0.28278.572.768.864.561.358.656.854.452.363.1254.56
SubNet [25]22.3980.075.771.667.864.761.459.357.155.164.8556.63
S3C [24]0.34178.073.970.266.163.360.158.356.754.064.5158.04
ALICE [42]45.2880.371.967.063.260.758.357.355.553.763.1059.72
NC-FSCIL [59]15.9382.977.573.769.065.662.059.757.855.467.0659.73
Ours11.6882.976.372.967.865.262.060.758.856.667.0260.32
", + "bbox": [ + 292, + 385, + 710, + 516 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "5.3 Ablation studies", + "text_level": 1, + "bbox": [ + 215, + 521, + 398, + 535 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Effectiveness of different losses. Ablation experiments of the effectiveness of $\\mathcal{L}_{IR}$ and $\\mathcal{L}_{CR}$ are conducted among CIFAR-100, miniImageNet, and CUB-200 datasets. Experimental results are shown in Table 4 and Table 5. The detailed $gAcc$ vs. $\\alpha$ as shown in Fig. 7. Results on CIFAR-100 show that introducing the FR module can effectively enhance the $gAcc$ by $3.28\\%$ . After further introducing the class-relation (CR) transfer loss, the performance at $gAcc$ is further improved by $1.07\\%$ . Adding the instance-relation (IR) transfer loss will further boost the $gAcc$ by $1.91\\%$ . The effectiveness of the proposed losses is also verified on the miniImageNet dataset. For miniImageNet dataset, the introduction of FR modules brings $3.02\\%$ improvement at $gAcc$ . Applying loss $\\mathcal{L}_{CR}$ and $\\mathcal{L}_{IR}$ can improve $gAcc$ by $0.49\\%$ and $0.82\\%$ , respectively. It is worth mentioning that $\\mathcal{L}_{CR}$ and $\\mathcal{L}_{IR}$ can also boost the $aAcc$ . Specifically, on the miniImageNet, adding these two losses results in an increase at $aAcc$ performance of $1.96\\%$ . On CIFAR-100, the $aAcc$ is improved by about $0.76\\%$ with the help of these two losses. See more experiments in the supplementary material.", + "bbox": [ + 212, + 551, + 787, + 777 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Effectiveness of multi-layer knowledge ensemble. In Sec. 4.2, we propose to exploit hierarchical knowledge from multiple layers for effective feature rectification. In practice, we select layers 8, 9 and 10 for the ensemble. Here, we take a more detailed discussion on each layer before the ensemble of multiple", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "Rethinking Few-shot Class-incremental Learning: Learning from Yourself", + "bbox": [ + 245, + 114, + 732, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/0aeb4b0bb7c1962593998918dc5e97a214fb980670be95f7f6a52b71543accd2.jpg", + "table_caption": [ + "Table 4: Ablation studies on CIFAR-100." + ], + "table_footnote": [], + "table_body": "
FR\\( {\\mathcal{L}}_{\\text{Cos }} \\)\\( {\\mathcal{L}}_{\\text{NovCE }} \\)\\( {\\mathcal{L}}_{CR} \\)\\( {\\mathcal{L}}_{IR} \\)aAcc. in each task (%)aAccgAcc
123456789
82.977.872.968.464.661.458.756.454.166.3554.20
✓✓82.976.173.167.964.860.859.257.253.966.2657.48
✓✓✓82.976.172.467.864.561.359.657.354.966.3358.55
✓✓✓✓82.975.772.567.464.861.960.458.756.367.0260.32
", + "bbox": [ + 199, + 178, + 552, + 224 + ], + "page_idx": 13 + }, + { + "type": "table", + "img_path": "images/21811813346d39895fb7a37b939ab0a5be4e674ab8f6be0dd14cf5e6db85d71d.jpg", + "table_caption": [ + "Table 5: Ablation studies on miniImageNet." + ], + "table_footnote": [], + "table_body": "
FRLcos+ LNoveCELCRLIRaAcc. in each task (%)aAccgAcc
123456789
84.079.475.071.167.864.461.358.856.968.7457.87
✓✓84.074.069.766.866.163.560.858.658.166.8460.89
✓✓✓84.077.773.669.967.464.661.659.858.768.6061.38
✓✓✓✓84.077.673.770.068.064.962.159.859.068.8062.20
", + "bbox": [ + 199, + 238, + 552, + 287 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/692d5722960dfe2f8914e6cd436a9210d4d5f02e817c8618a93feae34e9307d7.jpg", + "image_caption": [ + "Fig.7: Ablation studies on two datasets, evaluating using $gAcc$ ." + ], + "image_footnote": [], + "bbox": [ + 553, + 162, + 681, + 275 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/766a7f88dc627d7e4edf055268d6d7bd618febf9dd6f7f39e729e2290995840e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 681, + 162, + 808, + 275 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "layers. We conducted experiments on the miniImageNet, and the results are shown in Table 6. It is clear that for each layer, the corresponding FR module effectively calibrates the features and achieves improvement from $2.23\\%$ to $3.54\\%$ at $gAcc$ . Furthermore, the introduction of the FR module does not harm the $aAcc$ . Considering the fact that different layers contain hierarchical knowledge, the multi-layer ensemble is proposed. The results in the table show that the ensemble is effective in improving the $gAcc$ while maintaining $aAcc$ .", + "bbox": [ + 212, + 305, + 787, + 411 + ], + "page_idx": 13 + }, + { + "type": "table", + "img_path": "images/0559a85061f738971917e2db5cff81e042152d9cd6ca5e4cb2e147b3d47175e8.jpg", + "table_caption": [ + "Table 6: Detailed performance of each layer. Results are from miniImagenet dataset. 'l-Raw': vanilla ViT features from layer $L_{l}$ . 'l-Recti': Rectified features using the proposed FR module.", + "'Ensemble': Ensembled results from $l$ -Recti', $l \\in \\{8, 9, 10\\}$ ." + ], + "table_footnote": [], + "table_body": "
LayersaAcc in each session (%)aAccgAcc
012345678
8-Raw78.973.269.165.462.960.057.355.254.064.0156.61
8-Recti80.773.470.166.664.461.559.357.056.565.5260.15
9-raw82.577.373.269.765.363.760.958.957.567.6659.33
9-Recti83.076.673.269.767.564.661.959.959.068.3961.90
10-raw83.878.674.470.766.464.361.559.357.768.5558.90
10-Recti84.378.274.170.468.264.861.959.558.468.8461.13
Ensemble84.077.673.770.068.064.962.159.859.068.8062.20
", + "bbox": [ + 305, + 489, + 697, + 595 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "6 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 652, + 359, + 669 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In this work, we have revisited the challenges of the FSCIL, showing the limitation of the existing metric and introducing a novel metric denoted as generalized accuracy ( $gAcc$ ) for the comprehensive evaluation of FSCIL methods from a more balanced perspective. Furthermore, we incorporate the ViT into the FSCIL framework and unlock the potential of the ViT intermediate layer features. To achieve this, we have designed a feature rectification (FR) module aimed at integrating the valuable information from the intermediate layer features into the rectified feature space. Using the new metric $gAcc$ , we analyze the existing open-sourced FSCIL methods from a more balanced perspective and show the superiority of our method on three popular datasets.", + "bbox": [ + 212, + 688, + 787, + 840 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "Tang et al.", + "bbox": [ + 271, + 114, + 346, + 128 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 217, + 145, + 401, + 162 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "This work was supported partially by the National Key Research and Development Program of China (2023YFA1008503), NSFC(U21A20471, 62206315), Guangdong NSF Project (No. 2023B1515040025, 2020B1515120085, 2024A15150-10101), Guangzhou Basic and Applied Basic Research Scheme(2024A04J4067).", + "bbox": [ + 212, + 178, + 787, + 239 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 215, + 263, + 321, + 277 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Agarwal, A., Banerjee, B., Cuzzolin, F., Chaudhuri, S.: Semantics-driven generative replay for few-shot class incremental learning. In: Proceedings of the 30th ACM International Conference on Multimedia (2022)", + "2. Ahmad, T., Dhamija, A.R., Cruz, S., Rabinowitz, R., Li, C., Jafarzadeh, M., Boult, T.E.: Few-shot class incremental learning leveraging self-supervised features. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2022)", + "3. Akyurek, A.F., Akyurek, E., Wijaya, D.T., Andreas, J.: Subspace regularizers for few-shot class incremental learning. In: International Conference on Learning Representations (2022)", + "4. Castro, F.M., Marín-Jiménez, M.J., Guil, N., Schmid, C., Alahari, K.: End-to-end incremental learning. In: Proceedings of the European Conference on Computer Vision (2018)", + "5. Chen, H., Wang, Y., Xu, C., Xu, C., Tao, D.: Learning student networks via feature embedding. IEEE Transactions on Neural Networks and Learning Systems (2020)", + "6. Chen, K., Lee, C.G.: Incremental few-shot learning via vector quantization in deep embedded space. In: International Conference on Learning Representations (2020)", + "7. Cheraghian, A., Rahman, S., Fang, P., Roy, S.K., Petersson, L., Harandi, M.: Semantic-aware knowledge distillation for few-shot class-incremental learning. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition (2021)", + "8. Chi, Z., Gu, L., Liu, H., Wang, Y., Yu, Y., Tang, J.: Meta-scil: A meta-learning approach for few-shot class incremental learning. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition (2022)", + "9. Cui, Y., Deng, W., Chen, H., Liu, L.: Uncertainty-aware distillation for semi-supervised few-shot class-incremental learning. IEEE Transactions on Neural Networks and Learning Systems (2023)", + "0. Cui, Y., Deng, W., Xu, X., Liu, Z., Liu, Z., Pietikainen, M., Liu, L.: Uncertainty-guided semi-supervised few-shot class-incremental learning with knowledge distillation. IEEE Transactions on Multimedia (2022)", + "1. Cui, Y., Xiong, W., Tavakolian, M., Liu, L.: Semi-supervised few-shot class-incremental learning. In: IEEE International Conference on Image Processing (ICIP) (2021)", + "2. Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: Imagenet: A large-scale hierarchical image database. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2009)", + "3. Deng, Y., Xiang, X.: Expanding hyperspherical space for few-shot class-incremental learning. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV) (2024)" + ], + "bbox": [ + 225, + 295, + 784, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "Rethinking Few-shot Class-incremental Learning: Learning from Yourself", + "bbox": [ + 245, + 114, + 733, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 785, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "14. Dong, S., Hong, X., Tao, X., Chang, X., Wei, X., Gong, Y.: Few-shot class-incremental learning via relation knowledge distillation. In: Proceedings of the AAAI Conference on Artificial Intelligence (2021)", + "15. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. In: International Conference on Learning Representations (2021)", + "16. Douillard, A., Cord, M., Ollion, C., Robert, T., Valle, E.: Podnet: Pooled outputs distillation for small-tasks incremental learning. In: Proceedings of the European Conference on Computer Vision (2020)", + "17. Fu, S., Yan, J., Gao, Y., Xie, X., Zheng, W.S.: Asag: Building strong one-decoder-layer sparse detectors via adaptive sparse anchor generation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 6328-6338 (2023)", + "18. Gu, Z., Xu, C., Yang, J., Cui, Z.: Few-shot continual infomax learning. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (2023)", + "19. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2016)", + "20. Hersche, M., Karunaratne, G., Cherubini, G., Benini, L., Sebastian, A., Rahimi, A.: Constrained few-shot class-incremental learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2022)", + "21. Hou, S., Pan, X., Loy, C.C., Wang, Z., Lin, D.: Learning a unified classifier incrementally via rebalancing. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2019)", + "22. Ji, Z., Hou, Z., Liu, X., Pang, Y., Li, X.: Memorizing complementation network for few-shot class-incremental learning. IEEE Transactions on Image Processing (2023)", + "23. Jiang, W., Li, D., Hu, M., Zhai, G., Yang, X., Zhang, X.P.: Few-shot class-incremental learning with prior knowledge. arXiv preprint arXiv:2402.01201 (2024)", + "24. Kalla, J., Biswas, S.: S3c: Self-supervised stochastic classifiers for few-shot class-incremental learning. In: European Conference on Computer Vision (2022)", + "25. Kang, H., Yoon, J., Madjid, S.R.H., Hwang, S.J., Yoo, C.D.: On the soft-subnetwork for few-shot class incremental learning. In: International Conference on Learning Representations (2023)", + "26. Kim, D.Y., Han, D.J., Seo, J., Moon, J.: Warping the space: Weight space rotation for class-incremental few-shot learning. In: International Conference on Learning Representations (2023)", + "27. Kim, S., Jeong, Y., Park, J.S., Yoon, S.W.: Mics: Midpoint interpolation to learn compact and separated representations for few-shot class-incremental learning. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (2024)", + "28. Krizhevsky, A., Hinton, G., et al.: Learning multiple layers of features from tiny images. Technical report, University of Toronto (2009)", + "29. Lee, K., Lee, K., Shin, J., Lee, H.: Overcoming catastrophic forgetting with unlabeled data in the wild. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2019)", + "30. Li, J., Dong, S., Gong, Y., He, Y., Wei, X.: Analogical learning-based few-shot class-incremental learning. IEEE Transactions on Circuits and Systems for Video Technology (2024)" + ], + "bbox": [ + 215, + 146, + 785, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "Tang et al.", + "bbox": [ + 271, + 114, + 344, + 127 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "31. Li, Y.M., Huang, W.J., Wang, A.L., Zeng, L.A., Meng, J.K., Zheng, W.S.: Egoexofitness: Towards egocentric and exocentric full-body action understanding. arXiv preprint arXiv:2406.08877 (2024)", + "32. Li, Y.M., Zeng, L.A., Meng, J.K., Zheng, W.S.: Continual action assessment via task-consistent score-discriminative feature distribution modeling. IEEE Transactions on Circuits and Systems for Video Technology (2024)", + "33. Li, Z., Hoiem, D.: Learning without forgetting. IEEE transactions on pattern analysis and machine intelligence (2017)", + "34. Lin, K.Y., Ding, H., Zhou, J., Peng, Y.X., Zhao, Z., Loy, C.C., Zheng, W.S.: Rethinking clip-based video learners in cross-domain open-vocabulary action recognition. arXiv preprint arXiv:2403.01560 (2024)", + "35. Liu, H., Gu, L., Chi, Z., Wang, Y., Yu, Y., Chen, J., Tang, J.: Few-shot class-incremental learning via entropy-regularized data-free replay. In: European Conference on Computer Vision (2022)", + "36. Liu, S., Qi, L., Qin, H., Shi, J., Jia, J.: Path aggregation network for instance segmentation. In: Proceedings of the IEEE conference on computer vision and pattern recognition (2018)", + "37. Van der Maaten, L., Hinton, G.: Visualizing data using t-sne. Journal of machine learning research (2008)", + "38. Mazumder, P., Singh, P., Rai, P.: Few-shot lifelong learning. In: Proceedings of the AAAI Conference on Artificial Intelligence (2021)", + "39. Pan, Z., Yu, X., Zhang, M., Gao, Y.: Ssfe-net: Self-supervised feature enhancement for ultra-fine-grained few-shot class incremental learning. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (2023)", + "40. Park, W., Kim, D., Lu, Y., Cho, M.: Relational knowledge distillation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition (2019)", + "41. Passalis, N., Tzelepi, M., Tefas, A.: Heterogeneous knowledge distillation using information flow modeling. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2020)", + "42. Peng, C., Zhao, K., Wang, T., Li, M., Lovell, B.C.: Few-shot class-incremental learning from an open-set perspective. In: European Conference on Computer Vision (2022)", + "43. Peng, Y.X., Jiao, J., Feng, X., Zheng, W.S.: Consistent discrepancy learning for intra-camera supervised person re-identification. IEEE Transactions on Multimedia 25, 2393–2403 (2022)", + "44. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021)", + "45. Rebuffi, S.A., Kolesnikov, A., Sperl, G., Lampert, C.H.: icarl: Incremental classifier and representation learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2017)", + "46. Russakovsky, O., Deng, J., Su, H., Krause, J., Satheesh, S., Ma, S., Huang, Z., Karpathy, A., Khosla, A., Bernstein, M., et al.: Imagenet large scale visual recognition challenge. International journal of computer vision (2015)", + "47. Shi, G., Chen, J., Zhang, W., Zhan, L.M., Wu, X.M.: Overcoming catastrophic forgetting in incremental few-shot learning by finding flat minima. In: Advances in neural information processing systems (2021)" + ], + "bbox": [ + 212, + 146, + 787, + 840 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "Rethinking Few-shot Class-incremental Learning: Learning from Yourself", + "bbox": [ + 245, + 114, + 733, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "48. Smith, J., Hsu, Y.C., Balloch, J., Shen, Y., Jin, H., Kira, Z.: Always be dreaming: A new approach for data-free class-incremental learning. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (2021)", + "49. Song, Z., Zhao, Y., Shi, Y., Peng, P., Yuan, L., Tian, Y.: Learning with fantasy: Semantic-aware virtual contrastive constraint for few-shot class-incremental learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2023)", + "50. Tang, Y.M., Peng, Y.X., Zheng, W.S.: Learning to imagine: Diversify memory for incremental learning using unlabeled data. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2022)", + "51. Tang, Y.M., Peng, Y.X., Zheng, W.S.: When prompt-based incremental learning does not meet strong pretraining. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (2023)", + "52. Tao, X., Hong, X., Chang, X., Dong, S., Wei, X., Gong, Y.: Few-shot class incremental learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2020)", + "53. Touvron, H., Cord, M., Douze, M., Massa, F., Sablayrolles, A., Jégou, H.: Training data-efficient image transformers & distillation through attention. In: International Conference on Machine Learning (2021)", + "54. Vaze, S., Han, K., Vedaldi, A., Zisserman, A.: Open-set recognition: A good closed-set classifier is all you need? In: International Conference on Learning Representations (2021)", + "55. Wah, C., Branson, S., Welinder, P., Perona, P., Belongie, S.: The caltech-ucsd birds-200-2011 dataset (2011)", + "56. Wang, Q.W., Zhou, D.W., Zhang, Y.K., Zhan, D.C., Ye, H.J.: Few-shot class-incremental learning via training-free prototype calibration. In: Advances in Neural Information Processing Systems (2023)", + "57. Yan, S., Xie, J., He, X.: Der: Dynamically expandable representation for class incremental learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2021)", + "58. Yang, B., Lin, M., Zhang, Y., Liu, B., Liang, X., Ji, R., Ye, Q.: Dynamic support network for few-shot class incremental learning. IEEE Transactions on Pattern Analysis and Machine Intelligence (2022)", + "59. Yang, Y., Yuan, H., Li, X., Lin, Z., Torr, P., Tao, D.: Neural collapse inspired feature-classifier alignment for few-shot class incremental learning. In: International Conference on Learning Representations (2023)", + "60. Yao, G., Zhu, J., Zhou, W., Li, J.: Few-shot class-incremental learning based on representation enhancement. Journal of Electronic Imaging (2022)", + "61. Yu, L., Twardowski, B., Liu, X., Herranz, L., Wang, K., Cheng, Y., Jui, S., Weijer, J.v.d.: Semantic drift compensation for class-incremental learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2020)", + "62. Yuan, Y., Fu, R., Huang, L., Lin, W., Zhang, C., Chen, X., Wang, J.: Hrformer: High-resolution vision transformer for dense predict. Advances in Neural Information Processing Systems (2021)", + "63. Zhang, C., Song, N., Lin, G., Zheng, Y., Pan, P., Xu, Y.: Few-shot incremental learning with continually evolved classifiers. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition (2021)", + "64. Zhang, J., Zhang, J., Ghosh, S., Li, D., Tasci, S., Heck, L., Zhang, H., Kuo, C.C.J.: Class-incremental learning via deep model consolidation. In: Proceedings of the IEEE Winter Conference on Applications of Computer Vision (2020)" + ], + "bbox": [ + 215, + 146, + 784, + 839 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "Tang et al.", + "bbox": [ + 271, + 114, + 344, + 128 + ], + "page_idx": 17 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "65. Zhao, B., Xiao, X., Gan, G., Zhang, B., Xia, S.T.: Maintaining discrimination and fairness in class incremental learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2020)", + "66. Zhao, L., Lu, J., Xu, Y., Cheng, Z., Guo, D., Niu, Y., Fang, X.: Few-shot class-incremental learning via class-aware bilateral distillation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2023)", + "67. Zheng, G., Zhang, A.: Few-shot class-incremental learning with meta-learned class structures. In: International Conference on Data Mining Workshops (ICDMW) (2021)", + "68. Zheng, W.S., Yan, J., Peng, Y.X.: A versatile framework for multi-scene person re-identification. IEEE Transactions on Pattern Analysis and Machine Intelligence (2024)", + "69. Zhou, D.W., Wang, F.Y., Ye, H.J., Ma, L., Pu, S., Zhan, D.C.: Forward compatible few-shot class-incremental learning. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition (2022)", + "70. Zhou, D.W., Ye, H.J., Ma, L., Xie, D., Pu, S., Zhan, D.C.: Few-shot class-incremental learning by sampling multi-phase tasks. IEEE Transactions on Pattern Analysis and Machine Intelligence (2022)", + "71. Zhu, F., Zhang, X.Y., Wang, C., Yin, F., Liu, C.L.: Prototype augmentation and self-supervision for incremental learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2021)", + "72. Zhu, J., Yao, G., Zhou, W., Zhang, G., Ping, W., Zhang, W.: Feature distribution distillation-based few shot class incremental learning. In: the 5th International Conference on Pattern Recognition and Artificial Intelligence (PRAI) (2022)", + "73. Zhu, K., Cao, Y., Zhai, W., Cheng, J., Zha, Z.J.: Self-promoted prototype refinement for few-shot class-incremental learning. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition (2021)", + "74. Zhu, K., Zhai, W., Cao, Y., Luo, J., Zha, Z.J.: Self-sustaining representation expansion for non-exemplar class-incremental learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2022)", + "75. Zhu, M., Zhu, Z., Chen, S., Chen, C., Wu, B.: Enhanced few-shot class-incremental learning via ensemble models. arXiv preprint arXiv:2401.07208 (2024)", + "76. Zou, Y., Zhang, S., Li, Y., Li, R.: Margin-based few-shot class-incremental learning with class-level overfitting mitigation. In: Advances in neural information processing systems (2022)" + ], + "bbox": [ + 215, + 146, + 784, + 632 + ], + "page_idx": 18 + }, + { + "type": "header", + "text": "Rethinking Few-shot Class-incremental Learning: Learning from Yourself", + "bbox": [ + 245, + 114, + 733, + 128 + ], + "page_idx": 18 + }, + { + "type": "page_number", + "text": "19", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 18 + } +] \ No newline at end of file diff --git a/2024/Rethinking Few-shot Class-incremental Learning_ Learning from Yourself/a34c73cf-0382-420e-bbe9-4f775aa32eab_model.json b/2024/Rethinking Few-shot Class-incremental Learning_ Learning from Yourself/a34c73cf-0382-420e-bbe9-4f775aa32eab_model.json new file mode 100644 index 0000000000000000000000000000000000000000..0a227961560417252a2098271f1b5e6e0318ca04 --- /dev/null +++ b/2024/Rethinking Few-shot Class-incremental Learning_ Learning from Yourself/a34c73cf-0382-420e-bbe9-4f775aa32eab_model.json @@ -0,0 +1,2823 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.221, + 0.141, + 0.783, + 0.187 + ], + "angle": 0, + "content": "Rethinking Few-shot Class-incremental Learning: Learning from Yourself" + }, + { + "type": "text", + "bbox": [ + 0.243, + 0.212, + 0.761, + 0.245 + ], + "angle": 0, + "content": "Yu-Ming Tang\\(^{1,3}\\), Yi-Xing Peng\\(^{1,3,5}\\), Jingke Meng\\(^{\\star 1,3}\\), and Wei-Shi Zheng\\(^{1,2,3,4,5}\\)" + }, + { + "type": "text", + "bbox": [ + 0.238, + 0.255, + 0.761, + 0.282 + ], + "angle": 0, + "content": "\\(^{1}\\) School of Computer Science and Engineering, Sun Yat-sen University, China \n\\(^{2}\\) Peng Cheng Laboratory, Shenzhen, China" + }, + { + "type": "text", + "bbox": [ + 0.232, + 0.283, + 0.772, + 0.31 + ], + "angle": 0, + "content": "3 Key Laboratory of Machine Intelligence and Advanced Computing, Ministry of Education, China" + }, + { + "type": "text", + "bbox": [ + 0.232, + 0.31, + 0.769, + 0.338 + ], + "angle": 0, + "content": "4 Guangdong Province Key Laboratory of Information Security Technology, Sun Yat-sen University, China" + }, + { + "type": "text", + "bbox": [ + 0.255, + 0.339, + 0.745, + 0.38 + ], + "angle": 0, + "content": "\\(^{5}\\) Pazhou Laboratory (Huangpu), Guangzhou, Guangdong 510555, China {tangym9, pengyx23}@mail2.sysu.edu.cn, mengjke@gmail.com, wszheng@ieee.org" + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.418, + 0.744, + 0.695 + ], + "angle": 0, + "content": "Abstract. Few-shot class-incremental learning (FSCIL) aims to learn sequential classes with limited samples in a few-shot fashion. Inherited from the classical class-incremental learning setting, the popular benchmark of FSCIL uses averaged accuracy (\\(aAcc\\)) and last-task averaged accuracy (\\(lAcc\\)) as the evaluation metrics. However, we reveal that such evaluation metrics may not provide adequate emphasis on the novel class performance, and the continual learning ability of FSCIL methods could be ignored under this benchmark. In this work, as a complement to existing metrics, we offer a new metric called generalized average accuracy (\\(gAcc\\)) which is designed to provide an extra equitable evaluation by incorporating different perspectives of the performance under the guidance of a parameter \\(\\alpha\\). We also present an overall metric in the form of the area under the curve (AUC) along the \\(\\alpha\\). Under the guidance of \\(gAcc\\), we release the potential of intermediate features of the vision transformers to boost the novel-class performance. Taking information from intermediate layers which are less class-specific and more generalizable, we manage to rectify the final features, leading to a more generalizable transformer-based FSCIL framework. Without complex network designs or cumbersome training procedures, our method outperforms existing FSCIL methods at \\(aAcc\\) and \\(gAcc\\) on three datasets. See codes at 1" + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.708, + 0.741, + 0.736 + ], + "angle": 0, + "content": "Keywords: Continual Learning \\(\\cdot\\) Few-shot Learning \\(\\cdot\\) Catastrophic Forgetting \\(\\cdot\\) Vision Transformer" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.747, + 0.377, + 0.763 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.772, + 0.788, + 0.804 + ], + "angle": 0, + "content": "Although existing deep neural networks (DNNs) achieve great success in various real-world applications [17, 31, 32, 34, 43, 44, 54, 68], they suffer from catastrophic" + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.811, + 0.386, + 0.826 + ], + "angle": 0, + "content": "* Corresponding Author" + }, + { + "type": "page_footnote", + "bbox": [ + 0.222, + 0.827, + 0.623, + 0.84 + ], + "angle": 0, + "content": "1 https://github.com/iSEE-Laboratory/Revisiting_FSCIL" + }, + { + "type": "list", + "bbox": [ + 0.218, + 0.811, + 0.623, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.347, + 0.129 + ], + "angle": 0, + "content": "Tang et al." + }, + { + "type": "image_caption", + "bbox": [ + 0.301, + 0.152, + 0.465, + 0.163 + ], + "angle": 0, + "content": "(a) Performance after each task" + }, + { + "type": "image", + "bbox": [ + 0.248, + 0.168, + 0.497, + 0.287 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.57, + 0.145, + 0.733, + 0.169 + ], + "angle": 0, + "content": "(b) Detailed Performance after finishing the last task (Task 9)" + }, + { + "type": "image", + "bbox": [ + 0.51, + 0.17, + 0.758, + 0.287 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.297, + 0.788, + 0.371 + ], + "angle": 0, + "content": "Fig. 1: Performance under different metrics of two recent methods [24, 25]. (a): The dotted line denotes the average accuracy (\\( aAcc \\)) widely used in classical class-incremental learning. The solid line presents our proposed generalized accuracy (\\( gAcc \\)). We also show the average of each point on the right side. (Blue: \\( gAcc \\), Red: \\( aAcc \\)). (b): The detailed accuracies of each task after training on the last task (task 8). It is evident that S3C exhibits superior performance when confronted with novel classes. The conventional metric \\( aAcc \\) fails to reflect this due to the domination by the base-class performance. Models are trained on the CIFAR-100 dataset. Best viewed in color." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.393, + 0.789, + 0.53 + ], + "angle": 0, + "content": "forgetting when dealing with sequential data streams. The forgetting problem has attracted heavy attention in recent years [4,16,21,29,33,45,48,50,51,57,61,64,65, 71,74]. In this work, we focus on a more realistic and challenging problem: few-shot class-incremental learning (FSCIL), in which after training on a task with sufficient data (base task), only a few samples are provided during incremental tasks (novel tasks). Since Tao et al. [52] proposed the FSCIL benchmark in 2020, a large number of algorithms have emerged to tackle this problem, including meta-learning based [8, 20, 70], structure based [2, 25, 52, 58, 63], feature space based [3, 6, 20, 24, 26, 59, 69] methods, etc." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.538, + 0.789, + 0.842 + ], + "angle": 0, + "content": "However, almost all these methods are evaluated following the traditional class-incremental learning setting, i.e. using the average accuracy (\\(aAcc\\)) or the last-task averaged accuracy (\\(lAcc\\)) as the main metrics to critique the proposed method. Such accuracies are calculated by weighted averaging the base-task and novel-tasks accuracies with the weight proportional to the number of classes of each task. Considering the reality that the FSCIL benchmarks make the majority of the classes (about \\(50\\%\\) to \\(60\\%\\)) as the base task to establish a well-trained backbone for few-shot continual learning, the existing \\(aAcc\\) or \\(lAcc\\) metric is dominated by the base-class performance. In other words, current evaluation metrics can be largely boosted by only improving the performance of base classes. As shown in Fig. 1, a method like softNet [25] with higher base-class performance (see Fig. 1 (b)) usually gets higher \\(aAcc\\) (see red lines in Fig. 1 (a)) regardless of the inferior performance on the novel classes. Due to the inadequacy to capture performance changes in novel classes, current metrics are biased toward the performance of base classes and have limitations in guiding the FSCIL community. To tackle this, we propose a more balanced metric called generalized accuracy (\\(gAcc\\)). Parameterized by \\(\\alpha\\) ranging from 0 to 1, the \\(gAcc\\) offers a way to evaluate a model with explicit emphasis on performance on novel classes. Furthermore, we can comprehensively evaluate the model by calculating the area under the curve (AUC) along the parameter \\(\\alpha\\). It is worth mentioning that our intention" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.246, + 0.115, + 0.734, + 0.131 + ], + "angle": 0, + "content": "Rethinking Few-shot Class-incremental Learning: Learning from Yourself" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.177 + ], + "angle": 0, + "content": "is NOT to replace all previous metrics, but to provide a more fine-grained and balanced evaluation of FSCIL methods complementary to them." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.178, + 0.787, + 0.268 + ], + "angle": 0, + "content": "We empirically find that FSCIL can benefit from the powerful vision transformer (ViT) [15] by analyzing the performance of novel classes. Specifically, although vanilla ViT performs poorly in terms of unseen novel classes, we reveal that the intermediate features of ViT are less specified on base classes. It means the ViT can learn abundant semantic information but the information is mistakenly dropped in deep layers when extracting features for novel classes." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.268, + 0.788, + 0.403 + ], + "angle": 0, + "content": "Based on the observation, we propose a lightweight Feature Rectification (FR) module to rectify the output features of the ViT conditioned on the intermediate features. To achieve this, the FR module is supervised with two losses that work on two different levels, namely instance-level relation transfer and class-center relation transfer. The former transfers the relations between instances while the latter focuses on relations between instances and the class prototypes. We further enhance the FR module by leveraging hierarchical knowledge from multiple intermediate layers. Finally, our method boosts the novel-class accuracy and maintains a comparable performance of base classes." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.404, + 0.788, + 0.601 + ], + "angle": 0, + "content": "Our contributions are summarized as follows. (1) We revisit the existing FSCIL benchmark and propose a novel metric called generalized accuracy (\\(gAcc\\)) for a more fine-grained and balanced evaluation. We have implemented 14 FSCIL methods with codes released and show that the newly introduced SOTAs do not necessarily perform better on \\(gAcc\\). (2) Based on \\(gAcc\\), we reveal the potential of vision transformers in FSCIL. We empirically find that although features from the intermediate layer exhibit sub-optimal performance for the base classes, their transferrable ability helps them to have a distinct advantage in handling novel classes. (3) A lightweight FR module is proposed to rectify the final features to be more transferrable. The FR module manages to make the output feature more generalizable with assistance from intermediate features. (4) We conduct extensive experiments on 3 datasets and evaluate our method under both \\(aAcc\\) and \\(gAcc\\), showing the effectiveness and superiority of our method." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.61, + 0.388, + 0.626 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.629, + 0.787, + 0.72 + ], + "angle": 0, + "content": "Meta-learning based methods. Due to the limited samples of novel classes, some works [8,20,67,70] propose to use meta-learning to improve the transferability of the model. Concretely, MetaFscil [8] proposed to simulate novel tasks by sampling from the base task. C-FSCIL [20] uses meta-learning to train the backbone, which maps images from different classes to quasi-orthogonal vectors. ALFSCIL [30] formulate the process into a bi-level optimization problem." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Optimization-based methods. There is a large group of works focusing on the optimization of the DNNs in FSCIL. CLOM [76] considers the behavior of the backbone under different margins of the cosine loss. NC-FSCIL [59] designs a pre-defined classifier for efficient FSCIL without any explicit regularizer. SPPR [73] proposes a prototype refinement strategy for stronger expression ability. FACT [69] assigns virtual prototypes in the current stage for future updates. Alice [42] instead focuses on the compactness and the variety of the feature space. Based on the classifier learning, S3C [24] proposes a self-supervised stochastic" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.347, + 0.129 + ], + "angle": 0, + "content": "Tang et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.192 + ], + "angle": 0, + "content": "classifier to tackle the overfitting and forgetting problem. There are more works working on the feature space constraint [3, 6, 13, 26, 27, 56, 60], or knowledge distillation [7, 9-11, 14, 39, 66, 72]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.192, + 0.785, + 0.267 + ], + "angle": 0, + "content": "Data-driven methods. There is also a branch of work focusing on the data side of this problem, which focuses on the data replay [1,35]. FSIL-GAN [1] employs a GAN to generate features to compensate for the scarcity of data. DF-replay [35] uses synthesized data generated by a generator to replay. PL-FSCIL [23] seeks prior knowledge from unlabeled data." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.268, + 0.787, + 0.328 + ], + "angle": 0, + "content": "Structure-based methods. In addition to these works, a unique branch of works focuses on the network structure, like topology preserving [52], growing classifier [63], proposing a dynamic structure of the backbone [58], structure fusion [2, 22], sub-networks [25] and ensemble models [75]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.328, + 0.785, + 0.419 + ], + "angle": 0, + "content": "Compared with existing works, our work rephrases the challenge of the FSCIL setting and first reveals the potential of the intermediate features of the ViT backbone. There is also an existing work [42] offering an alternative metric. They did not provide a fine-grained evaluation with various perspectives and have potential risks in stability and symmetry problems. Please find more detailed discussions about related works in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.423, + 0.547, + 0.439 + ], + "angle": 0, + "content": "3 Generalized Average Accuracy" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.439, + 0.546, + 0.455 + ], + "angle": 0, + "content": "3.1 Limitation of the Existing Metrics" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.461, + 0.788, + 0.601 + ], + "angle": 0, + "content": "Accuracy metric on incremental tasks. Consider a classification model \\( f(\\cdot) \\) that takes an image as input and outputs the class predictions. For an incremental problem, there are multiple tasks and multiple test sets. Let us define test set at task \\( \\mathcal{T}_i \\) as \\( \\mathcal{D}_{test}^i = \\{x_t^i,y_t^i\\}_{t = 1}^n \\). Usually, the labeling space \\( \\mathcal{Y}^i\\triangleq \\{y_t^i |(x_t^i,y_t^i)\\in \\mathcal{D}_{test}^i\\} \\) of every task does not have any intersection, i.e. \\( \\forall i\\neq j,\\mathcal{Y}^i\\bigcap \\mathcal{Y}^j = \\emptyset \\). After trained on \\( \\mathcal{T}_i \\), the model can be denoted as \\( f_{i}(\\cdot) \\). The corresponding accuracy of model \\( f_{i}(\\cdot) \\) at task \\( \\mathcal{T}_j \\) (\\( j\\leq i \\)) can be denoted as \\( \\mathcal{A}_i^j\\triangleq \\frac{1}{n^j}\\sum_{t = 1}^{n^j}\\mathbb{1}(\\zeta (f_i(x_t^j)) = y_t^j) \\), where \\( \\mathbb{1}(\\cdot) = 1 \\) if the condition is true, otherwise 0. \\( \\zeta \\) is the argmax operation." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.602, + 0.788, + 0.693 + ], + "angle": 0, + "content": "Average accuracy. It is worth mentioning that for class-incremental learning, evaluating the model on all the previous tasks is essential. To test the model on previous tasks, an overall test set that contains all previous test sets is constructed as \\(\\mathcal{D}_{test}^{1\\rightarrow i} = \\bigcup_{m = 1}^{i}\\mathcal{D}_{test}^{m}\\). After trained on task \\(\\mathcal{T}_i\\), the accuracy of model \\(f_{i}(\\cdot)\\) on the overall test set \\(\\mathcal{D}_{test}^{1\\rightarrow i}\\) can be presented as \\(A_{i}^{1\\rightarrow i}\\). Once a \\(n_t\\)-task incremental training ends, the average accuracy (\\(aAcc\\)) can be calculated as:" + }, + { + "type": "equation", + "bbox": [ + 0.373, + 0.701, + 0.786, + 0.739 + ], + "angle": 0, + "content": "\\[\na A c c = \\frac {1}{n _ {t}} \\sum_ {i = 1} ^ {n _ {t}} a A c c _ {i} = \\frac {1}{n _ {t}} \\sum_ {i = 1} ^ {n _ {t}} A _ {i} ^ {1 \\rightarrow i}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.746, + 0.788, + 0.794 + ], + "angle": 0, + "content": "where the \\(aAcc_{i}\\) is the average accuracy at each task. We can also derive two additional metrics incidentally: last-task averaged accuracy \\(lAcc = aAcc_{n_t} = A_{n_t}^{1\\rightarrow n_t}\\), and task-wise average accuracy \\((tAcc)\\):" + }, + { + "type": "equation", + "bbox": [ + 0.364, + 0.803, + 0.786, + 0.843 + ], + "angle": 0, + "content": "\\[\nt A c c = \\frac {1}{n _ {t}} \\sum_ {i = 1} ^ {n _ {t}} t A c c _ {i} = \\frac {1}{n _ {t}} \\sum_ {i = 1} ^ {n _ {t}} \\frac {1}{i} \\sum_ {j = 1} ^ {i} A _ {i} ^ {j}, \\tag {2}\n\\]" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.246, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "Rethinking Few-shot Class-incremental Learning: Learning from Yourself" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.191 + ], + "angle": 0, + "content": "where the \\( tAcc_i \\) is the task-wise accuracy at each task. Note that the \\( aAcc \\) does not share the same value as \\( tAcc \\) normally since the number of classes of tasks is usually not the same. This will be elaborated on later in detail." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.192, + 0.788, + 0.329 + ], + "angle": 0, + "content": "Limitation. Since FSCIL usually sets the majority of the classes as the base task, existing metrics potentially neglect the performance changes of novel classes. To elaborate on this, we analyze the \\(aAcc\\) as follows. Firstly, we denote the size of the labeling space of task \\(\\mathcal{T}_i\\) as \\(|\\mathcal{V}_i|\\). In FSCIL, the number of classes of each novel task is usually the same, i.e. \\(|\\mathcal{V}_{novel}| = |\\mathcal{V}_2| = |\\mathcal{V}_3| = \\dots = |\\mathcal{V}_{n_t}| \\ll |\\mathcal{V}_1|\\). Assuming there are total \\(n_t\\) tasks, and task \\(\\mathcal{T}_1\\) is the base task with sufficient classes and samples (non-few-shot). The following tasks \\(\\{\\mathcal{T}_i\\}_{i=2}^{n_t}\\) are novel tasks, with \\(|\\mathcal{V}_{novel}|\\) classes for each task and a few samples per class. Now, we can rewrite average accuracy on task \\(\\mathcal{T}_i\\) as:" + }, + { + "type": "equation", + "bbox": [ + 0.286, + 0.334, + 0.786, + 0.378 + ], + "angle": 0, + "content": "\\[\na A c c _ {i} = \\frac {\\left| \\mathcal {Y} _ {1} \\right| A _ {i} ^ {1} + \\left| \\mathcal {Y} _ {n o v e l} \\right| \\sum_ {j = 2} ^ {i} A _ {i} ^ {j}}{\\left| \\mathcal {Y} _ {1} \\right| + (i - 1) \\left| \\mathcal {Y} _ {n o v e l} \\right|} = \\frac {\\frac {\\left| \\mathcal {Y} _ {1} \\right|}{\\left| \\mathcal {Y} _ {n o v e l} \\right|} A _ {i} ^ {1} + \\sum_ {j = 2} ^ {i} A _ {i} ^ {j}}{\\frac {\\left| \\mathcal {Y} _ {1} \\right|}{\\left| \\mathcal {Y} _ {n o v e l} \\right|} + (i - 1)}. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.383, + 0.784, + 0.468 + ], + "angle": 0, + "content": "From Eq. (3) we see that the value of \\( aAcc_i \\) will be dominated by the base class accuracy \\( A_i^1 \\) if the ratio \\( \\frac{|\\mathcal{Y}_1|}{|\\mathcal{Y}_{\\text{novel}}|} \\) is large. Unfortunately, the ratio is indeed large in the FSCIL settings. (e.g. \\( \\frac{|\\mathcal{Y}_1|}{|\\mathcal{Y}_{\\text{novel}}|} = 12 \\) for both CIFAR-100 and miniImageNet dataset, 10 for CUB-200.) Under such a large ratio, the changes (both improvement and degrades) of novel classes cannot be effectively evaluated by \\( aAcc \\)." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.473, + 0.553, + 0.488 + ], + "angle": 0, + "content": "3.2 Generalizing the Average Accuracy" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.49, + 0.787, + 0.52 + ], + "angle": 0, + "content": "To overcome the limitations of existing metrics, we now define the generalized accuracy (gAcc) at task \\(\\mathcal{T}_i\\) as:" + }, + { + "type": "equation", + "bbox": [ + 0.374, + 0.526, + 0.786, + 0.571 + ], + "angle": 0, + "content": "\\[\ng A c c _ {i} (\\alpha) = \\frac {\\alpha \\frac {| \\mathcal {Y} _ {1} |}{| \\mathcal {Y} _ {n o v e l} |} A _ {i} ^ {1} + \\sum_ {j = 2} ^ {i} A _ {i} ^ {j}}{\\alpha \\frac {| \\mathcal {Y} _ {1} |}{| \\mathcal {Y} _ {n o v e l} |} + (i - 1)}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.575, + 0.788, + 0.732 + ], + "angle": 0, + "content": "in which the \\( gAcc \\) is parameterized by \\( \\alpha \\in [0,1] \\). At task \\( \\mathcal{T}_i \\), \\( gAcc_i(0) \\) measures the accuracy of the novel tasks only. \\( gAcc_i(\\frac{|\\mathcal{Y}_{\\text{novel}}|}{|\\mathcal{Y}_1|}) \\) denotes the task-wise accuracy \\( tAcc_i \\) at task \\( \\mathcal{T}_i \\) while \\( gAcc(1) \\) represents the average accuracy \\( aAcc_i \\) at task \\( \\mathcal{T}_i \\). In addition to these discrete values, we further generalize the \\( \\alpha \\) to any rational numbers in [0,1]. Controlled by \\( \\alpha \\) in Eq. (4), when \\( \\alpha \\to 0 \\), \\( gAcc(\\alpha) \\) emphasizes to novel class while \\( \\alpha \\to 1 \\), \\( gAcc(\\alpha) \\) emphasizes the base class performance, i.e., the larger the \\( \\alpha \\), the greater the influence of the base class on the value of \\( gAcc(\\alpha) \\). To evaluate the performance of a model on a specific task \\( \\mathcal{T}_i \\), a standalone value is needed. Thus, we introduce the area under the curve (AUC) of \\( gAcc_i(\\alpha) \\) along the parameter \\( \\alpha \\) as a comprehensive metric:" + }, + { + "type": "equation", + "bbox": [ + 0.413, + 0.736, + 0.786, + 0.77 + ], + "angle": 0, + "content": "\\[\ng A c c _ {i} = \\int_ {0} ^ {1} g A c c _ {i} (\\alpha) d \\alpha . \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.774, + 0.784, + 0.804 + ], + "angle": 0, + "content": "Similar to \\( aA_{cc} \\), for multiple tasks, we average the \\( gA_{cc_i} \\) for each task and get an overall metric \\( gA_{cc} \\) for the whole learning process:" + }, + { + "type": "equation", + "bbox": [ + 0.432, + 0.81, + 0.786, + 0.843 + ], + "angle": 0, + "content": "\\[\ng A c c = \\frac {1}{n _ {t}} \\sum_ {i} ^ {n _ {t}} g A c c _ {i}. \\tag {6}\n\\]" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.347, + 0.129 + ], + "angle": 0, + "content": "Tang et al." + }, + { + "type": "image", + "bbox": [ + 0.26, + 0.145, + 0.578, + 0.294 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.584, + 0.147, + 0.728, + 0.16 + ], + "angle": 0, + "content": "(b) \\(gAcc_{i}\\) curve w.r.t Task i" + }, + { + "type": "image", + "bbox": [ + 0.58, + 0.169, + 0.728, + 0.293 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.303, + 0.788, + 0.378 + ], + "angle": 0, + "content": "Fig. 2: FSCIL performances are shown in generalized accuracy. (a): The \\(gAcc\\) curve (averaged across all \\(n_t\\) tasks) v.s. param \\(\\alpha\\). (b): The \\(gAcc\\) AUC of each task \\(\\mathcal{T}_i\\) (Eq. (5)). In the legend, we show the average AUC across tasks(Eq. (6)) of each method. We evaluate recent works including SAVC [49], NC [59], and Alice [42] on the miniImageNet dataset. Some corner cases: Lazy: The model maintains base-task performance while refusing to learn anything from novel tasks. Greedy: Regardless of the previous knowledge, the model only greedily focuses on the current task. Greedy-NF: A Non-Forget version of 'Greedy'. See more about these cases in our supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.386, + 0.788, + 0.416 + ], + "angle": 0, + "content": "The proposed \\(gAcc\\) provides different evaluation perspectives using different \\(\\alpha\\) so we believe such a metric can provide a more balanced evaluation of FSCIL." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.424, + 0.334, + 0.439 + ], + "angle": 0, + "content": "3.3 Analysis" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.447, + 0.788, + 0.568 + ], + "angle": 0, + "content": "We further study the necessity of the proposed \\( gAcc \\). We evaluate recent SOTAs in FSCIL including SAVC [49], Alice [42] and NC-FSCIL [59], and the results are in Fig. 2. We show that the \\( gAcc \\) provides more fine-grained evaluation than the classical accuracy metric. For example, in the left part of Fig. 2, although the \\( gAcc(1) \\), which is equal to the \\( aAcc \\), of SAVC is close to that of NC-FSCIL, the area under the curve(AUC) of SAVC is much smaller than NC-FSCIL. This indicates that the novel class performance is much worse for SAVC, which cannot be discovered by conventional metrics." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.569, + 0.789, + 0.842 + ], + "angle": 0, + "content": "Furthermore, to show the superiority of our metric, we consider three corner cases: Lazy, Greedy, and Greedy-NF. 'Lazy' stands for a type of algorithm that only focuses on maintaining base-class performance, i.e. \\(\\forall i \\geq j > 1\\), \\(A_{i}^{j} = 0\\), \\(A_{i}^{1} = A_{1}^{1}\\). By setting \\(A_{1}^{1} = 85\\%\\), we can see that when \\(\\alpha = 1\\) (i.e. the traditional aAcc metric), its performance is close to SOTA methods. Further, when considering popular forgetting metrics like performance dropping \\((PD = A_{i}^{j} - A_{j}^{j})\\) and Knowledge rate \\((KR = A_{i}^{j} / A_{j}^{j})\\), because the forgetting is zero, its PD is 0 and KR is 1. All these metrics indicate such 'Lazy' model is superior, which is contradictory to common sense. However, our proposed gAcc reveals that its performance is far from satisfying by setting \\(\\alpha < 1\\) and its AUC is far smaller than other methods. In other words, using the traditional metrics will indicate algorithms like 'Lazy' achieving SOTA performance while such methods perform poorly evaluated by our proposed gAcc. The gAcc can also discover other corner cases. 'Greedy' is to greedily learn the current knowledge (i.e. \\(A_{1}^{1} = 85\\%\\), \\(A_{i}^{j} = 100\\%\\), when \\(i = j\\) else 0). \"Greedy\" achieves low both aAcc and gAcc. We further eliminate the forgetting problem from 'Greedy' and define 'Greedy-NF' on top of it: set accuracy \\(A_{i}^{j}\\) to \\(100\\%\\) for \\(\\forall j > 1, j < i\\) (no forgetting at novel classes). We find that although 'Greedy-NF' performs" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.246, + 0.114, + 0.733, + 0.129 + ], + "angle": 0, + "content": "Rethinking Few-shot Class-incremental Learning: Learning from Yourself" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "image", + "bbox": [ + 0.292, + 0.145, + 0.523, + 0.314 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.549, + 0.149, + 0.717, + 0.311 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.325, + 0.788, + 0.378 + ], + "angle": 0, + "content": "Fig. 3: t-SNE [37] visualization of miniImageNet test set features from different layers of a ViT. Left: we choose layers \\( L_{9} \\), \\( L_{10} \\), \\( L_{11} \\), \\( L_{12} \\) (final) and show both base classes and novel classes features. It is observed that shallow features are more dispersed than deep features. Right: Accuracies of different tasks on features from various layers. It is clear that intermediate features achieve better performance on novel classes and thus have better generalization ability. Best viewed in color." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.389, + 0.784, + 0.434 + ], + "angle": 0, + "content": "well when majorly considering the novel-class performance \\((\\alpha \\in \\{0, \\frac{1}{12}, \\frac{2}{12}, \\frac{3}{12}\\})\\), the overall performance (AUC) is not satisfying. Please refer to Sec. A1 in the supplementary material for more details about these corner cases." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.436, + 0.785, + 0.57 + ], + "angle": 0, + "content": "By analyzing these cases, we show the superior properties of the proposed \\( gAcc \\). Firstly, the \\( gAcc \\) metric effectively distinguishes methods (e.g. 'Lazy') that only retain the base-task performance, thereby encouraging methods to learn from novel classes. Besides, \\( gAcc \\) avoids overemphasis on the performance of base classes or novel classes (corner case 'Greedy-NF' is still far worse than normal methods in terms of AUC), providing a more balanced evaluation outcome. In other words, the proposed \\( gAcc \\) metric provides objective and balanced results when evaluating FSCIL methods, eliminating 'shortcuts' in \\( aAcc \\) that focus only on the performance of the base classes." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.572, + 0.785, + 0.601 + ], + "angle": 0, + "content": "In conclusion, the \\(gAcc\\) curve and the AUC are necessary and powerful tools to evaluate the FSCIL methods." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.616, + 0.45, + 0.633 + ], + "angle": 0, + "content": "4 Learn from Yourself" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.645, + 0.563, + 0.661 + ], + "angle": 0, + "content": "4.1 Preliminary and The FSCIL Setting" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.673, + 0.788, + 0.844 + ], + "angle": 0, + "content": "For task \\(\\mathcal{T}_i\\in \\{\\mathcal{T}_i\\}_{i = 1}^T\\), the training set can be denoted as \\(\\mathcal{D}_{train}^{i} = \\{x_{t}^{i},y_{t}^{i}\\}_{t = 1}^{m^{i}}\\). For FSCIL settings, the base task \\(\\mathcal{T}_1\\) contains \\(|\\mathcal{V}_1|\\) classes and sufficient samples for training. For \\(\\forall i > 1\\), there are \\(|\\mathcal{V}_{novel}|\\) classes for each task and \\(k\\) samples per class. Typically, \\(|\\mathcal{V}_{novel}|\\) is set to 5, and \\(k = 5\\), i.e. a 5-way 5-shot setting. We rephrase the classification model \\(f(\\cdot)\\) in Sec. 3 as \\(f(\\cdot) = \\varPhi (f_B(\\cdot))\\) where \\(\\varPhi (\\cdot)\\) is the classifier and \\(f_{B}(\\cdot)\\) is the backbone. We follow previous works to use a cosine classifier: \\(\\varPhi(X)=P^\\top\\overline{X}\\), where \\(P\\in \\mathbb{R}^{d\\times |\\mathcal{V}|}\\) and \\(\\overline{X}\\in \\mathbb{R}^{d\\times 1}\\) are the L2-normalized inputs, and \\(d\\) is the feature dimension of \\(f_{B}(\\cdot)\\), \\(|\\mathcal{V}| = \\sum_{i}|\\mathcal{V}_{i}|\\) is the number of classes seen so far. In this work, we employ the Vision Transformer(ViT) [15] in the FSCIL setting. For a ViT with \\(N_{L}\\) layers \\(\\{L_l\\}_{l = 1}^{N_L}\\), given a input image \\(x\\), the output can be denoted as \\(X_{N_L} = L_{N_L}(L_{N_L - 1}(\\dots L_1(x))\\)." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.116, + 0.347, + 0.129 + ], + "angle": 0, + "content": "Tang et al." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.145, + 0.787, + 0.31 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.319, + 0.788, + 0.385 + ], + "angle": 0, + "content": "Fig. 4: The overall framework of our method. The Feature Rectification (FR) module takes the final and the intermediate features as input and outputs the rectified features for classification. The relation transfer losses are designed to convey valuable information from intermediate features to the rectified ones. The details of these two parts are shown in the left and lower parts of the figure. Further, a cosine constraint is proposed to maintain the base class performance and a classification loss is employed to adapt to the novel classes. Best viewed in color." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.394, + 0.788, + 0.469 + ], + "angle": 0, + "content": "Baseline. Recent works [3, 20, 24, 38, 47, 49, 59, 63, 76] suggest a training-frozen scheme, which trains the backbone \\( f_{B}(\\cdot) \\) and the classifier \\( \\varPhi(\\cdot) \\) (if available) on the base-task only, and then freeze the backbone during novel tasks. For novel classes, the classifier is extended by mean features (i.e. prototypes) extracted by the frozen backbone. We follow and employ the scheme as our baseline." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.47, + 0.789, + 0.756 + ], + "angle": 0, + "content": "Main idea of our method. Given that the backbone is initially trained for the base task and remains fixed during novel tasks, the main challenges lie in the potential overfitting on the base task and the poor feature extraction ability of the novel classes. We conduct a toy experiment to show these problems and show that the intermediate feature can convey valuable information to tackle these problems. Results are shown in Fig. 3, where we select 4 base classes and 5 novel classes and utilize the t-SNE [37] to visualize the features from different layers. It can be concluded that features extracted by the base-classes trained backbone are overfitted to the base classes and cannot be generalized well on novel classes, i.e. the final features are well aggregated, but it is hard to distinguish among unseen classes. (see \\( L_{12} \\) features on the left and the corresponding performance on the right). Compared with the final features (i.e. \\( L_{12} \\)), the shallower intermediate features \\( (L_9, L_{10}, L_{11}) \\) are more dispersed, which makes them more discriminative for unseen classes. This is cross-verified by the performance shown on the right side of Fig. 3. We can see that using intermediate features can largely boost the novel classes' performance (up to \\( 15\\% \\) improvement) with a price of performance decline of the base classes. Based on this observation, we propose to rectify the final feature based on the intermediate feature to better handle both base and novel classes." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.782, + 0.436, + 0.797 + ], + "angle": 0, + "content": "4.2 Feature Rectification" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.842 + ], + "angle": 0, + "content": "We propose a Feature Rectification (FR) module to rectify the final feature based on the relation information from the intermediate features. Specifically," + } + ], + [ + { + "type": "header", + "bbox": [ + 0.246, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "Rethinking Few-shot Class-incremental Learning: Learning from Yourself" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.147, + 0.784, + 0.176 + ], + "angle": 0, + "content": "given a pair of final feature \\( X_{N_L} \\) and the intermediate feature \\( X_l \\), we design an FR module conditioned on the intermediate features:" + }, + { + "type": "equation", + "bbox": [ + 0.365, + 0.188, + 0.786, + 0.204 + ], + "angle": 0, + "content": "\\[\nX _ {F R} = M _ {m i x} \\left(\\operatorname {c a t} \\left(M _ {F} \\left(X _ {N _ {L}}\\right), M _ {I} \\left(X _ {l}\\right)\\right), \\right. \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.212, + 0.784, + 0.331 + ], + "angle": 0, + "content": "where 'cat' means the concatenate operation. \\( M_{\\text{mix}} \\), \\( M_F \\) and \\( M_I \\) are 2-layer MLPs. By introducing the intermediate features, the FR module has access to a reference for extracting the valuable structural information lies in the intermediate features. A naive way to mine such information is to directly pull rectified features closer to shallow features on an instance-by-instance basis. Although such an instance-wise loss can bridge shallow features and rectified features, directly pulling the rectified feature to the shallow one could lead to noisy features since the shallow features are more scattered, as shown in Sec. 4.1." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.333, + 0.784, + 0.424 + ], + "angle": 0, + "content": "To effectively transfer discriminative cues among samples from shallow features to rectified features and avoid introducing noise, we propose two relation-based transfer losses at different levels. Firstly, we offer a Instance-level Relation transfer (IR) to mine the relation between feature instances. Given a pair of final features \\( X_{N_L}^1 \\) and \\( X_{N_L}^2 \\), and another pair of intermediate features \\( X_l^1 \\) and \\( x_{l}^{2} \\), the IR loss is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.381, + 0.433, + 0.785, + 0.45 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {I R} = \\delta \\left(\\epsilon \\left(X _ {F R} ^ {1}, X _ {F R} ^ {2}\\right), \\epsilon \\left(X _ {l} ^ {1}, x _ {l} ^ {2}\\right)\\right), \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.458, + 0.788, + 0.579 + ], + "angle": 0, + "content": "where \\(\\epsilon (\\cdot)\\) is the Euclidean distance and the \\(\\delta (\\cdot)\\) denotes the smooth L1 loss. Besides the relation between feature instances, another level of the relation information is the relation between features and class centers. Class-center Relation Transfer (CR) is proposed to mine this class-center-level information. Assume at task \\(T_{i}\\) with \\(|\\mathcal{V}_i|\\) classes and the feature dimension is \\(d\\). Given a set of class prototypes \\(P\\in \\mathbb{R}^{d\\times |\\mathcal{V}_i|}\\) from the classifier \\(\\varPhi(\\cdot)\\) as well as a pair of final feature \\(X_{N_L}\\in \\mathbb{R}^{d\\times 1}\\) and the intermediate feature \\(X_{l}\\) \\((l\\in [1,N_L))\\), the corresponding CR loss is formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.376, + 0.588, + 0.785, + 0.606 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {C R} = \\mathrm {K L} \\left(\\left(\\sigma \\left(P ^ {\\top} \\overline {{X _ {F R}}}\\right), \\sigma \\left(P ^ {\\top} \\overline {{X _ {l}}}\\right)\\right), \\right. \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.614, + 0.784, + 0.643 + ], + "angle": 0, + "content": "where \\(\\sigma (\\cdot)\\) is the softmax function, and \\(\\mathrm{KL}(\\cdot)\\) is the Kullback-Leibler divergence. \\(\\overline{X}\\) denotes the L2 normalization results of \\(X\\)." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.645, + 0.788, + 0.84 + ], + "angle": 0, + "content": "In the above equations, the two losses \\(\\mathcal{L}_{CR}\\) and \\(\\mathcal{L}_{IR}\\) work on different levels and serve different roles. \\(\\mathcal{L}_{CR}\\) works on the relations between a sample and all class centers, focusing on the coarse-grained class-specific patterns that distinguish different classes. Differently, \\(\\mathcal{L}_{IR}\\) provides knowledge about fine-grained discriminative cues that contribute to pairwise relations. The proposed relation-based transfer losses are related to previous knowledge distillation works [5, 40, 41]. However, these works do not share the same objective as ours. Our goal is to extract precious transferrable information hidden in various shallow layers hierarchically to rectify the final features while prior distillation methods focus on matching final outputs to teachers. Furthermore, there exists a branch of works using intermediate features [36, 62] directly for dense prediction. Differently, we are the first to extract intermediate knowledge to rectify the final feature for a classification task." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.347, + 0.129 + ], + "angle": 0, + "content": "Tang et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.193 + ], + "angle": 0, + "content": "Besides, we constrain the discriminative ability of rectified features on base classes to avoid negative rectification. Since the final feature of layer \\( L_{N_L} \\) is strong enough for the base classes, we apply a cosine constraint as:" + }, + { + "type": "equation", + "bbox": [ + 0.419, + 0.202, + 0.786, + 0.217 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {c o s}} = \\cos \\left(X _ {F R}, X _ {N _ {L}}\\right), \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.224, + 0.459, + 0.239 + ], + "angle": 0, + "content": "where \\(\\cos (\\cdot ,\\cdot)\\) is cosine similarity." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.24, + 0.784, + 0.27 + ], + "angle": 0, + "content": "Novel class training. To ensure the rectified features are centered around the class mean prototypes, we further introduce a classification loss for novel classes:" + }, + { + "type": "equation", + "bbox": [ + 0.339, + 0.283, + 0.786, + 0.3 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {N o v C E}} = - \\mathbb {1} (y \\in \\cup_ {n > 1} \\mathcal {Y} _ {n}) \\log \\sigma \\left(P ^ {\\top} \\overline {{X _ {F R}}}\\right) ^ {(y)}, \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.305, + 0.784, + 0.336 + ], + "angle": 0, + "content": "where \\(y\\) is the corresponding ground true label of \\(X_{FR}\\), \\((P^{\\top}\\overline{X_{FR}})^{(y)}\\) is the \\(y\\)-th element of \\((P^{\\top}\\overline{X_{FR}})\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.336, + 0.785, + 0.516 + ], + "angle": 0, + "content": "Anti-forgetting training. Another challenge of FSCIL is the forgetting problem. Thanks to our feature rectification design, the only trainable module is the FR module. As the FR module takes the final and the intermediate feature as the input, these kinds of features can be easily modeled by Gaussian distributions [51,71]. Hence, we construct a Gaussian distribution for each class on each layer. The Gaussian distribution of class \\(i\\) features at layer \\(L_{l}\\) can be denoted as \\(\\mathcal{N}_i^l (\\mu_i^l,\\Sigma_i^l)\\), where \\(\\Sigma_i^l\\) is obtained by calculating the covariance of each feature dimension if \\(i\\in \\mathcal{Y}_1\\) (base classes). For novel classes, the covariance matrix is computed by averaging the covariance matrices of the top-k most similar base classes. During the incremental learning phases, we utilize these Gaussian distributions to sample features, which serve as inputs from old tasks to the FR module, assisting the FR module in retaining previously learned knowledge." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.516, + 0.784, + 0.546 + ], + "angle": 0, + "content": "Overall training objectives. The total loss function for our feature rectification framework is as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.333, + 0.555, + 0.785, + 0.571 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\beta_ {\\text {c o s}} \\left(\\mathcal {L} _ {\\text {c o s}} + \\mathcal {L} _ {\\text {N o v C e}}\\right) + \\beta_ {C R} \\mathcal {L} _ {C R} + \\beta_ {I R} \\mathcal {L} _ {I R}, \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.578, + 0.579, + 0.593 + ], + "angle": 0, + "content": "where \\(\\beta_{cos},\\beta_{CR}\\) and \\(\\beta_{IR}\\) are trade-off parameters." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.594, + 0.788, + 0.735 + ], + "angle": 0, + "content": "Multi-layer knowledge ensemble. Since there exists hierarchical knowledge in different layers, we mine such information by knowledge ensemble from different layers. Specifically, we incorporate an FR module with each selected layer, resulting in a total of \\( n_{FR} \\) FR modules. During the training phase, these FR modules are optimized simultaneously. When evaluation, FR modules at layer \\( l \\) are employed to produce multiple rectified features denoted as \\( X_{FR}^{l} \\). These rectified features are subsequently input into the shared classifier \\( P \\) for generating multiple predictions: \\( \\mathrm{Pred}_l = \\sigma (P^\\top \\overline{X_{FR}^l}) \\). The final prediction is computed as the average of each FR-branch: \\( \\mathrm{Pred} = \\frac{1}{n_{FR}}\\sum_{l = 1}^{n_{FR}}\\mathrm{Pred}_l \\)." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.751, + 0.377, + 0.768 + ], + "angle": 0, + "content": "5 Experiments" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.773, + 0.414, + 0.789 + ], + "angle": 0, + "content": "5.1 Experiment Setup" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.787, + 0.842 + ], + "angle": 0, + "content": "Datasets. Following previous works, we conduct experiments on three popular datasets of FSCIL: CIFAR-100 [28], miniImageNet [46] and CUB-200 [55]. For statistics of the three datasets, please refer to the supplementary material." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.246, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "Rethinking Few-shot Class-incremental Learning: Learning from Yourself" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.144, + 0.788, + 0.176 + ], + "angle": 0, + "content": "Table 1: Comparison with FSCIL methods on miniImageNet. The best and second best results are marked bolded and underlined. Methods with†are reported with the original results in their paper. Other results are reimplemented by us." + }, + { + "type": "table", + "bbox": [ + 0.254, + 0.179, + 0.74, + 0.398 + ], + "angle": 0, + "content": "
MethodsNum. of \nParams.(M)aAcc in each task (%)
123456789aAccgAcc
TOPIC [52]†-61.350.145.241.237.535.532.229.524.439.64-
IDLVQ [6]†-64.859.855.952.649.947.644.843.141.851.16-
MetaFSCIL [8]†-72.068.063.860.357.655.253.050.849.258.85-
DF-Replay [35]†-71.867.163.259.857.054.051.649.548.258.02-
FCIL [18]†-76.3471.467.164.161.358.555.754.152.862.37-
Self-promoted [73]12.3463.664.159.555.552.149.245.643.140.952.6241.53
CEC [63]12.1272.167.063.059.656.953.951.449.547.857.9148.64
LIMIT [70]12.2973.367.663.259.556.753.751.449.748.258.1549.47
FACT [69]11.2676.371.366.962.959.656.353.551.149.260.7849.56
CLOM [76]14.2272.166.762.759.256.153.050.248.346.757.2346.11
C-FSCIL [20]12.7970.663.859.356.454.351.848.947.346.155.4050.46
BiDistFSCIL [66]11.5074.670.466.662.860.557.254.753.051.961.3253.49
S3C [24]0.3477.072.168.265.162.058.856.053.651.962.7453.70
Regularizer [3]26.3081.972.768.764.157.453.553.852.343.162.8255.62
SubNetwork [25]22.3878.173.569.866.263.260.357.655.754.564.3155.88
SAVC [49]12.1281.076.271.971.465.662.159.257.055.566.6656.35
ALICE [42]11.2081.171.167.864.162.460.157.956.856.064.1459.72
NC-FSCIL [59]15.8784.276.172.168.167.363.961.959.657.967.9060.93
YourSelf (Ours)11.5784.077.673.770.068.064.962.159.859.068.8062.20
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.401, + 0.793, + 0.552 + ], + "angle": 0, + "content": "Training and testing protocols. For training protocols, we follow prior works [3, 20, 24, 25, 42, 49, 59, 66] to split a large proportion of the training set to serve as the base task, and split the rest into small incremental tasks. Specifically, for CIFAR-100 and miniImageNet, the first task contains 60 classes out of the total 100 classes, and the rest 40 classes are split into 8 tasks, following a 5-way 5-shot scheme. For CUB-200, there are 100 classes in the first task and the number of novel tasks is 10 (in a 10-way 5-shot setting). As for the testing, we follow prior works to report the average accuracy (\\(aAcc\\)). In addition, we also provide detailed discussions on our metric generalized accuracy (\\(gAcc\\)). We further provide detailed results of novel classes only in our supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.552, + 0.79, + 0.75 + ], + "angle": 0, + "content": "Implementation details. For all experiments below, we use the ViT suggested by Touvron et, al. [53]. We adjust stuff like the dimensions of the backbone to match the number of parameters (around 10M) as the ResNet18 [19] which is widely used in this topic. The Feature Rectification (FR) is a lightweight module and uses popular MLPs implementation [53] with only 0.6M parameters. To ensure a fair comparison, we also re-implement other methods with the same backbone as ours, but found the performance is rarely improved. Please refer to Supplementary material for more details. We follow previous works to train from scratch on CIFAR-100 and miniImageNet dataset. Prior works [6,8,35, 52,63,73] adopt an ImageNet-pretrained ResNet for the CUB-200 dataset. To make a fair comparison, we also pretrained our backbone on the ImageNet [12] to evaluate on CUB-200. For evaluation, we reproduced 14 methods based on the public codes and conducted in-depth comparisons in the following sections." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.77, + 0.617, + 0.787 + ], + "angle": 0, + "content": "5.2 Comparison with State-of-the-art methods" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.789, + 0.842 + ], + "angle": 0, + "content": "Experimental results on miniImageNet are shown in Table 1. We report the classical accuracy (\\(aAcc\\)) in each section as well as its averaged results and our proposed \\(gAcc\\). From the table, we can conclude that our proposed method" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.347, + 0.129 + ], + "angle": 0, + "content": "Tang et al." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.148, + 0.357, + 0.291 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.361, + 0.151, + 0.498, + 0.292 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.22, + 0.292, + 0.502, + 0.32 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.506, + 0.145, + 0.646, + 0.292 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.648, + 0.147, + 0.787, + 0.292 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.546, + 0.293, + 0.756, + 0.325 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.33, + 0.499, + 0.382 + ], + "angle": 0, + "content": "Fig. 5: FSCIL results on miniImageNet. (a): The \\( gAcc \\) curve (averaged across all \\( n_t \\) tasks) vs. \\( \\alpha \\). (b): The \\( gAcc \\) AUC of each task \\( \\mathcal{T}_i \\) (Eq. (5)) vs. i. In the legend, we show the AUC value (Eq. (6)) of each method. Best viewed in color." + }, + { + "type": "image_caption", + "bbox": [ + 0.504, + 0.335, + 0.788, + 0.387 + ], + "angle": 0, + "content": "Fig. 6: FSCIL results on CUB-200. (a): The \\( gAcc \\) curve (averaged across all \\( n_t \\) tasks) vs. \\( \\alpha \\). (b): The \\( gAcc \\) AUC of each task \\( \\mathcal{T}_i \\) (Eq. (5)) vs. \\( i \\). In the legend, we show the AUC value (Eq. (6)) of each method. Best viewed in color." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.398, + 0.785, + 0.473 + ], + "angle": 0, + "content": "outperforms other methods by a large margin. Specifically, under the classical evaluation metric \\(aAcc\\), the proposed method achieves 68.80% accuracy, which is 0.90%, 4.66%, 2.14% higher than SOTA methods NC-FSCIL [59], ALICE [42] and SAVC [49]. The improvement is non-trivial considering that miniImageNet is a complex dataset." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.475, + 0.788, + 0.566 + ], + "angle": 0, + "content": "From a more balanced perspective, we also report \\( gAcc \\) on the Table 1 and Fig. 5. In the left figure of Fig. 5, except for the method ALICE when \\( \\alpha = 0 \\), our method outperforms most other methods by a clear margin (at least \\( 1.27\\% \\)) and achieves the highest AUC of all methods. This indicates that our method achieves a more balanced base-novel performance. To see raw numbers of the accuracy of novel classes, please refer to the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.567, + 0.788, + 0.717 + ], + "angle": 0, + "content": "Besides, Fig. 5 shows that methods that perform well at \\(\\alpha = 1\\) do not necessarily perform well when \\(\\alpha\\) is smaller. For example, although CLOM achieves \\(57.23\\%\\) average accuracy (\\(\\alpha = 1\\)) and is \\(1.83\\%\\) higher than method C-FSCIL (\\(55.4\\%\\)), we can see the performance drops significantly when \\(\\alpha\\) is small. It is mainly because CLOM focuses on the base-classes accuracy and performs poorly on novel classes. We can observe that the AUC of the CLOM is much smaller (\\(4.35\\%\\)) than that of C-FSCIL. The results confirm the necessity of our proposed \\(gAcc\\). Similar phenomena can be found when comparing other methods like SAVC and ALICE, and we can obtain similar conclusions on the other two datasets including CUB-200 and CIFAR-100." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Results on CUB-200 dataset are shown in Table 2 and Fig. 6. We can see from Table 2 that our method is better than any other existing methods in terms of \\( aAcc \\) and \\( gAcc \\) (1.52%@aAcc and 2.6%@gAcc improvement over the strong method SAVC). Results on CIFAR-100 dataset are shown in Table 3. For the traditional metric \\( aAcc \\), our method achieves comparable performance with SOTA methods, while our method outperforms all other methods by a clear margin when considering the newly introduced metric \\( gAcc \\). Please refer to Sec A5 of the supplementary material for more experiment results." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.246, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "Rethinking Few-shot Class-incremental Learning: Learning from Yourself" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "table_caption", + "bbox": [ + 0.215, + 0.144, + 0.788, + 0.176 + ], + "angle": 0, + "content": "Table 2: Comparison with FSCIL methods on CUB-200. The best and second best results are marked bolded and underlined. Methods with \\(\\dagger\\) are reported with the original results in their paper. Other results are reimplemented by us." + }, + { + "type": "table", + "bbox": [ + 0.256, + 0.177, + 0.744, + 0.35 + ], + "angle": 0, + "content": "
MethodsNum. of Params.(M)aAcc in each task (%)
1234567891011aAccgAcc
TOPIC [52]†-68.742.554.850.045.341.438.435.432.228.326.343.92-
DF-Replay [35]†-75.972.168.663.862.659.157.855.955.053.652.461.52-
BiDistFSCIL [66]†-79.175.472.869.167.565.164.063.561.961.560.967.34-
MetaFSCIL [8]†-75.972.468.864.863.060.058.356.854.853.852.661.93-
SubNetwork [25]†-78.774.671.467.565.462.661.159.457.557.256.764.68-
DSN [58]†-76.172.269.666.764.462.160.258.957.055.154.263.31-
Self-promoted [73]12.3662.654.751.647.344.340.938.839.035.935.133.544.0740.88
CEC [63]12.3475.771.468.164.062.058.857.355.553.653.252.161.1754.16
LIMIT [70]12.3379.576.473.269.267.164.362.861.959.859.157.666.5859.40
CLOM [76]18.8979.575.972.068.265.763.462.060.358.557.956.965.6157.40
FACT [69]11.3476.173.470.466.564.862.561.159.957.957.556.550.2957.93
NC-FSCIL [59]15.9181.176.472.869.366.163.762.261.059.058.057.366.1759.22
ALICE [42]41.7175.870.168.465.263.761.260.759.758.558.758.463.7760.28
S3C [24]13.0680.276.572.969.166.964.362.761.059.559.258.366.5361.05
SAVC [49]24.2981.177.675.071.269.666.765.864.862.962.762.268.3362.90
YourSelf (Ours)11.9083.477.075.372.269.066.866.065.664.164.563.669.8565.50
" + }, + { + "type": "table_caption", + "bbox": [ + 0.215, + 0.354, + 0.788, + 0.385 + ], + "angle": 0, + "content": "Table 3: Comparison results on CIFAR-100. The best and second best results are marked bolded and underlined. All results of other methods are reimplemented by us. See more results in the supplementary material." + }, + { + "type": "table", + "bbox": [ + 0.293, + 0.386, + 0.711, + 0.517 + ], + "angle": 0, + "content": "
MethodsNum. of Params.(M)aAcc in each task (%)aAccgAcc
123456789
CEC [63]0.29673.269.165.461.258.055.553.251.349.259.5650.85
LIMIT [70]0.29676.171.767.363.260.157.355.253.050.861.6452.16
CLOM [76]0.35574.269.866.162.259.056.154.051.949.860.3550.36
C-FSCIL [20]12.7976.872.067.263.159.756.754.451.949.761.2850.62
SAVC [49]0.84278.472.768.464.061.258.356.054.151.762.7653.29
FACT [69]0.28278.572.768.864.561.358.656.854.452.363.1254.56
SubNet [25]22.3980.075.771.667.864.761.459.357.155.164.8556.63
S3C [24]0.34178.073.970.266.163.360.158.356.754.064.5158.04
ALICE [42]45.2880.371.967.063.260.758.357.355.553.763.1059.72
NC-FSCIL [59]15.9382.977.573.769.065.662.059.757.855.467.0659.73
Ours11.6882.976.372.967.865.262.060.758.856.667.0260.32
" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.522, + 0.4, + 0.536 + ], + "angle": 0, + "content": "5.3 Ablation studies" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.553, + 0.788, + 0.779 + ], + "angle": 0, + "content": "Effectiveness of different losses. Ablation experiments of the effectiveness of \\(\\mathcal{L}_{IR}\\) and \\(\\mathcal{L}_{CR}\\) are conducted among CIFAR-100, miniImageNet, and CUB-200 datasets. Experimental results are shown in Table 4 and Table 5. The detailed \\(gAcc\\) vs. \\(\\alpha\\) as shown in Fig. 7. Results on CIFAR-100 show that introducing the FR module can effectively enhance the \\(gAcc\\) by \\(3.28\\%\\). After further introducing the class-relation (CR) transfer loss, the performance at \\(gAcc\\) is further improved by \\(1.07\\%\\). Adding the instance-relation (IR) transfer loss will further boost the \\(gAcc\\) by \\(1.91\\%\\). The effectiveness of the proposed losses is also verified on the miniImageNet dataset. For miniImageNet dataset, the introduction of FR modules brings \\(3.02\\%\\) improvement at \\(gAcc\\). Applying loss \\(\\mathcal{L}_{CR}\\) and \\(\\mathcal{L}_{IR}\\) can improve \\(gAcc\\) by \\(0.49\\%\\) and \\(0.82\\%\\), respectively. It is worth mentioning that \\(\\mathcal{L}_{CR}\\) and \\(\\mathcal{L}_{IR}\\) can also boost the \\(aAcc\\). Specifically, on the miniImageNet, adding these two losses results in an increase at \\(aAcc\\) performance of \\(1.96\\%\\). On CIFAR-100, the \\(aAcc\\) is improved by about \\(0.76\\%\\) with the help of these two losses. See more experiments in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Effectiveness of multi-layer knowledge ensemble. In Sec. 4.2, we propose to exploit hierarchical knowledge from multiple layers for effective feature rectification. In practice, we select layers 8, 9 and 10 for the ensemble. Here, we take a more detailed discussion on each layer before the ensemble of multiple" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.347, + 0.129 + ], + "angle": 0, + "content": "Tang et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.239, + 0.166, + 0.483, + 0.177 + ], + "angle": 0, + "content": "Table 4: Ablation studies on CIFAR-100." + }, + { + "type": "table", + "bbox": [ + 0.2, + 0.179, + 0.553, + 0.226 + ], + "angle": 0, + "content": "
FR\\( {\\mathcal{L}}_{\\text{Cos }} \\)\\( {\\mathcal{L}}_{\\text{NovCE }} \\)\\( {\\mathcal{L}}_{CR} \\)\\( {\\mathcal{L}}_{IR} \\)aAcc. in each task (%)aAccgAcc
123456789
82.977.872.968.464.661.458.756.454.166.3554.20
✓✓82.976.173.167.964.860.859.257.253.966.2657.48
✓✓✓82.976.172.467.864.561.359.657.354.966.3358.55
✓✓✓✓82.975.772.567.464.861.960.458.756.367.0260.32
" + }, + { + "type": "table_caption", + "bbox": [ + 0.23, + 0.228, + 0.492, + 0.239 + ], + "angle": 0, + "content": "Table 5: Ablation studies on miniImageNet." + }, + { + "type": "table", + "bbox": [ + 0.2, + 0.239, + 0.553, + 0.289 + ], + "angle": 0, + "content": "
FRLcos+ LNoveCELCRLIRaAcc. in each task (%)aAccgAcc
123456789
84.079.475.071.167.864.461.358.856.968.7457.87
✓✓84.074.069.766.866.163.560.858.658.166.8460.89
✓✓✓84.077.773.669.967.464.661.659.858.768.6061.38
✓✓✓✓84.077.673.770.068.064.962.159.859.068.8062.20
" + }, + { + "type": "image", + "bbox": [ + 0.555, + 0.164, + 0.683, + 0.276 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.683, + 0.164, + 0.81, + 0.276 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.554, + 0.277, + 0.812, + 0.299 + ], + "angle": 0, + "content": "Fig.7: Ablation studies on two datasets, evaluating using \\( gAcc \\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.306, + 0.788, + 0.412 + ], + "angle": 0, + "content": "layers. We conducted experiments on the miniImageNet, and the results are shown in Table 6. It is clear that for each layer, the corresponding FR module effectively calibrates the features and achieves improvement from \\(2.23\\%\\) to \\(3.54\\%\\) at \\(gAcc\\). Furthermore, the introduction of the FR module does not harm the \\(aAcc\\). Considering the fact that different layers contain hierarchical knowledge, the multi-layer ensemble is proposed. The results in the table show that the ensemble is effective in improving the \\(gAcc\\) while maintaining \\(aAcc\\)." + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.438, + 0.788, + 0.464 + ], + "angle": 0, + "content": "Table 6: Detailed performance of each layer. Results are from miniImagenet dataset. 'l-Raw': vanilla ViT features from layer \\(L_{l}\\). 'l-Recti': Rectified features using the proposed FR module." + }, + { + "type": "table_caption", + "bbox": [ + 0.221, + 0.466, + 0.568, + 0.479 + ], + "angle": 0, + "content": "'Ensemble': Ensembled results from \\(l\\)-Recti', \\(l \\in \\{8, 9, 10\\}\\)." + }, + { + "type": "table", + "bbox": [ + 0.306, + 0.491, + 0.698, + 0.597 + ], + "angle": 0, + "content": "
LayersaAcc in each session (%)aAccgAcc
012345678
8-Raw78.973.269.165.462.960.057.355.254.064.0156.61
8-Recti80.773.470.166.664.461.559.357.056.565.5260.15
9-raw82.577.373.269.765.363.760.958.957.567.6659.33
9-Recti83.076.673.269.767.564.661.959.959.068.3961.90
10-raw83.878.674.470.766.464.361.559.357.768.5558.90
10-Recti84.378.274.170.468.264.861.959.558.468.8461.13
Ensemble84.077.673.770.068.064.962.159.859.068.8062.20
" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.654, + 0.36, + 0.67 + ], + "angle": 0, + "content": "6 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.789, + 0.841 + ], + "angle": 0, + "content": "In this work, we have revisited the challenges of the FSCIL, showing the limitation of the existing metric and introducing a novel metric denoted as generalized accuracy (\\(gAcc\\)) for the comprehensive evaluation of FSCIL methods from a more balanced perspective. Furthermore, we incorporate the ViT into the FSCIL framework and unlock the potential of the ViT intermediate layer features. To achieve this, we have designed a feature rectification (FR) module aimed at integrating the valuable information from the intermediate layer features into the rectified feature space. Using the new metric \\(gAcc\\), we analyze the existing open-sourced FSCIL methods from a more balanced perspective and show the superiority of our method on three popular datasets." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.246, + 0.115, + 0.734, + 0.13 + ], + "angle": 0, + "content": "Rethinking Few-shot Class-incremental Learning: Learning from Yourself" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.146, + 0.403, + 0.163 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.179, + 0.788, + 0.24 + ], + "angle": 0, + "content": "This work was supported partially by the National Key Research and Development Program of China (2023YFA1008503), NSFC(U21A20471, 62206315), Guangdong NSF Project (No. 2023B1515040025, 2020B1515120085, 2024A15150-10101), Guangzhou Basic and Applied Basic Research Scheme(2024A04J4067)." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.264, + 0.323, + 0.279 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.296, + 0.785, + 0.338 + ], + "angle": 0, + "content": "1. Agarwal, A., Banerjee, B., Cuzzolin, F., Chaudhuri, S.: Semantics-driven generative replay for few-shot class incremental learning. In: Proceedings of the 30th ACM International Conference on Multimedia (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.34, + 0.785, + 0.393 + ], + "angle": 0, + "content": "2. Ahmad, T., Dhamija, A.R., Cruz, S., Rabinowitz, R., Li, C., Jafarzadeh, M., Boult, T.E.: Few-shot class incremental learning leveraging self-supervised features. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.395, + 0.785, + 0.435 + ], + "angle": 0, + "content": "3. Akyurek, A.F., Akyurek, E., Wijaya, D.T., Andreas, J.: Subspace regularizers for few-shot class incremental learning. In: International Conference on Learning Representations (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.437, + 0.785, + 0.478 + ], + "angle": 0, + "content": "4. Castro, F.M., Marín-Jiménez, M.J., Guil, N., Schmid, C., Alahari, K.: End-to-end incremental learning. In: Proceedings of the European Conference on Computer Vision (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.479, + 0.785, + 0.506 + ], + "angle": 0, + "content": "5. Chen, H., Wang, Y., Xu, C., Xu, C., Tao, D.: Learning student networks via feature embedding. IEEE Transactions on Neural Networks and Learning Systems (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.507, + 0.785, + 0.534 + ], + "angle": 0, + "content": "6. Chen, K., Lee, C.G.: Incremental few-shot learning via vector quantization in deep embedded space. In: International Conference on Learning Representations (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.535, + 0.785, + 0.589 + ], + "angle": 0, + "content": "7. Cheraghian, A., Rahman, S., Fang, P., Roy, S.K., Petersson, L., Harandi, M.: Semantic-aware knowledge distillation for few-shot class-incremental learning. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.59, + 0.785, + 0.631 + ], + "angle": 0, + "content": "8. Chi, Z., Gu, L., Liu, H., Wang, Y., Yu, Y., Tang, J.: Meta-scil: A meta-learning approach for few-shot class incremental learning. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.633, + 0.785, + 0.673 + ], + "angle": 0, + "content": "9. Cui, Y., Deng, W., Chen, H., Liu, L.: Uncertainty-aware distillation for semi-supervised few-shot class-incremental learning. IEEE Transactions on Neural Networks and Learning Systems (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.674, + 0.785, + 0.715 + ], + "angle": 0, + "content": "0. Cui, Y., Deng, W., Xu, X., Liu, Z., Liu, Z., Pietikainen, M., Liu, L.: Uncertainty-guided semi-supervised few-shot class-incremental learning with knowledge distillation. IEEE Transactions on Multimedia (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.716, + 0.785, + 0.757 + ], + "angle": 0, + "content": "1. Cui, Y., Xiong, W., Tavakolian, M., Liu, L.: Semi-supervised few-shot class-incremental learning. In: IEEE International Conference on Image Processing (ICIP) (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.758, + 0.785, + 0.798 + ], + "angle": 0, + "content": "2. Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: Imagenet: A large-scale hierarchical image database. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2009)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.799, + 0.785, + 0.84 + ], + "angle": 0, + "content": "3. Deng, Y., Xiang, X.: Expanding hyperspherical space for few-shot class-incremental learning. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV) (2024)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.296, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.346, + 0.128 + ], + "angle": 0, + "content": "Tang et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.787, + 0.189 + ], + "angle": 0, + "content": "14. Dong, S., Hong, X., Tao, X., Chang, X., Wei, X., Gong, Y.: Few-shot class-incremental learning via relation knowledge distillation. In: Proceedings of the AAAI Conference on Artificial Intelligence (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.19, + 0.787, + 0.245 + ], + "angle": 0, + "content": "15. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. In: International Conference on Learning Representations (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.247, + 0.787, + 0.288 + ], + "angle": 0, + "content": "16. Douillard, A., Cord, M., Ollion, C., Robert, T., Valle, E.: Podnet: Pooled outputs distillation for small-tasks incremental learning. In: Proceedings of the European Conference on Computer Vision (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.289, + 0.787, + 0.33 + ], + "angle": 0, + "content": "17. Fu, S., Yan, J., Gao, Y., Xie, X., Zheng, W.S.: Asag: Building strong one-decoder-layer sparse detectors via adaptive sparse anchor generation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 6328-6338 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.332, + 0.787, + 0.36 + ], + "angle": 0, + "content": "18. Gu, Z., Xu, C., Yang, J., Cui, Z.: Few-shot continual infomax learning. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.361, + 0.787, + 0.401 + ], + "angle": 0, + "content": "19. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.402, + 0.787, + 0.444 + ], + "angle": 0, + "content": "20. Hersche, M., Karunaratne, G., Cherubini, G., Benini, L., Sebastian, A., Rahimi, A.: Constrained few-shot class-incremental learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.445, + 0.787, + 0.487 + ], + "angle": 0, + "content": "21. Hou, S., Pan, X., Loy, C.C., Wang, Z., Lin, D.: Learning a unified classifier incrementally via rebalancing. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.488, + 0.787, + 0.528 + ], + "angle": 0, + "content": "22. Ji, Z., Hou, Z., Liu, X., Pang, Y., Li, X.: Memorizing complementation network for few-shot class-incremental learning. IEEE Transactions on Image Processing (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.53, + 0.787, + 0.557 + ], + "angle": 0, + "content": "23. Jiang, W., Li, D., Hu, M., Zhai, G., Yang, X., Zhang, X.P.: Few-shot class-incremental learning with prior knowledge. arXiv preprint arXiv:2402.01201 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.559, + 0.787, + 0.586 + ], + "angle": 0, + "content": "24. Kalla, J., Biswas, S.: S3c: Self-supervised stochastic classifiers for few-shot class-incremental learning. In: European Conference on Computer Vision (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.587, + 0.787, + 0.628 + ], + "angle": 0, + "content": "25. Kang, H., Yoon, J., Madjid, S.R.H., Hwang, S.J., Yoo, C.D.: On the soft-subnetwork for few-shot class incremental learning. In: International Conference on Learning Representations (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.63, + 0.787, + 0.67 + ], + "angle": 0, + "content": "26. Kim, D.Y., Han, D.J., Seo, J., Moon, J.: Warping the space: Weight space rotation for class-incremental few-shot learning. In: International Conference on Learning Representations (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.671, + 0.787, + 0.727 + ], + "angle": 0, + "content": "27. Kim, S., Jeong, Y., Park, J.S., Yoon, S.W.: Mics: Midpoint interpolation to learn compact and separated representations for few-shot class-incremental learning. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.728, + 0.787, + 0.755 + ], + "angle": 0, + "content": "28. Krizhevsky, A., Hinton, G., et al.: Learning multiple layers of features from tiny images. Technical report, University of Toronto (2009)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.757, + 0.787, + 0.797 + ], + "angle": 0, + "content": "29. Lee, K., Lee, K., Shin, J., Lee, H.: Overcoming catastrophic forgetting with unlabeled data in the wild. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.799, + 0.787, + 0.84 + ], + "angle": 0, + "content": "30. Li, J., Dong, S., Gong, Y., He, Y., Wei, X.: Analogical learning-based few-shot class-incremental learning. IEEE Transactions on Circuits and Systems for Video Technology (2024)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.246, + 0.115, + 0.734, + 0.129 + ], + "angle": 0, + "content": "Rethinking Few-shot Class-incremental Learning: Learning from Yourself" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.19 + ], + "angle": 0, + "content": "31. Li, Y.M., Huang, W.J., Wang, A.L., Zeng, L.A., Meng, J.K., Zheng, W.S.: Egoexofitness: Towards egocentric and exocentric full-body action understanding. arXiv preprint arXiv:2406.08877 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.19, + 0.788, + 0.233 + ], + "angle": 0, + "content": "32. Li, Y.M., Zeng, L.A., Meng, J.K., Zheng, W.S.: Continual action assessment via task-consistent score-discriminative feature distribution modeling. IEEE Transactions on Circuits and Systems for Video Technology (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.233, + 0.788, + 0.261 + ], + "angle": 0, + "content": "33. Li, Z., Hoiem, D.: Learning without forgetting. IEEE transactions on pattern analysis and machine intelligence (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.261, + 0.788, + 0.303 + ], + "angle": 0, + "content": "34. Lin, K.Y., Ding, H., Zhou, J., Peng, Y.X., Zhao, Z., Loy, C.C., Zheng, W.S.: Rethinking clip-based video learners in cross-domain open-vocabulary action recognition. arXiv preprint arXiv:2403.01560 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.303, + 0.788, + 0.346 + ], + "angle": 0, + "content": "35. Liu, H., Gu, L., Chi, Z., Wang, Y., Yu, Y., Chen, J., Tang, J.: Few-shot class-incremental learning via entropy-regularized data-free replay. In: European Conference on Computer Vision (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.346, + 0.788, + 0.389 + ], + "angle": 0, + "content": "36. Liu, S., Qi, L., Qin, H., Shi, J., Jia, J.: Path aggregation network for instance segmentation. In: Proceedings of the IEEE conference on computer vision and pattern recognition (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.389, + 0.788, + 0.416 + ], + "angle": 0, + "content": "37. Van der Maaten, L., Hinton, G.: Visualizing data using t-sne. Journal of machine learning research (2008)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.416, + 0.788, + 0.445 + ], + "angle": 0, + "content": "38. Mazumder, P., Singh, P., Rai, P.: Few-shot lifelong learning. In: Proceedings of the AAAI Conference on Artificial Intelligence (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.445, + 0.788, + 0.488 + ], + "angle": 0, + "content": "39. Pan, Z., Yu, X., Zhang, M., Gao, Y.: Ssfe-net: Self-supervised feature enhancement for ultra-fine-grained few-shot class incremental learning. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.488, + 0.788, + 0.53 + ], + "angle": 0, + "content": "40. Park, W., Kim, D., Lu, Y., Cho, M.: Relational knowledge distillation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.53, + 0.788, + 0.573 + ], + "angle": 0, + "content": "41. Passalis, N., Tzelepi, M., Tefas, A.: Heterogeneous knowledge distillation using information flow modeling. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.573, + 0.788, + 0.615 + ], + "angle": 0, + "content": "42. Peng, C., Zhao, K., Wang, T., Li, M., Lovell, B.C.: Few-shot class-incremental learning from an open-set perspective. In: European Conference on Computer Vision (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.615, + 0.788, + 0.657 + ], + "angle": 0, + "content": "43. Peng, Y.X., Jiao, J., Feng, X., Zheng, W.S.: Consistent discrepancy learning for intra-camera supervised person re-identification. IEEE Transactions on Multimedia 25, 2393–2403 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.657, + 0.788, + 0.713 + ], + "angle": 0, + "content": "44. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.713, + 0.788, + 0.756 + ], + "angle": 0, + "content": "45. Rebuffi, S.A., Kolesnikov, A., Sperl, G., Lampert, C.H.: icarl: Incremental classifier and representation learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.756, + 0.788, + 0.799 + ], + "angle": 0, + "content": "46. Russakovsky, O., Deng, J., Su, H., Krause, J., Satheesh, S., Ma, S., Huang, Z., Karpathy, A., Khosla, A., Bernstein, M., et al.: Imagenet large scale visual recognition challenge. International journal of computer vision (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.799, + 0.788, + 0.842 + ], + "angle": 0, + "content": "47. Shi, G., Chen, J., Zhang, W., Zhan, L.M., Wu, X.M.: Overcoming catastrophic forgetting in incremental few-shot learning by finding flat minima. In: Advances in neural information processing systems (2021)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.842 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.346, + 0.129 + ], + "angle": 0, + "content": "Tang et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.785, + 0.189 + ], + "angle": 0, + "content": "48. Smith, J., Hsu, Y.C., Balloch, J., Shen, Y., Jin, H., Kira, Z.: Always be dreaming: A new approach for data-free class-incremental learning. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.19, + 0.785, + 0.245 + ], + "angle": 0, + "content": "49. Song, Z., Zhao, Y., Shi, Y., Peng, P., Yuan, L., Tian, Y.: Learning with fantasy: Semantic-aware virtual contrastive constraint for few-shot class-incremental learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.246, + 0.785, + 0.286 + ], + "angle": 0, + "content": "50. Tang, Y.M., Peng, Y.X., Zheng, W.S.: Learning to imagine: Diversify memory for incremental learning using unlabeled data. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.287, + 0.785, + 0.327 + ], + "angle": 0, + "content": "51. Tang, Y.M., Peng, Y.X., Zheng, W.S.: When prompt-based incremental learning does not meet strong pretraining. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.328, + 0.785, + 0.37 + ], + "angle": 0, + "content": "52. Tao, X., Hong, X., Chang, X., Dong, S., Wei, X., Gong, Y.: Few-shot class incremental learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.371, + 0.785, + 0.412 + ], + "angle": 0, + "content": "53. Touvron, H., Cord, M., Douze, M., Massa, F., Sablayrolles, A., Jégou, H.: Training data-efficient image transformers & distillation through attention. In: International Conference on Machine Learning (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.412, + 0.785, + 0.452 + ], + "angle": 0, + "content": "54. Vaze, S., Han, K., Vedaldi, A., Zisserman, A.: Open-set recognition: A good closed-set classifier is all you need? In: International Conference on Learning Representations (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.453, + 0.785, + 0.479 + ], + "angle": 0, + "content": "55. Wah, C., Branson, S., Welinder, P., Perona, P., Belongie, S.: The caltech-ucsd birds-200-2011 dataset (2011)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.48, + 0.785, + 0.522 + ], + "angle": 0, + "content": "56. Wang, Q.W., Zhou, D.W., Zhang, Y.K., Zhan, D.C., Ye, H.J.: Few-shot class-incremental learning via training-free prototype calibration. In: Advances in Neural Information Processing Systems (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.523, + 0.785, + 0.563 + ], + "angle": 0, + "content": "57. Yan, S., Xie, J., He, X.: Der: Dynamically expandable representation for class incremental learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.565, + 0.785, + 0.605 + ], + "angle": 0, + "content": "58. Yang, B., Lin, M., Zhang, Y., Liu, B., Liang, X., Ji, R., Ye, Q.: Dynamic support network for few-shot class incremental learning. IEEE Transactions on Pattern Analysis and Machine Intelligence (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.606, + 0.785, + 0.646 + ], + "angle": 0, + "content": "59. Yang, Y., Yuan, H., Li, X., Lin, Z., Torr, P., Tao, D.: Neural collapse inspired feature-classifier alignment for few-shot class incremental learning. In: International Conference on Learning Representations (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.647, + 0.785, + 0.674 + ], + "angle": 0, + "content": "60. Yao, G., Zhu, J., Zhou, W., Li, J.: Few-shot class-incremental learning based on representation enhancement. Journal of Electronic Imaging (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.675, + 0.785, + 0.715 + ], + "angle": 0, + "content": "61. Yu, L., Twardowski, B., Liu, X., Herranz, L., Wang, K., Cheng, Y., Jui, S., Weijer, J.v.d.: Semantic drift compensation for class-incremental learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.716, + 0.785, + 0.757 + ], + "angle": 0, + "content": "62. Yuan, Y., Fu, R., Huang, L., Lin, W., Zhang, C., Chen, X., Wang, J.: Hrformer: High-resolution vision transformer for dense predict. Advances in Neural Information Processing Systems (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.758, + 0.785, + 0.799 + ], + "angle": 0, + "content": "63. Zhang, C., Song, N., Lin, G., Zheng, Y., Pan, P., Xu, Y.: Few-shot incremental learning with continually evolved classifiers. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.799, + 0.785, + 0.84 + ], + "angle": 0, + "content": "64. Zhang, J., Zhang, J., Ghosh, S., Li, D., Tasci, S., Heck, L., Zhang, H., Kuo, C.C.J.: Class-incremental learning via deep model consolidation. In: Proceedings of the IEEE Winter Conference on Applications of Computer Vision (2020)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.246, + 0.115, + 0.734, + 0.129 + ], + "angle": 0, + "content": "Rethinking Few-shot Class-incremental Learning: Learning from Yourself" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "19" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.189 + ], + "angle": 0, + "content": "65. Zhao, B., Xiao, X., Gan, G., Zhang, B., Xia, S.T.: Maintaining discrimination and fairness in class incremental learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.19, + 0.785, + 0.232 + ], + "angle": 0, + "content": "66. Zhao, L., Lu, J., Xu, Y., Cheng, Z., Guo, D., Niu, Y., Fang, X.: Few-shot class-incremental learning via class-aware bilateral distillation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.232, + 0.785, + 0.272 + ], + "angle": 0, + "content": "67. Zheng, G., Zhang, A.: Few-shot class-incremental learning with meta-learned class structures. In: International Conference on Data Mining Workshops (ICDMW) (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.273, + 0.785, + 0.313 + ], + "angle": 0, + "content": "68. Zheng, W.S., Yan, J., Peng, Y.X.: A versatile framework for multi-scene person re-identification. IEEE Transactions on Pattern Analysis and Machine Intelligence (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.315, + 0.785, + 0.355 + ], + "angle": 0, + "content": "69. Zhou, D.W., Wang, F.Y., Ye, H.J., Ma, L., Pu, S., Zhan, D.C.: Forward compatible few-shot class-incremental learning. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.357, + 0.785, + 0.397 + ], + "angle": 0, + "content": "70. Zhou, D.W., Ye, H.J., Ma, L., Xie, D., Pu, S., Zhan, D.C.: Few-shot class-incremental learning by sampling multi-phase tasks. IEEE Transactions on Pattern Analysis and Machine Intelligence (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.398, + 0.785, + 0.438 + ], + "angle": 0, + "content": "71. Zhu, F., Zhang, X.Y., Wang, C., Yin, F., Liu, C.L.: Prototype augmentation and self-supervision for incremental learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.439, + 0.785, + 0.481 + ], + "angle": 0, + "content": "72. Zhu, J., Yao, G., Zhou, W., Zhang, G., Ping, W., Zhang, W.: Feature distribution distillation-based few shot class incremental learning. In: the 5th International Conference on Pattern Recognition and Artificial Intelligence (PRAI) (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.481, + 0.785, + 0.522 + ], + "angle": 0, + "content": "73. Zhu, K., Cao, Y., Zhai, W., Cheng, J., Zha, Z.J.: Self-promoted prototype refinement for few-shot class-incremental learning. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.523, + 0.785, + 0.563 + ], + "angle": 0, + "content": "74. Zhu, K., Zhai, W., Cao, Y., Luo, J., Zha, Z.J.: Self-sustaining representation expansion for non-exemplar class-incremental learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.565, + 0.785, + 0.591 + ], + "angle": 0, + "content": "75. Zhu, M., Zhu, Z., Chen, S., Chen, C., Wu, B.: Enhanced few-shot class-incremental learning via ensemble models. arXiv preprint arXiv:2401.07208 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.592, + 0.785, + 0.633 + ], + "angle": 0, + "content": "76. Zou, Y., Zhang, S., Li, Y., Li, R.: Margin-based few-shot class-incremental learning with class-level overfitting mitigation. In: Advances in neural information processing systems (2022)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.633 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/Rethinking Few-shot Class-incremental Learning_ Learning from Yourself/a34c73cf-0382-420e-bbe9-4f775aa32eab_origin.pdf b/2024/Rethinking Few-shot Class-incremental Learning_ Learning from Yourself/a34c73cf-0382-420e-bbe9-4f775aa32eab_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..142e3783434eca52693a8db7d7a3d1451015f1da --- /dev/null +++ b/2024/Rethinking Few-shot Class-incremental Learning_ Learning from Yourself/a34c73cf-0382-420e-bbe9-4f775aa32eab_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:76ecf6a80070bcf5327fa9e9503f9c196df3e57340c6330e90f64578e80552d2 +size 1562275 diff --git a/2024/Rethinking Few-shot Class-incremental Learning_ Learning from Yourself/full.md b/2024/Rethinking Few-shot Class-incremental Learning_ Learning from Yourself/full.md new file mode 100644 index 0000000000000000000000000000000000000000..12b9b255068ebf8402fc21112b7613352481d613 --- /dev/null +++ b/2024/Rethinking Few-shot Class-incremental Learning_ Learning from Yourself/full.md @@ -0,0 +1,361 @@ +# Rethinking Few-shot Class-incremental Learning: Learning from Yourself + +Yu-Ming Tang $^{1,3}$ , Yi-Xing Peng $^{1,3,5}$ , Jingke Meng $^{\star 1,3}$ , and Wei-Shi Zheng $^{1,2,3,4,5}$ + +$^{1}$ School of Computer Science and Engineering, Sun Yat-sen University, China + $^{2}$ Peng Cheng Laboratory, Shenzhen, China + +3 Key Laboratory of Machine Intelligence and Advanced Computing, Ministry of Education, China + +4 Guangdong Province Key Laboratory of Information Security Technology, Sun Yat-sen University, China + +$^{5}$ Pazhou Laboratory (Huangpu), Guangzhou, Guangdong 510555, China {tangym9, pengyx23}@mail2.sysu.edu.cn, mengjke@gmail.com, wszheng@ieee.org + +Abstract. Few-shot class-incremental learning (FSCIL) aims to learn sequential classes with limited samples in a few-shot fashion. Inherited from the classical class-incremental learning setting, the popular benchmark of FSCIL uses averaged accuracy ( $aAcc$ ) and last-task averaged accuracy ( $lAcc$ ) as the evaluation metrics. However, we reveal that such evaluation metrics may not provide adequate emphasis on the novel class performance, and the continual learning ability of FSCIL methods could be ignored under this benchmark. In this work, as a complement to existing metrics, we offer a new metric called generalized average accuracy ( $gAcc$ ) which is designed to provide an extra equitable evaluation by incorporating different perspectives of the performance under the guidance of a parameter $\alpha$ . We also present an overall metric in the form of the area under the curve (AUC) along the $\alpha$ . Under the guidance of $gAcc$ , we release the potential of intermediate features of the vision transformers to boost the novel-class performance. Taking information from intermediate layers which are less class-specific and more generalizable, we manage to rectify the final features, leading to a more generalizable transformer-based FSCIL framework. Without complex network designs or cumbersome training procedures, our method outperforms existing FSCIL methods at $aAcc$ and $gAcc$ on three datasets. See codes at 1 + +Keywords: Continual Learning $\cdot$ Few-shot Learning $\cdot$ Catastrophic Forgetting $\cdot$ Vision Transformer + +# 1 Introduction + +Although existing deep neural networks (DNNs) achieve great success in various real-world applications [17, 31, 32, 34, 43, 44, 54, 68], they suffer from catastrophic + +![](images/11401f85782ee84c427d469061887ee428cbb72141ea6d2507449d58dbad8c37.jpg) +(a) Performance after each task + +![](images/ecb69e171671f49bdc7f0d5182231b32f89132f5e403e6781a9b76eacd65e3e0.jpg) +(b) Detailed Performance after finishing the last task (Task 9) +Fig. 1: Performance under different metrics of two recent methods [24, 25]. (a): The dotted line denotes the average accuracy ( $aAcc$ ) widely used in classical class-incremental learning. The solid line presents our proposed generalized accuracy ( $gAcc$ ). We also show the average of each point on the right side. (Blue: $gAcc$ , Red: $aAcc$ ). (b): The detailed accuracies of each task after training on the last task (task 8). It is evident that S3C exhibits superior performance when confronted with novel classes. The conventional metric $aAcc$ fails to reflect this due to the domination by the base-class performance. Models are trained on the CIFAR-100 dataset. Best viewed in color. + +forgetting when dealing with sequential data streams. The forgetting problem has attracted heavy attention in recent years [4,16,21,29,33,45,48,50,51,57,61,64,65, 71,74]. In this work, we focus on a more realistic and challenging problem: few-shot class-incremental learning (FSCIL), in which after training on a task with sufficient data (base task), only a few samples are provided during incremental tasks (novel tasks). Since Tao et al. [52] proposed the FSCIL benchmark in 2020, a large number of algorithms have emerged to tackle this problem, including meta-learning based [8, 20, 70], structure based [2, 25, 52, 58, 63], feature space based [3, 6, 20, 24, 26, 59, 69] methods, etc. + +However, almost all these methods are evaluated following the traditional class-incremental learning setting, i.e. using the average accuracy ( $aAcc$ ) or the last-task averaged accuracy ( $lAcc$ ) as the main metrics to critique the proposed method. Such accuracies are calculated by weighted averaging the base-task and novel-tasks accuracies with the weight proportional to the number of classes of each task. Considering the reality that the FSCIL benchmarks make the majority of the classes (about $50\%$ to $60\%$ ) as the base task to establish a well-trained backbone for few-shot continual learning, the existing $aAcc$ or $lAcc$ metric is dominated by the base-class performance. In other words, current evaluation metrics can be largely boosted by only improving the performance of base classes. As shown in Fig. 1, a method like softNet [25] with higher base-class performance (see Fig. 1 (b)) usually gets higher $aAcc$ (see red lines in Fig. 1 (a)) regardless of the inferior performance on the novel classes. Due to the inadequacy to capture performance changes in novel classes, current metrics are biased toward the performance of base classes and have limitations in guiding the FSCIL community. To tackle this, we propose a more balanced metric called generalized accuracy ( $gAcc$ ). Parameterized by $\alpha$ ranging from 0 to 1, the $gAcc$ offers a way to evaluate a model with explicit emphasis on performance on novel classes. Furthermore, we can comprehensively evaluate the model by calculating the area under the curve (AUC) along the parameter $\alpha$ . It is worth mentioning that our intention + +is NOT to replace all previous metrics, but to provide a more fine-grained and balanced evaluation of FSCIL methods complementary to them. + +We empirically find that FSCIL can benefit from the powerful vision transformer (ViT) [15] by analyzing the performance of novel classes. Specifically, although vanilla ViT performs poorly in terms of unseen novel classes, we reveal that the intermediate features of ViT are less specified on base classes. It means the ViT can learn abundant semantic information but the information is mistakenly dropped in deep layers when extracting features for novel classes. + +Based on the observation, we propose a lightweight Feature Rectification (FR) module to rectify the output features of the ViT conditioned on the intermediate features. To achieve this, the FR module is supervised with two losses that work on two different levels, namely instance-level relation transfer and class-center relation transfer. The former transfers the relations between instances while the latter focuses on relations between instances and the class prototypes. We further enhance the FR module by leveraging hierarchical knowledge from multiple intermediate layers. Finally, our method boosts the novel-class accuracy and maintains a comparable performance of base classes. + +Our contributions are summarized as follows. (1) We revisit the existing FSCIL benchmark and propose a novel metric called generalized accuracy ( $gAcc$ ) for a more fine-grained and balanced evaluation. We have implemented 14 FSCIL methods with codes released and show that the newly introduced SOTAs do not necessarily perform better on $gAcc$ . (2) Based on $gAcc$ , we reveal the potential of vision transformers in FSCIL. We empirically find that although features from the intermediate layer exhibit sub-optimal performance for the base classes, their transferrable ability helps them to have a distinct advantage in handling novel classes. (3) A lightweight FR module is proposed to rectify the final features to be more transferrable. The FR module manages to make the output feature more generalizable with assistance from intermediate features. (4) We conduct extensive experiments on 3 datasets and evaluate our method under both $aAcc$ and $gAcc$ , showing the effectiveness and superiority of our method. + +# 2 Related Work + +Meta-learning based methods. Due to the limited samples of novel classes, some works [8,20,67,70] propose to use meta-learning to improve the transferability of the model. Concretely, MetaFscil [8] proposed to simulate novel tasks by sampling from the base task. C-FSCIL [20] uses meta-learning to train the backbone, which maps images from different classes to quasi-orthogonal vectors. ALFSCIL [30] formulate the process into a bi-level optimization problem. + +Optimization-based methods. There is a large group of works focusing on the optimization of the DNNs in FSCIL. CLOM [76] considers the behavior of the backbone under different margins of the cosine loss. NC-FSCIL [59] designs a pre-defined classifier for efficient FSCIL without any explicit regularizer. SPPR [73] proposes a prototype refinement strategy for stronger expression ability. FACT [69] assigns virtual prototypes in the current stage for future updates. Alice [42] instead focuses on the compactness and the variety of the feature space. Based on the classifier learning, S3C [24] proposes a self-supervised stochastic + +classifier to tackle the overfitting and forgetting problem. There are more works working on the feature space constraint [3, 6, 13, 26, 27, 56, 60], or knowledge distillation [7, 9-11, 14, 39, 66, 72]. + +Data-driven methods. There is also a branch of work focusing on the data side of this problem, which focuses on the data replay [1,35]. FSIL-GAN [1] employs a GAN to generate features to compensate for the scarcity of data. DF-replay [35] uses synthesized data generated by a generator to replay. PL-FSCIL [23] seeks prior knowledge from unlabeled data. + +Structure-based methods. In addition to these works, a unique branch of works focuses on the network structure, like topology preserving [52], growing classifier [63], proposing a dynamic structure of the backbone [58], structure fusion [2, 22], sub-networks [25] and ensemble models [75]. + +Compared with existing works, our work rephrases the challenge of the FSCIL setting and first reveals the potential of the intermediate features of the ViT backbone. There is also an existing work [42] offering an alternative metric. They did not provide a fine-grained evaluation with various perspectives and have potential risks in stability and symmetry problems. Please find more detailed discussions about related works in the supplementary material. + +# 3 Generalized Average Accuracy + +# 3.1 Limitation of the Existing Metrics + +Accuracy metric on incremental tasks. Consider a classification model $f(\cdot)$ that takes an image as input and outputs the class predictions. For an incremental problem, there are multiple tasks and multiple test sets. Let us define test set at task $\mathcal{T}_i$ as $\mathcal{D}_{test}^i = \{x_t^i,y_t^i\}_{t = 1}^n$ . Usually, the labeling space $\mathcal{Y}^i\triangleq \{y_t^i |(x_t^i,y_t^i)\in \mathcal{D}_{test}^i\}$ of every task does not have any intersection, i.e. $\forall i\neq j,\mathcal{Y}^i\bigcap \mathcal{Y}^j = \emptyset$ . After trained on $\mathcal{T}_i$ , the model can be denoted as $f_{i}(\cdot)$ . The corresponding accuracy of model $f_{i}(\cdot)$ at task $\mathcal{T}_j$ ( $j\leq i$ ) can be denoted as $\mathcal{A}_i^j\triangleq \frac{1}{n^j}\sum_{t = 1}^{n^j}\mathbb{1}(\zeta (f_i(x_t^j)) = y_t^j)$ , where $\mathbb{1}(\cdot) = 1$ if the condition is true, otherwise 0. $\zeta$ is the argmax operation. + +Average accuracy. It is worth mentioning that for class-incremental learning, evaluating the model on all the previous tasks is essential. To test the model on previous tasks, an overall test set that contains all previous test sets is constructed as $\mathcal{D}_{test}^{1\rightarrow i} = \bigcup_{m = 1}^{i}\mathcal{D}_{test}^{m}$ . After trained on task $\mathcal{T}_i$ , the accuracy of model $f_{i}(\cdot)$ on the overall test set $\mathcal{D}_{test}^{1\rightarrow i}$ can be presented as $A_{i}^{1\rightarrow i}$ . Once a $n_t$ -task incremental training ends, the average accuracy ( $aAcc$ ) can be calculated as: + +$$ +a A c c = \frac {1}{n _ {t}} \sum_ {i = 1} ^ {n _ {t}} a A c c _ {i} = \frac {1}{n _ {t}} \sum_ {i = 1} ^ {n _ {t}} A _ {i} ^ {1 \rightarrow i}, \tag {1} +$$ + +where the $aAcc_{i}$ is the average accuracy at each task. We can also derive two additional metrics incidentally: last-task averaged accuracy $lAcc = aAcc_{n_t} = A_{n_t}^{1\rightarrow n_t}$ , and task-wise average accuracy $(tAcc)$ : + +$$ +t A c c = \frac {1}{n _ {t}} \sum_ {i = 1} ^ {n _ {t}} t A c c _ {i} = \frac {1}{n _ {t}} \sum_ {i = 1} ^ {n _ {t}} \frac {1}{i} \sum_ {j = 1} ^ {i} A _ {i} ^ {j}, \tag {2} +$$ + +where the $tAcc_i$ is the task-wise accuracy at each task. Note that the $aAcc$ does not share the same value as $tAcc$ normally since the number of classes of tasks is usually not the same. This will be elaborated on later in detail. + +Limitation. Since FSCIL usually sets the majority of the classes as the base task, existing metrics potentially neglect the performance changes of novel classes. To elaborate on this, we analyze the $aAcc$ as follows. Firstly, we denote the size of the labeling space of task $\mathcal{T}_i$ as $|\mathcal{V}_i|$ . In FSCIL, the number of classes of each novel task is usually the same, i.e. $|\mathcal{V}_{novel}| = |\mathcal{V}_2| = |\mathcal{V}_3| = \dots = |\mathcal{V}_{n_t}| \ll |\mathcal{V}_1|$ . Assuming there are total $n_t$ tasks, and task $\mathcal{T}_1$ is the base task with sufficient classes and samples (non-few-shot). The following tasks $\{\mathcal{T}_i\}_{i=2}^{n_t}$ are novel tasks, with $|\mathcal{V}_{novel}|$ classes for each task and a few samples per class. Now, we can rewrite average accuracy on task $\mathcal{T}_i$ as: + +$$ +a A c c _ {i} = \frac {\left| \mathcal {Y} _ {1} \right| A _ {i} ^ {1} + \left| \mathcal {Y} _ {n o v e l} \right| \sum_ {j = 2} ^ {i} A _ {i} ^ {j}}{\left| \mathcal {Y} _ {1} \right| + (i - 1) \left| \mathcal {Y} _ {n o v e l} \right|} = \frac {\frac {\left| \mathcal {Y} _ {1} \right|}{\left| \mathcal {Y} _ {n o v e l} \right|} A _ {i} ^ {1} + \sum_ {j = 2} ^ {i} A _ {i} ^ {j}}{\frac {\left| \mathcal {Y} _ {1} \right|}{\left| \mathcal {Y} _ {n o v e l} \right|} + (i - 1)}. \tag {3} +$$ + +From Eq. (3) we see that the value of $aAcc_i$ will be dominated by the base class accuracy $A_i^1$ if the ratio $\frac{|\mathcal{Y}_1|}{|\mathcal{Y}_{\text{novel}}|}$ is large. Unfortunately, the ratio is indeed large in the FSCIL settings. (e.g. $\frac{|\mathcal{Y}_1|}{|\mathcal{Y}_{\text{novel}}|} = 12$ for both CIFAR-100 and miniImageNet dataset, 10 for CUB-200.) Under such a large ratio, the changes (both improvement and degrades) of novel classes cannot be effectively evaluated by $aAcc$ . + +# 3.2 Generalizing the Average Accuracy + +To overcome the limitations of existing metrics, we now define the generalized accuracy (gAcc) at task $\mathcal{T}_i$ as: + +$$ +g A c c _ {i} (\alpha) = \frac {\alpha \frac {| \mathcal {Y} _ {1} |}{| \mathcal {Y} _ {n o v e l} |} A _ {i} ^ {1} + \sum_ {j = 2} ^ {i} A _ {i} ^ {j}}{\alpha \frac {| \mathcal {Y} _ {1} |}{| \mathcal {Y} _ {n o v e l} |} + (i - 1)}, \tag {4} +$$ + +in which the $gAcc$ is parameterized by $\alpha \in [0,1]$ . At task $\mathcal{T}_i$ , $gAcc_i(0)$ measures the accuracy of the novel tasks only. $gAcc_i(\frac{|\mathcal{Y}_{\text{novel}}|}{|\mathcal{Y}_1|})$ denotes the task-wise accuracy $tAcc_i$ at task $\mathcal{T}_i$ while $gAcc(1)$ represents the average accuracy $aAcc_i$ at task $\mathcal{T}_i$ . In addition to these discrete values, we further generalize the $\alpha$ to any rational numbers in [0,1]. Controlled by $\alpha$ in Eq. (4), when $\alpha \to 0$ , $gAcc(\alpha)$ emphasizes to novel class while $\alpha \to 1$ , $gAcc(\alpha)$ emphasizes the base class performance, i.e., the larger the $\alpha$ , the greater the influence of the base class on the value of $gAcc(\alpha)$ . To evaluate the performance of a model on a specific task $\mathcal{T}_i$ , a standalone value is needed. Thus, we introduce the area under the curve (AUC) of $gAcc_i(\alpha)$ along the parameter $\alpha$ as a comprehensive metric: + +$$ +g A c c _ {i} = \int_ {0} ^ {1} g A c c _ {i} (\alpha) d \alpha . \tag {5} +$$ + +Similar to $aA_{cc}$ , for multiple tasks, we average the $gA_{cc_i}$ for each task and get an overall metric $gA_{cc}$ for the whole learning process: + +$$ +g A c c = \frac {1}{n _ {t}} \sum_ {i} ^ {n _ {t}} g A c c _ {i}. \tag {6} +$$ + +![](images/a53e9f37d5a627240f619f39ff48fc55fcbfd334b536fff37414fda673768647.jpg) +(b) $gAcc_{i}$ curve w.r.t Task i + +![](images/f0b4f97fa5743c99779712f70fd62bb29d5c2ae4964f5a990daa41f26436e59b.jpg) +Fig. 2: FSCIL performances are shown in generalized accuracy. (a): The $gAcc$ curve (averaged across all $n_t$ tasks) v.s. param $\alpha$ . (b): The $gAcc$ AUC of each task $\mathcal{T}_i$ (Eq. (5)). In the legend, we show the average AUC across tasks(Eq. (6)) of each method. We evaluate recent works including SAVC [49], NC [59], and Alice [42] on the miniImageNet dataset. Some corner cases: Lazy: The model maintains base-task performance while refusing to learn anything from novel tasks. Greedy: Regardless of the previous knowledge, the model only greedily focuses on the current task. Greedy-NF: A Non-Forget version of 'Greedy'. See more about these cases in our supplementary material. + +The proposed $gAcc$ provides different evaluation perspectives using different $\alpha$ so we believe such a metric can provide a more balanced evaluation of FSCIL. + +# 3.3 Analysis + +We further study the necessity of the proposed $gAcc$ . We evaluate recent SOTAs in FSCIL including SAVC [49], Alice [42] and NC-FSCIL [59], and the results are in Fig. 2. We show that the $gAcc$ provides more fine-grained evaluation than the classical accuracy metric. For example, in the left part of Fig. 2, although the $gAcc(1)$ , which is equal to the $aAcc$ , of SAVC is close to that of NC-FSCIL, the area under the curve(AUC) of SAVC is much smaller than NC-FSCIL. This indicates that the novel class performance is much worse for SAVC, which cannot be discovered by conventional metrics. + +Furthermore, to show the superiority of our metric, we consider three corner cases: Lazy, Greedy, and Greedy-NF. 'Lazy' stands for a type of algorithm that only focuses on maintaining base-class performance, i.e. $\forall i \geq j > 1$ , $A_{i}^{j} = 0$ , $A_{i}^{1} = A_{1}^{1}$ . By setting $A_{1}^{1} = 85\%$ , we can see that when $\alpha = 1$ (i.e. the traditional aAcc metric), its performance is close to SOTA methods. Further, when considering popular forgetting metrics like performance dropping $(PD = A_{i}^{j} - A_{j}^{j})$ and Knowledge rate $(KR = A_{i}^{j} / A_{j}^{j})$ , because the forgetting is zero, its PD is 0 and KR is 1. All these metrics indicate such 'Lazy' model is superior, which is contradictory to common sense. However, our proposed gAcc reveals that its performance is far from satisfying by setting $\alpha < 1$ and its AUC is far smaller than other methods. In other words, using the traditional metrics will indicate algorithms like 'Lazy' achieving SOTA performance while such methods perform poorly evaluated by our proposed gAcc. The gAcc can also discover other corner cases. 'Greedy' is to greedily learn the current knowledge (i.e. $A_{1}^{1} = 85\%$ , $A_{i}^{j} = 100\%$ , when $i = j$ else 0). "Greedy" achieves low both aAcc and gAcc. We further eliminate the forgetting problem from 'Greedy' and define 'Greedy-NF' on top of it: set accuracy $A_{i}^{j}$ to $100\%$ for $\forall j > 1, j < i$ (no forgetting at novel classes). We find that although 'Greedy-NF' performs + +![](images/5c2caab1b7a2d3e639325541be9808aa8092de4b0332b1a53223ef1eb52dc1ca.jpg) +Fig. 3: t-SNE [37] visualization of miniImageNet test set features from different layers of a ViT. Left: we choose layers $L_{9}$ , $L_{10}$ , $L_{11}$ , $L_{12}$ (final) and show both base classes and novel classes features. It is observed that shallow features are more dispersed than deep features. Right: Accuracies of different tasks on features from various layers. It is clear that intermediate features achieve better performance on novel classes and thus have better generalization ability. Best viewed in color. + +![](images/879e596e7a8f7f01e548a584ae9d6152204d8d03633b5bd54d3bd86227312a83.jpg) + +well when majorly considering the novel-class performance $(\alpha \in \{0, \frac{1}{12}, \frac{2}{12}, \frac{3}{12}\})$ , the overall performance (AUC) is not satisfying. Please refer to Sec. A1 in the supplementary material for more details about these corner cases. + +By analyzing these cases, we show the superior properties of the proposed $gAcc$ . Firstly, the $gAcc$ metric effectively distinguishes methods (e.g. 'Lazy') that only retain the base-task performance, thereby encouraging methods to learn from novel classes. Besides, $gAcc$ avoids overemphasis on the performance of base classes or novel classes (corner case 'Greedy-NF' is still far worse than normal methods in terms of AUC), providing a more balanced evaluation outcome. In other words, the proposed $gAcc$ metric provides objective and balanced results when evaluating FSCIL methods, eliminating 'shortcuts' in $aAcc$ that focus only on the performance of the base classes. + +In conclusion, the $gAcc$ curve and the AUC are necessary and powerful tools to evaluate the FSCIL methods. + +# 4 Learn from Yourself + +# 4.1 Preliminary and The FSCIL Setting + +For task $\mathcal{T}_i\in \{\mathcal{T}_i\}_{i = 1}^T$ , the training set can be denoted as $\mathcal{D}_{train}^{i} = \{x_{t}^{i},y_{t}^{i}\}_{t = 1}^{m^{i}}$ . For FSCIL settings, the base task $\mathcal{T}_1$ contains $|\mathcal{V}_1|$ classes and sufficient samples for training. For $\forall i > 1$ , there are $|\mathcal{V}_{novel}|$ classes for each task and $k$ samples per class. Typically, $|\mathcal{V}_{novel}|$ is set to 5, and $k = 5$ , i.e. a 5-way 5-shot setting. We rephrase the classification model $f(\cdot)$ in Sec. 3 as $f(\cdot) = \varPhi (f_B(\cdot))$ where $\varPhi (\cdot)$ is the classifier and $f_{B}(\cdot)$ is the backbone. We follow previous works to use a cosine classifier: $\varPhi(X)=P^\top\overline{X}$ , where $P\in \mathbb{R}^{d\times |\mathcal{V}|}$ and $\overline{X}\in \mathbb{R}^{d\times 1}$ are the L2-normalized inputs, and $d$ is the feature dimension of $f_{B}(\cdot)$ , $|\mathcal{V}| = \sum_{i}|\mathcal{V}_{i}|$ is the number of classes seen so far. In this work, we employ the Vision Transformer(ViT) [15] in the FSCIL setting. For a ViT with $N_{L}$ layers $\{L_l\}_{l = 1}^{N_L}$ , given a input image $x$ , the output can be denoted as $X_{N_L} = L_{N_L}(L_{N_L - 1}(\dots L_1(x))$ . + +![](images/328db429ab746e1ab8b477ee54a346b95e253da5d3f0de2c69c16cde27945ca8.jpg) +Fig. 4: The overall framework of our method. The Feature Rectification (FR) module takes the final and the intermediate features as input and outputs the rectified features for classification. The relation transfer losses are designed to convey valuable information from intermediate features to the rectified ones. The details of these two parts are shown in the left and lower parts of the figure. Further, a cosine constraint is proposed to maintain the base class performance and a classification loss is employed to adapt to the novel classes. Best viewed in color. + +Baseline. Recent works [3, 20, 24, 38, 47, 49, 59, 63, 76] suggest a training-frozen scheme, which trains the backbone $f_{B}(\cdot)$ and the classifier $\varPhi(\cdot)$ (if available) on the base-task only, and then freeze the backbone during novel tasks. For novel classes, the classifier is extended by mean features (i.e. prototypes) extracted by the frozen backbone. We follow and employ the scheme as our baseline. + +Main idea of our method. Given that the backbone is initially trained for the base task and remains fixed during novel tasks, the main challenges lie in the potential overfitting on the base task and the poor feature extraction ability of the novel classes. We conduct a toy experiment to show these problems and show that the intermediate feature can convey valuable information to tackle these problems. Results are shown in Fig. 3, where we select 4 base classes and 5 novel classes and utilize the t-SNE [37] to visualize the features from different layers. It can be concluded that features extracted by the base-classes trained backbone are overfitted to the base classes and cannot be generalized well on novel classes, i.e. the final features are well aggregated, but it is hard to distinguish among unseen classes. (see $L_{12}$ features on the left and the corresponding performance on the right). Compared with the final features (i.e. $L_{12}$ ), the shallower intermediate features $(L_9, L_{10}, L_{11})$ are more dispersed, which makes them more discriminative for unseen classes. This is cross-verified by the performance shown on the right side of Fig. 3. We can see that using intermediate features can largely boost the novel classes' performance (up to $15\%$ improvement) with a price of performance decline of the base classes. Based on this observation, we propose to rectify the final feature based on the intermediate feature to better handle both base and novel classes. + +# 4.2 Feature Rectification + +We propose a Feature Rectification (FR) module to rectify the final feature based on the relation information from the intermediate features. Specifically, + +given a pair of final feature $X_{N_L}$ and the intermediate feature $X_l$ , we design an FR module conditioned on the intermediate features: + +$$ +X _ {F R} = M _ {m i x} \left(\operatorname {c a t} \left(M _ {F} \left(X _ {N _ {L}}\right), M _ {I} \left(X _ {l}\right)\right), \right. \tag {7} +$$ + +where 'cat' means the concatenate operation. $M_{\text{mix}}$ , $M_F$ and $M_I$ are 2-layer MLPs. By introducing the intermediate features, the FR module has access to a reference for extracting the valuable structural information lies in the intermediate features. A naive way to mine such information is to directly pull rectified features closer to shallow features on an instance-by-instance basis. Although such an instance-wise loss can bridge shallow features and rectified features, directly pulling the rectified feature to the shallow one could lead to noisy features since the shallow features are more scattered, as shown in Sec. 4.1. + +To effectively transfer discriminative cues among samples from shallow features to rectified features and avoid introducing noise, we propose two relation-based transfer losses at different levels. Firstly, we offer a Instance-level Relation transfer (IR) to mine the relation between feature instances. Given a pair of final features $X_{N_L}^1$ and $X_{N_L}^2$ , and another pair of intermediate features $X_l^1$ and $x_{l}^{2}$ , the IR loss is defined as: + +$$ +\mathcal {L} _ {I R} = \delta \left(\epsilon \left(X _ {F R} ^ {1}, X _ {F R} ^ {2}\right), \epsilon \left(X _ {l} ^ {1}, x _ {l} ^ {2}\right)\right), \tag {8} +$$ + +where $\epsilon (\cdot)$ is the Euclidean distance and the $\delta (\cdot)$ denotes the smooth L1 loss. Besides the relation between feature instances, another level of the relation information is the relation between features and class centers. Class-center Relation Transfer (CR) is proposed to mine this class-center-level information. Assume at task $T_{i}$ with $|\mathcal{V}_i|$ classes and the feature dimension is $d$ . Given a set of class prototypes $P\in \mathbb{R}^{d\times |\mathcal{V}_i|}$ from the classifier $\varPhi(\cdot)$ as well as a pair of final feature $X_{N_L}\in \mathbb{R}^{d\times 1}$ and the intermediate feature $X_{l}$ $(l\in [1,N_L))$ , the corresponding CR loss is formulated as: + +$$ +\mathcal {L} _ {C R} = \mathrm {K L} \left(\left(\sigma \left(P ^ {\top} \overline {{X _ {F R}}}\right), \sigma \left(P ^ {\top} \overline {{X _ {l}}}\right)\right), \right. \tag {9} +$$ + +where $\sigma (\cdot)$ is the softmax function, and $\mathrm{KL}(\cdot)$ is the Kullback-Leibler divergence. $\overline{X}$ denotes the L2 normalization results of $X$ . + +In the above equations, the two losses $\mathcal{L}_{CR}$ and $\mathcal{L}_{IR}$ work on different levels and serve different roles. $\mathcal{L}_{CR}$ works on the relations between a sample and all class centers, focusing on the coarse-grained class-specific patterns that distinguish different classes. Differently, $\mathcal{L}_{IR}$ provides knowledge about fine-grained discriminative cues that contribute to pairwise relations. The proposed relation-based transfer losses are related to previous knowledge distillation works [5, 40, 41]. However, these works do not share the same objective as ours. Our goal is to extract precious transferrable information hidden in various shallow layers hierarchically to rectify the final features while prior distillation methods focus on matching final outputs to teachers. Furthermore, there exists a branch of works using intermediate features [36, 62] directly for dense prediction. Differently, we are the first to extract intermediate knowledge to rectify the final feature for a classification task. + +Besides, we constrain the discriminative ability of rectified features on base classes to avoid negative rectification. Since the final feature of layer $L_{N_L}$ is strong enough for the base classes, we apply a cosine constraint as: + +$$ +\mathcal {L} _ {\text {c o s}} = \cos \left(X _ {F R}, X _ {N _ {L}}\right), \tag {10} +$$ + +where $\cos (\cdot ,\cdot)$ is cosine similarity. + +Novel class training. To ensure the rectified features are centered around the class mean prototypes, we further introduce a classification loss for novel classes: + +$$ +\mathcal {L} _ {\text {N o v C E}} = - \mathbb {1} (y \in \cup_ {n > 1} \mathcal {Y} _ {n}) \log \sigma \left(P ^ {\top} \overline {{X _ {F R}}}\right) ^ {(y)}, \tag {11} +$$ + +where $y$ is the corresponding ground true label of $X_{FR}$ , $(P^{\top}\overline{X_{FR}})^{(y)}$ is the $y$ -th element of $(P^{\top}\overline{X_{FR}})$ . + +Anti-forgetting training. Another challenge of FSCIL is the forgetting problem. Thanks to our feature rectification design, the only trainable module is the FR module. As the FR module takes the final and the intermediate feature as the input, these kinds of features can be easily modeled by Gaussian distributions [51,71]. Hence, we construct a Gaussian distribution for each class on each layer. The Gaussian distribution of class $i$ features at layer $L_{l}$ can be denoted as $\mathcal{N}_i^l (\mu_i^l,\Sigma_i^l)$ , where $\Sigma_i^l$ is obtained by calculating the covariance of each feature dimension if $i\in \mathcal{Y}_1$ (base classes). For novel classes, the covariance matrix is computed by averaging the covariance matrices of the top-k most similar base classes. During the incremental learning phases, we utilize these Gaussian distributions to sample features, which serve as inputs from old tasks to the FR module, assisting the FR module in retaining previously learned knowledge. + +Overall training objectives. The total loss function for our feature rectification framework is as follows: + +$$ +\mathcal {L} = \beta_ {\text {c o s}} \left(\mathcal {L} _ {\text {c o s}} + \mathcal {L} _ {\text {N o v C e}}\right) + \beta_ {C R} \mathcal {L} _ {C R} + \beta_ {I R} \mathcal {L} _ {I R}, \tag {12} +$$ + +where $\beta_{cos},\beta_{CR}$ and $\beta_{IR}$ are trade-off parameters. + +Multi-layer knowledge ensemble. Since there exists hierarchical knowledge in different layers, we mine such information by knowledge ensemble from different layers. Specifically, we incorporate an FR module with each selected layer, resulting in a total of $n_{FR}$ FR modules. During the training phase, these FR modules are optimized simultaneously. When evaluation, FR modules at layer $l$ are employed to produce multiple rectified features denoted as $X_{FR}^{l}$ . These rectified features are subsequently input into the shared classifier $P$ for generating multiple predictions: $\mathrm{Pred}_l = \sigma (P^\top \overline{X_{FR}^l})$ . The final prediction is computed as the average of each FR-branch: $\mathrm{Pred} = \frac{1}{n_{FR}}\sum_{l = 1}^{n_{FR}}\mathrm{Pred}_l$ . + +# 5 Experiments + +# 5.1 Experiment Setup + +Datasets. Following previous works, we conduct experiments on three popular datasets of FSCIL: CIFAR-100 [28], miniImageNet [46] and CUB-200 [55]. For statistics of the three datasets, please refer to the supplementary material. + +Table 1: Comparison with FSCIL methods on miniImageNet. The best and second best results are marked bolded and underlined. Methods with†are reported with the original results in their paper. Other results are reimplemented by us. + +
MethodsNum. of +Params.(M)aAcc in each task (%)
123456789aAccgAcc
TOPIC [52]†-61.350.145.241.237.535.532.229.524.439.64-
IDLVQ [6]†-64.859.855.952.649.947.644.843.141.851.16-
MetaFSCIL [8]†-72.068.063.860.357.655.253.050.849.258.85-
DF-Replay [35]†-71.867.163.259.857.054.051.649.548.258.02-
FCIL [18]†-76.3471.467.164.161.358.555.754.152.862.37-
Self-promoted [73]12.3463.664.159.555.552.149.245.643.140.952.6241.53
CEC [63]12.1272.167.063.059.656.953.951.449.547.857.9148.64
LIMIT [70]12.2973.367.663.259.556.753.751.449.748.258.1549.47
FACT [69]11.2676.371.366.962.959.656.353.551.149.260.7849.56
CLOM [76]14.2272.166.762.759.256.153.050.248.346.757.2346.11
C-FSCIL [20]12.7970.663.859.356.454.351.848.947.346.155.4050.46
BiDistFSCIL [66]11.5074.670.466.662.860.557.254.753.051.961.3253.49
S3C [24]0.3477.072.168.265.162.058.856.053.651.962.7453.70
Regularizer [3]26.3081.972.768.764.157.453.553.852.343.162.8255.62
SubNetwork [25]22.3878.173.569.866.263.260.357.655.754.564.3155.88
SAVC [49]12.1281.076.271.971.465.662.159.257.055.566.6656.35
ALICE [42]11.2081.171.167.864.162.460.157.956.856.064.1459.72
NC-FSCIL [59]15.8784.276.172.168.167.363.961.959.657.967.9060.93
YourSelf (Ours)11.5784.077.673.770.068.064.962.159.859.068.8062.20
+ +Training and testing protocols. For training protocols, we follow prior works [3, 20, 24, 25, 42, 49, 59, 66] to split a large proportion of the training set to serve as the base task, and split the rest into small incremental tasks. Specifically, for CIFAR-100 and miniImageNet, the first task contains 60 classes out of the total 100 classes, and the rest 40 classes are split into 8 tasks, following a 5-way 5-shot scheme. For CUB-200, there are 100 classes in the first task and the number of novel tasks is 10 (in a 10-way 5-shot setting). As for the testing, we follow prior works to report the average accuracy ( $aAcc$ ). In addition, we also provide detailed discussions on our metric generalized accuracy ( $gAcc$ ). We further provide detailed results of novel classes only in our supplementary material. + +Implementation details. For all experiments below, we use the ViT suggested by Touvron et, al. [53]. We adjust stuff like the dimensions of the backbone to match the number of parameters (around 10M) as the ResNet18 [19] which is widely used in this topic. The Feature Rectification (FR) is a lightweight module and uses popular MLPs implementation [53] with only 0.6M parameters. To ensure a fair comparison, we also re-implement other methods with the same backbone as ours, but found the performance is rarely improved. Please refer to Supplementary material for more details. We follow previous works to train from scratch on CIFAR-100 and miniImageNet dataset. Prior works [6,8,35, 52,63,73] adopt an ImageNet-pretrained ResNet for the CUB-200 dataset. To make a fair comparison, we also pretrained our backbone on the ImageNet [12] to evaluate on CUB-200. For evaluation, we reproduced 14 methods based on the public codes and conducted in-depth comparisons in the following sections. + +# 5.2 Comparison with State-of-the-art methods + +Experimental results on miniImageNet are shown in Table 1. We report the classical accuracy ( $aAcc$ ) in each section as well as its averaged results and our proposed $gAcc$ . From the table, we can conclude that our proposed method + +![](images/b03079e837c96ea84661f614d3031b6aac658f94c69deddf5e8be3102f2d926c.jpg) + +![](images/8076db5bbcad98b8b6136d43e0dc1d3cffc6a7eb1985f95e93b7a84eaef035db.jpg) + +![](images/c8680caf43746db8689de72a8a742b845a45a03f6dcf54a2f1e992f262a1b005.jpg) +Fig. 5: FSCIL results on miniImageNet. (a): The $gAcc$ curve (averaged across all $n_t$ tasks) vs. $\alpha$ . (b): The $gAcc$ AUC of each task $\mathcal{T}_i$ (Eq. (5)) vs. i. In the legend, we show the AUC value (Eq. (6)) of each method. Best viewed in color. + +![](images/d87760f4c26634ceadf17dbaa115231f4b0a64d14a49926304d0a7af4e2021e8.jpg) + +![](images/ca3f2436ef1d469c25d320d169fed211277cd1d4c018beaa09d71c415d596046.jpg) + +![](images/017231e1eb44ee990632f51a22c165f788e23cee0c4448460a125ecbed804d20.jpg) +Fig. 6: FSCIL results on CUB-200. (a): The $gAcc$ curve (averaged across all $n_t$ tasks) vs. $\alpha$ . (b): The $gAcc$ AUC of each task $\mathcal{T}_i$ (Eq. (5)) vs. $i$ . In the legend, we show the AUC value (Eq. (6)) of each method. Best viewed in color. + +outperforms other methods by a large margin. Specifically, under the classical evaluation metric $aAcc$ , the proposed method achieves 68.80% accuracy, which is 0.90%, 4.66%, 2.14% higher than SOTA methods NC-FSCIL [59], ALICE [42] and SAVC [49]. The improvement is non-trivial considering that miniImageNet is a complex dataset. + +From a more balanced perspective, we also report $gAcc$ on the Table 1 and Fig. 5. In the left figure of Fig. 5, except for the method ALICE when $\alpha = 0$ , our method outperforms most other methods by a clear margin (at least $1.27\%$ ) and achieves the highest AUC of all methods. This indicates that our method achieves a more balanced base-novel performance. To see raw numbers of the accuracy of novel classes, please refer to the supplementary material. + +Besides, Fig. 5 shows that methods that perform well at $\alpha = 1$ do not necessarily perform well when $\alpha$ is smaller. For example, although CLOM achieves $57.23\%$ average accuracy ( $\alpha = 1$ ) and is $1.83\%$ higher than method C-FSCIL ( $55.4\%$ ), we can see the performance drops significantly when $\alpha$ is small. It is mainly because CLOM focuses on the base-classes accuracy and performs poorly on novel classes. We can observe that the AUC of the CLOM is much smaller ( $4.35\%$ ) than that of C-FSCIL. The results confirm the necessity of our proposed $gAcc$ . Similar phenomena can be found when comparing other methods like SAVC and ALICE, and we can obtain similar conclusions on the other two datasets including CUB-200 and CIFAR-100. + +Results on CUB-200 dataset are shown in Table 2 and Fig. 6. We can see from Table 2 that our method is better than any other existing methods in terms of $aAcc$ and $gAcc$ (1.52%@aAcc and 2.6%@gAcc improvement over the strong method SAVC). Results on CIFAR-100 dataset are shown in Table 3. For the traditional metric $aAcc$ , our method achieves comparable performance with SOTA methods, while our method outperforms all other methods by a clear margin when considering the newly introduced metric $gAcc$ . Please refer to Sec A5 of the supplementary material for more experiment results. + +Table 2: Comparison with FSCIL methods on CUB-200. The best and second best results are marked bolded and underlined. Methods with $\dagger$ are reported with the original results in their paper. Other results are reimplemented by us. + +
MethodsNum. of Params.(M)aAcc in each task (%)
1234567891011aAccgAcc
TOPIC [52]†-68.742.554.850.045.341.438.435.432.228.326.343.92-
DF-Replay [35]†-75.972.168.663.862.659.157.855.955.053.652.461.52-
BiDistFSCIL [66]†-79.175.472.869.167.565.164.063.561.961.560.967.34-
MetaFSCIL [8]†-75.972.468.864.863.060.058.356.854.853.852.661.93-
SubNetwork [25]†-78.774.671.467.565.462.661.159.457.557.256.764.68-
DSN [58]†-76.172.269.666.764.462.160.258.957.055.154.263.31-
Self-promoted [73]12.3662.654.751.647.344.340.938.839.035.935.133.544.0740.88
CEC [63]12.3475.771.468.164.062.058.857.355.553.653.252.161.1754.16
LIMIT [70]12.3379.576.473.269.267.164.362.861.959.859.157.666.5859.40
CLOM [76]18.8979.575.972.068.265.763.462.060.358.557.956.965.6157.40
FACT [69]11.3476.173.470.466.564.862.561.159.957.957.556.550.2957.93
NC-FSCIL [59]15.9181.176.472.869.366.163.762.261.059.058.057.366.1759.22
ALICE [42]41.7175.870.168.465.263.761.260.759.758.558.758.463.7760.28
S3C [24]13.0680.276.572.969.166.964.362.761.059.559.258.366.5361.05
SAVC [49]24.2981.177.675.071.269.666.765.864.862.962.762.268.3362.90
YourSelf (Ours)11.9083.477.075.372.269.066.866.065.664.164.563.669.8565.50
+ +Table 3: Comparison results on CIFAR-100. The best and second best results are marked bolded and underlined. All results of other methods are reimplemented by us. See more results in the supplementary material. + +
MethodsNum. of Params.(M)aAcc in each task (%)aAccgAcc
123456789
CEC [63]0.29673.269.165.461.258.055.553.251.349.259.5650.85
LIMIT [70]0.29676.171.767.363.260.157.355.253.050.861.6452.16
CLOM [76]0.35574.269.866.162.259.056.154.051.949.860.3550.36
C-FSCIL [20]12.7976.872.067.263.159.756.754.451.949.761.2850.62
SAVC [49]0.84278.472.768.464.061.258.356.054.151.762.7653.29
FACT [69]0.28278.572.768.864.561.358.656.854.452.363.1254.56
SubNet [25]22.3980.075.771.667.864.761.459.357.155.164.8556.63
S3C [24]0.34178.073.970.266.163.360.158.356.754.064.5158.04
ALICE [42]45.2880.371.967.063.260.758.357.355.553.763.1059.72
NC-FSCIL [59]15.9382.977.573.769.065.662.059.757.855.467.0659.73
Ours11.6882.976.372.967.865.262.060.758.856.667.0260.32
+ +# 5.3 Ablation studies + +Effectiveness of different losses. Ablation experiments of the effectiveness of $\mathcal{L}_{IR}$ and $\mathcal{L}_{CR}$ are conducted among CIFAR-100, miniImageNet, and CUB-200 datasets. Experimental results are shown in Table 4 and Table 5. The detailed $gAcc$ vs. $\alpha$ as shown in Fig. 7. Results on CIFAR-100 show that introducing the FR module can effectively enhance the $gAcc$ by $3.28\%$ . After further introducing the class-relation (CR) transfer loss, the performance at $gAcc$ is further improved by $1.07\%$ . Adding the instance-relation (IR) transfer loss will further boost the $gAcc$ by $1.91\%$ . The effectiveness of the proposed losses is also verified on the miniImageNet dataset. For miniImageNet dataset, the introduction of FR modules brings $3.02\%$ improvement at $gAcc$ . Applying loss $\mathcal{L}_{CR}$ and $\mathcal{L}_{IR}$ can improve $gAcc$ by $0.49\%$ and $0.82\%$ , respectively. It is worth mentioning that $\mathcal{L}_{CR}$ and $\mathcal{L}_{IR}$ can also boost the $aAcc$ . Specifically, on the miniImageNet, adding these two losses results in an increase at $aAcc$ performance of $1.96\%$ . On CIFAR-100, the $aAcc$ is improved by about $0.76\%$ with the help of these two losses. See more experiments in the supplementary material. + +Effectiveness of multi-layer knowledge ensemble. In Sec. 4.2, we propose to exploit hierarchical knowledge from multiple layers for effective feature rectification. In practice, we select layers 8, 9 and 10 for the ensemble. Here, we take a more detailed discussion on each layer before the ensemble of multiple + +Table 4: Ablation studies on CIFAR-100. + +
FR\( {\mathcal{L}}_{\text{Cos }} \)\( {\mathcal{L}}_{\text{NovCE }} \)\( {\mathcal{L}}_{CR} \)\( {\mathcal{L}}_{IR} \)aAcc. in each task (%)aAccgAcc
123456789
82.977.872.968.464.661.458.756.454.166.3554.20
✓✓82.976.173.167.964.860.859.257.253.966.2657.48
✓✓✓82.976.172.467.864.561.359.657.354.966.3358.55
✓✓✓✓82.975.772.567.464.861.960.458.756.367.0260.32
+ +Table 5: Ablation studies on miniImageNet. + +
FRLcos+ LNoveCELCRLIRaAcc. in each task (%)aAccgAcc
123456789
84.079.475.071.167.864.461.358.856.968.7457.87
✓✓84.074.069.766.866.163.560.858.658.166.8460.89
✓✓✓84.077.773.669.967.464.661.659.858.768.6061.38
✓✓✓✓84.077.673.770.068.064.962.159.859.068.8062.20
+ +![](images/692d5722960dfe2f8914e6cd436a9210d4d5f02e817c8618a93feae34e9307d7.jpg) +Fig.7: Ablation studies on two datasets, evaluating using $gAcc$ . + +![](images/766a7f88dc627d7e4edf055268d6d7bd618febf9dd6f7f39e729e2290995840e.jpg) + +layers. We conducted experiments on the miniImageNet, and the results are shown in Table 6. It is clear that for each layer, the corresponding FR module effectively calibrates the features and achieves improvement from $2.23\%$ to $3.54\%$ at $gAcc$ . Furthermore, the introduction of the FR module does not harm the $aAcc$ . Considering the fact that different layers contain hierarchical knowledge, the multi-layer ensemble is proposed. The results in the table show that the ensemble is effective in improving the $gAcc$ while maintaining $aAcc$ . + +Table 6: Detailed performance of each layer. Results are from miniImagenet dataset. 'l-Raw': vanilla ViT features from layer $L_{l}$ . 'l-Recti': Rectified features using the proposed FR module. +'Ensemble': Ensembled results from $l$ -Recti', $l \in \{8, 9, 10\}$ . + +
LayersaAcc in each session (%)aAccgAcc
012345678
8-Raw78.973.269.165.462.960.057.355.254.064.0156.61
8-Recti80.773.470.166.664.461.559.357.056.565.5260.15
9-raw82.577.373.269.765.363.760.958.957.567.6659.33
9-Recti83.076.673.269.767.564.661.959.959.068.3961.90
10-raw83.878.674.470.766.464.361.559.357.768.5558.90
10-Recti84.378.274.170.468.264.861.959.558.468.8461.13
Ensemble84.077.673.770.068.064.962.159.859.068.8062.20
+ +# 6 Conclusion + +In this work, we have revisited the challenges of the FSCIL, showing the limitation of the existing metric and introducing a novel metric denoted as generalized accuracy ( $gAcc$ ) for the comprehensive evaluation of FSCIL methods from a more balanced perspective. Furthermore, we incorporate the ViT into the FSCIL framework and unlock the potential of the ViT intermediate layer features. To achieve this, we have designed a feature rectification (FR) module aimed at integrating the valuable information from the intermediate layer features into the rectified feature space. Using the new metric $gAcc$ , we analyze the existing open-sourced FSCIL methods from a more balanced perspective and show the superiority of our method on three popular datasets. + +# Acknowledgements + +This work was supported partially by the National Key Research and Development Program of China (2023YFA1008503), NSFC(U21A20471, 62206315), Guangdong NSF Project (No. 2023B1515040025, 2020B1515120085, 2024A15150-10101), Guangzhou Basic and Applied Basic Research Scheme(2024A04J4067). + +# References + +1. Agarwal, A., Banerjee, B., Cuzzolin, F., Chaudhuri, S.: Semantics-driven generative replay for few-shot class incremental learning. In: Proceedings of the 30th ACM International Conference on Multimedia (2022) +2. Ahmad, T., Dhamija, A.R., Cruz, S., Rabinowitz, R., Li, C., Jafarzadeh, M., Boult, T.E.: Few-shot class incremental learning leveraging self-supervised features. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2022) +3. Akyurek, A.F., Akyurek, E., Wijaya, D.T., Andreas, J.: Subspace regularizers for few-shot class incremental learning. In: International Conference on Learning Representations (2022) +4. Castro, F.M., Marín-Jiménez, M.J., Guil, N., Schmid, C., Alahari, K.: End-to-end incremental learning. In: Proceedings of the European Conference on Computer Vision (2018) +5. Chen, H., Wang, Y., Xu, C., Xu, C., Tao, D.: Learning student networks via feature embedding. IEEE Transactions on Neural Networks and Learning Systems (2020) +6. Chen, K., Lee, C.G.: Incremental few-shot learning via vector quantization in deep embedded space. In: International Conference on Learning Representations (2020) +7. Cheraghian, A., Rahman, S., Fang, P., Roy, S.K., Petersson, L., Harandi, M.: Semantic-aware knowledge distillation for few-shot class-incremental learning. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition (2021) +8. Chi, Z., Gu, L., Liu, H., Wang, Y., Yu, Y., Tang, J.: Meta-scil: A meta-learning approach for few-shot class incremental learning. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition (2022) +9. Cui, Y., Deng, W., Chen, H., Liu, L.: Uncertainty-aware distillation for semi-supervised few-shot class-incremental learning. IEEE Transactions on Neural Networks and Learning Systems (2023) +0. Cui, Y., Deng, W., Xu, X., Liu, Z., Liu, Z., Pietikainen, M., Liu, L.: Uncertainty-guided semi-supervised few-shot class-incremental learning with knowledge distillation. IEEE Transactions on Multimedia (2022) +1. Cui, Y., Xiong, W., Tavakolian, M., Liu, L.: Semi-supervised few-shot class-incremental learning. In: IEEE International Conference on Image Processing (ICIP) (2021) +2. Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: Imagenet: A large-scale hierarchical image database. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2009) +3. Deng, Y., Xiang, X.: Expanding hyperspherical space for few-shot class-incremental learning. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV) (2024) + +14. Dong, S., Hong, X., Tao, X., Chang, X., Wei, X., Gong, Y.: Few-shot class-incremental learning via relation knowledge distillation. In: Proceedings of the AAAI Conference on Artificial Intelligence (2021) +15. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. In: International Conference on Learning Representations (2021) +16. Douillard, A., Cord, M., Ollion, C., Robert, T., Valle, E.: Podnet: Pooled outputs distillation for small-tasks incremental learning. In: Proceedings of the European Conference on Computer Vision (2020) +17. Fu, S., Yan, J., Gao, Y., Xie, X., Zheng, W.S.: Asag: Building strong one-decoder-layer sparse detectors via adaptive sparse anchor generation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 6328-6338 (2023) +18. Gu, Z., Xu, C., Yang, J., Cui, Z.: Few-shot continual infomax learning. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (2023) +19. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2016) +20. Hersche, M., Karunaratne, G., Cherubini, G., Benini, L., Sebastian, A., Rahimi, A.: Constrained few-shot class-incremental learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2022) +21. Hou, S., Pan, X., Loy, C.C., Wang, Z., Lin, D.: Learning a unified classifier incrementally via rebalancing. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2019) +22. Ji, Z., Hou, Z., Liu, X., Pang, Y., Li, X.: Memorizing complementation network for few-shot class-incremental learning. IEEE Transactions on Image Processing (2023) +23. Jiang, W., Li, D., Hu, M., Zhai, G., Yang, X., Zhang, X.P.: Few-shot class-incremental learning with prior knowledge. arXiv preprint arXiv:2402.01201 (2024) +24. Kalla, J., Biswas, S.: S3c: Self-supervised stochastic classifiers for few-shot class-incremental learning. In: European Conference on Computer Vision (2022) +25. Kang, H., Yoon, J., Madjid, S.R.H., Hwang, S.J., Yoo, C.D.: On the soft-subnetwork for few-shot class incremental learning. In: International Conference on Learning Representations (2023) +26. Kim, D.Y., Han, D.J., Seo, J., Moon, J.: Warping the space: Weight space rotation for class-incremental few-shot learning. In: International Conference on Learning Representations (2023) +27. Kim, S., Jeong, Y., Park, J.S., Yoon, S.W.: Mics: Midpoint interpolation to learn compact and separated representations for few-shot class-incremental learning. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (2024) +28. Krizhevsky, A., Hinton, G., et al.: Learning multiple layers of features from tiny images. Technical report, University of Toronto (2009) +29. Lee, K., Lee, K., Shin, J., Lee, H.: Overcoming catastrophic forgetting with unlabeled data in the wild. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2019) +30. Li, J., Dong, S., Gong, Y., He, Y., Wei, X.: Analogical learning-based few-shot class-incremental learning. IEEE Transactions on Circuits and Systems for Video Technology (2024) + +31. Li, Y.M., Huang, W.J., Wang, A.L., Zeng, L.A., Meng, J.K., Zheng, W.S.: Egoexofitness: Towards egocentric and exocentric full-body action understanding. arXiv preprint arXiv:2406.08877 (2024) +32. Li, Y.M., Zeng, L.A., Meng, J.K., Zheng, W.S.: Continual action assessment via task-consistent score-discriminative feature distribution modeling. IEEE Transactions on Circuits and Systems for Video Technology (2024) +33. Li, Z., Hoiem, D.: Learning without forgetting. IEEE transactions on pattern analysis and machine intelligence (2017) +34. Lin, K.Y., Ding, H., Zhou, J., Peng, Y.X., Zhao, Z., Loy, C.C., Zheng, W.S.: Rethinking clip-based video learners in cross-domain open-vocabulary action recognition. arXiv preprint arXiv:2403.01560 (2024) +35. Liu, H., Gu, L., Chi, Z., Wang, Y., Yu, Y., Chen, J., Tang, J.: Few-shot class-incremental learning via entropy-regularized data-free replay. In: European Conference on Computer Vision (2022) +36. Liu, S., Qi, L., Qin, H., Shi, J., Jia, J.: Path aggregation network for instance segmentation. In: Proceedings of the IEEE conference on computer vision and pattern recognition (2018) +37. Van der Maaten, L., Hinton, G.: Visualizing data using t-sne. Journal of machine learning research (2008) +38. Mazumder, P., Singh, P., Rai, P.: Few-shot lifelong learning. In: Proceedings of the AAAI Conference on Artificial Intelligence (2021) +39. Pan, Z., Yu, X., Zhang, M., Gao, Y.: Ssfe-net: Self-supervised feature enhancement for ultra-fine-grained few-shot class incremental learning. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (2023) +40. Park, W., Kim, D., Lu, Y., Cho, M.: Relational knowledge distillation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition (2019) +41. Passalis, N., Tzelepi, M., Tefas, A.: Heterogeneous knowledge distillation using information flow modeling. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2020) +42. Peng, C., Zhao, K., Wang, T., Li, M., Lovell, B.C.: Few-shot class-incremental learning from an open-set perspective. In: European Conference on Computer Vision (2022) +43. Peng, Y.X., Jiao, J., Feng, X., Zheng, W.S.: Consistent discrepancy learning for intra-camera supervised person re-identification. IEEE Transactions on Multimedia 25, 2393–2403 (2022) +44. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021) +45. Rebuffi, S.A., Kolesnikov, A., Sperl, G., Lampert, C.H.: icarl: Incremental classifier and representation learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2017) +46. Russakovsky, O., Deng, J., Su, H., Krause, J., Satheesh, S., Ma, S., Huang, Z., Karpathy, A., Khosla, A., Bernstein, M., et al.: Imagenet large scale visual recognition challenge. International journal of computer vision (2015) +47. Shi, G., Chen, J., Zhang, W., Zhan, L.M., Wu, X.M.: Overcoming catastrophic forgetting in incremental few-shot learning by finding flat minima. In: Advances in neural information processing systems (2021) + +48. Smith, J., Hsu, Y.C., Balloch, J., Shen, Y., Jin, H., Kira, Z.: Always be dreaming: A new approach for data-free class-incremental learning. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (2021) +49. Song, Z., Zhao, Y., Shi, Y., Peng, P., Yuan, L., Tian, Y.: Learning with fantasy: Semantic-aware virtual contrastive constraint for few-shot class-incremental learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2023) +50. Tang, Y.M., Peng, Y.X., Zheng, W.S.: Learning to imagine: Diversify memory for incremental learning using unlabeled data. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2022) +51. Tang, Y.M., Peng, Y.X., Zheng, W.S.: When prompt-based incremental learning does not meet strong pretraining. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (2023) +52. Tao, X., Hong, X., Chang, X., Dong, S., Wei, X., Gong, Y.: Few-shot class incremental learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2020) +53. Touvron, H., Cord, M., Douze, M., Massa, F., Sablayrolles, A., Jégou, H.: Training data-efficient image transformers & distillation through attention. In: International Conference on Machine Learning (2021) +54. Vaze, S., Han, K., Vedaldi, A., Zisserman, A.: Open-set recognition: A good closed-set classifier is all you need? In: International Conference on Learning Representations (2021) +55. Wah, C., Branson, S., Welinder, P., Perona, P., Belongie, S.: The caltech-ucsd birds-200-2011 dataset (2011) +56. Wang, Q.W., Zhou, D.W., Zhang, Y.K., Zhan, D.C., Ye, H.J.: Few-shot class-incremental learning via training-free prototype calibration. In: Advances in Neural Information Processing Systems (2023) +57. Yan, S., Xie, J., He, X.: Der: Dynamically expandable representation for class incremental learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2021) +58. Yang, B., Lin, M., Zhang, Y., Liu, B., Liang, X., Ji, R., Ye, Q.: Dynamic support network for few-shot class incremental learning. IEEE Transactions on Pattern Analysis and Machine Intelligence (2022) +59. Yang, Y., Yuan, H., Li, X., Lin, Z., Torr, P., Tao, D.: Neural collapse inspired feature-classifier alignment for few-shot class incremental learning. In: International Conference on Learning Representations (2023) +60. Yao, G., Zhu, J., Zhou, W., Li, J.: Few-shot class-incremental learning based on representation enhancement. Journal of Electronic Imaging (2022) +61. Yu, L., Twardowski, B., Liu, X., Herranz, L., Wang, K., Cheng, Y., Jui, S., Weijer, J.v.d.: Semantic drift compensation for class-incremental learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2020) +62. Yuan, Y., Fu, R., Huang, L., Lin, W., Zhang, C., Chen, X., Wang, J.: Hrformer: High-resolution vision transformer for dense predict. Advances in Neural Information Processing Systems (2021) +63. Zhang, C., Song, N., Lin, G., Zheng, Y., Pan, P., Xu, Y.: Few-shot incremental learning with continually evolved classifiers. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition (2021) +64. Zhang, J., Zhang, J., Ghosh, S., Li, D., Tasci, S., Heck, L., Zhang, H., Kuo, C.C.J.: Class-incremental learning via deep model consolidation. In: Proceedings of the IEEE Winter Conference on Applications of Computer Vision (2020) + +65. Zhao, B., Xiao, X., Gan, G., Zhang, B., Xia, S.T.: Maintaining discrimination and fairness in class incremental learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2020) +66. Zhao, L., Lu, J., Xu, Y., Cheng, Z., Guo, D., Niu, Y., Fang, X.: Few-shot class-incremental learning via class-aware bilateral distillation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2023) +67. Zheng, G., Zhang, A.: Few-shot class-incremental learning with meta-learned class structures. In: International Conference on Data Mining Workshops (ICDMW) (2021) +68. Zheng, W.S., Yan, J., Peng, Y.X.: A versatile framework for multi-scene person re-identification. IEEE Transactions on Pattern Analysis and Machine Intelligence (2024) +69. Zhou, D.W., Wang, F.Y., Ye, H.J., Ma, L., Pu, S., Zhan, D.C.: Forward compatible few-shot class-incremental learning. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition (2022) +70. Zhou, D.W., Ye, H.J., Ma, L., Xie, D., Pu, S., Zhan, D.C.: Few-shot class-incremental learning by sampling multi-phase tasks. IEEE Transactions on Pattern Analysis and Machine Intelligence (2022) +71. Zhu, F., Zhang, X.Y., Wang, C., Yin, F., Liu, C.L.: Prototype augmentation and self-supervision for incremental learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2021) +72. Zhu, J., Yao, G., Zhou, W., Zhang, G., Ping, W., Zhang, W.: Feature distribution distillation-based few shot class incremental learning. In: the 5th International Conference on Pattern Recognition and Artificial Intelligence (PRAI) (2022) +73. Zhu, K., Cao, Y., Zhai, W., Cheng, J., Zha, Z.J.: Self-promoted prototype refinement for few-shot class-incremental learning. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition (2021) +74. Zhu, K., Zhai, W., Cao, Y., Luo, J., Zha, Z.J.: Self-sustaining representation expansion for non-exemplar class-incremental learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2022) +75. Zhu, M., Zhu, Z., Chen, S., Chen, C., Wu, B.: Enhanced few-shot class-incremental learning via ensemble models. arXiv preprint arXiv:2401.07208 (2024) +76. Zou, Y., Zhang, S., Li, Y., Li, R.: Margin-based few-shot class-incremental learning with class-level overfitting mitigation. In: Advances in neural information processing systems (2022) \ No newline at end of file diff --git a/2024/Rethinking Few-shot Class-incremental Learning_ Learning from Yourself/images.zip b/2024/Rethinking Few-shot Class-incremental Learning_ Learning from Yourself/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..316463efda9bd7dc74f35dd9a8b58d9017064331 --- /dev/null +++ b/2024/Rethinking Few-shot Class-incremental Learning_ Learning from Yourself/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7dbfbf07fc2190a58369970f90a4e065451e44bce14bd875d6f001c13610edff +size 683023 diff --git a/2024/Rethinking Few-shot Class-incremental Learning_ Learning from Yourself/layout.json b/2024/Rethinking Few-shot Class-incremental Learning_ Learning from Yourself/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..34d7368fb31cc6bd45e40c74a9c1e1389ea120fe --- /dev/null +++ b/2024/Rethinking Few-shot Class-incremental Learning_ Learning from Yourself/layout.json @@ -0,0 +1,14386 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 135, + 111, + 479, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 111, + 479, + 148 + ], + "spans": [ + { + "bbox": [ + 135, + 111, + 479, + 148 + ], + "type": "text", + "content": "Rethinking Few-shot Class-incremental Learning: Learning from Yourself" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 148, + 167, + 465, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 148, + 167, + 465, + 194 + ], + "spans": [ + { + "bbox": [ + 148, + 167, + 465, + 194 + ], + "type": "text", + "content": "Yu-Ming Tang" + }, + { + "bbox": [ + 148, + 167, + 465, + 194 + ], + "type": "inline_equation", + "content": "^{1,3}" + }, + { + "bbox": [ + 148, + 167, + 465, + 194 + ], + "type": "text", + "content": ", Yi-Xing Peng" + }, + { + "bbox": [ + 148, + 167, + 465, + 194 + ], + "type": "inline_equation", + "content": "^{1,3,5}" + }, + { + "bbox": [ + 148, + 167, + 465, + 194 + ], + "type": "text", + "content": ", Jingke Meng" + }, + { + "bbox": [ + 148, + 167, + 465, + 194 + ], + "type": "inline_equation", + "content": "^{\\star 1,3}" + }, + { + "bbox": [ + 148, + 167, + 465, + 194 + ], + "type": "text", + "content": ", and Wei-Shi Zheng" + }, + { + "bbox": [ + 148, + 167, + 465, + 194 + ], + "type": "inline_equation", + "content": "^{1,2,3,4,5}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 145, + 201, + 465, + 223 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 201, + 465, + 223 + ], + "spans": [ + { + "bbox": [ + 145, + 201, + 465, + 223 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 145, + 201, + 465, + 223 + ], + "type": "text", + "content": " School of Computer Science and Engineering, Sun Yat-sen University, China \n" + }, + { + "bbox": [ + 145, + 201, + 465, + 223 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 145, + 201, + 465, + 223 + ], + "type": "text", + "content": " Peng Cheng Laboratory, Shenzhen, China" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 141, + 224, + 472, + 245 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 224, + 472, + 245 + ], + "spans": [ + { + "bbox": [ + 141, + 224, + 472, + 245 + ], + "type": "text", + "content": "3 Key Laboratory of Machine Intelligence and Advanced Computing, Ministry of Education, China" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 141, + 245, + 470, + 267 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 245, + 470, + 267 + ], + "spans": [ + { + "bbox": [ + 141, + 245, + 470, + 267 + ], + "type": "text", + "content": "4 Guangdong Province Key Laboratory of Information Security Technology, Sun Yat-sen University, China" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 156, + 268, + 455, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 156, + 268, + 455, + 300 + ], + "spans": [ + { + "bbox": [ + 156, + 268, + 455, + 300 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 156, + 268, + 455, + 300 + ], + "type": "text", + "content": " Pazhou Laboratory (Huangpu), Guangzhou, Guangdong 510555, China {tangym9, pengyx23}@mail2.sysu.edu.cn, mengjke@gmail.com, wszheng@ieee.org" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 159, + 331, + 455, + 550 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 331, + 455, + 550 + ], + "spans": [ + { + "bbox": [ + 159, + 331, + 455, + 550 + ], + "type": "text", + "content": "Abstract. Few-shot class-incremental learning (FSCIL) aims to learn sequential classes with limited samples in a few-shot fashion. Inherited from the classical class-incremental learning setting, the popular benchmark of FSCIL uses averaged accuracy (" + }, + { + "bbox": [ + 159, + 331, + 455, + 550 + ], + "type": "inline_equation", + "content": "aAcc" + }, + { + "bbox": [ + 159, + 331, + 455, + 550 + ], + "type": "text", + "content": ") and last-task averaged accuracy (" + }, + { + "bbox": [ + 159, + 331, + 455, + 550 + ], + "type": "inline_equation", + "content": "lAcc" + }, + { + "bbox": [ + 159, + 331, + 455, + 550 + ], + "type": "text", + "content": ") as the evaluation metrics. However, we reveal that such evaluation metrics may not provide adequate emphasis on the novel class performance, and the continual learning ability of FSCIL methods could be ignored under this benchmark. In this work, as a complement to existing metrics, we offer a new metric called generalized average accuracy (" + }, + { + "bbox": [ + 159, + 331, + 455, + 550 + ], + "type": "inline_equation", + "content": "gAcc" + }, + { + "bbox": [ + 159, + 331, + 455, + 550 + ], + "type": "text", + "content": ") which is designed to provide an extra equitable evaluation by incorporating different perspectives of the performance under the guidance of a parameter " + }, + { + "bbox": [ + 159, + 331, + 455, + 550 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 159, + 331, + 455, + 550 + ], + "type": "text", + "content": ". We also present an overall metric in the form of the area under the curve (AUC) along the " + }, + { + "bbox": [ + 159, + 331, + 455, + 550 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 159, + 331, + 455, + 550 + ], + "type": "text", + "content": ". Under the guidance of " + }, + { + "bbox": [ + 159, + 331, + 455, + 550 + ], + "type": "inline_equation", + "content": "gAcc" + }, + { + "bbox": [ + 159, + 331, + 455, + 550 + ], + "type": "text", + "content": ", we release the potential of intermediate features of the vision transformers to boost the novel-class performance. Taking information from intermediate layers which are less class-specific and more generalizable, we manage to rectify the final features, leading to a more generalizable transformer-based FSCIL framework. Without complex network designs or cumbersome training procedures, our method outperforms existing FSCIL methods at " + }, + { + "bbox": [ + 159, + 331, + 455, + 550 + ], + "type": "inline_equation", + "content": "aAcc" + }, + { + "bbox": [ + 159, + 331, + 455, + 550 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 159, + 331, + 455, + 550 + ], + "type": "inline_equation", + "content": "gAcc" + }, + { + "bbox": [ + 159, + 331, + 455, + 550 + ], + "type": "text", + "content": " on three datasets. See codes at 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 159, + 560, + 453, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 560, + 453, + 582 + ], + "spans": [ + { + "bbox": [ + 159, + 560, + 453, + 582 + ], + "type": "text", + "content": "Keywords: Continual Learning " + }, + { + "bbox": [ + 159, + 560, + 453, + 582 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 560, + 453, + 582 + ], + "type": "text", + "content": " Few-shot Learning " + }, + { + "bbox": [ + 159, + 560, + 453, + 582 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 560, + 453, + 582 + ], + "type": "text", + "content": " Catastrophic Forgetting " + }, + { + "bbox": [ + 159, + 560, + 453, + 582 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 560, + 453, + 582 + ], + "type": "text", + "content": " Vision Transformer" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 591, + 230, + 604 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 591, + 230, + 604 + ], + "spans": [ + { + "bbox": [ + 132, + 591, + 230, + 604 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 611, + 482, + 636 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 611, + 482, + 636 + ], + "spans": [ + { + "bbox": [ + 130, + 611, + 482, + 636 + ], + "type": "text", + "content": "Although existing deep neural networks (DNNs) achieve great success in various real-world applications [17, 31, 32, 34, 43, 44, 54, 68], they suffer from catastrophic" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 642, + 236, + 654 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 642, + 236, + 654 + ], + "spans": [ + { + "bbox": [ + 133, + 642, + 236, + 654 + ], + "type": "text", + "content": "* Corresponding Author" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 135, + 654, + 381, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 654, + 381, + 665 + ], + "spans": [ + { + "bbox": [ + 135, + 654, + 381, + 665 + ], + "type": "text", + "content": "1 https://github.com/iSEE-Laboratory/Revisiting_FSCIL" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 151, + 133, + 304, + 227 + ], + "blocks": [ + { + "bbox": [ + 184, + 120, + 284, + 129 + ], + "lines": [ + { + "bbox": [ + 184, + 120, + 284, + 129 + ], + "spans": [ + { + "bbox": [ + 184, + 120, + 284, + 129 + ], + "type": "text", + "content": "(a) Performance after each task" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 151, + 133, + 304, + 227 + ], + "lines": [ + { + "bbox": [ + 151, + 133, + 304, + 227 + ], + "spans": [ + { + "bbox": [ + 151, + 133, + 304, + 227 + ], + "type": "image", + "image_path": "11401f85782ee84c427d469061887ee428cbb72141ea6d2507449d58dbad8c37.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 312, + 134, + 463, + 227 + ], + "blocks": [ + { + "bbox": [ + 348, + 114, + 448, + 133 + ], + "lines": [ + { + "bbox": [ + 348, + 114, + 448, + 133 + ], + "spans": [ + { + "bbox": [ + 348, + 114, + 448, + 133 + ], + "type": "text", + "content": "(b) Detailed Performance after finishing the last task (Task 9)" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 312, + 134, + 463, + 227 + ], + "lines": [ + { + "bbox": [ + 312, + 134, + 463, + 227 + ], + "spans": [ + { + "bbox": [ + 312, + 134, + 463, + 227 + ], + "type": "image", + "image_path": "ecb69e171671f49bdc7f0d5182231b32f89132f5e403e6781a9b76eacd65e3e0.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 235, + 482, + 293 + ], + "lines": [ + { + "bbox": [ + 130, + 235, + 482, + 293 + ], + "spans": [ + { + "bbox": [ + 130, + 235, + 482, + 293 + ], + "type": "text", + "content": "Fig. 1: Performance under different metrics of two recent methods [24, 25]. (a): The dotted line denotes the average accuracy (" + }, + { + "bbox": [ + 130, + 235, + 482, + 293 + ], + "type": "inline_equation", + "content": "aAcc" + }, + { + "bbox": [ + 130, + 235, + 482, + 293 + ], + "type": "text", + "content": ") widely used in classical class-incremental learning. The solid line presents our proposed generalized accuracy (" + }, + { + "bbox": [ + 130, + 235, + 482, + 293 + ], + "type": "inline_equation", + "content": "gAcc" + }, + { + "bbox": [ + 130, + 235, + 482, + 293 + ], + "type": "text", + "content": "). We also show the average of each point on the right side. (Blue: " + }, + { + "bbox": [ + 130, + 235, + 482, + 293 + ], + "type": "inline_equation", + "content": "gAcc" + }, + { + "bbox": [ + 130, + 235, + 482, + 293 + ], + "type": "text", + "content": ", Red: " + }, + { + "bbox": [ + 130, + 235, + 482, + 293 + ], + "type": "inline_equation", + "content": "aAcc" + }, + { + "bbox": [ + 130, + 235, + 482, + 293 + ], + "type": "text", + "content": "). (b): The detailed accuracies of each task after training on the last task (task 8). It is evident that S3C exhibits superior performance when confronted with novel classes. The conventional metric " + }, + { + "bbox": [ + 130, + 235, + 482, + 293 + ], + "type": "inline_equation", + "content": "aAcc" + }, + { + "bbox": [ + 130, + 235, + 482, + 293 + ], + "type": "text", + "content": " fails to reflect this due to the domination by the base-class performance. Models are trained on the CIFAR-100 dataset. Best viewed in color." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 311, + 482, + 419 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 311, + 482, + 419 + ], + "spans": [ + { + "bbox": [ + 130, + 311, + 482, + 419 + ], + "type": "text", + "content": "forgetting when dealing with sequential data streams. The forgetting problem has attracted heavy attention in recent years [4,16,21,29,33,45,48,50,51,57,61,64,65, 71,74]. In this work, we focus on a more realistic and challenging problem: few-shot class-incremental learning (FSCIL), in which after training on a task with sufficient data (base task), only a few samples are provided during incremental tasks (novel tasks). Since Tao et al. [52] proposed the FSCIL benchmark in 2020, a large number of algorithms have emerged to tackle this problem, including meta-learning based [8, 20, 70], structure based [2, 25, 52, 58, 63], feature space based [3, 6, 20, 24, 26, 59, 69] methods, etc." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 426, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 426, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 426, + 482, + 666 + ], + "type": "text", + "content": "However, almost all these methods are evaluated following the traditional class-incremental learning setting, i.e. using the average accuracy (" + }, + { + "bbox": [ + 130, + 426, + 482, + 666 + ], + "type": "inline_equation", + "content": "aAcc" + }, + { + "bbox": [ + 130, + 426, + 482, + 666 + ], + "type": "text", + "content": ") or the last-task averaged accuracy (" + }, + { + "bbox": [ + 130, + 426, + 482, + 666 + ], + "type": "inline_equation", + "content": "lAcc" + }, + { + "bbox": [ + 130, + 426, + 482, + 666 + ], + "type": "text", + "content": ") as the main metrics to critique the proposed method. Such accuracies are calculated by weighted averaging the base-task and novel-tasks accuracies with the weight proportional to the number of classes of each task. Considering the reality that the FSCIL benchmarks make the majority of the classes (about " + }, + { + "bbox": [ + 130, + 426, + 482, + 666 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 130, + 426, + 482, + 666 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 130, + 426, + 482, + 666 + ], + "type": "inline_equation", + "content": "60\\%" + }, + { + "bbox": [ + 130, + 426, + 482, + 666 + ], + "type": "text", + "content": ") as the base task to establish a well-trained backbone for few-shot continual learning, the existing " + }, + { + "bbox": [ + 130, + 426, + 482, + 666 + ], + "type": "inline_equation", + "content": "aAcc" + }, + { + "bbox": [ + 130, + 426, + 482, + 666 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 130, + 426, + 482, + 666 + ], + "type": "inline_equation", + "content": "lAcc" + }, + { + "bbox": [ + 130, + 426, + 482, + 666 + ], + "type": "text", + "content": " metric is dominated by the base-class performance. In other words, current evaluation metrics can be largely boosted by only improving the performance of base classes. As shown in Fig. 1, a method like softNet [25] with higher base-class performance (see Fig. 1 (b)) usually gets higher " + }, + { + "bbox": [ + 130, + 426, + 482, + 666 + ], + "type": "inline_equation", + "content": "aAcc" + }, + { + "bbox": [ + 130, + 426, + 482, + 666 + ], + "type": "text", + "content": " (see red lines in Fig. 1 (a)) regardless of the inferior performance on the novel classes. Due to the inadequacy to capture performance changes in novel classes, current metrics are biased toward the performance of base classes and have limitations in guiding the FSCIL community. To tackle this, we propose a more balanced metric called generalized accuracy (" + }, + { + "bbox": [ + 130, + 426, + 482, + 666 + ], + "type": "inline_equation", + "content": "gAcc" + }, + { + "bbox": [ + 130, + 426, + 482, + 666 + ], + "type": "text", + "content": "). Parameterized by " + }, + { + "bbox": [ + 130, + 426, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 130, + 426, + 482, + 666 + ], + "type": "text", + "content": " ranging from 0 to 1, the " + }, + { + "bbox": [ + 130, + 426, + 482, + 666 + ], + "type": "inline_equation", + "content": "gAcc" + }, + { + "bbox": [ + 130, + 426, + 482, + 666 + ], + "type": "text", + "content": " offers a way to evaluate a model with explicit emphasis on performance on novel classes. Furthermore, we can comprehensively evaluate the model by calculating the area under the curve (AUC) along the parameter " + }, + { + "bbox": [ + 130, + 426, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 130, + 426, + 482, + 666 + ], + "type": "text", + "content": ". It is worth mentioning that our intention" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 212, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 212, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 212, + 102 + ], + "type": "text", + "content": "Tang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": "is NOT to replace all previous metrics, but to provide a more fine-grained and balanced evaluation of FSCIL methods complementary to them." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 140, + 481, + 212 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 140, + 481, + 212 + ], + "spans": [ + { + "bbox": [ + 130, + 140, + 481, + 212 + ], + "type": "text", + "content": "We empirically find that FSCIL can benefit from the powerful vision transformer (ViT) [15] by analyzing the performance of novel classes. Specifically, although vanilla ViT performs poorly in terms of unseen novel classes, we reveal that the intermediate features of ViT are less specified on base classes. It means the ViT can learn abundant semantic information but the information is mistakenly dropped in deep layers when extracting features for novel classes." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 212, + 482, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 212, + 482, + 319 + ], + "spans": [ + { + "bbox": [ + 130, + 212, + 482, + 319 + ], + "type": "text", + "content": "Based on the observation, we propose a lightweight Feature Rectification (FR) module to rectify the output features of the ViT conditioned on the intermediate features. To achieve this, the FR module is supervised with two losses that work on two different levels, namely instance-level relation transfer and class-center relation transfer. The former transfers the relations between instances while the latter focuses on relations between instances and the class prototypes. We further enhance the FR module by leveraging hierarchical knowledge from multiple intermediate layers. Finally, our method boosts the novel-class accuracy and maintains a comparable performance of base classes." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 319, + 482, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 319, + 482, + 475 + ], + "spans": [ + { + "bbox": [ + 130, + 319, + 482, + 475 + ], + "type": "text", + "content": "Our contributions are summarized as follows. (1) We revisit the existing FSCIL benchmark and propose a novel metric called generalized accuracy (" + }, + { + "bbox": [ + 130, + 319, + 482, + 475 + ], + "type": "inline_equation", + "content": "gAcc" + }, + { + "bbox": [ + 130, + 319, + 482, + 475 + ], + "type": "text", + "content": ") for a more fine-grained and balanced evaluation. We have implemented 14 FSCIL methods with codes released and show that the newly introduced SOTAs do not necessarily perform better on " + }, + { + "bbox": [ + 130, + 319, + 482, + 475 + ], + "type": "inline_equation", + "content": "gAcc" + }, + { + "bbox": [ + 130, + 319, + 482, + 475 + ], + "type": "text", + "content": ". (2) Based on " + }, + { + "bbox": [ + 130, + 319, + 482, + 475 + ], + "type": "inline_equation", + "content": "gAcc" + }, + { + "bbox": [ + 130, + 319, + 482, + 475 + ], + "type": "text", + "content": ", we reveal the potential of vision transformers in FSCIL. We empirically find that although features from the intermediate layer exhibit sub-optimal performance for the base classes, their transferrable ability helps them to have a distinct advantage in handling novel classes. (3) A lightweight FR module is proposed to rectify the final features to be more transferrable. The FR module manages to make the output feature more generalizable with assistance from intermediate features. (4) We conduct extensive experiments on 3 datasets and evaluate our method under both " + }, + { + "bbox": [ + 130, + 319, + 482, + 475 + ], + "type": "inline_equation", + "content": "aAcc" + }, + { + "bbox": [ + 130, + 319, + 482, + 475 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 319, + 482, + 475 + ], + "type": "inline_equation", + "content": "gAcc" + }, + { + "bbox": [ + 130, + 319, + 482, + 475 + ], + "type": "text", + "content": ", showing the effectiveness and superiority of our method." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 483, + 237, + 495 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 483, + 237, + 495 + ], + "spans": [ + { + "bbox": [ + 131, + 483, + 237, + 495 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 498, + 481, + 570 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 498, + 481, + 570 + ], + "spans": [ + { + "bbox": [ + 130, + 498, + 481, + 570 + ], + "type": "text", + "content": "Meta-learning based methods. Due to the limited samples of novel classes, some works [8,20,67,70] propose to use meta-learning to improve the transferability of the model. Concretely, MetaFscil [8] proposed to simulate novel tasks by sampling from the base task. C-FSCIL [20] uses meta-learning to train the backbone, which maps images from different classes to quasi-orthogonal vectors. ALFSCIL [30] formulate the process into a bi-level optimization problem." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": "Optimization-based methods. There is a large group of works focusing on the optimization of the DNNs in FSCIL. CLOM [76] considers the behavior of the backbone under different margins of the cosine loss. NC-FSCIL [59] designs a pre-defined classifier for efficient FSCIL without any explicit regularizer. SPPR [73] proposes a prototype refinement strategy for stronger expression ability. FACT [69] assigns virtual prototypes in the current stage for future updates. Alice [42] instead focuses on the compactness and the variety of the feature space. Based on the classifier learning, S3C [24] proposes a self-supervised stochastic" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 150, + 91, + 449, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 150, + 91, + 449, + 103 + ], + "spans": [ + { + "bbox": [ + 150, + 91, + 449, + 103 + ], + "type": "text", + "content": "Rethinking Few-shot Class-incremental Learning: Learning from Yourself" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "content": "classifier to tackle the overfitting and forgetting problem. There are more works working on the feature space constraint [3, 6, 13, 26, 27, 56, 60], or knowledge distillation [7, 9-11, 14, 39, 66, 72]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 152, + 480, + 211 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 152, + 480, + 211 + ], + "spans": [ + { + "bbox": [ + 130, + 152, + 480, + 211 + ], + "type": "text", + "content": "Data-driven methods. There is also a branch of work focusing on the data side of this problem, which focuses on the data replay [1,35]. FSIL-GAN [1] employs a GAN to generate features to compensate for the scarcity of data. DF-replay [35] uses synthesized data generated by a generator to replay. PL-FSCIL [23] seeks prior knowledge from unlabeled data." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 212, + 481, + 259 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 212, + 481, + 259 + ], + "spans": [ + { + "bbox": [ + 130, + 212, + 481, + 259 + ], + "type": "text", + "content": "Structure-based methods. In addition to these works, a unique branch of works focuses on the network structure, like topology preserving [52], growing classifier [63], proposing a dynamic structure of the backbone [58], structure fusion [2, 22], sub-networks [25] and ensemble models [75]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 259, + 480, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 259, + 480, + 331 + ], + "spans": [ + { + "bbox": [ + 130, + 259, + 480, + 331 + ], + "type": "text", + "content": "Compared with existing works, our work rephrases the challenge of the FSCIL setting and first reveals the potential of the intermediate features of the ViT backbone. There is also an existing work [42] offering an alternative metric. They did not provide a fine-grained evaluation with various perspectives and have potential risks in stability and symmetry problems. Please find more detailed discussions about related works in the supplementary material." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 335, + 334, + 347 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 335, + 334, + 347 + ], + "spans": [ + { + "bbox": [ + 131, + 335, + 334, + 347 + ], + "type": "text", + "content": "3 Generalized Average Accuracy" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 347, + 334, + 360 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 347, + 334, + 360 + ], + "spans": [ + { + "bbox": [ + 131, + 347, + 334, + 360 + ], + "type": "text", + "content": "3.1 Limitation of the Existing Metrics" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 365, + 482, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 365, + 482, + 475 + ], + "spans": [ + { + "bbox": [ + 131, + 365, + 482, + 475 + ], + "type": "text", + "content": "Accuracy metric on incremental tasks. Consider a classification model " + }, + { + "bbox": [ + 131, + 365, + 482, + 475 + ], + "type": "inline_equation", + "content": "f(\\cdot)" + }, + { + "bbox": [ + 131, + 365, + 482, + 475 + ], + "type": "text", + "content": " that takes an image as input and outputs the class predictions. For an incremental problem, there are multiple tasks and multiple test sets. Let us define test set at task " + }, + { + "bbox": [ + 131, + 365, + 482, + 475 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_i" + }, + { + "bbox": [ + 131, + 365, + 482, + 475 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 131, + 365, + 482, + 475 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{test}^i = \\{x_t^i,y_t^i\\}_{t = 1}^n" + }, + { + "bbox": [ + 131, + 365, + 482, + 475 + ], + "type": "text", + "content": ". Usually, the labeling space " + }, + { + "bbox": [ + 131, + 365, + 482, + 475 + ], + "type": "inline_equation", + "content": "\\mathcal{Y}^i\\triangleq \\{y_t^i |(x_t^i,y_t^i)\\in \\mathcal{D}_{test}^i\\}" + }, + { + "bbox": [ + 131, + 365, + 482, + 475 + ], + "type": "text", + "content": " of every task does not have any intersection, i.e. " + }, + { + "bbox": [ + 131, + 365, + 482, + 475 + ], + "type": "inline_equation", + "content": "\\forall i\\neq j,\\mathcal{Y}^i\\bigcap \\mathcal{Y}^j = \\emptyset" + }, + { + "bbox": [ + 131, + 365, + 482, + 475 + ], + "type": "text", + "content": ". After trained on " + }, + { + "bbox": [ + 131, + 365, + 482, + 475 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_i" + }, + { + "bbox": [ + 131, + 365, + 482, + 475 + ], + "type": "text", + "content": ", the model can be denoted as " + }, + { + "bbox": [ + 131, + 365, + 482, + 475 + ], + "type": "inline_equation", + "content": "f_{i}(\\cdot)" + }, + { + "bbox": [ + 131, + 365, + 482, + 475 + ], + "type": "text", + "content": ". The corresponding accuracy of model " + }, + { + "bbox": [ + 131, + 365, + 482, + 475 + ], + "type": "inline_equation", + "content": "f_{i}(\\cdot)" + }, + { + "bbox": [ + 131, + 365, + 482, + 475 + ], + "type": "text", + "content": " at task " + }, + { + "bbox": [ + 131, + 365, + 482, + 475 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_j" + }, + { + "bbox": [ + 131, + 365, + 482, + 475 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 131, + 365, + 482, + 475 + ], + "type": "inline_equation", + "content": "j\\leq i" + }, + { + "bbox": [ + 131, + 365, + 482, + 475 + ], + "type": "text", + "content": ") can be denoted as " + }, + { + "bbox": [ + 131, + 365, + 482, + 475 + ], + "type": "inline_equation", + "content": "\\mathcal{A}_i^j\\triangleq \\frac{1}{n^j}\\sum_{t = 1}^{n^j}\\mathbb{1}(\\zeta (f_i(x_t^j)) = y_t^j)" + }, + { + "bbox": [ + 131, + 365, + 482, + 475 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 131, + 365, + 482, + 475 + ], + "type": "inline_equation", + "content": "\\mathbb{1}(\\cdot) = 1" + }, + { + "bbox": [ + 131, + 365, + 482, + 475 + ], + "type": "text", + "content": " if the condition is true, otherwise 0. " + }, + { + "bbox": [ + 131, + 365, + 482, + 475 + ], + "type": "inline_equation", + "content": "\\zeta" + }, + { + "bbox": [ + 131, + 365, + 482, + 475 + ], + "type": "text", + "content": " is the argmax operation." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 476, + 482, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 476, + 482, + 548 + ], + "spans": [ + { + "bbox": [ + 131, + 476, + 482, + 548 + ], + "type": "text", + "content": "Average accuracy. It is worth mentioning that for class-incremental learning, evaluating the model on all the previous tasks is essential. To test the model on previous tasks, an overall test set that contains all previous test sets is constructed as " + }, + { + "bbox": [ + 131, + 476, + 482, + 548 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{test}^{1\\rightarrow i} = \\bigcup_{m = 1}^{i}\\mathcal{D}_{test}^{m}" + }, + { + "bbox": [ + 131, + 476, + 482, + 548 + ], + "type": "text", + "content": ". After trained on task " + }, + { + "bbox": [ + 131, + 476, + 482, + 548 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_i" + }, + { + "bbox": [ + 131, + 476, + 482, + 548 + ], + "type": "text", + "content": ", the accuracy of model " + }, + { + "bbox": [ + 131, + 476, + 482, + 548 + ], + "type": "inline_equation", + "content": "f_{i}(\\cdot)" + }, + { + "bbox": [ + 131, + 476, + 482, + 548 + ], + "type": "text", + "content": " on the overall test set " + }, + { + "bbox": [ + 131, + 476, + 482, + 548 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{test}^{1\\rightarrow i}" + }, + { + "bbox": [ + 131, + 476, + 482, + 548 + ], + "type": "text", + "content": " can be presented as " + }, + { + "bbox": [ + 131, + 476, + 482, + 548 + ], + "type": "inline_equation", + "content": "A_{i}^{1\\rightarrow i}" + }, + { + "bbox": [ + 131, + 476, + 482, + 548 + ], + "type": "text", + "content": ". Once a " + }, + { + "bbox": [ + 131, + 476, + 482, + 548 + ], + "type": "inline_equation", + "content": "n_t" + }, + { + "bbox": [ + 131, + 476, + 482, + 548 + ], + "type": "text", + "content": "-task incremental training ends, the average accuracy (" + }, + { + "bbox": [ + 131, + 476, + 482, + 548 + ], + "type": "inline_equation", + "content": "aAcc" + }, + { + "bbox": [ + 131, + 476, + 482, + 548 + ], + "type": "text", + "content": ") can be calculated as:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 228, + 555, + 481, + 585 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 228, + 555, + 481, + 585 + ], + "spans": [ + { + "bbox": [ + 228, + 555, + 481, + 585 + ], + "type": "interline_equation", + "content": "a A c c = \\frac {1}{n _ {t}} \\sum_ {i = 1} ^ {n _ {t}} a A c c _ {i} = \\frac {1}{n _ {t}} \\sum_ {i = 1} ^ {n _ {t}} A _ {i} ^ {1 \\rightarrow i}, \\tag {1}", + "image_path": "aa81a0c070873256771338dcd7cac108c90ac13ebae7ed40bd7377c0e0c77dbb.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 131, + 590, + 482, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 590, + 482, + 628 + ], + "spans": [ + { + "bbox": [ + 131, + 590, + 482, + 628 + ], + "type": "text", + "content": "where the " + }, + { + "bbox": [ + 131, + 590, + 482, + 628 + ], + "type": "inline_equation", + "content": "aAcc_{i}" + }, + { + "bbox": [ + 131, + 590, + 482, + 628 + ], + "type": "text", + "content": " is the average accuracy at each task. We can also derive two additional metrics incidentally: last-task averaged accuracy " + }, + { + "bbox": [ + 131, + 590, + 482, + 628 + ], + "type": "inline_equation", + "content": "lAcc = aAcc_{n_t} = A_{n_t}^{1\\rightarrow n_t}" + }, + { + "bbox": [ + 131, + 590, + 482, + 628 + ], + "type": "text", + "content": ", and task-wise average accuracy " + }, + { + "bbox": [ + 131, + 590, + 482, + 628 + ], + "type": "inline_equation", + "content": "(tAcc)" + }, + { + "bbox": [ + 131, + 590, + 482, + 628 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 222, + 635, + 481, + 667 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 222, + 635, + 481, + 667 + ], + "spans": [ + { + "bbox": [ + 222, + 635, + 481, + 667 + ], + "type": "interline_equation", + "content": "t A c c = \\frac {1}{n _ {t}} \\sum_ {i = 1} ^ {n _ {t}} t A c c _ {i} = \\frac {1}{n _ {t}} \\sum_ {i = 1} ^ {n _ {t}} \\frac {1}{i} \\sum_ {j = 1} ^ {i} A _ {i} ^ {j}, \\tag {2}", + "image_path": "01861ed325ad979a43d840b9d980aa12c3afc51bad9c2c6f076a988b5bedbf38.jpg" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 212, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 212, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 212, + 102 + ], + "type": "text", + "content": "Tang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 151 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 151 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 151 + ], + "type": "text", + "content": "where the " + }, + { + "bbox": [ + 130, + 116, + 479, + 151 + ], + "type": "inline_equation", + "content": "tAcc_i" + }, + { + "bbox": [ + 130, + 116, + 479, + 151 + ], + "type": "text", + "content": " is the task-wise accuracy at each task. Note that the " + }, + { + "bbox": [ + 130, + 116, + 479, + 151 + ], + "type": "inline_equation", + "content": "aAcc" + }, + { + "bbox": [ + 130, + 116, + 479, + 151 + ], + "type": "text", + "content": " does not share the same value as " + }, + { + "bbox": [ + 130, + 116, + 479, + 151 + ], + "type": "inline_equation", + "content": "tAcc" + }, + { + "bbox": [ + 130, + 116, + 479, + 151 + ], + "type": "text", + "content": " normally since the number of classes of tasks is usually not the same. This will be elaborated on later in detail." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 152, + 482, + 260 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 152, + 482, + 260 + ], + "spans": [ + { + "bbox": [ + 130, + 152, + 482, + 260 + ], + "type": "text", + "content": "Limitation. Since FSCIL usually sets the majority of the classes as the base task, existing metrics potentially neglect the performance changes of novel classes. To elaborate on this, we analyze the " + }, + { + "bbox": [ + 130, + 152, + 482, + 260 + ], + "type": "inline_equation", + "content": "aAcc" + }, + { + "bbox": [ + 130, + 152, + 482, + 260 + ], + "type": "text", + "content": " as follows. Firstly, we denote the size of the labeling space of task " + }, + { + "bbox": [ + 130, + 152, + 482, + 260 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_i" + }, + { + "bbox": [ + 130, + 152, + 482, + 260 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 130, + 152, + 482, + 260 + ], + "type": "inline_equation", + "content": "|\\mathcal{V}_i|" + }, + { + "bbox": [ + 130, + 152, + 482, + 260 + ], + "type": "text", + "content": ". In FSCIL, the number of classes of each novel task is usually the same, i.e. " + }, + { + "bbox": [ + 130, + 152, + 482, + 260 + ], + "type": "inline_equation", + "content": "|\\mathcal{V}_{novel}| = |\\mathcal{V}_2| = |\\mathcal{V}_3| = \\dots = |\\mathcal{V}_{n_t}| \\ll |\\mathcal{V}_1|" + }, + { + "bbox": [ + 130, + 152, + 482, + 260 + ], + "type": "text", + "content": ". Assuming there are total " + }, + { + "bbox": [ + 130, + 152, + 482, + 260 + ], + "type": "inline_equation", + "content": "n_t" + }, + { + "bbox": [ + 130, + 152, + 482, + 260 + ], + "type": "text", + "content": " tasks, and task " + }, + { + "bbox": [ + 130, + 152, + 482, + 260 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_1" + }, + { + "bbox": [ + 130, + 152, + 482, + 260 + ], + "type": "text", + "content": " is the base task with sufficient classes and samples (non-few-shot). The following tasks " + }, + { + "bbox": [ + 130, + 152, + 482, + 260 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{T}_i\\}_{i=2}^{n_t}" + }, + { + "bbox": [ + 130, + 152, + 482, + 260 + ], + "type": "text", + "content": " are novel tasks, with " + }, + { + "bbox": [ + 130, + 152, + 482, + 260 + ], + "type": "inline_equation", + "content": "|\\mathcal{V}_{novel}|" + }, + { + "bbox": [ + 130, + 152, + 482, + 260 + ], + "type": "text", + "content": " classes for each task and a few samples per class. Now, we can rewrite average accuracy on task " + }, + { + "bbox": [ + 130, + 152, + 482, + 260 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_i" + }, + { + "bbox": [ + 130, + 152, + 482, + 260 + ], + "type": "text", + "content": " as:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 175, + 264, + 481, + 299 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 175, + 264, + 481, + 299 + ], + "spans": [ + { + "bbox": [ + 175, + 264, + 481, + 299 + ], + "type": "interline_equation", + "content": "a A c c _ {i} = \\frac {\\left| \\mathcal {Y} _ {1} \\right| A _ {i} ^ {1} + \\left| \\mathcal {Y} _ {n o v e l} \\right| \\sum_ {j = 2} ^ {i} A _ {i} ^ {j}}{\\left| \\mathcal {Y} _ {1} \\right| + (i - 1) \\left| \\mathcal {Y} _ {n o v e l} \\right|} = \\frac {\\frac {\\left| \\mathcal {Y} _ {1} \\right|}{\\left| \\mathcal {Y} _ {n o v e l} \\right|} A _ {i} ^ {1} + \\sum_ {j = 2} ^ {i} A _ {i} ^ {j}}{\\frac {\\left| \\mathcal {Y} _ {1} \\right|}{\\left| \\mathcal {Y} _ {n o v e l} \\right|} + (i - 1)}. \\tag {3}", + "image_path": "bd3e0e0811a76b02477801f7d02b241a1b6d6e593b8295d028ae122a66cca23d.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 303, + 479, + 370 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 303, + 479, + 370 + ], + "spans": [ + { + "bbox": [ + 131, + 303, + 479, + 370 + ], + "type": "text", + "content": "From Eq. (3) we see that the value of " + }, + { + "bbox": [ + 131, + 303, + 479, + 370 + ], + "type": "inline_equation", + "content": "aAcc_i" + }, + { + "bbox": [ + 131, + 303, + 479, + 370 + ], + "type": "text", + "content": " will be dominated by the base class accuracy " + }, + { + "bbox": [ + 131, + 303, + 479, + 370 + ], + "type": "inline_equation", + "content": "A_i^1" + }, + { + "bbox": [ + 131, + 303, + 479, + 370 + ], + "type": "text", + "content": " if the ratio " + }, + { + "bbox": [ + 131, + 303, + 479, + 370 + ], + "type": "inline_equation", + "content": "\\frac{|\\mathcal{Y}_1|}{|\\mathcal{Y}_{\\text{novel}}|}" + }, + { + "bbox": [ + 131, + 303, + 479, + 370 + ], + "type": "text", + "content": " is large. Unfortunately, the ratio is indeed large in the FSCIL settings. (e.g. " + }, + { + "bbox": [ + 131, + 303, + 479, + 370 + ], + "type": "inline_equation", + "content": "\\frac{|\\mathcal{Y}_1|}{|\\mathcal{Y}_{\\text{novel}}|} = 12" + }, + { + "bbox": [ + 131, + 303, + 479, + 370 + ], + "type": "text", + "content": " for both CIFAR-100 and miniImageNet dataset, 10 for CUB-200.) Under such a large ratio, the changes (both improvement and degrades) of novel classes cannot be effectively evaluated by " + }, + { + "bbox": [ + 131, + 303, + 479, + 370 + ], + "type": "inline_equation", + "content": "aAcc" + }, + { + "bbox": [ + 131, + 303, + 479, + 370 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 374, + 338, + 386 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 374, + 338, + 386 + ], + "spans": [ + { + "bbox": [ + 131, + 374, + 338, + 386 + ], + "type": "text", + "content": "3.2 Generalizing the Average Accuracy" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 388, + 481, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 388, + 481, + 411 + ], + "spans": [ + { + "bbox": [ + 131, + 388, + 481, + 411 + ], + "type": "text", + "content": "To overcome the limitations of existing metrics, we now define the generalized accuracy (gAcc) at task " + }, + { + "bbox": [ + 131, + 388, + 481, + 411 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_i" + }, + { + "bbox": [ + 131, + 388, + 481, + 411 + ], + "type": "text", + "content": " as:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 228, + 416, + 481, + 452 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 228, + 416, + 481, + 452 + ], + "spans": [ + { + "bbox": [ + 228, + 416, + 481, + 452 + ], + "type": "interline_equation", + "content": "g A c c _ {i} (\\alpha) = \\frac {\\alpha \\frac {| \\mathcal {Y} _ {1} |}{| \\mathcal {Y} _ {n o v e l} |} A _ {i} ^ {1} + \\sum_ {j = 2} ^ {i} A _ {i} ^ {j}}{\\alpha \\frac {| \\mathcal {Y} _ {1} |}{| \\mathcal {Y} _ {n o v e l} |} + (i - 1)}, \\tag {4}", + "image_path": "c44f62394f12829bde2fcbedc8900889366d66a724d72243dbbe67f26c465a13.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 455, + 482, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 455, + 482, + 579 + ], + "spans": [ + { + "bbox": [ + 131, + 455, + 482, + 579 + ], + "type": "text", + "content": "in which the " + }, + { + "bbox": [ + 131, + 455, + 482, + 579 + ], + "type": "inline_equation", + "content": "gAcc" + }, + { + "bbox": [ + 131, + 455, + 482, + 579 + ], + "type": "text", + "content": " is parameterized by " + }, + { + "bbox": [ + 131, + 455, + 482, + 579 + ], + "type": "inline_equation", + "content": "\\alpha \\in [0,1]" + }, + { + "bbox": [ + 131, + 455, + 482, + 579 + ], + "type": "text", + "content": ". At task " + }, + { + "bbox": [ + 131, + 455, + 482, + 579 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_i" + }, + { + "bbox": [ + 131, + 455, + 482, + 579 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 131, + 455, + 482, + 579 + ], + "type": "inline_equation", + "content": "gAcc_i(0)" + }, + { + "bbox": [ + 131, + 455, + 482, + 579 + ], + "type": "text", + "content": " measures the accuracy of the novel tasks only. " + }, + { + "bbox": [ + 131, + 455, + 482, + 579 + ], + "type": "inline_equation", + "content": "gAcc_i(\\frac{|\\mathcal{Y}_{\\text{novel}}|}{|\\mathcal{Y}_1|})" + }, + { + "bbox": [ + 131, + 455, + 482, + 579 + ], + "type": "text", + "content": " denotes the task-wise accuracy " + }, + { + "bbox": [ + 131, + 455, + 482, + 579 + ], + "type": "inline_equation", + "content": "tAcc_i" + }, + { + "bbox": [ + 131, + 455, + 482, + 579 + ], + "type": "text", + "content": " at task " + }, + { + "bbox": [ + 131, + 455, + 482, + 579 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_i" + }, + { + "bbox": [ + 131, + 455, + 482, + 579 + ], + "type": "text", + "content": " while " + }, + { + "bbox": [ + 131, + 455, + 482, + 579 + ], + "type": "inline_equation", + "content": "gAcc(1)" + }, + { + "bbox": [ + 131, + 455, + 482, + 579 + ], + "type": "text", + "content": " represents the average accuracy " + }, + { + "bbox": [ + 131, + 455, + 482, + 579 + ], + "type": "inline_equation", + "content": "aAcc_i" + }, + { + "bbox": [ + 131, + 455, + 482, + 579 + ], + "type": "text", + "content": " at task " + }, + { + "bbox": [ + 131, + 455, + 482, + 579 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_i" + }, + { + "bbox": [ + 131, + 455, + 482, + 579 + ], + "type": "text", + "content": ". In addition to these discrete values, we further generalize the " + }, + { + "bbox": [ + 131, + 455, + 482, + 579 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 131, + 455, + 482, + 579 + ], + "type": "text", + "content": " to any rational numbers in [0,1]. Controlled by " + }, + { + "bbox": [ + 131, + 455, + 482, + 579 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 131, + 455, + 482, + 579 + ], + "type": "text", + "content": " in Eq. (4), when " + }, + { + "bbox": [ + 131, + 455, + 482, + 579 + ], + "type": "inline_equation", + "content": "\\alpha \\to 0" + }, + { + "bbox": [ + 131, + 455, + 482, + 579 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 131, + 455, + 482, + 579 + ], + "type": "inline_equation", + "content": "gAcc(\\alpha)" + }, + { + "bbox": [ + 131, + 455, + 482, + 579 + ], + "type": "text", + "content": " emphasizes to novel class while " + }, + { + "bbox": [ + 131, + 455, + 482, + 579 + ], + "type": "inline_equation", + "content": "\\alpha \\to 1" + }, + { + "bbox": [ + 131, + 455, + 482, + 579 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 131, + 455, + 482, + 579 + ], + "type": "inline_equation", + "content": "gAcc(\\alpha)" + }, + { + "bbox": [ + 131, + 455, + 482, + 579 + ], + "type": "text", + "content": " emphasizes the base class performance, i.e., the larger the " + }, + { + "bbox": [ + 131, + 455, + 482, + 579 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 131, + 455, + 482, + 579 + ], + "type": "text", + "content": ", the greater the influence of the base class on the value of " + }, + { + "bbox": [ + 131, + 455, + 482, + 579 + ], + "type": "inline_equation", + "content": "gAcc(\\alpha)" + }, + { + "bbox": [ + 131, + 455, + 482, + 579 + ], + "type": "text", + "content": ". To evaluate the performance of a model on a specific task " + }, + { + "bbox": [ + 131, + 455, + 482, + 579 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_i" + }, + { + "bbox": [ + 131, + 455, + 482, + 579 + ], + "type": "text", + "content": ", a standalone value is needed. Thus, we introduce the area under the curve (AUC) of " + }, + { + "bbox": [ + 131, + 455, + 482, + 579 + ], + "type": "inline_equation", + "content": "gAcc_i(\\alpha)" + }, + { + "bbox": [ + 131, + 455, + 482, + 579 + ], + "type": "text", + "content": " along the parameter " + }, + { + "bbox": [ + 131, + 455, + 482, + 579 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 131, + 455, + 482, + 579 + ], + "type": "text", + "content": " as a comprehensive metric:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 252, + 582, + 481, + 609 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 252, + 582, + 481, + 609 + ], + "spans": [ + { + "bbox": [ + 252, + 582, + 481, + 609 + ], + "type": "interline_equation", + "content": "g A c c _ {i} = \\int_ {0} ^ {1} g A c c _ {i} (\\alpha) d \\alpha . \\tag {5}", + "image_path": "12bdee418e0c040a07863fa822a21e18a1c30e1d6102d7a6232fdcf71a9f59d1.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 131, + 613, + 479, + 636 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 613, + 479, + 636 + ], + "spans": [ + { + "bbox": [ + 131, + 613, + 479, + 636 + ], + "type": "text", + "content": "Similar to " + }, + { + "bbox": [ + 131, + 613, + 479, + 636 + ], + "type": "inline_equation", + "content": "aA_{cc}" + }, + { + "bbox": [ + 131, + 613, + 479, + 636 + ], + "type": "text", + "content": ", for multiple tasks, we average the " + }, + { + "bbox": [ + 131, + 613, + 479, + 636 + ], + "type": "inline_equation", + "content": "gA_{cc_i}" + }, + { + "bbox": [ + 131, + 613, + 479, + 636 + ], + "type": "text", + "content": " for each task and get an overall metric " + }, + { + "bbox": [ + 131, + 613, + 479, + 636 + ], + "type": "inline_equation", + "content": "gA_{cc}" + }, + { + "bbox": [ + 131, + 613, + 479, + 636 + ], + "type": "text", + "content": " for the whole learning process:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 264, + 641, + 481, + 667 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 264, + 641, + 481, + 667 + ], + "spans": [ + { + "bbox": [ + 264, + 641, + 481, + 667 + ], + "type": "interline_equation", + "content": "g A c c = \\frac {1}{n _ {t}} \\sum_ {i} ^ {n _ {t}} g A c c _ {i}. \\tag {6}", + "image_path": "fcb0f8fa473e93b3921888cf7bacd7a4def0c484ad2a9837b3607178bd61c578.jpg" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 150, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 150, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 150, + 91, + 448, + 102 + ], + "type": "text", + "content": "Rethinking Few-shot Class-incremental Learning: Learning from Yourself" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 159, + 114, + 353, + 232 + ], + "blocks": [ + { + "bbox": [ + 159, + 114, + 353, + 232 + ], + "lines": [ + { + "bbox": [ + 159, + 114, + 353, + 232 + ], + "spans": [ + { + "bbox": [ + 159, + 114, + 353, + 232 + ], + "type": "image", + "image_path": "a53e9f37d5a627240f619f39ff48fc55fcbfd334b536fff37414fda673768647.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 357, + 116, + 445, + 126 + ], + "lines": [ + { + "bbox": [ + 357, + 116, + 445, + 126 + ], + "spans": [ + { + "bbox": [ + 357, + 116, + 445, + 126 + ], + "type": "text", + "content": "(b) " + }, + { + "bbox": [ + 357, + 116, + 445, + 126 + ], + "type": "inline_equation", + "content": "gAcc_{i}" + }, + { + "bbox": [ + 357, + 116, + 445, + 126 + ], + "type": "text", + "content": " curve w.r.t Task i" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 354, + 133, + 445, + 232 + ], + "blocks": [ + { + "bbox": [ + 354, + 133, + 445, + 232 + ], + "lines": [ + { + "bbox": [ + 354, + 133, + 445, + 232 + ], + "spans": [ + { + "bbox": [ + 354, + 133, + 445, + 232 + ], + "type": "image", + "image_path": "f0b4f97fa5743c99779712f70fd62bb29d5c2ae4964f5a990daa41f26436e59b.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 239, + 482, + 299 + ], + "lines": [ + { + "bbox": [ + 130, + 239, + 482, + 299 + ], + "spans": [ + { + "bbox": [ + 130, + 239, + 482, + 299 + ], + "type": "text", + "content": "Fig. 2: FSCIL performances are shown in generalized accuracy. (a): The " + }, + { + "bbox": [ + 130, + 239, + 482, + 299 + ], + "type": "inline_equation", + "content": "gAcc" + }, + { + "bbox": [ + 130, + 239, + 482, + 299 + ], + "type": "text", + "content": " curve (averaged across all " + }, + { + "bbox": [ + 130, + 239, + 482, + 299 + ], + "type": "inline_equation", + "content": "n_t" + }, + { + "bbox": [ + 130, + 239, + 482, + 299 + ], + "type": "text", + "content": " tasks) v.s. param " + }, + { + "bbox": [ + 130, + 239, + 482, + 299 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 130, + 239, + 482, + 299 + ], + "type": "text", + "content": ". (b): The " + }, + { + "bbox": [ + 130, + 239, + 482, + 299 + ], + "type": "inline_equation", + "content": "gAcc" + }, + { + "bbox": [ + 130, + 239, + 482, + 299 + ], + "type": "text", + "content": " AUC of each task " + }, + { + "bbox": [ + 130, + 239, + 482, + 299 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_i" + }, + { + "bbox": [ + 130, + 239, + 482, + 299 + ], + "type": "text", + "content": " (Eq. (5)). In the legend, we show the average AUC across tasks(Eq. (6)) of each method. We evaluate recent works including SAVC [49], NC [59], and Alice [42] on the miniImageNet dataset. Some corner cases: Lazy: The model maintains base-task performance while refusing to learn anything from novel tasks. Greedy: Regardless of the previous knowledge, the model only greedily focuses on the current task. Greedy-NF: A Non-Forget version of 'Greedy'. See more about these cases in our supplementary material." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 305, + 482, + 329 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 305, + 482, + 329 + ], + "spans": [ + { + "bbox": [ + 130, + 305, + 482, + 329 + ], + "type": "text", + "content": "The proposed " + }, + { + "bbox": [ + 130, + 305, + 482, + 329 + ], + "type": "inline_equation", + "content": "gAcc" + }, + { + "bbox": [ + 130, + 305, + 482, + 329 + ], + "type": "text", + "content": " provides different evaluation perspectives using different " + }, + { + "bbox": [ + 130, + 305, + 482, + 329 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 130, + 305, + 482, + 329 + ], + "type": "text", + "content": " so we believe such a metric can provide a more balanced evaluation of FSCIL." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 335, + 204, + 347 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 335, + 204, + 347 + ], + "spans": [ + { + "bbox": [ + 131, + 335, + 204, + 347 + ], + "type": "text", + "content": "3.3 Analysis" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 354, + 482, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 354, + 482, + 449 + ], + "spans": [ + { + "bbox": [ + 130, + 354, + 482, + 449 + ], + "type": "text", + "content": "We further study the necessity of the proposed " + }, + { + "bbox": [ + 130, + 354, + 482, + 449 + ], + "type": "inline_equation", + "content": "gAcc" + }, + { + "bbox": [ + 130, + 354, + 482, + 449 + ], + "type": "text", + "content": ". We evaluate recent SOTAs in FSCIL including SAVC [49], Alice [42] and NC-FSCIL [59], and the results are in Fig. 2. We show that the " + }, + { + "bbox": [ + 130, + 354, + 482, + 449 + ], + "type": "inline_equation", + "content": "gAcc" + }, + { + "bbox": [ + 130, + 354, + 482, + 449 + ], + "type": "text", + "content": " provides more fine-grained evaluation than the classical accuracy metric. For example, in the left part of Fig. 2, although the " + }, + { + "bbox": [ + 130, + 354, + 482, + 449 + ], + "type": "inline_equation", + "content": "gAcc(1)" + }, + { + "bbox": [ + 130, + 354, + 482, + 449 + ], + "type": "text", + "content": ", which is equal to the " + }, + { + "bbox": [ + 130, + 354, + 482, + 449 + ], + "type": "inline_equation", + "content": "aAcc" + }, + { + "bbox": [ + 130, + 354, + 482, + 449 + ], + "type": "text", + "content": ", of SAVC is close to that of NC-FSCIL, the area under the curve(AUC) of SAVC is much smaller than NC-FSCIL. This indicates that the novel class performance is much worse for SAVC, which cannot be discovered by conventional metrics." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 450, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 450, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 450, + 482, + 666 + ], + "type": "text", + "content": "Furthermore, to show the superiority of our metric, we consider three corner cases: Lazy, Greedy, and Greedy-NF. 'Lazy' stands for a type of algorithm that only focuses on maintaining base-class performance, i.e. " + }, + { + "bbox": [ + 130, + 450, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\forall i \\geq j > 1" + }, + { + "bbox": [ + 130, + 450, + 482, + 666 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 450, + 482, + 666 + ], + "type": "inline_equation", + "content": "A_{i}^{j} = 0" + }, + { + "bbox": [ + 130, + 450, + 482, + 666 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 450, + 482, + 666 + ], + "type": "inline_equation", + "content": "A_{i}^{1} = A_{1}^{1}" + }, + { + "bbox": [ + 130, + 450, + 482, + 666 + ], + "type": "text", + "content": ". By setting " + }, + { + "bbox": [ + 130, + 450, + 482, + 666 + ], + "type": "inline_equation", + "content": "A_{1}^{1} = 85\\%" + }, + { + "bbox": [ + 130, + 450, + 482, + 666 + ], + "type": "text", + "content": ", we can see that when " + }, + { + "bbox": [ + 130, + 450, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\alpha = 1" + }, + { + "bbox": [ + 130, + 450, + 482, + 666 + ], + "type": "text", + "content": " (i.e. the traditional aAcc metric), its performance is close to SOTA methods. Further, when considering popular forgetting metrics like performance dropping " + }, + { + "bbox": [ + 130, + 450, + 482, + 666 + ], + "type": "inline_equation", + "content": "(PD = A_{i}^{j} - A_{j}^{j})" + }, + { + "bbox": [ + 130, + 450, + 482, + 666 + ], + "type": "text", + "content": " and Knowledge rate " + }, + { + "bbox": [ + 130, + 450, + 482, + 666 + ], + "type": "inline_equation", + "content": "(KR = A_{i}^{j} / A_{j}^{j})" + }, + { + "bbox": [ + 130, + 450, + 482, + 666 + ], + "type": "text", + "content": ", because the forgetting is zero, its PD is 0 and KR is 1. All these metrics indicate such 'Lazy' model is superior, which is contradictory to common sense. However, our proposed gAcc reveals that its performance is far from satisfying by setting " + }, + { + "bbox": [ + 130, + 450, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\alpha < 1" + }, + { + "bbox": [ + 130, + 450, + 482, + 666 + ], + "type": "text", + "content": " and its AUC is far smaller than other methods. In other words, using the traditional metrics will indicate algorithms like 'Lazy' achieving SOTA performance while such methods perform poorly evaluated by our proposed gAcc. The gAcc can also discover other corner cases. 'Greedy' is to greedily learn the current knowledge (i.e. " + }, + { + "bbox": [ + 130, + 450, + 482, + 666 + ], + "type": "inline_equation", + "content": "A_{1}^{1} = 85\\%" + }, + { + "bbox": [ + 130, + 450, + 482, + 666 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 450, + 482, + 666 + ], + "type": "inline_equation", + "content": "A_{i}^{j} = 100\\%" + }, + { + "bbox": [ + 130, + 450, + 482, + 666 + ], + "type": "text", + "content": ", when " + }, + { + "bbox": [ + 130, + 450, + 482, + 666 + ], + "type": "inline_equation", + "content": "i = j" + }, + { + "bbox": [ + 130, + 450, + 482, + 666 + ], + "type": "text", + "content": " else 0). \"Greedy\" achieves low both aAcc and gAcc. We further eliminate the forgetting problem from 'Greedy' and define 'Greedy-NF' on top of it: set accuracy " + }, + { + "bbox": [ + 130, + 450, + 482, + 666 + ], + "type": "inline_equation", + "content": "A_{i}^{j}" + }, + { + "bbox": [ + 130, + 450, + 482, + 666 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 130, + 450, + 482, + 666 + ], + "type": "inline_equation", + "content": "100\\%" + }, + { + "bbox": [ + 130, + 450, + 482, + 666 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 130, + 450, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\forall j > 1, j < i" + }, + { + "bbox": [ + 130, + 450, + 482, + 666 + ], + "type": "text", + "content": " (no forgetting at novel classes). We find that although 'Greedy-NF' performs" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 212, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 212, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 212, + 102 + ], + "type": "text", + "content": "Tang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 178, + 114, + 320, + 248 + ], + "blocks": [ + { + "bbox": [ + 178, + 114, + 320, + 248 + ], + "lines": [ + { + "bbox": [ + 178, + 114, + 320, + 248 + ], + "spans": [ + { + "bbox": [ + 178, + 114, + 320, + 248 + ], + "type": "image", + "image_path": "5c2caab1b7a2d3e639325541be9808aa8092de4b0332b1a53223ef1eb52dc1ca.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 257, + 482, + 299 + ], + "lines": [ + { + "bbox": [ + 130, + 257, + 482, + 299 + ], + "spans": [ + { + "bbox": [ + 130, + 257, + 482, + 299 + ], + "type": "text", + "content": "Fig. 3: t-SNE [37] visualization of miniImageNet test set features from different layers of a ViT. Left: we choose layers " + }, + { + "bbox": [ + 130, + 257, + 482, + 299 + ], + "type": "inline_equation", + "content": "L_{9}" + }, + { + "bbox": [ + 130, + 257, + 482, + 299 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 257, + 482, + 299 + ], + "type": "inline_equation", + "content": "L_{10}" + }, + { + "bbox": [ + 130, + 257, + 482, + 299 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 257, + 482, + 299 + ], + "type": "inline_equation", + "content": "L_{11}" + }, + { + "bbox": [ + 130, + 257, + 482, + 299 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 257, + 482, + 299 + ], + "type": "inline_equation", + "content": "L_{12}" + }, + { + "bbox": [ + 130, + 257, + 482, + 299 + ], + "type": "text", + "content": " (final) and show both base classes and novel classes features. It is observed that shallow features are more dispersed than deep features. Right: Accuracies of different tasks on features from various layers. It is clear that intermediate features achieve better performance on novel classes and thus have better generalization ability. Best viewed in color." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 335, + 118, + 438, + 246 + ], + "blocks": [ + { + "bbox": [ + 335, + 118, + 438, + 246 + ], + "lines": [ + { + "bbox": [ + 335, + 118, + 438, + 246 + ], + "spans": [ + { + "bbox": [ + 335, + 118, + 438, + 246 + ], + "type": "image", + "image_path": "879e596e7a8f7f01e548a584ae9d6152204d8d03633b5bd54d3bd86227312a83.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 308, + 479, + 343 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 308, + 479, + 343 + ], + "spans": [ + { + "bbox": [ + 130, + 308, + 479, + 343 + ], + "type": "text", + "content": "well when majorly considering the novel-class performance " + }, + { + "bbox": [ + 130, + 308, + 479, + 343 + ], + "type": "inline_equation", + "content": "(\\alpha \\in \\{0, \\frac{1}{12}, \\frac{2}{12}, \\frac{3}{12}\\})" + }, + { + "bbox": [ + 130, + 308, + 479, + 343 + ], + "type": "text", + "content": ", the overall performance (AUC) is not satisfying. Please refer to Sec. A1 in the supplementary material for more details about these corner cases." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 345, + 480, + 451 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 345, + 480, + 451 + ], + "spans": [ + { + "bbox": [ + 130, + 345, + 480, + 451 + ], + "type": "text", + "content": "By analyzing these cases, we show the superior properties of the proposed " + }, + { + "bbox": [ + 130, + 345, + 480, + 451 + ], + "type": "inline_equation", + "content": "gAcc" + }, + { + "bbox": [ + 130, + 345, + 480, + 451 + ], + "type": "text", + "content": ". Firstly, the " + }, + { + "bbox": [ + 130, + 345, + 480, + 451 + ], + "type": "inline_equation", + "content": "gAcc" + }, + { + "bbox": [ + 130, + 345, + 480, + 451 + ], + "type": "text", + "content": " metric effectively distinguishes methods (e.g. 'Lazy') that only retain the base-task performance, thereby encouraging methods to learn from novel classes. Besides, " + }, + { + "bbox": [ + 130, + 345, + 480, + 451 + ], + "type": "inline_equation", + "content": "gAcc" + }, + { + "bbox": [ + 130, + 345, + 480, + 451 + ], + "type": "text", + "content": " avoids overemphasis on the performance of base classes or novel classes (corner case 'Greedy-NF' is still far worse than normal methods in terms of AUC), providing a more balanced evaluation outcome. In other words, the proposed " + }, + { + "bbox": [ + 130, + 345, + 480, + 451 + ], + "type": "inline_equation", + "content": "gAcc" + }, + { + "bbox": [ + 130, + 345, + 480, + 451 + ], + "type": "text", + "content": " metric provides objective and balanced results when evaluating FSCIL methods, eliminating 'shortcuts' in " + }, + { + "bbox": [ + 130, + 345, + 480, + 451 + ], + "type": "inline_equation", + "content": "aAcc" + }, + { + "bbox": [ + 130, + 345, + 480, + 451 + ], + "type": "text", + "content": " that focus only on the performance of the base classes." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 453, + 480, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 453, + 480, + 475 + ], + "spans": [ + { + "bbox": [ + 130, + 453, + 480, + 475 + ], + "type": "text", + "content": "In conclusion, the " + }, + { + "bbox": [ + 130, + 453, + 480, + 475 + ], + "type": "inline_equation", + "content": "gAcc" + }, + { + "bbox": [ + 130, + 453, + 480, + 475 + ], + "type": "text", + "content": " curve and the AUC are necessary and powerful tools to evaluate the FSCIL methods." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 487, + 275, + 501 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 487, + 275, + 501 + ], + "spans": [ + { + "bbox": [ + 131, + 487, + 275, + 501 + ], + "type": "text", + "content": "4 Learn from Yourself" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 510, + 344, + 523 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 510, + 344, + 523 + ], + "spans": [ + { + "bbox": [ + 131, + 510, + 344, + 523 + ], + "type": "text", + "content": "4.1 Preliminary and The FSCIL Setting" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 533, + 482, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 482, + 668 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 482, + 668 + ], + "type": "text", + "content": "For task " + }, + { + "bbox": [ + 130, + 533, + 482, + 668 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_i\\in \\{\\mathcal{T}_i\\}_{i = 1}^T" + }, + { + "bbox": [ + 130, + 533, + 482, + 668 + ], + "type": "text", + "content": ", the training set can be denoted as " + }, + { + "bbox": [ + 130, + 533, + 482, + 668 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{train}^{i} = \\{x_{t}^{i},y_{t}^{i}\\}_{t = 1}^{m^{i}}" + }, + { + "bbox": [ + 130, + 533, + 482, + 668 + ], + "type": "text", + "content": ". For FSCIL settings, the base task " + }, + { + "bbox": [ + 130, + 533, + 482, + 668 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_1" + }, + { + "bbox": [ + 130, + 533, + 482, + 668 + ], + "type": "text", + "content": " contains " + }, + { + "bbox": [ + 130, + 533, + 482, + 668 + ], + "type": "inline_equation", + "content": "|\\mathcal{V}_1|" + }, + { + "bbox": [ + 130, + 533, + 482, + 668 + ], + "type": "text", + "content": " classes and sufficient samples for training. For " + }, + { + "bbox": [ + 130, + 533, + 482, + 668 + ], + "type": "inline_equation", + "content": "\\forall i > 1" + }, + { + "bbox": [ + 130, + 533, + 482, + 668 + ], + "type": "text", + "content": ", there are " + }, + { + "bbox": [ + 130, + 533, + 482, + 668 + ], + "type": "inline_equation", + "content": "|\\mathcal{V}_{novel}|" + }, + { + "bbox": [ + 130, + 533, + 482, + 668 + ], + "type": "text", + "content": " classes for each task and " + }, + { + "bbox": [ + 130, + 533, + 482, + 668 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 533, + 482, + 668 + ], + "type": "text", + "content": " samples per class. Typically, " + }, + { + "bbox": [ + 130, + 533, + 482, + 668 + ], + "type": "inline_equation", + "content": "|\\mathcal{V}_{novel}|" + }, + { + "bbox": [ + 130, + 533, + 482, + 668 + ], + "type": "text", + "content": " is set to 5, and " + }, + { + "bbox": [ + 130, + 533, + 482, + 668 + ], + "type": "inline_equation", + "content": "k = 5" + }, + { + "bbox": [ + 130, + 533, + 482, + 668 + ], + "type": "text", + "content": ", i.e. a 5-way 5-shot setting. We rephrase the classification model " + }, + { + "bbox": [ + 130, + 533, + 482, + 668 + ], + "type": "inline_equation", + "content": "f(\\cdot)" + }, + { + "bbox": [ + 130, + 533, + 482, + 668 + ], + "type": "text", + "content": " in Sec. 3 as " + }, + { + "bbox": [ + 130, + 533, + 482, + 668 + ], + "type": "inline_equation", + "content": "f(\\cdot) = \\varPhi (f_B(\\cdot))" + }, + { + "bbox": [ + 130, + 533, + 482, + 668 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 130, + 533, + 482, + 668 + ], + "type": "inline_equation", + "content": "\\varPhi (\\cdot)" + }, + { + "bbox": [ + 130, + 533, + 482, + 668 + ], + "type": "text", + "content": " is the classifier and " + }, + { + "bbox": [ + 130, + 533, + 482, + 668 + ], + "type": "inline_equation", + "content": "f_{B}(\\cdot)" + }, + { + "bbox": [ + 130, + 533, + 482, + 668 + ], + "type": "text", + "content": " is the backbone. We follow previous works to use a cosine classifier: " + }, + { + "bbox": [ + 130, + 533, + 482, + 668 + ], + "type": "inline_equation", + "content": "\\varPhi(X)=P^\\top\\overline{X}" + }, + { + "bbox": [ + 130, + 533, + 482, + 668 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 533, + 482, + 668 + ], + "type": "inline_equation", + "content": "P\\in \\mathbb{R}^{d\\times |\\mathcal{V}|}" + }, + { + "bbox": [ + 130, + 533, + 482, + 668 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 533, + 482, + 668 + ], + "type": "inline_equation", + "content": "\\overline{X}\\in \\mathbb{R}^{d\\times 1}" + }, + { + "bbox": [ + 130, + 533, + 482, + 668 + ], + "type": "text", + "content": " are the L2-normalized inputs, and " + }, + { + "bbox": [ + 130, + 533, + 482, + 668 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 130, + 533, + 482, + 668 + ], + "type": "text", + "content": " is the feature dimension of " + }, + { + "bbox": [ + 130, + 533, + 482, + 668 + ], + "type": "inline_equation", + "content": "f_{B}(\\cdot)" + }, + { + "bbox": [ + 130, + 533, + 482, + 668 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 533, + 482, + 668 + ], + "type": "inline_equation", + "content": "|\\mathcal{V}| = \\sum_{i}|\\mathcal{V}_{i}|" + }, + { + "bbox": [ + 130, + 533, + 482, + 668 + ], + "type": "text", + "content": " is the number of classes seen so far. In this work, we employ the Vision Transformer(ViT) [15] in the FSCIL setting. For a ViT with " + }, + { + "bbox": [ + 130, + 533, + 482, + 668 + ], + "type": "inline_equation", + "content": "N_{L}" + }, + { + "bbox": [ + 130, + 533, + 482, + 668 + ], + "type": "text", + "content": " layers " + }, + { + "bbox": [ + 130, + 533, + 482, + 668 + ], + "type": "inline_equation", + "content": "\\{L_l\\}_{l = 1}^{N_L}" + }, + { + "bbox": [ + 130, + 533, + 482, + 668 + ], + "type": "text", + "content": ", given a input image " + }, + { + "bbox": [ + 130, + 533, + 482, + 668 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 130, + 533, + 482, + 668 + ], + "type": "text", + "content": ", the output can be denoted as " + }, + { + "bbox": [ + 130, + 533, + 482, + 668 + ], + "type": "inline_equation", + "content": "X_{N_L} = L_{N_L}(L_{N_L - 1}(\\dots L_1(x))" + }, + { + "bbox": [ + 130, + 533, + 482, + 668 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 150, + 90, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 150, + 90, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 150, + 90, + 448, + 102 + ], + "type": "text", + "content": "Rethinking Few-shot Class-incremental Learning: Learning from Yourself" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 114, + 481, + 245 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 481, + 245 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 481, + 245 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 481, + 245 + ], + "type": "image", + "image_path": "328db429ab746e1ab8b477ee54a346b95e253da5d3f0de2c69c16cde27945ca8.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 252, + 482, + 304 + ], + "lines": [ + { + "bbox": [ + 130, + 252, + 482, + 304 + ], + "spans": [ + { + "bbox": [ + 130, + 252, + 482, + 304 + ], + "type": "text", + "content": "Fig. 4: The overall framework of our method. The Feature Rectification (FR) module takes the final and the intermediate features as input and outputs the rectified features for classification. The relation transfer losses are designed to convey valuable information from intermediate features to the rectified ones. The details of these two parts are shown in the left and lower parts of the figure. Further, a cosine constraint is proposed to maintain the base class performance and a classification loss is employed to adapt to the novel classes. Best viewed in color." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 312, + 482, + 371 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 312, + 482, + 371 + ], + "spans": [ + { + "bbox": [ + 130, + 312, + 482, + 371 + ], + "type": "text", + "content": "Baseline. Recent works [3, 20, 24, 38, 47, 49, 59, 63, 76] suggest a training-frozen scheme, which trains the backbone " + }, + { + "bbox": [ + 130, + 312, + 482, + 371 + ], + "type": "inline_equation", + "content": "f_{B}(\\cdot)" + }, + { + "bbox": [ + 130, + 312, + 482, + 371 + ], + "type": "text", + "content": " and the classifier " + }, + { + "bbox": [ + 130, + 312, + 482, + 371 + ], + "type": "inline_equation", + "content": "\\varPhi(\\cdot)" + }, + { + "bbox": [ + 130, + 312, + 482, + 371 + ], + "type": "text", + "content": " (if available) on the base-task only, and then freeze the backbone during novel tasks. For novel classes, the classifier is extended by mean features (i.e. prototypes) extracted by the frozen backbone. We follow and employ the scheme as our baseline." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 372, + 482, + 598 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 372, + 482, + 598 + ], + "spans": [ + { + "bbox": [ + 130, + 372, + 482, + 598 + ], + "type": "text", + "content": "Main idea of our method. Given that the backbone is initially trained for the base task and remains fixed during novel tasks, the main challenges lie in the potential overfitting on the base task and the poor feature extraction ability of the novel classes. We conduct a toy experiment to show these problems and show that the intermediate feature can convey valuable information to tackle these problems. Results are shown in Fig. 3, where we select 4 base classes and 5 novel classes and utilize the t-SNE [37] to visualize the features from different layers. It can be concluded that features extracted by the base-classes trained backbone are overfitted to the base classes and cannot be generalized well on novel classes, i.e. the final features are well aggregated, but it is hard to distinguish among unseen classes. (see " + }, + { + "bbox": [ + 130, + 372, + 482, + 598 + ], + "type": "inline_equation", + "content": "L_{12}" + }, + { + "bbox": [ + 130, + 372, + 482, + 598 + ], + "type": "text", + "content": " features on the left and the corresponding performance on the right). Compared with the final features (i.e. " + }, + { + "bbox": [ + 130, + 372, + 482, + 598 + ], + "type": "inline_equation", + "content": "L_{12}" + }, + { + "bbox": [ + 130, + 372, + 482, + 598 + ], + "type": "text", + "content": "), the shallower intermediate features " + }, + { + "bbox": [ + 130, + 372, + 482, + 598 + ], + "type": "inline_equation", + "content": "(L_9, L_{10}, L_{11})" + }, + { + "bbox": [ + 130, + 372, + 482, + 598 + ], + "type": "text", + "content": " are more dispersed, which makes them more discriminative for unseen classes. This is cross-verified by the performance shown on the right side of Fig. 3. We can see that using intermediate features can largely boost the novel classes' performance (up to " + }, + { + "bbox": [ + 130, + 372, + 482, + 598 + ], + "type": "inline_equation", + "content": "15\\%" + }, + { + "bbox": [ + 130, + 372, + 482, + 598 + ], + "type": "text", + "content": " improvement) with a price of performance decline of the base classes. Based on this observation, we propose to rectify the final feature based on the intermediate feature to better handle both base and novel classes." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 619, + 266, + 631 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 619, + 266, + 631 + ], + "spans": [ + { + "bbox": [ + 132, + 619, + 266, + 631 + ], + "type": "text", + "content": "4.2 Feature Rectification" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "content": "We propose a Feature Rectification (FR) module to rectify the final feature based on the relation information from the intermediate features. Specifically," + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 212, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 212, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 212, + 102 + ], + "type": "text", + "content": "Tang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 131, + 116, + 479, + 139 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 116, + 479, + 139 + ], + "spans": [ + { + "bbox": [ + 131, + 116, + 479, + 139 + ], + "type": "text", + "content": "given a pair of final feature " + }, + { + "bbox": [ + 131, + 116, + 479, + 139 + ], + "type": "inline_equation", + "content": "X_{N_L}" + }, + { + "bbox": [ + 131, + 116, + 479, + 139 + ], + "type": "text", + "content": " and the intermediate feature " + }, + { + "bbox": [ + 131, + 116, + 479, + 139 + ], + "type": "inline_equation", + "content": "X_l" + }, + { + "bbox": [ + 131, + 116, + 479, + 139 + ], + "type": "text", + "content": ", we design an FR module conditioned on the intermediate features:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 223, + 148, + 481, + 161 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 223, + 148, + 481, + 161 + ], + "spans": [ + { + "bbox": [ + 223, + 148, + 481, + 161 + ], + "type": "interline_equation", + "content": "X _ {F R} = M _ {m i x} \\left(\\operatorname {c a t} \\left(M _ {F} \\left(X _ {N _ {L}}\\right), M _ {I} \\left(X _ {l}\\right)\\right), \\right. \\tag {7}", + "image_path": "7c1062f8582b952cc7c5be40db4b7f97ff0a875093628173f04ddf1a4254ce36.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 167, + 479, + 262 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 167, + 479, + 262 + ], + "spans": [ + { + "bbox": [ + 130, + 167, + 479, + 262 + ], + "type": "text", + "content": "where 'cat' means the concatenate operation. " + }, + { + "bbox": [ + 130, + 167, + 479, + 262 + ], + "type": "inline_equation", + "content": "M_{\\text{mix}}" + }, + { + "bbox": [ + 130, + 167, + 479, + 262 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 167, + 479, + 262 + ], + "type": "inline_equation", + "content": "M_F" + }, + { + "bbox": [ + 130, + 167, + 479, + 262 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 167, + 479, + 262 + ], + "type": "inline_equation", + "content": "M_I" + }, + { + "bbox": [ + 130, + 167, + 479, + 262 + ], + "type": "text", + "content": " are 2-layer MLPs. By introducing the intermediate features, the FR module has access to a reference for extracting the valuable structural information lies in the intermediate features. A naive way to mine such information is to directly pull rectified features closer to shallow features on an instance-by-instance basis. Although such an instance-wise loss can bridge shallow features and rectified features, directly pulling the rectified feature to the shallow one could lead to noisy features since the shallow features are more scattered, as shown in Sec. 4.1." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 263, + 479, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 263, + 479, + 335 + ], + "spans": [ + { + "bbox": [ + 131, + 263, + 479, + 335 + ], + "type": "text", + "content": "To effectively transfer discriminative cues among samples from shallow features to rectified features and avoid introducing noise, we propose two relation-based transfer losses at different levels. Firstly, we offer a Instance-level Relation transfer (IR) to mine the relation between feature instances. Given a pair of final features " + }, + { + "bbox": [ + 131, + 263, + 479, + 335 + ], + "type": "inline_equation", + "content": "X_{N_L}^1" + }, + { + "bbox": [ + 131, + 263, + 479, + 335 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 263, + 479, + 335 + ], + "type": "inline_equation", + "content": "X_{N_L}^2" + }, + { + "bbox": [ + 131, + 263, + 479, + 335 + ], + "type": "text", + "content": ", and another pair of intermediate features " + }, + { + "bbox": [ + 131, + 263, + 479, + 335 + ], + "type": "inline_equation", + "content": "X_l^1" + }, + { + "bbox": [ + 131, + 263, + 479, + 335 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 263, + 479, + 335 + ], + "type": "inline_equation", + "content": "x_{l}^{2}" + }, + { + "bbox": [ + 131, + 263, + 479, + 335 + ], + "type": "text", + "content": ", the IR loss is defined as:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 233, + 342, + 480, + 356 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 233, + 342, + 480, + 356 + ], + "spans": [ + { + "bbox": [ + 233, + 342, + 480, + 356 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {I R} = \\delta \\left(\\epsilon \\left(X _ {F R} ^ {1}, X _ {F R} ^ {2}\\right), \\epsilon \\left(X _ {l} ^ {1}, x _ {l} ^ {2}\\right)\\right), \\tag {8}", + "image_path": "90bc69f5a2a0109a92d182682d5b92a63f9a73d7e85b2090ead352898e9c2914.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 362, + 482, + 458 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 362, + 482, + 458 + ], + "spans": [ + { + "bbox": [ + 131, + 362, + 482, + 458 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 362, + 482, + 458 + ], + "type": "inline_equation", + "content": "\\epsilon (\\cdot)" + }, + { + "bbox": [ + 131, + 362, + 482, + 458 + ], + "type": "text", + "content": " is the Euclidean distance and the " + }, + { + "bbox": [ + 131, + 362, + 482, + 458 + ], + "type": "inline_equation", + "content": "\\delta (\\cdot)" + }, + { + "bbox": [ + 131, + 362, + 482, + 458 + ], + "type": "text", + "content": " denotes the smooth L1 loss. Besides the relation between feature instances, another level of the relation information is the relation between features and class centers. Class-center Relation Transfer (CR) is proposed to mine this class-center-level information. Assume at task " + }, + { + "bbox": [ + 131, + 362, + 482, + 458 + ], + "type": "inline_equation", + "content": "T_{i}" + }, + { + "bbox": [ + 131, + 362, + 482, + 458 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 131, + 362, + 482, + 458 + ], + "type": "inline_equation", + "content": "|\\mathcal{V}_i|" + }, + { + "bbox": [ + 131, + 362, + 482, + 458 + ], + "type": "text", + "content": " classes and the feature dimension is " + }, + { + "bbox": [ + 131, + 362, + 482, + 458 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 131, + 362, + 482, + 458 + ], + "type": "text", + "content": ". Given a set of class prototypes " + }, + { + "bbox": [ + 131, + 362, + 482, + 458 + ], + "type": "inline_equation", + "content": "P\\in \\mathbb{R}^{d\\times |\\mathcal{V}_i|}" + }, + { + "bbox": [ + 131, + 362, + 482, + 458 + ], + "type": "text", + "content": " from the classifier " + }, + { + "bbox": [ + 131, + 362, + 482, + 458 + ], + "type": "inline_equation", + "content": "\\varPhi(\\cdot)" + }, + { + "bbox": [ + 131, + 362, + 482, + 458 + ], + "type": "text", + "content": " as well as a pair of final feature " + }, + { + "bbox": [ + 131, + 362, + 482, + 458 + ], + "type": "inline_equation", + "content": "X_{N_L}\\in \\mathbb{R}^{d\\times 1}" + }, + { + "bbox": [ + 131, + 362, + 482, + 458 + ], + "type": "text", + "content": " and the intermediate feature " + }, + { + "bbox": [ + 131, + 362, + 482, + 458 + ], + "type": "inline_equation", + "content": "X_{l}" + }, + { + "bbox": [ + 131, + 362, + 482, + 458 + ], + "type": "inline_equation", + "content": "(l\\in [1,N_L))" + }, + { + "bbox": [ + 131, + 362, + 482, + 458 + ], + "type": "text", + "content": ", the corresponding CR loss is formulated as:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 230, + 465, + 480, + 479 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 230, + 465, + 480, + 479 + ], + "spans": [ + { + "bbox": [ + 230, + 465, + 480, + 479 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {C R} = \\mathrm {K L} \\left(\\left(\\sigma \\left(P ^ {\\top} \\overline {{X _ {F R}}}\\right), \\sigma \\left(P ^ {\\top} \\overline {{X _ {l}}}\\right)\\right), \\right. \\tag {9}", + "image_path": "d06c642a4c19573fab8e59be5b105e5b268e023f7943bba1534a430e4359d9c8.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 486, + 479, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 486, + 479, + 509 + ], + "spans": [ + { + "bbox": [ + 131, + 486, + 479, + 509 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 486, + 479, + 509 + ], + "type": "inline_equation", + "content": "\\sigma (\\cdot)" + }, + { + "bbox": [ + 131, + 486, + 479, + 509 + ], + "type": "text", + "content": " is the softmax function, and " + }, + { + "bbox": [ + 131, + 486, + 479, + 509 + ], + "type": "inline_equation", + "content": "\\mathrm{KL}(\\cdot)" + }, + { + "bbox": [ + 131, + 486, + 479, + 509 + ], + "type": "text", + "content": " is the Kullback-Leibler divergence. " + }, + { + "bbox": [ + 131, + 486, + 479, + 509 + ], + "type": "inline_equation", + "content": "\\overline{X}" + }, + { + "bbox": [ + 131, + 486, + 479, + 509 + ], + "type": "text", + "content": " denotes the L2 normalization results of " + }, + { + "bbox": [ + 131, + 486, + 479, + 509 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 131, + 486, + 479, + 509 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 510, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 510, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 131, + 510, + 482, + 665 + ], + "type": "text", + "content": "In the above equations, the two losses " + }, + { + "bbox": [ + 131, + 510, + 482, + 665 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{CR}" + }, + { + "bbox": [ + 131, + 510, + 482, + 665 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 510, + 482, + 665 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{IR}" + }, + { + "bbox": [ + 131, + 510, + 482, + 665 + ], + "type": "text", + "content": " work on different levels and serve different roles. " + }, + { + "bbox": [ + 131, + 510, + 482, + 665 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{CR}" + }, + { + "bbox": [ + 131, + 510, + 482, + 665 + ], + "type": "text", + "content": " works on the relations between a sample and all class centers, focusing on the coarse-grained class-specific patterns that distinguish different classes. Differently, " + }, + { + "bbox": [ + 131, + 510, + 482, + 665 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{IR}" + }, + { + "bbox": [ + 131, + 510, + 482, + 665 + ], + "type": "text", + "content": " provides knowledge about fine-grained discriminative cues that contribute to pairwise relations. The proposed relation-based transfer losses are related to previous knowledge distillation works [5, 40, 41]. However, these works do not share the same objective as ours. Our goal is to extract precious transferrable information hidden in various shallow layers hierarchically to rectify the final features while prior distillation methods focus on matching final outputs to teachers. Furthermore, there exists a branch of works using intermediate features [36, 62] directly for dense prediction. Differently, we are the first to extract intermediate knowledge to rectify the final feature for a classification task." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 150, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 150, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 150, + 91, + 448, + 102 + ], + "type": "text", + "content": "Rethinking Few-shot Class-incremental Learning: Learning from Yourself" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 152 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 152 + ], + "type": "text", + "content": "Besides, we constrain the discriminative ability of rectified features on base classes to avoid negative rectification. Since the final feature of layer " + }, + { + "bbox": [ + 130, + 116, + 482, + 152 + ], + "type": "inline_equation", + "content": "L_{N_L}" + }, + { + "bbox": [ + 130, + 116, + 482, + 152 + ], + "type": "text", + "content": " is strong enough for the base classes, we apply a cosine constraint as:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 256, + 159, + 481, + 171 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 256, + 159, + 481, + 171 + ], + "spans": [ + { + "bbox": [ + 256, + 159, + 481, + 171 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {c o s}} = \\cos \\left(X _ {F R}, X _ {N _ {L}}\\right), \\tag {10}", + "image_path": "7431c5bdac57b2bec155ec8e3426904c93a15fe20a670aa5e1a6297c27ae0fff.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 177, + 280, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 177, + 280, + 189 + ], + "spans": [ + { + "bbox": [ + 130, + 177, + 280, + 189 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 177, + 280, + 189 + ], + "type": "inline_equation", + "content": "\\cos (\\cdot ,\\cdot)" + }, + { + "bbox": [ + 130, + 177, + 280, + 189 + ], + "type": "text", + "content": " is cosine similarity." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 190, + 479, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 190, + 479, + 213 + ], + "spans": [ + { + "bbox": [ + 130, + 190, + 479, + 213 + ], + "type": "text", + "content": "Novel class training. To ensure the rectified features are centered around the class mean prototypes, we further introduce a classification loss for novel classes:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 207, + 224, + 481, + 237 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 207, + 224, + 481, + 237 + ], + "spans": [ + { + "bbox": [ + 207, + 224, + 481, + 237 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {N o v C E}} = - \\mathbb {1} (y \\in \\cup_ {n > 1} \\mathcal {Y} _ {n}) \\log \\sigma \\left(P ^ {\\top} \\overline {{X _ {F R}}}\\right) ^ {(y)}, \\tag {11}", + "image_path": "d8eab4c9ceaf54a5b788918139c74d7bb3ef08f3806e782b04dea220189aa92d.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 241, + 479, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 241, + 479, + 266 + ], + "spans": [ + { + "bbox": [ + 130, + 241, + 479, + 266 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 241, + 479, + 266 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 130, + 241, + 479, + 266 + ], + "type": "text", + "content": " is the corresponding ground true label of " + }, + { + "bbox": [ + 130, + 241, + 479, + 266 + ], + "type": "inline_equation", + "content": "X_{FR}" + }, + { + "bbox": [ + 130, + 241, + 479, + 266 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 241, + 479, + 266 + ], + "type": "inline_equation", + "content": "(P^{\\top}\\overline{X_{FR}})^{(y)}" + }, + { + "bbox": [ + 130, + 241, + 479, + 266 + ], + "type": "text", + "content": " is the " + }, + { + "bbox": [ + 130, + 241, + 479, + 266 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 130, + 241, + 479, + 266 + ], + "type": "text", + "content": "-th element of " + }, + { + "bbox": [ + 130, + 241, + 479, + 266 + ], + "type": "inline_equation", + "content": "(P^{\\top}\\overline{X_{FR}})" + }, + { + "bbox": [ + 130, + 241, + 479, + 266 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 266, + 480, + 408 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 266, + 480, + 408 + ], + "spans": [ + { + "bbox": [ + 130, + 266, + 480, + 408 + ], + "type": "text", + "content": "Anti-forgetting training. Another challenge of FSCIL is the forgetting problem. Thanks to our feature rectification design, the only trainable module is the FR module. As the FR module takes the final and the intermediate feature as the input, these kinds of features can be easily modeled by Gaussian distributions [51,71]. Hence, we construct a Gaussian distribution for each class on each layer. The Gaussian distribution of class " + }, + { + "bbox": [ + 130, + 266, + 480, + 408 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 266, + 480, + 408 + ], + "type": "text", + "content": " features at layer " + }, + { + "bbox": [ + 130, + 266, + 480, + 408 + ], + "type": "inline_equation", + "content": "L_{l}" + }, + { + "bbox": [ + 130, + 266, + 480, + 408 + ], + "type": "text", + "content": " can be denoted as " + }, + { + "bbox": [ + 130, + 266, + 480, + 408 + ], + "type": "inline_equation", + "content": "\\mathcal{N}_i^l (\\mu_i^l,\\Sigma_i^l)" + }, + { + "bbox": [ + 130, + 266, + 480, + 408 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 266, + 480, + 408 + ], + "type": "inline_equation", + "content": "\\Sigma_i^l" + }, + { + "bbox": [ + 130, + 266, + 480, + 408 + ], + "type": "text", + "content": " is obtained by calculating the covariance of each feature dimension if " + }, + { + "bbox": [ + 130, + 266, + 480, + 408 + ], + "type": "inline_equation", + "content": "i\\in \\mathcal{Y}_1" + }, + { + "bbox": [ + 130, + 266, + 480, + 408 + ], + "type": "text", + "content": " (base classes). For novel classes, the covariance matrix is computed by averaging the covariance matrices of the top-k most similar base classes. During the incremental learning phases, we utilize these Gaussian distributions to sample features, which serve as inputs from old tasks to the FR module, assisting the FR module in retaining previously learned knowledge." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 408, + 479, + 432 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 408, + 479, + 432 + ], + "spans": [ + { + "bbox": [ + 130, + 408, + 479, + 432 + ], + "type": "text", + "content": "Overall training objectives. The total loss function for our feature rectification framework is as follows:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 203, + 439, + 480, + 452 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 203, + 439, + 480, + 452 + ], + "spans": [ + { + "bbox": [ + 203, + 439, + 480, + 452 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\beta_ {\\text {c o s}} \\left(\\mathcal {L} _ {\\text {c o s}} + \\mathcal {L} _ {\\text {N o v C e}}\\right) + \\beta_ {C R} \\mathcal {L} _ {C R} + \\beta_ {I R} \\mathcal {L} _ {I R}, \\tag {12}", + "image_path": "da05cf6ec0dd127610a12d94eb751afb59bcdb217a5b6f3ee6bcd23a9171443a.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 457, + 354, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 457, + 354, + 469 + ], + "spans": [ + { + "bbox": [ + 130, + 457, + 354, + 469 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 457, + 354, + 469 + ], + "type": "inline_equation", + "content": "\\beta_{cos},\\beta_{CR}" + }, + { + "bbox": [ + 130, + 457, + 354, + 469 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 457, + 354, + 469 + ], + "type": "inline_equation", + "content": "\\beta_{IR}" + }, + { + "bbox": [ + 130, + 457, + 354, + 469 + ], + "type": "text", + "content": " are trade-off parameters." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 470, + 482, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 470, + 482, + 582 + ], + "spans": [ + { + "bbox": [ + 130, + 470, + 482, + 582 + ], + "type": "text", + "content": "Multi-layer knowledge ensemble. Since there exists hierarchical knowledge in different layers, we mine such information by knowledge ensemble from different layers. Specifically, we incorporate an FR module with each selected layer, resulting in a total of " + }, + { + "bbox": [ + 130, + 470, + 482, + 582 + ], + "type": "inline_equation", + "content": "n_{FR}" + }, + { + "bbox": [ + 130, + 470, + 482, + 582 + ], + "type": "text", + "content": " FR modules. During the training phase, these FR modules are optimized simultaneously. When evaluation, FR modules at layer " + }, + { + "bbox": [ + 130, + 470, + 482, + 582 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 130, + 470, + 482, + 582 + ], + "type": "text", + "content": " are employed to produce multiple rectified features denoted as " + }, + { + "bbox": [ + 130, + 470, + 482, + 582 + ], + "type": "inline_equation", + "content": "X_{FR}^{l}" + }, + { + "bbox": [ + 130, + 470, + 482, + 582 + ], + "type": "text", + "content": ". These rectified features are subsequently input into the shared classifier " + }, + { + "bbox": [ + 130, + 470, + 482, + 582 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 130, + 470, + 482, + 582 + ], + "type": "text", + "content": " for generating multiple predictions: " + }, + { + "bbox": [ + 130, + 470, + 482, + 582 + ], + "type": "inline_equation", + "content": "\\mathrm{Pred}_l = \\sigma (P^\\top \\overline{X_{FR}^l})" + }, + { + "bbox": [ + 130, + 470, + 482, + 582 + ], + "type": "text", + "content": ". The final prediction is computed as the average of each FR-branch: " + }, + { + "bbox": [ + 130, + 470, + 482, + 582 + ], + "type": "inline_equation", + "content": "\\mathrm{Pred} = \\frac{1}{n_{FR}}\\sum_{l = 1}^{n_{FR}}\\mathrm{Pred}_l" + }, + { + "bbox": [ + 130, + 470, + 482, + 582 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 131, + 594, + 230, + 608 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 594, + 230, + 608 + ], + "spans": [ + { + "bbox": [ + 131, + 594, + 230, + 608 + ], + "type": "text", + "content": "5 Experiments" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 131, + 612, + 253, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 612, + 253, + 624 + ], + "spans": [ + { + "bbox": [ + 131, + 612, + 253, + 624 + ], + "type": "text", + "content": "5.1 Experiment Setup" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "type": "text", + "content": "Datasets. Following previous works, we conduct experiments on three popular datasets of FSCIL: CIFAR-100 [28], miniImageNet [46] and CUB-200 [55]. For statistics of the three datasets, please refer to the supplementary material." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 212, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 212, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 212, + 102 + ], + "type": "text", + "content": "Tang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 155, + 141, + 452, + 315 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 482, + 139 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 482, + 139 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 482, + 139 + ], + "type": "text", + "content": "Table 1: Comparison with FSCIL methods on miniImageNet. The best and second best results are marked bolded and underlined. Methods with†are reported with the original results in their paper. Other results are reimplemented by us." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 155, + 141, + 452, + 315 + ], + "lines": [ + { + "bbox": [ + 155, + 141, + 452, + 315 + ], + "spans": [ + { + "bbox": [ + 155, + 141, + 452, + 315 + ], + "type": "table", + "html": "
MethodsNum. of \nParams.(M)aAcc in each task (%)
123456789aAccgAcc
TOPIC [52]†-61.350.145.241.237.535.532.229.524.439.64-
IDLVQ [6]†-64.859.855.952.649.947.644.843.141.851.16-
MetaFSCIL [8]†-72.068.063.860.357.655.253.050.849.258.85-
DF-Replay [35]†-71.867.163.259.857.054.051.649.548.258.02-
FCIL [18]†-76.3471.467.164.161.358.555.754.152.862.37-
Self-promoted [73]12.3463.664.159.555.552.149.245.643.140.952.6241.53
CEC [63]12.1272.167.063.059.656.953.951.449.547.857.9148.64
LIMIT [70]12.2973.367.663.259.556.753.751.449.748.258.1549.47
FACT [69]11.2676.371.366.962.959.656.353.551.149.260.7849.56
CLOM [76]14.2272.166.762.759.256.153.050.248.346.757.2346.11
C-FSCIL [20]12.7970.663.859.356.454.351.848.947.346.155.4050.46
BiDistFSCIL [66]11.5074.670.466.662.860.557.254.753.051.961.3253.49
S3C [24]0.3477.072.168.265.162.058.856.053.651.962.7453.70
Regularizer [3]26.3081.972.768.764.157.453.553.852.343.162.8255.62
SubNetwork [25]22.3878.173.569.866.263.260.357.655.754.564.3155.88
SAVC [49]12.1281.076.271.971.465.662.159.257.055.566.6656.35
ALICE [42]11.2081.171.167.864.162.460.157.956.856.064.1459.72
NC-FSCIL [59]15.8784.276.172.168.167.363.961.959.657.967.9060.93
YourSelf (Ours)11.5784.077.673.770.068.064.962.159.859.068.8062.20
", + "image_path": "16ed78749bd485f56215e95935f94559d6bd7f5e0b4b85b4622c17378739ef06.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 317, + 485, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 317, + 485, + 437 + ], + "spans": [ + { + "bbox": [ + 130, + 317, + 485, + 437 + ], + "type": "text", + "content": "Training and testing protocols. For training protocols, we follow prior works [3, 20, 24, 25, 42, 49, 59, 66] to split a large proportion of the training set to serve as the base task, and split the rest into small incremental tasks. Specifically, for CIFAR-100 and miniImageNet, the first task contains 60 classes out of the total 100 classes, and the rest 40 classes are split into 8 tasks, following a 5-way 5-shot scheme. For CUB-200, there are 100 classes in the first task and the number of novel tasks is 10 (in a 10-way 5-shot setting). As for the testing, we follow prior works to report the average accuracy (" + }, + { + "bbox": [ + 130, + 317, + 485, + 437 + ], + "type": "inline_equation", + "content": "aAcc" + }, + { + "bbox": [ + 130, + 317, + 485, + 437 + ], + "type": "text", + "content": "). In addition, we also provide detailed discussions on our metric generalized accuracy (" + }, + { + "bbox": [ + 130, + 317, + 485, + 437 + ], + "type": "inline_equation", + "content": "gAcc" + }, + { + "bbox": [ + 130, + 317, + 485, + 437 + ], + "type": "text", + "content": "). We further provide detailed results of novel classes only in our supplementary material." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 437, + 483, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 437, + 483, + 594 + ], + "spans": [ + { + "bbox": [ + 130, + 437, + 483, + 594 + ], + "type": "text", + "content": "Implementation details. For all experiments below, we use the ViT suggested by Touvron et, al. [53]. We adjust stuff like the dimensions of the backbone to match the number of parameters (around 10M) as the ResNet18 [19] which is widely used in this topic. The Feature Rectification (FR) is a lightweight module and uses popular MLPs implementation [53] with only 0.6M parameters. To ensure a fair comparison, we also re-implement other methods with the same backbone as ours, but found the performance is rarely improved. Please refer to Supplementary material for more details. We follow previous works to train from scratch on CIFAR-100 and miniImageNet dataset. Prior works [6,8,35, 52,63,73] adopt an ImageNet-pretrained ResNet for the CUB-200 dataset. To make a fair comparison, we also pretrained our backbone on the ImageNet [12] to evaluate on CUB-200. For evaluation, we reproduced 14 methods based on the public codes and conducted in-depth comparisons in the following sections." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 609, + 377, + 623 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 609, + 377, + 623 + ], + "spans": [ + { + "bbox": [ + 131, + 609, + 377, + 623 + ], + "type": "text", + "content": "5.2 Comparison with State-of-the-art methods" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "text", + "content": "Experimental results on miniImageNet are shown in Table 1. We report the classical accuracy (" + }, + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "inline_equation", + "content": "aAcc" + }, + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "text", + "content": ") in each section as well as its averaged results and our proposed " + }, + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "inline_equation", + "content": "gAcc" + }, + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "text", + "content": ". From the table, we can conclude that our proposed method" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 150, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 150, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 150, + 91, + 448, + 102 + ], + "type": "text", + "content": "Rethinking Few-shot Class-incremental Learning: Learning from Yourself" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 117, + 218, + 230 + ], + "blocks": [ + { + "bbox": [ + 133, + 117, + 218, + 230 + ], + "lines": [ + { + "bbox": [ + 133, + 117, + 218, + 230 + ], + "spans": [ + { + "bbox": [ + 133, + 117, + 218, + 230 + ], + "type": "image", + "image_path": "b03079e837c96ea84661f614d3031b6aac658f94c69deddf5e8be3102f2d926c.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 220, + 119, + 304, + 231 + ], + "blocks": [ + { + "bbox": [ + 220, + 119, + 304, + 231 + ], + "lines": [ + { + "bbox": [ + 220, + 119, + 304, + 231 + ], + "spans": [ + { + "bbox": [ + 220, + 119, + 304, + 231 + ], + "type": "image", + "image_path": "8076db5bbcad98b8b6136d43e0dc1d3cffc6a7eb1985f95e93b7a84eaef035db.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 134, + 231, + 307, + 253 + ], + "blocks": [ + { + "bbox": [ + 134, + 231, + 307, + 253 + ], + "lines": [ + { + "bbox": [ + 134, + 231, + 307, + 253 + ], + "spans": [ + { + "bbox": [ + 134, + 231, + 307, + 253 + ], + "type": "image", + "image_path": "c8680caf43746db8689de72a8a742b845a45a03f6dcf54a2f1e992f262a1b005.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 261, + 305, + 302 + ], + "lines": [ + { + "bbox": [ + 132, + 261, + 305, + 302 + ], + "spans": [ + { + "bbox": [ + 132, + 261, + 305, + 302 + ], + "type": "text", + "content": "Fig. 5: FSCIL results on miniImageNet. (a): The " + }, + { + "bbox": [ + 132, + 261, + 305, + 302 + ], + "type": "inline_equation", + "content": "gAcc" + }, + { + "bbox": [ + 132, + 261, + 305, + 302 + ], + "type": "text", + "content": " curve (averaged across all " + }, + { + "bbox": [ + 132, + 261, + 305, + 302 + ], + "type": "inline_equation", + "content": "n_t" + }, + { + "bbox": [ + 132, + 261, + 305, + 302 + ], + "type": "text", + "content": " tasks) vs. " + }, + { + "bbox": [ + 132, + 261, + 305, + 302 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 132, + 261, + 305, + 302 + ], + "type": "text", + "content": ". (b): The " + }, + { + "bbox": [ + 132, + 261, + 305, + 302 + ], + "type": "inline_equation", + "content": "gAcc" + }, + { + "bbox": [ + 132, + 261, + 305, + 302 + ], + "type": "text", + "content": " AUC of each task " + }, + { + "bbox": [ + 132, + 261, + 305, + 302 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_i" + }, + { + "bbox": [ + 132, + 261, + 305, + 302 + ], + "type": "text", + "content": " (Eq. (5)) vs. i. In the legend, we show the AUC value (Eq. (6)) of each method. Best viewed in color." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 309, + 114, + 395, + 231 + ], + "blocks": [ + { + "bbox": [ + 309, + 114, + 395, + 231 + ], + "lines": [ + { + "bbox": [ + 309, + 114, + 395, + 231 + ], + "spans": [ + { + "bbox": [ + 309, + 114, + 395, + 231 + ], + "type": "image", + "image_path": "d87760f4c26634ceadf17dbaa115231f4b0a64d14a49926304d0a7af4e2021e8.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 396, + 116, + 481, + 231 + ], + "blocks": [ + { + "bbox": [ + 396, + 116, + 481, + 231 + ], + "lines": [ + { + "bbox": [ + 396, + 116, + 481, + 231 + ], + "spans": [ + { + "bbox": [ + 396, + 116, + 481, + 231 + ], + "type": "image", + "image_path": "ca3f2436ef1d469c25d320d169fed211277cd1d4c018beaa09d71c415d596046.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 334, + 232, + 462, + 257 + ], + "blocks": [ + { + "bbox": [ + 334, + 232, + 462, + 257 + ], + "lines": [ + { + "bbox": [ + 334, + 232, + 462, + 257 + ], + "spans": [ + { + "bbox": [ + 334, + 232, + 462, + 257 + ], + "type": "image", + "image_path": "017231e1eb44ee990632f51a22c165f788e23cee0c4448460a125ecbed804d20.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 308, + 265, + 482, + 306 + ], + "lines": [ + { + "bbox": [ + 308, + 265, + 482, + 306 + ], + "spans": [ + { + "bbox": [ + 308, + 265, + 482, + 306 + ], + "type": "text", + "content": "Fig. 6: FSCIL results on CUB-200. (a): The " + }, + { + "bbox": [ + 308, + 265, + 482, + 306 + ], + "type": "inline_equation", + "content": "gAcc" + }, + { + "bbox": [ + 308, + 265, + 482, + 306 + ], + "type": "text", + "content": " curve (averaged across all " + }, + { + "bbox": [ + 308, + 265, + 482, + 306 + ], + "type": "inline_equation", + "content": "n_t" + }, + { + "bbox": [ + 308, + 265, + 482, + 306 + ], + "type": "text", + "content": " tasks) vs. " + }, + { + "bbox": [ + 308, + 265, + 482, + 306 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 308, + 265, + 482, + 306 + ], + "type": "text", + "content": ". (b): The " + }, + { + "bbox": [ + 308, + 265, + 482, + 306 + ], + "type": "inline_equation", + "content": "gAcc" + }, + { + "bbox": [ + 308, + 265, + 482, + 306 + ], + "type": "text", + "content": " AUC of each task " + }, + { + "bbox": [ + 308, + 265, + 482, + 306 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_i" + }, + { + "bbox": [ + 308, + 265, + 482, + 306 + ], + "type": "text", + "content": " (Eq. (5)) vs. " + }, + { + "bbox": [ + 308, + 265, + 482, + 306 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 308, + 265, + 482, + 306 + ], + "type": "text", + "content": ". In the legend, we show the AUC value (Eq. (6)) of each method. Best viewed in color." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 315, + 480, + 374 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 315, + 480, + 374 + ], + "spans": [ + { + "bbox": [ + 130, + 315, + 480, + 374 + ], + "type": "text", + "content": "outperforms other methods by a large margin. Specifically, under the classical evaluation metric " + }, + { + "bbox": [ + 130, + 315, + 480, + 374 + ], + "type": "inline_equation", + "content": "aAcc" + }, + { + "bbox": [ + 130, + 315, + 480, + 374 + ], + "type": "text", + "content": ", the proposed method achieves 68.80% accuracy, which is 0.90%, 4.66%, 2.14% higher than SOTA methods NC-FSCIL [59], ALICE [42] and SAVC [49]. The improvement is non-trivial considering that miniImageNet is a complex dataset." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 376, + 482, + 448 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 376, + 482, + 448 + ], + "spans": [ + { + "bbox": [ + 130, + 376, + 482, + 448 + ], + "type": "text", + "content": "From a more balanced perspective, we also report " + }, + { + "bbox": [ + 130, + 376, + 482, + 448 + ], + "type": "inline_equation", + "content": "gAcc" + }, + { + "bbox": [ + 130, + 376, + 482, + 448 + ], + "type": "text", + "content": " on the Table 1 and Fig. 5. In the left figure of Fig. 5, except for the method ALICE when " + }, + { + "bbox": [ + 130, + 376, + 482, + 448 + ], + "type": "inline_equation", + "content": "\\alpha = 0" + }, + { + "bbox": [ + 130, + 376, + 482, + 448 + ], + "type": "text", + "content": ", our method outperforms most other methods by a clear margin (at least " + }, + { + "bbox": [ + 130, + 376, + 482, + 448 + ], + "type": "inline_equation", + "content": "1.27\\%" + }, + { + "bbox": [ + 130, + 376, + 482, + 448 + ], + "type": "text", + "content": ") and achieves the highest AUC of all methods. This indicates that our method achieves a more balanced base-novel performance. To see raw numbers of the accuracy of novel classes, please refer to the supplementary material." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 449, + 482, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 449, + 482, + 567 + ], + "spans": [ + { + "bbox": [ + 130, + 449, + 482, + 567 + ], + "type": "text", + "content": "Besides, Fig. 5 shows that methods that perform well at " + }, + { + "bbox": [ + 130, + 449, + 482, + 567 + ], + "type": "inline_equation", + "content": "\\alpha = 1" + }, + { + "bbox": [ + 130, + 449, + 482, + 567 + ], + "type": "text", + "content": " do not necessarily perform well when " + }, + { + "bbox": [ + 130, + 449, + 482, + 567 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 130, + 449, + 482, + 567 + ], + "type": "text", + "content": " is smaller. For example, although CLOM achieves " + }, + { + "bbox": [ + 130, + 449, + 482, + 567 + ], + "type": "inline_equation", + "content": "57.23\\%" + }, + { + "bbox": [ + 130, + 449, + 482, + 567 + ], + "type": "text", + "content": " average accuracy (" + }, + { + "bbox": [ + 130, + 449, + 482, + 567 + ], + "type": "inline_equation", + "content": "\\alpha = 1" + }, + { + "bbox": [ + 130, + 449, + 482, + 567 + ], + "type": "text", + "content": ") and is " + }, + { + "bbox": [ + 130, + 449, + 482, + 567 + ], + "type": "inline_equation", + "content": "1.83\\%" + }, + { + "bbox": [ + 130, + 449, + 482, + 567 + ], + "type": "text", + "content": " higher than method C-FSCIL (" + }, + { + "bbox": [ + 130, + 449, + 482, + 567 + ], + "type": "inline_equation", + "content": "55.4\\%" + }, + { + "bbox": [ + 130, + 449, + 482, + 567 + ], + "type": "text", + "content": "), we can see the performance drops significantly when " + }, + { + "bbox": [ + 130, + 449, + 482, + 567 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 130, + 449, + 482, + 567 + ], + "type": "text", + "content": " is small. It is mainly because CLOM focuses on the base-classes accuracy and performs poorly on novel classes. We can observe that the AUC of the CLOM is much smaller (" + }, + { + "bbox": [ + 130, + 449, + 482, + 567 + ], + "type": "inline_equation", + "content": "4.35\\%" + }, + { + "bbox": [ + 130, + 449, + 482, + 567 + ], + "type": "text", + "content": ") than that of C-FSCIL. The results confirm the necessity of our proposed " + }, + { + "bbox": [ + 130, + 449, + 482, + 567 + ], + "type": "inline_equation", + "content": "gAcc" + }, + { + "bbox": [ + 130, + 449, + 482, + 567 + ], + "type": "text", + "content": ". Similar phenomena can be found when comparing other methods like SAVC and ALICE, and we can obtain similar conclusions on the other two datasets including CUB-200 and CIFAR-100." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "text", + "content": "Results on CUB-200 dataset are shown in Table 2 and Fig. 6. We can see from Table 2 that our method is better than any other existing methods in terms of " + }, + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "inline_equation", + "content": "aAcc" + }, + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "inline_equation", + "content": "gAcc" + }, + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "text", + "content": " (1.52%@aAcc and 2.6%@gAcc improvement over the strong method SAVC). Results on CIFAR-100 dataset are shown in Table 3. For the traditional metric " + }, + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "inline_equation", + "content": "aAcc" + }, + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "text", + "content": ", our method achieves comparable performance with SOTA methods, while our method outperforms all other methods by a clear margin when considering the newly introduced metric " + }, + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "inline_equation", + "content": "gAcc" + }, + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "text", + "content": ". Please refer to Sec A5 of the supplementary material for more experiment results." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 212, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 212, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 212, + 102 + ], + "type": "text", + "content": "Tang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 156, + 140, + 455, + 277 + ], + "blocks": [ + { + "bbox": [ + 131, + 114, + 482, + 139 + ], + "lines": [ + { + "bbox": [ + 131, + 114, + 482, + 139 + ], + "spans": [ + { + "bbox": [ + 131, + 114, + 482, + 139 + ], + "type": "text", + "content": "Table 2: Comparison with FSCIL methods on CUB-200. The best and second best results are marked bolded and underlined. Methods with " + }, + { + "bbox": [ + 131, + 114, + 482, + 139 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 131, + 114, + 482, + 139 + ], + "type": "text", + "content": " are reported with the original results in their paper. Other results are reimplemented by us." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 156, + 140, + 455, + 277 + ], + "lines": [ + { + "bbox": [ + 156, + 140, + 455, + 277 + ], + "spans": [ + { + "bbox": [ + 156, + 140, + 455, + 277 + ], + "type": "table", + "html": "
MethodsNum. of Params.(M)aAcc in each task (%)
1234567891011aAccgAcc
TOPIC [52]†-68.742.554.850.045.341.438.435.432.228.326.343.92-
DF-Replay [35]†-75.972.168.663.862.659.157.855.955.053.652.461.52-
BiDistFSCIL [66]†-79.175.472.869.167.565.164.063.561.961.560.967.34-
MetaFSCIL [8]†-75.972.468.864.863.060.058.356.854.853.852.661.93-
SubNetwork [25]†-78.774.671.467.565.462.661.159.457.557.256.764.68-
DSN [58]†-76.172.269.666.764.462.160.258.957.055.154.263.31-
Self-promoted [73]12.3662.654.751.647.344.340.938.839.035.935.133.544.0740.88
CEC [63]12.3475.771.468.164.062.058.857.355.553.653.252.161.1754.16
LIMIT [70]12.3379.576.473.269.267.164.362.861.959.859.157.666.5859.40
CLOM [76]18.8979.575.972.068.265.763.462.060.358.557.956.965.6157.40
FACT [69]11.3476.173.470.466.564.862.561.159.957.957.556.550.2957.93
NC-FSCIL [59]15.9181.176.472.869.366.163.762.261.059.058.057.366.1759.22
ALICE [42]41.7175.870.168.465.263.761.260.759.758.558.758.463.7760.28
S3C [24]13.0680.276.572.969.166.964.362.761.059.559.258.366.5361.05
SAVC [49]24.2981.177.675.071.269.666.765.864.862.962.762.268.3362.90
YourSelf (Ours)11.9083.477.075.372.269.066.866.065.664.164.563.669.8565.50
", + "image_path": "d3d479cead14b913c3f2ed19d89771473caea4eefa1f1138a2ac7c4c0f3be90b.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 179, + 305, + 435, + 409 + ], + "blocks": [ + { + "bbox": [ + 131, + 280, + 482, + 304 + ], + "lines": [ + { + "bbox": [ + 131, + 280, + 482, + 304 + ], + "spans": [ + { + "bbox": [ + 131, + 280, + 482, + 304 + ], + "type": "text", + "content": "Table 3: Comparison results on CIFAR-100. The best and second best results are marked bolded and underlined. All results of other methods are reimplemented by us. See more results in the supplementary material." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 179, + 305, + 435, + 409 + ], + "lines": [ + { + "bbox": [ + 179, + 305, + 435, + 409 + ], + "spans": [ + { + "bbox": [ + 179, + 305, + 435, + 409 + ], + "type": "table", + "html": "
MethodsNum. of Params.(M)aAcc in each task (%)aAccgAcc
123456789
CEC [63]0.29673.269.165.461.258.055.553.251.349.259.5650.85
LIMIT [70]0.29676.171.767.363.260.157.355.253.050.861.6452.16
CLOM [76]0.35574.269.866.162.259.056.154.051.949.860.3550.36
C-FSCIL [20]12.7976.872.067.263.159.756.754.451.949.761.2850.62
SAVC [49]0.84278.472.768.464.061.258.356.054.151.762.7653.29
FACT [69]0.28278.572.768.864.561.358.656.854.452.363.1254.56
SubNet [25]22.3980.075.771.667.864.761.459.357.155.164.8556.63
S3C [24]0.34178.073.970.266.163.360.158.356.754.064.5158.04
ALICE [42]45.2880.371.967.063.260.758.357.355.553.763.1059.72
NC-FSCIL [59]15.9382.977.573.769.065.662.059.757.855.467.0659.73
Ours11.6882.976.372.967.865.262.060.758.856.667.0260.32
", + "image_path": "c7223bf47c26a2d7314510acf2df5b225cfc5f96e0206a91a1f11762077b975e.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 413, + 244, + 424 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 413, + 244, + 424 + ], + "spans": [ + { + "bbox": [ + 132, + 413, + 244, + 424 + ], + "type": "text", + "content": "5.3 Ablation studies" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 437, + 482, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 437, + 482, + 616 + ], + "spans": [ + { + "bbox": [ + 130, + 437, + 482, + 616 + ], + "type": "text", + "content": "Effectiveness of different losses. Ablation experiments of the effectiveness of " + }, + { + "bbox": [ + 130, + 437, + 482, + 616 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{IR}" + }, + { + "bbox": [ + 130, + 437, + 482, + 616 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 437, + 482, + 616 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{CR}" + }, + { + "bbox": [ + 130, + 437, + 482, + 616 + ], + "type": "text", + "content": " are conducted among CIFAR-100, miniImageNet, and CUB-200 datasets. Experimental results are shown in Table 4 and Table 5. The detailed " + }, + { + "bbox": [ + 130, + 437, + 482, + 616 + ], + "type": "inline_equation", + "content": "gAcc" + }, + { + "bbox": [ + 130, + 437, + 482, + 616 + ], + "type": "text", + "content": " vs. " + }, + { + "bbox": [ + 130, + 437, + 482, + 616 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 130, + 437, + 482, + 616 + ], + "type": "text", + "content": " as shown in Fig. 7. Results on CIFAR-100 show that introducing the FR module can effectively enhance the " + }, + { + "bbox": [ + 130, + 437, + 482, + 616 + ], + "type": "inline_equation", + "content": "gAcc" + }, + { + "bbox": [ + 130, + 437, + 482, + 616 + ], + "type": "text", + "content": " by " + }, + { + "bbox": [ + 130, + 437, + 482, + 616 + ], + "type": "inline_equation", + "content": "3.28\\%" + }, + { + "bbox": [ + 130, + 437, + 482, + 616 + ], + "type": "text", + "content": ". After further introducing the class-relation (CR) transfer loss, the performance at " + }, + { + "bbox": [ + 130, + 437, + 482, + 616 + ], + "type": "inline_equation", + "content": "gAcc" + }, + { + "bbox": [ + 130, + 437, + 482, + 616 + ], + "type": "text", + "content": " is further improved by " + }, + { + "bbox": [ + 130, + 437, + 482, + 616 + ], + "type": "inline_equation", + "content": "1.07\\%" + }, + { + "bbox": [ + 130, + 437, + 482, + 616 + ], + "type": "text", + "content": ". Adding the instance-relation (IR) transfer loss will further boost the " + }, + { + "bbox": [ + 130, + 437, + 482, + 616 + ], + "type": "inline_equation", + "content": "gAcc" + }, + { + "bbox": [ + 130, + 437, + 482, + 616 + ], + "type": "text", + "content": " by " + }, + { + "bbox": [ + 130, + 437, + 482, + 616 + ], + "type": "inline_equation", + "content": "1.91\\%" + }, + { + "bbox": [ + 130, + 437, + 482, + 616 + ], + "type": "text", + "content": ". The effectiveness of the proposed losses is also verified on the miniImageNet dataset. For miniImageNet dataset, the introduction of FR modules brings " + }, + { + "bbox": [ + 130, + 437, + 482, + 616 + ], + "type": "inline_equation", + "content": "3.02\\%" + }, + { + "bbox": [ + 130, + 437, + 482, + 616 + ], + "type": "text", + "content": " improvement at " + }, + { + "bbox": [ + 130, + 437, + 482, + 616 + ], + "type": "inline_equation", + "content": "gAcc" + }, + { + "bbox": [ + 130, + 437, + 482, + 616 + ], + "type": "text", + "content": ". Applying loss " + }, + { + "bbox": [ + 130, + 437, + 482, + 616 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{CR}" + }, + { + "bbox": [ + 130, + 437, + 482, + 616 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 437, + 482, + 616 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{IR}" + }, + { + "bbox": [ + 130, + 437, + 482, + 616 + ], + "type": "text", + "content": " can improve " + }, + { + "bbox": [ + 130, + 437, + 482, + 616 + ], + "type": "inline_equation", + "content": "gAcc" + }, + { + "bbox": [ + 130, + 437, + 482, + 616 + ], + "type": "text", + "content": " by " + }, + { + "bbox": [ + 130, + 437, + 482, + 616 + ], + "type": "inline_equation", + "content": "0.49\\%" + }, + { + "bbox": [ + 130, + 437, + 482, + 616 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 437, + 482, + 616 + ], + "type": "inline_equation", + "content": "0.82\\%" + }, + { + "bbox": [ + 130, + 437, + 482, + 616 + ], + "type": "text", + "content": ", respectively. It is worth mentioning that " + }, + { + "bbox": [ + 130, + 437, + 482, + 616 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{CR}" + }, + { + "bbox": [ + 130, + 437, + 482, + 616 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 437, + 482, + 616 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{IR}" + }, + { + "bbox": [ + 130, + 437, + 482, + 616 + ], + "type": "text", + "content": " can also boost the " + }, + { + "bbox": [ + 130, + 437, + 482, + 616 + ], + "type": "inline_equation", + "content": "aAcc" + }, + { + "bbox": [ + 130, + 437, + 482, + 616 + ], + "type": "text", + "content": ". Specifically, on the miniImageNet, adding these two losses results in an increase at " + }, + { + "bbox": [ + 130, + 437, + 482, + 616 + ], + "type": "inline_equation", + "content": "aAcc" + }, + { + "bbox": [ + 130, + 437, + 482, + 616 + ], + "type": "text", + "content": " performance of " + }, + { + "bbox": [ + 130, + 437, + 482, + 616 + ], + "type": "inline_equation", + "content": "1.96\\%" + }, + { + "bbox": [ + 130, + 437, + 482, + 616 + ], + "type": "text", + "content": ". On CIFAR-100, the " + }, + { + "bbox": [ + 130, + 437, + 482, + 616 + ], + "type": "inline_equation", + "content": "aAcc" + }, + { + "bbox": [ + 130, + 437, + 482, + 616 + ], + "type": "text", + "content": " is improved by about " + }, + { + "bbox": [ + 130, + 437, + 482, + 616 + ], + "type": "inline_equation", + "content": "0.76\\%" + }, + { + "bbox": [ + 130, + 437, + 482, + 616 + ], + "type": "text", + "content": " with the help of these two losses. See more experiments in the supplementary material." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "Effectiveness of multi-layer knowledge ensemble. In Sec. 4.2, we propose to exploit hierarchical knowledge from multiple layers for effective feature rectification. In practice, we select layers 8, 9 and 10 for the ensemble. Here, we take a more detailed discussion on each layer before the ensemble of multiple" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 150, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 150, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 150, + 91, + 448, + 102 + ], + "type": "text", + "content": "Rethinking Few-shot Class-incremental Learning: Learning from Yourself" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 122, + 141, + 338, + 178 + ], + "blocks": [ + { + "bbox": [ + 146, + 131, + 295, + 140 + ], + "lines": [ + { + "bbox": [ + 146, + 131, + 295, + 140 + ], + "spans": [ + { + "bbox": [ + 146, + 131, + 295, + 140 + ], + "type": "text", + "content": "Table 4: Ablation studies on CIFAR-100." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 122, + 141, + 338, + 178 + ], + "lines": [ + { + "bbox": [ + 122, + 141, + 338, + 178 + ], + "spans": [ + { + "bbox": [ + 122, + 141, + 338, + 178 + ], + "type": "table", + "html": "
FR\\( {\\mathcal{L}}_{\\text{Cos }} \\)\\( {\\mathcal{L}}_{\\text{NovCE }} \\)\\( {\\mathcal{L}}_{CR} \\)\\( {\\mathcal{L}}_{IR} \\)aAcc. in each task (%)aAccgAcc
123456789
82.977.872.968.464.661.458.756.454.166.3554.20
✓✓82.976.173.167.964.860.859.257.253.966.2657.48
✓✓✓82.976.172.467.864.561.359.657.354.966.3358.55
✓✓✓✓82.975.772.567.464.861.960.458.756.367.0260.32
", + "image_path": "0aeb4b0bb7c1962593998918dc5e97a214fb980670be95f7f6a52b71543accd2.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 122, + 189, + 338, + 228 + ], + "blocks": [ + { + "bbox": [ + 140, + 180, + 301, + 189 + ], + "lines": [ + { + "bbox": [ + 140, + 180, + 301, + 189 + ], + "spans": [ + { + "bbox": [ + 140, + 180, + 301, + 189 + ], + "type": "text", + "content": "Table 5: Ablation studies on miniImageNet." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 122, + 189, + 338, + 228 + ], + "lines": [ + { + "bbox": [ + 122, + 189, + 338, + 228 + ], + "spans": [ + { + "bbox": [ + 122, + 189, + 338, + 228 + ], + "type": "table", + "html": "
FRLcos+ LNoveCELCRLIRaAcc. in each task (%)aAccgAcc
123456789
84.079.475.071.167.864.461.358.856.968.7457.87
✓✓84.074.069.766.866.163.560.858.658.166.8460.89
✓✓✓84.077.773.669.967.464.661.659.858.768.6061.38
✓✓✓✓84.077.673.770.068.064.962.159.859.068.8062.20
", + "image_path": "21811813346d39895fb7a37b939ab0a5be4e674ab8f6be0dd14cf5e6db85d71d.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 339, + 129, + 417, + 218 + ], + "blocks": [ + { + "bbox": [ + 339, + 129, + 417, + 218 + ], + "lines": [ + { + "bbox": [ + 339, + 129, + 417, + 218 + ], + "spans": [ + { + "bbox": [ + 339, + 129, + 417, + 218 + ], + "type": "image", + "image_path": "692d5722960dfe2f8914e6cd436a9210d4d5f02e817c8618a93feae34e9307d7.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 339, + 219, + 496, + 236 + ], + "lines": [ + { + "bbox": [ + 339, + 219, + 496, + 236 + ], + "spans": [ + { + "bbox": [ + 339, + 219, + 496, + 236 + ], + "type": "text", + "content": "Fig.7: Ablation studies on two datasets, evaluating using " + }, + { + "bbox": [ + 339, + 219, + 496, + 236 + ], + "type": "inline_equation", + "content": "gAcc" + }, + { + "bbox": [ + 339, + 219, + 496, + 236 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 417, + 129, + 495, + 218 + ], + "blocks": [ + { + "bbox": [ + 417, + 129, + 495, + 218 + ], + "lines": [ + { + "bbox": [ + 417, + 129, + 495, + 218 + ], + "spans": [ + { + "bbox": [ + 417, + 129, + 495, + 218 + ], + "type": "image", + "image_path": "766a7f88dc627d7e4edf055268d6d7bd618febf9dd6f7f39e729e2290995840e.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 242, + 482, + 326 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 242, + 482, + 326 + ], + "spans": [ + { + "bbox": [ + 130, + 242, + 482, + 326 + ], + "type": "text", + "content": "layers. We conducted experiments on the miniImageNet, and the results are shown in Table 6. It is clear that for each layer, the corresponding FR module effectively calibrates the features and achieves improvement from " + }, + { + "bbox": [ + 130, + 242, + 482, + 326 + ], + "type": "inline_equation", + "content": "2.23\\%" + }, + { + "bbox": [ + 130, + 242, + 482, + 326 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 130, + 242, + 482, + 326 + ], + "type": "inline_equation", + "content": "3.54\\%" + }, + { + "bbox": [ + 130, + 242, + 482, + 326 + ], + "type": "text", + "content": " at " + }, + { + "bbox": [ + 130, + 242, + 482, + 326 + ], + "type": "inline_equation", + "content": "gAcc" + }, + { + "bbox": [ + 130, + 242, + 482, + 326 + ], + "type": "text", + "content": ". Furthermore, the introduction of the FR module does not harm the " + }, + { + "bbox": [ + 130, + 242, + 482, + 326 + ], + "type": "inline_equation", + "content": "aAcc" + }, + { + "bbox": [ + 130, + 242, + 482, + 326 + ], + "type": "text", + "content": ". Considering the fact that different layers contain hierarchical knowledge, the multi-layer ensemble is proposed. The results in the table show that the ensemble is effective in improving the " + }, + { + "bbox": [ + 130, + 242, + 482, + 326 + ], + "type": "inline_equation", + "content": "gAcc" + }, + { + "bbox": [ + 130, + 242, + 482, + 326 + ], + "type": "text", + "content": " while maintaining " + }, + { + "bbox": [ + 130, + 242, + 482, + 326 + ], + "type": "inline_equation", + "content": "aAcc" + }, + { + "bbox": [ + 130, + 242, + 482, + 326 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 187, + 388, + 427, + 472 + ], + "blocks": [ + { + "bbox": [ + 130, + 346, + 482, + 367 + ], + "lines": [ + { + "bbox": [ + 130, + 346, + 482, + 367 + ], + "spans": [ + { + "bbox": [ + 130, + 346, + 482, + 367 + ], + "type": "text", + "content": "Table 6: Detailed performance of each layer. Results are from miniImagenet dataset. 'l-Raw': vanilla ViT features from layer " + }, + { + "bbox": [ + 130, + 346, + 482, + 367 + ], + "type": "inline_equation", + "content": "L_{l}" + }, + { + "bbox": [ + 130, + 346, + 482, + 367 + ], + "type": "text", + "content": ". 'l-Recti': Rectified features using the proposed FR module." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 135, + 369, + 347, + 379 + ], + "lines": [ + { + "bbox": [ + 135, + 369, + 347, + 379 + ], + "spans": [ + { + "bbox": [ + 135, + 369, + 347, + 379 + ], + "type": "text", + "content": "'Ensemble': Ensembled results from " + }, + { + "bbox": [ + 135, + 369, + 347, + 379 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 135, + 369, + 347, + 379 + ], + "type": "text", + "content": "-Recti', " + }, + { + "bbox": [ + 135, + 369, + 347, + 379 + ], + "type": "inline_equation", + "content": "l \\in \\{8, 9, 10\\}" + }, + { + "bbox": [ + 135, + 369, + 347, + 379 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 187, + 388, + 427, + 472 + ], + "lines": [ + { + "bbox": [ + 187, + 388, + 427, + 472 + ], + "spans": [ + { + "bbox": [ + 187, + 388, + 427, + 472 + ], + "type": "table", + "html": "
LayersaAcc in each session (%)aAccgAcc
012345678
8-Raw78.973.269.165.462.960.057.355.254.064.0156.61
8-Recti80.773.470.166.664.461.559.357.056.565.5260.15
9-raw82.577.373.269.765.363.760.958.957.567.6659.33
9-Recti83.076.673.269.767.564.661.959.959.068.3961.90
10-raw83.878.674.470.766.464.361.559.357.768.5558.90
10-Recti84.378.274.170.468.264.861.959.558.468.8461.13
Ensemble84.077.673.770.068.064.962.159.859.068.8062.20
", + "image_path": "0559a85061f738971917e2db5cff81e042152d9cd6ca5e4cb2e147b3d47175e8.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 517, + 220, + 530 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 517, + 220, + 530 + ], + "spans": [ + { + "bbox": [ + 132, + 517, + 220, + 530 + ], + "type": "text", + "content": "6 Conclusion" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": "In this work, we have revisited the challenges of the FSCIL, showing the limitation of the existing metric and introducing a novel metric denoted as generalized accuracy (" + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "inline_equation", + "content": "gAcc" + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": ") for the comprehensive evaluation of FSCIL methods from a more balanced perspective. Furthermore, we incorporate the ViT into the FSCIL framework and unlock the potential of the ViT intermediate layer features. To achieve this, we have designed a feature rectification (FR) module aimed at integrating the valuable information from the intermediate layer features into the rectified feature space. Using the new metric " + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "inline_equation", + "content": "gAcc" + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": ", we analyze the existing open-sourced FSCIL methods from a more balanced perspective and show the superiority of our method on three popular datasets." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 212, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 212, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 212, + 102 + ], + "type": "text", + "content": "Tang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 115, + 246, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 115, + 246, + 129 + ], + "spans": [ + { + "bbox": [ + 133, + 115, + 246, + 129 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 141, + 482, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 141, + 482, + 190 + ], + "spans": [ + { + "bbox": [ + 130, + 141, + 482, + 190 + ], + "type": "text", + "content": "This work was supported partially by the National Key Research and Development Program of China (2023YFA1008503), NSFC(U21A20471, 62206315), Guangdong NSF Project (No. 2023B1515040025, 2020B1515120085, 2024A15150-10101), Guangzhou Basic and Applied Basic Research Scheme(2024A04J4067)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 209, + 197, + 220 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 209, + 197, + 220 + ], + "spans": [ + { + "bbox": [ + 132, + 209, + 197, + 220 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 234, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 138, + 234, + 480, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 234, + 480, + 267 + ], + "spans": [ + { + "bbox": [ + 138, + 234, + 480, + 267 + ], + "type": "text", + "content": "1. Agarwal, A., Banerjee, B., Cuzzolin, F., Chaudhuri, S.: Semantics-driven generative replay for few-shot class incremental learning. In: Proceedings of the 30th ACM International Conference on Multimedia (2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 269, + 480, + 311 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 269, + 480, + 311 + ], + "spans": [ + { + "bbox": [ + 138, + 269, + 480, + 311 + ], + "type": "text", + "content": "2. Ahmad, T., Dhamija, A.R., Cruz, S., Rabinowitz, R., Li, C., Jafarzadeh, M., Boult, T.E.: Few-shot class incremental learning leveraging self-supervised features. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2022)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 312, + 480, + 344 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 312, + 480, + 344 + ], + "spans": [ + { + "bbox": [ + 138, + 312, + 480, + 344 + ], + "type": "text", + "content": "3. Akyurek, A.F., Akyurek, E., Wijaya, D.T., Andreas, J.: Subspace regularizers for few-shot class incremental learning. In: International Conference on Learning Representations (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 346, + 480, + 378 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 346, + 480, + 378 + ], + "spans": [ + { + "bbox": [ + 138, + 346, + 480, + 378 + ], + "type": "text", + "content": "4. Castro, F.M., Marín-Jiménez, M.J., Guil, N., Schmid, C., Alahari, K.: End-to-end incremental learning. In: Proceedings of the European Conference on Computer Vision (2018)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 379, + 480, + 400 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 379, + 480, + 400 + ], + "spans": [ + { + "bbox": [ + 138, + 379, + 480, + 400 + ], + "type": "text", + "content": "5. Chen, H., Wang, Y., Xu, C., Xu, C., Tao, D.: Learning student networks via feature embedding. IEEE Transactions on Neural Networks and Learning Systems (2020)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 401, + 480, + 422 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 401, + 480, + 422 + ], + "spans": [ + { + "bbox": [ + 138, + 401, + 480, + 422 + ], + "type": "text", + "content": "6. Chen, K., Lee, C.G.: Incremental few-shot learning via vector quantization in deep embedded space. In: International Conference on Learning Representations (2020)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 423, + 480, + 466 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 423, + 480, + 466 + ], + "spans": [ + { + "bbox": [ + 138, + 423, + 480, + 466 + ], + "type": "text", + "content": "7. Cheraghian, A., Rahman, S., Fang, P., Roy, S.K., Petersson, L., Harandi, M.: Semantic-aware knowledge distillation for few-shot class-incremental learning. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition (2021)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 467, + 480, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 467, + 480, + 499 + ], + "spans": [ + { + "bbox": [ + 138, + 467, + 480, + 499 + ], + "type": "text", + "content": "8. Chi, Z., Gu, L., Liu, H., Wang, Y., Yu, Y., Tang, J.: Meta-scil: A meta-learning approach for few-shot class incremental learning. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition (2022)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 501, + 480, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 501, + 480, + 533 + ], + "spans": [ + { + "bbox": [ + 138, + 501, + 480, + 533 + ], + "type": "text", + "content": "9. Cui, Y., Deng, W., Chen, H., Liu, L.: Uncertainty-aware distillation for semi-supervised few-shot class-incremental learning. IEEE Transactions on Neural Networks and Learning Systems (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 533, + 480, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 533, + 480, + 566 + ], + "spans": [ + { + "bbox": [ + 138, + 533, + 480, + 566 + ], + "type": "text", + "content": "0. Cui, Y., Deng, W., Xu, X., Liu, Z., Liu, Z., Pietikainen, M., Liu, L.: Uncertainty-guided semi-supervised few-shot class-incremental learning with knowledge distillation. IEEE Transactions on Multimedia (2022)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 567, + 480, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 567, + 480, + 599 + ], + "spans": [ + { + "bbox": [ + 138, + 567, + 480, + 599 + ], + "type": "text", + "content": "1. Cui, Y., Xiong, W., Tavakolian, M., Liu, L.: Semi-supervised few-shot class-incremental learning. In: IEEE International Conference on Image Processing (ICIP) (2021)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 600, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 600, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 138, + 600, + 480, + 632 + ], + "type": "text", + "content": "2. Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: Imagenet: A large-scale hierarchical image database. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2009)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 632, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 632, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 632, + 480, + 665 + ], + "type": "text", + "content": "3. Deng, Y., Xiang, X.: Expanding hyperspherical space for few-shot class-incremental learning. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV) (2024)" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 150, + 91, + 449, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 150, + 91, + 449, + 102 + ], + "spans": [ + { + "bbox": [ + 150, + 91, + 449, + 102 + ], + "type": "text", + "content": "Rethinking Few-shot Class-incremental Learning: Learning from Yourself" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 481, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "type": "text", + "content": "14. Dong, S., Hong, X., Tao, X., Chang, X., Wei, X., Gong, Y.: Few-shot class-incremental learning via relation knowledge distillation. In: Proceedings of the AAAI Conference on Artificial Intelligence (2021)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 150, + 481, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 150, + 481, + 194 + ], + "spans": [ + { + "bbox": [ + 133, + 150, + 481, + 194 + ], + "type": "text", + "content": "15. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. In: International Conference on Learning Representations (2021)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 195, + 481, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 195, + 481, + 228 + ], + "spans": [ + { + "bbox": [ + 132, + 195, + 481, + 228 + ], + "type": "text", + "content": "16. Douillard, A., Cord, M., Ollion, C., Robert, T., Valle, E.: Podnet: Pooled outputs distillation for small-tasks incremental learning. In: Proceedings of the European Conference on Computer Vision (2020)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 228, + 481, + 261 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 228, + 481, + 261 + ], + "spans": [ + { + "bbox": [ + 132, + 228, + 481, + 261 + ], + "type": "text", + "content": "17. Fu, S., Yan, J., Gao, Y., Xie, X., Zheng, W.S.: Asag: Building strong one-decoder-layer sparse detectors via adaptive sparse anchor generation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 6328-6338 (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 262, + 481, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 262, + 481, + 285 + ], + "spans": [ + { + "bbox": [ + 132, + 262, + 481, + 285 + ], + "type": "text", + "content": "18. Gu, Z., Xu, C., Yang, J., Cui, Z.: Few-shot continual infomax learning. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 285, + 481, + 317 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 285, + 481, + 317 + ], + "spans": [ + { + "bbox": [ + 132, + 285, + 481, + 317 + ], + "type": "text", + "content": "19. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2016)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 318, + 481, + 351 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 318, + 481, + 351 + ], + "spans": [ + { + "bbox": [ + 132, + 318, + 481, + 351 + ], + "type": "text", + "content": "20. Hersche, M., Karunaratne, G., Cherubini, G., Benini, L., Sebastian, A., Rahimi, A.: Constrained few-shot class-incremental learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2022)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 352, + 481, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 352, + 481, + 385 + ], + "spans": [ + { + "bbox": [ + 132, + 352, + 481, + 385 + ], + "type": "text", + "content": "21. Hou, S., Pan, X., Loy, C.C., Wang, Z., Lin, D.: Learning a unified classifier incrementally via rebalancing. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2019)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 386, + 481, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 386, + 481, + 418 + ], + "spans": [ + { + "bbox": [ + 132, + 386, + 481, + 418 + ], + "type": "text", + "content": "22. Ji, Z., Hou, Z., Liu, X., Pang, Y., Li, X.: Memorizing complementation network for few-shot class-incremental learning. IEEE Transactions on Image Processing (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 419, + 481, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 419, + 481, + 441 + ], + "spans": [ + { + "bbox": [ + 132, + 419, + 481, + 441 + ], + "type": "text", + "content": "23. Jiang, W., Li, D., Hu, M., Zhai, G., Yang, X., Zhang, X.P.: Few-shot class-incremental learning with prior knowledge. arXiv preprint arXiv:2402.01201 (2024)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 442, + 481, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 442, + 481, + 464 + ], + "spans": [ + { + "bbox": [ + 132, + 442, + 481, + 464 + ], + "type": "text", + "content": "24. Kalla, J., Biswas, S.: S3c: Self-supervised stochastic classifiers for few-shot class-incremental learning. In: European Conference on Computer Vision (2022)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 464, + 481, + 497 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 464, + 481, + 497 + ], + "spans": [ + { + "bbox": [ + 132, + 464, + 481, + 497 + ], + "type": "text", + "content": "25. Kang, H., Yoon, J., Madjid, S.R.H., Hwang, S.J., Yoo, C.D.: On the soft-subnetwork for few-shot class incremental learning. In: International Conference on Learning Representations (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 498, + 481, + 530 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 498, + 481, + 530 + ], + "spans": [ + { + "bbox": [ + 132, + 498, + 481, + 530 + ], + "type": "text", + "content": "26. Kim, D.Y., Han, D.J., Seo, J., Moon, J.: Warping the space: Weight space rotation for class-incremental few-shot learning. In: International Conference on Learning Representations (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 531, + 481, + 575 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 531, + 481, + 575 + ], + "spans": [ + { + "bbox": [ + 132, + 531, + 481, + 575 + ], + "type": "text", + "content": "27. Kim, S., Jeong, Y., Park, J.S., Yoon, S.W.: Mics: Midpoint interpolation to learn compact and separated representations for few-shot class-incremental learning. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (2024)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 576, + 481, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 576, + 481, + 597 + ], + "spans": [ + { + "bbox": [ + 132, + 576, + 481, + 597 + ], + "type": "text", + "content": "28. Krizhevsky, A., Hinton, G., et al.: Learning multiple layers of features from tiny images. Technical report, University of Toronto (2009)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 599, + 481, + 631 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 599, + 481, + 631 + ], + "spans": [ + { + "bbox": [ + 132, + 599, + 481, + 631 + ], + "type": "text", + "content": "29. Lee, K., Lee, K., Shin, J., Lee, H.: Overcoming catastrophic forgetting with unlabeled data in the wild. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2019)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 632, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 481, + 665 + ], + "type": "text", + "content": "30. Li, J., Dong, S., Gong, Y., He, Y., Wei, X.: Analogical learning-based few-shot class-incremental learning. IEEE Transactions on Circuits and Systems for Video Technology (2024)" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 211, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 211, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 211, + 101 + ], + "type": "text", + "content": "Tang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 666 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 150 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 150 + ], + "type": "text", + "content": "31. Li, Y.M., Huang, W.J., Wang, A.L., Zeng, L.A., Meng, J.K., Zheng, W.S.: Egoexofitness: Towards egocentric and exocentric full-body action understanding. arXiv preprint arXiv:2406.08877 (2024)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 150, + 482, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 150, + 482, + 184 + ], + "spans": [ + { + "bbox": [ + 130, + 150, + 482, + 184 + ], + "type": "text", + "content": "32. Li, Y.M., Zeng, L.A., Meng, J.K., Zheng, W.S.: Continual action assessment via task-consistent score-discriminative feature distribution modeling. IEEE Transactions on Circuits and Systems for Video Technology (2024)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 184, + 482, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 184, + 482, + 206 + ], + "spans": [ + { + "bbox": [ + 130, + 184, + 482, + 206 + ], + "type": "text", + "content": "33. Li, Z., Hoiem, D.: Learning without forgetting. IEEE transactions on pattern analysis and machine intelligence (2017)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 206, + 482, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 206, + 482, + 239 + ], + "spans": [ + { + "bbox": [ + 130, + 206, + 482, + 239 + ], + "type": "text", + "content": "34. Lin, K.Y., Ding, H., Zhou, J., Peng, Y.X., Zhao, Z., Loy, C.C., Zheng, W.S.: Rethinking clip-based video learners in cross-domain open-vocabulary action recognition. arXiv preprint arXiv:2403.01560 (2024)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 239, + 482, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 239, + 482, + 274 + ], + "spans": [ + { + "bbox": [ + 130, + 239, + 482, + 274 + ], + "type": "text", + "content": "35. Liu, H., Gu, L., Chi, Z., Wang, Y., Yu, Y., Chen, J., Tang, J.: Few-shot class-incremental learning via entropy-regularized data-free replay. In: European Conference on Computer Vision (2022)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 274, + 482, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 274, + 482, + 308 + ], + "spans": [ + { + "bbox": [ + 130, + 274, + 482, + 308 + ], + "type": "text", + "content": "36. Liu, S., Qi, L., Qin, H., Shi, J., Jia, J.: Path aggregation network for instance segmentation. In: Proceedings of the IEEE conference on computer vision and pattern recognition (2018)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 308, + 482, + 329 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 308, + 482, + 329 + ], + "spans": [ + { + "bbox": [ + 130, + 308, + 482, + 329 + ], + "type": "text", + "content": "37. Van der Maaten, L., Hinton, G.: Visualizing data using t-sne. Journal of machine learning research (2008)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 329, + 482, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 329, + 482, + 352 + ], + "spans": [ + { + "bbox": [ + 130, + 329, + 482, + 352 + ], + "type": "text", + "content": "38. Mazumder, P., Singh, P., Rai, P.: Few-shot lifelong learning. In: Proceedings of the AAAI Conference on Artificial Intelligence (2021)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 352, + 482, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 352, + 482, + 386 + ], + "spans": [ + { + "bbox": [ + 130, + 352, + 482, + 386 + ], + "type": "text", + "content": "39. Pan, Z., Yu, X., Zhang, M., Gao, Y.: Ssfe-net: Self-supervised feature enhancement for ultra-fine-grained few-shot class incremental learning. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 386, + 482, + 419 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 386, + 482, + 419 + ], + "spans": [ + { + "bbox": [ + 130, + 386, + 482, + 419 + ], + "type": "text", + "content": "40. Park, W., Kim, D., Lu, Y., Cho, M.: Relational knowledge distillation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition (2019)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 419, + 482, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 419, + 482, + 453 + ], + "spans": [ + { + "bbox": [ + 130, + 419, + 482, + 453 + ], + "type": "text", + "content": "41. Passalis, N., Tzelepi, M., Tefas, A.: Heterogeneous knowledge distillation using information flow modeling. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2020)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 453, + 482, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 453, + 482, + 487 + ], + "spans": [ + { + "bbox": [ + 130, + 453, + 482, + 487 + ], + "type": "text", + "content": "42. Peng, C., Zhao, K., Wang, T., Li, M., Lovell, B.C.: Few-shot class-incremental learning from an open-set perspective. In: European Conference on Computer Vision (2022)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 487, + 482, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 487, + 482, + 520 + ], + "spans": [ + { + "bbox": [ + 130, + 487, + 482, + 520 + ], + "type": "text", + "content": "43. Peng, Y.X., Jiao, J., Feng, X., Zheng, W.S.: Consistent discrepancy learning for intra-camera supervised person re-identification. IEEE Transactions on Multimedia 25, 2393–2403 (2022)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 520, + 482, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 520, + 482, + 564 + ], + "spans": [ + { + "bbox": [ + 130, + 520, + 482, + 564 + ], + "type": "text", + "content": "44. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 564, + 482, + 598 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 564, + 482, + 598 + ], + "spans": [ + { + "bbox": [ + 130, + 564, + 482, + 598 + ], + "type": "text", + "content": "45. Rebuffi, S.A., Kolesnikov, A., Sperl, G., Lampert, C.H.: icarl: Incremental classifier and representation learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2017)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 130, + 598, + 482, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 598, + 482, + 632 + ], + "spans": [ + { + "bbox": [ + 130, + 598, + 482, + 632 + ], + "type": "text", + "content": "46. Russakovsky, O., Deng, J., Su, H., Krause, J., Satheesh, S., Ma, S., Huang, Z., Karpathy, A., Khosla, A., Bernstein, M., et al.: Imagenet large scale visual recognition challenge. International journal of computer vision (2015)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 130, + 632, + 482, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 632, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 632, + 482, + 666 + ], + "type": "text", + "content": "47. Shi, G., Chen, J., Zhang, W., Zhan, L.M., Wu, X.M.: Overcoming catastrophic forgetting in incremental few-shot learning by finding flat minima. In: Advances in neural information processing systems (2021)" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 150, + 91, + 449, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 150, + 91, + 449, + 102 + ], + "spans": [ + { + "bbox": [ + 150, + 91, + 449, + 102 + ], + "type": "text", + "content": "Rethinking Few-shot Class-incremental Learning: Learning from Yourself" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "type": "text", + "content": "48. Smith, J., Hsu, Y.C., Balloch, J., Shen, Y., Jin, H., Kira, Z.: Always be dreaming: A new approach for data-free class-incremental learning. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (2021)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 150, + 480, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 150, + 480, + 194 + ], + "spans": [ + { + "bbox": [ + 133, + 150, + 480, + 194 + ], + "type": "text", + "content": "49. Song, Z., Zhao, Y., Shi, Y., Peng, P., Yuan, L., Tian, Y.: Learning with fantasy: Semantic-aware virtual contrastive constraint for few-shot class-incremental learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 194, + 480, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 194, + 480, + 226 + ], + "spans": [ + { + "bbox": [ + 132, + 194, + 480, + 226 + ], + "type": "text", + "content": "50. Tang, Y.M., Peng, Y.X., Zheng, W.S.: Learning to imagine: Diversify memory for incremental learning using unlabeled data. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2022)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 227, + 480, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 227, + 480, + 258 + ], + "spans": [ + { + "bbox": [ + 132, + 227, + 480, + 258 + ], + "type": "text", + "content": "51. Tang, Y.M., Peng, Y.X., Zheng, W.S.: When prompt-based incremental learning does not meet strong pretraining. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 259, + 480, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 259, + 480, + 293 + ], + "spans": [ + { + "bbox": [ + 132, + 259, + 480, + 293 + ], + "type": "text", + "content": "52. Tao, X., Hong, X., Chang, X., Dong, S., Wei, X., Gong, Y.: Few-shot class incremental learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2020)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 293, + 480, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 293, + 480, + 326 + ], + "spans": [ + { + "bbox": [ + 132, + 293, + 480, + 326 + ], + "type": "text", + "content": "53. Touvron, H., Cord, M., Douze, M., Massa, F., Sablayrolles, A., Jégou, H.: Training data-efficient image transformers & distillation through attention. In: International Conference on Machine Learning (2021)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 326, + 480, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 326, + 480, + 357 + ], + "spans": [ + { + "bbox": [ + 132, + 326, + 480, + 357 + ], + "type": "text", + "content": "54. Vaze, S., Han, K., Vedaldi, A., Zisserman, A.: Open-set recognition: A good closed-set classifier is all you need? In: International Conference on Learning Representations (2021)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 358, + 480, + 379 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 358, + 480, + 379 + ], + "spans": [ + { + "bbox": [ + 132, + 358, + 480, + 379 + ], + "type": "text", + "content": "55. Wah, C., Branson, S., Welinder, P., Perona, P., Belongie, S.: The caltech-ucsd birds-200-2011 dataset (2011)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 380, + 480, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 380, + 480, + 413 + ], + "spans": [ + { + "bbox": [ + 132, + 380, + 480, + 413 + ], + "type": "text", + "content": "56. Wang, Q.W., Zhou, D.W., Zhang, Y.K., Zhan, D.C., Ye, H.J.: Few-shot class-incremental learning via training-free prototype calibration. In: Advances in Neural Information Processing Systems (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 414, + 480, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 414, + 480, + 445 + ], + "spans": [ + { + "bbox": [ + 132, + 414, + 480, + 445 + ], + "type": "text", + "content": "57. Yan, S., Xie, J., He, X.: Der: Dynamically expandable representation for class incremental learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2021)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 447, + 480, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 447, + 480, + 479 + ], + "spans": [ + { + "bbox": [ + 132, + 447, + 480, + 479 + ], + "type": "text", + "content": "58. Yang, B., Lin, M., Zhang, Y., Liu, B., Liang, X., Ji, R., Ye, Q.: Dynamic support network for few-shot class incremental learning. IEEE Transactions on Pattern Analysis and Machine Intelligence (2022)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 479, + 480, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 479, + 480, + 511 + ], + "spans": [ + { + "bbox": [ + 132, + 479, + 480, + 511 + ], + "type": "text", + "content": "59. Yang, Y., Yuan, H., Li, X., Lin, Z., Torr, P., Tao, D.: Neural collapse inspired feature-classifier alignment for few-shot class incremental learning. In: International Conference on Learning Representations (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 512, + 480, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 512, + 480, + 533 + ], + "spans": [ + { + "bbox": [ + 132, + 512, + 480, + 533 + ], + "type": "text", + "content": "60. Yao, G., Zhu, J., Zhou, W., Li, J.: Few-shot class-incremental learning based on representation enhancement. Journal of Electronic Imaging (2022)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 534, + 480, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 534, + 480, + 566 + ], + "spans": [ + { + "bbox": [ + 132, + 534, + 480, + 566 + ], + "type": "text", + "content": "61. Yu, L., Twardowski, B., Liu, X., Herranz, L., Wang, K., Cheng, Y., Jui, S., Weijer, J.v.d.: Semantic drift compensation for class-incremental learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2020)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 567, + 480, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 567, + 480, + 599 + ], + "spans": [ + { + "bbox": [ + 132, + 567, + 480, + 599 + ], + "type": "text", + "content": "62. Yuan, Y., Fu, R., Huang, L., Lin, W., Zhang, C., Chen, X., Wang, J.: Hrformer: High-resolution vision transformer for dense predict. Advances in Neural Information Processing Systems (2021)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 600, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 600, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 600, + 480, + 632 + ], + "type": "text", + "content": "63. Zhang, C., Song, N., Lin, G., Zheng, Y., Pan, P., Xu, Y.: Few-shot incremental learning with continually evolved classifiers. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition (2021)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "type": "text", + "content": "64. Zhang, J., Zhang, J., Ghosh, S., Li, D., Tasci, S., Heck, L., Zhang, H., Kuo, C.C.J.: Class-incremental learning via deep model consolidation. In: Proceedings of the IEEE Winter Conference on Applications of Computer Vision (2020)" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 211, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 211, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 211, + 102 + ], + "type": "text", + "content": "Tang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 501 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 480, + 149 + ], + "type": "text", + "content": "65. Zhao, B., Xiao, X., Gan, G., Zhang, B., Xia, S.T.: Maintaining discrimination and fairness in class incremental learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2020)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 150, + 480, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 150, + 480, + 183 + ], + "spans": [ + { + "bbox": [ + 132, + 150, + 480, + 183 + ], + "type": "text", + "content": "66. Zhao, L., Lu, J., Xu, Y., Cheng, Z., Guo, D., Niu, Y., Fang, X.: Few-shot class-incremental learning via class-aware bilateral distillation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 183, + 480, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 183, + 480, + 215 + ], + "spans": [ + { + "bbox": [ + 132, + 183, + 480, + 215 + ], + "type": "text", + "content": "67. Zheng, G., Zhang, A.: Few-shot class-incremental learning with meta-learned class structures. In: International Conference on Data Mining Workshops (ICDMW) (2021)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 216, + 480, + 247 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 216, + 480, + 247 + ], + "spans": [ + { + "bbox": [ + 132, + 216, + 480, + 247 + ], + "type": "text", + "content": "68. Zheng, W.S., Yan, J., Peng, Y.X.: A versatile framework for multi-scene person re-identification. IEEE Transactions on Pattern Analysis and Machine Intelligence (2024)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 249, + 480, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 249, + 480, + 281 + ], + "spans": [ + { + "bbox": [ + 132, + 249, + 480, + 281 + ], + "type": "text", + "content": "69. Zhou, D.W., Wang, F.Y., Ye, H.J., Ma, L., Pu, S., Zhan, D.C.: Forward compatible few-shot class-incremental learning. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition (2022)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 282, + 480, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 282, + 480, + 314 + ], + "spans": [ + { + "bbox": [ + 132, + 282, + 480, + 314 + ], + "type": "text", + "content": "70. Zhou, D.W., Ye, H.J., Ma, L., Xie, D., Pu, S., Zhan, D.C.: Few-shot class-incremental learning by sampling multi-phase tasks. IEEE Transactions on Pattern Analysis and Machine Intelligence (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 315, + 480, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 315, + 480, + 346 + ], + "spans": [ + { + "bbox": [ + 132, + 315, + 480, + 346 + ], + "type": "text", + "content": "71. Zhu, F., Zhang, X.Y., Wang, C., Yin, F., Liu, C.L.: Prototype augmentation and self-supervision for incremental learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2021)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 347, + 480, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 347, + 480, + 380 + ], + "spans": [ + { + "bbox": [ + 132, + 347, + 480, + 380 + ], + "type": "text", + "content": "72. Zhu, J., Yao, G., Zhou, W., Zhang, G., Ping, W., Zhang, W.: Feature distribution distillation-based few shot class incremental learning. In: the 5th International Conference on Pattern Recognition and Artificial Intelligence (PRAI) (2022)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 380, + 480, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 380, + 480, + 413 + ], + "spans": [ + { + "bbox": [ + 132, + 380, + 480, + 413 + ], + "type": "text", + "content": "73. Zhu, K., Cao, Y., Zhai, W., Cheng, J., Zha, Z.J.: Self-promoted prototype refinement for few-shot class-incremental learning. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition (2021)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 414, + 480, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 414, + 480, + 445 + ], + "spans": [ + { + "bbox": [ + 132, + 414, + 480, + 445 + ], + "type": "text", + "content": "74. Zhu, K., Zhai, W., Cao, Y., Luo, J., Zha, Z.J.: Self-sustaining representation expansion for non-exemplar class-incremental learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2022)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 447, + 480, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 447, + 480, + 468 + ], + "spans": [ + { + "bbox": [ + 132, + 447, + 480, + 468 + ], + "type": "text", + "content": "75. Zhu, M., Zhu, Z., Chen, S., Chen, C., Wu, B.: Enhanced few-shot class-incremental learning via ensemble models. arXiv preprint arXiv:2401.07208 (2024)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 468, + 480, + 501 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 468, + 480, + 501 + ], + "spans": [ + { + "bbox": [ + 132, + 468, + 480, + 501 + ], + "type": "text", + "content": "76. Zou, Y., Zhang, S., Li, Y., Li, R.: Margin-based few-shot class-incremental learning with class-level overfitting mitigation. In: Advances in neural information processing systems (2022)" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 150, + 91, + 449, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 150, + 91, + 449, + 102 + ], + "spans": [ + { + "bbox": [ + 150, + 91, + 449, + 102 + ], + "type": "text", + "content": "Rethinking Few-shot Class-incremental Learning: Learning from Yourself" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "19" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 18 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Rethinking Image Super Resolution from Training Data Perspectives/8d3fa075-356f-4f4c-b1a5-4318bed20e1d_content_list.json b/2024/Rethinking Image Super Resolution from Training Data Perspectives/8d3fa075-356f-4f4c-b1a5-4318bed20e1d_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..ce7aea2436b1b631a8110c83b2310fab06be0e02 --- /dev/null +++ b/2024/Rethinking Image Super Resolution from Training Data Perspectives/8d3fa075-356f-4f4c-b1a5-4318bed20e1d_content_list.json @@ -0,0 +1,1755 @@ +[ + { + "type": "text", + "text": "Rethinking Image Super-Resolution from Training Data Perspectives", + "text_level": 1, + "bbox": [ + 266, + 140, + 736, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Go Ohtani $^{1,2}$ , Ryu Tadokoro $^{2}$ , Ryosuke Yamada $^{2,3}$ , Yuki M. Asano $^{4}$ , Iro Laina $^{5}$ , Christian Rupprecht $^{5}$ , Nakamasasa Inoue $^{6,2}$ , Rio Yokota $^{6,2}$ , Hirokatsu Kataoka $^{2}$ , and Yoshimitsu Aoki $^{1}$", + "bbox": [ + 228, + 210, + 774, + 258 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Keio University", + "bbox": [ + 442, + 268, + 560, + 284 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ National Institute of Advanced Industrial Science and Technology (AIST)", + "bbox": [ + 248, + 282, + 751, + 297 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3 University of Tsukuba", + "bbox": [ + 419, + 297, + 583, + 311 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "4 University of Amsterdam, now at University of Technology Nuremberg", + "bbox": [ + 258, + 311, + 743, + 325 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "5 University of Oxford", + "bbox": [ + 424, + 325, + 576, + 338 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{6}$ Tokyo Institute of Technology", + "bbox": [ + 393, + 339, + 607, + 353 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. In this work, we investigate the understudied effect of the training data used for image super-resolution (SR). Most commonly, novel SR methods are developed and benchmarked on common training datasets such as DIV2K and DF2K. However, we investigate and rethink the training data from the perspectives of diversity and quality, thereby addressing the question of \"How important is SR training for SR models?\" To this end, we propose an automated image evaluation pipeline. With this, we stratify existing high-resolution image datasets and larger-scale image datasets such as ImageNet and PASS to compare their performances. We find that datasets with (i) low compression artifacts, (ii) high within-image diversity as judged by the number of different objects, and (iii) a large number of images from ImageNet or PASS all positively affect SR performance. We hope that the proposed simple-yet-effective dataset curation pipeline will inform the construction of SR datasets in the future and yield overall better models. Code is available at: https://github.com/gohtanii/DiverSeg-dataset", + "bbox": [ + 259, + 386, + 743, + 609 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Super-resolution dataset $\\cdot$ Image compression $\\cdot$ Image diversity", + "bbox": [ + 259, + 621, + 740, + 648 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 672, + 375, + 690 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Image super-resolution (SR) aims to reconstruct high-resolution images from low-resolution images. It has been considered as one of the most fundamental tasks in the field of computer vision, with applications ranging from autonomous driving to medical imaging. Deep learning methods have led to significant advances in SR over the last decade, focusing primarily on improvements in the neural network architectures. Early SR models rely on convolutional neural networks (CNNs) [9,10,12,14,20,26,28,29,35-37]. Recent innovations have given rise to transformer-based SR models [5,6,16,19,33,34,42], which have consistently improved the performances.", + "bbox": [ + 212, + 703, + 787, + 840 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/dfdd78fcf01ffef938e6622e91f2f74a50994f16ad0ad0b8a3eafbfdb911127a.jpg", + "image_caption": [ + "Fig. 1: We propose an automated image evaluation pipeline to curate a dataset for training SR models. The obtained dataset, namely DiverSeg, consists of low-resolution but high-quality images with many object regions. SR models trained on DiverSeg outperform those trained on high-resolution image datasets such as DF2K and LSDIR." + ], + "image_footnote": [], + "bbox": [ + 222, + 145, + 774, + 422 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "With the improvement of neural network architectures, the importance of training datasets has also increased, as discussed in [17]. Examples of high-resolution datasets include DIV2K [1] and Flickr2K [27]. The combined dataset of these two, referred to as DF2K, is often utilized for training SR models. Most recently, LSDIR [17] has been proposed, which consists of 84,991 high-resolution images. It has been confirmed that training on large high-resolution datasets contributes significantly to performance improvement [18, 19, 29, 38, 42]", + "bbox": [ + 212, + 520, + 787, + 626 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The conventional approach to constructing datasets relies on a manual evaluation step, where the following two perspectives are most commonly considered:", + "bbox": [ + 212, + 626, + 787, + 656 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. Resolution and quality [17, 31]. This perspective focuses on the pixel density of images. Images that do not meet the specified resolution threshold are excluded. Typically, HD, 2K, and 4K images are used to create a dataset. After the initial automatic filtering based on image size, the details of each image are manually evaluated to identify and exclude compressed images.", + "2. Variety and diversity [15]. This includes diversity in subjects (e.g., people, landscapes, urban scenes), lighting conditions, colors, textures, and other photographic elements. A diverse dataset is said to help train a model that is robust and performs well across a wide range of domains." + ], + "bbox": [ + 220, + 664, + 787, + 801 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Datasets constructed from these perspectives have been shown to significantly improve the performance of SR models. However, they also pose challenges in", + "bbox": [ + 212, + 809, + 787, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "G. Ohtani et al.", + "bbox": [ + 271, + 114, + 379, + 127 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "scaling the datasets, as collecting uncompressed high-resolution images is difficult and costly.", + "bbox": [ + 212, + 146, + 782, + 176 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To address this limitation, this paper rethinks these perspectives and proposes Diverse Segmentation dataset (DiverSeg), a low-resolution1 yet effective image dataset for training SR models. As shown in Figure 1, the dataset is constructed by applying filtering to a large set of low-resolution images, such as ImageNet-1k [8] and PASS [2]. In experiments, we demonstrate that models trained on DiverSeg outperform those trained on high-resolution image datasets such as DF2K and LSDIR. Based on this finding, our contributions are summarized as follows.", + "bbox": [ + 212, + 183, + 787, + 303 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1) Rethinking the resolution perspective. High-resolution images have been considered to be necessary for training SR models. In this work, we challenge this traditional perspective and show that SR models can be trained without high-resolution images. Specifically, we introduce a method to estimate image quality based on the kernel density estimation over blockiness values [4] that estimates the quantity distribution of blocking artifacts. We demonstrate that low-resolution images with high quality, indicated by reduced artifacts, can improve the performance of SR models. We also thoroughly analyze the impact of image quality on SR performance.", + "2) Rethinking the diversity perspective. When constructing datasets for training SR models, images containing only a small number of objects are often implicitly excluded during the manual resolution evaluation process. This is because evaluators typically focus on the details of objects or small objects in the images. Therefore, we explicitly calculate the number of objects in images and analyze how this number affects SR performance. In our experiments, we show that constructing datasets with images containing many objects improves the performance of SR models.", + "3) Dataset construction. To facilitate analysis from the above two perspectives, we introduce a framework that automatically creates a dataset from a set of low-resolution images collected from the web. Specifically, the framework consists of two steps, source selection and object-based filtering, which correspond to the first and second perspectives, respectively. Our framework eliminates the need for manual assessment, facilitating dataset scaling. In the first step, given several low-resolution image datasets, we estimate the image quality of them through their blockiness distributions and select datasets with quality larger than $90\\%$ . In the second step, we filter out images with a small number of object regions by using object detection and image segmentation models. The filtered dataset, which we refer to as the DiverSeg dataset, is utilized for training various SR models in our experiments. We apply our framework to a union set of ImageNet, Places365, and PASS to construct the DiverSeg dataset. We demonstrate that SR models trained on DiverSeg archive state-of-the-art performance." + ], + "bbox": [ + 212, + 310, + 787, + 792 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "Rethinking Image Super-Resolution from Training Data Perspectives", + "bbox": [ + 272, + 114, + 730, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "page_footnote", + "text": "1 We define images with a resolution lower than HD as low-resolution images.", + "bbox": [ + 217, + 824, + 736, + 839 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 143, + 388, + 160 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.1 Image super-resolution models", + "text_level": 1, + "bbox": [ + 215, + 178, + 516, + 194 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "A number of SR models have been proposed that take advantage of deep learning techniques. These models can be divided into two groups, CNN-based models [9, 10,12,14,20,26,28,29,35-37] and Transformer-based models [5,6,16,19,33,34,42]. Each group exploits different architectural strengths to enhance low-resolution images to high resolution.", + "bbox": [ + 212, + 203, + 785, + 277 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "CNN-based models. SRCNN [9] was the first model that integrates a deep convolutional architecture for SR. Subsequently, FSRCNN [10] significantly improved computational efficiency by performing convolutional processing in low resolution space and upsampling in the last layer. Furthermore, ESPCN [26] adopted an efficient upsampling method, sub-pixel convolution, to enhance performance while reducing computational costs. These methods established the basic structure of modern SR networks and laid the foundation for subsequent research developments. Later studies introduced various modules such as residual connections [12, 14, 20, 29], dense blocks [28, 29, 37], and attention mechanisms [7, 35, 36]. EDSR [20] eliminated batch normalization layers and introduced residual scaling to enable stable training of large models. MSRResNet [29] replaced the basic ResNet blocks with residual dense blocks to improve the balance between performance and computational efficiency. RCAN [36] incorporated channel attention mechanisms to adaptively weight feature representations, significantly enhancing SR performance.", + "bbox": [ + 212, + 279, + 785, + 503 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Transformer-based models. As the first Transformer-based image restoration model, IPT [5] was introduced as a large-scale model utilizing the Transformer's encoder and decoder architecture. By pre-training on ImageNet, IPT significantly improved SR performance, fully leveraging the capabilities of the Transformer. SwinIR [19] executes self-attention within local windows during feature extraction, demonstrating exceptional SR performance and establishing itself as the foundational model for Transformers in SR. Following this, models that build upon SwinIR have been developed [6,33,34,42], enhancing performance by extending and optimizing the self-attention mechanism. Among them, HAT [6] achieves state-of-the-art performance in SR by using a duplicated cross-attention module and pre-training on ImageNet. However, there may be potential for improvement, as ImageNet is primarily an image recognition dataset and may not be fully optimized for SR.", + "bbox": [ + 212, + 506, + 787, + 702 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.2 Image super-resolution datasets", + "text_level": 1, + "bbox": [ + 215, + 724, + 526, + 739 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "T91 [30] and BSDS200 [22] are early datasets used to train SR models, consisting of 91 and 200 images, respectively. The turning point in training datasets was ushered by the release of DIV2K [1], a compilation of 800 high-resolution images with minimal compression noise meticulously collected from the web. Subsequently, to accommodate the further scaling of the model, Flickr2K [27], consisting of 2,650 high-resolution images, is merged into DIV2K and is referred", + "bbox": [ + 212, + 750, + 785, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "G. Ohtani et al.", + "bbox": [ + 271, + 114, + 380, + 128 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/dea9793fed3a8a8306fa8e627c1f807638b424faceb95b2451095ca6ff31d578.jpg", + "image_caption": [ + "Fig. 2: Images of DiverSeg with their segmentation masks. DiverSeg is obtained from a large set of low-resolution images through the automated image evaluation pipeline." + ], + "image_footnote": [], + "bbox": [ + 217, + 145, + 782, + 282 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "to as DF2K. In recent years, to further expand the scale of SR datasets, datasets larger than DF2K have been released, such as LSDIR [17], comprising 84,991 images and HQ-50K [31] consisting of 50,000 images. These datasets persistently adhere to stringent criteria, ensuring high resolution and negligible compression noise in the imagery. However, since collecting images that meet the aforementioned conditions is quite challenging, these datasets still consist of only tens of thousands of images. In contrast, several approaches [5, 6, 16] adopt ImageNet, also known as ImageNet-1k, which consists of 1.28M images spanning diverse categories. These approaches leverage the diversity of texture patterns in ImageNet as an advantage. Nevertheless, ImageNet is pointed out to contain some images having low-resolution and JPEG-compression artifacts, which adversely affect the training results.", + "bbox": [ + 212, + 353, + 787, + 534 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3 DiverSeg dataset", + "text_level": 1, + "bbox": [ + 215, + 556, + 415, + 571 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "This section presents Diverse Segmentation dataset (DiverSeg), our dataset for training SR models without using high-resolution images. As shown in Figure 2, the DiverSeg dataset consists of low-resolution but high-quality images with diverse segmented object regions. The framework for constructing the dataset consists of two steps, source selection and object-based filtering, which are designed from our perspective of rethinking resolution and diversity, respectively. Our approach eliminates the manual cost of collecting and quality-checking high-resolution images.", + "bbox": [ + 212, + 585, + 787, + 705 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1 Source selection", + "text_level": 1, + "bbox": [ + 215, + 726, + 395, + 739 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Let $\\mathcal{X}$ be a set of low-resolution image datasets. This step filters out low-quality datasets by estimating the quality $\\hat{q}_X \\in [0,1]$ for each dataset $X \\in \\mathcal{X}$ and excluding those with $\\hat{q}_X < 0.9$ , under the assumption that low-quality images are detrimental when training SR models. We introduce a quality estimation method based on the Kullback-Leibler (KL) divergence between blockiness distributions as detailed below.", + "bbox": [ + 212, + 750, + 787, + 839 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "Rethinking Image Super-Resolution from Training Data Perspectives", + "bbox": [ + 272, + 114, + 730, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/51f21f57bc9f13566804ac593d29b853c03b7caaf49014cc9d08b4949e5874b4.jpg", + "image_caption": [ + "Fig. 3: Comparison of image degradation due to JPEG quality(blue). Blockiness values calculated from the images are marked. As the JPEG quality decreases and artifacts increase, we observe a corresponding rise in blockiness values." + ], + "image_footnote": [], + "bbox": [ + 223, + 143, + 787, + 256 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/7681ccaf48f120ec4f05e6755a7a09218b84820466595b61affb1bb84866e47b.jpg", + "image_caption": [ + "Fig. 4: (a) Blockiness distributions $p_{X,1.0}$ for $X = \\mathrm{ImageNet - 1k}$ , Places365 and PASS. (b) Basis distributions $p_{Z,q}$ for $Z = \\mathrm{DF2K}$ and $q = 0.5, 0.75, 0.85, 0.95, 1.0$ . We estimate the quality by comparing $p_{X,1.0}$ and $p_{Z,q}$ using the KL divergence." + ], + "image_footnote": [], + "bbox": [ + 220, + 318, + 500, + 455 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/731aae67800aee7ca10548feb9822e51a430967e4189ce57b9a7969c90690e8d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 504, + 318, + 784, + 455 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Image datasets. This work uses three web-collected low-resolution datasets $\\mathcal{X} = \\{\\mathrm{ImageNet - 1k},\\mathrm{Places365},\\mathrm{PASS}\\}$ . Training SR models on them is not straightforward because they may include highly compressed images that negatively affect training of SR models.", + "bbox": [ + 212, + 539, + 797, + 599 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Quality definition. Let $Y$ be a dataset of JPEG images. We define the quality $q_{Y}$ by the average JPEG quality, i.e., $q_{Y} \\triangleq \\frac{1}{|Y|} \\sum_{y \\in Y} Q(y)$ where $0 \\leq Q(y) \\leq 1$ is the JPEG quality of an image $y$ . The goal of quality estimation is to estimate $q_{X}$ given a dataset $X$ . Note that in the estimation phase, datasets may include images other than JPEG and the true quality $q_{X}$ is not observable.", + "bbox": [ + 212, + 599, + 787, + 676 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Blockiness distribution. To estimate the quality, we utilize the blockiness measure [4]. Specifically, for each image dataset $X \\in \\mathcal{X}$ , we estimate the distribution of blockiness values $p_{X,q}(b)$ by kernel density estimation as follows:", + "bbox": [ + 212, + 676, + 787, + 724 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\np _ {X, q} (b) = \\frac {1}{h | X |} \\sum_ {x \\in X} K \\left(\\frac {b - B \\left(\\phi_ {q} (x)\\right)}{h}\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 352, + 744, + 785, + 782 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $x$ is an image, $\\phi_q$ is the JPEG compression function, $q \\in [0,1]$ is a quality value, $K: \\mathbb{R} \\to \\mathbb{R}$ is a Gaussian kernel, and $h \\in \\mathbb{R}$ is the bandwidth determined by Scott's method. The function $B$ is the blockiness measure that measures the", + "bbox": [ + 212, + 794, + 787, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "G. Ohtani et al.", + "bbox": [ + 271, + 114, + 379, + 127 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "quantity of blocking artifacts by computing subband DCT coefficients. Specifically, $B$ is defined on images that are decomposed into $P \\times P$ patches as follows:", + "bbox": [ + 212, + 146, + 782, + 176 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\nB (x) = \\sum_ {i = 1} ^ {P} \\sum_ {j = 1} ^ {P} \\left| \\frac {\\bar {V} _ {\\text {c r o p}} (i , j) - \\bar {V} (i , j)}{\\bar {V} (i , j)} \\right|, \\quad \\bar {V} (i, j) = \\frac {1}{W H} \\sum_ {h = 1} ^ {H} \\sum_ {w = 1} ^ {W} V _ {w, h} (i, j) \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 233, + 200, + 785, + 244 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $W, H$ are the width and height of an image $x$ , $V_{w,h}(i,j)$ is the variation in the $(i,j)$ -th subband DCT coefficients within the $(h,w)$ -th patch being calculated and its four spatially adjacent patches. $V_{\\mathrm{crop}}(i,j)$ is the variation calculated similarly to $V_{w,h}(i,j)$ for the given image with the first 4 rows and 4 columns removed. $\\bar{V}$ and $\\bar{V}_{\\mathrm{crop}}$ are the average variations of $V_{w,h}(i,j)$ and $V_{\\mathrm{crop}}(i,j)$ calculated for each patch, respectively. This work uses $P = 8$ . The blockiness value $B(x)$ is expected to be low for uncompressed images and high for compressed images as shown in Figure 3.", + "bbox": [ + 212, + 255, + 787, + 377 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Quality estimation. We estimate the quality by comparing $p_{X,1.0}$ with $\\{p_{Z,q}\\}_{q \\in S}$ , where $p_{Z,q}$ is a basis distribution, a blockiness distribution of images of a fixed quality $q$ . More specifically, the estimated quality is given by", + "bbox": [ + 212, + 377, + 799, + 424 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {q} _ {X} = \\sum_ {q \\in S} q \\frac {\\exp \\left(- D _ {\\mathrm {K L}} \\left(p _ {X , 1 . 0} \\mid \\mid p _ {Z , q}\\right)\\right)}{\\sum_ {q ^ {\\prime} \\in S} \\exp \\left(- D _ {\\mathrm {K L}} \\left(p _ {X , 1 . 0} \\mid \\mid p _ {Z , q ^ {\\prime}}\\right)\\right)}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 341, + 434, + 785, + 472 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $Z$ is a small dataset that involves only uncompressed images. In this work, we use DF2K. $D_{\\mathrm{KL}}$ is the KL divergence, and $S = \\{1.0, 0.95, 0.85, 0.75, 0.5\\}$ are discretely sampled quality values. Figure 4 shows the blockiness distributions $p_{X,1.0}$ for the three low-resolution image datasets and the basis distributions $p_{Z,c}$ .", + "bbox": [ + 212, + 484, + 787, + 559 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Selection results. The estimated qualities for ImageNet-1k, Places365, and PASS were $95.5\\%$ , $75.0\\%$ and $99.8\\%$ , respectively. From this result, Places365 is filtered out.", + "bbox": [ + 212, + 560, + 787, + 604 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.2 Object-based filtering", + "text_level": 1, + "bbox": [ + 214, + 630, + 442, + 647 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Given a source dataset $X$ , this step applies filtering to refine it as a dataset for training SR models, under the assumption that images with diverse object regions are more effective than those with uniform or monotonous content. Specifically, the refined training dataset is given by $\\tilde{X} = \\{x \\in X : R(x) \\geq \\theta\\}$ , where $R(x)$ is the number of object regions and $\\theta$ is a threshold.", + "bbox": [ + 212, + 657, + 787, + 733 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We introduce two object-based filtering methods with different granularities to explore how the granularity of object detection, ranging from detailed identification of small objects or features to recognizing larger, more general objects, impacts SR performance.", + "bbox": [ + 212, + 733, + 787, + 794 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Segmentation-based filtering. This method counts the number of object parts by an image segmentation model. Specifically, we adopt the SAM [13] with the ViT-H backbone and define $R$ by the number of segmentation masks.", + "bbox": [ + 212, + 794, + 787, + 840 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "Rethinking Image Super-Resolution from Training Data Perspectives", + "bbox": [ + 272, + 114, + 732, + 130 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/156a6e7c22da197a23ac7adfeffe1e262a93ef2f3509fa50bc5c6cf226e0d6d2.jpg", + "table_caption": [ + "Table 1: Dataset statistics. HR: High resolution, LR: Low resolution, #Images: Number of images, #Pixels: Average number of pixels per image, Blockiness: median of blockiness measure indicating the intensity of JPEG compression noise, #Segments: Average number of segmentation masks." + ], + "table_footnote": [], + "table_body": "
DatasetTaskHR LR#Images#PixelsBlockiness#Segments
DIV2K [1]Super-resolution✓8002.8M0.47104
DF2K [1,27]Super-resolution✓3,4502.8M0.47103
LSDIR [17]Super-resolution✓84,9911.1M0.8292
ImageNet [8]Image recognition1,281,167237k4.3971
Places365 [39]Image recognition1,803,460366k80.71100
PASS [2]Image recognition1,439,589178k3.0374
DiverSeg-I (Ours)Super-resolution259,448233k2.83146
DiverSeg-P (Ours)Super-resolution267,055179k4.39146
DiverSeg-IP (Ours)Super-resolution526,503206k3.61146
", + "bbox": [ + 220, + 212, + 782, + 373 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We chose SAM because, unlike typical segmentation models that perform semantic segmentation based on class labels, SAM provides segments that do not impose semantic constraints, allowing for finer region segmentation of diverse objects. We set $\\theta = 100$ as the default, resulting in $260\\mathrm{k}$ images remaining after applying the filter to ImageNet-1k.", + "bbox": [ + 212, + 405, + 782, + 479 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Detection-based filtering. This method counts the number of objects per image. We adopt the Detic model [41] with the ViT-B backbone and define $R$ by the number of detected objects. We set $\\theta = 18$ , resulting in 260k images remaining after applying the filter to ImageNet-1k.", + "bbox": [ + 212, + 482, + 785, + 544 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "3.3 Dataset statistics", + "text_level": 1, + "bbox": [ + 215, + 571, + 405, + 587 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 1 shows dataset statistics. We created three variants of DiverSeg, namely DiverSeg-I, DiverSeg-P, and DiverSeg-IP, constructed from ImageNet-1k, PASS, and the union of the two, respectively. The segmentation-based filtering is applied to obtain these datasets. Compared to high-resolution datasets such as DF2K and LSDIR, our datasets have larger number of training images, but contain only low-resolution images. The median of blockiness values is decreased for ImageNet-1k and is increased for PASS after filtering.", + "bbox": [ + 212, + 604, + 787, + 712 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 215, + 739, + 375, + 758 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This section conducts experiments by training various SR models on DiverSeg datasets. We demonstrate that SR models can be trained without using high-resolution images and thoroughly analyze factors from a dataset perspective that are crucial for enhancing the performance of SR.", + "bbox": [ + 212, + 779, + 785, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "G. Ohtani et al.", + "bbox": [ + 271, + 114, + 380, + 128 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/f1af37c10faa956dc2e7194a6accb2ee32c04c70746bdf3050afbbef0306f9c8.jpg", + "table_caption": [ + "Table 2: Comparison of DiverSeg with high-resolution image datasets (DF2K and LSDIR). Models trained on DiverSeg-I and DiverSeg-P demonstrated superior performance despite not using any high-resolution images for training." + ], + "table_footnote": [], + "table_body": "
Model (Params)DatasetHR LRSet5Set14BSD100Urban100Manga109
PSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIM
MSRResNet [29] (1.5M)DF2K✓32.230.895528.670.783127.620.737426.230.789730.640.9108
LSDIR✓32.150.894828.660.783627.620.737426.310.791830.570.9105
DiverSeg-I✓32.270.896328.640.783727.640.737826.310.791830.530.9115
DiverSeg-P✓32.090.894328.610.783227.600.737126.280.791830.360.9101
RCAN [36] (15.5M)DF2K✓32.500.899028.870.788527.750.742126.730.805831.170.9165
LSDIR✓32.530.899228.890.789427.750.742526.910.809031.330.9180
DiverSeg-I✓32.700.901228.980.790827.810.744327.030.811631.580.9210
DiverSeg-P✓32.630.900028.950.789827.770.743526.990.813431.190.9190
EDSR [20] (43.0M)DF2K✓32.610.899828.910.789327.790.743426.840.808931.380.9176
LSDIR✓32.570.899228.970.790827.800.743827.050.813131.470.9192
DiverSeg-I✓32.710.901728.980.791327.850.745327.100.814231.720.9216
DiverSeg-P✓32.570.900229.060.791527.800.744727.100.816331.330.9191
SwinIR [19] (11.9M)DF2K✓32.920.904429.090.795027.920.748927.450.825432.030.9260
LSDIR✓32.860.903629.160.796327.920.749227.790.833131.980.9262
DiverSeg-I✓32.970.905329.230.797027.980.750827.830.833632.340.9283
DiverSeg-P✓32.850.904029.240.796127.960.750227.850.834932.280.9278
HAT [6] (20.7M)DF2K✓33.030.905629.160.796427.990.751427.930.836532.440.9292
LSDIR✓32.930.905329.290.798828.010.752528.450.846932.570.9306
DiverSeg-I✓33.150.907129.460.800428.070.754228.510.847732.900.9325
DiverSeg-P✓33.120.906829.500.800228.040.753628.530.849232.830.9320
", + "bbox": [ + 218, + 199, + 782, + 460 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.1 Experimental settings", + "text_level": 1, + "bbox": [ + 215, + 487, + 442, + 503 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "SR models. We use five models. Specifically, we use three CNN-based models: MSRResNet [29], EDSR [20], RCAN [36], and two Transformer-based models: SwinIR [19] and HAT [6].", + "bbox": [ + 212, + 508, + 782, + 555 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Training datasets. We compare the DiverSeg datasets with two high-resolution datasets: DF2K and LSDIR [17]. DF2K is a merged dataset of DIV2K [1] and Flickr2K [27].", + "bbox": [ + 212, + 555, + 782, + 599 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Evaluation datasets. We use five benchmark datasets: Set5 [3], Set14 [32], BSD100 [22], Urban100 [11], and Manga109 [23].", + "bbox": [ + 212, + 599, + 782, + 631 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Evaluation metrics. PSNR and SSIM on the Y channel (representing luminance) within the transformed YCbCr color space are used as evaluation metrics.", + "bbox": [ + 212, + 631, + 782, + 661 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Implementation settings. We follow the training setup of the original papers of the SR models [6, 19, 20, 29, 36]. Implementation details are provided in the supplementary materials.", + "bbox": [ + 212, + 661, + 782, + 705 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.2 Experimental results", + "text_level": 1, + "bbox": [ + 215, + 726, + 433, + 742 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Main results. To validate the effectiveness of our approach, we trained the five SR models on DiverSeg-I/P. The results are summarized in Table 2. As shown, models trained on DiverSeg-I/P achieved better performance than those trained on DF2K and LSDIR. This shows the effectiveness of the proposed datasets. To the best of our knowledge, we are the first to successfully train SR models without using high-resolution images.", + "bbox": [ + 212, + 750, + 784, + 839 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "Rethinking Image Super-Resolution from Training Data Perspectives", + "bbox": [ + 272, + 114, + 730, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/58f422521b66ab6d2656c18defe6bf07f9c80a44692ed832124098cb35211895.jpg", + "table_caption": [ + "Table 3: Quantitative comparison with state-of-the-art methods on five benchmark datasets. We applied our dataset to two Transformer-based models. Checkmarks for HR, LR indicate the use of high-resolution and low-resolution datasets, respectively." + ], + "table_footnote": [], + "table_body": "
MethodTraining DataHR LRSet5Set14BSD100Urban100Manga109
PSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIM
SAN [7]DIV2K✓32.640.900328.920.788827.780.743626.790.806831.180.9169
IGNN [40]DIV2K✓32.570.899828.850.789127.770.743426.840.809031.280.9182
HAN [25]DIV2K✓32.640.900228.900.789027.800.744226.850.809431.420.9177
NLSN [24]DIV2K✓32.590.900028.870.789127.780.744426.960.810931.270.9184
RRDB [29]DF2K✓32.730.901128.990.791727.850.745527.030.815331.660.9196
RCAN-it [21]DF2K✓32.690.900728.990.792227.870.745927.160.816831.780.9217
EDT [16]DF2K✓32.820.903129.090.793927.910.748327.460.824632.050.9254
HAT-S [6]DF2K✓32.920.904729.150.795827.970.750527.870.834632.350.9283
IPT [5]ImageNet✓32.64-29.01-27.82-27.26---
SwinIR [19]DF2K✓32.920.904429.090.795027.920.748927.450.825432.030.9260
SwinIR [19]DiverSeg-I (Ours)✓32.970.905329.230.797027.980.750827.830.833632.340.9283
HAT [6]DF2K✓33.040.905629.230.797328.000.751727.970.836832.480.9292
HAT [6]ImageNet→DF2K✓33.180.907329.380.800128.050.753428.370.844732.870.9319
HAT [6]DiverSeg-I (Ours)✓33.150.907129.460.800428.060.754228.510.847732.900.9325
HAT [6]DiverSeg-IP (Ours)✓33.140.907329.510.800728.070.754228.540.849232.960.9327
HAT-L [6]ImageNet→DF2K✓33.300.908329.470.801528.090.755128.600.849833.090.9335
HAT-L [6]DiverSeg-I (Ours)✓33.280.908329.540.802228.100.755628.750.852933.140.9340
HAT-L [6]DiverSeg-IP (Ours)✓33.200.908029.590.801928.110.755628.810.854733.190.9342
", + "bbox": [ + 218, + 202, + 782, + 426 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/56e01cf05478fe0d7e760f12c72728e011b1ad3304a304b24b8fc89b9b6acdb1.jpg", + "table_caption": [ + "Table 4: Filtering by blockiness (HAT model, ImageNet-1k). $\\theta^{\\prime}$ : threshold for blockiness values." + ], + "table_footnote": [], + "table_body": "
θ'#ImagesSet5Set14BSD100Urban100Manga109
PSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIM
10800k33.070.906829.340.800028.040.753828.410.846532.560.9310
30939k33.130.907229.390.800128.060.753828.430.846332.700.9316
1001.08M33.110.907129.420.800128.050.753728.410.845732.840.9321
-1.15M33.080.907129.400.799928.050.753528.410.845732.880.9323
", + "bbox": [ + 236, + 474, + 771, + 566 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Comparison with SOTA. Table 3 shows that DiverSeg improves the state-of-the-art performance. Specifically, the HAT and HAT-L models trained on DiverSeg datasets outperformed those trained with ImageNet-1k $\\rightarrow$ DF2K, which utilizes all ImageNet-1k images for pre-training and the DF2K images for fine-tuning, in terms of PSNR and SSIM on four of five benchmarking datasets. It is worth noting that DiverSeg-I filters out $77.5\\%$ of images from ImageNet-1k and thus training on it is more efficient than the approach relying on pre-training and fine-tuning. This confirmed the effectiveness and efficiency of our filtering approach.", + "bbox": [ + 212, + 593, + 787, + 731 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.3 Analysis 1: Effects of filtering", + "text_level": 1, + "bbox": [ + 215, + 753, + 506, + 768 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Filtering by blockiness measure. If the blockiness is an important factor for enchaining SR performance, filtering images by the blockiness values, i.e., constructing a training dataset by $\\hat{X} = \\{x\\in X:B(x)\\leq \\theta '\\}$ , would be a straightforward approach. However, as shown in Table 4, this filtering did not", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "G. Ohtani et al.", + "bbox": [ + 271, + 114, + 380, + 127 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/89d07a02fbd731f9f7f41c738f688d3539f0c08a5f84d0d2c5cf0700c7bb93cb.jpg", + "table_caption": [ + "Table 5: Comparison of filtering methods. Performance is evaluated using $260\\mathrm{k}$ filtered images (HAT model, ImageNet-1k)." + ], + "table_footnote": [], + "table_body": "
Filtering methodUrban100Manga109
PSNRSSIMPSNRSSIM
Blockiness28.390.846732.470.9304
Detection-based28.440.846232.870.9322
Seg.-based28.510.847732.900.9325
", + "bbox": [ + 227, + 191, + 514, + 268 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/7c456356d0e101aa2fcc8910a0c5ce7e49d098e1a21fd9df7511b70163ff2f3f.jpg", + "table_caption": [ + "Table 6: Performance comparison across different thresholds $\\theta$ (HAT model, ImageNet-1k)." + ], + "table_footnote": [], + "table_body": "
θ#ImagesUrban100Manga109
PSNRSSIMPSNRSSIM
01.2M28.410.845732.880.9323
50663k28.460.847232.900.9325
100259k28.510.847732.900.9325
15086k28.360.845232.830.9320
", + "bbox": [ + 558, + 191, + 787, + 268 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/8af3e0b5ee0687d5f4f1323378c3cf0fe395c5539eecff6991286b05af871b47.jpg", + "table_caption": [ + "Table 7: Analysis of effects of JPEG quality (HAT model)." + ], + "table_footnote": [], + "table_body": "
DatasetQuality (%)BlockinessSet5Set14BSD100Urban100Manga109
PSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIM
DF2K50165.3431.130.885527.540.753326.320.710325.040.751130.200.9024
7583.2032.630.900827.700.745727.700.745727.390.823831.880.9238
8546.9832.940.904329.070.794127.930.749927.720.831332.140.9266
9510.3332.980.904829.110.795127.990.752127.810.833832.210.9278
HR0.4733.030.905629.160.796427.990.751427.930.836532.440.9292
LSDIR50146.4328.500.848726.660.742124.720.667424.910.765729.390.8956
7557.1431.870.885328.530.780527.490.741727.300.823131.830.9223
8524.1432.840.904429.170.795527.980.751328.220.841232.380.9291
953.7132.970.904329.250.797928.050.753928.370.845232.420.9299
HR0.8232.930.905329.290.798828.010.752528.450.846932.570.9306
", + "bbox": [ + 220, + 303, + 782, + 450 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "improve PSNR and SSIM when reducing the threshold $\\theta^{\\prime}$ from 30 to 10 with an exception of SSIM on Urban100. This is because this method filters out images without considering the diversity of object regions.", + "bbox": [ + 212, + 479, + 784, + 526 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Object-based filtering. Table 5 compares object-based filtering methods described in Sec. 3.2, where ImageNet-1k is used as a source dataset. For a fair comparison, all dataset size after filtering are the same (260k images). As shown, the segmentation-based filtering performed the best. This indicates that images with diverse object regions are effective for SR training, and finer granularity leads to better performance.", + "bbox": [ + 212, + 527, + 784, + 617 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Filtering threshold. Table 6 shows the results obtained by varying the threshold $\\theta$ , which indicates the minimum number of segments. As shown, $\\theta = 100$ performed the best.", + "bbox": [ + 212, + 618, + 784, + 662 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4.4 Analysis 2: Effects of image quality", + "text_level": 1, + "bbox": [ + 215, + 675, + 550, + 691 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "In Sec. 3, we made an assumption that low-quality images are detrimental when training SR models. Here, we empirically justify this assumption and evaluate the impact of image quality on SR performance.", + "bbox": [ + 212, + 703, + 784, + 747 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Training with compressed images. This experiment applies JPEG compression to the two high-resolution datasets, DF2K and LSDIR, and trains a HAT model on each compressed dataset. As shown in Table 7, the performance decreases as the quality decreases on both datasets.", + "bbox": [ + 212, + 750, + 784, + 808 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Learning process. To further analyze why and how low quality images negatively affect training of SR models, Figure 5 compares learning processes. As can", + "bbox": [ + 212, + 809, + 784, + 839 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "Rethinking Image Super-Resolution from Training Data Perspectives", + "bbox": [ + 272, + 114, + 730, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/4fe7e02fcbb502ad20d83e2666ecc80b5bb0b42d3ad2219e74bc6e9ad45f3854.jpg", + "image_caption": [ + "Fig. 5: Comparison of learning processes obtained with various JPEG quality values." + ], + "image_footnote": [ + "Quality: HR 95% 85% 75% 50%" + ], + "bbox": [ + 217, + 143, + 496, + 295 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/9774f3a1c9fc6228546a0672961223950c0359cc76f84c234b1f0249a12a00a8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 496, + 143, + 776, + 295 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/40f63773650fba6630c4347c4d4cf97ee3ae59895185ef5144f749d12928e89c.jpg", + "image_caption": [ + "Fig. 6: Comparison of learning processes for ImageNet-1k, PASS, and Places365." + ], + "image_footnote": [ + "ImageNet PASS Places365" + ], + "bbox": [ + 220, + 339, + 406, + 443 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/c19f8097c1bcbb0ce6cfc774d4a09d4d923b91af515023267840f606f2228b97.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 406, + 340, + 591, + 443 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/7c9a6b211ee6ca901015927da0faf4f90ded647abab5082bd635f85931bb2006.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 591, + 340, + 776, + 443 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "be seen, PSNR decreases after 100k and 300k iterations with $50\\%$ and $75\\%$ qualities, respectively. With $50\\%$ quality, the final performance at the 800k iteration was worse than that at the 10k iteration. These results indicate that excluding low-quality images is crucial for enhancing SR performance. This is in contrast to other vision tasks such as image recognition, where increasing the number of training images, even with low quality images, often helps improve performance.", + "bbox": [ + 212, + 500, + 787, + 590 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Training with Places365. In the source selection step, the Places365 dataset was excluded because its quality estimated via the blockiness distribution was low. To justify this selection, we compare three datasets in Figure 6. As can be seen, the model trained on Places365 performed worse than those trained on ImageNet-1k and PASS. Similar to the learning processes obtained from low quality JPEG images, the PSNR decreased in the later phase of the learning process with Places365. These results confirmed the effectiveness of source selection before training based on the blockiness distributions.", + "bbox": [ + 212, + 590, + 787, + 710 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "4.5 Analysis 3: Visual comparison", + "text_level": 1, + "bbox": [ + 215, + 722, + 511, + 739 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Comparison of artifacts. Figure 7 compares the artifacts produced by JPEG compression with those produced by the SR models trained on DF2K using different image quality levels. Images obtained from models trained on $50\\%$ and $75\\%$ quality images show strong stripes or checkerboard patterns of artifacts. These artifacts appear to be more significant than those observed when the original image is compressed using JPEG, suggesting that they are induced by", + "bbox": [ + 212, + 750, + 787, + 842 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "G. Ohtani et al.", + "bbox": [ + 271, + 114, + 380, + 127 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/0d31d797598d68abc1e249599b9d1ac33565c56a4a8734043d41d62036345499.jpg", + "image_caption": [ + "AkkeraKanjinchou, Manga109" + ], + "image_footnote": [], + "bbox": [ + 230, + 147, + 390, + 339 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/0862b931f81a796c6f6b747088f909f80c76b8c9e7e5c9baed3349edaf2b1545.jpg", + "image_caption": [ + "Fig. 7: Comparison of artifacts produced by JPEG compression (blue) and SR models trained on various JPEG quality values (red, HAT model, DF2K). Numbers indicate JPEG image quality. First and third rows: cropped images. Second and fourth rows: differential images." + ], + "image_footnote": [], + "bbox": [ + 233, + 354, + 383, + 409 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/2d0325fc8fc994d65c7de56e725c520cae67d0ce7bd3c4265894d1a15d8a037e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 401, + 147, + 777, + 407 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "model training. The predisposition to stripes and checkerboard patterns is likely due to the inductive bias inherent in the architecture of the neural network; in particular, the square or rectangular shape of the filters in the convolutional operations could cause this. Improvements to the network architecture to allow training on lower quality images would be interesting as future work.", + "bbox": [ + 212, + 507, + 784, + 583 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Qualitative examples. Figure 8 shows visual comparisons of SR models trained on DF2K, LSDIR, DiverSeg-I, and DiverSeg-P. For the image \"img_011\" in Urban100, we observed that models trained on DF2K are unable to recover the horizontal stripes pattern, while the other three models successfully recovered it. With the three images of Manga109, we observed that models trained on DiverSeg-I/P exhibit noticeable improvement in the character region compared to those trained on DF2K or LSDIR.", + "bbox": [ + 212, + 583, + 787, + 688 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "4.6 Discussion and Limitations", + "text_level": 1, + "bbox": [ + 215, + 696, + 483, + 710 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "High-resolution datasets. In this paper, we applied the automated image evaluation pipeline to a large dataset of low-resolution images. We believe our finding contributes to the future construction of training datasets and the development of neural network architectures. Specifically, we demonstrated that SR models can be trained without high-resolution images. However, this does not imply that high-resolution image datasets are worthless. Rather, we believe that increasing the diversity of high-resolution images through our proposed filtering method could further improve SR performance. In Table 8, we examined this on", + "bbox": [ + 212, + 719, + 787, + 839 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "Rethinking Image Super-Resolution from Training Data Perspectives", + "bbox": [ + 272, + 114, + 730, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/beb5ae6da64ad56843148dee603f6ebba143057f079a8aefbcac6df2b9271b71.jpg", + "image_caption": [ + "Fig. 8: Visual comparison of $\\times 4$ SR models trained on DF2K, LSDIR, DiverSeg-I, and DiverSeg-P datasets. PSNR/SSIM is calculated for each cropped patch individually to better reflect the differences in performance." + ], + "image_footnote": [], + "bbox": [ + 218, + 141, + 787, + 364 + ], + "page_idx": 13 + }, + { + "type": "table", + "img_path": "images/82cbb55eb4833e51cfe5e61e1de05f984f9a28581222ce0d7eda8eb293f9ca8a.jpg", + "table_caption": [ + "Table 8: Applying object-based filtering to LSDIR with $\\theta = {100}$ (HAT model)." + ], + "table_footnote": [], + "table_body": "
Dataset#ImagesSet5Set14BSD100Urban100Manga109
PSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIM
LSDIR89,99132.930.905329.290.798828.010.752528.450.846932.570.9306
w/ filtering31,56132.950.905629.330.799128.020.752628.530.848732.600.9308
", + "bbox": [ + 233, + 450, + 769, + 510 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "the LSDIR dataset. Our results show that filtering out images with less than 100 object regions led to increased PSNR and SSIM with the HAT model. For future work, exploring a hybrid approach that utilizes both low- and high-resolution diverse images could be promising.", + "bbox": [ + 212, + 539, + 787, + 599 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Limitations. In this study, we focused on two perspectives: resolution and diversity. However, when collecting large datasets, there are other important perspectives, such as fairness and copyright. In addition, for benchmarking purposes, we used the five most commonly used SR datasets for a fair comparison with conventional methods. Nevertheless, real-world applications also face the challenge of blind super-resolution, where the degradation process is unknown. Creating datasets addressing this aspect would be an interesting future research direction.", + "bbox": [ + 212, + 599, + 787, + 707 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 722, + 359, + 738 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In this work, we investigated the effect of the training data for SR and showed that SR models are trainable even without using high-resolution images by applying the image evaluation pipeline to a set of large low-resolution images. In experiments, we thoroughly analyzed the effect of image quality and diversity to SR performance. We hope that this work will positively influence the future construction of training datasets and lead to better models.", + "bbox": [ + 212, + 748, + 787, + 839 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "G. Ohtani et al.", + "bbox": [ + 271, + 114, + 379, + 127 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 215, + 143, + 401, + 162 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "Computational resource of AI Bridging Cloud Infrastructure (ABCI) provided by National Institute of Advanced Industrial Science and Technology (AIST) was used.", + "bbox": [ + 212, + 175, + 784, + 220 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 215, + 243, + 321, + 258 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Agustsson, E., Timofte, R.: Ntire 2017 challenge on single image super-resolution: Dataset and study. In: CVPRW (2017) 2, 4, 8, 9", + "2. Asano, Y.M., Rupprecht, C., Zisserman, A., Vedaldi, A.: Pass: An imagenet replacement for self-supervised pretraining without humans. In: NeurIPS Track on Datasets and Benchmarks (2021) 3, 8", + "3. Bevilacqua, M., Roumy, A., Guillemot, C., Alberi-Morel, M.L.: Low-complexity single-image super-resolution based on nonnegative neighbor embedding. In: BMVC (2012) 9", + "4. Bhardwaj, D., Pankajakshan, V.: A jpeg blocking artifact detector for image forensics. Signal Processing: Image Communication 68, 155-161 (2018) 3, 6", + "5. Chen, H., Wang, Y., Guo, T., Xu, C., Deng, Y., Liu, Z., Ma, S., Xu, C., Xu, C., Gao, W.: Pre-trained image processing transformer. In: CVPR (2021) 1, 4, 5, 10", + "6. Chen, X., Wang, X., Zhou, J., Qiao, Y., Dong, C.: Activating more pixels in image super-resolution transformer. In: CVPR (2023) 1, 4, 5, 9, 10", + "7. Dai, T., Cai, J., Zhang, Y., Xia, S.T., Zhang, L.: Second-order attention network for single image super-resolution. In: CVPR (2019) 4, 10", + "8. Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: Imagenet: A large-scale hierarchical image database. In: CVPR (2009) 3, 8", + "9. Dong, C., Loy, C.C., He, K., Tang, X.: Learning a deep convolutional network for image super-resolution. In: ECCV (2014) 1, 4", + "0. Dong, C., Loy, C.C., Tang, X.: Accelerating the super-resolution convolutional neural network. In: ECCV (2016) 1, 4", + "1. Huang, J.B., Singh, A., Ahuja, N.: Single image super-resolution from transformed self-exemplars. In: CVPR (2015) 9", + "2. Kim, J., Lee, J.K., Lee, K.M.: Accurate image super-resolution using very deep convolutional networks. In: CVPR (2016) 1, 4", + "3. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., Dollar, P., Girshick, R.: Segment anything. In: ICCV (2023) 7", + "4. Ledig, C., Theis, L., Huszár, F., Caballero, J., Cunningham, A., Acosta, A., Aitken, A., Tejani, A., Totz, J., Wang, Z., et al.: Photo-realistic single image superresolution using a generative adversarial network. In: CVPR (2017) 1, 4", + "5. Li, A., Zhang, L., Liu, Y., Zhu, C.: Feature modulation transformer: Crossrefinement of global representation via high-frequency prior for image superresolution. In: CVPR (2023) 2", + "6. Li, W., Lu, X., Qian, S., Lu, J., Zhang, X., Jia, J.: On efficient transformer-based image pre-training for low-level vision. arXiv preprint arXiv:2112.10175 (2021) 1, 4, 5, 10", + "7. Li, Y., Zhang, K., Liang, J., Cao, J., Liu, C., Gong, R., Zhang, Y., Tang, H., Liu, Y., Demandolx, D., Ranjan, R., Timofte, R., Van Gool, L.: Lsdir: A large scale dataset for image restoration. In: CVPRW (2023) 2, 5, 8, 9" + ], + "bbox": [ + 225, + 273, + 784, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "Rethinking Image Super-Resolution from Training Data Perspectives", + "bbox": [ + 272, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "18. Li, Y., Zhang, Y., Timofte, R., Van Gool, L., Yu, L., Li, Y., Li, X., Jiang, T., Wu, Q., Han, M., et al.: Ntire 2023 challenge on efficient super-resolution: Methods and results. In: CVPRW (2023) 2", + "19. Liang, J., Cao, J., Sun, G., Zhang, K., Van Gool, L., Timofte, R.: Swinir: Image restoration using swin transformer. In: ICCVW (2021) 1, 2, 4, 9, 10", + "20. Lim, B., Son, S., Kim, H., Nah, S., Mu Lee, K.: Enhanced deep residual networks for single image super-resolution. In: CVPRW (2017) 1, 4, 9", + "21. Lin, Z., Garg, P., Banerjee, A., Magid, S.A., Sun, D., Zhang, Y., Van Gool, L., Wei, D., Pfister, H.: Revisiting rcan: Improved training for image super-resolution. arXiv preprint arXiv:2201.11279 (2022) 10", + "22. Martin, D., Fowlkes, C., Tal, D., Malik, J.: A database of human segmented natural images and its application to evaluating segmentation algorithms and measuring ecological statistics. In: ICCV (2001) 4, 9", + "23. Matsui, Y., Ito, K., Aramaki, Y., Fujimoto, A., Ogawa, T., Yamasaki, T., Aizawa, K.: Sketch-based manga retrieval using manga109 dataset. Multimedia Tools and Applications 76, 21811-21838 (2017) 9", + "24. Mei, Y., Fan, Y., Zhou, Y.: Image super-resolution with non-local sparse attention. In: CVPR (2021) 10", + "25. Niu, B., Wen, W., Ren, W., Zhang, X., Yang, L., Wang, S., Zhang, K., Cao, X., Shen, H.: Single image super-resolution via a holistic attention network. In: ECCV (2020) 10", + "26. Shi, W., Caballero, J., Huszar, F., Totz, J., Aitken, A.P., Bishop, R., Rueckert, D., Wang, Z.: Real-time single image and video super-resolution using an efficient sub-pixel convolutional neural network. In: CVPR (2016) 1, 4", + "27. Timofte, R., Agustsson, E., Van Gool, L., Yang, M.H., Zhang, L.: Ntire 2017 challenge on single image super-resolution: Methods and results. In: CVPRW (2017) 2, 4, 8, 9", + "28. Tong, T., Li, G., Liu, X., Gao, Q.: Image super-resolution using dense skip connections. In: ICCV (2017) 1, 4", + "29. Wang, X., Yu, K., Wu, S., Gu, J., Liu, Y., Dong, C., Qiao, Y., Change Loy, C.: Esrgan: Enhanced super-resolution generative adversarial networks. In: ECCVW (2018) 1, 2, 4, 9, 10", + "30. Yang, J., Wright, J., Huang, T., Ma, Y.: Image super-resolution as sparse representation of raw image patches. In: CVPR (2008) 4", + "31. Yang, Q., Chen, D., Tan, Z., Liu, Q., Chu, Q., Bao, J., Yuan, L., Hua, G., Yu, N.: Hq-50k: A large-scale, high-quality dataset for image restoration. arXiv preprint arXiv:2306.05390 (2023) 2, 5", + "32. Zeyde, R., Elad, M., Protter, M.: On single image scale-up using sparse-representations. In: Proc. 7th Int. Conf. Curves Surf. (2010) 9", + "33. Zhang, D., Huang, F., Liu, S., Wang, X., Jin, Z.: Swinfir: Revisiting the swinir with fast fourier convolution and improved training for image super-resolution. arXiv preprint arXiv:2208.11247 (2022) 1, 4", + "34. Zhang, X., Zeng, H., Guo, S., Zhang, L.: Efficient long-range attention network for image super-resolution. In: ECCV (2022) 1, 4", + "35. Zhang, Y., Li, K., Zhong, B., Fu, Y.: Residual non-local attention networks for image restoration. In: ICLR (2019) 1, 4", + "36. Zhang, Y., Li, K., Li, K., Wang, L., Zhong, B., Fu, Y.: Image super-resolution using very deep residual channel attention networks. In: ECCV (2018) 1, 4, 9", + "37. Zhang, Y., Tian, Y., Kong, Y., Zhong, B., Fu, Y.: Residual dense network for image super-resolution. In: CVPR (2018) 1, 4" + ], + "bbox": [ + 215, + 146, + 785, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "G. Ohtani et al.", + "bbox": [ + 271, + 114, + 380, + 127 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "38. Zhang, Y., Zhang, K., Chen, Z., Li, Y., Timofte, R., Zhang, J., Zhang, K., Peng, R., Ma, Y., Jia, L., et al.: Ntire 2023 challenge on image super-resolution (x4): Methods and results. In: CVPRW (2023) 2", + "39. Zhou, B., Lapedriza, A., Xiao, J., Torralba, A., Oliva, A.: Learning deep features for scene recognition using places database. Advances in neural information processing systems 27 (2014) 8", + "40. Zhou, S., Zhang, J., Zuo, W., Loy, C.C.: Cross-scale internal graph neural network for image super-resolution. Advances in neural information processing systems 33, 3499-3509 (2020) 10", + "41. Zhou, X., Girdhar, R., Joulin, A., Krahenbuhl, P., Misra, I.: Detecting twenty-thousand classes using image-level supervision. In: ECCV (2022) 8", + "42. Zhou, Y., Li, Z., Guo, C.L., Bai, S., Cheng, M.M., Hou, Q.: Srformer: Permuted self-attention for single image super-resolution. In: ICCV (2023) 1, 2, 4" + ], + "bbox": [ + 212, + 146, + 787, + 327 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "Rethinking Image Super-Resolution from Training Data Perspectives", + "bbox": [ + 272, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + } +] \ No newline at end of file diff --git a/2024/Rethinking Image Super Resolution from Training Data Perspectives/8d3fa075-356f-4f4c-b1a5-4318bed20e1d_model.json b/2024/Rethinking Image Super Resolution from Training Data Perspectives/8d3fa075-356f-4f4c-b1a5-4318bed20e1d_model.json new file mode 100644 index 0000000000000000000000000000000000000000..8b619a78f949f389b20ae6311d54d88e52a74285 --- /dev/null +++ b/2024/Rethinking Image Super Resolution from Training Data Perspectives/8d3fa075-356f-4f4c-b1a5-4318bed20e1d_model.json @@ -0,0 +1,2346 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.267, + 0.141, + 0.737, + 0.187 + ], + "angle": 0, + "content": "Rethinking Image Super-Resolution from Training Data Perspectives" + }, + { + "type": "text", + "bbox": [ + 0.229, + 0.212, + 0.775, + 0.259 + ], + "angle": 0, + "content": "Go Ohtani\\(^{1,2}\\), Ryu Tadokoro\\(^{2}\\), Ryosuke Yamada\\(^{2,3}\\), Yuki M. Asano\\(^{4}\\), Iro Laina\\(^{5}\\), Christian Rupprecht\\(^{5}\\), Nakamasasa Inoue\\(^{6,2}\\), Rio Yokota\\(^{6,2}\\), Hirokatsu Kataoka\\(^{2}\\), and Yoshimitsu Aoki\\(^{1}\\)" + }, + { + "type": "text", + "bbox": [ + 0.443, + 0.269, + 0.562, + 0.285 + ], + "angle": 0, + "content": "1 Keio University" + }, + { + "type": "text", + "bbox": [ + 0.249, + 0.284, + 0.753, + 0.299 + ], + "angle": 0, + "content": "\\(^{2}\\) National Institute of Advanced Industrial Science and Technology (AIST)" + }, + { + "type": "text", + "bbox": [ + 0.421, + 0.298, + 0.584, + 0.312 + ], + "angle": 0, + "content": "3 University of Tsukuba" + }, + { + "type": "text", + "bbox": [ + 0.259, + 0.312, + 0.745, + 0.326 + ], + "angle": 0, + "content": "4 University of Amsterdam, now at University of Technology Nuremberg" + }, + { + "type": "text", + "bbox": [ + 0.425, + 0.326, + 0.578, + 0.339 + ], + "angle": 0, + "content": "5 University of Oxford" + }, + { + "type": "text", + "bbox": [ + 0.394, + 0.34, + 0.609, + 0.354 + ], + "angle": 0, + "content": "\\(^{6}\\) Tokyo Institute of Technology" + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.387, + 0.744, + 0.61 + ], + "angle": 0, + "content": "Abstract. In this work, we investigate the understudied effect of the training data used for image super-resolution (SR). Most commonly, novel SR methods are developed and benchmarked on common training datasets such as DIV2K and DF2K. However, we investigate and rethink the training data from the perspectives of diversity and quality, thereby addressing the question of \"How important is SR training for SR models?\" To this end, we propose an automated image evaluation pipeline. With this, we stratify existing high-resolution image datasets and larger-scale image datasets such as ImageNet and PASS to compare their performances. We find that datasets with (i) low compression artifacts, (ii) high within-image diversity as judged by the number of different objects, and (iii) a large number of images from ImageNet or PASS all positively affect SR performance. We hope that the proposed simple-yet-effective dataset curation pipeline will inform the construction of SR datasets in the future and yield overall better models. Code is available at: https://github.com/gohtanii/DiverSeg-dataset" + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.622, + 0.741, + 0.65 + ], + "angle": 0, + "content": "Keywords: Super-resolution dataset \\(\\cdot\\) Image compression \\(\\cdot\\) Image diversity" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.674, + 0.377, + 0.691 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.704, + 0.789, + 0.841 + ], + "angle": 0, + "content": "Image super-resolution (SR) aims to reconstruct high-resolution images from low-resolution images. It has been considered as one of the most fundamental tasks in the field of computer vision, with applications ranging from autonomous driving to medical imaging. Deep learning methods have led to significant advances in SR over the last decade, focusing primarily on improvements in the neural network architectures. Early SR models rely on convolutional neural networks (CNNs) [9,10,12,14,20,26,28,29,35-37]. Recent innovations have given rise to transformer-based SR models [5,6,16,19,33,34,42], which have consistently improved the performances." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.38, + 0.128 + ], + "angle": 0, + "content": "G. Ohtani et al." + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.146, + 0.775, + 0.423 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.435, + 0.788, + 0.493 + ], + "angle": 0, + "content": "Fig. 1: We propose an automated image evaluation pipeline to curate a dataset for training SR models. The obtained dataset, namely DiverSeg, consists of low-resolution but high-quality images with many object regions. SR models trained on DiverSeg outperform those trained on high-resolution image datasets such as DF2K and LSDIR." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.521, + 0.788, + 0.627 + ], + "angle": 0, + "content": "With the improvement of neural network architectures, the importance of training datasets has also increased, as discussed in [17]. Examples of high-resolution datasets include DIV2K [1] and Flickr2K [27]. The combined dataset of these two, referred to as DF2K, is often utilized for training SR models. Most recently, LSDIR [17] has been proposed, which consists of 84,991 high-resolution images. It has been confirmed that training on large high-resolution datasets contributes significantly to performance improvement [18, 19, 29, 38, 42]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.627, + 0.788, + 0.657 + ], + "angle": 0, + "content": "The conventional approach to constructing datasets relies on a manual evaluation step, where the following two perspectives are most commonly considered:" + }, + { + "type": "text", + "bbox": [ + 0.222, + 0.665, + 0.788, + 0.741 + ], + "angle": 0, + "content": "1. Resolution and quality [17, 31]. This perspective focuses on the pixel density of images. Images that do not meet the specified resolution threshold are excluded. Typically, HD, 2K, and 4K images are used to create a dataset. After the initial automatic filtering based on image size, the details of each image are manually evaluated to identify and exclude compressed images." + }, + { + "type": "text", + "bbox": [ + 0.222, + 0.741, + 0.788, + 0.802 + ], + "angle": 0, + "content": "2. Variety and diversity [15]. This includes diversity in subjects (e.g., people, landscapes, urban scenes), lighting conditions, colors, textures, and other photographic elements. A diverse dataset is said to help train a model that is robust and performs well across a wide range of domains." + }, + { + "type": "list", + "bbox": [ + 0.222, + 0.665, + 0.788, + 0.802 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Datasets constructed from these perspectives have been shown to significantly improve the performance of SR models. However, they also pose challenges in" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.273, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Rethinking Image Super-Resolution from Training Data Perspectives" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.178 + ], + "angle": 0, + "content": "scaling the datasets, as collecting uncompressed high-resolution images is difficult and costly." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.184, + 0.788, + 0.304 + ], + "angle": 0, + "content": "To address this limitation, this paper rethinks these perspectives and proposes Diverse Segmentation dataset (DiverSeg), a low-resolution1 yet effective image dataset for training SR models. As shown in Figure 1, the dataset is constructed by applying filtering to a large set of low-resolution images, such as ImageNet-1k [8] and PASS [2]. In experiments, we demonstrate that models trained on DiverSeg outperform those trained on high-resolution image datasets such as DF2K and LSDIR. Based on this finding, our contributions are summarized as follows." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.311, + 0.788, + 0.448 + ], + "angle": 0, + "content": "1) Rethinking the resolution perspective. High-resolution images have been considered to be necessary for training SR models. In this work, we challenge this traditional perspective and show that SR models can be trained without high-resolution images. Specifically, we introduce a method to estimate image quality based on the kernel density estimation over blockiness values [4] that estimates the quantity distribution of blocking artifacts. We demonstrate that low-resolution images with high quality, indicated by reduced artifacts, can improve the performance of SR models. We also thoroughly analyze the impact of image quality on SR performance." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.453, + 0.788, + 0.575 + ], + "angle": 0, + "content": "2) Rethinking the diversity perspective. When constructing datasets for training SR models, images containing only a small number of objects are often implicitly excluded during the manual resolution evaluation process. This is because evaluators typically focus on the details of objects or small objects in the images. Therefore, we explicitly calculate the number of objects in images and analyze how this number affects SR performance. In our experiments, we show that constructing datasets with images containing many objects improves the performance of SR models." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.58, + 0.788, + 0.793 + ], + "angle": 0, + "content": "3) Dataset construction. To facilitate analysis from the above two perspectives, we introduce a framework that automatically creates a dataset from a set of low-resolution images collected from the web. Specifically, the framework consists of two steps, source selection and object-based filtering, which correspond to the first and second perspectives, respectively. Our framework eliminates the need for manual assessment, facilitating dataset scaling. In the first step, given several low-resolution image datasets, we estimate the image quality of them through their blockiness distributions and select datasets with quality larger than \\(90\\%\\). In the second step, we filter out images with a small number of object regions by using object detection and image segmentation models. The filtered dataset, which we refer to as the DiverSeg dataset, is utilized for training various SR models in our experiments. We apply our framework to a union set of ImageNet, Places365, and PASS to construct the DiverSeg dataset. We demonstrate that SR models trained on DiverSeg archive state-of-the-art performance." + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.311, + 0.788, + 0.793 + ], + "angle": 0, + "content": null + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.825, + 0.738, + 0.84 + ], + "angle": 0, + "content": "1 We define images with a resolution lower than HD as low-resolution images." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.381, + 0.129 + ], + "angle": 0, + "content": "G. Ohtani et al." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.145, + 0.39, + 0.161 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.179, + 0.517, + 0.195 + ], + "angle": 0, + "content": "2.1 Image super-resolution models" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.204, + 0.787, + 0.279 + ], + "angle": 0, + "content": "A number of SR models have been proposed that take advantage of deep learning techniques. These models can be divided into two groups, CNN-based models [9, 10,12,14,20,26,28,29,35-37] and Transformer-based models [5,6,16,19,33,34,42]. Each group exploits different architectural strengths to enhance low-resolution images to high resolution." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.28, + 0.787, + 0.505 + ], + "angle": 0, + "content": "CNN-based models. SRCNN [9] was the first model that integrates a deep convolutional architecture for SR. Subsequently, FSRCNN [10] significantly improved computational efficiency by performing convolutional processing in low resolution space and upsampling in the last layer. Furthermore, ESPCN [26] adopted an efficient upsampling method, sub-pixel convolution, to enhance performance while reducing computational costs. These methods established the basic structure of modern SR networks and laid the foundation for subsequent research developments. Later studies introduced various modules such as residual connections [12, 14, 20, 29], dense blocks [28, 29, 37], and attention mechanisms [7, 35, 36]. EDSR [20] eliminated batch normalization layers and introduced residual scaling to enable stable training of large models. MSRResNet [29] replaced the basic ResNet blocks with residual dense blocks to improve the balance between performance and computational efficiency. RCAN [36] incorporated channel attention mechanisms to adaptively weight feature representations, significantly enhancing SR performance." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.507, + 0.789, + 0.703 + ], + "angle": 0, + "content": "Transformer-based models. As the first Transformer-based image restoration model, IPT [5] was introduced as a large-scale model utilizing the Transformer's encoder and decoder architecture. By pre-training on ImageNet, IPT significantly improved SR performance, fully leveraging the capabilities of the Transformer. SwinIR [19] executes self-attention within local windows during feature extraction, demonstrating exceptional SR performance and establishing itself as the foundational model for Transformers in SR. Following this, models that build upon SwinIR have been developed [6,33,34,42], enhancing performance by extending and optimizing the self-attention mechanism. Among them, HAT [6] achieves state-of-the-art performance in SR by using a duplicated cross-attention module and pre-training on ImageNet. However, there may be potential for improvement, as ImageNet is primarily an image recognition dataset and may not be fully optimized for SR." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.725, + 0.527, + 0.74 + ], + "angle": 0, + "content": "2.2 Image super-resolution datasets" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.787, + 0.841 + ], + "angle": 0, + "content": "T91 [30] and BSDS200 [22] are early datasets used to train SR models, consisting of 91 and 200 images, respectively. The turning point in training datasets was ushered by the release of DIV2K [1], a compilation of 800 high-resolution images with minimal compression noise meticulously collected from the web. Subsequently, to accommodate the further scaling of the model, Flickr2K [27], consisting of 2,650 high-resolution images, is merged into DIV2K and is referred" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.273, + 0.115, + 0.731, + 0.129 + ], + "angle": 0, + "content": "Rethinking Image Super-Resolution from Training Data Perspectives" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.146, + 0.784, + 0.284 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.298, + 0.785, + 0.327 + ], + "angle": 0, + "content": "Fig. 2: Images of DiverSeg with their segmentation masks. DiverSeg is obtained from a large set of low-resolution images through the automated image evaluation pipeline." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.354, + 0.788, + 0.535 + ], + "angle": 0, + "content": "to as DF2K. In recent years, to further expand the scale of SR datasets, datasets larger than DF2K have been released, such as LSDIR [17], comprising 84,991 images and HQ-50K [31] consisting of 50,000 images. These datasets persistently adhere to stringent criteria, ensuring high resolution and negligible compression noise in the imagery. However, since collecting images that meet the aforementioned conditions is quite challenging, these datasets still consist of only tens of thousands of images. In contrast, several approaches [5, 6, 16] adopt ImageNet, also known as ImageNet-1k, which consists of 1.28M images spanning diverse categories. These approaches leverage the diversity of texture patterns in ImageNet as an advantage. Nevertheless, ImageNet is pointed out to contain some images having low-resolution and JPEG-compression artifacts, which adversely affect the training results." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.557, + 0.416, + 0.573 + ], + "angle": 0, + "content": "3 DiverSeg dataset" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.586, + 0.788, + 0.707 + ], + "angle": 0, + "content": "This section presents Diverse Segmentation dataset (DiverSeg), our dataset for training SR models without using high-resolution images. As shown in Figure 2, the DiverSeg dataset consists of low-resolution but high-quality images with diverse segmented object regions. The framework for constructing the dataset consists of two steps, source selection and object-based filtering, which are designed from our perspective of rethinking resolution and diversity, respectively. Our approach eliminates the manual cost of collecting and quality-checking high-resolution images." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.727, + 0.396, + 0.74 + ], + "angle": 0, + "content": "3.1 Source selection" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Let \\(\\mathcal{X}\\) be a set of low-resolution image datasets. This step filters out low-quality datasets by estimating the quality \\(\\hat{q}_X \\in [0,1]\\) for each dataset \\(X \\in \\mathcal{X}\\) and excluding those with \\(\\hat{q}_X < 0.9\\), under the assumption that low-quality images are detrimental when training SR models. We introduce a quality estimation method based on the Kullback-Leibler (KL) divergence between blockiness distributions as detailed below." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.38, + 0.128 + ], + "angle": 0, + "content": "G. Ohtani et al." + }, + { + "type": "image", + "bbox": [ + 0.224, + 0.144, + 0.788, + 0.257 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.261, + 0.789, + 0.304 + ], + "angle": 0, + "content": "Fig. 3: Comparison of image degradation due to JPEG quality(blue). Blockiness values calculated from the images are marked. As the JPEG quality decreases and artifacts increase, we observe a corresponding rise in blockiness values." + }, + { + "type": "image", + "bbox": [ + 0.222, + 0.319, + 0.501, + 0.457 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.319, + 0.785, + 0.457 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.467, + 0.788, + 0.511 + ], + "angle": 0, + "content": "Fig. 4: (a) Blockiness distributions \\( p_{X,1.0} \\) for \\( X = \\mathrm{ImageNet - 1k} \\), Places365 and PASS. (b) Basis distributions \\( p_{Z,q} \\) for \\( Z = \\mathrm{DF2K} \\) and \\( q = 0.5, 0.75, 0.85, 0.95, 1.0 \\). We estimate the quality by comparing \\( p_{X,1.0} \\) and \\( p_{Z,q} \\) using the KL divergence." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.54, + 0.799, + 0.6 + ], + "angle": 0, + "content": "Image datasets. This work uses three web-collected low-resolution datasets \\(\\mathcal{X} = \\{\\mathrm{ImageNet - 1k},\\mathrm{Places365},\\mathrm{PASS}\\}\\). Training SR models on them is not straightforward because they may include highly compressed images that negatively affect training of SR models." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.6, + 0.789, + 0.678 + ], + "angle": 0, + "content": "Quality definition. Let \\( Y \\) be a dataset of JPEG images. We define the quality \\( q_{Y} \\) by the average JPEG quality, i.e., \\( q_{Y} \\triangleq \\frac{1}{|Y|} \\sum_{y \\in Y} Q(y) \\) where \\( 0 \\leq Q(y) \\leq 1 \\) is the JPEG quality of an image \\( y \\). The goal of quality estimation is to estimate \\( q_{X} \\) given a dataset \\( X \\). Note that in the estimation phase, datasets may include images other than JPEG and the true quality \\( q_{X} \\) is not observable." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.677, + 0.789, + 0.725 + ], + "angle": 0, + "content": "Blockiness distribution. To estimate the quality, we utilize the blockiness measure [4]. Specifically, for each image dataset \\( X \\in \\mathcal{X} \\), we estimate the distribution of blockiness values \\( p_{X,q}(b) \\) by kernel density estimation as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.353, + 0.745, + 0.787, + 0.784 + ], + "angle": 0, + "content": "\\[\np _ {X, q} (b) = \\frac {1}{h | X |} \\sum_ {x \\in X} K \\left(\\frac {b - B \\left(\\phi_ {q} (x)\\right)}{h}\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.789, + 0.842 + ], + "angle": 0, + "content": "where \\( x \\) is an image, \\( \\phi_q \\) is the JPEG compression function, \\( q \\in [0,1] \\) is a quality value, \\( K: \\mathbb{R} \\to \\mathbb{R} \\) is a Gaussian kernel, and \\( h \\in \\mathbb{R} \\) is the bandwidth determined by Scott's method. The function \\( B \\) is the blockiness measure that measures the" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.273, + 0.115, + 0.733, + 0.131 + ], + "angle": 0, + "content": "Rethinking Image Super-Resolution from Training Data Perspectives" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.178 + ], + "angle": 0, + "content": "quantity of blocking artifacts by computing subband DCT coefficients. Specifically, \\( B \\) is defined on images that are decomposed into \\( P \\times P \\) patches as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.235, + 0.202, + 0.786, + 0.245 + ], + "angle": 0, + "content": "\\[\nB (x) = \\sum_ {i = 1} ^ {P} \\sum_ {j = 1} ^ {P} \\left| \\frac {\\bar {V} _ {\\text {c r o p}} (i , j) - \\bar {V} (i , j)}{\\bar {V} (i , j)} \\right|, \\quad \\bar {V} (i, j) = \\frac {1}{W H} \\sum_ {h = 1} ^ {H} \\sum_ {w = 1} ^ {W} V _ {w, h} (i, j) \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.256, + 0.788, + 0.378 + ], + "angle": 0, + "content": "where \\( W, H \\) are the width and height of an image \\( x \\), \\( V_{w,h}(i,j) \\) is the variation in the \\( (i,j) \\)-th subband DCT coefficients within the \\( (h,w) \\)-th patch being calculated and its four spatially adjacent patches. \\( V_{\\mathrm{crop}}(i,j) \\) is the variation calculated similarly to \\( V_{w,h}(i,j) \\) for the given image with the first 4 rows and 4 columns removed. \\( \\bar{V} \\) and \\( \\bar{V}_{\\mathrm{crop}} \\) are the average variations of \\( V_{w,h}(i,j) \\) and \\( V_{\\mathrm{crop}}(i,j) \\) calculated for each patch, respectively. This work uses \\( P = 8 \\). The blockiness value \\( B(x) \\) is expected to be low for uncompressed images and high for compressed images as shown in Figure 3." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.378, + 0.8, + 0.425 + ], + "angle": 0, + "content": "Quality estimation. We estimate the quality by comparing \\( p_{X,1.0} \\) with \\( \\{p_{Z,q}\\}_{q \\in S} \\), where \\( p_{Z,q} \\) is a basis distribution, a blockiness distribution of images of a fixed quality \\( q \\). More specifically, the estimated quality is given by" + }, + { + "type": "equation", + "bbox": [ + 0.342, + 0.435, + 0.786, + 0.473 + ], + "angle": 0, + "content": "\\[\n\\hat {q} _ {X} = \\sum_ {q \\in S} q \\frac {\\exp \\left(- D _ {\\mathrm {K L}} \\left(p _ {X , 1 . 0} \\mid \\mid p _ {Z , q}\\right)\\right)}{\\sum_ {q ^ {\\prime} \\in S} \\exp \\left(- D _ {\\mathrm {K L}} \\left(p _ {X , 1 . 0} \\mid \\mid p _ {Z , q ^ {\\prime}}\\right)\\right)}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.485, + 0.788, + 0.56 + ], + "angle": 0, + "content": "where \\( Z \\) is a small dataset that involves only uncompressed images. In this work, we use DF2K. \\( D_{\\mathrm{KL}} \\) is the KL divergence, and \\( S = \\{1.0, 0.95, 0.85, 0.75, 0.5\\} \\) are discretely sampled quality values. Figure 4 shows the blockiness distributions \\( p_{X,1.0} \\) for the three low-resolution image datasets and the basis distributions \\( p_{Z,c} \\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.561, + 0.788, + 0.605 + ], + "angle": 0, + "content": "Selection results. The estimated qualities for ImageNet-1k, Places365, and PASS were \\(95.5\\%\\), \\(75.0\\%\\) and \\(99.8\\%\\), respectively. From this result, Places365 is filtered out." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.631, + 0.444, + 0.648 + ], + "angle": 0, + "content": "3.2 Object-based filtering" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.658, + 0.788, + 0.734 + ], + "angle": 0, + "content": "Given a source dataset \\( X \\), this step applies filtering to refine it as a dataset for training SR models, under the assumption that images with diverse object regions are more effective than those with uniform or monotonous content. Specifically, the refined training dataset is given by \\( \\tilde{X} = \\{x \\in X : R(x) \\geq \\theta\\} \\), where \\( R(x) \\) is the number of object regions and \\( \\theta \\) is a threshold." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.734, + 0.788, + 0.795 + ], + "angle": 0, + "content": "We introduce two object-based filtering methods with different granularities to explore how the granularity of object detection, ranging from detailed identification of small objects or features to recognizing larger, more general objects, impacts SR performance." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Segmentation-based filtering. This method counts the number of object parts by an image segmentation model. Specifically, we adopt the SAM [13] with the ViT-H backbone and define \\( R \\) by the number of segmentation masks." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.381, + 0.129 + ], + "angle": 0, + "content": "G. Ohtani et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.788, + 0.201 + ], + "angle": 0, + "content": "Table 1: Dataset statistics. HR: High resolution, LR: Low resolution, #Images: Number of images, #Pixels: Average number of pixels per image, Blockiness: median of blockiness measure indicating the intensity of JPEG compression noise, #Segments: Average number of segmentation masks." + }, + { + "type": "table", + "bbox": [ + 0.222, + 0.213, + 0.784, + 0.374 + ], + "angle": 0, + "content": "
DatasetTaskHR LR#Images#PixelsBlockiness#Segments
DIV2K [1]Super-resolution✓8002.8M0.47104
DF2K [1,27]Super-resolution✓3,4502.8M0.47103
LSDIR [17]Super-resolution✓84,9911.1M0.8292
ImageNet [8]Image recognition1,281,167237k4.3971
Places365 [39]Image recognition1,803,460366k80.71100
PASS [2]Image recognition1,439,589178k3.0374
DiverSeg-I (Ours)Super-resolution259,448233k2.83146
DiverSeg-P (Ours)Super-resolution267,055179k4.39146
DiverSeg-IP (Ours)Super-resolution526,503206k3.61146
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.406, + 0.784, + 0.481 + ], + "angle": 0, + "content": "We chose SAM because, unlike typical segmentation models that perform semantic segmentation based on class labels, SAM provides segments that do not impose semantic constraints, allowing for finer region segmentation of diverse objects. We set \\(\\theta = 100\\) as the default, resulting in \\(260\\mathrm{k}\\) images remaining after applying the filter to ImageNet-1k." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.483, + 0.787, + 0.545 + ], + "angle": 0, + "content": "Detection-based filtering. This method counts the number of objects per image. We adopt the Detic model [41] with the ViT-B backbone and define \\( R \\) by the number of detected objects. We set \\( \\theta = 18 \\), resulting in 260k images remaining after applying the filter to ImageNet-1k." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.573, + 0.406, + 0.588 + ], + "angle": 0, + "content": "3.3 Dataset statistics" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.605, + 0.788, + 0.713 + ], + "angle": 0, + "content": "Table 1 shows dataset statistics. We created three variants of DiverSeg, namely DiverSeg-I, DiverSeg-P, and DiverSeg-IP, constructed from ImageNet-1k, PASS, and the union of the two, respectively. The segmentation-based filtering is applied to obtain these datasets. Compared to high-resolution datasets such as DF2K and LSDIR, our datasets have larger number of training images, but contain only low-resolution images. The median of blockiness values is decreased for ImageNet-1k and is increased for PASS after filtering." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.741, + 0.377, + 0.759 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.787, + 0.842 + ], + "angle": 0, + "content": "This section conducts experiments by training various SR models on DiverSeg datasets. We demonstrate that SR models can be trained without using high-resolution images and thoroughly analyze factors from a dataset perspective that are crucial for enhancing the performance of SR." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.273, + 0.115, + 0.732, + 0.13 + ], + "angle": 0, + "content": "Rethinking Image Super-Resolution from Training Data Perspectives" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "table_caption", + "bbox": [ + 0.215, + 0.145, + 0.784, + 0.187 + ], + "angle": 0, + "content": "Table 2: Comparison of DiverSeg with high-resolution image datasets (DF2K and LSDIR). Models trained on DiverSeg-I and DiverSeg-P demonstrated superior performance despite not using any high-resolution images for training." + }, + { + "type": "table", + "bbox": [ + 0.22, + 0.2, + 0.784, + 0.461 + ], + "angle": 0, + "content": "
Model (Params)DatasetHR LRSet5Set14BSD100Urban100Manga109
PSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIM
MSRResNet [29] (1.5M)DF2K✓32.230.895528.670.783127.620.737426.230.789730.640.9108
LSDIR✓32.150.894828.660.783627.620.737426.310.791830.570.9105
DiverSeg-I✓32.270.896328.640.783727.640.737826.310.791830.530.9115
DiverSeg-P✓32.090.894328.610.783227.600.737126.280.791830.360.9101
RCAN [36] (15.5M)DF2K✓32.500.899028.870.788527.750.742126.730.805831.170.9165
LSDIR✓32.530.899228.890.789427.750.742526.910.809031.330.9180
DiverSeg-I✓32.700.901228.980.790827.810.744327.030.811631.580.9210
DiverSeg-P✓32.630.900028.950.789827.770.743526.990.813431.190.9190
EDSR [20] (43.0M)DF2K✓32.610.899828.910.789327.790.743426.840.808931.380.9176
LSDIR✓32.570.899228.970.790827.800.743827.050.813131.470.9192
DiverSeg-I✓32.710.901728.980.791327.850.745327.100.814231.720.9216
DiverSeg-P✓32.570.900229.060.791527.800.744727.100.816331.330.9191
SwinIR [19] (11.9M)DF2K✓32.920.904429.090.795027.920.748927.450.825432.030.9260
LSDIR✓32.860.903629.160.796327.920.749227.790.833131.980.9262
DiverSeg-I✓32.970.905329.230.797027.980.750827.830.833632.340.9283
DiverSeg-P✓32.850.904029.240.796127.960.750227.850.834932.280.9278
HAT [6] (20.7M)DF2K✓33.030.905629.160.796427.990.751427.930.836532.440.9292
LSDIR✓32.930.905329.290.798828.010.752528.450.846932.570.9306
DiverSeg-I✓33.150.907129.460.800428.070.754228.510.847732.900.9325
DiverSeg-P✓33.120.906829.500.800228.040.753628.530.849232.830.9320
" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.488, + 0.443, + 0.504 + ], + "angle": 0, + "content": "4.1 Experimental settings" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.51, + 0.784, + 0.556 + ], + "angle": 0, + "content": "SR models. We use five models. Specifically, we use three CNN-based models: MSRResNet [29], EDSR [20], RCAN [36], and two Transformer-based models: SwinIR [19] and HAT [6]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.556, + 0.784, + 0.601 + ], + "angle": 0, + "content": "Training datasets. We compare the DiverSeg datasets with two high-resolution datasets: DF2K and LSDIR [17]. DF2K is a merged dataset of DIV2K [1] and Flickr2K [27]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.601, + 0.784, + 0.632 + ], + "angle": 0, + "content": "Evaluation datasets. We use five benchmark datasets: Set5 [3], Set14 [32], BSD100 [22], Urban100 [11], and Manga109 [23]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.632, + 0.784, + 0.662 + ], + "angle": 0, + "content": "Evaluation metrics. PSNR and SSIM on the Y channel (representing luminance) within the transformed YCbCr color space are used as evaluation metrics." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.662, + 0.784, + 0.707 + ], + "angle": 0, + "content": "Implementation settings. We follow the training setup of the original papers of the SR models [6, 19, 20, 29, 36]. Implementation details are provided in the supplementary materials." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.727, + 0.434, + 0.743 + ], + "angle": 0, + "content": "4.2 Experimental results" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.785, + 0.84 + ], + "angle": 0, + "content": "Main results. To validate the effectiveness of our approach, we trained the five SR models on DiverSeg-I/P. The results are summarized in Table 2. As shown, models trained on DiverSeg-I/P achieved better performance than those trained on DF2K and LSDIR. This shows the effectiveness of the proposed datasets. To the best of our knowledge, we are the first to successfully train SR models without using high-resolution images." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.381, + 0.128 + ], + "angle": 0, + "content": "G. Ohtani et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.788, + 0.189 + ], + "angle": 0, + "content": "Table 3: Quantitative comparison with state-of-the-art methods on five benchmark datasets. We applied our dataset to two Transformer-based models. Checkmarks for HR, LR indicate the use of high-resolution and low-resolution datasets, respectively." + }, + { + "type": "table", + "bbox": [ + 0.22, + 0.203, + 0.784, + 0.428 + ], + "angle": 0, + "content": "
MethodTraining DataHR LRSet5Set14BSD100Urban100Manga109
PSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIM
SAN [7]DIV2K✓32.640.900328.920.788827.780.743626.790.806831.180.9169
IGNN [40]DIV2K✓32.570.899828.850.789127.770.743426.840.809031.280.9182
HAN [25]DIV2K✓32.640.900228.900.789027.800.744226.850.809431.420.9177
NLSN [24]DIV2K✓32.590.900028.870.789127.780.744426.960.810931.270.9184
RRDB [29]DF2K✓32.730.901128.990.791727.850.745527.030.815331.660.9196
RCAN-it [21]DF2K✓32.690.900728.990.792227.870.745927.160.816831.780.9217
EDT [16]DF2K✓32.820.903129.090.793927.910.748327.460.824632.050.9254
HAT-S [6]DF2K✓32.920.904729.150.795827.970.750527.870.834632.350.9283
IPT [5]ImageNet✓32.64-29.01-27.82-27.26---
SwinIR [19]DF2K✓32.920.904429.090.795027.920.748927.450.825432.030.9260
SwinIR [19]DiverSeg-I (Ours)✓32.970.905329.230.797027.980.750827.830.833632.340.9283
HAT [6]DF2K✓33.040.905629.230.797328.000.751727.970.836832.480.9292
HAT [6]ImageNet→DF2K✓33.180.907329.380.800128.050.753428.370.844732.870.9319
HAT [6]DiverSeg-I (Ours)✓33.150.907129.460.800428.060.754228.510.847732.900.9325
HAT [6]DiverSeg-IP (Ours)✓33.140.907329.510.800728.070.754228.540.849232.960.9327
HAT-L [6]ImageNet→DF2K✓33.300.908329.470.801528.090.755128.600.849833.090.9335
HAT-L [6]DiverSeg-I (Ours)✓33.280.908329.540.802228.100.755628.750.852933.140.9340
HAT-L [6]DiverSeg-IP (Ours)✓33.200.908029.590.801928.110.755628.810.854733.190.9342
" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.441, + 0.788, + 0.469 + ], + "angle": 0, + "content": "Table 4: Filtering by blockiness (HAT model, ImageNet-1k). \\(\\theta^{\\prime}\\): threshold for blockiness values." + }, + { + "type": "table", + "bbox": [ + 0.238, + 0.475, + 0.772, + 0.567 + ], + "angle": 0, + "content": "
θ'#ImagesSet5Set14BSD100Urban100Manga109
PSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIM
10800k33.070.906829.340.800028.040.753828.410.846532.560.9310
30939k33.130.907229.390.800128.060.753828.430.846332.700.9316
1001.08M33.110.907129.420.800128.050.753728.410.845732.840.9321
-1.15M33.080.907129.400.799928.050.753528.410.845732.880.9323
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.594, + 0.789, + 0.732 + ], + "angle": 0, + "content": "Comparison with SOTA. Table 3 shows that DiverSeg improves the state-of-the-art performance. Specifically, the HAT and HAT-L models trained on DiverSeg datasets outperformed those trained with ImageNet-1k \\(\\rightarrow\\) DF2K, which utilizes all ImageNet-1k images for pre-training and the DF2K images for fine-tuning, in terms of PSNR and SSIM on four of five benchmarking datasets. It is worth noting that DiverSeg-I filters out \\(77.5\\%\\) of images from ImageNet-1k and thus training on it is more efficient than the approach relying on pre-training and fine-tuning. This confirmed the effectiveness and efficiency of our filtering approach." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.754, + 0.507, + 0.77 + ], + "angle": 0, + "content": "4.3 Analysis 1: Effects of filtering" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.789, + 0.842 + ], + "angle": 0, + "content": "Filtering by blockiness measure. If the blockiness is an important factor for enchaining SR performance, filtering images by the blockiness values, i.e., constructing a training dataset by \\(\\hat{X} = \\{x\\in X:B(x)\\leq \\theta '\\}\\), would be a straightforward approach. However, as shown in Table 4, this filtering did not" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.273, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Rethinking Image Super-Resolution from Training Data Perspectives" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.146, + 0.525, + 0.188 + ], + "angle": 0, + "content": "Table 5: Comparison of filtering methods. Performance is evaluated using \\(260\\mathrm{k}\\) filtered images (HAT model, ImageNet-1k)." + }, + { + "type": "table", + "bbox": [ + 0.228, + 0.192, + 0.516, + 0.27 + ], + "angle": 0, + "content": "
Filtering methodUrban100Manga109
PSNRSSIMPSNRSSIM
Blockiness28.390.846732.470.9304
Detection-based28.440.846232.870.9322
Seg.-based28.510.847732.900.9325
" + }, + { + "type": "table_caption", + "bbox": [ + 0.555, + 0.145, + 0.785, + 0.187 + ], + "angle": 0, + "content": "Table 6: Performance comparison across different thresholds \\( \\theta \\) (HAT model, ImageNet-1k)." + }, + { + "type": "table", + "bbox": [ + 0.56, + 0.192, + 0.789, + 0.27 + ], + "angle": 0, + "content": "
θ#ImagesUrban100Manga109
PSNRSSIMPSNRSSIM
01.2M28.410.845732.880.9323
50663k28.460.847232.900.9325
100259k28.510.847732.900.9325
15086k28.360.845232.830.9320
" + }, + { + "type": "table_caption", + "bbox": [ + 0.301, + 0.284, + 0.699, + 0.299 + ], + "angle": 0, + "content": "Table 7: Analysis of effects of JPEG quality (HAT model)." + }, + { + "type": "table", + "bbox": [ + 0.221, + 0.304, + 0.784, + 0.451 + ], + "angle": 0, + "content": "
DatasetQuality (%)BlockinessSet5Set14BSD100Urban100Manga109
PSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIM
DF2K50165.3431.130.885527.540.753326.320.710325.040.751130.200.9024
7583.2032.630.900827.700.745727.700.745727.390.823831.880.9238
8546.9832.940.904329.070.794127.930.749927.720.831332.140.9266
9510.3332.980.904829.110.795127.990.752127.810.833832.210.9278
HR0.4733.030.905629.160.796427.990.751427.930.836532.440.9292
LSDIR50146.4328.500.848726.660.742124.720.667424.910.765729.390.8956
7557.1431.870.885328.530.780527.490.741727.300.823131.830.9223
8524.1432.840.904429.170.795527.980.751328.220.841232.380.9291
953.7132.970.904329.250.797928.050.753928.370.845232.420.9299
HR0.8232.930.905329.290.798828.010.752528.450.846932.570.9306
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.481, + 0.785, + 0.527 + ], + "angle": 0, + "content": "improve PSNR and SSIM when reducing the threshold \\(\\theta^{\\prime}\\) from 30 to 10 with an exception of SSIM on Urban100. This is because this method filters out images without considering the diversity of object regions." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.528, + 0.785, + 0.618 + ], + "angle": 0, + "content": "Object-based filtering. Table 5 compares object-based filtering methods described in Sec. 3.2, where ImageNet-1k is used as a source dataset. For a fair comparison, all dataset size after filtering are the same (260k images). As shown, the segmentation-based filtering performed the best. This indicates that images with diverse object regions are effective for SR training, and finer granularity leads to better performance." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.619, + 0.785, + 0.663 + ], + "angle": 0, + "content": "Filtering threshold. Table 6 shows the results obtained by varying the threshold \\(\\theta\\), which indicates the minimum number of segments. As shown, \\(\\theta = 100\\) performed the best." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.676, + 0.551, + 0.692 + ], + "angle": 0, + "content": "4.4 Analysis 2: Effects of image quality" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.704, + 0.785, + 0.748 + ], + "angle": 0, + "content": "In Sec. 3, we made an assumption that low-quality images are detrimental when training SR models. Here, we empirically justify this assumption and evaluate the impact of image quality on SR performance." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.785, + 0.809 + ], + "angle": 0, + "content": "Training with compressed images. This experiment applies JPEG compression to the two high-resolution datasets, DF2K and LSDIR, and trains a HAT model on each compressed dataset. As shown in Table 7, the performance decreases as the quality decreases on both datasets." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.785, + 0.84 + ], + "angle": 0, + "content": "Learning process. To further analyze why and how low quality images negatively affect training of SR models, Figure 5 compares learning processes. As can" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.381, + 0.128 + ], + "angle": 0, + "content": "G. Ohtani et al." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.145, + 0.498, + 0.296 + ], + "angle": 0, + "content": null + }, + { + "type": "image_footnote", + "bbox": [ + 0.371, + 0.298, + 0.648, + 0.309 + ], + "angle": 0, + "content": "Quality: HR 95% 85% 75% 50%" + }, + { + "type": "image", + "bbox": [ + 0.498, + 0.145, + 0.777, + 0.296 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.217, + 0.312, + 0.784, + 0.327 + ], + "angle": 0, + "content": "Fig. 5: Comparison of learning processes obtained with various JPEG quality values." + }, + { + "type": "image", + "bbox": [ + 0.222, + 0.34, + 0.408, + 0.444 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.408, + 0.341, + 0.593, + 0.444 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.593, + 0.341, + 0.777, + 0.444 + ], + "angle": 0, + "content": null + }, + { + "type": "image_footnote", + "bbox": [ + 0.399, + 0.445, + 0.61, + 0.455 + ], + "angle": 0, + "content": "ImageNet PASS Places365" + }, + { + "type": "image_caption", + "bbox": [ + 0.23, + 0.456, + 0.771, + 0.47 + ], + "angle": 0, + "content": "Fig. 6: Comparison of learning processes for ImageNet-1k, PASS, and Places365." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.5, + 0.788, + 0.591 + ], + "angle": 0, + "content": "be seen, PSNR decreases after 100k and 300k iterations with \\(50\\%\\) and \\(75\\%\\) qualities, respectively. With \\(50\\%\\) quality, the final performance at the 800k iteration was worse than that at the 10k iteration. These results indicate that excluding low-quality images is crucial for enhancing SR performance. This is in contrast to other vision tasks such as image recognition, where increasing the number of training images, even with low quality images, often helps improve performance." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.591, + 0.788, + 0.712 + ], + "angle": 0, + "content": "Training with Places365. In the source selection step, the Places365 dataset was excluded because its quality estimated via the blockiness distribution was low. To justify this selection, we compare three datasets in Figure 6. As can be seen, the model trained on Places365 performed worse than those trained on ImageNet-1k and PASS. Similar to the learning processes obtained from low quality JPEG images, the PSNR decreased in the later phase of the learning process with Places365. These results confirmed the effectiveness of source selection before training based on the blockiness distributions." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.723, + 0.512, + 0.74 + ], + "angle": 0, + "content": "4.5 Analysis 3: Visual comparison" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.843 + ], + "angle": 0, + "content": "Comparison of artifacts. Figure 7 compares the artifacts produced by JPEG compression with those produced by the SR models trained on DF2K using different image quality levels. Images obtained from models trained on \\(50\\%\\) and \\(75\\%\\) quality images show strong stripes or checkerboard patterns of artifacts. These artifacts appear to be more significant than those observed when the original image is compressed using JPEG, suggesting that they are induced by" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.273, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Rethinking Image Super-Resolution from Training Data Perspectives" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "image", + "bbox": [ + 0.231, + 0.148, + 0.391, + 0.34 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.237, + 0.34, + 0.383, + 0.351 + ], + "angle": 0, + "content": "AkkeraKanjinchou, Manga109" + }, + { + "type": "image", + "bbox": [ + 0.235, + 0.355, + 0.385, + 0.41 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.403, + 0.148, + 0.778, + 0.409 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.424, + 0.788, + 0.48 + ], + "angle": 0, + "content": "Fig. 7: Comparison of artifacts produced by JPEG compression (blue) and SR models trained on various JPEG quality values (red, HAT model, DF2K). Numbers indicate JPEG image quality. First and third rows: cropped images. Second and fourth rows: differential images." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.508, + 0.785, + 0.584 + ], + "angle": 0, + "content": "model training. The predisposition to stripes and checkerboard patterns is likely due to the inductive bias inherent in the architecture of the neural network; in particular, the square or rectangular shape of the filters in the convolutional operations could cause this. Improvements to the network architecture to allow training on lower quality images would be interesting as future work." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.584, + 0.788, + 0.689 + ], + "angle": 0, + "content": "Qualitative examples. Figure 8 shows visual comparisons of SR models trained on DF2K, LSDIR, DiverSeg-I, and DiverSeg-P. For the image \"img_011\" in Urban100, we observed that models trained on DF2K are unable to recover the horizontal stripes pattern, while the other three models successfully recovered it. With the three images of Manga109, we observed that models trained on DiverSeg-I/P exhibit noticeable improvement in the character region compared to those trained on DF2K or LSDIR." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.697, + 0.485, + 0.711 + ], + "angle": 0, + "content": "4.6 Discussion and Limitations" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.788, + 0.84 + ], + "angle": 0, + "content": "High-resolution datasets. In this paper, we applied the automated image evaluation pipeline to a large dataset of low-resolution images. We believe our finding contributes to the future construction of training datasets and the development of neural network architectures. Specifically, we demonstrated that SR models can be trained without high-resolution images. However, this does not imply that high-resolution image datasets are worthless. Rather, we believe that increasing the diversity of high-resolution images through our proposed filtering method could further improve SR performance. In Table 8, we examined this on" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.38, + 0.128 + ], + "angle": 0, + "content": "G. Ohtani et al." + }, + { + "type": "image", + "bbox": [ + 0.219, + 0.142, + 0.788, + 0.366 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.375, + 0.789, + 0.418 + ], + "angle": 0, + "content": "Fig. 8: Visual comparison of \\(\\times 4\\) SR models trained on DF2K, LSDIR, DiverSeg-I, and DiverSeg-P datasets. PSNR/SSIM is calculated for each cropped patch individually to better reflect the differences in performance." + }, + { + "type": "table_caption", + "bbox": [ + 0.235, + 0.432, + 0.767, + 0.447 + ], + "angle": 0, + "content": "Table 8: Applying object-based filtering to LSDIR with \\( \\theta = {100} \\) (HAT model)." + }, + { + "type": "table", + "bbox": [ + 0.235, + 0.452, + 0.77, + 0.511 + ], + "angle": 0, + "content": "
Dataset#ImagesSet5Set14BSD100Urban100Manga109
PSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIM
LSDIR89,99132.930.905329.290.798828.010.752528.450.846932.570.9306
w/ filtering31,56132.950.905629.330.799128.020.752628.530.848732.600.9308
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.54, + 0.788, + 0.601 + ], + "angle": 0, + "content": "the LSDIR dataset. Our results show that filtering out images with less than 100 object regions led to increased PSNR and SSIM with the HAT model. For future work, exploring a hybrid approach that utilizes both low- and high-resolution diverse images could be promising." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.601, + 0.788, + 0.708 + ], + "angle": 0, + "content": "Limitations. In this study, we focused on two perspectives: resolution and diversity. However, when collecting large datasets, there are other important perspectives, such as fairness and copyright. In addition, for benchmarking purposes, we used the five most commonly used SR datasets for a fair comparison with conventional methods. Nevertheless, real-world applications also face the challenge of blind super-resolution, where the degradation process is unknown. Creating datasets addressing this aspect would be an interesting future research direction." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.723, + 0.36, + 0.739 + ], + "angle": 0, + "content": "5 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.749, + 0.788, + 0.84 + ], + "angle": 0, + "content": "In this work, we investigated the effect of the training data for SR and showed that SR models are trainable even without using high-resolution images by applying the image evaluation pipeline to a set of large low-resolution images. In experiments, we thoroughly analyzed the effect of image quality and diversity to SR performance. We hope that this work will positively influence the future construction of training datasets and lead to better models." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.273, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Rethinking Image Super-Resolution from Training Data Perspectives" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.145, + 0.403, + 0.163 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.176, + 0.785, + 0.221 + ], + "angle": 0, + "content": "Computational resource of AI Bridging Cloud Infrastructure (ABCI) provided by National Institute of Advanced Industrial Science and Technology (AIST) was used." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.244, + 0.323, + 0.26 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.274, + 0.785, + 0.302 + ], + "angle": 0, + "content": "1. Agustsson, E., Timofte, R.: Ntire 2017 challenge on single image super-resolution: Dataset and study. In: CVPRW (2017) 2, 4, 8, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.304, + 0.785, + 0.344 + ], + "angle": 0, + "content": "2. Asano, Y.M., Rupprecht, C., Zisserman, A., Vedaldi, A.: Pass: An imagenet replacement for self-supervised pretraining without humans. In: NeurIPS Track on Datasets and Benchmarks (2021) 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.346, + 0.785, + 0.385 + ], + "angle": 0, + "content": "3. Bevilacqua, M., Roumy, A., Guillemot, C., Alberi-Morel, M.L.: Low-complexity single-image super-resolution based on nonnegative neighbor embedding. In: BMVC (2012) 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.387, + 0.785, + 0.413 + ], + "angle": 0, + "content": "4. Bhardwaj, D., Pankajakshan, V.: A jpeg blocking artifact detector for image forensics. Signal Processing: Image Communication 68, 155-161 (2018) 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.415, + 0.785, + 0.441 + ], + "angle": 0, + "content": "5. Chen, H., Wang, Y., Guo, T., Xu, C., Deng, Y., Liu, Z., Ma, S., Xu, C., Xu, C., Gao, W.: Pre-trained image processing transformer. In: CVPR (2021) 1, 4, 5, 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.443, + 0.785, + 0.468 + ], + "angle": 0, + "content": "6. Chen, X., Wang, X., Zhou, J., Qiao, Y., Dong, C.: Activating more pixels in image super-resolution transformer. In: CVPR (2023) 1, 4, 5, 9, 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.47, + 0.785, + 0.496 + ], + "angle": 0, + "content": "7. Dai, T., Cai, J., Zhang, Y., Xia, S.T., Zhang, L.: Second-order attention network for single image super-resolution. In: CVPR (2019) 4, 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.498, + 0.785, + 0.523 + ], + "angle": 0, + "content": "8. Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: Imagenet: A large-scale hierarchical image database. In: CVPR (2009) 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.525, + 0.785, + 0.551 + ], + "angle": 0, + "content": "9. Dong, C., Loy, C.C., He, K., Tang, X.: Learning a deep convolutional network for image super-resolution. In: ECCV (2014) 1, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.553, + 0.785, + 0.578 + ], + "angle": 0, + "content": "0. Dong, C., Loy, C.C., Tang, X.: Accelerating the super-resolution convolutional neural network. In: ECCV (2016) 1, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.58, + 0.785, + 0.606 + ], + "angle": 0, + "content": "1. Huang, J.B., Singh, A., Ahuja, N.: Single image super-resolution from transformed self-exemplars. In: CVPR (2015) 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.608, + 0.785, + 0.634 + ], + "angle": 0, + "content": "2. Kim, J., Lee, J.K., Lee, K.M.: Accurate image super-resolution using very deep convolutional networks. In: CVPR (2016) 1, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.636, + 0.785, + 0.674 + ], + "angle": 0, + "content": "3. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., Dollar, P., Girshick, R.: Segment anything. In: ICCV (2023) 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.676, + 0.785, + 0.716 + ], + "angle": 0, + "content": "4. Ledig, C., Theis, L., Huszár, F., Caballero, J., Cunningham, A., Acosta, A., Aitken, A., Tejani, A., Totz, J., Wang, Z., et al.: Photo-realistic single image superresolution using a generative adversarial network. In: CVPR (2017) 1, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.718, + 0.785, + 0.757 + ], + "angle": 0, + "content": "5. Li, A., Zhang, L., Liu, Y., Zhu, C.: Feature modulation transformer: Crossrefinement of global representation via high-frequency prior for image superresolution. In: CVPR (2023) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.759, + 0.785, + 0.798 + ], + "angle": 0, + "content": "6. Li, W., Lu, X., Qian, S., Lu, J., Zhang, X., Jia, J.: On efficient transformer-based image pre-training for low-level vision. arXiv preprint arXiv:2112.10175 (2021) 1, 4, 5, 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.8, + 0.785, + 0.84 + ], + "angle": 0, + "content": "7. Li, Y., Zhang, K., Liang, J., Cao, J., Liu, C., Gong, R., Zhang, Y., Tang, H., Liu, Y., Demandolx, D., Ranjan, R., Timofte, R., Van Gool, L.: Lsdir: A large scale dataset for image restoration. In: CVPRW (2023) 2, 5, 8, 9" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.274, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.381, + 0.128 + ], + "angle": 0, + "content": "G. Ohtani et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.787, + 0.189 + ], + "angle": 0, + "content": "18. Li, Y., Zhang, Y., Timofte, R., Van Gool, L., Yu, L., Li, Y., Li, X., Jiang, T., Wu, Q., Han, M., et al.: Ntire 2023 challenge on efficient super-resolution: Methods and results. In: CVPRW (2023) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.19, + 0.786, + 0.217 + ], + "angle": 0, + "content": "19. Liang, J., Cao, J., Sun, G., Zhang, K., Van Gool, L., Timofte, R.: Swinir: Image restoration using swin transformer. In: ICCVW (2021) 1, 2, 4, 9, 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.218, + 0.787, + 0.244 + ], + "angle": 0, + "content": "20. Lim, B., Son, S., Kim, H., Nah, S., Mu Lee, K.: Enhanced deep residual networks for single image super-resolution. In: CVPRW (2017) 1, 4, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.245, + 0.786, + 0.286 + ], + "angle": 0, + "content": "21. Lin, Z., Garg, P., Banerjee, A., Magid, S.A., Sun, D., Zhang, Y., Van Gool, L., Wei, D., Pfister, H.: Revisiting rcan: Improved training for image super-resolution. arXiv preprint arXiv:2201.11279 (2022) 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.287, + 0.787, + 0.327 + ], + "angle": 0, + "content": "22. Martin, D., Fowlkes, C., Tal, D., Malik, J.: A database of human segmented natural images and its application to evaluating segmentation algorithms and measuring ecological statistics. In: ICCV (2001) 4, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.328, + 0.787, + 0.37 + ], + "angle": 0, + "content": "23. Matsui, Y., Ito, K., Aramaki, Y., Fujimoto, A., Ogawa, T., Yamasaki, T., Aizawa, K.: Sketch-based manga retrieval using manga109 dataset. Multimedia Tools and Applications 76, 21811-21838 (2017) 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.371, + 0.786, + 0.396 + ], + "angle": 0, + "content": "24. Mei, Y., Fan, Y., Zhou, Y.: Image super-resolution with non-local sparse attention. In: CVPR (2021) 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.398, + 0.787, + 0.438 + ], + "angle": 0, + "content": "25. Niu, B., Wen, W., Ren, W., Zhang, X., Yang, L., Wang, S., Zhang, K., Cao, X., Shen, H.: Single image super-resolution via a holistic attention network. In: ECCV (2020) 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.439, + 0.787, + 0.481 + ], + "angle": 0, + "content": "26. Shi, W., Caballero, J., Huszar, F., Totz, J., Aitken, A.P., Bishop, R., Rueckert, D., Wang, Z.: Real-time single image and video super-resolution using an efficient sub-pixel convolutional neural network. In: CVPR (2016) 1, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.481, + 0.787, + 0.521 + ], + "angle": 0, + "content": "27. Timofte, R., Agustsson, E., Van Gool, L., Yang, M.H., Zhang, L.: Ntire 2017 challenge on single image super-resolution: Methods and results. In: CVPRW (2017) 2, 4, 8, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.522, + 0.787, + 0.549 + ], + "angle": 0, + "content": "28. Tong, T., Li, G., Liu, X., Gao, Q.: Image super-resolution using dense skip connections. In: ICCV (2017) 1, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.549, + 0.787, + 0.591 + ], + "angle": 0, + "content": "29. Wang, X., Yu, K., Wu, S., Gu, J., Liu, Y., Dong, C., Qiao, Y., Change Loy, C.: Esrgan: Enhanced super-resolution generative adversarial networks. In: ECCVW (2018) 1, 2, 4, 9, 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.591, + 0.787, + 0.619 + ], + "angle": 0, + "content": "30. Yang, J., Wright, J., Huang, T., Ma, Y.: Image super-resolution as sparse representation of raw image patches. In: CVPR (2008) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.62, + 0.787, + 0.66 + ], + "angle": 0, + "content": "31. Yang, Q., Chen, D., Tan, Z., Liu, Q., Chu, Q., Bao, J., Yuan, L., Hua, G., Yu, N.: Hq-50k: A large-scale, high-quality dataset for image restoration. arXiv preprint arXiv:2306.05390 (2023) 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.66, + 0.787, + 0.688 + ], + "angle": 0, + "content": "32. Zeyde, R., Elad, M., Protter, M.: On single image scale-up using sparse-representations. In: Proc. 7th Int. Conf. Curves Surf. (2010) 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.689, + 0.787, + 0.73 + ], + "angle": 0, + "content": "33. Zhang, D., Huang, F., Liu, S., Wang, X., Jin, Z.: Swinfir: Revisiting the swinir with fast fourier convolution and improved training for image super-resolution. arXiv preprint arXiv:2208.11247 (2022) 1, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.73, + 0.787, + 0.757 + ], + "angle": 0, + "content": "34. Zhang, X., Zeng, H., Guo, S., Zhang, L.: Efficient long-range attention network for image super-resolution. In: ECCV (2022) 1, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.758, + 0.787, + 0.785 + ], + "angle": 0, + "content": "35. Zhang, Y., Li, K., Zhong, B., Fu, Y.: Residual non-local attention networks for image restoration. In: ICLR (2019) 1, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.785, + 0.787, + 0.813 + ], + "angle": 0, + "content": "36. Zhang, Y., Li, K., Li, K., Wang, L., Zhong, B., Fu, Y.: Image super-resolution using very deep residual channel attention networks. In: ECCV (2018) 1, 4, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.814, + 0.787, + 0.84 + ], + "angle": 0, + "content": "37. Zhang, Y., Tian, Y., Kong, Y., Zhong, B., Fu, Y.: Residual dense network for image super-resolution. In: CVPR (2018) 1, 4" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.273, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Rethinking Image Super-Resolution from Training Data Perspectives" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.189 + ], + "angle": 0, + "content": "38. Zhang, Y., Zhang, K., Chen, Z., Li, Y., Timofte, R., Zhang, J., Zhang, K., Peng, R., Ma, Y., Jia, L., et al.: Ntire 2023 challenge on image super-resolution (x4): Methods and results. In: CVPRW (2023) 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.19, + 0.788, + 0.231 + ], + "angle": 0, + "content": "39. Zhou, B., Lapedriza, A., Xiao, J., Torralba, A., Oliva, A.: Learning deep features for scene recognition using places database. Advances in neural information processing systems 27 (2014) 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.232, + 0.788, + 0.272 + ], + "angle": 0, + "content": "40. Zhou, S., Zhang, J., Zuo, W., Loy, C.C.: Cross-scale internal graph neural network for image super-resolution. Advances in neural information processing systems 33, 3499-3509 (2020) 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.273, + 0.788, + 0.3 + ], + "angle": 0, + "content": "41. Zhou, X., Girdhar, R., Joulin, A., Krahenbuhl, P., Misra, I.: Detecting twenty-thousand classes using image-level supervision. In: ECCV (2022) 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.3, + 0.788, + 0.328 + ], + "angle": 0, + "content": "42. Zhou, Y., Li, Z., Guo, C.L., Bai, S., Cheng, M.M., Hou, Q.: Srformer: Permuted self-attention for single image super-resolution. In: ICCV (2023) 1, 2, 4" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.328 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/Rethinking Image Super Resolution from Training Data Perspectives/8d3fa075-356f-4f4c-b1a5-4318bed20e1d_origin.pdf b/2024/Rethinking Image Super Resolution from Training Data Perspectives/8d3fa075-356f-4f4c-b1a5-4318bed20e1d_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..f3655a5dcf84a81a74ffa4fa322e9ba287e7e316 --- /dev/null +++ b/2024/Rethinking Image Super Resolution from Training Data Perspectives/8d3fa075-356f-4f4c-b1a5-4318bed20e1d_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:13a1930cd0629b978ff5a749e1a7218b5f4abf3e837cba073310f919475b8452 +size 23039261 diff --git a/2024/Rethinking Image Super Resolution from Training Data Perspectives/full.md b/2024/Rethinking Image Super Resolution from Training Data Perspectives/full.md new file mode 100644 index 0000000000000000000000000000000000000000..e962c80268a89170a3f67c6313388c62da3fba6e --- /dev/null +++ b/2024/Rethinking Image Super Resolution from Training Data Perspectives/full.md @@ -0,0 +1,296 @@ +# Rethinking Image Super-Resolution from Training Data Perspectives + +Go Ohtani $^{1,2}$ , Ryu Tadokoro $^{2}$ , Ryosuke Yamada $^{2,3}$ , Yuki M. Asano $^{4}$ , Iro Laina $^{5}$ , Christian Rupprecht $^{5}$ , Nakamasasa Inoue $^{6,2}$ , Rio Yokota $^{6,2}$ , Hirokatsu Kataoka $^{2}$ , and Yoshimitsu Aoki $^{1}$ + +1 Keio University + +$^{2}$ National Institute of Advanced Industrial Science and Technology (AIST) + +3 University of Tsukuba + +4 University of Amsterdam, now at University of Technology Nuremberg + +5 University of Oxford + +$^{6}$ Tokyo Institute of Technology + +Abstract. In this work, we investigate the understudied effect of the training data used for image super-resolution (SR). Most commonly, novel SR methods are developed and benchmarked on common training datasets such as DIV2K and DF2K. However, we investigate and rethink the training data from the perspectives of diversity and quality, thereby addressing the question of "How important is SR training for SR models?" To this end, we propose an automated image evaluation pipeline. With this, we stratify existing high-resolution image datasets and larger-scale image datasets such as ImageNet and PASS to compare their performances. We find that datasets with (i) low compression artifacts, (ii) high within-image diversity as judged by the number of different objects, and (iii) a large number of images from ImageNet or PASS all positively affect SR performance. We hope that the proposed simple-yet-effective dataset curation pipeline will inform the construction of SR datasets in the future and yield overall better models. Code is available at: https://github.com/gohtanii/DiverSeg-dataset + +Keywords: Super-resolution dataset $\cdot$ Image compression $\cdot$ Image diversity + +# 1 Introduction + +Image super-resolution (SR) aims to reconstruct high-resolution images from low-resolution images. It has been considered as one of the most fundamental tasks in the field of computer vision, with applications ranging from autonomous driving to medical imaging. Deep learning methods have led to significant advances in SR over the last decade, focusing primarily on improvements in the neural network architectures. Early SR models rely on convolutional neural networks (CNNs) [9,10,12,14,20,26,28,29,35-37]. Recent innovations have given rise to transformer-based SR models [5,6,16,19,33,34,42], which have consistently improved the performances. + +![](images/dfdd78fcf01ffef938e6622e91f2f74a50994f16ad0ad0b8a3eafbfdb911127a.jpg) +Fig. 1: We propose an automated image evaluation pipeline to curate a dataset for training SR models. The obtained dataset, namely DiverSeg, consists of low-resolution but high-quality images with many object regions. SR models trained on DiverSeg outperform those trained on high-resolution image datasets such as DF2K and LSDIR. + +With the improvement of neural network architectures, the importance of training datasets has also increased, as discussed in [17]. Examples of high-resolution datasets include DIV2K [1] and Flickr2K [27]. The combined dataset of these two, referred to as DF2K, is often utilized for training SR models. Most recently, LSDIR [17] has been proposed, which consists of 84,991 high-resolution images. It has been confirmed that training on large high-resolution datasets contributes significantly to performance improvement [18, 19, 29, 38, 42] + +The conventional approach to constructing datasets relies on a manual evaluation step, where the following two perspectives are most commonly considered: + +1. Resolution and quality [17, 31]. This perspective focuses on the pixel density of images. Images that do not meet the specified resolution threshold are excluded. Typically, HD, 2K, and 4K images are used to create a dataset. After the initial automatic filtering based on image size, the details of each image are manually evaluated to identify and exclude compressed images. +2. Variety and diversity [15]. This includes diversity in subjects (e.g., people, landscapes, urban scenes), lighting conditions, colors, textures, and other photographic elements. A diverse dataset is said to help train a model that is robust and performs well across a wide range of domains. + +Datasets constructed from these perspectives have been shown to significantly improve the performance of SR models. However, they also pose challenges in + +scaling the datasets, as collecting uncompressed high-resolution images is difficult and costly. + +To address this limitation, this paper rethinks these perspectives and proposes Diverse Segmentation dataset (DiverSeg), a low-resolution1 yet effective image dataset for training SR models. As shown in Figure 1, the dataset is constructed by applying filtering to a large set of low-resolution images, such as ImageNet-1k [8] and PASS [2]. In experiments, we demonstrate that models trained on DiverSeg outperform those trained on high-resolution image datasets such as DF2K and LSDIR. Based on this finding, our contributions are summarized as follows. + +1) Rethinking the resolution perspective. High-resolution images have been considered to be necessary for training SR models. In this work, we challenge this traditional perspective and show that SR models can be trained without high-resolution images. Specifically, we introduce a method to estimate image quality based on the kernel density estimation over blockiness values [4] that estimates the quantity distribution of blocking artifacts. We demonstrate that low-resolution images with high quality, indicated by reduced artifacts, can improve the performance of SR models. We also thoroughly analyze the impact of image quality on SR performance. +2) Rethinking the diversity perspective. When constructing datasets for training SR models, images containing only a small number of objects are often implicitly excluded during the manual resolution evaluation process. This is because evaluators typically focus on the details of objects or small objects in the images. Therefore, we explicitly calculate the number of objects in images and analyze how this number affects SR performance. In our experiments, we show that constructing datasets with images containing many objects improves the performance of SR models. +3) Dataset construction. To facilitate analysis from the above two perspectives, we introduce a framework that automatically creates a dataset from a set of low-resolution images collected from the web. Specifically, the framework consists of two steps, source selection and object-based filtering, which correspond to the first and second perspectives, respectively. Our framework eliminates the need for manual assessment, facilitating dataset scaling. In the first step, given several low-resolution image datasets, we estimate the image quality of them through their blockiness distributions and select datasets with quality larger than $90\%$ . In the second step, we filter out images with a small number of object regions by using object detection and image segmentation models. The filtered dataset, which we refer to as the DiverSeg dataset, is utilized for training various SR models in our experiments. We apply our framework to a union set of ImageNet, Places365, and PASS to construct the DiverSeg dataset. We demonstrate that SR models trained on DiverSeg archive state-of-the-art performance. + +# 2 Related Work + +# 2.1 Image super-resolution models + +A number of SR models have been proposed that take advantage of deep learning techniques. These models can be divided into two groups, CNN-based models [9, 10,12,14,20,26,28,29,35-37] and Transformer-based models [5,6,16,19,33,34,42]. Each group exploits different architectural strengths to enhance low-resolution images to high resolution. + +CNN-based models. SRCNN [9] was the first model that integrates a deep convolutional architecture for SR. Subsequently, FSRCNN [10] significantly improved computational efficiency by performing convolutional processing in low resolution space and upsampling in the last layer. Furthermore, ESPCN [26] adopted an efficient upsampling method, sub-pixel convolution, to enhance performance while reducing computational costs. These methods established the basic structure of modern SR networks and laid the foundation for subsequent research developments. Later studies introduced various modules such as residual connections [12, 14, 20, 29], dense blocks [28, 29, 37], and attention mechanisms [7, 35, 36]. EDSR [20] eliminated batch normalization layers and introduced residual scaling to enable stable training of large models. MSRResNet [29] replaced the basic ResNet blocks with residual dense blocks to improve the balance between performance and computational efficiency. RCAN [36] incorporated channel attention mechanisms to adaptively weight feature representations, significantly enhancing SR performance. + +Transformer-based models. As the first Transformer-based image restoration model, IPT [5] was introduced as a large-scale model utilizing the Transformer's encoder and decoder architecture. By pre-training on ImageNet, IPT significantly improved SR performance, fully leveraging the capabilities of the Transformer. SwinIR [19] executes self-attention within local windows during feature extraction, demonstrating exceptional SR performance and establishing itself as the foundational model for Transformers in SR. Following this, models that build upon SwinIR have been developed [6,33,34,42], enhancing performance by extending and optimizing the self-attention mechanism. Among them, HAT [6] achieves state-of-the-art performance in SR by using a duplicated cross-attention module and pre-training on ImageNet. However, there may be potential for improvement, as ImageNet is primarily an image recognition dataset and may not be fully optimized for SR. + +# 2.2 Image super-resolution datasets + +T91 [30] and BSDS200 [22] are early datasets used to train SR models, consisting of 91 and 200 images, respectively. The turning point in training datasets was ushered by the release of DIV2K [1], a compilation of 800 high-resolution images with minimal compression noise meticulously collected from the web. Subsequently, to accommodate the further scaling of the model, Flickr2K [27], consisting of 2,650 high-resolution images, is merged into DIV2K and is referred + +![](images/dea9793fed3a8a8306fa8e627c1f807638b424faceb95b2451095ca6ff31d578.jpg) +Fig. 2: Images of DiverSeg with their segmentation masks. DiverSeg is obtained from a large set of low-resolution images through the automated image evaluation pipeline. + +to as DF2K. In recent years, to further expand the scale of SR datasets, datasets larger than DF2K have been released, such as LSDIR [17], comprising 84,991 images and HQ-50K [31] consisting of 50,000 images. These datasets persistently adhere to stringent criteria, ensuring high resolution and negligible compression noise in the imagery. However, since collecting images that meet the aforementioned conditions is quite challenging, these datasets still consist of only tens of thousands of images. In contrast, several approaches [5, 6, 16] adopt ImageNet, also known as ImageNet-1k, which consists of 1.28M images spanning diverse categories. These approaches leverage the diversity of texture patterns in ImageNet as an advantage. Nevertheless, ImageNet is pointed out to contain some images having low-resolution and JPEG-compression artifacts, which adversely affect the training results. + +# 3 DiverSeg dataset + +This section presents Diverse Segmentation dataset (DiverSeg), our dataset for training SR models without using high-resolution images. As shown in Figure 2, the DiverSeg dataset consists of low-resolution but high-quality images with diverse segmented object regions. The framework for constructing the dataset consists of two steps, source selection and object-based filtering, which are designed from our perspective of rethinking resolution and diversity, respectively. Our approach eliminates the manual cost of collecting and quality-checking high-resolution images. + +# 3.1 Source selection + +Let $\mathcal{X}$ be a set of low-resolution image datasets. This step filters out low-quality datasets by estimating the quality $\hat{q}_X \in [0,1]$ for each dataset $X \in \mathcal{X}$ and excluding those with $\hat{q}_X < 0.9$ , under the assumption that low-quality images are detrimental when training SR models. We introduce a quality estimation method based on the Kullback-Leibler (KL) divergence between blockiness distributions as detailed below. + +![](images/51f21f57bc9f13566804ac593d29b853c03b7caaf49014cc9d08b4949e5874b4.jpg) +Fig. 3: Comparison of image degradation due to JPEG quality(blue). Blockiness values calculated from the images are marked. As the JPEG quality decreases and artifacts increase, we observe a corresponding rise in blockiness values. + +![](images/7681ccaf48f120ec4f05e6755a7a09218b84820466595b61affb1bb84866e47b.jpg) +Fig. 4: (a) Blockiness distributions $p_{X,1.0}$ for $X = \mathrm{ImageNet - 1k}$ , Places365 and PASS. (b) Basis distributions $p_{Z,q}$ for $Z = \mathrm{DF2K}$ and $q = 0.5, 0.75, 0.85, 0.95, 1.0$ . We estimate the quality by comparing $p_{X,1.0}$ and $p_{Z,q}$ using the KL divergence. + +![](images/731aae67800aee7ca10548feb9822e51a430967e4189ce57b9a7969c90690e8d.jpg) + +Image datasets. This work uses three web-collected low-resolution datasets $\mathcal{X} = \{\mathrm{ImageNet - 1k},\mathrm{Places365},\mathrm{PASS}\}$ . Training SR models on them is not straightforward because they may include highly compressed images that negatively affect training of SR models. + +Quality definition. Let $Y$ be a dataset of JPEG images. We define the quality $q_{Y}$ by the average JPEG quality, i.e., $q_{Y} \triangleq \frac{1}{|Y|} \sum_{y \in Y} Q(y)$ where $0 \leq Q(y) \leq 1$ is the JPEG quality of an image $y$ . The goal of quality estimation is to estimate $q_{X}$ given a dataset $X$ . Note that in the estimation phase, datasets may include images other than JPEG and the true quality $q_{X}$ is not observable. + +Blockiness distribution. To estimate the quality, we utilize the blockiness measure [4]. Specifically, for each image dataset $X \in \mathcal{X}$ , we estimate the distribution of blockiness values $p_{X,q}(b)$ by kernel density estimation as follows: + +$$ +p _ {X, q} (b) = \frac {1}{h | X |} \sum_ {x \in X} K \left(\frac {b - B \left(\phi_ {q} (x)\right)}{h}\right), \tag {1} +$$ + +where $x$ is an image, $\phi_q$ is the JPEG compression function, $q \in [0,1]$ is a quality value, $K: \mathbb{R} \to \mathbb{R}$ is a Gaussian kernel, and $h \in \mathbb{R}$ is the bandwidth determined by Scott's method. The function $B$ is the blockiness measure that measures the + +quantity of blocking artifacts by computing subband DCT coefficients. Specifically, $B$ is defined on images that are decomposed into $P \times P$ patches as follows: + +$$ +B (x) = \sum_ {i = 1} ^ {P} \sum_ {j = 1} ^ {P} \left| \frac {\bar {V} _ {\text {c r o p}} (i , j) - \bar {V} (i , j)}{\bar {V} (i , j)} \right|, \quad \bar {V} (i, j) = \frac {1}{W H} \sum_ {h = 1} ^ {H} \sum_ {w = 1} ^ {W} V _ {w, h} (i, j) \tag {2} +$$ + +where $W, H$ are the width and height of an image $x$ , $V_{w,h}(i,j)$ is the variation in the $(i,j)$ -th subband DCT coefficients within the $(h,w)$ -th patch being calculated and its four spatially adjacent patches. $V_{\mathrm{crop}}(i,j)$ is the variation calculated similarly to $V_{w,h}(i,j)$ for the given image with the first 4 rows and 4 columns removed. $\bar{V}$ and $\bar{V}_{\mathrm{crop}}$ are the average variations of $V_{w,h}(i,j)$ and $V_{\mathrm{crop}}(i,j)$ calculated for each patch, respectively. This work uses $P = 8$ . The blockiness value $B(x)$ is expected to be low for uncompressed images and high for compressed images as shown in Figure 3. + +Quality estimation. We estimate the quality by comparing $p_{X,1.0}$ with $\{p_{Z,q}\}_{q \in S}$ , where $p_{Z,q}$ is a basis distribution, a blockiness distribution of images of a fixed quality $q$ . More specifically, the estimated quality is given by + +$$ +\hat {q} _ {X} = \sum_ {q \in S} q \frac {\exp \left(- D _ {\mathrm {K L}} \left(p _ {X , 1 . 0} \mid \mid p _ {Z , q}\right)\right)}{\sum_ {q ^ {\prime} \in S} \exp \left(- D _ {\mathrm {K L}} \left(p _ {X , 1 . 0} \mid \mid p _ {Z , q ^ {\prime}}\right)\right)}, \tag {3} +$$ + +where $Z$ is a small dataset that involves only uncompressed images. In this work, we use DF2K. $D_{\mathrm{KL}}$ is the KL divergence, and $S = \{1.0, 0.95, 0.85, 0.75, 0.5\}$ are discretely sampled quality values. Figure 4 shows the blockiness distributions $p_{X,1.0}$ for the three low-resolution image datasets and the basis distributions $p_{Z,c}$ . + +Selection results. The estimated qualities for ImageNet-1k, Places365, and PASS were $95.5\%$ , $75.0\%$ and $99.8\%$ , respectively. From this result, Places365 is filtered out. + +# 3.2 Object-based filtering + +Given a source dataset $X$ , this step applies filtering to refine it as a dataset for training SR models, under the assumption that images with diverse object regions are more effective than those with uniform or monotonous content. Specifically, the refined training dataset is given by $\tilde{X} = \{x \in X : R(x) \geq \theta\}$ , where $R(x)$ is the number of object regions and $\theta$ is a threshold. + +We introduce two object-based filtering methods with different granularities to explore how the granularity of object detection, ranging from detailed identification of small objects or features to recognizing larger, more general objects, impacts SR performance. + +Segmentation-based filtering. This method counts the number of object parts by an image segmentation model. Specifically, we adopt the SAM [13] with the ViT-H backbone and define $R$ by the number of segmentation masks. + +Table 1: Dataset statistics. HR: High resolution, LR: Low resolution, #Images: Number of images, #Pixels: Average number of pixels per image, Blockiness: median of blockiness measure indicating the intensity of JPEG compression noise, #Segments: Average number of segmentation masks. + +
DatasetTaskHR LR#Images#PixelsBlockiness#Segments
DIV2K [1]Super-resolution✓8002.8M0.47104
DF2K [1,27]Super-resolution✓3,4502.8M0.47103
LSDIR [17]Super-resolution✓84,9911.1M0.8292
ImageNet [8]Image recognition1,281,167237k4.3971
Places365 [39]Image recognition1,803,460366k80.71100
PASS [2]Image recognition1,439,589178k3.0374
DiverSeg-I (Ours)Super-resolution259,448233k2.83146
DiverSeg-P (Ours)Super-resolution267,055179k4.39146
DiverSeg-IP (Ours)Super-resolution526,503206k3.61146
+ +We chose SAM because, unlike typical segmentation models that perform semantic segmentation based on class labels, SAM provides segments that do not impose semantic constraints, allowing for finer region segmentation of diverse objects. We set $\theta = 100$ as the default, resulting in $260\mathrm{k}$ images remaining after applying the filter to ImageNet-1k. + +Detection-based filtering. This method counts the number of objects per image. We adopt the Detic model [41] with the ViT-B backbone and define $R$ by the number of detected objects. We set $\theta = 18$ , resulting in 260k images remaining after applying the filter to ImageNet-1k. + +# 3.3 Dataset statistics + +Table 1 shows dataset statistics. We created three variants of DiverSeg, namely DiverSeg-I, DiverSeg-P, and DiverSeg-IP, constructed from ImageNet-1k, PASS, and the union of the two, respectively. The segmentation-based filtering is applied to obtain these datasets. Compared to high-resolution datasets such as DF2K and LSDIR, our datasets have larger number of training images, but contain only low-resolution images. The median of blockiness values is decreased for ImageNet-1k and is increased for PASS after filtering. + +# 4 Experiments + +This section conducts experiments by training various SR models on DiverSeg datasets. We demonstrate that SR models can be trained without using high-resolution images and thoroughly analyze factors from a dataset perspective that are crucial for enhancing the performance of SR. + +Table 2: Comparison of DiverSeg with high-resolution image datasets (DF2K and LSDIR). Models trained on DiverSeg-I and DiverSeg-P demonstrated superior performance despite not using any high-resolution images for training. + +
Model (Params)DatasetHR LRSet5Set14BSD100Urban100Manga109
PSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIM
MSRResNet [29] (1.5M)DF2K✓32.230.895528.670.783127.620.737426.230.789730.640.9108
LSDIR✓32.150.894828.660.783627.620.737426.310.791830.570.9105
DiverSeg-I✓32.270.896328.640.783727.640.737826.310.791830.530.9115
DiverSeg-P✓32.090.894328.610.783227.600.737126.280.791830.360.9101
RCAN [36] (15.5M)DF2K✓32.500.899028.870.788527.750.742126.730.805831.170.9165
LSDIR✓32.530.899228.890.789427.750.742526.910.809031.330.9180
DiverSeg-I✓32.700.901228.980.790827.810.744327.030.811631.580.9210
DiverSeg-P✓32.630.900028.950.789827.770.743526.990.813431.190.9190
EDSR [20] (43.0M)DF2K✓32.610.899828.910.789327.790.743426.840.808931.380.9176
LSDIR✓32.570.899228.970.790827.800.743827.050.813131.470.9192
DiverSeg-I✓32.710.901728.980.791327.850.745327.100.814231.720.9216
DiverSeg-P✓32.570.900229.060.791527.800.744727.100.816331.330.9191
SwinIR [19] (11.9M)DF2K✓32.920.904429.090.795027.920.748927.450.825432.030.9260
LSDIR✓32.860.903629.160.796327.920.749227.790.833131.980.9262
DiverSeg-I✓32.970.905329.230.797027.980.750827.830.833632.340.9283
DiverSeg-P✓32.850.904029.240.796127.960.750227.850.834932.280.9278
HAT [6] (20.7M)DF2K✓33.030.905629.160.796427.990.751427.930.836532.440.9292
LSDIR✓32.930.905329.290.798828.010.752528.450.846932.570.9306
DiverSeg-I✓33.150.907129.460.800428.070.754228.510.847732.900.9325
DiverSeg-P✓33.120.906829.500.800228.040.753628.530.849232.830.9320
+ +# 4.1 Experimental settings + +SR models. We use five models. Specifically, we use three CNN-based models: MSRResNet [29], EDSR [20], RCAN [36], and two Transformer-based models: SwinIR [19] and HAT [6]. + +Training datasets. We compare the DiverSeg datasets with two high-resolution datasets: DF2K and LSDIR [17]. DF2K is a merged dataset of DIV2K [1] and Flickr2K [27]. + +Evaluation datasets. We use five benchmark datasets: Set5 [3], Set14 [32], BSD100 [22], Urban100 [11], and Manga109 [23]. + +Evaluation metrics. PSNR and SSIM on the Y channel (representing luminance) within the transformed YCbCr color space are used as evaluation metrics. + +Implementation settings. We follow the training setup of the original papers of the SR models [6, 19, 20, 29, 36]. Implementation details are provided in the supplementary materials. + +# 4.2 Experimental results + +Main results. To validate the effectiveness of our approach, we trained the five SR models on DiverSeg-I/P. The results are summarized in Table 2. As shown, models trained on DiverSeg-I/P achieved better performance than those trained on DF2K and LSDIR. This shows the effectiveness of the proposed datasets. To the best of our knowledge, we are the first to successfully train SR models without using high-resolution images. + +Table 3: Quantitative comparison with state-of-the-art methods on five benchmark datasets. We applied our dataset to two Transformer-based models. Checkmarks for HR, LR indicate the use of high-resolution and low-resolution datasets, respectively. + +
MethodTraining DataHR LRSet5Set14BSD100Urban100Manga109
PSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIM
SAN [7]DIV2K✓32.640.900328.920.788827.780.743626.790.806831.180.9169
IGNN [40]DIV2K✓32.570.899828.850.789127.770.743426.840.809031.280.9182
HAN [25]DIV2K✓32.640.900228.900.789027.800.744226.850.809431.420.9177
NLSN [24]DIV2K✓32.590.900028.870.789127.780.744426.960.810931.270.9184
RRDB [29]DF2K✓32.730.901128.990.791727.850.745527.030.815331.660.9196
RCAN-it [21]DF2K✓32.690.900728.990.792227.870.745927.160.816831.780.9217
EDT [16]DF2K✓32.820.903129.090.793927.910.748327.460.824632.050.9254
HAT-S [6]DF2K✓32.920.904729.150.795827.970.750527.870.834632.350.9283
IPT [5]ImageNet✓32.64-29.01-27.82-27.26---
SwinIR [19]DF2K✓32.920.904429.090.795027.920.748927.450.825432.030.9260
SwinIR [19]DiverSeg-I (Ours)✓32.970.905329.230.797027.980.750827.830.833632.340.9283
HAT [6]DF2K✓33.040.905629.230.797328.000.751727.970.836832.480.9292
HAT [6]ImageNet→DF2K✓33.180.907329.380.800128.050.753428.370.844732.870.9319
HAT [6]DiverSeg-I (Ours)✓33.150.907129.460.800428.060.754228.510.847732.900.9325
HAT [6]DiverSeg-IP (Ours)✓33.140.907329.510.800728.070.754228.540.849232.960.9327
HAT-L [6]ImageNet→DF2K✓33.300.908329.470.801528.090.755128.600.849833.090.9335
HAT-L [6]DiverSeg-I (Ours)✓33.280.908329.540.802228.100.755628.750.852933.140.9340
HAT-L [6]DiverSeg-IP (Ours)✓33.200.908029.590.801928.110.755628.810.854733.190.9342
+ +Table 4: Filtering by blockiness (HAT model, ImageNet-1k). $\theta^{\prime}$ : threshold for blockiness values. + +
θ'#ImagesSet5Set14BSD100Urban100Manga109
PSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIM
10800k33.070.906829.340.800028.040.753828.410.846532.560.9310
30939k33.130.907229.390.800128.060.753828.430.846332.700.9316
1001.08M33.110.907129.420.800128.050.753728.410.845732.840.9321
-1.15M33.080.907129.400.799928.050.753528.410.845732.880.9323
+ +Comparison with SOTA. Table 3 shows that DiverSeg improves the state-of-the-art performance. Specifically, the HAT and HAT-L models trained on DiverSeg datasets outperformed those trained with ImageNet-1k $\rightarrow$ DF2K, which utilizes all ImageNet-1k images for pre-training and the DF2K images for fine-tuning, in terms of PSNR and SSIM on four of five benchmarking datasets. It is worth noting that DiverSeg-I filters out $77.5\%$ of images from ImageNet-1k and thus training on it is more efficient than the approach relying on pre-training and fine-tuning. This confirmed the effectiveness and efficiency of our filtering approach. + +# 4.3 Analysis 1: Effects of filtering + +Filtering by blockiness measure. If the blockiness is an important factor for enchaining SR performance, filtering images by the blockiness values, i.e., constructing a training dataset by $\hat{X} = \{x\in X:B(x)\leq \theta '\}$ , would be a straightforward approach. However, as shown in Table 4, this filtering did not + +Table 5: Comparison of filtering methods. Performance is evaluated using $260\mathrm{k}$ filtered images (HAT model, ImageNet-1k). + +
Filtering methodUrban100Manga109
PSNRSSIMPSNRSSIM
Blockiness28.390.846732.470.9304
Detection-based28.440.846232.870.9322
Seg.-based28.510.847732.900.9325
+ +Table 6: Performance comparison across different thresholds $\theta$ (HAT model, ImageNet-1k). + +
θ#ImagesUrban100Manga109
PSNRSSIMPSNRSSIM
01.2M28.410.845732.880.9323
50663k28.460.847232.900.9325
100259k28.510.847732.900.9325
15086k28.360.845232.830.9320
+ +Table 7: Analysis of effects of JPEG quality (HAT model). + +
DatasetQuality (%)BlockinessSet5Set14BSD100Urban100Manga109
PSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIM
DF2K50165.3431.130.885527.540.753326.320.710325.040.751130.200.9024
7583.2032.630.900827.700.745727.700.745727.390.823831.880.9238
8546.9832.940.904329.070.794127.930.749927.720.831332.140.9266
9510.3332.980.904829.110.795127.990.752127.810.833832.210.9278
HR0.4733.030.905629.160.796427.990.751427.930.836532.440.9292
LSDIR50146.4328.500.848726.660.742124.720.667424.910.765729.390.8956
7557.1431.870.885328.530.780527.490.741727.300.823131.830.9223
8524.1432.840.904429.170.795527.980.751328.220.841232.380.9291
953.7132.970.904329.250.797928.050.753928.370.845232.420.9299
HR0.8232.930.905329.290.798828.010.752528.450.846932.570.9306
+ +improve PSNR and SSIM when reducing the threshold $\theta^{\prime}$ from 30 to 10 with an exception of SSIM on Urban100. This is because this method filters out images without considering the diversity of object regions. + +Object-based filtering. Table 5 compares object-based filtering methods described in Sec. 3.2, where ImageNet-1k is used as a source dataset. For a fair comparison, all dataset size after filtering are the same (260k images). As shown, the segmentation-based filtering performed the best. This indicates that images with diverse object regions are effective for SR training, and finer granularity leads to better performance. + +Filtering threshold. Table 6 shows the results obtained by varying the threshold $\theta$ , which indicates the minimum number of segments. As shown, $\theta = 100$ performed the best. + +# 4.4 Analysis 2: Effects of image quality + +In Sec. 3, we made an assumption that low-quality images are detrimental when training SR models. Here, we empirically justify this assumption and evaluate the impact of image quality on SR performance. + +Training with compressed images. This experiment applies JPEG compression to the two high-resolution datasets, DF2K and LSDIR, and trains a HAT model on each compressed dataset. As shown in Table 7, the performance decreases as the quality decreases on both datasets. + +Learning process. To further analyze why and how low quality images negatively affect training of SR models, Figure 5 compares learning processes. As can + +Fig. 5: Comparison of learning processes obtained with various JPEG quality values. +![](images/4fe7e02fcbb502ad20d83e2666ecc80b5bb0b42d3ad2219e74bc6e9ad45f3854.jpg) +Quality: HR 95% 85% 75% 50% + +![](images/9774f3a1c9fc6228546a0672961223950c0359cc76f84c234b1f0249a12a00a8.jpg) + +Fig. 6: Comparison of learning processes for ImageNet-1k, PASS, and Places365. +![](images/40f63773650fba6630c4347c4d4cf97ee3ae59895185ef5144f749d12928e89c.jpg) +ImageNet PASS Places365 + +![](images/c19f8097c1bcbb0ce6cfc774d4a09d4d923b91af515023267840f606f2228b97.jpg) + +![](images/7c9a6b211ee6ca901015927da0faf4f90ded647abab5082bd635f85931bb2006.jpg) + +be seen, PSNR decreases after 100k and 300k iterations with $50\%$ and $75\%$ qualities, respectively. With $50\%$ quality, the final performance at the 800k iteration was worse than that at the 10k iteration. These results indicate that excluding low-quality images is crucial for enhancing SR performance. This is in contrast to other vision tasks such as image recognition, where increasing the number of training images, even with low quality images, often helps improve performance. + +Training with Places365. In the source selection step, the Places365 dataset was excluded because its quality estimated via the blockiness distribution was low. To justify this selection, we compare three datasets in Figure 6. As can be seen, the model trained on Places365 performed worse than those trained on ImageNet-1k and PASS. Similar to the learning processes obtained from low quality JPEG images, the PSNR decreased in the later phase of the learning process with Places365. These results confirmed the effectiveness of source selection before training based on the blockiness distributions. + +# 4.5 Analysis 3: Visual comparison + +Comparison of artifacts. Figure 7 compares the artifacts produced by JPEG compression with those produced by the SR models trained on DF2K using different image quality levels. Images obtained from models trained on $50\%$ and $75\%$ quality images show strong stripes or checkerboard patterns of artifacts. These artifacts appear to be more significant than those observed when the original image is compressed using JPEG, suggesting that they are induced by + +![](images/0d31d797598d68abc1e249599b9d1ac33565c56a4a8734043d41d62036345499.jpg) +AkkeraKanjinchou, Manga109 + +![](images/0862b931f81a796c6f6b747088f909f80c76b8c9e7e5c9baed3349edaf2b1545.jpg) +Fig. 7: Comparison of artifacts produced by JPEG compression (blue) and SR models trained on various JPEG quality values (red, HAT model, DF2K). Numbers indicate JPEG image quality. First and third rows: cropped images. Second and fourth rows: differential images. + +![](images/2d0325fc8fc994d65c7de56e725c520cae67d0ce7bd3c4265894d1a15d8a037e.jpg) + +model training. The predisposition to stripes and checkerboard patterns is likely due to the inductive bias inherent in the architecture of the neural network; in particular, the square or rectangular shape of the filters in the convolutional operations could cause this. Improvements to the network architecture to allow training on lower quality images would be interesting as future work. + +Qualitative examples. Figure 8 shows visual comparisons of SR models trained on DF2K, LSDIR, DiverSeg-I, and DiverSeg-P. For the image "img_011" in Urban100, we observed that models trained on DF2K are unable to recover the horizontal stripes pattern, while the other three models successfully recovered it. With the three images of Manga109, we observed that models trained on DiverSeg-I/P exhibit noticeable improvement in the character region compared to those trained on DF2K or LSDIR. + +# 4.6 Discussion and Limitations + +High-resolution datasets. In this paper, we applied the automated image evaluation pipeline to a large dataset of low-resolution images. We believe our finding contributes to the future construction of training datasets and the development of neural network architectures. Specifically, we demonstrated that SR models can be trained without high-resolution images. However, this does not imply that high-resolution image datasets are worthless. Rather, we believe that increasing the diversity of high-resolution images through our proposed filtering method could further improve SR performance. In Table 8, we examined this on + +![](images/beb5ae6da64ad56843148dee603f6ebba143057f079a8aefbcac6df2b9271b71.jpg) +Fig. 8: Visual comparison of $\times 4$ SR models trained on DF2K, LSDIR, DiverSeg-I, and DiverSeg-P datasets. PSNR/SSIM is calculated for each cropped patch individually to better reflect the differences in performance. + +Table 8: Applying object-based filtering to LSDIR with $\theta = {100}$ (HAT model). + +
Dataset#ImagesSet5Set14BSD100Urban100Manga109
PSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIM
LSDIR89,99132.930.905329.290.798828.010.752528.450.846932.570.9306
w/ filtering31,56132.950.905629.330.799128.020.752628.530.848732.600.9308
+ +the LSDIR dataset. Our results show that filtering out images with less than 100 object regions led to increased PSNR and SSIM with the HAT model. For future work, exploring a hybrid approach that utilizes both low- and high-resolution diverse images could be promising. + +Limitations. In this study, we focused on two perspectives: resolution and diversity. However, when collecting large datasets, there are other important perspectives, such as fairness and copyright. In addition, for benchmarking purposes, we used the five most commonly used SR datasets for a fair comparison with conventional methods. Nevertheless, real-world applications also face the challenge of blind super-resolution, where the degradation process is unknown. Creating datasets addressing this aspect would be an interesting future research direction. + +# 5 Conclusion + +In this work, we investigated the effect of the training data for SR and showed that SR models are trainable even without using high-resolution images by applying the image evaluation pipeline to a set of large low-resolution images. In experiments, we thoroughly analyzed the effect of image quality and diversity to SR performance. We hope that this work will positively influence the future construction of training datasets and lead to better models. + +# Acknowledgements + +Computational resource of AI Bridging Cloud Infrastructure (ABCI) provided by National Institute of Advanced Industrial Science and Technology (AIST) was used. + +# References + +1. Agustsson, E., Timofte, R.: Ntire 2017 challenge on single image super-resolution: Dataset and study. In: CVPRW (2017) 2, 4, 8, 9 +2. Asano, Y.M., Rupprecht, C., Zisserman, A., Vedaldi, A.: Pass: An imagenet replacement for self-supervised pretraining without humans. In: NeurIPS Track on Datasets and Benchmarks (2021) 3, 8 +3. Bevilacqua, M., Roumy, A., Guillemot, C., Alberi-Morel, M.L.: Low-complexity single-image super-resolution based on nonnegative neighbor embedding. In: BMVC (2012) 9 +4. Bhardwaj, D., Pankajakshan, V.: A jpeg blocking artifact detector for image forensics. Signal Processing: Image Communication 68, 155-161 (2018) 3, 6 +5. Chen, H., Wang, Y., Guo, T., Xu, C., Deng, Y., Liu, Z., Ma, S., Xu, C., Xu, C., Gao, W.: Pre-trained image processing transformer. In: CVPR (2021) 1, 4, 5, 10 +6. Chen, X., Wang, X., Zhou, J., Qiao, Y., Dong, C.: Activating more pixels in image super-resolution transformer. In: CVPR (2023) 1, 4, 5, 9, 10 +7. Dai, T., Cai, J., Zhang, Y., Xia, S.T., Zhang, L.: Second-order attention network for single image super-resolution. In: CVPR (2019) 4, 10 +8. Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: Imagenet: A large-scale hierarchical image database. In: CVPR (2009) 3, 8 +9. Dong, C., Loy, C.C., He, K., Tang, X.: Learning a deep convolutional network for image super-resolution. In: ECCV (2014) 1, 4 +0. Dong, C., Loy, C.C., Tang, X.: Accelerating the super-resolution convolutional neural network. In: ECCV (2016) 1, 4 +1. Huang, J.B., Singh, A., Ahuja, N.: Single image super-resolution from transformed self-exemplars. In: CVPR (2015) 9 +2. Kim, J., Lee, J.K., Lee, K.M.: Accurate image super-resolution using very deep convolutional networks. In: CVPR (2016) 1, 4 +3. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., Dollar, P., Girshick, R.: Segment anything. In: ICCV (2023) 7 +4. Ledig, C., Theis, L., Huszár, F., Caballero, J., Cunningham, A., Acosta, A., Aitken, A., Tejani, A., Totz, J., Wang, Z., et al.: Photo-realistic single image superresolution using a generative adversarial network. In: CVPR (2017) 1, 4 +5. Li, A., Zhang, L., Liu, Y., Zhu, C.: Feature modulation transformer: Crossrefinement of global representation via high-frequency prior for image superresolution. In: CVPR (2023) 2 +6. Li, W., Lu, X., Qian, S., Lu, J., Zhang, X., Jia, J.: On efficient transformer-based image pre-training for low-level vision. arXiv preprint arXiv:2112.10175 (2021) 1, 4, 5, 10 +7. Li, Y., Zhang, K., Liang, J., Cao, J., Liu, C., Gong, R., Zhang, Y., Tang, H., Liu, Y., Demandolx, D., Ranjan, R., Timofte, R., Van Gool, L.: Lsdir: A large scale dataset for image restoration. In: CVPRW (2023) 2, 5, 8, 9 + +18. Li, Y., Zhang, Y., Timofte, R., Van Gool, L., Yu, L., Li, Y., Li, X., Jiang, T., Wu, Q., Han, M., et al.: Ntire 2023 challenge on efficient super-resolution: Methods and results. In: CVPRW (2023) 2 +19. Liang, J., Cao, J., Sun, G., Zhang, K., Van Gool, L., Timofte, R.: Swinir: Image restoration using swin transformer. In: ICCVW (2021) 1, 2, 4, 9, 10 +20. Lim, B., Son, S., Kim, H., Nah, S., Mu Lee, K.: Enhanced deep residual networks for single image super-resolution. In: CVPRW (2017) 1, 4, 9 +21. Lin, Z., Garg, P., Banerjee, A., Magid, S.A., Sun, D., Zhang, Y., Van Gool, L., Wei, D., Pfister, H.: Revisiting rcan: Improved training for image super-resolution. arXiv preprint arXiv:2201.11279 (2022) 10 +22. Martin, D., Fowlkes, C., Tal, D., Malik, J.: A database of human segmented natural images and its application to evaluating segmentation algorithms and measuring ecological statistics. In: ICCV (2001) 4, 9 +23. Matsui, Y., Ito, K., Aramaki, Y., Fujimoto, A., Ogawa, T., Yamasaki, T., Aizawa, K.: Sketch-based manga retrieval using manga109 dataset. Multimedia Tools and Applications 76, 21811-21838 (2017) 9 +24. Mei, Y., Fan, Y., Zhou, Y.: Image super-resolution with non-local sparse attention. In: CVPR (2021) 10 +25. Niu, B., Wen, W., Ren, W., Zhang, X., Yang, L., Wang, S., Zhang, K., Cao, X., Shen, H.: Single image super-resolution via a holistic attention network. In: ECCV (2020) 10 +26. Shi, W., Caballero, J., Huszar, F., Totz, J., Aitken, A.P., Bishop, R., Rueckert, D., Wang, Z.: Real-time single image and video super-resolution using an efficient sub-pixel convolutional neural network. In: CVPR (2016) 1, 4 +27. Timofte, R., Agustsson, E., Van Gool, L., Yang, M.H., Zhang, L.: Ntire 2017 challenge on single image super-resolution: Methods and results. In: CVPRW (2017) 2, 4, 8, 9 +28. Tong, T., Li, G., Liu, X., Gao, Q.: Image super-resolution using dense skip connections. In: ICCV (2017) 1, 4 +29. Wang, X., Yu, K., Wu, S., Gu, J., Liu, Y., Dong, C., Qiao, Y., Change Loy, C.: Esrgan: Enhanced super-resolution generative adversarial networks. In: ECCVW (2018) 1, 2, 4, 9, 10 +30. Yang, J., Wright, J., Huang, T., Ma, Y.: Image super-resolution as sparse representation of raw image patches. In: CVPR (2008) 4 +31. Yang, Q., Chen, D., Tan, Z., Liu, Q., Chu, Q., Bao, J., Yuan, L., Hua, G., Yu, N.: Hq-50k: A large-scale, high-quality dataset for image restoration. arXiv preprint arXiv:2306.05390 (2023) 2, 5 +32. Zeyde, R., Elad, M., Protter, M.: On single image scale-up using sparse-representations. In: Proc. 7th Int. Conf. Curves Surf. (2010) 9 +33. Zhang, D., Huang, F., Liu, S., Wang, X., Jin, Z.: Swinfir: Revisiting the swinir with fast fourier convolution and improved training for image super-resolution. arXiv preprint arXiv:2208.11247 (2022) 1, 4 +34. Zhang, X., Zeng, H., Guo, S., Zhang, L.: Efficient long-range attention network for image super-resolution. In: ECCV (2022) 1, 4 +35. Zhang, Y., Li, K., Zhong, B., Fu, Y.: Residual non-local attention networks for image restoration. In: ICLR (2019) 1, 4 +36. Zhang, Y., Li, K., Li, K., Wang, L., Zhong, B., Fu, Y.: Image super-resolution using very deep residual channel attention networks. In: ECCV (2018) 1, 4, 9 +37. Zhang, Y., Tian, Y., Kong, Y., Zhong, B., Fu, Y.: Residual dense network for image super-resolution. In: CVPR (2018) 1, 4 + +38. Zhang, Y., Zhang, K., Chen, Z., Li, Y., Timofte, R., Zhang, J., Zhang, K., Peng, R., Ma, Y., Jia, L., et al.: Ntire 2023 challenge on image super-resolution (x4): Methods and results. In: CVPRW (2023) 2 +39. Zhou, B., Lapedriza, A., Xiao, J., Torralba, A., Oliva, A.: Learning deep features for scene recognition using places database. Advances in neural information processing systems 27 (2014) 8 +40. Zhou, S., Zhang, J., Zuo, W., Loy, C.C.: Cross-scale internal graph neural network for image super-resolution. Advances in neural information processing systems 33, 3499-3509 (2020) 10 +41. Zhou, X., Girdhar, R., Joulin, A., Krahenbuhl, P., Misra, I.: Detecting twenty-thousand classes using image-level supervision. In: ECCV (2022) 8 +42. Zhou, Y., Li, Z., Guo, C.L., Bai, S., Cheng, M.M., Hou, Q.: Srformer: Permuted self-attention for single image super-resolution. In: ICCV (2023) 1, 2, 4 \ No newline at end of file diff --git a/2024/Rethinking Image Super Resolution from Training Data Perspectives/images.zip b/2024/Rethinking Image Super Resolution from Training Data Perspectives/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..54a345de33520dbf3fcaa041a224b9c993cf94f0 --- /dev/null +++ b/2024/Rethinking Image Super Resolution from Training Data Perspectives/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6259ee29ea39ac35601c60530f35f3c5097c1245603730cb83db03b994e8a409 +size 1131863 diff --git a/2024/Rethinking Image Super Resolution from Training Data Perspectives/layout.json b/2024/Rethinking Image Super Resolution from Training Data Perspectives/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..0c05c7c67fdaab6da25025ca9f9842dbece5a4e2 --- /dev/null +++ b/2024/Rethinking Image Super Resolution from Training Data Perspectives/layout.json @@ -0,0 +1,9172 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 163, + 111, + 451, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 163, + 111, + 451, + 148 + ], + "spans": [ + { + "bbox": [ + 163, + 111, + 451, + 148 + ], + "type": "text", + "content": "Rethinking Image Super-Resolution from Training Data Perspectives" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 140, + 167, + 474, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 140, + 167, + 474, + 205 + ], + "spans": [ + { + "bbox": [ + 140, + 167, + 474, + 205 + ], + "type": "text", + "content": "Go Ohtani" + }, + { + "bbox": [ + 140, + 167, + 474, + 205 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 140, + 167, + 474, + 205 + ], + "type": "text", + "content": ", Ryu Tadokoro" + }, + { + "bbox": [ + 140, + 167, + 474, + 205 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 140, + 167, + 474, + 205 + ], + "type": "text", + "content": ", Ryosuke Yamada" + }, + { + "bbox": [ + 140, + 167, + 474, + 205 + ], + "type": "inline_equation", + "content": "^{2,3}" + }, + { + "bbox": [ + 140, + 167, + 474, + 205 + ], + "type": "text", + "content": ", Yuki M. Asano" + }, + { + "bbox": [ + 140, + 167, + 474, + 205 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 140, + 167, + 474, + 205 + ], + "type": "text", + "content": ", Iro Laina" + }, + { + "bbox": [ + 140, + 167, + 474, + 205 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 140, + 167, + 474, + 205 + ], + "type": "text", + "content": ", Christian Rupprecht" + }, + { + "bbox": [ + 140, + 167, + 474, + 205 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 140, + 167, + 474, + 205 + ], + "type": "text", + "content": ", Nakamasasa Inoue" + }, + { + "bbox": [ + 140, + 167, + 474, + 205 + ], + "type": "inline_equation", + "content": "^{6,2}" + }, + { + "bbox": [ + 140, + 167, + 474, + 205 + ], + "type": "text", + "content": ", Rio Yokota" + }, + { + "bbox": [ + 140, + 167, + 474, + 205 + ], + "type": "inline_equation", + "content": "^{6,2}" + }, + { + "bbox": [ + 140, + 167, + 474, + 205 + ], + "type": "text", + "content": ", Hirokatsu Kataoka" + }, + { + "bbox": [ + 140, + 167, + 474, + 205 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 140, + 167, + 474, + 205 + ], + "type": "text", + "content": ", and Yoshimitsu Aoki" + }, + { + "bbox": [ + 140, + 167, + 474, + 205 + ], + "type": "inline_equation", + "content": "^{1}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 271, + 213, + 343, + 225 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 271, + 213, + 343, + 225 + ], + "spans": [ + { + "bbox": [ + 271, + 213, + 343, + 225 + ], + "type": "text", + "content": "1 Keio University" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 152, + 224, + 460, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 224, + 460, + 236 + ], + "spans": [ + { + "bbox": [ + 152, + 224, + 460, + 236 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 152, + 224, + 460, + 236 + ], + "type": "text", + "content": " National Institute of Advanced Industrial Science and Technology (AIST)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 257, + 236, + 357, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 257, + 236, + 357, + 247 + ], + "spans": [ + { + "bbox": [ + 257, + 236, + 357, + 247 + ], + "type": "text", + "content": "3 University of Tsukuba" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 158, + 247, + 455, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 158, + 247, + 455, + 258 + ], + "spans": [ + { + "bbox": [ + 158, + 247, + 455, + 258 + ], + "type": "text", + "content": "4 University of Amsterdam, now at University of Technology Nuremberg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 260, + 258, + 353, + 268 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 260, + 258, + 353, + 268 + ], + "spans": [ + { + "bbox": [ + 260, + 258, + 353, + 268 + ], + "type": "text", + "content": "5 University of Oxford" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 241, + 269, + 372, + 280 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 241, + 269, + 372, + 280 + ], + "spans": [ + { + "bbox": [ + 241, + 269, + 372, + 280 + ], + "type": "inline_equation", + "content": "^{6}" + }, + { + "bbox": [ + 241, + 269, + 372, + 280 + ], + "type": "text", + "content": " Tokyo Institute of Technology" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 159, + 306, + 455, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 306, + 455, + 483 + ], + "spans": [ + { + "bbox": [ + 159, + 306, + 455, + 483 + ], + "type": "text", + "content": "Abstract. In this work, we investigate the understudied effect of the training data used for image super-resolution (SR). Most commonly, novel SR methods are developed and benchmarked on common training datasets such as DIV2K and DF2K. However, we investigate and rethink the training data from the perspectives of diversity and quality, thereby addressing the question of \"How important is SR training for SR models?\" To this end, we propose an automated image evaluation pipeline. With this, we stratify existing high-resolution image datasets and larger-scale image datasets such as ImageNet and PASS to compare their performances. We find that datasets with (i) low compression artifacts, (ii) high within-image diversity as judged by the number of different objects, and (iii) a large number of images from ImageNet or PASS all positively affect SR performance. We hope that the proposed simple-yet-effective dataset curation pipeline will inform the construction of SR datasets in the future and yield overall better models. Code is available at: https://github.com/gohtanii/DiverSeg-dataset" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 159, + 492, + 453, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 492, + 453, + 514 + ], + "spans": [ + { + "bbox": [ + 159, + 492, + 453, + 514 + ], + "type": "text", + "content": "Keywords: Super-resolution dataset " + }, + { + "bbox": [ + 159, + 492, + 453, + 514 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 492, + 453, + 514 + ], + "type": "text", + "content": " Image compression " + }, + { + "bbox": [ + 159, + 492, + 453, + 514 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 492, + 453, + 514 + ], + "type": "text", + "content": " Image diversity" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 533, + 230, + 547 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 533, + 230, + 547 + ], + "spans": [ + { + "bbox": [ + 132, + 533, + 230, + 547 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "text", + "content": "Image super-resolution (SR) aims to reconstruct high-resolution images from low-resolution images. It has been considered as one of the most fundamental tasks in the field of computer vision, with applications ranging from autonomous driving to medical imaging. Deep learning methods have led to significant advances in SR over the last decade, focusing primarily on improvements in the neural network architectures. Early SR models rely on convolutional neural networks (CNNs) [9,10,12,14,20,26,28,29,35-37]. Recent innovations have given rise to transformer-based SR models [5,6,16,19,33,34,42], which have consistently improved the performances." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 136, + 115, + 474, + 335 + ], + "blocks": [ + { + "bbox": [ + 136, + 115, + 474, + 335 + ], + "lines": [ + { + "bbox": [ + 136, + 115, + 474, + 335 + ], + "spans": [ + { + "bbox": [ + 136, + 115, + 474, + 335 + ], + "type": "image", + "image_path": "dfdd78fcf01ffef938e6622e91f2f74a50994f16ad0ad0b8a3eafbfdb911127a.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 344, + 482, + 390 + ], + "lines": [ + { + "bbox": [ + 130, + 344, + 482, + 390 + ], + "spans": [ + { + "bbox": [ + 130, + 344, + 482, + 390 + ], + "type": "text", + "content": "Fig. 1: We propose an automated image evaluation pipeline to curate a dataset for training SR models. The obtained dataset, namely DiverSeg, consists of low-resolution but high-quality images with many object regions. SR models trained on DiverSeg outperform those trained on high-resolution image datasets such as DF2K and LSDIR." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 412, + 482, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 412, + 482, + 496 + ], + "spans": [ + { + "bbox": [ + 130, + 412, + 482, + 496 + ], + "type": "text", + "content": "With the improvement of neural network architectures, the importance of training datasets has also increased, as discussed in [17]. Examples of high-resolution datasets include DIV2K [1] and Flickr2K [27]. The combined dataset of these two, referred to as DF2K, is often utilized for training SR models. Most recently, LSDIR [17] has been proposed, which consists of 84,991 high-resolution images. It has been confirmed that training on large high-resolution datasets contributes significantly to performance improvement [18, 19, 29, 38, 42]" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 496, + 482, + 520 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 496, + 482, + 520 + ], + "spans": [ + { + "bbox": [ + 130, + 496, + 482, + 520 + ], + "type": "text", + "content": "The conventional approach to constructing datasets relies on a manual evaluation step, where the following two perspectives are most commonly considered:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 135, + 526, + 482, + 635 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 135, + 526, + 482, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 526, + 482, + 586 + ], + "spans": [ + { + "bbox": [ + 135, + 526, + 482, + 586 + ], + "type": "text", + "content": "1. Resolution and quality [17, 31]. This perspective focuses on the pixel density of images. Images that do not meet the specified resolution threshold are excluded. Typically, HD, 2K, and 4K images are used to create a dataset. After the initial automatic filtering based on image size, the details of each image are manually evaluated to identify and exclude compressed images." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 135, + 586, + 482, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 586, + 482, + 635 + ], + "spans": [ + { + "bbox": [ + 135, + 586, + 482, + 635 + ], + "type": "text", + "content": "2. Variety and diversity [15]. This includes diversity in subjects (e.g., people, landscapes, urban scenes), lighting conditions, colors, textures, and other photographic elements. A diverse dataset is said to help train a model that is robust and performs well across a wide range of domains." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "content": "Datasets constructed from these perspectives have been shown to significantly improve the performance of SR models. However, they also pose challenges in" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 232, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 232, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 232, + 101 + ], + "type": "text", + "content": "G. Ohtani et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": "scaling the datasets, as collecting uncompressed high-resolution images is difficult and costly." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 145, + 482, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 145, + 482, + 240 + ], + "spans": [ + { + "bbox": [ + 130, + 145, + 482, + 240 + ], + "type": "text", + "content": "To address this limitation, this paper rethinks these perspectives and proposes Diverse Segmentation dataset (DiverSeg), a low-resolution1 yet effective image dataset for training SR models. As shown in Figure 1, the dataset is constructed by applying filtering to a large set of low-resolution images, such as ImageNet-1k [8] and PASS [2]. In experiments, we demonstrate that models trained on DiverSeg outperform those trained on high-resolution image datasets such as DF2K and LSDIR. Based on this finding, our contributions are summarized as follows." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 246, + 482, + 628 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 130, + 246, + 482, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 246, + 482, + 354 + ], + "spans": [ + { + "bbox": [ + 130, + 246, + 482, + 354 + ], + "type": "text", + "content": "1) Rethinking the resolution perspective. High-resolution images have been considered to be necessary for training SR models. In this work, we challenge this traditional perspective and show that SR models can be trained without high-resolution images. Specifically, we introduce a method to estimate image quality based on the kernel density estimation over blockiness values [4] that estimates the quantity distribution of blocking artifacts. We demonstrate that low-resolution images with high quality, indicated by reduced artifacts, can improve the performance of SR models. We also thoroughly analyze the impact of image quality on SR performance." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 358, + 482, + 455 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 358, + 482, + 455 + ], + "spans": [ + { + "bbox": [ + 130, + 358, + 482, + 455 + ], + "type": "text", + "content": "2) Rethinking the diversity perspective. When constructing datasets for training SR models, images containing only a small number of objects are often implicitly excluded during the manual resolution evaluation process. This is because evaluators typically focus on the details of objects or small objects in the images. Therefore, we explicitly calculate the number of objects in images and analyze how this number affects SR performance. In our experiments, we show that constructing datasets with images containing many objects improves the performance of SR models." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 459, + 482, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 459, + 482, + 628 + ], + "spans": [ + { + "bbox": [ + 130, + 459, + 482, + 628 + ], + "type": "text", + "content": "3) Dataset construction. To facilitate analysis from the above two perspectives, we introduce a framework that automatically creates a dataset from a set of low-resolution images collected from the web. Specifically, the framework consists of two steps, source selection and object-based filtering, which correspond to the first and second perspectives, respectively. Our framework eliminates the need for manual assessment, facilitating dataset scaling. In the first step, given several low-resolution image datasets, we estimate the image quality of them through their blockiness distributions and select datasets with quality larger than " + }, + { + "bbox": [ + 130, + 459, + 482, + 628 + ], + "type": "inline_equation", + "content": "90\\%" + }, + { + "bbox": [ + 130, + 459, + 482, + 628 + ], + "type": "text", + "content": ". In the second step, we filter out images with a small number of object regions by using object detection and image segmentation models. The filtered dataset, which we refer to as the DiverSeg dataset, is utilized for training various SR models in our experiments. We apply our framework to a union set of ImageNet, Places365, and PASS to construct the DiverSeg dataset. We demonstrate that SR models trained on DiverSeg archive state-of-the-art performance." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 167, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 167, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 167, + 91, + 447, + 102 + ], + "type": "text", + "content": "Rethinking Image Super-Resolution from Training Data Perspectives" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 133, + 653, + 451, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 653, + 451, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 653, + 451, + 665 + ], + "type": "text", + "content": "1 We define images with a resolution lower than HD as low-resolution images." + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 238, + 127 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 238, + 127 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 238, + 127 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 141, + 316, + 154 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 141, + 316, + 154 + ], + "spans": [ + { + "bbox": [ + 132, + 141, + 316, + 154 + ], + "type": "text", + "content": "2.1 Image super-resolution models" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 161, + 481, + 220 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 161, + 481, + 220 + ], + "spans": [ + { + "bbox": [ + 130, + 161, + 481, + 220 + ], + "type": "text", + "content": "A number of SR models have been proposed that take advantage of deep learning techniques. These models can be divided into two groups, CNN-based models [9, 10,12,14,20,26,28,29,35-37] and Transformer-based models [5,6,16,19,33,34,42]. Each group exploits different architectural strengths to enhance low-resolution images to high resolution." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 221, + 481, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 221, + 481, + 399 + ], + "spans": [ + { + "bbox": [ + 130, + 221, + 481, + 399 + ], + "type": "text", + "content": "CNN-based models. SRCNN [9] was the first model that integrates a deep convolutional architecture for SR. Subsequently, FSRCNN [10] significantly improved computational efficiency by performing convolutional processing in low resolution space and upsampling in the last layer. Furthermore, ESPCN [26] adopted an efficient upsampling method, sub-pixel convolution, to enhance performance while reducing computational costs. These methods established the basic structure of modern SR networks and laid the foundation for subsequent research developments. Later studies introduced various modules such as residual connections [12, 14, 20, 29], dense blocks [28, 29, 37], and attention mechanisms [7, 35, 36]. EDSR [20] eliminated batch normalization layers and introduced residual scaling to enable stable training of large models. MSRResNet [29] replaced the basic ResNet blocks with residual dense blocks to improve the balance between performance and computational efficiency. RCAN [36] incorporated channel attention mechanisms to adaptively weight feature representations, significantly enhancing SR performance." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 401, + 482, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 401, + 482, + 556 + ], + "spans": [ + { + "bbox": [ + 130, + 401, + 482, + 556 + ], + "type": "text", + "content": "Transformer-based models. As the first Transformer-based image restoration model, IPT [5] was introduced as a large-scale model utilizing the Transformer's encoder and decoder architecture. By pre-training on ImageNet, IPT significantly improved SR performance, fully leveraging the capabilities of the Transformer. SwinIR [19] executes self-attention within local windows during feature extraction, demonstrating exceptional SR performance and establishing itself as the foundational model for Transformers in SR. Following this, models that build upon SwinIR have been developed [6,33,34,42], enhancing performance by extending and optimizing the self-attention mechanism. Among them, HAT [6] achieves state-of-the-art performance in SR by using a duplicated cross-attention module and pre-training on ImageNet. However, there may be potential for improvement, as ImageNet is primarily an image recognition dataset and may not be fully optimized for SR." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 574, + 322, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 574, + 322, + 586 + ], + "spans": [ + { + "bbox": [ + 132, + 574, + 322, + 586 + ], + "type": "text", + "content": "2.2 Image super-resolution datasets" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 594, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 481, + 666 + ], + "type": "text", + "content": "T91 [30] and BSDS200 [22] are early datasets used to train SR models, consisting of 91 and 200 images, respectively. The turning point in training datasets was ushered by the release of DIV2K [1], a compilation of 800 high-resolution images with minimal compression noise meticulously collected from the web. Subsequently, to accommodate the further scaling of the model, Flickr2K [27], consisting of 2,650 high-resolution images, is merged into DIV2K and is referred" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 233, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 233, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 233, + 102 + ], + "type": "text", + "content": "G. Ohtani et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 115, + 479, + 224 + ], + "blocks": [ + { + "bbox": [ + 133, + 115, + 479, + 224 + ], + "lines": [ + { + "bbox": [ + 133, + 115, + 479, + 224 + ], + "spans": [ + { + "bbox": [ + 133, + 115, + 479, + 224 + ], + "type": "image", + "image_path": "dea9793fed3a8a8306fa8e627c1f807638b424faceb95b2451095ca6ff31d578.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 236, + 480, + 258 + ], + "lines": [ + { + "bbox": [ + 131, + 236, + 480, + 258 + ], + "spans": [ + { + "bbox": [ + 131, + 236, + 480, + 258 + ], + "type": "text", + "content": "Fig. 2: Images of DiverSeg with their segmentation masks. DiverSeg is obtained from a large set of low-resolution images through the automated image evaluation pipeline." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 280, + 482, + 423 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 280, + 482, + 423 + ], + "spans": [ + { + "bbox": [ + 130, + 280, + 482, + 423 + ], + "type": "text", + "content": "to as DF2K. In recent years, to further expand the scale of SR datasets, datasets larger than DF2K have been released, such as LSDIR [17], comprising 84,991 images and HQ-50K [31] consisting of 50,000 images. These datasets persistently adhere to stringent criteria, ensuring high resolution and negligible compression noise in the imagery. However, since collecting images that meet the aforementioned conditions is quite challenging, these datasets still consist of only tens of thousands of images. In contrast, several approaches [5, 6, 16] adopt ImageNet, also known as ImageNet-1k, which consists of 1.28M images spanning diverse categories. These approaches leverage the diversity of texture patterns in ImageNet as an advantage. Nevertheless, ImageNet is pointed out to contain some images having low-resolution and JPEG-compression artifacts, which adversely affect the training results." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 441, + 254, + 453 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 441, + 254, + 453 + ], + "spans": [ + { + "bbox": [ + 132, + 441, + 254, + 453 + ], + "type": "text", + "content": "3 DiverSeg dataset" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 464, + 482, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 464, + 482, + 559 + ], + "spans": [ + { + "bbox": [ + 130, + 464, + 482, + 559 + ], + "type": "text", + "content": "This section presents Diverse Segmentation dataset (DiverSeg), our dataset for training SR models without using high-resolution images. As shown in Figure 2, the DiverSeg dataset consists of low-resolution but high-quality images with diverse segmented object regions. The framework for constructing the dataset consists of two steps, source selection and object-based filtering, which are designed from our perspective of rethinking resolution and diversity, respectively. Our approach eliminates the manual cost of collecting and quality-checking high-resolution images." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 575, + 242, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 575, + 242, + 586 + ], + "spans": [ + { + "bbox": [ + 132, + 575, + 242, + 586 + ], + "type": "text", + "content": "3.1 Source selection" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 594, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 665 + ], + "type": "text", + "content": "Let " + }, + { + "bbox": [ + 130, + 594, + 482, + 665 + ], + "type": "inline_equation", + "content": "\\mathcal{X}" + }, + { + "bbox": [ + 130, + 594, + 482, + 665 + ], + "type": "text", + "content": " be a set of low-resolution image datasets. This step filters out low-quality datasets by estimating the quality " + }, + { + "bbox": [ + 130, + 594, + 482, + 665 + ], + "type": "inline_equation", + "content": "\\hat{q}_X \\in [0,1]" + }, + { + "bbox": [ + 130, + 594, + 482, + 665 + ], + "type": "text", + "content": " for each dataset " + }, + { + "bbox": [ + 130, + 594, + 482, + 665 + ], + "type": "inline_equation", + "content": "X \\in \\mathcal{X}" + }, + { + "bbox": [ + 130, + 594, + 482, + 665 + ], + "type": "text", + "content": " and excluding those with " + }, + { + "bbox": [ + 130, + 594, + 482, + 665 + ], + "type": "inline_equation", + "content": "\\hat{q}_X < 0.9" + }, + { + "bbox": [ + 130, + 594, + 482, + 665 + ], + "type": "text", + "content": ", under the assumption that low-quality images are detrimental when training SR models. We introduce a quality estimation method based on the Kullback-Leibler (KL) divergence between blockiness distributions as detailed below." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 167, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 167, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 167, + 91, + 447, + 102 + ], + "type": "text", + "content": "Rethinking Image Super-Resolution from Training Data Perspectives" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 137, + 114, + 482, + 203 + ], + "blocks": [ + { + "bbox": [ + 137, + 114, + 482, + 203 + ], + "lines": [ + { + "bbox": [ + 137, + 114, + 482, + 203 + ], + "spans": [ + { + "bbox": [ + 137, + 114, + 482, + 203 + ], + "type": "image", + "image_path": "51f21f57bc9f13566804ac593d29b853c03b7caaf49014cc9d08b4949e5874b4.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 206, + 482, + 240 + ], + "lines": [ + { + "bbox": [ + 131, + 206, + 482, + 240 + ], + "spans": [ + { + "bbox": [ + 131, + 206, + 482, + 240 + ], + "type": "text", + "content": "Fig. 3: Comparison of image degradation due to JPEG quality(blue). Blockiness values calculated from the images are marked. As the JPEG quality decreases and artifacts increase, we observe a corresponding rise in blockiness values." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 135, + 252, + 306, + 361 + ], + "blocks": [ + { + "bbox": [ + 135, + 252, + 306, + 361 + ], + "lines": [ + { + "bbox": [ + 135, + 252, + 306, + 361 + ], + "spans": [ + { + "bbox": [ + 135, + 252, + 306, + 361 + ], + "type": "image", + "image_path": "7681ccaf48f120ec4f05e6755a7a09218b84820466595b61affb1bb84866e47b.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 369, + 482, + 404 + ], + "lines": [ + { + "bbox": [ + 131, + 369, + 482, + 404 + ], + "spans": [ + { + "bbox": [ + 131, + 369, + 482, + 404 + ], + "type": "text", + "content": "Fig. 4: (a) Blockiness distributions " + }, + { + "bbox": [ + 131, + 369, + 482, + 404 + ], + "type": "inline_equation", + "content": "p_{X,1.0}" + }, + { + "bbox": [ + 131, + 369, + 482, + 404 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 131, + 369, + 482, + 404 + ], + "type": "inline_equation", + "content": "X = \\mathrm{ImageNet - 1k}" + }, + { + "bbox": [ + 131, + 369, + 482, + 404 + ], + "type": "text", + "content": ", Places365 and PASS. (b) Basis distributions " + }, + { + "bbox": [ + 131, + 369, + 482, + 404 + ], + "type": "inline_equation", + "content": "p_{Z,q}" + }, + { + "bbox": [ + 131, + 369, + 482, + 404 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 131, + 369, + 482, + 404 + ], + "type": "inline_equation", + "content": "Z = \\mathrm{DF2K}" + }, + { + "bbox": [ + 131, + 369, + 482, + 404 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 369, + 482, + 404 + ], + "type": "inline_equation", + "content": "q = 0.5, 0.75, 0.85, 0.95, 1.0" + }, + { + "bbox": [ + 131, + 369, + 482, + 404 + ], + "type": "text", + "content": ". We estimate the quality by comparing " + }, + { + "bbox": [ + 131, + 369, + 482, + 404 + ], + "type": "inline_equation", + "content": "p_{X,1.0}" + }, + { + "bbox": [ + 131, + 369, + 482, + 404 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 369, + 482, + 404 + ], + "type": "inline_equation", + "content": "p_{Z,q}" + }, + { + "bbox": [ + 131, + 369, + 482, + 404 + ], + "type": "text", + "content": " using the KL divergence." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 309, + 252, + 480, + 361 + ], + "blocks": [ + { + "bbox": [ + 309, + 252, + 480, + 361 + ], + "lines": [ + { + "bbox": [ + 309, + 252, + 480, + 361 + ], + "spans": [ + { + "bbox": [ + 309, + 252, + 480, + 361 + ], + "type": "image", + "image_path": "731aae67800aee7ca10548feb9822e51a430967e4189ce57b9a7969c90690e8d.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 427, + 488, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 427, + 488, + 475 + ], + "spans": [ + { + "bbox": [ + 130, + 427, + 488, + 475 + ], + "type": "text", + "content": "Image datasets. This work uses three web-collected low-resolution datasets " + }, + { + "bbox": [ + 130, + 427, + 488, + 475 + ], + "type": "inline_equation", + "content": "\\mathcal{X} = \\{\\mathrm{ImageNet - 1k},\\mathrm{Places365},\\mathrm{PASS}\\}" + }, + { + "bbox": [ + 130, + 427, + 488, + 475 + ], + "type": "text", + "content": ". Training SR models on them is not straightforward because they may include highly compressed images that negatively affect training of SR models." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 475, + 482, + 536 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 475, + 482, + 536 + ], + "spans": [ + { + "bbox": [ + 130, + 475, + 482, + 536 + ], + "type": "text", + "content": "Quality definition. Let " + }, + { + "bbox": [ + 130, + 475, + 482, + 536 + ], + "type": "inline_equation", + "content": "Y" + }, + { + "bbox": [ + 130, + 475, + 482, + 536 + ], + "type": "text", + "content": " be a dataset of JPEG images. We define the quality " + }, + { + "bbox": [ + 130, + 475, + 482, + 536 + ], + "type": "inline_equation", + "content": "q_{Y}" + }, + { + "bbox": [ + 130, + 475, + 482, + 536 + ], + "type": "text", + "content": " by the average JPEG quality, i.e., " + }, + { + "bbox": [ + 130, + 475, + 482, + 536 + ], + "type": "inline_equation", + "content": "q_{Y} \\triangleq \\frac{1}{|Y|} \\sum_{y \\in Y} Q(y)" + }, + { + "bbox": [ + 130, + 475, + 482, + 536 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 130, + 475, + 482, + 536 + ], + "type": "inline_equation", + "content": "0 \\leq Q(y) \\leq 1" + }, + { + "bbox": [ + 130, + 475, + 482, + 536 + ], + "type": "text", + "content": " is the JPEG quality of an image " + }, + { + "bbox": [ + 130, + 475, + 482, + 536 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 130, + 475, + 482, + 536 + ], + "type": "text", + "content": ". The goal of quality estimation is to estimate " + }, + { + "bbox": [ + 130, + 475, + 482, + 536 + ], + "type": "inline_equation", + "content": "q_{X}" + }, + { + "bbox": [ + 130, + 475, + 482, + 536 + ], + "type": "text", + "content": " given a dataset " + }, + { + "bbox": [ + 130, + 475, + 482, + 536 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 130, + 475, + 482, + 536 + ], + "type": "text", + "content": ". Note that in the estimation phase, datasets may include images other than JPEG and the true quality " + }, + { + "bbox": [ + 130, + 475, + 482, + 536 + ], + "type": "inline_equation", + "content": "q_{X}" + }, + { + "bbox": [ + 130, + 475, + 482, + 536 + ], + "type": "text", + "content": " is not observable." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 536, + 482, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 536, + 482, + 574 + ], + "spans": [ + { + "bbox": [ + 130, + 536, + 482, + 574 + ], + "type": "text", + "content": "Blockiness distribution. To estimate the quality, we utilize the blockiness measure [4]. Specifically, for each image dataset " + }, + { + "bbox": [ + 130, + 536, + 482, + 574 + ], + "type": "inline_equation", + "content": "X \\in \\mathcal{X}" + }, + { + "bbox": [ + 130, + 536, + 482, + 574 + ], + "type": "text", + "content": ", we estimate the distribution of blockiness values " + }, + { + "bbox": [ + 130, + 536, + 482, + 574 + ], + "type": "inline_equation", + "content": "p_{X,q}(b)" + }, + { + "bbox": [ + 130, + 536, + 482, + 574 + ], + "type": "text", + "content": " by kernel density estimation as follows:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 216, + 590, + 481, + 620 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 216, + 590, + 481, + 620 + ], + "spans": [ + { + "bbox": [ + 216, + 590, + 481, + 620 + ], + "type": "interline_equation", + "content": "p _ {X, q} (b) = \\frac {1}{h | X |} \\sum_ {x \\in X} K \\left(\\frac {b - B \\left(\\phi_ {q} (x)\\right)}{h}\\right), \\tag {1}", + "image_path": "1d28ab788afa44f7882da1815d97a262d7ea80b419ef8f4cd91adfc58e987b15.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "text", + "content": " is an image, " + }, + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\phi_q" + }, + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "text", + "content": " is the JPEG compression function, " + }, + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "inline_equation", + "content": "q \\in [0,1]" + }, + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "text", + "content": " is a quality value, " + }, + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "inline_equation", + "content": "K: \\mathbb{R} \\to \\mathbb{R}" + }, + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "text", + "content": " is a Gaussian kernel, and " + }, + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "inline_equation", + "content": "h \\in \\mathbb{R}" + }, + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "text", + "content": " is the bandwidth determined by Scott's method. The function " + }, + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "text", + "content": " is the blockiness measure that measures the" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 232, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 232, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 232, + 101 + ], + "type": "text", + "content": "G. Ohtani et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": "quantity of blocking artifacts by computing subband DCT coefficients. Specifically, " + }, + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": " is defined on images that are decomposed into " + }, + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "inline_equation", + "content": "P \\times P" + }, + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": " patches as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 143, + 159, + 481, + 194 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 159, + 481, + 194 + ], + "spans": [ + { + "bbox": [ + 143, + 159, + 481, + 194 + ], + "type": "interline_equation", + "content": "B (x) = \\sum_ {i = 1} ^ {P} \\sum_ {j = 1} ^ {P} \\left| \\frac {\\bar {V} _ {\\text {c r o p}} (i , j) - \\bar {V} (i , j)}{\\bar {V} (i , j)} \\right|, \\quad \\bar {V} (i, j) = \\frac {1}{W H} \\sum_ {h = 1} ^ {H} \\sum_ {w = 1} ^ {W} V _ {w, h} (i, j) \\tag {2}", + "image_path": "5b37aff40894f208c22106b0f04ed027d02e55b7427ebbc5b15cf34c42c7f157.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 202, + 482, + 299 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 202, + 482, + 299 + ], + "spans": [ + { + "bbox": [ + 130, + 202, + 482, + 299 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 202, + 482, + 299 + ], + "type": "inline_equation", + "content": "W, H" + }, + { + "bbox": [ + 130, + 202, + 482, + 299 + ], + "type": "text", + "content": " are the width and height of an image " + }, + { + "bbox": [ + 130, + 202, + 482, + 299 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 130, + 202, + 482, + 299 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 202, + 482, + 299 + ], + "type": "inline_equation", + "content": "V_{w,h}(i,j)" + }, + { + "bbox": [ + 130, + 202, + 482, + 299 + ], + "type": "text", + "content": " is the variation in the " + }, + { + "bbox": [ + 130, + 202, + 482, + 299 + ], + "type": "inline_equation", + "content": "(i,j)" + }, + { + "bbox": [ + 130, + 202, + 482, + 299 + ], + "type": "text", + "content": "-th subband DCT coefficients within the " + }, + { + "bbox": [ + 130, + 202, + 482, + 299 + ], + "type": "inline_equation", + "content": "(h,w)" + }, + { + "bbox": [ + 130, + 202, + 482, + 299 + ], + "type": "text", + "content": "-th patch being calculated and its four spatially adjacent patches. " + }, + { + "bbox": [ + 130, + 202, + 482, + 299 + ], + "type": "inline_equation", + "content": "V_{\\mathrm{crop}}(i,j)" + }, + { + "bbox": [ + 130, + 202, + 482, + 299 + ], + "type": "text", + "content": " is the variation calculated similarly to " + }, + { + "bbox": [ + 130, + 202, + 482, + 299 + ], + "type": "inline_equation", + "content": "V_{w,h}(i,j)" + }, + { + "bbox": [ + 130, + 202, + 482, + 299 + ], + "type": "text", + "content": " for the given image with the first 4 rows and 4 columns removed. " + }, + { + "bbox": [ + 130, + 202, + 482, + 299 + ], + "type": "inline_equation", + "content": "\\bar{V}" + }, + { + "bbox": [ + 130, + 202, + 482, + 299 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 202, + 482, + 299 + ], + "type": "inline_equation", + "content": "\\bar{V}_{\\mathrm{crop}}" + }, + { + "bbox": [ + 130, + 202, + 482, + 299 + ], + "type": "text", + "content": " are the average variations of " + }, + { + "bbox": [ + 130, + 202, + 482, + 299 + ], + "type": "inline_equation", + "content": "V_{w,h}(i,j)" + }, + { + "bbox": [ + 130, + 202, + 482, + 299 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 202, + 482, + 299 + ], + "type": "inline_equation", + "content": "V_{\\mathrm{crop}}(i,j)" + }, + { + "bbox": [ + 130, + 202, + 482, + 299 + ], + "type": "text", + "content": " calculated for each patch, respectively. This work uses " + }, + { + "bbox": [ + 130, + 202, + 482, + 299 + ], + "type": "inline_equation", + "content": "P = 8" + }, + { + "bbox": [ + 130, + 202, + 482, + 299 + ], + "type": "text", + "content": ". The blockiness value " + }, + { + "bbox": [ + 130, + 202, + 482, + 299 + ], + "type": "inline_equation", + "content": "B(x)" + }, + { + "bbox": [ + 130, + 202, + 482, + 299 + ], + "type": "text", + "content": " is expected to be low for uncompressed images and high for compressed images as shown in Figure 3." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 299, + 489, + 336 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 299, + 489, + 336 + ], + "spans": [ + { + "bbox": [ + 130, + 299, + 489, + 336 + ], + "type": "text", + "content": "Quality estimation. We estimate the quality by comparing " + }, + { + "bbox": [ + 130, + 299, + 489, + 336 + ], + "type": "inline_equation", + "content": "p_{X,1.0}" + }, + { + "bbox": [ + 130, + 299, + 489, + 336 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 130, + 299, + 489, + 336 + ], + "type": "inline_equation", + "content": "\\{p_{Z,q}\\}_{q \\in S}" + }, + { + "bbox": [ + 130, + 299, + 489, + 336 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 299, + 489, + 336 + ], + "type": "inline_equation", + "content": "p_{Z,q}" + }, + { + "bbox": [ + 130, + 299, + 489, + 336 + ], + "type": "text", + "content": " is a basis distribution, a blockiness distribution of images of a fixed quality " + }, + { + "bbox": [ + 130, + 299, + 489, + 336 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 130, + 299, + 489, + 336 + ], + "type": "text", + "content": ". More specifically, the estimated quality is given by" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 209, + 344, + 481, + 374 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 209, + 344, + 481, + 374 + ], + "spans": [ + { + "bbox": [ + 209, + 344, + 481, + 374 + ], + "type": "interline_equation", + "content": "\\hat {q} _ {X} = \\sum_ {q \\in S} q \\frac {\\exp \\left(- D _ {\\mathrm {K L}} \\left(p _ {X , 1 . 0} \\mid \\mid p _ {Z , q}\\right)\\right)}{\\sum_ {q ^ {\\prime} \\in S} \\exp \\left(- D _ {\\mathrm {K L}} \\left(p _ {X , 1 . 0} \\mid \\mid p _ {Z , q ^ {\\prime}}\\right)\\right)}, \\tag {3}", + "image_path": "5f58423b896ed69b66b8677f461c40cd4102e04ae2ac7180fa7a3fa81853d2a8.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 384, + 482, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 384, + 482, + 443 + ], + "spans": [ + { + "bbox": [ + 130, + 384, + 482, + 443 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 384, + 482, + 443 + ], + "type": "inline_equation", + "content": "Z" + }, + { + "bbox": [ + 130, + 384, + 482, + 443 + ], + "type": "text", + "content": " is a small dataset that involves only uncompressed images. In this work, we use DF2K. " + }, + { + "bbox": [ + 130, + 384, + 482, + 443 + ], + "type": "inline_equation", + "content": "D_{\\mathrm{KL}}" + }, + { + "bbox": [ + 130, + 384, + 482, + 443 + ], + "type": "text", + "content": " is the KL divergence, and " + }, + { + "bbox": [ + 130, + 384, + 482, + 443 + ], + "type": "inline_equation", + "content": "S = \\{1.0, 0.95, 0.85, 0.75, 0.5\\}" + }, + { + "bbox": [ + 130, + 384, + 482, + 443 + ], + "type": "text", + "content": " are discretely sampled quality values. Figure 4 shows the blockiness distributions " + }, + { + "bbox": [ + 130, + 384, + 482, + 443 + ], + "type": "inline_equation", + "content": "p_{X,1.0}" + }, + { + "bbox": [ + 130, + 384, + 482, + 443 + ], + "type": "text", + "content": " for the three low-resolution image datasets and the basis distributions " + }, + { + "bbox": [ + 130, + 384, + 482, + 443 + ], + "type": "inline_equation", + "content": "p_{Z,c}" + }, + { + "bbox": [ + 130, + 384, + 482, + 443 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 444, + 482, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 444, + 482, + 479 + ], + "spans": [ + { + "bbox": [ + 130, + 444, + 482, + 479 + ], + "type": "text", + "content": "Selection results. The estimated qualities for ImageNet-1k, Places365, and PASS were " + }, + { + "bbox": [ + 130, + 444, + 482, + 479 + ], + "type": "inline_equation", + "content": "95.5\\%" + }, + { + "bbox": [ + 130, + 444, + 482, + 479 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 444, + 482, + 479 + ], + "type": "inline_equation", + "content": "75.0\\%" + }, + { + "bbox": [ + 130, + 444, + 482, + 479 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 444, + 482, + 479 + ], + "type": "inline_equation", + "content": "99.8\\%" + }, + { + "bbox": [ + 130, + 444, + 482, + 479 + ], + "type": "text", + "content": ", respectively. From this result, Places365 is filtered out." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 499, + 271, + 513 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 499, + 271, + 513 + ], + "spans": [ + { + "bbox": [ + 131, + 499, + 271, + 513 + ], + "type": "text", + "content": "3.2 Object-based filtering" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 521, + 482, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 482, + 581 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 482, + 581 + ], + "type": "text", + "content": "Given a source dataset " + }, + { + "bbox": [ + 130, + 521, + 482, + 581 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 130, + 521, + 482, + 581 + ], + "type": "text", + "content": ", this step applies filtering to refine it as a dataset for training SR models, under the assumption that images with diverse object regions are more effective than those with uniform or monotonous content. Specifically, the refined training dataset is given by " + }, + { + "bbox": [ + 130, + 521, + 482, + 581 + ], + "type": "inline_equation", + "content": "\\tilde{X} = \\{x \\in X : R(x) \\geq \\theta\\}" + }, + { + "bbox": [ + 130, + 521, + 482, + 581 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 521, + 482, + 581 + ], + "type": "inline_equation", + "content": "R(x)" + }, + { + "bbox": [ + 130, + 521, + 482, + 581 + ], + "type": "text", + "content": " is the number of object regions and " + }, + { + "bbox": [ + 130, + 521, + 482, + 581 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 130, + 521, + 482, + 581 + ], + "type": "text", + "content": " is a threshold." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 581, + 482, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 581, + 482, + 629 + ], + "spans": [ + { + "bbox": [ + 130, + 581, + 482, + 629 + ], + "type": "text", + "content": "We introduce two object-based filtering methods with different granularities to explore how the granularity of object detection, ranging from detailed identification of small objects or features to recognizing larger, more general objects, impacts SR performance." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "text", + "content": "Segmentation-based filtering. This method counts the number of object parts by an image segmentation model. Specifically, we adopt the SAM [13] with the ViT-H backbone and define " + }, + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "text", + "content": " by the number of segmentation masks." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 167, + 91, + 448, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 167, + 91, + 448, + 103 + ], + "spans": [ + { + "bbox": [ + 167, + 91, + 448, + 103 + ], + "type": "text", + "content": "Rethinking Image Super-Resolution from Training Data Perspectives" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 135, + 168, + 479, + 296 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 482, + 159 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 482, + 159 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 482, + 159 + ], + "type": "text", + "content": "Table 1: Dataset statistics. HR: High resolution, LR: Low resolution, #Images: Number of images, #Pixels: Average number of pixels per image, Blockiness: median of blockiness measure indicating the intensity of JPEG compression noise, #Segments: Average number of segmentation masks." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 135, + 168, + 479, + 296 + ], + "lines": [ + { + "bbox": [ + 135, + 168, + 479, + 296 + ], + "spans": [ + { + "bbox": [ + 135, + 168, + 479, + 296 + ], + "type": "table", + "html": "
DatasetTaskHR LR#Images#PixelsBlockiness#Segments
DIV2K [1]Super-resolution✓8002.8M0.47104
DF2K [1,27]Super-resolution✓3,4502.8M0.47103
LSDIR [17]Super-resolution✓84,9911.1M0.8292
ImageNet [8]Image recognition1,281,167237k4.3971
Places365 [39]Image recognition1,803,460366k80.71100
PASS [2]Image recognition1,439,589178k3.0374
DiverSeg-I (Ours)Super-resolution259,448233k2.83146
DiverSeg-P (Ours)Super-resolution267,055179k4.39146
DiverSeg-IP (Ours)Super-resolution526,503206k3.61146
", + "image_path": "156a6e7c22da197a23ac7adfeffe1e262a93ef2f3509fa50bc5c6cf226e0d6d2.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 321, + 479, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 321, + 479, + 380 + ], + "spans": [ + { + "bbox": [ + 130, + 321, + 479, + 380 + ], + "type": "text", + "content": "We chose SAM because, unlike typical segmentation models that perform semantic segmentation based on class labels, SAM provides segments that do not impose semantic constraints, allowing for finer region segmentation of diverse objects. We set " + }, + { + "bbox": [ + 130, + 321, + 479, + 380 + ], + "type": "inline_equation", + "content": "\\theta = 100" + }, + { + "bbox": [ + 130, + 321, + 479, + 380 + ], + "type": "text", + "content": " as the default, resulting in " + }, + { + "bbox": [ + 130, + 321, + 479, + 380 + ], + "type": "inline_equation", + "content": "260\\mathrm{k}" + }, + { + "bbox": [ + 130, + 321, + 479, + 380 + ], + "type": "text", + "content": " images remaining after applying the filter to ImageNet-1k." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 382, + 481, + 431 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 382, + 481, + 431 + ], + "spans": [ + { + "bbox": [ + 130, + 382, + 481, + 431 + ], + "type": "text", + "content": "Detection-based filtering. This method counts the number of objects per image. We adopt the Detic model [41] with the ViT-B backbone and define " + }, + { + "bbox": [ + 130, + 382, + 481, + 431 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 130, + 382, + 481, + 431 + ], + "type": "text", + "content": " by the number of detected objects. We set " + }, + { + "bbox": [ + 130, + 382, + 481, + 431 + ], + "type": "inline_equation", + "content": "\\theta = 18" + }, + { + "bbox": [ + 130, + 382, + 481, + 431 + ], + "type": "text", + "content": ", resulting in 260k images remaining after applying the filter to ImageNet-1k." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 453, + 248, + 465 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 453, + 248, + 465 + ], + "spans": [ + { + "bbox": [ + 132, + 453, + 248, + 465 + ], + "type": "text", + "content": "3.3 Dataset statistics" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 479, + 482, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 479, + 482, + 564 + ], + "spans": [ + { + "bbox": [ + 130, + 479, + 482, + 564 + ], + "type": "text", + "content": "Table 1 shows dataset statistics. We created three variants of DiverSeg, namely DiverSeg-I, DiverSeg-P, and DiverSeg-IP, constructed from ImageNet-1k, PASS, and the union of the two, respectively. The segmentation-based filtering is applied to obtain these datasets. Compared to high-resolution datasets such as DF2K and LSDIR, our datasets have larger number of training images, but contain only low-resolution images. The median of blockiness values is decreased for ImageNet-1k and is increased for PASS after filtering." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 586, + 230, + 601 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 586, + 230, + 601 + ], + "spans": [ + { + "bbox": [ + 132, + 586, + 230, + 601 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "type": "text", + "content": "This section conducts experiments by training various SR models on DiverSeg datasets. We demonstrate that SR models can be trained without using high-resolution images and thoroughly analyze factors from a dataset perspective that are crucial for enhancing the performance of SR." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 233, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 233, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 233, + 102 + ], + "type": "text", + "content": "G. Ohtani et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 134, + 158, + 479, + 365 + ], + "blocks": [ + { + "bbox": [ + 131, + 114, + 479, + 148 + ], + "lines": [ + { + "bbox": [ + 131, + 114, + 479, + 148 + ], + "spans": [ + { + "bbox": [ + 131, + 114, + 479, + 148 + ], + "type": "text", + "content": "Table 2: Comparison of DiverSeg with high-resolution image datasets (DF2K and LSDIR). Models trained on DiverSeg-I and DiverSeg-P demonstrated superior performance despite not using any high-resolution images for training." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 134, + 158, + 479, + 365 + ], + "lines": [ + { + "bbox": [ + 134, + 158, + 479, + 365 + ], + "spans": [ + { + "bbox": [ + 134, + 158, + 479, + 365 + ], + "type": "table", + "html": "
Model (Params)DatasetHR LRSet5Set14BSD100Urban100Manga109
PSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIM
MSRResNet [29] (1.5M)DF2K✓32.230.895528.670.783127.620.737426.230.789730.640.9108
LSDIR✓32.150.894828.660.783627.620.737426.310.791830.570.9105
DiverSeg-I✓32.270.896328.640.783727.640.737826.310.791830.530.9115
DiverSeg-P✓32.090.894328.610.783227.600.737126.280.791830.360.9101
RCAN [36] (15.5M)DF2K✓32.500.899028.870.788527.750.742126.730.805831.170.9165
LSDIR✓32.530.899228.890.789427.750.742526.910.809031.330.9180
DiverSeg-I✓32.700.901228.980.790827.810.744327.030.811631.580.9210
DiverSeg-P✓32.630.900028.950.789827.770.743526.990.813431.190.9190
EDSR [20] (43.0M)DF2K✓32.610.899828.910.789327.790.743426.840.808931.380.9176
LSDIR✓32.570.899228.970.790827.800.743827.050.813131.470.9192
DiverSeg-I✓32.710.901728.980.791327.850.745327.100.814231.720.9216
DiverSeg-P✓32.570.900229.060.791527.800.744727.100.816331.330.9191
SwinIR [19] (11.9M)DF2K✓32.920.904429.090.795027.920.748927.450.825432.030.9260
LSDIR✓32.860.903629.160.796327.920.749227.790.833131.980.9262
DiverSeg-I✓32.970.905329.230.797027.980.750827.830.833632.340.9283
DiverSeg-P✓32.850.904029.240.796127.960.750227.850.834932.280.9278
HAT [6] (20.7M)DF2K✓33.030.905629.160.796427.990.751427.930.836532.440.9292
LSDIR✓32.930.905329.290.798828.010.752528.450.846932.570.9306
DiverSeg-I✓33.150.907129.460.800428.070.754228.510.847732.900.9325
DiverSeg-P✓33.120.906829.500.800228.040.753628.530.849232.830.9320
", + "image_path": "f1af37c10faa956dc2e7194a6accb2ee32c04c70746bdf3050afbbef0306f9c8.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 386, + 271, + 399 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 386, + 271, + 399 + ], + "spans": [ + { + "bbox": [ + 132, + 386, + 271, + 399 + ], + "type": "text", + "content": "4.1 Experimental settings" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 403, + 479, + 440 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 403, + 479, + 440 + ], + "spans": [ + { + "bbox": [ + 130, + 403, + 479, + 440 + ], + "type": "text", + "content": "SR models. We use five models. Specifically, we use three CNN-based models: MSRResNet [29], EDSR [20], RCAN [36], and two Transformer-based models: SwinIR [19] and HAT [6]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 440, + 479, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 440, + 479, + 475 + ], + "spans": [ + { + "bbox": [ + 130, + 440, + 479, + 475 + ], + "type": "text", + "content": "Training datasets. We compare the DiverSeg datasets with two high-resolution datasets: DF2K and LSDIR [17]. DF2K is a merged dataset of DIV2K [1] and Flickr2K [27]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 475, + 479, + 500 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 475, + 479, + 500 + ], + "spans": [ + { + "bbox": [ + 130, + 475, + 479, + 500 + ], + "type": "text", + "content": "Evaluation datasets. We use five benchmark datasets: Set5 [3], Set14 [32], BSD100 [22], Urban100 [11], and Manga109 [23]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 500, + 479, + 524 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 500, + 479, + 524 + ], + "spans": [ + { + "bbox": [ + 130, + 500, + 479, + 524 + ], + "type": "text", + "content": "Evaluation metrics. PSNR and SSIM on the Y channel (representing luminance) within the transformed YCbCr color space are used as evaluation metrics." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 524, + 479, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 524, + 479, + 559 + ], + "spans": [ + { + "bbox": [ + 130, + 524, + 479, + 559 + ], + "type": "text", + "content": "Implementation settings. We follow the training setup of the original papers of the SR models [6, 19, 20, 29, 36]. Implementation details are provided in the supplementary materials." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 575, + 265, + 588 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 575, + 265, + 588 + ], + "spans": [ + { + "bbox": [ + 132, + 575, + 265, + 588 + ], + "type": "text", + "content": "4.2 Experimental results" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 594, + 480, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 480, + 665 + ], + "type": "text", + "content": "Main results. To validate the effectiveness of our approach, we trained the five SR models on DiverSeg-I/P. The results are summarized in Table 2. As shown, models trained on DiverSeg-I/P achieved better performance than those trained on DF2K and LSDIR. This shows the effectiveness of the proposed datasets. To the best of our knowledge, we are the first to successfully train SR models without using high-resolution images." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 167, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 167, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 167, + 91, + 447, + 102 + ], + "type": "text", + "content": "Rethinking Image Super-Resolution from Training Data Perspectives" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 134, + 160, + 479, + 338 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 482, + 149 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 482, + 149 + ], + "type": "text", + "content": "Table 3: Quantitative comparison with state-of-the-art methods on five benchmark datasets. We applied our dataset to two Transformer-based models. Checkmarks for HR, LR indicate the use of high-resolution and low-resolution datasets, respectively." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 134, + 160, + 479, + 338 + ], + "lines": [ + { + "bbox": [ + 134, + 160, + 479, + 338 + ], + "spans": [ + { + "bbox": [ + 134, + 160, + 479, + 338 + ], + "type": "table", + "html": "
MethodTraining DataHR LRSet5Set14BSD100Urban100Manga109
PSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIM
SAN [7]DIV2K✓32.640.900328.920.788827.780.743626.790.806831.180.9169
IGNN [40]DIV2K✓32.570.899828.850.789127.770.743426.840.809031.280.9182
HAN [25]DIV2K✓32.640.900228.900.789027.800.744226.850.809431.420.9177
NLSN [24]DIV2K✓32.590.900028.870.789127.780.744426.960.810931.270.9184
RRDB [29]DF2K✓32.730.901128.990.791727.850.745527.030.815331.660.9196
RCAN-it [21]DF2K✓32.690.900728.990.792227.870.745927.160.816831.780.9217
EDT [16]DF2K✓32.820.903129.090.793927.910.748327.460.824632.050.9254
HAT-S [6]DF2K✓32.920.904729.150.795827.970.750527.870.834632.350.9283
IPT [5]ImageNet✓32.64-29.01-27.82-27.26---
SwinIR [19]DF2K✓32.920.904429.090.795027.920.748927.450.825432.030.9260
SwinIR [19]DiverSeg-I (Ours)✓32.970.905329.230.797027.980.750827.830.833632.340.9283
HAT [6]DF2K✓33.040.905629.230.797328.000.751727.970.836832.480.9292
HAT [6]ImageNet→DF2K✓33.180.907329.380.800128.050.753428.370.844732.870.9319
HAT [6]DiverSeg-I (Ours)✓33.150.907129.460.800428.060.754228.510.847732.900.9325
HAT [6]DiverSeg-IP (Ours)✓33.140.907329.510.800728.070.754228.540.849232.960.9327
HAT-L [6]ImageNet→DF2K✓33.300.908329.470.801528.090.755128.600.849833.090.9335
HAT-L [6]DiverSeg-I (Ours)✓33.280.908329.540.802228.100.755628.750.852933.140.9340
HAT-L [6]DiverSeg-IP (Ours)✓33.200.908029.590.801928.110.755628.810.854733.190.9342
", + "image_path": "58f422521b66ab6d2656c18defe6bf07f9c80a44692ed832124098cb35211895.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 145, + 376, + 472, + 449 + ], + "blocks": [ + { + "bbox": [ + 132, + 349, + 482, + 371 + ], + "lines": [ + { + "bbox": [ + 132, + 349, + 482, + 371 + ], + "spans": [ + { + "bbox": [ + 132, + 349, + 482, + 371 + ], + "type": "text", + "content": "Table 4: Filtering by blockiness (HAT model, ImageNet-1k). " + }, + { + "bbox": [ + 132, + 349, + 482, + 371 + ], + "type": "inline_equation", + "content": "\\theta^{\\prime}" + }, + { + "bbox": [ + 132, + 349, + 482, + 371 + ], + "type": "text", + "content": ": threshold for blockiness values." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 145, + 376, + 472, + 449 + ], + "lines": [ + { + "bbox": [ + 145, + 376, + 472, + 449 + ], + "spans": [ + { + "bbox": [ + 145, + 376, + 472, + 449 + ], + "type": "table", + "html": "
θ'#ImagesSet5Set14BSD100Urban100Manga109
PSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIM
10800k33.070.906829.340.800028.040.753828.410.846532.560.9310
30939k33.130.907229.390.800128.060.753828.430.846332.700.9316
1001.08M33.110.907129.420.800128.050.753728.410.845732.840.9321
-1.15M33.080.907129.400.799928.050.753528.410.845732.880.9323
", + "image_path": "56e01cf05478fe0d7e760f12c72728e011b1ad3304a304b24b8fc89b9b6acdb1.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 470, + 482, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 470, + 482, + 579 + ], + "spans": [ + { + "bbox": [ + 130, + 470, + 482, + 579 + ], + "type": "text", + "content": "Comparison with SOTA. Table 3 shows that DiverSeg improves the state-of-the-art performance. Specifically, the HAT and HAT-L models trained on DiverSeg datasets outperformed those trained with ImageNet-1k " + }, + { + "bbox": [ + 130, + 470, + 482, + 579 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 130, + 470, + 482, + 579 + ], + "type": "text", + "content": " DF2K, which utilizes all ImageNet-1k images for pre-training and the DF2K images for fine-tuning, in terms of PSNR and SSIM on four of five benchmarking datasets. It is worth noting that DiverSeg-I filters out " + }, + { + "bbox": [ + 130, + 470, + 482, + 579 + ], + "type": "inline_equation", + "content": "77.5\\%" + }, + { + "bbox": [ + 130, + 470, + 482, + 579 + ], + "type": "text", + "content": " of images from ImageNet-1k and thus training on it is more efficient than the approach relying on pre-training and fine-tuning. This confirmed the effectiveness and efficiency of our filtering approach." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 597, + 310, + 609 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 597, + 310, + 609 + ], + "spans": [ + { + "bbox": [ + 132, + 597, + 310, + 609 + ], + "type": "text", + "content": "4.3 Analysis 1: Effects of filtering" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "Filtering by blockiness measure. If the blockiness is an important factor for enchaining SR performance, filtering images by the blockiness values, i.e., constructing a training dataset by " + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\hat{X} = \\{x\\in X:B(x)\\leq \\theta '\\}" + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": ", would be a straightforward approach. However, as shown in Table 4, this filtering did not" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 233, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 233, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 233, + 101 + ], + "type": "text", + "content": "G. Ohtani et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 139, + 152, + 315, + 213 + ], + "blocks": [ + { + "bbox": [ + 132, + 115, + 321, + 148 + ], + "lines": [ + { + "bbox": [ + 132, + 115, + 321, + 148 + ], + "spans": [ + { + "bbox": [ + 132, + 115, + 321, + 148 + ], + "type": "text", + "content": "Table 5: Comparison of filtering methods. Performance is evaluated using " + }, + { + "bbox": [ + 132, + 115, + 321, + 148 + ], + "type": "inline_equation", + "content": "260\\mathrm{k}" + }, + { + "bbox": [ + 132, + 115, + 321, + 148 + ], + "type": "text", + "content": " filtered images (HAT model, ImageNet-1k)." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 139, + 152, + 315, + 213 + ], + "lines": [ + { + "bbox": [ + 139, + 152, + 315, + 213 + ], + "spans": [ + { + "bbox": [ + 139, + 152, + 315, + 213 + ], + "type": "table", + "html": "
Filtering methodUrban100Manga109
PSNRSSIMPSNRSSIM
Blockiness28.390.846732.470.9304
Detection-based28.440.846232.870.9322
Seg.-based28.510.847732.900.9325
", + "image_path": "89d07a02fbd731f9f7f41c738f688d3539f0c08a5f84d0d2c5cf0700c7bb93cb.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 342, + 152, + 482, + 213 + ], + "blocks": [ + { + "bbox": [ + 339, + 114, + 480, + 148 + ], + "lines": [ + { + "bbox": [ + 339, + 114, + 480, + 148 + ], + "spans": [ + { + "bbox": [ + 339, + 114, + 480, + 148 + ], + "type": "text", + "content": "Table 6: Performance comparison across different thresholds " + }, + { + "bbox": [ + 339, + 114, + 480, + 148 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 339, + 114, + 480, + 148 + ], + "type": "text", + "content": " (HAT model, ImageNet-1k)." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 342, + 152, + 482, + 213 + ], + "lines": [ + { + "bbox": [ + 342, + 152, + 482, + 213 + ], + "spans": [ + { + "bbox": [ + 342, + 152, + 482, + 213 + ], + "type": "table", + "html": "
θ#ImagesUrban100Manga109
PSNRSSIMPSNRSSIM
01.2M28.410.845732.880.9323
50663k28.460.847232.900.9325
100259k28.510.847732.900.9325
15086k28.360.845232.830.9320
", + "image_path": "7c456356d0e101aa2fcc8910a0c5ce7e49d098e1a21fd9df7511b70163ff2f3f.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 135, + 240, + 479, + 357 + ], + "blocks": [ + { + "bbox": [ + 184, + 224, + 427, + 236 + ], + "lines": [ + { + "bbox": [ + 184, + 224, + 427, + 236 + ], + "spans": [ + { + "bbox": [ + 184, + 224, + 427, + 236 + ], + "type": "text", + "content": "Table 7: Analysis of effects of JPEG quality (HAT model)." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 135, + 240, + 479, + 357 + ], + "lines": [ + { + "bbox": [ + 135, + 240, + 479, + 357 + ], + "spans": [ + { + "bbox": [ + 135, + 240, + 479, + 357 + ], + "type": "table", + "html": "
DatasetQuality (%)BlockinessSet5Set14BSD100Urban100Manga109
PSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIM
DF2K50165.3431.130.885527.540.753326.320.710325.040.751130.200.9024
7583.2032.630.900827.700.745727.700.745727.390.823831.880.9238
8546.9832.940.904329.070.794127.930.749927.720.831332.140.9266
9510.3332.980.904829.110.795127.990.752127.810.833832.210.9278
HR0.4733.030.905629.160.796427.990.751427.930.836532.440.9292
LSDIR50146.4328.500.848726.660.742124.720.667424.910.765729.390.8956
7557.1431.870.885328.530.780527.490.741727.300.823131.830.9223
8524.1432.840.904429.170.795527.980.751328.220.841232.380.9291
953.7132.970.904329.250.797928.050.753928.370.845232.420.9299
HR0.8232.930.905329.290.798828.010.752528.450.846932.570.9306
", + "image_path": "8af3e0b5ee0687d5f4f1323378c3cf0fe395c5539eecff6991286b05af871b47.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 380, + 480, + 417 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 380, + 480, + 417 + ], + "spans": [ + { + "bbox": [ + 130, + 380, + 480, + 417 + ], + "type": "text", + "content": "improve PSNR and SSIM when reducing the threshold " + }, + { + "bbox": [ + 130, + 380, + 480, + 417 + ], + "type": "inline_equation", + "content": "\\theta^{\\prime}" + }, + { + "bbox": [ + 130, + 380, + 480, + 417 + ], + "type": "text", + "content": " from 30 to 10 with an exception of SSIM on Urban100. This is because this method filters out images without considering the diversity of object regions." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 418, + 480, + 489 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 418, + 480, + 489 + ], + "spans": [ + { + "bbox": [ + 130, + 418, + 480, + 489 + ], + "type": "text", + "content": "Object-based filtering. Table 5 compares object-based filtering methods described in Sec. 3.2, where ImageNet-1k is used as a source dataset. For a fair comparison, all dataset size after filtering are the same (260k images). As shown, the segmentation-based filtering performed the best. This indicates that images with diverse object regions are effective for SR training, and finer granularity leads to better performance." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 490, + 480, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 490, + 480, + 525 + ], + "spans": [ + { + "bbox": [ + 130, + 490, + 480, + 525 + ], + "type": "text", + "content": "Filtering threshold. Table 6 shows the results obtained by varying the threshold " + }, + { + "bbox": [ + 130, + 490, + 480, + 525 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 130, + 490, + 480, + 525 + ], + "type": "text", + "content": ", which indicates the minimum number of segments. As shown, " + }, + { + "bbox": [ + 130, + 490, + 480, + 525 + ], + "type": "inline_equation", + "content": "\\theta = 100" + }, + { + "bbox": [ + 130, + 490, + 480, + 525 + ], + "type": "text", + "content": " performed the best." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 535, + 337, + 548 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 535, + 337, + 548 + ], + "spans": [ + { + "bbox": [ + 132, + 535, + 337, + 548 + ], + "type": "text", + "content": "4.4 Analysis 2: Effects of image quality" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 557, + 480, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 557, + 480, + 592 + ], + "spans": [ + { + "bbox": [ + 130, + 557, + 480, + 592 + ], + "type": "text", + "content": "In Sec. 3, we made an assumption that low-quality images are detrimental when training SR models. Here, we empirically justify this assumption and evaluate the impact of image quality on SR performance." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 594, + 480, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 480, + 640 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 480, + 640 + ], + "type": "text", + "content": "Training with compressed images. This experiment applies JPEG compression to the two high-resolution datasets, DF2K and LSDIR, and trains a HAT model on each compressed dataset. As shown in Table 7, the performance decreases as the quality decreases on both datasets." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 641, + 480, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 480, + 665 + ], + "type": "text", + "content": "Learning process. To further analyze why and how low quality images negatively affect training of SR models, Figure 5 compares learning processes. As can" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 167, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 167, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 167, + 91, + 447, + 102 + ], + "type": "text", + "content": "Rethinking Image Super-Resolution from Training Data Perspectives" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 114, + 304, + 234 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 304, + 234 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 304, + 234 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 304, + 234 + ], + "type": "image", + "image_path": "4fe7e02fcbb502ad20d83e2666ecc80b5bb0b42d3ad2219e74bc6e9ad45f3854.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 227, + 236, + 396, + 244 + ], + "lines": [ + { + "bbox": [ + 227, + 236, + 396, + 244 + ], + "spans": [ + { + "bbox": [ + 227, + 236, + 396, + 244 + ], + "type": "text", + "content": "Quality: HR 95% 85% 75% 50%" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 132, + 247, + 479, + 258 + ], + "lines": [ + { + "bbox": [ + 132, + 247, + 479, + 258 + ], + "spans": [ + { + "bbox": [ + 132, + 247, + 479, + 258 + ], + "type": "text", + "content": "Fig. 5: Comparison of learning processes obtained with various JPEG quality values." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 304, + 114, + 475, + 234 + ], + "blocks": [ + { + "bbox": [ + 304, + 114, + 475, + 234 + ], + "lines": [ + { + "bbox": [ + 304, + 114, + 475, + 234 + ], + "spans": [ + { + "bbox": [ + 304, + 114, + 475, + 234 + ], + "type": "image", + "image_path": "9774f3a1c9fc6228546a0672961223950c0359cc76f84c234b1f0249a12a00a8.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 135, + 269, + 249, + 351 + ], + "blocks": [ + { + "bbox": [ + 135, + 269, + 249, + 351 + ], + "lines": [ + { + "bbox": [ + 135, + 269, + 249, + 351 + ], + "spans": [ + { + "bbox": [ + 135, + 269, + 249, + 351 + ], + "type": "image", + "image_path": "40f63773650fba6630c4347c4d4cf97ee3ae59895185ef5144f749d12928e89c.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 244, + 352, + 373, + 360 + ], + "lines": [ + { + "bbox": [ + 244, + 352, + 373, + 360 + ], + "spans": [ + { + "bbox": [ + 244, + 352, + 373, + 360 + ], + "type": "text", + "content": "ImageNet PASS Places365" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 140, + 361, + 471, + 372 + ], + "lines": [ + { + "bbox": [ + 140, + 361, + 471, + 372 + ], + "spans": [ + { + "bbox": [ + 140, + 361, + 471, + 372 + ], + "type": "text", + "content": "Fig. 6: Comparison of learning processes for ImageNet-1k, PASS, and Places365." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 249, + 270, + 362, + 351 + ], + "blocks": [ + { + "bbox": [ + 249, + 270, + 362, + 351 + ], + "lines": [ + { + "bbox": [ + 249, + 270, + 362, + 351 + ], + "spans": [ + { + "bbox": [ + 249, + 270, + 362, + 351 + ], + "type": "image", + "image_path": "c19f8097c1bcbb0ce6cfc774d4a09d4d923b91af515023267840f606f2228b97.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 362, + 270, + 475, + 351 + ], + "blocks": [ + { + "bbox": [ + 362, + 270, + 475, + 351 + ], + "lines": [ + { + "bbox": [ + 362, + 270, + 475, + 351 + ], + "spans": [ + { + "bbox": [ + 362, + 270, + 475, + 351 + ], + "type": "image", + "image_path": "7c9a6b211ee6ca901015927da0faf4f90ded647abab5082bd635f85931bb2006.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 396, + 482, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 396, + 482, + 468 + ], + "spans": [ + { + "bbox": [ + 130, + 396, + 482, + 468 + ], + "type": "text", + "content": "be seen, PSNR decreases after 100k and 300k iterations with " + }, + { + "bbox": [ + 130, + 396, + 482, + 468 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 130, + 396, + 482, + 468 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 396, + 482, + 468 + ], + "type": "inline_equation", + "content": "75\\%" + }, + { + "bbox": [ + 130, + 396, + 482, + 468 + ], + "type": "text", + "content": " qualities, respectively. With " + }, + { + "bbox": [ + 130, + 396, + 482, + 468 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 130, + 396, + 482, + 468 + ], + "type": "text", + "content": " quality, the final performance at the 800k iteration was worse than that at the 10k iteration. These results indicate that excluding low-quality images is crucial for enhancing SR performance. This is in contrast to other vision tasks such as image recognition, where increasing the number of training images, even with low quality images, often helps improve performance." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 468, + 482, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 468, + 482, + 563 + ], + "spans": [ + { + "bbox": [ + 130, + 468, + 482, + 563 + ], + "type": "text", + "content": "Training with Places365. In the source selection step, the Places365 dataset was excluded because its quality estimated via the blockiness distribution was low. To justify this selection, we compare three datasets in Figure 6. As can be seen, the model trained on Places365 performed worse than those trained on ImageNet-1k and PASS. Similar to the learning processes obtained from low quality JPEG images, the PSNR decreased in the later phase of the learning process with Places365. These results confirmed the effectiveness of source selection before training based on the blockiness distributions." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 572, + 313, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 572, + 313, + 586 + ], + "spans": [ + { + "bbox": [ + 132, + 572, + 313, + 586 + ], + "type": "text", + "content": "4.5 Analysis 3: Visual comparison" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 594, + 482, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 667 + ], + "type": "text", + "content": "Comparison of artifacts. Figure 7 compares the artifacts produced by JPEG compression with those produced by the SR models trained on DF2K using different image quality levels. Images obtained from models trained on " + }, + { + "bbox": [ + 130, + 594, + 482, + 667 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 130, + 594, + 482, + 667 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 594, + 482, + 667 + ], + "type": "inline_equation", + "content": "75\\%" + }, + { + "bbox": [ + 130, + 594, + 482, + 667 + ], + "type": "text", + "content": " quality images show strong stripes or checkerboard patterns of artifacts. These artifacts appear to be more significant than those observed when the original image is compressed using JPEG, suggesting that they are induced by" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 233, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 233, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 233, + 101 + ], + "type": "text", + "content": "G. Ohtani et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 141, + 117, + 239, + 269 + ], + "blocks": [ + { + "bbox": [ + 141, + 117, + 239, + 269 + ], + "lines": [ + { + "bbox": [ + 141, + 117, + 239, + 269 + ], + "spans": [ + { + "bbox": [ + 141, + 117, + 239, + 269 + ], + "type": "image", + "image_path": "0d31d797598d68abc1e249599b9d1ac33565c56a4a8734043d41d62036345499.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 145, + 269, + 234, + 277 + ], + "lines": [ + { + "bbox": [ + 145, + 269, + 234, + 277 + ], + "spans": [ + { + "bbox": [ + 145, + 269, + 234, + 277 + ], + "type": "text", + "content": "AkkeraKanjinchou, Manga109" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 143, + 281, + 235, + 324 + ], + "blocks": [ + { + "bbox": [ + 143, + 281, + 235, + 324 + ], + "lines": [ + { + "bbox": [ + 143, + 281, + 235, + 324 + ], + "spans": [ + { + "bbox": [ + 143, + 281, + 235, + 324 + ], + "type": "image", + "image_path": "0862b931f81a796c6f6b747088f909f80c76b8c9e7e5c9baed3349edaf2b1545.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 335, + 482, + 380 + ], + "lines": [ + { + "bbox": [ + 131, + 335, + 482, + 380 + ], + "spans": [ + { + "bbox": [ + 131, + 335, + 482, + 380 + ], + "type": "text", + "content": "Fig. 7: Comparison of artifacts produced by JPEG compression (blue) and SR models trained on various JPEG quality values (red, HAT model, DF2K). Numbers indicate JPEG image quality. First and third rows: cropped images. Second and fourth rows: differential images." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 246, + 117, + 476, + 323 + ], + "blocks": [ + { + "bbox": [ + 246, + 117, + 476, + 323 + ], + "lines": [ + { + "bbox": [ + 246, + 117, + 476, + 323 + ], + "spans": [ + { + "bbox": [ + 246, + 117, + 476, + 323 + ], + "type": "image", + "image_path": "2d0325fc8fc994d65c7de56e725c520cae67d0ce7bd3c4265894d1a15d8a037e.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 402, + 480, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 402, + 480, + 462 + ], + "spans": [ + { + "bbox": [ + 130, + 402, + 480, + 462 + ], + "type": "text", + "content": "model training. The predisposition to stripes and checkerboard patterns is likely due to the inductive bias inherent in the architecture of the neural network; in particular, the square or rectangular shape of the filters in the convolutional operations could cause this. Improvements to the network architecture to allow training on lower quality images would be interesting as future work." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 462, + 482, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 462, + 482, + 545 + ], + "spans": [ + { + "bbox": [ + 130, + 462, + 482, + 545 + ], + "type": "text", + "content": "Qualitative examples. Figure 8 shows visual comparisons of SR models trained on DF2K, LSDIR, DiverSeg-I, and DiverSeg-P. For the image \"img_011\" in Urban100, we observed that models trained on DF2K are unable to recover the horizontal stripes pattern, while the other three models successfully recovered it. With the three images of Manga109, we observed that models trained on DiverSeg-I/P exhibit noticeable improvement in the character region compared to those trained on DF2K or LSDIR." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 552, + 296, + 563 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 552, + 296, + 563 + ], + "spans": [ + { + "bbox": [ + 132, + 552, + 296, + 563 + ], + "type": "text", + "content": "4.6 Discussion and Limitations" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 665 + ], + "type": "text", + "content": "High-resolution datasets. In this paper, we applied the automated image evaluation pipeline to a large dataset of low-resolution images. We believe our finding contributes to the future construction of training datasets and the development of neural network architectures. Specifically, we demonstrated that SR models can be trained without high-resolution images. However, this does not imply that high-resolution image datasets are worthless. Rather, we believe that increasing the diversity of high-resolution images through our proposed filtering method could further improve SR performance. In Table 8, we examined this on" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 167, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 167, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 167, + 91, + 447, + 102 + ], + "type": "text", + "content": "Rethinking Image Super-Resolution from Training Data Perspectives" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 134, + 112, + 482, + 289 + ], + "blocks": [ + { + "bbox": [ + 134, + 112, + 482, + 289 + ], + "lines": [ + { + "bbox": [ + 134, + 112, + 482, + 289 + ], + "spans": [ + { + "bbox": [ + 134, + 112, + 482, + 289 + ], + "type": "image", + "image_path": "beb5ae6da64ad56843148dee603f6ebba143057f079a8aefbcac6df2b9271b71.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 297, + 482, + 331 + ], + "lines": [ + { + "bbox": [ + 131, + 297, + 482, + 331 + ], + "spans": [ + { + "bbox": [ + 131, + 297, + 482, + 331 + ], + "type": "text", + "content": "Fig. 8: Visual comparison of " + }, + { + "bbox": [ + 131, + 297, + 482, + 331 + ], + "type": "inline_equation", + "content": "\\times 4" + }, + { + "bbox": [ + 131, + 297, + 482, + 331 + ], + "type": "text", + "content": " SR models trained on DF2K, LSDIR, DiverSeg-I, and DiverSeg-P datasets. PSNR/SSIM is calculated for each cropped patch individually to better reflect the differences in performance." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 143, + 357, + 471, + 404 + ], + "blocks": [ + { + "bbox": [ + 143, + 342, + 469, + 354 + ], + "lines": [ + { + "bbox": [ + 143, + 342, + 469, + 354 + ], + "spans": [ + { + "bbox": [ + 143, + 342, + 469, + 354 + ], + "type": "text", + "content": "Table 8: Applying object-based filtering to LSDIR with " + }, + { + "bbox": [ + 143, + 342, + 469, + 354 + ], + "type": "inline_equation", + "content": "\\theta = {100}" + }, + { + "bbox": [ + 143, + 342, + 469, + 354 + ], + "type": "text", + "content": " (HAT model)." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 143, + 357, + 471, + 404 + ], + "lines": [ + { + "bbox": [ + 143, + 357, + 471, + 404 + ], + "spans": [ + { + "bbox": [ + 143, + 357, + 471, + 404 + ], + "type": "table", + "html": "
Dataset#ImagesSet5Set14BSD100Urban100Manga109
PSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIM
LSDIR89,99132.930.905329.290.798828.010.752528.450.846932.570.9306
w/ filtering31,56132.950.905629.330.799128.020.752628.530.848732.600.9308
", + "image_path": "82cbb55eb4833e51cfe5e61e1de05f984f9a28581222ce0d7eda8eb293f9ca8a.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 427, + 482, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 427, + 482, + 475 + ], + "spans": [ + { + "bbox": [ + 130, + 427, + 482, + 475 + ], + "type": "text", + "content": "the LSDIR dataset. Our results show that filtering out images with less than 100 object regions led to increased PSNR and SSIM with the HAT model. For future work, exploring a hybrid approach that utilizes both low- and high-resolution diverse images could be promising." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 475, + 482, + 560 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 475, + 482, + 560 + ], + "spans": [ + { + "bbox": [ + 130, + 475, + 482, + 560 + ], + "type": "text", + "content": "Limitations. In this study, we focused on two perspectives: resolution and diversity. However, when collecting large datasets, there are other important perspectives, such as fairness and copyright. In addition, for benchmarking purposes, we used the five most commonly used SR datasets for a fair comparison with conventional methods. Nevertheless, real-world applications also face the challenge of blind super-resolution, where the degradation process is unknown. Creating datasets addressing this aspect would be an interesting future research direction." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 572, + 220, + 585 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 572, + 220, + 585 + ], + "spans": [ + { + "bbox": [ + 132, + 572, + 220, + 585 + ], + "type": "text", + "content": "5 Conclusion" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 593, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 593, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 593, + 482, + 665 + ], + "type": "text", + "content": "In this work, we investigated the effect of the training data for SR and showed that SR models are trainable even without using high-resolution images by applying the image evaluation pipeline to a set of large low-resolution images. In experiments, we thoroughly analyzed the effect of image quality and diversity to SR performance. We hope that this work will positively influence the future construction of training datasets and lead to better models." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 232, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 232, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 232, + 101 + ], + "type": "text", + "content": "G. Ohtani et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 246, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 246, + 129 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 246, + 129 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 139, + 480, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 139, + 480, + 175 + ], + "spans": [ + { + "bbox": [ + 130, + 139, + 480, + 175 + ], + "type": "text", + "content": "Computational resource of AI Bridging Cloud Infrastructure (ABCI) provided by National Institute of Advanced Industrial Science and Technology (AIST) was used." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 193, + 197, + 205 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 193, + 197, + 205 + ], + "spans": [ + { + "bbox": [ + 132, + 193, + 197, + 205 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 217, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 22, + "blocks": [ + { + "bbox": [ + 138, + 217, + 480, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 217, + 480, + 239 + ], + "spans": [ + { + "bbox": [ + 138, + 217, + 480, + 239 + ], + "type": "text", + "content": "1. Agustsson, E., Timofte, R.: Ntire 2017 challenge on single image super-resolution: Dataset and study. In: CVPRW (2017) 2, 4, 8, 9" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 240, + 480, + 272 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 240, + 480, + 272 + ], + "spans": [ + { + "bbox": [ + 138, + 240, + 480, + 272 + ], + "type": "text", + "content": "2. Asano, Y.M., Rupprecht, C., Zisserman, A., Vedaldi, A.: Pass: An imagenet replacement for self-supervised pretraining without humans. In: NeurIPS Track on Datasets and Benchmarks (2021) 3, 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 274, + 480, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 274, + 480, + 304 + ], + "spans": [ + { + "bbox": [ + 138, + 274, + 480, + 304 + ], + "type": "text", + "content": "3. Bevilacqua, M., Roumy, A., Guillemot, C., Alberi-Morel, M.L.: Low-complexity single-image super-resolution based on nonnegative neighbor embedding. In: BMVC (2012) 9" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 306, + 480, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 306, + 480, + 327 + ], + "spans": [ + { + "bbox": [ + 138, + 306, + 480, + 327 + ], + "type": "text", + "content": "4. Bhardwaj, D., Pankajakshan, V.: A jpeg blocking artifact detector for image forensics. Signal Processing: Image Communication 68, 155-161 (2018) 3, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 328, + 480, + 349 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 328, + 480, + 349 + ], + "spans": [ + { + "bbox": [ + 138, + 328, + 480, + 349 + ], + "type": "text", + "content": "5. Chen, H., Wang, Y., Guo, T., Xu, C., Deng, Y., Liu, Z., Ma, S., Xu, C., Xu, C., Gao, W.: Pre-trained image processing transformer. In: CVPR (2021) 1, 4, 5, 10" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 350, + 480, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 350, + 480, + 370 + ], + "spans": [ + { + "bbox": [ + 138, + 350, + 480, + 370 + ], + "type": "text", + "content": "6. Chen, X., Wang, X., Zhou, J., Qiao, Y., Dong, C.: Activating more pixels in image super-resolution transformer. In: CVPR (2023) 1, 4, 5, 9, 10" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 372, + 480, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 372, + 480, + 392 + ], + "spans": [ + { + "bbox": [ + 138, + 372, + 480, + 392 + ], + "type": "text", + "content": "7. Dai, T., Cai, J., Zhang, Y., Xia, S.T., Zhang, L.: Second-order attention network for single image super-resolution. In: CVPR (2019) 4, 10" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 394, + 480, + 414 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 394, + 480, + 414 + ], + "spans": [ + { + "bbox": [ + 138, + 394, + 480, + 414 + ], + "type": "text", + "content": "8. Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: Imagenet: A large-scale hierarchical image database. In: CVPR (2009) 3, 8" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 415, + 480, + 436 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 415, + 480, + 436 + ], + "spans": [ + { + "bbox": [ + 138, + 415, + 480, + 436 + ], + "type": "text", + "content": "9. Dong, C., Loy, C.C., He, K., Tang, X.: Learning a deep convolutional network for image super-resolution. In: ECCV (2014) 1, 4" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 437, + 480, + 457 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 437, + 480, + 457 + ], + "spans": [ + { + "bbox": [ + 138, + 437, + 480, + 457 + ], + "type": "text", + "content": "0. Dong, C., Loy, C.C., Tang, X.: Accelerating the super-resolution convolutional neural network. In: ECCV (2016) 1, 4" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 459, + 480, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 459, + 480, + 479 + ], + "spans": [ + { + "bbox": [ + 138, + 459, + 480, + 479 + ], + "type": "text", + "content": "1. Huang, J.B., Singh, A., Ahuja, N.: Single image super-resolution from transformed self-exemplars. In: CVPR (2015) 9" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 481, + 480, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 481, + 480, + 502 + ], + "spans": [ + { + "bbox": [ + 138, + 481, + 480, + 502 + ], + "type": "text", + "content": "2. Kim, J., Lee, J.K., Lee, K.M.: Accurate image super-resolution using very deep convolutional networks. In: CVPR (2016) 1, 4" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 503, + 480, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 503, + 480, + 533 + ], + "spans": [ + { + "bbox": [ + 138, + 503, + 480, + 533 + ], + "type": "text", + "content": "3. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., Dollar, P., Girshick, R.: Segment anything. In: ICCV (2023) 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 535, + 480, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 535, + 480, + 567 + ], + "spans": [ + { + "bbox": [ + 138, + 535, + 480, + 567 + ], + "type": "text", + "content": "4. Ledig, C., Theis, L., Huszár, F., Caballero, J., Cunningham, A., Acosta, A., Aitken, A., Tejani, A., Totz, J., Wang, Z., et al.: Photo-realistic single image superresolution using a generative adversarial network. In: CVPR (2017) 1, 4" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 138, + 568, + 480, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 568, + 480, + 599 + ], + "spans": [ + { + "bbox": [ + 138, + 568, + 480, + 599 + ], + "type": "text", + "content": "5. Li, A., Zhang, L., Liu, Y., Zhu, C.: Feature modulation transformer: Crossrefinement of global representation via high-frequency prior for image superresolution. In: CVPR (2023) 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 138, + 601, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 601, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 138, + 601, + 480, + 632 + ], + "type": "text", + "content": "6. Li, W., Lu, X., Qian, S., Lu, J., Zhang, X., Jia, J.: On efficient transformer-based image pre-training for low-level vision. arXiv preprint arXiv:2112.10175 (2021) 1, 4, 5, 10" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 138, + 633, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 633, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 633, + 480, + 665 + ], + "type": "text", + "content": "7. Li, Y., Zhang, K., Liang, J., Cao, J., Liu, C., Gong, R., Zhang, Y., Tang, H., Liu, Y., Demandolx, D., Ranjan, R., Timofte, R., Van Gool, L.: Lsdir: A large scale dataset for image restoration. In: CVPRW (2023) 2, 5, 8, 9" + } + ] + } + ], + "index": 21 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 167, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 167, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 167, + 91, + 447, + 102 + ], + "type": "text", + "content": "Rethinking Image Super-Resolution from Training Data Perspectives" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 22, + "blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "type": "text", + "content": "18. Li, Y., Zhang, Y., Timofte, R., Van Gool, L., Yu, L., Li, Y., Li, X., Jiang, T., Wu, Q., Han, M., et al.: Ntire 2023 challenge on efficient super-resolution: Methods and results. In: CVPRW (2023) 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 150, + 481, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 150, + 481, + 171 + ], + "spans": [ + { + "bbox": [ + 132, + 150, + 481, + 171 + ], + "type": "text", + "content": "19. Liang, J., Cao, J., Sun, G., Zhang, K., Van Gool, L., Timofte, R.: Swinir: Image restoration using swin transformer. In: ICCVW (2021) 1, 2, 4, 9, 10" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 172, + 481, + 193 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 172, + 481, + 193 + ], + "spans": [ + { + "bbox": [ + 132, + 172, + 481, + 193 + ], + "type": "text", + "content": "20. Lim, B., Son, S., Kim, H., Nah, S., Mu Lee, K.: Enhanced deep residual networks for single image super-resolution. In: CVPRW (2017) 1, 4, 9" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 194, + 481, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 194, + 481, + 226 + ], + "spans": [ + { + "bbox": [ + 132, + 194, + 481, + 226 + ], + "type": "text", + "content": "21. Lin, Z., Garg, P., Banerjee, A., Magid, S.A., Sun, D., Zhang, Y., Van Gool, L., Wei, D., Pfister, H.: Revisiting rcan: Improved training for image super-resolution. arXiv preprint arXiv:2201.11279 (2022) 10" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 227, + 481, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 227, + 481, + 258 + ], + "spans": [ + { + "bbox": [ + 132, + 227, + 481, + 258 + ], + "type": "text", + "content": "22. Martin, D., Fowlkes, C., Tal, D., Malik, J.: A database of human segmented natural images and its application to evaluating segmentation algorithms and measuring ecological statistics. In: ICCV (2001) 4, 9" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 259, + 481, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 259, + 481, + 293 + ], + "spans": [ + { + "bbox": [ + 132, + 259, + 481, + 293 + ], + "type": "text", + "content": "23. Matsui, Y., Ito, K., Aramaki, Y., Fujimoto, A., Ogawa, T., Yamasaki, T., Aizawa, K.: Sketch-based manga retrieval using manga109 dataset. Multimedia Tools and Applications 76, 21811-21838 (2017) 9" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 293, + 481, + 313 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 293, + 481, + 313 + ], + "spans": [ + { + "bbox": [ + 132, + 293, + 481, + 313 + ], + "type": "text", + "content": "24. Mei, Y., Fan, Y., Zhou, Y.: Image super-resolution with non-local sparse attention. In: CVPR (2021) 10" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 315, + 481, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 315, + 481, + 346 + ], + "spans": [ + { + "bbox": [ + 132, + 315, + 481, + 346 + ], + "type": "text", + "content": "25. Niu, B., Wen, W., Ren, W., Zhang, X., Yang, L., Wang, S., Zhang, K., Cao, X., Shen, H.: Single image super-resolution via a holistic attention network. In: ECCV (2020) 10" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 347, + 481, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 347, + 481, + 380 + ], + "spans": [ + { + "bbox": [ + 132, + 347, + 481, + 380 + ], + "type": "text", + "content": "26. Shi, W., Caballero, J., Huszar, F., Totz, J., Aitken, A.P., Bishop, R., Rueckert, D., Wang, Z.: Real-time single image and video super-resolution using an efficient sub-pixel convolutional neural network. In: CVPR (2016) 1, 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 380, + 481, + 412 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 380, + 481, + 412 + ], + "spans": [ + { + "bbox": [ + 132, + 380, + 481, + 412 + ], + "type": "text", + "content": "27. Timofte, R., Agustsson, E., Van Gool, L., Yang, M.H., Zhang, L.: Ntire 2017 challenge on single image super-resolution: Methods and results. In: CVPRW (2017) 2, 4, 8, 9" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 413, + 481, + 434 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 413, + 481, + 434 + ], + "spans": [ + { + "bbox": [ + 132, + 413, + 481, + 434 + ], + "type": "text", + "content": "28. Tong, T., Li, G., Liu, X., Gao, Q.: Image super-resolution using dense skip connections. In: ICCV (2017) 1, 4" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 434, + 481, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 434, + 481, + 468 + ], + "spans": [ + { + "bbox": [ + 132, + 434, + 481, + 468 + ], + "type": "text", + "content": "29. Wang, X., Yu, K., Wu, S., Gu, J., Liu, Y., Dong, C., Qiao, Y., Change Loy, C.: Esrgan: Enhanced super-resolution generative adversarial networks. In: ECCVW (2018) 1, 2, 4, 9, 10" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 468, + 481, + 490 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 468, + 481, + 490 + ], + "spans": [ + { + "bbox": [ + 132, + 468, + 481, + 490 + ], + "type": "text", + "content": "30. Yang, J., Wright, J., Huang, T., Ma, Y.: Image super-resolution as sparse representation of raw image patches. In: CVPR (2008) 4" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 491, + 481, + 522 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 491, + 481, + 522 + ], + "spans": [ + { + "bbox": [ + 132, + 491, + 481, + 522 + ], + "type": "text", + "content": "31. Yang, Q., Chen, D., Tan, Z., Liu, Q., Chu, Q., Bao, J., Yuan, L., Hua, G., Yu, N.: Hq-50k: A large-scale, high-quality dataset for image restoration. arXiv preprint arXiv:2306.05390 (2023) 2, 5" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 522, + 481, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 522, + 481, + 544 + ], + "spans": [ + { + "bbox": [ + 132, + 522, + 481, + 544 + ], + "type": "text", + "content": "32. Zeyde, R., Elad, M., Protter, M.: On single image scale-up using sparse-representations. In: Proc. 7th Int. Conf. Curves Surf. (2010) 9" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 545, + 481, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 545, + 481, + 578 + ], + "spans": [ + { + "bbox": [ + 132, + 545, + 481, + 578 + ], + "type": "text", + "content": "33. Zhang, D., Huang, F., Liu, S., Wang, X., Jin, Z.: Swinfir: Revisiting the swinir with fast fourier convolution and improved training for image super-resolution. arXiv preprint arXiv:2208.11247 (2022) 1, 4" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 578, + 481, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 578, + 481, + 599 + ], + "spans": [ + { + "bbox": [ + 132, + 578, + 481, + 599 + ], + "type": "text", + "content": "34. Zhang, X., Zeng, H., Guo, S., Zhang, L.: Efficient long-range attention network for image super-resolution. In: ECCV (2022) 1, 4" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 600, + 481, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 600, + 481, + 621 + ], + "spans": [ + { + "bbox": [ + 132, + 600, + 481, + 621 + ], + "type": "text", + "content": "35. Zhang, Y., Li, K., Zhong, B., Fu, Y.: Residual non-local attention networks for image restoration. In: ICLR (2019) 1, 4" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 132, + 621, + 481, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 621, + 481, + 643 + ], + "spans": [ + { + "bbox": [ + 132, + 621, + 481, + 643 + ], + "type": "text", + "content": "36. Zhang, Y., Li, K., Li, K., Wang, L., Zhong, B., Fu, Y.: Image super-resolution using very deep residual channel attention networks. In: ECCV (2018) 1, 4, 9" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 132, + 644, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 644, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 644, + 481, + 665 + ], + "type": "text", + "content": "37. Zhang, Y., Tian, Y., Kong, Y., Zhong, B., Fu, Y.: Residual dense network for image super-resolution. In: CVPR (2018) 1, 4" + } + ] + } + ], + "index": 21 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 233, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 233, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 233, + 101 + ], + "type": "text", + "content": "G. Ohtani et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 259 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "text", + "content": "38. Zhang, Y., Zhang, K., Chen, Z., Li, Y., Timofte, R., Zhang, J., Zhang, K., Peng, R., Ma, Y., Jia, L., et al.: Ntire 2023 challenge on image super-resolution (x4): Methods and results. In: CVPRW (2023) 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 150, + 482, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 150, + 482, + 182 + ], + "spans": [ + { + "bbox": [ + 130, + 150, + 482, + 182 + ], + "type": "text", + "content": "39. Zhou, B., Lapedriza, A., Xiao, J., Torralba, A., Oliva, A.: Learning deep features for scene recognition using places database. Advances in neural information processing systems 27 (2014) 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 183, + 482, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 183, + 482, + 215 + ], + "spans": [ + { + "bbox": [ + 130, + 183, + 482, + 215 + ], + "type": "text", + "content": "40. Zhou, S., Zhang, J., Zuo, W., Loy, C.C.: Cross-scale internal graph neural network for image super-resolution. Advances in neural information processing systems 33, 3499-3509 (2020) 10" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 216, + 482, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 216, + 482, + 237 + ], + "spans": [ + { + "bbox": [ + 130, + 216, + 482, + 237 + ], + "type": "text", + "content": "41. Zhou, X., Girdhar, R., Joulin, A., Krahenbuhl, P., Misra, I.: Detecting twenty-thousand classes using image-level supervision. In: ECCV (2022) 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 237, + 482, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 237, + 482, + 259 + ], + "spans": [ + { + "bbox": [ + 130, + 237, + 482, + 259 + ], + "type": "text", + "content": "42. Zhou, Y., Li, Z., Guo, C.L., Bai, S., Cheng, M.M., Hou, Q.: Srformer: Permuted self-attention for single image super-resolution. In: ICCV (2023) 1, 2, 4" + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 167, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 167, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 167, + 91, + 447, + 102 + ], + "type": "text", + "content": "Rethinking Image Super-Resolution from Training Data Perspectives" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Rethinking Image-to-Video Adaptation_ An Object-centric Perspective/17973f40-dece-4cfa-8ceb-59145fcea296_content_list.json b/2024/Rethinking Image-to-Video Adaptation_ An Object-centric Perspective/17973f40-dece-4cfa-8ceb-59145fcea296_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..cba7dfac61d8b785304461ffe0eaf12b6e74340a --- /dev/null +++ b/2024/Rethinking Image-to-Video Adaptation_ An Object-centric Perspective/17973f40-dece-4cfa-8ceb-59145fcea296_content_list.json @@ -0,0 +1,1704 @@ +[ + { + "type": "text", + "text": "Rethinking Image-to-Video Adaptation: An Object-centric Perspective", + "text_level": 1, + "bbox": [ + 274, + 140, + 728, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Rui Qian1, Shuangrui Ding1, and Dahua Lin1,2,3*", + "bbox": [ + 303, + 212, + 697, + 228 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 The Chinese University of Hong Kong, Hong Kong, China \n2 Shanghai Artificial Intelligence Laboratory, Shanghai, China \n3 Centre for Perceptual and Interactive Intelligence (CPII), Hong Kong, China {qr021, ds023, dhlin}@ie.cuhk.edu.hk", + "bbox": [ + 236, + 239, + 764, + 295 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Image-to-video adaptation seeks to efficiently adapt image models for use in the video domain. Instead of finetuning the entire image backbone, many image-to-video adaptation paradigms use lightweight adapters for temporal modeling on top of the spatial module. However, these attempts are subject to limitations in efficiency and interpretability. In this paper, we propose a novel and efficient image-to-video adaptation strategy from the object-centric perspective. Inspired by human perception, which identifies objects as key components for video understanding, we integrate a proxy task of object discovery into image-to-video transfer learning. Specifically, we adopt slot attention with learnable queries to distill each frame into a compact set of object tokens. These object-centric tokens are then processed through object-time interaction layers to model object state changes across time. Integrated with two novel object-level losses, we demonstrate the feasibility of performing efficient temporal reasoning solely on the compressed object-centric representations for video downstream tasks. Our method achieves state-of-the-art performance with fewer tunable parameters, only $5\\%$ of fully finetuned models and $50\\%$ of efficient tuning methods, on action recognition benchmarks. In addition, our model performs favorably in zero-shot video object segmentation without further retraining or object annotations, proving the effectiveness of object-centric video understanding.", + "bbox": [ + 261, + 339, + 743, + 630 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Image-to-video Adaptation $\\cdot$ Object-centric Learning", + "bbox": [ + 261, + 643, + 702, + 659 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 686, + 375, + 704 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recently, large-scale image pre-training has revolutionized the computer vision community, with the large foundation models dominating various image tasks [34, 56, 69, 78, 87, 94]. The remarkable success of these models can be attributed to the enormous amounts of image data [28, 56] or image-text pairs [34, 69, 78]. However, when we extend this discussion to the video domain, it becomes non-trivial. Acquiring large video datasets and managing high computational", + "bbox": [ + 212, + 723, + 787, + 815 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Corresponding author. Email: dhlin@ie.cuhk.edu.hk", + "bbox": [ + 217, + 824, + 580, + 839 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/ba3a23d8e4d481b67ba4e7d07107a6116f73242819352de7a69a14a14335a962.jpg", + "image_caption": [ + "Fig. 1: Illustration of our proposed object-centric video understanding pipeline for image-to-video adaptation. Specifically, we first parse each frame into several object components (represented in different colors) to form an interaction space. Then we respectively establish inter-object interactions within each frame (represented by dotted arrows), and temporal state changes of individual objects (depicted by solid arrows)." + ], + "image_footnote": [], + "bbox": [ + 232, + 142, + 772, + 311 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "costs make training video foundation models from scratch a significant challenge. To circumvent these difficulties, the community has developed an innovative concept dubbed image-to-video adaptation. This approach aims to leverage the power of pre-trained image models and apply them to video data, enhancing both performance and efficiency. The image-to-video adaptation establishes a connection between effective methodologies in image models and the realm of video understanding, thereby opening new avenues for research and exploration.", + "bbox": [ + 212, + 426, + 787, + 532 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Motivated by parameter-efficient transfer learning in natural language processing [31, 32, 41, 89], a series of recent studies [36, 46, 47, 55, 57, 86] have adopted an efficient tuning approach for image-to-video adaptation. These approaches maintain the image pre-trained model in a frozen state and only finetune a minimal set of additional parameters. For example, [47, 55, 57, 86] insert learnable spatio-temporal modules into Transformer encoder blocks. However, this formulation encounters two vital limitations. Firstly, tuning the in-between parameters in the shallow layers still requires a considerable amount of GPU memory during gradient backpropagation. Secondly, the introduction of basic temporal modules, like attention layers, makes temporal reasoning on highly redundant features and lacks the essential inductive bias to effectively benefit the image-to-video adaptation, thereby presenting challenges in interpreting the training's success.", + "bbox": [ + 212, + 535, + 787, + 717 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To tackle these challenges, we propose to transfer object knowledge from frozen image pre-trained models to videos for more efficient image-to-video adaptation. In essence, our approach entails two main steps: decoupling each frame into distinct object components, and subsequently identifying object state changes across time to facilitate comprehensive video comprehension as illustrated in Fig. 1. Specifically, we first pass the video frames through the image pre-trained model to obtain frame-wise features. Then, inspired by recent works on object discovery [51,82,83], we employ slot attention [51] with learnable queries [85] to", + "bbox": [ + 212, + 719, + 790, + 842 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "R. Qian et al.", + "bbox": [ + 271, + 114, + 364, + 128 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "parse each frame into a compact set of object tokens, serving as compressed representations for each frame. Thereafter, we feed these object tokens into object-time interaction layers to model the object state changes, which vividly depict the temporal dynamics in the video. Finally, we apply a linear head to the latent features representing these object state changes to produce the video-level prediction for action recognition. Unlike the existing works on object-centric representations, our method does not rely on object annotations [30,80,92], extra detectors [88] or additional modalities [91]. Instead, we develop two simple object-level losses to assist in distilling effective object-centric representations from image foundation models and establishing meaningful object state changes.", + "bbox": [ + 212, + 146, + 787, + 297 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this way, our method effectively mitigates the limitations of the previous work. Firstly, our method exhibits a notable enhancement in memory efficiency compared to conventional approaches. By exclusively finetuning the additional parameters on top of the image pre-trained model, we successfully reduce the memory required for gradient backpropagation. Additionally, compressing the video frames into object tokens significantly diminishes the computation requirements for temporal modeling, making our method more accessible and scalable. Secondly, by utilizing object-centric reasoning in the form of learnable object tokens and establishing the temporal state changes of individual objects, we inject strong inductive bias into the learning process that goes beyond basic temporal modules like attention layers. We demonstrate that our model can localize the specific objects in the video as a side-product, providing an insightful interpretation of the image-to-video adaptation.", + "bbox": [ + 212, + 297, + 787, + 493 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To summarize, our contributions are as follows. (1) We propose an efficient object-centric method to adapt image pre-trained models to the video domain. By parsing video frames into object tokens and establishing object state changes across time, we enhance temporal perception with reduced computation redundancy. (2) We achieve state-of-the-art performance on action recognition, presenting higher efficiency than existing efficient tuning counterparts. (3) Our method effectively discovers different objects in videos without object-level annotations and achieves robust zero-shot video object segmentation.", + "bbox": [ + 212, + 494, + 787, + 617 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 638, + 388, + 656 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Video action recognition is a fundamental problem in computer vision. Compared to image tasks, the extra temporal relationships between different frames are crucial for video understanding. To this end, prevalent methods build on existing image models [20, 29, 49], and make architectural modifications to enable temporal modeling [2, 10, 15, 16, 19, 21, 45, 50, 62, 64, 65, 74, 75]. These works either fully finetune the whole model [2, 5, 10, 21] or directly train on large-scale video data from scratch [66, 73], requiring huge computation cost. More recently, some works propose to recognize actions from the perspective of state changes [1, 54, 59, 71, 79]. This novel idea regards actions as transformations that make changes to the objects and environments. However, most of these works directly model the state changes on the representations of the entire scene, where", + "bbox": [ + 212, + 672, + 787, + 840 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "Object-centric Image-to-Video Adaptation", + "bbox": [ + 447, + 114, + 730, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "different semantics and objects are highly entangled. It is non-trivial to accurately perceive the state changes of different objects. In this work, we explicitly decouple different objects from the video scene, and respectively establish the state changes on individual objects to assist action recognition.", + "bbox": [ + 212, + 146, + 787, + 207 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Object-centric video representation learning aims at establishing robust representations for different objects to assist video understanding. To achieve this goal, a series of works borrow the idea from object discovery and employ iterative slot attention [51] to decompose different objects [3, 6, 12, 17, 18, 37, 39, 63, 85]. However, they are specifically trained for segmentation and can only deal with low-level segmentation tasks and lack the ability for high-level understanding. Another line of works uses the object-related features as a complement to the original video representations to enhance object awareness [24, 30, 53, 92, 93]. These methods achieve promising results on multiple action benchmarks [25, 26, 53], but still have two major limitations. First, they depend on bounding box annotations [30, 80, 92], additional detectors or modalities [88, 91] to identify different objects, restricting the scalability to unseen data. Second, the identified objects provide plentiful cues for video analysis, yet these methods still fuse the disentangled object representations with the original video features, imposing limitations on efficiency. In contrast, our process is free of object annotations or handcrafted priors but only adopts simple contrastive loss to distill general object knowledge from image foundation models and decouple object components in video frames. Then we directly model temporal state changes on compact object representations, enabling more intuitive and efficient video understanding.", + "bbox": [ + 212, + 220, + 787, + 508 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Efficient tuning becomes a tending direction with the development of large foundation models [7,14,69,94]. It is proposed in natural language processing to reduce the trainable parameters and improve efficiency when transferring pretrained models to downstream tasks [31,32,41,89]. Common strategies include inserting task-specific adapters into Transformer encoders [31,60,61], attaching prompt tokens to the input [42,48,67,70] and learning low-rank approximations [32]. Recently, the efficient tuning is increasingly explored in computer vision [4,11,22,35,44,72,95,96]. By freezing the image pre-trained model, [11,72] tune adapters with learnable parameters for efficient adaptation. [4,22,35,44] develop prompt tuning to adapt the frozen image model to various image tasks. Later, [36,46,47,55,57,58,68,77,86] take one step further to adapt image models to videos with temporal dynamics. In order to equip the model with temporal perception ability, [47,57,86] insert spatio-temporal modules into Transformer encoder blocks, but the tunable parameters in shallow Transformer layers require huge GPU memory in gradient backpropagation. [58,68] require two pathways to process spatial and temporal information. [36,46] attach spatio-temporal attention or motion decoders to the frame-wise features, but there exists high redundancy in attention computation [5,73]. Contrarily, our method retains the backbone architecture without modifications, distills object knowledge from image foundation models, and captures temporal state changes of discovered objects for video downstream tasks.", + "bbox": [ + 212, + 522, + 787, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "R. Qian et al.", + "bbox": [ + 271, + 114, + 364, + 128 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/dd7239b342f0ccbcc86a9401fad53167882a22250a0fdbf3c62af3b3fe820ab4.jpg", + "image_caption": [ + "Fig. 2: An overview of our object-centric image-to-video adaptation framework. We use a frozen image pre-trained model to extract frame-wise features and pass them through a lightweight temporal fusion block. Then we employ slot attention with learnable queries to decompose each frame into a compact set of object tokens. Thereafter, we develop object-time interaction layers to establish inter-object interactions and build temporal state changes of individual objects, which are then pooled into the video-level feature for action recognition." + ], + "image_footnote": [], + "bbox": [ + 215, + 143, + 781, + 253 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3 Method", + "text_level": 1, + "bbox": [ + 215, + 391, + 330, + 407 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The overall architecture of our object-centric image-to-video adaptation is shown in Fig. 2. We use a frozen pre-trained Vision Transformer [20] as the backbone to extract frame-wise features from the video clip. Then we apply slot attention with learnable query tokens to identify different object components and pass the object tokens through object-time interaction layers to model the temporal state changes of individual objects. The obtained object state change vectors encode rich semantics and essential temporal dynamics in the given video, and are subsequently pooled into the final video-level feature for downstream tasks.", + "bbox": [ + 212, + 422, + 787, + 544 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1 Object Decomposition", + "text_level": 1, + "bbox": [ + 215, + 565, + 447, + 580 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To start with, given a video clip $\\pmb{x} = \\{x_{1}, x_{2}, \\dots, x_{T}\\}$ consisting of $T$ frames, we use a frozen image pre-trained Vision Transformer $\\Phi_{\\mathrm{enc}}$ as the backbone to extract frame-wise features:", + "bbox": [ + 212, + 590, + 785, + 635 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\np _ {t}, f _ {t} = \\Phi_ {\\mathrm {e n c}} (x _ {t}), \\quad p _ {t} \\in \\mathbb {R} ^ {D}, \\quad f _ {t} \\in \\mathbb {R} ^ {H W \\times D}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 336, + 643, + 785, + 664 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $p_t$ and $f_t$ denote the output c1s token and feature map from the last Transformer block. $H, W, D$ respectively denote height, width, and channel dimension. These image pre-trained features encode rich but highly-entangled semantics. And the next step is to parse these frame features into different object components which are crucial for video perception and understanding [24,30,80,92].", + "bbox": [ + 212, + 672, + 787, + 750 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Inspired by recent works on unsupervised object discovery and segmentation [51,82,83], we develop a variant of slot attention on top of the feature map $f_{t}$ for object decomposition in each frame. To be specific, instead of random sampling from a prior distribution, we use $N$ learnable query tokens $q \\in \\mathbb{R}^{N \\times D}$ to initialize the slot vectors $S \\in \\mathbb{R}^{N \\times D}$ . The intuition of this design is to encourage each query token to capture specific semantics or concepts [33, 83] and produce", + "bbox": [ + 212, + 750, + 787, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "Object-centric Image-to-Video Adaptation", + "bbox": [ + 447, + 114, + 732, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "temporally coherent object decompositions, which provide essential inductive bias for temporal reasoning in later stages. Then at each slot attention iteration, following [51], we use three learnable linear transformations to project the slot vectors $S$ and frame features $f_{t}$ into query, key and value, i.e., $Q \\in \\mathbb{R}^{N \\times D}$ , $K \\in \\mathbb{R}^{HW \\times D}$ and $V \\in \\mathbb{R}^{HW \\times D}$ . After that, we compute the attention matrix $\\tilde{A} \\in \\mathbb{R}^{N \\times HW}$ with softmax normalization along the slot dimension as follows:", + "bbox": [ + 215, + 146, + 787, + 238 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {A} _ {i, j} := \\frac {e ^ {A _ {i , j}}}{\\sum_ {l = 1} ^ {N} e ^ {A _ {l , j}}}, \\quad \\text {w h e r e} \\quad A := \\frac {1}{\\sqrt {D}} Q K ^ {T}. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 338, + 261, + 785, + 299 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We use the weighted mean to aggregate the value and pass them through Gated Recurrent Unit (GRU) to update the slot vectors:", + "bbox": [ + 215, + 308, + 785, + 339 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nS := \\operatorname {G R U} (\\text {i n p u t s} = M V, \\text {s t a t e s} = S), \\quad \\text {w h e r e} \\quad M _ {i, j} := \\frac {\\tilde {A} _ {i , j}}{\\sum_ {l = 1} ^ {H W} \\tilde {A} _ {i , l}}. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 243, + 351, + 785, + 390 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We iterate the routing process three times, and take the final slot attention weights as potential object segmentation masks. The slot vectors from the final iteration are treated as the object tokens that distill object knowledge from the pre-trained image features. Note that we perform slot attention $\\Phi_{\\mathrm{sa}}$ on each frame in parallel, and denote the object tokens of each time stamp as:", + "bbox": [ + 215, + 398, + 785, + 474 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {o} = \\left\\{o _ {1} \\oplus o _ {2} \\oplus \\dots \\oplus o _ {T} \\right\\} \\in \\mathbb {R} ^ {T \\times N \\times D}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 372, + 484, + 785, + 503 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $o_t = \\Phi_{\\mathrm{sa}}(q, \\tilde{f}_t) \\in \\mathbb{R}^{N \\times D}$ , and $\\oplus$ denotes concatenation operation along time dimension. Since this attention mechanism introduces competition among slots, the obtained object tokens $o_t$ are expected to take over distinct semantic parts and separate each frame into $N$ different components.", + "bbox": [ + 212, + 516, + 785, + 577 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Surprisingly, we find that the distilled tokens and corresponding attention matrices perform remarkably well in zero-shot video object segmentation tasks. They are even comparable to some pixel-level supervised counterparts, as demonstrated in Section 4.3. This result verifies that our method effectively extracts robust object-centric features for temporal reasoning in the later stage. Note that since videos often contain partial object semantics, there tend to be some slots that carry background information and interact with the foreground object tokens for comprehensive video analysis.", + "bbox": [ + 212, + 579, + 787, + 700 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.2 Object-Time Interaction", + "text_level": 1, + "bbox": [ + 215, + 722, + 465, + 738 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "After obtaining the object tokens that reveal potential object semantics, we use them as compressed representations of video frames and further model the temporal dynamics from an object-centric perspective [30, 93]. Our intuition is that in continuous frames, the object tokens, $o_{t}^{n} \\in \\mathbb{R}^{D}, t = 1,2,\\dots,T$ , with the same index $n$ tend to represent the same objects across time as illustrated in Fig. 1. Hence, it is feasible to temporally track each object and conclude the", + "bbox": [ + 212, + 750, + 787, + 842 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "R. Qian et al.", + "bbox": [ + 271, + 114, + 364, + 128 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "object state changes based on the learned object tokens, thereby assisting the process of video understanding.", + "bbox": [ + 212, + 146, + 782, + 176 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To achieve this goal, we first apply self-attention to the object tokens within the same frame to introduce inter-object interactions at each time stamp:", + "bbox": [ + 212, + 176, + 784, + 205 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {o} _ {t} = \\Phi_ {\\mathrm {o i}} \\left(o _ {t}\\right) \\in \\mathbb {R} ^ {N \\times D}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 424, + 215, + 784, + 233 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $\\Phi_{\\mathrm{oi}}$ is the object interaction module instantiated by a standard Transformer encoder layer consisting of linear projections, self-attention calculation and a feed forward network. The resulting vector $\\tilde{\\sigma}_t^n$ not only encodes the attribute of object $n$ but also captures its relationships with contextual objects, serving as a representation for the latent state of object $n$ at time stamp $t$ .", + "bbox": [ + 212, + 243, + 782, + 318 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Considering that the temporal state changes are a concrete manifestation of the effects of actions, they serve as an important cue for analyzing the events occurring in a video and assisting video understanding [1,59,71,79]. Hence, based on the vectors representing latent object states, we take one step further to look into multiple time stamps and explicitly establish the state changes of each object. Without loss of generality, we take the object with index $n$ for illustration. Given the object states at different time stamps, we define a temporal interval $\\delta$ to sample the initial and final states, i.e., $\\tilde{o}_t^n$ as the initial state and $\\tilde{o}_{t + \\delta}^n$ as the final state. We concatenate them along channel dimension and pass them through the temporal interaction layer $\\Phi_{\\mathrm{ti}}$ to model the object state change between time $t$ and $t + \\delta$ :", + "bbox": [ + 212, + 320, + 784, + 484 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\ns _ {t} ^ {n} = \\Phi_ {\\mathrm {t i}} \\left(\\left[ \\tilde {\\alpha} _ {t} ^ {n} \\oplus \\tilde {\\alpha} _ {t + \\delta} ^ {n} \\right]\\right) \\in \\mathbb {R} ^ {D}, \\quad t = \\{1, 2, \\dots , T - \\delta \\}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 318, + 494, + 784, + 513 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $\\oplus$ denotes channel-wise concatenation. We instantiate $\\Phi_{\\mathrm{ti}}$ as a multi-layer perceptron of channel dimension $2D - D - D$ . In this way, for each object, we obtain $T' = T - \\delta$ state change vectors. The state change matrix covering all objects and time stamps can be denoted as $\\boldsymbol{s} \\in \\mathbb{R}^{T' \\times N \\times D}$ .", + "bbox": [ + 212, + 522, + 782, + 582 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Since the state change is a continuous process within a video sequence, it is feasible to conclude the overall state change of each object throughout the entire video by simple temporal average pooling on the established matrix $\\mathbf{s}$ . Thereafter, we further apply average pooling on object dimension to aggregate the state change information of all objects as the final video-level representation, $v \\in \\mathbb{R}^D$ , for downstream action recognition.", + "bbox": [ + 212, + 584, + 784, + 672 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Discussion. Unlike the existing works that use object features as complements to original video features [24,30,53,92,93] or model state changes on the highly entangled scene representations [1, 59, 71, 79], we perform temporal reasoning solely on distilled object tokens. Our object-centric state change modeling reduces redundancy, mitigates ambiguity and results in more fine-grained cues for video understanding. And comparing with the prevalent image-to-video adaptation methods, they introduce dense temporal fusion operations, e.g., temporal convolutions [46, 47, 57] and attention [46, 47, 58, 68, 86], to mix multi-frame features for action recognition. In contrast, we leverage the object state changes to perceive temporal dynamics, which closely resemble human logic and provide an effective inductive bias to the model. To be specific, it does not require dense", + "bbox": [ + 212, + 674, + 785, + 839 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "Object-centric Image-to-Video Adaptation", + "bbox": [ + 447, + 114, + 730, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "fusion or traversal of all frames. Instead, we respectively model state changes on separated objects across sparsely sampled frames, enabling more efficient scene parsing and temporal perception as verified in Section 4.4.", + "bbox": [ + 212, + 146, + 787, + 194 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "3.3 Training", + "text_level": 1, + "bbox": [ + 215, + 214, + 333, + 229 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In training, we freeze the parameters in the image pre-trained model $\\Phi_{\\mathrm{enc}}$ , and tune parameters that are responsible for object decomposition and temporal perception, i.e., $\\Phi_{\\mathrm{temp}}$ , $\\Phi_{\\mathrm{oi}}$ and $\\Phi_{\\mathrm{ti}}$ . Besides the standard cross-entropy loss for action classification, we introduce two object-level losses to further facilitate object decomposition and temporal state change modeling.", + "bbox": [ + 212, + 239, + 782, + 314 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Object Distillation Loss. Considering that the image pre-trained foundation models tend to encode rich semantics in the given scene [9, 56, 69], the output c1s token $p_t$ is expected to encompass the majority of the semantic components in frame $t$ . Hence, we use a contrastive grounding loss [23, 27, 90] to align the learned object tokens $o_t$ and the c1s token $p_t$ , encouraging $o_t$ to cover all objects in frame $t$ . Particularly, we first define the correspondence score between the $o_t$ and $p_t$ as:", + "bbox": [ + 212, + 315, + 784, + 421 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\nc \\left(o _ {t}, p _ {t}\\right) = \\sum_ {n = 1} ^ {N} \\frac {e ^ {\\left\\langle o _ {t} ^ {n} , p _ {t} \\right\\rangle / \\tau}}{\\sum_ {l = 1} ^ {N} e ^ {\\left\\langle o _ {t} ^ {l} , p _ {t} \\right\\rangle / \\tau}} \\left\\langle o _ {t} ^ {n}, p _ {t} \\right\\rangle , \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 364, + 433, + 785, + 474 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $\\langle o_t^n,p_t\\rangle = \\frac{o_t^n p_t}{\\|o_t^n\\|\\|p_t\\|}$ denotes cosine similarity calculation, $\\tau$ is the temperature hyper-parameter. This correspondence score adaptively aggregates each object component and avoids penalizing unrelated object semantics. Thereafter, we randomly sample c1s token vectors $p^\\prime$ from other videos within the minibatch to formulate a negative sample pool $\\mathcal{N}$ and apply the contrastive loss to maximize the agreement between the aligned object and c1s token pairs:", + "bbox": [ + 212, + 486, + 787, + 583 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {o b j} = - \\sum_ {t = 1} ^ {T} \\log \\frac {e ^ {c \\left(o _ {t} , p _ {t}\\right) / \\tau}}{e ^ {c \\left(o _ {t} , p _ {t}\\right) / \\tau} + \\sum_ {p ^ {\\prime} \\sim \\mathcal {N}} e ^ {c \\left(o _ {t} , p ^ {\\prime}\\right) / \\tau}}. \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 336, + 593, + 785, + 633 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This self-supervised alignment objective encourages the object tokens to cover diverse semantics and distill object knowledge from the image pre-trained model. Temporal Reasoning Loss. In our architecture, we analyze temporal dynamics from the perspective of object state changes. Therefore, it is crucial to formulate a learning objective that guides the temporal interaction module $\\Phi_{\\mathrm{ti}}$ to accurately capture the state changes between different time stamps. To achieve this goal, our intuition is that the same objects undergoing the same type of actions should possess consistent state changes. Conversely, different actions will result in distinct state changes for the same object, and different objects will exhibit diverse state changes under the same action. Based on this, we develop a margin loss to guide object state change modeling. In detail, given $s_t^n$ as query, we define the videos within the mini-batch that have the same action labels with the current video to form positive pool $\\mathcal{P}$ , and those of different action labels as", + "bbox": [ + 212, + 643, + 787, + 843 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "R. Qian et al.", + "bbox": [ + 271, + 114, + 364, + 128 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "negative pool $\\mathcal{N}$ . We sample the state change vectors of the same object and the same action, i.e., $\\{\\hat{s}_t^n |\\hat{s}\\in \\mathcal{P}\\}$ , as positive pairs, which should be aligned with $s_t^n$ . As for the negatives, we respectively sample the state change vectors of the same object but different actions, i.e., $\\{s_t^{l^{\\prime}n}|s^{\\prime}\\in \\mathcal{N}\\}$ , and those from the same video but of different objects, i.e., $\\{s_t^m,m\\neq n\\}$ , as hard negatives for discrimination. The margin loss with a margin hyper-parameter $\\lambda$ is formulated as:", + "bbox": [ + 214, + 146, + 782, + 237 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {\\text {t e m p}} = \\sum_ {t = 1} ^ {T ^ {\\prime}} \\sum_ {n = 1} ^ {N} \\left\\{\\sum_ {\\hat {s} \\in \\mathcal {P}} \\| s _ {t} ^ {n} - \\hat {s} _ {t} ^ {n} \\| _ {2} \\right. \\\\ + \\sum_ {s ^ {\\prime} \\in \\mathcal {N}} \\max \\left(\\lambda - \\| s _ {t} ^ {n} - s _ {t} ^ {\\prime n} \\| _ {2}, 0\\right) + \\sum_ {\\substack {m = 1 \\\\ m \\neq n}} ^ {N} \\max \\left(\\lambda - \\| s _ {t} ^ {n} - s _ {t} ^ {m} \\| _ {2}, 0\\right) \\}, \\tag{9} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 233, + 250, + 784, + 340 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where $\\| \\cdot \\|_2$ denotes $\\mathcal{L}_2$ distance. This term enforces the model to discriminate diverse object states and capture unique dynamic effects of each action category. Overall Objectives. Besides the above object-level losses, for action recognition tasks, we apply a linear classification head on the video-level representation $v$ with a supervised classification loss $\\mathcal{L}_{cls}$ in the form of cross-entropy. In this way, we formulate the overall loss function as:", + "bbox": [ + 214, + 352, + 782, + 443 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathcal {L} _ {o b j} + \\mathcal {L} _ {t e m p} + \\mathcal {L} _ {c l s}. \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 411, + 457, + 784, + 470 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 215, + 494, + 375, + 512 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.1 Implementation Details", + "text_level": 1, + "bbox": [ + 215, + 525, + 455, + 540 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We train and evaluate on three popular action recognition datasets: Kinetics-400 (K-400) [38], Something-Something-v2 (SSv2) [25], and Epic-Kitchens-100 (EK100) [13]. Additionally, we assess zero-shot video object segmentation on challenging multiple object segmentation dataset, DAVIS-2017-UVOS [8]. We use the frozen CLIP [69] pre-trained Vision Transformer [20] as $\\Phi_{\\mathrm{enc}}$ for framewise feature extraction. In training, we crop each frame into $224\\times 224$ , set the number of query tokens $N = 8$ , and the time interval for object state change $\\delta = T / 4$ in default. We use AdamW optimizer [52] with an initial learning rate $5\\times 10^{-4}$ and mini-batch size 512 to update the tunable parameters.", + "bbox": [ + 214, + 550, + 782, + 686 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.2 Video Action Recognition", + "text_level": 1, + "bbox": [ + 215, + 709, + 475, + 724 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "In this section, we present the comparisons between our method and recent state-of-the-art on K-400, SSv2 and EK100 in terms of the number of tunable parameters, inference GFLOPs and action recognition accuracy.", + "bbox": [ + 212, + 734, + 782, + 777 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "K-400. We first show the results on K-400 in Table 1 and conclude three observations. (1) Compared to fully finetuned models, our method achieves superior performance with significantly fewer tunable parameters. For example, our method of ViT-L/14 with only 21M tunable parameters outperforms all", + "bbox": [ + 214, + 779, + 782, + 839 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "Object-centric Image-to-Video Adaptation", + "bbox": [ + 447, + 114, + 730, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/bed9471378e3b04a4bb94b657ee919384878de40929b1216cfde8a8e42e1c1cc.jpg", + "table_caption": [ + "Table 1: Results on Kinetics-400. We report the pre-train model initialization, inference GFLOPs, the number of tunable parameters (M), Top-1 and Top-5 accuracy. The 'Views' indicates Frame Number $\\times$ Temporal Clips $\\times$ Spatial Crops." + ], + "table_footnote": [], + "table_body": "
MethodPre-trainViewsGFLOPs TunableTop-1Top-5
Full finetuning
TimeSformer-L [5]IN-21K64 × 1 × 3714012180.794.7
VideoSwin-L [50]IN-21K32 × 4 × 3724819783.195.9
MTV-L [84]JFT32 × 4 × 31805087684.396.3
PromptCLIP A7 [36]CLIP16 × 5 × 1--76.893.5
ActionCLIP [77]CLIP32 × 10 × 31689014283.897.1
X-CLIP-L/14 [55]CLIP8 × 4 × 3789042087.197.6
Efficient tuning
EVL ViT-L/14 [46]CLIP32 × 3 × 180885987.3-
ST-Adapter ViT-B/16 [57]CLIP32 × 3 × 118217.282.796.2
STAN-self-B/16 [47]CLIP16 × 3 × 11187-84.996.8
AIM ViT-B/16 [86]CLIP32 × 3 × 124281184.796.7
DualPath ViT-B/16 [58]CLIP32 × 3 × 17101085.497.1
DiST ViT-B/16 [68]CLIP32 × 3 × 11950-85.097.0
DiST ViT-L/14 [68]CLIP32 × 3 × 18490-88.097.9
Ours ViT-B/16CLIP8 × 3 × 12816.585.897.3
Ours ViT-B/16CLIP16 × 3 × 15626.586.197.5
Ours ViT-B/16CLIP32 × 3 × 111236.586.297.5
Ours ViT-L/14CLIP32 × 3 × 150682188.598.0
", + "bbox": [ + 230, + 199, + 767, + 506 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "the fully finetuned approaches that require over 100M tunable parameters. (2) Our method with only a visual encoder achieves better results than the multimodal methods, ActionCLIP [77] and X-CLIP [55], which employ an extra text branch to guide the finetuning process. (3) Among the efficient tuning methods [46, 47, 57, 58, 68, 86], EVL [46] attaches a motion decoder with multiple temporal convolution and attention layers for temporal modeling. AIM [86], ST-Adapter [57] and STAN [47] insert temporal blocks into intermediate CLIP encoder layers. DualPath [58] and DiST [68] introduce an additional temporal pathway to capture temporal dynamics. Our formulation models temporal relationships on decomposed objects in the form of state changes, which reduces computation redundancy but attains superior performance. For example, on ViT-B/16, our method with only an 8-frame input yields superior results to other approaches that employ a 32-frame input. On ViT-L/14, we reach the state-of-the-art results with $40\\%$ fewer GFLOPs (5068 vs 8490) than DiST [68]. SSv2. When extending to SSv2 which requires stronger temporal modeling, the observations are consistent. As depicted in Table 2, our object-centric formulation achieves state-of-the-art results and presents higher efficiency than both full finetuning and efficient tuning methods. In particular, our method with ViT-L/14 surpasses the advanced MViTv2-L [43] and DiST [68] with significantly reduced GFLOPs (5068 vs 8484, 8490). Moreover, on ViT-B/16, our method", + "bbox": [ + 212, + 536, + 787, + 843 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "R. Qian et al.", + "bbox": [ + 271, + 114, + 364, + 128 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/046dd8b6c61139186e4a1974ff79a8b84b174cec7b925dd057ac81bdf07aba7c.jpg", + "table_caption": [ + "Table 2: Results on Something-Something-v2. K-400†/K-600† indicates the model is pre-trained on both IN-21K and K-400/K-600." + ], + "table_footnote": [], + "table_body": "
MethodPre-trainViewsGFLOPsTunableTop-1Top-5
Full finetuning
TimeSformer-L [5]IN-21K64 × 1 × 1714012162.4-
MTV-B [84]IN-21K32 × 4 × 3479031067.690.4
ViViT-L/16×2 [2]K-400†16 × 4 × 31189231165.489.8
VideoSwin-B [50]K-400†32 × 1 × 19638969.692.7
MVViTv2-L (312↑) [43]K-400†32 × 1 × 3848421373.394.1
Efficient tuning
EVL ViT-L/14 [46]CLIP32 × 1 × 3964117566.7-
ST-Adapter ViT-B/16 [57]CLIP32 × 3 × 119557.269.592.6
STAN-self-B/16 [47]CLIP16 × 3 × 11376-69.592.7
AIM ViT-B/16 [86]CLIP32 × 1 × 324961469.192.2
DualPath ViT-B/16 [58]CLIP32 × 1 × 37161370.392.9
DiST ViT-B/16 [68]CLIP32 × 3 × 11950-70.992.1
DiST ViT-L/14 [68]CLIP32 × 3 × 18490-73.193.2
Ours ViT-B/16CLIP8 × 1 × 32816.571.293.1
Ours ViT-B/16CLIP16 × 1 × 35626.571.693.3
Ours ViT-B/16CLIP32 × 1 × 311236.571.893.4
Ours ViT-L/14CLIP32 × 1 × 350682173.694.3
", + "bbox": [ + 233, + 184, + 763, + 479 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "with only 8 frames as input outperforms others using 32 frames. It is non-trivial to achieve this on such a motion-heavy dataset, which in return reveals the significance of our temporal modeling in the form of object state changes. On the one hand, it enables accurate temporal perception with sparser sampling. On the other hand, our temporal modeling on individual objects eliminates the need for dense temporal fusion, substantially reducing computational redundancy.", + "bbox": [ + 212, + 506, + 784, + 595 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "EK100. On challenging egocentric EK100 that requires long-term object interactions, our method showcases substantial advantages over existing efficient tunning counterparts [46,57,68] in Table 3. The underlying reasons are two-fold. First, our method involves explicit object decomposition, which helps to capture the primary objects in egocentric videos and contributes to the significant improvements in nouns. Second, our temporal modeling on individual objects reduces the interference of camera motions and enforces the model to focus on the dynamics of each object, thus facilitating verb perception.", + "bbox": [ + 212, + 597, + 785, + 717 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4.3 Zero-shot Video Object Segmentation", + "text_level": 1, + "bbox": [ + 215, + 739, + 571, + 755 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "To verify whether our method understands videos in an object-centric way and interpret why the model works, we evaluate unsupervised video object segmentation in addition to action recognition. Particularly, we use the frozen ViT-B/16 backbone and adopt slot attention weights $\\{\\tilde{A}_i\\in \\mathbb{R}^{HW}\\}_{i = 1}^N$ in Eq. 2 as object masks candidate without bells and whistles. We report $\\mathcal{J}\\& \\mathcal{F}$ on challenging", + "bbox": [ + 212, + 763, + 785, + 840 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "Object-centric Image-to-Video Adaptation", + "bbox": [ + 447, + 114, + 730, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "multiple object discovery benchmark, DAVIS-2017-UVOS [8]. Note that we directly apply the model trained on K-400 or SSv2 to video object segmentation in a zero-shot manner.", + "bbox": [ + 212, + 146, + 782, + 191 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/bef05998a54056c73ca85c0fd6c3df90f15bfce75006fd704ef79bf32ba8c821.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodVerbNounAction
EVL ViT-B/16 [46]62.751.037.7
ST-Adapter ViT-B/16 [57]67.755.0-
DiST ViT-B/16 [68]69.558.145.8
DiST ViT-L/14 [68]70.761.648.9
Ours ViT-B/1671.160.949.0
Ours ViT-L/1473.264.551.7
", + "bbox": [ + 217, + 262, + 524, + 393 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/a7404338429f733595e158cb547e8ffc59fda522a17822bcd4758453733aec43.jpg", + "table_caption": [ + "Table 3: Results on Epic-Kitchens-100 Verb, Table 4: Results on multiple object Noun and Action. All compared methods take segmentation on DAVIS-2017-UVOS. an input of $8\\times 1\\times 3$" + ], + "table_footnote": [], + "table_body": "
ModelJ&FJF
OCLR [82]39.638.241.1
SMTC [63]40.536.444.6
BA [17]43.939.248.6
RVOS [76]41.236.845.7
Ours SSv241.037.644.4
Ours K-40044.742.347.1
Ours K-400 CRF49.346.552.1
", + "bbox": [ + 527, + 261, + 779, + 393 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Quantitative Results. We present the comparison with recent state-of-the-arts without object annotations [63,81,82] as well as a fully supervised baseline [76] (denoted in grey) in Table 4. Among the compared works without mask annotations, OCLR [82] takes optical flow as input to distinguish motion patterns of diverse objects. SMTC [63] designs two-stage slot attention to produce temporally consistent object segmentations. BA [17] also utilizes image foundation models like DION [9] to provide spatio-temporal correspondence and use clustering to predict the object segmentation masks. On the one hand, since K-400 covers more diverse scenes and objects than SSv2, it leads to better generalization ability in this zero-shot segmentation scenario. On the other hand, our formulation trained on K-400 achieves superior results to these counterparts, even surpassing the simple supervised baseline RVOS [76]. This demonstrates the feasibility of transferring object knowledge from image foundation models to the video domain, and our object-centric temporal reasoning further facilitates temporally coherent object decomposition. However, the gap to the state-of-the-art supervised method is still large because the mask annotations lead to much more accurate segmentation boundaries. A straightforward way is to apply Conditional Random Field (CRF) [40] for refinement, which leads to 4.6 points improvement on $\\mathcal{J}\\&\\mathcal{F}$ score. It is worth further exploration in the future.", + "bbox": [ + 212, + 430, + 787, + 717 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Qualitative Results. We also visualize some examples of the generated segmentation masks in Fig. 3, where each column presents the frames within the same video and the same color denotes the masks generated by the same slot query. Generally, our model discriminates different object semantics in videos. Each distilled object token can capture specific semantics or concepts that generalize across diverse time stamps and scenes. For example, the token marked in red focuses on humans, the blue one emphasizes quadruped animals, and the yellow one identifies two-wheeled vehicles. This temporally coherent object", + "bbox": [ + 212, + 719, + 787, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "R. Qian et al.", + "bbox": [ + 271, + 114, + 364, + 128 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "decomposition provides an effective inductive bias for subsequent temporal reasoning, leading to more efficient temporal modeling.", + "bbox": [ + 212, + 146, + 782, + 176 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "4.4 Ablation Study", + "text_level": 1, + "bbox": [ + 214, + 200, + 390, + 215 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "In this section, we use the CLIP ViT-B/16 as $\\Phi_{\\mathrm{enc}}$ with an input of 16 frames for all ablative experiments unless otherwise specified. More details and extra ablation studies are presented in Supplementary Materials.", + "bbox": [ + 212, + 227, + 784, + 272 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/d83ba3e1789da07a8fb77c42e0ba8d4bc92b5f9c55a1da1d0287ab2e4ebbe5e5.jpg", + "image_caption": [ + "Fig. 3: Visualization of object decomposition in Fig. 4: Comparison on training data the form of segmentation. Each column presents efficiency. Our method demonstrates three frames in a video. The same color denotes advantages, especially with a small the objects identified by the same object token. amount of training data." + ], + "image_footnote": [], + "bbox": [ + 217, + 300, + 767, + 407 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Training Losses. We explore the effectiveness of our newly introduced object distillation loss $\\mathcal{L}_{obj}$ and temporal reasoning loss $\\mathcal{L}_{temp}$ in Table 5. We observe that $\\mathcal{L}_{obj}$ leads to dramatic improvements on all three datasets, K-400, SSv2, and DAVIS-2017-UVOS. This is because $\\mathcal{L}_{obj}$ offers rich semantic knowledge reference from the image foundation model to facilitate object identification. And better object decomposition further enhances high-level action recognition. As for $\\mathcal{L}_{temp}$ , it guides the model to capture the state changes on each object. Without this guidance, relying solely on the action classification loss is insufficient to direct the temporal module to perceive representative temporal dynamics, thus resulting in a drastic performance drop on the motion-heavy dataset SSv2. And notably, this temporal loss also leads to improvements on DAVIS, which verifies that the temporal perception in return bolsters object decomposition.", + "bbox": [ + 212, + 507, + 787, + 688 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Temporal Sampling Strategy. We also study the effect of different temporal sampling strategies in training. Specifically, we vary the number of frames per clip $T$ and the time interval for object state change modeling $\\delta$ . In Table 6, we display the Top-1 accuracy on K-400 and SSv2, as well as the computation GFLOPs excluding ViT backbone for more intuitive comparison. An exciting phenomenon is that even on SSv2 that requires strong temporal modeling, we achieve promising results with very sparse sampling, i.e., only 4 frames as input with significantly reduced computation. Our method is generally robust to the temporal sampling hyper-parameters, demonstrating the effectiveness of our temporal reasoning on compressed object tokens.", + "bbox": [ + 212, + 688, + 787, + 840 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "Object-centric Image-to-Video Adaptation", + "bbox": [ + 447, + 114, + 730, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 784, + 126 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/f7d9665ef3ab0d9d5768c611b1dbf8c3747cd2dba97ddfed6d69dab7f9b0ce9f.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
LobjLtempK-400SSv2DAVIS
XX70.444.827.3
✓X80.355.239.5
X✓73.552.130.5
✓✓86.171.644.7
", + "bbox": [ + 222, + 198, + 480, + 284 + ], + "page_idx": 13 + }, + { + "type": "table", + "img_path": "images/80e2408a24c6c2c41b47dc27fc340a9bf13efbca970b2f43f870d24b718e29fb.jpg", + "table_caption": [ + "Table 5: Ablation studies on training loss Table 6: Ablation studies on the number items $\\mathcal{L}_{obj}$ and $\\mathcal{L}_{temp}$ . Note that $\\mathcal{L}_{cls}$ is of frames $T$ and sampling interval $\\delta$ . always applied in training." + ], + "table_footnote": [], + "table_body": "
TδGFLOPsK-400SSv2
414.285.470.7
828.485.871.2
848.285.871.1
16218.086.171.6
16417.886.171.6
", + "bbox": [ + 514, + 185, + 741, + 284 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Training Data Efficiency. Fig. 4 presents K-400 action recognition performance with various amounts of training data. For fair comparison with fully fine-tuned TimeSformer [5] as well as the efficient tuning methods, ST-Adapter [57] and AIM [86], we use CLIP ViT-B/16 with an input of 8 frames. Our method exhibits superior data efficiency to existing works. And the advantage becomes more significant with less training data, demonstrating the robust capacity of our object-centric formulation to transfer image pre-trained knowledge to videos.", + "bbox": [ + 212, + 313, + 787, + 420 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 441, + 359, + 458 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In this work, we propose an object-centric formulation to adapt the image pretrained models to the video domain. Specifically, we employ a frozen image pre-trained model to extract frame-wise features and use slot attention with learnable queries to parse each frame into a compact set of object tokens. Then we apply object-time interaction to these object tokens to explicitly establish object state changes across time span. By using the distilled object tokens as the compressed representations of video frames, we have reduced computation costs and improved temporal reasoning capabilities especially under sparse sampling scenarios. Our method demonstrates state-of-the-art performance across action recognition benchmarks while realizing robust zero-shot video object segmentation, signifying a promising direction for future research in video analysis.", + "bbox": [ + 212, + 474, + 785, + 640 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Limitations and Social Negative Impact. Our work demonstrates that the object-centric formulation serves as a reliable compression of video frames, which reduces redundancy and enables efficient temporal perception. But in some applications like a video QA system, it would be better to distill useful information according to specific questions. Incorporating questions as a condition to slot attention might address this problem. And there is a potential negative impact in privacy concerns. This work involves processing large amounts of visual data, which could include sensitive information and raise privacy concerns.", + "bbox": [ + 212, + 641, + 785, + 762 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "R. Qian et al.", + "bbox": [ + 271, + 114, + 364, + 128 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 217, + 143, + 401, + 162 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "This work is partially supported by the National Key R&D Program of China (2022ZD0160201), Shanghai Artificial Intelligence Laboratory, the Centre for Perceptual and Interactive Intelligence (CPII) Ltd under the Innovation and Technology Commission (ITC)'s InnoHK.", + "bbox": [ + 215, + 172, + 785, + 233 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 253, + 321, + 270 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Alayrac, J.B., Miech, A., Laptev, I., Sivic, J., et al.: Multi-task learning of object states and state-modifying actions from web videos. IEEE Transactions on Pattern Analysis and Machine Intelligence (2024)", + "2. Arnab, A., Dehghani, M., Heigold, G., Sun, C., Lucic, M., Schmid, C.: Vivit: A video vision transformer. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 6836-6846 (2021)", + "3. Aydemir, G., Xie, W., Guney, F.: Self-supervised object-centric learning for videos. In: Thirty-seventh Conference on Neural Information Processing Systems (2023), https://openreview.net/forum?id=919tWtJPXe", + "4. Bahng, H., Jahanian, A., Sankaranarayanan, S., Isola, P.: Visual prompting: Modifying pixel space to adapt pre-trained models. arXiv preprint arXiv:2203.17274 (2022)", + "5. Bertasius, G., Wang, H., Torresani, L.: Is space-time attention all you need for video understanding? In: ICML. vol. 2, p. 4 (2021)", + "6. Besbinar, B., Frossard, P.: Self-supervision by prediction for object discovery in videos. In: 2021 IEEE International Conference on Image Processing (ICIP). pp. 1509-1513. IEEE (2021)", + "7. Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J.D., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., et al.: Language models are few-shot learners. Advances in neural information processing systems 33, 1877-1901 (2020)", + "8. Caelles, S., Pont-Tuset, J., Perazzi, F., Montes, A., Maninis, K.K., Van Gool, L.: The 2019 davis challenge on vos: Unsupervised multi-object segmentation. arXiv preprint arXiv:1905.00737 (2019)", + "9. Caron, M., Touvron, H., Misra, I., Jégou, H., Mairal, J., Bojanowski, P., Joulin, A.: Emerging properties in self-supervised vision transformers. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 9650-9660 (2021)", + "0. Carreira, J., Zisserman, A.: Quo vadis, action recognition? a new model and the kinetics dataset. In: proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 6299-6308 (2017)", + "1. Chen, S., Ge, C., Tong, Z., Wang, J., Song, Y., Wang, J., Luo, P.: Adaptformer: Adapting vision transformers for scalable visual recognition. Advances in Neural Information Processing Systems 35, 16664-16678 (2022)", + "2. Crawford, E., Pineau, J.: Exploiting spatial invariance for scalable unsupervised object tracking. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 34, pp. 3684-3692 (2020)", + "3. Damen, D., Doughty, H., Farinella, G.M., Furnari, A., Kazakos, E., Ma, J., Moltisanti, D., Munro, J., Perrett, T., Price, W., et al.: Rescaling egocentric vision. arXiv preprint arXiv:2006.13256 (2020)", + "4. Devlin, J., Chang, M.W., Lee, K., Toutanova, K.: Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805 (2018)" + ], + "bbox": [ + 225, + 281, + 784, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "Object-centric Image-to-Video Adaptation", + "bbox": [ + 447, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "15. Ding, S., Li, M., Yang, T., Qian, R., Xu, H., Chen, Q., Wang, J., Xiong, H.: Motion-aware contrastive video representation learning via foreground-background merging. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 9716-9726 (2022)", + "16. Ding, S., Qian, R., Xiong, H.: Dual contrastive learning for spatio-temporal representation. In: Proceedings of the 30th ACM international conference on multimedia. pp. 5649-5658 (2022)", + "17. Ding, S., Qian, R., Xu, H., Lin, D., Xiong, H.: Betrayed by attention: A simple yet effective approach for self-supervised video object segmentation. arXiv preprint arXiv:2311.17893 (2023)", + "18. Ding, S., Xie, W., Chen, Y., Qian, R., Zhang, X., Xiong, H., Tian, Q.: Motion-inductive self-supervised object discovery in videos. arXiv preprint arXiv:2210.00221 (2022)", + "19. Ding, S., Zhao, P., Zhang, X., Qian, R., Xiong, H., Tian, Q.: Prune spatiotemporal tokens by semantic-aware temporal accumulation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 16945-16956 (2023)", + "20. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929 (2020)", + "21. Feichtenhofer, C., Fan, H., Malik, J., He, K.: Slowfast networks for video recognition. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 6202-6211 (2019)", + "22. Feng, C., Zhong, Y., Jie, Z., Chu, X., Ren, H., Wei, X., Xie, W., Ma, L.: Promptdet: Towards open-vocabulary detection using uncurated images. In: Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part IX. pp. 701-717. Springer (2022)", + "23. Ghiasi, G., Gu, X., Cui, Y., Lin, T.Y.: Scaling open-vocabulary image segmentation with image-level labels. In: Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXXVI. pp. 540-557. Springer (2022)", + "24. Girdhar, R., Carreira, J., Doersch, C., Zisserman, A.: Video action transformer network. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 244-253 (2019)", + "25. Goyal, R., Ebrahimi Kahou, S., Michalski, V., Materzynska, J., Westphal, S., Kim, H., Haenel, V., Fruend, I., Yianilos, P., Mueller-Freitag, M., et al.: The\" something something\" video database for learning and evaluating visual common sense. In: Proceedings of the IEEE international conference on computer vision. pp. 5842-5850 (2017)", + "26. Gu, C., Sun, C., Ross, D.A., Vondrick, C., Pantofaru, C., Li, Y., Vijayanarasimhan, S., Toderici, G., Ricco, S., Sukthankar, R., et al.: Ava: A video dataset of spatiotemporally localized atomic visual actions. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 6047-6056 (2018)", + "27. Gupta, T., Vahdat, A., Chechik, G., Yang, X., Kautz, J., Hoiem, D.: Contrastive learning for weakly supervised phrase grounding. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part III. pp. 752-768. Springer (2020)", + "28. He, K., Chen, X., Xie, S., Li, Y., Dollár, P., Girshick, R.: Masked autoencoders are scalable vision learners. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16000-16009 (2022)" + ], + "bbox": [ + 215, + 146, + 785, + 840 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "R. Qian et al.", + "bbox": [ + 271, + 114, + 364, + 128 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "29. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 770-778 (2016)", + "30. Herzig, R., Ben-Avraham, E., Mangalam, K., Bar, A., Chechik, G., Rohrbach, A., Darrell, T., Globerson, A.: Object-region video transformers. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3148-3159 (2022)", + "31. Houlsby, N., Giurgiu, A., Jastrzebski, S., Morrone, B., De Laroussilhe, Q., Gesmundo, A., Attariyan, M., Gelly, S.: Parameter-efficient transfer learning for nlp. In: International Conference on Machine Learning. pp. 2790-2799. PMLR (2019)", + "32. Hu, E.J., Shen, Y., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., Chen, W.: Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685 (2021)", + "33. Jia, B., Liu, Y., Huang, S.: Unsupervised object-centric learning with bi-level optimized query slot attention. arXiv preprint arXiv:2210.08990 (2022)", + "34. Jia, C., Yang, Y., Xia, Y., Chen, Y.T., Parekh, Z., Pham, H., Le, Q., Sung, Y.H., Li, Z., Duerig, T.: Scaling up visual and vision-language representation learning with noisy text supervision. In: International Conference on Machine Learning. pp. 4904-4916. PMLR (2021)", + "35. Jia, M., Tang, L., Chen, B.C., Cardie, C., Belongie, S., Hariharan, B., Lim, S.N.: Visual prompt tuning. In: Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXXIII. pp. 709-727. Springer (2022)", + "36. Ju, C., Han, T., Zheng, K., Zhang, Y., Xie, W.: Prompting visual-language models for efficient video understanding. In: Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXXV. pp. 105-124. Springer (2022)", + "37. Kabra, R., Zoran, D., Erdogan, G., Matthew, L., Creswell, A., Botvinick, M., Lerchner, A., Burgess, C.: Simone: View-invariant, temporally-abstracted object representations via unsupervised video decomposition. Advances in Neural Information Processing Systems 34, 20146-20159 (2021)", + "38. Kay, W., Carreira, J., Simonyan, K., Zhang, B., Hillier, C., Vijayanarasimhan, S., Viola, F., Green, T., Back, T., Natsev, P., et al.: The kinetics human action video dataset. arXiv preprint arXiv:1705.06950 (2017)", + "39. Kipf, T., Elsayed, G.F., Mahendran, A., Stone, A., Sabour, S., Heigold, G., Jonschkowski, R., Dosovitskiy, A., Greff, K.: Conditional object-centric learning from video. In: International Conference on Learning Representations (2022), https://openreview.net/forum?id=aD7uesX1GF_", + "40. Lafferty, J., McCallum, A., Pereira, F.C.: Conditional random fields: Probabilistic models for segmenting and labeling sequence data (2001)", + "41. Lester, B., Al-Rfou, R., Constant, N.: The power of scale for parameter-efficient prompt tuning. arXiv preprint arXiv:2104.08691 (2021)", + "42. Li, X.L., Liang, P.: Prefix-tuning: Optimizing continuous prompts for generation. arXiv preprint arXiv:2101.00190 (2021)", + "43. Li, Y., Wu, C.Y., Fan, H., Mangalam, K., Xiong, B., Malik, J., Feichtenhofer, C.: Mvitv2: Improved multiscale vision transformers for classification and detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4804-4814 (2022)", + "44. Liang, F., Wu, B., Dai, X., Li, K., Zhao, Y., Zhang, H., Zhang, P., Vajda, P., Marculescu, D.: Open-vocabulary semantic segmentation with mask-adapted clip. arXiv preprint arXiv:2210.04150 (2022)" + ], + "bbox": [ + 212, + 146, + 787, + 840 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "Object-centric Image-to-Video Adaptation", + "bbox": [ + 447, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "45. Lin, Z., Wu, Y.F., Peri, S.V., Sun, W., Singh, G., Deng, F., Jiang, J., Ahn, S.: Space: Unsupervised object-oriented scene representation via spatial attention and decomposition. arXiv preprint arXiv:2001.02407 (2020)", + "46. Lin, Z., Geng, S., Zhang, R., Gao, P., De Melo, G., Wang, X., Dai, J., Qiao, Y., Li, H.: Frozen clip models are efficient video learners. In: European Conference on Computer Vision. pp. 388-404. Springer (2022)", + "47. Liu, R., Huang, J., Li, G., Feng, J., Wu, X., Li, T.H.: Revisiting temporal modeling for clip-based image-to-video knowledge transferring. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 6555-6564 (June 2023)", + "48. Liu, X., Ji, K., Fu, Y., Tam, W.L., Du, Z., Yang, Z., Tang, J.: P-tuning v2: Prompt tuning can be comparable to fine-tuning universally across scales and tasks. arXiv preprint arXiv:2110.07602 (2021)", + "49. Liu, Z., Lin, Y., Cao, Y., Hu, H., Wei, Y., Zhang, Z., Lin, S., Guo, B.: Swin transformer: Hierarchical vision transformer using shifted windows. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 10012-10022 (2021)", + "50. Liu, Z., Ning, J., Cao, Y., Wei, Y., Zhang, Z., Lin, S., Hu, H.: Video swim transformer. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 3202-3211 (2022)", + "51. Locatello, F., Weissenborn, D., Unterthiner, T., Mahendran, A., Heigold, G., Uszkoreit, J., Dosovitskiy, A., Kipf, T.: Object-centric learning with slot attention. Advances in Neural Information Processing Systems 33, 11525-11538 (2020)", + "52. Loshchilov, I., Hutter, F.: Decoupled weight decay regularization. In: International Conference on Learning Representations (2018)", + "53. Materzynska, J., Xiao, T., Herzig, R., Xu, H., Wang, X., Darrell, T.: Something-else: Compositional action recognition with spatial-temporal interaction networks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1049–1059 (2020)", + "54. Nagarajan, T., Grauman, K.: Attributes as operators: factorizing unseen attribute-object compositions. In: Proceedings of the European Conference on Computer Vision (ECCV). pp. 169-185 (2018)", + "55. Ni, B., Peng, H., Chen, M., Zhang, S., Meng, G., Fu, J., Xiang, S., Ling, H.: Expanding language-image pretrained models for general video recognition. In: Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part IV. pp. 1-18. Springer (2022)", + "56. Oquab, M., Darcet, T., Moutakanni, T., Vo, H., Szafraniec, M., Khalidov, V., Fernandez, P., Haziza, D., Massa, F., El-Nouby, A., et al.: Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193 (2023)", + "57. Pan, J., Lin, Z., Zhu, X., Shao, J., Li, H.: St-adapter: Parameter-efficient image-to-video transfer learning for action recognition. arXiv preprint arXiv:2206.13559 (2022)", + "58. Park, J., Lee, J., Sohn, K.: Dual-path adaptation from image to video transformers. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 2203-2213 (June 2023)", + "59. Peh, E., Parmar, P., Fernando, B.: Learning to visually connect actions and their effects. arXiv preprint arXiv:2401.10805 (2024)", + "60. Pfeiffer, J., Kamath, A., Rückle, A., Cho, K., Gurevych, I.: Adapterfusion: Nondestructive task composition for transfer learning. In: Proceedings of the 16th Conference of the European Chapter of the Association for Computational Linguistics: Main Volume. pp. 487-503 (2021)" + ], + "bbox": [ + 215, + 146, + 784, + 839 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "R. Qian et al.", + "bbox": [ + 271, + 114, + 364, + 128 + ], + "page_idx": 17 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "61. Pfeiffer, J., Rückle, A., Poth, C., Kamath, A., Vulic, I., Ruder, S., Cho, K., Gurevych, I.: Adapterhub: A framework for adapting transformers. arXiv preprint arXiv:2007.07779 (2020)", + "62. Qian, R., Ding, S., Liu, X., Lin, D.: Static and dynamic concepts for self-supervised video representation learning. In: European Conference on Computer Vision. pp. 145-164. Springer (2022)", + "63. Qian, R., Ding, S., Liu, X., Lin, D.: Semantics meets temporal correspondence: Self-supervised object-centric learning in videos. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 16675-16687 (2023)", + "64. Qian, R., Dong, X., Zhang, P., Zang, Y., Ding, S., Lin, D., Wang, J.: Streaming long video understanding with large language models. arXiv preprint arXiv:2405.16009 (2024)", + "65. Qian, R., Li, Y., Liu, H., See, J., Ding, S., Liu, X., Li, D., Lin, W.: Enhancing self-supervised video representation learning via multi-level feature optimization. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 7990-8001 (2021)", + "66. Qian, R., Meng, T., Gong, B., Yang, M.H., Wang, H., Belongie, S., Cui, Y.: Spatiotemporal contrastive video representation learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6964-6974 (2021)", + "67. Qin, G., Eisner, J.: Learning how to ask: Querying lms with mixtures of soft prompts. arXiv preprint arXiv:2104.06599 (2021)", + "68. Qing, Z., Zhang, S., Huang, Z., Zhang, Y., Gao, C., Zhao, D., Sang, N.: Disentangling spatial and temporal learning for efficient image-to-video transfer learning. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 13934-13944 (October 2023)", + "69. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021)", + "70. Shin, T., Razeghi, Y., Logan IV, R.L., Wallace, E., Singh, S.: Autoprompt: Eliciting knowledge from language models with automatically generated prompts. arXiv preprint arXiv:2010.15980 (2020)", + "71. Souček, T., Alayrac, J.B., Miech, A., Laptev, I., Sivic, J.: Look for the change: Learning object states and state-modifying actions from untrimmed web videos. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13956-13966 (2022)", + "72. Sung, Y.L., Cho, J., Bansal, M.: Vl-adapter: Parameter-efficient transfer learning for vision-and-language tasks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5227-5237 (2022)", + "73. Tong, Z., Song, Y., Wang, J., Wang, L.: Videomae: Masked autoencoders are data-efficient learners for self-supervised video pre-training. Advances in neural information processing systems 35, 10078-10093 (2022)", + "74. Tran, D., Bourdev, L., Fergus, R., Torresani, L., Paluri, M.: Learning spatiotemporal features with 3d convolutional networks. In: Proceedings of the IEEE international conference on computer vision. pp. 4489-4497 (2015)", + "75. Tran, D., Wang, H., Torresani, L., Ray, J., LeCun, Y., Paluri, M.: A closer look at spatiotemporal convolutions for action recognition. In: Proceedings of the IEEE conference on Computer Vision and Pattern Recognition. pp. 6450-6459 (2018)" + ], + "bbox": [ + 212, + 146, + 787, + 840 + ], + "page_idx": 18 + }, + { + "type": "header", + "text": "Object-centric Image-to-Video Adaptation", + "bbox": [ + 447, + 114, + 730, + 128 + ], + "page_idx": 18 + }, + { + "type": "page_number", + "text": "19", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 18 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "76. Ventura, C., Bellver, M., Girbau, A., Salvador, A., Marques, F., Giro-i Nieto, X.: Rvos: End-to-end recurrent network for video object segmentation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 5277-5286 (2019)", + "77. Wang, M., Xing, J., Liu, Y.: Actionclip: A new paradigm for video action recognition. arXiv preprint arXiv:2109.08472 (2021)", + "78. Wang, W., Bao, H., Dong, L., Bjorck, J., Peng, Z., Liu, Q., Aggarwal, K., Mohammed, O.K., Singhal, S., Som, S., et al.: Image as a foreign language: Beit pretraining for vision and vision-language tasks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 19175-19186 (2023)", + "79. Wang, X., Farhadi, A., Gupta, A.: Actions~ transformations. In: Proceedings of the IEEE conference on Computer Vision and Pattern Recognition. pp. 2658-2667 (2016)", + "80. Wang, X., Gupta, A.: Videos as space-time region graphs. In: Proceedings of the European conference on computer vision (ECCV). pp. 399-417 (2018)", + "81. Wang, X., Misra, I., Zeng, Z., Girdhar, R., Darrell, T.: Videocutler: Surprisingly simple unsupervised video instance segmentation. arXiv preprint arXiv:2308.14710 (2023)", + "82. Xie, J., Xie, W., Zisserman, A.: Segmenting moving objects via an object-centric layered representation. In: Advances in Neural Information Processing Systems (2022)", + "83. Xu, J., De Mello, S., Liu, S., Byeon, W., Breuel, T., Kautz, J., Wang, X.: Groupvit: Semantic segmentation emerges from text supervision. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18134-18144 (2022)", + "84. Yan, S., Xiong, X., Arnab, A., Lu, Z., Zhang, M., Sun, C., Schmid, C.: Multiview transformers for video recognition. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3333-3343 (2022)", + "85. Yang, C., Lamdouar, H., Lu, E., Zisserman, A., Xie, W.: Self-supervised video object segmentation by motion grouping. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 7177-7188 (October 2021)", + "86. Yang, T., Zhu, Y., Xie, Y., Zhang, A., Chen, C., Li, M.: AIM: Adapting image models for efficient video action recognition. In: The Eleventh International Conference on Learning Representations (2023), https://openreview.net/forum?id=CIoSZ_HKHS7", + "87. Yuan, L., Chen, D., Chen, Y.L., Codella, N., Dai, X., Gao, J., Hu, H., Huang, X., Li, B., Li, C., et al.: Florence: A new foundation model for computer vision. arXiv preprint arXiv:2111.11432 (2021)", + "88. Zadaianchuk, A., Kleindessner, M., Zhu, Y., Locatello, F., Brox, T.: Unsupervised semantic segmentation with self-supervised object-centric representations. arXiv preprint arXiv:2207.05027 (2022)", + "89. Zaken, E.B., Ravfogel, S., Goldberg, Y.: Bitfit: Simple parameter-efficient fine-tuning for transformer-based masked language-models. arXiv preprint arXiv:2106.10199 (2021)", + "90. Zareian, A., Rosa, K.D., Hu, D.H., Chang, S.F.: Open-vocabulary object detection using captions. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14393-14402 (2021)", + "91. Zhang, C., Fu, C., Wang, S., Agarwal, N., Lee, K., Choi, C., Sun, C.: Object-centric video representation for long-term action anticipation. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 6751-6761 (2024)" + ], + "bbox": [ + 215, + 146, + 785, + 839 + ], + "page_idx": 19 + }, + { + "type": "page_number", + "text": "20", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 19 + }, + { + "type": "header", + "text": "R. Qian et al.", + "bbox": [ + 271, + 114, + 364, + 128 + ], + "page_idx": 19 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "92. Zhang, C., Gupta, A., Zisserman, A.: Is an object-centric video representation beneficial for transfer? In: Proceedings of the Asian Conference on Computer Vision. pp. 1976-1994 (2022)", + "93. Zhang, Y., Tokmakov, P., Hebert, M., Schmid, C.: A structured model for action detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9975-9984 (2019)", + "94. Zhou, J., Wei, C., Wang, H., Shen, W., Xie, C., Yuille, A., Kong, T.: ibot: Image bert pre-training with online tokenizer. arXiv preprint arXiv:2111.07832 (2021)", + "95. Zhou, K., Yang, J., Loy, C.C., Liu, Z.: Conditional prompt learning for vision-language models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16816-16825 (2022)", + "96. Zhou, K., Yang, J., Loy, C.C., Liu, Z.: Learning to prompt for vision-language models. International Journal of Computer Vision 130(9), 2337-2348 (2022)" + ], + "bbox": [ + 212, + 146, + 787, + 327 + ], + "page_idx": 20 + }, + { + "type": "header", + "text": "Object-centric Image-to-Video Adaptation", + "bbox": [ + 447, + 114, + 730, + 128 + ], + "page_idx": 20 + }, + { + "type": "page_number", + "text": "21", + "bbox": [ + 767, + 116, + 782, + 126 + ], + "page_idx": 20 + } +] \ No newline at end of file diff --git a/2024/Rethinking Image-to-Video Adaptation_ An Object-centric Perspective/17973f40-dece-4cfa-8ceb-59145fcea296_model.json b/2024/Rethinking Image-to-Video Adaptation_ An Object-centric Perspective/17973f40-dece-4cfa-8ceb-59145fcea296_model.json new file mode 100644 index 0000000000000000000000000000000000000000..347ec6a8703b4a042b4e404c20ad846f0d3f1407 --- /dev/null +++ b/2024/Rethinking Image-to-Video Adaptation_ An Object-centric Perspective/17973f40-dece-4cfa-8ceb-59145fcea296_model.json @@ -0,0 +1,2706 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.275, + 0.141, + 0.73, + 0.187 + ], + "angle": 0, + "content": "Rethinking Image-to-Video Adaptation: An Object-centric Perspective" + }, + { + "type": "text", + "bbox": [ + 0.304, + 0.213, + 0.699, + 0.229 + ], + "angle": 0, + "content": "Rui Qian1, Shuangrui Ding1, and Dahua Lin1,2,3*" + }, + { + "type": "text", + "bbox": [ + 0.238, + 0.24, + 0.765, + 0.296 + ], + "angle": 0, + "content": "1 The Chinese University of Hong Kong, Hong Kong, China \n2 Shanghai Artificial Intelligence Laboratory, Shanghai, China \n3 Centre for Perceptual and Interactive Intelligence (CPII), Hong Kong, China {qr021, ds023, dhlin}@ie.cuhk.edu.hk" + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.34, + 0.744, + 0.631 + ], + "angle": 0, + "content": "Abstract. Image-to-video adaptation seeks to efficiently adapt image models for use in the video domain. Instead of finetuning the entire image backbone, many image-to-video adaptation paradigms use lightweight adapters for temporal modeling on top of the spatial module. However, these attempts are subject to limitations in efficiency and interpretability. In this paper, we propose a novel and efficient image-to-video adaptation strategy from the object-centric perspective. Inspired by human perception, which identifies objects as key components for video understanding, we integrate a proxy task of object discovery into image-to-video transfer learning. Specifically, we adopt slot attention with learnable queries to distill each frame into a compact set of object tokens. These object-centric tokens are then processed through object-time interaction layers to model object state changes across time. Integrated with two novel object-level losses, we demonstrate the feasibility of performing efficient temporal reasoning solely on the compressed object-centric representations for video downstream tasks. Our method achieves state-of-the-art performance with fewer tunable parameters, only \\(5\\%\\) of fully finetuned models and \\(50\\%\\) of efficient tuning methods, on action recognition benchmarks. In addition, our model performs favorably in zero-shot video object segmentation without further retraining or object annotations, proving the effectiveness of object-centric video understanding." + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.645, + 0.704, + 0.66 + ], + "angle": 0, + "content": "Keywords: Image-to-video Adaptation \\(\\cdot\\) Object-centric Learning" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.688, + 0.377, + 0.705 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.724, + 0.788, + 0.816 + ], + "angle": 0, + "content": "Recently, large-scale image pre-training has revolutionized the computer vision community, with the large foundation models dominating various image tasks [34, 56, 69, 78, 87, 94]. The remarkable success of these models can be attributed to the enormous amounts of image data [28, 56] or image-text pairs [34, 69, 78]. However, when we extend this discussion to the video domain, it becomes non-trivial. Acquiring large video datasets and managing high computational" + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.825, + 0.581, + 0.84 + ], + "angle": 0, + "content": "* Corresponding author. Email: dhlin@ie.cuhk.edu.hk" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.365, + 0.129 + ], + "angle": 0, + "content": "R. Qian et al." + }, + { + "type": "image", + "bbox": [ + 0.233, + 0.143, + 0.773, + 0.313 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.322, + 0.788, + 0.392 + ], + "angle": 0, + "content": "Fig. 1: Illustration of our proposed object-centric video understanding pipeline for image-to-video adaptation. Specifically, we first parse each frame into several object components (represented in different colors) to form an interaction space. Then we respectively establish inter-object interactions within each frame (represented by dotted arrows), and temporal state changes of individual objects (depicted by solid arrows)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.427, + 0.788, + 0.534 + ], + "angle": 0, + "content": "costs make training video foundation models from scratch a significant challenge. To circumvent these difficulties, the community has developed an innovative concept dubbed image-to-video adaptation. This approach aims to leverage the power of pre-trained image models and apply them to video data, enhancing both performance and efficiency. The image-to-video adaptation establishes a connection between effective methodologies in image models and the realm of video understanding, thereby opening new avenues for research and exploration." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.536, + 0.788, + 0.718 + ], + "angle": 0, + "content": "Motivated by parameter-efficient transfer learning in natural language processing [31, 32, 41, 89], a series of recent studies [36, 46, 47, 55, 57, 86] have adopted an efficient tuning approach for image-to-video adaptation. These approaches maintain the image pre-trained model in a frozen state and only finetune a minimal set of additional parameters. For example, [47, 55, 57, 86] insert learnable spatio-temporal modules into Transformer encoder blocks. However, this formulation encounters two vital limitations. Firstly, tuning the in-between parameters in the shallow layers still requires a considerable amount of GPU memory during gradient backpropagation. Secondly, the introduction of basic temporal modules, like attention layers, makes temporal reasoning on highly redundant features and lacks the essential inductive bias to effectively benefit the image-to-video adaptation, thereby presenting challenges in interpreting the training's success." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.792, + 0.843 + ], + "angle": 0, + "content": "To tackle these challenges, we propose to transfer object knowledge from frozen image pre-trained models to videos for more efficient image-to-video adaptation. In essence, our approach entails two main steps: decoupling each frame into distinct object components, and subsequently identifying object state changes across time to facilitate comprehensive video comprehension as illustrated in Fig. 1. Specifically, we first pass the video frames through the image pre-trained model to obtain frame-wise features. Then, inspired by recent works on object discovery [51,82,83], we employ slot attention [51] with learnable queries [85] to" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.448, + 0.115, + 0.732, + 0.13 + ], + "angle": 0, + "content": "Object-centric Image-to-Video Adaptation" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.298 + ], + "angle": 0, + "content": "parse each frame into a compact set of object tokens, serving as compressed representations for each frame. Thereafter, we feed these object tokens into object-time interaction layers to model the object state changes, which vividly depict the temporal dynamics in the video. Finally, we apply a linear head to the latent features representing these object state changes to produce the video-level prediction for action recognition. Unlike the existing works on object-centric representations, our method does not rely on object annotations [30,80,92], extra detectors [88] or additional modalities [91]. Instead, we develop two simple object-level losses to assist in distilling effective object-centric representations from image foundation models and establishing meaningful object state changes." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.299, + 0.788, + 0.494 + ], + "angle": 0, + "content": "In this way, our method effectively mitigates the limitations of the previous work. Firstly, our method exhibits a notable enhancement in memory efficiency compared to conventional approaches. By exclusively finetuning the additional parameters on top of the image pre-trained model, we successfully reduce the memory required for gradient backpropagation. Additionally, compressing the video frames into object tokens significantly diminishes the computation requirements for temporal modeling, making our method more accessible and scalable. Secondly, by utilizing object-centric reasoning in the form of learnable object tokens and establishing the temporal state changes of individual objects, we inject strong inductive bias into the learning process that goes beyond basic temporal modules like attention layers. We demonstrate that our model can localize the specific objects in the video as a side-product, providing an insightful interpretation of the image-to-video adaptation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.496, + 0.788, + 0.618 + ], + "angle": 0, + "content": "To summarize, our contributions are as follows. (1) We propose an efficient object-centric method to adapt image pre-trained models to the video domain. By parsing video frames into object tokens and establishing object state changes across time, we enhance temporal perception with reduced computation redundancy. (2) We achieve state-of-the-art performance on action recognition, presenting higher efficiency than existing efficient tuning counterparts. (3) Our method effectively discovers different objects in videos without object-level annotations and achieves robust zero-shot video object segmentation." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.64, + 0.39, + 0.657 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Video action recognition is a fundamental problem in computer vision. Compared to image tasks, the extra temporal relationships between different frames are crucial for video understanding. To this end, prevalent methods build on existing image models [20, 29, 49], and make architectural modifications to enable temporal modeling [2, 10, 15, 16, 19, 21, 45, 50, 62, 64, 65, 74, 75]. These works either fully finetune the whole model [2, 5, 10, 21] or directly train on large-scale video data from scratch [66, 73], requiring huge computation cost. More recently, some works propose to recognize actions from the perspective of state changes [1, 54, 59, 71, 79]. This novel idea regards actions as transformations that make changes to the objects and environments. However, most of these works directly model the state changes on the representations of the entire scene, where" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.365, + 0.129 + ], + "angle": 0, + "content": "R. Qian et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.208 + ], + "angle": 0, + "content": "different semantics and objects are highly entangled. It is non-trivial to accurately perceive the state changes of different objects. In this work, we explicitly decouple different objects from the video scene, and respectively establish the state changes on individual objects to assist action recognition." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.221, + 0.789, + 0.51 + ], + "angle": 0, + "content": "Object-centric video representation learning aims at establishing robust representations for different objects to assist video understanding. To achieve this goal, a series of works borrow the idea from object discovery and employ iterative slot attention [51] to decompose different objects [3, 6, 12, 17, 18, 37, 39, 63, 85]. However, they are specifically trained for segmentation and can only deal with low-level segmentation tasks and lack the ability for high-level understanding. Another line of works uses the object-related features as a complement to the original video representations to enhance object awareness [24, 30, 53, 92, 93]. These methods achieve promising results on multiple action benchmarks [25, 26, 53], but still have two major limitations. First, they depend on bounding box annotations [30, 80, 92], additional detectors or modalities [88, 91] to identify different objects, restricting the scalability to unseen data. Second, the identified objects provide plentiful cues for video analysis, yet these methods still fuse the disentangled object representations with the original video features, imposing limitations on efficiency. In contrast, our process is free of object annotations or handcrafted priors but only adopts simple contrastive loss to distill general object knowledge from image foundation models and decouple object components in video frames. Then we directly model temporal state changes on compact object representations, enabling more intuitive and efficient video understanding." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.523, + 0.789, + 0.841 + ], + "angle": 0, + "content": "Efficient tuning becomes a tending direction with the development of large foundation models [7,14,69,94]. It is proposed in natural language processing to reduce the trainable parameters and improve efficiency when transferring pretrained models to downstream tasks [31,32,41,89]. Common strategies include inserting task-specific adapters into Transformer encoders [31,60,61], attaching prompt tokens to the input [42,48,67,70] and learning low-rank approximations [32]. Recently, the efficient tuning is increasingly explored in computer vision [4,11,22,35,44,72,95,96]. By freezing the image pre-trained model, [11,72] tune adapters with learnable parameters for efficient adaptation. [4,22,35,44] develop prompt tuning to adapt the frozen image model to various image tasks. Later, [36,46,47,55,57,58,68,77,86] take one step further to adapt image models to videos with temporal dynamics. In order to equip the model with temporal perception ability, [47,57,86] insert spatio-temporal modules into Transformer encoder blocks, but the tunable parameters in shallow Transformer layers require huge GPU memory in gradient backpropagation. [58,68] require two pathways to process spatial and temporal information. [36,46] attach spatio-temporal attention or motion decoders to the frame-wise features, but there exists high redundancy in attention computation [5,73]. Contrarily, our method retains the backbone architecture without modifications, distills object knowledge from image foundation models, and captures temporal state changes of discovered objects for video downstream tasks." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.448, + 0.115, + 0.733, + 0.13 + ], + "angle": 0, + "content": "Object-centric Image-to-Video Adaptation" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "image", + "bbox": [ + 0.217, + 0.144, + 0.782, + 0.254 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.267, + 0.788, + 0.366 + ], + "angle": 0, + "content": "Fig. 2: An overview of our object-centric image-to-video adaptation framework. We use a frozen image pre-trained model to extract frame-wise features and pass them through a lightweight temporal fusion block. Then we employ slot attention with learnable queries to decompose each frame into a compact set of object tokens. Thereafter, we develop object-time interaction layers to establish inter-object interactions and build temporal state changes of individual objects, which are then pooled into the video-level feature for action recognition." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.392, + 0.331, + 0.409 + ], + "angle": 0, + "content": "3 Method" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.423, + 0.788, + 0.545 + ], + "angle": 0, + "content": "The overall architecture of our object-centric image-to-video adaptation is shown in Fig. 2. We use a frozen pre-trained Vision Transformer [20] as the backbone to extract frame-wise features from the video clip. Then we apply slot attention with learnable query tokens to identify different object components and pass the object tokens through object-time interaction layers to model the temporal state changes of individual objects. The obtained object state change vectors encode rich semantics and essential temporal dynamics in the given video, and are subsequently pooled into the final video-level feature for downstream tasks." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.566, + 0.448, + 0.581 + ], + "angle": 0, + "content": "3.1 Object Decomposition" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.591, + 0.787, + 0.636 + ], + "angle": 0, + "content": "To start with, given a video clip \\( \\pmb{x} = \\{x_{1}, x_{2}, \\dots, x_{T}\\} \\) consisting of \\( T \\) frames, we use a frozen image pre-trained Vision Transformer \\( \\Phi_{\\mathrm{enc}} \\) as the backbone to extract frame-wise features:" + }, + { + "type": "equation", + "bbox": [ + 0.338, + 0.645, + 0.786, + 0.665 + ], + "angle": 0, + "content": "\\[\np _ {t}, f _ {t} = \\Phi_ {\\mathrm {e n c}} (x _ {t}), \\quad p _ {t} \\in \\mathbb {R} ^ {D}, \\quad f _ {t} \\in \\mathbb {R} ^ {H W \\times D}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.788, + 0.75 + ], + "angle": 0, + "content": "where \\( p_t \\) and \\( f_t \\) denote the output c1s token and feature map from the last Transformer block. \\( H, W, D \\) respectively denote height, width, and channel dimension. These image pre-trained features encode rich but highly-entangled semantics. And the next step is to parse these frame features into different object components which are crucial for video perception and understanding [24,30,80,92]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Inspired by recent works on unsupervised object discovery and segmentation [51,82,83], we develop a variant of slot attention on top of the feature map \\( f_{t} \\) for object decomposition in each frame. To be specific, instead of random sampling from a prior distribution, we use \\( N \\) learnable query tokens \\( q \\in \\mathbb{R}^{N \\times D} \\) to initialize the slot vectors \\( S \\in \\mathbb{R}^{N \\times D} \\). The intuition of this design is to encourage each query token to capture specific semantics or concepts [33, 83] and produce" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.365, + 0.129 + ], + "angle": 0, + "content": "R. Qian et al." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.147, + 0.788, + 0.239 + ], + "angle": 0, + "content": "temporally coherent object decompositions, which provide essential inductive bias for temporal reasoning in later stages. Then at each slot attention iteration, following [51], we use three learnable linear transformations to project the slot vectors \\( S \\) and frame features \\( f_{t} \\) into query, key and value, i.e., \\( Q \\in \\mathbb{R}^{N \\times D} \\), \\( K \\in \\mathbb{R}^{HW \\times D} \\) and \\( V \\in \\mathbb{R}^{HW \\times D} \\). After that, we compute the attention matrix \\( \\tilde{A} \\in \\mathbb{R}^{N \\times HW} \\) with softmax normalization along the slot dimension as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.339, + 0.262, + 0.786, + 0.3 + ], + "angle": 0, + "content": "\\[\n\\tilde {A} _ {i, j} := \\frac {e ^ {A _ {i , j}}}{\\sum_ {l = 1} ^ {N} e ^ {A _ {l , j}}}, \\quad \\text {w h e r e} \\quad A := \\frac {1}{\\sqrt {D}} Q K ^ {T}. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.309, + 0.787, + 0.34 + ], + "angle": 0, + "content": "We use the weighted mean to aggregate the value and pass them through Gated Recurrent Unit (GRU) to update the slot vectors:" + }, + { + "type": "equation", + "bbox": [ + 0.244, + 0.352, + 0.786, + 0.391 + ], + "angle": 0, + "content": "\\[\nS := \\operatorname {G R U} (\\text {i n p u t s} = M V, \\text {s t a t e s} = S), \\quad \\text {w h e r e} \\quad M _ {i, j} := \\frac {\\tilde {A} _ {i , j}}{\\sum_ {l = 1} ^ {H W} \\tilde {A} _ {i , l}}. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.399, + 0.787, + 0.476 + ], + "angle": 0, + "content": "We iterate the routing process three times, and take the final slot attention weights as potential object segmentation masks. The slot vectors from the final iteration are treated as the object tokens that distill object knowledge from the pre-trained image features. Note that we perform slot attention \\(\\Phi_{\\mathrm{sa}}\\) on each frame in parallel, and denote the object tokens of each time stamp as:" + }, + { + "type": "equation", + "bbox": [ + 0.373, + 0.486, + 0.786, + 0.505 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {o} = \\left\\{o _ {1} \\oplus o _ {2} \\oplus \\dots \\oplus o _ {T} \\right\\} \\in \\mathbb {R} ^ {T \\times N \\times D}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.517, + 0.787, + 0.578 + ], + "angle": 0, + "content": "where \\(o_t = \\Phi_{\\mathrm{sa}}(q, \\tilde{f}_t) \\in \\mathbb{R}^{N \\times D}\\), and \\(\\oplus\\) denotes concatenation operation along time dimension. Since this attention mechanism introduces competition among slots, the obtained object tokens \\(o_t\\) are expected to take over distinct semantic parts and separate each frame into \\(N\\) different components." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.58, + 0.788, + 0.701 + ], + "angle": 0, + "content": "Surprisingly, we find that the distilled tokens and corresponding attention matrices perform remarkably well in zero-shot video object segmentation tasks. They are even comparable to some pixel-level supervised counterparts, as demonstrated in Section 4.3. This result verifies that our method effectively extracts robust object-centric features for temporal reasoning in the later stage. Note that since videos often contain partial object semantics, there tend to be some slots that carry background information and interact with the foreground object tokens for comprehensive video analysis." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.723, + 0.466, + 0.739 + ], + "angle": 0, + "content": "3.2 Object-Time Interaction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.843 + ], + "angle": 0, + "content": "After obtaining the object tokens that reveal potential object semantics, we use them as compressed representations of video frames and further model the temporal dynamics from an object-centric perspective [30, 93]. Our intuition is that in continuous frames, the object tokens, \\( o_{t}^{n} \\in \\mathbb{R}^{D}, t = 1,2,\\dots,T \\), with the same index \\( n \\) tend to represent the same objects across time as illustrated in Fig. 1. Hence, it is feasible to temporally track each object and conclude the" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.448, + 0.115, + 0.731, + 0.129 + ], + "angle": 0, + "content": "Object-centric Image-to-Video Adaptation" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.177 + ], + "angle": 0, + "content": "object state changes based on the learned object tokens, thereby assisting the process of video understanding." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.178, + 0.785, + 0.207 + ], + "angle": 0, + "content": "To achieve this goal, we first apply self-attention to the object tokens within the same frame to introduce inter-object interactions at each time stamp:" + }, + { + "type": "equation", + "bbox": [ + 0.425, + 0.217, + 0.785, + 0.234 + ], + "angle": 0, + "content": "\\[\n\\tilde {o} _ {t} = \\Phi_ {\\mathrm {o i}} \\left(o _ {t}\\right) \\in \\mathbb {R} ^ {N \\times D}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.244, + 0.784, + 0.319 + ], + "angle": 0, + "content": "where \\(\\Phi_{\\mathrm{oi}}\\) is the object interaction module instantiated by a standard Transformer encoder layer consisting of linear projections, self-attention calculation and a feed forward network. The resulting vector \\(\\tilde{\\sigma}_t^n\\) not only encodes the attribute of object \\(n\\) but also captures its relationships with contextual objects, serving as a representation for the latent state of object \\(n\\) at time stamp \\(t\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.321, + 0.785, + 0.485 + ], + "angle": 0, + "content": "Considering that the temporal state changes are a concrete manifestation of the effects of actions, they serve as an important cue for analyzing the events occurring in a video and assisting video understanding [1,59,71,79]. Hence, based on the vectors representing latent object states, we take one step further to look into multiple time stamps and explicitly establish the state changes of each object. Without loss of generality, we take the object with index \\( n \\) for illustration. Given the object states at different time stamps, we define a temporal interval \\( \\delta \\) to sample the initial and final states, i.e., \\( \\tilde{o}_t^n \\) as the initial state and \\( \\tilde{o}_{t + \\delta}^n \\) as the final state. We concatenate them along channel dimension and pass them through the temporal interaction layer \\( \\Phi_{\\mathrm{ti}} \\) to model the object state change between time \\( t \\) and \\( t + \\delta \\):" + }, + { + "type": "equation", + "bbox": [ + 0.32, + 0.496, + 0.785, + 0.514 + ], + "angle": 0, + "content": "\\[\ns _ {t} ^ {n} = \\Phi_ {\\mathrm {t i}} \\left(\\left[ \\tilde {\\alpha} _ {t} ^ {n} \\oplus \\tilde {\\alpha} _ {t + \\delta} ^ {n} \\right]\\right) \\in \\mathbb {R} ^ {D}, \\quad t = \\{1, 2, \\dots , T - \\delta \\}, \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.523, + 0.784, + 0.583 + ], + "angle": 0, + "content": "where \\(\\oplus\\) denotes channel-wise concatenation. We instantiate \\(\\Phi_{\\mathrm{ti}}\\) as a multi-layer perceptron of channel dimension \\(2D - D - D\\). In this way, for each object, we obtain \\(T' = T - \\delta\\) state change vectors. The state change matrix covering all objects and time stamps can be denoted as \\(\\boldsymbol{s} \\in \\mathbb{R}^{T' \\times N \\times D}\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.585, + 0.785, + 0.674 + ], + "angle": 0, + "content": "Since the state change is a continuous process within a video sequence, it is feasible to conclude the overall state change of each object throughout the entire video by simple temporal average pooling on the established matrix \\( \\mathbf{s} \\). Thereafter, we further apply average pooling on object dimension to aggregate the state change information of all objects as the final video-level representation, \\( v \\in \\mathbb{R}^D \\), for downstream action recognition." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.675, + 0.786, + 0.84 + ], + "angle": 0, + "content": "Discussion. Unlike the existing works that use object features as complements to original video features [24,30,53,92,93] or model state changes on the highly entangled scene representations [1, 59, 71, 79], we perform temporal reasoning solely on distilled object tokens. Our object-centric state change modeling reduces redundancy, mitigates ambiguity and results in more fine-grained cues for video understanding. And comparing with the prevalent image-to-video adaptation methods, they introduce dense temporal fusion operations, e.g., temporal convolutions [46, 47, 57] and attention [46, 47, 58, 68, 86], to mix multi-frame features for action recognition. In contrast, we leverage the object state changes to perceive temporal dynamics, which closely resemble human logic and provide an effective inductive bias to the model. To be specific, it does not require dense" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.365, + 0.129 + ], + "angle": 0, + "content": "R. Qian et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.195 + ], + "angle": 0, + "content": "fusion or traversal of all frames. Instead, we respectively model state changes on separated objects across sparsely sampled frames, enabling more efficient scene parsing and temporal perception as verified in Section 4.4." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.215, + 0.334, + 0.23 + ], + "angle": 0, + "content": "3.3 Training" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.241, + 0.784, + 0.315 + ], + "angle": 0, + "content": "In training, we freeze the parameters in the image pre-trained model \\(\\Phi_{\\mathrm{enc}}\\), and tune parameters that are responsible for object decomposition and temporal perception, i.e., \\(\\Phi_{\\mathrm{temp}}\\), \\(\\Phi_{\\mathrm{oi}}\\) and \\(\\Phi_{\\mathrm{ti}}\\). Besides the standard cross-entropy loss for action classification, we introduce two object-level losses to further facilitate object decomposition and temporal state change modeling." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.316, + 0.785, + 0.422 + ], + "angle": 0, + "content": "Object Distillation Loss. Considering that the image pre-trained foundation models tend to encode rich semantics in the given scene [9, 56, 69], the output c1s token \\( p_t \\) is expected to encompass the majority of the semantic components in frame \\( t \\). Hence, we use a contrastive grounding loss [23, 27, 90] to align the learned object tokens \\( o_t \\) and the c1s token \\( p_t \\), encouraging \\( o_t \\) to cover all objects in frame \\( t \\). Particularly, we first define the correspondence score between the \\( o_t \\) and \\( p_t \\) as:" + }, + { + "type": "equation", + "bbox": [ + 0.366, + 0.434, + 0.786, + 0.476 + ], + "angle": 0, + "content": "\\[\nc \\left(o _ {t}, p _ {t}\\right) = \\sum_ {n = 1} ^ {N} \\frac {e ^ {\\left\\langle o _ {t} ^ {n} , p _ {t} \\right\\rangle / \\tau}}{\\sum_ {l = 1} ^ {N} e ^ {\\left\\langle o _ {t} ^ {l} , p _ {t} \\right\\rangle / \\tau}} \\left\\langle o _ {t} ^ {n}, p _ {t} \\right\\rangle , \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.487, + 0.788, + 0.584 + ], + "angle": 0, + "content": "where \\(\\langle o_t^n,p_t\\rangle = \\frac{o_t^n p_t}{\\|o_t^n\\|\\|p_t\\|}\\) denotes cosine similarity calculation, \\(\\tau\\) is the temperature hyper-parameter. This correspondence score adaptively aggregates each object component and avoids penalizing unrelated object semantics. Thereafter, we randomly sample c1s token vectors \\(p^\\prime\\) from other videos within the minibatch to formulate a negative sample pool \\(\\mathcal{N}\\) and apply the contrastive loss to maximize the agreement between the aligned object and c1s token pairs:" + }, + { + "type": "equation", + "bbox": [ + 0.337, + 0.594, + 0.787, + 0.634 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {o b j} = - \\sum_ {t = 1} ^ {T} \\log \\frac {e ^ {c \\left(o _ {t} , p _ {t}\\right) / \\tau}}{e ^ {c \\left(o _ {t} , p _ {t}\\right) / \\tau} + \\sum_ {p ^ {\\prime} \\sim \\mathcal {N}} e ^ {c \\left(o _ {t} , p ^ {\\prime}\\right) / \\tau}}. \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.644, + 0.788, + 0.844 + ], + "angle": 0, + "content": "This self-supervised alignment objective encourages the object tokens to cover diverse semantics and distill object knowledge from the image pre-trained model. Temporal Reasoning Loss. In our architecture, we analyze temporal dynamics from the perspective of object state changes. Therefore, it is crucial to formulate a learning objective that guides the temporal interaction module \\(\\Phi_{\\mathrm{ti}}\\) to accurately capture the state changes between different time stamps. To achieve this goal, our intuition is that the same objects undergoing the same type of actions should possess consistent state changes. Conversely, different actions will result in distinct state changes for the same object, and different objects will exhibit diverse state changes under the same action. Based on this, we develop a margin loss to guide object state change modeling. In detail, given \\(s_t^n\\) as query, we define the videos within the mini-batch that have the same action labels with the current video to form positive pool \\(\\mathcal{P}\\), and those of different action labels as" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.448, + 0.115, + 0.732, + 0.13 + ], + "angle": 0, + "content": "Object-centric Image-to-Video Adaptation" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.147, + 0.784, + 0.238 + ], + "angle": 0, + "content": "negative pool \\(\\mathcal{N}\\). We sample the state change vectors of the same object and the same action, i.e., \\(\\{\\hat{s}_t^n |\\hat{s}\\in \\mathcal{P}\\}\\), as positive pairs, which should be aligned with \\(s_t^n\\). As for the negatives, we respectively sample the state change vectors of the same object but different actions, i.e., \\(\\{s_t^{l^{\\prime}n}|s^{\\prime}\\in \\mathcal{N}\\}\\), and those from the same video but of different objects, i.e., \\(\\{s_t^m,m\\neq n\\}\\), as hard negatives for discrimination. The margin loss with a margin hyper-parameter \\(\\lambda\\) is formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.235, + 0.25, + 0.785, + 0.341 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {\\text {t e m p}} = \\sum_ {t = 1} ^ {T ^ {\\prime}} \\sum_ {n = 1} ^ {N} \\left\\{\\sum_ {\\hat {s} \\in \\mathcal {P}} \\| s _ {t} ^ {n} - \\hat {s} _ {t} ^ {n} \\| _ {2} \\right. \\\\ + \\sum_ {s ^ {\\prime} \\in \\mathcal {N}} \\max \\left(\\lambda - \\| s _ {t} ^ {n} - s _ {t} ^ {\\prime n} \\| _ {2}, 0\\right) + \\sum_ {\\substack {m = 1 \\\\ m \\neq n}} ^ {N} \\max \\left(\\lambda - \\| s _ {t} ^ {n} - s _ {t} ^ {m} \\| _ {2}, 0\\right) \\}, \\tag{9} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.353, + 0.784, + 0.444 + ], + "angle": 0, + "content": "where \\( \\| \\cdot \\|_2 \\) denotes \\( \\mathcal{L}_2 \\) distance. This term enforces the model to discriminate diverse object states and capture unique dynamic effects of each action category. Overall Objectives. Besides the above object-level losses, for action recognition tasks, we apply a linear classification head on the video-level representation \\( v \\) with a supervised classification loss \\( \\mathcal{L}_{cls} \\) in the form of cross-entropy. In this way, we formulate the overall loss function as:" + }, + { + "type": "equation", + "bbox": [ + 0.413, + 0.458, + 0.785, + 0.472 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathcal {L} _ {o b j} + \\mathcal {L} _ {t e m p} + \\mathcal {L} _ {c l s}. \\tag {10}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.495, + 0.376, + 0.513 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.526, + 0.457, + 0.541 + ], + "angle": 0, + "content": "4.1 Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.551, + 0.784, + 0.688 + ], + "angle": 0, + "content": "We train and evaluate on three popular action recognition datasets: Kinetics-400 (K-400) [38], Something-Something-v2 (SSv2) [25], and Epic-Kitchens-100 (EK100) [13]. Additionally, we assess zero-shot video object segmentation on challenging multiple object segmentation dataset, DAVIS-2017-UVOS [8]. We use the frozen CLIP [69] pre-trained Vision Transformer [20] as \\(\\Phi_{\\mathrm{enc}}\\) for framewise feature extraction. In training, we crop each frame into \\(224\\times 224\\), set the number of query tokens \\(N = 8\\), and the time interval for object state change \\(\\delta = T / 4\\) in default. We use AdamW optimizer [52] with an initial learning rate \\(5\\times 10^{-4}\\) and mini-batch size 512 to update the tunable parameters." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.71, + 0.476, + 0.725 + ], + "angle": 0, + "content": "4.2 Video Action Recognition" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.784, + 0.779 + ], + "angle": 0, + "content": "In this section, we present the comparisons between our method and recent state-of-the-art on K-400, SSv2 and EK100 in terms of the number of tunable parameters, inference GFLOPs and action recognition accuracy." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.78, + 0.784, + 0.84 + ], + "angle": 0, + "content": "K-400. We first show the results on K-400 in Table 1 and conclude three observations. (1) Compared to fully finetuned models, our method achieves superior performance with significantly fewer tunable parameters. For example, our method of ViT-L/14 with only 21M tunable parameters outperforms all" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.365, + 0.129 + ], + "angle": 0, + "content": "R. Qian et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.788, + 0.188 + ], + "angle": 0, + "content": "Table 1: Results on Kinetics-400. We report the pre-train model initialization, inference GFLOPs, the number of tunable parameters (M), Top-1 and Top-5 accuracy. The 'Views' indicates Frame Number \\(\\times\\) Temporal Clips \\(\\times\\) Spatial Crops." + }, + { + "type": "table", + "bbox": [ + 0.232, + 0.2, + 0.769, + 0.507 + ], + "angle": 0, + "content": "
MethodPre-trainViewsGFLOPs TunableTop-1Top-5
Full finetuning
TimeSformer-L [5]IN-21K64 × 1 × 3714012180.794.7
VideoSwin-L [50]IN-21K32 × 4 × 3724819783.195.9
MTV-L [84]JFT32 × 4 × 31805087684.396.3
PromptCLIP A7 [36]CLIP16 × 5 × 1--76.893.5
ActionCLIP [77]CLIP32 × 10 × 31689014283.897.1
X-CLIP-L/14 [55]CLIP8 × 4 × 3789042087.197.6
Efficient tuning
EVL ViT-L/14 [46]CLIP32 × 3 × 180885987.3-
ST-Adapter ViT-B/16 [57]CLIP32 × 3 × 118217.282.796.2
STAN-self-B/16 [47]CLIP16 × 3 × 11187-84.996.8
AIM ViT-B/16 [86]CLIP32 × 3 × 124281184.796.7
DualPath ViT-B/16 [58]CLIP32 × 3 × 17101085.497.1
DiST ViT-B/16 [68]CLIP32 × 3 × 11950-85.097.0
DiST ViT-L/14 [68]CLIP32 × 3 × 18490-88.097.9
Ours ViT-B/16CLIP8 × 3 × 12816.585.897.3
Ours ViT-B/16CLIP16 × 3 × 15626.586.197.5
Ours ViT-B/16CLIP32 × 3 × 111236.586.297.5
Ours ViT-L/14CLIP32 × 3 × 150682188.598.0
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.537, + 0.788, + 0.844 + ], + "angle": 0, + "content": "the fully finetuned approaches that require over 100M tunable parameters. (2) Our method with only a visual encoder achieves better results than the multimodal methods, ActionCLIP [77] and X-CLIP [55], which employ an extra text branch to guide the finetuning process. (3) Among the efficient tuning methods [46, 47, 57, 58, 68, 86], EVL [46] attaches a motion decoder with multiple temporal convolution and attention layers for temporal modeling. AIM [86], ST-Adapter [57] and STAN [47] insert temporal blocks into intermediate CLIP encoder layers. DualPath [58] and DiST [68] introduce an additional temporal pathway to capture temporal dynamics. Our formulation models temporal relationships on decomposed objects in the form of state changes, which reduces computation redundancy but attains superior performance. For example, on ViT-B/16, our method with only an 8-frame input yields superior results to other approaches that employ a 32-frame input. On ViT-L/14, we reach the state-of-the-art results with \\(40\\%\\) fewer GFLOPs (5068 vs 8490) than DiST [68]. SSv2. When extending to SSv2 which requires stronger temporal modeling, the observations are consistent. As depicted in Table 2, our object-centric formulation achieves state-of-the-art results and presents higher efficiency than both full finetuning and efficient tuning methods. In particular, our method with ViT-L/14 surpasses the advanced MViTv2-L [43] and DiST [68] with significantly reduced GFLOPs (5068 vs 8484, 8490). Moreover, on ViT-B/16, our method" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.448, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Object-centric Image-to-Video Adaptation" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.787, + 0.173 + ], + "angle": 0, + "content": "Table 2: Results on Something-Something-v2. K-400†/K-600† indicates the model is pre-trained on both IN-21K and K-400/K-600." + }, + { + "type": "table", + "bbox": [ + 0.235, + 0.185, + 0.764, + 0.48 + ], + "angle": 0, + "content": "
MethodPre-trainViewsGFLOPsTunableTop-1Top-5
Full finetuning
TimeSformer-L [5]IN-21K64 × 1 × 1714012162.4-
MTV-B [84]IN-21K32 × 4 × 3479031067.690.4
ViViT-L/16×2 [2]K-400†16 × 4 × 31189231165.489.8
VideoSwin-B [50]K-400†32 × 1 × 19638969.692.7
MVViTv2-L (312↑) [43]K-400†32 × 1 × 3848421373.394.1
Efficient tuning
EVL ViT-L/14 [46]CLIP32 × 1 × 3964117566.7-
ST-Adapter ViT-B/16 [57]CLIP32 × 3 × 119557.269.592.6
STAN-self-B/16 [47]CLIP16 × 3 × 11376-69.592.7
AIM ViT-B/16 [86]CLIP32 × 1 × 324961469.192.2
DualPath ViT-B/16 [58]CLIP32 × 1 × 37161370.392.9
DiST ViT-B/16 [68]CLIP32 × 3 × 11950-70.992.1
DiST ViT-L/14 [68]CLIP32 × 3 × 18490-73.193.2
Ours ViT-B/16CLIP8 × 1 × 32816.571.293.1
Ours ViT-B/16CLIP16 × 1 × 35626.571.693.3
Ours ViT-B/16CLIP32 × 1 × 311236.571.893.4
Ours ViT-L/14CLIP32 × 1 × 350682173.694.3
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.507, + 0.785, + 0.597 + ], + "angle": 0, + "content": "with only 8 frames as input outperforms others using 32 frames. It is non-trivial to achieve this on such a motion-heavy dataset, which in return reveals the significance of our temporal modeling in the form of object state changes. On the one hand, it enables accurate temporal perception with sparser sampling. On the other hand, our temporal modeling on individual objects eliminates the need for dense temporal fusion, substantially reducing computational redundancy." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.598, + 0.787, + 0.718 + ], + "angle": 0, + "content": "EK100. On challenging egocentric EK100 that requires long-term object interactions, our method showcases substantial advantages over existing efficient tunning counterparts [46,57,68] in Table 3. The underlying reasons are two-fold. First, our method involves explicit object decomposition, which helps to capture the primary objects in egocentric videos and contributes to the significant improvements in nouns. Second, our temporal modeling on individual objects reduces the interference of camera motions and enforces the model to focus on the dynamics of each object, thus facilitating verb perception." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.74, + 0.573, + 0.756 + ], + "angle": 0, + "content": "4.3 Zero-shot Video Object Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.787, + 0.841 + ], + "angle": 0, + "content": "To verify whether our method understands videos in an object-centric way and interpret why the model works, we evaluate unsupervised video object segmentation in addition to action recognition. Particularly, we use the frozen ViT-B/16 backbone and adopt slot attention weights \\(\\{\\tilde{A}_i\\in \\mathbb{R}^{HW}\\}_{i = 1}^N\\) in Eq. 2 as object masks candidate without bells and whistles. We report \\(\\mathcal{J}\\& \\mathcal{F}\\) on challenging" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.365, + 0.129 + ], + "angle": 0, + "content": "R. Qian et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.193 + ], + "angle": 0, + "content": "multiple object discovery benchmark, DAVIS-2017-UVOS [8]. Note that we directly apply the model trained on K-400 or SSv2 to video object segmentation in a zero-shot manner." + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.22, + 0.78, + 0.261 + ], + "angle": 0, + "content": "Table 3: Results on Epic-Kitchens-100 Verb, Table 4: Results on multiple object Noun and Action. All compared methods take segmentation on DAVIS-2017-UVOS. an input of \\(8\\times 1\\times 3\\)" + }, + { + "type": "table", + "bbox": [ + 0.218, + 0.263, + 0.526, + 0.395 + ], + "angle": 0, + "content": "
MethodVerbNounAction
EVL ViT-B/16 [46]62.751.037.7
ST-Adapter ViT-B/16 [57]67.755.0-
DiST ViT-B/16 [68]69.558.145.8
DiST ViT-L/14 [68]70.761.648.9
Ours ViT-B/1671.160.949.0
Ours ViT-L/1473.264.551.7
" + }, + { + "type": "table", + "bbox": [ + 0.528, + 0.262, + 0.781, + 0.394 + ], + "angle": 0, + "content": "
ModelJ&FJF
OCLR [82]39.638.241.1
SMTC [63]40.536.444.6
BA [17]43.939.248.6
RVOS [76]41.236.845.7
Ours SSv241.037.644.4
Ours K-40044.742.347.1
Ours K-400 CRF49.346.552.1
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.431, + 0.788, + 0.718 + ], + "angle": 0, + "content": "Quantitative Results. We present the comparison with recent state-of-the-arts without object annotations [63,81,82] as well as a fully supervised baseline [76] (denoted in grey) in Table 4. Among the compared works without mask annotations, OCLR [82] takes optical flow as input to distinguish motion patterns of diverse objects. SMTC [63] designs two-stage slot attention to produce temporally consistent object segmentations. BA [17] also utilizes image foundation models like DION [9] to provide spatio-temporal correspondence and use clustering to predict the object segmentation masks. On the one hand, since K-400 covers more diverse scenes and objects than SSv2, it leads to better generalization ability in this zero-shot segmentation scenario. On the other hand, our formulation trained on K-400 achieves superior results to these counterparts, even surpassing the simple supervised baseline RVOS [76]. This demonstrates the feasibility of transferring object knowledge from image foundation models to the video domain, and our object-centric temporal reasoning further facilitates temporally coherent object decomposition. However, the gap to the state-of-the-art supervised method is still large because the mask annotations lead to much more accurate segmentation boundaries. A straightforward way is to apply Conditional Random Field (CRF) [40] for refinement, which leads to 4.6 points improvement on \\(\\mathcal{J}\\&\\mathcal{F}\\) score. It is worth further exploration in the future." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Qualitative Results. We also visualize some examples of the generated segmentation masks in Fig. 3, where each column presents the frames within the same video and the same color denotes the masks generated by the same slot query. Generally, our model discriminates different object semantics in videos. Each distilled object token can capture specific semantics or concepts that generalize across diverse time stamps and scenes. For example, the token marked in red focuses on humans, the blue one emphasizes quadruped animals, and the yellow one identifies two-wheeled vehicles. This temporally coherent object" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.448, + 0.115, + 0.732, + 0.13 + ], + "angle": 0, + "content": "Object-centric Image-to-Video Adaptation" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.178 + ], + "angle": 0, + "content": "decomposition provides an effective inductive bias for subsequent temporal reasoning, leading to more efficient temporal modeling." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.201, + 0.391, + 0.217 + ], + "angle": 0, + "content": "4.4 Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.228, + 0.785, + 0.273 + ], + "angle": 0, + "content": "In this section, we use the CLIP ViT-B/16 as \\(\\Phi_{\\mathrm{enc}}\\) with an input of 16 frames for all ablative experiments unless otherwise specified. More details and extra ablation studies are presented in Supplementary Materials." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.301, + 0.768, + 0.409 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.42, + 0.778, + 0.477 + ], + "angle": 0, + "content": "Fig. 3: Visualization of object decomposition in Fig. 4: Comparison on training data the form of segmentation. Each column presents efficiency. Our method demonstrates three frames in a video. The same color denotes advantages, especially with a small the objects identified by the same object token. amount of training data." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.508, + 0.788, + 0.689 + ], + "angle": 0, + "content": "Training Losses. We explore the effectiveness of our newly introduced object distillation loss \\(\\mathcal{L}_{obj}\\) and temporal reasoning loss \\(\\mathcal{L}_{temp}\\) in Table 5. We observe that \\(\\mathcal{L}_{obj}\\) leads to dramatic improvements on all three datasets, K-400, SSv2, and DAVIS-2017-UVOS. This is because \\(\\mathcal{L}_{obj}\\) offers rich semantic knowledge reference from the image foundation model to facilitate object identification. And better object decomposition further enhances high-level action recognition. As for \\(\\mathcal{L}_{temp}\\), it guides the model to capture the state changes on each object. Without this guidance, relying solely on the action classification loss is insufficient to direct the temporal module to perceive representative temporal dynamics, thus resulting in a drastic performance drop on the motion-heavy dataset SSv2. And notably, this temporal loss also leads to improvements on DAVIS, which verifies that the temporal perception in return bolsters object decomposition." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Temporal Sampling Strategy. We also study the effect of different temporal sampling strategies in training. Specifically, we vary the number of frames per clip \\( T \\) and the time interval for object state change modeling \\( \\delta \\). In Table 6, we display the Top-1 accuracy on K-400 and SSv2, as well as the computation GFLOPs excluding ViT backbone for more intuitive comparison. An exciting phenomenon is that even on SSv2 that requires strong temporal modeling, we achieve promising results with very sparse sampling, i.e., only 4 frames as input with significantly reduced computation. Our method is generally robust to the temporal sampling hyper-parameters, demonstrating the effectiveness of our temporal reasoning on compressed object tokens." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.365, + 0.129 + ], + "angle": 0, + "content": "R. Qian et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.77, + 0.187 + ], + "angle": 0, + "content": "Table 5: Ablation studies on training loss Table 6: Ablation studies on the number items \\( \\mathcal{L}_{obj} \\) and \\( \\mathcal{L}_{temp} \\). Note that \\( \\mathcal{L}_{cls} \\) is of frames \\( T \\) and sampling interval \\( \\delta \\). always applied in training." + }, + { + "type": "table", + "bbox": [ + 0.223, + 0.199, + 0.482, + 0.285 + ], + "angle": 0, + "content": "
LobjLtempK-400SSv2DAVIS
XX70.444.827.3
✓X80.355.239.5
X✓73.552.130.5
✓✓86.171.644.7
" + }, + { + "type": "table", + "bbox": [ + 0.515, + 0.186, + 0.742, + 0.285 + ], + "angle": 0, + "content": "
TδGFLOPsK-400SSv2
414.285.470.7
828.485.871.2
848.285.871.1
16218.086.171.6
16417.886.171.6
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.314, + 0.788, + 0.421 + ], + "angle": 0, + "content": "Training Data Efficiency. Fig. 4 presents K-400 action recognition performance with various amounts of training data. For fair comparison with fully fine-tuned TimeSformer [5] as well as the efficient tuning methods, ST-Adapter [57] and AIM [86], we use CLIP ViT-B/16 with an input of 8 frames. Our method exhibits superior data efficiency to existing works. And the advantage becomes more significant with less training data, demonstrating the robust capacity of our object-centric formulation to transfer image pre-trained knowledge to videos." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.443, + 0.36, + 0.459 + ], + "angle": 0, + "content": "5 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.475, + 0.787, + 0.641 + ], + "angle": 0, + "content": "In this work, we propose an object-centric formulation to adapt the image pretrained models to the video domain. Specifically, we employ a frozen image pre-trained model to extract frame-wise features and use slot attention with learnable queries to parse each frame into a compact set of object tokens. Then we apply object-time interaction to these object tokens to explicitly establish object state changes across time span. By using the distilled object tokens as the compressed representations of video frames, we have reduced computation costs and improved temporal reasoning capabilities especially under sparse sampling scenarios. Our method demonstrates state-of-the-art performance across action recognition benchmarks while realizing robust zero-shot video object segmentation, signifying a promising direction for future research in video analysis." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.642, + 0.787, + 0.763 + ], + "angle": 0, + "content": "Limitations and Social Negative Impact. Our work demonstrates that the object-centric formulation serves as a reliable compression of video frames, which reduces redundancy and enables efficient temporal perception. But in some applications like a video QA system, it would be better to distill useful information according to specific questions. Incorporating questions as a condition to slot attention might address this problem. And there is a potential negative impact in privacy concerns. This work involves processing large amounts of visual data, which could include sensitive information and raise privacy concerns." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.448, + 0.115, + 0.732, + 0.13 + ], + "angle": 0, + "content": "Object-centric Image-to-Video Adaptation" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.403, + 0.163 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.174, + 0.787, + 0.234 + ], + "angle": 0, + "content": "This work is partially supported by the National Key R&D Program of China (2022ZD0160201), Shanghai Artificial Intelligence Laboratory, the Centre for Perceptual and Interactive Intelligence (CPII) Ltd under the Innovation and Technology Commission (ITC)'s InnoHK." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.255, + 0.323, + 0.271 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.282, + 0.785, + 0.324 + ], + "angle": 0, + "content": "1. Alayrac, J.B., Miech, A., Laptev, I., Sivic, J., et al.: Multi-task learning of object states and state-modifying actions from web videos. IEEE Transactions on Pattern Analysis and Machine Intelligence (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.325, + 0.785, + 0.364 + ], + "angle": 0, + "content": "2. Arnab, A., Dehghani, M., Heigold, G., Sun, C., Lucic, M., Schmid, C.: Vivit: A video vision transformer. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 6836-6846 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.365, + 0.785, + 0.405 + ], + "angle": 0, + "content": "3. Aydemir, G., Xie, W., Guney, F.: Self-supervised object-centric learning for videos. In: Thirty-seventh Conference on Neural Information Processing Systems (2023), https://openreview.net/forum?id=919tWtJPXe" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.406, + 0.785, + 0.446 + ], + "angle": 0, + "content": "4. Bahng, H., Jahanian, A., Sankaranarayanan, S., Isola, P.: Visual prompting: Modifying pixel space to adapt pre-trained models. arXiv preprint arXiv:2203.17274 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.447, + 0.785, + 0.472 + ], + "angle": 0, + "content": "5. Bertasius, G., Wang, H., Torresani, L.: Is space-time attention all you need for video understanding? In: ICML. vol. 2, p. 4 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.473, + 0.785, + 0.514 + ], + "angle": 0, + "content": "6. Besbinar, B., Frossard, P.: Self-supervision by prediction for object discovery in videos. In: 2021 IEEE International Conference on Image Processing (ICIP). pp. 1509-1513. IEEE (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.515, + 0.785, + 0.554 + ], + "angle": 0, + "content": "7. Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J.D., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., et al.: Language models are few-shot learners. Advances in neural information processing systems 33, 1877-1901 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.555, + 0.785, + 0.595 + ], + "angle": 0, + "content": "8. Caelles, S., Pont-Tuset, J., Perazzi, F., Montes, A., Maninis, K.K., Van Gool, L.: The 2019 davis challenge on vos: Unsupervised multi-object segmentation. arXiv preprint arXiv:1905.00737 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.596, + 0.785, + 0.636 + ], + "angle": 0, + "content": "9. Caron, M., Touvron, H., Misra, I., Jégou, H., Mairal, J., Bojanowski, P., Joulin, A.: Emerging properties in self-supervised vision transformers. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 9650-9660 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.637, + 0.785, + 0.677 + ], + "angle": 0, + "content": "0. Carreira, J., Zisserman, A.: Quo vadis, action recognition? a new model and the kinetics dataset. In: proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 6299-6308 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.678, + 0.785, + 0.717 + ], + "angle": 0, + "content": "1. Chen, S., Ge, C., Tong, Z., Wang, J., Song, Y., Wang, J., Luo, P.: Adaptformer: Adapting vision transformers for scalable visual recognition. Advances in Neural Information Processing Systems 35, 16664-16678 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.718, + 0.785, + 0.758 + ], + "angle": 0, + "content": "2. Crawford, E., Pineau, J.: Exploiting spatial invariance for scalable unsupervised object tracking. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 34, pp. 3684-3692 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.759, + 0.785, + 0.799 + ], + "angle": 0, + "content": "3. Damen, D., Doughty, H., Farinella, G.M., Furnari, A., Kazakos, E., Ma, J., Moltisanti, D., Munro, J., Perrett, T., Price, W., et al.: Rescaling egocentric vision. arXiv preprint arXiv:2006.13256 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.8, + 0.785, + 0.84 + ], + "angle": 0, + "content": "4. Devlin, J., Chang, M.W., Lee, K., Toutanova, K.: Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805 (2018)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.282, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.365, + 0.129 + ], + "angle": 0, + "content": "R. Qian et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.787, + 0.204 + ], + "angle": 0, + "content": "15. Ding, S., Li, M., Yang, T., Qian, R., Xu, H., Chen, Q., Wang, J., Xiong, H.: Motion-aware contrastive video representation learning via foreground-background merging. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 9716-9726 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.205, + 0.787, + 0.245 + ], + "angle": 0, + "content": "16. Ding, S., Qian, R., Xiong, H.: Dual contrastive learning for spatio-temporal representation. In: Proceedings of the 30th ACM international conference on multimedia. pp. 5649-5658 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.246, + 0.787, + 0.285 + ], + "angle": 0, + "content": "17. Ding, S., Qian, R., Xu, H., Lin, D., Xiong, H.: Betrayed by attention: A simple yet effective approach for self-supervised video object segmentation. arXiv preprint arXiv:2311.17893 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.287, + 0.787, + 0.327 + ], + "angle": 0, + "content": "18. Ding, S., Xie, W., Chen, Y., Qian, R., Zhang, X., Xiong, H., Tian, Q.: Motion-inductive self-supervised object discovery in videos. arXiv preprint arXiv:2210.00221 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.328, + 0.787, + 0.37 + ], + "angle": 0, + "content": "19. Ding, S., Zhao, P., Zhang, X., Qian, R., Xiong, H., Tian, Q.: Prune spatiotemporal tokens by semantic-aware temporal accumulation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 16945-16956 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.371, + 0.787, + 0.424 + ], + "angle": 0, + "content": "20. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.425, + 0.787, + 0.466 + ], + "angle": 0, + "content": "21. Feichtenhofer, C., Fan, H., Malik, J., He, K.: Slowfast networks for video recognition. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 6202-6211 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.467, + 0.787, + 0.522 + ], + "angle": 0, + "content": "22. Feng, C., Zhong, Y., Jie, Z., Chu, X., Ren, H., Wei, X., Xie, W., Ma, L.: Promptdet: Towards open-vocabulary detection using uncurated images. In: Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part IX. pp. 701-717. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.523, + 0.787, + 0.577 + ], + "angle": 0, + "content": "23. Ghiasi, G., Gu, X., Cui, Y., Lin, T.Y.: Scaling open-vocabulary image segmentation with image-level labels. In: Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXXVI. pp. 540-557. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.578, + 0.787, + 0.619 + ], + "angle": 0, + "content": "24. Girdhar, R., Carreira, J., Doersch, C., Zisserman, A.: Video action transformer network. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 244-253 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.62, + 0.787, + 0.688 + ], + "angle": 0, + "content": "25. Goyal, R., Ebrahimi Kahou, S., Michalski, V., Materzynska, J., Westphal, S., Kim, H., Haenel, V., Fruend, I., Yianilos, P., Mueller-Freitag, M., et al.: The\" something something\" video database for learning and evaluating visual common sense. In: Proceedings of the IEEE international conference on computer vision. pp. 5842-5850 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.689, + 0.787, + 0.743 + ], + "angle": 0, + "content": "26. Gu, C., Sun, C., Ross, D.A., Vondrick, C., Pantofaru, C., Li, Y., Vijayanarasimhan, S., Toderici, G., Ricco, S., Sukthankar, R., et al.: Ava: A video dataset of spatiotemporally localized atomic visual actions. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 6047-6056 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.744, + 0.787, + 0.799 + ], + "angle": 0, + "content": "27. Gupta, T., Vahdat, A., Chechik, G., Yang, X., Kautz, J., Hoiem, D.: Contrastive learning for weakly supervised phrase grounding. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part III. pp. 752-768. Springer (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.799, + 0.787, + 0.841 + ], + "angle": 0, + "content": "28. He, K., Chen, X., Xie, S., Li, Y., Dollár, P., Girshick, R.: Masked autoencoders are scalable vision learners. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16000-16009 (2022)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.448, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Object-centric Image-to-Video Adaptation" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.189 + ], + "angle": 0, + "content": "29. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 770-778 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.189, + 0.788, + 0.243 + ], + "angle": 0, + "content": "30. Herzig, R., Ben-Avraham, E., Mangalam, K., Bar, A., Chechik, G., Rohrbach, A., Darrell, T., Globerson, A.: Object-region video transformers. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3148-3159 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.243, + 0.788, + 0.285 + ], + "angle": 0, + "content": "31. Houlsby, N., Giurgiu, A., Jastrzebski, S., Morrone, B., De Laroussilhe, Q., Gesmundo, A., Attariyan, M., Gelly, S.: Parameter-efficient transfer learning for nlp. In: International Conference on Machine Learning. pp. 2790-2799. PMLR (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.285, + 0.788, + 0.325 + ], + "angle": 0, + "content": "32. Hu, E.J., Shen, Y., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., Chen, W.: Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.325, + 0.788, + 0.352 + ], + "angle": 0, + "content": "33. Jia, B., Liu, Y., Huang, S.: Unsupervised object-centric learning with bi-level optimized query slot attention. arXiv preprint arXiv:2210.08990 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.352, + 0.788, + 0.405 + ], + "angle": 0, + "content": "34. Jia, C., Yang, Y., Xia, Y., Chen, Y.T., Parekh, Z., Pham, H., Le, Q., Sung, Y.H., Li, Z., Duerig, T.: Scaling up visual and vision-language representation learning with noisy text supervision. In: International Conference on Machine Learning. pp. 4904-4916. PMLR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.405, + 0.788, + 0.461 + ], + "angle": 0, + "content": "35. Jia, M., Tang, L., Chen, B.C., Cardie, C., Belongie, S., Hariharan, B., Lim, S.N.: Visual prompt tuning. In: Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXXIII. pp. 709-727. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.461, + 0.788, + 0.515 + ], + "angle": 0, + "content": "36. Ju, C., Han, T., Zheng, K., Zhang, Y., Xie, W.: Prompting visual-language models for efficient video understanding. In: Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXXV. pp. 105-124. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.515, + 0.788, + 0.57 + ], + "angle": 0, + "content": "37. Kabra, R., Zoran, D., Erdogan, G., Matthew, L., Creswell, A., Botvinick, M., Lerchner, A., Burgess, C.: Simone: View-invariant, temporally-abstracted object representations via unsupervised video decomposition. Advances in Neural Information Processing Systems 34, 20146-20159 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.569, + 0.788, + 0.61 + ], + "angle": 0, + "content": "38. Kay, W., Carreira, J., Simonyan, K., Zhang, B., Hillier, C., Vijayanarasimhan, S., Viola, F., Green, T., Back, T., Natsev, P., et al.: The kinetics human action video dataset. arXiv preprint arXiv:1705.06950 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.61, + 0.788, + 0.664 + ], + "angle": 0, + "content": "39. Kipf, T., Elsayed, G.F., Mahendran, A., Stone, A., Sabour, S., Heigold, G., Jonschkowski, R., Dosovitskiy, A., Greff, K.: Conditional object-centric learning from video. In: International Conference on Learning Representations (2022), https://openreview.net/forum?id=aD7uesX1GF_" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.664, + 0.788, + 0.692 + ], + "angle": 0, + "content": "40. Lafferty, J., McCallum, A., Pereira, F.C.: Conditional random fields: Probabilistic models for segmenting and labeling sequence data (2001)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.692, + 0.788, + 0.718 + ], + "angle": 0, + "content": "41. Lester, B., Al-Rfou, R., Constant, N.: The power of scale for parameter-efficient prompt tuning. arXiv preprint arXiv:2104.08691 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.718, + 0.788, + 0.745 + ], + "angle": 0, + "content": "42. Li, X.L., Liang, P.: Prefix-tuning: Optimizing continuous prompts for generation. arXiv preprint arXiv:2101.00190 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.745, + 0.788, + 0.799 + ], + "angle": 0, + "content": "43. Li, Y., Wu, C.Y., Fan, H., Mangalam, K., Xiong, B., Malik, J., Feichtenhofer, C.: Mvitv2: Improved multiscale vision transformers for classification and detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4804-4814 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.799, + 0.788, + 0.841 + ], + "angle": 0, + "content": "44. Liang, F., Wu, B., Dai, X., Li, K., Zhao, Y., Zhang, H., Zhang, P., Vajda, P., Marculescu, D.: Open-vocabulary semantic segmentation with mask-adapted clip. arXiv preprint arXiv:2210.04150 (2022)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.365, + 0.129 + ], + "angle": 0, + "content": "R. Qian et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.785, + 0.189 + ], + "angle": 0, + "content": "45. Lin, Z., Wu, Y.F., Peri, S.V., Sun, W., Singh, G., Deng, F., Jiang, J., Ahn, S.: Space: Unsupervised object-oriented scene representation via spatial attention and decomposition. arXiv preprint arXiv:2001.02407 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.189, + 0.785, + 0.23 + ], + "angle": 0, + "content": "46. Lin, Z., Geng, S., Zhang, R., Gao, P., De Melo, G., Wang, X., Dai, J., Qiao, Y., Li, H.: Frozen clip models are efficient video learners. In: European Conference on Computer Vision. pp. 388-404. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.23, + 0.785, + 0.285 + ], + "angle": 0, + "content": "47. Liu, R., Huang, J., Li, G., Feng, J., Wu, X., Li, T.H.: Revisiting temporal modeling for clip-based image-to-video knowledge transferring. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 6555-6564 (June 2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.285, + 0.785, + 0.325 + ], + "angle": 0, + "content": "48. Liu, X., Ji, K., Fu, Y., Tam, W.L., Du, Z., Yang, Z., Tang, J.: P-tuning v2: Prompt tuning can be comparable to fine-tuning universally across scales and tasks. arXiv preprint arXiv:2110.07602 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.325, + 0.785, + 0.379 + ], + "angle": 0, + "content": "49. Liu, Z., Lin, Y., Cao, Y., Hu, H., Wei, Y., Zhang, Z., Lin, S., Guo, B.: Swin transformer: Hierarchical vision transformer using shifted windows. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 10012-10022 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.379, + 0.785, + 0.42 + ], + "angle": 0, + "content": "50. Liu, Z., Ning, J., Cao, Y., Wei, Y., Zhang, Z., Lin, S., Hu, H.: Video swim transformer. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 3202-3211 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.42, + 0.785, + 0.461 + ], + "angle": 0, + "content": "51. Locatello, F., Weissenborn, D., Unterthiner, T., Mahendran, A., Heigold, G., Uszkoreit, J., Dosovitskiy, A., Kipf, T.: Object-centric learning with slot attention. Advances in Neural Information Processing Systems 33, 11525-11538 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.461, + 0.785, + 0.487 + ], + "angle": 0, + "content": "52. Loshchilov, I., Hutter, F.: Decoupled weight decay regularization. In: International Conference on Learning Representations (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.487, + 0.785, + 0.542 + ], + "angle": 0, + "content": "53. Materzynska, J., Xiao, T., Herzig, R., Xu, H., Wang, X., Darrell, T.: Something-else: Compositional action recognition with spatial-temporal interaction networks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1049–1059 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.542, + 0.785, + 0.583 + ], + "angle": 0, + "content": "54. Nagarajan, T., Grauman, K.: Attributes as operators: factorizing unseen attribute-object compositions. In: Proceedings of the European Conference on Computer Vision (ECCV). pp. 169-185 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.583, + 0.785, + 0.637 + ], + "angle": 0, + "content": "55. Ni, B., Peng, H., Chen, M., Zhang, S., Meng, G., Fu, J., Xiang, S., Ling, H.: Expanding language-image pretrained models for general video recognition. In: Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part IV. pp. 1-18. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.637, + 0.785, + 0.678 + ], + "angle": 0, + "content": "56. Oquab, M., Darcet, T., Moutakanni, T., Vo, H., Szafraniec, M., Khalidov, V., Fernandez, P., Haziza, D., Massa, F., El-Nouby, A., et al.: Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.678, + 0.785, + 0.718 + ], + "angle": 0, + "content": "57. Pan, J., Lin, Z., Zhu, X., Shao, J., Li, H.: St-adapter: Parameter-efficient image-to-video transfer learning for action recognition. arXiv preprint arXiv:2206.13559 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.718, + 0.785, + 0.759 + ], + "angle": 0, + "content": "58. Park, J., Lee, J., Sohn, K.: Dual-path adaptation from image to video transformers. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 2203-2213 (June 2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.759, + 0.785, + 0.786 + ], + "angle": 0, + "content": "59. Peh, E., Parmar, P., Fernando, B.: Learning to visually connect actions and their effects. arXiv preprint arXiv:2401.10805 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.786, + 0.785, + 0.84 + ], + "angle": 0, + "content": "60. Pfeiffer, J., Kamath, A., Rückle, A., Cho, K., Gurevych, I.: Adapterfusion: Nondestructive task composition for transfer learning. In: Proceedings of the 16th Conference of the European Chapter of the Association for Computational Linguistics: Main Volume. pp. 487-503 (2021)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.448, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Object-centric Image-to-Video Adaptation" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "19" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.189 + ], + "angle": 0, + "content": "61. Pfeiffer, J., Rückle, A., Poth, C., Kamath, A., Vulic, I., Ruder, S., Cho, K., Gurevych, I.: Adapterhub: A framework for adapting transformers. arXiv preprint arXiv:2007.07779 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.19, + 0.788, + 0.233 + ], + "angle": 0, + "content": "62. Qian, R., Ding, S., Liu, X., Lin, D.: Static and dynamic concepts for self-supervised video representation learning. In: European Conference on Computer Vision. pp. 145-164. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.234, + 0.788, + 0.275 + ], + "angle": 0, + "content": "63. Qian, R., Ding, S., Liu, X., Lin, D.: Semantics meets temporal correspondence: Self-supervised object-centric learning in videos. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 16675-16687 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.276, + 0.788, + 0.317 + ], + "angle": 0, + "content": "64. Qian, R., Dong, X., Zhang, P., Zang, Y., Ding, S., Lin, D., Wang, J.: Streaming long video understanding with large language models. arXiv preprint arXiv:2405.16009 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.318, + 0.788, + 0.374 + ], + "angle": 0, + "content": "65. Qian, R., Li, Y., Liu, H., See, J., Ding, S., Liu, X., Li, D., Lin, W.: Enhancing self-supervised video representation learning via multi-level feature optimization. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 7990-8001 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.375, + 0.788, + 0.43 + ], + "angle": 0, + "content": "66. Qian, R., Meng, T., Gong, B., Yang, M.H., Wang, H., Belongie, S., Cui, Y.: Spatiotemporal contrastive video representation learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6964-6974 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.431, + 0.788, + 0.459 + ], + "angle": 0, + "content": "67. Qin, G., Eisner, J.: Learning how to ask: Querying lms with mixtures of soft prompts. arXiv preprint arXiv:2104.06599 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.46, + 0.788, + 0.515 + ], + "angle": 0, + "content": "68. Qing, Z., Zhang, S., Huang, Z., Zhang, Y., Gao, C., Zhao, D., Sang, N.: Disentangling spatial and temporal learning for efficient image-to-video transfer learning. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 13934-13944 (October 2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.516, + 0.788, + 0.571 + ], + "angle": 0, + "content": "69. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.572, + 0.788, + 0.614 + ], + "angle": 0, + "content": "70. Shin, T., Razeghi, Y., Logan IV, R.L., Wallace, E., Singh, S.: Autoprompt: Eliciting knowledge from language models with automatically generated prompts. arXiv preprint arXiv:2010.15980 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.615, + 0.788, + 0.671 + ], + "angle": 0, + "content": "71. Souček, T., Alayrac, J.B., Miech, A., Laptev, I., Sivic, J.: Look for the change: Learning object states and state-modifying actions from untrimmed web videos. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13956-13966 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.671, + 0.788, + 0.713 + ], + "angle": 0, + "content": "72. Sung, Y.L., Cho, J., Bansal, M.: Vl-adapter: Parameter-efficient transfer learning for vision-and-language tasks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5227-5237 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.714, + 0.788, + 0.755 + ], + "angle": 0, + "content": "73. Tong, Z., Song, Y., Wang, J., Wang, L.: Videomae: Masked autoencoders are data-efficient learners for self-supervised video pre-training. Advances in neural information processing systems 35, 10078-10093 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.757, + 0.788, + 0.798 + ], + "angle": 0, + "content": "74. Tran, D., Bourdev, L., Fergus, R., Torresani, L., Paluri, M.: Learning spatiotemporal features with 3d convolutional networks. In: Proceedings of the IEEE international conference on computer vision. pp. 4489-4497 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.799, + 0.788, + 0.842 + ], + "angle": 0, + "content": "75. Tran, D., Wang, H., Torresani, L., Ray, J., LeCun, Y., Paluri, M.: A closer look at spatiotemporal convolutions for action recognition. In: Proceedings of the IEEE conference on Computer Vision and Pattern Recognition. pp. 6450-6459 (2018)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.842 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "20" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.365, + 0.129 + ], + "angle": 0, + "content": "R. Qian et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.787, + 0.203 + ], + "angle": 0, + "content": "76. Ventura, C., Bellver, M., Girbau, A., Salvador, A., Marques, F., Giro-i Nieto, X.: Rvos: End-to-end recurrent network for video object segmentation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 5277-5286 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.203, + 0.786, + 0.23 + ], + "angle": 0, + "content": "77. Wang, M., Xing, J., Liu, Y.: Actionclip: A new paradigm for video action recognition. arXiv preprint arXiv:2109.08472 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.23, + 0.787, + 0.284 + ], + "angle": 0, + "content": "78. Wang, W., Bao, H., Dong, L., Bjorck, J., Peng, Z., Liu, Q., Aggarwal, K., Mohammed, O.K., Singhal, S., Som, S., et al.: Image as a foreign language: Beit pretraining for vision and vision-language tasks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 19175-19186 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.284, + 0.787, + 0.325 + ], + "angle": 0, + "content": "79. Wang, X., Farhadi, A., Gupta, A.: Actions~ transformations. In: Proceedings of the IEEE conference on Computer Vision and Pattern Recognition. pp. 2658-2667 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.325, + 0.787, + 0.352 + ], + "angle": 0, + "content": "80. Wang, X., Gupta, A.: Videos as space-time region graphs. In: Proceedings of the European conference on computer vision (ECCV). pp. 399-417 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.352, + 0.787, + 0.392 + ], + "angle": 0, + "content": "81. Wang, X., Misra, I., Zeng, Z., Girdhar, R., Darrell, T.: Videocutler: Surprisingly simple unsupervised video instance segmentation. arXiv preprint arXiv:2308.14710 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.392, + 0.787, + 0.433 + ], + "angle": 0, + "content": "82. Xie, J., Xie, W., Zisserman, A.: Segmenting moving objects via an object-centric layered representation. In: Advances in Neural Information Processing Systems (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.433, + 0.787, + 0.488 + ], + "angle": 0, + "content": "83. Xu, J., De Mello, S., Liu, S., Byeon, W., Breuel, T., Kautz, J., Wang, X.: Groupvit: Semantic segmentation emerges from text supervision. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18134-18144 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.488, + 0.787, + 0.528 + ], + "angle": 0, + "content": "84. Yan, S., Xiong, X., Arnab, A., Lu, Z., Zhang, M., Sun, C., Schmid, C.: Multiview transformers for video recognition. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3333-3343 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.528, + 0.787, + 0.569 + ], + "angle": 0, + "content": "85. Yang, C., Lamdouar, H., Lu, E., Zisserman, A., Xie, W.: Self-supervised video object segmentation by motion grouping. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 7177-7188 (October 2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.569, + 0.787, + 0.622 + ], + "angle": 0, + "content": "86. Yang, T., Zhu, Y., Xie, Y., Zhang, A., Chen, C., Li, M.: AIM: Adapting image models for efficient video action recognition. In: The Eleventh International Conference on Learning Representations (2023), https://openreview.net/forum?id=CIoSZ_HKHS7" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.622, + 0.787, + 0.664 + ], + "angle": 0, + "content": "87. Yuan, L., Chen, D., Chen, Y.L., Codella, N., Dai, X., Gao, J., Hu, H., Huang, X., Li, B., Li, C., et al.: Florence: A new foundation model for computer vision. arXiv preprint arXiv:2111.11432 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.664, + 0.787, + 0.704 + ], + "angle": 0, + "content": "88. Zadaianchuk, A., Kleindessner, M., Zhu, Y., Locatello, F., Brox, T.: Unsupervised semantic segmentation with self-supervised object-centric representations. arXiv preprint arXiv:2207.05027 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.704, + 0.787, + 0.744 + ], + "angle": 0, + "content": "89. Zaken, E.B., Ravfogel, S., Goldberg, Y.: Bitfit: Simple parameter-efficient fine-tuning for transformer-based masked language-models. arXiv preprint arXiv:2106.10199 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.744, + 0.787, + 0.786 + ], + "angle": 0, + "content": "90. Zareian, A., Rosa, K.D., Hu, D.H., Chang, S.F.: Open-vocabulary object detection using captions. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14393-14402 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.786, + 0.787, + 0.84 + ], + "angle": 0, + "content": "91. Zhang, C., Fu, C., Wang, S., Agarwal, N., Lee, K., Choi, C., Sun, C.: Object-centric video representation for long-term action anticipation. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 6751-6761 (2024)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.448, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Object-centric Image-to-Video Adaptation" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.117, + 0.784, + 0.127 + ], + "angle": 0, + "content": "21" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.189 + ], + "angle": 0, + "content": "92. Zhang, C., Gupta, A., Zisserman, A.: Is an object-centric video representation beneficial for transfer? In: Proceedings of the Asian Conference on Computer Vision. pp. 1976-1994 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.19, + 0.788, + 0.232 + ], + "angle": 0, + "content": "93. Zhang, Y., Tokmakov, P., Hebert, M., Schmid, C.: A structured model for action detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9975-9984 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.232, + 0.788, + 0.259 + ], + "angle": 0, + "content": "94. Zhou, J., Wei, C., Wang, H., Shen, W., Xie, C., Yuille, A., Kong, T.: ibot: Image bert pre-training with online tokenizer. arXiv preprint arXiv:2111.07832 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.26, + 0.788, + 0.3 + ], + "angle": 0, + "content": "95. Zhou, K., Yang, J., Loy, C.C., Liu, Z.: Conditional prompt learning for vision-language models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16816-16825 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.301, + 0.788, + 0.328 + ], + "angle": 0, + "content": "96. Zhou, K., Yang, J., Loy, C.C., Liu, Z.: Learning to prompt for vision-language models. International Journal of Computer Vision 130(9), 2337-2348 (2022)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.328 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/Rethinking Image-to-Video Adaptation_ An Object-centric Perspective/17973f40-dece-4cfa-8ceb-59145fcea296_origin.pdf b/2024/Rethinking Image-to-Video Adaptation_ An Object-centric Perspective/17973f40-dece-4cfa-8ceb-59145fcea296_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..f9ca834b4106246b32a2439bee97802e730cf51d --- /dev/null +++ b/2024/Rethinking Image-to-Video Adaptation_ An Object-centric Perspective/17973f40-dece-4cfa-8ceb-59145fcea296_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:43ca8614bd5ae7e3fb9b2739731647ed9fa836e3bf9b25cb1847fc7c74e63b45 +size 2271336 diff --git a/2024/Rethinking Image-to-Video Adaptation_ An Object-centric Perspective/full.md b/2024/Rethinking Image-to-Video Adaptation_ An Object-centric Perspective/full.md new file mode 100644 index 0000000000000000000000000000000000000000..d510886bb068eca08b5f8d076ca4a7ab9df3c905 --- /dev/null +++ b/2024/Rethinking Image-to-Video Adaptation_ An Object-centric Perspective/full.md @@ -0,0 +1,317 @@ +# Rethinking Image-to-Video Adaptation: An Object-centric Perspective + +Rui Qian1, Shuangrui Ding1, and Dahua Lin1,2,3* + +1 The Chinese University of Hong Kong, Hong Kong, China +2 Shanghai Artificial Intelligence Laboratory, Shanghai, China +3 Centre for Perceptual and Interactive Intelligence (CPII), Hong Kong, China {qr021, ds023, dhlin}@ie.cuhk.edu.hk + +Abstract. Image-to-video adaptation seeks to efficiently adapt image models for use in the video domain. Instead of finetuning the entire image backbone, many image-to-video adaptation paradigms use lightweight adapters for temporal modeling on top of the spatial module. However, these attempts are subject to limitations in efficiency and interpretability. In this paper, we propose a novel and efficient image-to-video adaptation strategy from the object-centric perspective. Inspired by human perception, which identifies objects as key components for video understanding, we integrate a proxy task of object discovery into image-to-video transfer learning. Specifically, we adopt slot attention with learnable queries to distill each frame into a compact set of object tokens. These object-centric tokens are then processed through object-time interaction layers to model object state changes across time. Integrated with two novel object-level losses, we demonstrate the feasibility of performing efficient temporal reasoning solely on the compressed object-centric representations for video downstream tasks. Our method achieves state-of-the-art performance with fewer tunable parameters, only $5\%$ of fully finetuned models and $50\%$ of efficient tuning methods, on action recognition benchmarks. In addition, our model performs favorably in zero-shot video object segmentation without further retraining or object annotations, proving the effectiveness of object-centric video understanding. + +Keywords: Image-to-video Adaptation $\cdot$ Object-centric Learning + +# 1 Introduction + +Recently, large-scale image pre-training has revolutionized the computer vision community, with the large foundation models dominating various image tasks [34, 56, 69, 78, 87, 94]. The remarkable success of these models can be attributed to the enormous amounts of image data [28, 56] or image-text pairs [34, 69, 78]. However, when we extend this discussion to the video domain, it becomes non-trivial. Acquiring large video datasets and managing high computational + +![](images/ba3a23d8e4d481b67ba4e7d07107a6116f73242819352de7a69a14a14335a962.jpg) +Fig. 1: Illustration of our proposed object-centric video understanding pipeline for image-to-video adaptation. Specifically, we first parse each frame into several object components (represented in different colors) to form an interaction space. Then we respectively establish inter-object interactions within each frame (represented by dotted arrows), and temporal state changes of individual objects (depicted by solid arrows). + +costs make training video foundation models from scratch a significant challenge. To circumvent these difficulties, the community has developed an innovative concept dubbed image-to-video adaptation. This approach aims to leverage the power of pre-trained image models and apply them to video data, enhancing both performance and efficiency. The image-to-video adaptation establishes a connection between effective methodologies in image models and the realm of video understanding, thereby opening new avenues for research and exploration. + +Motivated by parameter-efficient transfer learning in natural language processing [31, 32, 41, 89], a series of recent studies [36, 46, 47, 55, 57, 86] have adopted an efficient tuning approach for image-to-video adaptation. These approaches maintain the image pre-trained model in a frozen state and only finetune a minimal set of additional parameters. For example, [47, 55, 57, 86] insert learnable spatio-temporal modules into Transformer encoder blocks. However, this formulation encounters two vital limitations. Firstly, tuning the in-between parameters in the shallow layers still requires a considerable amount of GPU memory during gradient backpropagation. Secondly, the introduction of basic temporal modules, like attention layers, makes temporal reasoning on highly redundant features and lacks the essential inductive bias to effectively benefit the image-to-video adaptation, thereby presenting challenges in interpreting the training's success. + +To tackle these challenges, we propose to transfer object knowledge from frozen image pre-trained models to videos for more efficient image-to-video adaptation. In essence, our approach entails two main steps: decoupling each frame into distinct object components, and subsequently identifying object state changes across time to facilitate comprehensive video comprehension as illustrated in Fig. 1. Specifically, we first pass the video frames through the image pre-trained model to obtain frame-wise features. Then, inspired by recent works on object discovery [51,82,83], we employ slot attention [51] with learnable queries [85] to + +parse each frame into a compact set of object tokens, serving as compressed representations for each frame. Thereafter, we feed these object tokens into object-time interaction layers to model the object state changes, which vividly depict the temporal dynamics in the video. Finally, we apply a linear head to the latent features representing these object state changes to produce the video-level prediction for action recognition. Unlike the existing works on object-centric representations, our method does not rely on object annotations [30,80,92], extra detectors [88] or additional modalities [91]. Instead, we develop two simple object-level losses to assist in distilling effective object-centric representations from image foundation models and establishing meaningful object state changes. + +In this way, our method effectively mitigates the limitations of the previous work. Firstly, our method exhibits a notable enhancement in memory efficiency compared to conventional approaches. By exclusively finetuning the additional parameters on top of the image pre-trained model, we successfully reduce the memory required for gradient backpropagation. Additionally, compressing the video frames into object tokens significantly diminishes the computation requirements for temporal modeling, making our method more accessible and scalable. Secondly, by utilizing object-centric reasoning in the form of learnable object tokens and establishing the temporal state changes of individual objects, we inject strong inductive bias into the learning process that goes beyond basic temporal modules like attention layers. We demonstrate that our model can localize the specific objects in the video as a side-product, providing an insightful interpretation of the image-to-video adaptation. + +To summarize, our contributions are as follows. (1) We propose an efficient object-centric method to adapt image pre-trained models to the video domain. By parsing video frames into object tokens and establishing object state changes across time, we enhance temporal perception with reduced computation redundancy. (2) We achieve state-of-the-art performance on action recognition, presenting higher efficiency than existing efficient tuning counterparts. (3) Our method effectively discovers different objects in videos without object-level annotations and achieves robust zero-shot video object segmentation. + +# 2 Related Work + +Video action recognition is a fundamental problem in computer vision. Compared to image tasks, the extra temporal relationships between different frames are crucial for video understanding. To this end, prevalent methods build on existing image models [20, 29, 49], and make architectural modifications to enable temporal modeling [2, 10, 15, 16, 19, 21, 45, 50, 62, 64, 65, 74, 75]. These works either fully finetune the whole model [2, 5, 10, 21] or directly train on large-scale video data from scratch [66, 73], requiring huge computation cost. More recently, some works propose to recognize actions from the perspective of state changes [1, 54, 59, 71, 79]. This novel idea regards actions as transformations that make changes to the objects and environments. However, most of these works directly model the state changes on the representations of the entire scene, where + +different semantics and objects are highly entangled. It is non-trivial to accurately perceive the state changes of different objects. In this work, we explicitly decouple different objects from the video scene, and respectively establish the state changes on individual objects to assist action recognition. + +Object-centric video representation learning aims at establishing robust representations for different objects to assist video understanding. To achieve this goal, a series of works borrow the idea from object discovery and employ iterative slot attention [51] to decompose different objects [3, 6, 12, 17, 18, 37, 39, 63, 85]. However, they are specifically trained for segmentation and can only deal with low-level segmentation tasks and lack the ability for high-level understanding. Another line of works uses the object-related features as a complement to the original video representations to enhance object awareness [24, 30, 53, 92, 93]. These methods achieve promising results on multiple action benchmarks [25, 26, 53], but still have two major limitations. First, they depend on bounding box annotations [30, 80, 92], additional detectors or modalities [88, 91] to identify different objects, restricting the scalability to unseen data. Second, the identified objects provide plentiful cues for video analysis, yet these methods still fuse the disentangled object representations with the original video features, imposing limitations on efficiency. In contrast, our process is free of object annotations or handcrafted priors but only adopts simple contrastive loss to distill general object knowledge from image foundation models and decouple object components in video frames. Then we directly model temporal state changes on compact object representations, enabling more intuitive and efficient video understanding. + +Efficient tuning becomes a tending direction with the development of large foundation models [7,14,69,94]. It is proposed in natural language processing to reduce the trainable parameters and improve efficiency when transferring pretrained models to downstream tasks [31,32,41,89]. Common strategies include inserting task-specific adapters into Transformer encoders [31,60,61], attaching prompt tokens to the input [42,48,67,70] and learning low-rank approximations [32]. Recently, the efficient tuning is increasingly explored in computer vision [4,11,22,35,44,72,95,96]. By freezing the image pre-trained model, [11,72] tune adapters with learnable parameters for efficient adaptation. [4,22,35,44] develop prompt tuning to adapt the frozen image model to various image tasks. Later, [36,46,47,55,57,58,68,77,86] take one step further to adapt image models to videos with temporal dynamics. In order to equip the model with temporal perception ability, [47,57,86] insert spatio-temporal modules into Transformer encoder blocks, but the tunable parameters in shallow Transformer layers require huge GPU memory in gradient backpropagation. [58,68] require two pathways to process spatial and temporal information. [36,46] attach spatio-temporal attention or motion decoders to the frame-wise features, but there exists high redundancy in attention computation [5,73]. Contrarily, our method retains the backbone architecture without modifications, distills object knowledge from image foundation models, and captures temporal state changes of discovered objects for video downstream tasks. + +![](images/dd7239b342f0ccbcc86a9401fad53167882a22250a0fdbf3c62af3b3fe820ab4.jpg) +Fig. 2: An overview of our object-centric image-to-video adaptation framework. We use a frozen image pre-trained model to extract frame-wise features and pass them through a lightweight temporal fusion block. Then we employ slot attention with learnable queries to decompose each frame into a compact set of object tokens. Thereafter, we develop object-time interaction layers to establish inter-object interactions and build temporal state changes of individual objects, which are then pooled into the video-level feature for action recognition. + +# 3 Method + +The overall architecture of our object-centric image-to-video adaptation is shown in Fig. 2. We use a frozen pre-trained Vision Transformer [20] as the backbone to extract frame-wise features from the video clip. Then we apply slot attention with learnable query tokens to identify different object components and pass the object tokens through object-time interaction layers to model the temporal state changes of individual objects. The obtained object state change vectors encode rich semantics and essential temporal dynamics in the given video, and are subsequently pooled into the final video-level feature for downstream tasks. + +# 3.1 Object Decomposition + +To start with, given a video clip $\pmb{x} = \{x_{1}, x_{2}, \dots, x_{T}\}$ consisting of $T$ frames, we use a frozen image pre-trained Vision Transformer $\Phi_{\mathrm{enc}}$ as the backbone to extract frame-wise features: + +$$ +p _ {t}, f _ {t} = \Phi_ {\mathrm {e n c}} (x _ {t}), \quad p _ {t} \in \mathbb {R} ^ {D}, \quad f _ {t} \in \mathbb {R} ^ {H W \times D}, \tag {1} +$$ + +where $p_t$ and $f_t$ denote the output c1s token and feature map from the last Transformer block. $H, W, D$ respectively denote height, width, and channel dimension. These image pre-trained features encode rich but highly-entangled semantics. And the next step is to parse these frame features into different object components which are crucial for video perception and understanding [24,30,80,92]. + +Inspired by recent works on unsupervised object discovery and segmentation [51,82,83], we develop a variant of slot attention on top of the feature map $f_{t}$ for object decomposition in each frame. To be specific, instead of random sampling from a prior distribution, we use $N$ learnable query tokens $q \in \mathbb{R}^{N \times D}$ to initialize the slot vectors $S \in \mathbb{R}^{N \times D}$ . The intuition of this design is to encourage each query token to capture specific semantics or concepts [33, 83] and produce + +temporally coherent object decompositions, which provide essential inductive bias for temporal reasoning in later stages. Then at each slot attention iteration, following [51], we use three learnable linear transformations to project the slot vectors $S$ and frame features $f_{t}$ into query, key and value, i.e., $Q \in \mathbb{R}^{N \times D}$ , $K \in \mathbb{R}^{HW \times D}$ and $V \in \mathbb{R}^{HW \times D}$ . After that, we compute the attention matrix $\tilde{A} \in \mathbb{R}^{N \times HW}$ with softmax normalization along the slot dimension as follows: + +$$ +\tilde {A} _ {i, j} := \frac {e ^ {A _ {i , j}}}{\sum_ {l = 1} ^ {N} e ^ {A _ {l , j}}}, \quad \text {w h e r e} \quad A := \frac {1}{\sqrt {D}} Q K ^ {T}. \tag {2} +$$ + +We use the weighted mean to aggregate the value and pass them through Gated Recurrent Unit (GRU) to update the slot vectors: + +$$ +S := \operatorname {G R U} (\text {i n p u t s} = M V, \text {s t a t e s} = S), \quad \text {w h e r e} \quad M _ {i, j} := \frac {\tilde {A} _ {i , j}}{\sum_ {l = 1} ^ {H W} \tilde {A} _ {i , l}}. \tag {3} +$$ + +We iterate the routing process three times, and take the final slot attention weights as potential object segmentation masks. The slot vectors from the final iteration are treated as the object tokens that distill object knowledge from the pre-trained image features. Note that we perform slot attention $\Phi_{\mathrm{sa}}$ on each frame in parallel, and denote the object tokens of each time stamp as: + +$$ +\boldsymbol {o} = \left\{o _ {1} \oplus o _ {2} \oplus \dots \oplus o _ {T} \right\} \in \mathbb {R} ^ {T \times N \times D}, \tag {4} +$$ + +where $o_t = \Phi_{\mathrm{sa}}(q, \tilde{f}_t) \in \mathbb{R}^{N \times D}$ , and $\oplus$ denotes concatenation operation along time dimension. Since this attention mechanism introduces competition among slots, the obtained object tokens $o_t$ are expected to take over distinct semantic parts and separate each frame into $N$ different components. + +Surprisingly, we find that the distilled tokens and corresponding attention matrices perform remarkably well in zero-shot video object segmentation tasks. They are even comparable to some pixel-level supervised counterparts, as demonstrated in Section 4.3. This result verifies that our method effectively extracts robust object-centric features for temporal reasoning in the later stage. Note that since videos often contain partial object semantics, there tend to be some slots that carry background information and interact with the foreground object tokens for comprehensive video analysis. + +# 3.2 Object-Time Interaction + +After obtaining the object tokens that reveal potential object semantics, we use them as compressed representations of video frames and further model the temporal dynamics from an object-centric perspective [30, 93]. Our intuition is that in continuous frames, the object tokens, $o_{t}^{n} \in \mathbb{R}^{D}, t = 1,2,\dots,T$ , with the same index $n$ tend to represent the same objects across time as illustrated in Fig. 1. Hence, it is feasible to temporally track each object and conclude the + +object state changes based on the learned object tokens, thereby assisting the process of video understanding. + +To achieve this goal, we first apply self-attention to the object tokens within the same frame to introduce inter-object interactions at each time stamp: + +$$ +\tilde {o} _ {t} = \Phi_ {\mathrm {o i}} \left(o _ {t}\right) \in \mathbb {R} ^ {N \times D}, \tag {5} +$$ + +where $\Phi_{\mathrm{oi}}$ is the object interaction module instantiated by a standard Transformer encoder layer consisting of linear projections, self-attention calculation and a feed forward network. The resulting vector $\tilde{\sigma}_t^n$ not only encodes the attribute of object $n$ but also captures its relationships with contextual objects, serving as a representation for the latent state of object $n$ at time stamp $t$ . + +Considering that the temporal state changes are a concrete manifestation of the effects of actions, they serve as an important cue for analyzing the events occurring in a video and assisting video understanding [1,59,71,79]. Hence, based on the vectors representing latent object states, we take one step further to look into multiple time stamps and explicitly establish the state changes of each object. Without loss of generality, we take the object with index $n$ for illustration. Given the object states at different time stamps, we define a temporal interval $\delta$ to sample the initial and final states, i.e., $\tilde{o}_t^n$ as the initial state and $\tilde{o}_{t + \delta}^n$ as the final state. We concatenate them along channel dimension and pass them through the temporal interaction layer $\Phi_{\mathrm{ti}}$ to model the object state change between time $t$ and $t + \delta$ : + +$$ +s _ {t} ^ {n} = \Phi_ {\mathrm {t i}} \left(\left[ \tilde {\alpha} _ {t} ^ {n} \oplus \tilde {\alpha} _ {t + \delta} ^ {n} \right]\right) \in \mathbb {R} ^ {D}, \quad t = \{1, 2, \dots , T - \delta \}, \tag {6} +$$ + +where $\oplus$ denotes channel-wise concatenation. We instantiate $\Phi_{\mathrm{ti}}$ as a multi-layer perceptron of channel dimension $2D - D - D$ . In this way, for each object, we obtain $T' = T - \delta$ state change vectors. The state change matrix covering all objects and time stamps can be denoted as $\boldsymbol{s} \in \mathbb{R}^{T' \times N \times D}$ . + +Since the state change is a continuous process within a video sequence, it is feasible to conclude the overall state change of each object throughout the entire video by simple temporal average pooling on the established matrix $\mathbf{s}$ . Thereafter, we further apply average pooling on object dimension to aggregate the state change information of all objects as the final video-level representation, $v \in \mathbb{R}^D$ , for downstream action recognition. + +Discussion. Unlike the existing works that use object features as complements to original video features [24,30,53,92,93] or model state changes on the highly entangled scene representations [1, 59, 71, 79], we perform temporal reasoning solely on distilled object tokens. Our object-centric state change modeling reduces redundancy, mitigates ambiguity and results in more fine-grained cues for video understanding. And comparing with the prevalent image-to-video adaptation methods, they introduce dense temporal fusion operations, e.g., temporal convolutions [46, 47, 57] and attention [46, 47, 58, 68, 86], to mix multi-frame features for action recognition. In contrast, we leverage the object state changes to perceive temporal dynamics, which closely resemble human logic and provide an effective inductive bias to the model. To be specific, it does not require dense + +fusion or traversal of all frames. Instead, we respectively model state changes on separated objects across sparsely sampled frames, enabling more efficient scene parsing and temporal perception as verified in Section 4.4. + +# 3.3 Training + +In training, we freeze the parameters in the image pre-trained model $\Phi_{\mathrm{enc}}$ , and tune parameters that are responsible for object decomposition and temporal perception, i.e., $\Phi_{\mathrm{temp}}$ , $\Phi_{\mathrm{oi}}$ and $\Phi_{\mathrm{ti}}$ . Besides the standard cross-entropy loss for action classification, we introduce two object-level losses to further facilitate object decomposition and temporal state change modeling. + +Object Distillation Loss. Considering that the image pre-trained foundation models tend to encode rich semantics in the given scene [9, 56, 69], the output c1s token $p_t$ is expected to encompass the majority of the semantic components in frame $t$ . Hence, we use a contrastive grounding loss [23, 27, 90] to align the learned object tokens $o_t$ and the c1s token $p_t$ , encouraging $o_t$ to cover all objects in frame $t$ . Particularly, we first define the correspondence score between the $o_t$ and $p_t$ as: + +$$ +c \left(o _ {t}, p _ {t}\right) = \sum_ {n = 1} ^ {N} \frac {e ^ {\left\langle o _ {t} ^ {n} , p _ {t} \right\rangle / \tau}}{\sum_ {l = 1} ^ {N} e ^ {\left\langle o _ {t} ^ {l} , p _ {t} \right\rangle / \tau}} \left\langle o _ {t} ^ {n}, p _ {t} \right\rangle , \tag {7} +$$ + +where $\langle o_t^n,p_t\rangle = \frac{o_t^n p_t}{\|o_t^n\|\|p_t\|}$ denotes cosine similarity calculation, $\tau$ is the temperature hyper-parameter. This correspondence score adaptively aggregates each object component and avoids penalizing unrelated object semantics. Thereafter, we randomly sample c1s token vectors $p^\prime$ from other videos within the minibatch to formulate a negative sample pool $\mathcal{N}$ and apply the contrastive loss to maximize the agreement between the aligned object and c1s token pairs: + +$$ +\mathcal {L} _ {o b j} = - \sum_ {t = 1} ^ {T} \log \frac {e ^ {c \left(o _ {t} , p _ {t}\right) / \tau}}{e ^ {c \left(o _ {t} , p _ {t}\right) / \tau} + \sum_ {p ^ {\prime} \sim \mathcal {N}} e ^ {c \left(o _ {t} , p ^ {\prime}\right) / \tau}}. \tag {8} +$$ + +This self-supervised alignment objective encourages the object tokens to cover diverse semantics and distill object knowledge from the image pre-trained model. Temporal Reasoning Loss. In our architecture, we analyze temporal dynamics from the perspective of object state changes. Therefore, it is crucial to formulate a learning objective that guides the temporal interaction module $\Phi_{\mathrm{ti}}$ to accurately capture the state changes between different time stamps. To achieve this goal, our intuition is that the same objects undergoing the same type of actions should possess consistent state changes. Conversely, different actions will result in distinct state changes for the same object, and different objects will exhibit diverse state changes under the same action. Based on this, we develop a margin loss to guide object state change modeling. In detail, given $s_t^n$ as query, we define the videos within the mini-batch that have the same action labels with the current video to form positive pool $\mathcal{P}$ , and those of different action labels as + +negative pool $\mathcal{N}$ . We sample the state change vectors of the same object and the same action, i.e., $\{\hat{s}_t^n |\hat{s}\in \mathcal{P}\}$ , as positive pairs, which should be aligned with $s_t^n$ . As for the negatives, we respectively sample the state change vectors of the same object but different actions, i.e., $\{s_t^{l^{\prime}n}|s^{\prime}\in \mathcal{N}\}$ , and those from the same video but of different objects, i.e., $\{s_t^m,m\neq n\}$ , as hard negatives for discrimination. The margin loss with a margin hyper-parameter $\lambda$ is formulated as: + +$$ +\begin{array}{l} \mathcal {L} _ {\text {t e m p}} = \sum_ {t = 1} ^ {T ^ {\prime}} \sum_ {n = 1} ^ {N} \left\{\sum_ {\hat {s} \in \mathcal {P}} \| s _ {t} ^ {n} - \hat {s} _ {t} ^ {n} \| _ {2} \right. \\ + \sum_ {s ^ {\prime} \in \mathcal {N}} \max \left(\lambda - \| s _ {t} ^ {n} - s _ {t} ^ {\prime n} \| _ {2}, 0\right) + \sum_ {\substack {m = 1 \\ m \neq n}} ^ {N} \max \left(\lambda - \| s _ {t} ^ {n} - s _ {t} ^ {m} \| _ {2}, 0\right) \}, \tag{9} \\ \end{array} +$$ + +where $\| \cdot \|_2$ denotes $\mathcal{L}_2$ distance. This term enforces the model to discriminate diverse object states and capture unique dynamic effects of each action category. Overall Objectives. Besides the above object-level losses, for action recognition tasks, we apply a linear classification head on the video-level representation $v$ with a supervised classification loss $\mathcal{L}_{cls}$ in the form of cross-entropy. In this way, we formulate the overall loss function as: + +$$ +\mathcal {L} = \mathcal {L} _ {o b j} + \mathcal {L} _ {t e m p} + \mathcal {L} _ {c l s}. \tag {10} +$$ + +# 4 Experiments + +# 4.1 Implementation Details + +We train and evaluate on three popular action recognition datasets: Kinetics-400 (K-400) [38], Something-Something-v2 (SSv2) [25], and Epic-Kitchens-100 (EK100) [13]. Additionally, we assess zero-shot video object segmentation on challenging multiple object segmentation dataset, DAVIS-2017-UVOS [8]. We use the frozen CLIP [69] pre-trained Vision Transformer [20] as $\Phi_{\mathrm{enc}}$ for framewise feature extraction. In training, we crop each frame into $224\times 224$ , set the number of query tokens $N = 8$ , and the time interval for object state change $\delta = T / 4$ in default. We use AdamW optimizer [52] with an initial learning rate $5\times 10^{-4}$ and mini-batch size 512 to update the tunable parameters. + +# 4.2 Video Action Recognition + +In this section, we present the comparisons between our method and recent state-of-the-art on K-400, SSv2 and EK100 in terms of the number of tunable parameters, inference GFLOPs and action recognition accuracy. + +K-400. We first show the results on K-400 in Table 1 and conclude three observations. (1) Compared to fully finetuned models, our method achieves superior performance with significantly fewer tunable parameters. For example, our method of ViT-L/14 with only 21M tunable parameters outperforms all + +Table 1: Results on Kinetics-400. We report the pre-train model initialization, inference GFLOPs, the number of tunable parameters (M), Top-1 and Top-5 accuracy. The 'Views' indicates Frame Number $\times$ Temporal Clips $\times$ Spatial Crops. + +
MethodPre-trainViewsGFLOPs TunableTop-1Top-5
Full finetuning
TimeSformer-L [5]IN-21K64 × 1 × 3714012180.794.7
VideoSwin-L [50]IN-21K32 × 4 × 3724819783.195.9
MTV-L [84]JFT32 × 4 × 31805087684.396.3
PromptCLIP A7 [36]CLIP16 × 5 × 1--76.893.5
ActionCLIP [77]CLIP32 × 10 × 31689014283.897.1
X-CLIP-L/14 [55]CLIP8 × 4 × 3789042087.197.6
Efficient tuning
EVL ViT-L/14 [46]CLIP32 × 3 × 180885987.3-
ST-Adapter ViT-B/16 [57]CLIP32 × 3 × 118217.282.796.2
STAN-self-B/16 [47]CLIP16 × 3 × 11187-84.996.8
AIM ViT-B/16 [86]CLIP32 × 3 × 124281184.796.7
DualPath ViT-B/16 [58]CLIP32 × 3 × 17101085.497.1
DiST ViT-B/16 [68]CLIP32 × 3 × 11950-85.097.0
DiST ViT-L/14 [68]CLIP32 × 3 × 18490-88.097.9
Ours ViT-B/16CLIP8 × 3 × 12816.585.897.3
Ours ViT-B/16CLIP16 × 3 × 15626.586.197.5
Ours ViT-B/16CLIP32 × 3 × 111236.586.297.5
Ours ViT-L/14CLIP32 × 3 × 150682188.598.0
+ +the fully finetuned approaches that require over 100M tunable parameters. (2) Our method with only a visual encoder achieves better results than the multimodal methods, ActionCLIP [77] and X-CLIP [55], which employ an extra text branch to guide the finetuning process. (3) Among the efficient tuning methods [46, 47, 57, 58, 68, 86], EVL [46] attaches a motion decoder with multiple temporal convolution and attention layers for temporal modeling. AIM [86], ST-Adapter [57] and STAN [47] insert temporal blocks into intermediate CLIP encoder layers. DualPath [58] and DiST [68] introduce an additional temporal pathway to capture temporal dynamics. Our formulation models temporal relationships on decomposed objects in the form of state changes, which reduces computation redundancy but attains superior performance. For example, on ViT-B/16, our method with only an 8-frame input yields superior results to other approaches that employ a 32-frame input. On ViT-L/14, we reach the state-of-the-art results with $40\%$ fewer GFLOPs (5068 vs 8490) than DiST [68]. SSv2. When extending to SSv2 which requires stronger temporal modeling, the observations are consistent. As depicted in Table 2, our object-centric formulation achieves state-of-the-art results and presents higher efficiency than both full finetuning and efficient tuning methods. In particular, our method with ViT-L/14 surpasses the advanced MViTv2-L [43] and DiST [68] with significantly reduced GFLOPs (5068 vs 8484, 8490). Moreover, on ViT-B/16, our method + +Table 2: Results on Something-Something-v2. K-400†/K-600† indicates the model is pre-trained on both IN-21K and K-400/K-600. + +
MethodPre-trainViewsGFLOPsTunableTop-1Top-5
Full finetuning
TimeSformer-L [5]IN-21K64 × 1 × 1714012162.4-
MTV-B [84]IN-21K32 × 4 × 3479031067.690.4
ViViT-L/16×2 [2]K-400†16 × 4 × 31189231165.489.8
VideoSwin-B [50]K-400†32 × 1 × 19638969.692.7
MVViTv2-L (312↑) [43]K-400†32 × 1 × 3848421373.394.1
Efficient tuning
EVL ViT-L/14 [46]CLIP32 × 1 × 3964117566.7-
ST-Adapter ViT-B/16 [57]CLIP32 × 3 × 119557.269.592.6
STAN-self-B/16 [47]CLIP16 × 3 × 11376-69.592.7
AIM ViT-B/16 [86]CLIP32 × 1 × 324961469.192.2
DualPath ViT-B/16 [58]CLIP32 × 1 × 37161370.392.9
DiST ViT-B/16 [68]CLIP32 × 3 × 11950-70.992.1
DiST ViT-L/14 [68]CLIP32 × 3 × 18490-73.193.2
Ours ViT-B/16CLIP8 × 1 × 32816.571.293.1
Ours ViT-B/16CLIP16 × 1 × 35626.571.693.3
Ours ViT-B/16CLIP32 × 1 × 311236.571.893.4
Ours ViT-L/14CLIP32 × 1 × 350682173.694.3
+ +with only 8 frames as input outperforms others using 32 frames. It is non-trivial to achieve this on such a motion-heavy dataset, which in return reveals the significance of our temporal modeling in the form of object state changes. On the one hand, it enables accurate temporal perception with sparser sampling. On the other hand, our temporal modeling on individual objects eliminates the need for dense temporal fusion, substantially reducing computational redundancy. + +EK100. On challenging egocentric EK100 that requires long-term object interactions, our method showcases substantial advantages over existing efficient tunning counterparts [46,57,68] in Table 3. The underlying reasons are two-fold. First, our method involves explicit object decomposition, which helps to capture the primary objects in egocentric videos and contributes to the significant improvements in nouns. Second, our temporal modeling on individual objects reduces the interference of camera motions and enforces the model to focus on the dynamics of each object, thus facilitating verb perception. + +# 4.3 Zero-shot Video Object Segmentation + +To verify whether our method understands videos in an object-centric way and interpret why the model works, we evaluate unsupervised video object segmentation in addition to action recognition. Particularly, we use the frozen ViT-B/16 backbone and adopt slot attention weights $\{\tilde{A}_i\in \mathbb{R}^{HW}\}_{i = 1}^N$ in Eq. 2 as object masks candidate without bells and whistles. We report $\mathcal{J}\& \mathcal{F}$ on challenging + +multiple object discovery benchmark, DAVIS-2017-UVOS [8]. Note that we directly apply the model trained on K-400 or SSv2 to video object segmentation in a zero-shot manner. + +
MethodVerbNounAction
EVL ViT-B/16 [46]62.751.037.7
ST-Adapter ViT-B/16 [57]67.755.0-
DiST ViT-B/16 [68]69.558.145.8
DiST ViT-L/14 [68]70.761.648.9
Ours ViT-B/1671.160.949.0
Ours ViT-L/1473.264.551.7
+ +Table 3: Results on Epic-Kitchens-100 Verb, Table 4: Results on multiple object Noun and Action. All compared methods take segmentation on DAVIS-2017-UVOS. an input of $8\times 1\times 3$ + +
ModelJ&FJF
OCLR [82]39.638.241.1
SMTC [63]40.536.444.6
BA [17]43.939.248.6
RVOS [76]41.236.845.7
Ours SSv241.037.644.4
Ours K-40044.742.347.1
Ours K-400 CRF49.346.552.1
+ +Quantitative Results. We present the comparison with recent state-of-the-arts without object annotations [63,81,82] as well as a fully supervised baseline [76] (denoted in grey) in Table 4. Among the compared works without mask annotations, OCLR [82] takes optical flow as input to distinguish motion patterns of diverse objects. SMTC [63] designs two-stage slot attention to produce temporally consistent object segmentations. BA [17] also utilizes image foundation models like DION [9] to provide spatio-temporal correspondence and use clustering to predict the object segmentation masks. On the one hand, since K-400 covers more diverse scenes and objects than SSv2, it leads to better generalization ability in this zero-shot segmentation scenario. On the other hand, our formulation trained on K-400 achieves superior results to these counterparts, even surpassing the simple supervised baseline RVOS [76]. This demonstrates the feasibility of transferring object knowledge from image foundation models to the video domain, and our object-centric temporal reasoning further facilitates temporally coherent object decomposition. However, the gap to the state-of-the-art supervised method is still large because the mask annotations lead to much more accurate segmentation boundaries. A straightforward way is to apply Conditional Random Field (CRF) [40] for refinement, which leads to 4.6 points improvement on $\mathcal{J}\&\mathcal{F}$ score. It is worth further exploration in the future. + +Qualitative Results. We also visualize some examples of the generated segmentation masks in Fig. 3, where each column presents the frames within the same video and the same color denotes the masks generated by the same slot query. Generally, our model discriminates different object semantics in videos. Each distilled object token can capture specific semantics or concepts that generalize across diverse time stamps and scenes. For example, the token marked in red focuses on humans, the blue one emphasizes quadruped animals, and the yellow one identifies two-wheeled vehicles. This temporally coherent object + +decomposition provides an effective inductive bias for subsequent temporal reasoning, leading to more efficient temporal modeling. + +# 4.4 Ablation Study + +In this section, we use the CLIP ViT-B/16 as $\Phi_{\mathrm{enc}}$ with an input of 16 frames for all ablative experiments unless otherwise specified. More details and extra ablation studies are presented in Supplementary Materials. + +![](images/d83ba3e1789da07a8fb77c42e0ba8d4bc92b5f9c55a1da1d0287ab2e4ebbe5e5.jpg) +Fig. 3: Visualization of object decomposition in Fig. 4: Comparison on training data the form of segmentation. Each column presents efficiency. Our method demonstrates three frames in a video. The same color denotes advantages, especially with a small the objects identified by the same object token. amount of training data. + +Training Losses. We explore the effectiveness of our newly introduced object distillation loss $\mathcal{L}_{obj}$ and temporal reasoning loss $\mathcal{L}_{temp}$ in Table 5. We observe that $\mathcal{L}_{obj}$ leads to dramatic improvements on all three datasets, K-400, SSv2, and DAVIS-2017-UVOS. This is because $\mathcal{L}_{obj}$ offers rich semantic knowledge reference from the image foundation model to facilitate object identification. And better object decomposition further enhances high-level action recognition. As for $\mathcal{L}_{temp}$ , it guides the model to capture the state changes on each object. Without this guidance, relying solely on the action classification loss is insufficient to direct the temporal module to perceive representative temporal dynamics, thus resulting in a drastic performance drop on the motion-heavy dataset SSv2. And notably, this temporal loss also leads to improvements on DAVIS, which verifies that the temporal perception in return bolsters object decomposition. + +Temporal Sampling Strategy. We also study the effect of different temporal sampling strategies in training. Specifically, we vary the number of frames per clip $T$ and the time interval for object state change modeling $\delta$ . In Table 6, we display the Top-1 accuracy on K-400 and SSv2, as well as the computation GFLOPs excluding ViT backbone for more intuitive comparison. An exciting phenomenon is that even on SSv2 that requires strong temporal modeling, we achieve promising results with very sparse sampling, i.e., only 4 frames as input with significantly reduced computation. Our method is generally robust to the temporal sampling hyper-parameters, demonstrating the effectiveness of our temporal reasoning on compressed object tokens. + +
LobjLtempK-400SSv2DAVIS
XX70.444.827.3
✓X80.355.239.5
X✓73.552.130.5
✓✓86.171.644.7
+ +Table 5: Ablation studies on training loss Table 6: Ablation studies on the number items $\mathcal{L}_{obj}$ and $\mathcal{L}_{temp}$ . Note that $\mathcal{L}_{cls}$ is of frames $T$ and sampling interval $\delta$ . always applied in training. + +
TδGFLOPsK-400SSv2
414.285.470.7
828.485.871.2
848.285.871.1
16218.086.171.6
16417.886.171.6
+ +Training Data Efficiency. Fig. 4 presents K-400 action recognition performance with various amounts of training data. For fair comparison with fully fine-tuned TimeSformer [5] as well as the efficient tuning methods, ST-Adapter [57] and AIM [86], we use CLIP ViT-B/16 with an input of 8 frames. Our method exhibits superior data efficiency to existing works. And the advantage becomes more significant with less training data, demonstrating the robust capacity of our object-centric formulation to transfer image pre-trained knowledge to videos. + +# 5 Conclusion + +In this work, we propose an object-centric formulation to adapt the image pretrained models to the video domain. Specifically, we employ a frozen image pre-trained model to extract frame-wise features and use slot attention with learnable queries to parse each frame into a compact set of object tokens. Then we apply object-time interaction to these object tokens to explicitly establish object state changes across time span. By using the distilled object tokens as the compressed representations of video frames, we have reduced computation costs and improved temporal reasoning capabilities especially under sparse sampling scenarios. Our method demonstrates state-of-the-art performance across action recognition benchmarks while realizing robust zero-shot video object segmentation, signifying a promising direction for future research in video analysis. + +Limitations and Social Negative Impact. Our work demonstrates that the object-centric formulation serves as a reliable compression of video frames, which reduces redundancy and enables efficient temporal perception. But in some applications like a video QA system, it would be better to distill useful information according to specific questions. Incorporating questions as a condition to slot attention might address this problem. And there is a potential negative impact in privacy concerns. This work involves processing large amounts of visual data, which could include sensitive information and raise privacy concerns. + +# Acknowledgements + +This work is partially supported by the National Key R&D Program of China (2022ZD0160201), Shanghai Artificial Intelligence Laboratory, the Centre for Perceptual and Interactive Intelligence (CPII) Ltd under the Innovation and Technology Commission (ITC)'s InnoHK. + +# References + +1. Alayrac, J.B., Miech, A., Laptev, I., Sivic, J., et al.: Multi-task learning of object states and state-modifying actions from web videos. IEEE Transactions on Pattern Analysis and Machine Intelligence (2024) +2. Arnab, A., Dehghani, M., Heigold, G., Sun, C., Lucic, M., Schmid, C.: Vivit: A video vision transformer. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 6836-6846 (2021) +3. Aydemir, G., Xie, W., Guney, F.: Self-supervised object-centric learning for videos. In: Thirty-seventh Conference on Neural Information Processing Systems (2023), https://openreview.net/forum?id=919tWtJPXe +4. Bahng, H., Jahanian, A., Sankaranarayanan, S., Isola, P.: Visual prompting: Modifying pixel space to adapt pre-trained models. arXiv preprint arXiv:2203.17274 (2022) +5. Bertasius, G., Wang, H., Torresani, L.: Is space-time attention all you need for video understanding? In: ICML. vol. 2, p. 4 (2021) +6. Besbinar, B., Frossard, P.: Self-supervision by prediction for object discovery in videos. In: 2021 IEEE International Conference on Image Processing (ICIP). pp. 1509-1513. IEEE (2021) +7. Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J.D., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., et al.: Language models are few-shot learners. Advances in neural information processing systems 33, 1877-1901 (2020) +8. Caelles, S., Pont-Tuset, J., Perazzi, F., Montes, A., Maninis, K.K., Van Gool, L.: The 2019 davis challenge on vos: Unsupervised multi-object segmentation. arXiv preprint arXiv:1905.00737 (2019) +9. Caron, M., Touvron, H., Misra, I., Jégou, H., Mairal, J., Bojanowski, P., Joulin, A.: Emerging properties in self-supervised vision transformers. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 9650-9660 (2021) +0. Carreira, J., Zisserman, A.: Quo vadis, action recognition? a new model and the kinetics dataset. In: proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 6299-6308 (2017) +1. Chen, S., Ge, C., Tong, Z., Wang, J., Song, Y., Wang, J., Luo, P.: Adaptformer: Adapting vision transformers for scalable visual recognition. Advances in Neural Information Processing Systems 35, 16664-16678 (2022) +2. Crawford, E., Pineau, J.: Exploiting spatial invariance for scalable unsupervised object tracking. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 34, pp. 3684-3692 (2020) +3. Damen, D., Doughty, H., Farinella, G.M., Furnari, A., Kazakos, E., Ma, J., Moltisanti, D., Munro, J., Perrett, T., Price, W., et al.: Rescaling egocentric vision. arXiv preprint arXiv:2006.13256 (2020) +4. Devlin, J., Chang, M.W., Lee, K., Toutanova, K.: Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805 (2018) + +15. Ding, S., Li, M., Yang, T., Qian, R., Xu, H., Chen, Q., Wang, J., Xiong, H.: Motion-aware contrastive video representation learning via foreground-background merging. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 9716-9726 (2022) +16. Ding, S., Qian, R., Xiong, H.: Dual contrastive learning for spatio-temporal representation. In: Proceedings of the 30th ACM international conference on multimedia. pp. 5649-5658 (2022) +17. Ding, S., Qian, R., Xu, H., Lin, D., Xiong, H.: Betrayed by attention: A simple yet effective approach for self-supervised video object segmentation. arXiv preprint arXiv:2311.17893 (2023) +18. Ding, S., Xie, W., Chen, Y., Qian, R., Zhang, X., Xiong, H., Tian, Q.: Motion-inductive self-supervised object discovery in videos. arXiv preprint arXiv:2210.00221 (2022) +19. Ding, S., Zhao, P., Zhang, X., Qian, R., Xiong, H., Tian, Q.: Prune spatiotemporal tokens by semantic-aware temporal accumulation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 16945-16956 (2023) +20. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929 (2020) +21. Feichtenhofer, C., Fan, H., Malik, J., He, K.: Slowfast networks for video recognition. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 6202-6211 (2019) +22. Feng, C., Zhong, Y., Jie, Z., Chu, X., Ren, H., Wei, X., Xie, W., Ma, L.: Promptdet: Towards open-vocabulary detection using uncurated images. In: Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part IX. pp. 701-717. Springer (2022) +23. Ghiasi, G., Gu, X., Cui, Y., Lin, T.Y.: Scaling open-vocabulary image segmentation with image-level labels. In: Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXXVI. pp. 540-557. Springer (2022) +24. Girdhar, R., Carreira, J., Doersch, C., Zisserman, A.: Video action transformer network. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 244-253 (2019) +25. Goyal, R., Ebrahimi Kahou, S., Michalski, V., Materzynska, J., Westphal, S., Kim, H., Haenel, V., Fruend, I., Yianilos, P., Mueller-Freitag, M., et al.: The" something something" video database for learning and evaluating visual common sense. In: Proceedings of the IEEE international conference on computer vision. pp. 5842-5850 (2017) +26. Gu, C., Sun, C., Ross, D.A., Vondrick, C., Pantofaru, C., Li, Y., Vijayanarasimhan, S., Toderici, G., Ricco, S., Sukthankar, R., et al.: Ava: A video dataset of spatiotemporally localized atomic visual actions. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 6047-6056 (2018) +27. Gupta, T., Vahdat, A., Chechik, G., Yang, X., Kautz, J., Hoiem, D.: Contrastive learning for weakly supervised phrase grounding. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part III. pp. 752-768. Springer (2020) +28. He, K., Chen, X., Xie, S., Li, Y., Dollár, P., Girshick, R.: Masked autoencoders are scalable vision learners. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16000-16009 (2022) + +29. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 770-778 (2016) +30. Herzig, R., Ben-Avraham, E., Mangalam, K., Bar, A., Chechik, G., Rohrbach, A., Darrell, T., Globerson, A.: Object-region video transformers. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3148-3159 (2022) +31. Houlsby, N., Giurgiu, A., Jastrzebski, S., Morrone, B., De Laroussilhe, Q., Gesmundo, A., Attariyan, M., Gelly, S.: Parameter-efficient transfer learning for nlp. In: International Conference on Machine Learning. pp. 2790-2799. PMLR (2019) +32. Hu, E.J., Shen, Y., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., Chen, W.: Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685 (2021) +33. Jia, B., Liu, Y., Huang, S.: Unsupervised object-centric learning with bi-level optimized query slot attention. arXiv preprint arXiv:2210.08990 (2022) +34. Jia, C., Yang, Y., Xia, Y., Chen, Y.T., Parekh, Z., Pham, H., Le, Q., Sung, Y.H., Li, Z., Duerig, T.: Scaling up visual and vision-language representation learning with noisy text supervision. In: International Conference on Machine Learning. pp. 4904-4916. PMLR (2021) +35. Jia, M., Tang, L., Chen, B.C., Cardie, C., Belongie, S., Hariharan, B., Lim, S.N.: Visual prompt tuning. In: Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXXIII. pp. 709-727. Springer (2022) +36. Ju, C., Han, T., Zheng, K., Zhang, Y., Xie, W.: Prompting visual-language models for efficient video understanding. In: Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXXV. pp. 105-124. Springer (2022) +37. Kabra, R., Zoran, D., Erdogan, G., Matthew, L., Creswell, A., Botvinick, M., Lerchner, A., Burgess, C.: Simone: View-invariant, temporally-abstracted object representations via unsupervised video decomposition. Advances in Neural Information Processing Systems 34, 20146-20159 (2021) +38. Kay, W., Carreira, J., Simonyan, K., Zhang, B., Hillier, C., Vijayanarasimhan, S., Viola, F., Green, T., Back, T., Natsev, P., et al.: The kinetics human action video dataset. arXiv preprint arXiv:1705.06950 (2017) +39. Kipf, T., Elsayed, G.F., Mahendran, A., Stone, A., Sabour, S., Heigold, G., Jonschkowski, R., Dosovitskiy, A., Greff, K.: Conditional object-centric learning from video. In: International Conference on Learning Representations (2022), https://openreview.net/forum?id=aD7uesX1GF_ +40. Lafferty, J., McCallum, A., Pereira, F.C.: Conditional random fields: Probabilistic models for segmenting and labeling sequence data (2001) +41. Lester, B., Al-Rfou, R., Constant, N.: The power of scale for parameter-efficient prompt tuning. arXiv preprint arXiv:2104.08691 (2021) +42. Li, X.L., Liang, P.: Prefix-tuning: Optimizing continuous prompts for generation. arXiv preprint arXiv:2101.00190 (2021) +43. Li, Y., Wu, C.Y., Fan, H., Mangalam, K., Xiong, B., Malik, J., Feichtenhofer, C.: Mvitv2: Improved multiscale vision transformers for classification and detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4804-4814 (2022) +44. Liang, F., Wu, B., Dai, X., Li, K., Zhao, Y., Zhang, H., Zhang, P., Vajda, P., Marculescu, D.: Open-vocabulary semantic segmentation with mask-adapted clip. arXiv preprint arXiv:2210.04150 (2022) + +45. Lin, Z., Wu, Y.F., Peri, S.V., Sun, W., Singh, G., Deng, F., Jiang, J., Ahn, S.: Space: Unsupervised object-oriented scene representation via spatial attention and decomposition. arXiv preprint arXiv:2001.02407 (2020) +46. Lin, Z., Geng, S., Zhang, R., Gao, P., De Melo, G., Wang, X., Dai, J., Qiao, Y., Li, H.: Frozen clip models are efficient video learners. In: European Conference on Computer Vision. pp. 388-404. Springer (2022) +47. Liu, R., Huang, J., Li, G., Feng, J., Wu, X., Li, T.H.: Revisiting temporal modeling for clip-based image-to-video knowledge transferring. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 6555-6564 (June 2023) +48. Liu, X., Ji, K., Fu, Y., Tam, W.L., Du, Z., Yang, Z., Tang, J.: P-tuning v2: Prompt tuning can be comparable to fine-tuning universally across scales and tasks. arXiv preprint arXiv:2110.07602 (2021) +49. Liu, Z., Lin, Y., Cao, Y., Hu, H., Wei, Y., Zhang, Z., Lin, S., Guo, B.: Swin transformer: Hierarchical vision transformer using shifted windows. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 10012-10022 (2021) +50. Liu, Z., Ning, J., Cao, Y., Wei, Y., Zhang, Z., Lin, S., Hu, H.: Video swim transformer. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 3202-3211 (2022) +51. Locatello, F., Weissenborn, D., Unterthiner, T., Mahendran, A., Heigold, G., Uszkoreit, J., Dosovitskiy, A., Kipf, T.: Object-centric learning with slot attention. Advances in Neural Information Processing Systems 33, 11525-11538 (2020) +52. Loshchilov, I., Hutter, F.: Decoupled weight decay regularization. In: International Conference on Learning Representations (2018) +53. Materzynska, J., Xiao, T., Herzig, R., Xu, H., Wang, X., Darrell, T.: Something-else: Compositional action recognition with spatial-temporal interaction networks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1049–1059 (2020) +54. Nagarajan, T., Grauman, K.: Attributes as operators: factorizing unseen attribute-object compositions. In: Proceedings of the European Conference on Computer Vision (ECCV). pp. 169-185 (2018) +55. Ni, B., Peng, H., Chen, M., Zhang, S., Meng, G., Fu, J., Xiang, S., Ling, H.: Expanding language-image pretrained models for general video recognition. In: Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part IV. pp. 1-18. Springer (2022) +56. Oquab, M., Darcet, T., Moutakanni, T., Vo, H., Szafraniec, M., Khalidov, V., Fernandez, P., Haziza, D., Massa, F., El-Nouby, A., et al.: Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193 (2023) +57. Pan, J., Lin, Z., Zhu, X., Shao, J., Li, H.: St-adapter: Parameter-efficient image-to-video transfer learning for action recognition. arXiv preprint arXiv:2206.13559 (2022) +58. Park, J., Lee, J., Sohn, K.: Dual-path adaptation from image to video transformers. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 2203-2213 (June 2023) +59. Peh, E., Parmar, P., Fernando, B.: Learning to visually connect actions and their effects. arXiv preprint arXiv:2401.10805 (2024) +60. Pfeiffer, J., Kamath, A., Rückle, A., Cho, K., Gurevych, I.: Adapterfusion: Nondestructive task composition for transfer learning. In: Proceedings of the 16th Conference of the European Chapter of the Association for Computational Linguistics: Main Volume. pp. 487-503 (2021) + +61. Pfeiffer, J., Rückle, A., Poth, C., Kamath, A., Vulic, I., Ruder, S., Cho, K., Gurevych, I.: Adapterhub: A framework for adapting transformers. arXiv preprint arXiv:2007.07779 (2020) +62. Qian, R., Ding, S., Liu, X., Lin, D.: Static and dynamic concepts for self-supervised video representation learning. In: European Conference on Computer Vision. pp. 145-164. Springer (2022) +63. Qian, R., Ding, S., Liu, X., Lin, D.: Semantics meets temporal correspondence: Self-supervised object-centric learning in videos. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 16675-16687 (2023) +64. Qian, R., Dong, X., Zhang, P., Zang, Y., Ding, S., Lin, D., Wang, J.: Streaming long video understanding with large language models. arXiv preprint arXiv:2405.16009 (2024) +65. Qian, R., Li, Y., Liu, H., See, J., Ding, S., Liu, X., Li, D., Lin, W.: Enhancing self-supervised video representation learning via multi-level feature optimization. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 7990-8001 (2021) +66. Qian, R., Meng, T., Gong, B., Yang, M.H., Wang, H., Belongie, S., Cui, Y.: Spatiotemporal contrastive video representation learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6964-6974 (2021) +67. Qin, G., Eisner, J.: Learning how to ask: Querying lms with mixtures of soft prompts. arXiv preprint arXiv:2104.06599 (2021) +68. Qing, Z., Zhang, S., Huang, Z., Zhang, Y., Gao, C., Zhao, D., Sang, N.: Disentangling spatial and temporal learning for efficient image-to-video transfer learning. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 13934-13944 (October 2023) +69. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021) +70. Shin, T., Razeghi, Y., Logan IV, R.L., Wallace, E., Singh, S.: Autoprompt: Eliciting knowledge from language models with automatically generated prompts. arXiv preprint arXiv:2010.15980 (2020) +71. Souček, T., Alayrac, J.B., Miech, A., Laptev, I., Sivic, J.: Look for the change: Learning object states and state-modifying actions from untrimmed web videos. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13956-13966 (2022) +72. Sung, Y.L., Cho, J., Bansal, M.: Vl-adapter: Parameter-efficient transfer learning for vision-and-language tasks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5227-5237 (2022) +73. Tong, Z., Song, Y., Wang, J., Wang, L.: Videomae: Masked autoencoders are data-efficient learners for self-supervised video pre-training. Advances in neural information processing systems 35, 10078-10093 (2022) +74. Tran, D., Bourdev, L., Fergus, R., Torresani, L., Paluri, M.: Learning spatiotemporal features with 3d convolutional networks. In: Proceedings of the IEEE international conference on computer vision. pp. 4489-4497 (2015) +75. Tran, D., Wang, H., Torresani, L., Ray, J., LeCun, Y., Paluri, M.: A closer look at spatiotemporal convolutions for action recognition. In: Proceedings of the IEEE conference on Computer Vision and Pattern Recognition. pp. 6450-6459 (2018) + +76. Ventura, C., Bellver, M., Girbau, A., Salvador, A., Marques, F., Giro-i Nieto, X.: Rvos: End-to-end recurrent network for video object segmentation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 5277-5286 (2019) +77. Wang, M., Xing, J., Liu, Y.: Actionclip: A new paradigm for video action recognition. arXiv preprint arXiv:2109.08472 (2021) +78. Wang, W., Bao, H., Dong, L., Bjorck, J., Peng, Z., Liu, Q., Aggarwal, K., Mohammed, O.K., Singhal, S., Som, S., et al.: Image as a foreign language: Beit pretraining for vision and vision-language tasks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 19175-19186 (2023) +79. Wang, X., Farhadi, A., Gupta, A.: Actions~ transformations. In: Proceedings of the IEEE conference on Computer Vision and Pattern Recognition. pp. 2658-2667 (2016) +80. Wang, X., Gupta, A.: Videos as space-time region graphs. In: Proceedings of the European conference on computer vision (ECCV). pp. 399-417 (2018) +81. Wang, X., Misra, I., Zeng, Z., Girdhar, R., Darrell, T.: Videocutler: Surprisingly simple unsupervised video instance segmentation. arXiv preprint arXiv:2308.14710 (2023) +82. Xie, J., Xie, W., Zisserman, A.: Segmenting moving objects via an object-centric layered representation. In: Advances in Neural Information Processing Systems (2022) +83. Xu, J., De Mello, S., Liu, S., Byeon, W., Breuel, T., Kautz, J., Wang, X.: Groupvit: Semantic segmentation emerges from text supervision. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18134-18144 (2022) +84. Yan, S., Xiong, X., Arnab, A., Lu, Z., Zhang, M., Sun, C., Schmid, C.: Multiview transformers for video recognition. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3333-3343 (2022) +85. Yang, C., Lamdouar, H., Lu, E., Zisserman, A., Xie, W.: Self-supervised video object segmentation by motion grouping. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 7177-7188 (October 2021) +86. Yang, T., Zhu, Y., Xie, Y., Zhang, A., Chen, C., Li, M.: AIM: Adapting image models for efficient video action recognition. In: The Eleventh International Conference on Learning Representations (2023), https://openreview.net/forum?id=CIoSZ_HKHS7 +87. Yuan, L., Chen, D., Chen, Y.L., Codella, N., Dai, X., Gao, J., Hu, H., Huang, X., Li, B., Li, C., et al.: Florence: A new foundation model for computer vision. arXiv preprint arXiv:2111.11432 (2021) +88. Zadaianchuk, A., Kleindessner, M., Zhu, Y., Locatello, F., Brox, T.: Unsupervised semantic segmentation with self-supervised object-centric representations. arXiv preprint arXiv:2207.05027 (2022) +89. Zaken, E.B., Ravfogel, S., Goldberg, Y.: Bitfit: Simple parameter-efficient fine-tuning for transformer-based masked language-models. arXiv preprint arXiv:2106.10199 (2021) +90. Zareian, A., Rosa, K.D., Hu, D.H., Chang, S.F.: Open-vocabulary object detection using captions. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14393-14402 (2021) +91. Zhang, C., Fu, C., Wang, S., Agarwal, N., Lee, K., Choi, C., Sun, C.: Object-centric video representation for long-term action anticipation. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 6751-6761 (2024) + +92. Zhang, C., Gupta, A., Zisserman, A.: Is an object-centric video representation beneficial for transfer? In: Proceedings of the Asian Conference on Computer Vision. pp. 1976-1994 (2022) +93. Zhang, Y., Tokmakov, P., Hebert, M., Schmid, C.: A structured model for action detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9975-9984 (2019) +94. Zhou, J., Wei, C., Wang, H., Shen, W., Xie, C., Yuille, A., Kong, T.: ibot: Image bert pre-training with online tokenizer. arXiv preprint arXiv:2111.07832 (2021) +95. Zhou, K., Yang, J., Loy, C.C., Liu, Z.: Conditional prompt learning for vision-language models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16816-16825 (2022) +96. Zhou, K., Yang, J., Loy, C.C., Liu, Z.: Learning to prompt for vision-language models. International Journal of Computer Vision 130(9), 2337-2348 (2022) \ No newline at end of file diff --git a/2024/Rethinking Image-to-Video Adaptation_ An Object-centric Perspective/images.zip b/2024/Rethinking Image-to-Video Adaptation_ An Object-centric Perspective/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..b3a77390baefcb0722b17865f754d1128a325231 --- /dev/null +++ b/2024/Rethinking Image-to-Video Adaptation_ An Object-centric Perspective/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:35d9927dd43e87161d21958a7572d10233d082d45ac082bc4d9a6fa547af26ac +size 478364 diff --git a/2024/Rethinking Image-to-Video Adaptation_ An Object-centric Perspective/layout.json b/2024/Rethinking Image-to-Video Adaptation_ An Object-centric Perspective/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..060ab99a7cace7ff988f1901a091cbafc0fd93d6 --- /dev/null +++ b/2024/Rethinking Image-to-Video Adaptation_ An Object-centric Perspective/layout.json @@ -0,0 +1,10184 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 168, + 111, + 446, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 168, + 111, + 446, + 148 + ], + "spans": [ + { + "bbox": [ + 168, + 111, + 446, + 148 + ], + "type": "text", + "content": "Rethinking Image-to-Video Adaptation: An Object-centric Perspective" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 186, + 168, + 427, + 181 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 186, + 168, + 427, + 181 + ], + "spans": [ + { + "bbox": [ + 186, + 168, + 427, + 181 + ], + "type": "text", + "content": "Rui Qian1, Shuangrui Ding1, and Dahua Lin1,2,3*" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 145, + 190, + 468, + 234 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 190, + 468, + 234 + ], + "spans": [ + { + "bbox": [ + 145, + 190, + 468, + 234 + ], + "type": "text", + "content": "1 The Chinese University of Hong Kong, Hong Kong, China \n2 Shanghai Artificial Intelligence Laboratory, Shanghai, China \n3 Centre for Perceptual and Interactive Intelligence (CPII), Hong Kong, China {qr021, ds023, dhlin}@ie.cuhk.edu.hk" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 160, + 269, + 455, + 499 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 269, + 455, + 499 + ], + "spans": [ + { + "bbox": [ + 160, + 269, + 455, + 499 + ], + "type": "text", + "content": "Abstract. Image-to-video adaptation seeks to efficiently adapt image models for use in the video domain. Instead of finetuning the entire image backbone, many image-to-video adaptation paradigms use lightweight adapters for temporal modeling on top of the spatial module. However, these attempts are subject to limitations in efficiency and interpretability. In this paper, we propose a novel and efficient image-to-video adaptation strategy from the object-centric perspective. Inspired by human perception, which identifies objects as key components for video understanding, we integrate a proxy task of object discovery into image-to-video transfer learning. Specifically, we adopt slot attention with learnable queries to distill each frame into a compact set of object tokens. These object-centric tokens are then processed through object-time interaction layers to model object state changes across time. Integrated with two novel object-level losses, we demonstrate the feasibility of performing efficient temporal reasoning solely on the compressed object-centric representations for video downstream tasks. Our method achieves state-of-the-art performance with fewer tunable parameters, only " + }, + { + "bbox": [ + 160, + 269, + 455, + 499 + ], + "type": "inline_equation", + "content": "5\\%" + }, + { + "bbox": [ + 160, + 269, + 455, + 499 + ], + "type": "text", + "content": " of fully finetuned models and " + }, + { + "bbox": [ + 160, + 269, + 455, + 499 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 160, + 269, + 455, + 499 + ], + "type": "text", + "content": " of efficient tuning methods, on action recognition benchmarks. In addition, our model performs favorably in zero-shot video object segmentation without further retraining or object annotations, proving the effectiveness of object-centric video understanding." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 160, + 510, + 430, + 522 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 510, + 430, + 522 + ], + "spans": [ + { + "bbox": [ + 160, + 510, + 430, + 522 + ], + "type": "text", + "content": "Keywords: Image-to-video Adaptation " + }, + { + "bbox": [ + 160, + 510, + 430, + 522 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 510, + 430, + 522 + ], + "type": "text", + "content": " Object-centric Learning" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 544, + 230, + 558 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 544, + 230, + 558 + ], + "spans": [ + { + "bbox": [ + 132, + 544, + 230, + 558 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 573, + 482, + 646 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 573, + 482, + 646 + ], + "spans": [ + { + "bbox": [ + 130, + 573, + 482, + 646 + ], + "type": "text", + "content": "Recently, large-scale image pre-training has revolutionized the computer vision community, with the large foundation models dominating various image tasks [34, 56, 69, 78, 87, 94]. The remarkable success of these models can be attributed to the enormous amounts of image data [28, 56] or image-text pairs [34, 69, 78]. However, when we extend this discussion to the video domain, it becomes non-trivial. Acquiring large video datasets and managing high computational" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 653, + 355, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 653, + 355, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 653, + 355, + 665 + ], + "type": "text", + "content": "* Corresponding author. Email: dhlin@ie.cuhk.edu.hk" + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 142, + 113, + 473, + 247 + ], + "blocks": [ + { + "bbox": [ + 142, + 113, + 473, + 247 + ], + "lines": [ + { + "bbox": [ + 142, + 113, + 473, + 247 + ], + "spans": [ + { + "bbox": [ + 142, + 113, + 473, + 247 + ], + "type": "image", + "image_path": "ba3a23d8e4d481b67ba4e7d07107a6116f73242819352de7a69a14a14335a962.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 255, + 482, + 310 + ], + "lines": [ + { + "bbox": [ + 130, + 255, + 482, + 310 + ], + "spans": [ + { + "bbox": [ + 130, + 255, + 482, + 310 + ], + "type": "text", + "content": "Fig. 1: Illustration of our proposed object-centric video understanding pipeline for image-to-video adaptation. Specifically, we first parse each frame into several object components (represented in different colors) to form an interaction space. Then we respectively establish inter-object interactions within each frame (represented by dotted arrows), and temporal state changes of individual objects (depicted by solid arrows)." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 338, + 482, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 338, + 482, + 422 + ], + "spans": [ + { + "bbox": [ + 130, + 338, + 482, + 422 + ], + "type": "text", + "content": "costs make training video foundation models from scratch a significant challenge. To circumvent these difficulties, the community has developed an innovative concept dubbed image-to-video adaptation. This approach aims to leverage the power of pre-trained image models and apply them to video data, enhancing both performance and efficiency. The image-to-video adaptation establishes a connection between effective methodologies in image models and the realm of video understanding, thereby opening new avenues for research and exploration." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 424, + 482, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 424, + 482, + 568 + ], + "spans": [ + { + "bbox": [ + 130, + 424, + 482, + 568 + ], + "type": "text", + "content": "Motivated by parameter-efficient transfer learning in natural language processing [31, 32, 41, 89], a series of recent studies [36, 46, 47, 55, 57, 86] have adopted an efficient tuning approach for image-to-video adaptation. These approaches maintain the image pre-trained model in a frozen state and only finetune a minimal set of additional parameters. For example, [47, 55, 57, 86] insert learnable spatio-temporal modules into Transformer encoder blocks. However, this formulation encounters two vital limitations. Firstly, tuning the in-between parameters in the shallow layers still requires a considerable amount of GPU memory during gradient backpropagation. Secondly, the introduction of basic temporal modules, like attention layers, makes temporal reasoning on highly redundant features and lacks the essential inductive bias to effectively benefit the image-to-video adaptation, thereby presenting challenges in interpreting the training's success." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 570, + 484, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 484, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 484, + 667 + ], + "type": "text", + "content": "To tackle these challenges, we propose to transfer object knowledge from frozen image pre-trained models to videos for more efficient image-to-video adaptation. In essence, our approach entails two main steps: decoupling each frame into distinct object components, and subsequently identifying object state changes across time to facilitate comprehensive video comprehension as illustrated in Fig. 1. Specifically, we first pass the video frames through the image pre-trained model to obtain frame-wise features. Then, inspired by recent works on object discovery [51,82,83], we employ slot attention [51] with learnable queries [85] to" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "type": "text", + "content": "R. Qian et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "text", + "content": "parse each frame into a compact set of object tokens, serving as compressed representations for each frame. Thereafter, we feed these object tokens into object-time interaction layers to model the object state changes, which vividly depict the temporal dynamics in the video. Finally, we apply a linear head to the latent features representing these object state changes to produce the video-level prediction for action recognition. Unlike the existing works on object-centric representations, our method does not rely on object annotations [30,80,92], extra detectors [88] or additional modalities [91]. Instead, we develop two simple object-level losses to assist in distilling effective object-centric representations from image foundation models and establishing meaningful object state changes." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 236, + 482, + 391 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 236, + 482, + 391 + ], + "spans": [ + { + "bbox": [ + 130, + 236, + 482, + 391 + ], + "type": "text", + "content": "In this way, our method effectively mitigates the limitations of the previous work. Firstly, our method exhibits a notable enhancement in memory efficiency compared to conventional approaches. By exclusively finetuning the additional parameters on top of the image pre-trained model, we successfully reduce the memory required for gradient backpropagation. Additionally, compressing the video frames into object tokens significantly diminishes the computation requirements for temporal modeling, making our method more accessible and scalable. Secondly, by utilizing object-centric reasoning in the form of learnable object tokens and establishing the temporal state changes of individual objects, we inject strong inductive bias into the learning process that goes beyond basic temporal modules like attention layers. We demonstrate that our model can localize the specific objects in the video as a side-product, providing an insightful interpretation of the image-to-video adaptation." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 392, + 482, + 489 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 392, + 482, + 489 + ], + "spans": [ + { + "bbox": [ + 130, + 392, + 482, + 489 + ], + "type": "text", + "content": "To summarize, our contributions are as follows. (1) We propose an efficient object-centric method to adapt image pre-trained models to the video domain. By parsing video frames into object tokens and establishing object state changes across time, we enhance temporal perception with reduced computation redundancy. (2) We achieve state-of-the-art performance on action recognition, presenting higher efficiency than existing efficient tuning counterparts. (3) Our method effectively discovers different objects in videos without object-level annotations and achieves robust zero-shot video object segmentation." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 506, + 238, + 520 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 506, + 238, + 520 + ], + "spans": [ + { + "bbox": [ + 132, + 506, + 238, + 520 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": "Video action recognition is a fundamental problem in computer vision. Compared to image tasks, the extra temporal relationships between different frames are crucial for video understanding. To this end, prevalent methods build on existing image models [20, 29, 49], and make architectural modifications to enable temporal modeling [2, 10, 15, 16, 19, 21, 45, 50, 62, 64, 65, 74, 75]. These works either fully finetune the whole model [2, 5, 10, 21] or directly train on large-scale video data from scratch [66, 73], requiring huge computation cost. More recently, some works propose to recognize actions from the perspective of state changes [1, 54, 59, 71, 79]. This novel idea regards actions as transformations that make changes to the objects and environments. However, most of these works directly model the state changes on the representations of the entire scene, where" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 274, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 274, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 274, + 91, + 447, + 102 + ], + "type": "text", + "content": "Object-centric Image-to-Video Adaptation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 164 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 164 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 164 + ], + "type": "text", + "content": "different semantics and objects are highly entangled. It is non-trivial to accurately perceive the state changes of different objects. In this work, we explicitly decouple different objects from the video scene, and respectively establish the state changes on individual objects to assist action recognition." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 175, + 482, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 175, + 482, + 403 + ], + "spans": [ + { + "bbox": [ + 130, + 175, + 482, + 403 + ], + "type": "text", + "content": "Object-centric video representation learning aims at establishing robust representations for different objects to assist video understanding. To achieve this goal, a series of works borrow the idea from object discovery and employ iterative slot attention [51] to decompose different objects [3, 6, 12, 17, 18, 37, 39, 63, 85]. However, they are specifically trained for segmentation and can only deal with low-level segmentation tasks and lack the ability for high-level understanding. Another line of works uses the object-related features as a complement to the original video representations to enhance object awareness [24, 30, 53, 92, 93]. These methods achieve promising results on multiple action benchmarks [25, 26, 53], but still have two major limitations. First, they depend on bounding box annotations [30, 80, 92], additional detectors or modalities [88, 91] to identify different objects, restricting the scalability to unseen data. Second, the identified objects provide plentiful cues for video analysis, yet these methods still fuse the disentangled object representations with the original video features, imposing limitations on efficiency. In contrast, our process is free of object annotations or handcrafted priors but only adopts simple contrastive loss to distill general object knowledge from image foundation models and decouple object components in video frames. Then we directly model temporal state changes on compact object representations, enabling more intuitive and efficient video understanding." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 414, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 414, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 414, + 482, + 666 + ], + "type": "text", + "content": "Efficient tuning becomes a tending direction with the development of large foundation models [7,14,69,94]. It is proposed in natural language processing to reduce the trainable parameters and improve efficiency when transferring pretrained models to downstream tasks [31,32,41,89]. Common strategies include inserting task-specific adapters into Transformer encoders [31,60,61], attaching prompt tokens to the input [42,48,67,70] and learning low-rank approximations [32]. Recently, the efficient tuning is increasingly explored in computer vision [4,11,22,35,44,72,95,96]. By freezing the image pre-trained model, [11,72] tune adapters with learnable parameters for efficient adaptation. [4,22,35,44] develop prompt tuning to adapt the frozen image model to various image tasks. Later, [36,46,47,55,57,58,68,77,86] take one step further to adapt image models to videos with temporal dynamics. In order to equip the model with temporal perception ability, [47,57,86] insert spatio-temporal modules into Transformer encoder blocks, but the tunable parameters in shallow Transformer layers require huge GPU memory in gradient backpropagation. [58,68] require two pathways to process spatial and temporal information. [36,46] attach spatio-temporal attention or motion decoders to the frame-wise features, but there exists high redundancy in attention computation [5,73]. Contrarily, our method retains the backbone architecture without modifications, distills object knowledge from image foundation models, and captures temporal state changes of discovered objects for video downstream tasks." + } + ] + } + ], + "index": 4 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "type": "text", + "content": "R. Qian et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 132, + 114, + 478, + 201 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 478, + 201 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 478, + 201 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 478, + 201 + ], + "type": "image", + "image_path": "dd7239b342f0ccbcc86a9401fad53167882a22250a0fdbf3c62af3b3fe820ab4.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 211, + 482, + 289 + ], + "lines": [ + { + "bbox": [ + 130, + 211, + 482, + 289 + ], + "spans": [ + { + "bbox": [ + 130, + 211, + 482, + 289 + ], + "type": "text", + "content": "Fig. 2: An overview of our object-centric image-to-video adaptation framework. We use a frozen image pre-trained model to extract frame-wise features and pass them through a lightweight temporal fusion block. Then we employ slot attention with learnable queries to decompose each frame into a compact set of object tokens. Thereafter, we develop object-time interaction layers to establish inter-object interactions and build temporal state changes of individual objects, which are then pooled into the video-level feature for action recognition." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 310, + 202, + 323 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 310, + 202, + 323 + ], + "spans": [ + { + "bbox": [ + 132, + 310, + 202, + 323 + ], + "type": "text", + "content": "3 Method" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 335, + 482, + 431 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 335, + 482, + 431 + ], + "spans": [ + { + "bbox": [ + 130, + 335, + 482, + 431 + ], + "type": "text", + "content": "The overall architecture of our object-centric image-to-video adaptation is shown in Fig. 2. We use a frozen pre-trained Vision Transformer [20] as the backbone to extract frame-wise features from the video clip. Then we apply slot attention with learnable query tokens to identify different object components and pass the object tokens through object-time interaction layers to model the temporal state changes of individual objects. The obtained object state change vectors encode rich semantics and essential temporal dynamics in the given video, and are subsequently pooled into the final video-level feature for downstream tasks." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 448, + 274, + 460 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 448, + 274, + 460 + ], + "spans": [ + { + "bbox": [ + 132, + 448, + 274, + 460 + ], + "type": "text", + "content": "3.1 Object Decomposition" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 468, + 481, + 503 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 468, + 481, + 503 + ], + "spans": [ + { + "bbox": [ + 130, + 468, + 481, + 503 + ], + "type": "text", + "content": "To start with, given a video clip " + }, + { + "bbox": [ + 130, + 468, + 481, + 503 + ], + "type": "inline_equation", + "content": "\\pmb{x} = \\{x_{1}, x_{2}, \\dots, x_{T}\\}" + }, + { + "bbox": [ + 130, + 468, + 481, + 503 + ], + "type": "text", + "content": " consisting of " + }, + { + "bbox": [ + 130, + 468, + 481, + 503 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 130, + 468, + 481, + 503 + ], + "type": "text", + "content": " frames, we use a frozen image pre-trained Vision Transformer " + }, + { + "bbox": [ + 130, + 468, + 481, + 503 + ], + "type": "inline_equation", + "content": "\\Phi_{\\mathrm{enc}}" + }, + { + "bbox": [ + 130, + 468, + 481, + 503 + ], + "type": "text", + "content": " as the backbone to extract frame-wise features:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 206, + 510, + 481, + 526 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 206, + 510, + 481, + 526 + ], + "spans": [ + { + "bbox": [ + 206, + 510, + 481, + 526 + ], + "type": "interline_equation", + "content": "p _ {t}, f _ {t} = \\Phi_ {\\mathrm {e n c}} (x _ {t}), \\quad p _ {t} \\in \\mathbb {R} ^ {D}, \\quad f _ {t} \\in \\mathbb {R} ^ {H W \\times D}, \\tag {1}", + "image_path": "f5fc963ce5b898d4e816fe8f612662711a527e4a314946ecb7462efd6b47ec8b.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 533, + 482, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 482, + 594 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 482, + 594 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 533, + 482, + 594 + ], + "type": "inline_equation", + "content": "p_t" + }, + { + "bbox": [ + 130, + 533, + 482, + 594 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 533, + 482, + 594 + ], + "type": "inline_equation", + "content": "f_t" + }, + { + "bbox": [ + 130, + 533, + 482, + 594 + ], + "type": "text", + "content": " denote the output c1s token and feature map from the last Transformer block. " + }, + { + "bbox": [ + 130, + 533, + 482, + 594 + ], + "type": "inline_equation", + "content": "H, W, D" + }, + { + "bbox": [ + 130, + 533, + 482, + 594 + ], + "type": "text", + "content": " respectively denote height, width, and channel dimension. These image pre-trained features encode rich but highly-entangled semantics. And the next step is to parse these frame features into different object components which are crucial for video perception and understanding [24,30,80,92]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": "Inspired by recent works on unsupervised object discovery and segmentation [51,82,83], we develop a variant of slot attention on top of the feature map " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "f_{t}" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": " for object decomposition in each frame. To be specific, instead of random sampling from a prior distribution, we use " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": " learnable query tokens " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "q \\in \\mathbb{R}^{N \\times D}" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": " to initialize the slot vectors " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "S \\in \\mathbb{R}^{N \\times D}" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": ". The intuition of this design is to encourage each query token to capture specific semantics or concepts [33, 83] and produce" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 274, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 274, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 274, + 91, + 448, + 102 + ], + "type": "text", + "content": "Object-centric Image-to-Video Adaptation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 482, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 482, + 189 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 482, + 189 + ], + "type": "text", + "content": "temporally coherent object decompositions, which provide essential inductive bias for temporal reasoning in later stages. Then at each slot attention iteration, following [51], we use three learnable linear transformations to project the slot vectors " + }, + { + "bbox": [ + 132, + 116, + 482, + 189 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 132, + 116, + 482, + 189 + ], + "type": "text", + "content": " and frame features " + }, + { + "bbox": [ + 132, + 116, + 482, + 189 + ], + "type": "inline_equation", + "content": "f_{t}" + }, + { + "bbox": [ + 132, + 116, + 482, + 189 + ], + "type": "text", + "content": " into query, key and value, i.e., " + }, + { + "bbox": [ + 132, + 116, + 482, + 189 + ], + "type": "inline_equation", + "content": "Q \\in \\mathbb{R}^{N \\times D}" + }, + { + "bbox": [ + 132, + 116, + 482, + 189 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 132, + 116, + 482, + 189 + ], + "type": "inline_equation", + "content": "K \\in \\mathbb{R}^{HW \\times D}" + }, + { + "bbox": [ + 132, + 116, + 482, + 189 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 132, + 116, + 482, + 189 + ], + "type": "inline_equation", + "content": "V \\in \\mathbb{R}^{HW \\times D}" + }, + { + "bbox": [ + 132, + 116, + 482, + 189 + ], + "type": "text", + "content": ". After that, we compute the attention matrix " + }, + { + "bbox": [ + 132, + 116, + 482, + 189 + ], + "type": "inline_equation", + "content": "\\tilde{A} \\in \\mathbb{R}^{N \\times HW}" + }, + { + "bbox": [ + 132, + 116, + 482, + 189 + ], + "type": "text", + "content": " with softmax normalization along the slot dimension as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 207, + 207, + 481, + 237 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 207, + 207, + 481, + 237 + ], + "spans": [ + { + "bbox": [ + 207, + 207, + 481, + 237 + ], + "type": "interline_equation", + "content": "\\tilde {A} _ {i, j} := \\frac {e ^ {A _ {i , j}}}{\\sum_ {l = 1} ^ {N} e ^ {A _ {l , j}}}, \\quad \\text {w h e r e} \\quad A := \\frac {1}{\\sqrt {D}} Q K ^ {T}. \\tag {2}", + "image_path": "c0a7a87cabc97975a1c3b6319ac0aad530ce9b7ade2e39323466a127b94a0881.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 244, + 481, + 269 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 244, + 481, + 269 + ], + "spans": [ + { + "bbox": [ + 132, + 244, + 481, + 269 + ], + "type": "text", + "content": "We use the weighted mean to aggregate the value and pass them through Gated Recurrent Unit (GRU) to update the slot vectors:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 149, + 278, + 481, + 309 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 278, + 481, + 309 + ], + "spans": [ + { + "bbox": [ + 149, + 278, + 481, + 309 + ], + "type": "interline_equation", + "content": "S := \\operatorname {G R U} (\\text {i n p u t s} = M V, \\text {s t a t e s} = S), \\quad \\text {w h e r e} \\quad M _ {i, j} := \\frac {\\tilde {A} _ {i , j}}{\\sum_ {l = 1} ^ {H W} \\tilde {A} _ {i , l}}. \\tag {3}", + "image_path": "873c9b4134ce771ec23168fd7c0a36e379d0a2187c4153ea1e5dc63189649626.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 316, + 481, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 316, + 481, + 376 + ], + "spans": [ + { + "bbox": [ + 132, + 316, + 481, + 376 + ], + "type": "text", + "content": "We iterate the routing process three times, and take the final slot attention weights as potential object segmentation masks. The slot vectors from the final iteration are treated as the object tokens that distill object knowledge from the pre-trained image features. Note that we perform slot attention " + }, + { + "bbox": [ + 132, + 316, + 481, + 376 + ], + "type": "inline_equation", + "content": "\\Phi_{\\mathrm{sa}}" + }, + { + "bbox": [ + 132, + 316, + 481, + 376 + ], + "type": "text", + "content": " on each frame in parallel, and denote the object tokens of each time stamp as:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 228, + 384, + 481, + 399 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 228, + 384, + 481, + 399 + ], + "spans": [ + { + "bbox": [ + 228, + 384, + 481, + 399 + ], + "type": "interline_equation", + "content": "\\boldsymbol {o} = \\left\\{o _ {1} \\oplus o _ {2} \\oplus \\dots \\oplus o _ {T} \\right\\} \\in \\mathbb {R} ^ {T \\times N \\times D}, \\tag {4}", + "image_path": "68ce81a514ffea4cde0381f4a6a131b2775bee6013fb3057d1f6c3cb67a22d61.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 409, + 481, + 457 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 409, + 481, + 457 + ], + "spans": [ + { + "bbox": [ + 130, + 409, + 481, + 457 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 409, + 481, + 457 + ], + "type": "inline_equation", + "content": "o_t = \\Phi_{\\mathrm{sa}}(q, \\tilde{f}_t) \\in \\mathbb{R}^{N \\times D}" + }, + { + "bbox": [ + 130, + 409, + 481, + 457 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 409, + 481, + 457 + ], + "type": "inline_equation", + "content": "\\oplus" + }, + { + "bbox": [ + 130, + 409, + 481, + 457 + ], + "type": "text", + "content": " denotes concatenation operation along time dimension. Since this attention mechanism introduces competition among slots, the obtained object tokens " + }, + { + "bbox": [ + 130, + 409, + 481, + 457 + ], + "type": "inline_equation", + "content": "o_t" + }, + { + "bbox": [ + 130, + 409, + 481, + 457 + ], + "type": "text", + "content": " are expected to take over distinct semantic parts and separate each frame into " + }, + { + "bbox": [ + 130, + 409, + 481, + 457 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 409, + 481, + 457 + ], + "type": "text", + "content": " different components." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 459, + 482, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 459, + 482, + 555 + ], + "spans": [ + { + "bbox": [ + 130, + 459, + 482, + 555 + ], + "type": "text", + "content": "Surprisingly, we find that the distilled tokens and corresponding attention matrices perform remarkably well in zero-shot video object segmentation tasks. They are even comparable to some pixel-level supervised counterparts, as demonstrated in Section 4.3. This result verifies that our method effectively extracts robust object-centric features for temporal reasoning in the later stage. Note that since videos often contain partial object semantics, there tend to be some slots that carry background information and interact with the foreground object tokens for comprehensive video analysis." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 572, + 285, + 585 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 572, + 285, + 585 + ], + "spans": [ + { + "bbox": [ + 132, + 572, + 285, + 585 + ], + "type": "text", + "content": "3.2 Object-Time Interaction" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 594, + 482, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 667 + ], + "type": "text", + "content": "After obtaining the object tokens that reveal potential object semantics, we use them as compressed representations of video frames and further model the temporal dynamics from an object-centric perspective [30, 93]. Our intuition is that in continuous frames, the object tokens, " + }, + { + "bbox": [ + 130, + 594, + 482, + 667 + ], + "type": "inline_equation", + "content": "o_{t}^{n} \\in \\mathbb{R}^{D}, t = 1,2,\\dots,T" + }, + { + "bbox": [ + 130, + 594, + 482, + 667 + ], + "type": "text", + "content": ", with the same index " + }, + { + "bbox": [ + 130, + 594, + 482, + 667 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 130, + 594, + 482, + 667 + ], + "type": "text", + "content": " tend to represent the same objects across time as illustrated in Fig. 1. Hence, it is feasible to temporally track each object and conclude the" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "type": "text", + "content": "R. Qian et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": "object state changes based on the learned object tokens, thereby assisting the process of video understanding." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 140, + 480, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 140, + 480, + 163 + ], + "spans": [ + { + "bbox": [ + 130, + 140, + 480, + 163 + ], + "type": "text", + "content": "To achieve this goal, we first apply self-attention to the object tokens within the same frame to introduce inter-object interactions at each time stamp:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 260, + 171, + 480, + 185 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 260, + 171, + 480, + 185 + ], + "spans": [ + { + "bbox": [ + 260, + 171, + 480, + 185 + ], + "type": "interline_equation", + "content": "\\tilde {o} _ {t} = \\Phi_ {\\mathrm {o i}} \\left(o _ {t}\\right) \\in \\mathbb {R} ^ {N \\times D}, \\tag {5}", + "image_path": "7832503e279ca8fa0dc2d2520fd1e97b3dbe1591d9f6a19b7d1c44d5e85d4bd4.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 193, + 479, + 252 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 193, + 479, + 252 + ], + "spans": [ + { + "bbox": [ + 130, + 193, + 479, + 252 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 193, + 479, + 252 + ], + "type": "inline_equation", + "content": "\\Phi_{\\mathrm{oi}}" + }, + { + "bbox": [ + 130, + 193, + 479, + 252 + ], + "type": "text", + "content": " is the object interaction module instantiated by a standard Transformer encoder layer consisting of linear projections, self-attention calculation and a feed forward network. The resulting vector " + }, + { + "bbox": [ + 130, + 193, + 479, + 252 + ], + "type": "inline_equation", + "content": "\\tilde{\\sigma}_t^n" + }, + { + "bbox": [ + 130, + 193, + 479, + 252 + ], + "type": "text", + "content": " not only encodes the attribute of object " + }, + { + "bbox": [ + 130, + 193, + 479, + 252 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 130, + 193, + 479, + 252 + ], + "type": "text", + "content": " but also captures its relationships with contextual objects, serving as a representation for the latent state of object " + }, + { + "bbox": [ + 130, + 193, + 479, + 252 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 130, + 193, + 479, + 252 + ], + "type": "text", + "content": " at time stamp " + }, + { + "bbox": [ + 130, + 193, + 479, + 252 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 193, + 479, + 252 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 254, + 480, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 254, + 480, + 384 + ], + "spans": [ + { + "bbox": [ + 130, + 254, + 480, + 384 + ], + "type": "text", + "content": "Considering that the temporal state changes are a concrete manifestation of the effects of actions, they serve as an important cue for analyzing the events occurring in a video and assisting video understanding [1,59,71,79]. Hence, based on the vectors representing latent object states, we take one step further to look into multiple time stamps and explicitly establish the state changes of each object. Without loss of generality, we take the object with index " + }, + { + "bbox": [ + 130, + 254, + 480, + 384 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 130, + 254, + 480, + 384 + ], + "type": "text", + "content": " for illustration. Given the object states at different time stamps, we define a temporal interval " + }, + { + "bbox": [ + 130, + 254, + 480, + 384 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 130, + 254, + 480, + 384 + ], + "type": "text", + "content": " to sample the initial and final states, i.e., " + }, + { + "bbox": [ + 130, + 254, + 480, + 384 + ], + "type": "inline_equation", + "content": "\\tilde{o}_t^n" + }, + { + "bbox": [ + 130, + 254, + 480, + 384 + ], + "type": "text", + "content": " as the initial state and " + }, + { + "bbox": [ + 130, + 254, + 480, + 384 + ], + "type": "inline_equation", + "content": "\\tilde{o}_{t + \\delta}^n" + }, + { + "bbox": [ + 130, + 254, + 480, + 384 + ], + "type": "text", + "content": " as the final state. We concatenate them along channel dimension and pass them through the temporal interaction layer " + }, + { + "bbox": [ + 130, + 254, + 480, + 384 + ], + "type": "inline_equation", + "content": "\\Phi_{\\mathrm{ti}}" + }, + { + "bbox": [ + 130, + 254, + 480, + 384 + ], + "type": "text", + "content": " to model the object state change between time " + }, + { + "bbox": [ + 130, + 254, + 480, + 384 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 254, + 480, + 384 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 254, + 480, + 384 + ], + "type": "inline_equation", + "content": "t + \\delta" + }, + { + "bbox": [ + 130, + 254, + 480, + 384 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 195, + 392, + 480, + 407 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 195, + 392, + 480, + 407 + ], + "spans": [ + { + "bbox": [ + 195, + 392, + 480, + 407 + ], + "type": "interline_equation", + "content": "s _ {t} ^ {n} = \\Phi_ {\\mathrm {t i}} \\left(\\left[ \\tilde {\\alpha} _ {t} ^ {n} \\oplus \\tilde {\\alpha} _ {t + \\delta} ^ {n} \\right]\\right) \\in \\mathbb {R} ^ {D}, \\quad t = \\{1, 2, \\dots , T - \\delta \\}, \\tag {6}", + "image_path": "142d65275d828260c399c3c0bf8a334df3fe4c9cbb2f56d17b79789e3d5642f9.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 414, + 479, + 461 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 414, + 479, + 461 + ], + "spans": [ + { + "bbox": [ + 130, + 414, + 479, + 461 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 414, + 479, + 461 + ], + "type": "inline_equation", + "content": "\\oplus" + }, + { + "bbox": [ + 130, + 414, + 479, + 461 + ], + "type": "text", + "content": " denotes channel-wise concatenation. We instantiate " + }, + { + "bbox": [ + 130, + 414, + 479, + 461 + ], + "type": "inline_equation", + "content": "\\Phi_{\\mathrm{ti}}" + }, + { + "bbox": [ + 130, + 414, + 479, + 461 + ], + "type": "text", + "content": " as a multi-layer perceptron of channel dimension " + }, + { + "bbox": [ + 130, + 414, + 479, + 461 + ], + "type": "inline_equation", + "content": "2D - D - D" + }, + { + "bbox": [ + 130, + 414, + 479, + 461 + ], + "type": "text", + "content": ". In this way, for each object, we obtain " + }, + { + "bbox": [ + 130, + 414, + 479, + 461 + ], + "type": "inline_equation", + "content": "T' = T - \\delta" + }, + { + "bbox": [ + 130, + 414, + 479, + 461 + ], + "type": "text", + "content": " state change vectors. The state change matrix covering all objects and time stamps can be denoted as " + }, + { + "bbox": [ + 130, + 414, + 479, + 461 + ], + "type": "inline_equation", + "content": "\\boldsymbol{s} \\in \\mathbb{R}^{T' \\times N \\times D}" + }, + { + "bbox": [ + 130, + 414, + 479, + 461 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 463, + 480, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 463, + 480, + 533 + ], + "spans": [ + { + "bbox": [ + 130, + 463, + 480, + 533 + ], + "type": "text", + "content": "Since the state change is a continuous process within a video sequence, it is feasible to conclude the overall state change of each object throughout the entire video by simple temporal average pooling on the established matrix " + }, + { + "bbox": [ + 130, + 463, + 480, + 533 + ], + "type": "inline_equation", + "content": "\\mathbf{s}" + }, + { + "bbox": [ + 130, + 463, + 480, + 533 + ], + "type": "text", + "content": ". Thereafter, we further apply average pooling on object dimension to aggregate the state change information of all objects as the final video-level representation, " + }, + { + "bbox": [ + 130, + 463, + 480, + 533 + ], + "type": "inline_equation", + "content": "v \\in \\mathbb{R}^D" + }, + { + "bbox": [ + 130, + 463, + 480, + 533 + ], + "type": "text", + "content": ", for downstream action recognition." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 534, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 534, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 534, + 481, + 665 + ], + "type": "text", + "content": "Discussion. Unlike the existing works that use object features as complements to original video features [24,30,53,92,93] or model state changes on the highly entangled scene representations [1, 59, 71, 79], we perform temporal reasoning solely on distilled object tokens. Our object-centric state change modeling reduces redundancy, mitigates ambiguity and results in more fine-grained cues for video understanding. And comparing with the prevalent image-to-video adaptation methods, they introduce dense temporal fusion operations, e.g., temporal convolutions [46, 47, 57] and attention [46, 47, 58, 68, 86], to mix multi-frame features for action recognition. In contrast, we leverage the object state changes to perceive temporal dynamics, which closely resemble human logic and provide an effective inductive bias to the model. To be specific, it does not require dense" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 274, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 274, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 274, + 91, + 447, + 102 + ], + "type": "text", + "content": "Object-centric Image-to-Video Adaptation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 154 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 154 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 154 + ], + "type": "text", + "content": "fusion or traversal of all frames. Instead, we respectively model state changes on separated objects across sparsely sampled frames, enabling more efficient scene parsing and temporal perception as verified in Section 4.4." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 170, + 204, + 182 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 170, + 204, + 182 + ], + "spans": [ + { + "bbox": [ + 132, + 170, + 204, + 182 + ], + "type": "text", + "content": "3.3 Training" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 190, + 479, + 249 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 190, + 479, + 249 + ], + "spans": [ + { + "bbox": [ + 130, + 190, + 479, + 249 + ], + "type": "text", + "content": "In training, we freeze the parameters in the image pre-trained model " + }, + { + "bbox": [ + 130, + 190, + 479, + 249 + ], + "type": "inline_equation", + "content": "\\Phi_{\\mathrm{enc}}" + }, + { + "bbox": [ + 130, + 190, + 479, + 249 + ], + "type": "text", + "content": ", and tune parameters that are responsible for object decomposition and temporal perception, i.e., " + }, + { + "bbox": [ + 130, + 190, + 479, + 249 + ], + "type": "inline_equation", + "content": "\\Phi_{\\mathrm{temp}}" + }, + { + "bbox": [ + 130, + 190, + 479, + 249 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 190, + 479, + 249 + ], + "type": "inline_equation", + "content": "\\Phi_{\\mathrm{oi}}" + }, + { + "bbox": [ + 130, + 190, + 479, + 249 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 190, + 479, + 249 + ], + "type": "inline_equation", + "content": "\\Phi_{\\mathrm{ti}}" + }, + { + "bbox": [ + 130, + 190, + 479, + 249 + ], + "type": "text", + "content": ". Besides the standard cross-entropy loss for action classification, we introduce two object-level losses to further facilitate object decomposition and temporal state change modeling." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 250, + 480, + 334 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 250, + 480, + 334 + ], + "spans": [ + { + "bbox": [ + 130, + 250, + 480, + 334 + ], + "type": "text", + "content": "Object Distillation Loss. Considering that the image pre-trained foundation models tend to encode rich semantics in the given scene [9, 56, 69], the output c1s token " + }, + { + "bbox": [ + 130, + 250, + 480, + 334 + ], + "type": "inline_equation", + "content": "p_t" + }, + { + "bbox": [ + 130, + 250, + 480, + 334 + ], + "type": "text", + "content": " is expected to encompass the majority of the semantic components in frame " + }, + { + "bbox": [ + 130, + 250, + 480, + 334 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 250, + 480, + 334 + ], + "type": "text", + "content": ". Hence, we use a contrastive grounding loss [23, 27, 90] to align the learned object tokens " + }, + { + "bbox": [ + 130, + 250, + 480, + 334 + ], + "type": "inline_equation", + "content": "o_t" + }, + { + "bbox": [ + 130, + 250, + 480, + 334 + ], + "type": "text", + "content": " and the c1s token " + }, + { + "bbox": [ + 130, + 250, + 480, + 334 + ], + "type": "inline_equation", + "content": "p_t" + }, + { + "bbox": [ + 130, + 250, + 480, + 334 + ], + "type": "text", + "content": ", encouraging " + }, + { + "bbox": [ + 130, + 250, + 480, + 334 + ], + "type": "inline_equation", + "content": "o_t" + }, + { + "bbox": [ + 130, + 250, + 480, + 334 + ], + "type": "text", + "content": " to cover all objects in frame " + }, + { + "bbox": [ + 130, + 250, + 480, + 334 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 250, + 480, + 334 + ], + "type": "text", + "content": ". Particularly, we first define the correspondence score between the " + }, + { + "bbox": [ + 130, + 250, + 480, + 334 + ], + "type": "inline_equation", + "content": "o_t" + }, + { + "bbox": [ + 130, + 250, + 480, + 334 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 250, + 480, + 334 + ], + "type": "inline_equation", + "content": "p_t" + }, + { + "bbox": [ + 130, + 250, + 480, + 334 + ], + "type": "text", + "content": " as:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 223, + 343, + 481, + 376 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 223, + 343, + 481, + 376 + ], + "spans": [ + { + "bbox": [ + 223, + 343, + 481, + 376 + ], + "type": "interline_equation", + "content": "c \\left(o _ {t}, p _ {t}\\right) = \\sum_ {n = 1} ^ {N} \\frac {e ^ {\\left\\langle o _ {t} ^ {n} , p _ {t} \\right\\rangle / \\tau}}{\\sum_ {l = 1} ^ {N} e ^ {\\left\\langle o _ {t} ^ {l} , p _ {t} \\right\\rangle / \\tau}} \\left\\langle o _ {t} ^ {n}, p _ {t} \\right\\rangle , \\tag {7}", + "image_path": "8ca7167da84db7293f01e1d1f759c09253ef36bb6661436b2d8d9c8f869cdcc2.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 385, + 482, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 385, + 482, + 462 + ], + "spans": [ + { + "bbox": [ + 130, + 385, + 482, + 462 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 385, + 482, + 462 + ], + "type": "inline_equation", + "content": "\\langle o_t^n,p_t\\rangle = \\frac{o_t^n p_t}{\\|o_t^n\\|\\|p_t\\|}" + }, + { + "bbox": [ + 130, + 385, + 482, + 462 + ], + "type": "text", + "content": " denotes cosine similarity calculation, " + }, + { + "bbox": [ + 130, + 385, + 482, + 462 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 130, + 385, + 482, + 462 + ], + "type": "text", + "content": " is the temperature hyper-parameter. This correspondence score adaptively aggregates each object component and avoids penalizing unrelated object semantics. Thereafter, we randomly sample c1s token vectors " + }, + { + "bbox": [ + 130, + 385, + 482, + 462 + ], + "type": "inline_equation", + "content": "p^\\prime" + }, + { + "bbox": [ + 130, + 385, + 482, + 462 + ], + "type": "text", + "content": " from other videos within the minibatch to formulate a negative sample pool " + }, + { + "bbox": [ + 130, + 385, + 482, + 462 + ], + "type": "inline_equation", + "content": "\\mathcal{N}" + }, + { + "bbox": [ + 130, + 385, + 482, + 462 + ], + "type": "text", + "content": " and apply the contrastive loss to maximize the agreement between the aligned object and c1s token pairs:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 206, + 470, + 481, + 502 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 206, + 470, + 481, + 502 + ], + "spans": [ + { + "bbox": [ + 206, + 470, + 481, + 502 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {o b j} = - \\sum_ {t = 1} ^ {T} \\log \\frac {e ^ {c \\left(o _ {t} , p _ {t}\\right) / \\tau}}{e ^ {c \\left(o _ {t} , p _ {t}\\right) / \\tau} + \\sum_ {p ^ {\\prime} \\sim \\mathcal {N}} e ^ {c \\left(o _ {t} , p ^ {\\prime}\\right) / \\tau}}. \\tag {8}", + "image_path": "69059a4056fa7be045916e339328bad7a2afdc0e643f539d070aade4b6cdea6d.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 510, + 482, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 482, + 668 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 482, + 668 + ], + "type": "text", + "content": "This self-supervised alignment objective encourages the object tokens to cover diverse semantics and distill object knowledge from the image pre-trained model. Temporal Reasoning Loss. In our architecture, we analyze temporal dynamics from the perspective of object state changes. Therefore, it is crucial to formulate a learning objective that guides the temporal interaction module " + }, + { + "bbox": [ + 130, + 510, + 482, + 668 + ], + "type": "inline_equation", + "content": "\\Phi_{\\mathrm{ti}}" + }, + { + "bbox": [ + 130, + 510, + 482, + 668 + ], + "type": "text", + "content": " to accurately capture the state changes between different time stamps. To achieve this goal, our intuition is that the same objects undergoing the same type of actions should possess consistent state changes. Conversely, different actions will result in distinct state changes for the same object, and different objects will exhibit diverse state changes under the same action. Based on this, we develop a margin loss to guide object state change modeling. In detail, given " + }, + { + "bbox": [ + 130, + 510, + 482, + 668 + ], + "type": "inline_equation", + "content": "s_t^n" + }, + { + "bbox": [ + 130, + 510, + 482, + 668 + ], + "type": "text", + "content": " as query, we define the videos within the mini-batch that have the same action labels with the current video to form positive pool " + }, + { + "bbox": [ + 130, + 510, + 482, + 668 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 130, + 510, + 482, + 668 + ], + "type": "text", + "content": ", and those of different action labels as" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "type": "text", + "content": "R. Qian et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 131, + 116, + 479, + 188 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 116, + 479, + 188 + ], + "spans": [ + { + "bbox": [ + 131, + 116, + 479, + 188 + ], + "type": "text", + "content": "negative pool " + }, + { + "bbox": [ + 131, + 116, + 479, + 188 + ], + "type": "inline_equation", + "content": "\\mathcal{N}" + }, + { + "bbox": [ + 131, + 116, + 479, + 188 + ], + "type": "text", + "content": ". We sample the state change vectors of the same object and the same action, i.e., " + }, + { + "bbox": [ + 131, + 116, + 479, + 188 + ], + "type": "inline_equation", + "content": "\\{\\hat{s}_t^n |\\hat{s}\\in \\mathcal{P}\\}" + }, + { + "bbox": [ + 131, + 116, + 479, + 188 + ], + "type": "text", + "content": ", as positive pairs, which should be aligned with " + }, + { + "bbox": [ + 131, + 116, + 479, + 188 + ], + "type": "inline_equation", + "content": "s_t^n" + }, + { + "bbox": [ + 131, + 116, + 479, + 188 + ], + "type": "text", + "content": ". As for the negatives, we respectively sample the state change vectors of the same object but different actions, i.e., " + }, + { + "bbox": [ + 131, + 116, + 479, + 188 + ], + "type": "inline_equation", + "content": "\\{s_t^{l^{\\prime}n}|s^{\\prime}\\in \\mathcal{N}\\}" + }, + { + "bbox": [ + 131, + 116, + 479, + 188 + ], + "type": "text", + "content": ", and those from the same video but of different objects, i.e., " + }, + { + "bbox": [ + 131, + 116, + 479, + 188 + ], + "type": "inline_equation", + "content": "\\{s_t^m,m\\neq n\\}" + }, + { + "bbox": [ + 131, + 116, + 479, + 188 + ], + "type": "text", + "content": ", as hard negatives for discrimination. The margin loss with a margin hyper-parameter " + }, + { + "bbox": [ + 131, + 116, + 479, + 188 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 131, + 116, + 479, + 188 + ], + "type": "text", + "content": " is formulated as:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 143, + 198, + 480, + 270 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 198, + 480, + 270 + ], + "spans": [ + { + "bbox": [ + 143, + 198, + 480, + 270 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {\\text {t e m p}} = \\sum_ {t = 1} ^ {T ^ {\\prime}} \\sum_ {n = 1} ^ {N} \\left\\{\\sum_ {\\hat {s} \\in \\mathcal {P}} \\| s _ {t} ^ {n} - \\hat {s} _ {t} ^ {n} \\| _ {2} \\right. \\\\ + \\sum_ {s ^ {\\prime} \\in \\mathcal {N}} \\max \\left(\\lambda - \\| s _ {t} ^ {n} - s _ {t} ^ {\\prime n} \\| _ {2}, 0\\right) + \\sum_ {\\substack {m = 1 \\\\ m \\neq n}} ^ {N} \\max \\left(\\lambda - \\| s _ {t} ^ {n} - s _ {t} ^ {m} \\| _ {2}, 0\\right) \\}, \\tag{9} \\\\ \\end{array}", + "image_path": "d773868fc235f31152bfb170345b1da5799f2bb9bcce926fddd9b69b6c19c199.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 279, + 479, + 351 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 279, + 479, + 351 + ], + "spans": [ + { + "bbox": [ + 131, + 279, + 479, + 351 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 279, + 479, + 351 + ], + "type": "inline_equation", + "content": "\\| \\cdot \\|_2" + }, + { + "bbox": [ + 131, + 279, + 479, + 351 + ], + "type": "text", + "content": " denotes " + }, + { + "bbox": [ + 131, + 279, + 479, + 351 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_2" + }, + { + "bbox": [ + 131, + 279, + 479, + 351 + ], + "type": "text", + "content": " distance. This term enforces the model to discriminate diverse object states and capture unique dynamic effects of each action category. Overall Objectives. Besides the above object-level losses, for action recognition tasks, we apply a linear classification head on the video-level representation " + }, + { + "bbox": [ + 131, + 279, + 479, + 351 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 131, + 279, + 479, + 351 + ], + "type": "text", + "content": " with a supervised classification loss " + }, + { + "bbox": [ + 131, + 279, + 479, + 351 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{cls}" + }, + { + "bbox": [ + 131, + 279, + 479, + 351 + ], + "type": "text", + "content": " in the form of cross-entropy. In this way, we formulate the overall loss function as:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 252, + 362, + 480, + 373 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 252, + 362, + 480, + 373 + ], + "spans": [ + { + "bbox": [ + 252, + 362, + 480, + 373 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathcal {L} _ {o b j} + \\mathcal {L} _ {t e m p} + \\mathcal {L} _ {c l s}. \\tag {10}", + "image_path": "27cc582aec7a5aebd2f8643ff7da5c28816e47498168fb4a091fc231cd1501b5.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 392, + 230, + 406 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 392, + 230, + 406 + ], + "spans": [ + { + "bbox": [ + 132, + 392, + 230, + 406 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 416, + 279, + 428 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 416, + 279, + 428 + ], + "spans": [ + { + "bbox": [ + 132, + 416, + 279, + 428 + ], + "type": "text", + "content": "4.1 Implementation Details" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 436, + 479, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 436, + 479, + 544 + ], + "spans": [ + { + "bbox": [ + 131, + 436, + 479, + 544 + ], + "type": "text", + "content": "We train and evaluate on three popular action recognition datasets: Kinetics-400 (K-400) [38], Something-Something-v2 (SSv2) [25], and Epic-Kitchens-100 (EK100) [13]. Additionally, we assess zero-shot video object segmentation on challenging multiple object segmentation dataset, DAVIS-2017-UVOS [8]. We use the frozen CLIP [69] pre-trained Vision Transformer [20] as " + }, + { + "bbox": [ + 131, + 436, + 479, + 544 + ], + "type": "inline_equation", + "content": "\\Phi_{\\mathrm{enc}}" + }, + { + "bbox": [ + 131, + 436, + 479, + 544 + ], + "type": "text", + "content": " for framewise feature extraction. In training, we crop each frame into " + }, + { + "bbox": [ + 131, + 436, + 479, + 544 + ], + "type": "inline_equation", + "content": "224\\times 224" + }, + { + "bbox": [ + 131, + 436, + 479, + 544 + ], + "type": "text", + "content": ", set the number of query tokens " + }, + { + "bbox": [ + 131, + 436, + 479, + 544 + ], + "type": "inline_equation", + "content": "N = 8" + }, + { + "bbox": [ + 131, + 436, + 479, + 544 + ], + "type": "text", + "content": ", and the time interval for object state change " + }, + { + "bbox": [ + 131, + 436, + 479, + 544 + ], + "type": "inline_equation", + "content": "\\delta = T / 4" + }, + { + "bbox": [ + 131, + 436, + 479, + 544 + ], + "type": "text", + "content": " in default. We use AdamW optimizer [52] with an initial learning rate " + }, + { + "bbox": [ + 131, + 436, + 479, + 544 + ], + "type": "inline_equation", + "content": "5\\times 10^{-4}" + }, + { + "bbox": [ + 131, + 436, + 479, + 544 + ], + "type": "text", + "content": " and mini-batch size 512 to update the tunable parameters." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 562, + 291, + 574 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 562, + 291, + 574 + ], + "spans": [ + { + "bbox": [ + 132, + 562, + 291, + 574 + ], + "type": "text", + "content": "4.2 Video Action Recognition" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 582, + 479, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 479, + 616 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 479, + 616 + ], + "type": "text", + "content": "In this section, we present the comparisons between our method and recent state-of-the-art on K-400, SSv2 and EK100 in terms of the number of tunable parameters, inference GFLOPs and action recognition accuracy." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 131, + 617, + 479, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 617, + 479, + 665 + ], + "spans": [ + { + "bbox": [ + 131, + 617, + 479, + 665 + ], + "type": "text", + "content": "K-400. We first show the results on K-400 in Table 1 and conclude three observations. (1) Compared to fully finetuned models, our method achieves superior performance with significantly fewer tunable parameters. For example, our method of ViT-L/14 with only 21M tunable parameters outperforms all" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 274, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 274, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 274, + 91, + 447, + 102 + ], + "type": "text", + "content": "Object-centric Image-to-Video Adaptation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 141, + 158, + 470, + 401 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "type": "text", + "content": "Table 1: Results on Kinetics-400. We report the pre-train model initialization, inference GFLOPs, the number of tunable parameters (M), Top-1 and Top-5 accuracy. The 'Views' indicates Frame Number " + }, + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "type": "inline_equation", + "content": "\\times" + }, + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "type": "text", + "content": " Temporal Clips " + }, + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "type": "inline_equation", + "content": "\\times" + }, + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "type": "text", + "content": " Spatial Crops." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 141, + 158, + 470, + 401 + ], + "lines": [ + { + "bbox": [ + 141, + 158, + 470, + 401 + ], + "spans": [ + { + "bbox": [ + 141, + 158, + 470, + 401 + ], + "type": "table", + "html": "
MethodPre-trainViewsGFLOPs TunableTop-1Top-5
Full finetuning
TimeSformer-L [5]IN-21K64 × 1 × 3714012180.794.7
VideoSwin-L [50]IN-21K32 × 4 × 3724819783.195.9
MTV-L [84]JFT32 × 4 × 31805087684.396.3
PromptCLIP A7 [36]CLIP16 × 5 × 1--76.893.5
ActionCLIP [77]CLIP32 × 10 × 31689014283.897.1
X-CLIP-L/14 [55]CLIP8 × 4 × 3789042087.197.6
Efficient tuning
EVL ViT-L/14 [46]CLIP32 × 3 × 180885987.3-
ST-Adapter ViT-B/16 [57]CLIP32 × 3 × 118217.282.796.2
STAN-self-B/16 [47]CLIP16 × 3 × 11187-84.996.8
AIM ViT-B/16 [86]CLIP32 × 3 × 124281184.796.7
DualPath ViT-B/16 [58]CLIP32 × 3 × 17101085.497.1
DiST ViT-B/16 [68]CLIP32 × 3 × 11950-85.097.0
DiST ViT-L/14 [68]CLIP32 × 3 × 18490-88.097.9
Ours ViT-B/16CLIP8 × 3 × 12816.585.897.3
Ours ViT-B/16CLIP16 × 3 × 15626.586.197.5
Ours ViT-B/16CLIP32 × 3 × 111236.586.297.5
Ours ViT-L/14CLIP32 × 3 × 150682188.598.0
", + "image_path": "bed9471378e3b04a4bb94b657ee919384878de40929b1216cfde8a8e42e1c1cc.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 425, + 482, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 425, + 482, + 668 + ], + "spans": [ + { + "bbox": [ + 130, + 425, + 482, + 668 + ], + "type": "text", + "content": "the fully finetuned approaches that require over 100M tunable parameters. (2) Our method with only a visual encoder achieves better results than the multimodal methods, ActionCLIP [77] and X-CLIP [55], which employ an extra text branch to guide the finetuning process. (3) Among the efficient tuning methods [46, 47, 57, 58, 68, 86], EVL [46] attaches a motion decoder with multiple temporal convolution and attention layers for temporal modeling. AIM [86], ST-Adapter [57] and STAN [47] insert temporal blocks into intermediate CLIP encoder layers. DualPath [58] and DiST [68] introduce an additional temporal pathway to capture temporal dynamics. Our formulation models temporal relationships on decomposed objects in the form of state changes, which reduces computation redundancy but attains superior performance. For example, on ViT-B/16, our method with only an 8-frame input yields superior results to other approaches that employ a 32-frame input. On ViT-L/14, we reach the state-of-the-art results with " + }, + { + "bbox": [ + 130, + 425, + 482, + 668 + ], + "type": "inline_equation", + "content": "40\\%" + }, + { + "bbox": [ + 130, + 425, + 482, + 668 + ], + "type": "text", + "content": " fewer GFLOPs (5068 vs 8490) than DiST [68]. SSv2. When extending to SSv2 which requires stronger temporal modeling, the observations are consistent. As depicted in Table 2, our object-centric formulation achieves state-of-the-art results and presents higher efficiency than both full finetuning and efficient tuning methods. In particular, our method with ViT-L/14 surpasses the advanced MViTv2-L [43] and DiST [68] with significantly reduced GFLOPs (5068 vs 8484, 8490). Moreover, on ViT-B/16, our method" + } + ] + } + ], + "index": 4 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "type": "text", + "content": "R. Qian et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 143, + 146, + 467, + 380 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 481, + 137 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 481, + 137 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 481, + 137 + ], + "type": "text", + "content": "Table 2: Results on Something-Something-v2. K-400†/K-600† indicates the model is pre-trained on both IN-21K and K-400/K-600." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 143, + 146, + 467, + 380 + ], + "lines": [ + { + "bbox": [ + 143, + 146, + 467, + 380 + ], + "spans": [ + { + "bbox": [ + 143, + 146, + 467, + 380 + ], + "type": "table", + "html": "
MethodPre-trainViewsGFLOPsTunableTop-1Top-5
Full finetuning
TimeSformer-L [5]IN-21K64 × 1 × 1714012162.4-
MTV-B [84]IN-21K32 × 4 × 3479031067.690.4
ViViT-L/16×2 [2]K-400†16 × 4 × 31189231165.489.8
VideoSwin-B [50]K-400†32 × 1 × 19638969.692.7
MVViTv2-L (312↑) [43]K-400†32 × 1 × 3848421373.394.1
Efficient tuning
EVL ViT-L/14 [46]CLIP32 × 1 × 3964117566.7-
ST-Adapter ViT-B/16 [57]CLIP32 × 3 × 119557.269.592.6
STAN-self-B/16 [47]CLIP16 × 3 × 11376-69.592.7
AIM ViT-B/16 [86]CLIP32 × 1 × 324961469.192.2
DualPath ViT-B/16 [58]CLIP32 × 1 × 37161370.392.9
DiST ViT-B/16 [68]CLIP32 × 3 × 11950-70.992.1
DiST ViT-L/14 [68]CLIP32 × 3 × 18490-73.193.2
Ours ViT-B/16CLIP8 × 1 × 32816.571.293.1
Ours ViT-B/16CLIP16 × 1 × 35626.571.693.3
Ours ViT-B/16CLIP32 × 1 × 311236.571.893.4
Ours ViT-L/14CLIP32 × 1 × 350682173.694.3
", + "image_path": "046dd8b6c61139186e4a1974ff79a8b84b174cec7b925dd057ac81bdf07aba7c.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 401, + 480, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 401, + 480, + 472 + ], + "spans": [ + { + "bbox": [ + 130, + 401, + 480, + 472 + ], + "type": "text", + "content": "with only 8 frames as input outperforms others using 32 frames. It is non-trivial to achieve this on such a motion-heavy dataset, which in return reveals the significance of our temporal modeling in the form of object state changes. On the one hand, it enables accurate temporal perception with sparser sampling. On the other hand, our temporal modeling on individual objects eliminates the need for dense temporal fusion, substantially reducing computational redundancy." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 473, + 481, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 473, + 481, + 568 + ], + "spans": [ + { + "bbox": [ + 130, + 473, + 481, + 568 + ], + "type": "text", + "content": "EK100. On challenging egocentric EK100 that requires long-term object interactions, our method showcases substantial advantages over existing efficient tunning counterparts [46,57,68] in Table 3. The underlying reasons are two-fold. First, our method involves explicit object decomposition, which helps to capture the primary objects in egocentric videos and contributes to the significant improvements in nouns. Second, our temporal modeling on individual objects reduces the interference of camera motions and enforces the model to focus on the dynamics of each object, thus facilitating verb perception." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 586, + 350, + 598 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 586, + 350, + 598 + ], + "spans": [ + { + "bbox": [ + 132, + 586, + 350, + 598 + ], + "type": "text", + "content": "4.3 Zero-shot Video Object Segmentation" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 605, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 481, + 666 + ], + "type": "text", + "content": "To verify whether our method understands videos in an object-centric way and interpret why the model works, we evaluate unsupervised video object segmentation in addition to action recognition. Particularly, we use the frozen ViT-B/16 backbone and adopt slot attention weights " + }, + { + "bbox": [ + 130, + 605, + 481, + 666 + ], + "type": "inline_equation", + "content": "\\{\\tilde{A}_i\\in \\mathbb{R}^{HW}\\}_{i = 1}^N" + }, + { + "bbox": [ + 130, + 605, + 481, + 666 + ], + "type": "text", + "content": " in Eq. 2 as object masks candidate without bells and whistles. We report " + }, + { + "bbox": [ + 130, + 605, + 481, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{J}\\& \\mathcal{F}" + }, + { + "bbox": [ + 130, + 605, + 481, + 666 + ], + "type": "text", + "content": " on challenging" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 274, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 274, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 274, + 91, + 447, + 102 + ], + "type": "text", + "content": "Object-centric Image-to-Video Adaptation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "content": "multiple object discovery benchmark, DAVIS-2017-UVOS [8]. Note that we directly apply the model trained on K-400 or SSv2 to video object segmentation in a zero-shot manner." + } + ] + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 133, + 208, + 321, + 312 + ], + "blocks": [ + { + "bbox": [ + 133, + 208, + 321, + 312 + ], + "lines": [ + { + "bbox": [ + 133, + 208, + 321, + 312 + ], + "spans": [ + { + "bbox": [ + 133, + 208, + 321, + 312 + ], + "type": "table", + "html": "
MethodVerbNounAction
EVL ViT-B/16 [46]62.751.037.7
ST-Adapter ViT-B/16 [57]67.755.0-
DiST ViT-B/16 [68]69.558.145.8
DiST ViT-L/14 [68]70.761.648.9
Ours ViT-B/1671.160.949.0
Ours ViT-L/1473.264.551.7
", + "image_path": "bef05998a54056c73ca85c0fd6c3df90f15bfce75006fd704ef79bf32ba8c821.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 323, + 207, + 477, + 312 + ], + "blocks": [ + { + "bbox": [ + 130, + 174, + 477, + 206 + ], + "lines": [ + { + "bbox": [ + 130, + 174, + 477, + 206 + ], + "spans": [ + { + "bbox": [ + 130, + 174, + 477, + 206 + ], + "type": "text", + "content": "Table 3: Results on Epic-Kitchens-100 Verb, Table 4: Results on multiple object Noun and Action. All compared methods take segmentation on DAVIS-2017-UVOS. an input of " + }, + { + "bbox": [ + 130, + 174, + 477, + 206 + ], + "type": "inline_equation", + "content": "8\\times 1\\times 3" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 323, + 207, + 477, + 312 + ], + "lines": [ + { + "bbox": [ + 323, + 207, + 477, + 312 + ], + "spans": [ + { + "bbox": [ + 323, + 207, + 477, + 312 + ], + "type": "table", + "html": "
ModelJ&FJF
OCLR [82]39.638.241.1
SMTC [63]40.536.444.6
BA [17]43.939.248.6
RVOS [76]41.236.845.7
Ours SSv241.037.644.4
Ours K-40044.742.347.1
Ours K-400 CRF49.346.552.1
", + "image_path": "a7404338429f733595e158cb547e8ffc59fda522a17822bcd4758453733aec43.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 341, + 482, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 341, + 482, + 568 + ], + "spans": [ + { + "bbox": [ + 130, + 341, + 482, + 568 + ], + "type": "text", + "content": "Quantitative Results. We present the comparison with recent state-of-the-arts without object annotations [63,81,82] as well as a fully supervised baseline [76] (denoted in grey) in Table 4. Among the compared works without mask annotations, OCLR [82] takes optical flow as input to distinguish motion patterns of diverse objects. SMTC [63] designs two-stage slot attention to produce temporally consistent object segmentations. BA [17] also utilizes image foundation models like DION [9] to provide spatio-temporal correspondence and use clustering to predict the object segmentation masks. On the one hand, since K-400 covers more diverse scenes and objects than SSv2, it leads to better generalization ability in this zero-shot segmentation scenario. On the other hand, our formulation trained on K-400 achieves superior results to these counterparts, even surpassing the simple supervised baseline RVOS [76]. This demonstrates the feasibility of transferring object knowledge from image foundation models to the video domain, and our object-centric temporal reasoning further facilitates temporally coherent object decomposition. However, the gap to the state-of-the-art supervised method is still large because the mask annotations lead to much more accurate segmentation boundaries. A straightforward way is to apply Conditional Random Field (CRF) [40] for refinement, which leads to 4.6 points improvement on " + }, + { + "bbox": [ + 130, + 341, + 482, + 568 + ], + "type": "inline_equation", + "content": "\\mathcal{J}\\&\\mathcal{F}" + }, + { + "bbox": [ + 130, + 341, + 482, + 568 + ], + "type": "text", + "content": " score. It is worth further exploration in the future." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 666 + ], + "type": "text", + "content": "Qualitative Results. We also visualize some examples of the generated segmentation masks in Fig. 3, where each column presents the frames within the same video and the same color denotes the masks generated by the same slot query. Generally, our model discriminates different object semantics in videos. Each distilled object token can capture specific semantics or concepts that generalize across diverse time stamps and scenes. For example, the token marked in red focuses on humans, the blue one emphasizes quadruped animals, and the yellow one identifies two-wheeled vehicles. This temporally coherent object" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "type": "text", + "content": "R. Qian et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": "decomposition provides an effective inductive bias for subsequent temporal reasoning, leading to more efficient temporal modeling." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 159, + 239, + 171 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 159, + 239, + 171 + ], + "spans": [ + { + "bbox": [ + 131, + 159, + 239, + 171 + ], + "type": "text", + "content": "4.4 Ablation Study" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 180, + 480, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 180, + 480, + 216 + ], + "spans": [ + { + "bbox": [ + 130, + 180, + 480, + 216 + ], + "type": "text", + "content": "In this section, we use the CLIP ViT-B/16 as " + }, + { + "bbox": [ + 130, + 180, + 480, + 216 + ], + "type": "inline_equation", + "content": "\\Phi_{\\mathrm{enc}}" + }, + { + "bbox": [ + 130, + 180, + 480, + 216 + ], + "type": "text", + "content": " with an input of 16 frames for all ablative experiments unless otherwise specified. More details and extra ablation studies are presented in Supplementary Materials." + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 133, + 238, + 470, + 323 + ], + "blocks": [ + { + "bbox": [ + 133, + 238, + 470, + 323 + ], + "lines": [ + { + "bbox": [ + 133, + 238, + 470, + 323 + ], + "spans": [ + { + "bbox": [ + 133, + 238, + 470, + 323 + ], + "type": "image", + "image_path": "d83ba3e1789da07a8fb77c42e0ba8d4bc92b5f9c55a1da1d0287ab2e4ebbe5e5.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 332, + 476, + 377 + ], + "lines": [ + { + "bbox": [ + 130, + 332, + 476, + 377 + ], + "spans": [ + { + "bbox": [ + 130, + 332, + 476, + 377 + ], + "type": "text", + "content": "Fig. 3: Visualization of object decomposition in Fig. 4: Comparison on training data the form of segmentation. Each column presents efficiency. Our method demonstrates three frames in a video. The same color denotes advantages, especially with a small the objects identified by the same object token. amount of training data." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 402, + 482, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 402, + 482, + 545 + ], + "spans": [ + { + "bbox": [ + 130, + 402, + 482, + 545 + ], + "type": "text", + "content": "Training Losses. We explore the effectiveness of our newly introduced object distillation loss " + }, + { + "bbox": [ + 130, + 402, + 482, + 545 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{obj}" + }, + { + "bbox": [ + 130, + 402, + 482, + 545 + ], + "type": "text", + "content": " and temporal reasoning loss " + }, + { + "bbox": [ + 130, + 402, + 482, + 545 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{temp}" + }, + { + "bbox": [ + 130, + 402, + 482, + 545 + ], + "type": "text", + "content": " in Table 5. We observe that " + }, + { + "bbox": [ + 130, + 402, + 482, + 545 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{obj}" + }, + { + "bbox": [ + 130, + 402, + 482, + 545 + ], + "type": "text", + "content": " leads to dramatic improvements on all three datasets, K-400, SSv2, and DAVIS-2017-UVOS. This is because " + }, + { + "bbox": [ + 130, + 402, + 482, + 545 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{obj}" + }, + { + "bbox": [ + 130, + 402, + 482, + 545 + ], + "type": "text", + "content": " offers rich semantic knowledge reference from the image foundation model to facilitate object identification. And better object decomposition further enhances high-level action recognition. As for " + }, + { + "bbox": [ + 130, + 402, + 482, + 545 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{temp}" + }, + { + "bbox": [ + 130, + 402, + 482, + 545 + ], + "type": "text", + "content": ", it guides the model to capture the state changes on each object. Without this guidance, relying solely on the action classification loss is insufficient to direct the temporal module to perceive representative temporal dynamics, thus resulting in a drastic performance drop on the motion-heavy dataset SSv2. And notably, this temporal loss also leads to improvements on DAVIS, which verifies that the temporal perception in return bolsters object decomposition." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": "Temporal Sampling Strategy. We also study the effect of different temporal sampling strategies in training. Specifically, we vary the number of frames per clip " + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": " and the time interval for object state change modeling " + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": ". In Table 6, we display the Top-1 accuracy on K-400 and SSv2, as well as the computation GFLOPs excluding ViT backbone for more intuitive comparison. An exciting phenomenon is that even on SSv2 that requires strong temporal modeling, we achieve promising results with very sparse sampling, i.e., only 4 frames as input with significantly reduced computation. Our method is generally robust to the temporal sampling hyper-parameters, demonstrating the effectiveness of our temporal reasoning on compressed object tokens." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 274, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 274, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 274, + 91, + 447, + 102 + ], + "type": "text", + "content": "Object-centric Image-to-Video Adaptation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 136, + 157, + 294, + 225 + ], + "blocks": [ + { + "bbox": [ + 136, + 157, + 294, + 225 + ], + "lines": [ + { + "bbox": [ + 136, + 157, + 294, + 225 + ], + "spans": [ + { + "bbox": [ + 136, + 157, + 294, + 225 + ], + "type": "table", + "html": "
LobjLtempK-400SSv2DAVIS
XX70.444.827.3
✓X80.355.239.5
X✓73.552.130.5
✓✓86.171.644.7
", + "image_path": "f7d9665ef3ab0d9d5768c611b1dbf8c3747cd2dba97ddfed6d69dab7f9b0ce9f.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 315, + 147, + 454, + 225 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 471, + 148 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 471, + 148 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 471, + 148 + ], + "type": "text", + "content": "Table 5: Ablation studies on training loss Table 6: Ablation studies on the number items " + }, + { + "bbox": [ + 132, + 114, + 471, + 148 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{obj}" + }, + { + "bbox": [ + 132, + 114, + 471, + 148 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 132, + 114, + 471, + 148 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{temp}" + }, + { + "bbox": [ + 132, + 114, + 471, + 148 + ], + "type": "text", + "content": ". Note that " + }, + { + "bbox": [ + 132, + 114, + 471, + 148 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{cls}" + }, + { + "bbox": [ + 132, + 114, + 471, + 148 + ], + "type": "text", + "content": " is of frames " + }, + { + "bbox": [ + 132, + 114, + 471, + 148 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 132, + 114, + 471, + 148 + ], + "type": "text", + "content": " and sampling interval " + }, + { + "bbox": [ + 132, + 114, + 471, + 148 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 132, + 114, + 471, + 148 + ], + "type": "text", + "content": ". always applied in training." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 315, + 147, + 454, + 225 + ], + "lines": [ + { + "bbox": [ + 315, + 147, + 454, + 225 + ], + "spans": [ + { + "bbox": [ + 315, + 147, + 454, + 225 + ], + "type": "table", + "html": "
TδGFLOPsK-400SSv2
414.285.470.7
828.485.871.2
848.285.871.1
16218.086.171.6
16417.886.171.6
", + "image_path": "80e2408a24c6c2c41b47dc27fc340a9bf13efbca970b2f43f870d24b718e29fb.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 248, + 482, + 333 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 248, + 482, + 333 + ], + "spans": [ + { + "bbox": [ + 130, + 248, + 482, + 333 + ], + "type": "text", + "content": "Training Data Efficiency. Fig. 4 presents K-400 action recognition performance with various amounts of training data. For fair comparison with fully fine-tuned TimeSformer [5] as well as the efficient tuning methods, ST-Adapter [57] and AIM [86], we use CLIP ViT-B/16 with an input of 8 frames. Our method exhibits superior data efficiency to existing works. And the advantage becomes more significant with less training data, demonstrating the robust capacity of our object-centric formulation to transfer image pre-trained knowledge to videos." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 350, + 220, + 363 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 350, + 220, + 363 + ], + "spans": [ + { + "bbox": [ + 132, + 350, + 220, + 363 + ], + "type": "text", + "content": "5 Conclusion" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 376, + 481, + 507 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 376, + 481, + 507 + ], + "spans": [ + { + "bbox": [ + 130, + 376, + 481, + 507 + ], + "type": "text", + "content": "In this work, we propose an object-centric formulation to adapt the image pretrained models to the video domain. Specifically, we employ a frozen image pre-trained model to extract frame-wise features and use slot attention with learnable queries to parse each frame into a compact set of object tokens. Then we apply object-time interaction to these object tokens to explicitly establish object state changes across time span. By using the distilled object tokens as the compressed representations of video frames, we have reduced computation costs and improved temporal reasoning capabilities especially under sparse sampling scenarios. Our method demonstrates state-of-the-art performance across action recognition benchmarks while realizing robust zero-shot video object segmentation, signifying a promising direction for future research in video analysis." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 508, + 481, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 508, + 481, + 604 + ], + "spans": [ + { + "bbox": [ + 130, + 508, + 481, + 604 + ], + "type": "text", + "content": "Limitations and Social Negative Impact. Our work demonstrates that the object-centric formulation serves as a reliable compression of video frames, which reduces redundancy and enables efficient temporal perception. But in some applications like a video QA system, it would be better to distill useful information according to specific questions. Incorporating questions as a condition to slot attention might address this problem. And there is a potential negative impact in privacy concerns. This work involves processing large amounts of visual data, which could include sensitive information and raise privacy concerns." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "type": "text", + "content": "R. Qian et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 137, + 481, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 137, + 481, + 185 + ], + "spans": [ + { + "bbox": [ + 132, + 137, + 481, + 185 + ], + "type": "text", + "content": "This work is partially supported by the National Key R&D Program of China (2022ZD0160201), Shanghai Artificial Intelligence Laboratory, the Centre for Perceptual and Interactive Intelligence (CPII) Ltd under the Innovation and Technology Commission (ITC)'s InnoHK." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 201, + 197, + 214 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 201, + 197, + 214 + ], + "spans": [ + { + "bbox": [ + 133, + 201, + 197, + 214 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 223, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 138, + 223, + 480, + 256 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 223, + 480, + 256 + ], + "spans": [ + { + "bbox": [ + 138, + 223, + 480, + 256 + ], + "type": "text", + "content": "1. Alayrac, J.B., Miech, A., Laptev, I., Sivic, J., et al.: Multi-task learning of object states and state-modifying actions from web videos. IEEE Transactions on Pattern Analysis and Machine Intelligence (2024)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 257, + 480, + 288 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 257, + 480, + 288 + ], + "spans": [ + { + "bbox": [ + 138, + 257, + 480, + 288 + ], + "type": "text", + "content": "2. Arnab, A., Dehghani, M., Heigold, G., Sun, C., Lucic, M., Schmid, C.: Vivit: A video vision transformer. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 6836-6846 (2021)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 289, + 480, + 320 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 289, + 480, + 320 + ], + "spans": [ + { + "bbox": [ + 138, + 289, + 480, + 320 + ], + "type": "text", + "content": "3. Aydemir, G., Xie, W., Guney, F.: Self-supervised object-centric learning for videos. In: Thirty-seventh Conference on Neural Information Processing Systems (2023), https://openreview.net/forum?id=919tWtJPXe" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 321, + 480, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 321, + 480, + 353 + ], + "spans": [ + { + "bbox": [ + 138, + 321, + 480, + 353 + ], + "type": "text", + "content": "4. Bahng, H., Jahanian, A., Sankaranarayanan, S., Isola, P.: Visual prompting: Modifying pixel space to adapt pre-trained models. arXiv preprint arXiv:2203.17274 (2022)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 354, + 480, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 354, + 480, + 373 + ], + "spans": [ + { + "bbox": [ + 138, + 354, + 480, + 373 + ], + "type": "text", + "content": "5. Bertasius, G., Wang, H., Torresani, L.: Is space-time attention all you need for video understanding? In: ICML. vol. 2, p. 4 (2021)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 374, + 480, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 374, + 480, + 407 + ], + "spans": [ + { + "bbox": [ + 138, + 374, + 480, + 407 + ], + "type": "text", + "content": "6. Besbinar, B., Frossard, P.: Self-supervision by prediction for object discovery in videos. In: 2021 IEEE International Conference on Image Processing (ICIP). pp. 1509-1513. IEEE (2021)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 407, + 480, + 438 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 407, + 480, + 438 + ], + "spans": [ + { + "bbox": [ + 138, + 407, + 480, + 438 + ], + "type": "text", + "content": "7. Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J.D., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., et al.: Language models are few-shot learners. Advances in neural information processing systems 33, 1877-1901 (2020)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 439, + 480, + 471 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 439, + 480, + 471 + ], + "spans": [ + { + "bbox": [ + 138, + 439, + 480, + 471 + ], + "type": "text", + "content": "8. Caelles, S., Pont-Tuset, J., Perazzi, F., Montes, A., Maninis, K.K., Van Gool, L.: The 2019 davis challenge on vos: Unsupervised multi-object segmentation. arXiv preprint arXiv:1905.00737 (2019)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 472, + 480, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 472, + 480, + 503 + ], + "spans": [ + { + "bbox": [ + 138, + 472, + 480, + 503 + ], + "type": "text", + "content": "9. Caron, M., Touvron, H., Misra, I., Jégou, H., Mairal, J., Bojanowski, P., Joulin, A.: Emerging properties in self-supervised vision transformers. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 9650-9660 (2021)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 504, + 480, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 504, + 480, + 536 + ], + "spans": [ + { + "bbox": [ + 138, + 504, + 480, + 536 + ], + "type": "text", + "content": "0. Carreira, J., Zisserman, A.: Quo vadis, action recognition? a new model and the kinetics dataset. In: proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 6299-6308 (2017)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 536, + 480, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 536, + 480, + 567 + ], + "spans": [ + { + "bbox": [ + 138, + 536, + 480, + 567 + ], + "type": "text", + "content": "1. Chen, S., Ge, C., Tong, Z., Wang, J., Song, Y., Wang, J., Luo, P.: Adaptformer: Adapting vision transformers for scalable visual recognition. Advances in Neural Information Processing Systems 35, 16664-16678 (2022)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 568, + 480, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 568, + 480, + 600 + ], + "spans": [ + { + "bbox": [ + 138, + 568, + 480, + 600 + ], + "type": "text", + "content": "2. Crawford, E., Pineau, J.: Exploiting spatial invariance for scalable unsupervised object tracking. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 34, pp. 3684-3692 (2020)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 601, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 601, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 138, + 601, + 480, + 632 + ], + "type": "text", + "content": "3. Damen, D., Doughty, H., Farinella, G.M., Furnari, A., Kazakos, E., Ma, J., Moltisanti, D., Munro, J., Perrett, T., Price, W., et al.: Rescaling egocentric vision. arXiv preprint arXiv:2006.13256 (2020)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 633, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 633, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 633, + 480, + 665 + ], + "type": "text", + "content": "4. Devlin, J., Chang, M.W., Lee, K., Toutanova, K.: Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805 (2018)" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 274, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 274, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 274, + 91, + 447, + 102 + ], + "type": "text", + "content": "Object-centric Image-to-Video Adaptation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 666 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 481, + 161 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 481, + 161 + ], + "type": "text", + "content": "15. Ding, S., Li, M., Yang, T., Qian, R., Xu, H., Chen, Q., Wang, J., Xiong, H.: Motion-aware contrastive video representation learning via foreground-background merging. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 9716-9726 (2022)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 162, + 481, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 162, + 481, + 194 + ], + "spans": [ + { + "bbox": [ + 133, + 162, + 481, + 194 + ], + "type": "text", + "content": "16. Ding, S., Qian, R., Xiong, H.: Dual contrastive learning for spatio-temporal representation. In: Proceedings of the 30th ACM international conference on multimedia. pp. 5649-5658 (2022)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 194, + 481, + 225 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 194, + 481, + 225 + ], + "spans": [ + { + "bbox": [ + 133, + 194, + 481, + 225 + ], + "type": "text", + "content": "17. Ding, S., Qian, R., Xu, H., Lin, D., Xiong, H.: Betrayed by attention: A simple yet effective approach for self-supervised video object segmentation. arXiv preprint arXiv:2311.17893 (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 227, + 481, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 227, + 481, + 258 + ], + "spans": [ + { + "bbox": [ + 132, + 227, + 481, + 258 + ], + "type": "text", + "content": "18. Ding, S., Xie, W., Chen, Y., Qian, R., Zhang, X., Xiong, H., Tian, Q.: Motion-inductive self-supervised object discovery in videos. arXiv preprint arXiv:2210.00221 (2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 259, + 481, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 259, + 481, + 293 + ], + "spans": [ + { + "bbox": [ + 132, + 259, + 481, + 293 + ], + "type": "text", + "content": "19. Ding, S., Zhao, P., Zhang, X., Qian, R., Xiong, H., Tian, Q.: Prune spatiotemporal tokens by semantic-aware temporal accumulation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 16945-16956 (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 293, + 481, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 293, + 481, + 335 + ], + "spans": [ + { + "bbox": [ + 132, + 293, + 481, + 335 + ], + "type": "text", + "content": "20. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929 (2020)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 336, + 481, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 336, + 481, + 369 + ], + "spans": [ + { + "bbox": [ + 132, + 336, + 481, + 369 + ], + "type": "text", + "content": "21. Feichtenhofer, C., Fan, H., Malik, J., He, K.: Slowfast networks for video recognition. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 6202-6211 (2019)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 369, + 481, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 369, + 481, + 413 + ], + "spans": [ + { + "bbox": [ + 132, + 369, + 481, + 413 + ], + "type": "text", + "content": "22. Feng, C., Zhong, Y., Jie, Z., Chu, X., Ren, H., Wei, X., Xie, W., Ma, L.: Promptdet: Towards open-vocabulary detection using uncurated images. In: Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part IX. pp. 701-717. Springer (2022)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 414, + 481, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 414, + 481, + 456 + ], + "spans": [ + { + "bbox": [ + 132, + 414, + 481, + 456 + ], + "type": "text", + "content": "23. Ghiasi, G., Gu, X., Cui, Y., Lin, T.Y.: Scaling open-vocabulary image segmentation with image-level labels. In: Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXXVI. pp. 540-557. Springer (2022)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 457, + 481, + 490 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 457, + 481, + 490 + ], + "spans": [ + { + "bbox": [ + 132, + 457, + 481, + 490 + ], + "type": "text", + "content": "24. Girdhar, R., Carreira, J., Doersch, C., Zisserman, A.: Video action transformer network. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 244-253 (2019)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 491, + 481, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 491, + 481, + 544 + ], + "spans": [ + { + "bbox": [ + 132, + 491, + 481, + 544 + ], + "type": "text", + "content": "25. Goyal, R., Ebrahimi Kahou, S., Michalski, V., Materzynska, J., Westphal, S., Kim, H., Haenel, V., Fruend, I., Yianilos, P., Mueller-Freitag, M., et al.: The\" something something\" video database for learning and evaluating visual common sense. In: Proceedings of the IEEE international conference on computer vision. pp. 5842-5850 (2017)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 545, + 481, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 545, + 481, + 588 + ], + "spans": [ + { + "bbox": [ + 132, + 545, + 481, + 588 + ], + "type": "text", + "content": "26. Gu, C., Sun, C., Ross, D.A., Vondrick, C., Pantofaru, C., Li, Y., Vijayanarasimhan, S., Toderici, G., Ricco, S., Sukthankar, R., et al.: Ava: A video dataset of spatiotemporally localized atomic visual actions. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 6047-6056 (2018)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 589, + 481, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 589, + 481, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 589, + 481, + 632 + ], + "type": "text", + "content": "27. Gupta, T., Vahdat, A., Chechik, G., Yang, X., Kautz, J., Hoiem, D.: Contrastive learning for weakly supervised phrase grounding. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part III. pp. 752-768. Springer (2020)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 632, + 481, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 481, + 666 + ], + "type": "text", + "content": "28. He, K., Chen, X., Xie, S., Li, Y., Dollár, P., Girshick, R.: Masked autoencoders are scalable vision learners. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16000-16009 (2022)" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "type": "text", + "content": "R. Qian et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 666 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "text", + "content": "29. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 770-778 (2016)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 149, + 482, + 192 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 149, + 482, + 192 + ], + "spans": [ + { + "bbox": [ + 130, + 149, + 482, + 192 + ], + "type": "text", + "content": "30. Herzig, R., Ben-Avraham, E., Mangalam, K., Bar, A., Chechik, G., Rohrbach, A., Darrell, T., Globerson, A.: Object-region video transformers. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3148-3159 (2022)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 192, + 482, + 225 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 192, + 482, + 225 + ], + "spans": [ + { + "bbox": [ + 130, + 192, + 482, + 225 + ], + "type": "text", + "content": "31. Houlsby, N., Giurgiu, A., Jastrzebski, S., Morrone, B., De Laroussilhe, Q., Gesmundo, A., Attariyan, M., Gelly, S.: Parameter-efficient transfer learning for nlp. In: International Conference on Machine Learning. pp. 2790-2799. PMLR (2019)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 225, + 482, + 257 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 225, + 482, + 257 + ], + "spans": [ + { + "bbox": [ + 130, + 225, + 482, + 257 + ], + "type": "text", + "content": "32. Hu, E.J., Shen, Y., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., Chen, W.: Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685 (2021)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 257, + 482, + 278 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 257, + 482, + 278 + ], + "spans": [ + { + "bbox": [ + 130, + 257, + 482, + 278 + ], + "type": "text", + "content": "33. Jia, B., Liu, Y., Huang, S.: Unsupervised object-centric learning with bi-level optimized query slot attention. arXiv preprint arXiv:2210.08990 (2022)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 278, + 482, + 320 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 278, + 482, + 320 + ], + "spans": [ + { + "bbox": [ + 130, + 278, + 482, + 320 + ], + "type": "text", + "content": "34. Jia, C., Yang, Y., Xia, Y., Chen, Y.T., Parekh, Z., Pham, H., Le, Q., Sung, Y.H., Li, Z., Duerig, T.: Scaling up visual and vision-language representation learning with noisy text supervision. In: International Conference on Machine Learning. pp. 4904-4916. PMLR (2021)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 320, + 482, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 320, + 482, + 365 + ], + "spans": [ + { + "bbox": [ + 130, + 320, + 482, + 365 + ], + "type": "text", + "content": "35. Jia, M., Tang, L., Chen, B.C., Cardie, C., Belongie, S., Hariharan, B., Lim, S.N.: Visual prompt tuning. In: Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXXIII. pp. 709-727. Springer (2022)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 365, + 482, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 365, + 482, + 407 + ], + "spans": [ + { + "bbox": [ + 130, + 365, + 482, + 407 + ], + "type": "text", + "content": "36. Ju, C., Han, T., Zheng, K., Zhang, Y., Xie, W.: Prompting visual-language models for efficient video understanding. In: Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXXV. pp. 105-124. Springer (2022)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 407, + 482, + 451 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 407, + 482, + 451 + ], + "spans": [ + { + "bbox": [ + 130, + 407, + 482, + 451 + ], + "type": "text", + "content": "37. Kabra, R., Zoran, D., Erdogan, G., Matthew, L., Creswell, A., Botvinick, M., Lerchner, A., Burgess, C.: Simone: View-invariant, temporally-abstracted object representations via unsupervised video decomposition. Advances in Neural Information Processing Systems 34, 20146-20159 (2021)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 450, + 482, + 483 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 450, + 482, + 483 + ], + "spans": [ + { + "bbox": [ + 130, + 450, + 482, + 483 + ], + "type": "text", + "content": "38. Kay, W., Carreira, J., Simonyan, K., Zhang, B., Hillier, C., Vijayanarasimhan, S., Viola, F., Green, T., Back, T., Natsev, P., et al.: The kinetics human action video dataset. arXiv preprint arXiv:1705.06950 (2017)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 483, + 482, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 483, + 482, + 525 + ], + "spans": [ + { + "bbox": [ + 130, + 483, + 482, + 525 + ], + "type": "text", + "content": "39. Kipf, T., Elsayed, G.F., Mahendran, A., Stone, A., Sabour, S., Heigold, G., Jonschkowski, R., Dosovitskiy, A., Greff, K.: Conditional object-centric learning from video. In: International Conference on Learning Representations (2022), https://openreview.net/forum?id=aD7uesX1GF_" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 525, + 482, + 548 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 525, + 482, + 548 + ], + "spans": [ + { + "bbox": [ + 130, + 525, + 482, + 548 + ], + "type": "text", + "content": "40. Lafferty, J., McCallum, A., Pereira, F.C.: Conditional random fields: Probabilistic models for segmenting and labeling sequence data (2001)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 548, + 482, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 548, + 482, + 568 + ], + "spans": [ + { + "bbox": [ + 130, + 548, + 482, + 568 + ], + "type": "text", + "content": "41. Lester, B., Al-Rfou, R., Constant, N.: The power of scale for parameter-efficient prompt tuning. arXiv preprint arXiv:2104.08691 (2021)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 568, + 482, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 568, + 482, + 590 + ], + "spans": [ + { + "bbox": [ + 130, + 568, + 482, + 590 + ], + "type": "text", + "content": "42. Li, X.L., Liang, P.: Prefix-tuning: Optimizing continuous prompts for generation. arXiv preprint arXiv:2101.00190 (2021)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 590, + 482, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 590, + 482, + 632 + ], + "spans": [ + { + "bbox": [ + 130, + 590, + 482, + 632 + ], + "type": "text", + "content": "43. Li, Y., Wu, C.Y., Fan, H., Mangalam, K., Xiong, B., Malik, J., Feichtenhofer, C.: Mvitv2: Improved multiscale vision transformers for classification and detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4804-4814 (2022)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 130, + 632, + 482, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 632, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 632, + 482, + 666 + ], + "type": "text", + "content": "44. Liang, F., Wu, B., Dai, X., Li, K., Zhao, Y., Zhang, H., Zhang, P., Vajda, P., Marculescu, D.: Open-vocabulary semantic segmentation with mask-adapted clip. arXiv preprint arXiv:2210.04150 (2022)" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 274, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 274, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 274, + 91, + 447, + 102 + ], + "type": "text", + "content": "Object-centric Image-to-Video Adaptation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "type": "text", + "content": "45. Lin, Z., Wu, Y.F., Peri, S.V., Sun, W., Singh, G., Deng, F., Jiang, J., Ahn, S.: Space: Unsupervised object-oriented scene representation via spatial attention and decomposition. arXiv preprint arXiv:2001.02407 (2020)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 149, + 480, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 149, + 480, + 182 + ], + "spans": [ + { + "bbox": [ + 133, + 149, + 480, + 182 + ], + "type": "text", + "content": "46. Lin, Z., Geng, S., Zhang, R., Gao, P., De Melo, G., Wang, X., Dai, J., Qiao, Y., Li, H.: Frozen clip models are efficient video learners. In: European Conference on Computer Vision. pp. 388-404. Springer (2022)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 182, + 480, + 225 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 182, + 480, + 225 + ], + "spans": [ + { + "bbox": [ + 133, + 182, + 480, + 225 + ], + "type": "text", + "content": "47. Liu, R., Huang, J., Li, G., Feng, J., Wu, X., Li, T.H.: Revisiting temporal modeling for clip-based image-to-video knowledge transferring. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 6555-6564 (June 2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 225, + 480, + 257 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 225, + 480, + 257 + ], + "spans": [ + { + "bbox": [ + 132, + 225, + 480, + 257 + ], + "type": "text", + "content": "48. Liu, X., Ji, K., Fu, Y., Tam, W.L., Du, Z., Yang, Z., Tang, J.: P-tuning v2: Prompt tuning can be comparable to fine-tuning universally across scales and tasks. arXiv preprint arXiv:2110.07602 (2021)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 257, + 480, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 257, + 480, + 300 + ], + "spans": [ + { + "bbox": [ + 132, + 257, + 480, + 300 + ], + "type": "text", + "content": "49. Liu, Z., Lin, Y., Cao, Y., Hu, H., Wei, Y., Zhang, Z., Lin, S., Guo, B.: Swin transformer: Hierarchical vision transformer using shifted windows. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 10012-10022 (2021)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 300, + 480, + 332 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 300, + 480, + 332 + ], + "spans": [ + { + "bbox": [ + 132, + 300, + 480, + 332 + ], + "type": "text", + "content": "50. Liu, Z., Ning, J., Cao, Y., Wei, Y., Zhang, Z., Lin, S., Hu, H.: Video swim transformer. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 3202-3211 (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 332, + 480, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 332, + 480, + 365 + ], + "spans": [ + { + "bbox": [ + 132, + 332, + 480, + 365 + ], + "type": "text", + "content": "51. Locatello, F., Weissenborn, D., Unterthiner, T., Mahendran, A., Heigold, G., Uszkoreit, J., Dosovitskiy, A., Kipf, T.: Object-centric learning with slot attention. Advances in Neural Information Processing Systems 33, 11525-11538 (2020)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 365, + 480, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 365, + 480, + 385 + ], + "spans": [ + { + "bbox": [ + 132, + 365, + 480, + 385 + ], + "type": "text", + "content": "52. Loshchilov, I., Hutter, F.: Decoupled weight decay regularization. In: International Conference on Learning Representations (2018)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 385, + 480, + 429 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 385, + 480, + 429 + ], + "spans": [ + { + "bbox": [ + 132, + 385, + 480, + 429 + ], + "type": "text", + "content": "53. Materzynska, J., Xiao, T., Herzig, R., Xu, H., Wang, X., Darrell, T.: Something-else: Compositional action recognition with spatial-temporal interaction networks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1049–1059 (2020)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 429, + 480, + 461 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 429, + 480, + 461 + ], + "spans": [ + { + "bbox": [ + 132, + 429, + 480, + 461 + ], + "type": "text", + "content": "54. Nagarajan, T., Grauman, K.: Attributes as operators: factorizing unseen attribute-object compositions. In: Proceedings of the European Conference on Computer Vision (ECCV). pp. 169-185 (2018)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 461, + 480, + 504 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 461, + 480, + 504 + ], + "spans": [ + { + "bbox": [ + 132, + 461, + 480, + 504 + ], + "type": "text", + "content": "55. Ni, B., Peng, H., Chen, M., Zhang, S., Meng, G., Fu, J., Xiang, S., Ling, H.: Expanding language-image pretrained models for general video recognition. In: Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part IV. pp. 1-18. Springer (2022)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 504, + 480, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 504, + 480, + 536 + ], + "spans": [ + { + "bbox": [ + 132, + 504, + 480, + 536 + ], + "type": "text", + "content": "56. Oquab, M., Darcet, T., Moutakanni, T., Vo, H., Szafraniec, M., Khalidov, V., Fernandez, P., Haziza, D., Massa, F., El-Nouby, A., et al.: Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193 (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 536, + 480, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 536, + 480, + 568 + ], + "spans": [ + { + "bbox": [ + 132, + 536, + 480, + 568 + ], + "type": "text", + "content": "57. Pan, J., Lin, Z., Zhu, X., Shao, J., Li, H.: St-adapter: Parameter-efficient image-to-video transfer learning for action recognition. arXiv preprint arXiv:2206.13559 (2022)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 568, + 480, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 568, + 480, + 601 + ], + "spans": [ + { + "bbox": [ + 132, + 568, + 480, + 601 + ], + "type": "text", + "content": "58. Park, J., Lee, J., Sohn, K.: Dual-path adaptation from image to video transformers. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 2203-2213 (June 2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 601, + 480, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 601, + 480, + 622 + ], + "spans": [ + { + "bbox": [ + 132, + 601, + 480, + 622 + ], + "type": "text", + "content": "59. Peh, E., Parmar, P., Fernando, B.: Learning to visually connect actions and their effects. arXiv preprint arXiv:2401.10805 (2024)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 622, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 622, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 622, + 480, + 665 + ], + "type": "text", + "content": "60. Pfeiffer, J., Kamath, A., Rückle, A., Cho, K., Gurevych, I.: Adapterfusion: Nondestructive task composition for transfer learning. In: Proceedings of the 16th Conference of the European Chapter of the Association for Computational Linguistics: Main Volume. pp. 487-503 (2021)" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "type": "text", + "content": "R. Qian et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 666 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "text", + "content": "61. Pfeiffer, J., Rückle, A., Poth, C., Kamath, A., Vulic, I., Ruder, S., Cho, K., Gurevych, I.: Adapterhub: A framework for adapting transformers. arXiv preprint arXiv:2007.07779 (2020)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 150, + 482, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 150, + 482, + 184 + ], + "spans": [ + { + "bbox": [ + 130, + 150, + 482, + 184 + ], + "type": "text", + "content": "62. Qian, R., Ding, S., Liu, X., Lin, D.: Static and dynamic concepts for self-supervised video representation learning. In: European Conference on Computer Vision. pp. 145-164. Springer (2022)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 185, + 482, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 185, + 482, + 217 + ], + "spans": [ + { + "bbox": [ + 130, + 185, + 482, + 217 + ], + "type": "text", + "content": "63. Qian, R., Ding, S., Liu, X., Lin, D.: Semantics meets temporal correspondence: Self-supervised object-centric learning in videos. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 16675-16687 (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 218, + 482, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 218, + 482, + 251 + ], + "spans": [ + { + "bbox": [ + 130, + 218, + 482, + 251 + ], + "type": "text", + "content": "64. Qian, R., Dong, X., Zhang, P., Zang, Y., Ding, S., Lin, D., Wang, J.: Streaming long video understanding with large language models. arXiv preprint arXiv:2405.16009 (2024)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 251, + 482, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 251, + 482, + 296 + ], + "spans": [ + { + "bbox": [ + 130, + 251, + 482, + 296 + ], + "type": "text", + "content": "65. Qian, R., Li, Y., Liu, H., See, J., Ding, S., Liu, X., Li, D., Lin, W.: Enhancing self-supervised video representation learning via multi-level feature optimization. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 7990-8001 (2021)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 297, + 482, + 340 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 297, + 482, + 340 + ], + "spans": [ + { + "bbox": [ + 130, + 297, + 482, + 340 + ], + "type": "text", + "content": "66. Qian, R., Meng, T., Gong, B., Yang, M.H., Wang, H., Belongie, S., Cui, Y.: Spatiotemporal contrastive video representation learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6964-6974 (2021)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 341, + 482, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 341, + 482, + 363 + ], + "spans": [ + { + "bbox": [ + 130, + 341, + 482, + 363 + ], + "type": "text", + "content": "67. Qin, G., Eisner, J.: Learning how to ask: Querying lms with mixtures of soft prompts. arXiv preprint arXiv:2104.06599 (2021)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 364, + 482, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 364, + 482, + 407 + ], + "spans": [ + { + "bbox": [ + 130, + 364, + 482, + 407 + ], + "type": "text", + "content": "68. Qing, Z., Zhang, S., Huang, Z., Zhang, Y., Gao, C., Zhao, D., Sang, N.: Disentangling spatial and temporal learning for efficient image-to-video transfer learning. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 13934-13944 (October 2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 408, + 482, + 452 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 408, + 482, + 452 + ], + "spans": [ + { + "bbox": [ + 130, + 408, + 482, + 452 + ], + "type": "text", + "content": "69. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 453, + 482, + 486 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 453, + 482, + 486 + ], + "spans": [ + { + "bbox": [ + 130, + 453, + 482, + 486 + ], + "type": "text", + "content": "70. Shin, T., Razeghi, Y., Logan IV, R.L., Wallace, E., Singh, S.: Autoprompt: Eliciting knowledge from language models with automatically generated prompts. arXiv preprint arXiv:2010.15980 (2020)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 487, + 482, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 487, + 482, + 531 + ], + "spans": [ + { + "bbox": [ + 130, + 487, + 482, + 531 + ], + "type": "text", + "content": "71. Souček, T., Alayrac, J.B., Miech, A., Laptev, I., Sivic, J.: Look for the change: Learning object states and state-modifying actions from untrimmed web videos. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13956-13966 (2022)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 531, + 482, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 531, + 482, + 564 + ], + "spans": [ + { + "bbox": [ + 130, + 531, + 482, + 564 + ], + "type": "text", + "content": "72. Sung, Y.L., Cho, J., Bansal, M.: Vl-adapter: Parameter-efficient transfer learning for vision-and-language tasks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5227-5237 (2022)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 565, + 482, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 565, + 482, + 597 + ], + "spans": [ + { + "bbox": [ + 130, + 565, + 482, + 597 + ], + "type": "text", + "content": "73. Tong, Z., Song, Y., Wang, J., Wang, L.: Videomae: Masked autoencoders are data-efficient learners for self-supervised video pre-training. Advances in neural information processing systems 35, 10078-10093 (2022)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 599, + 482, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 599, + 482, + 632 + ], + "spans": [ + { + "bbox": [ + 130, + 599, + 482, + 632 + ], + "type": "text", + "content": "74. Tran, D., Bourdev, L., Fergus, R., Torresani, L., Paluri, M.: Learning spatiotemporal features with 3d convolutional networks. In: Proceedings of the IEEE international conference on computer vision. pp. 4489-4497 (2015)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 632, + 482, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 632, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 632, + 482, + 666 + ], + "type": "text", + "content": "75. Tran, D., Wang, H., Torresani, L., Ray, J., LeCun, Y., Paluri, M.: A closer look at spatiotemporal convolutions for action recognition. In: Proceedings of the IEEE conference on Computer Vision and Pattern Recognition. pp. 6450-6459 (2018)" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 274, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 274, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 274, + 91, + 447, + 102 + ], + "type": "text", + "content": "Object-centric Image-to-Video Adaptation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "19" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 18 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 481, + 160 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 481, + 160 + ], + "type": "text", + "content": "76. Ventura, C., Bellver, M., Girbau, A., Salvador, A., Marques, F., Giro-i Nieto, X.: Rvos: End-to-end recurrent network for video object segmentation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 5277-5286 (2019)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 160, + 481, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 160, + 481, + 182 + ], + "spans": [ + { + "bbox": [ + 132, + 160, + 481, + 182 + ], + "type": "text", + "content": "77. Wang, M., Xing, J., Liu, Y.: Actionclip: A new paradigm for video action recognition. arXiv preprint arXiv:2109.08472 (2021)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 182, + 481, + 224 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 182, + 481, + 224 + ], + "spans": [ + { + "bbox": [ + 132, + 182, + 481, + 224 + ], + "type": "text", + "content": "78. Wang, W., Bao, H., Dong, L., Bjorck, J., Peng, Z., Liu, Q., Aggarwal, K., Mohammed, O.K., Singhal, S., Som, S., et al.: Image as a foreign language: Beit pretraining for vision and vision-language tasks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 19175-19186 (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 224, + 481, + 257 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 224, + 481, + 257 + ], + "spans": [ + { + "bbox": [ + 132, + 224, + 481, + 257 + ], + "type": "text", + "content": "79. Wang, X., Farhadi, A., Gupta, A.: Actions~ transformations. In: Proceedings of the IEEE conference on Computer Vision and Pattern Recognition. pp. 2658-2667 (2016)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 257, + 481, + 278 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 257, + 481, + 278 + ], + "spans": [ + { + "bbox": [ + 132, + 257, + 481, + 278 + ], + "type": "text", + "content": "80. Wang, X., Gupta, A.: Videos as space-time region graphs. In: Proceedings of the European conference on computer vision (ECCV). pp. 399-417 (2018)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 278, + 481, + 310 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 278, + 481, + 310 + ], + "spans": [ + { + "bbox": [ + 132, + 278, + 481, + 310 + ], + "type": "text", + "content": "81. Wang, X., Misra, I., Zeng, Z., Girdhar, R., Darrell, T.: Videocutler: Surprisingly simple unsupervised video instance segmentation. arXiv preprint arXiv:2308.14710 (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 310, + 481, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 310, + 481, + 342 + ], + "spans": [ + { + "bbox": [ + 132, + 310, + 481, + 342 + ], + "type": "text", + "content": "82. Xie, J., Xie, W., Zisserman, A.: Segmenting moving objects via an object-centric layered representation. In: Advances in Neural Information Processing Systems (2022)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 342, + 481, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 342, + 481, + 386 + ], + "spans": [ + { + "bbox": [ + 132, + 342, + 481, + 386 + ], + "type": "text", + "content": "83. Xu, J., De Mello, S., Liu, S., Byeon, W., Breuel, T., Kautz, J., Wang, X.: Groupvit: Semantic segmentation emerges from text supervision. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18134-18144 (2022)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 386, + 481, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 386, + 481, + 418 + ], + "spans": [ + { + "bbox": [ + 132, + 386, + 481, + 418 + ], + "type": "text", + "content": "84. Yan, S., Xiong, X., Arnab, A., Lu, Z., Zhang, M., Sun, C., Schmid, C.: Multiview transformers for video recognition. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3333-3343 (2022)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 418, + 481, + 450 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 418, + 481, + 450 + ], + "spans": [ + { + "bbox": [ + 132, + 418, + 481, + 450 + ], + "type": "text", + "content": "85. Yang, C., Lamdouar, H., Lu, E., Zisserman, A., Xie, W.: Self-supervised video object segmentation by motion grouping. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 7177-7188 (October 2021)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 450, + 481, + 492 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 450, + 481, + 492 + ], + "spans": [ + { + "bbox": [ + 132, + 450, + 481, + 492 + ], + "type": "text", + "content": "86. Yang, T., Zhu, Y., Xie, Y., Zhang, A., Chen, C., Li, M.: AIM: Adapting image models for efficient video action recognition. In: The Eleventh International Conference on Learning Representations (2023), https://openreview.net/forum?id=CIoSZ_HKHS7" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 492, + 481, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 492, + 481, + 525 + ], + "spans": [ + { + "bbox": [ + 132, + 492, + 481, + 525 + ], + "type": "text", + "content": "87. Yuan, L., Chen, D., Chen, Y.L., Codella, N., Dai, X., Gao, J., Hu, H., Huang, X., Li, B., Li, C., et al.: Florence: A new foundation model for computer vision. arXiv preprint arXiv:2111.11432 (2021)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 525, + 481, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 525, + 481, + 557 + ], + "spans": [ + { + "bbox": [ + 132, + 525, + 481, + 557 + ], + "type": "text", + "content": "88. Zadaianchuk, A., Kleindessner, M., Zhu, Y., Locatello, F., Brox, T.: Unsupervised semantic segmentation with self-supervised object-centric representations. arXiv preprint arXiv:2207.05027 (2022)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 557, + 481, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 557, + 481, + 589 + ], + "spans": [ + { + "bbox": [ + 132, + 557, + 481, + 589 + ], + "type": "text", + "content": "89. Zaken, E.B., Ravfogel, S., Goldberg, Y.: Bitfit: Simple parameter-efficient fine-tuning for transformer-based masked language-models. arXiv preprint arXiv:2106.10199 (2021)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 589, + 481, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 589, + 481, + 622 + ], + "spans": [ + { + "bbox": [ + 132, + 589, + 481, + 622 + ], + "type": "text", + "content": "90. Zareian, A., Rosa, K.D., Hu, D.H., Chang, S.F.: Open-vocabulary object detection using captions. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14393-14402 (2021)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 622, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 622, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 622, + 481, + 665 + ], + "type": "text", + "content": "91. Zhang, C., Fu, C., Wang, S., Agarwal, N., Lee, K., Choi, C., Sun, C.: Object-centric video representation for long-term action anticipation. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 6751-6761 (2024)" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "20" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 223, + 102 + ], + "type": "text", + "content": "R. Qian et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 19 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 259 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "text", + "content": "92. Zhang, C., Gupta, A., Zisserman, A.: Is an object-centric video representation beneficial for transfer? In: Proceedings of the Asian Conference on Computer Vision. pp. 1976-1994 (2022)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 150, + 482, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 150, + 482, + 183 + ], + "spans": [ + { + "bbox": [ + 130, + 150, + 482, + 183 + ], + "type": "text", + "content": "93. Zhang, Y., Tokmakov, P., Hebert, M., Schmid, C.: A structured model for action detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9975-9984 (2019)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 183, + 482, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 183, + 482, + 205 + ], + "spans": [ + { + "bbox": [ + 130, + 183, + 482, + 205 + ], + "type": "text", + "content": "94. Zhou, J., Wei, C., Wang, H., Shen, W., Xie, C., Yuille, A., Kong, T.: ibot: Image bert pre-training with online tokenizer. arXiv preprint arXiv:2111.07832 (2021)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 205, + 482, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 205, + 482, + 237 + ], + "spans": [ + { + "bbox": [ + 130, + 205, + 482, + 237 + ], + "type": "text", + "content": "95. Zhou, K., Yang, J., Loy, C.C., Liu, Z.: Conditional prompt learning for vision-language models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16816-16825 (2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 238, + 482, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 238, + 482, + 259 + ], + "spans": [ + { + "bbox": [ + 130, + 238, + 482, + 259 + ], + "type": "text", + "content": "96. Zhou, K., Yang, J., Loy, C.C., Liu, Z.: Learning to prompt for vision-language models. International Journal of Computer Vision 130(9), 2337-2348 (2022)" + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 274, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 274, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 274, + 91, + 447, + 102 + ], + "type": "text", + "content": "Object-centric Image-to-Video Adaptation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "text", + "content": "21" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 20 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Rethinking LiDAR Domain Generalization_ Single Source as Multiple Density Domains/e618c429-3f75-446b-a895-903ce242d321_content_list.json b/2024/Rethinking LiDAR Domain Generalization_ Single Source as Multiple Density Domains/e618c429-3f75-446b-a895-903ce242d321_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..a1a47e3c7fe1bfcca0e8a08d68dd69089a0efe4a --- /dev/null +++ b/2024/Rethinking LiDAR Domain Generalization_ Single Source as Multiple Density Domains/e618c429-3f75-446b-a895-903ce242d321_content_list.json @@ -0,0 +1,1659 @@ +[ + { + "type": "text", + "text": "Rethinking LiDAR Domain Generalization: Single Source as Multiple Density Domains", + "text_level": 1, + "bbox": [ + 254, + 140, + 750, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jaeyeul Kim, Jungwan Woo, Jeonghoon Kim, and Sunghoon Im", + "bbox": [ + 236, + 212, + 766, + 228 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "DGIST, Daegu, South Korea {jykim94, friendship1, jeonghoon, sunghoonim}@dgist.ac.kr", + "bbox": [ + 279, + 239, + 723, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. In the realm of LiDAR-based perception, significant strides have been made, yet domain generalization remains a substantial challenge. The performance often deteriorates when models are applied to unfamiliar datasets with different LiDAR sensors or deployed in new environments, primarily due to variations in point cloud density distributions. To tackle this challenge, we propose a Density Discriminative Feature Embedding (DDFE) module, capitalizing on the observation that a single source LiDAR point cloud encompasses a spectrum of densities. The DDFE module is meticulously designed to extract density-specific features within a single source domain, facilitating the recognition of objects sharing similar density characteristics across different LiDAR sensors. In addition, we introduce a simple yet effective density augmentation technique aimed at expanding the spectrum of density in source data, thereby enhancing the capabilities of the DDFE. Our DDFE stands out as a versatile and lightweight domain generalization module. It can be seamlessly integrated into various 3D backbone networks, where it has demonstrated superior performance over current state-of-the-art domain generalization methods. Code is available at https://github.com/dgist-cvlab/MultiDensityDG.", + "bbox": [ + 261, + 306, + 743, + 570 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: LiDAR Semantic Segmentation $\\cdot$ Domain Generalization", + "bbox": [ + 261, + 583, + 727, + 597 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 647, + 375, + 662 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Light Detection and Ranging (LiDAR) provides detailed 3D data, making it indispensable for environmental perception in autonomous vehicles. Among the various LiDAR-based perception tasks, semantic segmentation plays a crucial role in understanding the driving scene by classifying each point into multiple classes. While LiDAR-based semantic segmentation [1, 13, 48] have been widely studied, their impressive performance is often constrained to scenarios where source and target datasets align perfectly. However, mismatches between these datasets can lead to significant performance declines. These are mainly attributed to two primary factors: environmental variations as exemplified by the", + "bbox": [ + 212, + 679, + 787, + 816 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "J. Kim and J. Woo—Both authors contributed equally to this work.", + "bbox": [ + 230, + 825, + 686, + 839 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/efa44f2c9d9f265d82a2c66341d6a608c733bc932655c2ab88d10c6a56dee2be.jpg", + "image_caption": [ + "Fig. 1: Motivation of our DDFE: Leveraging Diverse Densities in a Source LiDAR for Domain Generalization - Despite the apparent differences in density distributions between Waymo (64-channel), SemanticKITTI (64-channel), and nuScenes (32-channel), they share regions of overlapping density distributions. For example, the observation that a vehicle at 35 meters in Waymo (top) has a similar density to one at 25 meters in SemanticKITTI (middle) and 12 meters in nuScenes (bottom) underscores this phenomenon. This understanding of varying local density within a source domain serves as a foundation of our domain generalization method. The proposed DDFE transforms features from 3D space to a unified density space without additional training on unseen data, enhancing domain generalization performance." + ], + "image_footnote": [], + "bbox": [ + 218, + 143, + 781, + 299 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Waymo [35] in the USA, the SemanticKITTI [2] in Germany, and the nuScenes [4] in Singapore; and sensor-induced discrepancies, including differences in the number of beams and the field of view (FOV).", + "bbox": [ + 212, + 486, + 784, + 532 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Numerous Unsupervised Domain Adaptation (UDA) studies [17,30,45] address these issues but require additional fine-tuning whenever the target domain changes. In contrast, models deployed with robust Domain Generalization (DG) techniques are increasingly sought after for their potential to enable real-time autonomous driving systems without needing constant fine-tuning. Nevertheless, this crucial field remains under-researched, and its potential has yet to be fully realized. Existing study [16] pinpoints the point cloud density distribution as a prime performance hindrance. While some solutions like point cloud sampling and completion-based methods attempt to remedy this, they fall short in achieving the desired performance [44] and require sequentially labeled data and knowledge of ego-motion [29].", + "bbox": [ + 212, + 534, + 787, + 700 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we introduce a novel perspective in domain generalization by addressing the challenges posed by density variations across different LiDAR sensors. Previous studies [14, 16, 38] have primarily focused on global density differences, typically considering datasets from 64-channel LiDARs like Waymo [35] denser than those from 32-channel LiDARs like nuScenes [4]. However, these approaches oversimplify the inherent complexity of LiDAR data, which exhibits a wide range of density spectra over distance, and interpret them as a single global density value. Contrary to the simplified view of global density comparisons, our research recognizes that LiDAR point clouds are composed of regions", + "bbox": [ + 212, + 704, + 787, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "J. Kim et al.", + "bbox": [ + 271, + 114, + 357, + 126 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "with varying densities, as shown in Fig. 1. These variations are influenced by the distance of objects from the LiDAR sensor, leading to a more complex and nuanced understanding of density within LiDAR data.", + "bbox": [ + 212, + 146, + 782, + 191 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To do so, we propose a Density Discriminative Feature Embedding (DDFE) module, designed to enhance the domain generalization capabilities of LiDAR-based segmentation networks by exploiting these density variations. The DDFE module incorporates a beam density estimation module that encodes the specific densities for each 3D voxel, enabling refined density discrimination across regions. The features are then modulated using attention mechanisms guided by beam density, further enhancing the model's ability to generalize across domains with varying density characteristics. Furthermore, we introduce a density soft clipping technique. This method constrains the density spectrum, ensuring it does not encompass density distributions from unseen domains that are absent in source datasets. To supplement this, we use density augmentation to widen the density spectrum of the source data, thereby enhancing its domain generalization capabilities. Extensive experiments validate the superiority of our method over the conventional domain adaptation and generalization methods across multiple 3D backbone networks.", + "bbox": [ + 212, + 191, + 787, + 417 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In summary, our primary contributions include:", + "bbox": [ + 238, + 417, + 584, + 431 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We introduce a new perspective for domain generalization to overcome density variations caused by different LiDAR sensors, utilizing the diverse densities within point clouds observed in the source domain.", + "- We propose a density discriminative feature embedding module designed to identify areas with similar density distribution and amplify relevant features.", + "- We present a simple yet effective data augmentation strategy, aimed at broadening the density spectrum of source data.", + "- Extensive experiments demonstrate that our method outperforms state-of-the-art Domain Generalization and Domain Adaptation methods." + ], + "bbox": [ + 225, + 443, + 784, + 575 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 601, + 387, + 617 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.1 LiDAR-based Semantic Segmentation", + "text_level": 1, + "bbox": [ + 214, + 633, + 571, + 648 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "LiDAR point clouds pose unique challenges due to their irregular, unordered, and unstructured nature. Consequently, approaches to represent point cloud data have been broadly classified into three categories: Projection-based, point-based, and voxel-based methods. Projection-based methods transform a 3D point cloud into a 2D representation through either a spherical or bird's-eye view projection. By doing so, they can utilize lightweight models like 2D convolution [5,8,25,47] or transformer [1,6] rather than 3D convolution. Nevertheless, such methods face the inherent limitation of 2D kernels not preserving the 3D geometric information of the real world. Point-based methods [9,15,36] directly extract features from the point cloud. These methods utilize all the 3D spatial information without distortion. However, the trade-off is a substantial demand on memory and computational resources. In voxel-based approaches [7,11,13,48], point clouds", + "bbox": [ + 212, + 657, + 787, + 840 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "Rethinking LiDAR Domain Generalization", + "bbox": [ + 444, + 114, + 730, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "table", + "img_path": "images/5fbd05b24345b381007952556cc49952a4bcd689ad0c44a8fc2e7774f5789eef.jpg", + "table_caption": [ + "Table 1: LiDAR configuration for each dataset ( $f_{\\mathrm{min}}$ , $f_{\\mathrm{max}}$ : the minimum and maximum LiDAR field of view, $H_{\\mathrm{b}}$ , $V_{\\mathrm{b}}$ : the number of horizontal and vertical beams)." + ], + "table_footnote": [], + "table_body": "
Waymo [35]SemanticKITTI [2]nuScenes [4]Pandaset [42]SemanticPOSS [24]
Hb25602048108018001800
Vb6464326440
[fmin, fmax](°)[-17.6, +2.4][-24.8, +2.0][-30.0, +10.0][-25.0, +15.0][-16.0, +7.0]
", + "bbox": [ + 245, + 184, + 750, + 232 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "are quantized into 3D grids, known as voxels, and features are extracted using 3D convolutional methods. Alongside sparse convolution [12], voxel-based methods achieve high performance by maintaining an actual geometric receptive field while ensuring an efficient computational load.", + "bbox": [ + 212, + 260, + 782, + 320 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.2 LiDAR-based Domain Adaptation", + "text_level": 1, + "bbox": [ + 215, + 343, + 542, + 358 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Variations in LiDAR sensor configurations, as shown in Tab. 1, or alterations in operating environments can lead to differences in the distribution of point cloud data, which, in turn, may adversely affect the perception capabilities of autonomous driving systems. To tackle this challenge without constructing new training datasets for every sensor type and location, research has increasingly focused on unsupervised domain adaptation techniques [17,26,30,31,39,43,45]. These methods strive to preserve performance in the target domain by utilizing labeled data from a source domain alongside unlabeled data from the target domain. C&L [44] leverages sequential data to train a voxel completion network and segments on the reconstructed canonical domain to overcome domain discrepancies. Rochan et al. [28] propose a range-view-based unsupervised domain adaptation method, aligning beam positions between training and target data. LiDAR-UDA [34] employs beam subsampling to mimic different LiDAR sensors, and utilizes cross-frame ensembling and a Learned Aggregation Model to acquire improved pseudo labels. Although these DA methods demonstrate effectiveness, they share a common limitation: the need for individual fine-tuning with target data upon each domain shift.", + "bbox": [ + 212, + 368, + 787, + 625 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.3 LiDAR-based Domain Generalization", + "text_level": 1, + "bbox": [ + 215, + 648, + 570, + 662 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Domain Generalization [16, 19, 21, 29, 32, 33, 41] aims to improve the performance in scenarios where the target domain data, unseen during training, is encountered. DGLSS [16] leverages sparsity invariance feature consistency and bridges the semantic correlation consistency between the source data and the sparse domain generated by beam sampling. LiDomAug [29] aggregates multiple frames to generate dense world models and augment data by sampling through randomized LiDAR configurations with additional knowledge of ego-motion and the sequentially labeled data. BEV-DG [21] employs a bird's-eye view for enhanced cross-modal learning and develops density-maintained vector modeling for efficient learning of domain-invariant features. LiDOG [32] utilizes semantic priors in a 2D bird's-eye view to extract domain-agnostic features. Sanchez et", + "bbox": [ + 212, + 672, + 787, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "J. Kim et al.", + "bbox": [ + 271, + 114, + 357, + 126 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/7633e91ca9ca3c8c7c047cfbccb0ea6d849ab5c6135972114b1f96e79d9965dd.jpg", + "image_caption": [ + "Fig. 2: Overview of the DDFE pipeline." + ], + "image_footnote": [], + "bbox": [ + 225, + 143, + 779, + 433 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "al. [33] introduce a label propagation method that integrates multi-frame aggregation and ego-motion, although it demands significant computational resources due to the adaptation of KPCConv [36]. Distinctly, our approach advances a single-frame domain generalization approach that considers the inherent characteristics of LiDAR. Consequently, our method offers the distinct advantage of bypassing the need for ego-motion and sequentially labeled data during the training and inference phases.", + "bbox": [ + 212, + 494, + 787, + 602 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3 Method", + "text_level": 1, + "bbox": [ + 215, + 619, + 330, + 636 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we detail our domain generalization method for LiDAR semantic segmentation, which capitalizes on density variations observable within a single source LiDAR. The proposed Density Discriminative Feature Embedding (DDFE) module is primarily composed of four components: Point-voxel feature encoding as detailed in Sec. 3.1, Beam density estimation module in Sec. 3.2, Density soft clipping in Sec. 3.3, and Density-aware embedding module in Sec. 3.4. Additionally, Sec. 3.5 outlines our density augmentation technique. The overall framework of the proposed DDFE module is shown in Fig. 2.", + "bbox": [ + 212, + 648, + 787, + 772 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1 Point-voxel feature encoding", + "text_level": 1, + "bbox": [ + 214, + 787, + 500, + 804 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We introduce a technique to extract generalized representations across domains from point clouds. The intensity distribution in point clouds is influenced by", + "bbox": [ + 212, + 809, + 787, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "Rethinking LiDAR Domain Generalization", + "bbox": [ + 444, + 114, + 730, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "the specific LiDAR sensor used, which can adversely affect the performance of perception models during sensor transitions due to the variance in intensity information. Thus, we omit the use of LiDAR intensity values to enhance domain generalization performance. Furthermore, each LiDAR sensor has a unique distribution of measurement errors, particularly within the intra-voxel region. To minimize variance induced by localized sensing noise, we incorporate a direct voxel-wise feature encoding into the PointNet-based feature extraction [27]. This direct encoding, which bypasses localized information within the voxels, effectively dismisses grid size variations, such as those of $20\\mathrm{cm}$ . Given a point cloud $\\mathbf{P} = \\{p_i\\in \\mathbb{R}^3\\mid i = 1,\\dots,N\\}$ containing $N$ 3D points, we initially partition it into $M$ 3D voxels. For each voxel $v_{j}$ , it is defined as a set of points and is represented by $\\mathbf{V} = \\{v_{j}\\in \\mathbb{R}^{L\\times 3}\\mid j = 1,\\dots,M\\}$ , where $L$ indicates the number of points contained in each voxel. The coordinates of these points are subsequently transformed into the form of $(\\cos (\\theta_j),\\sin (\\theta_j),\\phi_j,r_j)$ , where $(\\theta_j,\\phi_j,r_j)$ correspond to spherical coordinates. These transformed coordinates are then encoded into voxel-wise features $F^{v}\\in \\mathbb{R}^{M\\times 16}$ through an MLP. We also encode point-wise features $F^{p}\\in \\mathbb{R}^{N\\times 16}$ , which are essential to produce point-wise outputs. For each point $k$ in voxel $v_{j}$ centered at $(x_{j}^{c},y_{j}^{c},z_{j}^{c})$ , we compute the offset as $(x_{jk} - x_{j}^{c},y_{jk} - y_{j}^{c},z_{jk} - z_{j}^{c})$ , with $k = 1,\\ldots ,L_{j}$ . Here, $L_{j}$ represents the number of points belonging to $v_{j}$ . These offsets are then processed by a point head to generate point-wise features.", + "bbox": [ + 212, + 146, + 787, + 465 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.2 Beam density estimation module", + "text_level": 1, + "bbox": [ + 215, + 484, + 532, + 500 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this section, we detail the method to compute the density expectation for each ray emitted from a LiDAR sensor. Since all rays from a LiDAR originate from a singular source and are emitted at a fixed angle, the density associated with each beam can be deduced using a spherical projection informed by the beam configuration. The LiDAR sensor configuration, accessible from low-level sources such as an SDK, determines the set of horizontal and vertical inclinations $\\mathbf{C}_h$ and $\\mathbf{C}_v$ for each beam in this manner:", + "bbox": [ + 212, + 508, + 787, + 616 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {C} _ {h} = \\left\\{\\frac {2 \\pi i}{H _ {\\mathrm {b}}} \\right\\} _ {i \\in \\{1, \\dots , H _ {\\mathrm {b}} \\}}, \\mathbf {C} _ {v} = \\left\\{\\frac {(f _ {\\operatorname* {m a x}} - f _ {\\operatorname* {m i n}}) j}{V _ {\\mathrm {b}}} + f _ {\\operatorname* {m i n}} \\right\\} _ {j \\in \\{1, \\dots , V _ {\\mathrm {b}} \\}}, \\qquad (1)\n$$\n", + "text_format": "latex", + "bbox": [ + 248, + 625, + 785, + 661 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $f_{\\mathrm{min}}$ and $f_{\\mathrm{max}}$ denote the minimum and maximum LiDAR field of view, while $H_{\\mathrm{b}}$ and $V_{\\mathrm{b}}$ represent the number of horizontal and vertical beams, detailed in Tab. 1. To project LiDAR beam inclinations onto spherical projected image coordinates, we use the projection function $\\pi : (\\theta, \\phi) \\to (\\dot{\\theta}, \\dot{\\phi})$ defined as:", + "bbox": [ + 212, + 667, + 787, + 731 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\dot {\\theta} = \\left\\lfloor \\frac {\\theta}{2 \\pi} W \\right\\rfloor , \\dot {\\phi} = \\left\\lfloor \\frac {\\phi - f _ {\\operatorname* {m i n}} ^ {\\operatorname* {p r o j}}}{f _ {\\operatorname* {m a x}} ^ {\\operatorname* {p r o j}} - f _ {\\operatorname* {m i n}} ^ {\\operatorname* {p r o j}}} H \\right\\rfloor , \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 366, + 741, + 785, + 781 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $H$ and $W$ are the height and width resolutions of the projected image, respectively set to $H = 512$ , $W = 5120$ . We define the projected image's field of view as $\\left[f_{\\mathrm{min}}^{\\mathrm{proj}}, f_{\\mathrm{max}}^{\\mathrm{proj}}\\right] = [-30.0, 15.0]$ to accommodate a range of LiDAR sensors.", + "bbox": [ + 212, + 789, + 787, + 844 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "J. Kim et al.", + "bbox": [ + 271, + 114, + 357, + 126 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Given this, beam configurations can be transformed into 1-D binary vectors $\\mathbf{B}_v\\in \\mathbb{R}^H$ and $\\mathbf{B}_h\\in \\mathbb{R}^W$ as follows:", + "bbox": [ + 212, + 146, + 782, + 176 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {B} _ {h} (\\dot {\\theta}) = \\mathbb {1} _ {\\mathbf {C} _ {h}} (\\theta), \\mathbf {B} _ {v} (\\dot {\\phi}) = \\mathbb {1} _ {\\mathbf {C} _ {v}} (\\phi), \\text {w h e r e} \\mathbb {1} _ {\\mathbf {C}} (x) = \\left\\{ \\begin{array}{l l} 1, & \\text {i f} x \\in \\mathbf {C} \\\\ 0, & \\text {o t h e r w i s e} \\end{array} , \\right. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 248, + 191, + 784, + 231 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where the indicator function $\\mathbb{1}(\\cdot)$ yields 1 when the azimuth or elevation $(\\theta, \\phi)$ of a pixel $(\\dot{\\theta}, \\dot{\\phi})$ corresponds to the beam inclinations specified in $\\mathbf{C}$ in Eq. (1). Here, $\\mathbf{C}$ can be either $\\mathbf{C}_h$ or $\\mathbf{C}_v$ that represent the locations of projected vertical and horizontal beams, respectively. To compute the beam density, we convolve the binary vectors $\\mathbf{B}_h$ and $\\mathbf{B}_v$ , with four distinct 1-D Gaussian kernels, each characterized by standard deviations $\\sigma_k = \\{10, 30, 50, 70\\}$ as follows:", + "bbox": [ + 212, + 243, + 782, + 335 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {\\mathbf {B}} _ {h} = \\mathbf {B} _ {h} * \\mathbf {G} _ {\\sigma_ {k}}, \\quad \\tilde {\\mathbf {B}} _ {v} = \\mathbf {B} _ {v} * \\mathbf {G} _ {\\sigma_ {k}}. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 383, + 351, + 784, + 366 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Finally, we define beam density $\\mathcal{D}_i$ of point $p_i$ as follows:", + "bbox": [ + 214, + 380, + 625, + 396 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {D} _ {i} = \\left[ \\sqrt {\\hat {\\mathbf {B}} _ {h} ^ {(k)} (\\dot {\\theta} _ {i}) \\cdot \\hat {\\mathbf {B}} _ {v} ^ {(k)} (\\dot {\\phi} _ {i}) / r _ {i} ^ {2}} \\right] _ {k = 1} ^ {4}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 372, + 411, + 784, + 446 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $r_i$ is the radial distance of a LiDAR point $p_i$ in the spherical coordinates, and $[\\cdot]_{k=1}^4$ denotes the concatenation operation. Taking into account that the density of the beam diminishes in proportion to the square of the distance due to its radial emission, we incorporate the inverse of $r^2$ in our computation.", + "bbox": [ + 212, + 458, + 782, + 518 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.3 Density soft clipping", + "text_level": 1, + "bbox": [ + 214, + 549, + 433, + 565 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "When exposed to densities that deviate from those in the source dataset, the model is susceptible to performance degradation in an unseen domain. To alleviate this issue, we introduce a density soft clipping method that confines the density spectrum using the $\\tanh(\\cdot)$ function as follows:", + "bbox": [ + 212, + 580, + 782, + 643 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {D} _ {i} ^ {c} = \\tanh \\left(\\frac {\\mathcal {D} _ {i} - m}{l}\\right) l + m, \\\\ m = \\frac {\\mathcal {P} _ {9 0} (\\mathcal {D}) + \\mathcal {P} _ {1 0} (\\mathcal {D})}{2}, l = \\frac {\\mathcal {P} _ {9 0} (\\mathcal {D}) - \\mathcal {P} _ {1 0} (\\mathcal {D})}{2}, \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 338, + 657, + 663, + 720 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $\\mathcal{D}_i^c$ is clipped density of point $p_i$ , and $\\mathcal{P}_{90}$ and $\\mathcal{P}_{10}$ are functions that extract $90^{\\mathrm{th}}$ and $10^{\\mathrm{th}}$ channel-wise percentile values from the density embedding of the training domain, respectively. Employing these percentile values allows us to discount the outlier density values. To calculate the percentile on the fly under memory constraints, we adopt the Reservoir Sampling [37] technique with a sample size $N = 1000$ . Please refer to the supplementary material for a more detailed algorithm description.", + "bbox": [ + 212, + 733, + 782, + 840 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "Rethinking LiDAR Domain Generalization", + "bbox": [ + 444, + 114, + 730, + 127 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 114, + 784, + 125 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.4 Density-aware embedding module", + "text_level": 1, + "bbox": [ + 215, + 146, + 540, + 162 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "For each point $p_i$ , a point-wise density embedding feature $\\mathcal{D}_i^c$ is obtained via the beam density estimation module incorporating density soft clipping. This feature serves as an input of both a point-wise attention function $f_p$ and a voxelwise attention function $f_v$ . These mechanisms are pivotal for crafting domain-invariant density-discriminative features. In the point-wise attention framework, $\\mathcal{D}_i^c$ is synchronized dimensionally with the point-wise feature $F_i^p$ , facilitating the creation of a density discriminative feature $\\hat{F}_i^p$ as follows:", + "bbox": [ + 212, + 170, + 787, + 279 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {F} _ {i} ^ {p} = f _ {p} \\left(\\mathcal {D} _ {i} ^ {c}\\right) \\odot F _ {i} ^ {p}, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 434, + 287, + 785, + 306 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $f_{p}$ consists of two 1D convolution layers followed by a sigmoid function. For voxel-wise attention, it begins by averaging the density embedding features for all points encapsulated within a voxel. This aggregated measure then undergoes a similar treatment as follows:", + "bbox": [ + 212, + 315, + 784, + 377 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {F} _ {j} ^ {v} = \\operatorname {C o n c a t} \\left(f _ {v} \\left(\\mathcal {D} _ {j} ^ {c}\\right) \\odot F _ {j} ^ {v}, g \\left(F _ {j} ^ {p}\\right)\\right), \\quad \\mathcal {D} _ {j} ^ {c} = \\frac {1}{\\left| L _ {j} \\right|} \\sum_ {p _ {i} \\in v _ {j}} \\mathcal {D} _ {i} ^ {c}, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 305, + 383, + 785, + 421 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $\\mathcal{D}_j^c$ is a clipped density of voxel $v_{j}$ and $\\hat{F}_j^v$ is a voxel-wise density discriminative feature and $f_{v}$ consists of two 1D convolution layers followed by a sigmoid function. An aggregation function $g$ for the voxel-wise feature $F_{j}^{v}$ is adjusted by voxel-wise attention and concatenated with the max pooled point-wise feature within the $v_{j}$ . Finally, the extracted features are passed through a single 1D convolutional layer, resulting in 32-channel features. These density-aware features are subsequently fed into a 3D backbone for LiDAR semantic segmentation.", + "bbox": [ + 212, + 431, + 787, + 540 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "3.5 Density Augmentation", + "text_level": 1, + "bbox": [ + 215, + 559, + 449, + 574 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The DDFE module aims to align densities across varied sensor domains by considering each dataset as a collection of multiple density domains. This method, while effective in many scenarios, may face challenges when there's little to no overlap in the density ranges between the source and unseen datasets. To address this potential issue, we introduce a density augmentation method aimed at broadening the density spectrum covered by the training data. Several existing 3D point cloud augmentation strategies, such as [16, 22, 38] outlined in recent studies, attempt to address density variations either directly or by implication. We adapt the Mix3D [22] for our purposes, which we refer to as enhanced-Mix3D, incorporating random translations along the direction of ego-vehicle movement, along with additional rotational transformations to simulate a wider range of density variations. We also employ beam sampling technique [16, 38], selectively eliminating specific LiDAR beams, to amplify the density in the lower direction. During the training phase, we apply both enhanced-Mix3D and beam sampling augmentations with a set probability of 0.5, aiming to effectively simulate diverse density conditions. Please refer to the supplementary material for more detailed implementation.", + "bbox": [ + 212, + 583, + 787, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "J. Kim et al.", + "bbox": [ + 271, + 114, + 357, + 127 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 215, + 143, + 375, + 162 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "In this section, we demonstrate the domain generalization performance of our method through extensive experiments. The implementation details and experimental settings of the proposed method are detailed in Sec. 4.1. The configuration of datasets used for evaluation is detailed in Sec. 4.2. A comparative analysis with recent domain generalization and domain adaptation methods is presented in Sec. 4.3. The effectiveness of individual components within the proposed DDFE and computational cost are analyzed in Section Sec. 4.4. Detailed per-class performance metrics are available in the supplementary material.", + "bbox": [ + 212, + 172, + 787, + 292 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.1 Implementation Details", + "text_level": 1, + "bbox": [ + 215, + 313, + 455, + 328 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We employ the point head inspired by Cylinder3D [48] to produce point-wise outputs as shown in Fig. 2. During the training phase, we incorporate the Lovasz-Softmax loss $\\mathcal{L}^{lovasz}$ [3] along with the Weighted Cross-Entropy loss $\\mathcal{L}^{wce}$ . The total loss is composed of an equal-weighted combination of point-wise and voxelwise losses as follows:", + "bbox": [ + 212, + 334, + 785, + 409 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {t o t a l}} = \\mathcal {L} _ {\\text {p o i n t}} ^ {\\text {l o v a s z}} + \\mathcal {L} _ {\\text {p o i n t}} ^ {\\text {w c e}} + \\mathcal {L} _ {\\text {v o x e l}} ^ {\\text {l o v a s z}} + \\mathcal {L} _ {\\text {v o x e l}} ^ {\\text {w c e}}. \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 346, + 417, + 785, + 435 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We use the Adam optimizer with an initial learning rate of 1e-3. This rate is decreased by a factor of 0.99 with every epoch. The training is conducted over 30 epochs with a batch size of two on a single NVIDIA RTX 3090, and the epoch yielding the highest source validation mIoU is chosen. For voxelization, we adopt a cubic size of [20cm, 20cm, 20cm]. In cases where multiple point labels are found within a voxel, the voxel label is determined based on the predominant label, aligning with the approach in Cylinder3D [48].", + "bbox": [ + 212, + 441, + 787, + 547 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.2 Datasets", + "text_level": 1, + "bbox": [ + 215, + 566, + 334, + 579 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "In our experimental setup, we employ three datasets: Waymo [35], nuScenes [4], and SemanticKITTI [2]. Notably, the Waymo and SemanticKITTI datasets use 64-channel LiDAR, while the nuScenes employs a 32-channel LiDAR. For detailed information, please refer to Tab. 1. In line with methods from previous studies [16,29], we split the nuScenes sequences into 700 for training and 150 for validation. The Waymo dataset is partitioned into 798 sequences for training and 202 for validation. Regarding the SemanticKITTI dataset, we use sequences 00 to 10 for training, setting aside sequence 08 exclusively for validation. In integrating class variations across datasets, we follow the class mapping configurations from previous studies for fair comparisons.", + "bbox": [ + 212, + 587, + 787, + 739 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.3 Comparison to State-of-the-Art DA/DG Methods", + "text_level": 1, + "bbox": [ + 215, + 757, + 671, + 772 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "The field of domain generalization of LiDAR-based semantic segmentation is still in its formative stages, with a lack of universally accepted experimental standards. This situation has led to a diversity of experimental designs across different studies. To facilitate fair comparisons, we align our experimental framework", + "bbox": [ + 212, + 779, + 785, + 839 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "Rethinking LiDAR Domain Generalization", + "bbox": [ + 444, + 114, + 730, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/9d1421215022dfa6138e74ccca9562741d63cd87b966cb42e87064228c917591.jpg", + "table_caption": [ + "Table 2: Comparison with domain generalization methods based on MinkowskiNet architecture using the mIoU. The best and the second best results are highlighted in bold and underline, respectively." + ], + "table_footnote": [], + "table_body": "
MethodDASourceWKNSourceKWNSourceNWK
Base75.3749.4047.8357.3135.2437.4265.7838.6536.24
IBN-Net [23]75.4751.1344.7257.7436.9938.7465.3136.5336.93
MLDG [20]72.4748.9448.6456.2635.3936.7761.3236.3332.70
COSMIX (W) [30]✓---K49.3539.4638.94N---
COSMIX (K) [30]✓66.6844.7149.96------
COSMIX (N) [30]✓65.6840.9947.9849.9838.0543.25---
DGLSS [16]75.2851.2349.6159.6240.6744.8365.3240.9338.98
Ours76.1557.0756.7562.5042.7349.4368.1645.9846.52
", + "bbox": [ + 217, + 196, + 776, + 300 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/aa77591b405832bb7019b813abe4833e3f4d73c0636381fa260c8e4176ba33e4.jpg", + "table_caption": [ + "Table 3: Comparison with domain adaptation and data augmentation methods." + ], + "table_footnote": [], + "table_body": "
BackboneMethodsK→NN→KBackboneMethodsK→NN→K
MinkNet42 [7]Base37.836.1C&L [44]Base27.923.5
CutMix [46]37.137.6SWD [18]27.724.5
Copy-Paste [10]38.541.13DGCA [39]27.423.9
Mix3D [22]43.144.7C&L [44]31.633.7
PolarMix [40]45.839.1LiDomAug [29]39.237.9
LiDomAug [29]45.948.3LiDAR-UDA [34]41.834.0
Ours (v=5cm)48.651.3Ours (v=5cm)42.541.0
Ours (v=20cm)50.146.3Ours (v=20cm)47.140.3
", + "bbox": [ + 271, + 338, + 723, + 439 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "with those utilized in seminal works such as LiDomAug [29] and DGLSS [16]. This approach allows us to benchmark our method against a range of Domain Adaptation (DA) and Domain Generalization (DG) techniques, showcasing the effectiveness of our method in the context of evolving domain generalization challenges.", + "bbox": [ + 212, + 467, + 784, + 544 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Experiments in the DGLSS Setting We compare our method with a domain adaptation method [30] and three domain generalization methods [16,20,23], following the experimental framework defined in DGLSS [16] using Waymo (W), SemanticKITTI (K), and nuScenes (N). We adopt MinkowskiNet [7] as our backbone network, aligning with the configuration utilized in DGLSS. The results in Tab. 2 indicate that the proposed method consistently outperforms DGLSS across all datasets. Impressively, our method not only demonstrates superior domain generalization performance but also dominates in the source-to-source settings $(\\mathrm{W} \\rightarrow \\mathrm{W}, \\mathrm{K} \\rightarrow \\mathrm{K}, \\mathrm{N} \\rightarrow \\mathrm{N})$ . Unlike DGLSS, which tailors its augmentation strategies for each specific dataset, our method employs a uniform hyperparameter setting across all datasets, achieving enhanced domain generalization results. Our method achieves an average increase of $+12.9\\%$ over DGLSS for unseen datasets using Waymo as the source data $(\\mathrm{W} \\rightarrow \\mathrm{K}, \\mathrm{W} \\rightarrow \\mathrm{N})$ . With SemanticKITTI as the source, there's an average enhancement of $+7.8\\%$ for other unseen domains $(\\mathrm{K} \\rightarrow \\mathrm{W}, \\mathrm{K} \\rightarrow \\mathrm{N})$ , and employing nuScenes as the source data yields an average increase of $+15.8\\%$ on other unseen datasets $(\\mathrm{N} \\rightarrow \\mathrm{W}, \\mathrm{N} \\rightarrow \\mathrm{K})$ .", + "bbox": [ + 212, + 566, + 785, + 809 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "These results demonstrate the robustness of the proposed method in addressing domain discrepancies.", + "bbox": [ + 212, + 810, + 784, + 839 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "J. Kim et al.", + "bbox": [ + 271, + 114, + 357, + 126 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/6787ce1d22ac50a87da0b61b6d136a34909358897286513db58cda2142795387.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 218, + 143, + 413, + 238 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/2e43adec9e88f675f3b4688dda717c5d4eff99b4d7bc1907473662c54c3536df.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 416, + 143, + 591, + 238 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/c1036684973bcba7064adb4c5b623e79ea12b333591a480939af40ccb6781e8c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 593, + 143, + 771, + 238 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/e98508730dd76647ca5502e4553fcd0d790de32b16677c8fe73d905b8b36f5ed.jpg", + "image_caption": [ + "(a) Ground-truth", + "Fig. 3: Qualitative comparison with MinkNet42 backbone. Top: Model trained on nuScenes, tested on SemanticKITTI $(\\mathrm{N} \\rightarrow \\mathrm{K})$ . Bottom: Trained on SemanticKITTI, tested on nuScenes $(\\mathrm{K} \\rightarrow \\mathrm{N})$ ." + ], + "image_footnote": [], + "bbox": [ + 217, + 241, + 413, + 335 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/c81395f531d12d51e25eec38a164efe7e3008364c6c5a09997dc916df4c834d3.jpg", + "image_caption": [ + "(b) MinkNet42" + ], + "image_footnote": [], + "bbox": [ + 415, + 239, + 591, + 335 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/225e7d0ce32039fdf0db98059c68904d7c80ebbd05d8ae1e2e99d3696907e1ee.jpg", + "image_caption": [ + "(c) MinkNet42+Ours" + ], + "image_footnote": [], + "bbox": [ + 593, + 239, + 771, + 335 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Experiments in the LiDomAug Setting We benchmark the proposed method against various augmentation methods [10,22,40,46], domain adaptation methods [18,34,39,44], and a domain generation method [29]. We adhere to the experimental setup established by LiDomAug [29], employing MinkNet42 [7] and C&L [44] as backbone networks, and utilizing SemanticKITTI and nuScenes datasets for evaluation. Unlike the DGLSS [16] setting which uses a voxel size of $20~\\mathrm{cm}$ , the LiDomAug setting uses a voxel size of $5\\mathrm{cm}$ . We benchmark the proposed method against various augmentation methods [10,22,40,46], domain adaptation methods [18,34,39,44], and a domain generation method [29]. We adhere to the experimental setup established by LiDomAug [29], employing MinkNet42 [7] and C&L [44] as backbone networks, and utilizing SemanticKITTI and nuScenes datasets for evaluation. The comparative analysis in Tab. 3 shows that our method surpasses all considered augmentation methods in both $(\\mathrm{K}\\rightarrow \\mathrm{N})$ and $(\\mathrm{N}\\rightarrow \\mathrm{K})$ configurations. Particularly with MinkNet42 as the backbone, our method achieves a $5.9\\%$ and $5.8\\%$ increase in performance over LiDomAug in the $(\\mathrm{K}\\rightarrow \\mathrm{N})$ and $(\\mathrm{N}\\rightarrow \\mathrm{K})$ scenarios, respectively. This enhancement is significant, considering LiDomAug's reliance on ego-motion and multi-frame data integration for domain generalization. When utilizing C&L [44] as the backbone, the proposed method significantly improves mIoU by $+34.5\\%$ and $+8.4\\%$ for the scenario of training on SemanticKITTI and evaluating on nuScenes $(\\mathrm{K}\\rightarrow \\mathrm{N})$ , compared to C&L and LiDomAug, respectively. Moreover, in the transition from the sparse 32-channel dataset (nuScenes) to the denser 64-channel dataset (SemanticKITTI) $(\\mathrm{N}\\rightarrow \\mathrm{K})$ , our method shows a $+21.7\\%$ increase in mIoU over C&L and $+8.2\\%$ improvement over LiDomAug.", + "bbox": [ + 212, + 431, + 787, + 792 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "We further conduct an analysis with a voxel size of $20~\\mathrm{cm}$ , unlike the comparative methods that all use a voxel size of $5\\mathrm{cm}$ . Increasing the voxel size from $5\\mathrm{cm}$ to $20\\mathrm{cm}$ results in reductions of $30.3\\%$ in training time and $62.5\\%$ in infer-", + "bbox": [ + 212, + 795, + 787, + 839 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "Rethinking LiDAR Domain Generalization", + "bbox": [ + 444, + 114, + 730, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/4194e98979e14ddf5fb84004a2ebcf4656897d1e713063a25c7dbf2813061609.jpg", + "table_caption": [ + "Table 4: Ablation study on individual components within our method using mIoU. Experiments evaluate the model's performance with and without the following: (a) Point-voxel encoding, (b) Density-aware embedding module, (c) Density clipping, and (d) Density-Augmentation. All experiments were conducted using a voxel size of 20 cm." + ], + "table_footnote": [], + "table_body": "
(a)(b)(c)(d)K→NN→K
40.731.4
✓43.0 (+5.7%)35.0 (+11.5%)
✓✓45.7 (+12.3%)40.5 (+29.0%)
✓✓✓46.2 (+13.5%)41.8 (+33.1%)
✓✓✓✓50.1 (+23.1%)46.3 (+47.5%)
", + "bbox": [ + 341, + 224, + 651, + 314 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "ence time, while maintaining comparable performance—with a slight decrease of $3.5\\%$ on the MinkNet42 backbone and an increase of $4.7\\%$ on the C&L backbone. Based on these results, we recommend a voxel size of $20~\\mathrm{cm}$ ( $\\mathrm{v} = 20\\mathrm{cm}$ ) as the default setting for our proposed method. The effectiveness is further illustrated in the qualitative results depicted in Fig. 3, highlighting the substantial improvements our method brings to the domain generalization performance of the 3D backbone network. Please refer to the supplementary material for detailed results, including per-class IoU.", + "bbox": [ + 212, + 338, + 787, + 460 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "4.4 Ablation Studies", + "text_level": 1, + "bbox": [ + 215, + 479, + 401, + 493 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "To validate the impact of individual components within our method, we conduct ablation studies. These studies concentrate on the evaluation of point-voxel encoding, the density-aware embedding module, density soft clipping, and density augmentation, providing insights into the significance of each element in enhancing the overall methodology.", + "bbox": [ + 212, + 502, + 785, + 578 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Analysis of the individual components of our method In our comprehensive ablation study in Tab. 4, we thoroughly investigate the individual components of our method. We use the experimental setup proposed by LiDomAug [29] and employ MinkowskiNet [7] as the backbone. This study dissects the impact of our DDFE, which includes (a) point-voxel encoding, (b) density-aware embedding module, and (c) density clipping. Additionally, it examines (d) the impact of our density augmentation method.", + "bbox": [ + 212, + 597, + 785, + 703 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "The integration of all these elements results in a significant performance uplift $+13.5\\%$ in the $(\\mathrm{K}\\rightarrow \\mathrm{N})$ scenario and $+33.1\\%$ in the $(\\mathrm{N}\\rightarrow \\mathrm{K})$ scenario over the baseline model. Notably, the density-aware embedding module stands out for its substantial effect, closely followed by the point-voxel encoding, highlighting its critical role. Furthermore, the study validates the utility of density soft clipping in enhancing model flexibility towards novel density configurations. Meanwhile, the gains observed with density clipping highlight the inherent challenge in adapting to unfamiliar density landscapes, emphasizing the nuanced contributions of each component towards achieving robust domain generalization. Lastly,", + "bbox": [ + 212, + 704, + 787, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "J. Kim et al.", + "bbox": [ + 271, + 114, + 357, + 126 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/276384e1d0fa284ec1809955af3ce82c2d527930ea45ffb90e932d2f711413f5.jpg", + "table_caption": [ + "Table 5: Comparison of data augmentation methods in the proposed DDFE. The best and the second best results are highlighted in bold and underline, respectively. 'D. V.' refers to Density Variation situation, and 'B. S.' to Beam Sampling augmentation." + ], + "table_footnote": [], + "table_body": "
MethodD. V.K→NN→K
DDFE45.941.8
+ PolarMix [40]48.2 (+5.0%)42.3 (+1.2%)
+ B. S. [38]✓49.1 (+7.0%)42.2 (+0.9%)
+ Mix3D [22]✓47.4 (+3.2%)44.8 (+7.1%)
+ Mix3D + B. S. [38]✓49.1 (+7.0%)44.2 (+5.7%)
+ E-Mix3D✓48.5 (+5.6%)46.4 (+11.0%)
+ E-Mix3D + B. S. [38]✓49.4 (+7.6%)46.5 (+11.2%)
", + "bbox": [ + 295, + 198, + 704, + 313 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/aaa37abd30689c928f31cf4afde08f9c61fc9cb7a99125d9da3c9f7ca4249fc1.jpg", + "image_caption": [ + "Fig. 4: Visualization of feature similarity matrices between the Waymo (64-channel) and the nuScenes (32-channel) datasets, with a focus on the distance of objects from the LiDAR. We utilize models trained with the nuScenes dataset for visualization. (a) A point-voxel feature encoding method. (b) A density-aware embedding module. (c) A density soft clipping. Our DDFE combines all modules (a), (b), and (c)." + ], + "image_footnote": [], + "bbox": [ + 246, + 327, + 480, + 440 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/8b0eedd32fe01fce7a217da4bb30d1e6e6ae5558322b10573c27e363a2c914b3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 486, + 327, + 756, + 439 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "incorporating density augmentation into DDFE leads to further enhancements, with an additional $8.4\\%$ performance increase in the $(\\mathrm{K} \\rightarrow \\mathrm{N})$ scenario and $10.8\\%$ in the $(\\mathrm{N} \\rightarrow \\mathrm{K})$ scenario. This demonstrates the effectiveness of the proposed density augmentation scheme in significantly boosting the performance of DDFE.", + "bbox": [ + 212, + 550, + 787, + 612 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Augmentation Tab. 5 shows the performance comparison when various augmentation methods are applied to the proposed DDFE. The experimental setup follows DGLSS [16], and the baseline involves applying the proposed DDFE to MinkowskiNet [7]. The proposed enhanced-Mix3D shows a $2.3\\%$ performance improvement over the original Mix3D [22] when trained on SemanticKITTI and tested on nuScenes $(\\mathrm{K}\\rightarrow \\mathrm{N})$ , and a $3.6\\%$ improvement when trained on nuScenes and tested on SemanticKITTI $(\\mathrm{N}\\rightarrow \\mathrm{K})$ . In conjunction with the beam sampling method, it achieves a $7.6\\%$ performance improvement in $(\\mathrm{K}\\rightarrow \\mathrm{N})$ and $11.2\\%$ in $(\\mathrm{N}\\rightarrow \\mathrm{K})$ compared to the baseline.", + "bbox": [ + 214, + 635, + 787, + 772 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Analysis of Density Discriminative Feature Embedding (DDFE) We delve into the effectiveness of the DDFE module by examining the feature similarity between the source dataset, nuScenes, and the unseen dataset, Waymo,", + "bbox": [ + 214, + 794, + 787, + 842 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "Rethinking LiDAR Domain Generalization", + "bbox": [ + 444, + 114, + 730, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 784, + 126 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "as depicted in Fig. 4. Our focus is on the input voxel features $\\hat{F}^v$ of the 3D backbone model generated by the DDFE. We initiate our analysis by setting a baseline that utilizes a point-voxel feature encoding method for comparison. Following this, we proceed to assess our enhanced method, which integrates a density-aware embedding module along with density soft clipping, to understand its impact on bridging the domain gap. Similarity metrics are derived from the L2 distance between averaged features across specified distance intervals (e.g., 0-5m, ..., 45-50m) for each dataset. According to Fig. 4, the baseline consistently shows higher feature similarity across distances, influenced by its direct embedding of 3D point coordinates. Conversely, our integrated model with density-aware embedding and soft clipping significantly aligns feature similarity across consistent density distributions. It demonstrates the efficacy of the proposed modules in adapting to differences between the source and unseen datasets.", + "bbox": [ + 212, + 145, + 787, + 343 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Computational Cost The inference time of our method for processing a single frame with the nuScenes is 44ms on an NVIDIA RTX 3090. The inclusion of our DDFE module into the MinkNet42 architecture accounts for an extra 8ms of this computation time, while the base MinkNet42 architecture alone requires 36ms. This integration modestly elevates the model's complexity, introducing approximately 23.8k parameters $(+0.06\\%)$ . Furthermore, the application of density augmentation during training adds an extra 60ms per scan on the nuScenes.", + "bbox": [ + 212, + 378, + 787, + 486 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 522, + 359, + 539 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In this paper, we introduce a new perspective for domain generalization of LiDAR semantic segmentation by exploiting the concept of density diversity within a source domain. Based on this perspective, we propose the Density Discriminative Feature Embedding (DDFE) module. DDFE is designed to incorporate expected densities derived from LiDAR beams into a density-aware feature space, significantly enhancing the model's capability to distinguish between different densities. This novel approach improves the adaptability and accuracy of LiDAR semantic segmentation, enabling them to perform effectively across diverse domains, including those previously unseen. In addition to the core methodology, we also present a simple and effective data augmentation technique that extends the density spectrum of the source data. Extensive experiments on the SemanticKITTI, Waymo, and nuScenes datasets demonstrate the effectiveness of the proposed method in elevating the domain generalization performance of LiDAR semantic segmentation. A limitation of this study is its exclusive focus on semantic segmentation tasks. Thus, in future work, we plan to extend our domain generalization method to encompass other LiDAR-based perception tasks, such as 3D object detection. This extension can involve utilizing object-centric density discriminative features alongside point-wise density embedding.", + "bbox": [ + 212, + 566, + 787, + 840 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "J. Kim et al.", + "bbox": [ + 271, + 114, + 357, + 126 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 217, + 143, + 401, + 162 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "This work was supported by the Digital Innovation Hub project supervised by the Daegu Digital Innovation Promotion Agency(DIP) grant funded by the Korea government(MSIT and Daegu Metropolitan City) in 2023(DBSD1-02, Vision picking system for the logistics industry based on artificial intelligence object recognition) and the National Research Foundation of Korea (NRF) grant funded by the Korea government (MSIT) (No. RS-2023-00210908).", + "bbox": [ + 212, + 176, + 787, + 268 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 215, + 292, + 325, + 308 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Ando, A., Gidaris, S., Bursuc, A., Puy, G., Boulch, A., Marlet, R.: Rangevit: Towards vision transformers for 3d semantic segmentation in autonomous driving. In: Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 5240-5250 (2023)", + "2. Behley, J., Garbade, M., Milioto, A., Quenzel, J., Behnke, S., Stachniss, C., Gall, J.: Semantickitti: A dataset for semantic scene understanding of lidar sequences. In: Proceedings of IEEE International Conference on Computer Vision (ICCV). pp. 9297-9307 (2019)", + "3. Berman, M., Triki, A.R., Blaschko, M.B.: The lovasz-softmax loss: A tractable surrogate for the optimization of the intersection-over-union measure in neural networks. In: Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 4413-4421 (2018)", + "4. Caesar, H., et al.: nuscenes: A multimodal dataset for autonomous driving. In: Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 11621-11631 (2020)", + "5. Chen, T.H., Chang, T.S.: Rangeseg: Range-aware real time segmentation of 3d lidar point clouds. IEEE Transactions on Intelligent Vehicles 7(1), 93-101 (2022). https://doi.org/10.1109/TIV.2021.3085827", + "6. Cheng, H.X., Han, X.F., Xiao, G.Q.: Transrvnet: Lidar semantic segmentation with transformer. IEEE Transactions on Intelligent Transportation Systems 24(6), 5895-5907 (2023). https://doi.org/10.1109/TITS.2023.3248117", + "7. Choy, C., Gwak, J., Savarese, S.: 4d spatio-temporal convnets: Minkowski convolutional neural networks. In: Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 3075-3084 (2019)", + "8. Cortinhal, T., Tzelepis, G., Erdal Aksoy, E.: Salsanext: Fast, uncertainty-aware semantic segmentation of lidar point clouds. In: Bebis, G., Yin, Z., Kim, E., Bender, J., Subr, K., Kwon, B.C., Zhao, J., Kalkofen, D., Baciu, G. (eds.) Advances in Visual Computing. pp. 207-222. Springer International Publishing, Cham (2020). https://doi.org/10.1007/978-3-030-64559-5_16", + "9. Fan, S., Dong, Q., Zhu, F., Lv, Y., Ye, P., Wang, F.Y.: Scf-net: Learning spatial contextual features for large-scale point cloud segmentation. In: Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 14504-14513 (2021)", + "10. Ghiasi, G., Cui, Y., Srinivas, A.J., Qian, R., Lin, T.Y., Cubuk, E.D., Le, Q.V., Zoph, B.: Simple copy-paste is a strong data augmentation method for instance segmentation. In: Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 2918-2928 (2021)" + ], + "bbox": [ + 218, + 325, + 784, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "Rethinking LiDAR Domain Generalization", + "bbox": [ + 444, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "11. Graham, B., Engelcke, M., Van Der Maaten, L.: 3d semantic segmentation with submanifold sparse convolutional networks. In: Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 9224-9232 (2018)", + "12. Graham, B., Van der Maaten, L.: Submanifold sparse convolutional networks. arXiv preprint arXiv:1706.01307 (2017)", + "13. Hou, Y., Zhu, X., Ma, Y., Loy, C.C., Li, Y.: Point-to-voxel knowledge distillation for lidar semantic segmentation. In: Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 8479-8488 (2022)", + "14. Hu, Q., Liu, D., Hu, W.: Density-insensitive unsupervised domain adaption on 3d object detection. In: Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 17556-17566 (2023)", + "15. Hu, Q., Yang, B., Xie, L., Rosa, S., Guo, Y., Wang, Z., Trigoni, N., Markham, A.: Randla-net: Efficient semantic segmentation of large-scale point clouds. In: Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 11108-11117 (2020)", + "16. Kim, H., Kang, Y., Oh, C., Yoon, K.J.: Single domain generalization for lidar semantic segmentation. In: Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 17587-17598 (2023)", + "17. Kong, L., Quader, N., Liong, V.E.: Conda: Unsupervised domain adaptation for lidar segmentation via regularized domain concatenation. In: Proceedings of IEEE International Conference on Robotics and Automation (ICRA). pp. 9338-9345. IEEE (2023)", + "18. Lee, C.Y., Batra, T., Baig, M.H., Ulbricht, D.: Sliced wasserstein discrepancy for unsupervised domain adaptation. In: Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 10285-10295 (2019)", + "19. Lehner, A., Gasperini, S., Marcos-Ramiro, A., Schmidt, M., Mahani, M.A.N., Navab, N., Busam, B., Tombari, F.: 3d-vfield: Adversarial augmentation of point clouds for domain generalization in 3d object detection. In: Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 17295-17304 (2022)", + "20. Li, D., Yang, Y., Song, Y.Z., Hospedales, T.: Learning to generalize: Meta-learning for domain generalization. In: Proceedings of the AAAI Conference on Artificial Intelligence (AAAI). vol. 32 (2018)", + "21. Li, M., Zhang, Y., Ma, X., Qu, Y., Fu, Y.: Bev-dg: Cross-modal learning under bird's-eye view for domain generalization of 3d semantic segmentation. In: Proceedings of IEEE International Conference on Computer Vision (ICCV). pp. 11632-11642 (2023)", + "22. Nekrasov, A., Schult, J., Litany, O., Leibe, B., Engelmann, F.: Mix3d: Out-of-context data augmentation for 3d scenes. In: Proceedings of International Conference on 3D Vision (3DV). pp. 116-125 (2021)", + "23. Pan, X., Luo, P., Shi, J., Tang, X.: Two at once: Enhancing learning and generalization capacities via ibn-net. In: Proceedings of European Conference on Computer Vision (ECCV). pp. 464-479 (2018)", + "24. Pan, Y., Gao, B., Mei, J., Geng, S., Li, C., Zhao, H.: Semanticposs: A point cloud dataset with large quantity of dynamic instances. In: 2020 IEEE Intelligent Vehicles Symposium (IV). pp. 687-693 (2020). https://doi.org/10.1109/IV47402.2020.9304596", + "25. Peng, K., Fei, J., Yang, K., Roitberg, A., Zhang, J., Bieder, F., Heidenreich, P., Stiller, C., Stiefelhagen, R.: Mass: Multi-attentional semantic segmentation of lidar" + ], + "bbox": [ + 217, + 147, + 784, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "J. Kim et al.", + "bbox": [ + 271, + 114, + 357, + 126 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "data for dense top-view understanding. IEEE Transactions on Intelligent Transportation Systems 23(9), 15824-15840 (2022). https://doi.org/10.1109/TITS.2022.3145588", + "26. Peng, X., Zhu, X., Ma, Y.: Cl3d: Unsupervised domain adaptation for cross-lidar 3d detection. In: Proceedings of the AAAI Conference on Artificial Intelligence (AAAI). vol. 37, pp. 2047-2055 (2023)", + "27. Qi, C.R., Su, H., Mo, K., Guibas, L.J.: Pointnet: Deep learning on point sets for 3d classification and segmentation. In: Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 652-660 (2017)", + "28. Rochan, M., Aich, S., Corral-Soto, E.R., Nabatchian, A., Liu, B.: Unsupervised domain adaptation in lidar semantic segmentation with self-supervision and gated adapters. In: Proceedings of IEEE International Conference on Robotics and Automation (ICRA). pp. 2649-2655 (2022)", + "29. Ryu, K., Hwang, S., Park, J.: Instant domain augmentation for lidar semantic segmentation. In: Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 9350-9360 (2023)", + "30. Saltori, C., Galasso, F., Fiameni, G., Sebe, N., Ricci, E., Poiesi, F.: Cosmix: Compositional semantic mix for domain adaptation in 3d lidar segmentation. In: Proceedings of European Conference on Computer Vision (ECCV). pp. 586-602. Springer (2022)", + "31. Saltori, C., Lathuilière, S., Sebe, N., Ricci, E., Galasso, F.: Sf-uda 3d: Source-free unsupervised domain adaptation for lidar-based 3d object detection. In: Proceedings of International Conference on 3D Vision (3DV). pp. 771–780 (2020)", + "32. Saltori, C., Osep, A., Ricci, E., Leal-Taixe, L.: Walking your lidog: A journey through multiple domains for lidar semantic segmentation. In: Proceedings of IEEE International Conference on Computer Vision (ICCV). pp. 196-206 (2023)", + "33. Sanchez, J., Deschaud, J.E., Goulette, F.: Domain generalization of 3d semantic segmentation in autonomous driving. In: Proceedings of IEEE International Conference on Computer Vision (ICCV). pp. 18077-18087 (2023)", + "34. Shaban, A., Lee, J., Jung, S., Meng, X., Boots, B.: Lidar-uda: Self-ensembling through time for unsupervised lidar domain adaptation. In: Proceedings of IEEE International Conference on Computer Vision (ICCV). pp. 19784-19794 (2023)", + "35. Sun, P., et al.: Scalability in perception for autonomous driving: Waymo open dataset. In: Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 2446-2454 (2020)", + "36. Thomas, H., Qi, C.R., Deschaud, J.E., Marcotegui, B., Goulette, F., Guibas, L.J.: Kpconv: Flexible and deformable convolution for point clouds. In: Proceedings of IEEE International Conference on Computer Vision (ICCV). pp. 6411-6420 (2019)", + "37. Vitter, J.S.: Random sampling with a reservoir. ACM Trans. Math. Softw. 11(1), 37-57 (mar 1985). https://doi.org/10.1145/3147.3165", + "38. Wei, Y., Wei, Z., Rao, Y., Li, J., Zhou, J., Lu, J.: Lidar distillation: Bridging the beam-induced domain gap for 3d object detection. In: Proceedings of European Conference on Computer Vision (ECCV). pp. 179-195. Springer (2022)", + "39. Wu, B., Zhou, X., Zhao, S., Yue, X., Keutzer, K.: Squeezesegv2: Improved model structure and unsupervised domain adaptation for road-object segmentation from a lidar point cloud. In: Proceedings of IEEE International Conference on Robotics and Automation (ICRA). pp. 4376-4382 (2019)", + "40. Xiao, A., Huang, J., Guan, D., Cui, K., Lu, S., Shao, L.: Polarmix: A general data augmentation technique for lidar point clouds. In: Advances in Neural Information Processing Systems (NeurIPS). pp. 11035-11048 (2022)" + ], + "bbox": [ + 215, + 147, + 784, + 840 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "Rethinking LiDAR Domain Generalization", + "bbox": [ + 444, + 114, + 730, + 127 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "41. Xiao, A., Huang, J., Xuan, W., Ren, R., Liu, K., Guan, D., El Saddik, A., Lu, S., Xing, E.P.: 3d semantic segmentation in the wild: Learning generalized models for adverse-condition point clouds. In: Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 9382-9392 (2023)", + "42. Xiao, P., Shao, Z., Hao, S., Zhang, Z., Chai, X., Jiao, J., Li, Z., Wu, J., Sun, K., Jiang, K., Wang, Y., Yang, D.: Pandaset: Advanced sensor suite dataset for autonomous driving. In: ITSC. pp. 3095-3101 (2021). https://doi.org/10.1109/ITSC48978.2021.9565009", + "43. Yang, J., Shi, S., Wang, Z., Li, H., Qi, X.: St3d: Self-training for unsupervised domain adaptation on 3d object detection. In: Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 10368-10378 (2021)", + "44. Yi, L., Gong, B., Funkhouser, T.: Complete & label: A domain adaptation approach to semantic segmentation of lidar point clouds. In: Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 15363-15373 (2021)", + "45. Yuan, Z., Wen, C., Cheng, M., Su, Y., Liu, W., Yu, S., Wang, C.: Category-level adversaries for outdoor lidar point clouds cross-domain semantic segmentation. IEEE Transactions on Intelligent Transportation Systems 24(2), 1982-1993 (2023). https://doi.org/10.1109/TITS.2022.3219853", + "46. Yun, S., Han, D., Oh, S.J., Chun, S., Choe, J., Yoo, Y.: Cutmix: Regularization strategy to train strong classifiers with localizable features. In: Proceedings of IEEE International Conference on Computer Vision (ICCV). pp. 6023-6032 (2019)", + "47. Zhang, Y., Zhou, Z., David, P., Yue, X., Xi, Z., Gong, B., Foroosh, H.: Polarnet: An improved grid representation for online lidar point clouds semantic segmentation. In: Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 9601-9610 (2020)", + "48. Zhu, X., Zhou, H., Wang, T., Hong, F., Ma, Y., Li, W., Li, H., Lin, D.: Cylindrical and asymmetrical 3d convolution networks for lidar segmentation. In: Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 9939-9948 (2021)" + ], + "bbox": [ + 215, + 146, + 785, + 549 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "J. Kim et al.", + "bbox": [ + 271, + 114, + 357, + 126 + ], + "page_idx": 17 + } +] \ No newline at end of file diff --git a/2024/Rethinking LiDAR Domain Generalization_ Single Source as Multiple Density Domains/e618c429-3f75-446b-a895-903ce242d321_model.json b/2024/Rethinking LiDAR Domain Generalization_ Single Source as Multiple Density Domains/e618c429-3f75-446b-a895-903ce242d321_model.json new file mode 100644 index 0000000000000000000000000000000000000000..5fae640ea79fb0502a44503db0f9320176fac4ba --- /dev/null +++ b/2024/Rethinking LiDAR Domain Generalization_ Single Source as Multiple Density Domains/e618c429-3f75-446b-a895-903ce242d321_model.json @@ -0,0 +1,2260 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.255, + 0.141, + 0.75, + 0.187 + ], + "angle": 0, + "content": "Rethinking LiDAR Domain Generalization: Single Source as Multiple Density Domains" + }, + { + "type": "text", + "bbox": [ + 0.238, + 0.213, + 0.767, + 0.229 + ], + "angle": 0, + "content": "Jaeyeul Kim, Jungwan Woo, Jeonghoon Kim, and Sunghoon Im" + }, + { + "type": "text", + "bbox": [ + 0.28, + 0.24, + 0.724, + 0.27 + ], + "angle": 0, + "content": "DGIST, Daegu, South Korea {jykim94, friendship1, jeonghoon, sunghoonim}@dgist.ac.kr" + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.307, + 0.744, + 0.571 + ], + "angle": 0, + "content": "Abstract. In the realm of LiDAR-based perception, significant strides have been made, yet domain generalization remains a substantial challenge. The performance often deteriorates when models are applied to unfamiliar datasets with different LiDAR sensors or deployed in new environments, primarily due to variations in point cloud density distributions. To tackle this challenge, we propose a Density Discriminative Feature Embedding (DDFE) module, capitalizing on the observation that a single source LiDAR point cloud encompasses a spectrum of densities. The DDFE module is meticulously designed to extract density-specific features within a single source domain, facilitating the recognition of objects sharing similar density characteristics across different LiDAR sensors. In addition, we introduce a simple yet effective density augmentation technique aimed at expanding the spectrum of density in source data, thereby enhancing the capabilities of the DDFE. Our DDFE stands out as a versatile and lightweight domain generalization module. It can be seamlessly integrated into various 3D backbone networks, where it has demonstrated superior performance over current state-of-the-art domain generalization methods. Code is available at https://github.com/dgist-cvlab/MultiDensityDG." + }, + { + "type": "text", + "bbox": [ + 0.263, + 0.584, + 0.728, + 0.598 + ], + "angle": 0, + "content": "Keywords: LiDAR Semantic Segmentation \\(\\cdot\\) Domain Generalization" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.648, + 0.377, + 0.663 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.68, + 0.788, + 0.817 + ], + "angle": 0, + "content": "Light Detection and Ranging (LiDAR) provides detailed 3D data, making it indispensable for environmental perception in autonomous vehicles. Among the various LiDAR-based perception tasks, semantic segmentation plays a crucial role in understanding the driving scene by classifying each point into multiple classes. While LiDAR-based semantic segmentation [1, 13, 48] have been widely studied, their impressive performance is often constrained to scenarios where source and target datasets align perfectly. However, mismatches between these datasets can lead to significant performance declines. These are mainly attributed to two primary factors: environmental variations as exemplified by the" + }, + { + "type": "page_footnote", + "bbox": [ + 0.232, + 0.826, + 0.687, + 0.84 + ], + "angle": 0, + "content": "J. Kim and J. Woo—Both authors contributed equally to this work." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.358, + 0.127 + ], + "angle": 0, + "content": "J. Kim et al." + }, + { + "type": "image", + "bbox": [ + 0.22, + 0.144, + 0.782, + 0.3 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.312, + 0.788, + 0.452 + ], + "angle": 0, + "content": "Fig. 1: Motivation of our DDFE: Leveraging Diverse Densities in a Source LiDAR for Domain Generalization - Despite the apparent differences in density distributions between Waymo (64-channel), SemanticKITTI (64-channel), and nuScenes (32-channel), they share regions of overlapping density distributions. For example, the observation that a vehicle at 35 meters in Waymo (top) has a similar density to one at 25 meters in SemanticKITTI (middle) and 12 meters in nuScenes (bottom) underscores this phenomenon. This understanding of varying local density within a source domain serves as a foundation of our domain generalization method. The proposed DDFE transforms features from 3D space to a unified density space without additional training on unseen data, enhancing domain generalization performance." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.487, + 0.785, + 0.533 + ], + "angle": 0, + "content": "Waymo [35] in the USA, the SemanticKITTI [2] in Germany, and the nuScenes [4] in Singapore; and sensor-induced discrepancies, including differences in the number of beams and the field of view (FOV)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.535, + 0.788, + 0.702 + ], + "angle": 0, + "content": "Numerous Unsupervised Domain Adaptation (UDA) studies [17,30,45] address these issues but require additional fine-tuning whenever the target domain changes. In contrast, models deployed with robust Domain Generalization (DG) techniques are increasingly sought after for their potential to enable real-time autonomous driving systems without needing constant fine-tuning. Nevertheless, this crucial field remains under-researched, and its potential has yet to be fully realized. Existing study [16] pinpoints the point cloud density distribution as a prime performance hindrance. While some solutions like point cloud sampling and completion-based methods attempt to remedy this, they fall short in achieving the desired performance [44] and require sequentially labeled data and knowledge of ego-motion [29]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.705, + 0.788, + 0.842 + ], + "angle": 0, + "content": "In this paper, we introduce a novel perspective in domain generalization by addressing the challenges posed by density variations across different LiDAR sensors. Previous studies [14, 16, 38] have primarily focused on global density differences, typically considering datasets from 64-channel LiDARs like Waymo [35] denser than those from 32-channel LiDARs like nuScenes [4]. However, these approaches oversimplify the inherent complexity of LiDAR data, which exhibits a wide range of density spectra over distance, and interpret them as a single global density value. Contrary to the simplified view of global density comparisons, our research recognizes that LiDAR point clouds are composed of regions" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.446, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Rethinking LiDAR Domain Generalization" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.192 + ], + "angle": 0, + "content": "with varying densities, as shown in Fig. 1. These variations are influenced by the distance of objects from the LiDAR sensor, leading to a more complex and nuanced understanding of density within LiDAR data." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.192, + 0.788, + 0.418 + ], + "angle": 0, + "content": "To do so, we propose a Density Discriminative Feature Embedding (DDFE) module, designed to enhance the domain generalization capabilities of LiDAR-based segmentation networks by exploiting these density variations. The DDFE module incorporates a beam density estimation module that encodes the specific densities for each 3D voxel, enabling refined density discrimination across regions. The features are then modulated using attention mechanisms guided by beam density, further enhancing the model's ability to generalize across domains with varying density characteristics. Furthermore, we introduce a density soft clipping technique. This method constrains the density spectrum, ensuring it does not encompass density distributions from unseen domains that are absent in source datasets. To supplement this, we use density augmentation to widen the density spectrum of the source data, thereby enhancing its domain generalization capabilities. Extensive experiments validate the superiority of our method over the conventional domain adaptation and generalization methods across multiple 3D backbone networks." + }, + { + "type": "text", + "bbox": [ + 0.24, + 0.419, + 0.585, + 0.433 + ], + "angle": 0, + "content": "In summary, our primary contributions include:" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.444, + 0.784, + 0.486 + ], + "angle": 0, + "content": "- We introduce a new perspective for domain generalization to overcome density variations caused by different LiDAR sensors, utilizing the diverse densities within point clouds observed in the source domain." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.489, + 0.785, + 0.517 + ], + "angle": 0, + "content": "- We propose a density discriminative feature embedding module designed to identify areas with similar density distribution and amplify relevant features." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.519, + 0.784, + 0.547 + ], + "angle": 0, + "content": "- We present a simple yet effective data augmentation strategy, aimed at broadening the density spectrum of source data." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.549, + 0.784, + 0.577 + ], + "angle": 0, + "content": "- Extensive experiments demonstrate that our method outperforms state-of-the-art Domain Generalization and Domain Adaptation methods." + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.444, + 0.785, + 0.577 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.602, + 0.388, + 0.618 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.634, + 0.572, + 0.649 + ], + "angle": 0, + "content": "2.1 LiDAR-based Semantic Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.659, + 0.788, + 0.841 + ], + "angle": 0, + "content": "LiDAR point clouds pose unique challenges due to their irregular, unordered, and unstructured nature. Consequently, approaches to represent point cloud data have been broadly classified into three categories: Projection-based, point-based, and voxel-based methods. Projection-based methods transform a 3D point cloud into a 2D representation through either a spherical or bird's-eye view projection. By doing so, they can utilize lightweight models like 2D convolution [5,8,25,47] or transformer [1,6] rather than 3D convolution. Nevertheless, such methods face the inherent limitation of 2D kernels not preserving the 3D geometric information of the real world. Point-based methods [9,15,36] directly extract features from the point cloud. These methods utilize all the 3D spatial information without distortion. However, the trade-off is a substantial demand on memory and computational resources. In voxel-based approaches [7,11,13,48], point clouds" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.358, + 0.127 + ], + "angle": 0, + "content": "J. Kim et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.785, + 0.174 + ], + "angle": 0, + "content": "Table 1: LiDAR configuration for each dataset (\\( f_{\\mathrm{min}} \\), \\( f_{\\mathrm{max}} \\): the minimum and maximum LiDAR field of view, \\( H_{\\mathrm{b}} \\), \\( V_{\\mathrm{b}} \\): the number of horizontal and vertical beams)." + }, + { + "type": "table", + "bbox": [ + 0.246, + 0.185, + 0.751, + 0.233 + ], + "angle": 0, + "content": "
Waymo [35]SemanticKITTI [2]nuScenes [4]Pandaset [42]SemanticPOSS [24]
Hb25602048108018001800
Vb6464326440
[fmin, fmax](°)[-17.6, +2.4][-24.8, +2.0][-30.0, +10.0][-25.0, +15.0][-16.0, +7.0]
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.261, + 0.784, + 0.321 + ], + "angle": 0, + "content": "are quantized into 3D grids, known as voxels, and features are extracted using 3D convolutional methods. Alongside sparse convolution [12], voxel-based methods achieve high performance by maintaining an actual geometric receptive field while ensuring an efficient computational load." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.344, + 0.543, + 0.359 + ], + "angle": 0, + "content": "2.2 LiDAR-based Domain Adaptation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.369, + 0.788, + 0.626 + ], + "angle": 0, + "content": "Variations in LiDAR sensor configurations, as shown in Tab. 1, or alterations in operating environments can lead to differences in the distribution of point cloud data, which, in turn, may adversely affect the perception capabilities of autonomous driving systems. To tackle this challenge without constructing new training datasets for every sensor type and location, research has increasingly focused on unsupervised domain adaptation techniques [17,26,30,31,39,43,45]. These methods strive to preserve performance in the target domain by utilizing labeled data from a source domain alongside unlabeled data from the target domain. C&L [44] leverages sequential data to train a voxel completion network and segments on the reconstructed canonical domain to overcome domain discrepancies. Rochan et al. [28] propose a range-view-based unsupervised domain adaptation method, aligning beam positions between training and target data. LiDAR-UDA [34] employs beam subsampling to mimic different LiDAR sensors, and utilizes cross-frame ensembling and a Learned Aggregation Model to acquire improved pseudo labels. Although these DA methods demonstrate effectiveness, they share a common limitation: the need for individual fine-tuning with target data upon each domain shift." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.649, + 0.571, + 0.663 + ], + "angle": 0, + "content": "2.3 LiDAR-based Domain Generalization" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Domain Generalization [16, 19, 21, 29, 32, 33, 41] aims to improve the performance in scenarios where the target domain data, unseen during training, is encountered. DGLSS [16] leverages sparsity invariance feature consistency and bridges the semantic correlation consistency between the source data and the sparse domain generated by beam sampling. LiDomAug [29] aggregates multiple frames to generate dense world models and augment data by sampling through randomized LiDAR configurations with additional knowledge of ego-motion and the sequentially labeled data. BEV-DG [21] employs a bird's-eye view for enhanced cross-modal learning and develops density-maintained vector modeling for efficient learning of domain-invariant features. LiDOG [32] utilizes semantic priors in a 2D bird's-eye view to extract domain-agnostic features. Sanchez et" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.446, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Rethinking LiDAR Domain Generalization" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "image", + "bbox": [ + 0.227, + 0.144, + 0.78, + 0.434 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.366, + 0.455, + 0.637, + 0.471 + ], + "angle": 0, + "content": "Fig. 2: Overview of the DDFE pipeline." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.496, + 0.788, + 0.603 + ], + "angle": 0, + "content": "al. [33] introduce a label propagation method that integrates multi-frame aggregation and ego-motion, although it demands significant computational resources due to the adaptation of KPCConv [36]. Distinctly, our approach advances a single-frame domain generalization approach that considers the inherent characteristics of LiDAR. Consequently, our method offers the distinct advantage of bypassing the need for ego-motion and sequentially labeled data during the training and inference phases." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.621, + 0.331, + 0.637 + ], + "angle": 0, + "content": "3 Method" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.649, + 0.789, + 0.773 + ], + "angle": 0, + "content": "In this section, we detail our domain generalization method for LiDAR semantic segmentation, which capitalizes on density variations observable within a single source LiDAR. The proposed Density Discriminative Feature Embedding (DDFE) module is primarily composed of four components: Point-voxel feature encoding as detailed in Sec. 3.1, Beam density estimation module in Sec. 3.2, Density soft clipping in Sec. 3.3, and Density-aware embedding module in Sec. 3.4. Additionally, Sec. 3.5 outlines our density augmentation technique. The overall framework of the proposed DDFE module is shown in Fig. 2." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.789, + 0.5, + 0.805 + ], + "angle": 0, + "content": "3.1 Point-voxel feature encoding" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.789, + 0.842 + ], + "angle": 0, + "content": "We introduce a technique to extract generalized representations across domains from point clouds. The intensity distribution in point clouds is influenced by" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.358, + 0.127 + ], + "angle": 0, + "content": "J. Kim et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.466 + ], + "angle": 0, + "content": "the specific LiDAR sensor used, which can adversely affect the performance of perception models during sensor transitions due to the variance in intensity information. Thus, we omit the use of LiDAR intensity values to enhance domain generalization performance. Furthermore, each LiDAR sensor has a unique distribution of measurement errors, particularly within the intra-voxel region. To minimize variance induced by localized sensing noise, we incorporate a direct voxel-wise feature encoding into the PointNet-based feature extraction [27]. This direct encoding, which bypasses localized information within the voxels, effectively dismisses grid size variations, such as those of \\(20\\mathrm{cm}\\). Given a point cloud \\(\\mathbf{P} = \\{p_i\\in \\mathbb{R}^3\\mid i = 1,\\dots,N\\}\\) containing \\(N\\) 3D points, we initially partition it into \\(M\\) 3D voxels. For each voxel \\(v_{j}\\), it is defined as a set of points and is represented by \\(\\mathbf{V} = \\{v_{j}\\in \\mathbb{R}^{L\\times 3}\\mid j = 1,\\dots,M\\}\\), where \\(L\\) indicates the number of points contained in each voxel. The coordinates of these points are subsequently transformed into the form of \\((\\cos (\\theta_j),\\sin (\\theta_j),\\phi_j,r_j)\\), where \\((\\theta_j,\\phi_j,r_j)\\) correspond to spherical coordinates. These transformed coordinates are then encoded into voxel-wise features \\(F^{v}\\in \\mathbb{R}^{M\\times 16}\\) through an MLP. We also encode point-wise features \\(F^{p}\\in \\mathbb{R}^{N\\times 16}\\), which are essential to produce point-wise outputs. For each point \\(k\\) in voxel \\(v_{j}\\) centered at \\((x_{j}^{c},y_{j}^{c},z_{j}^{c})\\), we compute the offset as \\((x_{jk} - x_{j}^{c},y_{jk} - y_{j}^{c},z_{jk} - z_{j}^{c})\\), with \\(k = 1,\\ldots ,L_{j}\\). Here, \\(L_{j}\\) represents the number of points belonging to \\(v_{j}\\). These offsets are then processed by a point head to generate point-wise features." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.486, + 0.533, + 0.501 + ], + "angle": 0, + "content": "3.2 Beam density estimation module" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.51, + 0.788, + 0.617 + ], + "angle": 0, + "content": "In this section, we detail the method to compute the density expectation for each ray emitted from a LiDAR sensor. Since all rays from a LiDAR originate from a singular source and are emitted at a fixed angle, the density associated with each beam can be deduced using a spherical projection informed by the beam configuration. The LiDAR sensor configuration, accessible from low-level sources such as an SDK, determines the set of horizontal and vertical inclinations \\(\\mathbf{C}_h\\) and \\(\\mathbf{C}_v\\) for each beam in this manner:" + }, + { + "type": "equation", + "bbox": [ + 0.249, + 0.625, + 0.786, + 0.662 + ], + "angle": 0, + "content": "\\[\n\\mathbf {C} _ {h} = \\left\\{\\frac {2 \\pi i}{H _ {\\mathrm {b}}} \\right\\} _ {i \\in \\{1, \\dots , H _ {\\mathrm {b}} \\}}, \\mathbf {C} _ {v} = \\left\\{\\frac {(f _ {\\operatorname* {m a x}} - f _ {\\operatorname* {m i n}}) j}{V _ {\\mathrm {b}}} + f _ {\\operatorname* {m i n}} \\right\\} _ {j \\in \\{1, \\dots , V _ {\\mathrm {b}} \\}}, \\qquad (1)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.669, + 0.788, + 0.732 + ], + "angle": 0, + "content": "where \\( f_{\\mathrm{min}} \\) and \\( f_{\\mathrm{max}} \\) denote the minimum and maximum LiDAR field of view, while \\( H_{\\mathrm{b}} \\) and \\( V_{\\mathrm{b}} \\) represent the number of horizontal and vertical beams, detailed in Tab. 1. To project LiDAR beam inclinations onto spherical projected image coordinates, we use the projection function \\( \\pi : (\\theta, \\phi) \\to (\\dot{\\theta}, \\dot{\\phi}) \\) defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.367, + 0.742, + 0.786, + 0.782 + ], + "angle": 0, + "content": "\\[\n\\dot {\\theta} = \\left\\lfloor \\frac {\\theta}{2 \\pi} W \\right\\rfloor , \\dot {\\phi} = \\left\\lfloor \\frac {\\phi - f _ {\\operatorname* {m i n}} ^ {\\operatorname* {p r o j}}}{f _ {\\operatorname* {m a x}} ^ {\\operatorname* {p r o j}} - f _ {\\operatorname* {m i n}} ^ {\\operatorname* {p r o j}}} H \\right\\rfloor , \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.79, + 0.788, + 0.845 + ], + "angle": 0, + "content": "where \\( H \\) and \\( W \\) are the height and width resolutions of the projected image, respectively set to \\( H = 512 \\), \\( W = 5120 \\). We define the projected image's field of view as \\( \\left[f_{\\mathrm{min}}^{\\mathrm{proj}}, f_{\\mathrm{max}}^{\\mathrm{proj}}\\right] = [-30.0, 15.0] \\) to accommodate a range of LiDAR sensors." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.446, + 0.115, + 0.731, + 0.128 + ], + "angle": 0, + "content": "Rethinking LiDAR Domain Generalization" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.126 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.178 + ], + "angle": 0, + "content": "Given this, beam configurations can be transformed into 1-D binary vectors \\(\\mathbf{B}_v\\in \\mathbb{R}^H\\) and \\(\\mathbf{B}_h\\in \\mathbb{R}^W\\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.249, + 0.192, + 0.785, + 0.232 + ], + "angle": 0, + "content": "\\[\n\\mathbf {B} _ {h} (\\dot {\\theta}) = \\mathbb {1} _ {\\mathbf {C} _ {h}} (\\theta), \\mathbf {B} _ {v} (\\dot {\\phi}) = \\mathbb {1} _ {\\mathbf {C} _ {v}} (\\phi), \\text {w h e r e} \\mathbb {1} _ {\\mathbf {C}} (x) = \\left\\{ \\begin{array}{l l} 1, & \\text {i f} x \\in \\mathbf {C} \\\\ 0, & \\text {o t h e r w i s e} \\end{array} , \\right. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.244, + 0.784, + 0.336 + ], + "angle": 0, + "content": "where the indicator function \\(\\mathbb{1}(\\cdot)\\) yields 1 when the azimuth or elevation \\((\\theta, \\phi)\\) of a pixel \\((\\dot{\\theta}, \\dot{\\phi})\\) corresponds to the beam inclinations specified in \\(\\mathbf{C}\\) in Eq. (1). Here, \\(\\mathbf{C}\\) can be either \\(\\mathbf{C}_h\\) or \\(\\mathbf{C}_v\\) that represent the locations of projected vertical and horizontal beams, respectively. To compute the beam density, we convolve the binary vectors \\(\\mathbf{B}_h\\) and \\(\\mathbf{B}_v\\), with four distinct 1-D Gaussian kernels, each characterized by standard deviations \\(\\sigma_k = \\{10, 30, 50, 70\\}\\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.384, + 0.352, + 0.785, + 0.367 + ], + "angle": 0, + "content": "\\[\n\\tilde {\\mathbf {B}} _ {h} = \\mathbf {B} _ {h} * \\mathbf {G} _ {\\sigma_ {k}}, \\quad \\tilde {\\mathbf {B}} _ {v} = \\mathbf {B} _ {v} * \\mathbf {G} _ {\\sigma_ {k}}. \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.381, + 0.627, + 0.397 + ], + "angle": 0, + "content": "Finally, we define beam density \\(\\mathcal{D}_i\\) of point \\(p_i\\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.374, + 0.412, + 0.785, + 0.448 + ], + "angle": 0, + "content": "\\[\n\\mathcal {D} _ {i} = \\left[ \\sqrt {\\hat {\\mathbf {B}} _ {h} ^ {(k)} (\\dot {\\theta} _ {i}) \\cdot \\hat {\\mathbf {B}} _ {v} ^ {(k)} (\\dot {\\phi} _ {i}) / r _ {i} ^ {2}} \\right] _ {k = 1} ^ {4}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.459, + 0.784, + 0.52 + ], + "angle": 0, + "content": "where \\( r_i \\) is the radial distance of a LiDAR point \\( p_i \\) in the spherical coordinates, and \\( [\\cdot]_{k=1}^4 \\) denotes the concatenation operation. Taking into account that the density of the beam diminishes in proportion to the square of the distance due to its radial emission, we incorporate the inverse of \\( r^2 \\) in our computation." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.55, + 0.434, + 0.566 + ], + "angle": 0, + "content": "3.3 Density soft clipping" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.582, + 0.784, + 0.644 + ], + "angle": 0, + "content": "When exposed to densities that deviate from those in the source dataset, the model is susceptible to performance degradation in an unseen domain. To alleviate this issue, we introduce a density soft clipping method that confines the density spectrum using the \\(\\tanh(\\cdot)\\) function as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.339, + 0.658, + 0.665, + 0.721 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {D} _ {i} ^ {c} = \\tanh \\left(\\frac {\\mathcal {D} _ {i} - m}{l}\\right) l + m, \\\\ m = \\frac {\\mathcal {P} _ {9 0} (\\mathcal {D}) + \\mathcal {P} _ {1 0} (\\mathcal {D})}{2}, l = \\frac {\\mathcal {P} _ {9 0} (\\mathcal {D}) - \\mathcal {P} _ {1 0} (\\mathcal {D})}{2}, \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.734, + 0.784, + 0.841 + ], + "angle": 0, + "content": "where \\(\\mathcal{D}_i^c\\) is clipped density of point \\(p_i\\), and \\(\\mathcal{P}_{90}\\) and \\(\\mathcal{P}_{10}\\) are functions that extract \\(90^{\\mathrm{th}}\\) and \\(10^{\\mathrm{th}}\\) channel-wise percentile values from the density embedding of the training domain, respectively. Employing these percentile values allows us to discount the outlier density values. To calculate the percentile on the fly under memory constraints, we adopt the Reservoir Sampling [37] technique with a sample size \\(N = 1000\\). Please refer to the supplementary material for a more detailed algorithm description." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.358, + 0.128 + ], + "angle": 0, + "content": "J. Kim et al." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.147, + 0.542, + 0.163 + ], + "angle": 0, + "content": "3.4 Density-aware embedding module" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.171, + 0.788, + 0.28 + ], + "angle": 0, + "content": "For each point \\( p_i \\), a point-wise density embedding feature \\( \\mathcal{D}_i^c \\) is obtained via the beam density estimation module incorporating density soft clipping. This feature serves as an input of both a point-wise attention function \\( f_p \\) and a voxelwise attention function \\( f_v \\). These mechanisms are pivotal for crafting domain-invariant density-discriminative features. In the point-wise attention framework, \\( \\mathcal{D}_i^c \\) is synchronized dimensionally with the point-wise feature \\( F_i^p \\), facilitating the creation of a density discriminative feature \\( \\hat{F}_i^p \\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.436, + 0.289, + 0.786, + 0.308 + ], + "angle": 0, + "content": "\\[\n\\hat {F} _ {i} ^ {p} = f _ {p} \\left(\\mathcal {D} _ {i} ^ {c}\\right) \\odot F _ {i} ^ {p}, \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.316, + 0.785, + 0.378 + ], + "angle": 0, + "content": "where \\( f_{p} \\) consists of two 1D convolution layers followed by a sigmoid function. For voxel-wise attention, it begins by averaging the density embedding features for all points encapsulated within a voxel. This aggregated measure then undergoes a similar treatment as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.307, + 0.385, + 0.786, + 0.422 + ], + "angle": 0, + "content": "\\[\n\\hat {F} _ {j} ^ {v} = \\operatorname {C o n c a t} \\left(f _ {v} \\left(\\mathcal {D} _ {j} ^ {c}\\right) \\odot F _ {j} ^ {v}, g \\left(F _ {j} ^ {p}\\right)\\right), \\quad \\mathcal {D} _ {j} ^ {c} = \\frac {1}{\\left| L _ {j} \\right|} \\sum_ {p _ {i} \\in v _ {j}} \\mathcal {D} _ {i} ^ {c}, \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.432, + 0.788, + 0.541 + ], + "angle": 0, + "content": "where \\(\\mathcal{D}_j^c\\) is a clipped density of voxel \\(v_{j}\\) and \\(\\hat{F}_j^v\\) is a voxel-wise density discriminative feature and \\(f_{v}\\) consists of two 1D convolution layers followed by a sigmoid function. An aggregation function \\(g\\) for the voxel-wise feature \\(F_{j}^{v}\\) is adjusted by voxel-wise attention and concatenated with the max pooled point-wise feature within the \\(v_{j}\\). Finally, the extracted features are passed through a single 1D convolutional layer, resulting in 32-channel features. These density-aware features are subsequently fed into a 3D backbone for LiDAR semantic segmentation." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.56, + 0.45, + 0.575 + ], + "angle": 0, + "content": "3.5 Density Augmentation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.584, + 0.789, + 0.841 + ], + "angle": 0, + "content": "The DDFE module aims to align densities across varied sensor domains by considering each dataset as a collection of multiple density domains. This method, while effective in many scenarios, may face challenges when there's little to no overlap in the density ranges between the source and unseen datasets. To address this potential issue, we introduce a density augmentation method aimed at broadening the density spectrum covered by the training data. Several existing 3D point cloud augmentation strategies, such as [16, 22, 38] outlined in recent studies, attempt to address density variations either directly or by implication. We adapt the Mix3D [22] for our purposes, which we refer to as enhanced-Mix3D, incorporating random translations along the direction of ego-vehicle movement, along with additional rotational transformations to simulate a wider range of density variations. We also employ beam sampling technique [16, 38], selectively eliminating specific LiDAR beams, to amplify the density in the lower direction. During the training phase, we apply both enhanced-Mix3D and beam sampling augmentations with a set probability of 0.5, aiming to effectively simulate diverse density conditions. Please refer to the supplementary material for more detailed implementation." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.446, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Rethinking LiDAR Domain Generalization" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.145, + 0.376, + 0.163 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.174, + 0.788, + 0.294 + ], + "angle": 0, + "content": "In this section, we demonstrate the domain generalization performance of our method through extensive experiments. The implementation details and experimental settings of the proposed method are detailed in Sec. 4.1. The configuration of datasets used for evaluation is detailed in Sec. 4.2. A comparative analysis with recent domain generalization and domain adaptation methods is presented in Sec. 4.3. The effectiveness of individual components within the proposed DDFE and computational cost are analyzed in Section Sec. 4.4. Detailed per-class performance metrics are available in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.314, + 0.457, + 0.329 + ], + "angle": 0, + "content": "4.1 Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.335, + 0.787, + 0.41 + ], + "angle": 0, + "content": "We employ the point head inspired by Cylinder3D [48] to produce point-wise outputs as shown in Fig. 2. During the training phase, we incorporate the Lovasz-Softmax loss \\(\\mathcal{L}^{lovasz}\\) [3] along with the Weighted Cross-Entropy loss \\(\\mathcal{L}^{wce}\\). The total loss is composed of an equal-weighted combination of point-wise and voxelwise losses as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.347, + 0.418, + 0.786, + 0.436 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {t o t a l}} = \\mathcal {L} _ {\\text {p o i n t}} ^ {\\text {l o v a s z}} + \\mathcal {L} _ {\\text {p o i n t}} ^ {\\text {w c e}} + \\mathcal {L} _ {\\text {v o x e l}} ^ {\\text {l o v a s z}} + \\mathcal {L} _ {\\text {v o x e l}} ^ {\\text {w c e}}. \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.442, + 0.788, + 0.549 + ], + "angle": 0, + "content": "We use the Adam optimizer with an initial learning rate of 1e-3. This rate is decreased by a factor of 0.99 with every epoch. The training is conducted over 30 epochs with a batch size of two on a single NVIDIA RTX 3090, and the epoch yielding the highest source validation mIoU is chosen. For voxelization, we adopt a cubic size of [20cm, 20cm, 20cm]. In cases where multiple point labels are found within a voxel, the voxel label is determined based on the predominant label, aligning with the approach in Cylinder3D [48]." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.567, + 0.335, + 0.58 + ], + "angle": 0, + "content": "4.2 Datasets" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.588, + 0.788, + 0.74 + ], + "angle": 0, + "content": "In our experimental setup, we employ three datasets: Waymo [35], nuScenes [4], and SemanticKITTI [2]. Notably, the Waymo and SemanticKITTI datasets use 64-channel LiDAR, while the nuScenes employs a 32-channel LiDAR. For detailed information, please refer to Tab. 1. In line with methods from previous studies [16,29], we split the nuScenes sequences into 700 for training and 150 for validation. The Waymo dataset is partitioned into 798 sequences for training and 202 for validation. Regarding the SemanticKITTI dataset, we use sequences 00 to 10 for training, setting aside sequence 08 exclusively for validation. In integrating class variations across datasets, we follow the class mapping configurations from previous studies for fair comparisons." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.758, + 0.673, + 0.773 + ], + "angle": 0, + "content": "4.3 Comparison to State-of-the-Art DA/DG Methods" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.787, + 0.84 + ], + "angle": 0, + "content": "The field of domain generalization of LiDAR-based semantic segmentation is still in its formative stages, with a lack of universally accepted experimental standards. This situation has led to a diversity of experimental designs across different studies. To facilitate fair comparisons, we align our experimental framework" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.358, + 0.127 + ], + "angle": 0, + "content": "J. Kim et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.785, + 0.187 + ], + "angle": 0, + "content": "Table 2: Comparison with domain generalization methods based on MinkowskiNet architecture using the mIoU. The best and the second best results are highlighted in bold and underline, respectively." + }, + { + "type": "table", + "bbox": [ + 0.218, + 0.198, + 0.777, + 0.301 + ], + "angle": 0, + "content": "
MethodDASourceWKNSourceKWNSourceNWK
Base75.3749.4047.8357.3135.2437.4265.7838.6536.24
IBN-Net [23]75.4751.1344.7257.7436.9938.7465.3136.5336.93
MLDG [20]72.4748.9448.6456.2635.3936.7761.3236.3332.70
COSMIX (W) [30]✓---K49.3539.4638.94N---
COSMIX (K) [30]✓66.6844.7149.96------
COSMIX (N) [30]✓65.6840.9947.9849.9838.0543.25---
DGLSS [16]75.2851.2349.6159.6240.6744.8365.3240.9338.98
Ours76.1557.0756.7562.5042.7349.4368.1645.9846.52
" + }, + { + "type": "table_caption", + "bbox": [ + 0.233, + 0.313, + 0.768, + 0.327 + ], + "angle": 0, + "content": "Table 3: Comparison with domain adaptation and data augmentation methods." + }, + { + "type": "table", + "bbox": [ + 0.272, + 0.339, + 0.724, + 0.44 + ], + "angle": 0, + "content": "
BackboneMethodsK→NN→KBackboneMethodsK→NN→K
MinkNet42 [7]Base37.836.1C&L [44]Base27.923.5
CutMix [46]37.137.6SWD [18]27.724.5
Copy-Paste [10]38.541.13DGCA [39]27.423.9
Mix3D [22]43.144.7C&L [44]31.633.7
PolarMix [40]45.839.1LiDomAug [29]39.237.9
LiDomAug [29]45.948.3LiDAR-UDA [34]41.834.0
Ours (v=5cm)48.651.3Ours (v=5cm)42.541.0
Ours (v=20cm)50.146.3Ours (v=20cm)47.140.3
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.468, + 0.785, + 0.545 + ], + "angle": 0, + "content": "with those utilized in seminal works such as LiDomAug [29] and DGLSS [16]. This approach allows us to benchmark our method against a range of Domain Adaptation (DA) and Domain Generalization (DG) techniques, showcasing the effectiveness of our method in the context of evolving domain generalization challenges." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.568, + 0.787, + 0.81 + ], + "angle": 0, + "content": "Experiments in the DGLSS Setting We compare our method with a domain adaptation method [30] and three domain generalization methods [16,20,23], following the experimental framework defined in DGLSS [16] using Waymo (W), SemanticKITTI (K), and nuScenes (N). We adopt MinkowskiNet [7] as our backbone network, aligning with the configuration utilized in DGLSS. The results in Tab. 2 indicate that the proposed method consistently outperforms DGLSS across all datasets. Impressively, our method not only demonstrates superior domain generalization performance but also dominates in the source-to-source settings \\((\\mathrm{W} \\rightarrow \\mathrm{W}, \\mathrm{K} \\rightarrow \\mathrm{K}, \\mathrm{N} \\rightarrow \\mathrm{N})\\). Unlike DGLSS, which tailors its augmentation strategies for each specific dataset, our method employs a uniform hyperparameter setting across all datasets, achieving enhanced domain generalization results. Our method achieves an average increase of \\(+12.9\\%\\) over DGLSS for unseen datasets using Waymo as the source data \\((\\mathrm{W} \\rightarrow \\mathrm{K}, \\mathrm{W} \\rightarrow \\mathrm{N})\\). With SemanticKITTI as the source, there's an average enhancement of \\(+7.8\\%\\) for other unseen domains \\((\\mathrm{K} \\rightarrow \\mathrm{W}, \\mathrm{K} \\rightarrow \\mathrm{N})\\), and employing nuScenes as the source data yields an average increase of \\(+15.8\\%\\) on other unseen datasets \\((\\mathrm{N} \\rightarrow \\mathrm{W}, \\mathrm{N} \\rightarrow \\mathrm{K})\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.811, + 0.785, + 0.84 + ], + "angle": 0, + "content": "These results demonstrate the robustness of the proposed method in addressing domain discrepancies." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.446, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Rethinking LiDAR Domain Generalization" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "image", + "bbox": [ + 0.219, + 0.144, + 0.414, + 0.239 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.417, + 0.144, + 0.593, + 0.239 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.594, + 0.144, + 0.772, + 0.239 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.242, + 0.414, + 0.336 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.275, + 0.337, + 0.376, + 0.348 + ], + "angle": 0, + "content": "(a) Ground-truth" + }, + { + "type": "image", + "bbox": [ + 0.416, + 0.241, + 0.593, + 0.336 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.461, + 0.337, + 0.548, + 0.348 + ], + "angle": 0, + "content": "(b) MinkNet42" + }, + { + "type": "image", + "bbox": [ + 0.594, + 0.241, + 0.772, + 0.336 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.622, + 0.337, + 0.744, + 0.348 + ], + "angle": 0, + "content": "(c) MinkNet42+Ours" + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.36, + 0.788, + 0.402 + ], + "angle": 0, + "content": "Fig. 3: Qualitative comparison with MinkNet42 backbone. Top: Model trained on nuScenes, tested on SemanticKITTI \\((\\mathrm{N} \\rightarrow \\mathrm{K})\\). Bottom: Trained on SemanticKITTI, tested on nuScenes \\((\\mathrm{K} \\rightarrow \\mathrm{N})\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.433, + 0.789, + 0.794 + ], + "angle": 0, + "content": "Experiments in the LiDomAug Setting We benchmark the proposed method against various augmentation methods [10,22,40,46], domain adaptation methods [18,34,39,44], and a domain generation method [29]. We adhere to the experimental setup established by LiDomAug [29], employing MinkNet42 [7] and C&L [44] as backbone networks, and utilizing SemanticKITTI and nuScenes datasets for evaluation. Unlike the DGLSS [16] setting which uses a voxel size of \\(20~\\mathrm{cm}\\), the LiDomAug setting uses a voxel size of \\(5\\mathrm{cm}\\). We benchmark the proposed method against various augmentation methods [10,22,40,46], domain adaptation methods [18,34,39,44], and a domain generation method [29]. We adhere to the experimental setup established by LiDomAug [29], employing MinkNet42 [7] and C&L [44] as backbone networks, and utilizing SemanticKITTI and nuScenes datasets for evaluation. The comparative analysis in Tab. 3 shows that our method surpasses all considered augmentation methods in both \\((\\mathrm{K}\\rightarrow \\mathrm{N})\\) and \\((\\mathrm{N}\\rightarrow \\mathrm{K})\\) configurations. Particularly with MinkNet42 as the backbone, our method achieves a \\(5.9\\%\\) and \\(5.8\\%\\) increase in performance over LiDomAug in the \\((\\mathrm{K}\\rightarrow \\mathrm{N})\\) and \\((\\mathrm{N}\\rightarrow \\mathrm{K})\\) scenarios, respectively. This enhancement is significant, considering LiDomAug's reliance on ego-motion and multi-frame data integration for domain generalization. When utilizing C&L [44] as the backbone, the proposed method significantly improves mIoU by \\(+34.5\\%\\) and \\(+8.4\\%\\) for the scenario of training on SemanticKITTI and evaluating on nuScenes \\((\\mathrm{K}\\rightarrow \\mathrm{N})\\), compared to C&L and LiDomAug, respectively. Moreover, in the transition from the sparse 32-channel dataset (nuScenes) to the denser 64-channel dataset (SemanticKITTI) \\((\\mathrm{N}\\rightarrow \\mathrm{K})\\), our method shows a \\(+21.7\\%\\) increase in mIoU over C&L and \\(+8.2\\%\\) improvement over LiDomAug." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.796, + 0.788, + 0.84 + ], + "angle": 0, + "content": "We further conduct an analysis with a voxel size of \\(20~\\mathrm{cm}\\), unlike the comparative methods that all use a voxel size of \\(5\\mathrm{cm}\\). Increasing the voxel size from \\(5\\mathrm{cm}\\) to \\(20\\mathrm{cm}\\) results in reductions of \\(30.3\\%\\) in training time and \\(62.5\\%\\) in infer-" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.358, + 0.127 + ], + "angle": 0, + "content": "J. Kim et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.788, + 0.214 + ], + "angle": 0, + "content": "Table 4: Ablation study on individual components within our method using mIoU. Experiments evaluate the model's performance with and without the following: (a) Point-voxel encoding, (b) Density-aware embedding module, (c) Density clipping, and (d) Density-Augmentation. All experiments were conducted using a voxel size of 20 cm." + }, + { + "type": "table", + "bbox": [ + 0.342, + 0.226, + 0.653, + 0.315 + ], + "angle": 0, + "content": "
(a)(b)(c)(d)K→NN→K
40.731.4
✓43.0 (+5.7%)35.0 (+11.5%)
✓✓45.7 (+12.3%)40.5 (+29.0%)
✓✓✓46.2 (+13.5%)41.8 (+33.1%)
✓✓✓✓50.1 (+23.1%)46.3 (+47.5%)
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.339, + 0.788, + 0.462 + ], + "angle": 0, + "content": "ence time, while maintaining comparable performance—with a slight decrease of \\(3.5\\%\\) on the MinkNet42 backbone and an increase of \\(4.7\\%\\) on the C&L backbone. Based on these results, we recommend a voxel size of \\(20~\\mathrm{cm}\\) (\\(\\mathrm{v} = 20\\mathrm{cm}\\)) as the default setting for our proposed method. The effectiveness is further illustrated in the qualitative results depicted in Fig. 3, highlighting the substantial improvements our method brings to the domain generalization performance of the 3D backbone network. Please refer to the supplementary material for detailed results, including per-class IoU." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.481, + 0.402, + 0.494 + ], + "angle": 0, + "content": "4.4 Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.503, + 0.787, + 0.579 + ], + "angle": 0, + "content": "To validate the impact of individual components within our method, we conduct ablation studies. These studies concentrate on the evaluation of point-voxel encoding, the density-aware embedding module, density soft clipping, and density augmentation, providing insights into the significance of each element in enhancing the overall methodology." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.598, + 0.787, + 0.704 + ], + "angle": 0, + "content": "Analysis of the individual components of our method In our comprehensive ablation study in Tab. 4, we thoroughly investigate the individual components of our method. We use the experimental setup proposed by LiDomAug [29] and employ MinkowskiNet [7] as the backbone. This study dissects the impact of our DDFE, which includes (a) point-voxel encoding, (b) density-aware embedding module, and (c) density clipping. Additionally, it examines (d) the impact of our density augmentation method." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.705, + 0.788, + 0.841 + ], + "angle": 0, + "content": "The integration of all these elements results in a significant performance uplift \\(+13.5\\%\\) in the \\((\\mathrm{K}\\rightarrow \\mathrm{N})\\) scenario and \\(+33.1\\%\\) in the \\((\\mathrm{N}\\rightarrow \\mathrm{K})\\) scenario over the baseline model. Notably, the density-aware embedding module stands out for its substantial effect, closely followed by the point-voxel encoding, highlighting its critical role. Furthermore, the study validates the utility of density soft clipping in enhancing model flexibility towards novel density configurations. Meanwhile, the gains observed with density clipping highlight the inherent challenge in adapting to unfamiliar density landscapes, emphasizing the nuanced contributions of each component towards achieving robust domain generalization. Lastly," + } + ], + [ + { + "type": "header", + "bbox": [ + 0.446, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Rethinking LiDAR Domain Generalization" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "table_caption", + "bbox": [ + 0.215, + 0.145, + 0.788, + 0.189 + ], + "angle": 0, + "content": "Table 5: Comparison of data augmentation methods in the proposed DDFE. The best and the second best results are highlighted in bold and underline, respectively. 'D. V.' refers to Density Variation situation, and 'B. S.' to Beam Sampling augmentation." + }, + { + "type": "table", + "bbox": [ + 0.296, + 0.199, + 0.705, + 0.314 + ], + "angle": 0, + "content": "
MethodD. V.K→NN→K
DDFE45.941.8
+ PolarMix [40]48.2 (+5.0%)42.3 (+1.2%)
+ B. S. [38]✓49.1 (+7.0%)42.2 (+0.9%)
+ Mix3D [22]✓47.4 (+3.2%)44.8 (+7.1%)
+ Mix3D + B. S. [38]✓49.1 (+7.0%)44.2 (+5.7%)
+ E-Mix3D✓48.5 (+5.6%)46.4 (+11.0%)
+ E-Mix3D + B. S. [38]✓49.4 (+7.6%)46.5 (+11.2%)
" + }, + { + "type": "image", + "bbox": [ + 0.248, + 0.328, + 0.482, + 0.441 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.487, + 0.328, + 0.758, + 0.44 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.451, + 0.788, + 0.522 + ], + "angle": 0, + "content": "Fig. 4: Visualization of feature similarity matrices between the Waymo (64-channel) and the nuScenes (32-channel) datasets, with a focus on the distance of objects from the LiDAR. We utilize models trained with the nuScenes dataset for visualization. (a) A point-voxel feature encoding method. (b) A density-aware embedding module. (c) A density soft clipping. Our DDFE combines all modules (a), (b), and (c)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.551, + 0.788, + 0.613 + ], + "angle": 0, + "content": "incorporating density augmentation into DDFE leads to further enhancements, with an additional \\(8.4\\%\\) performance increase in the \\((\\mathrm{K} \\rightarrow \\mathrm{N})\\) scenario and \\(10.8\\%\\) in the \\((\\mathrm{N} \\rightarrow \\mathrm{K})\\) scenario. This demonstrates the effectiveness of the proposed density augmentation scheme in significantly boosting the performance of DDFE." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.636, + 0.788, + 0.773 + ], + "angle": 0, + "content": "Augmentation Tab. 5 shows the performance comparison when various augmentation methods are applied to the proposed DDFE. The experimental setup follows DGLSS [16], and the baseline involves applying the proposed DDFE to MinkowskiNet [7]. The proposed enhanced-Mix3D shows a \\(2.3\\%\\) performance improvement over the original Mix3D [22] when trained on SemanticKITTI and tested on nuScenes \\((\\mathrm{K}\\rightarrow \\mathrm{N})\\), and a \\(3.6\\%\\) improvement when trained on nuScenes and tested on SemanticKITTI \\((\\mathrm{N}\\rightarrow \\mathrm{K})\\). In conjunction with the beam sampling method, it achieves a \\(7.6\\%\\) performance improvement in \\((\\mathrm{K}\\rightarrow \\mathrm{N})\\) and \\(11.2\\%\\) in \\((\\mathrm{N}\\rightarrow \\mathrm{K})\\) compared to the baseline." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.795, + 0.788, + 0.843 + ], + "angle": 0, + "content": "Analysis of Density Discriminative Feature Embedding (DDFE) We delve into the effectiveness of the DDFE module by examining the feature similarity between the source dataset, nuScenes, and the unseen dataset, Waymo," + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.358, + 0.127 + ], + "angle": 0, + "content": "J. Kim et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.146, + 0.788, + 0.344 + ], + "angle": 0, + "content": "as depicted in Fig. 4. Our focus is on the input voxel features \\(\\hat{F}^v\\) of the 3D backbone model generated by the DDFE. We initiate our analysis by setting a baseline that utilizes a point-voxel feature encoding method for comparison. Following this, we proceed to assess our enhanced method, which integrates a density-aware embedding module along with density soft clipping, to understand its impact on bridging the domain gap. Similarity metrics are derived from the L2 distance between averaged features across specified distance intervals (e.g., 0-5m, ..., 45-50m) for each dataset. According to Fig. 4, the baseline consistently shows higher feature similarity across distances, influenced by its direct embedding of 3D point coordinates. Conversely, our integrated model with density-aware embedding and soft clipping significantly aligns feature similarity across consistent density distributions. It demonstrates the efficacy of the proposed modules in adapting to differences between the source and unseen datasets." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.38, + 0.788, + 0.487 + ], + "angle": 0, + "content": "Computational Cost The inference time of our method for processing a single frame with the nuScenes is 44ms on an NVIDIA RTX 3090. The inclusion of our DDFE module into the MinkNet42 architecture accounts for an extra 8ms of this computation time, while the base MinkNet42 architecture alone requires 36ms. This integration modestly elevates the model's complexity, introducing approximately 23.8k parameters \\((+0.06\\%)\\). Furthermore, the application of density augmentation during training adds an extra 60ms per scan on the nuScenes." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.523, + 0.36, + 0.54 + ], + "angle": 0, + "content": "5 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.568, + 0.788, + 0.842 + ], + "angle": 0, + "content": "In this paper, we introduce a new perspective for domain generalization of LiDAR semantic segmentation by exploiting the concept of density diversity within a source domain. Based on this perspective, we propose the Density Discriminative Feature Embedding (DDFE) module. DDFE is designed to incorporate expected densities derived from LiDAR beams into a density-aware feature space, significantly enhancing the model's capability to distinguish between different densities. This novel approach improves the adaptability and accuracy of LiDAR semantic segmentation, enabling them to perform effectively across diverse domains, including those previously unseen. In addition to the core methodology, we also present a simple and effective data augmentation technique that extends the density spectrum of the source data. Extensive experiments on the SemanticKITTI, Waymo, and nuScenes datasets demonstrate the effectiveness of the proposed method in elevating the domain generalization performance of LiDAR semantic segmentation. A limitation of this study is its exclusive focus on semantic segmentation tasks. Thus, in future work, we plan to extend our domain generalization method to encompass other LiDAR-based perception tasks, such as 3D object detection. This extension can involve utilizing object-centric density discriminative features alongside point-wise density embedding." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.446, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Rethinking LiDAR Domain Generalization" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.403, + 0.163 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.178, + 0.788, + 0.27 + ], + "angle": 0, + "content": "This work was supported by the Digital Innovation Hub project supervised by the Daegu Digital Innovation Promotion Agency(DIP) grant funded by the Korea government(MSIT and Daegu Metropolitan City) in 2023(DBSD1-02, Vision picking system for the logistics industry based on artificial intelligence object recognition) and the National Research Foundation of Korea (NRF) grant funded by the Korea government (MSIT) (No. RS-2023-00210908)." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.293, + 0.326, + 0.309 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.326, + 0.785, + 0.38 + ], + "angle": 0, + "content": "1. Ando, A., Gidaris, S., Bursuc, A., Puy, G., Boulch, A., Marlet, R.: Rangevit: Towards vision transformers for 3d semantic segmentation in autonomous driving. In: Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 5240-5250 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.383, + 0.785, + 0.436 + ], + "angle": 0, + "content": "2. Behley, J., Garbade, M., Milioto, A., Quenzel, J., Behnke, S., Stachniss, C., Gall, J.: Semantickitti: A dataset for semantic scene understanding of lidar sequences. In: Proceedings of IEEE International Conference on Computer Vision (ICCV). pp. 9297-9307 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.438, + 0.785, + 0.492 + ], + "angle": 0, + "content": "3. Berman, M., Triki, A.R., Blaschko, M.B.: The lovasz-softmax loss: A tractable surrogate for the optimization of the intersection-over-union measure in neural networks. In: Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 4413-4421 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.494, + 0.785, + 0.534 + ], + "angle": 0, + "content": "4. Caesar, H., et al.: nuscenes: A multimodal dataset for autonomous driving. In: Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 11621-11631 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.535, + 0.785, + 0.576 + ], + "angle": 0, + "content": "5. Chen, T.H., Chang, T.S.: Rangeseg: Range-aware real time segmentation of 3d lidar point clouds. IEEE Transactions on Intelligent Vehicles 7(1), 93-101 (2022). https://doi.org/10.1109/TIV.2021.3085827" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.578, + 0.785, + 0.617 + ], + "angle": 0, + "content": "6. Cheng, H.X., Han, X.F., Xiao, G.Q.: Transrvnet: Lidar semantic segmentation with transformer. IEEE Transactions on Intelligent Transportation Systems 24(6), 5895-5907 (2023). https://doi.org/10.1109/TITS.2023.3248117" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.619, + 0.785, + 0.659 + ], + "angle": 0, + "content": "7. Choy, C., Gwak, J., Savarese, S.: 4d spatio-temporal convnets: Minkowski convolutional neural networks. In: Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 3075-3084 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.661, + 0.785, + 0.728 + ], + "angle": 0, + "content": "8. Cortinhal, T., Tzelepis, G., Erdal Aksoy, E.: Salsanext: Fast, uncertainty-aware semantic segmentation of lidar point clouds. In: Bebis, G., Yin, Z., Kim, E., Bender, J., Subr, K., Kwon, B.C., Zhao, J., Kalkofen, D., Baciu, G. (eds.) Advances in Visual Computing. pp. 207-222. Springer International Publishing, Cham (2020). https://doi.org/10.1007/978-3-030-64559-5_16" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.73, + 0.785, + 0.784 + ], + "angle": 0, + "content": "9. Fan, S., Dong, Q., Zhu, F., Lv, Y., Ye, P., Wang, F.Y.: Scf-net: Learning spatial contextual features for large-scale point cloud segmentation. In: Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 14504-14513 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.785, + 0.785, + 0.84 + ], + "angle": 0, + "content": "10. Ghiasi, G., Cui, Y., Srinivas, A.J., Qian, R., Lin, T.Y., Cubuk, E.D., Le, Q.V., Zoph, B.: Simple copy-paste is a strong data augmentation method for instance segmentation. In: Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 2918-2928 (2021)" + }, + { + "type": "list", + "bbox": [ + 0.22, + 0.326, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.358, + 0.127 + ], + "angle": 0, + "content": "J. Kim et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.148, + 0.785, + 0.189 + ], + "angle": 0, + "content": "11. Graham, B., Engelcke, M., Van Der Maaten, L.: 3d semantic segmentation with submanifold sparse convolutional networks. In: Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 9224-9232 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.191, + 0.785, + 0.218 + ], + "angle": 0, + "content": "12. Graham, B., Van der Maaten, L.: Submanifold sparse convolutional networks. arXiv preprint arXiv:1706.01307 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.22, + 0.785, + 0.261 + ], + "angle": 0, + "content": "13. Hou, Y., Zhu, X., Ma, Y., Loy, C.C., Li, Y.: Point-to-voxel knowledge distillation for lidar semantic segmentation. In: Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 8479-8488 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.263, + 0.785, + 0.303 + ], + "angle": 0, + "content": "14. Hu, Q., Liu, D., Hu, W.: Density-insensitive unsupervised domain adaption on 3d object detection. In: Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 17556-17566 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.305, + 0.785, + 0.359 + ], + "angle": 0, + "content": "15. Hu, Q., Yang, B., Xie, L., Rosa, S., Guo, Y., Wang, Z., Trigoni, N., Markham, A.: Randla-net: Efficient semantic segmentation of large-scale point clouds. In: Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 11108-11117 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.361, + 0.785, + 0.402 + ], + "angle": 0, + "content": "16. Kim, H., Kang, Y., Oh, C., Yoon, K.J.: Single domain generalization for lidar semantic segmentation. In: Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 17587-17598 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.404, + 0.785, + 0.458 + ], + "angle": 0, + "content": "17. Kong, L., Quader, N., Liong, V.E.: Conda: Unsupervised domain adaptation for lidar segmentation via regularized domain concatenation. In: Proceedings of IEEE International Conference on Robotics and Automation (ICRA). pp. 9338-9345. IEEE (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.46, + 0.785, + 0.5 + ], + "angle": 0, + "content": "18. Lee, C.Y., Batra, T., Baig, M.H., Ulbricht, D.: Sliced wasserstein discrepancy for unsupervised domain adaptation. In: Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 10285-10295 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.502, + 0.785, + 0.571 + ], + "angle": 0, + "content": "19. Lehner, A., Gasperini, S., Marcos-Ramiro, A., Schmidt, M., Mahani, M.A.N., Navab, N., Busam, B., Tombari, F.: 3d-vfield: Adversarial augmentation of point clouds for domain generalization in 3d object detection. In: Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 17295-17304 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.573, + 0.785, + 0.613 + ], + "angle": 0, + "content": "20. Li, D., Yang, Y., Song, Y.Z., Hospedales, T.: Learning to generalize: Meta-learning for domain generalization. In: Proceedings of the AAAI Conference on Artificial Intelligence (AAAI). vol. 32 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.615, + 0.785, + 0.669 + ], + "angle": 0, + "content": "21. Li, M., Zhang, Y., Ma, X., Qu, Y., Fu, Y.: Bev-dg: Cross-modal learning under bird's-eye view for domain generalization of 3d semantic segmentation. In: Proceedings of IEEE International Conference on Computer Vision (ICCV). pp. 11632-11642 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.672, + 0.785, + 0.712 + ], + "angle": 0, + "content": "22. Nekrasov, A., Schult, J., Litany, O., Leibe, B., Engelmann, F.: Mix3d: Out-of-context data augmentation for 3d scenes. In: Proceedings of International Conference on 3D Vision (3DV). pp. 116-125 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.714, + 0.785, + 0.755 + ], + "angle": 0, + "content": "23. Pan, X., Luo, P., Shi, J., Tang, X.: Two at once: Enhancing learning and generalization capacities via ibn-net. In: Proceedings of European Conference on Computer Vision (ECCV). pp. 464-479 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.757, + 0.785, + 0.809 + ], + "angle": 0, + "content": "24. Pan, Y., Gao, B., Mei, J., Geng, S., Li, C., Zhao, H.: Semanticposs: A point cloud dataset with large quantity of dynamic instances. In: 2020 IEEE Intelligent Vehicles Symposium (IV). pp. 687-693 (2020). https://doi.org/10.1109/IV47402.2020.9304596" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.812, + 0.785, + 0.84 + ], + "angle": 0, + "content": "25. Peng, K., Fei, J., Yang, K., Roitberg, A., Zhang, J., Bieder, F., Heidenreich, P., Stiller, C., Stiefelhagen, R.: Mass: Multi-attentional semantic segmentation of lidar" + }, + { + "type": "list", + "bbox": [ + 0.218, + 0.148, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.446, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "Rethinking LiDAR Domain Generalization" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.245, + 0.148, + 0.785, + 0.188 + ], + "angle": 0, + "content": "data for dense top-view understanding. IEEE Transactions on Intelligent Transportation Systems 23(9), 15824-15840 (2022). https://doi.org/10.1109/TITS.2022.3145588" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.19, + 0.785, + 0.232 + ], + "angle": 0, + "content": "26. Peng, X., Zhu, X., Ma, Y.: Cl3d: Unsupervised domain adaptation for cross-lidar 3d detection. In: Proceedings of the AAAI Conference on Artificial Intelligence (AAAI). vol. 37, pp. 2047-2055 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.233, + 0.785, + 0.273 + ], + "angle": 0, + "content": "27. Qi, C.R., Su, H., Mo, K., Guibas, L.J.: Pointnet: Deep learning on point sets for 3d classification and segmentation. In: Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 652-660 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.273, + 0.785, + 0.327 + ], + "angle": 0, + "content": "28. Rochan, M., Aich, S., Corral-Soto, E.R., Nabatchian, A., Liu, B.: Unsupervised domain adaptation in lidar semantic segmentation with self-supervision and gated adapters. In: Proceedings of IEEE International Conference on Robotics and Automation (ICRA). pp. 2649-2655 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.328, + 0.785, + 0.37 + ], + "angle": 0, + "content": "29. Ryu, K., Hwang, S., Park, J.: Instant domain augmentation for lidar semantic segmentation. In: Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 9350-9360 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.371, + 0.785, + 0.424 + ], + "angle": 0, + "content": "30. Saltori, C., Galasso, F., Fiameni, G., Sebe, N., Ricci, E., Poiesi, F.: Cosmix: Compositional semantic mix for domain adaptation in 3d lidar segmentation. In: Proceedings of European Conference on Computer Vision (ECCV). pp. 586-602. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.425, + 0.785, + 0.467 + ], + "angle": 0, + "content": "31. Saltori, C., Lathuilière, S., Sebe, N., Ricci, E., Galasso, F.: Sf-uda 3d: Source-free unsupervised domain adaptation for lidar-based 3d object detection. In: Proceedings of International Conference on 3D Vision (3DV). pp. 771–780 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.468, + 0.785, + 0.509 + ], + "angle": 0, + "content": "32. Saltori, C., Osep, A., Ricci, E., Leal-Taixe, L.: Walking your lidog: A journey through multiple domains for lidar semantic segmentation. In: Proceedings of IEEE International Conference on Computer Vision (ICCV). pp. 196-206 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.509, + 0.785, + 0.55 + ], + "angle": 0, + "content": "33. Sanchez, J., Deschaud, J.E., Goulette, F.: Domain generalization of 3d semantic segmentation in autonomous driving. In: Proceedings of IEEE International Conference on Computer Vision (ICCV). pp. 18077-18087 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.55, + 0.785, + 0.592 + ], + "angle": 0, + "content": "34. Shaban, A., Lee, J., Jung, S., Meng, X., Boots, B.: Lidar-uda: Self-ensembling through time for unsupervised lidar domain adaptation. In: Proceedings of IEEE International Conference on Computer Vision (ICCV). pp. 19784-19794 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.592, + 0.785, + 0.633 + ], + "angle": 0, + "content": "35. Sun, P., et al.: Scalability in perception for autonomous driving: Waymo open dataset. In: Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 2446-2454 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.633, + 0.785, + 0.675 + ], + "angle": 0, + "content": "36. Thomas, H., Qi, C.R., Deschaud, J.E., Marcotegui, B., Goulette, F., Guibas, L.J.: Kpconv: Flexible and deformable convolution for point clouds. In: Proceedings of IEEE International Conference on Computer Vision (ICCV). pp. 6411-6420 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.675, + 0.785, + 0.702 + ], + "angle": 0, + "content": "37. Vitter, J.S.: Random sampling with a reservoir. ACM Trans. Math. Softw. 11(1), 37-57 (mar 1985). https://doi.org/10.1145/3147.3165" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.702, + 0.785, + 0.743 + ], + "angle": 0, + "content": "38. Wei, Y., Wei, Z., Rao, Y., Li, J., Zhou, J., Lu, J.: Lidar distillation: Bridging the beam-induced domain gap for 3d object detection. In: Proceedings of European Conference on Computer Vision (ECCV). pp. 179-195. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.744, + 0.785, + 0.799 + ], + "angle": 0, + "content": "39. Wu, B., Zhou, X., Zhao, S., Yue, X., Keutzer, K.: Squeezesegv2: Improved model structure and unsupervised domain adaptation for road-object segmentation from a lidar point cloud. In: Proceedings of IEEE International Conference on Robotics and Automation (ICRA). pp. 4376-4382 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.799, + 0.785, + 0.841 + ], + "angle": 0, + "content": "40. Xiao, A., Huang, J., Guan, D., Cui, K., Lu, S., Shao, L.: Polarmix: A general data augmentation technique for lidar point clouds. In: Advances in Neural Information Processing Systems (NeurIPS). pp. 11035-11048 (2022)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.148, + 0.785, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.358, + 0.127 + ], + "angle": 0, + "content": "J. Kim et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.787, + 0.203 + ], + "angle": 0, + "content": "41. Xiao, A., Huang, J., Xuan, W., Ren, R., Liu, K., Guan, D., El Saddik, A., Lu, S., Xing, E.P.: 3d semantic segmentation in the wild: Learning generalized models for adverse-condition point clouds. In: Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 9382-9392 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.204, + 0.787, + 0.257 + ], + "angle": 0, + "content": "42. Xiao, P., Shao, Z., Hao, S., Zhang, Z., Chai, X., Jiao, J., Li, Z., Wu, J., Sun, K., Jiang, K., Wang, Y., Yang, D.: Pandaset: Advanced sensor suite dataset for autonomous driving. In: ITSC. pp. 3095-3101 (2021). https://doi.org/10.1109/ITSC48978.2021.9565009" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.259, + 0.787, + 0.3 + ], + "angle": 0, + "content": "43. Yang, J., Shi, S., Wang, Z., Li, H., Qi, X.: St3d: Self-training for unsupervised domain adaptation on 3d object detection. In: Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 10368-10378 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.3, + 0.787, + 0.342 + ], + "angle": 0, + "content": "44. Yi, L., Gong, B., Funkhouser, T.: Complete & label: A domain adaptation approach to semantic segmentation of lidar point clouds. In: Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 15363-15373 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.342, + 0.787, + 0.397 + ], + "angle": 0, + "content": "45. Yuan, Z., Wen, C., Cheng, M., Su, Y., Liu, W., Yu, S., Wang, C.: Category-level adversaries for outdoor lidar point clouds cross-domain semantic segmentation. IEEE Transactions on Intelligent Transportation Systems 24(2), 1982-1993 (2023). https://doi.org/10.1109/TITS.2022.3219853" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.397, + 0.787, + 0.439 + ], + "angle": 0, + "content": "46. Yun, S., Han, D., Oh, S.J., Chun, S., Choe, J., Yoo, Y.: Cutmix: Regularization strategy to train strong classifiers with localizable features. In: Proceedings of IEEE International Conference on Computer Vision (ICCV). pp. 6023-6032 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.439, + 0.787, + 0.493 + ], + "angle": 0, + "content": "47. Zhang, Y., Zhou, Z., David, P., Yue, X., Xi, Z., Gong, B., Foroosh, H.: Polarnet: An improved grid representation for online lidar point clouds semantic segmentation. In: Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 9601-9610 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.493, + 0.787, + 0.55 + ], + "angle": 0, + "content": "48. Zhu, X., Zhou, H., Wang, T., Hong, F., Ma, Y., Li, W., Li, H., Lin, D.: Cylindrical and asymmetrical 3d convolution networks for lidar segmentation. In: Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 9939-9948 (2021)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.55 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/Rethinking LiDAR Domain Generalization_ Single Source as Multiple Density Domains/e618c429-3f75-446b-a895-903ce242d321_origin.pdf b/2024/Rethinking LiDAR Domain Generalization_ Single Source as Multiple Density Domains/e618c429-3f75-446b-a895-903ce242d321_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..8f756ed48923a5a7d9b52c02c5c13cd69ff0e923 --- /dev/null +++ b/2024/Rethinking LiDAR Domain Generalization_ Single Source as Multiple Density Domains/e618c429-3f75-446b-a895-903ce242d321_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:148b1ff7d21f59b958f000d69d5874cd191e1ff9562764ad244c04d3c4e774e6 +size 7859160 diff --git a/2024/Rethinking LiDAR Domain Generalization_ Single Source as Multiple Density Domains/full.md b/2024/Rethinking LiDAR Domain Generalization_ Single Source as Multiple Density Domains/full.md new file mode 100644 index 0000000000000000000000000000000000000000..08d25299a9f3bb05971ca7c681b682a55abb2141 --- /dev/null +++ b/2024/Rethinking LiDAR Domain Generalization_ Single Source as Multiple Density Domains/full.md @@ -0,0 +1,286 @@ +# Rethinking LiDAR Domain Generalization: Single Source as Multiple Density Domains + +Jaeyeul Kim, Jungwan Woo, Jeonghoon Kim, and Sunghoon Im + +DGIST, Daegu, South Korea {jykim94, friendship1, jeonghoon, sunghoonim}@dgist.ac.kr + +Abstract. In the realm of LiDAR-based perception, significant strides have been made, yet domain generalization remains a substantial challenge. The performance often deteriorates when models are applied to unfamiliar datasets with different LiDAR sensors or deployed in new environments, primarily due to variations in point cloud density distributions. To tackle this challenge, we propose a Density Discriminative Feature Embedding (DDFE) module, capitalizing on the observation that a single source LiDAR point cloud encompasses a spectrum of densities. The DDFE module is meticulously designed to extract density-specific features within a single source domain, facilitating the recognition of objects sharing similar density characteristics across different LiDAR sensors. In addition, we introduce a simple yet effective density augmentation technique aimed at expanding the spectrum of density in source data, thereby enhancing the capabilities of the DDFE. Our DDFE stands out as a versatile and lightweight domain generalization module. It can be seamlessly integrated into various 3D backbone networks, where it has demonstrated superior performance over current state-of-the-art domain generalization methods. Code is available at https://github.com/dgist-cvlab/MultiDensityDG. + +Keywords: LiDAR Semantic Segmentation $\cdot$ Domain Generalization + +# 1 Introduction + +Light Detection and Ranging (LiDAR) provides detailed 3D data, making it indispensable for environmental perception in autonomous vehicles. Among the various LiDAR-based perception tasks, semantic segmentation plays a crucial role in understanding the driving scene by classifying each point into multiple classes. While LiDAR-based semantic segmentation [1, 13, 48] have been widely studied, their impressive performance is often constrained to scenarios where source and target datasets align perfectly. However, mismatches between these datasets can lead to significant performance declines. These are mainly attributed to two primary factors: environmental variations as exemplified by the + +![](images/efa44f2c9d9f265d82a2c66341d6a608c733bc932655c2ab88d10c6a56dee2be.jpg) +Fig. 1: Motivation of our DDFE: Leveraging Diverse Densities in a Source LiDAR for Domain Generalization - Despite the apparent differences in density distributions between Waymo (64-channel), SemanticKITTI (64-channel), and nuScenes (32-channel), they share regions of overlapping density distributions. For example, the observation that a vehicle at 35 meters in Waymo (top) has a similar density to one at 25 meters in SemanticKITTI (middle) and 12 meters in nuScenes (bottom) underscores this phenomenon. This understanding of varying local density within a source domain serves as a foundation of our domain generalization method. The proposed DDFE transforms features from 3D space to a unified density space without additional training on unseen data, enhancing domain generalization performance. + +Waymo [35] in the USA, the SemanticKITTI [2] in Germany, and the nuScenes [4] in Singapore; and sensor-induced discrepancies, including differences in the number of beams and the field of view (FOV). + +Numerous Unsupervised Domain Adaptation (UDA) studies [17,30,45] address these issues but require additional fine-tuning whenever the target domain changes. In contrast, models deployed with robust Domain Generalization (DG) techniques are increasingly sought after for their potential to enable real-time autonomous driving systems without needing constant fine-tuning. Nevertheless, this crucial field remains under-researched, and its potential has yet to be fully realized. Existing study [16] pinpoints the point cloud density distribution as a prime performance hindrance. While some solutions like point cloud sampling and completion-based methods attempt to remedy this, they fall short in achieving the desired performance [44] and require sequentially labeled data and knowledge of ego-motion [29]. + +In this paper, we introduce a novel perspective in domain generalization by addressing the challenges posed by density variations across different LiDAR sensors. Previous studies [14, 16, 38] have primarily focused on global density differences, typically considering datasets from 64-channel LiDARs like Waymo [35] denser than those from 32-channel LiDARs like nuScenes [4]. However, these approaches oversimplify the inherent complexity of LiDAR data, which exhibits a wide range of density spectra over distance, and interpret them as a single global density value. Contrary to the simplified view of global density comparisons, our research recognizes that LiDAR point clouds are composed of regions + +with varying densities, as shown in Fig. 1. These variations are influenced by the distance of objects from the LiDAR sensor, leading to a more complex and nuanced understanding of density within LiDAR data. + +To do so, we propose a Density Discriminative Feature Embedding (DDFE) module, designed to enhance the domain generalization capabilities of LiDAR-based segmentation networks by exploiting these density variations. The DDFE module incorporates a beam density estimation module that encodes the specific densities for each 3D voxel, enabling refined density discrimination across regions. The features are then modulated using attention mechanisms guided by beam density, further enhancing the model's ability to generalize across domains with varying density characteristics. Furthermore, we introduce a density soft clipping technique. This method constrains the density spectrum, ensuring it does not encompass density distributions from unseen domains that are absent in source datasets. To supplement this, we use density augmentation to widen the density spectrum of the source data, thereby enhancing its domain generalization capabilities. Extensive experiments validate the superiority of our method over the conventional domain adaptation and generalization methods across multiple 3D backbone networks. + +In summary, our primary contributions include: + +- We introduce a new perspective for domain generalization to overcome density variations caused by different LiDAR sensors, utilizing the diverse densities within point clouds observed in the source domain. +- We propose a density discriminative feature embedding module designed to identify areas with similar density distribution and amplify relevant features. +- We present a simple yet effective data augmentation strategy, aimed at broadening the density spectrum of source data. +- Extensive experiments demonstrate that our method outperforms state-of-the-art Domain Generalization and Domain Adaptation methods. + +# 2 Related Work + +# 2.1 LiDAR-based Semantic Segmentation + +LiDAR point clouds pose unique challenges due to their irregular, unordered, and unstructured nature. Consequently, approaches to represent point cloud data have been broadly classified into three categories: Projection-based, point-based, and voxel-based methods. Projection-based methods transform a 3D point cloud into a 2D representation through either a spherical or bird's-eye view projection. By doing so, they can utilize lightweight models like 2D convolution [5,8,25,47] or transformer [1,6] rather than 3D convolution. Nevertheless, such methods face the inherent limitation of 2D kernels not preserving the 3D geometric information of the real world. Point-based methods [9,15,36] directly extract features from the point cloud. These methods utilize all the 3D spatial information without distortion. However, the trade-off is a substantial demand on memory and computational resources. In voxel-based approaches [7,11,13,48], point clouds + +Table 1: LiDAR configuration for each dataset ( $f_{\mathrm{min}}$ , $f_{\mathrm{max}}$ : the minimum and maximum LiDAR field of view, $H_{\mathrm{b}}$ , $V_{\mathrm{b}}$ : the number of horizontal and vertical beams). + +
Waymo [35]SemanticKITTI [2]nuScenes [4]Pandaset [42]SemanticPOSS [24]
Hb25602048108018001800
Vb6464326440
[fmin, fmax](°)[-17.6, +2.4][-24.8, +2.0][-30.0, +10.0][-25.0, +15.0][-16.0, +7.0]
+ +are quantized into 3D grids, known as voxels, and features are extracted using 3D convolutional methods. Alongside sparse convolution [12], voxel-based methods achieve high performance by maintaining an actual geometric receptive field while ensuring an efficient computational load. + +# 2.2 LiDAR-based Domain Adaptation + +Variations in LiDAR sensor configurations, as shown in Tab. 1, or alterations in operating environments can lead to differences in the distribution of point cloud data, which, in turn, may adversely affect the perception capabilities of autonomous driving systems. To tackle this challenge without constructing new training datasets for every sensor type and location, research has increasingly focused on unsupervised domain adaptation techniques [17,26,30,31,39,43,45]. These methods strive to preserve performance in the target domain by utilizing labeled data from a source domain alongside unlabeled data from the target domain. C&L [44] leverages sequential data to train a voxel completion network and segments on the reconstructed canonical domain to overcome domain discrepancies. Rochan et al. [28] propose a range-view-based unsupervised domain adaptation method, aligning beam positions between training and target data. LiDAR-UDA [34] employs beam subsampling to mimic different LiDAR sensors, and utilizes cross-frame ensembling and a Learned Aggregation Model to acquire improved pseudo labels. Although these DA methods demonstrate effectiveness, they share a common limitation: the need for individual fine-tuning with target data upon each domain shift. + +# 2.3 LiDAR-based Domain Generalization + +Domain Generalization [16, 19, 21, 29, 32, 33, 41] aims to improve the performance in scenarios where the target domain data, unseen during training, is encountered. DGLSS [16] leverages sparsity invariance feature consistency and bridges the semantic correlation consistency between the source data and the sparse domain generated by beam sampling. LiDomAug [29] aggregates multiple frames to generate dense world models and augment data by sampling through randomized LiDAR configurations with additional knowledge of ego-motion and the sequentially labeled data. BEV-DG [21] employs a bird's-eye view for enhanced cross-modal learning and develops density-maintained vector modeling for efficient learning of domain-invariant features. LiDOG [32] utilizes semantic priors in a 2D bird's-eye view to extract domain-agnostic features. Sanchez et + +![](images/7633e91ca9ca3c8c7c047cfbccb0ea6d849ab5c6135972114b1f96e79d9965dd.jpg) +Fig. 2: Overview of the DDFE pipeline. + +al. [33] introduce a label propagation method that integrates multi-frame aggregation and ego-motion, although it demands significant computational resources due to the adaptation of KPCConv [36]. Distinctly, our approach advances a single-frame domain generalization approach that considers the inherent characteristics of LiDAR. Consequently, our method offers the distinct advantage of bypassing the need for ego-motion and sequentially labeled data during the training and inference phases. + +# 3 Method + +In this section, we detail our domain generalization method for LiDAR semantic segmentation, which capitalizes on density variations observable within a single source LiDAR. The proposed Density Discriminative Feature Embedding (DDFE) module is primarily composed of four components: Point-voxel feature encoding as detailed in Sec. 3.1, Beam density estimation module in Sec. 3.2, Density soft clipping in Sec. 3.3, and Density-aware embedding module in Sec. 3.4. Additionally, Sec. 3.5 outlines our density augmentation technique. The overall framework of the proposed DDFE module is shown in Fig. 2. + +# 3.1 Point-voxel feature encoding + +We introduce a technique to extract generalized representations across domains from point clouds. The intensity distribution in point clouds is influenced by + +the specific LiDAR sensor used, which can adversely affect the performance of perception models during sensor transitions due to the variance in intensity information. Thus, we omit the use of LiDAR intensity values to enhance domain generalization performance. Furthermore, each LiDAR sensor has a unique distribution of measurement errors, particularly within the intra-voxel region. To minimize variance induced by localized sensing noise, we incorporate a direct voxel-wise feature encoding into the PointNet-based feature extraction [27]. This direct encoding, which bypasses localized information within the voxels, effectively dismisses grid size variations, such as those of $20\mathrm{cm}$ . Given a point cloud $\mathbf{P} = \{p_i\in \mathbb{R}^3\mid i = 1,\dots,N\}$ containing $N$ 3D points, we initially partition it into $M$ 3D voxels. For each voxel $v_{j}$ , it is defined as a set of points and is represented by $\mathbf{V} = \{v_{j}\in \mathbb{R}^{L\times 3}\mid j = 1,\dots,M\}$ , where $L$ indicates the number of points contained in each voxel. The coordinates of these points are subsequently transformed into the form of $(\cos (\theta_j),\sin (\theta_j),\phi_j,r_j)$ , where $(\theta_j,\phi_j,r_j)$ correspond to spherical coordinates. These transformed coordinates are then encoded into voxel-wise features $F^{v}\in \mathbb{R}^{M\times 16}$ through an MLP. We also encode point-wise features $F^{p}\in \mathbb{R}^{N\times 16}$ , which are essential to produce point-wise outputs. For each point $k$ in voxel $v_{j}$ centered at $(x_{j}^{c},y_{j}^{c},z_{j}^{c})$ , we compute the offset as $(x_{jk} - x_{j}^{c},y_{jk} - y_{j}^{c},z_{jk} - z_{j}^{c})$ , with $k = 1,\ldots ,L_{j}$ . Here, $L_{j}$ represents the number of points belonging to $v_{j}$ . These offsets are then processed by a point head to generate point-wise features. + +# 3.2 Beam density estimation module + +In this section, we detail the method to compute the density expectation for each ray emitted from a LiDAR sensor. Since all rays from a LiDAR originate from a singular source and are emitted at a fixed angle, the density associated with each beam can be deduced using a spherical projection informed by the beam configuration. The LiDAR sensor configuration, accessible from low-level sources such as an SDK, determines the set of horizontal and vertical inclinations $\mathbf{C}_h$ and $\mathbf{C}_v$ for each beam in this manner: + +$$ +\mathbf {C} _ {h} = \left\{\frac {2 \pi i}{H _ {\mathrm {b}}} \right\} _ {i \in \{1, \dots , H _ {\mathrm {b}} \}}, \mathbf {C} _ {v} = \left\{\frac {(f _ {\operatorname* {m a x}} - f _ {\operatorname* {m i n}}) j}{V _ {\mathrm {b}}} + f _ {\operatorname* {m i n}} \right\} _ {j \in \{1, \dots , V _ {\mathrm {b}} \}}, \qquad (1) +$$ + +where $f_{\mathrm{min}}$ and $f_{\mathrm{max}}$ denote the minimum and maximum LiDAR field of view, while $H_{\mathrm{b}}$ and $V_{\mathrm{b}}$ represent the number of horizontal and vertical beams, detailed in Tab. 1. To project LiDAR beam inclinations onto spherical projected image coordinates, we use the projection function $\pi : (\theta, \phi) \to (\dot{\theta}, \dot{\phi})$ defined as: + +$$ +\dot {\theta} = \left\lfloor \frac {\theta}{2 \pi} W \right\rfloor , \dot {\phi} = \left\lfloor \frac {\phi - f _ {\operatorname* {m i n}} ^ {\operatorname* {p r o j}}}{f _ {\operatorname* {m a x}} ^ {\operatorname* {p r o j}} - f _ {\operatorname* {m i n}} ^ {\operatorname* {p r o j}}} H \right\rfloor , \tag {2} +$$ + +where $H$ and $W$ are the height and width resolutions of the projected image, respectively set to $H = 512$ , $W = 5120$ . We define the projected image's field of view as $\left[f_{\mathrm{min}}^{\mathrm{proj}}, f_{\mathrm{max}}^{\mathrm{proj}}\right] = [-30.0, 15.0]$ to accommodate a range of LiDAR sensors. + +Given this, beam configurations can be transformed into 1-D binary vectors $\mathbf{B}_v\in \mathbb{R}^H$ and $\mathbf{B}_h\in \mathbb{R}^W$ as follows: + +$$ +\mathbf {B} _ {h} (\dot {\theta}) = \mathbb {1} _ {\mathbf {C} _ {h}} (\theta), \mathbf {B} _ {v} (\dot {\phi}) = \mathbb {1} _ {\mathbf {C} _ {v}} (\phi), \text {w h e r e} \mathbb {1} _ {\mathbf {C}} (x) = \left\{ \begin{array}{l l} 1, & \text {i f} x \in \mathbf {C} \\ 0, & \text {o t h e r w i s e} \end{array} , \right. \tag {3} +$$ + +where the indicator function $\mathbb{1}(\cdot)$ yields 1 when the azimuth or elevation $(\theta, \phi)$ of a pixel $(\dot{\theta}, \dot{\phi})$ corresponds to the beam inclinations specified in $\mathbf{C}$ in Eq. (1). Here, $\mathbf{C}$ can be either $\mathbf{C}_h$ or $\mathbf{C}_v$ that represent the locations of projected vertical and horizontal beams, respectively. To compute the beam density, we convolve the binary vectors $\mathbf{B}_h$ and $\mathbf{B}_v$ , with four distinct 1-D Gaussian kernels, each characterized by standard deviations $\sigma_k = \{10, 30, 50, 70\}$ as follows: + +$$ +\tilde {\mathbf {B}} _ {h} = \mathbf {B} _ {h} * \mathbf {G} _ {\sigma_ {k}}, \quad \tilde {\mathbf {B}} _ {v} = \mathbf {B} _ {v} * \mathbf {G} _ {\sigma_ {k}}. \tag {4} +$$ + +Finally, we define beam density $\mathcal{D}_i$ of point $p_i$ as follows: + +$$ +\mathcal {D} _ {i} = \left[ \sqrt {\hat {\mathbf {B}} _ {h} ^ {(k)} (\dot {\theta} _ {i}) \cdot \hat {\mathbf {B}} _ {v} ^ {(k)} (\dot {\phi} _ {i}) / r _ {i} ^ {2}} \right] _ {k = 1} ^ {4}, \tag {5} +$$ + +where $r_i$ is the radial distance of a LiDAR point $p_i$ in the spherical coordinates, and $[\cdot]_{k=1}^4$ denotes the concatenation operation. Taking into account that the density of the beam diminishes in proportion to the square of the distance due to its radial emission, we incorporate the inverse of $r^2$ in our computation. + +# 3.3 Density soft clipping + +When exposed to densities that deviate from those in the source dataset, the model is susceptible to performance degradation in an unseen domain. To alleviate this issue, we introduce a density soft clipping method that confines the density spectrum using the $\tanh(\cdot)$ function as follows: + +$$ +\begin{array}{l} \mathcal {D} _ {i} ^ {c} = \tanh \left(\frac {\mathcal {D} _ {i} - m}{l}\right) l + m, \\ m = \frac {\mathcal {P} _ {9 0} (\mathcal {D}) + \mathcal {P} _ {1 0} (\mathcal {D})}{2}, l = \frac {\mathcal {P} _ {9 0} (\mathcal {D}) - \mathcal {P} _ {1 0} (\mathcal {D})}{2}, \\ \end{array} +$$ + +where $\mathcal{D}_i^c$ is clipped density of point $p_i$ , and $\mathcal{P}_{90}$ and $\mathcal{P}_{10}$ are functions that extract $90^{\mathrm{th}}$ and $10^{\mathrm{th}}$ channel-wise percentile values from the density embedding of the training domain, respectively. Employing these percentile values allows us to discount the outlier density values. To calculate the percentile on the fly under memory constraints, we adopt the Reservoir Sampling [37] technique with a sample size $N = 1000$ . Please refer to the supplementary material for a more detailed algorithm description. + +# 3.4 Density-aware embedding module + +For each point $p_i$ , a point-wise density embedding feature $\mathcal{D}_i^c$ is obtained via the beam density estimation module incorporating density soft clipping. This feature serves as an input of both a point-wise attention function $f_p$ and a voxelwise attention function $f_v$ . These mechanisms are pivotal for crafting domain-invariant density-discriminative features. In the point-wise attention framework, $\mathcal{D}_i^c$ is synchronized dimensionally with the point-wise feature $F_i^p$ , facilitating the creation of a density discriminative feature $\hat{F}_i^p$ as follows: + +$$ +\hat {F} _ {i} ^ {p} = f _ {p} \left(\mathcal {D} _ {i} ^ {c}\right) \odot F _ {i} ^ {p}, \tag {7} +$$ + +where $f_{p}$ consists of two 1D convolution layers followed by a sigmoid function. For voxel-wise attention, it begins by averaging the density embedding features for all points encapsulated within a voxel. This aggregated measure then undergoes a similar treatment as follows: + +$$ +\hat {F} _ {j} ^ {v} = \operatorname {C o n c a t} \left(f _ {v} \left(\mathcal {D} _ {j} ^ {c}\right) \odot F _ {j} ^ {v}, g \left(F _ {j} ^ {p}\right)\right), \quad \mathcal {D} _ {j} ^ {c} = \frac {1}{\left| L _ {j} \right|} \sum_ {p _ {i} \in v _ {j}} \mathcal {D} _ {i} ^ {c}, \tag {8} +$$ + +where $\mathcal{D}_j^c$ is a clipped density of voxel $v_{j}$ and $\hat{F}_j^v$ is a voxel-wise density discriminative feature and $f_{v}$ consists of two 1D convolution layers followed by a sigmoid function. An aggregation function $g$ for the voxel-wise feature $F_{j}^{v}$ is adjusted by voxel-wise attention and concatenated with the max pooled point-wise feature within the $v_{j}$ . Finally, the extracted features are passed through a single 1D convolutional layer, resulting in 32-channel features. These density-aware features are subsequently fed into a 3D backbone for LiDAR semantic segmentation. + +# 3.5 Density Augmentation + +The DDFE module aims to align densities across varied sensor domains by considering each dataset as a collection of multiple density domains. This method, while effective in many scenarios, may face challenges when there's little to no overlap in the density ranges between the source and unseen datasets. To address this potential issue, we introduce a density augmentation method aimed at broadening the density spectrum covered by the training data. Several existing 3D point cloud augmentation strategies, such as [16, 22, 38] outlined in recent studies, attempt to address density variations either directly or by implication. We adapt the Mix3D [22] for our purposes, which we refer to as enhanced-Mix3D, incorporating random translations along the direction of ego-vehicle movement, along with additional rotational transformations to simulate a wider range of density variations. We also employ beam sampling technique [16, 38], selectively eliminating specific LiDAR beams, to amplify the density in the lower direction. During the training phase, we apply both enhanced-Mix3D and beam sampling augmentations with a set probability of 0.5, aiming to effectively simulate diverse density conditions. Please refer to the supplementary material for more detailed implementation. + +# 4 Experiments + +In this section, we demonstrate the domain generalization performance of our method through extensive experiments. The implementation details and experimental settings of the proposed method are detailed in Sec. 4.1. The configuration of datasets used for evaluation is detailed in Sec. 4.2. A comparative analysis with recent domain generalization and domain adaptation methods is presented in Sec. 4.3. The effectiveness of individual components within the proposed DDFE and computational cost are analyzed in Section Sec. 4.4. Detailed per-class performance metrics are available in the supplementary material. + +# 4.1 Implementation Details + +We employ the point head inspired by Cylinder3D [48] to produce point-wise outputs as shown in Fig. 2. During the training phase, we incorporate the Lovasz-Softmax loss $\mathcal{L}^{lovasz}$ [3] along with the Weighted Cross-Entropy loss $\mathcal{L}^{wce}$ . The total loss is composed of an equal-weighted combination of point-wise and voxelwise losses as follows: + +$$ +\mathcal {L} _ {\text {t o t a l}} = \mathcal {L} _ {\text {p o i n t}} ^ {\text {l o v a s z}} + \mathcal {L} _ {\text {p o i n t}} ^ {\text {w c e}} + \mathcal {L} _ {\text {v o x e l}} ^ {\text {l o v a s z}} + \mathcal {L} _ {\text {v o x e l}} ^ {\text {w c e}}. \tag {9} +$$ + +We use the Adam optimizer with an initial learning rate of 1e-3. This rate is decreased by a factor of 0.99 with every epoch. The training is conducted over 30 epochs with a batch size of two on a single NVIDIA RTX 3090, and the epoch yielding the highest source validation mIoU is chosen. For voxelization, we adopt a cubic size of [20cm, 20cm, 20cm]. In cases where multiple point labels are found within a voxel, the voxel label is determined based on the predominant label, aligning with the approach in Cylinder3D [48]. + +# 4.2 Datasets + +In our experimental setup, we employ three datasets: Waymo [35], nuScenes [4], and SemanticKITTI [2]. Notably, the Waymo and SemanticKITTI datasets use 64-channel LiDAR, while the nuScenes employs a 32-channel LiDAR. For detailed information, please refer to Tab. 1. In line with methods from previous studies [16,29], we split the nuScenes sequences into 700 for training and 150 for validation. The Waymo dataset is partitioned into 798 sequences for training and 202 for validation. Regarding the SemanticKITTI dataset, we use sequences 00 to 10 for training, setting aside sequence 08 exclusively for validation. In integrating class variations across datasets, we follow the class mapping configurations from previous studies for fair comparisons. + +# 4.3 Comparison to State-of-the-Art DA/DG Methods + +The field of domain generalization of LiDAR-based semantic segmentation is still in its formative stages, with a lack of universally accepted experimental standards. This situation has led to a diversity of experimental designs across different studies. To facilitate fair comparisons, we align our experimental framework + +Table 2: Comparison with domain generalization methods based on MinkowskiNet architecture using the mIoU. The best and the second best results are highlighted in bold and underline, respectively. + +
MethodDASourceWKNSourceKWNSourceNWK
Base75.3749.4047.8357.3135.2437.4265.7838.6536.24
IBN-Net [23]75.4751.1344.7257.7436.9938.7465.3136.5336.93
MLDG [20]72.4748.9448.6456.2635.3936.7761.3236.3332.70
COSMIX (W) [30]✓---K49.3539.4638.94N---
COSMIX (K) [30]✓66.6844.7149.96------
COSMIX (N) [30]✓65.6840.9947.9849.9838.0543.25---
DGLSS [16]75.2851.2349.6159.6240.6744.8365.3240.9338.98
Ours76.1557.0756.7562.5042.7349.4368.1645.9846.52
+ +Table 3: Comparison with domain adaptation and data augmentation methods. + +
BackboneMethodsK→NN→KBackboneMethodsK→NN→K
MinkNet42 [7]Base37.836.1C&L [44]Base27.923.5
CutMix [46]37.137.6SWD [18]27.724.5
Copy-Paste [10]38.541.13DGCA [39]27.423.9
Mix3D [22]43.144.7C&L [44]31.633.7
PolarMix [40]45.839.1LiDomAug [29]39.237.9
LiDomAug [29]45.948.3LiDAR-UDA [34]41.834.0
Ours (v=5cm)48.651.3Ours (v=5cm)42.541.0
Ours (v=20cm)50.146.3Ours (v=20cm)47.140.3
+ +with those utilized in seminal works such as LiDomAug [29] and DGLSS [16]. This approach allows us to benchmark our method against a range of Domain Adaptation (DA) and Domain Generalization (DG) techniques, showcasing the effectiveness of our method in the context of evolving domain generalization challenges. + +Experiments in the DGLSS Setting We compare our method with a domain adaptation method [30] and three domain generalization methods [16,20,23], following the experimental framework defined in DGLSS [16] using Waymo (W), SemanticKITTI (K), and nuScenes (N). We adopt MinkowskiNet [7] as our backbone network, aligning with the configuration utilized in DGLSS. The results in Tab. 2 indicate that the proposed method consistently outperforms DGLSS across all datasets. Impressively, our method not only demonstrates superior domain generalization performance but also dominates in the source-to-source settings $(\mathrm{W} \rightarrow \mathrm{W}, \mathrm{K} \rightarrow \mathrm{K}, \mathrm{N} \rightarrow \mathrm{N})$ . Unlike DGLSS, which tailors its augmentation strategies for each specific dataset, our method employs a uniform hyperparameter setting across all datasets, achieving enhanced domain generalization results. Our method achieves an average increase of $+12.9\%$ over DGLSS for unseen datasets using Waymo as the source data $(\mathrm{W} \rightarrow \mathrm{K}, \mathrm{W} \rightarrow \mathrm{N})$ . With SemanticKITTI as the source, there's an average enhancement of $+7.8\%$ for other unseen domains $(\mathrm{K} \rightarrow \mathrm{W}, \mathrm{K} \rightarrow \mathrm{N})$ , and employing nuScenes as the source data yields an average increase of $+15.8\%$ on other unseen datasets $(\mathrm{N} \rightarrow \mathrm{W}, \mathrm{N} \rightarrow \mathrm{K})$ . + +These results demonstrate the robustness of the proposed method in addressing domain discrepancies. + +![](images/6787ce1d22ac50a87da0b61b6d136a34909358897286513db58cda2142795387.jpg) + +![](images/2e43adec9e88f675f3b4688dda717c5d4eff99b4d7bc1907473662c54c3536df.jpg) + +![](images/c1036684973bcba7064adb4c5b623e79ea12b333591a480939af40ccb6781e8c.jpg) + +![](images/e98508730dd76647ca5502e4553fcd0d790de32b16677c8fe73d905b8b36f5ed.jpg) +(a) Ground-truth +Fig. 3: Qualitative comparison with MinkNet42 backbone. Top: Model trained on nuScenes, tested on SemanticKITTI $(\mathrm{N} \rightarrow \mathrm{K})$ . Bottom: Trained on SemanticKITTI, tested on nuScenes $(\mathrm{K} \rightarrow \mathrm{N})$ . + +![](images/c81395f531d12d51e25eec38a164efe7e3008364c6c5a09997dc916df4c834d3.jpg) +(b) MinkNet42 + +![](images/225e7d0ce32039fdf0db98059c68904d7c80ebbd05d8ae1e2e99d3696907e1ee.jpg) +(c) MinkNet42+Ours + +Experiments in the LiDomAug Setting We benchmark the proposed method against various augmentation methods [10,22,40,46], domain adaptation methods [18,34,39,44], and a domain generation method [29]. We adhere to the experimental setup established by LiDomAug [29], employing MinkNet42 [7] and C&L [44] as backbone networks, and utilizing SemanticKITTI and nuScenes datasets for evaluation. Unlike the DGLSS [16] setting which uses a voxel size of $20~\mathrm{cm}$ , the LiDomAug setting uses a voxel size of $5\mathrm{cm}$ . We benchmark the proposed method against various augmentation methods [10,22,40,46], domain adaptation methods [18,34,39,44], and a domain generation method [29]. We adhere to the experimental setup established by LiDomAug [29], employing MinkNet42 [7] and C&L [44] as backbone networks, and utilizing SemanticKITTI and nuScenes datasets for evaluation. The comparative analysis in Tab. 3 shows that our method surpasses all considered augmentation methods in both $(\mathrm{K}\rightarrow \mathrm{N})$ and $(\mathrm{N}\rightarrow \mathrm{K})$ configurations. Particularly with MinkNet42 as the backbone, our method achieves a $5.9\%$ and $5.8\%$ increase in performance over LiDomAug in the $(\mathrm{K}\rightarrow \mathrm{N})$ and $(\mathrm{N}\rightarrow \mathrm{K})$ scenarios, respectively. This enhancement is significant, considering LiDomAug's reliance on ego-motion and multi-frame data integration for domain generalization. When utilizing C&L [44] as the backbone, the proposed method significantly improves mIoU by $+34.5\%$ and $+8.4\%$ for the scenario of training on SemanticKITTI and evaluating on nuScenes $(\mathrm{K}\rightarrow \mathrm{N})$ , compared to C&L and LiDomAug, respectively. Moreover, in the transition from the sparse 32-channel dataset (nuScenes) to the denser 64-channel dataset (SemanticKITTI) $(\mathrm{N}\rightarrow \mathrm{K})$ , our method shows a $+21.7\%$ increase in mIoU over C&L and $+8.2\%$ improvement over LiDomAug. + +We further conduct an analysis with a voxel size of $20~\mathrm{cm}$ , unlike the comparative methods that all use a voxel size of $5\mathrm{cm}$ . Increasing the voxel size from $5\mathrm{cm}$ to $20\mathrm{cm}$ results in reductions of $30.3\%$ in training time and $62.5\%$ in infer- + +Table 4: Ablation study on individual components within our method using mIoU. Experiments evaluate the model's performance with and without the following: (a) Point-voxel encoding, (b) Density-aware embedding module, (c) Density clipping, and (d) Density-Augmentation. All experiments were conducted using a voxel size of 20 cm. + +
(a)(b)(c)(d)K→NN→K
40.731.4
✓43.0 (+5.7%)35.0 (+11.5%)
✓✓45.7 (+12.3%)40.5 (+29.0%)
✓✓✓46.2 (+13.5%)41.8 (+33.1%)
✓✓✓✓50.1 (+23.1%)46.3 (+47.5%)
+ +ence time, while maintaining comparable performance—with a slight decrease of $3.5\%$ on the MinkNet42 backbone and an increase of $4.7\%$ on the C&L backbone. Based on these results, we recommend a voxel size of $20~\mathrm{cm}$ ( $\mathrm{v} = 20\mathrm{cm}$ ) as the default setting for our proposed method. The effectiveness is further illustrated in the qualitative results depicted in Fig. 3, highlighting the substantial improvements our method brings to the domain generalization performance of the 3D backbone network. Please refer to the supplementary material for detailed results, including per-class IoU. + +# 4.4 Ablation Studies + +To validate the impact of individual components within our method, we conduct ablation studies. These studies concentrate on the evaluation of point-voxel encoding, the density-aware embedding module, density soft clipping, and density augmentation, providing insights into the significance of each element in enhancing the overall methodology. + +Analysis of the individual components of our method In our comprehensive ablation study in Tab. 4, we thoroughly investigate the individual components of our method. We use the experimental setup proposed by LiDomAug [29] and employ MinkowskiNet [7] as the backbone. This study dissects the impact of our DDFE, which includes (a) point-voxel encoding, (b) density-aware embedding module, and (c) density clipping. Additionally, it examines (d) the impact of our density augmentation method. + +The integration of all these elements results in a significant performance uplift $+13.5\%$ in the $(\mathrm{K}\rightarrow \mathrm{N})$ scenario and $+33.1\%$ in the $(\mathrm{N}\rightarrow \mathrm{K})$ scenario over the baseline model. Notably, the density-aware embedding module stands out for its substantial effect, closely followed by the point-voxel encoding, highlighting its critical role. Furthermore, the study validates the utility of density soft clipping in enhancing model flexibility towards novel density configurations. Meanwhile, the gains observed with density clipping highlight the inherent challenge in adapting to unfamiliar density landscapes, emphasizing the nuanced contributions of each component towards achieving robust domain generalization. Lastly, + +Table 5: Comparison of data augmentation methods in the proposed DDFE. The best and the second best results are highlighted in bold and underline, respectively. 'D. V.' refers to Density Variation situation, and 'B. S.' to Beam Sampling augmentation. + +
MethodD. V.K→NN→K
DDFE45.941.8
+ PolarMix [40]48.2 (+5.0%)42.3 (+1.2%)
+ B. S. [38]✓49.1 (+7.0%)42.2 (+0.9%)
+ Mix3D [22]✓47.4 (+3.2%)44.8 (+7.1%)
+ Mix3D + B. S. [38]✓49.1 (+7.0%)44.2 (+5.7%)
+ E-Mix3D✓48.5 (+5.6%)46.4 (+11.0%)
+ E-Mix3D + B. S. [38]✓49.4 (+7.6%)46.5 (+11.2%)
+ +![](images/aaa37abd30689c928f31cf4afde08f9c61fc9cb7a99125d9da3c9f7ca4249fc1.jpg) +Fig. 4: Visualization of feature similarity matrices between the Waymo (64-channel) and the nuScenes (32-channel) datasets, with a focus on the distance of objects from the LiDAR. We utilize models trained with the nuScenes dataset for visualization. (a) A point-voxel feature encoding method. (b) A density-aware embedding module. (c) A density soft clipping. Our DDFE combines all modules (a), (b), and (c). + +![](images/8b0eedd32fe01fce7a217da4bb30d1e6e6ae5558322b10573c27e363a2c914b3.jpg) + +incorporating density augmentation into DDFE leads to further enhancements, with an additional $8.4\%$ performance increase in the $(\mathrm{K} \rightarrow \mathrm{N})$ scenario and $10.8\%$ in the $(\mathrm{N} \rightarrow \mathrm{K})$ scenario. This demonstrates the effectiveness of the proposed density augmentation scheme in significantly boosting the performance of DDFE. + +Augmentation Tab. 5 shows the performance comparison when various augmentation methods are applied to the proposed DDFE. The experimental setup follows DGLSS [16], and the baseline involves applying the proposed DDFE to MinkowskiNet [7]. The proposed enhanced-Mix3D shows a $2.3\%$ performance improvement over the original Mix3D [22] when trained on SemanticKITTI and tested on nuScenes $(\mathrm{K}\rightarrow \mathrm{N})$ , and a $3.6\%$ improvement when trained on nuScenes and tested on SemanticKITTI $(\mathrm{N}\rightarrow \mathrm{K})$ . In conjunction with the beam sampling method, it achieves a $7.6\%$ performance improvement in $(\mathrm{K}\rightarrow \mathrm{N})$ and $11.2\%$ in $(\mathrm{N}\rightarrow \mathrm{K})$ compared to the baseline. + +Analysis of Density Discriminative Feature Embedding (DDFE) We delve into the effectiveness of the DDFE module by examining the feature similarity between the source dataset, nuScenes, and the unseen dataset, Waymo, + +as depicted in Fig. 4. Our focus is on the input voxel features $\hat{F}^v$ of the 3D backbone model generated by the DDFE. We initiate our analysis by setting a baseline that utilizes a point-voxel feature encoding method for comparison. Following this, we proceed to assess our enhanced method, which integrates a density-aware embedding module along with density soft clipping, to understand its impact on bridging the domain gap. Similarity metrics are derived from the L2 distance between averaged features across specified distance intervals (e.g., 0-5m, ..., 45-50m) for each dataset. According to Fig. 4, the baseline consistently shows higher feature similarity across distances, influenced by its direct embedding of 3D point coordinates. Conversely, our integrated model with density-aware embedding and soft clipping significantly aligns feature similarity across consistent density distributions. It demonstrates the efficacy of the proposed modules in adapting to differences between the source and unseen datasets. + +Computational Cost The inference time of our method for processing a single frame with the nuScenes is 44ms on an NVIDIA RTX 3090. The inclusion of our DDFE module into the MinkNet42 architecture accounts for an extra 8ms of this computation time, while the base MinkNet42 architecture alone requires 36ms. This integration modestly elevates the model's complexity, introducing approximately 23.8k parameters $(+0.06\%)$ . Furthermore, the application of density augmentation during training adds an extra 60ms per scan on the nuScenes. + +# 5 Conclusion + +In this paper, we introduce a new perspective for domain generalization of LiDAR semantic segmentation by exploiting the concept of density diversity within a source domain. Based on this perspective, we propose the Density Discriminative Feature Embedding (DDFE) module. DDFE is designed to incorporate expected densities derived from LiDAR beams into a density-aware feature space, significantly enhancing the model's capability to distinguish between different densities. This novel approach improves the adaptability and accuracy of LiDAR semantic segmentation, enabling them to perform effectively across diverse domains, including those previously unseen. In addition to the core methodology, we also present a simple and effective data augmentation technique that extends the density spectrum of the source data. Extensive experiments on the SemanticKITTI, Waymo, and nuScenes datasets demonstrate the effectiveness of the proposed method in elevating the domain generalization performance of LiDAR semantic segmentation. A limitation of this study is its exclusive focus on semantic segmentation tasks. Thus, in future work, we plan to extend our domain generalization method to encompass other LiDAR-based perception tasks, such as 3D object detection. This extension can involve utilizing object-centric density discriminative features alongside point-wise density embedding. + +# Acknowledgements + +This work was supported by the Digital Innovation Hub project supervised by the Daegu Digital Innovation Promotion Agency(DIP) grant funded by the Korea government(MSIT and Daegu Metropolitan City) in 2023(DBSD1-02, Vision picking system for the logistics industry based on artificial intelligence object recognition) and the National Research Foundation of Korea (NRF) grant funded by the Korea government (MSIT) (No. RS-2023-00210908). + +# References + +1. Ando, A., Gidaris, S., Bursuc, A., Puy, G., Boulch, A., Marlet, R.: Rangevit: Towards vision transformers for 3d semantic segmentation in autonomous driving. In: Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 5240-5250 (2023) +2. Behley, J., Garbade, M., Milioto, A., Quenzel, J., Behnke, S., Stachniss, C., Gall, J.: Semantickitti: A dataset for semantic scene understanding of lidar sequences. In: Proceedings of IEEE International Conference on Computer Vision (ICCV). pp. 9297-9307 (2019) +3. Berman, M., Triki, A.R., Blaschko, M.B.: The lovasz-softmax loss: A tractable surrogate for the optimization of the intersection-over-union measure in neural networks. In: Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 4413-4421 (2018) +4. Caesar, H., et al.: nuscenes: A multimodal dataset for autonomous driving. In: Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 11621-11631 (2020) +5. Chen, T.H., Chang, T.S.: Rangeseg: Range-aware real time segmentation of 3d lidar point clouds. IEEE Transactions on Intelligent Vehicles 7(1), 93-101 (2022). https://doi.org/10.1109/TIV.2021.3085827 +6. Cheng, H.X., Han, X.F., Xiao, G.Q.: Transrvnet: Lidar semantic segmentation with transformer. IEEE Transactions on Intelligent Transportation Systems 24(6), 5895-5907 (2023). https://doi.org/10.1109/TITS.2023.3248117 +7. Choy, C., Gwak, J., Savarese, S.: 4d spatio-temporal convnets: Minkowski convolutional neural networks. In: Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 3075-3084 (2019) +8. Cortinhal, T., Tzelepis, G., Erdal Aksoy, E.: Salsanext: Fast, uncertainty-aware semantic segmentation of lidar point clouds. In: Bebis, G., Yin, Z., Kim, E., Bender, J., Subr, K., Kwon, B.C., Zhao, J., Kalkofen, D., Baciu, G. (eds.) Advances in Visual Computing. pp. 207-222. Springer International Publishing, Cham (2020). https://doi.org/10.1007/978-3-030-64559-5_16 +9. Fan, S., Dong, Q., Zhu, F., Lv, Y., Ye, P., Wang, F.Y.: Scf-net: Learning spatial contextual features for large-scale point cloud segmentation. In: Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 14504-14513 (2021) +10. Ghiasi, G., Cui, Y., Srinivas, A.J., Qian, R., Lin, T.Y., Cubuk, E.D., Le, Q.V., Zoph, B.: Simple copy-paste is a strong data augmentation method for instance segmentation. In: Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 2918-2928 (2021) + +11. Graham, B., Engelcke, M., Van Der Maaten, L.: 3d semantic segmentation with submanifold sparse convolutional networks. In: Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 9224-9232 (2018) +12. Graham, B., Van der Maaten, L.: Submanifold sparse convolutional networks. arXiv preprint arXiv:1706.01307 (2017) +13. Hou, Y., Zhu, X., Ma, Y., Loy, C.C., Li, Y.: Point-to-voxel knowledge distillation for lidar semantic segmentation. In: Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 8479-8488 (2022) +14. Hu, Q., Liu, D., Hu, W.: Density-insensitive unsupervised domain adaption on 3d object detection. In: Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 17556-17566 (2023) +15. Hu, Q., Yang, B., Xie, L., Rosa, S., Guo, Y., Wang, Z., Trigoni, N., Markham, A.: Randla-net: Efficient semantic segmentation of large-scale point clouds. In: Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 11108-11117 (2020) +16. Kim, H., Kang, Y., Oh, C., Yoon, K.J.: Single domain generalization for lidar semantic segmentation. In: Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 17587-17598 (2023) +17. Kong, L., Quader, N., Liong, V.E.: Conda: Unsupervised domain adaptation for lidar segmentation via regularized domain concatenation. In: Proceedings of IEEE International Conference on Robotics and Automation (ICRA). pp. 9338-9345. IEEE (2023) +18. Lee, C.Y., Batra, T., Baig, M.H., Ulbricht, D.: Sliced wasserstein discrepancy for unsupervised domain adaptation. In: Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 10285-10295 (2019) +19. Lehner, A., Gasperini, S., Marcos-Ramiro, A., Schmidt, M., Mahani, M.A.N., Navab, N., Busam, B., Tombari, F.: 3d-vfield: Adversarial augmentation of point clouds for domain generalization in 3d object detection. In: Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 17295-17304 (2022) +20. Li, D., Yang, Y., Song, Y.Z., Hospedales, T.: Learning to generalize: Meta-learning for domain generalization. In: Proceedings of the AAAI Conference on Artificial Intelligence (AAAI). vol. 32 (2018) +21. Li, M., Zhang, Y., Ma, X., Qu, Y., Fu, Y.: Bev-dg: Cross-modal learning under bird's-eye view for domain generalization of 3d semantic segmentation. In: Proceedings of IEEE International Conference on Computer Vision (ICCV). pp. 11632-11642 (2023) +22. Nekrasov, A., Schult, J., Litany, O., Leibe, B., Engelmann, F.: Mix3d: Out-of-context data augmentation for 3d scenes. In: Proceedings of International Conference on 3D Vision (3DV). pp. 116-125 (2021) +23. Pan, X., Luo, P., Shi, J., Tang, X.: Two at once: Enhancing learning and generalization capacities via ibn-net. In: Proceedings of European Conference on Computer Vision (ECCV). pp. 464-479 (2018) +24. Pan, Y., Gao, B., Mei, J., Geng, S., Li, C., Zhao, H.: Semanticposs: A point cloud dataset with large quantity of dynamic instances. In: 2020 IEEE Intelligent Vehicles Symposium (IV). pp. 687-693 (2020). https://doi.org/10.1109/IV47402.2020.9304596 +25. Peng, K., Fei, J., Yang, K., Roitberg, A., Zhang, J., Bieder, F., Heidenreich, P., Stiller, C., Stiefelhagen, R.: Mass: Multi-attentional semantic segmentation of lidar + +data for dense top-view understanding. IEEE Transactions on Intelligent Transportation Systems 23(9), 15824-15840 (2022). https://doi.org/10.1109/TITS.2022.3145588 +26. Peng, X., Zhu, X., Ma, Y.: Cl3d: Unsupervised domain adaptation for cross-lidar 3d detection. In: Proceedings of the AAAI Conference on Artificial Intelligence (AAAI). vol. 37, pp. 2047-2055 (2023) +27. Qi, C.R., Su, H., Mo, K., Guibas, L.J.: Pointnet: Deep learning on point sets for 3d classification and segmentation. In: Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 652-660 (2017) +28. Rochan, M., Aich, S., Corral-Soto, E.R., Nabatchian, A., Liu, B.: Unsupervised domain adaptation in lidar semantic segmentation with self-supervision and gated adapters. In: Proceedings of IEEE International Conference on Robotics and Automation (ICRA). pp. 2649-2655 (2022) +29. Ryu, K., Hwang, S., Park, J.: Instant domain augmentation for lidar semantic segmentation. In: Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 9350-9360 (2023) +30. Saltori, C., Galasso, F., Fiameni, G., Sebe, N., Ricci, E., Poiesi, F.: Cosmix: Compositional semantic mix for domain adaptation in 3d lidar segmentation. In: Proceedings of European Conference on Computer Vision (ECCV). pp. 586-602. Springer (2022) +31. Saltori, C., Lathuilière, S., Sebe, N., Ricci, E., Galasso, F.: Sf-uda 3d: Source-free unsupervised domain adaptation for lidar-based 3d object detection. In: Proceedings of International Conference on 3D Vision (3DV). pp. 771–780 (2020) +32. Saltori, C., Osep, A., Ricci, E., Leal-Taixe, L.: Walking your lidog: A journey through multiple domains for lidar semantic segmentation. In: Proceedings of IEEE International Conference on Computer Vision (ICCV). pp. 196-206 (2023) +33. Sanchez, J., Deschaud, J.E., Goulette, F.: Domain generalization of 3d semantic segmentation in autonomous driving. In: Proceedings of IEEE International Conference on Computer Vision (ICCV). pp. 18077-18087 (2023) +34. Shaban, A., Lee, J., Jung, S., Meng, X., Boots, B.: Lidar-uda: Self-ensembling through time for unsupervised lidar domain adaptation. In: Proceedings of IEEE International Conference on Computer Vision (ICCV). pp. 19784-19794 (2023) +35. Sun, P., et al.: Scalability in perception for autonomous driving: Waymo open dataset. In: Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 2446-2454 (2020) +36. Thomas, H., Qi, C.R., Deschaud, J.E., Marcotegui, B., Goulette, F., Guibas, L.J.: Kpconv: Flexible and deformable convolution for point clouds. In: Proceedings of IEEE International Conference on Computer Vision (ICCV). pp. 6411-6420 (2019) +37. Vitter, J.S.: Random sampling with a reservoir. ACM Trans. Math. Softw. 11(1), 37-57 (mar 1985). https://doi.org/10.1145/3147.3165 +38. Wei, Y., Wei, Z., Rao, Y., Li, J., Zhou, J., Lu, J.: Lidar distillation: Bridging the beam-induced domain gap for 3d object detection. In: Proceedings of European Conference on Computer Vision (ECCV). pp. 179-195. Springer (2022) +39. Wu, B., Zhou, X., Zhao, S., Yue, X., Keutzer, K.: Squeezesegv2: Improved model structure and unsupervised domain adaptation for road-object segmentation from a lidar point cloud. In: Proceedings of IEEE International Conference on Robotics and Automation (ICRA). pp. 4376-4382 (2019) +40. Xiao, A., Huang, J., Guan, D., Cui, K., Lu, S., Shao, L.: Polarmix: A general data augmentation technique for lidar point clouds. In: Advances in Neural Information Processing Systems (NeurIPS). pp. 11035-11048 (2022) + +41. Xiao, A., Huang, J., Xuan, W., Ren, R., Liu, K., Guan, D., El Saddik, A., Lu, S., Xing, E.P.: 3d semantic segmentation in the wild: Learning generalized models for adverse-condition point clouds. In: Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 9382-9392 (2023) +42. Xiao, P., Shao, Z., Hao, S., Zhang, Z., Chai, X., Jiao, J., Li, Z., Wu, J., Sun, K., Jiang, K., Wang, Y., Yang, D.: Pandaset: Advanced sensor suite dataset for autonomous driving. In: ITSC. pp. 3095-3101 (2021). https://doi.org/10.1109/ITSC48978.2021.9565009 +43. Yang, J., Shi, S., Wang, Z., Li, H., Qi, X.: St3d: Self-training for unsupervised domain adaptation on 3d object detection. In: Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 10368-10378 (2021) +44. Yi, L., Gong, B., Funkhouser, T.: Complete & label: A domain adaptation approach to semantic segmentation of lidar point clouds. In: Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 15363-15373 (2021) +45. Yuan, Z., Wen, C., Cheng, M., Su, Y., Liu, W., Yu, S., Wang, C.: Category-level adversaries for outdoor lidar point clouds cross-domain semantic segmentation. IEEE Transactions on Intelligent Transportation Systems 24(2), 1982-1993 (2023). https://doi.org/10.1109/TITS.2022.3219853 +46. Yun, S., Han, D., Oh, S.J., Chun, S., Choe, J., Yoo, Y.: Cutmix: Regularization strategy to train strong classifiers with localizable features. In: Proceedings of IEEE International Conference on Computer Vision (ICCV). pp. 6023-6032 (2019) +47. Zhang, Y., Zhou, Z., David, P., Yue, X., Xi, Z., Gong, B., Foroosh, H.: Polarnet: An improved grid representation for online lidar point clouds semantic segmentation. In: Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 9601-9610 (2020) +48. Zhu, X., Zhou, H., Wang, T., Hong, F., Ma, Y., Li, W., Li, H., Lin, D.: Cylindrical and asymmetrical 3d convolution networks for lidar segmentation. In: Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 9939-9948 (2021) \ No newline at end of file diff --git a/2024/Rethinking LiDAR Domain Generalization_ Single Source as Multiple Density Domains/images.zip b/2024/Rethinking LiDAR Domain Generalization_ Single Source as Multiple Density Domains/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..4bb9f0b190a5baa0663eec08b5a233c61e23027a --- /dev/null +++ b/2024/Rethinking LiDAR Domain Generalization_ Single Source as Multiple Density Domains/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:90253ee79d2be7652c9737f52bdc12150d834c7e08c86ac0c0f92415e44cbbe8 +size 538888 diff --git a/2024/Rethinking LiDAR Domain Generalization_ Single Source as Multiple Density Domains/layout.json b/2024/Rethinking LiDAR Domain Generalization_ Single Source as Multiple Density Domains/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..7d45c72119959beb26f4b9e70120659a90d1d2a9 --- /dev/null +++ b/2024/Rethinking LiDAR Domain Generalization_ Single Source as Multiple Density Domains/layout.json @@ -0,0 +1,9549 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 156, + 111, + 459, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 156, + 111, + 459, + 148 + ], + "spans": [ + { + "bbox": [ + 156, + 111, + 459, + 148 + ], + "type": "text", + "content": "Rethinking LiDAR Domain Generalization: Single Source as Multiple Density Domains" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 168, + 469, + 181 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 168, + 469, + 181 + ], + "spans": [ + { + "bbox": [ + 145, + 168, + 469, + 181 + ], + "type": "text", + "content": "Jaeyeul Kim, Jungwan Woo, Jeonghoon Kim, and Sunghoon Im" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 171, + 190, + 443, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 171, + 190, + 443, + 213 + ], + "spans": [ + { + "bbox": [ + 171, + 190, + 443, + 213 + ], + "type": "text", + "content": "DGIST, Daegu, South Korea {jykim94, friendship1, jeonghoon, sunghoonim}@dgist.ac.kr" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 160, + 243, + 455, + 452 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 243, + 455, + 452 + ], + "spans": [ + { + "bbox": [ + 160, + 243, + 455, + 452 + ], + "type": "text", + "content": "Abstract. In the realm of LiDAR-based perception, significant strides have been made, yet domain generalization remains a substantial challenge. The performance often deteriorates when models are applied to unfamiliar datasets with different LiDAR sensors or deployed in new environments, primarily due to variations in point cloud density distributions. To tackle this challenge, we propose a Density Discriminative Feature Embedding (DDFE) module, capitalizing on the observation that a single source LiDAR point cloud encompasses a spectrum of densities. The DDFE module is meticulously designed to extract density-specific features within a single source domain, facilitating the recognition of objects sharing similar density characteristics across different LiDAR sensors. In addition, we introduce a simple yet effective density augmentation technique aimed at expanding the spectrum of density in source data, thereby enhancing the capabilities of the DDFE. Our DDFE stands out as a versatile and lightweight domain generalization module. It can be seamlessly integrated into various 3D backbone networks, where it has demonstrated superior performance over current state-of-the-art domain generalization methods. Code is available at https://github.com/dgist-cvlab/MultiDensityDG." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 160, + 462, + 445, + 473 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 462, + 445, + 473 + ], + "spans": [ + { + "bbox": [ + 160, + 462, + 445, + 473 + ], + "type": "text", + "content": "Keywords: LiDAR Semantic Segmentation " + }, + { + "bbox": [ + 160, + 462, + 445, + 473 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 462, + 445, + 473 + ], + "type": "text", + "content": " Domain Generalization" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 513, + 230, + 525 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 513, + 230, + 525 + ], + "spans": [ + { + "bbox": [ + 132, + 513, + 230, + 525 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 538, + 482, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 538, + 482, + 647 + ], + "spans": [ + { + "bbox": [ + 130, + 538, + 482, + 647 + ], + "type": "text", + "content": "Light Detection and Ranging (LiDAR) provides detailed 3D data, making it indispensable for environmental perception in autonomous vehicles. Among the various LiDAR-based perception tasks, semantic segmentation plays a crucial role in understanding the driving scene by classifying each point into multiple classes. While LiDAR-based semantic segmentation [1, 13, 48] have been widely studied, their impressive performance is often constrained to scenarios where source and target datasets align perfectly. However, mismatches between these datasets can lead to significant performance declines. These are mainly attributed to two primary factors: environmental variations as exemplified by the" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 141, + 654, + 420, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 654, + 420, + 665 + ], + "spans": [ + { + "bbox": [ + 141, + 654, + 420, + 665 + ], + "type": "text", + "content": "J. Kim and J. Woo—Both authors contributed equally to this work." + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 134, + 114, + 478, + 237 + ], + "blocks": [ + { + "bbox": [ + 134, + 114, + 478, + 237 + ], + "lines": [ + { + "bbox": [ + 134, + 114, + 478, + 237 + ], + "spans": [ + { + "bbox": [ + 134, + 114, + 478, + 237 + ], + "type": "image", + "image_path": "efa44f2c9d9f265d82a2c66341d6a608c733bc932655c2ab88d10c6a56dee2be.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 247, + 482, + 357 + ], + "lines": [ + { + "bbox": [ + 130, + 247, + 482, + 357 + ], + "spans": [ + { + "bbox": [ + 130, + 247, + 482, + 357 + ], + "type": "text", + "content": "Fig. 1: Motivation of our DDFE: Leveraging Diverse Densities in a Source LiDAR for Domain Generalization - Despite the apparent differences in density distributions between Waymo (64-channel), SemanticKITTI (64-channel), and nuScenes (32-channel), they share regions of overlapping density distributions. For example, the observation that a vehicle at 35 meters in Waymo (top) has a similar density to one at 25 meters in SemanticKITTI (middle) and 12 meters in nuScenes (bottom) underscores this phenomenon. This understanding of varying local density within a source domain serves as a foundation of our domain generalization method. The proposed DDFE transforms features from 3D space to a unified density space without additional training on unseen data, enhancing domain generalization performance." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 385, + 480, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 385, + 480, + 422 + ], + "spans": [ + { + "bbox": [ + 130, + 385, + 480, + 422 + ], + "type": "text", + "content": "Waymo [35] in the USA, the SemanticKITTI [2] in Germany, and the nuScenes [4] in Singapore; and sensor-induced discrepancies, including differences in the number of beams and the field of view (FOV)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 423, + 482, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 423, + 482, + 555 + ], + "spans": [ + { + "bbox": [ + 130, + 423, + 482, + 555 + ], + "type": "text", + "content": "Numerous Unsupervised Domain Adaptation (UDA) studies [17,30,45] address these issues but require additional fine-tuning whenever the target domain changes. In contrast, models deployed with robust Domain Generalization (DG) techniques are increasingly sought after for their potential to enable real-time autonomous driving systems without needing constant fine-tuning. Nevertheless, this crucial field remains under-researched, and its potential has yet to be fully realized. Existing study [16] pinpoints the point cloud density distribution as a prime performance hindrance. While some solutions like point cloud sampling and completion-based methods attempt to remedy this, they fall short in achieving the desired performance [44] and require sequentially labeled data and knowledge of ego-motion [29]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "text", + "content": "In this paper, we introduce a novel perspective in domain generalization by addressing the challenges posed by density variations across different LiDAR sensors. Previous studies [14, 16, 38] have primarily focused on global density differences, typically considering datasets from 64-channel LiDARs like Waymo [35] denser than those from 32-channel LiDARs like nuScenes [4]. However, these approaches oversimplify the inherent complexity of LiDAR data, which exhibits a wide range of density spectra over distance, and interpret them as a single global density value. Contrary to the simplified view of global density comparisons, our research recognizes that LiDAR point clouds are composed of regions" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "text", + "content": "J. Kim et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "content": "with varying densities, as shown in Fig. 1. These variations are influenced by the distance of objects from the LiDAR sensor, leading to a more complex and nuanced understanding of density within LiDAR data." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 152, + 482, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 152, + 482, + 331 + ], + "spans": [ + { + "bbox": [ + 130, + 152, + 482, + 331 + ], + "type": "text", + "content": "To do so, we propose a Density Discriminative Feature Embedding (DDFE) module, designed to enhance the domain generalization capabilities of LiDAR-based segmentation networks by exploiting these density variations. The DDFE module incorporates a beam density estimation module that encodes the specific densities for each 3D voxel, enabling refined density discrimination across regions. The features are then modulated using attention mechanisms guided by beam density, further enhancing the model's ability to generalize across domains with varying density characteristics. Furthermore, we introduce a density soft clipping technique. This method constrains the density spectrum, ensuring it does not encompass density distributions from unseen domains that are absent in source datasets. To supplement this, we use density augmentation to widen the density spectrum of the source data, thereby enhancing its domain generalization capabilities. Extensive experiments validate the superiority of our method over the conventional domain adaptation and generalization methods across multiple 3D backbone networks." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 146, + 331, + 358, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 331, + 358, + 342 + ], + "spans": [ + { + "bbox": [ + 146, + 331, + 358, + 342 + ], + "type": "text", + "content": "In summary, our primary contributions include:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 351, + 480, + 456 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 138, + 351, + 479, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 351, + 479, + 384 + ], + "spans": [ + { + "bbox": [ + 138, + 351, + 479, + 384 + ], + "type": "text", + "content": "- We introduce a new perspective for domain generalization to overcome density variations caused by different LiDAR sensors, utilizing the diverse densities within point clouds observed in the source domain." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 387, + 480, + 409 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 387, + 480, + 409 + ], + "spans": [ + { + "bbox": [ + 138, + 387, + 480, + 409 + ], + "type": "text", + "content": "- We propose a density discriminative feature embedding module designed to identify areas with similar density distribution and amplify relevant features." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 411, + 479, + 433 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 411, + 479, + 433 + ], + "spans": [ + { + "bbox": [ + 138, + 411, + 479, + 433 + ], + "type": "text", + "content": "- We present a simple yet effective data augmentation strategy, aimed at broadening the density spectrum of source data." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 434, + 479, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 434, + 479, + 456 + ], + "spans": [ + { + "bbox": [ + 138, + 434, + 479, + 456 + ], + "type": "text", + "content": "- Extensive experiments demonstrate that our method outperforms state-of-the-art Domain Generalization and Domain Adaptation methods." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 132, + 476, + 237, + 489 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 476, + 237, + 489 + ], + "spans": [ + { + "bbox": [ + 132, + 476, + 237, + 489 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 131, + 502, + 350, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 502, + 350, + 514 + ], + "spans": [ + { + "bbox": [ + 131, + 502, + 350, + 514 + ], + "type": "text", + "content": "2.1 LiDAR-based Semantic Segmentation" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "content": "LiDAR point clouds pose unique challenges due to their irregular, unordered, and unstructured nature. Consequently, approaches to represent point cloud data have been broadly classified into three categories: Projection-based, point-based, and voxel-based methods. Projection-based methods transform a 3D point cloud into a 2D representation through either a spherical or bird's-eye view projection. By doing so, they can utilize lightweight models like 2D convolution [5,8,25,47] or transformer [1,6] rather than 3D convolution. Nevertheless, such methods face the inherent limitation of 2D kernels not preserving the 3D geometric information of the real world. Point-based methods [9,15,36] directly extract features from the point cloud. These methods utilize all the 3D spatial information without distortion. However, the trade-off is a substantial demand on memory and computational resources. In voxel-based approaches [7,11,13,48], point clouds" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 272, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 272, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 272, + 91, + 447, + 102 + ], + "type": "text", + "content": "Rethinking LiDAR Domain Generalization" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 150, + 146, + 459, + 184 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 480, + 137 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 480, + 137 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 480, + 137 + ], + "type": "text", + "content": "Table 1: LiDAR configuration for each dataset (" + }, + { + "bbox": [ + 132, + 114, + 480, + 137 + ], + "type": "inline_equation", + "content": "f_{\\mathrm{min}}" + }, + { + "bbox": [ + 132, + 114, + 480, + 137 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 132, + 114, + 480, + 137 + ], + "type": "inline_equation", + "content": "f_{\\mathrm{max}}" + }, + { + "bbox": [ + 132, + 114, + 480, + 137 + ], + "type": "text", + "content": ": the minimum and maximum LiDAR field of view, " + }, + { + "bbox": [ + 132, + 114, + 480, + 137 + ], + "type": "inline_equation", + "content": "H_{\\mathrm{b}}" + }, + { + "bbox": [ + 132, + 114, + 480, + 137 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 132, + 114, + 480, + 137 + ], + "type": "inline_equation", + "content": "V_{\\mathrm{b}}" + }, + { + "bbox": [ + 132, + 114, + 480, + 137 + ], + "type": "text", + "content": ": the number of horizontal and vertical beams)." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 150, + 146, + 459, + 184 + ], + "lines": [ + { + "bbox": [ + 150, + 146, + 459, + 184 + ], + "spans": [ + { + "bbox": [ + 150, + 146, + 459, + 184 + ], + "type": "table", + "html": "
Waymo [35]SemanticKITTI [2]nuScenes [4]Pandaset [42]SemanticPOSS [24]
Hb25602048108018001800
Vb6464326440
[fmin, fmax](°)[-17.6, +2.4][-24.8, +2.0][-30.0, +10.0][-25.0, +15.0][-16.0, +7.0]
", + "image_path": "5fbd05b24345b381007952556cc49952a4bcd689ad0c44a8fc2e7774f5789eef.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 206, + 479, + 254 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 206, + 479, + 254 + ], + "spans": [ + { + "bbox": [ + 130, + 206, + 479, + 254 + ], + "type": "text", + "content": "are quantized into 3D grids, known as voxels, and features are extracted using 3D convolutional methods. Alongside sparse convolution [12], voxel-based methods achieve high performance by maintaining an actual geometric receptive field while ensuring an efficient computational load." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 272, + 332, + 284 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 272, + 332, + 284 + ], + "spans": [ + { + "bbox": [ + 132, + 272, + 332, + 284 + ], + "type": "text", + "content": "2.2 LiDAR-based Domain Adaptation" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 292, + 482, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 292, + 482, + 495 + ], + "spans": [ + { + "bbox": [ + 130, + 292, + 482, + 495 + ], + "type": "text", + "content": "Variations in LiDAR sensor configurations, as shown in Tab. 1, or alterations in operating environments can lead to differences in the distribution of point cloud data, which, in turn, may adversely affect the perception capabilities of autonomous driving systems. To tackle this challenge without constructing new training datasets for every sensor type and location, research has increasingly focused on unsupervised domain adaptation techniques [17,26,30,31,39,43,45]. These methods strive to preserve performance in the target domain by utilizing labeled data from a source domain alongside unlabeled data from the target domain. C&L [44] leverages sequential data to train a voxel completion network and segments on the reconstructed canonical domain to overcome domain discrepancies. Rochan et al. [28] propose a range-view-based unsupervised domain adaptation method, aligning beam positions between training and target data. LiDAR-UDA [34] employs beam subsampling to mimic different LiDAR sensors, and utilizes cross-frame ensembling and a Learned Aggregation Model to acquire improved pseudo labels. Although these DA methods demonstrate effectiveness, they share a common limitation: the need for individual fine-tuning with target data upon each domain shift." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 514, + 349, + 525 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 514, + 349, + 525 + ], + "spans": [ + { + "bbox": [ + 132, + 514, + 349, + 525 + ], + "type": "text", + "content": "2.3 LiDAR-based Domain Generalization" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": "Domain Generalization [16, 19, 21, 29, 32, 33, 41] aims to improve the performance in scenarios where the target domain data, unseen during training, is encountered. DGLSS [16] leverages sparsity invariance feature consistency and bridges the semantic correlation consistency between the source data and the sparse domain generated by beam sampling. LiDomAug [29] aggregates multiple frames to generate dense world models and augment data by sampling through randomized LiDAR configurations with additional knowledge of ego-motion and the sequentially labeled data. BEV-DG [21] employs a bird's-eye view for enhanced cross-modal learning and develops density-maintained vector modeling for efficient learning of domain-invariant features. LiDOG [32] utilizes semantic priors in a 2D bird's-eye view to extract domain-agnostic features. Sanchez et" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "text", + "content": "J. Kim et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 138, + 114, + 477, + 343 + ], + "blocks": [ + { + "bbox": [ + 138, + 114, + 477, + 343 + ], + "lines": [ + { + "bbox": [ + 138, + 114, + 477, + 343 + ], + "spans": [ + { + "bbox": [ + 138, + 114, + 477, + 343 + ], + "type": "image", + "image_path": "7633e91ca9ca3c8c7c047cfbccb0ea6d849ab5c6135972114b1f96e79d9965dd.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 223, + 360, + 389, + 373 + ], + "lines": [ + { + "bbox": [ + 223, + 360, + 389, + 373 + ], + "spans": [ + { + "bbox": [ + 223, + 360, + 389, + 373 + ], + "type": "text", + "content": "Fig. 2: Overview of the DDFE pipeline." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 392, + 482, + 477 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 392, + 482, + 477 + ], + "spans": [ + { + "bbox": [ + 130, + 392, + 482, + 477 + ], + "type": "text", + "content": "al. [33] introduce a label propagation method that integrates multi-frame aggregation and ego-motion, although it demands significant computational resources due to the adaptation of KPCConv [36]. Distinctly, our approach advances a single-frame domain generalization approach that considers the inherent characteristics of LiDAR. Consequently, our method offers the distinct advantage of bypassing the need for ego-motion and sequentially labeled data during the training and inference phases." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 491, + 202, + 504 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 491, + 202, + 504 + ], + "spans": [ + { + "bbox": [ + 132, + 491, + 202, + 504 + ], + "type": "text", + "content": "3 Method" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 514, + 482, + 612 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 514, + 482, + 612 + ], + "spans": [ + { + "bbox": [ + 130, + 514, + 482, + 612 + ], + "type": "text", + "content": "In this section, we detail our domain generalization method for LiDAR semantic segmentation, which capitalizes on density variations observable within a single source LiDAR. The proposed Density Discriminative Feature Embedding (DDFE) module is primarily composed of four components: Point-voxel feature encoding as detailed in Sec. 3.1, Beam density estimation module in Sec. 3.2, Density soft clipping in Sec. 3.3, and Density-aware embedding module in Sec. 3.4. Additionally, Sec. 3.5 outlines our density augmentation technique. The overall framework of the proposed DDFE module is shown in Fig. 2." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 624, + 306, + 637 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 624, + 306, + 637 + ], + "spans": [ + { + "bbox": [ + 131, + 624, + 306, + 637 + ], + "type": "text", + "content": "3.1 Point-voxel feature encoding" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "content": "We introduce a technique to extract generalized representations across domains from point clouds. The intensity distribution in point clouds is influenced by" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 272, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 272, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 272, + 91, + 447, + 102 + ], + "type": "text", + "content": "Rethinking LiDAR Domain Generalization" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 369 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 369 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 369 + ], + "type": "text", + "content": "the specific LiDAR sensor used, which can adversely affect the performance of perception models during sensor transitions due to the variance in intensity information. Thus, we omit the use of LiDAR intensity values to enhance domain generalization performance. Furthermore, each LiDAR sensor has a unique distribution of measurement errors, particularly within the intra-voxel region. To minimize variance induced by localized sensing noise, we incorporate a direct voxel-wise feature encoding into the PointNet-based feature extraction [27]. This direct encoding, which bypasses localized information within the voxels, effectively dismisses grid size variations, such as those of " + }, + { + "bbox": [ + 130, + 116, + 482, + 369 + ], + "type": "inline_equation", + "content": "20\\mathrm{cm}" + }, + { + "bbox": [ + 130, + 116, + 482, + 369 + ], + "type": "text", + "content": ". Given a point cloud " + }, + { + "bbox": [ + 130, + 116, + 482, + 369 + ], + "type": "inline_equation", + "content": "\\mathbf{P} = \\{p_i\\in \\mathbb{R}^3\\mid i = 1,\\dots,N\\}" + }, + { + "bbox": [ + 130, + 116, + 482, + 369 + ], + "type": "text", + "content": " containing " + }, + { + "bbox": [ + 130, + 116, + 482, + 369 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 116, + 482, + 369 + ], + "type": "text", + "content": " 3D points, we initially partition it into " + }, + { + "bbox": [ + 130, + 116, + 482, + 369 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 130, + 116, + 482, + 369 + ], + "type": "text", + "content": " 3D voxels. For each voxel " + }, + { + "bbox": [ + 130, + 116, + 482, + 369 + ], + "type": "inline_equation", + "content": "v_{j}" + }, + { + "bbox": [ + 130, + 116, + 482, + 369 + ], + "type": "text", + "content": ", it is defined as a set of points and is represented by " + }, + { + "bbox": [ + 130, + 116, + 482, + 369 + ], + "type": "inline_equation", + "content": "\\mathbf{V} = \\{v_{j}\\in \\mathbb{R}^{L\\times 3}\\mid j = 1,\\dots,M\\}" + }, + { + "bbox": [ + 130, + 116, + 482, + 369 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 116, + 482, + 369 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 130, + 116, + 482, + 369 + ], + "type": "text", + "content": " indicates the number of points contained in each voxel. The coordinates of these points are subsequently transformed into the form of " + }, + { + "bbox": [ + 130, + 116, + 482, + 369 + ], + "type": "inline_equation", + "content": "(\\cos (\\theta_j),\\sin (\\theta_j),\\phi_j,r_j)" + }, + { + "bbox": [ + 130, + 116, + 482, + 369 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 116, + 482, + 369 + ], + "type": "inline_equation", + "content": "(\\theta_j,\\phi_j,r_j)" + }, + { + "bbox": [ + 130, + 116, + 482, + 369 + ], + "type": "text", + "content": " correspond to spherical coordinates. These transformed coordinates are then encoded into voxel-wise features " + }, + { + "bbox": [ + 130, + 116, + 482, + 369 + ], + "type": "inline_equation", + "content": "F^{v}\\in \\mathbb{R}^{M\\times 16}" + }, + { + "bbox": [ + 130, + 116, + 482, + 369 + ], + "type": "text", + "content": " through an MLP. We also encode point-wise features " + }, + { + "bbox": [ + 130, + 116, + 482, + 369 + ], + "type": "inline_equation", + "content": "F^{p}\\in \\mathbb{R}^{N\\times 16}" + }, + { + "bbox": [ + 130, + 116, + 482, + 369 + ], + "type": "text", + "content": ", which are essential to produce point-wise outputs. For each point " + }, + { + "bbox": [ + 130, + 116, + 482, + 369 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 116, + 482, + 369 + ], + "type": "text", + "content": " in voxel " + }, + { + "bbox": [ + 130, + 116, + 482, + 369 + ], + "type": "inline_equation", + "content": "v_{j}" + }, + { + "bbox": [ + 130, + 116, + 482, + 369 + ], + "type": "text", + "content": " centered at " + }, + { + "bbox": [ + 130, + 116, + 482, + 369 + ], + "type": "inline_equation", + "content": "(x_{j}^{c},y_{j}^{c},z_{j}^{c})" + }, + { + "bbox": [ + 130, + 116, + 482, + 369 + ], + "type": "text", + "content": ", we compute the offset as " + }, + { + "bbox": [ + 130, + 116, + 482, + 369 + ], + "type": "inline_equation", + "content": "(x_{jk} - x_{j}^{c},y_{jk} - y_{j}^{c},z_{jk} - z_{j}^{c})" + }, + { + "bbox": [ + 130, + 116, + 482, + 369 + ], + "type": "text", + "content": ", with " + }, + { + "bbox": [ + 130, + 116, + 482, + 369 + ], + "type": "inline_equation", + "content": "k = 1,\\ldots ,L_{j}" + }, + { + "bbox": [ + 130, + 116, + 482, + 369 + ], + "type": "text", + "content": ". Here, " + }, + { + "bbox": [ + 130, + 116, + 482, + 369 + ], + "type": "inline_equation", + "content": "L_{j}" + }, + { + "bbox": [ + 130, + 116, + 482, + 369 + ], + "type": "text", + "content": " represents the number of points belonging to " + }, + { + "bbox": [ + 130, + 116, + 482, + 369 + ], + "type": "inline_equation", + "content": "v_{j}" + }, + { + "bbox": [ + 130, + 116, + 482, + 369 + ], + "type": "text", + "content": ". These offsets are then processed by a point head to generate point-wise features." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 384, + 326, + 396 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 384, + 326, + 396 + ], + "spans": [ + { + "bbox": [ + 132, + 384, + 326, + 396 + ], + "type": "text", + "content": "3.2 Beam density estimation module" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 403, + 482, + 488 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 403, + 482, + 488 + ], + "spans": [ + { + "bbox": [ + 130, + 403, + 482, + 488 + ], + "type": "text", + "content": "In this section, we detail the method to compute the density expectation for each ray emitted from a LiDAR sensor. Since all rays from a LiDAR originate from a singular source and are emitted at a fixed angle, the density associated with each beam can be deduced using a spherical projection informed by the beam configuration. The LiDAR sensor configuration, accessible from low-level sources such as an SDK, determines the set of horizontal and vertical inclinations " + }, + { + "bbox": [ + 130, + 403, + 482, + 488 + ], + "type": "inline_equation", + "content": "\\mathbf{C}_h" + }, + { + "bbox": [ + 130, + 403, + 482, + 488 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 403, + 482, + 488 + ], + "type": "inline_equation", + "content": "\\mathbf{C}_v" + }, + { + "bbox": [ + 130, + 403, + 482, + 488 + ], + "type": "text", + "content": " for each beam in this manner:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 152, + 495, + 481, + 524 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 495, + 481, + 524 + ], + "spans": [ + { + "bbox": [ + 152, + 495, + 481, + 524 + ], + "type": "interline_equation", + "content": "\\mathbf {C} _ {h} = \\left\\{\\frac {2 \\pi i}{H _ {\\mathrm {b}}} \\right\\} _ {i \\in \\{1, \\dots , H _ {\\mathrm {b}} \\}}, \\mathbf {C} _ {v} = \\left\\{\\frac {(f _ {\\operatorname* {m a x}} - f _ {\\operatorname* {m i n}}) j}{V _ {\\mathrm {b}}} + f _ {\\operatorname* {m i n}} \\right\\} _ {j \\in \\{1, \\dots , V _ {\\mathrm {b}} \\}}, \\qquad (1)", + "image_path": "9e574d4cc841a4df4bf90c237b30b84ba171de0bbd9a2f6c397d80e975259e7e.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 529, + 482, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 529, + 482, + 579 + ], + "spans": [ + { + "bbox": [ + 130, + 529, + 482, + 579 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 529, + 482, + 579 + ], + "type": "inline_equation", + "content": "f_{\\mathrm{min}}" + }, + { + "bbox": [ + 130, + 529, + 482, + 579 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 529, + 482, + 579 + ], + "type": "inline_equation", + "content": "f_{\\mathrm{max}}" + }, + { + "bbox": [ + 130, + 529, + 482, + 579 + ], + "type": "text", + "content": " denote the minimum and maximum LiDAR field of view, while " + }, + { + "bbox": [ + 130, + 529, + 482, + 579 + ], + "type": "inline_equation", + "content": "H_{\\mathrm{b}}" + }, + { + "bbox": [ + 130, + 529, + 482, + 579 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 529, + 482, + 579 + ], + "type": "inline_equation", + "content": "V_{\\mathrm{b}}" + }, + { + "bbox": [ + 130, + 529, + 482, + 579 + ], + "type": "text", + "content": " represent the number of horizontal and vertical beams, detailed in Tab. 1. To project LiDAR beam inclinations onto spherical projected image coordinates, we use the projection function " + }, + { + "bbox": [ + 130, + 529, + 482, + 579 + ], + "type": "inline_equation", + "content": "\\pi : (\\theta, \\phi) \\to (\\dot{\\theta}, \\dot{\\phi})" + }, + { + "bbox": [ + 130, + 529, + 482, + 579 + ], + "type": "text", + "content": " defined as:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 224, + 587, + 481, + 619 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 224, + 587, + 481, + 619 + ], + "spans": [ + { + "bbox": [ + 224, + 587, + 481, + 619 + ], + "type": "interline_equation", + "content": "\\dot {\\theta} = \\left\\lfloor \\frac {\\theta}{2 \\pi} W \\right\\rfloor , \\dot {\\phi} = \\left\\lfloor \\frac {\\phi - f _ {\\operatorname* {m i n}} ^ {\\operatorname* {p r o j}}}{f _ {\\operatorname* {m a x}} ^ {\\operatorname* {p r o j}} - f _ {\\operatorname* {m i n}} ^ {\\operatorname* {p r o j}}} H \\right\\rfloor , \\tag {2}", + "image_path": "174f757f9d04e0b37f72b71d96769695a2d455392d1effec3ae6931ac4bb61ec.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 625, + 482, + 669 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 625, + 482, + 669 + ], + "spans": [ + { + "bbox": [ + 130, + 625, + 482, + 669 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 625, + 482, + 669 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 130, + 625, + 482, + 669 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 625, + 482, + 669 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 130, + 625, + 482, + 669 + ], + "type": "text", + "content": " are the height and width resolutions of the projected image, respectively set to " + }, + { + "bbox": [ + 130, + 625, + 482, + 669 + ], + "type": "inline_equation", + "content": "H = 512" + }, + { + "bbox": [ + 130, + 625, + 482, + 669 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 625, + 482, + 669 + ], + "type": "inline_equation", + "content": "W = 5120" + }, + { + "bbox": [ + 130, + 625, + 482, + 669 + ], + "type": "text", + "content": ". We define the projected image's field of view as " + }, + { + "bbox": [ + 130, + 625, + 482, + 669 + ], + "type": "inline_equation", + "content": "\\left[f_{\\mathrm{min}}^{\\mathrm{proj}}, f_{\\mathrm{max}}^{\\mathrm{proj}}\\right] = [-30.0, 15.0]" + }, + { + "bbox": [ + 130, + 625, + 482, + 669 + ], + "type": "text", + "content": " to accommodate a range of LiDAR sensors." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "text", + "content": "J. Kim et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": "Given this, beam configurations can be transformed into 1-D binary vectors " + }, + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "inline_equation", + "content": "\\mathbf{B}_v\\in \\mathbb{R}^H" + }, + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "inline_equation", + "content": "\\mathbf{B}_h\\in \\mathbb{R}^W" + }, + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 152, + 152, + 480, + 183 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 152, + 480, + 183 + ], + "spans": [ + { + "bbox": [ + 152, + 152, + 480, + 183 + ], + "type": "interline_equation", + "content": "\\mathbf {B} _ {h} (\\dot {\\theta}) = \\mathbb {1} _ {\\mathbf {C} _ {h}} (\\theta), \\mathbf {B} _ {v} (\\dot {\\phi}) = \\mathbb {1} _ {\\mathbf {C} _ {v}} (\\phi), \\text {w h e r e} \\mathbb {1} _ {\\mathbf {C}} (x) = \\left\\{ \\begin{array}{l l} 1, & \\text {i f} x \\in \\mathbf {C} \\\\ 0, & \\text {o t h e r w i s e} \\end{array} , \\right. \\tag {3}", + "image_path": "b78699189504a202c7366e778d73c2f97b7e837f9d3d3fe52ec252846fa9610f.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 193, + 479, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 193, + 479, + 266 + ], + "spans": [ + { + "bbox": [ + 130, + 193, + 479, + 266 + ], + "type": "text", + "content": "where the indicator function " + }, + { + "bbox": [ + 130, + 193, + 479, + 266 + ], + "type": "inline_equation", + "content": "\\mathbb{1}(\\cdot)" + }, + { + "bbox": [ + 130, + 193, + 479, + 266 + ], + "type": "text", + "content": " yields 1 when the azimuth or elevation " + }, + { + "bbox": [ + 130, + 193, + 479, + 266 + ], + "type": "inline_equation", + "content": "(\\theta, \\phi)" + }, + { + "bbox": [ + 130, + 193, + 479, + 266 + ], + "type": "text", + "content": " of a pixel " + }, + { + "bbox": [ + 130, + 193, + 479, + 266 + ], + "type": "inline_equation", + "content": "(\\dot{\\theta}, \\dot{\\phi})" + }, + { + "bbox": [ + 130, + 193, + 479, + 266 + ], + "type": "text", + "content": " corresponds to the beam inclinations specified in " + }, + { + "bbox": [ + 130, + 193, + 479, + 266 + ], + "type": "inline_equation", + "content": "\\mathbf{C}" + }, + { + "bbox": [ + 130, + 193, + 479, + 266 + ], + "type": "text", + "content": " in Eq. (1). Here, " + }, + { + "bbox": [ + 130, + 193, + 479, + 266 + ], + "type": "inline_equation", + "content": "\\mathbf{C}" + }, + { + "bbox": [ + 130, + 193, + 479, + 266 + ], + "type": "text", + "content": " can be either " + }, + { + "bbox": [ + 130, + 193, + 479, + 266 + ], + "type": "inline_equation", + "content": "\\mathbf{C}_h" + }, + { + "bbox": [ + 130, + 193, + 479, + 266 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 130, + 193, + 479, + 266 + ], + "type": "inline_equation", + "content": "\\mathbf{C}_v" + }, + { + "bbox": [ + 130, + 193, + 479, + 266 + ], + "type": "text", + "content": " that represent the locations of projected vertical and horizontal beams, respectively. To compute the beam density, we convolve the binary vectors " + }, + { + "bbox": [ + 130, + 193, + 479, + 266 + ], + "type": "inline_equation", + "content": "\\mathbf{B}_h" + }, + { + "bbox": [ + 130, + 193, + 479, + 266 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 193, + 479, + 266 + ], + "type": "inline_equation", + "content": "\\mathbf{B}_v" + }, + { + "bbox": [ + 130, + 193, + 479, + 266 + ], + "type": "text", + "content": ", with four distinct 1-D Gaussian kernels, each characterized by standard deviations " + }, + { + "bbox": [ + 130, + 193, + 479, + 266 + ], + "type": "inline_equation", + "content": "\\sigma_k = \\{10, 30, 50, 70\\}" + }, + { + "bbox": [ + 130, + 193, + 479, + 266 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 235, + 278, + 480, + 290 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 235, + 278, + 480, + 290 + ], + "spans": [ + { + "bbox": [ + 235, + 278, + 480, + 290 + ], + "type": "interline_equation", + "content": "\\tilde {\\mathbf {B}} _ {h} = \\mathbf {B} _ {h} * \\mathbf {G} _ {\\sigma_ {k}}, \\quad \\tilde {\\mathbf {B}} _ {v} = \\mathbf {B} _ {v} * \\mathbf {G} _ {\\sigma_ {k}}. \\tag {4}", + "image_path": "79b121a211788d89f82cc5c56d05abbe047b2781351d5b5dd882f3c9d26e8c3e.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 301, + 383, + 314 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 301, + 383, + 314 + ], + "spans": [ + { + "bbox": [ + 131, + 301, + 383, + 314 + ], + "type": "text", + "content": "Finally, we define beam density " + }, + { + "bbox": [ + 131, + 301, + 383, + 314 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_i" + }, + { + "bbox": [ + 131, + 301, + 383, + 314 + ], + "type": "text", + "content": " of point " + }, + { + "bbox": [ + 131, + 301, + 383, + 314 + ], + "type": "inline_equation", + "content": "p_i" + }, + { + "bbox": [ + 131, + 301, + 383, + 314 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 228, + 326, + 480, + 354 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 228, + 326, + 480, + 354 + ], + "spans": [ + { + "bbox": [ + 228, + 326, + 480, + 354 + ], + "type": "interline_equation", + "content": "\\mathcal {D} _ {i} = \\left[ \\sqrt {\\hat {\\mathbf {B}} _ {h} ^ {(k)} (\\dot {\\theta} _ {i}) \\cdot \\hat {\\mathbf {B}} _ {v} ^ {(k)} (\\dot {\\phi} _ {i}) / r _ {i} ^ {2}} \\right] _ {k = 1} ^ {4}, \\tag {5}", + "image_path": "8eeb2f39f8c4a0cbebdd3fdab0323b7d4deacc537839ced67f970a738c0c54b4.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 363, + 479, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 363, + 479, + 411 + ], + "spans": [ + { + "bbox": [ + 130, + 363, + 479, + 411 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 363, + 479, + 411 + ], + "type": "inline_equation", + "content": "r_i" + }, + { + "bbox": [ + 130, + 363, + 479, + 411 + ], + "type": "text", + "content": " is the radial distance of a LiDAR point " + }, + { + "bbox": [ + 130, + 363, + 479, + 411 + ], + "type": "inline_equation", + "content": "p_i" + }, + { + "bbox": [ + 130, + 363, + 479, + 411 + ], + "type": "text", + "content": " in the spherical coordinates, and " + }, + { + "bbox": [ + 130, + 363, + 479, + 411 + ], + "type": "inline_equation", + "content": "[\\cdot]_{k=1}^4" + }, + { + "bbox": [ + 130, + 363, + 479, + 411 + ], + "type": "text", + "content": " denotes the concatenation operation. Taking into account that the density of the beam diminishes in proportion to the square of the distance due to its radial emission, we incorporate the inverse of " + }, + { + "bbox": [ + 130, + 363, + 479, + 411 + ], + "type": "inline_equation", + "content": "r^2" + }, + { + "bbox": [ + 130, + 363, + 479, + 411 + ], + "type": "text", + "content": " in our computation." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 435, + 265, + 448 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 435, + 265, + 448 + ], + "spans": [ + { + "bbox": [ + 131, + 435, + 265, + 448 + ], + "type": "text", + "content": "3.3 Density soft clipping" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 460, + 479, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 460, + 479, + 510 + ], + "spans": [ + { + "bbox": [ + 130, + 460, + 479, + 510 + ], + "type": "text", + "content": "When exposed to densities that deviate from those in the source dataset, the model is susceptible to performance degradation in an unseen domain. To alleviate this issue, we introduce a density soft clipping method that confines the density spectrum using the " + }, + { + "bbox": [ + 130, + 460, + 479, + 510 + ], + "type": "inline_equation", + "content": "\\tanh(\\cdot)" + }, + { + "bbox": [ + 130, + 460, + 479, + 510 + ], + "type": "text", + "content": " function as follows:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 207, + 521, + 406, + 571 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 207, + 521, + 406, + 571 + ], + "spans": [ + { + "bbox": [ + 207, + 521, + 406, + 571 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {D} _ {i} ^ {c} = \\tanh \\left(\\frac {\\mathcal {D} _ {i} - m}{l}\\right) l + m, \\\\ m = \\frac {\\mathcal {P} _ {9 0} (\\mathcal {D}) + \\mathcal {P} _ {1 0} (\\mathcal {D})}{2}, l = \\frac {\\mathcal {P} _ {9 0} (\\mathcal {D}) - \\mathcal {P} _ {1 0} (\\mathcal {D})}{2}, \\\\ \\end{array}", + "image_path": "aec4677362de41a0dd933aa5116ebbcff55dab02d33faae0d5a6b4d64419e16b.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 581, + 479, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 581, + 479, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 581, + 479, + 666 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 581, + 479, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_i^c" + }, + { + "bbox": [ + 130, + 581, + 479, + 666 + ], + "type": "text", + "content": " is clipped density of point " + }, + { + "bbox": [ + 130, + 581, + 479, + 666 + ], + "type": "inline_equation", + "content": "p_i" + }, + { + "bbox": [ + 130, + 581, + 479, + 666 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 581, + 479, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_{90}" + }, + { + "bbox": [ + 130, + 581, + 479, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 581, + 479, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_{10}" + }, + { + "bbox": [ + 130, + 581, + 479, + 666 + ], + "type": "text", + "content": " are functions that extract " + }, + { + "bbox": [ + 130, + 581, + 479, + 666 + ], + "type": "inline_equation", + "content": "90^{\\mathrm{th}}" + }, + { + "bbox": [ + 130, + 581, + 479, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 581, + 479, + 666 + ], + "type": "inline_equation", + "content": "10^{\\mathrm{th}}" + }, + { + "bbox": [ + 130, + 581, + 479, + 666 + ], + "type": "text", + "content": " channel-wise percentile values from the density embedding of the training domain, respectively. Employing these percentile values allows us to discount the outlier density values. To calculate the percentile on the fly under memory constraints, we adopt the Reservoir Sampling [37] technique with a sample size " + }, + { + "bbox": [ + 130, + 581, + 479, + 666 + ], + "type": "inline_equation", + "content": "N = 1000" + }, + { + "bbox": [ + 130, + 581, + 479, + 666 + ], + "type": "text", + "content": ". Please refer to the supplementary material for a more detailed algorithm description." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 272, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 272, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 272, + 91, + 447, + 101 + ], + "type": "text", + "content": "Rethinking LiDAR Domain Generalization" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 99 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 99 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 99 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 331, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 331, + 129 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 331, + 129 + ], + "type": "text", + "content": "3.4 Density-aware embedding module" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 135, + 482, + 221 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 135, + 482, + 221 + ], + "spans": [ + { + "bbox": [ + 130, + 135, + 482, + 221 + ], + "type": "text", + "content": "For each point " + }, + { + "bbox": [ + 130, + 135, + 482, + 221 + ], + "type": "inline_equation", + "content": "p_i" + }, + { + "bbox": [ + 130, + 135, + 482, + 221 + ], + "type": "text", + "content": ", a point-wise density embedding feature " + }, + { + "bbox": [ + 130, + 135, + 482, + 221 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_i^c" + }, + { + "bbox": [ + 130, + 135, + 482, + 221 + ], + "type": "text", + "content": " is obtained via the beam density estimation module incorporating density soft clipping. This feature serves as an input of both a point-wise attention function " + }, + { + "bbox": [ + 130, + 135, + 482, + 221 + ], + "type": "inline_equation", + "content": "f_p" + }, + { + "bbox": [ + 130, + 135, + 482, + 221 + ], + "type": "text", + "content": " and a voxelwise attention function " + }, + { + "bbox": [ + 130, + 135, + 482, + 221 + ], + "type": "inline_equation", + "content": "f_v" + }, + { + "bbox": [ + 130, + 135, + 482, + 221 + ], + "type": "text", + "content": ". These mechanisms are pivotal for crafting domain-invariant density-discriminative features. In the point-wise attention framework, " + }, + { + "bbox": [ + 130, + 135, + 482, + 221 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_i^c" + }, + { + "bbox": [ + 130, + 135, + 482, + 221 + ], + "type": "text", + "content": " is synchronized dimensionally with the point-wise feature " + }, + { + "bbox": [ + 130, + 135, + 482, + 221 + ], + "type": "inline_equation", + "content": "F_i^p" + }, + { + "bbox": [ + 130, + 135, + 482, + 221 + ], + "type": "text", + "content": ", facilitating the creation of a density discriminative feature " + }, + { + "bbox": [ + 130, + 135, + 482, + 221 + ], + "type": "inline_equation", + "content": "\\hat{F}_i^p" + }, + { + "bbox": [ + 130, + 135, + 482, + 221 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 266, + 228, + 481, + 243 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 266, + 228, + 481, + 243 + ], + "spans": [ + { + "bbox": [ + 266, + 228, + 481, + 243 + ], + "type": "interline_equation", + "content": "\\hat {F} _ {i} ^ {p} = f _ {p} \\left(\\mathcal {D} _ {i} ^ {c}\\right) \\odot F _ {i} ^ {p}, \\tag {7}", + "image_path": "0d8f5cf3102e2e6404512338873786e4145e5d729a7c676c2bf1f6bba5049155.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 250, + 480, + 299 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 250, + 480, + 299 + ], + "spans": [ + { + "bbox": [ + 130, + 250, + 480, + 299 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 250, + 480, + 299 + ], + "type": "inline_equation", + "content": "f_{p}" + }, + { + "bbox": [ + 130, + 250, + 480, + 299 + ], + "type": "text", + "content": " consists of two 1D convolution layers followed by a sigmoid function. For voxel-wise attention, it begins by averaging the density embedding features for all points encapsulated within a voxel. This aggregated measure then undergoes a similar treatment as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 187, + 304, + 481, + 334 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 187, + 304, + 481, + 334 + ], + "spans": [ + { + "bbox": [ + 187, + 304, + 481, + 334 + ], + "type": "interline_equation", + "content": "\\hat {F} _ {j} ^ {v} = \\operatorname {C o n c a t} \\left(f _ {v} \\left(\\mathcal {D} _ {j} ^ {c}\\right) \\odot F _ {j} ^ {v}, g \\left(F _ {j} ^ {p}\\right)\\right), \\quad \\mathcal {D} _ {j} ^ {c} = \\frac {1}{\\left| L _ {j} \\right|} \\sum_ {p _ {i} \\in v _ {j}} \\mathcal {D} _ {i} ^ {c}, \\tag {8}", + "image_path": "17bba7b06f62528239fa0e13e754e180408d9e339ec29a33eecf4904bf6c45a6.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 342, + 482, + 428 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 342, + 482, + 428 + ], + "spans": [ + { + "bbox": [ + 130, + 342, + 482, + 428 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 342, + 482, + 428 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_j^c" + }, + { + "bbox": [ + 130, + 342, + 482, + 428 + ], + "type": "text", + "content": " is a clipped density of voxel " + }, + { + "bbox": [ + 130, + 342, + 482, + 428 + ], + "type": "inline_equation", + "content": "v_{j}" + }, + { + "bbox": [ + 130, + 342, + 482, + 428 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 342, + 482, + 428 + ], + "type": "inline_equation", + "content": "\\hat{F}_j^v" + }, + { + "bbox": [ + 130, + 342, + 482, + 428 + ], + "type": "text", + "content": " is a voxel-wise density discriminative feature and " + }, + { + "bbox": [ + 130, + 342, + 482, + 428 + ], + "type": "inline_equation", + "content": "f_{v}" + }, + { + "bbox": [ + 130, + 342, + 482, + 428 + ], + "type": "text", + "content": " consists of two 1D convolution layers followed by a sigmoid function. An aggregation function " + }, + { + "bbox": [ + 130, + 342, + 482, + 428 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 130, + 342, + 482, + 428 + ], + "type": "text", + "content": " for the voxel-wise feature " + }, + { + "bbox": [ + 130, + 342, + 482, + 428 + ], + "type": "inline_equation", + "content": "F_{j}^{v}" + }, + { + "bbox": [ + 130, + 342, + 482, + 428 + ], + "type": "text", + "content": " is adjusted by voxel-wise attention and concatenated with the max pooled point-wise feature within the " + }, + { + "bbox": [ + 130, + 342, + 482, + 428 + ], + "type": "inline_equation", + "content": "v_{j}" + }, + { + "bbox": [ + 130, + 342, + 482, + 428 + ], + "type": "text", + "content": ". Finally, the extracted features are passed through a single 1D convolutional layer, resulting in 32-channel features. These density-aware features are subsequently fed into a 3D backbone for LiDAR semantic segmentation." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 443, + 275, + 455 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 443, + 275, + 455 + ], + "spans": [ + { + "bbox": [ + 132, + 443, + 275, + 455 + ], + "type": "text", + "content": "3.5 Density Augmentation" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 462, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 462, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 462, + 482, + 666 + ], + "type": "text", + "content": "The DDFE module aims to align densities across varied sensor domains by considering each dataset as a collection of multiple density domains. This method, while effective in many scenarios, may face challenges when there's little to no overlap in the density ranges between the source and unseen datasets. To address this potential issue, we introduce a density augmentation method aimed at broadening the density spectrum covered by the training data. Several existing 3D point cloud augmentation strategies, such as [16, 22, 38] outlined in recent studies, attempt to address density variations either directly or by implication. We adapt the Mix3D [22] for our purposes, which we refer to as enhanced-Mix3D, incorporating random translations along the direction of ego-vehicle movement, along with additional rotational transformations to simulate a wider range of density variations. We also employ beam sampling technique [16, 38], selectively eliminating specific LiDAR beams, to amplify the density in the lower direction. During the training phase, we apply both enhanced-Mix3D and beam sampling augmentations with a set probability of 0.5, aiming to effectively simulate diverse density conditions. Please refer to the supplementary material for more detailed implementation." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 219, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 219, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 219, + 101 + ], + "type": "text", + "content": "J. Kim et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 230, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 230, + 129 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 230, + 129 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 137, + 482, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 137, + 482, + 232 + ], + "spans": [ + { + "bbox": [ + 130, + 137, + 482, + 232 + ], + "type": "text", + "content": "In this section, we demonstrate the domain generalization performance of our method through extensive experiments. The implementation details and experimental settings of the proposed method are detailed in Sec. 4.1. The configuration of datasets used for evaluation is detailed in Sec. 4.2. A comparative analysis with recent domain generalization and domain adaptation methods is presented in Sec. 4.3. The effectiveness of individual components within the proposed DDFE and computational cost are analyzed in Section Sec. 4.4. Detailed per-class performance metrics are available in the supplementary material." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 248, + 279, + 260 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 248, + 279, + 260 + ], + "spans": [ + { + "bbox": [ + 132, + 248, + 279, + 260 + ], + "type": "text", + "content": "4.1 Implementation Details" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 265, + 481, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 265, + 481, + 324 + ], + "spans": [ + { + "bbox": [ + 130, + 265, + 481, + 324 + ], + "type": "text", + "content": "We employ the point head inspired by Cylinder3D [48] to produce point-wise outputs as shown in Fig. 2. During the training phase, we incorporate the Lovasz-Softmax loss " + }, + { + "bbox": [ + 130, + 265, + 481, + 324 + ], + "type": "inline_equation", + "content": "\\mathcal{L}^{lovasz}" + }, + { + "bbox": [ + 130, + 265, + 481, + 324 + ], + "type": "text", + "content": " [3] along with the Weighted Cross-Entropy loss " + }, + { + "bbox": [ + 130, + 265, + 481, + 324 + ], + "type": "inline_equation", + "content": "\\mathcal{L}^{wce}" + }, + { + "bbox": [ + 130, + 265, + 481, + 324 + ], + "type": "text", + "content": ". The total loss is composed of an equal-weighted combination of point-wise and voxelwise losses as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 212, + 331, + 481, + 345 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 212, + 331, + 481, + 345 + ], + "spans": [ + { + "bbox": [ + 212, + 331, + 481, + 345 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {t o t a l}} = \\mathcal {L} _ {\\text {p o i n t}} ^ {\\text {l o v a s z}} + \\mathcal {L} _ {\\text {p o i n t}} ^ {\\text {w c e}} + \\mathcal {L} _ {\\text {v o x e l}} ^ {\\text {l o v a s z}} + \\mathcal {L} _ {\\text {v o x e l}} ^ {\\text {w c e}}. \\tag {9}", + "image_path": "38a73ba1cdde9b49290ddd07ce2b155d0c25547d17f2033f6c009624fca76ffc.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 350, + 482, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 350, + 482, + 434 + ], + "spans": [ + { + "bbox": [ + 130, + 350, + 482, + 434 + ], + "type": "text", + "content": "We use the Adam optimizer with an initial learning rate of 1e-3. This rate is decreased by a factor of 0.99 with every epoch. The training is conducted over 30 epochs with a batch size of two on a single NVIDIA RTX 3090, and the epoch yielding the highest source validation mIoU is chosen. For voxelization, we adopt a cubic size of [20cm, 20cm, 20cm]. In cases where multiple point labels are found within a voxel, the voxel label is determined based on the predominant label, aligning with the approach in Cylinder3D [48]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 449, + 205, + 459 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 449, + 205, + 459 + ], + "spans": [ + { + "bbox": [ + 132, + 449, + 205, + 459 + ], + "type": "text", + "content": "4.2 Datasets" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 465, + 482, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 465, + 482, + 586 + ], + "spans": [ + { + "bbox": [ + 130, + 465, + 482, + 586 + ], + "type": "text", + "content": "In our experimental setup, we employ three datasets: Waymo [35], nuScenes [4], and SemanticKITTI [2]. Notably, the Waymo and SemanticKITTI datasets use 64-channel LiDAR, while the nuScenes employs a 32-channel LiDAR. For detailed information, please refer to Tab. 1. In line with methods from previous studies [16,29], we split the nuScenes sequences into 700 for training and 150 for validation. The Waymo dataset is partitioned into 798 sequences for training and 202 for validation. Regarding the SemanticKITTI dataset, we use sequences 00 to 10 for training, setting aside sequence 08 exclusively for validation. In integrating class variations across datasets, we follow the class mapping configurations from previous studies for fair comparisons." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 600, + 411, + 612 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 600, + 411, + 612 + ], + "spans": [ + { + "bbox": [ + 132, + 600, + 411, + 612 + ], + "type": "text", + "content": "4.3 Comparison to State-of-the-Art DA/DG Methods" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 617, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 481, + 665 + ], + "type": "text", + "content": "The field of domain generalization of LiDAR-based semantic segmentation is still in its formative stages, with a lack of universally accepted experimental standards. This situation has led to a diversity of experimental designs across different studies. To facilitate fair comparisons, we align our experimental framework" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 272, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 272, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 272, + 91, + 447, + 102 + ], + "type": "text", + "content": "Rethinking LiDAR Domain Generalization" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 133, + 156, + 475, + 238 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 480, + 148 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 480, + 148 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 480, + 148 + ], + "type": "text", + "content": "Table 2: Comparison with domain generalization methods based on MinkowskiNet architecture using the mIoU. The best and the second best results are highlighted in bold and underline, respectively." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 133, + 156, + 475, + 238 + ], + "lines": [ + { + "bbox": [ + 133, + 156, + 475, + 238 + ], + "spans": [ + { + "bbox": [ + 133, + 156, + 475, + 238 + ], + "type": "table", + "html": "
MethodDASourceWKNSourceKWNSourceNWK
Base75.3749.4047.8357.3135.2437.4265.7838.6536.24
IBN-Net [23]75.4751.1344.7257.7436.9938.7465.3136.5336.93
MLDG [20]72.4748.9448.6456.2635.3936.7761.3236.3332.70
COSMIX (W) [30]✓---K49.3539.4638.94N---
COSMIX (K) [30]✓66.6844.7149.96------
COSMIX (N) [30]✓65.6840.9947.9849.9838.0543.25---
DGLSS [16]75.2851.2349.6159.6240.6744.8365.3240.9338.98
Ours76.1557.0756.7562.5042.7349.4368.1645.9846.52
", + "image_path": "9d1421215022dfa6138e74ccca9562741d63cd87b966cb42e87064228c917591.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 166, + 268, + 443, + 348 + ], + "blocks": [ + { + "bbox": [ + 142, + 247, + 470, + 258 + ], + "lines": [ + { + "bbox": [ + 142, + 247, + 470, + 258 + ], + "spans": [ + { + "bbox": [ + 142, + 247, + 470, + 258 + ], + "type": "text", + "content": "Table 3: Comparison with domain adaptation and data augmentation methods." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 166, + 268, + 443, + 348 + ], + "lines": [ + { + "bbox": [ + 166, + 268, + 443, + 348 + ], + "spans": [ + { + "bbox": [ + 166, + 268, + 443, + 348 + ], + "type": "table", + "html": "
BackboneMethodsK→NN→KBackboneMethodsK→NN→K
MinkNet42 [7]Base37.836.1C&L [44]Base27.923.5
CutMix [46]37.137.6SWD [18]27.724.5
Copy-Paste [10]38.541.13DGCA [39]27.423.9
Mix3D [22]43.144.7C&L [44]31.633.7
PolarMix [40]45.839.1LiDomAug [29]39.237.9
LiDomAug [29]45.948.3LiDAR-UDA [34]41.834.0
Ours (v=5cm)48.651.3Ours (v=5cm)42.541.0
Ours (v=20cm)50.146.3Ours (v=20cm)47.140.3
", + "image_path": "aa77591b405832bb7019b813abe4833e3f4d73c0636381fa260c8e4176ba33e4.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 370, + 480, + 431 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 370, + 480, + 431 + ], + "spans": [ + { + "bbox": [ + 130, + 370, + 480, + 431 + ], + "type": "text", + "content": "with those utilized in seminal works such as LiDomAug [29] and DGLSS [16]. This approach allows us to benchmark our method against a range of Domain Adaptation (DA) and Domain Generalization (DG) techniques, showcasing the effectiveness of our method in the context of evolving domain generalization challenges." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 449, + 481, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 449, + 481, + 641 + ], + "spans": [ + { + "bbox": [ + 130, + 449, + 481, + 641 + ], + "type": "text", + "content": "Experiments in the DGLSS Setting We compare our method with a domain adaptation method [30] and three domain generalization methods [16,20,23], following the experimental framework defined in DGLSS [16] using Waymo (W), SemanticKITTI (K), and nuScenes (N). We adopt MinkowskiNet [7] as our backbone network, aligning with the configuration utilized in DGLSS. The results in Tab. 2 indicate that the proposed method consistently outperforms DGLSS across all datasets. Impressively, our method not only demonstrates superior domain generalization performance but also dominates in the source-to-source settings " + }, + { + "bbox": [ + 130, + 449, + 481, + 641 + ], + "type": "inline_equation", + "content": "(\\mathrm{W} \\rightarrow \\mathrm{W}, \\mathrm{K} \\rightarrow \\mathrm{K}, \\mathrm{N} \\rightarrow \\mathrm{N})" + }, + { + "bbox": [ + 130, + 449, + 481, + 641 + ], + "type": "text", + "content": ". Unlike DGLSS, which tailors its augmentation strategies for each specific dataset, our method employs a uniform hyperparameter setting across all datasets, achieving enhanced domain generalization results. Our method achieves an average increase of " + }, + { + "bbox": [ + 130, + 449, + 481, + 641 + ], + "type": "inline_equation", + "content": "+12.9\\%" + }, + { + "bbox": [ + 130, + 449, + 481, + 641 + ], + "type": "text", + "content": " over DGLSS for unseen datasets using Waymo as the source data " + }, + { + "bbox": [ + 130, + 449, + 481, + 641 + ], + "type": "inline_equation", + "content": "(\\mathrm{W} \\rightarrow \\mathrm{K}, \\mathrm{W} \\rightarrow \\mathrm{N})" + }, + { + "bbox": [ + 130, + 449, + 481, + 641 + ], + "type": "text", + "content": ". With SemanticKITTI as the source, there's an average enhancement of " + }, + { + "bbox": [ + 130, + 449, + 481, + 641 + ], + "type": "inline_equation", + "content": "+7.8\\%" + }, + { + "bbox": [ + 130, + 449, + 481, + 641 + ], + "type": "text", + "content": " for other unseen domains " + }, + { + "bbox": [ + 130, + 449, + 481, + 641 + ], + "type": "inline_equation", + "content": "(\\mathrm{K} \\rightarrow \\mathrm{W}, \\mathrm{K} \\rightarrow \\mathrm{N})" + }, + { + "bbox": [ + 130, + 449, + 481, + 641 + ], + "type": "text", + "content": ", and employing nuScenes as the source data yields an average increase of " + }, + { + "bbox": [ + 130, + 449, + 481, + 641 + ], + "type": "inline_equation", + "content": "+15.8\\%" + }, + { + "bbox": [ + 130, + 449, + 481, + 641 + ], + "type": "text", + "content": " on other unseen datasets " + }, + { + "bbox": [ + 130, + 449, + 481, + 641 + ], + "type": "inline_equation", + "content": "(\\mathrm{N} \\rightarrow \\mathrm{W}, \\mathrm{N} \\rightarrow \\mathrm{K})" + }, + { + "bbox": [ + 130, + 449, + 481, + 641 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 642, + 480, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 642, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 642, + 480, + 665 + ], + "type": "text", + "content": "These results demonstrate the robustness of the proposed method in addressing domain discrepancies." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "text", + "content": "J. Kim et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 134, + 114, + 253, + 189 + ], + "blocks": [ + { + "bbox": [ + 134, + 114, + 253, + 189 + ], + "lines": [ + { + "bbox": [ + 134, + 114, + 253, + 189 + ], + "spans": [ + { + "bbox": [ + 134, + 114, + 253, + 189 + ], + "type": "image", + "image_path": "6787ce1d22ac50a87da0b61b6d136a34909358897286513db58cda2142795387.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 255, + 114, + 362, + 189 + ], + "blocks": [ + { + "bbox": [ + 255, + 114, + 362, + 189 + ], + "lines": [ + { + "bbox": [ + 255, + 114, + 362, + 189 + ], + "spans": [ + { + "bbox": [ + 255, + 114, + 362, + 189 + ], + "type": "image", + "image_path": "2e43adec9e88f675f3b4688dda717c5d4eff99b4d7bc1907473662c54c3536df.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 363, + 114, + 472, + 189 + ], + "blocks": [ + { + "bbox": [ + 363, + 114, + 472, + 189 + ], + "lines": [ + { + "bbox": [ + 363, + 114, + 472, + 189 + ], + "spans": [ + { + "bbox": [ + 363, + 114, + 472, + 189 + ], + "type": "image", + "image_path": "c1036684973bcba7064adb4c5b623e79ea12b333591a480939af40ccb6781e8c.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 133, + 191, + 253, + 266 + ], + "blocks": [ + { + "bbox": [ + 133, + 191, + 253, + 266 + ], + "lines": [ + { + "bbox": [ + 133, + 191, + 253, + 266 + ], + "spans": [ + { + "bbox": [ + 133, + 191, + 253, + 266 + ], + "type": "image", + "image_path": "e98508730dd76647ca5502e4553fcd0d790de32b16677c8fe73d905b8b36f5ed.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 168, + 266, + 230, + 275 + ], + "lines": [ + { + "bbox": [ + 168, + 266, + 230, + 275 + ], + "spans": [ + { + "bbox": [ + 168, + 266, + 230, + 275 + ], + "type": "text", + "content": "(a) Ground-truth" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 131, + 285, + 482, + 318 + ], + "lines": [ + { + "bbox": [ + 131, + 285, + 482, + 318 + ], + "spans": [ + { + "bbox": [ + 131, + 285, + 482, + 318 + ], + "type": "text", + "content": "Fig. 3: Qualitative comparison with MinkNet42 backbone. Top: Model trained on nuScenes, tested on SemanticKITTI " + }, + { + "bbox": [ + 131, + 285, + 482, + 318 + ], + "type": "inline_equation", + "content": "(\\mathrm{N} \\rightarrow \\mathrm{K})" + }, + { + "bbox": [ + 131, + 285, + 482, + 318 + ], + "type": "text", + "content": ". Bottom: Trained on SemanticKITTI, tested on nuScenes " + }, + { + "bbox": [ + 131, + 285, + 482, + 318 + ], + "type": "inline_equation", + "content": "(\\mathrm{K} \\rightarrow \\mathrm{N})" + }, + { + "bbox": [ + 131, + 285, + 482, + 318 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 254, + 190, + 362, + 266 + ], + "blocks": [ + { + "bbox": [ + 254, + 190, + 362, + 266 + ], + "lines": [ + { + "bbox": [ + 254, + 190, + 362, + 266 + ], + "spans": [ + { + "bbox": [ + 254, + 190, + 362, + 266 + ], + "type": "image", + "image_path": "c81395f531d12d51e25eec38a164efe7e3008364c6c5a09997dc916df4c834d3.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 282, + 266, + 335, + 275 + ], + "lines": [ + { + "bbox": [ + 282, + 266, + 335, + 275 + ], + "spans": [ + { + "bbox": [ + 282, + 266, + 335, + 275 + ], + "type": "text", + "content": "(b) MinkNet42" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 363, + 190, + 472, + 266 + ], + "blocks": [ + { + "bbox": [ + 363, + 190, + 472, + 266 + ], + "lines": [ + { + "bbox": [ + 363, + 190, + 472, + 266 + ], + "spans": [ + { + "bbox": [ + 363, + 190, + 472, + 266 + ], + "type": "image", + "image_path": "225e7d0ce32039fdf0db98059c68904d7c80ebbd05d8ae1e2e99d3696907e1ee.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 380, + 266, + 455, + 275 + ], + "lines": [ + { + "bbox": [ + 380, + 266, + 455, + 275 + ], + "spans": [ + { + "bbox": [ + 380, + 266, + 455, + 275 + ], + "type": "text", + "content": "(c) MinkNet42+Ours" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 342, + 482, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 342, + 482, + 628 + ], + "spans": [ + { + "bbox": [ + 130, + 342, + 482, + 628 + ], + "type": "text", + "content": "Experiments in the LiDomAug Setting We benchmark the proposed method against various augmentation methods [10,22,40,46], domain adaptation methods [18,34,39,44], and a domain generation method [29]. We adhere to the experimental setup established by LiDomAug [29], employing MinkNet42 [7] and C&L [44] as backbone networks, and utilizing SemanticKITTI and nuScenes datasets for evaluation. Unlike the DGLSS [16] setting which uses a voxel size of " + }, + { + "bbox": [ + 130, + 342, + 482, + 628 + ], + "type": "inline_equation", + "content": "20~\\mathrm{cm}" + }, + { + "bbox": [ + 130, + 342, + 482, + 628 + ], + "type": "text", + "content": ", the LiDomAug setting uses a voxel size of " + }, + { + "bbox": [ + 130, + 342, + 482, + 628 + ], + "type": "inline_equation", + "content": "5\\mathrm{cm}" + }, + { + "bbox": [ + 130, + 342, + 482, + 628 + ], + "type": "text", + "content": ". We benchmark the proposed method against various augmentation methods [10,22,40,46], domain adaptation methods [18,34,39,44], and a domain generation method [29]. We adhere to the experimental setup established by LiDomAug [29], employing MinkNet42 [7] and C&L [44] as backbone networks, and utilizing SemanticKITTI and nuScenes datasets for evaluation. The comparative analysis in Tab. 3 shows that our method surpasses all considered augmentation methods in both " + }, + { + "bbox": [ + 130, + 342, + 482, + 628 + ], + "type": "inline_equation", + "content": "(\\mathrm{K}\\rightarrow \\mathrm{N})" + }, + { + "bbox": [ + 130, + 342, + 482, + 628 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 342, + 482, + 628 + ], + "type": "inline_equation", + "content": "(\\mathrm{N}\\rightarrow \\mathrm{K})" + }, + { + "bbox": [ + 130, + 342, + 482, + 628 + ], + "type": "text", + "content": " configurations. Particularly with MinkNet42 as the backbone, our method achieves a " + }, + { + "bbox": [ + 130, + 342, + 482, + 628 + ], + "type": "inline_equation", + "content": "5.9\\%" + }, + { + "bbox": [ + 130, + 342, + 482, + 628 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 342, + 482, + 628 + ], + "type": "inline_equation", + "content": "5.8\\%" + }, + { + "bbox": [ + 130, + 342, + 482, + 628 + ], + "type": "text", + "content": " increase in performance over LiDomAug in the " + }, + { + "bbox": [ + 130, + 342, + 482, + 628 + ], + "type": "inline_equation", + "content": "(\\mathrm{K}\\rightarrow \\mathrm{N})" + }, + { + "bbox": [ + 130, + 342, + 482, + 628 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 342, + 482, + 628 + ], + "type": "inline_equation", + "content": "(\\mathrm{N}\\rightarrow \\mathrm{K})" + }, + { + "bbox": [ + 130, + 342, + 482, + 628 + ], + "type": "text", + "content": " scenarios, respectively. This enhancement is significant, considering LiDomAug's reliance on ego-motion and multi-frame data integration for domain generalization. When utilizing C&L [44] as the backbone, the proposed method significantly improves mIoU by " + }, + { + "bbox": [ + 130, + 342, + 482, + 628 + ], + "type": "inline_equation", + "content": "+34.5\\%" + }, + { + "bbox": [ + 130, + 342, + 482, + 628 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 342, + 482, + 628 + ], + "type": "inline_equation", + "content": "+8.4\\%" + }, + { + "bbox": [ + 130, + 342, + 482, + 628 + ], + "type": "text", + "content": " for the scenario of training on SemanticKITTI and evaluating on nuScenes " + }, + { + "bbox": [ + 130, + 342, + 482, + 628 + ], + "type": "inline_equation", + "content": "(\\mathrm{K}\\rightarrow \\mathrm{N})" + }, + { + "bbox": [ + 130, + 342, + 482, + 628 + ], + "type": "text", + "content": ", compared to C&L and LiDomAug, respectively. Moreover, in the transition from the sparse 32-channel dataset (nuScenes) to the denser 64-channel dataset (SemanticKITTI) " + }, + { + "bbox": [ + 130, + 342, + 482, + 628 + ], + "type": "inline_equation", + "content": "(\\mathrm{N}\\rightarrow \\mathrm{K})" + }, + { + "bbox": [ + 130, + 342, + 482, + 628 + ], + "type": "text", + "content": ", our method shows a " + }, + { + "bbox": [ + 130, + 342, + 482, + 628 + ], + "type": "inline_equation", + "content": "+21.7\\%" + }, + { + "bbox": [ + 130, + 342, + 482, + 628 + ], + "type": "text", + "content": " increase in mIoU over C&L and " + }, + { + "bbox": [ + 130, + 342, + 482, + 628 + ], + "type": "inline_equation", + "content": "+8.2\\%" + }, + { + "bbox": [ + 130, + 342, + 482, + 628 + ], + "type": "text", + "content": " improvement over LiDomAug." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 630, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 630, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 630, + 482, + 665 + ], + "type": "text", + "content": "We further conduct an analysis with a voxel size of " + }, + { + "bbox": [ + 130, + 630, + 482, + 665 + ], + "type": "inline_equation", + "content": "20~\\mathrm{cm}" + }, + { + "bbox": [ + 130, + 630, + 482, + 665 + ], + "type": "text", + "content": ", unlike the comparative methods that all use a voxel size of " + }, + { + "bbox": [ + 130, + 630, + 482, + 665 + ], + "type": "inline_equation", + "content": "5\\mathrm{cm}" + }, + { + "bbox": [ + 130, + 630, + 482, + 665 + ], + "type": "text", + "content": ". Increasing the voxel size from " + }, + { + "bbox": [ + 130, + 630, + 482, + 665 + ], + "type": "inline_equation", + "content": "5\\mathrm{cm}" + }, + { + "bbox": [ + 130, + 630, + 482, + 665 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 130, + 630, + 482, + 665 + ], + "type": "inline_equation", + "content": "20\\mathrm{cm}" + }, + { + "bbox": [ + 130, + 630, + 482, + 665 + ], + "type": "text", + "content": " results in reductions of " + }, + { + "bbox": [ + 130, + 630, + 482, + 665 + ], + "type": "inline_equation", + "content": "30.3\\%" + }, + { + "bbox": [ + 130, + 630, + 482, + 665 + ], + "type": "text", + "content": " in training time and " + }, + { + "bbox": [ + 130, + 630, + 482, + 665 + ], + "type": "inline_equation", + "content": "62.5\\%" + }, + { + "bbox": [ + 130, + 630, + 482, + 665 + ], + "type": "text", + "content": " in infer-" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 272, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 272, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 272, + 91, + 447, + 102 + ], + "type": "text", + "content": "Rethinking LiDAR Domain Generalization" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 209, + 178, + 399, + 249 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 482, + 169 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 482, + 169 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 482, + 169 + ], + "type": "text", + "content": "Table 4: Ablation study on individual components within our method using mIoU. Experiments evaluate the model's performance with and without the following: (a) Point-voxel encoding, (b) Density-aware embedding module, (c) Density clipping, and (d) Density-Augmentation. All experiments were conducted using a voxel size of 20 cm." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 209, + 178, + 399, + 249 + ], + "lines": [ + { + "bbox": [ + 209, + 178, + 399, + 249 + ], + "spans": [ + { + "bbox": [ + 209, + 178, + 399, + 249 + ], + "type": "table", + "html": "
(a)(b)(c)(d)K→NN→K
40.731.4
✓43.0 (+5.7%)35.0 (+11.5%)
✓✓45.7 (+12.3%)40.5 (+29.0%)
✓✓✓46.2 (+13.5%)41.8 (+33.1%)
✓✓✓✓50.1 (+23.1%)46.3 (+47.5%)
", + "image_path": "4194e98979e14ddf5fb84004a2ebcf4656897d1e713063a25c7dbf2813061609.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 268, + 482, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 268, + 482, + 365 + ], + "spans": [ + { + "bbox": [ + 130, + 268, + 482, + 365 + ], + "type": "text", + "content": "ence time, while maintaining comparable performance—with a slight decrease of " + }, + { + "bbox": [ + 130, + 268, + 482, + 365 + ], + "type": "inline_equation", + "content": "3.5\\%" + }, + { + "bbox": [ + 130, + 268, + 482, + 365 + ], + "type": "text", + "content": " on the MinkNet42 backbone and an increase of " + }, + { + "bbox": [ + 130, + 268, + 482, + 365 + ], + "type": "inline_equation", + "content": "4.7\\%" + }, + { + "bbox": [ + 130, + 268, + 482, + 365 + ], + "type": "text", + "content": " on the C&L backbone. Based on these results, we recommend a voxel size of " + }, + { + "bbox": [ + 130, + 268, + 482, + 365 + ], + "type": "inline_equation", + "content": "20~\\mathrm{cm}" + }, + { + "bbox": [ + 130, + 268, + 482, + 365 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 130, + 268, + 482, + 365 + ], + "type": "inline_equation", + "content": "\\mathrm{v} = 20\\mathrm{cm}" + }, + { + "bbox": [ + 130, + 268, + 482, + 365 + ], + "type": "text", + "content": ") as the default setting for our proposed method. The effectiveness is further illustrated in the qualitative results depicted in Fig. 3, highlighting the substantial improvements our method brings to the domain generalization performance of the 3D backbone network. Please refer to the supplementary material for detailed results, including per-class IoU." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 380, + 246, + 391 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 380, + 246, + 391 + ], + "spans": [ + { + "bbox": [ + 132, + 380, + 246, + 391 + ], + "type": "text", + "content": "4.4 Ablation Studies" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 398, + 481, + 458 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 398, + 481, + 458 + ], + "spans": [ + { + "bbox": [ + 130, + 398, + 481, + 458 + ], + "type": "text", + "content": "To validate the impact of individual components within our method, we conduct ablation studies. These studies concentrate on the evaluation of point-voxel encoding, the density-aware embedding module, density soft clipping, and density augmentation, providing insights into the significance of each element in enhancing the overall methodology." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 473, + 481, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 473, + 481, + 557 + ], + "spans": [ + { + "bbox": [ + 130, + 473, + 481, + 557 + ], + "type": "text", + "content": "Analysis of the individual components of our method In our comprehensive ablation study in Tab. 4, we thoroughly investigate the individual components of our method. We use the experimental setup proposed by LiDomAug [29] and employ MinkowskiNet [7] as the backbone. This study dissects the impact of our DDFE, which includes (a) point-voxel encoding, (b) density-aware embedding module, and (c) density clipping. Additionally, it examines (d) the impact of our density augmentation method." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "text", + "content": "The integration of all these elements results in a significant performance uplift " + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "inline_equation", + "content": "+13.5\\%" + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "text", + "content": " in the " + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "inline_equation", + "content": "(\\mathrm{K}\\rightarrow \\mathrm{N})" + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "text", + "content": " scenario and " + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "inline_equation", + "content": "+33.1\\%" + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "text", + "content": " in the " + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "inline_equation", + "content": "(\\mathrm{N}\\rightarrow \\mathrm{K})" + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "text", + "content": " scenario over the baseline model. Notably, the density-aware embedding module stands out for its substantial effect, closely followed by the point-voxel encoding, highlighting its critical role. Furthermore, the study validates the utility of density soft clipping in enhancing model flexibility towards novel density configurations. Meanwhile, the gains observed with density clipping highlight the inherent challenge in adapting to unfamiliar density landscapes, emphasizing the nuanced contributions of each component towards achieving robust domain generalization. Lastly," + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "text", + "content": "J. Kim et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 181, + 157, + 431, + 248 + ], + "blocks": [ + { + "bbox": [ + 131, + 114, + 482, + 149 + ], + "lines": [ + { + "bbox": [ + 131, + 114, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 131, + 114, + 482, + 149 + ], + "type": "text", + "content": "Table 5: Comparison of data augmentation methods in the proposed DDFE. The best and the second best results are highlighted in bold and underline, respectively. 'D. V.' refers to Density Variation situation, and 'B. S.' to Beam Sampling augmentation." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 181, + 157, + 431, + 248 + ], + "lines": [ + { + "bbox": [ + 181, + 157, + 431, + 248 + ], + "spans": [ + { + "bbox": [ + 181, + 157, + 431, + 248 + ], + "type": "table", + "html": "
MethodD. V.K→NN→K
DDFE45.941.8
+ PolarMix [40]48.2 (+5.0%)42.3 (+1.2%)
+ B. S. [38]✓49.1 (+7.0%)42.2 (+0.9%)
+ Mix3D [22]✓47.4 (+3.2%)44.8 (+7.1%)
+ Mix3D + B. S. [38]✓49.1 (+7.0%)44.2 (+5.7%)
+ E-Mix3D✓48.5 (+5.6%)46.4 (+11.0%)
+ E-Mix3D + B. S. [38]✓49.4 (+7.6%)46.5 (+11.2%)
", + "image_path": "276384e1d0fa284ec1809955af3ce82c2d527930ea45ffb90e932d2f711413f5.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 151, + 259, + 294, + 349 + ], + "blocks": [ + { + "bbox": [ + 151, + 259, + 294, + 349 + ], + "lines": [ + { + "bbox": [ + 151, + 259, + 294, + 349 + ], + "spans": [ + { + "bbox": [ + 151, + 259, + 294, + 349 + ], + "type": "image", + "image_path": "aaa37abd30689c928f31cf4afde08f9c61fc9cb7a99125d9da3c9f7ca4249fc1.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 357, + 482, + 413 + ], + "lines": [ + { + "bbox": [ + 131, + 357, + 482, + 413 + ], + "spans": [ + { + "bbox": [ + 131, + 357, + 482, + 413 + ], + "type": "text", + "content": "Fig. 4: Visualization of feature similarity matrices between the Waymo (64-channel) and the nuScenes (32-channel) datasets, with a focus on the distance of objects from the LiDAR. We utilize models trained with the nuScenes dataset for visualization. (a) A point-voxel feature encoding method. (b) A density-aware embedding module. (c) A density soft clipping. Our DDFE combines all modules (a), (b), and (c)." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 298, + 259, + 463, + 348 + ], + "blocks": [ + { + "bbox": [ + 298, + 259, + 463, + 348 + ], + "lines": [ + { + "bbox": [ + 298, + 259, + 463, + 348 + ], + "spans": [ + { + "bbox": [ + 298, + 259, + 463, + 348 + ], + "type": "image", + "image_path": "8b0eedd32fe01fce7a217da4bb30d1e6e6ae5558322b10573c27e363a2c914b3.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 436, + 482, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 436, + 482, + 485 + ], + "spans": [ + { + "bbox": [ + 130, + 436, + 482, + 485 + ], + "type": "text", + "content": "incorporating density augmentation into DDFE leads to further enhancements, with an additional " + }, + { + "bbox": [ + 130, + 436, + 482, + 485 + ], + "type": "inline_equation", + "content": "8.4\\%" + }, + { + "bbox": [ + 130, + 436, + 482, + 485 + ], + "type": "text", + "content": " performance increase in the " + }, + { + "bbox": [ + 130, + 436, + 482, + 485 + ], + "type": "inline_equation", + "content": "(\\mathrm{K} \\rightarrow \\mathrm{N})" + }, + { + "bbox": [ + 130, + 436, + 482, + 485 + ], + "type": "text", + "content": " scenario and " + }, + { + "bbox": [ + 130, + 436, + 482, + 485 + ], + "type": "inline_equation", + "content": "10.8\\%" + }, + { + "bbox": [ + 130, + 436, + 482, + 485 + ], + "type": "text", + "content": " in the " + }, + { + "bbox": [ + 130, + 436, + 482, + 485 + ], + "type": "inline_equation", + "content": "(\\mathrm{N} \\rightarrow \\mathrm{K})" + }, + { + "bbox": [ + 130, + 436, + 482, + 485 + ], + "type": "text", + "content": " scenario. This demonstrates the effectiveness of the proposed density augmentation scheme in significantly boosting the performance of DDFE." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 503, + 482, + 612 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 503, + 482, + 612 + ], + "spans": [ + { + "bbox": [ + 131, + 503, + 482, + 612 + ], + "type": "text", + "content": "Augmentation Tab. 5 shows the performance comparison when various augmentation methods are applied to the proposed DDFE. The experimental setup follows DGLSS [16], and the baseline involves applying the proposed DDFE to MinkowskiNet [7]. The proposed enhanced-Mix3D shows a " + }, + { + "bbox": [ + 131, + 503, + 482, + 612 + ], + "type": "inline_equation", + "content": "2.3\\%" + }, + { + "bbox": [ + 131, + 503, + 482, + 612 + ], + "type": "text", + "content": " performance improvement over the original Mix3D [22] when trained on SemanticKITTI and tested on nuScenes " + }, + { + "bbox": [ + 131, + 503, + 482, + 612 + ], + "type": "inline_equation", + "content": "(\\mathrm{K}\\rightarrow \\mathrm{N})" + }, + { + "bbox": [ + 131, + 503, + 482, + 612 + ], + "type": "text", + "content": ", and a " + }, + { + "bbox": [ + 131, + 503, + 482, + 612 + ], + "type": "inline_equation", + "content": "3.6\\%" + }, + { + "bbox": [ + 131, + 503, + 482, + 612 + ], + "type": "text", + "content": " improvement when trained on nuScenes and tested on SemanticKITTI " + }, + { + "bbox": [ + 131, + 503, + 482, + 612 + ], + "type": "inline_equation", + "content": "(\\mathrm{N}\\rightarrow \\mathrm{K})" + }, + { + "bbox": [ + 131, + 503, + 482, + 612 + ], + "type": "text", + "content": ". In conjunction with the beam sampling method, it achieves a " + }, + { + "bbox": [ + 131, + 503, + 482, + 612 + ], + "type": "inline_equation", + "content": "7.6\\%" + }, + { + "bbox": [ + 131, + 503, + 482, + 612 + ], + "type": "text", + "content": " performance improvement in " + }, + { + "bbox": [ + 131, + 503, + 482, + 612 + ], + "type": "inline_equation", + "content": "(\\mathrm{K}\\rightarrow \\mathrm{N})" + }, + { + "bbox": [ + 131, + 503, + 482, + 612 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 503, + 482, + 612 + ], + "type": "inline_equation", + "content": "11.2\\%" + }, + { + "bbox": [ + 131, + 503, + 482, + 612 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 131, + 503, + 482, + 612 + ], + "type": "inline_equation", + "content": "(\\mathrm{N}\\rightarrow \\mathrm{K})" + }, + { + "bbox": [ + 131, + 503, + 482, + 612 + ], + "type": "text", + "content": " compared to the baseline." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 629, + 482, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 629, + 482, + 667 + ], + "spans": [ + { + "bbox": [ + 131, + 629, + 482, + 667 + ], + "type": "text", + "content": "Analysis of Density Discriminative Feature Embedding (DDFE) We delve into the effectiveness of the DDFE module by examining the feature similarity between the source dataset, nuScenes, and the unseen dataset, Waymo," + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 272, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 272, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 272, + 91, + 447, + 102 + ], + "type": "text", + "content": "Rethinking LiDAR Domain Generalization" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 115, + 482, + 272 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 115, + 482, + 272 + ], + "spans": [ + { + "bbox": [ + 130, + 115, + 482, + 272 + ], + "type": "text", + "content": "as depicted in Fig. 4. Our focus is on the input voxel features " + }, + { + "bbox": [ + 130, + 115, + 482, + 272 + ], + "type": "inline_equation", + "content": "\\hat{F}^v" + }, + { + "bbox": [ + 130, + 115, + 482, + 272 + ], + "type": "text", + "content": " of the 3D backbone model generated by the DDFE. We initiate our analysis by setting a baseline that utilizes a point-voxel feature encoding method for comparison. Following this, we proceed to assess our enhanced method, which integrates a density-aware embedding module along with density soft clipping, to understand its impact on bridging the domain gap. Similarity metrics are derived from the L2 distance between averaged features across specified distance intervals (e.g., 0-5m, ..., 45-50m) for each dataset. According to Fig. 4, the baseline consistently shows higher feature similarity across distances, influenced by its direct embedding of 3D point coordinates. Conversely, our integrated model with density-aware embedding and soft clipping significantly aligns feature similarity across consistent density distributions. It demonstrates the efficacy of the proposed modules in adapting to differences between the source and unseen datasets." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 300, + 482, + 385 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 300, + 482, + 385 + ], + "spans": [ + { + "bbox": [ + 130, + 300, + 482, + 385 + ], + "type": "text", + "content": "Computational Cost The inference time of our method for processing a single frame with the nuScenes is 44ms on an NVIDIA RTX 3090. The inclusion of our DDFE module into the MinkNet42 architecture accounts for an extra 8ms of this computation time, while the base MinkNet42 architecture alone requires 36ms. This integration modestly elevates the model's complexity, introducing approximately 23.8k parameters " + }, + { + "bbox": [ + 130, + 300, + 482, + 385 + ], + "type": "inline_equation", + "content": "(+0.06\\%)" + }, + { + "bbox": [ + 130, + 300, + 482, + 385 + ], + "type": "text", + "content": ". Furthermore, the application of density augmentation during training adds an extra 60ms per scan on the nuScenes." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 414, + 220, + 427 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 414, + 220, + 427 + ], + "spans": [ + { + "bbox": [ + 132, + 414, + 220, + 427 + ], + "type": "text", + "content": "5 Conclusion" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 449, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 449, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 449, + 482, + 666 + ], + "type": "text", + "content": "In this paper, we introduce a new perspective for domain generalization of LiDAR semantic segmentation by exploiting the concept of density diversity within a source domain. Based on this perspective, we propose the Density Discriminative Feature Embedding (DDFE) module. DDFE is designed to incorporate expected densities derived from LiDAR beams into a density-aware feature space, significantly enhancing the model's capability to distinguish between different densities. This novel approach improves the adaptability and accuracy of LiDAR semantic segmentation, enabling them to perform effectively across diverse domains, including those previously unseen. In addition to the core methodology, we also present a simple and effective data augmentation technique that extends the density spectrum of the source data. Extensive experiments on the SemanticKITTI, Waymo, and nuScenes datasets demonstrate the effectiveness of the proposed method in elevating the domain generalization performance of LiDAR semantic segmentation. A limitation of this study is its exclusive focus on semantic segmentation tasks. Thus, in future work, we plan to extend our domain generalization method to encompass other LiDAR-based perception tasks, such as 3D object detection. This extension can involve utilizing object-centric density discriminative features alongside point-wise density embedding." + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "text", + "content": "J. Kim et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 140, + 482, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 140, + 482, + 213 + ], + "spans": [ + { + "bbox": [ + 130, + 140, + 482, + 213 + ], + "type": "text", + "content": "This work was supported by the Digital Innovation Hub project supervised by the Daegu Digital Innovation Promotion Agency(DIP) grant funded by the Korea government(MSIT and Daegu Metropolitan City) in 2023(DBSD1-02, Vision picking system for the logistics industry based on artificial intelligence object recognition) and the National Research Foundation of Korea (NRF) grant funded by the Korea government (MSIT) (No. RS-2023-00210908)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 232, + 199, + 244 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 232, + 199, + 244 + ], + "spans": [ + { + "bbox": [ + 132, + 232, + 199, + 244 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 134, + 258, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 138, + 258, + 480, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 258, + 480, + 300 + ], + "spans": [ + { + "bbox": [ + 138, + 258, + 480, + 300 + ], + "type": "text", + "content": "1. Ando, A., Gidaris, S., Bursuc, A., Puy, G., Boulch, A., Marlet, R.: Rangevit: Towards vision transformers for 3d semantic segmentation in autonomous driving. In: Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 5240-5250 (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 303, + 480, + 345 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 303, + 480, + 345 + ], + "spans": [ + { + "bbox": [ + 138, + 303, + 480, + 345 + ], + "type": "text", + "content": "2. Behley, J., Garbade, M., Milioto, A., Quenzel, J., Behnke, S., Stachniss, C., Gall, J.: Semantickitti: A dataset for semantic scene understanding of lidar sequences. In: Proceedings of IEEE International Conference on Computer Vision (ICCV). pp. 9297-9307 (2019)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 346, + 480, + 389 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 346, + 480, + 389 + ], + "spans": [ + { + "bbox": [ + 138, + 346, + 480, + 389 + ], + "type": "text", + "content": "3. Berman, M., Triki, A.R., Blaschko, M.B.: The lovasz-softmax loss: A tractable surrogate for the optimization of the intersection-over-union measure in neural networks. In: Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 4413-4421 (2018)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 391, + 480, + 422 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 391, + 480, + 422 + ], + "spans": [ + { + "bbox": [ + 138, + 391, + 480, + 422 + ], + "type": "text", + "content": "4. Caesar, H., et al.: nuscenes: A multimodal dataset for autonomous driving. In: Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 11621-11631 (2020)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 423, + 480, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 423, + 480, + 456 + ], + "spans": [ + { + "bbox": [ + 138, + 423, + 480, + 456 + ], + "type": "text", + "content": "5. Chen, T.H., Chang, T.S.: Rangeseg: Range-aware real time segmentation of 3d lidar point clouds. IEEE Transactions on Intelligent Vehicles 7(1), 93-101 (2022). https://doi.org/10.1109/TIV.2021.3085827" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 457, + 480, + 488 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 457, + 480, + 488 + ], + "spans": [ + { + "bbox": [ + 138, + 457, + 480, + 488 + ], + "type": "text", + "content": "6. Cheng, H.X., Han, X.F., Xiao, G.Q.: Transrvnet: Lidar semantic segmentation with transformer. IEEE Transactions on Intelligent Transportation Systems 24(6), 5895-5907 (2023). https://doi.org/10.1109/TITS.2023.3248117" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 490, + 480, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 490, + 480, + 521 + ], + "spans": [ + { + "bbox": [ + 138, + 490, + 480, + 521 + ], + "type": "text", + "content": "7. Choy, C., Gwak, J., Savarese, S.: 4d spatio-temporal convnets: Minkowski convolutional neural networks. In: Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 3075-3084 (2019)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 523, + 480, + 576 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 523, + 480, + 576 + ], + "spans": [ + { + "bbox": [ + 138, + 523, + 480, + 576 + ], + "type": "text", + "content": "8. Cortinhal, T., Tzelepis, G., Erdal Aksoy, E.: Salsanext: Fast, uncertainty-aware semantic segmentation of lidar point clouds. In: Bebis, G., Yin, Z., Kim, E., Bender, J., Subr, K., Kwon, B.C., Zhao, J., Kalkofen, D., Baciu, G. (eds.) Advances in Visual Computing. pp. 207-222. Springer International Publishing, Cham (2020). https://doi.org/10.1007/978-3-030-64559-5_16" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 578, + 480, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 578, + 480, + 620 + ], + "spans": [ + { + "bbox": [ + 138, + 578, + 480, + 620 + ], + "type": "text", + "content": "9. Fan, S., Dong, Q., Zhu, F., Lv, Y., Ye, P., Wang, F.Y.: Scf-net: Learning spatial contextual features for large-scale point cloud segmentation. In: Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 14504-14513 (2021)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 134, + 621, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 621, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 134, + 621, + 480, + 665 + ], + "type": "text", + "content": "10. Ghiasi, G., Cui, Y., Srinivas, A.J., Qian, R., Lin, T.Y., Cubuk, E.D., Le, Q.V., Zoph, B.: Simple copy-paste is a strong data augmentation method for instance segmentation. In: Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 2918-2928 (2021)" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 272, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 272, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 272, + 91, + 447, + 102 + ], + "type": "text", + "content": "Rethinking LiDAR Domain Generalization" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 117, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 133, + 117, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 117, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 117, + 480, + 149 + ], + "type": "text", + "content": "11. Graham, B., Engelcke, M., Van Der Maaten, L.: 3d semantic segmentation with submanifold sparse convolutional networks. In: Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 9224-9232 (2018)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 151, + 480, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 151, + 480, + 172 + ], + "spans": [ + { + "bbox": [ + 133, + 151, + 480, + 172 + ], + "type": "text", + "content": "12. Graham, B., Van der Maaten, L.: Submanifold sparse convolutional networks. arXiv preprint arXiv:1706.01307 (2017)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 174, + 480, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 174, + 480, + 206 + ], + "spans": [ + { + "bbox": [ + 133, + 174, + 480, + 206 + ], + "type": "text", + "content": "13. Hou, Y., Zhu, X., Ma, Y., Loy, C.C., Li, Y.: Point-to-voxel knowledge distillation for lidar semantic segmentation. In: Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 8479-8488 (2022)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 134, + 208, + 480, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 208, + 480, + 239 + ], + "spans": [ + { + "bbox": [ + 134, + 208, + 480, + 239 + ], + "type": "text", + "content": "14. Hu, Q., Liu, D., Hu, W.: Density-insensitive unsupervised domain adaption on 3d object detection. In: Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 17556-17566 (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 134, + 241, + 480, + 284 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 241, + 480, + 284 + ], + "spans": [ + { + "bbox": [ + 134, + 241, + 480, + 284 + ], + "type": "text", + "content": "15. Hu, Q., Yang, B., Xie, L., Rosa, S., Guo, Y., Wang, Z., Trigoni, N., Markham, A.: Randla-net: Efficient semantic segmentation of large-scale point clouds. In: Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 11108-11117 (2020)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 134, + 285, + 480, + 318 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 285, + 480, + 318 + ], + "spans": [ + { + "bbox": [ + 134, + 285, + 480, + 318 + ], + "type": "text", + "content": "16. Kim, H., Kang, Y., Oh, C., Yoon, K.J.: Single domain generalization for lidar semantic segmentation. In: Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 17587-17598 (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 134, + 319, + 480, + 362 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 319, + 480, + 362 + ], + "spans": [ + { + "bbox": [ + 134, + 319, + 480, + 362 + ], + "type": "text", + "content": "17. Kong, L., Quader, N., Liong, V.E.: Conda: Unsupervised domain adaptation for lidar segmentation via regularized domain concatenation. In: Proceedings of IEEE International Conference on Robotics and Automation (ICRA). pp. 9338-9345. IEEE (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 134, + 364, + 480, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 364, + 480, + 396 + ], + "spans": [ + { + "bbox": [ + 134, + 364, + 480, + 396 + ], + "type": "text", + "content": "18. Lee, C.Y., Batra, T., Baig, M.H., Ulbricht, D.: Sliced wasserstein discrepancy for unsupervised domain adaptation. In: Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 10285-10295 (2019)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 134, + 397, + 480, + 452 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 397, + 480, + 452 + ], + "spans": [ + { + "bbox": [ + 134, + 397, + 480, + 452 + ], + "type": "text", + "content": "19. Lehner, A., Gasperini, S., Marcos-Ramiro, A., Schmidt, M., Mahani, M.A.N., Navab, N., Busam, B., Tombari, F.: 3d-vfield: Adversarial augmentation of point clouds for domain generalization in 3d object detection. In: Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 17295-17304 (2022)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 134, + 453, + 480, + 485 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 453, + 480, + 485 + ], + "spans": [ + { + "bbox": [ + 134, + 453, + 480, + 485 + ], + "type": "text", + "content": "20. Li, D., Yang, Y., Song, Y.Z., Hospedales, T.: Learning to generalize: Meta-learning for domain generalization. In: Proceedings of the AAAI Conference on Artificial Intelligence (AAAI). vol. 32 (2018)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 134, + 487, + 480, + 529 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 487, + 480, + 529 + ], + "spans": [ + { + "bbox": [ + 134, + 487, + 480, + 529 + ], + "type": "text", + "content": "21. Li, M., Zhang, Y., Ma, X., Qu, Y., Fu, Y.: Bev-dg: Cross-modal learning under bird's-eye view for domain generalization of 3d semantic segmentation. In: Proceedings of IEEE International Conference on Computer Vision (ICCV). pp. 11632-11642 (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 134, + 532, + 480, + 563 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 532, + 480, + 563 + ], + "spans": [ + { + "bbox": [ + 134, + 532, + 480, + 563 + ], + "type": "text", + "content": "22. Nekrasov, A., Schult, J., Litany, O., Leibe, B., Engelmann, F.: Mix3d: Out-of-context data augmentation for 3d scenes. In: Proceedings of International Conference on 3D Vision (3DV). pp. 116-125 (2021)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 134, + 565, + 480, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 565, + 480, + 597 + ], + "spans": [ + { + "bbox": [ + 134, + 565, + 480, + 597 + ], + "type": "text", + "content": "23. Pan, X., Luo, P., Shi, J., Tang, X.: Two at once: Enhancing learning and generalization capacities via ibn-net. In: Proceedings of European Conference on Computer Vision (ECCV). pp. 464-479 (2018)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 134, + 599, + 480, + 640 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 599, + 480, + 640 + ], + "spans": [ + { + "bbox": [ + 134, + 599, + 480, + 640 + ], + "type": "text", + "content": "24. Pan, Y., Gao, B., Mei, J., Geng, S., Li, C., Zhao, H.: Semanticposs: A point cloud dataset with large quantity of dynamic instances. In: 2020 IEEE Intelligent Vehicles Symposium (IV). pp. 687-693 (2020). https://doi.org/10.1109/IV47402.2020.9304596" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 134, + 643, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 643, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 134, + 643, + 480, + 665 + ], + "type": "text", + "content": "25. Peng, K., Fei, J., Yang, K., Roitberg, A., Zhang, J., Bieder, F., Heidenreich, P., Stiller, C., Stiefelhagen, R.: Mass: Multi-attentional semantic segmentation of lidar" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "text", + "content": "J. Kim et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 117, + 480, + 666 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 149, + 117, + 480, + 148 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 117, + 480, + 148 + ], + "spans": [ + { + "bbox": [ + 149, + 117, + 480, + 148 + ], + "type": "text", + "content": "data for dense top-view understanding. IEEE Transactions on Intelligent Transportation Systems 23(9), 15824-15840 (2022). https://doi.org/10.1109/TITS.2022.3145588" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 150, + 480, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 150, + 480, + 183 + ], + "spans": [ + { + "bbox": [ + 133, + 150, + 480, + 183 + ], + "type": "text", + "content": "26. Peng, X., Zhu, X., Ma, Y.: Cl3d: Unsupervised domain adaptation for cross-lidar 3d detection. In: Proceedings of the AAAI Conference on Artificial Intelligence (AAAI). vol. 37, pp. 2047-2055 (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 184, + 480, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 184, + 480, + 216 + ], + "spans": [ + { + "bbox": [ + 133, + 184, + 480, + 216 + ], + "type": "text", + "content": "27. Qi, C.R., Su, H., Mo, K., Guibas, L.J.: Pointnet: Deep learning on point sets for 3d classification and segmentation. In: Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 652-660 (2017)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 216, + 480, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 216, + 480, + 258 + ], + "spans": [ + { + "bbox": [ + 133, + 216, + 480, + 258 + ], + "type": "text", + "content": "28. Rochan, M., Aich, S., Corral-Soto, E.R., Nabatchian, A., Liu, B.: Unsupervised domain adaptation in lidar semantic segmentation with self-supervision and gated adapters. In: Proceedings of IEEE International Conference on Robotics and Automation (ICRA). pp. 2649-2655 (2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 133, + 259, + 480, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 259, + 480, + 293 + ], + "spans": [ + { + "bbox": [ + 133, + 259, + 480, + 293 + ], + "type": "text", + "content": "29. Ryu, K., Hwang, S., Park, J.: Instant domain augmentation for lidar semantic segmentation. In: Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 9350-9360 (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 293, + 480, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 293, + 480, + 335 + ], + "spans": [ + { + "bbox": [ + 132, + 293, + 480, + 335 + ], + "type": "text", + "content": "30. Saltori, C., Galasso, F., Fiameni, G., Sebe, N., Ricci, E., Poiesi, F.: Cosmix: Compositional semantic mix for domain adaptation in 3d lidar segmentation. In: Proceedings of European Conference on Computer Vision (ECCV). pp. 586-602. Springer (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 336, + 480, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 336, + 480, + 369 + ], + "spans": [ + { + "bbox": [ + 132, + 336, + 480, + 369 + ], + "type": "text", + "content": "31. Saltori, C., Lathuilière, S., Sebe, N., Ricci, E., Galasso, F.: Sf-uda 3d: Source-free unsupervised domain adaptation for lidar-based 3d object detection. In: Proceedings of International Conference on 3D Vision (3DV). pp. 771–780 (2020)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 370, + 480, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 370, + 480, + 403 + ], + "spans": [ + { + "bbox": [ + 132, + 370, + 480, + 403 + ], + "type": "text", + "content": "32. Saltori, C., Osep, A., Ricci, E., Leal-Taixe, L.: Walking your lidog: A journey through multiple domains for lidar semantic segmentation. In: Proceedings of IEEE International Conference on Computer Vision (ICCV). pp. 196-206 (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 403, + 480, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 403, + 480, + 435 + ], + "spans": [ + { + "bbox": [ + 132, + 403, + 480, + 435 + ], + "type": "text", + "content": "33. Sanchez, J., Deschaud, J.E., Goulette, F.: Domain generalization of 3d semantic segmentation in autonomous driving. In: Proceedings of IEEE International Conference on Computer Vision (ICCV). pp. 18077-18087 (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 435, + 480, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 435, + 480, + 468 + ], + "spans": [ + { + "bbox": [ + 132, + 435, + 480, + 468 + ], + "type": "text", + "content": "34. Shaban, A., Lee, J., Jung, S., Meng, X., Boots, B.: Lidar-uda: Self-ensembling through time for unsupervised lidar domain adaptation. In: Proceedings of IEEE International Conference on Computer Vision (ICCV). pp. 19784-19794 (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 468, + 480, + 501 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 468, + 480, + 501 + ], + "spans": [ + { + "bbox": [ + 132, + 468, + 480, + 501 + ], + "type": "text", + "content": "35. Sun, P., et al.: Scalability in perception for autonomous driving: Waymo open dataset. In: Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 2446-2454 (2020)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 501, + 480, + 534 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 501, + 480, + 534 + ], + "spans": [ + { + "bbox": [ + 132, + 501, + 480, + 534 + ], + "type": "text", + "content": "36. Thomas, H., Qi, C.R., Deschaud, J.E., Marcotegui, B., Goulette, F., Guibas, L.J.: Kpconv: Flexible and deformable convolution for point clouds. In: Proceedings of IEEE International Conference on Computer Vision (ICCV). pp. 6411-6420 (2019)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 534, + 480, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 534, + 480, + 555 + ], + "spans": [ + { + "bbox": [ + 132, + 534, + 480, + 555 + ], + "type": "text", + "content": "37. Vitter, J.S.: Random sampling with a reservoir. ACM Trans. Math. Softw. 11(1), 37-57 (mar 1985). https://doi.org/10.1145/3147.3165" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 555, + 480, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 555, + 480, + 588 + ], + "spans": [ + { + "bbox": [ + 132, + 555, + 480, + 588 + ], + "type": "text", + "content": "38. Wei, Y., Wei, Z., Rao, Y., Li, J., Zhou, J., Lu, J.: Lidar distillation: Bridging the beam-induced domain gap for 3d object detection. In: Proceedings of European Conference on Computer Vision (ECCV). pp. 179-195. Springer (2022)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 589, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 589, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 589, + 480, + 632 + ], + "type": "text", + "content": "39. Wu, B., Zhou, X., Zhao, S., Yue, X., Keutzer, K.: Squeezesegv2: Improved model structure and unsupervised domain adaptation for road-object segmentation from a lidar point cloud. In: Proceedings of IEEE International Conference on Robotics and Automation (ICRA). pp. 4376-4382 (2019)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 632, + 480, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 480, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 480, + 666 + ], + "type": "text", + "content": "40. Xiao, A., Huang, J., Guan, D., Cui, K., Lu, S., Shao, L.: Polarmix: A general data augmentation technique for lidar point clouds. In: Advances in Neural Information Processing Systems (NeurIPS). pp. 11035-11048 (2022)" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 272, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 272, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 272, + 91, + 447, + 101 + ], + "type": "text", + "content": "Rethinking LiDAR Domain Generalization" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 435 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 481, + 160 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 481, + 160 + ], + "type": "text", + "content": "41. Xiao, A., Huang, J., Xuan, W., Ren, R., Liu, K., Guan, D., El Saddik, A., Lu, S., Xing, E.P.: 3d semantic segmentation in the wild: Learning generalized models for adverse-condition point clouds. In: Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 9382-9392 (2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 161, + 481, + 203 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 161, + 481, + 203 + ], + "spans": [ + { + "bbox": [ + 133, + 161, + 481, + 203 + ], + "type": "text", + "content": "42. Xiao, P., Shao, Z., Hao, S., Zhang, Z., Chai, X., Jiao, J., Li, Z., Wu, J., Sun, K., Jiang, K., Wang, Y., Yang, D.: Pandaset: Advanced sensor suite dataset for autonomous driving. In: ITSC. pp. 3095-3101 (2021). https://doi.org/10.1109/ITSC48978.2021.9565009" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 205, + 481, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 205, + 481, + 237 + ], + "spans": [ + { + "bbox": [ + 132, + 205, + 481, + 237 + ], + "type": "text", + "content": "43. Yang, J., Shi, S., Wang, Z., Li, H., Qi, X.: St3d: Self-training for unsupervised domain adaptation on 3d object detection. In: Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 10368-10378 (2021)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 237, + 481, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 237, + 481, + 270 + ], + "spans": [ + { + "bbox": [ + 132, + 237, + 481, + 270 + ], + "type": "text", + "content": "44. Yi, L., Gong, B., Funkhouser, T.: Complete & label: A domain adaptation approach to semantic segmentation of lidar point clouds. In: Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 15363-15373 (2021)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 270, + 481, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 270, + 481, + 314 + ], + "spans": [ + { + "bbox": [ + 132, + 270, + 481, + 314 + ], + "type": "text", + "content": "45. Yuan, Z., Wen, C., Cheng, M., Su, Y., Liu, W., Yu, S., Wang, C.: Category-level adversaries for outdoor lidar point clouds cross-domain semantic segmentation. IEEE Transactions on Intelligent Transportation Systems 24(2), 1982-1993 (2023). https://doi.org/10.1109/TITS.2022.3219853" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 314, + 481, + 347 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 314, + 481, + 347 + ], + "spans": [ + { + "bbox": [ + 132, + 314, + 481, + 347 + ], + "type": "text", + "content": "46. Yun, S., Han, D., Oh, S.J., Chun, S., Choe, J., Yoo, Y.: Cutmix: Regularization strategy to train strong classifiers with localizable features. In: Proceedings of IEEE International Conference on Computer Vision (ICCV). pp. 6023-6032 (2019)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 347, + 481, + 390 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 347, + 481, + 390 + ], + "spans": [ + { + "bbox": [ + 132, + 347, + 481, + 390 + ], + "type": "text", + "content": "47. Zhang, Y., Zhou, Z., David, P., Yue, X., Xi, Z., Gong, B., Foroosh, H.: Polarnet: An improved grid representation for online lidar point clouds semantic segmentation. In: Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 9601-9610 (2020)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 390, + 481, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 390, + 481, + 435 + ], + "spans": [ + { + "bbox": [ + 132, + 390, + 481, + 435 + ], + "type": "text", + "content": "48. Zhu, X., Zhou, H., Wang, T., Hong, F., Ma, Y., Li, W., Li, H., Lin, D.: Cylindrical and asymmetrical 3d convolution networks for lidar segmentation. In: Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 9939-9948 (2021)" + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "text", + "content": "J. Kim et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Rethinking Normalization Layers for Domain Generalizable Person Re-identification/02192c3d-6e2f-475e-b7a0-eaf65cc6bd4c_content_list.json b/2024/Rethinking Normalization Layers for Domain Generalizable Person Re-identification/02192c3d-6e2f-475e-b7a0-eaf65cc6bd4c_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..ef36223720f5290d990ac6f9ae53ef5d7032e21e --- /dev/null +++ b/2024/Rethinking Normalization Layers for Domain Generalizable Person Re-identification/02192c3d-6e2f-475e-b7a0-eaf65cc6bd4c_content_list.json @@ -0,0 +1,1688 @@ +[ + { + "type": "text", + "text": "Rethinking Normalization Layers for Domain Generalizable Person Re-identification", + "text_level": 1, + "bbox": [ + 243, + 140, + 759, + 184 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Ren Nie $^{1}$ , Jin Ding $^{1}$ , Xue Zhou $^{2,1*}$ , and Xi Li $^{3}$", + "bbox": [ + 330, + 212, + 671, + 228 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 University of Electronic Science and Technology of China, Chengdu, China \nnieren@std.uestc.edu.cn, dingjin1998@outlook.com, zhouxue@uestc.edu.cn", + "bbox": [ + 228, + 239, + 772, + 267 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "2 Shenzhen Institute for Advanced Study, UESTC", + "bbox": [ + 333, + 267, + 669, + 281 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3 College of Computer Science & Technology, Zhejiang University xilizju@zju.edu.cn", + "bbox": [ + 281, + 281, + 720, + 310 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Domain Generalizable Person Re-Identification (DG-ReID) strives to transfer learned feature representation from source domains to unseen target domains, despite significant distribution shifts. While most existing methods enhance model generalization and discriminative feature extraction capability by introducing Instance Normalization (IN) in combination with Batch Normalization (BN), these approaches still struggle with the overfitting of normalization layers to the source domains, posing challenges in domain generalization. To address this issue, we propose ReNorm, a purely normalization-based framework that integrates two complementary normalization layers through two forward propagations for the same weight matrix. In the first forward propagation, Remix Normalization (RN) combines IN and BN in a concise manner to ensure the feature extraction capability. As an effective complement to RN, Emulation Normalization (EN) simulates the testing process of RN, implicitly mitigating the domain shifts caused by the absence of target domain information and actively guiding the model in learning how to generalize the feature extraction capability to unseen target domains. Meanwhile, we propose Domain Frozen (DF), freezing updates to affine parameters to reduce the impact of these less robust parameters on overfitting to the source domains. Extensive experiments show that our framework achieves state-of-the-art performance on the popular benchmarks. The code is available at https://github.com/3699nr/ReNorm.", + "bbox": [ + 259, + 353, + 743, + 672 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Normalization $\\cdot$ Domain generalization $\\cdot$ Re-identification", + "bbox": [ + 261, + 686, + 732, + 699 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 729, + 375, + 744 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Person Re-identification (ReID) is a practical task that has been widely applied in video surveillance and security. It aims at retrieving identical persons across a multi-camera network with non-overlapping fields of view. While current approaches [6,10,12,28,35,43,49] have achieved impressive performance within the same source domains during both training and testing phases. However, when", + "bbox": [ + 212, + 763, + 787, + 840 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/986784d482117644a269c053cf0f6f4d1012444bd7c3ed8725ea9786404f98a6.jpg", + "image_caption": [ + "Fig. 1: Differences between existing normalization-based DG-ReID methods and ReNorm. To illustrate more clearly, we use the example of training with three source domains. (a) Existing normalization-based methods combine IN and BN in a specific way to ensure both the generalization and discrimination capability of the weight matrix. (b) ReNorm integrates two complementary normalization layers through two forward propagations for the same weight matrix. RN enables the weight matrix to extract discriminative features within the source domains, while EN is employed to actively guide the weight matrix to generalize the feature extraction capability to unseen target domains. (c) Some details of ReNorm. Detailed explanations can be found in Sec 3." + ], + "image_footnote": [], + "bbox": [ + 217, + 146, + 785, + 358 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "directly deploying these well-trained models on an unseen target domain, an issue known as \"domain shift\" [39] arises, significantly impacting the performance of the models. To alleviate this problem, recent efforts have been devoted to two approaches: unsupervised domain adaptive (DA) ReID [5,17,20,26], and domain generalizable (DG) ReID [2,3,19,37,44,45]. In contrast to DA ReID, DG ReID is more practical and challenging as it trains on the multiple source domains and then directly transfers the learned representation to the unseen target domain without any target data for fine-tuning.", + "bbox": [ + 212, + 532, + 787, + 654 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Among numerous DG ReID methods, normalization-based approaches [2,3,8, 19,37] stand out for their simplicity and versatility, often integrating seamlessly with other advanced techniques such as meta-learning, distribution alignment, and ensemble learning. Illustrated in Fig.1(a), the combination of Batch Normalization (BN) [16] for capturing discriminative features and Instance Normalization (IN) [38] for eliminating domain-specific information facilitates normalization-based DG-ReID methods in achieving both high generalization and discrimination capabilities. During training, the statistical parameters of both IN and BN within the normalization layers are derived from the same data, ensuring a consistent distribution. However, this consistency is disrupted when the model is deployed on the target domain, as IN statistical parameters originate from the target domain itself, while BN statistical parameters are derived from population", + "bbox": [ + 212, + 657, + 792, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "R. Nie and J. Ding et al.", + "bbox": [ + 271, + 114, + 437, + 128 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "statistics accumulated from source domains during training. The distribution of the target domain significantly differs from that of the source domains, resulting in a marked decline in the model's performance. Furthermore, the vulnerable and sensitive learnable affine parameters are also prone to overfitting to the source domains, leading to performance degeneration in both training and testing phase.", + "bbox": [ + 212, + 146, + 784, + 222 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Moreover, emerging researches [24, 40] have shown that the identity and domain information are separately embedded in the weight matrix (learnable parameters in convolutional layers) and normalization layers. Consequently, we rethink the key challenges of DG ReID, which are rooted in normalization layers. We believe that the capability of the weight matrix to extract discriminative features sets the upper limit of the model's performance. Simultaneously, the normalization layers, responsible for fitting and scaling the input from target domains, establish the lower limit for the model's generalization performance.", + "bbox": [ + 212, + 223, + 785, + 343 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To address the issue of normalization layers being prone to overfitting to the source domains, we restructure the normalization layers, proposing ReNorm, a concise yet effective DG ReID framework entirely based on the design of normalization layers, as shown in Fig.1(b). In each iteration, the network undergoes two forward propagations and then one backpropagation after aggregating two forward losses. The two forward propagations apply different processing to the normalization layers, utilizing two complementary normalization layers to facilitate the weight matrix in extracting discriminative features from familiar source domains, while simultaneously guiding the same weight matrix to learn how to transfer this acquired feature extraction capability to unfamiliar target domains.", + "bbox": [ + 212, + 344, + 787, + 494 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Specifically, in our ReNorm framework, the two forward propagations correspond to two different but complementary normalization strategies: Remix Normalization (RN) and Emulation Normalization (EN). In the first forward propagation, RN introduces Instance Normalization (IN) in combination with Batch Normalization (BN) by directly summing their statistical parameters for standardization. Meanwhile, the BN statistical parameters of the current source domain are stored in the corresponding buffer, as shown in Fig.1(c). In the second forward propagation, EN adopts a similar summing approach but emulates the inference process of RN. It treats the current input as a simulated target domain, and calculates the IN statistical parameters accordingly. Simultaneously, the required BN statistical parameters are obtained from a remaining source domain's buffer stored in RN. This pairwise normalization strategy enables the network to actively learn how to transfer knowledge from source domains to the target domain with a distinct distribution. Additionally, we introduce Domain Frozen (DF), where the vulnerable settings of learnable affine parameters within both normalization layers are frozen and exempt from updates. By doing so, ReNorm greatly raises the lower limit of the model's generalization ability.", + "bbox": [ + 212, + 496, + 787, + 751 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We summarize our contributions as follows:", + "bbox": [ + 240, + 752, + 553, + 766 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "- We introduce ReNorm, a simple yet effective framework for DG ReID, entirely and solely based on the design of normalization layers. This framework significantly mitigates performance degradation resulting from distribution shifts between source and target domains in a concise yet effective way.", + "bbox": [ + 225, + 779, + 785, + 839 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "Rethinking Normalization Layers for DG-ReID", + "bbox": [ + 413, + 114, + 727, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "- We introduce two complementary normalization strategies (i.e., Remix Normalization and Emulation Normalization) equipped with Domain Frozen, actively enabling the network to learn how to transfer the feature extraction capability acquired from source domains to unseen target domains.", + "bbox": [ + 223, + 146, + 782, + 205 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "- Extensive experiments demonstrate the simplicity yet state-of-the-art performance of our ReNorm framework under multiple testing protocols.", + "bbox": [ + 225, + 205, + 784, + 236 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2 Generalizable Person Re-Identification.", + "text_level": 1, + "bbox": [ + 215, + 253, + 630, + 271 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Deep learning-based Person ReID approaches [6,10,12,28,35,43,49] have made great progress in recent years, achieving impressive accuracy in Person ReID. However, when these well-trained models are applied to unseen target domains that have distribution shifts with source domains, the performance decreases significantly. To alleviate this problem, recent efforts are devoted to two approaches: unsupervised domain adaptive (DA) ReID [5,17,20,26] and domain generalizable (DG) ReID. Diverging from the DA ReID, which needs unlabeled target domain datasets for unsupervised fine-tuning. DG ReID trains on the source domains and then directly tests on the unseen target domain. As a result, DG ReID has attracted attention for its robustness and generalization.", + "bbox": [ + 212, + 281, + 787, + 431 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Existing DG ReID methods can be briefly divided into four categories:", + "bbox": [ + 238, + 431, + 750, + 446 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Normalization strategies [2, 33, 34]. Batch Normalization (BN) can preserve diversity among samples, capturing more discriminative features and accelerating convergence. However, it tends to be prone to overfitting to source domains. Instance Normalization (IN) can eliminate domain-specific information to enhance generalization but inevitably removes discriminative information. Thus, normalization-based methods typically combine IN and BN in various ways to ensure both high generalization and discrimination capability of the network. Jiao et al. [18] proposed Dynamically Transformed Instance Normalization (DTIN) to alleviate the drawback of IN by transforming normalized features into appropriate representations in a learnable and adaptive manner. Chen et al. [2] designed a Cluster-Instance Normalization (CINorm) to eliminate the effect of outliers in original BN that causes severe overfitting.", + "bbox": [ + 212, + 446, + 787, + 628 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Domain alignment [19,32]. These methods try to minimize the difference among source domains for learning domain-invariant representations. They assume that the invariant features from the source domains should also be robust to any unseen target domains. Jin et al. [19] proposed a dual causality loss to align the outputs and a Style Normalization and Restitution (SNR) module to distill identity-relevant features via a restitution module for IN compensation.", + "bbox": [ + 212, + 628, + 787, + 718 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Meta-learning [3,37,46]. These methods improve generalized performance through virtual simulations of real train-test domain shifts. Following the concept of 'learning to lean', Choi et al. [3]proposed MetaBIN which utilized a meta-learning training strategy with sets of learnable balancing parameters to integrate BN and IN.", + "bbox": [ + 212, + 719, + 787, + 794 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Ensemble learning [4,25,44]. These methods exploit the relationship among multiple source domains by using diverse experts to extract domain-specific features from the corresponding domain. To avoid the large model size, Xu et al. [44]", + "bbox": [ + 212, + 794, + 787, + 842 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 230, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "R. Nie and J. Ding et al.", + "bbox": [ + 271, + 114, + 437, + 128 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "proposed a lightweight model META, which leverages IN and introduces it into a global branch to pursue invariant features across domains.", + "bbox": [ + 212, + 146, + 782, + 176 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "By introducing IN in combination with BN, normalization-based methods are simple and effective and can be combined with other techniques. However, when tested on unknown target domains, the distribution shifts between source and target domains significantly affect the fitting and scaling ability of normalization layers. Thus, leading to a decline in the model's generalization performance.", + "bbox": [ + 212, + 176, + 787, + 253 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3 Methodology", + "text_level": 1, + "bbox": [ + 214, + 277, + 380, + 295 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Inspired by the observation that the weight matrix tends to encapsulate identity information while domain-specific knowledge is embedded within the normalization layers [24, 40], we propose a novel ReNorm framework for tackling the domain shifts problem for DG-ReID which is rooted in normalization layers. As shown in Fig.2, ReNorm integrates two complementary normalization layers through two forward propagations for the same weight matrix. In each iteration, we sample a mini-batch from a source domain. In the first forward propagation, Remix Normalization (RN) combines IN and BN to ensure the discriminative feature extraction capability of the weight matrix. At the same time, the BN statistical parameters of the current source domain are stored in the corresponding buffer by Exponential Moving Average (EMA). During the second forward propagation, we treat the current input as a simulated target domain of the remaining source domains. Thus, Emulation Normalization (EN) can emulate the inference process of RN to guide the weight matrix to learn how to transfer the acquired feature extraction capability to an unseen domain. Meanwhile, we introduce Domain Frozen (DF) to stabilize the weight matrix by freezing the updates of affine parameters in normalization layers.", + "bbox": [ + 212, + 311, + 787, + 568 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1 Preliminary", + "text_level": 1, + "bbox": [ + 214, + 590, + 361, + 606 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In the DG ReID task, we have access to $K$ source domains (datasets) for training. We denote the $k$ -th source domain as $D_{k}, k \\in 1,2,\\dots,K$ , which have completely non-overlapping label spaces. Given a feature map $X_{k} \\in \\mathbb{R}^{N \\times C \\times H \\times W}$ of a minibatch extracted from a certain domain $D_{k}$ , where $N,C,H$ , and $W$ denote the batch size, the number of channels, the height, and the width, respectively. Following the definition from [41], we decompose the normalization layers into two steps as follows:", + "bbox": [ + 212, + 618, + 787, + 722 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {X} _ {k} = \\frac {X _ {k} - \\mu_ {k}}{\\sqrt {\\sigma_ {k} ^ {2} + \\epsilon}}, \\quad \\tilde {X} _ {k} = \\gamma \\hat {X} _ {k} + \\beta , \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 382, + 723, + 785, + 757 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The first step involves standardization using the mean $\\mu_{k}$ and variance $\\sigma_k^2$ calculated from the input feature map. After that, channel-wise learnable affine parameters $\\gamma$ and $\\beta$ are defined for scaling and shifting during end-to-end training. $\\epsilon >0$ is a constant to avoid numerical problems. The main distinction among various normalization techniques primarily lies in the first step.", + "bbox": [ + 212, + 763, + 787, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "Rethinking Normalization Layers for DG-ReID", + "bbox": [ + 411, + 114, + 727, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/05a152d63138a59f14f1ae9a7fc5ed43aaa261a9becef2a98bde9879893dfb50.jpg", + "image_caption": [ + "Fig. 2: The intra-layer structure of our ReNorm framework. The upper part corresponds to the first forward propagation with Remix Normalization, while the lower part corresponds to the second forward propagation with Emulation Normalization. After standardization, we employ Domain Frozen strategy for the affine parameters. Best viewed in color, detailed explanations can be found in Sec 3.2 and Sec 3.3." + ], + "image_footnote": [], + "bbox": [ + 218, + 145, + 784, + 465 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Instance Normalization [38] can remove style information from input features. So it has been widely used for the DG-ReID task to extract domain-invariant representations by removing their statistical contrast across feature channels. IN calculates mean $\\mu_k^{in} \\in \\mathbb{R}^{N \\times C}$ and variance $\\sigma_k^{in^2} \\in \\mathbb{R}^{N \\times C}$ with respect to each sample and each channel for standardization:", + "bbox": [ + 212, + 574, + 787, + 652 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mu_ {k} ^ {i n} = \\frac {\\sum_ {h , w} ^ {H W} X _ {k}}{H W} \\quad \\text {a n d} \\quad \\sigma_ {k} ^ {i n 2} = \\frac {\\sum_ {h , w} ^ {H W} \\left(X _ {k} - \\mu_ {k} ^ {i n}\\right) ^ {2}}{H W}. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 305, + 657, + 785, + 694 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Batch Normalization [16] is widely used in convolutional neural networks to capture discriminative features and mitigate gradient vanishing or exploding. It calculates mean $\\mu_k^{bn} \\in \\mathbb{R}^C$ and variance $\\sigma_k^{bn^2} \\in \\mathbb{R}^C$ with respect each channel in a mini-batch:", + "bbox": [ + 212, + 699, + 787, + 760 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mu_ {k} ^ {b n} = \\frac {\\sum_ {n , h , w} ^ {N H W} X _ {k}}{N H W} \\quad \\text {a n d} \\quad \\sigma_ {k} ^ {b n ^ {2}} = \\frac {\\sum_ {n , h , w} ^ {N H W} \\left(X _ {k} - \\mu_ {k} ^ {b n}\\right) ^ {2}}{N H W}. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 292, + 765, + 785, + 801 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "It is worth noting that different from IN where the mean $\\mu^{in}$ and variance $\\sigma^{in^2}$ are computed from the input feature map $X$ during both training and test-", + "bbox": [ + 212, + 808, + 787, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "R. Nie and J. Ding et al.", + "bbox": [ + 271, + 114, + 436, + 128 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "ing. During testing, BN standardizes the input feature map $X$ by population statistics which are estimated using Exponential Moving Average (EMA) during training. This special mechanism of BN enriched with domain-specific information of source domains poses a significant risk to the generalization performance of the model.", + "bbox": [ + 212, + 146, + 787, + 220 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.2 Remix Normalization", + "text_level": 1, + "bbox": [ + 215, + 244, + 439, + 258 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Previous studies [15,31] have shown that Batch Normalization (BN) can extract domain-specific discriminative information, but it can easily overfit to a particular domain data in a mini-batch, while Instance Normalization (IN) can reduce the domain-specific style information, but it results in the loss of some discriminative information. Therefore, we combine these two normalization methods for a trade-off. Inspired by DSON [36] and SN [29], we design Remix Normalization (RN) which combines the statistical parameters of BN and IN in a simple way. Given an input feature map $X_{k}$ , RN calculates mean $\\mu_k^{rn} \\in \\mathbb{R}^{N \\times C}$ and variance $\\sigma_k^{rn^2} \\in \\mathbb{R}^{N \\times C}$ for standardization as follows:", + "bbox": [ + 212, + 268, + 787, + 407 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mu_ {k} ^ {r n} = \\left(\\mu_ {k} ^ {\\mathrm {i n}} + \\mu_ {k} ^ {\\mathrm {b n}}\\right) / 2,\n$$\n", + "text_format": "latex", + "bbox": [ + 424, + 417, + 785, + 445 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\sigma_ {k} ^ {r n 2} = (\\sigma_ {k} ^ {i n 2} + \\sigma_ {k} ^ {b n 2}) / 2.\n$$\n", + "text_format": "latex", + "bbox": [ + 415, + 439, + 588, + 459 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Meanwhile, unlike previous approaches where each normalization layer corresponds to one BN buffer, we set up corresponding $Buffer_{k}:\\mu_{k}^{pop},\\sigma_{k}^{pop2}\\in \\mathbb{R}^{C},k\\in 1,2,\\ldots ,K$ for each RN for $K$ source domains to store their domain-specific information as follows:", + "bbox": [ + 212, + 469, + 787, + 529 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mu_ {k} ^ {\\text {p o p}} \\leftarrow \\lambda \\mu_ {k} ^ {\\text {p o p}} + (1 - \\lambda) \\mu_ {k} ^ {b n},\n$$\n", + "text_format": "latex", + "bbox": [ + 403, + 541, + 611, + 559 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\sigma_ {k} ^ {p o p 2} \\leftarrow \\lambda \\sigma_ {k} ^ {p o p 2} + (1 - \\lambda) \\sigma_ {k} ^ {b n 2}. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 392, + 554, + 784, + 582 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "As shown in Fig.2, we use three primary colors (yellow, red, blue) to represent the corresponding source domains $D_{1}$ , $D_{2}$ , $D_{K}$ . The input mini-batch is from $D_{1}$ . After calculating the required IN and BN statistical parameters, we follow Eq.5 to update the corresponding BN population statistics (i.e., $Buffer_{1}:\\mu_{1}^{pop},\\sigma_{1}^{pop2}$ ). In Fig.2, we represent this update process with solid lines.", + "bbox": [ + 212, + 593, + 787, + 669 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The efficacy of combining BN and IN has been verified in many previous DG ReID works [2, 33, 34]. Nevertheless, a common challenge persists across these approaches: overfitting of the normalization layers to the source domains. Specifically, during the training process, following Eq.4, the statistical parameters of IN and BN within the normalization layer are derived from the same source domain, resulting in a consistent distribution. However, when the model is directly tested on the target domain, there is no concept of mini-batch anymore [41]. Thus, RN calculates the parameters for standardization as follows:", + "bbox": [ + 212, + 670, + 787, + 791 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mu_ {k} ^ {r n} = \\left(\\mu_ {t a r} ^ {\\text {i n}} + \\mu_ {k} ^ {\\text {p o p}}\\right) / 2, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 416, + 803, + 785, + 830 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\sigma_ {k} ^ {r n 2} = (\\sigma_ {t a r} ^ {i n 2} + \\sigma_ {k} ^ {p o p 2}) / 2,\n$$\n", + "text_format": "latex", + "bbox": [ + 410, + 824, + 593, + 842 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "Rethinking Normalization Layers for DG-ReID", + "bbox": [ + 411, + 113, + 727, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 114, + 785, + 126 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "the IN statistical parameters (i.e., $\\mu_{tar}^{\\mathrm{in}},\\sigma_{tar}^{in2}$ ) are derived from the input target domain itself. Conversely, the BN statistical parameters are based on population statistics (i.e., $\\mu_k^{pop},\\sigma_k^{pop2}$ ) of source domains. The actual distribution between the target domain and the source domain is completely different, which leads to the inability of the statistics parameters (i.e., $\\mu_k^{rn},\\sigma_k^{rn2}$ ) to fit the input $X_{tar}$ from the target domain. Concurrently, the model lacks the adaptability to address this mismatch, resulting in a performance decline. Furthermore, the learnable affine parameters also tend to overfit to the source domains, leading to a decrease in performance during both training and testing.", + "bbox": [ + 212, + 145, + 787, + 282 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "3.3 Emulation Normalization", + "text_level": 1, + "bbox": [ + 215, + 301, + 470, + 315 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Remix Normalization enables the weight matrix in learning to extract discriminative features while also possessing some degree of domain invariance. However, a crucial limitation emerges when the model encounters the distribution shifts introduced by the unseen target domain, as it lacks the knowledge of how to adapt. Hence, a natural consideration arises: given the entirely distinct distributions among different source domains, we can treat the current source domain as a simulated target domain of a remaining source domain.", + "bbox": [ + 212, + 321, + 784, + 428 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Thus, we introduce a novel normalization operation named Emulation Normalization (EN), which emulates the inference procedure of Remix Normalization during training. By doing so, EN actively guides the weight matrix to acquire the capability of handling issues arising from the inability of RN to fit the input from unseen target domains. EN calculates mean $\\mu_k^{en} \\in \\mathbb{R}^{N \\times C}$ and variance $\\sigma_k^{en2} \\in \\mathbb{R}^{N \\times C}$ as follows:", + "bbox": [ + 212, + 429, + 785, + 518 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mu_ {k} ^ {e n} = \\left(1 - w _ {j} ^ {m}\\right) \\mu_ {k} ^ {\\mathrm {i n}} + w _ {j} ^ {m} \\mu_ {j} ^ {\\mathrm {p o p}} \\\\ \\sigma_ {k} ^ {e n 2} = \\left(1 - w _ {j} ^ {v}\\right) \\sigma_ {k} ^ {i n 2} + w _ {j} ^ {v} \\sigma_ {j} ^ {p o p 2}, \\tag {7} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 383, + 526, + 784, + 569 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where we treat the current input source domain $D_{k}$ as the simulated target domain of $D_{j}$ , which is randomly selected from the remaining. By combining IN statistical parameters of $X_{k}$ from this simulated target domain $D_{k}$ and BN population statistics from a remaining source domain $D_{j}$ , our EN can actively guide the weight matrix to transfer discriminative feature extraction from familiar source domains to unfamiliar target domain with different distribution. Notably, the learnable balancing ratios (i.e., $w_{j}^{m}, w_{j}^{v}$ ) correspond individually to the randomly selected source domain $D_{j}$ .", + "bbox": [ + 212, + 573, + 787, + 694 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "As shown in Fig.2, We use the current input $D_{1}$ as the simulated target domain of $D_{2}$ which supports the required BN statistical parameters stored in Buffer2 (i.e., $\\mu_2^{pop}, \\sigma_2^{pop^2}$ ), as specified in Eq.7. It is worth noting that Buffer2 could combine IN statistical parameters from $D_{1}$ and $D_K$ , this pattern holds true for all the buffers.", + "bbox": [ + 212, + 695, + 787, + 768 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "3.4 Domain Frozen", + "text_level": 1, + "bbox": [ + 215, + 787, + 387, + 801 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Another factor that may lead to a decrease in the model's generalization capability lies in the learnable affine parameters of the normalization layers. Compared", + "bbox": [ + 212, + 809, + 785, + 839 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "R. Nie and J. Ding et al.", + "bbox": [ + 271, + 114, + 437, + 128 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "to the weight matrix, the number of learnable affine parameters is significantly smaller, implying a weaker robustness when handling the target domain with a distribution divergence from the source domains. Moreover, ReID can be approximated as a channel-wise classification task, indicating that channel-wise affine parameters exert a more profound influence.", + "bbox": [ + 212, + 146, + 784, + 220 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Furthermore, as shown in Fig.3, we visualize the affine parameters for each source domain, approximating the domain-specific information within. It is evident that significant disparities exist among different source domains. These channel-wise divergences will accumulate into the weight matrix through backward propagation, leading to confusion in feature learning.", + "bbox": [ + 212, + 222, + 785, + 297 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Consequently, we propose Domain Frozen (DF), a strategy that stabilizes the weight matrix by freezing the updates of affine parameters in every normalization layer. For Remix Normalization, we maintain the affine parameters in a frozen state following their initialization with pre-trained weights. In the case of Emulation Normalization, the input feature map $X_{k}$ is standardized by statistical parameters from two different domains. The output $\\hat{X}_k$ inherently deviates from a normal distribution. Consequently, with the integration of Domain Frozen, we have opted to directly eliminate the original affine parameters, which is tantamount to setting the affine parameters $\\gamma$ and $\\beta$ to 1 and 0, respectively.", + "bbox": [ + 212, + 297, + 787, + 433 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Moreover, a further apparent advantage of DF is that it indirectly shares the affine parameters between the source domains (during training) and the target domain (during testing). This sharing further mitigates the domain shifts caused by the absence of target domain information, enhancing the model's ability to generalize to unseen domains.", + "bbox": [ + 212, + 433, + 787, + 508 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "3.5 Training and testing", + "text_level": 1, + "bbox": [ + 214, + 527, + 429, + 544 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "As illustrated in Fig.1(b), we propose a novel ReNorm framework to integrate the aforementioned normalization strategies. For clarification purposes, we refer to the weight matrix equipped with Remix Normalization (RN) as Network- $R$ and the same weight matrix equipped with Emulation Normalization (EN) as Network- $E$ . Meanwhile, Domain Frozen (DF) is applied uniformly in both RN and EN.", + "bbox": [ + 212, + 551, + 787, + 640 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Model Training. In our methodology, there is no need for additional design of loss functions. The employment of two fundamental ReID loss functions, triplet loss $(L_{tri})$ [13] and cross-entropy loss $(L_{ce})$ suffices for achieving optimal performance. In each iteration, we randomly select one source domain $D_{k}$ from a set of $K$ source domains and randomly sample a mini-batch samples of size $N_{b}$ from this domain. As illustrated in Fig.1, this set of data is then fed through the Network $R$ for first forward propagation and calculate corresponding triplet loss $L_{tri}^{R}$ and cross-entropy loss $L_{ce}^{R}$ , which is represented by the solid line. Subsequently, we treat the current source domain as a simulated target domain and randomly select one $D_{j}$ from the remaining source domains to serve as a simulated source domain, providing Batch Normalization (BN) statistical parameters (i.e., $\\mu_j^{pop},\\sigma_j^{pop^2}$ ) for the second forward propagation, which is indicated by the dashed line. The same mini-batch will be fed into Network $E$ to compute", + "bbox": [ + 212, + 642, + 787, + 840 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "Rethinking Normalization Layers for DG-ReID", + "bbox": [ + 413, + 114, + 727, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "the corresponding triplet loss $L_{tri}^{E}$ and cross-entropy loss $L_{ce}^{E}$ . Combing these two networks mentioned above, we have the following total loss for backward propagation:", + "bbox": [ + 212, + 145, + 782, + 191 + ], + "page_idx": 9 + }, + { + "type": "equation", + "text": "\n$$\nL _ {t o t a l} = L _ {t r i} ^ {R} + L _ {c e} ^ {R} + L _ {t r i} ^ {E} + L _ {c e} ^ {E}. \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 387, + 191, + 784, + 208 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Testing. During testing, we feed the target domain to both networks $K$ times each. Subsequently, we compute the average of the $2*K$ features obtained from two networks to obtain the final inference result.", + "bbox": [ + 212, + 214, + 782, + 258 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 214, + 282, + 375, + 300 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.1 Datasets and Settings", + "text_level": 1, + "bbox": [ + 214, + 314, + 442, + 329 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Datesets. Our approach is evaluated on 9 person ReID datasets, including 5 large-scale datasets: Market1501 (M) [47], MSMT17 (MS) [39], CUHK02 (C2) [22], CUHK03 (C3) [23], and CUHK-SYSU (CS) [42], and 4 small-scale datasets: PRID [14], GRID [27], VIPeR [7], and iLIDs [48]. We do not use DukeMTMCreID (D) [50] since it has been retracted. We adopt Cumulative Matching Characteristics (CMC) and mean Average Precision (mAP) as evaluation metrics.", + "bbox": [ + 212, + 339, + 782, + 429 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Evaluation Protocols. We set three evaluation protocols following [37,44]. For Protocol-1, all images including both training and testing sets from four large-scale datasets (i.e., M, CS, C2, C3) are used for training, and the four small-scale datasets (i.e., PRID, GRID, VIPeR, iLIDs) are used for testing. The final results for each dataset are evaluated on the average of 10 repeated random splits of gallery and query sets. For Protocol-2 and Protocol-3, following the leave-one-out setting in [44], we select one domain from $\\mathrm{M} + \\mathrm{MS} + \\mathrm{CS} + \\mathrm{C3}$ for testing, while the remaining three domains serve as the training sets. CS is excluded as a target domain due to its images all come from one camera view. Additionally, under Protocol-2, only the training sets of the source domains are used for training, while all images are exploited under Protocol-3.", + "bbox": [ + 212, + 430, + 784, + 595 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Implementation Details. We employ ResNet50 [9] pre-trained on ImageNet as our backbone. The size of a mini-batch is set to 64, including 16 identities and 4 images per identity. Each image is resized to $256 \\times 128$ . We use random horizontal flipping, random cropping, zero padding, color jittering, and auto-augmentation [11] as data augmentation. We also adopt a warm-up strategy in the first 10 epochs. The learning rate is initialized as $3.5 \\times 10^{-4}$ and divided by 10 at the 30th, 60th and, 90th epochs, respectively. All experiments are conducted using Pytorch on a single RTX 4090 GPU.", + "bbox": [ + 212, + 595, + 784, + 715 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.2 Comparison with the State-of-the-Arts", + "text_level": 1, + "bbox": [ + 214, + 739, + 581, + 753 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Comparison with DG-ReID Methods Under Protocol-1. As shown in Table 1, we compare our method with other state-of-the-arts under Protocol-1. Since DukeMTMC-reID (D) [50] has been retracted, we only use other large-scale datasets for training. Notably, our method demonstrates superior performance on PRID, GRID, and VIPeR. Overall, ReNorm significantly outperforms other", + "bbox": [ + 212, + 763, + 782, + 839 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "R. Nie and J. Ding et al.", + "bbox": [ + 271, + 114, + 437, + 128 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/cfdcbb0a6e66c96d2972cf2cb86bd33e0af1d12c80ab89fa06f710d0a68caec4.jpg", + "table_caption": [ + "Table 1: Comparison with the SOTA methods in DG ReID under the setting of Protocol-1. All results are the average of 10 random sampling. The best and the second best results are highlighted in bold and underline." + ], + "table_footnote": [], + "table_body": "
MethodReferenceSource domainsTarget domainAverage
PRIDGRIDVIPeRiLIDs
mAPRank-1mAPRank-1mAPRank-1mAPRank-1mAPRank-1
SNR [19]CVPR'20M+D+C266.552.147.740.261.352.989.984.166.457.3
RaMoE [4]CVPR'2167.357.754.246.864.656.690.285.069.161.5
DMG-Net [1]CVPR'21+C3+CS68.460.656.651.060.453.983.979.367.361.2
DTIN-Net [18]ECCV'2279.771.060.651.870.762.987.281.874.666.9
MetaBIN [3]CVPR'21M+C270.861.257.950.264.355.982.774.768.960.5
META [44]ECCV'2271.761.960.152.468.461.583.579.268.960.5
ACL [45]ECCV'2273.463.065.755.275.166.486.581.875.266.6
CINorm [2]TMM'23+C3+CS80.572.459.949.476.767.688.082.276.267.9
PAOA [21]WACV'2474.065.667.256.376.666.787.183.176.267.9
ReNormThis paper81.373.967.559.477.069.187.082.278.271.2
", + "bbox": [ + 220, + 196, + 785, + 344 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "methods by at least $2.0\\%$ and $3.3\\%$ in average mAP and Rank-1, respectively. The experimental findings substantiate the efficacy of our ReNorm, which combines simplicity with substantial impact.", + "bbox": [ + 212, + 375, + 782, + 420 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Comparison Under Protocol-2 and Protocol-3. We compare our method with previous state-of-the-arts under protocol-2 and protocol-3 to evaluate our method on large-scale datasets. As shown in Table 2, we can find that our ReNorm achieves superior performance compared with previous methods on almost all datasets under the two protocols. Overall, ReNorm outperforms other approaches by an average of at least $2.0\\%$ mAP and $4.1\\%$ Rank-1 under Protocol-2, and by $1.7\\%$ mAP and $2.9\\%$ Rank-1 under Protocol-3, respectively. Notably, when MSMT17 (MS) serves as the target domain, our method surpasses the second best by $3.1\\%$ mAP and $5.7\\%$ Rank-1 under Protocol-2, and by $2.7\\%$ mAP and $5.7\\%$ Rank-1 under Protocol-3. Collectively, these results further illustrate our model's superiority in domain generalization.", + "bbox": [ + 212, + 421, + 787, + 587 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4.3 Ablation Study", + "text_level": 1, + "bbox": [ + 215, + 613, + 388, + 628 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Effectiveness of different modules. We investigate the effects of different components in ReNorm design to study their individual contributions, as shown in Table 3. The first row corresponds to a basic ResNet50 framework. In the table, every even row compared to the previous row incorporates the Domain Frozen (DF) strategy, which straightforwardly boosts the network by at least $1.3\\%$ mAP and $1.5\\%$ Rank-1, on average. Comparing first and second rows with the third and fourth rows, even when Remix Normalization (RN) simply combines IN and BN, it significantly enhances the network's generalization performance. Furthermore, with the addition of Emulation Normalization (EN), the two complementary normalization layers further improve the model's generalization ability, resulting in an average improvement of at least $2.0\\%$ mAP and $2.5\\%$ Rank-1. The experiments thoroughly demonstrate the simplicity and effectiveness of our approach.", + "bbox": [ + 212, + 643, + 787, + 840 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "Rethinking Normalization Layers for DG-ReID", + "bbox": [ + 411, + 113, + 727, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/14bcc82f1472b265f3bc9d733cd84f6cb3abd6d9fcadfaf2e9f3c619ed68d9c1.jpg", + "table_caption": [ + "Table 2: Comparison with the SOTA methods in DG ReID under the setting of Protocol-2 and Protocol-3. All results are the average of 10 random sampling. The best and the second best results are highlighted in bold and underline." + ], + "table_footnote": [], + "table_body": "
MethodReferenceSettingM+MS+CS→C3M+CS+C3→MSMS+CS+C3→MAverage
mAP Rank-1mAP Rank-1mAP Rank-1mAP Rank-1mAP Rank-1mAP Rank-1mAP Rank-1mAP Rank-1
SNR [19]CVPR'20Protocol-28.98.96.819.934.662.716.830.5
M3L [46]CVPR'2120.931.915.936.958.479.931.749.6
MetaBIN [3]CVPR'2128.828.117.840.257.980.134.849.5
META [44]ECCV'2236.335.122.549.967.586.142.157.0
ACL [45]ECCV'2241.241.820.445.974.389.345.359.0
ReNormThis Paper43.644.725.655.672.789.147.363.1
SNR [19]CVPR'20Protocol-317.517.17.722.052.477.825.939.0
M3L [46]CVPR'2135.736.517.438.662.482.738.552.6
MetaBIN [3]CVPR'2143.043.118.841.267.284.543.056.3
META [44]ECCV'2247.146.224.452.176.590.549.362.9
ACL [45]ECCV'2249.950.121.747.376.890.649.362.7
PAOA [21]WACV'2449.850.525.151.577.190.850.764.3
ReNormThis Paper51.952.527.857.877.491.252.467.2
", + "bbox": [ + 220, + 196, + 785, + 407 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/ee17437053028f2036b20061c43a1c1e116109a3fd0b548cf0c54d0775d3d298.jpg", + "table_caption": [ + "Table 3: Ablation studies for different components of our framework. The experiment is conducted under protocol-2. The best results are highlighted in bold." + ], + "table_footnote": [], + "table_body": "
RNENDFTarget: C3Target: MSTarget: MAverage
mAP Rank-1mAP Rank-1mAP Rank-1mAP Rank-1
35.334.919.944.966.385.140.555.0
✓36.536.922.148.568.286.642.357.3
✓40.540.522.149.469.087.143.959.0
✓✓40.440.223.752.771.988.945.360.6
✓✓42.543.624.453.471.187.946.061.6
✓✓✓43.644.725.655.672.789.147.363.1
", + "bbox": [ + 240, + 463, + 763, + 579 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Visualization of the domain-specific information within the affine parameters. In this experiment, we establish corresponding RN for each source domain and reinstate the updates of all affine parameters. After training, we sum the affine parameters $\\gamma$ of each source domain at each RN layer to approximate the domain-specific information within. As shown in Fig.3, The horizontal axis represents the layer position of RN in the backbone, while the vertical axis represents the corresponding proportions. It is evident that significant disparities exist among source domains across all the normalization layers. During training, the disparities accumulate in the weight matrix, potentially leading to misguided learning.", + "bbox": [ + 215, + 611, + 785, + 762 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Effectiveness on where to plug the modules. As normalization-based modules, the proposed ReNorm and DF can be plugged into any stage of the network by replacing the corresponding BN layers. In our experiments, we employ ResNet50 as our backbone, which consists of four groups of residual blocks and is organized into four stages. We replace our ReNorm and DF with diverse stages'", + "bbox": [ + 215, + 763, + 785, + 839 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "R. Nie and J. Ding et al.", + "bbox": [ + 271, + 114, + 437, + 128 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/8bcb600a2b32df48b6e33eff9ebfba688aac5dd196b5f5ceb4afb567a5787cdc.jpg", + "image_caption": [ + "Fig. 3: Visualization of the domain-specific information within the affine parameters under Protocol-2, the source domains used for training being MSMT17, CUHK-SYSU and Market1501." + ], + "image_footnote": [], + "bbox": [ + 222, + 148, + 777, + 282 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/670f466b2dc0af418813204d9247729ad9d7bce9baea9e1ea1bc00b74f55171f.jpg", + "table_caption": [ + "Table 4: Ablation study on where to plug the modules into ResNet50." + ], + "table_footnote": [], + "table_body": "
StageTarget: C3
Module: ReNormModule: DF
mAPRank-1mAPRank-1
None34.935.342.543.6
439.439.843.043.3
3 - 441.942.043.343.5
2 - 442.243.343.144.0
1 - 443.644.743.644.7
", + "bbox": [ + 232, + 393, + 486, + 508 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/d1cd79c0750ed3afb75efd4400bec45b6095ce314778aba5f01006105353ad59.jpg", + "table_caption": [ + "Table 5: Ablation study on different combination of outputs into the result." + ], + "table_footnote": [], + "table_body": "
MethodAverage
mAPRank-1
concat(R)46.261.8
concat(E)45.860.8
mean(R)46.361.9
mean(E)45.860.9
concat(R,E)47.062.8
mean(R,E)47.363.1
", + "bbox": [ + 527, + 393, + 781, + 508 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "BN layers to explore the effects of different plugging stages. As shown in Table 4, the results demonstrate that employing ReNorm and DF across all normalization layers performs better than others. Additionally, incorporating ReNorm and DF from any stage could improve overall performance, further demonstrating the effectiveness of our modules.", + "bbox": [ + 212, + 537, + 784, + 611 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Combination of outputs into the final result. The experiment is conducted under Protocol-2. During testing, each image will go through Network- $R$ (R) and Network- $E$ (E) separately $K$ times, resulting in a total of $2*K$ outputs. As shown in Table 5, from the first four rows, it is evident that network- $R$ generally outperforms network- $E$ , and both have surpassed the previous SOTA with $45.3\\%$ mAP and $59.0\\%$ Rank-1 on average. After combining the outputs of the two networks, there is a significant improvement in the final results. Ultimately, taking the mean of the $2*K$ outputs directly yields the best result with $47.3\\%$ mAP and $63.1\\%$ Rank-1. Experimental results confirm the enhanced generalization capability of our framework in domain-invariant feature extraction.", + "bbox": [ + 212, + 613, + 785, + 763 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Visualization of the target domain distribution. To further demonstrate the effectiveness of our ReNorm framework, we visualize the feature distribution of unseen target domains under Protocol-1. We conduct t-SNE [30] experiments on the same query and gallery image pairs from four target domains by implementing BN, IN, and our ReNorm, respectively. In Fig.4 (a), the", + "bbox": [ + 212, + 763, + 785, + 839 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "Rethinking Normalization Layers for DG-ReID", + "bbox": [ + 413, + 114, + 727, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/a2237fb5379cbc87bf9176ae98899e168605bf1247f29f20f597f214b58ca8b3.jpg", + "image_caption": [ + "(a) BN" + ], + "image_footnote": [], + "bbox": [ + 222, + 143, + 390, + 273 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/f3105fdb03b99fb47ff956e2397506bf429106bf2710ad4846e04b12f4777628.jpg", + "image_caption": [ + "(b) IN", + "Fig. 4: The t-SNE results of features under Protocol-1 on three models with different normalization layers. (a) indicates ResNet50 with BN for all normalization layers, (b) denotes using instance normalization, and (c) utilizes our ReNorm for all normalization layers. Various colors illustrate samples from different target domains, and different shapes denote sample pairs from query or gallery sets." + ], + "image_footnote": [], + "bbox": [ + 418, + 143, + 586, + 273 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/a181018657d9b21f620839e461fbcdaf07355feb381fac66458922e675d8a97f.jpg", + "image_caption": [ + "(c) ReNorm (ours)" + ], + "image_footnote": [], + "bbox": [ + 614, + 143, + 784, + 273 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "\"BN\" model exhibits smaller distances between query and gallery sets of the same identity but is heavily influenced by domain factors, leading to clear domain boundaries. In contrast, Fig.4 (b) shows that IN removes these boundaries but increases the distance between query and gallery sample pairs. The observations prove that BN can capture more discriminative features but tends to overfit to source domains, while IN can eliminate domain-specific information to enhance generalization but inevitably removes discriminative information. Fig.4 (c) shows that the distances between query and gallery sample pairs are much smaller, demonstrating that our approach achieves better performance. Simultaneously, there are no apparent boundaries among target domains, indicating that our method effectively eliminates domain-specific information.", + "bbox": [ + 212, + 407, + 787, + 574 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 611, + 357, + 628 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In this paper, we present a purely normalization-based framework called ReNorm for tackling the problem of domain generalizable person re-identification (DG-ReID). Our framework integrates two complementary normalization layers through two forward propagations for the same weight matrix. Remix Normalization (RN) enables the weight matrix to extract discriminative features within the source domains, while Emulation Normalization (EN) is employed to actively guide the weight matrix to learn how to generalize the feature extraction capability to unseen target domains. Meanwhile, we introduce Domain Frozen (DF), a strategy that stabilizes the weight matrix by freezing the updates of affine parameters in normalization layers. Extensive experiments demonstrate the simplicity yet state-of-the-art performance of ReNorm among all the popular benchmarks.", + "bbox": [ + 212, + 657, + 795, + 839 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "R. Nie and J. Ding et al.", + "bbox": [ + 271, + 114, + 437, + 128 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 217, + 143, + 401, + 162 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "This work was supported by the Natural Science Foundation of China (No. 62372082, 61972071, U20A20184), Natural Science Foundation of Sichuan Province (No.2023NSFSC0485), National Science Foundation for Distinguished Young Scholars under Grant 62225605 and Zhejiang Key Research and Development Program under Grant 2023C03196.", + "bbox": [ + 212, + 176, + 797, + 253 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 277, + 321, + 292 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Bai, Y., Jiao, J., Ce, W., Liu, J., Lou, Y., Feng, X., Duan, L.Y.: Person30k: A dual-meta generalization network for person re-identification. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2123-2132 (2021)", + "2. Chen, Z., Wang, W., Zhao, Z., Su, F., Men, A., Dong, Y.: Cluster-instance normalization: A statistical relation-aware normalization for generalizable person re-identification. IEEE Transactions on Multimedia (2023)", + "3. Choi, S., Kim, T., Jeong, M., Park, H., Kim, C.: Meta batch-instance normalization for generalizable person re-identification. In: Proceedings of the IEEE/CVF conference on Computer Vision and Pattern Recognition. pp. 3425-3435 (2021)", + "4. Dai, Y., Li, X., Liu, J., Tong, Z., Duan, L.Y.: Generalizable person re-identification with relevance-aware mixture of experts. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16145-16154 (2021)", + "5. Dai, Y., Liu, J., Sun, Y., Tong, Z., Zhang, C., Duan, L.Y.: Idm: An intermediate domain module for domain adaptive person re-id. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 11864-11874 (2021)", + "6. Fu, D., Chen, D., Bao, J., Yang, H., Yuan, L., Zhang, L., Li, H., Chen, D.: Unsupervised pre-training for person re-identification. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 14750-14759 (2021)", + "7. Gray, D., Tao, H.: Viewpoint invariant pedestrian recognition with an ensemble of localized features. ECCV (1) 2008, 262-275 (2008)", + "8. Han, K., Si, C., Huang, Y., Wang, L., Tan, T.: Generalizable person re-identification via self-supervised batch norm test-time adaption. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 36, pp. 817-825 (2022)", + "9. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 770-778 (2016)", + "0. He, L., Liang, J., Li, H., Sun, Z.: Deep spatial feature reconstruction for partial person re-identification: Alignment-free approach. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 7073-7082 (2018)", + "1. He, L., Liao, X., Liu, W., Liu, X., Cheng, P., Mei, T.: Fastreid: A python toolbox for general instance re-identification. arXiv preprint arXiv:2006.02631 (2020)", + "12. He, S., Luo, H., Chen, W., Zhang, M., Zhang, Y., Wang, F., Li, H., Jiang, W.: Multi-domain learning and identity mining for vehicle re-identification. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops. pp. 582-583 (2020)", + "13. Hermans, A., Beyer, L., Leibe, B.: In defense of the triplet loss for person re-identification. arXiv preprint arXiv:1703.07737 (2017)" + ], + "bbox": [ + 218, + 310, + 785, + 840 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "Rethinking Normalization Layers for DG-ReID", + "bbox": [ + 411, + 114, + 727, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "14. Hirzer, M., Beleznai, C., Roth, P.M., Bischof, H.: Person re-identification by descriptive and discriminative classification. In: Image Analysis: 17th Scandinavian Conference, SCIA 2011, Ystad, Sweden, May 2011. Proceedings 17. pp. 91-102. Springer (2011)", + "15. Huang, X., Belongie, S.: Arbitrary style transfer in real-time with adaptive instance normalization. In: Proceedings of the IEEE international conference on computer vision. pp. 1501-1510 (2017)", + "16. Ioffe, S., Szegedy, C.: Batch normalization: Accelerating deep network training by reducing internal covariate shift. In: International conference on machine learning. pp. 448-456. pmlr (2015)", + "17. Jiang, K., Zhang, T., Zhang, Y., Wu, F., Rui, Y.: Self-supervised agent learning for unsupervised cross-domain person re-identification. IEEE Transactions on Image Processing 29, 8549–8560 (2020)", + "18. Jiao, B., Liu, L., Gao, L., Lin, G., Yang, L., Zhang, S., Wang, P., Zhang, Y.: Dynamically transformed instance normalization network for generalizable person re-identification. In: European Conference on Computer Vision. pp. 285–301. Springer (2022)", + "19. Jin, X., Lan, C., Zeng, W., Chen, Z., Zhang, L.: Style normalization and restitution for generalizable person re-identification. In: proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 3143-3152 (2020)", + "20. Lee, G., Lee, S., Kim, D., Shin, Y., Yoon, Y., Ham, B.: Camera-driven representation learning for unsupervised domain adaptive person re-identification. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 11453-11462 (2023)", + "21. Li, Q., Gong, S.: Mitigate domain shift by primary-auxiliary objectives association for generalizing person reid. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 394-403 (2024)", + "22. Li, W., Wang, X.: Locally aligned feature transforms across views. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 3594-3601 (2013)", + "23. Li, W., Zhao, R., Xiao, T., Wang, X.: Deepreid: Deep filter pairing neural network for person re-identification. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 152-159 (2014)", + "24. Li, Y., Wang, N., Shi, J., Hou, X., Liu, J.: Adaptive batch normalization for practical domain adaptation. Pattern Recognition 80, 109-117 (2018)", + "25. Lin, C.S., Cheng, Y.C., Wang, Y.C.F.: Domain generalized person re-identification via cross-domain episodic learning. In: 2020 25th International Conference on Pattern Recognition (ICPR). pp. 6758-6763 (2021). https://doi.org/10.1109/ICPR48806.2021.9413013", + "26. Liu, J., Zha, Z.J., Chen, D., Hong, R., Wang, M.: Adaptive transfer network for cross-domain person re-identification. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 7202-7211 (2019)", + "27. Loy, C.C., Xiang, T., Gong, S.: Time-delayed correlation analysis for multi-camera activity understanding. International Journal of Computer Vision 90, 106-129 (2010)", + "28. Luo, H., Jiang, W., Gu, Y., Liu, F., Liao, X., Lai, S., Gu, J.: A strong baseline and batch normalization neck for deep person re-identification. IEEE Transactions on Multimedia 22(10), 2597-2609 (2019)", + "29. Luo, P., Ren, J., Peng, Z., Zhang, R., Li, J.: Differentiable learning-to-normalize via switchable normalization. arXiv preprint arXiv:1806.10779 (2018)" + ], + "bbox": [ + 215, + 146, + 785, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "R. Nie and J. Ding et al.", + "bbox": [ + 271, + 114, + 437, + 128 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "30. Van der Maaten, L., Hinton, G.: Visualizing data using t-sne. Journal of machine learning research 9(11) (2008)", + "31. Nam, H., Kim, H.E.: Batch-instance normalization for adaptively style-invariant neural networks. Advances in Neural Information Processing Systems 31 (2018)", + "32. Ni, H., Song, J., Luo, X., Zheng, F., Li, W., Shen, H.T.: Meta distribution alignment for generalizable person re-identification. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2487-2496 (2022)", + "33. Pan, X., Luo, P., Shi, J., Tang, X.: Two at once: Enhancing learning and generalization capacities via ibn-net. In: Proceedings of the European Conference on Computer Vision (ECCV). pp. 464-479 (2018)", + "34. Qi, L., Wang, L., Shi, Y., Geng, X.: A novel mix-normalization method for generalizable multi-source person re-identification. IEEE Transactions on Multimedia (2022)", + "35. Ren, M., He, L., Liao, X., Liu, W., Wang, Y., Tan, T.: Learning instance-level spatial-temporal patterns for person re-identification. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 14930-14939 (2021)", + "36. Seo, S., Suh, Y., Kim, D., Kim, G., Han, J., Han, B.: Learning to optimize domain specific normalization for domain generalization. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXII 16. pp. 68-83. Springer (2020)", + "37. Song, J., Yang, Y., Song, Y.Z., Xiang, T., Hospedales, T.M.: Generalizable person re-identification by domain-invariant mapping network. In: Proceedings of the IEEE/CVF conference on Computer Vision and Pattern Recognition. pp. 719-728 (2019)", + "38. Ulyanov, D., Vedaldi, A., Lempitsky, V.: Instance normalization: The missing ingredient for fast stylization. arXiv preprint arXiv:1607.08022 (2016)", + "39. Wei, L., Zhang, S., Gao, W., Tian, Q.: Person transfer gan to bridge domain gap for person re-identification. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 79-88 (2018)", + "40. Wu, Y., Chi, Z., Wang, Y., Plataniotis, K.N., Feng, S.: Test-time domain adaptation by learning domain-aware batch normalization. arXiv preprint arXiv:2312.10165 (2023)", + "41. Wu, Y., Johnson, J.: Rethinking\" batch\" in batchnorm. arXiv preprint arXiv:2105.07576 (2021)", + "42. Xiao, T., Li, S., Wang, B., Lin, L., Wang, X.: End-to-end deep learning for person search. arXiv preprint arXiv:1604.01850 2(2), 4 (2016)", + "43. Xu, B., He, L., Liao, X., Liu, W., Sun, Z., Mei, T.: Black re-id: A head-shoulder descriptor for the challenging problem of person re-identification. In: Proceedings of the 28th ACM International Conference on Multimedia. pp. 673–681 (2020)", + "44. Xu, B., Liang, J., He, L., Sun, Z.: Mimic embedding via adaptive aggregation: Learning generalizable person re-identification. In: Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XIV. pp. 372-388. Springer (2022)", + "45. Zhang, P., Dou, H., Yu, Y., Li, X.: Adaptive cross-domain learning for generalizable person re-identification. In: European Conference on Computer Vision. pp. 215-232. Springer (2022)", + "46. Zhao, Y., Zhong, Z., Yang, F., Luo, Z., Lin, Y., Li, S., Sebe, N.: Learning to generalize unseen domains via memory-based multi-source meta-learning for person re-identification. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6277-6286 (2021)" + ], + "bbox": [ + 212, + 146, + 787, + 839 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "Rethinking Normalization Layers for DG-ReID", + "bbox": [ + 411, + 114, + 727, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "47. Zheng, L., Shen, L., Tian, L., Wang, S., Wang, J., Tian, Q.: Scalable person re-identification: A benchmark. In: Proceedings of the IEEE international conference on computer vision. pp. 1116-1124 (2015)", + "48. Zheng, W., Gong, S., Xiang, T.: Associating groups of people. In: British Machine Vision Conference (2009)", + "49. Zheng, Z., Ruan, T., Wei, Y., Yang, Y., Mei, T.: VehicleNet: Learning robust visual representation for vehicle re-identification. IEEE Transactions on Multimedia 23, 2683-2693 (2020)", + "50. Zheng, Z., Zheng, L., Yang, Y.: Unlabeled samples generated by gan improve the person re-identification baseline in vitro. In: Proceedings of the IEEE international conference on computer vision. pp. 3754-3762 (2017)" + ], + "bbox": [ + 215, + 146, + 784, + 300 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "R. Nie and J. Ding et al.", + "bbox": [ + 271, + 114, + 437, + 128 + ], + "page_idx": 17 + } +] \ No newline at end of file diff --git a/2024/Rethinking Normalization Layers for Domain Generalizable Person Re-identification/02192c3d-6e2f-475e-b7a0-eaf65cc6bd4c_model.json b/2024/Rethinking Normalization Layers for Domain Generalizable Person Re-identification/02192c3d-6e2f-475e-b7a0-eaf65cc6bd4c_model.json new file mode 100644 index 0000000000000000000000000000000000000000..ceb83f611f4a4e70d034c41cfb047185e8db8eb2 --- /dev/null +++ b/2024/Rethinking Normalization Layers for Domain Generalizable Person Re-identification/02192c3d-6e2f-475e-b7a0-eaf65cc6bd4c_model.json @@ -0,0 +1,2271 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.244, + 0.141, + 0.761, + 0.185 + ], + "angle": 0, + "content": "Rethinking Normalization Layers for Domain Generalizable Person Re-identification" + }, + { + "type": "text", + "bbox": [ + 0.331, + 0.213, + 0.672, + 0.229 + ], + "angle": 0, + "content": "Ren Nie\\(^{1}\\), Jin Ding\\(^{1}\\), Xue Zhou\\(^{2,1*}\\), and Xi Li\\(^{3}\\)" + }, + { + "type": "text", + "bbox": [ + 0.229, + 0.24, + 0.774, + 0.268 + ], + "angle": 0, + "content": "1 University of Electronic Science and Technology of China, Chengdu, China \nnieren@std.uestc.edu.cn, dingjin1998@outlook.com, zhouxue@uestc.edu.cn" + }, + { + "type": "text", + "bbox": [ + 0.334, + 0.268, + 0.671, + 0.282 + ], + "angle": 0, + "content": "2 Shenzhen Institute for Advanced Study, UESTC" + }, + { + "type": "text", + "bbox": [ + 0.282, + 0.282, + 0.721, + 0.311 + ], + "angle": 0, + "content": "3 College of Computer Science & Technology, Zhejiang University xilizju@zju.edu.cn" + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.354, + 0.744, + 0.673 + ], + "angle": 0, + "content": "Abstract. Domain Generalizable Person Re-Identification (DG-ReID) strives to transfer learned feature representation from source domains to unseen target domains, despite significant distribution shifts. While most existing methods enhance model generalization and discriminative feature extraction capability by introducing Instance Normalization (IN) in combination with Batch Normalization (BN), these approaches still struggle with the overfitting of normalization layers to the source domains, posing challenges in domain generalization. To address this issue, we propose ReNorm, a purely normalization-based framework that integrates two complementary normalization layers through two forward propagations for the same weight matrix. In the first forward propagation, Remix Normalization (RN) combines IN and BN in a concise manner to ensure the feature extraction capability. As an effective complement to RN, Emulation Normalization (EN) simulates the testing process of RN, implicitly mitigating the domain shifts caused by the absence of target domain information and actively guiding the model in learning how to generalize the feature extraction capability to unseen target domains. Meanwhile, we propose Domain Frozen (DF), freezing updates to affine parameters to reduce the impact of these less robust parameters on overfitting to the source domains. Extensive experiments show that our framework achieves state-of-the-art performance on the popular benchmarks. The code is available at https://github.com/3699nr/ReNorm." + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.687, + 0.733, + 0.7 + ], + "angle": 0, + "content": "Keywords: Normalization \\(\\cdot\\) Domain generalization \\(\\cdot\\) Re-identification" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.73, + 0.377, + 0.746 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Person Re-identification (ReID) is a practical task that has been widely applied in video surveillance and security. It aims at retrieving identical persons across a multi-camera network with non-overlapping fields of view. While current approaches [6,10,12,28,35,43,49] have achieved impressive performance within the same source domains during both training and testing phases. However, when" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.438, + 0.129 + ], + "angle": 0, + "content": "R. Nie and J. Ding et al." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.147, + 0.787, + 0.359 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.37, + 0.788, + 0.495 + ], + "angle": 0, + "content": "Fig. 1: Differences between existing normalization-based DG-ReID methods and ReNorm. To illustrate more clearly, we use the example of training with three source domains. (a) Existing normalization-based methods combine IN and BN in a specific way to ensure both the generalization and discrimination capability of the weight matrix. (b) ReNorm integrates two complementary normalization layers through two forward propagations for the same weight matrix. RN enables the weight matrix to extract discriminative features within the source domains, while EN is employed to actively guide the weight matrix to generalize the feature extraction capability to unseen target domains. (c) Some details of ReNorm. Detailed explanations can be found in Sec 3." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.534, + 0.788, + 0.655 + ], + "angle": 0, + "content": "directly deploying these well-trained models on an unseen target domain, an issue known as \"domain shift\" [39] arises, significantly impacting the performance of the models. To alleviate this problem, recent efforts have been devoted to two approaches: unsupervised domain adaptive (DA) ReID [5,17,20,26], and domain generalizable (DG) ReID [2,3,19,37,44,45]. In contrast to DA ReID, DG ReID is more practical and challenging as it trains on the multiple source domains and then directly transfers the learned representation to the unseen target domain without any target data for fine-tuning." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.659, + 0.794, + 0.842 + ], + "angle": 0, + "content": "Among numerous DG ReID methods, normalization-based approaches [2,3,8, 19,37] stand out for their simplicity and versatility, often integrating seamlessly with other advanced techniques such as meta-learning, distribution alignment, and ensemble learning. Illustrated in Fig.1(a), the combination of Batch Normalization (BN) [16] for capturing discriminative features and Instance Normalization (IN) [38] for eliminating domain-specific information facilitates normalization-based DG-ReID methods in achieving both high generalization and discrimination capabilities. During training, the statistical parameters of both IN and BN within the normalization layers are derived from the same data, ensuring a consistent distribution. However, this consistency is disrupted when the model is deployed on the target domain, as IN statistical parameters originate from the target domain itself, while BN statistical parameters are derived from population" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.414, + 0.115, + 0.728, + 0.129 + ], + "angle": 0, + "content": "Rethinking Normalization Layers for DG-ReID" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.785, + 0.223 + ], + "angle": 0, + "content": "statistics accumulated from source domains during training. The distribution of the target domain significantly differs from that of the source domains, resulting in a marked decline in the model's performance. Furthermore, the vulnerable and sensitive learnable affine parameters are also prone to overfitting to the source domains, leading to performance degeneration in both training and testing phase." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.224, + 0.787, + 0.344 + ], + "angle": 0, + "content": "Moreover, emerging researches [24, 40] have shown that the identity and domain information are separately embedded in the weight matrix (learnable parameters in convolutional layers) and normalization layers. Consequently, we rethink the key challenges of DG ReID, which are rooted in normalization layers. We believe that the capability of the weight matrix to extract discriminative features sets the upper limit of the model's performance. Simultaneously, the normalization layers, responsible for fitting and scaling the input from target domains, establish the lower limit for the model's generalization performance." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.345, + 0.788, + 0.496 + ], + "angle": 0, + "content": "To address the issue of normalization layers being prone to overfitting to the source domains, we restructure the normalization layers, proposing ReNorm, a concise yet effective DG ReID framework entirely based on the design of normalization layers, as shown in Fig.1(b). In each iteration, the network undergoes two forward propagations and then one backpropagation after aggregating two forward losses. The two forward propagations apply different processing to the normalization layers, utilizing two complementary normalization layers to facilitate the weight matrix in extracting discriminative features from familiar source domains, while simultaneously guiding the same weight matrix to learn how to transfer this acquired feature extraction capability to unfamiliar target domains." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.497, + 0.788, + 0.752 + ], + "angle": 0, + "content": "Specifically, in our ReNorm framework, the two forward propagations correspond to two different but complementary normalization strategies: Remix Normalization (RN) and Emulation Normalization (EN). In the first forward propagation, RN introduces Instance Normalization (IN) in combination with Batch Normalization (BN) by directly summing their statistical parameters for standardization. Meanwhile, the BN statistical parameters of the current source domain are stored in the corresponding buffer, as shown in Fig.1(c). In the second forward propagation, EN adopts a similar summing approach but emulates the inference process of RN. It treats the current input as a simulated target domain, and calculates the IN statistical parameters accordingly. Simultaneously, the required BN statistical parameters are obtained from a remaining source domain's buffer stored in RN. This pairwise normalization strategy enables the network to actively learn how to transfer knowledge from source domains to the target domain with a distinct distribution. Additionally, we introduce Domain Frozen (DF), where the vulnerable settings of learnable affine parameters within both normalization layers are frozen and exempt from updates. By doing so, ReNorm greatly raises the lower limit of the model's generalization ability." + }, + { + "type": "text", + "bbox": [ + 0.241, + 0.753, + 0.555, + 0.767 + ], + "angle": 0, + "content": "We summarize our contributions as follows:" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.78, + 0.787, + 0.84 + ], + "angle": 0, + "content": "- We introduce ReNorm, a simple yet effective framework for DG ReID, entirely and solely based on the design of normalization layers. This framework significantly mitigates performance degradation resulting from distribution shifts between source and target domains in a concise yet effective way." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.232, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.438, + 0.129 + ], + "angle": 0, + "content": "R. Nie and J. Ding et al." + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.147, + 0.784, + 0.207 + ], + "angle": 0, + "content": "- We introduce two complementary normalization strategies (i.e., Remix Normalization and Emulation Normalization) equipped with Domain Frozen, actively enabling the network to learn how to transfer the feature extraction capability acquired from source domains to unseen target domains." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.207, + 0.785, + 0.237 + ], + "angle": 0, + "content": "- Extensive experiments demonstrate the simplicity yet state-of-the-art performance of our ReNorm framework under multiple testing protocols." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.255, + 0.631, + 0.272 + ], + "angle": 0, + "content": "2 Generalizable Person Re-Identification." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.282, + 0.788, + 0.433 + ], + "angle": 0, + "content": "Deep learning-based Person ReID approaches [6,10,12,28,35,43,49] have made great progress in recent years, achieving impressive accuracy in Person ReID. However, when these well-trained models are applied to unseen target domains that have distribution shifts with source domains, the performance decreases significantly. To alleviate this problem, recent efforts are devoted to two approaches: unsupervised domain adaptive (DA) ReID [5,17,20,26] and domain generalizable (DG) ReID. Diverging from the DA ReID, which needs unlabeled target domain datasets for unsupervised fine-tuning. DG ReID trains on the source domains and then directly tests on the unseen target domain. As a result, DG ReID has attracted attention for its robustness and generalization." + }, + { + "type": "text", + "bbox": [ + 0.24, + 0.433, + 0.751, + 0.447 + ], + "angle": 0, + "content": "Existing DG ReID methods can be briefly divided into four categories:" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.448, + 0.788, + 0.629 + ], + "angle": 0, + "content": "Normalization strategies [2, 33, 34]. Batch Normalization (BN) can preserve diversity among samples, capturing more discriminative features and accelerating convergence. However, it tends to be prone to overfitting to source domains. Instance Normalization (IN) can eliminate domain-specific information to enhance generalization but inevitably removes discriminative information. Thus, normalization-based methods typically combine IN and BN in various ways to ensure both high generalization and discrimination capability of the network. Jiao et al. [18] proposed Dynamically Transformed Instance Normalization (DTIN) to alleviate the drawback of IN by transforming normalized features into appropriate representations in a learnable and adaptive manner. Chen et al. [2] designed a Cluster-Instance Normalization (CINorm) to eliminate the effect of outliers in original BN that causes severe overfitting." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.63, + 0.788, + 0.719 + ], + "angle": 0, + "content": "Domain alignment [19,32]. These methods try to minimize the difference among source domains for learning domain-invariant representations. They assume that the invariant features from the source domains should also be robust to any unseen target domains. Jin et al. [19] proposed a dual causality loss to align the outputs and a Style Normalization and Restitution (SNR) module to distill identity-relevant features via a restitution module for IN compensation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.788, + 0.795 + ], + "angle": 0, + "content": "Meta-learning [3,37,46]. These methods improve generalized performance through virtual simulations of real train-test domain shifts. Following the concept of 'learning to lean', Choi et al. [3]proposed MetaBIN which utilized a meta-learning training strategy with sets of learnable balancing parameters to integrate BN and IN." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.788, + 0.843 + ], + "angle": 0, + "content": "Ensemble learning [4,25,44]. These methods exploit the relationship among multiple source domains by using diverse experts to extract domain-specific features from the corresponding domain. To avoid the large model size, Xu et al. [44]" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.413, + 0.115, + 0.728, + 0.129 + ], + "angle": 0, + "content": "Rethinking Normalization Layers for DG-ReID" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.177 + ], + "angle": 0, + "content": "proposed a lightweight model META, which leverages IN and introduces it into a global branch to pursue invariant features across domains." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.178, + 0.788, + 0.254 + ], + "angle": 0, + "content": "By introducing IN in combination with BN, normalization-based methods are simple and effective and can be combined with other techniques. However, when tested on unknown target domains, the distribution shifts between source and target domains significantly affect the fitting and scaling ability of normalization layers. Thus, leading to a decline in the model's generalization performance." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.278, + 0.381, + 0.296 + ], + "angle": 0, + "content": "3 Methodology" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.312, + 0.789, + 0.569 + ], + "angle": 0, + "content": "Inspired by the observation that the weight matrix tends to encapsulate identity information while domain-specific knowledge is embedded within the normalization layers [24, 40], we propose a novel ReNorm framework for tackling the domain shifts problem for DG-ReID which is rooted in normalization layers. As shown in Fig.2, ReNorm integrates two complementary normalization layers through two forward propagations for the same weight matrix. In each iteration, we sample a mini-batch from a source domain. In the first forward propagation, Remix Normalization (RN) combines IN and BN to ensure the discriminative feature extraction capability of the weight matrix. At the same time, the BN statistical parameters of the current source domain are stored in the corresponding buffer by Exponential Moving Average (EMA). During the second forward propagation, we treat the current input as a simulated target domain of the remaining source domains. Thus, Emulation Normalization (EN) can emulate the inference process of RN to guide the weight matrix to learn how to transfer the acquired feature extraction capability to an unseen domain. Meanwhile, we introduce Domain Frozen (DF) to stabilize the weight matrix by freezing the updates of affine parameters in normalization layers." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.592, + 0.362, + 0.607 + ], + "angle": 0, + "content": "3.1 Preliminary" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.619, + 0.788, + 0.723 + ], + "angle": 0, + "content": "In the DG ReID task, we have access to \\(K\\) source domains (datasets) for training. We denote the \\(k\\)-th source domain as \\(D_{k}, k \\in 1,2,\\dots,K\\), which have completely non-overlapping label spaces. Given a feature map \\(X_{k} \\in \\mathbb{R}^{N \\times C \\times H \\times W}\\) of a minibatch extracted from a certain domain \\(D_{k}\\), where \\(N,C,H\\), and \\(W\\) denote the batch size, the number of channels, the height, and the width, respectively. Following the definition from [41], we decompose the normalization layers into two steps as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.383, + 0.724, + 0.786, + 0.758 + ], + "angle": 0, + "content": "\\[\n\\hat {X} _ {k} = \\frac {X _ {k} - \\mu_ {k}}{\\sqrt {\\sigma_ {k} ^ {2} + \\epsilon}}, \\quad \\tilde {X} _ {k} = \\gamma \\hat {X} _ {k} + \\beta , \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.788, + 0.841 + ], + "angle": 0, + "content": "The first step involves standardization using the mean \\(\\mu_{k}\\) and variance \\(\\sigma_k^2\\) calculated from the input feature map. After that, channel-wise learnable affine parameters \\(\\gamma\\) and \\(\\beta\\) are defined for scaling and shifting during end-to-end training. \\(\\epsilon >0\\) is a constant to avoid numerical problems. The main distinction among various normalization techniques primarily lies in the first step." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.437, + 0.129 + ], + "angle": 0, + "content": "R. Nie and J. Ding et al." + }, + { + "type": "image", + "bbox": [ + 0.219, + 0.146, + 0.785, + 0.467 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.479, + 0.788, + 0.549 + ], + "angle": 0, + "content": "Fig. 2: The intra-layer structure of our ReNorm framework. The upper part corresponds to the first forward propagation with Remix Normalization, while the lower part corresponds to the second forward propagation with Emulation Normalization. After standardization, we employ Domain Frozen strategy for the affine parameters. Best viewed in color, detailed explanations can be found in Sec 3.2 and Sec 3.3." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.575, + 0.788, + 0.653 + ], + "angle": 0, + "content": "Instance Normalization [38] can remove style information from input features. So it has been widely used for the DG-ReID task to extract domain-invariant representations by removing their statistical contrast across feature channels. IN calculates mean \\(\\mu_k^{in} \\in \\mathbb{R}^{N \\times C}\\) and variance \\(\\sigma_k^{in^2} \\in \\mathbb{R}^{N \\times C}\\) with respect to each sample and each channel for standardization:" + }, + { + "type": "equation", + "bbox": [ + 0.307, + 0.658, + 0.786, + 0.695 + ], + "angle": 0, + "content": "\\[\n\\mu_ {k} ^ {i n} = \\frac {\\sum_ {h , w} ^ {H W} X _ {k}}{H W} \\quad \\text {a n d} \\quad \\sigma_ {k} ^ {i n 2} = \\frac {\\sum_ {h , w} ^ {H W} \\left(X _ {k} - \\mu_ {k} ^ {i n}\\right) ^ {2}}{H W}. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.7, + 0.788, + 0.761 + ], + "angle": 0, + "content": "Batch Normalization [16] is widely used in convolutional neural networks to capture discriminative features and mitigate gradient vanishing or exploding. It calculates mean \\(\\mu_k^{bn} \\in \\mathbb{R}^C\\) and variance \\(\\sigma_k^{bn^2} \\in \\mathbb{R}^C\\) with respect each channel in a mini-batch:" + }, + { + "type": "equation", + "bbox": [ + 0.294, + 0.766, + 0.786, + 0.803 + ], + "angle": 0, + "content": "\\[\n\\mu_ {k} ^ {b n} = \\frac {\\sum_ {n , h , w} ^ {N H W} X _ {k}}{N H W} \\quad \\text {a n d} \\quad \\sigma_ {k} ^ {b n ^ {2}} = \\frac {\\sum_ {n , h , w} ^ {N H W} \\left(X _ {k} - \\mu_ {k} ^ {b n}\\right) ^ {2}}{N H W}. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.809, + 0.788, + 0.842 + ], + "angle": 0, + "content": "It is worth noting that different from IN where the mean \\(\\mu^{in}\\) and variance \\(\\sigma^{in^2}\\) are computed from the input feature map \\(X\\) during both training and test-" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.413, + 0.114, + 0.728, + 0.129 + ], + "angle": 0, + "content": "Rethinking Normalization Layers for DG-ReID" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.787, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.222 + ], + "angle": 0, + "content": "ing. During testing, BN standardizes the input feature map \\( X \\) by population statistics which are estimated using Exponential Moving Average (EMA) during training. This special mechanism of BN enriched with domain-specific information of source domains poses a significant risk to the generalization performance of the model." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.245, + 0.441, + 0.259 + ], + "angle": 0, + "content": "3.2 Remix Normalization" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.27, + 0.789, + 0.409 + ], + "angle": 0, + "content": "Previous studies [15,31] have shown that Batch Normalization (BN) can extract domain-specific discriminative information, but it can easily overfit to a particular domain data in a mini-batch, while Instance Normalization (IN) can reduce the domain-specific style information, but it results in the loss of some discriminative information. Therefore, we combine these two normalization methods for a trade-off. Inspired by DSON [36] and SN [29], we design Remix Normalization (RN) which combines the statistical parameters of BN and IN in a simple way. Given an input feature map \\(X_{k}\\), RN calculates mean \\(\\mu_k^{rn} \\in \\mathbb{R}^{N \\times C}\\) and variance \\(\\sigma_k^{rn^2} \\in \\mathbb{R}^{N \\times C}\\) for standardization as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.425, + 0.419, + 0.786, + 0.446 + ], + "angle": 0, + "content": "\\[\n\\mu_ {k} ^ {r n} = \\left(\\mu_ {k} ^ {\\mathrm {i n}} + \\mu_ {k} ^ {\\mathrm {b n}}\\right) / 2,\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.416, + 0.44, + 0.589, + 0.46 + ], + "angle": 0, + "content": "\\[\n\\sigma_ {k} ^ {r n 2} = (\\sigma_ {k} ^ {i n 2} + \\sigma_ {k} ^ {b n 2}) / 2.\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.47, + 0.788, + 0.53 + ], + "angle": 0, + "content": "Meanwhile, unlike previous approaches where each normalization layer corresponds to one BN buffer, we set up corresponding \\( Buffer_{k}:\\mu_{k}^{pop},\\sigma_{k}^{pop2}\\in \\mathbb{R}^{C},k\\in 1,2,\\ldots ,K \\) for each RN for \\( K \\) source domains to store their domain-specific information as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.404, + 0.542, + 0.612, + 0.56 + ], + "angle": 0, + "content": "\\[\n\\mu_ {k} ^ {\\text {p o p}} \\leftarrow \\lambda \\mu_ {k} ^ {\\text {p o p}} + (1 - \\lambda) \\mu_ {k} ^ {b n},\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.393, + 0.555, + 0.785, + 0.583 + ], + "angle": 0, + "content": "\\[\n\\sigma_ {k} ^ {p o p 2} \\leftarrow \\lambda \\sigma_ {k} ^ {p o p 2} + (1 - \\lambda) \\sigma_ {k} ^ {b n 2}. \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.594, + 0.788, + 0.67 + ], + "angle": 0, + "content": "As shown in Fig.2, we use three primary colors (yellow, red, blue) to represent the corresponding source domains \\(D_{1}\\), \\(D_{2}\\), \\(D_{K}\\). The input mini-batch is from \\(D_{1}\\). After calculating the required IN and BN statistical parameters, we follow Eq.5 to update the corresponding BN population statistics (i.e., \\(Buffer_{1}:\\mu_{1}^{pop},\\sigma_{1}^{pop2}\\)). In Fig.2, we represent this update process with solid lines." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.671, + 0.789, + 0.792 + ], + "angle": 0, + "content": "The efficacy of combining BN and IN has been verified in many previous DG ReID works [2, 33, 34]. Nevertheless, a common challenge persists across these approaches: overfitting of the normalization layers to the source domains. Specifically, during the training process, following Eq.4, the statistical parameters of IN and BN within the normalization layer are derived from the same source domain, resulting in a consistent distribution. However, when the model is directly tested on the target domain, there is no concept of mini-batch anymore [41]. Thus, RN calculates the parameters for standardization as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.418, + 0.804, + 0.786, + 0.831 + ], + "angle": 0, + "content": "\\[\n\\mu_ {k} ^ {r n} = \\left(\\mu_ {t a r} ^ {\\text {i n}} + \\mu_ {k} ^ {\\text {p o p}}\\right) / 2, \\tag {6}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.411, + 0.825, + 0.594, + 0.843 + ], + "angle": 0, + "content": "\\[\n\\sigma_ {k} ^ {r n 2} = (\\sigma_ {t a r} ^ {i n 2} + \\sigma_ {k} ^ {p o p 2}) / 2,\n\\]" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.438, + 0.129 + ], + "angle": 0, + "content": "R. Nie and J. Ding et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.146, + 0.788, + 0.284 + ], + "angle": 0, + "content": "the IN statistical parameters (i.e., \\(\\mu_{tar}^{\\mathrm{in}},\\sigma_{tar}^{in2}\\)) are derived from the input target domain itself. Conversely, the BN statistical parameters are based on population statistics (i.e., \\(\\mu_k^{pop},\\sigma_k^{pop2}\\)) of source domains. The actual distribution between the target domain and the source domain is completely different, which leads to the inability of the statistics parameters (i.e., \\(\\mu_k^{rn},\\sigma_k^{rn2}\\)) to fit the input \\(X_{tar}\\) from the target domain. Concurrently, the model lacks the adaptability to address this mismatch, resulting in a performance decline. Furthermore, the learnable affine parameters also tend to overfit to the source domains, leading to a decrease in performance during both training and testing." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.302, + 0.471, + 0.316 + ], + "angle": 0, + "content": "3.3 Emulation Normalization" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.323, + 0.785, + 0.429 + ], + "angle": 0, + "content": "Remix Normalization enables the weight matrix in learning to extract discriminative features while also possessing some degree of domain invariance. However, a crucial limitation emerges when the model encounters the distribution shifts introduced by the unseen target domain, as it lacks the knowledge of how to adapt. Hence, a natural consideration arises: given the entirely distinct distributions among different source domains, we can treat the current source domain as a simulated target domain of a remaining source domain." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.43, + 0.787, + 0.52 + ], + "angle": 0, + "content": "Thus, we introduce a novel normalization operation named Emulation Normalization (EN), which emulates the inference procedure of Remix Normalization during training. By doing so, EN actively guides the weight matrix to acquire the capability of handling issues arising from the inability of RN to fit the input from unseen target domains. EN calculates mean \\(\\mu_k^{en} \\in \\mathbb{R}^{N \\times C}\\) and variance \\(\\sigma_k^{en2} \\in \\mathbb{R}^{N \\times C}\\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.385, + 0.527, + 0.785, + 0.57 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mu_ {k} ^ {e n} = \\left(1 - w _ {j} ^ {m}\\right) \\mu_ {k} ^ {\\mathrm {i n}} + w _ {j} ^ {m} \\mu_ {j} ^ {\\mathrm {p o p}} \\\\ \\sigma_ {k} ^ {e n 2} = \\left(1 - w _ {j} ^ {v}\\right) \\sigma_ {k} ^ {i n 2} + w _ {j} ^ {v} \\sigma_ {j} ^ {p o p 2}, \\tag {7} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.574, + 0.788, + 0.695 + ], + "angle": 0, + "content": "where we treat the current input source domain \\( D_{k} \\) as the simulated target domain of \\( D_{j} \\), which is randomly selected from the remaining. By combining IN statistical parameters of \\( X_{k} \\) from this simulated target domain \\( D_{k} \\) and BN population statistics from a remaining source domain \\( D_{j} \\), our EN can actively guide the weight matrix to transfer discriminative feature extraction from familiar source domains to unfamiliar target domain with different distribution. Notably, the learnable balancing ratios (i.e., \\( w_{j}^{m}, w_{j}^{v} \\)) correspond individually to the randomly selected source domain \\( D_{j} \\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.696, + 0.788, + 0.77 + ], + "angle": 0, + "content": "As shown in Fig.2, We use the current input \\( D_{1} \\) as the simulated target domain of \\( D_{2} \\) which supports the required BN statistical parameters stored in Buffer2 (i.e., \\( \\mu_2^{pop}, \\sigma_2^{pop^2} \\)), as specified in Eq.7. It is worth noting that Buffer2 could combine IN statistical parameters from \\( D_{1} \\) and \\( D_K \\), this pattern holds true for all the buffers." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.789, + 0.388, + 0.803 + ], + "angle": 0, + "content": "3.4 Domain Frozen" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Another factor that may lead to a decrease in the model's generalization capability lies in the learnable affine parameters of the normalization layers. Compared" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.414, + 0.115, + 0.728, + 0.129 + ], + "angle": 0, + "content": "Rethinking Normalization Layers for DG-ReID" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.785, + 0.221 + ], + "angle": 0, + "content": "to the weight matrix, the number of learnable affine parameters is significantly smaller, implying a weaker robustness when handling the target domain with a distribution divergence from the source domains. Moreover, ReID can be approximated as a channel-wise classification task, indicating that channel-wise affine parameters exert a more profound influence." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.223, + 0.787, + 0.298 + ], + "angle": 0, + "content": "Furthermore, as shown in Fig.3, we visualize the affine parameters for each source domain, approximating the domain-specific information within. It is evident that significant disparities exist among different source domains. These channel-wise divergences will accumulate into the weight matrix through backward propagation, leading to confusion in feature learning." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.298, + 0.788, + 0.434 + ], + "angle": 0, + "content": "Consequently, we propose Domain Frozen (DF), a strategy that stabilizes the weight matrix by freezing the updates of affine parameters in every normalization layer. For Remix Normalization, we maintain the affine parameters in a frozen state following their initialization with pre-trained weights. In the case of Emulation Normalization, the input feature map \\( X_{k} \\) is standardized by statistical parameters from two different domains. The output \\( \\hat{X}_k \\) inherently deviates from a normal distribution. Consequently, with the integration of Domain Frozen, we have opted to directly eliminate the original affine parameters, which is tantamount to setting the affine parameters \\( \\gamma \\) and \\( \\beta \\) to 1 and 0, respectively." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.434, + 0.788, + 0.509 + ], + "angle": 0, + "content": "Moreover, a further apparent advantage of DF is that it indirectly shares the affine parameters between the source domains (during training) and the target domain (during testing). This sharing further mitigates the domain shifts caused by the absence of target domain information, enhancing the model's ability to generalize to unseen domains." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.529, + 0.431, + 0.545 + ], + "angle": 0, + "content": "3.5 Training and testing" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.552, + 0.788, + 0.641 + ], + "angle": 0, + "content": "As illustrated in Fig.1(b), we propose a novel ReNorm framework to integrate the aforementioned normalization strategies. For clarification purposes, we refer to the weight matrix equipped with Remix Normalization (RN) as Network- \\(R\\) and the same weight matrix equipped with Emulation Normalization (EN) as Network- \\(E\\). Meanwhile, Domain Frozen (DF) is applied uniformly in both RN and EN." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.643, + 0.789, + 0.841 + ], + "angle": 0, + "content": "Model Training. In our methodology, there is no need for additional design of loss functions. The employment of two fundamental ReID loss functions, triplet loss \\((L_{tri})\\) [13] and cross-entropy loss \\((L_{ce})\\) suffices for achieving optimal performance. In each iteration, we randomly select one source domain \\(D_{k}\\) from a set of \\(K\\) source domains and randomly sample a mini-batch samples of size \\(N_{b}\\) from this domain. As illustrated in Fig.1, this set of data is then fed through the Network \\(R\\) for first forward propagation and calculate corresponding triplet loss \\(L_{tri}^{R}\\) and cross-entropy loss \\(L_{ce}^{R}\\), which is represented by the solid line. Subsequently, we treat the current source domain as a simulated target domain and randomly select one \\(D_{j}\\) from the remaining source domains to serve as a simulated source domain, providing Batch Normalization (BN) statistical parameters (i.e., \\(\\mu_j^{pop},\\sigma_j^{pop^2}\\)) for the second forward propagation, which is indicated by the dashed line. The same mini-batch will be fed into Network \\(E\\) to compute" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.438, + 0.129 + ], + "angle": 0, + "content": "R. Nie and J. Ding et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.146, + 0.784, + 0.192 + ], + "angle": 0, + "content": "the corresponding triplet loss \\( L_{tri}^{E} \\) and cross-entropy loss \\( L_{ce}^{E} \\). Combing these two networks mentioned above, we have the following total loss for backward propagation:" + }, + { + "type": "equation", + "bbox": [ + 0.388, + 0.192, + 0.785, + 0.209 + ], + "angle": 0, + "content": "\\[\nL _ {t o t a l} = L _ {t r i} ^ {R} + L _ {c e} ^ {R} + L _ {t r i} ^ {E} + L _ {c e} ^ {E}. \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.215, + 0.784, + 0.26 + ], + "angle": 0, + "content": "Testing. During testing, we feed the target domain to both networks \\( K \\) times each. Subsequently, we compute the average of the \\( 2*K \\) features obtained from two networks to obtain the final inference result." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.284, + 0.376, + 0.301 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.315, + 0.443, + 0.33 + ], + "angle": 0, + "content": "4.1 Datasets and Settings" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.34, + 0.784, + 0.43 + ], + "angle": 0, + "content": "Datesets. Our approach is evaluated on 9 person ReID datasets, including 5 large-scale datasets: Market1501 (M) [47], MSMT17 (MS) [39], CUHK02 (C2) [22], CUHK03 (C3) [23], and CUHK-SYSU (CS) [42], and 4 small-scale datasets: PRID [14], GRID [27], VIPeR [7], and iLIDs [48]. We do not use DukeMTMCreID (D) [50] since it has been retracted. We adopt Cumulative Matching Characteristics (CMC) and mean Average Precision (mAP) as evaluation metrics." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.431, + 0.785, + 0.596 + ], + "angle": 0, + "content": "Evaluation Protocols. We set three evaluation protocols following [37,44]. For Protocol-1, all images including both training and testing sets from four large-scale datasets (i.e., M, CS, C2, C3) are used for training, and the four small-scale datasets (i.e., PRID, GRID, VIPeR, iLIDs) are used for testing. The final results for each dataset are evaluated on the average of 10 repeated random splits of gallery and query sets. For Protocol-2 and Protocol-3, following the leave-one-out setting in [44], we select one domain from \\(\\mathrm{M} + \\mathrm{MS} + \\mathrm{CS} + \\mathrm{C3}\\) for testing, while the remaining three domains serve as the training sets. CS is excluded as a target domain due to its images all come from one camera view. Additionally, under Protocol-2, only the training sets of the source domains are used for training, while all images are exploited under Protocol-3." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.597, + 0.785, + 0.716 + ], + "angle": 0, + "content": "Implementation Details. We employ ResNet50 [9] pre-trained on ImageNet as our backbone. The size of a mini-batch is set to 64, including 16 identities and 4 images per identity. Each image is resized to \\(256 \\times 128\\). We use random horizontal flipping, random cropping, zero padding, color jittering, and auto-augmentation [11] as data augmentation. We also adopt a warm-up strategy in the first 10 epochs. The learning rate is initialized as \\(3.5 \\times 10^{-4}\\) and divided by 10 at the 30th, 60th and, 90th epochs, respectively. All experiments are conducted using Pytorch on a single RTX 4090 GPU." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.74, + 0.583, + 0.755 + ], + "angle": 0, + "content": "4.2 Comparison with the State-of-the-Arts" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.784, + 0.84 + ], + "angle": 0, + "content": "Comparison with DG-ReID Methods Under Protocol-1. As shown in Table 1, we compare our method with other state-of-the-arts under Protocol-1. Since DukeMTMC-reID (D) [50] has been retracted, we only use other large-scale datasets for training. Notably, our method demonstrates superior performance on PRID, GRID, and VIPeR. Overall, ReNorm significantly outperforms other" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.413, + 0.114, + 0.728, + 0.129 + ], + "angle": 0, + "content": "Rethinking Normalization Layers for DG-ReID" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.788, + 0.187 + ], + "angle": 0, + "content": "Table 1: Comparison with the SOTA methods in DG ReID under the setting of Protocol-1. All results are the average of 10 random sampling. The best and the second best results are highlighted in bold and underline." + }, + { + "type": "table", + "bbox": [ + 0.222, + 0.198, + 0.787, + 0.345 + ], + "angle": 0, + "content": "
MethodReferenceSource domainsTarget domainAverage
PRIDGRIDVIPeRiLIDs
mAPRank-1mAPRank-1mAPRank-1mAPRank-1mAPRank-1
SNR [19]CVPR'20M+D+C266.552.147.740.261.352.989.984.166.457.3
RaMoE [4]CVPR'2167.357.754.246.864.656.690.285.069.161.5
DMG-Net [1]CVPR'21+C3+CS68.460.656.651.060.453.983.979.367.361.2
DTIN-Net [18]ECCV'2279.771.060.651.870.762.987.281.874.666.9
MetaBIN [3]CVPR'21M+C270.861.257.950.264.355.982.774.768.960.5
META [44]ECCV'2271.761.960.152.468.461.583.579.268.960.5
ACL [45]ECCV'2273.463.065.755.275.166.486.581.875.266.6
CINorm [2]TMM'23+C3+CS80.572.459.949.476.767.688.082.276.267.9
PAOA [21]WACV'2474.065.667.256.376.666.787.183.176.267.9
ReNormThis paper81.373.967.559.477.069.187.082.278.271.2
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.375, + 0.784, + 0.421 + ], + "angle": 0, + "content": "methods by at least \\(2.0\\%\\) and \\(3.3\\%\\) in average mAP and Rank-1, respectively. The experimental findings substantiate the efficacy of our ReNorm, which combines simplicity with substantial impact." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.422, + 0.789, + 0.588 + ], + "angle": 0, + "content": "Comparison Under Protocol-2 and Protocol-3. We compare our method with previous state-of-the-arts under protocol-2 and protocol-3 to evaluate our method on large-scale datasets. As shown in Table 2, we can find that our ReNorm achieves superior performance compared with previous methods on almost all datasets under the two protocols. Overall, ReNorm outperforms other approaches by an average of at least \\(2.0\\%\\) mAP and \\(4.1\\%\\) Rank-1 under Protocol-2, and by \\(1.7\\%\\) mAP and \\(2.9\\%\\) Rank-1 under Protocol-3, respectively. Notably, when MSMT17 (MS) serves as the target domain, our method surpasses the second best by \\(3.1\\%\\) mAP and \\(5.7\\%\\) Rank-1 under Protocol-2, and by \\(2.7\\%\\) mAP and \\(5.7\\%\\) Rank-1 under Protocol-3. Collectively, these results further illustrate our model's superiority in domain generalization." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.614, + 0.39, + 0.63 + ], + "angle": 0, + "content": "4.3 Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.644, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Effectiveness of different modules. We investigate the effects of different components in ReNorm design to study their individual contributions, as shown in Table 3. The first row corresponds to a basic ResNet50 framework. In the table, every even row compared to the previous row incorporates the Domain Frozen (DF) strategy, which straightforwardly boosts the network by at least \\(1.3\\%\\) mAP and \\(1.5\\%\\) Rank-1, on average. Comparing first and second rows with the third and fourth rows, even when Remix Normalization (RN) simply combines IN and BN, it significantly enhances the network's generalization performance. Furthermore, with the addition of Emulation Normalization (EN), the two complementary normalization layers further improve the model's generalization ability, resulting in an average improvement of at least \\(2.0\\%\\) mAP and \\(2.5\\%\\) Rank-1. The experiments thoroughly demonstrate the simplicity and effectiveness of our approach." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.438, + 0.129 + ], + "angle": 0, + "content": "R. Nie and J. Ding et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.217, + 0.145, + 0.788, + 0.187 + ], + "angle": 0, + "content": "Table 2: Comparison with the SOTA methods in DG ReID under the setting of Protocol-2 and Protocol-3. All results are the average of 10 random sampling. The best and the second best results are highlighted in bold and underline." + }, + { + "type": "table", + "bbox": [ + 0.222, + 0.198, + 0.787, + 0.408 + ], + "angle": 0, + "content": "
MethodReferenceSettingM+MS+CS→C3M+CS+C3→MSMS+CS+C3→MAverage
mAP Rank-1mAP Rank-1mAP Rank-1mAP Rank-1mAP Rank-1mAP Rank-1mAP Rank-1mAP Rank-1
SNR [19]CVPR'20Protocol-28.98.96.819.934.662.716.830.5
M3L [46]CVPR'2120.931.915.936.958.479.931.749.6
MetaBIN [3]CVPR'2128.828.117.840.257.980.134.849.5
META [44]ECCV'2236.335.122.549.967.586.142.157.0
ACL [45]ECCV'2241.241.820.445.974.389.345.359.0
ReNormThis Paper43.644.725.655.672.789.147.363.1
SNR [19]CVPR'20Protocol-317.517.17.722.052.477.825.939.0
M3L [46]CVPR'2135.736.517.438.662.482.738.552.6
MetaBIN [3]CVPR'2143.043.118.841.267.284.543.056.3
META [44]ECCV'2247.146.224.452.176.590.549.362.9
ACL [45]ECCV'2249.950.121.747.376.890.649.362.7
PAOA [21]WACV'2449.850.525.151.577.190.850.764.3
ReNormThis Paper51.952.527.857.877.491.252.467.2
" + }, + { + "type": "table_caption", + "bbox": [ + 0.217, + 0.425, + 0.785, + 0.453 + ], + "angle": 0, + "content": "Table 3: Ablation studies for different components of our framework. The experiment is conducted under protocol-2. The best results are highlighted in bold." + }, + { + "type": "table", + "bbox": [ + 0.241, + 0.464, + 0.764, + 0.58 + ], + "angle": 0, + "content": "
RNENDFTarget: C3Target: MSTarget: MAverage
mAP Rank-1mAP Rank-1mAP Rank-1mAP Rank-1
35.334.919.944.966.385.140.555.0
✓36.536.922.148.568.286.642.357.3
✓40.540.522.149.469.087.143.959.0
✓✓40.440.223.752.771.988.945.360.6
✓✓42.543.624.453.471.187.946.061.6
✓✓✓43.644.725.655.672.789.147.363.1
" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.612, + 0.787, + 0.763 + ], + "angle": 0, + "content": "Visualization of the domain-specific information within the affine parameters. In this experiment, we establish corresponding RN for each source domain and reinstate the updates of all affine parameters. After training, we sum the affine parameters \\(\\gamma\\) of each source domain at each RN layer to approximate the domain-specific information within. As shown in Fig.3, The horizontal axis represents the layer position of RN in the backbone, while the vertical axis represents the corresponding proportions. It is evident that significant disparities exist among source domains across all the normalization layers. During training, the disparities accumulate in the weight matrix, potentially leading to misguided learning." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.765, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Effectiveness on where to plug the modules. As normalization-based modules, the proposed ReNorm and DF can be plugged into any stage of the network by replacing the corresponding BN layers. In our experiments, we employ ResNet50 as our backbone, which consists of four groups of residual blocks and is organized into four stages. We replace our ReNorm and DF with diverse stages'" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.414, + 0.115, + 0.728, + 0.129 + ], + "angle": 0, + "content": "Rethinking Normalization Layers for DG-ReID" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.149, + 0.778, + 0.284 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.299, + 0.785, + 0.34 + ], + "angle": 0, + "content": "Fig. 3: Visualization of the domain-specific information within the affine parameters under Protocol-2, the source domains used for training being MSMT17, CUHK-SYSU and Market1501." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.356, + 0.5, + 0.383 + ], + "angle": 0, + "content": "Table 4: Ablation study on where to plug the modules into ResNet50." + }, + { + "type": "table", + "bbox": [ + 0.233, + 0.395, + 0.487, + 0.51 + ], + "angle": 0, + "content": "
StageTarget: C3
Module: ReNormModule: DF
mAPRank-1mAPRank-1
None34.935.342.543.6
439.439.843.043.3
3 - 441.942.043.343.5
2 - 442.243.343.144.0
1 - 443.644.743.644.7
" + }, + { + "type": "table_caption", + "bbox": [ + 0.512, + 0.356, + 0.796, + 0.384 + ], + "angle": 0, + "content": "Table 5: Ablation study on different combination of outputs into the result." + }, + { + "type": "table", + "bbox": [ + 0.528, + 0.395, + 0.782, + 0.509 + ], + "angle": 0, + "content": "
MethodAverage
mAPRank-1
concat(R)46.261.8
concat(E)45.860.8
mean(R)46.361.9
mean(E)45.860.9
concat(R,E)47.062.8
mean(R,E)47.363.1
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.538, + 0.785, + 0.612 + ], + "angle": 0, + "content": "BN layers to explore the effects of different plugging stages. As shown in Table 4, the results demonstrate that employing ReNorm and DF across all normalization layers performs better than others. Additionally, incorporating ReNorm and DF from any stage could improve overall performance, further demonstrating the effectiveness of our modules." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.614, + 0.787, + 0.764 + ], + "angle": 0, + "content": "Combination of outputs into the final result. The experiment is conducted under Protocol-2. During testing, each image will go through Network-\\(R\\) (R) and Network-\\(E\\) (E) separately \\(K\\) times, resulting in a total of \\(2*K\\) outputs. As shown in Table 5, from the first four rows, it is evident that network-\\(R\\) generally outperforms network-\\(E\\), and both have surpassed the previous SOTA with \\(45.3\\%\\) mAP and \\(59.0\\%\\) Rank-1 on average. After combining the outputs of the two networks, there is a significant improvement in the final results. Ultimately, taking the mean of the \\(2*K\\) outputs directly yields the best result with \\(47.3\\%\\) mAP and \\(63.1\\%\\) Rank-1. Experimental results confirm the enhanced generalization capability of our framework in domain-invariant feature extraction." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Visualization of the target domain distribution. To further demonstrate the effectiveness of our ReNorm framework, we visualize the feature distribution of unseen target domains under Protocol-1. We conduct t-SNE [30] experiments on the same query and gallery image pairs from four target domains by implementing BN, IN, and our ReNorm, respectively. In Fig.4 (a), the" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.438, + 0.129 + ], + "angle": 0, + "content": "R. Nie and J. Ding et al." + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.144, + 0.391, + 0.274 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.283, + 0.279, + 0.326, + 0.291 + ], + "angle": 0, + "content": "(a) BN" + }, + { + "type": "image", + "bbox": [ + 0.419, + 0.144, + 0.588, + 0.274 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.482, + 0.279, + 0.523, + 0.291 + ], + "angle": 0, + "content": "(b) IN" + }, + { + "type": "image", + "bbox": [ + 0.616, + 0.144, + 0.785, + 0.274 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.645, + 0.279, + 0.754, + 0.291 + ], + "angle": 0, + "content": "(c) ReNorm (ours)" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.302, + 0.785, + 0.372 + ], + "angle": 0, + "content": "Fig. 4: The t-SNE results of features under Protocol-1 on three models with different normalization layers. (a) indicates ResNet50 with BN for all normalization layers, (b) denotes using instance normalization, and (c) utilizes our ReNorm for all normalization layers. Various colors illustrate samples from different target domains, and different shapes denote sample pairs from query or gallery sets." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.408, + 0.788, + 0.575 + ], + "angle": 0, + "content": "\"BN\" model exhibits smaller distances between query and gallery sets of the same identity but is heavily influenced by domain factors, leading to clear domain boundaries. In contrast, Fig.4 (b) shows that IN removes these boundaries but increases the distance between query and gallery sample pairs. The observations prove that BN can capture more discriminative features but tends to overfit to source domains, while IN can eliminate domain-specific information to enhance generalization but inevitably removes discriminative information. Fig.4 (c) shows that the distances between query and gallery sample pairs are much smaller, demonstrating that our approach achieves better performance. Simultaneously, there are no apparent boundaries among target domains, indicating that our method effectively eliminates domain-specific information." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.612, + 0.359, + 0.629 + ], + "angle": 0, + "content": "5 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.659, + 0.797, + 0.84 + ], + "angle": 0, + "content": "In this paper, we present a purely normalization-based framework called ReNorm for tackling the problem of domain generalizable person re-identification (DG-ReID). Our framework integrates two complementary normalization layers through two forward propagations for the same weight matrix. Remix Normalization (RN) enables the weight matrix to extract discriminative features within the source domains, while Emulation Normalization (EN) is employed to actively guide the weight matrix to learn how to generalize the feature extraction capability to unseen target domains. Meanwhile, we introduce Domain Frozen (DF), a strategy that stabilizes the weight matrix by freezing the updates of affine parameters in normalization layers. Extensive experiments demonstrate the simplicity yet state-of-the-art performance of ReNorm among all the popular benchmarks." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.413, + 0.115, + 0.728, + 0.129 + ], + "angle": 0, + "content": "Rethinking Normalization Layers for DG-ReID" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.403, + 0.163 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.178, + 0.798, + 0.254 + ], + "angle": 0, + "content": "This work was supported by the Natural Science Foundation of China (No. 62372082, 61972071, U20A20184), Natural Science Foundation of Sichuan Province (No.2023NSFSC0485), National Science Foundation for Distinguished Young Scholars under Grant 62225605 and Zhejiang Key Research and Development Program under Grant 2023C03196." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.279, + 0.323, + 0.294 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.311, + 0.787, + 0.367 + ], + "angle": 0, + "content": "1. Bai, Y., Jiao, J., Ce, W., Liu, J., Lou, Y., Feng, X., Duan, L.Y.: Person30k: A dual-meta generalization network for person re-identification. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2123-2132 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.368, + 0.787, + 0.409 + ], + "angle": 0, + "content": "2. Chen, Z., Wang, W., Zhao, Z., Su, F., Men, A., Dong, Y.: Cluster-instance normalization: A statistical relation-aware normalization for generalizable person re-identification. IEEE Transactions on Multimedia (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.41, + 0.787, + 0.451 + ], + "angle": 0, + "content": "3. Choi, S., Kim, T., Jeong, M., Park, H., Kim, C.: Meta batch-instance normalization for generalizable person re-identification. In: Proceedings of the IEEE/CVF conference on Computer Vision and Pattern Recognition. pp. 3425-3435 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.452, + 0.787, + 0.492 + ], + "angle": 0, + "content": "4. Dai, Y., Li, X., Liu, J., Tong, Z., Duan, L.Y.: Generalizable person re-identification with relevance-aware mixture of experts. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16145-16154 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.493, + 0.787, + 0.535 + ], + "angle": 0, + "content": "5. Dai, Y., Liu, J., Sun, Y., Tong, Z., Zhang, C., Duan, L.Y.: Idm: An intermediate domain module for domain adaptive person re-id. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 11864-11874 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.535, + 0.787, + 0.576 + ], + "angle": 0, + "content": "6. Fu, D., Chen, D., Bao, J., Yang, H., Yuan, L., Zhang, L., Li, H., Chen, D.: Unsupervised pre-training for person re-identification. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 14750-14759 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.577, + 0.787, + 0.604 + ], + "angle": 0, + "content": "7. Gray, D., Tao, H.: Viewpoint invariant pedestrian recognition with an ensemble of localized features. ECCV (1) 2008, 262-275 (2008)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.605, + 0.787, + 0.646 + ], + "angle": 0, + "content": "8. Han, K., Si, C., Huang, Y., Wang, L., Tan, T.: Generalizable person re-identification via self-supervised batch norm test-time adaption. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 36, pp. 817-825 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.647, + 0.787, + 0.687 + ], + "angle": 0, + "content": "9. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 770-778 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.688, + 0.787, + 0.729 + ], + "angle": 0, + "content": "0. He, L., Liang, J., Li, H., Sun, Z.: Deep spatial feature reconstruction for partial person re-identification: Alignment-free approach. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 7073-7082 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.729, + 0.787, + 0.757 + ], + "angle": 0, + "content": "1. He, L., Liao, X., Liu, W., Liu, X., Cheng, P., Mei, T.: Fastreid: A python toolbox for general instance re-identification. arXiv preprint arXiv:2006.02631 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.758, + 0.787, + 0.813 + ], + "angle": 0, + "content": "12. He, S., Luo, H., Chen, W., Zhang, M., Zhang, Y., Wang, F., Li, H., Jiang, W.: Multi-domain learning and identity mining for vehicle re-identification. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops. pp. 582-583 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.814, + 0.787, + 0.841 + ], + "angle": 0, + "content": "13. Hermans, A., Beyer, L., Leibe, B.: In defense of the triplet loss for person re-identification. arXiv preprint arXiv:1703.07737 (2017)" + }, + { + "type": "list", + "bbox": [ + 0.22, + 0.311, + 0.787, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.438, + 0.129 + ], + "angle": 0, + "content": "R. Nie and J. Ding et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.787, + 0.204 + ], + "angle": 0, + "content": "14. Hirzer, M., Beleznai, C., Roth, P.M., Bischof, H.: Person re-identification by descriptive and discriminative classification. In: Image Analysis: 17th Scandinavian Conference, SCIA 2011, Ystad, Sweden, May 2011. Proceedings 17. pp. 91-102. Springer (2011)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.205, + 0.787, + 0.245 + ], + "angle": 0, + "content": "15. Huang, X., Belongie, S.: Arbitrary style transfer in real-time with adaptive instance normalization. In: Proceedings of the IEEE international conference on computer vision. pp. 1501-1510 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.246, + 0.786, + 0.286 + ], + "angle": 0, + "content": "16. Ioffe, S., Szegedy, C.: Batch normalization: Accelerating deep network training by reducing internal covariate shift. In: International conference on machine learning. pp. 448-456. pmlr (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.288, + 0.787, + 0.327 + ], + "angle": 0, + "content": "17. Jiang, K., Zhang, T., Zhang, Y., Wu, F., Rui, Y.: Self-supervised agent learning for unsupervised cross-domain person re-identification. IEEE Transactions on Image Processing 29, 8549–8560 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.328, + 0.786, + 0.383 + ], + "angle": 0, + "content": "18. Jiao, B., Liu, L., Gao, L., Lin, G., Yang, L., Zhang, S., Wang, P., Zhang, Y.: Dynamically transformed instance normalization network for generalizable person re-identification. In: European Conference on Computer Vision. pp. 285–301. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.384, + 0.786, + 0.425 + ], + "angle": 0, + "content": "19. Jin, X., Lan, C., Zeng, W., Chen, Z., Zhang, L.: Style normalization and restitution for generalizable person re-identification. In: proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 3143-3152 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.426, + 0.787, + 0.48 + ], + "angle": 0, + "content": "20. Lee, G., Lee, S., Kim, D., Shin, Y., Yoon, Y., Ham, B.: Camera-driven representation learning for unsupervised domain adaptive person re-identification. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 11453-11462 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.481, + 0.787, + 0.522 + ], + "angle": 0, + "content": "21. Li, Q., Gong, S.: Mitigate domain shift by primary-auxiliary objectives association for generalizing person reid. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 394-403 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.523, + 0.787, + 0.563 + ], + "angle": 0, + "content": "22. Li, W., Wang, X.: Locally aligned feature transforms across views. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 3594-3601 (2013)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.565, + 0.787, + 0.605 + ], + "angle": 0, + "content": "23. Li, W., Zhao, R., Xiao, T., Wang, X.: Deepreid: Deep filter pairing neural network for person re-identification. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 152-159 (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.606, + 0.787, + 0.633 + ], + "angle": 0, + "content": "24. Li, Y., Wang, N., Shi, J., Hou, X., Liu, J.: Adaptive batch normalization for practical domain adaptation. Pattern Recognition 80, 109-117 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.634, + 0.787, + 0.687 + ], + "angle": 0, + "content": "25. Lin, C.S., Cheng, Y.C., Wang, Y.C.F.: Domain generalized person re-identification via cross-domain episodic learning. In: 2020 25th International Conference on Pattern Recognition (ICPR). pp. 6758-6763 (2021). https://doi.org/10.1109/ICPR48806.2021.9413013" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.689, + 0.787, + 0.73 + ], + "angle": 0, + "content": "26. Liu, J., Zha, Z.J., Chen, D., Hong, R., Wang, M.: Adaptive transfer network for cross-domain person re-identification. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 7202-7211 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.731, + 0.787, + 0.771 + ], + "angle": 0, + "content": "27. Loy, C.C., Xiang, T., Gong, S.: Time-delayed correlation analysis for multi-camera activity understanding. International Journal of Computer Vision 90, 106-129 (2010)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.772, + 0.787, + 0.812 + ], + "angle": 0, + "content": "28. Luo, H., Jiang, W., Gu, Y., Liu, F., Liao, X., Lai, S., Gu, J.: A strong baseline and batch normalization neck for deep person re-identification. IEEE Transactions on Multimedia 22(10), 2597-2609 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.813, + 0.787, + 0.84 + ], + "angle": 0, + "content": "29. Luo, P., Ren, J., Peng, Z., Zhang, R., Li, J.: Differentiable learning-to-normalize via switchable normalization. arXiv preprint arXiv:1806.10779 (2018)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.413, + 0.115, + 0.728, + 0.129 + ], + "angle": 0, + "content": "Rethinking Normalization Layers for DG-ReID" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.175 + ], + "angle": 0, + "content": "30. Van der Maaten, L., Hinton, G.: Visualizing data using t-sne. Journal of machine learning research 9(11) (2008)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.175, + 0.787, + 0.203 + ], + "angle": 0, + "content": "31. Nam, H., Kim, H.E.: Batch-instance normalization for adaptively style-invariant neural networks. Advances in Neural Information Processing Systems 31 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.203, + 0.787, + 0.245 + ], + "angle": 0, + "content": "32. Ni, H., Song, J., Luo, X., Zheng, F., Li, W., Shen, H.T.: Meta distribution alignment for generalizable person re-identification. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2487-2496 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.245, + 0.787, + 0.286 + ], + "angle": 0, + "content": "33. Pan, X., Luo, P., Shi, J., Tang, X.: Two at once: Enhancing learning and generalization capacities via ibn-net. In: Proceedings of the European Conference on Computer Vision (ECCV). pp. 464-479 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.286, + 0.787, + 0.327 + ], + "angle": 0, + "content": "34. Qi, L., Wang, L., Shi, Y., Geng, X.: A novel mix-normalization method for generalizable multi-source person re-identification. IEEE Transactions on Multimedia (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.327, + 0.787, + 0.37 + ], + "angle": 0, + "content": "35. Ren, M., He, L., Liao, X., Liu, W., Wang, Y., Tan, T.: Learning instance-level spatial-temporal patterns for person re-identification. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 14930-14939 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.37, + 0.787, + 0.425 + ], + "angle": 0, + "content": "36. Seo, S., Suh, Y., Kim, D., Kim, G., Han, J., Han, B.: Learning to optimize domain specific normalization for domain generalization. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXII 16. pp. 68-83. Springer (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.425, + 0.787, + 0.48 + ], + "angle": 0, + "content": "37. Song, J., Yang, Y., Song, Y.Z., Xiang, T., Hospedales, T.M.: Generalizable person re-identification by domain-invariant mapping network. In: Proceedings of the IEEE/CVF conference on Computer Vision and Pattern Recognition. pp. 719-728 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.48, + 0.787, + 0.509 + ], + "angle": 0, + "content": "38. Ulyanov, D., Vedaldi, A., Lempitsky, V.: Instance normalization: The missing ingredient for fast stylization. arXiv preprint arXiv:1607.08022 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.509, + 0.787, + 0.55 + ], + "angle": 0, + "content": "39. Wei, L., Zhang, S., Gao, W., Tian, Q.: Person transfer gan to bridge domain gap for person re-identification. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 79-88 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.55, + 0.787, + 0.591 + ], + "angle": 0, + "content": "40. Wu, Y., Chi, Z., Wang, Y., Plataniotis, K.N., Feng, S.: Test-time domain adaptation by learning domain-aware batch normalization. arXiv preprint arXiv:2312.10165 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.591, + 0.787, + 0.619 + ], + "angle": 0, + "content": "41. Wu, Y., Johnson, J.: Rethinking\" batch\" in batchnorm. arXiv preprint arXiv:2105.07576 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.619, + 0.787, + 0.647 + ], + "angle": 0, + "content": "42. Xiao, T., Li, S., Wang, B., Lin, L., Wang, X.: End-to-end deep learning for person search. arXiv preprint arXiv:1604.01850 2(2), 4 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.647, + 0.787, + 0.689 + ], + "angle": 0, + "content": "43. Xu, B., He, L., Liao, X., Liu, W., Sun, Z., Mei, T.: Black re-id: A head-shoulder descriptor for the challenging problem of person re-identification. In: Proceedings of the 28th ACM International Conference on Multimedia. pp. 673–681 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.689, + 0.787, + 0.744 + ], + "angle": 0, + "content": "44. Xu, B., Liang, J., He, L., Sun, Z.: Mimic embedding via adaptive aggregation: Learning generalizable person re-identification. In: Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XIV. pp. 372-388. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.744, + 0.787, + 0.785 + ], + "angle": 0, + "content": "45. Zhang, P., Dou, H., Yu, Y., Li, X.: Adaptive cross-domain learning for generalizable person re-identification. In: European Conference on Computer Vision. pp. 215-232. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.785, + 0.787, + 0.84 + ], + "angle": 0, + "content": "46. Zhao, Y., Zhong, Z., Yang, F., Luo, Z., Lin, Y., Li, S., Sebe, N.: Learning to generalize unseen domains via memory-based multi-source meta-learning for person re-identification. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6277-6286 (2021)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.438, + 0.129 + ], + "angle": 0, + "content": "R. Nie and J. Ding et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.189 + ], + "angle": 0, + "content": "47. Zheng, L., Shen, L., Tian, L., Wang, S., Wang, J., Tian, Q.: Scalable person re-identification: A benchmark. In: Proceedings of the IEEE international conference on computer vision. pp. 1116-1124 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.19, + 0.785, + 0.218 + ], + "angle": 0, + "content": "48. Zheng, W., Gong, S., Xiang, T.: Associating groups of people. In: British Machine Vision Conference (2009)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.218, + 0.785, + 0.259 + ], + "angle": 0, + "content": "49. Zheng, Z., Ruan, T., Wei, Y., Yang, Y., Mei, T.: VehicleNet: Learning robust visual representation for vehicle re-identification. IEEE Transactions on Multimedia 23, 2683-2693 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.259, + 0.785, + 0.301 + ], + "angle": 0, + "content": "50. Zheng, Z., Zheng, L., Yang, Y.: Unlabeled samples generated by gan improve the person re-identification baseline in vitro. In: Proceedings of the IEEE international conference on computer vision. pp. 3754-3762 (2017)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.301 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/Rethinking Normalization Layers for Domain Generalizable Person Re-identification/02192c3d-6e2f-475e-b7a0-eaf65cc6bd4c_origin.pdf b/2024/Rethinking Normalization Layers for Domain Generalizable Person Re-identification/02192c3d-6e2f-475e-b7a0-eaf65cc6bd4c_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..7239e04d3a14d2078bbfcff696b17622ef480436 --- /dev/null +++ b/2024/Rethinking Normalization Layers for Domain Generalizable Person Re-identification/02192c3d-6e2f-475e-b7a0-eaf65cc6bd4c_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e1dac3170b5c8d62d9c073457c014637fbbdcabeaace03370e7cb6456a23fd8f +size 1204314 diff --git a/2024/Rethinking Normalization Layers for Domain Generalizable Person Re-identification/full.md b/2024/Rethinking Normalization Layers for Domain Generalizable Person Re-identification/full.md new file mode 100644 index 0000000000000000000000000000000000000000..045b0151d50138c99e1484e8b7224815a8daa5b0 --- /dev/null +++ b/2024/Rethinking Normalization Layers for Domain Generalizable Person Re-identification/full.md @@ -0,0 +1,299 @@ +# Rethinking Normalization Layers for Domain Generalizable Person Re-identification + +Ren Nie $^{1}$ , Jin Ding $^{1}$ , Xue Zhou $^{2,1*}$ , and Xi Li $^{3}$ + +1 University of Electronic Science and Technology of China, Chengdu, China +nieren@std.uestc.edu.cn, dingjin1998@outlook.com, zhouxue@uestc.edu.cn + +2 Shenzhen Institute for Advanced Study, UESTC + +3 College of Computer Science & Technology, Zhejiang University xilizju@zju.edu.cn + +Abstract. Domain Generalizable Person Re-Identification (DG-ReID) strives to transfer learned feature representation from source domains to unseen target domains, despite significant distribution shifts. While most existing methods enhance model generalization and discriminative feature extraction capability by introducing Instance Normalization (IN) in combination with Batch Normalization (BN), these approaches still struggle with the overfitting of normalization layers to the source domains, posing challenges in domain generalization. To address this issue, we propose ReNorm, a purely normalization-based framework that integrates two complementary normalization layers through two forward propagations for the same weight matrix. In the first forward propagation, Remix Normalization (RN) combines IN and BN in a concise manner to ensure the feature extraction capability. As an effective complement to RN, Emulation Normalization (EN) simulates the testing process of RN, implicitly mitigating the domain shifts caused by the absence of target domain information and actively guiding the model in learning how to generalize the feature extraction capability to unseen target domains. Meanwhile, we propose Domain Frozen (DF), freezing updates to affine parameters to reduce the impact of these less robust parameters on overfitting to the source domains. Extensive experiments show that our framework achieves state-of-the-art performance on the popular benchmarks. The code is available at https://github.com/3699nr/ReNorm. + +Keywords: Normalization $\cdot$ Domain generalization $\cdot$ Re-identification + +# 1 Introduction + +Person Re-identification (ReID) is a practical task that has been widely applied in video surveillance and security. It aims at retrieving identical persons across a multi-camera network with non-overlapping fields of view. While current approaches [6,10,12,28,35,43,49] have achieved impressive performance within the same source domains during both training and testing phases. However, when + +![](images/986784d482117644a269c053cf0f6f4d1012444bd7c3ed8725ea9786404f98a6.jpg) +Fig. 1: Differences between existing normalization-based DG-ReID methods and ReNorm. To illustrate more clearly, we use the example of training with three source domains. (a) Existing normalization-based methods combine IN and BN in a specific way to ensure both the generalization and discrimination capability of the weight matrix. (b) ReNorm integrates two complementary normalization layers through two forward propagations for the same weight matrix. RN enables the weight matrix to extract discriminative features within the source domains, while EN is employed to actively guide the weight matrix to generalize the feature extraction capability to unseen target domains. (c) Some details of ReNorm. Detailed explanations can be found in Sec 3. + +directly deploying these well-trained models on an unseen target domain, an issue known as "domain shift" [39] arises, significantly impacting the performance of the models. To alleviate this problem, recent efforts have been devoted to two approaches: unsupervised domain adaptive (DA) ReID [5,17,20,26], and domain generalizable (DG) ReID [2,3,19,37,44,45]. In contrast to DA ReID, DG ReID is more practical and challenging as it trains on the multiple source domains and then directly transfers the learned representation to the unseen target domain without any target data for fine-tuning. + +Among numerous DG ReID methods, normalization-based approaches [2,3,8, 19,37] stand out for their simplicity and versatility, often integrating seamlessly with other advanced techniques such as meta-learning, distribution alignment, and ensemble learning. Illustrated in Fig.1(a), the combination of Batch Normalization (BN) [16] for capturing discriminative features and Instance Normalization (IN) [38] for eliminating domain-specific information facilitates normalization-based DG-ReID methods in achieving both high generalization and discrimination capabilities. During training, the statistical parameters of both IN and BN within the normalization layers are derived from the same data, ensuring a consistent distribution. However, this consistency is disrupted when the model is deployed on the target domain, as IN statistical parameters originate from the target domain itself, while BN statistical parameters are derived from population + +statistics accumulated from source domains during training. The distribution of the target domain significantly differs from that of the source domains, resulting in a marked decline in the model's performance. Furthermore, the vulnerable and sensitive learnable affine parameters are also prone to overfitting to the source domains, leading to performance degeneration in both training and testing phase. + +Moreover, emerging researches [24, 40] have shown that the identity and domain information are separately embedded in the weight matrix (learnable parameters in convolutional layers) and normalization layers. Consequently, we rethink the key challenges of DG ReID, which are rooted in normalization layers. We believe that the capability of the weight matrix to extract discriminative features sets the upper limit of the model's performance. Simultaneously, the normalization layers, responsible for fitting and scaling the input from target domains, establish the lower limit for the model's generalization performance. + +To address the issue of normalization layers being prone to overfitting to the source domains, we restructure the normalization layers, proposing ReNorm, a concise yet effective DG ReID framework entirely based on the design of normalization layers, as shown in Fig.1(b). In each iteration, the network undergoes two forward propagations and then one backpropagation after aggregating two forward losses. The two forward propagations apply different processing to the normalization layers, utilizing two complementary normalization layers to facilitate the weight matrix in extracting discriminative features from familiar source domains, while simultaneously guiding the same weight matrix to learn how to transfer this acquired feature extraction capability to unfamiliar target domains. + +Specifically, in our ReNorm framework, the two forward propagations correspond to two different but complementary normalization strategies: Remix Normalization (RN) and Emulation Normalization (EN). In the first forward propagation, RN introduces Instance Normalization (IN) in combination with Batch Normalization (BN) by directly summing their statistical parameters for standardization. Meanwhile, the BN statistical parameters of the current source domain are stored in the corresponding buffer, as shown in Fig.1(c). In the second forward propagation, EN adopts a similar summing approach but emulates the inference process of RN. It treats the current input as a simulated target domain, and calculates the IN statistical parameters accordingly. Simultaneously, the required BN statistical parameters are obtained from a remaining source domain's buffer stored in RN. This pairwise normalization strategy enables the network to actively learn how to transfer knowledge from source domains to the target domain with a distinct distribution. Additionally, we introduce Domain Frozen (DF), where the vulnerable settings of learnable affine parameters within both normalization layers are frozen and exempt from updates. By doing so, ReNorm greatly raises the lower limit of the model's generalization ability. + +We summarize our contributions as follows: + +- We introduce ReNorm, a simple yet effective framework for DG ReID, entirely and solely based on the design of normalization layers. This framework significantly mitigates performance degradation resulting from distribution shifts between source and target domains in a concise yet effective way. + +- We introduce two complementary normalization strategies (i.e., Remix Normalization and Emulation Normalization) equipped with Domain Frozen, actively enabling the network to learn how to transfer the feature extraction capability acquired from source domains to unseen target domains. + +- Extensive experiments demonstrate the simplicity yet state-of-the-art performance of our ReNorm framework under multiple testing protocols. + +# 2 Generalizable Person Re-Identification. + +Deep learning-based Person ReID approaches [6,10,12,28,35,43,49] have made great progress in recent years, achieving impressive accuracy in Person ReID. However, when these well-trained models are applied to unseen target domains that have distribution shifts with source domains, the performance decreases significantly. To alleviate this problem, recent efforts are devoted to two approaches: unsupervised domain adaptive (DA) ReID [5,17,20,26] and domain generalizable (DG) ReID. Diverging from the DA ReID, which needs unlabeled target domain datasets for unsupervised fine-tuning. DG ReID trains on the source domains and then directly tests on the unseen target domain. As a result, DG ReID has attracted attention for its robustness and generalization. + +Existing DG ReID methods can be briefly divided into four categories: + +Normalization strategies [2, 33, 34]. Batch Normalization (BN) can preserve diversity among samples, capturing more discriminative features and accelerating convergence. However, it tends to be prone to overfitting to source domains. Instance Normalization (IN) can eliminate domain-specific information to enhance generalization but inevitably removes discriminative information. Thus, normalization-based methods typically combine IN and BN in various ways to ensure both high generalization and discrimination capability of the network. Jiao et al. [18] proposed Dynamically Transformed Instance Normalization (DTIN) to alleviate the drawback of IN by transforming normalized features into appropriate representations in a learnable and adaptive manner. Chen et al. [2] designed a Cluster-Instance Normalization (CINorm) to eliminate the effect of outliers in original BN that causes severe overfitting. + +Domain alignment [19,32]. These methods try to minimize the difference among source domains for learning domain-invariant representations. They assume that the invariant features from the source domains should also be robust to any unseen target domains. Jin et al. [19] proposed a dual causality loss to align the outputs and a Style Normalization and Restitution (SNR) module to distill identity-relevant features via a restitution module for IN compensation. + +Meta-learning [3,37,46]. These methods improve generalized performance through virtual simulations of real train-test domain shifts. Following the concept of 'learning to lean', Choi et al. [3]proposed MetaBIN which utilized a meta-learning training strategy with sets of learnable balancing parameters to integrate BN and IN. + +Ensemble learning [4,25,44]. These methods exploit the relationship among multiple source domains by using diverse experts to extract domain-specific features from the corresponding domain. To avoid the large model size, Xu et al. [44] + +proposed a lightweight model META, which leverages IN and introduces it into a global branch to pursue invariant features across domains. + +By introducing IN in combination with BN, normalization-based methods are simple and effective and can be combined with other techniques. However, when tested on unknown target domains, the distribution shifts between source and target domains significantly affect the fitting and scaling ability of normalization layers. Thus, leading to a decline in the model's generalization performance. + +# 3 Methodology + +Inspired by the observation that the weight matrix tends to encapsulate identity information while domain-specific knowledge is embedded within the normalization layers [24, 40], we propose a novel ReNorm framework for tackling the domain shifts problem for DG-ReID which is rooted in normalization layers. As shown in Fig.2, ReNorm integrates two complementary normalization layers through two forward propagations for the same weight matrix. In each iteration, we sample a mini-batch from a source domain. In the first forward propagation, Remix Normalization (RN) combines IN and BN to ensure the discriminative feature extraction capability of the weight matrix. At the same time, the BN statistical parameters of the current source domain are stored in the corresponding buffer by Exponential Moving Average (EMA). During the second forward propagation, we treat the current input as a simulated target domain of the remaining source domains. Thus, Emulation Normalization (EN) can emulate the inference process of RN to guide the weight matrix to learn how to transfer the acquired feature extraction capability to an unseen domain. Meanwhile, we introduce Domain Frozen (DF) to stabilize the weight matrix by freezing the updates of affine parameters in normalization layers. + +# 3.1 Preliminary + +In the DG ReID task, we have access to $K$ source domains (datasets) for training. We denote the $k$ -th source domain as $D_{k}, k \in 1,2,\dots,K$ , which have completely non-overlapping label spaces. Given a feature map $X_{k} \in \mathbb{R}^{N \times C \times H \times W}$ of a minibatch extracted from a certain domain $D_{k}$ , where $N,C,H$ , and $W$ denote the batch size, the number of channels, the height, and the width, respectively. Following the definition from [41], we decompose the normalization layers into two steps as follows: + +$$ +\hat {X} _ {k} = \frac {X _ {k} - \mu_ {k}}{\sqrt {\sigma_ {k} ^ {2} + \epsilon}}, \quad \tilde {X} _ {k} = \gamma \hat {X} _ {k} + \beta , \tag {1} +$$ + +The first step involves standardization using the mean $\mu_{k}$ and variance $\sigma_k^2$ calculated from the input feature map. After that, channel-wise learnable affine parameters $\gamma$ and $\beta$ are defined for scaling and shifting during end-to-end training. $\epsilon >0$ is a constant to avoid numerical problems. The main distinction among various normalization techniques primarily lies in the first step. + +![](images/05a152d63138a59f14f1ae9a7fc5ed43aaa261a9becef2a98bde9879893dfb50.jpg) +Fig. 2: The intra-layer structure of our ReNorm framework. The upper part corresponds to the first forward propagation with Remix Normalization, while the lower part corresponds to the second forward propagation with Emulation Normalization. After standardization, we employ Domain Frozen strategy for the affine parameters. Best viewed in color, detailed explanations can be found in Sec 3.2 and Sec 3.3. + +Instance Normalization [38] can remove style information from input features. So it has been widely used for the DG-ReID task to extract domain-invariant representations by removing their statistical contrast across feature channels. IN calculates mean $\mu_k^{in} \in \mathbb{R}^{N \times C}$ and variance $\sigma_k^{in^2} \in \mathbb{R}^{N \times C}$ with respect to each sample and each channel for standardization: + +$$ +\mu_ {k} ^ {i n} = \frac {\sum_ {h , w} ^ {H W} X _ {k}}{H W} \quad \text {a n d} \quad \sigma_ {k} ^ {i n 2} = \frac {\sum_ {h , w} ^ {H W} \left(X _ {k} - \mu_ {k} ^ {i n}\right) ^ {2}}{H W}. \tag {2} +$$ + +Batch Normalization [16] is widely used in convolutional neural networks to capture discriminative features and mitigate gradient vanishing or exploding. It calculates mean $\mu_k^{bn} \in \mathbb{R}^C$ and variance $\sigma_k^{bn^2} \in \mathbb{R}^C$ with respect each channel in a mini-batch: + +$$ +\mu_ {k} ^ {b n} = \frac {\sum_ {n , h , w} ^ {N H W} X _ {k}}{N H W} \quad \text {a n d} \quad \sigma_ {k} ^ {b n ^ {2}} = \frac {\sum_ {n , h , w} ^ {N H W} \left(X _ {k} - \mu_ {k} ^ {b n}\right) ^ {2}}{N H W}. \tag {3} +$$ + +It is worth noting that different from IN where the mean $\mu^{in}$ and variance $\sigma^{in^2}$ are computed from the input feature map $X$ during both training and test- + +ing. During testing, BN standardizes the input feature map $X$ by population statistics which are estimated using Exponential Moving Average (EMA) during training. This special mechanism of BN enriched with domain-specific information of source domains poses a significant risk to the generalization performance of the model. + +# 3.2 Remix Normalization + +Previous studies [15,31] have shown that Batch Normalization (BN) can extract domain-specific discriminative information, but it can easily overfit to a particular domain data in a mini-batch, while Instance Normalization (IN) can reduce the domain-specific style information, but it results in the loss of some discriminative information. Therefore, we combine these two normalization methods for a trade-off. Inspired by DSON [36] and SN [29], we design Remix Normalization (RN) which combines the statistical parameters of BN and IN in a simple way. Given an input feature map $X_{k}$ , RN calculates mean $\mu_k^{rn} \in \mathbb{R}^{N \times C}$ and variance $\sigma_k^{rn^2} \in \mathbb{R}^{N \times C}$ for standardization as follows: + +$$ +\mu_ {k} ^ {r n} = \left(\mu_ {k} ^ {\mathrm {i n}} + \mu_ {k} ^ {\mathrm {b n}}\right) / 2, +$$ + +$$ +\sigma_ {k} ^ {r n 2} = (\sigma_ {k} ^ {i n 2} + \sigma_ {k} ^ {b n 2}) / 2. +$$ + +Meanwhile, unlike previous approaches where each normalization layer corresponds to one BN buffer, we set up corresponding $Buffer_{k}:\mu_{k}^{pop},\sigma_{k}^{pop2}\in \mathbb{R}^{C},k\in 1,2,\ldots ,K$ for each RN for $K$ source domains to store their domain-specific information as follows: + +$$ +\mu_ {k} ^ {\text {p o p}} \leftarrow \lambda \mu_ {k} ^ {\text {p o p}} + (1 - \lambda) \mu_ {k} ^ {b n}, +$$ + +$$ +\sigma_ {k} ^ {p o p 2} \leftarrow \lambda \sigma_ {k} ^ {p o p 2} + (1 - \lambda) \sigma_ {k} ^ {b n 2}. \tag {5} +$$ + +As shown in Fig.2, we use three primary colors (yellow, red, blue) to represent the corresponding source domains $D_{1}$ , $D_{2}$ , $D_{K}$ . The input mini-batch is from $D_{1}$ . After calculating the required IN and BN statistical parameters, we follow Eq.5 to update the corresponding BN population statistics (i.e., $Buffer_{1}:\mu_{1}^{pop},\sigma_{1}^{pop2}$ ). In Fig.2, we represent this update process with solid lines. + +The efficacy of combining BN and IN has been verified in many previous DG ReID works [2, 33, 34]. Nevertheless, a common challenge persists across these approaches: overfitting of the normalization layers to the source domains. Specifically, during the training process, following Eq.4, the statistical parameters of IN and BN within the normalization layer are derived from the same source domain, resulting in a consistent distribution. However, when the model is directly tested on the target domain, there is no concept of mini-batch anymore [41]. Thus, RN calculates the parameters for standardization as follows: + +$$ +\mu_ {k} ^ {r n} = \left(\mu_ {t a r} ^ {\text {i n}} + \mu_ {k} ^ {\text {p o p}}\right) / 2, \tag {6} +$$ + +$$ +\sigma_ {k} ^ {r n 2} = (\sigma_ {t a r} ^ {i n 2} + \sigma_ {k} ^ {p o p 2}) / 2, +$$ + +the IN statistical parameters (i.e., $\mu_{tar}^{\mathrm{in}},\sigma_{tar}^{in2}$ ) are derived from the input target domain itself. Conversely, the BN statistical parameters are based on population statistics (i.e., $\mu_k^{pop},\sigma_k^{pop2}$ ) of source domains. The actual distribution between the target domain and the source domain is completely different, which leads to the inability of the statistics parameters (i.e., $\mu_k^{rn},\sigma_k^{rn2}$ ) to fit the input $X_{tar}$ from the target domain. Concurrently, the model lacks the adaptability to address this mismatch, resulting in a performance decline. Furthermore, the learnable affine parameters also tend to overfit to the source domains, leading to a decrease in performance during both training and testing. + +# 3.3 Emulation Normalization + +Remix Normalization enables the weight matrix in learning to extract discriminative features while also possessing some degree of domain invariance. However, a crucial limitation emerges when the model encounters the distribution shifts introduced by the unseen target domain, as it lacks the knowledge of how to adapt. Hence, a natural consideration arises: given the entirely distinct distributions among different source domains, we can treat the current source domain as a simulated target domain of a remaining source domain. + +Thus, we introduce a novel normalization operation named Emulation Normalization (EN), which emulates the inference procedure of Remix Normalization during training. By doing so, EN actively guides the weight matrix to acquire the capability of handling issues arising from the inability of RN to fit the input from unseen target domains. EN calculates mean $\mu_k^{en} \in \mathbb{R}^{N \times C}$ and variance $\sigma_k^{en2} \in \mathbb{R}^{N \times C}$ as follows: + +$$ +\begin{array}{l} \mu_ {k} ^ {e n} = \left(1 - w _ {j} ^ {m}\right) \mu_ {k} ^ {\mathrm {i n}} + w _ {j} ^ {m} \mu_ {j} ^ {\mathrm {p o p}} \\ \sigma_ {k} ^ {e n 2} = \left(1 - w _ {j} ^ {v}\right) \sigma_ {k} ^ {i n 2} + w _ {j} ^ {v} \sigma_ {j} ^ {p o p 2}, \tag {7} \\ \end{array} +$$ + +where we treat the current input source domain $D_{k}$ as the simulated target domain of $D_{j}$ , which is randomly selected from the remaining. By combining IN statistical parameters of $X_{k}$ from this simulated target domain $D_{k}$ and BN population statistics from a remaining source domain $D_{j}$ , our EN can actively guide the weight matrix to transfer discriminative feature extraction from familiar source domains to unfamiliar target domain with different distribution. Notably, the learnable balancing ratios (i.e., $w_{j}^{m}, w_{j}^{v}$ ) correspond individually to the randomly selected source domain $D_{j}$ . + +As shown in Fig.2, We use the current input $D_{1}$ as the simulated target domain of $D_{2}$ which supports the required BN statistical parameters stored in Buffer2 (i.e., $\mu_2^{pop}, \sigma_2^{pop^2}$ ), as specified in Eq.7. It is worth noting that Buffer2 could combine IN statistical parameters from $D_{1}$ and $D_K$ , this pattern holds true for all the buffers. + +# 3.4 Domain Frozen + +Another factor that may lead to a decrease in the model's generalization capability lies in the learnable affine parameters of the normalization layers. Compared + +to the weight matrix, the number of learnable affine parameters is significantly smaller, implying a weaker robustness when handling the target domain with a distribution divergence from the source domains. Moreover, ReID can be approximated as a channel-wise classification task, indicating that channel-wise affine parameters exert a more profound influence. + +Furthermore, as shown in Fig.3, we visualize the affine parameters for each source domain, approximating the domain-specific information within. It is evident that significant disparities exist among different source domains. These channel-wise divergences will accumulate into the weight matrix through backward propagation, leading to confusion in feature learning. + +Consequently, we propose Domain Frozen (DF), a strategy that stabilizes the weight matrix by freezing the updates of affine parameters in every normalization layer. For Remix Normalization, we maintain the affine parameters in a frozen state following their initialization with pre-trained weights. In the case of Emulation Normalization, the input feature map $X_{k}$ is standardized by statistical parameters from two different domains. The output $\hat{X}_k$ inherently deviates from a normal distribution. Consequently, with the integration of Domain Frozen, we have opted to directly eliminate the original affine parameters, which is tantamount to setting the affine parameters $\gamma$ and $\beta$ to 1 and 0, respectively. + +Moreover, a further apparent advantage of DF is that it indirectly shares the affine parameters between the source domains (during training) and the target domain (during testing). This sharing further mitigates the domain shifts caused by the absence of target domain information, enhancing the model's ability to generalize to unseen domains. + +# 3.5 Training and testing + +As illustrated in Fig.1(b), we propose a novel ReNorm framework to integrate the aforementioned normalization strategies. For clarification purposes, we refer to the weight matrix equipped with Remix Normalization (RN) as Network- $R$ and the same weight matrix equipped with Emulation Normalization (EN) as Network- $E$ . Meanwhile, Domain Frozen (DF) is applied uniformly in both RN and EN. + +Model Training. In our methodology, there is no need for additional design of loss functions. The employment of two fundamental ReID loss functions, triplet loss $(L_{tri})$ [13] and cross-entropy loss $(L_{ce})$ suffices for achieving optimal performance. In each iteration, we randomly select one source domain $D_{k}$ from a set of $K$ source domains and randomly sample a mini-batch samples of size $N_{b}$ from this domain. As illustrated in Fig.1, this set of data is then fed through the Network $R$ for first forward propagation and calculate corresponding triplet loss $L_{tri}^{R}$ and cross-entropy loss $L_{ce}^{R}$ , which is represented by the solid line. Subsequently, we treat the current source domain as a simulated target domain and randomly select one $D_{j}$ from the remaining source domains to serve as a simulated source domain, providing Batch Normalization (BN) statistical parameters (i.e., $\mu_j^{pop},\sigma_j^{pop^2}$ ) for the second forward propagation, which is indicated by the dashed line. The same mini-batch will be fed into Network $E$ to compute + +the corresponding triplet loss $L_{tri}^{E}$ and cross-entropy loss $L_{ce}^{E}$ . Combing these two networks mentioned above, we have the following total loss for backward propagation: + +$$ +L _ {t o t a l} = L _ {t r i} ^ {R} + L _ {c e} ^ {R} + L _ {t r i} ^ {E} + L _ {c e} ^ {E}. \tag {8} +$$ + +Testing. During testing, we feed the target domain to both networks $K$ times each. Subsequently, we compute the average of the $2*K$ features obtained from two networks to obtain the final inference result. + +# 4 Experiments + +# 4.1 Datasets and Settings + +Datesets. Our approach is evaluated on 9 person ReID datasets, including 5 large-scale datasets: Market1501 (M) [47], MSMT17 (MS) [39], CUHK02 (C2) [22], CUHK03 (C3) [23], and CUHK-SYSU (CS) [42], and 4 small-scale datasets: PRID [14], GRID [27], VIPeR [7], and iLIDs [48]. We do not use DukeMTMCreID (D) [50] since it has been retracted. We adopt Cumulative Matching Characteristics (CMC) and mean Average Precision (mAP) as evaluation metrics. + +Evaluation Protocols. We set three evaluation protocols following [37,44]. For Protocol-1, all images including both training and testing sets from four large-scale datasets (i.e., M, CS, C2, C3) are used for training, and the four small-scale datasets (i.e., PRID, GRID, VIPeR, iLIDs) are used for testing. The final results for each dataset are evaluated on the average of 10 repeated random splits of gallery and query sets. For Protocol-2 and Protocol-3, following the leave-one-out setting in [44], we select one domain from $\mathrm{M} + \mathrm{MS} + \mathrm{CS} + \mathrm{C3}$ for testing, while the remaining three domains serve as the training sets. CS is excluded as a target domain due to its images all come from one camera view. Additionally, under Protocol-2, only the training sets of the source domains are used for training, while all images are exploited under Protocol-3. + +Implementation Details. We employ ResNet50 [9] pre-trained on ImageNet as our backbone. The size of a mini-batch is set to 64, including 16 identities and 4 images per identity. Each image is resized to $256 \times 128$ . We use random horizontal flipping, random cropping, zero padding, color jittering, and auto-augmentation [11] as data augmentation. We also adopt a warm-up strategy in the first 10 epochs. The learning rate is initialized as $3.5 \times 10^{-4}$ and divided by 10 at the 30th, 60th and, 90th epochs, respectively. All experiments are conducted using Pytorch on a single RTX 4090 GPU. + +# 4.2 Comparison with the State-of-the-Arts + +Comparison with DG-ReID Methods Under Protocol-1. As shown in Table 1, we compare our method with other state-of-the-arts under Protocol-1. Since DukeMTMC-reID (D) [50] has been retracted, we only use other large-scale datasets for training. Notably, our method demonstrates superior performance on PRID, GRID, and VIPeR. Overall, ReNorm significantly outperforms other + +Table 1: Comparison with the SOTA methods in DG ReID under the setting of Protocol-1. All results are the average of 10 random sampling. The best and the second best results are highlighted in bold and underline. + +
MethodReferenceSource domainsTarget domainAverage
PRIDGRIDVIPeRiLIDs
mAPRank-1mAPRank-1mAPRank-1mAPRank-1mAPRank-1
SNR [19]CVPR'20M+D+C266.552.147.740.261.352.989.984.166.457.3
RaMoE [4]CVPR'2167.357.754.246.864.656.690.285.069.161.5
DMG-Net [1]CVPR'21+C3+CS68.460.656.651.060.453.983.979.367.361.2
DTIN-Net [18]ECCV'2279.771.060.651.870.762.987.281.874.666.9
MetaBIN [3]CVPR'21M+C270.861.257.950.264.355.982.774.768.960.5
META [44]ECCV'2271.761.960.152.468.461.583.579.268.960.5
ACL [45]ECCV'2273.463.065.755.275.166.486.581.875.266.6
CINorm [2]TMM'23+C3+CS80.572.459.949.476.767.688.082.276.267.9
PAOA [21]WACV'2474.065.667.256.376.666.787.183.176.267.9
ReNormThis paper81.373.967.559.477.069.187.082.278.271.2
+ +methods by at least $2.0\%$ and $3.3\%$ in average mAP and Rank-1, respectively. The experimental findings substantiate the efficacy of our ReNorm, which combines simplicity with substantial impact. + +Comparison Under Protocol-2 and Protocol-3. We compare our method with previous state-of-the-arts under protocol-2 and protocol-3 to evaluate our method on large-scale datasets. As shown in Table 2, we can find that our ReNorm achieves superior performance compared with previous methods on almost all datasets under the two protocols. Overall, ReNorm outperforms other approaches by an average of at least $2.0\%$ mAP and $4.1\%$ Rank-1 under Protocol-2, and by $1.7\%$ mAP and $2.9\%$ Rank-1 under Protocol-3, respectively. Notably, when MSMT17 (MS) serves as the target domain, our method surpasses the second best by $3.1\%$ mAP and $5.7\%$ Rank-1 under Protocol-2, and by $2.7\%$ mAP and $5.7\%$ Rank-1 under Protocol-3. Collectively, these results further illustrate our model's superiority in domain generalization. + +# 4.3 Ablation Study + +Effectiveness of different modules. We investigate the effects of different components in ReNorm design to study their individual contributions, as shown in Table 3. The first row corresponds to a basic ResNet50 framework. In the table, every even row compared to the previous row incorporates the Domain Frozen (DF) strategy, which straightforwardly boosts the network by at least $1.3\%$ mAP and $1.5\%$ Rank-1, on average. Comparing first and second rows with the third and fourth rows, even when Remix Normalization (RN) simply combines IN and BN, it significantly enhances the network's generalization performance. Furthermore, with the addition of Emulation Normalization (EN), the two complementary normalization layers further improve the model's generalization ability, resulting in an average improvement of at least $2.0\%$ mAP and $2.5\%$ Rank-1. The experiments thoroughly demonstrate the simplicity and effectiveness of our approach. + +Table 2: Comparison with the SOTA methods in DG ReID under the setting of Protocol-2 and Protocol-3. All results are the average of 10 random sampling. The best and the second best results are highlighted in bold and underline. + +
MethodReferenceSettingM+MS+CS→C3M+CS+C3→MSMS+CS+C3→MAverage
mAP Rank-1mAP Rank-1mAP Rank-1mAP Rank-1mAP Rank-1mAP Rank-1mAP Rank-1mAP Rank-1
SNR [19]CVPR'20Protocol-28.98.96.819.934.662.716.830.5
M3L [46]CVPR'2120.931.915.936.958.479.931.749.6
MetaBIN [3]CVPR'2128.828.117.840.257.980.134.849.5
META [44]ECCV'2236.335.122.549.967.586.142.157.0
ACL [45]ECCV'2241.241.820.445.974.389.345.359.0
ReNormThis Paper43.644.725.655.672.789.147.363.1
SNR [19]CVPR'20Protocol-317.517.17.722.052.477.825.939.0
M3L [46]CVPR'2135.736.517.438.662.482.738.552.6
MetaBIN [3]CVPR'2143.043.118.841.267.284.543.056.3
META [44]ECCV'2247.146.224.452.176.590.549.362.9
ACL [45]ECCV'2249.950.121.747.376.890.649.362.7
PAOA [21]WACV'2449.850.525.151.577.190.850.764.3
ReNormThis Paper51.952.527.857.877.491.252.467.2
+ +Table 3: Ablation studies for different components of our framework. The experiment is conducted under protocol-2. The best results are highlighted in bold. + +
RNENDFTarget: C3Target: MSTarget: MAverage
mAP Rank-1mAP Rank-1mAP Rank-1mAP Rank-1
35.334.919.944.966.385.140.555.0
✓36.536.922.148.568.286.642.357.3
✓40.540.522.149.469.087.143.959.0
✓✓40.440.223.752.771.988.945.360.6
✓✓42.543.624.453.471.187.946.061.6
✓✓✓43.644.725.655.672.789.147.363.1
+ +Visualization of the domain-specific information within the affine parameters. In this experiment, we establish corresponding RN for each source domain and reinstate the updates of all affine parameters. After training, we sum the affine parameters $\gamma$ of each source domain at each RN layer to approximate the domain-specific information within. As shown in Fig.3, The horizontal axis represents the layer position of RN in the backbone, while the vertical axis represents the corresponding proportions. It is evident that significant disparities exist among source domains across all the normalization layers. During training, the disparities accumulate in the weight matrix, potentially leading to misguided learning. + +Effectiveness on where to plug the modules. As normalization-based modules, the proposed ReNorm and DF can be plugged into any stage of the network by replacing the corresponding BN layers. In our experiments, we employ ResNet50 as our backbone, which consists of four groups of residual blocks and is organized into four stages. We replace our ReNorm and DF with diverse stages' + +![](images/8bcb600a2b32df48b6e33eff9ebfba688aac5dd196b5f5ceb4afb567a5787cdc.jpg) +Fig. 3: Visualization of the domain-specific information within the affine parameters under Protocol-2, the source domains used for training being MSMT17, CUHK-SYSU and Market1501. + +Table 4: Ablation study on where to plug the modules into ResNet50. + +
StageTarget: C3
Module: ReNormModule: DF
mAPRank-1mAPRank-1
None34.935.342.543.6
439.439.843.043.3
3 - 441.942.043.343.5
2 - 442.243.343.144.0
1 - 443.644.743.644.7
+ +Table 5: Ablation study on different combination of outputs into the result. + +
MethodAverage
mAPRank-1
concat(R)46.261.8
concat(E)45.860.8
mean(R)46.361.9
mean(E)45.860.9
concat(R,E)47.062.8
mean(R,E)47.363.1
+ +BN layers to explore the effects of different plugging stages. As shown in Table 4, the results demonstrate that employing ReNorm and DF across all normalization layers performs better than others. Additionally, incorporating ReNorm and DF from any stage could improve overall performance, further demonstrating the effectiveness of our modules. + +Combination of outputs into the final result. The experiment is conducted under Protocol-2. During testing, each image will go through Network- $R$ (R) and Network- $E$ (E) separately $K$ times, resulting in a total of $2*K$ outputs. As shown in Table 5, from the first four rows, it is evident that network- $R$ generally outperforms network- $E$ , and both have surpassed the previous SOTA with $45.3\%$ mAP and $59.0\%$ Rank-1 on average. After combining the outputs of the two networks, there is a significant improvement in the final results. Ultimately, taking the mean of the $2*K$ outputs directly yields the best result with $47.3\%$ mAP and $63.1\%$ Rank-1. Experimental results confirm the enhanced generalization capability of our framework in domain-invariant feature extraction. + +Visualization of the target domain distribution. To further demonstrate the effectiveness of our ReNorm framework, we visualize the feature distribution of unseen target domains under Protocol-1. We conduct t-SNE [30] experiments on the same query and gallery image pairs from four target domains by implementing BN, IN, and our ReNorm, respectively. In Fig.4 (a), the + +![](images/a2237fb5379cbc87bf9176ae98899e168605bf1247f29f20f597f214b58ca8b3.jpg) +(a) BN + +![](images/f3105fdb03b99fb47ff956e2397506bf429106bf2710ad4846e04b12f4777628.jpg) +(b) IN +Fig. 4: The t-SNE results of features under Protocol-1 on three models with different normalization layers. (a) indicates ResNet50 with BN for all normalization layers, (b) denotes using instance normalization, and (c) utilizes our ReNorm for all normalization layers. Various colors illustrate samples from different target domains, and different shapes denote sample pairs from query or gallery sets. + +![](images/a181018657d9b21f620839e461fbcdaf07355feb381fac66458922e675d8a97f.jpg) +(c) ReNorm (ours) + +"BN" model exhibits smaller distances between query and gallery sets of the same identity but is heavily influenced by domain factors, leading to clear domain boundaries. In contrast, Fig.4 (b) shows that IN removes these boundaries but increases the distance between query and gallery sample pairs. The observations prove that BN can capture more discriminative features but tends to overfit to source domains, while IN can eliminate domain-specific information to enhance generalization but inevitably removes discriminative information. Fig.4 (c) shows that the distances between query and gallery sample pairs are much smaller, demonstrating that our approach achieves better performance. Simultaneously, there are no apparent boundaries among target domains, indicating that our method effectively eliminates domain-specific information. + +# 5 Conclusion + +In this paper, we present a purely normalization-based framework called ReNorm for tackling the problem of domain generalizable person re-identification (DG-ReID). Our framework integrates two complementary normalization layers through two forward propagations for the same weight matrix. Remix Normalization (RN) enables the weight matrix to extract discriminative features within the source domains, while Emulation Normalization (EN) is employed to actively guide the weight matrix to learn how to generalize the feature extraction capability to unseen target domains. Meanwhile, we introduce Domain Frozen (DF), a strategy that stabilizes the weight matrix by freezing the updates of affine parameters in normalization layers. Extensive experiments demonstrate the simplicity yet state-of-the-art performance of ReNorm among all the popular benchmarks. + +# Acknowledgements + +This work was supported by the Natural Science Foundation of China (No. 62372082, 61972071, U20A20184), Natural Science Foundation of Sichuan Province (No.2023NSFSC0485), National Science Foundation for Distinguished Young Scholars under Grant 62225605 and Zhejiang Key Research and Development Program under Grant 2023C03196. + +# References + +1. Bai, Y., Jiao, J., Ce, W., Liu, J., Lou, Y., Feng, X., Duan, L.Y.: Person30k: A dual-meta generalization network for person re-identification. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2123-2132 (2021) +2. Chen, Z., Wang, W., Zhao, Z., Su, F., Men, A., Dong, Y.: Cluster-instance normalization: A statistical relation-aware normalization for generalizable person re-identification. IEEE Transactions on Multimedia (2023) +3. Choi, S., Kim, T., Jeong, M., Park, H., Kim, C.: Meta batch-instance normalization for generalizable person re-identification. In: Proceedings of the IEEE/CVF conference on Computer Vision and Pattern Recognition. pp. 3425-3435 (2021) +4. Dai, Y., Li, X., Liu, J., Tong, Z., Duan, L.Y.: Generalizable person re-identification with relevance-aware mixture of experts. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16145-16154 (2021) +5. Dai, Y., Liu, J., Sun, Y., Tong, Z., Zhang, C., Duan, L.Y.: Idm: An intermediate domain module for domain adaptive person re-id. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 11864-11874 (2021) +6. Fu, D., Chen, D., Bao, J., Yang, H., Yuan, L., Zhang, L., Li, H., Chen, D.: Unsupervised pre-training for person re-identification. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 14750-14759 (2021) +7. Gray, D., Tao, H.: Viewpoint invariant pedestrian recognition with an ensemble of localized features. ECCV (1) 2008, 262-275 (2008) +8. Han, K., Si, C., Huang, Y., Wang, L., Tan, T.: Generalizable person re-identification via self-supervised batch norm test-time adaption. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 36, pp. 817-825 (2022) +9. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 770-778 (2016) +0. He, L., Liang, J., Li, H., Sun, Z.: Deep spatial feature reconstruction for partial person re-identification: Alignment-free approach. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 7073-7082 (2018) +1. He, L., Liao, X., Liu, W., Liu, X., Cheng, P., Mei, T.: Fastreid: A python toolbox for general instance re-identification. arXiv preprint arXiv:2006.02631 (2020) +12. He, S., Luo, H., Chen, W., Zhang, M., Zhang, Y., Wang, F., Li, H., Jiang, W.: Multi-domain learning and identity mining for vehicle re-identification. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops. pp. 582-583 (2020) +13. Hermans, A., Beyer, L., Leibe, B.: In defense of the triplet loss for person re-identification. arXiv preprint arXiv:1703.07737 (2017) + +14. Hirzer, M., Beleznai, C., Roth, P.M., Bischof, H.: Person re-identification by descriptive and discriminative classification. In: Image Analysis: 17th Scandinavian Conference, SCIA 2011, Ystad, Sweden, May 2011. Proceedings 17. pp. 91-102. Springer (2011) +15. Huang, X., Belongie, S.: Arbitrary style transfer in real-time with adaptive instance normalization. In: Proceedings of the IEEE international conference on computer vision. pp. 1501-1510 (2017) +16. Ioffe, S., Szegedy, C.: Batch normalization: Accelerating deep network training by reducing internal covariate shift. In: International conference on machine learning. pp. 448-456. pmlr (2015) +17. Jiang, K., Zhang, T., Zhang, Y., Wu, F., Rui, Y.: Self-supervised agent learning for unsupervised cross-domain person re-identification. IEEE Transactions on Image Processing 29, 8549–8560 (2020) +18. Jiao, B., Liu, L., Gao, L., Lin, G., Yang, L., Zhang, S., Wang, P., Zhang, Y.: Dynamically transformed instance normalization network for generalizable person re-identification. In: European Conference on Computer Vision. pp. 285–301. Springer (2022) +19. Jin, X., Lan, C., Zeng, W., Chen, Z., Zhang, L.: Style normalization and restitution for generalizable person re-identification. In: proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 3143-3152 (2020) +20. Lee, G., Lee, S., Kim, D., Shin, Y., Yoon, Y., Ham, B.: Camera-driven representation learning for unsupervised domain adaptive person re-identification. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 11453-11462 (2023) +21. Li, Q., Gong, S.: Mitigate domain shift by primary-auxiliary objectives association for generalizing person reid. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 394-403 (2024) +22. Li, W., Wang, X.: Locally aligned feature transforms across views. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 3594-3601 (2013) +23. Li, W., Zhao, R., Xiao, T., Wang, X.: Deepreid: Deep filter pairing neural network for person re-identification. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 152-159 (2014) +24. Li, Y., Wang, N., Shi, J., Hou, X., Liu, J.: Adaptive batch normalization for practical domain adaptation. Pattern Recognition 80, 109-117 (2018) +25. Lin, C.S., Cheng, Y.C., Wang, Y.C.F.: Domain generalized person re-identification via cross-domain episodic learning. In: 2020 25th International Conference on Pattern Recognition (ICPR). pp. 6758-6763 (2021). https://doi.org/10.1109/ICPR48806.2021.9413013 +26. Liu, J., Zha, Z.J., Chen, D., Hong, R., Wang, M.: Adaptive transfer network for cross-domain person re-identification. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 7202-7211 (2019) +27. Loy, C.C., Xiang, T., Gong, S.: Time-delayed correlation analysis for multi-camera activity understanding. International Journal of Computer Vision 90, 106-129 (2010) +28. Luo, H., Jiang, W., Gu, Y., Liu, F., Liao, X., Lai, S., Gu, J.: A strong baseline and batch normalization neck for deep person re-identification. IEEE Transactions on Multimedia 22(10), 2597-2609 (2019) +29. Luo, P., Ren, J., Peng, Z., Zhang, R., Li, J.: Differentiable learning-to-normalize via switchable normalization. arXiv preprint arXiv:1806.10779 (2018) + +30. Van der Maaten, L., Hinton, G.: Visualizing data using t-sne. Journal of machine learning research 9(11) (2008) +31. Nam, H., Kim, H.E.: Batch-instance normalization for adaptively style-invariant neural networks. Advances in Neural Information Processing Systems 31 (2018) +32. Ni, H., Song, J., Luo, X., Zheng, F., Li, W., Shen, H.T.: Meta distribution alignment for generalizable person re-identification. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2487-2496 (2022) +33. Pan, X., Luo, P., Shi, J., Tang, X.: Two at once: Enhancing learning and generalization capacities via ibn-net. In: Proceedings of the European Conference on Computer Vision (ECCV). pp. 464-479 (2018) +34. Qi, L., Wang, L., Shi, Y., Geng, X.: A novel mix-normalization method for generalizable multi-source person re-identification. IEEE Transactions on Multimedia (2022) +35. Ren, M., He, L., Liao, X., Liu, W., Wang, Y., Tan, T.: Learning instance-level spatial-temporal patterns for person re-identification. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 14930-14939 (2021) +36. Seo, S., Suh, Y., Kim, D., Kim, G., Han, J., Han, B.: Learning to optimize domain specific normalization for domain generalization. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXII 16. pp. 68-83. Springer (2020) +37. Song, J., Yang, Y., Song, Y.Z., Xiang, T., Hospedales, T.M.: Generalizable person re-identification by domain-invariant mapping network. In: Proceedings of the IEEE/CVF conference on Computer Vision and Pattern Recognition. pp. 719-728 (2019) +38. Ulyanov, D., Vedaldi, A., Lempitsky, V.: Instance normalization: The missing ingredient for fast stylization. arXiv preprint arXiv:1607.08022 (2016) +39. Wei, L., Zhang, S., Gao, W., Tian, Q.: Person transfer gan to bridge domain gap for person re-identification. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 79-88 (2018) +40. Wu, Y., Chi, Z., Wang, Y., Plataniotis, K.N., Feng, S.: Test-time domain adaptation by learning domain-aware batch normalization. arXiv preprint arXiv:2312.10165 (2023) +41. Wu, Y., Johnson, J.: Rethinking" batch" in batchnorm. arXiv preprint arXiv:2105.07576 (2021) +42. Xiao, T., Li, S., Wang, B., Lin, L., Wang, X.: End-to-end deep learning for person search. arXiv preprint arXiv:1604.01850 2(2), 4 (2016) +43. Xu, B., He, L., Liao, X., Liu, W., Sun, Z., Mei, T.: Black re-id: A head-shoulder descriptor for the challenging problem of person re-identification. In: Proceedings of the 28th ACM International Conference on Multimedia. pp. 673–681 (2020) +44. Xu, B., Liang, J., He, L., Sun, Z.: Mimic embedding via adaptive aggregation: Learning generalizable person re-identification. In: Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XIV. pp. 372-388. Springer (2022) +45. Zhang, P., Dou, H., Yu, Y., Li, X.: Adaptive cross-domain learning for generalizable person re-identification. In: European Conference on Computer Vision. pp. 215-232. Springer (2022) +46. Zhao, Y., Zhong, Z., Yang, F., Luo, Z., Lin, Y., Li, S., Sebe, N.: Learning to generalize unseen domains via memory-based multi-source meta-learning for person re-identification. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6277-6286 (2021) + +47. Zheng, L., Shen, L., Tian, L., Wang, S., Wang, J., Tian, Q.: Scalable person re-identification: A benchmark. In: Proceedings of the IEEE international conference on computer vision. pp. 1116-1124 (2015) +48. Zheng, W., Gong, S., Xiang, T.: Associating groups of people. In: British Machine Vision Conference (2009) +49. Zheng, Z., Ruan, T., Wei, Y., Yang, Y., Mei, T.: VehicleNet: Learning robust visual representation for vehicle re-identification. IEEE Transactions on Multimedia 23, 2683-2693 (2020) +50. Zheng, Z., Zheng, L., Yang, Y.: Unlabeled samples generated by gan improve the person re-identification baseline in vitro. In: Proceedings of the IEEE international conference on computer vision. pp. 3754-3762 (2017) \ No newline at end of file diff --git a/2024/Rethinking Normalization Layers for Domain Generalizable Person Re-identification/images.zip b/2024/Rethinking Normalization Layers for Domain Generalizable Person Re-identification/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..6038228a3be505bda08d65b706d592dce3004cfd --- /dev/null +++ b/2024/Rethinking Normalization Layers for Domain Generalizable Person Re-identification/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e5ce39edbe1ff89219d469a14004008d23a4446eb5379b6798520011a1688af4 +size 532399 diff --git a/2024/Rethinking Normalization Layers for Domain Generalizable Person Re-identification/layout.json b/2024/Rethinking Normalization Layers for Domain Generalizable Person Re-identification/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..fa7a021c50f891be60933349a11553838adbc98f --- /dev/null +++ b/2024/Rethinking Normalization Layers for Domain Generalizable Person Re-identification/layout.json @@ -0,0 +1,9044 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 149, + 111, + 465, + 146 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 111, + 465, + 146 + ], + "spans": [ + { + "bbox": [ + 149, + 111, + 465, + 146 + ], + "type": "text", + "content": "Rethinking Normalization Layers for Domain Generalizable Person Re-identification" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 202, + 168, + 411, + 181 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 202, + 168, + 411, + 181 + ], + "spans": [ + { + "bbox": [ + 202, + 168, + 411, + 181 + ], + "type": "text", + "content": "Ren Nie" + }, + { + "bbox": [ + 202, + 168, + 411, + 181 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 202, + 168, + 411, + 181 + ], + "type": "text", + "content": ", Jin Ding" + }, + { + "bbox": [ + 202, + 168, + 411, + 181 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 202, + 168, + 411, + 181 + ], + "type": "text", + "content": ", Xue Zhou" + }, + { + "bbox": [ + 202, + 168, + 411, + 181 + ], + "type": "inline_equation", + "content": "^{2,1*}" + }, + { + "bbox": [ + 202, + 168, + 411, + 181 + ], + "type": "text", + "content": ", and Xi Li" + }, + { + "bbox": [ + 202, + 168, + 411, + 181 + ], + "type": "inline_equation", + "content": "^{3}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 140, + 190, + 473, + 212 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 140, + 190, + 473, + 212 + ], + "spans": [ + { + "bbox": [ + 140, + 190, + 473, + 212 + ], + "type": "text", + "content": "1 University of Electronic Science and Technology of China, Chengdu, China \nnieren@std.uestc.edu.cn, dingjin1998@outlook.com, zhouxue@uestc.edu.cn" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 204, + 212, + 410, + 223 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 204, + 212, + 410, + 223 + ], + "spans": [ + { + "bbox": [ + 204, + 212, + 410, + 223 + ], + "type": "text", + "content": "2 Shenzhen Institute for Advanced Study, UESTC" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 172, + 223, + 441, + 246 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 172, + 223, + 441, + 246 + ], + "spans": [ + { + "bbox": [ + 172, + 223, + 441, + 246 + ], + "type": "text", + "content": "3 College of Computer Science & Technology, Zhejiang University xilizju@zju.edu.cn" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 159, + 280, + 455, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 280, + 455, + 533 + ], + "spans": [ + { + "bbox": [ + 159, + 280, + 455, + 533 + ], + "type": "text", + "content": "Abstract. Domain Generalizable Person Re-Identification (DG-ReID) strives to transfer learned feature representation from source domains to unseen target domains, despite significant distribution shifts. While most existing methods enhance model generalization and discriminative feature extraction capability by introducing Instance Normalization (IN) in combination with Batch Normalization (BN), these approaches still struggle with the overfitting of normalization layers to the source domains, posing challenges in domain generalization. To address this issue, we propose ReNorm, a purely normalization-based framework that integrates two complementary normalization layers through two forward propagations for the same weight matrix. In the first forward propagation, Remix Normalization (RN) combines IN and BN in a concise manner to ensure the feature extraction capability. As an effective complement to RN, Emulation Normalization (EN) simulates the testing process of RN, implicitly mitigating the domain shifts caused by the absence of target domain information and actively guiding the model in learning how to generalize the feature extraction capability to unseen target domains. Meanwhile, we propose Domain Frozen (DF), freezing updates to affine parameters to reduce the impact of these less robust parameters on overfitting to the source domains. Extensive experiments show that our framework achieves state-of-the-art performance on the popular benchmarks. The code is available at https://github.com/3699nr/ReNorm." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 160, + 544, + 448, + 554 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 544, + 448, + 554 + ], + "spans": [ + { + "bbox": [ + 160, + 544, + 448, + 554 + ], + "type": "text", + "content": "Keywords: Normalization " + }, + { + "bbox": [ + 160, + 544, + 448, + 554 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 544, + 448, + 554 + ], + "type": "text", + "content": " Domain generalization " + }, + { + "bbox": [ + 160, + 544, + 448, + 554 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 544, + 448, + 554 + ], + "type": "text", + "content": " Re-identification" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 578, + 230, + 590 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 578, + 230, + 590 + ], + "spans": [ + { + "bbox": [ + 132, + 578, + 230, + 590 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": "Person Re-identification (ReID) is a practical task that has been widely applied in video surveillance and security. It aims at retrieving identical persons across a multi-camera network with non-overlapping fields of view. While current approaches [6,10,12,28,35,43,49] have achieved impressive performance within the same source domains during both training and testing phases. However, when" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 116, + 481, + 284 + ], + "blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 284 + ], + "lines": [ + { + "bbox": [ + 133, + 116, + 481, + 284 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 481, + 284 + ], + "type": "image", + "image_path": "986784d482117644a269c053cf0f6f4d1012444bd7c3ed8725ea9786404f98a6.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 293, + 482, + 392 + ], + "lines": [ + { + "bbox": [ + 130, + 293, + 482, + 392 + ], + "spans": [ + { + "bbox": [ + 130, + 293, + 482, + 392 + ], + "type": "text", + "content": "Fig. 1: Differences between existing normalization-based DG-ReID methods and ReNorm. To illustrate more clearly, we use the example of training with three source domains. (a) Existing normalization-based methods combine IN and BN in a specific way to ensure both the generalization and discrimination capability of the weight matrix. (b) ReNorm integrates two complementary normalization layers through two forward propagations for the same weight matrix. RN enables the weight matrix to extract discriminative features within the source domains, while EN is employed to actively guide the weight matrix to generalize the feature extraction capability to unseen target domains. (c) Some details of ReNorm. Detailed explanations can be found in Sec 3." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 422, + 482, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 422, + 482, + 518 + ], + "spans": [ + { + "bbox": [ + 130, + 422, + 482, + 518 + ], + "type": "text", + "content": "directly deploying these well-trained models on an unseen target domain, an issue known as \"domain shift\" [39] arises, significantly impacting the performance of the models. To alleviate this problem, recent efforts have been devoted to two approaches: unsupervised domain adaptive (DA) ReID [5,17,20,26], and domain generalizable (DG) ReID [2,3,19,37,44,45]. In contrast to DA ReID, DG ReID is more practical and challenging as it trains on the multiple source domains and then directly transfers the learned representation to the unseen target domain without any target data for fine-tuning." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 521, + 485, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 485, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 485, + 666 + ], + "type": "text", + "content": "Among numerous DG ReID methods, normalization-based approaches [2,3,8, 19,37] stand out for their simplicity and versatility, often integrating seamlessly with other advanced techniques such as meta-learning, distribution alignment, and ensemble learning. Illustrated in Fig.1(a), the combination of Batch Normalization (BN) [16] for capturing discriminative features and Instance Normalization (IN) [38] for eliminating domain-specific information facilitates normalization-based DG-ReID methods in achieving both high generalization and discrimination capabilities. During training, the statistical parameters of both IN and BN within the normalization layers are derived from the same data, ensuring a consistent distribution. However, this consistency is disrupted when the model is deployed on the target domain, as IN statistical parameters originate from the target domain itself, while BN statistical parameters are derived from population" + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 268, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 268, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 268, + 102 + ], + "type": "text", + "content": "R. Nie and J. Ding et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 480, + 176 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 480, + 176 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 480, + 176 + ], + "type": "text", + "content": "statistics accumulated from source domains during training. The distribution of the target domain significantly differs from that of the source domains, resulting in a marked decline in the model's performance. Furthermore, the vulnerable and sensitive learnable affine parameters are also prone to overfitting to the source domains, leading to performance degeneration in both training and testing phase." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 177, + 481, + 272 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 177, + 481, + 272 + ], + "spans": [ + { + "bbox": [ + 130, + 177, + 481, + 272 + ], + "type": "text", + "content": "Moreover, emerging researches [24, 40] have shown that the identity and domain information are separately embedded in the weight matrix (learnable parameters in convolutional layers) and normalization layers. Consequently, we rethink the key challenges of DG ReID, which are rooted in normalization layers. We believe that the capability of the weight matrix to extract discriminative features sets the upper limit of the model's performance. Simultaneously, the normalization layers, responsible for fitting and scaling the input from target domains, establish the lower limit for the model's generalization performance." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 273, + 482, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 273, + 482, + 392 + ], + "spans": [ + { + "bbox": [ + 130, + 273, + 482, + 392 + ], + "type": "text", + "content": "To address the issue of normalization layers being prone to overfitting to the source domains, we restructure the normalization layers, proposing ReNorm, a concise yet effective DG ReID framework entirely based on the design of normalization layers, as shown in Fig.1(b). In each iteration, the network undergoes two forward propagations and then one backpropagation after aggregating two forward losses. The two forward propagations apply different processing to the normalization layers, utilizing two complementary normalization layers to facilitate the weight matrix in extracting discriminative features from familiar source domains, while simultaneously guiding the same weight matrix to learn how to transfer this acquired feature extraction capability to unfamiliar target domains." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 393, + 482, + 595 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 393, + 482, + 595 + ], + "spans": [ + { + "bbox": [ + 130, + 393, + 482, + 595 + ], + "type": "text", + "content": "Specifically, in our ReNorm framework, the two forward propagations correspond to two different but complementary normalization strategies: Remix Normalization (RN) and Emulation Normalization (EN). In the first forward propagation, RN introduces Instance Normalization (IN) in combination with Batch Normalization (BN) by directly summing their statistical parameters for standardization. Meanwhile, the BN statistical parameters of the current source domain are stored in the corresponding buffer, as shown in Fig.1(c). In the second forward propagation, EN adopts a similar summing approach but emulates the inference process of RN. It treats the current input as a simulated target domain, and calculates the IN statistical parameters accordingly. Simultaneously, the required BN statistical parameters are obtained from a remaining source domain's buffer stored in RN. This pairwise normalization strategy enables the network to actively learn how to transfer knowledge from source domains to the target domain with a distinct distribution. Additionally, we introduce Domain Frozen (DF), where the vulnerable settings of learnable affine parameters within both normalization layers are frozen and exempt from updates. By doing so, ReNorm greatly raises the lower limit of the model's generalization ability." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 147, + 596, + 339, + 607 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 147, + 596, + 339, + 607 + ], + "spans": [ + { + "bbox": [ + 147, + 596, + 339, + 607 + ], + "type": "text", + "content": "We summarize our contributions as follows:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 617, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 617, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 617, + 481, + 665 + ], + "type": "text", + "content": "- We introduce ReNorm, a simple yet effective framework for DG ReID, entirely and solely based on the design of normalization layers. This framework significantly mitigates performance degradation resulting from distribution shifts between source and target domains in a concise yet effective way." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 253, + 91, + 445, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 253, + 91, + 445, + 102 + ], + "spans": [ + { + "bbox": [ + 253, + 91, + 445, + 102 + ], + "type": "text", + "content": "Rethinking Normalization Layers for DG-ReID" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 137, + 116, + 479, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 116, + 479, + 163 + ], + "spans": [ + { + "bbox": [ + 137, + 116, + 479, + 163 + ], + "type": "text", + "content": "- We introduce two complementary normalization strategies (i.e., Remix Normalization and Emulation Normalization) equipped with Domain Frozen, actively enabling the network to learn how to transfer the feature extraction capability acquired from source domains to unseen target domains." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 163, + 480, + 187 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 163, + 480, + 187 + ], + "spans": [ + { + "bbox": [ + 138, + 163, + 480, + 187 + ], + "type": "text", + "content": "- Extensive experiments demonstrate the simplicity yet state-of-the-art performance of our ReNorm framework under multiple testing protocols." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 201, + 386, + 215 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 201, + 386, + 215 + ], + "spans": [ + { + "bbox": [ + 132, + 201, + 386, + 215 + ], + "type": "text", + "content": "2 Generalizable Person Re-Identification." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 223, + 482, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 223, + 482, + 342 + ], + "spans": [ + { + "bbox": [ + 130, + 223, + 482, + 342 + ], + "type": "text", + "content": "Deep learning-based Person ReID approaches [6,10,12,28,35,43,49] have made great progress in recent years, achieving impressive accuracy in Person ReID. However, when these well-trained models are applied to unseen target domains that have distribution shifts with source domains, the performance decreases significantly. To alleviate this problem, recent efforts are devoted to two approaches: unsupervised domain adaptive (DA) ReID [5,17,20,26] and domain generalizable (DG) ReID. Diverging from the DA ReID, which needs unlabeled target domain datasets for unsupervised fine-tuning. DG ReID trains on the source domains and then directly tests on the unseen target domain. As a result, DG ReID has attracted attention for its robustness and generalization." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 146, + 342, + 459, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 342, + 459, + 354 + ], + "spans": [ + { + "bbox": [ + 146, + 342, + 459, + 354 + ], + "type": "text", + "content": "Existing DG ReID methods can be briefly divided into four categories:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 354, + 482, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 354, + 482, + 498 + ], + "spans": [ + { + "bbox": [ + 130, + 354, + 482, + 498 + ], + "type": "text", + "content": "Normalization strategies [2, 33, 34]. Batch Normalization (BN) can preserve diversity among samples, capturing more discriminative features and accelerating convergence. However, it tends to be prone to overfitting to source domains. Instance Normalization (IN) can eliminate domain-specific information to enhance generalization but inevitably removes discriminative information. Thus, normalization-based methods typically combine IN and BN in various ways to ensure both high generalization and discrimination capability of the network. Jiao et al. [18] proposed Dynamically Transformed Instance Normalization (DTIN) to alleviate the drawback of IN by transforming normalized features into appropriate representations in a learnable and adaptive manner. Chen et al. [2] designed a Cluster-Instance Normalization (CINorm) to eliminate the effect of outliers in original BN that causes severe overfitting." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 498, + 482, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 498, + 482, + 569 + ], + "spans": [ + { + "bbox": [ + 130, + 498, + 482, + 569 + ], + "type": "text", + "content": "Domain alignment [19,32]. These methods try to minimize the difference among source domains for learning domain-invariant representations. They assume that the invariant features from the source domains should also be robust to any unseen target domains. Jin et al. [19] proposed a dual causality loss to align the outputs and a Style Normalization and Restitution (SNR) module to distill identity-relevant features via a restitution module for IN compensation." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 570, + 482, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 629 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 629 + ], + "type": "text", + "content": "Meta-learning [3,37,46]. These methods improve generalized performance through virtual simulations of real train-test domain shifts. Following the concept of 'learning to lean', Choi et al. [3]proposed MetaBIN which utilized a meta-learning training strategy with sets of learnable balancing parameters to integrate BN and IN." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "text", + "content": "Ensemble learning [4,25,44]. These methods exploit the relationship among multiple source domains by using diverse experts to extract domain-specific features from the corresponding domain. To avoid the large model size, Xu et al. [44]" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 141, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 141, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 141, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 268, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 268, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 268, + 102 + ], + "type": "text", + "content": "R. Nie and J. Ding et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 140 + ], + "type": "text", + "content": "proposed a lightweight model META, which leverages IN and introduces it into a global branch to pursue invariant features across domains." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 140, + 482, + 201 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 140, + 482, + 201 + ], + "spans": [ + { + "bbox": [ + 130, + 140, + 482, + 201 + ], + "type": "text", + "content": "By introducing IN in combination with BN, normalization-based methods are simple and effective and can be combined with other techniques. However, when tested on unknown target domains, the distribution shifts between source and target domains significantly affect the fitting and scaling ability of normalization layers. Thus, leading to a decline in the model's generalization performance." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 220, + 233, + 234 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 220, + 233, + 234 + ], + "spans": [ + { + "bbox": [ + 131, + 220, + 233, + 234 + ], + "type": "text", + "content": "3 Methodology" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 247, + 482, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 247, + 482, + 450 + ], + "spans": [ + { + "bbox": [ + 130, + 247, + 482, + 450 + ], + "type": "text", + "content": "Inspired by the observation that the weight matrix tends to encapsulate identity information while domain-specific knowledge is embedded within the normalization layers [24, 40], we propose a novel ReNorm framework for tackling the domain shifts problem for DG-ReID which is rooted in normalization layers. As shown in Fig.2, ReNorm integrates two complementary normalization layers through two forward propagations for the same weight matrix. In each iteration, we sample a mini-batch from a source domain. In the first forward propagation, Remix Normalization (RN) combines IN and BN to ensure the discriminative feature extraction capability of the weight matrix. At the same time, the BN statistical parameters of the current source domain are stored in the corresponding buffer by Exponential Moving Average (EMA). During the second forward propagation, we treat the current input as a simulated target domain of the remaining source domains. Thus, Emulation Normalization (EN) can emulate the inference process of RN to guide the weight matrix to learn how to transfer the acquired feature extraction capability to an unseen domain. Meanwhile, we introduce Domain Frozen (DF) to stabilize the weight matrix by freezing the updates of affine parameters in normalization layers." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 468, + 221, + 480 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 468, + 221, + 480 + ], + "spans": [ + { + "bbox": [ + 131, + 468, + 221, + 480 + ], + "type": "text", + "content": "3.1 Preliminary" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 490, + 482, + 572 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 490, + 482, + 572 + ], + "spans": [ + { + "bbox": [ + 130, + 490, + 482, + 572 + ], + "type": "text", + "content": "In the DG ReID task, we have access to " + }, + { + "bbox": [ + 130, + 490, + 482, + 572 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 490, + 482, + 572 + ], + "type": "text", + "content": " source domains (datasets) for training. We denote the " + }, + { + "bbox": [ + 130, + 490, + 482, + 572 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 490, + 482, + 572 + ], + "type": "text", + "content": "-th source domain as " + }, + { + "bbox": [ + 130, + 490, + 482, + 572 + ], + "type": "inline_equation", + "content": "D_{k}, k \\in 1,2,\\dots,K" + }, + { + "bbox": [ + 130, + 490, + 482, + 572 + ], + "type": "text", + "content": ", which have completely non-overlapping label spaces. Given a feature map " + }, + { + "bbox": [ + 130, + 490, + 482, + 572 + ], + "type": "inline_equation", + "content": "X_{k} \\in \\mathbb{R}^{N \\times C \\times H \\times W}" + }, + { + "bbox": [ + 130, + 490, + 482, + 572 + ], + "type": "text", + "content": " of a minibatch extracted from a certain domain " + }, + { + "bbox": [ + 130, + 490, + 482, + 572 + ], + "type": "inline_equation", + "content": "D_{k}" + }, + { + "bbox": [ + 130, + 490, + 482, + 572 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 490, + 482, + 572 + ], + "type": "inline_equation", + "content": "N,C,H" + }, + { + "bbox": [ + 130, + 490, + 482, + 572 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 490, + 482, + 572 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 130, + 490, + 482, + 572 + ], + "type": "text", + "content": " denote the batch size, the number of channels, the height, and the width, respectively. Following the definition from [41], we decompose the normalization layers into two steps as follows:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 234, + 573, + 481, + 600 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 234, + 573, + 481, + 600 + ], + "spans": [ + { + "bbox": [ + 234, + 573, + 481, + 600 + ], + "type": "interline_equation", + "content": "\\hat {X} _ {k} = \\frac {X _ {k} - \\mu_ {k}}{\\sqrt {\\sigma_ {k} ^ {2} + \\epsilon}}, \\quad \\tilde {X} _ {k} = \\gamma \\hat {X} _ {k} + \\beta , \\tag {1}", + "image_path": "ae4f9cab66c79acf0dcd98369bf20d7063954728a6349383bfb0a74444e4a9ad.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": "The first step involves standardization using the mean " + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mu_{k}" + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": " and variance " + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\sigma_k^2" + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": " calculated from the input feature map. After that, channel-wise learnable affine parameters " + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": " are defined for scaling and shifting during end-to-end training. " + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\epsilon >0" + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": " is a constant to avoid numerical problems. The main distinction among various normalization techniques primarily lies in the first step." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 252, + 91, + 445, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 252, + 91, + 445, + 102 + ], + "spans": [ + { + "bbox": [ + 252, + 91, + 445, + 102 + ], + "type": "text", + "content": "Rethinking Normalization Layers for DG-ReID" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 134, + 115, + 480, + 369 + ], + "blocks": [ + { + "bbox": [ + 134, + 115, + 480, + 369 + ], + "lines": [ + { + "bbox": [ + 134, + 115, + 480, + 369 + ], + "spans": [ + { + "bbox": [ + 134, + 115, + 480, + 369 + ], + "type": "image", + "image_path": "05a152d63138a59f14f1ae9a7fc5ed43aaa261a9becef2a98bde9879893dfb50.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 379, + 482, + 434 + ], + "lines": [ + { + "bbox": [ + 130, + 379, + 482, + 434 + ], + "spans": [ + { + "bbox": [ + 130, + 379, + 482, + 434 + ], + "type": "text", + "content": "Fig. 2: The intra-layer structure of our ReNorm framework. The upper part corresponds to the first forward propagation with Remix Normalization, while the lower part corresponds to the second forward propagation with Emulation Normalization. After standardization, we employ Domain Frozen strategy for the affine parameters. Best viewed in color, detailed explanations can be found in Sec 3.2 and Sec 3.3." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 455, + 482, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 455, + 482, + 517 + ], + "spans": [ + { + "bbox": [ + 130, + 455, + 482, + 517 + ], + "type": "text", + "content": "Instance Normalization [38] can remove style information from input features. So it has been widely used for the DG-ReID task to extract domain-invariant representations by removing their statistical contrast across feature channels. IN calculates mean " + }, + { + "bbox": [ + 130, + 455, + 482, + 517 + ], + "type": "inline_equation", + "content": "\\mu_k^{in} \\in \\mathbb{R}^{N \\times C}" + }, + { + "bbox": [ + 130, + 455, + 482, + 517 + ], + "type": "text", + "content": " and variance " + }, + { + "bbox": [ + 130, + 455, + 482, + 517 + ], + "type": "inline_equation", + "content": "\\sigma_k^{in^2} \\in \\mathbb{R}^{N \\times C}" + }, + { + "bbox": [ + 130, + 455, + 482, + 517 + ], + "type": "text", + "content": " with respect to each sample and each channel for standardization:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 187, + 521, + 481, + 550 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 187, + 521, + 481, + 550 + ], + "spans": [ + { + "bbox": [ + 187, + 521, + 481, + 550 + ], + "type": "interline_equation", + "content": "\\mu_ {k} ^ {i n} = \\frac {\\sum_ {h , w} ^ {H W} X _ {k}}{H W} \\quad \\text {a n d} \\quad \\sigma_ {k} ^ {i n 2} = \\frac {\\sum_ {h , w} ^ {H W} \\left(X _ {k} - \\mu_ {k} ^ {i n}\\right) ^ {2}}{H W}. \\tag {2}", + "image_path": "d2ac982f13ce18d73d06c0311f7daf180c748c9717f1e5b3e675aecebb2a390a.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 554, + 482, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 554, + 482, + 602 + ], + "spans": [ + { + "bbox": [ + 130, + 554, + 482, + 602 + ], + "type": "text", + "content": "Batch Normalization [16] is widely used in convolutional neural networks to capture discriminative features and mitigate gradient vanishing or exploding. It calculates mean " + }, + { + "bbox": [ + 130, + 554, + 482, + 602 + ], + "type": "inline_equation", + "content": "\\mu_k^{bn} \\in \\mathbb{R}^C" + }, + { + "bbox": [ + 130, + 554, + 482, + 602 + ], + "type": "text", + "content": " and variance " + }, + { + "bbox": [ + 130, + 554, + 482, + 602 + ], + "type": "inline_equation", + "content": "\\sigma_k^{bn^2} \\in \\mathbb{R}^C" + }, + { + "bbox": [ + 130, + 554, + 482, + 602 + ], + "type": "text", + "content": " with respect each channel in a mini-batch:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 179, + 606, + 481, + 635 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 179, + 606, + 481, + 635 + ], + "spans": [ + { + "bbox": [ + 179, + 606, + 481, + 635 + ], + "type": "interline_equation", + "content": "\\mu_ {k} ^ {b n} = \\frac {\\sum_ {n , h , w} ^ {N H W} X _ {k}}{N H W} \\quad \\text {a n d} \\quad \\sigma_ {k} ^ {b n ^ {2}} = \\frac {\\sum_ {n , h , w} ^ {N H W} \\left(X _ {k} - \\mu_ {k} ^ {b n}\\right) ^ {2}}{N H W}. \\tag {3}", + "image_path": "50e6f08fa680f60232fbc1cd32c891428cb2727ad601c5bb17d5ed60f11d272a.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 640, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 640, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 640, + 482, + 666 + ], + "type": "text", + "content": "It is worth noting that different from IN where the mean " + }, + { + "bbox": [ + 130, + 640, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mu^{in}" + }, + { + "bbox": [ + 130, + 640, + 482, + 666 + ], + "type": "text", + "content": " and variance " + }, + { + "bbox": [ + 130, + 640, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\sigma^{in^2}" + }, + { + "bbox": [ + 130, + 640, + 482, + 666 + ], + "type": "text", + "content": " are computed from the input feature map " + }, + { + "bbox": [ + 130, + 640, + 482, + 666 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 130, + 640, + 482, + 666 + ], + "type": "text", + "content": " during both training and test-" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 267, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 267, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 267, + 102 + ], + "type": "text", + "content": "R. Nie and J. Ding et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 175 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 175 + ], + "type": "text", + "content": "ing. During testing, BN standardizes the input feature map " + }, + { + "bbox": [ + 130, + 116, + 482, + 175 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 130, + 116, + 482, + 175 + ], + "type": "text", + "content": " by population statistics which are estimated using Exponential Moving Average (EMA) during training. This special mechanism of BN enriched with domain-specific information of source domains poses a significant risk to the generalization performance of the model." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 194, + 269, + 205 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 194, + 269, + 205 + ], + "spans": [ + { + "bbox": [ + 132, + 194, + 269, + 205 + ], + "type": "text", + "content": "3.2 Remix Normalization" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 213, + 482, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 213, + 482, + 323 + ], + "spans": [ + { + "bbox": [ + 130, + 213, + 482, + 323 + ], + "type": "text", + "content": "Previous studies [15,31] have shown that Batch Normalization (BN) can extract domain-specific discriminative information, but it can easily overfit to a particular domain data in a mini-batch, while Instance Normalization (IN) can reduce the domain-specific style information, but it results in the loss of some discriminative information. Therefore, we combine these two normalization methods for a trade-off. Inspired by DSON [36] and SN [29], we design Remix Normalization (RN) which combines the statistical parameters of BN and IN in a simple way. Given an input feature map " + }, + { + "bbox": [ + 130, + 213, + 482, + 323 + ], + "type": "inline_equation", + "content": "X_{k}" + }, + { + "bbox": [ + 130, + 213, + 482, + 323 + ], + "type": "text", + "content": ", RN calculates mean " + }, + { + "bbox": [ + 130, + 213, + 482, + 323 + ], + "type": "inline_equation", + "content": "\\mu_k^{rn} \\in \\mathbb{R}^{N \\times C}" + }, + { + "bbox": [ + 130, + 213, + 482, + 323 + ], + "type": "text", + "content": " and variance " + }, + { + "bbox": [ + 130, + 213, + 482, + 323 + ], + "type": "inline_equation", + "content": "\\sigma_k^{rn^2} \\in \\mathbb{R}^{N \\times C}" + }, + { + "bbox": [ + 130, + 213, + 482, + 323 + ], + "type": "text", + "content": " for standardization as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 260, + 331, + 481, + 353 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 260, + 331, + 481, + 353 + ], + "spans": [ + { + "bbox": [ + 260, + 331, + 481, + 353 + ], + "type": "interline_equation", + "content": "\\mu_ {k} ^ {r n} = \\left(\\mu_ {k} ^ {\\mathrm {i n}} + \\mu_ {k} ^ {\\mathrm {b n}}\\right) / 2,", + "image_path": "94ba4c9dc1cadf5b90b0dc39aea672a29f191dfa19cea275755ad05ae1132c7e.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 254, + 348, + 360, + 364 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 254, + 348, + 360, + 364 + ], + "spans": [ + { + "bbox": [ + 254, + 348, + 360, + 364 + ], + "type": "interline_equation", + "content": "\\sigma_ {k} ^ {r n 2} = (\\sigma_ {k} ^ {i n 2} + \\sigma_ {k} ^ {b n 2}) / 2.", + "image_path": "dfbf76dab4c1653f3c6c4be90649cf214f9f08caa65d0cdecba93b9255ffd123.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 372, + 482, + 419 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 372, + 482, + 419 + ], + "spans": [ + { + "bbox": [ + 130, + 372, + 482, + 419 + ], + "type": "text", + "content": "Meanwhile, unlike previous approaches where each normalization layer corresponds to one BN buffer, we set up corresponding " + }, + { + "bbox": [ + 130, + 372, + 482, + 419 + ], + "type": "inline_equation", + "content": "Buffer_{k}:\\mu_{k}^{pop},\\sigma_{k}^{pop2}\\in \\mathbb{R}^{C},k\\in 1,2,\\ldots ,K" + }, + { + "bbox": [ + 130, + 372, + 482, + 419 + ], + "type": "text", + "content": " for each RN for " + }, + { + "bbox": [ + 130, + 372, + 482, + 419 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 372, + 482, + 419 + ], + "type": "text", + "content": " source domains to store their domain-specific information as follows:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 247, + 429, + 374, + 443 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 247, + 429, + 374, + 443 + ], + "spans": [ + { + "bbox": [ + 247, + 429, + 374, + 443 + ], + "type": "interline_equation", + "content": "\\mu_ {k} ^ {\\text {p o p}} \\leftarrow \\lambda \\mu_ {k} ^ {\\text {p o p}} + (1 - \\lambda) \\mu_ {k} ^ {b n},", + "image_path": "0f5db49856eff8e627e66443aeee6fd6be4b9828ac2c9278b0ba213b29aed62b.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 240, + 439, + 480, + 461 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 240, + 439, + 480, + 461 + ], + "spans": [ + { + "bbox": [ + 240, + 439, + 480, + 461 + ], + "type": "interline_equation", + "content": "\\sigma_ {k} ^ {p o p 2} \\leftarrow \\lambda \\sigma_ {k} ^ {p o p 2} + (1 - \\lambda) \\sigma_ {k} ^ {b n 2}. \\tag {5}", + "image_path": "75d565e2120d9fbb317c2cf5daf898ca0738a0bead9f1fd89ef831f620f31cbb.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 470, + 482, + 530 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 470, + 482, + 530 + ], + "spans": [ + { + "bbox": [ + 130, + 470, + 482, + 530 + ], + "type": "text", + "content": "As shown in Fig.2, we use three primary colors (yellow, red, blue) to represent the corresponding source domains " + }, + { + "bbox": [ + 130, + 470, + 482, + 530 + ], + "type": "inline_equation", + "content": "D_{1}" + }, + { + "bbox": [ + 130, + 470, + 482, + 530 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 470, + 482, + 530 + ], + "type": "inline_equation", + "content": "D_{2}" + }, + { + "bbox": [ + 130, + 470, + 482, + 530 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 470, + 482, + 530 + ], + "type": "inline_equation", + "content": "D_{K}" + }, + { + "bbox": [ + 130, + 470, + 482, + 530 + ], + "type": "text", + "content": ". The input mini-batch is from " + }, + { + "bbox": [ + 130, + 470, + 482, + 530 + ], + "type": "inline_equation", + "content": "D_{1}" + }, + { + "bbox": [ + 130, + 470, + 482, + 530 + ], + "type": "text", + "content": ". After calculating the required IN and BN statistical parameters, we follow Eq.5 to update the corresponding BN population statistics (i.e., " + }, + { + "bbox": [ + 130, + 470, + 482, + 530 + ], + "type": "inline_equation", + "content": "Buffer_{1}:\\mu_{1}^{pop},\\sigma_{1}^{pop2}" + }, + { + "bbox": [ + 130, + 470, + 482, + 530 + ], + "type": "text", + "content": "). In Fig.2, we represent this update process with solid lines." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 531, + 482, + 627 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 531, + 482, + 627 + ], + "spans": [ + { + "bbox": [ + 130, + 531, + 482, + 627 + ], + "type": "text", + "content": "The efficacy of combining BN and IN has been verified in many previous DG ReID works [2, 33, 34]. Nevertheless, a common challenge persists across these approaches: overfitting of the normalization layers to the source domains. Specifically, during the training process, following Eq.4, the statistical parameters of IN and BN within the normalization layer are derived from the same source domain, resulting in a consistent distribution. However, when the model is directly tested on the target domain, there is no concept of mini-batch anymore [41]. Thus, RN calculates the parameters for standardization as follows:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 255, + 636, + 481, + 658 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 255, + 636, + 481, + 658 + ], + "spans": [ + { + "bbox": [ + 255, + 636, + 481, + 658 + ], + "type": "interline_equation", + "content": "\\mu_ {k} ^ {r n} = \\left(\\mu_ {t a r} ^ {\\text {i n}} + \\mu_ {k} ^ {\\text {p o p}}\\right) / 2, \\tag {6}", + "image_path": "d17f0117c170a9a09cb4e612d043f613cc57c32c13d30df92d03e22b1623a86a.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 251, + 653, + 363, + 667 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 251, + 653, + 363, + 667 + ], + "spans": [ + { + "bbox": [ + 251, + 653, + 363, + 667 + ], + "type": "interline_equation", + "content": "\\sigma_ {k} ^ {r n 2} = (\\sigma_ {t a r} ^ {i n 2} + \\sigma_ {k} ^ {p o p 2}) / 2,", + "image_path": "8beb96ebf7c20eecbb7e2fcac0929a3658c5de4379f62ab36dce3b1206c4e299.jpg" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 252, + 90, + 445, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 252, + 90, + 445, + 102 + ], + "spans": [ + { + "bbox": [ + 252, + 90, + 445, + 102 + ], + "type": "text", + "content": "Rethinking Normalization Layers for DG-ReID" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 115, + 482, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 115, + 482, + 224 + ], + "spans": [ + { + "bbox": [ + 130, + 115, + 482, + 224 + ], + "type": "text", + "content": "the IN statistical parameters (i.e., " + }, + { + "bbox": [ + 130, + 115, + 482, + 224 + ], + "type": "inline_equation", + "content": "\\mu_{tar}^{\\mathrm{in}},\\sigma_{tar}^{in2}" + }, + { + "bbox": [ + 130, + 115, + 482, + 224 + ], + "type": "text", + "content": ") are derived from the input target domain itself. Conversely, the BN statistical parameters are based on population statistics (i.e., " + }, + { + "bbox": [ + 130, + 115, + 482, + 224 + ], + "type": "inline_equation", + "content": "\\mu_k^{pop},\\sigma_k^{pop2}" + }, + { + "bbox": [ + 130, + 115, + 482, + 224 + ], + "type": "text", + "content": ") of source domains. The actual distribution between the target domain and the source domain is completely different, which leads to the inability of the statistics parameters (i.e., " + }, + { + "bbox": [ + 130, + 115, + 482, + 224 + ], + "type": "inline_equation", + "content": "\\mu_k^{rn},\\sigma_k^{rn2}" + }, + { + "bbox": [ + 130, + 115, + 482, + 224 + ], + "type": "text", + "content": ") to fit the input " + }, + { + "bbox": [ + 130, + 115, + 482, + 224 + ], + "type": "inline_equation", + "content": "X_{tar}" + }, + { + "bbox": [ + 130, + 115, + 482, + 224 + ], + "type": "text", + "content": " from the target domain. Concurrently, the model lacks the adaptability to address this mismatch, resulting in a performance decline. Furthermore, the learnable affine parameters also tend to overfit to the source domains, leading to a decrease in performance during both training and testing." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 239, + 288, + 250 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 239, + 288, + 250 + ], + "spans": [ + { + "bbox": [ + 132, + 239, + 288, + 250 + ], + "type": "text", + "content": "3.3 Emulation Normalization" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 255, + 480, + 339 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 255, + 480, + 339 + ], + "spans": [ + { + "bbox": [ + 130, + 255, + 480, + 339 + ], + "type": "text", + "content": "Remix Normalization enables the weight matrix in learning to extract discriminative features while also possessing some degree of domain invariance. However, a crucial limitation emerges when the model encounters the distribution shifts introduced by the unseen target domain, as it lacks the knowledge of how to adapt. Hence, a natural consideration arises: given the entirely distinct distributions among different source domains, we can treat the current source domain as a simulated target domain of a remaining source domain." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 340, + 481, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 340, + 481, + 411 + ], + "spans": [ + { + "bbox": [ + 130, + 340, + 481, + 411 + ], + "type": "text", + "content": "Thus, we introduce a novel normalization operation named Emulation Normalization (EN), which emulates the inference procedure of Remix Normalization during training. By doing so, EN actively guides the weight matrix to acquire the capability of handling issues arising from the inability of RN to fit the input from unseen target domains. EN calculates mean " + }, + { + "bbox": [ + 130, + 340, + 481, + 411 + ], + "type": "inline_equation", + "content": "\\mu_k^{en} \\in \\mathbb{R}^{N \\times C}" + }, + { + "bbox": [ + 130, + 340, + 481, + 411 + ], + "type": "text", + "content": " and variance " + }, + { + "bbox": [ + 130, + 340, + 481, + 411 + ], + "type": "inline_equation", + "content": "\\sigma_k^{en2} \\in \\mathbb{R}^{N \\times C}" + }, + { + "bbox": [ + 130, + 340, + 481, + 411 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 235, + 417, + 480, + 451 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 235, + 417, + 480, + 451 + ], + "spans": [ + { + "bbox": [ + 235, + 417, + 480, + 451 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mu_ {k} ^ {e n} = \\left(1 - w _ {j} ^ {m}\\right) \\mu_ {k} ^ {\\mathrm {i n}} + w _ {j} ^ {m} \\mu_ {j} ^ {\\mathrm {p o p}} \\\\ \\sigma_ {k} ^ {e n 2} = \\left(1 - w _ {j} ^ {v}\\right) \\sigma_ {k} ^ {i n 2} + w _ {j} ^ {v} \\sigma_ {j} ^ {p o p 2}, \\tag {7} \\\\ \\end{array}", + "image_path": "c66243b1399ef10a24e6c4102e9d8dc9416f90af6f619703b056d952b300dfa7.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 454, + 482, + 550 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 454, + 482, + 550 + ], + "spans": [ + { + "bbox": [ + 130, + 454, + 482, + 550 + ], + "type": "text", + "content": "where we treat the current input source domain " + }, + { + "bbox": [ + 130, + 454, + 482, + 550 + ], + "type": "inline_equation", + "content": "D_{k}" + }, + { + "bbox": [ + 130, + 454, + 482, + 550 + ], + "type": "text", + "content": " as the simulated target domain of " + }, + { + "bbox": [ + 130, + 454, + 482, + 550 + ], + "type": "inline_equation", + "content": "D_{j}" + }, + { + "bbox": [ + 130, + 454, + 482, + 550 + ], + "type": "text", + "content": ", which is randomly selected from the remaining. By combining IN statistical parameters of " + }, + { + "bbox": [ + 130, + 454, + 482, + 550 + ], + "type": "inline_equation", + "content": "X_{k}" + }, + { + "bbox": [ + 130, + 454, + 482, + 550 + ], + "type": "text", + "content": " from this simulated target domain " + }, + { + "bbox": [ + 130, + 454, + 482, + 550 + ], + "type": "inline_equation", + "content": "D_{k}" + }, + { + "bbox": [ + 130, + 454, + 482, + 550 + ], + "type": "text", + "content": " and BN population statistics from a remaining source domain " + }, + { + "bbox": [ + 130, + 454, + 482, + 550 + ], + "type": "inline_equation", + "content": "D_{j}" + }, + { + "bbox": [ + 130, + 454, + 482, + 550 + ], + "type": "text", + "content": ", our EN can actively guide the weight matrix to transfer discriminative feature extraction from familiar source domains to unfamiliar target domain with different distribution. Notably, the learnable balancing ratios (i.e., " + }, + { + "bbox": [ + 130, + 454, + 482, + 550 + ], + "type": "inline_equation", + "content": "w_{j}^{m}, w_{j}^{v}" + }, + { + "bbox": [ + 130, + 454, + 482, + 550 + ], + "type": "text", + "content": ") correspond individually to the randomly selected source domain " + }, + { + "bbox": [ + 130, + 454, + 482, + 550 + ], + "type": "inline_equation", + "content": "D_{j}" + }, + { + "bbox": [ + 130, + 454, + 482, + 550 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 551, + 482, + 609 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 551, + 482, + 609 + ], + "spans": [ + { + "bbox": [ + 130, + 551, + 482, + 609 + ], + "type": "text", + "content": "As shown in Fig.2, We use the current input " + }, + { + "bbox": [ + 130, + 551, + 482, + 609 + ], + "type": "inline_equation", + "content": "D_{1}" + }, + { + "bbox": [ + 130, + 551, + 482, + 609 + ], + "type": "text", + "content": " as the simulated target domain of " + }, + { + "bbox": [ + 130, + 551, + 482, + 609 + ], + "type": "inline_equation", + "content": "D_{2}" + }, + { + "bbox": [ + 130, + 551, + 482, + 609 + ], + "type": "text", + "content": " which supports the required BN statistical parameters stored in Buffer2 (i.e., " + }, + { + "bbox": [ + 130, + 551, + 482, + 609 + ], + "type": "inline_equation", + "content": "\\mu_2^{pop}, \\sigma_2^{pop^2}" + }, + { + "bbox": [ + 130, + 551, + 482, + 609 + ], + "type": "text", + "content": "), as specified in Eq.7. It is worth noting that Buffer2 could combine IN statistical parameters from " + }, + { + "bbox": [ + 130, + 551, + 482, + 609 + ], + "type": "inline_equation", + "content": "D_{1}" + }, + { + "bbox": [ + 130, + 551, + 482, + 609 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 551, + 482, + 609 + ], + "type": "inline_equation", + "content": "D_K" + }, + { + "bbox": [ + 130, + 551, + 482, + 609 + ], + "type": "text", + "content": ", this pattern holds true for all the buffers." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 624, + 237, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 624, + 237, + 635 + ], + "spans": [ + { + "bbox": [ + 132, + 624, + 237, + 635 + ], + "type": "text", + "content": "3.4 Domain Frozen" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "content": "Another factor that may lead to a decrease in the model's generalization capability lies in the learnable affine parameters of the normalization layers. Compared" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 268, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 268, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 268, + 102 + ], + "type": "text", + "content": "R. Nie and J. Ding et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 480, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 480, + 175 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 480, + 175 + ], + "type": "text", + "content": "to the weight matrix, the number of learnable affine parameters is significantly smaller, implying a weaker robustness when handling the target domain with a distribution divergence from the source domains. Moreover, ReID can be approximated as a channel-wise classification task, indicating that channel-wise affine parameters exert a more profound influence." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 176, + 481, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 176, + 481, + 236 + ], + "spans": [ + { + "bbox": [ + 130, + 176, + 481, + 236 + ], + "type": "text", + "content": "Furthermore, as shown in Fig.3, we visualize the affine parameters for each source domain, approximating the domain-specific information within. It is evident that significant disparities exist among different source domains. These channel-wise divergences will accumulate into the weight matrix through backward propagation, leading to confusion in feature learning." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 236, + 482, + 343 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 236, + 482, + 343 + ], + "spans": [ + { + "bbox": [ + 130, + 236, + 482, + 343 + ], + "type": "text", + "content": "Consequently, we propose Domain Frozen (DF), a strategy that stabilizes the weight matrix by freezing the updates of affine parameters in every normalization layer. For Remix Normalization, we maintain the affine parameters in a frozen state following their initialization with pre-trained weights. In the case of Emulation Normalization, the input feature map " + }, + { + "bbox": [ + 130, + 236, + 482, + 343 + ], + "type": "inline_equation", + "content": "X_{k}" + }, + { + "bbox": [ + 130, + 236, + 482, + 343 + ], + "type": "text", + "content": " is standardized by statistical parameters from two different domains. The output " + }, + { + "bbox": [ + 130, + 236, + 482, + 343 + ], + "type": "inline_equation", + "content": "\\hat{X}_k" + }, + { + "bbox": [ + 130, + 236, + 482, + 343 + ], + "type": "text", + "content": " inherently deviates from a normal distribution. Consequently, with the integration of Domain Frozen, we have opted to directly eliminate the original affine parameters, which is tantamount to setting the affine parameters " + }, + { + "bbox": [ + 130, + 236, + 482, + 343 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 130, + 236, + 482, + 343 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 236, + 482, + 343 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 130, + 236, + 482, + 343 + ], + "type": "text", + "content": " to 1 and 0, respectively." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 343, + 482, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 343, + 482, + 403 + ], + "spans": [ + { + "bbox": [ + 130, + 343, + 482, + 403 + ], + "type": "text", + "content": "Moreover, a further apparent advantage of DF is that it indirectly shares the affine parameters between the source domains (during training) and the target domain (during testing). This sharing further mitigates the domain shifts caused by the absence of target domain information, enhancing the model's ability to generalize to unseen domains." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 418, + 263, + 431 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 418, + 263, + 431 + ], + "spans": [ + { + "bbox": [ + 131, + 418, + 263, + 431 + ], + "type": "text", + "content": "3.5 Training and testing" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 437, + 482, + 507 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 437, + 482, + 507 + ], + "spans": [ + { + "bbox": [ + 130, + 437, + 482, + 507 + ], + "type": "text", + "content": "As illustrated in Fig.1(b), we propose a novel ReNorm framework to integrate the aforementioned normalization strategies. For clarification purposes, we refer to the weight matrix equipped with Remix Normalization (RN) as Network- " + }, + { + "bbox": [ + 130, + 437, + 482, + 507 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 130, + 437, + 482, + 507 + ], + "type": "text", + "content": " and the same weight matrix equipped with Emulation Normalization (EN) as Network- " + }, + { + "bbox": [ + 130, + 437, + 482, + 507 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 130, + 437, + 482, + 507 + ], + "type": "text", + "content": ". Meanwhile, Domain Frozen (DF) is applied uniformly in both RN and EN." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 509, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 509, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 509, + 482, + 666 + ], + "type": "text", + "content": "Model Training. In our methodology, there is no need for additional design of loss functions. The employment of two fundamental ReID loss functions, triplet loss " + }, + { + "bbox": [ + 130, + 509, + 482, + 666 + ], + "type": "inline_equation", + "content": "(L_{tri})" + }, + { + "bbox": [ + 130, + 509, + 482, + 666 + ], + "type": "text", + "content": " [13] and cross-entropy loss " + }, + { + "bbox": [ + 130, + 509, + 482, + 666 + ], + "type": "inline_equation", + "content": "(L_{ce})" + }, + { + "bbox": [ + 130, + 509, + 482, + 666 + ], + "type": "text", + "content": " suffices for achieving optimal performance. In each iteration, we randomly select one source domain " + }, + { + "bbox": [ + 130, + 509, + 482, + 666 + ], + "type": "inline_equation", + "content": "D_{k}" + }, + { + "bbox": [ + 130, + 509, + 482, + 666 + ], + "type": "text", + "content": " from a set of " + }, + { + "bbox": [ + 130, + 509, + 482, + 666 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 509, + 482, + 666 + ], + "type": "text", + "content": " source domains and randomly sample a mini-batch samples of size " + }, + { + "bbox": [ + 130, + 509, + 482, + 666 + ], + "type": "inline_equation", + "content": "N_{b}" + }, + { + "bbox": [ + 130, + 509, + 482, + 666 + ], + "type": "text", + "content": " from this domain. As illustrated in Fig.1, this set of data is then fed through the Network " + }, + { + "bbox": [ + 130, + 509, + 482, + 666 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 130, + 509, + 482, + 666 + ], + "type": "text", + "content": " for first forward propagation and calculate corresponding triplet loss " + }, + { + "bbox": [ + 130, + 509, + 482, + 666 + ], + "type": "inline_equation", + "content": "L_{tri}^{R}" + }, + { + "bbox": [ + 130, + 509, + 482, + 666 + ], + "type": "text", + "content": " and cross-entropy loss " + }, + { + "bbox": [ + 130, + 509, + 482, + 666 + ], + "type": "inline_equation", + "content": "L_{ce}^{R}" + }, + { + "bbox": [ + 130, + 509, + 482, + 666 + ], + "type": "text", + "content": ", which is represented by the solid line. Subsequently, we treat the current source domain as a simulated target domain and randomly select one " + }, + { + "bbox": [ + 130, + 509, + 482, + 666 + ], + "type": "inline_equation", + "content": "D_{j}" + }, + { + "bbox": [ + 130, + 509, + 482, + 666 + ], + "type": "text", + "content": " from the remaining source domains to serve as a simulated source domain, providing Batch Normalization (BN) statistical parameters (i.e., " + }, + { + "bbox": [ + 130, + 509, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mu_j^{pop},\\sigma_j^{pop^2}" + }, + { + "bbox": [ + 130, + 509, + 482, + 666 + ], + "type": "text", + "content": ") for the second forward propagation, which is indicated by the dashed line. The same mini-batch will be fed into Network " + }, + { + "bbox": [ + 130, + 509, + 482, + 666 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 130, + 509, + 482, + 666 + ], + "type": "text", + "content": " to compute" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 253, + 91, + 445, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 253, + 91, + 445, + 102 + ], + "spans": [ + { + "bbox": [ + 253, + 91, + 445, + 102 + ], + "type": "text", + "content": "Rethinking Normalization Layers for DG-ReID" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 115, + 479, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 115, + 479, + 152 + ], + "spans": [ + { + "bbox": [ + 130, + 115, + 479, + 152 + ], + "type": "text", + "content": "the corresponding triplet loss " + }, + { + "bbox": [ + 130, + 115, + 479, + 152 + ], + "type": "inline_equation", + "content": "L_{tri}^{E}" + }, + { + "bbox": [ + 130, + 115, + 479, + 152 + ], + "type": "text", + "content": " and cross-entropy loss " + }, + { + "bbox": [ + 130, + 115, + 479, + 152 + ], + "type": "inline_equation", + "content": "L_{ce}^{E}" + }, + { + "bbox": [ + 130, + 115, + 479, + 152 + ], + "type": "text", + "content": ". Combing these two networks mentioned above, we have the following total loss for backward propagation:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 237, + 152, + 480, + 165 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 237, + 152, + 480, + 165 + ], + "spans": [ + { + "bbox": [ + 237, + 152, + 480, + 165 + ], + "type": "interline_equation", + "content": "L _ {t o t a l} = L _ {t r i} ^ {R} + L _ {c e} ^ {R} + L _ {t r i} ^ {E} + L _ {c e} ^ {E}. \\tag {8}", + "image_path": "2abba124225cb906bd3740416a577a933edb7f62dcdfed6fc717fd6139cdd738.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 170, + 479, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 170, + 479, + 205 + ], + "spans": [ + { + "bbox": [ + 130, + 170, + 479, + 205 + ], + "type": "text", + "content": "Testing. During testing, we feed the target domain to both networks " + }, + { + "bbox": [ + 130, + 170, + 479, + 205 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 170, + 479, + 205 + ], + "type": "text", + "content": " times each. Subsequently, we compute the average of the " + }, + { + "bbox": [ + 130, + 170, + 479, + 205 + ], + "type": "inline_equation", + "content": "2*K" + }, + { + "bbox": [ + 130, + 170, + 479, + 205 + ], + "type": "text", + "content": " features obtained from two networks to obtain the final inference result." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 224, + 230, + 238 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 224, + 230, + 238 + ], + "spans": [ + { + "bbox": [ + 131, + 224, + 230, + 238 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 249, + 271, + 261 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 249, + 271, + 261 + ], + "spans": [ + { + "bbox": [ + 131, + 249, + 271, + 261 + ], + "type": "text", + "content": "4.1 Datasets and Settings" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 269, + 479, + 340 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 269, + 479, + 340 + ], + "spans": [ + { + "bbox": [ + 130, + 269, + 479, + 340 + ], + "type": "text", + "content": "Datesets. Our approach is evaluated on 9 person ReID datasets, including 5 large-scale datasets: Market1501 (M) [47], MSMT17 (MS) [39], CUHK02 (C2) [22], CUHK03 (C3) [23], and CUHK-SYSU (CS) [42], and 4 small-scale datasets: PRID [14], GRID [27], VIPeR [7], and iLIDs [48]. We do not use DukeMTMCreID (D) [50] since it has been retracted. We adopt Cumulative Matching Characteristics (CMC) and mean Average Precision (mAP) as evaluation metrics." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 341, + 480, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 341, + 480, + 472 + ], + "spans": [ + { + "bbox": [ + 130, + 341, + 480, + 472 + ], + "type": "text", + "content": "Evaluation Protocols. We set three evaluation protocols following [37,44]. For Protocol-1, all images including both training and testing sets from four large-scale datasets (i.e., M, CS, C2, C3) are used for training, and the four small-scale datasets (i.e., PRID, GRID, VIPeR, iLIDs) are used for testing. The final results for each dataset are evaluated on the average of 10 repeated random splits of gallery and query sets. For Protocol-2 and Protocol-3, following the leave-one-out setting in [44], we select one domain from " + }, + { + "bbox": [ + 130, + 341, + 480, + 472 + ], + "type": "inline_equation", + "content": "\\mathrm{M} + \\mathrm{MS} + \\mathrm{CS} + \\mathrm{C3}" + }, + { + "bbox": [ + 130, + 341, + 480, + 472 + ], + "type": "text", + "content": " for testing, while the remaining three domains serve as the training sets. CS is excluded as a target domain due to its images all come from one camera view. Additionally, under Protocol-2, only the training sets of the source domains are used for training, while all images are exploited under Protocol-3." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 472, + 480, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 472, + 480, + 567 + ], + "spans": [ + { + "bbox": [ + 130, + 472, + 480, + 567 + ], + "type": "text", + "content": "Implementation Details. We employ ResNet50 [9] pre-trained on ImageNet as our backbone. The size of a mini-batch is set to 64, including 16 identities and 4 images per identity. Each image is resized to " + }, + { + "bbox": [ + 130, + 472, + 480, + 567 + ], + "type": "inline_equation", + "content": "256 \\times 128" + }, + { + "bbox": [ + 130, + 472, + 480, + 567 + ], + "type": "text", + "content": ". We use random horizontal flipping, random cropping, zero padding, color jittering, and auto-augmentation [11] as data augmentation. We also adopt a warm-up strategy in the first 10 epochs. The learning rate is initialized as " + }, + { + "bbox": [ + 130, + 472, + 480, + 567 + ], + "type": "inline_equation", + "content": "3.5 \\times 10^{-4}" + }, + { + "bbox": [ + 130, + 472, + 480, + 567 + ], + "type": "text", + "content": " and divided by 10 at the 30th, 60th and, 90th epochs, respectively. All experiments are conducted using Pytorch on a single RTX 4090 GPU." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 586, + 356, + 597 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 586, + 356, + 597 + ], + "spans": [ + { + "bbox": [ + 131, + 586, + 356, + 597 + ], + "type": "text", + "content": "4.2 Comparison with the State-of-the-Arts" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 605, + 479, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 479, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 479, + 665 + ], + "type": "text", + "content": "Comparison with DG-ReID Methods Under Protocol-1. As shown in Table 1, we compare our method with other state-of-the-arts under Protocol-1. Since DukeMTMC-reID (D) [50] has been retracted, we only use other large-scale datasets for training. Notably, our method demonstrates superior performance on PRID, GRID, and VIPeR. Overall, ReNorm significantly outperforms other" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 268, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 268, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 268, + 102 + ], + "type": "text", + "content": "R. Nie and J. Ding et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 135, + 156, + 481, + 273 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "type": "text", + "content": "Table 1: Comparison with the SOTA methods in DG ReID under the setting of Protocol-1. All results are the average of 10 random sampling. The best and the second best results are highlighted in bold and underline." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 135, + 156, + 481, + 273 + ], + "lines": [ + { + "bbox": [ + 135, + 156, + 481, + 273 + ], + "spans": [ + { + "bbox": [ + 135, + 156, + 481, + 273 + ], + "type": "table", + "html": "
MethodReferenceSource domainsTarget domainAverage
PRIDGRIDVIPeRiLIDs
mAPRank-1mAPRank-1mAPRank-1mAPRank-1mAPRank-1
SNR [19]CVPR'20M+D+C266.552.147.740.261.352.989.984.166.457.3
RaMoE [4]CVPR'2167.357.754.246.864.656.690.285.069.161.5
DMG-Net [1]CVPR'21+C3+CS68.460.656.651.060.453.983.979.367.361.2
DTIN-Net [18]ECCV'2279.771.060.651.870.762.987.281.874.666.9
MetaBIN [3]CVPR'21M+C270.861.257.950.264.355.982.774.768.960.5
META [44]ECCV'2271.761.960.152.468.461.583.579.268.960.5
ACL [45]ECCV'2273.463.065.755.275.166.486.581.875.266.6
CINorm [2]TMM'23+C3+CS80.572.459.949.476.767.688.082.276.267.9
PAOA [21]WACV'2474.065.667.256.376.666.787.183.176.267.9
ReNormThis paper81.373.967.559.477.069.187.082.278.271.2
", + "image_path": "cfdcbb0a6e66c96d2972cf2cb86bd33e0af1d12c80ab89fa06f710d0a68caec4.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 297, + 479, + 333 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 297, + 479, + 333 + ], + "spans": [ + { + "bbox": [ + 130, + 297, + 479, + 333 + ], + "type": "text", + "content": "methods by at least " + }, + { + "bbox": [ + 130, + 297, + 479, + 333 + ], + "type": "inline_equation", + "content": "2.0\\%" + }, + { + "bbox": [ + 130, + 297, + 479, + 333 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 297, + 479, + 333 + ], + "type": "inline_equation", + "content": "3.3\\%" + }, + { + "bbox": [ + 130, + 297, + 479, + 333 + ], + "type": "text", + "content": " in average mAP and Rank-1, respectively. The experimental findings substantiate the efficacy of our ReNorm, which combines simplicity with substantial impact." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 334, + 482, + 465 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 334, + 482, + 465 + ], + "spans": [ + { + "bbox": [ + 130, + 334, + 482, + 465 + ], + "type": "text", + "content": "Comparison Under Protocol-2 and Protocol-3. We compare our method with previous state-of-the-arts under protocol-2 and protocol-3 to evaluate our method on large-scale datasets. As shown in Table 2, we can find that our ReNorm achieves superior performance compared with previous methods on almost all datasets under the two protocols. Overall, ReNorm outperforms other approaches by an average of at least " + }, + { + "bbox": [ + 130, + 334, + 482, + 465 + ], + "type": "inline_equation", + "content": "2.0\\%" + }, + { + "bbox": [ + 130, + 334, + 482, + 465 + ], + "type": "text", + "content": " mAP and " + }, + { + "bbox": [ + 130, + 334, + 482, + 465 + ], + "type": "inline_equation", + "content": "4.1\\%" + }, + { + "bbox": [ + 130, + 334, + 482, + 465 + ], + "type": "text", + "content": " Rank-1 under Protocol-2, and by " + }, + { + "bbox": [ + 130, + 334, + 482, + 465 + ], + "type": "inline_equation", + "content": "1.7\\%" + }, + { + "bbox": [ + 130, + 334, + 482, + 465 + ], + "type": "text", + "content": " mAP and " + }, + { + "bbox": [ + 130, + 334, + 482, + 465 + ], + "type": "inline_equation", + "content": "2.9\\%" + }, + { + "bbox": [ + 130, + 334, + 482, + 465 + ], + "type": "text", + "content": " Rank-1 under Protocol-3, respectively. Notably, when MSMT17 (MS) serves as the target domain, our method surpasses the second best by " + }, + { + "bbox": [ + 130, + 334, + 482, + 465 + ], + "type": "inline_equation", + "content": "3.1\\%" + }, + { + "bbox": [ + 130, + 334, + 482, + 465 + ], + "type": "text", + "content": " mAP and " + }, + { + "bbox": [ + 130, + 334, + 482, + 465 + ], + "type": "inline_equation", + "content": "5.7\\%" + }, + { + "bbox": [ + 130, + 334, + 482, + 465 + ], + "type": "text", + "content": " Rank-1 under Protocol-2, and by " + }, + { + "bbox": [ + 130, + 334, + 482, + 465 + ], + "type": "inline_equation", + "content": "2.7\\%" + }, + { + "bbox": [ + 130, + 334, + 482, + 465 + ], + "type": "text", + "content": " mAP and " + }, + { + "bbox": [ + 130, + 334, + 482, + 465 + ], + "type": "inline_equation", + "content": "5.7\\%" + }, + { + "bbox": [ + 130, + 334, + 482, + 465 + ], + "type": "text", + "content": " Rank-1 under Protocol-3. Collectively, these results further illustrate our model's superiority in domain generalization." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 486, + 238, + 498 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 486, + 238, + 498 + ], + "spans": [ + { + "bbox": [ + 132, + 486, + 238, + 498 + ], + "type": "text", + "content": "4.3 Ablation Study" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "content": "Effectiveness of different modules. We investigate the effects of different components in ReNorm design to study their individual contributions, as shown in Table 3. The first row corresponds to a basic ResNet50 framework. In the table, every even row compared to the previous row incorporates the Domain Frozen (DF) strategy, which straightforwardly boosts the network by at least " + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "inline_equation", + "content": "1.3\\%" + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "content": " mAP and " + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "inline_equation", + "content": "1.5\\%" + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "content": " Rank-1, on average. Comparing first and second rows with the third and fourth rows, even when Remix Normalization (RN) simply combines IN and BN, it significantly enhances the network's generalization performance. Furthermore, with the addition of Emulation Normalization (EN), the two complementary normalization layers further improve the model's generalization ability, resulting in an average improvement of at least " + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "inline_equation", + "content": "2.0\\%" + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "content": " mAP and " + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "inline_equation", + "content": "2.5\\%" + }, + { + "bbox": [ + 130, + 510, + 482, + 666 + ], + "type": "text", + "content": " Rank-1. The experiments thoroughly demonstrate the simplicity and effectiveness of our approach." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 252, + 90, + 445, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 252, + 90, + 445, + 102 + ], + "spans": [ + { + "bbox": [ + 252, + 90, + 445, + 102 + ], + "type": "text", + "content": "Rethinking Normalization Layers for DG-ReID" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 135, + 156, + 481, + 323 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 482, + 148 + ], + "type": "text", + "content": "Table 2: Comparison with the SOTA methods in DG ReID under the setting of Protocol-2 and Protocol-3. All results are the average of 10 random sampling. The best and the second best results are highlighted in bold and underline." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 135, + 156, + 481, + 323 + ], + "lines": [ + { + "bbox": [ + 135, + 156, + 481, + 323 + ], + "spans": [ + { + "bbox": [ + 135, + 156, + 481, + 323 + ], + "type": "table", + "html": "
MethodReferenceSettingM+MS+CS→C3M+CS+C3→MSMS+CS+C3→MAverage
mAP Rank-1mAP Rank-1mAP Rank-1mAP Rank-1mAP Rank-1mAP Rank-1mAP Rank-1mAP Rank-1
SNR [19]CVPR'20Protocol-28.98.96.819.934.662.716.830.5
M3L [46]CVPR'2120.931.915.936.958.479.931.749.6
MetaBIN [3]CVPR'2128.828.117.840.257.980.134.849.5
META [44]ECCV'2236.335.122.549.967.586.142.157.0
ACL [45]ECCV'2241.241.820.445.974.389.345.359.0
ReNormThis Paper43.644.725.655.672.789.147.363.1
SNR [19]CVPR'20Protocol-317.517.17.722.052.477.825.939.0
M3L [46]CVPR'2135.736.517.438.662.482.738.552.6
MetaBIN [3]CVPR'2143.043.118.841.267.284.543.056.3
META [44]ECCV'2247.146.224.452.176.590.549.362.9
ACL [45]ECCV'2249.950.121.747.376.890.649.362.7
PAOA [21]WACV'2449.850.525.151.577.190.850.764.3
ReNormThis Paper51.952.527.857.877.491.252.467.2
", + "image_path": "14bcc82f1472b265f3bc9d733cd84f6cb3abd6d9fcadfaf2e9f3c619ed68d9c1.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 147, + 367, + 467, + 459 + ], + "blocks": [ + { + "bbox": [ + 132, + 336, + 480, + 358 + ], + "lines": [ + { + "bbox": [ + 132, + 336, + 480, + 358 + ], + "spans": [ + { + "bbox": [ + 132, + 336, + 480, + 358 + ], + "type": "text", + "content": "Table 3: Ablation studies for different components of our framework. The experiment is conducted under protocol-2. The best results are highlighted in bold." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 147, + 367, + 467, + 459 + ], + "lines": [ + { + "bbox": [ + 147, + 367, + 467, + 459 + ], + "spans": [ + { + "bbox": [ + 147, + 367, + 467, + 459 + ], + "type": "table", + "html": "
RNENDFTarget: C3Target: MSTarget: MAverage
mAP Rank-1mAP Rank-1mAP Rank-1mAP Rank-1
35.334.919.944.966.385.140.555.0
✓36.536.922.148.568.286.642.357.3
✓40.540.522.149.469.087.143.959.0
✓✓40.440.223.752.771.988.945.360.6
✓✓42.543.624.453.471.187.946.061.6
✓✓✓43.644.725.655.672.789.147.363.1
", + "image_path": "ee17437053028f2036b20061c43a1c1e116109a3fd0b548cf0c54d0775d3d298.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 484, + 481, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 484, + 481, + 604 + ], + "spans": [ + { + "bbox": [ + 132, + 484, + 481, + 604 + ], + "type": "text", + "content": "Visualization of the domain-specific information within the affine parameters. In this experiment, we establish corresponding RN for each source domain and reinstate the updates of all affine parameters. After training, we sum the affine parameters " + }, + { + "bbox": [ + 132, + 484, + 481, + 604 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 132, + 484, + 481, + 604 + ], + "type": "text", + "content": " of each source domain at each RN layer to approximate the domain-specific information within. As shown in Fig.3, The horizontal axis represents the layer position of RN in the backbone, while the vertical axis represents the corresponding proportions. It is evident that significant disparities exist among source domains across all the normalization layers. During training, the disparities accumulate in the weight matrix, potentially leading to misguided learning." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 605, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 605, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 605, + 481, + 665 + ], + "type": "text", + "content": "Effectiveness on where to plug the modules. As normalization-based modules, the proposed ReNorm and DF can be plugged into any stage of the network by replacing the corresponding BN layers. In our experiments, we employ ResNet50 as our backbone, which consists of four groups of residual blocks and is organized into four stages. We replace our ReNorm and DF with diverse stages'" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 268, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 268, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 268, + 102 + ], + "type": "text", + "content": "R. Nie and J. Ding et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 136, + 118, + 476, + 224 + ], + "blocks": [ + { + "bbox": [ + 136, + 118, + 476, + 224 + ], + "lines": [ + { + "bbox": [ + 136, + 118, + 476, + 224 + ], + "spans": [ + { + "bbox": [ + 136, + 118, + 476, + 224 + ], + "type": "image", + "image_path": "8bcb600a2b32df48b6e33eff9ebfba688aac5dd196b5f5ceb4afb567a5787cdc.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 236, + 480, + 269 + ], + "lines": [ + { + "bbox": [ + 131, + 236, + 480, + 269 + ], + "spans": [ + { + "bbox": [ + 131, + 236, + 480, + 269 + ], + "type": "text", + "content": "Fig. 3: Visualization of the domain-specific information within the affine parameters under Protocol-2, the source domains used for training being MSMT17, CUHK-SYSU and Market1501." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 142, + 312, + 298, + 403 + ], + "blocks": [ + { + "bbox": [ + 132, + 281, + 306, + 303 + ], + "lines": [ + { + "bbox": [ + 132, + 281, + 306, + 303 + ], + "spans": [ + { + "bbox": [ + 132, + 281, + 306, + 303 + ], + "type": "text", + "content": "Table 4: Ablation study on where to plug the modules into ResNet50." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 142, + 312, + 298, + 403 + ], + "lines": [ + { + "bbox": [ + 142, + 312, + 298, + 403 + ], + "spans": [ + { + "bbox": [ + 142, + 312, + 298, + 403 + ], + "type": "table", + "html": "
StageTarget: C3
Module: ReNormModule: DF
mAPRank-1mAPRank-1
None34.935.342.543.6
439.439.843.043.3
3 - 441.942.043.343.5
2 - 442.243.343.144.0
1 - 443.644.743.644.7
", + "image_path": "670f466b2dc0af418813204d9247729ad9d7bce9baea9e1ea1bc00b74f55171f.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 323, + 312, + 478, + 403 + ], + "blocks": [ + { + "bbox": [ + 313, + 281, + 487, + 304 + ], + "lines": [ + { + "bbox": [ + 313, + 281, + 487, + 304 + ], + "spans": [ + { + "bbox": [ + 313, + 281, + 487, + 304 + ], + "type": "text", + "content": "Table 5: Ablation study on different combination of outputs into the result." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 323, + 312, + 478, + 403 + ], + "lines": [ + { + "bbox": [ + 323, + 312, + 478, + 403 + ], + "spans": [ + { + "bbox": [ + 323, + 312, + 478, + 403 + ], + "type": "table", + "html": "
MethodAverage
mAPRank-1
concat(R)46.261.8
concat(E)45.860.8
mean(R)46.361.9
mean(E)45.860.9
concat(R,E)47.062.8
mean(R,E)47.363.1
", + "image_path": "d1cd79c0750ed3afb75efd4400bec45b6095ce314778aba5f01006105353ad59.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 426, + 480, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 426, + 480, + 484 + ], + "spans": [ + { + "bbox": [ + 130, + 426, + 480, + 484 + ], + "type": "text", + "content": "BN layers to explore the effects of different plugging stages. As shown in Table 4, the results demonstrate that employing ReNorm and DF across all normalization layers performs better than others. Additionally, incorporating ReNorm and DF from any stage could improve overall performance, further demonstrating the effectiveness of our modules." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 486, + 481, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 486, + 481, + 605 + ], + "spans": [ + { + "bbox": [ + 130, + 486, + 481, + 605 + ], + "type": "text", + "content": "Combination of outputs into the final result. The experiment is conducted under Protocol-2. During testing, each image will go through Network-" + }, + { + "bbox": [ + 130, + 486, + 481, + 605 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 130, + 486, + 481, + 605 + ], + "type": "text", + "content": " (R) and Network-" + }, + { + "bbox": [ + 130, + 486, + 481, + 605 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 130, + 486, + 481, + 605 + ], + "type": "text", + "content": " (E) separately " + }, + { + "bbox": [ + 130, + 486, + 481, + 605 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 486, + 481, + 605 + ], + "type": "text", + "content": " times, resulting in a total of " + }, + { + "bbox": [ + 130, + 486, + 481, + 605 + ], + "type": "inline_equation", + "content": "2*K" + }, + { + "bbox": [ + 130, + 486, + 481, + 605 + ], + "type": "text", + "content": " outputs. As shown in Table 5, from the first four rows, it is evident that network-" + }, + { + "bbox": [ + 130, + 486, + 481, + 605 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 130, + 486, + 481, + 605 + ], + "type": "text", + "content": " generally outperforms network-" + }, + { + "bbox": [ + 130, + 486, + 481, + 605 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 130, + 486, + 481, + 605 + ], + "type": "text", + "content": ", and both have surpassed the previous SOTA with " + }, + { + "bbox": [ + 130, + 486, + 481, + 605 + ], + "type": "inline_equation", + "content": "45.3\\%" + }, + { + "bbox": [ + 130, + 486, + 481, + 605 + ], + "type": "text", + "content": " mAP and " + }, + { + "bbox": [ + 130, + 486, + 481, + 605 + ], + "type": "inline_equation", + "content": "59.0\\%" + }, + { + "bbox": [ + 130, + 486, + 481, + 605 + ], + "type": "text", + "content": " Rank-1 on average. After combining the outputs of the two networks, there is a significant improvement in the final results. Ultimately, taking the mean of the " + }, + { + "bbox": [ + 130, + 486, + 481, + 605 + ], + "type": "inline_equation", + "content": "2*K" + }, + { + "bbox": [ + 130, + 486, + 481, + 605 + ], + "type": "text", + "content": " outputs directly yields the best result with " + }, + { + "bbox": [ + 130, + 486, + 481, + 605 + ], + "type": "inline_equation", + "content": "47.3\\%" + }, + { + "bbox": [ + 130, + 486, + 481, + 605 + ], + "type": "text", + "content": " mAP and " + }, + { + "bbox": [ + 130, + 486, + 481, + 605 + ], + "type": "inline_equation", + "content": "63.1\\%" + }, + { + "bbox": [ + 130, + 486, + 481, + 605 + ], + "type": "text", + "content": " Rank-1. Experimental results confirm the enhanced generalization capability of our framework in domain-invariant feature extraction." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 605, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 481, + 665 + ], + "type": "text", + "content": "Visualization of the target domain distribution. To further demonstrate the effectiveness of our ReNorm framework, we visualize the feature distribution of unseen target domains under Protocol-1. We conduct t-SNE [30] experiments on the same query and gallery image pairs from four target domains by implementing BN, IN, and our ReNorm, respectively. In Fig.4 (a), the" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 253, + 91, + 445, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 253, + 91, + 445, + 102 + ], + "spans": [ + { + "bbox": [ + 253, + 91, + 445, + 102 + ], + "type": "text", + "content": "Rethinking Normalization Layers for DG-ReID" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 136, + 114, + 239, + 217 + ], + "blocks": [ + { + "bbox": [ + 136, + 114, + 239, + 217 + ], + "lines": [ + { + "bbox": [ + 136, + 114, + 239, + 217 + ], + "spans": [ + { + "bbox": [ + 136, + 114, + 239, + 217 + ], + "type": "image", + "image_path": "a2237fb5379cbc87bf9176ae98899e168605bf1247f29f20f597f214b58ca8b3.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 173, + 220, + 199, + 230 + ], + "lines": [ + { + "bbox": [ + 173, + 220, + 199, + 230 + ], + "spans": [ + { + "bbox": [ + 173, + 220, + 199, + 230 + ], + "type": "text", + "content": "(a) BN" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 256, + 114, + 359, + 217 + ], + "blocks": [ + { + "bbox": [ + 256, + 114, + 359, + 217 + ], + "lines": [ + { + "bbox": [ + 256, + 114, + 359, + 217 + ], + "spans": [ + { + "bbox": [ + 256, + 114, + 359, + 217 + ], + "type": "image", + "image_path": "f3105fdb03b99fb47ff956e2397506bf429106bf2710ad4846e04b12f4777628.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 294, + 220, + 320, + 230 + ], + "lines": [ + { + "bbox": [ + 294, + 220, + 320, + 230 + ], + "spans": [ + { + "bbox": [ + 294, + 220, + 320, + 230 + ], + "type": "text", + "content": "(b) IN" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 239, + 480, + 294 + ], + "lines": [ + { + "bbox": [ + 130, + 239, + 480, + 294 + ], + "spans": [ + { + "bbox": [ + 130, + 239, + 480, + 294 + ], + "type": "text", + "content": "Fig. 4: The t-SNE results of features under Protocol-1 on three models with different normalization layers. (a) indicates ResNet50 with BN for all normalization layers, (b) denotes using instance normalization, and (c) utilizes our ReNorm for all normalization layers. Various colors illustrate samples from different target domains, and different shapes denote sample pairs from query or gallery sets." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 376, + 114, + 480, + 217 + ], + "blocks": [ + { + "bbox": [ + 376, + 114, + 480, + 217 + ], + "lines": [ + { + "bbox": [ + 376, + 114, + 480, + 217 + ], + "spans": [ + { + "bbox": [ + 376, + 114, + 480, + 217 + ], + "type": "image", + "image_path": "a181018657d9b21f620839e461fbcdaf07355feb381fac66458922e675d8a97f.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 394, + 220, + 461, + 230 + ], + "lines": [ + { + "bbox": [ + 394, + 220, + 461, + 230 + ], + "spans": [ + { + "bbox": [ + 394, + 220, + 461, + 230 + ], + "type": "text", + "content": "(c) ReNorm (ours)" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 323, + 482, + 455 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 323, + 482, + 455 + ], + "spans": [ + { + "bbox": [ + 130, + 323, + 482, + 455 + ], + "type": "text", + "content": "\"BN\" model exhibits smaller distances between query and gallery sets of the same identity but is heavily influenced by domain factors, leading to clear domain boundaries. In contrast, Fig.4 (b) shows that IN removes these boundaries but increases the distance between query and gallery sample pairs. The observations prove that BN can capture more discriminative features but tends to overfit to source domains, while IN can eliminate domain-specific information to enhance generalization but inevitably removes discriminative information. Fig.4 (c) shows that the distances between query and gallery sample pairs are much smaller, demonstrating that our approach achieves better performance. Simultaneously, there are no apparent boundaries among target domains, indicating that our method effectively eliminates domain-specific information." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 484, + 219, + 498 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 484, + 219, + 498 + ], + "spans": [ + { + "bbox": [ + 132, + 484, + 219, + 498 + ], + "type": "text", + "content": "5 Conclusion" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 521, + 487, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 487, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 487, + 665 + ], + "type": "text", + "content": "In this paper, we present a purely normalization-based framework called ReNorm for tackling the problem of domain generalizable person re-identification (DG-ReID). Our framework integrates two complementary normalization layers through two forward propagations for the same weight matrix. Remix Normalization (RN) enables the weight matrix to extract discriminative features within the source domains, while Emulation Normalization (EN) is employed to actively guide the weight matrix to learn how to generalize the feature extraction capability to unseen target domains. Meanwhile, we introduce Domain Frozen (DF), a strategy that stabilizes the weight matrix by freezing the updates of affine parameters in normalization layers. Extensive experiments demonstrate the simplicity yet state-of-the-art performance of ReNorm among all the popular benchmarks." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 268, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 268, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 268, + 102 + ], + "type": "text", + "content": "R. Nie and J. Ding et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 140, + 488, + 201 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 140, + 488, + 201 + ], + "spans": [ + { + "bbox": [ + 130, + 140, + 488, + 201 + ], + "type": "text", + "content": "This work was supported by the Natural Science Foundation of China (No. 62372082, 61972071, U20A20184), Natural Science Foundation of Sichuan Province (No.2023NSFSC0485), National Science Foundation for Distinguished Young Scholars under Grant 62225605 and Zhejiang Key Research and Development Program under Grant 2023C03196." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 220, + 197, + 232 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 220, + 197, + 232 + ], + "spans": [ + { + "bbox": [ + 133, + 220, + 197, + 232 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 134, + 246, + 481, + 666 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 138, + 246, + 481, + 290 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 246, + 481, + 290 + ], + "spans": [ + { + "bbox": [ + 138, + 246, + 481, + 290 + ], + "type": "text", + "content": "1. Bai, Y., Jiao, J., Ce, W., Liu, J., Lou, Y., Feng, X., Duan, L.Y.: Person30k: A dual-meta generalization network for person re-identification. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2123-2132 (2021)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 291, + 481, + 323 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 291, + 481, + 323 + ], + "spans": [ + { + "bbox": [ + 138, + 291, + 481, + 323 + ], + "type": "text", + "content": "2. Chen, Z., Wang, W., Zhao, Z., Su, F., Men, A., Dong, Y.: Cluster-instance normalization: A statistical relation-aware normalization for generalizable person re-identification. IEEE Transactions on Multimedia (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 324, + 481, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 324, + 481, + 357 + ], + "spans": [ + { + "bbox": [ + 138, + 324, + 481, + 357 + ], + "type": "text", + "content": "3. Choi, S., Kim, T., Jeong, M., Park, H., Kim, C.: Meta batch-instance normalization for generalizable person re-identification. In: Proceedings of the IEEE/CVF conference on Computer Vision and Pattern Recognition. pp. 3425-3435 (2021)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 357, + 481, + 389 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 357, + 481, + 389 + ], + "spans": [ + { + "bbox": [ + 138, + 357, + 481, + 389 + ], + "type": "text", + "content": "4. Dai, Y., Li, X., Liu, J., Tong, Z., Duan, L.Y.: Generalizable person re-identification with relevance-aware mixture of experts. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16145-16154 (2021)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 390, + 481, + 423 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 390, + 481, + 423 + ], + "spans": [ + { + "bbox": [ + 138, + 390, + 481, + 423 + ], + "type": "text", + "content": "5. Dai, Y., Liu, J., Sun, Y., Tong, Z., Zhang, C., Duan, L.Y.: Idm: An intermediate domain module for domain adaptive person re-id. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 11864-11874 (2021)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 423, + 481, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 423, + 481, + 456 + ], + "spans": [ + { + "bbox": [ + 138, + 423, + 481, + 456 + ], + "type": "text", + "content": "6. Fu, D., Chen, D., Bao, J., Yang, H., Yuan, L., Zhang, L., Li, H., Chen, D.: Unsupervised pre-training for person re-identification. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 14750-14759 (2021)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 456, + 481, + 478 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 456, + 481, + 478 + ], + "spans": [ + { + "bbox": [ + 138, + 456, + 481, + 478 + ], + "type": "text", + "content": "7. Gray, D., Tao, H.: Viewpoint invariant pedestrian recognition with an ensemble of localized features. ECCV (1) 2008, 262-275 (2008)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 479, + 481, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 479, + 481, + 511 + ], + "spans": [ + { + "bbox": [ + 138, + 479, + 481, + 511 + ], + "type": "text", + "content": "8. Han, K., Si, C., Huang, Y., Wang, L., Tan, T.: Generalizable person re-identification via self-supervised batch norm test-time adaption. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 36, pp. 817-825 (2022)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 512, + 481, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 512, + 481, + 544 + ], + "spans": [ + { + "bbox": [ + 138, + 512, + 481, + 544 + ], + "type": "text", + "content": "9. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 770-778 (2016)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 544, + 481, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 544, + 481, + 577 + ], + "spans": [ + { + "bbox": [ + 138, + 544, + 481, + 577 + ], + "type": "text", + "content": "0. He, L., Liang, J., Li, H., Sun, Z.: Deep spatial feature reconstruction for partial person re-identification: Alignment-free approach. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 7073-7082 (2018)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 577, + 481, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 577, + 481, + 599 + ], + "spans": [ + { + "bbox": [ + 138, + 577, + 481, + 599 + ], + "type": "text", + "content": "1. He, L., Liao, X., Liu, W., Liu, X., Cheng, P., Mei, T.: Fastreid: A python toolbox for general instance re-identification. arXiv preprint arXiv:2006.02631 (2020)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 134, + 600, + 481, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 600, + 481, + 643 + ], + "spans": [ + { + "bbox": [ + 134, + 600, + 481, + 643 + ], + "type": "text", + "content": "12. He, S., Luo, H., Chen, W., Zhang, M., Zhang, Y., Wang, F., Li, H., Jiang, W.: Multi-domain learning and identity mining for vehicle re-identification. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops. pp. 582-583 (2020)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 134, + 644, + 481, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 644, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 134, + 644, + 481, + 666 + ], + "type": "text", + "content": "13. Hermans, A., Beyer, L., Leibe, B.: In defense of the triplet loss for person re-identification. arXiv preprint arXiv:1703.07737 (2017)" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 252, + 91, + 445, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 252, + 91, + 445, + 102 + ], + "spans": [ + { + "bbox": [ + 252, + 91, + 445, + 102 + ], + "type": "text", + "content": "Rethinking Normalization Layers for DG-ReID" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 481, + 161 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 481, + 161 + ], + "type": "text", + "content": "14. Hirzer, M., Beleznai, C., Roth, P.M., Bischof, H.: Person re-identification by descriptive and discriminative classification. In: Image Analysis: 17th Scandinavian Conference, SCIA 2011, Ystad, Sweden, May 2011. Proceedings 17. pp. 91-102. Springer (2011)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 162, + 481, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 162, + 481, + 194 + ], + "spans": [ + { + "bbox": [ + 132, + 162, + 481, + 194 + ], + "type": "text", + "content": "15. Huang, X., Belongie, S.: Arbitrary style transfer in real-time with adaptive instance normalization. In: Proceedings of the IEEE international conference on computer vision. pp. 1501-1510 (2017)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 194, + 481, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 194, + 481, + 226 + ], + "spans": [ + { + "bbox": [ + 133, + 194, + 481, + 226 + ], + "type": "text", + "content": "16. Ioffe, S., Szegedy, C.: Batch normalization: Accelerating deep network training by reducing internal covariate shift. In: International conference on machine learning. pp. 448-456. pmlr (2015)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 134, + 228, + 481, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 228, + 481, + 258 + ], + "spans": [ + { + "bbox": [ + 134, + 228, + 481, + 258 + ], + "type": "text", + "content": "17. Jiang, K., Zhang, T., Zhang, Y., Wu, F., Rui, Y.: Self-supervised agent learning for unsupervised cross-domain person re-identification. IEEE Transactions on Image Processing 29, 8549–8560 (2020)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 134, + 259, + 481, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 259, + 481, + 303 + ], + "spans": [ + { + "bbox": [ + 134, + 259, + 481, + 303 + ], + "type": "text", + "content": "18. Jiao, B., Liu, L., Gao, L., Lin, G., Yang, L., Zhang, S., Wang, P., Zhang, Y.: Dynamically transformed instance normalization network for generalizable person re-identification. In: European Conference on Computer Vision. pp. 285–301. Springer (2022)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 134, + 304, + 481, + 336 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 304, + 481, + 336 + ], + "spans": [ + { + "bbox": [ + 134, + 304, + 481, + 336 + ], + "type": "text", + "content": "19. Jin, X., Lan, C., Zeng, W., Chen, Z., Zhang, L.: Style normalization and restitution for generalizable person re-identification. In: proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 3143-3152 (2020)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 133, + 337, + 481, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 337, + 481, + 380 + ], + "spans": [ + { + "bbox": [ + 133, + 337, + 481, + 380 + ], + "type": "text", + "content": "20. Lee, G., Lee, S., Kim, D., Shin, Y., Yoon, Y., Ham, B.: Camera-driven representation learning for unsupervised domain adaptive person re-identification. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 11453-11462 (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 133, + 380, + 481, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 380, + 481, + 413 + ], + "spans": [ + { + "bbox": [ + 133, + 380, + 481, + 413 + ], + "type": "text", + "content": "21. Li, Q., Gong, S.: Mitigate domain shift by primary-auxiliary objectives association for generalizing person reid. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 394-403 (2024)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 133, + 414, + 481, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 414, + 481, + 445 + ], + "spans": [ + { + "bbox": [ + 133, + 414, + 481, + 445 + ], + "type": "text", + "content": "22. Li, W., Wang, X.: Locally aligned feature transforms across views. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 3594-3601 (2013)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 133, + 447, + 481, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 447, + 481, + 479 + ], + "spans": [ + { + "bbox": [ + 133, + 447, + 481, + 479 + ], + "type": "text", + "content": "23. Li, W., Zhao, R., Xiao, T., Wang, X.: Deepreid: Deep filter pairing neural network for person re-identification. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 152-159 (2014)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 133, + 479, + 481, + 501 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 479, + 481, + 501 + ], + "spans": [ + { + "bbox": [ + 133, + 479, + 481, + 501 + ], + "type": "text", + "content": "24. Li, Y., Wang, N., Shi, J., Hou, X., Liu, J.: Adaptive batch normalization for practical domain adaptation. Pattern Recognition 80, 109-117 (2018)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 133, + 502, + 481, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 502, + 481, + 544 + ], + "spans": [ + { + "bbox": [ + 133, + 502, + 481, + 544 + ], + "type": "text", + "content": "25. Lin, C.S., Cheng, Y.C., Wang, Y.C.F.: Domain generalized person re-identification via cross-domain episodic learning. In: 2020 25th International Conference on Pattern Recognition (ICPR). pp. 6758-6763 (2021). https://doi.org/10.1109/ICPR48806.2021.9413013" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 133, + 545, + 481, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 545, + 481, + 578 + ], + "spans": [ + { + "bbox": [ + 133, + 545, + 481, + 578 + ], + "type": "text", + "content": "26. Liu, J., Zha, Z.J., Chen, D., Hong, R., Wang, M.: Adaptive transfer network for cross-domain person re-identification. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 7202-7211 (2019)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 133, + 578, + 481, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 578, + 481, + 610 + ], + "spans": [ + { + "bbox": [ + 133, + 578, + 481, + 610 + ], + "type": "text", + "content": "27. Loy, C.C., Xiang, T., Gong, S.: Time-delayed correlation analysis for multi-camera activity understanding. International Journal of Computer Vision 90, 106-129 (2010)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 133, + 611, + 481, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 611, + 481, + 643 + ], + "spans": [ + { + "bbox": [ + 133, + 611, + 481, + 643 + ], + "type": "text", + "content": "28. Luo, H., Jiang, W., Gu, Y., Liu, F., Liao, X., Lai, S., Gu, J.: A strong baseline and batch normalization neck for deep person re-identification. IEEE Transactions on Multimedia 22(10), 2597-2609 (2019)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 133, + 643, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 643, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 643, + 481, + 665 + ], + "type": "text", + "content": "29. Luo, P., Ren, J., Peng, Z., Zhang, R., Li, J.: Differentiable learning-to-normalize via switchable normalization. arXiv preprint arXiv:1806.10779 (2018)" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 268, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 268, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 268, + 102 + ], + "type": "text", + "content": "R. Nie and J. Ding et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 665 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 138 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 138 + ], + "type": "text", + "content": "30. Van der Maaten, L., Hinton, G.: Visualizing data using t-sne. Journal of machine learning research 9(11) (2008)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 138, + 481, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 138, + 481, + 160 + ], + "spans": [ + { + "bbox": [ + 130, + 138, + 481, + 160 + ], + "type": "text", + "content": "31. Nam, H., Kim, H.E.: Batch-instance normalization for adaptively style-invariant neural networks. Advances in Neural Information Processing Systems 31 (2018)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 160, + 481, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 160, + 481, + 194 + ], + "spans": [ + { + "bbox": [ + 130, + 160, + 481, + 194 + ], + "type": "text", + "content": "32. Ni, H., Song, J., Luo, X., Zheng, F., Li, W., Shen, H.T.: Meta distribution alignment for generalizable person re-identification. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2487-2496 (2022)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 194, + 481, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 194, + 481, + 226 + ], + "spans": [ + { + "bbox": [ + 130, + 194, + 481, + 226 + ], + "type": "text", + "content": "33. Pan, X., Luo, P., Shi, J., Tang, X.: Two at once: Enhancing learning and generalization capacities via ibn-net. In: Proceedings of the European Conference on Computer Vision (ECCV). pp. 464-479 (2018)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 226, + 481, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 226, + 481, + 258 + ], + "spans": [ + { + "bbox": [ + 130, + 226, + 481, + 258 + ], + "type": "text", + "content": "34. Qi, L., Wang, L., Shi, Y., Geng, X.: A novel mix-normalization method for generalizable multi-source person re-identification. IEEE Transactions on Multimedia (2022)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 258, + 481, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 258, + 481, + 293 + ], + "spans": [ + { + "bbox": [ + 130, + 258, + 481, + 293 + ], + "type": "text", + "content": "35. Ren, M., He, L., Liao, X., Liu, W., Wang, Y., Tan, T.: Learning instance-level spatial-temporal patterns for person re-identification. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 14930-14939 (2021)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 293, + 481, + 336 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 293, + 481, + 336 + ], + "spans": [ + { + "bbox": [ + 130, + 293, + 481, + 336 + ], + "type": "text", + "content": "36. Seo, S., Suh, Y., Kim, D., Kim, G., Han, J., Han, B.: Learning to optimize domain specific normalization for domain generalization. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXII 16. pp. 68-83. Springer (2020)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 336, + 481, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 336, + 481, + 380 + ], + "spans": [ + { + "bbox": [ + 130, + 336, + 481, + 380 + ], + "type": "text", + "content": "37. Song, J., Yang, Y., Song, Y.Z., Xiang, T., Hospedales, T.M.: Generalizable person re-identification by domain-invariant mapping network. In: Proceedings of the IEEE/CVF conference on Computer Vision and Pattern Recognition. pp. 719-728 (2019)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 380, + 481, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 380, + 481, + 403 + ], + "spans": [ + { + "bbox": [ + 130, + 380, + 481, + 403 + ], + "type": "text", + "content": "38. Ulyanov, D., Vedaldi, A., Lempitsky, V.: Instance normalization: The missing ingredient for fast stylization. arXiv preprint arXiv:1607.08022 (2016)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 403, + 481, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 403, + 481, + 435 + ], + "spans": [ + { + "bbox": [ + 130, + 403, + 481, + 435 + ], + "type": "text", + "content": "39. Wei, L., Zhang, S., Gao, W., Tian, Q.: Person transfer gan to bridge domain gap for person re-identification. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 79-88 (2018)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 435, + 481, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 435, + 481, + 468 + ], + "spans": [ + { + "bbox": [ + 130, + 435, + 481, + 468 + ], + "type": "text", + "content": "40. Wu, Y., Chi, Z., Wang, Y., Plataniotis, K.N., Feng, S.: Test-time domain adaptation by learning domain-aware batch normalization. arXiv preprint arXiv:2312.10165 (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 468, + 481, + 490 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 468, + 481, + 490 + ], + "spans": [ + { + "bbox": [ + 130, + 468, + 481, + 490 + ], + "type": "text", + "content": "41. Wu, Y., Johnson, J.: Rethinking\" batch\" in batchnorm. arXiv preprint arXiv:2105.07576 (2021)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 490, + 481, + 512 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 490, + 481, + 512 + ], + "spans": [ + { + "bbox": [ + 130, + 490, + 481, + 512 + ], + "type": "text", + "content": "42. Xiao, T., Li, S., Wang, B., Lin, L., Wang, X.: End-to-end deep learning for person search. arXiv preprint arXiv:1604.01850 2(2), 4 (2016)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 512, + 481, + 545 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 512, + 481, + 545 + ], + "spans": [ + { + "bbox": [ + 130, + 512, + 481, + 545 + ], + "type": "text", + "content": "43. Xu, B., He, L., Liao, X., Liu, W., Sun, Z., Mei, T.: Black re-id: A head-shoulder descriptor for the challenging problem of person re-identification. In: Proceedings of the 28th ACM International Conference on Multimedia. pp. 673–681 (2020)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 545, + 481, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 481, + 589 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 481, + 589 + ], + "type": "text", + "content": "44. Xu, B., Liang, J., He, L., Sun, Z.: Mimic embedding via adaptive aggregation: Learning generalizable person re-identification. In: Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XIV. pp. 372-388. Springer (2022)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 130, + 589, + 481, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 589, + 481, + 621 + ], + "spans": [ + { + "bbox": [ + 130, + 589, + 481, + 621 + ], + "type": "text", + "content": "45. Zhang, P., Dou, H., Yu, Y., Li, X.: Adaptive cross-domain learning for generalizable person re-identification. In: European Conference on Computer Vision. pp. 215-232. Springer (2022)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 130, + 621, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 621, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 621, + 481, + 665 + ], + "type": "text", + "content": "46. Zhao, Y., Zhong, Z., Yang, F., Luo, Z., Lin, Y., Li, S., Sebe, N.: Learning to generalize unseen domains via memory-based multi-source meta-learning for person re-identification. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6277-6286 (2021)" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 252, + 91, + 445, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 252, + 91, + 445, + 102 + ], + "spans": [ + { + "bbox": [ + 252, + 91, + 445, + 102 + ], + "type": "text", + "content": "Rethinking Normalization Layers for DG-ReID" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 238 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 480, + 149 + ], + "type": "text", + "content": "47. Zheng, L., Shen, L., Tian, L., Wang, S., Wang, J., Tian, Q.: Scalable person re-identification: A benchmark. In: Proceedings of the IEEE international conference on computer vision. pp. 1116-1124 (2015)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 150, + 480, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 150, + 480, + 172 + ], + "spans": [ + { + "bbox": [ + 132, + 150, + 480, + 172 + ], + "type": "text", + "content": "48. Zheng, W., Gong, S., Xiang, T.: Associating groups of people. In: British Machine Vision Conference (2009)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 172, + 480, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 172, + 480, + 205 + ], + "spans": [ + { + "bbox": [ + 132, + 172, + 480, + 205 + ], + "type": "text", + "content": "49. Zheng, Z., Ruan, T., Wei, Y., Yang, Y., Mei, T.: VehicleNet: Learning robust visual representation for vehicle re-identification. IEEE Transactions on Multimedia 23, 2683-2693 (2020)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 205, + 480, + 238 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 205, + 480, + 238 + ], + "spans": [ + { + "bbox": [ + 132, + 205, + 480, + 238 + ], + "type": "text", + "content": "50. Zheng, Z., Zheng, L., Yang, Y.: Unlabeled samples generated by gan improve the person re-identification baseline in vitro. In: Proceedings of the IEEE international conference on computer vision. pp. 3754-3762 (2017)" + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 268, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 268, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 268, + 102 + ], + "type": "text", + "content": "R. Nie and J. Ding et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Rethinking Tree-Ring Watermarking for Enhanced Multi-Key Identification/dd0cc8f0-17b3-4928-9aeb-d79e0a1054db_content_list.json b/2024/Rethinking Tree-Ring Watermarking for Enhanced Multi-Key Identification/dd0cc8f0-17b3-4928-9aeb-d79e0a1054db_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..443e1daf0d413c806682a48622f0f996f98ca7ab --- /dev/null +++ b/2024/Rethinking Tree-Ring Watermarking for Enhanced Multi-Key Identification/dd0cc8f0-17b3-4928-9aeb-d79e0a1054db_content_list.json @@ -0,0 +1,1782 @@ +[ + { + "type": "text", + "text": "RingID: Rethinking Tree-Ring Watermarking for Enhanced Multi-Key Identification", + "text_level": 1, + "bbox": [ + 222, + 140, + 779, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Hai $\\mathrm{Ci}^{*}$ Pei Yang\\*, Yiren Song\\*, and Mike Zheng Shou", + "bbox": [ + 269, + 212, + 730, + 227 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Show Lab, National University of Singapore cihai03@gmail.com yangpei@u.nus.edu yiren@nus.edu.sg mike.zheng.shou@gmail.com", + "bbox": [ + 330, + 239, + 671, + 282 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. We revisit Tree-Ring Watermarking, a recent diffusion model watermarking method that demonstrates great robustness to various attacks. We conduct an in-depth study on it and reveal that the distribution shift unintentionally introduced by the watermarking process, apart from watermark pattern matching, contributes to its exceptional robustness. Our investigation further exposes inherent flaws in its original design, particularly in its ability to identify multiple distinct keys, where distribution shift offers no assistance. Based on these findings and analysis, we present RingID for enhanced multi-key identification. It consists of a novel multi-channel heterogeneous watermarking approach designed to seamlessly amalgamate distinctive advantages from diverse watermarks. Coupled with a series of suggested enhancements, RingID exhibits substantial advancements in multi-key identification. Github page: https://github.com/showlab/RingID", + "bbox": [ + 261, + 321, + 743, + 517 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Diffusion Models $\\cdot$ Tree-Ring Watermarking $\\cdot$ Multi-Key Identification", + "bbox": [ + 261, + 529, + 740, + 556 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 585, + 375, + 599 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "With the popularity of diffusion models, a vast number of high-quality diffusion generated images are circulating on the internet. It has become increasingly important to identify AI-generated images to detect fabrication and to track and protect copyrights. Luckily, watermarking [1, 13] offers a reliable solution.", + "bbox": [ + 212, + 617, + 785, + 678 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Research on watermarking has a long history in computer vision [9, 34]. A series of works [13, 14, 35, 48, 49] imprint imperceptible watermarks on images through small pixel perturbations. Recently, [47] demonstrates that pixel-level perturbations are provably removable by regeneration attacks. In contrast, Tree-Ring Watermarking [38] has recently introduced a novel approach, proposing to imprint a specially designed tree-ring pattern into the initial noise of a diffusion model. Different patterns are also called different \"keys\". The verification of the watermark's existence involves recovering the initial diffusion noise from the image and comparing the extracted key with the injected one. Tree-Ring", + "bbox": [ + 212, + 679, + 787, + 816 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Equal Contribution. ${}^{\\boxtimes }$ Corresponding Author.", + "bbox": [ + 217, + 823, + 537, + 840 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/31852c4dd9f7b48c201208e99c5e85b08affb1a9b1424b4e40c6061a03e13fd8.jpg", + "image_caption": [ + "Tree-Ring" + ], + "image_footnote": [], + "bbox": [ + 218, + 162, + 367, + 253 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/88e1bbd7d21261cb169f4e01d9250ef3323af44ea05d1127d6b6e3c9c1d40587.jpg", + "image_caption": [ + "RingID" + ], + "image_footnote": [], + "bbox": [ + 383, + 164, + 529, + 252 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/4d13a9715ed9f8335ef0e6a85f9504bc171571aef2017eab31f1e22196ec6e11.jpg", + "image_caption": [ + "Number of Keys", + "Fig. 1: Tree-Ring imprints a broken watermark to a single channel. Our method RingID achieves stronger robustness by imprinting intact watermarks to multiple channels. In the task of multi-key identification, RingID shows overwhelming advantages." + ], + "image_footnote": [], + "bbox": [ + 550, + 146, + 787, + 258 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "is considered a semantic watermark since the watermark changes the layout of the generated image and semantically hidden in it. It has demonstrated strong robustness against regeneration attacks [47] and various other image transformations [1, 38], credited to the meticulously designed tree-ring pattern. However, previous research only studies its performance in the verification task, which aims at distinguishing between watermarked and unwatermarked images. It's still unclear if Tree-Ring can be used to distribute multiple distinct keys and distinguish between them, a.k.a. watermark identification, which is crucial for source tracing and user attribution. Will Tree-Ring still be strong and robust as in the verification task? In this work, we will explore and answer this question.", + "bbox": [ + 212, + 366, + 787, + 517 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We start by studying the source of Tree-Ring's robustness under verification setting. For the first time, we uncover that the robustness of Tree-Ring stems not only from the carefully designed tree-ring pattern but also from the unintentionally introduced distribution shift during watermark imprinting. Such shift makes a big difference, especially in its robustness to Rotation and Crop/Scale image transformations, which cannot be handled by tree-ring's pattern design. More importantly, this overlooked power does not help in the identification scenario, because different watermarks experience the same shift.", + "bbox": [ + 212, + 518, + 787, + 640 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We further evaluate Tree-Ring in the identification task and surprisingly find that Tree-Ring struggles to distinguish between even dozens of different keys. Identification accuracy is reported in Tab. 2. The disappearance of the extra power of distribution shift has clearly exposed more design flaws in Tree-Ring. First, we find that tree-ring patterns are not discriminative enough to distinguish between different keys. This results in Tree-Ring's bad accuracy under various attacks or even under the clean setting. Meanwhile, the defective imprinting process leads to the injection of broken watermarks, further weakening its distinguishing ability. In addition, without the help of distribution shift, Tree-Ring is completely incapable of handling rotation and crop/scale transformations. Identification accuracy under these attacks is close to zero.", + "bbox": [ + 212, + 642, + 787, + 806 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In order to address the above issues, we propose RingID for enhanced identification capability. It bases on a novel Multi-Channel Heterogeneous watermarking", + "bbox": [ + 214, + 809, + 787, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "H. Ci et al.", + "bbox": [ + 271, + 114, + 349, + 127 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "framework, which can amalgamate distinctive advantages from different types of watermarks to resist to various kinds of attacks. Coupled with discretization and lossless imprinting methods, RingID gains substantial improvement on distinguishing ability. What's more, for the specific attack such as rotation, we locate the flaws in Tree-Ring's original design and propose systematic solutions, which improves the identification accuracy under rotation from nearly 0 to 0.86. To summarize our contributions:", + "bbox": [ + 212, + 146, + 787, + 252 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. For the first time, we elucidate how the unintentional introduction of distribution shift substantially contributes to Tree-Ring's robustness. We demonstrate this effect from both mathematical and empirical perspectives.", + "2. We systematically explore Tree-Ring's performance in the identification scenario and reveal its limitations through comprehensive analysis.", + "3. We present RingID, a systematic solution that significantly enhances watermark verification and identification, delivering impressive results. The verification AUC improves from 0.975 to 0.995, while the identification accuracy rises from 0.07 to 0.82." + ], + "bbox": [ + 220, + 261, + 782, + 393 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 215, + 419, + 387, + 434 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.1 Diffusion Models", + "text_level": 1, + "bbox": [ + 215, + 443, + 401, + 457 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Diffusion models [12, 18, 31] signify a swiftly advancing category of generative models. They define a forward diffusion process that gradually add small amount of Gaussian noise to the input data and learn to recover the clean data from the distortion [18]. Diffusion models initially demonstrate huge potential in generating high-resolution images [18, 32], resulting in globally recognized products like Midjourney [25], DALL-E [3, 26] and StableDiffusion [29]. Thereafter, they are broadly applied to modeling various types of data, e.g. video [4, 16, 30, 33, 39, 44], audio [19], 3d object [22, 27], humans [7, 8, 36, 50], etc.", + "bbox": [ + 212, + 464, + 787, + 585 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2 Watermarking Diffusion Models", + "text_level": 1, + "bbox": [ + 215, + 604, + 526, + 619 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "When it comes to watermarking diffusion models, there are two different types of objectives: watermarking model weights [2, 23, 37, 45, 48] or generated content [6, 14, 35, 43, 46, 49]. Model weights watermarking methods aim to protect the intellectual property of the model itself and identify whether another model is an instance of plagiarism. While content watermarking methods protect the copyright of generated images [13, 42, 43] or videos [24, 46] by adding imperceptible perturbations onto pixels. They either imprint the watermark into the frequency via DFT [20], DCT [9, 11] and wavelet transformations [10, 40], or directly onto pixels via an encoder network [15, 35, 41, 49]. Recently, Tree-Ring [38] proposes to imprint a tree-ring watermark pattern into the initial diffusion noise. It demonstrates strong robustness to various attacks. In this work, we delve into Tree-Ring, uncovering its source of power in watermark verification and pinpointing its vulnerability in identifying multiple keys. Subsequently, we introduce RingID, a systematic solution fortified with enhanced identification capability.", + "bbox": [ + 212, + 628, + 787, + 840 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "RingID: Enhanced Multi-Key Identification", + "bbox": [ + 441, + 114, + 730, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3 Preliminaries", + "text_level": 1, + "bbox": [ + 215, + 143, + 380, + 160 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1 Threat Model", + "text_level": 1, + "bbox": [ + 215, + 184, + 377, + 199 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Model Owner: The owner of the diffusion model provides the image generation service through API access. He would like to imprint imperceptible watermarks into every generated image for copyright protection and source tracking. Given an arbitrary image, the model owner would like to check whether the image contains his own watermark (a.k.a. Watermark Verification) and which watermark it is if he distributes multiple different watermarks (a.k.a. Identification). The goal of the model owner is to ensure high verification and identification accuracy regardless of any image transformations.", + "bbox": [ + 212, + 217, + 787, + 338 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Attacker: The attacker uses the model owner's service to generate an image. He attempts to apply various image transformations such as rotation and JPEG compression to disrupt the watermark, in order to evade the detection and identification check of the model owner. Subsequently, he can claim ownership of the image and profit from it illegally", + "bbox": [ + 212, + 340, + 787, + 416 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2 Notation", + "text_level": 1, + "bbox": [ + 215, + 446, + 336, + 459 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We denote the initial diffusion noise by $x_{T}$ , and its frequency counterpart after FFT by $X_{T}$ . $T$ is the diffusion step. Sometimes we may omit the subscript $T$ for simplicity. We denote the real part, imaginary part and conjugate symmetry of $X_{T}$ by $X_{\\mathrm{re}}$ , $X_{\\mathrm{im}}$ , and $X_{\\mathrm{cs}}$ . In the context of StableDiffusion [29], $X_{T} \\in R^{h \\times w \\times c}$ has 4 channels $\\mathcal{C} = \\{0,1,2,3\\}$ . We denote the watermarked channels by $\\mathcal{C}_w$ , where $\\mathcal{C}_w \\subset \\mathcal{C}$ . We denote the imprinted key by $w$ , the recovered key at detection time by $\\hat{w}$ , and the null watermark extracted from the unwatermarked image by $\\hat{w}_{\\emptyset}$ . We use $\\mathcal{W} = \\{w_1, w_2, \\ldots\\}$ to denote a set of different keys. $|\\mathcal{W}|$ is its capacity.", + "bbox": [ + 212, + 479, + 787, + 602 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3 Task Formulation: Verification vs. Identification", + "text_level": 1, + "bbox": [ + 215, + 627, + 656, + 642 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this section, we formulate the verification and identification task following the problem statement of Tree-Ring [38]. The task of watermark verification is to detect the existence of watermarks. Verification algorithms attempt to distinguish between watermarked images carrying key $\\hat{w}$ and unwatermarked images carrying null key $\\hat{w}_{\\emptyset}$ . More formally, verification can be formulated as distinguishing between the distance to the reference key $w$ : $d(\\hat{w}, w) \\neq d(\\hat{w}_{\\emptyset}, w)$ . In Tree-Ring, $d$ is chosen to be $\\ell_{1}$ distance and $w$ is the watermark to be imprinted.", + "bbox": [ + 212, + 657, + 787, + 763 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "When a set of different keys $w_{i} \\in \\mathcal{W}$ are distributed, the goal of watermark identification is to find the best match for the recovered key $\\hat{w}$ among all candidates in $\\mathcal{W}$ , i.e. $\\mathrm{ID}(\\hat{w}) = \\arg \\min_{i} [d(\\hat{w}, w_{i})]$ . In practice, different users are usually assigned different keys. Thus, each generated image can be source tracked and attributed to a unique user via watermark identification.", + "bbox": [ + 212, + 763, + 787, + 839 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "H. Ci et al.", + "bbox": [ + 271, + 114, + 346, + 126 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/96908b8f6a6d835c80c90827bb5c6c76a73b7986b046f2384401e79b5775d6b4.jpg", + "image_caption": [ + "Fig. 2: Framework of the watermarking process. RingID introduces a series of approaches that can help to imprint a lossless and robust watermark. In contrast, TreeRing injects a lossy and less robust watermark." + ], + "image_footnote": [], + "bbox": [ + 222, + 148, + 781, + 344 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4 Delve into Tree-Ring Watermarking", + "text_level": 1, + "bbox": [ + 215, + 430, + 602, + 448 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we will reveal why Tree-Ring [38] is robust to various image transformations, even some of them are not considered by design. We further evaluate Tree-Ring in the multi-key identification task and pinpoint its vulnerabilities.", + "bbox": [ + 215, + 464, + 787, + 510 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1 Recap of Tree-Ring", + "text_level": 1, + "bbox": [ + 215, + 534, + 423, + 550 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Tree-Ring [38] originally studies the verification setting. Its goal is to distinguish watermarked images from non-watermarked images. As illustrated in Fig. 2, Tree-Ring transforms the initial diffusion noise $x_{T}$ into the frequency $X_{T}$ and injects a tree-ring pattern $w$ into the center of $X_{T}$ , where the value of each ring in pattern $w$ is sampled from Gaussian. It then transforms the watermarked $X_{T}$ back to the spatial domain, retaining only the real part to feed the diffusion network for image generation. Notably, the accompanying imaginary part is discarded to be compatible with the network input. For detection of the watermark, Tree-Ring inverts the diffusion process to estimate $x_{T}$ from an image, transforms it to the frequency domain, and extracts the watermark pattern $\\hat{w}$ . Similarly, nonsense pattern $\\hat{w}_{\\emptyset}$ can be extracted from the non-watermarked image. Tree-Ring distinguishes between these two $\\ell_{1}$ -to-reference distances: $\\| \\hat{w} - w \\|_{1}$ and $\\| \\hat{w}_{\\emptyset} - w \\|_{1}$ to determine whether an image is watermarked or not.", + "bbox": [ + 215, + 561, + 787, + 758 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2 Distribution Shift in Watermarking", + "text_level": 1, + "bbox": [ + 215, + 782, + 555, + 797 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Tree-Ring's detection algorithm bases on the intuition that the recovered pattern $\\hat{w}$ from a watermarked image should match the injected reference pattern $w$", + "bbox": [ + 215, + 809, + 785, + 839 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "RingID: Enhanced Multi-Key Identification", + "bbox": [ + 441, + 114, + 730, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/3282c7e1061491af0c189db50e1099cac935da4f5b890c5a9f3f983a5adc5bcf.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 246, + 145, + 754, + 241 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/4833b9f577866c4ada8aabea1bc2c47997ec67079d34f606000be252bde96986.jpg", + "image_caption": [ + "Fig. 3: The effect of distribution shift in both verification and identification tasks. Discarding the imaginary part during watermark imprinting causes the distribution shift in $\\ell_1$ distance. Distribution shift aids verification but not identification." + ], + "image_footnote": [], + "bbox": [ + 246, + 243, + 754, + 299 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "to some extent, leading to a smaller $\\ell_1$ -to-reference distance, i.e. $\\| \\hat{w} - w \\|_1 < \\| \\hat{w}_{\\emptyset} - w \\|_1$ .", + "bbox": [ + 214, + 383, + 787, + 412 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "However, we have discovered that apart from the pattern matching, another crucial factor contributes to the separation of $\\| \\hat{w} - w \\|_1$ v.s. $\\| \\hat{w}_{\\emptyset} - w \\|_1$ . This factor is the distribution shift introduced by discarding the imaginary part during the watermark injection process, which is highlighted in Fig. 2. This affects all watermarked images, further reducing the expectation of $\\mathbb{E}\\left[\\| \\hat{w} - w \\|_1\\right]$ . In other words, it shifts the distribution of $\\| \\hat{w} - w \\|_1$ towards a smaller value, making it easier to distinguish between $\\| \\hat{w} - w \\|_1$ and $\\| \\hat{w}_{\\emptyset} - w \\|_1$ .", + "bbox": [ + 212, + 414, + 787, + 521 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Mathematically, we prove that the shift factor is $\\frac{\\sqrt{3}}{2}$ in a general and simplified setting. Let $\\hat{w}_n$ denote the recovered watermark that has never been shifted, thus we have the following:", + "bbox": [ + 212, + 521, + 787, + 568 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbb {E} \\left[ \\| \\hat {w} - w \\| _ {1} \\right] = \\frac {\\sqrt {3}}{2} \\mathbb {E} \\left[ \\| \\hat {w} _ {n} - w \\| _ {1} \\right]. \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 374, + 580, + 785, + 612 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Detailed derivation is elaborated in Supplementary A. Actually, we find any operation leading to a sufficient distribution shift can be considered a good standalone watermarking approach. We demonstrate the operation of discarding the imaginary part performs well on its own in Supplementary B.", + "bbox": [ + 214, + 618, + 787, + 680 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Effect in Different Tasks and Attacks Fig. 3 illustrates the effect of distribution shift under different tasks. In Verification, it helps distinguish between watermarked and non-watermarked images. Next, we go one step further and consider the situation when the image undergoes different attacks like JPEG or Resize. If the tree-ring pattern itself can cope with a certain attack, the improvement brought by the distribution shift will be small, because the two distributions are already separable even without the help of shift. On the contrary, if the tree-ring pattern itself cannot deal with a certain attack, the improvement brought by the distribution shift will be significant. Empirically (Supplementary", + "bbox": [ + 214, + 703, + 789, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "H. Ci et al.", + "bbox": [ + 271, + 114, + 349, + 127 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "A.6), we find distribution shift helps a lot under Rotation and Crop/Scale attacks. This implies that the original design of tree-ring cannot handle these two attacks. When it comes to Identification, things are a little different. Since different watermarks experience the same distribution shift, discarding the imaginary part doesn't help distinguishing them.", + "bbox": [ + 212, + 146, + 782, + 220 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "This sparks our curiosity about Tree-Ring's performance in identification, since distribution shift doesn't help and identification is more difficult. We would like to know if Tree-Ring still works under Rotation and Crop/Scale attacks. In the next section, we will explore these questions.", + "bbox": [ + 212, + 222, + 782, + 282 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3 Is Tree-Ring Good at Identification?", + "text_level": 1, + "bbox": [ + 214, + 301, + 562, + 316 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We report the performance of Tree-Ring when identifying different numbers of keys in Tab. 2. Tree-Ring struggles to identify as few as 32 keys. What's more, identification accuracy is nearly zero under Rotation and Crop/Scale attacks.", + "bbox": [ + 212, + 324, + 784, + 369 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We summarize 2 reasons as follows: (1) Distribution shift doesn't help. As tree-ring watermark cannot handle Rotation and Crop/Scale attacks, identification accuracy under these attacks drops to nearly zero. (2) Tee-ring pattern doesn't have enough distinguishing ability. Since identification is more difficult than verification, poor distinguishing ability results in poor performance under most attacks or even under no attack setting.", + "bbox": [ + 212, + 369, + 782, + 460 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4 Why is Tree-Ring Vulnerable to Rotation and Crop/Scale?", + "text_level": 1, + "bbox": [ + 214, + 479, + 746, + 494 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "All previous experimental results and analysis show one thing, tree-ring pattern cannot handle these two attacks. In this section, we will see what happens.", + "bbox": [ + 212, + 503, + 782, + 532 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Vulnerable to Crop & Scale Scaling an image by a factor of $a$ corresponds to compressing the frequency spectrum $X[u,v]$ in both magnitudes and positions: $\\frac{1}{a^2} X\\left[\\frac{u}{a},\\frac{v}{a}\\right]$ . Cropping an image modulates the frequency magnitudes by a sinc function. It is straightforward that these transformations easily disrupt the tree-ring pattern on frequency domain. As shown in Fig. 4(e), the pattern is completely disrupted.", + "bbox": [ + 212, + 532, + 782, + 622 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Vulnerable to Rotation Although Tree-Ring is designed as ring patterns to seek rotation robustness, we find the existing design ignores many details, leading to no rotation robustness. In short, a combination of corner cropping when rotating, a lossy injection process, and the rotationally asymmetric shape cause the problem. We will explain in detail and propose solutions in Sec. 5.2.", + "bbox": [ + 212, + 623, + 782, + 698 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5 RingID: Enhanced Watermarking for Identification", + "text_level": 1, + "bbox": [ + 214, + 720, + 745, + 736 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this section, we show how our proposed method RingID advances in identifying multiple distinct keys. RingID originates from Tree-Ring but incorporates systematic enhancements. We introduce a multi-channel heterogeneous watermarking framework, a discretization approach and a method of lossless imprinting to generally improve its distinguishing ability. We also discuss a series of designs regarding stronger rotation robustness and increased capacity.", + "bbox": [ + 212, + 750, + 782, + 839 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "RingID: Enhanced Multi-Key Identification", + "bbox": [ + 441, + 114, + 730, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.1 Multi-Channel Heterogeneous Watermarking", + "text_level": 1, + "bbox": [ + 215, + 146, + 633, + 162 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Multi-channel heterogeneous (MCH) watermarking serves as a general framework to comprehensively improve watermarking robustness to a battery of attacks. Intuitively, we imprint different types of watermarks on different channels of the initial noise to amalgamate their distinctive advantages, as illustrated in Fig. 1. During matching, we calculate the minimum of channel-wise $\\ell_1$ distance to the reference watermark for all watermarked channels:", + "bbox": [ + 212, + 172, + 787, + 263 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {I D} (\\hat {w}) = \\arg \\min _ {i} \\left\\{\\min _ {c \\in C _ {w}} \\left[ \\lambda_ {c} \\| \\hat {w} ^ {c} - w _ {i} ^ {c} \\| _ {1} \\right] \\right\\} \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 356, + 272, + 785, + 305 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $c$ denotes the channel index and traverses all watermarked channels $C_w$ . $i$ denotes the index of candidate reference watermarks. $\\lambda_c$ is a channel-wise normalizing factor.", + "bbox": [ + 212, + 314, + 782, + 359 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Eq. (2) bases on the intuition that if a watermark is robust to a certain attack, then its distance from the ground-truth reference is smaller than that of other watermarks that are not robust. Ideally, it enables adaptive selection of the most robust watermark under different attacks. Notably, the total capacity of MCH is the minimum of each watermark. Therefore, it's recommended to use a watermark pattern with sufficient capacity on each channel. At the same time, in order to reduce the impact on the generation quality, we need to carefully select the type of watermarks for combination. We find the Gaussian noise pattern is a perfect candidate. It has infinite capacity and the same distribution as the initial noise. Meanwhile, it has good robustness to non-geometric attacks [38] and is complementary to the tree-ring pattern. So RingID explores the combination of a Gaussian noise watermark and a tree-ring watermark. Empirical study in Sec. 6.3 demonstrates that this combination is able to perfectly amalgamate unique advantages from the noise and ring watermark.", + "bbox": [ + 212, + 359, + 787, + 571 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.2 Being Invariant to Rotation", + "text_level": 1, + "bbox": [ + 215, + 594, + 493, + 609 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Although the tree-ring pattern is designed to be rotationally symmetric, it exhibits unexpected vulnerability to rotation as shown in Tab. 2. Below, we identify the causes of the issue and propose a series of measures to address it.", + "bbox": [ + 212, + 619, + 785, + 666 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Spatial Shift to Avoid Corner Cropping We visualize the imprinted tree-ring watermark in spatial domain in Fig. 4(c). Due to the properties of FFT, tree-ring patterns in the frequency tends to exhibit concentrated energy in the four corners in the spatial domain, rendering them vulnerable to cropping during rotation. We propose to circularly shift the watermark pattern in the spatial domain to the center of the image, thus the majority of watermark information is kept while rotating. For an $N \\times N$ latent image, we circularly shift the pattern by $\\frac{N}{2}$ pixels in both width and height dimension. This is equivalent to directly multiplying the frequency watermark with a chessboard pattern given by $H[u,v] = e^{j\\frac{\\pi}{2}(u + v)} = (-1)^{u + v}$ . The change of the watermark can be clearly", + "bbox": [ + 212, + 688, + 787, + 843 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "H. Ci et al.", + "bbox": [ + 271, + 114, + 349, + 127 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/0818580d2bedb26f0e04968fab624344aeee8267d1b8b36a501bcd2be207d2a8.jpg", + "image_caption": [ + "(a) Discretization" + ], + "image_footnote": [], + "bbox": [ + 250, + 186, + 316, + 257 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/01c7fb2d70c7e839178ac58ed1b6bbd852b1046a240bbdfe20975ad884cb1067.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 330, + 188, + 395, + 258 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/8b5d329ded6bcd03c6f203559fadbd4f6df61a8db74be7ab7544c3ee75e6fc9f.jpg", + "image_caption": [ + "(b) Rounder Ring" + ], + "image_footnote": [], + "bbox": [ + 426, + 166, + 493, + 286 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/d6e9a183641f6ce5934fc7c18aaf07b7d9e5cc80704d272ab81ca4e00eeed782.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 508, + 166, + 575, + 292 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/fb9afe1dece5f3b4fedf427f6ea6e1397b4f51b21af015a3908ff649aae20d43.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 602, + 152, + 754, + 294 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/f22f1968a17b015f0dfbb5599709215f959d29fcee5f9283ed9499e1a5bbb261.jpg", + "image_caption": [ + "(d) Lossless Imprinting" + ], + "image_footnote": [], + "bbox": [ + 284, + 318, + 529, + 444 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/65252de08df1416b6425d5274a3c928ce6c4aa3ca64b9ae457a77b03f03e1476.jpg", + "image_caption": [ + "(e) Crop & Scale", + "Fig. 4: Detailed illustration of different acts and their impacts. (a) Discretization: Discretize values to $\\pm \\alpha$ for each ring, increasing the distinguishing ability of the watermark. (b) Rounder ring: Render smoother edges and reduce artifacts. Increase rotational robustness. (c) Spatial Shift: Shift energy on the corners to the center, avoiding potential cropping during rotation. (d) Lossless Imprinting: Ensure the imprinted pattern the same as the original design. (e) Crop & Scale: Zooming in the spatial domain compresses the imprinted pattern in the frequency domain, causing pattern mismatch and hinders identification." + ], + "image_footnote": [], + "bbox": [ + 596, + 318, + 746, + 444 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "observed in Fig. 4(c). Although spatial-shifting helps a lot for rotation invariance, we noticed that this often leads to the generation of circular artifacts at the center of the image. Keeping generation quality in mind, we further multiply the shifted pattern by a factor $\\eta$ to suppress the peak at the center. $\\eta$ trades-off the robustness, as smaller $\\eta$ weakens the watermark pattern. In practice, we find $\\eta \\in [0.8,0.9]$ strikes a good balance between image quality and robustness.", + "bbox": [ + 212, + 599, + 787, + 691 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Lossless Imprinting The operation of discarding the imaginary part during watermarking not only shifts the distribution of $\\ell_1$ -to-reference, but also disrupt the pattern, resulting in the loss of its rotational symmetry. Fig. 4(d) visualizes the actually carried pattern. More concretely, recall from Fig. 2, when the watermarked noise $X[u,v]$ is transformed back to the spatial $x[m,n]$ , only its real part $x_{\\mathrm{re}}[m,n]$ is kept for image generation. From a frequency domain perspective, this is equivalent to generation with its conjugate symmetry $X_{\\mathrm{cs}}[u,v]$ , which is not necessarily equal to $X[u,v]$ . To leverage the rotational symmetry of the", + "bbox": [ + 212, + 718, + 787, + 840 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "RingID: Enhanced Multi-Key Identification", + "bbox": [ + 441, + 114, + 730, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 114, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "ring patterns, we attempt to eliminate the information loss during imprinting by solving $X[u,v] = X_{\\mathrm{cs}}[u,v]$ . The results indicate that $X_{\\mathrm{re}}[u,v]$ should be an even function and $X_{\\mathrm{im}}[u,v]$ should be an odd function about the Fourier center. To satisfy these constraints, we imprint the tree-ring pattern only in the real part and leave the imaginary part blank. We also align the center of the tree-ring pattern with the Fourier center (32,32). Originally, the tree-ring pattern was put at (31,32) and imprinted on both the real and the imaginary.", + "bbox": [ + 212, + 146, + 787, + 253 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Making A Rounder Ring Rings are originally determined by locating pixels equidistant from a give circle center $(x_0, y_0)$ : $(x - x_0)^2 + (y - y_0)^2 = r^2$ . We observed that this simple approach does not yield a well-rounded circular ring, particularly in images with lower resolutions like initial noise of shape $64 \\times 64$ . Artifacts and alias are shown in Fig. 4(b). Although it is centrosymmetric, it lacks rotational symmetry. Well, the solution is quite straightforward: drawing a rounder ring. There are a number of tools that can help draw a much rounder ring. Here, we employ a very simple idea to get a rounder ring. We place a white pixel on a black background at a distance $r$ from the rotation center. By rotating the low resolution image 360 degrees and recording the trajectory of this pixel, we obtain a rounder ring with a radius of $r$ . Despite its simplicity, empirical results in Tab. 3 show that it is crucial for rotational invariance.", + "bbox": [ + 212, + 277, + 789, + 459 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "5.3 Discretization for Enhanced Distinguishability", + "text_level": 1, + "bbox": [ + 214, + 484, + 643, + 501 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Tree-Ring samples values for each ring from a Gaussian distribution to keep consistency with the distribution of initial noise. We find that this random sampling makes it exceedingly challenging to distinguish between two keys. To address this issue, we propose a discretization approach, specifying that the value for each ring can only be either $\\alpha$ or $-\\alpha$ , illustrated in Fig. 4(a). Consequently, the watermark pattern for $n$ rings has a capacity of $2^n$ . While this seemingly reduces the upper limit of watermark capacity, it significantly enhances its valid capacity, ensuring optimal utilization of all available slots. Our experiments reveal that setting $\\alpha$ as the standard deviation of the initial noise ensures substantial distinguishability between different keys while reducing the impact on generation quality to a small level.", + "bbox": [ + 212, + 513, + 787, + 679 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "5.4 Increasing Capacity", + "text_level": 1, + "bbox": [ + 214, + 705, + 426, + 722 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "We consider two simple approaches to increase capacity: adjusting the number of rings in a single channel or imprinting rings in extra channels. We evaluate the advantages and drawbacks of both methods in Sec. 6.4. In summary, increasing the number of rings $|R|$ in a single channel can effectively boost capacity, leading to $2^{|R|}$ keys. However, excessive ring count results in a noticeable decline in robustness and generation quality. Imprinting watermarks in multiple channels exponentially increases capacity by $n^{|c|}$ , where $c$ is number of channels and $n$ is", + "bbox": [ + 212, + 734, + 787, + 840 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "H. Ci et al.", + "bbox": [ + 271, + 114, + 346, + 126 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "the capacity of a single channel. However, it's likely to introduce ring-like artifacts in the center of generated images. Therefore, we opt to place the watermark in a single channel and enhance capacity by adjusting the range of rings.", + "bbox": [ + 212, + 146, + 782, + 191 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "6 Experiment", + "text_level": 1, + "bbox": [ + 214, + 210, + 364, + 228 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "6.1 Experimental Setup", + "text_level": 1, + "bbox": [ + 214, + 237, + 426, + 253 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Following Tree-Ring [38], we experiment with the state-of-the-art opensource diffusion model StableDiffusion-V2 [29]. By default, RingID imprints a tree-ring watermark with radius 3-14 on the channel 3 of the initial noise and a Gaussian noise watermark on channels 0. We adopt discretization (to 64 and -64), spatial shift, rounder drawing, and lossless imprinting as ordered in Fig. 2. We report the ROC-AUC between 1000 watermarked and 1000 unwatermarked images as the verification metric. In identification, we report identification accuracy with different numbers of keys. Evaluation is performed under various image distortions, following the original setup of Tree-Ring [38], i.e., JPEG 25, rotation $75^{\\circ}$ , $75\\%$ random C&S, blur kernel 8, noise std 0.1, brightness [0, 6]. We assess the generation quality with CLIP scores [28] by OpenCLIP-ViT/G [5] between the generated image and the prompts to measure text-image alignment, and Frechet Inception Distance (FID) [17] to quantify the similarity between generated and natural images. We generate 5000 images and evaluate FID on the COCO 2017 training dataset [21]. We set both generation and inversion steps to 50.", + "bbox": [ + 212, + 258, + 787, + 484 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "6.2 Comparison with Baselines", + "text_level": 1, + "bbox": [ + 214, + 502, + 485, + 517 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Robustness In verification (Tab. 1), both RingID and Tree-Ring perfectly distinguish between watermarked and unwatermarked images when no attacks are performed, achieving AUC score of 1. However, RingID exhibit stronger robustness to various image distortions than Tree-Ring, achieving near-perfect average AUC score, even though it is not optimized for verification. In the more challenging identification task, (Tab. 2), RingID significantly outperforms Tree-Ring. When identifying 32 keys, Tree-Ring's average identification accuracy (excluding C&S) is 0.465, while RingID's accuracy reaches up to 0.992. As the number of keys increases to 2048, this gap widens drastically (0.077 v.s. 0.942), further highlighting RingID's superior robustness and larger valid key capacity.", + "bbox": [ + 212, + 522, + 787, + 672 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Notably, both methods struggle with the C&S distortion in identification (Tab. 2). This aligns with our analysis in Sec. 4.2. Scaling in the spatial domain causes inverse scaling in the frequency domain, leading to pattern disruption and rendering identification challenging under such distortions. However, in verification, with the help of distribution shift, watermarked images can still be effectively distinguished from unwatermarked ones under C&S. Both Tree-Ring and RingID achieve average AUC higher than 0.97.", + "bbox": [ + 212, + 674, + 787, + 779 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Image Quality RingID achieves strong performance without compromising the generation quality. It maintains comparable text-image alignments with TreeRing. Both methods achieve similar CLIP scores (0.364 v.s. 0.365). RingID obtains an FID of 26.13, which is also comparable to Tree-Ring's FID of 25.93.", + "bbox": [ + 214, + 779, + 787, + 839 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "RingID: Enhanced Multi-Key Identification", + "bbox": [ + 441, + 114, + 730, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 116, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/a3d75da98c77a8370364d9ad9997c080fa184ae1f6dca297dfe93487c1d56b62.jpg", + "table_caption": [ + "Table 1: Comparison with Tree-Ring in the verification task. The table shows the ROC-AUC values under various image distortions and CLIP Scores." + ], + "table_footnote": [], + "table_body": "
MethodsRing RadiusCleanRotateJPEGC&SBlurNoiseBrightnessAvgCLIP Score
Tree-Ring0-101.0000.9350.9990.9610.9990.9440.9830.9750.364
RingID0-101.0001.0001.0000.9790.9940.9690.9910.9900.359
3-141.0001.0001.0000.9870.9890.9980.9940.9950.365
", + "bbox": [ + 215, + 185, + 785, + 250 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/5155663580f91040a0e77e131db16eeaca572d4ebebf553d6856550321a05f85.jpg", + "table_caption": [ + "Table 2: Comparison with Tree-Ring in the identification task. We report the identification accuracy under various distortions for different numbers of keys." + ], + "table_footnote": [], + "table_body": "
Methods#KeysCleanRotateJPEGC&SBlurNoiseBrightnessAvg\\(Avg_{noC&S}\\)
Tree-Ring320.7900.0200.4200.0400.6100.5300.4200.4040.465
1280.4500.0100.1200.0200.2800.2300.1700.1830.210
20480.2000.0000.0400.0000.0900.0700.0600.0660.077
RingID321.0001.0001.0000.5300.9901.0000.9600.9260.992
1281.0000.9801.0000.2800.9801.0000.9400.8830.983
20481.0000.8601.0000.0800.9700.9500.8700.8190.942
", + "bbox": [ + 217, + 295, + 785, + 402 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "6.3 Ablation Study", + "text_level": 1, + "bbox": [ + 215, + 431, + 388, + 448 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Components of RingID Tab. 3 ablates all proposed components of RingID. We can find that all designs are crucial for the final success of RingID. Spatial shift, lossless imprinting, and discretization play a significant role in the robustness of rotation. While the use of rounder rings also increase the accuracy by approximately $40\\%$ (from 0.620 to 0.860). Notably, spatial shift sacrifices the robustness to blur and brightness attacks for rotation robustness.", + "bbox": [ + 212, + 457, + 785, + 547 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Heterogeneous Watermarking RingID employs the combination of a Gaussian noise watermark and a ring watermark. Here, we ablate the choice of imprinting the noise watermark on different channels. By comparing the first 3 rows of Tab. 4, we find that heterogeneous watermarking can effectively combine the advantage from both the noise watermark and the ring watermark. Among single-channel configurations, imprinting noise watermark on channel 0 (0.819) achieves the best performance. The combination of channels 0 and 1 yields the highest accuracy among all trials. Interestingly, imprinting noise watermarks on all available channels (0, 1, 2) decrease the accuracy from 0.830 to 0.807.", + "bbox": [ + 212, + 547, + 785, + 685 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "6.4 Discussion", + "text_level": 1, + "bbox": [ + 215, + 708, + 349, + 722 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Where is the Best Place to Put Rings? We could imprint all rings on one channel or distribute them evenly across multiple channels. In Tab. 5, we fix the total number of rings to 12 and distribute them to 1-4 channels. we can find that performance drops as we distribute rings to more channels. What's more, imprinting rings on multiple channels cause artifacts to appear more frequently (shown in Supplementary C). So given fixed capacity, it's better to imprint all rings on a single channel. This raises another question: what are the optimal", + "bbox": [ + 212, + 733, + 785, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "H. Ci et al.", + "bbox": [ + 271, + 114, + 346, + 126 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/48e0a0b89bf7d5ad6541983b190b5aabd0c68636e1c329d17792c1d8e2b6ce61.jpg", + "table_caption": [ + "Table 3: Ablation study on the effect of each proposed module in RingID. We report the identification accuracy for 2048 keys when removing individual components." + ], + "table_footnote": [], + "table_body": "
SettingsCleanRotateJPEGC&SBlurNoiseBrightnessAvg
RingID1.0000.8601.0000.0800.9700.9500.8700.819
- Spatial Shift1.0000.0001.0000.0500.9900.9300.9400.701
- Lossless Imprinting1.0000.0100.9700.1600.9500.9800.8300.700
- Rounder Ring1.0000.6200.9900.1000.8900.9700.8500.774
- Discretization0.9800.1200.3800.0200.4500.6500.3900.427
- Heterogeneous1.0000.8200.9400.0300.9600.7100.7200.740
", + "bbox": [ + 217, + 185, + 785, + 292 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/10678218a115fed06926b9c49683dca3c97a992f3ea374e84a9f95d8ba834953.jpg", + "table_caption": [ + "Table 4: Ablation study on multi-channel heterogeneous watermarking. This table reports the identification accuracy when the noise watermark is imprinted on different channels (abbr. CHs). “-” means the watermark is not imprinted." + ], + "table_footnote": [], + "table_body": "
Ring CHsNoise CHsCleanRotateJPEGC&SBlurNoiseBrightnessAvg
-01.0000.0001.0000.1700.9900.9500.9200.719
3-1.0000.8200.9400.0300.9600.7100.7200.740
301.0000.8601.0000.0800.9700.9500.8700.819
311.0000.8201.0000.1301.0000.8800.7700.800
321.0000.8400.9800.1000.9200.9900.7900.803
30,11.0000.8500.9500.1500.9900.9900.8800.830
30,21.0000.8400.9900.1200.9500.9800.8700.821
30,1,21.0000.8300.9600.1600.9300.9800.7900.807
", + "bbox": [ + 217, + 378, + 785, + 511 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "inner and outer radii if all rings are placed on a single channel? From Tab. 6 we can see that 3-14 forms a good choice for 11 rings with capacity 2048. The optimal ring radius also varies for different ring numbers.", + "bbox": [ + 215, + 558, + 784, + 603 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "More Keys We would like to explore whether RingID can handle a larger number of keys. Fig. 5 shows the average identification accuracy under various attacks $v.s.$ number of keys. We observe that RingID maintains good accuracy even with a large number of keys. And identification accuracy linearly decreases as the number of keys exponentially increase.", + "bbox": [ + 215, + 612, + 784, + 686 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/0dd4abc8edaeba4494308ff7d018cbc9c5859adcadc80f1b86ede17b4e933dbe.jpg", + "table_caption": [ + "Table 6: Influence of ring radius on identification accuracy and text-image alignment." + ], + "table_footnote": [], + "table_body": "
Radius#KeysAccCLIP ScoreRadius#KeysAccCLIP ScoreRadius#KeysAccCLIP Score
1-6320.8860.3631-92560.8190.3621-1220480.7860.361
2-7320.8760.3662-102560.8560.3582-1320480.8040.361
3-8320.8970.3653-112560.8500.3623-1420480.8190.368
4-9320.8970.3684-122560.8470.3634-1520480.8130.363
5-10320.8810.3615-132560.8390.3605-1620480.8030.360
", + "bbox": [ + 217, + 761, + 785, + 835 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "RingID: Enhanced Multi-Key Identification", + "bbox": [ + 441, + 114, + 730, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/35d36e710a17cdaa343af00aebce0fbd5a40b4fb01e59ca0244dd9dae0c0dbd0.jpg", + "table_caption": [ + "Table 5: Distributing 12 rings across different number of channels (#CHs). This table reports the average identification accuracy and CLIP Score." + ], + "table_footnote": [], + "table_body": "
Ring Radius #CHsAvgCLIP Score
3-1510.7400.357
3-920.7370.351
3-730.6640.351
3-640.5440.340
", + "bbox": [ + 217, + 226, + 486, + 306 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/dc91f09a8e3332f8fef9acaaed2e773f7422009913d4fd33282cd11ca1654349.jpg", + "image_caption": [ + "Fig. 5: Average identification accuracy $v.s.$ number of keys." + ], + "image_footnote": [], + "bbox": [ + 532, + 164, + 776, + 299 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/52bb288e0054b399f62eb9e2ebab2ee3e14701a8ddaa8ee532e2a35388fdd284.jpg", + "image_caption": [ + "Fig. 6: Qualitative results. The subtle changes in image layouts show that both TreeRing and RingID embed watermarks into the image's semantic content." + ], + "image_footnote": [], + "bbox": [ + 241, + 347, + 767, + 513 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "6.5 Qualitative Results", + "text_level": 1, + "bbox": [ + 215, + 588, + 421, + 603 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "As shown in Fig. 6, both RingID and Tree-Ring watermarking can generate images of good visual quality. Unlike low-level watermarking techniques, both RingID and Tree-Ring cause subtle changes in the image layout. However, RingID generates images with layout that is closer to the unwatermarked image. We owe this to avoiding watermarking on the low-frequency part (radius 0-3).", + "bbox": [ + 214, + 614, + 787, + 691 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "7 Conclusion and Limitation", + "text_level": 1, + "bbox": [ + 215, + 715, + 509, + 732 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "We conduct a comprehensive investigation on Tree-Ring Watermarking, revealing its limitations in identifying multiple keys. We then propose RingID to enhance identification capability.", + "bbox": [ + 212, + 750, + 785, + 792 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Limitation: Under multi-key identification, one limitation is that both TreeRing and our RingID are vulnerable to cropping and scaling attacks. Future work may explore other transform domains to resist these attacks.", + "bbox": [ + 214, + 794, + 785, + 839 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "H. Ci et al.", + "bbox": [ + 271, + 114, + 346, + 126 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgment: This project is supported by the Ministry of Education, Singapore, under the Academic Research Fund Tier 1 (FY2022) Award 22-5406-A0001.", + "bbox": [ + 215, + 146, + 784, + 191 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 217, + 321, + 233 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. An, B., Ding, M., Rabbani, T., Agrawal, A., Xu, Y., Deng, C., Zhu, S., Mohamed, A., Wen, Y., Goldstein, T., et al.: Benchmarking the robustness of image watermarks. arXiv preprint arXiv:2401.08573 (2024)", + "2. Bansal, A., Chiang, P.y., Curry, M.J., Jain, R., Wigington, C., Manjunatha, V., Dickerson, J.P., Goldstein, T.: Certified neural network watermarks with randomized smoothing. In: International Conference on Machine Learning. pp. 1450-1465. PMLR (2022)", + "3. Betker, J., Goh, G., Jing, L., Brooks, T., Wang, J., Li, L., Ouyang, L., Zhuang, J., Lee, J., Guo, Y., et al.: Improving image generation with better captions. Computer Science. https://cdn.opencai.com/papers/dall-e-3.pdf 2(3), 8 (2023)", + "4. Blattmann, A., Dockhorn, T., Kulal, S., Mendelevitch, D., Kilian, M., Lorenz, D., Levi, Y., English, Z., Voleti, V., Letts, A., et al.: Stable video diffusion: Scaling latent video diffusion models to large datasets. arXiv preprint arXiv:2311.15127 (2023)", + "5. Cherti, M., Beaumont, R., Wightman, R., Wortsman, M., Ilharco, G., Gordon, C., Schuhmann, C., Schmidt, L., Jitsev, J.: Reproducible scaling laws for contrastive language-image learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2818-2829 (2023)", + "6. Ci, H., Song, Y., Yang, P., Xie, J., Shou, M.Z.: Wmadapter: Adding watermark control to latent diffusion models. arXiv preprint arXiv:2406.08337 (2024)", + "7. Ci, H., Wang, C., Ma, X., Wang, Y.: Optimizing network structure for 3d human pose estimation. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 2262-2271 (2019)", + "8. Ci, H., Wu, M., Zhu, W., Ma, X., Dong, H., Zhong, F., Wang, Y.: Gfpose: Learning 3d human pose prior with gradient fields. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4800-4810 (2023)", + "9. Cox, I.: Digital watermarking and steganography. Morgan Kaufmann google scholar 2, 893-914 (2007)", + "0. Cui, S., Guo, X., Zhang, Z.: Estimation and inference in ultrahigh dimensional partially linear single-index models. arXiv preprint arXiv:2404.04471 (2024)", + "1. Cui, S., Sudjianto, A., Zhang, A., Li, R.: Enhancing robustness of gradientboosted decision trees through one-hot encoding and regularization. arXiv preprint arXiv:2304.13761 (2023)", + "2. Dhariwal, P., Nichol, A.: Diffusion models beat gans on image synthesis. Advances in neural information processing systems 34, 8780-8794 (2021)", + "3. Fernandez, P., Couairon, G., Jégou, H., Douze, M., Furon, T.: The stable signature: Rooting watermarks in latent diffusion models. arXiv preprint arXiv:2303.15435 (2023)", + "4. Fernandez, P., Sablayrolles, A., Furon, T., Jégou, H., Douze, M.: Watermarking images in self-supervised latent spaces. In: ICASSP 2022-2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). pp. 3054-3058. IEEE (2022)" + ], + "bbox": [ + 225, + 250, + 784, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "RingID: Enhanced Multi-Key Identification", + "bbox": [ + 441, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "15. Guo, H., Zhang, Q., Luo, J., Guo, F., Zhang, W., Su, X., Li, M.: Practical deep dispersed watermarking with synchronization and fusion. In: Proceedings of the 31st ACM International Conference on Multimedia. pp. 7922-7932 (2023)", + "16. Guo, Y., Yang, C., Rao, A., Liang, Z., Wang, Y., Qiao, Y., Agrawala, M., Lin, D., Dai, B.: Animatediff:Animate your personalized text-to-image diffusion models without specific tuning. International Conference on Learning Representations (2024)", + "17. Heusel, M., Ramsauer, H., Unterthiner, T., Nessler, B., Hochreiter, S.: Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems 30 (2017)", + "18. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. Advances in neural information processing systems 33, 6840-6851 (2020)", + "19. Kong, Z., Ping, W., Huang, J., Zhao, K., Catanzaro, B.: Diffwave: A versatile diffusion model for audio synthesis. arXiv preprint arXiv:2009.09761 (2020)", + "20. Lin, C.Y., Wu, M., Bloom, J.A., Cox, I.J., Miller, M.L., Lui, Y.M.: Rotation, scale, and translation resilient watermarking for images. IEEE Transactions on image processing 10(5), 767-782 (2001)", + "21. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollar, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13. pp. 740-755. Springer (2014)", + "22. Liu, R., Wu, R., Van Hoorick, B., Tokmakov, P., Zakharov, S., Vondrick, C.: Zero-1-to-3: Zero-shot one image to 3d object. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 9298-9309 (2023)", + "23. Liu, Y., Li, Z., Backes, M., Shen, Y., Zhang, Y.: Watermarking diffusion model. arXiv preprint arXiv:2305.12502 (2023)", + "24. Luo, X., Li, Y., Chang, H., Liu, C., Milanfar, P., Yang, F.: Dvmark: a deep multiscale framework for video watermarking. IEEE Transactions on Image Processing (2023)", + "25. Midjourney.com: Midjourney. https://www.midjourney.com/home", + "26. OpenAI: Dalle2. https://openai.com/dall-e-2", + "27. Poole, B., Jain, A., Barron, J.T., Mildenhall, B.: Dreamfusion: Text-to-3d using 2d diffusion. arXiv preprint arXiv:2209.14988 (2022)", + "28. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021)", + "29. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 10684-10695 (June 2022)", + "30. Singer, U., Polyak, A., Hayes, T., Yin, X., An, J., Zhang, S., Hu, Q., Yang, H., Ashual, O., Gafni, O., et al.: Make-a-video: Text-to-video generation without text-video data. arXiv preprint arXiv:2209.14792 (2022)", + "31. Song, J., Meng, C., Ermon, S.: Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502 (2020)", + "32. Song, Y., Sohl-Dickstein, J., Kingma, D.P., Kumar, A., Ermon, S., Poole, B.: Score-based generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456 (2020)" + ], + "bbox": [ + 215, + 146, + 785, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "H. Ci et al.", + "bbox": [ + 271, + 114, + 346, + 126 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "33. Song, Y., Huang, S., Yao, C., Ye, X., Ci, H., Liu, J., Zhang, Y., Shou, M.Z.: Processpainter: Learn painting process from sequence data. arXiv preprint arXiv:2406.06062 (2024)", + "34. Steinberg, Y., Merhav, N.: Identification in the presence of side information with application to watermarking. IEEE Transactions on Information Theory 47(4), 1410-1422 (2001)", + "35. Tancik, M., Mildenhall, B., Ng, R.: Stegastamp: Invisible hyperlinks in physical photographs. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 2117-2126 (2020)", + "36. Tevet, G., Raab, S., Gordon, B., Shafir, Y., Cohen-Or, D., Bermano, A.H.: Human motion diffusion model. arXiv preprint arXiv:2209.14916 (2022)", + "37. Uchida, Y., Nagai, Y., Sakazawa, S., Satoh, S.: Embedding watermarks into deep neural networks. In: Proceedings of the 2017 ACM on international conference on multimedia retrieval. pp. 269-277 (2017)", + "38. Wen, Y., Kirchenbauer, J., Geiping, J., Goldstein, T.: Tree-ring watermarks: Fingerprints for diffusion images that are invisible and robust. arXiv preprint arXiv:2305.20030 (2023)", + "39. Wu, J.Z., Ge, Y., Wang, X., Lei, S.W., Gu, Y., Shi, Y., Hsu, W., Shan, Y., Qie, X., Shou, M.Z.: Tune-a-video: One-shot tuning of image diffusion models for text-to-video generation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 7623-7633 (2023)", + "40. Xia, X.G., Boncelet, C.G., Arce, G.R.: Wavelet transform based watermark for digital images. Optics Express 3(12), 497-511 (1998)", + "41. Xiong, C., Qin, C., Feng, G., Zhang, X.: Flexible and secure watermarking for latent diffusion model. In: Proceedings of the 31st ACM International Conference on Multimedia. pp. 1668-1676 (2023)", + "42. Yang, P., Ci, H., Song, Y., Shou, M.Z.: Steganalysis on digital watermarking: Is your defense truly impervious? arXiv preprint arXiv:2406.09026 (2024)", + "43. Yu, N., Skripniuk, V., Abdelnabi, S., Fritz, M.: Artificial fingerprinting for generative models: Rooting deepfake attribution in training data. In: Proceedings of the IEEE/CVF International conference on computer vision. pp. 14448-14457 (2021)", + "44. Zhang, D.J., Wu, J.Z., Liu, J.W., Zhao, R., Ran, L., Gu, Y., Gao, D., Shou, M.Z.: Show-1: Marrying pixel and latent diffusion models for text-to-video generation. arXiv preprint arXiv:2309.15818 (2023)", + "45. Zhang, J., Gu, Z., Jang, J., Wu, H., Stoecklin, M.P., Huang, H., Molloy, I.: Protecting intellectual property of deep neural networks with watermarking. In: Proceedings of the 2018 on Asia conference on computer and communications security. pp. 159-172 (2018)", + "46. Zhang, K.A., Xu, L., Cuesta-Infante, A., Veeramachaneni, K.: Robust invisible video watermarking with attention. arXiv preprint arXiv:1909.01285 (2019)", + "47. Zhao, X., Zhang, K., Wang, Y.X., Li, L.: Generative autoencoders as watermark attackers: Analyses of vulnerabilities and threats. arXiv preprint arXiv:2306.01953 (2023)", + "48. Zhao, Y., Pang, T., Du, C., Yang, X., Cheung, N.M., Lin, M.: A recipe for watermarking diffusion models. arXiv preprint arXiv:2303.10137 (2023)", + "49. Zhu, J., Kaplan, R., Johnson, J., Fei-Fei, L.: Hidden: Hiding data with deep networks. In: Proceedings of the European conference on computer vision (ECCV). pp. 657-672 (2018)", + "50. Zhu, W., Ma, X., Ro, D., Ci, H., Zhang, J., Shi, J., Gao, F., Tian, Q., Wang, Y.: Human motion generation: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence (2023)" + ], + "bbox": [ + 212, + 146, + 787, + 833 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "RingID: Enhanced Multi-Key Identification", + "bbox": [ + 441, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 16 + } +] \ No newline at end of file diff --git a/2024/Rethinking Tree-Ring Watermarking for Enhanced Multi-Key Identification/dd0cc8f0-17b3-4928-9aeb-d79e0a1054db_model.json b/2024/Rethinking Tree-Ring Watermarking for Enhanced Multi-Key Identification/dd0cc8f0-17b3-4928-9aeb-d79e0a1054db_model.json new file mode 100644 index 0000000000000000000000000000000000000000..83025eb1f85463c7a83caad6935ac579b2e28551 --- /dev/null +++ b/2024/Rethinking Tree-Ring Watermarking for Enhanced Multi-Key Identification/dd0cc8f0-17b3-4928-9aeb-d79e0a1054db_model.json @@ -0,0 +1,2434 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.223, + 0.141, + 0.78, + 0.187 + ], + "angle": 0, + "content": "RingID: Rethinking Tree-Ring Watermarking for Enhanced Multi-Key Identification" + }, + { + "type": "text", + "bbox": [ + 0.271, + 0.213, + 0.732, + 0.228 + ], + "angle": 0, + "content": "Hai \\(\\mathrm{Ci}^{*}\\) Pei Yang\\*, Yiren Song\\*, and Mike Zheng Shou" + }, + { + "type": "text", + "bbox": [ + 0.331, + 0.24, + 0.672, + 0.284 + ], + "angle": 0, + "content": "Show Lab, National University of Singapore cihai03@gmail.com yangpei@u.nus.edu yiren@nus.edu.sg mike.zheng.shou@gmail.com" + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.322, + 0.744, + 0.518 + ], + "angle": 0, + "content": "Abstract. We revisit Tree-Ring Watermarking, a recent diffusion model watermarking method that demonstrates great robustness to various attacks. We conduct an in-depth study on it and reveal that the distribution shift unintentionally introduced by the watermarking process, apart from watermark pattern matching, contributes to its exceptional robustness. Our investigation further exposes inherent flaws in its original design, particularly in its ability to identify multiple distinct keys, where distribution shift offers no assistance. Based on these findings and analysis, we present RingID for enhanced multi-key identification. It consists of a novel multi-channel heterogeneous watermarking approach designed to seamlessly amalgamate distinctive advantages from diverse watermarks. Coupled with a series of suggested enhancements, RingID exhibits substantial advancements in multi-key identification. Github page: https://github.com/showlab/RingID" + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.53, + 0.741, + 0.558 + ], + "angle": 0, + "content": "Keywords: Diffusion Models \\(\\cdot\\) Tree-Ring Watermarking \\(\\cdot\\) Multi-Key Identification" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.586, + 0.377, + 0.601 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.618, + 0.787, + 0.679 + ], + "angle": 0, + "content": "With the popularity of diffusion models, a vast number of high-quality diffusion generated images are circulating on the internet. It has become increasingly important to identify AI-generated images to detect fabrication and to track and protect copyrights. Luckily, watermarking [1, 13] offers a reliable solution." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.68, + 0.788, + 0.817 + ], + "angle": 0, + "content": "Research on watermarking has a long history in computer vision [9, 34]. A series of works [13, 14, 35, 48, 49] imprint imperceptible watermarks on images through small pixel perturbations. Recently, [47] demonstrates that pixel-level perturbations are provably removable by regeneration attacks. In contrast, Tree-Ring Watermarking [38] has recently introduced a novel approach, proposing to imprint a specially designed tree-ring pattern into the initial noise of a diffusion model. Different patterns are also called different \"keys\". The verification of the watermark's existence involves recovering the initial diffusion noise from the image and comparing the extracted key with the injected one. Tree-Ring" + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.824, + 0.538, + 0.841 + ], + "angle": 0, + "content": "* Equal Contribution. \\( {}^{\\boxtimes } \\) Corresponding Author." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.35, + 0.128 + ], + "angle": 0, + "content": "H. Ci et al." + }, + { + "type": "image", + "bbox": [ + 0.219, + 0.164, + 0.368, + 0.254 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.255, + 0.262, + 0.327, + 0.277 + ], + "angle": 0, + "content": "Tree-Ring" + }, + { + "type": "image", + "bbox": [ + 0.384, + 0.165, + 0.531, + 0.253 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.428, + 0.262, + 0.48, + 0.277 + ], + "angle": 0, + "content": "RingID" + }, + { + "type": "image", + "bbox": [ + 0.552, + 0.147, + 0.788, + 0.26 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.637, + 0.264, + 0.746, + 0.276 + ], + "angle": 0, + "content": "Number of Keys" + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.291, + 0.789, + 0.335 + ], + "angle": 0, + "content": "Fig. 1: Tree-Ring imprints a broken watermark to a single channel. Our method RingID achieves stronger robustness by imprinting intact watermarks to multiple channels. In the task of multi-key identification, RingID shows overwhelming advantages." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.367, + 0.789, + 0.518 + ], + "angle": 0, + "content": "is considered a semantic watermark since the watermark changes the layout of the generated image and semantically hidden in it. It has demonstrated strong robustness against regeneration attacks [47] and various other image transformations [1, 38], credited to the meticulously designed tree-ring pattern. However, previous research only studies its performance in the verification task, which aims at distinguishing between watermarked and unwatermarked images. It's still unclear if Tree-Ring can be used to distribute multiple distinct keys and distinguish between them, a.k.a. watermark identification, which is crucial for source tracing and user attribution. Will Tree-Ring still be strong and robust as in the verification task? In this work, we will explore and answer this question." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.52, + 0.789, + 0.641 + ], + "angle": 0, + "content": "We start by studying the source of Tree-Ring's robustness under verification setting. For the first time, we uncover that the robustness of Tree-Ring stems not only from the carefully designed tree-ring pattern but also from the unintentionally introduced distribution shift during watermark imprinting. Such shift makes a big difference, especially in its robustness to Rotation and Crop/Scale image transformations, which cannot be handled by tree-ring's pattern design. More importantly, this overlooked power does not help in the identification scenario, because different watermarks experience the same shift." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.643, + 0.789, + 0.808 + ], + "angle": 0, + "content": "We further evaluate Tree-Ring in the identification task and surprisingly find that Tree-Ring struggles to distinguish between even dozens of different keys. Identification accuracy is reported in Tab. 2. The disappearance of the extra power of distribution shift has clearly exposed more design flaws in Tree-Ring. First, we find that tree-ring patterns are not discriminative enough to distinguish between different keys. This results in Tree-Ring's bad accuracy under various attacks or even under the clean setting. Meanwhile, the defective imprinting process leads to the injection of broken watermarks, further weakening its distinguishing ability. In addition, without the help of distribution shift, Tree-Ring is completely incapable of handling rotation and crop/scale transformations. Identification accuracy under these attacks is close to zero." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.81, + 0.789, + 0.842 + ], + "angle": 0, + "content": "In order to address the above issues, we propose RingID for enhanced identification capability. It bases on a novel Multi-Channel Heterogeneous watermarking" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.442, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "RingID: Enhanced Multi-Key Identification" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.253 + ], + "angle": 0, + "content": "framework, which can amalgamate distinctive advantages from different types of watermarks to resist to various kinds of attacks. Coupled with discretization and lossless imprinting methods, RingID gains substantial improvement on distinguishing ability. What's more, for the specific attack such as rotation, we locate the flaws in Tree-Ring's original design and propose systematic solutions, which improves the identification accuracy under rotation from nearly 0 to 0.86. To summarize our contributions:" + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.262, + 0.784, + 0.304 + ], + "angle": 0, + "content": "1. For the first time, we elucidate how the unintentional introduction of distribution shift substantially contributes to Tree-Ring's robustness. We demonstrate this effect from both mathematical and empirical perspectives." + }, + { + "type": "text", + "bbox": [ + 0.222, + 0.308, + 0.784, + 0.335 + ], + "angle": 0, + "content": "2. We systematically explore Tree-Ring's performance in the identification scenario and reveal its limitations through comprehensive analysis." + }, + { + "type": "text", + "bbox": [ + 0.222, + 0.338, + 0.784, + 0.394 + ], + "angle": 0, + "content": "3. We present RingID, a systematic solution that significantly enhances watermark verification and identification, delivering impressive results. The verification AUC improves from 0.975 to 0.995, while the identification accuracy rises from 0.07 to 0.82." + }, + { + "type": "list", + "bbox": [ + 0.222, + 0.262, + 0.784, + 0.394 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.42, + 0.388, + 0.435 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.444, + 0.403, + 0.458 + ], + "angle": 0, + "content": "2.1 Diffusion Models" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.465, + 0.788, + 0.587 + ], + "angle": 0, + "content": "Diffusion models [12, 18, 31] signify a swiftly advancing category of generative models. They define a forward diffusion process that gradually add small amount of Gaussian noise to the input data and learn to recover the clean data from the distortion [18]. Diffusion models initially demonstrate huge potential in generating high-resolution images [18, 32], resulting in globally recognized products like Midjourney [25], DALL-E [3, 26] and StableDiffusion [29]. Thereafter, they are broadly applied to modeling various types of data, e.g. video [4, 16, 30, 33, 39, 44], audio [19], 3d object [22, 27], humans [7, 8, 36, 50], etc." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.605, + 0.527, + 0.62 + ], + "angle": 0, + "content": "2.2 Watermarking Diffusion Models" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.629, + 0.789, + 0.841 + ], + "angle": 0, + "content": "When it comes to watermarking diffusion models, there are two different types of objectives: watermarking model weights [2, 23, 37, 45, 48] or generated content [6, 14, 35, 43, 46, 49]. Model weights watermarking methods aim to protect the intellectual property of the model itself and identify whether another model is an instance of plagiarism. While content watermarking methods protect the copyright of generated images [13, 42, 43] or videos [24, 46] by adding imperceptible perturbations onto pixels. They either imprint the watermark into the frequency via DFT [20], DCT [9, 11] and wavelet transformations [10, 40], or directly onto pixels via an encoder network [15, 35, 41, 49]. Recently, Tree-Ring [38] proposes to imprint a tree-ring watermark pattern into the initial diffusion noise. It demonstrates strong robustness to various attacks. In this work, we delve into Tree-Ring, uncovering its source of power in watermark verification and pinpointing its vulnerability in identifying multiple keys. Subsequently, we introduce RingID, a systematic solution fortified with enhanced identification capability." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.348, + 0.127 + ], + "angle": 0, + "content": "H. Ci et al." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.145, + 0.382, + 0.161 + ], + "angle": 0, + "content": "3 Preliminaries" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.185, + 0.379, + 0.2 + ], + "angle": 0, + "content": "3.1 Threat Model" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.218, + 0.788, + 0.339 + ], + "angle": 0, + "content": "Model Owner: The owner of the diffusion model provides the image generation service through API access. He would like to imprint imperceptible watermarks into every generated image for copyright protection and source tracking. Given an arbitrary image, the model owner would like to check whether the image contains his own watermark (a.k.a. Watermark Verification) and which watermark it is if he distributes multiple different watermarks (a.k.a. Identification). The goal of the model owner is to ensure high verification and identification accuracy regardless of any image transformations." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.341, + 0.788, + 0.417 + ], + "angle": 0, + "content": "Attacker: The attacker uses the model owner's service to generate an image. He attempts to apply various image transformations such as rotation and JPEG compression to disrupt the watermark, in order to evade the detection and identification check of the model owner. Subsequently, he can claim ownership of the image and profit from it illegally" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.447, + 0.337, + 0.46 + ], + "angle": 0, + "content": "3.2 Notation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.48, + 0.788, + 0.603 + ], + "angle": 0, + "content": "We denote the initial diffusion noise by \\( x_{T} \\), and its frequency counterpart after FFT by \\( X_{T} \\). \\( T \\) is the diffusion step. Sometimes we may omit the subscript \\( T \\) for simplicity. We denote the real part, imaginary part and conjugate symmetry of \\( X_{T} \\) by \\( X_{\\mathrm{re}} \\), \\( X_{\\mathrm{im}} \\), and \\( X_{\\mathrm{cs}} \\). In the context of StableDiffusion [29], \\( X_{T} \\in R^{h \\times w \\times c} \\) has 4 channels \\( \\mathcal{C} = \\{0,1,2,3\\} \\). We denote the watermarked channels by \\( \\mathcal{C}_w \\), where \\( \\mathcal{C}_w \\subset \\mathcal{C} \\). We denote the imprinted key by \\( w \\), the recovered key at detection time by \\( \\hat{w} \\), and the null watermark extracted from the unwatermarked image by \\( \\hat{w}_{\\emptyset} \\). We use \\( \\mathcal{W} = \\{w_1, w_2, \\ldots\\} \\) to denote a set of different keys. \\( |\\mathcal{W}| \\) is its capacity." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.628, + 0.657, + 0.643 + ], + "angle": 0, + "content": "3.3 Task Formulation: Verification vs. Identification" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.658, + 0.788, + 0.764 + ], + "angle": 0, + "content": "In this section, we formulate the verification and identification task following the problem statement of Tree-Ring [38]. The task of watermark verification is to detect the existence of watermarks. Verification algorithms attempt to distinguish between watermarked images carrying key \\(\\hat{w}\\) and unwatermarked images carrying null key \\(\\hat{w}_{\\emptyset}\\). More formally, verification can be formulated as distinguishing between the distance to the reference key \\(w\\): \\(d(\\hat{w}, w) \\neq d(\\hat{w}_{\\emptyset}, w)\\). In Tree-Ring, \\(d\\) is chosen to be \\(\\ell_{1}\\) distance and \\(w\\) is the watermark to be imprinted." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.788, + 0.84 + ], + "angle": 0, + "content": "When a set of different keys \\( w_{i} \\in \\mathcal{W} \\) are distributed, the goal of watermark identification is to find the best match for the recovered key \\( \\hat{w} \\) among all candidates in \\( \\mathcal{W} \\), i.e. \\( \\mathrm{ID}(\\hat{w}) = \\arg \\min_{i} [d(\\hat{w}, w_{i})] \\). In practice, different users are usually assigned different keys. Thus, each generated image can be source tracked and attributed to a unique user via watermark identification." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.442, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "RingID: Enhanced Multi-Key Identification" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.149, + 0.782, + 0.345 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.36, + 0.788, + 0.403 + ], + "angle": 0, + "content": "Fig. 2: Framework of the watermarking process. RingID introduces a series of approaches that can help to imprint a lossless and robust watermark. In contrast, TreeRing injects a lossy and less robust watermark." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.431, + 0.603, + 0.449 + ], + "angle": 0, + "content": "4 Delve into Tree-Ring Watermarking" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.465, + 0.788, + 0.511 + ], + "angle": 0, + "content": "In this section, we will reveal why Tree-Ring [38] is robust to various image transformations, even some of them are not considered by design. We further evaluate Tree-Ring in the multi-key identification task and pinpoint its vulnerabilities." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.535, + 0.424, + 0.551 + ], + "angle": 0, + "content": "4.1 Recap of Tree-Ring" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.562, + 0.788, + 0.76 + ], + "angle": 0, + "content": "Tree-Ring [38] originally studies the verification setting. Its goal is to distinguish watermarked images from non-watermarked images. As illustrated in Fig. 2, Tree-Ring transforms the initial diffusion noise \\( x_{T} \\) into the frequency \\( X_{T} \\) and injects a tree-ring pattern \\( w \\) into the center of \\( X_{T} \\), where the value of each ring in pattern \\( w \\) is sampled from Gaussian. It then transforms the watermarked \\( X_{T} \\) back to the spatial domain, retaining only the real part to feed the diffusion network for image generation. Notably, the accompanying imaginary part is discarded to be compatible with the network input. For detection of the watermark, Tree-Ring inverts the diffusion process to estimate \\( x_{T} \\) from an image, transforms it to the frequency domain, and extracts the watermark pattern \\( \\hat{w} \\). Similarly, nonsense pattern \\( \\hat{w}_{\\emptyset} \\) can be extracted from the non-watermarked image. Tree-Ring distinguishes between these two \\( \\ell_{1} \\)-to-reference distances: \\( \\| \\hat{w} - w \\|_{1} \\) and \\( \\| \\hat{w}_{\\emptyset} - w \\|_{1} \\) to determine whether an image is watermarked or not." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.783, + 0.557, + 0.799 + ], + "angle": 0, + "content": "4.2 Distribution Shift in Watermarking" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.81, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Tree-Ring's detection algorithm bases on the intuition that the recovered pattern \\(\\hat{w}\\) from a watermarked image should match the injected reference pattern \\(w\\)" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.35, + 0.128 + ], + "angle": 0, + "content": "H. Ci et al." + }, + { + "type": "image", + "bbox": [ + 0.248, + 0.146, + 0.756, + 0.242 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.248, + 0.244, + 0.756, + 0.3 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.311, + 0.788, + 0.354 + ], + "angle": 0, + "content": "Fig. 3: The effect of distribution shift in both verification and identification tasks. Discarding the imaginary part during watermark imprinting causes the distribution shift in \\(\\ell_1\\) distance. Distribution shift aids verification but not identification." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.384, + 0.789, + 0.414 + ], + "angle": 0, + "content": "to some extent, leading to a smaller \\(\\ell_1\\)-to-reference distance, i.e. \\(\\| \\hat{w} - w \\|_1 < \\| \\hat{w}_{\\emptyset} - w \\|_1\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.415, + 0.789, + 0.522 + ], + "angle": 0, + "content": "However, we have discovered that apart from the pattern matching, another crucial factor contributes to the separation of \\( \\| \\hat{w} - w \\|_1 \\) v.s. \\( \\| \\hat{w}_{\\emptyset} - w \\|_1 \\). This factor is the distribution shift introduced by discarding the imaginary part during the watermark injection process, which is highlighted in Fig. 2. This affects all watermarked images, further reducing the expectation of \\( \\mathbb{E}\\left[\\| \\hat{w} - w \\|_1\\right] \\). In other words, it shifts the distribution of \\( \\| \\hat{w} - w \\|_1 \\) towards a smaller value, making it easier to distinguish between \\( \\| \\hat{w} - w \\|_1 \\) and \\( \\| \\hat{w}_{\\emptyset} - w \\|_1 \\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.522, + 0.789, + 0.569 + ], + "angle": 0, + "content": "Mathematically, we prove that the shift factor is \\(\\frac{\\sqrt{3}}{2}\\) in a general and simplified setting. Let \\(\\hat{w}_n\\) denote the recovered watermark that has never been shifted, thus we have the following:" + }, + { + "type": "equation", + "bbox": [ + 0.375, + 0.581, + 0.787, + 0.613 + ], + "angle": 0, + "content": "\\[\n\\mathbb {E} \\left[ \\| \\hat {w} - w \\| _ {1} \\right] = \\frac {\\sqrt {3}}{2} \\mathbb {E} \\left[ \\| \\hat {w} _ {n} - w \\| _ {1} \\right]. \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.619, + 0.789, + 0.681 + ], + "angle": 0, + "content": "Detailed derivation is elaborated in Supplementary A. Actually, we find any operation leading to a sufficient distribution shift can be considered a good standalone watermarking approach. We demonstrate the operation of discarding the imaginary part performs well on its own in Supplementary B." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.704, + 0.79, + 0.842 + ], + "angle": 0, + "content": "Effect in Different Tasks and Attacks Fig. 3 illustrates the effect of distribution shift under different tasks. In Verification, it helps distinguish between watermarked and non-watermarked images. Next, we go one step further and consider the situation when the image undergoes different attacks like JPEG or Resize. If the tree-ring pattern itself can cope with a certain attack, the improvement brought by the distribution shift will be small, because the two distributions are already separable even without the help of shift. On the contrary, if the tree-ring pattern itself cannot deal with a certain attack, the improvement brought by the distribution shift will be significant. Empirically (Supplementary" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.442, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "RingID: Enhanced Multi-Key Identification" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.783, + 0.221 + ], + "angle": 0, + "content": "A.6), we find distribution shift helps a lot under Rotation and Crop/Scale attacks. This implies that the original design of tree-ring cannot handle these two attacks. When it comes to Identification, things are a little different. Since different watermarks experience the same distribution shift, discarding the imaginary part doesn't help distinguishing them." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.223, + 0.784, + 0.283 + ], + "angle": 0, + "content": "This sparks our curiosity about Tree-Ring's performance in identification, since distribution shift doesn't help and identification is more difficult. We would like to know if Tree-Ring still works under Rotation and Crop/Scale attacks. In the next section, we will explore these questions." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.303, + 0.563, + 0.318 + ], + "angle": 0, + "content": "4.3 Is Tree-Ring Good at Identification?" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.325, + 0.785, + 0.37 + ], + "angle": 0, + "content": "We report the performance of Tree-Ring when identifying different numbers of keys in Tab. 2. Tree-Ring struggles to identify as few as 32 keys. What's more, identification accuracy is nearly zero under Rotation and Crop/Scale attacks." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.371, + 0.784, + 0.461 + ], + "angle": 0, + "content": "We summarize 2 reasons as follows: (1) Distribution shift doesn't help. As tree-ring watermark cannot handle Rotation and Crop/Scale attacks, identification accuracy under these attacks drops to nearly zero. (2) Tee-ring pattern doesn't have enough distinguishing ability. Since identification is more difficult than verification, poor distinguishing ability results in poor performance under most attacks or even under no attack setting." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.481, + 0.747, + 0.496 + ], + "angle": 0, + "content": "4.4 Why is Tree-Ring Vulnerable to Rotation and Crop/Scale?" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.504, + 0.783, + 0.533 + ], + "angle": 0, + "content": "All previous experimental results and analysis show one thing, tree-ring pattern cannot handle these two attacks. In this section, we will see what happens." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.534, + 0.784, + 0.623 + ], + "angle": 0, + "content": "Vulnerable to Crop & Scale Scaling an image by a factor of \\(a\\) corresponds to compressing the frequency spectrum \\(X[u,v]\\) in both magnitudes and positions: \\(\\frac{1}{a^2} X\\left[\\frac{u}{a},\\frac{v}{a}\\right]\\). Cropping an image modulates the frequency magnitudes by a sinc function. It is straightforward that these transformations easily disrupt the tree-ring pattern on frequency domain. As shown in Fig. 4(e), the pattern is completely disrupted." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.624, + 0.784, + 0.699 + ], + "angle": 0, + "content": "Vulnerable to Rotation Although Tree-Ring is designed as ring patterns to seek rotation robustness, we find the existing design ignores many details, leading to no rotation robustness. In short, a combination of corner cropping when rotating, a lossy injection process, and the rotationally asymmetric shape cause the problem. We will explain in detail and propose solutions in Sec. 5.2." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.721, + 0.746, + 0.737 + ], + "angle": 0, + "content": "5 RingID: Enhanced Watermarking for Identification" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.784, + 0.84 + ], + "angle": 0, + "content": "In this section, we show how our proposed method RingID advances in identifying multiple distinct keys. RingID originates from Tree-Ring but incorporates systematic enhancements. We introduce a multi-channel heterogeneous watermarking framework, a discretization approach and a method of lossless imprinting to generally improve its distinguishing ability. We also discuss a series of designs regarding stronger rotation robustness and increased capacity." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.35, + 0.128 + ], + "angle": 0, + "content": "H. Ci et al." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.147, + 0.635, + 0.163 + ], + "angle": 0, + "content": "5.1 Multi-Channel Heterogeneous Watermarking" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.173, + 0.788, + 0.264 + ], + "angle": 0, + "content": "Multi-channel heterogeneous (MCH) watermarking serves as a general framework to comprehensively improve watermarking robustness to a battery of attacks. Intuitively, we imprint different types of watermarks on different channels of the initial noise to amalgamate their distinctive advantages, as illustrated in Fig. 1. During matching, we calculate the minimum of channel-wise \\(\\ell_1\\) distance to the reference watermark for all watermarked channels:" + }, + { + "type": "equation", + "bbox": [ + 0.357, + 0.273, + 0.786, + 0.306 + ], + "angle": 0, + "content": "\\[\n\\operatorname {I D} (\\hat {w}) = \\arg \\min _ {i} \\left\\{\\min _ {c \\in C _ {w}} \\left[ \\lambda_ {c} \\| \\hat {w} ^ {c} - w _ {i} ^ {c} \\| _ {1} \\right] \\right\\} \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.315, + 0.784, + 0.36 + ], + "angle": 0, + "content": "where \\( c \\) denotes the channel index and traverses all watermarked channels \\( C_w \\). \\( i \\) denotes the index of candidate reference watermarks. \\( \\lambda_c \\) is a channel-wise normalizing factor." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.361, + 0.788, + 0.572 + ], + "angle": 0, + "content": "Eq. (2) bases on the intuition that if a watermark is robust to a certain attack, then its distance from the ground-truth reference is smaller than that of other watermarks that are not robust. Ideally, it enables adaptive selection of the most robust watermark under different attacks. Notably, the total capacity of MCH is the minimum of each watermark. Therefore, it's recommended to use a watermark pattern with sufficient capacity on each channel. At the same time, in order to reduce the impact on the generation quality, we need to carefully select the type of watermarks for combination. We find the Gaussian noise pattern is a perfect candidate. It has infinite capacity and the same distribution as the initial noise. Meanwhile, it has good robustness to non-geometric attacks [38] and is complementary to the tree-ring pattern. So RingID explores the combination of a Gaussian noise watermark and a tree-ring watermark. Empirical study in Sec. 6.3 demonstrates that this combination is able to perfectly amalgamate unique advantages from the noise and ring watermark." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.595, + 0.494, + 0.611 + ], + "angle": 0, + "content": "5.2 Being Invariant to Rotation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.621, + 0.787, + 0.667 + ], + "angle": 0, + "content": "Although the tree-ring pattern is designed to be rotationally symmetric, it exhibits unexpected vulnerability to rotation as shown in Tab. 2. Below, we identify the causes of the issue and propose a series of measures to address it." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.789, + 0.844 + ], + "angle": 0, + "content": "Spatial Shift to Avoid Corner Cropping We visualize the imprinted tree-ring watermark in spatial domain in Fig. 4(c). Due to the properties of FFT, tree-ring patterns in the frequency tends to exhibit concentrated energy in the four corners in the spatial domain, rendering them vulnerable to cropping during rotation. We propose to circularly shift the watermark pattern in the spatial domain to the center of the image, thus the majority of watermark information is kept while rotating. For an \\( N \\times N \\) latent image, we circularly shift the pattern by \\( \\frac{N}{2} \\) pixels in both width and height dimension. This is equivalent to directly multiplying the frequency watermark with a chessboard pattern given by \\( H[u,v] = e^{j\\frac{\\pi}{2}(u + v)} = (-1)^{u + v} \\). The change of the watermark can be clearly" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.442, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "RingID: Enhanced Multi-Key Identification" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "image_caption", + "bbox": [ + 0.266, + 0.153, + 0.381, + 0.166 + ], + "angle": 0, + "content": "(a) Discretization" + }, + { + "type": "image", + "bbox": [ + 0.25, + 0.188, + 0.317, + 0.258 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.331, + 0.189, + 0.397, + 0.259 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.444, + 0.153, + 0.56, + 0.166 + ], + "angle": 0, + "content": "(b) Rounder Ring" + }, + { + "type": "image", + "bbox": [ + 0.428, + 0.167, + 0.495, + 0.287 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.509, + 0.167, + 0.576, + 0.294 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.604, + 0.153, + 0.756, + 0.295 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.338, + 0.305, + 0.491, + 0.318 + ], + "angle": 0, + "content": "(d) Lossless Imprinting" + }, + { + "type": "image", + "bbox": [ + 0.285, + 0.319, + 0.53, + 0.445 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.618, + 0.305, + 0.726, + 0.318 + ], + "angle": 0, + "content": "(e) Crop & Scale" + }, + { + "type": "image", + "bbox": [ + 0.598, + 0.319, + 0.747, + 0.445 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.457, + 0.788, + 0.569 + ], + "angle": 0, + "content": "Fig. 4: Detailed illustration of different acts and their impacts. (a) Discretization: Discretize values to \\(\\pm \\alpha\\) for each ring, increasing the distinguishing ability of the watermark. (b) Rounder ring: Render smoother edges and reduce artifacts. Increase rotational robustness. (c) Spatial Shift: Shift energy on the corners to the center, avoiding potential cropping during rotation. (d) Lossless Imprinting: Ensure the imprinted pattern the same as the original design. (e) Crop & Scale: Zooming in the spatial domain compresses the imprinted pattern in the frequency domain, causing pattern mismatch and hinders identification." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.601, + 0.788, + 0.693 + ], + "angle": 0, + "content": "observed in Fig. 4(c). Although spatial-shifting helps a lot for rotation invariance, we noticed that this often leads to the generation of circular artifacts at the center of the image. Keeping generation quality in mind, we further multiply the shifted pattern by a factor \\(\\eta\\) to suppress the peak at the center. \\(\\eta\\) trades-off the robustness, as smaller \\(\\eta\\) weakens the watermark pattern. In practice, we find \\(\\eta \\in [0.8,0.9]\\) strikes a good balance between image quality and robustness." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.719, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Lossless Imprinting The operation of discarding the imaginary part during watermarking not only shifts the distribution of \\(\\ell_1\\)-to-reference, but also disrupt the pattern, resulting in the loss of its rotational symmetry. Fig. 4(d) visualizes the actually carried pattern. More concretely, recall from Fig. 2, when the watermarked noise \\(X[u,v]\\) is transformed back to the spatial \\(x[m,n]\\), only its real part \\(x_{\\mathrm{re}}[m,n]\\) is kept for image generation. From a frequency domain perspective, this is equivalent to generation with its conjugate symmetry \\(X_{\\mathrm{cs}}[u,v]\\), which is not necessarily equal to \\(X[u,v]\\). To leverage the rotational symmetry of the" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.348, + 0.127 + ], + "angle": 0, + "content": "H. Ci et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.254 + ], + "angle": 0, + "content": "ring patterns, we attempt to eliminate the information loss during imprinting by solving \\( X[u,v] = X_{\\mathrm{cs}}[u,v] \\). The results indicate that \\( X_{\\mathrm{re}}[u,v] \\) should be an even function and \\( X_{\\mathrm{im}}[u,v] \\) should be an odd function about the Fourier center. To satisfy these constraints, we imprint the tree-ring pattern only in the real part and leave the imaginary part blank. We also align the center of the tree-ring pattern with the Fourier center (32,32). Originally, the tree-ring pattern was put at (31,32) and imprinted on both the real and the imaginary." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.279, + 0.79, + 0.46 + ], + "angle": 0, + "content": "Making A Rounder Ring Rings are originally determined by locating pixels equidistant from a give circle center \\((x_0, y_0)\\): \\((x - x_0)^2 + (y - y_0)^2 = r^2\\). We observed that this simple approach does not yield a well-rounded circular ring, particularly in images with lower resolutions like initial noise of shape \\(64 \\times 64\\). Artifacts and alias are shown in Fig. 4(b). Although it is centrosymmetric, it lacks rotational symmetry. Well, the solution is quite straightforward: drawing a rounder ring. There are a number of tools that can help draw a much rounder ring. Here, we employ a very simple idea to get a rounder ring. We place a white pixel on a black background at a distance \\(r\\) from the rotation center. By rotating the low resolution image 360 degrees and recording the trajectory of this pixel, we obtain a rounder ring with a radius of \\(r\\). Despite its simplicity, empirical results in Tab. 3 show that it is crucial for rotational invariance." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.486, + 0.644, + 0.502 + ], + "angle": 0, + "content": "5.3 Discretization for Enhanced Distinguishability" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.514, + 0.789, + 0.68 + ], + "angle": 0, + "content": "Tree-Ring samples values for each ring from a Gaussian distribution to keep consistency with the distribution of initial noise. We find that this random sampling makes it exceedingly challenging to distinguish between two keys. To address this issue, we propose a discretization approach, specifying that the value for each ring can only be either \\(\\alpha\\) or \\(-\\alpha\\), illustrated in Fig. 4(a). Consequently, the watermark pattern for \\(n\\) rings has a capacity of \\(2^n\\). While this seemingly reduces the upper limit of watermark capacity, it significantly enhances its valid capacity, ensuring optimal utilization of all available slots. Our experiments reveal that setting \\(\\alpha\\) as the standard deviation of the initial noise ensures substantial distinguishability between different keys while reducing the impact on generation quality to a small level." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.707, + 0.427, + 0.723 + ], + "angle": 0, + "content": "5.4 Increasing Capacity" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.789, + 0.841 + ], + "angle": 0, + "content": "We consider two simple approaches to increase capacity: adjusting the number of rings in a single channel or imprinting rings in extra channels. We evaluate the advantages and drawbacks of both methods in Sec. 6.4. In summary, increasing the number of rings \\( |R| \\) in a single channel can effectively boost capacity, leading to \\( 2^{|R|} \\) keys. However, excessive ring count results in a noticeable decline in robustness and generation quality. Imprinting watermarks in multiple channels exponentially increases capacity by \\( n^{|c|} \\), where \\( c \\) is number of channels and \\( n \\) is" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.442, + 0.115, + 0.732, + 0.13 + ], + "angle": 0, + "content": "RingID: Enhanced Multi-Key Identification" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.193 + ], + "angle": 0, + "content": "the capacity of a single channel. However, it's likely to introduce ring-like artifacts in the center of generated images. Therefore, we opt to place the watermark in a single channel and enhance capacity by adjusting the range of rings." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.211, + 0.366, + 0.229 + ], + "angle": 0, + "content": "6 Experiment" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.238, + 0.428, + 0.254 + ], + "angle": 0, + "content": "6.1 Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.259, + 0.788, + 0.486 + ], + "angle": 0, + "content": "Following Tree-Ring [38], we experiment with the state-of-the-art opensource diffusion model StableDiffusion-V2 [29]. By default, RingID imprints a tree-ring watermark with radius 3-14 on the channel 3 of the initial noise and a Gaussian noise watermark on channels 0. We adopt discretization (to 64 and -64), spatial shift, rounder drawing, and lossless imprinting as ordered in Fig. 2. We report the ROC-AUC between 1000 watermarked and 1000 unwatermarked images as the verification metric. In identification, we report identification accuracy with different numbers of keys. Evaluation is performed under various image distortions, following the original setup of Tree-Ring [38], i.e., JPEG 25, rotation \\(75^{\\circ}\\), \\(75\\%\\) random C&S, blur kernel 8, noise std 0.1, brightness [0, 6]. We assess the generation quality with CLIP scores [28] by OpenCLIP-ViT/G [5] between the generated image and the prompts to measure text-image alignment, and Frechet Inception Distance (FID) [17] to quantify the similarity between generated and natural images. We generate 5000 images and evaluate FID on the COCO 2017 training dataset [21]. We set both generation and inversion steps to 50." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.503, + 0.486, + 0.518 + ], + "angle": 0, + "content": "6.2 Comparison with Baselines" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.523, + 0.788, + 0.674 + ], + "angle": 0, + "content": "Robustness In verification (Tab. 1), both RingID and Tree-Ring perfectly distinguish between watermarked and unwatermarked images when no attacks are performed, achieving AUC score of 1. However, RingID exhibit stronger robustness to various image distortions than Tree-Ring, achieving near-perfect average AUC score, even though it is not optimized for verification. In the more challenging identification task, (Tab. 2), RingID significantly outperforms Tree-Ring. When identifying 32 keys, Tree-Ring's average identification accuracy (excluding C&S) is 0.465, while RingID's accuracy reaches up to 0.992. As the number of keys increases to 2048, this gap widens drastically (0.077 v.s. 0.942), further highlighting RingID's superior robustness and larger valid key capacity." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.675, + 0.788, + 0.78 + ], + "angle": 0, + "content": "Notably, both methods struggle with the C&S distortion in identification (Tab. 2). This aligns with our analysis in Sec. 4.2. Scaling in the spatial domain causes inverse scaling in the frequency domain, leading to pattern disruption and rendering identification challenging under such distortions. However, in verification, with the help of distribution shift, watermarked images can still be effectively distinguished from unwatermarked ones under C&S. Both Tree-Ring and RingID achieve average AUC higher than 0.97." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.78, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Image Quality RingID achieves strong performance without compromising the generation quality. It maintains comparable text-image alignments with TreeRing. Both methods achieve similar CLIP scores (0.364 v.s. 0.365). RingID obtains an FID of 26.13, which is also comparable to Tree-Ring's FID of 25.93." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.348, + 0.127 + ], + "angle": 0, + "content": "H. Ci et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.788, + 0.174 + ], + "angle": 0, + "content": "Table 1: Comparison with Tree-Ring in the verification task. The table shows the ROC-AUC values under various image distortions and CLIP Scores." + }, + { + "type": "table", + "bbox": [ + 0.217, + 0.186, + 0.787, + 0.251 + ], + "angle": 0, + "content": "
MethodsRing RadiusCleanRotateJPEGC&SBlurNoiseBrightnessAvgCLIP Score
Tree-Ring0-101.0000.9350.9990.9610.9990.9440.9830.9750.364
RingID0-101.0001.0001.0000.9790.9940.9690.9910.9900.359
3-141.0001.0001.0000.9870.9890.9980.9940.9950.365
" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.255, + 0.787, + 0.283 + ], + "angle": 0, + "content": "Table 2: Comparison with Tree-Ring in the identification task. We report the identification accuracy under various distortions for different numbers of keys." + }, + { + "type": "table", + "bbox": [ + 0.218, + 0.296, + 0.787, + 0.403 + ], + "angle": 0, + "content": "
Methods#KeysCleanRotateJPEGC&SBlurNoiseBrightnessAvg\\(Avg_{noC&S}\\)
Tree-Ring320.7900.0200.4200.0400.6100.5300.4200.4040.465
1280.4500.0100.1200.0200.2800.2300.1700.1830.210
20480.2000.0000.0400.0000.0900.0700.0600.0660.077
RingID321.0001.0001.0000.5300.9901.0000.9600.9260.992
1281.0000.9801.0000.2800.9801.0000.9400.8830.983
20481.0000.8601.0000.0800.9700.9500.8700.8190.942
" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.433, + 0.389, + 0.449 + ], + "angle": 0, + "content": "6.3 Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.458, + 0.787, + 0.548 + ], + "angle": 0, + "content": "Components of RingID Tab. 3 ablates all proposed components of RingID. We can find that all designs are crucial for the final success of RingID. Spatial shift, lossless imprinting, and discretization play a significant role in the robustness of rotation. While the use of rounder rings also increase the accuracy by approximately \\(40\\%\\) (from 0.620 to 0.860). Notably, spatial shift sacrifices the robustness to blur and brightness attacks for rotation robustness." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.549, + 0.787, + 0.686 + ], + "angle": 0, + "content": "Heterogeneous Watermarking RingID employs the combination of a Gaussian noise watermark and a ring watermark. Here, we ablate the choice of imprinting the noise watermark on different channels. By comparing the first 3 rows of Tab. 4, we find that heterogeneous watermarking can effectively combine the advantage from both the noise watermark and the ring watermark. Among single-channel configurations, imprinting noise watermark on channel 0 (0.819) achieves the best performance. The combination of channels 0 and 1 yields the highest accuracy among all trials. Interestingly, imprinting noise watermarks on all available channels (0, 1, 2) decrease the accuracy from 0.830 to 0.807." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.709, + 0.35, + 0.723 + ], + "angle": 0, + "content": "6.4 Discussion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.734, + 0.787, + 0.841 + ], + "angle": 0, + "content": "Where is the Best Place to Put Rings? We could imprint all rings on one channel or distribute them evenly across multiple channels. In Tab. 5, we fix the total number of rings to 12 and distribute them to 1-4 channels. we can find that performance drops as we distribute rings to more channels. What's more, imprinting rings on multiple channels cause artifacts to appear more frequently (shown in Supplementary C). So given fixed capacity, it's better to imprint all rings on a single channel. This raises another question: what are the optimal" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.442, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "RingID: Enhanced Multi-Key Identification" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "table_caption", + "bbox": [ + 0.217, + 0.145, + 0.785, + 0.174 + ], + "angle": 0, + "content": "Table 3: Ablation study on the effect of each proposed module in RingID. We report the identification accuracy for 2048 keys when removing individual components." + }, + { + "type": "table", + "bbox": [ + 0.218, + 0.186, + 0.787, + 0.294 + ], + "angle": 0, + "content": "
SettingsCleanRotateJPEGC&SBlurNoiseBrightnessAvg
RingID1.0000.8601.0000.0800.9700.9500.8700.819
- Spatial Shift1.0000.0001.0000.0500.9900.9300.9400.701
- Lossless Imprinting1.0000.0100.9700.1600.9500.9800.8300.700
- Rounder Ring1.0000.6200.9900.1000.8900.9700.8500.774
- Discretization0.9800.1200.3800.0200.4500.6500.3900.427
- Heterogeneous1.0000.8200.9400.0300.9600.7100.7200.740
" + }, + { + "type": "table_caption", + "bbox": [ + 0.217, + 0.325, + 0.785, + 0.368 + ], + "angle": 0, + "content": "Table 4: Ablation study on multi-channel heterogeneous watermarking. This table reports the identification accuracy when the noise watermark is imprinted on different channels (abbr. CHs). “-” means the watermark is not imprinted." + }, + { + "type": "table", + "bbox": [ + 0.218, + 0.38, + 0.787, + 0.512 + ], + "angle": 0, + "content": "
Ring CHsNoise CHsCleanRotateJPEGC&SBlurNoiseBrightnessAvg
-01.0000.0001.0000.1700.9900.9500.9200.719
3-1.0000.8200.9400.0300.9600.7100.7200.740
301.0000.8601.0000.0800.9700.9500.8700.819
311.0000.8201.0000.1301.0000.8800.7700.800
321.0000.8400.9800.1000.9200.9900.7900.803
30,11.0000.8500.9500.1500.9900.9900.8800.830
30,21.0000.8400.9900.1200.9500.9800.8700.821
30,1,21.0000.8300.9600.1600.9300.9800.7900.807
" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.559, + 0.785, + 0.604 + ], + "angle": 0, + "content": "inner and outer radii if all rings are placed on a single channel? From Tab. 6 we can see that 3-14 forms a good choice for 11 rings with capacity 2048. The optimal ring radius also varies for different ring numbers." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.613, + 0.785, + 0.688 + ], + "angle": 0, + "content": "More Keys We would like to explore whether RingID can handle a larger number of keys. Fig. 5 shows the average identification accuracy under various attacks \\( v.s. \\) number of keys. We observe that RingID maintains good accuracy even with a large number of keys. And identification accuracy linearly decreases as the number of keys exponentially increase." + }, + { + "type": "table_caption", + "bbox": [ + 0.218, + 0.733, + 0.784, + 0.748 + ], + "angle": 0, + "content": "Table 6: Influence of ring radius on identification accuracy and text-image alignment." + }, + { + "type": "table", + "bbox": [ + 0.218, + 0.762, + 0.787, + 0.837 + ], + "angle": 0, + "content": "
Radius#KeysAccCLIP ScoreRadius#KeysAccCLIP ScoreRadius#KeysAccCLIP Score
1-6320.8860.3631-92560.8190.3621-1220480.7860.361
2-7320.8760.3662-102560.8560.3582-1320480.8040.361
3-8320.8970.3653-112560.8500.3623-1420480.8190.368
4-9320.8970.3684-122560.8470.3634-1520480.8130.363
5-10320.8810.3615-132560.8390.3605-1620480.8030.360
" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.348, + 0.127 + ], + "angle": 0, + "content": "H. Ci et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.167, + 0.488, + 0.223 + ], + "angle": 0, + "content": "Table 5: Distributing 12 rings across different number of channels (#CHs). This table reports the average identification accuracy and CLIP Score." + }, + { + "type": "table", + "bbox": [ + 0.218, + 0.227, + 0.487, + 0.307 + ], + "angle": 0, + "content": "
Ring Radius #CHsAvgCLIP Score
3-1510.7400.357
3-920.7370.351
3-730.6640.351
3-640.5440.340
" + }, + { + "type": "image", + "bbox": [ + 0.534, + 0.165, + 0.777, + 0.3 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.527, + 0.301, + 0.786, + 0.328 + ], + "angle": 0, + "content": "Fig. 5: Average identification accuracy \\( v.s. \\) number of keys." + }, + { + "type": "image", + "bbox": [ + 0.242, + 0.348, + 0.769, + 0.515 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.53, + 0.788, + 0.559 + ], + "angle": 0, + "content": "Fig. 6: Qualitative results. The subtle changes in image layouts show that both TreeRing and RingID embed watermarks into the image's semantic content." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.589, + 0.422, + 0.604 + ], + "angle": 0, + "content": "6.5 Qualitative Results" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.615, + 0.788, + 0.692 + ], + "angle": 0, + "content": "As shown in Fig. 6, both RingID and Tree-Ring watermarking can generate images of good visual quality. Unlike low-level watermarking techniques, both RingID and Tree-Ring cause subtle changes in the image layout. However, RingID generates images with layout that is closer to the unwatermarked image. We owe this to avoiding watermarking on the low-frequency part (radius 0-3)." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.716, + 0.51, + 0.733 + ], + "angle": 0, + "content": "7 Conclusion and Limitation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.786, + 0.794 + ], + "angle": 0, + "content": "We conduct a comprehensive investigation on Tree-Ring Watermarking, revealing its limitations in identifying multiple keys. We then propose RingID to enhance identification capability." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.795, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Limitation: Under multi-key identification, one limitation is that both TreeRing and our RingID are vulnerable to cropping and scaling attacks. Future work may explore other transform domains to resist these attacks." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.442, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "RingID: Enhanced Multi-Key Identification" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.192 + ], + "angle": 0, + "content": "Acknowledgment: This project is supported by the Ministry of Education, Singapore, under the Academic Research Fund Tier 1 (FY2022) Award 22-5406-A0001." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.218, + 0.323, + 0.234 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.251, + 0.785, + 0.294 + ], + "angle": 0, + "content": "1. An, B., Ding, M., Rabbani, T., Agrawal, A., Xu, Y., Deng, C., Zhu, S., Mohamed, A., Wen, Y., Goldstein, T., et al.: Benchmarking the robustness of image watermarks. arXiv preprint arXiv:2401.08573 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.295, + 0.785, + 0.35 + ], + "angle": 0, + "content": "2. Bansal, A., Chiang, P.y., Curry, M.J., Jain, R., Wigington, C., Manjunatha, V., Dickerson, J.P., Goldstein, T.: Certified neural network watermarks with randomized smoothing. In: International Conference on Machine Learning. pp. 1450-1465. PMLR (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.351, + 0.785, + 0.392 + ], + "angle": 0, + "content": "3. Betker, J., Goh, G., Jing, L., Brooks, T., Wang, J., Li, L., Ouyang, L., Zhuang, J., Lee, J., Guo, Y., et al.: Improving image generation with better captions. Computer Science. https://cdn.opencai.com/papers/dall-e-3.pdf 2(3), 8 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.393, + 0.785, + 0.448 + ], + "angle": 0, + "content": "4. Blattmann, A., Dockhorn, T., Kulal, S., Mendelevitch, D., Kilian, M., Lorenz, D., Levi, Y., English, Z., Voleti, V., Letts, A., et al.: Stable video diffusion: Scaling latent video diffusion models to large datasets. arXiv preprint arXiv:2311.15127 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.449, + 0.785, + 0.504 + ], + "angle": 0, + "content": "5. Cherti, M., Beaumont, R., Wightman, R., Wortsman, M., Ilharco, G., Gordon, C., Schuhmann, C., Schmidt, L., Jitsev, J.: Reproducible scaling laws for contrastive language-image learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2818-2829 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.505, + 0.785, + 0.533 + ], + "angle": 0, + "content": "6. Ci, H., Song, Y., Yang, P., Xie, J., Shou, M.Z.: Wmadapter: Adding watermark control to latent diffusion models. arXiv preprint arXiv:2406.08337 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.533, + 0.785, + 0.574 + ], + "angle": 0, + "content": "7. Ci, H., Wang, C., Ma, X., Wang, Y.: Optimizing network structure for 3d human pose estimation. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 2262-2271 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.575, + 0.785, + 0.616 + ], + "angle": 0, + "content": "8. Ci, H., Wu, M., Zhu, W., Ma, X., Dong, H., Zhong, F., Wang, Y.: Gfpose: Learning 3d human pose prior with gradient fields. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4800-4810 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.617, + 0.785, + 0.644 + ], + "angle": 0, + "content": "9. Cox, I.: Digital watermarking and steganography. Morgan Kaufmann google scholar 2, 893-914 (2007)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.645, + 0.785, + 0.673 + ], + "angle": 0, + "content": "0. Cui, S., Guo, X., Zhang, Z.: Estimation and inference in ultrahigh dimensional partially linear single-index models. arXiv preprint arXiv:2404.04471 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.673, + 0.785, + 0.714 + ], + "angle": 0, + "content": "1. Cui, S., Sudjianto, A., Zhang, A., Li, R.: Enhancing robustness of gradientboosted decision trees through one-hot encoding and regularization. arXiv preprint arXiv:2304.13761 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.715, + 0.785, + 0.743 + ], + "angle": 0, + "content": "2. Dhariwal, P., Nichol, A.: Diffusion models beat gans on image synthesis. Advances in neural information processing systems 34, 8780-8794 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.744, + 0.785, + 0.785 + ], + "angle": 0, + "content": "3. Fernandez, P., Couairon, G., Jégou, H., Douze, M., Furon, T.: The stable signature: Rooting watermarks in latent diffusion models. arXiv preprint arXiv:2303.15435 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.785, + 0.785, + 0.84 + ], + "angle": 0, + "content": "4. Fernandez, P., Sablayrolles, A., Furon, T., Jégou, H., Douze, M.: Watermarking images in self-supervised latent spaces. In: ICASSP 2022-2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). pp. 3054-3058. IEEE (2022)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.251, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.348, + 0.127 + ], + "angle": 0, + "content": "H. Ci et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.787, + 0.189 + ], + "angle": 0, + "content": "15. Guo, H., Zhang, Q., Luo, J., Guo, F., Zhang, W., Su, X., Li, M.: Practical deep dispersed watermarking with synchronization and fusion. In: Proceedings of the 31st ACM International Conference on Multimedia. pp. 7922-7932 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.19, + 0.787, + 0.245 + ], + "angle": 0, + "content": "16. Guo, Y., Yang, C., Rao, A., Liang, Z., Wang, Y., Qiao, Y., Agrawala, M., Lin, D., Dai, B.: Animatediff:Animate your personalized text-to-image diffusion models without specific tuning. International Conference on Learning Representations (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.247, + 0.787, + 0.288 + ], + "angle": 0, + "content": "17. Heusel, M., Ramsauer, H., Unterthiner, T., Nessler, B., Hochreiter, S.: Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems 30 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.289, + 0.787, + 0.317 + ], + "angle": 0, + "content": "18. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. Advances in neural information processing systems 33, 6840-6851 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.318, + 0.787, + 0.346 + ], + "angle": 0, + "content": "19. Kong, Z., Ping, W., Huang, J., Zhao, K., Catanzaro, B.: Diffwave: A versatile diffusion model for audio synthesis. arXiv preprint arXiv:2009.09761 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.346, + 0.787, + 0.388 + ], + "angle": 0, + "content": "20. Lin, C.Y., Wu, M., Bloom, J.A., Cox, I.J., Miller, M.L., Lui, Y.M.: Rotation, scale, and translation resilient watermarking for images. IEEE Transactions on image processing 10(5), 767-782 (2001)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.389, + 0.787, + 0.444 + ], + "angle": 0, + "content": "21. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollar, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13. pp. 740-755. Springer (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.445, + 0.787, + 0.487 + ], + "angle": 0, + "content": "22. Liu, R., Wu, R., Van Hoorick, B., Tokmakov, P., Zakharov, S., Vondrick, C.: Zero-1-to-3: Zero-shot one image to 3d object. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 9298-9309 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.488, + 0.787, + 0.515 + ], + "angle": 0, + "content": "23. Liu, Y., Li, Z., Backes, M., Shen, Y., Zhang, Y.: Watermarking diffusion model. arXiv preprint arXiv:2305.12502 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.516, + 0.787, + 0.557 + ], + "angle": 0, + "content": "24. Luo, X., Li, Y., Chang, H., Liu, C., Milanfar, P., Yang, F.: Dvmark: a deep multiscale framework for video watermarking. IEEE Transactions on Image Processing (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.558, + 0.685, + 0.572 + ], + "angle": 0, + "content": "25. Midjourney.com: Midjourney. https://www.midjourney.com/home" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.573, + 0.57, + 0.587 + ], + "angle": 0, + "content": "26. OpenAI: Dalle2. https://openai.com/dall-e-2" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.588, + 0.787, + 0.615 + ], + "angle": 0, + "content": "27. Poole, B., Jain, A., Barron, J.T., Mildenhall, B.: Dreamfusion: Text-to-3d using 2d diffusion. arXiv preprint arXiv:2209.14988 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.616, + 0.787, + 0.671 + ], + "angle": 0, + "content": "28. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.672, + 0.787, + 0.727 + ], + "angle": 0, + "content": "29. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 10684-10695 (June 2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.728, + 0.787, + 0.77 + ], + "angle": 0, + "content": "30. Singer, U., Polyak, A., Hayes, T., Yin, X., An, J., Zhang, S., Hu, Q., Yang, H., Ashual, O., Gafni, O., et al.: Make-a-video: Text-to-video generation without text-video data. arXiv preprint arXiv:2209.14792 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.771, + 0.787, + 0.798 + ], + "angle": 0, + "content": "31. Song, J., Meng, C., Ermon, S.: Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.799, + 0.787, + 0.84 + ], + "angle": 0, + "content": "32. Song, Y., Sohl-Dickstein, J., Kingma, D.P., Kumar, A., Ermon, S., Poole, B.: Score-based generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456 (2020)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.442, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "RingID: Enhanced Multi-Key Identification" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.189 + ], + "angle": 0, + "content": "33. Song, Y., Huang, S., Yao, C., Ye, X., Ci, H., Liu, J., Zhang, Y., Shou, M.Z.: Processpainter: Learn painting process from sequence data. arXiv preprint arXiv:2406.06062 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.189, + 0.788, + 0.229 + ], + "angle": 0, + "content": "34. Steinberg, Y., Merhav, N.: Identification in the presence of side information with application to watermarking. IEEE Transactions on Information Theory 47(4), 1410-1422 (2001)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.229, + 0.788, + 0.27 + ], + "angle": 0, + "content": "35. Tancik, M., Mildenhall, B., Ng, R.: Stegastamp: Invisible hyperlinks in physical photographs. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 2117-2126 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.27, + 0.788, + 0.297 + ], + "angle": 0, + "content": "36. Tevet, G., Raab, S., Gordon, B., Shafir, Y., Cohen-Or, D., Bermano, A.H.: Human motion diffusion model. arXiv preprint arXiv:2209.14916 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.297, + 0.788, + 0.337 + ], + "angle": 0, + "content": "37. Uchida, Y., Nagai, Y., Sakazawa, S., Satoh, S.: Embedding watermarks into deep neural networks. In: Proceedings of the 2017 ACM on international conference on multimedia retrieval. pp. 269-277 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.337, + 0.788, + 0.377 + ], + "angle": 0, + "content": "38. Wen, Y., Kirchenbauer, J., Geiping, J., Goldstein, T.: Tree-ring watermarks: Fingerprints for diffusion images that are invisible and robust. arXiv preprint arXiv:2305.20030 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.377, + 0.788, + 0.432 + ], + "angle": 0, + "content": "39. Wu, J.Z., Ge, Y., Wang, X., Lei, S.W., Gu, Y., Shi, Y., Hsu, W., Shan, Y., Qie, X., Shou, M.Z.: Tune-a-video: One-shot tuning of image diffusion models for text-to-video generation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 7623-7633 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.432, + 0.788, + 0.458 + ], + "angle": 0, + "content": "40. Xia, X.G., Boncelet, C.G., Arce, G.R.: Wavelet transform based watermark for digital images. Optics Express 3(12), 497-511 (1998)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.458, + 0.788, + 0.498 + ], + "angle": 0, + "content": "41. Xiong, C., Qin, C., Feng, G., Zhang, X.: Flexible and secure watermarking for latent diffusion model. In: Proceedings of the 31st ACM International Conference on Multimedia. pp. 1668-1676 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.498, + 0.788, + 0.525 + ], + "angle": 0, + "content": "42. Yang, P., Ci, H., Song, Y., Shou, M.Z.: Steganalysis on digital watermarking: Is your defense truly impervious? arXiv preprint arXiv:2406.09026 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.525, + 0.788, + 0.565 + ], + "angle": 0, + "content": "43. Yu, N., Skripniuk, V., Abdelnabi, S., Fritz, M.: Artificial fingerprinting for generative models: Rooting deepfake attribution in training data. In: Proceedings of the IEEE/CVF International conference on computer vision. pp. 14448-14457 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.565, + 0.788, + 0.605 + ], + "angle": 0, + "content": "44. Zhang, D.J., Wu, J.Z., Liu, J.W., Zhao, R., Ran, L., Gu, Y., Gao, D., Shou, M.Z.: Show-1: Marrying pixel and latent diffusion models for text-to-video generation. arXiv preprint arXiv:2309.15818 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.605, + 0.788, + 0.659 + ], + "angle": 0, + "content": "45. Zhang, J., Gu, Z., Jang, J., Wu, H., Stoecklin, M.P., Huang, H., Molloy, I.: Protecting intellectual property of deep neural networks with watermarking. In: Proceedings of the 2018 on Asia conference on computer and communications security. pp. 159-172 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.659, + 0.788, + 0.686 + ], + "angle": 0, + "content": "46. Zhang, K.A., Xu, L., Cuesta-Infante, A., Veeramachaneni, K.: Robust invisible video watermarking with attention. arXiv preprint arXiv:1909.01285 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.686, + 0.788, + 0.726 + ], + "angle": 0, + "content": "47. Zhao, X., Zhang, K., Wang, Y.X., Li, L.: Generative autoencoders as watermark attackers: Analyses of vulnerabilities and threats. arXiv preprint arXiv:2306.01953 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.726, + 0.788, + 0.752 + ], + "angle": 0, + "content": "48. Zhao, Y., Pang, T., Du, C., Yang, X., Cheung, N.M., Lin, M.: A recipe for watermarking diffusion models. arXiv preprint arXiv:2303.10137 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.752, + 0.788, + 0.793 + ], + "angle": 0, + "content": "49. Zhu, J., Kaplan, R., Johnson, J., Fei-Fei, L.: Hidden: Hiding data with deep networks. In: Proceedings of the European conference on computer vision (ECCV). pp. 657-672 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.793, + 0.788, + 0.834 + ], + "angle": 0, + "content": "50. Zhu, W., Ma, X., Ro, D., Ci, H., Zhang, J., Shi, J., Gao, F., Tian, Q., Wang, Y.: Human motion generation: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence (2023)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.834 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/Rethinking Tree-Ring Watermarking for Enhanced Multi-Key Identification/dd0cc8f0-17b3-4928-9aeb-d79e0a1054db_origin.pdf b/2024/Rethinking Tree-Ring Watermarking for Enhanced Multi-Key Identification/dd0cc8f0-17b3-4928-9aeb-d79e0a1054db_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..e4e2ada6bba26291d44f48cb1d2e267aa80def2c --- /dev/null +++ b/2024/Rethinking Tree-Ring Watermarking for Enhanced Multi-Key Identification/dd0cc8f0-17b3-4928-9aeb-d79e0a1054db_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:37686a24e92d83cc3127049d78468729f2bb47e3990877b0bef4689113f664df +size 2223539 diff --git a/2024/Rethinking Tree-Ring Watermarking for Enhanced Multi-Key Identification/full.md b/2024/Rethinking Tree-Ring Watermarking for Enhanced Multi-Key Identification/full.md new file mode 100644 index 0000000000000000000000000000000000000000..351204208947b4272ac9ba1b11ffc37fa16754e8 --- /dev/null +++ b/2024/Rethinking Tree-Ring Watermarking for Enhanced Multi-Key Identification/full.md @@ -0,0 +1,303 @@ +# RingID: Rethinking Tree-Ring Watermarking for Enhanced Multi-Key Identification + +Hai $\mathrm{Ci}^{*}$ Pei Yang\*, Yiren Song\*, and Mike Zheng Shou + +Show Lab, National University of Singapore cihai03@gmail.com yangpei@u.nus.edu yiren@nus.edu.sg mike.zheng.shou@gmail.com + +Abstract. We revisit Tree-Ring Watermarking, a recent diffusion model watermarking method that demonstrates great robustness to various attacks. We conduct an in-depth study on it and reveal that the distribution shift unintentionally introduced by the watermarking process, apart from watermark pattern matching, contributes to its exceptional robustness. Our investigation further exposes inherent flaws in its original design, particularly in its ability to identify multiple distinct keys, where distribution shift offers no assistance. Based on these findings and analysis, we present RingID for enhanced multi-key identification. It consists of a novel multi-channel heterogeneous watermarking approach designed to seamlessly amalgamate distinctive advantages from diverse watermarks. Coupled with a series of suggested enhancements, RingID exhibits substantial advancements in multi-key identification. Github page: https://github.com/showlab/RingID + +Keywords: Diffusion Models $\cdot$ Tree-Ring Watermarking $\cdot$ Multi-Key Identification + +# 1 Introduction + +With the popularity of diffusion models, a vast number of high-quality diffusion generated images are circulating on the internet. It has become increasingly important to identify AI-generated images to detect fabrication and to track and protect copyrights. Luckily, watermarking [1, 13] offers a reliable solution. + +Research on watermarking has a long history in computer vision [9, 34]. A series of works [13, 14, 35, 48, 49] imprint imperceptible watermarks on images through small pixel perturbations. Recently, [47] demonstrates that pixel-level perturbations are provably removable by regeneration attacks. In contrast, Tree-Ring Watermarking [38] has recently introduced a novel approach, proposing to imprint a specially designed tree-ring pattern into the initial noise of a diffusion model. Different patterns are also called different "keys". The verification of the watermark's existence involves recovering the initial diffusion noise from the image and comparing the extracted key with the injected one. Tree-Ring + +![](images/31852c4dd9f7b48c201208e99c5e85b08affb1a9b1424b4e40c6061a03e13fd8.jpg) +Tree-Ring + +![](images/88e1bbd7d21261cb169f4e01d9250ef3323af44ea05d1127d6b6e3c9c1d40587.jpg) +RingID + +![](images/4d13a9715ed9f8335ef0e6a85f9504bc171571aef2017eab31f1e22196ec6e11.jpg) +Number of Keys +Fig. 1: Tree-Ring imprints a broken watermark to a single channel. Our method RingID achieves stronger robustness by imprinting intact watermarks to multiple channels. In the task of multi-key identification, RingID shows overwhelming advantages. + +is considered a semantic watermark since the watermark changes the layout of the generated image and semantically hidden in it. It has demonstrated strong robustness against regeneration attacks [47] and various other image transformations [1, 38], credited to the meticulously designed tree-ring pattern. However, previous research only studies its performance in the verification task, which aims at distinguishing between watermarked and unwatermarked images. It's still unclear if Tree-Ring can be used to distribute multiple distinct keys and distinguish between them, a.k.a. watermark identification, which is crucial for source tracing and user attribution. Will Tree-Ring still be strong and robust as in the verification task? In this work, we will explore and answer this question. + +We start by studying the source of Tree-Ring's robustness under verification setting. For the first time, we uncover that the robustness of Tree-Ring stems not only from the carefully designed tree-ring pattern but also from the unintentionally introduced distribution shift during watermark imprinting. Such shift makes a big difference, especially in its robustness to Rotation and Crop/Scale image transformations, which cannot be handled by tree-ring's pattern design. More importantly, this overlooked power does not help in the identification scenario, because different watermarks experience the same shift. + +We further evaluate Tree-Ring in the identification task and surprisingly find that Tree-Ring struggles to distinguish between even dozens of different keys. Identification accuracy is reported in Tab. 2. The disappearance of the extra power of distribution shift has clearly exposed more design flaws in Tree-Ring. First, we find that tree-ring patterns are not discriminative enough to distinguish between different keys. This results in Tree-Ring's bad accuracy under various attacks or even under the clean setting. Meanwhile, the defective imprinting process leads to the injection of broken watermarks, further weakening its distinguishing ability. In addition, without the help of distribution shift, Tree-Ring is completely incapable of handling rotation and crop/scale transformations. Identification accuracy under these attacks is close to zero. + +In order to address the above issues, we propose RingID for enhanced identification capability. It bases on a novel Multi-Channel Heterogeneous watermarking + +framework, which can amalgamate distinctive advantages from different types of watermarks to resist to various kinds of attacks. Coupled with discretization and lossless imprinting methods, RingID gains substantial improvement on distinguishing ability. What's more, for the specific attack such as rotation, we locate the flaws in Tree-Ring's original design and propose systematic solutions, which improves the identification accuracy under rotation from nearly 0 to 0.86. To summarize our contributions: + +1. For the first time, we elucidate how the unintentional introduction of distribution shift substantially contributes to Tree-Ring's robustness. We demonstrate this effect from both mathematical and empirical perspectives. +2. We systematically explore Tree-Ring's performance in the identification scenario and reveal its limitations through comprehensive analysis. +3. We present RingID, a systematic solution that significantly enhances watermark verification and identification, delivering impressive results. The verification AUC improves from 0.975 to 0.995, while the identification accuracy rises from 0.07 to 0.82. + +# 2 Related Work + +# 2.1 Diffusion Models + +Diffusion models [12, 18, 31] signify a swiftly advancing category of generative models. They define a forward diffusion process that gradually add small amount of Gaussian noise to the input data and learn to recover the clean data from the distortion [18]. Diffusion models initially demonstrate huge potential in generating high-resolution images [18, 32], resulting in globally recognized products like Midjourney [25], DALL-E [3, 26] and StableDiffusion [29]. Thereafter, they are broadly applied to modeling various types of data, e.g. video [4, 16, 30, 33, 39, 44], audio [19], 3d object [22, 27], humans [7, 8, 36, 50], etc. + +# 2.2 Watermarking Diffusion Models + +When it comes to watermarking diffusion models, there are two different types of objectives: watermarking model weights [2, 23, 37, 45, 48] or generated content [6, 14, 35, 43, 46, 49]. Model weights watermarking methods aim to protect the intellectual property of the model itself and identify whether another model is an instance of plagiarism. While content watermarking methods protect the copyright of generated images [13, 42, 43] or videos [24, 46] by adding imperceptible perturbations onto pixels. They either imprint the watermark into the frequency via DFT [20], DCT [9, 11] and wavelet transformations [10, 40], or directly onto pixels via an encoder network [15, 35, 41, 49]. Recently, Tree-Ring [38] proposes to imprint a tree-ring watermark pattern into the initial diffusion noise. It demonstrates strong robustness to various attacks. In this work, we delve into Tree-Ring, uncovering its source of power in watermark verification and pinpointing its vulnerability in identifying multiple keys. Subsequently, we introduce RingID, a systematic solution fortified with enhanced identification capability. + +# 3 Preliminaries + +# 3.1 Threat Model + +Model Owner: The owner of the diffusion model provides the image generation service through API access. He would like to imprint imperceptible watermarks into every generated image for copyright protection and source tracking. Given an arbitrary image, the model owner would like to check whether the image contains his own watermark (a.k.a. Watermark Verification) and which watermark it is if he distributes multiple different watermarks (a.k.a. Identification). The goal of the model owner is to ensure high verification and identification accuracy regardless of any image transformations. + +Attacker: The attacker uses the model owner's service to generate an image. He attempts to apply various image transformations such as rotation and JPEG compression to disrupt the watermark, in order to evade the detection and identification check of the model owner. Subsequently, he can claim ownership of the image and profit from it illegally + +# 3.2 Notation + +We denote the initial diffusion noise by $x_{T}$ , and its frequency counterpart after FFT by $X_{T}$ . $T$ is the diffusion step. Sometimes we may omit the subscript $T$ for simplicity. We denote the real part, imaginary part and conjugate symmetry of $X_{T}$ by $X_{\mathrm{re}}$ , $X_{\mathrm{im}}$ , and $X_{\mathrm{cs}}$ . In the context of StableDiffusion [29], $X_{T} \in R^{h \times w \times c}$ has 4 channels $\mathcal{C} = \{0,1,2,3\}$ . We denote the watermarked channels by $\mathcal{C}_w$ , where $\mathcal{C}_w \subset \mathcal{C}$ . We denote the imprinted key by $w$ , the recovered key at detection time by $\hat{w}$ , and the null watermark extracted from the unwatermarked image by $\hat{w}_{\emptyset}$ . We use $\mathcal{W} = \{w_1, w_2, \ldots\}$ to denote a set of different keys. $|\mathcal{W}|$ is its capacity. + +# 3.3 Task Formulation: Verification vs. Identification + +In this section, we formulate the verification and identification task following the problem statement of Tree-Ring [38]. The task of watermark verification is to detect the existence of watermarks. Verification algorithms attempt to distinguish between watermarked images carrying key $\hat{w}$ and unwatermarked images carrying null key $\hat{w}_{\emptyset}$ . More formally, verification can be formulated as distinguishing between the distance to the reference key $w$ : $d(\hat{w}, w) \neq d(\hat{w}_{\emptyset}, w)$ . In Tree-Ring, $d$ is chosen to be $\ell_{1}$ distance and $w$ is the watermark to be imprinted. + +When a set of different keys $w_{i} \in \mathcal{W}$ are distributed, the goal of watermark identification is to find the best match for the recovered key $\hat{w}$ among all candidates in $\mathcal{W}$ , i.e. $\mathrm{ID}(\hat{w}) = \arg \min_{i} [d(\hat{w}, w_{i})]$ . In practice, different users are usually assigned different keys. Thus, each generated image can be source tracked and attributed to a unique user via watermark identification. + +![](images/96908b8f6a6d835c80c90827bb5c6c76a73b7986b046f2384401e79b5775d6b4.jpg) +Fig. 2: Framework of the watermarking process. RingID introduces a series of approaches that can help to imprint a lossless and robust watermark. In contrast, TreeRing injects a lossy and less robust watermark. + +# 4 Delve into Tree-Ring Watermarking + +In this section, we will reveal why Tree-Ring [38] is robust to various image transformations, even some of them are not considered by design. We further evaluate Tree-Ring in the multi-key identification task and pinpoint its vulnerabilities. + +# 4.1 Recap of Tree-Ring + +Tree-Ring [38] originally studies the verification setting. Its goal is to distinguish watermarked images from non-watermarked images. As illustrated in Fig. 2, Tree-Ring transforms the initial diffusion noise $x_{T}$ into the frequency $X_{T}$ and injects a tree-ring pattern $w$ into the center of $X_{T}$ , where the value of each ring in pattern $w$ is sampled from Gaussian. It then transforms the watermarked $X_{T}$ back to the spatial domain, retaining only the real part to feed the diffusion network for image generation. Notably, the accompanying imaginary part is discarded to be compatible with the network input. For detection of the watermark, Tree-Ring inverts the diffusion process to estimate $x_{T}$ from an image, transforms it to the frequency domain, and extracts the watermark pattern $\hat{w}$ . Similarly, nonsense pattern $\hat{w}_{\emptyset}$ can be extracted from the non-watermarked image. Tree-Ring distinguishes between these two $\ell_{1}$ -to-reference distances: $\| \hat{w} - w \|_{1}$ and $\| \hat{w}_{\emptyset} - w \|_{1}$ to determine whether an image is watermarked or not. + +# 4.2 Distribution Shift in Watermarking + +Tree-Ring's detection algorithm bases on the intuition that the recovered pattern $\hat{w}$ from a watermarked image should match the injected reference pattern $w$ + +![](images/3282c7e1061491af0c189db50e1099cac935da4f5b890c5a9f3f983a5adc5bcf.jpg) + +![](images/4833b9f577866c4ada8aabea1bc2c47997ec67079d34f606000be252bde96986.jpg) +Fig. 3: The effect of distribution shift in both verification and identification tasks. Discarding the imaginary part during watermark imprinting causes the distribution shift in $\ell_1$ distance. Distribution shift aids verification but not identification. + +to some extent, leading to a smaller $\ell_1$ -to-reference distance, i.e. $\| \hat{w} - w \|_1 < \| \hat{w}_{\emptyset} - w \|_1$ . + +However, we have discovered that apart from the pattern matching, another crucial factor contributes to the separation of $\| \hat{w} - w \|_1$ v.s. $\| \hat{w}_{\emptyset} - w \|_1$ . This factor is the distribution shift introduced by discarding the imaginary part during the watermark injection process, which is highlighted in Fig. 2. This affects all watermarked images, further reducing the expectation of $\mathbb{E}\left[\| \hat{w} - w \|_1\right]$ . In other words, it shifts the distribution of $\| \hat{w} - w \|_1$ towards a smaller value, making it easier to distinguish between $\| \hat{w} - w \|_1$ and $\| \hat{w}_{\emptyset} - w \|_1$ . + +Mathematically, we prove that the shift factor is $\frac{\sqrt{3}}{2}$ in a general and simplified setting. Let $\hat{w}_n$ denote the recovered watermark that has never been shifted, thus we have the following: + +$$ +\mathbb {E} \left[ \| \hat {w} - w \| _ {1} \right] = \frac {\sqrt {3}}{2} \mathbb {E} \left[ \| \hat {w} _ {n} - w \| _ {1} \right]. \tag {1} +$$ + +Detailed derivation is elaborated in Supplementary A. Actually, we find any operation leading to a sufficient distribution shift can be considered a good standalone watermarking approach. We demonstrate the operation of discarding the imaginary part performs well on its own in Supplementary B. + +Effect in Different Tasks and Attacks Fig. 3 illustrates the effect of distribution shift under different tasks. In Verification, it helps distinguish between watermarked and non-watermarked images. Next, we go one step further and consider the situation when the image undergoes different attacks like JPEG or Resize. If the tree-ring pattern itself can cope with a certain attack, the improvement brought by the distribution shift will be small, because the two distributions are already separable even without the help of shift. On the contrary, if the tree-ring pattern itself cannot deal with a certain attack, the improvement brought by the distribution shift will be significant. Empirically (Supplementary + +A.6), we find distribution shift helps a lot under Rotation and Crop/Scale attacks. This implies that the original design of tree-ring cannot handle these two attacks. When it comes to Identification, things are a little different. Since different watermarks experience the same distribution shift, discarding the imaginary part doesn't help distinguishing them. + +This sparks our curiosity about Tree-Ring's performance in identification, since distribution shift doesn't help and identification is more difficult. We would like to know if Tree-Ring still works under Rotation and Crop/Scale attacks. In the next section, we will explore these questions. + +# 4.3 Is Tree-Ring Good at Identification? + +We report the performance of Tree-Ring when identifying different numbers of keys in Tab. 2. Tree-Ring struggles to identify as few as 32 keys. What's more, identification accuracy is nearly zero under Rotation and Crop/Scale attacks. + +We summarize 2 reasons as follows: (1) Distribution shift doesn't help. As tree-ring watermark cannot handle Rotation and Crop/Scale attacks, identification accuracy under these attacks drops to nearly zero. (2) Tee-ring pattern doesn't have enough distinguishing ability. Since identification is more difficult than verification, poor distinguishing ability results in poor performance under most attacks or even under no attack setting. + +# 4.4 Why is Tree-Ring Vulnerable to Rotation and Crop/Scale? + +All previous experimental results and analysis show one thing, tree-ring pattern cannot handle these two attacks. In this section, we will see what happens. + +Vulnerable to Crop & Scale Scaling an image by a factor of $a$ corresponds to compressing the frequency spectrum $X[u,v]$ in both magnitudes and positions: $\frac{1}{a^2} X\left[\frac{u}{a},\frac{v}{a}\right]$ . Cropping an image modulates the frequency magnitudes by a sinc function. It is straightforward that these transformations easily disrupt the tree-ring pattern on frequency domain. As shown in Fig. 4(e), the pattern is completely disrupted. + +Vulnerable to Rotation Although Tree-Ring is designed as ring patterns to seek rotation robustness, we find the existing design ignores many details, leading to no rotation robustness. In short, a combination of corner cropping when rotating, a lossy injection process, and the rotationally asymmetric shape cause the problem. We will explain in detail and propose solutions in Sec. 5.2. + +# 5 RingID: Enhanced Watermarking for Identification + +In this section, we show how our proposed method RingID advances in identifying multiple distinct keys. RingID originates from Tree-Ring but incorporates systematic enhancements. We introduce a multi-channel heterogeneous watermarking framework, a discretization approach and a method of lossless imprinting to generally improve its distinguishing ability. We also discuss a series of designs regarding stronger rotation robustness and increased capacity. + +# 5.1 Multi-Channel Heterogeneous Watermarking + +Multi-channel heterogeneous (MCH) watermarking serves as a general framework to comprehensively improve watermarking robustness to a battery of attacks. Intuitively, we imprint different types of watermarks on different channels of the initial noise to amalgamate their distinctive advantages, as illustrated in Fig. 1. During matching, we calculate the minimum of channel-wise $\ell_1$ distance to the reference watermark for all watermarked channels: + +$$ +\operatorname {I D} (\hat {w}) = \arg \min _ {i} \left\{\min _ {c \in C _ {w}} \left[ \lambda_ {c} \| \hat {w} ^ {c} - w _ {i} ^ {c} \| _ {1} \right] \right\} \tag {2} +$$ + +where $c$ denotes the channel index and traverses all watermarked channels $C_w$ . $i$ denotes the index of candidate reference watermarks. $\lambda_c$ is a channel-wise normalizing factor. + +Eq. (2) bases on the intuition that if a watermark is robust to a certain attack, then its distance from the ground-truth reference is smaller than that of other watermarks that are not robust. Ideally, it enables adaptive selection of the most robust watermark under different attacks. Notably, the total capacity of MCH is the minimum of each watermark. Therefore, it's recommended to use a watermark pattern with sufficient capacity on each channel. At the same time, in order to reduce the impact on the generation quality, we need to carefully select the type of watermarks for combination. We find the Gaussian noise pattern is a perfect candidate. It has infinite capacity and the same distribution as the initial noise. Meanwhile, it has good robustness to non-geometric attacks [38] and is complementary to the tree-ring pattern. So RingID explores the combination of a Gaussian noise watermark and a tree-ring watermark. Empirical study in Sec. 6.3 demonstrates that this combination is able to perfectly amalgamate unique advantages from the noise and ring watermark. + +# 5.2 Being Invariant to Rotation + +Although the tree-ring pattern is designed to be rotationally symmetric, it exhibits unexpected vulnerability to rotation as shown in Tab. 2. Below, we identify the causes of the issue and propose a series of measures to address it. + +Spatial Shift to Avoid Corner Cropping We visualize the imprinted tree-ring watermark in spatial domain in Fig. 4(c). Due to the properties of FFT, tree-ring patterns in the frequency tends to exhibit concentrated energy in the four corners in the spatial domain, rendering them vulnerable to cropping during rotation. We propose to circularly shift the watermark pattern in the spatial domain to the center of the image, thus the majority of watermark information is kept while rotating. For an $N \times N$ latent image, we circularly shift the pattern by $\frac{N}{2}$ pixels in both width and height dimension. This is equivalent to directly multiplying the frequency watermark with a chessboard pattern given by $H[u,v] = e^{j\frac{\pi}{2}(u + v)} = (-1)^{u + v}$ . The change of the watermark can be clearly + +![](images/0818580d2bedb26f0e04968fab624344aeee8267d1b8b36a501bcd2be207d2a8.jpg) +(a) Discretization + +![](images/01c7fb2d70c7e839178ac58ed1b6bbd852b1046a240bbdfe20975ad884cb1067.jpg) + +![](images/8b5d329ded6bcd03c6f203559fadbd4f6df61a8db74be7ab7544c3ee75e6fc9f.jpg) +(b) Rounder Ring + +![](images/d6e9a183641f6ce5934fc7c18aaf07b7d9e5cc80704d272ab81ca4e00eeed782.jpg) + +![](images/fb9afe1dece5f3b4fedf427f6ea6e1397b4f51b21af015a3908ff649aae20d43.jpg) + +![](images/f22f1968a17b015f0dfbb5599709215f959d29fcee5f9283ed9499e1a5bbb261.jpg) +(d) Lossless Imprinting + +![](images/65252de08df1416b6425d5274a3c928ce6c4aa3ca64b9ae457a77b03f03e1476.jpg) +(e) Crop & Scale +Fig. 4: Detailed illustration of different acts and their impacts. (a) Discretization: Discretize values to $\pm \alpha$ for each ring, increasing the distinguishing ability of the watermark. (b) Rounder ring: Render smoother edges and reduce artifacts. Increase rotational robustness. (c) Spatial Shift: Shift energy on the corners to the center, avoiding potential cropping during rotation. (d) Lossless Imprinting: Ensure the imprinted pattern the same as the original design. (e) Crop & Scale: Zooming in the spatial domain compresses the imprinted pattern in the frequency domain, causing pattern mismatch and hinders identification. + +observed in Fig. 4(c). Although spatial-shifting helps a lot for rotation invariance, we noticed that this often leads to the generation of circular artifacts at the center of the image. Keeping generation quality in mind, we further multiply the shifted pattern by a factor $\eta$ to suppress the peak at the center. $\eta$ trades-off the robustness, as smaller $\eta$ weakens the watermark pattern. In practice, we find $\eta \in [0.8,0.9]$ strikes a good balance between image quality and robustness. + +Lossless Imprinting The operation of discarding the imaginary part during watermarking not only shifts the distribution of $\ell_1$ -to-reference, but also disrupt the pattern, resulting in the loss of its rotational symmetry. Fig. 4(d) visualizes the actually carried pattern. More concretely, recall from Fig. 2, when the watermarked noise $X[u,v]$ is transformed back to the spatial $x[m,n]$ , only its real part $x_{\mathrm{re}}[m,n]$ is kept for image generation. From a frequency domain perspective, this is equivalent to generation with its conjugate symmetry $X_{\mathrm{cs}}[u,v]$ , which is not necessarily equal to $X[u,v]$ . To leverage the rotational symmetry of the + +ring patterns, we attempt to eliminate the information loss during imprinting by solving $X[u,v] = X_{\mathrm{cs}}[u,v]$ . The results indicate that $X_{\mathrm{re}}[u,v]$ should be an even function and $X_{\mathrm{im}}[u,v]$ should be an odd function about the Fourier center. To satisfy these constraints, we imprint the tree-ring pattern only in the real part and leave the imaginary part blank. We also align the center of the tree-ring pattern with the Fourier center (32,32). Originally, the tree-ring pattern was put at (31,32) and imprinted on both the real and the imaginary. + +Making A Rounder Ring Rings are originally determined by locating pixels equidistant from a give circle center $(x_0, y_0)$ : $(x - x_0)^2 + (y - y_0)^2 = r^2$ . We observed that this simple approach does not yield a well-rounded circular ring, particularly in images with lower resolutions like initial noise of shape $64 \times 64$ . Artifacts and alias are shown in Fig. 4(b). Although it is centrosymmetric, it lacks rotational symmetry. Well, the solution is quite straightforward: drawing a rounder ring. There are a number of tools that can help draw a much rounder ring. Here, we employ a very simple idea to get a rounder ring. We place a white pixel on a black background at a distance $r$ from the rotation center. By rotating the low resolution image 360 degrees and recording the trajectory of this pixel, we obtain a rounder ring with a radius of $r$ . Despite its simplicity, empirical results in Tab. 3 show that it is crucial for rotational invariance. + +# 5.3 Discretization for Enhanced Distinguishability + +Tree-Ring samples values for each ring from a Gaussian distribution to keep consistency with the distribution of initial noise. We find that this random sampling makes it exceedingly challenging to distinguish between two keys. To address this issue, we propose a discretization approach, specifying that the value for each ring can only be either $\alpha$ or $-\alpha$ , illustrated in Fig. 4(a). Consequently, the watermark pattern for $n$ rings has a capacity of $2^n$ . While this seemingly reduces the upper limit of watermark capacity, it significantly enhances its valid capacity, ensuring optimal utilization of all available slots. Our experiments reveal that setting $\alpha$ as the standard deviation of the initial noise ensures substantial distinguishability between different keys while reducing the impact on generation quality to a small level. + +# 5.4 Increasing Capacity + +We consider two simple approaches to increase capacity: adjusting the number of rings in a single channel or imprinting rings in extra channels. We evaluate the advantages and drawbacks of both methods in Sec. 6.4. In summary, increasing the number of rings $|R|$ in a single channel can effectively boost capacity, leading to $2^{|R|}$ keys. However, excessive ring count results in a noticeable decline in robustness and generation quality. Imprinting watermarks in multiple channels exponentially increases capacity by $n^{|c|}$ , where $c$ is number of channels and $n$ is + +the capacity of a single channel. However, it's likely to introduce ring-like artifacts in the center of generated images. Therefore, we opt to place the watermark in a single channel and enhance capacity by adjusting the range of rings. + +# 6 Experiment + +# 6.1 Experimental Setup + +Following Tree-Ring [38], we experiment with the state-of-the-art opensource diffusion model StableDiffusion-V2 [29]. By default, RingID imprints a tree-ring watermark with radius 3-14 on the channel 3 of the initial noise and a Gaussian noise watermark on channels 0. We adopt discretization (to 64 and -64), spatial shift, rounder drawing, and lossless imprinting as ordered in Fig. 2. We report the ROC-AUC between 1000 watermarked and 1000 unwatermarked images as the verification metric. In identification, we report identification accuracy with different numbers of keys. Evaluation is performed under various image distortions, following the original setup of Tree-Ring [38], i.e., JPEG 25, rotation $75^{\circ}$ , $75\%$ random C&S, blur kernel 8, noise std 0.1, brightness [0, 6]. We assess the generation quality with CLIP scores [28] by OpenCLIP-ViT/G [5] between the generated image and the prompts to measure text-image alignment, and Frechet Inception Distance (FID) [17] to quantify the similarity between generated and natural images. We generate 5000 images and evaluate FID on the COCO 2017 training dataset [21]. We set both generation and inversion steps to 50. + +# 6.2 Comparison with Baselines + +Robustness In verification (Tab. 1), both RingID and Tree-Ring perfectly distinguish between watermarked and unwatermarked images when no attacks are performed, achieving AUC score of 1. However, RingID exhibit stronger robustness to various image distortions than Tree-Ring, achieving near-perfect average AUC score, even though it is not optimized for verification. In the more challenging identification task, (Tab. 2), RingID significantly outperforms Tree-Ring. When identifying 32 keys, Tree-Ring's average identification accuracy (excluding C&S) is 0.465, while RingID's accuracy reaches up to 0.992. As the number of keys increases to 2048, this gap widens drastically (0.077 v.s. 0.942), further highlighting RingID's superior robustness and larger valid key capacity. + +Notably, both methods struggle with the C&S distortion in identification (Tab. 2). This aligns with our analysis in Sec. 4.2. Scaling in the spatial domain causes inverse scaling in the frequency domain, leading to pattern disruption and rendering identification challenging under such distortions. However, in verification, with the help of distribution shift, watermarked images can still be effectively distinguished from unwatermarked ones under C&S. Both Tree-Ring and RingID achieve average AUC higher than 0.97. + +Image Quality RingID achieves strong performance without compromising the generation quality. It maintains comparable text-image alignments with TreeRing. Both methods achieve similar CLIP scores (0.364 v.s. 0.365). RingID obtains an FID of 26.13, which is also comparable to Tree-Ring's FID of 25.93. + +Table 1: Comparison with Tree-Ring in the verification task. The table shows the ROC-AUC values under various image distortions and CLIP Scores. + +
MethodsRing RadiusCleanRotateJPEGC&SBlurNoiseBrightnessAvgCLIP Score
Tree-Ring0-101.0000.9350.9990.9610.9990.9440.9830.9750.364
RingID0-101.0001.0001.0000.9790.9940.9690.9910.9900.359
3-141.0001.0001.0000.9870.9890.9980.9940.9950.365
+ +Table 2: Comparison with Tree-Ring in the identification task. We report the identification accuracy under various distortions for different numbers of keys. + +
Methods#KeysCleanRotateJPEGC&SBlurNoiseBrightnessAvg\(Avg_{noC&S}\)
Tree-Ring320.7900.0200.4200.0400.6100.5300.4200.4040.465
1280.4500.0100.1200.0200.2800.2300.1700.1830.210
20480.2000.0000.0400.0000.0900.0700.0600.0660.077
RingID321.0001.0001.0000.5300.9901.0000.9600.9260.992
1281.0000.9801.0000.2800.9801.0000.9400.8830.983
20481.0000.8601.0000.0800.9700.9500.8700.8190.942
+ +# 6.3 Ablation Study + +Components of RingID Tab. 3 ablates all proposed components of RingID. We can find that all designs are crucial for the final success of RingID. Spatial shift, lossless imprinting, and discretization play a significant role in the robustness of rotation. While the use of rounder rings also increase the accuracy by approximately $40\%$ (from 0.620 to 0.860). Notably, spatial shift sacrifices the robustness to blur and brightness attacks for rotation robustness. + +Heterogeneous Watermarking RingID employs the combination of a Gaussian noise watermark and a ring watermark. Here, we ablate the choice of imprinting the noise watermark on different channels. By comparing the first 3 rows of Tab. 4, we find that heterogeneous watermarking can effectively combine the advantage from both the noise watermark and the ring watermark. Among single-channel configurations, imprinting noise watermark on channel 0 (0.819) achieves the best performance. The combination of channels 0 and 1 yields the highest accuracy among all trials. Interestingly, imprinting noise watermarks on all available channels (0, 1, 2) decrease the accuracy from 0.830 to 0.807. + +# 6.4 Discussion + +Where is the Best Place to Put Rings? We could imprint all rings on one channel or distribute them evenly across multiple channels. In Tab. 5, we fix the total number of rings to 12 and distribute them to 1-4 channels. we can find that performance drops as we distribute rings to more channels. What's more, imprinting rings on multiple channels cause artifacts to appear more frequently (shown in Supplementary C). So given fixed capacity, it's better to imprint all rings on a single channel. This raises another question: what are the optimal + +Table 3: Ablation study on the effect of each proposed module in RingID. We report the identification accuracy for 2048 keys when removing individual components. + +
SettingsCleanRotateJPEGC&SBlurNoiseBrightnessAvg
RingID1.0000.8601.0000.0800.9700.9500.8700.819
- Spatial Shift1.0000.0001.0000.0500.9900.9300.9400.701
- Lossless Imprinting1.0000.0100.9700.1600.9500.9800.8300.700
- Rounder Ring1.0000.6200.9900.1000.8900.9700.8500.774
- Discretization0.9800.1200.3800.0200.4500.6500.3900.427
- Heterogeneous1.0000.8200.9400.0300.9600.7100.7200.740
+ +Table 4: Ablation study on multi-channel heterogeneous watermarking. This table reports the identification accuracy when the noise watermark is imprinted on different channels (abbr. CHs). “-” means the watermark is not imprinted. + +
Ring CHsNoise CHsCleanRotateJPEGC&SBlurNoiseBrightnessAvg
-01.0000.0001.0000.1700.9900.9500.9200.719
3-1.0000.8200.9400.0300.9600.7100.7200.740
301.0000.8601.0000.0800.9700.9500.8700.819
311.0000.8201.0000.1301.0000.8800.7700.800
321.0000.8400.9800.1000.9200.9900.7900.803
30,11.0000.8500.9500.1500.9900.9900.8800.830
30,21.0000.8400.9900.1200.9500.9800.8700.821
30,1,21.0000.8300.9600.1600.9300.9800.7900.807
+ +inner and outer radii if all rings are placed on a single channel? From Tab. 6 we can see that 3-14 forms a good choice for 11 rings with capacity 2048. The optimal ring radius also varies for different ring numbers. + +More Keys We would like to explore whether RingID can handle a larger number of keys. Fig. 5 shows the average identification accuracy under various attacks $v.s.$ number of keys. We observe that RingID maintains good accuracy even with a large number of keys. And identification accuracy linearly decreases as the number of keys exponentially increase. + +Table 6: Influence of ring radius on identification accuracy and text-image alignment. + +
Radius#KeysAccCLIP ScoreRadius#KeysAccCLIP ScoreRadius#KeysAccCLIP Score
1-6320.8860.3631-92560.8190.3621-1220480.7860.361
2-7320.8760.3662-102560.8560.3582-1320480.8040.361
3-8320.8970.3653-112560.8500.3623-1420480.8190.368
4-9320.8970.3684-122560.8470.3634-1520480.8130.363
5-10320.8810.3615-132560.8390.3605-1620480.8030.360
+ +Table 5: Distributing 12 rings across different number of channels (#CHs). This table reports the average identification accuracy and CLIP Score. + +
Ring Radius #CHsAvgCLIP Score
3-1510.7400.357
3-920.7370.351
3-730.6640.351
3-640.5440.340
+ +![](images/dc91f09a8e3332f8fef9acaaed2e773f7422009913d4fd33282cd11ca1654349.jpg) +Fig. 5: Average identification accuracy $v.s.$ number of keys. + +![](images/52bb288e0054b399f62eb9e2ebab2ee3e14701a8ddaa8ee532e2a35388fdd284.jpg) +Fig. 6: Qualitative results. The subtle changes in image layouts show that both TreeRing and RingID embed watermarks into the image's semantic content. + +# 6.5 Qualitative Results + +As shown in Fig. 6, both RingID and Tree-Ring watermarking can generate images of good visual quality. Unlike low-level watermarking techniques, both RingID and Tree-Ring cause subtle changes in the image layout. However, RingID generates images with layout that is closer to the unwatermarked image. We owe this to avoiding watermarking on the low-frequency part (radius 0-3). + +# 7 Conclusion and Limitation + +We conduct a comprehensive investigation on Tree-Ring Watermarking, revealing its limitations in identifying multiple keys. We then propose RingID to enhance identification capability. + +Limitation: Under multi-key identification, one limitation is that both TreeRing and our RingID are vulnerable to cropping and scaling attacks. Future work may explore other transform domains to resist these attacks. + +Acknowledgment: This project is supported by the Ministry of Education, Singapore, under the Academic Research Fund Tier 1 (FY2022) Award 22-5406-A0001. + +# References + +1. An, B., Ding, M., Rabbani, T., Agrawal, A., Xu, Y., Deng, C., Zhu, S., Mohamed, A., Wen, Y., Goldstein, T., et al.: Benchmarking the robustness of image watermarks. arXiv preprint arXiv:2401.08573 (2024) +2. Bansal, A., Chiang, P.y., Curry, M.J., Jain, R., Wigington, C., Manjunatha, V., Dickerson, J.P., Goldstein, T.: Certified neural network watermarks with randomized smoothing. In: International Conference on Machine Learning. pp. 1450-1465. PMLR (2022) +3. Betker, J., Goh, G., Jing, L., Brooks, T., Wang, J., Li, L., Ouyang, L., Zhuang, J., Lee, J., Guo, Y., et al.: Improving image generation with better captions. Computer Science. https://cdn.opencai.com/papers/dall-e-3.pdf 2(3), 8 (2023) +4. Blattmann, A., Dockhorn, T., Kulal, S., Mendelevitch, D., Kilian, M., Lorenz, D., Levi, Y., English, Z., Voleti, V., Letts, A., et al.: Stable video diffusion: Scaling latent video diffusion models to large datasets. arXiv preprint arXiv:2311.15127 (2023) +5. Cherti, M., Beaumont, R., Wightman, R., Wortsman, M., Ilharco, G., Gordon, C., Schuhmann, C., Schmidt, L., Jitsev, J.: Reproducible scaling laws for contrastive language-image learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2818-2829 (2023) +6. Ci, H., Song, Y., Yang, P., Xie, J., Shou, M.Z.: Wmadapter: Adding watermark control to latent diffusion models. arXiv preprint arXiv:2406.08337 (2024) +7. Ci, H., Wang, C., Ma, X., Wang, Y.: Optimizing network structure for 3d human pose estimation. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 2262-2271 (2019) +8. Ci, H., Wu, M., Zhu, W., Ma, X., Dong, H., Zhong, F., Wang, Y.: Gfpose: Learning 3d human pose prior with gradient fields. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4800-4810 (2023) +9. Cox, I.: Digital watermarking and steganography. Morgan Kaufmann google scholar 2, 893-914 (2007) +0. Cui, S., Guo, X., Zhang, Z.: Estimation and inference in ultrahigh dimensional partially linear single-index models. arXiv preprint arXiv:2404.04471 (2024) +1. Cui, S., Sudjianto, A., Zhang, A., Li, R.: Enhancing robustness of gradientboosted decision trees through one-hot encoding and regularization. arXiv preprint arXiv:2304.13761 (2023) +2. Dhariwal, P., Nichol, A.: Diffusion models beat gans on image synthesis. Advances in neural information processing systems 34, 8780-8794 (2021) +3. Fernandez, P., Couairon, G., Jégou, H., Douze, M., Furon, T.: The stable signature: Rooting watermarks in latent diffusion models. arXiv preprint arXiv:2303.15435 (2023) +4. Fernandez, P., Sablayrolles, A., Furon, T., Jégou, H., Douze, M.: Watermarking images in self-supervised latent spaces. In: ICASSP 2022-2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). pp. 3054-3058. IEEE (2022) + +15. Guo, H., Zhang, Q., Luo, J., Guo, F., Zhang, W., Su, X., Li, M.: Practical deep dispersed watermarking with synchronization and fusion. In: Proceedings of the 31st ACM International Conference on Multimedia. pp. 7922-7932 (2023) +16. Guo, Y., Yang, C., Rao, A., Liang, Z., Wang, Y., Qiao, Y., Agrawala, M., Lin, D., Dai, B.: Animatediff:Animate your personalized text-to-image diffusion models without specific tuning. International Conference on Learning Representations (2024) +17. Heusel, M., Ramsauer, H., Unterthiner, T., Nessler, B., Hochreiter, S.: Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems 30 (2017) +18. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. Advances in neural information processing systems 33, 6840-6851 (2020) +19. Kong, Z., Ping, W., Huang, J., Zhao, K., Catanzaro, B.: Diffwave: A versatile diffusion model for audio synthesis. arXiv preprint arXiv:2009.09761 (2020) +20. Lin, C.Y., Wu, M., Bloom, J.A., Cox, I.J., Miller, M.L., Lui, Y.M.: Rotation, scale, and translation resilient watermarking for images. IEEE Transactions on image processing 10(5), 767-782 (2001) +21. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollar, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13. pp. 740-755. Springer (2014) +22. Liu, R., Wu, R., Van Hoorick, B., Tokmakov, P., Zakharov, S., Vondrick, C.: Zero-1-to-3: Zero-shot one image to 3d object. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 9298-9309 (2023) +23. Liu, Y., Li, Z., Backes, M., Shen, Y., Zhang, Y.: Watermarking diffusion model. arXiv preprint arXiv:2305.12502 (2023) +24. Luo, X., Li, Y., Chang, H., Liu, C., Milanfar, P., Yang, F.: Dvmark: a deep multiscale framework for video watermarking. IEEE Transactions on Image Processing (2023) +25. Midjourney.com: Midjourney. https://www.midjourney.com/home +26. OpenAI: Dalle2. https://openai.com/dall-e-2 +27. Poole, B., Jain, A., Barron, J.T., Mildenhall, B.: Dreamfusion: Text-to-3d using 2d diffusion. arXiv preprint arXiv:2209.14988 (2022) +28. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021) +29. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 10684-10695 (June 2022) +30. Singer, U., Polyak, A., Hayes, T., Yin, X., An, J., Zhang, S., Hu, Q., Yang, H., Ashual, O., Gafni, O., et al.: Make-a-video: Text-to-video generation without text-video data. arXiv preprint arXiv:2209.14792 (2022) +31. Song, J., Meng, C., Ermon, S.: Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502 (2020) +32. Song, Y., Sohl-Dickstein, J., Kingma, D.P., Kumar, A., Ermon, S., Poole, B.: Score-based generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456 (2020) + +33. Song, Y., Huang, S., Yao, C., Ye, X., Ci, H., Liu, J., Zhang, Y., Shou, M.Z.: Processpainter: Learn painting process from sequence data. arXiv preprint arXiv:2406.06062 (2024) +34. Steinberg, Y., Merhav, N.: Identification in the presence of side information with application to watermarking. IEEE Transactions on Information Theory 47(4), 1410-1422 (2001) +35. Tancik, M., Mildenhall, B., Ng, R.: Stegastamp: Invisible hyperlinks in physical photographs. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 2117-2126 (2020) +36. Tevet, G., Raab, S., Gordon, B., Shafir, Y., Cohen-Or, D., Bermano, A.H.: Human motion diffusion model. arXiv preprint arXiv:2209.14916 (2022) +37. Uchida, Y., Nagai, Y., Sakazawa, S., Satoh, S.: Embedding watermarks into deep neural networks. In: Proceedings of the 2017 ACM on international conference on multimedia retrieval. pp. 269-277 (2017) +38. Wen, Y., Kirchenbauer, J., Geiping, J., Goldstein, T.: Tree-ring watermarks: Fingerprints for diffusion images that are invisible and robust. arXiv preprint arXiv:2305.20030 (2023) +39. Wu, J.Z., Ge, Y., Wang, X., Lei, S.W., Gu, Y., Shi, Y., Hsu, W., Shan, Y., Qie, X., Shou, M.Z.: Tune-a-video: One-shot tuning of image diffusion models for text-to-video generation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 7623-7633 (2023) +40. Xia, X.G., Boncelet, C.G., Arce, G.R.: Wavelet transform based watermark for digital images. Optics Express 3(12), 497-511 (1998) +41. Xiong, C., Qin, C., Feng, G., Zhang, X.: Flexible and secure watermarking for latent diffusion model. In: Proceedings of the 31st ACM International Conference on Multimedia. pp. 1668-1676 (2023) +42. Yang, P., Ci, H., Song, Y., Shou, M.Z.: Steganalysis on digital watermarking: Is your defense truly impervious? arXiv preprint arXiv:2406.09026 (2024) +43. Yu, N., Skripniuk, V., Abdelnabi, S., Fritz, M.: Artificial fingerprinting for generative models: Rooting deepfake attribution in training data. In: Proceedings of the IEEE/CVF International conference on computer vision. pp. 14448-14457 (2021) +44. Zhang, D.J., Wu, J.Z., Liu, J.W., Zhao, R., Ran, L., Gu, Y., Gao, D., Shou, M.Z.: Show-1: Marrying pixel and latent diffusion models for text-to-video generation. arXiv preprint arXiv:2309.15818 (2023) +45. Zhang, J., Gu, Z., Jang, J., Wu, H., Stoecklin, M.P., Huang, H., Molloy, I.: Protecting intellectual property of deep neural networks with watermarking. In: Proceedings of the 2018 on Asia conference on computer and communications security. pp. 159-172 (2018) +46. Zhang, K.A., Xu, L., Cuesta-Infante, A., Veeramachaneni, K.: Robust invisible video watermarking with attention. arXiv preprint arXiv:1909.01285 (2019) +47. Zhao, X., Zhang, K., Wang, Y.X., Li, L.: Generative autoencoders as watermark attackers: Analyses of vulnerabilities and threats. arXiv preprint arXiv:2306.01953 (2023) +48. Zhao, Y., Pang, T., Du, C., Yang, X., Cheung, N.M., Lin, M.: A recipe for watermarking diffusion models. arXiv preprint arXiv:2303.10137 (2023) +49. Zhu, J., Kaplan, R., Johnson, J., Fei-Fei, L.: Hidden: Hiding data with deep networks. In: Proceedings of the European conference on computer vision (ECCV). pp. 657-672 (2018) +50. Zhu, W., Ma, X., Ro, D., Ci, H., Zhang, J., Shi, J., Gao, F., Tian, Q., Wang, Y.: Human motion generation: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence (2023) \ No newline at end of file diff --git a/2024/Rethinking Tree-Ring Watermarking for Enhanced Multi-Key Identification/images.zip b/2024/Rethinking Tree-Ring Watermarking for Enhanced Multi-Key Identification/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..058e070688495bc3e2d1c5377389581febf21f52 --- /dev/null +++ b/2024/Rethinking Tree-Ring Watermarking for Enhanced Multi-Key Identification/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:24d8c923c28789ef0ffd9f95e61fc019202ec304ea25a5f9b2578495b4938f34 +size 512788 diff --git a/2024/Rethinking Tree-Ring Watermarking for Enhanced Multi-Key Identification/layout.json b/2024/Rethinking Tree-Ring Watermarking for Enhanced Multi-Key Identification/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..ddb5924c92be9bd7425e860cb52a68c6b67d3b45 --- /dev/null +++ b/2024/Rethinking Tree-Ring Watermarking for Enhanced Multi-Key Identification/layout.json @@ -0,0 +1,9601 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 136, + 111, + 477, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 111, + 477, + 148 + ], + "spans": [ + { + "bbox": [ + 136, + 111, + 477, + 148 + ], + "type": "text", + "content": "RingID: Rethinking Tree-Ring Watermarking for Enhanced Multi-Key Identification" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 165, + 168, + 447, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 165, + 168, + 447, + 180 + ], + "spans": [ + { + "bbox": [ + 165, + 168, + 447, + 180 + ], + "type": "text", + "content": "Hai " + }, + { + "bbox": [ + 165, + 168, + 447, + 180 + ], + "type": "inline_equation", + "content": "\\mathrm{Ci}^{*}" + }, + { + "bbox": [ + 165, + 168, + 447, + 180 + ], + "type": "text", + "content": " Pei Yang\\*, Yiren Song\\*, and Mike Zheng Shou" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 202, + 190, + 411, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 202, + 190, + 411, + 224 + ], + "spans": [ + { + "bbox": [ + 202, + 190, + 411, + 224 + ], + "type": "text", + "content": "Show Lab, National University of Singapore cihai03@gmail.com yangpei@u.nus.edu yiren@nus.edu.sg mike.zheng.shou@gmail.com" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 160, + 255, + 455, + 410 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 255, + 455, + 410 + ], + "spans": [ + { + "bbox": [ + 160, + 255, + 455, + 410 + ], + "type": "text", + "content": "Abstract. We revisit Tree-Ring Watermarking, a recent diffusion model watermarking method that demonstrates great robustness to various attacks. We conduct an in-depth study on it and reveal that the distribution shift unintentionally introduced by the watermarking process, apart from watermark pattern matching, contributes to its exceptional robustness. Our investigation further exposes inherent flaws in its original design, particularly in its ability to identify multiple distinct keys, where distribution shift offers no assistance. Based on these findings and analysis, we present RingID for enhanced multi-key identification. It consists of a novel multi-channel heterogeneous watermarking approach designed to seamlessly amalgamate distinctive advantages from diverse watermarks. Coupled with a series of suggested enhancements, RingID exhibits substantial advancements in multi-key identification. Github page: https://github.com/showlab/RingID" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 160, + 419, + 453, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 419, + 453, + 441 + ], + "spans": [ + { + "bbox": [ + 160, + 419, + 453, + 441 + ], + "type": "text", + "content": "Keywords: Diffusion Models " + }, + { + "bbox": [ + 160, + 419, + 453, + 441 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 419, + 453, + 441 + ], + "type": "text", + "content": " Tree-Ring Watermarking " + }, + { + "bbox": [ + 160, + 419, + 453, + 441 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 419, + 453, + 441 + ], + "type": "text", + "content": " Multi-Key Identification" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 464, + 230, + 475 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 464, + 230, + 475 + ], + "spans": [ + { + "bbox": [ + 132, + 464, + 230, + 475 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 489, + 481, + 537 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 489, + 481, + 537 + ], + "spans": [ + { + "bbox": [ + 130, + 489, + 481, + 537 + ], + "type": "text", + "content": "With the popularity of diffusion models, a vast number of high-quality diffusion generated images are circulating on the internet. It has become increasingly important to identify AI-generated images to detect fabrication and to track and protect copyrights. Luckily, watermarking [1, 13] offers a reliable solution." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 538, + 482, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 538, + 482, + 647 + ], + "spans": [ + { + "bbox": [ + 130, + 538, + 482, + 647 + ], + "type": "text", + "content": "Research on watermarking has a long history in computer vision [9, 34]. A series of works [13, 14, 35, 48, 49] imprint imperceptible watermarks on images through small pixel perturbations. Recently, [47] demonstrates that pixel-level perturbations are provably removable by regeneration attacks. In contrast, Tree-Ring Watermarking [38] has recently introduced a novel approach, proposing to imprint a specially designed tree-ring pattern into the initial noise of a diffusion model. Different patterns are also called different \"keys\". The verification of the watermark's existence involves recovering the initial diffusion noise from the image and comparing the extracted key with the injected one. Tree-Ring" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 652, + 329, + 666 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 652, + 329, + 666 + ], + "spans": [ + { + "bbox": [ + 133, + 652, + 329, + 666 + ], + "type": "text", + "content": "* Equal Contribution. " + }, + { + "bbox": [ + 133, + 652, + 329, + 666 + ], + "type": "inline_equation", + "content": "{}^{\\boxtimes }" + }, + { + "bbox": [ + 133, + 652, + 329, + 666 + ], + "type": "text", + "content": " Corresponding Author." + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 134, + 129, + 225, + 201 + ], + "blocks": [ + { + "bbox": [ + 134, + 129, + 225, + 201 + ], + "lines": [ + { + "bbox": [ + 134, + 129, + 225, + 201 + ], + "spans": [ + { + "bbox": [ + 134, + 129, + 225, + 201 + ], + "type": "image", + "image_path": "31852c4dd9f7b48c201208e99c5e85b08affb1a9b1424b4e40c6061a03e13fd8.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 156, + 207, + 200, + 219 + ], + "lines": [ + { + "bbox": [ + 156, + 207, + 200, + 219 + ], + "spans": [ + { + "bbox": [ + 156, + 207, + 200, + 219 + ], + "type": "text", + "content": "Tree-Ring" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 235, + 130, + 324, + 200 + ], + "blocks": [ + { + "bbox": [ + 235, + 130, + 324, + 200 + ], + "lines": [ + { + "bbox": [ + 235, + 130, + 324, + 200 + ], + "spans": [ + { + "bbox": [ + 235, + 130, + 324, + 200 + ], + "type": "image", + "image_path": "88e1bbd7d21261cb169f4e01d9250ef3323af44ea05d1127d6b6e3c9c1d40587.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 261, + 207, + 293, + 219 + ], + "lines": [ + { + "bbox": [ + 261, + 207, + 293, + 219 + ], + "spans": [ + { + "bbox": [ + 261, + 207, + 293, + 219 + ], + "type": "text", + "content": "RingID" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 337, + 116, + 482, + 205 + ], + "blocks": [ + { + "bbox": [ + 337, + 116, + 482, + 205 + ], + "lines": [ + { + "bbox": [ + 337, + 116, + 482, + 205 + ], + "spans": [ + { + "bbox": [ + 337, + 116, + 482, + 205 + ], + "type": "image", + "image_path": "4d13a9715ed9f8335ef0e6a85f9504bc171571aef2017eab31f1e22196ec6e11.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 389, + 209, + 456, + 218 + ], + "lines": [ + { + "bbox": [ + 389, + 209, + 456, + 218 + ], + "spans": [ + { + "bbox": [ + 389, + 209, + 456, + 218 + ], + "type": "text", + "content": "Number of Keys" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 131, + 230, + 482, + 265 + ], + "lines": [ + { + "bbox": [ + 131, + 230, + 482, + 265 + ], + "spans": [ + { + "bbox": [ + 131, + 230, + 482, + 265 + ], + "type": "text", + "content": "Fig. 1: Tree-Ring imprints a broken watermark to a single channel. Our method RingID achieves stronger robustness by imprinting intact watermarks to multiple channels. In the task of multi-key identification, RingID shows overwhelming advantages." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 290, + 482, + 410 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 290, + 482, + 410 + ], + "spans": [ + { + "bbox": [ + 130, + 290, + 482, + 410 + ], + "type": "text", + "content": "is considered a semantic watermark since the watermark changes the layout of the generated image and semantically hidden in it. It has demonstrated strong robustness against regeneration attacks [47] and various other image transformations [1, 38], credited to the meticulously designed tree-ring pattern. However, previous research only studies its performance in the verification task, which aims at distinguishing between watermarked and unwatermarked images. It's still unclear if Tree-Ring can be used to distribute multiple distinct keys and distinguish between them, a.k.a. watermark identification, which is crucial for source tracing and user attribution. Will Tree-Ring still be strong and robust as in the verification task? In this work, we will explore and answer this question." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 411, + 482, + 507 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 411, + 482, + 507 + ], + "spans": [ + { + "bbox": [ + 130, + 411, + 482, + 507 + ], + "type": "text", + "content": "We start by studying the source of Tree-Ring's robustness under verification setting. For the first time, we uncover that the robustness of Tree-Ring stems not only from the carefully designed tree-ring pattern but also from the unintentionally introduced distribution shift during watermark imprinting. Such shift makes a big difference, especially in its robustness to Rotation and Crop/Scale image transformations, which cannot be handled by tree-ring's pattern design. More importantly, this overlooked power does not help in the identification scenario, because different watermarks experience the same shift." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 509, + 482, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 509, + 482, + 639 + ], + "spans": [ + { + "bbox": [ + 130, + 509, + 482, + 639 + ], + "type": "text", + "content": "We further evaluate Tree-Ring in the identification task and surprisingly find that Tree-Ring struggles to distinguish between even dozens of different keys. Identification accuracy is reported in Tab. 2. The disappearance of the extra power of distribution shift has clearly exposed more design flaws in Tree-Ring. First, we find that tree-ring patterns are not discriminative enough to distinguish between different keys. This results in Tree-Ring's bad accuracy under various attacks or even under the clean setting. Meanwhile, the defective imprinting process leads to the injection of broken watermarks, further weakening its distinguishing ability. In addition, without the help of distribution shift, Tree-Ring is completely incapable of handling rotation and crop/scale transformations. Identification accuracy under these attacks is close to zero." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 131, + 641, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 641, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 131, + 641, + 482, + 666 + ], + "type": "text", + "content": "In order to address the above issues, we propose RingID for enhanced identification capability. It bases on a novel Multi-Channel Heterogeneous watermarking" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 214, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 214, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 214, + 101 + ], + "type": "text", + "content": "H. Ci et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 200 + ], + "type": "text", + "content": "framework, which can amalgamate distinctive advantages from different types of watermarks to resist to various kinds of attacks. Coupled with discretization and lossless imprinting methods, RingID gains substantial improvement on distinguishing ability. What's more, for the specific attack such as rotation, we locate the flaws in Tree-Ring's original design and propose systematic solutions, which improves the identification accuracy under rotation from nearly 0 to 0.86. To summarize our contributions:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 135, + 207, + 479, + 312 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 136, + 207, + 479, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 207, + 479, + 240 + ], + "spans": [ + { + "bbox": [ + 136, + 207, + 479, + 240 + ], + "type": "text", + "content": "1. For the first time, we elucidate how the unintentional introduction of distribution shift substantially contributes to Tree-Ring's robustness. We demonstrate this effect from both mathematical and empirical perspectives." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 135, + 243, + 479, + 265 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 243, + 479, + 265 + ], + "spans": [ + { + "bbox": [ + 135, + 243, + 479, + 265 + ], + "type": "text", + "content": "2. We systematically explore Tree-Ring's performance in the identification scenario and reveal its limitations through comprehensive analysis." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 135, + 267, + 479, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 267, + 479, + 312 + ], + "spans": [ + { + "bbox": [ + 135, + 267, + 479, + 312 + ], + "type": "text", + "content": "3. We present RingID, a systematic solution that significantly enhances watermark verification and identification, delivering impressive results. The verification AUC improves from 0.975 to 0.995, while the identification accuracy rises from 0.07 to 0.82." + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 132, + 332, + 237, + 344 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 332, + 237, + 344 + ], + "spans": [ + { + "bbox": [ + 132, + 332, + 237, + 344 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 351, + 246, + 362 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 351, + 246, + 362 + ], + "spans": [ + { + "bbox": [ + 132, + 351, + 246, + 362 + ], + "type": "text", + "content": "2.1 Diffusion Models" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 368, + 482, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 368, + 482, + 464 + ], + "spans": [ + { + "bbox": [ + 130, + 368, + 482, + 464 + ], + "type": "text", + "content": "Diffusion models [12, 18, 31] signify a swiftly advancing category of generative models. They define a forward diffusion process that gradually add small amount of Gaussian noise to the input data and learn to recover the clean data from the distortion [18]. Diffusion models initially demonstrate huge potential in generating high-resolution images [18, 32], resulting in globally recognized products like Midjourney [25], DALL-E [3, 26] and StableDiffusion [29]. Thereafter, they are broadly applied to modeling various types of data, e.g. video [4, 16, 30, 33, 39, 44], audio [19], 3d object [22, 27], humans [7, 8, 36, 50], etc." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 479, + 322, + 491 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 479, + 322, + 491 + ], + "spans": [ + { + "bbox": [ + 132, + 479, + 322, + 491 + ], + "type": "text", + "content": "2.2 Watermarking Diffusion Models" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 498, + 482, + 666 + ], + "type": "text", + "content": "When it comes to watermarking diffusion models, there are two different types of objectives: watermarking model weights [2, 23, 37, 45, 48] or generated content [6, 14, 35, 43, 46, 49]. Model weights watermarking methods aim to protect the intellectual property of the model itself and identify whether another model is an instance of plagiarism. While content watermarking methods protect the copyright of generated images [13, 42, 43] or videos [24, 46] by adding imperceptible perturbations onto pixels. They either imprint the watermark into the frequency via DFT [20], DCT [9, 11] and wavelet transformations [10, 40], or directly onto pixels via an encoder network [15, 35, 41, 49]. Recently, Tree-Ring [38] proposes to imprint a tree-ring watermark pattern into the initial diffusion noise. It demonstrates strong robustness to various attacks. In this work, we delve into Tree-Ring, uncovering its source of power in watermark verification and pinpointing its vulnerability in identifying multiple keys. Subsequently, we introduce RingID, a systematic solution fortified with enhanced identification capability." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 270, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 270, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 270, + 91, + 447, + 102 + ], + "type": "text", + "content": "RingID: Enhanced Multi-Key Identification" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 233, + 127 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 233, + 127 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 233, + 127 + ], + "type": "text", + "content": "3 Preliminaries" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 146, + 231, + 158 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 146, + 231, + 158 + ], + "spans": [ + { + "bbox": [ + 132, + 146, + 231, + 158 + ], + "type": "text", + "content": "3.1 Threat Model" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 172, + 482, + 268 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 172, + 482, + 268 + ], + "spans": [ + { + "bbox": [ + 130, + 172, + 482, + 268 + ], + "type": "text", + "content": "Model Owner: The owner of the diffusion model provides the image generation service through API access. He would like to imprint imperceptible watermarks into every generated image for copyright protection and source tracking. Given an arbitrary image, the model owner would like to check whether the image contains his own watermark (a.k.a. Watermark Verification) and which watermark it is if he distributes multiple different watermarks (a.k.a. Identification). The goal of the model owner is to ensure high verification and identification accuracy regardless of any image transformations." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 270, + 482, + 330 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 270, + 482, + 330 + ], + "spans": [ + { + "bbox": [ + 130, + 270, + 482, + 330 + ], + "type": "text", + "content": "Attacker: The attacker uses the model owner's service to generate an image. He attempts to apply various image transformations such as rotation and JPEG compression to disrupt the watermark, in order to evade the detection and identification check of the model owner. Subsequently, he can claim ownership of the image and profit from it illegally" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 354, + 206, + 364 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 354, + 206, + 364 + ], + "spans": [ + { + "bbox": [ + 132, + 354, + 206, + 364 + ], + "type": "text", + "content": "3.2 Notation" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 380, + 482, + 477 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 380, + 482, + 477 + ], + "spans": [ + { + "bbox": [ + 130, + 380, + 482, + 477 + ], + "type": "text", + "content": "We denote the initial diffusion noise by " + }, + { + "bbox": [ + 130, + 380, + 482, + 477 + ], + "type": "inline_equation", + "content": "x_{T}" + }, + { + "bbox": [ + 130, + 380, + 482, + 477 + ], + "type": "text", + "content": ", and its frequency counterpart after FFT by " + }, + { + "bbox": [ + 130, + 380, + 482, + 477 + ], + "type": "inline_equation", + "content": "X_{T}" + }, + { + "bbox": [ + 130, + 380, + 482, + 477 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 130, + 380, + 482, + 477 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 130, + 380, + 482, + 477 + ], + "type": "text", + "content": " is the diffusion step. Sometimes we may omit the subscript " + }, + { + "bbox": [ + 130, + 380, + 482, + 477 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 130, + 380, + 482, + 477 + ], + "type": "text", + "content": " for simplicity. We denote the real part, imaginary part and conjugate symmetry of " + }, + { + "bbox": [ + 130, + 380, + 482, + 477 + ], + "type": "inline_equation", + "content": "X_{T}" + }, + { + "bbox": [ + 130, + 380, + 482, + 477 + ], + "type": "text", + "content": " by " + }, + { + "bbox": [ + 130, + 380, + 482, + 477 + ], + "type": "inline_equation", + "content": "X_{\\mathrm{re}}" + }, + { + "bbox": [ + 130, + 380, + 482, + 477 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 380, + 482, + 477 + ], + "type": "inline_equation", + "content": "X_{\\mathrm{im}}" + }, + { + "bbox": [ + 130, + 380, + 482, + 477 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 380, + 482, + 477 + ], + "type": "inline_equation", + "content": "X_{\\mathrm{cs}}" + }, + { + "bbox": [ + 130, + 380, + 482, + 477 + ], + "type": "text", + "content": ". In the context of StableDiffusion [29], " + }, + { + "bbox": [ + 130, + 380, + 482, + 477 + ], + "type": "inline_equation", + "content": "X_{T} \\in R^{h \\times w \\times c}" + }, + { + "bbox": [ + 130, + 380, + 482, + 477 + ], + "type": "text", + "content": " has 4 channels " + }, + { + "bbox": [ + 130, + 380, + 482, + 477 + ], + "type": "inline_equation", + "content": "\\mathcal{C} = \\{0,1,2,3\\}" + }, + { + "bbox": [ + 130, + 380, + 482, + 477 + ], + "type": "text", + "content": ". We denote the watermarked channels by " + }, + { + "bbox": [ + 130, + 380, + 482, + 477 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_w" + }, + { + "bbox": [ + 130, + 380, + 482, + 477 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 380, + 482, + 477 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_w \\subset \\mathcal{C}" + }, + { + "bbox": [ + 130, + 380, + 482, + 477 + ], + "type": "text", + "content": ". We denote the imprinted key by " + }, + { + "bbox": [ + 130, + 380, + 482, + 477 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 130, + 380, + 482, + 477 + ], + "type": "text", + "content": ", the recovered key at detection time by " + }, + { + "bbox": [ + 130, + 380, + 482, + 477 + ], + "type": "inline_equation", + "content": "\\hat{w}" + }, + { + "bbox": [ + 130, + 380, + 482, + 477 + ], + "type": "text", + "content": ", and the null watermark extracted from the unwatermarked image by " + }, + { + "bbox": [ + 130, + 380, + 482, + 477 + ], + "type": "inline_equation", + "content": "\\hat{w}_{\\emptyset}" + }, + { + "bbox": [ + 130, + 380, + 482, + 477 + ], + "type": "text", + "content": ". We use " + }, + { + "bbox": [ + 130, + 380, + 482, + 477 + ], + "type": "inline_equation", + "content": "\\mathcal{W} = \\{w_1, w_2, \\ldots\\}" + }, + { + "bbox": [ + 130, + 380, + 482, + 477 + ], + "type": "text", + "content": " to denote a set of different keys. " + }, + { + "bbox": [ + 130, + 380, + 482, + 477 + ], + "type": "inline_equation", + "content": "|\\mathcal{W}|" + }, + { + "bbox": [ + 130, + 380, + 482, + 477 + ], + "type": "text", + "content": " is its capacity." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 497, + 402, + 509 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 497, + 402, + 509 + ], + "spans": [ + { + "bbox": [ + 132, + 497, + 402, + 509 + ], + "type": "text", + "content": "3.3 Task Formulation: Verification vs. Identification" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 521, + 482, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 482, + 605 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 482, + 605 + ], + "type": "text", + "content": "In this section, we formulate the verification and identification task following the problem statement of Tree-Ring [38]. The task of watermark verification is to detect the existence of watermarks. Verification algorithms attempt to distinguish between watermarked images carrying key " + }, + { + "bbox": [ + 130, + 521, + 482, + 605 + ], + "type": "inline_equation", + "content": "\\hat{w}" + }, + { + "bbox": [ + 130, + 521, + 482, + 605 + ], + "type": "text", + "content": " and unwatermarked images carrying null key " + }, + { + "bbox": [ + 130, + 521, + 482, + 605 + ], + "type": "inline_equation", + "content": "\\hat{w}_{\\emptyset}" + }, + { + "bbox": [ + 130, + 521, + 482, + 605 + ], + "type": "text", + "content": ". More formally, verification can be formulated as distinguishing between the distance to the reference key " + }, + { + "bbox": [ + 130, + 521, + 482, + 605 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 130, + 521, + 482, + 605 + ], + "type": "text", + "content": ": " + }, + { + "bbox": [ + 130, + 521, + 482, + 605 + ], + "type": "inline_equation", + "content": "d(\\hat{w}, w) \\neq d(\\hat{w}_{\\emptyset}, w)" + }, + { + "bbox": [ + 130, + 521, + 482, + 605 + ], + "type": "text", + "content": ". In Tree-Ring, " + }, + { + "bbox": [ + 130, + 521, + 482, + 605 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 130, + 521, + 482, + 605 + ], + "type": "text", + "content": " is chosen to be " + }, + { + "bbox": [ + 130, + 521, + 482, + 605 + ], + "type": "inline_equation", + "content": "\\ell_{1}" + }, + { + "bbox": [ + 130, + 521, + 482, + 605 + ], + "type": "text", + "content": " distance and " + }, + { + "bbox": [ + 130, + 521, + 482, + 605 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 130, + 521, + 482, + 605 + ], + "type": "text", + "content": " is the watermark to be imprinted." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "type": "text", + "content": "When a set of different keys " + }, + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "type": "inline_equation", + "content": "w_{i} \\in \\mathcal{W}" + }, + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "type": "text", + "content": " are distributed, the goal of watermark identification is to find the best match for the recovered key " + }, + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "type": "inline_equation", + "content": "\\hat{w}" + }, + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "type": "text", + "content": " among all candidates in " + }, + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "type": "inline_equation", + "content": "\\mathcal{W}" + }, + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "type": "text", + "content": ", i.e. " + }, + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "type": "inline_equation", + "content": "\\mathrm{ID}(\\hat{w}) = \\arg \\min_{i} [d(\\hat{w}, w_{i})]" + }, + { + "bbox": [ + 130, + 605, + 482, + 665 + ], + "type": "text", + "content": ". In practice, different users are usually assigned different keys. Thus, each generated image can be source tracked and attributed to a unique user via watermark identification." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 212, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 212, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 212, + 100 + ], + "type": "text", + "content": "H. Ci et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 136, + 118, + 478, + 273 + ], + "blocks": [ + { + "bbox": [ + 136, + 118, + 478, + 273 + ], + "lines": [ + { + "bbox": [ + 136, + 118, + 478, + 273 + ], + "spans": [ + { + "bbox": [ + 136, + 118, + 478, + 273 + ], + "type": "image", + "image_path": "96908b8f6a6d835c80c90827bb5c6c76a73b7986b046f2384401e79b5775d6b4.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 285, + 482, + 319 + ], + "lines": [ + { + "bbox": [ + 132, + 285, + 482, + 319 + ], + "spans": [ + { + "bbox": [ + 132, + 285, + 482, + 319 + ], + "type": "text", + "content": "Fig. 2: Framework of the watermarking process. RingID introduces a series of approaches that can help to imprint a lossless and robust watermark. In contrast, TreeRing injects a lossy and less robust watermark." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 341, + 369, + 355 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 341, + 369, + 355 + ], + "spans": [ + { + "bbox": [ + 132, + 341, + 369, + 355 + ], + "type": "text", + "content": "4 Delve into Tree-Ring Watermarking" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 368, + 482, + 404 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 368, + 482, + 404 + ], + "spans": [ + { + "bbox": [ + 132, + 368, + 482, + 404 + ], + "type": "text", + "content": "In this section, we will reveal why Tree-Ring [38] is robust to various image transformations, even some of them are not considered by design. We further evaluate Tree-Ring in the multi-key identification task and pinpoint its vulnerabilities." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 423, + 259, + 436 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 423, + 259, + 436 + ], + "spans": [ + { + "bbox": [ + 132, + 423, + 259, + 436 + ], + "type": "text", + "content": "4.1 Recap of Tree-Ring" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 445, + 482, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 445, + 482, + 601 + ], + "spans": [ + { + "bbox": [ + 132, + 445, + 482, + 601 + ], + "type": "text", + "content": "Tree-Ring [38] originally studies the verification setting. Its goal is to distinguish watermarked images from non-watermarked images. As illustrated in Fig. 2, Tree-Ring transforms the initial diffusion noise " + }, + { + "bbox": [ + 132, + 445, + 482, + 601 + ], + "type": "inline_equation", + "content": "x_{T}" + }, + { + "bbox": [ + 132, + 445, + 482, + 601 + ], + "type": "text", + "content": " into the frequency " + }, + { + "bbox": [ + 132, + 445, + 482, + 601 + ], + "type": "inline_equation", + "content": "X_{T}" + }, + { + "bbox": [ + 132, + 445, + 482, + 601 + ], + "type": "text", + "content": " and injects a tree-ring pattern " + }, + { + "bbox": [ + 132, + 445, + 482, + 601 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 132, + 445, + 482, + 601 + ], + "type": "text", + "content": " into the center of " + }, + { + "bbox": [ + 132, + 445, + 482, + 601 + ], + "type": "inline_equation", + "content": "X_{T}" + }, + { + "bbox": [ + 132, + 445, + 482, + 601 + ], + "type": "text", + "content": ", where the value of each ring in pattern " + }, + { + "bbox": [ + 132, + 445, + 482, + 601 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 132, + 445, + 482, + 601 + ], + "type": "text", + "content": " is sampled from Gaussian. It then transforms the watermarked " + }, + { + "bbox": [ + 132, + 445, + 482, + 601 + ], + "type": "inline_equation", + "content": "X_{T}" + }, + { + "bbox": [ + 132, + 445, + 482, + 601 + ], + "type": "text", + "content": " back to the spatial domain, retaining only the real part to feed the diffusion network for image generation. Notably, the accompanying imaginary part is discarded to be compatible with the network input. For detection of the watermark, Tree-Ring inverts the diffusion process to estimate " + }, + { + "bbox": [ + 132, + 445, + 482, + 601 + ], + "type": "inline_equation", + "content": "x_{T}" + }, + { + "bbox": [ + 132, + 445, + 482, + 601 + ], + "type": "text", + "content": " from an image, transforms it to the frequency domain, and extracts the watermark pattern " + }, + { + "bbox": [ + 132, + 445, + 482, + 601 + ], + "type": "inline_equation", + "content": "\\hat{w}" + }, + { + "bbox": [ + 132, + 445, + 482, + 601 + ], + "type": "text", + "content": ". Similarly, nonsense pattern " + }, + { + "bbox": [ + 132, + 445, + 482, + 601 + ], + "type": "inline_equation", + "content": "\\hat{w}_{\\emptyset}" + }, + { + "bbox": [ + 132, + 445, + 482, + 601 + ], + "type": "text", + "content": " can be extracted from the non-watermarked image. Tree-Ring distinguishes between these two " + }, + { + "bbox": [ + 132, + 445, + 482, + 601 + ], + "type": "inline_equation", + "content": "\\ell_{1}" + }, + { + "bbox": [ + 132, + 445, + 482, + 601 + ], + "type": "text", + "content": "-to-reference distances: " + }, + { + "bbox": [ + 132, + 445, + 482, + 601 + ], + "type": "inline_equation", + "content": "\\| \\hat{w} - w \\|_{1}" + }, + { + "bbox": [ + 132, + 445, + 482, + 601 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 132, + 445, + 482, + 601 + ], + "type": "inline_equation", + "content": "\\| \\hat{w}_{\\emptyset} - w \\|_{1}" + }, + { + "bbox": [ + 132, + 445, + 482, + 601 + ], + "type": "text", + "content": " to determine whether an image is watermarked or not." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 620, + 340, + 632 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 620, + 340, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 620, + 340, + 632 + ], + "type": "text", + "content": "4.2 Distribution Shift in Watermarking" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 641, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 641, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 641, + 481, + 665 + ], + "type": "text", + "content": "Tree-Ring's detection algorithm bases on the intuition that the recovered pattern " + }, + { + "bbox": [ + 132, + 641, + 481, + 665 + ], + "type": "inline_equation", + "content": "\\hat{w}" + }, + { + "bbox": [ + 132, + 641, + 481, + 665 + ], + "type": "text", + "content": " from a watermarked image should match the injected reference pattern " + }, + { + "bbox": [ + 132, + 641, + 481, + 665 + ], + "type": "inline_equation", + "content": "w" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 270, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 270, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 270, + 91, + 447, + 102 + ], + "type": "text", + "content": "RingID: Enhanced Multi-Key Identification" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 151, + 115, + 462, + 191 + ], + "blocks": [ + { + "bbox": [ + 151, + 115, + 462, + 191 + ], + "lines": [ + { + "bbox": [ + 151, + 115, + 462, + 191 + ], + "spans": [ + { + "bbox": [ + 151, + 115, + 462, + 191 + ], + "type": "image", + "image_path": "3282c7e1061491af0c189db50e1099cac935da4f5b890c5a9f3f983a5adc5bcf.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 151, + 193, + 462, + 237 + ], + "blocks": [ + { + "bbox": [ + 151, + 193, + 462, + 237 + ], + "lines": [ + { + "bbox": [ + 151, + 193, + 462, + 237 + ], + "spans": [ + { + "bbox": [ + 151, + 193, + 462, + 237 + ], + "type": "image", + "image_path": "4833b9f577866c4ada8aabea1bc2c47997ec67079d34f606000be252bde96986.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 246, + 482, + 280 + ], + "lines": [ + { + "bbox": [ + 131, + 246, + 482, + 280 + ], + "spans": [ + { + "bbox": [ + 131, + 246, + 482, + 280 + ], + "type": "text", + "content": "Fig. 3: The effect of distribution shift in both verification and identification tasks. Discarding the imaginary part during watermark imprinting causes the distribution shift in " + }, + { + "bbox": [ + 131, + 246, + 482, + 280 + ], + "type": "inline_equation", + "content": "\\ell_1" + }, + { + "bbox": [ + 131, + 246, + 482, + 280 + ], + "type": "text", + "content": " distance. Distribution shift aids verification but not identification." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 304, + 482, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 304, + 482, + 327 + ], + "spans": [ + { + "bbox": [ + 131, + 304, + 482, + 327 + ], + "type": "text", + "content": "to some extent, leading to a smaller " + }, + { + "bbox": [ + 131, + 304, + 482, + 327 + ], + "type": "inline_equation", + "content": "\\ell_1" + }, + { + "bbox": [ + 131, + 304, + 482, + 327 + ], + "type": "text", + "content": "-to-reference distance, i.e. " + }, + { + "bbox": [ + 131, + 304, + 482, + 327 + ], + "type": "inline_equation", + "content": "\\| \\hat{w} - w \\|_1 < \\| \\hat{w}_{\\emptyset} - w \\|_1" + }, + { + "bbox": [ + 131, + 304, + 482, + 327 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 328, + 482, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 328, + 482, + 413 + ], + "spans": [ + { + "bbox": [ + 130, + 328, + 482, + 413 + ], + "type": "text", + "content": "However, we have discovered that apart from the pattern matching, another crucial factor contributes to the separation of " + }, + { + "bbox": [ + 130, + 328, + 482, + 413 + ], + "type": "inline_equation", + "content": "\\| \\hat{w} - w \\|_1" + }, + { + "bbox": [ + 130, + 328, + 482, + 413 + ], + "type": "text", + "content": " v.s. " + }, + { + "bbox": [ + 130, + 328, + 482, + 413 + ], + "type": "inline_equation", + "content": "\\| \\hat{w}_{\\emptyset} - w \\|_1" + }, + { + "bbox": [ + 130, + 328, + 482, + 413 + ], + "type": "text", + "content": ". This factor is the distribution shift introduced by discarding the imaginary part during the watermark injection process, which is highlighted in Fig. 2. This affects all watermarked images, further reducing the expectation of " + }, + { + "bbox": [ + 130, + 328, + 482, + 413 + ], + "type": "inline_equation", + "content": "\\mathbb{E}\\left[\\| \\hat{w} - w \\|_1\\right]" + }, + { + "bbox": [ + 130, + 328, + 482, + 413 + ], + "type": "text", + "content": ". In other words, it shifts the distribution of " + }, + { + "bbox": [ + 130, + 328, + 482, + 413 + ], + "type": "inline_equation", + "content": "\\| \\hat{w} - w \\|_1" + }, + { + "bbox": [ + 130, + 328, + 482, + 413 + ], + "type": "text", + "content": " towards a smaller value, making it easier to distinguish between " + }, + { + "bbox": [ + 130, + 328, + 482, + 413 + ], + "type": "inline_equation", + "content": "\\| \\hat{w} - w \\|_1" + }, + { + "bbox": [ + 130, + 328, + 482, + 413 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 328, + 482, + 413 + ], + "type": "inline_equation", + "content": "\\| \\hat{w}_{\\emptyset} - w \\|_1" + }, + { + "bbox": [ + 130, + 328, + 482, + 413 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 413, + 482, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 413, + 482, + 450 + ], + "spans": [ + { + "bbox": [ + 130, + 413, + 482, + 450 + ], + "type": "text", + "content": "Mathematically, we prove that the shift factor is " + }, + { + "bbox": [ + 130, + 413, + 482, + 450 + ], + "type": "inline_equation", + "content": "\\frac{\\sqrt{3}}{2}" + }, + { + "bbox": [ + 130, + 413, + 482, + 450 + ], + "type": "text", + "content": " in a general and simplified setting. Let " + }, + { + "bbox": [ + 130, + 413, + 482, + 450 + ], + "type": "inline_equation", + "content": "\\hat{w}_n" + }, + { + "bbox": [ + 130, + 413, + 482, + 450 + ], + "type": "text", + "content": " denote the recovered watermark that has never been shifted, thus we have the following:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 229, + 460, + 481, + 485 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 229, + 460, + 481, + 485 + ], + "spans": [ + { + "bbox": [ + 229, + 460, + 481, + 485 + ], + "type": "interline_equation", + "content": "\\mathbb {E} \\left[ \\| \\hat {w} - w \\| _ {1} \\right] = \\frac {\\sqrt {3}}{2} \\mathbb {E} \\left[ \\| \\hat {w} _ {n} - w \\| _ {1} \\right]. \\tag {1}", + "image_path": "4ce94c96fee2e66a5690923540cd09fbbe2f0bbdaafa86f6667da02ee7dae9ed.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 490, + 482, + 539 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 490, + 482, + 539 + ], + "spans": [ + { + "bbox": [ + 131, + 490, + 482, + 539 + ], + "type": "text", + "content": "Detailed derivation is elaborated in Supplementary A. Actually, we find any operation leading to a sufficient distribution shift can be considered a good standalone watermarking approach. We demonstrate the operation of discarding the imaginary part performs well on its own in Supplementary B." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 557, + 483, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 557, + 483, + 666 + ], + "spans": [ + { + "bbox": [ + 131, + 557, + 483, + 666 + ], + "type": "text", + "content": "Effect in Different Tasks and Attacks Fig. 3 illustrates the effect of distribution shift under different tasks. In Verification, it helps distinguish between watermarked and non-watermarked images. Next, we go one step further and consider the situation when the image undergoes different attacks like JPEG or Resize. If the tree-ring pattern itself can cope with a certain attack, the improvement brought by the distribution shift will be small, because the two distributions are already separable even without the help of shift. On the contrary, if the tree-ring pattern itself cannot deal with a certain attack, the improvement brought by the distribution shift will be significant. Empirically (Supplementary" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 214, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 214, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 214, + 101 + ], + "type": "text", + "content": "H. Ci et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 175 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 175 + ], + "type": "text", + "content": "A.6), we find distribution shift helps a lot under Rotation and Crop/Scale attacks. This implies that the original design of tree-ring cannot handle these two attacks. When it comes to Identification, things are a little different. Since different watermarks experience the same distribution shift, discarding the imaginary part doesn't help distinguishing them." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 176, + 479, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 176, + 479, + 224 + ], + "spans": [ + { + "bbox": [ + 130, + 176, + 479, + 224 + ], + "type": "text", + "content": "This sparks our curiosity about Tree-Ring's performance in identification, since distribution shift doesn't help and identification is more difficult. We would like to know if Tree-Ring still works under Rotation and Crop/Scale attacks. In the next section, we will explore these questions." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 239, + 344, + 251 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 239, + 344, + 251 + ], + "spans": [ + { + "bbox": [ + 131, + 239, + 344, + 251 + ], + "type": "text", + "content": "4.3 Is Tree-Ring Good at Identification?" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 257, + 480, + 293 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 257, + 480, + 293 + ], + "spans": [ + { + "bbox": [ + 130, + 257, + 480, + 293 + ], + "type": "text", + "content": "We report the performance of Tree-Ring when identifying different numbers of keys in Tab. 2. Tree-Ring struggles to identify as few as 32 keys. What's more, identification accuracy is nearly zero under Rotation and Crop/Scale attacks." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 293, + 479, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 293, + 479, + 365 + ], + "spans": [ + { + "bbox": [ + 130, + 293, + 479, + 365 + ], + "type": "text", + "content": "We summarize 2 reasons as follows: (1) Distribution shift doesn't help. As tree-ring watermark cannot handle Rotation and Crop/Scale attacks, identification accuracy under these attacks drops to nearly zero. (2) Tee-ring pattern doesn't have enough distinguishing ability. Since identification is more difficult than verification, poor distinguishing ability results in poor performance under most attacks or even under no attack setting." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 380, + 457, + 392 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 380, + 457, + 392 + ], + "spans": [ + { + "bbox": [ + 131, + 380, + 457, + 392 + ], + "type": "text", + "content": "4.4 Why is Tree-Ring Vulnerable to Rotation and Crop/Scale?" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 399, + 479, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 399, + 479, + 422 + ], + "spans": [ + { + "bbox": [ + 130, + 399, + 479, + 422 + ], + "type": "text", + "content": "All previous experimental results and analysis show one thing, tree-ring pattern cannot handle these two attacks. In this section, we will see what happens." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 422, + 479, + 493 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 422, + 479, + 493 + ], + "spans": [ + { + "bbox": [ + 130, + 422, + 479, + 493 + ], + "type": "text", + "content": "Vulnerable to Crop & Scale Scaling an image by a factor of " + }, + { + "bbox": [ + 130, + 422, + 479, + 493 + ], + "type": "inline_equation", + "content": "a" + }, + { + "bbox": [ + 130, + 422, + 479, + 493 + ], + "type": "text", + "content": " corresponds to compressing the frequency spectrum " + }, + { + "bbox": [ + 130, + 422, + 479, + 493 + ], + "type": "inline_equation", + "content": "X[u,v]" + }, + { + "bbox": [ + 130, + 422, + 479, + 493 + ], + "type": "text", + "content": " in both magnitudes and positions: " + }, + { + "bbox": [ + 130, + 422, + 479, + 493 + ], + "type": "inline_equation", + "content": "\\frac{1}{a^2} X\\left[\\frac{u}{a},\\frac{v}{a}\\right]" + }, + { + "bbox": [ + 130, + 422, + 479, + 493 + ], + "type": "text", + "content": ". Cropping an image modulates the frequency magnitudes by a sinc function. It is straightforward that these transformations easily disrupt the tree-ring pattern on frequency domain. As shown in Fig. 4(e), the pattern is completely disrupted." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 494, + 479, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 494, + 479, + 553 + ], + "spans": [ + { + "bbox": [ + 130, + 494, + 479, + 553 + ], + "type": "text", + "content": "Vulnerable to Rotation Although Tree-Ring is designed as ring patterns to seek rotation robustness, we find the existing design ignores many details, leading to no rotation robustness. In short, a combination of corner cropping when rotating, a lossy injection process, and the rotationally asymmetric shape cause the problem. We will explain in detail and propose solutions in Sec. 5.2." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 131, + 571, + 456, + 583 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 571, + 456, + 583 + ], + "spans": [ + { + "bbox": [ + 131, + 571, + 456, + 583 + ], + "type": "text", + "content": "5 RingID: Enhanced Watermarking for Identification" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 594, + 479, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 479, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 479, + 665 + ], + "type": "text", + "content": "In this section, we show how our proposed method RingID advances in identifying multiple distinct keys. RingID originates from Tree-Ring but incorporates systematic enhancements. We introduce a multi-channel heterogeneous watermarking framework, a discretization approach and a method of lossless imprinting to generally improve its distinguishing ability. We also discuss a series of designs regarding stronger rotation robustness and increased capacity." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 270, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 270, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 270, + 91, + 447, + 102 + ], + "type": "text", + "content": "RingID: Enhanced Multi-Key Identification" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 388, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 388, + 129 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 388, + 129 + ], + "type": "text", + "content": "5.1 Multi-Channel Heterogeneous Watermarking" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 137, + 482, + 209 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 137, + 482, + 209 + ], + "spans": [ + { + "bbox": [ + 130, + 137, + 482, + 209 + ], + "type": "text", + "content": "Multi-channel heterogeneous (MCH) watermarking serves as a general framework to comprehensively improve watermarking robustness to a battery of attacks. Intuitively, we imprint different types of watermarks on different channels of the initial noise to amalgamate their distinctive advantages, as illustrated in Fig. 1. During matching, we calculate the minimum of channel-wise " + }, + { + "bbox": [ + 130, + 137, + 482, + 209 + ], + "type": "inline_equation", + "content": "\\ell_1" + }, + { + "bbox": [ + 130, + 137, + 482, + 209 + ], + "type": "text", + "content": " distance to the reference watermark for all watermarked channels:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 218, + 216, + 481, + 242 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 218, + 216, + 481, + 242 + ], + "spans": [ + { + "bbox": [ + 218, + 216, + 481, + 242 + ], + "type": "interline_equation", + "content": "\\operatorname {I D} (\\hat {w}) = \\arg \\min _ {i} \\left\\{\\min _ {c \\in C _ {w}} \\left[ \\lambda_ {c} \\| \\hat {w} ^ {c} - w _ {i} ^ {c} \\| _ {1} \\right] \\right\\} \\tag {2}", + "image_path": "3febbdc412626f829ffc3e25b4062983ab1b0a8bfcf526f910bd38d6f0e5a49a.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 249, + 479, + 285 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 249, + 479, + 285 + ], + "spans": [ + { + "bbox": [ + 130, + 249, + 479, + 285 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 249, + 479, + 285 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 130, + 249, + 479, + 285 + ], + "type": "text", + "content": " denotes the channel index and traverses all watermarked channels " + }, + { + "bbox": [ + 130, + 249, + 479, + 285 + ], + "type": "inline_equation", + "content": "C_w" + }, + { + "bbox": [ + 130, + 249, + 479, + 285 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 130, + 249, + 479, + 285 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 249, + 479, + 285 + ], + "type": "text", + "content": " denotes the index of candidate reference watermarks. " + }, + { + "bbox": [ + 130, + 249, + 479, + 285 + ], + "type": "inline_equation", + "content": "\\lambda_c" + }, + { + "bbox": [ + 130, + 249, + 479, + 285 + ], + "type": "text", + "content": " is a channel-wise normalizing factor." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 285, + 482, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 285, + 482, + 453 + ], + "spans": [ + { + "bbox": [ + 130, + 285, + 482, + 453 + ], + "type": "text", + "content": "Eq. (2) bases on the intuition that if a watermark is robust to a certain attack, then its distance from the ground-truth reference is smaller than that of other watermarks that are not robust. Ideally, it enables adaptive selection of the most robust watermark under different attacks. Notably, the total capacity of MCH is the minimum of each watermark. Therefore, it's recommended to use a watermark pattern with sufficient capacity on each channel. At the same time, in order to reduce the impact on the generation quality, we need to carefully select the type of watermarks for combination. We find the Gaussian noise pattern is a perfect candidate. It has infinite capacity and the same distribution as the initial noise. Meanwhile, it has good robustness to non-geometric attacks [38] and is complementary to the tree-ring pattern. So RingID explores the combination of a Gaussian noise watermark and a tree-ring watermark. Empirical study in Sec. 6.3 demonstrates that this combination is able to perfectly amalgamate unique advantages from the noise and ring watermark." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 471, + 302, + 483 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 471, + 302, + 483 + ], + "spans": [ + { + "bbox": [ + 132, + 471, + 302, + 483 + ], + "type": "text", + "content": "5.2 Being Invariant to Rotation" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 491, + 481, + 528 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 491, + 481, + 528 + ], + "spans": [ + { + "bbox": [ + 130, + 491, + 481, + 528 + ], + "type": "text", + "content": "Although the tree-ring pattern is designed to be rotationally symmetric, it exhibits unexpected vulnerability to rotation as shown in Tab. 2. Below, we identify the causes of the issue and propose a series of measures to address it." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 545, + 482, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 482, + 668 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 482, + 668 + ], + "type": "text", + "content": "Spatial Shift to Avoid Corner Cropping We visualize the imprinted tree-ring watermark in spatial domain in Fig. 4(c). Due to the properties of FFT, tree-ring patterns in the frequency tends to exhibit concentrated energy in the four corners in the spatial domain, rendering them vulnerable to cropping during rotation. We propose to circularly shift the watermark pattern in the spatial domain to the center of the image, thus the majority of watermark information is kept while rotating. For an " + }, + { + "bbox": [ + 130, + 545, + 482, + 668 + ], + "type": "inline_equation", + "content": "N \\times N" + }, + { + "bbox": [ + 130, + 545, + 482, + 668 + ], + "type": "text", + "content": " latent image, we circularly shift the pattern by " + }, + { + "bbox": [ + 130, + 545, + 482, + 668 + ], + "type": "inline_equation", + "content": "\\frac{N}{2}" + }, + { + "bbox": [ + 130, + 545, + 482, + 668 + ], + "type": "text", + "content": " pixels in both width and height dimension. This is equivalent to directly multiplying the frequency watermark with a chessboard pattern given by " + }, + { + "bbox": [ + 130, + 545, + 482, + 668 + ], + "type": "inline_equation", + "content": "H[u,v] = e^{j\\frac{\\pi}{2}(u + v)} = (-1)^{u + v}" + }, + { + "bbox": [ + 130, + 545, + 482, + 668 + ], + "type": "text", + "content": ". The change of the watermark can be clearly" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 214, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 214, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 214, + 101 + ], + "type": "text", + "content": "H. Ci et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 153, + 148, + 194, + 204 + ], + "blocks": [ + { + "bbox": [ + 162, + 121, + 233, + 131 + ], + "lines": [ + { + "bbox": [ + 162, + 121, + 233, + 131 + ], + "spans": [ + { + "bbox": [ + 162, + 121, + 233, + 131 + ], + "type": "text", + "content": "(a) Discretization" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 153, + 148, + 194, + 204 + ], + "lines": [ + { + "bbox": [ + 153, + 148, + 194, + 204 + ], + "spans": [ + { + "bbox": [ + 153, + 148, + 194, + 204 + ], + "type": "image", + "image_path": "0818580d2bedb26f0e04968fab624344aeee8267d1b8b36a501bcd2be207d2a8.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 202, + 149, + 242, + 205 + ], + "blocks": [ + { + "bbox": [ + 202, + 149, + 242, + 205 + ], + "lines": [ + { + "bbox": [ + 202, + 149, + 242, + 205 + ], + "spans": [ + { + "bbox": [ + 202, + 149, + 242, + 205 + ], + "type": "image", + "image_path": "01c7fb2d70c7e839178ac58ed1b6bbd852b1046a240bbdfe20975ad884cb1067.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 261, + 132, + 302, + 227 + ], + "blocks": [ + { + "bbox": [ + 271, + 121, + 342, + 131 + ], + "lines": [ + { + "bbox": [ + 271, + 121, + 342, + 131 + ], + "spans": [ + { + "bbox": [ + 271, + 121, + 342, + 131 + ], + "type": "text", + "content": "(b) Rounder Ring" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 261, + 132, + 302, + 227 + ], + "lines": [ + { + "bbox": [ + 261, + 132, + 302, + 227 + ], + "spans": [ + { + "bbox": [ + 261, + 132, + 302, + 227 + ], + "type": "image", + "image_path": "8b5d329ded6bcd03c6f203559fadbd4f6df61a8db74be7ab7544c3ee75e6fc9f.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 311, + 132, + 352, + 232 + ], + "blocks": [ + { + "bbox": [ + 311, + 132, + 352, + 232 + ], + "lines": [ + { + "bbox": [ + 311, + 132, + 352, + 232 + ], + "spans": [ + { + "bbox": [ + 311, + 132, + 352, + 232 + ], + "type": "image", + "image_path": "d6e9a183641f6ce5934fc7c18aaf07b7d9e5cc80704d272ab81ca4e00eeed782.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 369, + 121, + 462, + 233 + ], + "blocks": [ + { + "bbox": [ + 369, + 121, + 462, + 233 + ], + "lines": [ + { + "bbox": [ + 369, + 121, + 462, + 233 + ], + "spans": [ + { + "bbox": [ + 369, + 121, + 462, + 233 + ], + "type": "image", + "image_path": "fb9afe1dece5f3b4fedf427f6ea6e1397b4f51b21af015a3908ff649aae20d43.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 174, + 252, + 324, + 352 + ], + "blocks": [ + { + "bbox": [ + 206, + 241, + 300, + 251 + ], + "lines": [ + { + "bbox": [ + 206, + 241, + 300, + 251 + ], + "spans": [ + { + "bbox": [ + 206, + 241, + 300, + 251 + ], + "type": "text", + "content": "(d) Lossless Imprinting" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 174, + 252, + 324, + 352 + ], + "lines": [ + { + "bbox": [ + 174, + 252, + 324, + 352 + ], + "spans": [ + { + "bbox": [ + 174, + 252, + 324, + 352 + ], + "type": "image", + "image_path": "f22f1968a17b015f0dfbb5599709215f959d29fcee5f9283ed9499e1a5bbb261.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 365, + 252, + 457, + 352 + ], + "blocks": [ + { + "bbox": [ + 378, + 241, + 444, + 251 + ], + "lines": [ + { + "bbox": [ + 378, + 241, + 444, + 251 + ], + "spans": [ + { + "bbox": [ + 378, + 241, + 444, + 251 + ], + "type": "text", + "content": "(e) Crop & Scale" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 365, + 252, + 457, + 352 + ], + "lines": [ + { + "bbox": [ + 365, + 252, + 457, + 352 + ], + "spans": [ + { + "bbox": [ + 365, + 252, + 457, + 352 + ], + "type": "image", + "image_path": "65252de08df1416b6425d5274a3c928ce6c4aa3ca64b9ae457a77b03f03e1476.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 361, + 482, + 450 + ], + "lines": [ + { + "bbox": [ + 130, + 361, + 482, + 450 + ], + "spans": [ + { + "bbox": [ + 130, + 361, + 482, + 450 + ], + "type": "text", + "content": "Fig. 4: Detailed illustration of different acts and their impacts. (a) Discretization: Discretize values to " + }, + { + "bbox": [ + 130, + 361, + 482, + 450 + ], + "type": "inline_equation", + "content": "\\pm \\alpha" + }, + { + "bbox": [ + 130, + 361, + 482, + 450 + ], + "type": "text", + "content": " for each ring, increasing the distinguishing ability of the watermark. (b) Rounder ring: Render smoother edges and reduce artifacts. Increase rotational robustness. (c) Spatial Shift: Shift energy on the corners to the center, avoiding potential cropping during rotation. (d) Lossless Imprinting: Ensure the imprinted pattern the same as the original design. (e) Crop & Scale: Zooming in the spatial domain compresses the imprinted pattern in the frequency domain, causing pattern mismatch and hinders identification." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 475, + 482, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 475, + 482, + 548 + ], + "spans": [ + { + "bbox": [ + 130, + 475, + 482, + 548 + ], + "type": "text", + "content": "observed in Fig. 4(c). Although spatial-shifting helps a lot for rotation invariance, we noticed that this often leads to the generation of circular artifacts at the center of the image. Keeping generation quality in mind, we further multiply the shifted pattern by a factor " + }, + { + "bbox": [ + 130, + 475, + 482, + 548 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 130, + 475, + 482, + 548 + ], + "type": "text", + "content": " to suppress the peak at the center. " + }, + { + "bbox": [ + 130, + 475, + 482, + 548 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 130, + 475, + 482, + 548 + ], + "type": "text", + "content": " trades-off the robustness, as smaller " + }, + { + "bbox": [ + 130, + 475, + 482, + 548 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 130, + 475, + 482, + 548 + ], + "type": "text", + "content": " weakens the watermark pattern. In practice, we find " + }, + { + "bbox": [ + 130, + 475, + 482, + 548 + ], + "type": "inline_equation", + "content": "\\eta \\in [0.8,0.9]" + }, + { + "bbox": [ + 130, + 475, + 482, + 548 + ], + "type": "text", + "content": " strikes a good balance between image quality and robustness." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "type": "text", + "content": "Lossless Imprinting The operation of discarding the imaginary part during watermarking not only shifts the distribution of " + }, + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\ell_1" + }, + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "type": "text", + "content": "-to-reference, but also disrupt the pattern, resulting in the loss of its rotational symmetry. Fig. 4(d) visualizes the actually carried pattern. More concretely, recall from Fig. 2, when the watermarked noise " + }, + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "type": "inline_equation", + "content": "X[u,v]" + }, + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "type": "text", + "content": " is transformed back to the spatial " + }, + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "type": "inline_equation", + "content": "x[m,n]" + }, + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "type": "text", + "content": ", only its real part " + }, + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "type": "inline_equation", + "content": "x_{\\mathrm{re}}[m,n]" + }, + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "type": "text", + "content": " is kept for image generation. From a frequency domain perspective, this is equivalent to generation with its conjugate symmetry " + }, + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "type": "inline_equation", + "content": "X_{\\mathrm{cs}}[u,v]" + }, + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "type": "text", + "content": ", which is not necessarily equal to " + }, + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "type": "inline_equation", + "content": "X[u,v]" + }, + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "type": "text", + "content": ". To leverage the rotational symmetry of the" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 270, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 270, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 270, + 91, + 447, + 102 + ], + "type": "text", + "content": "RingID: Enhanced Multi-Key Identification" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 201 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 201 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 201 + ], + "type": "text", + "content": "ring patterns, we attempt to eliminate the information loss during imprinting by solving " + }, + { + "bbox": [ + 130, + 116, + 482, + 201 + ], + "type": "inline_equation", + "content": "X[u,v] = X_{\\mathrm{cs}}[u,v]" + }, + { + "bbox": [ + 130, + 116, + 482, + 201 + ], + "type": "text", + "content": ". The results indicate that " + }, + { + "bbox": [ + 130, + 116, + 482, + 201 + ], + "type": "inline_equation", + "content": "X_{\\mathrm{re}}[u,v]" + }, + { + "bbox": [ + 130, + 116, + 482, + 201 + ], + "type": "text", + "content": " should be an even function and " + }, + { + "bbox": [ + 130, + 116, + 482, + 201 + ], + "type": "inline_equation", + "content": "X_{\\mathrm{im}}[u,v]" + }, + { + "bbox": [ + 130, + 116, + 482, + 201 + ], + "type": "text", + "content": " should be an odd function about the Fourier center. To satisfy these constraints, we imprint the tree-ring pattern only in the real part and leave the imaginary part blank. We also align the center of the tree-ring pattern with the Fourier center (32,32). Originally, the tree-ring pattern was put at (31,32) and imprinted on both the real and the imaginary." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 220, + 483, + 364 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 220, + 483, + 364 + ], + "spans": [ + { + "bbox": [ + 130, + 220, + 483, + 364 + ], + "type": "text", + "content": "Making A Rounder Ring Rings are originally determined by locating pixels equidistant from a give circle center " + }, + { + "bbox": [ + 130, + 220, + 483, + 364 + ], + "type": "inline_equation", + "content": "(x_0, y_0)" + }, + { + "bbox": [ + 130, + 220, + 483, + 364 + ], + "type": "text", + "content": ": " + }, + { + "bbox": [ + 130, + 220, + 483, + 364 + ], + "type": "inline_equation", + "content": "(x - x_0)^2 + (y - y_0)^2 = r^2" + }, + { + "bbox": [ + 130, + 220, + 483, + 364 + ], + "type": "text", + "content": ". We observed that this simple approach does not yield a well-rounded circular ring, particularly in images with lower resolutions like initial noise of shape " + }, + { + "bbox": [ + 130, + 220, + 483, + 364 + ], + "type": "inline_equation", + "content": "64 \\times 64" + }, + { + "bbox": [ + 130, + 220, + 483, + 364 + ], + "type": "text", + "content": ". Artifacts and alias are shown in Fig. 4(b). Although it is centrosymmetric, it lacks rotational symmetry. Well, the solution is quite straightforward: drawing a rounder ring. There are a number of tools that can help draw a much rounder ring. Here, we employ a very simple idea to get a rounder ring. We place a white pixel on a black background at a distance " + }, + { + "bbox": [ + 130, + 220, + 483, + 364 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 130, + 220, + 483, + 364 + ], + "type": "text", + "content": " from the rotation center. By rotating the low resolution image 360 degrees and recording the trajectory of this pixel, we obtain a rounder ring with a radius of " + }, + { + "bbox": [ + 130, + 220, + 483, + 364 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 130, + 220, + 483, + 364 + ], + "type": "text", + "content": ". Despite its simplicity, empirical results in Tab. 3 show that it is crucial for rotational invariance." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 384, + 394, + 397 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 384, + 394, + 397 + ], + "spans": [ + { + "bbox": [ + 131, + 384, + 394, + 397 + ], + "type": "text", + "content": "5.3 Discretization for Enhanced Distinguishability" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 407, + 482, + 538 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 407, + 482, + 538 + ], + "spans": [ + { + "bbox": [ + 130, + 407, + 482, + 538 + ], + "type": "text", + "content": "Tree-Ring samples values for each ring from a Gaussian distribution to keep consistency with the distribution of initial noise. We find that this random sampling makes it exceedingly challenging to distinguish between two keys. To address this issue, we propose a discretization approach, specifying that the value for each ring can only be either " + }, + { + "bbox": [ + 130, + 407, + 482, + 538 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 130, + 407, + 482, + 538 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 130, + 407, + 482, + 538 + ], + "type": "inline_equation", + "content": "-\\alpha" + }, + { + "bbox": [ + 130, + 407, + 482, + 538 + ], + "type": "text", + "content": ", illustrated in Fig. 4(a). Consequently, the watermark pattern for " + }, + { + "bbox": [ + 130, + 407, + 482, + 538 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 130, + 407, + 482, + 538 + ], + "type": "text", + "content": " rings has a capacity of " + }, + { + "bbox": [ + 130, + 407, + 482, + 538 + ], + "type": "inline_equation", + "content": "2^n" + }, + { + "bbox": [ + 130, + 407, + 482, + 538 + ], + "type": "text", + "content": ". While this seemingly reduces the upper limit of watermark capacity, it significantly enhances its valid capacity, ensuring optimal utilization of all available slots. Our experiments reveal that setting " + }, + { + "bbox": [ + 130, + 407, + 482, + 538 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 130, + 407, + 482, + 538 + ], + "type": "text", + "content": " as the standard deviation of the initial noise ensures substantial distinguishability between different keys while reducing the impact on generation quality to a small level." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 559, + 261, + 572 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 559, + 261, + 572 + ], + "spans": [ + { + "bbox": [ + 131, + 559, + 261, + 572 + ], + "type": "text", + "content": "5.4 Increasing Capacity" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "content": "We consider two simple approaches to increase capacity: adjusting the number of rings in a single channel or imprinting rings in extra channels. We evaluate the advantages and drawbacks of both methods in Sec. 6.4. In summary, increasing the number of rings " + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "inline_equation", + "content": "|R|" + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "content": " in a single channel can effectively boost capacity, leading to " + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "inline_equation", + "content": "2^{|R|}" + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "content": " keys. However, excessive ring count results in a noticeable decline in robustness and generation quality. Imprinting watermarks in multiple channels exponentially increases capacity by " + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "inline_equation", + "content": "n^{|c|}" + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "content": " is number of channels and " + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "content": " is" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 212, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 212, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 212, + 100 + ], + "type": "text", + "content": "H. Ci et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "content": "the capacity of a single channel. However, it's likely to introduce ring-like artifacts in the center of generated images. Therefore, we opt to place the watermark in a single channel and enhance capacity by adjusting the range of rings." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 167, + 223, + 181 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 167, + 223, + 181 + ], + "spans": [ + { + "bbox": [ + 131, + 167, + 223, + 181 + ], + "type": "text", + "content": "6 Experiment" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 188, + 261, + 201 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 188, + 261, + 201 + ], + "spans": [ + { + "bbox": [ + 131, + 188, + 261, + 201 + ], + "type": "text", + "content": "6.1 Experimental Setup" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 205, + 482, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 205, + 482, + 384 + ], + "spans": [ + { + "bbox": [ + 130, + 205, + 482, + 384 + ], + "type": "text", + "content": "Following Tree-Ring [38], we experiment with the state-of-the-art opensource diffusion model StableDiffusion-V2 [29]. By default, RingID imprints a tree-ring watermark with radius 3-14 on the channel 3 of the initial noise and a Gaussian noise watermark on channels 0. We adopt discretization (to 64 and -64), spatial shift, rounder drawing, and lossless imprinting as ordered in Fig. 2. We report the ROC-AUC between 1000 watermarked and 1000 unwatermarked images as the verification metric. In identification, we report identification accuracy with different numbers of keys. Evaluation is performed under various image distortions, following the original setup of Tree-Ring [38], i.e., JPEG 25, rotation " + }, + { + "bbox": [ + 130, + 205, + 482, + 384 + ], + "type": "inline_equation", + "content": "75^{\\circ}" + }, + { + "bbox": [ + 130, + 205, + 482, + 384 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 205, + 482, + 384 + ], + "type": "inline_equation", + "content": "75\\%" + }, + { + "bbox": [ + 130, + 205, + 482, + 384 + ], + "type": "text", + "content": " random C&S, blur kernel 8, noise std 0.1, brightness [0, 6]. We assess the generation quality with CLIP scores [28] by OpenCLIP-ViT/G [5] between the generated image and the prompts to measure text-image alignment, and Frechet Inception Distance (FID) [17] to quantify the similarity between generated and natural images. We generate 5000 images and evaluate FID on the COCO 2017 training dataset [21]. We set both generation and inversion steps to 50." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 398, + 297, + 410 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 398, + 297, + 410 + ], + "spans": [ + { + "bbox": [ + 131, + 398, + 297, + 410 + ], + "type": "text", + "content": "6.2 Comparison with Baselines" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 414, + 482, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 414, + 482, + 533 + ], + "spans": [ + { + "bbox": [ + 130, + 414, + 482, + 533 + ], + "type": "text", + "content": "Robustness In verification (Tab. 1), both RingID and Tree-Ring perfectly distinguish between watermarked and unwatermarked images when no attacks are performed, achieving AUC score of 1. However, RingID exhibit stronger robustness to various image distortions than Tree-Ring, achieving near-perfect average AUC score, even though it is not optimized for verification. In the more challenging identification task, (Tab. 2), RingID significantly outperforms Tree-Ring. When identifying 32 keys, Tree-Ring's average identification accuracy (excluding C&S) is 0.465, while RingID's accuracy reaches up to 0.992. As the number of keys increases to 2048, this gap widens drastically (0.077 v.s. 0.942), further highlighting RingID's superior robustness and larger valid key capacity." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 534, + 482, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 534, + 482, + 617 + ], + "spans": [ + { + "bbox": [ + 130, + 534, + 482, + 617 + ], + "type": "text", + "content": "Notably, both methods struggle with the C&S distortion in identification (Tab. 2). This aligns with our analysis in Sec. 4.2. Scaling in the spatial domain causes inverse scaling in the frequency domain, leading to pattern disruption and rendering identification challenging under such distortions. However, in verification, with the help of distribution shift, watermarked images can still be effectively distinguished from unwatermarked ones under C&S. Both Tree-Ring and RingID achieve average AUC higher than 0.97." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 617, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 617, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 131, + 617, + 482, + 665 + ], + "type": "text", + "content": "Image Quality RingID achieves strong performance without compromising the generation quality. It maintains comparable text-image alignments with TreeRing. Both methods achieve similar CLIP scores (0.364 v.s. 0.365). RingID obtains an FID of 26.13, which is also comparable to Tree-Ring's FID of 25.93." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 270, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 270, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 270, + 91, + 447, + 102 + ], + "type": "text", + "content": "RingID: Enhanced Multi-Key Identification" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 132, + 147, + 481, + 198 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 482, + 137 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 482, + 137 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 482, + 137 + ], + "type": "text", + "content": "Table 1: Comparison with Tree-Ring in the verification task. The table shows the ROC-AUC values under various image distortions and CLIP Scores." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 132, + 147, + 481, + 198 + ], + "lines": [ + { + "bbox": [ + 132, + 147, + 481, + 198 + ], + "spans": [ + { + "bbox": [ + 132, + 147, + 481, + 198 + ], + "type": "table", + "html": "
MethodsRing RadiusCleanRotateJPEGC&SBlurNoiseBrightnessAvgCLIP Score
Tree-Ring0-101.0000.9350.9990.9610.9990.9440.9830.9750.364
RingID0-101.0001.0001.0000.9790.9940.9690.9910.9900.359
3-141.0001.0001.0000.9870.9890.9980.9940.9950.365
", + "image_path": "a3d75da98c77a8370364d9ad9997c080fa184ae1f6dca297dfe93487c1d56b62.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 133, + 234, + 481, + 319 + ], + "blocks": [ + { + "bbox": [ + 132, + 201, + 481, + 224 + ], + "lines": [ + { + "bbox": [ + 132, + 201, + 481, + 224 + ], + "spans": [ + { + "bbox": [ + 132, + 201, + 481, + 224 + ], + "type": "text", + "content": "Table 2: Comparison with Tree-Ring in the identification task. We report the identification accuracy under various distortions for different numbers of keys." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 133, + 234, + 481, + 319 + ], + "lines": [ + { + "bbox": [ + 133, + 234, + 481, + 319 + ], + "spans": [ + { + "bbox": [ + 133, + 234, + 481, + 319 + ], + "type": "table", + "html": "
Methods#KeysCleanRotateJPEGC&SBlurNoiseBrightnessAvg\\(Avg_{noC&S}\\)
Tree-Ring320.7900.0200.4200.0400.6100.5300.4200.4040.465
1280.4500.0100.1200.0200.2800.2300.1700.1830.210
20480.2000.0000.0400.0000.0900.0700.0600.0660.077
RingID321.0001.0001.0000.5300.9901.0000.9600.9260.992
1281.0000.9801.0000.2800.9801.0000.9400.8830.983
20481.0000.8601.0000.0800.9700.9500.8700.8190.942
", + "image_path": "5155663580f91040a0e77e131db16eeaca572d4ebebf553d6856550321a05f85.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 342, + 238, + 355 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 342, + 238, + 355 + ], + "spans": [ + { + "bbox": [ + 132, + 342, + 238, + 355 + ], + "type": "text", + "content": "6.3 Ablation Study" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 362, + 481, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 362, + 481, + 434 + ], + "spans": [ + { + "bbox": [ + 130, + 362, + 481, + 434 + ], + "type": "text", + "content": "Components of RingID Tab. 3 ablates all proposed components of RingID. We can find that all designs are crucial for the final success of RingID. Spatial shift, lossless imprinting, and discretization play a significant role in the robustness of rotation. While the use of rounder rings also increase the accuracy by approximately " + }, + { + "bbox": [ + 130, + 362, + 481, + 434 + ], + "type": "inline_equation", + "content": "40\\%" + }, + { + "bbox": [ + 130, + 362, + 481, + 434 + ], + "type": "text", + "content": " (from 0.620 to 0.860). Notably, spatial shift sacrifices the robustness to blur and brightness attacks for rotation robustness." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 434, + 481, + 543 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 434, + 481, + 543 + ], + "spans": [ + { + "bbox": [ + 130, + 434, + 481, + 543 + ], + "type": "text", + "content": "Heterogeneous Watermarking RingID employs the combination of a Gaussian noise watermark and a ring watermark. Here, we ablate the choice of imprinting the noise watermark on different channels. By comparing the first 3 rows of Tab. 4, we find that heterogeneous watermarking can effectively combine the advantage from both the noise watermark and the ring watermark. Among single-channel configurations, imprinting noise watermark on channel 0 (0.819) achieves the best performance. The combination of channels 0 and 1 yields the highest accuracy among all trials. Interestingly, imprinting noise watermarks on all available channels (0, 1, 2) decrease the accuracy from 0.830 to 0.807." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 561, + 214, + 572 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 561, + 214, + 572 + ], + "spans": [ + { + "bbox": [ + 132, + 561, + 214, + 572 + ], + "type": "text", + "content": "6.4 Discussion" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 581, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 581, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 581, + 481, + 666 + ], + "type": "text", + "content": "Where is the Best Place to Put Rings? We could imprint all rings on one channel or distribute them evenly across multiple channels. In Tab. 5, we fix the total number of rings to 12 and distribute them to 1-4 channels. we can find that performance drops as we distribute rings to more channels. What's more, imprinting rings on multiple channels cause artifacts to appear more frequently (shown in Supplementary C). So given fixed capacity, it's better to imprint all rings on a single channel. This raises another question: what are the optimal" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 212, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 212, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 212, + 100 + ], + "type": "text", + "content": "H. Ci et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 133, + 147, + 481, + 232 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 480, + 137 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 480, + 137 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 480, + 137 + ], + "type": "text", + "content": "Table 3: Ablation study on the effect of each proposed module in RingID. We report the identification accuracy for 2048 keys when removing individual components." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 133, + 147, + 481, + 232 + ], + "lines": [ + { + "bbox": [ + 133, + 147, + 481, + 232 + ], + "spans": [ + { + "bbox": [ + 133, + 147, + 481, + 232 + ], + "type": "table", + "html": "
SettingsCleanRotateJPEGC&SBlurNoiseBrightnessAvg
RingID1.0000.8601.0000.0800.9700.9500.8700.819
- Spatial Shift1.0000.0001.0000.0500.9900.9300.9400.701
- Lossless Imprinting1.0000.0100.9700.1600.9500.9800.8300.700
- Rounder Ring1.0000.6200.9900.1000.8900.9700.8500.774
- Discretization0.9800.1200.3800.0200.4500.6500.3900.427
- Heterogeneous1.0000.8200.9400.0300.9600.7100.7200.740
", + "image_path": "48e0a0b89bf7d5ad6541983b190b5aabd0c68636e1c329d17792c1d8e2b6ce61.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 133, + 300, + 481, + 405 + ], + "blocks": [ + { + "bbox": [ + 132, + 257, + 480, + 291 + ], + "lines": [ + { + "bbox": [ + 132, + 257, + 480, + 291 + ], + "spans": [ + { + "bbox": [ + 132, + 257, + 480, + 291 + ], + "type": "text", + "content": "Table 4: Ablation study on multi-channel heterogeneous watermarking. This table reports the identification accuracy when the noise watermark is imprinted on different channels (abbr. CHs). “-” means the watermark is not imprinted." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 133, + 300, + 481, + 405 + ], + "lines": [ + { + "bbox": [ + 133, + 300, + 481, + 405 + ], + "spans": [ + { + "bbox": [ + 133, + 300, + 481, + 405 + ], + "type": "table", + "html": "
Ring CHsNoise CHsCleanRotateJPEGC&SBlurNoiseBrightnessAvg
-01.0000.0001.0000.1700.9900.9500.9200.719
3-1.0000.8200.9400.0300.9600.7100.7200.740
301.0000.8601.0000.0800.9700.9500.8700.819
311.0000.8201.0000.1301.0000.8800.7700.800
321.0000.8400.9800.1000.9200.9900.7900.803
30,11.0000.8500.9500.1500.9900.9900.8800.830
30,21.0000.8400.9900.1200.9500.9800.8700.821
30,1,21.0000.8300.9600.1600.9300.9800.7900.807
", + "image_path": "10678218a115fed06926b9c49683dca3c97a992f3ea374e84a9f95d8ba834953.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 442, + 480, + 478 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 442, + 480, + 478 + ], + "spans": [ + { + "bbox": [ + 132, + 442, + 480, + 478 + ], + "type": "text", + "content": "inner and outer radii if all rings are placed on a single channel? From Tab. 6 we can see that 3-14 forms a good choice for 11 rings with capacity 2048. The optimal ring radius also varies for different ring numbers." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 485, + 480, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 485, + 480, + 544 + ], + "spans": [ + { + "bbox": [ + 132, + 485, + 480, + 544 + ], + "type": "text", + "content": "More Keys We would like to explore whether RingID can handle a larger number of keys. Fig. 5 shows the average identification accuracy under various attacks " + }, + { + "bbox": [ + 132, + 485, + 480, + 544 + ], + "type": "inline_equation", + "content": "v.s." + }, + { + "bbox": [ + 132, + 485, + 480, + 544 + ], + "type": "text", + "content": " number of keys. We observe that RingID maintains good accuracy even with a large number of keys. And identification accuracy linearly decreases as the number of keys exponentially increase." + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 133, + 603, + 481, + 662 + ], + "blocks": [ + { + "bbox": [ + 133, + 580, + 479, + 592 + ], + "lines": [ + { + "bbox": [ + 133, + 580, + 479, + 592 + ], + "spans": [ + { + "bbox": [ + 133, + 580, + 479, + 592 + ], + "type": "text", + "content": "Table 6: Influence of ring radius on identification accuracy and text-image alignment." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 133, + 603, + 481, + 662 + ], + "lines": [ + { + "bbox": [ + 133, + 603, + 481, + 662 + ], + "spans": [ + { + "bbox": [ + 133, + 603, + 481, + 662 + ], + "type": "table", + "html": "
Radius#KeysAccCLIP ScoreRadius#KeysAccCLIP ScoreRadius#KeysAccCLIP Score
1-6320.8860.3631-92560.8190.3621-1220480.7860.361
2-7320.8760.3662-102560.8560.3582-1320480.8040.361
3-8320.8970.3653-112560.8500.3623-1420480.8190.368
4-9320.8970.3684-122560.8470.3634-1520480.8130.363
5-10320.8810.3615-132560.8390.3605-1620480.8030.360
", + "image_path": "0dd4abc8edaeba4494308ff7d018cbc9c5859adcadc80f1b86ede17b4e933dbe.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 270, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 270, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 270, + 91, + 447, + 102 + ], + "type": "text", + "content": "RingID: Enhanced Multi-Key Identification" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 133, + 179, + 298, + 243 + ], + "blocks": [ + { + "bbox": [ + 132, + 132, + 298, + 176 + ], + "lines": [ + { + "bbox": [ + 132, + 132, + 298, + 176 + ], + "spans": [ + { + "bbox": [ + 132, + 132, + 298, + 176 + ], + "type": "text", + "content": "Table 5: Distributing 12 rings across different number of channels (#CHs). This table reports the average identification accuracy and CLIP Score." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 133, + 179, + 298, + 243 + ], + "lines": [ + { + "bbox": [ + 133, + 179, + 298, + 243 + ], + "spans": [ + { + "bbox": [ + 133, + 179, + 298, + 243 + ], + "type": "table", + "html": "
Ring Radius #CHsAvgCLIP Score
3-1510.7400.357
3-920.7370.351
3-730.6640.351
3-640.5440.340
", + "image_path": "35d36e710a17cdaa343af00aebce0fbd5a40b4fb01e59ca0244dd9dae0c0dbd0.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 326, + 130, + 475, + 237 + ], + "blocks": [ + { + "bbox": [ + 326, + 130, + 475, + 237 + ], + "lines": [ + { + "bbox": [ + 326, + 130, + 475, + 237 + ], + "spans": [ + { + "bbox": [ + 326, + 130, + 475, + 237 + ], + "type": "image", + "image_path": "dc91f09a8e3332f8fef9acaaed2e773f7422009913d4fd33282cd11ca1654349.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 322, + 238, + 481, + 259 + ], + "lines": [ + { + "bbox": [ + 322, + 238, + 481, + 259 + ], + "spans": [ + { + "bbox": [ + 322, + 238, + 481, + 259 + ], + "type": "text", + "content": "Fig. 5: Average identification accuracy " + }, + { + "bbox": [ + 322, + 238, + 481, + 259 + ], + "type": "inline_equation", + "content": "v.s." + }, + { + "bbox": [ + 322, + 238, + 481, + 259 + ], + "type": "text", + "content": " number of keys." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 148, + 275, + 470, + 407 + ], + "blocks": [ + { + "bbox": [ + 148, + 275, + 470, + 407 + ], + "lines": [ + { + "bbox": [ + 148, + 275, + 470, + 407 + ], + "spans": [ + { + "bbox": [ + 148, + 275, + 470, + 407 + ], + "type": "image", + "image_path": "52bb288e0054b399f62eb9e2ebab2ee3e14701a8ddaa8ee532e2a35388fdd284.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 419, + 482, + 442 + ], + "lines": [ + { + "bbox": [ + 132, + 419, + 482, + 442 + ], + "spans": [ + { + "bbox": [ + 132, + 419, + 482, + 442 + ], + "type": "text", + "content": "Fig. 6: Qualitative results. The subtle changes in image layouts show that both TreeRing and RingID embed watermarks into the image's semantic content." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 466, + 258, + 478 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 466, + 258, + 478 + ], + "spans": [ + { + "bbox": [ + 132, + 466, + 258, + 478 + ], + "type": "text", + "content": "6.5 Qualitative Results" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 487, + 482, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 487, + 482, + 548 + ], + "spans": [ + { + "bbox": [ + 131, + 487, + 482, + 548 + ], + "type": "text", + "content": "As shown in Fig. 6, both RingID and Tree-Ring watermarking can generate images of good visual quality. Unlike low-level watermarking techniques, both RingID and Tree-Ring cause subtle changes in the image layout. However, RingID generates images with layout that is closer to the unwatermarked image. We owe this to avoiding watermarking on the low-frequency part (radius 0-3)." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 567, + 312, + 580 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 567, + 312, + 580 + ], + "spans": [ + { + "bbox": [ + 132, + 567, + 312, + 580 + ], + "type": "text", + "content": "7 Conclusion and Limitation" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 594, + 481, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 481, + 628 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 481, + 628 + ], + "type": "text", + "content": "We conduct a comprehensive investigation on Tree-Ring Watermarking, revealing its limitations in identifying multiple keys. We then propose RingID to enhance identification capability." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 131, + 629, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 629, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 131, + 629, + 481, + 665 + ], + "type": "text", + "content": "Limitation: Under multi-key identification, one limitation is that both TreeRing and our RingID are vulnerable to cropping and scaling attacks. Future work may explore other transform domains to resist these attacks." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 212, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 212, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 212, + 100 + ], + "type": "text", + "content": "H. Ci et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 480, + 152 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 480, + 152 + ], + "type": "text", + "content": "Acknowledgment: This project is supported by the Ministry of Education, Singapore, under the Academic Research Fund Tier 1 (FY2022) Award 22-5406-A0001." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 172, + 197, + 185 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 172, + 197, + 185 + ], + "spans": [ + { + "bbox": [ + 133, + 172, + 197, + 185 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 198, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 138, + 198, + 480, + 232 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 198, + 480, + 232 + ], + "spans": [ + { + "bbox": [ + 138, + 198, + 480, + 232 + ], + "type": "text", + "content": "1. An, B., Ding, M., Rabbani, T., Agrawal, A., Xu, Y., Deng, C., Zhu, S., Mohamed, A., Wen, Y., Goldstein, T., et al.: Benchmarking the robustness of image watermarks. arXiv preprint arXiv:2401.08573 (2024)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 233, + 480, + 277 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 233, + 480, + 277 + ], + "spans": [ + { + "bbox": [ + 138, + 233, + 480, + 277 + ], + "type": "text", + "content": "2. Bansal, A., Chiang, P.y., Curry, M.J., Jain, R., Wigington, C., Manjunatha, V., Dickerson, J.P., Goldstein, T.: Certified neural network watermarks with randomized smoothing. In: International Conference on Machine Learning. pp. 1450-1465. PMLR (2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 277, + 480, + 310 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 277, + 480, + 310 + ], + "spans": [ + { + "bbox": [ + 138, + 277, + 480, + 310 + ], + "type": "text", + "content": "3. Betker, J., Goh, G., Jing, L., Brooks, T., Wang, J., Li, L., Ouyang, L., Zhuang, J., Lee, J., Guo, Y., et al.: Improving image generation with better captions. Computer Science. https://cdn.opencai.com/papers/dall-e-3.pdf 2(3), 8 (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 311, + 480, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 311, + 480, + 354 + ], + "spans": [ + { + "bbox": [ + 138, + 311, + 480, + 354 + ], + "type": "text", + "content": "4. Blattmann, A., Dockhorn, T., Kulal, S., Mendelevitch, D., Kilian, M., Lorenz, D., Levi, Y., English, Z., Voleti, V., Letts, A., et al.: Stable video diffusion: Scaling latent video diffusion models to large datasets. arXiv preprint arXiv:2311.15127 (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 355, + 480, + 399 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 355, + 480, + 399 + ], + "spans": [ + { + "bbox": [ + 138, + 355, + 480, + 399 + ], + "type": "text", + "content": "5. Cherti, M., Beaumont, R., Wightman, R., Wortsman, M., Ilharco, G., Gordon, C., Schuhmann, C., Schmidt, L., Jitsev, J.: Reproducible scaling laws for contrastive language-image learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2818-2829 (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 399, + 480, + 422 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 399, + 480, + 422 + ], + "spans": [ + { + "bbox": [ + 138, + 399, + 480, + 422 + ], + "type": "text", + "content": "6. Ci, H., Song, Y., Yang, P., Xie, J., Shou, M.Z.: Wmadapter: Adding watermark control to latent diffusion models. arXiv preprint arXiv:2406.08337 (2024)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 422, + 480, + 454 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 422, + 480, + 454 + ], + "spans": [ + { + "bbox": [ + 138, + 422, + 480, + 454 + ], + "type": "text", + "content": "7. Ci, H., Wang, C., Ma, X., Wang, Y.: Optimizing network structure for 3d human pose estimation. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 2262-2271 (2019)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 455, + 480, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 455, + 480, + 487 + ], + "spans": [ + { + "bbox": [ + 138, + 455, + 480, + 487 + ], + "type": "text", + "content": "8. Ci, H., Wu, M., Zhu, W., Ma, X., Dong, H., Zhong, F., Wang, Y.: Gfpose: Learning 3d human pose prior with gradient fields. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4800-4810 (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 488, + 480, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 488, + 480, + 510 + ], + "spans": [ + { + "bbox": [ + 138, + 488, + 480, + 510 + ], + "type": "text", + "content": "9. Cox, I.: Digital watermarking and steganography. Morgan Kaufmann google scholar 2, 893-914 (2007)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 510, + 480, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 510, + 480, + 533 + ], + "spans": [ + { + "bbox": [ + 138, + 510, + 480, + 533 + ], + "type": "text", + "content": "0. Cui, S., Guo, X., Zhang, Z.: Estimation and inference in ultrahigh dimensional partially linear single-index models. arXiv preprint arXiv:2404.04471 (2024)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 533, + 480, + 565 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 533, + 480, + 565 + ], + "spans": [ + { + "bbox": [ + 138, + 533, + 480, + 565 + ], + "type": "text", + "content": "1. Cui, S., Sudjianto, A., Zhang, A., Li, R.: Enhancing robustness of gradientboosted decision trees through one-hot encoding and regularization. arXiv preprint arXiv:2304.13761 (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 566, + 480, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 566, + 480, + 588 + ], + "spans": [ + { + "bbox": [ + 138, + 566, + 480, + 588 + ], + "type": "text", + "content": "2. Dhariwal, P., Nichol, A.: Diffusion models beat gans on image synthesis. Advances in neural information processing systems 34, 8780-8794 (2021)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 589, + 480, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 589, + 480, + 621 + ], + "spans": [ + { + "bbox": [ + 138, + 589, + 480, + 621 + ], + "type": "text", + "content": "3. Fernandez, P., Couairon, G., Jégou, H., Douze, M., Furon, T.: The stable signature: Rooting watermarks in latent diffusion models. arXiv preprint arXiv:2303.15435 (2023)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 621, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 621, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 621, + 480, + 665 + ], + "type": "text", + "content": "4. Fernandez, P., Sablayrolles, A., Furon, T., Jégou, H., Douze, M.: Watermarking images in self-supervised latent spaces. In: ICASSP 2022-2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). pp. 3054-3058. IEEE (2022)" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 270, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 270, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 270, + 91, + 447, + 102 + ], + "type": "text", + "content": "RingID: Enhanced Multi-Key Identification" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "type": "text", + "content": "15. Guo, H., Zhang, Q., Luo, J., Guo, F., Zhang, W., Su, X., Li, M.: Practical deep dispersed watermarking with synchronization and fusion. In: Proceedings of the 31st ACM International Conference on Multimedia. pp. 7922-7932 (2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 150, + 481, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 150, + 481, + 194 + ], + "spans": [ + { + "bbox": [ + 133, + 150, + 481, + 194 + ], + "type": "text", + "content": "16. Guo, Y., Yang, C., Rao, A., Liang, Z., Wang, Y., Qiao, Y., Agrawala, M., Lin, D., Dai, B.: Animatediff:Animate your personalized text-to-image diffusion models without specific tuning. International Conference on Learning Representations (2024)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 195, + 481, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 195, + 481, + 228 + ], + "spans": [ + { + "bbox": [ + 132, + 195, + 481, + 228 + ], + "type": "text", + "content": "17. Heusel, M., Ramsauer, H., Unterthiner, T., Nessler, B., Hochreiter, S.: Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems 30 (2017)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 228, + 481, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 228, + 481, + 251 + ], + "spans": [ + { + "bbox": [ + 132, + 228, + 481, + 251 + ], + "type": "text", + "content": "18. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. Advances in neural information processing systems 33, 6840-6851 (2020)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 251, + 481, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 251, + 481, + 274 + ], + "spans": [ + { + "bbox": [ + 132, + 251, + 481, + 274 + ], + "type": "text", + "content": "19. Kong, Z., Ping, W., Huang, J., Zhao, K., Catanzaro, B.: Diffwave: A versatile diffusion model for audio synthesis. arXiv preprint arXiv:2009.09761 (2020)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 274, + 481, + 307 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 274, + 481, + 307 + ], + "spans": [ + { + "bbox": [ + 132, + 274, + 481, + 307 + ], + "type": "text", + "content": "20. Lin, C.Y., Wu, M., Bloom, J.A., Cox, I.J., Miller, M.L., Lui, Y.M.: Rotation, scale, and translation resilient watermarking for images. IEEE Transactions on image processing 10(5), 767-782 (2001)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 308, + 481, + 351 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 308, + 481, + 351 + ], + "spans": [ + { + "bbox": [ + 132, + 308, + 481, + 351 + ], + "type": "text", + "content": "21. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollar, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13. pp. 740-755. Springer (2014)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 352, + 481, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 352, + 481, + 385 + ], + "spans": [ + { + "bbox": [ + 132, + 352, + 481, + 385 + ], + "type": "text", + "content": "22. Liu, R., Wu, R., Van Hoorick, B., Tokmakov, P., Zakharov, S., Vondrick, C.: Zero-1-to-3: Zero-shot one image to 3d object. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 9298-9309 (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 386, + 481, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 386, + 481, + 407 + ], + "spans": [ + { + "bbox": [ + 132, + 386, + 481, + 407 + ], + "type": "text", + "content": "23. Liu, Y., Li, Z., Backes, M., Shen, Y., Zhang, Y.: Watermarking diffusion model. arXiv preprint arXiv:2305.12502 (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 408, + 481, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 408, + 481, + 441 + ], + "spans": [ + { + "bbox": [ + 132, + 408, + 481, + 441 + ], + "type": "text", + "content": "24. Luo, X., Li, Y., Chang, H., Liu, C., Milanfar, P., Yang, F.: Dvmark: a deep multiscale framework for video watermarking. IEEE Transactions on Image Processing (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 441, + 419, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 441, + 419, + 453 + ], + "spans": [ + { + "bbox": [ + 132, + 441, + 419, + 453 + ], + "type": "text", + "content": "25. Midjourney.com: Midjourney. https://www.midjourney.com/home" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 453, + 348, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 453, + 348, + 464 + ], + "spans": [ + { + "bbox": [ + 132, + 453, + 348, + 464 + ], + "type": "text", + "content": "26. OpenAI: Dalle2. https://openai.com/dall-e-2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 465, + 481, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 465, + 481, + 487 + ], + "spans": [ + { + "bbox": [ + 132, + 465, + 481, + 487 + ], + "type": "text", + "content": "27. Poole, B., Jain, A., Barron, J.T., Mildenhall, B.: Dreamfusion: Text-to-3d using 2d diffusion. arXiv preprint arXiv:2209.14988 (2022)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 487, + 481, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 487, + 481, + 531 + ], + "spans": [ + { + "bbox": [ + 132, + 487, + 481, + 531 + ], + "type": "text", + "content": "28. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 532, + 481, + 575 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 532, + 481, + 575 + ], + "spans": [ + { + "bbox": [ + 132, + 532, + 481, + 575 + ], + "type": "text", + "content": "29. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 10684-10695 (June 2022)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 576, + 481, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 576, + 481, + 609 + ], + "spans": [ + { + "bbox": [ + 132, + 576, + 481, + 609 + ], + "type": "text", + "content": "30. Singer, U., Polyak, A., Hayes, T., Yin, X., An, J., Zhang, S., Hu, Q., Yang, H., Ashual, O., Gafni, O., et al.: Make-a-video: Text-to-video generation without text-video data. arXiv preprint arXiv:2209.14792 (2022)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 610, + 481, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 610, + 481, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 610, + 481, + 632 + ], + "type": "text", + "content": "31. Song, J., Meng, C., Ermon, S.: Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502 (2020)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 632, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 481, + 665 + ], + "type": "text", + "content": "32. Song, Y., Sohl-Dickstein, J., Kingma, D.P., Kumar, A., Ermon, S., Poole, B.: Score-based generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456 (2020)" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 212, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 212, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 212, + 100 + ], + "type": "text", + "content": "H. Ci et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 660 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "text", + "content": "33. Song, Y., Huang, S., Yao, C., Ye, X., Ci, H., Liu, J., Zhang, Y., Shou, M.Z.: Processpainter: Learn painting process from sequence data. arXiv preprint arXiv:2406.06062 (2024)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 149, + 482, + 181 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 149, + 482, + 181 + ], + "spans": [ + { + "bbox": [ + 130, + 149, + 482, + 181 + ], + "type": "text", + "content": "34. Steinberg, Y., Merhav, N.: Identification in the presence of side information with application to watermarking. IEEE Transactions on Information Theory 47(4), 1410-1422 (2001)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 181, + 482, + 213 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 181, + 482, + 213 + ], + "spans": [ + { + "bbox": [ + 130, + 181, + 482, + 213 + ], + "type": "text", + "content": "35. Tancik, M., Mildenhall, B., Ng, R.: Stegastamp: Invisible hyperlinks in physical photographs. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 2117-2126 (2020)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 213, + 482, + 235 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 213, + 482, + 235 + ], + "spans": [ + { + "bbox": [ + 130, + 213, + 482, + 235 + ], + "type": "text", + "content": "36. Tevet, G., Raab, S., Gordon, B., Shafir, Y., Cohen-Or, D., Bermano, A.H.: Human motion diffusion model. arXiv preprint arXiv:2209.14916 (2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 235, + 482, + 266 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 235, + 482, + 266 + ], + "spans": [ + { + "bbox": [ + 130, + 235, + 482, + 266 + ], + "type": "text", + "content": "37. Uchida, Y., Nagai, Y., Sakazawa, S., Satoh, S.: Embedding watermarks into deep neural networks. In: Proceedings of the 2017 ACM on international conference on multimedia retrieval. pp. 269-277 (2017)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 266, + 482, + 298 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 266, + 482, + 298 + ], + "spans": [ + { + "bbox": [ + 130, + 266, + 482, + 298 + ], + "type": "text", + "content": "38. Wen, Y., Kirchenbauer, J., Geiping, J., Goldstein, T.: Tree-ring watermarks: Fingerprints for diffusion images that are invisible and robust. arXiv preprint arXiv:2305.20030 (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 298, + 482, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 298, + 482, + 342 + ], + "spans": [ + { + "bbox": [ + 130, + 298, + 482, + 342 + ], + "type": "text", + "content": "39. Wu, J.Z., Ge, Y., Wang, X., Lei, S.W., Gu, Y., Shi, Y., Hsu, W., Shan, Y., Qie, X., Shou, M.Z.: Tune-a-video: One-shot tuning of image diffusion models for text-to-video generation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 7623-7633 (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 342, + 482, + 362 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 342, + 482, + 362 + ], + "spans": [ + { + "bbox": [ + 130, + 342, + 482, + 362 + ], + "type": "text", + "content": "40. Xia, X.G., Boncelet, C.G., Arce, G.R.: Wavelet transform based watermark for digital images. Optics Express 3(12), 497-511 (1998)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 362, + 482, + 394 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 362, + 482, + 394 + ], + "spans": [ + { + "bbox": [ + 130, + 362, + 482, + 394 + ], + "type": "text", + "content": "41. Xiong, C., Qin, C., Feng, G., Zhang, X.: Flexible and secure watermarking for latent diffusion model. In: Proceedings of the 31st ACM International Conference on Multimedia. pp. 1668-1676 (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 394, + 482, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 394, + 482, + 415 + ], + "spans": [ + { + "bbox": [ + 130, + 394, + 482, + 415 + ], + "type": "text", + "content": "42. Yang, P., Ci, H., Song, Y., Shou, M.Z.: Steganalysis on digital watermarking: Is your defense truly impervious? arXiv preprint arXiv:2406.09026 (2024)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 415, + 482, + 447 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 415, + 482, + 447 + ], + "spans": [ + { + "bbox": [ + 130, + 415, + 482, + 447 + ], + "type": "text", + "content": "43. Yu, N., Skripniuk, V., Abdelnabi, S., Fritz, M.: Artificial fingerprinting for generative models: Rooting deepfake attribution in training data. In: Proceedings of the IEEE/CVF International conference on computer vision. pp. 14448-14457 (2021)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 447, + 482, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 447, + 482, + 479 + ], + "spans": [ + { + "bbox": [ + 130, + 447, + 482, + 479 + ], + "type": "text", + "content": "44. Zhang, D.J., Wu, J.Z., Liu, J.W., Zhao, R., Ran, L., Gu, Y., Gao, D., Shou, M.Z.: Show-1: Marrying pixel and latent diffusion models for text-to-video generation. arXiv preprint arXiv:2309.15818 (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 479, + 482, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 479, + 482, + 521 + ], + "spans": [ + { + "bbox": [ + 130, + 479, + 482, + 521 + ], + "type": "text", + "content": "45. Zhang, J., Gu, Z., Jang, J., Wu, H., Stoecklin, M.P., Huang, H., Molloy, I.: Protecting intellectual property of deep neural networks with watermarking. In: Proceedings of the 2018 on Asia conference on computer and communications security. pp. 159-172 (2018)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 521, + 482, + 543 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 482, + 543 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 482, + 543 + ], + "type": "text", + "content": "46. Zhang, K.A., Xu, L., Cuesta-Infante, A., Veeramachaneni, K.: Robust invisible video watermarking with attention. arXiv preprint arXiv:1909.01285 (2019)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 543, + 482, + 574 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 543, + 482, + 574 + ], + "spans": [ + { + "bbox": [ + 130, + 543, + 482, + 574 + ], + "type": "text", + "content": "47. Zhao, X., Zhang, K., Wang, Y.X., Li, L.: Generative autoencoders as watermark attackers: Analyses of vulnerabilities and threats. arXiv preprint arXiv:2306.01953 (2023)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 130, + 574, + 482, + 595 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 574, + 482, + 595 + ], + "spans": [ + { + "bbox": [ + 130, + 574, + 482, + 595 + ], + "type": "text", + "content": "48. Zhao, Y., Pang, T., Du, C., Yang, X., Cheung, N.M., Lin, M.: A recipe for watermarking diffusion models. arXiv preprint arXiv:2303.10137 (2023)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 130, + 595, + 482, + 628 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 595, + 482, + 628 + ], + "spans": [ + { + "bbox": [ + 130, + 595, + 482, + 628 + ], + "type": "text", + "content": "49. Zhu, J., Kaplan, R., Johnson, J., Fei-Fei, L.: Hidden: Hiding data with deep networks. In: Proceedings of the European conference on computer vision (ECCV). pp. 657-672 (2018)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 130, + 628, + 482, + 660 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 628, + 482, + 660 + ], + "spans": [ + { + "bbox": [ + 130, + 628, + 482, + 660 + ], + "type": "text", + "content": "50. Zhu, W., Ma, X., Ro, D., Ci, H., Zhang, J., Shi, J., Gao, F., Tian, Q., Wang, Y.: Human motion generation: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence (2023)" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 270, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 270, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 270, + 91, + 447, + 102 + ], + "type": "text", + "content": "RingID: Enhanced Multi-Key Identification" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Rethinking Unsupervised Outlier Detection via Multiple Thresholding/7bb2035e-e079-470b-b56b-69d7d8631ccd_content_list.json b/2024/Rethinking Unsupervised Outlier Detection via Multiple Thresholding/7bb2035e-e079-470b-b56b-69d7d8631ccd_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..af54d0fac8fb428d8b608193d401b36b308c5a61 --- /dev/null +++ b/2024/Rethinking Unsupervised Outlier Detection via Multiple Thresholding/7bb2035e-e079-470b-b56b-69d7d8631ccd_content_list.json @@ -0,0 +1,2061 @@ +[ + { + "type": "text", + "text": "Rethinking Unsupervised Outlier Detection via Multiple Thresholding", + "text_level": 1, + "bbox": [ + 230, + 140, + 772, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zhonghang Liu1, Panzhong Lu2, Guoyang Xie3,4, Zhichao Lu4, and Wen-Yan Lin1", + "bbox": [ + 225, + 212, + 777, + 244 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Singapore Management University $^{2}$ Westlake University", + "bbox": [ + 379, + 255, + 625, + 282 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3 Contemporary Amperex Technology Co., Limited", + "bbox": [ + 328, + 282, + 673, + 299 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{4}$ City University of Hong Kong", + "bbox": [ + 395, + 297, + 609, + 311 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "zhliu.2020@phdcs.smu.edu.sg", + "bbox": [ + 395, + 311, + 607, + 325 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "lupanzhong@westlake.edu.cn", + "bbox": [ + 398, + 327, + 602, + 339 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "guoyang.xie@ieee.org", + "bbox": [ + 421, + 340, + 581, + 354 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "luzhichao.cn@gmail.com", + "bbox": [ + 419, + 354, + 583, + 367 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "daniellin@smu.edu.sg", + "bbox": [ + 421, + 369, + 580, + 381 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. In the realm of unsupervised image outlier detection, assigning outlier scores holds greater significance than its subsequent task: thresholding for predicting labels. This is because determining the optimal threshold on non-separable outlier score functions is an ill-posed problem. However, the lack of predicted labels not only hinders some real applications of current outlier detectors but also causes these methods not to be enhanced by leveraging the dataset's self-supervision. To advance existing scoring methods, we propose a multiple thresholding (Multi-T) module. It generates two thresholds that isolate inliers and outliers from the unlabelled target dataset, whereas outliers are employed to obtain better feature representation while inliers provide an uncontaminated manifold. Extensive experiments verify that Multi-T can significantly improve proposed outlier scoring methods. Moreover, Multi-T contributes to a naive distance-based method being state-of-the-art. Code is available at: https://github.com/zhliu-uod/Multi-T.", + "bbox": [ + 259, + 429, + 743, + 638 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Unsupervised outlier detection, multiple thresholding, outlier scoring", + "bbox": [ + 259, + 651, + 740, + 680 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 710, + 375, + 726 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Which is more important for the unsupervised outlier detection (UOD), outlier score or label? Currently, the mainstream UOD approaches [23, 28, 31, 56] focus on the first, i.e., learning a discriminative outlier score function. However, those recent outlier detectors are usually complex and non-iterative since the absence of predicted labels limits these methods to be further advanced with the dataset's self-supervision. Such that, this work tends to label the unlabelled tar", + "bbox": [ + 212, + 750, + 787, + 840 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/9080d1e2120f51744f2d3474d25e21c9a67b27c903f68baf00619bd1eb1b6ef1.jpg", + "image_caption": [ + "Fig. 1: Unlike the default UOD paradigm (a) concerned about learning a discriminative score function, our perspective (b) is to explore the dataset's prior knowledge via thresholding, to advance the previously proposed method. (c) illustrates the significant improvement of DeepSVDD [50] with Multi-T module." + ], + "image_footnote": [], + "bbox": [ + 215, + 143, + 787, + 239 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "get dataset, thereby leveraging the dataset's prior knowledge to enhance simple outlier scoring methods $^{1}$ , achieving state-of-the-art (SOTA) results efficiently.", + "bbox": [ + 212, + 338, + 782, + 368 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To understand the importance of the target dataset's self-supervision, we first decompose the predominant image outlier scoring approaches [9,32,50,51] into two continuous stages: inlier/normal manifold learning and distance/similarity inference. There is no doubt that identifying inliers will contribute to learning an uncontaminated inlier manifold (e.g. the hypersphere of DeepSVDD [50]). Additionally, motivated by Shell Theory [32,33] in the high-dimensional space, image feature representation for distance computation can be significantly improved via outliers-based Shell Normalization [32]. Such that it is critical to obtain the thresholds that separate both inliers and outliers from target datasets.", + "bbox": [ + 212, + 369, + 784, + 503 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To align with our above objective, we present a Multi-T module, which follows a multiple thresholding mechanism. Compared with conventional single thresholding methods [19, 24, 48, 49], Multi-T generates two distinct thresholds that separately isolate outliers and inliers from the target dataset, which is capable of dealing with the inevitable inseparability of the initial given score distribution. This allows us to directly implement Multi-T on a simple outlier score function.", + "bbox": [ + 212, + 506, + 784, + 595 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "By understanding the impact of the outlier ratio, Multi-T involves two stages: Uncontaminated inliers. We employ an iterative two-step process (i) identifying a noisy inlier distribution through the analysis of sorted initial score function, i.e., Ergodic-set normalized [34] distance to the mean of target dataset; (ii) subsequently filtering outliers with 3-sigma rule [42]. When converged, we can identify uncontaminated inliers, thereby training an inlier/normal manifold.", + "bbox": [ + 212, + 598, + 784, + 688 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Adaptive outliers. The uncontaminated inliers identification process also outputs a series of outlier threshold candidates. Subsequently, we compare the ranking-index similarity between Ergodic-set Normalization [34] and Shell Normalization [32] on the initial outlier score function. The structural consistency and contrastive properties of the two normalization procedures motivate us to implicitly estimate a rough outlier ratio, thereby detecting the adaptive outliers.", + "bbox": [ + 212, + 690, + 784, + 780 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The Multi-T module is training-free, highly efficient and grants the adaptiveness to choose thresholds suited to downstream implementations. As evaluated", + "bbox": [ + 212, + 781, + 784, + 811 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "Z. Liu et al.", + "bbox": [ + 271, + 114, + 352, + 126 + ], + "page_idx": 1 + }, + { + "type": "page_footnote", + "text": "1 Outlier scoring methods refers to the classical outlier detectors.", + "bbox": [ + 217, + 824, + 653, + 839 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "with comprehensive experimental settings, Multi-T helps to boost significant performance (efficacy and efficiency) improvement. For example, the AUC score of DeepSVDD [50] is improved from 0.622 to 0.925 on STL-10 dataset [15]. Moreover, the naive distance-to-the-target dataset's mean [32] metric integrated with Multi-T achieves SOTA results with only 1.2 second consumption for 10,000 ResNet-50 [21] samples that are orders of magnitude faster than baselines. Our primary contributions are summarized below:", + "bbox": [ + 212, + 146, + 782, + 251 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We introduce a novel perspective for UOD about enhancing previously proposed scoring methods via thresholding on a simple initial outlier score function.", + "- We present a multiple thresholding (Multi-T) module that generates two different thresholds to separate both inliers and outliers from the target dataset.", + "- The efficacy of integrated scoring methods can be significantly improved without the external complexity increase, as evaluated in extensive experiments." + ], + "bbox": [ + 212, + 252, + 782, + 340 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 214, + 367, + 387, + 382 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.1 Unsupervised Outlier Detection", + "text_level": 1, + "bbox": [ + 214, + 400, + 524, + 416 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Classic UOD task aims to assign an outlier score/likelihood to an image sample. The recent models can be divided into inlier-manifold learning and outlier exposure. Inlier-manifold learning assumes that inliers are the majority. Deep learning-based outlier detectors, as illustrated in studies [14, 28, 30, 38, 56, 58, 62, 63], typically focus on reducing the dimensionality of image data by projecting it into a latent and discriminative space. Additionally, various statistics-based methods model the inlier manifold using discrimination-based [32, 34] or density-based approaches [11, 26, 31, 40]. In applications where outliers are not the minority [13, 57], different approaches are required. Some outlier exposure methods [22, 36] employ out-of-distribution (OOD) data to train networks for detecting unseen OOD samples. For example, Shell-Renormalization [32] employs predicted outlier candidates to iteratively refine feature representation. It is particularly useful in scenarios where classic manifold learners, assuming outliers as a minority, may not be effective. However, these two paradigms are separable, this work will attempt to design a thresholding framework that generates both inliers and outliers, leveraging the benefits of both. Inlier-manifold learning and outlier exposure.", + "bbox": [ + 212, + 426, + 787, + 683 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2 Thresholding", + "text_level": 1, + "bbox": [ + 214, + 707, + 372, + 723 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Some traditional outlier detection methods [32, 35, 52], provide both the outlier score and its corresponding threshold concurrently. In this context, the threshold manifests as a hypersphere that accepts the inliers while rejecting those outliers. Unquestionably, discerning it presents its challenges in high-dimensional space [34]. In response, we venture into an alternate perspective. Some other threshold detectors are mostly based on statistical analysis that can be utilized for any given $1 - d$ outlier score function [28, 31, 32]. The representative", + "bbox": [ + 212, + 733, + 785, + 839 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "Rethinking Unsupervised Outlier Detection via Multiple Thresholding", + "bbox": [ + 264, + 114, + 730, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "works involve Kernel-based [43, 48]; Curve-based [17]; Normality-based: [5, 10]; Filtering-based [19, 55]; Statistical-based [1, 3, 4, 6-8, 11, 12, 16, 24, 37, 49] and Transformation-based [25, 46]. In practice, the default thresholding follows a single thresholding paradigm is not realistic since the perfect discriminativeness of the outlier score function is usually not assumed. Therefore, this work introduces a new perspective: multiple thresholding.", + "bbox": [ + 212, + 146, + 787, + 238 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3 Overview", + "text_level": 1, + "bbox": [ + 215, + 256, + 346, + 273 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1 Problem Definition", + "text_level": 1, + "bbox": [ + 215, + 285, + 421, + 299 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Before introducing our method, we first formally define the problem. The target dataset $\\mathbb{Z} = \\mathbb{Z}_{in}\\cup \\mathbb{Z}_{out}$ involves $n$ unlabelled images, here $\\mathbb{Z}_{in}$ and $\\mathbb{Z}_{out}$ represent inliers (normal data) and outliers (anomalous data), respectively. The outlier (contamination) ratio $\\gamma \\in (0,1)$ is denoted as $\\frac{\\#\\mathbb{Z}_{out}}{\\#\\mathbb{Z}_{in} + \\#\\mathbb{Z}_{out}}$ . Notably, this setup differs from the related semi-supervised outlier detection task by eliminating the need for a prior train/test split for the target dataset. Aligning with previous works [28, 32, 34], the raw images are extracted to feature representation $\\mathbf{X} = \\mathbf{X}_{\\mathrm{in}}\\cup \\mathbf{X}_{\\mathrm{out}} = \\{\\mathbf{x}_i\\}_{i = 1}^n$ with pixel representation [32, 34] or pre-trained neural networks (e.g., ResNet [21] and CLIP [45]). The main objective of UOD is to create an outlier score function $\\mathrm{F}(\\cdot)$ to evaluate the likelihood of an image feature $\\mathbf{x}_i\\in \\mathbf{X}$ being outliers:", + "bbox": [ + 212, + 306, + 787, + 470 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {Y} (\\mathbf {x} _ {i}) = \\left\\{ \\begin{array}{l} 0, \\mathrm {F} (\\mathbf {x} _ {i}) < \\phi \\\\ 1, \\mathrm {F} (\\mathbf {x} _ {i}) \\geq \\phi , \\end{array} \\right. \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 416, + 473, + 787, + 503 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathrm{Y} = 0$ (inlier)/1 (outlier) refers to predicted labels and $\\phi$ is the threshold (decision boundary). In this study, we not only center on evaluating the ranking accuracy of $\\mathrm{F}(\\cdot)$ [28,32,34,50], but also measure the efficacy of the threshold $\\phi$ .", + "bbox": [ + 214, + 507, + 787, + 554 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2 Rethinking UOD", + "text_level": 1, + "bbox": [ + 215, + 571, + 403, + 585 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As there is no prior train/test split for the target dataset, a classic outlier scoring method $\\mathrm{M}(\\cdot)$ predicts the outlier score function $\\mathrm{F}(\\mathbf{X}) = \\{\\mathrm{F}(\\mathbf{x}_i)\\}_{i=1}^n$ follows the below process:", + "bbox": [ + 212, + 593, + 785, + 637 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {F} (\\mathbf {X}) = \\mathrm {M}. \\operatorname {f i t} (\\mathbf {X}). \\operatorname {p r e d i c t} (\\mathbf {X}), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 395, + 638, + 784, + 654 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathrm{fit}(\\cdot)$ and $\\mathrm{predict}(\\cdot)$ refer to fitting the target dataset and predicting outlier scores, respectively. Motivated by previous research [32, 33], Shell normalization S-norm $(\\cdot)$ illustrated below is an ideal feature de-noising/refining paradigm.", + "bbox": [ + 212, + 659, + 787, + 705 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {S - n o r m} \\left(\\mathbf {x} _ {i}, \\mathbf {v} _ {\\mathrm {S}}\\right) = \\frac {\\mathbf {x} _ {i} - \\mathbf {v} _ {\\mathrm {S}}}{\\left| \\left| \\mathbf {x} _ {i} - \\mathbf {v} _ {\\mathrm {S}} \\right| \\right| _ {2}}, \\mathbf {v} _ {\\mathrm {S}} = \\left[ \\frac {1}{n} \\sum_ {i = 1} ^ {n} \\mathbf {X} _ {\\text {o u t}} [ i ] [ 1 ], \\dots , \\frac {1}{n} \\sum_ {i = 1} ^ {n} \\mathbf {X} _ {\\text {o u t}} [ i ] [ d ] \\right], \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 233, + 713, + 785, + 763 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $||\\cdot ||_2$ refers to $\\ell_2$ -norm and $d$ is the feature dimension. Thus, our first stage is to identify outlier candidates $\\mathbf{X}_{\\mathrm{out}}^{\\prime}$ . Subsequently, as how some one-class learning-based methods (e.g., OCVSM [51], DeepSVDD [50]) usually perform, we secondly predict the inliers $\\mathbf{X}_{\\mathrm{in}}^{\\prime}$ to fit an inlier/normal manifold. Thus, our objective is to identify both $\\mathbf{X}_{\\mathrm{in}}^{\\prime}$ and $\\mathbf{X}_{\\mathrm{out}}^{\\prime}$ .", + "bbox": [ + 212, + 765, + 787, + 842 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "Z. Liu et al.", + "bbox": [ + 271, + 114, + 354, + 127 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/83c573f03cd01b2363343e817f095bb8d99b2e5f646cfeae7bf746ecdadac16c.jpg", + "image_caption": [ + "Fig. 2: The overall paradigm of adopting the multi-thresholds learning (Multi-T) module to advance the existing outlier scoring methods. (a) The preparation consists of feature extraction and an initial outlier score function. (b) Visualization of our Multi-T module. (c) Integrating the predicted inliers and outliers with the previously proposed outlier detectors and obtaining an enhanced outlier score function." + ], + "image_footnote": [], + "bbox": [ + 212, + 143, + 787, + 383 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4 Method", + "text_level": 1, + "bbox": [ + 215, + 493, + 328, + 508 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our solution is displayed as below subsections. Sec. 4.1: we introduce the initial outlier score function; Sec. 4.2: we present the Multi-T module, to generate two thresholds for separating inliers and outliers from the target dataset; Sec. 4.3: we leverage predicted inliers and outliers to advance previously proposed methods.", + "bbox": [ + 212, + 530, + 787, + 592 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1 Initial Outlier Score Function", + "text_level": 1, + "bbox": [ + 215, + 619, + 504, + 633 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In our approach, we employ LVAD [34], one of the UOD SOTAs, as the initial outlier score function. LVAD assumes each image feature $\\mathbf{x}_i \\in \\mathbf{X}$ comes from one of $T$ high-dimensional generative processes. Thus, the outlier score becomes a sum of the weighted distance of the given instance, $\\mathbf{x}_i$ , arising from each generative process $\\{w_t, \\mathbf{m}_t \\mid t \\in \\{1, \\dots, T\\}\\}$ , which can be simplified as:", + "bbox": [ + 212, + 650, + 787, + 727 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {F} _ {\\mathrm {L V A D}} (\\mathbf {x} _ {i}) = \\sum_ {t = 1} ^ {T} w _ {t} \\cdot \\operatorname {D i s t} (\\mathrm {E} - \\operatorname {n o r m} (\\mathbf {x} _ {i}), \\mathrm {E} - \\operatorname {n o r m} (\\mathbf {m} _ {t})), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 308, + 739, + 785, + 781 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathrm{Dist}(\\cdot)$ refers to the $\\ell_2$ -norm metric (Euclidean distance), $w_{t}$ is the weight of the $t$ -th generative process, $\\mathbf{m}_t$ is the $t$ -th generative process's centroid and $\\mathrm{E - norm}(\\cdot)$ refers to the Ergodic-set normalization, an effective and outlier ratio", + "bbox": [ + 212, + 794, + 787, + 842 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "Rethinking Unsupervised Outlier Detection via Multiple Thresholding", + "bbox": [ + 264, + 114, + 732, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "code", + "sub_type": "algorithm", + "code_caption": [ + "Algorithm 1: MTL" + ], + "code_body": "1 Input: Initial Outlier Score Function: $\\mathrm{F_{init}(X)}$ , Outlier Scoring Method: M(·) \n2 Output: Advanced Outlier Score Function: $\\mathrm{F_{M + Multi - T}(X)}$ \n3 // Stage 1: Identifying Uncontaminated Inliers \n4 for $b$ in iterations do \n5 if not converged then \n6 $\\begin{array}{rlr}\\widehat{\\mathrm{F}}_{\\mathrm{init}}(\\mathbf{X}) = \\mathrm{sort}(\\mathrm{F}_{\\mathrm{init}}(\\mathbf{X})) & \\quad / & \\mathrm{sort - transform(Eq.9)}\\\\ I^b = \\left\\{\\widehat{\\mathrm{F}}_{\\mathrm{init}}(\\mathbf{x}_i)|i < \\max_i\\left\\{g(a_i) > \\widehat{\\mathrm{F}}_{\\mathrm{init}}(\\mathbf{x}_i)\\right\\} \\right\\} & \\quad / & \\mathrm{linear - fit(Eq.10)}\\\\ \\phi_{\\mathrm{out}}^b = \\mathrm{mean}(I^b) + 3\\cdot \\mathrm{std}(I^b) & \\quad / & \\mathrm{outlier~threshold~(Eq.11)}\\\\ I^{b + 1} = I^b.\\mathrm{remove}(\\{\\mathrm{F}_{\\mathrm{init}}(\\mathbf{x}_i)|\\mathrm{F}_{\\mathrm{init}}(\\mathbf{x}_i) > \\phi_{\\mathrm{out}}^b\\}) & \\quad / & \\mathrm{filtering(Eq.11)}\\\\ b = b + 1 \\end{array}$ \n10 \n11 Converged $b:b^{*}$ \n12 $\\phi_{\\mathrm{in}} = \\max (I^{b^{*}})$ // inlier threshold(Eq.12) \n13 $\\mathbf{X}_{\\mathrm{in}}^{\\prime} = \\{\\mathbf{x}_{i}|F_{\\mathrm{init}}(\\mathbf{x}_{i})\\leq \\phi_{\\mathrm{in}}\\}$ // predicted inliers(Eq.13) \n14 // Stage 2: Identifying Adaptive Outliers \n15 $\\rho \\gets$ similarity(Fs(X),FE(X)) // outlier ratio estimation(Eq.15,16) \n16 if $\\rho >0.3$ then \n17 $\\phi_{\\mathrm{out}} = \\phi_{\\mathrm{out}}^{*}$ \n18 else \n19 if $0.1\\leq \\rho \\leq 0.3$ then \n20 $\\phi_{\\mathrm{out}} = \\phi_{\\mathrm{out}}^{1}$ \n21 else \n22 $\\begin{array}{rlr}\\lfloor & \\phi_{\\mathrm{out}} = \\phi_{\\mathrm{out}}^{0}\\\\ & \\end{array}$ \n23 $\\mathbf{X}_{\\mathrm{out}}^{\\prime} = \\{\\mathbf{x}_{i}|F_{\\mathrm{init}}(\\mathbf{x}_{i}) > \\phi_{\\mathrm{out}}\\}$ // predicted outliers(Eq.20) \n24 // Stage 3: Adopting Multi-T to Outlier Scoring Method \n25 return M.fit({S-norm(xi,v's)|xi∈X'in}).predict({S-norm(xi,v's)|xi∈X})", + "bbox": [ + 218, + 166, + 787, + 551 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "invariant normalization procedure, illustrated as follows:", + "bbox": [ + 215, + 590, + 624, + 606 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {E} - \\operatorname {n o r m} \\left(\\mathbf {x} _ {i}, \\mathbf {v} _ {\\mathrm {E}}\\right) = \\frac {\\mathbf {x} _ {i} - \\mathbf {v} _ {\\mathrm {E}}}{\\left| \\left| \\mathbf {x} _ {i} - \\mathbf {v} _ {\\mathrm {E}} \\right| \\right| _ {2}}, \\mathbf {v} _ {\\mathrm {E}} = \\left[ \\frac {1}{n \\cdot d} \\sum_ {i = 1} ^ {n} \\sum_ {j = 1} ^ {d} \\mathbf {x} _ {i, j}, \\dots , \\frac {1}{n \\cdot d} \\sum_ {i = 1} ^ {n} \\sum_ {j = 1} ^ {d} \\mathbf {x} _ {i, j} \\right], \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 220, + 619, + 784, + 676 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\mathbf{v}_{\\mathrm{E}}$ is its reference vector and $j$ is the dimension index. To gain higher efficiency and maintain efficacy, we set $T = 1$ , such that the initial outlier score function (LVAD-S) is formulated as:", + "bbox": [ + 212, + 676, + 787, + 722 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathrm {F} _ {\\text {i n i t}} (\\mathbf {X}) = \\mathrm {F} _ {\\text {L V A D - S}} (\\mathbf {X}) = \\left. \\left\\{\\operatorname {D i s t} \\left(\\mathrm {E} - \\operatorname {n o r m} \\left(\\mathbf {x} _ {i}, \\mathbf {v} _ {\\mathrm {E}}\\right), \\operatorname {E} - \\operatorname {n o r m} \\left(\\mathbf {m} _ {\\mathbf {X}}, \\mathbf {v} _ {\\mathrm {E}}\\right)\\right) \\right\\} _ {i = 1} ^ {n}, \\right. \\\\ \\mathbf {m} _ {\\mathbf {X}} = \\left[ \\frac {1}{n} \\sum_ {i = 1} ^ {n} \\mathbf {x} _ {i, 1}, \\dots , \\frac {1}{n} \\sum_ {i = 1} ^ {n} \\mathbf {x} _ {i, d} \\right], \\tag {6} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 233, + 736, + 785, + 796 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\mathbf{m}_{\\mathbf{X}}$ is the mean of target dataset's features. Despite its simplicity, Eq. 6 is still a reliable outlier scoring method [34].", + "bbox": [ + 212, + 809, + 785, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "Z. Liu et al.", + "bbox": [ + 271, + 114, + 354, + 127 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2 Multi-T: Multiple Thresholding", + "text_level": 1, + "bbox": [ + 215, + 146, + 524, + 162 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Motivation. In the field of UOD, the conventional thresholding paradigm is to learn a single threshold $\\phi$ on the initial outlier score function $\\mathrm{F}_{\\mathrm{init}}(\\mathbf{X})$ . However, it ignores the impact of the outlier ratio $\\gamma$ . As the mean of target dataset $\\mathbf{m}_{\\mathbf{X}}$ will be shifted to outliers when $\\gamma$ becomes higher, i.e., mean-shift problem [47], there will be an inevitable overlap between inlier and outlier score distributions. So we define the distribution of the initial outlier score function $D$ as follows:", + "bbox": [ + 212, + 170, + 782, + 261 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\nD = \\operatorname {F} _ {\\text {i n i t}} (\\mathbf {X}) = \\operatorname {F} _ {\\text {i n i t}} \\left(\\mathbf {X} _ {\\text {i n}}\\right) \\cup \\operatorname {F} _ {\\text {i n i t}} \\left(\\mathbf {X} _ {\\text {o u t}}\\right) = I \\cup A \\cup O, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 313, + 272, + 785, + 287 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $I$ and $O$ refer to the uncontaminated inlier and outlier score distributions, $A$ refers to the overlap. Such that we are concerned about multi-thresholding instead of the conventional way that explicitly finding a single threshold between $\\mathrm{F_{init}(\\mathbf{X}_{in})}$ and $\\mathrm{F_{init}(\\mathbf{X}_{out})}$ . Specifically, our objective is identifying the inlier threshold $\\phi_{\\mathrm{in}}$ and outlier threshold $\\phi_{\\mathrm{out}}$ that isolate $I$ and $O$ from $D$ .", + "bbox": [ + 214, + 297, + 782, + 372 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Identifying uncontaminated inliers. We identify $I$ by understanding the initial outlier score function (Eq. 6), which computes the normalized $\\ell_2$ -norm of each instance $\\mathbf{x}_i \\in \\mathbf{X}$ relative to the mean of $\\mathbf{X}$ . Inspired by Shell Theory [32], $I$ satisfies a Gaussian-like distribution, the classic statistical estimation of $I$ is:", + "bbox": [ + 214, + 373, + 785, + 434 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\nI = \\left\\{\\mathrm {F} _ {\\text {i n i t}} (\\mathbf {x} _ {i}) \\mid \\mathrm {F} _ {\\text {i n i t}} (\\mathbf {x} _ {i}) < \\operatorname {m e a n} (D) + k \\cdot \\operatorname {s t d} (D) \\right\\}, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 326, + 444, + 784, + 460 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $\\mathrm{mean}(D)$ and $\\mathrm{std}(D)$ refer to the mean and standard deviation, $k$ is the hyper-parameter of the \" $k$ -sigma\" rule. In practice, optimizing $k$ without supervision is challenging. Thus, we convert to a non-parametric way by utilizing an ascending sort projection $\\mathrm{sort}(\\cdot)$ to $\\mathrm{F}_{\\mathrm{init}}(\\mathbf{X})$ , which projects $D$ into a 2- $d$ space ( $x$ -axis: instance index, $y$ -axis: sorted outlier score), the sorted outlier score is:", + "bbox": [ + 214, + 469, + 782, + 546 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\widehat {\\mathrm {F}} _ {\\text {i n i t}} (\\mathbf {X}) = \\operatorname {s o r t} \\left(\\widehat {\\mathrm {F}} _ {\\text {i n i t}} (\\mathbf {X})\\right), i. e., \\widehat {\\mathrm {F}} _ {\\text {i n i t}} [ i + 1 ] > \\widehat {\\mathrm {F}} _ {\\text {i n i t}} [ i ]. \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 320, + 556, + 784, + 574 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Intuitively, $\\widehat{I} = \\mathrm{sort}(I)$ can be fitted with a naive linear regressor $g(a) = \\boldsymbol{\\beta}^{\\top} \\cdot a = \\beta_{0} + \\beta_{1} \\cdot a$ , as shown in Fig. 2 (b). So we identify the potential inliers as:", + "bbox": [ + 214, + 584, + 782, + 616 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\nI = \\left\\{\\widehat {\\mathrm {F}} _ {\\text {i n i t}} (\\mathbf {x} _ {i}) | i < \\max _ {i} \\left\\{g \\left(a _ {i}\\right) > \\widehat {\\mathrm {F}} _ {\\text {i n i t}} (\\mathbf {x} _ {i}) \\right\\} \\right\\}, s. t. \\min _ {\\beta} \\left| \\left| a ^ {\\top} \\beta - \\widehat {\\mathrm {F}} _ {\\text {i n i t}} (\\mathbf {X}) \\right| \\right| _ {2}, \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 233, + 626, + 784, + 667 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $\\beta$ is the coefficients. Inevitably, this fitting process will be shifted with outliers, which is addressed together with the following outliers' identification.", + "bbox": [ + 212, + 667, + 782, + 696 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Identifying adaptive outliers. In statistics, 3-sigma rule [42] declares variables $C$ has a large probability within three standard deviations from the mean if $C$ follows a Gaussian-like distribution. So we employ it with two aspects:", + "bbox": [ + 214, + 696, + 782, + 741 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "(i) iteratively filter outliers during the phase of inliers identification:", + "bbox": [ + 215, + 742, + 705, + 758 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\phi_ {\\text {o u t}} ^ {b} = \\operatorname {m e a n} \\left(I ^ {b}\\right) + 3 \\cdot \\operatorname {s t d} \\left(I ^ {b}\\right), I ^ {b + 1} = I ^ {b}. \\text {r e m o v e} \\left(\\left\\{\\mathrm {F} _ {\\text {i n i t}} \\left(\\mathbf {x} _ {i}\\right) \\mid \\mathrm {F} _ {\\text {i n i t}} \\left(\\mathbf {x} _ {i}\\right) > \\phi_ {\\text {o u t}} ^ {b} \\right\\}\\right), \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 228, + 767, + 784, + 797 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $b$ refers to the $b$ -th iteration. When converged, the inlier threshold is:", + "bbox": [ + 215, + 799, + 761, + 813 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\phi_ {\\text {i n}} = \\max \\left(I ^ {b ^ {*}}\\right), \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 444, + 824, + 784, + 840 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "Rethinking Unsupervised Outlier Detection via Multiple Thresholding", + "bbox": [ + 264, + 114, + 732, + 130 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $I^{b^*}$ refers to the converged inlier prediction and we can identify inliers as:", + "bbox": [ + 215, + 145, + 782, + 162 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {X} _ {\\mathrm {i n}} ^ {\\prime} = \\left\\{\\mathbf {x} _ {i} \\mid \\mathrm {F} _ {\\text {i n i t}} \\left(\\mathbf {x} _ {i}\\right) \\leq \\phi_ {\\mathrm {i n}} \\right\\}. \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 406, + 176, + 785, + 193 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "(ii) During the above phase, we estimate the outlier threshold candidates:", + "bbox": [ + 215, + 196, + 746, + 213 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\phi_ {\\mathrm {o u t}} ^ {1} = \\operatorname {m e a n} \\left(I ^ {1}\\right) + 3 \\cdot \\operatorname {s t d} \\left(I ^ {1}\\right), \\phi_ {\\mathrm {o u t}} ^ {*} = \\operatorname {m e a n} \\left(I ^ {b ^ {*}}\\right) + 3 \\cdot \\operatorname {s t d} \\left(I ^ {b ^ {*}}\\right), \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 282, + 222, + 785, + 242 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $\\phi_{\\mathrm{out}}^{1}$ , $\\phi_{\\mathrm{out}}^{*}$ refer to the outlier threshold candidates at the first and final iterations, respectively, i.e., $\\phi_{\\mathrm{out}}^{1} > \\phi_{\\mathrm{out}}^{*}$ . To obtain a $\\gamma$ -adaptive outlier threshold, we consider the impact and relationship of two mentioned normalization procedures: Shell Normalization S-norm( $\\cdot$ ) (Eq. 3) and Ergodic-set Normalization E-norm( $\\cdot$ ) (Eq. 5). Both them follow a similar normalization formula (denominator: $\\ell_2$ -norm), the difference is their reference vectors $\\mathbf{v}_{\\mathrm{S}}'$ and $\\mathbf{v}_{\\mathrm{E}}$ , contrast as:", + "bbox": [ + 215, + 248, + 785, + 340 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {v} _ {\\mathrm {S}} ^ {\\prime} = \\left[ \\frac {1}{n} \\sum_ {i = 1} ^ {n} \\mathbf {X} _ {\\text {o u t}} ^ {\\prime} [ i ] [ 1 ], \\dots , \\frac {1}{n} \\sum_ {i = 1} ^ {n} \\mathbf {X} _ {\\text {o u t}} ^ {\\prime} [ i ] [ d ] \\right], \\tag {15}\n$$\n", + "text_format": "latex", + "bbox": [ + 338, + 352, + 785, + 400 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {v} _ {\\mathrm {E}} = \\left[ \\frac {1}{n \\cdot d} \\sum_ {i = 1} ^ {n} \\sum_ {j = 1} ^ {d} \\mathbf {x} _ {i, j}, \\dots , \\frac {1}{n \\cdot d} \\sum_ {i = 1} ^ {n} \\sum_ {j = 1} ^ {d} \\mathbf {x} _ {i, j} \\right].\n$$\n", + "text_format": "latex", + "bbox": [ + 336, + 393, + 668, + 435 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Obviously, Shell normalization is strongly subjective to the efficacy of outlier prediction $\\mathbf{X}_{\\mathrm{out}}^{\\prime}$ while Ergodic-set Normalization is independent with $\\gamma$ that is a sub-optimal but stable operation. To compare these two normalization procedures, we first employ them to Eq. 6, and obtain two outlier score functions:", + "bbox": [ + 215, + 439, + 785, + 501 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathrm {F} _ {\\mathrm {S}} (\\mathbf {X}) = \\left\\{\\operatorname {D i s t} \\left(\\mathrm {S} - \\operatorname {n o r m} \\left(\\mathbf {x} _ {i}, \\mathbf {v} _ {\\mathrm {S}} ^ {\\prime}\\right), \\operatorname {S} - \\operatorname {n o r m} \\left(\\mathbf {m} _ {\\mathbf {X}}, \\mathbf {v} _ {\\mathrm {S}} ^ {\\prime}\\right)\\right) \\right\\} _ {i = 1} ^ {n}, \\\\ \\mathrm {F} _ {\\mathrm {E}} (\\mathbf {X}) = \\left\\{\\operatorname {D i s t} \\left(\\mathrm {E} - \\operatorname {n o r m} \\left(\\mathbf {v} _ {\\mathrm {E}} - \\mathbf {v} _ {\\mathrm {I}}\\right), \\operatorname {E} - \\operatorname {n o r m} \\left(\\mathbf {m} _ {\\mathbf {E}} - \\mathbf {v} _ {\\mathrm {I}}\\right)\\right) \\right\\} ^ {n}. \\end{array} \\tag {16}\n$$\n", + "text_format": "latex", + "bbox": [ + 307, + 512, + 785, + 537 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {F} _ {\\mathrm {E}} (\\mathbf {X}) = \\left\\{\\operatorname {D i s t} \\left(\\operatorname {E} - \\operatorname {n o r m} \\left(\\mathbf {x} _ {i}, \\mathbf {v} _ {\\mathrm {E}}\\right), \\operatorname {E} - \\operatorname {n o r m} \\left(\\mathbf {m} _ {\\mathbf {X}}, \\mathbf {v} _ {\\mathrm {E}}\\right)\\right) \\right\\} _ {i = 1} ^ {n},\n$$\n", + "text_format": "latex", + "bbox": [ + 305, + 531, + 696, + 547 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $\\mathbf{v}_{\\mathrm{S}}^{\\prime}$ comes from $\\mathbf{X}_{\\mathrm{out}}^{\\prime}$ obtained by MAD [6] (Shell Normalization's default thresholding method that is only effective on high- $\\gamma$ ). Subsequently, two outlier score functions $\\mathrm{F_S}(\\mathbf{X})$ and $\\mathrm{F_E}(\\mathbf{X})$ are arranged in ascending order, and we denote the ranked-index lists as $\\mathcal{R}_{\\mathrm{S}}$ and $\\mathcal{R}_{\\mathrm{E}}$ . The outlier ratio $\\gamma$ can be approximately estimated with the similarity between $\\mathcal{R}_{\\mathrm{S}}$ and $\\mathcal{R}_{\\mathrm{E}}$ since $\\mathrm{F_S}(\\mathbf{X})$ and $\\mathrm{F_E}(\\mathbf{X})$ share a consistent structure (distance-to-the-mean), making $\\mathcal{R}_{\\mathrm{S}}$ and $\\mathcal{R}_{\\mathrm{E}}$ comparable. Besides, $\\mathrm{F_E}(\\mathbf{X})$ serves as a reliably effective baseline. When $\\gamma$ is low, S-norm $(\\mathbf{X},\\mathbf{v}_{\\mathrm{S}}^{\\prime})$ tends to under-perform, lead the similarity between $\\mathcal{R}_{\\mathrm{S}}$ and $\\mathcal{R}_{\\mathrm{E}}$ becomes low. In contrast, with a high $\\gamma$ , S-norm $(\\mathbf{X},\\mathbf{v}_{\\mathrm{S}}^{\\prime})$ merely refines the ranking of a small number of outliers, indicative of high similarity. The similarity $\\rho \\in [-1,1]$ is computed with the Pearson correlation coefficient:", + "bbox": [ + 215, + 551, + 787, + 720 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\rho = \\frac {\\operatorname {c o v} \\left(\\mathcal {R} _ {\\mathrm {S}} , \\mathcal {R} _ {\\mathrm {E}}\\right)}{\\operatorname {s t d} \\left(\\mathcal {R} _ {\\mathrm {S}}\\right) \\cdot \\operatorname {s t d} \\left(\\mathcal {R} _ {\\mathrm {E}}\\right)}, \\tag {17}\n$$\n", + "text_format": "latex", + "bbox": [ + 418, + 728, + 785, + 762 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $\\operatorname{cov}(\\mathcal{R}_{\\mathrm{S}},\\mathcal{R}_{\\mathrm{E}})$ and $\\mathrm{std}(\\mathcal{R}_{\\mathrm{S}}) / \\mathrm{std}(\\mathcal{R}_{\\mathrm{E}})$ are the covariance and standard deviation of the rank variables. If the similarity is relatively low ( $< 0.1$ ), we compute the 3-sigma of the entire outlier score distribution $D$ as the outlier threshold:", + "bbox": [ + 215, + 768, + 785, + 814 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\phi_ {\\text {o u t}} ^ {0} = \\operatorname {m e a n} (D) + 3 \\cdot \\operatorname {s t d} (D). \\tag {18}\n$$\n", + "text_format": "latex", + "bbox": [ + 398, + 824, + 785, + 842 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "Z. Liu et al.", + "bbox": [ + 271, + 114, + 354, + 127 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "It is a conservative estimator for outliers since the above linear fitting might suffer from some challenges at low- $\\gamma$ since there is no explicit outlier score distribution. If the similarity is large ( $>0.3$ ), we choose the relatively smooth outlier threshold $\\phi_{\\mathrm{out}}^{*}$ , while $\\phi_{\\mathrm{out}}^{1}$ is used for middle- $\\gamma$ cases. The outlier threshold is described as:", + "bbox": [ + 214, + 146, + 782, + 208 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\phi_ {\\text {o u t}} = \\left\\{ \\begin{array}{l} \\phi_ {\\text {o u t}} ^ {*}, \\text {i f} \\rho > 0. 3 \\\\ \\phi_ {\\text {o u t}} ^ {1}, \\text {i f} 0. 1 \\leq \\rho \\leq 0. 3 \\\\ \\phi_ {\\text {o u t}} ^ {0}, \\text {o t h e r w i s e} \\end{array} \\right. \\tag {19}\n$$\n", + "text_format": "latex", + "bbox": [ + 387, + 219, + 782, + 265 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Therefore, we can subsequently identify outlier ratio invariant outliers as:", + "bbox": [ + 215, + 268, + 743, + 282 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {X} _ {\\text {o u t}} ^ {\\prime} = \\left\\{\\mathbf {x} _ {i} \\mid \\mathrm {F} _ {\\text {i n i t}} \\left(\\mathbf {x} _ {i}\\right) > \\phi_ {\\text {o u t}} \\right\\}. \\tag {20}\n$$\n", + "text_format": "latex", + "bbox": [ + 398, + 290, + 782, + 306 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.3 Implementation", + "text_level": 1, + "bbox": [ + 215, + 323, + 392, + 338 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Firstly, we can directly employ Multi-T with the distance-to-the-mean paradigm:", + "bbox": [ + 215, + 344, + 782, + 359 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {F} _ {\\mathrm {M u l t i - T}} (\\mathbf {X}) = \\{\\mathrm {D i s t} (\\mathrm {S - n o r m} (\\mathbf {x} _ {i}, \\mathbf {v} _ {\\mathrm {S}} ^ {\\prime}), \\mathrm {S - n o r m} (\\mathbf {m} _ {\\mathbf {X} _ {\\mathrm {i n}} ^ {\\prime}}, \\mathbf {v} _ {\\mathrm {S}} ^ {\\prime})) \\} _ {i = 1} ^ {n},\n$$\n", + "text_format": "latex", + "bbox": [ + 287, + 371, + 715, + 388 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {v} _ {\\mathrm {S}} ^ {\\prime} = \\left[ \\frac {1}{n} \\sum_ {i = 1} ^ {n} \\mathbf {X} _ {\\text {o u t}} ^ {\\prime} [ i ] [ 1 ], \\dots , \\frac {1}{n} \\sum_ {i = 1} ^ {n} \\mathbf {X} _ {\\text {o u t}} ^ {\\prime} [ i ] [ d ] \\right], \\tag {21}\n$$\n", + "text_format": "latex", + "bbox": [ + 352, + 392, + 782, + 429 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {m} _ {\\mathbf {X} _ {\\mathrm {i n}} ^ {\\prime}} = \\left[ \\frac {1}{n} \\sum_ {i = 1} ^ {n} \\mathbf {X} _ {\\mathrm {i n}} ^ {\\prime} [ i ] [ 1 ], \\dots , \\frac {1}{n} \\sum_ {i = 1} ^ {n} \\mathbf {X} _ {\\mathrm {i n}} ^ {\\prime} [ i ] [ d ] \\right].\n$$\n", + "text_format": "latex", + "bbox": [ + 331, + 434, + 653, + 470 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Secondly, Multi-T can be integrated with UOD methods $\\mathrm{M}(\\cdot)$ , illustrated as:", + "bbox": [ + 215, + 474, + 766, + 489 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathrm {F} _ {\\mathrm {M} + \\text {M u l t i - T}} (\\mathbf {X}) = \\mathrm {M}. \\operatorname {f i t} \\left(\\left\\{\\mathrm {S} - \\operatorname {n o r m} \\left(\\mathbf {x} _ {i}, \\mathbf {v} _ {\\mathrm {S}} ^ {\\prime}\\right) \\mid \\mathbf {x} _ {i} \\in \\mathbf {X} _ {i n} ^ {\\prime} \\right\\}\\right) \\\\ . \\operatorname {p r e d i c t} \\left(\\left\\{\\mathrm {S} - \\operatorname {n o r m} \\left(\\mathbf {x} _ {i}, \\mathbf {v} _ {\\mathrm {S}} ^ {\\prime}\\right) \\mid \\mathbf {x} _ {i} \\in \\mathbf {X} \\right\\}\\right). \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 305, + 501, + 784, + 536 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "5 Experiments", + "text_level": 1, + "bbox": [ + 215, + 553, + 375, + 570 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We cover our experiments for (i) comparing with other threshold learners on dataset splitting (Sec. 5.1); (ii) measuring the performance improvement of adopting the Multi-T to previously proposed outlier scoring methods (Sec. 5.2).", + "bbox": [ + 212, + 580, + 785, + 625 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Benchmarks and Feature Extraction. Based on the existing research [32,34], we use the raw pixel representation for grayscale datasets, including MNIST [29] and Fashion-MNIST [59]. For RGB datasets such as STL-10 [15], Internet [32], CIFAR-10 [27], MIT-Places-Small [61], we adopt two deep feature extractors: ImageNet pretrained [20] ResNet-50 [21] (ResNet)2, and CLIP [45].", + "bbox": [ + 212, + 626, + 782, + 700 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Target Dataset. In our experiments, each class within a benchmark dataset is alternatively regarded as inliers, with instances from all other classes considered outliers. The target dataset includes all inliers along with a randomly selected subset of outliers. Results for each dataset are averaged across all classes. Additionally, for every round, we further average the results over a wide range of outlier ratios: $[0.05, 0.1, 0.2, 0.3, 0.4]^3$ , to assess the method's robustness/stability.", + "bbox": [ + 212, + 700, + 785, + 792 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "Rethinking Unsupervised Outlier Detection via Multiple Thresholding", + "bbox": [ + 264, + 114, + 732, + 130 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 114, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "page_footnote", + "text": "2 Unless otherwise stated, ResNet-50 is the default feature extractor.", + "bbox": [ + 217, + 810, + 679, + 824 + ], + "page_idx": 8 + }, + { + "type": "page_footnote", + "text": "3 Unless otherwise stated, the experimental results are averaged across $[0.05,\\dots ,0.4]$", + "bbox": [ + 218, + 825, + 782, + 840 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/0ea2d465b66a5324d1e7a34f36b80e91c9d0a7e96f996495888d2e720529f2ed.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 215, + 142, + 787, + 268 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/ba6b4d142c17f8360942e86c2b65bf5a75fe039c58dc6e41df6d9f4276f2a032.jpg", + "image_caption": [ + "Fig.3: Qualitative results compared with SOTA threshold learners. The experiments are conducted on STL-10 (Inlier class: Monkey)." + ], + "image_footnote": [], + "bbox": [ + 215, + 270, + 787, + 401 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "5.1 Threshold Learning", + "text_level": 1, + "bbox": [ + 215, + 464, + 426, + 479 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Competing Methods. We compare our proposed Multi-T with those effective thresholding (TL) methods involving Kernel-based: AUCP [48], FGD [43]; Curve-based: EB [17]; Normality-based: DSN [5], CHAU [10]; Filtering-based: FILTER [19], HIST [55]; Statistical-based: MTT [49], BOOT [37], QMCD [24], CLF [7], IQR [8], KARCH [1], MCST [16], GESD [4], REGR [3], MAD [6], CLUST [11], CPD [12] and Transformation-based: MOLL [25], YJ [46].", + "bbox": [ + 212, + 488, + 787, + 579 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Evaluation Metric. The performance of general thresholding is measured with $\\mathrm{F}_{\\beta}$ -score, defined as follows:", + "bbox": [ + 212, + 579, + 785, + 609 + ], + "page_idx": 9 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {F} _ {\\beta} = \\left(1 + \\beta^ {2}\\right) \\cdot \\frac {\\text {p r e c i s i o n} \\cdot \\text {r e c a l l}}{\\left(\\beta^ {2} \\cdot \\text {p r e c i s i o n}\\right) + \\text {r e c a l l}}, \\tag {23}\n$$\n", + "text_format": "latex", + "bbox": [ + 352, + 618, + 785, + 650 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "where $\\beta < 1$ tilts towards precision while $\\beta > 1$ prioritizes recall. In this work, we utilize $\\mathrm{F}_{0.1}$ -score $(\\mathrm{F}_{0.1})$ and $\\mathrm{F}_{10}$ -score $(\\mathrm{F}_{10})$ that measure the accuracy of predicted outliers and inliers. Besides, $\\mathrm{F}_{1}$ -score $(\\beta = 1)$ , a harmonic mean of the precision and recall, is not practical for this task since the compete separability between inliers' and outliers' score distributions cannot be guaranteed.", + "bbox": [ + 212, + 657, + 787, + 734 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Main Results. Across a series of competing threshold learners, while one method may achieve a commendable $\\mathrm{F}_{0.1}$ -score of 0.976, the corresponding $\\mathrm{F}_{10}$ -score is mere 0.452, and vice versa ( $\\mathrm{F}_{10}$ -score: 0.979; $\\mathrm{F}_{0.1}$ -score: 0.447). By contrast, our solution maintains its efficacy across various evaluation criteria ( $\\mathrm{F}_{0.1}$ -score: 0.917, $\\mathrm{F}_{10}$ -score: 0.980). It exhibits remarkable stability across a diverse spectrum of outlier ratios, as illustrated in Fig. 3. To further investigate the benefits of our technique, Tab. 1 categorizes two groups, each representing a different", + "bbox": [ + 212, + 734, + 787, + 840 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "Z. Liu et al.", + "bbox": [ + 271, + 114, + 354, + 127 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/fbe891ae5a6a4c1bacb10ab1a83a9c2e6423ad602b48d2658c85819d5638ce4b.jpg", + "table_caption": [ + "Table 1: Average $\\mathrm{F}_{\\beta}$ -score of thresholding methods on STL-10. \"+\"GT Norm\" refers to the ideal initial outlier score function. \"Highest $\\mathrm{F}_{\\beta}$ \" ( $\\beta: 0.1, 10$ ) refers to the method that garners the highest $\\mathrm{F}_{\\beta}$ -score across all baseline models. The better average result is highlighted in bold." + ], + "table_footnote": [], + "table_body": "
Outlier Score Fun.Method (TL)ResNet-50CLIP
F0.1F10Avg.F0.1F10Avg.
LVAD-S [34]Highest F0.10.9110.4540.6820.9360.4010.669
Highest F100.3820.9670.6740.3570.9540.655
Multi-T(Ours)0.8400.8690.8550.8600.8660.863
+GT Norm [32]Highest F0.10.9760.4520.7140.9770.4480.713
Highest F100.4470.9790.7130.6910.9860.839
Multi-T(Ours)0.9170.9800.9490.8570.9840.920
", + "bbox": [ + 241, + 212, + 751, + 363 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "outlier score function. Such categorization is pivotal as threshold methods ought to accommodate various foundational bases.", + "bbox": [ + 212, + 404, + 782, + 433 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/c4e8b38f24b3da10bfa2d939081c0f8c16ef44437a4aa2342032a0685b5a04e1.jpg", + "table_caption": [ + "Table 2: Average AUC results compared with SOTA outlier scoring methods (outlier detectors). Blue and Orange indicates the best and second-best results, respectively." + ], + "table_footnote": [], + "table_body": "
Method (OS)STL-10CIFAR-10CIFAR-100MIT-PlacesMNIST
ResNetCLIPResNetCLIPResNetCLIPResNetCLIPPixel
IF [35]0.8360.9430.7800.8910.7900.8660.6870.8680.776
LOF [11]0.6280.6260.6730.6210.8390.8490.5560.5200.754
RSRAE [28]0.9620.9380.8620.8790.9140.8850.8740.8760.793
ECOD [31]0.9070.9810.8730.9350.8730.9180.7770.9430.734
LUNAR [18]0.7760.8210.7670.7740.8380.8990.6430.8710.797
Shell-Re. [32]0.8620.8380.8600.8130.8350.8130.8260.9140.776
LVAD [34]0.9540.9680.8600.9170.9210.9170.8440.9190.867
Multi-T0.9680.9890.8950.9570.9380.9560.8670.9740.897
DeepSVDD [50]0.6220.5970.5600.5090.5630.5810.5830.5490.513
+Multi-T0.9250.9210.7690.8190.8350.8260.7550.8320.732
Improve.48.7%54.3%37.3%60.9%48.3%42.2%29.5%51.6%37.9%
OCSVM [51]0.9270.9210.8260.8500.8790.8500.8260.8710.831
+Multi-T0.9570.9650.8590.9160.9160.8990.8460.9240.863
Improve.3.24%4.78%4.00%7.76%4.21%5.76%2.42%6.08%3.85%
", + "bbox": [ + 220, + 512, + 777, + 801 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "Rethinking Unsupervised Outlier Detection via Multiple Thresholding", + "bbox": [ + 264, + 114, + 730, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "5.2 Outlier Scoring", + "text_level": 1, + "bbox": [ + 215, + 145, + 388, + 162 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Competing Methods. We compare two themes of baseline outlier scoring (OS) methods: (i) statistical-based methods: IF [35], OCSVM [51], DeepSVDD [51], ECOD [31], LUNAR [28], RSRAE [28], Shell-Re. [32], LVAD [34]. (ii) deep-learning-based models: GOAD [9], ICL [53], REPEN [39], NeuTraL [44], SLAD [60]. For a fair comparison, we apply Ergodic-set normalization [34] if it improves the performance of the baseline algorithms, such as IF [35], OCSVM [51].", + "bbox": [ + 212, + 172, + 782, + 262 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Evaluation Metric. The performance of outlier scoring (ranking accuracy) is primarily assessed using the Area Under the Receiver Operating Characteristic curve (AUC). This metric provides a thorough assessment of ranking accuracy. Main Results. Tab. 2 shows that Multi-T can be seamlessly integrated with two classic outlier scoring methods: DeepSVDD [51], and OCSVM [51] while exhibiting significant improvements in both efficacy and stability across a diverse range of outlier ratios and various benchmarks. In most of our experiments, Multi-T itself achieves SOTA results. Notably, even in the case of non-aligned and low-resolution datasets like CIFAR-10 [27] and CIFAR-100 [27], which are known to pose challenges [41]. It surpasses the current SOTA AUC scores by margins of $7.76\\%$ and $5.76\\%$ for CIFAR-10 and CIFAR-100, respectively. Additionally, we observe a logical and significant enhancement in performance with improved feature representation, e.g., for the MIT-Places [61] dataset, the AUC improves from 0.867 using ResNet-50 [21] to 0.974 with CLIP [45].", + "bbox": [ + 212, + 265, + 782, + 474 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/41634fbb02a0421af99d58e7121887fd0f609778178bc237078d3bfa187d74c2.jpg", + "table_caption": [ + "Table 3: Average AUC results, for more outlier scoring methods with $(\\mathrm{w} / )$ or without(w/o) our Multi-T module, conducted on STL-10." + ], + "table_footnote": [], + "table_body": "
FeatureIF [35]ECOD [11]ABOD [26]PCA [54]GMM [2]
w/ow/w/ow/w/ow/w/ow/w/ow/
ResNet0.8360.8990.9070.9190.6650.8830.8650.9450.8590.952
CLIP0.9430.9830.9810.9840.7150.9090.9840.9940.8920.962
Avg.0.8900.9410.9440.9510.6900.8960.9250.9700.8760.957
FeatureGOAD [9]ICL [53]REPEN [39]NeuTraL [44]SLAD [60]
w/ow/w/ow/w/ow/w/ow/w/ow/
ResNet0.9520.9620.9340.9570.8770.8890.8540.9500.9410.962
CLIP0.9610.9890.9510.9820.8790.9300.8510.9710.9450.985
Avg.0.9570.9750.9430.9700.8780.9090.8530.9610.9430.973
", + "bbox": [ + 240, + 545, + 756, + 734 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "In Tab. 3, we present the results of Multi-T integrated with more methods involving both statistical and deep models, which shows its broad applications. Moreover, our solution excels not just in detection accuracy but also in running speed since Multi-T compresses the size of fitting data, as illustrated in Tab. 4, 5. In Fig. 4, we compare with the most related work Shell-Re. [32], which", + "bbox": [ + 212, + 763, + 784, + 839 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "Z. Liu et al.", + "bbox": [ + 271, + 114, + 352, + 126 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "has a built-in Robust-Least-Square (RLS) thresholding procedure with MAD [6]. Apparently, our method is capable of estimating multiple thresholds that facilitate a clear demarcation of diverse scenarios (outlier ratio, dataset domain, and feature representation).", + "bbox": [ + 212, + 146, + 787, + 207 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/1676022d9cc44fc572692e70d1ed5eda9ec2e7789e270ad99c767eac28872c57.jpg", + "table_caption": [ + "Table 4: Efficiency comparison for outlier Table 5: Efficiency comparison for scoring. Timing is measured with 10,000 thresholding. All thresholding methods samples, GPU: NVIDIA RTX 3080. are conducted on CPU." + ], + "table_footnote": [], + "table_body": "
Method (OS)DeviceTime (↓)
LVAD [34]CPU645.2s
RSRAE [28]GPU121.6s
Multi-TCPU1.2s
OCSVM [51]CPU154.8s
+Multi-T105.4s
GOAD [9]GPU268.1s
+Multi-T211.7s
", + "bbox": [ + 276, + 272, + 488, + 412 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/9ceb9270064564fc4133a27d53e60742b87f65de820f47642651cfa7a6e76733.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Method (TL)Time (↓)
CPD [12]3.61s
FWFM [19]1.26s
DSN [5]3.79s
CLUST [11]7.36s
AUCP [48]1.26s
Multi-T0.97s
", + "bbox": [ + 575, + 272, + 738, + 412 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/e1672596f140f350da057673c08f03e6b135ce868f81e8e8d74e03bb65541534.jpg", + "image_caption": [ + "Fig. 4: AUC results of Multi-T compared with Shell-Re. [32] over various outlier ratios." + ], + "image_footnote": [], + "bbox": [ + 217, + 462, + 784, + 672 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Discussions. We conclude the feasibility of adopting Multi-T to outlier scoring methods into three aspects: (i) both normalization with outliers and manifold learning with inliers are necessary for most of outlier scoring methods; (ii) Despite the initial outlier score distribution (Eq. 6) is usually not perfectly separable, it provides a reliable baseline. (iii) Multi-T is able to isolate the overlap region and predict both uncontaminated inliers.", + "bbox": [ + 212, + 748, + 787, + 840 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "Rethinking Unsupervised Outlier Detection via Multiple Thresholding", + "bbox": [ + 264, + 114, + 730, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/dafb589ace953eb14b59ed10b465fcd8cb8d669835e1be0502e37fb92d2c110c.jpg", + "table_caption": [ + "Table 6: Ablation study for DeepSVDD and LVAD-S with Multi-T ( $\\gamma$ : 0.2)." + ], + "table_footnote": [], + "table_body": "
DeepSVDDAUCLVAD-SAUC
+Multi-T0.950+Multi-T0.971
w/o In0.643w/o In0.965
w/o Out0.939w/o Out0.951
", + "bbox": [ + 263, + 183, + 500, + 267 + ], + "page_idx": 13 + }, + { + "type": "table", + "img_path": "images/26f614312681aa46a9c613f2782e1d7963fc19302f50853b0ce9c470392819ec.jpg", + "table_caption": [ + "Table 7: Comparison with $k$ -sigma." + ], + "table_footnote": [], + "table_body": "
MetricMethodScore
F0.13-sigma0.406
Multi-T0.840
F101-sigma0.806
Multi-T0.869
", + "bbox": [ + 576, + 170, + 736, + 267 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5.3 Ablation Study", + "text_level": 1, + "bbox": [ + 215, + 287, + 388, + 303 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "The experiments thus far have established the effectiveness of Multi-T. However, there remains a concern regarding how sensitive the performance improvement is to Multi-T's two primary components: adaptive outliers (Out) and uncontaminated inliers (In), which refer to the normalization and manifold learning procedures, respectively. We select two representative outlier scoring models: DeepSVDD [50] and LVAD-S [34], whereas DeepSVDD follows a widely-used outlier detection mechanism, i.e., learning the normality (hyper-sphere), and identifying inliers is of great significance. Without predicted inliers, the AUC result decreases from 0.950 to 0.643 on STL-10, shown in Tab. 6. Additionally, since the normalization procedure can be considered as a distance de-noising procedure [33], Multi-T will contribute to those methods with distance computation, e.g., the result of LVAD-S is improved from 0.951 to 0.971 on STL-10 and 0.838 to 0.882 on CIFAR-10, which verify our prior assumption that the identification of both inliers and outliers is of great significance, which enhances the reliability of multiple thresholding perspective. Moreover, Tab. 7 indicates that our thresholding process markedly surpasses the classical $k$ -sigma rule [42].", + "bbox": [ + 212, + 311, + 787, + 555 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5.4 Limitation", + "text_level": 1, + "bbox": [ + 215, + 577, + 349, + 590 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Our method is tied to the ranking accuracy (separability) of the initial outlier score function. Based on most related works, we consider UOD as an one-class learning task (only one inlier/normal class in the target dataset). However we find in some specific cases, there might exist multi-normal classes, e.g., the digit 5 class of the MNIST dataset. In that case, the efficacy of our method might be decreased. Additionally, this statistical-based method has limitations in very small-scale datasets since the \"three-sigma\" rule will be less effective.", + "bbox": [ + 212, + 602, + 787, + 708 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "6 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 732, + 359, + 748 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "This work introduces a novel perspective for UOD about advancing existing outlier detectors (scoring methods) via thresholding. To this end, we propose a multiple thresholding (Multi-T) module, to label the unlabelled target dataset. Comprehensive experiments verify that the Multi-T can significantly improve both the efficacy and efficiency of previously proposed outlier scoring methods.", + "bbox": [ + 212, + 763, + 787, + 840 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "Z. Liu et al.", + "bbox": [ + 271, + 114, + 352, + 126 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgement", + "text_level": 1, + "bbox": [ + 217, + 143, + 403, + 162 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "We would like to thank the reviewers for their valuable comments and suggestions.", + "bbox": [ + 215, + 178, + 782, + 208 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 233, + 321, + 250 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Afsari, B.: Riemannian $L^p$ center of mass: existence, uniqueness, and convexity. Proceedings of the American Mathematical Society pp. 655-673 (2011)", + "2. Aggarwal, C.C.: Outlier analysis second edition (2016)", + "3. Aggarwal, C.C., Aggarwal, C.C.: An introduction to outlier analysis. Springer (2017)", + "4. Alrawashdeh, M.J.: An adjusted grubbs' and generalized extreme studentized deviation. Demonstratio Mathematica pp. 548-557 (2021)", + "5. Amagata, D., Onizuka, M., Hara, T.: Fast and exact outlier detection in metric spaces: a proximity graph-based approach. In: Proceedings of the 2021 International Conference on Management of Data. pp. 36-48 (2021)", + "6. Archana, N., Pawar, S.: Periodicity detection of outlier sequences using constraint based pattern tree with mad. International Journal of Advanced Studies in Computers, Science and Engineering p. 34 (2015)", + "7. Barbado, A., Corcho, O., Benjamins, R.: Rule extraction in unsupervised anomaly detection for model explainability: Application to oneclasssvm. Expert Systems with Applications p. 116100 (2022)", + "8. Bardet, J.M., Dimby, S.F.: A new non-parametric detector of univariate outliers for distributions with unbounded support. Extremes pp. 751-775 (2017)", + "9. Bergman, L., Hoshen, Y.: Classification-based anomaly detection for general data. arXiv preprint arXiv:2005.02359 (2020)", + "10. Bol'shev, L., Ubaidullaeva, M.: Chauvenet's test in the classical theory of errors. Theory of Probability & Its Applications pp. 683-692 (1975)", + "1. Breunig, M.M., Kriegel, H.P., Ng, R.T., Sander, J.: Lof: identifying density-based local outliers. In: Proceedings of the ACM SIGMOD International Conference on Management of Data. pp. 93-104 (2000)", + "2. Van den Burg, G.J., Williams, C.K.: An evaluation of change point detection algorithms. arXiv preprint arXiv:2003.06222 (2020)", + "3. Camposeco, F., Sattler, T., Cohen, A., Geiger, A., Pollefeys, M.: Toroidal constraints for two-point localization under high outlier ratios. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 4545-4553 (2017)", + "4. Chalapathy, R., Menon, A.K., Chawla, S.: Robust, deep and inductive anomaly detection. In: Proceedings of Joint European Conference on Machine Learning and Knowledge Discovery in Databases. pp. 36-51 (2017)", + "5. Coates, A., Ng, A., Lee, H.: An analysis of single-layer networks in unsupervised feature learning. In: Proceedings of the International Conference on Artificial Intelligence and Statistics. pp. 215-223 (2011)", + "6. Coin, D.: Testing normality in the presence of outliers. Statistical Methods and Applications pp. 3-12 (2008)", + "7. Friendly, M., Monette, G., Fox, J.: Elliptical insights: understanding statistical methods through elliptical geometry. Statistical Science pp. 1-39 (2013)" + ], + "bbox": [ + 225, + 266, + 784, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "Rethinking Unsupervised Outlier Detection via Multiple Thresholding", + "bbox": [ + 264, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "18. Goodge, A., Hooi, B., Ng, S.K., Ng, W.S.: Lunar: Unifying local outlier detection methods via graph neural networks. In: Proceedings of the AAAI Conference on Artificial Intelligence. pp. 6737-6745 (2022)", + "19. Hashemi, N., German, E.V., Ramirez, J.P., Ruths, J.: Filtering approaches for dealing with noise in anomaly detection. In: 2019 IEEE 58th Conference on Decision and Control. pp. 5356-5361 (2019)", + "20. He, K., Girshick, R., Dollar, P.: Rethinking imagenet pre-training. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 4918-4927 (2019)", + "21. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 770-778 (2016)", + "22. Hendrycks, D., Mazeika, M., Dietterich, T.: Deep anomaly detection with outlier exposure. arXiv preprint arXiv:1812.04606 (2018)", + "23. Huyan, N., Quan, D., Zhang, X., Liang, X., Chanussot, J., Jiao, L.: Unsupervised outlier detection using memory and contrastive learning. IEEE Transactions on Image Processing pp. 6440-6454 (2022)", + "24. Iouchtchenko, D., Raymond, N., Roy, P.N., Nooijen, M.: Deterministic and quasi-random sampling of optimized gaussian mixture distributions for vibronic monte carlo. arXiv preprint arXiv:1912.11594 (2019)", + "25. Keyzer, M.A., Sonneveld, B.: Using the mollifier method to characterize datasets and models: the case of the universal soil loss equation. ITC Journal pp. 263-272 (1997)", + "26. Kriegel, H.P., Schubert, M., Zimek, A.: Angle-based outlier detection in high-dimensional data. In: Proceedings of the 14th ACM SIGKDD international conference on Knowledge discovery and data mining. pp. 444-452 (2008)", + "27. Krizhevsky, A., Hinton, G., et al.: Learning multiple layers of features from tiny images (2009)", + "28. Lai, C.H., Zou, D., Lerman, G.: Robust subspace recovery layer for unsupervised anomaly detection. arXiv preprint arXiv:1904.00152 (2019)", + "29. LeCun, Y., Cortes, C.: MNIST handwritten digit database (2010)", + "30. Li, T., Wang, Z., Liu, S., Lin, W.Y.: Deep unsupervised anomaly detection. In: Proceedings of IEEE Winter Conference on Applications of Computer Vision. pp. 3636-3645 (2021)", + "31. Li, Z., Zhao, Y., Hu, X., Botta, N., Ionescu, C., Chen, G.H.: Ecod: Unsupervised outlier detection using empirical cumulative distribution functions. arXiv preprint arXiv:2201.00382 (2022)", + "32. Lin, D., Liu, S., Li, H., Cheung, N.M., Ren, C., Matsushita, Y.: Shell theory: A statistical model of reality. IEEE Transactions on Pattern Analysis and Machine Intelligence (2021)", + "33. Lin, W.Y., Liu, S., Dai, B.T., Li, H.: Distance based image classification: A solution to generative classification's conundrum? International Journal of Computer Vision pp. 177-198 (2023)", + "34. Lin, W.Y., Liu, Z., Liu, S.: Locally varying distance transform for unsupervised visual anomaly detection. In: European Conference on Computer Vision. pp. 354-371. Springer (2022)", + "35. Liu, F.T., Ting, K.M., Zhou, Z.H.: Isolation forest. In: Proceedings of International Conference on Data Mining. pp. 413-422 (2008)", + "36. Liznerski, P., Ruff, L., Vandermeulen, R.A., Franks, B.J., Müller, K.R., Kloft, M.: Exposing outlier exposure: What can be learned from few, one, and zero outlier images. arXiv preprint arXiv:2205.11474 (2022)" + ], + "bbox": [ + 215, + 146, + 784, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "Z. Liu et al.", + "bbox": [ + 271, + 114, + 352, + 126 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "37. Martin, M.A., Roberts, S.: An evaluation of bootstrap methods for outlier detection in least squares regression. Journal of Applied Statistics pp. 703-720 (2006)", + "38. Masci, J., Meier, U., Ciresan, D., Schmidhuber, J.: Stacked convolutional autoencoders for hierarchical feature extraction. In: Proceedings of International Conference on Artificial Neural Networks. pp. 52-59 (2011)", + "39. Pang, G., Cao, L., Chen, L., Liu, H.: Learning representations of ultrahigh-dimensional data for random distance-based outlier detection. In: Proceedings of the 24th ACM SIGKDD international conference on knowledge discovery & data mining. pp. 2041-2050 (2018)", + "40. Parzen, E.: On estimation of a probability density function and mode. The annals of mathematical statistics pp. 1065-1076 (1962)", + "41. Perera, P., Nallapati, R., Xiang, B.: Ocgan: One-class novelty detection using gans with constrained latent representations. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 2898-2906 (2019)", + "42. Pukelsheim, F.: The three sigma rule. The American Statistician pp. 88-91 (1994)", + "43. Qi, Z., Jiang, D., Chen, X.: Iterative gradient descent for outlier detection. International Journal of Wavelets, Multiresolution and Information Processing p. 2150004 (2021)", + "44. Qiu, C., Pfrommer, T., Kloft, M., Mandt, S., Rudolph, M.: Neural transformation learning for deep anomaly detection beyond images. In: International Conference on Machine Learning. pp. 8703-8714 (2021)", + "45. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: Proceedings of International Conference on Machine Learning. pp. 8748-8763 (2021)", + "46. Raymaekers, J., Rousseeuw, P.J.: Transforming variables to central normality. Machine Learning pp. 1-23 (2021)", + "47. Reiss, T., Hoshen, Y.: Mean-shifted contrastive loss for anomaly detection. In: Proceedings of the AAAI Conference on Artificial Intelligence. pp. 2155-2162 (2023)", + "48. Ren, K., Yang, H., Zhao, Y., Chen, W., Xue, M., Miao, H., Huang, S., Liu, J.: A robust auc maximization framework with simultaneous outlier detection and feature selection for positive-unlabeled classification. IEEE Transactions on Neural Networks and Learning Systems pp. 3072-3083 (2018)", + "49. Rengasamy, D., Rothwell, B.C., Figueredo, G.P.: Towards a more reliable interpretation of machine learning outputs for safety-critical systems using feature importance fusion. Applied Sciences p. 11854 (2021)", + "50. Ruff, L., Vandermeulen, R., Goernitz, N., Deecke, L., Siddiqui, S.A., Binder, A., Müller, E., Kloft, M.: Deep one-class classification. In: Proceedings of International Conference on Machine Learning. pp. 4393-4402 (2018)", + "51. Scholkopf, B., Platt, J.C., Shawe-Taylor, J., Smola, A.J., Williamson, R.C.: Estimating the support of a high-dimensional distribution. Neural Computation 13(7), 1443-1471 (2001)", + "52. Scholkopf, B., Smola, A.J., Bach, F., et al.: Learning with kernels: support vector machines, regularization, optimization, and beyond (2002)", + "53. Shenkar, T., Wolf, L.: Anomaly detection for tabular data with internal contrastive learning. In: International Conference on Learning Representations (2021)", + "54. Shyu, M.L., Chen, S.C., Sarinnapakorn, K., Chang, L.: A novel anomaly detection scheme based on principal component classifier. In: Proceedings of the IEEE foundations and new directions of data mining workshop. pp. 172-179 (2003)" + ], + "bbox": [ + 212, + 146, + 785, + 840 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "Rethinking Unsupervised Outlier Detection via Multiple Thresholding", + "bbox": [ + 264, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "55. Thanammal, K., Jayasudha, J., Vijayalakshmi, R., Arumugaperumal, S.: Effective histogram thresholding techniques for natural images using segmentation. Journal of Image and Graphics pp. 113-116 (2014)", + "56. Wang, S., Zeng, Y., Liu, X., Zhu, E., Yin, J., Xu, C., Kloft, M.: Effective end-to-end unsupervised outlier detection via inlier priority of discriminative network. Proceedings of Advances in Neural Information Processing Systems (2019)", + "57. Wang, S., Zhu, E., Hu, X., Liu, X., Liu, Q., Yin, J., Wang, F.: Robustness can be cheap: A highly efficient approach to discover outliers under high outlier ratios. In: Proceedings of the AAAI Conference on Artificial Intelligence. pp. 5313-5320 (2019)", + "58. Xia, Y., Cao, X., Wen, F., Hua, G., Sun, J.: Learning discriminative reconstructions for unsupervised outlier removal. In: Proceedings of the IEEE International Conference on Computer Vision. pp. 1511-1519 (2015)", + "59. Xiao, H., Rasul, K., Vollgraf, R.: Fashion-mnist: a novel image dataset for benchmarking machine learning algorithms (2017)", + "60. Xu, H., Wang, Y., Wei, J., Jian, S., Li, Y., Liu, N.: Fascinating supervisory signals and where to find them: Deep anomaly detection with scale learning. arXiv preprint arXiv:2305.16114 (2023)", + "61. Zhou, B., Lapedriza, A., Khosla, A., Oliva, A., Torralba, A.: Places: A 10 million image database for scene recognition. IEEE Transactions on Pattern Analysis and Machine Intelligence pp. 1452-1464 (2017)", + "62. Zhou, C., Paffenroth, R.C.: Anomaly detection with robust deep autoencoders. In: Proceedings of the ACM SIGKDD International Conference on Knowledge Discovery and Data Mining. pp. 665-674 (2017)", + "63. Zong, B., Song, Q., Min, M.R., Cheng, W., Lumezanu, C., Cho, D., Chen, H.: Deep autoencoding gaussian mixture model for unsupervised anomaly detection. In: Proceedings of International Conference on Learning Representations (2018)" + ], + "bbox": [ + 215, + 146, + 785, + 522 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "Z. Liu et al.", + "bbox": [ + 271, + 114, + 352, + 126 + ], + "page_idx": 17 + } +] \ No newline at end of file diff --git a/2024/Rethinking Unsupervised Outlier Detection via Multiple Thresholding/7bb2035e-e079-470b-b56b-69d7d8631ccd_model.json b/2024/Rethinking Unsupervised Outlier Detection via Multiple Thresholding/7bb2035e-e079-470b-b56b-69d7d8631ccd_model.json new file mode 100644 index 0000000000000000000000000000000000000000..517cc5881617028007afb10f550f381123d051b8 --- /dev/null +++ b/2024/Rethinking Unsupervised Outlier Detection via Multiple Thresholding/7bb2035e-e079-470b-b56b-69d7d8631ccd_model.json @@ -0,0 +1,2777 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.232, + 0.141, + 0.773, + 0.187 + ], + "angle": 0, + "content": "Rethinking Unsupervised Outlier Detection via Multiple Thresholding" + }, + { + "type": "text", + "bbox": [ + 0.227, + 0.213, + 0.778, + 0.245 + ], + "angle": 0, + "content": "Zhonghang Liu1, Panzhong Lu2, Guoyang Xie3,4, Zhichao Lu4, and Wen-Yan Lin1" + }, + { + "type": "text", + "bbox": [ + 0.38, + 0.256, + 0.626, + 0.284 + ], + "angle": 0, + "content": "\\(^{1}\\) Singapore Management University \\(^{2}\\) Westlake University" + }, + { + "type": "text", + "bbox": [ + 0.329, + 0.284, + 0.674, + 0.3 + ], + "angle": 0, + "content": "3 Contemporary Amperex Technology Co., Limited" + }, + { + "type": "text", + "bbox": [ + 0.397, + 0.299, + 0.61, + 0.313 + ], + "angle": 0, + "content": "\\(^{4}\\) City University of Hong Kong" + }, + { + "type": "text", + "bbox": [ + 0.397, + 0.313, + 0.609, + 0.327 + ], + "angle": 0, + "content": "zhliu.2020@phdcs.smu.edu.sg" + }, + { + "type": "text", + "bbox": [ + 0.4, + 0.328, + 0.603, + 0.34 + ], + "angle": 0, + "content": "lupanzhong@westlake.edu.cn" + }, + { + "type": "text", + "bbox": [ + 0.422, + 0.342, + 0.582, + 0.355 + ], + "angle": 0, + "content": "guoyang.xie@ieee.org" + }, + { + "type": "text", + "bbox": [ + 0.421, + 0.356, + 0.584, + 0.368 + ], + "angle": 0, + "content": "luzhichao.cn@gmail.com" + }, + { + "type": "text", + "bbox": [ + 0.423, + 0.37, + 0.581, + 0.382 + ], + "angle": 0, + "content": "daniellin@smu.edu.sg" + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.43, + 0.744, + 0.639 + ], + "angle": 0, + "content": "Abstract. In the realm of unsupervised image outlier detection, assigning outlier scores holds greater significance than its subsequent task: thresholding for predicting labels. This is because determining the optimal threshold on non-separable outlier score functions is an ill-posed problem. However, the lack of predicted labels not only hinders some real applications of current outlier detectors but also causes these methods not to be enhanced by leveraging the dataset's self-supervision. To advance existing scoring methods, we propose a multiple thresholding (Multi-T) module. It generates two thresholds that isolate inliers and outliers from the unlabelled target dataset, whereas outliers are employed to obtain better feature representation while inliers provide an uncontaminated manifold. Extensive experiments verify that Multi-T can significantly improve proposed outlier scoring methods. Moreover, Multi-T contributes to a naive distance-based method being state-of-the-art. Code is available at: https://github.com/zhliu-uod/Multi-T." + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.652, + 0.741, + 0.681 + ], + "angle": 0, + "content": "Keywords: Unsupervised outlier detection, multiple thresholding, outlier scoring" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.712, + 0.376, + 0.727 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Which is more important for the unsupervised outlier detection (UOD), outlier score or label? Currently, the mainstream UOD approaches [23, 28, 31, 56] focus on the first, i.e., learning a discriminative outlier score function. However, those recent outlier detectors are usually complex and non-iterative since the absence of predicted labels limits these methods to be further advanced with the dataset's self-supervision. Such that, this work tends to label the unlabelled tar" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.353, + 0.127 + ], + "angle": 0, + "content": "Z. Liu et al." + }, + { + "type": "image", + "bbox": [ + 0.216, + 0.145, + 0.788, + 0.241 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.251, + 0.788, + 0.307 + ], + "angle": 0, + "content": "Fig. 1: Unlike the default UOD paradigm (a) concerned about learning a discriminative score function, our perspective (b) is to explore the dataset's prior knowledge via thresholding, to advance the previously proposed method. (c) illustrates the significant improvement of DeepSVDD [50] with Multi-T module." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.339, + 0.784, + 0.369 + ], + "angle": 0, + "content": "get dataset, thereby leveraging the dataset's prior knowledge to enhance simple outlier scoring methods\\(^{1}\\), achieving state-of-the-art (SOTA) results efficiently." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.37, + 0.785, + 0.505 + ], + "angle": 0, + "content": "To understand the importance of the target dataset's self-supervision, we first decompose the predominant image outlier scoring approaches [9,32,50,51] into two continuous stages: inlier/normal manifold learning and distance/similarity inference. There is no doubt that identifying inliers will contribute to learning an uncontaminated inlier manifold (e.g. the hypersphere of DeepSVDD [50]). Additionally, motivated by Shell Theory [32,33] in the high-dimensional space, image feature representation for distance computation can be significantly improved via outliers-based Shell Normalization [32]. Such that it is critical to obtain the thresholds that separate both inliers and outliers from target datasets." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.507, + 0.785, + 0.597 + ], + "angle": 0, + "content": "To align with our above objective, we present a Multi-T module, which follows a multiple thresholding mechanism. Compared with conventional single thresholding methods [19, 24, 48, 49], Multi-T generates two distinct thresholds that separately isolate outliers and inliers from the target dataset, which is capable of dealing with the inevitable inseparability of the initial given score distribution. This allows us to directly implement Multi-T on a simple outlier score function." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.599, + 0.785, + 0.689 + ], + "angle": 0, + "content": "By understanding the impact of the outlier ratio, Multi-T involves two stages: Uncontaminated inliers. We employ an iterative two-step process (i) identifying a noisy inlier distribution through the analysis of sorted initial score function, i.e., Ergodic-set normalized [34] distance to the mean of target dataset; (ii) subsequently filtering outliers with 3-sigma rule [42]. When converged, we can identify uncontaminated inliers, thereby training an inlier/normal manifold." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.691, + 0.785, + 0.781 + ], + "angle": 0, + "content": "Adaptive outliers. The uncontaminated inliers identification process also outputs a series of outlier threshold candidates. Subsequently, we compare the ranking-index similarity between Ergodic-set Normalization [34] and Shell Normalization [32] on the initial outlier score function. The structural consistency and contrastive properties of the two normalization procedures motivate us to implicitly estimate a rough outlier ratio, thereby detecting the adaptive outliers." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.782, + 0.785, + 0.813 + ], + "angle": 0, + "content": "The Multi-T module is training-free, highly efficient and grants the adaptiveness to choose thresholds suited to downstream implementations. As evaluated" + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.825, + 0.655, + 0.84 + ], + "angle": 0, + "content": "1 Outlier scoring methods refers to the classical outlier detectors." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.265, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Rethinking Unsupervised Outlier Detection via Multiple Thresholding" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.252 + ], + "angle": 0, + "content": "with comprehensive experimental settings, Multi-T helps to boost significant performance (efficacy and efficiency) improvement. For example, the AUC score of DeepSVDD [50] is improved from 0.622 to 0.925 on STL-10 dataset [15]. Moreover, the naive distance-to-the-target dataset's mean [32] metric integrated with Multi-T achieves SOTA results with only 1.2 second consumption for 10,000 ResNet-50 [21] samples that are orders of magnitude faster than baselines. Our primary contributions are summarized below:" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.253, + 0.784, + 0.282 + ], + "angle": 0, + "content": "- We introduce a novel perspective for UOD about enhancing previously proposed scoring methods via thresholding on a simple initial outlier score function." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.283, + 0.784, + 0.311 + ], + "angle": 0, + "content": "- We present a multiple thresholding (Multi-T) module that generates two different thresholds to separate both inliers and outliers from the target dataset." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.313, + 0.784, + 0.342 + ], + "angle": 0, + "content": "- The efficacy of integrated scoring methods can be significantly improved without the external complexity increase, as evaluated in extensive experiments." + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.253, + 0.784, + 0.342 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.368, + 0.388, + 0.383 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.401, + 0.525, + 0.417 + ], + "angle": 0, + "content": "2.1 Unsupervised Outlier Detection" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.428, + 0.788, + 0.684 + ], + "angle": 0, + "content": "Classic UOD task aims to assign an outlier score/likelihood to an image sample. The recent models can be divided into inlier-manifold learning and outlier exposure. Inlier-manifold learning assumes that inliers are the majority. Deep learning-based outlier detectors, as illustrated in studies [14, 28, 30, 38, 56, 58, 62, 63], typically focus on reducing the dimensionality of image data by projecting it into a latent and discriminative space. Additionally, various statistics-based methods model the inlier manifold using discrimination-based [32, 34] or density-based approaches [11, 26, 31, 40]. In applications where outliers are not the minority [13, 57], different approaches are required. Some outlier exposure methods [22, 36] employ out-of-distribution (OOD) data to train networks for detecting unseen OOD samples. For example, Shell-Renormalization [32] employs predicted outlier candidates to iteratively refine feature representation. It is particularly useful in scenarios where classic manifold learners, assuming outliers as a minority, may not be effective. However, these two paradigms are separable, this work will attempt to design a thresholding framework that generates both inliers and outliers, leveraging the benefits of both. Inlier-manifold learning and outlier exposure." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.708, + 0.373, + 0.724 + ], + "angle": 0, + "content": "2.2 Thresholding" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.734, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Some traditional outlier detection methods [32, 35, 52], provide both the outlier score and its corresponding threshold concurrently. In this context, the threshold manifests as a hypersphere that accepts the inliers while rejecting those outliers. Unquestionably, discerning it presents its challenges in high-dimensional space [34]. In response, we venture into an alternate perspective. Some other threshold detectors are mostly based on statistical analysis that can be utilized for any given \\(1 - d\\) outlier score function [28, 31, 32]. The representative" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.355, + 0.128 + ], + "angle": 0, + "content": "Z. Liu et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.239 + ], + "angle": 0, + "content": "works involve Kernel-based [43, 48]; Curve-based [17]; Normality-based: [5, 10]; Filtering-based [19, 55]; Statistical-based [1, 3, 4, 6-8, 11, 12, 16, 24, 37, 49] and Transformation-based [25, 46]. In practice, the default thresholding follows a single thresholding paradigm is not realistic since the perfect discriminativeness of the outlier score function is usually not assumed. Therefore, this work introduces a new perspective: multiple thresholding." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.257, + 0.347, + 0.274 + ], + "angle": 0, + "content": "3 Overview" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.286, + 0.422, + 0.3 + ], + "angle": 0, + "content": "3.1 Problem Definition" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.308, + 0.789, + 0.472 + ], + "angle": 0, + "content": "Before introducing our method, we first formally define the problem. The target dataset \\(\\mathbb{Z} = \\mathbb{Z}_{in}\\cup \\mathbb{Z}_{out}\\) involves \\(n\\) unlabelled images, here \\(\\mathbb{Z}_{in}\\) and \\(\\mathbb{Z}_{out}\\) represent inliers (normal data) and outliers (anomalous data), respectively. The outlier (contamination) ratio \\(\\gamma \\in (0,1)\\) is denoted as \\(\\frac{\\#\\mathbb{Z}_{out}}{\\#\\mathbb{Z}_{in} + \\#\\mathbb{Z}_{out}}\\). Notably, this setup differs from the related semi-supervised outlier detection task by eliminating the need for a prior train/test split for the target dataset. Aligning with previous works [28, 32, 34], the raw images are extracted to feature representation \\(\\mathbf{X} = \\mathbf{X}_{\\mathrm{in}}\\cup \\mathbf{X}_{\\mathrm{out}} = \\{\\mathbf{x}_i\\}_{i = 1}^n\\) with pixel representation [32, 34] or pre-trained neural networks (e.g., ResNet [21] and CLIP [45]). The main objective of UOD is to create an outlier score function \\(\\mathrm{F}(\\cdot)\\) to evaluate the likelihood of an image feature \\(\\mathbf{x}_i\\in \\mathbf{X}\\) being outliers:" + }, + { + "type": "equation", + "bbox": [ + 0.417, + 0.474, + 0.788, + 0.505 + ], + "angle": 0, + "content": "\\[\n\\mathrm {Y} (\\mathbf {x} _ {i}) = \\left\\{ \\begin{array}{l} 0, \\mathrm {F} (\\mathbf {x} _ {i}) < \\phi \\\\ 1, \\mathrm {F} (\\mathbf {x} _ {i}) \\geq \\phi , \\end{array} \\right. \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.508, + 0.788, + 0.555 + ], + "angle": 0, + "content": "where \\( \\mathrm{Y} = 0 \\) (inlier)/1 (outlier) refers to predicted labels and \\( \\phi \\) is the threshold (decision boundary). In this study, we not only center on evaluating the ranking accuracy of \\( \\mathrm{F}(\\cdot) \\) [28,32,34,50], but also measure the efficacy of the threshold \\( \\phi \\)." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.572, + 0.405, + 0.587 + ], + "angle": 0, + "content": "3.2 Rethinking UOD" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.594, + 0.787, + 0.638 + ], + "angle": 0, + "content": "As there is no prior train/test split for the target dataset, a classic outlier scoring method \\( \\mathrm{M}(\\cdot) \\) predicts the outlier score function \\( \\mathrm{F}(\\mathbf{X}) = \\{\\mathrm{F}(\\mathbf{x}_i)\\}_{i=1}^n \\) follows the below process:" + }, + { + "type": "equation", + "bbox": [ + 0.396, + 0.64, + 0.785, + 0.655 + ], + "angle": 0, + "content": "\\[\n\\mathrm {F} (\\mathbf {X}) = \\mathrm {M}. \\operatorname {f i t} (\\mathbf {X}). \\operatorname {p r e d i c t} (\\mathbf {X}), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.66, + 0.788, + 0.707 + ], + "angle": 0, + "content": "where \\(\\mathrm{fit}(\\cdot)\\) and \\(\\mathrm{predict}(\\cdot)\\) refer to fitting the target dataset and predicting outlier scores, respectively. Motivated by previous research [32, 33], Shell normalization S-norm \\((\\cdot)\\) illustrated below is an ideal feature de-noising/refining paradigm." + }, + { + "type": "equation", + "bbox": [ + 0.235, + 0.714, + 0.786, + 0.765 + ], + "angle": 0, + "content": "\\[\n\\operatorname {S - n o r m} \\left(\\mathbf {x} _ {i}, \\mathbf {v} _ {\\mathrm {S}}\\right) = \\frac {\\mathbf {x} _ {i} - \\mathbf {v} _ {\\mathrm {S}}}{\\left| \\left| \\mathbf {x} _ {i} - \\mathbf {v} _ {\\mathrm {S}} \\right| \\right| _ {2}}, \\mathbf {v} _ {\\mathrm {S}} = \\left[ \\frac {1}{n} \\sum_ {i = 1} ^ {n} \\mathbf {X} _ {\\text {o u t}} [ i ] [ 1 ], \\dots , \\frac {1}{n} \\sum_ {i = 1} ^ {n} \\mathbf {X} _ {\\text {o u t}} [ i ] [ d ] \\right], \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.766, + 0.788, + 0.843 + ], + "angle": 0, + "content": "where \\(||\\cdot ||_2\\) refers to \\(\\ell_2\\)-norm and \\(d\\) is the feature dimension. Thus, our first stage is to identify outlier candidates \\(\\mathbf{X}_{\\mathrm{out}}^{\\prime}\\). Subsequently, as how some one-class learning-based methods (e.g., OCVSM [51], DeepSVDD [50]) usually perform, we secondly predict the inliers \\(\\mathbf{X}_{\\mathrm{in}}^{\\prime}\\) to fit an inlier/normal manifold. Thus, our objective is to identify both \\(\\mathbf{X}_{\\mathrm{in}}^{\\prime}\\) and \\(\\mathbf{X}_{\\mathrm{out}}^{\\prime}\\)." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.265, + 0.115, + 0.733, + 0.13 + ], + "angle": 0, + "content": "Rethinking Unsupervised Outlier Detection via Multiple Thresholding" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "image", + "bbox": [ + 0.214, + 0.144, + 0.788, + 0.384 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.393, + 0.788, + 0.464 + ], + "angle": 0, + "content": "Fig. 2: The overall paradigm of adopting the multi-thresholds learning (Multi-T) module to advance the existing outlier scoring methods. (a) The preparation consists of feature extraction and an initial outlier score function. (b) Visualization of our Multi-T module. (c) Integrating the predicted inliers and outliers with the previously proposed outlier detectors and obtaining an enhanced outlier score function." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.494, + 0.33, + 0.509 + ], + "angle": 0, + "content": "4 Method" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.531, + 0.788, + 0.593 + ], + "angle": 0, + "content": "Our solution is displayed as below subsections. Sec. 4.1: we introduce the initial outlier score function; Sec. 4.2: we present the Multi-T module, to generate two thresholds for separating inliers and outliers from the target dataset; Sec. 4.3: we leverage predicted inliers and outliers to advance previously proposed methods." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.62, + 0.506, + 0.635 + ], + "angle": 0, + "content": "4.1 Initial Outlier Score Function" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.651, + 0.788, + 0.728 + ], + "angle": 0, + "content": "In our approach, we employ LVAD [34], one of the UOD SOTAs, as the initial outlier score function. LVAD assumes each image feature \\(\\mathbf{x}_i \\in \\mathbf{X}\\) comes from one of \\(T\\) high-dimensional generative processes. Thus, the outlier score becomes a sum of the weighted distance of the given instance, \\(\\mathbf{x}_i\\), arising from each generative process \\(\\{w_t, \\mathbf{m}_t \\mid t \\in \\{1, \\dots, T\\}\\}\\), which can be simplified as:" + }, + { + "type": "equation", + "bbox": [ + 0.31, + 0.74, + 0.787, + 0.782 + ], + "angle": 0, + "content": "\\[\n\\mathrm {F} _ {\\mathrm {L V A D}} (\\mathbf {x} _ {i}) = \\sum_ {t = 1} ^ {T} w _ {t} \\cdot \\operatorname {D i s t} (\\mathrm {E} - \\operatorname {n o r m} (\\mathbf {x} _ {i}), \\mathrm {E} - \\operatorname {n o r m} (\\mathbf {m} _ {t})), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.788, + 0.843 + ], + "angle": 0, + "content": "where \\(\\mathrm{Dist}(\\cdot)\\) refers to the \\(\\ell_2\\)-norm metric (Euclidean distance), \\(w_{t}\\) is the weight of the \\(t\\)-th generative process, \\(\\mathbf{m}_t\\) is the \\(t\\)-th generative process's centroid and \\(\\mathrm{E - norm}(\\cdot)\\) refers to the Ergodic-set normalization, an effective and outlier ratio" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.355, + 0.128 + ], + "angle": 0, + "content": "Z. Liu et al." + }, + { + "type": "code_caption", + "bbox": [ + 0.23, + 0.151, + 0.381, + 0.165 + ], + "angle": 0, + "content": "Algorithm 1: MTL" + }, + { + "type": "algorithm", + "bbox": [ + 0.219, + 0.167, + 0.788, + 0.553 + ], + "angle": 0, + "content": "1 Input: Initial Outlier Score Function: \\(\\mathrm{F_{init}(X)}\\) , Outlier Scoring Method: M(·) \n2 Output: Advanced Outlier Score Function: \\(\\mathrm{F_{M + Multi - T}(X)}\\) \n3 // Stage 1: Identifying Uncontaminated Inliers \n4 for \\(b\\) in iterations do \n5 if not converged then \n6 \\(\\begin{array}{rlr}\\widehat{\\mathrm{F}}_{\\mathrm{init}}(\\mathbf{X}) = \\mathrm{sort}(\\mathrm{F}_{\\mathrm{init}}(\\mathbf{X})) & \\quad / & \\mathrm{sort - transform(Eq.9)}\\\\ I^b = \\left\\{\\widehat{\\mathrm{F}}_{\\mathrm{init}}(\\mathbf{x}_i)|i < \\max_i\\left\\{g(a_i) > \\widehat{\\mathrm{F}}_{\\mathrm{init}}(\\mathbf{x}_i)\\right\\} \\right\\} & \\quad / & \\mathrm{linear - fit(Eq.10)}\\\\ \\phi_{\\mathrm{out}}^b = \\mathrm{mean}(I^b) + 3\\cdot \\mathrm{std}(I^b) & \\quad / & \\mathrm{outlier~threshold~(Eq.11)}\\\\ I^{b + 1} = I^b.\\mathrm{remove}(\\{\\mathrm{F}_{\\mathrm{init}}(\\mathbf{x}_i)|\\mathrm{F}_{\\mathrm{init}}(\\mathbf{x}_i) > \\phi_{\\mathrm{out}}^b\\}) & \\quad / & \\mathrm{filtering(Eq.11)}\\\\ b = b + 1 \\end{array}\\) \n10 \n11 Converged \\(b:b^{*}\\) \n12 \\(\\phi_{\\mathrm{in}} = \\max (I^{b^{*}})\\) // inlier threshold(Eq.12) \n13 \\(\\mathbf{X}_{\\mathrm{in}}^{\\prime} = \\{\\mathbf{x}_{i}|F_{\\mathrm{init}}(\\mathbf{x}_{i})\\leq \\phi_{\\mathrm{in}}\\}\\) // predicted inliers(Eq.13) \n14 // Stage 2: Identifying Adaptive Outliers \n15 \\(\\rho \\gets\\) similarity(Fs(X),FE(X)) // outlier ratio estimation(Eq.15,16) \n16 if \\(\\rho >0.3\\) then \n17 \\(\\phi_{\\mathrm{out}} = \\phi_{\\mathrm{out}}^{*}\\) \n18 else \n19 if \\(0.1\\leq \\rho \\leq 0.3\\) then \n20 \\(\\phi_{\\mathrm{out}} = \\phi_{\\mathrm{out}}^{1}\\) \n21 else \n22 \\(\\begin{array}{rlr}\\lfloor & \\phi_{\\mathrm{out}} = \\phi_{\\mathrm{out}}^{0}\\\\ & \\end{array}\\) \n23 \\(\\mathbf{X}_{\\mathrm{out}}^{\\prime} = \\{\\mathbf{x}_{i}|F_{\\mathrm{init}}(\\mathbf{x}_{i}) > \\phi_{\\mathrm{out}}\\}\\) // predicted outliers(Eq.20) \n24 // Stage 3: Adopting Multi-T to Outlier Scoring Method \n25 return M.fit({S-norm(xi,v's)|xi∈X'in}).predict({S-norm(xi,v's)|xi∈X})" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.591, + 0.625, + 0.607 + ], + "angle": 0, + "content": "invariant normalization procedure, illustrated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.222, + 0.621, + 0.785, + 0.677 + ], + "angle": 0, + "content": "\\[\n\\operatorname {E} - \\operatorname {n o r m} \\left(\\mathbf {x} _ {i}, \\mathbf {v} _ {\\mathrm {E}}\\right) = \\frac {\\mathbf {x} _ {i} - \\mathbf {v} _ {\\mathrm {E}}}{\\left| \\left| \\mathbf {x} _ {i} - \\mathbf {v} _ {\\mathrm {E}} \\right| \\right| _ {2}}, \\mathbf {v} _ {\\mathrm {E}} = \\left[ \\frac {1}{n \\cdot d} \\sum_ {i = 1} ^ {n} \\sum_ {j = 1} ^ {d} \\mathbf {x} _ {i, j}, \\dots , \\frac {1}{n \\cdot d} \\sum_ {i = 1} ^ {n} \\sum_ {j = 1} ^ {d} \\mathbf {x} _ {i, j} \\right], \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.678, + 0.788, + 0.723 + ], + "angle": 0, + "content": "where \\(\\mathbf{v}_{\\mathrm{E}}\\) is its reference vector and \\(j\\) is the dimension index. To gain higher efficiency and maintain efficacy, we set \\(T = 1\\), such that the initial outlier score function (LVAD-S) is formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.234, + 0.737, + 0.786, + 0.797 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathrm {F} _ {\\text {i n i t}} (\\mathbf {X}) = \\mathrm {F} _ {\\text {L V A D - S}} (\\mathbf {X}) = \\left. \\left\\{\\operatorname {D i s t} \\left(\\mathrm {E} - \\operatorname {n o r m} \\left(\\mathbf {x} _ {i}, \\mathbf {v} _ {\\mathrm {E}}\\right), \\operatorname {E} - \\operatorname {n o r m} \\left(\\mathbf {m} _ {\\mathbf {X}}, \\mathbf {v} _ {\\mathrm {E}}\\right)\\right) \\right\\} _ {i = 1} ^ {n}, \\right. \\\\ \\mathbf {m} _ {\\mathbf {X}} = \\left[ \\frac {1}{n} \\sum_ {i = 1} ^ {n} \\mathbf {x} _ {i, 1}, \\dots , \\frac {1}{n} \\sum_ {i = 1} ^ {n} \\mathbf {x} _ {i, d} \\right], \\tag {6} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.842 + ], + "angle": 0, + "content": "where \\(\\mathbf{m}_{\\mathbf{X}}\\) is the mean of target dataset's features. Despite its simplicity, Eq. 6 is still a reliable outlier scoring method [34]." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.265, + 0.115, + 0.733, + 0.131 + ], + "angle": 0, + "content": "Rethinking Unsupervised Outlier Detection via Multiple Thresholding" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.147, + 0.525, + 0.163 + ], + "angle": 0, + "content": "4.2 Multi-T: Multiple Thresholding" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.171, + 0.784, + 0.262 + ], + "angle": 0, + "content": "Motivation. In the field of UOD, the conventional thresholding paradigm is to learn a single threshold \\(\\phi\\) on the initial outlier score function \\(\\mathrm{F}_{\\mathrm{init}}(\\mathbf{X})\\). However, it ignores the impact of the outlier ratio \\(\\gamma\\). As the mean of target dataset \\(\\mathbf{m}_{\\mathbf{X}}\\) will be shifted to outliers when \\(\\gamma\\) becomes higher, i.e., mean-shift problem [47], there will be an inevitable overlap between inlier and outlier score distributions. So we define the distribution of the initial outlier score function \\(D\\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.315, + 0.273, + 0.786, + 0.289 + ], + "angle": 0, + "content": "\\[\nD = \\operatorname {F} _ {\\text {i n i t}} (\\mathbf {X}) = \\operatorname {F} _ {\\text {i n i t}} \\left(\\mathbf {X} _ {\\text {i n}}\\right) \\cup \\operatorname {F} _ {\\text {i n i t}} \\left(\\mathbf {X} _ {\\text {o u t}}\\right) = I \\cup A \\cup O, \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.298, + 0.784, + 0.373 + ], + "angle": 0, + "content": "where \\( I \\) and \\( O \\) refer to the uncontaminated inlier and outlier score distributions, \\( A \\) refers to the overlap. Such that we are concerned about multi-thresholding instead of the conventional way that explicitly finding a single threshold between \\( \\mathrm{F_{init}(\\mathbf{X}_{in})} \\) and \\( \\mathrm{F_{init}(\\mathbf{X}_{out})} \\). Specifically, our objective is identifying the inlier threshold \\( \\phi_{\\mathrm{in}} \\) and outlier threshold \\( \\phi_{\\mathrm{out}} \\) that isolate \\( I \\) and \\( O \\) from \\( D \\)." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.374, + 0.787, + 0.435 + ], + "angle": 0, + "content": "Identifying uncontaminated inliers. We identify \\(I\\) by understanding the initial outlier score function (Eq. 6), which computes the normalized \\(\\ell_2\\)-norm of each instance \\(\\mathbf{x}_i \\in \\mathbf{X}\\) relative to the mean of \\(\\mathbf{X}\\). Inspired by Shell Theory [32], \\(I\\) satisfies a Gaussian-like distribution, the classic statistical estimation of \\(I\\) is:" + }, + { + "type": "equation", + "bbox": [ + 0.327, + 0.445, + 0.785, + 0.461 + ], + "angle": 0, + "content": "\\[\nI = \\left\\{\\mathrm {F} _ {\\text {i n i t}} (\\mathbf {x} _ {i}) \\mid \\mathrm {F} _ {\\text {i n i t}} (\\mathbf {x} _ {i}) < \\operatorname {m e a n} (D) + k \\cdot \\operatorname {s t d} (D) \\right\\}, \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.47, + 0.784, + 0.547 + ], + "angle": 0, + "content": "where \\(\\mathrm{mean}(D)\\) and \\(\\mathrm{std}(D)\\) refer to the mean and standard deviation, \\(k\\) is the hyper-parameter of the \" \\(k\\)-sigma\" rule. In practice, optimizing \\(k\\) without supervision is challenging. Thus, we convert to a non-parametric way by utilizing an ascending sort projection \\(\\mathrm{sort}(\\cdot)\\) to \\(\\mathrm{F}_{\\mathrm{init}}(\\mathbf{X})\\), which projects \\(D\\) into a 2-\\(d\\) space (\\(x\\)-axis: instance index, \\(y\\)-axis: sorted outlier score), the sorted outlier score is:" + }, + { + "type": "equation", + "bbox": [ + 0.321, + 0.557, + 0.785, + 0.575 + ], + "angle": 0, + "content": "\\[\n\\widehat {\\mathrm {F}} _ {\\text {i n i t}} (\\mathbf {X}) = \\operatorname {s o r t} \\left(\\widehat {\\mathrm {F}} _ {\\text {i n i t}} (\\mathbf {X})\\right), i. e., \\widehat {\\mathrm {F}} _ {\\text {i n i t}} [ i + 1 ] > \\widehat {\\mathrm {F}} _ {\\text {i n i t}} [ i ]. \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.585, + 0.784, + 0.617 + ], + "angle": 0, + "content": "Intuitively, \\(\\widehat{I} = \\mathrm{sort}(I)\\) can be fitted with a naive linear regressor \\(g(a) = \\boldsymbol{\\beta}^{\\top} \\cdot a = \\beta_{0} + \\beta_{1} \\cdot a\\), as shown in Fig. 2 (b). So we identify the potential inliers as:" + }, + { + "type": "equation", + "bbox": [ + 0.235, + 0.627, + 0.785, + 0.668 + ], + "angle": 0, + "content": "\\[\nI = \\left\\{\\widehat {\\mathrm {F}} _ {\\text {i n i t}} (\\mathbf {x} _ {i}) | i < \\max _ {i} \\left\\{g \\left(a _ {i}\\right) > \\widehat {\\mathrm {F}} _ {\\text {i n i t}} (\\mathbf {x} _ {i}) \\right\\} \\right\\}, s. t. \\min _ {\\beta} \\left| \\left| a ^ {\\top} \\beta - \\widehat {\\mathrm {F}} _ {\\text {i n i t}} (\\mathbf {X}) \\right| \\right| _ {2}, \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.668, + 0.784, + 0.697 + ], + "angle": 0, + "content": "where \\(\\beta\\) is the coefficients. Inevitably, this fitting process will be shifted with outliers, which is addressed together with the following outliers' identification." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.697, + 0.784, + 0.742 + ], + "angle": 0, + "content": "Identifying adaptive outliers. In statistics, 3-sigma rule [42] declares variables \\( C \\) has a large probability within three standard deviations from the mean if \\( C \\) follows a Gaussian-like distribution. So we employ it with two aspects:" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.743, + 0.706, + 0.759 + ], + "angle": 0, + "content": "(i) iteratively filter outliers during the phase of inliers identification:" + }, + { + "type": "equation", + "bbox": [ + 0.229, + 0.768, + 0.785, + 0.799 + ], + "angle": 0, + "content": "\\[\n\\phi_ {\\text {o u t}} ^ {b} = \\operatorname {m e a n} \\left(I ^ {b}\\right) + 3 \\cdot \\operatorname {s t d} \\left(I ^ {b}\\right), I ^ {b + 1} = I ^ {b}. \\text {r e m o v e} \\left(\\left\\{\\mathrm {F} _ {\\text {i n i t}} \\left(\\mathbf {x} _ {i}\\right) \\mid \\mathrm {F} _ {\\text {i n i t}} \\left(\\mathbf {x} _ {i}\\right) > \\phi_ {\\text {o u t}} ^ {b} \\right\\}\\right), \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.8, + 0.762, + 0.814 + ], + "angle": 0, + "content": "where \\( b \\) refers to the \\( b \\)-th iteration. When converged, the inlier threshold is:" + }, + { + "type": "equation", + "bbox": [ + 0.446, + 0.825, + 0.785, + 0.841 + ], + "angle": 0, + "content": "\\[\n\\phi_ {\\text {i n}} = \\max \\left(I ^ {b ^ {*}}\\right), \\tag {12}\n\\]" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.355, + 0.128 + ], + "angle": 0, + "content": "Z. Liu et al." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.146, + 0.784, + 0.163 + ], + "angle": 0, + "content": "where \\(I^{b^*}\\) refers to the converged inlier prediction and we can identify inliers as:" + }, + { + "type": "equation", + "bbox": [ + 0.407, + 0.178, + 0.786, + 0.194 + ], + "angle": 0, + "content": "\\[\n\\mathbf {X} _ {\\mathrm {i n}} ^ {\\prime} = \\left\\{\\mathbf {x} _ {i} \\mid \\mathrm {F} _ {\\text {i n i t}} \\left(\\mathbf {x} _ {i}\\right) \\leq \\phi_ {\\mathrm {i n}} \\right\\}. \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.198, + 0.747, + 0.214 + ], + "angle": 0, + "content": "(ii) During the above phase, we estimate the outlier threshold candidates:" + }, + { + "type": "equation", + "bbox": [ + 0.283, + 0.223, + 0.786, + 0.243 + ], + "angle": 0, + "content": "\\[\n\\phi_ {\\mathrm {o u t}} ^ {1} = \\operatorname {m e a n} \\left(I ^ {1}\\right) + 3 \\cdot \\operatorname {s t d} \\left(I ^ {1}\\right), \\phi_ {\\mathrm {o u t}} ^ {*} = \\operatorname {m e a n} \\left(I ^ {b ^ {*}}\\right) + 3 \\cdot \\operatorname {s t d} \\left(I ^ {b ^ {*}}\\right), \\tag {14}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.249, + 0.787, + 0.342 + ], + "angle": 0, + "content": "where \\(\\phi_{\\mathrm{out}}^{1}\\), \\(\\phi_{\\mathrm{out}}^{*}\\) refer to the outlier threshold candidates at the first and final iterations, respectively, i.e., \\(\\phi_{\\mathrm{out}}^{1} > \\phi_{\\mathrm{out}}^{*}\\). To obtain a \\(\\gamma\\)-adaptive outlier threshold, we consider the impact and relationship of two mentioned normalization procedures: Shell Normalization S-norm(\\(\\cdot\\)) (Eq. 3) and Ergodic-set Normalization E-norm(\\(\\cdot\\)) (Eq. 5). Both them follow a similar normalization formula (denominator: \\(\\ell_2\\)-norm), the difference is their reference vectors \\(\\mathbf{v}_{\\mathrm{S}}'\\) and \\(\\mathbf{v}_{\\mathrm{E}}\\), contrast as:" + }, + { + "type": "equation", + "bbox": [ + 0.339, + 0.353, + 0.786, + 0.401 + ], + "angle": 0, + "content": "\\[\n\\mathbf {v} _ {\\mathrm {S}} ^ {\\prime} = \\left[ \\frac {1}{n} \\sum_ {i = 1} ^ {n} \\mathbf {X} _ {\\text {o u t}} ^ {\\prime} [ i ] [ 1 ], \\dots , \\frac {1}{n} \\sum_ {i = 1} ^ {n} \\mathbf {X} _ {\\text {o u t}} ^ {\\prime} [ i ] [ d ] \\right], \\tag {15}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.337, + 0.395, + 0.669, + 0.436 + ], + "angle": 0, + "content": "\\[\n\\mathbf {v} _ {\\mathrm {E}} = \\left[ \\frac {1}{n \\cdot d} \\sum_ {i = 1} ^ {n} \\sum_ {j = 1} ^ {d} \\mathbf {x} _ {i, j}, \\dots , \\frac {1}{n \\cdot d} \\sum_ {i = 1} ^ {n} \\sum_ {j = 1} ^ {d} \\mathbf {x} _ {i, j} \\right].\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.44, + 0.787, + 0.502 + ], + "angle": 0, + "content": "Obviously, Shell normalization is strongly subjective to the efficacy of outlier prediction \\(\\mathbf{X}_{\\mathrm{out}}^{\\prime}\\) while Ergodic-set Normalization is independent with \\(\\gamma\\) that is a sub-optimal but stable operation. To compare these two normalization procedures, we first employ them to Eq. 6, and obtain two outlier score functions:" + }, + { + "type": "equation", + "bbox": [ + 0.308, + 0.513, + 0.786, + 0.538 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathrm {F} _ {\\mathrm {S}} (\\mathbf {X}) = \\left\\{\\operatorname {D i s t} \\left(\\mathrm {S} - \\operatorname {n o r m} \\left(\\mathbf {x} _ {i}, \\mathbf {v} _ {\\mathrm {S}} ^ {\\prime}\\right), \\operatorname {S} - \\operatorname {n o r m} \\left(\\mathbf {m} _ {\\mathbf {X}}, \\mathbf {v} _ {\\mathrm {S}} ^ {\\prime}\\right)\\right) \\right\\} _ {i = 1} ^ {n}, \\\\ \\mathrm {F} _ {\\mathrm {E}} (\\mathbf {X}) = \\left\\{\\operatorname {D i s t} \\left(\\mathrm {E} - \\operatorname {n o r m} \\left(\\mathbf {v} _ {\\mathrm {E}} - \\mathbf {v} _ {\\mathrm {I}}\\right), \\operatorname {E} - \\operatorname {n o r m} \\left(\\mathbf {m} _ {\\mathbf {E}} - \\mathbf {v} _ {\\mathrm {I}}\\right)\\right) \\right\\} ^ {n}. \\end{array} \\tag {16}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.307, + 0.532, + 0.697, + 0.548 + ], + "angle": 0, + "content": "\\[\n\\mathrm {F} _ {\\mathrm {E}} (\\mathbf {X}) = \\left\\{\\operatorname {D i s t} \\left(\\operatorname {E} - \\operatorname {n o r m} \\left(\\mathbf {x} _ {i}, \\mathbf {v} _ {\\mathrm {E}}\\right), \\operatorname {E} - \\operatorname {n o r m} \\left(\\mathbf {m} _ {\\mathbf {X}}, \\mathbf {v} _ {\\mathrm {E}}\\right)\\right) \\right\\} _ {i = 1} ^ {n},\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.552, + 0.788, + 0.721 + ], + "angle": 0, + "content": "where \\(\\mathbf{v}_{\\mathrm{S}}^{\\prime}\\) comes from \\(\\mathbf{X}_{\\mathrm{out}}^{\\prime}\\) obtained by MAD [6] (Shell Normalization's default thresholding method that is only effective on high- \\(\\gamma\\)). Subsequently, two outlier score functions \\(\\mathrm{F_S}(\\mathbf{X})\\) and \\(\\mathrm{F_E}(\\mathbf{X})\\) are arranged in ascending order, and we denote the ranked-index lists as \\(\\mathcal{R}_{\\mathrm{S}}\\) and \\(\\mathcal{R}_{\\mathrm{E}}\\). The outlier ratio \\(\\gamma\\) can be approximately estimated with the similarity between \\(\\mathcal{R}_{\\mathrm{S}}\\) and \\(\\mathcal{R}_{\\mathrm{E}}\\) since \\(\\mathrm{F_S}(\\mathbf{X})\\) and \\(\\mathrm{F_E}(\\mathbf{X})\\) share a consistent structure (distance-to-the-mean), making \\(\\mathcal{R}_{\\mathrm{S}}\\) and \\(\\mathcal{R}_{\\mathrm{E}}\\) comparable. Besides, \\(\\mathrm{F_E}(\\mathbf{X})\\) serves as a reliably effective baseline. When \\(\\gamma\\) is low, S-norm \\((\\mathbf{X},\\mathbf{v}_{\\mathrm{S}}^{\\prime})\\) tends to under-perform, lead the similarity between \\(\\mathcal{R}_{\\mathrm{S}}\\) and \\(\\mathcal{R}_{\\mathrm{E}}\\) becomes low. In contrast, with a high \\(\\gamma\\), S-norm \\((\\mathbf{X},\\mathbf{v}_{\\mathrm{S}}^{\\prime})\\) merely refines the ranking of a small number of outliers, indicative of high similarity. The similarity \\(\\rho \\in [-1,1]\\) is computed with the Pearson correlation coefficient:" + }, + { + "type": "equation", + "bbox": [ + 0.419, + 0.729, + 0.786, + 0.763 + ], + "angle": 0, + "content": "\\[\n\\rho = \\frac {\\operatorname {c o v} \\left(\\mathcal {R} _ {\\mathrm {S}} , \\mathcal {R} _ {\\mathrm {E}}\\right)}{\\operatorname {s t d} \\left(\\mathcal {R} _ {\\mathrm {S}}\\right) \\cdot \\operatorname {s t d} \\left(\\mathcal {R} _ {\\mathrm {E}}\\right)}, \\tag {17}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.769, + 0.787, + 0.815 + ], + "angle": 0, + "content": "where \\(\\operatorname{cov}(\\mathcal{R}_{\\mathrm{S}},\\mathcal{R}_{\\mathrm{E}})\\) and \\(\\mathrm{std}(\\mathcal{R}_{\\mathrm{S}}) / \\mathrm{std}(\\mathcal{R}_{\\mathrm{E}})\\) are the covariance and standard deviation of the rank variables. If the similarity is relatively low (\\(< 0.1\\)), we compute the 3-sigma of the entire outlier score distribution \\(D\\) as the outlier threshold:" + }, + { + "type": "equation", + "bbox": [ + 0.399, + 0.825, + 0.786, + 0.843 + ], + "angle": 0, + "content": "\\[\n\\phi_ {\\text {o u t}} ^ {0} = \\operatorname {m e a n} (D) + 3 \\cdot \\operatorname {s t d} (D). \\tag {18}\n\\]" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.265, + 0.115, + 0.733, + 0.131 + ], + "angle": 0, + "content": "Rethinking Unsupervised Outlier Detection via Multiple Thresholding" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.147, + 0.784, + 0.209 + ], + "angle": 0, + "content": "It is a conservative estimator for outliers since the above linear fitting might suffer from some challenges at low- \\(\\gamma\\) since there is no explicit outlier score distribution. If the similarity is large (\\(>0.3\\)), we choose the relatively smooth outlier threshold \\(\\phi_{\\mathrm{out}}^{*}\\), while \\(\\phi_{\\mathrm{out}}^{1}\\) is used for middle- \\(\\gamma\\) cases. The outlier threshold is described as:" + }, + { + "type": "equation", + "bbox": [ + 0.388, + 0.22, + 0.784, + 0.266 + ], + "angle": 0, + "content": "\\[\n\\phi_ {\\text {o u t}} = \\left\\{ \\begin{array}{l} \\phi_ {\\text {o u t}} ^ {*}, \\text {i f} \\rho > 0. 3 \\\\ \\phi_ {\\text {o u t}} ^ {1}, \\text {i f} 0. 1 \\leq \\rho \\leq 0. 3 \\\\ \\phi_ {\\text {o u t}} ^ {0}, \\text {o t h e r w i s e} \\end{array} \\right. \\tag {19}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.269, + 0.744, + 0.283 + ], + "angle": 0, + "content": "Therefore, we can subsequently identify outlier ratio invariant outliers as:" + }, + { + "type": "equation", + "bbox": [ + 0.399, + 0.291, + 0.784, + 0.307 + ], + "angle": 0, + "content": "\\[\n\\mathbf {X} _ {\\text {o u t}} ^ {\\prime} = \\left\\{\\mathbf {x} _ {i} \\mid \\mathrm {F} _ {\\text {i n i t}} \\left(\\mathbf {x} _ {i}\\right) > \\phi_ {\\text {o u t}} \\right\\}. \\tag {20}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.324, + 0.393, + 0.339 + ], + "angle": 0, + "content": "4.3 Implementation" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.345, + 0.784, + 0.361 + ], + "angle": 0, + "content": "Firstly, we can directly employ Multi-T with the distance-to-the-mean paradigm:" + }, + { + "type": "equation", + "bbox": [ + 0.288, + 0.372, + 0.717, + 0.389 + ], + "angle": 0, + "content": "\\[\n\\mathrm {F} _ {\\mathrm {M u l t i - T}} (\\mathbf {X}) = \\{\\mathrm {D i s t} (\\mathrm {S - n o r m} (\\mathbf {x} _ {i}, \\mathbf {v} _ {\\mathrm {S}} ^ {\\prime}), \\mathrm {S - n o r m} (\\mathbf {m} _ {\\mathbf {X} _ {\\mathrm {i n}} ^ {\\prime}}, \\mathbf {v} _ {\\mathrm {S}} ^ {\\prime})) \\} _ {i = 1} ^ {n},\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.354, + 0.393, + 0.784, + 0.43 + ], + "angle": 0, + "content": "\\[\n\\mathbf {v} _ {\\mathrm {S}} ^ {\\prime} = \\left[ \\frac {1}{n} \\sum_ {i = 1} ^ {n} \\mathbf {X} _ {\\text {o u t}} ^ {\\prime} [ i ] [ 1 ], \\dots , \\frac {1}{n} \\sum_ {i = 1} ^ {n} \\mathbf {X} _ {\\text {o u t}} ^ {\\prime} [ i ] [ d ] \\right], \\tag {21}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.333, + 0.435, + 0.655, + 0.471 + ], + "angle": 0, + "content": "\\[\n\\mathbf {m} _ {\\mathbf {X} _ {\\mathrm {i n}} ^ {\\prime}} = \\left[ \\frac {1}{n} \\sum_ {i = 1} ^ {n} \\mathbf {X} _ {\\mathrm {i n}} ^ {\\prime} [ i ] [ 1 ], \\dots , \\frac {1}{n} \\sum_ {i = 1} ^ {n} \\mathbf {X} _ {\\mathrm {i n}} ^ {\\prime} [ i ] [ d ] \\right].\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.475, + 0.767, + 0.491 + ], + "angle": 0, + "content": "Secondly, Multi-T can be integrated with UOD methods \\(\\mathrm{M}(\\cdot)\\), illustrated as:" + }, + { + "type": "equation", + "bbox": [ + 0.307, + 0.502, + 0.785, + 0.537 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathrm {F} _ {\\mathrm {M} + \\text {M u l t i - T}} (\\mathbf {X}) = \\mathrm {M}. \\operatorname {f i t} \\left(\\left\\{\\mathrm {S} - \\operatorname {n o r m} \\left(\\mathbf {x} _ {i}, \\mathbf {v} _ {\\mathrm {S}} ^ {\\prime}\\right) \\mid \\mathbf {x} _ {i} \\in \\mathbf {X} _ {i n} ^ {\\prime} \\right\\}\\right) \\\\ . \\operatorname {p r e d i c t} \\left(\\left\\{\\mathrm {S} - \\operatorname {n o r m} \\left(\\mathbf {x} _ {i}, \\mathbf {v} _ {\\mathrm {S}} ^ {\\prime}\\right) \\mid \\mathbf {x} _ {i} \\in \\mathbf {X} \\right\\}\\right). \\\\ \\end{array}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.554, + 0.376, + 0.571 + ], + "angle": 0, + "content": "5 Experiments" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.581, + 0.787, + 0.626 + ], + "angle": 0, + "content": "We cover our experiments for (i) comparing with other threshold learners on dataset splitting (Sec. 5.1); (ii) measuring the performance improvement of adopting the Multi-T to previously proposed outlier scoring methods (Sec. 5.2)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.627, + 0.784, + 0.702 + ], + "angle": 0, + "content": "Benchmarks and Feature Extraction. Based on the existing research [32,34], we use the raw pixel representation for grayscale datasets, including MNIST [29] and Fashion-MNIST [59]. For RGB datasets such as STL-10 [15], Internet [32], CIFAR-10 [27], MIT-Places-Small [61], we adopt two deep feature extractors: ImageNet pretrained [20] ResNet-50 [21] (ResNet)2, and CLIP [45]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.702, + 0.786, + 0.793 + ], + "angle": 0, + "content": "Target Dataset. In our experiments, each class within a benchmark dataset is alternatively regarded as inliers, with instances from all other classes considered outliers. The target dataset includes all inliers along with a randomly selected subset of outliers. Results for each dataset are averaged across all classes. Additionally, for every round, we further average the results over a wide range of outlier ratios: \\([0.05, 0.1, 0.2, 0.3, 0.4]^3\\), to assess the method's robustness/stability." + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.811, + 0.68, + 0.825 + ], + "angle": 0, + "content": "2 Unless otherwise stated, ResNet-50 is the default feature extractor." + }, + { + "type": "page_footnote", + "bbox": [ + 0.22, + 0.826, + 0.784, + 0.841 + ], + "angle": 0, + "content": "3 Unless otherwise stated, the experimental results are averaged across \\([0.05,\\dots ,0.4]\\)" + }, + { + "type": "list", + "bbox": [ + 0.218, + 0.811, + 0.784, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.355, + 0.128 + ], + "angle": 0, + "content": "Z. Liu et al." + }, + { + "type": "image", + "bbox": [ + 0.216, + 0.143, + 0.788, + 0.269 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.216, + 0.271, + 0.788, + 0.402 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.409, + 0.788, + 0.438 + ], + "angle": 0, + "content": "Fig.3: Qualitative results compared with SOTA threshold learners. The experiments are conducted on STL-10 (Inlier class: Monkey)." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.465, + 0.427, + 0.481 + ], + "angle": 0, + "content": "5.1 Threshold Learning" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.489, + 0.788, + 0.58 + ], + "angle": 0, + "content": "Competing Methods. We compare our proposed Multi-T with those effective thresholding (TL) methods involving Kernel-based: AUCP [48], FGD [43]; Curve-based: EB [17]; Normality-based: DSN [5], CHAU [10]; Filtering-based: FILTER [19], HIST [55]; Statistical-based: MTT [49], BOOT [37], QMCD [24], CLF [7], IQR [8], KARCH [1], MCST [16], GESD [4], REGR [3], MAD [6], CLUST [11], CPD [12] and Transformation-based: MOLL [25], YJ [46]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.58, + 0.787, + 0.61 + ], + "angle": 0, + "content": "Evaluation Metric. The performance of general thresholding is measured with \\(\\mathrm{F}_{\\beta}\\)-score, defined as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.354, + 0.619, + 0.786, + 0.651 + ], + "angle": 0, + "content": "\\[\n\\mathrm {F} _ {\\beta} = \\left(1 + \\beta^ {2}\\right) \\cdot \\frac {\\text {p r e c i s i o n} \\cdot \\text {r e c a l l}}{\\left(\\beta^ {2} \\cdot \\text {p r e c i s i o n}\\right) + \\text {r e c a l l}}, \\tag {23}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.659, + 0.788, + 0.735 + ], + "angle": 0, + "content": "where \\(\\beta < 1\\) tilts towards precision while \\(\\beta > 1\\) prioritizes recall. In this work, we utilize \\(\\mathrm{F}_{0.1}\\)-score \\((\\mathrm{F}_{0.1})\\) and \\(\\mathrm{F}_{10}\\)-score \\((\\mathrm{F}_{10})\\) that measure the accuracy of predicted outliers and inliers. Besides, \\(\\mathrm{F}_{1}\\)-score \\((\\beta = 1)\\), a harmonic mean of the precision and recall, is not practical for this task since the compete separability between inliers' and outliers' score distributions cannot be guaranteed." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.789, + 0.841 + ], + "angle": 0, + "content": "Main Results. Across a series of competing threshold learners, while one method may achieve a commendable \\( \\mathrm{F}_{0.1} \\)-score of 0.976, the corresponding \\( \\mathrm{F}_{10} \\)-score is mere 0.452, and vice versa (\\( \\mathrm{F}_{10} \\)-score: 0.979; \\( \\mathrm{F}_{0.1} \\)-score: 0.447). By contrast, our solution maintains its efficacy across various evaluation criteria (\\( \\mathrm{F}_{0.1} \\)-score: 0.917, \\( \\mathrm{F}_{10} \\)-score: 0.980). It exhibits remarkable stability across a diverse spectrum of outlier ratios, as illustrated in Fig. 3. To further investigate the benefits of our technique, Tab. 1 categorizes two groups, each representing a different" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.265, + 0.115, + 0.732, + 0.13 + ], + "angle": 0, + "content": "Rethinking Unsupervised Outlier Detection via Multiple Thresholding" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.145, + 0.788, + 0.201 + ], + "angle": 0, + "content": "Table 1: Average \\( \\mathrm{F}_{\\beta} \\)-score of thresholding methods on STL-10. \"+\"GT Norm\" refers to the ideal initial outlier score function. \"Highest \\( \\mathrm{F}_{\\beta} \\)\" (\\( \\beta: 0.1, 10 \\)) refers to the method that garners the highest \\( \\mathrm{F}_{\\beta} \\)-score across all baseline models. The better average result is highlighted in bold." + }, + { + "type": "table", + "bbox": [ + 0.242, + 0.213, + 0.753, + 0.364 + ], + "angle": 0, + "content": "
Outlier Score Fun.Method (TL)ResNet-50CLIP
F0.1F10Avg.F0.1F10Avg.
LVAD-S [34]Highest F0.10.9110.4540.6820.9360.4010.669
Highest F100.3820.9670.6740.3570.9540.655
Multi-T(Ours)0.8400.8690.8550.8600.8660.863
+GT Norm [32]Highest F0.10.9760.4520.7140.9770.4480.713
Highest F100.4470.9790.7130.6910.9860.839
Multi-T(Ours)0.9170.9800.9490.8570.9840.920
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.405, + 0.784, + 0.434 + ], + "angle": 0, + "content": "outlier score function. Such categorization is pivotal as threshold methods ought to accommodate various foundational bases." + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.472, + 0.788, + 0.502 + ], + "angle": 0, + "content": "Table 2: Average AUC results compared with SOTA outlier scoring methods (outlier detectors). Blue and Orange indicates the best and second-best results, respectively." + }, + { + "type": "table", + "bbox": [ + 0.221, + 0.513, + 0.779, + 0.803 + ], + "angle": 0, + "content": "
Method (OS)STL-10CIFAR-10CIFAR-100MIT-PlacesMNIST
ResNetCLIPResNetCLIPResNetCLIPResNetCLIPPixel
IF [35]0.8360.9430.7800.8910.7900.8660.6870.8680.776
LOF [11]0.6280.6260.6730.6210.8390.8490.5560.5200.754
RSRAE [28]0.9620.9380.8620.8790.9140.8850.8740.8760.793
ECOD [31]0.9070.9810.8730.9350.8730.9180.7770.9430.734
LUNAR [18]0.7760.8210.7670.7740.8380.8990.6430.8710.797
Shell-Re. [32]0.8620.8380.8600.8130.8350.8130.8260.9140.776
LVAD [34]0.9540.9680.8600.9170.9210.9170.8440.9190.867
Multi-T0.9680.9890.8950.9570.9380.9560.8670.9740.897
DeepSVDD [50]0.6220.5970.5600.5090.5630.5810.5830.5490.513
+Multi-T0.9250.9210.7690.8190.8350.8260.7550.8320.732
Improve.48.7%54.3%37.3%60.9%48.3%42.2%29.5%51.6%37.9%
OCSVM [51]0.9270.9210.8260.8500.8790.8500.8260.8710.831
+Multi-T0.9570.9650.8590.9160.9160.8990.8460.9240.863
Improve.3.24%4.78%4.00%7.76%4.21%5.76%2.42%6.08%3.85%
" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.353, + 0.127 + ], + "angle": 0, + "content": "Z. Liu et al." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.146, + 0.39, + 0.163 + ], + "angle": 0, + "content": "5.2 Outlier Scoring" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.174, + 0.783, + 0.263 + ], + "angle": 0, + "content": "Competing Methods. We compare two themes of baseline outlier scoring (OS) methods: (i) statistical-based methods: IF [35], OCSVM [51], DeepSVDD [51], ECOD [31], LUNAR [28], RSRAE [28], Shell-Re. [32], LVAD [34]. (ii) deep-learning-based models: GOAD [9], ICL [53], REPEN [39], NeuTraL [44], SLAD [60]. For a fair comparison, we apply Ergodic-set normalization [34] if it improves the performance of the baseline algorithms, such as IF [35], OCSVM [51]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.266, + 0.784, + 0.476 + ], + "angle": 0, + "content": "Evaluation Metric. The performance of outlier scoring (ranking accuracy) is primarily assessed using the Area Under the Receiver Operating Characteristic curve (AUC). This metric provides a thorough assessment of ranking accuracy. Main Results. Tab. 2 shows that Multi-T can be seamlessly integrated with two classic outlier scoring methods: DeepSVDD [51], and OCSVM [51] while exhibiting significant improvements in both efficacy and stability across a diverse range of outlier ratios and various benchmarks. In most of our experiments, Multi-T itself achieves SOTA results. Notably, even in the case of non-aligned and low-resolution datasets like CIFAR-10 [27] and CIFAR-100 [27], which are known to pose challenges [41]. It surpasses the current SOTA AUC scores by margins of \\(7.76\\%\\) and \\(5.76\\%\\) for CIFAR-10 and CIFAR-100, respectively. Additionally, we observe a logical and significant enhancement in performance with improved feature representation, e.g., for the MIT-Places [61] dataset, the AUC improves from 0.867 using ResNet-50 [21] to 0.974 with CLIP [45]." + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.504, + 0.784, + 0.534 + ], + "angle": 0, + "content": "Table 3: Average AUC results, for more outlier scoring methods with \\((\\mathrm{w} / )\\) or without(w/o) our Multi-T module, conducted on STL-10." + }, + { + "type": "table", + "bbox": [ + 0.241, + 0.546, + 0.758, + 0.735 + ], + "angle": 0, + "content": "
FeatureIF [35]ECOD [11]ABOD [26]PCA [54]GMM [2]
w/ow/w/ow/w/ow/w/ow/w/ow/
ResNet0.8360.8990.9070.9190.6650.8830.8650.9450.8590.952
CLIP0.9430.9830.9810.9840.7150.9090.9840.9940.8920.962
Avg.0.8900.9410.9440.9510.6900.8960.9250.9700.8760.957
FeatureGOAD [9]ICL [53]REPEN [39]NeuTraL [44]SLAD [60]
w/ow/w/ow/w/ow/w/ow/w/ow/
ResNet0.9520.9620.9340.9570.8770.8890.8540.9500.9410.962
CLIP0.9610.9890.9510.9820.8790.9300.8510.9710.9450.985
Avg.0.9570.9750.9430.9700.8780.9090.8530.9610.9430.973
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.785, + 0.84 + ], + "angle": 0, + "content": "In Tab. 3, we present the results of Multi-T integrated with more methods involving both statistical and deep models, which shows its broad applications. Moreover, our solution excels not just in detection accuracy but also in running speed since Multi-T compresses the size of fitting data, as illustrated in Tab. 4, 5. In Fig. 4, we compare with the most related work Shell-Re. [32], which" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.265, + 0.115, + 0.732, + 0.13 + ], + "angle": 0, + "content": "Rethinking Unsupervised Outlier Detection via Multiple Thresholding" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.208 + ], + "angle": 0, + "content": "has a built-in Robust-Least-Square (RLS) thresholding procedure with MAD [6]. Apparently, our method is capable of estimating multiple thresholds that facilitate a clear demarcation of diverse scenarios (outlier ratio, dataset domain, and feature representation)." + }, + { + "type": "table_caption", + "bbox": [ + 0.239, + 0.22, + 0.79, + 0.262 + ], + "angle": 0, + "content": "Table 4: Efficiency comparison for outlier Table 5: Efficiency comparison for scoring. Timing is measured with 10,000 thresholding. All thresholding methods samples, GPU: NVIDIA RTX 3080. are conducted on CPU." + }, + { + "type": "table", + "bbox": [ + 0.277, + 0.273, + 0.489, + 0.414 + ], + "angle": 0, + "content": "
Method (OS)DeviceTime (↓)
LVAD [34]CPU645.2s
RSRAE [28]GPU121.6s
Multi-TCPU1.2s
OCSVM [51]CPU154.8s
+Multi-T105.4s
GOAD [9]GPU268.1s
+Multi-T211.7s
" + }, + { + "type": "table", + "bbox": [ + 0.576, + 0.273, + 0.739, + 0.413 + ], + "angle": 0, + "content": "
Method (TL)Time (↓)
CPD [12]3.61s
FWFM [19]1.26s
DSN [5]3.79s
CLUST [11]7.36s
AUCP [48]1.26s
Multi-T0.97s
" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.463, + 0.785, + 0.673 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.682, + 0.785, + 0.697 + ], + "angle": 0, + "content": "Fig. 4: AUC results of Multi-T compared with Shell-Re. [32] over various outlier ratios." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.749, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Discussions. We conclude the feasibility of adopting Multi-T to outlier scoring methods into three aspects: (i) both normalization with outliers and manifold learning with inliers are necessary for most of outlier scoring methods; (ii) Despite the initial outlier score distribution (Eq. 6) is usually not perfectly separable, it provides a reliable baseline. (iii) Multi-T is able to isolate the overlap region and predict both uncontaminated inliers." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.353, + 0.127 + ], + "angle": 0, + "content": "Z. Liu et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.24, + 0.145, + 0.528, + 0.173 + ], + "angle": 0, + "content": "Table 6: Ablation study for DeepSVDD and LVAD-S with Multi-T (\\(\\gamma\\): 0.2)." + }, + { + "type": "table", + "bbox": [ + 0.264, + 0.184, + 0.501, + 0.268 + ], + "angle": 0, + "content": "
DeepSVDDAUCLVAD-SAUC
+Multi-T0.950+Multi-T0.971
w/o In0.643w/o In0.965
w/o Out0.939w/o Out0.951
" + }, + { + "type": "table_caption", + "bbox": [ + 0.537, + 0.145, + 0.779, + 0.159 + ], + "angle": 0, + "content": "Table 7: Comparison with \\( k \\)-sigma." + }, + { + "type": "table", + "bbox": [ + 0.577, + 0.171, + 0.737, + 0.268 + ], + "angle": 0, + "content": "
MetricMethodScore
F0.13-sigma0.406
Multi-T0.840
F101-sigma0.806
Multi-T0.869
" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.288, + 0.389, + 0.304 + ], + "angle": 0, + "content": "5.3 Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.313, + 0.788, + 0.556 + ], + "angle": 0, + "content": "The experiments thus far have established the effectiveness of Multi-T. However, there remains a concern regarding how sensitive the performance improvement is to Multi-T's two primary components: adaptive outliers (Out) and uncontaminated inliers (In), which refer to the normalization and manifold learning procedures, respectively. We select two representative outlier scoring models: DeepSVDD [50] and LVAD-S [34], whereas DeepSVDD follows a widely-used outlier detection mechanism, i.e., learning the normality (hyper-sphere), and identifying inliers is of great significance. Without predicted inliers, the AUC result decreases from 0.950 to 0.643 on STL-10, shown in Tab. 6. Additionally, since the normalization procedure can be considered as a distance de-noising procedure [33], Multi-T will contribute to those methods with distance computation, e.g., the result of LVAD-S is improved from 0.951 to 0.971 on STL-10 and 0.838 to 0.882 on CIFAR-10, which verify our prior assumption that the identification of both inliers and outliers is of great significance, which enhances the reliability of multiple thresholding perspective. Moreover, Tab. 7 indicates that our thresholding process markedly surpasses the classical \\(k\\)-sigma rule [42]." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.578, + 0.35, + 0.592 + ], + "angle": 0, + "content": "5.4 Limitation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.603, + 0.788, + 0.709 + ], + "angle": 0, + "content": "Our method is tied to the ranking accuracy (separability) of the initial outlier score function. Based on most related works, we consider UOD as an one-class learning task (only one inlier/normal class in the target dataset). However we find in some specific cases, there might exist multi-normal classes, e.g., the digit 5 class of the MNIST dataset. In that case, the efficacy of our method might be decreased. Additionally, this statistical-based method has limitations in very small-scale datasets since the \"three-sigma\" rule will be less effective." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.733, + 0.36, + 0.749 + ], + "angle": 0, + "content": "6 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.788, + 0.841 + ], + "angle": 0, + "content": "This work introduces a novel perspective for UOD about advancing existing outlier detectors (scoring methods) via thresholding. To this end, we propose a multiple thresholding (Multi-T) module, to label the unlabelled target dataset. Comprehensive experiments verify that the Multi-T can significantly improve both the efficacy and efficiency of previously proposed outlier scoring methods." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.265, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Rethinking Unsupervised Outlier Detection via Multiple Thresholding" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.404, + 0.163 + ], + "angle": 0, + "content": "Acknowledgement" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.179, + 0.784, + 0.209 + ], + "angle": 0, + "content": "We would like to thank the reviewers for their valuable comments and suggestions." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.234, + 0.323, + 0.25 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.267, + 0.785, + 0.295 + ], + "angle": 0, + "content": "1. Afsari, B.: Riemannian \\(L^p\\) center of mass: existence, uniqueness, and convexity. Proceedings of the American Mathematical Society pp. 655-673 (2011)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.296, + 0.607, + 0.309 + ], + "angle": 0, + "content": "2. Aggarwal, C.C.: Outlier analysis second edition (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.31, + 0.785, + 0.337 + ], + "angle": 0, + "content": "3. Aggarwal, C.C., Aggarwal, C.C.: An introduction to outlier analysis. Springer (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.338, + 0.785, + 0.365 + ], + "angle": 0, + "content": "4. Alrawashdeh, M.J.: An adjusted grubbs' and generalized extreme studentized deviation. Demonstratio Mathematica pp. 548-557 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.367, + 0.785, + 0.407 + ], + "angle": 0, + "content": "5. Amagata, D., Onizuka, M., Hara, T.: Fast and exact outlier detection in metric spaces: a proximity graph-based approach. In: Proceedings of the 2021 International Conference on Management of Data. pp. 36-48 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.408, + 0.785, + 0.449 + ], + "angle": 0, + "content": "6. Archana, N., Pawar, S.: Periodicity detection of outlier sequences using constraint based pattern tree with mad. International Journal of Advanced Studies in Computers, Science and Engineering p. 34 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.449, + 0.785, + 0.491 + ], + "angle": 0, + "content": "7. Barbado, A., Corcho, O., Benjamins, R.: Rule extraction in unsupervised anomaly detection for model explainability: Application to oneclasssvm. Expert Systems with Applications p. 116100 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.491, + 0.785, + 0.519 + ], + "angle": 0, + "content": "8. Bardet, J.M., Dimby, S.F.: A new non-parametric detector of univariate outliers for distributions with unbounded support. Extremes pp. 751-775 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.52, + 0.785, + 0.547 + ], + "angle": 0, + "content": "9. Bergman, L., Hoshen, Y.: Classification-based anomaly detection for general data. arXiv preprint arXiv:2005.02359 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.548, + 0.785, + 0.575 + ], + "angle": 0, + "content": "10. Bol'shev, L., Ubaidullaeva, M.: Chauvenet's test in the classical theory of errors. Theory of Probability & Its Applications pp. 683-692 (1975)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.576, + 0.785, + 0.617 + ], + "angle": 0, + "content": "1. Breunig, M.M., Kriegel, H.P., Ng, R.T., Sander, J.: Lof: identifying density-based local outliers. In: Proceedings of the ACM SIGMOD International Conference on Management of Data. pp. 93-104 (2000)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.618, + 0.785, + 0.645 + ], + "angle": 0, + "content": "2. Van den Burg, G.J., Williams, C.K.: An evaluation of change point detection algorithms. arXiv preprint arXiv:2003.06222 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.646, + 0.785, + 0.699 + ], + "angle": 0, + "content": "3. Camposeco, F., Sattler, T., Cohen, A., Geiger, A., Pollefeys, M.: Toroidal constraints for two-point localization under high outlier ratios. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 4545-4553 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.701, + 0.785, + 0.742 + ], + "angle": 0, + "content": "4. Chalapathy, R., Menon, A.K., Chawla, S.: Robust, deep and inductive anomaly detection. In: Proceedings of Joint European Conference on Machine Learning and Knowledge Discovery in Databases. pp. 36-51 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.743, + 0.785, + 0.784 + ], + "angle": 0, + "content": "5. Coates, A., Ng, A., Lee, H.: An analysis of single-layer networks in unsupervised feature learning. In: Proceedings of the International Conference on Artificial Intelligence and Statistics. pp. 215-223 (2011)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.785, + 0.785, + 0.812 + ], + "angle": 0, + "content": "6. Coin, D.: Testing normality in the presence of outliers. Statistical Methods and Applications pp. 3-12 (2008)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.813, + 0.785, + 0.84 + ], + "angle": 0, + "content": "7. Friendly, M., Monette, G., Fox, J.: Elliptical insights: understanding statistical methods through elliptical geometry. Statistical Science pp. 1-39 (2013)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.267, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.353, + 0.127 + ], + "angle": 0, + "content": "Z. Liu et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.785, + 0.189 + ], + "angle": 0, + "content": "18. Goodge, A., Hooi, B., Ng, S.K., Ng, W.S.: Lunar: Unifying local outlier detection methods via graph neural networks. In: Proceedings of the AAAI Conference on Artificial Intelligence. pp. 6737-6745 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.189, + 0.785, + 0.23 + ], + "angle": 0, + "content": "19. Hashemi, N., German, E.V., Ramirez, J.P., Ruths, J.: Filtering approaches for dealing with noise in anomaly detection. In: 2019 IEEE 58th Conference on Decision and Control. pp. 5356-5361 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.23, + 0.785, + 0.271 + ], + "angle": 0, + "content": "20. He, K., Girshick, R., Dollar, P.: Rethinking imagenet pre-training. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 4918-4927 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.271, + 0.785, + 0.312 + ], + "angle": 0, + "content": "21. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 770-778 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.312, + 0.785, + 0.338 + ], + "angle": 0, + "content": "22. Hendrycks, D., Mazeika, M., Dietterich, T.: Deep anomaly detection with outlier exposure. arXiv preprint arXiv:1812.04606 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.338, + 0.785, + 0.379 + ], + "angle": 0, + "content": "23. Huyan, N., Quan, D., Zhang, X., Liang, X., Chanussot, J., Jiao, L.: Unsupervised outlier detection using memory and contrastive learning. IEEE Transactions on Image Processing pp. 6440-6454 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.379, + 0.785, + 0.42 + ], + "angle": 0, + "content": "24. Iouchtchenko, D., Raymond, N., Roy, P.N., Nooijen, M.: Deterministic and quasi-random sampling of optimized gaussian mixture distributions for vibronic monte carlo. arXiv preprint arXiv:1912.11594 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.421, + 0.785, + 0.461 + ], + "angle": 0, + "content": "25. Keyzer, M.A., Sonneveld, B.: Using the mollifier method to characterize datasets and models: the case of the universal soil loss equation. ITC Journal pp. 263-272 (1997)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.461, + 0.785, + 0.502 + ], + "angle": 0, + "content": "26. Kriegel, H.P., Schubert, M., Zimek, A.: Angle-based outlier detection in high-dimensional data. In: Proceedings of the 14th ACM SIGKDD international conference on Knowledge discovery and data mining. pp. 444-452 (2008)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.502, + 0.785, + 0.528 + ], + "angle": 0, + "content": "27. Krizhevsky, A., Hinton, G., et al.: Learning multiple layers of features from tiny images (2009)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.528, + 0.785, + 0.556 + ], + "angle": 0, + "content": "28. Lai, C.H., Zou, D., Lerman, G.: Robust subspace recovery layer for unsupervised anomaly detection. arXiv preprint arXiv:1904.00152 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.556, + 0.681, + 0.569 + ], + "angle": 0, + "content": "29. LeCun, Y., Cortes, C.: MNIST handwritten digit database (2010)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.569, + 0.785, + 0.609 + ], + "angle": 0, + "content": "30. Li, T., Wang, Z., Liu, S., Lin, W.Y.: Deep unsupervised anomaly detection. In: Proceedings of IEEE Winter Conference on Applications of Computer Vision. pp. 3636-3645 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.609, + 0.785, + 0.65 + ], + "angle": 0, + "content": "31. Li, Z., Zhao, Y., Hu, X., Botta, N., Ionescu, C., Chen, G.H.: Ecod: Unsupervised outlier detection using empirical cumulative distribution functions. arXiv preprint arXiv:2201.00382 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.65, + 0.785, + 0.691 + ], + "angle": 0, + "content": "32. Lin, D., Liu, S., Li, H., Cheung, N.M., Ren, C., Matsushita, Y.: Shell theory: A statistical model of reality. IEEE Transactions on Pattern Analysis and Machine Intelligence (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.691, + 0.785, + 0.732 + ], + "angle": 0, + "content": "33. Lin, W.Y., Liu, S., Dai, B.T., Li, H.: Distance based image classification: A solution to generative classification's conundrum? International Journal of Computer Vision pp. 177-198 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.732, + 0.785, + 0.772 + ], + "angle": 0, + "content": "34. Lin, W.Y., Liu, Z., Liu, S.: Locally varying distance transform for unsupervised visual anomaly detection. In: European Conference on Computer Vision. pp. 354-371. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.772, + 0.785, + 0.8 + ], + "angle": 0, + "content": "35. Liu, F.T., Ting, K.M., Zhou, Z.H.: Isolation forest. In: Proceedings of International Conference on Data Mining. pp. 413-422 (2008)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.8, + 0.785, + 0.84 + ], + "angle": 0, + "content": "36. Liznerski, P., Ruff, L., Vandermeulen, R.A., Franks, B.J., Müller, K.R., Kloft, M.: Exposing outlier exposure: What can be learned from few, one, and zero outlier images. arXiv preprint arXiv:2205.11474 (2022)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.265, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Rethinking Unsupervised Outlier Detection via Multiple Thresholding" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.785, + 0.175 + ], + "angle": 0, + "content": "37. Martin, M.A., Roberts, S.: An evaluation of bootstrap methods for outlier detection in least squares regression. Journal of Applied Statistics pp. 703-720 (2006)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.177, + 0.787, + 0.218 + ], + "angle": 0, + "content": "38. Masci, J., Meier, U., Ciresan, D., Schmidhuber, J.: Stacked convolutional autoencoders for hierarchical feature extraction. In: Proceedings of International Conference on Artificial Neural Networks. pp. 52-59 (2011)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.219, + 0.787, + 0.274 + ], + "angle": 0, + "content": "39. Pang, G., Cao, L., Chen, L., Liu, H.: Learning representations of ultrahigh-dimensional data for random distance-based outlier detection. In: Proceedings of the 24th ACM SIGKDD international conference on knowledge discovery & data mining. pp. 2041-2050 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.276, + 0.786, + 0.303 + ], + "angle": 0, + "content": "40. Parzen, E.: On estimation of a probability density function and mode. The annals of mathematical statistics pp. 1065-1076 (1962)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.304, + 0.786, + 0.346 + ], + "angle": 0, + "content": "41. Perera, P., Nallapati, R., Xiang, B.: Ocgan: One-class novelty detection using gans with constrained latent representations. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 2898-2906 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.346, + 0.786, + 0.361 + ], + "angle": 0, + "content": "42. Pukelsheim, F.: The three sigma rule. The American Statistician pp. 88-91 (1994)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.361, + 0.786, + 0.401 + ], + "angle": 0, + "content": "43. Qi, Z., Jiang, D., Chen, X.: Iterative gradient descent for outlier detection. International Journal of Wavelets, Multiresolution and Information Processing p. 2150004 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.403, + 0.786, + 0.445 + ], + "angle": 0, + "content": "44. Qiu, C., Pfrommer, T., Kloft, M., Mandt, S., Rudolph, M.: Neural transformation learning for deep anomaly detection beyond images. In: International Conference on Machine Learning. pp. 8703-8714 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.445, + 0.786, + 0.5 + ], + "angle": 0, + "content": "45. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: Proceedings of International Conference on Machine Learning. pp. 8748-8763 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.501, + 0.786, + 0.529 + ], + "angle": 0, + "content": "46. Raymaekers, J., Rousseeuw, P.J.: Transforming variables to central normality. Machine Learning pp. 1-23 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.53, + 0.786, + 0.558 + ], + "angle": 0, + "content": "47. Reiss, T., Hoshen, Y.: Mean-shifted contrastive loss for anomaly detection. In: Proceedings of the AAAI Conference on Artificial Intelligence. pp. 2155-2162 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.559, + 0.786, + 0.613 + ], + "angle": 0, + "content": "48. Ren, K., Yang, H., Zhao, Y., Chen, W., Xue, M., Miao, H., Huang, S., Liu, J.: A robust auc maximization framework with simultaneous outlier detection and feature selection for positive-unlabeled classification. IEEE Transactions on Neural Networks and Learning Systems pp. 3072-3083 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.615, + 0.786, + 0.657 + ], + "angle": 0, + "content": "49. Rengasamy, D., Rothwell, B.C., Figueredo, G.P.: Towards a more reliable interpretation of machine learning outputs for safety-critical systems using feature importance fusion. Applied Sciences p. 11854 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.657, + 0.786, + 0.699 + ], + "angle": 0, + "content": "50. Ruff, L., Vandermeulen, R., Goernitz, N., Deecke, L., Siddiqui, S.A., Binder, A., Müller, E., Kloft, M.: Deep one-class classification. In: Proceedings of International Conference on Machine Learning. pp. 4393-4402 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.7, + 0.786, + 0.741 + ], + "angle": 0, + "content": "51. Scholkopf, B., Platt, J.C., Shawe-Taylor, J., Smola, A.J., Williamson, R.C.: Estimating the support of a high-dimensional distribution. Neural Computation 13(7), 1443-1471 (2001)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.741, + 0.786, + 0.77 + ], + "angle": 0, + "content": "52. Scholkopf, B., Smola, A.J., Bach, F., et al.: Learning with kernels: support vector machines, regularization, optimization, and beyond (2002)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.77, + 0.786, + 0.798 + ], + "angle": 0, + "content": "53. Shenkar, T., Wolf, L.: Anomaly detection for tabular data with internal contrastive learning. In: International Conference on Learning Representations (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.799, + 0.786, + 0.841 + ], + "angle": 0, + "content": "54. Shyu, M.L., Chen, S.C., Sarinnapakorn, K., Chang, L.: A novel anomaly detection scheme based on principal component classifier. In: Proceedings of the IEEE foundations and new directions of data mining workshop. pp. 172-179 (2003)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.787, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.353, + 0.127 + ], + "angle": 0, + "content": "Z. Liu et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.787, + 0.189 + ], + "angle": 0, + "content": "55. Thanammal, K., Jayasudha, J., Vijayalakshmi, R., Arumugaperumal, S.: Effective histogram thresholding techniques for natural images using segmentation. Journal of Image and Graphics pp. 113-116 (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.19, + 0.786, + 0.231 + ], + "angle": 0, + "content": "56. Wang, S., Zeng, Y., Liu, X., Zhu, E., Yin, J., Xu, C., Kloft, M.: Effective end-to-end unsupervised outlier detection via inlier priority of discriminative network. Proceedings of Advances in Neural Information Processing Systems (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.232, + 0.787, + 0.286 + ], + "angle": 0, + "content": "57. Wang, S., Zhu, E., Hu, X., Liu, X., Liu, Q., Yin, J., Wang, F.: Robustness can be cheap: A highly efficient approach to discover outliers under high outlier ratios. In: Proceedings of the AAAI Conference on Artificial Intelligence. pp. 5313-5320 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.287, + 0.787, + 0.328 + ], + "angle": 0, + "content": "58. Xia, Y., Cao, X., Wen, F., Hua, G., Sun, J.: Learning discriminative reconstructions for unsupervised outlier removal. In: Proceedings of the IEEE International Conference on Computer Vision. pp. 1511-1519 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.328, + 0.787, + 0.356 + ], + "angle": 0, + "content": "59. Xiao, H., Rasul, K., Vollgraf, R.: Fashion-mnist: a novel image dataset for benchmarking machine learning algorithms (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.356, + 0.787, + 0.397 + ], + "angle": 0, + "content": "60. Xu, H., Wang, Y., Wei, J., Jian, S., Li, Y., Liu, N.: Fascinating supervisory signals and where to find them: Deep anomaly detection with scale learning. arXiv preprint arXiv:2305.16114 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.398, + 0.787, + 0.439 + ], + "angle": 0, + "content": "61. Zhou, B., Lapedriza, A., Khosla, A., Oliva, A., Torralba, A.: Places: A 10 million image database for scene recognition. IEEE Transactions on Pattern Analysis and Machine Intelligence pp. 1452-1464 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.439, + 0.787, + 0.481 + ], + "angle": 0, + "content": "62. Zhou, C., Paffenroth, R.C.: Anomaly detection with robust deep autoencoders. In: Proceedings of the ACM SIGKDD International Conference on Knowledge Discovery and Data Mining. pp. 665-674 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.481, + 0.787, + 0.523 + ], + "angle": 0, + "content": "63. Zong, B., Song, Q., Min, M.R., Cheng, W., Lumezanu, C., Cho, D., Chen, H.: Deep autoencoding gaussian mixture model for unsupervised anomaly detection. In: Proceedings of International Conference on Learning Representations (2018)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.523 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/Rethinking Unsupervised Outlier Detection via Multiple Thresholding/7bb2035e-e079-470b-b56b-69d7d8631ccd_origin.pdf b/2024/Rethinking Unsupervised Outlier Detection via Multiple Thresholding/7bb2035e-e079-470b-b56b-69d7d8631ccd_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..37e60ba9f884ea2c9406b0b2ab10e3e36c5495f9 --- /dev/null +++ b/2024/Rethinking Unsupervised Outlier Detection via Multiple Thresholding/7bb2035e-e079-470b-b56b-69d7d8631ccd_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9e157c0c67af79d806d0fa444c7cb8d2789a5b226987c9aa0d68ed95edac2d2b +size 1336909 diff --git a/2024/Rethinking Unsupervised Outlier Detection via Multiple Thresholding/full.md b/2024/Rethinking Unsupervised Outlier Detection via Multiple Thresholding/full.md new file mode 100644 index 0000000000000000000000000000000000000000..e5101b317f5fcd539f23fe81cd59717df3b777eb --- /dev/null +++ b/2024/Rethinking Unsupervised Outlier Detection via Multiple Thresholding/full.md @@ -0,0 +1,420 @@ +# Rethinking Unsupervised Outlier Detection via Multiple Thresholding + +Zhonghang Liu1, Panzhong Lu2, Guoyang Xie3,4, Zhichao Lu4, and Wen-Yan Lin1 + +$^{1}$ Singapore Management University $^{2}$ Westlake University + +3 Contemporary Amperex Technology Co., Limited + +$^{4}$ City University of Hong Kong + +zhliu.2020@phdcs.smu.edu.sg + +lupanzhong@westlake.edu.cn + +guoyang.xie@ieee.org + +luzhichao.cn@gmail.com + +daniellin@smu.edu.sg + +Abstract. In the realm of unsupervised image outlier detection, assigning outlier scores holds greater significance than its subsequent task: thresholding for predicting labels. This is because determining the optimal threshold on non-separable outlier score functions is an ill-posed problem. However, the lack of predicted labels not only hinders some real applications of current outlier detectors but also causes these methods not to be enhanced by leveraging the dataset's self-supervision. To advance existing scoring methods, we propose a multiple thresholding (Multi-T) module. It generates two thresholds that isolate inliers and outliers from the unlabelled target dataset, whereas outliers are employed to obtain better feature representation while inliers provide an uncontaminated manifold. Extensive experiments verify that Multi-T can significantly improve proposed outlier scoring methods. Moreover, Multi-T contributes to a naive distance-based method being state-of-the-art. Code is available at: https://github.com/zhliu-uod/Multi-T. + +Keywords: Unsupervised outlier detection, multiple thresholding, outlier scoring + +# 1 Introduction + +Which is more important for the unsupervised outlier detection (UOD), outlier score or label? Currently, the mainstream UOD approaches [23, 28, 31, 56] focus on the first, i.e., learning a discriminative outlier score function. However, those recent outlier detectors are usually complex and non-iterative since the absence of predicted labels limits these methods to be further advanced with the dataset's self-supervision. Such that, this work tends to label the unlabelled tar + +![](images/9080d1e2120f51744f2d3474d25e21c9a67b27c903f68baf00619bd1eb1b6ef1.jpg) +Fig. 1: Unlike the default UOD paradigm (a) concerned about learning a discriminative score function, our perspective (b) is to explore the dataset's prior knowledge via thresholding, to advance the previously proposed method. (c) illustrates the significant improvement of DeepSVDD [50] with Multi-T module. + +get dataset, thereby leveraging the dataset's prior knowledge to enhance simple outlier scoring methods $^{1}$ , achieving state-of-the-art (SOTA) results efficiently. + +To understand the importance of the target dataset's self-supervision, we first decompose the predominant image outlier scoring approaches [9,32,50,51] into two continuous stages: inlier/normal manifold learning and distance/similarity inference. There is no doubt that identifying inliers will contribute to learning an uncontaminated inlier manifold (e.g. the hypersphere of DeepSVDD [50]). Additionally, motivated by Shell Theory [32,33] in the high-dimensional space, image feature representation for distance computation can be significantly improved via outliers-based Shell Normalization [32]. Such that it is critical to obtain the thresholds that separate both inliers and outliers from target datasets. + +To align with our above objective, we present a Multi-T module, which follows a multiple thresholding mechanism. Compared with conventional single thresholding methods [19, 24, 48, 49], Multi-T generates two distinct thresholds that separately isolate outliers and inliers from the target dataset, which is capable of dealing with the inevitable inseparability of the initial given score distribution. This allows us to directly implement Multi-T on a simple outlier score function. + +By understanding the impact of the outlier ratio, Multi-T involves two stages: Uncontaminated inliers. We employ an iterative two-step process (i) identifying a noisy inlier distribution through the analysis of sorted initial score function, i.e., Ergodic-set normalized [34] distance to the mean of target dataset; (ii) subsequently filtering outliers with 3-sigma rule [42]. When converged, we can identify uncontaminated inliers, thereby training an inlier/normal manifold. + +Adaptive outliers. The uncontaminated inliers identification process also outputs a series of outlier threshold candidates. Subsequently, we compare the ranking-index similarity between Ergodic-set Normalization [34] and Shell Normalization [32] on the initial outlier score function. The structural consistency and contrastive properties of the two normalization procedures motivate us to implicitly estimate a rough outlier ratio, thereby detecting the adaptive outliers. + +The Multi-T module is training-free, highly efficient and grants the adaptiveness to choose thresholds suited to downstream implementations. As evaluated + +with comprehensive experimental settings, Multi-T helps to boost significant performance (efficacy and efficiency) improvement. For example, the AUC score of DeepSVDD [50] is improved from 0.622 to 0.925 on STL-10 dataset [15]. Moreover, the naive distance-to-the-target dataset's mean [32] metric integrated with Multi-T achieves SOTA results with only 1.2 second consumption for 10,000 ResNet-50 [21] samples that are orders of magnitude faster than baselines. Our primary contributions are summarized below: + +- We introduce a novel perspective for UOD about enhancing previously proposed scoring methods via thresholding on a simple initial outlier score function. +- We present a multiple thresholding (Multi-T) module that generates two different thresholds to separate both inliers and outliers from the target dataset. +- The efficacy of integrated scoring methods can be significantly improved without the external complexity increase, as evaluated in extensive experiments. + +# 2 Related Work + +# 2.1 Unsupervised Outlier Detection + +Classic UOD task aims to assign an outlier score/likelihood to an image sample. The recent models can be divided into inlier-manifold learning and outlier exposure. Inlier-manifold learning assumes that inliers are the majority. Deep learning-based outlier detectors, as illustrated in studies [14, 28, 30, 38, 56, 58, 62, 63], typically focus on reducing the dimensionality of image data by projecting it into a latent and discriminative space. Additionally, various statistics-based methods model the inlier manifold using discrimination-based [32, 34] or density-based approaches [11, 26, 31, 40]. In applications where outliers are not the minority [13, 57], different approaches are required. Some outlier exposure methods [22, 36] employ out-of-distribution (OOD) data to train networks for detecting unseen OOD samples. For example, Shell-Renormalization [32] employs predicted outlier candidates to iteratively refine feature representation. It is particularly useful in scenarios where classic manifold learners, assuming outliers as a minority, may not be effective. However, these two paradigms are separable, this work will attempt to design a thresholding framework that generates both inliers and outliers, leveraging the benefits of both. Inlier-manifold learning and outlier exposure. + +# 2.2 Thresholding + +Some traditional outlier detection methods [32, 35, 52], provide both the outlier score and its corresponding threshold concurrently. In this context, the threshold manifests as a hypersphere that accepts the inliers while rejecting those outliers. Unquestionably, discerning it presents its challenges in high-dimensional space [34]. In response, we venture into an alternate perspective. Some other threshold detectors are mostly based on statistical analysis that can be utilized for any given $1 - d$ outlier score function [28, 31, 32]. The representative + +works involve Kernel-based [43, 48]; Curve-based [17]; Normality-based: [5, 10]; Filtering-based [19, 55]; Statistical-based [1, 3, 4, 6-8, 11, 12, 16, 24, 37, 49] and Transformation-based [25, 46]. In practice, the default thresholding follows a single thresholding paradigm is not realistic since the perfect discriminativeness of the outlier score function is usually not assumed. Therefore, this work introduces a new perspective: multiple thresholding. + +# 3 Overview + +# 3.1 Problem Definition + +Before introducing our method, we first formally define the problem. The target dataset $\mathbb{Z} = \mathbb{Z}_{in}\cup \mathbb{Z}_{out}$ involves $n$ unlabelled images, here $\mathbb{Z}_{in}$ and $\mathbb{Z}_{out}$ represent inliers (normal data) and outliers (anomalous data), respectively. The outlier (contamination) ratio $\gamma \in (0,1)$ is denoted as $\frac{\#\mathbb{Z}_{out}}{\#\mathbb{Z}_{in} + \#\mathbb{Z}_{out}}$ . Notably, this setup differs from the related semi-supervised outlier detection task by eliminating the need for a prior train/test split for the target dataset. Aligning with previous works [28, 32, 34], the raw images are extracted to feature representation $\mathbf{X} = \mathbf{X}_{\mathrm{in}}\cup \mathbf{X}_{\mathrm{out}} = \{\mathbf{x}_i\}_{i = 1}^n$ with pixel representation [32, 34] or pre-trained neural networks (e.g., ResNet [21] and CLIP [45]). The main objective of UOD is to create an outlier score function $\mathrm{F}(\cdot)$ to evaluate the likelihood of an image feature $\mathbf{x}_i\in \mathbf{X}$ being outliers: + +$$ +\mathrm {Y} (\mathbf {x} _ {i}) = \left\{ \begin{array}{l} 0, \mathrm {F} (\mathbf {x} _ {i}) < \phi \\ 1, \mathrm {F} (\mathbf {x} _ {i}) \geq \phi , \end{array} \right. \tag {1} +$$ + +where $\mathrm{Y} = 0$ (inlier)/1 (outlier) refers to predicted labels and $\phi$ is the threshold (decision boundary). In this study, we not only center on evaluating the ranking accuracy of $\mathrm{F}(\cdot)$ [28,32,34,50], but also measure the efficacy of the threshold $\phi$ . + +# 3.2 Rethinking UOD + +As there is no prior train/test split for the target dataset, a classic outlier scoring method $\mathrm{M}(\cdot)$ predicts the outlier score function $\mathrm{F}(\mathbf{X}) = \{\mathrm{F}(\mathbf{x}_i)\}_{i=1}^n$ follows the below process: + +$$ +\mathrm {F} (\mathbf {X}) = \mathrm {M}. \operatorname {f i t} (\mathbf {X}). \operatorname {p r e d i c t} (\mathbf {X}), \tag {2} +$$ + +where $\mathrm{fit}(\cdot)$ and $\mathrm{predict}(\cdot)$ refer to fitting the target dataset and predicting outlier scores, respectively. Motivated by previous research [32, 33], Shell normalization S-norm $(\cdot)$ illustrated below is an ideal feature de-noising/refining paradigm. + +$$ +\operatorname {S - n o r m} \left(\mathbf {x} _ {i}, \mathbf {v} _ {\mathrm {S}}\right) = \frac {\mathbf {x} _ {i} - \mathbf {v} _ {\mathrm {S}}}{\left| \left| \mathbf {x} _ {i} - \mathbf {v} _ {\mathrm {S}} \right| \right| _ {2}}, \mathbf {v} _ {\mathrm {S}} = \left[ \frac {1}{n} \sum_ {i = 1} ^ {n} \mathbf {X} _ {\text {o u t}} [ i ] [ 1 ], \dots , \frac {1}{n} \sum_ {i = 1} ^ {n} \mathbf {X} _ {\text {o u t}} [ i ] [ d ] \right], \tag {3} +$$ + +where $||\cdot ||_2$ refers to $\ell_2$ -norm and $d$ is the feature dimension. Thus, our first stage is to identify outlier candidates $\mathbf{X}_{\mathrm{out}}^{\prime}$ . Subsequently, as how some one-class learning-based methods (e.g., OCVSM [51], DeepSVDD [50]) usually perform, we secondly predict the inliers $\mathbf{X}_{\mathrm{in}}^{\prime}$ to fit an inlier/normal manifold. Thus, our objective is to identify both $\mathbf{X}_{\mathrm{in}}^{\prime}$ and $\mathbf{X}_{\mathrm{out}}^{\prime}$ . + +![](images/83c573f03cd01b2363343e817f095bb8d99b2e5f646cfeae7bf746ecdadac16c.jpg) +Fig. 2: The overall paradigm of adopting the multi-thresholds learning (Multi-T) module to advance the existing outlier scoring methods. (a) The preparation consists of feature extraction and an initial outlier score function. (b) Visualization of our Multi-T module. (c) Integrating the predicted inliers and outliers with the previously proposed outlier detectors and obtaining an enhanced outlier score function. + +# 4 Method + +Our solution is displayed as below subsections. Sec. 4.1: we introduce the initial outlier score function; Sec. 4.2: we present the Multi-T module, to generate two thresholds for separating inliers and outliers from the target dataset; Sec. 4.3: we leverage predicted inliers and outliers to advance previously proposed methods. + +# 4.1 Initial Outlier Score Function + +In our approach, we employ LVAD [34], one of the UOD SOTAs, as the initial outlier score function. LVAD assumes each image feature $\mathbf{x}_i \in \mathbf{X}$ comes from one of $T$ high-dimensional generative processes. Thus, the outlier score becomes a sum of the weighted distance of the given instance, $\mathbf{x}_i$ , arising from each generative process $\{w_t, \mathbf{m}_t \mid t \in \{1, \dots, T\}\}$ , which can be simplified as: + +$$ +\mathrm {F} _ {\mathrm {L V A D}} (\mathbf {x} _ {i}) = \sum_ {t = 1} ^ {T} w _ {t} \cdot \operatorname {D i s t} (\mathrm {E} - \operatorname {n o r m} (\mathbf {x} _ {i}), \mathrm {E} - \operatorname {n o r m} (\mathbf {m} _ {t})), \tag {4} +$$ + +where $\mathrm{Dist}(\cdot)$ refers to the $\ell_2$ -norm metric (Euclidean distance), $w_{t}$ is the weight of the $t$ -th generative process, $\mathbf{m}_t$ is the $t$ -th generative process's centroid and $\mathrm{E - norm}(\cdot)$ refers to the Ergodic-set normalization, an effective and outlier ratio + +Algorithm 1: MTL +1 Input: Initial Outlier Score Function: $\mathrm{F_{init}(X)}$ , Outlier Scoring Method: M(·) +2 Output: Advanced Outlier Score Function: $\mathrm{F_{M + Multi - T}(X)}$ +3 // Stage 1: Identifying Uncontaminated Inliers +4 for $b$ in iterations do +5 if not converged then +6 $\begin{array}{rlr}\widehat{\mathrm{F}}_{\mathrm{init}}(\mathbf{X}) = \mathrm{sort}(\mathrm{F}_{\mathrm{init}}(\mathbf{X})) & \quad / & \mathrm{sort - transform(Eq.9)}\\ I^b = \left\{\widehat{\mathrm{F}}_{\mathrm{init}}(\mathbf{x}_i)|i < \max_i\left\{g(a_i) > \widehat{\mathrm{F}}_{\mathrm{init}}(\mathbf{x}_i)\right\} \right\} & \quad / & \mathrm{linear - fit(Eq.10)}\\ \phi_{\mathrm{out}}^b = \mathrm{mean}(I^b) + 3\cdot \mathrm{std}(I^b) & \quad / & \mathrm{outlier~threshold~(Eq.11)}\\ I^{b + 1} = I^b.\mathrm{remove}(\{\mathrm{F}_{\mathrm{init}}(\mathbf{x}_i)|\mathrm{F}_{\mathrm{init}}(\mathbf{x}_i) > \phi_{\mathrm{out}}^b\}) & \quad / & \mathrm{filtering(Eq.11)}\\ b = b + 1 \end{array}$ +10 +11 Converged $b:b^{*}$ +12 $\phi_{\mathrm{in}} = \max (I^{b^{*}})$ // inlier threshold(Eq.12) +13 $\mathbf{X}_{\mathrm{in}}^{\prime} = \{\mathbf{x}_{i}|F_{\mathrm{init}}(\mathbf{x}_{i})\leq \phi_{\mathrm{in}}\}$ // predicted inliers(Eq.13) +14 // Stage 2: Identifying Adaptive Outliers +15 $\rho \gets$ similarity(Fs(X),FE(X)) // outlier ratio estimation(Eq.15,16) +16 if $\rho >0.3$ then +17 $\phi_{\mathrm{out}} = \phi_{\mathrm{out}}^{*}$ +18 else +19 if $0.1\leq \rho \leq 0.3$ then +20 $\phi_{\mathrm{out}} = \phi_{\mathrm{out}}^{1}$ +21 else +22 $\begin{array}{rlr}\lfloor & \phi_{\mathrm{out}} = \phi_{\mathrm{out}}^{0}\\ & \end{array}$ +23 $\mathbf{X}_{\mathrm{out}}^{\prime} = \{\mathbf{x}_{i}|F_{\mathrm{init}}(\mathbf{x}_{i}) > \phi_{\mathrm{out}}\}$ // predicted outliers(Eq.20) +24 // Stage 3: Adopting Multi-T to Outlier Scoring Method +25 return M.fit({S-norm(xi,v's)|xi∈X'in}).predict({S-norm(xi,v's)|xi∈X}) + +invariant normalization procedure, illustrated as follows: + +$$ +\operatorname {E} - \operatorname {n o r m} \left(\mathbf {x} _ {i}, \mathbf {v} _ {\mathrm {E}}\right) = \frac {\mathbf {x} _ {i} - \mathbf {v} _ {\mathrm {E}}}{\left| \left| \mathbf {x} _ {i} - \mathbf {v} _ {\mathrm {E}} \right| \right| _ {2}}, \mathbf {v} _ {\mathrm {E}} = \left[ \frac {1}{n \cdot d} \sum_ {i = 1} ^ {n} \sum_ {j = 1} ^ {d} \mathbf {x} _ {i, j}, \dots , \frac {1}{n \cdot d} \sum_ {i = 1} ^ {n} \sum_ {j = 1} ^ {d} \mathbf {x} _ {i, j} \right], \tag {5} +$$ + +where $\mathbf{v}_{\mathrm{E}}$ is its reference vector and $j$ is the dimension index. To gain higher efficiency and maintain efficacy, we set $T = 1$ , such that the initial outlier score function (LVAD-S) is formulated as: + +$$ +\begin{array}{l} \mathrm {F} _ {\text {i n i t}} (\mathbf {X}) = \mathrm {F} _ {\text {L V A D - S}} (\mathbf {X}) = \left. \left\{\operatorname {D i s t} \left(\mathrm {E} - \operatorname {n o r m} \left(\mathbf {x} _ {i}, \mathbf {v} _ {\mathrm {E}}\right), \operatorname {E} - \operatorname {n o r m} \left(\mathbf {m} _ {\mathbf {X}}, \mathbf {v} _ {\mathrm {E}}\right)\right) \right\} _ {i = 1} ^ {n}, \right. \\ \mathbf {m} _ {\mathbf {X}} = \left[ \frac {1}{n} \sum_ {i = 1} ^ {n} \mathbf {x} _ {i, 1}, \dots , \frac {1}{n} \sum_ {i = 1} ^ {n} \mathbf {x} _ {i, d} \right], \tag {6} \\ \end{array} +$$ + +where $\mathbf{m}_{\mathbf{X}}$ is the mean of target dataset's features. Despite its simplicity, Eq. 6 is still a reliable outlier scoring method [34]. + +# 4.2 Multi-T: Multiple Thresholding + +Motivation. In the field of UOD, the conventional thresholding paradigm is to learn a single threshold $\phi$ on the initial outlier score function $\mathrm{F}_{\mathrm{init}}(\mathbf{X})$ . However, it ignores the impact of the outlier ratio $\gamma$ . As the mean of target dataset $\mathbf{m}_{\mathbf{X}}$ will be shifted to outliers when $\gamma$ becomes higher, i.e., mean-shift problem [47], there will be an inevitable overlap between inlier and outlier score distributions. So we define the distribution of the initial outlier score function $D$ as follows: + +$$ +D = \operatorname {F} _ {\text {i n i t}} (\mathbf {X}) = \operatorname {F} _ {\text {i n i t}} \left(\mathbf {X} _ {\text {i n}}\right) \cup \operatorname {F} _ {\text {i n i t}} \left(\mathbf {X} _ {\text {o u t}}\right) = I \cup A \cup O, \tag {7} +$$ + +where $I$ and $O$ refer to the uncontaminated inlier and outlier score distributions, $A$ refers to the overlap. Such that we are concerned about multi-thresholding instead of the conventional way that explicitly finding a single threshold between $\mathrm{F_{init}(\mathbf{X}_{in})}$ and $\mathrm{F_{init}(\mathbf{X}_{out})}$ . Specifically, our objective is identifying the inlier threshold $\phi_{\mathrm{in}}$ and outlier threshold $\phi_{\mathrm{out}}$ that isolate $I$ and $O$ from $D$ . + +Identifying uncontaminated inliers. We identify $I$ by understanding the initial outlier score function (Eq. 6), which computes the normalized $\ell_2$ -norm of each instance $\mathbf{x}_i \in \mathbf{X}$ relative to the mean of $\mathbf{X}$ . Inspired by Shell Theory [32], $I$ satisfies a Gaussian-like distribution, the classic statistical estimation of $I$ is: + +$$ +I = \left\{\mathrm {F} _ {\text {i n i t}} (\mathbf {x} _ {i}) \mid \mathrm {F} _ {\text {i n i t}} (\mathbf {x} _ {i}) < \operatorname {m e a n} (D) + k \cdot \operatorname {s t d} (D) \right\}, \tag {8} +$$ + +where $\mathrm{mean}(D)$ and $\mathrm{std}(D)$ refer to the mean and standard deviation, $k$ is the hyper-parameter of the " $k$ -sigma" rule. In practice, optimizing $k$ without supervision is challenging. Thus, we convert to a non-parametric way by utilizing an ascending sort projection $\mathrm{sort}(\cdot)$ to $\mathrm{F}_{\mathrm{init}}(\mathbf{X})$ , which projects $D$ into a 2- $d$ space ( $x$ -axis: instance index, $y$ -axis: sorted outlier score), the sorted outlier score is: + +$$ +\widehat {\mathrm {F}} _ {\text {i n i t}} (\mathbf {X}) = \operatorname {s o r t} \left(\widehat {\mathrm {F}} _ {\text {i n i t}} (\mathbf {X})\right), i. e., \widehat {\mathrm {F}} _ {\text {i n i t}} [ i + 1 ] > \widehat {\mathrm {F}} _ {\text {i n i t}} [ i ]. \tag {9} +$$ + +Intuitively, $\widehat{I} = \mathrm{sort}(I)$ can be fitted with a naive linear regressor $g(a) = \boldsymbol{\beta}^{\top} \cdot a = \beta_{0} + \beta_{1} \cdot a$ , as shown in Fig. 2 (b). So we identify the potential inliers as: + +$$ +I = \left\{\widehat {\mathrm {F}} _ {\text {i n i t}} (\mathbf {x} _ {i}) | i < \max _ {i} \left\{g \left(a _ {i}\right) > \widehat {\mathrm {F}} _ {\text {i n i t}} (\mathbf {x} _ {i}) \right\} \right\}, s. t. \min _ {\beta} \left| \left| a ^ {\top} \beta - \widehat {\mathrm {F}} _ {\text {i n i t}} (\mathbf {X}) \right| \right| _ {2}, \tag {10} +$$ + +where $\beta$ is the coefficients. Inevitably, this fitting process will be shifted with outliers, which is addressed together with the following outliers' identification. + +Identifying adaptive outliers. In statistics, 3-sigma rule [42] declares variables $C$ has a large probability within three standard deviations from the mean if $C$ follows a Gaussian-like distribution. So we employ it with two aspects: + +(i) iteratively filter outliers during the phase of inliers identification: + +$$ +\phi_ {\text {o u t}} ^ {b} = \operatorname {m e a n} \left(I ^ {b}\right) + 3 \cdot \operatorname {s t d} \left(I ^ {b}\right), I ^ {b + 1} = I ^ {b}. \text {r e m o v e} \left(\left\{\mathrm {F} _ {\text {i n i t}} \left(\mathbf {x} _ {i}\right) \mid \mathrm {F} _ {\text {i n i t}} \left(\mathbf {x} _ {i}\right) > \phi_ {\text {o u t}} ^ {b} \right\}\right), \tag {11} +$$ + +where $b$ refers to the $b$ -th iteration. When converged, the inlier threshold is: + +$$ +\phi_ {\text {i n}} = \max \left(I ^ {b ^ {*}}\right), \tag {12} +$$ + +where $I^{b^*}$ refers to the converged inlier prediction and we can identify inliers as: + +$$ +\mathbf {X} _ {\mathrm {i n}} ^ {\prime} = \left\{\mathbf {x} _ {i} \mid \mathrm {F} _ {\text {i n i t}} \left(\mathbf {x} _ {i}\right) \leq \phi_ {\mathrm {i n}} \right\}. \tag {13} +$$ + +(ii) During the above phase, we estimate the outlier threshold candidates: + +$$ +\phi_ {\mathrm {o u t}} ^ {1} = \operatorname {m e a n} \left(I ^ {1}\right) + 3 \cdot \operatorname {s t d} \left(I ^ {1}\right), \phi_ {\mathrm {o u t}} ^ {*} = \operatorname {m e a n} \left(I ^ {b ^ {*}}\right) + 3 \cdot \operatorname {s t d} \left(I ^ {b ^ {*}}\right), \tag {14} +$$ + +where $\phi_{\mathrm{out}}^{1}$ , $\phi_{\mathrm{out}}^{*}$ refer to the outlier threshold candidates at the first and final iterations, respectively, i.e., $\phi_{\mathrm{out}}^{1} > \phi_{\mathrm{out}}^{*}$ . To obtain a $\gamma$ -adaptive outlier threshold, we consider the impact and relationship of two mentioned normalization procedures: Shell Normalization S-norm( $\cdot$ ) (Eq. 3) and Ergodic-set Normalization E-norm( $\cdot$ ) (Eq. 5). Both them follow a similar normalization formula (denominator: $\ell_2$ -norm), the difference is their reference vectors $\mathbf{v}_{\mathrm{S}}'$ and $\mathbf{v}_{\mathrm{E}}$ , contrast as: + +$$ +\mathbf {v} _ {\mathrm {S}} ^ {\prime} = \left[ \frac {1}{n} \sum_ {i = 1} ^ {n} \mathbf {X} _ {\text {o u t}} ^ {\prime} [ i ] [ 1 ], \dots , \frac {1}{n} \sum_ {i = 1} ^ {n} \mathbf {X} _ {\text {o u t}} ^ {\prime} [ i ] [ d ] \right], \tag {15} +$$ + +$$ +\mathbf {v} _ {\mathrm {E}} = \left[ \frac {1}{n \cdot d} \sum_ {i = 1} ^ {n} \sum_ {j = 1} ^ {d} \mathbf {x} _ {i, j}, \dots , \frac {1}{n \cdot d} \sum_ {i = 1} ^ {n} \sum_ {j = 1} ^ {d} \mathbf {x} _ {i, j} \right]. +$$ + +Obviously, Shell normalization is strongly subjective to the efficacy of outlier prediction $\mathbf{X}_{\mathrm{out}}^{\prime}$ while Ergodic-set Normalization is independent with $\gamma$ that is a sub-optimal but stable operation. To compare these two normalization procedures, we first employ them to Eq. 6, and obtain two outlier score functions: + +$$ +\begin{array}{l} \mathrm {F} _ {\mathrm {S}} (\mathbf {X}) = \left\{\operatorname {D i s t} \left(\mathrm {S} - \operatorname {n o r m} \left(\mathbf {x} _ {i}, \mathbf {v} _ {\mathrm {S}} ^ {\prime}\right), \operatorname {S} - \operatorname {n o r m} \left(\mathbf {m} _ {\mathbf {X}}, \mathbf {v} _ {\mathrm {S}} ^ {\prime}\right)\right) \right\} _ {i = 1} ^ {n}, \\ \mathrm {F} _ {\mathrm {E}} (\mathbf {X}) = \left\{\operatorname {D i s t} \left(\mathrm {E} - \operatorname {n o r m} \left(\mathbf {v} _ {\mathrm {E}} - \mathbf {v} _ {\mathrm {I}}\right), \operatorname {E} - \operatorname {n o r m} \left(\mathbf {m} _ {\mathbf {E}} - \mathbf {v} _ {\mathrm {I}}\right)\right) \right\} ^ {n}. \end{array} \tag {16} +$$ + +$$ +\mathrm {F} _ {\mathrm {E}} (\mathbf {X}) = \left\{\operatorname {D i s t} \left(\operatorname {E} - \operatorname {n o r m} \left(\mathbf {x} _ {i}, \mathbf {v} _ {\mathrm {E}}\right), \operatorname {E} - \operatorname {n o r m} \left(\mathbf {m} _ {\mathbf {X}}, \mathbf {v} _ {\mathrm {E}}\right)\right) \right\} _ {i = 1} ^ {n}, +$$ + +where $\mathbf{v}_{\mathrm{S}}^{\prime}$ comes from $\mathbf{X}_{\mathrm{out}}^{\prime}$ obtained by MAD [6] (Shell Normalization's default thresholding method that is only effective on high- $\gamma$ ). Subsequently, two outlier score functions $\mathrm{F_S}(\mathbf{X})$ and $\mathrm{F_E}(\mathbf{X})$ are arranged in ascending order, and we denote the ranked-index lists as $\mathcal{R}_{\mathrm{S}}$ and $\mathcal{R}_{\mathrm{E}}$ . The outlier ratio $\gamma$ can be approximately estimated with the similarity between $\mathcal{R}_{\mathrm{S}}$ and $\mathcal{R}_{\mathrm{E}}$ since $\mathrm{F_S}(\mathbf{X})$ and $\mathrm{F_E}(\mathbf{X})$ share a consistent structure (distance-to-the-mean), making $\mathcal{R}_{\mathrm{S}}$ and $\mathcal{R}_{\mathrm{E}}$ comparable. Besides, $\mathrm{F_E}(\mathbf{X})$ serves as a reliably effective baseline. When $\gamma$ is low, S-norm $(\mathbf{X},\mathbf{v}_{\mathrm{S}}^{\prime})$ tends to under-perform, lead the similarity between $\mathcal{R}_{\mathrm{S}}$ and $\mathcal{R}_{\mathrm{E}}$ becomes low. In contrast, with a high $\gamma$ , S-norm $(\mathbf{X},\mathbf{v}_{\mathrm{S}}^{\prime})$ merely refines the ranking of a small number of outliers, indicative of high similarity. The similarity $\rho \in [-1,1]$ is computed with the Pearson correlation coefficient: + +$$ +\rho = \frac {\operatorname {c o v} \left(\mathcal {R} _ {\mathrm {S}} , \mathcal {R} _ {\mathrm {E}}\right)}{\operatorname {s t d} \left(\mathcal {R} _ {\mathrm {S}}\right) \cdot \operatorname {s t d} \left(\mathcal {R} _ {\mathrm {E}}\right)}, \tag {17} +$$ + +where $\operatorname{cov}(\mathcal{R}_{\mathrm{S}},\mathcal{R}_{\mathrm{E}})$ and $\mathrm{std}(\mathcal{R}_{\mathrm{S}}) / \mathrm{std}(\mathcal{R}_{\mathrm{E}})$ are the covariance and standard deviation of the rank variables. If the similarity is relatively low ( $< 0.1$ ), we compute the 3-sigma of the entire outlier score distribution $D$ as the outlier threshold: + +$$ +\phi_ {\text {o u t}} ^ {0} = \operatorname {m e a n} (D) + 3 \cdot \operatorname {s t d} (D). \tag {18} +$$ + +It is a conservative estimator for outliers since the above linear fitting might suffer from some challenges at low- $\gamma$ since there is no explicit outlier score distribution. If the similarity is large ( $>0.3$ ), we choose the relatively smooth outlier threshold $\phi_{\mathrm{out}}^{*}$ , while $\phi_{\mathrm{out}}^{1}$ is used for middle- $\gamma$ cases. The outlier threshold is described as: + +$$ +\phi_ {\text {o u t}} = \left\{ \begin{array}{l} \phi_ {\text {o u t}} ^ {*}, \text {i f} \rho > 0. 3 \\ \phi_ {\text {o u t}} ^ {1}, \text {i f} 0. 1 \leq \rho \leq 0. 3 \\ \phi_ {\text {o u t}} ^ {0}, \text {o t h e r w i s e} \end{array} \right. \tag {19} +$$ + +Therefore, we can subsequently identify outlier ratio invariant outliers as: + +$$ +\mathbf {X} _ {\text {o u t}} ^ {\prime} = \left\{\mathbf {x} _ {i} \mid \mathrm {F} _ {\text {i n i t}} \left(\mathbf {x} _ {i}\right) > \phi_ {\text {o u t}} \right\}. \tag {20} +$$ + +# 4.3 Implementation + +Firstly, we can directly employ Multi-T with the distance-to-the-mean paradigm: + +$$ +\mathrm {F} _ {\mathrm {M u l t i - T}} (\mathbf {X}) = \{\mathrm {D i s t} (\mathrm {S - n o r m} (\mathbf {x} _ {i}, \mathbf {v} _ {\mathrm {S}} ^ {\prime}), \mathrm {S - n o r m} (\mathbf {m} _ {\mathbf {X} _ {\mathrm {i n}} ^ {\prime}}, \mathbf {v} _ {\mathrm {S}} ^ {\prime})) \} _ {i = 1} ^ {n}, +$$ + +$$ +\mathbf {v} _ {\mathrm {S}} ^ {\prime} = \left[ \frac {1}{n} \sum_ {i = 1} ^ {n} \mathbf {X} _ {\text {o u t}} ^ {\prime} [ i ] [ 1 ], \dots , \frac {1}{n} \sum_ {i = 1} ^ {n} \mathbf {X} _ {\text {o u t}} ^ {\prime} [ i ] [ d ] \right], \tag {21} +$$ + +$$ +\mathbf {m} _ {\mathbf {X} _ {\mathrm {i n}} ^ {\prime}} = \left[ \frac {1}{n} \sum_ {i = 1} ^ {n} \mathbf {X} _ {\mathrm {i n}} ^ {\prime} [ i ] [ 1 ], \dots , \frac {1}{n} \sum_ {i = 1} ^ {n} \mathbf {X} _ {\mathrm {i n}} ^ {\prime} [ i ] [ d ] \right]. +$$ + +Secondly, Multi-T can be integrated with UOD methods $\mathrm{M}(\cdot)$ , illustrated as: + +$$ +\begin{array}{l} \mathrm {F} _ {\mathrm {M} + \text {M u l t i - T}} (\mathbf {X}) = \mathrm {M}. \operatorname {f i t} \left(\left\{\mathrm {S} - \operatorname {n o r m} \left(\mathbf {x} _ {i}, \mathbf {v} _ {\mathrm {S}} ^ {\prime}\right) \mid \mathbf {x} _ {i} \in \mathbf {X} _ {i n} ^ {\prime} \right\}\right) \\ . \operatorname {p r e d i c t} \left(\left\{\mathrm {S} - \operatorname {n o r m} \left(\mathbf {x} _ {i}, \mathbf {v} _ {\mathrm {S}} ^ {\prime}\right) \mid \mathbf {x} _ {i} \in \mathbf {X} \right\}\right). \\ \end{array} +$$ + +# 5 Experiments + +We cover our experiments for (i) comparing with other threshold learners on dataset splitting (Sec. 5.1); (ii) measuring the performance improvement of adopting the Multi-T to previously proposed outlier scoring methods (Sec. 5.2). + +Benchmarks and Feature Extraction. Based on the existing research [32,34], we use the raw pixel representation for grayscale datasets, including MNIST [29] and Fashion-MNIST [59]. For RGB datasets such as STL-10 [15], Internet [32], CIFAR-10 [27], MIT-Places-Small [61], we adopt two deep feature extractors: ImageNet pretrained [20] ResNet-50 [21] (ResNet)2, and CLIP [45]. + +Target Dataset. In our experiments, each class within a benchmark dataset is alternatively regarded as inliers, with instances from all other classes considered outliers. The target dataset includes all inliers along with a randomly selected subset of outliers. Results for each dataset are averaged across all classes. Additionally, for every round, we further average the results over a wide range of outlier ratios: $[0.05, 0.1, 0.2, 0.3, 0.4]^3$ , to assess the method's robustness/stability. + +![](images/0ea2d465b66a5324d1e7a34f36b80e91c9d0a7e96f996495888d2e720529f2ed.jpg) + +![](images/ba6b4d142c17f8360942e86c2b65bf5a75fe039c58dc6e41df6d9f4276f2a032.jpg) +Fig.3: Qualitative results compared with SOTA threshold learners. The experiments are conducted on STL-10 (Inlier class: Monkey). + +# 5.1 Threshold Learning + +Competing Methods. We compare our proposed Multi-T with those effective thresholding (TL) methods involving Kernel-based: AUCP [48], FGD [43]; Curve-based: EB [17]; Normality-based: DSN [5], CHAU [10]; Filtering-based: FILTER [19], HIST [55]; Statistical-based: MTT [49], BOOT [37], QMCD [24], CLF [7], IQR [8], KARCH [1], MCST [16], GESD [4], REGR [3], MAD [6], CLUST [11], CPD [12] and Transformation-based: MOLL [25], YJ [46]. + +Evaluation Metric. The performance of general thresholding is measured with $\mathrm{F}_{\beta}$ -score, defined as follows: + +$$ +\mathrm {F} _ {\beta} = \left(1 + \beta^ {2}\right) \cdot \frac {\text {p r e c i s i o n} \cdot \text {r e c a l l}}{\left(\beta^ {2} \cdot \text {p r e c i s i o n}\right) + \text {r e c a l l}}, \tag {23} +$$ + +where $\beta < 1$ tilts towards precision while $\beta > 1$ prioritizes recall. In this work, we utilize $\mathrm{F}_{0.1}$ -score $(\mathrm{F}_{0.1})$ and $\mathrm{F}_{10}$ -score $(\mathrm{F}_{10})$ that measure the accuracy of predicted outliers and inliers. Besides, $\mathrm{F}_{1}$ -score $(\beta = 1)$ , a harmonic mean of the precision and recall, is not practical for this task since the compete separability between inliers' and outliers' score distributions cannot be guaranteed. + +Main Results. Across a series of competing threshold learners, while one method may achieve a commendable $\mathrm{F}_{0.1}$ -score of 0.976, the corresponding $\mathrm{F}_{10}$ -score is mere 0.452, and vice versa ( $\mathrm{F}_{10}$ -score: 0.979; $\mathrm{F}_{0.1}$ -score: 0.447). By contrast, our solution maintains its efficacy across various evaluation criteria ( $\mathrm{F}_{0.1}$ -score: 0.917, $\mathrm{F}_{10}$ -score: 0.980). It exhibits remarkable stability across a diverse spectrum of outlier ratios, as illustrated in Fig. 3. To further investigate the benefits of our technique, Tab. 1 categorizes two groups, each representing a different + +Table 1: Average $\mathrm{F}_{\beta}$ -score of thresholding methods on STL-10. "+"GT Norm" refers to the ideal initial outlier score function. "Highest $\mathrm{F}_{\beta}$ " ( $\beta: 0.1, 10$ ) refers to the method that garners the highest $\mathrm{F}_{\beta}$ -score across all baseline models. The better average result is highlighted in bold. + +
Outlier Score Fun.Method (TL)ResNet-50CLIP
F0.1F10Avg.F0.1F10Avg.
LVAD-S [34]Highest F0.10.9110.4540.6820.9360.4010.669
Highest F100.3820.9670.6740.3570.9540.655
Multi-T(Ours)0.8400.8690.8550.8600.8660.863
+GT Norm [32]Highest F0.10.9760.4520.7140.9770.4480.713
Highest F100.4470.9790.7130.6910.9860.839
Multi-T(Ours)0.9170.9800.9490.8570.9840.920
+ +outlier score function. Such categorization is pivotal as threshold methods ought to accommodate various foundational bases. + +Table 2: Average AUC results compared with SOTA outlier scoring methods (outlier detectors). Blue and Orange indicates the best and second-best results, respectively. + +
Method (OS)STL-10CIFAR-10CIFAR-100MIT-PlacesMNIST
ResNetCLIPResNetCLIPResNetCLIPResNetCLIPPixel
IF [35]0.8360.9430.7800.8910.7900.8660.6870.8680.776
LOF [11]0.6280.6260.6730.6210.8390.8490.5560.5200.754
RSRAE [28]0.9620.9380.8620.8790.9140.8850.8740.8760.793
ECOD [31]0.9070.9810.8730.9350.8730.9180.7770.9430.734
LUNAR [18]0.7760.8210.7670.7740.8380.8990.6430.8710.797
Shell-Re. [32]0.8620.8380.8600.8130.8350.8130.8260.9140.776
LVAD [34]0.9540.9680.8600.9170.9210.9170.8440.9190.867
Multi-T0.9680.9890.8950.9570.9380.9560.8670.9740.897
DeepSVDD [50]0.6220.5970.5600.5090.5630.5810.5830.5490.513
+Multi-T0.9250.9210.7690.8190.8350.8260.7550.8320.732
Improve.48.7%54.3%37.3%60.9%48.3%42.2%29.5%51.6%37.9%
OCSVM [51]0.9270.9210.8260.8500.8790.8500.8260.8710.831
+Multi-T0.9570.9650.8590.9160.9160.8990.8460.9240.863
Improve.3.24%4.78%4.00%7.76%4.21%5.76%2.42%6.08%3.85%
+ +# 5.2 Outlier Scoring + +Competing Methods. We compare two themes of baseline outlier scoring (OS) methods: (i) statistical-based methods: IF [35], OCSVM [51], DeepSVDD [51], ECOD [31], LUNAR [28], RSRAE [28], Shell-Re. [32], LVAD [34]. (ii) deep-learning-based models: GOAD [9], ICL [53], REPEN [39], NeuTraL [44], SLAD [60]. For a fair comparison, we apply Ergodic-set normalization [34] if it improves the performance of the baseline algorithms, such as IF [35], OCSVM [51]. + +Evaluation Metric. The performance of outlier scoring (ranking accuracy) is primarily assessed using the Area Under the Receiver Operating Characteristic curve (AUC). This metric provides a thorough assessment of ranking accuracy. Main Results. Tab. 2 shows that Multi-T can be seamlessly integrated with two classic outlier scoring methods: DeepSVDD [51], and OCSVM [51] while exhibiting significant improvements in both efficacy and stability across a diverse range of outlier ratios and various benchmarks. In most of our experiments, Multi-T itself achieves SOTA results. Notably, even in the case of non-aligned and low-resolution datasets like CIFAR-10 [27] and CIFAR-100 [27], which are known to pose challenges [41]. It surpasses the current SOTA AUC scores by margins of $7.76\%$ and $5.76\%$ for CIFAR-10 and CIFAR-100, respectively. Additionally, we observe a logical and significant enhancement in performance with improved feature representation, e.g., for the MIT-Places [61] dataset, the AUC improves from 0.867 using ResNet-50 [21] to 0.974 with CLIP [45]. + +Table 3: Average AUC results, for more outlier scoring methods with $(\mathrm{w} / )$ or without(w/o) our Multi-T module, conducted on STL-10. + +
FeatureIF [35]ECOD [11]ABOD [26]PCA [54]GMM [2]
w/ow/w/ow/w/ow/w/ow/w/ow/
ResNet0.8360.8990.9070.9190.6650.8830.8650.9450.8590.952
CLIP0.9430.9830.9810.9840.7150.9090.9840.9940.8920.962
Avg.0.8900.9410.9440.9510.6900.8960.9250.9700.8760.957
FeatureGOAD [9]ICL [53]REPEN [39]NeuTraL [44]SLAD [60]
w/ow/w/ow/w/ow/w/ow/w/ow/
ResNet0.9520.9620.9340.9570.8770.8890.8540.9500.9410.962
CLIP0.9610.9890.9510.9820.8790.9300.8510.9710.9450.985
Avg.0.9570.9750.9430.9700.8780.9090.8530.9610.9430.973
+ +In Tab. 3, we present the results of Multi-T integrated with more methods involving both statistical and deep models, which shows its broad applications. Moreover, our solution excels not just in detection accuracy but also in running speed since Multi-T compresses the size of fitting data, as illustrated in Tab. 4, 5. In Fig. 4, we compare with the most related work Shell-Re. [32], which + +has a built-in Robust-Least-Square (RLS) thresholding procedure with MAD [6]. Apparently, our method is capable of estimating multiple thresholds that facilitate a clear demarcation of diverse scenarios (outlier ratio, dataset domain, and feature representation). + +Table 4: Efficiency comparison for outlier Table 5: Efficiency comparison for scoring. Timing is measured with 10,000 thresholding. All thresholding methods samples, GPU: NVIDIA RTX 3080. are conducted on CPU. + +
Method (OS)DeviceTime (↓)
LVAD [34]CPU645.2s
RSRAE [28]GPU121.6s
Multi-TCPU1.2s
OCSVM [51]CPU154.8s
+Multi-T105.4s
GOAD [9]GPU268.1s
+Multi-T211.7s
+ +
Method (TL)Time (↓)
CPD [12]3.61s
FWFM [19]1.26s
DSN [5]3.79s
CLUST [11]7.36s
AUCP [48]1.26s
Multi-T0.97s
+ +![](images/e1672596f140f350da057673c08f03e6b135ce868f81e8e8d74e03bb65541534.jpg) +Fig. 4: AUC results of Multi-T compared with Shell-Re. [32] over various outlier ratios. + +Discussions. We conclude the feasibility of adopting Multi-T to outlier scoring methods into three aspects: (i) both normalization with outliers and manifold learning with inliers are necessary for most of outlier scoring methods; (ii) Despite the initial outlier score distribution (Eq. 6) is usually not perfectly separable, it provides a reliable baseline. (iii) Multi-T is able to isolate the overlap region and predict both uncontaminated inliers. + +Table 6: Ablation study for DeepSVDD and LVAD-S with Multi-T ( $\gamma$ : 0.2). + +
DeepSVDDAUCLVAD-SAUC
+Multi-T0.950+Multi-T0.971
w/o In0.643w/o In0.965
w/o Out0.939w/o Out0.951
+ +Table 7: Comparison with $k$ -sigma. + +
MetricMethodScore
F0.13-sigma0.406
Multi-T0.840
F101-sigma0.806
Multi-T0.869
+ +# 5.3 Ablation Study + +The experiments thus far have established the effectiveness of Multi-T. However, there remains a concern regarding how sensitive the performance improvement is to Multi-T's two primary components: adaptive outliers (Out) and uncontaminated inliers (In), which refer to the normalization and manifold learning procedures, respectively. We select two representative outlier scoring models: DeepSVDD [50] and LVAD-S [34], whereas DeepSVDD follows a widely-used outlier detection mechanism, i.e., learning the normality (hyper-sphere), and identifying inliers is of great significance. Without predicted inliers, the AUC result decreases from 0.950 to 0.643 on STL-10, shown in Tab. 6. Additionally, since the normalization procedure can be considered as a distance de-noising procedure [33], Multi-T will contribute to those methods with distance computation, e.g., the result of LVAD-S is improved from 0.951 to 0.971 on STL-10 and 0.838 to 0.882 on CIFAR-10, which verify our prior assumption that the identification of both inliers and outliers is of great significance, which enhances the reliability of multiple thresholding perspective. Moreover, Tab. 7 indicates that our thresholding process markedly surpasses the classical $k$ -sigma rule [42]. + +# 5.4 Limitation + +Our method is tied to the ranking accuracy (separability) of the initial outlier score function. Based on most related works, we consider UOD as an one-class learning task (only one inlier/normal class in the target dataset). However we find in some specific cases, there might exist multi-normal classes, e.g., the digit 5 class of the MNIST dataset. In that case, the efficacy of our method might be decreased. Additionally, this statistical-based method has limitations in very small-scale datasets since the "three-sigma" rule will be less effective. + +# 6 Conclusion + +This work introduces a novel perspective for UOD about advancing existing outlier detectors (scoring methods) via thresholding. To this end, we propose a multiple thresholding (Multi-T) module, to label the unlabelled target dataset. Comprehensive experiments verify that the Multi-T can significantly improve both the efficacy and efficiency of previously proposed outlier scoring methods. + +# Acknowledgement + +We would like to thank the reviewers for their valuable comments and suggestions. + +# References + +1. Afsari, B.: Riemannian $L^p$ center of mass: existence, uniqueness, and convexity. Proceedings of the American Mathematical Society pp. 655-673 (2011) +2. Aggarwal, C.C.: Outlier analysis second edition (2016) +3. Aggarwal, C.C., Aggarwal, C.C.: An introduction to outlier analysis. Springer (2017) +4. Alrawashdeh, M.J.: An adjusted grubbs' and generalized extreme studentized deviation. Demonstratio Mathematica pp. 548-557 (2021) +5. Amagata, D., Onizuka, M., Hara, T.: Fast and exact outlier detection in metric spaces: a proximity graph-based approach. In: Proceedings of the 2021 International Conference on Management of Data. pp. 36-48 (2021) +6. Archana, N., Pawar, S.: Periodicity detection of outlier sequences using constraint based pattern tree with mad. International Journal of Advanced Studies in Computers, Science and Engineering p. 34 (2015) +7. Barbado, A., Corcho, O., Benjamins, R.: Rule extraction in unsupervised anomaly detection for model explainability: Application to oneclasssvm. Expert Systems with Applications p. 116100 (2022) +8. Bardet, J.M., Dimby, S.F.: A new non-parametric detector of univariate outliers for distributions with unbounded support. Extremes pp. 751-775 (2017) +9. Bergman, L., Hoshen, Y.: Classification-based anomaly detection for general data. arXiv preprint arXiv:2005.02359 (2020) +10. Bol'shev, L., Ubaidullaeva, M.: Chauvenet's test in the classical theory of errors. Theory of Probability & Its Applications pp. 683-692 (1975) +1. Breunig, M.M., Kriegel, H.P., Ng, R.T., Sander, J.: Lof: identifying density-based local outliers. In: Proceedings of the ACM SIGMOD International Conference on Management of Data. pp. 93-104 (2000) +2. Van den Burg, G.J., Williams, C.K.: An evaluation of change point detection algorithms. arXiv preprint arXiv:2003.06222 (2020) +3. Camposeco, F., Sattler, T., Cohen, A., Geiger, A., Pollefeys, M.: Toroidal constraints for two-point localization under high outlier ratios. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 4545-4553 (2017) +4. Chalapathy, R., Menon, A.K., Chawla, S.: Robust, deep and inductive anomaly detection. In: Proceedings of Joint European Conference on Machine Learning and Knowledge Discovery in Databases. pp. 36-51 (2017) +5. Coates, A., Ng, A., Lee, H.: An analysis of single-layer networks in unsupervised feature learning. In: Proceedings of the International Conference on Artificial Intelligence and Statistics. pp. 215-223 (2011) +6. Coin, D.: Testing normality in the presence of outliers. Statistical Methods and Applications pp. 3-12 (2008) +7. Friendly, M., Monette, G., Fox, J.: Elliptical insights: understanding statistical methods through elliptical geometry. Statistical Science pp. 1-39 (2013) + +18. Goodge, A., Hooi, B., Ng, S.K., Ng, W.S.: Lunar: Unifying local outlier detection methods via graph neural networks. In: Proceedings of the AAAI Conference on Artificial Intelligence. pp. 6737-6745 (2022) +19. Hashemi, N., German, E.V., Ramirez, J.P., Ruths, J.: Filtering approaches for dealing with noise in anomaly detection. In: 2019 IEEE 58th Conference on Decision and Control. pp. 5356-5361 (2019) +20. He, K., Girshick, R., Dollar, P.: Rethinking imagenet pre-training. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 4918-4927 (2019) +21. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 770-778 (2016) +22. Hendrycks, D., Mazeika, M., Dietterich, T.: Deep anomaly detection with outlier exposure. arXiv preprint arXiv:1812.04606 (2018) +23. Huyan, N., Quan, D., Zhang, X., Liang, X., Chanussot, J., Jiao, L.: Unsupervised outlier detection using memory and contrastive learning. IEEE Transactions on Image Processing pp. 6440-6454 (2022) +24. Iouchtchenko, D., Raymond, N., Roy, P.N., Nooijen, M.: Deterministic and quasi-random sampling of optimized gaussian mixture distributions for vibronic monte carlo. arXiv preprint arXiv:1912.11594 (2019) +25. Keyzer, M.A., Sonneveld, B.: Using the mollifier method to characterize datasets and models: the case of the universal soil loss equation. ITC Journal pp. 263-272 (1997) +26. Kriegel, H.P., Schubert, M., Zimek, A.: Angle-based outlier detection in high-dimensional data. In: Proceedings of the 14th ACM SIGKDD international conference on Knowledge discovery and data mining. pp. 444-452 (2008) +27. Krizhevsky, A., Hinton, G., et al.: Learning multiple layers of features from tiny images (2009) +28. Lai, C.H., Zou, D., Lerman, G.: Robust subspace recovery layer for unsupervised anomaly detection. arXiv preprint arXiv:1904.00152 (2019) +29. LeCun, Y., Cortes, C.: MNIST handwritten digit database (2010) +30. Li, T., Wang, Z., Liu, S., Lin, W.Y.: Deep unsupervised anomaly detection. In: Proceedings of IEEE Winter Conference on Applications of Computer Vision. pp. 3636-3645 (2021) +31. Li, Z., Zhao, Y., Hu, X., Botta, N., Ionescu, C., Chen, G.H.: Ecod: Unsupervised outlier detection using empirical cumulative distribution functions. arXiv preprint arXiv:2201.00382 (2022) +32. Lin, D., Liu, S., Li, H., Cheung, N.M., Ren, C., Matsushita, Y.: Shell theory: A statistical model of reality. IEEE Transactions on Pattern Analysis and Machine Intelligence (2021) +33. Lin, W.Y., Liu, S., Dai, B.T., Li, H.: Distance based image classification: A solution to generative classification's conundrum? International Journal of Computer Vision pp. 177-198 (2023) +34. Lin, W.Y., Liu, Z., Liu, S.: Locally varying distance transform for unsupervised visual anomaly detection. In: European Conference on Computer Vision. pp. 354-371. Springer (2022) +35. Liu, F.T., Ting, K.M., Zhou, Z.H.: Isolation forest. In: Proceedings of International Conference on Data Mining. pp. 413-422 (2008) +36. Liznerski, P., Ruff, L., Vandermeulen, R.A., Franks, B.J., Müller, K.R., Kloft, M.: Exposing outlier exposure: What can be learned from few, one, and zero outlier images. arXiv preprint arXiv:2205.11474 (2022) + +37. Martin, M.A., Roberts, S.: An evaluation of bootstrap methods for outlier detection in least squares regression. Journal of Applied Statistics pp. 703-720 (2006) +38. Masci, J., Meier, U., Ciresan, D., Schmidhuber, J.: Stacked convolutional autoencoders for hierarchical feature extraction. In: Proceedings of International Conference on Artificial Neural Networks. pp. 52-59 (2011) +39. Pang, G., Cao, L., Chen, L., Liu, H.: Learning representations of ultrahigh-dimensional data for random distance-based outlier detection. In: Proceedings of the 24th ACM SIGKDD international conference on knowledge discovery & data mining. pp. 2041-2050 (2018) +40. Parzen, E.: On estimation of a probability density function and mode. The annals of mathematical statistics pp. 1065-1076 (1962) +41. Perera, P., Nallapati, R., Xiang, B.: Ocgan: One-class novelty detection using gans with constrained latent representations. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 2898-2906 (2019) +42. Pukelsheim, F.: The three sigma rule. The American Statistician pp. 88-91 (1994) +43. Qi, Z., Jiang, D., Chen, X.: Iterative gradient descent for outlier detection. International Journal of Wavelets, Multiresolution and Information Processing p. 2150004 (2021) +44. Qiu, C., Pfrommer, T., Kloft, M., Mandt, S., Rudolph, M.: Neural transformation learning for deep anomaly detection beyond images. In: International Conference on Machine Learning. pp. 8703-8714 (2021) +45. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: Proceedings of International Conference on Machine Learning. pp. 8748-8763 (2021) +46. Raymaekers, J., Rousseeuw, P.J.: Transforming variables to central normality. Machine Learning pp. 1-23 (2021) +47. Reiss, T., Hoshen, Y.: Mean-shifted contrastive loss for anomaly detection. In: Proceedings of the AAAI Conference on Artificial Intelligence. pp. 2155-2162 (2023) +48. Ren, K., Yang, H., Zhao, Y., Chen, W., Xue, M., Miao, H., Huang, S., Liu, J.: A robust auc maximization framework with simultaneous outlier detection and feature selection for positive-unlabeled classification. IEEE Transactions on Neural Networks and Learning Systems pp. 3072-3083 (2018) +49. Rengasamy, D., Rothwell, B.C., Figueredo, G.P.: Towards a more reliable interpretation of machine learning outputs for safety-critical systems using feature importance fusion. Applied Sciences p. 11854 (2021) +50. Ruff, L., Vandermeulen, R., Goernitz, N., Deecke, L., Siddiqui, S.A., Binder, A., Müller, E., Kloft, M.: Deep one-class classification. In: Proceedings of International Conference on Machine Learning. pp. 4393-4402 (2018) +51. Scholkopf, B., Platt, J.C., Shawe-Taylor, J., Smola, A.J., Williamson, R.C.: Estimating the support of a high-dimensional distribution. Neural Computation 13(7), 1443-1471 (2001) +52. Scholkopf, B., Smola, A.J., Bach, F., et al.: Learning with kernels: support vector machines, regularization, optimization, and beyond (2002) +53. Shenkar, T., Wolf, L.: Anomaly detection for tabular data with internal contrastive learning. In: International Conference on Learning Representations (2021) +54. Shyu, M.L., Chen, S.C., Sarinnapakorn, K., Chang, L.: A novel anomaly detection scheme based on principal component classifier. In: Proceedings of the IEEE foundations and new directions of data mining workshop. pp. 172-179 (2003) + +55. Thanammal, K., Jayasudha, J., Vijayalakshmi, R., Arumugaperumal, S.: Effective histogram thresholding techniques for natural images using segmentation. Journal of Image and Graphics pp. 113-116 (2014) +56. Wang, S., Zeng, Y., Liu, X., Zhu, E., Yin, J., Xu, C., Kloft, M.: Effective end-to-end unsupervised outlier detection via inlier priority of discriminative network. Proceedings of Advances in Neural Information Processing Systems (2019) +57. Wang, S., Zhu, E., Hu, X., Liu, X., Liu, Q., Yin, J., Wang, F.: Robustness can be cheap: A highly efficient approach to discover outliers under high outlier ratios. In: Proceedings of the AAAI Conference on Artificial Intelligence. pp. 5313-5320 (2019) +58. Xia, Y., Cao, X., Wen, F., Hua, G., Sun, J.: Learning discriminative reconstructions for unsupervised outlier removal. In: Proceedings of the IEEE International Conference on Computer Vision. pp. 1511-1519 (2015) +59. Xiao, H., Rasul, K., Vollgraf, R.: Fashion-mnist: a novel image dataset for benchmarking machine learning algorithms (2017) +60. Xu, H., Wang, Y., Wei, J., Jian, S., Li, Y., Liu, N.: Fascinating supervisory signals and where to find them: Deep anomaly detection with scale learning. arXiv preprint arXiv:2305.16114 (2023) +61. Zhou, B., Lapedriza, A., Khosla, A., Oliva, A., Torralba, A.: Places: A 10 million image database for scene recognition. IEEE Transactions on Pattern Analysis and Machine Intelligence pp. 1452-1464 (2017) +62. Zhou, C., Paffenroth, R.C.: Anomaly detection with robust deep autoencoders. In: Proceedings of the ACM SIGKDD International Conference on Knowledge Discovery and Data Mining. pp. 665-674 (2017) +63. Zong, B., Song, Q., Min, M.R., Cheng, W., Lumezanu, C., Cho, D., Chen, H.: Deep autoencoding gaussian mixture model for unsupervised anomaly detection. In: Proceedings of International Conference on Learning Representations (2018) \ No newline at end of file diff --git a/2024/Rethinking Unsupervised Outlier Detection via Multiple Thresholding/images.zip b/2024/Rethinking Unsupervised Outlier Detection via Multiple Thresholding/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..eb40422b5bf608cf24c9611b78a517a784653925 --- /dev/null +++ b/2024/Rethinking Unsupervised Outlier Detection via Multiple Thresholding/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:689cbb9f918fb24f5439f3b7cb475a5cc20114ca35d0d23189f12daca7c4c5dc +size 811018 diff --git a/2024/Rethinking Unsupervised Outlier Detection via Multiple Thresholding/layout.json b/2024/Rethinking Unsupervised Outlier Detection via Multiple Thresholding/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..db969812068884a74558491c51a92cf7fb971ecc --- /dev/null +++ b/2024/Rethinking Unsupervised Outlier Detection via Multiple Thresholding/layout.json @@ -0,0 +1,11873 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 141, + 111, + 473, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 111, + 473, + 148 + ], + "spans": [ + { + "bbox": [ + 141, + 111, + 473, + 148 + ], + "type": "text", + "content": "Rethinking Unsupervised Outlier Detection via Multiple Thresholding" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 138, + 168, + 476, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 168, + 476, + 194 + ], + "spans": [ + { + "bbox": [ + 138, + 168, + 476, + 194 + ], + "type": "text", + "content": "Zhonghang Liu1, Panzhong Lu2, Guoyang Xie3,4, Zhichao Lu4, and Wen-Yan Lin1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 232, + 202, + 383, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 232, + 202, + 383, + 224 + ], + "spans": [ + { + "bbox": [ + 232, + 202, + 383, + 224 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 232, + 202, + 383, + 224 + ], + "type": "text", + "content": " Singapore Management University " + }, + { + "bbox": [ + 232, + 202, + 383, + 224 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 232, + 202, + 383, + 224 + ], + "type": "text", + "content": " Westlake University" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 201, + 224, + 412, + 237 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 201, + 224, + 412, + 237 + ], + "spans": [ + { + "bbox": [ + 201, + 224, + 412, + 237 + ], + "type": "text", + "content": "3 Contemporary Amperex Technology Co., Limited" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 242, + 236, + 373, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 242, + 236, + 373, + 247 + ], + "spans": [ + { + "bbox": [ + 242, + 236, + 373, + 247 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 242, + 236, + 373, + 247 + ], + "type": "text", + "content": " City University of Hong Kong" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 242, + 247, + 372, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 242, + 247, + 372, + 258 + ], + "spans": [ + { + "bbox": [ + 242, + 247, + 372, + 258 + ], + "type": "text", + "content": "zhliu.2020@phdcs.smu.edu.sg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 244, + 259, + 369, + 269 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 244, + 259, + 369, + 269 + ], + "spans": [ + { + "bbox": [ + 244, + 259, + 369, + 269 + ], + "type": "text", + "content": "lupanzhong@westlake.edu.cn" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 258, + 270, + 356, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 258, + 270, + 356, + 281 + ], + "spans": [ + { + "bbox": [ + 258, + 270, + 356, + 281 + ], + "type": "text", + "content": "guoyang.xie@ieee.org" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 257, + 281, + 357, + 291 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 257, + 281, + 357, + 291 + ], + "spans": [ + { + "bbox": [ + 257, + 281, + 357, + 291 + ], + "type": "text", + "content": "luzhichao.cn@gmail.com" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 258, + 293, + 355, + 302 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 258, + 293, + 355, + 302 + ], + "spans": [ + { + "bbox": [ + 258, + 293, + 355, + 302 + ], + "type": "text", + "content": "daniellin@smu.edu.sg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 159, + 340, + 455, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 340, + 455, + 506 + ], + "spans": [ + { + "bbox": [ + 159, + 340, + 455, + 506 + ], + "type": "text", + "content": "Abstract. In the realm of unsupervised image outlier detection, assigning outlier scores holds greater significance than its subsequent task: thresholding for predicting labels. This is because determining the optimal threshold on non-separable outlier score functions is an ill-posed problem. However, the lack of predicted labels not only hinders some real applications of current outlier detectors but also causes these methods not to be enhanced by leveraging the dataset's self-supervision. To advance existing scoring methods, we propose a multiple thresholding (Multi-T) module. It generates two thresholds that isolate inliers and outliers from the unlabelled target dataset, whereas outliers are employed to obtain better feature representation while inliers provide an uncontaminated manifold. Extensive experiments verify that Multi-T can significantly improve proposed outlier scoring methods. Moreover, Multi-T contributes to a naive distance-based method being state-of-the-art. Code is available at: https://github.com/zhliu-uod/Multi-T." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 159, + 516, + 453, + 539 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 516, + 453, + 539 + ], + "spans": [ + { + "bbox": [ + 159, + 516, + 453, + 539 + ], + "type": "text", + "content": "Keywords: Unsupervised outlier detection, multiple thresholding, outlier scoring" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 563, + 230, + 575 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 563, + 230, + 575 + ], + "spans": [ + { + "bbox": [ + 132, + 563, + 230, + 575 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": "Which is more important for the unsupervised outlier detection (UOD), outlier score or label? Currently, the mainstream UOD approaches [23, 28, 31, 56] focus on the first, i.e., learning a discriminative outlier score function. However, those recent outlier detectors are usually complex and non-iterative since the absence of predicted labels limits these methods to be further advanced with the dataset's self-supervision. Such that, this work tends to label the unlabelled tar" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 132, + 114, + 482, + 190 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 482, + 190 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 482, + 190 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 482, + 190 + ], + "type": "image", + "image_path": "9080d1e2120f51744f2d3474d25e21c9a67b27c903f68baf00619bd1eb1b6ef1.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 198, + 482, + 243 + ], + "lines": [ + { + "bbox": [ + 130, + 198, + 482, + 243 + ], + "spans": [ + { + "bbox": [ + 130, + 198, + 482, + 243 + ], + "type": "text", + "content": "Fig. 1: Unlike the default UOD paradigm (a) concerned about learning a discriminative score function, our perspective (b) is to explore the dataset's prior knowledge via thresholding, to advance the previously proposed method. (c) illustrates the significant improvement of DeepSVDD [50] with Multi-T module." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 268, + 479, + 292 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 268, + 479, + 292 + ], + "spans": [ + { + "bbox": [ + 130, + 268, + 479, + 292 + ], + "type": "text", + "content": "get dataset, thereby leveraging the dataset's prior knowledge to enhance simple outlier scoring methods" + }, + { + "bbox": [ + 130, + 268, + 479, + 292 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 130, + 268, + 479, + 292 + ], + "type": "text", + "content": ", achieving state-of-the-art (SOTA) results efficiently." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 293, + 480, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 293, + 480, + 399 + ], + "spans": [ + { + "bbox": [ + 130, + 293, + 480, + 399 + ], + "type": "text", + "content": "To understand the importance of the target dataset's self-supervision, we first decompose the predominant image outlier scoring approaches [9,32,50,51] into two continuous stages: inlier/normal manifold learning and distance/similarity inference. There is no doubt that identifying inliers will contribute to learning an uncontaminated inlier manifold (e.g. the hypersphere of DeepSVDD [50]). Additionally, motivated by Shell Theory [32,33] in the high-dimensional space, image feature representation for distance computation can be significantly improved via outliers-based Shell Normalization [32]. Such that it is critical to obtain the thresholds that separate both inliers and outliers from target datasets." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 401, + 480, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 401, + 480, + 472 + ], + "spans": [ + { + "bbox": [ + 130, + 401, + 480, + 472 + ], + "type": "text", + "content": "To align with our above objective, we present a Multi-T module, which follows a multiple thresholding mechanism. Compared with conventional single thresholding methods [19, 24, 48, 49], Multi-T generates two distinct thresholds that separately isolate outliers and inliers from the target dataset, which is capable of dealing with the inevitable inseparability of the initial given score distribution. This allows us to directly implement Multi-T on a simple outlier score function." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 474, + 480, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 474, + 480, + 545 + ], + "spans": [ + { + "bbox": [ + 130, + 474, + 480, + 545 + ], + "type": "text", + "content": "By understanding the impact of the outlier ratio, Multi-T involves two stages: Uncontaminated inliers. We employ an iterative two-step process (i) identifying a noisy inlier distribution through the analysis of sorted initial score function, i.e., Ergodic-set normalized [34] distance to the mean of target dataset; (ii) subsequently filtering outliers with 3-sigma rule [42]. When converged, we can identify uncontaminated inliers, thereby training an inlier/normal manifold." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 547, + 480, + 618 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 547, + 480, + 618 + ], + "spans": [ + { + "bbox": [ + 130, + 547, + 480, + 618 + ], + "type": "text", + "content": "Adaptive outliers. The uncontaminated inliers identification process also outputs a series of outlier threshold candidates. Subsequently, we compare the ranking-index similarity between Ergodic-set Normalization [34] and Shell Normalization [32] on the initial outlier score function. The structural consistency and contrastive properties of the two normalization procedures motivate us to implicitly estimate a rough outlier ratio, thereby detecting the adaptive outliers." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 619, + 480, + 643 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 619, + 480, + 643 + ], + "spans": [ + { + "bbox": [ + 130, + 619, + 480, + 643 + ], + "type": "text", + "content": "The Multi-T module is training-free, highly efficient and grants the adaptiveness to choose thresholds suited to downstream implementations. As evaluated" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 216, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 216, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 216, + 100 + ], + "type": "text", + "content": "Z. Liu et al." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 133, + 653, + 400, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 653, + 400, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 653, + 400, + 665 + ], + "type": "text", + "content": "1 Outlier scoring methods refers to the classical outlier detectors." + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 199 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 199 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 199 + ], + "type": "text", + "content": "with comprehensive experimental settings, Multi-T helps to boost significant performance (efficacy and efficiency) improvement. For example, the AUC score of DeepSVDD [50] is improved from 0.622 to 0.925 on STL-10 dataset [15]. Moreover, the naive distance-to-the-target dataset's mean [32] metric integrated with Multi-T achieves SOTA results with only 1.2 second consumption for 10,000 ResNet-50 [21] samples that are orders of magnitude faster than baselines. Our primary contributions are summarized below:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 200, + 479, + 270 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 130, + 200, + 479, + 223 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 200, + 479, + 223 + ], + "spans": [ + { + "bbox": [ + 130, + 200, + 479, + 223 + ], + "type": "text", + "content": "- We introduce a novel perspective for UOD about enhancing previously proposed scoring methods via thresholding on a simple initial outlier score function." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 224, + 479, + 246 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 224, + 479, + 246 + ], + "spans": [ + { + "bbox": [ + 130, + 224, + 479, + 246 + ], + "type": "text", + "content": "- We present a multiple thresholding (Multi-T) module that generates two different thresholds to separate both inliers and outliers from the target dataset." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 247, + 479, + 270 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 247, + 479, + 270 + ], + "spans": [ + { + "bbox": [ + 130, + 247, + 479, + 270 + ], + "type": "text", + "content": "- The efficacy of integrated scoring methods can be significantly improved without the external complexity increase, as evaluated in extensive experiments." + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 131, + 291, + 237, + 303 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 291, + 237, + 303 + ], + "spans": [ + { + "bbox": [ + 131, + 291, + 237, + 303 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 317, + 321, + 330 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 317, + 321, + 330 + ], + "spans": [ + { + "bbox": [ + 131, + 317, + 321, + 330 + ], + "type": "text", + "content": "2.1 Unsupervised Outlier Detection" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 338, + 482, + 541 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 338, + 482, + 541 + ], + "spans": [ + { + "bbox": [ + 130, + 338, + 482, + 541 + ], + "type": "text", + "content": "Classic UOD task aims to assign an outlier score/likelihood to an image sample. The recent models can be divided into inlier-manifold learning and outlier exposure. Inlier-manifold learning assumes that inliers are the majority. Deep learning-based outlier detectors, as illustrated in studies [14, 28, 30, 38, 56, 58, 62, 63], typically focus on reducing the dimensionality of image data by projecting it into a latent and discriminative space. Additionally, various statistics-based methods model the inlier manifold using discrimination-based [32, 34] or density-based approaches [11, 26, 31, 40]. In applications where outliers are not the minority [13, 57], different approaches are required. Some outlier exposure methods [22, 36] employ out-of-distribution (OOD) data to train networks for detecting unseen OOD samples. For example, Shell-Renormalization [32] employs predicted outlier candidates to iteratively refine feature representation. It is particularly useful in scenarios where classic manifold learners, assuming outliers as a minority, may not be effective. However, these two paradigms are separable, this work will attempt to design a thresholding framework that generates both inliers and outliers, leveraging the benefits of both. Inlier-manifold learning and outlier exposure." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 560, + 228, + 573 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 560, + 228, + 573 + ], + "spans": [ + { + "bbox": [ + 131, + 560, + 228, + 573 + ], + "type": "text", + "content": "2.2 Thresholding" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 581, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 581, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 581, + 481, + 665 + ], + "type": "text", + "content": "Some traditional outlier detection methods [32, 35, 52], provide both the outlier score and its corresponding threshold concurrently. In this context, the threshold manifests as a hypersphere that accepts the inliers while rejecting those outliers. Unquestionably, discerning it presents its challenges in high-dimensional space [34]. In response, we venture into an alternate perspective. Some other threshold detectors are mostly based on statistical analysis that can be utilized for any given " + }, + { + "bbox": [ + 130, + 581, + 481, + 665 + ], + "type": "inline_equation", + "content": "1 - d" + }, + { + "bbox": [ + 130, + 581, + 481, + 665 + ], + "type": "text", + "content": " outlier score function [28, 31, 32]. The representative" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 162, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 162, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 162, + 91, + 447, + 102 + ], + "type": "text", + "content": "Rethinking Unsupervised Outlier Detection via Multiple Thresholding" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "text", + "content": "works involve Kernel-based [43, 48]; Curve-based [17]; Normality-based: [5, 10]; Filtering-based [19, 55]; Statistical-based [1, 3, 4, 6-8, 11, 12, 16, 24, 37, 49] and Transformation-based [25, 46]. In practice, the default thresholding follows a single thresholding paradigm is not realistic since the perfect discriminativeness of the outlier score function is usually not assumed. Therefore, this work introduces a new perspective: multiple thresholding." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 203, + 212, + 217 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 203, + 212, + 217 + ], + "spans": [ + { + "bbox": [ + 132, + 203, + 212, + 217 + ], + "type": "text", + "content": "3 Overview" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 226, + 258, + 237 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 226, + 258, + 237 + ], + "spans": [ + { + "bbox": [ + 132, + 226, + 258, + 237 + ], + "type": "text", + "content": "3.1 Problem Definition" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 243, + 482, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 243, + 482, + 373 + ], + "spans": [ + { + "bbox": [ + 130, + 243, + 482, + 373 + ], + "type": "text", + "content": "Before introducing our method, we first formally define the problem. The target dataset " + }, + { + "bbox": [ + 130, + 243, + 482, + 373 + ], + "type": "inline_equation", + "content": "\\mathbb{Z} = \\mathbb{Z}_{in}\\cup \\mathbb{Z}_{out}" + }, + { + "bbox": [ + 130, + 243, + 482, + 373 + ], + "type": "text", + "content": " involves " + }, + { + "bbox": [ + 130, + 243, + 482, + 373 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 130, + 243, + 482, + 373 + ], + "type": "text", + "content": " unlabelled images, here " + }, + { + "bbox": [ + 130, + 243, + 482, + 373 + ], + "type": "inline_equation", + "content": "\\mathbb{Z}_{in}" + }, + { + "bbox": [ + 130, + 243, + 482, + 373 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 243, + 482, + 373 + ], + "type": "inline_equation", + "content": "\\mathbb{Z}_{out}" + }, + { + "bbox": [ + 130, + 243, + 482, + 373 + ], + "type": "text", + "content": " represent inliers (normal data) and outliers (anomalous data), respectively. The outlier (contamination) ratio " + }, + { + "bbox": [ + 130, + 243, + 482, + 373 + ], + "type": "inline_equation", + "content": "\\gamma \\in (0,1)" + }, + { + "bbox": [ + 130, + 243, + 482, + 373 + ], + "type": "text", + "content": " is denoted as " + }, + { + "bbox": [ + 130, + 243, + 482, + 373 + ], + "type": "inline_equation", + "content": "\\frac{\\#\\mathbb{Z}_{out}}{\\#\\mathbb{Z}_{in} + \\#\\mathbb{Z}_{out}}" + }, + { + "bbox": [ + 130, + 243, + 482, + 373 + ], + "type": "text", + "content": ". Notably, this setup differs from the related semi-supervised outlier detection task by eliminating the need for a prior train/test split for the target dataset. Aligning with previous works [28, 32, 34], the raw images are extracted to feature representation " + }, + { + "bbox": [ + 130, + 243, + 482, + 373 + ], + "type": "inline_equation", + "content": "\\mathbf{X} = \\mathbf{X}_{\\mathrm{in}}\\cup \\mathbf{X}_{\\mathrm{out}} = \\{\\mathbf{x}_i\\}_{i = 1}^n" + }, + { + "bbox": [ + 130, + 243, + 482, + 373 + ], + "type": "text", + "content": " with pixel representation [32, 34] or pre-trained neural networks (e.g., ResNet [21] and CLIP [45]). The main objective of UOD is to create an outlier score function " + }, + { + "bbox": [ + 130, + 243, + 482, + 373 + ], + "type": "inline_equation", + "content": "\\mathrm{F}(\\cdot)" + }, + { + "bbox": [ + 130, + 243, + 482, + 373 + ], + "type": "text", + "content": " to evaluate the likelihood of an image feature " + }, + { + "bbox": [ + 130, + 243, + 482, + 373 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_i\\in \\mathbf{X}" + }, + { + "bbox": [ + 130, + 243, + 482, + 373 + ], + "type": "text", + "content": " being outliers:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 255, + 375, + 482, + 399 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 255, + 375, + 482, + 399 + ], + "spans": [ + { + "bbox": [ + 255, + 375, + 482, + 399 + ], + "type": "interline_equation", + "content": "\\mathrm {Y} (\\mathbf {x} _ {i}) = \\left\\{ \\begin{array}{l} 0, \\mathrm {F} (\\mathbf {x} _ {i}) < \\phi \\\\ 1, \\mathrm {F} (\\mathbf {x} _ {i}) \\geq \\phi , \\end{array} \\right. \\tag {1}", + "image_path": "53afe8321d1acc29beb11a6befdae7ab7f32d43aaadefdcf04d19c893f769706.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 402, + 482, + 439 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 402, + 482, + 439 + ], + "spans": [ + { + "bbox": [ + 131, + 402, + 482, + 439 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 402, + 482, + 439 + ], + "type": "inline_equation", + "content": "\\mathrm{Y} = 0" + }, + { + "bbox": [ + 131, + 402, + 482, + 439 + ], + "type": "text", + "content": " (inlier)/1 (outlier) refers to predicted labels and " + }, + { + "bbox": [ + 131, + 402, + 482, + 439 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 131, + 402, + 482, + 439 + ], + "type": "text", + "content": " is the threshold (decision boundary). In this study, we not only center on evaluating the ranking accuracy of " + }, + { + "bbox": [ + 131, + 402, + 482, + 439 + ], + "type": "inline_equation", + "content": "\\mathrm{F}(\\cdot)" + }, + { + "bbox": [ + 131, + 402, + 482, + 439 + ], + "type": "text", + "content": " [28,32,34,50], but also measure the efficacy of the threshold " + }, + { + "bbox": [ + 131, + 402, + 482, + 439 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 131, + 402, + 482, + 439 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 453, + 247, + 464 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 453, + 247, + 464 + ], + "spans": [ + { + "bbox": [ + 132, + 453, + 247, + 464 + ], + "type": "text", + "content": "3.2 Rethinking UOD" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 470, + 481, + 505 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 470, + 481, + 505 + ], + "spans": [ + { + "bbox": [ + 130, + 470, + 481, + 505 + ], + "type": "text", + "content": "As there is no prior train/test split for the target dataset, a classic outlier scoring method " + }, + { + "bbox": [ + 130, + 470, + 481, + 505 + ], + "type": "inline_equation", + "content": "\\mathrm{M}(\\cdot)" + }, + { + "bbox": [ + 130, + 470, + 481, + 505 + ], + "type": "text", + "content": " predicts the outlier score function " + }, + { + "bbox": [ + 130, + 470, + 481, + 505 + ], + "type": "inline_equation", + "content": "\\mathrm{F}(\\mathbf{X}) = \\{\\mathrm{F}(\\mathbf{x}_i)\\}_{i=1}^n" + }, + { + "bbox": [ + 130, + 470, + 481, + 505 + ], + "type": "text", + "content": " follows the below process:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 242, + 506, + 480, + 518 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 242, + 506, + 480, + 518 + ], + "spans": [ + { + "bbox": [ + 242, + 506, + 480, + 518 + ], + "type": "interline_equation", + "content": "\\mathrm {F} (\\mathbf {X}) = \\mathrm {M}. \\operatorname {f i t} (\\mathbf {X}). \\operatorname {p r e d i c t} (\\mathbf {X}), \\tag {2}", + "image_path": "be5bdfb7c51e9cfb6b5a014c65051a44983a136caec649cfae6010bfc447bef9.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 522, + 482, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 522, + 482, + 559 + ], + "spans": [ + { + "bbox": [ + 130, + 522, + 482, + 559 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 522, + 482, + 559 + ], + "type": "inline_equation", + "content": "\\mathrm{fit}(\\cdot)" + }, + { + "bbox": [ + 130, + 522, + 482, + 559 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 522, + 482, + 559 + ], + "type": "inline_equation", + "content": "\\mathrm{predict}(\\cdot)" + }, + { + "bbox": [ + 130, + 522, + 482, + 559 + ], + "type": "text", + "content": " refer to fitting the target dataset and predicting outlier scores, respectively. Motivated by previous research [32, 33], Shell normalization S-norm " + }, + { + "bbox": [ + 130, + 522, + 482, + 559 + ], + "type": "inline_equation", + "content": "(\\cdot)" + }, + { + "bbox": [ + 130, + 522, + 482, + 559 + ], + "type": "text", + "content": " illustrated below is an ideal feature de-noising/refining paradigm." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 143, + 565, + 481, + 605 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 565, + 481, + 605 + ], + "spans": [ + { + "bbox": [ + 143, + 565, + 481, + 605 + ], + "type": "interline_equation", + "content": "\\operatorname {S - n o r m} \\left(\\mathbf {x} _ {i}, \\mathbf {v} _ {\\mathrm {S}}\\right) = \\frac {\\mathbf {x} _ {i} - \\mathbf {v} _ {\\mathrm {S}}}{\\left| \\left| \\mathbf {x} _ {i} - \\mathbf {v} _ {\\mathrm {S}} \\right| \\right| _ {2}}, \\mathbf {v} _ {\\mathrm {S}} = \\left[ \\frac {1}{n} \\sum_ {i = 1} ^ {n} \\mathbf {X} _ {\\text {o u t}} [ i ] [ 1 ], \\dots , \\frac {1}{n} \\sum_ {i = 1} ^ {n} \\mathbf {X} _ {\\text {o u t}} [ i ] [ d ] \\right], \\tag {3}", + "image_path": "3c44682eda9b8edf68f01c460e8afb687125b3f11dc973a5e2123fd1370bcfa2.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 606, + 482, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 606, + 482, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 606, + 482, + 667 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 606, + 482, + 667 + ], + "type": "inline_equation", + "content": "||\\cdot ||_2" + }, + { + "bbox": [ + 130, + 606, + 482, + 667 + ], + "type": "text", + "content": " refers to " + }, + { + "bbox": [ + 130, + 606, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\ell_2" + }, + { + "bbox": [ + 130, + 606, + 482, + 667 + ], + "type": "text", + "content": "-norm and " + }, + { + "bbox": [ + 130, + 606, + 482, + 667 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 130, + 606, + 482, + 667 + ], + "type": "text", + "content": " is the feature dimension. Thus, our first stage is to identify outlier candidates " + }, + { + "bbox": [ + 130, + 606, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_{\\mathrm{out}}^{\\prime}" + }, + { + "bbox": [ + 130, + 606, + 482, + 667 + ], + "type": "text", + "content": ". Subsequently, as how some one-class learning-based methods (e.g., OCVSM [51], DeepSVDD [50]) usually perform, we secondly predict the inliers " + }, + { + "bbox": [ + 130, + 606, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_{\\mathrm{in}}^{\\prime}" + }, + { + "bbox": [ + 130, + 606, + 482, + 667 + ], + "type": "text", + "content": " to fit an inlier/normal manifold. Thus, our objective is to identify both " + }, + { + "bbox": [ + 130, + 606, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_{\\mathrm{in}}^{\\prime}" + }, + { + "bbox": [ + 130, + 606, + 482, + 667 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 606, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_{\\mathrm{out}}^{\\prime}" + }, + { + "bbox": [ + 130, + 606, + 482, + 667 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "type": "text", + "content": "Z. Liu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 130, + 114, + 482, + 304 + ], + "blocks": [ + { + "bbox": [ + 130, + 114, + 482, + 304 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 482, + 304 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 482, + 304 + ], + "type": "image", + "image_path": "83c573f03cd01b2363343e817f095bb8d99b2e5f646cfeae7bf746ecdadac16c.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 311, + 482, + 367 + ], + "lines": [ + { + "bbox": [ + 130, + 311, + 482, + 367 + ], + "spans": [ + { + "bbox": [ + 130, + 311, + 482, + 367 + ], + "type": "text", + "content": "Fig. 2: The overall paradigm of adopting the multi-thresholds learning (Multi-T) module to advance the existing outlier scoring methods. (a) The preparation consists of feature extraction and an initial outlier score function. (b) Visualization of our Multi-T module. (c) Integrating the predicted inliers and outliers with the previously proposed outlier detectors and obtaining an enhanced outlier score function." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 391, + 201, + 403 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 391, + 201, + 403 + ], + "spans": [ + { + "bbox": [ + 132, + 391, + 201, + 403 + ], + "type": "text", + "content": "4 Method" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 420, + 482, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 420, + 482, + 469 + ], + "spans": [ + { + "bbox": [ + 130, + 420, + 482, + 469 + ], + "type": "text", + "content": "Our solution is displayed as below subsections. Sec. 4.1: we introduce the initial outlier score function; Sec. 4.2: we present the Multi-T module, to generate two thresholds for separating inliers and outliers from the target dataset; Sec. 4.3: we leverage predicted inliers and outliers to advance previously proposed methods." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 491, + 309, + 502 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 491, + 309, + 502 + ], + "spans": [ + { + "bbox": [ + 132, + 491, + 309, + 502 + ], + "type": "text", + "content": "4.1 Initial Outlier Score Function" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 515, + 482, + 576 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 515, + 482, + 576 + ], + "spans": [ + { + "bbox": [ + 130, + 515, + 482, + 576 + ], + "type": "text", + "content": "In our approach, we employ LVAD [34], one of the UOD SOTAs, as the initial outlier score function. LVAD assumes each image feature " + }, + { + "bbox": [ + 130, + 515, + 482, + 576 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_i \\in \\mathbf{X}" + }, + { + "bbox": [ + 130, + 515, + 482, + 576 + ], + "type": "text", + "content": " comes from one of " + }, + { + "bbox": [ + 130, + 515, + 482, + 576 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 130, + 515, + 482, + 576 + ], + "type": "text", + "content": " high-dimensional generative processes. Thus, the outlier score becomes a sum of the weighted distance of the given instance, " + }, + { + "bbox": [ + 130, + 515, + 482, + 576 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_i" + }, + { + "bbox": [ + 130, + 515, + 482, + 576 + ], + "type": "text", + "content": ", arising from each generative process " + }, + { + "bbox": [ + 130, + 515, + 482, + 576 + ], + "type": "inline_equation", + "content": "\\{w_t, \\mathbf{m}_t \\mid t \\in \\{1, \\dots, T\\}\\}" + }, + { + "bbox": [ + 130, + 515, + 482, + 576 + ], + "type": "text", + "content": ", which can be simplified as:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 189, + 586, + 481, + 619 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 189, + 586, + 481, + 619 + ], + "spans": [ + { + "bbox": [ + 189, + 586, + 481, + 619 + ], + "type": "interline_equation", + "content": "\\mathrm {F} _ {\\mathrm {L V A D}} (\\mathbf {x} _ {i}) = \\sum_ {t = 1} ^ {T} w _ {t} \\cdot \\operatorname {D i s t} (\\mathrm {E} - \\operatorname {n o r m} (\\mathbf {x} _ {i}), \\mathrm {E} - \\operatorname {n o r m} (\\mathbf {m} _ {t})), \\tag {4}", + "image_path": "7622bfe323db9d46668d93e0984e7d11916d088ff445bd8fc6d9a2cafafdf3a2.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\mathrm{Dist}(\\cdot)" + }, + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "text", + "content": " refers to the " + }, + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\ell_2" + }, + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "text", + "content": "-norm metric (Euclidean distance), " + }, + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "inline_equation", + "content": "w_{t}" + }, + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "text", + "content": " is the weight of the " + }, + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "text", + "content": "-th generative process, " + }, + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\mathbf{m}_t" + }, + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "text", + "content": " is the " + }, + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "text", + "content": "-th generative process's centroid and " + }, + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\mathrm{E - norm}(\\cdot)" + }, + { + "bbox": [ + 130, + 629, + 482, + 667 + ], + "type": "text", + "content": " refers to the Ergodic-set normalization, an effective and outlier ratio" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 162, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 162, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 162, + 91, + 448, + 102 + ], + "type": "text", + "content": "Rethinking Unsupervised Outlier Detection via Multiple Thresholding" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "code", + "bbox": [ + 134, + 132, + 482, + 437 + ], + "blocks": [ + { + "bbox": [ + 140, + 119, + 233, + 130 + ], + "lines": [ + { + "bbox": [ + 140, + 119, + 233, + 130 + ], + "spans": [ + { + "bbox": [ + 140, + 119, + 233, + 130 + ], + "type": "text", + "content": "Algorithm 1: MTL" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "code_caption" + }, + { + "bbox": [ + 134, + 132, + 482, + 437 + ], + "lines": [ + { + "bbox": [ + 134, + 132, + 482, + 437 + ], + "spans": [ + { + "bbox": [ + 134, + 132, + 482, + 437 + ], + "type": "text", + "content": "1 Input: Initial Outlier Score Function: " + }, + { + "bbox": [ + 134, + 132, + 482, + 437 + ], + "type": "inline_equation", + "content": "\\mathrm{F_{init}(X)}" + }, + { + "bbox": [ + 134, + 132, + 482, + 437 + ], + "type": "text", + "content": " , Outlier Scoring Method: M(·) \n2 Output: Advanced Outlier Score Function: " + }, + { + "bbox": [ + 134, + 132, + 482, + 437 + ], + "type": "inline_equation", + "content": "\\mathrm{F_{M + Multi - T}(X)}" + }, + { + "bbox": [ + 134, + 132, + 482, + 437 + ], + "type": "text", + "content": " \n3 // Stage 1: Identifying Uncontaminated Inliers \n4 for " + }, + { + "bbox": [ + 134, + 132, + 482, + 437 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 134, + 132, + 482, + 437 + ], + "type": "text", + "content": " in iterations do \n5 if not converged then \n6 " + }, + { + "bbox": [ + 134, + 132, + 482, + 437 + ], + "type": "inline_equation", + "content": "\\begin{array}{rlr}\\widehat{\\mathrm{F}}_{\\mathrm{init}}(\\mathbf{X}) = \\mathrm{sort}(\\mathrm{F}_{\\mathrm{init}}(\\mathbf{X})) & \\quad / & \\mathrm{sort - transform(Eq.9)}\\\\ I^b = \\left\\{\\widehat{\\mathrm{F}}_{\\mathrm{init}}(\\mathbf{x}_i)|i < \\max_i\\left\\{g(a_i) > \\widehat{\\mathrm{F}}_{\\mathrm{init}}(\\mathbf{x}_i)\\right\\} \\right\\} & \\quad / & \\mathrm{linear - fit(Eq.10)}\\\\ \\phi_{\\mathrm{out}}^b = \\mathrm{mean}(I^b) + 3\\cdot \\mathrm{std}(I^b) & \\quad / & \\mathrm{outlier~threshold~(Eq.11)}\\\\ I^{b + 1} = I^b.\\mathrm{remove}(\\{\\mathrm{F}_{\\mathrm{init}}(\\mathbf{x}_i)|\\mathrm{F}_{\\mathrm{init}}(\\mathbf{x}_i) > \\phi_{\\mathrm{out}}^b\\}) & \\quad / & \\mathrm{filtering(Eq.11)}\\\\ b = b + 1 \\end{array}" + }, + { + "bbox": [ + 134, + 132, + 482, + 437 + ], + "type": "text", + "content": " \n10 \n11 Converged " + }, + { + "bbox": [ + 134, + 132, + 482, + 437 + ], + "type": "inline_equation", + "content": "b:b^{*}" + }, + { + "bbox": [ + 134, + 132, + 482, + 437 + ], + "type": "text", + "content": " \n12 " + }, + { + "bbox": [ + 134, + 132, + 482, + 437 + ], + "type": "inline_equation", + "content": "\\phi_{\\mathrm{in}} = \\max (I^{b^{*}})" + }, + { + "bbox": [ + 134, + 132, + 482, + 437 + ], + "type": "text", + "content": " // inlier threshold(Eq.12) \n13 " + }, + { + "bbox": [ + 134, + 132, + 482, + 437 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_{\\mathrm{in}}^{\\prime} = \\{\\mathbf{x}_{i}|F_{\\mathrm{init}}(\\mathbf{x}_{i})\\leq \\phi_{\\mathrm{in}}\\}" + }, + { + "bbox": [ + 134, + 132, + 482, + 437 + ], + "type": "text", + "content": " // predicted inliers(Eq.13) \n14 // Stage 2: Identifying Adaptive Outliers \n15 " + }, + { + "bbox": [ + 134, + 132, + 482, + 437 + ], + "type": "inline_equation", + "content": "\\rho \\gets" + }, + { + "bbox": [ + 134, + 132, + 482, + 437 + ], + "type": "text", + "content": " similarity(Fs(X),FE(X)) // outlier ratio estimation(Eq.15,16) \n16 if " + }, + { + "bbox": [ + 134, + 132, + 482, + 437 + ], + "type": "inline_equation", + "content": "\\rho >0.3" + }, + { + "bbox": [ + 134, + 132, + 482, + 437 + ], + "type": "text", + "content": " then \n17 " + }, + { + "bbox": [ + 134, + 132, + 482, + 437 + ], + "type": "inline_equation", + "content": "\\phi_{\\mathrm{out}} = \\phi_{\\mathrm{out}}^{*}" + }, + { + "bbox": [ + 134, + 132, + 482, + 437 + ], + "type": "text", + "content": " \n18 else \n19 if " + }, + { + "bbox": [ + 134, + 132, + 482, + 437 + ], + "type": "inline_equation", + "content": "0.1\\leq \\rho \\leq 0.3" + }, + { + "bbox": [ + 134, + 132, + 482, + 437 + ], + "type": "text", + "content": " then \n20 " + }, + { + "bbox": [ + 134, + 132, + 482, + 437 + ], + "type": "inline_equation", + "content": "\\phi_{\\mathrm{out}} = \\phi_{\\mathrm{out}}^{1}" + }, + { + "bbox": [ + 134, + 132, + 482, + 437 + ], + "type": "text", + "content": " \n21 else \n22 " + }, + { + "bbox": [ + 134, + 132, + 482, + 437 + ], + "type": "inline_equation", + "content": "\\begin{array}{rlr}\\lfloor & \\phi_{\\mathrm{out}} = \\phi_{\\mathrm{out}}^{0}\\\\ & \\end{array}" + }, + { + "bbox": [ + 134, + 132, + 482, + 437 + ], + "type": "text", + "content": " \n23 " + }, + { + "bbox": [ + 134, + 132, + 482, + 437 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_{\\mathrm{out}}^{\\prime} = \\{\\mathbf{x}_{i}|F_{\\mathrm{init}}(\\mathbf{x}_{i}) > \\phi_{\\mathrm{out}}\\}" + }, + { + "bbox": [ + 134, + 132, + 482, + 437 + ], + "type": "text", + "content": " // predicted outliers(Eq.20) \n24 // Stage 3: Adopting Multi-T to Outlier Scoring Method \n25 return M.fit({S-norm(xi,v's)|xi∈X'in}).predict({S-norm(xi,v's)|xi∈X})" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "code_body" + } + ], + "index": 3, + "sub_type": "algorithm" + }, + { + "bbox": [ + 132, + 468, + 382, + 480 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 468, + 382, + 480 + ], + "spans": [ + { + "bbox": [ + 132, + 468, + 382, + 480 + ], + "type": "text", + "content": "invariant normalization procedure, illustrated as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 135, + 491, + 480, + 536 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 491, + 480, + 536 + ], + "spans": [ + { + "bbox": [ + 135, + 491, + 480, + 536 + ], + "type": "interline_equation", + "content": "\\operatorname {E} - \\operatorname {n o r m} \\left(\\mathbf {x} _ {i}, \\mathbf {v} _ {\\mathrm {E}}\\right) = \\frac {\\mathbf {x} _ {i} - \\mathbf {v} _ {\\mathrm {E}}}{\\left| \\left| \\mathbf {x} _ {i} - \\mathbf {v} _ {\\mathrm {E}} \\right| \\right| _ {2}}, \\mathbf {v} _ {\\mathrm {E}} = \\left[ \\frac {1}{n \\cdot d} \\sum_ {i = 1} ^ {n} \\sum_ {j = 1} ^ {d} \\mathbf {x} _ {i, j}, \\dots , \\frac {1}{n \\cdot d} \\sum_ {i = 1} ^ {n} \\sum_ {j = 1} ^ {d} \\mathbf {x} _ {i, j} \\right], \\tag {5}", + "image_path": "f07b8ee8bfa989d5535f15e1f6c32a54c1d5cc997315bc52ab9356e56ec7c22a.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 536, + 482, + 572 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 536, + 482, + 572 + ], + "spans": [ + { + "bbox": [ + 130, + 536, + 482, + 572 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 536, + 482, + 572 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_{\\mathrm{E}}" + }, + { + "bbox": [ + 130, + 536, + 482, + 572 + ], + "type": "text", + "content": " is its reference vector and " + }, + { + "bbox": [ + 130, + 536, + 482, + 572 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 130, + 536, + 482, + 572 + ], + "type": "text", + "content": " is the dimension index. To gain higher efficiency and maintain efficacy, we set " + }, + { + "bbox": [ + 130, + 536, + 482, + 572 + ], + "type": "inline_equation", + "content": "T = 1" + }, + { + "bbox": [ + 130, + 536, + 482, + 572 + ], + "type": "text", + "content": ", such that the initial outlier score function (LVAD-S) is formulated as:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 143, + 583, + 481, + 631 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 583, + 481, + 631 + ], + "spans": [ + { + "bbox": [ + 143, + 583, + 481, + 631 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathrm {F} _ {\\text {i n i t}} (\\mathbf {X}) = \\mathrm {F} _ {\\text {L V A D - S}} (\\mathbf {X}) = \\left. \\left\\{\\operatorname {D i s t} \\left(\\mathrm {E} - \\operatorname {n o r m} \\left(\\mathbf {x} _ {i}, \\mathbf {v} _ {\\mathrm {E}}\\right), \\operatorname {E} - \\operatorname {n o r m} \\left(\\mathbf {m} _ {\\mathbf {X}}, \\mathbf {v} _ {\\mathrm {E}}\\right)\\right) \\right\\} _ {i = 1} ^ {n}, \\right. \\\\ \\mathbf {m} _ {\\mathbf {X}} = \\left[ \\frac {1}{n} \\sum_ {i = 1} ^ {n} \\mathbf {x} _ {i, 1}, \\dots , \\frac {1}{n} \\sum_ {i = 1} ^ {n} \\mathbf {x} _ {i, d} \\right], \\tag {6} \\\\ \\end{array}", + "image_path": "759c593e5428f1177c309f41e5812bde3fd9f1a59a18e7dccb25eb26e2ecccc3.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{m}_{\\mathbf{X}}" + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "content": " is the mean of target dataset's features. Despite its simplicity, Eq. 6 is still a reliable outlier scoring method [34]." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "type": "text", + "content": "Z. Liu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 321, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 321, + 129 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 321, + 129 + ], + "type": "text", + "content": "4.2 Multi-T: Multiple Thresholding" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 135, + 479, + 207 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 135, + 479, + 207 + ], + "spans": [ + { + "bbox": [ + 130, + 135, + 479, + 207 + ], + "type": "text", + "content": "Motivation. In the field of UOD, the conventional thresholding paradigm is to learn a single threshold " + }, + { + "bbox": [ + 130, + 135, + 479, + 207 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 130, + 135, + 479, + 207 + ], + "type": "text", + "content": " on the initial outlier score function " + }, + { + "bbox": [ + 130, + 135, + 479, + 207 + ], + "type": "inline_equation", + "content": "\\mathrm{F}_{\\mathrm{init}}(\\mathbf{X})" + }, + { + "bbox": [ + 130, + 135, + 479, + 207 + ], + "type": "text", + "content": ". However, it ignores the impact of the outlier ratio " + }, + { + "bbox": [ + 130, + 135, + 479, + 207 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 130, + 135, + 479, + 207 + ], + "type": "text", + "content": ". As the mean of target dataset " + }, + { + "bbox": [ + 130, + 135, + 479, + 207 + ], + "type": "inline_equation", + "content": "\\mathbf{m}_{\\mathbf{X}}" + }, + { + "bbox": [ + 130, + 135, + 479, + 207 + ], + "type": "text", + "content": " will be shifted to outliers when " + }, + { + "bbox": [ + 130, + 135, + 479, + 207 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 130, + 135, + 479, + 207 + ], + "type": "text", + "content": " becomes higher, i.e., mean-shift problem [47], there will be an inevitable overlap between inlier and outlier score distributions. So we define the distribution of the initial outlier score function " + }, + { + "bbox": [ + 130, + 135, + 479, + 207 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 130, + 135, + 479, + 207 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 192, + 216, + 481, + 228 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 192, + 216, + 481, + 228 + ], + "spans": [ + { + "bbox": [ + 192, + 216, + 481, + 228 + ], + "type": "interline_equation", + "content": "D = \\operatorname {F} _ {\\text {i n i t}} (\\mathbf {X}) = \\operatorname {F} _ {\\text {i n i t}} \\left(\\mathbf {X} _ {\\text {i n}}\\right) \\cup \\operatorname {F} _ {\\text {i n i t}} \\left(\\mathbf {X} _ {\\text {o u t}}\\right) = I \\cup A \\cup O, \\tag {7}", + "image_path": "59e4f3ce3b248634cc5d142258fa9e17f0ee6ced280aa52543299ce177217a96.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 236, + 479, + 295 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 236, + 479, + 295 + ], + "spans": [ + { + "bbox": [ + 131, + 236, + 479, + 295 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 236, + 479, + 295 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 131, + 236, + 479, + 295 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 236, + 479, + 295 + ], + "type": "inline_equation", + "content": "O" + }, + { + "bbox": [ + 131, + 236, + 479, + 295 + ], + "type": "text", + "content": " refer to the uncontaminated inlier and outlier score distributions, " + }, + { + "bbox": [ + 131, + 236, + 479, + 295 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 131, + 236, + 479, + 295 + ], + "type": "text", + "content": " refers to the overlap. Such that we are concerned about multi-thresholding instead of the conventional way that explicitly finding a single threshold between " + }, + { + "bbox": [ + 131, + 236, + 479, + 295 + ], + "type": "inline_equation", + "content": "\\mathrm{F_{init}(\\mathbf{X}_{in})}" + }, + { + "bbox": [ + 131, + 236, + 479, + 295 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 236, + 479, + 295 + ], + "type": "inline_equation", + "content": "\\mathrm{F_{init}(\\mathbf{X}_{out})}" + }, + { + "bbox": [ + 131, + 236, + 479, + 295 + ], + "type": "text", + "content": ". Specifically, our objective is identifying the inlier threshold " + }, + { + "bbox": [ + 131, + 236, + 479, + 295 + ], + "type": "inline_equation", + "content": "\\phi_{\\mathrm{in}}" + }, + { + "bbox": [ + 131, + 236, + 479, + 295 + ], + "type": "text", + "content": " and outlier threshold " + }, + { + "bbox": [ + 131, + 236, + 479, + 295 + ], + "type": "inline_equation", + "content": "\\phi_{\\mathrm{out}}" + }, + { + "bbox": [ + 131, + 236, + 479, + 295 + ], + "type": "text", + "content": " that isolate " + }, + { + "bbox": [ + 131, + 236, + 479, + 295 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 131, + 236, + 479, + 295 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 236, + 479, + 295 + ], + "type": "inline_equation", + "content": "O" + }, + { + "bbox": [ + 131, + 236, + 479, + 295 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 131, + 236, + 479, + 295 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 131, + 236, + 479, + 295 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 296, + 481, + 344 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 296, + 481, + 344 + ], + "spans": [ + { + "bbox": [ + 131, + 296, + 481, + 344 + ], + "type": "text", + "content": "Identifying uncontaminated inliers. We identify " + }, + { + "bbox": [ + 131, + 296, + 481, + 344 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 131, + 296, + 481, + 344 + ], + "type": "text", + "content": " by understanding the initial outlier score function (Eq. 6), which computes the normalized " + }, + { + "bbox": [ + 131, + 296, + 481, + 344 + ], + "type": "inline_equation", + "content": "\\ell_2" + }, + { + "bbox": [ + 131, + 296, + 481, + 344 + ], + "type": "text", + "content": "-norm of each instance " + }, + { + "bbox": [ + 131, + 296, + 481, + 344 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_i \\in \\mathbf{X}" + }, + { + "bbox": [ + 131, + 296, + 481, + 344 + ], + "type": "text", + "content": " relative to the mean of " + }, + { + "bbox": [ + 131, + 296, + 481, + 344 + ], + "type": "inline_equation", + "content": "\\mathbf{X}" + }, + { + "bbox": [ + 131, + 296, + 481, + 344 + ], + "type": "text", + "content": ". Inspired by Shell Theory [32], " + }, + { + "bbox": [ + 131, + 296, + 481, + 344 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 131, + 296, + 481, + 344 + ], + "type": "text", + "content": " satisfies a Gaussian-like distribution, the classic statistical estimation of " + }, + { + "bbox": [ + 131, + 296, + 481, + 344 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 131, + 296, + 481, + 344 + ], + "type": "text", + "content": " is:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 200, + 352, + 480, + 365 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 200, + 352, + 480, + 365 + ], + "spans": [ + { + "bbox": [ + 200, + 352, + 480, + 365 + ], + "type": "interline_equation", + "content": "I = \\left\\{\\mathrm {F} _ {\\text {i n i t}} (\\mathbf {x} _ {i}) \\mid \\mathrm {F} _ {\\text {i n i t}} (\\mathbf {x} _ {i}) < \\operatorname {m e a n} (D) + k \\cdot \\operatorname {s t d} (D) \\right\\}, \\tag {8}", + "image_path": "0d93d7e64b2790e519c5495bab62ce22a4486988ec55f8800ff02f88a02f62fe.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 372, + 479, + 433 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 372, + 479, + 433 + ], + "spans": [ + { + "bbox": [ + 131, + 372, + 479, + 433 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 372, + 479, + 433 + ], + "type": "inline_equation", + "content": "\\mathrm{mean}(D)" + }, + { + "bbox": [ + 131, + 372, + 479, + 433 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 372, + 479, + 433 + ], + "type": "inline_equation", + "content": "\\mathrm{std}(D)" + }, + { + "bbox": [ + 131, + 372, + 479, + 433 + ], + "type": "text", + "content": " refer to the mean and standard deviation, " + }, + { + "bbox": [ + 131, + 372, + 479, + 433 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 131, + 372, + 479, + 433 + ], + "type": "text", + "content": " is the hyper-parameter of the \" " + }, + { + "bbox": [ + 131, + 372, + 479, + 433 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 131, + 372, + 479, + 433 + ], + "type": "text", + "content": "-sigma\" rule. In practice, optimizing " + }, + { + "bbox": [ + 131, + 372, + 479, + 433 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 131, + 372, + 479, + 433 + ], + "type": "text", + "content": " without supervision is challenging. Thus, we convert to a non-parametric way by utilizing an ascending sort projection " + }, + { + "bbox": [ + 131, + 372, + 479, + 433 + ], + "type": "inline_equation", + "content": "\\mathrm{sort}(\\cdot)" + }, + { + "bbox": [ + 131, + 372, + 479, + 433 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 131, + 372, + 479, + 433 + ], + "type": "inline_equation", + "content": "\\mathrm{F}_{\\mathrm{init}}(\\mathbf{X})" + }, + { + "bbox": [ + 131, + 372, + 479, + 433 + ], + "type": "text", + "content": ", which projects " + }, + { + "bbox": [ + 131, + 372, + 479, + 433 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 131, + 372, + 479, + 433 + ], + "type": "text", + "content": " into a 2-" + }, + { + "bbox": [ + 131, + 372, + 479, + 433 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 131, + 372, + 479, + 433 + ], + "type": "text", + "content": " space (" + }, + { + "bbox": [ + 131, + 372, + 479, + 433 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 131, + 372, + 479, + 433 + ], + "type": "text", + "content": "-axis: instance index, " + }, + { + "bbox": [ + 131, + 372, + 479, + 433 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 131, + 372, + 479, + 433 + ], + "type": "text", + "content": "-axis: sorted outlier score), the sorted outlier score is:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 196, + 441, + 480, + 455 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 196, + 441, + 480, + 455 + ], + "spans": [ + { + "bbox": [ + 196, + 441, + 480, + 455 + ], + "type": "interline_equation", + "content": "\\widehat {\\mathrm {F}} _ {\\text {i n i t}} (\\mathbf {X}) = \\operatorname {s o r t} \\left(\\widehat {\\mathrm {F}} _ {\\text {i n i t}} (\\mathbf {X})\\right), i. e., \\widehat {\\mathrm {F}} _ {\\text {i n i t}} [ i + 1 ] > \\widehat {\\mathrm {F}} _ {\\text {i n i t}} [ i ]. \\tag {9}", + "image_path": "487fb0894e4c33ee0100f0f994ecf2edf15d1406957df3e644774a3cd0afb44d.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 463, + 479, + 488 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 463, + 479, + 488 + ], + "spans": [ + { + "bbox": [ + 131, + 463, + 479, + 488 + ], + "type": "text", + "content": "Intuitively, " + }, + { + "bbox": [ + 131, + 463, + 479, + 488 + ], + "type": "inline_equation", + "content": "\\widehat{I} = \\mathrm{sort}(I)" + }, + { + "bbox": [ + 131, + 463, + 479, + 488 + ], + "type": "text", + "content": " can be fitted with a naive linear regressor " + }, + { + "bbox": [ + 131, + 463, + 479, + 488 + ], + "type": "inline_equation", + "content": "g(a) = \\boldsymbol{\\beta}^{\\top} \\cdot a = \\beta_{0} + \\beta_{1} \\cdot a" + }, + { + "bbox": [ + 131, + 463, + 479, + 488 + ], + "type": "text", + "content": ", as shown in Fig. 2 (b). So we identify the potential inliers as:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 143, + 496, + 480, + 529 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 496, + 480, + 529 + ], + "spans": [ + { + "bbox": [ + 143, + 496, + 480, + 529 + ], + "type": "interline_equation", + "content": "I = \\left\\{\\widehat {\\mathrm {F}} _ {\\text {i n i t}} (\\mathbf {x} _ {i}) | i < \\max _ {i} \\left\\{g \\left(a _ {i}\\right) > \\widehat {\\mathrm {F}} _ {\\text {i n i t}} (\\mathbf {x} _ {i}) \\right\\} \\right\\}, s. t. \\min _ {\\beta} \\left| \\left| a ^ {\\top} \\beta - \\widehat {\\mathrm {F}} _ {\\text {i n i t}} (\\mathbf {X}) \\right| \\right| _ {2}, \\tag {10}", + "image_path": "0c90ca80a490cc4c548846625e44c24b6b6798862e1e1934e2b1ed7139c31feb.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 529, + 479, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 529, + 479, + 552 + ], + "spans": [ + { + "bbox": [ + 130, + 529, + 479, + 552 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 529, + 479, + 552 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 130, + 529, + 479, + 552 + ], + "type": "text", + "content": " is the coefficients. Inevitably, this fitting process will be shifted with outliers, which is addressed together with the following outliers' identification." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 131, + 552, + 479, + 587 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 552, + 479, + 587 + ], + "spans": [ + { + "bbox": [ + 131, + 552, + 479, + 587 + ], + "type": "text", + "content": "Identifying adaptive outliers. In statistics, 3-sigma rule [42] declares variables " + }, + { + "bbox": [ + 131, + 552, + 479, + 587 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 131, + 552, + 479, + 587 + ], + "type": "text", + "content": " has a large probability within three standard deviations from the mean if " + }, + { + "bbox": [ + 131, + 552, + 479, + 587 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 131, + 552, + 479, + 587 + ], + "type": "text", + "content": " follows a Gaussian-like distribution. So we employ it with two aspects:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 588, + 432, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 588, + 432, + 601 + ], + "spans": [ + { + "bbox": [ + 132, + 588, + 432, + 601 + ], + "type": "text", + "content": "(i) iteratively filter outliers during the phase of inliers identification:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 140, + 608, + 480, + 632 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 140, + 608, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 140, + 608, + 480, + 632 + ], + "type": "interline_equation", + "content": "\\phi_ {\\text {o u t}} ^ {b} = \\operatorname {m e a n} \\left(I ^ {b}\\right) + 3 \\cdot \\operatorname {s t d} \\left(I ^ {b}\\right), I ^ {b + 1} = I ^ {b}. \\text {r e m o v e} \\left(\\left\\{\\mathrm {F} _ {\\text {i n i t}} \\left(\\mathbf {x} _ {i}\\right) \\mid \\mathrm {F} _ {\\text {i n i t}} \\left(\\mathbf {x} _ {i}\\right) > \\phi_ {\\text {o u t}} ^ {b} \\right\\}\\right), \\tag {11}", + "image_path": "8319f5ee75f724483d582470a7a001a43ea559f1f306618136095d9439927562.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 633, + 466, + 644 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 633, + 466, + 644 + ], + "spans": [ + { + "bbox": [ + 132, + 633, + 466, + 644 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 132, + 633, + 466, + 644 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 132, + 633, + 466, + 644 + ], + "type": "text", + "content": " refers to the " + }, + { + "bbox": [ + 132, + 633, + 466, + 644 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 132, + 633, + 466, + 644 + ], + "type": "text", + "content": "-th iteration. When converged, the inlier threshold is:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 272, + 653, + 480, + 666 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 272, + 653, + 480, + 666 + ], + "spans": [ + { + "bbox": [ + 272, + 653, + 480, + 666 + ], + "type": "interline_equation", + "content": "\\phi_ {\\text {i n}} = \\max \\left(I ^ {b ^ {*}}\\right), \\tag {12}", + "image_path": "dfefa87c24de3714d575c8f5f0adf68b88902497cdd65f584b29792971a75a6f.jpg" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 162, + 91, + 448, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 162, + 91, + 448, + 103 + ], + "spans": [ + { + "bbox": [ + 162, + 91, + 448, + 103 + ], + "type": "text", + "content": "Rethinking Unsupervised Outlier Detection via Multiple Thresholding" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 115, + 479, + 129 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 115, + 479, + 129 + ], + "spans": [ + { + "bbox": [ + 132, + 115, + 479, + 129 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 132, + 115, + 479, + 129 + ], + "type": "inline_equation", + "content": "I^{b^*}" + }, + { + "bbox": [ + 132, + 115, + 479, + 129 + ], + "type": "text", + "content": " refers to the converged inlier prediction and we can identify inliers as:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 249, + 140, + 481, + 153 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 249, + 140, + 481, + 153 + ], + "spans": [ + { + "bbox": [ + 249, + 140, + 481, + 153 + ], + "type": "interline_equation", + "content": "\\mathbf {X} _ {\\mathrm {i n}} ^ {\\prime} = \\left\\{\\mathbf {x} _ {i} \\mid \\mathrm {F} _ {\\text {i n i t}} \\left(\\mathbf {x} _ {i}\\right) \\leq \\phi_ {\\mathrm {i n}} \\right\\}. \\tag {13}", + "image_path": "b535d8705556b671d0351b21600d24dffd2452fd51a34a78c9c85eef7a921f9a.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 156, + 457, + 169 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 156, + 457, + 169 + ], + "spans": [ + { + "bbox": [ + 132, + 156, + 457, + 169 + ], + "type": "text", + "content": "(ii) During the above phase, we estimate the outlier threshold candidates:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 173, + 176, + 481, + 192 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 173, + 176, + 481, + 192 + ], + "spans": [ + { + "bbox": [ + 173, + 176, + 481, + 192 + ], + "type": "interline_equation", + "content": "\\phi_ {\\mathrm {o u t}} ^ {1} = \\operatorname {m e a n} \\left(I ^ {1}\\right) + 3 \\cdot \\operatorname {s t d} \\left(I ^ {1}\\right), \\phi_ {\\mathrm {o u t}} ^ {*} = \\operatorname {m e a n} \\left(I ^ {b ^ {*}}\\right) + 3 \\cdot \\operatorname {s t d} \\left(I ^ {b ^ {*}}\\right), \\tag {14}", + "image_path": "edb0f94ec13f81196995c99a423110d228c84075fc23ebcc0f320101d371bfb1.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 197, + 481, + 270 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 197, + 481, + 270 + ], + "spans": [ + { + "bbox": [ + 132, + 197, + 481, + 270 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 132, + 197, + 481, + 270 + ], + "type": "inline_equation", + "content": "\\phi_{\\mathrm{out}}^{1}" + }, + { + "bbox": [ + 132, + 197, + 481, + 270 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 132, + 197, + 481, + 270 + ], + "type": "inline_equation", + "content": "\\phi_{\\mathrm{out}}^{*}" + }, + { + "bbox": [ + 132, + 197, + 481, + 270 + ], + "type": "text", + "content": " refer to the outlier threshold candidates at the first and final iterations, respectively, i.e., " + }, + { + "bbox": [ + 132, + 197, + 481, + 270 + ], + "type": "inline_equation", + "content": "\\phi_{\\mathrm{out}}^{1} > \\phi_{\\mathrm{out}}^{*}" + }, + { + "bbox": [ + 132, + 197, + 481, + 270 + ], + "type": "text", + "content": ". To obtain a " + }, + { + "bbox": [ + 132, + 197, + 481, + 270 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 132, + 197, + 481, + 270 + ], + "type": "text", + "content": "-adaptive outlier threshold, we consider the impact and relationship of two mentioned normalization procedures: Shell Normalization S-norm(" + }, + { + "bbox": [ + 132, + 197, + 481, + 270 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 132, + 197, + 481, + 270 + ], + "type": "text", + "content": ") (Eq. 3) and Ergodic-set Normalization E-norm(" + }, + { + "bbox": [ + 132, + 197, + 481, + 270 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 132, + 197, + 481, + 270 + ], + "type": "text", + "content": ") (Eq. 5). Both them follow a similar normalization formula (denominator: " + }, + { + "bbox": [ + 132, + 197, + 481, + 270 + ], + "type": "inline_equation", + "content": "\\ell_2" + }, + { + "bbox": [ + 132, + 197, + 481, + 270 + ], + "type": "text", + "content": "-norm), the difference is their reference vectors " + }, + { + "bbox": [ + 132, + 197, + 481, + 270 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_{\\mathrm{S}}'" + }, + { + "bbox": [ + 132, + 197, + 481, + 270 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 132, + 197, + 481, + 270 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_{\\mathrm{E}}" + }, + { + "bbox": [ + 132, + 197, + 481, + 270 + ], + "type": "text", + "content": ", contrast as:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 207, + 279, + 481, + 317 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 207, + 279, + 481, + 317 + ], + "spans": [ + { + "bbox": [ + 207, + 279, + 481, + 317 + ], + "type": "interline_equation", + "content": "\\mathbf {v} _ {\\mathrm {S}} ^ {\\prime} = \\left[ \\frac {1}{n} \\sum_ {i = 1} ^ {n} \\mathbf {X} _ {\\text {o u t}} ^ {\\prime} [ i ] [ 1 ], \\dots , \\frac {1}{n} \\sum_ {i = 1} ^ {n} \\mathbf {X} _ {\\text {o u t}} ^ {\\prime} [ i ] [ d ] \\right], \\tag {15}", + "image_path": "e7f49c041a5a59765bcd41d54fc39090a155bf9335f35ff62f88fbf49b1b498b.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 206, + 312, + 409, + 345 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 206, + 312, + 409, + 345 + ], + "spans": [ + { + "bbox": [ + 206, + 312, + 409, + 345 + ], + "type": "interline_equation", + "content": "\\mathbf {v} _ {\\mathrm {E}} = \\left[ \\frac {1}{n \\cdot d} \\sum_ {i = 1} ^ {n} \\sum_ {j = 1} ^ {d} \\mathbf {x} _ {i, j}, \\dots , \\frac {1}{n \\cdot d} \\sum_ {i = 1} ^ {n} \\sum_ {j = 1} ^ {d} \\mathbf {x} _ {i, j} \\right].", + "image_path": "6b6332567893b9b06aa77678ae3519901f59e941e29fb63f451363c3c6190be0.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 348, + 481, + 397 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 348, + 481, + 397 + ], + "spans": [ + { + "bbox": [ + 132, + 348, + 481, + 397 + ], + "type": "text", + "content": "Obviously, Shell normalization is strongly subjective to the efficacy of outlier prediction " + }, + { + "bbox": [ + 132, + 348, + 481, + 397 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_{\\mathrm{out}}^{\\prime}" + }, + { + "bbox": [ + 132, + 348, + 481, + 397 + ], + "type": "text", + "content": " while Ergodic-set Normalization is independent with " + }, + { + "bbox": [ + 132, + 348, + 481, + 397 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 132, + 348, + 481, + 397 + ], + "type": "text", + "content": " that is a sub-optimal but stable operation. To compare these two normalization procedures, we first employ them to Eq. 6, and obtain two outlier score functions:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 188, + 406, + 481, + 426 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 188, + 406, + 481, + 426 + ], + "spans": [ + { + "bbox": [ + 188, + 406, + 481, + 426 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathrm {F} _ {\\mathrm {S}} (\\mathbf {X}) = \\left\\{\\operatorname {D i s t} \\left(\\mathrm {S} - \\operatorname {n o r m} \\left(\\mathbf {x} _ {i}, \\mathbf {v} _ {\\mathrm {S}} ^ {\\prime}\\right), \\operatorname {S} - \\operatorname {n o r m} \\left(\\mathbf {m} _ {\\mathbf {X}}, \\mathbf {v} _ {\\mathrm {S}} ^ {\\prime}\\right)\\right) \\right\\} _ {i = 1} ^ {n}, \\\\ \\mathrm {F} _ {\\mathrm {E}} (\\mathbf {X}) = \\left\\{\\operatorname {D i s t} \\left(\\mathrm {E} - \\operatorname {n o r m} \\left(\\mathbf {v} _ {\\mathrm {E}} - \\mathbf {v} _ {\\mathrm {I}}\\right), \\operatorname {E} - \\operatorname {n o r m} \\left(\\mathbf {m} _ {\\mathbf {E}} - \\mathbf {v} _ {\\mathrm {I}}\\right)\\right) \\right\\} ^ {n}. \\end{array} \\tag {16}", + "image_path": "cc8ec9af88db8587d90b062f4fe14cc1953ce3f330241e803d2caa2e23c0124a.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 187, + 421, + 426, + 434 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 187, + 421, + 426, + 434 + ], + "spans": [ + { + "bbox": [ + 187, + 421, + 426, + 434 + ], + "type": "interline_equation", + "content": "\\mathrm {F} _ {\\mathrm {E}} (\\mathbf {X}) = \\left\\{\\operatorname {D i s t} \\left(\\operatorname {E} - \\operatorname {n o r m} \\left(\\mathbf {x} _ {i}, \\mathbf {v} _ {\\mathrm {E}}\\right), \\operatorname {E} - \\operatorname {n o r m} \\left(\\mathbf {m} _ {\\mathbf {X}}, \\mathbf {v} _ {\\mathrm {E}}\\right)\\right) \\right\\} _ {i = 1} ^ {n},", + "image_path": "260c06d5ebec9579b58c93b537868429134686da7f3bebc48b17df7b39e9ee56.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 437, + 482, + 571 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 437, + 482, + 571 + ], + "spans": [ + { + "bbox": [ + 132, + 437, + 482, + 571 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 132, + 437, + 482, + 571 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_{\\mathrm{S}}^{\\prime}" + }, + { + "bbox": [ + 132, + 437, + 482, + 571 + ], + "type": "text", + "content": " comes from " + }, + { + "bbox": [ + 132, + 437, + 482, + 571 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_{\\mathrm{out}}^{\\prime}" + }, + { + "bbox": [ + 132, + 437, + 482, + 571 + ], + "type": "text", + "content": " obtained by MAD [6] (Shell Normalization's default thresholding method that is only effective on high- " + }, + { + "bbox": [ + 132, + 437, + 482, + 571 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 132, + 437, + 482, + 571 + ], + "type": "text", + "content": "). Subsequently, two outlier score functions " + }, + { + "bbox": [ + 132, + 437, + 482, + 571 + ], + "type": "inline_equation", + "content": "\\mathrm{F_S}(\\mathbf{X})" + }, + { + "bbox": [ + 132, + 437, + 482, + 571 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 132, + 437, + 482, + 571 + ], + "type": "inline_equation", + "content": "\\mathrm{F_E}(\\mathbf{X})" + }, + { + "bbox": [ + 132, + 437, + 482, + 571 + ], + "type": "text", + "content": " are arranged in ascending order, and we denote the ranked-index lists as " + }, + { + "bbox": [ + 132, + 437, + 482, + 571 + ], + "type": "inline_equation", + "content": "\\mathcal{R}_{\\mathrm{S}}" + }, + { + "bbox": [ + 132, + 437, + 482, + 571 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 132, + 437, + 482, + 571 + ], + "type": "inline_equation", + "content": "\\mathcal{R}_{\\mathrm{E}}" + }, + { + "bbox": [ + 132, + 437, + 482, + 571 + ], + "type": "text", + "content": ". The outlier ratio " + }, + { + "bbox": [ + 132, + 437, + 482, + 571 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 132, + 437, + 482, + 571 + ], + "type": "text", + "content": " can be approximately estimated with the similarity between " + }, + { + "bbox": [ + 132, + 437, + 482, + 571 + ], + "type": "inline_equation", + "content": "\\mathcal{R}_{\\mathrm{S}}" + }, + { + "bbox": [ + 132, + 437, + 482, + 571 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 132, + 437, + 482, + 571 + ], + "type": "inline_equation", + "content": "\\mathcal{R}_{\\mathrm{E}}" + }, + { + "bbox": [ + 132, + 437, + 482, + 571 + ], + "type": "text", + "content": " since " + }, + { + "bbox": [ + 132, + 437, + 482, + 571 + ], + "type": "inline_equation", + "content": "\\mathrm{F_S}(\\mathbf{X})" + }, + { + "bbox": [ + 132, + 437, + 482, + 571 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 132, + 437, + 482, + 571 + ], + "type": "inline_equation", + "content": "\\mathrm{F_E}(\\mathbf{X})" + }, + { + "bbox": [ + 132, + 437, + 482, + 571 + ], + "type": "text", + "content": " share a consistent structure (distance-to-the-mean), making " + }, + { + "bbox": [ + 132, + 437, + 482, + 571 + ], + "type": "inline_equation", + "content": "\\mathcal{R}_{\\mathrm{S}}" + }, + { + "bbox": [ + 132, + 437, + 482, + 571 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 132, + 437, + 482, + 571 + ], + "type": "inline_equation", + "content": "\\mathcal{R}_{\\mathrm{E}}" + }, + { + "bbox": [ + 132, + 437, + 482, + 571 + ], + "type": "text", + "content": " comparable. Besides, " + }, + { + "bbox": [ + 132, + 437, + 482, + 571 + ], + "type": "inline_equation", + "content": "\\mathrm{F_E}(\\mathbf{X})" + }, + { + "bbox": [ + 132, + 437, + 482, + 571 + ], + "type": "text", + "content": " serves as a reliably effective baseline. When " + }, + { + "bbox": [ + 132, + 437, + 482, + 571 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 132, + 437, + 482, + 571 + ], + "type": "text", + "content": " is low, S-norm " + }, + { + "bbox": [ + 132, + 437, + 482, + 571 + ], + "type": "inline_equation", + "content": "(\\mathbf{X},\\mathbf{v}_{\\mathrm{S}}^{\\prime})" + }, + { + "bbox": [ + 132, + 437, + 482, + 571 + ], + "type": "text", + "content": " tends to under-perform, lead the similarity between " + }, + { + "bbox": [ + 132, + 437, + 482, + 571 + ], + "type": "inline_equation", + "content": "\\mathcal{R}_{\\mathrm{S}}" + }, + { + "bbox": [ + 132, + 437, + 482, + 571 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 132, + 437, + 482, + 571 + ], + "type": "inline_equation", + "content": "\\mathcal{R}_{\\mathrm{E}}" + }, + { + "bbox": [ + 132, + 437, + 482, + 571 + ], + "type": "text", + "content": " becomes low. In contrast, with a high " + }, + { + "bbox": [ + 132, + 437, + 482, + 571 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 132, + 437, + 482, + 571 + ], + "type": "text", + "content": ", S-norm " + }, + { + "bbox": [ + 132, + 437, + 482, + 571 + ], + "type": "inline_equation", + "content": "(\\mathbf{X},\\mathbf{v}_{\\mathrm{S}}^{\\prime})" + }, + { + "bbox": [ + 132, + 437, + 482, + 571 + ], + "type": "text", + "content": " merely refines the ranking of a small number of outliers, indicative of high similarity. The similarity " + }, + { + "bbox": [ + 132, + 437, + 482, + 571 + ], + "type": "inline_equation", + "content": "\\rho \\in [-1,1]" + }, + { + "bbox": [ + 132, + 437, + 482, + 571 + ], + "type": "text", + "content": " is computed with the Pearson correlation coefficient:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 256, + 577, + 481, + 604 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 256, + 577, + 481, + 604 + ], + "spans": [ + { + "bbox": [ + 256, + 577, + 481, + 604 + ], + "type": "interline_equation", + "content": "\\rho = \\frac {\\operatorname {c o v} \\left(\\mathcal {R} _ {\\mathrm {S}} , \\mathcal {R} _ {\\mathrm {E}}\\right)}{\\operatorname {s t d} \\left(\\mathcal {R} _ {\\mathrm {S}}\\right) \\cdot \\operatorname {s t d} \\left(\\mathcal {R} _ {\\mathrm {E}}\\right)}, \\tag {17}", + "image_path": "57240db46e980d03f7b0c0585f086ba006d1d91f300a54e2864e137784e7c5d0.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 609, + 481, + 645 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 609, + 481, + 645 + ], + "spans": [ + { + "bbox": [ + 132, + 609, + 481, + 645 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 132, + 609, + 481, + 645 + ], + "type": "inline_equation", + "content": "\\operatorname{cov}(\\mathcal{R}_{\\mathrm{S}},\\mathcal{R}_{\\mathrm{E}})" + }, + { + "bbox": [ + 132, + 609, + 481, + 645 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 132, + 609, + 481, + 645 + ], + "type": "inline_equation", + "content": "\\mathrm{std}(\\mathcal{R}_{\\mathrm{S}}) / \\mathrm{std}(\\mathcal{R}_{\\mathrm{E}})" + }, + { + "bbox": [ + 132, + 609, + 481, + 645 + ], + "type": "text", + "content": " are the covariance and standard deviation of the rank variables. If the similarity is relatively low (" + }, + { + "bbox": [ + 132, + 609, + 481, + 645 + ], + "type": "inline_equation", + "content": "< 0.1" + }, + { + "bbox": [ + 132, + 609, + 481, + 645 + ], + "type": "text", + "content": "), we compute the 3-sigma of the entire outlier score distribution " + }, + { + "bbox": [ + 132, + 609, + 481, + 645 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 132, + 609, + 481, + 645 + ], + "type": "text", + "content": " as the outlier threshold:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 244, + 653, + 481, + 667 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 244, + 653, + 481, + 667 + ], + "spans": [ + { + "bbox": [ + 244, + 653, + 481, + 667 + ], + "type": "interline_equation", + "content": "\\phi_ {\\text {o u t}} ^ {0} = \\operatorname {m e a n} (D) + 3 \\cdot \\operatorname {s t d} (D). \\tag {18}", + "image_path": "4c0cc3417472e0e881f15a62aa9af6fdd4c75425e0c6f2010351a4ebdcebba49.jpg" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "type": "text", + "content": "Z. Liu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 131, + 116, + 479, + 165 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 116, + 479, + 165 + ], + "spans": [ + { + "bbox": [ + 131, + 116, + 479, + 165 + ], + "type": "text", + "content": "It is a conservative estimator for outliers since the above linear fitting might suffer from some challenges at low- " + }, + { + "bbox": [ + 131, + 116, + 479, + 165 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 131, + 116, + 479, + 165 + ], + "type": "text", + "content": " since there is no explicit outlier score distribution. If the similarity is large (" + }, + { + "bbox": [ + 131, + 116, + 479, + 165 + ], + "type": "inline_equation", + "content": ">0.3" + }, + { + "bbox": [ + 131, + 116, + 479, + 165 + ], + "type": "text", + "content": "), we choose the relatively smooth outlier threshold " + }, + { + "bbox": [ + 131, + 116, + 479, + 165 + ], + "type": "inline_equation", + "content": "\\phi_{\\mathrm{out}}^{*}" + }, + { + "bbox": [ + 131, + 116, + 479, + 165 + ], + "type": "text", + "content": ", while " + }, + { + "bbox": [ + 131, + 116, + 479, + 165 + ], + "type": "inline_equation", + "content": "\\phi_{\\mathrm{out}}^{1}" + }, + { + "bbox": [ + 131, + 116, + 479, + 165 + ], + "type": "text", + "content": " is used for middle- " + }, + { + "bbox": [ + 131, + 116, + 479, + 165 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 131, + 116, + 479, + 165 + ], + "type": "text", + "content": " cases. The outlier threshold is described as:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 237, + 174, + 479, + 210 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 237, + 174, + 479, + 210 + ], + "spans": [ + { + "bbox": [ + 237, + 174, + 479, + 210 + ], + "type": "interline_equation", + "content": "\\phi_ {\\text {o u t}} = \\left\\{ \\begin{array}{l} \\phi_ {\\text {o u t}} ^ {*}, \\text {i f} \\rho > 0. 3 \\\\ \\phi_ {\\text {o u t}} ^ {1}, \\text {i f} 0. 1 \\leq \\rho \\leq 0. 3 \\\\ \\phi_ {\\text {o u t}} ^ {0}, \\text {o t h e r w i s e} \\end{array} \\right. \\tag {19}", + "image_path": "451133826d0b253988b9b52b2db3b479a7bbc8db68284b54664cf064f309922b.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 213, + 455, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 213, + 455, + 224 + ], + "spans": [ + { + "bbox": [ + 132, + 213, + 455, + 224 + ], + "type": "text", + "content": "Therefore, we can subsequently identify outlier ratio invariant outliers as:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 244, + 230, + 479, + 243 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 244, + 230, + 479, + 243 + ], + "spans": [ + { + "bbox": [ + 244, + 230, + 479, + 243 + ], + "type": "interline_equation", + "content": "\\mathbf {X} _ {\\text {o u t}} ^ {\\prime} = \\left\\{\\mathbf {x} _ {i} \\mid \\mathrm {F} _ {\\text {i n i t}} \\left(\\mathbf {x} _ {i}\\right) > \\phi_ {\\text {o u t}} \\right\\}. \\tag {20}", + "image_path": "870b643f2df05cbd76bd0350e30b71844f33fbd3ab564871f984a2366c20593e.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 256, + 240, + 268 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 256, + 240, + 268 + ], + "spans": [ + { + "bbox": [ + 132, + 256, + 240, + 268 + ], + "type": "text", + "content": "4.3 Implementation" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 273, + 479, + 285 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 273, + 479, + 285 + ], + "spans": [ + { + "bbox": [ + 132, + 273, + 479, + 285 + ], + "type": "text", + "content": "Firstly, we can directly employ Multi-T with the distance-to-the-mean paradigm:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 176, + 294, + 438, + 308 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 176, + 294, + 438, + 308 + ], + "spans": [ + { + "bbox": [ + 176, + 294, + 438, + 308 + ], + "type": "interline_equation", + "content": "\\mathrm {F} _ {\\mathrm {M u l t i - T}} (\\mathbf {X}) = \\{\\mathrm {D i s t} (\\mathrm {S - n o r m} (\\mathbf {x} _ {i}, \\mathbf {v} _ {\\mathrm {S}} ^ {\\prime}), \\mathrm {S - n o r m} (\\mathbf {m} _ {\\mathbf {X} _ {\\mathrm {i n}} ^ {\\prime}}, \\mathbf {v} _ {\\mathrm {S}} ^ {\\prime})) \\} _ {i = 1} ^ {n},", + "image_path": "f58f479ee35af8ed1ec048116c6831e6d175ca19922e1d7352c9dfe391ce284f.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 216, + 311, + 479, + 340 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 216, + 311, + 479, + 340 + ], + "spans": [ + { + "bbox": [ + 216, + 311, + 479, + 340 + ], + "type": "interline_equation", + "content": "\\mathbf {v} _ {\\mathrm {S}} ^ {\\prime} = \\left[ \\frac {1}{n} \\sum_ {i = 1} ^ {n} \\mathbf {X} _ {\\text {o u t}} ^ {\\prime} [ i ] [ 1 ], \\dots , \\frac {1}{n} \\sum_ {i = 1} ^ {n} \\mathbf {X} _ {\\text {o u t}} ^ {\\prime} [ i ] [ d ] \\right], \\tag {21}", + "image_path": "922e68f6818b6ecef03e0eb04517a74818f453a9b795ddd1b0718f0339ba269b.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 203, + 344, + 400, + 373 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 203, + 344, + 400, + 373 + ], + "spans": [ + { + "bbox": [ + 203, + 344, + 400, + 373 + ], + "type": "interline_equation", + "content": "\\mathbf {m} _ {\\mathbf {X} _ {\\mathrm {i n}} ^ {\\prime}} = \\left[ \\frac {1}{n} \\sum_ {i = 1} ^ {n} \\mathbf {X} _ {\\mathrm {i n}} ^ {\\prime} [ i ] [ 1 ], \\dots , \\frac {1}{n} \\sum_ {i = 1} ^ {n} \\mathbf {X} _ {\\mathrm {i n}} ^ {\\prime} [ i ] [ d ] \\right].", + "image_path": "781427223f3cf7c77cd384e0b3de1f914227449a1b87dcd304f7223d117cd2d8.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 376, + 469, + 388 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 376, + 469, + 388 + ], + "spans": [ + { + "bbox": [ + 132, + 376, + 469, + 388 + ], + "type": "text", + "content": "Secondly, Multi-T can be integrated with UOD methods " + }, + { + "bbox": [ + 132, + 376, + 469, + 388 + ], + "type": "inline_equation", + "content": "\\mathrm{M}(\\cdot)" + }, + { + "bbox": [ + 132, + 376, + 469, + 388 + ], + "type": "text", + "content": ", illustrated as:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 187, + 397, + 480, + 425 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 187, + 397, + 480, + 425 + ], + "spans": [ + { + "bbox": [ + 187, + 397, + 480, + 425 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathrm {F} _ {\\mathrm {M} + \\text {M u l t i - T}} (\\mathbf {X}) = \\mathrm {M}. \\operatorname {f i t} \\left(\\left\\{\\mathrm {S} - \\operatorname {n o r m} \\left(\\mathbf {x} _ {i}, \\mathbf {v} _ {\\mathrm {S}} ^ {\\prime}\\right) \\mid \\mathbf {x} _ {i} \\in \\mathbf {X} _ {i n} ^ {\\prime} \\right\\}\\right) \\\\ . \\operatorname {p r e d i c t} \\left(\\left\\{\\mathrm {S} - \\operatorname {n o r m} \\left(\\mathbf {x} _ {i}, \\mathbf {v} _ {\\mathrm {S}} ^ {\\prime}\\right) \\mid \\mathbf {x} _ {i} \\in \\mathbf {X} \\right\\}\\right). \\\\ \\end{array}", + "image_path": "348a0aa0853342f68f3bd095eea0abdaf2086a4b6a46ba1c8cbc0c526776bde5.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 438, + 230, + 452 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 438, + 230, + 452 + ], + "spans": [ + { + "bbox": [ + 132, + 438, + 230, + 452 + ], + "type": "text", + "content": "5 Experiments" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 460, + 481, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 460, + 481, + 495 + ], + "spans": [ + { + "bbox": [ + 130, + 460, + 481, + 495 + ], + "type": "text", + "content": "We cover our experiments for (i) comparing with other threshold learners on dataset splitting (Sec. 5.1); (ii) measuring the performance improvement of adopting the Multi-T to previously proposed outlier scoring methods (Sec. 5.2)." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 496, + 479, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 496, + 479, + 555 + ], + "spans": [ + { + "bbox": [ + 130, + 496, + 479, + 555 + ], + "type": "text", + "content": "Benchmarks and Feature Extraction. Based on the existing research [32,34], we use the raw pixel representation for grayscale datasets, including MNIST [29] and Fashion-MNIST [59]. For RGB datasets such as STL-10 [15], Internet [32], CIFAR-10 [27], MIT-Places-Small [61], we adopt two deep feature extractors: ImageNet pretrained [20] ResNet-50 [21] (ResNet)2, and CLIP [45]." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 555, + 481, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 555, + 481, + 628 + ], + "spans": [ + { + "bbox": [ + 130, + 555, + 481, + 628 + ], + "type": "text", + "content": "Target Dataset. In our experiments, each class within a benchmark dataset is alternatively regarded as inliers, with instances from all other classes considered outliers. The target dataset includes all inliers along with a randomly selected subset of outliers. Results for each dataset are averaged across all classes. Additionally, for every round, we further average the results over a wide range of outlier ratios: " + }, + { + "bbox": [ + 130, + 555, + 481, + 628 + ], + "type": "inline_equation", + "content": "[0.05, 0.1, 0.2, 0.3, 0.4]^3" + }, + { + "bbox": [ + 130, + 555, + 481, + 628 + ], + "type": "text", + "content": ", to assess the method's robustness/stability." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 162, + 91, + 448, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 162, + 91, + 448, + 103 + ], + "spans": [ + { + "bbox": [ + 162, + 91, + 448, + 103 + ], + "type": "text", + "content": "Rethinking Unsupervised Outlier Detection via Multiple Thresholding" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 133, + 642, + 416, + 653 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 642, + 416, + 653 + ], + "spans": [ + { + "bbox": [ + 133, + 642, + 416, + 653 + ], + "type": "text", + "content": "2 Unless otherwise stated, ResNet-50 is the default feature extractor." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 134, + 654, + 479, + 666 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 654, + 479, + 666 + ], + "spans": [ + { + "bbox": [ + 134, + 654, + 479, + 666 + ], + "type": "text", + "content": "3 Unless otherwise stated, the experimental results are averaged across " + }, + { + "bbox": [ + 134, + 654, + 479, + 666 + ], + "type": "inline_equation", + "content": "[0.05,\\dots ,0.4]" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 132, + 113, + 482, + 213 + ], + "blocks": [ + { + "bbox": [ + 132, + 113, + 482, + 213 + ], + "lines": [ + { + "bbox": [ + 132, + 113, + 482, + 213 + ], + "spans": [ + { + "bbox": [ + 132, + 113, + 482, + 213 + ], + "type": "image", + "image_path": "0ea2d465b66a5324d1e7a34f36b80e91c9d0a7e96f996495888d2e720529f2ed.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 132, + 214, + 482, + 318 + ], + "blocks": [ + { + "bbox": [ + 132, + 214, + 482, + 318 + ], + "lines": [ + { + "bbox": [ + 132, + 214, + 482, + 318 + ], + "spans": [ + { + "bbox": [ + 132, + 214, + 482, + 318 + ], + "type": "image", + "image_path": "ba6b4d142c17f8360942e86c2b65bf5a75fe039c58dc6e41df6d9f4276f2a032.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 323, + 482, + 346 + ], + "lines": [ + { + "bbox": [ + 130, + 323, + 482, + 346 + ], + "spans": [ + { + "bbox": [ + 130, + 323, + 482, + 346 + ], + "type": "text", + "content": "Fig.3: Qualitative results compared with SOTA threshold learners. The experiments are conducted on STL-10 (Inlier class: Monkey)." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 368, + 261, + 380 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 368, + 261, + 380 + ], + "spans": [ + { + "bbox": [ + 132, + 368, + 261, + 380 + ], + "type": "text", + "content": "5.1 Threshold Learning" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 387, + 482, + 459 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 387, + 482, + 459 + ], + "spans": [ + { + "bbox": [ + 130, + 387, + 482, + 459 + ], + "type": "text", + "content": "Competing Methods. We compare our proposed Multi-T with those effective thresholding (TL) methods involving Kernel-based: AUCP [48], FGD [43]; Curve-based: EB [17]; Normality-based: DSN [5], CHAU [10]; Filtering-based: FILTER [19], HIST [55]; Statistical-based: MTT [49], BOOT [37], QMCD [24], CLF [7], IQR [8], KARCH [1], MCST [16], GESD [4], REGR [3], MAD [6], CLUST [11], CPD [12] and Transformation-based: MOLL [25], YJ [46]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 459, + 481, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 459, + 481, + 483 + ], + "spans": [ + { + "bbox": [ + 130, + 459, + 481, + 483 + ], + "type": "text", + "content": "Evaluation Metric. The performance of general thresholding is measured with " + }, + { + "bbox": [ + 130, + 459, + 481, + 483 + ], + "type": "inline_equation", + "content": "\\mathrm{F}_{\\beta}" + }, + { + "bbox": [ + 130, + 459, + 481, + 483 + ], + "type": "text", + "content": "-score, defined as follows:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 216, + 490, + 481, + 515 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 216, + 490, + 481, + 515 + ], + "spans": [ + { + "bbox": [ + 216, + 490, + 481, + 515 + ], + "type": "interline_equation", + "content": "\\mathrm {F} _ {\\beta} = \\left(1 + \\beta^ {2}\\right) \\cdot \\frac {\\text {p r e c i s i o n} \\cdot \\text {r e c a l l}}{\\left(\\beta^ {2} \\cdot \\text {p r e c i s i o n}\\right) + \\text {r e c a l l}}, \\tag {23}", + "image_path": "905e25c1b2f7671e343b77b14b5e79829cd7f8143af445ad887df1ab7c7b5614.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 521, + 482, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 482, + 582 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 482, + 582 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 521, + 482, + 582 + ], + "type": "inline_equation", + "content": "\\beta < 1" + }, + { + "bbox": [ + 130, + 521, + 482, + 582 + ], + "type": "text", + "content": " tilts towards precision while " + }, + { + "bbox": [ + 130, + 521, + 482, + 582 + ], + "type": "inline_equation", + "content": "\\beta > 1" + }, + { + "bbox": [ + 130, + 521, + 482, + 582 + ], + "type": "text", + "content": " prioritizes recall. In this work, we utilize " + }, + { + "bbox": [ + 130, + 521, + 482, + 582 + ], + "type": "inline_equation", + "content": "\\mathrm{F}_{0.1}" + }, + { + "bbox": [ + 130, + 521, + 482, + 582 + ], + "type": "text", + "content": "-score " + }, + { + "bbox": [ + 130, + 521, + 482, + 582 + ], + "type": "inline_equation", + "content": "(\\mathrm{F}_{0.1})" + }, + { + "bbox": [ + 130, + 521, + 482, + 582 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 521, + 482, + 582 + ], + "type": "inline_equation", + "content": "\\mathrm{F}_{10}" + }, + { + "bbox": [ + 130, + 521, + 482, + 582 + ], + "type": "text", + "content": "-score " + }, + { + "bbox": [ + 130, + 521, + 482, + 582 + ], + "type": "inline_equation", + "content": "(\\mathrm{F}_{10})" + }, + { + "bbox": [ + 130, + 521, + 482, + 582 + ], + "type": "text", + "content": " that measure the accuracy of predicted outliers and inliers. Besides, " + }, + { + "bbox": [ + 130, + 521, + 482, + 582 + ], + "type": "inline_equation", + "content": "\\mathrm{F}_{1}" + }, + { + "bbox": [ + 130, + 521, + 482, + 582 + ], + "type": "text", + "content": "-score " + }, + { + "bbox": [ + 130, + 521, + 482, + 582 + ], + "type": "inline_equation", + "content": "(\\beta = 1)" + }, + { + "bbox": [ + 130, + 521, + 482, + 582 + ], + "type": "text", + "content": ", a harmonic mean of the precision and recall, is not practical for this task since the compete separability between inliers' and outliers' score distributions cannot be guaranteed." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "content": "Main Results. Across a series of competing threshold learners, while one method may achieve a commendable " + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathrm{F}_{0.1}" + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "content": "-score of 0.976, the corresponding " + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathrm{F}_{10}" + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "content": "-score is mere 0.452, and vice versa (" + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathrm{F}_{10}" + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "content": "-score: 0.979; " + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathrm{F}_{0.1}" + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "content": "-score: 0.447). By contrast, our solution maintains its efficacy across various evaluation criteria (" + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathrm{F}_{0.1}" + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "content": "-score: 0.917, " + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathrm{F}_{10}" + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "content": "-score: 0.980). It exhibits remarkable stability across a diverse spectrum of outlier ratios, as illustrated in Fig. 3. To further investigate the benefits of our technique, Tab. 1 categorizes two groups, each representing a different" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 101 + ], + "type": "text", + "content": "Z. Liu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 148, + 168, + 460, + 288 + ], + "blocks": [ + { + "bbox": [ + 130, + 114, + 482, + 159 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 482, + 159 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 482, + 159 + ], + "type": "text", + "content": "Table 1: Average " + }, + { + "bbox": [ + 130, + 114, + 482, + 159 + ], + "type": "inline_equation", + "content": "\\mathrm{F}_{\\beta}" + }, + { + "bbox": [ + 130, + 114, + 482, + 159 + ], + "type": "text", + "content": "-score of thresholding methods on STL-10. \"+\"GT Norm\" refers to the ideal initial outlier score function. \"Highest " + }, + { + "bbox": [ + 130, + 114, + 482, + 159 + ], + "type": "inline_equation", + "content": "\\mathrm{F}_{\\beta}" + }, + { + "bbox": [ + 130, + 114, + 482, + 159 + ], + "type": "text", + "content": "\" (" + }, + { + "bbox": [ + 130, + 114, + 482, + 159 + ], + "type": "inline_equation", + "content": "\\beta: 0.1, 10" + }, + { + "bbox": [ + 130, + 114, + 482, + 159 + ], + "type": "text", + "content": ") refers to the method that garners the highest " + }, + { + "bbox": [ + 130, + 114, + 482, + 159 + ], + "type": "inline_equation", + "content": "\\mathrm{F}_{\\beta}" + }, + { + "bbox": [ + 130, + 114, + 482, + 159 + ], + "type": "text", + "content": "-score across all baseline models. The better average result is highlighted in bold." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 148, + 168, + 460, + 288 + ], + "lines": [ + { + "bbox": [ + 148, + 168, + 460, + 288 + ], + "spans": [ + { + "bbox": [ + 148, + 168, + 460, + 288 + ], + "type": "table", + "html": "
Outlier Score Fun.Method (TL)ResNet-50CLIP
F0.1F10Avg.F0.1F10Avg.
LVAD-S [34]Highest F0.10.9110.4540.6820.9360.4010.669
Highest F100.3820.9670.6740.3570.9540.655
Multi-T(Ours)0.8400.8690.8550.8600.8660.863
+GT Norm [32]Highest F0.10.9760.4520.7140.9770.4480.713
Highest F100.4470.9790.7130.6910.9860.839
Multi-T(Ours)0.9170.9800.9490.8570.9840.920
", + "image_path": "fbe891ae5a6a4c1bacb10ab1a83a9c2e6423ad602b48d2658c85819d5638ce4b.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 320, + 479, + 343 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 320, + 479, + 343 + ], + "spans": [ + { + "bbox": [ + 130, + 320, + 479, + 343 + ], + "type": "text", + "content": "outlier score function. Such categorization is pivotal as threshold methods ought to accommodate various foundational bases." + } + ] + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 135, + 406, + 476, + 635 + ], + "blocks": [ + { + "bbox": [ + 130, + 373, + 482, + 397 + ], + "lines": [ + { + "bbox": [ + 130, + 373, + 482, + 397 + ], + "spans": [ + { + "bbox": [ + 130, + 373, + 482, + 397 + ], + "type": "text", + "content": "Table 2: Average AUC results compared with SOTA outlier scoring methods (outlier detectors). Blue and Orange indicates the best and second-best results, respectively." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 135, + 406, + 476, + 635 + ], + "lines": [ + { + "bbox": [ + 135, + 406, + 476, + 635 + ], + "spans": [ + { + "bbox": [ + 135, + 406, + 476, + 635 + ], + "type": "table", + "html": "
Method (OS)STL-10CIFAR-10CIFAR-100MIT-PlacesMNIST
ResNetCLIPResNetCLIPResNetCLIPResNetCLIPPixel
IF [35]0.8360.9430.7800.8910.7900.8660.6870.8680.776
LOF [11]0.6280.6260.6730.6210.8390.8490.5560.5200.754
RSRAE [28]0.9620.9380.8620.8790.9140.8850.8740.8760.793
ECOD [31]0.9070.9810.8730.9350.8730.9180.7770.9430.734
LUNAR [18]0.7760.8210.7670.7740.8380.8990.6430.8710.797
Shell-Re. [32]0.8620.8380.8600.8130.8350.8130.8260.9140.776
LVAD [34]0.9540.9680.8600.9170.9210.9170.8440.9190.867
Multi-T0.9680.9890.8950.9570.9380.9560.8670.9740.897
DeepSVDD [50]0.6220.5970.5600.5090.5630.5810.5830.5490.513
+Multi-T0.9250.9210.7690.8190.8350.8260.7550.8320.732
Improve.48.7%54.3%37.3%60.9%48.3%42.2%29.5%51.6%37.9%
OCSVM [51]0.9270.9210.8260.8500.8790.8500.8260.8710.831
+Multi-T0.9570.9650.8590.9160.9160.8990.8460.9240.863
Improve.3.24%4.78%4.00%7.76%4.21%5.76%2.42%6.08%3.85%
", + "image_path": "c4e8b38f24b3da10bfa2d939081c0f8c16ef44437a4aa2342032a0685b5a04e1.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 162, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 162, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 162, + 91, + 447, + 102 + ], + "type": "text", + "content": "Rethinking Unsupervised Outlier Detection via Multiple Thresholding" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 115, + 238, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 115, + 238, + 129 + ], + "spans": [ + { + "bbox": [ + 132, + 115, + 238, + 129 + ], + "type": "text", + "content": "5.2 Outlier Scoring" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 137, + 479, + 208 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 137, + 479, + 208 + ], + "spans": [ + { + "bbox": [ + 130, + 137, + 479, + 208 + ], + "type": "text", + "content": "Competing Methods. We compare two themes of baseline outlier scoring (OS) methods: (i) statistical-based methods: IF [35], OCSVM [51], DeepSVDD [51], ECOD [31], LUNAR [28], RSRAE [28], Shell-Re. [32], LVAD [34]. (ii) deep-learning-based models: GOAD [9], ICL [53], REPEN [39], NeuTraL [44], SLAD [60]. For a fair comparison, we apply Ergodic-set normalization [34] if it improves the performance of the baseline algorithms, such as IF [35], OCSVM [51]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 210, + 479, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 210, + 479, + 376 + ], + "spans": [ + { + "bbox": [ + 130, + 210, + 479, + 376 + ], + "type": "text", + "content": "Evaluation Metric. The performance of outlier scoring (ranking accuracy) is primarily assessed using the Area Under the Receiver Operating Characteristic curve (AUC). This metric provides a thorough assessment of ranking accuracy. Main Results. Tab. 2 shows that Multi-T can be seamlessly integrated with two classic outlier scoring methods: DeepSVDD [51], and OCSVM [51] while exhibiting significant improvements in both efficacy and stability across a diverse range of outlier ratios and various benchmarks. In most of our experiments, Multi-T itself achieves SOTA results. Notably, even in the case of non-aligned and low-resolution datasets like CIFAR-10 [27] and CIFAR-100 [27], which are known to pose challenges [41]. It surpasses the current SOTA AUC scores by margins of " + }, + { + "bbox": [ + 130, + 210, + 479, + 376 + ], + "type": "inline_equation", + "content": "7.76\\%" + }, + { + "bbox": [ + 130, + 210, + 479, + 376 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 210, + 479, + 376 + ], + "type": "inline_equation", + "content": "5.76\\%" + }, + { + "bbox": [ + 130, + 210, + 479, + 376 + ], + "type": "text", + "content": " for CIFAR-10 and CIFAR-100, respectively. Additionally, we observe a logical and significant enhancement in performance with improved feature representation, e.g., for the MIT-Places [61] dataset, the AUC improves from 0.867 using ResNet-50 [21] to 0.974 with CLIP [45]." + } + ] + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 147, + 432, + 463, + 582 + ], + "blocks": [ + { + "bbox": [ + 130, + 399, + 479, + 422 + ], + "lines": [ + { + "bbox": [ + 130, + 399, + 479, + 422 + ], + "spans": [ + { + "bbox": [ + 130, + 399, + 479, + 422 + ], + "type": "text", + "content": "Table 3: Average AUC results, for more outlier scoring methods with " + }, + { + "bbox": [ + 130, + 399, + 479, + 422 + ], + "type": "inline_equation", + "content": "(\\mathrm{w} / )" + }, + { + "bbox": [ + 130, + 399, + 479, + 422 + ], + "type": "text", + "content": " or without(w/o) our Multi-T module, conducted on STL-10." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 147, + 432, + 463, + 582 + ], + "lines": [ + { + "bbox": [ + 147, + 432, + 463, + 582 + ], + "spans": [ + { + "bbox": [ + 147, + 432, + 463, + 582 + ], + "type": "table", + "html": "
FeatureIF [35]ECOD [11]ABOD [26]PCA [54]GMM [2]
w/ow/w/ow/w/ow/w/ow/w/ow/
ResNet0.8360.8990.9070.9190.6650.8830.8650.9450.8590.952
CLIP0.9430.9830.9810.9840.7150.9090.9840.9940.8920.962
Avg.0.8900.9410.9440.9510.6900.8960.9250.9700.8760.957
FeatureGOAD [9]ICL [53]REPEN [39]NeuTraL [44]SLAD [60]
w/ow/w/ow/w/ow/w/ow/w/ow/
ResNet0.9520.9620.9340.9570.8770.8890.8540.9500.9410.962
CLIP0.9610.9890.9510.9820.8790.9300.8510.9710.9450.985
Avg.0.9570.9750.9430.9700.8780.9090.8530.9610.9430.973
", + "image_path": "41634fbb02a0421af99d58e7121887fd0f609778178bc237078d3bfa187d74c2.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 605, + 480, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 480, + 665 + ], + "type": "text", + "content": "In Tab. 3, we present the results of Multi-T integrated with more methods involving both statistical and deep models, which shows its broad applications. Moreover, our solution excels not just in detection accuracy but also in running speed since Multi-T compresses the size of fitting data, as illustrated in Tab. 4, 5. In Fig. 4, we compare with the most related work Shell-Re. [32], which" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 216, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 216, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 216, + 100 + ], + "type": "text", + "content": "Z. Liu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 164 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 164 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 164 + ], + "type": "text", + "content": "has a built-in Robust-Least-Square (RLS) thresholding procedure with MAD [6]. Apparently, our method is capable of estimating multiple thresholds that facilitate a clear demarcation of diverse scenarios (outlier ratio, dataset domain, and feature representation)." + } + ] + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 169, + 216, + 299, + 327 + ], + "blocks": [ + { + "bbox": [ + 146, + 174, + 483, + 207 + ], + "lines": [ + { + "bbox": [ + 146, + 174, + 483, + 207 + ], + "spans": [ + { + "bbox": [ + 146, + 174, + 483, + 207 + ], + "type": "text", + "content": "Table 4: Efficiency comparison for outlier Table 5: Efficiency comparison for scoring. Timing is measured with 10,000 thresholding. All thresholding methods samples, GPU: NVIDIA RTX 3080. are conducted on CPU." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 169, + 216, + 299, + 327 + ], + "lines": [ + { + "bbox": [ + 169, + 216, + 299, + 327 + ], + "spans": [ + { + "bbox": [ + 169, + 216, + 299, + 327 + ], + "type": "table", + "html": "
Method (OS)DeviceTime (↓)
LVAD [34]CPU645.2s
RSRAE [28]GPU121.6s
Multi-TCPU1.2s
OCSVM [51]CPU154.8s
+Multi-T105.4s
GOAD [9]GPU268.1s
+Multi-T211.7s
", + "image_path": "1676022d9cc44fc572692e70d1ed5eda9ec2e7789e270ad99c767eac28872c57.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 352, + 216, + 452, + 327 + ], + "blocks": [ + { + "bbox": [ + 352, + 216, + 452, + 327 + ], + "lines": [ + { + "bbox": [ + 352, + 216, + 452, + 327 + ], + "spans": [ + { + "bbox": [ + 352, + 216, + 452, + 327 + ], + "type": "table", + "html": "
Method (TL)Time (↓)
CPD [12]3.61s
FWFM [19]1.26s
DSN [5]3.79s
CLUST [11]7.36s
AUCP [48]1.26s
Multi-T0.97s
", + "image_path": "9ceb9270064564fc4133a27d53e60742b87f65de820f47642651cfa7a6e76733.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 133, + 366, + 480, + 533 + ], + "blocks": [ + { + "bbox": [ + 133, + 366, + 480, + 533 + ], + "lines": [ + { + "bbox": [ + 133, + 366, + 480, + 533 + ], + "spans": [ + { + "bbox": [ + 133, + 366, + 480, + 533 + ], + "type": "image", + "image_path": "e1672596f140f350da057673c08f03e6b135ce868f81e8e8d74e03bb65541534.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 540, + 480, + 552 + ], + "lines": [ + { + "bbox": [ + 132, + 540, + 480, + 552 + ], + "spans": [ + { + "bbox": [ + 132, + 540, + 480, + 552 + ], + "type": "text", + "content": "Fig. 4: AUC results of Multi-T compared with Shell-Re. [32] over various outlier ratios." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 593, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 593, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 593, + 482, + 666 + ], + "type": "text", + "content": "Discussions. We conclude the feasibility of adopting Multi-T to outlier scoring methods into three aspects: (i) both normalization with outliers and manifold learning with inliers are necessary for most of outlier scoring methods; (ii) Despite the initial outlier score distribution (Eq. 6) is usually not perfectly separable, it provides a reliable baseline. (iii) Multi-T is able to isolate the overlap region and predict both uncontaminated inliers." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 162, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 162, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 162, + 91, + 447, + 102 + ], + "type": "text", + "content": "Rethinking Unsupervised Outlier Detection via Multiple Thresholding" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 161, + 145, + 306, + 212 + ], + "blocks": [ + { + "bbox": [ + 146, + 114, + 323, + 137 + ], + "lines": [ + { + "bbox": [ + 146, + 114, + 323, + 137 + ], + "spans": [ + { + "bbox": [ + 146, + 114, + 323, + 137 + ], + "type": "text", + "content": "Table 6: Ablation study for DeepSVDD and LVAD-S with Multi-T (" + }, + { + "bbox": [ + 146, + 114, + 323, + 137 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 146, + 114, + 323, + 137 + ], + "type": "text", + "content": ": 0.2)." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 161, + 145, + 306, + 212 + ], + "lines": [ + { + "bbox": [ + 161, + 145, + 306, + 212 + ], + "spans": [ + { + "bbox": [ + 161, + 145, + 306, + 212 + ], + "type": "table", + "html": "
DeepSVDDAUCLVAD-SAUC
+Multi-T0.950+Multi-T0.971
w/o In0.643w/o In0.965
w/o Out0.939w/o Out0.951
", + "image_path": "dafb589ace953eb14b59ed10b465fcd8cb8d669835e1be0502e37fb92d2c110c.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 353, + 135, + 451, + 212 + ], + "blocks": [ + { + "bbox": [ + 328, + 114, + 476, + 125 + ], + "lines": [ + { + "bbox": [ + 328, + 114, + 476, + 125 + ], + "spans": [ + { + "bbox": [ + 328, + 114, + 476, + 125 + ], + "type": "text", + "content": "Table 7: Comparison with " + }, + { + "bbox": [ + 328, + 114, + 476, + 125 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 328, + 114, + 476, + 125 + ], + "type": "text", + "content": "-sigma." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 353, + 135, + 451, + 212 + ], + "lines": [ + { + "bbox": [ + 353, + 135, + 451, + 212 + ], + "spans": [ + { + "bbox": [ + 353, + 135, + 451, + 212 + ], + "type": "table", + "html": "
MetricMethodScore
F0.13-sigma0.406
Multi-T0.840
F101-sigma0.806
Multi-T0.869
", + "image_path": "26f614312681aa46a9c613f2782e1d7963fc19302f50853b0ce9c470392819ec.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 228, + 238, + 240 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 228, + 238, + 240 + ], + "spans": [ + { + "bbox": [ + 132, + 228, + 238, + 240 + ], + "type": "text", + "content": "5.3 Ablation Study" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 247, + 482, + 440 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 247, + 482, + 440 + ], + "spans": [ + { + "bbox": [ + 130, + 247, + 482, + 440 + ], + "type": "text", + "content": "The experiments thus far have established the effectiveness of Multi-T. However, there remains a concern regarding how sensitive the performance improvement is to Multi-T's two primary components: adaptive outliers (Out) and uncontaminated inliers (In), which refer to the normalization and manifold learning procedures, respectively. We select two representative outlier scoring models: DeepSVDD [50] and LVAD-S [34], whereas DeepSVDD follows a widely-used outlier detection mechanism, i.e., learning the normality (hyper-sphere), and identifying inliers is of great significance. Without predicted inliers, the AUC result decreases from 0.950 to 0.643 on STL-10, shown in Tab. 6. Additionally, since the normalization procedure can be considered as a distance de-noising procedure [33], Multi-T will contribute to those methods with distance computation, e.g., the result of LVAD-S is improved from 0.951 to 0.971 on STL-10 and 0.838 to 0.882 on CIFAR-10, which verify our prior assumption that the identification of both inliers and outliers is of great significance, which enhances the reliability of multiple thresholding perspective. Moreover, Tab. 7 indicates that our thresholding process markedly surpasses the classical " + }, + { + "bbox": [ + 130, + 247, + 482, + 440 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 247, + 482, + 440 + ], + "type": "text", + "content": "-sigma rule [42]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 457, + 214, + 468 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 457, + 214, + 468 + ], + "spans": [ + { + "bbox": [ + 132, + 457, + 214, + 468 + ], + "type": "text", + "content": "5.4 Limitation" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 477, + 482, + 561 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 477, + 482, + 561 + ], + "spans": [ + { + "bbox": [ + 130, + 477, + 482, + 561 + ], + "type": "text", + "content": "Our method is tied to the ranking accuracy (separability) of the initial outlier score function. Based on most related works, we consider UOD as an one-class learning task (only one inlier/normal class in the target dataset). However we find in some specific cases, there might exist multi-normal classes, e.g., the digit 5 class of the MNIST dataset. In that case, the efficacy of our method might be decreased. Additionally, this statistical-based method has limitations in very small-scale datasets since the \"three-sigma\" rule will be less effective." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 580, + 220, + 593 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 580, + 220, + 593 + ], + "spans": [ + { + "bbox": [ + 132, + 580, + 220, + 593 + ], + "type": "text", + "content": "6 Conclusion" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": "This work introduces a novel perspective for UOD about advancing existing outlier detectors (scoring methods) via thresholding. To this end, we propose a multiple thresholding (Multi-T) module, to label the unlabelled target dataset. Comprehensive experiments verify that the Multi-T can significantly improve both the efficacy and efficiency of previously proposed outlier scoring methods." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 216, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 216, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 216, + 100 + ], + "type": "text", + "content": "Z. Liu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 247, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 247, + 129 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 247, + 129 + ], + "type": "text", + "content": "Acknowledgement" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 141, + 479, + 165 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 141, + 479, + 165 + ], + "spans": [ + { + "bbox": [ + 132, + 141, + 479, + 165 + ], + "type": "text", + "content": "We would like to thank the reviewers for their valuable comments and suggestions." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 185, + 197, + 198 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 185, + 197, + 198 + ], + "spans": [ + { + "bbox": [ + 133, + 185, + 197, + 198 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 211, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 22, + "blocks": [ + { + "bbox": [ + 138, + 211, + 480, + 233 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 211, + 480, + 233 + ], + "spans": [ + { + "bbox": [ + 138, + 211, + 480, + 233 + ], + "type": "text", + "content": "1. Afsari, B.: Riemannian " + }, + { + "bbox": [ + 138, + 211, + 480, + 233 + ], + "type": "inline_equation", + "content": "L^p" + }, + { + "bbox": [ + 138, + 211, + 480, + 233 + ], + "type": "text", + "content": " center of mass: existence, uniqueness, and convexity. Proceedings of the American Mathematical Society pp. 655-673 (2011)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 234, + 371, + 244 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 234, + 371, + 244 + ], + "spans": [ + { + "bbox": [ + 138, + 234, + 371, + 244 + ], + "type": "text", + "content": "2. Aggarwal, C.C.: Outlier analysis second edition (2016)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 245, + 480, + 266 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 245, + 480, + 266 + ], + "spans": [ + { + "bbox": [ + 138, + 245, + 480, + 266 + ], + "type": "text", + "content": "3. Aggarwal, C.C., Aggarwal, C.C.: An introduction to outlier analysis. Springer (2017)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 267, + 480, + 289 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 267, + 480, + 289 + ], + "spans": [ + { + "bbox": [ + 138, + 267, + 480, + 289 + ], + "type": "text", + "content": "4. Alrawashdeh, M.J.: An adjusted grubbs' and generalized extreme studentized deviation. Demonstratio Mathematica pp. 548-557 (2021)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 290, + 480, + 322 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 290, + 480, + 322 + ], + "spans": [ + { + "bbox": [ + 138, + 290, + 480, + 322 + ], + "type": "text", + "content": "5. Amagata, D., Onizuka, M., Hara, T.: Fast and exact outlier detection in metric spaces: a proximity graph-based approach. In: Proceedings of the 2021 International Conference on Management of Data. pp. 36-48 (2021)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 323, + 480, + 355 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 323, + 480, + 355 + ], + "spans": [ + { + "bbox": [ + 138, + 323, + 480, + 355 + ], + "type": "text", + "content": "6. Archana, N., Pawar, S.: Periodicity detection of outlier sequences using constraint based pattern tree with mad. International Journal of Advanced Studies in Computers, Science and Engineering p. 34 (2015)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 355, + 480, + 388 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 355, + 480, + 388 + ], + "spans": [ + { + "bbox": [ + 138, + 355, + 480, + 388 + ], + "type": "text", + "content": "7. Barbado, A., Corcho, O., Benjamins, R.: Rule extraction in unsupervised anomaly detection for model explainability: Application to oneclasssvm. Expert Systems with Applications p. 116100 (2022)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 388, + 480, + 411 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 388, + 480, + 411 + ], + "spans": [ + { + "bbox": [ + 138, + 388, + 480, + 411 + ], + "type": "text", + "content": "8. Bardet, J.M., Dimby, S.F.: A new non-parametric detector of univariate outliers for distributions with unbounded support. Extremes pp. 751-775 (2017)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 411, + 480, + 433 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 411, + 480, + 433 + ], + "spans": [ + { + "bbox": [ + 138, + 411, + 480, + 433 + ], + "type": "text", + "content": "9. Bergman, L., Hoshen, Y.: Classification-based anomaly detection for general data. arXiv preprint arXiv:2005.02359 (2020)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 434, + 480, + 455 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 434, + 480, + 455 + ], + "spans": [ + { + "bbox": [ + 138, + 434, + 480, + 455 + ], + "type": "text", + "content": "10. Bol'shev, L., Ubaidullaeva, M.: Chauvenet's test in the classical theory of errors. Theory of Probability & Its Applications pp. 683-692 (1975)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 456, + 480, + 488 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 456, + 480, + 488 + ], + "spans": [ + { + "bbox": [ + 138, + 456, + 480, + 488 + ], + "type": "text", + "content": "1. Breunig, M.M., Kriegel, H.P., Ng, R.T., Sander, J.: Lof: identifying density-based local outliers. In: Proceedings of the ACM SIGMOD International Conference on Management of Data. pp. 93-104 (2000)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 489, + 480, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 489, + 480, + 510 + ], + "spans": [ + { + "bbox": [ + 138, + 489, + 480, + 510 + ], + "type": "text", + "content": "2. Van den Burg, G.J., Williams, C.K.: An evaluation of change point detection algorithms. arXiv preprint arXiv:2003.06222 (2020)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 511, + 480, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 511, + 480, + 553 + ], + "spans": [ + { + "bbox": [ + 138, + 511, + 480, + 553 + ], + "type": "text", + "content": "3. Camposeco, F., Sattler, T., Cohen, A., Geiger, A., Pollefeys, M.: Toroidal constraints for two-point localization under high outlier ratios. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 4545-4553 (2017)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 555, + 480, + 587 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 555, + 480, + 587 + ], + "spans": [ + { + "bbox": [ + 138, + 555, + 480, + 587 + ], + "type": "text", + "content": "4. Chalapathy, R., Menon, A.K., Chawla, S.: Robust, deep and inductive anomaly detection. In: Proceedings of Joint European Conference on Machine Learning and Knowledge Discovery in Databases. pp. 36-51 (2017)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 138, + 588, + 480, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 588, + 480, + 620 + ], + "spans": [ + { + "bbox": [ + 138, + 588, + 480, + 620 + ], + "type": "text", + "content": "5. Coates, A., Ng, A., Lee, H.: An analysis of single-layer networks in unsupervised feature learning. In: Proceedings of the International Conference on Artificial Intelligence and Statistics. pp. 215-223 (2011)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 138, + 621, + 480, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 621, + 480, + 643 + ], + "spans": [ + { + "bbox": [ + 138, + 621, + 480, + 643 + ], + "type": "text", + "content": "6. Coin, D.: Testing normality in the presence of outliers. Statistical Methods and Applications pp. 3-12 (2008)" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 138, + 643, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 643, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 643, + 480, + 665 + ], + "type": "text", + "content": "7. Friendly, M., Monette, G., Fox, J.: Elliptical insights: understanding statistical methods through elliptical geometry. Statistical Science pp. 1-39 (2013)" + } + ] + } + ], + "index": 21 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 162, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 162, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 162, + 91, + 447, + 102 + ], + "type": "text", + "content": "Rethinking Unsupervised Outlier Detection via Multiple Thresholding" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "type": "text", + "content": "18. Goodge, A., Hooi, B., Ng, S.K., Ng, W.S.: Lunar: Unifying local outlier detection methods via graph neural networks. In: Proceedings of the AAAI Conference on Artificial Intelligence. pp. 6737-6745 (2022)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 149, + 480, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 149, + 480, + 182 + ], + "spans": [ + { + "bbox": [ + 133, + 149, + 480, + 182 + ], + "type": "text", + "content": "19. Hashemi, N., German, E.V., Ramirez, J.P., Ruths, J.: Filtering approaches for dealing with noise in anomaly detection. In: 2019 IEEE 58th Conference on Decision and Control. pp. 5356-5361 (2019)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 182, + 480, + 214 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 182, + 480, + 214 + ], + "spans": [ + { + "bbox": [ + 132, + 182, + 480, + 214 + ], + "type": "text", + "content": "20. He, K., Girshick, R., Dollar, P.: Rethinking imagenet pre-training. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 4918-4927 (2019)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 214, + 480, + 247 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 214, + 480, + 247 + ], + "spans": [ + { + "bbox": [ + 132, + 214, + 480, + 247 + ], + "type": "text", + "content": "21. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 770-778 (2016)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 247, + 480, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 247, + 480, + 267 + ], + "spans": [ + { + "bbox": [ + 132, + 247, + 480, + 267 + ], + "type": "text", + "content": "22. Hendrycks, D., Mazeika, M., Dietterich, T.: Deep anomaly detection with outlier exposure. arXiv preprint arXiv:1812.04606 (2018)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 267, + 480, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 267, + 480, + 300 + ], + "spans": [ + { + "bbox": [ + 132, + 267, + 480, + 300 + ], + "type": "text", + "content": "23. Huyan, N., Quan, D., Zhang, X., Liang, X., Chanussot, J., Jiao, L.: Unsupervised outlier detection using memory and contrastive learning. IEEE Transactions on Image Processing pp. 6440-6454 (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 300, + 480, + 332 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 300, + 480, + 332 + ], + "spans": [ + { + "bbox": [ + 132, + 300, + 480, + 332 + ], + "type": "text", + "content": "24. Iouchtchenko, D., Raymond, N., Roy, P.N., Nooijen, M.: Deterministic and quasi-random sampling of optimized gaussian mixture distributions for vibronic monte carlo. arXiv preprint arXiv:1912.11594 (2019)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 333, + 480, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 333, + 480, + 365 + ], + "spans": [ + { + "bbox": [ + 132, + 333, + 480, + 365 + ], + "type": "text", + "content": "25. Keyzer, M.A., Sonneveld, B.: Using the mollifier method to characterize datasets and models: the case of the universal soil loss equation. ITC Journal pp. 263-272 (1997)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 365, + 480, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 365, + 480, + 397 + ], + "spans": [ + { + "bbox": [ + 132, + 365, + 480, + 397 + ], + "type": "text", + "content": "26. Kriegel, H.P., Schubert, M., Zimek, A.: Angle-based outlier detection in high-dimensional data. In: Proceedings of the 14th ACM SIGKDD international conference on Knowledge discovery and data mining. pp. 444-452 (2008)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 397, + 480, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 397, + 480, + 418 + ], + "spans": [ + { + "bbox": [ + 132, + 397, + 480, + 418 + ], + "type": "text", + "content": "27. Krizhevsky, A., Hinton, G., et al.: Learning multiple layers of features from tiny images (2009)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 418, + 480, + 440 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 418, + 480, + 440 + ], + "spans": [ + { + "bbox": [ + 132, + 418, + 480, + 440 + ], + "type": "text", + "content": "28. Lai, C.H., Zou, D., Lerman, G.: Robust subspace recovery layer for unsupervised anomaly detection. arXiv preprint arXiv:1904.00152 (2019)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 440, + 416, + 450 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 440, + 416, + 450 + ], + "spans": [ + { + "bbox": [ + 132, + 440, + 416, + 450 + ], + "type": "text", + "content": "29. LeCun, Y., Cortes, C.: MNIST handwritten digit database (2010)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 450, + 480, + 482 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 450, + 480, + 482 + ], + "spans": [ + { + "bbox": [ + 132, + 450, + 480, + 482 + ], + "type": "text", + "content": "30. Li, T., Wang, Z., Liu, S., Lin, W.Y.: Deep unsupervised anomaly detection. In: Proceedings of IEEE Winter Conference on Applications of Computer Vision. pp. 3636-3645 (2021)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 482, + 480, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 482, + 480, + 514 + ], + "spans": [ + { + "bbox": [ + 132, + 482, + 480, + 514 + ], + "type": "text", + "content": "31. Li, Z., Zhao, Y., Hu, X., Botta, N., Ionescu, C., Chen, G.H.: Ecod: Unsupervised outlier detection using empirical cumulative distribution functions. arXiv preprint arXiv:2201.00382 (2022)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 514, + 480, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 514, + 480, + 547 + ], + "spans": [ + { + "bbox": [ + 132, + 514, + 480, + 547 + ], + "type": "text", + "content": "32. Lin, D., Liu, S., Li, H., Cheung, N.M., Ren, C., Matsushita, Y.: Shell theory: A statistical model of reality. IEEE Transactions on Pattern Analysis and Machine Intelligence (2021)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 547, + 480, + 579 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 547, + 480, + 579 + ], + "spans": [ + { + "bbox": [ + 132, + 547, + 480, + 579 + ], + "type": "text", + "content": "33. Lin, W.Y., Liu, S., Dai, B.T., Li, H.: Distance based image classification: A solution to generative classification's conundrum? International Journal of Computer Vision pp. 177-198 (2023)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 579, + 480, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 579, + 480, + 611 + ], + "spans": [ + { + "bbox": [ + 132, + 579, + 480, + 611 + ], + "type": "text", + "content": "34. Lin, W.Y., Liu, Z., Liu, S.: Locally varying distance transform for unsupervised visual anomaly detection. In: European Conference on Computer Vision. pp. 354-371. Springer (2022)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 611, + 480, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 611, + 480, + 633 + ], + "spans": [ + { + "bbox": [ + 132, + 611, + 480, + 633 + ], + "type": "text", + "content": "35. Liu, F.T., Ting, K.M., Zhou, Z.H.: Isolation forest. In: Proceedings of International Conference on Data Mining. pp. 413-422 (2008)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 132, + 633, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 633, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 633, + 480, + 665 + ], + "type": "text", + "content": "36. Liznerski, P., Ruff, L., Vandermeulen, R.A., Franks, B.J., Müller, K.R., Kloft, M.: Exposing outlier exposure: What can be learned from few, one, and zero outlier images. arXiv preprint arXiv:2205.11474 (2022)" + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 216, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 216, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 216, + 100 + ], + "type": "text", + "content": "Z. Liu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 481, + 666 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 130, + 116, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 480, + 138 + ], + "type": "text", + "content": "37. Martin, M.A., Roberts, S.: An evaluation of bootstrap methods for outlier detection in least squares regression. Journal of Applied Statistics pp. 703-720 (2006)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 140, + 481, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 140, + 481, + 172 + ], + "spans": [ + { + "bbox": [ + 132, + 140, + 481, + 172 + ], + "type": "text", + "content": "38. Masci, J., Meier, U., Ciresan, D., Schmidhuber, J.: Stacked convolutional autoencoders for hierarchical feature extraction. In: Proceedings of International Conference on Artificial Neural Networks. pp. 52-59 (2011)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 173, + 481, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 173, + 481, + 217 + ], + "spans": [ + { + "bbox": [ + 132, + 173, + 481, + 217 + ], + "type": "text", + "content": "39. Pang, G., Cao, L., Chen, L., Liu, H.: Learning representations of ultrahigh-dimensional data for random distance-based outlier detection. In: Proceedings of the 24th ACM SIGKDD international conference on knowledge discovery & data mining. pp. 2041-2050 (2018)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 218, + 481, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 218, + 481, + 239 + ], + "spans": [ + { + "bbox": [ + 132, + 218, + 481, + 239 + ], + "type": "text", + "content": "40. Parzen, E.: On estimation of a probability density function and mode. The annals of mathematical statistics pp. 1065-1076 (1962)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 240, + 481, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 240, + 481, + 274 + ], + "spans": [ + { + "bbox": [ + 132, + 240, + 481, + 274 + ], + "type": "text", + "content": "41. Perera, P., Nallapati, R., Xiang, B.: Ocgan: One-class novelty detection using gans with constrained latent representations. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 2898-2906 (2019)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 274, + 481, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 274, + 481, + 285 + ], + "spans": [ + { + "bbox": [ + 132, + 274, + 481, + 285 + ], + "type": "text", + "content": "42. Pukelsheim, F.: The three sigma rule. The American Statistician pp. 88-91 (1994)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 285, + 481, + 317 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 285, + 481, + 317 + ], + "spans": [ + { + "bbox": [ + 132, + 285, + 481, + 317 + ], + "type": "text", + "content": "43. Qi, Z., Jiang, D., Chen, X.: Iterative gradient descent for outlier detection. International Journal of Wavelets, Multiresolution and Information Processing p. 2150004 (2021)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 319, + 481, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 319, + 481, + 352 + ], + "spans": [ + { + "bbox": [ + 132, + 319, + 481, + 352 + ], + "type": "text", + "content": "44. Qiu, C., Pfrommer, T., Kloft, M., Mandt, S., Rudolph, M.: Neural transformation learning for deep anomaly detection beyond images. In: International Conference on Machine Learning. pp. 8703-8714 (2021)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 352, + 481, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 352, + 481, + 396 + ], + "spans": [ + { + "bbox": [ + 132, + 352, + 481, + 396 + ], + "type": "text", + "content": "45. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: Proceedings of International Conference on Machine Learning. pp. 8748-8763 (2021)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 396, + 481, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 396, + 481, + 418 + ], + "spans": [ + { + "bbox": [ + 132, + 396, + 481, + 418 + ], + "type": "text", + "content": "46. Raymaekers, J., Rousseeuw, P.J.: Transforming variables to central normality. Machine Learning pp. 1-23 (2021)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 419, + 481, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 419, + 481, + 441 + ], + "spans": [ + { + "bbox": [ + 132, + 419, + 481, + 441 + ], + "type": "text", + "content": "47. Reiss, T., Hoshen, Y.: Mean-shifted contrastive loss for anomaly detection. In: Proceedings of the AAAI Conference on Artificial Intelligence. pp. 2155-2162 (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 442, + 481, + 485 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 442, + 481, + 485 + ], + "spans": [ + { + "bbox": [ + 132, + 442, + 481, + 485 + ], + "type": "text", + "content": "48. Ren, K., Yang, H., Zhao, Y., Chen, W., Xue, M., Miao, H., Huang, S., Liu, J.: A robust auc maximization framework with simultaneous outlier detection and feature selection for positive-unlabeled classification. IEEE Transactions on Neural Networks and Learning Systems pp. 3072-3083 (2018)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 487, + 481, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 487, + 481, + 520 + ], + "spans": [ + { + "bbox": [ + 132, + 487, + 481, + 520 + ], + "type": "text", + "content": "49. Rengasamy, D., Rothwell, B.C., Figueredo, G.P.: Towards a more reliable interpretation of machine learning outputs for safety-critical systems using feature importance fusion. Applied Sciences p. 11854 (2021)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 520, + 481, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 520, + 481, + 553 + ], + "spans": [ + { + "bbox": [ + 132, + 520, + 481, + 553 + ], + "type": "text", + "content": "50. Ruff, L., Vandermeulen, R., Goernitz, N., Deecke, L., Siddiqui, S.A., Binder, A., Müller, E., Kloft, M.: Deep one-class classification. In: Proceedings of International Conference on Machine Learning. pp. 4393-4402 (2018)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 554, + 481, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 554, + 481, + 586 + ], + "spans": [ + { + "bbox": [ + 132, + 554, + 481, + 586 + ], + "type": "text", + "content": "51. Scholkopf, B., Platt, J.C., Shawe-Taylor, J., Smola, A.J., Williamson, R.C.: Estimating the support of a high-dimensional distribution. Neural Computation 13(7), 1443-1471 (2001)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 586, + 481, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 586, + 481, + 609 + ], + "spans": [ + { + "bbox": [ + 132, + 586, + 481, + 609 + ], + "type": "text", + "content": "52. Scholkopf, B., Smola, A.J., Bach, F., et al.: Learning with kernels: support vector machines, regularization, optimization, and beyond (2002)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 609, + 481, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 609, + 481, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 609, + 481, + 632 + ], + "type": "text", + "content": "53. Shenkar, T., Wolf, L.: Anomaly detection for tabular data with internal contrastive learning. In: International Conference on Learning Representations (2021)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 632, + 481, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 481, + 666 + ], + "type": "text", + "content": "54. Shyu, M.L., Chen, S.C., Sarinnapakorn, K., Chang, L.: A novel anomaly detection scheme based on principal component classifier. In: Proceedings of the IEEE foundations and new directions of data mining workshop. pp. 172-179 (2003)" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 162, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 162, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 162, + 91, + 447, + 102 + ], + "type": "text", + "content": "Rethinking Unsupervised Outlier Detection via Multiple Thresholding" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 414 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "type": "text", + "content": "55. Thanammal, K., Jayasudha, J., Vijayalakshmi, R., Arumugaperumal, S.: Effective histogram thresholding techniques for natural images using segmentation. Journal of Image and Graphics pp. 113-116 (2014)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 150, + 481, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 150, + 481, + 182 + ], + "spans": [ + { + "bbox": [ + 132, + 150, + 481, + 182 + ], + "type": "text", + "content": "56. Wang, S., Zeng, Y., Liu, X., Zhu, E., Yin, J., Xu, C., Kloft, M.: Effective end-to-end unsupervised outlier detection via inlier priority of discriminative network. Proceedings of Advances in Neural Information Processing Systems (2019)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 183, + 481, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 183, + 481, + 226 + ], + "spans": [ + { + "bbox": [ + 132, + 183, + 481, + 226 + ], + "type": "text", + "content": "57. Wang, S., Zhu, E., Hu, X., Liu, X., Liu, Q., Yin, J., Wang, F.: Robustness can be cheap: A highly efficient approach to discover outliers under high outlier ratios. In: Proceedings of the AAAI Conference on Artificial Intelligence. pp. 5313-5320 (2019)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 227, + 481, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 227, + 481, + 259 + ], + "spans": [ + { + "bbox": [ + 132, + 227, + 481, + 259 + ], + "type": "text", + "content": "58. Xia, Y., Cao, X., Wen, F., Hua, G., Sun, J.: Learning discriminative reconstructions for unsupervised outlier removal. In: Proceedings of the IEEE International Conference on Computer Vision. pp. 1511-1519 (2015)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 259, + 481, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 259, + 481, + 281 + ], + "spans": [ + { + "bbox": [ + 132, + 259, + 481, + 281 + ], + "type": "text", + "content": "59. Xiao, H., Rasul, K., Vollgraf, R.: Fashion-mnist: a novel image dataset for benchmarking machine learning algorithms (2017)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 281, + 481, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 281, + 481, + 314 + ], + "spans": [ + { + "bbox": [ + 132, + 281, + 481, + 314 + ], + "type": "text", + "content": "60. Xu, H., Wang, Y., Wei, J., Jian, S., Li, Y., Liu, N.: Fascinating supervisory signals and where to find them: Deep anomaly detection with scale learning. arXiv preprint arXiv:2305.16114 (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 315, + 481, + 347 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 315, + 481, + 347 + ], + "spans": [ + { + "bbox": [ + 132, + 315, + 481, + 347 + ], + "type": "text", + "content": "61. Zhou, B., Lapedriza, A., Khosla, A., Oliva, A., Torralba, A.: Places: A 10 million image database for scene recognition. IEEE Transactions on Pattern Analysis and Machine Intelligence pp. 1452-1464 (2017)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 347, + 481, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 347, + 481, + 380 + ], + "spans": [ + { + "bbox": [ + 132, + 347, + 481, + 380 + ], + "type": "text", + "content": "62. Zhou, C., Paffenroth, R.C.: Anomaly detection with robust deep autoencoders. In: Proceedings of the ACM SIGKDD International Conference on Knowledge Discovery and Data Mining. pp. 665-674 (2017)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 380, + 481, + 414 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 380, + 481, + 414 + ], + "spans": [ + { + "bbox": [ + 132, + 380, + 481, + 414 + ], + "type": "text", + "content": "63. Zong, B., Song, Q., Min, M.R., Cheng, W., Lumezanu, C., Cho, D., Chen, H.: Deep autoencoding gaussian mixture model for unsupervised anomaly detection. In: Proceedings of International Conference on Learning Representations (2018)" + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 216, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 216, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 216, + 100 + ], + "type": "text", + "content": "Z. Liu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Rethinking Video Deblurring with Wavelet-Aware Dynamic Transformer and Diffusion Model/7afdf756-63a6-48fd-b654-3bb54bd5d976_content_list.json b/2024/Rethinking Video Deblurring with Wavelet-Aware Dynamic Transformer and Diffusion Model/7afdf756-63a6-48fd-b654-3bb54bd5d976_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..5eae5e8436a58390218b6845fd492eff826ee877 --- /dev/null +++ b/2024/Rethinking Video Deblurring with Wavelet-Aware Dynamic Transformer and Diffusion Model/7afdf756-63a6-48fd-b654-3bb54bd5d976_content_list.json @@ -0,0 +1,2602 @@ +[ + { + "type": "text", + "text": "Rethinking Video Deblurring with Wavelet-Aware Dynamic Transformer and Diffusion Model", + "text_level": 1, + "bbox": [ + 217, + 140, + 785, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Chen Rao $^{1\\dagger}$ , Guangyuan Li $^{1\\dagger}$ , Zehua Lan $^{1}$ , Jiakai Sun $^{1}$ , Junsheng Luan $^{1}$ , Wei Xing $^{1*}$ , Lei Zhao $^{1*}$ , Huaizhong Lin $^{1*}$ , Jianfeng Dong $^{2}$ , and Dalong Zhang $^{1}$", + "bbox": [ + 222, + 210, + 779, + 258 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 College of Computer Science and Technology, Zhejiang University", + "bbox": [ + 272, + 268, + 725, + 282 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ Zhejiang Gongshang University", + "bbox": [ + 388, + 282, + 612, + 297 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Current video deblurring methods have limitations in recovering high-frequency information since the regression losses are conservative with high-frequency details. Since Diffusion Models (DMs) have strong capabilities in generating high-frequency details, we consider introducing DMs into the video deblurring task. However, we found that directly applying DMs to the video deblurring task has the following problems: (1) DMs require many iteration steps to generate videos from Gaussian noise, which consumes many computational resources. (2) DMs are easily misled by the blurry artifacts in the video, resulting in irrational content and distortion of the deblurred video. To address the above issues, we propose a novel video deblurring framework VD-Diff that integrates the diffusion model into the Wavelet-Aware Dynamic Transformer (WADT). Specifically, we perform the diffusion model in a highly compact latent space to generate prior features containing high-frequency information that conforms to the ground truth distribution. We design the WADT to preserve and recover the low-frequency information in the video while utilizing the high-frequency information generated by the diffusion model. Extensive experiments show that our proposed VD-Diff outperforms SOTA methods on GoPro, DVD, BSD, and Real-World Video datasets. The codes will be available at https://github.com/Chen-Rao/VD-Diff.", + "bbox": [ + 261, + 335, + 743, + 625 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Diffusion model $\\cdot$ Prior feature $\\cdot$ Video deblurring $\\cdot$ Wavelet-based transformer", + "bbox": [ + 259, + 638, + 740, + 667 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 693, + 375, + 709 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Portable devices are widely used to capture videos. However, camera shake and fast movement of objects lead to undesirable blur in video. To reduce the impact of video blur, researchers have done a lot of work on video deblurring. Traditional methods are mainly based on hand-crafted assumptions and priors, which are difficult to design. Besides, inappropriate priors will reduce the quality of video deblurring.", + "bbox": [ + 212, + 724, + 787, + 816 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† co-first authors, equal contribution; * Corresponding author", + "bbox": [ + 230, + 824, + 643, + 839 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In the past decade, video deblurring has witnessed significant progresses with the development of deep learning. Early deep learning methods mainly use CNN-based models [13, 14, 26, 28, 36, 42, 46] to extract video features for capturing the spatio-temporal information between frames. Nevertheless, these approaches have limitations in modelling long-range spatial dependencies and capturing non-local self-similarities.", + "bbox": [ + 212, + 146, + 787, + 236 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recently, Transformer has been employed in video deblurring tasks [4, 21, 22, 37] to address limitations in CNN-based methods, which can establish long-range spatial dependencies and capture non-local spatial information. Although Transformer-based methods exhibit impressive performance, they still have the following issues. On the one hand, the methods using standard global Transformer [8] tend to deblur videos based on low-frequency details and have limited perception of high-frequency information. On the other hand, when the local window-based Transformer [37] is used to reduce the computational cost, self-attention is calculated within a location-specific window, resulting in a limited receptive field. Consequently, it may ignore some key elements of similar and clearer scene patches in the spatiotemporal neighborhood. Therefore, the local window-based self-attention fails to capture valuable high-frequency information from the spatiotemporal neighborhood, resulting in the loss of high-frequency information.", + "bbox": [ + 212, + 239, + 787, + 450 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Very recently, diffusion models (DMs) have shown their outstanding capabilities in video synthesis tasks [2,9,23,24,31,32,40], which can denoise pure Gaussian noise into high-quality video with abundant high-frequency details. Therefore, we consider employing DMs in video deblurring to restore high-frequency information. However, directly applying DMs to video deblurring tasks has the following problems. First, DMs demand a large number of iteration steps to generate videos from Gaussian noise, which costs a lot of computational resources. Second, DMs are sensitive to motion blur in videos, resulting in irrational content and distortion of the deblurred video.", + "bbox": [ + 212, + 455, + 787, + 590 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To solve the above problems, we propose a novel method called VD-Diff, which combines our proposed Wavelet-Aware Dynamic Transformer (WADT) and diffusion model for video deblurring. Specifically, we employ the diffusion model in a highly compact latent space to generate prior features containing high-frequency information that conforms to the ground truth distribution, which facilitates video deblurring. Applying DM in a highly compressed latent space can significantly reduce the number of denoising iteration steps.", + "bbox": [ + 212, + 595, + 787, + 700 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In addition, we design WADT to retain and restore the low-frequency information in the video while exacting the high-frequency information from the prior features generated by the diffusion model, which ensures the deblurred video is undistorted. WADT first decomposes the blurred video frames into the approximation coefficient and detail coefficients by Wavelet Transform (WT) and extracts low-frequency information from the approximation coefficient by Wavelet-Aware Dynamic Transformer Layers (WADTL). Meanwhile, WADT employs WADTL to fuse the prior features generated by DM containing high-frequency information to generate artifact-free and distortion-free videos. Furthermore, we", + "bbox": [ + 212, + 704, + 787, + 839 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "C. Rao et al.", + "bbox": [ + 271, + 114, + 359, + 128 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/810cf35c3598d5d428cf97f7db2b4bf6a027e7752d709a6ce97aa11d954f7ffc.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 215, + 143, + 782, + 301 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/b60cb3f2a0343fc49bdd654572a345deb4375720a04a7c77fc098585e2463ac8.jpg", + "image_caption": [ + "Wavelet-Aware Dynamic Transformer Layer" + ], + "image_footnote": [], + "bbox": [ + 220, + 306, + 236, + 319 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/7545eba5884d0c00867b75e62d5916668bfffa9eadc11326f8a846f1aefe4025.jpg", + "image_caption": [ + "Wavelet-based Bidirectional Propagation Fuse", + "Fig. 1: The overall architecture of the proposed VD-Diff, which consists of Wavelet-Aware Dynamic Transformer (WADT) and Diffusion Model (DM). Specifically, WADT adopts Wavelet Transform (WT) for feature separation, WADT Layer (WADTL) for deep feature extraction, and Wavelet-based Bidirectional Propagation Fuse (WBPF) for spatio-temporal information propagation between frames. The DM generates prior features to supplement high-frequency information for the deblurring process in WADTL.WT: Wavelet Transform. IWT: Inverse Wavelet Transform." + ], + "image_footnote": [], + "bbox": [ + 401, + 308, + 418, + 318 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/9fc35a009037c40cd4b578e81cbe60655a8d6fb51e8f963eb7eb96a1c05035c0.jpg", + "image_caption": [ + "Concatenation" + ], + "image_footnote": [], + "bbox": [ + 596, + 308, + 611, + 318 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/17f159f7cd8d1f37efb1d969d1be38115041955441c8167bb3f13b8ec94c4539.jpg", + "image_caption": [ + "Element-wise Summation" + ], + "image_footnote": [], + "bbox": [ + 669, + 308, + 686, + 318 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "introduce a Wavelet-based Bidirectional Propagation Fuse (WBPF) module in WADT to fully use the spatio-temporal information between frames, which ensure that the video is undistorted in the spatio-temporal dimension. Our main contributions are as follows:", + "bbox": [ + 212, + 460, + 782, + 518 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "(1) We propose a novel method combining the Wavelet-Aware Dynamic Transformer (WADT) and the diffusion model for video deblurring. To the best of our knowledge, our study is the first to employ DM for video deblurring, which requires only a few iteration steps to restore artifact-free and distortion-free videos.", + "(2) Our diffusion model can generate prior features containing high-frequency information that conforms to the GT distribution, which facilitates video deblurring.", + "(3) Our proposed WADT can capture low-frequency information while utilizing the prior features generated by the DM to supplement high-frequency information, which further enhance the quality of deblurred video.", + "(4) Extensive experiments show that our proposed VD-Diff outperforms SOTA methods on GoPro, DVD, BSD and Real-World Video datasets." + ], + "bbox": [ + 212, + 521, + 789, + 715 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 214, + 739, + 387, + 755 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.1 Video Deblurring", + "text_level": 1, + "bbox": [ + 214, + 770, + 406, + 786 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "With the development of deep learning [20], CNN is widely used in video deblurring. For instance, Zhong et al. [46] propose residual dense blocks and a global spatio-temporal attention module for video deblurring; Pan et al. [28] develop a", + "bbox": [ + 212, + 794, + 782, + 840 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "Video Deblurring with Transformer and Diffusion Model", + "bbox": [ + 354, + 114, + 730, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "deep CNN model to estimate optical flow from intermediate latent frames. CNN-based methods have driven the development of video deblurring, but CNN has limitations in capturing long-range dependencies. To address these issues, some researchers applied the Transformer to video deblurring and achieved impressive results. For example, Lin et al. [22] propose a method to calculate self-attention by capturing key elements of adjacent frames through optical flow. Liang et al. [21] propose a recurrent video restoration transformer to process local neighboring frames in parallel within a globally recurrent framework. However, the above Transformer-based methods have limited ability to capture high-frequency information, leading to the deblurred videos lacking details.", + "bbox": [ + 212, + 146, + 787, + 297 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.2 Diffusion Models", + "text_level": 1, + "bbox": [ + 214, + 316, + 403, + 330 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "DMs have shown excellent performance in the field of synthesis. LDM [31] reaches a near-optimal point between complexity reduction and detail preservation by applying DM in the latent space. Blattmann et al. [2] turns the image generator into a video generator by introducing a temporal dimension to the latent space diffusion model. Artbank [43] utilize textual embedding to learn the visual priors from the images, synthesizing highly realistic artistic stylized images. DMs have also shown strong capabilities in the field of image restoration. For instance, DiffIR [38], Hi-diff [5] and DiffMSR [18] use DM to generate prior representations for image restoration. Although some methods, such as DDIM [45] and LDM [31], have reduced the computational complexity of DM by changing the denoising process or compressing the image, using existing DM for video deblurring still requires unaffordable computational resources. To cope with the above issue, we apply DM in a highly compact latent space to reduce the iteration steps. We were the first to apply diffusion models to video deblurring, and our models achieved outstanding performance.", + "bbox": [ + 212, + 337, + 787, + 565 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.3 Wavelet-based Restoration Methods", + "text_level": 1, + "bbox": [ + 214, + 583, + 563, + 598 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Wavelet is widely used in low-visual restoration tasks [7, 15, 19, 44, 48] as it can separate low-frequency and high-frequency information. For example, Huang et al. [12] propose a wavelet-based CNN for multi-scale face super resolution. WavTrans [17] combines wavelet transform with Transformer for feature extraction and fusion; DSTNet [29] develops a wavelet-based feature propagation method to effectively propagate main structures from long-range frames for better video deblurring. Inspired by the above methods, we proposed Wavelet-Aware Dynamic Transformer, which can extract the global structure of the video and utilize the prior features generated by DM to restore high-frequency details.", + "bbox": [ + 212, + 604, + 787, + 742 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3 Methodology", + "text_level": 1, + "bbox": [ + 214, + 760, + 380, + 777 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1 Overall Architecture", + "text_level": 1, + "bbox": [ + 214, + 787, + 433, + 803 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The overall framework of our proposed VD-Diff is shown in Fig. 1, which consists of two parts: (1) Wavelet-Aware Dynamic Transformer (WADT); (2) Diffusion", + "bbox": [ + 212, + 809, + 785, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "C. Rao et al.", + "bbox": [ + 271, + 114, + 359, + 128 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Model. Specifically, we first employ Diffusion Model to obtain a highly compact latent prior feature $z \\in \\mathbb{R}^{T \\times C'}$ , which contains supplementary high-frequency information. Note that $C'$ is the number of channels of latent features. Then, we employ WADT to deblur the blurry video frames $V_{blur} \\in R^{T \\times 3 \\times H \\times W}$ into $V_{HQ} \\in R^{T \\times 3 \\times H \\times W}$ with the help of the prior feature $z$ .", + "bbox": [ + 212, + 146, + 787, + 223 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2 Wavelet-Aware Dynamic Transformer", + "text_level": 1, + "bbox": [ + 214, + 243, + 571, + 258 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As shown in the WADT part of Fig. 1, we first utilize Conv3D to perform shallow feature extraction on the blurry video frames $V_{blur} \\in R^{T \\times 3 \\times H \\times W}$ to obtain a global feature $F_{in} \\in \\mathbb{R}^{T \\times C \\times H \\times W}$ . Then, we use Wavelet Transform (WT) to decompose it into approximate coefficients $F_{a} \\in \\mathbb{R}^{T \\times C / 4 \\times H \\times W}$ and detail coefficients $F_{d} \\in \\mathbb{R}^{T \\times 3C / 4 \\times H \\times W}$ and again use WT on $F_{a}$ to obtain $F_{aa} \\in \\mathbb{R}^{T \\times \\hat{C} \\times H \\times W}$ and $F_{ad} \\in \\mathbb{R}^{T \\times 3\\hat{C} \\times H \\times W}$ , where $\\hat{C} = C / 16$ . Next, we employ WADTL to obtain the deblurred features $F_{out}$ from the approximate coefficients $F_{aa}$ and the prior feature $z'$ . Then, we use WBPF to explore spatiotemporal information further. Finally, we use two IWTs to fuse features to obtain artifact-free and distortion-free videos $V_{HQ}$ . We will elaborate on the structure of WADTL and WBPF as follows.", + "bbox": [ + 212, + 267, + 787, + 434 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Wavelet-Aware Dynamic Transformer Layer. As shown in Fig. 1, we employ a stack of multiple WADT layer (WADTL) modules hierarchically to fuse approximate coefficients $F_{aa}$ and prior features $z'$ . As shown in Fig. 2, WADTL consists of a Wavelet-Aware Dynamic Multi-head Self-Attention (WAD-MSA) and a Wavelet-Aware Dynamic Feed-Forward Network (WAD-FFN).", + "bbox": [ + 212, + 435, + 787, + 508 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In the WAD-MSA part, given a approximate coefficients $F_{aa}$ , we use the highly compact latent prior feature $z'$ to guide video restoration:", + "bbox": [ + 212, + 510, + 787, + 541 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {F} = W _ {l} ^ {1} z ^ {\\prime} \\odot L N \\left(F _ {a a}\\right) + W _ {l} ^ {2} z ^ {\\prime}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 392, + 550, + 784, + 566 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\odot$ indicates element-wise multiplication, $LN$ denotes layer normalization [1], $W_{l}$ represents linear layer. Then, we utilize a 3D convolution layer and a 2D convolution layer to project $\\hat{F}$ into query $Q = W_2^Q W_3^Q\\hat{F}$ , key $K = W_2^K W_3^K\\hat{F}$ , value $V = W_2^V W_3^V\\hat{F}$ where $W_{3}$ is the channel-wise 3D convolution and $W_{2}$ is the depth-wise 2D convolution. Note that we use 3D convolution in the $T$ , $H$ , and $W$ dimensions and the same 2D convolution for the $T$ features in the temporal dimension $T$ . Next, we perform the attention calculation:", + "bbox": [ + 214, + 577, + 787, + 686 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {F} _ {\\text {o u t}} ^ {\\prime} = W _ {a} \\hat {V} \\cdot \\operatorname {S o f t m a x} (\\hat {K} \\cdot \\hat {Q} / \\gamma) + F _ {a a}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 357, + 696, + 784, + 713 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\gamma$ is a learnable scaling parameter.", + "bbox": [ + 214, + 723, + 513, + 737 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In the WDA-FFN part, we first integrate the prior $z'$ into $\\mathcal{F}_{out}'$ to obtain $\\hat{\\mathcal{F}}_{out}'$ . And then we use Conv3D to aggregate temporal information and use Conv2D to aggregate spatial information from neighboring pixels. Besides, we employ a gating mechanism to obtain more valuable information. The overall process of WDA-FFN can be described as:", + "bbox": [ + 212, + 737, + 787, + 813 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nF _ {o u t} = G \\left(W _ {2} ^ {1} W _ {3} ^ {1} \\hat {\\mathcal {F}} _ {o u t} ^ {\\prime}\\right) \\odot W _ {2} ^ {2} W _ {3} ^ {2} \\hat {\\mathcal {F}} _ {o u t} ^ {\\prime} + \\hat {\\mathcal {F}} _ {o u t} ^ {\\prime}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 344, + 824, + 784, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "Video Deblurring with Transformer and Diffusion Model", + "bbox": [ + 354, + 114, + 732, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/d28e1dae8032e1ea2cb2493272ec1cc6267ae2a5c209484a50333d34f2f06ca5.jpg", + "image_caption": [ + "Fig. 2: The illustration of Wavelet-Aware Dynamic Transformer Layer (WADTL), which consists of Wavelet-Aware Dynamic Multi-head Self-Attention (WAD-MSA) and Wavelet-Aware Dynamic Feed-Forward Network (WAD-FFN)." + ], + "image_footnote": [], + "bbox": [ + 220, + 143, + 782, + 231 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/7f1466d096027a7e03b75b2d7e41e271be9fee6ef117286783d77d2db398e0b8.jpg", + "image_caption": [ + "Fig. 3: The structure of the Forward Process in the WBPF. The Backward Process has the same network structure as the Forward Process, but the direction of information propagation between frames is opposite." + ], + "image_footnote": [], + "bbox": [ + 217, + 305, + 782, + 481 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $G$ denotes Gaussian Error Linear Units (GELU).", + "bbox": [ + 212, + 574, + 616, + 589 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Wavelet-based Bidirectional Propagation Fuse. Given the generated features $\\{X_{i}\\}_{i = 1}^{N}$ by the first Inverse Wavelet Transform (IwT), to explore useful contents from the features of other frames, exiting methods usually simply stack $\\{X_{i}\\}_{i = 1}^{N}$ or the alignment results of $\\{X_{i}\\}_{i = 1}^{N}$ . However, directly stacking frames without accurately estimating the features of $\\{X_{i}\\}_{i = 1}^{N}$ or the alignment results of $\\{X_{i}\\}_{i = 1}^{N}$ can only partially exploit the valuable information between frames and may also cause the propagation of erroneous information between frames. To further capture more useful long-range spatio-temporal information and reduce the influence of inaccurate information, we propose Wavelet-based Bidirectional Propagation Fuse (WBPF). As shown in Fig. 1, WBPF consists of a forward process and a backward process.", + "bbox": [ + 212, + 589, + 787, + 760 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We take the forward process as an example to introduce the WBPF module in detail, as shown in Fig. 3. First, we connect $Y_{i}$ and $X_{i + 1}$ to get $\\bar{X}$ , where $Y_{i}$ is the output of the previous forward process and $X_{i + 1}$ is the input of the current forward process. We then use a convolution layer with LeakyReLU to project $\\bar{X}$ into $\\hat{X}$ . Then we split $\\hat{X}$ into $\\hat{X}_1, \\hat{X}_2$ along the channel dimension. Next, we", + "bbox": [ + 212, + 763, + 789, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "C. Rao et al.", + "bbox": [ + 271, + 114, + 359, + 127 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "obtain the fused feature from $\\hat{X}_1, \\hat{X}_2$ by:", + "bbox": [ + 214, + 143, + 511, + 162 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\nF _ {p r o} = \\hat {X} _ {1} \\odot S \\left(W _ {1} \\hat {X} _ {1}\\right) + \\hat {X} _ {2} \\odot S \\left(W _ {2} \\hat {X} _ {2}\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 354, + 172, + 785, + 191 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $S$ denotes the Sigmoid function and $W$ is the $3 \\times 3$ 2D convolution. We utilize AvgPool, ResBlock, MaxPool, and Conv2D to better explore spatial information. The overall process can be described as follows:", + "bbox": [ + 214, + 200, + 785, + 247 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\bar {Y} _ {i + 1} = F _ {p r o} \\otimes \\text {u p s a m p l e} (P \\left(F _ {p r o}\\right)), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 377, + 258, + 785, + 277 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $\\mathrm{P}$ represents the stack of AvgPool, ResBlock, MaxPool, ResBlock and Conv2D. In the last, we employ 30 ResBlocks on $\\bar{Y}_{i + 1}$ to obtain $Y_{i + 1}$ .", + "bbox": [ + 214, + 287, + 785, + 319 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.3 Diffusion Model", + "text_level": 1, + "bbox": [ + 215, + 340, + 395, + 354 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Based on conditional denoising diffusion probabilistic models [11,27,31,35], the DM involves a forward diffusion process and a reverse denoising process. We first adopt the Latent Encoder (LE) on ground truth video and the Condition Encoder (CE) on blurry video to generate the ground truth feature $z \\in R^{T \\times 4C'}$ and the condition feature $c \\in R^{T \\times 4C'}$ . Then, we use the diffusion model to generate the prior feature $z' \\in R^{T \\times 4C'}$ using $c \\in R^{T \\times 4C'}$ as the condition and $z \\in R^{T \\times 4C'}$ as the target. We detail the diffusion process as follows.", + "bbox": [ + 214, + 364, + 787, + 470 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In the forward diffusion process, we transform $z$ into Gaussian noise by $T$ iterations, which can be described as follows:", + "bbox": [ + 214, + 472, + 785, + 501 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\nq \\left(z _ {t} \\mid z _ {t - 1}\\right) = \\mathcal {N} \\left(z _ {t}; \\sqrt {1 - \\beta_ {t}} z _ {t - 1}, \\beta_ {t} I\\right), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 367, + 512, + 785, + 531 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $t = 1,2,\\dots,T$ ; $z_0 = z$ ; $\\beta_{t} \\in (0,1)$ are hyperparameters that control the variance of the noise. Using the notation $\\alpha_{t} = 1 - \\beta_{t}$ , $\\bar{\\alpha}_{t} = \\prod_{i=1}^{t} \\alpha_{i}$ and through iterative derivation with reparameterization, Eq. (6) can be written as:", + "bbox": [ + 214, + 542, + 785, + 588 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\nq \\left(z _ {T} \\mid z _ {0}\\right) = \\mathcal {N} \\left(z _ {T}; \\sqrt {\\bar {\\alpha} _ {T}} z _ {0}, (1 - \\bar {\\alpha} _ {T}) I\\right), \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 367, + 599, + 785, + 617 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "that is", + "bbox": [ + 215, + 628, + 267, + 641 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\nz _ {T} = \\sqrt {\\bar {\\alpha} _ {T}} z _ {0} + \\sqrt {1 - \\bar {\\alpha} _ {T}} \\epsilon . \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 410, + 642, + 785, + 659 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In the reverse denoising process, DM sample a Gaussian random noise $z_{T}$ and then gradually denoise $z_{T}$ to $z_{0}$ :", + "bbox": [ + 214, + 666, + 785, + 696 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\nq \\left(z _ {t - 1} \\mid z _ {t}, z _ {0}\\right) = \\mathcal {N} \\left(z _ {t - 1}; \\mu_ {t} \\left(z _ {t}, z _ {0}\\right), \\tilde {\\beta} _ {t} I\\right), \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 362, + 708, + 785, + 724 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mu_ {t} \\left(z _ {t}, z _ {0}\\right) = \\frac {\\sqrt {\\bar {\\alpha} _ {t - 1}} \\beta_ {t}}{1 - \\bar {\\alpha} _ {t}} z _ {0} + \\frac {\\sqrt {\\alpha_ {t}} \\left(1 - \\bar {\\alpha} _ {t - 1}\\right)}{1 - \\bar {\\alpha} _ {t}} z _ {t}. \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 344, + 748, + 785, + 779 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We can use Eq. (8) to represent $z_0$ in Eq. (10) and obtain:", + "bbox": [ + 215, + 785, + 637, + 800 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mu_ {t} \\left(z _ {t}, z _ {0}\\right) = \\frac {1}{\\sqrt {\\alpha_ {t}}} \\left(z _ {t} - \\epsilon \\frac {1 - \\alpha_ {t}}{\\sqrt {1 - \\bar {\\alpha} _ {t}}}\\right). \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 380, + 811, + 785, + 844 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "Video Deblurring with Transformer and Diffusion Model", + "bbox": [ + 354, + 114, + 732, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 114, + 785, + 126 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We remove the variance estimation in Eq. (9):", + "bbox": [ + 215, + 146, + 550, + 162 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\nz _ {t - 1} = \\frac {1}{\\sqrt {\\alpha_ {t}}} (z _ {t} - \\epsilon \\frac {1 - \\alpha_ {t}}{\\sqrt {1 - \\bar {\\alpha} _ {t}}}). \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 398, + 172, + 785, + 205 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Then the noise $\\epsilon$ is the only uncertain variable. Following previous work [11, 31, 38], we adopt a neural network to predict the noise $\\epsilon$ on $z_{t}$ , $c$ and $t$ .", + "bbox": [ + 215, + 214, + 784, + 244 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "3.4 Training and Inference Strategies", + "text_level": 1, + "bbox": [ + 215, + 266, + 537, + 282 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Inspired by [5,38], we divide the training process into three stages: the first is to train the Latent Encoder with Wavelet-Aware Dynamic Transformer for ground truth feature extraction, the second is to train the Diffusion Model to generate the prior feature, and the third is to jointly train DM and WADT to generate artifact-free and high-realistic videos.", + "bbox": [ + 212, + 291, + 782, + 366 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Training Stage One. The first stage aims to train the Latent Encoder (LE) to extract valuable ground truth prior feature $z$ from ground truth (GT). We train LE and WADT as a whole model by inputting $z$ into WADTL instead of $z'$ and ignoring the condition encoder and the diffusion model in Fig. 1. We employ L1 loss and Multi-scale Frequency Reconstruction (MSFR) [10] loss. L1 loss is defined as:", + "bbox": [ + 214, + 366, + 787, + 454 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\nL _ {1} = \\sum_ {i = 1} ^ {N} \\| V _ {H Q} ^ {i} - V _ {G T} ^ {i} \\| _ {1}, \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 415, + 455, + 785, + 494 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $V_{HQ}^{i}$ is the $i$ -th frame of the deblurred HQ video and $V_{GT}^{i}$ is the $i$ -th frame of the GT video. The MSFR loss measures the L1 distance between multi-scale GT and deblurred frame in the frequency domain as follows:", + "bbox": [ + 214, + 501, + 787, + 547 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\nL _ {M S F R} = \\sum_ {i = 1} ^ {N} \\sum_ {k = 1} ^ {K} \\frac {1}{t _ {k}} \\| \\mathcal {F T} \\left(\\hat {S} _ {k} ^ {i}\\right) - \\mathcal {F T} \\left(S _ {k} ^ {i}\\right) \\| _ {1}, \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 346, + 556, + 785, + 599 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $\\mathcal{FT}$ denotes the Fast Fourier Transform that transfers image signal to the frequency domain, $K$ denotes the number of different scales and the $\\tilde{S}_k^i$ indicates the image signal of the $i$ -th frame at the $k$ -th scale. The final optimization function of stage one is determined as follows:", + "bbox": [ + 214, + 607, + 787, + 669 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\nL _ {d e b l u r} = L _ {1} + \\lambda L _ {M S F R}, \\tag {15}\n$$\n", + "text_format": "latex", + "bbox": [ + 411, + 681, + 785, + 698 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where we experimentally set $\\lambda = 0.1$ .", + "bbox": [ + 215, + 707, + 483, + 720 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Training Stage Two. In the second stage, we first use a Condition Encoder (CE), whose struct is the same as the Latent Encoder (LE), to generate condition feature $c$ from the blur video. Then, we train a diffusion model (DM) to generate prior feature $z'$ using condition feature $c$ .", + "bbox": [ + 214, + 722, + 782, + 781 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We run the complete $T$ iteration reverse processes Eq. (12) to generate prior feature $\\hat{z}$ . We denoted this loss as:", + "bbox": [ + 214, + 782, + 785, + 811 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\nL _ {d i f f} = \\left\\| z ^ {\\prime} - \\hat {z} \\right\\| _ {1}. \\tag {16}\n$$\n", + "text_format": "latex", + "bbox": [ + 436, + 824, + 785, + 842 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "C. Rao et al.", + "bbox": [ + 271, + 114, + 359, + 127 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/125b4a395c134fea66e38887fda90aaeaa3ae00e30870831ccb096055610a083.jpg", + "image_caption": [ + "DSTNet" + ], + "image_footnote": [], + "bbox": [ + 218, + 143, + 477, + 258 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/34467c743e513dbf72fb9fe5d571113f7994645057868fe5d75cf547f43893c8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 478, + 143, + 540, + 189 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/2e5151c314db403abf170db16f105bd4fc0f14c1f4e81a0dc99cbaae795173d9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 540, + 143, + 599, + 189 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/ae14aba8a6e4a35d2d5ca17b145376bb576b748578d7a67baa58ef7c5a0c6e6d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 602, + 143, + 663, + 189 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/2619cda6ee0e89f4ac8cd9e4d56599faf8712413b8a1786f16a1eebbd0b5a57a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 665, + 143, + 723, + 189 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/970968f7d221ab30225ed5ac812c905974a85ef5e93c54122643e10c0f18fbd6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 725, + 143, + 785, + 189 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/3f0a5bb7d668cdbf660b8244ca65c6571e7e0381b541a7e2ece03a8c76060c6c.jpg", + "image_caption": [ + "Blur" + ], + "image_footnote": [], + "bbox": [ + 478, + 200, + 540, + 246 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/776ad1b560723080c0b2ac9a379706ed3d3c561ac401e4a73d976954e232a9e3.jpg", + "image_caption": [ + "EDVR", + "FGST" + ], + "image_footnote": [], + "bbox": [ + 540, + 200, + 599, + 246 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/6b586679166e3f3733f95f917fed921e26e88f46c14f91601545e975a612b8f5.jpg", + "image_caption": [ + "MPRNet", + "DSTNet-L" + ], + "image_footnote": [], + "bbox": [ + 601, + 200, + 663, + 246 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/f95625964bf76fc6c1f8d6427bb9ede1fcde619c7ee0a690526481d968348ef1.jpg", + "image_caption": [ + "TSP", + "Ours" + ], + "image_footnote": [], + "bbox": [ + 665, + 200, + 723, + 246 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/5ee8b9e951f18e52fbe8fc3f6344e2f02942aab9bfb980ae3b0582688ac032f4.jpg", + "image_caption": [ + "SFE", + "GT" + ], + "image_footnote": [], + "bbox": [ + 725, + 200, + 785, + 246 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/b32b1633ef79986b5c7a045015039ba0711f7a8f60abbcc0ef6b3912412a0548.jpg", + "image_caption": [ + "Fig. 4: Visual comparison on GoPro [25] dataset. The deblurred results of previous work still contain significant blur effects. Our method generates much clearer frames.", + "Fig. 5: Visual comparison on DVD [33] dataset. The deblurring effect of our proposed model is significantly better." + ], + "image_footnote": [], + "bbox": [ + 218, + 313, + 477, + 426 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/2fa8e5bed21ca76f4140a3515925d540d5e4b063f5152035b79cec00903bb7ec.jpg", + "image_caption": [ + "Blur" + ], + "image_footnote": [], + "bbox": [ + 478, + 313, + 539, + 359 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/50c4fcbefc9a5f8999bdf252c08062a8d948324d4f811ff9b34fa2819e49bb5d.jpg", + "image_caption": [ + "EDVR" + ], + "image_footnote": [], + "bbox": [ + 540, + 313, + 599, + 359 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/7eba2eb8ae04825a2284d4718e4fc2cf449945ff8c3e55dc1f62c8f1ab50d424.jpg", + "image_caption": [ + "MPRNet" + ], + "image_footnote": [], + "bbox": [ + 601, + 313, + 661, + 359 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/62f0c44f826e22d1e35ee7d785eaa1aaee51c73a570161751696ac0e0e787cb6.jpg", + "image_caption": [ + "TSP" + ], + "image_footnote": [], + "bbox": [ + 663, + 313, + 723, + 359 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/1cf4cc5f99b5858eeec3b889c7f5f2eac0354255fdaaf2ea55fa975ddabfafaa.jpg", + "image_caption": [ + "SFE" + ], + "image_footnote": [], + "bbox": [ + 723, + 313, + 782, + 359 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/238a4422bdbd826570a704a88c36ad75f6fc3b5a0454ee4a493fd7e7276b62aa.jpg", + "image_caption": [ + "BasicVSR" + ], + "image_footnote": [], + "bbox": [ + 478, + 371, + 540, + 416 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/b64927086d94787828917df63b393568bb79074df8fbdb0af4f0d8ad26502add.jpg", + "image_caption": [ + "FGST" + ], + "image_footnote": [], + "bbox": [ + 540, + 371, + 599, + 416 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/84bf2448fa6734706b9e366fa6f046c00cf374a2e6572f219aaf636ab33b4bf5.jpg", + "image_caption": [ + "DSTNet" + ], + "image_footnote": [], + "bbox": [ + 601, + 371, + 661, + 416 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/0c03ceb7eb36a91a384be2137c5aa20158089bd4df25fd48a384d4a497ce18ee.jpg", + "image_caption": [ + "Ours" + ], + "image_footnote": [], + "bbox": [ + 663, + 371, + 723, + 416 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/64a517332c6fb4ed9c341c121ea99be52dbde3531a7da0bdd9847297e63e33fb.jpg", + "image_caption": [ + "GT" + ], + "image_footnote": [], + "bbox": [ + 723, + 371, + 782, + 416 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Training Stage Three. In Training Stage Two, we train the DM to generate $z'$ to approximate $z$ . However, there inevitably is a slight deviation between the prior features $z'$ and the ground truth prior feature $z$ . To make the DM and WADT work better together to acquire information from different frequency domains for video deblurring and improve the robustness of WADT to the prior feature $z'$ , we jointly train DM and WADT in Training Stage Three. The training loss $L_{total}$ is given by the sum of the deblurring loss $L_{deblur}$ and the diffusion loss $L_{diff}$ :", + "bbox": [ + 212, + 500, + 787, + 619 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\nL _ {t o t a l} = L _ {d e b l u r} + L _ {d i f f}. \\tag {17}\n$$\n", + "text_format": "latex", + "bbox": [ + 411, + 622, + 784, + 638 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Inference. Given a blurry input video $V_{blur} \\in R^{T \\times 3 \\times H \\times W}$ , we first compress $V_{blur}$ into a condition feature $c \\in R^{T \\times 4C'}$ . Then, the diffusion model generates the prior feature $z'$ from pure Gaussian noise $\\epsilon$ and condition feature $c$ . Finally, VD-diff deblurs the blurry video with $z'$ using WADT.", + "bbox": [ + 214, + 645, + 787, + 705 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 214, + 733, + 375, + 750 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.1 Datasets", + "text_level": 1, + "bbox": [ + 214, + 766, + 333, + 780 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "GoPro. The GoPro [25] benchmark is composed of over 3,300 blurry-sharp image pairs of dynamic scenes. It is obtained by a high-speed camera. The training and testing subsets are split in proportional to 2:1.", + "bbox": [ + 212, + 794, + 784, + 839 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "Video Deblurring with Transformer and Diffusion Model", + "bbox": [ + 354, + 114, + 732, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/e777d341f88b32ac1ab52d0d52bd2067e78c0df1076a2a82e56c3d7d9625ad22.jpg", + "image_caption": [ + "Blur" + ], + "image_footnote": [], + "bbox": [ + 220, + 143, + 357, + 242 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/4a4c941a0566c193cd96daabeb3ce965bcd00b9991475b4372680fae7a5b8b41.jpg", + "image_caption": [ + "DVD [33]" + ], + "image_footnote": [], + "bbox": [ + 357, + 143, + 495, + 242 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/7c24433315b348357e2485f62f59b59cc6f8a169698852f57dd758ef6c8b3346.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 495, + 143, + 630, + 242 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/865029901e1066fc4bdfe4f3d124c0f9b2632a3a30f655f0328800d39e4544cf.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 630, + 143, + 774, + 242 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/efbac54d69d8f13105e14fb834623f22042b2e63ab5a111fd257ce32e3572668.jpg", + "image_caption": [ + "TSP [28]" + ], + "image_footnote": [], + "bbox": [ + 220, + 255, + 357, + 352 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/22200fbd5a291171e3eea3f76e1214106d8eb8621ce6d1bd46a69678b4418b35.jpg", + "image_caption": [ + "SFE [39]" + ], + "image_footnote": [], + "bbox": [ + 357, + 255, + 495, + 352 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/874ecc640e9f669020c65163817202607f6889547c22c72be6580916f9d30b7e.jpg", + "image_caption": [ + "EDVR [36]", + "DSTNet [29]" + ], + "image_footnote": [], + "bbox": [ + 495, + 255, + 630, + 352 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/9ea57fc86eb8151b6f223207f911492332e77b338abb92c4d5161a351d8f43b0.jpg", + "image_caption": [ + "MPRNet [41]", + "VD-Diff (Ours)" + ], + "image_footnote": [], + "bbox": [ + 630, + 255, + 774, + 352 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/d32b83ab696924aaf341a27841053d837d771479a2920fa3fb7387c23f9fc179.jpg", + "image_caption": [ + "Fig. 6: Visual results of VD-Diff and SOTA methods on the real blurry videos of [33]." + ], + "image_footnote": [], + "bbox": [ + 225, + 407, + 390, + 513 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/0decf592cd7d2adbf19f250e3939315c0903a7a7a0679d5242b7fb1c886946d0.jpg", + "image_caption": [ + "Fig. 7: Comparison of various methods in terms of time dimension." + ], + "image_footnote": [], + "bbox": [ + 401, + 407, + 457, + 515 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/320c671e9a598d4221bd5131164e7e3a2c5f5fcae13092003561b5688d37e596.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 468, + 410, + 522, + 515 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/28979012f28514e8b93d683d295c3b0a5187c8a252bcee1b5f9580c7e12dfc2d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 532, + 410, + 584, + 515 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/2d9c6f5184ee7c681a4596d8fd8813f4bd2bd35422a1c6b5b936f3e5593312d4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 598, + 410, + 651, + 515 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/16f6d5ef7b030d8ae96258ef62e14ca97eec7c053fa8d2133cb070b20cd0a933.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 661, + 410, + 714, + 515 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/252d7ae0a05d0cc82d61f707d3fd8a092218923bc5775541698e2455b9267448.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 725, + 411, + 779, + 515 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "DVD. The DVD [33] dataset consists of 71 videos with 6,708 blurry-sharp image pairs. It is divided into train/test subsets with 61 videos (5,708 image pairs) and 10 videos (1,000 image pairs). DVD is captured with mobile phones and DSLR at a frame rate of 240 FPS.", + "bbox": [ + 212, + 582, + 787, + 642 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "BSD. The Beam-Splitter Deblurring Dataset (BSD) [46] consists of 240 videos with 24k blurry-sharp image pairs. Cross-validation experiments between real-world and synthetic datasets demonstrate the high generality of the BSD dataset.", + "bbox": [ + 212, + 643, + 787, + 688 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Real-World Videos. We evaluate the generality of our proposed approach on the real blurry datasets [6]. Due to the inaccessibility of the ground truth, we only compare VD-Diff with other models on visual results.", + "bbox": [ + 212, + 688, + 787, + 734 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.2 Implementation Details", + "text_level": 1, + "bbox": [ + 214, + 755, + 457, + 770 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "We implement VD-Diff in PyTorch [30] using a single NVIDIA RTX A6000 GPU. The training process takes 7 days and 10 hours. We set the batch size as 2, the learning rate as $8 \\times 10^{-5}$ and weight decay as $4 \\times 10^{-5}$ . Patches at the size of $256 \\times 256$ cropped from training frames are fed into the models. The sequence", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "C. Rao et al.", + "bbox": [ + 271, + 114, + 359, + 127 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/273968429ab4dcf48ba4a49d95efa3888f1806a535940052610198baa32abfd9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 215, + 143, + 352, + 247 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/663957c0f29f7d6cdb5aaa055089c788488e9cef9219fb30623a99c8ccd142b8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 361, + 143, + 496, + 247 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/5e6d92ab721868fb6249f25e1f95cb4cd8f9920452d84f899096c0c781421582.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 506, + 143, + 640, + 247 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/b35cde1bf0a75cb1339908d4d57e6ae57b2040669040772fc244c660a9cfd57a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 651, + 143, + 785, + 247 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/9e6570368fd84698e1d6ad41df0389c40af7bd47748e59e404739d764f83c3bb.jpg", + "image_caption": [ + "Blur" + ], + "image_footnote": [], + "bbox": [ + 215, + 252, + 351, + 354 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/9ba28251ab1265077dbc045c9ebe82331ceb5c696f4905d5bcf1f08d6daae92b.jpg", + "image_caption": [ + "LDM" + ], + "image_footnote": [], + "bbox": [ + 361, + 252, + 496, + 354 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/4c6262d3caf1ae557e900ac308d766f1f9489b47fad83881fad34439d4c028f1.jpg", + "image_caption": [ + "Ours" + ], + "image_footnote": [], + "bbox": [ + 506, + 252, + 640, + 354 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/14ff7918093f633e6c43de34e33771634f717c61b9c391c681af4640d0b17bfc.jpg", + "image_caption": [ + "GT", + "Fig. 8: Qualitative comparison with pre-trained large diffusion models [31]." + ], + "image_footnote": [], + "bbox": [ + 651, + 252, + 785, + 354 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "length is set to 15 in training and 16 in testing. Since we use diffusion within a highly compressed latent space, the denoising iteration steps $T$ only needs 4. We set $C$ to 64 and $C'$ to 256, respectively. In the WADT module, we set $N_1$ and $N_2$ to 7 and 2, respectively, to balance the calculation amount and the deblurring effect. The horizontal and vertical flips are performed for data augmentation.", + "bbox": [ + 212, + 429, + 787, + 506 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4.3 Quantitative Results", + "text_level": 1, + "bbox": [ + 215, + 527, + 433, + 541 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "We compare our proposed VD-Diff with other SOTA methods, as shown in Tables 1, 2, and 3. As can be seen, our method outperforms SOTA methods on GoPro [25] and DVD [33] benchmark datasets. Specifically, as shown in Table 1, Our method outperforms DSTNet-L, RVRT, and FGST by $0.50\\mathrm{dB}$ , $0.63\\mathrm{dB}$ , and $2.65\\mathrm{dB}$ in PSNR on the GoPro [25] dataset, respectively. Our model achieves SOTA performance on the GoPro [25] dataset. Table 2 shows that our proposed method generates the deblurred videos with higher PSNR and SSIM values on the DVD [33] dataset. As shown in Table 3, we also achieve SOTA on the BSD [46] dataset. As seen from Table 4, our model has a relatively low parameter size and less inference time than other models. Although our model's parameter size and inference time are slightly higher than DSTNet and BasicVSR++, the PSNR metric on the GoPro [25] dataset is at least 1.3dB higher than the two models. As shown in Fig. 9(a), our model achieves favorable results in terms of accuracy and FLOPs. The above results show the effectiveness of our model.", + "bbox": [ + 212, + 551, + 787, + 763 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4.4 Qualitative Results", + "text_level": 1, + "bbox": [ + 215, + 784, + 419, + 799 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "We provide visual comparisons on GoPro [25], DVD [33], and Real-World Video datasets [33] as shown in Fig. 4, 5, and 6, respectively. Some models sacrifice", + "bbox": [ + 212, + 809, + 785, + 839 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "Video Deblurring with Transformer and Diffusion Model", + "bbox": [ + 354, + 114, + 730, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 116, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/ab5e5d3e94321f0481c888f4918d56f8f043bcb229032996891a623592a668d9.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodDVD [33]EDVR [36]STFAN [47]ESTRNN [46]MPRNet [41]TSP [28]
PSNR ↑27.3126.8328.5931.0732.7331.67
SSIM ↑0.82550.84260.86080.90230.93660.9279
BasicVSR [3]NAFNet [4]FGST [22]DSTNet [29]DSTNet-L [29]RVRT [21]Ours
34.0133.7132.9034.1635.0534.9235.55
0.95200.96680.96100.96790.97330.97380.9752
", + "bbox": [ + 212, + 143, + 774, + 234 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/26dc8ad5c84bb2e3e4996b6700fb8ea0dfdd983c6feebe00e2186626619faba4.jpg", + "table_caption": [ + "Table 1: VD-Diff achieves SOTA results on the GoPro dataset [25]." + ], + "table_footnote": [], + "table_body": "
MethodDVD [33]EDVR [36]STFAN [47]ESTRNN [46]TSP [28]SEF [39]
PSNR ↑30.0128.5131.1532.0132.1331.71
SSIM ↑0.88770.86370.90490.91620.92680.9160
GSTA [34]MPRNet [41]ARVo [16]FGST [22]DSTNet [29]RVRT [21]Ours
32.5332.7332.8033.3633.7934.3034.56
0.94680.93660.93520.95000.96150.96550.9684
", + "bbox": [ + 236, + 268, + 766, + 361 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Table 2: VD-Diff achieves SOTA results on the DVD dataset [33].", + "bbox": [ + 276, + 361, + 723, + 375 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "texture and details to maintain overall structure, resulting in over-smooth videos that lose high-frequency detail. Other models ignore global information to preserve high-frequency detail, introducing redundant speckle textures and artifacts that do not match the real video in fast-motion video. In contrast, our VD-Diff has the best visual effect as it can leverage WADT to model long-range dependencies to capture low-frequency information and employ DM to supplement valuable high-frequency information. Therefore, as seen from the visual comparison, our model can restore texture and details while maintaining the overall framework of the video.", + "bbox": [ + 212, + 411, + 784, + 545 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Besides, we conduct a qualitative comparison in terms of the time dimension, as shown in Fig. 7. As can be seen, previous models such as EDVR [36], MPRNet [41], TSP [28], and DSTNet [29] all exhibit unclear outlines, indicating that they cannot guarantee temporal consistency. In contrast, our method effectively maintains temporal consistency.", + "bbox": [ + 212, + 547, + 784, + 622 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Furthermore, we conduct a comparison with the direct application of a pretrained large diffusion model (i.e., LDM [31]) to video deblurring, as shown in Fig. 8. We notice that the videos deblurred by LDM contain a significant number of artifacts and inconsistencies with the content of the original videos, which is attributed to motion blur present in the blurred videos. In contrast, our method leverages the diffusion model and WADT to maximally restore the details in the video while preserving the video content.", + "bbox": [ + 212, + 623, + 785, + 729 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "4.5 Ablation Study", + "text_level": 1, + "bbox": [ + 215, + 752, + 390, + 768 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "This section explores the significance of each key component of our proposed VD-Diff. All variants are tested on the GoPro [25] dataset.", + "bbox": [ + 212, + 779, + 784, + 808 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Effect of Diffusion Model. To validate the effectiveness of the diffusion model, we design a variant that does not utilize DM to generate prior features", + "bbox": [ + 212, + 809, + 785, + 839 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "C. Rao et al.", + "bbox": [ + 271, + 114, + 359, + 127 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/4b627d2a73a995ae0d8944bd2aab3dac29bbeee15c77d5c8acb9b42722c42db6.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodDVD [33]DTBN [14]Nah et al. [26]ESTRNN [46]TSP [28]DSTNet [29]Ours
PSNR ↑29.9531.8433.0033.3632.8434.4535.61
SSIM ↑0.86920.91700.93300.93700.93980.95480.9758
", + "bbox": [ + 220, + 145, + 789, + 191 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/5c2dd916ad8719c9f00db4d698f504f18f416c29728a9aac1a5befb8a8f0df27.jpg", + "table_caption": [ + "Table 3: VD-Diff achieves SOTA results on the BSD dataset [46]." + ], + "table_footnote": [], + "table_body": "
MethodDVDEDVRTSPFGSTDSTNetBasicVSR++RVRTDSTNet-LOurs
Parameters (M)15.3023.6016.199.707.459.4613.6016.9611.93
Running time (s)0.882.701.900.720.080.100.410.280.31
", + "bbox": [ + 220, + 227, + 802, + 275 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/73dd92f7f128c44514264e7d95dad46075ceab3b61e05180575544a831cab57c.jpg", + "table_caption": [ + "Table 4: Quantitative evaluations of the video deblurring methods regarding network parameters and running time. The running time is obtained on a machine with a single NVIDIA RTX A6000 GPU. The size of the test images is $1280 \\times 720$ pixels." + ], + "table_footnote": [], + "table_body": "
VariantsComponentsMetrics
Diffusion ModelWaveletJoint TrainingWBPFPSNR ↑SSIM ↑
w/o DMX✓✓✓34.380.9684
w/o Wavelet✓X✓✓34.750.9709
w/o Joint Training✓✓X✓34.920.9717
w/o WBPF✓✓✓X33.580.9344
Full model✓✓✓✓35.550.9752
", + "bbox": [ + 232, + 358, + 769, + 455 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Table 5: Ablation experiments on various variants of VD-Diif on the GoPro [25] dataset, and the best quantitative results are marked in bold.", + "bbox": [ + 215, + 455, + 782, + 483 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/516623607e2afb29e59b91eac0e5b7b0288129edcf7c9d7e5a21e6b270f36bfd.jpg", + "image_caption": [ + "Fig. 9: (a) Floating point operations (FLOPs) vs video deblurring performance on the GoPro [25] dataset. Our model achieves favorable results in terms of accuracy and FLOPs. (b) Ablation study of the number of iteration steps. (c) Ablation study of the sequence length during inference." + ], + "image_footnote": [], + "bbox": [ + 218, + 513, + 408, + 633 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/edec444189fc7d2e64b0ffbcbb461bbf0efd8f498909e5ca7ac29d335e4835c8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 410, + 513, + 593, + 630 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/70fcc8cfaa5e45b0bc13359405234be3819008fba7d5859e161ed4bc475d12d1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 596, + 513, + 782, + 633 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "but solely relies on WADT for video deblurring, named as $w / o$ DM. As seen in Table 5, without employing DM, the deblurring performance of the network significantly declines, with a decrease of 1.17 in PSNR, which indicates that DM can provide valuable prior features to enhance the deblurring performance.", + "bbox": [ + 212, + 733, + 782, + 792 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Effect of Wavelet Transform. We conduct ablation experiments to study the impact of wavelet transform on VD-Diff. We replace wavelet transform and inverse wavelet transform with bilinear downsample and upsample to evaluate", + "bbox": [ + 212, + 794, + 784, + 839 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "Video Deblurring with Transformer and Diffusion Model", + "bbox": [ + 354, + 114, + 730, + 127 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 784, + 126 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "the effect of wavelet transform, named $w / o$ Wavelet. We adopt the same training strategy as VD-Diff. Table 5 shows $w / o$ Wavelet decreases by 0.8dB regarding the PSNR metrics. This experiment shows the wavelet transform method for feature extraction is better than the general upsampling and downsampling methods.", + "bbox": [ + 212, + 146, + 782, + 205 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Effect of Joint Training Strategy. In Sec. 3.4, we analyze the importance of joint training. To verify our analysis, we remove joint training stage i.e., Training Stage Three, as shown in Table 5. We notice that VD-Diff is significantly better than $w / o$ Joint Training with an improvement of 0.63dB in terms of PSNR metrics, which demonstrates that joint training can improve the collaborative working ability of DM and WADT, thereby improving the deblurring effect.", + "bbox": [ + 212, + 208, + 784, + 297 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Effect of Wavelet-based Bidirectional Propagation Fuse. To investigate the effect of Wavelet-based Bidirectional Propagation Fuse (WBPF), we design a variant model by removing the WBPF in WADT, termed as $w / o$ WBPF. Table 5 shows that WBPF can effectively transmit useful information between frames, improving the model's ability to capture spatio-temporal information.", + "bbox": [ + 212, + 300, + 784, + 375 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Effect of Iteration Steps. To explore the impact of the number of iterations on the diffusion model, we set different iteration steps to train the model and then compare the PSNR metrics on the GoPro dataset. The results are shown in Fig. 9(b). The value of PSNR increases as the number of iteration steps increases, but when the number of iterations is greater than 4, the value of PSNR almost remains stable. It shows that in a highly compressed latent space, just setting the number of iteration steps to 4 can generate suitable prior features.", + "bbox": [ + 212, + 377, + 784, + 482 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Effect of Sequence Length. Intuitively, the sequence length will affect the deblurring effect of our proposed model. If the sequence length is too short, the model fails to capture essential information from distant frames. Conversely, if the sequence length is too long, the computational load becomes excessive, potentially causing error accumulation. To investigate this, we experimented with various sequence lengths. As shown in Fig. 9(c), the PSNR value increases steadily with the number of frames. However, after 16 frames, the calculation becomes very large as the sequence length increases, and the PSNR does not change significantly. Therefore, we set the sequence length to 16 to reach a near-optimal point between the deblurring effect and computational reduction.", + "bbox": [ + 212, + 484, + 784, + 633 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 665, + 357, + 680 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In this paper, we propose a brand-new method that innovatively integrates the diffusion model with the Transformer for video deblurring. Specifically, the diffusion model generates prior features containing high-frequency information that conforms to the ground truth (GT) distribution. Meanwhile, the Wavelet-Aware Dynamic Transformer captures low-frequency information and utilizes the prior features generated by the diffusion model to supplement high-frequency information, thereby enhancing the quality of the deblurred video. Comprehensive experimental results show that our proposed method outperforms the state-of-the-art methods.", + "bbox": [ + 212, + 704, + 782, + 838 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "C. Rao et al.", + "bbox": [ + 271, + 114, + 357, + 126 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 217, + 143, + 401, + 162 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "This work was supported in part by National Social Science Foundation Major Project \"Research on Virtual Restoration of Tang and Song Painting Colors and Construction of Traditional Color Resource Library from a Digital Perspective\" (19ZDA046), Zhejiang Province Program (2022C01222, 2023C03199, 2023C03201), the National Program of China (62172365, 2021YFF0900604, 19ZDA197), Ningbo Science and Technology Plan Project (2022Z167, 2023Z137), and MOE Frontier Science Center for Brain Science & Brain-Machine Integration (Zhejiang University).", + "bbox": [ + 212, + 174, + 787, + 297 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 316, + 321, + 332 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Ba, J.L., Kiros, J.R., Hinton, G.E.: Layer normalization. arXiv preprint arXiv:1607.06450 (2016)", + "2. Blattmann, A., Rombach, R., Ling, H., Dockhorn, T., Kim, S.W., Fidler, S., Kreis, K.: Align your latents: High-resolution video synthesis with latent diffusion models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 22563-22575 (2023)", + "3. Chan, K.C., Zhou, S., Xu, X., Loy, C.C.: On the generalization of basicvsr++ to video deblurring and denoising. arXiv preprint arXiv:2204.05308 (2022)", + "4. Chen, L., Chu, X., Zhang, X., Sun, J.: Simple baselines for image restoration. In: European Conference on Computer Vision. pp. 17-33. Springer (2022)", + "5. Chen, Z., Zhang, Y., Liu, D., Xia, B., Gu, J., Kong, L., Yuan, X.: Hierarchical integration diffusion model for realistic image deblurring. arXiv preprint arXiv:2305.12966 (2023)", + "6. Cho, S., Wang, J., Lee, S.: Video deblurring for hand-held cameras using patch-based synthesis. ACM Transactions on Graphics (TOG) 31(4), 1-9 (2012)", + "7. Chu, T., Chen, J., Sun, J., Lian, S., Wang, Z., Zuo, Z., Zhao, L., Xing, W., Lu, D.: Rethinking fast fourier convolution in image inpainting. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 23195-23205 (2023)", + "8. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929 (2020)", + "9. Esser, P., Chiu, J., Atighehchian, P., Granskog, J., Germanidis, A.: Structure and content-guided video synthesis with diffusion models. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 7346-7356 (2023)", + "10. Fan, Y., Hong, C., Wang, X., Zeng, Z., Guo, Z.: Multi-input-output fusion attention module for deblurring networks. In: 2021 IEEE international conference on big data (Big Data). pp. 3176-3182. IEEE (2021)", + "11. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. Advances in neural information processing systems 33, 6840-6851 (2020)", + "12. Huang, H., He, R., Sun, Z., Tan, T.: Wavelet-srnet: A wavelet-based cnn for multiscale face super resolution. In: Proceedings of the IEEE international conference on computer vision. pp. 1689-1697 (2017)", + "13. Hyun Kim, T., Mu Lee, K.: Generalized video deblurring for dynamic scenes. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 5426-5434 (2015)" + ], + "bbox": [ + 218, + 347, + 784, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "Video Deblurring with Transformer and Diffusion Model", + "bbox": [ + 354, + 114, + 730, + 127 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "14. Hyun Kim, T., Mu Lee, K., Scholkopf, B., Hirsch, M.: Online video deblurring via dynamic temporal blending network. In: Proceedings of the IEEE international conference on computer vision. pp. 4038-4047 (2017)", + "15. Li, A., Zhao, L., Zuo, Z., Wang, Z., Xing, W., Lu, D.: Migt: Multi-modal image inpainting guided with text. Neurocomputing 520, 376-385 (2023)", + "16. Li, D., Xu, C., Zhang, K., Yu, X., Zhong, Y., Ren, W., Suominen, H., Li, H.: Arvo: Learning all-range volumetric correspondence for video deblurring. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7721-7731 (2021)", + "17. Li, G., Lyu, J., Wang, C., Dou, Q., Qin, J.: Wavtrans: Synergizing wavelet and cross-attention transformer for multi-contrast mri super-resolution. In: International Conference on Medical Image Computing and Computer-Assisted Intervention. pp. 463–473. Springer (2022)", + "18. Li, G., Rao, C., Mo, J., Zhang, Z., Xing, W., Zhao, L.: Rethinking diffusion model for multi-contrast mri super-resolution. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11365-11374 (2024)", + "19. Li, W., Zhao, L., Lin, Z., Xu, D., Lu, D.: Non-local image inpainting using low-rank matrix completion. In: Computer Graphics Forum. vol. 34, pp. 111-122. Wiley Online Library (2015)", + "20. Li, W., Zhao, L., Xu, D., Lu, D.: Efficient image completion method based on alternating direction theory. In: 2013 IEEE International Conference on Image Processing. pp. 700-703. IEEE (2013)", + "21. Liang, J., Fan, Y., Xiang, X., Ranjan, R., Ilg, E., Green, S., Cao, J., Zhang, K., Timofte, R., Gool, L.V.: Recurrent video restoration transformer with guided deformable attention. Advances in Neural Information Processing Systems 35, 378-393 (2022)", + "22. Lin, J., Cai, Y., Hu, X., Wang, H., Yan, Y., Zou, X., Ding, H., Zhang, Y., Timofte, R., Van Gool, L.: Flow-guided sparse transformer for video deblurring. arXiv preprint arXiv:2201.01893 (2022)", + "23. Liu, X., Park, D.H., Azadi, S., Zhang, G., Chopikyan, A., Hu, Y., Shi, H., Rohrbach, A., Darrell, T.: More control for free! image synthesis with semantic diffusion guidance. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV). pp. 289-299 (January 2023)", + "24. Luo, Z., Chen, D., Zhang, Y., Huang, Y., Wang, L., Shen, Y., Zhao, D., Zhou, J., Tan, T.: Videofusion: Decomposed diffusion models for high-quality video generation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10209-10218 (2023)", + "25. Nah, S., Hyun Kim, T., Mu Lee, K.: Deep multi-scale convolutional neural network for dynamic scene deblurring. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 3883-3891 (2017)", + "26. Nah, S., Son, S., Lee, K.M.: Recurrent neural networks with intra-frame iterations for video deblurring. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 8102-8111 (2019)", + "27. Ozdenizci, O., Legenstein, R.: Restoring vision in adverse weather conditions with patch-based denoising diffusion models. IEEE Transactions on Pattern Analysis and Machine Intelligence (2023)", + "28. Pan, J., Bai, H., Tang, J.: Cascaded deep video deblurring using temporal sharpness prior. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 3043-3051 (2020)" + ], + "bbox": [ + 215, + 146, + 785, + 840 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "C. Rao et al.", + "bbox": [ + 271, + 114, + 357, + 126 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "29. Pan, J., Xu, B., Dong, J., Ge, J., Tang, J.: Deep discriminative spatial and temporal network for efficient video deblurring. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 22191-22200 (2023)", + "30. Paszke, A., Gross, S., Massa, F., Lerer, A., Bradbury, J., Chanan, G., Killeen, T., Lin, Z., Gimelshein, N., Antiga, L., et al.: Pytorch: An imperative style, high-performance deep learning library. Advances in neural information processing systems 32 (2019)", + "31. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10684-10695 (2022)", + "32. Ruan, L., Ma, Y., Yang, H., He, H., Liu, B., Fu, J., Yuan, N.J., Jin, Q., Guo, B.: Mm-diffusion: Learning multi-modal diffusion models for joint audio and video generation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10219-10228 (2023)", + "33. Su, S., Delbracio, M., Wang, J., Sapiro, G., Heidrich, W., Wang, O.: Deep video deblurring for hand-held cameras. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1279-1288 (2017)", + "34. Suin, M., Rajagopalan, A.: Gated spatio-temporal attention-guided video deblurring. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7802-7811 (2021)", + "35. Tashiro, Y., Song, J., Song, Y., Ermon, S.: Csdi: Conditional score-based diffusion models for probabilistic time series imputation. Advances in Neural Information Processing Systems 34, 24804-24816 (2021)", + "36. Wang, X., Chan, K.C., Yu, K., Dong, C., Change Loy, C.: Edvr: Video restoration with enhanced deformable convolutional networks. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops. pp. 0-0 (2019)", + "37. Wang, Z., Cun, X., Bao, J., Zhou, W., Liu, J., Li, H.: Uformer: A general u-shaped transformer for image restoration. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 17683-17693 (2022)", + "38. Xia, B., Zhang, Y., Wang, S., Wang, Y., Wu, X., Tian, Y., Yang, W., Van Gool, L.: Diffir: Efficient diffusion model for image restoration. arXiv preprint arXiv:2303.09472 (2023)", + "39. Xiang, X., Wei, H., Pan, J.: Deep video deblurring using sharpness features from exemplars. IEEE Transactions on Image Processing 29, 8976-8987 (2020)", + "40. Yu, S., Sohn, K., Kim, S., Shin, J.: Video probabilistic diffusion models in projected latent space. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 18456-18466 (June 2023)", + "41. Zamir, S.W., Arora, A., Khan, S., Hayat, M., Khan, F.S., Yang, M.H., Shao, L.: Multi-stage progressive image restoration. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 14821-14831 (2021)", + "42. Zhang, K., Luo, W., Zhong, Y., Ma, L., Liu, W., Li, H.: Adversarial spatio-temporal learning for video deblurring. IEEE Transactions on Image Processing 28(1), 291-301 (2018)", + "43. Zhang, Z., Zhang, Q., Xing, W., Li, G., Zhao, L., Sun, J., Lan, Z., Luan, J., Huang, Y., Lin, H.: Artbank: Artistic style transfer with pre-trained diffusion model and implicit style prompt bank. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 38, pp. 7396-7404 (2024)", + "44. Zhao, L., Mo, Q., Lin, S., Wang, Z., Zuo, Z., Chen, H., Xing, W., Lu, D.: Uctgan: Diverse image inpainting based on unsupervised cross-space translation. In: Pro" + ], + "bbox": [ + 212, + 146, + 787, + 839 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "Video Deblurring with Transformer and Diffusion Model", + "bbox": [ + 354, + 114, + 730, + 127 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "ceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 5741-5750 (2020)", + "45. Zheng, B., Yuan, S., Slabaugh, G., Leonardis, A.: Image demoireing with learnable bandpass filters. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3636-3645 (2020)", + "46. Zhong, Z., Gao, Y., Zheng, Y., Zheng, B.: Efficient spatio-temporal recurrent neural network for video deblurring. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part VI 16. pp. 191-207. Springer (2020)", + "47. Zhou, S., Zhang, J., Pan, J., Xie, H., Zuo, W., Ren, J.: Spatio-temporal filter adaptive network for video deblurring. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 2482-2491 (2019)", + "48. Zuo, Z., Zhao, L., Li, A., Wang, Z., Zhang, Z., Chen, J., Xing, W., Lu, D.: Generative image inpainting with segmentation confusion adversarial training and contrastive learning. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 37, pp. 3888-3896 (2023)" + ], + "bbox": [ + 217, + 147, + 785, + 369 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "C.Rao et al.", + "bbox": [ + 271, + 114, + 357, + 126 + ], + "page_idx": 17 + } +] \ No newline at end of file diff --git a/2024/Rethinking Video Deblurring with Wavelet-Aware Dynamic Transformer and Diffusion Model/7afdf756-63a6-48fd-b654-3bb54bd5d976_model.json b/2024/Rethinking Video Deblurring with Wavelet-Aware Dynamic Transformer and Diffusion Model/7afdf756-63a6-48fd-b654-3bb54bd5d976_model.json new file mode 100644 index 0000000000000000000000000000000000000000..d038393d587511c8ec7d1fd94436f31da1822cbb --- /dev/null +++ b/2024/Rethinking Video Deblurring with Wavelet-Aware Dynamic Transformer and Diffusion Model/7afdf756-63a6-48fd-b654-3bb54bd5d976_model.json @@ -0,0 +1,3437 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.218, + 0.141, + 0.787, + 0.187 + ], + "angle": 0, + "content": "Rethinking Video Deblurring with Wavelet-Aware Dynamic Transformer and Diffusion Model" + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.212, + 0.781, + 0.259 + ], + "angle": 0, + "content": "Chen Rao\\(^{1\\dagger}\\), Guangyuan Li\\(^{1\\dagger}\\), Zehua Lan\\(^{1}\\), Jiakai Sun\\(^{1}\\), Junsheng Luan\\(^{1}\\), Wei Xing\\(^{1*}\\), Lei Zhao\\(^{1*}\\), Huaizhong Lin\\(^{1*}\\), Jianfeng Dong\\(^{2}\\), and Dalong Zhang\\(^{1}\\)" + }, + { + "type": "text", + "bbox": [ + 0.274, + 0.269, + 0.727, + 0.283 + ], + "angle": 0, + "content": "1 College of Computer Science and Technology, Zhejiang University" + }, + { + "type": "text", + "bbox": [ + 0.389, + 0.284, + 0.613, + 0.298 + ], + "angle": 0, + "content": "\\(^{2}\\) Zhejiang Gongshang University" + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.336, + 0.744, + 0.626 + ], + "angle": 0, + "content": "Abstract. Current video deblurring methods have limitations in recovering high-frequency information since the regression losses are conservative with high-frequency details. Since Diffusion Models (DMs) have strong capabilities in generating high-frequency details, we consider introducing DMs into the video deblurring task. However, we found that directly applying DMs to the video deblurring task has the following problems: (1) DMs require many iteration steps to generate videos from Gaussian noise, which consumes many computational resources. (2) DMs are easily misled by the blurry artifacts in the video, resulting in irrational content and distortion of the deblurred video. To address the above issues, we propose a novel video deblurring framework VD-Diff that integrates the diffusion model into the Wavelet-Aware Dynamic Transformer (WADT). Specifically, we perform the diffusion model in a highly compact latent space to generate prior features containing high-frequency information that conforms to the ground truth distribution. We design the WADT to preserve and recover the low-frequency information in the video while utilizing the high-frequency information generated by the diffusion model. Extensive experiments show that our proposed VD-Diff outperforms SOTA methods on GoPro, DVD, BSD, and Real-World Video datasets. The codes will be available at https://github.com/Chen-Rao/VD-Diff." + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.64, + 0.741, + 0.668 + ], + "angle": 0, + "content": "Keywords: Diffusion model \\(\\cdot\\) Prior feature \\(\\cdot\\) Video deblurring \\(\\cdot\\) Wavelet-based transformer" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.694, + 0.377, + 0.71 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.725, + 0.788, + 0.817 + ], + "angle": 0, + "content": "Portable devices are widely used to capture videos. However, camera shake and fast movement of objects lead to undesirable blur in video. To reduce the impact of video blur, researchers have done a lot of work on video deblurring. Traditional methods are mainly based on hand-crafted assumptions and priors, which are difficult to design. Besides, inappropriate priors will reduce the quality of video deblurring." + }, + { + "type": "page_footnote", + "bbox": [ + 0.232, + 0.825, + 0.644, + 0.84 + ], + "angle": 0, + "content": "† co-first authors, equal contribution; * Corresponding author" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.129 + ], + "angle": 0, + "content": "C. Rao et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.237 + ], + "angle": 0, + "content": "In the past decade, video deblurring has witnessed significant progresses with the development of deep learning. Early deep learning methods mainly use CNN-based models [13, 14, 26, 28, 36, 42, 46] to extract video features for capturing the spatio-temporal information between frames. Nevertheless, these approaches have limitations in modelling long-range spatial dependencies and capturing non-local self-similarities." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.241, + 0.788, + 0.452 + ], + "angle": 0, + "content": "Recently, Transformer has been employed in video deblurring tasks [4, 21, 22, 37] to address limitations in CNN-based methods, which can establish long-range spatial dependencies and capture non-local spatial information. Although Transformer-based methods exhibit impressive performance, they still have the following issues. On the one hand, the methods using standard global Transformer [8] tend to deblur videos based on low-frequency details and have limited perception of high-frequency information. On the other hand, when the local window-based Transformer [37] is used to reduce the computational cost, self-attention is calculated within a location-specific window, resulting in a limited receptive field. Consequently, it may ignore some key elements of similar and clearer scene patches in the spatiotemporal neighborhood. Therefore, the local window-based self-attention fails to capture valuable high-frequency information from the spatiotemporal neighborhood, resulting in the loss of high-frequency information." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.456, + 0.788, + 0.592 + ], + "angle": 0, + "content": "Very recently, diffusion models (DMs) have shown their outstanding capabilities in video synthesis tasks [2,9,23,24,31,32,40], which can denoise pure Gaussian noise into high-quality video with abundant high-frequency details. Therefore, we consider employing DMs in video deblurring to restore high-frequency information. However, directly applying DMs to video deblurring tasks has the following problems. First, DMs demand a large number of iteration steps to generate videos from Gaussian noise, which costs a lot of computational resources. Second, DMs are sensitive to motion blur in videos, resulting in irrational content and distortion of the deblurred video." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.596, + 0.788, + 0.702 + ], + "angle": 0, + "content": "To solve the above problems, we propose a novel method called VD-Diff, which combines our proposed Wavelet-Aware Dynamic Transformer (WADT) and diffusion model for video deblurring. Specifically, we employ the diffusion model in a highly compact latent space to generate prior features containing high-frequency information that conforms to the ground truth distribution, which facilitates video deblurring. Applying DM in a highly compressed latent space can significantly reduce the number of denoising iteration steps." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.705, + 0.788, + 0.84 + ], + "angle": 0, + "content": "In addition, we design WADT to retain and restore the low-frequency information in the video while exacting the high-frequency information from the prior features generated by the diffusion model, which ensures the deblurred video is undistorted. WADT first decomposes the blurred video frames into the approximation coefficient and detail coefficients by Wavelet Transform (WT) and extracts low-frequency information from the approximation coefficient by Wavelet-Aware Dynamic Transformer Layers (WADTL). Meanwhile, WADT employs WADTL to fuse the prior features generated by DM containing high-frequency information to generate artifact-free and distortion-free videos. Furthermore, we" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.356, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Video Deblurring with Transformer and Diffusion Model" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "image", + "bbox": [ + 0.217, + 0.144, + 0.784, + 0.303 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.221, + 0.308, + 0.238, + 0.32 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.24, + 0.309, + 0.39, + 0.318 + ], + "angle": 0, + "content": "Wavelet-Aware Dynamic Transformer Layer" + }, + { + "type": "image", + "bbox": [ + 0.403, + 0.309, + 0.419, + 0.319 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.421, + 0.309, + 0.58, + 0.319 + ], + "angle": 0, + "content": "Wavelet-based Bidirectional Propagation Fuse" + }, + { + "type": "image", + "bbox": [ + 0.597, + 0.309, + 0.612, + 0.319 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.613, + 0.309, + 0.663, + 0.318 + ], + "angle": 0, + "content": "Concatenation" + }, + { + "type": "image", + "bbox": [ + 0.671, + 0.309, + 0.687, + 0.319 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.69, + 0.31, + 0.779, + 0.318 + ], + "angle": 0, + "content": "Element-wise Summation" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.334, + 0.785, + 0.431 + ], + "angle": 0, + "content": "Fig. 1: The overall architecture of the proposed VD-Diff, which consists of Wavelet-Aware Dynamic Transformer (WADT) and Diffusion Model (DM). Specifically, WADT adopts Wavelet Transform (WT) for feature separation, WADT Layer (WADTL) for deep feature extraction, and Wavelet-based Bidirectional Propagation Fuse (WBPF) for spatio-temporal information propagation between frames. The DM generates prior features to supplement high-frequency information for the deblurring process in WADTL.WT: Wavelet Transform. IWT: Inverse Wavelet Transform." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.461, + 0.784, + 0.52 + ], + "angle": 0, + "content": "introduce a Wavelet-based Bidirectional Propagation Fuse (WBPF) module in WADT to fully use the spatio-temporal information between frames, which ensure that the video is undistorted in the spatio-temporal dimension. Our main contributions are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.522, + 0.784, + 0.595 + ], + "angle": 0, + "content": "(1) We propose a novel method combining the Wavelet-Aware Dynamic Transformer (WADT) and the diffusion model for video deblurring. To the best of our knowledge, our study is the first to employ DM for video deblurring, which requires only a few iteration steps to restore artifact-free and distortion-free videos." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.597, + 0.784, + 0.641 + ], + "angle": 0, + "content": "(2) Our diffusion model can generate prior features containing high-frequency information that conforms to the GT distribution, which facilitates video deblurring." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.642, + 0.784, + 0.687 + ], + "angle": 0, + "content": "(3) Our proposed WADT can capture low-frequency information while utilizing the prior features generated by the DM to supplement high-frequency information, which further enhance the quality of deblurred video." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.687, + 0.79, + 0.717 + ], + "angle": 0, + "content": "(4) Extensive experiments show that our proposed VD-Diff outperforms SOTA methods on GoPro, DVD, BSD and Real-World Video datasets." + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.522, + 0.79, + 0.717 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.74, + 0.388, + 0.756 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.771, + 0.407, + 0.787 + ], + "angle": 0, + "content": "2.1 Video Deblurring" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.784, + 0.842 + ], + "angle": 0, + "content": "With the development of deep learning [20], CNN is widely used in video deblurring. For instance, Zhong et al. [46] propose residual dense blocks and a global spatio-temporal attention module for video deblurring; Pan et al. [28] develop a" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.129 + ], + "angle": 0, + "content": "C. Rao et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.299 + ], + "angle": 0, + "content": "deep CNN model to estimate optical flow from intermediate latent frames. CNN-based methods have driven the development of video deblurring, but CNN has limitations in capturing long-range dependencies. To address these issues, some researchers applied the Transformer to video deblurring and achieved impressive results. For example, Lin et al. [22] propose a method to calculate self-attention by capturing key elements of adjacent frames through optical flow. Liang et al. [21] propose a recurrent video restoration transformer to process local neighboring frames in parallel within a globally recurrent framework. However, the above Transformer-based methods have limited ability to capture high-frequency information, leading to the deblurred videos lacking details." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.317, + 0.405, + 0.332 + ], + "angle": 0, + "content": "2.2 Diffusion Models" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.338, + 0.789, + 0.566 + ], + "angle": 0, + "content": "DMs have shown excellent performance in the field of synthesis. LDM [31] reaches a near-optimal point between complexity reduction and detail preservation by applying DM in the latent space. Blattmann et al. [2] turns the image generator into a video generator by introducing a temporal dimension to the latent space diffusion model. Artbank [43] utilize textual embedding to learn the visual priors from the images, synthesizing highly realistic artistic stylized images. DMs have also shown strong capabilities in the field of image restoration. For instance, DiffIR [38], Hi-diff [5] and DiffMSR [18] use DM to generate prior representations for image restoration. Although some methods, such as DDIM [45] and LDM [31], have reduced the computational complexity of DM by changing the denoising process or compressing the image, using existing DM for video deblurring still requires unaffordable computational resources. To cope with the above issue, we apply DM in a highly compact latent space to reduce the iteration steps. We were the first to apply diffusion models to video deblurring, and our models achieved outstanding performance." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.584, + 0.564, + 0.599 + ], + "angle": 0, + "content": "2.3 Wavelet-based Restoration Methods" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.605, + 0.789, + 0.743 + ], + "angle": 0, + "content": "Wavelet is widely used in low-visual restoration tasks [7, 15, 19, 44, 48] as it can separate low-frequency and high-frequency information. For example, Huang et al. [12] propose a wavelet-based CNN for multi-scale face super resolution. WavTrans [17] combines wavelet transform with Transformer for feature extraction and fusion; DSTNet [29] develops a wavelet-based feature propagation method to effectively propagate main structures from long-range frames for better video deblurring. Inspired by the above methods, we proposed Wavelet-Aware Dynamic Transformer, which can extract the global structure of the video and utilize the prior features generated by DM to restore high-frequency details." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.761, + 0.382, + 0.779 + ], + "angle": 0, + "content": "3 Methodology" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.789, + 0.434, + 0.804 + ], + "angle": 0, + "content": "3.1 Overall Architecture" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.842 + ], + "angle": 0, + "content": "The overall framework of our proposed VD-Diff is shown in Fig. 1, which consists of two parts: (1) Wavelet-Aware Dynamic Transformer (WADT); (2) Diffusion" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.355, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "Video Deblurring with Transformer and Diffusion Model" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.224 + ], + "angle": 0, + "content": "Model. Specifically, we first employ Diffusion Model to obtain a highly compact latent prior feature \\(z \\in \\mathbb{R}^{T \\times C'}\\), which contains supplementary high-frequency information. Note that \\(C'\\) is the number of channels of latent features. Then, we employ WADT to deblur the blurry video frames \\(V_{blur} \\in R^{T \\times 3 \\times H \\times W}\\) into \\(V_{HQ} \\in R^{T \\times 3 \\times H \\times W}\\) with the help of the prior feature \\(z\\)." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.244, + 0.573, + 0.26 + ], + "angle": 0, + "content": "3.2 Wavelet-Aware Dynamic Transformer" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.268, + 0.789, + 0.435 + ], + "angle": 0, + "content": "As shown in the WADT part of Fig. 1, we first utilize Conv3D to perform shallow feature extraction on the blurry video frames \\( V_{blur} \\in R^{T \\times 3 \\times H \\times W} \\) to obtain a global feature \\( F_{in} \\in \\mathbb{R}^{T \\times C \\times H \\times W} \\). Then, we use Wavelet Transform (WT) to decompose it into approximate coefficients \\( F_{a} \\in \\mathbb{R}^{T \\times C / 4 \\times H \\times W} \\) and detail coefficients \\( F_{d} \\in \\mathbb{R}^{T \\times 3C / 4 \\times H \\times W} \\) and again use WT on \\( F_{a} \\) to obtain \\( F_{aa} \\in \\mathbb{R}^{T \\times \\hat{C} \\times H \\times W} \\) and \\( F_{ad} \\in \\mathbb{R}^{T \\times 3\\hat{C} \\times H \\times W} \\), where \\( \\hat{C} = C / 16 \\). Next, we employ WADTL to obtain the deblurred features \\( F_{out} \\) from the approximate coefficients \\( F_{aa} \\) and the prior feature \\( z' \\). Then, we use WBPF to explore spatiotemporal information further. Finally, we use two IWTs to fuse features to obtain artifact-free and distortion-free videos \\( V_{HQ} \\). We will elaborate on the structure of WADTL and WBPF as follows." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.436, + 0.789, + 0.51 + ], + "angle": 0, + "content": "Wavelet-Aware Dynamic Transformer Layer. As shown in Fig. 1, we employ a stack of multiple WADT layer (WADTL) modules hierarchically to fuse approximate coefficients \\( F_{aa} \\) and prior features \\( z' \\). As shown in Fig. 2, WADTL consists of a Wavelet-Aware Dynamic Multi-head Self-Attention (WAD-MSA) and a Wavelet-Aware Dynamic Feed-Forward Network (WAD-FFN)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.511, + 0.789, + 0.542 + ], + "angle": 0, + "content": "In the WAD-MSA part, given a approximate coefficients \\( F_{aa} \\), we use the highly compact latent prior feature \\( z' \\) to guide video restoration:" + }, + { + "type": "equation", + "bbox": [ + 0.393, + 0.551, + 0.785, + 0.568 + ], + "angle": 0, + "content": "\\[\n\\hat {F} = W _ {l} ^ {1} z ^ {\\prime} \\odot L N \\left(F _ {a a}\\right) + W _ {l} ^ {2} z ^ {\\prime}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.578, + 0.788, + 0.687 + ], + "angle": 0, + "content": "where \\(\\odot\\) indicates element-wise multiplication, \\(LN\\) denotes layer normalization [1], \\(W_{l}\\) represents linear layer. Then, we utilize a 3D convolution layer and a 2D convolution layer to project \\(\\hat{F}\\) into query \\(Q = W_2^Q W_3^Q\\hat{F}\\), key \\(K = W_2^K W_3^K\\hat{F}\\), value \\(V = W_2^V W_3^V\\hat{F}\\) where \\(W_{3}\\) is the channel-wise 3D convolution and \\(W_{2}\\) is the depth-wise 2D convolution. Note that we use 3D convolution in the \\(T\\), \\(H\\), and \\(W\\) dimensions and the same 2D convolution for the \\(T\\) features in the temporal dimension \\(T\\). Next, we perform the attention calculation:" + }, + { + "type": "equation", + "bbox": [ + 0.358, + 0.697, + 0.785, + 0.714 + ], + "angle": 0, + "content": "\\[\n\\mathcal {F} _ {\\text {o u t}} ^ {\\prime} = W _ {a} \\hat {V} \\cdot \\operatorname {S o f t m a x} (\\hat {K} \\cdot \\hat {Q} / \\gamma) + F _ {a a}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.724, + 0.514, + 0.738 + ], + "angle": 0, + "content": "where \\(\\gamma\\) is a learnable scaling parameter." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.738, + 0.788, + 0.814 + ], + "angle": 0, + "content": "In the WDA-FFN part, we first integrate the prior \\( z' \\) into \\( \\mathcal{F}_{out}' \\) to obtain \\( \\hat{\\mathcal{F}}_{out}' \\). And then we use Conv3D to aggregate temporal information and use Conv2D to aggregate spatial information from neighboring pixels. Besides, we employ a gating mechanism to obtain more valuable information. The overall process of WDA-FFN can be described as:" + }, + { + "type": "equation", + "bbox": [ + 0.345, + 0.825, + 0.785, + 0.842 + ], + "angle": 0, + "content": "\\[\nF _ {o u t} = G \\left(W _ {2} ^ {1} W _ {3} ^ {1} \\hat {\\mathcal {F}} _ {o u t} ^ {\\prime}\\right) \\odot W _ {2} ^ {2} W _ {3} ^ {2} \\hat {\\mathcal {F}} _ {o u t} ^ {\\prime} + \\hat {\\mathcal {F}} _ {o u t} ^ {\\prime}, \\tag {3}\n\\]" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.128 + ], + "angle": 0, + "content": "C. Rao et al." + }, + { + "type": "image", + "bbox": [ + 0.221, + 0.144, + 0.784, + 0.232 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.243, + 0.788, + 0.288 + ], + "angle": 0, + "content": "Fig. 2: The illustration of Wavelet-Aware Dynamic Transformer Layer (WADTL), which consists of Wavelet-Aware Dynamic Multi-head Self-Attention (WAD-MSA) and Wavelet-Aware Dynamic Feed-Forward Network (WAD-FFN)." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.306, + 0.784, + 0.482 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.495, + 0.789, + 0.54 + ], + "angle": 0, + "content": "Fig. 3: The structure of the Forward Process in the WBPF. The Backward Process has the same network structure as the Forward Process, but the direction of information propagation between frames is opposite." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.575, + 0.617, + 0.59 + ], + "angle": 0, + "content": "where \\(G\\) denotes Gaussian Error Linear Units (GELU)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.59, + 0.789, + 0.761 + ], + "angle": 0, + "content": "Wavelet-based Bidirectional Propagation Fuse. Given the generated features \\(\\{X_{i}\\}_{i = 1}^{N}\\) by the first Inverse Wavelet Transform (IwT), to explore useful contents from the features of other frames, exiting methods usually simply stack \\(\\{X_{i}\\}_{i = 1}^{N}\\) or the alignment results of \\(\\{X_{i}\\}_{i = 1}^{N}\\). However, directly stacking frames without accurately estimating the features of \\(\\{X_{i}\\}_{i = 1}^{N}\\) or the alignment results of \\(\\{X_{i}\\}_{i = 1}^{N}\\) can only partially exploit the valuable information between frames and may also cause the propagation of erroneous information between frames. To further capture more useful long-range spatio-temporal information and reduce the influence of inaccurate information, we propose Wavelet-based Bidirectional Propagation Fuse (WBPF). As shown in Fig. 1, WBPF consists of a forward process and a backward process." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.79, + 0.842 + ], + "angle": 0, + "content": "We take the forward process as an example to introduce the WBPF module in detail, as shown in Fig. 3. First, we connect \\( Y_{i} \\) and \\( X_{i + 1} \\) to get \\( \\bar{X} \\), where \\( Y_{i} \\) is the output of the previous forward process and \\( X_{i + 1} \\) is the input of the current forward process. We then use a convolution layer with LeakyReLU to project \\( \\bar{X} \\) into \\( \\hat{X} \\). Then we split \\( \\hat{X} \\) into \\( \\hat{X}_1, \\hat{X}_2 \\) along the channel dimension. Next, we" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.355, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "Video Deblurring with Transformer and Diffusion Model" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.787, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.145, + 0.512, + 0.163 + ], + "angle": 0, + "content": "obtain the fused feature from \\(\\hat{X}_1, \\hat{X}_2\\) by:" + }, + { + "type": "equation", + "bbox": [ + 0.355, + 0.173, + 0.786, + 0.192 + ], + "angle": 0, + "content": "\\[\nF _ {p r o} = \\hat {X} _ {1} \\odot S \\left(W _ {1} \\hat {X} _ {1}\\right) + \\hat {X} _ {2} \\odot S \\left(W _ {2} \\hat {X} _ {2}\\right), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.202, + 0.787, + 0.248 + ], + "angle": 0, + "content": "where \\( S \\) denotes the Sigmoid function and \\( W \\) is the \\( 3 \\times 3 \\) 2D convolution. We utilize AvgPool, ResBlock, MaxPool, and Conv2D to better explore spatial information. The overall process can be described as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.379, + 0.26, + 0.786, + 0.278 + ], + "angle": 0, + "content": "\\[\n\\bar {Y} _ {i + 1} = F _ {p r o} \\otimes \\text {u p s a m p l e} (P \\left(F _ {p r o}\\right)), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.288, + 0.787, + 0.32 + ], + "angle": 0, + "content": "where \\(\\mathrm{P}\\) represents the stack of AvgPool, ResBlock, MaxPool, ResBlock and Conv2D. In the last, we employ 30 ResBlocks on \\(\\bar{Y}_{i + 1}\\) to obtain \\(Y_{i + 1}\\)." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.341, + 0.396, + 0.355 + ], + "angle": 0, + "content": "3.3 Diffusion Model" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.366, + 0.788, + 0.472 + ], + "angle": 0, + "content": "Based on conditional denoising diffusion probabilistic models [11,27,31,35], the DM involves a forward diffusion process and a reverse denoising process. We first adopt the Latent Encoder (LE) on ground truth video and the Condition Encoder (CE) on blurry video to generate the ground truth feature \\( z \\in R^{T \\times 4C'} \\) and the condition feature \\( c \\in R^{T \\times 4C'} \\). Then, we use the diffusion model to generate the prior feature \\( z' \\in R^{T \\times 4C'} \\) using \\( c \\in R^{T \\times 4C'} \\) as the condition and \\( z \\in R^{T \\times 4C'} \\) as the target. We detail the diffusion process as follows." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.473, + 0.787, + 0.502 + ], + "angle": 0, + "content": "In the forward diffusion process, we transform \\(z\\) into Gaussian noise by \\(T\\) iterations, which can be described as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.369, + 0.513, + 0.786, + 0.532 + ], + "angle": 0, + "content": "\\[\nq \\left(z _ {t} \\mid z _ {t - 1}\\right) = \\mathcal {N} \\left(z _ {t}; \\sqrt {1 - \\beta_ {t}} z _ {t - 1}, \\beta_ {t} I\\right), \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.543, + 0.787, + 0.589 + ], + "angle": 0, + "content": "where \\( t = 1,2,\\dots,T \\); \\( z_0 = z \\); \\( \\beta_{t} \\in (0,1) \\) are hyperparameters that control the variance of the noise. Using the notation \\( \\alpha_{t} = 1 - \\beta_{t} \\), \\( \\bar{\\alpha}_{t} = \\prod_{i=1}^{t} \\alpha_{i} \\) and through iterative derivation with reparameterization, Eq. (6) can be written as:" + }, + { + "type": "equation", + "bbox": [ + 0.369, + 0.601, + 0.786, + 0.618 + ], + "angle": 0, + "content": "\\[\nq \\left(z _ {T} \\mid z _ {0}\\right) = \\mathcal {N} \\left(z _ {T}; \\sqrt {\\bar {\\alpha} _ {T}} z _ {0}, (1 - \\bar {\\alpha} _ {T}) I\\right), \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.629, + 0.268, + 0.642 + ], + "angle": 0, + "content": "that is" + }, + { + "type": "equation", + "bbox": [ + 0.411, + 0.643, + 0.786, + 0.66 + ], + "angle": 0, + "content": "\\[\nz _ {T} = \\sqrt {\\bar {\\alpha} _ {T}} z _ {0} + \\sqrt {1 - \\bar {\\alpha} _ {T}} \\epsilon . \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.667, + 0.787, + 0.697 + ], + "angle": 0, + "content": "In the reverse denoising process, DM sample a Gaussian random noise \\( z_{T} \\) and then gradually denoise \\( z_{T} \\) to \\( z_{0} \\):" + }, + { + "type": "equation", + "bbox": [ + 0.363, + 0.709, + 0.786, + 0.726 + ], + "angle": 0, + "content": "\\[\nq \\left(z _ {t - 1} \\mid z _ {t}, z _ {0}\\right) = \\mathcal {N} \\left(z _ {t - 1}; \\mu_ {t} \\left(z _ {t}, z _ {0}\\right), \\tilde {\\beta} _ {t} I\\right), \\tag {9}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.346, + 0.749, + 0.786, + 0.78 + ], + "angle": 0, + "content": "\\[\n\\mu_ {t} \\left(z _ {t}, z _ {0}\\right) = \\frac {\\sqrt {\\bar {\\alpha} _ {t - 1}} \\beta_ {t}}{1 - \\bar {\\alpha} _ {t}} z _ {0} + \\frac {\\sqrt {\\alpha_ {t}} \\left(1 - \\bar {\\alpha} _ {t - 1}\\right)}{1 - \\bar {\\alpha} _ {t}} z _ {t}. \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.786, + 0.638, + 0.801 + ], + "angle": 0, + "content": "We can use Eq. (8) to represent \\( z_0 \\) in Eq. (10) and obtain:" + }, + { + "type": "equation", + "bbox": [ + 0.382, + 0.812, + 0.786, + 0.845 + ], + "angle": 0, + "content": "\\[\n\\mu_ {t} \\left(z _ {t}, z _ {0}\\right) = \\frac {1}{\\sqrt {\\alpha_ {t}}} \\left(z _ {t} - \\epsilon \\frac {1 - \\alpha_ {t}}{\\sqrt {1 - \\bar {\\alpha} _ {t}}}\\right). \\tag {11}\n\\]" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.128 + ], + "angle": 0, + "content": "C. Rao et al." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.147, + 0.551, + 0.163 + ], + "angle": 0, + "content": "We remove the variance estimation in Eq. (9):" + }, + { + "type": "equation", + "bbox": [ + 0.399, + 0.173, + 0.786, + 0.206 + ], + "angle": 0, + "content": "\\[\nz _ {t - 1} = \\frac {1}{\\sqrt {\\alpha_ {t}}} (z _ {t} - \\epsilon \\frac {1 - \\alpha_ {t}}{\\sqrt {1 - \\bar {\\alpha} _ {t}}}). \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.215, + 0.785, + 0.246 + ], + "angle": 0, + "content": "Then the noise \\(\\epsilon\\) is the only uncertain variable. Following previous work [11, 31, 38], we adopt a neural network to predict the noise \\(\\epsilon\\) on \\(z_{t}\\), \\(c\\) and \\(t\\)." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.267, + 0.538, + 0.284 + ], + "angle": 0, + "content": "3.4 Training and Inference Strategies" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.292, + 0.784, + 0.367 + ], + "angle": 0, + "content": "Inspired by [5,38], we divide the training process into three stages: the first is to train the Latent Encoder with Wavelet-Aware Dynamic Transformer for ground truth feature extraction, the second is to train the Diffusion Model to generate the prior feature, and the third is to jointly train DM and WADT to generate artifact-free and high-realistic videos." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.367, + 0.788, + 0.455 + ], + "angle": 0, + "content": "Training Stage One. The first stage aims to train the Latent Encoder (LE) to extract valuable ground truth prior feature \\( z \\) from ground truth (GT). We train LE and WADT as a whole model by inputting \\( z \\) into WADTL instead of \\( z' \\) and ignoring the condition encoder and the diffusion model in Fig. 1. We employ L1 loss and Multi-scale Frequency Reconstruction (MSFR) [10] loss. L1 loss is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.416, + 0.456, + 0.786, + 0.495 + ], + "angle": 0, + "content": "\\[\nL _ {1} = \\sum_ {i = 1} ^ {N} \\| V _ {H Q} ^ {i} - V _ {G T} ^ {i} \\| _ {1}, \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.502, + 0.788, + 0.548 + ], + "angle": 0, + "content": "where \\( V_{HQ}^{i} \\) is the \\( i \\)-th frame of the deblurred HQ video and \\( V_{GT}^{i} \\) is the \\( i \\)-th frame of the GT video. The MSFR loss measures the L1 distance between multi-scale GT and deblurred frame in the frequency domain as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.348, + 0.558, + 0.786, + 0.6 + ], + "angle": 0, + "content": "\\[\nL _ {M S F R} = \\sum_ {i = 1} ^ {N} \\sum_ {k = 1} ^ {K} \\frac {1}{t _ {k}} \\| \\mathcal {F T} \\left(\\hat {S} _ {k} ^ {i}\\right) - \\mathcal {F T} \\left(S _ {k} ^ {i}\\right) \\| _ {1}, \\tag {14}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.608, + 0.788, + 0.67 + ], + "angle": 0, + "content": "where \\(\\mathcal{FT}\\) denotes the Fast Fourier Transform that transfers image signal to the frequency domain, \\(K\\) denotes the number of different scales and the \\(\\tilde{S}_k^i\\) indicates the image signal of the \\(i\\)-th frame at the \\(k\\)-th scale. The final optimization function of stage one is determined as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.413, + 0.682, + 0.786, + 0.699 + ], + "angle": 0, + "content": "\\[\nL _ {d e b l u r} = L _ {1} + \\lambda L _ {M S F R}, \\tag {15}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.708, + 0.484, + 0.722 + ], + "angle": 0, + "content": "where we experimentally set \\(\\lambda = 0.1\\)." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.723, + 0.784, + 0.782 + ], + "angle": 0, + "content": "Training Stage Two. In the second stage, we first use a Condition Encoder (CE), whose struct is the same as the Latent Encoder (LE), to generate condition feature \\( c \\) from the blur video. Then, we train a diffusion model (DM) to generate prior feature \\( z' \\) using condition feature \\( c \\)." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.784, + 0.786, + 0.813 + ], + "angle": 0, + "content": "We run the complete \\(T\\) iteration reverse processes Eq. (12) to generate prior feature \\(\\hat{z}\\). We denoted this loss as:" + }, + { + "type": "equation", + "bbox": [ + 0.437, + 0.825, + 0.786, + 0.843 + ], + "angle": 0, + "content": "\\[\nL _ {d i f f} = \\left\\| z ^ {\\prime} - \\hat {z} \\right\\| _ {1}. \\tag {16}\n\\]" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.356, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "Video Deblurring with Transformer and Diffusion Model" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "image", + "bbox": [ + 0.22, + 0.144, + 0.478, + 0.259 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.48, + 0.144, + 0.541, + 0.19 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.497, + 0.192, + 0.523, + 0.201 + ], + "angle": 0, + "content": "Blur" + }, + { + "type": "image", + "bbox": [ + 0.542, + 0.144, + 0.601, + 0.19 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.551, + 0.192, + 0.59, + 0.201 + ], + "angle": 0, + "content": "EDVR" + }, + { + "type": "image", + "bbox": [ + 0.603, + 0.144, + 0.664, + 0.19 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.607, + 0.192, + 0.66, + 0.201 + ], + "angle": 0, + "content": "MPRNet" + }, + { + "type": "image", + "bbox": [ + 0.666, + 0.144, + 0.725, + 0.19 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.682, + 0.192, + 0.708, + 0.201 + ], + "angle": 0, + "content": "TSP" + }, + { + "type": "image", + "bbox": [ + 0.727, + 0.144, + 0.787, + 0.19 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.743, + 0.192, + 0.77, + 0.201 + ], + "angle": 0, + "content": "SFE" + }, + { + "type": "image", + "bbox": [ + 0.48, + 0.201, + 0.542, + 0.247 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.485, + 0.249, + 0.535, + 0.258 + ], + "angle": 0, + "content": "DSTNet" + }, + { + "type": "image", + "bbox": [ + 0.542, + 0.201, + 0.601, + 0.247 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.553, + 0.249, + 0.589, + 0.258 + ], + "angle": 0, + "content": "FGST" + }, + { + "type": "image", + "bbox": [ + 0.602, + 0.201, + 0.664, + 0.247 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.604, + 0.249, + 0.664, + 0.258 + ], + "angle": 0, + "content": "DSTNet-L" + }, + { + "type": "image", + "bbox": [ + 0.666, + 0.201, + 0.725, + 0.247 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.68, + 0.249, + 0.71, + 0.258 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image", + "bbox": [ + 0.727, + 0.201, + 0.787, + 0.247 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.746, + 0.249, + 0.767, + 0.258 + ], + "angle": 0, + "content": "GT" + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.271, + 0.785, + 0.299 + ], + "angle": 0, + "content": "Fig. 4: Visual comparison on GoPro [25] dataset. The deblurred results of previous work still contain significant blur effects. Our method generates much clearer frames." + }, + { + "type": "image", + "bbox": [ + 0.219, + 0.314, + 0.478, + 0.427 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.48, + 0.314, + 0.54, + 0.36 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.362, + 0.522, + 0.37 + ], + "angle": 0, + "content": "Blur" + }, + { + "type": "image", + "bbox": [ + 0.542, + 0.314, + 0.601, + 0.36 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.552, + 0.362, + 0.59, + 0.37 + ], + "angle": 0, + "content": "EDVR" + }, + { + "type": "image", + "bbox": [ + 0.602, + 0.314, + 0.662, + 0.36 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.606, + 0.362, + 0.657, + 0.37 + ], + "angle": 0, + "content": "MPRNet" + }, + { + "type": "image", + "bbox": [ + 0.665, + 0.314, + 0.724, + 0.36 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.68, + 0.362, + 0.705, + 0.37 + ], + "angle": 0, + "content": "TSP" + }, + { + "type": "image", + "bbox": [ + 0.725, + 0.314, + 0.784, + 0.36 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.741, + 0.362, + 0.767, + 0.37 + ], + "angle": 0, + "content": "SFE" + }, + { + "type": "image", + "bbox": [ + 0.48, + 0.372, + 0.541, + 0.417 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.482, + 0.418, + 0.54, + 0.428 + ], + "angle": 0, + "content": "BasicVSR" + }, + { + "type": "image", + "bbox": [ + 0.542, + 0.372, + 0.601, + 0.417 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.553, + 0.418, + 0.59, + 0.427 + ], + "angle": 0, + "content": "FGST" + }, + { + "type": "image", + "bbox": [ + 0.602, + 0.372, + 0.663, + 0.417 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.607, + 0.418, + 0.657, + 0.427 + ], + "angle": 0, + "content": "DSTNet" + }, + { + "type": "image", + "bbox": [ + 0.665, + 0.372, + 0.724, + 0.417 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.678, + 0.418, + 0.708, + 0.427 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image", + "bbox": [ + 0.725, + 0.372, + 0.784, + 0.417 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.743, + 0.418, + 0.764, + 0.427 + ], + "angle": 0, + "content": "GT" + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.441, + 0.785, + 0.469 + ], + "angle": 0, + "content": "Fig. 5: Visual comparison on DVD [33] dataset. The deblurring effect of our proposed model is significantly better." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.501, + 0.788, + 0.621 + ], + "angle": 0, + "content": "Training Stage Three. In Training Stage Two, we train the DM to generate \\( z' \\) to approximate \\( z \\). However, there inevitably is a slight deviation between the prior features \\( z' \\) and the ground truth prior feature \\( z \\). To make the DM and WADT work better together to acquire information from different frequency domains for video deblurring and improve the robustness of WADT to the prior feature \\( z' \\), we jointly train DM and WADT in Training Stage Three. The training loss \\( L_{total} \\) is given by the sum of the deblurring loss \\( L_{deblur} \\) and the diffusion loss \\( L_{diff} \\):" + }, + { + "type": "equation", + "bbox": [ + 0.412, + 0.623, + 0.785, + 0.639 + ], + "angle": 0, + "content": "\\[\nL _ {t o t a l} = L _ {d e b l u r} + L _ {d i f f}. \\tag {17}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.646, + 0.788, + 0.707 + ], + "angle": 0, + "content": "Inference. Given a blurry input video \\( V_{blur} \\in R^{T \\times 3 \\times H \\times W} \\), we first compress \\( V_{blur} \\) into a condition feature \\( c \\in R^{T \\times 4C'} \\). Then, the diffusion model generates the prior feature \\( z' \\) from pure Gaussian noise \\( \\epsilon \\) and condition feature \\( c \\). Finally, VD-diff deblurs the blurry video with \\( z' \\) using WADT." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.734, + 0.376, + 0.751 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.767, + 0.334, + 0.781 + ], + "angle": 0, + "content": "4.1 Datasets" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.785, + 0.84 + ], + "angle": 0, + "content": "GoPro. The GoPro [25] benchmark is composed of over 3,300 blurry-sharp image pairs of dynamic scenes. It is obtained by a high-speed camera. The training and testing subsets are split in proportional to 2:1." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.128 + ], + "angle": 0, + "content": "C. Rao et al." + }, + { + "type": "image", + "bbox": [ + 0.222, + 0.144, + 0.359, + 0.243 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.278, + 0.244, + 0.306, + 0.254 + ], + "angle": 0, + "content": "Blur" + }, + { + "type": "image", + "bbox": [ + 0.359, + 0.145, + 0.496, + 0.243 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.402, + 0.244, + 0.457, + 0.255 + ], + "angle": 0, + "content": "DVD [33]" + }, + { + "type": "image", + "bbox": [ + 0.496, + 0.145, + 0.632, + 0.243 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.535, + 0.244, + 0.598, + 0.256 + ], + "angle": 0, + "content": "EDVR [36]" + }, + { + "type": "image", + "bbox": [ + 0.632, + 0.145, + 0.775, + 0.243 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.665, + 0.244, + 0.741, + 0.256 + ], + "angle": 0, + "content": "MPRNet [41]" + }, + { + "type": "image", + "bbox": [ + 0.221, + 0.256, + 0.359, + 0.353 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.268, + 0.355, + 0.319, + 0.367 + ], + "angle": 0, + "content": "TSP [28]" + }, + { + "type": "image", + "bbox": [ + 0.359, + 0.256, + 0.496, + 0.353 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.404, + 0.355, + 0.455, + 0.367 + ], + "angle": 0, + "content": "SFE [39]" + }, + { + "type": "image", + "bbox": [ + 0.496, + 0.256, + 0.632, + 0.353 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.529, + 0.355, + 0.602, + 0.367 + ], + "angle": 0, + "content": "DSTNet [29]" + }, + { + "type": "image", + "bbox": [ + 0.632, + 0.256, + 0.775, + 0.353 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.656, + 0.355, + 0.744, + 0.367 + ], + "angle": 0, + "content": "VD-Diff (Ours)" + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.378, + 0.785, + 0.394 + ], + "angle": 0, + "content": "Fig. 6: Visual results of VD-Diff and SOTA methods on the real blurry videos of [33]." + }, + { + "type": "image", + "bbox": [ + 0.226, + 0.409, + 0.391, + 0.515 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.403, + 0.409, + 0.458, + 0.516 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.469, + 0.411, + 0.524, + 0.516 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.534, + 0.411, + 0.586, + 0.516 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.599, + 0.411, + 0.652, + 0.516 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.662, + 0.411, + 0.715, + 0.516 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.727, + 0.412, + 0.78, + 0.516 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.275, + 0.541, + 0.726, + 0.555 + ], + "angle": 0, + "content": "Fig. 7: Comparison of various methods in terms of time dimension." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.583, + 0.788, + 0.643 + ], + "angle": 0, + "content": "DVD. The DVD [33] dataset consists of 71 videos with 6,708 blurry-sharp image pairs. It is divided into train/test subsets with 61 videos (5,708 image pairs) and 10 videos (1,000 image pairs). DVD is captured with mobile phones and DSLR at a frame rate of 240 FPS." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.644, + 0.788, + 0.689 + ], + "angle": 0, + "content": "BSD. The Beam-Splitter Deblurring Dataset (BSD) [46] consists of 240 videos with 24k blurry-sharp image pairs. Cross-validation experiments between real-world and synthetic datasets demonstrate the high generality of the BSD dataset." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.788, + 0.735 + ], + "angle": 0, + "content": "Real-World Videos. We evaluate the generality of our proposed approach on the real blurry datasets [6]. Due to the inaccessibility of the ground truth, we only compare VD-Diff with other models on visual results." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.756, + 0.458, + 0.771 + ], + "angle": 0, + "content": "4.2 Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.789, + 0.842 + ], + "angle": 0, + "content": "We implement VD-Diff in PyTorch [30] using a single NVIDIA RTX A6000 GPU. The training process takes 7 days and 10 hours. We set the batch size as 2, the learning rate as \\(8 \\times 10^{-5}\\) and weight decay as \\(4 \\times 10^{-5}\\). Patches at the size of \\(256 \\times 256\\) cropped from training frames are fed into the models. The sequence" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.356, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Video Deblurring with Transformer and Diffusion Model" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "image", + "bbox": [ + 0.217, + 0.144, + 0.353, + 0.248 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.362, + 0.144, + 0.497, + 0.248 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.507, + 0.144, + 0.642, + 0.248 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.653, + 0.144, + 0.787, + 0.248 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.217, + 0.253, + 0.352, + 0.356 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.271, + 0.358, + 0.297, + 0.368 + ], + "angle": 0, + "content": "Blur" + }, + { + "type": "image", + "bbox": [ + 0.362, + 0.253, + 0.497, + 0.356 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.414, + 0.358, + 0.445, + 0.368 + ], + "angle": 0, + "content": "LDM" + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.253, + 0.642, + 0.356 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.561, + 0.358, + 0.587, + 0.368 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image", + "bbox": [ + 0.653, + 0.253, + 0.787, + 0.356 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.708, + 0.358, + 0.726, + 0.368 + ], + "angle": 0, + "content": "GT" + }, + { + "type": "image_caption", + "bbox": [ + 0.25, + 0.386, + 0.751, + 0.402 + ], + "angle": 0, + "content": "Fig. 8: Qualitative comparison with pre-trained large diffusion models [31]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.43, + 0.788, + 0.507 + ], + "angle": 0, + "content": "length is set to 15 in training and 16 in testing. Since we use diffusion within a highly compressed latent space, the denoising iteration steps \\( T \\) only needs 4. We set \\( C \\) to 64 and \\( C' \\) to 256, respectively. In the WADT module, we set \\( N_1 \\) and \\( N_2 \\) to 7 and 2, respectively, to balance the calculation amount and the deblurring effect. The horizontal and vertical flips are performed for data augmentation." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.528, + 0.434, + 0.542 + ], + "angle": 0, + "content": "4.3 Quantitative Results" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.552, + 0.789, + 0.765 + ], + "angle": 0, + "content": "We compare our proposed VD-Diff with other SOTA methods, as shown in Tables 1, 2, and 3. As can be seen, our method outperforms SOTA methods on GoPro [25] and DVD [33] benchmark datasets. Specifically, as shown in Table 1, Our method outperforms DSTNet-L, RVRT, and FGST by \\(0.50\\mathrm{dB}\\), \\(0.63\\mathrm{dB}\\), and \\(2.65\\mathrm{dB}\\) in PSNR on the GoPro [25] dataset, respectively. Our model achieves SOTA performance on the GoPro [25] dataset. Table 2 shows that our proposed method generates the deblurred videos with higher PSNR and SSIM values on the DVD [33] dataset. As shown in Table 3, we also achieve SOTA on the BSD [46] dataset. As seen from Table 4, our model has a relatively low parameter size and less inference time than other models. Although our model's parameter size and inference time are slightly higher than DSTNet and BasicVSR++, the PSNR metric on the GoPro [25] dataset is at least 1.3dB higher than the two models. As shown in Fig. 9(a), our model achieves favorable results in terms of accuracy and FLOPs. The above results show the effectiveness of our model." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.785, + 0.421, + 0.8 + ], + "angle": 0, + "content": "4.4 Qualitative Results" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.84 + ], + "angle": 0, + "content": "We provide visual comparisons on GoPro [25], DVD [33], and Real-World Video datasets [33] as shown in Fig. 4, 5, and 6, respectively. Some models sacrifice" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.128 + ], + "angle": 0, + "content": "C. Rao et al." + }, + { + "type": "table", + "bbox": [ + 0.214, + 0.145, + 0.775, + 0.236 + ], + "angle": 0, + "content": "
MethodDVD [33]EDVR [36]STFAN [47]ESTRNN [46]MPRNet [41]TSP [28]
PSNR ↑27.3126.8328.5931.0732.7331.67
SSIM ↑0.82550.84260.86080.90230.93660.9279
BasicVSR [3]NAFNet [4]FGST [22]DSTNet [29]DSTNet-L [29]RVRT [21]Ours
34.0133.7132.9034.1635.0534.9235.55
0.95200.96680.96100.96790.97330.97380.9752
" + }, + { + "type": "table_caption", + "bbox": [ + 0.273, + 0.236, + 0.728, + 0.251 + ], + "angle": 0, + "content": "Table 1: VD-Diff achieves SOTA results on the GoPro dataset [25]." + }, + { + "type": "table", + "bbox": [ + 0.238, + 0.27, + 0.767, + 0.362 + ], + "angle": 0, + "content": "
MethodDVD [33]EDVR [36]STFAN [47]ESTRNN [46]TSP [28]SEF [39]
PSNR ↑30.0128.5131.1532.0132.1331.71
SSIM ↑0.88770.86370.90490.91620.92680.9160
GSTA [34]MPRNet [41]ARVo [16]FGST [22]DSTNet [29]RVRT [21]Ours
32.5332.7332.8033.3633.7934.3034.56
0.94680.93660.93520.95000.96150.96550.9684
" + }, + { + "type": "table_caption", + "bbox": [ + 0.277, + 0.362, + 0.724, + 0.375 + ], + "angle": 0, + "content": "Table 2: VD-Diff achieves SOTA results on the DVD dataset [33]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.412, + 0.785, + 0.546 + ], + "angle": 0, + "content": "texture and details to maintain overall structure, resulting in over-smooth videos that lose high-frequency detail. Other models ignore global information to preserve high-frequency detail, introducing redundant speckle textures and artifacts that do not match the real video in fast-motion video. In contrast, our VD-Diff has the best visual effect as it can leverage WADT to model long-range dependencies to capture low-frequency information and employ DM to supplement valuable high-frequency information. Therefore, as seen from the visual comparison, our model can restore texture and details while maintaining the overall framework of the video." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.548, + 0.785, + 0.623 + ], + "angle": 0, + "content": "Besides, we conduct a qualitative comparison in terms of the time dimension, as shown in Fig. 7. As can be seen, previous models such as EDVR [36], MPRNet [41], TSP [28], and DSTNet [29] all exhibit unclear outlines, indicating that they cannot guarantee temporal consistency. In contrast, our method effectively maintains temporal consistency." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.624, + 0.787, + 0.73 + ], + "angle": 0, + "content": "Furthermore, we conduct a comparison with the direct application of a pretrained large diffusion model (i.e., LDM [31]) to video deblurring, as shown in Fig. 8. We notice that the videos deblurred by LDM contain a significant number of artifacts and inconsistencies with the content of the original videos, which is attributed to motion blur present in the blurred videos. In contrast, our method leverages the diffusion model and WADT to maximally restore the details in the video while preserving the video content." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.753, + 0.391, + 0.769 + ], + "angle": 0, + "content": "4.5 Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.785, + 0.809 + ], + "angle": 0, + "content": "This section explores the significance of each key component of our proposed VD-Diff. All variants are tested on the GoPro [25] dataset." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Effect of Diffusion Model. To validate the effectiveness of the diffusion model, we design a variant that does not utilize DM to generate prior features" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.355, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "Video Deblurring with Transformer and Diffusion Model" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "table", + "bbox": [ + 0.222, + 0.146, + 0.79, + 0.193 + ], + "angle": 0, + "content": "
MethodDVD [33]DTBN [14]Nah et al. [26]ESTRNN [46]TSP [28]DSTNet [29]Ours
PSNR ↑29.9531.8433.0033.3632.8434.4535.61
SSIM ↑0.86920.91700.93300.93700.93980.95480.9758
" + }, + { + "type": "table_caption", + "bbox": [ + 0.28, + 0.194, + 0.722, + 0.208 + ], + "angle": 0, + "content": "Table 3: VD-Diff achieves SOTA results on the BSD dataset [46]." + }, + { + "type": "table", + "bbox": [ + 0.221, + 0.228, + 0.803, + 0.276 + ], + "angle": 0, + "content": "
MethodDVDEDVRTSPFGSTDSTNetBasicVSR++RVRTDSTNet-LOurs
Parameters (M)15.3023.6016.199.707.459.4613.6016.9611.93
Running time (s)0.882.701.900.720.080.100.410.280.31
" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.278, + 0.785, + 0.319 + ], + "angle": 0, + "content": "Table 4: Quantitative evaluations of the video deblurring methods regarding network parameters and running time. The running time is obtained on a machine with a single NVIDIA RTX A6000 GPU. The size of the test images is \\(1280 \\times 720\\) pixels." + }, + { + "type": "table", + "bbox": [ + 0.233, + 0.359, + 0.77, + 0.456 + ], + "angle": 0, + "content": "
VariantsComponentsMetrics
Diffusion ModelWaveletJoint TrainingWBPFPSNR ↑SSIM ↑
w/o DMX✓✓✓34.380.9684
w/o Wavelet✓X✓✓34.750.9709
w/o Joint Training✓✓X✓34.920.9717
w/o WBPF✓✓✓X33.580.9344
Full model✓✓✓✓35.550.9752
" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.457, + 0.784, + 0.484 + ], + "angle": 0, + "content": "Table 5: Ablation experiments on various variants of VD-Diif on the GoPro [25] dataset, and the best quantitative results are marked in bold." + }, + { + "type": "image", + "bbox": [ + 0.22, + 0.514, + 0.409, + 0.634 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.411, + 0.514, + 0.594, + 0.631 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.597, + 0.515, + 0.784, + 0.634 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.648, + 0.785, + 0.704 + ], + "angle": 0, + "content": "Fig. 9: (a) Floating point operations (FLOPs) vs video deblurring performance on the GoPro [25] dataset. Our model achieves favorable results in terms of accuracy and FLOPs. (b) Ablation study of the number of iteration steps. (c) Ablation study of the sequence length during inference." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.734, + 0.784, + 0.794 + ], + "angle": 0, + "content": "but solely relies on WADT for video deblurring, named as \\( w / o \\) DM. As seen in Table 5, without employing DM, the deblurring performance of the network significantly declines, with a decrease of 1.17 in PSNR, which indicates that DM can provide valuable prior features to enhance the deblurring performance." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.785, + 0.84 + ], + "angle": 0, + "content": "Effect of Wavelet Transform. We conduct ablation experiments to study the impact of wavelet transform on VD-Diff. We replace wavelet transform and inverse wavelet transform with bilinear downsample and upsample to evaluate" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.359, + 0.127 + ], + "angle": 0, + "content": "C. Rao et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.207 + ], + "angle": 0, + "content": "the effect of wavelet transform, named \\( w / o \\) Wavelet. We adopt the same training strategy as VD-Diff. Table 5 shows \\( w / o \\) Wavelet decreases by 0.8dB regarding the PSNR metrics. This experiment shows the wavelet transform method for feature extraction is better than the general upsampling and downsampling methods." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.209, + 0.785, + 0.299 + ], + "angle": 0, + "content": "Effect of Joint Training Strategy. In Sec. 3.4, we analyze the importance of joint training. To verify our analysis, we remove joint training stage i.e., Training Stage Three, as shown in Table 5. We notice that VD-Diff is significantly better than \\( w / o \\) Joint Training with an improvement of 0.63dB in terms of PSNR metrics, which demonstrates that joint training can improve the collaborative working ability of DM and WADT, thereby improving the deblurring effect." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.301, + 0.785, + 0.376 + ], + "angle": 0, + "content": "Effect of Wavelet-based Bidirectional Propagation Fuse. To investigate the effect of Wavelet-based Bidirectional Propagation Fuse (WBPF), we design a variant model by removing the WBPF in WADT, termed as \\( w / o \\) WBPF. Table 5 shows that WBPF can effectively transmit useful information between frames, improving the model's ability to capture spatio-temporal information." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.378, + 0.785, + 0.483 + ], + "angle": 0, + "content": "Effect of Iteration Steps. To explore the impact of the number of iterations on the diffusion model, we set different iteration steps to train the model and then compare the PSNR metrics on the GoPro dataset. The results are shown in Fig. 9(b). The value of PSNR increases as the number of iteration steps increases, but when the number of iterations is greater than 4, the value of PSNR almost remains stable. It shows that in a highly compressed latent space, just setting the number of iteration steps to 4 can generate suitable prior features." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.485, + 0.785, + 0.635 + ], + "angle": 0, + "content": "Effect of Sequence Length. Intuitively, the sequence length will affect the deblurring effect of our proposed model. If the sequence length is too short, the model fails to capture essential information from distant frames. Conversely, if the sequence length is too long, the computational load becomes excessive, potentially causing error accumulation. To investigate this, we experimented with various sequence lengths. As shown in Fig. 9(c), the PSNR value increases steadily with the number of frames. However, after 16 frames, the calculation becomes very large as the sequence length increases, and the PSNR does not change significantly. Therefore, we set the sequence length to 16 to reach a near-optimal point between the deblurring effect and computational reduction." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.666, + 0.359, + 0.681 + ], + "angle": 0, + "content": "5 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.705, + 0.784, + 0.839 + ], + "angle": 0, + "content": "In this paper, we propose a brand-new method that innovatively integrates the diffusion model with the Transformer for video deblurring. Specifically, the diffusion model generates prior features containing high-frequency information that conforms to the ground truth (GT) distribution. Meanwhile, the Wavelet-Aware Dynamic Transformer captures low-frequency information and utilizes the prior features generated by the diffusion model to supplement high-frequency information, thereby enhancing the quality of the deblurred video. Comprehensive experimental results show that our proposed method outperforms the state-of-the-art methods." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.356, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "Video Deblurring with Transformer and Diffusion Model" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.403, + 0.163 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.175, + 0.788, + 0.298 + ], + "angle": 0, + "content": "This work was supported in part by National Social Science Foundation Major Project \"Research on Virtual Restoration of Tang and Song Painting Colors and Construction of Traditional Color Resource Library from a Digital Perspective\" (19ZDA046), Zhejiang Province Program (2022C01222, 2023C03199, 2023C03201), the National Program of China (62172365, 2021YFF0900604, 19ZDA197), Ningbo Science and Technology Plan Project (2022Z167, 2023Z137), and MOE Frontier Science Center for Brain Science & Brain-Machine Integration (Zhejiang University)." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.318, + 0.323, + 0.333 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.348, + 0.785, + 0.374 + ], + "angle": 0, + "content": "1. Ba, J.L., Kiros, J.R., Hinton, G.E.: Layer normalization. arXiv preprint arXiv:1607.06450 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.376, + 0.785, + 0.429 + ], + "angle": 0, + "content": "2. Blattmann, A., Rombach, R., Ling, H., Dockhorn, T., Kim, S.W., Fidler, S., Kreis, K.: Align your latents: High-resolution video synthesis with latent diffusion models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 22563-22575 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.431, + 0.785, + 0.457 + ], + "angle": 0, + "content": "3. Chan, K.C., Zhou, S., Xu, X., Loy, C.C.: On the generalization of basicvsr++ to video deblurring and denoising. arXiv preprint arXiv:2204.05308 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.458, + 0.785, + 0.484 + ], + "angle": 0, + "content": "4. Chen, L., Chu, X., Zhang, X., Sun, J.: Simple baselines for image restoration. In: European Conference on Computer Vision. pp. 17-33. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.485, + 0.785, + 0.525 + ], + "angle": 0, + "content": "5. Chen, Z., Zhang, Y., Liu, D., Xia, B., Gu, J., Kong, L., Yuan, X.: Hierarchical integration diffusion model for realistic image deblurring. arXiv preprint arXiv:2305.12966 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.527, + 0.785, + 0.553 + ], + "angle": 0, + "content": "6. Cho, S., Wang, J., Lee, S.: Video deblurring for hand-held cameras using patch-based synthesis. ACM Transactions on Graphics (TOG) 31(4), 1-9 (2012)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.554, + 0.785, + 0.593 + ], + "angle": 0, + "content": "7. Chu, T., Chen, J., Sun, J., Lian, S., Wang, Z., Zuo, Z., Zhao, L., Xing, W., Lu, D.: Rethinking fast fourier convolution in image inpainting. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 23195-23205 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.595, + 0.785, + 0.649 + ], + "angle": 0, + "content": "8. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.65, + 0.785, + 0.69 + ], + "angle": 0, + "content": "9. Esser, P., Chiu, J., Atighehchian, P., Granskog, J., Germanidis, A.: Structure and content-guided video synthesis with diffusion models. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 7346-7356 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.691, + 0.785, + 0.731 + ], + "angle": 0, + "content": "10. Fan, Y., Hong, C., Wang, X., Zeng, Z., Guo, Z.: Multi-input-output fusion attention module for deblurring networks. In: 2021 IEEE international conference on big data (Big Data). pp. 3176-3182. IEEE (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.732, + 0.785, + 0.758 + ], + "angle": 0, + "content": "11. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. Advances in neural information processing systems 33, 6840-6851 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.759, + 0.785, + 0.799 + ], + "angle": 0, + "content": "12. Huang, H., He, R., Sun, Z., Tan, T.: Wavelet-srnet: A wavelet-based cnn for multiscale face super resolution. In: Proceedings of the IEEE international conference on computer vision. pp. 1689-1697 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.8, + 0.785, + 0.84 + ], + "angle": 0, + "content": "13. Hyun Kim, T., Mu Lee, K.: Generalized video deblurring for dynamic scenes. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 5426-5434 (2015)" + }, + { + "type": "list", + "bbox": [ + 0.22, + 0.348, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.359, + 0.127 + ], + "angle": 0, + "content": "C. Rao et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.787, + 0.189 + ], + "angle": 0, + "content": "14. Hyun Kim, T., Mu Lee, K., Scholkopf, B., Hirsch, M.: Online video deblurring via dynamic temporal blending network. In: Proceedings of the IEEE international conference on computer vision. pp. 4038-4047 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.191, + 0.786, + 0.218 + ], + "angle": 0, + "content": "15. Li, A., Zhao, L., Zuo, Z., Wang, Z., Xing, W., Lu, D.: Migt: Multi-modal image inpainting guided with text. Neurocomputing 520, 376-385 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.219, + 0.786, + 0.274 + ], + "angle": 0, + "content": "16. Li, D., Xu, C., Zhang, K., Yu, X., Zhong, Y., Ren, W., Suominen, H., Li, H.: Arvo: Learning all-range volumetric correspondence for video deblurring. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7721-7731 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.276, + 0.787, + 0.331 + ], + "angle": 0, + "content": "17. Li, G., Lyu, J., Wang, C., Dou, Q., Qin, J.: Wavtrans: Synergizing wavelet and cross-attention transformer for multi-contrast mri super-resolution. In: International Conference on Medical Image Computing and Computer-Assisted Intervention. pp. 463–473. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.332, + 0.787, + 0.374 + ], + "angle": 0, + "content": "18. Li, G., Rao, C., Mo, J., Zhang, Z., Xing, W., Zhao, L.: Rethinking diffusion model for multi-contrast mri super-resolution. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11365-11374 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.375, + 0.787, + 0.416 + ], + "angle": 0, + "content": "19. Li, W., Zhao, L., Lin, Z., Xu, D., Lu, D.: Non-local image inpainting using low-rank matrix completion. In: Computer Graphics Forum. vol. 34, pp. 111-122. Wiley Online Library (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.417, + 0.787, + 0.459 + ], + "angle": 0, + "content": "20. Li, W., Zhao, L., Xu, D., Lu, D.: Efficient image completion method based on alternating direction theory. In: 2013 IEEE International Conference on Image Processing. pp. 700-703. IEEE (2013)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.46, + 0.787, + 0.515 + ], + "angle": 0, + "content": "21. Liang, J., Fan, Y., Xiang, X., Ranjan, R., Ilg, E., Green, S., Cao, J., Zhang, K., Timofte, R., Gool, L.V.: Recurrent video restoration transformer with guided deformable attention. Advances in Neural Information Processing Systems 35, 378-393 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.517, + 0.787, + 0.558 + ], + "angle": 0, + "content": "22. Lin, J., Cai, Y., Hu, X., Wang, H., Yan, Y., Zou, X., Ding, H., Zhang, Y., Timofte, R., Van Gool, L.: Flow-guided sparse transformer for video deblurring. arXiv preprint arXiv:2201.01893 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.559, + 0.787, + 0.613 + ], + "angle": 0, + "content": "23. Liu, X., Park, D.H., Azadi, S., Zhang, G., Chopikyan, A., Hu, Y., Shi, H., Rohrbach, A., Darrell, T.: More control for free! image synthesis with semantic diffusion guidance. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV). pp. 289-299 (January 2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.615, + 0.787, + 0.671 + ], + "angle": 0, + "content": "24. Luo, Z., Chen, D., Zhang, Y., Huang, Y., Wang, L., Shen, Y., Zhao, D., Zhou, J., Tan, T.: Videofusion: Decomposed diffusion models for high-quality video generation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10209-10218 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.672, + 0.787, + 0.713 + ], + "angle": 0, + "content": "25. Nah, S., Hyun Kim, T., Mu Lee, K.: Deep multi-scale convolutional neural network for dynamic scene deblurring. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 3883-3891 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.714, + 0.787, + 0.755 + ], + "angle": 0, + "content": "26. Nah, S., Son, S., Lee, K.M.: Recurrent neural networks with intra-frame iterations for video deblurring. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 8102-8111 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.757, + 0.787, + 0.798 + ], + "angle": 0, + "content": "27. Ozdenizci, O., Legenstein, R.: Restoring vision in adverse weather conditions with patch-based denoising diffusion models. IEEE Transactions on Pattern Analysis and Machine Intelligence (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.799, + 0.787, + 0.841 + ], + "angle": 0, + "content": "28. Pan, J., Bai, H., Tang, J.: Cascaded deep video deblurring using temporal sharpness prior. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 3043-3051 (2020)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.356, + 0.115, + 0.732, + 0.128 + ], + "angle": 0, + "content": "Video Deblurring with Transformer and Diffusion Model" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.189 + ], + "angle": 0, + "content": "29. Pan, J., Xu, B., Dong, J., Ge, J., Tang, J.: Deep discriminative spatial and temporal network for efficient video deblurring. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 22191-22200 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.19, + 0.788, + 0.245 + ], + "angle": 0, + "content": "30. Paszke, A., Gross, S., Massa, F., Lerer, A., Bradbury, J., Chanan, G., Killeen, T., Lin, Z., Gimelshein, N., Antiga, L., et al.: Pytorch: An imperative style, high-performance deep learning library. Advances in neural information processing systems 32 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.246, + 0.788, + 0.286 + ], + "angle": 0, + "content": "31. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10684-10695 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.287, + 0.788, + 0.342 + ], + "angle": 0, + "content": "32. Ruan, L., Ma, Y., Yang, H., He, H., Liu, B., Fu, J., Yuan, N.J., Jin, Q., Guo, B.: Mm-diffusion: Learning multi-modal diffusion models for joint audio and video generation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10219-10228 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.343, + 0.788, + 0.384 + ], + "angle": 0, + "content": "33. Su, S., Delbracio, M., Wang, J., Sapiro, G., Heidrich, W., Wang, O.: Deep video deblurring for hand-held cameras. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1279-1288 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.384, + 0.788, + 0.425 + ], + "angle": 0, + "content": "34. Suin, M., Rajagopalan, A.: Gated spatio-temporal attention-guided video deblurring. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7802-7811 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.425, + 0.788, + 0.467 + ], + "angle": 0, + "content": "35. Tashiro, Y., Song, J., Song, Y., Ermon, S.: Csdi: Conditional score-based diffusion models for probabilistic time series imputation. Advances in Neural Information Processing Systems 34, 24804-24816 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.467, + 0.788, + 0.522 + ], + "angle": 0, + "content": "36. Wang, X., Chan, K.C., Yu, K., Dong, C., Change Loy, C.: Edvr: Video restoration with enhanced deformable convolutional networks. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops. pp. 0-0 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.522, + 0.788, + 0.563 + ], + "angle": 0, + "content": "37. Wang, Z., Cun, X., Bao, J., Zhou, W., Liu, J., Li, H.: Uformer: A general u-shaped transformer for image restoration. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 17683-17693 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.564, + 0.788, + 0.605 + ], + "angle": 0, + "content": "38. Xia, B., Zhang, Y., Wang, S., Wang, Y., Wu, X., Tian, Y., Yang, W., Van Gool, L.: Diffir: Efficient diffusion model for image restoration. arXiv preprint arXiv:2303.09472 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.606, + 0.788, + 0.633 + ], + "angle": 0, + "content": "39. Xiang, X., Wei, H., Pan, J.: Deep video deblurring using sharpness features from exemplars. IEEE Transactions on Image Processing 29, 8976-8987 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.633, + 0.788, + 0.674 + ], + "angle": 0, + "content": "40. Yu, S., Sohn, K., Kim, S., Shin, J.: Video probabilistic diffusion models in projected latent space. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 18456-18466 (June 2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.675, + 0.788, + 0.715 + ], + "angle": 0, + "content": "41. Zamir, S.W., Arora, A., Khan, S., Hayat, M., Khan, F.S., Yang, M.H., Shao, L.: Multi-stage progressive image restoration. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 14821-14831 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.716, + 0.788, + 0.757 + ], + "angle": 0, + "content": "42. Zhang, K., Luo, W., Zhong, Y., Ma, L., Liu, W., Li, H.: Adversarial spatio-temporal learning for video deblurring. IEEE Transactions on Image Processing 28(1), 291-301 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.758, + 0.788, + 0.813 + ], + "angle": 0, + "content": "43. Zhang, Z., Zhang, Q., Xing, W., Li, G., Zhao, L., Sun, J., Lan, Z., Luan, J., Huang, Y., Lin, H.: Artbank: Artistic style transfer with pre-trained diffusion model and implicit style prompt bank. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 38, pp. 7396-7404 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.813, + 0.788, + 0.84 + ], + "angle": 0, + "content": "44. Zhao, L., Mo, Q., Lin, S., Wang, Z., Zuo, Z., Chen, H., Xing, W., Lu, D.: Uctgan: Diverse image inpainting based on unsupervised cross-space translation. In: Pro" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.359, + 0.127 + ], + "angle": 0, + "content": "C.Rao et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.245, + 0.148, + 0.785, + 0.175 + ], + "angle": 0, + "content": "ceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 5741-5750 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.177, + 0.786, + 0.217 + ], + "angle": 0, + "content": "45. Zheng, B., Yuan, S., Slabaugh, G., Leonardis, A.: Image demoireing with learnable bandpass filters. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3636-3645 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.218, + 0.786, + 0.272 + ], + "angle": 0, + "content": "46. Zhong, Z., Gao, Y., Zheng, Y., Zheng, B.: Efficient spatio-temporal recurrent neural network for video deblurring. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part VI 16. pp. 191-207. Springer (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.273, + 0.786, + 0.314 + ], + "angle": 0, + "content": "47. Zhou, S., Zhang, J., Pan, J., Xie, H., Zuo, W., Ren, J.: Spatio-temporal filter adaptive network for video deblurring. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 2482-2491 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.315, + 0.786, + 0.37 + ], + "angle": 0, + "content": "48. Zuo, Z., Zhao, L., Li, A., Wang, Z., Zhang, Z., Chen, J., Xing, W., Lu, D.: Generative image inpainting with segmentation confusion adversarial training and contrastive learning. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 37, pp. 3888-3896 (2023)" + }, + { + "type": "list", + "bbox": [ + 0.218, + 0.148, + 0.786, + 0.37 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/Rethinking Video Deblurring with Wavelet-Aware Dynamic Transformer and Diffusion Model/7afdf756-63a6-48fd-b654-3bb54bd5d976_origin.pdf b/2024/Rethinking Video Deblurring with Wavelet-Aware Dynamic Transformer and Diffusion Model/7afdf756-63a6-48fd-b654-3bb54bd5d976_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..7eb54d46e8224066961500e122b84326fac05578 --- /dev/null +++ b/2024/Rethinking Video Deblurring with Wavelet-Aware Dynamic Transformer and Diffusion Model/7afdf756-63a6-48fd-b654-3bb54bd5d976_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8e7ea04120c635e174db65c0dfa142f7ea7d8ce581e3f7955780974ecb66346c +size 3199604 diff --git a/2024/Rethinking Video Deblurring with Wavelet-Aware Dynamic Transformer and Diffusion Model/full.md b/2024/Rethinking Video Deblurring with Wavelet-Aware Dynamic Transformer and Diffusion Model/full.md new file mode 100644 index 0000000000000000000000000000000000000000..57b88ff67aeeef8ddf05d69a35de8e7fa891e25f --- /dev/null +++ b/2024/Rethinking Video Deblurring with Wavelet-Aware Dynamic Transformer and Diffusion Model/full.md @@ -0,0 +1,478 @@ +# Rethinking Video Deblurring with Wavelet-Aware Dynamic Transformer and Diffusion Model + +Chen Rao $^{1\dagger}$ , Guangyuan Li $^{1\dagger}$ , Zehua Lan $^{1}$ , Jiakai Sun $^{1}$ , Junsheng Luan $^{1}$ , Wei Xing $^{1*}$ , Lei Zhao $^{1*}$ , Huaizhong Lin $^{1*}$ , Jianfeng Dong $^{2}$ , and Dalong Zhang $^{1}$ + +1 College of Computer Science and Technology, Zhejiang University + +$^{2}$ Zhejiang Gongshang University + +Abstract. Current video deblurring methods have limitations in recovering high-frequency information since the regression losses are conservative with high-frequency details. Since Diffusion Models (DMs) have strong capabilities in generating high-frequency details, we consider introducing DMs into the video deblurring task. However, we found that directly applying DMs to the video deblurring task has the following problems: (1) DMs require many iteration steps to generate videos from Gaussian noise, which consumes many computational resources. (2) DMs are easily misled by the blurry artifacts in the video, resulting in irrational content and distortion of the deblurred video. To address the above issues, we propose a novel video deblurring framework VD-Diff that integrates the diffusion model into the Wavelet-Aware Dynamic Transformer (WADT). Specifically, we perform the diffusion model in a highly compact latent space to generate prior features containing high-frequency information that conforms to the ground truth distribution. We design the WADT to preserve and recover the low-frequency information in the video while utilizing the high-frequency information generated by the diffusion model. Extensive experiments show that our proposed VD-Diff outperforms SOTA methods on GoPro, DVD, BSD, and Real-World Video datasets. The codes will be available at https://github.com/Chen-Rao/VD-Diff. + +Keywords: Diffusion model $\cdot$ Prior feature $\cdot$ Video deblurring $\cdot$ Wavelet-based transformer + +# 1 Introduction + +Portable devices are widely used to capture videos. However, camera shake and fast movement of objects lead to undesirable blur in video. To reduce the impact of video blur, researchers have done a lot of work on video deblurring. Traditional methods are mainly based on hand-crafted assumptions and priors, which are difficult to design. Besides, inappropriate priors will reduce the quality of video deblurring. + +In the past decade, video deblurring has witnessed significant progresses with the development of deep learning. Early deep learning methods mainly use CNN-based models [13, 14, 26, 28, 36, 42, 46] to extract video features for capturing the spatio-temporal information between frames. Nevertheless, these approaches have limitations in modelling long-range spatial dependencies and capturing non-local self-similarities. + +Recently, Transformer has been employed in video deblurring tasks [4, 21, 22, 37] to address limitations in CNN-based methods, which can establish long-range spatial dependencies and capture non-local spatial information. Although Transformer-based methods exhibit impressive performance, they still have the following issues. On the one hand, the methods using standard global Transformer [8] tend to deblur videos based on low-frequency details and have limited perception of high-frequency information. On the other hand, when the local window-based Transformer [37] is used to reduce the computational cost, self-attention is calculated within a location-specific window, resulting in a limited receptive field. Consequently, it may ignore some key elements of similar and clearer scene patches in the spatiotemporal neighborhood. Therefore, the local window-based self-attention fails to capture valuable high-frequency information from the spatiotemporal neighborhood, resulting in the loss of high-frequency information. + +Very recently, diffusion models (DMs) have shown their outstanding capabilities in video synthesis tasks [2,9,23,24,31,32,40], which can denoise pure Gaussian noise into high-quality video with abundant high-frequency details. Therefore, we consider employing DMs in video deblurring to restore high-frequency information. However, directly applying DMs to video deblurring tasks has the following problems. First, DMs demand a large number of iteration steps to generate videos from Gaussian noise, which costs a lot of computational resources. Second, DMs are sensitive to motion blur in videos, resulting in irrational content and distortion of the deblurred video. + +To solve the above problems, we propose a novel method called VD-Diff, which combines our proposed Wavelet-Aware Dynamic Transformer (WADT) and diffusion model for video deblurring. Specifically, we employ the diffusion model in a highly compact latent space to generate prior features containing high-frequency information that conforms to the ground truth distribution, which facilitates video deblurring. Applying DM in a highly compressed latent space can significantly reduce the number of denoising iteration steps. + +In addition, we design WADT to retain and restore the low-frequency information in the video while exacting the high-frequency information from the prior features generated by the diffusion model, which ensures the deblurred video is undistorted. WADT first decomposes the blurred video frames into the approximation coefficient and detail coefficients by Wavelet Transform (WT) and extracts low-frequency information from the approximation coefficient by Wavelet-Aware Dynamic Transformer Layers (WADTL). Meanwhile, WADT employs WADTL to fuse the prior features generated by DM containing high-frequency information to generate artifact-free and distortion-free videos. Furthermore, we + +![](images/810cf35c3598d5d428cf97f7db2b4bf6a027e7752d709a6ce97aa11d954f7ffc.jpg) + +![](images/b60cb3f2a0343fc49bdd654572a345deb4375720a04a7c77fc098585e2463ac8.jpg) +Wavelet-Aware Dynamic Transformer Layer + +![](images/7545eba5884d0c00867b75e62d5916668bfffa9eadc11326f8a846f1aefe4025.jpg) +Wavelet-based Bidirectional Propagation Fuse +Fig. 1: The overall architecture of the proposed VD-Diff, which consists of Wavelet-Aware Dynamic Transformer (WADT) and Diffusion Model (DM). Specifically, WADT adopts Wavelet Transform (WT) for feature separation, WADT Layer (WADTL) for deep feature extraction, and Wavelet-based Bidirectional Propagation Fuse (WBPF) for spatio-temporal information propagation between frames. The DM generates prior features to supplement high-frequency information for the deblurring process in WADTL.WT: Wavelet Transform. IWT: Inverse Wavelet Transform. + +![](images/9fc35a009037c40cd4b578e81cbe60655a8d6fb51e8f963eb7eb96a1c05035c0.jpg) +Concatenation + +![](images/17f159f7cd8d1f37efb1d969d1be38115041955441c8167bb3f13b8ec94c4539.jpg) +Element-wise Summation + +introduce a Wavelet-based Bidirectional Propagation Fuse (WBPF) module in WADT to fully use the spatio-temporal information between frames, which ensure that the video is undistorted in the spatio-temporal dimension. Our main contributions are as follows: + +(1) We propose a novel method combining the Wavelet-Aware Dynamic Transformer (WADT) and the diffusion model for video deblurring. To the best of our knowledge, our study is the first to employ DM for video deblurring, which requires only a few iteration steps to restore artifact-free and distortion-free videos. +(2) Our diffusion model can generate prior features containing high-frequency information that conforms to the GT distribution, which facilitates video deblurring. +(3) Our proposed WADT can capture low-frequency information while utilizing the prior features generated by the DM to supplement high-frequency information, which further enhance the quality of deblurred video. +(4) Extensive experiments show that our proposed VD-Diff outperforms SOTA methods on GoPro, DVD, BSD and Real-World Video datasets. + +# 2 Related Work + +# 2.1 Video Deblurring + +With the development of deep learning [20], CNN is widely used in video deblurring. For instance, Zhong et al. [46] propose residual dense blocks and a global spatio-temporal attention module for video deblurring; Pan et al. [28] develop a + +deep CNN model to estimate optical flow from intermediate latent frames. CNN-based methods have driven the development of video deblurring, but CNN has limitations in capturing long-range dependencies. To address these issues, some researchers applied the Transformer to video deblurring and achieved impressive results. For example, Lin et al. [22] propose a method to calculate self-attention by capturing key elements of adjacent frames through optical flow. Liang et al. [21] propose a recurrent video restoration transformer to process local neighboring frames in parallel within a globally recurrent framework. However, the above Transformer-based methods have limited ability to capture high-frequency information, leading to the deblurred videos lacking details. + +# 2.2 Diffusion Models + +DMs have shown excellent performance in the field of synthesis. LDM [31] reaches a near-optimal point between complexity reduction and detail preservation by applying DM in the latent space. Blattmann et al. [2] turns the image generator into a video generator by introducing a temporal dimension to the latent space diffusion model. Artbank [43] utilize textual embedding to learn the visual priors from the images, synthesizing highly realistic artistic stylized images. DMs have also shown strong capabilities in the field of image restoration. For instance, DiffIR [38], Hi-diff [5] and DiffMSR [18] use DM to generate prior representations for image restoration. Although some methods, such as DDIM [45] and LDM [31], have reduced the computational complexity of DM by changing the denoising process or compressing the image, using existing DM for video deblurring still requires unaffordable computational resources. To cope with the above issue, we apply DM in a highly compact latent space to reduce the iteration steps. We were the first to apply diffusion models to video deblurring, and our models achieved outstanding performance. + +# 2.3 Wavelet-based Restoration Methods + +Wavelet is widely used in low-visual restoration tasks [7, 15, 19, 44, 48] as it can separate low-frequency and high-frequency information. For example, Huang et al. [12] propose a wavelet-based CNN for multi-scale face super resolution. WavTrans [17] combines wavelet transform with Transformer for feature extraction and fusion; DSTNet [29] develops a wavelet-based feature propagation method to effectively propagate main structures from long-range frames for better video deblurring. Inspired by the above methods, we proposed Wavelet-Aware Dynamic Transformer, which can extract the global structure of the video and utilize the prior features generated by DM to restore high-frequency details. + +# 3 Methodology + +# 3.1 Overall Architecture + +The overall framework of our proposed VD-Diff is shown in Fig. 1, which consists of two parts: (1) Wavelet-Aware Dynamic Transformer (WADT); (2) Diffusion + +Model. Specifically, we first employ Diffusion Model to obtain a highly compact latent prior feature $z \in \mathbb{R}^{T \times C'}$ , which contains supplementary high-frequency information. Note that $C'$ is the number of channels of latent features. Then, we employ WADT to deblur the blurry video frames $V_{blur} \in R^{T \times 3 \times H \times W}$ into $V_{HQ} \in R^{T \times 3 \times H \times W}$ with the help of the prior feature $z$ . + +# 3.2 Wavelet-Aware Dynamic Transformer + +As shown in the WADT part of Fig. 1, we first utilize Conv3D to perform shallow feature extraction on the blurry video frames $V_{blur} \in R^{T \times 3 \times H \times W}$ to obtain a global feature $F_{in} \in \mathbb{R}^{T \times C \times H \times W}$ . Then, we use Wavelet Transform (WT) to decompose it into approximate coefficients $F_{a} \in \mathbb{R}^{T \times C / 4 \times H \times W}$ and detail coefficients $F_{d} \in \mathbb{R}^{T \times 3C / 4 \times H \times W}$ and again use WT on $F_{a}$ to obtain $F_{aa} \in \mathbb{R}^{T \times \hat{C} \times H \times W}$ and $F_{ad} \in \mathbb{R}^{T \times 3\hat{C} \times H \times W}$ , where $\hat{C} = C / 16$ . Next, we employ WADTL to obtain the deblurred features $F_{out}$ from the approximate coefficients $F_{aa}$ and the prior feature $z'$ . Then, we use WBPF to explore spatiotemporal information further. Finally, we use two IWTs to fuse features to obtain artifact-free and distortion-free videos $V_{HQ}$ . We will elaborate on the structure of WADTL and WBPF as follows. + +Wavelet-Aware Dynamic Transformer Layer. As shown in Fig. 1, we employ a stack of multiple WADT layer (WADTL) modules hierarchically to fuse approximate coefficients $F_{aa}$ and prior features $z'$ . As shown in Fig. 2, WADTL consists of a Wavelet-Aware Dynamic Multi-head Self-Attention (WAD-MSA) and a Wavelet-Aware Dynamic Feed-Forward Network (WAD-FFN). + +In the WAD-MSA part, given a approximate coefficients $F_{aa}$ , we use the highly compact latent prior feature $z'$ to guide video restoration: + +$$ +\hat {F} = W _ {l} ^ {1} z ^ {\prime} \odot L N \left(F _ {a a}\right) + W _ {l} ^ {2} z ^ {\prime}, \tag {1} +$$ + +where $\odot$ indicates element-wise multiplication, $LN$ denotes layer normalization [1], $W_{l}$ represents linear layer. Then, we utilize a 3D convolution layer and a 2D convolution layer to project $\hat{F}$ into query $Q = W_2^Q W_3^Q\hat{F}$ , key $K = W_2^K W_3^K\hat{F}$ , value $V = W_2^V W_3^V\hat{F}$ where $W_{3}$ is the channel-wise 3D convolution and $W_{2}$ is the depth-wise 2D convolution. Note that we use 3D convolution in the $T$ , $H$ , and $W$ dimensions and the same 2D convolution for the $T$ features in the temporal dimension $T$ . Next, we perform the attention calculation: + +$$ +\mathcal {F} _ {\text {o u t}} ^ {\prime} = W _ {a} \hat {V} \cdot \operatorname {S o f t m a x} (\hat {K} \cdot \hat {Q} / \gamma) + F _ {a a}, \tag {2} +$$ + +where $\gamma$ is a learnable scaling parameter. + +In the WDA-FFN part, we first integrate the prior $z'$ into $\mathcal{F}_{out}'$ to obtain $\hat{\mathcal{F}}_{out}'$ . And then we use Conv3D to aggregate temporal information and use Conv2D to aggregate spatial information from neighboring pixels. Besides, we employ a gating mechanism to obtain more valuable information. The overall process of WDA-FFN can be described as: + +$$ +F _ {o u t} = G \left(W _ {2} ^ {1} W _ {3} ^ {1} \hat {\mathcal {F}} _ {o u t} ^ {\prime}\right) \odot W _ {2} ^ {2} W _ {3} ^ {2} \hat {\mathcal {F}} _ {o u t} ^ {\prime} + \hat {\mathcal {F}} _ {o u t} ^ {\prime}, \tag {3} +$$ + +![](images/d28e1dae8032e1ea2cb2493272ec1cc6267ae2a5c209484a50333d34f2f06ca5.jpg) +Fig. 2: The illustration of Wavelet-Aware Dynamic Transformer Layer (WADTL), which consists of Wavelet-Aware Dynamic Multi-head Self-Attention (WAD-MSA) and Wavelet-Aware Dynamic Feed-Forward Network (WAD-FFN). + +![](images/7f1466d096027a7e03b75b2d7e41e271be9fee6ef117286783d77d2db398e0b8.jpg) +Fig. 3: The structure of the Forward Process in the WBPF. The Backward Process has the same network structure as the Forward Process, but the direction of information propagation between frames is opposite. + +where $G$ denotes Gaussian Error Linear Units (GELU). + +Wavelet-based Bidirectional Propagation Fuse. Given the generated features $\{X_{i}\}_{i = 1}^{N}$ by the first Inverse Wavelet Transform (IwT), to explore useful contents from the features of other frames, exiting methods usually simply stack $\{X_{i}\}_{i = 1}^{N}$ or the alignment results of $\{X_{i}\}_{i = 1}^{N}$ . However, directly stacking frames without accurately estimating the features of $\{X_{i}\}_{i = 1}^{N}$ or the alignment results of $\{X_{i}\}_{i = 1}^{N}$ can only partially exploit the valuable information between frames and may also cause the propagation of erroneous information between frames. To further capture more useful long-range spatio-temporal information and reduce the influence of inaccurate information, we propose Wavelet-based Bidirectional Propagation Fuse (WBPF). As shown in Fig. 1, WBPF consists of a forward process and a backward process. + +We take the forward process as an example to introduce the WBPF module in detail, as shown in Fig. 3. First, we connect $Y_{i}$ and $X_{i + 1}$ to get $\bar{X}$ , where $Y_{i}$ is the output of the previous forward process and $X_{i + 1}$ is the input of the current forward process. We then use a convolution layer with LeakyReLU to project $\bar{X}$ into $\hat{X}$ . Then we split $\hat{X}$ into $\hat{X}_1, \hat{X}_2$ along the channel dimension. Next, we + +obtain the fused feature from $\hat{X}_1, \hat{X}_2$ by: + +$$ +F _ {p r o} = \hat {X} _ {1} \odot S \left(W _ {1} \hat {X} _ {1}\right) + \hat {X} _ {2} \odot S \left(W _ {2} \hat {X} _ {2}\right), \tag {4} +$$ + +where $S$ denotes the Sigmoid function and $W$ is the $3 \times 3$ 2D convolution. We utilize AvgPool, ResBlock, MaxPool, and Conv2D to better explore spatial information. The overall process can be described as follows: + +$$ +\bar {Y} _ {i + 1} = F _ {p r o} \otimes \text {u p s a m p l e} (P \left(F _ {p r o}\right)), \tag {5} +$$ + +where $\mathrm{P}$ represents the stack of AvgPool, ResBlock, MaxPool, ResBlock and Conv2D. In the last, we employ 30 ResBlocks on $\bar{Y}_{i + 1}$ to obtain $Y_{i + 1}$ . + +# 3.3 Diffusion Model + +Based on conditional denoising diffusion probabilistic models [11,27,31,35], the DM involves a forward diffusion process and a reverse denoising process. We first adopt the Latent Encoder (LE) on ground truth video and the Condition Encoder (CE) on blurry video to generate the ground truth feature $z \in R^{T \times 4C'}$ and the condition feature $c \in R^{T \times 4C'}$ . Then, we use the diffusion model to generate the prior feature $z' \in R^{T \times 4C'}$ using $c \in R^{T \times 4C'}$ as the condition and $z \in R^{T \times 4C'}$ as the target. We detail the diffusion process as follows. + +In the forward diffusion process, we transform $z$ into Gaussian noise by $T$ iterations, which can be described as follows: + +$$ +q \left(z _ {t} \mid z _ {t - 1}\right) = \mathcal {N} \left(z _ {t}; \sqrt {1 - \beta_ {t}} z _ {t - 1}, \beta_ {t} I\right), \tag {6} +$$ + +where $t = 1,2,\dots,T$ ; $z_0 = z$ ; $\beta_{t} \in (0,1)$ are hyperparameters that control the variance of the noise. Using the notation $\alpha_{t} = 1 - \beta_{t}$ , $\bar{\alpha}_{t} = \prod_{i=1}^{t} \alpha_{i}$ and through iterative derivation with reparameterization, Eq. (6) can be written as: + +$$ +q \left(z _ {T} \mid z _ {0}\right) = \mathcal {N} \left(z _ {T}; \sqrt {\bar {\alpha} _ {T}} z _ {0}, (1 - \bar {\alpha} _ {T}) I\right), \tag {7} +$$ + +that is + +$$ +z _ {T} = \sqrt {\bar {\alpha} _ {T}} z _ {0} + \sqrt {1 - \bar {\alpha} _ {T}} \epsilon . \tag {8} +$$ + +In the reverse denoising process, DM sample a Gaussian random noise $z_{T}$ and then gradually denoise $z_{T}$ to $z_{0}$ : + +$$ +q \left(z _ {t - 1} \mid z _ {t}, z _ {0}\right) = \mathcal {N} \left(z _ {t - 1}; \mu_ {t} \left(z _ {t}, z _ {0}\right), \tilde {\beta} _ {t} I\right), \tag {9} +$$ + +$$ +\mu_ {t} \left(z _ {t}, z _ {0}\right) = \frac {\sqrt {\bar {\alpha} _ {t - 1}} \beta_ {t}}{1 - \bar {\alpha} _ {t}} z _ {0} + \frac {\sqrt {\alpha_ {t}} \left(1 - \bar {\alpha} _ {t - 1}\right)}{1 - \bar {\alpha} _ {t}} z _ {t}. \tag {10} +$$ + +We can use Eq. (8) to represent $z_0$ in Eq. (10) and obtain: + +$$ +\mu_ {t} \left(z _ {t}, z _ {0}\right) = \frac {1}{\sqrt {\alpha_ {t}}} \left(z _ {t} - \epsilon \frac {1 - \alpha_ {t}}{\sqrt {1 - \bar {\alpha} _ {t}}}\right). \tag {11} +$$ + +We remove the variance estimation in Eq. (9): + +$$ +z _ {t - 1} = \frac {1}{\sqrt {\alpha_ {t}}} (z _ {t} - \epsilon \frac {1 - \alpha_ {t}}{\sqrt {1 - \bar {\alpha} _ {t}}}). \tag {12} +$$ + +Then the noise $\epsilon$ is the only uncertain variable. Following previous work [11, 31, 38], we adopt a neural network to predict the noise $\epsilon$ on $z_{t}$ , $c$ and $t$ . + +# 3.4 Training and Inference Strategies + +Inspired by [5,38], we divide the training process into three stages: the first is to train the Latent Encoder with Wavelet-Aware Dynamic Transformer for ground truth feature extraction, the second is to train the Diffusion Model to generate the prior feature, and the third is to jointly train DM and WADT to generate artifact-free and high-realistic videos. + +Training Stage One. The first stage aims to train the Latent Encoder (LE) to extract valuable ground truth prior feature $z$ from ground truth (GT). We train LE and WADT as a whole model by inputting $z$ into WADTL instead of $z'$ and ignoring the condition encoder and the diffusion model in Fig. 1. We employ L1 loss and Multi-scale Frequency Reconstruction (MSFR) [10] loss. L1 loss is defined as: + +$$ +L _ {1} = \sum_ {i = 1} ^ {N} \| V _ {H Q} ^ {i} - V _ {G T} ^ {i} \| _ {1}, \tag {13} +$$ + +where $V_{HQ}^{i}$ is the $i$ -th frame of the deblurred HQ video and $V_{GT}^{i}$ is the $i$ -th frame of the GT video. The MSFR loss measures the L1 distance between multi-scale GT and deblurred frame in the frequency domain as follows: + +$$ +L _ {M S F R} = \sum_ {i = 1} ^ {N} \sum_ {k = 1} ^ {K} \frac {1}{t _ {k}} \| \mathcal {F T} \left(\hat {S} _ {k} ^ {i}\right) - \mathcal {F T} \left(S _ {k} ^ {i}\right) \| _ {1}, \tag {14} +$$ + +where $\mathcal{FT}$ denotes the Fast Fourier Transform that transfers image signal to the frequency domain, $K$ denotes the number of different scales and the $\tilde{S}_k^i$ indicates the image signal of the $i$ -th frame at the $k$ -th scale. The final optimization function of stage one is determined as follows: + +$$ +L _ {d e b l u r} = L _ {1} + \lambda L _ {M S F R}, \tag {15} +$$ + +where we experimentally set $\lambda = 0.1$ . + +Training Stage Two. In the second stage, we first use a Condition Encoder (CE), whose struct is the same as the Latent Encoder (LE), to generate condition feature $c$ from the blur video. Then, we train a diffusion model (DM) to generate prior feature $z'$ using condition feature $c$ . + +We run the complete $T$ iteration reverse processes Eq. (12) to generate prior feature $\hat{z}$ . We denoted this loss as: + +$$ +L _ {d i f f} = \left\| z ^ {\prime} - \hat {z} \right\| _ {1}. \tag {16} +$$ + +![](images/125b4a395c134fea66e38887fda90aaeaa3ae00e30870831ccb096055610a083.jpg) +DSTNet + +![](images/34467c743e513dbf72fb9fe5d571113f7994645057868fe5d75cf547f43893c8.jpg) + +![](images/2e5151c314db403abf170db16f105bd4fc0f14c1f4e81a0dc99cbaae795173d9.jpg) + +![](images/ae14aba8a6e4a35d2d5ca17b145376bb576b748578d7a67baa58ef7c5a0c6e6d.jpg) + +![](images/2619cda6ee0e89f4ac8cd9e4d56599faf8712413b8a1786f16a1eebbd0b5a57a.jpg) + +![](images/970968f7d221ab30225ed5ac812c905974a85ef5e93c54122643e10c0f18fbd6.jpg) + +![](images/3f0a5bb7d668cdbf660b8244ca65c6571e7e0381b541a7e2ece03a8c76060c6c.jpg) +Blur + +![](images/776ad1b560723080c0b2ac9a379706ed3d3c561ac401e4a73d976954e232a9e3.jpg) +EDVR +FGST + +![](images/6b586679166e3f3733f95f917fed921e26e88f46c14f91601545e975a612b8f5.jpg) +MPRNet +DSTNet-L + +![](images/f95625964bf76fc6c1f8d6427bb9ede1fcde619c7ee0a690526481d968348ef1.jpg) +TSP +Ours + +![](images/5ee8b9e951f18e52fbe8fc3f6344e2f02942aab9bfb980ae3b0582688ac032f4.jpg) +SFE +GT + +![](images/b32b1633ef79986b5c7a045015039ba0711f7a8f60abbcc0ef6b3912412a0548.jpg) +Fig. 4: Visual comparison on GoPro [25] dataset. The deblurred results of previous work still contain significant blur effects. Our method generates much clearer frames. +Fig. 5: Visual comparison on DVD [33] dataset. The deblurring effect of our proposed model is significantly better. + +![](images/2fa8e5bed21ca76f4140a3515925d540d5e4b063f5152035b79cec00903bb7ec.jpg) +Blur + +![](images/50c4fcbefc9a5f8999bdf252c08062a8d948324d4f811ff9b34fa2819e49bb5d.jpg) +EDVR + +![](images/7eba2eb8ae04825a2284d4718e4fc2cf449945ff8c3e55dc1f62c8f1ab50d424.jpg) +MPRNet + +![](images/62f0c44f826e22d1e35ee7d785eaa1aaee51c73a570161751696ac0e0e787cb6.jpg) +TSP + +![](images/1cf4cc5f99b5858eeec3b889c7f5f2eac0354255fdaaf2ea55fa975ddabfafaa.jpg) +SFE + +![](images/238a4422bdbd826570a704a88c36ad75f6fc3b5a0454ee4a493fd7e7276b62aa.jpg) +BasicVSR + +![](images/b64927086d94787828917df63b393568bb79074df8fbdb0af4f0d8ad26502add.jpg) +FGST + +![](images/84bf2448fa6734706b9e366fa6f046c00cf374a2e6572f219aaf636ab33b4bf5.jpg) +DSTNet + +![](images/0c03ceb7eb36a91a384be2137c5aa20158089bd4df25fd48a384d4a497ce18ee.jpg) +Ours + +![](images/64a517332c6fb4ed9c341c121ea99be52dbde3531a7da0bdd9847297e63e33fb.jpg) +GT + +Training Stage Three. In Training Stage Two, we train the DM to generate $z'$ to approximate $z$ . However, there inevitably is a slight deviation between the prior features $z'$ and the ground truth prior feature $z$ . To make the DM and WADT work better together to acquire information from different frequency domains for video deblurring and improve the robustness of WADT to the prior feature $z'$ , we jointly train DM and WADT in Training Stage Three. The training loss $L_{total}$ is given by the sum of the deblurring loss $L_{deblur}$ and the diffusion loss $L_{diff}$ : + +$$ +L _ {t o t a l} = L _ {d e b l u r} + L _ {d i f f}. \tag {17} +$$ + +Inference. Given a blurry input video $V_{blur} \in R^{T \times 3 \times H \times W}$ , we first compress $V_{blur}$ into a condition feature $c \in R^{T \times 4C'}$ . Then, the diffusion model generates the prior feature $z'$ from pure Gaussian noise $\epsilon$ and condition feature $c$ . Finally, VD-diff deblurs the blurry video with $z'$ using WADT. + +# 4 Experiments + +# 4.1 Datasets + +GoPro. The GoPro [25] benchmark is composed of over 3,300 blurry-sharp image pairs of dynamic scenes. It is obtained by a high-speed camera. The training and testing subsets are split in proportional to 2:1. + +![](images/e777d341f88b32ac1ab52d0d52bd2067e78c0df1076a2a82e56c3d7d9625ad22.jpg) +Blur + +![](images/4a4c941a0566c193cd96daabeb3ce965bcd00b9991475b4372680fae7a5b8b41.jpg) +DVD [33] + +![](images/7c24433315b348357e2485f62f59b59cc6f8a169698852f57dd758ef6c8b3346.jpg) + +![](images/865029901e1066fc4bdfe4f3d124c0f9b2632a3a30f655f0328800d39e4544cf.jpg) + +![](images/efbac54d69d8f13105e14fb834623f22042b2e63ab5a111fd257ce32e3572668.jpg) +TSP [28] + +![](images/22200fbd5a291171e3eea3f76e1214106d8eb8621ce6d1bd46a69678b4418b35.jpg) +SFE [39] + +![](images/874ecc640e9f669020c65163817202607f6889547c22c72be6580916f9d30b7e.jpg) +EDVR [36] +DSTNet [29] + +![](images/9ea57fc86eb8151b6f223207f911492332e77b338abb92c4d5161a351d8f43b0.jpg) +MPRNet [41] +VD-Diff (Ours) + +![](images/d32b83ab696924aaf341a27841053d837d771479a2920fa3fb7387c23f9fc179.jpg) +Fig. 6: Visual results of VD-Diff and SOTA methods on the real blurry videos of [33]. + +![](images/0decf592cd7d2adbf19f250e3939315c0903a7a7a0679d5242b7fb1c886946d0.jpg) +Fig. 7: Comparison of various methods in terms of time dimension. + +![](images/320c671e9a598d4221bd5131164e7e3a2c5f5fcae13092003561b5688d37e596.jpg) + +![](images/28979012f28514e8b93d683d295c3b0a5187c8a252bcee1b5f9580c7e12dfc2d.jpg) + +![](images/2d9c6f5184ee7c681a4596d8fd8813f4bd2bd35422a1c6b5b936f3e5593312d4.jpg) + +![](images/16f6d5ef7b030d8ae96258ef62e14ca97eec7c053fa8d2133cb070b20cd0a933.jpg) + +![](images/252d7ae0a05d0cc82d61f707d3fd8a092218923bc5775541698e2455b9267448.jpg) + +DVD. The DVD [33] dataset consists of 71 videos with 6,708 blurry-sharp image pairs. It is divided into train/test subsets with 61 videos (5,708 image pairs) and 10 videos (1,000 image pairs). DVD is captured with mobile phones and DSLR at a frame rate of 240 FPS. + +BSD. The Beam-Splitter Deblurring Dataset (BSD) [46] consists of 240 videos with 24k blurry-sharp image pairs. Cross-validation experiments between real-world and synthetic datasets demonstrate the high generality of the BSD dataset. + +Real-World Videos. We evaluate the generality of our proposed approach on the real blurry datasets [6]. Due to the inaccessibility of the ground truth, we only compare VD-Diff with other models on visual results. + +# 4.2 Implementation Details + +We implement VD-Diff in PyTorch [30] using a single NVIDIA RTX A6000 GPU. The training process takes 7 days and 10 hours. We set the batch size as 2, the learning rate as $8 \times 10^{-5}$ and weight decay as $4 \times 10^{-5}$ . Patches at the size of $256 \times 256$ cropped from training frames are fed into the models. The sequence + +![](images/273968429ab4dcf48ba4a49d95efa3888f1806a535940052610198baa32abfd9.jpg) + +![](images/663957c0f29f7d6cdb5aaa055089c788488e9cef9219fb30623a99c8ccd142b8.jpg) + +![](images/5e6d92ab721868fb6249f25e1f95cb4cd8f9920452d84f899096c0c781421582.jpg) + +![](images/b35cde1bf0a75cb1339908d4d57e6ae57b2040669040772fc244c660a9cfd57a.jpg) + +![](images/9e6570368fd84698e1d6ad41df0389c40af7bd47748e59e404739d764f83c3bb.jpg) +Blur + +![](images/9ba28251ab1265077dbc045c9ebe82331ceb5c696f4905d5bcf1f08d6daae92b.jpg) +LDM + +![](images/4c6262d3caf1ae557e900ac308d766f1f9489b47fad83881fad34439d4c028f1.jpg) +Ours + +![](images/14ff7918093f633e6c43de34e33771634f717c61b9c391c681af4640d0b17bfc.jpg) +GT +Fig. 8: Qualitative comparison with pre-trained large diffusion models [31]. + +length is set to 15 in training and 16 in testing. Since we use diffusion within a highly compressed latent space, the denoising iteration steps $T$ only needs 4. We set $C$ to 64 and $C'$ to 256, respectively. In the WADT module, we set $N_1$ and $N_2$ to 7 and 2, respectively, to balance the calculation amount and the deblurring effect. The horizontal and vertical flips are performed for data augmentation. + +# 4.3 Quantitative Results + +We compare our proposed VD-Diff with other SOTA methods, as shown in Tables 1, 2, and 3. As can be seen, our method outperforms SOTA methods on GoPro [25] and DVD [33] benchmark datasets. Specifically, as shown in Table 1, Our method outperforms DSTNet-L, RVRT, and FGST by $0.50\mathrm{dB}$ , $0.63\mathrm{dB}$ , and $2.65\mathrm{dB}$ in PSNR on the GoPro [25] dataset, respectively. Our model achieves SOTA performance on the GoPro [25] dataset. Table 2 shows that our proposed method generates the deblurred videos with higher PSNR and SSIM values on the DVD [33] dataset. As shown in Table 3, we also achieve SOTA on the BSD [46] dataset. As seen from Table 4, our model has a relatively low parameter size and less inference time than other models. Although our model's parameter size and inference time are slightly higher than DSTNet and BasicVSR++, the PSNR metric on the GoPro [25] dataset is at least 1.3dB higher than the two models. As shown in Fig. 9(a), our model achieves favorable results in terms of accuracy and FLOPs. The above results show the effectiveness of our model. + +# 4.4 Qualitative Results + +We provide visual comparisons on GoPro [25], DVD [33], and Real-World Video datasets [33] as shown in Fig. 4, 5, and 6, respectively. Some models sacrifice + +
MethodDVD [33]EDVR [36]STFAN [47]ESTRNN [46]MPRNet [41]TSP [28]
PSNR ↑27.3126.8328.5931.0732.7331.67
SSIM ↑0.82550.84260.86080.90230.93660.9279
BasicVSR [3]NAFNet [4]FGST [22]DSTNet [29]DSTNet-L [29]RVRT [21]Ours
34.0133.7132.9034.1635.0534.9235.55
0.95200.96680.96100.96790.97330.97380.9752
+ +Table 1: VD-Diff achieves SOTA results on the GoPro dataset [25]. + +
MethodDVD [33]EDVR [36]STFAN [47]ESTRNN [46]TSP [28]SEF [39]
PSNR ↑30.0128.5131.1532.0132.1331.71
SSIM ↑0.88770.86370.90490.91620.92680.9160
GSTA [34]MPRNet [41]ARVo [16]FGST [22]DSTNet [29]RVRT [21]Ours
32.5332.7332.8033.3633.7934.3034.56
0.94680.93660.93520.95000.96150.96550.9684
+ +Table 2: VD-Diff achieves SOTA results on the DVD dataset [33]. + +texture and details to maintain overall structure, resulting in over-smooth videos that lose high-frequency detail. Other models ignore global information to preserve high-frequency detail, introducing redundant speckle textures and artifacts that do not match the real video in fast-motion video. In contrast, our VD-Diff has the best visual effect as it can leverage WADT to model long-range dependencies to capture low-frequency information and employ DM to supplement valuable high-frequency information. Therefore, as seen from the visual comparison, our model can restore texture and details while maintaining the overall framework of the video. + +Besides, we conduct a qualitative comparison in terms of the time dimension, as shown in Fig. 7. As can be seen, previous models such as EDVR [36], MPRNet [41], TSP [28], and DSTNet [29] all exhibit unclear outlines, indicating that they cannot guarantee temporal consistency. In contrast, our method effectively maintains temporal consistency. + +Furthermore, we conduct a comparison with the direct application of a pretrained large diffusion model (i.e., LDM [31]) to video deblurring, as shown in Fig. 8. We notice that the videos deblurred by LDM contain a significant number of artifacts and inconsistencies with the content of the original videos, which is attributed to motion blur present in the blurred videos. In contrast, our method leverages the diffusion model and WADT to maximally restore the details in the video while preserving the video content. + +# 4.5 Ablation Study + +This section explores the significance of each key component of our proposed VD-Diff. All variants are tested on the GoPro [25] dataset. + +Effect of Diffusion Model. To validate the effectiveness of the diffusion model, we design a variant that does not utilize DM to generate prior features + +
MethodDVD [33]DTBN [14]Nah et al. [26]ESTRNN [46]TSP [28]DSTNet [29]Ours
PSNR ↑29.9531.8433.0033.3632.8434.4535.61
SSIM ↑0.86920.91700.93300.93700.93980.95480.9758
+ +Table 3: VD-Diff achieves SOTA results on the BSD dataset [46]. + +
MethodDVDEDVRTSPFGSTDSTNetBasicVSR++RVRTDSTNet-LOurs
Parameters (M)15.3023.6016.199.707.459.4613.6016.9611.93
Running time (s)0.882.701.900.720.080.100.410.280.31
+ +Table 4: Quantitative evaluations of the video deblurring methods regarding network parameters and running time. The running time is obtained on a machine with a single NVIDIA RTX A6000 GPU. The size of the test images is $1280 \times 720$ pixels. + +
VariantsComponentsMetrics
Diffusion ModelWaveletJoint TrainingWBPFPSNR ↑SSIM ↑
w/o DMX✓✓✓34.380.9684
w/o Wavelet✓X✓✓34.750.9709
w/o Joint Training✓✓X✓34.920.9717
w/o WBPF✓✓✓X33.580.9344
Full model✓✓✓✓35.550.9752
+ +Table 5: Ablation experiments on various variants of VD-Diif on the GoPro [25] dataset, and the best quantitative results are marked in bold. + +![](images/516623607e2afb29e59b91eac0e5b7b0288129edcf7c9d7e5a21e6b270f36bfd.jpg) +Fig. 9: (a) Floating point operations (FLOPs) vs video deblurring performance on the GoPro [25] dataset. Our model achieves favorable results in terms of accuracy and FLOPs. (b) Ablation study of the number of iteration steps. (c) Ablation study of the sequence length during inference. + +![](images/edec444189fc7d2e64b0ffbcbb461bbf0efd8f498909e5ca7ac29d335e4835c8.jpg) + +![](images/70fcc8cfaa5e45b0bc13359405234be3819008fba7d5859e161ed4bc475d12d1.jpg) + +but solely relies on WADT for video deblurring, named as $w / o$ DM. As seen in Table 5, without employing DM, the deblurring performance of the network significantly declines, with a decrease of 1.17 in PSNR, which indicates that DM can provide valuable prior features to enhance the deblurring performance. + +Effect of Wavelet Transform. We conduct ablation experiments to study the impact of wavelet transform on VD-Diff. We replace wavelet transform and inverse wavelet transform with bilinear downsample and upsample to evaluate + +the effect of wavelet transform, named $w / o$ Wavelet. We adopt the same training strategy as VD-Diff. Table 5 shows $w / o$ Wavelet decreases by 0.8dB regarding the PSNR metrics. This experiment shows the wavelet transform method for feature extraction is better than the general upsampling and downsampling methods. + +Effect of Joint Training Strategy. In Sec. 3.4, we analyze the importance of joint training. To verify our analysis, we remove joint training stage i.e., Training Stage Three, as shown in Table 5. We notice that VD-Diff is significantly better than $w / o$ Joint Training with an improvement of 0.63dB in terms of PSNR metrics, which demonstrates that joint training can improve the collaborative working ability of DM and WADT, thereby improving the deblurring effect. + +Effect of Wavelet-based Bidirectional Propagation Fuse. To investigate the effect of Wavelet-based Bidirectional Propagation Fuse (WBPF), we design a variant model by removing the WBPF in WADT, termed as $w / o$ WBPF. Table 5 shows that WBPF can effectively transmit useful information between frames, improving the model's ability to capture spatio-temporal information. + +Effect of Iteration Steps. To explore the impact of the number of iterations on the diffusion model, we set different iteration steps to train the model and then compare the PSNR metrics on the GoPro dataset. The results are shown in Fig. 9(b). The value of PSNR increases as the number of iteration steps increases, but when the number of iterations is greater than 4, the value of PSNR almost remains stable. It shows that in a highly compressed latent space, just setting the number of iteration steps to 4 can generate suitable prior features. + +Effect of Sequence Length. Intuitively, the sequence length will affect the deblurring effect of our proposed model. If the sequence length is too short, the model fails to capture essential information from distant frames. Conversely, if the sequence length is too long, the computational load becomes excessive, potentially causing error accumulation. To investigate this, we experimented with various sequence lengths. As shown in Fig. 9(c), the PSNR value increases steadily with the number of frames. However, after 16 frames, the calculation becomes very large as the sequence length increases, and the PSNR does not change significantly. Therefore, we set the sequence length to 16 to reach a near-optimal point between the deblurring effect and computational reduction. + +# 5 Conclusion + +In this paper, we propose a brand-new method that innovatively integrates the diffusion model with the Transformer for video deblurring. Specifically, the diffusion model generates prior features containing high-frequency information that conforms to the ground truth (GT) distribution. Meanwhile, the Wavelet-Aware Dynamic Transformer captures low-frequency information and utilizes the prior features generated by the diffusion model to supplement high-frequency information, thereby enhancing the quality of the deblurred video. Comprehensive experimental results show that our proposed method outperforms the state-of-the-art methods. + +# Acknowledgements + +This work was supported in part by National Social Science Foundation Major Project "Research on Virtual Restoration of Tang and Song Painting Colors and Construction of Traditional Color Resource Library from a Digital Perspective" (19ZDA046), Zhejiang Province Program (2022C01222, 2023C03199, 2023C03201), the National Program of China (62172365, 2021YFF0900604, 19ZDA197), Ningbo Science and Technology Plan Project (2022Z167, 2023Z137), and MOE Frontier Science Center for Brain Science & Brain-Machine Integration (Zhejiang University). + +# References + +1. Ba, J.L., Kiros, J.R., Hinton, G.E.: Layer normalization. arXiv preprint arXiv:1607.06450 (2016) +2. Blattmann, A., Rombach, R., Ling, H., Dockhorn, T., Kim, S.W., Fidler, S., Kreis, K.: Align your latents: High-resolution video synthesis with latent diffusion models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 22563-22575 (2023) +3. Chan, K.C., Zhou, S., Xu, X., Loy, C.C.: On the generalization of basicvsr++ to video deblurring and denoising. arXiv preprint arXiv:2204.05308 (2022) +4. Chen, L., Chu, X., Zhang, X., Sun, J.: Simple baselines for image restoration. In: European Conference on Computer Vision. pp. 17-33. Springer (2022) +5. Chen, Z., Zhang, Y., Liu, D., Xia, B., Gu, J., Kong, L., Yuan, X.: Hierarchical integration diffusion model for realistic image deblurring. arXiv preprint arXiv:2305.12966 (2023) +6. Cho, S., Wang, J., Lee, S.: Video deblurring for hand-held cameras using patch-based synthesis. ACM Transactions on Graphics (TOG) 31(4), 1-9 (2012) +7. Chu, T., Chen, J., Sun, J., Lian, S., Wang, Z., Zuo, Z., Zhao, L., Xing, W., Lu, D.: Rethinking fast fourier convolution in image inpainting. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 23195-23205 (2023) +8. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929 (2020) +9. Esser, P., Chiu, J., Atighehchian, P., Granskog, J., Germanidis, A.: Structure and content-guided video synthesis with diffusion models. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 7346-7356 (2023) +10. Fan, Y., Hong, C., Wang, X., Zeng, Z., Guo, Z.: Multi-input-output fusion attention module for deblurring networks. In: 2021 IEEE international conference on big data (Big Data). pp. 3176-3182. IEEE (2021) +11. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. Advances in neural information processing systems 33, 6840-6851 (2020) +12. Huang, H., He, R., Sun, Z., Tan, T.: Wavelet-srnet: A wavelet-based cnn for multiscale face super resolution. In: Proceedings of the IEEE international conference on computer vision. pp. 1689-1697 (2017) +13. Hyun Kim, T., Mu Lee, K.: Generalized video deblurring for dynamic scenes. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 5426-5434 (2015) + +14. Hyun Kim, T., Mu Lee, K., Scholkopf, B., Hirsch, M.: Online video deblurring via dynamic temporal blending network. In: Proceedings of the IEEE international conference on computer vision. pp. 4038-4047 (2017) +15. Li, A., Zhao, L., Zuo, Z., Wang, Z., Xing, W., Lu, D.: Migt: Multi-modal image inpainting guided with text. Neurocomputing 520, 376-385 (2023) +16. Li, D., Xu, C., Zhang, K., Yu, X., Zhong, Y., Ren, W., Suominen, H., Li, H.: Arvo: Learning all-range volumetric correspondence for video deblurring. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7721-7731 (2021) +17. Li, G., Lyu, J., Wang, C., Dou, Q., Qin, J.: Wavtrans: Synergizing wavelet and cross-attention transformer for multi-contrast mri super-resolution. In: International Conference on Medical Image Computing and Computer-Assisted Intervention. pp. 463–473. Springer (2022) +18. Li, G., Rao, C., Mo, J., Zhang, Z., Xing, W., Zhao, L.: Rethinking diffusion model for multi-contrast mri super-resolution. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11365-11374 (2024) +19. Li, W., Zhao, L., Lin, Z., Xu, D., Lu, D.: Non-local image inpainting using low-rank matrix completion. In: Computer Graphics Forum. vol. 34, pp. 111-122. Wiley Online Library (2015) +20. Li, W., Zhao, L., Xu, D., Lu, D.: Efficient image completion method based on alternating direction theory. In: 2013 IEEE International Conference on Image Processing. pp. 700-703. IEEE (2013) +21. Liang, J., Fan, Y., Xiang, X., Ranjan, R., Ilg, E., Green, S., Cao, J., Zhang, K., Timofte, R., Gool, L.V.: Recurrent video restoration transformer with guided deformable attention. Advances in Neural Information Processing Systems 35, 378-393 (2022) +22. Lin, J., Cai, Y., Hu, X., Wang, H., Yan, Y., Zou, X., Ding, H., Zhang, Y., Timofte, R., Van Gool, L.: Flow-guided sparse transformer for video deblurring. arXiv preprint arXiv:2201.01893 (2022) +23. Liu, X., Park, D.H., Azadi, S., Zhang, G., Chopikyan, A., Hu, Y., Shi, H., Rohrbach, A., Darrell, T.: More control for free! image synthesis with semantic diffusion guidance. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV). pp. 289-299 (January 2023) +24. Luo, Z., Chen, D., Zhang, Y., Huang, Y., Wang, L., Shen, Y., Zhao, D., Zhou, J., Tan, T.: Videofusion: Decomposed diffusion models for high-quality video generation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10209-10218 (2023) +25. Nah, S., Hyun Kim, T., Mu Lee, K.: Deep multi-scale convolutional neural network for dynamic scene deblurring. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 3883-3891 (2017) +26. Nah, S., Son, S., Lee, K.M.: Recurrent neural networks with intra-frame iterations for video deblurring. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 8102-8111 (2019) +27. Ozdenizci, O., Legenstein, R.: Restoring vision in adverse weather conditions with patch-based denoising diffusion models. IEEE Transactions on Pattern Analysis and Machine Intelligence (2023) +28. Pan, J., Bai, H., Tang, J.: Cascaded deep video deblurring using temporal sharpness prior. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 3043-3051 (2020) + +29. Pan, J., Xu, B., Dong, J., Ge, J., Tang, J.: Deep discriminative spatial and temporal network for efficient video deblurring. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 22191-22200 (2023) +30. Paszke, A., Gross, S., Massa, F., Lerer, A., Bradbury, J., Chanan, G., Killeen, T., Lin, Z., Gimelshein, N., Antiga, L., et al.: Pytorch: An imperative style, high-performance deep learning library. Advances in neural information processing systems 32 (2019) +31. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10684-10695 (2022) +32. Ruan, L., Ma, Y., Yang, H., He, H., Liu, B., Fu, J., Yuan, N.J., Jin, Q., Guo, B.: Mm-diffusion: Learning multi-modal diffusion models for joint audio and video generation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10219-10228 (2023) +33. Su, S., Delbracio, M., Wang, J., Sapiro, G., Heidrich, W., Wang, O.: Deep video deblurring for hand-held cameras. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1279-1288 (2017) +34. Suin, M., Rajagopalan, A.: Gated spatio-temporal attention-guided video deblurring. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7802-7811 (2021) +35. Tashiro, Y., Song, J., Song, Y., Ermon, S.: Csdi: Conditional score-based diffusion models for probabilistic time series imputation. Advances in Neural Information Processing Systems 34, 24804-24816 (2021) +36. Wang, X., Chan, K.C., Yu, K., Dong, C., Change Loy, C.: Edvr: Video restoration with enhanced deformable convolutional networks. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops. pp. 0-0 (2019) +37. Wang, Z., Cun, X., Bao, J., Zhou, W., Liu, J., Li, H.: Uformer: A general u-shaped transformer for image restoration. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 17683-17693 (2022) +38. Xia, B., Zhang, Y., Wang, S., Wang, Y., Wu, X., Tian, Y., Yang, W., Van Gool, L.: Diffir: Efficient diffusion model for image restoration. arXiv preprint arXiv:2303.09472 (2023) +39. Xiang, X., Wei, H., Pan, J.: Deep video deblurring using sharpness features from exemplars. IEEE Transactions on Image Processing 29, 8976-8987 (2020) +40. Yu, S., Sohn, K., Kim, S., Shin, J.: Video probabilistic diffusion models in projected latent space. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 18456-18466 (June 2023) +41. Zamir, S.W., Arora, A., Khan, S., Hayat, M., Khan, F.S., Yang, M.H., Shao, L.: Multi-stage progressive image restoration. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 14821-14831 (2021) +42. Zhang, K., Luo, W., Zhong, Y., Ma, L., Liu, W., Li, H.: Adversarial spatio-temporal learning for video deblurring. IEEE Transactions on Image Processing 28(1), 291-301 (2018) +43. Zhang, Z., Zhang, Q., Xing, W., Li, G., Zhao, L., Sun, J., Lan, Z., Luan, J., Huang, Y., Lin, H.: Artbank: Artistic style transfer with pre-trained diffusion model and implicit style prompt bank. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 38, pp. 7396-7404 (2024) +44. Zhao, L., Mo, Q., Lin, S., Wang, Z., Zuo, Z., Chen, H., Xing, W., Lu, D.: Uctgan: Diverse image inpainting based on unsupervised cross-space translation. In: Pro + +ceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 5741-5750 (2020) +45. Zheng, B., Yuan, S., Slabaugh, G., Leonardis, A.: Image demoireing with learnable bandpass filters. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3636-3645 (2020) +46. Zhong, Z., Gao, Y., Zheng, Y., Zheng, B.: Efficient spatio-temporal recurrent neural network for video deblurring. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part VI 16. pp. 191-207. Springer (2020) +47. Zhou, S., Zhang, J., Pan, J., Xie, H., Zuo, W., Ren, J.: Spatio-temporal filter adaptive network for video deblurring. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 2482-2491 (2019) +48. Zuo, Z., Zhao, L., Li, A., Wang, Z., Zhang, Z., Chen, J., Xing, W., Lu, D.: Generative image inpainting with segmentation confusion adversarial training and contrastive learning. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 37, pp. 3888-3896 (2023) \ No newline at end of file diff --git a/2024/Rethinking Video Deblurring with Wavelet-Aware Dynamic Transformer and Diffusion Model/images.zip b/2024/Rethinking Video Deblurring with Wavelet-Aware Dynamic Transformer and Diffusion Model/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..af211404d3860a9e01143040b09c6175cd49823a --- /dev/null +++ b/2024/Rethinking Video Deblurring with Wavelet-Aware Dynamic Transformer and Diffusion Model/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:11226daa6c2049a1854ba32a672679c6219849ce5d575ad3f338ed6de61f171d +size 664086 diff --git a/2024/Rethinking Video Deblurring with Wavelet-Aware Dynamic Transformer and Diffusion Model/layout.json b/2024/Rethinking Video Deblurring with Wavelet-Aware Dynamic Transformer and Diffusion Model/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..9f8b870f97f2da1e757d5f0c5769e1bc613c47f7 --- /dev/null +++ b/2024/Rethinking Video Deblurring with Wavelet-Aware Dynamic Transformer and Diffusion Model/layout.json @@ -0,0 +1,13948 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 133, + 111, + 481, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 111, + 481, + 148 + ], + "spans": [ + { + "bbox": [ + 133, + 111, + 481, + 148 + ], + "type": "text", + "content": "Rethinking Video Deblurring with Wavelet-Aware Dynamic Transformer and Diffusion Model" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 136, + 167, + 477, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 167, + 477, + 205 + ], + "spans": [ + { + "bbox": [ + 136, + 167, + 477, + 205 + ], + "type": "text", + "content": "Chen Rao" + }, + { + "bbox": [ + 136, + 167, + 477, + 205 + ], + "type": "inline_equation", + "content": "^{1\\dagger}" + }, + { + "bbox": [ + 136, + 167, + 477, + 205 + ], + "type": "text", + "content": ", Guangyuan Li" + }, + { + "bbox": [ + 136, + 167, + 477, + 205 + ], + "type": "inline_equation", + "content": "^{1\\dagger}" + }, + { + "bbox": [ + 136, + 167, + 477, + 205 + ], + "type": "text", + "content": ", Zehua Lan" + }, + { + "bbox": [ + 136, + 167, + 477, + 205 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 136, + 167, + 477, + 205 + ], + "type": "text", + "content": ", Jiakai Sun" + }, + { + "bbox": [ + 136, + 167, + 477, + 205 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 136, + 167, + 477, + 205 + ], + "type": "text", + "content": ", Junsheng Luan" + }, + { + "bbox": [ + 136, + 167, + 477, + 205 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 136, + 167, + 477, + 205 + ], + "type": "text", + "content": ", Wei Xing" + }, + { + "bbox": [ + 136, + 167, + 477, + 205 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 136, + 167, + 477, + 205 + ], + "type": "text", + "content": ", Lei Zhao" + }, + { + "bbox": [ + 136, + 167, + 477, + 205 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 136, + 167, + 477, + 205 + ], + "type": "text", + "content": ", Huaizhong Lin" + }, + { + "bbox": [ + 136, + 167, + 477, + 205 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 136, + 167, + 477, + 205 + ], + "type": "text", + "content": ", Jianfeng Dong" + }, + { + "bbox": [ + 136, + 167, + 477, + 205 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 136, + 167, + 477, + 205 + ], + "type": "text", + "content": ", and Dalong Zhang" + }, + { + "bbox": [ + 136, + 167, + 477, + 205 + ], + "type": "inline_equation", + "content": "^{1}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 167, + 213, + 444, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 167, + 213, + 444, + 224 + ], + "spans": [ + { + "bbox": [ + 167, + 213, + 444, + 224 + ], + "type": "text", + "content": "1 College of Computer Science and Technology, Zhejiang University" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 238, + 224, + 375, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 238, + 224, + 375, + 236 + ], + "spans": [ + { + "bbox": [ + 238, + 224, + 375, + 236 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 238, + 224, + 375, + 236 + ], + "type": "text", + "content": " Zhejiang Gongshang University" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 160, + 266, + 455, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 266, + 455, + 495 + ], + "spans": [ + { + "bbox": [ + 160, + 266, + 455, + 495 + ], + "type": "text", + "content": "Abstract. Current video deblurring methods have limitations in recovering high-frequency information since the regression losses are conservative with high-frequency details. Since Diffusion Models (DMs) have strong capabilities in generating high-frequency details, we consider introducing DMs into the video deblurring task. However, we found that directly applying DMs to the video deblurring task has the following problems: (1) DMs require many iteration steps to generate videos from Gaussian noise, which consumes many computational resources. (2) DMs are easily misled by the blurry artifacts in the video, resulting in irrational content and distortion of the deblurred video. To address the above issues, we propose a novel video deblurring framework VD-Diff that integrates the diffusion model into the Wavelet-Aware Dynamic Transformer (WADT). Specifically, we perform the diffusion model in a highly compact latent space to generate prior features containing high-frequency information that conforms to the ground truth distribution. We design the WADT to preserve and recover the low-frequency information in the video while utilizing the high-frequency information generated by the diffusion model. Extensive experiments show that our proposed VD-Diff outperforms SOTA methods on GoPro, DVD, BSD, and Real-World Video datasets. The codes will be available at https://github.com/Chen-Rao/VD-Diff." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 159, + 506, + 453, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 506, + 453, + 529 + ], + "spans": [ + { + "bbox": [ + 159, + 506, + 453, + 529 + ], + "type": "text", + "content": "Keywords: Diffusion model " + }, + { + "bbox": [ + 159, + 506, + 453, + 529 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 506, + 453, + 529 + ], + "type": "text", + "content": " Prior feature " + }, + { + "bbox": [ + 159, + 506, + 453, + 529 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 506, + 453, + 529 + ], + "type": "text", + "content": " Video deblurring " + }, + { + "bbox": [ + 159, + 506, + 453, + 529 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 506, + 453, + 529 + ], + "type": "text", + "content": " Wavelet-based transformer" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 549, + 230, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 549, + 230, + 562 + ], + "spans": [ + { + "bbox": [ + 132, + 549, + 230, + 562 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 574, + 482, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 574, + 482, + 647 + ], + "spans": [ + { + "bbox": [ + 130, + 574, + 482, + 647 + ], + "type": "text", + "content": "Portable devices are widely used to capture videos. However, camera shake and fast movement of objects lead to undesirable blur in video. To reduce the impact of video blur, researchers have done a lot of work on video deblurring. Traditional methods are mainly based on hand-crafted assumptions and priors, which are difficult to design. Besides, inappropriate priors will reduce the quality of video deblurring." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 141, + 653, + 394, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 653, + 394, + 665 + ], + "spans": [ + { + "bbox": [ + 141, + 653, + 394, + 665 + ], + "type": "text", + "content": "† co-first authors, equal contribution; * Corresponding author" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 187 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 187 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 187 + ], + "type": "text", + "content": "In the past decade, video deblurring has witnessed significant progresses with the development of deep learning. Early deep learning methods mainly use CNN-based models [13, 14, 26, 28, 36, 42, 46] to extract video features for capturing the spatio-temporal information between frames. Nevertheless, these approaches have limitations in modelling long-range spatial dependencies and capturing non-local self-similarities." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 190, + 482, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 190, + 482, + 357 + ], + "spans": [ + { + "bbox": [ + 130, + 190, + 482, + 357 + ], + "type": "text", + "content": "Recently, Transformer has been employed in video deblurring tasks [4, 21, 22, 37] to address limitations in CNN-based methods, which can establish long-range spatial dependencies and capture non-local spatial information. Although Transformer-based methods exhibit impressive performance, they still have the following issues. On the one hand, the methods using standard global Transformer [8] tend to deblur videos based on low-frequency details and have limited perception of high-frequency information. On the other hand, when the local window-based Transformer [37] is used to reduce the computational cost, self-attention is calculated within a location-specific window, resulting in a limited receptive field. Consequently, it may ignore some key elements of similar and clearer scene patches in the spatiotemporal neighborhood. Therefore, the local window-based self-attention fails to capture valuable high-frequency information from the spatiotemporal neighborhood, resulting in the loss of high-frequency information." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 361, + 482, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 361, + 482, + 468 + ], + "spans": [ + { + "bbox": [ + 130, + 361, + 482, + 468 + ], + "type": "text", + "content": "Very recently, diffusion models (DMs) have shown their outstanding capabilities in video synthesis tasks [2,9,23,24,31,32,40], which can denoise pure Gaussian noise into high-quality video with abundant high-frequency details. Therefore, we consider employing DMs in video deblurring to restore high-frequency information. However, directly applying DMs to video deblurring tasks has the following problems. First, DMs demand a large number of iteration steps to generate videos from Gaussian noise, which costs a lot of computational resources. Second, DMs are sensitive to motion blur in videos, resulting in irrational content and distortion of the deblurred video." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 472, + 482, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 472, + 482, + 555 + ], + "spans": [ + { + "bbox": [ + 130, + 472, + 482, + 555 + ], + "type": "text", + "content": "To solve the above problems, we propose a novel method called VD-Diff, which combines our proposed Wavelet-Aware Dynamic Transformer (WADT) and diffusion model for video deblurring. Specifically, we employ the diffusion model in a highly compact latent space to generate prior features containing high-frequency information that conforms to the ground truth distribution, which facilitates video deblurring. Applying DM in a highly compressed latent space can significantly reduce the number of denoising iteration steps." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 558, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 558, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 558, + 482, + 665 + ], + "type": "text", + "content": "In addition, we design WADT to retain and restore the low-frequency information in the video while exacting the high-frequency information from the prior features generated by the diffusion model, which ensures the deblurred video is undistorted. WADT first decomposes the blurred video frames into the approximation coefficient and detail coefficients by Wavelet Transform (WT) and extracts low-frequency information from the approximation coefficient by Wavelet-Aware Dynamic Transformer Layers (WADTL). Meanwhile, WADT employs WADTL to fuse the prior features generated by DM containing high-frequency information to generate artifact-free and distortion-free videos. Furthermore, we" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 102 + ], + "type": "text", + "content": "C. Rao et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 132, + 114, + 479, + 239 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 479, + 239 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 479, + 239 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 479, + 239 + ], + "type": "image", + "image_path": "810cf35c3598d5d428cf97f7db2b4bf6a027e7752d709a6ce97aa11d954f7ffc.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 135, + 243, + 145, + 253 + ], + "blocks": [ + { + "bbox": [ + 135, + 243, + 145, + 253 + ], + "lines": [ + { + "bbox": [ + 135, + 243, + 145, + 253 + ], + "spans": [ + { + "bbox": [ + 135, + 243, + 145, + 253 + ], + "type": "image", + "image_path": "b60cb3f2a0343fc49bdd654572a345deb4375720a04a7c77fc098585e2463ac8.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 146, + 244, + 238, + 251 + ], + "lines": [ + { + "bbox": [ + 146, + 244, + 238, + 251 + ], + "spans": [ + { + "bbox": [ + 146, + 244, + 238, + 251 + ], + "type": "text", + "content": "Wavelet-Aware Dynamic Transformer Layer" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 246, + 244, + 256, + 252 + ], + "blocks": [ + { + "bbox": [ + 246, + 244, + 256, + 252 + ], + "lines": [ + { + "bbox": [ + 246, + 244, + 256, + 252 + ], + "spans": [ + { + "bbox": [ + 246, + 244, + 256, + 252 + ], + "type": "image", + "image_path": "7545eba5884d0c00867b75e62d5916668bfffa9eadc11326f8a846f1aefe4025.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 257, + 244, + 354, + 252 + ], + "lines": [ + { + "bbox": [ + 257, + 244, + 354, + 252 + ], + "spans": [ + { + "bbox": [ + 257, + 244, + 354, + 252 + ], + "type": "text", + "content": "Wavelet-based Bidirectional Propagation Fuse" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 264, + 480, + 341 + ], + "lines": [ + { + "bbox": [ + 130, + 264, + 480, + 341 + ], + "spans": [ + { + "bbox": [ + 130, + 264, + 480, + 341 + ], + "type": "text", + "content": "Fig. 1: The overall architecture of the proposed VD-Diff, which consists of Wavelet-Aware Dynamic Transformer (WADT) and Diffusion Model (DM). Specifically, WADT adopts Wavelet Transform (WT) for feature separation, WADT Layer (WADTL) for deep feature extraction, and Wavelet-based Bidirectional Propagation Fuse (WBPF) for spatio-temporal information propagation between frames. The DM generates prior features to supplement high-frequency information for the deblurring process in WADTL.WT: Wavelet Transform. IWT: Inverse Wavelet Transform." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 365, + 244, + 374, + 252 + ], + "blocks": [ + { + "bbox": [ + 365, + 244, + 374, + 252 + ], + "lines": [ + { + "bbox": [ + 365, + 244, + 374, + 252 + ], + "spans": [ + { + "bbox": [ + 365, + 244, + 374, + 252 + ], + "type": "image", + "image_path": "9fc35a009037c40cd4b578e81cbe60655a8d6fb51e8f963eb7eb96a1c05035c0.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 375, + 244, + 405, + 251 + ], + "lines": [ + { + "bbox": [ + 375, + 244, + 405, + 251 + ], + "spans": [ + { + "bbox": [ + 375, + 244, + 405, + 251 + ], + "type": "text", + "content": "Concatenation" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 410, + 244, + 420, + 252 + ], + "blocks": [ + { + "bbox": [ + 410, + 244, + 420, + 252 + ], + "lines": [ + { + "bbox": [ + 410, + 244, + 420, + 252 + ], + "spans": [ + { + "bbox": [ + 410, + 244, + 420, + 252 + ], + "type": "image", + "image_path": "17f159f7cd8d1f37efb1d969d1be38115041955441c8167bb3f13b8ec94c4539.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 422, + 245, + 476, + 251 + ], + "lines": [ + { + "bbox": [ + 422, + 245, + 476, + 251 + ], + "spans": [ + { + "bbox": [ + 422, + 245, + 476, + 251 + ], + "type": "text", + "content": "Element-wise Summation" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 365, + 479, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 365, + 479, + 411 + ], + "spans": [ + { + "bbox": [ + 130, + 365, + 479, + 411 + ], + "type": "text", + "content": "introduce a Wavelet-based Bidirectional Propagation Fuse (WBPF) module in WADT to fully use the spatio-temporal information between frames, which ensure that the video is undistorted in the spatio-temporal dimension. Our main contributions are as follows:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 413, + 483, + 567 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 130, + 413, + 479, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 413, + 479, + 471 + ], + "spans": [ + { + "bbox": [ + 130, + 413, + 479, + 471 + ], + "type": "text", + "content": "(1) We propose a novel method combining the Wavelet-Aware Dynamic Transformer (WADT) and the diffusion model for video deblurring. To the best of our knowledge, our study is the first to employ DM for video deblurring, which requires only a few iteration steps to restore artifact-free and distortion-free videos." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 472, + 479, + 507 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 472, + 479, + 507 + ], + "spans": [ + { + "bbox": [ + 130, + 472, + 479, + 507 + ], + "type": "text", + "content": "(2) Our diffusion model can generate prior features containing high-frequency information that conforms to the GT distribution, which facilitates video deblurring." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 508, + 479, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 508, + 479, + 544 + ], + "spans": [ + { + "bbox": [ + 130, + 508, + 479, + 544 + ], + "type": "text", + "content": "(3) Our proposed WADT can capture low-frequency information while utilizing the prior features generated by the DM to supplement high-frequency information, which further enhance the quality of deblurred video." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 544, + 483, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 544, + 483, + 567 + ], + "spans": [ + { + "bbox": [ + 130, + 544, + 483, + 567 + ], + "type": "text", + "content": "(4) Extensive experiments show that our proposed VD-Diff outperforms SOTA methods on GoPro, DVD, BSD and Real-World Video datasets." + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 131, + 586, + 237, + 598 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 586, + 237, + 598 + ], + "spans": [ + { + "bbox": [ + 131, + 586, + 237, + 598 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 131, + 610, + 249, + 623 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 610, + 249, + 623 + ], + "spans": [ + { + "bbox": [ + 131, + 610, + 249, + 623 + ], + "type": "text", + "content": "2.1 Video Deblurring" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 130, + 629, + 479, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 479, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 479, + 666 + ], + "type": "text", + "content": "With the development of deep learning [20], CNN is widely used in video deblurring. For instance, Zhong et al. [46] propose residual dense blocks and a global spatio-temporal attention module for video deblurring; Pan et al. [28] develop a" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 217, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 217, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 217, + 91, + 447, + 102 + ], + "type": "text", + "content": "Video Deblurring with Transformer and Diffusion Model" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "text", + "content": "deep CNN model to estimate optical flow from intermediate latent frames. CNN-based methods have driven the development of video deblurring, but CNN has limitations in capturing long-range dependencies. To address these issues, some researchers applied the Transformer to video deblurring and achieved impressive results. For example, Lin et al. [22] propose a method to calculate self-attention by capturing key elements of adjacent frames through optical flow. Liang et al. [21] propose a recurrent video restoration transformer to process local neighboring frames in parallel within a globally recurrent framework. However, the above Transformer-based methods have limited ability to capture high-frequency information, leading to the deblurred videos lacking details." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 251, + 247, + 262 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 251, + 247, + 262 + ], + "spans": [ + { + "bbox": [ + 131, + 251, + 247, + 262 + ], + "type": "text", + "content": "2.2 Diffusion Models" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 267, + 482, + 448 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 267, + 482, + 448 + ], + "spans": [ + { + "bbox": [ + 130, + 267, + 482, + 448 + ], + "type": "text", + "content": "DMs have shown excellent performance in the field of synthesis. LDM [31] reaches a near-optimal point between complexity reduction and detail preservation by applying DM in the latent space. Blattmann et al. [2] turns the image generator into a video generator by introducing a temporal dimension to the latent space diffusion model. Artbank [43] utilize textual embedding to learn the visual priors from the images, synthesizing highly realistic artistic stylized images. DMs have also shown strong capabilities in the field of image restoration. For instance, DiffIR [38], Hi-diff [5] and DiffMSR [18] use DM to generate prior representations for image restoration. Although some methods, such as DDIM [45] and LDM [31], have reduced the computational complexity of DM by changing the denoising process or compressing the image, using existing DM for video deblurring still requires unaffordable computational resources. To cope with the above issue, we apply DM in a highly compact latent space to reduce the iteration steps. We were the first to apply diffusion models to video deblurring, and our models achieved outstanding performance." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 462, + 345, + 474 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 462, + 345, + 474 + ], + "spans": [ + { + "bbox": [ + 131, + 462, + 345, + 474 + ], + "type": "text", + "content": "2.3 Wavelet-based Restoration Methods" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 479, + 482, + 588 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 479, + 482, + 588 + ], + "spans": [ + { + "bbox": [ + 130, + 479, + 482, + 588 + ], + "type": "text", + "content": "Wavelet is widely used in low-visual restoration tasks [7, 15, 19, 44, 48] as it can separate low-frequency and high-frequency information. For example, Huang et al. [12] propose a wavelet-based CNN for multi-scale face super resolution. WavTrans [17] combines wavelet transform with Transformer for feature extraction and fusion; DSTNet [29] develops a wavelet-based feature propagation method to effectively propagate main structures from long-range frames for better video deblurring. Inspired by the above methods, we proposed Wavelet-Aware Dynamic Transformer, which can extract the global structure of the video and utilize the prior features generated by DM to restore high-frequency details." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 602, + 233, + 616 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 602, + 233, + 616 + ], + "spans": [ + { + "bbox": [ + 131, + 602, + 233, + 616 + ], + "type": "text", + "content": "3 Methodology" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 624, + 265, + 636 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 624, + 265, + 636 + ], + "spans": [ + { + "bbox": [ + 131, + 624, + 265, + 636 + ], + "type": "text", + "content": "3.1 Overall Architecture" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "content": "The overall framework of our proposed VD-Diff is shown in Fig. 1, which consists of two parts: (1) Wavelet-Aware Dynamic Transformer (WADT); (2) Diffusion" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 102 + ], + "type": "text", + "content": "C. Rao et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "type": "text", + "content": "Model. Specifically, we first employ Diffusion Model to obtain a highly compact latent prior feature " + }, + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "type": "inline_equation", + "content": "z \\in \\mathbb{R}^{T \\times C'}" + }, + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "type": "text", + "content": ", which contains supplementary high-frequency information. Note that " + }, + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "type": "inline_equation", + "content": "C'" + }, + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "type": "text", + "content": " is the number of channels of latent features. Then, we employ WADT to deblur the blurry video frames " + }, + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "type": "inline_equation", + "content": "V_{blur} \\in R^{T \\times 3 \\times H \\times W}" + }, + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "type": "text", + "content": " into " + }, + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "type": "inline_equation", + "content": "V_{HQ} \\in R^{T \\times 3 \\times H \\times W}" + }, + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "type": "text", + "content": " with the help of the prior feature " + }, + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 193, + 350, + 205 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 193, + 350, + 205 + ], + "spans": [ + { + "bbox": [ + 131, + 193, + 350, + 205 + ], + "type": "text", + "content": "3.2 Wavelet-Aware Dynamic Transformer" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 212, + 482, + 344 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 212, + 482, + 344 + ], + "spans": [ + { + "bbox": [ + 130, + 212, + 482, + 344 + ], + "type": "text", + "content": "As shown in the WADT part of Fig. 1, we first utilize Conv3D to perform shallow feature extraction on the blurry video frames " + }, + { + "bbox": [ + 130, + 212, + 482, + 344 + ], + "type": "inline_equation", + "content": "V_{blur} \\in R^{T \\times 3 \\times H \\times W}" + }, + { + "bbox": [ + 130, + 212, + 482, + 344 + ], + "type": "text", + "content": " to obtain a global feature " + }, + { + "bbox": [ + 130, + 212, + 482, + 344 + ], + "type": "inline_equation", + "content": "F_{in} \\in \\mathbb{R}^{T \\times C \\times H \\times W}" + }, + { + "bbox": [ + 130, + 212, + 482, + 344 + ], + "type": "text", + "content": ". Then, we use Wavelet Transform (WT) to decompose it into approximate coefficients " + }, + { + "bbox": [ + 130, + 212, + 482, + 344 + ], + "type": "inline_equation", + "content": "F_{a} \\in \\mathbb{R}^{T \\times C / 4 \\times H \\times W}" + }, + { + "bbox": [ + 130, + 212, + 482, + 344 + ], + "type": "text", + "content": " and detail coefficients " + }, + { + "bbox": [ + 130, + 212, + 482, + 344 + ], + "type": "inline_equation", + "content": "F_{d} \\in \\mathbb{R}^{T \\times 3C / 4 \\times H \\times W}" + }, + { + "bbox": [ + 130, + 212, + 482, + 344 + ], + "type": "text", + "content": " and again use WT on " + }, + { + "bbox": [ + 130, + 212, + 482, + 344 + ], + "type": "inline_equation", + "content": "F_{a}" + }, + { + "bbox": [ + 130, + 212, + 482, + 344 + ], + "type": "text", + "content": " to obtain " + }, + { + "bbox": [ + 130, + 212, + 482, + 344 + ], + "type": "inline_equation", + "content": "F_{aa} \\in \\mathbb{R}^{T \\times \\hat{C} \\times H \\times W}" + }, + { + "bbox": [ + 130, + 212, + 482, + 344 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 212, + 482, + 344 + ], + "type": "inline_equation", + "content": "F_{ad} \\in \\mathbb{R}^{T \\times 3\\hat{C} \\times H \\times W}" + }, + { + "bbox": [ + 130, + 212, + 482, + 344 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 212, + 482, + 344 + ], + "type": "inline_equation", + "content": "\\hat{C} = C / 16" + }, + { + "bbox": [ + 130, + 212, + 482, + 344 + ], + "type": "text", + "content": ". Next, we employ WADTL to obtain the deblurred features " + }, + { + "bbox": [ + 130, + 212, + 482, + 344 + ], + "type": "inline_equation", + "content": "F_{out}" + }, + { + "bbox": [ + 130, + 212, + 482, + 344 + ], + "type": "text", + "content": " from the approximate coefficients " + }, + { + "bbox": [ + 130, + 212, + 482, + 344 + ], + "type": "inline_equation", + "content": "F_{aa}" + }, + { + "bbox": [ + 130, + 212, + 482, + 344 + ], + "type": "text", + "content": " and the prior feature " + }, + { + "bbox": [ + 130, + 212, + 482, + 344 + ], + "type": "inline_equation", + "content": "z'" + }, + { + "bbox": [ + 130, + 212, + 482, + 344 + ], + "type": "text", + "content": ". Then, we use WBPF to explore spatiotemporal information further. Finally, we use two IWTs to fuse features to obtain artifact-free and distortion-free videos " + }, + { + "bbox": [ + 130, + 212, + 482, + 344 + ], + "type": "inline_equation", + "content": "V_{HQ}" + }, + { + "bbox": [ + 130, + 212, + 482, + 344 + ], + "type": "text", + "content": ". We will elaborate on the structure of WADTL and WBPF as follows." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 345, + 482, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 345, + 482, + 403 + ], + "spans": [ + { + "bbox": [ + 130, + 345, + 482, + 403 + ], + "type": "text", + "content": "Wavelet-Aware Dynamic Transformer Layer. As shown in Fig. 1, we employ a stack of multiple WADT layer (WADTL) modules hierarchically to fuse approximate coefficients " + }, + { + "bbox": [ + 130, + 345, + 482, + 403 + ], + "type": "inline_equation", + "content": "F_{aa}" + }, + { + "bbox": [ + 130, + 345, + 482, + 403 + ], + "type": "text", + "content": " and prior features " + }, + { + "bbox": [ + 130, + 345, + 482, + 403 + ], + "type": "inline_equation", + "content": "z'" + }, + { + "bbox": [ + 130, + 345, + 482, + 403 + ], + "type": "text", + "content": ". As shown in Fig. 2, WADTL consists of a Wavelet-Aware Dynamic Multi-head Self-Attention (WAD-MSA) and a Wavelet-Aware Dynamic Feed-Forward Network (WAD-FFN)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 404, + 482, + 429 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 404, + 482, + 429 + ], + "spans": [ + { + "bbox": [ + 130, + 404, + 482, + 429 + ], + "type": "text", + "content": "In the WAD-MSA part, given a approximate coefficients " + }, + { + "bbox": [ + 130, + 404, + 482, + 429 + ], + "type": "inline_equation", + "content": "F_{aa}" + }, + { + "bbox": [ + 130, + 404, + 482, + 429 + ], + "type": "text", + "content": ", we use the highly compact latent prior feature " + }, + { + "bbox": [ + 130, + 404, + 482, + 429 + ], + "type": "inline_equation", + "content": "z'" + }, + { + "bbox": [ + 130, + 404, + 482, + 429 + ], + "type": "text", + "content": " to guide video restoration:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 240, + 436, + 480, + 449 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 240, + 436, + 480, + 449 + ], + "spans": [ + { + "bbox": [ + 240, + 436, + 480, + 449 + ], + "type": "interline_equation", + "content": "\\hat {F} = W _ {l} ^ {1} z ^ {\\prime} \\odot L N \\left(F _ {a a}\\right) + W _ {l} ^ {2} z ^ {\\prime}, \\tag {1}", + "image_path": "1c20cc46604daa26b93df7e1603a240363505f5a35f8dac2edda956fc58530f6.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 457, + 482, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 457, + 482, + 544 + ], + "spans": [ + { + "bbox": [ + 131, + 457, + 482, + 544 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 457, + 482, + 544 + ], + "type": "inline_equation", + "content": "\\odot" + }, + { + "bbox": [ + 131, + 457, + 482, + 544 + ], + "type": "text", + "content": " indicates element-wise multiplication, " + }, + { + "bbox": [ + 131, + 457, + 482, + 544 + ], + "type": "inline_equation", + "content": "LN" + }, + { + "bbox": [ + 131, + 457, + 482, + 544 + ], + "type": "text", + "content": " denotes layer normalization [1], " + }, + { + "bbox": [ + 131, + 457, + 482, + 544 + ], + "type": "inline_equation", + "content": "W_{l}" + }, + { + "bbox": [ + 131, + 457, + 482, + 544 + ], + "type": "text", + "content": " represents linear layer. Then, we utilize a 3D convolution layer and a 2D convolution layer to project " + }, + { + "bbox": [ + 131, + 457, + 482, + 544 + ], + "type": "inline_equation", + "content": "\\hat{F}" + }, + { + "bbox": [ + 131, + 457, + 482, + 544 + ], + "type": "text", + "content": " into query " + }, + { + "bbox": [ + 131, + 457, + 482, + 544 + ], + "type": "inline_equation", + "content": "Q = W_2^Q W_3^Q\\hat{F}" + }, + { + "bbox": [ + 131, + 457, + 482, + 544 + ], + "type": "text", + "content": ", key " + }, + { + "bbox": [ + 131, + 457, + 482, + 544 + ], + "type": "inline_equation", + "content": "K = W_2^K W_3^K\\hat{F}" + }, + { + "bbox": [ + 131, + 457, + 482, + 544 + ], + "type": "text", + "content": ", value " + }, + { + "bbox": [ + 131, + 457, + 482, + 544 + ], + "type": "inline_equation", + "content": "V = W_2^V W_3^V\\hat{F}" + }, + { + "bbox": [ + 131, + 457, + 482, + 544 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 131, + 457, + 482, + 544 + ], + "type": "inline_equation", + "content": "W_{3}" + }, + { + "bbox": [ + 131, + 457, + 482, + 544 + ], + "type": "text", + "content": " is the channel-wise 3D convolution and " + }, + { + "bbox": [ + 131, + 457, + 482, + 544 + ], + "type": "inline_equation", + "content": "W_{2}" + }, + { + "bbox": [ + 131, + 457, + 482, + 544 + ], + "type": "text", + "content": " is the depth-wise 2D convolution. Note that we use 3D convolution in the " + }, + { + "bbox": [ + 131, + 457, + 482, + 544 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 131, + 457, + 482, + 544 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 131, + 457, + 482, + 544 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 131, + 457, + 482, + 544 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 131, + 457, + 482, + 544 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 131, + 457, + 482, + 544 + ], + "type": "text", + "content": " dimensions and the same 2D convolution for the " + }, + { + "bbox": [ + 131, + 457, + 482, + 544 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 131, + 457, + 482, + 544 + ], + "type": "text", + "content": " features in the temporal dimension " + }, + { + "bbox": [ + 131, + 457, + 482, + 544 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 131, + 457, + 482, + 544 + ], + "type": "text", + "content": ". Next, we perform the attention calculation:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 219, + 552, + 480, + 565 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 219, + 552, + 480, + 565 + ], + "spans": [ + { + "bbox": [ + 219, + 552, + 480, + 565 + ], + "type": "interline_equation", + "content": "\\mathcal {F} _ {\\text {o u t}} ^ {\\prime} = W _ {a} \\hat {V} \\cdot \\operatorname {S o f t m a x} (\\hat {K} \\cdot \\hat {Q} / \\gamma) + F _ {a a}, \\tag {2}", + "image_path": "aa9b53f621ea27d7158ce39e07a03200634787daa4427b7ab164f42e618601c8.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 573, + 314, + 584 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 573, + 314, + 584 + ], + "spans": [ + { + "bbox": [ + 131, + 573, + 314, + 584 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 573, + 314, + 584 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 131, + 573, + 314, + 584 + ], + "type": "text", + "content": " is a learnable scaling parameter." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 584, + 482, + 644 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 584, + 482, + 644 + ], + "spans": [ + { + "bbox": [ + 130, + 584, + 482, + 644 + ], + "type": "text", + "content": "In the WDA-FFN part, we first integrate the prior " + }, + { + "bbox": [ + 130, + 584, + 482, + 644 + ], + "type": "inline_equation", + "content": "z'" + }, + { + "bbox": [ + 130, + 584, + 482, + 644 + ], + "type": "text", + "content": " into " + }, + { + "bbox": [ + 130, + 584, + 482, + 644 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{out}'" + }, + { + "bbox": [ + 130, + 584, + 482, + 644 + ], + "type": "text", + "content": " to obtain " + }, + { + "bbox": [ + 130, + 584, + 482, + 644 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{F}}_{out}'" + }, + { + "bbox": [ + 130, + 584, + 482, + 644 + ], + "type": "text", + "content": ". And then we use Conv3D to aggregate temporal information and use Conv2D to aggregate spatial information from neighboring pixels. Besides, we employ a gating mechanism to obtain more valuable information. The overall process of WDA-FFN can be described as:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 211, + 653, + 480, + 666 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 211, + 653, + 480, + 666 + ], + "spans": [ + { + "bbox": [ + 211, + 653, + 480, + 666 + ], + "type": "interline_equation", + "content": "F _ {o u t} = G \\left(W _ {2} ^ {1} W _ {3} ^ {1} \\hat {\\mathcal {F}} _ {o u t} ^ {\\prime}\\right) \\odot W _ {2} ^ {2} W _ {3} ^ {2} \\hat {\\mathcal {F}} _ {o u t} ^ {\\prime} + \\hat {\\mathcal {F}} _ {o u t} ^ {\\prime}, \\tag {3}", + "image_path": "ca4351e0ed3607c67942cc1942cabc5fd850f9d02bfd8545dcb28b7c6493e93d.jpg" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 217, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 217, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 217, + 91, + 448, + 102 + ], + "type": "text", + "content": "Video Deblurring with Transformer and Diffusion Model" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 135, + 114, + 479, + 183 + ], + "blocks": [ + { + "bbox": [ + 135, + 114, + 479, + 183 + ], + "lines": [ + { + "bbox": [ + 135, + 114, + 479, + 183 + ], + "spans": [ + { + "bbox": [ + 135, + 114, + 479, + 183 + ], + "type": "image", + "image_path": "d28e1dae8032e1ea2cb2493272ec1cc6267ae2a5c209484a50333d34f2f06ca5.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 192, + 482, + 228 + ], + "lines": [ + { + "bbox": [ + 132, + 192, + 482, + 228 + ], + "spans": [ + { + "bbox": [ + 132, + 192, + 482, + 228 + ], + "type": "text", + "content": "Fig. 2: The illustration of Wavelet-Aware Dynamic Transformer Layer (WADTL), which consists of Wavelet-Aware Dynamic Multi-head Self-Attention (WAD-MSA) and Wavelet-Aware Dynamic Feed-Forward Network (WAD-FFN)." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 133, + 242, + 479, + 381 + ], + "blocks": [ + { + "bbox": [ + 133, + 242, + 479, + 381 + ], + "lines": [ + { + "bbox": [ + 133, + 242, + 479, + 381 + ], + "spans": [ + { + "bbox": [ + 133, + 242, + 479, + 381 + ], + "type": "image", + "image_path": "7f1466d096027a7e03b75b2d7e41e271be9fee6ef117286783d77d2db398e0b8.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 392, + 482, + 427 + ], + "lines": [ + { + "bbox": [ + 131, + 392, + 482, + 427 + ], + "spans": [ + { + "bbox": [ + 131, + 392, + 482, + 427 + ], + "type": "text", + "content": "Fig. 3: The structure of the Forward Process in the WBPF. The Backward Process has the same network structure as the Forward Process, but the direction of information propagation between frames is opposite." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 455, + 377, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 455, + 377, + 467 + ], + "spans": [ + { + "bbox": [ + 130, + 455, + 377, + 467 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 455, + 377, + 467 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 130, + 455, + 377, + 467 + ], + "type": "text", + "content": " denotes Gaussian Error Linear Units (GELU)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 467, + 482, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 467, + 482, + 602 + ], + "spans": [ + { + "bbox": [ + 130, + 467, + 482, + 602 + ], + "type": "text", + "content": "Wavelet-based Bidirectional Propagation Fuse. Given the generated features " + }, + { + "bbox": [ + 130, + 467, + 482, + 602 + ], + "type": "inline_equation", + "content": "\\{X_{i}\\}_{i = 1}^{N}" + }, + { + "bbox": [ + 130, + 467, + 482, + 602 + ], + "type": "text", + "content": " by the first Inverse Wavelet Transform (IwT), to explore useful contents from the features of other frames, exiting methods usually simply stack " + }, + { + "bbox": [ + 130, + 467, + 482, + 602 + ], + "type": "inline_equation", + "content": "\\{X_{i}\\}_{i = 1}^{N}" + }, + { + "bbox": [ + 130, + 467, + 482, + 602 + ], + "type": "text", + "content": " or the alignment results of " + }, + { + "bbox": [ + 130, + 467, + 482, + 602 + ], + "type": "inline_equation", + "content": "\\{X_{i}\\}_{i = 1}^{N}" + }, + { + "bbox": [ + 130, + 467, + 482, + 602 + ], + "type": "text", + "content": ". However, directly stacking frames without accurately estimating the features of " + }, + { + "bbox": [ + 130, + 467, + 482, + 602 + ], + "type": "inline_equation", + "content": "\\{X_{i}\\}_{i = 1}^{N}" + }, + { + "bbox": [ + 130, + 467, + 482, + 602 + ], + "type": "text", + "content": " or the alignment results of " + }, + { + "bbox": [ + 130, + 467, + 482, + 602 + ], + "type": "inline_equation", + "content": "\\{X_{i}\\}_{i = 1}^{N}" + }, + { + "bbox": [ + 130, + 467, + 482, + 602 + ], + "type": "text", + "content": " can only partially exploit the valuable information between frames and may also cause the propagation of erroneous information between frames. To further capture more useful long-range spatio-temporal information and reduce the influence of inaccurate information, we propose Wavelet-based Bidirectional Propagation Fuse (WBPF). As shown in Fig. 1, WBPF consists of a forward process and a backward process." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 605, + 483, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 483, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 483, + 666 + ], + "type": "text", + "content": "We take the forward process as an example to introduce the WBPF module in detail, as shown in Fig. 3. First, we connect " + }, + { + "bbox": [ + 130, + 605, + 483, + 666 + ], + "type": "inline_equation", + "content": "Y_{i}" + }, + { + "bbox": [ + 130, + 605, + 483, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 605, + 483, + 666 + ], + "type": "inline_equation", + "content": "X_{i + 1}" + }, + { + "bbox": [ + 130, + 605, + 483, + 666 + ], + "type": "text", + "content": " to get " + }, + { + "bbox": [ + 130, + 605, + 483, + 666 + ], + "type": "inline_equation", + "content": "\\bar{X}" + }, + { + "bbox": [ + 130, + 605, + 483, + 666 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 605, + 483, + 666 + ], + "type": "inline_equation", + "content": "Y_{i}" + }, + { + "bbox": [ + 130, + 605, + 483, + 666 + ], + "type": "text", + "content": " is the output of the previous forward process and " + }, + { + "bbox": [ + 130, + 605, + 483, + 666 + ], + "type": "inline_equation", + "content": "X_{i + 1}" + }, + { + "bbox": [ + 130, + 605, + 483, + 666 + ], + "type": "text", + "content": " is the input of the current forward process. We then use a convolution layer with LeakyReLU to project " + }, + { + "bbox": [ + 130, + 605, + 483, + 666 + ], + "type": "inline_equation", + "content": "\\bar{X}" + }, + { + "bbox": [ + 130, + 605, + 483, + 666 + ], + "type": "text", + "content": " into " + }, + { + "bbox": [ + 130, + 605, + 483, + 666 + ], + "type": "inline_equation", + "content": "\\hat{X}" + }, + { + "bbox": [ + 130, + 605, + 483, + 666 + ], + "type": "text", + "content": ". Then we split " + }, + { + "bbox": [ + 130, + 605, + 483, + 666 + ], + "type": "inline_equation", + "content": "\\hat{X}" + }, + { + "bbox": [ + 130, + 605, + 483, + 666 + ], + "type": "text", + "content": " into " + }, + { + "bbox": [ + 130, + 605, + 483, + 666 + ], + "type": "inline_equation", + "content": "\\hat{X}_1, \\hat{X}_2" + }, + { + "bbox": [ + 130, + 605, + 483, + 666 + ], + "type": "text", + "content": " along the channel dimension. Next, we" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "type": "text", + "content": "C. Rao et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 131, + 114, + 313, + 129 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 114, + 313, + 129 + ], + "spans": [ + { + "bbox": [ + 131, + 114, + 313, + 129 + ], + "type": "text", + "content": "obtain the fused feature from " + }, + { + "bbox": [ + 131, + 114, + 313, + 129 + ], + "type": "inline_equation", + "content": "\\hat{X}_1, \\hat{X}_2" + }, + { + "bbox": [ + 131, + 114, + 313, + 129 + ], + "type": "text", + "content": " by:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 217, + 137, + 481, + 152 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 217, + 137, + 481, + 152 + ], + "spans": [ + { + "bbox": [ + 217, + 137, + 481, + 152 + ], + "type": "interline_equation", + "content": "F _ {p r o} = \\hat {X} _ {1} \\odot S \\left(W _ {1} \\hat {X} _ {1}\\right) + \\hat {X} _ {2} \\odot S \\left(W _ {2} \\hat {X} _ {2}\\right), \\tag {4}", + "image_path": "db8247d6a5553417ab3e86b28c1694c445d2c6c2143d843d1e774d5a6e378f88.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 159, + 481, + 196 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 159, + 481, + 196 + ], + "spans": [ + { + "bbox": [ + 131, + 159, + 481, + 196 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 159, + 481, + 196 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 131, + 159, + 481, + 196 + ], + "type": "text", + "content": " denotes the Sigmoid function and " + }, + { + "bbox": [ + 131, + 159, + 481, + 196 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 131, + 159, + 481, + 196 + ], + "type": "text", + "content": " is the " + }, + { + "bbox": [ + 131, + 159, + 481, + 196 + ], + "type": "inline_equation", + "content": "3 \\times 3" + }, + { + "bbox": [ + 131, + 159, + 481, + 196 + ], + "type": "text", + "content": " 2D convolution. We utilize AvgPool, ResBlock, MaxPool, and Conv2D to better explore spatial information. The overall process can be described as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 231, + 205, + 481, + 220 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 231, + 205, + 481, + 220 + ], + "spans": [ + { + "bbox": [ + 231, + 205, + 481, + 220 + ], + "type": "interline_equation", + "content": "\\bar {Y} _ {i + 1} = F _ {p r o} \\otimes \\text {u p s a m p l e} (P \\left(F _ {p r o}\\right)), \\tag {5}", + "image_path": "527d984dc86f1855058b1a4f503915ef315ae3d2067b788f0f85ce14163aca59.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 228, + 481, + 253 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 228, + 481, + 253 + ], + "spans": [ + { + "bbox": [ + 131, + 228, + 481, + 253 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 228, + 481, + 253 + ], + "type": "inline_equation", + "content": "\\mathrm{P}" + }, + { + "bbox": [ + 131, + 228, + 481, + 253 + ], + "type": "text", + "content": " represents the stack of AvgPool, ResBlock, MaxPool, ResBlock and Conv2D. In the last, we employ 30 ResBlocks on " + }, + { + "bbox": [ + 131, + 228, + 481, + 253 + ], + "type": "inline_equation", + "content": "\\bar{Y}_{i + 1}" + }, + { + "bbox": [ + 131, + 228, + 481, + 253 + ], + "type": "text", + "content": " to obtain " + }, + { + "bbox": [ + 131, + 228, + 481, + 253 + ], + "type": "inline_equation", + "content": "Y_{i + 1}" + }, + { + "bbox": [ + 131, + 228, + 481, + 253 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 270, + 242, + 281 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 270, + 242, + 281 + ], + "spans": [ + { + "bbox": [ + 132, + 270, + 242, + 281 + ], + "type": "text", + "content": "3.3 Diffusion Model" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 289, + 482, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 289, + 482, + 373 + ], + "spans": [ + { + "bbox": [ + 131, + 289, + 482, + 373 + ], + "type": "text", + "content": "Based on conditional denoising diffusion probabilistic models [11,27,31,35], the DM involves a forward diffusion process and a reverse denoising process. We first adopt the Latent Encoder (LE) on ground truth video and the Condition Encoder (CE) on blurry video to generate the ground truth feature " + }, + { + "bbox": [ + 131, + 289, + 482, + 373 + ], + "type": "inline_equation", + "content": "z \\in R^{T \\times 4C'}" + }, + { + "bbox": [ + 131, + 289, + 482, + 373 + ], + "type": "text", + "content": " and the condition feature " + }, + { + "bbox": [ + 131, + 289, + 482, + 373 + ], + "type": "inline_equation", + "content": "c \\in R^{T \\times 4C'}" + }, + { + "bbox": [ + 131, + 289, + 482, + 373 + ], + "type": "text", + "content": ". Then, we use the diffusion model to generate the prior feature " + }, + { + "bbox": [ + 131, + 289, + 482, + 373 + ], + "type": "inline_equation", + "content": "z' \\in R^{T \\times 4C'}" + }, + { + "bbox": [ + 131, + 289, + 482, + 373 + ], + "type": "text", + "content": " using " + }, + { + "bbox": [ + 131, + 289, + 482, + 373 + ], + "type": "inline_equation", + "content": "c \\in R^{T \\times 4C'}" + }, + { + "bbox": [ + 131, + 289, + 482, + 373 + ], + "type": "text", + "content": " as the condition and " + }, + { + "bbox": [ + 131, + 289, + 482, + 373 + ], + "type": "inline_equation", + "content": "z \\in R^{T \\times 4C'}" + }, + { + "bbox": [ + 131, + 289, + 482, + 373 + ], + "type": "text", + "content": " as the target. We detail the diffusion process as follows." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 374, + 481, + 397 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 374, + 481, + 397 + ], + "spans": [ + { + "bbox": [ + 131, + 374, + 481, + 397 + ], + "type": "text", + "content": "In the forward diffusion process, we transform " + }, + { + "bbox": [ + 131, + 374, + 481, + 397 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 131, + 374, + 481, + 397 + ], + "type": "text", + "content": " into Gaussian noise by " + }, + { + "bbox": [ + 131, + 374, + 481, + 397 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 131, + 374, + 481, + 397 + ], + "type": "text", + "content": " iterations, which can be described as follows:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 225, + 406, + 481, + 421 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 225, + 406, + 481, + 421 + ], + "spans": [ + { + "bbox": [ + 225, + 406, + 481, + 421 + ], + "type": "interline_equation", + "content": "q \\left(z _ {t} \\mid z _ {t - 1}\\right) = \\mathcal {N} \\left(z _ {t}; \\sqrt {1 - \\beta_ {t}} z _ {t - 1}, \\beta_ {t} I\\right), \\tag {6}", + "image_path": "0cfe8e59b5186aaad8567d02741e33b6a4d7e05c88be89f3603784572d20811f.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 131, + 430, + 481, + 466 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 430, + 481, + 466 + ], + "spans": [ + { + "bbox": [ + 131, + 430, + 481, + 466 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 430, + 481, + 466 + ], + "type": "inline_equation", + "content": "t = 1,2,\\dots,T" + }, + { + "bbox": [ + 131, + 430, + 481, + 466 + ], + "type": "text", + "content": "; " + }, + { + "bbox": [ + 131, + 430, + 481, + 466 + ], + "type": "inline_equation", + "content": "z_0 = z" + }, + { + "bbox": [ + 131, + 430, + 481, + 466 + ], + "type": "text", + "content": "; " + }, + { + "bbox": [ + 131, + 430, + 481, + 466 + ], + "type": "inline_equation", + "content": "\\beta_{t} \\in (0,1)" + }, + { + "bbox": [ + 131, + 430, + 481, + 466 + ], + "type": "text", + "content": " are hyperparameters that control the variance of the noise. Using the notation " + }, + { + "bbox": [ + 131, + 430, + 481, + 466 + ], + "type": "inline_equation", + "content": "\\alpha_{t} = 1 - \\beta_{t}" + }, + { + "bbox": [ + 131, + 430, + 481, + 466 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 131, + 430, + 481, + 466 + ], + "type": "inline_equation", + "content": "\\bar{\\alpha}_{t} = \\prod_{i=1}^{t} \\alpha_{i}" + }, + { + "bbox": [ + 131, + 430, + 481, + 466 + ], + "type": "text", + "content": " and through iterative derivation with reparameterization, Eq. (6) can be written as:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 225, + 475, + 481, + 489 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 225, + 475, + 481, + 489 + ], + "spans": [ + { + "bbox": [ + 225, + 475, + 481, + 489 + ], + "type": "interline_equation", + "content": "q \\left(z _ {T} \\mid z _ {0}\\right) = \\mathcal {N} \\left(z _ {T}; \\sqrt {\\bar {\\alpha} _ {T}} z _ {0}, (1 - \\bar {\\alpha} _ {T}) I\\right), \\tag {7}", + "image_path": "463125453c2e27ea7dfd9df95f2266e041c93ed089efe542c2a575ec86ee1b83.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 498, + 164, + 508 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 498, + 164, + 508 + ], + "spans": [ + { + "bbox": [ + 132, + 498, + 164, + 508 + ], + "type": "text", + "content": "that is" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 251, + 509, + 481, + 522 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 251, + 509, + 481, + 522 + ], + "spans": [ + { + "bbox": [ + 251, + 509, + 481, + 522 + ], + "type": "interline_equation", + "content": "z _ {T} = \\sqrt {\\bar {\\alpha} _ {T}} z _ {0} + \\sqrt {1 - \\bar {\\alpha} _ {T}} \\epsilon . \\tag {8}", + "image_path": "43d15846f6f5458e2b9257d9c3b671f02ea0551d08a504c893c2832c2b94d73b.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 131, + 528, + 481, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 528, + 481, + 552 + ], + "spans": [ + { + "bbox": [ + 131, + 528, + 481, + 552 + ], + "type": "text", + "content": "In the reverse denoising process, DM sample a Gaussian random noise " + }, + { + "bbox": [ + 131, + 528, + 481, + 552 + ], + "type": "inline_equation", + "content": "z_{T}" + }, + { + "bbox": [ + 131, + 528, + 481, + 552 + ], + "type": "text", + "content": " and then gradually denoise " + }, + { + "bbox": [ + 131, + 528, + 481, + 552 + ], + "type": "inline_equation", + "content": "z_{T}" + }, + { + "bbox": [ + 131, + 528, + 481, + 552 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 131, + 528, + 481, + 552 + ], + "type": "inline_equation", + "content": "z_{0}" + }, + { + "bbox": [ + 131, + 528, + 481, + 552 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 222, + 561, + 481, + 574 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 222, + 561, + 481, + 574 + ], + "spans": [ + { + "bbox": [ + 222, + 561, + 481, + 574 + ], + "type": "interline_equation", + "content": "q \\left(z _ {t - 1} \\mid z _ {t}, z _ {0}\\right) = \\mathcal {N} \\left(z _ {t - 1}; \\mu_ {t} \\left(z _ {t}, z _ {0}\\right), \\tilde {\\beta} _ {t} I\\right), \\tag {9}", + "image_path": "cc507a3874a3e493cc249ac03a165e53fb052188a45d301b96a9da22e5592fc3.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 211, + 593, + 481, + 617 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 211, + 593, + 481, + 617 + ], + "spans": [ + { + "bbox": [ + 211, + 593, + 481, + 617 + ], + "type": "interline_equation", + "content": "\\mu_ {t} \\left(z _ {t}, z _ {0}\\right) = \\frac {\\sqrt {\\bar {\\alpha} _ {t - 1}} \\beta_ {t}}{1 - \\bar {\\alpha} _ {t}} z _ {0} + \\frac {\\sqrt {\\alpha_ {t}} \\left(1 - \\bar {\\alpha} _ {t - 1}\\right)}{1 - \\bar {\\alpha} _ {t}} z _ {t}. \\tag {10}", + "image_path": "7b12dd22a41d4346aceb2a46fb321339751a8b5890232c6b163ee60ca4f1503b.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 622, + 390, + 634 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 622, + 390, + 634 + ], + "spans": [ + { + "bbox": [ + 132, + 622, + 390, + 634 + ], + "type": "text", + "content": "We can use Eq. (8) to represent " + }, + { + "bbox": [ + 132, + 622, + 390, + 634 + ], + "type": "inline_equation", + "content": "z_0" + }, + { + "bbox": [ + 132, + 622, + 390, + 634 + ], + "type": "text", + "content": " in Eq. (10) and obtain:" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 233, + 643, + 481, + 669 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 233, + 643, + 481, + 669 + ], + "spans": [ + { + "bbox": [ + 233, + 643, + 481, + 669 + ], + "type": "interline_equation", + "content": "\\mu_ {t} \\left(z _ {t}, z _ {0}\\right) = \\frac {1}{\\sqrt {\\alpha_ {t}}} \\left(z _ {t} - \\epsilon \\frac {1 - \\alpha_ {t}}{\\sqrt {1 - \\bar {\\alpha} _ {t}}}\\right). \\tag {11}", + "image_path": "70adb426bb2ff80a41ddc8154318d030e0441bb82df317c950720951bf9c568c.jpg" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 217, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 217, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 217, + 91, + 448, + 102 + ], + "type": "text", + "content": "Video Deblurring with Transformer and Diffusion Model" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 337, + 129 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 337, + 129 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 337, + 129 + ], + "type": "text", + "content": "We remove the variance estimation in Eq. (9):" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 244, + 137, + 481, + 163 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 244, + 137, + 481, + 163 + ], + "spans": [ + { + "bbox": [ + 244, + 137, + 481, + 163 + ], + "type": "interline_equation", + "content": "z _ {t - 1} = \\frac {1}{\\sqrt {\\alpha_ {t}}} (z _ {t} - \\epsilon \\frac {1 - \\alpha_ {t}}{\\sqrt {1 - \\bar {\\alpha} _ {t}}}). \\tag {12}", + "image_path": "68cd754d4ca78c68ec00289bd185bd9280b567d60f1110c976720d22fc10c03f.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 170, + 480, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 170, + 480, + 194 + ], + "spans": [ + { + "bbox": [ + 132, + 170, + 480, + 194 + ], + "type": "text", + "content": "Then the noise " + }, + { + "bbox": [ + 132, + 170, + 480, + 194 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 132, + 170, + 480, + 194 + ], + "type": "text", + "content": " is the only uncertain variable. Following previous work [11, 31, 38], we adopt a neural network to predict the noise " + }, + { + "bbox": [ + 132, + 170, + 480, + 194 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 132, + 170, + 480, + 194 + ], + "type": "text", + "content": " on " + }, + { + "bbox": [ + 132, + 170, + 480, + 194 + ], + "type": "inline_equation", + "content": "z_{t}" + }, + { + "bbox": [ + 132, + 170, + 480, + 194 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 132, + 170, + 480, + 194 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 132, + 170, + 480, + 194 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 132, + 170, + 480, + 194 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 132, + 170, + 480, + 194 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 211, + 329, + 224 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 211, + 329, + 224 + ], + "spans": [ + { + "bbox": [ + 132, + 211, + 329, + 224 + ], + "type": "text", + "content": "3.4 Training and Inference Strategies" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 231, + 479, + 290 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 231, + 479, + 290 + ], + "spans": [ + { + "bbox": [ + 130, + 231, + 479, + 290 + ], + "type": "text", + "content": "Inspired by [5,38], we divide the training process into three stages: the first is to train the Latent Encoder with Wavelet-Aware Dynamic Transformer for ground truth feature extraction, the second is to train the Diffusion Model to generate the prior feature, and the third is to jointly train DM and WADT to generate artifact-free and high-realistic videos." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 290, + 482, + 360 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 290, + 482, + 360 + ], + "spans": [ + { + "bbox": [ + 131, + 290, + 482, + 360 + ], + "type": "text", + "content": "Training Stage One. The first stage aims to train the Latent Encoder (LE) to extract valuable ground truth prior feature " + }, + { + "bbox": [ + 131, + 290, + 482, + 360 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 131, + 290, + 482, + 360 + ], + "type": "text", + "content": " from ground truth (GT). We train LE and WADT as a whole model by inputting " + }, + { + "bbox": [ + 131, + 290, + 482, + 360 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 131, + 290, + 482, + 360 + ], + "type": "text", + "content": " into WADTL instead of " + }, + { + "bbox": [ + 131, + 290, + 482, + 360 + ], + "type": "inline_equation", + "content": "z'" + }, + { + "bbox": [ + 131, + 290, + 482, + 360 + ], + "type": "text", + "content": " and ignoring the condition encoder and the diffusion model in Fig. 1. We employ L1 loss and Multi-scale Frequency Reconstruction (MSFR) [10] loss. L1 loss is defined as:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 254, + 361, + 481, + 392 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 254, + 361, + 481, + 392 + ], + "spans": [ + { + "bbox": [ + 254, + 361, + 481, + 392 + ], + "type": "interline_equation", + "content": "L _ {1} = \\sum_ {i = 1} ^ {N} \\| V _ {H Q} ^ {i} - V _ {G T} ^ {i} \\| _ {1}, \\tag {13}", + "image_path": "ebec943c2fa958f46e0dd406d36bdbf9126aa7f82f8f02017c61f9c7b727cd42.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 397, + 482, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 397, + 482, + 434 + ], + "spans": [ + { + "bbox": [ + 131, + 397, + 482, + 434 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 397, + 482, + 434 + ], + "type": "inline_equation", + "content": "V_{HQ}^{i}" + }, + { + "bbox": [ + 131, + 397, + 482, + 434 + ], + "type": "text", + "content": " is the " + }, + { + "bbox": [ + 131, + 397, + 482, + 434 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 131, + 397, + 482, + 434 + ], + "type": "text", + "content": "-th frame of the deblurred HQ video and " + }, + { + "bbox": [ + 131, + 397, + 482, + 434 + ], + "type": "inline_equation", + "content": "V_{GT}^{i}" + }, + { + "bbox": [ + 131, + 397, + 482, + 434 + ], + "type": "text", + "content": " is the " + }, + { + "bbox": [ + 131, + 397, + 482, + 434 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 131, + 397, + 482, + 434 + ], + "type": "text", + "content": "-th frame of the GT video. The MSFR loss measures the L1 distance between multi-scale GT and deblurred frame in the frequency domain as follows:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 212, + 441, + 481, + 475 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 212, + 441, + 481, + 475 + ], + "spans": [ + { + "bbox": [ + 212, + 441, + 481, + 475 + ], + "type": "interline_equation", + "content": "L _ {M S F R} = \\sum_ {i = 1} ^ {N} \\sum_ {k = 1} ^ {K} \\frac {1}{t _ {k}} \\| \\mathcal {F T} \\left(\\hat {S} _ {k} ^ {i}\\right) - \\mathcal {F T} \\left(S _ {k} ^ {i}\\right) \\| _ {1}, \\tag {14}", + "image_path": "5aa4d09030612a44fb70665fd3fe621c38dad5f89a38c99f4817954b5f6bff53.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 131, + 481, + 482, + 530 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 481, + 482, + 530 + ], + "spans": [ + { + "bbox": [ + 131, + 481, + 482, + 530 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 481, + 482, + 530 + ], + "type": "inline_equation", + "content": "\\mathcal{FT}" + }, + { + "bbox": [ + 131, + 481, + 482, + 530 + ], + "type": "text", + "content": " denotes the Fast Fourier Transform that transfers image signal to the frequency domain, " + }, + { + "bbox": [ + 131, + 481, + 482, + 530 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 131, + 481, + 482, + 530 + ], + "type": "text", + "content": " denotes the number of different scales and the " + }, + { + "bbox": [ + 131, + 481, + 482, + 530 + ], + "type": "inline_equation", + "content": "\\tilde{S}_k^i" + }, + { + "bbox": [ + 131, + 481, + 482, + 530 + ], + "type": "text", + "content": " indicates the image signal of the " + }, + { + "bbox": [ + 131, + 481, + 482, + 530 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 131, + 481, + 482, + 530 + ], + "type": "text", + "content": "-th frame at the " + }, + { + "bbox": [ + 131, + 481, + 482, + 530 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 131, + 481, + 482, + 530 + ], + "type": "text", + "content": "-th scale. The final optimization function of stage one is determined as follows:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 252, + 540, + 481, + 553 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 252, + 540, + 481, + 553 + ], + "spans": [ + { + "bbox": [ + 252, + 540, + 481, + 553 + ], + "type": "interline_equation", + "content": "L _ {d e b l u r} = L _ {1} + \\lambda L _ {M S F R}, \\tag {15}", + "image_path": "d59918f660c47f20ff1a71e9bbcd5f5a8e49e6d9d419eb7d01d5e9251d2f076a.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 560, + 296, + 571 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 560, + 296, + 571 + ], + "spans": [ + { + "bbox": [ + 132, + 560, + 296, + 571 + ], + "type": "text", + "content": "where we experimentally set " + }, + { + "bbox": [ + 132, + 560, + 296, + 571 + ], + "type": "inline_equation", + "content": "\\lambda = 0.1" + }, + { + "bbox": [ + 132, + 560, + 296, + 571 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 131, + 572, + 479, + 619 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 572, + 479, + 619 + ], + "spans": [ + { + "bbox": [ + 131, + 572, + 479, + 619 + ], + "type": "text", + "content": "Training Stage Two. In the second stage, we first use a Condition Encoder (CE), whose struct is the same as the Latent Encoder (LE), to generate condition feature " + }, + { + "bbox": [ + 131, + 572, + 479, + 619 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 131, + 572, + 479, + 619 + ], + "type": "text", + "content": " from the blur video. Then, we train a diffusion model (DM) to generate prior feature " + }, + { + "bbox": [ + 131, + 572, + 479, + 619 + ], + "type": "inline_equation", + "content": "z'" + }, + { + "bbox": [ + 131, + 572, + 479, + 619 + ], + "type": "text", + "content": " using condition feature " + }, + { + "bbox": [ + 131, + 572, + 479, + 619 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 131, + 572, + 479, + 619 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 131, + 620, + 481, + 643 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 620, + 481, + 643 + ], + "spans": [ + { + "bbox": [ + 131, + 620, + 481, + 643 + ], + "type": "text", + "content": "We run the complete " + }, + { + "bbox": [ + 131, + 620, + 481, + 643 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 131, + 620, + 481, + 643 + ], + "type": "text", + "content": " iteration reverse processes Eq. (12) to generate prior feature " + }, + { + "bbox": [ + 131, + 620, + 481, + 643 + ], + "type": "inline_equation", + "content": "\\hat{z}" + }, + { + "bbox": [ + 131, + 620, + 481, + 643 + ], + "type": "text", + "content": ". We denoted this loss as:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 267, + 653, + 481, + 667 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 267, + 653, + 481, + 667 + ], + "spans": [ + { + "bbox": [ + 267, + 653, + 481, + 667 + ], + "type": "interline_equation", + "content": "L _ {d i f f} = \\left\\| z ^ {\\prime} - \\hat {z} \\right\\| _ {1}. \\tag {16}", + "image_path": "9dc11e2d1bcbdab570cbca6379075303f74ff62420c31b79f42877029635dcfc.jpg" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "type": "text", + "content": "C. Rao et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 134, + 114, + 292, + 205 + ], + "blocks": [ + { + "bbox": [ + 134, + 114, + 292, + 205 + ], + "lines": [ + { + "bbox": [ + 134, + 114, + 292, + 205 + ], + "spans": [ + { + "bbox": [ + 134, + 114, + 292, + 205 + ], + "type": "image", + "image_path": "125b4a395c134fea66e38887fda90aaeaa3ae00e30870831ccb096055610a083.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 296, + 197, + 327, + 204 + ], + "lines": [ + { + "bbox": [ + 296, + 197, + 327, + 204 + ], + "spans": [ + { + "bbox": [ + 296, + 197, + 327, + 204 + ], + "type": "text", + "content": "DSTNet" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 293, + 114, + 331, + 150 + ], + "blocks": [ + { + "bbox": [ + 293, + 114, + 331, + 150 + ], + "lines": [ + { + "bbox": [ + 293, + 114, + 331, + 150 + ], + "spans": [ + { + "bbox": [ + 293, + 114, + 331, + 150 + ], + "type": "image", + "image_path": "34467c743e513dbf72fb9fe5d571113f7994645057868fe5d75cf547f43893c8.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 331, + 114, + 367, + 150 + ], + "blocks": [ + { + "bbox": [ + 331, + 114, + 367, + 150 + ], + "lines": [ + { + "bbox": [ + 331, + 114, + 367, + 150 + ], + "spans": [ + { + "bbox": [ + 331, + 114, + 367, + 150 + ], + "type": "image", + "image_path": "2e5151c314db403abf170db16f105bd4fc0f14c1f4e81a0dc99cbaae795173d9.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 369, + 114, + 406, + 150 + ], + "blocks": [ + { + "bbox": [ + 369, + 114, + 406, + 150 + ], + "lines": [ + { + "bbox": [ + 369, + 114, + 406, + 150 + ], + "spans": [ + { + "bbox": [ + 369, + 114, + 406, + 150 + ], + "type": "image", + "image_path": "ae14aba8a6e4a35d2d5ca17b145376bb576b748578d7a67baa58ef7c5a0c6e6d.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 407, + 114, + 443, + 150 + ], + "blocks": [ + { + "bbox": [ + 407, + 114, + 443, + 150 + ], + "lines": [ + { + "bbox": [ + 407, + 114, + 443, + 150 + ], + "spans": [ + { + "bbox": [ + 407, + 114, + 443, + 150 + ], + "type": "image", + "image_path": "2619cda6ee0e89f4ac8cd9e4d56599faf8712413b8a1786f16a1eebbd0b5a57a.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 444, + 114, + 481, + 150 + ], + "blocks": [ + { + "bbox": [ + 444, + 114, + 481, + 150 + ], + "lines": [ + { + "bbox": [ + 444, + 114, + 481, + 150 + ], + "spans": [ + { + "bbox": [ + 444, + 114, + 481, + 150 + ], + "type": "image", + "image_path": "970968f7d221ab30225ed5ac812c905974a85ef5e93c54122643e10c0f18fbd6.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 293, + 159, + 331, + 195 + ], + "blocks": [ + { + "bbox": [ + 304, + 152, + 320, + 159 + ], + "lines": [ + { + "bbox": [ + 304, + 152, + 320, + 159 + ], + "spans": [ + { + "bbox": [ + 304, + 152, + 320, + 159 + ], + "type": "text", + "content": "Blur" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 293, + 159, + 331, + 195 + ], + "lines": [ + { + "bbox": [ + 293, + 159, + 331, + 195 + ], + "spans": [ + { + "bbox": [ + 293, + 159, + 331, + 195 + ], + "type": "image", + "image_path": "3f0a5bb7d668cdbf660b8244ca65c6571e7e0381b541a7e2ece03a8c76060c6c.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 331, + 159, + 367, + 195 + ], + "blocks": [ + { + "bbox": [ + 337, + 152, + 361, + 159 + ], + "lines": [ + { + "bbox": [ + 337, + 152, + 361, + 159 + ], + "spans": [ + { + "bbox": [ + 337, + 152, + 361, + 159 + ], + "type": "text", + "content": "EDVR" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 331, + 159, + 367, + 195 + ], + "lines": [ + { + "bbox": [ + 331, + 159, + 367, + 195 + ], + "spans": [ + { + "bbox": [ + 331, + 159, + 367, + 195 + ], + "type": "image", + "image_path": "776ad1b560723080c0b2ac9a379706ed3d3c561ac401e4a73d976954e232a9e3.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 338, + 197, + 360, + 204 + ], + "lines": [ + { + "bbox": [ + 338, + 197, + 360, + 204 + ], + "spans": [ + { + "bbox": [ + 338, + 197, + 360, + 204 + ], + "type": "text", + "content": "FGST" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 368, + 159, + 406, + 195 + ], + "blocks": [ + { + "bbox": [ + 371, + 152, + 403, + 159 + ], + "lines": [ + { + "bbox": [ + 371, + 152, + 403, + 159 + ], + "spans": [ + { + "bbox": [ + 371, + 152, + 403, + 159 + ], + "type": "text", + "content": "MPRNet" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 368, + 159, + 406, + 195 + ], + "lines": [ + { + "bbox": [ + 368, + 159, + 406, + 195 + ], + "spans": [ + { + "bbox": [ + 368, + 159, + 406, + 195 + ], + "type": "image", + "image_path": "6b586679166e3f3733f95f917fed921e26e88f46c14f91601545e975a612b8f5.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 369, + 197, + 406, + 204 + ], + "lines": [ + { + "bbox": [ + 369, + 197, + 406, + 204 + ], + "spans": [ + { + "bbox": [ + 369, + 197, + 406, + 204 + ], + "type": "text", + "content": "DSTNet-L" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 407, + 159, + 443, + 195 + ], + "blocks": [ + { + "bbox": [ + 417, + 152, + 433, + 159 + ], + "lines": [ + { + "bbox": [ + 417, + 152, + 433, + 159 + ], + "spans": [ + { + "bbox": [ + 417, + 152, + 433, + 159 + ], + "type": "text", + "content": "TSP" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 407, + 159, + 443, + 195 + ], + "lines": [ + { + "bbox": [ + 407, + 159, + 443, + 195 + ], + "spans": [ + { + "bbox": [ + 407, + 159, + 443, + 195 + ], + "type": "image", + "image_path": "f95625964bf76fc6c1f8d6427bb9ede1fcde619c7ee0a690526481d968348ef1.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 416, + 197, + 434, + 204 + ], + "lines": [ + { + "bbox": [ + 416, + 197, + 434, + 204 + ], + "spans": [ + { + "bbox": [ + 416, + 197, + 434, + 204 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 444, + 159, + 481, + 195 + ], + "blocks": [ + { + "bbox": [ + 454, + 152, + 471, + 159 + ], + "lines": [ + { + "bbox": [ + 454, + 152, + 471, + 159 + ], + "spans": [ + { + "bbox": [ + 454, + 152, + 471, + 159 + ], + "type": "text", + "content": "SFE" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 444, + 159, + 481, + 195 + ], + "lines": [ + { + "bbox": [ + 444, + 159, + 481, + 195 + ], + "spans": [ + { + "bbox": [ + 444, + 159, + 481, + 195 + ], + "type": "image", + "image_path": "5ee8b9e951f18e52fbe8fc3f6344e2f02942aab9bfb980ae3b0582688ac032f4.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 456, + 197, + 469, + 204 + ], + "lines": [ + { + "bbox": [ + 456, + 197, + 469, + 204 + ], + "spans": [ + { + "bbox": [ + 456, + 197, + 469, + 204 + ], + "type": "text", + "content": "GT" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 134, + 248, + 292, + 338 + ], + "blocks": [ + { + "bbox": [ + 131, + 214, + 480, + 236 + ], + "lines": [ + { + "bbox": [ + 131, + 214, + 480, + 236 + ], + "spans": [ + { + "bbox": [ + 131, + 214, + 480, + 236 + ], + "type": "text", + "content": "Fig. 4: Visual comparison on GoPro [25] dataset. The deblurred results of previous work still contain significant blur effects. Our method generates much clearer frames." + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 134, + 248, + 292, + 338 + ], + "lines": [ + { + "bbox": [ + 134, + 248, + 292, + 338 + ], + "spans": [ + { + "bbox": [ + 134, + 248, + 292, + 338 + ], + "type": "image", + "image_path": "b32b1633ef79986b5c7a045015039ba0711f7a8f60abbcc0ef6b3912412a0548.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 349, + 480, + 371 + ], + "lines": [ + { + "bbox": [ + 131, + 349, + 480, + 371 + ], + "spans": [ + { + "bbox": [ + 131, + 349, + 480, + 371 + ], + "type": "text", + "content": "Fig. 5: Visual comparison on DVD [33] dataset. The deblurring effect of our proposed model is significantly better." + } + ] + } + ], + "index": 45, + "angle": 0, + "type": "image_caption" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 293, + 248, + 330, + 285 + ], + "blocks": [ + { + "bbox": [ + 293, + 248, + 330, + 285 + ], + "lines": [ + { + "bbox": [ + 293, + 248, + 330, + 285 + ], + "spans": [ + { + "bbox": [ + 293, + 248, + 330, + 285 + ], + "type": "image", + "image_path": "2fa8e5bed21ca76f4140a3515925d540d5e4b063f5152035b79cec00903bb7ec.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 286, + 319, + 293 + ], + "lines": [ + { + "bbox": [ + 304, + 286, + 319, + 293 + ], + "spans": [ + { + "bbox": [ + 304, + 286, + 319, + 293 + ], + "type": "text", + "content": "Blur" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_caption" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 331, + 248, + 367, + 285 + ], + "blocks": [ + { + "bbox": [ + 331, + 248, + 367, + 285 + ], + "lines": [ + { + "bbox": [ + 331, + 248, + 367, + 285 + ], + "spans": [ + { + "bbox": [ + 331, + 248, + 367, + 285 + ], + "type": "image", + "image_path": "50c4fcbefc9a5f8999bdf252c08062a8d948324d4f811ff9b34fa2819e49bb5d.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 337, + 286, + 361, + 293 + ], + "lines": [ + { + "bbox": [ + 337, + 286, + 361, + 293 + ], + "spans": [ + { + "bbox": [ + 337, + 286, + 361, + 293 + ], + "type": "text", + "content": "EDVR" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_caption" + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 368, + 248, + 405, + 285 + ], + "blocks": [ + { + "bbox": [ + 368, + 248, + 405, + 285 + ], + "lines": [ + { + "bbox": [ + 368, + 248, + 405, + 285 + ], + "spans": [ + { + "bbox": [ + 368, + 248, + 405, + 285 + ], + "type": "image", + "image_path": "7eba2eb8ae04825a2284d4718e4fc2cf449945ff8c3e55dc1f62c8f1ab50d424.jpg" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 370, + 286, + 402, + 293 + ], + "lines": [ + { + "bbox": [ + 370, + 286, + 402, + 293 + ], + "spans": [ + { + "bbox": [ + 370, + 286, + 402, + 293 + ], + "type": "text", + "content": "MPRNet" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_caption" + } + ], + "index": 29 + }, + { + "type": "image", + "bbox": [ + 406, + 248, + 443, + 285 + ], + "blocks": [ + { + "bbox": [ + 406, + 248, + 443, + 285 + ], + "lines": [ + { + "bbox": [ + 406, + 248, + 443, + 285 + ], + "spans": [ + { + "bbox": [ + 406, + 248, + 443, + 285 + ], + "type": "image", + "image_path": "62f0c44f826e22d1e35ee7d785eaa1aaee51c73a570161751696ac0e0e787cb6.jpg" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 416, + 286, + 431, + 293 + ], + "lines": [ + { + "bbox": [ + 416, + 286, + 431, + 293 + ], + "spans": [ + { + "bbox": [ + 416, + 286, + 431, + 293 + ], + "type": "text", + "content": "TSP" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_caption" + } + ], + "index": 31 + }, + { + "type": "image", + "bbox": [ + 443, + 248, + 479, + 285 + ], + "blocks": [ + { + "bbox": [ + 443, + 248, + 479, + 285 + ], + "lines": [ + { + "bbox": [ + 443, + 248, + 479, + 285 + ], + "spans": [ + { + "bbox": [ + 443, + 248, + 479, + 285 + ], + "type": "image", + "image_path": "1cf4cc5f99b5858eeec3b889c7f5f2eac0354255fdaaf2ea55fa975ddabfafaa.jpg" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 453, + 286, + 469, + 293 + ], + "lines": [ + { + "bbox": [ + 453, + 286, + 469, + 293 + ], + "spans": [ + { + "bbox": [ + 453, + 286, + 469, + 293 + ], + "type": "text", + "content": "SFE" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_caption" + } + ], + "index": 33 + }, + { + "type": "image", + "bbox": [ + 293, + 294, + 331, + 330 + ], + "blocks": [ + { + "bbox": [ + 293, + 294, + 331, + 330 + ], + "lines": [ + { + "bbox": [ + 293, + 294, + 331, + 330 + ], + "spans": [ + { + "bbox": [ + 293, + 294, + 331, + 330 + ], + "type": "image", + "image_path": "238a4422bdbd826570a704a88c36ad75f6fc3b5a0454ee4a493fd7e7276b62aa.jpg" + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 294, + 331, + 330, + 338 + ], + "lines": [ + { + "bbox": [ + 294, + 331, + 330, + 338 + ], + "spans": [ + { + "bbox": [ + 294, + 331, + 330, + 338 + ], + "type": "text", + "content": "BasicVSR" + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_caption" + } + ], + "index": 35 + }, + { + "type": "image", + "bbox": [ + 331, + 294, + 367, + 330 + ], + "blocks": [ + { + "bbox": [ + 331, + 294, + 367, + 330 + ], + "lines": [ + { + "bbox": [ + 331, + 294, + 367, + 330 + ], + "spans": [ + { + "bbox": [ + 331, + 294, + 367, + 330 + ], + "type": "image", + "image_path": "b64927086d94787828917df63b393568bb79074df8fbdb0af4f0d8ad26502add.jpg" + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 338, + 331, + 361, + 338 + ], + "lines": [ + { + "bbox": [ + 338, + 331, + 361, + 338 + ], + "spans": [ + { + "bbox": [ + 338, + 331, + 361, + 338 + ], + "type": "text", + "content": "FGST" + } + ] + } + ], + "index": 38, + "angle": 0, + "type": "image_caption" + } + ], + "index": 37 + }, + { + "type": "image", + "bbox": [ + 368, + 294, + 405, + 330 + ], + "blocks": [ + { + "bbox": [ + 368, + 294, + 405, + 330 + ], + "lines": [ + { + "bbox": [ + 368, + 294, + 405, + 330 + ], + "spans": [ + { + "bbox": [ + 368, + 294, + 405, + 330 + ], + "type": "image", + "image_path": "84bf2448fa6734706b9e366fa6f046c00cf374a2e6572f219aaf636ab33b4bf5.jpg" + } + ] + } + ], + "index": 39, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 371, + 331, + 402, + 338 + ], + "lines": [ + { + "bbox": [ + 371, + 331, + 402, + 338 + ], + "spans": [ + { + "bbox": [ + 371, + 331, + 402, + 338 + ], + "type": "text", + "content": "DSTNet" + } + ] + } + ], + "index": 40, + "angle": 0, + "type": "image_caption" + } + ], + "index": 39 + }, + { + "type": "image", + "bbox": [ + 406, + 294, + 443, + 330 + ], + "blocks": [ + { + "bbox": [ + 406, + 294, + 443, + 330 + ], + "lines": [ + { + "bbox": [ + 406, + 294, + 443, + 330 + ], + "spans": [ + { + "bbox": [ + 406, + 294, + 443, + 330 + ], + "type": "image", + "image_path": "0c03ceb7eb36a91a384be2137c5aa20158089bd4df25fd48a384d4a497ce18ee.jpg" + } + ] + } + ], + "index": 41, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 414, + 331, + 433, + 338 + ], + "lines": [ + { + "bbox": [ + 414, + 331, + 433, + 338 + ], + "spans": [ + { + "bbox": [ + 414, + 331, + 433, + 338 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 42, + "angle": 0, + "type": "image_caption" + } + ], + "index": 41 + }, + { + "type": "image", + "bbox": [ + 443, + 294, + 479, + 330 + ], + "blocks": [ + { + "bbox": [ + 443, + 294, + 479, + 330 + ], + "lines": [ + { + "bbox": [ + 443, + 294, + 479, + 330 + ], + "spans": [ + { + "bbox": [ + 443, + 294, + 479, + 330 + ], + "type": "image", + "image_path": "64a517332c6fb4ed9c341c121ea99be52dbde3531a7da0bdd9847297e63e33fb.jpg" + } + ] + } + ], + "index": 43, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 454, + 331, + 467, + 338 + ], + "lines": [ + { + "bbox": [ + 454, + 331, + 467, + 338 + ], + "spans": [ + { + "bbox": [ + 454, + 331, + 467, + 338 + ], + "type": "text", + "content": "GT" + } + ] + } + ], + "index": 44, + "angle": 0, + "type": "image_caption" + } + ], + "index": 43 + }, + { + "bbox": [ + 130, + 396, + 482, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 396, + 482, + 491 + ], + "spans": [ + { + "bbox": [ + 130, + 396, + 482, + 491 + ], + "type": "text", + "content": "Training Stage Three. In Training Stage Two, we train the DM to generate " + }, + { + "bbox": [ + 130, + 396, + 482, + 491 + ], + "type": "inline_equation", + "content": "z'" + }, + { + "bbox": [ + 130, + 396, + 482, + 491 + ], + "type": "text", + "content": " to approximate " + }, + { + "bbox": [ + 130, + 396, + 482, + 491 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 130, + 396, + 482, + 491 + ], + "type": "text", + "content": ". However, there inevitably is a slight deviation between the prior features " + }, + { + "bbox": [ + 130, + 396, + 482, + 491 + ], + "type": "inline_equation", + "content": "z'" + }, + { + "bbox": [ + 130, + 396, + 482, + 491 + ], + "type": "text", + "content": " and the ground truth prior feature " + }, + { + "bbox": [ + 130, + 396, + 482, + 491 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 130, + 396, + 482, + 491 + ], + "type": "text", + "content": ". To make the DM and WADT work better together to acquire information from different frequency domains for video deblurring and improve the robustness of WADT to the prior feature " + }, + { + "bbox": [ + 130, + 396, + 482, + 491 + ], + "type": "inline_equation", + "content": "z'" + }, + { + "bbox": [ + 130, + 396, + 482, + 491 + ], + "type": "text", + "content": ", we jointly train DM and WADT in Training Stage Three. The training loss " + }, + { + "bbox": [ + 130, + 396, + 482, + 491 + ], + "type": "inline_equation", + "content": "L_{total}" + }, + { + "bbox": [ + 130, + 396, + 482, + 491 + ], + "type": "text", + "content": " is given by the sum of the deblurring loss " + }, + { + "bbox": [ + 130, + 396, + 482, + 491 + ], + "type": "inline_equation", + "content": "L_{deblur}" + }, + { + "bbox": [ + 130, + 396, + 482, + 491 + ], + "type": "text", + "content": " and the diffusion loss " + }, + { + "bbox": [ + 130, + 396, + 482, + 491 + ], + "type": "inline_equation", + "content": "L_{diff}" + }, + { + "bbox": [ + 130, + 396, + 482, + 491 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 46 + }, + { + "bbox": [ + 252, + 493, + 480, + 506 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 252, + 493, + 480, + 506 + ], + "spans": [ + { + "bbox": [ + 252, + 493, + 480, + 506 + ], + "type": "interline_equation", + "content": "L _ {t o t a l} = L _ {d e b l u r} + L _ {d i f f}. \\tag {17}", + "image_path": "dc24d767728e8359fa9912780f2484990b5fafd612a51e5e8cf4b56dca1570a4.jpg" + } + ] + } + ], + "index": 47 + }, + { + "bbox": [ + 131, + 511, + 482, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 511, + 482, + 559 + ], + "spans": [ + { + "bbox": [ + 131, + 511, + 482, + 559 + ], + "type": "text", + "content": "Inference. Given a blurry input video " + }, + { + "bbox": [ + 131, + 511, + 482, + 559 + ], + "type": "inline_equation", + "content": "V_{blur} \\in R^{T \\times 3 \\times H \\times W}" + }, + { + "bbox": [ + 131, + 511, + 482, + 559 + ], + "type": "text", + "content": ", we first compress " + }, + { + "bbox": [ + 131, + 511, + 482, + 559 + ], + "type": "inline_equation", + "content": "V_{blur}" + }, + { + "bbox": [ + 131, + 511, + 482, + 559 + ], + "type": "text", + "content": " into a condition feature " + }, + { + "bbox": [ + 131, + 511, + 482, + 559 + ], + "type": "inline_equation", + "content": "c \\in R^{T \\times 4C'}" + }, + { + "bbox": [ + 131, + 511, + 482, + 559 + ], + "type": "text", + "content": ". Then, the diffusion model generates the prior feature " + }, + { + "bbox": [ + 131, + 511, + 482, + 559 + ], + "type": "inline_equation", + "content": "z'" + }, + { + "bbox": [ + 131, + 511, + 482, + 559 + ], + "type": "text", + "content": " from pure Gaussian noise " + }, + { + "bbox": [ + 131, + 511, + 482, + 559 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 131, + 511, + 482, + 559 + ], + "type": "text", + "content": " and condition feature " + }, + { + "bbox": [ + 131, + 511, + 482, + 559 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 131, + 511, + 482, + 559 + ], + "type": "text", + "content": ". Finally, VD-diff deblurs the blurry video with " + }, + { + "bbox": [ + 131, + 511, + 482, + 559 + ], + "type": "inline_equation", + "content": "z'" + }, + { + "bbox": [ + 131, + 511, + 482, + 559 + ], + "type": "text", + "content": " using WADT." + } + ] + } + ], + "index": 48 + }, + { + "bbox": [ + 131, + 581, + 230, + 594 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 581, + 230, + 594 + ], + "spans": [ + { + "bbox": [ + 131, + 581, + 230, + 594 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 49 + }, + { + "bbox": [ + 131, + 607, + 204, + 618 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 607, + 204, + 618 + ], + "spans": [ + { + "bbox": [ + 131, + 607, + 204, + 618 + ], + "type": "text", + "content": "4.1 Datasets" + } + ] + } + ], + "index": 50 + }, + { + "bbox": [ + 130, + 629, + 480, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 480, + 665 + ], + "type": "text", + "content": "GoPro. The GoPro [25] benchmark is composed of over 3,300 blurry-sharp image pairs of dynamic scenes. It is obtained by a high-speed camera. The training and testing subsets are split in proportional to 2:1." + } + ] + } + ], + "index": 51 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 217, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 217, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 217, + 91, + 448, + 102 + ], + "type": "text", + "content": "Video Deblurring with Transformer and Diffusion Model" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 135, + 114, + 219, + 192 + ], + "blocks": [ + { + "bbox": [ + 135, + 114, + 219, + 192 + ], + "lines": [ + { + "bbox": [ + 135, + 114, + 219, + 192 + ], + "spans": [ + { + "bbox": [ + 135, + 114, + 219, + 192 + ], + "type": "image", + "image_path": "e777d341f88b32ac1ab52d0d52bd2067e78c0df1076a2a82e56c3d7d9625ad22.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 170, + 193, + 187, + 201 + ], + "lines": [ + { + "bbox": [ + 170, + 193, + 187, + 201 + ], + "spans": [ + { + "bbox": [ + 170, + 193, + 187, + 201 + ], + "type": "text", + "content": "Blur" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 219, + 114, + 303, + 192 + ], + "blocks": [ + { + "bbox": [ + 219, + 114, + 303, + 192 + ], + "lines": [ + { + "bbox": [ + 219, + 114, + 303, + 192 + ], + "spans": [ + { + "bbox": [ + 219, + 114, + 303, + 192 + ], + "type": "image", + "image_path": "4a4c941a0566c193cd96daabeb3ce965bcd00b9991475b4372680fae7a5b8b41.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 246, + 193, + 279, + 201 + ], + "lines": [ + { + "bbox": [ + 246, + 193, + 279, + 201 + ], + "spans": [ + { + "bbox": [ + 246, + 193, + 279, + 201 + ], + "type": "text", + "content": "DVD [33]" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 303, + 114, + 386, + 192 + ], + "blocks": [ + { + "bbox": [ + 303, + 114, + 386, + 192 + ], + "lines": [ + { + "bbox": [ + 303, + 114, + 386, + 192 + ], + "spans": [ + { + "bbox": [ + 303, + 114, + 386, + 192 + ], + "type": "image", + "image_path": "7c24433315b348357e2485f62f59b59cc6f8a169698852f57dd758ef6c8b3346.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 386, + 114, + 474, + 192 + ], + "blocks": [ + { + "bbox": [ + 386, + 114, + 474, + 192 + ], + "lines": [ + { + "bbox": [ + 386, + 114, + 474, + 192 + ], + "spans": [ + { + "bbox": [ + 386, + 114, + 474, + 192 + ], + "type": "image", + "image_path": "865029901e1066fc4bdfe4f3d124c0f9b2632a3a30f655f0328800d39e4544cf.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 135, + 202, + 219, + 279 + ], + "blocks": [ + { + "bbox": [ + 135, + 202, + 219, + 279 + ], + "lines": [ + { + "bbox": [ + 135, + 202, + 219, + 279 + ], + "spans": [ + { + "bbox": [ + 135, + 202, + 219, + 279 + ], + "type": "image", + "image_path": "efbac54d69d8f13105e14fb834623f22042b2e63ab5a111fd257ce32e3572668.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 164, + 281, + 195, + 290 + ], + "lines": [ + { + "bbox": [ + 164, + 281, + 195, + 290 + ], + "spans": [ + { + "bbox": [ + 164, + 281, + 195, + 290 + ], + "type": "text", + "content": "TSP [28]" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 219, + 202, + 303, + 279 + ], + "blocks": [ + { + "bbox": [ + 219, + 202, + 303, + 279 + ], + "lines": [ + { + "bbox": [ + 219, + 202, + 303, + 279 + ], + "spans": [ + { + "bbox": [ + 219, + 202, + 303, + 279 + ], + "type": "image", + "image_path": "22200fbd5a291171e3eea3f76e1214106d8eb8621ce6d1bd46a69678b4418b35.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 247, + 281, + 278, + 290 + ], + "lines": [ + { + "bbox": [ + 247, + 281, + 278, + 290 + ], + "spans": [ + { + "bbox": [ + 247, + 281, + 278, + 290 + ], + "type": "text", + "content": "SFE [39]" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 303, + 202, + 386, + 279 + ], + "blocks": [ + { + "bbox": [ + 327, + 193, + 365, + 202 + ], + "lines": [ + { + "bbox": [ + 327, + 193, + 365, + 202 + ], + "spans": [ + { + "bbox": [ + 327, + 193, + 365, + 202 + ], + "type": "text", + "content": "EDVR [36]" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 303, + 202, + 386, + 279 + ], + "lines": [ + { + "bbox": [ + 303, + 202, + 386, + 279 + ], + "spans": [ + { + "bbox": [ + 303, + 202, + 386, + 279 + ], + "type": "image", + "image_path": "874ecc640e9f669020c65163817202607f6889547c22c72be6580916f9d30b7e.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 323, + 281, + 368, + 290 + ], + "lines": [ + { + "bbox": [ + 323, + 281, + 368, + 290 + ], + "spans": [ + { + "bbox": [ + 323, + 281, + 368, + 290 + ], + "type": "text", + "content": "DSTNet [29]" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 386, + 202, + 474, + 279 + ], + "blocks": [ + { + "bbox": [ + 406, + 193, + 453, + 202 + ], + "lines": [ + { + "bbox": [ + 406, + 193, + 453, + 202 + ], + "spans": [ + { + "bbox": [ + 406, + 193, + 453, + 202 + ], + "type": "text", + "content": "MPRNet [41]" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 386, + 202, + 474, + 279 + ], + "lines": [ + { + "bbox": [ + 386, + 202, + 474, + 279 + ], + "spans": [ + { + "bbox": [ + 386, + 202, + 474, + 279 + ], + "type": "image", + "image_path": "9ea57fc86eb8151b6f223207f911492332e77b338abb92c4d5161a351d8f43b0.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 401, + 281, + 455, + 290 + ], + "lines": [ + { + "bbox": [ + 401, + 281, + 455, + 290 + ], + "spans": [ + { + "bbox": [ + 401, + 281, + 455, + 290 + ], + "type": "text", + "content": "VD-Diff (Ours)" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 138, + 323, + 239, + 407 + ], + "blocks": [ + { + "bbox": [ + 132, + 299, + 480, + 312 + ], + "lines": [ + { + "bbox": [ + 132, + 299, + 480, + 312 + ], + "spans": [ + { + "bbox": [ + 132, + 299, + 480, + 312 + ], + "type": "text", + "content": "Fig. 6: Visual results of VD-Diff and SOTA methods on the real blurry videos of [33]." + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 138, + 323, + 239, + 407 + ], + "lines": [ + { + "bbox": [ + 138, + 323, + 239, + 407 + ], + "spans": [ + { + "bbox": [ + 138, + 323, + 239, + 407 + ], + "type": "image", + "image_path": "d32b83ab696924aaf341a27841053d837d771479a2920fa3fb7387c23f9fc179.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 246, + 323, + 280, + 408 + ], + "blocks": [ + { + "bbox": [ + 246, + 323, + 280, + 408 + ], + "lines": [ + { + "bbox": [ + 246, + 323, + 280, + 408 + ], + "spans": [ + { + "bbox": [ + 246, + 323, + 280, + 408 + ], + "type": "image", + "image_path": "0decf592cd7d2adbf19f250e3939315c0903a7a7a0679d5242b7fb1c886946d0.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 168, + 428, + 444, + 439 + ], + "lines": [ + { + "bbox": [ + 168, + 428, + 444, + 439 + ], + "spans": [ + { + "bbox": [ + 168, + 428, + 444, + 439 + ], + "type": "text", + "content": "Fig. 7: Comparison of various methods in terms of time dimension." + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 287, + 325, + 320, + 408 + ], + "blocks": [ + { + "bbox": [ + 287, + 325, + 320, + 408 + ], + "lines": [ + { + "bbox": [ + 287, + 325, + 320, + 408 + ], + "spans": [ + { + "bbox": [ + 287, + 325, + 320, + 408 + ], + "type": "image", + "image_path": "320c671e9a598d4221bd5131164e7e3a2c5f5fcae13092003561b5688d37e596.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 326, + 325, + 358, + 408 + ], + "blocks": [ + { + "bbox": [ + 326, + 325, + 358, + 408 + ], + "lines": [ + { + "bbox": [ + 326, + 325, + 358, + 408 + ], + "spans": [ + { + "bbox": [ + 326, + 325, + 358, + 408 + ], + "type": "image", + "image_path": "28979012f28514e8b93d683d295c3b0a5187c8a252bcee1b5f9580c7e12dfc2d.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 366, + 325, + 399, + 408 + ], + "blocks": [ + { + "bbox": [ + 366, + 325, + 399, + 408 + ], + "lines": [ + { + "bbox": [ + 366, + 325, + 399, + 408 + ], + "spans": [ + { + "bbox": [ + 366, + 325, + 399, + 408 + ], + "type": "image", + "image_path": "2d9c6f5184ee7c681a4596d8fd8813f4bd2bd35422a1c6b5b936f3e5593312d4.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 405, + 325, + 437, + 408 + ], + "blocks": [ + { + "bbox": [ + 405, + 325, + 437, + 408 + ], + "lines": [ + { + "bbox": [ + 405, + 325, + 437, + 408 + ], + "spans": [ + { + "bbox": [ + 405, + 325, + 437, + 408 + ], + "type": "image", + "image_path": "16f6d5ef7b030d8ae96258ef62e14ca97eec7c053fa8d2133cb070b20cd0a933.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 444, + 326, + 477, + 408 + ], + "blocks": [ + { + "bbox": [ + 444, + 326, + 477, + 408 + ], + "lines": [ + { + "bbox": [ + 444, + 326, + 477, + 408 + ], + "spans": [ + { + "bbox": [ + 444, + 326, + 477, + 408 + ], + "type": "image", + "image_path": "252d7ae0a05d0cc82d61f707d3fd8a092218923bc5775541698e2455b9267448.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + } + ], + "index": 25 + }, + { + "bbox": [ + 130, + 461, + 482, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 461, + 482, + 509 + ], + "spans": [ + { + "bbox": [ + 130, + 461, + 482, + 509 + ], + "type": "text", + "content": "DVD. The DVD [33] dataset consists of 71 videos with 6,708 blurry-sharp image pairs. It is divided into train/test subsets with 61 videos (5,708 image pairs) and 10 videos (1,000 image pairs). DVD is captured with mobile phones and DSLR at a frame rate of 240 FPS." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 130, + 510, + 482, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 482, + 545 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 482, + 545 + ], + "type": "text", + "content": "BSD. The Beam-Splitter Deblurring Dataset (BSD) [46] consists of 240 videos with 24k blurry-sharp image pairs. Cross-validation experiments between real-world and synthetic datasets demonstrate the high generality of the BSD dataset." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 130, + 545, + 482, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 482, + 582 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 482, + 582 + ], + "type": "text", + "content": "Real-World Videos. We evaluate the generality of our proposed approach on the real blurry datasets [6]. Due to the inaccessibility of the ground truth, we only compare VD-Diff with other models on visual results." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 131, + 598, + 280, + 610 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 598, + 280, + 610 + ], + "spans": [ + { + "bbox": [ + 131, + 598, + 280, + 610 + ], + "type": "text", + "content": "4.2 Implementation Details" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "We implement VD-Diff in PyTorch [30] using a single NVIDIA RTX A6000 GPU. The training process takes 7 days and 10 hours. We set the batch size as 2, the learning rate as " + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "inline_equation", + "content": "8 \\times 10^{-5}" + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": " and weight decay as " + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "inline_equation", + "content": "4 \\times 10^{-5}" + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": ". Patches at the size of " + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": " cropped from training frames are fed into the models. The sequence" + } + ] + } + ], + "index": 31 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "type": "text", + "content": "C. Rao et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 132, + 114, + 216, + 196 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 216, + 196 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 216, + 196 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 216, + 196 + ], + "type": "image", + "image_path": "273968429ab4dcf48ba4a49d95efa3888f1806a535940052610198baa32abfd9.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 221, + 114, + 304, + 196 + ], + "blocks": [ + { + "bbox": [ + 221, + 114, + 304, + 196 + ], + "lines": [ + { + "bbox": [ + 221, + 114, + 304, + 196 + ], + "spans": [ + { + "bbox": [ + 221, + 114, + 304, + 196 + ], + "type": "image", + "image_path": "663957c0f29f7d6cdb5aaa055089c788488e9cef9219fb30623a99c8ccd142b8.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 310, + 114, + 392, + 196 + ], + "blocks": [ + { + "bbox": [ + 310, + 114, + 392, + 196 + ], + "lines": [ + { + "bbox": [ + 310, + 114, + 392, + 196 + ], + "spans": [ + { + "bbox": [ + 310, + 114, + 392, + 196 + ], + "type": "image", + "image_path": "5e6d92ab721868fb6249f25e1f95cb4cd8f9920452d84f899096c0c781421582.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 399, + 114, + 481, + 196 + ], + "blocks": [ + { + "bbox": [ + 399, + 114, + 481, + 196 + ], + "lines": [ + { + "bbox": [ + 399, + 114, + 481, + 196 + ], + "spans": [ + { + "bbox": [ + 399, + 114, + 481, + 196 + ], + "type": "image", + "image_path": "b35cde1bf0a75cb1339908d4d57e6ae57b2040669040772fc244c660a9cfd57a.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 132, + 200, + 215, + 281 + ], + "blocks": [ + { + "bbox": [ + 132, + 200, + 215, + 281 + ], + "lines": [ + { + "bbox": [ + 132, + 200, + 215, + 281 + ], + "spans": [ + { + "bbox": [ + 132, + 200, + 215, + 281 + ], + "type": "image", + "image_path": "9e6570368fd84698e1d6ad41df0389c40af7bd47748e59e404739d764f83c3bb.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 165, + 283, + 181, + 291 + ], + "lines": [ + { + "bbox": [ + 165, + 283, + 181, + 291 + ], + "spans": [ + { + "bbox": [ + 165, + 283, + 181, + 291 + ], + "type": "text", + "content": "Blur" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 221, + 200, + 304, + 281 + ], + "blocks": [ + { + "bbox": [ + 221, + 200, + 304, + 281 + ], + "lines": [ + { + "bbox": [ + 221, + 200, + 304, + 281 + ], + "spans": [ + { + "bbox": [ + 221, + 200, + 304, + 281 + ], + "type": "image", + "image_path": "9ba28251ab1265077dbc045c9ebe82331ceb5c696f4905d5bcf1f08d6daae92b.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 253, + 283, + 272, + 291 + ], + "lines": [ + { + "bbox": [ + 253, + 283, + 272, + 291 + ], + "spans": [ + { + "bbox": [ + 253, + 283, + 272, + 291 + ], + "type": "text", + "content": "LDM" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 310, + 200, + 392, + 281 + ], + "blocks": [ + { + "bbox": [ + 310, + 200, + 392, + 281 + ], + "lines": [ + { + "bbox": [ + 310, + 200, + 392, + 281 + ], + "spans": [ + { + "bbox": [ + 310, + 200, + 392, + 281 + ], + "type": "image", + "image_path": "4c6262d3caf1ae557e900ac308d766f1f9489b47fad83881fad34439d4c028f1.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 343, + 283, + 359, + 291 + ], + "lines": [ + { + "bbox": [ + 343, + 283, + 359, + 291 + ], + "spans": [ + { + "bbox": [ + 343, + 283, + 359, + 291 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 399, + 200, + 481, + 281 + ], + "blocks": [ + { + "bbox": [ + 399, + 200, + 481, + 281 + ], + "lines": [ + { + "bbox": [ + 399, + 200, + 481, + 281 + ], + "spans": [ + { + "bbox": [ + 399, + 200, + 481, + 281 + ], + "type": "image", + "image_path": "14ff7918093f633e6c43de34e33771634f717c61b9c391c681af4640d0b17bfc.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 433, + 283, + 444, + 291 + ], + "lines": [ + { + "bbox": [ + 433, + 283, + 444, + 291 + ], + "spans": [ + { + "bbox": [ + 433, + 283, + 444, + 291 + ], + "type": "text", + "content": "GT" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 153, + 305, + 459, + 318 + ], + "lines": [ + { + "bbox": [ + 153, + 305, + 459, + 318 + ], + "spans": [ + { + "bbox": [ + 153, + 305, + 459, + 318 + ], + "type": "text", + "content": "Fig. 8: Qualitative comparison with pre-trained large diffusion models [31]." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 340, + 482, + 401 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 340, + 482, + 401 + ], + "spans": [ + { + "bbox": [ + 130, + 340, + 482, + 401 + ], + "type": "text", + "content": "length is set to 15 in training and 16 in testing. Since we use diffusion within a highly compressed latent space, the denoising iteration steps " + }, + { + "bbox": [ + 130, + 340, + 482, + 401 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 130, + 340, + 482, + 401 + ], + "type": "text", + "content": " only needs 4. We set " + }, + { + "bbox": [ + 130, + 340, + 482, + 401 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 130, + 340, + 482, + 401 + ], + "type": "text", + "content": " to 64 and " + }, + { + "bbox": [ + 130, + 340, + 482, + 401 + ], + "type": "inline_equation", + "content": "C'" + }, + { + "bbox": [ + 130, + 340, + 482, + 401 + ], + "type": "text", + "content": " to 256, respectively. In the WADT module, we set " + }, + { + "bbox": [ + 130, + 340, + 482, + 401 + ], + "type": "inline_equation", + "content": "N_1" + }, + { + "bbox": [ + 130, + 340, + 482, + 401 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 340, + 482, + 401 + ], + "type": "inline_equation", + "content": "N_2" + }, + { + "bbox": [ + 130, + 340, + 482, + 401 + ], + "type": "text", + "content": " to 7 and 2, respectively, to balance the calculation amount and the deblurring effect. The horizontal and vertical flips are performed for data augmentation." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 418, + 265, + 429 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 418, + 265, + 429 + ], + "spans": [ + { + "bbox": [ + 132, + 418, + 265, + 429 + ], + "type": "text", + "content": "4.3 Quantitative Results" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 130, + 437, + 482, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 437, + 482, + 605 + ], + "spans": [ + { + "bbox": [ + 130, + 437, + 482, + 605 + ], + "type": "text", + "content": "We compare our proposed VD-Diff with other SOTA methods, as shown in Tables 1, 2, and 3. As can be seen, our method outperforms SOTA methods on GoPro [25] and DVD [33] benchmark datasets. Specifically, as shown in Table 1, Our method outperforms DSTNet-L, RVRT, and FGST by " + }, + { + "bbox": [ + 130, + 437, + 482, + 605 + ], + "type": "inline_equation", + "content": "0.50\\mathrm{dB}" + }, + { + "bbox": [ + 130, + 437, + 482, + 605 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 437, + 482, + 605 + ], + "type": "inline_equation", + "content": "0.63\\mathrm{dB}" + }, + { + "bbox": [ + 130, + 437, + 482, + 605 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 437, + 482, + 605 + ], + "type": "inline_equation", + "content": "2.65\\mathrm{dB}" + }, + { + "bbox": [ + 130, + 437, + 482, + 605 + ], + "type": "text", + "content": " in PSNR on the GoPro [25] dataset, respectively. Our model achieves SOTA performance on the GoPro [25] dataset. Table 2 shows that our proposed method generates the deblurred videos with higher PSNR and SSIM values on the DVD [33] dataset. As shown in Table 3, we also achieve SOTA on the BSD [46] dataset. As seen from Table 4, our model has a relatively low parameter size and less inference time than other models. Although our model's parameter size and inference time are slightly higher than DSTNet and BasicVSR++, the PSNR metric on the GoPro [25] dataset is at least 1.3dB higher than the two models. As shown in Fig. 9(a), our model achieves favorable results in terms of accuracy and FLOPs. The above results show the effectiveness of our model." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 621, + 257, + 633 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 621, + 257, + 633 + ], + "spans": [ + { + "bbox": [ + 132, + 621, + 257, + 633 + ], + "type": "text", + "content": "4.4 Qualitative Results" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "content": "We provide visual comparisons on GoPro [25], DVD [33], and Real-World Video datasets [33] as shown in Fig. 4, 5, and 6, respectively. Some models sacrifice" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 217, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 217, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 217, + 91, + 447, + 102 + ], + "type": "text", + "content": "Video Deblurring with Transformer and Diffusion Model" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 130, + 114, + 474, + 186 + ], + "blocks": [ + { + "bbox": [ + 130, + 114, + 474, + 186 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 474, + 186 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 474, + 186 + ], + "type": "table", + "html": "
MethodDVD [33]EDVR [36]STFAN [47]ESTRNN [46]MPRNet [41]TSP [28]
PSNR ↑27.3126.8328.5931.0732.7331.67
SSIM ↑0.82550.84260.86080.90230.93660.9279
BasicVSR [3]NAFNet [4]FGST [22]DSTNet [29]DSTNet-L [29]RVRT [21]Ours
34.0133.7132.9034.1635.0534.9235.55
0.95200.96680.96100.96790.97330.97380.9752
", + "image_path": "ab5e5d3e94321f0481c888f4918d56f8f043bcb229032996891a623592a668d9.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 145, + 213, + 469, + 286 + ], + "blocks": [ + { + "bbox": [ + 167, + 186, + 445, + 198 + ], + "lines": [ + { + "bbox": [ + 167, + 186, + 445, + 198 + ], + "spans": [ + { + "bbox": [ + 167, + 186, + 445, + 198 + ], + "type": "text", + "content": "Table 1: VD-Diff achieves SOTA results on the GoPro dataset [25]." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 145, + 213, + 469, + 286 + ], + "lines": [ + { + "bbox": [ + 145, + 213, + 469, + 286 + ], + "spans": [ + { + "bbox": [ + 145, + 213, + 469, + 286 + ], + "type": "table", + "html": "
MethodDVD [33]EDVR [36]STFAN [47]ESTRNN [46]TSP [28]SEF [39]
PSNR ↑30.0128.5131.1532.0132.1331.71
SSIM ↑0.88770.86370.90490.91620.92680.9160
GSTA [34]MPRNet [41]ARVo [16]FGST [22]DSTNet [29]RVRT [21]Ours
32.5332.7332.8033.3633.7934.3034.56
0.94680.93660.93520.95000.96150.96550.9684
", + "image_path": "26dc8ad5c84bb2e3e4996b6700fb8ea0dfdd983c6feebe00e2186626619faba4.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 169, + 286, + 443, + 297 + ], + "lines": [ + { + "bbox": [ + 169, + 286, + 443, + 297 + ], + "spans": [ + { + "bbox": [ + 169, + 286, + 443, + 297 + ], + "type": "text", + "content": "Table 2: VD-Diff achieves SOTA results on the DVD dataset [33]." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 130, + 326, + 480, + 432 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 326, + 480, + 432 + ], + "spans": [ + { + "bbox": [ + 130, + 326, + 480, + 432 + ], + "type": "text", + "content": "texture and details to maintain overall structure, resulting in over-smooth videos that lose high-frequency detail. Other models ignore global information to preserve high-frequency detail, introducing redundant speckle textures and artifacts that do not match the real video in fast-motion video. In contrast, our VD-Diff has the best visual effect as it can leverage WADT to model long-range dependencies to capture low-frequency information and employ DM to supplement valuable high-frequency information. Therefore, as seen from the visual comparison, our model can restore texture and details while maintaining the overall framework of the video." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 434, + 480, + 493 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 434, + 480, + 493 + ], + "spans": [ + { + "bbox": [ + 130, + 434, + 480, + 493 + ], + "type": "text", + "content": "Besides, we conduct a qualitative comparison in terms of the time dimension, as shown in Fig. 7. As can be seen, previous models such as EDVR [36], MPRNet [41], TSP [28], and DSTNet [29] all exhibit unclear outlines, indicating that they cannot guarantee temporal consistency. In contrast, our method effectively maintains temporal consistency." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 494, + 481, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 494, + 481, + 578 + ], + "spans": [ + { + "bbox": [ + 130, + 494, + 481, + 578 + ], + "type": "text", + "content": "Furthermore, we conduct a comparison with the direct application of a pretrained large diffusion model (i.e., LDM [31]) to video deblurring, as shown in Fig. 8. We notice that the videos deblurred by LDM contain a significant number of artifacts and inconsistencies with the content of the original videos, which is attributed to motion blur present in the blurred videos. In contrast, our method leverages the diffusion model and WADT to maximally restore the details in the video while preserving the video content." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 596, + 239, + 609 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 596, + 239, + 609 + ], + "spans": [ + { + "bbox": [ + 132, + 596, + 239, + 609 + ], + "type": "text", + "content": "4.5 Ablation Study" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 617, + 480, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 480, + 640 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 480, + 640 + ], + "type": "text", + "content": "This section explores the significance of each key component of our proposed VD-Diff. All variants are tested on the GoPro [25] dataset." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "content": "Effect of Diffusion Model. To validate the effectiveness of the diffusion model, we design a variant that does not utilize DM to generate prior features" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "type": "text", + "content": "C. Rao et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 135, + 115, + 483, + 152 + ], + "blocks": [ + { + "bbox": [ + 135, + 115, + 483, + 152 + ], + "lines": [ + { + "bbox": [ + 135, + 115, + 483, + 152 + ], + "spans": [ + { + "bbox": [ + 135, + 115, + 483, + 152 + ], + "type": "table", + "html": "
MethodDVD [33]DTBN [14]Nah et al. [26]ESTRNN [46]TSP [28]DSTNet [29]Ours
PSNR ↑29.9531.8433.0033.3632.8434.4535.61
SSIM ↑0.86920.91700.93300.93700.93980.95480.9758
", + "image_path": "4b627d2a73a995ae0d8944bd2aab3dac29bbeee15c77d5c8acb9b42722c42db6.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 135, + 180, + 491, + 218 + ], + "blocks": [ + { + "bbox": [ + 171, + 153, + 441, + 164 + ], + "lines": [ + { + "bbox": [ + 171, + 153, + 441, + 164 + ], + "spans": [ + { + "bbox": [ + 171, + 153, + 441, + 164 + ], + "type": "text", + "content": "Table 3: VD-Diff achieves SOTA results on the BSD dataset [46]." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 135, + 180, + 491, + 218 + ], + "lines": [ + { + "bbox": [ + 135, + 180, + 491, + 218 + ], + "spans": [ + { + "bbox": [ + 135, + 180, + 491, + 218 + ], + "type": "table", + "html": "
MethodDVDEDVRTSPFGSTDSTNetBasicVSR++RVRTDSTNet-LOurs
Parameters (M)15.3023.6016.199.707.459.4613.6016.9611.93
Running time (s)0.882.701.900.720.080.100.410.280.31
", + "image_path": "5c2dd916ad8719c9f00db4d698f504f18f416c29728a9aac1a5befb8a8f0df27.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 142, + 284, + 471, + 361 + ], + "blocks": [ + { + "bbox": [ + 132, + 220, + 480, + 252 + ], + "lines": [ + { + "bbox": [ + 132, + 220, + 480, + 252 + ], + "spans": [ + { + "bbox": [ + 132, + 220, + 480, + 252 + ], + "type": "text", + "content": "Table 4: Quantitative evaluations of the video deblurring methods regarding network parameters and running time. The running time is obtained on a machine with a single NVIDIA RTX A6000 GPU. The size of the test images is " + }, + { + "bbox": [ + 132, + 220, + 480, + 252 + ], + "type": "inline_equation", + "content": "1280 \\times 720" + }, + { + "bbox": [ + 132, + 220, + 480, + 252 + ], + "type": "text", + "content": " pixels." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 142, + 284, + 471, + 361 + ], + "lines": [ + { + "bbox": [ + 142, + 284, + 471, + 361 + ], + "spans": [ + { + "bbox": [ + 142, + 284, + 471, + 361 + ], + "type": "table", + "html": "
VariantsComponentsMetrics
Diffusion ModelWaveletJoint TrainingWBPFPSNR ↑SSIM ↑
w/o DMX✓✓✓34.380.9684
w/o Wavelet✓X✓✓34.750.9709
w/o Joint Training✓✓X✓34.920.9717
w/o WBPF✓✓✓X33.580.9344
Full model✓✓✓✓35.550.9752
", + "image_path": "73dd92f7f128c44514264e7d95dad46075ceab3b61e05180575544a831cab57c.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 361, + 479, + 383 + ], + "lines": [ + { + "bbox": [ + 132, + 361, + 479, + 383 + ], + "spans": [ + { + "bbox": [ + 132, + 361, + 479, + 383 + ], + "type": "text", + "content": "Table 5: Ablation experiments on various variants of VD-Diif on the GoPro [25] dataset, and the best quantitative results are marked in bold." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 134, + 407, + 250, + 502 + ], + "blocks": [ + { + "bbox": [ + 134, + 407, + 250, + 502 + ], + "lines": [ + { + "bbox": [ + 134, + 407, + 250, + 502 + ], + "spans": [ + { + "bbox": [ + 134, + 407, + 250, + 502 + ], + "type": "image", + "image_path": "516623607e2afb29e59b91eac0e5b7b0288129edcf7c9d7e5a21e6b270f36bfd.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 513, + 480, + 557 + ], + "lines": [ + { + "bbox": [ + 130, + 513, + 480, + 557 + ], + "spans": [ + { + "bbox": [ + 130, + 513, + 480, + 557 + ], + "type": "text", + "content": "Fig. 9: (a) Floating point operations (FLOPs) vs video deblurring performance on the GoPro [25] dataset. Our model achieves favorable results in terms of accuracy and FLOPs. (b) Ablation study of the number of iteration steps. (c) Ablation study of the sequence length during inference." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 251, + 407, + 363, + 499 + ], + "blocks": [ + { + "bbox": [ + 251, + 407, + 363, + 499 + ], + "lines": [ + { + "bbox": [ + 251, + 407, + 363, + 499 + ], + "spans": [ + { + "bbox": [ + 251, + 407, + 363, + 499 + ], + "type": "image", + "image_path": "edec444189fc7d2e64b0ffbcbb461bbf0efd8f498909e5ca7ac29d335e4835c8.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 365, + 407, + 479, + 502 + ], + "blocks": [ + { + "bbox": [ + 365, + 407, + 479, + 502 + ], + "lines": [ + { + "bbox": [ + 365, + 407, + 479, + 502 + ], + "spans": [ + { + "bbox": [ + 365, + 407, + 479, + 502 + ], + "type": "image", + "image_path": "70fcc8cfaa5e45b0bc13359405234be3819008fba7d5859e161ed4bc475d12d1.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 581, + 479, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 581, + 479, + 628 + ], + "spans": [ + { + "bbox": [ + 130, + 581, + 479, + 628 + ], + "type": "text", + "content": "but solely relies on WADT for video deblurring, named as " + }, + { + "bbox": [ + 130, + 581, + 479, + 628 + ], + "type": "inline_equation", + "content": "w / o" + }, + { + "bbox": [ + 130, + 581, + 479, + 628 + ], + "type": "text", + "content": " DM. As seen in Table 5, without employing DM, the deblurring performance of the network significantly declines, with a decrease of 1.17 in PSNR, which indicates that DM can provide valuable prior features to enhance the deblurring performance." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 629, + 480, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 480, + 665 + ], + "type": "text", + "content": "Effect of Wavelet Transform. We conduct ablation experiments to study the impact of wavelet transform on VD-Diff. We replace wavelet transform and inverse wavelet transform with bilinear downsample and upsample to evaluate" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 217, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 217, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 217, + 91, + 447, + 101 + ], + "type": "text", + "content": "Video Deblurring with Transformer and Diffusion Model" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "text", + "content": "the effect of wavelet transform, named " + }, + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "inline_equation", + "content": "w / o" + }, + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "text", + "content": " Wavelet. We adopt the same training strategy as VD-Diff. Table 5 shows " + }, + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "inline_equation", + "content": "w / o" + }, + { + "bbox": [ + 130, + 116, + 479, + 163 + ], + "type": "text", + "content": " Wavelet decreases by 0.8dB regarding the PSNR metrics. This experiment shows the wavelet transform method for feature extraction is better than the general upsampling and downsampling methods." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 165, + 480, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 165, + 480, + 236 + ], + "spans": [ + { + "bbox": [ + 130, + 165, + 480, + 236 + ], + "type": "text", + "content": "Effect of Joint Training Strategy. In Sec. 3.4, we analyze the importance of joint training. To verify our analysis, we remove joint training stage i.e., Training Stage Three, as shown in Table 5. We notice that VD-Diff is significantly better than " + }, + { + "bbox": [ + 130, + 165, + 480, + 236 + ], + "type": "inline_equation", + "content": "w / o" + }, + { + "bbox": [ + 130, + 165, + 480, + 236 + ], + "type": "text", + "content": " Joint Training with an improvement of 0.63dB in terms of PSNR metrics, which demonstrates that joint training can improve the collaborative working ability of DM and WADT, thereby improving the deblurring effect." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 238, + 480, + 297 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 238, + 480, + 297 + ], + "spans": [ + { + "bbox": [ + 130, + 238, + 480, + 297 + ], + "type": "text", + "content": "Effect of Wavelet-based Bidirectional Propagation Fuse. To investigate the effect of Wavelet-based Bidirectional Propagation Fuse (WBPF), we design a variant model by removing the WBPF in WADT, termed as " + }, + { + "bbox": [ + 130, + 238, + 480, + 297 + ], + "type": "inline_equation", + "content": "w / o" + }, + { + "bbox": [ + 130, + 238, + 480, + 297 + ], + "type": "text", + "content": " WBPF. Table 5 shows that WBPF can effectively transmit useful information between frames, improving the model's ability to capture spatio-temporal information." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 299, + 480, + 382 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 299, + 480, + 382 + ], + "spans": [ + { + "bbox": [ + 130, + 299, + 480, + 382 + ], + "type": "text", + "content": "Effect of Iteration Steps. To explore the impact of the number of iterations on the diffusion model, we set different iteration steps to train the model and then compare the PSNR metrics on the GoPro dataset. The results are shown in Fig. 9(b). The value of PSNR increases as the number of iteration steps increases, but when the number of iterations is greater than 4, the value of PSNR almost remains stable. It shows that in a highly compressed latent space, just setting the number of iteration steps to 4 can generate suitable prior features." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 384, + 480, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 384, + 480, + 502 + ], + "spans": [ + { + "bbox": [ + 130, + 384, + 480, + 502 + ], + "type": "text", + "content": "Effect of Sequence Length. Intuitively, the sequence length will affect the deblurring effect of our proposed model. If the sequence length is too short, the model fails to capture essential information from distant frames. Conversely, if the sequence length is too long, the computational load becomes excessive, potentially causing error accumulation. To investigate this, we experimented with various sequence lengths. As shown in Fig. 9(c), the PSNR value increases steadily with the number of frames. However, after 16 frames, the calculation becomes very large as the sequence length increases, and the PSNR does not change significantly. Therefore, we set the sequence length to 16 to reach a near-optimal point between the deblurring effect and computational reduction." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 527, + 219, + 539 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 527, + 219, + 539 + ], + "spans": [ + { + "bbox": [ + 132, + 527, + 219, + 539 + ], + "type": "text", + "content": "5 Conclusion" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 558, + 479, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 558, + 479, + 664 + ], + "spans": [ + { + "bbox": [ + 130, + 558, + 479, + 664 + ], + "type": "text", + "content": "In this paper, we propose a brand-new method that innovatively integrates the diffusion model with the Transformer for video deblurring. Specifically, the diffusion model generates prior features containing high-frequency information that conforms to the ground truth (GT) distribution. Meanwhile, the Wavelet-Aware Dynamic Transformer captures low-frequency information and utilizes the prior features generated by the diffusion model to supplement high-frequency information, thereby enhancing the quality of the deblurred video. Comprehensive experimental results show that our proposed method outperforms the state-of-the-art methods." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "text", + "content": "C. Rao et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 246, + 129 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 138, + 482, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 138, + 482, + 236 + ], + "spans": [ + { + "bbox": [ + 130, + 138, + 482, + 236 + ], + "type": "text", + "content": "This work was supported in part by National Social Science Foundation Major Project \"Research on Virtual Restoration of Tang and Song Painting Colors and Construction of Traditional Color Resource Library from a Digital Perspective\" (19ZDA046), Zhejiang Province Program (2022C01222, 2023C03199, 2023C03201), the National Program of China (62172365, 2021YFF0900604, 19ZDA197), Ningbo Science and Technology Plan Project (2022Z167, 2023Z137), and MOE Frontier Science Center for Brain Science & Brain-Machine Integration (Zhejiang University)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 251, + 197, + 263 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 251, + 197, + 263 + ], + "spans": [ + { + "bbox": [ + 133, + 251, + 197, + 263 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 134, + 275, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 138, + 275, + 480, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 275, + 480, + 296 + ], + "spans": [ + { + "bbox": [ + 138, + 275, + 480, + 296 + ], + "type": "text", + "content": "1. Ba, J.L., Kiros, J.R., Hinton, G.E.: Layer normalization. arXiv preprint arXiv:1607.06450 (2016)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 297, + 480, + 339 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 297, + 480, + 339 + ], + "spans": [ + { + "bbox": [ + 138, + 297, + 480, + 339 + ], + "type": "text", + "content": "2. Blattmann, A., Rombach, R., Ling, H., Dockhorn, T., Kim, S.W., Fidler, S., Kreis, K.: Align your latents: High-resolution video synthesis with latent diffusion models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 22563-22575 (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 341, + 480, + 361 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 341, + 480, + 361 + ], + "spans": [ + { + "bbox": [ + 138, + 341, + 480, + 361 + ], + "type": "text", + "content": "3. Chan, K.C., Zhou, S., Xu, X., Loy, C.C.: On the generalization of basicvsr++ to video deblurring and denoising. arXiv preprint arXiv:2204.05308 (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 362, + 480, + 383 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 362, + 480, + 383 + ], + "spans": [ + { + "bbox": [ + 138, + 362, + 480, + 383 + ], + "type": "text", + "content": "4. Chen, L., Chu, X., Zhang, X., Sun, J.: Simple baselines for image restoration. In: European Conference on Computer Vision. pp. 17-33. Springer (2022)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 384, + 480, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 384, + 480, + 415 + ], + "spans": [ + { + "bbox": [ + 138, + 384, + 480, + 415 + ], + "type": "text", + "content": "5. Chen, Z., Zhang, Y., Liu, D., Xia, B., Gu, J., Kong, L., Yuan, X.: Hierarchical integration diffusion model for realistic image deblurring. arXiv preprint arXiv:2305.12966 (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 417, + 480, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 417, + 480, + 437 + ], + "spans": [ + { + "bbox": [ + 138, + 417, + 480, + 437 + ], + "type": "text", + "content": "6. Cho, S., Wang, J., Lee, S.: Video deblurring for hand-held cameras using patch-based synthesis. ACM Transactions on Graphics (TOG) 31(4), 1-9 (2012)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 438, + 480, + 469 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 438, + 480, + 469 + ], + "spans": [ + { + "bbox": [ + 138, + 438, + 480, + 469 + ], + "type": "text", + "content": "7. Chu, T., Chen, J., Sun, J., Lian, S., Wang, Z., Zuo, Z., Zhao, L., Xing, W., Lu, D.: Rethinking fast fourier convolution in image inpainting. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 23195-23205 (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 471, + 480, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 471, + 480, + 514 + ], + "spans": [ + { + "bbox": [ + 138, + 471, + 480, + 514 + ], + "type": "text", + "content": "8. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929 (2020)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 514, + 480, + 546 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 514, + 480, + 546 + ], + "spans": [ + { + "bbox": [ + 138, + 514, + 480, + 546 + ], + "type": "text", + "content": "9. Esser, P., Chiu, J., Atighehchian, P., Granskog, J., Germanidis, A.: Structure and content-guided video synthesis with diffusion models. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 7346-7356 (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 134, + 547, + 480, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 547, + 480, + 578 + ], + "spans": [ + { + "bbox": [ + 134, + 547, + 480, + 578 + ], + "type": "text", + "content": "10. Fan, Y., Hong, C., Wang, X., Zeng, Z., Guo, Z.: Multi-input-output fusion attention module for deblurring networks. In: 2021 IEEE international conference on big data (Big Data). pp. 3176-3182. IEEE (2021)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 134, + 579, + 480, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 579, + 480, + 600 + ], + "spans": [ + { + "bbox": [ + 134, + 579, + 480, + 600 + ], + "type": "text", + "content": "11. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. Advances in neural information processing systems 33, 6840-6851 (2020)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 134, + 601, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 601, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 134, + 601, + 480, + 632 + ], + "type": "text", + "content": "12. Huang, H., He, R., Sun, Z., Tan, T.: Wavelet-srnet: A wavelet-based cnn for multiscale face super resolution. In: Proceedings of the IEEE international conference on computer vision. pp. 1689-1697 (2017)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 134, + 633, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 633, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 134, + 633, + 480, + 665 + ], + "type": "text", + "content": "13. Hyun Kim, T., Mu Lee, K.: Generalized video deblurring for dynamic scenes. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 5426-5434 (2015)" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 217, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 217, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 217, + 91, + 447, + 101 + ], + "type": "text", + "content": "Video Deblurring with Transformer and Diffusion Model" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 666 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "type": "text", + "content": "14. Hyun Kim, T., Mu Lee, K., Scholkopf, B., Hirsch, M.: Online video deblurring via dynamic temporal blending network. In: Proceedings of the IEEE international conference on computer vision. pp. 4038-4047 (2017)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 151, + 481, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 151, + 481, + 172 + ], + "spans": [ + { + "bbox": [ + 133, + 151, + 481, + 172 + ], + "type": "text", + "content": "15. Li, A., Zhao, L., Zuo, Z., Wang, Z., Xing, W., Lu, D.: Migt: Multi-modal image inpainting guided with text. Neurocomputing 520, 376-385 (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 173, + 481, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 173, + 481, + 217 + ], + "spans": [ + { + "bbox": [ + 133, + 173, + 481, + 217 + ], + "type": "text", + "content": "16. Li, D., Xu, C., Zhang, K., Yu, X., Zhong, Y., Ren, W., Suominen, H., Li, H.: Arvo: Learning all-range volumetric correspondence for video deblurring. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7721-7731 (2021)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 218, + 481, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 218, + 481, + 262 + ], + "spans": [ + { + "bbox": [ + 132, + 218, + 481, + 262 + ], + "type": "text", + "content": "17. Li, G., Lyu, J., Wang, C., Dou, Q., Qin, J.: Wavtrans: Synergizing wavelet and cross-attention transformer for multi-contrast mri super-resolution. In: International Conference on Medical Image Computing and Computer-Assisted Intervention. pp. 463–473. Springer (2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 262, + 481, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 262, + 481, + 296 + ], + "spans": [ + { + "bbox": [ + 132, + 262, + 481, + 296 + ], + "type": "text", + "content": "18. Li, G., Rao, C., Mo, J., Zhang, Z., Xing, W., Zhao, L.: Rethinking diffusion model for multi-contrast mri super-resolution. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11365-11374 (2024)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 297, + 481, + 329 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 297, + 481, + 329 + ], + "spans": [ + { + "bbox": [ + 132, + 297, + 481, + 329 + ], + "type": "text", + "content": "19. Li, W., Zhao, L., Lin, Z., Xu, D., Lu, D.: Non-local image inpainting using low-rank matrix completion. In: Computer Graphics Forum. vol. 34, pp. 111-122. Wiley Online Library (2015)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 330, + 481, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 330, + 481, + 363 + ], + "spans": [ + { + "bbox": [ + 132, + 330, + 481, + 363 + ], + "type": "text", + "content": "20. Li, W., Zhao, L., Xu, D., Lu, D.: Efficient image completion method based on alternating direction theory. In: 2013 IEEE International Conference on Image Processing. pp. 700-703. IEEE (2013)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 364, + 481, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 364, + 481, + 407 + ], + "spans": [ + { + "bbox": [ + 132, + 364, + 481, + 407 + ], + "type": "text", + "content": "21. Liang, J., Fan, Y., Xiang, X., Ranjan, R., Ilg, E., Green, S., Cao, J., Zhang, K., Timofte, R., Gool, L.V.: Recurrent video restoration transformer with guided deformable attention. Advances in Neural Information Processing Systems 35, 378-393 (2022)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 409, + 481, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 409, + 481, + 441 + ], + "spans": [ + { + "bbox": [ + 132, + 409, + 481, + 441 + ], + "type": "text", + "content": "22. Lin, J., Cai, Y., Hu, X., Wang, H., Yan, Y., Zou, X., Ding, H., Zhang, Y., Timofte, R., Van Gool, L.: Flow-guided sparse transformer for video deblurring. arXiv preprint arXiv:2201.01893 (2022)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 442, + 481, + 485 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 442, + 481, + 485 + ], + "spans": [ + { + "bbox": [ + 132, + 442, + 481, + 485 + ], + "type": "text", + "content": "23. Liu, X., Park, D.H., Azadi, S., Zhang, G., Chopikyan, A., Hu, Y., Shi, H., Rohrbach, A., Darrell, T.: More control for free! image synthesis with semantic diffusion guidance. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV). pp. 289-299 (January 2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 487, + 481, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 487, + 481, + 531 + ], + "spans": [ + { + "bbox": [ + 132, + 487, + 481, + 531 + ], + "type": "text", + "content": "24. Luo, Z., Chen, D., Zhang, Y., Huang, Y., Wang, L., Shen, Y., Zhao, D., Zhou, J., Tan, T.: Videofusion: Decomposed diffusion models for high-quality video generation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10209-10218 (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 532, + 481, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 532, + 481, + 564 + ], + "spans": [ + { + "bbox": [ + 132, + 532, + 481, + 564 + ], + "type": "text", + "content": "25. Nah, S., Hyun Kim, T., Mu Lee, K.: Deep multi-scale convolutional neural network for dynamic scene deblurring. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 3883-3891 (2017)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 565, + 481, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 565, + 481, + 597 + ], + "spans": [ + { + "bbox": [ + 132, + 565, + 481, + 597 + ], + "type": "text", + "content": "26. Nah, S., Son, S., Lee, K.M.: Recurrent neural networks with intra-frame iterations for video deblurring. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 8102-8111 (2019)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 599, + 481, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 599, + 481, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 599, + 481, + 632 + ], + "type": "text", + "content": "27. Ozdenizci, O., Legenstein, R.: Restoring vision in adverse weather conditions with patch-based denoising diffusion models. IEEE Transactions on Pattern Analysis and Machine Intelligence (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 632, + 481, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 481, + 666 + ], + "type": "text", + "content": "28. Pan, J., Bai, H., Tang, J.: Cascaded deep video deblurring using temporal sharpness prior. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 3043-3051 (2020)" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "text", + "content": "C. Rao et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 665 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "text", + "content": "29. Pan, J., Xu, B., Dong, J., Ge, J., Tang, J.: Deep discriminative spatial and temporal network for efficient video deblurring. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 22191-22200 (2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 150, + 482, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 150, + 482, + 194 + ], + "spans": [ + { + "bbox": [ + 130, + 150, + 482, + 194 + ], + "type": "text", + "content": "30. Paszke, A., Gross, S., Massa, F., Lerer, A., Bradbury, J., Chanan, G., Killeen, T., Lin, Z., Gimelshein, N., Antiga, L., et al.: Pytorch: An imperative style, high-performance deep learning library. Advances in neural information processing systems 32 (2019)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 194, + 482, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 194, + 482, + 226 + ], + "spans": [ + { + "bbox": [ + 130, + 194, + 482, + 226 + ], + "type": "text", + "content": "31. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10684-10695 (2022)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 227, + 482, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 227, + 482, + 270 + ], + "spans": [ + { + "bbox": [ + 130, + 227, + 482, + 270 + ], + "type": "text", + "content": "32. Ruan, L., Ma, Y., Yang, H., He, H., Liu, B., Fu, J., Yuan, N.J., Jin, Q., Guo, B.: Mm-diffusion: Learning multi-modal diffusion models for joint audio and video generation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10219-10228 (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 271, + 482, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 271, + 482, + 304 + ], + "spans": [ + { + "bbox": [ + 130, + 271, + 482, + 304 + ], + "type": "text", + "content": "33. Su, S., Delbracio, M., Wang, J., Sapiro, G., Heidrich, W., Wang, O.: Deep video deblurring for hand-held cameras. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1279-1288 (2017)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 304, + 482, + 336 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 304, + 482, + 336 + ], + "spans": [ + { + "bbox": [ + 130, + 304, + 482, + 336 + ], + "type": "text", + "content": "34. Suin, M., Rajagopalan, A.: Gated spatio-temporal attention-guided video deblurring. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7802-7811 (2021)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 336, + 482, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 336, + 482, + 369 + ], + "spans": [ + { + "bbox": [ + 130, + 336, + 482, + 369 + ], + "type": "text", + "content": "35. Tashiro, Y., Song, J., Song, Y., Ermon, S.: Csdi: Conditional score-based diffusion models for probabilistic time series imputation. Advances in Neural Information Processing Systems 34, 24804-24816 (2021)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 369, + 482, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 369, + 482, + 413 + ], + "spans": [ + { + "bbox": [ + 130, + 369, + 482, + 413 + ], + "type": "text", + "content": "36. Wang, X., Chan, K.C., Yu, K., Dong, C., Change Loy, C.: Edvr: Video restoration with enhanced deformable convolutional networks. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops. pp. 0-0 (2019)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 413, + 482, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 413, + 482, + 445 + ], + "spans": [ + { + "bbox": [ + 130, + 413, + 482, + 445 + ], + "type": "text", + "content": "37. Wang, Z., Cun, X., Bao, J., Zhou, W., Liu, J., Li, H.: Uformer: A general u-shaped transformer for image restoration. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 17683-17693 (2022)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 446, + 482, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 446, + 482, + 479 + ], + "spans": [ + { + "bbox": [ + 130, + 446, + 482, + 479 + ], + "type": "text", + "content": "38. Xia, B., Zhang, Y., Wang, S., Wang, Y., Wu, X., Tian, Y., Yang, W., Van Gool, L.: Diffir: Efficient diffusion model for image restoration. arXiv preprint arXiv:2303.09472 (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 479, + 482, + 501 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 479, + 482, + 501 + ], + "spans": [ + { + "bbox": [ + 130, + 479, + 482, + 501 + ], + "type": "text", + "content": "39. Xiang, X., Wei, H., Pan, J.: Deep video deblurring using sharpness features from exemplars. IEEE Transactions on Image Processing 29, 8976-8987 (2020)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 501, + 482, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 501, + 482, + 533 + ], + "spans": [ + { + "bbox": [ + 130, + 501, + 482, + 533 + ], + "type": "text", + "content": "40. Yu, S., Sohn, K., Kim, S., Shin, J.: Video probabilistic diffusion models in projected latent space. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 18456-18466 (June 2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 534, + 482, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 534, + 482, + 566 + ], + "spans": [ + { + "bbox": [ + 130, + 534, + 482, + 566 + ], + "type": "text", + "content": "41. Zamir, S.W., Arora, A., Khan, S., Hayat, M., Khan, F.S., Yang, M.H., Shao, L.: Multi-stage progressive image restoration. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 14821-14831 (2021)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 567, + 482, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 567, + 482, + 599 + ], + "spans": [ + { + "bbox": [ + 130, + 567, + 482, + 599 + ], + "type": "text", + "content": "42. Zhang, K., Luo, W., Zhong, Y., Ma, L., Liu, W., Li, H.: Adversarial spatio-temporal learning for video deblurring. IEEE Transactions on Image Processing 28(1), 291-301 (2018)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 600, + 482, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 600, + 482, + 643 + ], + "spans": [ + { + "bbox": [ + 130, + 600, + 482, + 643 + ], + "type": "text", + "content": "43. Zhang, Z., Zhang, Q., Xing, W., Li, G., Zhao, L., Sun, J., Lan, Z., Luan, J., Huang, Y., Lin, H.: Artbank: Artistic style transfer with pre-trained diffusion model and implicit style prompt bank. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 38, pp. 7396-7404 (2024)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 130, + 643, + 482, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 643, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 643, + 482, + 665 + ], + "type": "text", + "content": "44. Zhao, L., Mo, Q., Lin, S., Wang, Z., Zuo, Z., Chen, H., Xing, W., Lu, D.: Uctgan: Diverse image inpainting based on unsupervised cross-space translation. In: Pro" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 217, + 91, + 447, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 217, + 91, + 447, + 101 + ], + "spans": [ + { + "bbox": [ + 217, + 91, + 447, + 101 + ], + "type": "text", + "content": "Video Deblurring with Transformer and Diffusion Model" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 117, + 481, + 293 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 149, + 117, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 117, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 149, + 117, + 480, + 138 + ], + "type": "text", + "content": "ceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 5741-5750 (2020)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 140, + 481, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 140, + 481, + 171 + ], + "spans": [ + { + "bbox": [ + 133, + 140, + 481, + 171 + ], + "type": "text", + "content": "45. Zheng, B., Yuan, S., Slabaugh, G., Leonardis, A.: Image demoireing with learnable bandpass filters. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3636-3645 (2020)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 172, + 481, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 172, + 481, + 215 + ], + "spans": [ + { + "bbox": [ + 133, + 172, + 481, + 215 + ], + "type": "text", + "content": "46. Zhong, Z., Gao, Y., Zheng, Y., Zheng, B.: Efficient spatio-temporal recurrent neural network for video deblurring. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part VI 16. pp. 191-207. Springer (2020)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 216, + 481, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 216, + 481, + 248 + ], + "spans": [ + { + "bbox": [ + 133, + 216, + 481, + 248 + ], + "type": "text", + "content": "47. Zhou, S., Zhang, J., Pan, J., Xie, H., Zuo, W., Ren, J.: Spatio-temporal filter adaptive network for video deblurring. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 2482-2491 (2019)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 133, + 249, + 481, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 249, + 481, + 293 + ], + "spans": [ + { + "bbox": [ + 133, + 249, + 481, + 293 + ], + "type": "text", + "content": "48. Zuo, Z., Zhao, L., Li, A., Wang, Z., Zhang, Z., Chen, J., Xing, W., Lu, D.: Generative image inpainting with segmentation confusion adversarial training and contrastive learning. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 37, pp. 3888-3896 (2023)" + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "text", + "content": "C.Rao et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Rethinking Video-Text Understanding_ Retrieval from Counterfactually Augmented Data/21cc04b7-ea52-4351-82c2-826a88e983b2_content_list.json b/2024/Rethinking Video-Text Understanding_ Retrieval from Counterfactually Augmented Data/21cc04b7-ea52-4351-82c2-826a88e983b2_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..d24ef2562beba8673ef8703672b91ed4e8ca8309 --- /dev/null +++ b/2024/Rethinking Video-Text Understanding_ Retrieval from Counterfactually Augmented Data/21cc04b7-ea52-4351-82c2-826a88e983b2_content_list.json @@ -0,0 +1,1775 @@ +[ + { + "type": "text", + "text": "Rethinking Video-Text Understanding: Retrieval from Counterfactually Augmented Data", + "text_level": 1, + "bbox": [ + 217, + 140, + 787, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Wufei Ma $^{1,2\\star}$ , Kai Li $^{1}$ , Zhongshi Jiang $^{1}$ , Moustafa Meshry $^{1}$ , Qihao Liu $^{2}$ , Huiyu Wang $^{3}$ , Christian Hane $^{1}$ , and Alan Yuille $^{2}$", + "bbox": [ + 241, + 210, + 759, + 244 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Meta Reality Labs", + "bbox": [ + 431, + 253, + 571, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "2 Johns Hopkins University", + "bbox": [ + 406, + 268, + 594, + 282 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3 Meta AI", + "bbox": [ + 464, + 282, + 537, + 295 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Recent video-text foundation models have demonstrated strong performance on a wide variety of downstream video understanding tasks. Can these video-text models genuinely understand the contents of natural videos? Standard video-text evaluations could be misleading as many questions can be inferred merely from the objects and contexts in a single frame or biases inherent in the datasets. In this paper, we aim to better assess the capabilities of current video-text models and understand their limitations. We propose a novel evaluation task for video-text understanding, namely retrieval from counterfactually augmented data (RCAD), and a new Feint6K dataset. To succeed on our new evaluation task, models must derive a comprehensive understanding of the video from cross-frame reasoning. Analyses show that previous video-text foundation models can be easily fooled by counterfactually augmented data and are far behind human-level performance. In order to narrow the gap between video-text models and human performance on RCAD, we identify a key limitation of current contrastive approaches on video-text data and introduce LLM-teacher, a more effective approach to learn action semantics by leveraging knowledge obtained from a pretrained large language model. Experiments and analyses show that our approach successfully learn more discriminative action embeddings and improves results on Feint6K when applied to multiple video-text models. Our Feint6K dataset and project page is available here.", + "bbox": [ + 261, + 339, + 754, + 645 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Video-Text Understanding $\\cdot$ Retrieval", + "bbox": [ + 261, + 659, + 604, + 674 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 702, + 375, + 718 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Video-text foundation models have gained increasing attention due to their simple formulation and strong transferability [12, 16, 23]. By pretraining on web scale video-text datasets, these models demonstrate strong performance across a wide range of downstream tasks, such as video-text retrieval [22, 28] and video question answering [25].", + "bbox": [ + 212, + 737, + 787, + 814 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* All data collection and experiments were conducted at JHU.", + "bbox": [ + 217, + 824, + 635, + 839 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "As video-text foundation models evolve and achieve increasingly better performance on various benchmarks, we raise the following questions: Can these video-text model truly grasp the semantics of natural videos? Are these models genuinely reaching a level of understanding comparable to humans? Despite the remarkable achievements made in previous works, our study suggests that current video-text models still fall far behind human-level understanding.", + "bbox": [ + 212, + 146, + 787, + 238 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We argue that existing prominent results on standard video-text tasks can be misleading as models largely exploit the shortcuts and biases inherent in the dataset. Many of the questions can be answered by objects or context extracted from a single frame without capturing cross-frame relations. For the examples in Fig. 2a, the video-text alignment can be easily inferred from shortcuts such as \"cymbals\" or \"football\". Moreover, the models may utilize biases in the datasets, such as the spurious correlation between \"outdoor\" and \"football\". Current evaluations of video-text understanding are compromised by shortcuts and biases, which obscure us from analyzing the limitations of current models. As we proceed from image-text understanding to video-text understanding, we should focus on more challenging semantics in the video domain that require cross-frame reasoning to solve, such as the interaction between person and object or the change of appearances over a sequence of frames (see Fig. 1a).", + "bbox": [ + 212, + 239, + 789, + 436 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper we propose a novel evaluation paradigm of video-text understanding, i.e., retrieval from counterfactually augmented data (RCAD). As demonstrated in Fig. 2b, the goal is to retrieve the only positive caption with matched semantics among \"hard\" negative captions. Negative captions are counterfactually augmented so that video-text models must derive a holistic understanding of the video sequence for both objects and actions, in order to retrieve the correct caption. As a comparison, negative examples from standard video-to-text retrieval [22, 24, 28] are captions of different videos in the same dataset, often containing different object entities that are easy to distinguish.", + "bbox": [ + 212, + 440, + 789, + 575 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We follow the previous human-in-the-loop system [9] and develop a benchmark dataset for our new RCAD task, i.e., the Feint6K dataset. We test a wide range of public video-text methods with various pretraining strategies on this new benchmark. Two failure examples are visualized in Fig. 1a and quantitative results are summarized in Fig. 1b. Note how state-of-the-art video-text model, InternVideo [23], features a $87.9\\%$ rank-1 accuracy on standard video-to-text retrieval but drops by $32.8\\%$ (pretraining only) and $29.7\\%$ (finetuned) when tested on our benchmark with counterfactually augmented data. By establishing a human-level baseline on our benchmark, the InternVideo model surprisingly falls far behind human performance by $38.6\\%$ . Our findings sharply contrast to the prominent performance obtained on previous video-text benchmarks and question the common belief that latest video-text models can develop a fairly effective representation for texts and videos in existing datasets.", + "bbox": [ + 212, + 579, + 789, + 776 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In light of our evaluation results on Feint6K, we identify a significant limitation of the widely adopted contrastive representation learning on video-text data, which is the issue of shortcut learning (see Sec. 4.1). To address this, we propose LLM-teacher that enables a more effective learning of action se", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "W. Ma et al.", + "bbox": [ + 271, + 114, + 357, + 126 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/3c0eebf4e7165d4fe3da696685d2cb5d5f4da7c6217ea066f34ace79e7dd6847.jpg", + "image_caption": [ + "Groundtruth: \n $\\checkmark$ A person sitting outside with some tools is examining a rock on a leather piece." + ], + "image_footnote": [], + "bbox": [ + 217, + 143, + 336, + 200 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/29b9e994213addfa806d0ae37029d51d199cb91363f3fcf207ffe1da2aa622d9.jpg", + "image_caption": [ + "Predicted: \n $x$ A person sitting outside with some tools is throwing a rock on a leather piece.", + "Groundtruth: A person is building a space jet toy at high speed." + ], + "image_footnote": [], + "bbox": [ + 344, + 143, + 465, + 200 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/1c44aa3ac092405d659aecdd7275730fe461e6676e77870a1a2cb5cc0172905c.jpg", + "image_caption": [ + "Predicted: \n $\\times$ A person is coloring a space jet toy at high speed.", + "(a) Failures of InternVideo [23] on Feint6K", + "(b) Quantitative results on Feint6K dataset", + "Fig. 1: (a): Although with large-scale pretraining on web-scale data, current video-text model like [23] can be easily fooled by counterfactually augmented data. (b): The performance of InternVideo on retrieval from counterfactually augmented data (RCAD) drops by over $30\\%$ when compared to the standard video-to-text retrieval and by $38.6\\%$ when compared to human-level performance. We also evaluate models on standard video-text retrieval from only 6 sampled candidates and show that our RCAD task is indeed much more challenging. Our LLM-teacher successfully improves the performance on RCAD by enforcing a more effective learning of action semantics." + ], + "image_footnote": [], + "bbox": [ + 477, + 148, + 779, + 260 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "mantics by introducing extra knowledge from pretrained large language models (LLMs). Specifically we generate synthetic captions by modifying the contents of an existing caption and LLM serves as the teacher to determine if a synthetic caption matches the semantics in the original caption using binary pseudo-labels or continuous logits. This approach enables a more effective learning of action embeddings from the available video-text data. Experimental results show that LLM-teacher learns more discriminative action embeddings from our advanced contrastive objectives and improves the performance of retrieval from counter-factually augmented data when applied to multiple video-text models.", + "bbox": [ + 212, + 431, + 787, + 568 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In summary, our key contributions are as follows. (1) We propose to evaluate video-text models on questions that require cross-frame reasoning and develop a new task, retrieval from counterfactually augmented data, and a new dataset, Feint6K (Sec. 3). (2) Extensive experimental results on our new evaluation paradigm suggest that existing video-text models demonstrate very limited understanding of the action semantics in a video, which is contrary to the prominent performance achieved on standard video-text retrieval benchmarks. (3) From our results on Feint6K dataset, we identify a key shortcoming of contrastive learning approaches on video-text data. We present LLM-teacher to enforce a more effective learning of action embeddings by injecting knowledge from pretrained LLMs (Sec. 4). Our approach effectively improves the results on Feint6K when applied to multiple text-video models.", + "bbox": [ + 212, + 569, + 787, + 750 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Works", + "text_level": 1, + "bbox": [ + 215, + 775, + 395, + 790 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Video-text pretraining. With the availability of web-scale video-text paired datasets, such as WebVid2M [1] and HowTo100M [17], recent developments of", + "bbox": [ + 212, + 809, + 787, + 839 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "Rethinking Video-Text Understanding: RCAD and Feint6K", + "bbox": [ + 334, + 114, + 730, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/c336959575ac3049fd973da454260476b894ece1d94dd44934d3745a55dbe9bc.jpg", + "image_caption": [ + "(a) Standard video-to-text retrieval", + "Fig. 2: Different evaluations of video-text understanding. (a): In standard video-to-text retrieval, negative captions are sampled from different videos in the same dataset. Image-text models can achieve good performance by exploiting shortcuts (e.g., \"football\" and \"cymbals\") and biases (e.g., spurious correlation between \"outdoor\" and \"football\"). (b): In our proposed RCAD paradigm, we adopt a human-in-the-loop system (see Sec. 3.2) to obtain \"hard\" negative captions with unchanged object entities but modified actions. Models must develop a holistic understanding of the semantics from the sequence of frames to retrieve the matched caption." + ], + "image_footnote": [], + "bbox": [ + 217, + 142, + 460, + 303 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/8e0b68883a179cd5a7ce1e1fc09f0c0d71cb9e161e78ff68219104f946d93be2.jpg", + "image_caption": [ + "(b) RCAD" + ], + "image_footnote": [], + "bbox": [ + 542, + 146, + 790, + 301 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "video-text pretraining models achieved improved results on a wide range of video understanding tasks. Visual and textual embeddings jointly learned from the large-scale data demonstrate strong transferability and largely reduce the efforts in downstream tasks. Mainstream pretraining objectives can be categorized into discriminative and generative. Discriminative approaches extend the objective of CLIP [18] to the video-text domain and learn multi-modal representations by contrasting between matched and unmatched pairs [15, 16, 23, 26] or simply predicting if a video-text pair is matched [16]. Generative methods follow the masked modeling idea in BERT [5] and adopted masked language modeling (MLM) and masked video modeling (MVM) for video-text pretraining [16, 21, 23]. These generative approaches are often considered superior for action recognition, while discriminative approaches learn better semantics from texts [21].", + "bbox": [ + 212, + 468, + 787, + 650 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Evaluation of video-text models. In order to analyze the effectiveness of video-text pretraining, previous works [15,16,21,23,26,29] focused on the following tasks: (1) video-text retrieval (zero-shot or finetuned) [22, 28]: retrieving the matched video (or text) given a text (or video) as query where the negative candidates are unmatched pairs from the same dataset, (2) video question answering (finetuned) [25]: predicting the answer with a classification head, (3) video classification (zero-shot or finetuned) [6,10,20]: classifying videos using label names as text prompts, and (4) video captioning (zero-shot or finetuned) [11, 28]: summarizing the contents of a video.", + "bbox": [ + 212, + 672, + 787, + 808 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We argue that results on these video-text tasks could be deceitful as most tasks heavily rely on the alignment of object entities in the video and text", + "bbox": [ + 212, + 809, + 787, + 839 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "W. Ma et al.", + "bbox": [ + 271, + 114, + 359, + 127 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "(which image-text models are also capable of), and “action understanding” can often be developed through shortcuts and biases. For instance, a commonly-used evaluation for “action recognition” is video classification on ActivityNet [6], where a model classifies a given video into classes such as “playing badminton”, “kayaking”, or “volleyball”. However, video-text models would classify the videos by exploiting objects and contexts as shortcuts (e.g., “badminton” and “kayak”), as well as other biases (e.g., spurious correlation between “kayak” and “on the water”) without genuinely understanding the semantics of the action represented by interactions between the person and the objects over time. For instance, classifying the action as “kayaking” is essentially a “kayak” detection problem in ActivityNet. Therefore we propose retrieval from counterfactually augmented data, a new evaluation paradigm where we aim to eliminate the shortcuts from the questions so models must establish a comprehensive understanding of the semantics from cross-frame reasoning in order to predict the correct answer.", + "bbox": [ + 212, + 146, + 787, + 358 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3 Retrieval from Counterfactually Augmented Data", + "text_level": 1, + "bbox": [ + 215, + 381, + 732, + 398 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Previous evaluation tasks of video-text models focused on the alignment of feature embeddings of the video-text pairs (e.g., video-text retrieval [22, 28] and video classification [10, 20]) or reconstruction of the semantics with text (e.g., video question answering [25] and video captioning [11, 28]). However, these evaluation tasks are largely limited by the paired data available in existing video-text datasets. As demonstrated by in Fig. 2a, we can extract most of the semantics of a video, such as the objects, people, and contexts in the video, by looking at only one frame from the video. The action in the video can be further inferred by exploiting biases in the datasets. Therefore, models pretrained on image-text data often perform surprisingly well on existing video-text tasks.", + "bbox": [ + 212, + 412, + 787, + 564 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To effectively evaluate how video-text models can understand the contents and semantics of videos beyond images, we must develop video understanding tasks that are free from shortcuts and require cross-frame reasoning to solve. To this end, we propose a new evaluation task for video-text models, i.e., retrieval from counterfactually augmented data (Sec. 3.1), and a new dataset Feint6K (Sec. 3.2). In contrast to previous datasets with matched video-text pairs scrapped from the web [24, 28] or annotated by human [22], we adopt a human-in-the-loop system and annotate counterfactually manipulated texts as negative pairs with the original video. This allows us to evaluate video-text models with novel tasks and better understand their limitations. Lastly we measure human performance on our dataset as a direct comparison with the state-of-the-art video-text models (Sec. 3.3).", + "bbox": [ + 212, + 565, + 787, + 746 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1 Task Formulation", + "text_level": 1, + "bbox": [ + 215, + 768, + 406, + 782 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Retrieval from counterfactually augmented data is a variant of the standard video-to-text retrieval. As shown in Fig. 2b, given a video sequence and a list of candidate captions, the model is asked to retrieve the caption that best matches", + "bbox": [ + 212, + 794, + 785, + 839 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "Rethinking Video-Text Understanding: RCAD and Feint6K", + "bbox": [ + 334, + 113, + 730, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Video A", + "text_level": 1, + "bbox": [ + 212, + 154, + 274, + 169 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/f917624d6973ec7537a39238f822e2d9cbcada4184c6e42855970622fc90204a.jpg", + "image_caption": [ + "Fig. 3: Examples of $RCAD$ on our Feint6K dataset." + ], + "image_footnote": [], + "bbox": [ + 290, + 142, + 787, + 183 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Video B", + "text_level": 1, + "bbox": [ + 215, + 200, + 272, + 213 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/b784d2ebffcb678ab1a7ba413eb75cf19f603be00ab0b2f003e25f872a3d3e4d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 290, + 186, + 785, + 227 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Video C", + "text_level": 1, + "bbox": [ + 215, + 239, + 274, + 253 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/adee20fde29321c8721323115aae2802dfc2fe5728d8f8d8839a2efd253409d3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 290, + 232, + 785, + 262 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Captions of Video A", + "text_level": 1, + "bbox": [ + 290, + 267, + 401, + 277 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "[Positive] A person is shinning a shoe with a shoe shinning machine.", + "bbox": [ + 290, + 281, + 444, + 310 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "[Negative 1] A person is making a shoe with a shoe shimming machine.", + "bbox": [ + 290, + 313, + 431, + 342 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "[Negative 2] A person is wearing a shoe with a shoe shinning machine.", + "bbox": [ + 290, + 345, + 434, + 375 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Captions of Video B", + "text_level": 1, + "bbox": [ + 459, + 267, + 568, + 277 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "[Positive] A man is standing on the field attempting to kick a football.", + "bbox": [ + 457, + 281, + 614, + 310 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "[Negative 1] A man is standing on the field attempting to catch a football.", + "bbox": [ + 457, + 313, + 614, + 342 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "[Negative 2] A man is standing on the field attempting to throw a football.", + "bbox": [ + 457, + 345, + 616, + 375 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Captions of Video C", + "text_level": 1, + "bbox": [ + 627, + 267, + 736, + 277 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "[Positive] A person wearing a pink jacket is feeding birds with his hand.", + "bbox": [ + 625, + 281, + 781, + 310 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "[Negative 1] A person wearing a pink jacket is petting birds with his hand.", + "bbox": [ + 625, + 313, + 785, + 342 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "[Negative 2] A person wearing a pink jacket is holding birds with his hand.", + "bbox": [ + 625, + 345, + 785, + 375 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "the semantics in the video. However, unlike standard video-to-text retrieval where negative captions come from other video-text pairs in the same dataset, negative captions in our task are modified from the positive captions, with the same text structure and object entities but different actions (see Fig. 3).", + "bbox": [ + 212, + 429, + 784, + 489 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Similar to video-text retrieval, our proposed task provides the benefit of zero-shot evaluation. This allows us to quantitatively analyze the effectiveness of video-text pretraining without downstream finetuning.", + "bbox": [ + 212, + 489, + 785, + 535 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "What distinguishes our task from all previous video-text tasks is its focus on questions that require cross-frame reasoning. Consider \"Video B\" in Fig. 3 as an example. All captions contain the same object entities present in the video, i.e., \"a man\", \"the field\", and \"a football\", but differ in the action specified. We cannot determine the most accurate caption because all suggested actions, \"kick\", \"catch\", and \"throw\", are plausible given the video's context. To retrieve the corresponding caption, it is crucial to grasp the action's semantics by inspecting the interactions between the man and the football across a series of frames. This underscores the necessity of understanding the video semantics over time in order to succeed in our task, rather than relying shortcuts or biases.", + "bbox": [ + 212, + 536, + 787, + 686 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.2 Feint6K: Data Collection", + "text_level": 1, + "bbox": [ + 214, + 709, + 470, + 723 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We utilize a human-in-the-loop system [9] to counterfactually manipulate the positive captions in existing video-text datasets [22, 28]. The generated counterfactually augmented data, when paired with the original videos, forms negative pairs to be used in our proposed task. We recruit 40 annotators to manually manipulate the actions in existing captions. Specifically, the new actions introduced must be plausible within the context in the caption but are not occurring in the corresponding video. In summary, a total of 6,243 videos from the validation", + "bbox": [ + 212, + 733, + 787, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "W. Ma et al.", + "bbox": [ + 271, + 114, + 359, + 126 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/2e45d571385c4daed4a61b7268a9f80b13fcbbc644dcfb1e49b93b565c82733c.jpg", + "image_caption": [ + "Fig. 4: Overview of our data collection pipeline for Feint6K dataset featuring a human-in-the-loop system." + ], + "image_footnote": [], + "bbox": [ + 243, + 141, + 759, + 266 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "set of the MSR-VTT dataset [28] and the test set of the VATEX dataset [22]. For detailed statistics of our Feint6K dataset, please refer to our supplementary materials.", + "bbox": [ + 212, + 334, + 782, + 378 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The overview of our human-in-the-loop system is demonstrated in Fig. 4. We start by sampling a small subset of video-text pairs and manually annotate the counterfactually augmented captions for the purpose of training and demonstration. Specific guidance and feedback are given to the annotators on the practice questions during the training stage. Moreover, we adopt a reviewing process during the annotation stage to further ensure the quality of our annotated data. Each annotation is reviewed before acceptance, where rejected annotations are marked and sent back for refinement. We refer the readers to the supplementary materials regarding the annotation interface and annotator guidance.", + "bbox": [ + 212, + 380, + 784, + 516 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.3 Human Performance", + "text_level": 1, + "bbox": [ + 215, + 541, + 431, + 555 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We employ the same group of annotators to establish a human-level baseline on our Feint6K dataset. Analyzing the human performance serves two goals: (i) verifying the legitimacy of our Feint6K dataset – if each question is answerable and has one unique answer, and (ii) comparing the state-of-the-art video-text models with human-level performance.", + "bbox": [ + 212, + 568, + 784, + 643 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Each annotator is first presented with the full video and then asked to select the one out of six captions that best matches the semantics in the video. To avoid leakage, we make sure the annotators assigned to the testing questions are not the same annotators annotating this video. We also randomly shuffle the order of the candidate captions to remove any biases. In Sec. 5.2 we present the quantitative comparisons between human-level performance and current state-of-the-art video-text models on our Feint6K dataset.", + "bbox": [ + 212, + 643, + 785, + 750 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4 LLM-teacher", + "text_level": 1, + "bbox": [ + 215, + 775, + 377, + 790 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Contrastive learning are widely adopted in previous video-text representation learning works [16, 23, 26]. The contrastive models learn to distinguish between", + "bbox": [ + 212, + 809, + 782, + 839 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "Rethinking Video-Text Understanding: RCAD and Feint6K", + "bbox": [ + 334, + 113, + 732, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "similar and dissimilar pairs of video-text data by contrasting their feature representations. Despite the simple formulation and great generalization ability, our results on Feint6K show that current video-text models built heavily on contrastive objectives have very limited understanding of action semantics in natural videos and can be easily fooled by counterfactually augmented data. Therefore we aim to develop a more effective contrastive approach and learn more discriminative action representations from existing video-text data.", + "bbox": [ + 212, + 146, + 782, + 251 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Sec. 4.1 discusses the limitations of current contrastive approaches due to numerous shortcuts in video-text dataset. Inspired by this finding, we present LLM-teacher in Sec. 4.2, a simple but effective approach to learn more discriminative action embeddings by incorporating knowledge from pretrained large language models.", + "bbox": [ + 212, + 252, + 784, + 328 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.1 Shortcuts in Video-Text Data", + "text_level": 1, + "bbox": [ + 215, + 352, + 504, + 366 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The CLIP model [18] trained on image-text data with a contrastive objective demonstrated strong zero-shot capabilities in many vision tasks [13,30] and was widely adopted as the vision encoder in many large vision-language models [14]. However, directly generalizing this idea to video-text data may face unprecedented challenges due to shortcuts in the video-text data.", + "bbox": [ + 212, + 378, + 782, + 454 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "As we proceed from image-text to video-text representation learning, we aim to learn a powerful embedding for not only the object entities, but also the actions specified by texts or sequence of frames. Naturally we hope to achieve this by contrasting videos with different actions. However, as we are optimizing the contrastive objectives in a mini-batch of videos, object entities become the shortcuts that contrastive models exploit to saturate the contrastive objective, leading to an ineffective action embeddings. Consider the example in Fig. 2a, the contrastive loss would be small as long as the model learns discriminative embeddings for shortcut objects such as \"cymbal\" and \"football\", even without any understanding of the actions. Since many video-text models start with pretrained CLIP encoders [23,26], they often possess strong embeddings for shortcut objects from the beginning of the video-text pretraining, which hinders the model from further learning effective action representations.", + "bbox": [ + 212, + 455, + 782, + 651 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In Fig. 5 we present quantitative results of changes in cosine similarities when the video is unchanged but the object or action in the caption is manipulated. We find that state-of-the-art video-text models learn less discriminative embeddings for actions as compared to objects, which also explains the small gap between image-text and video-text models on standard video-text tasks.", + "bbox": [ + 212, + 652, + 782, + 727 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.2 LLM-teacher", + "text_level": 1, + "bbox": [ + 215, + 752, + 370, + 766 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We introduce LLM-teacher, an LLM-powered approach to learn better action representations with contrastive objectives. Specifically, LLM serves as the \"teacher\" and provides synthesized captions as extra knowledge for the video-text model to learn from. These new captions serve as negative captions of the original video,", + "bbox": [ + 212, + 779, + 802, + 839 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "W. Ma et al.", + "bbox": [ + 271, + 114, + 359, + 127 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/49fff5b43c9cdadc257ceae8112a25ec80fe30eb7d114c6ddecce79a0869f22a.jpg", + "image_caption": [ + "(a) Comparison between $\\Delta s$ w.r.t. object or action changes." + ], + "image_footnote": [], + "bbox": [ + 223, + 147, + 493, + 263 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/c89c032b08ca7476da73e19c348b841d3f223e885edbc7f98ffb0aa738177e97.jpg", + "image_caption": [ + "(b) Comparison between $\\Delta s$ w.r.t. action using InternVideo and our LLM-teacher.", + "Fig. 5: Change of cosine similarity w.r.t. objects or actions. (a): Comparison between the change in cosine similarity when the action or object is swapped. Results show that current video-text models learn a more effective embedding for objects than for actions. (b): Comparison between the change in cosine similarity using InternVideo or our LLM-teacher. This demonstrate that LLM-teacher learns a more discriminative embedding for actions by enabling a more effective contrastive learning using knowledge from LLMs. Refer to Sec. 5.2 for more details." + ], + "image_footnote": [], + "bbox": [ + 511, + 148, + 781, + 263 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "removing possible shortcuts and biases and enabling a more effective learning of action semantics.", + "bbox": [ + 212, + 435, + 784, + 464 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We start with positive captions from common video-text datasets and run an abstract meaning representation (AMR) parser, resulting in a list of action or object tokens. To generate the hard negative captions, two methods are considered to find novel actions or objects given the original caption and the substitution token $w$ .", + "bbox": [ + 212, + 468, + 784, + 542 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Method I: Mask filling. Masked language modeling (MLM) [5] is a widely adopted self-supervised task for LLMs. In the pretraining stage, a certain percentage of word tokens are substituted with a special token [mask] and the LLM is optimized to recover the masked tokens. To find negative captions with novel actions given the original caption $c_{0}$ and substitution token $a$ , we first substitute the token $a$ with [mask] and use a LLM pretrained with MLM to predict a list of $k$ possible tokens $\\{\\hat{w}_i\\}_{i=1}^k$ such that $\\hat{w}_i \\neq w$ and $\\hat{w}_i$ is an action token.", + "bbox": [ + 212, + 578, + 784, + 684 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Method II: LLM-powered chatbot. Although mask filling can find appropriate words to substitute $w$ , it is limited by one single token. In many cases, we want to update the prepositions following the change of verbs. For example, when \"install\" is substituted by \"uninstall\", we should also update the preposition from \"install ... to\" to \"uninstall ... from\". Here we consider a more flexible approach that builds on LLM-powered chatbots. Besides a prompt describing the text substitution, we leverage the in-context learning ability of LLMs [2, 7] and provide multiple in-context examples to obtain the desirable negative captions.", + "bbox": [ + 212, + 718, + 784, + 839 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "Rethinking Video-Text Understanding: RCAD and Feint6K", + "bbox": [ + 334, + 113, + 732, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 114, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "LLM-teacher. With the procedure above, we obtain $k$ synthesized negative captions $\\{n_i\\}_{i=1}^k$ besides the original positive caption $p$ for each video $v$ in existing datasets [22,28]. For each of the original and generated captions, LLM serves as a \"teacher\" and provide binary pseudo-labels - if the caption have matched (\"1\") or unmatched (\"0\") semantics with the video. Now we may optimize the video-text model with the normalized temperature-scaled cross entropy loss", + "bbox": [ + 215, + 146, + 782, + 237 + ], + "page_idx": 9 + }, + { + "type": "equation", + "text": "\n$$\nl = - \\log \\frac {\\exp \\left(\\sin \\left(f _ {v} , f _ {p}\\right) / \\tau\\right)}{\\exp \\left(\\sin \\left(f _ {v} , f _ {p}\\right) / \\tau\\right) + \\sum_ {i = 1} ^ {k} \\exp \\left(\\sin \\left(f _ {v} , f _ {n _ {i}}\\right) / \\tau\\right)} \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 295, + 246, + 785, + 284 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "where $f_{v}, f_{p}, f_{ni}$ are the visual and textual embeddings.", + "bbox": [ + 215, + 292, + 616, + 308 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "In practice, we find that certain negative captions have unmatched but similar semantics to the video. It is undesirable to regard them as a strictly negative pair in the contrastive loss. Therefore, we extend the binary pseudo-labels to soft logits by computing caption similarities with a pretrained LLM and then optimize the video-text model to match model outputs with the soft logits from the LLM teacher model [8].", + "bbox": [ + 212, + 309, + 782, + 398 + ], + "page_idx": 9 + }, + { + "type": "equation", + "text": "\n$$\nl = \\mathcal {L} _ {\\mathrm {K L}} \\left(z _ {\\text {v i d e o - t e x t}}, z _ {\\text {L L M}}\\right) \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 352, + 412, + 784, + 428 + ], + "page_idx": 9 + }, + { + "type": "equation", + "text": "\n$$\nz _ {\\mathrm {v i d e o - t e x t , t}} = \\frac {\\exp (\\sin (f _ {v} , f _ {t}) / \\tau)}{\\exp (\\sin (f _ {v} , f _ {p}) / \\tau) + \\sum_ {i = 1} ^ {k} \\exp (\\sin (f _ {v} , f _ {n _ {i}}) / \\tau)}\n$$\n", + "text_format": "latex", + "bbox": [ + 284, + 431, + 715, + 467 + ], + "page_idx": 9 + }, + { + "type": "equation", + "text": "\n$$\nz _ {\\mathrm {L L M , t}} = \\frac {\\exp (\\sin (e _ {p} , e _ {t}) / \\tau)}{\\exp (\\sin (e _ {p} , e _ {p}) / \\tau) + \\sum_ {i = 1} ^ {k} \\exp (\\sin (e _ {p} , e _ {n _ {i}}) / \\tau)}\n$$\n", + "text_format": "latex", + "bbox": [ + 313, + 469, + 714, + 506 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "where $t \\in \\{n_i\\}_{i=1}^k \\cup \\{p\\}$ , $f$ 's are textual and visual embeddings of the video-text model, and $e$ 's are textual embeddings of a pretrained LLM.", + "bbox": [ + 215, + 516, + 782, + 547 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "5 Experiments", + "text_level": 1, + "bbox": [ + 215, + 570, + 374, + 587 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "5.1 Experimental Setup", + "text_level": 1, + "bbox": [ + 215, + 602, + 426, + 618 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Datasets. We use MSR-VTT dataset [28] and VATEX dataset [22] for standard video-to-text retrieval and our Feint6K dataset for retrieval from counter-factually augmented data. Note that Feint6K dataset contains the same video sequences for evaluation as in MSR-VTT dataset and VATEX dataset.", + "bbox": [ + 215, + 627, + 782, + 688 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Evaluation metrics. For standard video-to-text retrieval we report rank-1 accuracy $(\\mathbb{R}@\\mathbb{1})$ as this is the most challenging metric and distinguish the performance of previous video-text models. For retrieval from counterfactually-augmented data we use rank-1 accuracy $(\\mathbb{R}@\\mathbb{1})$ , rank-2 accuracy $(\\mathbb{R}@\\mathbb{2})$ , and mean rank (MeanR).", + "bbox": [ + 215, + 710, + 782, + 786 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Video-text models. We consider multiple state-of-the-art public models in this work. CLIP4Clip [15] extends CLIP model [18] to the video domain using a", + "bbox": [ + 215, + 809, + 782, + 839 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "W. Ma et al.", + "bbox": [ + 271, + 114, + 357, + 126 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/5f876245145b0bb24a0ae0bbaf7421a7cceed97753ecae41e714329559088ab0.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelMSR-VTTFeint6K (MSR-VTT)VATEXFeint6K (VATEX)
R@1R@1R@2MeanRR@1R@1R@2MeanR
Human95.296.8
Random<1e-316.7<1e-316.7
Zero-Shot
CLIP [18]26.337.355.32.638.834.854.32.7
VideoCLIP [27]14.535.171.32.613.733.070.72.7
InternVideo [23]37.545.863.62.376.944.163.92.3
LanguageBind [31]42.841.377.02.442.377.02.4
Finetuned
LanguageBind [31]
CLIP4Clip [15]43.150.872.42.0
VindLU [3]46.653.470.92.0
SimVTP [16]50.235.770.82.676.633.668.42.6
w/ LLM-teacher-lbl49.0 ↓1.240.0 ↑4.373.1 ↑2.32.3 ↓0.374.8 ↓1.837.3 ↑3.772.2 ↑3.82.4 ↓0.2
w/ LLM-teacher-lgt49.5 ↓0.743.5 ↑7.875.0 ↑4.22.2 ↓0.475.3 ↓1.340.1 ↑6.573.5 ↑5.12.3 ↓0.3
InternVideo [23]49.158.680.21.887.958.276.91.9
w/ LLM-teacher-lbl48.2 ↓0.964.2 ↑5.682.5 ↑2.31.7 ↓0.185.2 ↓2.763.8 ↑5.680.5 ↑3.61.7 ↓0.2
w/ LLM-teacher-lgt48.9 ↓0.265.8 ↑7.283.8 ↑3.61.7 ↓0.187.3 ↓0.665.6 ↑7.481.7 ↑4.81.7 ↓0.2
", + "bbox": [ + 215, + 143, + 785, + 393 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Table 1: Performance of standard video-to-text retrieval on MSR-VTT dataset [28] and VATEX dataset [22], as compared to performance of retrieval from counterfactually-augmented data on our Feint6K dataset. In the \"zero-shot\" setting, models are evaluated directly after the pretraining stage, while in the \"finetuned\" setting, models are finetuned on the training set of MSR-VTT or VATEX. With LLM-teacher, we enable a more efficient learning of action semantics and effectively improves R@1 accuracy on Feint6K. Here \"LLM-teacher-lbl\" stands for our approach with binary pseudo-labels and \"LLM-teacher-lgt\" uses soft logits.", + "bbox": [ + 212, + 405, + 787, + 517 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "mean pooling mechanism for zero-shot video-text retrieval and a contrastive loss for finetuning. SimVTP [16] is pretrained on the WebVid2M dataset [1] with a combination of contrastive learning and masked modeling. InternVideo [23] is a video-text foundation model pretrained on a combination of 7 large-scale videotext datasets, with a supervised video post-pretraining for better video recognition. LanguageBind [31] is a language-based multi-modal pretraining model with remarkable performance on a wide range of benchmarks.", + "bbox": [ + 212, + 547, + 787, + 652 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "To demonstrate the efficacy of our LLM-teacher, we adopt this method on two state-of-the-art video-text models, SimVTP and InternVideo. In the default setting we generate 10 action-based captions with LLM for each training video. For text generation we use a pretrained XLM-RoBERTa [4] for mask filling. To measure text similarities and compute soft logits, we utilize a pretrained Sentence-BERT model [19]. In Sec. 5.3 we ablate on the choices of parameters and caption generation.", + "bbox": [ + 212, + 654, + 787, + 758 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "5.2 Main Results", + "text_level": 1, + "bbox": [ + 215, + 782, + 372, + 796 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Performance of previous state-of-the-art. In Tab. 1 we report the quantitative results of standard video-text retrieval on MSR-VTT and VATEX, and of", + "bbox": [ + 212, + 809, + 787, + 839 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "Rethinking Video-Text Understanding: RCAD and Feint6K", + "bbox": [ + 334, + 114, + 730, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 116, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "retrieval from counterfactually augmented data on our Feint6K. We also report a \"Random\" performance where a model predicts random guesses. We make the following observations: (i) Results show that previous state-of-the-art video-text models demonstrate limited understanding of the action semantics in a video with less than $60\\%$ R@1 accuracy on RCAD, given a $16.7\\%$ R@1 accuracy when taking random guesses. (ii) Although it is hard to compare the quantitative results between two tasks, we note that our RCAD is much more challenging given the performance gap between \"Random\" and previous state-of-the-arts. This supports our previous arguments that our RCAD removes shortcuts from standard video-text datasets and focuses on harder questions in the video domain that require cross-frame reasoning.", + "bbox": [ + 212, + 146, + 787, + 313 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Performance of LLM-teacher. We apply our LLM-teacher to two pretrained video-text models, SimVTP [16] and InternVideo [23]. Specifically, we consider the two objectives in Eq. (1) and Eq. (2), labeled as \"LLM-teacher-lbl\" and \"LLM-teacher-lgt\" respectively. Results in Tab. 1 show that by exploiting knowledge from pretrained LLMs, LLM-teacher enables a more efficient leaning of action semantics and achieved improved results on various metrics of RCAD, increasing the R@1 accuracy on RCAD by $7.2\\%$ and $7.4\\%$ . We also note that with binary pseudo-labels, results demonstrate a trade-off between the performance on standard video-to-text retrieval and RCAD; and in comparison, LLM-teacher with soft logits achieves the highest performance on RCAD ( $65.8\\%$ and $65.6\\%$ ) with negligible drops ( $0.2\\%$ and $0.6\\%$ ) on standard retrieval. This is because soft logits account for the text similarities between different actions and avoid overfitting on binary pseudo-labels.", + "bbox": [ + 212, + 335, + 789, + 534 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Comparison with human-level performance. The human-level performance are reported in Tab. 1. We see that human annotators achieved an almost perfect performance on Feint6K dataset, with a $95.2\\%$ accuracy on MSR-VTT videos and a $96.8\\%$ accuracy on VATEX videos. These results show that: (i) most questions for retrieval from counterfactually-augmented data are answerable with an unique answer, and (ii) this task is fairly simple for human. As a comparison, all previous video-text models with large-scale pretraining fall behind by a wide margin. Specifically, the state-of-the-art video-text model InternVideo [23] features a heavy two-stage pretraining on a collection of 7 large-scale video-text datasets but falls behind by $36.6\\%$ and $38.6\\%$ even with downstream finetuning.", + "bbox": [ + 212, + 558, + 789, + 709 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Analyses of failure cases. In Fig. 1b we present two failure cases by the InternVideo [23] model on our Feint6K dataset. Both examples are trivial for human, yet challenging from video-text models due to incapability of cross-frame reasoning. For the first example, a model must analyze the interactions between the person and the rock over a sequence of frames to predict the correct answer. And in the second example, the model must observe the change of appearances over time. As we can see, with the counterfactually augmented data in our Feint6K,", + "bbox": [ + 212, + 734, + 787, + 839 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "W. Ma et al.", + "bbox": [ + 271, + 114, + 357, + 126 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "models cannot \"guess\" the correct answer by exploiting shortcuts extracted from a single frame. The findings on Feint6K highlight significant weaknesses in current video-text models, offering valuable insights for future researches in this area. Additional qualitative examples of video-text models on Feint6K dataset are provided in the supplementary materials for readers' reference.", + "bbox": [ + 212, + 146, + 787, + 223 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Sensitivity to the change of object or action. With the counterfactually augmented data collected, we investigate how the cosine similarities between the video and text change when the object or action in the text are swapped with an object or action not present in the image. Given the original video-text pair $(v,t)$ and a counterfactually-augmented text $\\hat{t}$ , the change in cosine similarity is given by $\\Delta s = s(v,\\hat{t}) - s(v,t)$ . As $\\hat{t}$ contains objects or actions not present in the video, $\\Delta s$ should be negative for an ideal video-text model. Moreover, larger $|\\Delta s|$ implies the model being more sensitive to the changes. We compute $\\Delta s$ using the InternVideo with pretraining only (\"IntVid-pt\"), InternVideo with downstream finetuning (\"IntVid-ft\"), and our LLM-teacher (\"LLM-teacher\").", + "bbox": [ + 212, + 243, + 787, + 393 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Results in Fig. 5a show that the $\\Delta s$ are almost always negative when the objects are swapped and the absolute changes are larger. In comparison, $\\Delta s$ are sometimes positive when actions are modified and the absolute changes are smaller. This demonstrate that InternVideo learns a more effective embedding for objects than for actions. This further highlights the necessity of our new evaluation paradigm for exploring the limitations of video-text models beyond existing tasks. Moreover, in Fig. 5b we note that with LLM-teacher the action embeddings are more effective than the ones learned by InternVideo, as shown by the more reasonable changes in cosine similarity when actions are swapped.", + "bbox": [ + 212, + 395, + 787, + 530 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Please refer to the supplementary materials for a detailed discussion regarding the influence of textual encoders.", + "bbox": [ + 212, + 530, + 785, + 560 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "5.3 Ablation Study", + "text_level": 1, + "bbox": [ + 214, + 582, + 388, + 597 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "We conduct ablation studies on the VATEX dataset [22] and our Feint6K dataset. We follow the same settings as above and report rank-1 (R@1) accuracy for standard video-to-text retrieval and rank-1 accuracy (R@1), rank-2 accuracy (R@2), and mean rank (MeanR) for counterfactual augmented data.", + "bbox": [ + 212, + 606, + 787, + 667 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Choice of captions in LLM-teacher. Tab. 2 shows the comparison between different numbers and types of LLM captions generated. In the \"default\" setting we use 10 action-based LLM captions and compare to settings with 5 action-based LLM captions or 5 action-based and 5 object-based LLM captions. Results show with 5 action-based LLM captions the R@1 accuracy on Feint6K drops by $0.9\\%$ while the R@1 accuracy on VATEX increases by $0.3\\%$ . We also experiment on object-based LLM captions and find them not beneficial for both standard video-to-text retrieval or our RCAD. This is consistent with our assumption that pretrained video-text models already learns a discriminative embedding for objects and a more efficient training objective for action is needed.", + "bbox": [ + 212, + 688, + 787, + 840 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "Rethinking Video-Text Understanding: RCAD and Feint6K", + "bbox": [ + 334, + 113, + 730, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/00452429d88f66ca60d637193a99280c5245a7fe844cae2ccf3e65b266f9e729.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelVATEXFeint6K (VATEX)
R@1R@1R@2MeanR
Default(10 action captions; XLM-RoBERTa)87.365.681.71.7
5 action captions87.6 ↑0.364.7 ↓0.981.0 ↓0.71.70.0
5 object + 5 object captions87.5 ↑0.264.2 ↓1.480.6 ↓1.11.70.0
LLM Chatbot87.0 ↓0.365.9 ↑0.381.8 ↑0.11.70.0
", + "bbox": [ + 287, + 143, + 718, + 252 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Table 2: Ablation studies on the number of LLM captions and caption generation methods. In the default setting we use a total of 10 action-based LLM captions and utilize a pretrained XLM-RoBERTa [4] for mask filling.", + "bbox": [ + 215, + 261, + 787, + 305 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Choice of caption generation. Besides a pretrained XLM-RoBERTa model for mask filling, we also experiment on generating captions with LLM-powered agents that are finetuned on chat datasets for dialogue applications. Specifically we use the \"chat\" model finetuned for dialogue applications. Empirically we find that captions generated by LLM chatbots achieve a higher overall quality, exploring a more diverse caption space and better flexibility. However we don't observe significant improvements from our ablation study experiments – the R@1 accuracy on VATEX drops by $0.3\\%$ and the R@1 accuracy on Feint6K increases by $0.3\\%$ . We choose XLM-RoBERTa in our main experiments as the model runs faster and scale up easily to bigger settings. We provide qualitative comparisons between the two types of caption generation in our supplementary materials.", + "bbox": [ + 212, + 330, + 787, + 500 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "6 Conclusions", + "text_level": 1, + "bbox": [ + 215, + 518, + 370, + 535 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In this work we propose a new evaluation task, retrieval from counterfactually augmented data, and a benchmark dataset, Feint6K. The idea is to remove shortcuts in the video-text questions with a human-in-the-loop system and produce more challenging questions where the video-text model must derive a comprehensive understanding of the video with cross-frame reasoning. Quantitative and qualitative evaluation results on our Feint6K dataset show that despite the task is trivial for human, previous state-of-the-art video-text models can be easily fooled by the counterfactually augmented data. This implies that the prominent results on previous video-text benchmarks could be misleading - models may largely exploit shortcuts without genuinely understand the video contents. Moreover, we identify a key limitation of current contrastive learning on videotext data being the shortcut learning and propose LLM-teacher that enables a more effective learning of action semantics by utilizing knowledge from pretrained LLMs. Experimental results and analyses show that our method can learn a more discriminative representation for actions.", + "bbox": [ + 212, + 547, + 789, + 776 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Supplementary materials. We present the following: (1) limitations of our work, (2) more details about our Feint6K dataset, (3) ethics statements, and (4) extra quantitative and qualitative results.", + "bbox": [ + 215, + 794, + 787, + 840 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "W. Ma et al.", + "bbox": [ + 271, + 114, + 359, + 127 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgements. We would like to thank Daniel Khashabi, Yiyan Li, and the anonymous reviewers for their helpful comments and suggestions. Wufei Ma is supported by ONR with N00014-23-1-2641.", + "bbox": [ + 215, + 146, + 787, + 191 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 214, + 321, + 229 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Bain, M., Nagrani, A., Varol, G., Zisserman, A.: Frozen in time: A joint video and image encoder for end-to-end retrieval. In: IEEE International Conference on Computer Vision (2021) 3, 11", + "2. Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J.D., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., et al.: Language models are few-shot learners. Advances in neural information processing systems 33, 1877-1901 (2020) 9", + "3. Cheng, F., Wang, X., Lei, J., Crandall, D., Bansal, M., Bertasius, G.: Vindlu: A recipe for effective video-and-language pretraining. arXiv preprint arXiv:2212.05051 (2022) 11", + "4. Conneau, A., Khandelwal, K., Goyal, N., Chaudhary, V., Wenzek, G., Guzmán, F., Grave, E., Ott, M., Zettlemoyer, L., Stoyanov, V.: Unsupervised cross-lingual representation learning at scale. arXiv preprint arXiv:1911.02116 (2019) 11, 14", + "5. Devlin, J., Chang, M.W., Lee, K., Toutanova, K.: Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805 (2018) 4, 9", + "6. Fabian Caba Heilbron, Victor Escorcia, B.G., Niebles, J.C.: Activitynet: A large-scale video benchmark for human activity understanding. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 961-970 (2015) 4, 5", + "7. Gupta, T., Kembhavi, A.: Visual programming: Compositional visual reasoning without training. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14953-14962 (2023) 9", + "8. Hinton, G., Vinyals, O., Dean, J.: Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531 (2015) 10", + "9. Kaushik, D., Hovy, E., Lipton, Z.C.: Learning the difference that makes a difference with counterfactually-augmented data. arXiv preprint arXiv:1909.12434 (2019) 2, 6", + "0. Kay, W., Carreira, J., Simonyan, K., Zhang, B., Hillier, C., Vijayanarasimhan, S., Viola, F., Green, T., Back, T., Natsev, P., et al.: The kinetics human action video dataset. arXiv preprint arXiv:1705.06950 (2017) 4, 5", + "1. Krishna, R., Hata, K., Ren, F., Fei-Fei, L., Carlos Niebles, J.: Dense-captioning events in videos. In: Proceedings of the IEEE international conference on computer vision. pp. 706-715 (2017) 4, 5", + "2. Li, K., Wang, Y., Li, Y., Wang, Y., He, Y., Wang, L., Qiao, Y.: Unmasked teacher: Towards training-efficient video foundation models. arXiv preprint arXiv:2303.16058 (2023) 1", + "3. Lin, J., Gong, S.: Gridclip: One-stage object detection by grid-level clip representation learning. arXiv preprint arXiv:2303.09252 (2023) 8", + "4. Liu, H., Li, C., Wu, Q., Lee, Y.J.: Visual instruction tuning. In: NeurIPS (2023) 8", + "5. Luo, H., Ji, L., Zhong, M., Chen, Y., Lei, W., Duan, N., Li, T.: Clip4clip: An empirical study of clip for end to end video clip retrieval and captioning. Neurocomputing 508, 293-304 (2022) 4, 10, 11" + ], + "bbox": [ + 225, + 244, + 784, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "Rethinking Video-Text Understanding: RCAD and Feint6K", + "bbox": [ + 334, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "16. Ma, Y., Yang, T., Shan, Y., Li, X.: Simvtp: Simple video text pre-training with masked autoencoders. arXiv preprint arXiv:2212.03490 (2022) 1, 4, 7, 11, 12", + "17. Miech, A., Zhukov, D., Alayrac, J.B., Tapaswi, M., Laptev, I., Sivic, J.: HowTo100M: Learning a Text-Video Embedding by Watching Hundred Million Narrated Video Clips. In: ICCV (2019) 3", + "18. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021) 4, 8, 10, 11", + "19. Reimers, N., Gurevych, I.: Sentence-bert: Sentence embeddings using siamese bert-networks. In: Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing. Association for Computational Linguistics (11 2019), https://arxiv.org/abs/1908.10084 11", + "20. Soomro, K., Zamir, A.R., Shah, M.: Ucf101: A dataset of 101 human actions classes from videos in the wild. arXiv preprint arXiv:1212.0402 (2012) 4, 5", + "21. Tong, Z., Song, Y., Wang, J., Wang, L.: Videomae: Masked autoencoders are data-efficient learners for self-supervised video pre-training. Advances in neural information processing systems 35, 10078-10093 (2022) 4", + "22. Wang, X., Wu, J., Chen, J., Li, L., Wang, Y.F., Wang, W.Y.: Latex: A large-scale, high-quality multilingual dataset for video-and-language research. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 4581-4591 (2019) 1, 2, 4, 5, 6, 7, 10, 11, 13", + "23. Wang, Y., Li, K., Li, Y., He, Y., Huang, B., Zhao, Z., Zhang, H., Xu, J., Liu, Y., Wang, Z., et al.: Internvideo: General video foundation models via generative and discriminative learning. arXiv preprint arXiv:2212.03191 (2022) 1, 2, 3, 4, 7, 8, 11, 12", + "24. Wu, Z., Yao, T., Fu, Y., Jiang, Y.G.: Deep learning for video classification and captioning. In: Frontiers of multimedia research, pp. 3-29 (2017) 2, 5", + "25. Xu, D., Zhao, Z., Xiao, J., Wu, F., Zhang, H., He, X., Zhuang, Y.: Video question answering via gradually refined attention over appearance and motion. In: Proceedings of the 25th ACM international conference on Multimedia. pp. 1645-1653 (2017) 1, 4, 5", + "26. Xu, H., Ghosh, G., Huang, P.Y., Okhonko, D., Aghajanyan, A., Metze, F., Zettlemoyer, L., Feichtenhofer, C.: Videoclip: Contrastive pre-training for zero-shot video-text understanding. arXiv preprint arXiv:2109.14084 (2021) 4, 7, 8", + "27. Xu, H., Ghosh, G., Huang, P.Y., Okhonko, D., Aghajanyan, A., Metze, F., Zettlemoyer, L., Feichtenhofer, C.: VideoCLIP: Contrastive pre-training for zero-shot video-text understanding. In: Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing (EMNLP). Association for Computational Linguistics, Online (Nov 2021) 11", + "28. Xu, J., Mei, T., Yao, T., Rui, Y.: Msr-vtt: A large video description dataset for bridging video and language. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 5288-5296 (2016) 1, 2, 4, 5, 6, 7, 10, 11", + "29. Yan, S., Zhu, T., Wang, Z., Cao, Y., Zhang, M., Ghosh, S., Wu, Y., Yu, J.: Videotext modeling with zero-shot transfer from contrastive captioners. arXiv preprint arXiv:2212.04979 (2022) 4", + "30. Zhou, Z., Lei, Y., Zhang, B., Liu, L., Liu, Y.: Zegclip: Towards adapting clip for zero-shot semantic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11175-11185 (2023) 8" + ], + "bbox": [ + 215, + 147, + 784, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "W. Ma et al.", + "bbox": [ + 271, + 114, + 357, + 126 + ], + "page_idx": 15 + }, + { + "type": "ref_text", + "text": "31. Zhu, B., Lin, B., Ning, M., Yan, Y., Cui, J., HongFa, W., Pang, Y., Jiang, W., Zhang, J., Li, Z., Zhang, C.W., Li, Z., Liu, W., Yuan, L.: Languagebind: Extending video-language pretraining to n-modality by language-based semantic alignment. In: The Twelfth International Conference on Learning Representations (2024), https://openreview.net/forum?id=QmZKc7UZCy 11", + "bbox": [ + 215, + 146, + 787, + 218 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "Rethinking Video-Text Understanding: RCAD and Feint6K", + "bbox": [ + 334, + 113, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + } +] \ No newline at end of file diff --git a/2024/Rethinking Video-Text Understanding_ Retrieval from Counterfactually Augmented Data/21cc04b7-ea52-4351-82c2-826a88e983b2_model.json b/2024/Rethinking Video-Text Understanding_ Retrieval from Counterfactually Augmented Data/21cc04b7-ea52-4351-82c2-826a88e983b2_model.json new file mode 100644 index 0000000000000000000000000000000000000000..4cf7020af2f6a593c0ca1d0913f70831bdbb9ed6 --- /dev/null +++ b/2024/Rethinking Video-Text Understanding_ Retrieval from Counterfactually Augmented Data/21cc04b7-ea52-4351-82c2-826a88e983b2_model.json @@ -0,0 +1,2192 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.218, + 0.141, + 0.788, + 0.187 + ], + "angle": 0, + "content": "Rethinking Video-Text Understanding: Retrieval from Counterfactually Augmented Data" + }, + { + "type": "text", + "bbox": [ + 0.243, + 0.212, + 0.761, + 0.245 + ], + "angle": 0, + "content": "Wufei Ma\\(^{1,2\\star}\\), Kai Li\\(^{1}\\), Zhongshi Jiang\\(^{1}\\), Moustafa Meshry\\(^{1}\\), Qihao Liu\\(^{2}\\), Huiyu Wang\\(^{3}\\), Christian Hane\\(^{1}\\), and Alan Yuille\\(^{2}\\)" + }, + { + "type": "text", + "bbox": [ + 0.432, + 0.255, + 0.573, + 0.269 + ], + "angle": 0, + "content": "Meta Reality Labs" + }, + { + "type": "text", + "bbox": [ + 0.407, + 0.27, + 0.596, + 0.284 + ], + "angle": 0, + "content": "2 Johns Hopkins University" + }, + { + "type": "text", + "bbox": [ + 0.465, + 0.284, + 0.539, + 0.296 + ], + "angle": 0, + "content": "3 Meta AI" + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.34, + 0.756, + 0.646 + ], + "angle": 0, + "content": "Abstract. Recent video-text foundation models have demonstrated strong performance on a wide variety of downstream video understanding tasks. Can these video-text models genuinely understand the contents of natural videos? Standard video-text evaluations could be misleading as many questions can be inferred merely from the objects and contexts in a single frame or biases inherent in the datasets. In this paper, we aim to better assess the capabilities of current video-text models and understand their limitations. We propose a novel evaluation task for video-text understanding, namely retrieval from counterfactually augmented data (RCAD), and a new Feint6K dataset. To succeed on our new evaluation task, models must derive a comprehensive understanding of the video from cross-frame reasoning. Analyses show that previous video-text foundation models can be easily fooled by counterfactually augmented data and are far behind human-level performance. In order to narrow the gap between video-text models and human performance on RCAD, we identify a key limitation of current contrastive approaches on video-text data and introduce LLM-teacher, a more effective approach to learn action semantics by leveraging knowledge obtained from a pretrained large language model. Experiments and analyses show that our approach successfully learn more discriminative action embeddings and improves results on Feint6K when applied to multiple video-text models. Our Feint6K dataset and project page is available here." + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.66, + 0.605, + 0.675 + ], + "angle": 0, + "content": "Keywords: Video-Text Understanding \\(\\cdot\\) Retrieval" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.703, + 0.377, + 0.719 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.738, + 0.788, + 0.815 + ], + "angle": 0, + "content": "Video-text foundation models have gained increasing attention due to their simple formulation and strong transferability [12, 16, 23]. By pretraining on web scale video-text datasets, these models demonstrate strong performance across a wide range of downstream tasks, such as video-text retrieval [22, 28] and video question answering [25]." + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.825, + 0.637, + 0.84 + ], + "angle": 0, + "content": "* All data collection and experiments were conducted at JHU." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.359, + 0.127 + ], + "angle": 0, + "content": "W. Ma et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.239 + ], + "angle": 0, + "content": "As video-text foundation models evolve and achieve increasingly better performance on various benchmarks, we raise the following questions: Can these video-text model truly grasp the semantics of natural videos? Are these models genuinely reaching a level of understanding comparable to humans? Despite the remarkable achievements made in previous works, our study suggests that current video-text models still fall far behind human-level understanding." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.241, + 0.79, + 0.438 + ], + "angle": 0, + "content": "We argue that existing prominent results on standard video-text tasks can be misleading as models largely exploit the shortcuts and biases inherent in the dataset. Many of the questions can be answered by objects or context extracted from a single frame without capturing cross-frame relations. For the examples in Fig. 2a, the video-text alignment can be easily inferred from shortcuts such as \"cymbals\" or \"football\". Moreover, the models may utilize biases in the datasets, such as the spurious correlation between \"outdoor\" and \"football\". Current evaluations of video-text understanding are compromised by shortcuts and biases, which obscure us from analyzing the limitations of current models. As we proceed from image-text understanding to video-text understanding, we should focus on more challenging semantics in the video domain that require cross-frame reasoning to solve, such as the interaction between person and object or the change of appearances over a sequence of frames (see Fig. 1a)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.441, + 0.79, + 0.577 + ], + "angle": 0, + "content": "In this paper we propose a novel evaluation paradigm of video-text understanding, i.e., retrieval from counterfactually augmented data (RCAD). As demonstrated in Fig. 2b, the goal is to retrieve the only positive caption with matched semantics among \"hard\" negative captions. Negative captions are counterfactually augmented so that video-text models must derive a holistic understanding of the video sequence for both objects and actions, in order to retrieve the correct caption. As a comparison, negative examples from standard video-to-text retrieval [22, 24, 28] are captions of different videos in the same dataset, often containing different object entities that are easy to distinguish." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.58, + 0.79, + 0.777 + ], + "angle": 0, + "content": "We follow the previous human-in-the-loop system [9] and develop a benchmark dataset for our new RCAD task, i.e., the Feint6K dataset. We test a wide range of public video-text methods with various pretraining strategies on this new benchmark. Two failure examples are visualized in Fig. 1a and quantitative results are summarized in Fig. 1b. Note how state-of-the-art video-text model, InternVideo [23], features a \\(87.9\\%\\) rank-1 accuracy on standard video-to-text retrieval but drops by \\(32.8\\%\\) (pretraining only) and \\(29.7\\%\\) (finetuned) when tested on our benchmark with counterfactually augmented data. By establishing a human-level baseline on our benchmark, the InternVideo model surprisingly falls far behind human performance by \\(38.6\\%\\). Our findings sharply contrast to the prominent performance obtained on previous video-text benchmarks and question the common belief that latest video-text models can develop a fairly effective representation for texts and videos in existing datasets." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.841 + ], + "angle": 0, + "content": "In light of our evaluation results on Feint6K, we identify a significant limitation of the widely adopted contrastive representation learning on video-text data, which is the issue of shortcut learning (see Sec. 4.1). To address this, we propose LLM-teacher that enables a more effective learning of action se" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.336, + 0.115, + 0.731, + 0.129 + ], + "angle": 0, + "content": "Rethinking Video-Text Understanding: RCAD and Feint6K" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.144, + 0.338, + 0.201 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.218, + 0.205, + 0.335, + 0.232 + ], + "angle": 0, + "content": "Groundtruth: \n\\(\\checkmark\\) A person sitting outside with some tools is examining a rock on a leather piece." + }, + { + "type": "image_caption", + "bbox": [ + 0.218, + 0.235, + 0.334, + 0.263 + ], + "angle": 0, + "content": "Predicted: \n\\(x\\) A person sitting outside with some tools is throwing a rock on a leather piece." + }, + { + "type": "image", + "bbox": [ + 0.346, + 0.144, + 0.466, + 0.201 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.346, + 0.204, + 0.461, + 0.224 + ], + "angle": 0, + "content": "Groundtruth: A person is building a space jet toy at high speed." + }, + { + "type": "image_caption", + "bbox": [ + 0.346, + 0.229, + 0.462, + 0.25 + ], + "angle": 0, + "content": "Predicted: \n\\(\\times\\) A person is coloring a space jet toy at high speed." + }, + { + "type": "image_caption", + "bbox": [ + 0.218, + 0.267, + 0.466, + 0.28 + ], + "angle": 0, + "content": "(a) Failures of InternVideo [23] on Feint6K" + }, + { + "type": "image", + "bbox": [ + 0.478, + 0.149, + 0.78, + 0.261 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.504, + 0.267, + 0.757, + 0.279 + ], + "angle": 0, + "content": "(b) Quantitative results on Feint6K dataset" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.291, + 0.788, + 0.402 + ], + "angle": 0, + "content": "Fig. 1: (a): Although with large-scale pretraining on web-scale data, current video-text model like [23] can be easily fooled by counterfactually augmented data. (b): The performance of InternVideo on retrieval from counterfactually augmented data (RCAD) drops by over \\(30\\%\\) when compared to the standard video-to-text retrieval and by \\(38.6\\%\\) when compared to human-level performance. We also evaluate models on standard video-text retrieval from only 6 sampled candidates and show that our RCAD task is indeed much more challenging. Our LLM-teacher successfully improves the performance on RCAD by enforcing a more effective learning of action semantics." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.433, + 0.788, + 0.569 + ], + "angle": 0, + "content": "mantics by introducing extra knowledge from pretrained large language models (LLMs). Specifically we generate synthetic captions by modifying the contents of an existing caption and LLM serves as the teacher to determine if a synthetic caption matches the semantics in the original caption using binary pseudo-labels or continuous logits. This approach enables a more effective learning of action embeddings from the available video-text data. Experimental results show that LLM-teacher learns more discriminative action embeddings from our advanced contrastive objectives and improves the performance of retrieval from counter-factually augmented data when applied to multiple video-text models." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.57, + 0.788, + 0.751 + ], + "angle": 0, + "content": "In summary, our key contributions are as follows. (1) We propose to evaluate video-text models on questions that require cross-frame reasoning and develop a new task, retrieval from counterfactually augmented data, and a new dataset, Feint6K (Sec. 3). (2) Extensive experimental results on our new evaluation paradigm suggest that existing video-text models demonstrate very limited understanding of the action semantics in a video, which is contrary to the prominent performance achieved on standard video-text retrieval benchmarks. (3) From our results on Feint6K dataset, we identify a key shortcoming of contrastive learning approaches on video-text data. We present LLM-teacher to enforce a more effective learning of action embeddings by injecting knowledge from pretrained LLMs (Sec. 4). Our approach effectively improves the results on Feint6K when applied to multiple text-video models." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.776, + 0.396, + 0.791 + ], + "angle": 0, + "content": "2 Related Works" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Video-text pretraining. With the availability of web-scale video-text paired datasets, such as WebVid2M [1] and HowTo100M [17], recent developments of" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.128 + ], + "angle": 0, + "content": "W. Ma et al." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.143, + 0.462, + 0.304 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.235, + 0.304, + 0.442, + 0.316 + ], + "angle": 0, + "content": "(a) Standard video-to-text retrieval" + }, + { + "type": "image", + "bbox": [ + 0.543, + 0.147, + 0.791, + 0.302 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.634, + 0.303, + 0.699, + 0.315 + ], + "angle": 0, + "content": "(b) RCAD" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.326, + 0.788, + 0.438 + ], + "angle": 0, + "content": "Fig. 2: Different evaluations of video-text understanding. (a): In standard video-to-text retrieval, negative captions are sampled from different videos in the same dataset. Image-text models can achieve good performance by exploiting shortcuts (e.g., \"football\" and \"cymbals\") and biases (e.g., spurious correlation between \"outdoor\" and \"football\"). (b): In our proposed RCAD paradigm, we adopt a human-in-the-loop system (see Sec. 3.2) to obtain \"hard\" negative captions with unchanged object entities but modified actions. Models must develop a holistic understanding of the semantics from the sequence of frames to retrieve the matched caption." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.469, + 0.789, + 0.651 + ], + "angle": 0, + "content": "video-text pretraining models achieved improved results on a wide range of video understanding tasks. Visual and textual embeddings jointly learned from the large-scale data demonstrate strong transferability and largely reduce the efforts in downstream tasks. Mainstream pretraining objectives can be categorized into discriminative and generative. Discriminative approaches extend the objective of CLIP [18] to the video-text domain and learn multi-modal representations by contrasting between matched and unmatched pairs [15, 16, 23, 26] or simply predicting if a video-text pair is matched [16]. Generative methods follow the masked modeling idea in BERT [5] and adopted masked language modeling (MLM) and masked video modeling (MVM) for video-text pretraining [16, 21, 23]. These generative approaches are often considered superior for action recognition, while discriminative approaches learn better semantics from texts [21]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.789, + 0.809 + ], + "angle": 0, + "content": "Evaluation of video-text models. In order to analyze the effectiveness of video-text pretraining, previous works [15,16,21,23,26,29] focused on the following tasks: (1) video-text retrieval (zero-shot or finetuned) [22, 28]: retrieving the matched video (or text) given a text (or video) as query where the negative candidates are unmatched pairs from the same dataset, (2) video question answering (finetuned) [25]: predicting the answer with a classification head, (3) video classification (zero-shot or finetuned) [6,10,20]: classifying videos using label names as text prompts, and (4) video captioning (zero-shot or finetuned) [11, 28]: summarizing the contents of a video." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.789, + 0.84 + ], + "angle": 0, + "content": "We argue that results on these video-text tasks could be deceitful as most tasks heavily rely on the alignment of object entities in the video and text" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.336, + 0.114, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Rethinking Video-Text Understanding: RCAD and Feint6K" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.359 + ], + "angle": 0, + "content": "(which image-text models are also capable of), and “action understanding” can often be developed through shortcuts and biases. For instance, a commonly-used evaluation for “action recognition” is video classification on ActivityNet [6], where a model classifies a given video into classes such as “playing badminton”, “kayaking”, or “volleyball”. However, video-text models would classify the videos by exploiting objects and contexts as shortcuts (e.g., “badminton” and “kayak”), as well as other biases (e.g., spurious correlation between “kayak” and “on the water”) without genuinely understanding the semantics of the action represented by interactions between the person and the objects over time. For instance, classifying the action as “kayaking” is essentially a “kayak” detection problem in ActivityNet. Therefore we propose retrieval from counterfactually augmented data, a new evaluation paradigm where we aim to eliminate the shortcuts from the questions so models must establish a comprehensive understanding of the semantics from cross-frame reasoning in order to predict the correct answer." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.382, + 0.733, + 0.399 + ], + "angle": 0, + "content": "3 Retrieval from Counterfactually Augmented Data" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.414, + 0.788, + 0.565 + ], + "angle": 0, + "content": "Previous evaluation tasks of video-text models focused on the alignment of feature embeddings of the video-text pairs (e.g., video-text retrieval [22, 28] and video classification [10, 20]) or reconstruction of the semantics with text (e.g., video question answering [25] and video captioning [11, 28]). However, these evaluation tasks are largely limited by the paired data available in existing video-text datasets. As demonstrated by in Fig. 2a, we can extract most of the semantics of a video, such as the objects, people, and contexts in the video, by looking at only one frame from the video. The action in the video can be further inferred by exploiting biases in the datasets. Therefore, models pretrained on image-text data often perform surprisingly well on existing video-text tasks." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.566, + 0.788, + 0.747 + ], + "angle": 0, + "content": "To effectively evaluate how video-text models can understand the contents and semantics of videos beyond images, we must develop video understanding tasks that are free from shortcuts and require cross-frame reasoning to solve. To this end, we propose a new evaluation task for video-text models, i.e., retrieval from counterfactually augmented data (Sec. 3.1), and a new dataset Feint6K (Sec. 3.2). In contrast to previous datasets with matched video-text pairs scrapped from the web [24, 28] or annotated by human [22], we adopt a human-in-the-loop system and annotate counterfactually manipulated texts as negative pairs with the original video. This allows us to evaluate video-text models with novel tasks and better understand their limitations. Lastly we measure human performance on our dataset as a direct comparison with the state-of-the-art video-text models (Sec. 3.3)." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.77, + 0.407, + 0.784 + ], + "angle": 0, + "content": "3.1 Task Formulation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Retrieval from counterfactually augmented data is a variant of the standard video-to-text retrieval. As shown in Fig. 2b, given a video sequence and a list of candidate captions, the model is asked to retrieve the caption that best matches" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.127 + ], + "angle": 0, + "content": "W. Ma et al." + }, + { + "type": "title", + "bbox": [ + 0.214, + 0.155, + 0.275, + 0.17 + ], + "angle": 0, + "content": "Video A" + }, + { + "type": "image", + "bbox": [ + 0.292, + 0.143, + 0.788, + 0.184 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.201, + 0.274, + 0.214 + ], + "angle": 0, + "content": "Video B" + }, + { + "type": "image", + "bbox": [ + 0.291, + 0.188, + 0.787, + 0.228 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.241, + 0.275, + 0.255 + ], + "angle": 0, + "content": "Video C" + }, + { + "type": "image", + "bbox": [ + 0.292, + 0.233, + 0.787, + 0.263 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.291, + 0.268, + 0.402, + 0.279 + ], + "angle": 0, + "content": "Captions of Video A" + }, + { + "type": "text", + "bbox": [ + 0.291, + 0.282, + 0.446, + 0.311 + ], + "angle": 0, + "content": "[Positive] A person is shinning a shoe with a shoe shinning machine." + }, + { + "type": "text", + "bbox": [ + 0.291, + 0.314, + 0.432, + 0.343 + ], + "angle": 0, + "content": "[Negative 1] A person is making a shoe with a shoe shimming machine." + }, + { + "type": "text", + "bbox": [ + 0.291, + 0.346, + 0.436, + 0.375 + ], + "angle": 0, + "content": "[Negative 2] A person is wearing a shoe with a shoe shinning machine." + }, + { + "type": "title", + "bbox": [ + 0.46, + 0.268, + 0.57, + 0.279 + ], + "angle": 0, + "content": "Captions of Video B" + }, + { + "type": "text", + "bbox": [ + 0.459, + 0.282, + 0.615, + 0.311 + ], + "angle": 0, + "content": "[Positive] A man is standing on the field attempting to kick a football." + }, + { + "type": "text", + "bbox": [ + 0.459, + 0.314, + 0.616, + 0.343 + ], + "angle": 0, + "content": "[Negative 1] A man is standing on the field attempting to catch a football." + }, + { + "type": "text", + "bbox": [ + 0.459, + 0.346, + 0.617, + 0.375 + ], + "angle": 0, + "content": "[Negative 2] A man is standing on the field attempting to throw a football." + }, + { + "type": "title", + "bbox": [ + 0.628, + 0.268, + 0.738, + 0.279 + ], + "angle": 0, + "content": "Captions of Video C" + }, + { + "type": "text", + "bbox": [ + 0.627, + 0.282, + 0.782, + 0.311 + ], + "angle": 0, + "content": "[Positive] A person wearing a pink jacket is feeding birds with his hand." + }, + { + "type": "text", + "bbox": [ + 0.627, + 0.314, + 0.786, + 0.343 + ], + "angle": 0, + "content": "[Negative 1] A person wearing a pink jacket is petting birds with his hand." + }, + { + "type": "text", + "bbox": [ + 0.627, + 0.346, + 0.786, + 0.375 + ], + "angle": 0, + "content": "[Negative 2] A person wearing a pink jacket is holding birds with his hand." + }, + { + "type": "image_caption", + "bbox": [ + 0.327, + 0.386, + 0.674, + 0.4 + ], + "angle": 0, + "content": "Fig. 3: Examples of \\( RCAD \\) on our Feint6K dataset." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.43, + 0.785, + 0.49 + ], + "angle": 0, + "content": "the semantics in the video. However, unlike standard video-to-text retrieval where negative captions come from other video-text pairs in the same dataset, negative captions in our task are modified from the positive captions, with the same text structure and object entities but different actions (see Fig. 3)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.491, + 0.787, + 0.536 + ], + "angle": 0, + "content": "Similar to video-text retrieval, our proposed task provides the benefit of zero-shot evaluation. This allows us to quantitatively analyze the effectiveness of video-text pretraining without downstream finetuning." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.537, + 0.788, + 0.687 + ], + "angle": 0, + "content": "What distinguishes our task from all previous video-text tasks is its focus on questions that require cross-frame reasoning. Consider \"Video B\" in Fig. 3 as an example. All captions contain the same object entities present in the video, i.e., \"a man\", \"the field\", and \"a football\", but differ in the action specified. We cannot determine the most accurate caption because all suggested actions, \"kick\", \"catch\", and \"throw\", are plausible given the video's context. To retrieve the corresponding caption, it is crucial to grasp the action's semantics by inspecting the interactions between the man and the football across a series of frames. This underscores the necessity of understanding the video semantics over time in order to succeed in our task, rather than relying shortcuts or biases." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.71, + 0.471, + 0.724 + ], + "angle": 0, + "content": "3.2 Feint6K: Data Collection" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.734, + 0.788, + 0.841 + ], + "angle": 0, + "content": "We utilize a human-in-the-loop system [9] to counterfactually manipulate the positive captions in existing video-text datasets [22, 28]. The generated counterfactually augmented data, when paired with the original videos, forms negative pairs to be used in our proposed task. We recruit 40 annotators to manually manipulate the actions in existing captions. Specifically, the new actions introduced must be plausible within the context in the caption but are not occurring in the corresponding video. In summary, a total of 6,243 videos from the validation" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.335, + 0.114, + 0.733, + 0.129 + ], + "angle": 0, + "content": "Rethinking Video-Text Understanding: RCAD and Feint6K" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "image", + "bbox": [ + 0.245, + 0.142, + 0.761, + 0.267 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.276, + 0.785, + 0.304 + ], + "angle": 0, + "content": "Fig. 4: Overview of our data collection pipeline for Feint6K dataset featuring a human-in-the-loop system." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.335, + 0.784, + 0.379 + ], + "angle": 0, + "content": "set of the MSR-VTT dataset [28] and the test set of the VATEX dataset [22]. For detailed statistics of our Feint6K dataset, please refer to our supplementary materials." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.381, + 0.785, + 0.517 + ], + "angle": 0, + "content": "The overview of our human-in-the-loop system is demonstrated in Fig. 4. We start by sampling a small subset of video-text pairs and manually annotate the counterfactually augmented captions for the purpose of training and demonstration. Specific guidance and feedback are given to the annotators on the practice questions during the training stage. Moreover, we adopt a reviewing process during the annotation stage to further ensure the quality of our annotated data. Each annotation is reviewed before acceptance, where rejected annotations are marked and sent back for refinement. We refer the readers to the supplementary materials regarding the annotation interface and annotator guidance." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.542, + 0.433, + 0.556 + ], + "angle": 0, + "content": "3.3 Human Performance" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.569, + 0.785, + 0.644 + ], + "angle": 0, + "content": "We employ the same group of annotators to establish a human-level baseline on our Feint6K dataset. Analyzing the human performance serves two goals: (i) verifying the legitimacy of our Feint6K dataset – if each question is answerable and has one unique answer, and (ii) comparing the state-of-the-art video-text models with human-level performance." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.645, + 0.786, + 0.75 + ], + "angle": 0, + "content": "Each annotator is first presented with the full video and then asked to select the one out of six captions that best matches the semantics in the video. To avoid leakage, we make sure the annotators assigned to the testing questions are not the same annotators annotating this video. We also randomly shuffle the order of the candidate captions to remove any biases. In Sec. 5.2 we present the quantitative comparisons between human-level performance and current state-of-the-art video-text models on our Feint6K dataset." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.776, + 0.379, + 0.791 + ], + "angle": 0, + "content": "4 LLM-teacher" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.784, + 0.84 + ], + "angle": 0, + "content": "Contrastive learning are widely adopted in previous video-text representation learning works [16, 23, 26]. The contrastive models learn to distinguish between" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.128 + ], + "angle": 0, + "content": "W. Ma et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.783, + 0.252 + ], + "angle": 0, + "content": "similar and dissimilar pairs of video-text data by contrasting their feature representations. Despite the simple formulation and great generalization ability, our results on Feint6K show that current video-text models built heavily on contrastive objectives have very limited understanding of action semantics in natural videos and can be easily fooled by counterfactually augmented data. Therefore we aim to develop a more effective contrastive approach and learn more discriminative action representations from existing video-text data." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.253, + 0.785, + 0.329 + ], + "angle": 0, + "content": "Sec. 4.1 discusses the limitations of current contrastive approaches due to numerous shortcuts in video-text dataset. Inspired by this finding, we present LLM-teacher in Sec. 4.2, a simple but effective approach to learn more discriminative action embeddings by incorporating knowledge from pretrained large language models." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.353, + 0.506, + 0.367 + ], + "angle": 0, + "content": "4.1 Shortcuts in Video-Text Data" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.38, + 0.783, + 0.455 + ], + "angle": 0, + "content": "The CLIP model [18] trained on image-text data with a contrastive objective demonstrated strong zero-shot capabilities in many vision tasks [13,30] and was widely adopted as the vision encoder in many large vision-language models [14]. However, directly generalizing this idea to video-text data may face unprecedented challenges due to shortcuts in the video-text data." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.457, + 0.784, + 0.652 + ], + "angle": 0, + "content": "As we proceed from image-text to video-text representation learning, we aim to learn a powerful embedding for not only the object entities, but also the actions specified by texts or sequence of frames. Naturally we hope to achieve this by contrasting videos with different actions. However, as we are optimizing the contrastive objectives in a mini-batch of videos, object entities become the shortcuts that contrastive models exploit to saturate the contrastive objective, leading to an ineffective action embeddings. Consider the example in Fig. 2a, the contrastive loss would be small as long as the model learns discriminative embeddings for shortcut objects such as \"cymbal\" and \"football\", even without any understanding of the actions. Since many video-text models start with pretrained CLIP encoders [23,26], they often possess strong embeddings for shortcut objects from the beginning of the video-text pretraining, which hinders the model from further learning effective action representations." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.653, + 0.784, + 0.728 + ], + "angle": 0, + "content": "In Fig. 5 we present quantitative results of changes in cosine similarities when the video is unchanged but the object or action in the caption is manipulated. We find that state-of-the-art video-text models learn less discriminative embeddings for actions as compared to objects, which also explains the small gap between image-text and video-text models on standard video-text tasks." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.753, + 0.371, + 0.767 + ], + "angle": 0, + "content": "4.2 LLM-teacher" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.803, + 0.84 + ], + "angle": 0, + "content": "We introduce LLM-teacher, an LLM-powered approach to learn better action representations with contrastive objectives. Specifically, LLM serves as the \"teacher\" and provides synthesized captions as extra knowledge for the video-text model to learn from. These new captions serve as negative captions of the original video," + } + ], + [ + { + "type": "header", + "bbox": [ + 0.336, + 0.114, + 0.733, + 0.129 + ], + "angle": 0, + "content": "Rethinking Video-Text Understanding: RCAD and Feint6K" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "image", + "bbox": [ + 0.224, + 0.148, + 0.495, + 0.264 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.218, + 0.27, + 0.498, + 0.293 + ], + "angle": 0, + "content": "(a) Comparison between \\(\\Delta s\\) w.r.t. object or action changes." + }, + { + "type": "image", + "bbox": [ + 0.512, + 0.149, + 0.782, + 0.264 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.506, + 0.27, + 0.786, + 0.292 + ], + "angle": 0, + "content": "(b) Comparison between \\(\\Delta s\\) w.r.t. action using InternVideo and our LLM-teacher." + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.303, + 0.785, + 0.4 + ], + "angle": 0, + "content": "Fig. 5: Change of cosine similarity w.r.t. objects or actions. (a): Comparison between the change in cosine similarity when the action or object is swapped. Results show that current video-text models learn a more effective embedding for objects than for actions. (b): Comparison between the change in cosine similarity using InternVideo or our LLM-teacher. This demonstrate that LLM-teacher learns a more discriminative embedding for actions by enabling a more effective contrastive learning using knowledge from LLMs. Refer to Sec. 5.2 for more details." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.436, + 0.785, + 0.465 + ], + "angle": 0, + "content": "removing possible shortcuts and biases and enabling a more effective learning of action semantics." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.469, + 0.785, + 0.543 + ], + "angle": 0, + "content": "We start with positive captions from common video-text datasets and run an abstract meaning representation (AMR) parser, resulting in a list of action or object tokens. To generate the hard negative captions, two methods are considered to find novel actions or objects given the original caption and the substitution token \\( w \\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.579, + 0.785, + 0.685 + ], + "angle": 0, + "content": "Method I: Mask filling. Masked language modeling (MLM) [5] is a widely adopted self-supervised task for LLMs. In the pretraining stage, a certain percentage of word tokens are substituted with a special token [mask] and the LLM is optimized to recover the masked tokens. To find negative captions with novel actions given the original caption \\( c_{0} \\) and substitution token \\( a \\), we first substitute the token \\( a \\) with [mask] and use a LLM pretrained with MLM to predict a list of \\( k \\) possible tokens \\( \\{\\hat{w}_i\\}_{i=1}^k \\) such that \\( \\hat{w}_i \\neq w \\) and \\( \\hat{w}_i \\) is an action token." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.719, + 0.785, + 0.84 + ], + "angle": 0, + "content": "Method II: LLM-powered chatbot. Although mask filling can find appropriate words to substitute \\( w \\), it is limited by one single token. In many cases, we want to update the prepositions following the change of verbs. For example, when \"install\" is substituted by \"uninstall\", we should also update the preposition from \"install ... to\" to \"uninstall ... from\". Here we consider a more flexible approach that builds on LLM-powered chatbots. Besides a prompt describing the text substitution, we leverage the in-context learning ability of LLMs [2, 7] and provide multiple in-context examples to obtain the desirable negative captions." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.358, + 0.127 + ], + "angle": 0, + "content": "W. Ma et al." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.147, + 0.784, + 0.238 + ], + "angle": 0, + "content": "LLM-teacher. With the procedure above, we obtain \\( k \\) synthesized negative captions \\( \\{n_i\\}_{i=1}^k \\) besides the original positive caption \\( p \\) for each video \\( v \\) in existing datasets [22,28]. For each of the original and generated captions, LLM serves as a \"teacher\" and provide binary pseudo-labels - if the caption have matched (\"1\") or unmatched (\"0\") semantics with the video. Now we may optimize the video-text model with the normalized temperature-scaled cross entropy loss" + }, + { + "type": "equation", + "bbox": [ + 0.297, + 0.247, + 0.786, + 0.285 + ], + "angle": 0, + "content": "\\[\nl = - \\log \\frac {\\exp \\left(\\sin \\left(f _ {v} , f _ {p}\\right) / \\tau\\right)}{\\exp \\left(\\sin \\left(f _ {v} , f _ {p}\\right) / \\tau\\right) + \\sum_ {i = 1} ^ {k} \\exp \\left(\\sin \\left(f _ {v} , f _ {n _ {i}}\\right) / \\tau\\right)} \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.294, + 0.617, + 0.309 + ], + "angle": 0, + "content": "where \\(f_{v}, f_{p}, f_{ni}\\) are the visual and textual embeddings." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.31, + 0.784, + 0.4 + ], + "angle": 0, + "content": "In practice, we find that certain negative captions have unmatched but similar semantics to the video. It is undesirable to regard them as a strictly negative pair in the contrastive loss. Therefore, we extend the binary pseudo-labels to soft logits by computing caption similarities with a pretrained LLM and then optimize the video-text model to match model outputs with the soft logits from the LLM teacher model [8]." + }, + { + "type": "equation", + "bbox": [ + 0.354, + 0.413, + 0.785, + 0.429 + ], + "angle": 0, + "content": "\\[\nl = \\mathcal {L} _ {\\mathrm {K L}} \\left(z _ {\\text {v i d e o - t e x t}}, z _ {\\text {L L M}}\\right) \\tag {2}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.285, + 0.432, + 0.717, + 0.468 + ], + "angle": 0, + "content": "\\[\nz _ {\\mathrm {v i d e o - t e x t , t}} = \\frac {\\exp (\\sin (f _ {v} , f _ {t}) / \\tau)}{\\exp (\\sin (f _ {v} , f _ {p}) / \\tau) + \\sum_ {i = 1} ^ {k} \\exp (\\sin (f _ {v} , f _ {n _ {i}}) / \\tau)}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.315, + 0.47, + 0.715, + 0.507 + ], + "angle": 0, + "content": "\\[\nz _ {\\mathrm {L L M , t}} = \\frac {\\exp (\\sin (e _ {p} , e _ {t}) / \\tau)}{\\exp (\\sin (e _ {p} , e _ {p}) / \\tau) + \\sum_ {i = 1} ^ {k} \\exp (\\sin (e _ {p} , e _ {n _ {i}}) / \\tau)}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.517, + 0.784, + 0.548 + ], + "angle": 0, + "content": "where \\( t \\in \\{n_i\\}_{i=1}^k \\cup \\{p\\} \\), \\( f \\)'s are textual and visual embeddings of the video-text model, and \\( e \\)'s are textual embeddings of a pretrained LLM." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.571, + 0.375, + 0.588 + ], + "angle": 0, + "content": "5 Experiments" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.603, + 0.427, + 0.619 + ], + "angle": 0, + "content": "5.1 Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.628, + 0.784, + 0.689 + ], + "angle": 0, + "content": "Datasets. We use MSR-VTT dataset [28] and VATEX dataset [22] for standard video-to-text retrieval and our Feint6K dataset for retrieval from counter-factually augmented data. Note that Feint6K dataset contains the same video sequences for evaluation as in MSR-VTT dataset and VATEX dataset." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.712, + 0.784, + 0.787 + ], + "angle": 0, + "content": "Evaluation metrics. For standard video-to-text retrieval we report rank-1 accuracy \\((\\mathbb{R}@\\mathbb{1})\\) as this is the most challenging metric and distinguish the performance of previous video-text models. For retrieval from counterfactually-augmented data we use rank-1 accuracy \\((\\mathbb{R}@\\mathbb{1})\\), rank-2 accuracy \\((\\mathbb{R}@\\mathbb{2})\\), and mean rank (MeanR)." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.81, + 0.784, + 0.84 + ], + "angle": 0, + "content": "Video-text models. We consider multiple state-of-the-art public models in this work. CLIP4Clip [15] extends CLIP model [18] to the video domain using a" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.336, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Rethinking Video-Text Understanding: RCAD and Feint6K" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "table", + "bbox": [ + 0.217, + 0.144, + 0.787, + 0.395 + ], + "angle": 0, + "content": "
ModelMSR-VTTFeint6K (MSR-VTT)VATEXFeint6K (VATEX)
R@1R@1R@2MeanRR@1R@1R@2MeanR
Human95.296.8
Random<1e-316.7<1e-316.7
Zero-Shot
CLIP [18]26.337.355.32.638.834.854.32.7
VideoCLIP [27]14.535.171.32.613.733.070.72.7
InternVideo [23]37.545.863.62.376.944.163.92.3
LanguageBind [31]42.841.377.02.442.377.02.4
Finetuned
LanguageBind [31]
CLIP4Clip [15]43.150.872.42.0
VindLU [3]46.653.470.92.0
SimVTP [16]50.235.770.82.676.633.668.42.6
w/ LLM-teacher-lbl49.0 ↓1.240.0 ↑4.373.1 ↑2.32.3 ↓0.374.8 ↓1.837.3 ↑3.772.2 ↑3.82.4 ↓0.2
w/ LLM-teacher-lgt49.5 ↓0.743.5 ↑7.875.0 ↑4.22.2 ↓0.475.3 ↓1.340.1 ↑6.573.5 ↑5.12.3 ↓0.3
InternVideo [23]49.158.680.21.887.958.276.91.9
w/ LLM-teacher-lbl48.2 ↓0.964.2 ↑5.682.5 ↑2.31.7 ↓0.185.2 ↓2.763.8 ↑5.680.5 ↑3.61.7 ↓0.2
w/ LLM-teacher-lgt48.9 ↓0.265.8 ↑7.283.8 ↑3.61.7 ↓0.187.3 ↓0.665.6 ↑7.481.7 ↑4.81.7 ↓0.2
" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.406, + 0.788, + 0.518 + ], + "angle": 0, + "content": "Table 1: Performance of standard video-to-text retrieval on MSR-VTT dataset [28] and VATEX dataset [22], as compared to performance of retrieval from counterfactually-augmented data on our Feint6K dataset. In the \"zero-shot\" setting, models are evaluated directly after the pretraining stage, while in the \"finetuned\" setting, models are finetuned on the training set of MSR-VTT or VATEX. With LLM-teacher, we enable a more efficient learning of action semantics and effectively improves R@1 accuracy on Feint6K. Here \"LLM-teacher-lbl\" stands for our approach with binary pseudo-labels and \"LLM-teacher-lgt\" uses soft logits." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.548, + 0.788, + 0.653 + ], + "angle": 0, + "content": "mean pooling mechanism for zero-shot video-text retrieval and a contrastive loss for finetuning. SimVTP [16] is pretrained on the WebVid2M dataset [1] with a combination of contrastive learning and masked modeling. InternVideo [23] is a video-text foundation model pretrained on a combination of 7 large-scale videotext datasets, with a supervised video post-pretraining for better video recognition. LanguageBind [31] is a language-based multi-modal pretraining model with remarkable performance on a wide range of benchmarks." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.655, + 0.788, + 0.76 + ], + "angle": 0, + "content": "To demonstrate the efficacy of our LLM-teacher, we adopt this method on two state-of-the-art video-text models, SimVTP and InternVideo. In the default setting we generate 10 action-based captions with LLM for each training video. For text generation we use a pretrained XLM-RoBERTa [4] for mask filling. To measure text similarities and compute soft logits, we utilize a pretrained Sentence-BERT model [19]. In Sec. 5.3 we ablate on the choices of parameters and caption generation." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.784, + 0.373, + 0.797 + ], + "angle": 0, + "content": "5.2 Main Results" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.789, + 0.84 + ], + "angle": 0, + "content": "Performance of previous state-of-the-art. In Tab. 1 we report the quantitative results of standard video-text retrieval on MSR-VTT and VATEX, and of" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.359, + 0.127 + ], + "angle": 0, + "content": "W. Ma et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.314 + ], + "angle": 0, + "content": "retrieval from counterfactually augmented data on our Feint6K. We also report a \"Random\" performance where a model predicts random guesses. We make the following observations: (i) Results show that previous state-of-the-art video-text models demonstrate limited understanding of the action semantics in a video with less than \\(60\\%\\) R@1 accuracy on RCAD, given a \\(16.7\\%\\) R@1 accuracy when taking random guesses. (ii) Although it is hard to compare the quantitative results between two tasks, we note that our RCAD is much more challenging given the performance gap between \"Random\" and previous state-of-the-arts. This supports our previous arguments that our RCAD removes shortcuts from standard video-text datasets and focuses on harder questions in the video domain that require cross-frame reasoning." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.337, + 0.79, + 0.535 + ], + "angle": 0, + "content": "Performance of LLM-teacher. We apply our LLM-teacher to two pretrained video-text models, SimVTP [16] and InternVideo [23]. Specifically, we consider the two objectives in Eq. (1) and Eq. (2), labeled as \"LLM-teacher-lbl\" and \"LLM-teacher-lgt\" respectively. Results in Tab. 1 show that by exploiting knowledge from pretrained LLMs, LLM-teacher enables a more efficient leaning of action semantics and achieved improved results on various metrics of RCAD, increasing the R@1 accuracy on RCAD by \\(7.2\\%\\) and \\(7.4\\%\\). We also note that with binary pseudo-labels, results demonstrate a trade-off between the performance on standard video-to-text retrieval and RCAD; and in comparison, LLM-teacher with soft logits achieves the highest performance on RCAD (\\(65.8\\%\\) and \\(65.6\\%\\)) with negligible drops (\\(0.2\\%\\) and \\(0.6\\%\\)) on standard retrieval. This is because soft logits account for the text similarities between different actions and avoid overfitting on binary pseudo-labels." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.559, + 0.79, + 0.71 + ], + "angle": 0, + "content": "Comparison with human-level performance. The human-level performance are reported in Tab. 1. We see that human annotators achieved an almost perfect performance on Feint6K dataset, with a \\(95.2\\%\\) accuracy on MSR-VTT videos and a \\(96.8\\%\\) accuracy on VATEX videos. These results show that: (i) most questions for retrieval from counterfactually-augmented data are answerable with an unique answer, and (ii) this task is fairly simple for human. As a comparison, all previous video-text models with large-scale pretraining fall behind by a wide margin. Specifically, the state-of-the-art video-text model InternVideo [23] features a heavy two-stage pretraining on a collection of 7 large-scale video-text datasets but falls behind by \\(36.6\\%\\) and \\(38.6\\%\\) even with downstream finetuning." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Analyses of failure cases. In Fig. 1b we present two failure cases by the InternVideo [23] model on our Feint6K dataset. Both examples are trivial for human, yet challenging from video-text models due to incapability of cross-frame reasoning. For the first example, a model must analyze the interactions between the person and the rock over a sequence of frames to predict the correct answer. And in the second example, the model must observe the change of appearances over time. As we can see, with the counterfactually augmented data in our Feint6K," + } + ], + [ + { + "type": "header", + "bbox": [ + 0.335, + 0.114, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Rethinking Video-Text Understanding: RCAD and Feint6K" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.224 + ], + "angle": 0, + "content": "models cannot \"guess\" the correct answer by exploiting shortcuts extracted from a single frame. The findings on Feint6K highlight significant weaknesses in current video-text models, offering valuable insights for future researches in this area. Additional qualitative examples of video-text models on Feint6K dataset are provided in the supplementary materials for readers' reference." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.244, + 0.788, + 0.394 + ], + "angle": 0, + "content": "Sensitivity to the change of object or action. With the counterfactually augmented data collected, we investigate how the cosine similarities between the video and text change when the object or action in the text are swapped with an object or action not present in the image. Given the original video-text pair \\((v,t)\\) and a counterfactually-augmented text \\(\\hat{t}\\), the change in cosine similarity is given by \\(\\Delta s = s(v,\\hat{t}) - s(v,t)\\). As \\(\\hat{t}\\) contains objects or actions not present in the video, \\(\\Delta s\\) should be negative for an ideal video-text model. Moreover, larger \\(|\\Delta s|\\) implies the model being more sensitive to the changes. We compute \\(\\Delta s\\) using the InternVideo with pretraining only (\"IntVid-pt\"), InternVideo with downstream finetuning (\"IntVid-ft\"), and our LLM-teacher (\"LLM-teacher\")." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.396, + 0.788, + 0.531 + ], + "angle": 0, + "content": "Results in Fig. 5a show that the \\(\\Delta s\\) are almost always negative when the objects are swapped and the absolute changes are larger. In comparison, \\(\\Delta s\\) are sometimes positive when actions are modified and the absolute changes are smaller. This demonstrate that InternVideo learns a more effective embedding for objects than for actions. This further highlights the necessity of our new evaluation paradigm for exploring the limitations of video-text models beyond existing tasks. Moreover, in Fig. 5b we note that with LLM-teacher the action embeddings are more effective than the ones learned by InternVideo, as shown by the more reasonable changes in cosine similarity when actions are swapped." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.531, + 0.787, + 0.561 + ], + "angle": 0, + "content": "Please refer to the supplementary materials for a detailed discussion regarding the influence of textual encoders." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.583, + 0.39, + 0.598 + ], + "angle": 0, + "content": "5.3 Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.607, + 0.788, + 0.669 + ], + "angle": 0, + "content": "We conduct ablation studies on the VATEX dataset [22] and our Feint6K dataset. We follow the same settings as above and report rank-1 (R@1) accuracy for standard video-to-text retrieval and rank-1 accuracy (R@1), rank-2 accuracy (R@2), and mean rank (MeanR) for counterfactual augmented data." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Choice of captions in LLM-teacher. Tab. 2 shows the comparison between different numbers and types of LLM captions generated. In the \"default\" setting we use 10 action-based LLM captions and compare to settings with 5 action-based LLM captions or 5 action-based and 5 object-based LLM captions. Results show with 5 action-based LLM captions the R@1 accuracy on Feint6K drops by \\(0.9\\%\\) while the R@1 accuracy on VATEX increases by \\(0.3\\%\\). We also experiment on object-based LLM captions and find them not beneficial for both standard video-to-text retrieval or our RCAD. This is consistent with our assumption that pretrained video-text models already learns a discriminative embedding for objects and a more efficient training objective for action is needed." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.128 + ], + "angle": 0, + "content": "W. Ma et al." + }, + { + "type": "table", + "bbox": [ + 0.288, + 0.144, + 0.719, + 0.253 + ], + "angle": 0, + "content": "
ModelVATEXFeint6K (VATEX)
R@1R@1R@2MeanR
Default(10 action captions; XLM-RoBERTa)87.365.681.71.7
5 action captions87.6 ↑0.364.7 ↓0.981.0 ↓0.71.70.0
5 object + 5 object captions87.5 ↑0.264.2 ↓1.480.6 ↓1.11.70.0
LLM Chatbot87.0 ↓0.365.9 ↑0.381.8 ↑0.11.70.0
" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.262, + 0.788, + 0.306 + ], + "angle": 0, + "content": "Table 2: Ablation studies on the number of LLM captions and caption generation methods. In the default setting we use a total of 10 action-based LLM captions and utilize a pretrained XLM-RoBERTa [4] for mask filling." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.332, + 0.789, + 0.501 + ], + "angle": 0, + "content": "Choice of caption generation. Besides a pretrained XLM-RoBERTa model for mask filling, we also experiment on generating captions with LLM-powered agents that are finetuned on chat datasets for dialogue applications. Specifically we use the \"chat\" model finetuned for dialogue applications. Empirically we find that captions generated by LLM chatbots achieve a higher overall quality, exploring a more diverse caption space and better flexibility. However we don't observe significant improvements from our ablation study experiments – the R@1 accuracy on VATEX drops by \\(0.3\\%\\) and the R@1 accuracy on Feint6K increases by \\(0.3\\%\\). We choose XLM-RoBERTa in our main experiments as the model runs faster and scale up easily to bigger settings. We provide qualitative comparisons between the two types of caption generation in our supplementary materials." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.519, + 0.371, + 0.536 + ], + "angle": 0, + "content": "6 Conclusions" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.548, + 0.79, + 0.777 + ], + "angle": 0, + "content": "In this work we propose a new evaluation task, retrieval from counterfactually augmented data, and a benchmark dataset, Feint6K. The idea is to remove shortcuts in the video-text questions with a human-in-the-loop system and produce more challenging questions where the video-text model must derive a comprehensive understanding of the video with cross-frame reasoning. Quantitative and qualitative evaluation results on our Feint6K dataset show that despite the task is trivial for human, previous state-of-the-art video-text models can be easily fooled by the counterfactually augmented data. This implies that the prominent results on previous video-text benchmarks could be misleading - models may largely exploit shortcuts without genuinely understand the video contents. Moreover, we identify a key limitation of current contrastive learning on videotext data being the shortcut learning and propose LLM-teacher that enables a more effective learning of action semantics by utilizing knowledge from pretrained LLMs. Experimental results and analyses show that our method can learn a more discriminative representation for actions." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.795, + 0.789, + 0.842 + ], + "angle": 0, + "content": "Supplementary materials. We present the following: (1) limitations of our work, (2) more details about our Feint6K dataset, (3) ethics statements, and (4) extra quantitative and qualitative results." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.336, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Rethinking Video-Text Understanding: RCAD and Feint6K" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.147, + 0.788, + 0.193 + ], + "angle": 0, + "content": "Acknowledgements. We would like to thank Daniel Khashabi, Yiyan Li, and the anonymous reviewers for their helpful comments and suggestions. Wufei Ma is supported by ONR with N00014-23-1-2641." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.215, + 0.323, + 0.231 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.246, + 0.785, + 0.287 + ], + "angle": 0, + "content": "1. Bain, M., Nagrani, A., Varol, G., Zisserman, A.: Frozen in time: A joint video and image encoder for end-to-end retrieval. In: IEEE International Conference on Computer Vision (2021) 3, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.288, + 0.785, + 0.34 + ], + "angle": 0, + "content": "2. Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J.D., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., et al.: Language models are few-shot learners. Advances in neural information processing systems 33, 1877-1901 (2020) 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.343, + 0.785, + 0.384 + ], + "angle": 0, + "content": "3. Cheng, F., Wang, X., Lei, J., Crandall, D., Bansal, M., Bertasius, G.: Vindlu: A recipe for effective video-and-language pretraining. arXiv preprint arXiv:2212.05051 (2022) 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.385, + 0.785, + 0.426 + ], + "angle": 0, + "content": "4. Conneau, A., Khandelwal, K., Goyal, N., Chaudhary, V., Wenzek, G., Guzmán, F., Grave, E., Ott, M., Zettlemoyer, L., Stoyanov, V.: Unsupervised cross-lingual representation learning at scale. arXiv preprint arXiv:1911.02116 (2019) 11, 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.427, + 0.785, + 0.468 + ], + "angle": 0, + "content": "5. Devlin, J., Chang, M.W., Lee, K., Toutanova, K.: Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805 (2018) 4, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.468, + 0.785, + 0.521 + ], + "angle": 0, + "content": "6. Fabian Caba Heilbron, Victor Escorcia, B.G., Niebles, J.C.: Activitynet: A large-scale video benchmark for human activity understanding. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 961-970 (2015) 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.523, + 0.785, + 0.564 + ], + "angle": 0, + "content": "7. Gupta, T., Kembhavi, A.: Visual programming: Compositional visual reasoning without training. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14953-14962 (2023) 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.565, + 0.785, + 0.592 + ], + "angle": 0, + "content": "8. Hinton, G., Vinyals, O., Dean, J.: Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531 (2015) 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.592, + 0.785, + 0.631 + ], + "angle": 0, + "content": "9. Kaushik, D., Hovy, E., Lipton, Z.C.: Learning the difference that makes a difference with counterfactually-augmented data. arXiv preprint arXiv:1909.12434 (2019) 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.633, + 0.785, + 0.674 + ], + "angle": 0, + "content": "0. Kay, W., Carreira, J., Simonyan, K., Zhang, B., Hillier, C., Vijayanarasimhan, S., Viola, F., Green, T., Back, T., Natsev, P., et al.: The kinetics human action video dataset. arXiv preprint arXiv:1705.06950 (2017) 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.675, + 0.785, + 0.715 + ], + "angle": 0, + "content": "1. Krishna, R., Hata, K., Ren, F., Fei-Fei, L., Carlos Niebles, J.: Dense-captioning events in videos. In: Proceedings of the IEEE international conference on computer vision. pp. 706-715 (2017) 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.716, + 0.785, + 0.757 + ], + "angle": 0, + "content": "2. Li, K., Wang, Y., Li, Y., Wang, Y., He, Y., Wang, L., Qiao, Y.: Unmasked teacher: Towards training-efficient video foundation models. arXiv preprint arXiv:2303.16058 (2023) 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.758, + 0.785, + 0.785 + ], + "angle": 0, + "content": "3. Lin, J., Gong, S.: Gridclip: One-stage object detection by grid-level clip representation learning. arXiv preprint arXiv:2303.09252 (2023) 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.785, + 0.785, + 0.798 + ], + "angle": 0, + "content": "4. Liu, H., Li, C., Wu, Q., Lee, Y.J.: Visual instruction tuning. In: NeurIPS (2023) 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.799, + 0.785, + 0.84 + ], + "angle": 0, + "content": "5. Luo, H., Ji, L., Zhong, M., Chen, Y., Lei, W., Duan, N., Li, T.: Clip4clip: An empirical study of clip for end to end video clip retrieval and captioning. Neurocomputing 508, 293-304 (2022) 4, 10, 11" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.246, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.359, + 0.127 + ], + "angle": 0, + "content": "W. Ma et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.148, + 0.785, + 0.175 + ], + "angle": 0, + "content": "16. Ma, Y., Yang, T., Shan, Y., Li, X.: Simvtp: Simple video text pre-training with masked autoencoders. arXiv preprint arXiv:2212.03490 (2022) 1, 4, 7, 11, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.178, + 0.785, + 0.218 + ], + "angle": 0, + "content": "17. Miech, A., Zhukov, D., Alayrac, J.B., Tapaswi, M., Laptev, I., Sivic, J.: HowTo100M: Learning a Text-Video Embedding by Watching Hundred Million Narrated Video Clips. In: ICCV (2019) 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.22, + 0.785, + 0.274 + ], + "angle": 0, + "content": "18. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021) 4, 8, 10, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.219, + 0.276, + 0.785, + 0.33 + ], + "angle": 0, + "content": "19. Reimers, N., Gurevych, I.: Sentence-bert: Sentence embeddings using siamese bert-networks. In: Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing. Association for Computational Linguistics (11 2019), https://arxiv.org/abs/1908.10084 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.332, + 0.785, + 0.36 + ], + "angle": 0, + "content": "20. Soomro, K., Zamir, A.R., Shah, M.: Ucf101: A dataset of 101 human actions classes from videos in the wild. arXiv preprint arXiv:1212.0402 (2012) 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.361, + 0.785, + 0.402 + ], + "angle": 0, + "content": "21. Tong, Z., Song, Y., Wang, J., Wang, L.: Videomae: Masked autoencoders are data-efficient learners for self-supervised video pre-training. Advances in neural information processing systems 35, 10078-10093 (2022) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.404, + 0.785, + 0.458 + ], + "angle": 0, + "content": "22. Wang, X., Wu, J., Chen, J., Li, L., Wang, Y.F., Wang, W.Y.: Latex: A large-scale, high-quality multilingual dataset for video-and-language research. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 4581-4591 (2019) 1, 2, 4, 5, 6, 7, 10, 11, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.46, + 0.785, + 0.514 + ], + "angle": 0, + "content": "23. Wang, Y., Li, K., Li, Y., He, Y., Huang, B., Zhao, Z., Zhang, H., Xu, J., Liu, Y., Wang, Z., et al.: Internvideo: General video foundation models via generative and discriminative learning. arXiv preprint arXiv:2212.03191 (2022) 1, 2, 3, 4, 7, 8, 11, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.516, + 0.785, + 0.544 + ], + "angle": 0, + "content": "24. Wu, Z., Yao, T., Fu, Y., Jiang, Y.G.: Deep learning for video classification and captioning. In: Frontiers of multimedia research, pp. 3-29 (2017) 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.545, + 0.785, + 0.599 + ], + "angle": 0, + "content": "25. Xu, D., Zhao, Z., Xiao, J., Wu, F., Zhang, H., He, X., Zhuang, Y.: Video question answering via gradually refined attention over appearance and motion. In: Proceedings of the 25th ACM international conference on Multimedia. pp. 1645-1653 (2017) 1, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.601, + 0.785, + 0.642 + ], + "angle": 0, + "content": "26. Xu, H., Ghosh, G., Huang, P.Y., Okhonko, D., Aghajanyan, A., Metze, F., Zettlemoyer, L., Feichtenhofer, C.: Videoclip: Contrastive pre-training for zero-shot video-text understanding. arXiv preprint arXiv:2109.14084 (2021) 4, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.644, + 0.785, + 0.712 + ], + "angle": 0, + "content": "27. Xu, H., Ghosh, G., Huang, P.Y., Okhonko, D., Aghajanyan, A., Metze, F., Zettlemoyer, L., Feichtenhofer, C.: VideoCLIP: Contrastive pre-training for zero-shot video-text understanding. In: Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing (EMNLP). Association for Computational Linguistics, Online (Nov 2021) 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.714, + 0.785, + 0.755 + ], + "angle": 0, + "content": "28. Xu, J., Mei, T., Yao, T., Rui, Y.: Msr-vtt: A large video description dataset for bridging video and language. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 5288-5296 (2016) 1, 2, 4, 5, 6, 7, 10, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.757, + 0.785, + 0.797 + ], + "angle": 0, + "content": "29. Yan, S., Zhu, T., Wang, Z., Cao, Y., Zhang, M., Ghosh, S., Wu, Y., Yu, J.: Videotext modeling with zero-shot transfer from contrastive captioners. arXiv preprint arXiv:2212.04979 (2022) 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.799, + 0.785, + 0.84 + ], + "angle": 0, + "content": "30. Zhou, Z., Lei, Y., Zhang, B., Liu, L., Liu, Y.: Zegclip: Towards adapting clip for zero-shot semantic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11175-11185 (2023) 8" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.148, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.335, + 0.114, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Rethinking Video-Text Understanding: RCAD and Feint6K" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.788, + 0.219 + ], + "angle": 0, + "content": "31. Zhu, B., Lin, B., Ning, M., Yan, Y., Cui, J., HongFa, W., Pang, Y., Jiang, W., Zhang, J., Li, Z., Zhang, C.W., Li, Z., Liu, W., Yuan, L.: Languagebind: Extending video-language pretraining to n-modality by language-based semantic alignment. In: The Twelfth International Conference on Learning Representations (2024), https://openreview.net/forum?id=QmZKc7UZCy 11" + } + ] +] \ No newline at end of file diff --git a/2024/Rethinking Video-Text Understanding_ Retrieval from Counterfactually Augmented Data/21cc04b7-ea52-4351-82c2-826a88e983b2_origin.pdf b/2024/Rethinking Video-Text Understanding_ Retrieval from Counterfactually Augmented Data/21cc04b7-ea52-4351-82c2-826a88e983b2_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..ad1262ec4757a45d2c1d528dcd87e3cf99761fc3 --- /dev/null +++ b/2024/Rethinking Video-Text Understanding_ Retrieval from Counterfactually Augmented Data/21cc04b7-ea52-4351-82c2-826a88e983b2_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d11dfe5800a9cf14e0fabbc003e8dc6c171ae4a2bd6dc53bb9977ebc221c999b +size 10999526 diff --git a/2024/Rethinking Video-Text Understanding_ Retrieval from Counterfactually Augmented Data/full.md b/2024/Rethinking Video-Text Understanding_ Retrieval from Counterfactually Augmented Data/full.md new file mode 100644 index 0000000000000000000000000000000000000000..1412517fa0373e4114cb4d3241f00b0fa5bb6cc7 --- /dev/null +++ b/2024/Rethinking Video-Text Understanding_ Retrieval from Counterfactually Augmented Data/full.md @@ -0,0 +1,289 @@ +# Rethinking Video-Text Understanding: Retrieval from Counterfactually Augmented Data + +Wufei Ma $^{1,2\star}$ , Kai Li $^{1}$ , Zhongshi Jiang $^{1}$ , Moustafa Meshry $^{1}$ , Qihao Liu $^{2}$ , Huiyu Wang $^{3}$ , Christian Hane $^{1}$ , and Alan Yuille $^{2}$ + +Meta Reality Labs + +2 Johns Hopkins University + +3 Meta AI + +Abstract. Recent video-text foundation models have demonstrated strong performance on a wide variety of downstream video understanding tasks. Can these video-text models genuinely understand the contents of natural videos? Standard video-text evaluations could be misleading as many questions can be inferred merely from the objects and contexts in a single frame or biases inherent in the datasets. In this paper, we aim to better assess the capabilities of current video-text models and understand their limitations. We propose a novel evaluation task for video-text understanding, namely retrieval from counterfactually augmented data (RCAD), and a new Feint6K dataset. To succeed on our new evaluation task, models must derive a comprehensive understanding of the video from cross-frame reasoning. Analyses show that previous video-text foundation models can be easily fooled by counterfactually augmented data and are far behind human-level performance. In order to narrow the gap between video-text models and human performance on RCAD, we identify a key limitation of current contrastive approaches on video-text data and introduce LLM-teacher, a more effective approach to learn action semantics by leveraging knowledge obtained from a pretrained large language model. Experiments and analyses show that our approach successfully learn more discriminative action embeddings and improves results on Feint6K when applied to multiple video-text models. Our Feint6K dataset and project page is available here. + +Keywords: Video-Text Understanding $\cdot$ Retrieval + +# 1 Introduction + +Video-text foundation models have gained increasing attention due to their simple formulation and strong transferability [12, 16, 23]. By pretraining on web scale video-text datasets, these models demonstrate strong performance across a wide range of downstream tasks, such as video-text retrieval [22, 28] and video question answering [25]. + +As video-text foundation models evolve and achieve increasingly better performance on various benchmarks, we raise the following questions: Can these video-text model truly grasp the semantics of natural videos? Are these models genuinely reaching a level of understanding comparable to humans? Despite the remarkable achievements made in previous works, our study suggests that current video-text models still fall far behind human-level understanding. + +We argue that existing prominent results on standard video-text tasks can be misleading as models largely exploit the shortcuts and biases inherent in the dataset. Many of the questions can be answered by objects or context extracted from a single frame without capturing cross-frame relations. For the examples in Fig. 2a, the video-text alignment can be easily inferred from shortcuts such as "cymbals" or "football". Moreover, the models may utilize biases in the datasets, such as the spurious correlation between "outdoor" and "football". Current evaluations of video-text understanding are compromised by shortcuts and biases, which obscure us from analyzing the limitations of current models. As we proceed from image-text understanding to video-text understanding, we should focus on more challenging semantics in the video domain that require cross-frame reasoning to solve, such as the interaction between person and object or the change of appearances over a sequence of frames (see Fig. 1a). + +In this paper we propose a novel evaluation paradigm of video-text understanding, i.e., retrieval from counterfactually augmented data (RCAD). As demonstrated in Fig. 2b, the goal is to retrieve the only positive caption with matched semantics among "hard" negative captions. Negative captions are counterfactually augmented so that video-text models must derive a holistic understanding of the video sequence for both objects and actions, in order to retrieve the correct caption. As a comparison, negative examples from standard video-to-text retrieval [22, 24, 28] are captions of different videos in the same dataset, often containing different object entities that are easy to distinguish. + +We follow the previous human-in-the-loop system [9] and develop a benchmark dataset for our new RCAD task, i.e., the Feint6K dataset. We test a wide range of public video-text methods with various pretraining strategies on this new benchmark. Two failure examples are visualized in Fig. 1a and quantitative results are summarized in Fig. 1b. Note how state-of-the-art video-text model, InternVideo [23], features a $87.9\%$ rank-1 accuracy on standard video-to-text retrieval but drops by $32.8\%$ (pretraining only) and $29.7\%$ (finetuned) when tested on our benchmark with counterfactually augmented data. By establishing a human-level baseline on our benchmark, the InternVideo model surprisingly falls far behind human performance by $38.6\%$ . Our findings sharply contrast to the prominent performance obtained on previous video-text benchmarks and question the common belief that latest video-text models can develop a fairly effective representation for texts and videos in existing datasets. + +In light of our evaluation results on Feint6K, we identify a significant limitation of the widely adopted contrastive representation learning on video-text data, which is the issue of shortcut learning (see Sec. 4.1). To address this, we propose LLM-teacher that enables a more effective learning of action se + +![](images/3c0eebf4e7165d4fe3da696685d2cb5d5f4da7c6217ea066f34ace79e7dd6847.jpg) +Groundtruth: + $\checkmark$ A person sitting outside with some tools is examining a rock on a leather piece. + +![](images/29b9e994213addfa806d0ae37029d51d199cb91363f3fcf207ffe1da2aa622d9.jpg) +Predicted: + $x$ A person sitting outside with some tools is throwing a rock on a leather piece. +Groundtruth: A person is building a space jet toy at high speed. + +![](images/1c44aa3ac092405d659aecdd7275730fe461e6676e77870a1a2cb5cc0172905c.jpg) +Predicted: + $\times$ A person is coloring a space jet toy at high speed. +(a) Failures of InternVideo [23] on Feint6K +(b) Quantitative results on Feint6K dataset +Fig. 1: (a): Although with large-scale pretraining on web-scale data, current video-text model like [23] can be easily fooled by counterfactually augmented data. (b): The performance of InternVideo on retrieval from counterfactually augmented data (RCAD) drops by over $30\%$ when compared to the standard video-to-text retrieval and by $38.6\%$ when compared to human-level performance. We also evaluate models on standard video-text retrieval from only 6 sampled candidates and show that our RCAD task is indeed much more challenging. Our LLM-teacher successfully improves the performance on RCAD by enforcing a more effective learning of action semantics. + +mantics by introducing extra knowledge from pretrained large language models (LLMs). Specifically we generate synthetic captions by modifying the contents of an existing caption and LLM serves as the teacher to determine if a synthetic caption matches the semantics in the original caption using binary pseudo-labels or continuous logits. This approach enables a more effective learning of action embeddings from the available video-text data. Experimental results show that LLM-teacher learns more discriminative action embeddings from our advanced contrastive objectives and improves the performance of retrieval from counter-factually augmented data when applied to multiple video-text models. + +In summary, our key contributions are as follows. (1) We propose to evaluate video-text models on questions that require cross-frame reasoning and develop a new task, retrieval from counterfactually augmented data, and a new dataset, Feint6K (Sec. 3). (2) Extensive experimental results on our new evaluation paradigm suggest that existing video-text models demonstrate very limited understanding of the action semantics in a video, which is contrary to the prominent performance achieved on standard video-text retrieval benchmarks. (3) From our results on Feint6K dataset, we identify a key shortcoming of contrastive learning approaches on video-text data. We present LLM-teacher to enforce a more effective learning of action embeddings by injecting knowledge from pretrained LLMs (Sec. 4). Our approach effectively improves the results on Feint6K when applied to multiple text-video models. + +# 2 Related Works + +Video-text pretraining. With the availability of web-scale video-text paired datasets, such as WebVid2M [1] and HowTo100M [17], recent developments of + +![](images/c336959575ac3049fd973da454260476b894ece1d94dd44934d3745a55dbe9bc.jpg) +(a) Standard video-to-text retrieval +Fig. 2: Different evaluations of video-text understanding. (a): In standard video-to-text retrieval, negative captions are sampled from different videos in the same dataset. Image-text models can achieve good performance by exploiting shortcuts (e.g., "football" and "cymbals") and biases (e.g., spurious correlation between "outdoor" and "football"). (b): In our proposed RCAD paradigm, we adopt a human-in-the-loop system (see Sec. 3.2) to obtain "hard" negative captions with unchanged object entities but modified actions. Models must develop a holistic understanding of the semantics from the sequence of frames to retrieve the matched caption. + +![](images/8e0b68883a179cd5a7ce1e1fc09f0c0d71cb9e161e78ff68219104f946d93be2.jpg) +(b) RCAD + +video-text pretraining models achieved improved results on a wide range of video understanding tasks. Visual and textual embeddings jointly learned from the large-scale data demonstrate strong transferability and largely reduce the efforts in downstream tasks. Mainstream pretraining objectives can be categorized into discriminative and generative. Discriminative approaches extend the objective of CLIP [18] to the video-text domain and learn multi-modal representations by contrasting between matched and unmatched pairs [15, 16, 23, 26] or simply predicting if a video-text pair is matched [16]. Generative methods follow the masked modeling idea in BERT [5] and adopted masked language modeling (MLM) and masked video modeling (MVM) for video-text pretraining [16, 21, 23]. These generative approaches are often considered superior for action recognition, while discriminative approaches learn better semantics from texts [21]. + +Evaluation of video-text models. In order to analyze the effectiveness of video-text pretraining, previous works [15,16,21,23,26,29] focused on the following tasks: (1) video-text retrieval (zero-shot or finetuned) [22, 28]: retrieving the matched video (or text) given a text (or video) as query where the negative candidates are unmatched pairs from the same dataset, (2) video question answering (finetuned) [25]: predicting the answer with a classification head, (3) video classification (zero-shot or finetuned) [6,10,20]: classifying videos using label names as text prompts, and (4) video captioning (zero-shot or finetuned) [11, 28]: summarizing the contents of a video. + +We argue that results on these video-text tasks could be deceitful as most tasks heavily rely on the alignment of object entities in the video and text + +(which image-text models are also capable of), and “action understanding” can often be developed through shortcuts and biases. For instance, a commonly-used evaluation for “action recognition” is video classification on ActivityNet [6], where a model classifies a given video into classes such as “playing badminton”, “kayaking”, or “volleyball”. However, video-text models would classify the videos by exploiting objects and contexts as shortcuts (e.g., “badminton” and “kayak”), as well as other biases (e.g., spurious correlation between “kayak” and “on the water”) without genuinely understanding the semantics of the action represented by interactions between the person and the objects over time. For instance, classifying the action as “kayaking” is essentially a “kayak” detection problem in ActivityNet. Therefore we propose retrieval from counterfactually augmented data, a new evaluation paradigm where we aim to eliminate the shortcuts from the questions so models must establish a comprehensive understanding of the semantics from cross-frame reasoning in order to predict the correct answer. + +# 3 Retrieval from Counterfactually Augmented Data + +Previous evaluation tasks of video-text models focused on the alignment of feature embeddings of the video-text pairs (e.g., video-text retrieval [22, 28] and video classification [10, 20]) or reconstruction of the semantics with text (e.g., video question answering [25] and video captioning [11, 28]). However, these evaluation tasks are largely limited by the paired data available in existing video-text datasets. As demonstrated by in Fig. 2a, we can extract most of the semantics of a video, such as the objects, people, and contexts in the video, by looking at only one frame from the video. The action in the video can be further inferred by exploiting biases in the datasets. Therefore, models pretrained on image-text data often perform surprisingly well on existing video-text tasks. + +To effectively evaluate how video-text models can understand the contents and semantics of videos beyond images, we must develop video understanding tasks that are free from shortcuts and require cross-frame reasoning to solve. To this end, we propose a new evaluation task for video-text models, i.e., retrieval from counterfactually augmented data (Sec. 3.1), and a new dataset Feint6K (Sec. 3.2). In contrast to previous datasets with matched video-text pairs scrapped from the web [24, 28] or annotated by human [22], we adopt a human-in-the-loop system and annotate counterfactually manipulated texts as negative pairs with the original video. This allows us to evaluate video-text models with novel tasks and better understand their limitations. Lastly we measure human performance on our dataset as a direct comparison with the state-of-the-art video-text models (Sec. 3.3). + +# 3.1 Task Formulation + +Retrieval from counterfactually augmented data is a variant of the standard video-to-text retrieval. As shown in Fig. 2b, given a video sequence and a list of candidate captions, the model is asked to retrieve the caption that best matches + +# Video A + +![](images/f917624d6973ec7537a39238f822e2d9cbcada4184c6e42855970622fc90204a.jpg) +Fig. 3: Examples of $RCAD$ on our Feint6K dataset. + +# Video B + +![](images/b784d2ebffcb678ab1a7ba413eb75cf19f603be00ab0b2f003e25f872a3d3e4d.jpg) + +# Video C + +![](images/adee20fde29321c8721323115aae2802dfc2fe5728d8f8d8839a2efd253409d3.jpg) + +# Captions of Video A + +[Positive] A person is shinning a shoe with a shoe shinning machine. + +[Negative 1] A person is making a shoe with a shoe shimming machine. + +[Negative 2] A person is wearing a shoe with a shoe shinning machine. + +# Captions of Video B + +[Positive] A man is standing on the field attempting to kick a football. + +[Negative 1] A man is standing on the field attempting to catch a football. + +[Negative 2] A man is standing on the field attempting to throw a football. + +# Captions of Video C + +[Positive] A person wearing a pink jacket is feeding birds with his hand. + +[Negative 1] A person wearing a pink jacket is petting birds with his hand. + +[Negative 2] A person wearing a pink jacket is holding birds with his hand. + +the semantics in the video. However, unlike standard video-to-text retrieval where negative captions come from other video-text pairs in the same dataset, negative captions in our task are modified from the positive captions, with the same text structure and object entities but different actions (see Fig. 3). + +Similar to video-text retrieval, our proposed task provides the benefit of zero-shot evaluation. This allows us to quantitatively analyze the effectiveness of video-text pretraining without downstream finetuning. + +What distinguishes our task from all previous video-text tasks is its focus on questions that require cross-frame reasoning. Consider "Video B" in Fig. 3 as an example. All captions contain the same object entities present in the video, i.e., "a man", "the field", and "a football", but differ in the action specified. We cannot determine the most accurate caption because all suggested actions, "kick", "catch", and "throw", are plausible given the video's context. To retrieve the corresponding caption, it is crucial to grasp the action's semantics by inspecting the interactions between the man and the football across a series of frames. This underscores the necessity of understanding the video semantics over time in order to succeed in our task, rather than relying shortcuts or biases. + +# 3.2 Feint6K: Data Collection + +We utilize a human-in-the-loop system [9] to counterfactually manipulate the positive captions in existing video-text datasets [22, 28]. The generated counterfactually augmented data, when paired with the original videos, forms negative pairs to be used in our proposed task. We recruit 40 annotators to manually manipulate the actions in existing captions. Specifically, the new actions introduced must be plausible within the context in the caption but are not occurring in the corresponding video. In summary, a total of 6,243 videos from the validation + +![](images/2e45d571385c4daed4a61b7268a9f80b13fcbbc644dcfb1e49b93b565c82733c.jpg) +Fig. 4: Overview of our data collection pipeline for Feint6K dataset featuring a human-in-the-loop system. + +set of the MSR-VTT dataset [28] and the test set of the VATEX dataset [22]. For detailed statistics of our Feint6K dataset, please refer to our supplementary materials. + +The overview of our human-in-the-loop system is demonstrated in Fig. 4. We start by sampling a small subset of video-text pairs and manually annotate the counterfactually augmented captions for the purpose of training and demonstration. Specific guidance and feedback are given to the annotators on the practice questions during the training stage. Moreover, we adopt a reviewing process during the annotation stage to further ensure the quality of our annotated data. Each annotation is reviewed before acceptance, where rejected annotations are marked and sent back for refinement. We refer the readers to the supplementary materials regarding the annotation interface and annotator guidance. + +# 3.3 Human Performance + +We employ the same group of annotators to establish a human-level baseline on our Feint6K dataset. Analyzing the human performance serves two goals: (i) verifying the legitimacy of our Feint6K dataset – if each question is answerable and has one unique answer, and (ii) comparing the state-of-the-art video-text models with human-level performance. + +Each annotator is first presented with the full video and then asked to select the one out of six captions that best matches the semantics in the video. To avoid leakage, we make sure the annotators assigned to the testing questions are not the same annotators annotating this video. We also randomly shuffle the order of the candidate captions to remove any biases. In Sec. 5.2 we present the quantitative comparisons between human-level performance and current state-of-the-art video-text models on our Feint6K dataset. + +# 4 LLM-teacher + +Contrastive learning are widely adopted in previous video-text representation learning works [16, 23, 26]. The contrastive models learn to distinguish between + +similar and dissimilar pairs of video-text data by contrasting their feature representations. Despite the simple formulation and great generalization ability, our results on Feint6K show that current video-text models built heavily on contrastive objectives have very limited understanding of action semantics in natural videos and can be easily fooled by counterfactually augmented data. Therefore we aim to develop a more effective contrastive approach and learn more discriminative action representations from existing video-text data. + +Sec. 4.1 discusses the limitations of current contrastive approaches due to numerous shortcuts in video-text dataset. Inspired by this finding, we present LLM-teacher in Sec. 4.2, a simple but effective approach to learn more discriminative action embeddings by incorporating knowledge from pretrained large language models. + +# 4.1 Shortcuts in Video-Text Data + +The CLIP model [18] trained on image-text data with a contrastive objective demonstrated strong zero-shot capabilities in many vision tasks [13,30] and was widely adopted as the vision encoder in many large vision-language models [14]. However, directly generalizing this idea to video-text data may face unprecedented challenges due to shortcuts in the video-text data. + +As we proceed from image-text to video-text representation learning, we aim to learn a powerful embedding for not only the object entities, but also the actions specified by texts or sequence of frames. Naturally we hope to achieve this by contrasting videos with different actions. However, as we are optimizing the contrastive objectives in a mini-batch of videos, object entities become the shortcuts that contrastive models exploit to saturate the contrastive objective, leading to an ineffective action embeddings. Consider the example in Fig. 2a, the contrastive loss would be small as long as the model learns discriminative embeddings for shortcut objects such as "cymbal" and "football", even without any understanding of the actions. Since many video-text models start with pretrained CLIP encoders [23,26], they often possess strong embeddings for shortcut objects from the beginning of the video-text pretraining, which hinders the model from further learning effective action representations. + +In Fig. 5 we present quantitative results of changes in cosine similarities when the video is unchanged but the object or action in the caption is manipulated. We find that state-of-the-art video-text models learn less discriminative embeddings for actions as compared to objects, which also explains the small gap between image-text and video-text models on standard video-text tasks. + +# 4.2 LLM-teacher + +We introduce LLM-teacher, an LLM-powered approach to learn better action representations with contrastive objectives. Specifically, LLM serves as the "teacher" and provides synthesized captions as extra knowledge for the video-text model to learn from. These new captions serve as negative captions of the original video, + +![](images/49fff5b43c9cdadc257ceae8112a25ec80fe30eb7d114c6ddecce79a0869f22a.jpg) +(a) Comparison between $\Delta s$ w.r.t. object or action changes. + +![](images/c89c032b08ca7476da73e19c348b841d3f223e885edbc7f98ffb0aa738177e97.jpg) +(b) Comparison between $\Delta s$ w.r.t. action using InternVideo and our LLM-teacher. +Fig. 5: Change of cosine similarity w.r.t. objects or actions. (a): Comparison between the change in cosine similarity when the action or object is swapped. Results show that current video-text models learn a more effective embedding for objects than for actions. (b): Comparison between the change in cosine similarity using InternVideo or our LLM-teacher. This demonstrate that LLM-teacher learns a more discriminative embedding for actions by enabling a more effective contrastive learning using knowledge from LLMs. Refer to Sec. 5.2 for more details. + +removing possible shortcuts and biases and enabling a more effective learning of action semantics. + +We start with positive captions from common video-text datasets and run an abstract meaning representation (AMR) parser, resulting in a list of action or object tokens. To generate the hard negative captions, two methods are considered to find novel actions or objects given the original caption and the substitution token $w$ . + +Method I: Mask filling. Masked language modeling (MLM) [5] is a widely adopted self-supervised task for LLMs. In the pretraining stage, a certain percentage of word tokens are substituted with a special token [mask] and the LLM is optimized to recover the masked tokens. To find negative captions with novel actions given the original caption $c_{0}$ and substitution token $a$ , we first substitute the token $a$ with [mask] and use a LLM pretrained with MLM to predict a list of $k$ possible tokens $\{\hat{w}_i\}_{i=1}^k$ such that $\hat{w}_i \neq w$ and $\hat{w}_i$ is an action token. + +Method II: LLM-powered chatbot. Although mask filling can find appropriate words to substitute $w$ , it is limited by one single token. In many cases, we want to update the prepositions following the change of verbs. For example, when "install" is substituted by "uninstall", we should also update the preposition from "install ... to" to "uninstall ... from". Here we consider a more flexible approach that builds on LLM-powered chatbots. Besides a prompt describing the text substitution, we leverage the in-context learning ability of LLMs [2, 7] and provide multiple in-context examples to obtain the desirable negative captions. + +LLM-teacher. With the procedure above, we obtain $k$ synthesized negative captions $\{n_i\}_{i=1}^k$ besides the original positive caption $p$ for each video $v$ in existing datasets [22,28]. For each of the original and generated captions, LLM serves as a "teacher" and provide binary pseudo-labels - if the caption have matched ("1") or unmatched ("0") semantics with the video. Now we may optimize the video-text model with the normalized temperature-scaled cross entropy loss + +$$ +l = - \log \frac {\exp \left(\sin \left(f _ {v} , f _ {p}\right) / \tau\right)}{\exp \left(\sin \left(f _ {v} , f _ {p}\right) / \tau\right) + \sum_ {i = 1} ^ {k} \exp \left(\sin \left(f _ {v} , f _ {n _ {i}}\right) / \tau\right)} \tag {1} +$$ + +where $f_{v}, f_{p}, f_{ni}$ are the visual and textual embeddings. + +In practice, we find that certain negative captions have unmatched but similar semantics to the video. It is undesirable to regard them as a strictly negative pair in the contrastive loss. Therefore, we extend the binary pseudo-labels to soft logits by computing caption similarities with a pretrained LLM and then optimize the video-text model to match model outputs with the soft logits from the LLM teacher model [8]. + +$$ +l = \mathcal {L} _ {\mathrm {K L}} \left(z _ {\text {v i d e o - t e x t}}, z _ {\text {L L M}}\right) \tag {2} +$$ + +$$ +z _ {\mathrm {v i d e o - t e x t , t}} = \frac {\exp (\sin (f _ {v} , f _ {t}) / \tau)}{\exp (\sin (f _ {v} , f _ {p}) / \tau) + \sum_ {i = 1} ^ {k} \exp (\sin (f _ {v} , f _ {n _ {i}}) / \tau)} +$$ + +$$ +z _ {\mathrm {L L M , t}} = \frac {\exp (\sin (e _ {p} , e _ {t}) / \tau)}{\exp (\sin (e _ {p} , e _ {p}) / \tau) + \sum_ {i = 1} ^ {k} \exp (\sin (e _ {p} , e _ {n _ {i}}) / \tau)} +$$ + +where $t \in \{n_i\}_{i=1}^k \cup \{p\}$ , $f$ 's are textual and visual embeddings of the video-text model, and $e$ 's are textual embeddings of a pretrained LLM. + +# 5 Experiments + +# 5.1 Experimental Setup + +Datasets. We use MSR-VTT dataset [28] and VATEX dataset [22] for standard video-to-text retrieval and our Feint6K dataset for retrieval from counter-factually augmented data. Note that Feint6K dataset contains the same video sequences for evaluation as in MSR-VTT dataset and VATEX dataset. + +Evaluation metrics. For standard video-to-text retrieval we report rank-1 accuracy $(\mathbb{R}@\mathbb{1})$ as this is the most challenging metric and distinguish the performance of previous video-text models. For retrieval from counterfactually-augmented data we use rank-1 accuracy $(\mathbb{R}@\mathbb{1})$ , rank-2 accuracy $(\mathbb{R}@\mathbb{2})$ , and mean rank (MeanR). + +Video-text models. We consider multiple state-of-the-art public models in this work. CLIP4Clip [15] extends CLIP model [18] to the video domain using a + +
ModelMSR-VTTFeint6K (MSR-VTT)VATEXFeint6K (VATEX)
R@1R@1R@2MeanRR@1R@1R@2MeanR
Human95.296.8
Random<1e-316.7<1e-316.7
Zero-Shot
CLIP [18]26.337.355.32.638.834.854.32.7
VideoCLIP [27]14.535.171.32.613.733.070.72.7
InternVideo [23]37.545.863.62.376.944.163.92.3
LanguageBind [31]42.841.377.02.442.377.02.4
Finetuned
LanguageBind [31]
CLIP4Clip [15]43.150.872.42.0
VindLU [3]46.653.470.92.0
SimVTP [16]50.235.770.82.676.633.668.42.6
w/ LLM-teacher-lbl49.0 ↓1.240.0 ↑4.373.1 ↑2.32.3 ↓0.374.8 ↓1.837.3 ↑3.772.2 ↑3.82.4 ↓0.2
w/ LLM-teacher-lgt49.5 ↓0.743.5 ↑7.875.0 ↑4.22.2 ↓0.475.3 ↓1.340.1 ↑6.573.5 ↑5.12.3 ↓0.3
InternVideo [23]49.158.680.21.887.958.276.91.9
w/ LLM-teacher-lbl48.2 ↓0.964.2 ↑5.682.5 ↑2.31.7 ↓0.185.2 ↓2.763.8 ↑5.680.5 ↑3.61.7 ↓0.2
w/ LLM-teacher-lgt48.9 ↓0.265.8 ↑7.283.8 ↑3.61.7 ↓0.187.3 ↓0.665.6 ↑7.481.7 ↑4.81.7 ↓0.2
+ +Table 1: Performance of standard video-to-text retrieval on MSR-VTT dataset [28] and VATEX dataset [22], as compared to performance of retrieval from counterfactually-augmented data on our Feint6K dataset. In the "zero-shot" setting, models are evaluated directly after the pretraining stage, while in the "finetuned" setting, models are finetuned on the training set of MSR-VTT or VATEX. With LLM-teacher, we enable a more efficient learning of action semantics and effectively improves R@1 accuracy on Feint6K. Here "LLM-teacher-lbl" stands for our approach with binary pseudo-labels and "LLM-teacher-lgt" uses soft logits. + +mean pooling mechanism for zero-shot video-text retrieval and a contrastive loss for finetuning. SimVTP [16] is pretrained on the WebVid2M dataset [1] with a combination of contrastive learning and masked modeling. InternVideo [23] is a video-text foundation model pretrained on a combination of 7 large-scale videotext datasets, with a supervised video post-pretraining for better video recognition. LanguageBind [31] is a language-based multi-modal pretraining model with remarkable performance on a wide range of benchmarks. + +To demonstrate the efficacy of our LLM-teacher, we adopt this method on two state-of-the-art video-text models, SimVTP and InternVideo. In the default setting we generate 10 action-based captions with LLM for each training video. For text generation we use a pretrained XLM-RoBERTa [4] for mask filling. To measure text similarities and compute soft logits, we utilize a pretrained Sentence-BERT model [19]. In Sec. 5.3 we ablate on the choices of parameters and caption generation. + +# 5.2 Main Results + +Performance of previous state-of-the-art. In Tab. 1 we report the quantitative results of standard video-text retrieval on MSR-VTT and VATEX, and of + +retrieval from counterfactually augmented data on our Feint6K. We also report a "Random" performance where a model predicts random guesses. We make the following observations: (i) Results show that previous state-of-the-art video-text models demonstrate limited understanding of the action semantics in a video with less than $60\%$ R@1 accuracy on RCAD, given a $16.7\%$ R@1 accuracy when taking random guesses. (ii) Although it is hard to compare the quantitative results between two tasks, we note that our RCAD is much more challenging given the performance gap between "Random" and previous state-of-the-arts. This supports our previous arguments that our RCAD removes shortcuts from standard video-text datasets and focuses on harder questions in the video domain that require cross-frame reasoning. + +Performance of LLM-teacher. We apply our LLM-teacher to two pretrained video-text models, SimVTP [16] and InternVideo [23]. Specifically, we consider the two objectives in Eq. (1) and Eq. (2), labeled as "LLM-teacher-lbl" and "LLM-teacher-lgt" respectively. Results in Tab. 1 show that by exploiting knowledge from pretrained LLMs, LLM-teacher enables a more efficient leaning of action semantics and achieved improved results on various metrics of RCAD, increasing the R@1 accuracy on RCAD by $7.2\%$ and $7.4\%$ . We also note that with binary pseudo-labels, results demonstrate a trade-off between the performance on standard video-to-text retrieval and RCAD; and in comparison, LLM-teacher with soft logits achieves the highest performance on RCAD ( $65.8\%$ and $65.6\%$ ) with negligible drops ( $0.2\%$ and $0.6\%$ ) on standard retrieval. This is because soft logits account for the text similarities between different actions and avoid overfitting on binary pseudo-labels. + +Comparison with human-level performance. The human-level performance are reported in Tab. 1. We see that human annotators achieved an almost perfect performance on Feint6K dataset, with a $95.2\%$ accuracy on MSR-VTT videos and a $96.8\%$ accuracy on VATEX videos. These results show that: (i) most questions for retrieval from counterfactually-augmented data are answerable with an unique answer, and (ii) this task is fairly simple for human. As a comparison, all previous video-text models with large-scale pretraining fall behind by a wide margin. Specifically, the state-of-the-art video-text model InternVideo [23] features a heavy two-stage pretraining on a collection of 7 large-scale video-text datasets but falls behind by $36.6\%$ and $38.6\%$ even with downstream finetuning. + +Analyses of failure cases. In Fig. 1b we present two failure cases by the InternVideo [23] model on our Feint6K dataset. Both examples are trivial for human, yet challenging from video-text models due to incapability of cross-frame reasoning. For the first example, a model must analyze the interactions between the person and the rock over a sequence of frames to predict the correct answer. And in the second example, the model must observe the change of appearances over time. As we can see, with the counterfactually augmented data in our Feint6K, + +models cannot "guess" the correct answer by exploiting shortcuts extracted from a single frame. The findings on Feint6K highlight significant weaknesses in current video-text models, offering valuable insights for future researches in this area. Additional qualitative examples of video-text models on Feint6K dataset are provided in the supplementary materials for readers' reference. + +Sensitivity to the change of object or action. With the counterfactually augmented data collected, we investigate how the cosine similarities between the video and text change when the object or action in the text are swapped with an object or action not present in the image. Given the original video-text pair $(v,t)$ and a counterfactually-augmented text $\hat{t}$ , the change in cosine similarity is given by $\Delta s = s(v,\hat{t}) - s(v,t)$ . As $\hat{t}$ contains objects or actions not present in the video, $\Delta s$ should be negative for an ideal video-text model. Moreover, larger $|\Delta s|$ implies the model being more sensitive to the changes. We compute $\Delta s$ using the InternVideo with pretraining only ("IntVid-pt"), InternVideo with downstream finetuning ("IntVid-ft"), and our LLM-teacher ("LLM-teacher"). + +Results in Fig. 5a show that the $\Delta s$ are almost always negative when the objects are swapped and the absolute changes are larger. In comparison, $\Delta s$ are sometimes positive when actions are modified and the absolute changes are smaller. This demonstrate that InternVideo learns a more effective embedding for objects than for actions. This further highlights the necessity of our new evaluation paradigm for exploring the limitations of video-text models beyond existing tasks. Moreover, in Fig. 5b we note that with LLM-teacher the action embeddings are more effective than the ones learned by InternVideo, as shown by the more reasonable changes in cosine similarity when actions are swapped. + +Please refer to the supplementary materials for a detailed discussion regarding the influence of textual encoders. + +# 5.3 Ablation Study + +We conduct ablation studies on the VATEX dataset [22] and our Feint6K dataset. We follow the same settings as above and report rank-1 (R@1) accuracy for standard video-to-text retrieval and rank-1 accuracy (R@1), rank-2 accuracy (R@2), and mean rank (MeanR) for counterfactual augmented data. + +Choice of captions in LLM-teacher. Tab. 2 shows the comparison between different numbers and types of LLM captions generated. In the "default" setting we use 10 action-based LLM captions and compare to settings with 5 action-based LLM captions or 5 action-based and 5 object-based LLM captions. Results show with 5 action-based LLM captions the R@1 accuracy on Feint6K drops by $0.9\%$ while the R@1 accuracy on VATEX increases by $0.3\%$ . We also experiment on object-based LLM captions and find them not beneficial for both standard video-to-text retrieval or our RCAD. This is consistent with our assumption that pretrained video-text models already learns a discriminative embedding for objects and a more efficient training objective for action is needed. + +
ModelVATEXFeint6K (VATEX)
R@1R@1R@2MeanR
Default(10 action captions; XLM-RoBERTa)87.365.681.71.7
5 action captions87.6 ↑0.364.7 ↓0.981.0 ↓0.71.70.0
5 object + 5 object captions87.5 ↑0.264.2 ↓1.480.6 ↓1.11.70.0
LLM Chatbot87.0 ↓0.365.9 ↑0.381.8 ↑0.11.70.0
+ +Table 2: Ablation studies on the number of LLM captions and caption generation methods. In the default setting we use a total of 10 action-based LLM captions and utilize a pretrained XLM-RoBERTa [4] for mask filling. + +Choice of caption generation. Besides a pretrained XLM-RoBERTa model for mask filling, we also experiment on generating captions with LLM-powered agents that are finetuned on chat datasets for dialogue applications. Specifically we use the "chat" model finetuned for dialogue applications. Empirically we find that captions generated by LLM chatbots achieve a higher overall quality, exploring a more diverse caption space and better flexibility. However we don't observe significant improvements from our ablation study experiments – the R@1 accuracy on VATEX drops by $0.3\%$ and the R@1 accuracy on Feint6K increases by $0.3\%$ . We choose XLM-RoBERTa in our main experiments as the model runs faster and scale up easily to bigger settings. We provide qualitative comparisons between the two types of caption generation in our supplementary materials. + +# 6 Conclusions + +In this work we propose a new evaluation task, retrieval from counterfactually augmented data, and a benchmark dataset, Feint6K. The idea is to remove shortcuts in the video-text questions with a human-in-the-loop system and produce more challenging questions where the video-text model must derive a comprehensive understanding of the video with cross-frame reasoning. Quantitative and qualitative evaluation results on our Feint6K dataset show that despite the task is trivial for human, previous state-of-the-art video-text models can be easily fooled by the counterfactually augmented data. This implies that the prominent results on previous video-text benchmarks could be misleading - models may largely exploit shortcuts without genuinely understand the video contents. Moreover, we identify a key limitation of current contrastive learning on videotext data being the shortcut learning and propose LLM-teacher that enables a more effective learning of action semantics by utilizing knowledge from pretrained LLMs. Experimental results and analyses show that our method can learn a more discriminative representation for actions. + +Supplementary materials. We present the following: (1) limitations of our work, (2) more details about our Feint6K dataset, (3) ethics statements, and (4) extra quantitative and qualitative results. + +Acknowledgements. We would like to thank Daniel Khashabi, Yiyan Li, and the anonymous reviewers for their helpful comments and suggestions. Wufei Ma is supported by ONR with N00014-23-1-2641. + +# References + +1. Bain, M., Nagrani, A., Varol, G., Zisserman, A.: Frozen in time: A joint video and image encoder for end-to-end retrieval. In: IEEE International Conference on Computer Vision (2021) 3, 11 +2. Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J.D., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., et al.: Language models are few-shot learners. Advances in neural information processing systems 33, 1877-1901 (2020) 9 +3. Cheng, F., Wang, X., Lei, J., Crandall, D., Bansal, M., Bertasius, G.: Vindlu: A recipe for effective video-and-language pretraining. arXiv preprint arXiv:2212.05051 (2022) 11 +4. Conneau, A., Khandelwal, K., Goyal, N., Chaudhary, V., Wenzek, G., Guzmán, F., Grave, E., Ott, M., Zettlemoyer, L., Stoyanov, V.: Unsupervised cross-lingual representation learning at scale. arXiv preprint arXiv:1911.02116 (2019) 11, 14 +5. Devlin, J., Chang, M.W., Lee, K., Toutanova, K.: Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805 (2018) 4, 9 +6. Fabian Caba Heilbron, Victor Escorcia, B.G., Niebles, J.C.: Activitynet: A large-scale video benchmark for human activity understanding. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 961-970 (2015) 4, 5 +7. Gupta, T., Kembhavi, A.: Visual programming: Compositional visual reasoning without training. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14953-14962 (2023) 9 +8. Hinton, G., Vinyals, O., Dean, J.: Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531 (2015) 10 +9. Kaushik, D., Hovy, E., Lipton, Z.C.: Learning the difference that makes a difference with counterfactually-augmented data. arXiv preprint arXiv:1909.12434 (2019) 2, 6 +0. Kay, W., Carreira, J., Simonyan, K., Zhang, B., Hillier, C., Vijayanarasimhan, S., Viola, F., Green, T., Back, T., Natsev, P., et al.: The kinetics human action video dataset. arXiv preprint arXiv:1705.06950 (2017) 4, 5 +1. Krishna, R., Hata, K., Ren, F., Fei-Fei, L., Carlos Niebles, J.: Dense-captioning events in videos. In: Proceedings of the IEEE international conference on computer vision. pp. 706-715 (2017) 4, 5 +2. Li, K., Wang, Y., Li, Y., Wang, Y., He, Y., Wang, L., Qiao, Y.: Unmasked teacher: Towards training-efficient video foundation models. arXiv preprint arXiv:2303.16058 (2023) 1 +3. Lin, J., Gong, S.: Gridclip: One-stage object detection by grid-level clip representation learning. arXiv preprint arXiv:2303.09252 (2023) 8 +4. Liu, H., Li, C., Wu, Q., Lee, Y.J.: Visual instruction tuning. In: NeurIPS (2023) 8 +5. Luo, H., Ji, L., Zhong, M., Chen, Y., Lei, W., Duan, N., Li, T.: Clip4clip: An empirical study of clip for end to end video clip retrieval and captioning. Neurocomputing 508, 293-304 (2022) 4, 10, 11 + +16. Ma, Y., Yang, T., Shan, Y., Li, X.: Simvtp: Simple video text pre-training with masked autoencoders. arXiv preprint arXiv:2212.03490 (2022) 1, 4, 7, 11, 12 +17. Miech, A., Zhukov, D., Alayrac, J.B., Tapaswi, M., Laptev, I., Sivic, J.: HowTo100M: Learning a Text-Video Embedding by Watching Hundred Million Narrated Video Clips. In: ICCV (2019) 3 +18. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021) 4, 8, 10, 11 +19. Reimers, N., Gurevych, I.: Sentence-bert: Sentence embeddings using siamese bert-networks. In: Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing. Association for Computational Linguistics (11 2019), https://arxiv.org/abs/1908.10084 11 +20. Soomro, K., Zamir, A.R., Shah, M.: Ucf101: A dataset of 101 human actions classes from videos in the wild. arXiv preprint arXiv:1212.0402 (2012) 4, 5 +21. Tong, Z., Song, Y., Wang, J., Wang, L.: Videomae: Masked autoencoders are data-efficient learners for self-supervised video pre-training. Advances in neural information processing systems 35, 10078-10093 (2022) 4 +22. Wang, X., Wu, J., Chen, J., Li, L., Wang, Y.F., Wang, W.Y.: Latex: A large-scale, high-quality multilingual dataset for video-and-language research. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 4581-4591 (2019) 1, 2, 4, 5, 6, 7, 10, 11, 13 +23. Wang, Y., Li, K., Li, Y., He, Y., Huang, B., Zhao, Z., Zhang, H., Xu, J., Liu, Y., Wang, Z., et al.: Internvideo: General video foundation models via generative and discriminative learning. arXiv preprint arXiv:2212.03191 (2022) 1, 2, 3, 4, 7, 8, 11, 12 +24. Wu, Z., Yao, T., Fu, Y., Jiang, Y.G.: Deep learning for video classification and captioning. In: Frontiers of multimedia research, pp. 3-29 (2017) 2, 5 +25. Xu, D., Zhao, Z., Xiao, J., Wu, F., Zhang, H., He, X., Zhuang, Y.: Video question answering via gradually refined attention over appearance and motion. In: Proceedings of the 25th ACM international conference on Multimedia. pp. 1645-1653 (2017) 1, 4, 5 +26. Xu, H., Ghosh, G., Huang, P.Y., Okhonko, D., Aghajanyan, A., Metze, F., Zettlemoyer, L., Feichtenhofer, C.: Videoclip: Contrastive pre-training for zero-shot video-text understanding. arXiv preprint arXiv:2109.14084 (2021) 4, 7, 8 +27. Xu, H., Ghosh, G., Huang, P.Y., Okhonko, D., Aghajanyan, A., Metze, F., Zettlemoyer, L., Feichtenhofer, C.: VideoCLIP: Contrastive pre-training for zero-shot video-text understanding. In: Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing (EMNLP). Association for Computational Linguistics, Online (Nov 2021) 11 +28. Xu, J., Mei, T., Yao, T., Rui, Y.: Msr-vtt: A large video description dataset for bridging video and language. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 5288-5296 (2016) 1, 2, 4, 5, 6, 7, 10, 11 +29. Yan, S., Zhu, T., Wang, Z., Cao, Y., Zhang, M., Ghosh, S., Wu, Y., Yu, J.: Videotext modeling with zero-shot transfer from contrastive captioners. arXiv preprint arXiv:2212.04979 (2022) 4 +30. Zhou, Z., Lei, Y., Zhang, B., Liu, L., Liu, Y.: Zegclip: Towards adapting clip for zero-shot semantic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11175-11185 (2023) 8 + +31. Zhu, B., Lin, B., Ning, M., Yan, Y., Cui, J., HongFa, W., Pang, Y., Jiang, W., Zhang, J., Li, Z., Zhang, C.W., Li, Z., Liu, W., Yuan, L.: Languagebind: Extending video-language pretraining to n-modality by language-based semantic alignment. In: The Twelfth International Conference on Learning Representations (2024), https://openreview.net/forum?id=QmZKc7UZCy 11 \ No newline at end of file diff --git a/2024/Rethinking Video-Text Understanding_ Retrieval from Counterfactually Augmented Data/images.zip b/2024/Rethinking Video-Text Understanding_ Retrieval from Counterfactually Augmented Data/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..fe6569f311f45a5889ef6ae966932fe2c8f0d0a8 --- /dev/null +++ b/2024/Rethinking Video-Text Understanding_ Retrieval from Counterfactually Augmented Data/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:19f3e94266c605a2e9ca4ec1f430c114332b145a1c000c70b480b5ecb60efed9 +size 310343 diff --git a/2024/Rethinking Video-Text Understanding_ Retrieval from Counterfactually Augmented Data/layout.json b/2024/Rethinking Video-Text Understanding_ Retrieval from Counterfactually Augmented Data/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..031b8fe1b07e34c75a93404fd3857b3e3b671ea0 --- /dev/null +++ b/2024/Rethinking Video-Text Understanding_ Retrieval from Counterfactually Augmented Data/layout.json @@ -0,0 +1,8095 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 133, + 111, + 482, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 111, + 482, + 148 + ], + "spans": [ + { + "bbox": [ + 133, + 111, + 482, + 148 + ], + "type": "text", + "content": "Rethinking Video-Text Understanding: Retrieval from Counterfactually Augmented Data" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 148, + 167, + 465, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 148, + 167, + 465, + 194 + ], + "spans": [ + { + "bbox": [ + 148, + 167, + 465, + 194 + ], + "type": "text", + "content": "Wufei Ma" + }, + { + "bbox": [ + 148, + 167, + 465, + 194 + ], + "type": "inline_equation", + "content": "^{1,2\\star}" + }, + { + "bbox": [ + 148, + 167, + 465, + 194 + ], + "type": "text", + "content": ", Kai Li" + }, + { + "bbox": [ + 148, + 167, + 465, + 194 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 148, + 167, + 465, + 194 + ], + "type": "text", + "content": ", Zhongshi Jiang" + }, + { + "bbox": [ + 148, + 167, + 465, + 194 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 148, + 167, + 465, + 194 + ], + "type": "text", + "content": ", Moustafa Meshry" + }, + { + "bbox": [ + 148, + 167, + 465, + 194 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 148, + 167, + 465, + 194 + ], + "type": "text", + "content": ", Qihao Liu" + }, + { + "bbox": [ + 148, + 167, + 465, + 194 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 148, + 167, + 465, + 194 + ], + "type": "text", + "content": ", Huiyu Wang" + }, + { + "bbox": [ + 148, + 167, + 465, + 194 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 148, + 167, + 465, + 194 + ], + "type": "text", + "content": ", Christian Hane" + }, + { + "bbox": [ + 148, + 167, + 465, + 194 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 148, + 167, + 465, + 194 + ], + "type": "text", + "content": ", and Alan Yuille" + }, + { + "bbox": [ + 148, + 167, + 465, + 194 + ], + "type": "inline_equation", + "content": "^{2}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 264, + 201, + 350, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 264, + 201, + 350, + 213 + ], + "spans": [ + { + "bbox": [ + 264, + 201, + 350, + 213 + ], + "type": "text", + "content": "Meta Reality Labs" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 249, + 213, + 364, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 249, + 213, + 364, + 224 + ], + "spans": [ + { + "bbox": [ + 249, + 213, + 364, + 224 + ], + "type": "text", + "content": "2 Johns Hopkins University" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 284, + 224, + 329, + 234 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 284, + 224, + 329, + 234 + ], + "spans": [ + { + "bbox": [ + 284, + 224, + 329, + 234 + ], + "type": "text", + "content": "3 Meta AI" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 160, + 269, + 462, + 511 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 269, + 462, + 511 + ], + "spans": [ + { + "bbox": [ + 160, + 269, + 462, + 511 + ], + "type": "text", + "content": "Abstract. Recent video-text foundation models have demonstrated strong performance on a wide variety of downstream video understanding tasks. Can these video-text models genuinely understand the contents of natural videos? Standard video-text evaluations could be misleading as many questions can be inferred merely from the objects and contexts in a single frame or biases inherent in the datasets. In this paper, we aim to better assess the capabilities of current video-text models and understand their limitations. We propose a novel evaluation task for video-text understanding, namely retrieval from counterfactually augmented data (RCAD), and a new Feint6K dataset. To succeed on our new evaluation task, models must derive a comprehensive understanding of the video from cross-frame reasoning. Analyses show that previous video-text foundation models can be easily fooled by counterfactually augmented data and are far behind human-level performance. In order to narrow the gap between video-text models and human performance on RCAD, we identify a key limitation of current contrastive approaches on video-text data and introduce LLM-teacher, a more effective approach to learn action semantics by leveraging knowledge obtained from a pretrained large language model. Experiments and analyses show that our approach successfully learn more discriminative action embeddings and improves results on Feint6K when applied to multiple video-text models. Our Feint6K dataset and project page is available here." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 160, + 522, + 370, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 522, + 370, + 534 + ], + "spans": [ + { + "bbox": [ + 160, + 522, + 370, + 534 + ], + "type": "text", + "content": "Keywords: Video-Text Understanding " + }, + { + "bbox": [ + 160, + 522, + 370, + 534 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 522, + 370, + 534 + ], + "type": "text", + "content": " Retrieval" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 556, + 230, + 569 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 556, + 230, + 569 + ], + "spans": [ + { + "bbox": [ + 132, + 556, + 230, + 569 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 584, + 482, + 645 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 584, + 482, + 645 + ], + "spans": [ + { + "bbox": [ + 130, + 584, + 482, + 645 + ], + "type": "text", + "content": "Video-text foundation models have gained increasing attention due to their simple formulation and strong transferability [12, 16, 23]. By pretraining on web scale video-text datasets, these models demonstrate strong performance across a wide range of downstream tasks, such as video-text retrieval [22, 28] and video question answering [25]." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 653, + 389, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 653, + 389, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 653, + 389, + 665 + ], + "type": "text", + "content": "* All data collection and experiments were conducted at JHU." + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "text", + "content": "As video-text foundation models evolve and achieve increasingly better performance on various benchmarks, we raise the following questions: Can these video-text model truly grasp the semantics of natural videos? Are these models genuinely reaching a level of understanding comparable to humans? Despite the remarkable achievements made in previous works, our study suggests that current video-text models still fall far behind human-level understanding." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 190, + 483, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 190, + 483, + 346 + ], + "spans": [ + { + "bbox": [ + 130, + 190, + 483, + 346 + ], + "type": "text", + "content": "We argue that existing prominent results on standard video-text tasks can be misleading as models largely exploit the shortcuts and biases inherent in the dataset. Many of the questions can be answered by objects or context extracted from a single frame without capturing cross-frame relations. For the examples in Fig. 2a, the video-text alignment can be easily inferred from shortcuts such as \"cymbals\" or \"football\". Moreover, the models may utilize biases in the datasets, such as the spurious correlation between \"outdoor\" and \"football\". Current evaluations of video-text understanding are compromised by shortcuts and biases, which obscure us from analyzing the limitations of current models. As we proceed from image-text understanding to video-text understanding, we should focus on more challenging semantics in the video domain that require cross-frame reasoning to solve, such as the interaction between person and object or the change of appearances over a sequence of frames (see Fig. 1a)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 349, + 483, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 349, + 483, + 456 + ], + "spans": [ + { + "bbox": [ + 130, + 349, + 483, + 456 + ], + "type": "text", + "content": "In this paper we propose a novel evaluation paradigm of video-text understanding, i.e., retrieval from counterfactually augmented data (RCAD). As demonstrated in Fig. 2b, the goal is to retrieve the only positive caption with matched semantics among \"hard\" negative captions. Negative captions are counterfactually augmented so that video-text models must derive a holistic understanding of the video sequence for both objects and actions, in order to retrieve the correct caption. As a comparison, negative examples from standard video-to-text retrieval [22, 24, 28] are captions of different videos in the same dataset, often containing different object entities that are easy to distinguish." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 459, + 483, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 459, + 483, + 615 + ], + "spans": [ + { + "bbox": [ + 130, + 459, + 483, + 615 + ], + "type": "text", + "content": "We follow the previous human-in-the-loop system [9] and develop a benchmark dataset for our new RCAD task, i.e., the Feint6K dataset. We test a wide range of public video-text methods with various pretraining strategies on this new benchmark. Two failure examples are visualized in Fig. 1a and quantitative results are summarized in Fig. 1b. Note how state-of-the-art video-text model, InternVideo [23], features a " + }, + { + "bbox": [ + 130, + 459, + 483, + 615 + ], + "type": "inline_equation", + "content": "87.9\\%" + }, + { + "bbox": [ + 130, + 459, + 483, + 615 + ], + "type": "text", + "content": " rank-1 accuracy on standard video-to-text retrieval but drops by " + }, + { + "bbox": [ + 130, + 459, + 483, + 615 + ], + "type": "inline_equation", + "content": "32.8\\%" + }, + { + "bbox": [ + 130, + 459, + 483, + 615 + ], + "type": "text", + "content": " (pretraining only) and " + }, + { + "bbox": [ + 130, + 459, + 483, + 615 + ], + "type": "inline_equation", + "content": "29.7\\%" + }, + { + "bbox": [ + 130, + 459, + 483, + 615 + ], + "type": "text", + "content": " (finetuned) when tested on our benchmark with counterfactually augmented data. By establishing a human-level baseline on our benchmark, the InternVideo model surprisingly falls far behind human performance by " + }, + { + "bbox": [ + 130, + 459, + 483, + 615 + ], + "type": "inline_equation", + "content": "38.6\\%" + }, + { + "bbox": [ + 130, + 459, + 483, + 615 + ], + "type": "text", + "content": ". Our findings sharply contrast to the prominent performance obtained on previous video-text benchmarks and question the common belief that latest video-text models can develop a fairly effective representation for texts and videos in existing datasets." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "In light of our evaluation results on Feint6K, we identify a significant limitation of the widely adopted contrastive representation learning on video-text data, which is the issue of shortcut learning (see Sec. 4.1). To address this, we propose LLM-teacher that enables a more effective learning of action se" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "text", + "content": "W. Ma et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 114, + 206, + 159 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 206, + 159 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 206, + 159 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 206, + 159 + ], + "type": "image", + "image_path": "3c0eebf4e7165d4fe3da696685d2cb5d5f4da7c6217ea066f34ace79e7dd6847.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 133, + 162, + 205, + 183 + ], + "lines": [ + { + "bbox": [ + 133, + 162, + 205, + 183 + ], + "spans": [ + { + "bbox": [ + 133, + 162, + 205, + 183 + ], + "type": "text", + "content": "Groundtruth: \n" + }, + { + "bbox": [ + 133, + 162, + 205, + 183 + ], + "type": "inline_equation", + "content": "\\checkmark" + }, + { + "bbox": [ + 133, + 162, + 205, + 183 + ], + "type": "text", + "content": " A person sitting outside with some tools is examining a rock on a leather piece." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 211, + 114, + 285, + 159 + ], + "blocks": [ + { + "bbox": [ + 133, + 186, + 204, + 208 + ], + "lines": [ + { + "bbox": [ + 133, + 186, + 204, + 208 + ], + "spans": [ + { + "bbox": [ + 133, + 186, + 204, + 208 + ], + "type": "text", + "content": "Predicted: \n" + }, + { + "bbox": [ + 133, + 186, + 204, + 208 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 133, + 186, + 204, + 208 + ], + "type": "text", + "content": " A person sitting outside with some tools is throwing a rock on a leather piece." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 211, + 114, + 285, + 159 + ], + "lines": [ + { + "bbox": [ + 211, + 114, + 285, + 159 + ], + "spans": [ + { + "bbox": [ + 211, + 114, + 285, + 159 + ], + "type": "image", + "image_path": "29b9e994213addfa806d0ae37029d51d199cb91363f3fcf207ffe1da2aa622d9.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 211, + 161, + 282, + 177 + ], + "lines": [ + { + "bbox": [ + 211, + 161, + 282, + 177 + ], + "spans": [ + { + "bbox": [ + 211, + 161, + 282, + 177 + ], + "type": "text", + "content": "Groundtruth: A person is building a space jet toy at high speed." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 292, + 118, + 477, + 206 + ], + "blocks": [ + { + "bbox": [ + 211, + 181, + 282, + 198 + ], + "lines": [ + { + "bbox": [ + 211, + 181, + 282, + 198 + ], + "spans": [ + { + "bbox": [ + 211, + 181, + 282, + 198 + ], + "type": "text", + "content": "Predicted: \n" + }, + { + "bbox": [ + 211, + 181, + 282, + 198 + ], + "type": "inline_equation", + "content": "\\times" + }, + { + "bbox": [ + 211, + 181, + 282, + 198 + ], + "type": "text", + "content": " A person is coloring a space jet toy at high speed." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 133, + 211, + 285, + 221 + ], + "lines": [ + { + "bbox": [ + 133, + 211, + 285, + 221 + ], + "spans": [ + { + "bbox": [ + 133, + 211, + 285, + 221 + ], + "type": "text", + "content": "(a) Failures of InternVideo [23] on Feint6K" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 292, + 118, + 477, + 206 + ], + "lines": [ + { + "bbox": [ + 292, + 118, + 477, + 206 + ], + "spans": [ + { + "bbox": [ + 292, + 118, + 477, + 206 + ], + "type": "image", + "image_path": "1c44aa3ac092405d659aecdd7275730fe461e6676e77870a1a2cb5cc0172905c.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 308, + 211, + 463, + 220 + ], + "lines": [ + { + "bbox": [ + 308, + 211, + 463, + 220 + ], + "spans": [ + { + "bbox": [ + 308, + 211, + 463, + 220 + ], + "type": "text", + "content": "(b) Quantitative results on Feint6K dataset" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 230, + 482, + 318 + ], + "lines": [ + { + "bbox": [ + 130, + 230, + 482, + 318 + ], + "spans": [ + { + "bbox": [ + 130, + 230, + 482, + 318 + ], + "type": "text", + "content": "Fig. 1: (a): Although with large-scale pretraining on web-scale data, current video-text model like [23] can be easily fooled by counterfactually augmented data. (b): The performance of InternVideo on retrieval from counterfactually augmented data (RCAD) drops by over " + }, + { + "bbox": [ + 130, + 230, + 482, + 318 + ], + "type": "inline_equation", + "content": "30\\%" + }, + { + "bbox": [ + 130, + 230, + 482, + 318 + ], + "type": "text", + "content": " when compared to the standard video-to-text retrieval and by " + }, + { + "bbox": [ + 130, + 230, + 482, + 318 + ], + "type": "inline_equation", + "content": "38.6\\%" + }, + { + "bbox": [ + 130, + 230, + 482, + 318 + ], + "type": "text", + "content": " when compared to human-level performance. We also evaluate models on standard video-text retrieval from only 6 sampled candidates and show that our RCAD task is indeed much more challenging. Our LLM-teacher successfully improves the performance on RCAD by enforcing a more effective learning of action semantics." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 342, + 482, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 342, + 482, + 450 + ], + "spans": [ + { + "bbox": [ + 130, + 342, + 482, + 450 + ], + "type": "text", + "content": "mantics by introducing extra knowledge from pretrained large language models (LLMs). Specifically we generate synthetic captions by modifying the contents of an existing caption and LLM serves as the teacher to determine if a synthetic caption matches the semantics in the original caption using binary pseudo-labels or continuous logits. This approach enables a more effective learning of action embeddings from the available video-text data. Experimental results show that LLM-teacher learns more discriminative action embeddings from our advanced contrastive objectives and improves the performance of retrieval from counter-factually augmented data when applied to multiple video-text models." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 451, + 482, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 451, + 482, + 594 + ], + "spans": [ + { + "bbox": [ + 130, + 451, + 482, + 594 + ], + "type": "text", + "content": "In summary, our key contributions are as follows. (1) We propose to evaluate video-text models on questions that require cross-frame reasoning and develop a new task, retrieval from counterfactually augmented data, and a new dataset, Feint6K (Sec. 3). (2) Extensive experimental results on our new evaluation paradigm suggest that existing video-text models demonstrate very limited understanding of the action semantics in a video, which is contrary to the prominent performance achieved on standard video-text retrieval benchmarks. (3) From our results on Feint6K dataset, we identify a key shortcoming of contrastive learning approaches on video-text data. We present LLM-teacher to enforce a more effective learning of action embeddings by injecting knowledge from pretrained LLMs (Sec. 4). Our approach effectively improves the results on Feint6K when applied to multiple text-video models." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 614, + 242, + 626 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 614, + 242, + 626 + ], + "spans": [ + { + "bbox": [ + 132, + 614, + 242, + 626 + ], + "type": "text", + "content": "2 Related Works" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 641, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 482, + 665 + ], + "type": "text", + "content": "Video-text pretraining. With the availability of web-scale video-text paired datasets, such as WebVid2M [1] and HowTo100M [17], recent developments of" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 205, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 205, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 205, + 91, + 447, + 102 + ], + "type": "text", + "content": "Rethinking Video-Text Understanding: RCAD and Feint6K" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 113, + 282, + 240 + ], + "blocks": [ + { + "bbox": [ + 133, + 113, + 282, + 240 + ], + "lines": [ + { + "bbox": [ + 133, + 113, + 282, + 240 + ], + "spans": [ + { + "bbox": [ + 133, + 113, + 282, + 240 + ], + "type": "image", + "image_path": "c336959575ac3049fd973da454260476b894ece1d94dd44934d3745a55dbe9bc.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 143, + 240, + 270, + 250 + ], + "lines": [ + { + "bbox": [ + 143, + 240, + 270, + 250 + ], + "spans": [ + { + "bbox": [ + 143, + 240, + 270, + 250 + ], + "type": "text", + "content": "(a) Standard video-to-text retrieval" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 258, + 482, + 346 + ], + "lines": [ + { + "bbox": [ + 130, + 258, + 482, + 346 + ], + "spans": [ + { + "bbox": [ + 130, + 258, + 482, + 346 + ], + "type": "text", + "content": "Fig. 2: Different evaluations of video-text understanding. (a): In standard video-to-text retrieval, negative captions are sampled from different videos in the same dataset. Image-text models can achieve good performance by exploiting shortcuts (e.g., \"football\" and \"cymbals\") and biases (e.g., spurious correlation between \"outdoor\" and \"football\"). (b): In our proposed RCAD paradigm, we adopt a human-in-the-loop system (see Sec. 3.2) to obtain \"hard\" negative captions with unchanged object entities but modified actions. Models must develop a holistic understanding of the semantics from the sequence of frames to retrieve the matched caption." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 332, + 116, + 484, + 239 + ], + "blocks": [ + { + "bbox": [ + 332, + 116, + 484, + 239 + ], + "lines": [ + { + "bbox": [ + 332, + 116, + 484, + 239 + ], + "spans": [ + { + "bbox": [ + 332, + 116, + 484, + 239 + ], + "type": "image", + "image_path": "8e0b68883a179cd5a7ce1e1fc09f0c0d71cb9e161e78ff68219104f946d93be2.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 388, + 239, + 427, + 249 + ], + "lines": [ + { + "bbox": [ + 388, + 239, + 427, + 249 + ], + "spans": [ + { + "bbox": [ + 388, + 239, + 427, + 249 + ], + "type": "text", + "content": "(b) RCAD" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 371, + 482, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 371, + 482, + 515 + ], + "spans": [ + { + "bbox": [ + 130, + 371, + 482, + 515 + ], + "type": "text", + "content": "video-text pretraining models achieved improved results on a wide range of video understanding tasks. Visual and textual embeddings jointly learned from the large-scale data demonstrate strong transferability and largely reduce the efforts in downstream tasks. Mainstream pretraining objectives can be categorized into discriminative and generative. Discriminative approaches extend the objective of CLIP [18] to the video-text domain and learn multi-modal representations by contrasting between matched and unmatched pairs [15, 16, 23, 26] or simply predicting if a video-text pair is matched [16]. Generative methods follow the masked modeling idea in BERT [5] and adopted masked language modeling (MLM) and masked video modeling (MVM) for video-text pretraining [16, 21, 23]. These generative approaches are often considered superior for action recognition, while discriminative approaches learn better semantics from texts [21]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 533, + 482, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 482, + 640 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 482, + 640 + ], + "type": "text", + "content": "Evaluation of video-text models. In order to analyze the effectiveness of video-text pretraining, previous works [15,16,21,23,26,29] focused on the following tasks: (1) video-text retrieval (zero-shot or finetuned) [22, 28]: retrieving the matched video (or text) given a text (or video) as query where the negative candidates are unmatched pairs from the same dataset, (2) video question answering (finetuned) [25]: predicting the answer with a classification head, (3) video classification (zero-shot or finetuned) [6,10,20]: classifying videos using label names as text prompts, and (4) video captioning (zero-shot or finetuned) [11, 28]: summarizing the contents of a video." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 641, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 482, + 665 + ], + "type": "text", + "content": "We argue that results on these video-text tasks could be deceitful as most tasks heavily rely on the alignment of object entities in the video and text" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "type": "text", + "content": "W. Ma et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 284 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 284 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 284 + ], + "type": "text", + "content": "(which image-text models are also capable of), and “action understanding” can often be developed through shortcuts and biases. For instance, a commonly-used evaluation for “action recognition” is video classification on ActivityNet [6], where a model classifies a given video into classes such as “playing badminton”, “kayaking”, or “volleyball”. However, video-text models would classify the videos by exploiting objects and contexts as shortcuts (e.g., “badminton” and “kayak”), as well as other biases (e.g., spurious correlation between “kayak” and “on the water”) without genuinely understanding the semantics of the action represented by interactions between the person and the objects over time. For instance, classifying the action as “kayaking” is essentially a “kayak” detection problem in ActivityNet. Therefore we propose retrieval from counterfactually augmented data, a new evaluation paradigm where we aim to eliminate the shortcuts from the questions so models must establish a comprehensive understanding of the semantics from cross-frame reasoning in order to predict the correct answer." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 302, + 448, + 316 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 302, + 448, + 316 + ], + "spans": [ + { + "bbox": [ + 132, + 302, + 448, + 316 + ], + "type": "text", + "content": "3 Retrieval from Counterfactually Augmented Data" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 327, + 482, + 447 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 327, + 482, + 447 + ], + "spans": [ + { + "bbox": [ + 130, + 327, + 482, + 447 + ], + "type": "text", + "content": "Previous evaluation tasks of video-text models focused on the alignment of feature embeddings of the video-text pairs (e.g., video-text retrieval [22, 28] and video classification [10, 20]) or reconstruction of the semantics with text (e.g., video question answering [25] and video captioning [11, 28]). However, these evaluation tasks are largely limited by the paired data available in existing video-text datasets. As demonstrated by in Fig. 2a, we can extract most of the semantics of a video, such as the objects, people, and contexts in the video, by looking at only one frame from the video. The action in the video can be further inferred by exploiting biases in the datasets. Therefore, models pretrained on image-text data often perform surprisingly well on existing video-text tasks." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 448, + 482, + 591 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 448, + 482, + 591 + ], + "spans": [ + { + "bbox": [ + 130, + 448, + 482, + 591 + ], + "type": "text", + "content": "To effectively evaluate how video-text models can understand the contents and semantics of videos beyond images, we must develop video understanding tasks that are free from shortcuts and require cross-frame reasoning to solve. To this end, we propose a new evaluation task for video-text models, i.e., retrieval from counterfactually augmented data (Sec. 3.1), and a new dataset Feint6K (Sec. 3.2). In contrast to previous datasets with matched video-text pairs scrapped from the web [24, 28] or annotated by human [22], we adopt a human-in-the-loop system and annotate counterfactually manipulated texts as negative pairs with the original video. This allows us to evaluate video-text models with novel tasks and better understand their limitations. Lastly we measure human performance on our dataset as a direct comparison with the state-of-the-art video-text models (Sec. 3.3)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 609, + 249, + 620 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 609, + 249, + 620 + ], + "spans": [ + { + "bbox": [ + 132, + 609, + 249, + 620 + ], + "type": "text", + "content": "3.1 Task Formulation" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 629, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 481, + 665 + ], + "type": "text", + "content": "Retrieval from counterfactually augmented data is a variant of the standard video-to-text retrieval. As shown in Fig. 2b, given a video sequence and a list of candidate captions, the model is asked to retrieve the caption that best matches" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 205, + 90, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 205, + 90, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 205, + 90, + 447, + 102 + ], + "type": "text", + "content": "Rethinking Video-Text Understanding: RCAD and Feint6K" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 122, + 168, + 134 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 122, + 168, + 134 + ], + "spans": [ + { + "bbox": [ + 130, + 122, + 168, + 134 + ], + "type": "text", + "content": "Video A" + } + ] + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 178, + 113, + 482, + 145 + ], + "blocks": [ + { + "bbox": [ + 178, + 113, + 482, + 145 + ], + "lines": [ + { + "bbox": [ + 178, + 113, + 482, + 145 + ], + "spans": [ + { + "bbox": [ + 178, + 113, + 482, + 145 + ], + "type": "image", + "image_path": "f917624d6973ec7537a39238f822e2d9cbcada4184c6e42855970622fc90204a.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 200, + 305, + 412, + 316 + ], + "lines": [ + { + "bbox": [ + 200, + 305, + 412, + 316 + ], + "spans": [ + { + "bbox": [ + 200, + 305, + 412, + 316 + ], + "type": "text", + "content": "Fig. 3: Examples of " + }, + { + "bbox": [ + 200, + 305, + 412, + 316 + ], + "type": "inline_equation", + "content": "RCAD" + }, + { + "bbox": [ + 200, + 305, + 412, + 316 + ], + "type": "text", + "content": " on our Feint6K dataset." + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 159, + 167, + 169 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 159, + 167, + 169 + ], + "spans": [ + { + "bbox": [ + 132, + 159, + 167, + 169 + ], + "type": "text", + "content": "Video B" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 178, + 148, + 481, + 180 + ], + "blocks": [ + { + "bbox": [ + 178, + 148, + 481, + 180 + ], + "lines": [ + { + "bbox": [ + 178, + 148, + 481, + 180 + ], + "spans": [ + { + "bbox": [ + 178, + 148, + 481, + 180 + ], + "type": "image", + "image_path": "b784d2ebffcb678ab1a7ba413eb75cf19f603be00ab0b2f003e25f872a3d3e4d.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 190, + 168, + 201 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 190, + 168, + 201 + ], + "spans": [ + { + "bbox": [ + 132, + 190, + 168, + 201 + ], + "type": "text", + "content": "Video C" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 178, + 184, + 481, + 208 + ], + "blocks": [ + { + "bbox": [ + 178, + 184, + 481, + 208 + ], + "lines": [ + { + "bbox": [ + 178, + 184, + 481, + 208 + ], + "spans": [ + { + "bbox": [ + 178, + 184, + 481, + 208 + ], + "type": "image", + "image_path": "adee20fde29321c8721323115aae2802dfc2fe5728d8f8d8839a2efd253409d3.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 178, + 212, + 246, + 220 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 178, + 212, + 246, + 220 + ], + "spans": [ + { + "bbox": [ + 178, + 212, + 246, + 220 + ], + "type": "text", + "content": "Captions of Video A" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 178, + 223, + 272, + 246 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 178, + 223, + 272, + 246 + ], + "spans": [ + { + "bbox": [ + 178, + 223, + 272, + 246 + ], + "type": "text", + "content": "[Positive] A person is shinning a shoe with a shoe shinning machine." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 178, + 248, + 264, + 271 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 178, + 248, + 264, + 271 + ], + "spans": [ + { + "bbox": [ + 178, + 248, + 264, + 271 + ], + "type": "text", + "content": "[Negative 1] A person is making a shoe with a shoe shimming machine." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 178, + 274, + 266, + 297 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 178, + 274, + 266, + 297 + ], + "spans": [ + { + "bbox": [ + 178, + 274, + 266, + 297 + ], + "type": "text", + "content": "[Negative 2] A person is wearing a shoe with a shoe shinning machine." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 281, + 212, + 348, + 220 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 281, + 212, + 348, + 220 + ], + "spans": [ + { + "bbox": [ + 281, + 212, + 348, + 220 + ], + "type": "text", + "content": "Captions of Video B" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 280, + 223, + 376, + 246 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 280, + 223, + 376, + 246 + ], + "spans": [ + { + "bbox": [ + 280, + 223, + 376, + 246 + ], + "type": "text", + "content": "[Positive] A man is standing on the field attempting to kick a football." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 280, + 248, + 376, + 271 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 280, + 248, + 376, + 271 + ], + "spans": [ + { + "bbox": [ + 280, + 248, + 376, + 271 + ], + "type": "text", + "content": "[Negative 1] A man is standing on the field attempting to catch a football." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 280, + 274, + 377, + 297 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 280, + 274, + 377, + 297 + ], + "spans": [ + { + "bbox": [ + 280, + 274, + 377, + 297 + ], + "type": "text", + "content": "[Negative 2] A man is standing on the field attempting to throw a football." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 384, + 212, + 451, + 220 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 384, + 212, + 451, + 220 + ], + "spans": [ + { + "bbox": [ + 384, + 212, + 451, + 220 + ], + "type": "text", + "content": "Captions of Video C" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 383, + 223, + 478, + 246 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 383, + 223, + 478, + 246 + ], + "spans": [ + { + "bbox": [ + 383, + 223, + 478, + 246 + ], + "type": "text", + "content": "[Positive] A person wearing a pink jacket is feeding birds with his hand." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 383, + 248, + 481, + 271 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 383, + 248, + 481, + 271 + ], + "spans": [ + { + "bbox": [ + 383, + 248, + 481, + 271 + ], + "type": "text", + "content": "[Negative 1] A person wearing a pink jacket is petting birds with his hand." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 383, + 274, + 481, + 297 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 383, + 274, + 481, + 297 + ], + "spans": [ + { + "bbox": [ + 383, + 274, + 481, + 297 + ], + "type": "text", + "content": "[Negative 2] A person wearing a pink jacket is holding birds with his hand." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 130, + 340, + 480, + 388 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 340, + 480, + 388 + ], + "spans": [ + { + "bbox": [ + 130, + 340, + 480, + 388 + ], + "type": "text", + "content": "the semantics in the video. However, unlike standard video-to-text retrieval where negative captions come from other video-text pairs in the same dataset, negative captions in our task are modified from the positive captions, with the same text structure and object entities but different actions (see Fig. 3)." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 130, + 388, + 481, + 424 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 388, + 481, + 424 + ], + "spans": [ + { + "bbox": [ + 130, + 388, + 481, + 424 + ], + "type": "text", + "content": "Similar to video-text retrieval, our proposed task provides the benefit of zero-shot evaluation. This allows us to quantitatively analyze the effectiveness of video-text pretraining without downstream finetuning." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 130, + 425, + 482, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 425, + 482, + 544 + ], + "spans": [ + { + "bbox": [ + 130, + 425, + 482, + 544 + ], + "type": "text", + "content": "What distinguishes our task from all previous video-text tasks is its focus on questions that require cross-frame reasoning. Consider \"Video B\" in Fig. 3 as an example. All captions contain the same object entities present in the video, i.e., \"a man\", \"the field\", and \"a football\", but differ in the action specified. We cannot determine the most accurate caption because all suggested actions, \"kick\", \"catch\", and \"throw\", are plausible given the video's context. To retrieve the corresponding caption, it is crucial to grasp the action's semantics by inspecting the interactions between the man and the football across a series of frames. This underscores the necessity of understanding the video semantics over time in order to succeed in our task, rather than relying shortcuts or biases." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 131, + 562, + 288, + 573 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 562, + 288, + 573 + ], + "spans": [ + { + "bbox": [ + 131, + 562, + 288, + 573 + ], + "type": "text", + "content": "3.2 Feint6K: Data Collection" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "content": "We utilize a human-in-the-loop system [9] to counterfactually manipulate the positive captions in existing video-text datasets [22, 28]. The generated counterfactually augmented data, when paired with the original videos, forms negative pairs to be used in our proposed task. We recruit 40 annotators to manually manipulate the actions in existing captions. Specifically, the new actions introduced must be plausible within the context in the caption but are not occurring in the corresponding video. In summary, a total of 6,243 videos from the validation" + } + ] + } + ], + "index": 25 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "text", + "content": "W. Ma et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 149, + 112, + 465, + 211 + ], + "blocks": [ + { + "bbox": [ + 149, + 112, + 465, + 211 + ], + "lines": [ + { + "bbox": [ + 149, + 112, + 465, + 211 + ], + "spans": [ + { + "bbox": [ + 149, + 112, + 465, + 211 + ], + "type": "image", + "image_path": "2e45d571385c4daed4a61b7268a9f80b13fcbbc644dcfb1e49b93b565c82733c.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 218, + 480, + 240 + ], + "lines": [ + { + "bbox": [ + 131, + 218, + 480, + 240 + ], + "spans": [ + { + "bbox": [ + 131, + 218, + 480, + 240 + ], + "type": "text", + "content": "Fig. 4: Overview of our data collection pipeline for Feint6K dataset featuring a human-in-the-loop system." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 265, + 479, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 265, + 479, + 300 + ], + "spans": [ + { + "bbox": [ + 130, + 265, + 479, + 300 + ], + "type": "text", + "content": "set of the MSR-VTT dataset [28] and the test set of the VATEX dataset [22]. For detailed statistics of our Feint6K dataset, please refer to our supplementary materials." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 301, + 480, + 409 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 301, + 480, + 409 + ], + "spans": [ + { + "bbox": [ + 130, + 301, + 480, + 409 + ], + "type": "text", + "content": "The overview of our human-in-the-loop system is demonstrated in Fig. 4. We start by sampling a small subset of video-text pairs and manually annotate the counterfactually augmented captions for the purpose of training and demonstration. Specific guidance and feedback are given to the annotators on the practice questions during the training stage. Moreover, we adopt a reviewing process during the annotation stage to further ensure the quality of our annotated data. Each annotation is reviewed before acceptance, where rejected annotations are marked and sent back for refinement. We refer the readers to the supplementary materials regarding the annotation interface and annotator guidance." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 429, + 264, + 440 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 429, + 264, + 440 + ], + "spans": [ + { + "bbox": [ + 132, + 429, + 264, + 440 + ], + "type": "text", + "content": "3.3 Human Performance" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 450, + 480, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 450, + 480, + 510 + ], + "spans": [ + { + "bbox": [ + 130, + 450, + 480, + 510 + ], + "type": "text", + "content": "We employ the same group of annotators to establish a human-level baseline on our Feint6K dataset. Analyzing the human performance serves two goals: (i) verifying the legitimacy of our Feint6K dataset – if each question is answerable and has one unique answer, and (ii) comparing the state-of-the-art video-text models with human-level performance." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 510, + 481, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 510, + 481, + 594 + ], + "spans": [ + { + "bbox": [ + 130, + 510, + 481, + 594 + ], + "type": "text", + "content": "Each annotator is first presented with the full video and then asked to select the one out of six captions that best matches the semantics in the video. To avoid leakage, we make sure the annotators assigned to the testing questions are not the same annotators annotating this video. We also randomly shuffle the order of the candidate captions to remove any biases. In Sec. 5.2 we present the quantitative comparisons between human-level performance and current state-of-the-art video-text models on our Feint6K dataset." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 614, + 231, + 626 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 614, + 231, + 626 + ], + "spans": [ + { + "bbox": [ + 132, + 614, + 231, + 626 + ], + "type": "text", + "content": "4 LLM-teacher" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 641, + 479, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 479, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 479, + 665 + ], + "type": "text", + "content": "Contrastive learning are widely adopted in previous video-text representation learning works [16, 23, 26]. The contrastive models learn to distinguish between" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 205, + 90, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 205, + 90, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 205, + 90, + 448, + 102 + ], + "type": "text", + "content": "Rethinking Video-Text Understanding: RCAD and Feint6K" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 199 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 199 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 199 + ], + "type": "text", + "content": "similar and dissimilar pairs of video-text data by contrasting their feature representations. Despite the simple formulation and great generalization ability, our results on Feint6K show that current video-text models built heavily on contrastive objectives have very limited understanding of action semantics in natural videos and can be easily fooled by counterfactually augmented data. Therefore we aim to develop a more effective contrastive approach and learn more discriminative action representations from existing video-text data." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 200, + 480, + 260 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 200, + 480, + 260 + ], + "spans": [ + { + "bbox": [ + 130, + 200, + 480, + 260 + ], + "type": "text", + "content": "Sec. 4.1 discusses the limitations of current contrastive approaches due to numerous shortcuts in video-text dataset. Inspired by this finding, we present LLM-teacher in Sec. 4.2, a simple but effective approach to learn more discriminative action embeddings by incorporating knowledge from pretrained large language models." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 279, + 309, + 290 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 279, + 309, + 290 + ], + "spans": [ + { + "bbox": [ + 132, + 279, + 309, + 290 + ], + "type": "text", + "content": "4.1 Shortcuts in Video-Text Data" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 300, + 479, + 360 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 300, + 479, + 360 + ], + "spans": [ + { + "bbox": [ + 130, + 300, + 479, + 360 + ], + "type": "text", + "content": "The CLIP model [18] trained on image-text data with a contrastive objective demonstrated strong zero-shot capabilities in many vision tasks [13,30] and was widely adopted as the vision encoder in many large vision-language models [14]. However, directly generalizing this idea to video-text data may face unprecedented challenges due to shortcuts in the video-text data." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 361, + 479, + 516 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 361, + 479, + 516 + ], + "spans": [ + { + "bbox": [ + 130, + 361, + 479, + 516 + ], + "type": "text", + "content": "As we proceed from image-text to video-text representation learning, we aim to learn a powerful embedding for not only the object entities, but also the actions specified by texts or sequence of frames. Naturally we hope to achieve this by contrasting videos with different actions. However, as we are optimizing the contrastive objectives in a mini-batch of videos, object entities become the shortcuts that contrastive models exploit to saturate the contrastive objective, leading to an ineffective action embeddings. Consider the example in Fig. 2a, the contrastive loss would be small as long as the model learns discriminative embeddings for shortcut objects such as \"cymbal\" and \"football\", even without any understanding of the actions. Since many video-text models start with pretrained CLIP encoders [23,26], they often possess strong embeddings for shortcut objects from the beginning of the video-text pretraining, which hinders the model from further learning effective action representations." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 517, + 479, + 576 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 517, + 479, + 576 + ], + "spans": [ + { + "bbox": [ + 130, + 517, + 479, + 576 + ], + "type": "text", + "content": "In Fig. 5 we present quantitative results of changes in cosine similarities when the video is unchanged but the object or action in the caption is manipulated. We find that state-of-the-art video-text models learn less discriminative embeddings for actions as compared to objects, which also explains the small gap between image-text and video-text models on standard video-text tasks." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 596, + 227, + 607 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 596, + 227, + 607 + ], + "spans": [ + { + "bbox": [ + 132, + 596, + 227, + 607 + ], + "type": "text", + "content": "4.2 LLM-teacher" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 617, + 491, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 491, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 491, + 665 + ], + "type": "text", + "content": "We introduce LLM-teacher, an LLM-powered approach to learn better action representations with contrastive objectives. Specifically, LLM serves as the \"teacher\" and provides synthesized captions as extra knowledge for the video-text model to learn from. These new captions serve as negative captions of the original video," + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "type": "text", + "content": "W. Ma et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 137, + 117, + 302, + 209 + ], + "blocks": [ + { + "bbox": [ + 137, + 117, + 302, + 209 + ], + "lines": [ + { + "bbox": [ + 137, + 117, + 302, + 209 + ], + "spans": [ + { + "bbox": [ + 137, + 117, + 302, + 209 + ], + "type": "image", + "image_path": "49fff5b43c9cdadc257ceae8112a25ec80fe30eb7d114c6ddecce79a0869f22a.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 133, + 213, + 304, + 232 + ], + "lines": [ + { + "bbox": [ + 133, + 213, + 304, + 232 + ], + "spans": [ + { + "bbox": [ + 133, + 213, + 304, + 232 + ], + "type": "text", + "content": "(a) Comparison between " + }, + { + "bbox": [ + 133, + 213, + 304, + 232 + ], + "type": "inline_equation", + "content": "\\Delta s" + }, + { + "bbox": [ + 133, + 213, + 304, + 232 + ], + "type": "text", + "content": " w.r.t. object or action changes." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 313, + 118, + 478, + 209 + ], + "blocks": [ + { + "bbox": [ + 313, + 118, + 478, + 209 + ], + "lines": [ + { + "bbox": [ + 313, + 118, + 478, + 209 + ], + "spans": [ + { + "bbox": [ + 313, + 118, + 478, + 209 + ], + "type": "image", + "image_path": "c89c032b08ca7476da73e19c348b841d3f223e885edbc7f98ffb0aa738177e97.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 309, + 213, + 481, + 231 + ], + "lines": [ + { + "bbox": [ + 309, + 213, + 481, + 231 + ], + "spans": [ + { + "bbox": [ + 309, + 213, + 481, + 231 + ], + "type": "text", + "content": "(b) Comparison between " + }, + { + "bbox": [ + 309, + 213, + 481, + 231 + ], + "type": "inline_equation", + "content": "\\Delta s" + }, + { + "bbox": [ + 309, + 213, + 481, + 231 + ], + "type": "text", + "content": " w.r.t. action using InternVideo and our LLM-teacher." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 239, + 480, + 316 + ], + "lines": [ + { + "bbox": [ + 130, + 239, + 480, + 316 + ], + "spans": [ + { + "bbox": [ + 130, + 239, + 480, + 316 + ], + "type": "text", + "content": "Fig. 5: Change of cosine similarity w.r.t. objects or actions. (a): Comparison between the change in cosine similarity when the action or object is swapped. Results show that current video-text models learn a more effective embedding for objects than for actions. (b): Comparison between the change in cosine similarity using InternVideo or our LLM-teacher. This demonstrate that LLM-teacher learns a more discriminative embedding for actions by enabling a more effective contrastive learning using knowledge from LLMs. Refer to Sec. 5.2 for more details." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 345, + 480, + 368 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 345, + 480, + 368 + ], + "spans": [ + { + "bbox": [ + 130, + 345, + 480, + 368 + ], + "type": "text", + "content": "removing possible shortcuts and biases and enabling a more effective learning of action semantics." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 371, + 480, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 371, + 480, + 430 + ], + "spans": [ + { + "bbox": [ + 130, + 371, + 480, + 430 + ], + "type": "text", + "content": "We start with positive captions from common video-text datasets and run an abstract meaning representation (AMR) parser, resulting in a list of action or object tokens. To generate the hard negative captions, two methods are considered to find novel actions or objects given the original caption and the substitution token " + }, + { + "bbox": [ + 130, + 371, + 480, + 430 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 130, + 371, + 480, + 430 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 458, + 480, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 458, + 480, + 542 + ], + "spans": [ + { + "bbox": [ + 130, + 458, + 480, + 542 + ], + "type": "text", + "content": "Method I: Mask filling. Masked language modeling (MLM) [5] is a widely adopted self-supervised task for LLMs. In the pretraining stage, a certain percentage of word tokens are substituted with a special token [mask] and the LLM is optimized to recover the masked tokens. To find negative captions with novel actions given the original caption " + }, + { + "bbox": [ + 130, + 458, + 480, + 542 + ], + "type": "inline_equation", + "content": "c_{0}" + }, + { + "bbox": [ + 130, + 458, + 480, + 542 + ], + "type": "text", + "content": " and substitution token " + }, + { + "bbox": [ + 130, + 458, + 480, + 542 + ], + "type": "inline_equation", + "content": "a" + }, + { + "bbox": [ + 130, + 458, + 480, + 542 + ], + "type": "text", + "content": ", we first substitute the token " + }, + { + "bbox": [ + 130, + 458, + 480, + 542 + ], + "type": "inline_equation", + "content": "a" + }, + { + "bbox": [ + 130, + 458, + 480, + 542 + ], + "type": "text", + "content": " with [mask] and use a LLM pretrained with MLM to predict a list of " + }, + { + "bbox": [ + 130, + 458, + 480, + 542 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 130, + 458, + 480, + 542 + ], + "type": "text", + "content": " possible tokens " + }, + { + "bbox": [ + 130, + 458, + 480, + 542 + ], + "type": "inline_equation", + "content": "\\{\\hat{w}_i\\}_{i=1}^k" + }, + { + "bbox": [ + 130, + 458, + 480, + 542 + ], + "type": "text", + "content": " such that " + }, + { + "bbox": [ + 130, + 458, + 480, + 542 + ], + "type": "inline_equation", + "content": "\\hat{w}_i \\neq w" + }, + { + "bbox": [ + 130, + 458, + 480, + 542 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 458, + 480, + 542 + ], + "type": "inline_equation", + "content": "\\hat{w}_i" + }, + { + "bbox": [ + 130, + 458, + 480, + 542 + ], + "type": "text", + "content": " is an action token." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 569, + 480, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 569, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 569, + 480, + 665 + ], + "type": "text", + "content": "Method II: LLM-powered chatbot. Although mask filling can find appropriate words to substitute " + }, + { + "bbox": [ + 130, + 569, + 480, + 665 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 130, + 569, + 480, + 665 + ], + "type": "text", + "content": ", it is limited by one single token. In many cases, we want to update the prepositions following the change of verbs. For example, when \"install\" is substituted by \"uninstall\", we should also update the preposition from \"install ... to\" to \"uninstall ... from\". Here we consider a more flexible approach that builds on LLM-powered chatbots. Besides a prompt describing the text substitution, we leverage the in-context learning ability of LLMs [2, 7] and provide multiple in-context examples to obtain the desirable negative captions." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 205, + 90, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 205, + 90, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 205, + 90, + 448, + 102 + ], + "type": "text", + "content": "Rethinking Video-Text Understanding: RCAD and Feint6K" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 479, + 188 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 479, + 188 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 479, + 188 + ], + "type": "text", + "content": "LLM-teacher. With the procedure above, we obtain " + }, + { + "bbox": [ + 132, + 116, + 479, + 188 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 132, + 116, + 479, + 188 + ], + "type": "text", + "content": " synthesized negative captions " + }, + { + "bbox": [ + 132, + 116, + 479, + 188 + ], + "type": "inline_equation", + "content": "\\{n_i\\}_{i=1}^k" + }, + { + "bbox": [ + 132, + 116, + 479, + 188 + ], + "type": "text", + "content": " besides the original positive caption " + }, + { + "bbox": [ + 132, + 116, + 479, + 188 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 132, + 116, + 479, + 188 + ], + "type": "text", + "content": " for each video " + }, + { + "bbox": [ + 132, + 116, + 479, + 188 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 132, + 116, + 479, + 188 + ], + "type": "text", + "content": " in existing datasets [22,28]. For each of the original and generated captions, LLM serves as a \"teacher\" and provide binary pseudo-labels - if the caption have matched (\"1\") or unmatched (\"0\") semantics with the video. Now we may optimize the video-text model with the normalized temperature-scaled cross entropy loss" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 181, + 195, + 481, + 225 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 181, + 195, + 481, + 225 + ], + "spans": [ + { + "bbox": [ + 181, + 195, + 481, + 225 + ], + "type": "interline_equation", + "content": "l = - \\log \\frac {\\exp \\left(\\sin \\left(f _ {v} , f _ {p}\\right) / \\tau\\right)}{\\exp \\left(\\sin \\left(f _ {v} , f _ {p}\\right) / \\tau\\right) + \\sum_ {i = 1} ^ {k} \\exp \\left(\\sin \\left(f _ {v} , f _ {n _ {i}}\\right) / \\tau\\right)} \\tag {1}", + "image_path": "ffbcc9cb5c4555232fee2744a2b661430d72aa95a3098b333ea768590f2b5e27.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 232, + 377, + 244 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 232, + 377, + 244 + ], + "spans": [ + { + "bbox": [ + 132, + 232, + 377, + 244 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 132, + 232, + 377, + 244 + ], + "type": "inline_equation", + "content": "f_{v}, f_{p}, f_{ni}" + }, + { + "bbox": [ + 132, + 232, + 377, + 244 + ], + "type": "text", + "content": " are the visual and textual embeddings." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 245, + 479, + 316 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 245, + 479, + 316 + ], + "spans": [ + { + "bbox": [ + 130, + 245, + 479, + 316 + ], + "type": "text", + "content": "In practice, we find that certain negative captions have unmatched but similar semantics to the video. It is undesirable to regard them as a strictly negative pair in the contrastive loss. Therefore, we extend the binary pseudo-labels to soft logits by computing caption similarities with a pretrained LLM and then optimize the video-text model to match model outputs with the soft logits from the LLM teacher model [8]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 216, + 327, + 480, + 339 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 216, + 327, + 480, + 339 + ], + "spans": [ + { + "bbox": [ + 216, + 327, + 480, + 339 + ], + "type": "interline_equation", + "content": "l = \\mathcal {L} _ {\\mathrm {K L}} \\left(z _ {\\text {v i d e o - t e x t}}, z _ {\\text {L L M}}\\right) \\tag {2}", + "image_path": "8dd5727b3b9b85b0ad6270c10d1ca4f904012ff7454eb3be55fa01a8827484a3.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 174, + 342, + 438, + 370 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 174, + 342, + 438, + 370 + ], + "spans": [ + { + "bbox": [ + 174, + 342, + 438, + 370 + ], + "type": "interline_equation", + "content": "z _ {\\mathrm {v i d e o - t e x t , t}} = \\frac {\\exp (\\sin (f _ {v} , f _ {t}) / \\tau)}{\\exp (\\sin (f _ {v} , f _ {p}) / \\tau) + \\sum_ {i = 1} ^ {k} \\exp (\\sin (f _ {v} , f _ {n _ {i}}) / \\tau)}", + "image_path": "6f0e61eea8e23264c0ff8062d540b6912e398b2ce91ab25f75d2f67c2a1ad819.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 192, + 372, + 437, + 401 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 192, + 372, + 437, + 401 + ], + "spans": [ + { + "bbox": [ + 192, + 372, + 437, + 401 + ], + "type": "interline_equation", + "content": "z _ {\\mathrm {L L M , t}} = \\frac {\\exp (\\sin (e _ {p} , e _ {t}) / \\tau)}{\\exp (\\sin (e _ {p} , e _ {p}) / \\tau) + \\sum_ {i = 1} ^ {k} \\exp (\\sin (e _ {p} , e _ {n _ {i}}) / \\tau)}", + "image_path": "dc7f408a71bae2ffb74f49e417e61b0980de2581ccbee64929cd5a8ad3aac1af.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 409, + 479, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 409, + 479, + 434 + ], + "spans": [ + { + "bbox": [ + 132, + 409, + 479, + 434 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 132, + 409, + 479, + 434 + ], + "type": "inline_equation", + "content": "t \\in \\{n_i\\}_{i=1}^k \\cup \\{p\\}" + }, + { + "bbox": [ + 132, + 409, + 479, + 434 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 132, + 409, + 479, + 434 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 132, + 409, + 479, + 434 + ], + "type": "text", + "content": "'s are textual and visual embeddings of the video-text model, and " + }, + { + "bbox": [ + 132, + 409, + 479, + 434 + ], + "type": "inline_equation", + "content": "e" + }, + { + "bbox": [ + 132, + 409, + 479, + 434 + ], + "type": "text", + "content": "'s are textual embeddings of a pretrained LLM." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 452, + 229, + 465 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 452, + 229, + 465 + ], + "spans": [ + { + "bbox": [ + 132, + 452, + 229, + 465 + ], + "type": "text", + "content": "5 Experiments" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 477, + 261, + 490 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 477, + 261, + 490 + ], + "spans": [ + { + "bbox": [ + 132, + 477, + 261, + 490 + ], + "type": "text", + "content": "5.1 Experimental Setup" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 497, + 479, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 497, + 479, + 545 + ], + "spans": [ + { + "bbox": [ + 132, + 497, + 479, + 545 + ], + "type": "text", + "content": "Datasets. We use MSR-VTT dataset [28] and VATEX dataset [22] for standard video-to-text retrieval and our Feint6K dataset for retrieval from counter-factually augmented data. Note that Feint6K dataset contains the same video sequences for evaluation as in MSR-VTT dataset and VATEX dataset." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 563, + 479, + 623 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 563, + 479, + 623 + ], + "spans": [ + { + "bbox": [ + 132, + 563, + 479, + 623 + ], + "type": "text", + "content": "Evaluation metrics. For standard video-to-text retrieval we report rank-1 accuracy " + }, + { + "bbox": [ + 132, + 563, + 479, + 623 + ], + "type": "inline_equation", + "content": "(\\mathbb{R}@\\mathbb{1})" + }, + { + "bbox": [ + 132, + 563, + 479, + 623 + ], + "type": "text", + "content": " as this is the most challenging metric and distinguish the performance of previous video-text models. For retrieval from counterfactually-augmented data we use rank-1 accuracy " + }, + { + "bbox": [ + 132, + 563, + 479, + 623 + ], + "type": "inline_equation", + "content": "(\\mathbb{R}@\\mathbb{1})" + }, + { + "bbox": [ + 132, + 563, + 479, + 623 + ], + "type": "text", + "content": ", rank-2 accuracy " + }, + { + "bbox": [ + 132, + 563, + 479, + 623 + ], + "type": "inline_equation", + "content": "(\\mathbb{R}@\\mathbb{2})" + }, + { + "bbox": [ + 132, + 563, + 479, + 623 + ], + "type": "text", + "content": ", and mean rank (MeanR)." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 641, + 479, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 641, + 479, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 641, + 479, + 665 + ], + "type": "text", + "content": "Video-text models. We consider multiple state-of-the-art public models in this work. CLIP4Clip [15] extends CLIP model [18] to the video domain using a" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "text", + "content": "W. Ma et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 132, + 114, + 481, + 312 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 481, + 312 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 481, + 312 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 481, + 312 + ], + "type": "table", + "html": "
ModelMSR-VTTFeint6K (MSR-VTT)VATEXFeint6K (VATEX)
R@1R@1R@2MeanRR@1R@1R@2MeanR
Human95.296.8
Random<1e-316.7<1e-316.7
Zero-Shot
CLIP [18]26.337.355.32.638.834.854.32.7
VideoCLIP [27]14.535.171.32.613.733.070.72.7
InternVideo [23]37.545.863.62.376.944.163.92.3
LanguageBind [31]42.841.377.02.442.377.02.4
Finetuned
LanguageBind [31]
CLIP4Clip [15]43.150.872.42.0
VindLU [3]46.653.470.92.0
SimVTP [16]50.235.770.82.676.633.668.42.6
w/ LLM-teacher-lbl49.0 ↓1.240.0 ↑4.373.1 ↑2.32.3 ↓0.374.8 ↓1.837.3 ↑3.772.2 ↑3.82.4 ↓0.2
w/ LLM-teacher-lgt49.5 ↓0.743.5 ↑7.875.0 ↑4.22.2 ↓0.475.3 ↓1.340.1 ↑6.573.5 ↑5.12.3 ↓0.3
InternVideo [23]49.158.680.21.887.958.276.91.9
w/ LLM-teacher-lbl48.2 ↓0.964.2 ↑5.682.5 ↑2.31.7 ↓0.185.2 ↓2.763.8 ↑5.680.5 ↑3.61.7 ↓0.2
w/ LLM-teacher-lgt48.9 ↓0.265.8 ↑7.283.8 ↑3.61.7 ↓0.187.3 ↓0.665.6 ↑7.481.7 ↑4.81.7 ↓0.2
", + "image_path": "5f876245145b0bb24a0ae0bbaf7421a7cceed97753ecae41e714329559088ab0.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 321, + 482, + 410 + ], + "lines": [ + { + "bbox": [ + 130, + 321, + 482, + 410 + ], + "spans": [ + { + "bbox": [ + 130, + 321, + 482, + 410 + ], + "type": "text", + "content": "Table 1: Performance of standard video-to-text retrieval on MSR-VTT dataset [28] and VATEX dataset [22], as compared to performance of retrieval from counterfactually-augmented data on our Feint6K dataset. In the \"zero-shot\" setting, models are evaluated directly after the pretraining stage, while in the \"finetuned\" setting, models are finetuned on the training set of MSR-VTT or VATEX. With LLM-teacher, we enable a more efficient learning of action semantics and effectively improves R@1 accuracy on Feint6K. Here \"LLM-teacher-lbl\" stands for our approach with binary pseudo-labels and \"LLM-teacher-lgt\" uses soft logits." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 130, + 434, + 482, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 434, + 482, + 517 + ], + "spans": [ + { + "bbox": [ + 130, + 434, + 482, + 517 + ], + "type": "text", + "content": "mean pooling mechanism for zero-shot video-text retrieval and a contrastive loss for finetuning. SimVTP [16] is pretrained on the WebVid2M dataset [1] with a combination of contrastive learning and masked modeling. InternVideo [23] is a video-text foundation model pretrained on a combination of 7 large-scale videotext datasets, with a supervised video post-pretraining for better video recognition. LanguageBind [31] is a language-based multi-modal pretraining model with remarkable performance on a wide range of benchmarks." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 518, + 482, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 518, + 482, + 601 + ], + "spans": [ + { + "bbox": [ + 130, + 518, + 482, + 601 + ], + "type": "text", + "content": "To demonstrate the efficacy of our LLM-teacher, we adopt this method on two state-of-the-art video-text models, SimVTP and InternVideo. In the default setting we generate 10 action-based captions with LLM for each training video. For text generation we use a pretrained XLM-RoBERTa [4] for mask filling. To measure text similarities and compute soft logits, we utilize a pretrained Sentence-BERT model [19]. In Sec. 5.3 we ablate on the choices of parameters and caption generation." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 620, + 228, + 631 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 620, + 228, + 631 + ], + "spans": [ + { + "bbox": [ + 132, + 620, + 228, + 631 + ], + "type": "text", + "content": "5.2 Main Results" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 641, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 482, + 665 + ], + "type": "text", + "content": "Performance of previous state-of-the-art. In Tab. 1 we report the quantitative results of standard video-text retrieval on MSR-VTT and VATEX, and of" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 205, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 205, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 205, + 91, + 447, + 102 + ], + "type": "text", + "content": "Rethinking Video-Text Understanding: RCAD and Feint6K" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 248 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 248 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 248 + ], + "type": "text", + "content": "retrieval from counterfactually augmented data on our Feint6K. We also report a \"Random\" performance where a model predicts random guesses. We make the following observations: (i) Results show that previous state-of-the-art video-text models demonstrate limited understanding of the action semantics in a video with less than " + }, + { + "bbox": [ + 130, + 116, + 482, + 248 + ], + "type": "inline_equation", + "content": "60\\%" + }, + { + "bbox": [ + 130, + 116, + 482, + 248 + ], + "type": "text", + "content": " R@1 accuracy on RCAD, given a " + }, + { + "bbox": [ + 130, + 116, + 482, + 248 + ], + "type": "inline_equation", + "content": "16.7\\%" + }, + { + "bbox": [ + 130, + 116, + 482, + 248 + ], + "type": "text", + "content": " R@1 accuracy when taking random guesses. (ii) Although it is hard to compare the quantitative results between two tasks, we note that our RCAD is much more challenging given the performance gap between \"Random\" and previous state-of-the-arts. This supports our previous arguments that our RCAD removes shortcuts from standard video-text datasets and focuses on harder questions in the video domain that require cross-frame reasoning." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 266, + 483, + 423 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 266, + 483, + 423 + ], + "spans": [ + { + "bbox": [ + 130, + 266, + 483, + 423 + ], + "type": "text", + "content": "Performance of LLM-teacher. We apply our LLM-teacher to two pretrained video-text models, SimVTP [16] and InternVideo [23]. Specifically, we consider the two objectives in Eq. (1) and Eq. (2), labeled as \"LLM-teacher-lbl\" and \"LLM-teacher-lgt\" respectively. Results in Tab. 1 show that by exploiting knowledge from pretrained LLMs, LLM-teacher enables a more efficient leaning of action semantics and achieved improved results on various metrics of RCAD, increasing the R@1 accuracy on RCAD by " + }, + { + "bbox": [ + 130, + 266, + 483, + 423 + ], + "type": "inline_equation", + "content": "7.2\\%" + }, + { + "bbox": [ + 130, + 266, + 483, + 423 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 266, + 483, + 423 + ], + "type": "inline_equation", + "content": "7.4\\%" + }, + { + "bbox": [ + 130, + 266, + 483, + 423 + ], + "type": "text", + "content": ". We also note that with binary pseudo-labels, results demonstrate a trade-off between the performance on standard video-to-text retrieval and RCAD; and in comparison, LLM-teacher with soft logits achieves the highest performance on RCAD (" + }, + { + "bbox": [ + 130, + 266, + 483, + 423 + ], + "type": "inline_equation", + "content": "65.8\\%" + }, + { + "bbox": [ + 130, + 266, + 483, + 423 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 266, + 483, + 423 + ], + "type": "inline_equation", + "content": "65.6\\%" + }, + { + "bbox": [ + 130, + 266, + 483, + 423 + ], + "type": "text", + "content": ") with negligible drops (" + }, + { + "bbox": [ + 130, + 266, + 483, + 423 + ], + "type": "inline_equation", + "content": "0.2\\%" + }, + { + "bbox": [ + 130, + 266, + 483, + 423 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 266, + 483, + 423 + ], + "type": "inline_equation", + "content": "0.6\\%" + }, + { + "bbox": [ + 130, + 266, + 483, + 423 + ], + "type": "text", + "content": ") on standard retrieval. This is because soft logits account for the text similarities between different actions and avoid overfitting on binary pseudo-labels." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 442, + 483, + 562 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 442, + 483, + 562 + ], + "spans": [ + { + "bbox": [ + 130, + 442, + 483, + 562 + ], + "type": "text", + "content": "Comparison with human-level performance. The human-level performance are reported in Tab. 1. We see that human annotators achieved an almost perfect performance on Feint6K dataset, with a " + }, + { + "bbox": [ + 130, + 442, + 483, + 562 + ], + "type": "inline_equation", + "content": "95.2\\%" + }, + { + "bbox": [ + 130, + 442, + 483, + 562 + ], + "type": "text", + "content": " accuracy on MSR-VTT videos and a " + }, + { + "bbox": [ + 130, + 442, + 483, + 562 + ], + "type": "inline_equation", + "content": "96.8\\%" + }, + { + "bbox": [ + 130, + 442, + 483, + 562 + ], + "type": "text", + "content": " accuracy on VATEX videos. These results show that: (i) most questions for retrieval from counterfactually-augmented data are answerable with an unique answer, and (ii) this task is fairly simple for human. As a comparison, all previous video-text models with large-scale pretraining fall behind by a wide margin. Specifically, the state-of-the-art video-text model InternVideo [23] features a heavy two-stage pretraining on a collection of 7 large-scale video-text datasets but falls behind by " + }, + { + "bbox": [ + 130, + 442, + 483, + 562 + ], + "type": "inline_equation", + "content": "36.6\\%" + }, + { + "bbox": [ + 130, + 442, + 483, + 562 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 442, + 483, + 562 + ], + "type": "inline_equation", + "content": "38.6\\%" + }, + { + "bbox": [ + 130, + 442, + 483, + 562 + ], + "type": "text", + "content": " even with downstream finetuning." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 582, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 482, + 665 + ], + "type": "text", + "content": "Analyses of failure cases. In Fig. 1b we present two failure cases by the InternVideo [23] model on our Feint6K dataset. Both examples are trivial for human, yet challenging from video-text models due to incapability of cross-frame reasoning. For the first example, a model must analyze the interactions between the person and the rock over a sequence of frames to predict the correct answer. And in the second example, the model must observe the change of appearances over time. As we can see, with the counterfactually augmented data in our Feint6K," + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "text", + "content": "W. Ma et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 177 + ], + "type": "text", + "content": "models cannot \"guess\" the correct answer by exploiting shortcuts extracted from a single frame. The findings on Feint6K highlight significant weaknesses in current video-text models, offering valuable insights for future researches in this area. Additional qualitative examples of video-text models on Feint6K dataset are provided in the supplementary materials for readers' reference." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 193, + 482, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 193, + 482, + 312 + ], + "spans": [ + { + "bbox": [ + 130, + 193, + 482, + 312 + ], + "type": "text", + "content": "Sensitivity to the change of object or action. With the counterfactually augmented data collected, we investigate how the cosine similarities between the video and text change when the object or action in the text are swapped with an object or action not present in the image. Given the original video-text pair " + }, + { + "bbox": [ + 130, + 193, + 482, + 312 + ], + "type": "inline_equation", + "content": "(v,t)" + }, + { + "bbox": [ + 130, + 193, + 482, + 312 + ], + "type": "text", + "content": " and a counterfactually-augmented text " + }, + { + "bbox": [ + 130, + 193, + 482, + 312 + ], + "type": "inline_equation", + "content": "\\hat{t}" + }, + { + "bbox": [ + 130, + 193, + 482, + 312 + ], + "type": "text", + "content": ", the change in cosine similarity is given by " + }, + { + "bbox": [ + 130, + 193, + 482, + 312 + ], + "type": "inline_equation", + "content": "\\Delta s = s(v,\\hat{t}) - s(v,t)" + }, + { + "bbox": [ + 130, + 193, + 482, + 312 + ], + "type": "text", + "content": ". As " + }, + { + "bbox": [ + 130, + 193, + 482, + 312 + ], + "type": "inline_equation", + "content": "\\hat{t}" + }, + { + "bbox": [ + 130, + 193, + 482, + 312 + ], + "type": "text", + "content": " contains objects or actions not present in the video, " + }, + { + "bbox": [ + 130, + 193, + 482, + 312 + ], + "type": "inline_equation", + "content": "\\Delta s" + }, + { + "bbox": [ + 130, + 193, + 482, + 312 + ], + "type": "text", + "content": " should be negative for an ideal video-text model. Moreover, larger " + }, + { + "bbox": [ + 130, + 193, + 482, + 312 + ], + "type": "inline_equation", + "content": "|\\Delta s|" + }, + { + "bbox": [ + 130, + 193, + 482, + 312 + ], + "type": "text", + "content": " implies the model being more sensitive to the changes. We compute " + }, + { + "bbox": [ + 130, + 193, + 482, + 312 + ], + "type": "inline_equation", + "content": "\\Delta s" + }, + { + "bbox": [ + 130, + 193, + 482, + 312 + ], + "type": "text", + "content": " using the InternVideo with pretraining only (\"IntVid-pt\"), InternVideo with downstream finetuning (\"IntVid-ft\"), and our LLM-teacher (\"LLM-teacher\")." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 313, + 482, + 420 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 313, + 482, + 420 + ], + "spans": [ + { + "bbox": [ + 130, + 313, + 482, + 420 + ], + "type": "text", + "content": "Results in Fig. 5a show that the " + }, + { + "bbox": [ + 130, + 313, + 482, + 420 + ], + "type": "inline_equation", + "content": "\\Delta s" + }, + { + "bbox": [ + 130, + 313, + 482, + 420 + ], + "type": "text", + "content": " are almost always negative when the objects are swapped and the absolute changes are larger. In comparison, " + }, + { + "bbox": [ + 130, + 313, + 482, + 420 + ], + "type": "inline_equation", + "content": "\\Delta s" + }, + { + "bbox": [ + 130, + 313, + 482, + 420 + ], + "type": "text", + "content": " are sometimes positive when actions are modified and the absolute changes are smaller. This demonstrate that InternVideo learns a more effective embedding for objects than for actions. This further highlights the necessity of our new evaluation paradigm for exploring the limitations of video-text models beyond existing tasks. Moreover, in Fig. 5b we note that with LLM-teacher the action embeddings are more effective than the ones learned by InternVideo, as shown by the more reasonable changes in cosine similarity when actions are swapped." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 420, + 481, + 444 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 420, + 481, + 444 + ], + "spans": [ + { + "bbox": [ + 130, + 420, + 481, + 444 + ], + "type": "text", + "content": "Please refer to the supplementary materials for a detailed discussion regarding the influence of textual encoders." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 461, + 238, + 473 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 461, + 238, + 473 + ], + "spans": [ + { + "bbox": [ + 131, + 461, + 238, + 473 + ], + "type": "text", + "content": "5.3 Ablation Study" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 480, + 482, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 480, + 482, + 529 + ], + "spans": [ + { + "bbox": [ + 130, + 480, + 482, + 529 + ], + "type": "text", + "content": "We conduct ablation studies on the VATEX dataset [22] and our Feint6K dataset. We follow the same settings as above and report rank-1 (R@1) accuracy for standard video-to-text retrieval and rank-1 accuracy (R@1), rank-2 accuracy (R@2), and mean rank (MeanR) for counterfactual augmented data." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": "Choice of captions in LLM-teacher. Tab. 2 shows the comparison between different numbers and types of LLM captions generated. In the \"default\" setting we use 10 action-based LLM captions and compare to settings with 5 action-based LLM captions or 5 action-based and 5 object-based LLM captions. Results show with 5 action-based LLM captions the R@1 accuracy on Feint6K drops by " + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "inline_equation", + "content": "0.9\\%" + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": " while the R@1 accuracy on VATEX increases by " + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "inline_equation", + "content": "0.3\\%" + }, + { + "bbox": [ + 130, + 545, + 482, + 666 + ], + "type": "text", + "content": ". We also experiment on object-based LLM captions and find them not beneficial for both standard video-to-text retrieval or our RCAD. This is consistent with our assumption that pretrained video-text models already learns a discriminative embedding for objects and a more efficient training objective for action is needed." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 205, + 90, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 205, + 90, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 205, + 90, + 447, + 102 + ], + "type": "text", + "content": "Rethinking Video-Text Understanding: RCAD and Feint6K" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 176, + 114, + 440, + 200 + ], + "blocks": [ + { + "bbox": [ + 176, + 114, + 440, + 200 + ], + "lines": [ + { + "bbox": [ + 176, + 114, + 440, + 200 + ], + "spans": [ + { + "bbox": [ + 176, + 114, + 440, + 200 + ], + "type": "table", + "html": "
ModelVATEXFeint6K (VATEX)
R@1R@1R@2MeanR
Default(10 action captions; XLM-RoBERTa)87.365.681.71.7
5 action captions87.6 ↑0.364.7 ↓0.981.0 ↓0.71.70.0
5 object + 5 object captions87.5 ↑0.264.2 ↓1.480.6 ↓1.11.70.0
LLM Chatbot87.0 ↓0.365.9 ↑0.381.8 ↑0.11.70.0
", + "image_path": "00452429d88f66ca60d637193a99280c5245a7fe844cae2ccf3e65b266f9e729.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 207, + 482, + 242 + ], + "lines": [ + { + "bbox": [ + 132, + 207, + 482, + 242 + ], + "spans": [ + { + "bbox": [ + 132, + 207, + 482, + 242 + ], + "type": "text", + "content": "Table 2: Ablation studies on the number of LLM captions and caption generation methods. In the default setting we use a total of 10 action-based LLM captions and utilize a pretrained XLM-RoBERTa [4] for mask filling." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 130, + 262, + 482, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 262, + 482, + 396 + ], + "spans": [ + { + "bbox": [ + 130, + 262, + 482, + 396 + ], + "type": "text", + "content": "Choice of caption generation. Besides a pretrained XLM-RoBERTa model for mask filling, we also experiment on generating captions with LLM-powered agents that are finetuned on chat datasets for dialogue applications. Specifically we use the \"chat\" model finetuned for dialogue applications. Empirically we find that captions generated by LLM chatbots achieve a higher overall quality, exploring a more diverse caption space and better flexibility. However we don't observe significant improvements from our ablation study experiments – the R@1 accuracy on VATEX drops by " + }, + { + "bbox": [ + 130, + 262, + 482, + 396 + ], + "type": "inline_equation", + "content": "0.3\\%" + }, + { + "bbox": [ + 130, + 262, + 482, + 396 + ], + "type": "text", + "content": " and the R@1 accuracy on Feint6K increases by " + }, + { + "bbox": [ + 130, + 262, + 482, + 396 + ], + "type": "inline_equation", + "content": "0.3\\%" + }, + { + "bbox": [ + 130, + 262, + 482, + 396 + ], + "type": "text", + "content": ". We choose XLM-RoBERTa in our main experiments as the model runs faster and scale up easily to bigger settings. We provide qualitative comparisons between the two types of caption generation in our supplementary materials." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 411, + 227, + 424 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 411, + 227, + 424 + ], + "spans": [ + { + "bbox": [ + 132, + 411, + 227, + 424 + ], + "type": "text", + "content": "6 Conclusions" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 434, + 483, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 434, + 483, + 615 + ], + "spans": [ + { + "bbox": [ + 130, + 434, + 483, + 615 + ], + "type": "text", + "content": "In this work we propose a new evaluation task, retrieval from counterfactually augmented data, and a benchmark dataset, Feint6K. The idea is to remove shortcuts in the video-text questions with a human-in-the-loop system and produce more challenging questions where the video-text model must derive a comprehensive understanding of the video with cross-frame reasoning. Quantitative and qualitative evaluation results on our Feint6K dataset show that despite the task is trivial for human, previous state-of-the-art video-text models can be easily fooled by the counterfactually augmented data. This implies that the prominent results on previous video-text benchmarks could be misleading - models may largely exploit shortcuts without genuinely understand the video contents. Moreover, we identify a key limitation of current contrastive learning on videotext data being the shortcut learning and propose LLM-teacher that enables a more effective learning of action semantics by utilizing knowledge from pretrained LLMs. Experimental results and analyses show that our method can learn a more discriminative representation for actions." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 629, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 629, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 629, + 482, + 666 + ], + "type": "text", + "content": "Supplementary materials. We present the following: (1) limitations of our work, (2) more details about our Feint6K dataset, (3) ethics statements, and (4) extra quantitative and qualitative results." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 101 + ], + "type": "text", + "content": "W. Ma et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 482, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 482, + 152 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 482, + 152 + ], + "type": "text", + "content": "Acknowledgements. We would like to thank Daniel Khashabi, Yiyan Li, and the anonymous reviewers for their helpful comments and suggestions. Wufei Ma is supported by ONR with N00014-23-1-2641." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 170, + 197, + 182 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 170, + 197, + 182 + ], + "spans": [ + { + "bbox": [ + 133, + 170, + 197, + 182 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 194, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 138, + 194, + 480, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 194, + 480, + 227 + ], + "spans": [ + { + "bbox": [ + 138, + 194, + 480, + 227 + ], + "type": "text", + "content": "1. Bain, M., Nagrani, A., Varol, G., Zisserman, A.: Frozen in time: A joint video and image encoder for end-to-end retrieval. In: IEEE International Conference on Computer Vision (2021) 3, 11" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 228, + 480, + 269 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 228, + 480, + 269 + ], + "spans": [ + { + "bbox": [ + 138, + 228, + 480, + 269 + ], + "type": "text", + "content": "2. Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J.D., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., et al.: Language models are few-shot learners. Advances in neural information processing systems 33, 1877-1901 (2020) 9" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 271, + 480, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 271, + 480, + 304 + ], + "spans": [ + { + "bbox": [ + 138, + 271, + 480, + 304 + ], + "type": "text", + "content": "3. Cheng, F., Wang, X., Lei, J., Crandall, D., Bansal, M., Bertasius, G.: Vindlu: A recipe for effective video-and-language pretraining. arXiv preprint arXiv:2212.05051 (2022) 11" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 304, + 480, + 337 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 304, + 480, + 337 + ], + "spans": [ + { + "bbox": [ + 138, + 304, + 480, + 337 + ], + "type": "text", + "content": "4. Conneau, A., Khandelwal, K., Goyal, N., Chaudhary, V., Wenzek, G., Guzmán, F., Grave, E., Ott, M., Zettlemoyer, L., Stoyanov, V.: Unsupervised cross-lingual representation learning at scale. arXiv preprint arXiv:1911.02116 (2019) 11, 14" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 338, + 480, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 338, + 480, + 370 + ], + "spans": [ + { + "bbox": [ + 138, + 338, + 480, + 370 + ], + "type": "text", + "content": "5. Devlin, J., Chang, M.W., Lee, K., Toutanova, K.: Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805 (2018) 4, 9" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 370, + 480, + 412 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 370, + 480, + 412 + ], + "spans": [ + { + "bbox": [ + 138, + 370, + 480, + 412 + ], + "type": "text", + "content": "6. Fabian Caba Heilbron, Victor Escorcia, B.G., Niebles, J.C.: Activitynet: A large-scale video benchmark for human activity understanding. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 961-970 (2015) 4, 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 414, + 480, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 414, + 480, + 446 + ], + "spans": [ + { + "bbox": [ + 138, + 414, + 480, + 446 + ], + "type": "text", + "content": "7. Gupta, T., Kembhavi, A.: Visual programming: Compositional visual reasoning without training. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14953-14962 (2023) 9" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 447, + 480, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 447, + 480, + 468 + ], + "spans": [ + { + "bbox": [ + 138, + 447, + 480, + 468 + ], + "type": "text", + "content": "8. Hinton, G., Vinyals, O., Dean, J.: Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531 (2015) 10" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 468, + 480, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 468, + 480, + 499 + ], + "spans": [ + { + "bbox": [ + 138, + 468, + 480, + 499 + ], + "type": "text", + "content": "9. Kaushik, D., Hovy, E., Lipton, Z.C.: Learning the difference that makes a difference with counterfactually-augmented data. arXiv preprint arXiv:1909.12434 (2019) 2, 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 501, + 480, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 501, + 480, + 533 + ], + "spans": [ + { + "bbox": [ + 138, + 501, + 480, + 533 + ], + "type": "text", + "content": "0. Kay, W., Carreira, J., Simonyan, K., Zhang, B., Hillier, C., Vijayanarasimhan, S., Viola, F., Green, T., Back, T., Natsev, P., et al.: The kinetics human action video dataset. arXiv preprint arXiv:1705.06950 (2017) 4, 5" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 534, + 480, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 534, + 480, + 566 + ], + "spans": [ + { + "bbox": [ + 138, + 534, + 480, + 566 + ], + "type": "text", + "content": "1. Krishna, R., Hata, K., Ren, F., Fei-Fei, L., Carlos Niebles, J.: Dense-captioning events in videos. In: Proceedings of the IEEE international conference on computer vision. pp. 706-715 (2017) 4, 5" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 567, + 480, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 567, + 480, + 599 + ], + "spans": [ + { + "bbox": [ + 138, + 567, + 480, + 599 + ], + "type": "text", + "content": "2. Li, K., Wang, Y., Li, Y., Wang, Y., He, Y., Wang, L., Qiao, Y.: Unmasked teacher: Towards training-efficient video foundation models. arXiv preprint arXiv:2303.16058 (2023) 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 600, + 480, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 600, + 480, + 621 + ], + "spans": [ + { + "bbox": [ + 138, + 600, + 480, + 621 + ], + "type": "text", + "content": "3. Lin, J., Gong, S.: Gridclip: One-stage object detection by grid-level clip representation learning. arXiv preprint arXiv:2303.09252 (2023) 8" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 621, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 621, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 138, + 621, + 480, + 632 + ], + "type": "text", + "content": "4. Liu, H., Li, C., Wu, Q., Lee, Y.J.: Visual instruction tuning. In: NeurIPS (2023) 8" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 632, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 632, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 632, + 480, + 665 + ], + "type": "text", + "content": "5. Luo, H., Ji, L., Zhong, M., Chen, Y., Lei, W., Duan, N., Li, T.: Clip4clip: An empirical study of clip for end to end video clip retrieval and captioning. Neurocomputing 508, 293-304 (2022) 4, 10, 11" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 205, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 205, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 205, + 91, + 447, + 102 + ], + "type": "text", + "content": "Rethinking Video-Text Understanding: RCAD and Feint6K" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 117, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 133, + 117, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 117, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 133, + 117, + 480, + 138 + ], + "type": "text", + "content": "16. Ma, Y., Yang, T., Shan, Y., Li, X.: Simvtp: Simple video text pre-training with masked autoencoders. arXiv preprint arXiv:2212.03490 (2022) 1, 4, 7, 11, 12" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 140, + 480, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 140, + 480, + 172 + ], + "spans": [ + { + "bbox": [ + 133, + 140, + 480, + 172 + ], + "type": "text", + "content": "17. Miech, A., Zhukov, D., Alayrac, J.B., Tapaswi, M., Laptev, I., Sivic, J.: HowTo100M: Learning a Text-Video Embedding by Watching Hundred Million Narrated Video Clips. In: ICCV (2019) 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 174, + 480, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 174, + 480, + 217 + ], + "spans": [ + { + "bbox": [ + 133, + 174, + 480, + 217 + ], + "type": "text", + "content": "18. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021) 4, 8, 10, 11" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 134, + 218, + 480, + 261 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 218, + 480, + 261 + ], + "spans": [ + { + "bbox": [ + 134, + 218, + 480, + 261 + ], + "type": "text", + "content": "19. Reimers, N., Gurevych, I.: Sentence-bert: Sentence embeddings using siamese bert-networks. In: Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing. Association for Computational Linguistics (11 2019), https://arxiv.org/abs/1908.10084 11" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 133, + 262, + 480, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 262, + 480, + 285 + ], + "spans": [ + { + "bbox": [ + 133, + 262, + 480, + 285 + ], + "type": "text", + "content": "20. Soomro, K., Zamir, A.R., Shah, M.: Ucf101: A dataset of 101 human actions classes from videos in the wild. arXiv preprint arXiv:1212.0402 (2012) 4, 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 285, + 480, + 318 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 285, + 480, + 318 + ], + "spans": [ + { + "bbox": [ + 132, + 285, + 480, + 318 + ], + "type": "text", + "content": "21. Tong, Z., Song, Y., Wang, J., Wang, L.: Videomae: Masked autoencoders are data-efficient learners for self-supervised video pre-training. Advances in neural information processing systems 35, 10078-10093 (2022) 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 319, + 480, + 362 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 319, + 480, + 362 + ], + "spans": [ + { + "bbox": [ + 132, + 319, + 480, + 362 + ], + "type": "text", + "content": "22. Wang, X., Wu, J., Chen, J., Li, L., Wang, Y.F., Wang, W.Y.: Latex: A large-scale, high-quality multilingual dataset for video-and-language research. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 4581-4591 (2019) 1, 2, 4, 5, 6, 7, 10, 11, 13" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 364, + 480, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 364, + 480, + 407 + ], + "spans": [ + { + "bbox": [ + 132, + 364, + 480, + 407 + ], + "type": "text", + "content": "23. Wang, Y., Li, K., Li, Y., He, Y., Huang, B., Zhao, Z., Zhang, H., Xu, J., Liu, Y., Wang, Z., et al.: Internvideo: General video foundation models via generative and discriminative learning. arXiv preprint arXiv:2212.03191 (2022) 1, 2, 3, 4, 7, 8, 11, 12" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 408, + 480, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 408, + 480, + 430 + ], + "spans": [ + { + "bbox": [ + 132, + 408, + 480, + 430 + ], + "type": "text", + "content": "24. Wu, Z., Yao, T., Fu, Y., Jiang, Y.G.: Deep learning for video classification and captioning. In: Frontiers of multimedia research, pp. 3-29 (2017) 2, 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 431, + 480, + 474 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 431, + 480, + 474 + ], + "spans": [ + { + "bbox": [ + 132, + 431, + 480, + 474 + ], + "type": "text", + "content": "25. Xu, D., Zhao, Z., Xiao, J., Wu, F., Zhang, H., He, X., Zhuang, Y.: Video question answering via gradually refined attention over appearance and motion. In: Proceedings of the 25th ACM international conference on Multimedia. pp. 1645-1653 (2017) 1, 4, 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 475, + 480, + 508 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 475, + 480, + 508 + ], + "spans": [ + { + "bbox": [ + 132, + 475, + 480, + 508 + ], + "type": "text", + "content": "26. Xu, H., Ghosh, G., Huang, P.Y., Okhonko, D., Aghajanyan, A., Metze, F., Zettlemoyer, L., Feichtenhofer, C.: Videoclip: Contrastive pre-training for zero-shot video-text understanding. arXiv preprint arXiv:2109.14084 (2021) 4, 7, 8" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 510, + 480, + 563 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 510, + 480, + 563 + ], + "spans": [ + { + "bbox": [ + 132, + 510, + 480, + 563 + ], + "type": "text", + "content": "27. Xu, H., Ghosh, G., Huang, P.Y., Okhonko, D., Aghajanyan, A., Metze, F., Zettlemoyer, L., Feichtenhofer, C.: VideoCLIP: Contrastive pre-training for zero-shot video-text understanding. In: Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing (EMNLP). Association for Computational Linguistics, Online (Nov 2021) 11" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 565, + 480, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 565, + 480, + 597 + ], + "spans": [ + { + "bbox": [ + 132, + 565, + 480, + 597 + ], + "type": "text", + "content": "28. Xu, J., Mei, T., Yao, T., Rui, Y.: Msr-vtt: A large video description dataset for bridging video and language. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 5288-5296 (2016) 1, 2, 4, 5, 6, 7, 10, 11" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 599, + 480, + 631 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 599, + 480, + 631 + ], + "spans": [ + { + "bbox": [ + 132, + 599, + 480, + 631 + ], + "type": "text", + "content": "29. Yan, S., Zhu, T., Wang, Z., Cao, Y., Zhang, M., Ghosh, S., Wu, Y., Yu, J.: Videotext modeling with zero-shot transfer from contrastive captioners. arXiv preprint arXiv:2212.04979 (2022) 4" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "type": "text", + "content": "30. Zhou, Z., Lei, Y., Zhang, B., Liu, L., Liu, Y.: Zegclip: Towards adapting clip for zero-shot semantic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11175-11185 (2023) 8" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "text", + "content": "W. Ma et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 482, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 482, + 173 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 482, + 173 + ], + "type": "text", + "content": "31. Zhu, B., Lin, B., Ning, M., Yan, Y., Cui, J., HongFa, W., Pang, Y., Jiang, W., Zhang, J., Li, Z., Zhang, C.W., Li, Z., Liu, W., Yuan, L.: Languagebind: Extending video-language pretraining to n-modality by language-based semantic alignment. In: The Twelfth International Conference on Learning Representations (2024), https://openreview.net/forum?id=QmZKc7UZCy 11" + } + ] + } + ], + "index": 2 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 205, + 90, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 205, + 90, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 205, + 90, + 447, + 102 + ], + "type": "text", + "content": "Rethinking Video-Text Understanding: RCAD and Feint6K" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Rethinking Weakly-supervised Video Temporal Grounding From a Game Perspective/4f2540b2-441f-49e9-a4f9-06c96b70a305_content_list.json b/2024/Rethinking Weakly-supervised Video Temporal Grounding From a Game Perspective/4f2540b2-441f-49e9-a4f9-06c96b70a305_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..88937dea12e2c1b0d43f855868b333ca0df6ddb3 --- /dev/null +++ b/2024/Rethinking Weakly-supervised Video Temporal Grounding From a Game Perspective/4f2540b2-441f-49e9-a4f9-06c96b70a305_content_list.json @@ -0,0 +1,2009 @@ +[ + { + "type": "text", + "text": "Rethinking Weakly-supervised Video Temporal Grounding From a Game Perspective", + "text_level": 1, + "bbox": [ + 230, + 140, + 772, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Xiang Fang $^{1*}$ , Zeyu Xiong $^{1*}$ , Wanlong Fang $^{1*}$ , Xiaoye Qu $^{1}$ , Chen Chen $^{2}$ , Jianfeng Dong $^{3}$ , Keke Tang $^{4}$ , Pan Zhou $^{1\\boxtimes}$ , Yu Cheng $^{5}$ , and Daizong Liu $^{6\\boxtimes}$", + "bbox": [ + 225, + 215, + 777, + 247 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Hubei Key Laboratory of Distributed System Security, Hubei Engineering Research", + "bbox": [ + 215, + 257, + 785, + 272 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Center on Big Data Security, School of Cyber Science and Engineering, Huazhong", + "bbox": [ + 228, + 273, + 772, + 287 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "University of Science and Technology", + "bbox": [ + 377, + 287, + 625, + 300 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "2 University of Central Florida", + "bbox": [ + 395, + 300, + 604, + 313 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3 Zhejiang Gongshang University", + "bbox": [ + 388, + 313, + 612, + 327 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "4 Guangzhou University", + "bbox": [ + 419, + 328, + 583, + 342 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "5 The Chinese University of Hong Kong", + "bbox": [ + 367, + 342, + 635, + 354 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "6 Peking University", + "bbox": [ + 433, + 356, + 568, + 369 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "xfang9508@gmail.com, zeyuxiong@hust.edu.cn, wanlongfang@gmail.com,", + "bbox": [ + 243, + 369, + 754, + 383 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "xiaoye@hust.edu.cn, chen.chen@crcv.ucf.edu, dongjf24@gmail.com,", + "bbox": [ + 256, + 385, + 740, + 397 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "tangbohutbh@gmail.com, panzhou@hust.edu.cn, chengyu@cse.cuhk.edu.hk,", + "bbox": [ + 238, + 398, + 761, + 411 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "dzliu@stu.pku.edu.cn", + "bbox": [ + 423, + 412, + 578, + 424 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. This paper addresses the challenging task of weakly-supervised video temporal grounding. Existing approaches are generally based on the moment proposal selection framework that utilizes contrastive learning and reconstruction paradigm for scoring the pre-defined moment proposals. Although they have achieved significant progress, we argue that their current frameworks have overlooked two indispensable issues: 1) Coarse-grained cross-modal learning: previous methods solely capture the global video-level alignment with the query, failing to model the detailed consistency between video frames and query words for accurately grounding the moment boundaries. 2) Complex moment proposals: their performance severely relies on the quality of proposals, which are also time-consuming and complicated for selection. To this end, in this paper, we make the first attempt to tackle this task from a novel game perspective, which effectively learns the uncertain relationship between each vision-language pair with diverse granularity and flexible combination for multi-level cross-modal interaction. Specifically, we creatively model each video frame and query word as game players with multivariate cooperative game theory to learn their contribution to the cross-modal similarity score. By quantifying the trend of frame-word cooperation within a coalition via the game-theoretic interaction, we are able to value all uncertain but possible correspondence between frames and words. Finally, instead of using moment proposals, we utilize the learned query-guided frame-wise scores for better moment localization. Experiments show that our method achieves superior performance on both Charades-STA and ActivityNet Caption datasets.", + "bbox": [ + 259, + 455, + 743, + 801 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Xiang Fang, Zeyu Xiong, and Wanlong Fang contributed equally to this work.", + "bbox": [ + 212, + 810, + 746, + 825 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "Corresponding authors: Pan Zhou and Daizong Liu.", + "bbox": [ + 212, + 825, + 576, + 839 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/1f60eb03a3c257228783a23a232c2c9f38ea109e0f0ad74cc5278ab11062fd33.jpg", + "image_caption": [ + "Fig. 1: (a) Most existing WS-VTG methods define moment proposals for query matching. They not only suffer from the coarse-grained alignment but also are limited to the quality of the proposals. (b) Instead of using proposals, our method reviews the WS-VTG from a brand-new game-interaction perspective to learn more fine-grained frame-aware self- and cross-modal alignment for accurate boundary localization." + ], + "image_footnote": [], + "bbox": [ + 217, + 145, + 558, + 243 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/1efc26c781866c4d7d0c1625873b46e5ad008da2e5befe11042c33ba90443ae0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 558, + 145, + 784, + 243 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 321, + 374, + 335 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Video temporal grounding (VTG) aims to localize the start and end timestamps of a specific event moment described by a given language query in an untrimmed video [2,32,39,75-77,99-102,113,119,120]. This task enables us to efficiently find video moments of human interest instead of traversing the entire video, which has broad application potential in video surveillance [13-15, 18, 24-26, 50-52, 59, 90-93, 125, 126], video summarization [12, 16, 17, 21-23, 36, 48, 94, 106, 107], etc. Most VTG methods [27-31, 53-55, 57, 58, 60-62, 65, 66, 127] follow the fully-supervised setting, where each frame is manually annotated as query-relevant or not. Although these methods achieve significant breakthroughs, they severely rely on extensive laborious manual annotations of moment boundaries, thus limiting their scalability and practicability in real-world applications. To alleviate the dense reliance, weakly-supervised VTG (WS-VTG) has received increasing attention [10, 49, 73, 88, 89, 123], which only requires the knowledge of matched video-query pairs without detailed frame-wise annotations, which is more challenging than the fully-supervised setting.", + "bbox": [ + 212, + 356, + 787, + 582 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Since there is no frame-wise annotation in WS-VTG, learning detailed frame-query alignment is difficult. As shown in Fig. 1(a), almost all existing WS-VTG solutions directly employ the proposal-based framework [10,33,73,89] that first generates multiple moment proposals and then learns the scores to indicate the potential alignment between video proposals and language query. Some approaches further utilize reconstruction-based paradigm [49,88,123] to minimize the reconstruction loss between the partially masked query and moment proposals to identify the proposal which best reconstructs the query. Although existing WS-VTG methods have achieved great progress, we argue that they still suffer from two inescapable limitations: 1) Firstly, as frame-wise annotation is not provided, these WS-VTG methods generally generate the moment proposals for all videos via sliding windows. In this way, they only model the coarse-grained cross-modal interaction of each proposal-query pair. Actually, in most cases, we expect to capture more fine-grained information for accurate boundary localization, such as how much the semantics of a specific frame helps or harms the query-guided cross-modal alignment. Unfortunately, existing methods rely on proposal-level cross-modal learning cannot achieve this goal. 2) Secondly, the", + "bbox": [ + 212, + 583, + 787, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "X. Fang, Z. Xiong, W. Fang et al.", + "bbox": [ + 271, + 114, + 496, + 128 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "performance of these proposal-based methods is severely limited by the quality of moment proposals as they devise moment proposals regardless of the specific contents and difficulty of each video.", + "bbox": [ + 212, + 146, + 782, + 191 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To tackle the above limitations, in this paper, we propose to address the WS-VTG task from a new game-theory perspective, which creatively models the subtle video frames and query words as game players to learn their uncertain but possible cross-modal interaction with diverse granularity and flexible combination. The game learning helps to generate query-guided frame-aware knowledge for fine-grained video grounding. As shown in Fig. 1(b), if visual representations and textual representations have strong semantic correspondence, they tend to cooperate together and contribute to a high cross-modal similarity score. By forming these cooperated frame-word representations as a coalition and quantifying the trend of cooperation within a coalition via the game-theoretic interaction index, we can not only learn the coalition contribution to the cross-modal semantic similarity, but also measure the additional benefits brought by the coalition compared with the costs of the lost coalitions of these players with others. In this manner, our cooperative WS-VTG game is able to value all possible correspondence between frames and words for sensitive and explainable cross-modal contrast. By formulating these detailed cross-modal correspondences as fine-grained query-guided frame-wise scores, we can efficiently generate the moment instead of using complicated proposals. In particular, we implement the cooperative game in both self-modal and cross-modal scenarios for self-modal attentive learning and cross-modal fine-grained alignment. We further extend the cross-modal game with multi-level query-semantic interaction for more comprehensive video understanding.", + "bbox": [ + 212, + 191, + 787, + 523 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our main contributions are summarized as follows:", + "bbox": [ + 238, + 523, + 607, + 537 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We reveal the limitations of existing WS-VTG methods, which rely on complex moment proposals and fail to capture fine-grained frame-word alignment for accurate boundary grounding. To this end, we propose a novel game-theory based framework to learn the detailed frame-wise query-relevance scores for efficiently constructing the accurate moment.", + "- We apply the game learning to both self- and cross-modal scenarios for contextual self-enhancement and cross-alignment. The cross-modal game is also extended to multi-level for comprehensive learning.", + "- To verify the effectiveness of our framework, we conduct experiments on two challenging WS-VTG benchmarks, i.e., Charades-STA and ActivityNet Caption. Experiments show that our method achieves superior performance. In representative cases, our method outperforms all compared methods by $4.53\\%$ on Charades-STA." + ], + "bbox": [ + 225, + 545, + 782, + 737 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Works", + "text_level": 1, + "bbox": [ + 215, + 763, + 395, + 779 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Fully-supervised video temporal grounding. Existing methods generally address the video temporal grounding in a fully-supervised manner, where both the annotations of video-sentence pairs and corresponding moment boundaries", + "bbox": [ + 212, + 794, + 782, + 839 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "Rethinking WS-VTG From a Game Perspective", + "bbox": [ + 411, + 114, + 730, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "are given. Most of them [32, 35, 37, 56, 61, 64, 66, 95, 103, 110, 111, 114, 116, 121, 122] utilize the proposal-based framework, which first integrate the sentence representation with those pre-defined moment proposals individually, and then evaluate their matching relationships. The proposal with the highest matching score is selected as the prediction. Instead of utilizing the moment proposals, recent proposal-free methods [8, 9, 74, 115] directly regress the temporal locations of the target moment.", + "bbox": [ + 212, + 146, + 782, + 251 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Weakly-supervised video temporal grounding. As manually annotating temporal boundaries of target moments is time-consuming, recent research attentions shift to weakly-supervised video temporal grounding [10, 73, 88, 89], which only requires video-level annotations. Mithun et al. [73] proposed the first weakly-supervised model to learn a joint embedding space for video and query representations. Gao et al. [33] developed a two-stream structure to measure the moment-query consistency and conduct moment selection simultaneously. Although the above methods have achieved promising performance, they are two-stage approaches that utilize multi-scale sliding windows to generate moment candidates, therefore suffering from inferior effectiveness and efficiency. To address this issue, [49, 68, 123] score all the moments sampled at different scales in a single pass and further improve the moment-sentence matching accuracy. However, almost all of them rely on moment proposals for matching and selection, which fail to capture and distinguish more fine-grained details among visually similar frames for acquiring more accurate moment boundaries.", + "bbox": [ + 212, + 252, + 785, + 478 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Cooperative game theory. Cooperative game theory focuses on the formation of coalitions and cooperation among rational players or groups to achieve common goals [1,6,40,41,47,69,80-82,96]. A typical cooperative game consists of a set of players with a game function. The game function maps all possible subsets of players, called group or coalition, to a number which represents the total payoff earned by these players working cooperatively to achieve the goal [19, 20, 46, 84]. To measure the contributions of each player and allocate different payoffs to these individuals fairly, a few researchers have proposed various concepts of value by computing the average added worth that one player brings to all possible coalitions, such as Shapley value [72, 85, 108] and Banzhaf value [4, 45, 78]. Recently, some concepts of cooperative game theory have been interpreted in the field of deep learning [11, 67, 118]. For example, Li et al. [47] propose a fine-grained image-text semantic alignment pre-training framework based on Shapley interaction. Different from these works, we propose for the first time to model the detailed consistency between video frames and query words by recasting multivariate game-theoretic indices to address more complicated WS-VTG task.", + "bbox": [ + 212, + 479, + 787, + 734 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3 Methodology", + "text_level": 1, + "bbox": [ + 215, + 760, + 380, + 779 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this section, we elaborate on the proposed method, which learns the uncertain and fine-grained cross-modal alignment in WS-VTG task. Specifically, our method conceptualizes the interaction between the video and query as a game,", + "bbox": [ + 212, + 794, + 785, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "X. Fang, Z. Xiong, W. Fang et al.", + "bbox": [ + 271, + 114, + 496, + 128 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/b26a4d14c5506599fbbb0c8864015dd78a5b7e1da1b8ceed3957a70cf47c5568.jpg", + "image_caption": [ + "Fig. 2: Overview of our framework, where \"SQAN\" denotes Sequential Query Attention Network. Specifically, we first conduct self-modal game to enhance the self-modal instance semantics, then devise multi-level cross-modal games to learn the fine-grained and uncertain cross-modal correspondence. At last, we predict the moment based on the query-guided frame-wise scores." + ], + "image_footnote": [], + "bbox": [ + 215, + 143, + 787, + 348 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where each frame/word acts as a player and forms coalitions. During the game learning, visual frame and linguistic word representations with strong semantic relevance will cooperate to form a new coalition, the benefits of which are quantified through the game interaction index. Our method maximizes payoff by strategically forming coalitions to achieve fine-grained semantic alignment between video and query. The overall framework is shown in Fig. 2.", + "bbox": [ + 212, + 433, + 787, + 525 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1 Preliminaries", + "text_level": 1, + "bbox": [ + 215, + 558, + 372, + 571 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Problem formulation. Generally, given an untrimmed video and a sentence query, WS-VTG aims to localize the specific video moment semantically related to the query with only the video-level annotation. Let $V = \\{v_{1},\\dots,v_{i}\\dots,v_{T}\\}$ denote the untrimmed video with $T$ frames, and $Q = \\{q_{1},\\dots,q_{j},\\dots,q_{M}\\}$ represent the sentence query containing $M$ words, where $v_{i}$ and $q_{j}$ are the $i$ -th frame and $j$ -th word, respectively. The objective function $\\mathcal{F}$ can be formulated as $\\mathcal{F}: (V,Q) \\to (\\tau_s,\\tau_e)$ , where $\\tau_s,\\tau_e$ denote the start and end timestamps of the video moment boundary in video $V$ semantically corresponding to query $Q$ .", + "bbox": [ + 212, + 595, + 787, + 717 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Video and query representations. To extract contextual video representation, following [32, 117], each input video $V$ is fed into a pre-trained 3D convolutional network [5, 97] with a stack of multi-head self-attention [98] layers for video encoding, then we get frame-wise video representation $\\mathbf{V} = \\{\\pmb{v}_i\\}_{i=1}^T \\in \\mathbb{R}^{T \\times D_h}$ , where $D_h$ is the hidden dimension. For query representation, we also follow [59, 117] to embed each word of the query by the Glove [83] model with a further multi-head self-attention module and a BiLSTM layer. Similarly, we can obtain the word-wise query representation $\\mathbf{Q} = \\{\\pmb{q}_j\\}_{j=1}^M \\in \\mathbb{R}^{M \\times D_h}$ .", + "bbox": [ + 212, + 719, + 787, + 843 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "Rethinking WS-VTG From a Game Perspective", + "bbox": [ + 411, + 114, + 730, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2 Game-theoretic Learning for Self-modal Semantic Enhancement", + "text_level": 1, + "bbox": [ + 215, + 146, + 784, + 161 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We argue that the instance relations within each modality are important to infer the consecutive visual event contents or linguistic phrase contexts. To achieve this self-modal learning, we review previous classic game-theoretic value [34, 85] to explore the weighted semantic contribution of each frame/word to the whole video/query semantics for enhancing the self-modal representation of both video and query.", + "bbox": [ + 212, + 179, + 782, + 268 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Visual self-modal game. Generally, a cooperative game consists of a player set $\\mathcal{P} = \\{1,2,\\dots,n\\}$ and a game function $g(\\cdot)$ . In detail, function $g(\\cdot)$ maps each player subset $S \\subseteq \\mathcal{P}$ to a score value, which indicates the payoff when all players in coalition $S$ work together in the game. To capture the inherent correlations within the video modality, we directly take the video frames $\\mathbf{V} = \\{\\pmb{v}_i\\}_{i=1}^T$ as frame players $\\mathcal{P} = \\mathbf{V}, n = T$ to play the game and evaluate frame-wise contribution to frame-level video semantics.", + "bbox": [ + 214, + 271, + 784, + 375 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Since different frames contribute and weight differently to video understanding, we attempt to adopt a game-theoretic value to measure the importance of each frame player in the cooperative game. Specifically, each frame player $\\pmb{v}_i$ in player set $\\pmb{V}$ has a weighted marginal contribution to the entire game, which can be measured by the game-theoretic value (e.g., Banzhaf value [34]) as:", + "bbox": [ + 212, + 377, + 784, + 448 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nB \\left(\\boldsymbol {v} _ {i} \\mid \\boldsymbol {V}\\right) = \\frac {1}{2 ^ {T - 1}} \\sum_ {\\mathcal {S} \\subseteq \\boldsymbol {V} \\backslash \\boldsymbol {v} _ {i}} [ g (\\mathcal {S} \\cup \\{\\boldsymbol {v} _ {i} \\}) - g (\\mathcal {S}) ], \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 352, + 462, + 784, + 494 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $S \\subseteq V \\backslash \\mathbf{v}_i$ represents a group formed without $\\mathbf{v}_i$ , $\\frac{1}{2^{T-1}}$ is the likelihood of $S$ being sampled, and $g(S \\cup \\{\\mathbf{v}_i\\}) - g(S)$ denotes the marginal contribution brought by player $\\mathbf{v}_i$ joining the coalition $S$ . $B(\\mathbf{v}_i | V)$ indicates the ability of player $\\mathbf{v}_i$ to influence the video-level semantic by calculating the weighted marginal contribution of player $\\mathbf{v}_i$ to all possible coalitions $S$ . Generally, within the context of our visual game, coalitions in the video modality represent consecutive frames of the same event or frames that are semantically similar. In this way, we can enumerate and learn all possible relations between the frames within the video for more contextual visual feature learning.", + "bbox": [ + 212, + 510, + 784, + 646 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Visual soft supervision. To assist the above game-theoretic value learning, we first develop an additional learnable video-domain relevance header to predict the same self-modal relevance $\\hat{R}^v = \\{\\hat{r}_i^v\\}_{i=1}^T = \\text{header}(\\mathbf{V})$ as Banzhaf values $\\{B(\\mathbf{v}_i|\\mathbf{V})\\}_{i=1}^T$ among all frames for soft supervision (more discussions will be illustrated later). Specifically, the header is implemented by: 1) 1D convolutional layers for local self-consistency modeling, 2) a self-attention module for global self-relevance capturing, 3) a convolutional layer for decoding. Then, we optimize the Kullback-Leibler Divergence (KLD) between the predicted relevance $\\hat{r}_i^v$ and the Banzhaf value $B(\\mathbf{v}_i|\\mathbf{V})$ for learning the consistent self-modal similarity. In detail, we define the game-guided frame-wise probability distribution as $D^v = [p_1^v,p_2^v,\\dots,p_i^v,\\dots,p_T^v]^\\top$ , where $p_i^v = \\frac{\\exp(B(\\mathbf{v}_i|\\mathbf{V}))}{\\sum_{l=1}^T\\exp(B(\\mathbf{v}_l|\\mathbf{V}))}$ . Similarly, the predicted frame-wise probability distribution calculated by the output of the self-relevance header can be denoted as $\\hat{D}^v = [\\hat{p}_1^v,\\hat{p}_2^v,\\dots,\\hat{p}_i^v,\\dots,\\hat{p}_T^v]^\\top$ , where", + "bbox": [ + 214, + 647, + 785, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "X. Fang, Z. Xiong, W. Fang et al.", + "bbox": [ + 271, + 114, + 496, + 128 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "$\\hat{p}_i^v = \\frac{\\exp(\\hat{r}_i^v)}{\\sum_{l=1}^{T}\\exp(\\hat{r}_l^v)}$ . Finally, the self-relevance loss $\\mathcal{L}_{self}^v$ for each video is:", + "bbox": [ + 212, + 143, + 730, + 167 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {s e l f} ^ {v} = \\operatorname {K L D} \\left(\\hat {D} ^ {v} | | D ^ {v}\\right) = - \\frac {1}{T} \\sum_ {i = 1} ^ {T} p _ {i} ^ {v} \\left(\\log \\hat {p} _ {i} ^ {v} - \\log p _ {i} ^ {v}\\right). \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 318, + 176, + 785, + 200 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Based on the learned game-theoretic value $B(\\pmb{v}_i|\\pmb{V})$ for each player $\\pmb{v}_i$ in player set $\\pmb{V}$ , we deploy a Softmax function to obtain the self-modal enhanced video features $\\widetilde{\\pmb{v}}_i$ as:", + "bbox": [ + 212, + 208, + 782, + 251 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\widetilde {\\boldsymbol {v}} _ {i} = \\boldsymbol {v} _ {i} + \\sum_ {\\mathcal {S} \\subseteq \\mathcal {V} \\backslash \\boldsymbol {v} _ {i}} \\operatorname {S o f t m a x} [ g (\\mathcal {S} \\cup \\{\\boldsymbol {v} _ {i} \\}) - g (\\mathcal {S}) ] \\cdot (\\mathcal {S} \\cup \\{\\boldsymbol {v} _ {i} \\}). \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 307, + 258, + 785, + 287 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The enhanced video features $\\widetilde{\\pmb{V}} = \\{\\widetilde{\\pmb{v}}_i\\}_{i=1}^T$ aggregates the relevant contexts from the whole video.", + "bbox": [ + 212, + 297, + 782, + 327 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Linguistic self-modal game. Similarly, we deem query words $\\mathbf{Q} = \\{\\mathbf{q}_j\\}_{j=1}^M$ as players $\\mathcal{P} = \\mathbf{Q}, n = M$ and calculate each word's marginal contribution $B(\\mathbf{q}_j|\\mathbf{Q})$ to the whole query $\\mathbf{Q}$ by formulating the linguistic self-modal game as:", + "bbox": [ + 212, + 328, + 782, + 373 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\nB \\left(\\boldsymbol {q} _ {j} \\mid \\boldsymbol {Q}\\right) = \\frac {1}{2 ^ {M - 1}} \\sum_ {\\mathcal {S} \\subseteq \\boldsymbol {Q} \\backslash \\boldsymbol {q} _ {j}} [ g (\\mathcal {S} \\cup \\{\\boldsymbol {q} _ {j} \\}) - g (\\mathcal {S}) ]. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 349, + 381, + 785, + 414 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Linguistic soft supervision. We first take the game probability distributions $D^{q} = [p_{1}^{q}, p_{2}^{q}, \\dots, p_{j}^{q}, \\dots, p_{M}^{q}]^{\\top}$ as the soft label, and then supervise the predicted distributions $\\hat{D}^{q} = [\\hat{p}_{1}^{q}, \\hat{p}_{2}^{q}, \\dots, \\hat{p}_{j}^{q}, \\dots, \\hat{p}_{M}^{q}]^{\\top}$ of another header by KLD loss $\\mathcal{L}_{self}^{q}$ similar to Eq. (2). At last, we can obtain the enhanced query features $\\tilde{\\boldsymbol{Q}} = \\{\\tilde{\\boldsymbol{q}}_{j}\\}_{j=1}^{M}$ by:", + "bbox": [ + 214, + 422, + 782, + 503 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\widetilde {\\boldsymbol {q}} _ {j} = \\boldsymbol {q} _ {j} + \\sum_ {\\mathcal {S} \\subseteq \\mathcal {Q} \\backslash \\boldsymbol {q} _ {j}} \\operatorname {S o f t m a x} [ g (\\mathcal {S} \\cup \\{\\boldsymbol {q} _ {j} \\}) - g (\\mathcal {S}) ] \\cdot (\\mathcal {S} \\cup \\{\\boldsymbol {q} _ {j} \\}). \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 307, + 513, + 785, + 545 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Discussion on learnable header. Although we can directly utilize the game interaction value to enumerate the relations between different players via Eqs. (1) and (4), there are still two challenges: 1) No supervision signal: The performance of the game interaction severely depends on the quality of the learned player features. However, there is no supervision for learning the representative features. 2) Complex and time-consuming: The game process is complex and costs much time and resources. To this end, we develop a learnable header to mimic the game interaction by utilizing the KLD loss. This KLD function serves as a soft supervision label to not only learn the consistency between the game interaction and header, but also potentially train more distinguishing player features of each modality. During the inference, we can solely utilize the prediction of header to model the game interaction instead of Eqs. (1) and (4).", + "bbox": [ + 212, + 551, + 785, + 734 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.3 Game-theoretic Interaction for Cross-modal Multi-level Alignment", + "text_level": 1, + "bbox": [ + 214, + 755, + 723, + 785 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "After obtaining the self-enhanced video and query features, we further develop a cross-modal game between frames and words to handle their uncertainty during their fine-grained semantic alignment.", + "bbox": [ + 212, + 794, + 782, + 839 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "Rethinking WS-VTG From a Game Perspective", + "bbox": [ + 411, + 114, + 732, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Cross-modal game. From a game-theoretic perspective, we attempt to conceptualize the problem of fine-grained cross-modal alignment as a collaborative effort between video frames and query words. Specifically, when a video frame and a query word exhibit a high degree of semantic similarity, they are more likely to collaborate and their union will have a larger score contributed to the final moment. To this end, we take the enhanced video features $\\widetilde{V} = \\{\\widetilde{v}_i\\}_{i=1}^T$ and enhanced query features $\\widetilde{Q} = \\{\\widetilde{q}_j\\}_{j=1}^M$ as two different kinds of players $\\mathcal{P} = \\{\\widetilde{v}_i\\}_{i=1}^T \\cup \\{\\widetilde{q}_j\\}_{j=1}^M$ , $n = T + M$ . In a cooperative game, different players tend to work together in groups, called coalitions, to achieve a common goal, i.e., representing the same semantics of the target event. In the case of two players $\\widetilde{v}_i$ and $\\widetilde{q}_j$ in set $\\mathcal{P}$ , it may occur that $g(\\{\\widetilde{v}_i\\})$ and $g(\\{\\widetilde{q}_j\\})$ are small individually, but at the same time the reward of their forming coalition $g(\\{\\widetilde{v}_i, \\widetilde{q}_j\\})$ is considerable. This is because players in the coalition interact and collaborate with each other, which may bring additional benefits (or costs) to the game. Therefore, we also need to measure the additional benefits brought by the coalition compared with the costs of the lost coalitions of these players with others. To form a better coalition $\\{\\widetilde{v}_i, \\widetilde{q}_j\\} \\subseteq \\mathcal{P}$ , we utilize the effective interaction index $I_B(\\{\\widetilde{v}_i, \\widetilde{q}_j\\})$ like [34] to compute the additional benefit as follows:", + "bbox": [ + 212, + 146, + 787, + 406 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\nI _ {B} \\left(\\left\\{\\widetilde {\\boldsymbol {v}} _ {i}, \\widetilde {\\boldsymbol {q}} _ {j} \\right\\}\\right) = \\frac {1}{2 ^ {n - 2}} \\sum_ {\\mathcal {S} \\subseteq \\mathcal {P} \\backslash \\left\\{\\widetilde {\\boldsymbol {v}} _ {i}, \\widetilde {\\boldsymbol {q}} _ {j} \\right\\}} \\left[ g \\left(\\mathcal {S} \\cup \\left\\{\\widetilde {\\boldsymbol {v}} _ {i}, \\widetilde {\\boldsymbol {q}} _ {j} \\right\\}\\right) + g (\\mathcal {S}) - g \\left(\\mathcal {S} \\cup \\left\\{\\widetilde {\\boldsymbol {v}} _ {i} \\right\\}\\right) - g \\left(\\mathcal {S} \\cup \\left\\{\\widetilde {\\boldsymbol {q}} _ {j} \\right\\}\\right) \\right], \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 233, + 412, + 785, + 446 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $\\frac{1}{2^{n-2}}$ is the likelihood of $\\mathcal{S} \\subseteq \\mathcal{P} \\backslash \\{\\widetilde{v}_i, \\widetilde{q}_j\\}$ being sampled. Intuitively, $I_B(\\{\\widetilde{v}_i, \\widetilde{q}_j\\})$ quantifies this additional benefit compared to when players $\\widetilde{v}_i$ and $\\widetilde{q}_j$ work independently. It also embodies the tendency of interaction between frame-player and word-player. A higher value of $I_B(\\{\\widetilde{v}_i, \\widetilde{q}_j\\})$ signifies that players $\\widetilde{v}_i$ and $\\widetilde{q}_j$ cooperate closely with each other, and the formed coalition will bring additional high returns. By this interaction index, we can measure the closeness between plays $\\widetilde{v}_i$ and $\\widetilde{q}_j$ to achieve fine-grained alignment between video and query.", + "bbox": [ + 212, + 455, + 787, + 575 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Game details. To ensure complete consistency between the cooperative game and cross-modal alignment learning, the game function $g$ should meet the following criteria: 1) the game payoff benefits from strongly corresponding semantic pairs $\\{\\pmb{v}_i^+, \\pmb{q}_j^+\\}$ , i.e., $g(\\mathcal{P}) - g(\\mathcal{P} \\setminus \\{\\pmb{v}_i^+, \\pmb{q}_j^+\\} \\cup \\{[\\{\\pmb{v}_i^+, \\pmb{q}_j^+\\}]\\}) < 0$ ; 2) the game payoff is compromised by semantically opposite pairs $\\{\\pmb{v}_i^-, \\pmb{q}_j^-\\}$ , i.e., $g(\\mathcal{P}) - g(\\mathcal{P} \\setminus \\{\\pmb{v}_i^-, \\pmb{q}_j^-\\} \\cup \\{[\\{\\pmb{v}_i^-, \\pmb{q}_j^-\\}]\\}) > 0$ ; 3) when there are no players to cooperate, the payoff is zero, i.e., $g(\\{\\pmb{v}_i\\}_{i=1}^T) = g(\\{\\pmb{q}_j\\}_{j=1}^M) = g(\\phi) = 0$ , where $\\phi$ is the empty set. It should be noted that any function that satisfies the aforementioned conditions can be employed as the game function $g(\\cdot)$ . To simplify matters, we utilize the cosine similarity as $g$ in all games.", + "bbox": [ + 212, + 577, + 787, + 734 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "During the game, when players $\\mathbf{v}_i$ and $\\mathbf{q}_j$ form a coalition $\\{\\mathbf{v}_i, \\mathbf{q}_j\\}$ , we deem $[\\{\\mathbf{v}_i, \\mathbf{q}_j\\}]$ as a singular hypothetical player, which is the union of the player in $\\{\\mathbf{v}_i, \\mathbf{q}_j\\}$ . Then, the reduced game is formed by removing the individual players in $\\{\\mathbf{v}_i, \\mathbf{q}_j\\}$ from the game and adding $[\\{\\mathbf{v}_i, \\mathbf{q}_j\\}]$ to the game.", + "bbox": [ + 212, + 734, + 787, + 795 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Word-level alignment with local game interaction. Similar to the self-modal game, we develop an alignment header to predict the word-frame alignment matrix $A = [a_{i,j}]^{T \\times M}$ between frame $\\widetilde{\\boldsymbol{v}}_i$ and word $\\widetilde{\\boldsymbol{q}}_j$ , supervised by", + "bbox": [ + 212, + 795, + 787, + 842 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "X. Fang, Z. Xiong, W. Fang et al.", + "bbox": [ + 271, + 114, + 496, + 128 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "$I_B(\\{\\widetilde{v}_i,\\widetilde{q}_j\\})$ . Then, we optimize the Kullback-Leibler Divergence (KLD) between the $I_B(\\{\\widetilde{v}_i,\\widetilde{q}_j\\})$ and $a_{i,j}$ , to bring the probability distributions of the output of the alignment header and interaction index close together to establish word-level semantic alignment between frame players and word players.", + "bbox": [ + 212, + 145, + 782, + 205 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "In detail, we define the word-to-frame probability distribution $d_{w2f}^{j}$ of the $j$ -th word-to-frame alignment as $d_{w2f}^{j} = [p_{1,j}, \\dots, p_{t,j}]$ , where $p_{i,j} = \\frac{\\exp(I_B(\\{\\widetilde{v}_i, \\widetilde{q}_j\\}))}{\\sum_{l=1}^T \\exp(I_B(\\{\\widetilde{v}_l, \\widetilde{q}_j\\}))}$ and the total query-to-video probability distribution $D_{q2v} = [d_{w2f}^1, \\dots, d_{w2f}^M]^\\top$ . Similarly, the predicted fine-grained probability distribution can be represented as $\\hat{d}_{w2f}^{j} = [\\hat{p}_{1,j}, \\dots, \\hat{p}_{t,j}]$ , where $\\hat{p}_{i,j} = \\frac{\\exp(a_{i,j})}{\\sum_{l=1}^t \\exp(a_{l,j})}$ , and the total predicted probability distribution is defined as $\\hat{D}_{q2v} = [\\hat{d}_{w2f}^1, \\dots, \\hat{d}_{w2f}^M]^\\top$ . At last, the word-level alignment loss $\\mathcal{L}_{word}$ is formulated as:", + "bbox": [ + 212, + 207, + 787, + 332 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {w o r d}} = \\operatorname {K L D} \\left(\\hat {D} _ {q 2 v} | | D _ {q 2 v}\\right) = - \\frac {1}{T M} \\sum_ {i = 1} ^ {T} \\sum_ {j = 1} ^ {M} p _ {i, j} \\left(\\log \\hat {p} _ {i, j} - \\log p _ {i, j}\\right). \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 264, + 343, + 785, + 368 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Phrase- and sentence-level alignment with global game interaction. Although the above cross-modal game-theoretic interaction between each possible player-pair $\\{\\widetilde{v}_i,\\widetilde{q}_j\\}$ addresses the problem of word-level alignment between word player $\\widetilde{v}_i$ and frame player $\\widetilde{q}_j$ , it is still limited to aligning frames with isolated words, and fails to explore the relations between frames with more contextual phrase- or sentence-level semantics for better understanding temporal events. To this end, we introduce a multi-level game interaction to measure the semantic consistency between frames with different-grained textual semantics. Specifically, we follow [63, 74] to utilize Sequential Query Attention Network (SQAN) to extract the phrase-level textual semantics $\\tilde{\\mathbf{Q}}^p = \\{\\tilde{\\mathbf{q}}_j^p\\}_{j = 1}^P$ , and $P$ is the phrase number. The sentence-level semantics $\\tilde{\\mathbf{Q}}^s$ is obtained by concatenating the last hidden states in both the forward and backward BiLSTM. Similar to worl-level alignment, we compute both phrase- and sentence-level text-to-frame game interaction as $I_B(\\{\\widetilde{v}_i,\\widetilde{q}_j^p\\})$ , $I_B(\\{\\widetilde{v}_i,\\tilde{Q}^s\\})$ via Eq. (6). We design corresponding alignment headers (phrase-level cross-modal header and sentence-level cross-modal header) to predict their alignment matrix via $\\mathcal{L}_{phra},\\mathcal{L}_{sent}$ .", + "bbox": [ + 212, + 378, + 787, + 627 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "3.4 Model Training and Inference", + "text_level": 1, + "bbox": [ + 215, + 651, + 508, + 666 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Weak supervision for video-query matching. We first interact each frame with the semantics of words, phrases and the whole sentence to calculate the overall query-to-frame matching degree by:", + "bbox": [ + 212, + 678, + 784, + 722 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\nm _ {i} = \\frac {1}{3} \\left(I _ {B} \\left(\\{\\widetilde {\\boldsymbol {v}} _ {i}, \\widetilde {\\boldsymbol {Q}} ^ {s} \\}\\right) + \\frac {1}{M} \\sum_ {j = 1} ^ {M} I _ {B} \\left(\\{\\widetilde {\\boldsymbol {v}} _ {i}, \\widetilde {\\boldsymbol {q}} _ {j} \\}\\right) + \\frac {1}{P} \\sum_ {l = 1} ^ {P} I _ {B} \\left(\\left\\{\\widetilde {\\boldsymbol {v}} _ {i}, \\widetilde {\\boldsymbol {q}} _ {l} ^ {p} \\right\\}\\right)\\right), \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 263, + 732, + 785, + 758 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where $m_{i}$ is the overall query-to-frame matching degree of $i$ -th frame.", + "bbox": [ + 214, + 768, + 715, + 782 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Then, the whole video-sentence matching score is calculated by:", + "bbox": [ + 238, + 782, + 700, + 797 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\phi^ {c} = \\sum_ {i = 1} ^ {T} w _ {i} \\cdot m _ {i}, w _ {i} = \\frac {\\exp \\left(I _ {B} \\left(\\left\\{\\widetilde {\\boldsymbol {v}} _ {i} , \\widetilde {\\boldsymbol {Q}} ^ {s} \\right\\}\\right)\\right)}{\\sum_ {i = 1} ^ {T} \\exp \\left(I _ {B} \\left(\\left\\{\\widetilde {\\boldsymbol {v}} _ {i} , \\widetilde {\\boldsymbol {Q}} ^ {s} \\right\\}\\right)\\right)}, \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 331, + 811, + 785, + 844 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "Rethinking WS-VTG From a Game Perspective", + "bbox": [ + 411, + 114, + 732, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where we utilize the sentence-level similarity to represent the contribution of each frame to the target moment among the entire video. For matched video $V_{k}$ and query $Q_{k}$ , we calculate the video-sentence matching score $\\phi_{V_k,Q_k}^c$ . The scores of unmatched pairs are expressed as $\\phi_{V_k,Q_l}^c$ and $\\phi_{Q_k,V_l}^c$ where $l\\neq k$ .", + "bbox": [ + 212, + 146, + 782, + 208 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Finally, we utilize the cross-modal contrastive loss [79] for weak supervision:", + "bbox": [ + 238, + 208, + 784, + 223 + ], + "page_idx": 9 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {C} = - \\frac {1}{N _ {b}} \\sum_ {k = 1} ^ {N _ {b}} \\log \\frac {\\exp \\left(\\phi_ {V _ {k} , Q _ {k}} ^ {c} / \\tau\\right)}{\\sum_ {l = 1} ^ {N _ {b}} \\exp \\left(\\phi_ {V _ {k} , Q _ {l}} ^ {c} / \\tau\\right)} - \\frac {1}{N _ {b}} \\sum_ {k = 1} ^ {N _ {b}} \\log \\frac {\\exp \\left(\\phi_ {Q _ {k} , V _ {k}} ^ {c} / \\tau\\right)}{\\sum_ {l = 1} ^ {N _ {b}} \\exp \\left(\\phi_ {Q _ {k} , V _ {l}} ^ {c} / \\tau\\right)}, \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 228, + 241, + 782, + 275 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "where $\\tau$ is the temperature parameter, $N_{b}$ is the batch size.", + "bbox": [ + 214, + 281, + 645, + 295 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Reconstruction loss. To further enhance the semantic relevance of the highlighted frames to the sentence, we reconstruct the query conditioned on highlighted frames. Following popular reconstruction strategy [49,124], we randomly replace the 1/3 words in the original query with a specific symbol, and predict the next word by given a prefix of the query and highlighted frames feature. Specifically, we embed the masked query $\\mathbf{Q}^m$ by Glove and predict the next word $\\hat{\\pmb{q}}_{i+1}$ using the conditioned transformer following CNM [124] with input $(\\tilde{\\mathbf{V}}, C_m, Q_{1:i}^m)$ , where $C_m = \\{m_1, m_2, \\dots, m_T\\}$ represents the highlighted frame mask, $Q_{1:i}^m$ denotes the masked query words from the 1-st to $i$ -th.", + "bbox": [ + 212, + 296, + 784, + 431 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Then, we introduce the cross-entropy loss $\\mathcal{L}_{rec}$ to calculate the probability distribution $\\mathcal{P}^q$ different between the predict word $\\hat{\\pmb{q}}_{i + 1}$ and the real word as:", + "bbox": [ + 212, + 431, + 782, + 460 + ], + "page_idx": 9 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {r e c} = - \\sum_ {i = 1} ^ {M - 1} \\log \\mathcal {P} ^ {q} \\left(\\boldsymbol {q} _ {i + 1} \\mid \\hat {\\boldsymbol {q}} _ {i + 1}\\right). \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 385, + 468, + 782, + 491 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Overall training. To train the whole model, we combine game-guided self-relevance losses $\\mathcal{L}_{self} = \\mathcal{L}_{self}^v +\\mathcal{L}_{self}^q$ , game alignment loss $\\mathcal{L}_{align} = \\mathcal{L}_{word} + \\mathcal{L}_{phra} + \\mathcal{L}_{sent}$ , video-level contrastive loss $\\mathcal{L}_C$ , and reconstruction loss $\\mathcal{L}_{rec}$ with trade-off hyper-parameters $(\\alpha_{1},\\alpha_{2},\\alpha_{3}$ and $\\alpha_4)$ to define the overall loss as:", + "bbox": [ + 214, + 497, + 782, + 555 + ], + "page_idx": 9 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\alpha_ {1} \\mathcal {L} _ {\\text {s e l f}} + \\alpha_ {2} \\mathcal {L} _ {\\text {a l i g n}} + \\alpha_ {3} \\mathcal {L} _ {C} + \\alpha_ {4} \\mathcal {L} _ {\\text {r e c}}. \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 362, + 564, + 782, + 577 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Inference. During inference, we directly utilize the self-relevance headers to enhance the video and query features, and exploit the multi-level alignment headers to predict the overall query-to-frame matching degree. We locate the frame with the highest score as the basic predicted moment, and add the left/right frames into the moment if the ratio of their scores to the frame score of the closest moment boundary is higher than a threshold. We repeat this step until no frame can be added. We can obtain more moments by locating different initial frames with different scores.", + "bbox": [ + 212, + 585, + 784, + 704 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 214, + 727, + 375, + 744 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.1 Datasets and Evaluation Metrics", + "text_level": 1, + "bbox": [ + 214, + 756, + 532, + 771 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Charades-STA. The Charades-STA dataset [32] is built based on the Charades [86], which contains 6,672 videos of indoor activities and involves 16,128 query-video pairs. There are 12,408 pairs used for training and 3,720 used for testing. The average duration of each video is 29.76 seconds.", + "bbox": [ + 212, + 779, + 782, + 839 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "X. Fang, Z. Xiong, W. Fang et al.", + "bbox": [ + 271, + 114, + 496, + 128 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "ActivityNet Caption. The ActivityNet Caption dataset [43] contains 20,000 videos with 100,000 queries, where 37,421 query-video pairs are used for training and 34,536 are used for testing. On average, each video in ActivityNet Caption has 3.65 annotated moments and each annotated moment lasts for 36 seconds. For fair comparison, we follow [10, 73, 88, 89, 123, 124] to use val1 as valuation and report the test performance on val2.", + "bbox": [ + 212, + 146, + 782, + 234 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Evaluation metrics. Following previous works, we adopt the metrics “R@n, IoU=m” to evaluate our model, which presents the proportion of the top $n$ predicted moments with IoU larger than $m$ .", + "bbox": [ + 212, + 237, + 784, + 282 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4.2 Implementation Details", + "text_level": 1, + "bbox": [ + 214, + 303, + 455, + 318 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "For fair comparison, we follow previous works to apply C3D [97] to encode the videos on ActivityNet Caption and I3D [5] on Charades-STA. We uniformly downsample the length of video feature sequences to $T = 200$ for ActivityNet Caption and $T = 64$ for Charades-STA. As for query encoding, we also follow previous works to utilize set the length of word feature sequences to $M = 20$ , and utilize Glove [83] to embed each word. The dimension $D_h$ is set to 512. These trade-off hyper-parameters are set to $\\alpha_1 = 1.0, \\alpha_2 = 1.0, \\alpha_3 = 1.0, \\alpha_4 = 10$ . Since the calculation of the exact game-theoretic interaction is an NP-hard problem [71], we follow previous works to utilize sampling-based method [3,44] to obtain unbiased estimates for approximating it. We train the whole model for 100 epochs with batch size of 16 and early stopping strategy. Parameter optimization is performed by Adam [42] optimizer with learning rate $3 \\times 10^{-4}$ , and linear decay of learning rate and gradient clipping of 1.0. The inference threshold is set to 0.8 in ActivityNet Caption and 0.9 in Charades-STA.", + "bbox": [ + 212, + 327, + 784, + 537 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4.3 Main Results", + "text_level": 1, + "bbox": [ + 214, + 559, + 372, + 573 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Quantitative comparison. We compare our method with the state-of-the-art methods, including fully-supervised (FS) methods and weakly-supervised (WS) methods. Best results are in bold. We introduce two types of cooperative game indices, i.e., Banzhaf index [34] and Shapley index [85], to our model. As summarized in Tab. 1, both two different variants of our method surpass all existing methods on both Charades-STA and ActivityNet Captions. Particularly, on Charades-STA in terms of \"R@5, IoU=0.5\", Ours(Banzhaf) outperforms the best compared method CPL by $4.53\\%$ . The main reason is that our game-theoretic-based approach can learn the local alignment between words and frames for accurately predicting the moment boundaries. Besides, the Shapley variant performs similar to the Banzhaf variant for this specific WS-VTG task, since they share the same game interaction formulation with different normalization. Overall, this table verifies the effectiveness of our game-based framework.", + "bbox": [ + 212, + 583, + 784, + 777 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Generalization on cross-dataset evaluation. At last, we explore the generalization of models by cross-dataset evaluation in Tab. 2. Concretely, we first train a model on one dataset and then evaluate its performance on the other dataset. We can find that: 1) For all WS-VTG methods, the performance of", + "bbox": [ + 212, + 780, + 785, + 839 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "Rethinking WS-VTG From a Game Perspective", + "bbox": [ + 411, + 114, + 730, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 116, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/6a1648e71c33565161b93736f84c0a88b6eca6acfe597373e3d60ab363258d57.jpg", + "table_caption": [ + "Table 1: Comparisons with state-of-the-art methods. \"WS\" and \"FS\" denotes weakly-supervised and fully-supervised methods, respectively." + ], + "table_footnote": [], + "table_body": "
TypeMethodCharades-STAActivityNet Caption
R@1, IoU=0.5R@1, IoU=0.7R@5, IoU=0.5R@5, IoU=0.7R@1, IoU=0.3R@1, IoU=0.5R@5, IoU=0.3R@5, IoU=0.5
FSCTRL [32]23.638.8958.9229.52-29.01-59.17
2DTAN [121]39.8123.2579.3351.1559.4544.5185.5377.13
DRN [115]53.0931.7589.0660.05-45.45-77.97
WSSCN [49]23.589.9771.8038.8747.2329.2271.4555.69
WSTAN [104]29.3512.2876.1341.5352.4530.0179.3863.42
ICVC [7]31.0216.5377.5341.9146.6229.5280.9266.61
MARN [87]33.8715.5473.9041.9448.5231.3775.9160.00
CRM [38]34.7616.37--55.2632.19--
CNM [124]35.1514.95--55.6833.33--
ACN [109]37.0215.26--57.6634.18--
VCA [105]38.1319.5778.7537.7550.4531.0071.7953.83
LCNet [112]39.1918.1780.5645.2448.4926.3382.5162.66
CPL [125]49.2422.3984.7152.3755.7331.3763.0543.13
DM2 [70]51.3923.72--54.7331.85--
Ours(Banzhaf)54.1726.4688.9355.7958.9435.6586.4170.26
Ours(Shapley)54.3126.2889.2455.6259.1735.8086.5970.03
", + "bbox": [ + 225, + 172, + 787, + 431 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/ca98ec7bf1b6aa9e824541fc6d73e533f6209cdb8ece23b2943a85625af102c9.jpg", + "table_caption": [ + "Table 2: Cross-dataset evaluation. $A\\rightarrow C$ denotes training on ActivityNet Caption and evaluation on CharadesSTA, and vice versa." + ], + "table_footnote": [], + "table_body": "
MethodA→CC→A
R@1, IoU=0.7R@5, IoU=0.7R@1, IoU=0.5R@5, IoU=0.5
LCNet [112]12.0631.4710.6531.76
CPL [125]14.3235.8111.5929.88
Ours(Banzhaf)18.7442.5820.3742.96
Ours(Shapley)18.2643.1320.0441.42
", + "bbox": [ + 223, + 497, + 500, + 588 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/096937f3d5d36a18f2cabc2e31684e3bba05139f06634375f7d8d5d49b67561b.jpg", + "image_caption": [ + "Fig. 3: Qualitative results. The colors of green lines represent different degrees of confidence." + ], + "image_footnote": [], + "bbox": [ + 509, + 434, + 784, + 544 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "$\\mathrm{A} \\rightarrow \\mathrm{C}$ achieves better than $\\mathrm{C} \\rightarrow \\mathrm{A}$ . This is because that ActivityNet Caption is larger and more complex than Charades-STA, thus the former is able to bring more generalized knowledge to the latter while the latter fails to provide general knowledge to the former. 2) Our two variants outperform state-of-the-art methods (e.g., LCNet) with clear margins on cross-dataset evaluation. We speculate it to: Previous works severely rely on the moment proposals, however, LCNet defines different sliding windows on different datasets for proposal generation. Therefore, its non-uniform proposal structure limits its generalization on cross-dataset evaluation. Instead, we learn multi-level frame-aware alignment and utilize frame-wise scores for predicting moment, which is more flexible and adaptable, thus improving the generalization-ability.", + "bbox": [ + 212, + 597, + 785, + 763 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Qualitative comparison. In Fig. 3, we provide a qualitative example on ActivityNet Caption. The colors of green lines represent different degrees of interaction confidence, demonstrating that the game interaction can well learn the uncertain and fine-grained frame-word correspondence in the WS setting. Besides, all three-level interaction (word-level, phrase-level and sentence-level) contributes", + "bbox": [ + 212, + 763, + 785, + 839 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "X. Fang, Z. Xiong, W. Fang et al.", + "bbox": [ + 271, + 114, + 496, + 128 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/397665a7d8b62e35860f44918d9b05ad2294d4044d5cb695835f5e61819416e6.jpg", + "table_caption": [ + "Table 3: Comparison on speed per sample (s) and GPU memory cost (MB)." + ], + "table_footnote": [], + "table_body": "
MetricStageLCNet [112]CPL [125]Ours(Banzhaf)Ours(Shapley)
SpeedTraining0.79s0.27s0.94s0.98s
Inference0.62s0.21s0.10s0.11s
GPUTraining8647MB4156MB5237MB5482MB
MemoryInference7462MB3428MB2058MB2165MB
", + "bbox": [ + 236, + 159, + 761, + 232 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/e3f8afea7878f290bed22e84fbcbe5ad9ed6432213b2d4a281d798c7658f0133.jpg", + "table_caption": [ + "Table 4: Main ablation study." + ], + "table_footnote": [], + "table_body": "
Self-modal GameMulti-level Cross-modal GameBanzhaf-based GameShapley-based Game
Charades-STAActivityNet CaptionCharades-STAActivityNet Caption
Video Query GameWord- Phrase- Sentence- levelIoU=0.7IoU=0.5IoU=0.5R@1,R@5,R@1,R@5,R@1,R@5,R@1,R@5,
IoU=0.7IoU=0.5IoU=0.5IoU=0.5IoU=0.7IoU=0.5IoU=0.5IoU=0.5
××✓××19.3444.6729.8361.7219.2944.3130.1561.56
✓×✓××21.2547.8631.3963.9821.1047.7331.5463.67
✓✓✓××22.9750.6432.7265.8022.6950.3832.9765.65
✓✓✓✓×24.8153.2834.4668.0324.6453.0334.5967.82
✓✓✓✓✓26.4655.7935.6570.2626.2855.6235.8070.03
", + "bbox": [ + 220, + 256, + 779, + 371 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "differently to the grounding, and our final method is able to localize more accurate moment boundaries than state-of-the-art method CPL.", + "bbox": [ + 212, + 381, + 784, + 410 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Complexity comparison. In Tab. 3, our method costs more training time/memory as we need to enumerate coalition samples. Specifically, we sampled 5500 samples. We find that, the model instability decreases along with the increase of the sampling number, and when the sample size ranges from 4500 to 5500, the error with the true value of banzhaf gradually decreases from $6.9\\%$ to $1.1\\%$ . After exceeding 5500 samples, more samples will not affect the result. During inference, since we can directly utilize the learned game headers to generate game value instead of using complex game learning, our model is faster with lower memory.", + "bbox": [ + 212, + 412, + 807, + 532 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "4.4 Ablation Study", + "text_level": 1, + "bbox": [ + 215, + 556, + 388, + 571 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "In this section, we perform in-depth ablation studies to evaluate the effect of each component. We implement our model with different game settings.", + "bbox": [ + 212, + 582, + 785, + 612 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Main ablation. As shown in Tab. 4, we conduct ablation studies regarding the components (i.e., self-modal games in two modalities, and multi-level cross-modal games). We observe the following findings: 1) Our baseline model (only contains word-level cross-modal game) is powerful and outperforms some WS methods, demonstrating that it is effective to capture the uncertain yet fine-grained alignment for better determining the boundaries. 2) Both video and query self-modal games bring significant improvement to the overall model, indicating that self-enhancement is crucial for context integration. 3) Each query level of cross-modal game further boost the performance, verifying that phrase-and sentence-level semantics help to better understand temporal events. Overall, all components contribute a lot to final performance, showing their effectiveness.", + "bbox": [ + 212, + 613, + 785, + 777 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Effect of game headers. The headers in both self- and cross-modal games are utilized to provide soft supervision for better game interaction and player-representation learning. To explore the impact of the structure of these headers, we compare several popular structures in Tab. 5. We find that the combination of", + "bbox": [ + 212, + 779, + 785, + 839 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "Rethinking WS-VTG From a Game Perspective", + "bbox": [ + 411, + 114, + 730, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/7bcddf95c78911d5c6e75ceece003030b116e017d8ede6f10448b3b571fe9576.jpg", + "table_caption": [ + "Table 5: Ablation study on each component, where \"LL\" means \"Linear Layer\", \"CL\" means \"Convolution Layer\", \"SA\" means \"Self-attention\", \"MSE\" means \"Mean Squared Error\", \"CE\" means \"Cross Entropy\", \"KLD\" means \"Kullback-Leibler Divergence\"; \"RL\" means \"Reconstruction Loss\"." + ], + "table_footnote": [], + "table_body": "
ComponentSettingBanzhaf-based GameShapley-based Game
Charades-STAActivityNet CaptionCharades-STAActivityNet Caption
R@1, IoU=0.7R@5, IoU=0.7R@1, IoU=0.5R@5, IoU=0.5R@1, IoU=0.7R@5, IoU=0.7R@1, IoU=0.5R@5, IoU=0.5
Game HeadersLL25.2853.9034.1568.0324.9753.6834.2567.94
CL25.5254.3134.5668.6425.2654.0334.8268.47
LL + SA25.9454.8335.2769.1825.7154.5935.5369.06
CL + SA26.4655.7935.6570.2626.2855.6235.8070.03
Soft SupervisionNone23.2151.6432.3666.4022.9951.4532.7166.18
MSE24.4253.1933.8567.5724.2452.8940.0667.33
CE24.7353.5834.1668.0424.5253.2634.3767.80
KLD26.4655.7935.6570.2626.2855.6235.8070.03
Grounding Lossesw/o RL24.3553.5734.1468.4224.0853.3134.4568.19
w/ RL26.4655.7935.6570.2626.2855.6235.8070.03
", + "bbox": [ + 218, + 199, + 787, + 400 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "convolution layer and self-attention can capture both local and global interaction, so it achieves the best performance and is beneficial for cross-modal multi-level interaction, illustrating the significance of our game headers.", + "bbox": [ + 215, + 417, + 784, + 460 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Necessity of soft supervision. To explore the effect of soft supervision, we conduct the ablation study to investigate the contribution of the KLD loss during the header learning in Tab. 5. The KLD loss achieves better performance than both MSE and CE, showing that it is more suitable for game headers to mimic game-theoretic interaction learning. Moreover, without soft supervision, the performance will drop a lot, showing the importance of the soft supervision.", + "bbox": [ + 215, + 464, + 784, + 553 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Analysis on reconstruction loss. Most previous WS-VTG works also additionally utilize the reconstruction paradigm to assist the model training. As shown in Tab. 5, we further conduct such ablation study to investigate the impact of reconstruction loss. It shows that the reconstruction loss can improve significantly the performance, demonstrating its effectiveness.", + "bbox": [ + 215, + 556, + 784, + 630 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Conclusion", + "text_level": 1, + "bbox": [ + 217, + 662, + 356, + 679 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In this paper, we rethink the limitations of previous weakly-supervised video temporal grounding works, and creatively propose to utilize the game-theoretic interaction to learn the uncertain relationship between video-query pairs with diverse granularity. Specifically, we first introduce two self-modal games to correlate the frames/words to enhance their contextual semantics. Then, we propose a cross-modal game to value fine-grained correspondence between frames and words with multiple levels for better determining the query-related moment boundaries. Extensive experiments demonstrate the effectiveness of our game-based framework on two challenging datasets.", + "bbox": [ + 215, + 703, + 784, + 839 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "X. Fang, Z. Xiong, W. Fang et al.", + "bbox": [ + 271, + 114, + 496, + 128 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 143, + 321, + 159 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Albarelli, A., Rodola, E., Torsello, A.: A game-theoretic approach to fine surface registration without initial motion estimation. In: 2010 IEEE computer society conference on computer vision and pattern recognition. pp. 430-437. IEEE (2010)", + "2. Anne Hendricks, L., Wang, O., Shechtman, E., Sivic, J., Darrell, T., Russell, B.: Localizing moments in video with natural language. In: Proceedings of the IEEE International Conference on Computer Vision. pp. 5803-5812 (2017)", + "3. Bachrach, Y., Markakis, E., Resnick, E., Procaccia, A.D., Rosenschein, J.S., Saberi, A.: Approximating power indices: theoretical and empirical analysis. Autonomous Agents and Multi-Agent Systems 20, 105-122 (2010)", + "4. Banzhaf III, J.F.: Weighted voting doesn't work: A mathematical analysis. Rutgers L. Rev. 19, 317 (1964)", + "5. Carreira, J., Zisserman, A.: Quo vadis, action recognition? a new model and the kinetics dataset. In: proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 6299-6308 (2017)", + "6. Chalkiadakis, G., Elkind, E., Wooldridge, M.: Computational aspects of cooperative game theory. Synthesis Lectures on Artificial Intelligence and Machine Learning 5(6), 1-168 (2011)", + "7. Chen, J., Luo, W., Zhang, W., Ma, L.: Explore inter-contrast between videos via composition for weakly supervised temporal sentence grounding. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 36, pp. 267-275 (2022)", + "8. Chen, J., Ma, L., Chen, X., Jie, Z., Luo, J.: Localizing natural language in videos. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 33, pp. 8175-8182 (2019)", + "9. Chen, L., Lu, C., Tang, S., Xiao, J., Zhang, D., Tan, C., Li, X.: Rethinking the bottom-up framework for query-based video localization. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 34, pp. 10551-10558 (2020)", + "10. Chen, Z., Ma, L., Luo, W., Tang, P., Wong, K.Y.K.: Look closer to ground better: Weakly-supervised temporal grounding of sentence in video. arXiv preprint arXiv:2001.09308 (2020)", + "11. Datta, A., Sen, S., Zick, Y.: Algorithmic transparency via quantitative input influence: Theory and experiments with learning systems. In: 2016 IEEE symposium on security and privacy. pp. 598-617. IEEE (2016)", + "12. Deng, S., Wen, J., Liu, C., Yan, K., Xu, G., Xu, Y.: Projective incomplete multiview clustering. IEEE Transactions on Neural Networks and Learning Systems pp. 1-13 (2023). https://doi.org/10.1109/TNNLS.2023.3242473", + "13. Dong, J., Chen, X., Zhang, M., Yang, X., Chen, S., Li, X., Wang, X.: Partially relevant video retrieval. In: Proceedings of the 30th ACM International Conference on Multimedia. pp. 246-257 (2022)", + "14. Dong, J., Li, X., Snoek, C.G.: Predicting visual features from text for image and video caption retrieval. IEEE Transactions on Multimedia 20(12), 3377-3388 (2018)", + "15. Dong, J., Li, X., Xu, C., Yang, X., Yang, G., Wang, X., Wang, M.: Dual encoding for video retrieval by text. IEEE Transactions on Pattern Analysis and Machine Intelligence 44(8), 4065-4080 (2022)", + "16. Dong, J., Peng, X., Ma, Z., Liu, D., Qu, X., Yang, X., Zhu, J., Liu, B.: From region to patch: Attribute-aware foreground-background contrastive learning for fine-grained fashion retrieval. In: Proceedings of the 46th International ACM SIGIR Conference on Research and Development in Information Retrieval. pp. 1273-1282 (2023)" + ], + "bbox": [ + 225, + 172, + 784, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "Rethinking WS-VTG From a Game Perspective", + "bbox": [ + 411, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "17. Dong, J., Sun, S., Liu, Z., Chen, S., Liu, B., Wang, X.: Hierarchical contrast for unsupervised skeleton-based action representation learning. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 37, pp. 525-533 (2023)", + "18. Dong, J., Wang, Y., Chen, X., Qu, X., Li, X., He, Y., Wang, X.: Reading-strategy inspired visual representation learning for text-to-video retrieval. IEEE Transactions on Circuits and Systems for Video Technology 32(8), 5680-5694 (2022)", + "19. Donoser, M., Bischof, H.: Diffusion processes for retrieval revisited. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1320-1327 (2013)", + "20. Dowdall, J., Pavlidis, I.T., Tsiamiyrtzis, P.: Coalitional tracking in facial infrared imaging and beyond. In: 2006 Conference on Computer Vision and Pattern Recognition Workshop. pp. 134-134. IEEE (2006)", + "21. Fang, X., Easwaran, A., Genest, B.: Uncertainty-guided appearance-motion association network for out-of-distribution action detection. In: IEEE International Conference on Multimedia Information Processing and Retrieval (2024)", + "22. Fang, X., Fang, W., Liu, D., Qu, X., Dong, J., Zhou, P., Li, R., Xu, Z., Chen, L., Zheng, P., Cheng, Y.: Not all inputs are valid: Towards open-set video moment retrieval using language. In: Proceedings of the 32th ACM International Conference on Multimedia (2024)", + "23. Fang, X., Hu, Y.: Double self-weighted multi-view clustering via adaptive view fusion. arXiv preprint arXiv:2011.10396 (2020)", + "24. Fang, X., Hu, Y., Zhou, P., Wu, D.: Animc: A soft approach for autoweighted noisy and incomplete multiview clustering. IEEE Transactions on Artificial Intelligence 3(2), 192-206 (2021)", + "25. Fang, X., Hu, Y., Zhou, P., Wu, D.O.: V3h: View variation and view heredity for incomplete multiview clustering. IEEE Transactions on Artificial Intelligence 1(3), 233-247 (2020)", + "26. Fang, X., Hu, Y., Zhou, P., Wu, D.O.: Unbalanced incomplete multi-view clustering via the scheme of view evolution: Weak views are meat; strong views do eat. IEEE Transactions on Emerging Topics in Computational Intelligence 6(4), 913-927 (2021)", + "27. Fang, X., Liu, D., Fang, W., Zhou, P., Cheng, Y., Tang, K., Zou, K.: Annotations are not all you need: A cross-modal knowledge transfer network for unsupervised temporal sentence grounding. In: Findings of the Association for Computational Linguistics: EMNLP 2023. pp. 8721-8733 (2023)", + "28. Fang, X., Liu, D., Fang, W., Zhou, P., Xu, Z., Xu, W., Chen, J., Li, R.: Fewer steps, better performance: Efficient cross-modal clip trimming for video moment retrieval using language. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 38, pp. 1735-1743 (2024)", + "29. Fang, X., Liu, D., Zhou, P., Hu, Y.: Multi-modal cross-domain alignment network for video moment retrieval. IEEE Transactions on Multimedia 25, 7517-7532 (2022)", + "30. Fang, X., Liu, D., Zhou, P., Nan, G.: You can ground earlier than see: An effective and efficient pipeline for temporal sentence grounding in compressed videos. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2448-2460 (2023)", + "31. Fang, X., Liu, D., Zhou, P., Xu, Z., Li, R.: Hierarchical local-global transformer for temporal sentence grounding. IEEE Transactions on Multimedia (2023)", + "32. Gao, J., Sun, C., Yang, Z., Nevatia, R.: Tall: Temporal activity localization via language query. In: Proceedings of the IEEE International Conference on Computer Vision. pp. 5267-5275 (2017)" + ], + "bbox": [ + 225, + 146, + 784, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "X. Fang, Z. Xiong, W. Fang et al.", + "bbox": [ + 269, + 114, + 496, + 128 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "33. Gao, M., Davis, L., Socher, R., Xiong, C.: Wslln: Weakly supervised natural language localization networks. In: Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing. pp. 1481-1487 (2019)", + "34. Grabisch, M., Roubens, M.: An axiomatic approach to the concept of interaction among players in cooperative games. International Journal of game theory 28, 547-565 (1999)", + "35. Guo, C., Liu, D., Zhou, P.: A hybrid alignment loss for temporal moment localization with natural language. In: 2022 IEEE International Conference on Multimedia and Expo (ICME). pp. 1-6. IEEE (2022)", + "36. Guo, D., Li, K., Hu, B., Zhang, Y., Wang, M.: Benchmarking micro-action recognition: Dataset, method, and application. IEEE Transactions on Circuits and Systems for Video Technology 34(7), 6238-6252 (2024)", + "37. Hendricks, L.A., Wang, O., Shechtman, E., Sivic, J., Darrell, T., Russell, B.: Localizing moments in video with temporal language. Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing p. 1380-1390 (2018)", + "38. Huang, J., Liu, Y., Gong, S., Jin, H.: Cross-sentence temporal and semantic relations in video activity localisation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 7199-7208 (2021)", + "39. Jiang, L., Wang, C., Ning, X., Yu, Z.: Lttpoint: A mlp-based point cloud classification method with local topology transformation module. In: 2023 7th Asian Conference on Artificial Intelligence Technology (ACAIT). pp. 783-789. IEEE (2023)", + "40. Jin, P., Huang, J., Xiong, P., Tian, S., Liu, C., Ji, X., Yuan, L., Chen, J.: Videotext as game players: Hierarchical banzhaf interaction for cross-modal representation learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2472-2482 (2023)", + "41. Jin, S., Wang, S., Fang, F.: Game theoretical analysis on capacity configuration for microgrid based on multi-agent system. International Journal of Electrical Power & Energy Systems 125, 106485 (2021)", + "42. Kingma, D.P., Ba, J.: Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980 (2014)", + "43. Krishna, R., Hata, K., Ren, F., Fei-Fei, L., Carlos Niebles, J.: Dense-c captioning events in videos. In: Proceedings of the IEEE International Conference on Computer Vision. pp. 706-715 (2017)", + "44. Leech, D.: Computation of power indices (2002)", + "45. Lehrer, E.: An axiomatization of the banzhaf value. International Journal of Game Theory 17, 89-99 (1988)", + "46. Li, H., Cao, M., Cheng, X., Li, Y., Zhu, Z., Zou, Y.: G2l: Semantically aligned and uniform video grounding via geodesic and game theory. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 12032-12042 (2023)", + "47. Li, J., HE, X., Wei, L., Qian, L., Zhu, L., Xie, L., Zhuang, Y., Tian, Q., Tang, S.: Fine-grained semantically aligned vision-language pre-training. In: Advances in Neural Information Processing Systems (2022)", + "48. Lin, K.Q., Zhang, P., Chen, J., Pramanick, S., Gao, D., Wang, A.J., Yan, R., Shou, M.Z.: Univtg: Towards unified video-language temporal grounding. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 2794-2804 (2023)" + ], + "bbox": [ + 225, + 146, + 784, + 839 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "Rethinking WS-VTG From a Game Perspective", + "bbox": [ + 411, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "49. Lin, Z., Zhao, Z., Zhang, Z., Wang, Q., Liu, H.: Weakly-supervised video moment retrieval via semantic completion network. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 34, pp. 11539-11546 (2020)", + "50. Liu, C., Wen, J., Luo, X., Huang, C., Wu, Z., Xu, Y.: Dicnet: Deep instance-level contrastive network for double incomplete multi-view multi-label classification. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 37, pp. 8807-8815 (2023)", + "51. Liu, C., Wen, J., Luo, X., Xu, Y.: Incomplete multi-view multi-label learning via label-guided masked view- and category-aware transformers. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 37, pp. 8816-8824 (2023)", + "52. Liu, C., Wen, J., Wu, Z., Luo, X., Huang, C., Xu, Y.: Information recovery-driven deep incomplete multiview clustering network. IEEE Transactions on Neural Networks and Learning Systems pp. 1-11 (2023)", + "53. Liu, D., Fang, X., Hu, W., Zhou, P.: Exploring optical-flow-guided motion and detection-based appearance for temporal sentence grounding. IEEE Transactions on Multimedia 25, 8539-8553 (2023)", + "54. Liu, D., Fang, X., Qu, X., Dong, J., Yan, H., Yang, Y., Zhou, P., Cheng, Y.: Unsupervised domain adaptative temporal sentence localization with mutual information maximization. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 38, pp. 3567-3575 (2024)", + "55. Liu, D., Fang, X., Zhou, P., Di, X., Lu, W., Cheng, Y.: Hypotheses tree building for one-shot temporal sentence localization. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 37, pp. 1640-1648 (2023)", + "56. Liu, D., Hu, W.: Learning to focus on the foreground for temporal sentence grounding. In: Proceedings of the 29th International Conference on Computational Linguistics. pp. 5532-5541 (2022)", + "57. Liu, D., Hu, W.: Skimming, locating, then perusing: A human-like framework for natural language video localization. In: Proceedings of the 30th ACM International Conference on Multimedia. pp. 4536-4545 (2022)", + "58. Liu, D., Qu, X., Dong, J., Nan, G., Zhou, P., Xu, Z., Chen, L., Yan, H., Cheng, Y.: Filling the information gap between video and query for language-driven moment retrieval. In: Proceedings of the 31st ACM International Conference on Multimedia. pp. 4190-4199 (2023)", + "59. Liu, D., Qu, X., Dong, J., Zhou, P., Cheng, Y., Wei, W., Xu, Z., Xie, Y.: Context-aware biaffine localizing network for temporal sentence grounding. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 11235-11244 (2021)", + "60. Liu, D., Qu, X., Dong, J., Zhou, P., Xu, Z., Wang, H., Di, X., Lu, W., Cheng, Y.: Transform-equivariant consistency learning for temporal sentence grounding. ACM Transactions on Multimedia Computing, Communications and Applications 20(4), 1-19 (2024)", + "61. Liu, D., Qu, X., Fang, X., Dong, J., Zhou, P., Nan, G., Tang, K., Fang, W., Cheng, Y.: Towards robust temporal activity localization learning with noisy labels. In: Proceedings of the 2024 Joint International Conference on Computational Linguistics, Language Resources and Evaluation (LREC-COLING 2024). pp. 16630-16642 (2024)", + "62. Liu, D., Qu, X., Hu, W.: Reducing the vision and language bias for temporal sentence grounding. In: Proceedings of the 30th ACM International Conference on Multimedia. pp. 4092-4101 (2022)" + ], + "bbox": [ + 225, + 146, + 784, + 839 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "X. Fang, Z. Xiong, W. Fang et al.", + "bbox": [ + 271, + 114, + 496, + 128 + ], + "page_idx": 17 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "63. Liu, D., Qu, X., Liu, X.Y., Dong, J., Zhou, P., Xu, Z.: Jointly cross-and self-modal graph attention network for query-based moment localization. In: Proceedings of the 28th ACM International Conference on Multimedia. pp. 4070-4078 (2020)", + "64. Liu, D., Zhou, P.: Jointly visual-and semantic-aware graph memory networks for temporal sentence localization in videos. In: ICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). pp. 1-5. IEEE (2023)", + "65. Liu, D., Zhou, P., Xu, Z., Wang, H., Li, R.: Few-shot temporal sentence grounding via memory-guided semantic learning. IEEE Transactions on Circuits and Systems for Video Technology 33(5), 2491-2505 (2022)", + "66. Liu, D., Zhu, J., Fang, X., Xiong, Z., Wang, H., Li, R., Zhou, P.: Conditional video diffusion network for fine-grained temporal sentence grounding. IEEE Transactions on Multimedia (2023)", + "67. Lundberg, S.M., Lee, S.I.: A unified approach to interpreting model predictions. In: Proceedings of the 31st International Conference on Neural Information Processing Systems. pp. 4768-4777 (2017)", + "68. Ma, M., Yoon, S., Kim, J., Lee, Y., Kang, S., Yoo, C.D.: VLANet: Video-language alignment network for weakly-supervised video moment retrieval. In: Proceedings of the European Conference on Computer Vision. pp. 156-171 (2020)", + "69. Ma, W.C., Huang, D.A., Lee, N., Kitani, K.M.: Forecasting interactive dynamics of pedestrians with fictitious play. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 774-782 (2017)", + "70. Ma, Y., Liu, Y., Wang, L., Kang, W., Qiao, Y., Wang, Y.: Dual masked modeling for weakly-supervised temporal boundary discovery. IEEE Transactions on Multimedia (2023)", + "71. Matsui, Y., Matsui, T.: Np-completeness for calculating power indices of weighted majority games. Theoretical Computer Science 263(1-2), 305-310 (2001)", + "72. Michalak, T.P., Aadithya, K.V., Szczepanski, P.L., Ravindran, B., Jennings, N.R.: Efficient computation of the shapley value for game-theoretic network centrality. Journal of Artificial Intelligence Research 46, 607-650 (2013)", + "73. Mithun, N.C., Paul, S., Roy-Chowdhury, A.K.: Weakly supervised video moment retrieval from text queries. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 11592-11601 (2019)", + "74. Mun, J., Cho, M., Han, B.: Local-global video-text interactions for temporal grounding. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 10810-10819 (2020)", + "75. Ning, E., Wang, C., Zhang, H., Ning, X., Tiwari, P.: Occluded person re-identification with deep learning: a survey and perspectives. Expert Systems with Applications p. 122419 (2023)", + "76. Ning, E., Wang, Y., Wang, C., Zhang, H., Ning, X.: Enhancement, integration, expansion: Activating representation of detailed features for occluded person re-identification. Neural Networks 169, 532-541 (2024)", + "77. Ning, E., Zhang, C., Wang, C., Ning, X., Chen, H., Bai, X.: Pedestrian re-id based on feature consistency and contrast enhancement. Displays 79, 102467 (2023)", + "78. Nowak, A.S.: On an axiomatization of the banzhaf value without the additivity axiom. International Journal of Game Theory 26, 137-141 (1997)", + "79. Oord, A.v.d., Li, Y., Vinyals, O.: Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748 (2018)", + "80. Osborne, M.J., Rubinstein, A.: A course in game theory. MIT press (1994)" + ], + "bbox": [ + 225, + 146, + 785, + 839 + ], + "page_idx": 18 + }, + { + "type": "header", + "text": "Rethinking WS-VTG From a Game Perspective", + "bbox": [ + 411, + 114, + 730, + 128 + ], + "page_idx": 18 + }, + { + "type": "page_number", + "text": "19", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 18 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "81. Patel, R., Garnelo, M., Gemp, I., Dyer, C., Bachrach, Y.: Game-theoretic vocabulary selection via the shapley value and banzhaf index. In: Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies. pp. 2789-2798 (2021)", + "82. Pavan, M., Pelillo, M.: A new graph-theoretic approach to clustering and segmentation. In: 2003 IEEE Computer Society Conference on Computer Vision and Pattern Recognition, 2003. Proceedings. vol. 1, pp. I-I. IEEE (2003)", + "83. Pennington, J., Socher, R., Manning, C.D.: Glove: Global vectors for word representation. In: Proceedings of the Conference on Empirical Methods in Natural Language Processing. pp. 1532-1543 (2014)", + "84. Rodola, E., Bronstein, A.M., Albarelli, A., Bergamasco, F., Torsello, A.: A game-theoretic approach to deformable shape matching. In: 2012 IEEE Conference on Computer Vision and Pattern Recognition. pp. 182-189. IEEE (2012)", + "85. Shapley, L.S., et al.: A value for n-person games (1953)", + "86. Sigurdsson, G.A., Varol, G., Wang, X., Farhadi, A., Laptev, I., Gupta, A.: Hollywood in homes: Crowdsourcing data collection for activity understanding. In: European Conference on Computer Vision. pp. 510-526 (2016)", + "87. Song, Y., Wang, J., Ma, L., Yu, J., Liang, J., Yuan, L., Yu, Z.: Marn: Multilevel attentional reconstruction networks for weakly supervised video temporal grounding. Neurocomputing 554, 126625 (2023)", + "88. Song, Y., Wang, J., Ma, L., Yu, Z., Yu, J.: Weakly-supervised multi-level attentional reconstruction network for grounding textual queries in videos. arXiv preprint arXiv:2003.07048 (2020)", + "89. Tan, R., Xu, H., Saenko, K., Plummer, B.A.: Logan: Latent graph co-attention network for weakly-supervised video moment retrieval. In: Proceedings of the IEEE Winter Conference on Applications of Computer Vision. pp. 2083–2092 (2021)", + "90. Tang, K., Chen, Y., Peng, W., Zhang, Y., Fang, M., Wang, Z., Song, P.: Reppvconv: attentively fusing reparameterized voxel features for efficient 3d point cloud perception. The Visual Computer 39(11), 5577-5588 (2023)", + "91. Tang, K., Lou, T., Peng, W., Chen, N., Shi, Y., Wang, W.: Effective single-step adversarial training with energy-based models. IEEE Transactions on Emerging Topics in Computational Intelligence (2024). https://doi.org/10.1109/TETCI.2024.3378652", + "92. Tang, K., Ma, Y., Miao, D., Song, P., Gu, Z., Tian, Z., Wang, W.: Decision fusion networks for image classification. IEEE Transactions on Neural Networks and Learning Systems (2022). https://doi.org/10.1109/TNNLS.2022.3196129", + "93. Tang, K., Shi, Y., Lou, T., Peng, W., He, X., Zhu, P., Gu, Z., Tian, Z.: Rethinking perturbation directions for imperceptible adversarial attacks on point clouds. IEEE Internet of Things Journal 10(6), 5158-5169 (2022)", + "94. Tang, K., Zhao, W., Peng, W., Fang, X., Cui, X., Zhu, P., Tian, Z.: Reparameterization head for efficient multi-input networks. In: ICASSP 2024 - 2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). pp. 6190-6194 (2024). https://doi.org/10.1109/ICASSP48485.2024.10447574", + "95. Tang, K., Zhao, W., Peng, W., Fang, X., Cui, X., Zhu, P., Tian, Z.: Reparameterization head for efficient multi-input networks. In: ICASSP 2024-2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). pp. 6190-6194. IEEE (2024)", + "96. Torsello, A., Bulo, S.R., Pelillo, M.: Grouping with asymmetric affinities: A game-theoretic perspective. In: 2006 IEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR'06). vol. 1, pp. 292-299. IEEE (2006)" + ], + "bbox": [ + 225, + 147, + 784, + 839 + ], + "page_idx": 19 + }, + { + "type": "page_number", + "text": "20", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 19 + }, + { + "type": "header", + "text": "X. Fang, Z. Xiong, W. Fang et al.", + "bbox": [ + 271, + 114, + 496, + 128 + ], + "page_idx": 19 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "97. Tran, D., Bourdev, L., Fergus, R., Torresani, L., Paluri, M.: Learning spatiotemporal features with 3d convolutional networks. In: Proceedings of the IEEE International Conference on Computer Vision. pp. 4489-4497 (2015)", + "98. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. In: Advances in Neural Information Processing Systems. pp. 5998-6008 (2017)", + "99. Wang, C., Ning, X., Li, W., Bai, X., Gao, X.: 3d person re-identification based on global semantic guidance and local feature aggregation. IEEE Transactions on Circuits and Systems for Video Technology (2023)", + "100. Wang, C., Ning, X., Sun, L., Zhang, L., Li, W., Bai, X.: Learning discriminative features by covering local geometric space for point cloud analysis. IEEE Transactions on Geoscience and Remote Sensing 60, 1-15 (2022)", + "101. Wang, C., Wang, C., Li, W., Wang, H.: A brief survey on rgb-d semantic segmentation using deep learning. Displays 70, 102080 (2021)", + "102. Wang, C., Wang, H., Ning, X., Shengwei, T., Li, W.: 3d point cloud classification method based on dynamic coverage of local area. Journal of Software 34(4), 1962-1976 (2022)", + "103. Wang, J., Ma, L., Jiang, W.: Temporally grounding language queries in videos by contextual boundary-aware prediction. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 34, pp. 12168-12175 (2020)", + "104. Wang, Y., Deng, J., Zhou, W., Li, H.: Weakly supervised temporal adjacent network for language grounding. IEEE Transactions on Multimedia 24, 3276-3286 (2021)", + "105. Wang, Z., Chen, J., Jiang, Y.G.: Visual co-occurrence alignment learning for weakly-supervised video moment retrieval. In: Proceedings of the 29th ACM International Conference on Multimedia. pp. 1459–1468 (2021)", + "106. Wen, J., Liu, C., Deng, S., Liu, Y., Fei, L., Yan, K., Xu, Y.: Deep double incomplete multi-view multi-label learning with incomplete labels and missing views. IEEE Transactions on Neural Networks and Learning Systems pp. 1-13 (2023). https://doi.org/10.1109/TNNLS.2023.3260349", + "107. Wen, J., Zhang, Z., Li, Z.J.: A survey on incomplete multiview clustering. IEEE transactions on systems, man, and cybernetics. Systems 53(2 Pt.2), 1136-1149 (2023)", + "108. Winter, E.: The shapley value. Handbook of game theory with economic applications 3, 2025-2054 (2002)", + "109. Wu, H., Lyu, Y., Shen, X., Zhao, X., Wang, M., Zhang, X., Luo, Z.: Atomic-action-based contrastive network for weakly supervised temporal language grounding. In: 2023 IEEE International Conference on Multimedia and Expo (ICME). pp. 1523-1528. IEEE (2023)", + "110. Xiong, Z., Liu, D., Zhou, P.: Gaussian kernel-based cross modal network for spatio-temporal video grounding. In: IEEE International Conference on Image Processing (ICIP). pp. 2481-2485 (2022)", + "111. Xiong, Z., Liu, D., Zhou, P., Zhu, J.: Tracking objects and activities with attention for temporal sentence grounding. In: ICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). pp. 1-5. IEEE (2023)", + "112. Yang, W., Zhang, T., Zhang, Y., Wu, F.: Local correspondence network for weakly supervised temporal sentence grounding. IEEE Transactions on Image Processing 30, 3252-3262 (2021)" + ], + "bbox": [ + 218, + 146, + 784, + 839 + ], + "page_idx": 20 + }, + { + "type": "header", + "text": "Rethinking WS-VTG From a Game Perspective", + "bbox": [ + 411, + 114, + 730, + 128 + ], + "page_idx": 20 + }, + { + "type": "page_number", + "text": "21", + "bbox": [ + 767, + 116, + 782, + 126 + ], + "page_idx": 20 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "113. Yu, Z., Li, L., Xie, J., Wang, C., Li, W., Ning, X.: Pedestrian 3d shape understanding for person re-identification via multi-view learning. IEEE Transactions on Circuits and Systems for Video Technology (2024)", + "114. Yuan, Y., Ma, L., Wang, J., Liu, W., Zhu, W.: Semantic conditioned dynamic modulation for temporal sentence grounding in videos. In: Proceedings of the 33rd International Conference on Neural Information Processing Systems. pp. 536-546 (2019)", + "115. Zeng, R., Xu, H., Huang, W., Chen, P., Tan, M., Gan, C.: Dense regression network for video grounding. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 10287-10296 (2020)", + "116. Zhang, D., Dai, X., Wang, X., Wang, Y.F., Davis, L.S.: Man: Moment alignment network for natural language moment retrieval via iterative graph adjustment. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 1247-1257 (2019)", + "117. Zhang, H., Sun, A., Jing, W., Zhou, J.T.: Span-based localizing network for natural language video localization. In: Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics. pp. 6543-6554 (2020)", + "118. Zhang, H., Xie, Y., Zheng, L., Zhang, D., Zhang, Q.: Interpreting multivariate shapley interactions in dnns. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 35, pp. 10877-10886 (2021)", + "119. Zhang, H., Wang, C., Tian, S., Lu, B., Zhang, L., Ning, X., Bai, X.: Deep learning-based 3d point cloud classification: A systematic survey and outlook. Displays 79, 102456 (2023)", + "120. Zhang, H., Wang, C., Yu, L., Tian, S., Ning, X., Rodrigues, J.: Pointgt: A method for point-cloud classification and segmentation based on local geometric transformation. IEEE Transactions on Multimedia (2024)", + "121. Zhang, S., Peng, H., Fu, J., Luo, J.: Learning 2d temporal adjacent networks for moment localization with natural language. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 34, pp. 12870-12877 (2020)", + "122. Zhang, Z., Lin, Z., Zhao, Z., Xiao, Z.: Cross-modal interaction networks for query-based moment retrieval in videos. In: Proceedings of the 42nd International ACM SIGIR Conference on Research and Development in Information Retrieval. pp. 655-664 (2019)", + "123. Zhang, Z., Zhao, Z., Lin, Z., He, X., et al.: Counterfactual contrastive learning for weakly-supervised vision-language grounding. Advances in Neural Information Processing Systems 33, 18123-18134 (2020)", + "124. Zheng, M., Huang, Y., Chen, Q., Liu, Y.: Weakly supervised video moment localization with contrastive negative sample mining. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 36, pp. 3517-3525 (2022)", + "125. Zheng, M., Huang, Y., Chen, Q., Peng, Y., Liu, Y.: Weakly supervised temporal sentence grounding with gaussian-based contrastive proposal learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15555-15564 (2022)", + "126. Zheng, Q., Dong, J., Qu, X., Yang, X., Wang, Y., Zhou, P., Liu, B., Wang, X.: Progressive localization networks for language-based moment localization. ACM Transactions on Multimedia Computing, Communications and Applications 19(2), 1-21 (2023)", + "127. Zhu, J., Liu, D., Zhou, P., Di, X., Cheng, Y., Yang, S., Xu, W., Xu, Z., Wan, Y., Sun, L., et al.: Rethinking the video sampling and reasoning strategies for temporal sentence grounding. arXiv preprint arXiv:2301.00514 (2023)" + ], + "bbox": [ + 215, + 146, + 785, + 830 + ], + "page_idx": 21 + }, + { + "type": "page_number", + "text": "22", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 21 + }, + { + "type": "header", + "text": "X. Fang, Z. Xiong, W. Fang et al.", + "bbox": [ + 271, + 114, + 496, + 128 + ], + "page_idx": 21 + } +] \ No newline at end of file diff --git a/2024/Rethinking Weakly-supervised Video Temporal Grounding From a Game Perspective/4f2540b2-441f-49e9-a4f9-06c96b70a305_model.json b/2024/Rethinking Weakly-supervised Video Temporal Grounding From a Game Perspective/4f2540b2-441f-49e9-a4f9-06c96b70a305_model.json new file mode 100644 index 0000000000000000000000000000000000000000..10b467ebf589cd9396dbccab5f56a2b8bac24750 --- /dev/null +++ b/2024/Rethinking Weakly-supervised Video Temporal Grounding From a Game Perspective/4f2540b2-441f-49e9-a4f9-06c96b70a305_model.json @@ -0,0 +1,3368 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.232, + 0.141, + 0.773, + 0.187 + ], + "angle": 0, + "content": "Rethinking Weakly-supervised Video Temporal Grounding From a Game Perspective" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.217, + 0.778, + 0.248 + ], + "angle": 0, + "content": "Xiang Fang\\(^{1*}\\), Zeyu Xiong\\(^{1*}\\), Wanlong Fang\\(^{1*}\\), Xiaoye Qu\\(^{1}\\), Chen Chen\\(^{2}\\), Jianfeng Dong\\(^{3}\\), Keke Tang\\(^{4}\\), Pan Zhou\\(^{1\\boxtimes}\\), Yu Cheng\\(^{5}\\), and Daizong Liu\\(^{6\\boxtimes}\\)" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.258, + 0.786, + 0.273 + ], + "angle": 0, + "content": "1 Hubei Key Laboratory of Distributed System Security, Hubei Engineering Research" + }, + { + "type": "text", + "bbox": [ + 0.229, + 0.274, + 0.774, + 0.288 + ], + "angle": 0, + "content": "Center on Big Data Security, School of Cyber Science and Engineering, Huazhong" + }, + { + "type": "text", + "bbox": [ + 0.378, + 0.288, + 0.626, + 0.301 + ], + "angle": 0, + "content": "University of Science and Technology" + }, + { + "type": "text", + "bbox": [ + 0.397, + 0.301, + 0.605, + 0.314 + ], + "angle": 0, + "content": "2 University of Central Florida" + }, + { + "type": "text", + "bbox": [ + 0.39, + 0.314, + 0.613, + 0.328 + ], + "angle": 0, + "content": "3 Zhejiang Gongshang University" + }, + { + "type": "text", + "bbox": [ + 0.42, + 0.329, + 0.584, + 0.343 + ], + "angle": 0, + "content": "4 Guangzhou University" + }, + { + "type": "text", + "bbox": [ + 0.368, + 0.343, + 0.636, + 0.356 + ], + "angle": 0, + "content": "5 The Chinese University of Hong Kong" + }, + { + "type": "text", + "bbox": [ + 0.434, + 0.357, + 0.569, + 0.37 + ], + "angle": 0, + "content": "6 Peking University" + }, + { + "type": "text", + "bbox": [ + 0.245, + 0.371, + 0.756, + 0.385 + ], + "angle": 0, + "content": "xfang9508@gmail.com, zeyuxiong@hust.edu.cn, wanlongfang@gmail.com," + }, + { + "type": "text", + "bbox": [ + 0.258, + 0.386, + 0.741, + 0.398 + ], + "angle": 0, + "content": "xiaoye@hust.edu.cn, chen.chen@crcv.ucf.edu, dongjf24@gmail.com," + }, + { + "type": "text", + "bbox": [ + 0.24, + 0.4, + 0.762, + 0.412 + ], + "angle": 0, + "content": "tangbohutbh@gmail.com, panzhou@hust.edu.cn, chengyu@cse.cuhk.edu.hk," + }, + { + "type": "text", + "bbox": [ + 0.424, + 0.414, + 0.58, + 0.425 + ], + "angle": 0, + "content": "dzliu@stu.pku.edu.cn" + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.457, + 0.745, + 0.803 + ], + "angle": 0, + "content": "Abstract. This paper addresses the challenging task of weakly-supervised video temporal grounding. Existing approaches are generally based on the moment proposal selection framework that utilizes contrastive learning and reconstruction paradigm for scoring the pre-defined moment proposals. Although they have achieved significant progress, we argue that their current frameworks have overlooked two indispensable issues: 1) Coarse-grained cross-modal learning: previous methods solely capture the global video-level alignment with the query, failing to model the detailed consistency between video frames and query words for accurately grounding the moment boundaries. 2) Complex moment proposals: their performance severely relies on the quality of proposals, which are also time-consuming and complicated for selection. To this end, in this paper, we make the first attempt to tackle this task from a novel game perspective, which effectively learns the uncertain relationship between each vision-language pair with diverse granularity and flexible combination for multi-level cross-modal interaction. Specifically, we creatively model each video frame and query word as game players with multivariate cooperative game theory to learn their contribution to the cross-modal similarity score. By quantifying the trend of frame-word cooperation within a coalition via the game-theoretic interaction, we are able to value all uncertain but possible correspondence between frames and words. Finally, instead of using moment proposals, we utilize the learned query-guided frame-wise scores for better moment localization. Experiments show that our method achieves superior performance on both Charades-STA and ActivityNet Caption datasets." + }, + { + "type": "page_footnote", + "bbox": [ + 0.214, + 0.811, + 0.747, + 0.826 + ], + "angle": 0, + "content": "* Xiang Fang, Zeyu Xiong, and Wanlong Fang contributed equally to this work." + }, + { + "type": "page_footnote", + "bbox": [ + 0.214, + 0.826, + 0.577, + 0.84 + ], + "angle": 0, + "content": "Corresponding authors: Pan Zhou and Daizong Liu." + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.811, + 0.747, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.498, + 0.129 + ], + "angle": 0, + "content": "X. Fang, Z. Xiong, W. Fang et al." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.146, + 0.559, + 0.244 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.56, + 0.146, + 0.785, + 0.244 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.245, + 0.788, + 0.314 + ], + "angle": 0, + "content": "Fig. 1: (a) Most existing WS-VTG methods define moment proposals for query matching. They not only suffer from the coarse-grained alignment but also are limited to the quality of the proposals. (b) Instead of using proposals, our method reviews the WS-VTG from a brand-new game-interaction perspective to learn more fine-grained frame-aware self- and cross-modal alignment for accurate boundary localization." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.322, + 0.375, + 0.337 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.357, + 0.788, + 0.583 + ], + "angle": 0, + "content": "Video temporal grounding (VTG) aims to localize the start and end timestamps of a specific event moment described by a given language query in an untrimmed video [2,32,39,75-77,99-102,113,119,120]. This task enables us to efficiently find video moments of human interest instead of traversing the entire video, which has broad application potential in video surveillance [13-15, 18, 24-26, 50-52, 59, 90-93, 125, 126], video summarization [12, 16, 17, 21-23, 36, 48, 94, 106, 107], etc. Most VTG methods [27-31, 53-55, 57, 58, 60-62, 65, 66, 127] follow the fully-supervised setting, where each frame is manually annotated as query-relevant or not. Although these methods achieve significant breakthroughs, they severely rely on extensive laborious manual annotations of moment boundaries, thus limiting their scalability and practicability in real-world applications. To alleviate the dense reliance, weakly-supervised VTG (WS-VTG) has received increasing attention [10, 49, 73, 88, 89, 123], which only requires the knowledge of matched video-query pairs without detailed frame-wise annotations, which is more challenging than the fully-supervised setting." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.584, + 0.789, + 0.842 + ], + "angle": 0, + "content": "Since there is no frame-wise annotation in WS-VTG, learning detailed frame-query alignment is difficult. As shown in Fig. 1(a), almost all existing WS-VTG solutions directly employ the proposal-based framework [10,33,73,89] that first generates multiple moment proposals and then learns the scores to indicate the potential alignment between video proposals and language query. Some approaches further utilize reconstruction-based paradigm [49,88,123] to minimize the reconstruction loss between the partially masked query and moment proposals to identify the proposal which best reconstructs the query. Although existing WS-VTG methods have achieved great progress, we argue that they still suffer from two inescapable limitations: 1) Firstly, as frame-wise annotation is not provided, these WS-VTG methods generally generate the moment proposals for all videos via sliding windows. In this way, they only model the coarse-grained cross-modal interaction of each proposal-query pair. Actually, in most cases, we expect to capture more fine-grained information for accurate boundary localization, such as how much the semantics of a specific frame helps or harms the query-guided cross-modal alignment. Unfortunately, existing methods rely on proposal-level cross-modal learning cannot achieve this goal. 2) Secondly, the" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.412, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Rethinking WS-VTG From a Game Perspective" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.192 + ], + "angle": 0, + "content": "performance of these proposal-based methods is severely limited by the quality of moment proposals as they devise moment proposals regardless of the specific contents and difficulty of each video." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.193, + 0.788, + 0.524 + ], + "angle": 0, + "content": "To tackle the above limitations, in this paper, we propose to address the WS-VTG task from a new game-theory perspective, which creatively models the subtle video frames and query words as game players to learn their uncertain but possible cross-modal interaction with diverse granularity and flexible combination. The game learning helps to generate query-guided frame-aware knowledge for fine-grained video grounding. As shown in Fig. 1(b), if visual representations and textual representations have strong semantic correspondence, they tend to cooperate together and contribute to a high cross-modal similarity score. By forming these cooperated frame-word representations as a coalition and quantifying the trend of cooperation within a coalition via the game-theoretic interaction index, we can not only learn the coalition contribution to the cross-modal semantic similarity, but also measure the additional benefits brought by the coalition compared with the costs of the lost coalitions of these players with others. In this manner, our cooperative WS-VTG game is able to value all possible correspondence between frames and words for sensitive and explainable cross-modal contrast. By formulating these detailed cross-modal correspondences as fine-grained query-guided frame-wise scores, we can efficiently generate the moment instead of using complicated proposals. In particular, we implement the cooperative game in both self-modal and cross-modal scenarios for self-modal attentive learning and cross-modal fine-grained alignment. We further extend the cross-modal game with multi-level query-semantic interaction for more comprehensive video understanding." + }, + { + "type": "text", + "bbox": [ + 0.24, + 0.525, + 0.608, + 0.539 + ], + "angle": 0, + "content": "Our main contributions are summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.546, + 0.784, + 0.619 + ], + "angle": 0, + "content": "- We reveal the limitations of existing WS-VTG methods, which rely on complex moment proposals and fail to capture fine-grained frame-word alignment for accurate boundary grounding. To this end, we propose a novel game-theory based framework to learn the detailed frame-wise query-relevance scores for efficiently constructing the accurate moment." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.621, + 0.784, + 0.663 + ], + "angle": 0, + "content": "- We apply the game learning to both self- and cross-modal scenarios for contextual self-enhancement and cross-alignment. The cross-modal game is also extended to multi-level for comprehensive learning." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.666, + 0.784, + 0.738 + ], + "angle": 0, + "content": "- To verify the effectiveness of our framework, we conduct experiments on two challenging WS-VTG benchmarks, i.e., Charades-STA and ActivityNet Caption. Experiments show that our method achieves superior performance. In representative cases, our method outperforms all compared methods by \\(4.53\\%\\) on Charades-STA." + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.546, + 0.784, + 0.738 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.764, + 0.396, + 0.78 + ], + "angle": 0, + "content": "2 Related Works" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.784, + 0.84 + ], + "angle": 0, + "content": "Fully-supervised video temporal grounding. Existing methods generally address the video temporal grounding in a fully-supervised manner, where both the annotations of video-sentence pairs and corresponding moment boundaries" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.498, + 0.129 + ], + "angle": 0, + "content": "X. Fang, Z. Xiong, W. Fang et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.252 + ], + "angle": 0, + "content": "are given. Most of them [32, 35, 37, 56, 61, 64, 66, 95, 103, 110, 111, 114, 116, 121, 122] utilize the proposal-based framework, which first integrate the sentence representation with those pre-defined moment proposals individually, and then evaluate their matching relationships. The proposal with the highest matching score is selected as the prediction. Instead of utilizing the moment proposals, recent proposal-free methods [8, 9, 74, 115] directly regress the temporal locations of the target moment." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.253, + 0.787, + 0.479 + ], + "angle": 0, + "content": "Weakly-supervised video temporal grounding. As manually annotating temporal boundaries of target moments is time-consuming, recent research attentions shift to weakly-supervised video temporal grounding [10, 73, 88, 89], which only requires video-level annotations. Mithun et al. [73] proposed the first weakly-supervised model to learn a joint embedding space for video and query representations. Gao et al. [33] developed a two-stream structure to measure the moment-query consistency and conduct moment selection simultaneously. Although the above methods have achieved promising performance, they are two-stage approaches that utilize multi-scale sliding windows to generate moment candidates, therefore suffering from inferior effectiveness and efficiency. To address this issue, [49, 68, 123] score all the moments sampled at different scales in a single pass and further improve the moment-sentence matching accuracy. However, almost all of them rely on moment proposals for matching and selection, which fail to capture and distinguish more fine-grained details among visually similar frames for acquiring more accurate moment boundaries." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.48, + 0.788, + 0.736 + ], + "angle": 0, + "content": "Cooperative game theory. Cooperative game theory focuses on the formation of coalitions and cooperation among rational players or groups to achieve common goals [1,6,40,41,47,69,80-82,96]. A typical cooperative game consists of a set of players with a game function. The game function maps all possible subsets of players, called group or coalition, to a number which represents the total payoff earned by these players working cooperatively to achieve the goal [19, 20, 46, 84]. To measure the contributions of each player and allocate different payoffs to these individuals fairly, a few researchers have proposed various concepts of value by computing the average added worth that one player brings to all possible coalitions, such as Shapley value [72, 85, 108] and Banzhaf value [4, 45, 78]. Recently, some concepts of cooperative game theory have been interpreted in the field of deep learning [11, 67, 118]. For example, Li et al. [47] propose a fine-grained image-text semantic alignment pre-training framework based on Shapley interaction. Different from these works, we propose for the first time to model the detailed consistency between video frames and query words by recasting multivariate game-theoretic indices to address more complicated WS-VTG task." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.761, + 0.381, + 0.78 + ], + "angle": 0, + "content": "3 Methodology" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.787, + 0.842 + ], + "angle": 0, + "content": "In this section, we elaborate on the proposed method, which learns the uncertain and fine-grained cross-modal alignment in WS-VTG task. Specifically, our method conceptualizes the interaction between the video and query as a game," + } + ], + [ + { + "type": "header", + "bbox": [ + 0.413, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Rethinking WS-VTG From a Game Perspective" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "image", + "bbox": [ + 0.216, + 0.144, + 0.788, + 0.349 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.351, + 0.788, + 0.42 + ], + "angle": 0, + "content": "Fig. 2: Overview of our framework, where \"SQAN\" denotes Sequential Query Attention Network. Specifically, we first conduct self-modal game to enhance the self-modal instance semantics, then devise multi-level cross-modal games to learn the fine-grained and uncertain cross-modal correspondence. At last, we predict the moment based on the query-guided frame-wise scores." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.434, + 0.788, + 0.526 + ], + "angle": 0, + "content": "where each frame/word acts as a player and forms coalitions. During the game learning, visual frame and linguistic word representations with strong semantic relevance will cooperate to form a new coalition, the benefits of which are quantified through the game interaction index. Our method maximizes payoff by strategically forming coalitions to achieve fine-grained semantic alignment between video and query. The overall framework is shown in Fig. 2." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.559, + 0.373, + 0.573 + ], + "angle": 0, + "content": "3.1 Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.596, + 0.788, + 0.718 + ], + "angle": 0, + "content": "Problem formulation. Generally, given an untrimmed video and a sentence query, WS-VTG aims to localize the specific video moment semantically related to the query with only the video-level annotation. Let \\( V = \\{v_{1},\\dots,v_{i}\\dots,v_{T}\\} \\) denote the untrimmed video with \\( T \\) frames, and \\( Q = \\{q_{1},\\dots,q_{j},\\dots,q_{M}\\} \\) represent the sentence query containing \\( M \\) words, where \\( v_{i} \\) and \\( q_{j} \\) are the \\( i \\)-th frame and \\( j \\)-th word, respectively. The objective function \\( \\mathcal{F} \\) can be formulated as \\( \\mathcal{F}: (V,Q) \\to (\\tau_s,\\tau_e) \\), where \\( \\tau_s,\\tau_e \\) denote the start and end timestamps of the video moment boundary in video \\( V \\) semantically corresponding to query \\( Q \\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.788, + 0.844 + ], + "angle": 0, + "content": "Video and query representations. To extract contextual video representation, following [32, 117], each input video \\(V\\) is fed into a pre-trained 3D convolutional network [5, 97] with a stack of multi-head self-attention [98] layers for video encoding, then we get frame-wise video representation \\(\\mathbf{V} = \\{\\pmb{v}_i\\}_{i=1}^T \\in \\mathbb{R}^{T \\times D_h}\\), where \\(D_h\\) is the hidden dimension. For query representation, we also follow [59, 117] to embed each word of the query by the Glove [83] model with a further multi-head self-attention module and a BiLSTM layer. Similarly, we can obtain the word-wise query representation \\(\\mathbf{Q} = \\{\\pmb{q}_j\\}_{j=1}^M \\in \\mathbb{R}^{M \\times D_h}\\)." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.498, + 0.129 + ], + "angle": 0, + "content": "X. Fang, Z. Xiong, W. Fang et al." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.147, + 0.785, + 0.162 + ], + "angle": 0, + "content": "3.2 Game-theoretic Learning for Self-modal Semantic Enhancement" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.18, + 0.784, + 0.27 + ], + "angle": 0, + "content": "We argue that the instance relations within each modality are important to infer the consecutive visual event contents or linguistic phrase contexts. To achieve this self-modal learning, we review previous classic game-theoretic value [34, 85] to explore the weighted semantic contribution of each frame/word to the whole video/query semantics for enhancing the self-modal representation of both video and query." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.272, + 0.785, + 0.376 + ], + "angle": 0, + "content": "Visual self-modal game. Generally, a cooperative game consists of a player set \\(\\mathcal{P} = \\{1,2,\\dots,n\\}\\) and a game function \\(g(\\cdot)\\). In detail, function \\(g(\\cdot)\\) maps each player subset \\(S \\subseteq \\mathcal{P}\\) to a score value, which indicates the payoff when all players in coalition \\(S\\) work together in the game. To capture the inherent correlations within the video modality, we directly take the video frames \\(\\mathbf{V} = \\{\\pmb{v}_i\\}_{i=1}^T\\) as frame players \\(\\mathcal{P} = \\mathbf{V}, n = T\\) to play the game and evaluate frame-wise contribution to frame-level video semantics." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.378, + 0.785, + 0.449 + ], + "angle": 0, + "content": "Since different frames contribute and weight differently to video understanding, we attempt to adopt a game-theoretic value to measure the importance of each frame player in the cooperative game. Specifically, each frame player \\( \\pmb{v}_i \\) in player set \\( \\pmb{V} \\) has a weighted marginal contribution to the entire game, which can be measured by the game-theoretic value (e.g., Banzhaf value [34]) as:" + }, + { + "type": "equation", + "bbox": [ + 0.354, + 0.463, + 0.785, + 0.496 + ], + "angle": 0, + "content": "\\[\nB \\left(\\boldsymbol {v} _ {i} \\mid \\boldsymbol {V}\\right) = \\frac {1}{2 ^ {T - 1}} \\sum_ {\\mathcal {S} \\subseteq \\boldsymbol {V} \\backslash \\boldsymbol {v} _ {i}} [ g (\\mathcal {S} \\cup \\{\\boldsymbol {v} _ {i} \\}) - g (\\mathcal {S}) ], \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.511, + 0.785, + 0.647 + ], + "angle": 0, + "content": "where \\( S \\subseteq V \\backslash \\mathbf{v}_i \\) represents a group formed without \\( \\mathbf{v}_i \\), \\( \\frac{1}{2^{T-1}} \\) is the likelihood of \\( S \\) being sampled, and \\( g(S \\cup \\{\\mathbf{v}_i\\}) - g(S) \\) denotes the marginal contribution brought by player \\( \\mathbf{v}_i \\) joining the coalition \\( S \\). \\( B(\\mathbf{v}_i | V) \\) indicates the ability of player \\( \\mathbf{v}_i \\) to influence the video-level semantic by calculating the weighted marginal contribution of player \\( \\mathbf{v}_i \\) to all possible coalitions \\( S \\). Generally, within the context of our visual game, coalitions in the video modality represent consecutive frames of the same event or frames that are semantically similar. In this way, we can enumerate and learn all possible relations between the frames within the video for more contextual visual feature learning." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.648, + 0.786, + 0.842 + ], + "angle": 0, + "content": "Visual soft supervision. To assist the above game-theoretic value learning, we first develop an additional learnable video-domain relevance header to predict the same self-modal relevance \\(\\hat{R}^v = \\{\\hat{r}_i^v\\}_{i=1}^T = \\text{header}(\\mathbf{V})\\) as Banzhaf values \\(\\{B(\\mathbf{v}_i|\\mathbf{V})\\}_{i=1}^T\\) among all frames for soft supervision (more discussions will be illustrated later). Specifically, the header is implemented by: 1) 1D convolutional layers for local self-consistency modeling, 2) a self-attention module for global self-relevance capturing, 3) a convolutional layer for decoding. Then, we optimize the Kullback-Leibler Divergence (KLD) between the predicted relevance \\(\\hat{r}_i^v\\) and the Banzhaf value \\(B(\\mathbf{v}_i|\\mathbf{V})\\) for learning the consistent self-modal similarity. In detail, we define the game-guided frame-wise probability distribution as \\(D^v = [p_1^v,p_2^v,\\dots,p_i^v,\\dots,p_T^v]^\\top\\), where \\(p_i^v = \\frac{\\exp(B(\\mathbf{v}_i|\\mathbf{V}))}{\\sum_{l=1}^T\\exp(B(\\mathbf{v}_l|\\mathbf{V}))}\\). Similarly, the predicted frame-wise probability distribution calculated by the output of the self-relevance header can be denoted as \\(\\hat{D}^v = [\\hat{p}_1^v,\\hat{p}_2^v,\\dots,\\hat{p}_i^v,\\dots,\\hat{p}_T^v]^\\top\\), where" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.412, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "Rethinking WS-VTG From a Game Perspective" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.144, + 0.731, + 0.169 + ], + "angle": 0, + "content": "\\(\\hat{p}_i^v = \\frac{\\exp(\\hat{r}_i^v)}{\\sum_{l=1}^{T}\\exp(\\hat{r}_l^v)}\\). Finally, the self-relevance loss \\(\\mathcal{L}_{self}^v\\) for each video is:" + }, + { + "type": "equation", + "bbox": [ + 0.32, + 0.177, + 0.786, + 0.202 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {s e l f} ^ {v} = \\operatorname {K L D} \\left(\\hat {D} ^ {v} | | D ^ {v}\\right) = - \\frac {1}{T} \\sum_ {i = 1} ^ {T} p _ {i} ^ {v} \\left(\\log \\hat {p} _ {i} ^ {v} - \\log p _ {i} ^ {v}\\right). \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.209, + 0.784, + 0.252 + ], + "angle": 0, + "content": "Based on the learned game-theoretic value \\( B(\\pmb{v}_i|\\pmb{V}) \\) for each player \\( \\pmb{v}_i \\) in player set \\( \\pmb{V} \\), we deploy a Softmax function to obtain the self-modal enhanced video features \\( \\widetilde{\\pmb{v}}_i \\) as:" + }, + { + "type": "equation", + "bbox": [ + 0.308, + 0.259, + 0.786, + 0.289 + ], + "angle": 0, + "content": "\\[\n\\widetilde {\\boldsymbol {v}} _ {i} = \\boldsymbol {v} _ {i} + \\sum_ {\\mathcal {S} \\subseteq \\mathcal {V} \\backslash \\boldsymbol {v} _ {i}} \\operatorname {S o f t m a x} [ g (\\mathcal {S} \\cup \\{\\boldsymbol {v} _ {i} \\}) - g (\\mathcal {S}) ] \\cdot (\\mathcal {S} \\cup \\{\\boldsymbol {v} _ {i} \\}). \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.299, + 0.784, + 0.328 + ], + "angle": 0, + "content": "The enhanced video features \\(\\widetilde{\\pmb{V}} = \\{\\widetilde{\\pmb{v}}_i\\}_{i=1}^T\\) aggregates the relevant contexts from the whole video." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.329, + 0.784, + 0.374 + ], + "angle": 0, + "content": "Linguistic self-modal game. Similarly, we deem query words \\( \\mathbf{Q} = \\{\\mathbf{q}_j\\}_{j=1}^M \\) as players \\( \\mathcal{P} = \\mathbf{Q}, n = M \\) and calculate each word's marginal contribution \\( B(\\mathbf{q}_j|\\mathbf{Q}) \\) to the whole query \\( \\mathbf{Q} \\) by formulating the linguistic self-modal game as:" + }, + { + "type": "equation", + "bbox": [ + 0.351, + 0.382, + 0.786, + 0.415 + ], + "angle": 0, + "content": "\\[\nB \\left(\\boldsymbol {q} _ {j} \\mid \\boldsymbol {Q}\\right) = \\frac {1}{2 ^ {M - 1}} \\sum_ {\\mathcal {S} \\subseteq \\boldsymbol {Q} \\backslash \\boldsymbol {q} _ {j}} [ g (\\mathcal {S} \\cup \\{\\boldsymbol {q} _ {j} \\}) - g (\\mathcal {S}) ]. \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.423, + 0.784, + 0.505 + ], + "angle": 0, + "content": "Linguistic soft supervision. We first take the game probability distributions \\( D^{q} = [p_{1}^{q}, p_{2}^{q}, \\dots, p_{j}^{q}, \\dots, p_{M}^{q}]^{\\top} \\) as the soft label, and then supervise the predicted distributions \\( \\hat{D}^{q} = [\\hat{p}_{1}^{q}, \\hat{p}_{2}^{q}, \\dots, \\hat{p}_{j}^{q}, \\dots, \\hat{p}_{M}^{q}]^{\\top} \\) of another header by KLD loss \\( \\mathcal{L}_{self}^{q} \\) similar to Eq. (2). At last, we can obtain the enhanced query features \\( \\tilde{\\boldsymbol{Q}} = \\{\\tilde{\\boldsymbol{q}}_{j}\\}_{j=1}^{M} \\) by:" + }, + { + "type": "equation", + "bbox": [ + 0.308, + 0.515, + 0.786, + 0.546 + ], + "angle": 0, + "content": "\\[\n\\widetilde {\\boldsymbol {q}} _ {j} = \\boldsymbol {q} _ {j} + \\sum_ {\\mathcal {S} \\subseteq \\mathcal {Q} \\backslash \\boldsymbol {q} _ {j}} \\operatorname {S o f t m a x} [ g (\\mathcal {S} \\cup \\{\\boldsymbol {q} _ {j} \\}) - g (\\mathcal {S}) ] \\cdot (\\mathcal {S} \\cup \\{\\boldsymbol {q} _ {j} \\}). \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.553, + 0.786, + 0.735 + ], + "angle": 0, + "content": "Discussion on learnable header. Although we can directly utilize the game interaction value to enumerate the relations between different players via Eqs. (1) and (4), there are still two challenges: 1) No supervision signal: The performance of the game interaction severely depends on the quality of the learned player features. However, there is no supervision for learning the representative features. 2) Complex and time-consuming: The game process is complex and costs much time and resources. To this end, we develop a learnable header to mimic the game interaction by utilizing the KLD loss. This KLD function serves as a soft supervision label to not only learn the consistency between the game interaction and header, but also potentially train more distinguishing player features of each modality. During the inference, we can solely utilize the prediction of header to model the game interaction instead of Eqs. (1) and (4)." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.756, + 0.724, + 0.786 + ], + "angle": 0, + "content": "3.3 Game-theoretic Interaction for Cross-modal Multi-level Alignment" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.784, + 0.84 + ], + "angle": 0, + "content": "After obtaining the self-enhanced video and query features, we further develop a cross-modal game between frames and words to handle their uncertainty during their fine-grained semantic alignment." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.498, + 0.129 + ], + "angle": 0, + "content": "X. Fang, Z. Xiong, W. Fang et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.407 + ], + "angle": 0, + "content": "Cross-modal game. From a game-theoretic perspective, we attempt to conceptualize the problem of fine-grained cross-modal alignment as a collaborative effort between video frames and query words. Specifically, when a video frame and a query word exhibit a high degree of semantic similarity, they are more likely to collaborate and their union will have a larger score contributed to the final moment. To this end, we take the enhanced video features \\(\\widetilde{V} = \\{\\widetilde{v}_i\\}_{i=1}^T\\) and enhanced query features \\(\\widetilde{Q} = \\{\\widetilde{q}_j\\}_{j=1}^M\\) as two different kinds of players \\(\\mathcal{P} = \\{\\widetilde{v}_i\\}_{i=1}^T \\cup \\{\\widetilde{q}_j\\}_{j=1}^M\\), \\(n = T + M\\). In a cooperative game, different players tend to work together in groups, called coalitions, to achieve a common goal, i.e., representing the same semantics of the target event. In the case of two players \\(\\widetilde{v}_i\\) and \\(\\widetilde{q}_j\\) in set \\(\\mathcal{P}\\), it may occur that \\(g(\\{\\widetilde{v}_i\\})\\) and \\(g(\\{\\widetilde{q}_j\\})\\) are small individually, but at the same time the reward of their forming coalition \\(g(\\{\\widetilde{v}_i, \\widetilde{q}_j\\})\\) is considerable. This is because players in the coalition interact and collaborate with each other, which may bring additional benefits (or costs) to the game. Therefore, we also need to measure the additional benefits brought by the coalition compared with the costs of the lost coalitions of these players with others. To form a better coalition \\(\\{\\widetilde{v}_i, \\widetilde{q}_j\\} \\subseteq \\mathcal{P}\\), we utilize the effective interaction index \\(I_B(\\{\\widetilde{v}_i, \\widetilde{q}_j\\})\\) like [34] to compute the additional benefit as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.235, + 0.414, + 0.786, + 0.448 + ], + "angle": 0, + "content": "\\[\nI _ {B} \\left(\\left\\{\\widetilde {\\boldsymbol {v}} _ {i}, \\widetilde {\\boldsymbol {q}} _ {j} \\right\\}\\right) = \\frac {1}{2 ^ {n - 2}} \\sum_ {\\mathcal {S} \\subseteq \\mathcal {P} \\backslash \\left\\{\\widetilde {\\boldsymbol {v}} _ {i}, \\widetilde {\\boldsymbol {q}} _ {j} \\right\\}} \\left[ g \\left(\\mathcal {S} \\cup \\left\\{\\widetilde {\\boldsymbol {v}} _ {i}, \\widetilde {\\boldsymbol {q}} _ {j} \\right\\}\\right) + g (\\mathcal {S}) - g \\left(\\mathcal {S} \\cup \\left\\{\\widetilde {\\boldsymbol {v}} _ {i} \\right\\}\\right) - g \\left(\\mathcal {S} \\cup \\left\\{\\widetilde {\\boldsymbol {q}} _ {j} \\right\\}\\right) \\right], \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.456, + 0.788, + 0.577 + ], + "angle": 0, + "content": "where \\(\\frac{1}{2^{n-2}}\\) is the likelihood of \\(\\mathcal{S} \\subseteq \\mathcal{P} \\backslash \\{\\widetilde{v}_i, \\widetilde{q}_j\\}\\) being sampled. Intuitively, \\(I_B(\\{\\widetilde{v}_i, \\widetilde{q}_j\\})\\) quantifies this additional benefit compared to when players \\(\\widetilde{v}_i\\) and \\(\\widetilde{q}_j\\) work independently. It also embodies the tendency of interaction between frame-player and word-player. A higher value of \\(I_B(\\{\\widetilde{v}_i, \\widetilde{q}_j\\})\\) signifies that players \\(\\widetilde{v}_i\\) and \\(\\widetilde{q}_j\\) cooperate closely with each other, and the formed coalition will bring additional high returns. By this interaction index, we can measure the closeness between plays \\(\\widetilde{v}_i\\) and \\(\\widetilde{q}_j\\) to achieve fine-grained alignment between video and query." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.578, + 0.788, + 0.735 + ], + "angle": 0, + "content": "Game details. To ensure complete consistency between the cooperative game and cross-modal alignment learning, the game function \\( g \\) should meet the following criteria: 1) the game payoff benefits from strongly corresponding semantic pairs \\( \\{\\pmb{v}_i^+, \\pmb{q}_j^+\\} \\), i.e., \\( g(\\mathcal{P}) - g(\\mathcal{P} \\setminus \\{\\pmb{v}_i^+, \\pmb{q}_j^+\\} \\cup \\{[\\{\\pmb{v}_i^+, \\pmb{q}_j^+\\}]\\}) < 0 \\); 2) the game payoff is compromised by semantically opposite pairs \\( \\{\\pmb{v}_i^-, \\pmb{q}_j^-\\} \\), i.e., \\( g(\\mathcal{P}) - g(\\mathcal{P} \\setminus \\{\\pmb{v}_i^-, \\pmb{q}_j^-\\} \\cup \\{[\\{\\pmb{v}_i^-, \\pmb{q}_j^-\\}]\\}) > 0 \\); 3) when there are no players to cooperate, the payoff is zero, i.e., \\( g(\\{\\pmb{v}_i\\}_{i=1}^T) = g(\\{\\pmb{q}_j\\}_{j=1}^M) = g(\\phi) = 0 \\), where \\( \\phi \\) is the empty set. It should be noted that any function that satisfies the aforementioned conditions can be employed as the game function \\( g(\\cdot) \\). To simplify matters, we utilize the cosine similarity as \\( g \\) in all games." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.788, + 0.796 + ], + "angle": 0, + "content": "During the game, when players \\( \\mathbf{v}_i \\) and \\( \\mathbf{q}_j \\) form a coalition \\( \\{\\mathbf{v}_i, \\mathbf{q}_j\\} \\), we deem \\( [\\{\\mathbf{v}_i, \\mathbf{q}_j\\}] \\) as a singular hypothetical player, which is the union of the player in \\( \\{\\mathbf{v}_i, \\mathbf{q}_j\\} \\). Then, the reduced game is formed by removing the individual players in \\( \\{\\mathbf{v}_i, \\mathbf{q}_j\\} \\) from the game and adding \\( [\\{\\mathbf{v}_i, \\mathbf{q}_j\\}] \\) to the game." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.796, + 0.788, + 0.843 + ], + "angle": 0, + "content": "Word-level alignment with local game interaction. Similar to the self-modal game, we develop an alignment header to predict the word-frame alignment matrix \\( A = [a_{i,j}]^{T \\times M} \\) between frame \\( \\widetilde{\\boldsymbol{v}}_i \\) and word \\( \\widetilde{\\boldsymbol{q}}_j \\), supervised by" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.412, + 0.115, + 0.733, + 0.13 + ], + "angle": 0, + "content": "Rethinking WS-VTG From a Game Perspective" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.146, + 0.784, + 0.207 + ], + "angle": 0, + "content": "\\(I_B(\\{\\widetilde{v}_i,\\widetilde{q}_j\\})\\) . Then, we optimize the Kullback-Leibler Divergence (KLD) between the \\(I_B(\\{\\widetilde{v}_i,\\widetilde{q}_j\\})\\) and \\(a_{i,j}\\) , to bring the probability distributions of the output of the alignment header and interaction index close together to establish word-level semantic alignment between frame players and word players." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.208, + 0.788, + 0.333 + ], + "angle": 0, + "content": "In detail, we define the word-to-frame probability distribution \\(d_{w2f}^{j}\\) of the \\(j\\)-th word-to-frame alignment as \\(d_{w2f}^{j} = [p_{1,j}, \\dots, p_{t,j}]\\), where \\(p_{i,j} = \\frac{\\exp(I_B(\\{\\widetilde{v}_i, \\widetilde{q}_j\\}))}{\\sum_{l=1}^T \\exp(I_B(\\{\\widetilde{v}_l, \\widetilde{q}_j\\}))}\\) and the total query-to-video probability distribution \\(D_{q2v} = [d_{w2f}^1, \\dots, d_{w2f}^M]^\\top\\). Similarly, the predicted fine-grained probability distribution can be represented as \\(\\hat{d}_{w2f}^{j} = [\\hat{p}_{1,j}, \\dots, \\hat{p}_{t,j}]\\), where \\(\\hat{p}_{i,j} = \\frac{\\exp(a_{i,j})}{\\sum_{l=1}^t \\exp(a_{l,j})}\\), and the total predicted probability distribution is defined as \\(\\hat{D}_{q2v} = [\\hat{d}_{w2f}^1, \\dots, \\hat{d}_{w2f}^M]^\\top\\). At last, the word-level alignment loss \\(\\mathcal{L}_{word}\\) is formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.265, + 0.344, + 0.786, + 0.369 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {w o r d}} = \\operatorname {K L D} \\left(\\hat {D} _ {q 2 v} | | D _ {q 2 v}\\right) = - \\frac {1}{T M} \\sum_ {i = 1} ^ {T} \\sum_ {j = 1} ^ {M} p _ {i, j} \\left(\\log \\hat {p} _ {i, j} - \\log p _ {i, j}\\right). \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.379, + 0.788, + 0.628 + ], + "angle": 0, + "content": "Phrase- and sentence-level alignment with global game interaction. Although the above cross-modal game-theoretic interaction between each possible player-pair \\(\\{\\widetilde{v}_i,\\widetilde{q}_j\\}\\) addresses the problem of word-level alignment between word player \\(\\widetilde{v}_i\\) and frame player \\(\\widetilde{q}_j\\), it is still limited to aligning frames with isolated words, and fails to explore the relations between frames with more contextual phrase- or sentence-level semantics for better understanding temporal events. To this end, we introduce a multi-level game interaction to measure the semantic consistency between frames with different-grained textual semantics. Specifically, we follow [63, 74] to utilize Sequential Query Attention Network (SQAN) to extract the phrase-level textual semantics \\(\\tilde{\\mathbf{Q}}^p = \\{\\tilde{\\mathbf{q}}_j^p\\}_{j = 1}^P\\), and \\(P\\) is the phrase number. The sentence-level semantics \\(\\tilde{\\mathbf{Q}}^s\\) is obtained by concatenating the last hidden states in both the forward and backward BiLSTM. Similar to worl-level alignment, we compute both phrase- and sentence-level text-to-frame game interaction as \\(I_B(\\{\\widetilde{v}_i,\\widetilde{q}_j^p\\})\\), \\(I_B(\\{\\widetilde{v}_i,\\tilde{Q}^s\\})\\) via Eq. (6). We design corresponding alignment headers (phrase-level cross-modal header and sentence-level cross-modal header) to predict their alignment matrix via \\(\\mathcal{L}_{phra},\\mathcal{L}_{sent}\\)." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.652, + 0.509, + 0.667 + ], + "angle": 0, + "content": "3.4 Model Training and Inference" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.679, + 0.785, + 0.723 + ], + "angle": 0, + "content": "Weak supervision for video-query matching. We first interact each frame with the semantics of words, phrases and the whole sentence to calculate the overall query-to-frame matching degree by:" + }, + { + "type": "equation", + "bbox": [ + 0.264, + 0.733, + 0.786, + 0.759 + ], + "angle": 0, + "content": "\\[\nm _ {i} = \\frac {1}{3} \\left(I _ {B} \\left(\\{\\widetilde {\\boldsymbol {v}} _ {i}, \\widetilde {\\boldsymbol {Q}} ^ {s} \\}\\right) + \\frac {1}{M} \\sum_ {j = 1} ^ {M} I _ {B} \\left(\\{\\widetilde {\\boldsymbol {v}} _ {i}, \\widetilde {\\boldsymbol {q}} _ {j} \\}\\right) + \\frac {1}{P} \\sum_ {l = 1} ^ {P} I _ {B} \\left(\\left\\{\\widetilde {\\boldsymbol {v}} _ {i}, \\widetilde {\\boldsymbol {q}} _ {l} ^ {p} \\right\\}\\right)\\right), \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.769, + 0.717, + 0.783 + ], + "angle": 0, + "content": "where \\(m_{i}\\) is the overall query-to-frame matching degree of \\(i\\)-th frame." + }, + { + "type": "text", + "bbox": [ + 0.24, + 0.784, + 0.702, + 0.798 + ], + "angle": 0, + "content": "Then, the whole video-sentence matching score is calculated by:" + }, + { + "type": "equation", + "bbox": [ + 0.333, + 0.812, + 0.786, + 0.845 + ], + "angle": 0, + "content": "\\[\n\\phi^ {c} = \\sum_ {i = 1} ^ {T} w _ {i} \\cdot m _ {i}, w _ {i} = \\frac {\\exp \\left(I _ {B} \\left(\\left\\{\\widetilde {\\boldsymbol {v}} _ {i} , \\widetilde {\\boldsymbol {Q}} ^ {s} \\right\\}\\right)\\right)}{\\sum_ {i = 1} ^ {T} \\exp \\left(I _ {B} \\left(\\left\\{\\widetilde {\\boldsymbol {v}} _ {i} , \\widetilde {\\boldsymbol {Q}} ^ {s} \\right\\}\\right)\\right)}, \\tag {9}\n\\]" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.498, + 0.129 + ], + "angle": 0, + "content": "X. Fang, Z. Xiong, W. Fang et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.209 + ], + "angle": 0, + "content": "where we utilize the sentence-level similarity to represent the contribution of each frame to the target moment among the entire video. For matched video \\( V_{k} \\) and query \\( Q_{k} \\), we calculate the video-sentence matching score \\( \\phi_{V_k,Q_k}^c \\). The scores of unmatched pairs are expressed as \\( \\phi_{V_k,Q_l}^c \\) and \\( \\phi_{Q_k,V_l}^c \\) where \\( l\\neq k \\)." + }, + { + "type": "text", + "bbox": [ + 0.239, + 0.209, + 0.785, + 0.224 + ], + "angle": 0, + "content": "Finally, we utilize the cross-modal contrastive loss [79] for weak supervision:" + }, + { + "type": "equation", + "bbox": [ + 0.23, + 0.242, + 0.784, + 0.276 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {C} = - \\frac {1}{N _ {b}} \\sum_ {k = 1} ^ {N _ {b}} \\log \\frac {\\exp \\left(\\phi_ {V _ {k} , Q _ {k}} ^ {c} / \\tau\\right)}{\\sum_ {l = 1} ^ {N _ {b}} \\exp \\left(\\phi_ {V _ {k} , Q _ {l}} ^ {c} / \\tau\\right)} - \\frac {1}{N _ {b}} \\sum_ {k = 1} ^ {N _ {b}} \\log \\frac {\\exp \\left(\\phi_ {Q _ {k} , V _ {k}} ^ {c} / \\tau\\right)}{\\sum_ {l = 1} ^ {N _ {b}} \\exp \\left(\\phi_ {Q _ {k} , V _ {l}} ^ {c} / \\tau\\right)}, \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.282, + 0.646, + 0.296 + ], + "angle": 0, + "content": "where \\(\\tau\\) is the temperature parameter, \\(N_{b}\\) is the batch size." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.297, + 0.785, + 0.433 + ], + "angle": 0, + "content": "Reconstruction loss. To further enhance the semantic relevance of the highlighted frames to the sentence, we reconstruct the query conditioned on highlighted frames. Following popular reconstruction strategy [49,124], we randomly replace the 1/3 words in the original query with a specific symbol, and predict the next word by given a prefix of the query and highlighted frames feature. Specifically, we embed the masked query \\( \\mathbf{Q}^m \\) by Glove and predict the next word \\( \\hat{\\pmb{q}}_{i+1} \\) using the conditioned transformer following CNM [124] with input \\( (\\tilde{\\mathbf{V}}, C_m, Q_{1:i}^m) \\), where \\( C_m = \\{m_1, m_2, \\dots, m_T\\} \\) represents the highlighted frame mask, \\( Q_{1:i}^m \\) denotes the masked query words from the 1-st to \\( i \\)-th." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.433, + 0.784, + 0.462 + ], + "angle": 0, + "content": "Then, we introduce the cross-entropy loss \\(\\mathcal{L}_{rec}\\) to calculate the probability distribution \\(\\mathcal{P}^q\\) different between the predict word \\(\\hat{\\pmb{q}}_{i + 1}\\) and the real word as:" + }, + { + "type": "equation", + "bbox": [ + 0.386, + 0.469, + 0.784, + 0.492 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {r e c} = - \\sum_ {i = 1} ^ {M - 1} \\log \\mathcal {P} ^ {q} \\left(\\boldsymbol {q} _ {i + 1} \\mid \\hat {\\boldsymbol {q}} _ {i + 1}\\right). \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.498, + 0.784, + 0.556 + ], + "angle": 0, + "content": "Overall training. To train the whole model, we combine game-guided self-relevance losses \\(\\mathcal{L}_{self} = \\mathcal{L}_{self}^v +\\mathcal{L}_{self}^q\\), game alignment loss \\(\\mathcal{L}_{align} = \\mathcal{L}_{word} + \\mathcal{L}_{phra} + \\mathcal{L}_{sent}\\), video-level contrastive loss \\(\\mathcal{L}_C\\), and reconstruction loss \\(\\mathcal{L}_{rec}\\) with trade-off hyper-parameters \\((\\alpha_{1},\\alpha_{2},\\alpha_{3}\\) and \\(\\alpha_4)\\) to define the overall loss as:" + }, + { + "type": "equation", + "bbox": [ + 0.363, + 0.565, + 0.784, + 0.578 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\alpha_ {1} \\mathcal {L} _ {\\text {s e l f}} + \\alpha_ {2} \\mathcal {L} _ {\\text {a l i g n}} + \\alpha_ {3} \\mathcal {L} _ {C} + \\alpha_ {4} \\mathcal {L} _ {\\text {r e c}}. \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.586, + 0.785, + 0.705 + ], + "angle": 0, + "content": "Inference. During inference, we directly utilize the self-relevance headers to enhance the video and query features, and exploit the multi-level alignment headers to predict the overall query-to-frame matching degree. We locate the frame with the highest score as the basic predicted moment, and add the left/right frames into the moment if the ratio of their scores to the frame score of the closest moment boundary is higher than a threshold. We repeat this step until no frame can be added. We can obtain more moments by locating different initial frames with different scores." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.728, + 0.376, + 0.745 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.757, + 0.533, + 0.772 + ], + "angle": 0, + "content": "4.1 Datasets and Evaluation Metrics" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.784, + 0.84 + ], + "angle": 0, + "content": "Charades-STA. The Charades-STA dataset [32] is built based on the Charades [86], which contains 6,672 videos of indoor activities and involves 16,128 query-video pairs. There are 12,408 pairs used for training and 3,720 used for testing. The average duration of each video is 29.76 seconds." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.412, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Rethinking WS-VTG From a Game Perspective" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.236 + ], + "angle": 0, + "content": "ActivityNet Caption. The ActivityNet Caption dataset [43] contains 20,000 videos with 100,000 queries, where 37,421 query-video pairs are used for training and 34,536 are used for testing. On average, each video in ActivityNet Caption has 3.65 annotated moments and each annotated moment lasts for 36 seconds. For fair comparison, we follow [10, 73, 88, 89, 123, 124] to use val1 as valuation and report the test performance on val2." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.238, + 0.785, + 0.283 + ], + "angle": 0, + "content": "Evaluation metrics. Following previous works, we adopt the metrics “R@n, IoU=m” to evaluate our model, which presents the proportion of the top \\( n \\) predicted moments with IoU larger than \\( m \\)." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.304, + 0.457, + 0.319 + ], + "angle": 0, + "content": "4.2 Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.328, + 0.785, + 0.538 + ], + "angle": 0, + "content": "For fair comparison, we follow previous works to apply C3D [97] to encode the videos on ActivityNet Caption and I3D [5] on Charades-STA. We uniformly downsample the length of video feature sequences to \\( T = 200 \\) for ActivityNet Caption and \\( T = 64 \\) for Charades-STA. As for query encoding, we also follow previous works to utilize set the length of word feature sequences to \\( M = 20 \\), and utilize Glove [83] to embed each word. The dimension \\( D_h \\) is set to 512. These trade-off hyper-parameters are set to \\( \\alpha_1 = 1.0, \\alpha_2 = 1.0, \\alpha_3 = 1.0, \\alpha_4 = 10 \\). Since the calculation of the exact game-theoretic interaction is an NP-hard problem [71], we follow previous works to utilize sampling-based method [3,44] to obtain unbiased estimates for approximating it. We train the whole model for 100 epochs with batch size of 16 and early stopping strategy. Parameter optimization is performed by Adam [42] optimizer with learning rate \\( 3 \\times 10^{-4} \\), and linear decay of learning rate and gradient clipping of 1.0. The inference threshold is set to 0.8 in ActivityNet Caption and 0.9 in Charades-STA." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.56, + 0.373, + 0.574 + ], + "angle": 0, + "content": "4.3 Main Results" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.584, + 0.785, + 0.779 + ], + "angle": 0, + "content": "Quantitative comparison. We compare our method with the state-of-the-art methods, including fully-supervised (FS) methods and weakly-supervised (WS) methods. Best results are in bold. We introduce two types of cooperative game indices, i.e., Banzhaf index [34] and Shapley index [85], to our model. As summarized in Tab. 1, both two different variants of our method surpass all existing methods on both Charades-STA and ActivityNet Captions. Particularly, on Charades-STA in terms of \"R@5, IoU=0.5\", Ours(Banzhaf) outperforms the best compared method CPL by \\(4.53\\%\\). The main reason is that our game-theoretic-based approach can learn the local alignment between words and frames for accurately predicting the moment boundaries. Besides, the Shapley variant performs similar to the Banzhaf variant for this specific WS-VTG task, since they share the same game interaction formulation with different normalization. Overall, this table verifies the effectiveness of our game-based framework." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.781, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Generalization on cross-dataset evaluation. At last, we explore the generalization of models by cross-dataset evaluation in Tab. 2. Concretely, we first train a model on one dataset and then evaluate its performance on the other dataset. We can find that: 1) For all WS-VTG methods, the performance of" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.498, + 0.129 + ], + "angle": 0, + "content": "X. Fang, Z. Xiong, W. Fang et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.787, + 0.172 + ], + "angle": 0, + "content": "Table 1: Comparisons with state-of-the-art methods. \"WS\" and \"FS\" denotes weakly-supervised and fully-supervised methods, respectively." + }, + { + "type": "table", + "bbox": [ + 0.226, + 0.173, + 0.789, + 0.432 + ], + "angle": 0, + "content": "
TypeMethodCharades-STAActivityNet Caption
R@1, IoU=0.5R@1, IoU=0.7R@5, IoU=0.5R@5, IoU=0.7R@1, IoU=0.3R@1, IoU=0.5R@5, IoU=0.3R@5, IoU=0.5
FSCTRL [32]23.638.8958.9229.52-29.01-59.17
2DTAN [121]39.8123.2579.3351.1559.4544.5185.5377.13
DRN [115]53.0931.7589.0660.05-45.45-77.97
WSSCN [49]23.589.9771.8038.8747.2329.2271.4555.69
WSTAN [104]29.3512.2876.1341.5352.4530.0179.3863.42
ICVC [7]31.0216.5377.5341.9146.6229.5280.9266.61
MARN [87]33.8715.5473.9041.9448.5231.3775.9160.00
CRM [38]34.7616.37--55.2632.19--
CNM [124]35.1514.95--55.6833.33--
ACN [109]37.0215.26--57.6634.18--
VCA [105]38.1319.5778.7537.7550.4531.0071.7953.83
LCNet [112]39.1918.1780.5645.2448.4926.3382.5162.66
CPL [125]49.2422.3984.7152.3755.7331.3763.0543.13
DM2 [70]51.3923.72--54.7331.85--
Ours(Banzhaf)54.1726.4688.9355.7958.9435.6586.4170.26
Ours(Shapley)54.3126.2889.2455.6259.1735.8086.5970.03
" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.443, + 0.492, + 0.496 + ], + "angle": 0, + "content": "Table 2: Cross-dataset evaluation. \\(A\\rightarrow C\\) denotes training on ActivityNet Caption and evaluation on CharadesSTA, and vice versa." + }, + { + "type": "table", + "bbox": [ + 0.225, + 0.498, + 0.5, + 0.589 + ], + "angle": 0, + "content": "
MethodA→CC→A
R@1, IoU=0.7R@5, IoU=0.7R@1, IoU=0.5R@5, IoU=0.5
LCNet [112]12.0631.4710.6531.76
CPL [125]14.3235.8111.5929.88
Ours(Banzhaf)18.7442.5820.3742.96
Ours(Shapley)18.2643.1320.0441.42
" + }, + { + "type": "image", + "bbox": [ + 0.51, + 0.435, + 0.785, + 0.545 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.509, + 0.55, + 0.785, + 0.589 + ], + "angle": 0, + "content": "Fig. 3: Qualitative results. The colors of green lines represent different degrees of confidence." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.598, + 0.787, + 0.764 + ], + "angle": 0, + "content": "\\(\\mathrm{A} \\rightarrow \\mathrm{C}\\) achieves better than \\(\\mathrm{C} \\rightarrow \\mathrm{A}\\). This is because that ActivityNet Caption is larger and more complex than Charades-STA, thus the former is able to bring more generalized knowledge to the latter while the latter fails to provide general knowledge to the former. 2) Our two variants outperform state-of-the-art methods (e.g., LCNet) with clear margins on cross-dataset evaluation. We speculate it to: Previous works severely rely on the moment proposals, however, LCNet defines different sliding windows on different datasets for proposal generation. Therefore, its non-uniform proposal structure limits its generalization on cross-dataset evaluation. Instead, we learn multi-level frame-aware alignment and utilize frame-wise scores for predicting moment, which is more flexible and adaptable, thus improving the generalization-ability." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Qualitative comparison. In Fig. 3, we provide a qualitative example on ActivityNet Caption. The colors of green lines represent different degrees of interaction confidence, demonstrating that the game interaction can well learn the uncertain and fine-grained frame-word correspondence in the WS setting. Besides, all three-level interaction (word-level, phrase-level and sentence-level) contributes" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.412, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Rethinking WS-VTG From a Game Perspective" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "table_caption", + "bbox": [ + 0.245, + 0.145, + 0.757, + 0.159 + ], + "angle": 0, + "content": "Table 3: Comparison on speed per sample (s) and GPU memory cost (MB)." + }, + { + "type": "table", + "bbox": [ + 0.237, + 0.16, + 0.763, + 0.233 + ], + "angle": 0, + "content": "
MetricStageLCNet [112]CPL [125]Ours(Banzhaf)Ours(Shapley)
SpeedTraining0.79s0.27s0.94s0.98s
Inference0.62s0.21s0.10s0.11s
GPUTraining8647MB4156MB5237MB5482MB
MemoryInference7462MB3428MB2058MB2165MB
" + }, + { + "type": "table_caption", + "bbox": [ + 0.388, + 0.241, + 0.615, + 0.255 + ], + "angle": 0, + "content": "Table 4: Main ablation study." + }, + { + "type": "table", + "bbox": [ + 0.222, + 0.257, + 0.781, + 0.372 + ], + "angle": 0, + "content": "
Self-modal GameMulti-level Cross-modal GameBanzhaf-based GameShapley-based Game
Charades-STAActivityNet CaptionCharades-STAActivityNet Caption
Video Query GameWord- Phrase- Sentence- levelIoU=0.7IoU=0.5IoU=0.5R@1,R@5,R@1,R@5,R@1,R@5,R@1,R@5,
IoU=0.7IoU=0.5IoU=0.5IoU=0.5IoU=0.7IoU=0.5IoU=0.5IoU=0.5
××✓××19.3444.6729.8361.7219.2944.3130.1561.56
✓×✓××21.2547.8631.3963.9821.1047.7331.5463.67
✓✓✓××22.9750.6432.7265.8022.6950.3832.9765.65
✓✓✓✓×24.8153.2834.4668.0324.6453.0334.5967.82
✓✓✓✓✓26.4655.7935.6570.2626.2855.6235.8070.03
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.382, + 0.785, + 0.411 + ], + "angle": 0, + "content": "differently to the grounding, and our final method is able to localize more accurate moment boundaries than state-of-the-art method CPL." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.413, + 0.808, + 0.534 + ], + "angle": 0, + "content": "Complexity comparison. In Tab. 3, our method costs more training time/memory as we need to enumerate coalition samples. Specifically, we sampled 5500 samples. We find that, the model instability decreases along with the increase of the sampling number, and when the sample size ranges from 4500 to 5500, the error with the true value of banzhaf gradually decreases from \\(6.9\\%\\) to \\(1.1\\%\\). After exceeding 5500 samples, more samples will not affect the result. During inference, since we can directly utilize the learned game headers to generate game value instead of using complex game learning, our model is faster with lower memory." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.557, + 0.389, + 0.573 + ], + "angle": 0, + "content": "4.4 Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.583, + 0.787, + 0.613 + ], + "angle": 0, + "content": "In this section, we perform in-depth ablation studies to evaluate the effect of each component. We implement our model with different game settings." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.614, + 0.787, + 0.779 + ], + "angle": 0, + "content": "Main ablation. As shown in Tab. 4, we conduct ablation studies regarding the components (i.e., self-modal games in two modalities, and multi-level cross-modal games). We observe the following findings: 1) Our baseline model (only contains word-level cross-modal game) is powerful and outperforms some WS methods, demonstrating that it is effective to capture the uncertain yet fine-grained alignment for better determining the boundaries. 2) Both video and query self-modal games bring significant improvement to the overall model, indicating that self-enhancement is crucial for context integration. 3) Each query level of cross-modal game further boost the performance, verifying that phrase-and sentence-level semantics help to better understand temporal events. Overall, all components contribute a lot to final performance, showing their effectiveness." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Effect of game headers. The headers in both self- and cross-modal games are utilized to provide soft supervision for better game interaction and player-representation learning. To explore the impact of the structure of these headers, we compare several popular structures in Tab. 5. We find that the combination of" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.498, + 0.129 + ], + "angle": 0, + "content": "X. Fang, Z. Xiong, W. Fang et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.217, + 0.145, + 0.785, + 0.199 + ], + "angle": 0, + "content": "Table 5: Ablation study on each component, where \"LL\" means \"Linear Layer\", \"CL\" means \"Convolution Layer\", \"SA\" means \"Self-attention\", \"MSE\" means \"Mean Squared Error\", \"CE\" means \"Cross Entropy\", \"KLD\" means \"Kullback-Leibler Divergence\"; \"RL\" means \"Reconstruction Loss\"." + }, + { + "type": "table", + "bbox": [ + 0.22, + 0.2, + 0.789, + 0.401 + ], + "angle": 0, + "content": "
ComponentSettingBanzhaf-based GameShapley-based Game
Charades-STAActivityNet CaptionCharades-STAActivityNet Caption
R@1, IoU=0.7R@5, IoU=0.7R@1, IoU=0.5R@5, IoU=0.5R@1, IoU=0.7R@5, IoU=0.7R@1, IoU=0.5R@5, IoU=0.5
Game HeadersLL25.2853.9034.1568.0324.9753.6834.2567.94
CL25.5254.3134.5668.6425.2654.0334.8268.47
LL + SA25.9454.8335.2769.1825.7154.5935.5369.06
CL + SA26.4655.7935.6570.2626.2855.6235.8070.03
Soft SupervisionNone23.2151.6432.3666.4022.9951.4532.7166.18
MSE24.4253.1933.8567.5724.2452.8940.0667.33
CE24.7353.5834.1668.0424.5253.2634.3767.80
KLD26.4655.7935.6570.2626.2855.6235.8070.03
Grounding Lossesw/o RL24.3553.5734.1468.4224.0853.3134.4568.19
w/ RL26.4655.7935.6570.2626.2855.6235.8070.03
" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.418, + 0.785, + 0.461 + ], + "angle": 0, + "content": "convolution layer and self-attention can capture both local and global interaction, so it achieves the best performance and is beneficial for cross-modal multi-level interaction, illustrating the significance of our game headers." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.465, + 0.785, + 0.554 + ], + "angle": 0, + "content": "Necessity of soft supervision. To explore the effect of soft supervision, we conduct the ablation study to investigate the contribution of the KLD loss during the header learning in Tab. 5. The KLD loss achieves better performance than both MSE and CE, showing that it is more suitable for game headers to mimic game-theoretic interaction learning. Moreover, without soft supervision, the performance will drop a lot, showing the importance of the soft supervision." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.557, + 0.785, + 0.631 + ], + "angle": 0, + "content": "Analysis on reconstruction loss. Most previous WS-VTG works also additionally utilize the reconstruction paradigm to assist the model training. As shown in Tab. 5, we further conduct such ablation study to investigate the impact of reconstruction loss. It shows that the reconstruction loss can improve significantly the performance, demonstrating its effectiveness." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.664, + 0.357, + 0.68 + ], + "angle": 0, + "content": "5 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.704, + 0.785, + 0.84 + ], + "angle": 0, + "content": "In this paper, we rethink the limitations of previous weakly-supervised video temporal grounding works, and creatively propose to utilize the game-theoretic interaction to learn the uncertain relationship between video-query pairs with diverse granularity. Specifically, we first introduce two self-modal games to correlate the frames/words to enhance their contextual semantics. Then, we propose a cross-modal game to value fine-grained correspondence between frames and words with multiple levels for better determining the query-related moment boundaries. Extensive experiments demonstrate the effectiveness of our game-based framework on two challenging datasets." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.412, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Rethinking WS-VTG From a Game Perspective" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.323, + 0.16 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.233, + 0.173, + 0.785, + 0.214 + ], + "angle": 0, + "content": "1. Albarelli, A., Rodola, E., Torsello, A.: A game-theoretic approach to fine surface registration without initial motion estimation. In: 2010 IEEE computer society conference on computer vision and pattern recognition. pp. 430-437. IEEE (2010)" + }, + { + "type": "ref_text", + "bbox": [ + 0.233, + 0.215, + 0.785, + 0.255 + ], + "angle": 0, + "content": "2. Anne Hendricks, L., Wang, O., Shechtman, E., Sivic, J., Darrell, T., Russell, B.: Localizing moments in video with natural language. In: Proceedings of the IEEE International Conference on Computer Vision. pp. 5803-5812 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.233, + 0.256, + 0.785, + 0.296 + ], + "angle": 0, + "content": "3. Bachrach, Y., Markakis, E., Resnick, E., Procaccia, A.D., Rosenschein, J.S., Saberi, A.: Approximating power indices: theoretical and empirical analysis. Autonomous Agents and Multi-Agent Systems 20, 105-122 (2010)" + }, + { + "type": "ref_text", + "bbox": [ + 0.233, + 0.297, + 0.785, + 0.323 + ], + "angle": 0, + "content": "4. Banzhaf III, J.F.: Weighted voting doesn't work: A mathematical analysis. Rutgers L. Rev. 19, 317 (1964)" + }, + { + "type": "ref_text", + "bbox": [ + 0.233, + 0.324, + 0.785, + 0.363 + ], + "angle": 0, + "content": "5. Carreira, J., Zisserman, A.: Quo vadis, action recognition? a new model and the kinetics dataset. In: proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 6299-6308 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.233, + 0.364, + 0.785, + 0.404 + ], + "angle": 0, + "content": "6. Chalkiadakis, G., Elkind, E., Wooldridge, M.: Computational aspects of cooperative game theory. Synthesis Lectures on Artificial Intelligence and Machine Learning 5(6), 1-168 (2011)" + }, + { + "type": "ref_text", + "bbox": [ + 0.233, + 0.405, + 0.785, + 0.445 + ], + "angle": 0, + "content": "7. Chen, J., Luo, W., Zhang, W., Ma, L.: Explore inter-contrast between videos via composition for weakly supervised temporal sentence grounding. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 36, pp. 267-275 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.233, + 0.446, + 0.785, + 0.485 + ], + "angle": 0, + "content": "8. Chen, J., Ma, L., Chen, X., Jie, Z., Luo, J.: Localizing natural language in videos. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 33, pp. 8175-8182 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.233, + 0.486, + 0.785, + 0.527 + ], + "angle": 0, + "content": "9. Chen, L., Lu, C., Tang, S., Xiao, J., Zhang, D., Tan, C., Li, X.: Rethinking the bottom-up framework for query-based video localization. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 34, pp. 10551-10558 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.227, + 0.528, + 0.785, + 0.567 + ], + "angle": 0, + "content": "10. Chen, Z., Ma, L., Luo, W., Tang, P., Wong, K.Y.K.: Look closer to ground better: Weakly-supervised temporal grounding of sentence in video. arXiv preprint arXiv:2001.09308 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.227, + 0.568, + 0.785, + 0.609 + ], + "angle": 0, + "content": "11. Datta, A., Sen, S., Zick, Y.: Algorithmic transparency via quantitative input influence: Theory and experiments with learning systems. In: 2016 IEEE symposium on security and privacy. pp. 598-617. IEEE (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.227, + 0.61, + 0.785, + 0.649 + ], + "angle": 0, + "content": "12. Deng, S., Wen, J., Liu, C., Yan, K., Xu, G., Xu, Y.: Projective incomplete multiview clustering. IEEE Transactions on Neural Networks and Learning Systems pp. 1-13 (2023). https://doi.org/10.1109/TNNLS.2023.3242473" + }, + { + "type": "ref_text", + "bbox": [ + 0.227, + 0.65, + 0.785, + 0.69 + ], + "angle": 0, + "content": "13. Dong, J., Chen, X., Zhang, M., Yang, X., Chen, S., Li, X., Wang, X.: Partially relevant video retrieval. In: Proceedings of the 30th ACM International Conference on Multimedia. pp. 246-257 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.227, + 0.691, + 0.785, + 0.731 + ], + "angle": 0, + "content": "14. Dong, J., Li, X., Snoek, C.G.: Predicting visual features from text for image and video caption retrieval. IEEE Transactions on Multimedia 20(12), 3377-3388 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.227, + 0.731, + 0.785, + 0.772 + ], + "angle": 0, + "content": "15. Dong, J., Li, X., Xu, C., Yang, X., Yang, G., Wang, X., Wang, M.: Dual encoding for video retrieval by text. IEEE Transactions on Pattern Analysis and Machine Intelligence 44(8), 4065-4080 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.227, + 0.772, + 0.785, + 0.84 + ], + "angle": 0, + "content": "16. Dong, J., Peng, X., Ma, Z., Liu, D., Qu, X., Yang, X., Zhu, J., Liu, B.: From region to patch: Attribute-aware foreground-background contrastive learning for fine-grained fashion retrieval. In: Proceedings of the 46th International ACM SIGIR Conference on Research and Development in Information Retrieval. pp. 1273-1282 (2023)" + }, + { + "type": "list", + "bbox": [ + 0.227, + 0.173, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.271, + 0.115, + 0.498, + 0.129 + ], + "angle": 0, + "content": "X. Fang, Z. Xiong, W. Fang et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.147, + 0.785, + 0.189 + ], + "angle": 0, + "content": "17. Dong, J., Sun, S., Liu, Z., Chen, S., Liu, B., Wang, X.: Hierarchical contrast for unsupervised skeleton-based action representation learning. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 37, pp. 525-533 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.189, + 0.785, + 0.23 + ], + "angle": 0, + "content": "18. Dong, J., Wang, Y., Chen, X., Qu, X., Li, X., He, Y., Wang, X.: Reading-strategy inspired visual representation learning for text-to-video retrieval. IEEE Transactions on Circuits and Systems for Video Technology 32(8), 5680-5694 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.23, + 0.785, + 0.271 + ], + "angle": 0, + "content": "19. Donoser, M., Bischof, H.: Diffusion processes for retrieval revisited. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1320-1327 (2013)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.271, + 0.785, + 0.311 + ], + "angle": 0, + "content": "20. Dowdall, J., Pavlidis, I.T., Tsiamiyrtzis, P.: Coalitional tracking in facial infrared imaging and beyond. In: 2006 Conference on Computer Vision and Pattern Recognition Workshop. pp. 134-134. IEEE (2006)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.311, + 0.785, + 0.352 + ], + "angle": 0, + "content": "21. Fang, X., Easwaran, A., Genest, B.: Uncertainty-guided appearance-motion association network for out-of-distribution action detection. In: IEEE International Conference on Multimedia Information Processing and Retrieval (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.352, + 0.785, + 0.405 + ], + "angle": 0, + "content": "22. Fang, X., Fang, W., Liu, D., Qu, X., Dong, J., Zhou, P., Li, R., Xu, Z., Chen, L., Zheng, P., Cheng, Y.: Not all inputs are valid: Towards open-set video moment retrieval using language. In: Proceedings of the 32th ACM International Conference on Multimedia (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.406, + 0.785, + 0.433 + ], + "angle": 0, + "content": "23. Fang, X., Hu, Y.: Double self-weighted multi-view clustering via adaptive view fusion. arXiv preprint arXiv:2011.10396 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.433, + 0.785, + 0.473 + ], + "angle": 0, + "content": "24. Fang, X., Hu, Y., Zhou, P., Wu, D.: Animc: A soft approach for autoweighted noisy and incomplete multiview clustering. IEEE Transactions on Artificial Intelligence 3(2), 192-206 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.473, + 0.785, + 0.514 + ], + "angle": 0, + "content": "25. Fang, X., Hu, Y., Zhou, P., Wu, D.O.: V3h: View variation and view heredity for incomplete multiview clustering. IEEE Transactions on Artificial Intelligence 1(3), 233-247 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.514, + 0.785, + 0.568 + ], + "angle": 0, + "content": "26. Fang, X., Hu, Y., Zhou, P., Wu, D.O.: Unbalanced incomplete multi-view clustering via the scheme of view evolution: Weak views are meat; strong views do eat. IEEE Transactions on Emerging Topics in Computational Intelligence 6(4), 913-927 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.568, + 0.785, + 0.623 + ], + "angle": 0, + "content": "27. Fang, X., Liu, D., Fang, W., Zhou, P., Cheng, Y., Tang, K., Zou, K.: Annotations are not all you need: A cross-modal knowledge transfer network for unsupervised temporal sentence grounding. In: Findings of the Association for Computational Linguistics: EMNLP 2023. pp. 8721-8733 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.623, + 0.785, + 0.678 + ], + "angle": 0, + "content": "28. Fang, X., Liu, D., Fang, W., Zhou, P., Xu, Z., Xu, W., Chen, J., Li, R.: Fewer steps, better performance: Efficient cross-modal clip trimming for video moment retrieval using language. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 38, pp. 1735-1743 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.678, + 0.785, + 0.718 + ], + "angle": 0, + "content": "29. Fang, X., Liu, D., Zhou, P., Hu, Y.: Multi-modal cross-domain alignment network for video moment retrieval. IEEE Transactions on Multimedia 25, 7517-7532 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.718, + 0.785, + 0.772 + ], + "angle": 0, + "content": "30. Fang, X., Liu, D., Zhou, P., Nan, G.: You can ground earlier than see: An effective and efficient pipeline for temporal sentence grounding in compressed videos. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2448-2460 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.772, + 0.785, + 0.8 + ], + "angle": 0, + "content": "31. Fang, X., Liu, D., Zhou, P., Xu, Z., Li, R.: Hierarchical local-global transformer for temporal sentence grounding. IEEE Transactions on Multimedia (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.8, + 0.785, + 0.84 + ], + "angle": 0, + "content": "32. Gao, J., Sun, C., Yang, Z., Nevatia, R.: Tall: Temporal activity localization via language query. In: Proceedings of the IEEE International Conference on Computer Vision. pp. 5267-5275 (2017)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.147, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.412, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Rethinking WS-VTG From a Game Perspective" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.147, + 0.785, + 0.204 + ], + "angle": 0, + "content": "33. Gao, M., Davis, L., Socher, R., Xiong, C.: Wslln: Weakly supervised natural language localization networks. In: Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing. pp. 1481-1487 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.205, + 0.785, + 0.245 + ], + "angle": 0, + "content": "34. Grabisch, M., Roubens, M.: An axiomatic approach to the concept of interaction among players in cooperative games. International Journal of game theory 28, 547-565 (1999)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.246, + 0.785, + 0.286 + ], + "angle": 0, + "content": "35. Guo, C., Liu, D., Zhou, P.: A hybrid alignment loss for temporal moment localization with natural language. In: 2022 IEEE International Conference on Multimedia and Expo (ICME). pp. 1-6. IEEE (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.287, + 0.785, + 0.328 + ], + "angle": 0, + "content": "36. Guo, D., Li, K., Hu, B., Zhang, Y., Wang, M.: Benchmarking micro-action recognition: Dataset, method, and application. IEEE Transactions on Circuits and Systems for Video Technology 34(7), 6238-6252 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.328, + 0.785, + 0.383 + ], + "angle": 0, + "content": "37. Hendricks, L.A., Wang, O., Shechtman, E., Sivic, J., Darrell, T., Russell, B.: Localizing moments in video with temporal language. Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing p. 1380-1390 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.384, + 0.785, + 0.425 + ], + "angle": 0, + "content": "38. Huang, J., Liu, Y., Gong, S., Jin, H.: Cross-sentence temporal and semantic relations in video activity localisation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 7199-7208 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.425, + 0.785, + 0.48 + ], + "angle": 0, + "content": "39. Jiang, L., Wang, C., Ning, X., Yu, Z.: Lttpoint: A mlp-based point cloud classification method with local topology transformation module. In: 2023 7th Asian Conference on Artificial Intelligence Technology (ACAIT). pp. 783-789. IEEE (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.481, + 0.785, + 0.536 + ], + "angle": 0, + "content": "40. Jin, P., Huang, J., Xiong, P., Tian, S., Liu, C., Ji, X., Yuan, L., Chen, J.: Videotext as game players: Hierarchical banzhaf interaction for cross-modal representation learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2472-2482 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.536, + 0.785, + 0.577 + ], + "angle": 0, + "content": "41. Jin, S., Wang, S., Fang, F.: Game theoretical analysis on capacity configuration for microgrid based on multi-agent system. International Journal of Electrical Power & Energy Systems 125, 106485 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.577, + 0.785, + 0.605 + ], + "angle": 0, + "content": "42. Kingma, D.P., Ba, J.: Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980 (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.606, + 0.785, + 0.646 + ], + "angle": 0, + "content": "43. Krishna, R., Hata, K., Ren, F., Fei-Fei, L., Carlos Niebles, J.: Dense-c captioning events in videos. In: Proceedings of the IEEE International Conference on Computer Vision. pp. 706-715 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.647, + 0.571, + 0.66 + ], + "angle": 0, + "content": "44. Leech, D.: Computation of power indices (2002)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.661, + 0.785, + 0.688 + ], + "angle": 0, + "content": "45. Lehrer, E.: An axiomatization of the banzhaf value. International Journal of Game Theory 17, 89-99 (1988)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.689, + 0.785, + 0.743 + ], + "angle": 0, + "content": "46. Li, H., Cao, M., Cheng, X., Li, Y., Zhu, Z., Zou, Y.: G2l: Semantically aligned and uniform video grounding via geodesic and game theory. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 12032-12042 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.744, + 0.785, + 0.785 + ], + "angle": 0, + "content": "47. Li, J., HE, X., Wei, L., Qian, L., Zhu, L., Xie, L., Zhuang, Y., Tian, Q., Tang, S.: Fine-grained semantically aligned vision-language pre-training. In: Advances in Neural Information Processing Systems (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.786, + 0.785, + 0.84 + ], + "angle": 0, + "content": "48. Lin, K.Q., Zhang, P., Chen, J., Pramanick, S., Gao, D., Wang, A.J., Yan, R., Shou, M.Z.: Univtg: Towards unified video-language temporal grounding. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 2794-2804 (2023)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.147, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.498, + 0.129 + ], + "angle": 0, + "content": "X. Fang, Z. Xiong, W. Fang et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.147, + 0.785, + 0.189 + ], + "angle": 0, + "content": "49. Lin, Z., Zhao, Z., Zhang, Z., Wang, Q., Liu, H.: Weakly-supervised video moment retrieval via semantic completion network. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 34, pp. 11539-11546 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.191, + 0.785, + 0.245 + ], + "angle": 0, + "content": "50. Liu, C., Wen, J., Luo, X., Huang, C., Wu, Z., Xu, Y.: Dicnet: Deep instance-level contrastive network for double incomplete multi-view multi-label classification. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 37, pp. 8807-8815 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.247, + 0.785, + 0.288 + ], + "angle": 0, + "content": "51. Liu, C., Wen, J., Luo, X., Xu, Y.: Incomplete multi-view multi-label learning via label-guided masked view- and category-aware transformers. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 37, pp. 8816-8824 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.29, + 0.785, + 0.331 + ], + "angle": 0, + "content": "52. Liu, C., Wen, J., Wu, Z., Luo, X., Huang, C., Xu, Y.: Information recovery-driven deep incomplete multiview clustering network. IEEE Transactions on Neural Networks and Learning Systems pp. 1-11 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.333, + 0.785, + 0.374 + ], + "angle": 0, + "content": "53. Liu, D., Fang, X., Hu, W., Zhou, P.: Exploring optical-flow-guided motion and detection-based appearance for temporal sentence grounding. IEEE Transactions on Multimedia 25, 8539-8553 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.375, + 0.785, + 0.431 + ], + "angle": 0, + "content": "54. Liu, D., Fang, X., Qu, X., Dong, J., Yan, H., Yang, Y., Zhou, P., Cheng, Y.: Unsupervised domain adaptative temporal sentence localization with mutual information maximization. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 38, pp. 3567-3575 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.432, + 0.785, + 0.472 + ], + "angle": 0, + "content": "55. Liu, D., Fang, X., Zhou, P., Di, X., Lu, W., Cheng, Y.: Hypotheses tree building for one-shot temporal sentence localization. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 37, pp. 1640-1648 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.474, + 0.785, + 0.515 + ], + "angle": 0, + "content": "56. Liu, D., Hu, W.: Learning to focus on the foreground for temporal sentence grounding. In: Proceedings of the 29th International Conference on Computational Linguistics. pp. 5532-5541 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.517, + 0.785, + 0.558 + ], + "angle": 0, + "content": "57. Liu, D., Hu, W.: Skimming, locating, then perusing: A human-like framework for natural language video localization. In: Proceedings of the 30th ACM International Conference on Multimedia. pp. 4536-4545 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.56, + 0.785, + 0.613 + ], + "angle": 0, + "content": "58. Liu, D., Qu, X., Dong, J., Nan, G., Zhou, P., Xu, Z., Chen, L., Yan, H., Cheng, Y.: Filling the information gap between video and query for language-driven moment retrieval. In: Proceedings of the 31st ACM International Conference on Multimedia. pp. 4190-4199 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.615, + 0.785, + 0.671 + ], + "angle": 0, + "content": "59. Liu, D., Qu, X., Dong, J., Zhou, P., Cheng, Y., Wei, W., Xu, Z., Xie, Y.: Context-aware biaffine localizing network for temporal sentence grounding. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 11235-11244 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.672, + 0.785, + 0.727 + ], + "angle": 0, + "content": "60. Liu, D., Qu, X., Dong, J., Zhou, P., Xu, Z., Wang, H., Di, X., Lu, W., Cheng, Y.: Transform-equivariant consistency learning for temporal sentence grounding. ACM Transactions on Multimedia Computing, Communications and Applications 20(4), 1-19 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.728, + 0.785, + 0.798 + ], + "angle": 0, + "content": "61. Liu, D., Qu, X., Fang, X., Dong, J., Zhou, P., Nan, G., Tang, K., Fang, W., Cheng, Y.: Towards robust temporal activity localization learning with noisy labels. In: Proceedings of the 2024 Joint International Conference on Computational Linguistics, Language Resources and Evaluation (LREC-COLING 2024). pp. 16630-16642 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.799, + 0.785, + 0.84 + ], + "angle": 0, + "content": "62. Liu, D., Qu, X., Hu, W.: Reducing the vision and language bias for temporal sentence grounding. In: Proceedings of the 30th ACM International Conference on Multimedia. pp. 4092-4101 (2022)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.147, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.412, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Rethinking WS-VTG From a Game Perspective" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "19" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.147, + 0.787, + 0.189 + ], + "angle": 0, + "content": "63. Liu, D., Qu, X., Liu, X.Y., Dong, J., Zhou, P., Xu, Z.: Jointly cross-and self-modal graph attention network for query-based moment localization. In: Proceedings of the 28th ACM International Conference on Multimedia. pp. 4070-4078 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.191, + 0.787, + 0.245 + ], + "angle": 0, + "content": "64. Liu, D., Zhou, P.: Jointly visual-and semantic-aware graph memory networks for temporal sentence localization in videos. In: ICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). pp. 1-5. IEEE (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.247, + 0.787, + 0.288 + ], + "angle": 0, + "content": "65. Liu, D., Zhou, P., Xu, Z., Wang, H., Li, R.: Few-shot temporal sentence grounding via memory-guided semantic learning. IEEE Transactions on Circuits and Systems for Video Technology 33(5), 2491-2505 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.289, + 0.787, + 0.33 + ], + "angle": 0, + "content": "66. Liu, D., Zhu, J., Fang, X., Xiong, Z., Wang, H., Li, R., Zhou, P.: Conditional video diffusion network for fine-grained temporal sentence grounding. IEEE Transactions on Multimedia (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.332, + 0.787, + 0.373 + ], + "angle": 0, + "content": "67. Lundberg, S.M., Lee, S.I.: A unified approach to interpreting model predictions. In: Proceedings of the 31st International Conference on Neural Information Processing Systems. pp. 4768-4777 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.374, + 0.787, + 0.415 + ], + "angle": 0, + "content": "68. Ma, M., Yoon, S., Kim, J., Lee, Y., Kang, S., Yoo, C.D.: VLANet: Video-language alignment network for weakly-supervised video moment retrieval. In: Proceedings of the European Conference on Computer Vision. pp. 156-171 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.417, + 0.787, + 0.457 + ], + "angle": 0, + "content": "69. Ma, W.C., Huang, D.A., Lee, N., Kitani, K.M.: Forecasting interactive dynamics of pedestrians with fictitious play. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 774-782 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.459, + 0.787, + 0.5 + ], + "angle": 0, + "content": "70. Ma, Y., Liu, Y., Wang, L., Kang, W., Qiao, Y., Wang, Y.: Dual masked modeling for weakly-supervised temporal boundary discovery. IEEE Transactions on Multimedia (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.501, + 0.787, + 0.528 + ], + "angle": 0, + "content": "71. Matsui, Y., Matsui, T.: Np-completeness for calculating power indices of weighted majority games. Theoretical Computer Science 263(1-2), 305-310 (2001)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.53, + 0.787, + 0.571 + ], + "angle": 0, + "content": "72. Michalak, T.P., Aadithya, K.V., Szczepanski, P.L., Ravindran, B., Jennings, N.R.: Efficient computation of the shapley value for game-theoretic network centrality. Journal of Artificial Intelligence Research 46, 607-650 (2013)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.572, + 0.787, + 0.613 + ], + "angle": 0, + "content": "73. Mithun, N.C., Paul, S., Roy-Chowdhury, A.K.: Weakly supervised video moment retrieval from text queries. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 11592-11601 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.614, + 0.787, + 0.655 + ], + "angle": 0, + "content": "74. Mun, J., Cho, M., Han, B.: Local-global video-text interactions for temporal grounding. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 10810-10819 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.657, + 0.787, + 0.698 + ], + "angle": 0, + "content": "75. Ning, E., Wang, C., Zhang, H., Ning, X., Tiwari, P.: Occluded person re-identification with deep learning: a survey and perspectives. Expert Systems with Applications p. 122419 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.699, + 0.787, + 0.74 + ], + "angle": 0, + "content": "76. Ning, E., Wang, Y., Wang, C., Zhang, H., Ning, X.: Enhancement, integration, expansion: Activating representation of detailed features for occluded person re-identification. Neural Networks 169, 532-541 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.741, + 0.787, + 0.769 + ], + "angle": 0, + "content": "77. Ning, E., Zhang, C., Wang, C., Ning, X., Chen, H., Bai, X.: Pedestrian re-id based on feature consistency and contrast enhancement. Displays 79, 102467 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.77, + 0.787, + 0.797 + ], + "angle": 0, + "content": "78. Nowak, A.S.: On an axiomatization of the banzhaf value without the additivity axiom. International Journal of Game Theory 26, 137-141 (1997)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.798, + 0.787, + 0.826 + ], + "angle": 0, + "content": "79. Oord, A.v.d., Li, Y., Vinyals, O.: Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.827, + 0.749, + 0.84 + ], + "angle": 0, + "content": "80. Osborne, M.J., Rubinstein, A.: A course in game theory. MIT press (1994)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.147, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "20" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.498, + 0.129 + ], + "angle": 0, + "content": "X. Fang, Z. Xiong, W. Fang et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.148, + 0.784, + 0.203 + ], + "angle": 0, + "content": "81. Patel, R., Garnelo, M., Gemp, I., Dyer, C., Bachrach, Y.: Game-theoretic vocabulary selection via the shapley value and banzhaf index. In: Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies. pp. 2789-2798 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.204, + 0.785, + 0.244 + ], + "angle": 0, + "content": "82. Pavan, M., Pelillo, M.: A new graph-theoretic approach to clustering and segmentation. In: 2003 IEEE Computer Society Conference on Computer Vision and Pattern Recognition, 2003. Proceedings. vol. 1, pp. I-I. IEEE (2003)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.245, + 0.785, + 0.284 + ], + "angle": 0, + "content": "83. Pennington, J., Socher, R., Manning, C.D.: Glove: Global vectors for word representation. In: Proceedings of the Conference on Empirical Methods in Natural Language Processing. pp. 1532-1543 (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.285, + 0.785, + 0.325 + ], + "angle": 0, + "content": "84. Rodola, E., Bronstein, A.M., Albarelli, A., Bergamasco, F., Torsello, A.: A game-theoretic approach to deformable shape matching. In: 2012 IEEE Conference on Computer Vision and Pattern Recognition. pp. 182-189. IEEE (2012)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.326, + 0.619, + 0.338 + ], + "angle": 0, + "content": "85. Shapley, L.S., et al.: A value for n-person games (1953)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.339, + 0.785, + 0.378 + ], + "angle": 0, + "content": "86. Sigurdsson, G.A., Varol, G., Wang, X., Farhadi, A., Laptev, I., Gupta, A.: Hollywood in homes: Crowdsourcing data collection for activity understanding. In: European Conference on Computer Vision. pp. 510-526 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.379, + 0.785, + 0.419 + ], + "angle": 0, + "content": "87. Song, Y., Wang, J., Ma, L., Yu, J., Liang, J., Yuan, L., Yu, Z.: Marn: Multilevel attentional reconstruction networks for weakly supervised video temporal grounding. Neurocomputing 554, 126625 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.42, + 0.785, + 0.46 + ], + "angle": 0, + "content": "88. Song, Y., Wang, J., Ma, L., Yu, Z., Yu, J.: Weakly-supervised multi-level attentional reconstruction network for grounding textual queries in videos. arXiv preprint arXiv:2003.07048 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.461, + 0.785, + 0.514 + ], + "angle": 0, + "content": "89. Tan, R., Xu, H., Saenko, K., Plummer, B.A.: Logan: Latent graph co-attention network for weakly-supervised video moment retrieval. In: Proceedings of the IEEE Winter Conference on Applications of Computer Vision. pp. 2083–2092 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.515, + 0.785, + 0.554 + ], + "angle": 0, + "content": "90. Tang, K., Chen, Y., Peng, W., Zhang, Y., Fang, M., Wang, Z., Song, P.: Reppvconv: attentively fusing reparameterized voxel features for efficient 3d point cloud perception. The Visual Computer 39(11), 5577-5588 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.555, + 0.785, + 0.607 + ], + "angle": 0, + "content": "91. Tang, K., Lou, T., Peng, W., Chen, N., Shi, Y., Wang, W.: Effective single-step adversarial training with energy-based models. IEEE Transactions on Emerging Topics in Computational Intelligence (2024). https://doi.org/10.1109/TETCI.2024.3378652" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.608, + 0.785, + 0.65 + ], + "angle": 0, + "content": "92. Tang, K., Ma, Y., Miao, D., Song, P., Gu, Z., Tian, Z., Wang, W.: Decision fusion networks for image classification. IEEE Transactions on Neural Networks and Learning Systems (2022). https://doi.org/10.1109/TNNLS.2022.3196129" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.65, + 0.785, + 0.69 + ], + "angle": 0, + "content": "93. Tang, K., Shi, Y., Lou, T., Peng, W., He, X., Zhu, P., Gu, Z., Tian, Z.: Rethinking perturbation directions for imperceptible adversarial attacks on point clouds. IEEE Internet of Things Journal 10(6), 5158-5169 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.691, + 0.785, + 0.745 + ], + "angle": 0, + "content": "94. Tang, K., Zhao, W., Peng, W., Fang, X., Cui, X., Zhu, P., Tian, Z.: Reparameterization head for efficient multi-input networks. In: ICASSP 2024 - 2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). pp. 6190-6194 (2024). https://doi.org/10.1109/ICASSP48485.2024.10447574" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.746, + 0.785, + 0.799 + ], + "angle": 0, + "content": "95. Tang, K., Zhao, W., Peng, W., Fang, X., Cui, X., Zhu, P., Tian, Z.: Reparameterization head for efficient multi-input networks. In: ICASSP 2024-2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). pp. 6190-6194. IEEE (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.8, + 0.785, + 0.84 + ], + "angle": 0, + "content": "96. Torsello, A., Bulo, S.R., Pelillo, M.: Grouping with asymmetric affinities: A game-theoretic perspective. In: 2006 IEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR'06). vol. 1, pp. 292-299. IEEE (2006)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.148, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.412, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Rethinking WS-VTG From a Game Perspective" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.117, + 0.784, + 0.127 + ], + "angle": 0, + "content": "21" + }, + { + "type": "ref_text", + "bbox": [ + 0.223, + 0.147, + 0.785, + 0.189 + ], + "angle": 0, + "content": "97. Tran, D., Bourdev, L., Fergus, R., Torresani, L., Paluri, M.: Learning spatiotemporal features with 3d convolutional networks. In: Proceedings of the IEEE International Conference on Computer Vision. pp. 4489-4497 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.191, + 0.785, + 0.233 + ], + "angle": 0, + "content": "98. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. In: Advances in Neural Information Processing Systems. pp. 5998-6008 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.234, + 0.785, + 0.274 + ], + "angle": 0, + "content": "99. Wang, C., Ning, X., Li, W., Bai, X., Gao, X.: 3d person re-identification based on global semantic guidance and local feature aggregation. IEEE Transactions on Circuits and Systems for Video Technology (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.276, + 0.785, + 0.317 + ], + "angle": 0, + "content": "100. Wang, C., Ning, X., Sun, L., Zhang, L., Li, W., Bai, X.: Learning discriminative features by covering local geometric space for point cloud analysis. IEEE Transactions on Geoscience and Remote Sensing 60, 1-15 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.318, + 0.785, + 0.346 + ], + "angle": 0, + "content": "101. Wang, C., Wang, C., Li, W., Wang, H.: A brief survey on rgb-d semantic segmentation using deep learning. Displays 70, 102080 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.347, + 0.785, + 0.388 + ], + "angle": 0, + "content": "102. Wang, C., Wang, H., Ning, X., Shengwei, T., Li, W.: 3d point cloud classification method based on dynamic coverage of local area. Journal of Software 34(4), 1962-1976 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.389, + 0.785, + 0.431 + ], + "angle": 0, + "content": "103. Wang, J., Ma, L., Jiang, W.: Temporally grounding language queries in videos by contextual boundary-aware prediction. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 34, pp. 12168-12175 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.432, + 0.785, + 0.472 + ], + "angle": 0, + "content": "104. Wang, Y., Deng, J., Zhou, W., Li, H.: Weakly supervised temporal adjacent network for language grounding. IEEE Transactions on Multimedia 24, 3276-3286 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.474, + 0.785, + 0.515 + ], + "angle": 0, + "content": "105. Wang, Z., Chen, J., Jiang, Y.G.: Visual co-occurrence alignment learning for weakly-supervised video moment retrieval. In: Proceedings of the 29th ACM International Conference on Multimedia. pp. 1459–1468 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.517, + 0.785, + 0.572 + ], + "angle": 0, + "content": "106. Wen, J., Liu, C., Deng, S., Liu, Y., Fei, L., Yan, K., Xu, Y.: Deep double incomplete multi-view multi-label learning with incomplete labels and missing views. IEEE Transactions on Neural Networks and Learning Systems pp. 1-13 (2023). https://doi.org/10.1109/TNNLS.2023.3260349" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.573, + 0.785, + 0.614 + ], + "angle": 0, + "content": "107. Wen, J., Zhang, Z., Li, Z.J.: A survey on incomplete multiview clustering. IEEE transactions on systems, man, and cybernetics. Systems 53(2 Pt.2), 1136-1149 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.615, + 0.785, + 0.643 + ], + "angle": 0, + "content": "108. Winter, E.: The shapley value. Handbook of game theory with economic applications 3, 2025-2054 (2002)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.644, + 0.785, + 0.699 + ], + "angle": 0, + "content": "109. Wu, H., Lyu, Y., Shen, X., Zhao, X., Wang, M., Zhang, X., Luo, Z.: Atomic-action-based contrastive network for weakly supervised temporal language grounding. In: 2023 IEEE International Conference on Multimedia and Expo (ICME). pp. 1523-1528. IEEE (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.7, + 0.785, + 0.741 + ], + "angle": 0, + "content": "110. Xiong, Z., Liu, D., Zhou, P.: Gaussian kernel-based cross modal network for spatio-temporal video grounding. In: IEEE International Conference on Image Processing (ICIP). pp. 2481-2485 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.742, + 0.785, + 0.798 + ], + "angle": 0, + "content": "111. Xiong, Z., Liu, D., Zhou, P., Zhu, J.: Tracking objects and activities with attention for temporal sentence grounding. In: ICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). pp. 1-5. IEEE (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.799, + 0.785, + 0.84 + ], + "angle": 0, + "content": "112. Yang, W., Zhang, T., Zhang, Y., Wu, F.: Local correspondence network for weakly supervised temporal sentence grounding. IEEE Transactions on Image Processing 30, 3252-3262 (2021)" + }, + { + "type": "list", + "bbox": [ + 0.22, + 0.147, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "22" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.498, + 0.129 + ], + "angle": 0, + "content": "X. Fang, Z. Xiong, W. Fang et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.785, + 0.189 + ], + "angle": 0, + "content": "113. Yu, Z., Li, L., Xie, J., Wang, C., Li, W., Ning, X.: Pedestrian 3d shape understanding for person re-identification via multi-view learning. IEEE Transactions on Circuits and Systems for Video Technology (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.189, + 0.787, + 0.243 + ], + "angle": 0, + "content": "114. Yuan, Y., Ma, L., Wang, J., Liu, W., Zhu, W.: Semantic conditioned dynamic modulation for temporal sentence grounding in videos. In: Proceedings of the 33rd International Conference on Neural Information Processing Systems. pp. 536-546 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.243, + 0.785, + 0.285 + ], + "angle": 0, + "content": "115. Zeng, R., Xu, H., Huang, W., Chen, P., Tan, M., Gan, C.: Dense regression network for video grounding. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 10287-10296 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.285, + 0.785, + 0.34 + ], + "angle": 0, + "content": "116. Zhang, D., Dai, X., Wang, X., Wang, Y.F., Davis, L.S.: Man: Moment alignment network for natural language moment retrieval via iterative graph adjustment. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 1247-1257 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.34, + 0.787, + 0.381 + ], + "angle": 0, + "content": "117. Zhang, H., Sun, A., Jing, W., Zhou, J.T.: Span-based localizing network for natural language video localization. In: Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics. pp. 6543-6554 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.381, + 0.785, + 0.422 + ], + "angle": 0, + "content": "118. Zhang, H., Xie, Y., Zheng, L., Zhang, D., Zhang, Q.: Interpreting multivariate shapley interactions in dnns. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 35, pp. 10877-10886 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.422, + 0.785, + 0.462 + ], + "angle": 0, + "content": "119. Zhang, H., Wang, C., Tian, S., Lu, B., Zhang, L., Ning, X., Bai, X.: Deep learning-based 3d point cloud classification: A systematic survey and outlook. Displays 79, 102456 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.462, + 0.785, + 0.502 + ], + "angle": 0, + "content": "120. Zhang, H., Wang, C., Yu, L., Tian, S., Ning, X., Rodrigues, J.: Pointgt: A method for point-cloud classification and segmentation based on local geometric transformation. IEEE Transactions on Multimedia (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.502, + 0.785, + 0.544 + ], + "angle": 0, + "content": "121. Zhang, S., Peng, H., Fu, J., Luo, J.: Learning 2d temporal adjacent networks for moment localization with natural language. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 34, pp. 12870-12877 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.544, + 0.785, + 0.599 + ], + "angle": 0, + "content": "122. Zhang, Z., Lin, Z., Zhao, Z., Xiao, Z.: Cross-modal interaction networks for query-based moment retrieval in videos. In: Proceedings of the 42nd International ACM SIGIR Conference on Research and Development in Information Retrieval. pp. 655-664 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.599, + 0.785, + 0.639 + ], + "angle": 0, + "content": "123. Zhang, Z., Zhao, Z., Lin, Z., He, X., et al.: Counterfactual contrastive learning for weakly-supervised vision-language grounding. Advances in Neural Information Processing Systems 33, 18123-18134 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.639, + 0.785, + 0.68 + ], + "angle": 0, + "content": "124. Zheng, M., Huang, Y., Chen, Q., Liu, Y.: Weakly supervised video moment localization with contrastive negative sample mining. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 36, pp. 3517-3525 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.68, + 0.785, + 0.735 + ], + "angle": 0, + "content": "125. Zheng, M., Huang, Y., Chen, Q., Peng, Y., Liu, Y.: Weakly supervised temporal sentence grounding with gaussian-based contrastive proposal learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15555-15564 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.735, + 0.785, + 0.789 + ], + "angle": 0, + "content": "126. Zheng, Q., Dong, J., Qu, X., Yang, X., Wang, Y., Zhou, P., Liu, B., Wang, X.: Progressive localization networks for language-based moment localization. ACM Transactions on Multimedia Computing, Communications and Applications 19(2), 1-21 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.789, + 0.785, + 0.831 + ], + "angle": 0, + "content": "127. Zhu, J., Liu, D., Zhou, P., Di, X., Cheng, Y., Yang, S., Xu, W., Xu, Z., Wan, Y., Sun, L., et al.: Rethinking the video sampling and reasoning strategies for temporal sentence grounding. arXiv preprint arXiv:2301.00514 (2023)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.831 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/Rethinking Weakly-supervised Video Temporal Grounding From a Game Perspective/4f2540b2-441f-49e9-a4f9-06c96b70a305_origin.pdf b/2024/Rethinking Weakly-supervised Video Temporal Grounding From a Game Perspective/4f2540b2-441f-49e9-a4f9-06c96b70a305_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..4791872aca4c11875e9243b6d25c552fd577184f --- /dev/null +++ b/2024/Rethinking Weakly-supervised Video Temporal Grounding From a Game Perspective/4f2540b2-441f-49e9-a4f9-06c96b70a305_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bc693d8a5dc9d2c84128c6776ac857fd9b256230109ac2e0d3f0fafda59dbd61 +size 1204258 diff --git a/2024/Rethinking Weakly-supervised Video Temporal Grounding From a Game Perspective/full.md b/2024/Rethinking Weakly-supervised Video Temporal Grounding From a Game Perspective/full.md new file mode 100644 index 0000000000000000000000000000000000000000..5a8801e19d9963be2b829a721b4e264dbcf8f761 --- /dev/null +++ b/2024/Rethinking Weakly-supervised Video Temporal Grounding From a Game Perspective/full.md @@ -0,0 +1,395 @@ +# Rethinking Weakly-supervised Video Temporal Grounding From a Game Perspective + +Xiang Fang $^{1*}$ , Zeyu Xiong $^{1*}$ , Wanlong Fang $^{1*}$ , Xiaoye Qu $^{1}$ , Chen Chen $^{2}$ , Jianfeng Dong $^{3}$ , Keke Tang $^{4}$ , Pan Zhou $^{1\boxtimes}$ , Yu Cheng $^{5}$ , and Daizong Liu $^{6\boxtimes}$ + +1 Hubei Key Laboratory of Distributed System Security, Hubei Engineering Research + +Center on Big Data Security, School of Cyber Science and Engineering, Huazhong + +University of Science and Technology + +2 University of Central Florida + +3 Zhejiang Gongshang University + +4 Guangzhou University + +5 The Chinese University of Hong Kong + +6 Peking University + +xfang9508@gmail.com, zeyuxiong@hust.edu.cn, wanlongfang@gmail.com, + +xiaoye@hust.edu.cn, chen.chen@crcv.ucf.edu, dongjf24@gmail.com, + +tangbohutbh@gmail.com, panzhou@hust.edu.cn, chengyu@cse.cuhk.edu.hk, + +dzliu@stu.pku.edu.cn + +Abstract. This paper addresses the challenging task of weakly-supervised video temporal grounding. Existing approaches are generally based on the moment proposal selection framework that utilizes contrastive learning and reconstruction paradigm for scoring the pre-defined moment proposals. Although they have achieved significant progress, we argue that their current frameworks have overlooked two indispensable issues: 1) Coarse-grained cross-modal learning: previous methods solely capture the global video-level alignment with the query, failing to model the detailed consistency between video frames and query words for accurately grounding the moment boundaries. 2) Complex moment proposals: their performance severely relies on the quality of proposals, which are also time-consuming and complicated for selection. To this end, in this paper, we make the first attempt to tackle this task from a novel game perspective, which effectively learns the uncertain relationship between each vision-language pair with diverse granularity and flexible combination for multi-level cross-modal interaction. Specifically, we creatively model each video frame and query word as game players with multivariate cooperative game theory to learn their contribution to the cross-modal similarity score. By quantifying the trend of frame-word cooperation within a coalition via the game-theoretic interaction, we are able to value all uncertain but possible correspondence between frames and words. Finally, instead of using moment proposals, we utilize the learned query-guided frame-wise scores for better moment localization. Experiments show that our method achieves superior performance on both Charades-STA and ActivityNet Caption datasets. + +![](images/1f60eb03a3c257228783a23a232c2c9f38ea109e0f0ad74cc5278ab11062fd33.jpg) +Fig. 1: (a) Most existing WS-VTG methods define moment proposals for query matching. They not only suffer from the coarse-grained alignment but also are limited to the quality of the proposals. (b) Instead of using proposals, our method reviews the WS-VTG from a brand-new game-interaction perspective to learn more fine-grained frame-aware self- and cross-modal alignment for accurate boundary localization. + +![](images/1efc26c781866c4d7d0c1625873b46e5ad008da2e5befe11042c33ba90443ae0.jpg) + +# 1 Introduction + +Video temporal grounding (VTG) aims to localize the start and end timestamps of a specific event moment described by a given language query in an untrimmed video [2,32,39,75-77,99-102,113,119,120]. This task enables us to efficiently find video moments of human interest instead of traversing the entire video, which has broad application potential in video surveillance [13-15, 18, 24-26, 50-52, 59, 90-93, 125, 126], video summarization [12, 16, 17, 21-23, 36, 48, 94, 106, 107], etc. Most VTG methods [27-31, 53-55, 57, 58, 60-62, 65, 66, 127] follow the fully-supervised setting, where each frame is manually annotated as query-relevant or not. Although these methods achieve significant breakthroughs, they severely rely on extensive laborious manual annotations of moment boundaries, thus limiting their scalability and practicability in real-world applications. To alleviate the dense reliance, weakly-supervised VTG (WS-VTG) has received increasing attention [10, 49, 73, 88, 89, 123], which only requires the knowledge of matched video-query pairs without detailed frame-wise annotations, which is more challenging than the fully-supervised setting. + +Since there is no frame-wise annotation in WS-VTG, learning detailed frame-query alignment is difficult. As shown in Fig. 1(a), almost all existing WS-VTG solutions directly employ the proposal-based framework [10,33,73,89] that first generates multiple moment proposals and then learns the scores to indicate the potential alignment between video proposals and language query. Some approaches further utilize reconstruction-based paradigm [49,88,123] to minimize the reconstruction loss between the partially masked query and moment proposals to identify the proposal which best reconstructs the query. Although existing WS-VTG methods have achieved great progress, we argue that they still suffer from two inescapable limitations: 1) Firstly, as frame-wise annotation is not provided, these WS-VTG methods generally generate the moment proposals for all videos via sliding windows. In this way, they only model the coarse-grained cross-modal interaction of each proposal-query pair. Actually, in most cases, we expect to capture more fine-grained information for accurate boundary localization, such as how much the semantics of a specific frame helps or harms the query-guided cross-modal alignment. Unfortunately, existing methods rely on proposal-level cross-modal learning cannot achieve this goal. 2) Secondly, the + +performance of these proposal-based methods is severely limited by the quality of moment proposals as they devise moment proposals regardless of the specific contents and difficulty of each video. + +To tackle the above limitations, in this paper, we propose to address the WS-VTG task from a new game-theory perspective, which creatively models the subtle video frames and query words as game players to learn their uncertain but possible cross-modal interaction with diverse granularity and flexible combination. The game learning helps to generate query-guided frame-aware knowledge for fine-grained video grounding. As shown in Fig. 1(b), if visual representations and textual representations have strong semantic correspondence, they tend to cooperate together and contribute to a high cross-modal similarity score. By forming these cooperated frame-word representations as a coalition and quantifying the trend of cooperation within a coalition via the game-theoretic interaction index, we can not only learn the coalition contribution to the cross-modal semantic similarity, but also measure the additional benefits brought by the coalition compared with the costs of the lost coalitions of these players with others. In this manner, our cooperative WS-VTG game is able to value all possible correspondence between frames and words for sensitive and explainable cross-modal contrast. By formulating these detailed cross-modal correspondences as fine-grained query-guided frame-wise scores, we can efficiently generate the moment instead of using complicated proposals. In particular, we implement the cooperative game in both self-modal and cross-modal scenarios for self-modal attentive learning and cross-modal fine-grained alignment. We further extend the cross-modal game with multi-level query-semantic interaction for more comprehensive video understanding. + +Our main contributions are summarized as follows: + +- We reveal the limitations of existing WS-VTG methods, which rely on complex moment proposals and fail to capture fine-grained frame-word alignment for accurate boundary grounding. To this end, we propose a novel game-theory based framework to learn the detailed frame-wise query-relevance scores for efficiently constructing the accurate moment. +- We apply the game learning to both self- and cross-modal scenarios for contextual self-enhancement and cross-alignment. The cross-modal game is also extended to multi-level for comprehensive learning. +- To verify the effectiveness of our framework, we conduct experiments on two challenging WS-VTG benchmarks, i.e., Charades-STA and ActivityNet Caption. Experiments show that our method achieves superior performance. In representative cases, our method outperforms all compared methods by $4.53\%$ on Charades-STA. + +# 2 Related Works + +Fully-supervised video temporal grounding. Existing methods generally address the video temporal grounding in a fully-supervised manner, where both the annotations of video-sentence pairs and corresponding moment boundaries + +are given. Most of them [32, 35, 37, 56, 61, 64, 66, 95, 103, 110, 111, 114, 116, 121, 122] utilize the proposal-based framework, which first integrate the sentence representation with those pre-defined moment proposals individually, and then evaluate their matching relationships. The proposal with the highest matching score is selected as the prediction. Instead of utilizing the moment proposals, recent proposal-free methods [8, 9, 74, 115] directly regress the temporal locations of the target moment. + +Weakly-supervised video temporal grounding. As manually annotating temporal boundaries of target moments is time-consuming, recent research attentions shift to weakly-supervised video temporal grounding [10, 73, 88, 89], which only requires video-level annotations. Mithun et al. [73] proposed the first weakly-supervised model to learn a joint embedding space for video and query representations. Gao et al. [33] developed a two-stream structure to measure the moment-query consistency and conduct moment selection simultaneously. Although the above methods have achieved promising performance, they are two-stage approaches that utilize multi-scale sliding windows to generate moment candidates, therefore suffering from inferior effectiveness and efficiency. To address this issue, [49, 68, 123] score all the moments sampled at different scales in a single pass and further improve the moment-sentence matching accuracy. However, almost all of them rely on moment proposals for matching and selection, which fail to capture and distinguish more fine-grained details among visually similar frames for acquiring more accurate moment boundaries. + +Cooperative game theory. Cooperative game theory focuses on the formation of coalitions and cooperation among rational players or groups to achieve common goals [1,6,40,41,47,69,80-82,96]. A typical cooperative game consists of a set of players with a game function. The game function maps all possible subsets of players, called group or coalition, to a number which represents the total payoff earned by these players working cooperatively to achieve the goal [19, 20, 46, 84]. To measure the contributions of each player and allocate different payoffs to these individuals fairly, a few researchers have proposed various concepts of value by computing the average added worth that one player brings to all possible coalitions, such as Shapley value [72, 85, 108] and Banzhaf value [4, 45, 78]. Recently, some concepts of cooperative game theory have been interpreted in the field of deep learning [11, 67, 118]. For example, Li et al. [47] propose a fine-grained image-text semantic alignment pre-training framework based on Shapley interaction. Different from these works, we propose for the first time to model the detailed consistency between video frames and query words by recasting multivariate game-theoretic indices to address more complicated WS-VTG task. + +# 3 Methodology + +In this section, we elaborate on the proposed method, which learns the uncertain and fine-grained cross-modal alignment in WS-VTG task. Specifically, our method conceptualizes the interaction between the video and query as a game, + +![](images/b26a4d14c5506599fbbb0c8864015dd78a5b7e1da1b8ceed3957a70cf47c5568.jpg) +Fig. 2: Overview of our framework, where "SQAN" denotes Sequential Query Attention Network. Specifically, we first conduct self-modal game to enhance the self-modal instance semantics, then devise multi-level cross-modal games to learn the fine-grained and uncertain cross-modal correspondence. At last, we predict the moment based on the query-guided frame-wise scores. + +where each frame/word acts as a player and forms coalitions. During the game learning, visual frame and linguistic word representations with strong semantic relevance will cooperate to form a new coalition, the benefits of which are quantified through the game interaction index. Our method maximizes payoff by strategically forming coalitions to achieve fine-grained semantic alignment between video and query. The overall framework is shown in Fig. 2. + +# 3.1 Preliminaries + +Problem formulation. Generally, given an untrimmed video and a sentence query, WS-VTG aims to localize the specific video moment semantically related to the query with only the video-level annotation. Let $V = \{v_{1},\dots,v_{i}\dots,v_{T}\}$ denote the untrimmed video with $T$ frames, and $Q = \{q_{1},\dots,q_{j},\dots,q_{M}\}$ represent the sentence query containing $M$ words, where $v_{i}$ and $q_{j}$ are the $i$ -th frame and $j$ -th word, respectively. The objective function $\mathcal{F}$ can be formulated as $\mathcal{F}: (V,Q) \to (\tau_s,\tau_e)$ , where $\tau_s,\tau_e$ denote the start and end timestamps of the video moment boundary in video $V$ semantically corresponding to query $Q$ . + +Video and query representations. To extract contextual video representation, following [32, 117], each input video $V$ is fed into a pre-trained 3D convolutional network [5, 97] with a stack of multi-head self-attention [98] layers for video encoding, then we get frame-wise video representation $\mathbf{V} = \{\pmb{v}_i\}_{i=1}^T \in \mathbb{R}^{T \times D_h}$ , where $D_h$ is the hidden dimension. For query representation, we also follow [59, 117] to embed each word of the query by the Glove [83] model with a further multi-head self-attention module and a BiLSTM layer. Similarly, we can obtain the word-wise query representation $\mathbf{Q} = \{\pmb{q}_j\}_{j=1}^M \in \mathbb{R}^{M \times D_h}$ . + +# 3.2 Game-theoretic Learning for Self-modal Semantic Enhancement + +We argue that the instance relations within each modality are important to infer the consecutive visual event contents or linguistic phrase contexts. To achieve this self-modal learning, we review previous classic game-theoretic value [34, 85] to explore the weighted semantic contribution of each frame/word to the whole video/query semantics for enhancing the self-modal representation of both video and query. + +Visual self-modal game. Generally, a cooperative game consists of a player set $\mathcal{P} = \{1,2,\dots,n\}$ and a game function $g(\cdot)$ . In detail, function $g(\cdot)$ maps each player subset $S \subseteq \mathcal{P}$ to a score value, which indicates the payoff when all players in coalition $S$ work together in the game. To capture the inherent correlations within the video modality, we directly take the video frames $\mathbf{V} = \{\pmb{v}_i\}_{i=1}^T$ as frame players $\mathcal{P} = \mathbf{V}, n = T$ to play the game and evaluate frame-wise contribution to frame-level video semantics. + +Since different frames contribute and weight differently to video understanding, we attempt to adopt a game-theoretic value to measure the importance of each frame player in the cooperative game. Specifically, each frame player $\pmb{v}_i$ in player set $\pmb{V}$ has a weighted marginal contribution to the entire game, which can be measured by the game-theoretic value (e.g., Banzhaf value [34]) as: + +$$ +B \left(\boldsymbol {v} _ {i} \mid \boldsymbol {V}\right) = \frac {1}{2 ^ {T - 1}} \sum_ {\mathcal {S} \subseteq \boldsymbol {V} \backslash \boldsymbol {v} _ {i}} [ g (\mathcal {S} \cup \{\boldsymbol {v} _ {i} \}) - g (\mathcal {S}) ], \tag {1} +$$ + +where $S \subseteq V \backslash \mathbf{v}_i$ represents a group formed without $\mathbf{v}_i$ , $\frac{1}{2^{T-1}}$ is the likelihood of $S$ being sampled, and $g(S \cup \{\mathbf{v}_i\}) - g(S)$ denotes the marginal contribution brought by player $\mathbf{v}_i$ joining the coalition $S$ . $B(\mathbf{v}_i | V)$ indicates the ability of player $\mathbf{v}_i$ to influence the video-level semantic by calculating the weighted marginal contribution of player $\mathbf{v}_i$ to all possible coalitions $S$ . Generally, within the context of our visual game, coalitions in the video modality represent consecutive frames of the same event or frames that are semantically similar. In this way, we can enumerate and learn all possible relations between the frames within the video for more contextual visual feature learning. + +Visual soft supervision. To assist the above game-theoretic value learning, we first develop an additional learnable video-domain relevance header to predict the same self-modal relevance $\hat{R}^v = \{\hat{r}_i^v\}_{i=1}^T = \text{header}(\mathbf{V})$ as Banzhaf values $\{B(\mathbf{v}_i|\mathbf{V})\}_{i=1}^T$ among all frames for soft supervision (more discussions will be illustrated later). Specifically, the header is implemented by: 1) 1D convolutional layers for local self-consistency modeling, 2) a self-attention module for global self-relevance capturing, 3) a convolutional layer for decoding. Then, we optimize the Kullback-Leibler Divergence (KLD) between the predicted relevance $\hat{r}_i^v$ and the Banzhaf value $B(\mathbf{v}_i|\mathbf{V})$ for learning the consistent self-modal similarity. In detail, we define the game-guided frame-wise probability distribution as $D^v = [p_1^v,p_2^v,\dots,p_i^v,\dots,p_T^v]^\top$ , where $p_i^v = \frac{\exp(B(\mathbf{v}_i|\mathbf{V}))}{\sum_{l=1}^T\exp(B(\mathbf{v}_l|\mathbf{V}))}$ . Similarly, the predicted frame-wise probability distribution calculated by the output of the self-relevance header can be denoted as $\hat{D}^v = [\hat{p}_1^v,\hat{p}_2^v,\dots,\hat{p}_i^v,\dots,\hat{p}_T^v]^\top$ , where + +$\hat{p}_i^v = \frac{\exp(\hat{r}_i^v)}{\sum_{l=1}^{T}\exp(\hat{r}_l^v)}$ . Finally, the self-relevance loss $\mathcal{L}_{self}^v$ for each video is: + +$$ +\mathcal {L} _ {s e l f} ^ {v} = \operatorname {K L D} \left(\hat {D} ^ {v} | | D ^ {v}\right) = - \frac {1}{T} \sum_ {i = 1} ^ {T} p _ {i} ^ {v} \left(\log \hat {p} _ {i} ^ {v} - \log p _ {i} ^ {v}\right). \tag {2} +$$ + +Based on the learned game-theoretic value $B(\pmb{v}_i|\pmb{V})$ for each player $\pmb{v}_i$ in player set $\pmb{V}$ , we deploy a Softmax function to obtain the self-modal enhanced video features $\widetilde{\pmb{v}}_i$ as: + +$$ +\widetilde {\boldsymbol {v}} _ {i} = \boldsymbol {v} _ {i} + \sum_ {\mathcal {S} \subseteq \mathcal {V} \backslash \boldsymbol {v} _ {i}} \operatorname {S o f t m a x} [ g (\mathcal {S} \cup \{\boldsymbol {v} _ {i} \}) - g (\mathcal {S}) ] \cdot (\mathcal {S} \cup \{\boldsymbol {v} _ {i} \}). \tag {3} +$$ + +The enhanced video features $\widetilde{\pmb{V}} = \{\widetilde{\pmb{v}}_i\}_{i=1}^T$ aggregates the relevant contexts from the whole video. + +Linguistic self-modal game. Similarly, we deem query words $\mathbf{Q} = \{\mathbf{q}_j\}_{j=1}^M$ as players $\mathcal{P} = \mathbf{Q}, n = M$ and calculate each word's marginal contribution $B(\mathbf{q}_j|\mathbf{Q})$ to the whole query $\mathbf{Q}$ by formulating the linguistic self-modal game as: + +$$ +B \left(\boldsymbol {q} _ {j} \mid \boldsymbol {Q}\right) = \frac {1}{2 ^ {M - 1}} \sum_ {\mathcal {S} \subseteq \boldsymbol {Q} \backslash \boldsymbol {q} _ {j}} [ g (\mathcal {S} \cup \{\boldsymbol {q} _ {j} \}) - g (\mathcal {S}) ]. \tag {4} +$$ + +Linguistic soft supervision. We first take the game probability distributions $D^{q} = [p_{1}^{q}, p_{2}^{q}, \dots, p_{j}^{q}, \dots, p_{M}^{q}]^{\top}$ as the soft label, and then supervise the predicted distributions $\hat{D}^{q} = [\hat{p}_{1}^{q}, \hat{p}_{2}^{q}, \dots, \hat{p}_{j}^{q}, \dots, \hat{p}_{M}^{q}]^{\top}$ of another header by KLD loss $\mathcal{L}_{self}^{q}$ similar to Eq. (2). At last, we can obtain the enhanced query features $\tilde{\boldsymbol{Q}} = \{\tilde{\boldsymbol{q}}_{j}\}_{j=1}^{M}$ by: + +$$ +\widetilde {\boldsymbol {q}} _ {j} = \boldsymbol {q} _ {j} + \sum_ {\mathcal {S} \subseteq \mathcal {Q} \backslash \boldsymbol {q} _ {j}} \operatorname {S o f t m a x} [ g (\mathcal {S} \cup \{\boldsymbol {q} _ {j} \}) - g (\mathcal {S}) ] \cdot (\mathcal {S} \cup \{\boldsymbol {q} _ {j} \}). \tag {5} +$$ + +Discussion on learnable header. Although we can directly utilize the game interaction value to enumerate the relations between different players via Eqs. (1) and (4), there are still two challenges: 1) No supervision signal: The performance of the game interaction severely depends on the quality of the learned player features. However, there is no supervision for learning the representative features. 2) Complex and time-consuming: The game process is complex and costs much time and resources. To this end, we develop a learnable header to mimic the game interaction by utilizing the KLD loss. This KLD function serves as a soft supervision label to not only learn the consistency between the game interaction and header, but also potentially train more distinguishing player features of each modality. During the inference, we can solely utilize the prediction of header to model the game interaction instead of Eqs. (1) and (4). + +# 3.3 Game-theoretic Interaction for Cross-modal Multi-level Alignment + +After obtaining the self-enhanced video and query features, we further develop a cross-modal game between frames and words to handle their uncertainty during their fine-grained semantic alignment. + +Cross-modal game. From a game-theoretic perspective, we attempt to conceptualize the problem of fine-grained cross-modal alignment as a collaborative effort between video frames and query words. Specifically, when a video frame and a query word exhibit a high degree of semantic similarity, they are more likely to collaborate and their union will have a larger score contributed to the final moment. To this end, we take the enhanced video features $\widetilde{V} = \{\widetilde{v}_i\}_{i=1}^T$ and enhanced query features $\widetilde{Q} = \{\widetilde{q}_j\}_{j=1}^M$ as two different kinds of players $\mathcal{P} = \{\widetilde{v}_i\}_{i=1}^T \cup \{\widetilde{q}_j\}_{j=1}^M$ , $n = T + M$ . In a cooperative game, different players tend to work together in groups, called coalitions, to achieve a common goal, i.e., representing the same semantics of the target event. In the case of two players $\widetilde{v}_i$ and $\widetilde{q}_j$ in set $\mathcal{P}$ , it may occur that $g(\{\widetilde{v}_i\})$ and $g(\{\widetilde{q}_j\})$ are small individually, but at the same time the reward of their forming coalition $g(\{\widetilde{v}_i, \widetilde{q}_j\})$ is considerable. This is because players in the coalition interact and collaborate with each other, which may bring additional benefits (or costs) to the game. Therefore, we also need to measure the additional benefits brought by the coalition compared with the costs of the lost coalitions of these players with others. To form a better coalition $\{\widetilde{v}_i, \widetilde{q}_j\} \subseteq \mathcal{P}$ , we utilize the effective interaction index $I_B(\{\widetilde{v}_i, \widetilde{q}_j\})$ like [34] to compute the additional benefit as follows: + +$$ +I _ {B} \left(\left\{\widetilde {\boldsymbol {v}} _ {i}, \widetilde {\boldsymbol {q}} _ {j} \right\}\right) = \frac {1}{2 ^ {n - 2}} \sum_ {\mathcal {S} \subseteq \mathcal {P} \backslash \left\{\widetilde {\boldsymbol {v}} _ {i}, \widetilde {\boldsymbol {q}} _ {j} \right\}} \left[ g \left(\mathcal {S} \cup \left\{\widetilde {\boldsymbol {v}} _ {i}, \widetilde {\boldsymbol {q}} _ {j} \right\}\right) + g (\mathcal {S}) - g \left(\mathcal {S} \cup \left\{\widetilde {\boldsymbol {v}} _ {i} \right\}\right) - g \left(\mathcal {S} \cup \left\{\widetilde {\boldsymbol {q}} _ {j} \right\}\right) \right], \tag {6} +$$ + +where $\frac{1}{2^{n-2}}$ is the likelihood of $\mathcal{S} \subseteq \mathcal{P} \backslash \{\widetilde{v}_i, \widetilde{q}_j\}$ being sampled. Intuitively, $I_B(\{\widetilde{v}_i, \widetilde{q}_j\})$ quantifies this additional benefit compared to when players $\widetilde{v}_i$ and $\widetilde{q}_j$ work independently. It also embodies the tendency of interaction between frame-player and word-player. A higher value of $I_B(\{\widetilde{v}_i, \widetilde{q}_j\})$ signifies that players $\widetilde{v}_i$ and $\widetilde{q}_j$ cooperate closely with each other, and the formed coalition will bring additional high returns. By this interaction index, we can measure the closeness between plays $\widetilde{v}_i$ and $\widetilde{q}_j$ to achieve fine-grained alignment between video and query. + +Game details. To ensure complete consistency between the cooperative game and cross-modal alignment learning, the game function $g$ should meet the following criteria: 1) the game payoff benefits from strongly corresponding semantic pairs $\{\pmb{v}_i^+, \pmb{q}_j^+\}$ , i.e., $g(\mathcal{P}) - g(\mathcal{P} \setminus \{\pmb{v}_i^+, \pmb{q}_j^+\} \cup \{[\{\pmb{v}_i^+, \pmb{q}_j^+\}]\}) < 0$ ; 2) the game payoff is compromised by semantically opposite pairs $\{\pmb{v}_i^-, \pmb{q}_j^-\}$ , i.e., $g(\mathcal{P}) - g(\mathcal{P} \setminus \{\pmb{v}_i^-, \pmb{q}_j^-\} \cup \{[\{\pmb{v}_i^-, \pmb{q}_j^-\}]\}) > 0$ ; 3) when there are no players to cooperate, the payoff is zero, i.e., $g(\{\pmb{v}_i\}_{i=1}^T) = g(\{\pmb{q}_j\}_{j=1}^M) = g(\phi) = 0$ , where $\phi$ is the empty set. It should be noted that any function that satisfies the aforementioned conditions can be employed as the game function $g(\cdot)$ . To simplify matters, we utilize the cosine similarity as $g$ in all games. + +During the game, when players $\mathbf{v}_i$ and $\mathbf{q}_j$ form a coalition $\{\mathbf{v}_i, \mathbf{q}_j\}$ , we deem $[\{\mathbf{v}_i, \mathbf{q}_j\}]$ as a singular hypothetical player, which is the union of the player in $\{\mathbf{v}_i, \mathbf{q}_j\}$ . Then, the reduced game is formed by removing the individual players in $\{\mathbf{v}_i, \mathbf{q}_j\}$ from the game and adding $[\{\mathbf{v}_i, \mathbf{q}_j\}]$ to the game. + +Word-level alignment with local game interaction. Similar to the self-modal game, we develop an alignment header to predict the word-frame alignment matrix $A = [a_{i,j}]^{T \times M}$ between frame $\widetilde{\boldsymbol{v}}_i$ and word $\widetilde{\boldsymbol{q}}_j$ , supervised by + +$I_B(\{\widetilde{v}_i,\widetilde{q}_j\})$ . Then, we optimize the Kullback-Leibler Divergence (KLD) between the $I_B(\{\widetilde{v}_i,\widetilde{q}_j\})$ and $a_{i,j}$ , to bring the probability distributions of the output of the alignment header and interaction index close together to establish word-level semantic alignment between frame players and word players. + +In detail, we define the word-to-frame probability distribution $d_{w2f}^{j}$ of the $j$ -th word-to-frame alignment as $d_{w2f}^{j} = [p_{1,j}, \dots, p_{t,j}]$ , where $p_{i,j} = \frac{\exp(I_B(\{\widetilde{v}_i, \widetilde{q}_j\}))}{\sum_{l=1}^T \exp(I_B(\{\widetilde{v}_l, \widetilde{q}_j\}))}$ and the total query-to-video probability distribution $D_{q2v} = [d_{w2f}^1, \dots, d_{w2f}^M]^\top$ . Similarly, the predicted fine-grained probability distribution can be represented as $\hat{d}_{w2f}^{j} = [\hat{p}_{1,j}, \dots, \hat{p}_{t,j}]$ , where $\hat{p}_{i,j} = \frac{\exp(a_{i,j})}{\sum_{l=1}^t \exp(a_{l,j})}$ , and the total predicted probability distribution is defined as $\hat{D}_{q2v} = [\hat{d}_{w2f}^1, \dots, \hat{d}_{w2f}^M]^\top$ . At last, the word-level alignment loss $\mathcal{L}_{word}$ is formulated as: + +$$ +\mathcal {L} _ {\text {w o r d}} = \operatorname {K L D} \left(\hat {D} _ {q 2 v} | | D _ {q 2 v}\right) = - \frac {1}{T M} \sum_ {i = 1} ^ {T} \sum_ {j = 1} ^ {M} p _ {i, j} \left(\log \hat {p} _ {i, j} - \log p _ {i, j}\right). \tag {7} +$$ + +Phrase- and sentence-level alignment with global game interaction. Although the above cross-modal game-theoretic interaction between each possible player-pair $\{\widetilde{v}_i,\widetilde{q}_j\}$ addresses the problem of word-level alignment between word player $\widetilde{v}_i$ and frame player $\widetilde{q}_j$ , it is still limited to aligning frames with isolated words, and fails to explore the relations between frames with more contextual phrase- or sentence-level semantics for better understanding temporal events. To this end, we introduce a multi-level game interaction to measure the semantic consistency between frames with different-grained textual semantics. Specifically, we follow [63, 74] to utilize Sequential Query Attention Network (SQAN) to extract the phrase-level textual semantics $\tilde{\mathbf{Q}}^p = \{\tilde{\mathbf{q}}_j^p\}_{j = 1}^P$ , and $P$ is the phrase number. The sentence-level semantics $\tilde{\mathbf{Q}}^s$ is obtained by concatenating the last hidden states in both the forward and backward BiLSTM. Similar to worl-level alignment, we compute both phrase- and sentence-level text-to-frame game interaction as $I_B(\{\widetilde{v}_i,\widetilde{q}_j^p\})$ , $I_B(\{\widetilde{v}_i,\tilde{Q}^s\})$ via Eq. (6). We design corresponding alignment headers (phrase-level cross-modal header and sentence-level cross-modal header) to predict their alignment matrix via $\mathcal{L}_{phra},\mathcal{L}_{sent}$ . + +# 3.4 Model Training and Inference + +Weak supervision for video-query matching. We first interact each frame with the semantics of words, phrases and the whole sentence to calculate the overall query-to-frame matching degree by: + +$$ +m _ {i} = \frac {1}{3} \left(I _ {B} \left(\{\widetilde {\boldsymbol {v}} _ {i}, \widetilde {\boldsymbol {Q}} ^ {s} \}\right) + \frac {1}{M} \sum_ {j = 1} ^ {M} I _ {B} \left(\{\widetilde {\boldsymbol {v}} _ {i}, \widetilde {\boldsymbol {q}} _ {j} \}\right) + \frac {1}{P} \sum_ {l = 1} ^ {P} I _ {B} \left(\left\{\widetilde {\boldsymbol {v}} _ {i}, \widetilde {\boldsymbol {q}} _ {l} ^ {p} \right\}\right)\right), \tag {8} +$$ + +where $m_{i}$ is the overall query-to-frame matching degree of $i$ -th frame. + +Then, the whole video-sentence matching score is calculated by: + +$$ +\phi^ {c} = \sum_ {i = 1} ^ {T} w _ {i} \cdot m _ {i}, w _ {i} = \frac {\exp \left(I _ {B} \left(\left\{\widetilde {\boldsymbol {v}} _ {i} , \widetilde {\boldsymbol {Q}} ^ {s} \right\}\right)\right)}{\sum_ {i = 1} ^ {T} \exp \left(I _ {B} \left(\left\{\widetilde {\boldsymbol {v}} _ {i} , \widetilde {\boldsymbol {Q}} ^ {s} \right\}\right)\right)}, \tag {9} +$$ + +where we utilize the sentence-level similarity to represent the contribution of each frame to the target moment among the entire video. For matched video $V_{k}$ and query $Q_{k}$ , we calculate the video-sentence matching score $\phi_{V_k,Q_k}^c$ . The scores of unmatched pairs are expressed as $\phi_{V_k,Q_l}^c$ and $\phi_{Q_k,V_l}^c$ where $l\neq k$ . + +Finally, we utilize the cross-modal contrastive loss [79] for weak supervision: + +$$ +\mathcal {L} _ {C} = - \frac {1}{N _ {b}} \sum_ {k = 1} ^ {N _ {b}} \log \frac {\exp \left(\phi_ {V _ {k} , Q _ {k}} ^ {c} / \tau\right)}{\sum_ {l = 1} ^ {N _ {b}} \exp \left(\phi_ {V _ {k} , Q _ {l}} ^ {c} / \tau\right)} - \frac {1}{N _ {b}} \sum_ {k = 1} ^ {N _ {b}} \log \frac {\exp \left(\phi_ {Q _ {k} , V _ {k}} ^ {c} / \tau\right)}{\sum_ {l = 1} ^ {N _ {b}} \exp \left(\phi_ {Q _ {k} , V _ {l}} ^ {c} / \tau\right)}, \tag {10} +$$ + +where $\tau$ is the temperature parameter, $N_{b}$ is the batch size. + +Reconstruction loss. To further enhance the semantic relevance of the highlighted frames to the sentence, we reconstruct the query conditioned on highlighted frames. Following popular reconstruction strategy [49,124], we randomly replace the 1/3 words in the original query with a specific symbol, and predict the next word by given a prefix of the query and highlighted frames feature. Specifically, we embed the masked query $\mathbf{Q}^m$ by Glove and predict the next word $\hat{\pmb{q}}_{i+1}$ using the conditioned transformer following CNM [124] with input $(\tilde{\mathbf{V}}, C_m, Q_{1:i}^m)$ , where $C_m = \{m_1, m_2, \dots, m_T\}$ represents the highlighted frame mask, $Q_{1:i}^m$ denotes the masked query words from the 1-st to $i$ -th. + +Then, we introduce the cross-entropy loss $\mathcal{L}_{rec}$ to calculate the probability distribution $\mathcal{P}^q$ different between the predict word $\hat{\pmb{q}}_{i + 1}$ and the real word as: + +$$ +\mathcal {L} _ {r e c} = - \sum_ {i = 1} ^ {M - 1} \log \mathcal {P} ^ {q} \left(\boldsymbol {q} _ {i + 1} \mid \hat {\boldsymbol {q}} _ {i + 1}\right). \tag {11} +$$ + +Overall training. To train the whole model, we combine game-guided self-relevance losses $\mathcal{L}_{self} = \mathcal{L}_{self}^v +\mathcal{L}_{self}^q$ , game alignment loss $\mathcal{L}_{align} = \mathcal{L}_{word} + \mathcal{L}_{phra} + \mathcal{L}_{sent}$ , video-level contrastive loss $\mathcal{L}_C$ , and reconstruction loss $\mathcal{L}_{rec}$ with trade-off hyper-parameters $(\alpha_{1},\alpha_{2},\alpha_{3}$ and $\alpha_4)$ to define the overall loss as: + +$$ +\mathcal {L} = \alpha_ {1} \mathcal {L} _ {\text {s e l f}} + \alpha_ {2} \mathcal {L} _ {\text {a l i g n}} + \alpha_ {3} \mathcal {L} _ {C} + \alpha_ {4} \mathcal {L} _ {\text {r e c}}. \tag {12} +$$ + +Inference. During inference, we directly utilize the self-relevance headers to enhance the video and query features, and exploit the multi-level alignment headers to predict the overall query-to-frame matching degree. We locate the frame with the highest score as the basic predicted moment, and add the left/right frames into the moment if the ratio of their scores to the frame score of the closest moment boundary is higher than a threshold. We repeat this step until no frame can be added. We can obtain more moments by locating different initial frames with different scores. + +# 4 Experiments + +# 4.1 Datasets and Evaluation Metrics + +Charades-STA. The Charades-STA dataset [32] is built based on the Charades [86], which contains 6,672 videos of indoor activities and involves 16,128 query-video pairs. There are 12,408 pairs used for training and 3,720 used for testing. The average duration of each video is 29.76 seconds. + +ActivityNet Caption. The ActivityNet Caption dataset [43] contains 20,000 videos with 100,000 queries, where 37,421 query-video pairs are used for training and 34,536 are used for testing. On average, each video in ActivityNet Caption has 3.65 annotated moments and each annotated moment lasts for 36 seconds. For fair comparison, we follow [10, 73, 88, 89, 123, 124] to use val1 as valuation and report the test performance on val2. + +Evaluation metrics. Following previous works, we adopt the metrics “R@n, IoU=m” to evaluate our model, which presents the proportion of the top $n$ predicted moments with IoU larger than $m$ . + +# 4.2 Implementation Details + +For fair comparison, we follow previous works to apply C3D [97] to encode the videos on ActivityNet Caption and I3D [5] on Charades-STA. We uniformly downsample the length of video feature sequences to $T = 200$ for ActivityNet Caption and $T = 64$ for Charades-STA. As for query encoding, we also follow previous works to utilize set the length of word feature sequences to $M = 20$ , and utilize Glove [83] to embed each word. The dimension $D_h$ is set to 512. These trade-off hyper-parameters are set to $\alpha_1 = 1.0, \alpha_2 = 1.0, \alpha_3 = 1.0, \alpha_4 = 10$ . Since the calculation of the exact game-theoretic interaction is an NP-hard problem [71], we follow previous works to utilize sampling-based method [3,44] to obtain unbiased estimates for approximating it. We train the whole model for 100 epochs with batch size of 16 and early stopping strategy. Parameter optimization is performed by Adam [42] optimizer with learning rate $3 \times 10^{-4}$ , and linear decay of learning rate and gradient clipping of 1.0. The inference threshold is set to 0.8 in ActivityNet Caption and 0.9 in Charades-STA. + +# 4.3 Main Results + +Quantitative comparison. We compare our method with the state-of-the-art methods, including fully-supervised (FS) methods and weakly-supervised (WS) methods. Best results are in bold. We introduce two types of cooperative game indices, i.e., Banzhaf index [34] and Shapley index [85], to our model. As summarized in Tab. 1, both two different variants of our method surpass all existing methods on both Charades-STA and ActivityNet Captions. Particularly, on Charades-STA in terms of "R@5, IoU=0.5", Ours(Banzhaf) outperforms the best compared method CPL by $4.53\%$ . The main reason is that our game-theoretic-based approach can learn the local alignment between words and frames for accurately predicting the moment boundaries. Besides, the Shapley variant performs similar to the Banzhaf variant for this specific WS-VTG task, since they share the same game interaction formulation with different normalization. Overall, this table verifies the effectiveness of our game-based framework. + +Generalization on cross-dataset evaluation. At last, we explore the generalization of models by cross-dataset evaluation in Tab. 2. Concretely, we first train a model on one dataset and then evaluate its performance on the other dataset. We can find that: 1) For all WS-VTG methods, the performance of + +Table 1: Comparisons with state-of-the-art methods. "WS" and "FS" denotes weakly-supervised and fully-supervised methods, respectively. + +
TypeMethodCharades-STAActivityNet Caption
R@1, IoU=0.5R@1, IoU=0.7R@5, IoU=0.5R@5, IoU=0.7R@1, IoU=0.3R@1, IoU=0.5R@5, IoU=0.3R@5, IoU=0.5
FSCTRL [32]23.638.8958.9229.52-29.01-59.17
2DTAN [121]39.8123.2579.3351.1559.4544.5185.5377.13
DRN [115]53.0931.7589.0660.05-45.45-77.97
WSSCN [49]23.589.9771.8038.8747.2329.2271.4555.69
WSTAN [104]29.3512.2876.1341.5352.4530.0179.3863.42
ICVC [7]31.0216.5377.5341.9146.6229.5280.9266.61
MARN [87]33.8715.5473.9041.9448.5231.3775.9160.00
CRM [38]34.7616.37--55.2632.19--
CNM [124]35.1514.95--55.6833.33--
ACN [109]37.0215.26--57.6634.18--
VCA [105]38.1319.5778.7537.7550.4531.0071.7953.83
LCNet [112]39.1918.1780.5645.2448.4926.3382.5162.66
CPL [125]49.2422.3984.7152.3755.7331.3763.0543.13
DM2 [70]51.3923.72--54.7331.85--
Ours(Banzhaf)54.1726.4688.9355.7958.9435.6586.4170.26
Ours(Shapley)54.3126.2889.2455.6259.1735.8086.5970.03
+ +Table 2: Cross-dataset evaluation. $A\rightarrow C$ denotes training on ActivityNet Caption and evaluation on CharadesSTA, and vice versa. + +
MethodA→CC→A
R@1, IoU=0.7R@5, IoU=0.7R@1, IoU=0.5R@5, IoU=0.5
LCNet [112]12.0631.4710.6531.76
CPL [125]14.3235.8111.5929.88
Ours(Banzhaf)18.7442.5820.3742.96
Ours(Shapley)18.2643.1320.0441.42
+ +![](images/096937f3d5d36a18f2cabc2e31684e3bba05139f06634375f7d8d5d49b67561b.jpg) +Fig. 3: Qualitative results. The colors of green lines represent different degrees of confidence. + +$\mathrm{A} \rightarrow \mathrm{C}$ achieves better than $\mathrm{C} \rightarrow \mathrm{A}$ . This is because that ActivityNet Caption is larger and more complex than Charades-STA, thus the former is able to bring more generalized knowledge to the latter while the latter fails to provide general knowledge to the former. 2) Our two variants outperform state-of-the-art methods (e.g., LCNet) with clear margins on cross-dataset evaluation. We speculate it to: Previous works severely rely on the moment proposals, however, LCNet defines different sliding windows on different datasets for proposal generation. Therefore, its non-uniform proposal structure limits its generalization on cross-dataset evaluation. Instead, we learn multi-level frame-aware alignment and utilize frame-wise scores for predicting moment, which is more flexible and adaptable, thus improving the generalization-ability. + +Qualitative comparison. In Fig. 3, we provide a qualitative example on ActivityNet Caption. The colors of green lines represent different degrees of interaction confidence, demonstrating that the game interaction can well learn the uncertain and fine-grained frame-word correspondence in the WS setting. Besides, all three-level interaction (word-level, phrase-level and sentence-level) contributes + +Table 3: Comparison on speed per sample (s) and GPU memory cost (MB). + +
MetricStageLCNet [112]CPL [125]Ours(Banzhaf)Ours(Shapley)
SpeedTraining0.79s0.27s0.94s0.98s
Inference0.62s0.21s0.10s0.11s
GPUTraining8647MB4156MB5237MB5482MB
MemoryInference7462MB3428MB2058MB2165MB
+ +Table 4: Main ablation study. + +
Self-modal GameMulti-level Cross-modal GameBanzhaf-based GameShapley-based Game
Charades-STAActivityNet CaptionCharades-STAActivityNet Caption
Video Query GameWord- Phrase- Sentence- levelIoU=0.7IoU=0.5IoU=0.5R@1,R@5,R@1,R@5,R@1,R@5,R@1,R@5,
IoU=0.7IoU=0.5IoU=0.5IoU=0.5IoU=0.7IoU=0.5IoU=0.5IoU=0.5
××✓××19.3444.6729.8361.7219.2944.3130.1561.56
✓×✓××21.2547.8631.3963.9821.1047.7331.5463.67
✓✓✓××22.9750.6432.7265.8022.6950.3832.9765.65
✓✓✓✓×24.8153.2834.4668.0324.6453.0334.5967.82
✓✓✓✓✓26.4655.7935.6570.2626.2855.6235.8070.03
+ +differently to the grounding, and our final method is able to localize more accurate moment boundaries than state-of-the-art method CPL. + +Complexity comparison. In Tab. 3, our method costs more training time/memory as we need to enumerate coalition samples. Specifically, we sampled 5500 samples. We find that, the model instability decreases along with the increase of the sampling number, and when the sample size ranges from 4500 to 5500, the error with the true value of banzhaf gradually decreases from $6.9\%$ to $1.1\%$ . After exceeding 5500 samples, more samples will not affect the result. During inference, since we can directly utilize the learned game headers to generate game value instead of using complex game learning, our model is faster with lower memory. + +# 4.4 Ablation Study + +In this section, we perform in-depth ablation studies to evaluate the effect of each component. We implement our model with different game settings. + +Main ablation. As shown in Tab. 4, we conduct ablation studies regarding the components (i.e., self-modal games in two modalities, and multi-level cross-modal games). We observe the following findings: 1) Our baseline model (only contains word-level cross-modal game) is powerful and outperforms some WS methods, demonstrating that it is effective to capture the uncertain yet fine-grained alignment for better determining the boundaries. 2) Both video and query self-modal games bring significant improvement to the overall model, indicating that self-enhancement is crucial for context integration. 3) Each query level of cross-modal game further boost the performance, verifying that phrase-and sentence-level semantics help to better understand temporal events. Overall, all components contribute a lot to final performance, showing their effectiveness. + +Effect of game headers. The headers in both self- and cross-modal games are utilized to provide soft supervision for better game interaction and player-representation learning. To explore the impact of the structure of these headers, we compare several popular structures in Tab. 5. We find that the combination of + +Table 5: Ablation study on each component, where "LL" means "Linear Layer", "CL" means "Convolution Layer", "SA" means "Self-attention", "MSE" means "Mean Squared Error", "CE" means "Cross Entropy", "KLD" means "Kullback-Leibler Divergence"; "RL" means "Reconstruction Loss". + +
ComponentSettingBanzhaf-based GameShapley-based Game
Charades-STAActivityNet CaptionCharades-STAActivityNet Caption
R@1, IoU=0.7R@5, IoU=0.7R@1, IoU=0.5R@5, IoU=0.5R@1, IoU=0.7R@5, IoU=0.7R@1, IoU=0.5R@5, IoU=0.5
Game HeadersLL25.2853.9034.1568.0324.9753.6834.2567.94
CL25.5254.3134.5668.6425.2654.0334.8268.47
LL + SA25.9454.8335.2769.1825.7154.5935.5369.06
CL + SA26.4655.7935.6570.2626.2855.6235.8070.03
Soft SupervisionNone23.2151.6432.3666.4022.9951.4532.7166.18
MSE24.4253.1933.8567.5724.2452.8940.0667.33
CE24.7353.5834.1668.0424.5253.2634.3767.80
KLD26.4655.7935.6570.2626.2855.6235.8070.03
Grounding Lossesw/o RL24.3553.5734.1468.4224.0853.3134.4568.19
w/ RL26.4655.7935.6570.2626.2855.6235.8070.03
+ +convolution layer and self-attention can capture both local and global interaction, so it achieves the best performance and is beneficial for cross-modal multi-level interaction, illustrating the significance of our game headers. + +Necessity of soft supervision. To explore the effect of soft supervision, we conduct the ablation study to investigate the contribution of the KLD loss during the header learning in Tab. 5. The KLD loss achieves better performance than both MSE and CE, showing that it is more suitable for game headers to mimic game-theoretic interaction learning. Moreover, without soft supervision, the performance will drop a lot, showing the importance of the soft supervision. + +Analysis on reconstruction loss. Most previous WS-VTG works also additionally utilize the reconstruction paradigm to assist the model training. As shown in Tab. 5, we further conduct such ablation study to investigate the impact of reconstruction loss. It shows that the reconstruction loss can improve significantly the performance, demonstrating its effectiveness. + +# 5 Conclusion + +In this paper, we rethink the limitations of previous weakly-supervised video temporal grounding works, and creatively propose to utilize the game-theoretic interaction to learn the uncertain relationship between video-query pairs with diverse granularity. Specifically, we first introduce two self-modal games to correlate the frames/words to enhance their contextual semantics. Then, we propose a cross-modal game to value fine-grained correspondence between frames and words with multiple levels for better determining the query-related moment boundaries. Extensive experiments demonstrate the effectiveness of our game-based framework on two challenging datasets. + +# References + +1. Albarelli, A., Rodola, E., Torsello, A.: A game-theoretic approach to fine surface registration without initial motion estimation. In: 2010 IEEE computer society conference on computer vision and pattern recognition. pp. 430-437. IEEE (2010) +2. Anne Hendricks, L., Wang, O., Shechtman, E., Sivic, J., Darrell, T., Russell, B.: Localizing moments in video with natural language. In: Proceedings of the IEEE International Conference on Computer Vision. pp. 5803-5812 (2017) +3. Bachrach, Y., Markakis, E., Resnick, E., Procaccia, A.D., Rosenschein, J.S., Saberi, A.: Approximating power indices: theoretical and empirical analysis. Autonomous Agents and Multi-Agent Systems 20, 105-122 (2010) +4. Banzhaf III, J.F.: Weighted voting doesn't work: A mathematical analysis. Rutgers L. Rev. 19, 317 (1964) +5. Carreira, J., Zisserman, A.: Quo vadis, action recognition? a new model and the kinetics dataset. In: proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 6299-6308 (2017) +6. Chalkiadakis, G., Elkind, E., Wooldridge, M.: Computational aspects of cooperative game theory. Synthesis Lectures on Artificial Intelligence and Machine Learning 5(6), 1-168 (2011) +7. Chen, J., Luo, W., Zhang, W., Ma, L.: Explore inter-contrast between videos via composition for weakly supervised temporal sentence grounding. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 36, pp. 267-275 (2022) +8. Chen, J., Ma, L., Chen, X., Jie, Z., Luo, J.: Localizing natural language in videos. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 33, pp. 8175-8182 (2019) +9. Chen, L., Lu, C., Tang, S., Xiao, J., Zhang, D., Tan, C., Li, X.: Rethinking the bottom-up framework for query-based video localization. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 34, pp. 10551-10558 (2020) +10. Chen, Z., Ma, L., Luo, W., Tang, P., Wong, K.Y.K.: Look closer to ground better: Weakly-supervised temporal grounding of sentence in video. arXiv preprint arXiv:2001.09308 (2020) +11. Datta, A., Sen, S., Zick, Y.: Algorithmic transparency via quantitative input influence: Theory and experiments with learning systems. In: 2016 IEEE symposium on security and privacy. pp. 598-617. IEEE (2016) +12. Deng, S., Wen, J., Liu, C., Yan, K., Xu, G., Xu, Y.: Projective incomplete multiview clustering. IEEE Transactions on Neural Networks and Learning Systems pp. 1-13 (2023). https://doi.org/10.1109/TNNLS.2023.3242473 +13. Dong, J., Chen, X., Zhang, M., Yang, X., Chen, S., Li, X., Wang, X.: Partially relevant video retrieval. In: Proceedings of the 30th ACM International Conference on Multimedia. pp. 246-257 (2022) +14. Dong, J., Li, X., Snoek, C.G.: Predicting visual features from text for image and video caption retrieval. IEEE Transactions on Multimedia 20(12), 3377-3388 (2018) +15. Dong, J., Li, X., Xu, C., Yang, X., Yang, G., Wang, X., Wang, M.: Dual encoding for video retrieval by text. IEEE Transactions on Pattern Analysis and Machine Intelligence 44(8), 4065-4080 (2022) +16. Dong, J., Peng, X., Ma, Z., Liu, D., Qu, X., Yang, X., Zhu, J., Liu, B.: From region to patch: Attribute-aware foreground-background contrastive learning for fine-grained fashion retrieval. In: Proceedings of the 46th International ACM SIGIR Conference on Research and Development in Information Retrieval. pp. 1273-1282 (2023) + +17. Dong, J., Sun, S., Liu, Z., Chen, S., Liu, B., Wang, X.: Hierarchical contrast for unsupervised skeleton-based action representation learning. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 37, pp. 525-533 (2023) +18. Dong, J., Wang, Y., Chen, X., Qu, X., Li, X., He, Y., Wang, X.: Reading-strategy inspired visual representation learning for text-to-video retrieval. IEEE Transactions on Circuits and Systems for Video Technology 32(8), 5680-5694 (2022) +19. Donoser, M., Bischof, H.: Diffusion processes for retrieval revisited. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1320-1327 (2013) +20. Dowdall, J., Pavlidis, I.T., Tsiamiyrtzis, P.: Coalitional tracking in facial infrared imaging and beyond. In: 2006 Conference on Computer Vision and Pattern Recognition Workshop. pp. 134-134. IEEE (2006) +21. Fang, X., Easwaran, A., Genest, B.: Uncertainty-guided appearance-motion association network for out-of-distribution action detection. In: IEEE International Conference on Multimedia Information Processing and Retrieval (2024) +22. Fang, X., Fang, W., Liu, D., Qu, X., Dong, J., Zhou, P., Li, R., Xu, Z., Chen, L., Zheng, P., Cheng, Y.: Not all inputs are valid: Towards open-set video moment retrieval using language. In: Proceedings of the 32th ACM International Conference on Multimedia (2024) +23. Fang, X., Hu, Y.: Double self-weighted multi-view clustering via adaptive view fusion. arXiv preprint arXiv:2011.10396 (2020) +24. Fang, X., Hu, Y., Zhou, P., Wu, D.: Animc: A soft approach for autoweighted noisy and incomplete multiview clustering. IEEE Transactions on Artificial Intelligence 3(2), 192-206 (2021) +25. Fang, X., Hu, Y., Zhou, P., Wu, D.O.: V3h: View variation and view heredity for incomplete multiview clustering. IEEE Transactions on Artificial Intelligence 1(3), 233-247 (2020) +26. Fang, X., Hu, Y., Zhou, P., Wu, D.O.: Unbalanced incomplete multi-view clustering via the scheme of view evolution: Weak views are meat; strong views do eat. IEEE Transactions on Emerging Topics in Computational Intelligence 6(4), 913-927 (2021) +27. Fang, X., Liu, D., Fang, W., Zhou, P., Cheng, Y., Tang, K., Zou, K.: Annotations are not all you need: A cross-modal knowledge transfer network for unsupervised temporal sentence grounding. In: Findings of the Association for Computational Linguistics: EMNLP 2023. pp. 8721-8733 (2023) +28. Fang, X., Liu, D., Fang, W., Zhou, P., Xu, Z., Xu, W., Chen, J., Li, R.: Fewer steps, better performance: Efficient cross-modal clip trimming for video moment retrieval using language. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 38, pp. 1735-1743 (2024) +29. Fang, X., Liu, D., Zhou, P., Hu, Y.: Multi-modal cross-domain alignment network for video moment retrieval. IEEE Transactions on Multimedia 25, 7517-7532 (2022) +30. Fang, X., Liu, D., Zhou, P., Nan, G.: You can ground earlier than see: An effective and efficient pipeline for temporal sentence grounding in compressed videos. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2448-2460 (2023) +31. Fang, X., Liu, D., Zhou, P., Xu, Z., Li, R.: Hierarchical local-global transformer for temporal sentence grounding. IEEE Transactions on Multimedia (2023) +32. Gao, J., Sun, C., Yang, Z., Nevatia, R.: Tall: Temporal activity localization via language query. In: Proceedings of the IEEE International Conference on Computer Vision. pp. 5267-5275 (2017) + +33. Gao, M., Davis, L., Socher, R., Xiong, C.: Wslln: Weakly supervised natural language localization networks. In: Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing. pp. 1481-1487 (2019) +34. Grabisch, M., Roubens, M.: An axiomatic approach to the concept of interaction among players in cooperative games. International Journal of game theory 28, 547-565 (1999) +35. Guo, C., Liu, D., Zhou, P.: A hybrid alignment loss for temporal moment localization with natural language. In: 2022 IEEE International Conference on Multimedia and Expo (ICME). pp. 1-6. IEEE (2022) +36. Guo, D., Li, K., Hu, B., Zhang, Y., Wang, M.: Benchmarking micro-action recognition: Dataset, method, and application. IEEE Transactions on Circuits and Systems for Video Technology 34(7), 6238-6252 (2024) +37. Hendricks, L.A., Wang, O., Shechtman, E., Sivic, J., Darrell, T., Russell, B.: Localizing moments in video with temporal language. Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing p. 1380-1390 (2018) +38. Huang, J., Liu, Y., Gong, S., Jin, H.: Cross-sentence temporal and semantic relations in video activity localisation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 7199-7208 (2021) +39. Jiang, L., Wang, C., Ning, X., Yu, Z.: Lttpoint: A mlp-based point cloud classification method with local topology transformation module. In: 2023 7th Asian Conference on Artificial Intelligence Technology (ACAIT). pp. 783-789. IEEE (2023) +40. Jin, P., Huang, J., Xiong, P., Tian, S., Liu, C., Ji, X., Yuan, L., Chen, J.: Videotext as game players: Hierarchical banzhaf interaction for cross-modal representation learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2472-2482 (2023) +41. Jin, S., Wang, S., Fang, F.: Game theoretical analysis on capacity configuration for microgrid based on multi-agent system. International Journal of Electrical Power & Energy Systems 125, 106485 (2021) +42. Kingma, D.P., Ba, J.: Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980 (2014) +43. Krishna, R., Hata, K., Ren, F., Fei-Fei, L., Carlos Niebles, J.: Dense-c captioning events in videos. In: Proceedings of the IEEE International Conference on Computer Vision. pp. 706-715 (2017) +44. Leech, D.: Computation of power indices (2002) +45. Lehrer, E.: An axiomatization of the banzhaf value. International Journal of Game Theory 17, 89-99 (1988) +46. Li, H., Cao, M., Cheng, X., Li, Y., Zhu, Z., Zou, Y.: G2l: Semantically aligned and uniform video grounding via geodesic and game theory. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 12032-12042 (2023) +47. Li, J., HE, X., Wei, L., Qian, L., Zhu, L., Xie, L., Zhuang, Y., Tian, Q., Tang, S.: Fine-grained semantically aligned vision-language pre-training. In: Advances in Neural Information Processing Systems (2022) +48. Lin, K.Q., Zhang, P., Chen, J., Pramanick, S., Gao, D., Wang, A.J., Yan, R., Shou, M.Z.: Univtg: Towards unified video-language temporal grounding. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 2794-2804 (2023) + +49. Lin, Z., Zhao, Z., Zhang, Z., Wang, Q., Liu, H.: Weakly-supervised video moment retrieval via semantic completion network. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 34, pp. 11539-11546 (2020) +50. Liu, C., Wen, J., Luo, X., Huang, C., Wu, Z., Xu, Y.: Dicnet: Deep instance-level contrastive network for double incomplete multi-view multi-label classification. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 37, pp. 8807-8815 (2023) +51. Liu, C., Wen, J., Luo, X., Xu, Y.: Incomplete multi-view multi-label learning via label-guided masked view- and category-aware transformers. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 37, pp. 8816-8824 (2023) +52. Liu, C., Wen, J., Wu, Z., Luo, X., Huang, C., Xu, Y.: Information recovery-driven deep incomplete multiview clustering network. IEEE Transactions on Neural Networks and Learning Systems pp. 1-11 (2023) +53. Liu, D., Fang, X., Hu, W., Zhou, P.: Exploring optical-flow-guided motion and detection-based appearance for temporal sentence grounding. IEEE Transactions on Multimedia 25, 8539-8553 (2023) +54. Liu, D., Fang, X., Qu, X., Dong, J., Yan, H., Yang, Y., Zhou, P., Cheng, Y.: Unsupervised domain adaptative temporal sentence localization with mutual information maximization. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 38, pp. 3567-3575 (2024) +55. Liu, D., Fang, X., Zhou, P., Di, X., Lu, W., Cheng, Y.: Hypotheses tree building for one-shot temporal sentence localization. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 37, pp. 1640-1648 (2023) +56. Liu, D., Hu, W.: Learning to focus on the foreground for temporal sentence grounding. In: Proceedings of the 29th International Conference on Computational Linguistics. pp. 5532-5541 (2022) +57. Liu, D., Hu, W.: Skimming, locating, then perusing: A human-like framework for natural language video localization. In: Proceedings of the 30th ACM International Conference on Multimedia. pp. 4536-4545 (2022) +58. Liu, D., Qu, X., Dong, J., Nan, G., Zhou, P., Xu, Z., Chen, L., Yan, H., Cheng, Y.: Filling the information gap between video and query for language-driven moment retrieval. In: Proceedings of the 31st ACM International Conference on Multimedia. pp. 4190-4199 (2023) +59. Liu, D., Qu, X., Dong, J., Zhou, P., Cheng, Y., Wei, W., Xu, Z., Xie, Y.: Context-aware biaffine localizing network for temporal sentence grounding. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 11235-11244 (2021) +60. Liu, D., Qu, X., Dong, J., Zhou, P., Xu, Z., Wang, H., Di, X., Lu, W., Cheng, Y.: Transform-equivariant consistency learning for temporal sentence grounding. ACM Transactions on Multimedia Computing, Communications and Applications 20(4), 1-19 (2024) +61. Liu, D., Qu, X., Fang, X., Dong, J., Zhou, P., Nan, G., Tang, K., Fang, W., Cheng, Y.: Towards robust temporal activity localization learning with noisy labels. In: Proceedings of the 2024 Joint International Conference on Computational Linguistics, Language Resources and Evaluation (LREC-COLING 2024). pp. 16630-16642 (2024) +62. Liu, D., Qu, X., Hu, W.: Reducing the vision and language bias for temporal sentence grounding. In: Proceedings of the 30th ACM International Conference on Multimedia. pp. 4092-4101 (2022) + +63. Liu, D., Qu, X., Liu, X.Y., Dong, J., Zhou, P., Xu, Z.: Jointly cross-and self-modal graph attention network for query-based moment localization. In: Proceedings of the 28th ACM International Conference on Multimedia. pp. 4070-4078 (2020) +64. Liu, D., Zhou, P.: Jointly visual-and semantic-aware graph memory networks for temporal sentence localization in videos. In: ICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). pp. 1-5. IEEE (2023) +65. Liu, D., Zhou, P., Xu, Z., Wang, H., Li, R.: Few-shot temporal sentence grounding via memory-guided semantic learning. IEEE Transactions on Circuits and Systems for Video Technology 33(5), 2491-2505 (2022) +66. Liu, D., Zhu, J., Fang, X., Xiong, Z., Wang, H., Li, R., Zhou, P.: Conditional video diffusion network for fine-grained temporal sentence grounding. IEEE Transactions on Multimedia (2023) +67. Lundberg, S.M., Lee, S.I.: A unified approach to interpreting model predictions. In: Proceedings of the 31st International Conference on Neural Information Processing Systems. pp. 4768-4777 (2017) +68. Ma, M., Yoon, S., Kim, J., Lee, Y., Kang, S., Yoo, C.D.: VLANet: Video-language alignment network for weakly-supervised video moment retrieval. In: Proceedings of the European Conference on Computer Vision. pp. 156-171 (2020) +69. Ma, W.C., Huang, D.A., Lee, N., Kitani, K.M.: Forecasting interactive dynamics of pedestrians with fictitious play. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 774-782 (2017) +70. Ma, Y., Liu, Y., Wang, L., Kang, W., Qiao, Y., Wang, Y.: Dual masked modeling for weakly-supervised temporal boundary discovery. IEEE Transactions on Multimedia (2023) +71. Matsui, Y., Matsui, T.: Np-completeness for calculating power indices of weighted majority games. Theoretical Computer Science 263(1-2), 305-310 (2001) +72. Michalak, T.P., Aadithya, K.V., Szczepanski, P.L., Ravindran, B., Jennings, N.R.: Efficient computation of the shapley value for game-theoretic network centrality. Journal of Artificial Intelligence Research 46, 607-650 (2013) +73. Mithun, N.C., Paul, S., Roy-Chowdhury, A.K.: Weakly supervised video moment retrieval from text queries. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 11592-11601 (2019) +74. Mun, J., Cho, M., Han, B.: Local-global video-text interactions for temporal grounding. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 10810-10819 (2020) +75. Ning, E., Wang, C., Zhang, H., Ning, X., Tiwari, P.: Occluded person re-identification with deep learning: a survey and perspectives. Expert Systems with Applications p. 122419 (2023) +76. Ning, E., Wang, Y., Wang, C., Zhang, H., Ning, X.: Enhancement, integration, expansion: Activating representation of detailed features for occluded person re-identification. Neural Networks 169, 532-541 (2024) +77. Ning, E., Zhang, C., Wang, C., Ning, X., Chen, H., Bai, X.: Pedestrian re-id based on feature consistency and contrast enhancement. Displays 79, 102467 (2023) +78. Nowak, A.S.: On an axiomatization of the banzhaf value without the additivity axiom. International Journal of Game Theory 26, 137-141 (1997) +79. Oord, A.v.d., Li, Y., Vinyals, O.: Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748 (2018) +80. Osborne, M.J., Rubinstein, A.: A course in game theory. MIT press (1994) + +81. Patel, R., Garnelo, M., Gemp, I., Dyer, C., Bachrach, Y.: Game-theoretic vocabulary selection via the shapley value and banzhaf index. In: Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies. pp. 2789-2798 (2021) +82. Pavan, M., Pelillo, M.: A new graph-theoretic approach to clustering and segmentation. In: 2003 IEEE Computer Society Conference on Computer Vision and Pattern Recognition, 2003. Proceedings. vol. 1, pp. I-I. IEEE (2003) +83. Pennington, J., Socher, R., Manning, C.D.: Glove: Global vectors for word representation. In: Proceedings of the Conference on Empirical Methods in Natural Language Processing. pp. 1532-1543 (2014) +84. Rodola, E., Bronstein, A.M., Albarelli, A., Bergamasco, F., Torsello, A.: A game-theoretic approach to deformable shape matching. In: 2012 IEEE Conference on Computer Vision and Pattern Recognition. pp. 182-189. IEEE (2012) +85. Shapley, L.S., et al.: A value for n-person games (1953) +86. Sigurdsson, G.A., Varol, G., Wang, X., Farhadi, A., Laptev, I., Gupta, A.: Hollywood in homes: Crowdsourcing data collection for activity understanding. In: European Conference on Computer Vision. pp. 510-526 (2016) +87. Song, Y., Wang, J., Ma, L., Yu, J., Liang, J., Yuan, L., Yu, Z.: Marn: Multilevel attentional reconstruction networks for weakly supervised video temporal grounding. Neurocomputing 554, 126625 (2023) +88. Song, Y., Wang, J., Ma, L., Yu, Z., Yu, J.: Weakly-supervised multi-level attentional reconstruction network for grounding textual queries in videos. arXiv preprint arXiv:2003.07048 (2020) +89. Tan, R., Xu, H., Saenko, K., Plummer, B.A.: Logan: Latent graph co-attention network for weakly-supervised video moment retrieval. In: Proceedings of the IEEE Winter Conference on Applications of Computer Vision. pp. 2083–2092 (2021) +90. Tang, K., Chen, Y., Peng, W., Zhang, Y., Fang, M., Wang, Z., Song, P.: Reppvconv: attentively fusing reparameterized voxel features for efficient 3d point cloud perception. The Visual Computer 39(11), 5577-5588 (2023) +91. Tang, K., Lou, T., Peng, W., Chen, N., Shi, Y., Wang, W.: Effective single-step adversarial training with energy-based models. IEEE Transactions on Emerging Topics in Computational Intelligence (2024). https://doi.org/10.1109/TETCI.2024.3378652 +92. Tang, K., Ma, Y., Miao, D., Song, P., Gu, Z., Tian, Z., Wang, W.: Decision fusion networks for image classification. IEEE Transactions on Neural Networks and Learning Systems (2022). https://doi.org/10.1109/TNNLS.2022.3196129 +93. Tang, K., Shi, Y., Lou, T., Peng, W., He, X., Zhu, P., Gu, Z., Tian, Z.: Rethinking perturbation directions for imperceptible adversarial attacks on point clouds. IEEE Internet of Things Journal 10(6), 5158-5169 (2022) +94. Tang, K., Zhao, W., Peng, W., Fang, X., Cui, X., Zhu, P., Tian, Z.: Reparameterization head for efficient multi-input networks. In: ICASSP 2024 - 2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). pp. 6190-6194 (2024). https://doi.org/10.1109/ICASSP48485.2024.10447574 +95. Tang, K., Zhao, W., Peng, W., Fang, X., Cui, X., Zhu, P., Tian, Z.: Reparameterization head for efficient multi-input networks. In: ICASSP 2024-2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). pp. 6190-6194. IEEE (2024) +96. Torsello, A., Bulo, S.R., Pelillo, M.: Grouping with asymmetric affinities: A game-theoretic perspective. In: 2006 IEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR'06). vol. 1, pp. 292-299. IEEE (2006) + +97. Tran, D., Bourdev, L., Fergus, R., Torresani, L., Paluri, M.: Learning spatiotemporal features with 3d convolutional networks. In: Proceedings of the IEEE International Conference on Computer Vision. pp. 4489-4497 (2015) +98. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. In: Advances in Neural Information Processing Systems. pp. 5998-6008 (2017) +99. Wang, C., Ning, X., Li, W., Bai, X., Gao, X.: 3d person re-identification based on global semantic guidance and local feature aggregation. IEEE Transactions on Circuits and Systems for Video Technology (2023) +100. Wang, C., Ning, X., Sun, L., Zhang, L., Li, W., Bai, X.: Learning discriminative features by covering local geometric space for point cloud analysis. IEEE Transactions on Geoscience and Remote Sensing 60, 1-15 (2022) +101. Wang, C., Wang, C., Li, W., Wang, H.: A brief survey on rgb-d semantic segmentation using deep learning. Displays 70, 102080 (2021) +102. Wang, C., Wang, H., Ning, X., Shengwei, T., Li, W.: 3d point cloud classification method based on dynamic coverage of local area. Journal of Software 34(4), 1962-1976 (2022) +103. Wang, J., Ma, L., Jiang, W.: Temporally grounding language queries in videos by contextual boundary-aware prediction. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 34, pp. 12168-12175 (2020) +104. Wang, Y., Deng, J., Zhou, W., Li, H.: Weakly supervised temporal adjacent network for language grounding. IEEE Transactions on Multimedia 24, 3276-3286 (2021) +105. Wang, Z., Chen, J., Jiang, Y.G.: Visual co-occurrence alignment learning for weakly-supervised video moment retrieval. In: Proceedings of the 29th ACM International Conference on Multimedia. pp. 1459–1468 (2021) +106. Wen, J., Liu, C., Deng, S., Liu, Y., Fei, L., Yan, K., Xu, Y.: Deep double incomplete multi-view multi-label learning with incomplete labels and missing views. IEEE Transactions on Neural Networks and Learning Systems pp. 1-13 (2023). https://doi.org/10.1109/TNNLS.2023.3260349 +107. Wen, J., Zhang, Z., Li, Z.J.: A survey on incomplete multiview clustering. IEEE transactions on systems, man, and cybernetics. Systems 53(2 Pt.2), 1136-1149 (2023) +108. Winter, E.: The shapley value. Handbook of game theory with economic applications 3, 2025-2054 (2002) +109. Wu, H., Lyu, Y., Shen, X., Zhao, X., Wang, M., Zhang, X., Luo, Z.: Atomic-action-based contrastive network for weakly supervised temporal language grounding. In: 2023 IEEE International Conference on Multimedia and Expo (ICME). pp. 1523-1528. IEEE (2023) +110. Xiong, Z., Liu, D., Zhou, P.: Gaussian kernel-based cross modal network for spatio-temporal video grounding. In: IEEE International Conference on Image Processing (ICIP). pp. 2481-2485 (2022) +111. Xiong, Z., Liu, D., Zhou, P., Zhu, J.: Tracking objects and activities with attention for temporal sentence grounding. In: ICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). pp. 1-5. IEEE (2023) +112. Yang, W., Zhang, T., Zhang, Y., Wu, F.: Local correspondence network for weakly supervised temporal sentence grounding. IEEE Transactions on Image Processing 30, 3252-3262 (2021) + +113. Yu, Z., Li, L., Xie, J., Wang, C., Li, W., Ning, X.: Pedestrian 3d shape understanding for person re-identification via multi-view learning. IEEE Transactions on Circuits and Systems for Video Technology (2024) +114. Yuan, Y., Ma, L., Wang, J., Liu, W., Zhu, W.: Semantic conditioned dynamic modulation for temporal sentence grounding in videos. In: Proceedings of the 33rd International Conference on Neural Information Processing Systems. pp. 536-546 (2019) +115. Zeng, R., Xu, H., Huang, W., Chen, P., Tan, M., Gan, C.: Dense regression network for video grounding. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 10287-10296 (2020) +116. Zhang, D., Dai, X., Wang, X., Wang, Y.F., Davis, L.S.: Man: Moment alignment network for natural language moment retrieval via iterative graph adjustment. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 1247-1257 (2019) +117. Zhang, H., Sun, A., Jing, W., Zhou, J.T.: Span-based localizing network for natural language video localization. In: Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics. pp. 6543-6554 (2020) +118. Zhang, H., Xie, Y., Zheng, L., Zhang, D., Zhang, Q.: Interpreting multivariate shapley interactions in dnns. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 35, pp. 10877-10886 (2021) +119. Zhang, H., Wang, C., Tian, S., Lu, B., Zhang, L., Ning, X., Bai, X.: Deep learning-based 3d point cloud classification: A systematic survey and outlook. Displays 79, 102456 (2023) +120. Zhang, H., Wang, C., Yu, L., Tian, S., Ning, X., Rodrigues, J.: Pointgt: A method for point-cloud classification and segmentation based on local geometric transformation. IEEE Transactions on Multimedia (2024) +121. Zhang, S., Peng, H., Fu, J., Luo, J.: Learning 2d temporal adjacent networks for moment localization with natural language. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 34, pp. 12870-12877 (2020) +122. Zhang, Z., Lin, Z., Zhao, Z., Xiao, Z.: Cross-modal interaction networks for query-based moment retrieval in videos. In: Proceedings of the 42nd International ACM SIGIR Conference on Research and Development in Information Retrieval. pp. 655-664 (2019) +123. Zhang, Z., Zhao, Z., Lin, Z., He, X., et al.: Counterfactual contrastive learning for weakly-supervised vision-language grounding. Advances in Neural Information Processing Systems 33, 18123-18134 (2020) +124. Zheng, M., Huang, Y., Chen, Q., Liu, Y.: Weakly supervised video moment localization with contrastive negative sample mining. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 36, pp. 3517-3525 (2022) +125. Zheng, M., Huang, Y., Chen, Q., Peng, Y., Liu, Y.: Weakly supervised temporal sentence grounding with gaussian-based contrastive proposal learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15555-15564 (2022) +126. Zheng, Q., Dong, J., Qu, X., Yang, X., Wang, Y., Zhou, P., Liu, B., Wang, X.: Progressive localization networks for language-based moment localization. ACM Transactions on Multimedia Computing, Communications and Applications 19(2), 1-21 (2023) +127. Zhu, J., Liu, D., Zhou, P., Di, X., Cheng, Y., Yang, S., Xu, W., Xu, Z., Wan, Y., Sun, L., et al.: Rethinking the video sampling and reasoning strategies for temporal sentence grounding. arXiv preprint arXiv:2301.00514 (2023) \ No newline at end of file diff --git a/2024/Rethinking Weakly-supervised Video Temporal Grounding From a Game Perspective/images.zip b/2024/Rethinking Weakly-supervised Video Temporal Grounding From a Game Perspective/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..c431c3f71526bdb8393982c6b41f8f016809aee1 --- /dev/null +++ b/2024/Rethinking Weakly-supervised Video Temporal Grounding From a Game Perspective/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0984ba94730a62a408888d4b4e7a44d23f785b4fff2804a012ecb5857a93add2 +size 554223 diff --git a/2024/Rethinking Weakly-supervised Video Temporal Grounding From a Game Perspective/layout.json b/2024/Rethinking Weakly-supervised Video Temporal Grounding From a Game Perspective/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..454eafb0aa8c49e8ebf7b98d067080b979882ff7 --- /dev/null +++ b/2024/Rethinking Weakly-supervised Video Temporal Grounding From a Game Perspective/layout.json @@ -0,0 +1,13515 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 141, + 111, + 473, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 111, + 473, + 148 + ], + "spans": [ + { + "bbox": [ + 141, + 111, + 473, + 148 + ], + "type": "text", + "content": "Rethinking Weakly-supervised Video Temporal Grounding From a Game Perspective" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 138, + 171, + 476, + 196 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 171, + 476, + 196 + ], + "spans": [ + { + "bbox": [ + 138, + 171, + 476, + 196 + ], + "type": "text", + "content": "Xiang Fang" + }, + { + "bbox": [ + 138, + 171, + 476, + 196 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 138, + 171, + 476, + 196 + ], + "type": "text", + "content": ", Zeyu Xiong" + }, + { + "bbox": [ + 138, + 171, + 476, + 196 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 138, + 171, + 476, + 196 + ], + "type": "text", + "content": ", Wanlong Fang" + }, + { + "bbox": [ + 138, + 171, + 476, + 196 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 138, + 171, + 476, + 196 + ], + "type": "text", + "content": ", Xiaoye Qu" + }, + { + "bbox": [ + 138, + 171, + 476, + 196 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 138, + 171, + 476, + 196 + ], + "type": "text", + "content": ", Chen Chen" + }, + { + "bbox": [ + 138, + 171, + 476, + 196 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 138, + 171, + 476, + 196 + ], + "type": "text", + "content": ", Jianfeng Dong" + }, + { + "bbox": [ + 138, + 171, + 476, + 196 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 138, + 171, + 476, + 196 + ], + "type": "text", + "content": ", Keke Tang" + }, + { + "bbox": [ + 138, + 171, + 476, + 196 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 138, + 171, + 476, + 196 + ], + "type": "text", + "content": ", Pan Zhou" + }, + { + "bbox": [ + 138, + 171, + 476, + 196 + ], + "type": "inline_equation", + "content": "^{1\\boxtimes}" + }, + { + "bbox": [ + 138, + 171, + 476, + 196 + ], + "type": "text", + "content": ", Yu Cheng" + }, + { + "bbox": [ + 138, + 171, + 476, + 196 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 138, + 171, + 476, + 196 + ], + "type": "text", + "content": ", and Daizong Liu" + }, + { + "bbox": [ + 138, + 171, + 476, + 196 + ], + "type": "inline_equation", + "content": "^{6\\boxtimes}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 132, + 204, + 481, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 204, + 481, + 216 + ], + "spans": [ + { + "bbox": [ + 132, + 204, + 481, + 216 + ], + "type": "text", + "content": "1 Hubei Key Laboratory of Distributed System Security, Hubei Engineering Research" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 140, + 217, + 473, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 140, + 217, + 473, + 228 + ], + "spans": [ + { + "bbox": [ + 140, + 217, + 473, + 228 + ], + "type": "text", + "content": "Center on Big Data Security, School of Cyber Science and Engineering, Huazhong" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 231, + 228, + 383, + 238 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 231, + 228, + 383, + 238 + ], + "spans": [ + { + "bbox": [ + 231, + 228, + 383, + 238 + ], + "type": "text", + "content": "University of Science and Technology" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 242, + 238, + 370, + 248 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 242, + 238, + 370, + 248 + ], + "spans": [ + { + "bbox": [ + 242, + 238, + 370, + 248 + ], + "type": "text", + "content": "2 University of Central Florida" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 238, + 248, + 375, + 259 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 238, + 248, + 375, + 259 + ], + "spans": [ + { + "bbox": [ + 238, + 248, + 375, + 259 + ], + "type": "text", + "content": "3 Zhejiang Gongshang University" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 257, + 260, + 357, + 271 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 257, + 260, + 357, + 271 + ], + "spans": [ + { + "bbox": [ + 257, + 260, + 357, + 271 + ], + "type": "text", + "content": "4 Guangzhou University" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 225, + 271, + 389, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 225, + 271, + 389, + 281 + ], + "spans": [ + { + "bbox": [ + 225, + 271, + 389, + 281 + ], + "type": "text", + "content": "5 The Chinese University of Hong Kong" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 265, + 282, + 348, + 293 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 265, + 282, + 348, + 293 + ], + "spans": [ + { + "bbox": [ + 265, + 282, + 348, + 293 + ], + "type": "text", + "content": "6 Peking University" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 149, + 293, + 462, + 304 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 293, + 462, + 304 + ], + "spans": [ + { + "bbox": [ + 149, + 293, + 462, + 304 + ], + "type": "text", + "content": "xfang9508@gmail.com, zeyuxiong@hust.edu.cn, wanlongfang@gmail.com," + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 157, + 305, + 453, + 315 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 157, + 305, + 453, + 315 + ], + "spans": [ + { + "bbox": [ + 157, + 305, + 453, + 315 + ], + "type": "text", + "content": "xiaoye@hust.edu.cn, chen.chen@crcv.ucf.edu, dongjf24@gmail.com," + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 146, + 316, + 466, + 326 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 316, + 466, + 326 + ], + "spans": [ + { + "bbox": [ + 146, + 316, + 466, + 326 + ], + "type": "text", + "content": "tangbohutbh@gmail.com, panzhou@hust.edu.cn, chengyu@cse.cuhk.edu.hk," + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 259, + 327, + 354, + 336 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 259, + 327, + 354, + 336 + ], + "spans": [ + { + "bbox": [ + 259, + 327, + 354, + 336 + ], + "type": "text", + "content": "dzliu@stu.pku.edu.cn" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 159, + 361, + 455, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 361, + 455, + 635 + ], + "spans": [ + { + "bbox": [ + 159, + 361, + 455, + 635 + ], + "type": "text", + "content": "Abstract. This paper addresses the challenging task of weakly-supervised video temporal grounding. Existing approaches are generally based on the moment proposal selection framework that utilizes contrastive learning and reconstruction paradigm for scoring the pre-defined moment proposals. Although they have achieved significant progress, we argue that their current frameworks have overlooked two indispensable issues: 1) Coarse-grained cross-modal learning: previous methods solely capture the global video-level alignment with the query, failing to model the detailed consistency between video frames and query words for accurately grounding the moment boundaries. 2) Complex moment proposals: their performance severely relies on the quality of proposals, which are also time-consuming and complicated for selection. To this end, in this paper, we make the first attempt to tackle this task from a novel game perspective, which effectively learns the uncertain relationship between each vision-language pair with diverse granularity and flexible combination for multi-level cross-modal interaction. Specifically, we creatively model each video frame and query word as game players with multivariate cooperative game theory to learn their contribution to the cross-modal similarity score. By quantifying the trend of frame-word cooperation within a coalition via the game-theoretic interaction, we are able to value all uncertain but possible correspondence between frames and words. Finally, instead of using moment proposals, we utilize the learned query-guided frame-wise scores for better moment localization. Experiments show that our method achieves superior performance on both Charades-STA and ActivityNet Caption datasets." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 130, + 642, + 457, + 654 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 642, + 457, + 654 + ], + "spans": [ + { + "bbox": [ + 130, + 642, + 457, + 654 + ], + "type": "text", + "content": "* Xiang Fang, Zeyu Xiong, and Wanlong Fang contributed equally to this work." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 654, + 353, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 654, + 353, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 654, + 353, + 665 + ], + "type": "text", + "content": "Corresponding authors: Pan Zhou and Daizong Liu." + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 115, + 342, + 193 + ], + "blocks": [ + { + "bbox": [ + 133, + 115, + 342, + 193 + ], + "lines": [ + { + "bbox": [ + 133, + 115, + 342, + 193 + ], + "spans": [ + { + "bbox": [ + 133, + 115, + 342, + 193 + ], + "type": "image", + "image_path": "1f60eb03a3c257228783a23a232c2c9f38ea109e0f0ad74cc5278ab11062fd33.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 194, + 482, + 248 + ], + "lines": [ + { + "bbox": [ + 130, + 194, + 482, + 248 + ], + "spans": [ + { + "bbox": [ + 130, + 194, + 482, + 248 + ], + "type": "text", + "content": "Fig. 1: (a) Most existing WS-VTG methods define moment proposals for query matching. They not only suffer from the coarse-grained alignment but also are limited to the quality of the proposals. (b) Instead of using proposals, our method reviews the WS-VTG from a brand-new game-interaction perspective to learn more fine-grained frame-aware self- and cross-modal alignment for accurate boundary localization." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 342, + 115, + 480, + 193 + ], + "blocks": [ + { + "bbox": [ + 342, + 115, + 480, + 193 + ], + "lines": [ + { + "bbox": [ + 342, + 115, + 480, + 193 + ], + "spans": [ + { + "bbox": [ + 342, + 115, + 480, + 193 + ], + "type": "image", + "image_path": "1efc26c781866c4d7d0c1625873b46e5ad008da2e5befe11042c33ba90443ae0.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 255, + 229, + 266 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 255, + 229, + 266 + ], + "spans": [ + { + "bbox": [ + 132, + 255, + 229, + 266 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 282, + 482, + 461 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 282, + 482, + 461 + ], + "spans": [ + { + "bbox": [ + 130, + 282, + 482, + 461 + ], + "type": "text", + "content": "Video temporal grounding (VTG) aims to localize the start and end timestamps of a specific event moment described by a given language query in an untrimmed video [2,32,39,75-77,99-102,113,119,120]. This task enables us to efficiently find video moments of human interest instead of traversing the entire video, which has broad application potential in video surveillance [13-15, 18, 24-26, 50-52, 59, 90-93, 125, 126], video summarization [12, 16, 17, 21-23, 36, 48, 94, 106, 107], etc. Most VTG methods [27-31, 53-55, 57, 58, 60-62, 65, 66, 127] follow the fully-supervised setting, where each frame is manually annotated as query-relevant or not. Although these methods achieve significant breakthroughs, they severely rely on extensive laborious manual annotations of moment boundaries, thus limiting their scalability and practicability in real-world applications. To alleviate the dense reliance, weakly-supervised VTG (WS-VTG) has received increasing attention [10, 49, 73, 88, 89, 123], which only requires the knowledge of matched video-query pairs without detailed frame-wise annotations, which is more challenging than the fully-supervised setting." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 462, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 462, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 462, + 482, + 666 + ], + "type": "text", + "content": "Since there is no frame-wise annotation in WS-VTG, learning detailed frame-query alignment is difficult. As shown in Fig. 1(a), almost all existing WS-VTG solutions directly employ the proposal-based framework [10,33,73,89] that first generates multiple moment proposals and then learns the scores to indicate the potential alignment between video proposals and language query. Some approaches further utilize reconstruction-based paradigm [49,88,123] to minimize the reconstruction loss between the partially masked query and moment proposals to identify the proposal which best reconstructs the query. Although existing WS-VTG methods have achieved great progress, we argue that they still suffer from two inescapable limitations: 1) Firstly, as frame-wise annotation is not provided, these WS-VTG methods generally generate the moment proposals for all videos via sliding windows. In this way, they only model the coarse-grained cross-modal interaction of each proposal-query pair. Actually, in most cases, we expect to capture more fine-grained information for accurate boundary localization, such as how much the semantics of a specific frame helps or harms the query-guided cross-modal alignment. Unfortunately, existing methods rely on proposal-level cross-modal learning cannot achieve this goal. 2) Secondly, the" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 304, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 304, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 304, + 102 + ], + "type": "text", + "content": "X. Fang, Z. Xiong, W. Fang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "content": "performance of these proposal-based methods is severely limited by the quality of moment proposals as they devise moment proposals regardless of the specific contents and difficulty of each video." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 152, + 482, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 152, + 482, + 415 + ], + "spans": [ + { + "bbox": [ + 130, + 152, + 482, + 415 + ], + "type": "text", + "content": "To tackle the above limitations, in this paper, we propose to address the WS-VTG task from a new game-theory perspective, which creatively models the subtle video frames and query words as game players to learn their uncertain but possible cross-modal interaction with diverse granularity and flexible combination. The game learning helps to generate query-guided frame-aware knowledge for fine-grained video grounding. As shown in Fig. 1(b), if visual representations and textual representations have strong semantic correspondence, they tend to cooperate together and contribute to a high cross-modal similarity score. By forming these cooperated frame-word representations as a coalition and quantifying the trend of cooperation within a coalition via the game-theoretic interaction index, we can not only learn the coalition contribution to the cross-modal semantic similarity, but also measure the additional benefits brought by the coalition compared with the costs of the lost coalitions of these players with others. In this manner, our cooperative WS-VTG game is able to value all possible correspondence between frames and words for sensitive and explainable cross-modal contrast. By formulating these detailed cross-modal correspondences as fine-grained query-guided frame-wise scores, we can efficiently generate the moment instead of using complicated proposals. In particular, we implement the cooperative game in both self-modal and cross-modal scenarios for self-modal attentive learning and cross-modal fine-grained alignment. We further extend the cross-modal game with multi-level query-semantic interaction for more comprehensive video understanding." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 146, + 415, + 372, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 415, + 372, + 426 + ], + "spans": [ + { + "bbox": [ + 146, + 415, + 372, + 426 + ], + "type": "text", + "content": "Our main contributions are summarized as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 432, + 479, + 584 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 138, + 432, + 479, + 490 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 432, + 479, + 490 + ], + "spans": [ + { + "bbox": [ + 138, + 432, + 479, + 490 + ], + "type": "text", + "content": "- We reveal the limitations of existing WS-VTG methods, which rely on complex moment proposals and fail to capture fine-grained frame-word alignment for accurate boundary grounding. To this end, we propose a novel game-theory based framework to learn the detailed frame-wise query-relevance scores for efficiently constructing the accurate moment." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 491, + 479, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 491, + 479, + 525 + ], + "spans": [ + { + "bbox": [ + 138, + 491, + 479, + 525 + ], + "type": "text", + "content": "- We apply the game learning to both self- and cross-modal scenarios for contextual self-enhancement and cross-alignment. The cross-modal game is also extended to multi-level for comprehensive learning." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 527, + 479, + 584 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 527, + 479, + 584 + ], + "spans": [ + { + "bbox": [ + 138, + 527, + 479, + 584 + ], + "type": "text", + "content": "- To verify the effectiveness of our framework, we conduct experiments on two challenging WS-VTG benchmarks, i.e., Charades-STA and ActivityNet Caption. Experiments show that our method achieves superior performance. In representative cases, our method outperforms all compared methods by " + }, + { + "bbox": [ + 138, + 527, + 479, + 584 + ], + "type": "inline_equation", + "content": "4.53\\%" + }, + { + "bbox": [ + 138, + 527, + 479, + 584 + ], + "type": "text", + "content": " on Charades-STA." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 132, + 605, + 242, + 617 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 605, + 242, + 617 + ], + "spans": [ + { + "bbox": [ + 132, + 605, + 242, + 617 + ], + "type": "text", + "content": "2 Related Works" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 629, + 479, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 479, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 479, + 665 + ], + "type": "text", + "content": "Fully-supervised video temporal grounding. Existing methods generally address the video temporal grounding in a fully-supervised manner, where both the annotations of video-sentence pairs and corresponding moment boundaries" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 252, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 252, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 252, + 91, + 447, + 102 + ], + "type": "text", + "content": "Rethinking WS-VTG From a Game Perspective" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 199 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 199 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 199 + ], + "type": "text", + "content": "are given. Most of them [32, 35, 37, 56, 61, 64, 66, 95, 103, 110, 111, 114, 116, 121, 122] utilize the proposal-based framework, which first integrate the sentence representation with those pre-defined moment proposals individually, and then evaluate their matching relationships. The proposal with the highest matching score is selected as the prediction. Instead of utilizing the moment proposals, recent proposal-free methods [8, 9, 74, 115] directly regress the temporal locations of the target moment." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 200, + 481, + 379 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 200, + 481, + 379 + ], + "spans": [ + { + "bbox": [ + 130, + 200, + 481, + 379 + ], + "type": "text", + "content": "Weakly-supervised video temporal grounding. As manually annotating temporal boundaries of target moments is time-consuming, recent research attentions shift to weakly-supervised video temporal grounding [10, 73, 88, 89], which only requires video-level annotations. Mithun et al. [73] proposed the first weakly-supervised model to learn a joint embedding space for video and query representations. Gao et al. [33] developed a two-stream structure to measure the moment-query consistency and conduct moment selection simultaneously. Although the above methods have achieved promising performance, they are two-stage approaches that utilize multi-scale sliding windows to generate moment candidates, therefore suffering from inferior effectiveness and efficiency. To address this issue, [49, 68, 123] score all the moments sampled at different scales in a single pass and further improve the moment-sentence matching accuracy. However, almost all of them rely on moment proposals for matching and selection, which fail to capture and distinguish more fine-grained details among visually similar frames for acquiring more accurate moment boundaries." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 380, + 482, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 380, + 482, + 582 + ], + "spans": [ + { + "bbox": [ + 130, + 380, + 482, + 582 + ], + "type": "text", + "content": "Cooperative game theory. Cooperative game theory focuses on the formation of coalitions and cooperation among rational players or groups to achieve common goals [1,6,40,41,47,69,80-82,96]. A typical cooperative game consists of a set of players with a game function. The game function maps all possible subsets of players, called group or coalition, to a number which represents the total payoff earned by these players working cooperatively to achieve the goal [19, 20, 46, 84]. To measure the contributions of each player and allocate different payoffs to these individuals fairly, a few researchers have proposed various concepts of value by computing the average added worth that one player brings to all possible coalitions, such as Shapley value [72, 85, 108] and Banzhaf value [4, 45, 78]. Recently, some concepts of cooperative game theory have been interpreted in the field of deep learning [11, 67, 118]. For example, Li et al. [47] propose a fine-grained image-text semantic alignment pre-training framework based on Shapley interaction. Different from these works, we propose for the first time to model the detailed consistency between video frames and query words by recasting multivariate game-theoretic indices to address more complicated WS-VTG task." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 602, + 233, + 617 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 602, + 233, + 617 + ], + "spans": [ + { + "bbox": [ + 132, + 602, + 233, + 617 + ], + "type": "text", + "content": "3 Methodology" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "type": "text", + "content": "In this section, we elaborate on the proposed method, which learns the uncertain and fine-grained cross-modal alignment in WS-VTG task. Specifically, our method conceptualizes the interaction between the video and query as a game," + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 304, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 304, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 304, + 102 + ], + "type": "text", + "content": "X. Fang, Z. Xiong, W. Fang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 132, + 114, + 482, + 276 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 482, + 276 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 482, + 276 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 482, + 276 + ], + "type": "image", + "image_path": "b26a4d14c5506599fbbb0c8864015dd78a5b7e1da1b8ceed3957a70cf47c5568.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 277, + 482, + 332 + ], + "lines": [ + { + "bbox": [ + 130, + 277, + 482, + 332 + ], + "spans": [ + { + "bbox": [ + 130, + 277, + 482, + 332 + ], + "type": "text", + "content": "Fig. 2: Overview of our framework, where \"SQAN\" denotes Sequential Query Attention Network. Specifically, we first conduct self-modal game to enhance the self-modal instance semantics, then devise multi-level cross-modal games to learn the fine-grained and uncertain cross-modal correspondence. At last, we predict the moment based on the query-guided frame-wise scores." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 343, + 482, + 416 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 343, + 482, + 416 + ], + "spans": [ + { + "bbox": [ + 130, + 343, + 482, + 416 + ], + "type": "text", + "content": "where each frame/word acts as a player and forms coalitions. During the game learning, visual frame and linguistic word representations with strong semantic relevance will cooperate to form a new coalition, the benefits of which are quantified through the game interaction index. Our method maximizes payoff by strategically forming coalitions to achieve fine-grained semantic alignment between video and query. The overall framework is shown in Fig. 2." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 442, + 228, + 453 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 442, + 228, + 453 + ], + "spans": [ + { + "bbox": [ + 132, + 442, + 228, + 453 + ], + "type": "text", + "content": "3.1 Preliminaries" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 472, + 482, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 472, + 482, + 568 + ], + "spans": [ + { + "bbox": [ + 130, + 472, + 482, + 568 + ], + "type": "text", + "content": "Problem formulation. Generally, given an untrimmed video and a sentence query, WS-VTG aims to localize the specific video moment semantically related to the query with only the video-level annotation. Let " + }, + { + "bbox": [ + 130, + 472, + 482, + 568 + ], + "type": "inline_equation", + "content": "V = \\{v_{1},\\dots,v_{i}\\dots,v_{T}\\}" + }, + { + "bbox": [ + 130, + 472, + 482, + 568 + ], + "type": "text", + "content": " denote the untrimmed video with " + }, + { + "bbox": [ + 130, + 472, + 482, + 568 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 130, + 472, + 482, + 568 + ], + "type": "text", + "content": " frames, and " + }, + { + "bbox": [ + 130, + 472, + 482, + 568 + ], + "type": "inline_equation", + "content": "Q = \\{q_{1},\\dots,q_{j},\\dots,q_{M}\\}" + }, + { + "bbox": [ + 130, + 472, + 482, + 568 + ], + "type": "text", + "content": " represent the sentence query containing " + }, + { + "bbox": [ + 130, + 472, + 482, + 568 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 130, + 472, + 482, + 568 + ], + "type": "text", + "content": " words, where " + }, + { + "bbox": [ + 130, + 472, + 482, + 568 + ], + "type": "inline_equation", + "content": "v_{i}" + }, + { + "bbox": [ + 130, + 472, + 482, + 568 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 472, + 482, + 568 + ], + "type": "inline_equation", + "content": "q_{j}" + }, + { + "bbox": [ + 130, + 472, + 482, + 568 + ], + "type": "text", + "content": " are the " + }, + { + "bbox": [ + 130, + 472, + 482, + 568 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 472, + 482, + 568 + ], + "type": "text", + "content": "-th frame and " + }, + { + "bbox": [ + 130, + 472, + 482, + 568 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 130, + 472, + 482, + 568 + ], + "type": "text", + "content": "-th word, respectively. The objective function " + }, + { + "bbox": [ + 130, + 472, + 482, + 568 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 130, + 472, + 482, + 568 + ], + "type": "text", + "content": " can be formulated as " + }, + { + "bbox": [ + 130, + 472, + 482, + 568 + ], + "type": "inline_equation", + "content": "\\mathcal{F}: (V,Q) \\to (\\tau_s,\\tau_e)" + }, + { + "bbox": [ + 130, + 472, + 482, + 568 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 472, + 482, + 568 + ], + "type": "inline_equation", + "content": "\\tau_s,\\tau_e" + }, + { + "bbox": [ + 130, + 472, + 482, + 568 + ], + "type": "text", + "content": " denote the start and end timestamps of the video moment boundary in video " + }, + { + "bbox": [ + 130, + 472, + 482, + 568 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 130, + 472, + 482, + 568 + ], + "type": "text", + "content": " semantically corresponding to query " + }, + { + "bbox": [ + 130, + 472, + 482, + 568 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 130, + 472, + 482, + 568 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 570, + 482, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 482, + 668 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 482, + 668 + ], + "type": "text", + "content": "Video and query representations. To extract contextual video representation, following [32, 117], each input video " + }, + { + "bbox": [ + 130, + 570, + 482, + 668 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 130, + 570, + 482, + 668 + ], + "type": "text", + "content": " is fed into a pre-trained 3D convolutional network [5, 97] with a stack of multi-head self-attention [98] layers for video encoding, then we get frame-wise video representation " + }, + { + "bbox": [ + 130, + 570, + 482, + 668 + ], + "type": "inline_equation", + "content": "\\mathbf{V} = \\{\\pmb{v}_i\\}_{i=1}^T \\in \\mathbb{R}^{T \\times D_h}" + }, + { + "bbox": [ + 130, + 570, + 482, + 668 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 570, + 482, + 668 + ], + "type": "inline_equation", + "content": "D_h" + }, + { + "bbox": [ + 130, + 570, + 482, + 668 + ], + "type": "text", + "content": " is the hidden dimension. For query representation, we also follow [59, 117] to embed each word of the query by the Glove [83] model with a further multi-head self-attention module and a BiLSTM layer. Similarly, we can obtain the word-wise query representation " + }, + { + "bbox": [ + 130, + 570, + 482, + 668 + ], + "type": "inline_equation", + "content": "\\mathbf{Q} = \\{\\pmb{q}_j\\}_{j=1}^M \\in \\mathbb{R}^{M \\times D_h}" + }, + { + "bbox": [ + 130, + 570, + 482, + 668 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 252, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 252, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 252, + 91, + 447, + 102 + ], + "type": "text", + "content": "Rethinking WS-VTG From a Game Perspective" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 128 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 480, + 128 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 480, + 128 + ], + "type": "text", + "content": "3.2 Game-theoretic Learning for Self-modal Semantic Enhancement" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 142, + 479, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 142, + 479, + 213 + ], + "spans": [ + { + "bbox": [ + 130, + 142, + 479, + 213 + ], + "type": "text", + "content": "We argue that the instance relations within each modality are important to infer the consecutive visual event contents or linguistic phrase contexts. To achieve this self-modal learning, we review previous classic game-theoretic value [34, 85] to explore the weighted semantic contribution of each frame/word to the whole video/query semantics for enhancing the self-modal representation of both video and query." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 215, + 480, + 297 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 215, + 480, + 297 + ], + "spans": [ + { + "bbox": [ + 131, + 215, + 480, + 297 + ], + "type": "text", + "content": "Visual self-modal game. Generally, a cooperative game consists of a player set " + }, + { + "bbox": [ + 131, + 215, + 480, + 297 + ], + "type": "inline_equation", + "content": "\\mathcal{P} = \\{1,2,\\dots,n\\}" + }, + { + "bbox": [ + 131, + 215, + 480, + 297 + ], + "type": "text", + "content": " and a game function " + }, + { + "bbox": [ + 131, + 215, + 480, + 297 + ], + "type": "inline_equation", + "content": "g(\\cdot)" + }, + { + "bbox": [ + 131, + 215, + 480, + 297 + ], + "type": "text", + "content": ". In detail, function " + }, + { + "bbox": [ + 131, + 215, + 480, + 297 + ], + "type": "inline_equation", + "content": "g(\\cdot)" + }, + { + "bbox": [ + 131, + 215, + 480, + 297 + ], + "type": "text", + "content": " maps each player subset " + }, + { + "bbox": [ + 131, + 215, + 480, + 297 + ], + "type": "inline_equation", + "content": "S \\subseteq \\mathcal{P}" + }, + { + "bbox": [ + 131, + 215, + 480, + 297 + ], + "type": "text", + "content": " to a score value, which indicates the payoff when all players in coalition " + }, + { + "bbox": [ + 131, + 215, + 480, + 297 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 131, + 215, + 480, + 297 + ], + "type": "text", + "content": " work together in the game. To capture the inherent correlations within the video modality, we directly take the video frames " + }, + { + "bbox": [ + 131, + 215, + 480, + 297 + ], + "type": "inline_equation", + "content": "\\mathbf{V} = \\{\\pmb{v}_i\\}_{i=1}^T" + }, + { + "bbox": [ + 131, + 215, + 480, + 297 + ], + "type": "text", + "content": " as frame players " + }, + { + "bbox": [ + 131, + 215, + 480, + 297 + ], + "type": "inline_equation", + "content": "\\mathcal{P} = \\mathbf{V}, n = T" + }, + { + "bbox": [ + 131, + 215, + 480, + 297 + ], + "type": "text", + "content": " to play the game and evaluate frame-wise contribution to frame-level video semantics." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 299, + 480, + 355 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 299, + 480, + 355 + ], + "spans": [ + { + "bbox": [ + 130, + 299, + 480, + 355 + ], + "type": "text", + "content": "Since different frames contribute and weight differently to video understanding, we attempt to adopt a game-theoretic value to measure the importance of each frame player in the cooperative game. Specifically, each frame player " + }, + { + "bbox": [ + 130, + 299, + 480, + 355 + ], + "type": "inline_equation", + "content": "\\pmb{v}_i" + }, + { + "bbox": [ + 130, + 299, + 480, + 355 + ], + "type": "text", + "content": " in player set " + }, + { + "bbox": [ + 130, + 299, + 480, + 355 + ], + "type": "inline_equation", + "content": "\\pmb{V}" + }, + { + "bbox": [ + 130, + 299, + 480, + 355 + ], + "type": "text", + "content": " has a weighted marginal contribution to the entire game, which can be measured by the game-theoretic value (e.g., Banzhaf value [34]) as:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 216, + 366, + 480, + 392 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 216, + 366, + 480, + 392 + ], + "spans": [ + { + "bbox": [ + 216, + 366, + 480, + 392 + ], + "type": "interline_equation", + "content": "B \\left(\\boldsymbol {v} _ {i} \\mid \\boldsymbol {V}\\right) = \\frac {1}{2 ^ {T - 1}} \\sum_ {\\mathcal {S} \\subseteq \\boldsymbol {V} \\backslash \\boldsymbol {v} _ {i}} [ g (\\mathcal {S} \\cup \\{\\boldsymbol {v} _ {i} \\}) - g (\\mathcal {S}) ], \\tag {1}", + "image_path": "ec4a538d1e3d838b570d88d7e43375ebd456669e77185e3be90aa62ade3680d1.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 404, + 480, + 512 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 404, + 480, + 512 + ], + "spans": [ + { + "bbox": [ + 130, + 404, + 480, + 512 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 404, + 480, + 512 + ], + "type": "inline_equation", + "content": "S \\subseteq V \\backslash \\mathbf{v}_i" + }, + { + "bbox": [ + 130, + 404, + 480, + 512 + ], + "type": "text", + "content": " represents a group formed without " + }, + { + "bbox": [ + 130, + 404, + 480, + 512 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_i" + }, + { + "bbox": [ + 130, + 404, + 480, + 512 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 404, + 480, + 512 + ], + "type": "inline_equation", + "content": "\\frac{1}{2^{T-1}}" + }, + { + "bbox": [ + 130, + 404, + 480, + 512 + ], + "type": "text", + "content": " is the likelihood of " + }, + { + "bbox": [ + 130, + 404, + 480, + 512 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 130, + 404, + 480, + 512 + ], + "type": "text", + "content": " being sampled, and " + }, + { + "bbox": [ + 130, + 404, + 480, + 512 + ], + "type": "inline_equation", + "content": "g(S \\cup \\{\\mathbf{v}_i\\}) - g(S)" + }, + { + "bbox": [ + 130, + 404, + 480, + 512 + ], + "type": "text", + "content": " denotes the marginal contribution brought by player " + }, + { + "bbox": [ + 130, + 404, + 480, + 512 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_i" + }, + { + "bbox": [ + 130, + 404, + 480, + 512 + ], + "type": "text", + "content": " joining the coalition " + }, + { + "bbox": [ + 130, + 404, + 480, + 512 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 130, + 404, + 480, + 512 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 130, + 404, + 480, + 512 + ], + "type": "inline_equation", + "content": "B(\\mathbf{v}_i | V)" + }, + { + "bbox": [ + 130, + 404, + 480, + 512 + ], + "type": "text", + "content": " indicates the ability of player " + }, + { + "bbox": [ + 130, + 404, + 480, + 512 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_i" + }, + { + "bbox": [ + 130, + 404, + 480, + 512 + ], + "type": "text", + "content": " to influence the video-level semantic by calculating the weighted marginal contribution of player " + }, + { + "bbox": [ + 130, + 404, + 480, + 512 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_i" + }, + { + "bbox": [ + 130, + 404, + 480, + 512 + ], + "type": "text", + "content": " to all possible coalitions " + }, + { + "bbox": [ + 130, + 404, + 480, + 512 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 130, + 404, + 480, + 512 + ], + "type": "text", + "content": ". Generally, within the context of our visual game, coalitions in the video modality represent consecutive frames of the same event or frames that are semantically similar. In this way, we can enumerate and learn all possible relations between the frames within the video for more contextual visual feature learning." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 513, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 513, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 131, + 513, + 481, + 666 + ], + "type": "text", + "content": "Visual soft supervision. To assist the above game-theoretic value learning, we first develop an additional learnable video-domain relevance header to predict the same self-modal relevance " + }, + { + "bbox": [ + 131, + 513, + 481, + 666 + ], + "type": "inline_equation", + "content": "\\hat{R}^v = \\{\\hat{r}_i^v\\}_{i=1}^T = \\text{header}(\\mathbf{V})" + }, + { + "bbox": [ + 131, + 513, + 481, + 666 + ], + "type": "text", + "content": " as Banzhaf values " + }, + { + "bbox": [ + 131, + 513, + 481, + 666 + ], + "type": "inline_equation", + "content": "\\{B(\\mathbf{v}_i|\\mathbf{V})\\}_{i=1}^T" + }, + { + "bbox": [ + 131, + 513, + 481, + 666 + ], + "type": "text", + "content": " among all frames for soft supervision (more discussions will be illustrated later). Specifically, the header is implemented by: 1) 1D convolutional layers for local self-consistency modeling, 2) a self-attention module for global self-relevance capturing, 3) a convolutional layer for decoding. Then, we optimize the Kullback-Leibler Divergence (KLD) between the predicted relevance " + }, + { + "bbox": [ + 131, + 513, + 481, + 666 + ], + "type": "inline_equation", + "content": "\\hat{r}_i^v" + }, + { + "bbox": [ + 131, + 513, + 481, + 666 + ], + "type": "text", + "content": " and the Banzhaf value " + }, + { + "bbox": [ + 131, + 513, + 481, + 666 + ], + "type": "inline_equation", + "content": "B(\\mathbf{v}_i|\\mathbf{V})" + }, + { + "bbox": [ + 131, + 513, + 481, + 666 + ], + "type": "text", + "content": " for learning the consistent self-modal similarity. In detail, we define the game-guided frame-wise probability distribution as " + }, + { + "bbox": [ + 131, + 513, + 481, + 666 + ], + "type": "inline_equation", + "content": "D^v = [p_1^v,p_2^v,\\dots,p_i^v,\\dots,p_T^v]^\\top" + }, + { + "bbox": [ + 131, + 513, + 481, + 666 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 131, + 513, + 481, + 666 + ], + "type": "inline_equation", + "content": "p_i^v = \\frac{\\exp(B(\\mathbf{v}_i|\\mathbf{V}))}{\\sum_{l=1}^T\\exp(B(\\mathbf{v}_l|\\mathbf{V}))}" + }, + { + "bbox": [ + 131, + 513, + 481, + 666 + ], + "type": "text", + "content": ". Similarly, the predicted frame-wise probability distribution calculated by the output of the self-relevance header can be denoted as " + }, + { + "bbox": [ + 131, + 513, + 481, + 666 + ], + "type": "inline_equation", + "content": "\\hat{D}^v = [\\hat{p}_1^v,\\hat{p}_2^v,\\dots,\\hat{p}_i^v,\\dots,\\hat{p}_T^v]^\\top" + }, + { + "bbox": [ + 131, + 513, + 481, + 666 + ], + "type": "text", + "content": ", where" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 304, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 304, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 304, + 102 + ], + "type": "text", + "content": "X. Fang, Z. Xiong, W. Fang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 114, + 447, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 114, + 447, + 133 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 447, + 133 + ], + "type": "inline_equation", + "content": "\\hat{p}_i^v = \\frac{\\exp(\\hat{r}_i^v)}{\\sum_{l=1}^{T}\\exp(\\hat{r}_l^v)}" + }, + { + "bbox": [ + 130, + 114, + 447, + 133 + ], + "type": "text", + "content": ". Finally, the self-relevance loss " + }, + { + "bbox": [ + 130, + 114, + 447, + 133 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{self}^v" + }, + { + "bbox": [ + 130, + 114, + 447, + 133 + ], + "type": "text", + "content": " for each video is:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 195, + 140, + 481, + 159 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 195, + 140, + 481, + 159 + ], + "spans": [ + { + "bbox": [ + 195, + 140, + 481, + 159 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {s e l f} ^ {v} = \\operatorname {K L D} \\left(\\hat {D} ^ {v} | | D ^ {v}\\right) = - \\frac {1}{T} \\sum_ {i = 1} ^ {T} p _ {i} ^ {v} \\left(\\log \\hat {p} _ {i} ^ {v} - \\log p _ {i} ^ {v}\\right). \\tag {2}", + "image_path": "ca11a8a5455117a351756e5edf38ea38fa15ad7fcc2b46c52c94ee00039de987.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 165, + 479, + 199 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 165, + 479, + 199 + ], + "spans": [ + { + "bbox": [ + 130, + 165, + 479, + 199 + ], + "type": "text", + "content": "Based on the learned game-theoretic value " + }, + { + "bbox": [ + 130, + 165, + 479, + 199 + ], + "type": "inline_equation", + "content": "B(\\pmb{v}_i|\\pmb{V})" + }, + { + "bbox": [ + 130, + 165, + 479, + 199 + ], + "type": "text", + "content": " for each player " + }, + { + "bbox": [ + 130, + 165, + 479, + 199 + ], + "type": "inline_equation", + "content": "\\pmb{v}_i" + }, + { + "bbox": [ + 130, + 165, + 479, + 199 + ], + "type": "text", + "content": " in player set " + }, + { + "bbox": [ + 130, + 165, + 479, + 199 + ], + "type": "inline_equation", + "content": "\\pmb{V}" + }, + { + "bbox": [ + 130, + 165, + 479, + 199 + ], + "type": "text", + "content": ", we deploy a Softmax function to obtain the self-modal enhanced video features " + }, + { + "bbox": [ + 130, + 165, + 479, + 199 + ], + "type": "inline_equation", + "content": "\\widetilde{\\pmb{v}}_i" + }, + { + "bbox": [ + 130, + 165, + 479, + 199 + ], + "type": "text", + "content": " as:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 188, + 205, + 481, + 228 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 188, + 205, + 481, + 228 + ], + "spans": [ + { + "bbox": [ + 188, + 205, + 481, + 228 + ], + "type": "interline_equation", + "content": "\\widetilde {\\boldsymbol {v}} _ {i} = \\boldsymbol {v} _ {i} + \\sum_ {\\mathcal {S} \\subseteq \\mathcal {V} \\backslash \\boldsymbol {v} _ {i}} \\operatorname {S o f t m a x} [ g (\\mathcal {S} \\cup \\{\\boldsymbol {v} _ {i} \\}) - g (\\mathcal {S}) ] \\cdot (\\mathcal {S} \\cup \\{\\boldsymbol {v} _ {i} \\}). \\tag {3}", + "image_path": "80bae2bb375a123906613830b9ca3a9689db5296e24cf9e4e30bc85a7fb86aa7.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 236, + 479, + 259 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 236, + 479, + 259 + ], + "spans": [ + { + "bbox": [ + 130, + 236, + 479, + 259 + ], + "type": "text", + "content": "The enhanced video features " + }, + { + "bbox": [ + 130, + 236, + 479, + 259 + ], + "type": "inline_equation", + "content": "\\widetilde{\\pmb{V}} = \\{\\widetilde{\\pmb{v}}_i\\}_{i=1}^T" + }, + { + "bbox": [ + 130, + 236, + 479, + 259 + ], + "type": "text", + "content": " aggregates the relevant contexts from the whole video." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 260, + 479, + 296 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 260, + 479, + 296 + ], + "spans": [ + { + "bbox": [ + 130, + 260, + 479, + 296 + ], + "type": "text", + "content": "Linguistic self-modal game. Similarly, we deem query words " + }, + { + "bbox": [ + 130, + 260, + 479, + 296 + ], + "type": "inline_equation", + "content": "\\mathbf{Q} = \\{\\mathbf{q}_j\\}_{j=1}^M" + }, + { + "bbox": [ + 130, + 260, + 479, + 296 + ], + "type": "text", + "content": " as players " + }, + { + "bbox": [ + 130, + 260, + 479, + 296 + ], + "type": "inline_equation", + "content": "\\mathcal{P} = \\mathbf{Q}, n = M" + }, + { + "bbox": [ + 130, + 260, + 479, + 296 + ], + "type": "text", + "content": " and calculate each word's marginal contribution " + }, + { + "bbox": [ + 130, + 260, + 479, + 296 + ], + "type": "inline_equation", + "content": "B(\\mathbf{q}_j|\\mathbf{Q})" + }, + { + "bbox": [ + 130, + 260, + 479, + 296 + ], + "type": "text", + "content": " to the whole query " + }, + { + "bbox": [ + 130, + 260, + 479, + 296 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}" + }, + { + "bbox": [ + 130, + 260, + 479, + 296 + ], + "type": "text", + "content": " by formulating the linguistic self-modal game as:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 214, + 302, + 481, + 328 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 214, + 302, + 481, + 328 + ], + "spans": [ + { + "bbox": [ + 214, + 302, + 481, + 328 + ], + "type": "interline_equation", + "content": "B \\left(\\boldsymbol {q} _ {j} \\mid \\boldsymbol {Q}\\right) = \\frac {1}{2 ^ {M - 1}} \\sum_ {\\mathcal {S} \\subseteq \\boldsymbol {Q} \\backslash \\boldsymbol {q} _ {j}} [ g (\\mathcal {S} \\cup \\{\\boldsymbol {q} _ {j} \\}) - g (\\mathcal {S}) ]. \\tag {4}", + "image_path": "8966b6fa631f153bd05bbd638b2dec847a1d4fef4ecd7a25f9ee12d598f128ff.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 335, + 479, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 335, + 479, + 399 + ], + "spans": [ + { + "bbox": [ + 131, + 335, + 479, + 399 + ], + "type": "text", + "content": "Linguistic soft supervision. We first take the game probability distributions " + }, + { + "bbox": [ + 131, + 335, + 479, + 399 + ], + "type": "inline_equation", + "content": "D^{q} = [p_{1}^{q}, p_{2}^{q}, \\dots, p_{j}^{q}, \\dots, p_{M}^{q}]^{\\top}" + }, + { + "bbox": [ + 131, + 335, + 479, + 399 + ], + "type": "text", + "content": " as the soft label, and then supervise the predicted distributions " + }, + { + "bbox": [ + 131, + 335, + 479, + 399 + ], + "type": "inline_equation", + "content": "\\hat{D}^{q} = [\\hat{p}_{1}^{q}, \\hat{p}_{2}^{q}, \\dots, \\hat{p}_{j}^{q}, \\dots, \\hat{p}_{M}^{q}]^{\\top}" + }, + { + "bbox": [ + 131, + 335, + 479, + 399 + ], + "type": "text", + "content": " of another header by KLD loss " + }, + { + "bbox": [ + 131, + 335, + 479, + 399 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{self}^{q}" + }, + { + "bbox": [ + 131, + 335, + 479, + 399 + ], + "type": "text", + "content": " similar to Eq. (2). At last, we can obtain the enhanced query features " + }, + { + "bbox": [ + 131, + 335, + 479, + 399 + ], + "type": "inline_equation", + "content": "\\tilde{\\boldsymbol{Q}} = \\{\\tilde{\\boldsymbol{q}}_{j}\\}_{j=1}^{M}" + }, + { + "bbox": [ + 131, + 335, + 479, + 399 + ], + "type": "text", + "content": " by:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 188, + 407, + 481, + 432 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 188, + 407, + 481, + 432 + ], + "spans": [ + { + "bbox": [ + 188, + 407, + 481, + 432 + ], + "type": "interline_equation", + "content": "\\widetilde {\\boldsymbol {q}} _ {j} = \\boldsymbol {q} _ {j} + \\sum_ {\\mathcal {S} \\subseteq \\mathcal {Q} \\backslash \\boldsymbol {q} _ {j}} \\operatorname {S o f t m a x} [ g (\\mathcal {S} \\cup \\{\\boldsymbol {q} _ {j} \\}) - g (\\mathcal {S}) ] \\cdot (\\mathcal {S} \\cup \\{\\boldsymbol {q} _ {j} \\}). \\tag {5}", + "image_path": "326f4545a63db2e66ad8f767fa2007c7d73602961f0ea482bf097741c0f8c0d0.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 437, + 481, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 437, + 481, + 582 + ], + "spans": [ + { + "bbox": [ + 130, + 437, + 481, + 582 + ], + "type": "text", + "content": "Discussion on learnable header. Although we can directly utilize the game interaction value to enumerate the relations between different players via Eqs. (1) and (4), there are still two challenges: 1) No supervision signal: The performance of the game interaction severely depends on the quality of the learned player features. However, there is no supervision for learning the representative features. 2) Complex and time-consuming: The game process is complex and costs much time and resources. To this end, we develop a learnable header to mimic the game interaction by utilizing the KLD loss. This KLD function serves as a soft supervision label to not only learn the consistency between the game interaction and header, but also potentially train more distinguishing player features of each modality. During the inference, we can solely utilize the prediction of header to model the game interaction instead of Eqs. (1) and (4)." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 131, + 598, + 443, + 622 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 598, + 443, + 622 + ], + "spans": [ + { + "bbox": [ + 131, + 598, + 443, + 622 + ], + "type": "text", + "content": "3.3 Game-theoretic Interaction for Cross-modal Multi-level Alignment" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 629, + 479, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 479, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 479, + 665 + ], + "type": "text", + "content": "After obtaining the self-enhanced video and query features, we further develop a cross-modal game between frames and words to handle their uncertainty during their fine-grained semantic alignment." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 252, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 252, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 252, + 91, + 448, + 102 + ], + "type": "text", + "content": "Rethinking WS-VTG From a Game Perspective" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 322 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 322 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 322 + ], + "type": "text", + "content": "Cross-modal game. From a game-theoretic perspective, we attempt to conceptualize the problem of fine-grained cross-modal alignment as a collaborative effort between video frames and query words. Specifically, when a video frame and a query word exhibit a high degree of semantic similarity, they are more likely to collaborate and their union will have a larger score contributed to the final moment. To this end, we take the enhanced video features " + }, + { + "bbox": [ + 130, + 116, + 482, + 322 + ], + "type": "inline_equation", + "content": "\\widetilde{V} = \\{\\widetilde{v}_i\\}_{i=1}^T" + }, + { + "bbox": [ + 130, + 116, + 482, + 322 + ], + "type": "text", + "content": " and enhanced query features " + }, + { + "bbox": [ + 130, + 116, + 482, + 322 + ], + "type": "inline_equation", + "content": "\\widetilde{Q} = \\{\\widetilde{q}_j\\}_{j=1}^M" + }, + { + "bbox": [ + 130, + 116, + 482, + 322 + ], + "type": "text", + "content": " as two different kinds of players " + }, + { + "bbox": [ + 130, + 116, + 482, + 322 + ], + "type": "inline_equation", + "content": "\\mathcal{P} = \\{\\widetilde{v}_i\\}_{i=1}^T \\cup \\{\\widetilde{q}_j\\}_{j=1}^M" + }, + { + "bbox": [ + 130, + 116, + 482, + 322 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 116, + 482, + 322 + ], + "type": "inline_equation", + "content": "n = T + M" + }, + { + "bbox": [ + 130, + 116, + 482, + 322 + ], + "type": "text", + "content": ". In a cooperative game, different players tend to work together in groups, called coalitions, to achieve a common goal, i.e., representing the same semantics of the target event. In the case of two players " + }, + { + "bbox": [ + 130, + 116, + 482, + 322 + ], + "type": "inline_equation", + "content": "\\widetilde{v}_i" + }, + { + "bbox": [ + 130, + 116, + 482, + 322 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 116, + 482, + 322 + ], + "type": "inline_equation", + "content": "\\widetilde{q}_j" + }, + { + "bbox": [ + 130, + 116, + 482, + 322 + ], + "type": "text", + "content": " in set " + }, + { + "bbox": [ + 130, + 116, + 482, + 322 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 130, + 116, + 482, + 322 + ], + "type": "text", + "content": ", it may occur that " + }, + { + "bbox": [ + 130, + 116, + 482, + 322 + ], + "type": "inline_equation", + "content": "g(\\{\\widetilde{v}_i\\})" + }, + { + "bbox": [ + 130, + 116, + 482, + 322 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 116, + 482, + 322 + ], + "type": "inline_equation", + "content": "g(\\{\\widetilde{q}_j\\})" + }, + { + "bbox": [ + 130, + 116, + 482, + 322 + ], + "type": "text", + "content": " are small individually, but at the same time the reward of their forming coalition " + }, + { + "bbox": [ + 130, + 116, + 482, + 322 + ], + "type": "inline_equation", + "content": "g(\\{\\widetilde{v}_i, \\widetilde{q}_j\\})" + }, + { + "bbox": [ + 130, + 116, + 482, + 322 + ], + "type": "text", + "content": " is considerable. This is because players in the coalition interact and collaborate with each other, which may bring additional benefits (or costs) to the game. Therefore, we also need to measure the additional benefits brought by the coalition compared with the costs of the lost coalitions of these players with others. To form a better coalition " + }, + { + "bbox": [ + 130, + 116, + 482, + 322 + ], + "type": "inline_equation", + "content": "\\{\\widetilde{v}_i, \\widetilde{q}_j\\} \\subseteq \\mathcal{P}" + }, + { + "bbox": [ + 130, + 116, + 482, + 322 + ], + "type": "text", + "content": ", we utilize the effective interaction index " + }, + { + "bbox": [ + 130, + 116, + 482, + 322 + ], + "type": "inline_equation", + "content": "I_B(\\{\\widetilde{v}_i, \\widetilde{q}_j\\})" + }, + { + "bbox": [ + 130, + 116, + 482, + 322 + ], + "type": "text", + "content": " like [34] to compute the additional benefit as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 143, + 327, + 481, + 354 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 327, + 481, + 354 + ], + "spans": [ + { + "bbox": [ + 143, + 327, + 481, + 354 + ], + "type": "interline_equation", + "content": "I _ {B} \\left(\\left\\{\\widetilde {\\boldsymbol {v}} _ {i}, \\widetilde {\\boldsymbol {q}} _ {j} \\right\\}\\right) = \\frac {1}{2 ^ {n - 2}} \\sum_ {\\mathcal {S} \\subseteq \\mathcal {P} \\backslash \\left\\{\\widetilde {\\boldsymbol {v}} _ {i}, \\widetilde {\\boldsymbol {q}} _ {j} \\right\\}} \\left[ g \\left(\\mathcal {S} \\cup \\left\\{\\widetilde {\\boldsymbol {v}} _ {i}, \\widetilde {\\boldsymbol {q}} _ {j} \\right\\}\\right) + g (\\mathcal {S}) - g \\left(\\mathcal {S} \\cup \\left\\{\\widetilde {\\boldsymbol {v}} _ {i} \\right\\}\\right) - g \\left(\\mathcal {S} \\cup \\left\\{\\widetilde {\\boldsymbol {q}} _ {j} \\right\\}\\right) \\right], \\tag {6}", + "image_path": "44744a64f840be5c10704d714a6eeb74f168d6769ffebced78c8485f57e614d4.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 361, + 482, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 361, + 482, + 456 + ], + "spans": [ + { + "bbox": [ + 130, + 361, + 482, + 456 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 361, + 482, + 456 + ], + "type": "inline_equation", + "content": "\\frac{1}{2^{n-2}}" + }, + { + "bbox": [ + 130, + 361, + 482, + 456 + ], + "type": "text", + "content": " is the likelihood of " + }, + { + "bbox": [ + 130, + 361, + 482, + 456 + ], + "type": "inline_equation", + "content": "\\mathcal{S} \\subseteq \\mathcal{P} \\backslash \\{\\widetilde{v}_i, \\widetilde{q}_j\\}" + }, + { + "bbox": [ + 130, + 361, + 482, + 456 + ], + "type": "text", + "content": " being sampled. Intuitively, " + }, + { + "bbox": [ + 130, + 361, + 482, + 456 + ], + "type": "inline_equation", + "content": "I_B(\\{\\widetilde{v}_i, \\widetilde{q}_j\\})" + }, + { + "bbox": [ + 130, + 361, + 482, + 456 + ], + "type": "text", + "content": " quantifies this additional benefit compared to when players " + }, + { + "bbox": [ + 130, + 361, + 482, + 456 + ], + "type": "inline_equation", + "content": "\\widetilde{v}_i" + }, + { + "bbox": [ + 130, + 361, + 482, + 456 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 361, + 482, + 456 + ], + "type": "inline_equation", + "content": "\\widetilde{q}_j" + }, + { + "bbox": [ + 130, + 361, + 482, + 456 + ], + "type": "text", + "content": " work independently. It also embodies the tendency of interaction between frame-player and word-player. A higher value of " + }, + { + "bbox": [ + 130, + 361, + 482, + 456 + ], + "type": "inline_equation", + "content": "I_B(\\{\\widetilde{v}_i, \\widetilde{q}_j\\})" + }, + { + "bbox": [ + 130, + 361, + 482, + 456 + ], + "type": "text", + "content": " signifies that players " + }, + { + "bbox": [ + 130, + 361, + 482, + 456 + ], + "type": "inline_equation", + "content": "\\widetilde{v}_i" + }, + { + "bbox": [ + 130, + 361, + 482, + 456 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 361, + 482, + 456 + ], + "type": "inline_equation", + "content": "\\widetilde{q}_j" + }, + { + "bbox": [ + 130, + 361, + 482, + 456 + ], + "type": "text", + "content": " cooperate closely with each other, and the formed coalition will bring additional high returns. By this interaction index, we can measure the closeness between plays " + }, + { + "bbox": [ + 130, + 361, + 482, + 456 + ], + "type": "inline_equation", + "content": "\\widetilde{v}_i" + }, + { + "bbox": [ + 130, + 361, + 482, + 456 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 361, + 482, + 456 + ], + "type": "inline_equation", + "content": "\\widetilde{q}_j" + }, + { + "bbox": [ + 130, + 361, + 482, + 456 + ], + "type": "text", + "content": " to achieve fine-grained alignment between video and query." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 457, + 482, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 457, + 482, + 582 + ], + "spans": [ + { + "bbox": [ + 130, + 457, + 482, + 582 + ], + "type": "text", + "content": "Game details. To ensure complete consistency between the cooperative game and cross-modal alignment learning, the game function " + }, + { + "bbox": [ + 130, + 457, + 482, + 582 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 130, + 457, + 482, + 582 + ], + "type": "text", + "content": " should meet the following criteria: 1) the game payoff benefits from strongly corresponding semantic pairs " + }, + { + "bbox": [ + 130, + 457, + 482, + 582 + ], + "type": "inline_equation", + "content": "\\{\\pmb{v}_i^+, \\pmb{q}_j^+\\}" + }, + { + "bbox": [ + 130, + 457, + 482, + 582 + ], + "type": "text", + "content": ", i.e., " + }, + { + "bbox": [ + 130, + 457, + 482, + 582 + ], + "type": "inline_equation", + "content": "g(\\mathcal{P}) - g(\\mathcal{P} \\setminus \\{\\pmb{v}_i^+, \\pmb{q}_j^+\\} \\cup \\{[\\{\\pmb{v}_i^+, \\pmb{q}_j^+\\}]\\}) < 0" + }, + { + "bbox": [ + 130, + 457, + 482, + 582 + ], + "type": "text", + "content": "; 2) the game payoff is compromised by semantically opposite pairs " + }, + { + "bbox": [ + 130, + 457, + 482, + 582 + ], + "type": "inline_equation", + "content": "\\{\\pmb{v}_i^-, \\pmb{q}_j^-\\}" + }, + { + "bbox": [ + 130, + 457, + 482, + 582 + ], + "type": "text", + "content": ", i.e., " + }, + { + "bbox": [ + 130, + 457, + 482, + 582 + ], + "type": "inline_equation", + "content": "g(\\mathcal{P}) - g(\\mathcal{P} \\setminus \\{\\pmb{v}_i^-, \\pmb{q}_j^-\\} \\cup \\{[\\{\\pmb{v}_i^-, \\pmb{q}_j^-\\}]\\}) > 0" + }, + { + "bbox": [ + 130, + 457, + 482, + 582 + ], + "type": "text", + "content": "; 3) when there are no players to cooperate, the payoff is zero, i.e., " + }, + { + "bbox": [ + 130, + 457, + 482, + 582 + ], + "type": "inline_equation", + "content": "g(\\{\\pmb{v}_i\\}_{i=1}^T) = g(\\{\\pmb{q}_j\\}_{j=1}^M) = g(\\phi) = 0" + }, + { + "bbox": [ + 130, + 457, + 482, + 582 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 457, + 482, + 582 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 130, + 457, + 482, + 582 + ], + "type": "text", + "content": " is the empty set. It should be noted that any function that satisfies the aforementioned conditions can be employed as the game function " + }, + { + "bbox": [ + 130, + 457, + 482, + 582 + ], + "type": "inline_equation", + "content": "g(\\cdot)" + }, + { + "bbox": [ + 130, + 457, + 482, + 582 + ], + "type": "text", + "content": ". To simplify matters, we utilize the cosine similarity as " + }, + { + "bbox": [ + 130, + 457, + 482, + 582 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 130, + 457, + 482, + 582 + ], + "type": "text", + "content": " in all games." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 582, + 482, + 630 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 482, + 630 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 482, + 630 + ], + "type": "text", + "content": "During the game, when players " + }, + { + "bbox": [ + 130, + 582, + 482, + 630 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_i" + }, + { + "bbox": [ + 130, + 582, + 482, + 630 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 582, + 482, + 630 + ], + "type": "inline_equation", + "content": "\\mathbf{q}_j" + }, + { + "bbox": [ + 130, + 582, + 482, + 630 + ], + "type": "text", + "content": " form a coalition " + }, + { + "bbox": [ + 130, + 582, + 482, + 630 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{v}_i, \\mathbf{q}_j\\}" + }, + { + "bbox": [ + 130, + 582, + 482, + 630 + ], + "type": "text", + "content": ", we deem " + }, + { + "bbox": [ + 130, + 582, + 482, + 630 + ], + "type": "inline_equation", + "content": "[\\{\\mathbf{v}_i, \\mathbf{q}_j\\}]" + }, + { + "bbox": [ + 130, + 582, + 482, + 630 + ], + "type": "text", + "content": " as a singular hypothetical player, which is the union of the player in " + }, + { + "bbox": [ + 130, + 582, + 482, + 630 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{v}_i, \\mathbf{q}_j\\}" + }, + { + "bbox": [ + 130, + 582, + 482, + 630 + ], + "type": "text", + "content": ". Then, the reduced game is formed by removing the individual players in " + }, + { + "bbox": [ + 130, + 582, + 482, + 630 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{v}_i, \\mathbf{q}_j\\}" + }, + { + "bbox": [ + 130, + 582, + 482, + 630 + ], + "type": "text", + "content": " from the game and adding " + }, + { + "bbox": [ + 130, + 582, + 482, + 630 + ], + "type": "inline_equation", + "content": "[\\{\\mathbf{v}_i, \\mathbf{q}_j\\}]" + }, + { + "bbox": [ + 130, + 582, + 482, + 630 + ], + "type": "text", + "content": " to the game." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 630, + 482, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 630, + 482, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 630, + 482, + 667 + ], + "type": "text", + "content": "Word-level alignment with local game interaction. Similar to the self-modal game, we develop an alignment header to predict the word-frame alignment matrix " + }, + { + "bbox": [ + 130, + 630, + 482, + 667 + ], + "type": "inline_equation", + "content": "A = [a_{i,j}]^{T \\times M}" + }, + { + "bbox": [ + 130, + 630, + 482, + 667 + ], + "type": "text", + "content": " between frame " + }, + { + "bbox": [ + 130, + 630, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\widetilde{\\boldsymbol{v}}_i" + }, + { + "bbox": [ + 130, + 630, + 482, + 667 + ], + "type": "text", + "content": " and word " + }, + { + "bbox": [ + 130, + 630, + 482, + 667 + ], + "type": "inline_equation", + "content": "\\widetilde{\\boldsymbol{q}}_j" + }, + { + "bbox": [ + 130, + 630, + 482, + 667 + ], + "type": "text", + "content": ", supervised by" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 304, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 304, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 304, + 102 + ], + "type": "text", + "content": "X. Fang, Z. Xiong, W. Fang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 115, + 479, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 115, + 479, + 163 + ], + "spans": [ + { + "bbox": [ + 130, + 115, + 479, + 163 + ], + "type": "inline_equation", + "content": "I_B(\\{\\widetilde{v}_i,\\widetilde{q}_j\\})" + }, + { + "bbox": [ + 130, + 115, + 479, + 163 + ], + "type": "text", + "content": " . Then, we optimize the Kullback-Leibler Divergence (KLD) between the " + }, + { + "bbox": [ + 130, + 115, + 479, + 163 + ], + "type": "inline_equation", + "content": "I_B(\\{\\widetilde{v}_i,\\widetilde{q}_j\\})" + }, + { + "bbox": [ + 130, + 115, + 479, + 163 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 115, + 479, + 163 + ], + "type": "inline_equation", + "content": "a_{i,j}" + }, + { + "bbox": [ + 130, + 115, + 479, + 163 + ], + "type": "text", + "content": " , to bring the probability distributions of the output of the alignment header and interaction index close together to establish word-level semantic alignment between frame players and word players." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 164, + 482, + 263 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 164, + 482, + 263 + ], + "spans": [ + { + "bbox": [ + 130, + 164, + 482, + 263 + ], + "type": "text", + "content": "In detail, we define the word-to-frame probability distribution " + }, + { + "bbox": [ + 130, + 164, + 482, + 263 + ], + "type": "inline_equation", + "content": "d_{w2f}^{j}" + }, + { + "bbox": [ + 130, + 164, + 482, + 263 + ], + "type": "text", + "content": " of the " + }, + { + "bbox": [ + 130, + 164, + 482, + 263 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 130, + 164, + 482, + 263 + ], + "type": "text", + "content": "-th word-to-frame alignment as " + }, + { + "bbox": [ + 130, + 164, + 482, + 263 + ], + "type": "inline_equation", + "content": "d_{w2f}^{j} = [p_{1,j}, \\dots, p_{t,j}]" + }, + { + "bbox": [ + 130, + 164, + 482, + 263 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 164, + 482, + 263 + ], + "type": "inline_equation", + "content": "p_{i,j} = \\frac{\\exp(I_B(\\{\\widetilde{v}_i, \\widetilde{q}_j\\}))}{\\sum_{l=1}^T \\exp(I_B(\\{\\widetilde{v}_l, \\widetilde{q}_j\\}))}" + }, + { + "bbox": [ + 130, + 164, + 482, + 263 + ], + "type": "text", + "content": " and the total query-to-video probability distribution " + }, + { + "bbox": [ + 130, + 164, + 482, + 263 + ], + "type": "inline_equation", + "content": "D_{q2v} = [d_{w2f}^1, \\dots, d_{w2f}^M]^\\top" + }, + { + "bbox": [ + 130, + 164, + 482, + 263 + ], + "type": "text", + "content": ". Similarly, the predicted fine-grained probability distribution can be represented as " + }, + { + "bbox": [ + 130, + 164, + 482, + 263 + ], + "type": "inline_equation", + "content": "\\hat{d}_{w2f}^{j} = [\\hat{p}_{1,j}, \\dots, \\hat{p}_{t,j}]" + }, + { + "bbox": [ + 130, + 164, + 482, + 263 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 164, + 482, + 263 + ], + "type": "inline_equation", + "content": "\\hat{p}_{i,j} = \\frac{\\exp(a_{i,j})}{\\sum_{l=1}^t \\exp(a_{l,j})}" + }, + { + "bbox": [ + 130, + 164, + 482, + 263 + ], + "type": "text", + "content": ", and the total predicted probability distribution is defined as " + }, + { + "bbox": [ + 130, + 164, + 482, + 263 + ], + "type": "inline_equation", + "content": "\\hat{D}_{q2v} = [\\hat{d}_{w2f}^1, \\dots, \\hat{d}_{w2f}^M]^\\top" + }, + { + "bbox": [ + 130, + 164, + 482, + 263 + ], + "type": "text", + "content": ". At last, the word-level alignment loss " + }, + { + "bbox": [ + 130, + 164, + 482, + 263 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{word}" + }, + { + "bbox": [ + 130, + 164, + 482, + 263 + ], + "type": "text", + "content": " is formulated as:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 162, + 272, + 481, + 292 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 162, + 272, + 481, + 292 + ], + "spans": [ + { + "bbox": [ + 162, + 272, + 481, + 292 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {w o r d}} = \\operatorname {K L D} \\left(\\hat {D} _ {q 2 v} | | D _ {q 2 v}\\right) = - \\frac {1}{T M} \\sum_ {i = 1} ^ {T} \\sum_ {j = 1} ^ {M} p _ {i, j} \\left(\\log \\hat {p} _ {i, j} - \\log p _ {i, j}\\right). \\tag {7}", + "image_path": "1bc1c14b81793572f6554010187446deb4460b27e636c9439daffae9b44e5723.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 300, + 482, + 497 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 300, + 482, + 497 + ], + "spans": [ + { + "bbox": [ + 130, + 300, + 482, + 497 + ], + "type": "text", + "content": "Phrase- and sentence-level alignment with global game interaction. Although the above cross-modal game-theoretic interaction between each possible player-pair " + }, + { + "bbox": [ + 130, + 300, + 482, + 497 + ], + "type": "inline_equation", + "content": "\\{\\widetilde{v}_i,\\widetilde{q}_j\\}" + }, + { + "bbox": [ + 130, + 300, + 482, + 497 + ], + "type": "text", + "content": " addresses the problem of word-level alignment between word player " + }, + { + "bbox": [ + 130, + 300, + 482, + 497 + ], + "type": "inline_equation", + "content": "\\widetilde{v}_i" + }, + { + "bbox": [ + 130, + 300, + 482, + 497 + ], + "type": "text", + "content": " and frame player " + }, + { + "bbox": [ + 130, + 300, + 482, + 497 + ], + "type": "inline_equation", + "content": "\\widetilde{q}_j" + }, + { + "bbox": [ + 130, + 300, + 482, + 497 + ], + "type": "text", + "content": ", it is still limited to aligning frames with isolated words, and fails to explore the relations between frames with more contextual phrase- or sentence-level semantics for better understanding temporal events. To this end, we introduce a multi-level game interaction to measure the semantic consistency between frames with different-grained textual semantics. Specifically, we follow [63, 74] to utilize Sequential Query Attention Network (SQAN) to extract the phrase-level textual semantics " + }, + { + "bbox": [ + 130, + 300, + 482, + 497 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{Q}}^p = \\{\\tilde{\\mathbf{q}}_j^p\\}_{j = 1}^P" + }, + { + "bbox": [ + 130, + 300, + 482, + 497 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 300, + 482, + 497 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 130, + 300, + 482, + 497 + ], + "type": "text", + "content": " is the phrase number. The sentence-level semantics " + }, + { + "bbox": [ + 130, + 300, + 482, + 497 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{Q}}^s" + }, + { + "bbox": [ + 130, + 300, + 482, + 497 + ], + "type": "text", + "content": " is obtained by concatenating the last hidden states in both the forward and backward BiLSTM. Similar to worl-level alignment, we compute both phrase- and sentence-level text-to-frame game interaction as " + }, + { + "bbox": [ + 130, + 300, + 482, + 497 + ], + "type": "inline_equation", + "content": "I_B(\\{\\widetilde{v}_i,\\widetilde{q}_j^p\\})" + }, + { + "bbox": [ + 130, + 300, + 482, + 497 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 300, + 482, + 497 + ], + "type": "inline_equation", + "content": "I_B(\\{\\widetilde{v}_i,\\tilde{Q}^s\\})" + }, + { + "bbox": [ + 130, + 300, + 482, + 497 + ], + "type": "text", + "content": " via Eq. (6). We design corresponding alignment headers (phrase-level cross-modal header and sentence-level cross-modal header) to predict their alignment matrix via " + }, + { + "bbox": [ + 130, + 300, + 482, + 497 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{phra},\\mathcal{L}_{sent}" + }, + { + "bbox": [ + 130, + 300, + 482, + 497 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 516, + 311, + 528 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 516, + 311, + 528 + ], + "spans": [ + { + "bbox": [ + 132, + 516, + 311, + 528 + ], + "type": "text", + "content": "3.4 Model Training and Inference" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 537, + 480, + 572 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 537, + 480, + 572 + ], + "spans": [ + { + "bbox": [ + 130, + 537, + 480, + 572 + ], + "type": "text", + "content": "Weak supervision for video-query matching. We first interact each frame with the semantics of words, phrases and the whole sentence to calculate the overall query-to-frame matching degree by:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 161, + 580, + 481, + 601 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 161, + 580, + 481, + 601 + ], + "spans": [ + { + "bbox": [ + 161, + 580, + 481, + 601 + ], + "type": "interline_equation", + "content": "m _ {i} = \\frac {1}{3} \\left(I _ {B} \\left(\\{\\widetilde {\\boldsymbol {v}} _ {i}, \\widetilde {\\boldsymbol {Q}} ^ {s} \\}\\right) + \\frac {1}{M} \\sum_ {j = 1} ^ {M} I _ {B} \\left(\\{\\widetilde {\\boldsymbol {v}} _ {i}, \\widetilde {\\boldsymbol {q}} _ {j} \\}\\right) + \\frac {1}{P} \\sum_ {l = 1} ^ {P} I _ {B} \\left(\\left\\{\\widetilde {\\boldsymbol {v}} _ {i}, \\widetilde {\\boldsymbol {q}} _ {l} ^ {p} \\right\\}\\right)\\right), \\tag {8}", + "image_path": "a56af99af356294745f8ff0a376c31d80d99785aac4c344945fc80a9c1cba951.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 609, + 438, + 620 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 609, + 438, + 620 + ], + "spans": [ + { + "bbox": [ + 131, + 609, + 438, + 620 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 609, + 438, + 620 + ], + "type": "inline_equation", + "content": "m_{i}" + }, + { + "bbox": [ + 131, + 609, + 438, + 620 + ], + "type": "text", + "content": " is the overall query-to-frame matching degree of " + }, + { + "bbox": [ + 131, + 609, + 438, + 620 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 131, + 609, + 438, + 620 + ], + "type": "text", + "content": "-th frame." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 146, + 620, + 429, + 632 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 620, + 429, + 632 + ], + "spans": [ + { + "bbox": [ + 146, + 620, + 429, + 632 + ], + "type": "text", + "content": "Then, the whole video-sentence matching score is calculated by:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 203, + 643, + 481, + 669 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 203, + 643, + 481, + 669 + ], + "spans": [ + { + "bbox": [ + 203, + 643, + 481, + 669 + ], + "type": "interline_equation", + "content": "\\phi^ {c} = \\sum_ {i = 1} ^ {T} w _ {i} \\cdot m _ {i}, w _ {i} = \\frac {\\exp \\left(I _ {B} \\left(\\left\\{\\widetilde {\\boldsymbol {v}} _ {i} , \\widetilde {\\boldsymbol {Q}} ^ {s} \\right\\}\\right)\\right)}{\\sum_ {i = 1} ^ {T} \\exp \\left(I _ {B} \\left(\\left\\{\\widetilde {\\boldsymbol {v}} _ {i} , \\widetilde {\\boldsymbol {Q}} ^ {s} \\right\\}\\right)\\right)}, \\tag {9}", + "image_path": "b67918aee34ee48c7a323403c8a92d1d3ba4e6b36fb9b01d5e1dcfe502a3e5d1.jpg" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 252, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 252, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 252, + 91, + 448, + 102 + ], + "type": "text", + "content": "Rethinking WS-VTG From a Game Perspective" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 165 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 165 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 165 + ], + "type": "text", + "content": "where we utilize the sentence-level similarity to represent the contribution of each frame to the target moment among the entire video. For matched video " + }, + { + "bbox": [ + 130, + 116, + 479, + 165 + ], + "type": "inline_equation", + "content": "V_{k}" + }, + { + "bbox": [ + 130, + 116, + 479, + 165 + ], + "type": "text", + "content": " and query " + }, + { + "bbox": [ + 130, + 116, + 479, + 165 + ], + "type": "inline_equation", + "content": "Q_{k}" + }, + { + "bbox": [ + 130, + 116, + 479, + 165 + ], + "type": "text", + "content": ", we calculate the video-sentence matching score " + }, + { + "bbox": [ + 130, + 116, + 479, + 165 + ], + "type": "inline_equation", + "content": "\\phi_{V_k,Q_k}^c" + }, + { + "bbox": [ + 130, + 116, + 479, + 165 + ], + "type": "text", + "content": ". The scores of unmatched pairs are expressed as " + }, + { + "bbox": [ + 130, + 116, + 479, + 165 + ], + "type": "inline_equation", + "content": "\\phi_{V_k,Q_l}^c" + }, + { + "bbox": [ + 130, + 116, + 479, + 165 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 116, + 479, + 165 + ], + "type": "inline_equation", + "content": "\\phi_{Q_k,V_l}^c" + }, + { + "bbox": [ + 130, + 116, + 479, + 165 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 130, + 116, + 479, + 165 + ], + "type": "inline_equation", + "content": "l\\neq k" + }, + { + "bbox": [ + 130, + 116, + 479, + 165 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 146, + 165, + 480, + 177 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 165, + 480, + 177 + ], + "spans": [ + { + "bbox": [ + 146, + 165, + 480, + 177 + ], + "type": "text", + "content": "Finally, we utilize the cross-modal contrastive loss [79] for weak supervision:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 140, + 191, + 479, + 218 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 140, + 191, + 479, + 218 + ], + "spans": [ + { + "bbox": [ + 140, + 191, + 479, + 218 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {C} = - \\frac {1}{N _ {b}} \\sum_ {k = 1} ^ {N _ {b}} \\log \\frac {\\exp \\left(\\phi_ {V _ {k} , Q _ {k}} ^ {c} / \\tau\\right)}{\\sum_ {l = 1} ^ {N _ {b}} \\exp \\left(\\phi_ {V _ {k} , Q _ {l}} ^ {c} / \\tau\\right)} - \\frac {1}{N _ {b}} \\sum_ {k = 1} ^ {N _ {b}} \\log \\frac {\\exp \\left(\\phi_ {Q _ {k} , V _ {k}} ^ {c} / \\tau\\right)}{\\sum_ {l = 1} ^ {N _ {b}} \\exp \\left(\\phi_ {Q _ {k} , V _ {l}} ^ {c} / \\tau\\right)}, \\tag {10}", + "image_path": "788ac94f03df8c94ef08e0d130a35ec8062705345b833169f320921de52503d6.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 223, + 395, + 234 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 223, + 395, + 234 + ], + "spans": [ + { + "bbox": [ + 131, + 223, + 395, + 234 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 223, + 395, + 234 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 131, + 223, + 395, + 234 + ], + "type": "text", + "content": " is the temperature parameter, " + }, + { + "bbox": [ + 131, + 223, + 395, + 234 + ], + "type": "inline_equation", + "content": "N_{b}" + }, + { + "bbox": [ + 131, + 223, + 395, + 234 + ], + "type": "text", + "content": " is the batch size." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 235, + 480, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 235, + 480, + 342 + ], + "spans": [ + { + "bbox": [ + 130, + 235, + 480, + 342 + ], + "type": "text", + "content": "Reconstruction loss. To further enhance the semantic relevance of the highlighted frames to the sentence, we reconstruct the query conditioned on highlighted frames. Following popular reconstruction strategy [49,124], we randomly replace the 1/3 words in the original query with a specific symbol, and predict the next word by given a prefix of the query and highlighted frames feature. Specifically, we embed the masked query " + }, + { + "bbox": [ + 130, + 235, + 480, + 342 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}^m" + }, + { + "bbox": [ + 130, + 235, + 480, + 342 + ], + "type": "text", + "content": " by Glove and predict the next word " + }, + { + "bbox": [ + 130, + 235, + 480, + 342 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{q}}_{i+1}" + }, + { + "bbox": [ + 130, + 235, + 480, + 342 + ], + "type": "text", + "content": " using the conditioned transformer following CNM [124] with input " + }, + { + "bbox": [ + 130, + 235, + 480, + 342 + ], + "type": "inline_equation", + "content": "(\\tilde{\\mathbf{V}}, C_m, Q_{1:i}^m)" + }, + { + "bbox": [ + 130, + 235, + 480, + 342 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 235, + 480, + 342 + ], + "type": "inline_equation", + "content": "C_m = \\{m_1, m_2, \\dots, m_T\\}" + }, + { + "bbox": [ + 130, + 235, + 480, + 342 + ], + "type": "text", + "content": " represents the highlighted frame mask, " + }, + { + "bbox": [ + 130, + 235, + 480, + 342 + ], + "type": "inline_equation", + "content": "Q_{1:i}^m" + }, + { + "bbox": [ + 130, + 235, + 480, + 342 + ], + "type": "text", + "content": " denotes the masked query words from the 1-st to " + }, + { + "bbox": [ + 130, + 235, + 480, + 342 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 235, + 480, + 342 + ], + "type": "text", + "content": "-th." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 342, + 479, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 342, + 479, + 365 + ], + "spans": [ + { + "bbox": [ + 130, + 342, + 479, + 365 + ], + "type": "text", + "content": "Then, we introduce the cross-entropy loss " + }, + { + "bbox": [ + 130, + 342, + 479, + 365 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{rec}" + }, + { + "bbox": [ + 130, + 342, + 479, + 365 + ], + "type": "text", + "content": " to calculate the probability distribution " + }, + { + "bbox": [ + 130, + 342, + 479, + 365 + ], + "type": "inline_equation", + "content": "\\mathcal{P}^q" + }, + { + "bbox": [ + 130, + 342, + 479, + 365 + ], + "type": "text", + "content": " different between the predict word " + }, + { + "bbox": [ + 130, + 342, + 479, + 365 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{q}}_{i + 1}" + }, + { + "bbox": [ + 130, + 342, + 479, + 365 + ], + "type": "text", + "content": " and the real word as:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 236, + 371, + 479, + 389 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 236, + 371, + 479, + 389 + ], + "spans": [ + { + "bbox": [ + 236, + 371, + 479, + 389 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {r e c} = - \\sum_ {i = 1} ^ {M - 1} \\log \\mathcal {P} ^ {q} \\left(\\boldsymbol {q} _ {i + 1} \\mid \\hat {\\boldsymbol {q}} _ {i + 1}\\right). \\tag {11}", + "image_path": "98437a7144851e278beb839a8d3965cc7d8b84e8089a2ce33e07d3bfeef77b4e.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 394, + 479, + 440 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 394, + 479, + 440 + ], + "spans": [ + { + "bbox": [ + 131, + 394, + 479, + 440 + ], + "type": "text", + "content": "Overall training. To train the whole model, we combine game-guided self-relevance losses " + }, + { + "bbox": [ + 131, + 394, + 479, + 440 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{self} = \\mathcal{L}_{self}^v +\\mathcal{L}_{self}^q" + }, + { + "bbox": [ + 131, + 394, + 479, + 440 + ], + "type": "text", + "content": ", game alignment loss " + }, + { + "bbox": [ + 131, + 394, + 479, + 440 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{align} = \\mathcal{L}_{word} + \\mathcal{L}_{phra} + \\mathcal{L}_{sent}" + }, + { + "bbox": [ + 131, + 394, + 479, + 440 + ], + "type": "text", + "content": ", video-level contrastive loss " + }, + { + "bbox": [ + 131, + 394, + 479, + 440 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_C" + }, + { + "bbox": [ + 131, + 394, + 479, + 440 + ], + "type": "text", + "content": ", and reconstruction loss " + }, + { + "bbox": [ + 131, + 394, + 479, + 440 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{rec}" + }, + { + "bbox": [ + 131, + 394, + 479, + 440 + ], + "type": "text", + "content": " with trade-off hyper-parameters " + }, + { + "bbox": [ + 131, + 394, + 479, + 440 + ], + "type": "inline_equation", + "content": "(\\alpha_{1},\\alpha_{2},\\alpha_{3}" + }, + { + "bbox": [ + 131, + 394, + 479, + 440 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 394, + 479, + 440 + ], + "type": "inline_equation", + "content": "\\alpha_4)" + }, + { + "bbox": [ + 131, + 394, + 479, + 440 + ], + "type": "text", + "content": " to define the overall loss as:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 222, + 447, + 479, + 457 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 222, + 447, + 479, + 457 + ], + "spans": [ + { + "bbox": [ + 222, + 447, + 479, + 457 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\alpha_ {1} \\mathcal {L} _ {\\text {s e l f}} + \\alpha_ {2} \\mathcal {L} _ {\\text {a l i g n}} + \\alpha_ {3} \\mathcal {L} _ {C} + \\alpha_ {4} \\mathcal {L} _ {\\text {r e c}}. \\tag {12}", + "image_path": "613652c197a0348de3a687ceff520e9ee252557a0667e682da9864882fffa2d4.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 464, + 480, + 558 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 464, + 480, + 558 + ], + "spans": [ + { + "bbox": [ + 130, + 464, + 480, + 558 + ], + "type": "text", + "content": "Inference. During inference, we directly utilize the self-relevance headers to enhance the video and query features, and exploit the multi-level alignment headers to predict the overall query-to-frame matching degree. We locate the frame with the highest score as the basic predicted moment, and add the left/right frames into the moment if the ratio of their scores to the frame score of the closest moment boundary is higher than a threshold. We repeat this step until no frame can be added. We can obtain more moments by locating different initial frames with different scores." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 131, + 576, + 230, + 590 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 576, + 230, + 590 + ], + "spans": [ + { + "bbox": [ + 131, + 576, + 230, + 590 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 131, + 599, + 326, + 611 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 599, + 326, + 611 + ], + "spans": [ + { + "bbox": [ + 131, + 599, + 326, + 611 + ], + "type": "text", + "content": "4.1 Datasets and Evaluation Metrics" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 617, + 479, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 479, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 479, + 665 + ], + "type": "text", + "content": "Charades-STA. The Charades-STA dataset [32] is built based on the Charades [86], which contains 6,672 videos of indoor activities and involves 16,128 query-video pairs. There are 12,408 pairs used for training and 3,720 used for testing. The average duration of each video is 29.76 seconds." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 304, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 304, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 304, + 102 + ], + "type": "text", + "content": "X. Fang, Z. Xiong, W. Fang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 186 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 186 + ], + "type": "text", + "content": "ActivityNet Caption. The ActivityNet Caption dataset [43] contains 20,000 videos with 100,000 queries, where 37,421 query-video pairs are used for training and 34,536 are used for testing. On average, each video in ActivityNet Caption has 3.65 annotated moments and each annotated moment lasts for 36 seconds. For fair comparison, we follow [10, 73, 88, 89, 123, 124] to use val1 as valuation and report the test performance on val2." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 188, + 480, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 188, + 480, + 224 + ], + "spans": [ + { + "bbox": [ + 130, + 188, + 480, + 224 + ], + "type": "text", + "content": "Evaluation metrics. Following previous works, we adopt the metrics “R@n, IoU=m” to evaluate our model, which presents the proportion of the top " + }, + { + "bbox": [ + 130, + 188, + 480, + 224 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 130, + 188, + 480, + 224 + ], + "type": "text", + "content": " predicted moments with IoU larger than " + }, + { + "bbox": [ + 130, + 188, + 480, + 224 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 130, + 188, + 480, + 224 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 240, + 279, + 252 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 240, + 279, + 252 + ], + "spans": [ + { + "bbox": [ + 131, + 240, + 279, + 252 + ], + "type": "text", + "content": "4.2 Implementation Details" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 259, + 480, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 259, + 480, + 426 + ], + "spans": [ + { + "bbox": [ + 130, + 259, + 480, + 426 + ], + "type": "text", + "content": "For fair comparison, we follow previous works to apply C3D [97] to encode the videos on ActivityNet Caption and I3D [5] on Charades-STA. We uniformly downsample the length of video feature sequences to " + }, + { + "bbox": [ + 130, + 259, + 480, + 426 + ], + "type": "inline_equation", + "content": "T = 200" + }, + { + "bbox": [ + 130, + 259, + 480, + 426 + ], + "type": "text", + "content": " for ActivityNet Caption and " + }, + { + "bbox": [ + 130, + 259, + 480, + 426 + ], + "type": "inline_equation", + "content": "T = 64" + }, + { + "bbox": [ + 130, + 259, + 480, + 426 + ], + "type": "text", + "content": " for Charades-STA. As for query encoding, we also follow previous works to utilize set the length of word feature sequences to " + }, + { + "bbox": [ + 130, + 259, + 480, + 426 + ], + "type": "inline_equation", + "content": "M = 20" + }, + { + "bbox": [ + 130, + 259, + 480, + 426 + ], + "type": "text", + "content": ", and utilize Glove [83] to embed each word. The dimension " + }, + { + "bbox": [ + 130, + 259, + 480, + 426 + ], + "type": "inline_equation", + "content": "D_h" + }, + { + "bbox": [ + 130, + 259, + 480, + 426 + ], + "type": "text", + "content": " is set to 512. These trade-off hyper-parameters are set to " + }, + { + "bbox": [ + 130, + 259, + 480, + 426 + ], + "type": "inline_equation", + "content": "\\alpha_1 = 1.0, \\alpha_2 = 1.0, \\alpha_3 = 1.0, \\alpha_4 = 10" + }, + { + "bbox": [ + 130, + 259, + 480, + 426 + ], + "type": "text", + "content": ". Since the calculation of the exact game-theoretic interaction is an NP-hard problem [71], we follow previous works to utilize sampling-based method [3,44] to obtain unbiased estimates for approximating it. We train the whole model for 100 epochs with batch size of 16 and early stopping strategy. Parameter optimization is performed by Adam [42] optimizer with learning rate " + }, + { + "bbox": [ + 130, + 259, + 480, + 426 + ], + "type": "inline_equation", + "content": "3 \\times 10^{-4}" + }, + { + "bbox": [ + 130, + 259, + 480, + 426 + ], + "type": "text", + "content": ", and linear decay of learning rate and gradient clipping of 1.0. The inference threshold is set to 0.8 in ActivityNet Caption and 0.9 in Charades-STA." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 443, + 228, + 454 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 443, + 228, + 454 + ], + "spans": [ + { + "bbox": [ + 131, + 443, + 228, + 454 + ], + "type": "text", + "content": "4.3 Main Results" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 462, + 480, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 462, + 480, + 616 + ], + "spans": [ + { + "bbox": [ + 130, + 462, + 480, + 616 + ], + "type": "text", + "content": "Quantitative comparison. We compare our method with the state-of-the-art methods, including fully-supervised (FS) methods and weakly-supervised (WS) methods. Best results are in bold. We introduce two types of cooperative game indices, i.e., Banzhaf index [34] and Shapley index [85], to our model. As summarized in Tab. 1, both two different variants of our method surpass all existing methods on both Charades-STA and ActivityNet Captions. Particularly, on Charades-STA in terms of \"R@5, IoU=0.5\", Ours(Banzhaf) outperforms the best compared method CPL by " + }, + { + "bbox": [ + 130, + 462, + 480, + 616 + ], + "type": "inline_equation", + "content": "4.53\\%" + }, + { + "bbox": [ + 130, + 462, + 480, + 616 + ], + "type": "text", + "content": ". The main reason is that our game-theoretic-based approach can learn the local alignment between words and frames for accurately predicting the moment boundaries. Besides, the Shapley variant performs similar to the Banzhaf variant for this specific WS-VTG task, since they share the same game interaction formulation with different normalization. Overall, this table verifies the effectiveness of our game-based framework." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 618, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 618, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 618, + 481, + 665 + ], + "type": "text", + "content": "Generalization on cross-dataset evaluation. At last, we explore the generalization of models by cross-dataset evaluation in Tab. 2. Concretely, we first train a model on one dataset and then evaluate its performance on the other dataset. We can find that: 1) For all WS-VTG methods, the performance of" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 252, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 252, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 252, + 91, + 447, + 102 + ], + "type": "text", + "content": "Rethinking WS-VTG From a Game Perspective" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 138, + 137, + 482, + 342 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 481, + 136 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 481, + 136 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 481, + 136 + ], + "type": "text", + "content": "Table 1: Comparisons with state-of-the-art methods. \"WS\" and \"FS\" denotes weakly-supervised and fully-supervised methods, respectively." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 138, + 137, + 482, + 342 + ], + "lines": [ + { + "bbox": [ + 138, + 137, + 482, + 342 + ], + "spans": [ + { + "bbox": [ + 138, + 137, + 482, + 342 + ], + "type": "table", + "html": "
TypeMethodCharades-STAActivityNet Caption
R@1, IoU=0.5R@1, IoU=0.7R@5, IoU=0.5R@5, IoU=0.7R@1, IoU=0.3R@1, IoU=0.5R@5, IoU=0.3R@5, IoU=0.5
FSCTRL [32]23.638.8958.9229.52-29.01-59.17
2DTAN [121]39.8123.2579.3351.1559.4544.5185.5377.13
DRN [115]53.0931.7589.0660.05-45.45-77.97
WSSCN [49]23.589.9771.8038.8747.2329.2271.4555.69
WSTAN [104]29.3512.2876.1341.5352.4530.0179.3863.42
ICVC [7]31.0216.5377.5341.9146.6229.5280.9266.61
MARN [87]33.8715.5473.9041.9448.5231.3775.9160.00
CRM [38]34.7616.37--55.2632.19--
CNM [124]35.1514.95--55.6833.33--
ACN [109]37.0215.26--57.6634.18--
VCA [105]38.1319.5778.7537.7550.4531.0071.7953.83
LCNet [112]39.1918.1780.5645.2448.4926.3382.5162.66
CPL [125]49.2422.3984.7152.3755.7331.3763.0543.13
DM2 [70]51.3923.72--54.7331.85--
Ours(Banzhaf)54.1726.4688.9355.7958.9435.6586.4170.26
Ours(Shapley)54.3126.2889.2455.6259.1735.8086.5970.03
", + "image_path": "6a1648e71c33565161b93736f84c0a88b6eca6acfe597373e3d60ab363258d57.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 137, + 394, + 306, + 466 + ], + "blocks": [ + { + "bbox": [ + 132, + 350, + 301, + 392 + ], + "lines": [ + { + "bbox": [ + 132, + 350, + 301, + 392 + ], + "spans": [ + { + "bbox": [ + 132, + 350, + 301, + 392 + ], + "type": "text", + "content": "Table 2: Cross-dataset evaluation. " + }, + { + "bbox": [ + 132, + 350, + 301, + 392 + ], + "type": "inline_equation", + "content": "A\\rightarrow C" + }, + { + "bbox": [ + 132, + 350, + 301, + 392 + ], + "type": "text", + "content": " denotes training on ActivityNet Caption and evaluation on CharadesSTA, and vice versa." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 137, + 394, + 306, + 466 + ], + "lines": [ + { + "bbox": [ + 137, + 394, + 306, + 466 + ], + "spans": [ + { + "bbox": [ + 137, + 394, + 306, + 466 + ], + "type": "table", + "html": "
MethodA→CC→A
R@1, IoU=0.7R@5, IoU=0.7R@1, IoU=0.5R@5, IoU=0.5
LCNet [112]12.0631.4710.6531.76
CPL [125]14.3235.8111.5929.88
Ours(Banzhaf)18.7442.5820.3742.96
Ours(Shapley)18.2643.1320.0441.42
", + "image_path": "ca98ec7bf1b6aa9e824541fc6d73e533f6209cdb8ece23b2943a85625af102c9.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 312, + 344, + 480, + 431 + ], + "blocks": [ + { + "bbox": [ + 312, + 344, + 480, + 431 + ], + "lines": [ + { + "bbox": [ + 312, + 344, + 480, + 431 + ], + "spans": [ + { + "bbox": [ + 312, + 344, + 480, + 431 + ], + "type": "image", + "image_path": "096937f3d5d36a18f2cabc2e31684e3bba05139f06634375f7d8d5d49b67561b.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 311, + 435, + 480, + 466 + ], + "lines": [ + { + "bbox": [ + 311, + 435, + 480, + 466 + ], + "spans": [ + { + "bbox": [ + 311, + 435, + 480, + 466 + ], + "type": "text", + "content": "Fig. 3: Qualitative results. The colors of green lines represent different degrees of confidence." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 473, + 481, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 473, + 481, + 605 + ], + "spans": [ + { + "bbox": [ + 130, + 473, + 481, + 605 + ], + "type": "inline_equation", + "content": "\\mathrm{A} \\rightarrow \\mathrm{C}" + }, + { + "bbox": [ + 130, + 473, + 481, + 605 + ], + "type": "text", + "content": " achieves better than " + }, + { + "bbox": [ + 130, + 473, + 481, + 605 + ], + "type": "inline_equation", + "content": "\\mathrm{C} \\rightarrow \\mathrm{A}" + }, + { + "bbox": [ + 130, + 473, + 481, + 605 + ], + "type": "text", + "content": ". This is because that ActivityNet Caption is larger and more complex than Charades-STA, thus the former is able to bring more generalized knowledge to the latter while the latter fails to provide general knowledge to the former. 2) Our two variants outperform state-of-the-art methods (e.g., LCNet) with clear margins on cross-dataset evaluation. We speculate it to: Previous works severely rely on the moment proposals, however, LCNet defines different sliding windows on different datasets for proposal generation. Therefore, its non-uniform proposal structure limits its generalization on cross-dataset evaluation. Instead, we learn multi-level frame-aware alignment and utilize frame-wise scores for predicting moment, which is more flexible and adaptable, thus improving the generalization-ability." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 605, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 481, + 665 + ], + "type": "text", + "content": "Qualitative comparison. In Fig. 3, we provide a qualitative example on ActivityNet Caption. The colors of green lines represent different degrees of interaction confidence, demonstrating that the game interaction can well learn the uncertain and fine-grained frame-word correspondence in the WS setting. Besides, all three-level interaction (word-level, phrase-level and sentence-level) contributes" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 304, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 304, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 304, + 102 + ], + "type": "text", + "content": "X. Fang, Z. Xiong, W. Fang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 145, + 126, + 466, + 184 + ], + "blocks": [ + { + "bbox": [ + 149, + 114, + 463, + 125 + ], + "lines": [ + { + "bbox": [ + 149, + 114, + 463, + 125 + ], + "spans": [ + { + "bbox": [ + 149, + 114, + 463, + 125 + ], + "type": "text", + "content": "Table 3: Comparison on speed per sample (s) and GPU memory cost (MB)." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 145, + 126, + 466, + 184 + ], + "lines": [ + { + "bbox": [ + 145, + 126, + 466, + 184 + ], + "spans": [ + { + "bbox": [ + 145, + 126, + 466, + 184 + ], + "type": "table", + "html": "
MetricStageLCNet [112]CPL [125]Ours(Banzhaf)Ours(Shapley)
SpeedTraining0.79s0.27s0.94s0.98s
Inference0.62s0.21s0.10s0.11s
GPUTraining8647MB4156MB5237MB5482MB
MemoryInference7462MB3428MB2058MB2165MB
", + "image_path": "397665a7d8b62e35860f44918d9b05ad2294d4044d5cb695835f5e61819416e6.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 135, + 203, + 477, + 294 + ], + "blocks": [ + { + "bbox": [ + 237, + 190, + 376, + 201 + ], + "lines": [ + { + "bbox": [ + 237, + 190, + 376, + 201 + ], + "spans": [ + { + "bbox": [ + 237, + 190, + 376, + 201 + ], + "type": "text", + "content": "Table 4: Main ablation study." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 135, + 203, + 477, + 294 + ], + "lines": [ + { + "bbox": [ + 135, + 203, + 477, + 294 + ], + "spans": [ + { + "bbox": [ + 135, + 203, + 477, + 294 + ], + "type": "table", + "html": "
Self-modal GameMulti-level Cross-modal GameBanzhaf-based GameShapley-based Game
Charades-STAActivityNet CaptionCharades-STAActivityNet Caption
Video Query GameWord- Phrase- Sentence- levelIoU=0.7IoU=0.5IoU=0.5R@1,R@5,R@1,R@5,R@1,R@5,R@1,R@5,
IoU=0.7IoU=0.5IoU=0.5IoU=0.5IoU=0.7IoU=0.5IoU=0.5IoU=0.5
××✓××19.3444.6729.8361.7219.2944.3130.1561.56
✓×✓××21.2547.8631.3963.9821.1047.7331.5463.67
✓✓✓××22.9750.6432.7265.8022.6950.3832.9765.65
✓✓✓✓×24.8153.2834.4668.0324.6453.0334.5967.82
✓✓✓✓✓26.4655.7935.6570.2626.2855.6235.8070.03
", + "image_path": "e3f8afea7878f290bed22e84fbcbe5ad9ed6432213b2d4a281d798c7658f0133.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 302, + 480, + 325 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 302, + 480, + 325 + ], + "spans": [ + { + "bbox": [ + 130, + 302, + 480, + 325 + ], + "type": "text", + "content": "differently to the grounding, and our final method is able to localize more accurate moment boundaries than state-of-the-art method CPL." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 327, + 494, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 327, + 494, + 422 + ], + "spans": [ + { + "bbox": [ + 130, + 327, + 494, + 422 + ], + "type": "text", + "content": "Complexity comparison. In Tab. 3, our method costs more training time/memory as we need to enumerate coalition samples. Specifically, we sampled 5500 samples. We find that, the model instability decreases along with the increase of the sampling number, and when the sample size ranges from 4500 to 5500, the error with the true value of banzhaf gradually decreases from " + }, + { + "bbox": [ + 130, + 327, + 494, + 422 + ], + "type": "inline_equation", + "content": "6.9\\%" + }, + { + "bbox": [ + 130, + 327, + 494, + 422 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 130, + 327, + 494, + 422 + ], + "type": "inline_equation", + "content": "1.1\\%" + }, + { + "bbox": [ + 130, + 327, + 494, + 422 + ], + "type": "text", + "content": ". After exceeding 5500 samples, more samples will not affect the result. During inference, since we can directly utilize the learned game headers to generate game value instead of using complex game learning, our model is faster with lower memory." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 441, + 238, + 453 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 441, + 238, + 453 + ], + "spans": [ + { + "bbox": [ + 132, + 441, + 238, + 453 + ], + "type": "text", + "content": "4.4 Ablation Study" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 461, + 481, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 461, + 481, + 485 + ], + "spans": [ + { + "bbox": [ + 130, + 461, + 481, + 485 + ], + "type": "text", + "content": "In this section, we perform in-depth ablation studies to evaluate the effect of each component. We implement our model with different game settings." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 486, + 481, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 486, + 481, + 616 + ], + "spans": [ + { + "bbox": [ + 130, + 486, + 481, + 616 + ], + "type": "text", + "content": "Main ablation. As shown in Tab. 4, we conduct ablation studies regarding the components (i.e., self-modal games in two modalities, and multi-level cross-modal games). We observe the following findings: 1) Our baseline model (only contains word-level cross-modal game) is powerful and outperforms some WS methods, demonstrating that it is effective to capture the uncertain yet fine-grained alignment for better determining the boundaries. 2) Both video and query self-modal games bring significant improvement to the overall model, indicating that self-enhancement is crucial for context integration. 3) Each query level of cross-modal game further boost the performance, verifying that phrase-and sentence-level semantics help to better understand temporal events. Overall, all components contribute a lot to final performance, showing their effectiveness." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 617, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 481, + 665 + ], + "type": "text", + "content": "Effect of game headers. The headers in both self- and cross-modal games are utilized to provide soft supervision for better game interaction and player-representation learning. To explore the impact of the structure of these headers, we compare several popular structures in Tab. 5. We find that the combination of" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 252, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 252, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 252, + 91, + 447, + 102 + ], + "type": "text", + "content": "Rethinking WS-VTG From a Game Perspective" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 134, + 158, + 482, + 317 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 480, + 157 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 480, + 157 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 480, + 157 + ], + "type": "text", + "content": "Table 5: Ablation study on each component, where \"LL\" means \"Linear Layer\", \"CL\" means \"Convolution Layer\", \"SA\" means \"Self-attention\", \"MSE\" means \"Mean Squared Error\", \"CE\" means \"Cross Entropy\", \"KLD\" means \"Kullback-Leibler Divergence\"; \"RL\" means \"Reconstruction Loss\"." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 134, + 158, + 482, + 317 + ], + "lines": [ + { + "bbox": [ + 134, + 158, + 482, + 317 + ], + "spans": [ + { + "bbox": [ + 134, + 158, + 482, + 317 + ], + "type": "table", + "html": "
ComponentSettingBanzhaf-based GameShapley-based Game
Charades-STAActivityNet CaptionCharades-STAActivityNet Caption
R@1, IoU=0.7R@5, IoU=0.7R@1, IoU=0.5R@5, IoU=0.5R@1, IoU=0.7R@5, IoU=0.7R@1, IoU=0.5R@5, IoU=0.5
Game HeadersLL25.2853.9034.1568.0324.9753.6834.2567.94
CL25.5254.3134.5668.6425.2654.0334.8268.47
LL + SA25.9454.8335.2769.1825.7154.5935.5369.06
CL + SA26.4655.7935.6570.2626.2855.6235.8070.03
Soft SupervisionNone23.2151.6432.3666.4022.9951.4532.7166.18
MSE24.4253.1933.8567.5724.2452.8940.0667.33
CE24.7353.5834.1668.0424.5253.2634.3767.80
KLD26.4655.7935.6570.2626.2855.6235.8070.03
Grounding Lossesw/o RL24.3553.5734.1468.4224.0853.3134.4568.19
w/ RL26.4655.7935.6570.2626.2855.6235.8070.03
", + "image_path": "7bcddf95c78911d5c6e75ceece003030b116e017d8ede6f10448b3b571fe9576.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 331, + 480, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 331, + 480, + 365 + ], + "spans": [ + { + "bbox": [ + 132, + 331, + 480, + 365 + ], + "type": "text", + "content": "convolution layer and self-attention can capture both local and global interaction, so it achieves the best performance and is beneficial for cross-modal multi-level interaction, illustrating the significance of our game headers." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 368, + 480, + 438 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 368, + 480, + 438 + ], + "spans": [ + { + "bbox": [ + 132, + 368, + 480, + 438 + ], + "type": "text", + "content": "Necessity of soft supervision. To explore the effect of soft supervision, we conduct the ablation study to investigate the contribution of the KLD loss during the header learning in Tab. 5. The KLD loss achieves better performance than both MSE and CE, showing that it is more suitable for game headers to mimic game-theoretic interaction learning. Moreover, without soft supervision, the performance will drop a lot, showing the importance of the soft supervision." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 441, + 480, + 499 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 441, + 480, + 499 + ], + "spans": [ + { + "bbox": [ + 132, + 441, + 480, + 499 + ], + "type": "text", + "content": "Analysis on reconstruction loss. Most previous WS-VTG works also additionally utilize the reconstruction paradigm to assist the model training. As shown in Tab. 5, we further conduct such ablation study to investigate the impact of reconstruction loss. It shows that the reconstruction loss can improve significantly the performance, demonstrating its effectiveness." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 133, + 525, + 218, + 538 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 525, + 218, + 538 + ], + "spans": [ + { + "bbox": [ + 133, + 525, + 218, + 538 + ], + "type": "text", + "content": "5 Conclusion" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 557, + 480, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 557, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 557, + 480, + 665 + ], + "type": "text", + "content": "In this paper, we rethink the limitations of previous weakly-supervised video temporal grounding works, and creatively propose to utilize the game-theoretic interaction to learn the uncertain relationship between video-query pairs with diverse granularity. Specifically, we first introduce two self-modal games to correlate the frames/words to enhance their contextual semantics. Then, we propose a cross-modal game to value fine-grained correspondence between frames and words with multiple levels for better determining the query-related moment boundaries. Extensive experiments demonstrate the effectiveness of our game-based framework on two challenging datasets." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 304, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 304, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 304, + 102 + ], + "type": "text", + "content": "X. Fang, Z. Xiong, W. Fang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 197, + 126 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 197, + 126 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 197, + 126 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 137, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 142, + 137, + 480, + 169 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 142, + 137, + 480, + 169 + ], + "spans": [ + { + "bbox": [ + 142, + 137, + 480, + 169 + ], + "type": "text", + "content": "1. Albarelli, A., Rodola, E., Torsello, A.: A game-theoretic approach to fine surface registration without initial motion estimation. In: 2010 IEEE computer society conference on computer vision and pattern recognition. pp. 430-437. IEEE (2010)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 142, + 170, + 480, + 201 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 142, + 170, + 480, + 201 + ], + "spans": [ + { + "bbox": [ + 142, + 170, + 480, + 201 + ], + "type": "text", + "content": "2. Anne Hendricks, L., Wang, O., Shechtman, E., Sivic, J., Darrell, T., Russell, B.: Localizing moments in video with natural language. In: Proceedings of the IEEE International Conference on Computer Vision. pp. 5803-5812 (2017)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 142, + 202, + 480, + 234 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 142, + 202, + 480, + 234 + ], + "spans": [ + { + "bbox": [ + 142, + 202, + 480, + 234 + ], + "type": "text", + "content": "3. Bachrach, Y., Markakis, E., Resnick, E., Procaccia, A.D., Rosenschein, J.S., Saberi, A.: Approximating power indices: theoretical and empirical analysis. Autonomous Agents and Multi-Agent Systems 20, 105-122 (2010)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 142, + 235, + 480, + 255 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 142, + 235, + 480, + 255 + ], + "spans": [ + { + "bbox": [ + 142, + 235, + 480, + 255 + ], + "type": "text", + "content": "4. Banzhaf III, J.F.: Weighted voting doesn't work: A mathematical analysis. Rutgers L. Rev. 19, 317 (1964)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 142, + 256, + 480, + 287 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 142, + 256, + 480, + 287 + ], + "spans": [ + { + "bbox": [ + 142, + 256, + 480, + 287 + ], + "type": "text", + "content": "5. Carreira, J., Zisserman, A.: Quo vadis, action recognition? a new model and the kinetics dataset. In: proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 6299-6308 (2017)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 142, + 288, + 480, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 142, + 288, + 480, + 319 + ], + "spans": [ + { + "bbox": [ + 142, + 288, + 480, + 319 + ], + "type": "text", + "content": "6. Chalkiadakis, G., Elkind, E., Wooldridge, M.: Computational aspects of cooperative game theory. Synthesis Lectures on Artificial Intelligence and Machine Learning 5(6), 1-168 (2011)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 142, + 320, + 480, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 142, + 320, + 480, + 352 + ], + "spans": [ + { + "bbox": [ + 142, + 320, + 480, + 352 + ], + "type": "text", + "content": "7. Chen, J., Luo, W., Zhang, W., Ma, L.: Explore inter-contrast between videos via composition for weakly supervised temporal sentence grounding. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 36, pp. 267-275 (2022)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 142, + 353, + 480, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 142, + 353, + 480, + 384 + ], + "spans": [ + { + "bbox": [ + 142, + 353, + 480, + 384 + ], + "type": "text", + "content": "8. Chen, J., Ma, L., Chen, X., Jie, Z., Luo, J.: Localizing natural language in videos. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 33, pp. 8175-8182 (2019)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 142, + 384, + 480, + 417 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 142, + 384, + 480, + 417 + ], + "spans": [ + { + "bbox": [ + 142, + 384, + 480, + 417 + ], + "type": "text", + "content": "9. Chen, L., Lu, C., Tang, S., Xiao, J., Zhang, D., Tan, C., Li, X.: Rethinking the bottom-up framework for query-based video localization. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 34, pp. 10551-10558 (2020)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 418, + 480, + 449 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 418, + 480, + 449 + ], + "spans": [ + { + "bbox": [ + 138, + 418, + 480, + 449 + ], + "type": "text", + "content": "10. Chen, Z., Ma, L., Luo, W., Tang, P., Wong, K.Y.K.: Look closer to ground better: Weakly-supervised temporal grounding of sentence in video. arXiv preprint arXiv:2001.09308 (2020)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 449, + 480, + 482 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 449, + 480, + 482 + ], + "spans": [ + { + "bbox": [ + 138, + 449, + 480, + 482 + ], + "type": "text", + "content": "11. Datta, A., Sen, S., Zick, Y.: Algorithmic transparency via quantitative input influence: Theory and experiments with learning systems. In: 2016 IEEE symposium on security and privacy. pp. 598-617. IEEE (2016)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 483, + 480, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 483, + 480, + 514 + ], + "spans": [ + { + "bbox": [ + 138, + 483, + 480, + 514 + ], + "type": "text", + "content": "12. Deng, S., Wen, J., Liu, C., Yan, K., Xu, G., Xu, Y.: Projective incomplete multiview clustering. IEEE Transactions on Neural Networks and Learning Systems pp. 1-13 (2023). https://doi.org/10.1109/TNNLS.2023.3242473" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 514, + 480, + 546 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 514, + 480, + 546 + ], + "spans": [ + { + "bbox": [ + 138, + 514, + 480, + 546 + ], + "type": "text", + "content": "13. Dong, J., Chen, X., Zhang, M., Yang, X., Chen, S., Li, X., Wang, X.: Partially relevant video retrieval. In: Proceedings of the 30th ACM International Conference on Multimedia. pp. 246-257 (2022)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 547, + 480, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 547, + 480, + 578 + ], + "spans": [ + { + "bbox": [ + 138, + 547, + 480, + 578 + ], + "type": "text", + "content": "14. Dong, J., Li, X., Snoek, C.G.: Predicting visual features from text for image and video caption retrieval. IEEE Transactions on Multimedia 20(12), 3377-3388 (2018)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 578, + 480, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 578, + 480, + 611 + ], + "spans": [ + { + "bbox": [ + 138, + 578, + 480, + 611 + ], + "type": "text", + "content": "15. Dong, J., Li, X., Xu, C., Yang, X., Yang, G., Wang, X., Wang, M.: Dual encoding for video retrieval by text. IEEE Transactions on Pattern Analysis and Machine Intelligence 44(8), 4065-4080 (2022)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 611, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 611, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 611, + 480, + 665 + ], + "type": "text", + "content": "16. Dong, J., Peng, X., Ma, Z., Liu, D., Qu, X., Yang, X., Zhu, J., Liu, B.: From region to patch: Attribute-aware foreground-background contrastive learning for fine-grained fashion retrieval. In: Proceedings of the 46th International ACM SIGIR Conference on Research and Development in Information Retrieval. pp. 1273-1282 (2023)" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 252, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 252, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 252, + 91, + 447, + 102 + ], + "type": "text", + "content": "Rethinking WS-VTG From a Game Perspective" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 138, + 116, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 138, + 116, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 116, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 138, + 116, + 480, + 149 + ], + "type": "text", + "content": "17. Dong, J., Sun, S., Liu, Z., Chen, S., Liu, B., Wang, X.: Hierarchical contrast for unsupervised skeleton-based action representation learning. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 37, pp. 525-533 (2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 149, + 480, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 149, + 480, + 182 + ], + "spans": [ + { + "bbox": [ + 138, + 149, + 480, + 182 + ], + "type": "text", + "content": "18. Dong, J., Wang, Y., Chen, X., Qu, X., Li, X., He, Y., Wang, X.: Reading-strategy inspired visual representation learning for text-to-video retrieval. IEEE Transactions on Circuits and Systems for Video Technology 32(8), 5680-5694 (2022)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 182, + 480, + 214 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 182, + 480, + 214 + ], + "spans": [ + { + "bbox": [ + 138, + 182, + 480, + 214 + ], + "type": "text", + "content": "19. Donoser, M., Bischof, H.: Diffusion processes for retrieval revisited. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1320-1327 (2013)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 214, + 480, + 246 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 214, + 480, + 246 + ], + "spans": [ + { + "bbox": [ + 138, + 214, + 480, + 246 + ], + "type": "text", + "content": "20. Dowdall, J., Pavlidis, I.T., Tsiamiyrtzis, P.: Coalitional tracking in facial infrared imaging and beyond. In: 2006 Conference on Computer Vision and Pattern Recognition Workshop. pp. 134-134. IEEE (2006)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 246, + 480, + 278 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 246, + 480, + 278 + ], + "spans": [ + { + "bbox": [ + 138, + 246, + 480, + 278 + ], + "type": "text", + "content": "21. Fang, X., Easwaran, A., Genest, B.: Uncertainty-guided appearance-motion association network for out-of-distribution action detection. In: IEEE International Conference on Multimedia Information Processing and Retrieval (2024)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 278, + 480, + 320 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 278, + 480, + 320 + ], + "spans": [ + { + "bbox": [ + 138, + 278, + 480, + 320 + ], + "type": "text", + "content": "22. Fang, X., Fang, W., Liu, D., Qu, X., Dong, J., Zhou, P., Li, R., Xu, Z., Chen, L., Zheng, P., Cheng, Y.: Not all inputs are valid: Towards open-set video moment retrieval using language. In: Proceedings of the 32th ACM International Conference on Multimedia (2024)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 321, + 480, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 321, + 480, + 342 + ], + "spans": [ + { + "bbox": [ + 138, + 321, + 480, + 342 + ], + "type": "text", + "content": "23. Fang, X., Hu, Y.: Double self-weighted multi-view clustering via adaptive view fusion. arXiv preprint arXiv:2011.10396 (2020)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 342, + 480, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 342, + 480, + 374 + ], + "spans": [ + { + "bbox": [ + 138, + 342, + 480, + 374 + ], + "type": "text", + "content": "24. Fang, X., Hu, Y., Zhou, P., Wu, D.: Animc: A soft approach for autoweighted noisy and incomplete multiview clustering. IEEE Transactions on Artificial Intelligence 3(2), 192-206 (2021)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 374, + 480, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 374, + 480, + 407 + ], + "spans": [ + { + "bbox": [ + 138, + 374, + 480, + 407 + ], + "type": "text", + "content": "25. Fang, X., Hu, Y., Zhou, P., Wu, D.O.: V3h: View variation and view heredity for incomplete multiview clustering. IEEE Transactions on Artificial Intelligence 1(3), 233-247 (2020)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 407, + 480, + 449 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 407, + 480, + 449 + ], + "spans": [ + { + "bbox": [ + 138, + 407, + 480, + 449 + ], + "type": "text", + "content": "26. Fang, X., Hu, Y., Zhou, P., Wu, D.O.: Unbalanced incomplete multi-view clustering via the scheme of view evolution: Weak views are meat; strong views do eat. IEEE Transactions on Emerging Topics in Computational Intelligence 6(4), 913-927 (2021)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 449, + 480, + 493 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 449, + 480, + 493 + ], + "spans": [ + { + "bbox": [ + 138, + 449, + 480, + 493 + ], + "type": "text", + "content": "27. Fang, X., Liu, D., Fang, W., Zhou, P., Cheng, Y., Tang, K., Zou, K.: Annotations are not all you need: A cross-modal knowledge transfer network for unsupervised temporal sentence grounding. In: Findings of the Association for Computational Linguistics: EMNLP 2023. pp. 8721-8733 (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 493, + 480, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 493, + 480, + 536 + ], + "spans": [ + { + "bbox": [ + 138, + 493, + 480, + 536 + ], + "type": "text", + "content": "28. Fang, X., Liu, D., Fang, W., Zhou, P., Xu, Z., Xu, W., Chen, J., Li, R.: Fewer steps, better performance: Efficient cross-modal clip trimming for video moment retrieval using language. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 38, pp. 1735-1743 (2024)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 536, + 480, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 536, + 480, + 568 + ], + "spans": [ + { + "bbox": [ + 138, + 536, + 480, + 568 + ], + "type": "text", + "content": "29. Fang, X., Liu, D., Zhou, P., Hu, Y.: Multi-modal cross-domain alignment network for video moment retrieval. IEEE Transactions on Multimedia 25, 7517-7532 (2022)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 568, + 480, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 568, + 480, + 611 + ], + "spans": [ + { + "bbox": [ + 138, + 568, + 480, + 611 + ], + "type": "text", + "content": "30. Fang, X., Liu, D., Zhou, P., Nan, G.: You can ground earlier than see: An effective and efficient pipeline for temporal sentence grounding in compressed videos. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2448-2460 (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 611, + 480, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 611, + 480, + 633 + ], + "spans": [ + { + "bbox": [ + 138, + 611, + 480, + 633 + ], + "type": "text", + "content": "31. Fang, X., Liu, D., Zhou, P., Xu, Z., Li, R.: Hierarchical local-global transformer for temporal sentence grounding. IEEE Transactions on Multimedia (2023)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 633, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 633, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 633, + 480, + 665 + ], + "type": "text", + "content": "32. Gao, J., Sun, C., Yang, Z., Nevatia, R.: Tall: Temporal activity localization via language query. In: Proceedings of the IEEE International Conference on Computer Vision. pp. 5267-5275 (2017)" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 165, + 91, + 304, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 165, + 91, + 304, + 102 + ], + "spans": [ + { + "bbox": [ + 165, + 91, + 304, + 102 + ], + "type": "text", + "content": "X. Fang, Z. Xiong, W. Fang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 138, + 116, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 138, + 116, + 480, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 116, + 480, + 161 + ], + "spans": [ + { + "bbox": [ + 138, + 116, + 480, + 161 + ], + "type": "text", + "content": "33. Gao, M., Davis, L., Socher, R., Xiong, C.: Wslln: Weakly supervised natural language localization networks. In: Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing. pp. 1481-1487 (2019)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 162, + 480, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 162, + 480, + 194 + ], + "spans": [ + { + "bbox": [ + 138, + 162, + 480, + 194 + ], + "type": "text", + "content": "34. Grabisch, M., Roubens, M.: An axiomatic approach to the concept of interaction among players in cooperative games. International Journal of game theory 28, 547-565 (1999)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 194, + 480, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 194, + 480, + 226 + ], + "spans": [ + { + "bbox": [ + 138, + 194, + 480, + 226 + ], + "type": "text", + "content": "35. Guo, C., Liu, D., Zhou, P.: A hybrid alignment loss for temporal moment localization with natural language. In: 2022 IEEE International Conference on Multimedia and Expo (ICME). pp. 1-6. IEEE (2022)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 227, + 480, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 227, + 480, + 259 + ], + "spans": [ + { + "bbox": [ + 138, + 227, + 480, + 259 + ], + "type": "text", + "content": "36. Guo, D., Li, K., Hu, B., Zhang, Y., Wang, M.: Benchmarking micro-action recognition: Dataset, method, and application. IEEE Transactions on Circuits and Systems for Video Technology 34(7), 6238-6252 (2024)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 259, + 480, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 259, + 480, + 303 + ], + "spans": [ + { + "bbox": [ + 138, + 259, + 480, + 303 + ], + "type": "text", + "content": "37. Hendricks, L.A., Wang, O., Shechtman, E., Sivic, J., Darrell, T., Russell, B.: Localizing moments in video with temporal language. Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing p. 1380-1390 (2018)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 304, + 480, + 336 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 304, + 480, + 336 + ], + "spans": [ + { + "bbox": [ + 138, + 304, + 480, + 336 + ], + "type": "text", + "content": "38. Huang, J., Liu, Y., Gong, S., Jin, H.: Cross-sentence temporal and semantic relations in video activity localisation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 7199-7208 (2021)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 336, + 480, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 336, + 480, + 380 + ], + "spans": [ + { + "bbox": [ + 138, + 336, + 480, + 380 + ], + "type": "text", + "content": "39. Jiang, L., Wang, C., Ning, X., Yu, Z.: Lttpoint: A mlp-based point cloud classification method with local topology transformation module. In: 2023 7th Asian Conference on Artificial Intelligence Technology (ACAIT). pp. 783-789. IEEE (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 380, + 480, + 424 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 380, + 480, + 424 + ], + "spans": [ + { + "bbox": [ + 138, + 380, + 480, + 424 + ], + "type": "text", + "content": "40. Jin, P., Huang, J., Xiong, P., Tian, S., Liu, C., Ji, X., Yuan, L., Chen, J.: Videotext as game players: Hierarchical banzhaf interaction for cross-modal representation learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2472-2482 (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 424, + 480, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 424, + 480, + 456 + ], + "spans": [ + { + "bbox": [ + 138, + 424, + 480, + 456 + ], + "type": "text", + "content": "41. Jin, S., Wang, S., Fang, F.: Game theoretical analysis on capacity configuration for microgrid based on multi-agent system. International Journal of Electrical Power & Energy Systems 125, 106485 (2021)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 456, + 480, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 456, + 480, + 479 + ], + "spans": [ + { + "bbox": [ + 138, + 456, + 480, + 479 + ], + "type": "text", + "content": "42. Kingma, D.P., Ba, J.: Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980 (2014)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 479, + 480, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 479, + 480, + 511 + ], + "spans": [ + { + "bbox": [ + 138, + 479, + 480, + 511 + ], + "type": "text", + "content": "43. Krishna, R., Hata, K., Ren, F., Fei-Fei, L., Carlos Niebles, J.: Dense-c captioning events in videos. In: Proceedings of the IEEE International Conference on Computer Vision. pp. 706-715 (2017)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 512, + 349, + 522 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 512, + 349, + 522 + ], + "spans": [ + { + "bbox": [ + 138, + 512, + 349, + 522 + ], + "type": "text", + "content": "44. Leech, D.: Computation of power indices (2002)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 523, + 480, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 523, + 480, + 544 + ], + "spans": [ + { + "bbox": [ + 138, + 523, + 480, + 544 + ], + "type": "text", + "content": "45. Lehrer, E.: An axiomatization of the banzhaf value. International Journal of Game Theory 17, 89-99 (1988)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 545, + 480, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 545, + 480, + 588 + ], + "spans": [ + { + "bbox": [ + 138, + 545, + 480, + 588 + ], + "type": "text", + "content": "46. Li, H., Cao, M., Cheng, X., Li, Y., Zhu, Z., Zou, Y.: G2l: Semantically aligned and uniform video grounding via geodesic and game theory. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 12032-12042 (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 589, + 480, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 589, + 480, + 621 + ], + "spans": [ + { + "bbox": [ + 138, + 589, + 480, + 621 + ], + "type": "text", + "content": "47. Li, J., HE, X., Wei, L., Qian, L., Zhu, L., Xie, L., Zhuang, Y., Tian, Q., Tang, S.: Fine-grained semantically aligned vision-language pre-training. In: Advances in Neural Information Processing Systems (2022)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 622, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 622, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 622, + 480, + 665 + ], + "type": "text", + "content": "48. Lin, K.Q., Zhang, P., Chen, J., Pramanick, S., Gao, D., Wang, A.J., Yan, R., Shou, M.Z.: Univtg: Towards unified video-language temporal grounding. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 2794-2804 (2023)" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 252, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 252, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 252, + 91, + 447, + 102 + ], + "type": "text", + "content": "Rethinking WS-VTG From a Game Perspective" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 138, + 116, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 138, + 116, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 116, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 138, + 116, + 480, + 149 + ], + "type": "text", + "content": "49. Lin, Z., Zhao, Z., Zhang, Z., Wang, Q., Liu, H.: Weakly-supervised video moment retrieval via semantic completion network. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 34, pp. 11539-11546 (2020)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 151, + 480, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 151, + 480, + 194 + ], + "spans": [ + { + "bbox": [ + 138, + 151, + 480, + 194 + ], + "type": "text", + "content": "50. Liu, C., Wen, J., Luo, X., Huang, C., Wu, Z., Xu, Y.: Dicnet: Deep instance-level contrastive network for double incomplete multi-view multi-label classification. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 37, pp. 8807-8815 (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 195, + 480, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 195, + 480, + 228 + ], + "spans": [ + { + "bbox": [ + 138, + 195, + 480, + 228 + ], + "type": "text", + "content": "51. Liu, C., Wen, J., Luo, X., Xu, Y.: Incomplete multi-view multi-label learning via label-guided masked view- and category-aware transformers. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 37, pp. 8816-8824 (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 229, + 480, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 229, + 480, + 262 + ], + "spans": [ + { + "bbox": [ + 138, + 229, + 480, + 262 + ], + "type": "text", + "content": "52. Liu, C., Wen, J., Wu, Z., Luo, X., Huang, C., Xu, Y.: Information recovery-driven deep incomplete multiview clustering network. IEEE Transactions on Neural Networks and Learning Systems pp. 1-11 (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 263, + 480, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 263, + 480, + 296 + ], + "spans": [ + { + "bbox": [ + 138, + 263, + 480, + 296 + ], + "type": "text", + "content": "53. Liu, D., Fang, X., Hu, W., Zhou, P.: Exploring optical-flow-guided motion and detection-based appearance for temporal sentence grounding. IEEE Transactions on Multimedia 25, 8539-8553 (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 297, + 480, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 297, + 480, + 341 + ], + "spans": [ + { + "bbox": [ + 138, + 297, + 480, + 341 + ], + "type": "text", + "content": "54. Liu, D., Fang, X., Qu, X., Dong, J., Yan, H., Yang, Y., Zhou, P., Cheng, Y.: Unsupervised domain adaptative temporal sentence localization with mutual information maximization. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 38, pp. 3567-3575 (2024)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 342, + 480, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 342, + 480, + 373 + ], + "spans": [ + { + "bbox": [ + 138, + 342, + 480, + 373 + ], + "type": "text", + "content": "55. Liu, D., Fang, X., Zhou, P., Di, X., Lu, W., Cheng, Y.: Hypotheses tree building for one-shot temporal sentence localization. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 37, pp. 1640-1648 (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 375, + 480, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 375, + 480, + 407 + ], + "spans": [ + { + "bbox": [ + 138, + 375, + 480, + 407 + ], + "type": "text", + "content": "56. Liu, D., Hu, W.: Learning to focus on the foreground for temporal sentence grounding. In: Proceedings of the 29th International Conference on Computational Linguistics. pp. 5532-5541 (2022)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 409, + 480, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 409, + 480, + 441 + ], + "spans": [ + { + "bbox": [ + 138, + 409, + 480, + 441 + ], + "type": "text", + "content": "57. Liu, D., Hu, W.: Skimming, locating, then perusing: A human-like framework for natural language video localization. In: Proceedings of the 30th ACM International Conference on Multimedia. pp. 4536-4545 (2022)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 443, + 480, + 485 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 443, + 480, + 485 + ], + "spans": [ + { + "bbox": [ + 138, + 443, + 480, + 485 + ], + "type": "text", + "content": "58. Liu, D., Qu, X., Dong, J., Nan, G., Zhou, P., Xu, Z., Chen, L., Yan, H., Cheng, Y.: Filling the information gap between video and query for language-driven moment retrieval. In: Proceedings of the 31st ACM International Conference on Multimedia. pp. 4190-4199 (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 487, + 480, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 487, + 480, + 531 + ], + "spans": [ + { + "bbox": [ + 138, + 487, + 480, + 531 + ], + "type": "text", + "content": "59. Liu, D., Qu, X., Dong, J., Zhou, P., Cheng, Y., Wei, W., Xu, Z., Xie, Y.: Context-aware biaffine localizing network for temporal sentence grounding. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 11235-11244 (2021)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 532, + 480, + 575 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 532, + 480, + 575 + ], + "spans": [ + { + "bbox": [ + 138, + 532, + 480, + 575 + ], + "type": "text", + "content": "60. Liu, D., Qu, X., Dong, J., Zhou, P., Xu, Z., Wang, H., Di, X., Lu, W., Cheng, Y.: Transform-equivariant consistency learning for temporal sentence grounding. ACM Transactions on Multimedia Computing, Communications and Applications 20(4), 1-19 (2024)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 576, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 576, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 138, + 576, + 480, + 632 + ], + "type": "text", + "content": "61. Liu, D., Qu, X., Fang, X., Dong, J., Zhou, P., Nan, G., Tang, K., Fang, W., Cheng, Y.: Towards robust temporal activity localization learning with noisy labels. In: Proceedings of the 2024 Joint International Conference on Computational Linguistics, Language Resources and Evaluation (LREC-COLING 2024). pp. 16630-16642 (2024)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 632, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 632, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 632, + 480, + 665 + ], + "type": "text", + "content": "62. Liu, D., Qu, X., Hu, W.: Reducing the vision and language bias for temporal sentence grounding. In: Proceedings of the 30th ACM International Conference on Multimedia. pp. 4092-4101 (2022)" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 304, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 304, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 304, + 102 + ], + "type": "text", + "content": "X. Fang, Z. Xiong, W. Fang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + }, + { + "para_blocks": [ + { + "bbox": [ + 138, + 116, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 138, + 116, + 481, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 116, + 481, + 149 + ], + "spans": [ + { + "bbox": [ + 138, + 116, + 481, + 149 + ], + "type": "text", + "content": "63. Liu, D., Qu, X., Liu, X.Y., Dong, J., Zhou, P., Xu, Z.: Jointly cross-and self-modal graph attention network for query-based moment localization. In: Proceedings of the 28th ACM International Conference on Multimedia. pp. 4070-4078 (2020)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 151, + 481, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 151, + 481, + 194 + ], + "spans": [ + { + "bbox": [ + 138, + 151, + 481, + 194 + ], + "type": "text", + "content": "64. Liu, D., Zhou, P.: Jointly visual-and semantic-aware graph memory networks for temporal sentence localization in videos. In: ICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). pp. 1-5. IEEE (2023)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 195, + 481, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 195, + 481, + 228 + ], + "spans": [ + { + "bbox": [ + 138, + 195, + 481, + 228 + ], + "type": "text", + "content": "65. Liu, D., Zhou, P., Xu, Z., Wang, H., Li, R.: Few-shot temporal sentence grounding via memory-guided semantic learning. IEEE Transactions on Circuits and Systems for Video Technology 33(5), 2491-2505 (2022)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 228, + 481, + 261 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 228, + 481, + 261 + ], + "spans": [ + { + "bbox": [ + 138, + 228, + 481, + 261 + ], + "type": "text", + "content": "66. Liu, D., Zhu, J., Fang, X., Xiong, Z., Wang, H., Li, R., Zhou, P.: Conditional video diffusion network for fine-grained temporal sentence grounding. IEEE Transactions on Multimedia (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 262, + 481, + 295 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 262, + 481, + 295 + ], + "spans": [ + { + "bbox": [ + 138, + 262, + 481, + 295 + ], + "type": "text", + "content": "67. Lundberg, S.M., Lee, S.I.: A unified approach to interpreting model predictions. In: Proceedings of the 31st International Conference on Neural Information Processing Systems. pp. 4768-4777 (2017)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 296, + 481, + 328 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 296, + 481, + 328 + ], + "spans": [ + { + "bbox": [ + 138, + 296, + 481, + 328 + ], + "type": "text", + "content": "68. Ma, M., Yoon, S., Kim, J., Lee, Y., Kang, S., Yoo, C.D.: VLANet: Video-language alignment network for weakly-supervised video moment retrieval. In: Proceedings of the European Conference on Computer Vision. pp. 156-171 (2020)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 330, + 481, + 361 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 330, + 481, + 361 + ], + "spans": [ + { + "bbox": [ + 138, + 330, + 481, + 361 + ], + "type": "text", + "content": "69. Ma, W.C., Huang, D.A., Lee, N., Kitani, K.M.: Forecasting interactive dynamics of pedestrians with fictitious play. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 774-782 (2017)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 363, + 481, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 363, + 481, + 396 + ], + "spans": [ + { + "bbox": [ + 138, + 363, + 481, + 396 + ], + "type": "text", + "content": "70. Ma, Y., Liu, Y., Wang, L., Kang, W., Qiao, Y., Wang, Y.: Dual masked modeling for weakly-supervised temporal boundary discovery. IEEE Transactions on Multimedia (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 396, + 481, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 396, + 481, + 418 + ], + "spans": [ + { + "bbox": [ + 138, + 396, + 481, + 418 + ], + "type": "text", + "content": "71. Matsui, Y., Matsui, T.: Np-completeness for calculating power indices of weighted majority games. Theoretical Computer Science 263(1-2), 305-310 (2001)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 419, + 481, + 452 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 419, + 481, + 452 + ], + "spans": [ + { + "bbox": [ + 138, + 419, + 481, + 452 + ], + "type": "text", + "content": "72. Michalak, T.P., Aadithya, K.V., Szczepanski, P.L., Ravindran, B., Jennings, N.R.: Efficient computation of the shapley value for game-theoretic network centrality. Journal of Artificial Intelligence Research 46, 607-650 (2013)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 453, + 481, + 485 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 453, + 481, + 485 + ], + "spans": [ + { + "bbox": [ + 138, + 453, + 481, + 485 + ], + "type": "text", + "content": "73. Mithun, N.C., Paul, S., Roy-Chowdhury, A.K.: Weakly supervised video moment retrieval from text queries. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 11592-11601 (2019)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 486, + 481, + 518 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 486, + 481, + 518 + ], + "spans": [ + { + "bbox": [ + 138, + 486, + 481, + 518 + ], + "type": "text", + "content": "74. Mun, J., Cho, M., Han, B.: Local-global video-text interactions for temporal grounding. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 10810-10819 (2020)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 520, + 481, + 552 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 520, + 481, + 552 + ], + "spans": [ + { + "bbox": [ + 138, + 520, + 481, + 552 + ], + "type": "text", + "content": "75. Ning, E., Wang, C., Zhang, H., Ning, X., Tiwari, P.: Occluded person re-identification with deep learning: a survey and perspectives. Expert Systems with Applications p. 122419 (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 553, + 481, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 553, + 481, + 586 + ], + "spans": [ + { + "bbox": [ + 138, + 553, + 481, + 586 + ], + "type": "text", + "content": "76. Ning, E., Wang, Y., Wang, C., Zhang, H., Ning, X.: Enhancement, integration, expansion: Activating representation of detailed features for occluded person re-identification. Neural Networks 169, 532-541 (2024)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 586, + 481, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 586, + 481, + 609 + ], + "spans": [ + { + "bbox": [ + 138, + 586, + 481, + 609 + ], + "type": "text", + "content": "77. Ning, E., Zhang, C., Wang, C., Ning, X., Chen, H., Bai, X.: Pedestrian re-id based on feature consistency and contrast enhancement. Displays 79, 102467 (2023)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 609, + 481, + 631 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 609, + 481, + 631 + ], + "spans": [ + { + "bbox": [ + 138, + 609, + 481, + 631 + ], + "type": "text", + "content": "78. Nowak, A.S.: On an axiomatization of the banzhaf value without the additivity axiom. International Journal of Game Theory 26, 137-141 (1997)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 632, + 481, + 654 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 632, + 481, + 654 + ], + "spans": [ + { + "bbox": [ + 138, + 632, + 481, + 654 + ], + "type": "text", + "content": "79. Oord, A.v.d., Li, Y., Vinyals, O.: Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748 (2018)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 138, + 654, + 458, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 654, + 458, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 654, + 458, + 665 + ], + "type": "text", + "content": "80. Osborne, M.J., Rubinstein, A.: A course in game theory. MIT press (1994)" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 252, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 252, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 252, + 91, + 447, + 102 + ], + "type": "text", + "content": "Rethinking WS-VTG From a Game Perspective" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "19" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 18 + }, + { + "para_blocks": [ + { + "bbox": [ + 138, + 117, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 138, + 117, + 479, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 117, + 479, + 160 + ], + "spans": [ + { + "bbox": [ + 138, + 117, + 479, + 160 + ], + "type": "text", + "content": "81. Patel, R., Garnelo, M., Gemp, I., Dyer, C., Bachrach, Y.: Game-theoretic vocabulary selection via the shapley value and banzhaf index. In: Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies. pp. 2789-2798 (2021)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 161, + 480, + 193 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 161, + 480, + 193 + ], + "spans": [ + { + "bbox": [ + 138, + 161, + 480, + 193 + ], + "type": "text", + "content": "82. Pavan, M., Pelillo, M.: A new graph-theoretic approach to clustering and segmentation. In: 2003 IEEE Computer Society Conference on Computer Vision and Pattern Recognition, 2003. Proceedings. vol. 1, pp. I-I. IEEE (2003)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 194, + 480, + 224 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 194, + 480, + 224 + ], + "spans": [ + { + "bbox": [ + 138, + 194, + 480, + 224 + ], + "type": "text", + "content": "83. Pennington, J., Socher, R., Manning, C.D.: Glove: Global vectors for word representation. In: Proceedings of the Conference on Empirical Methods in Natural Language Processing. pp. 1532-1543 (2014)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 225, + 480, + 257 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 225, + 480, + 257 + ], + "spans": [ + { + "bbox": [ + 138, + 225, + 480, + 257 + ], + "type": "text", + "content": "84. Rodola, E., Bronstein, A.M., Albarelli, A., Bergamasco, F., Torsello, A.: A game-theoretic approach to deformable shape matching. In: 2012 IEEE Conference on Computer Vision and Pattern Recognition. pp. 182-189. IEEE (2012)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 258, + 378, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 258, + 378, + 267 + ], + "spans": [ + { + "bbox": [ + 138, + 258, + 378, + 267 + ], + "type": "text", + "content": "85. Shapley, L.S., et al.: A value for n-person games (1953)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 268, + 480, + 299 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 268, + 480, + 299 + ], + "spans": [ + { + "bbox": [ + 138, + 268, + 480, + 299 + ], + "type": "text", + "content": "86. Sigurdsson, G.A., Varol, G., Wang, X., Farhadi, A., Laptev, I., Gupta, A.: Hollywood in homes: Crowdsourcing data collection for activity understanding. In: European Conference on Computer Vision. pp. 510-526 (2016)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 300, + 480, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 300, + 480, + 331 + ], + "spans": [ + { + "bbox": [ + 138, + 300, + 480, + 331 + ], + "type": "text", + "content": "87. Song, Y., Wang, J., Ma, L., Yu, J., Liang, J., Yuan, L., Yu, Z.: Marn: Multilevel attentional reconstruction networks for weakly supervised video temporal grounding. Neurocomputing 554, 126625 (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 332, + 480, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 332, + 480, + 364 + ], + "spans": [ + { + "bbox": [ + 138, + 332, + 480, + 364 + ], + "type": "text", + "content": "88. Song, Y., Wang, J., Ma, L., Yu, Z., Yu, J.: Weakly-supervised multi-level attentional reconstruction network for grounding textual queries in videos. arXiv preprint arXiv:2003.07048 (2020)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 365, + 480, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 365, + 480, + 407 + ], + "spans": [ + { + "bbox": [ + 138, + 365, + 480, + 407 + ], + "type": "text", + "content": "89. Tan, R., Xu, H., Saenko, K., Plummer, B.A.: Logan: Latent graph co-attention network for weakly-supervised video moment retrieval. In: Proceedings of the IEEE Winter Conference on Applications of Computer Vision. pp. 2083–2092 (2021)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 407, + 480, + 438 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 407, + 480, + 438 + ], + "spans": [ + { + "bbox": [ + 138, + 407, + 480, + 438 + ], + "type": "text", + "content": "90. Tang, K., Chen, Y., Peng, W., Zhang, Y., Fang, M., Wang, Z., Song, P.: Reppvconv: attentively fusing reparameterized voxel features for efficient 3d point cloud perception. The Visual Computer 39(11), 5577-5588 (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 439, + 480, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 439, + 480, + 480 + ], + "spans": [ + { + "bbox": [ + 138, + 439, + 480, + 480 + ], + "type": "text", + "content": "91. Tang, K., Lou, T., Peng, W., Chen, N., Shi, Y., Wang, W.: Effective single-step adversarial training with energy-based models. IEEE Transactions on Emerging Topics in Computational Intelligence (2024). https://doi.org/10.1109/TETCI.2024.3378652" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 481, + 480, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 481, + 480, + 514 + ], + "spans": [ + { + "bbox": [ + 138, + 481, + 480, + 514 + ], + "type": "text", + "content": "92. Tang, K., Ma, Y., Miao, D., Song, P., Gu, Z., Tian, Z., Wang, W.: Decision fusion networks for image classification. IEEE Transactions on Neural Networks and Learning Systems (2022). https://doi.org/10.1109/TNNLS.2022.3196129" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 514, + 480, + 546 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 514, + 480, + 546 + ], + "spans": [ + { + "bbox": [ + 138, + 514, + 480, + 546 + ], + "type": "text", + "content": "93. Tang, K., Shi, Y., Lou, T., Peng, W., He, X., Zhu, P., Gu, Z., Tian, Z.: Rethinking perturbation directions for imperceptible adversarial attacks on point clouds. IEEE Internet of Things Journal 10(6), 5158-5169 (2022)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 547, + 480, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 547, + 480, + 590 + ], + "spans": [ + { + "bbox": [ + 138, + 547, + 480, + 590 + ], + "type": "text", + "content": "94. Tang, K., Zhao, W., Peng, W., Fang, X., Cui, X., Zhu, P., Tian, Z.: Reparameterization head for efficient multi-input networks. In: ICASSP 2024 - 2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). pp. 6190-6194 (2024). https://doi.org/10.1109/ICASSP48485.2024.10447574" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 590, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 590, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 138, + 590, + 480, + 632 + ], + "type": "text", + "content": "95. Tang, K., Zhao, W., Peng, W., Fang, X., Cui, X., Zhu, P., Tian, Z.: Reparameterization head for efficient multi-input networks. In: ICASSP 2024-2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). pp. 6190-6194. IEEE (2024)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 633, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 633, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 633, + 480, + 665 + ], + "type": "text", + "content": "96. Torsello, A., Bulo, S.R., Pelillo, M.: Grouping with asymmetric affinities: A game-theoretic perspective. In: 2006 IEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR'06). vol. 1, pp. 292-299. IEEE (2006)" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "20" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 304, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 304, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 304, + 102 + ], + "type": "text", + "content": "X. Fang, Z. Xiong, W. Fang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 19 + }, + { + "para_blocks": [ + { + "bbox": [ + 134, + 116, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 136, + 116, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 116, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 136, + 116, + 480, + 149 + ], + "type": "text", + "content": "97. Tran, D., Bourdev, L., Fergus, R., Torresani, L., Paluri, M.: Learning spatiotemporal features with 3d convolutional networks. In: Proceedings of the IEEE International Conference on Computer Vision. pp. 4489-4497 (2015)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 151, + 480, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 151, + 480, + 184 + ], + "spans": [ + { + "bbox": [ + 138, + 151, + 480, + 184 + ], + "type": "text", + "content": "98. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. In: Advances in Neural Information Processing Systems. pp. 5998-6008 (2017)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 185, + 480, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 185, + 480, + 217 + ], + "spans": [ + { + "bbox": [ + 138, + 185, + 480, + 217 + ], + "type": "text", + "content": "99. Wang, C., Ning, X., Li, W., Bai, X., Gao, X.: 3d person re-identification based on global semantic guidance and local feature aggregation. IEEE Transactions on Circuits and Systems for Video Technology (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 134, + 218, + 480, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 218, + 480, + 251 + ], + "spans": [ + { + "bbox": [ + 134, + 218, + 480, + 251 + ], + "type": "text", + "content": "100. Wang, C., Ning, X., Sun, L., Zhang, L., Li, W., Bai, X.: Learning discriminative features by covering local geometric space for point cloud analysis. IEEE Transactions on Geoscience and Remote Sensing 60, 1-15 (2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 134, + 251, + 480, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 251, + 480, + 274 + ], + "spans": [ + { + "bbox": [ + 134, + 251, + 480, + 274 + ], + "type": "text", + "content": "101. Wang, C., Wang, C., Li, W., Wang, H.: A brief survey on rgb-d semantic segmentation using deep learning. Displays 70, 102080 (2021)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 134, + 274, + 480, + 307 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 274, + 480, + 307 + ], + "spans": [ + { + "bbox": [ + 134, + 274, + 480, + 307 + ], + "type": "text", + "content": "102. Wang, C., Wang, H., Ning, X., Shengwei, T., Li, W.: 3d point cloud classification method based on dynamic coverage of local area. Journal of Software 34(4), 1962-1976 (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 134, + 308, + 480, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 308, + 480, + 341 + ], + "spans": [ + { + "bbox": [ + 134, + 308, + 480, + 341 + ], + "type": "text", + "content": "103. Wang, J., Ma, L., Jiang, W.: Temporally grounding language queries in videos by contextual boundary-aware prediction. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 34, pp. 12168-12175 (2020)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 134, + 342, + 480, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 342, + 480, + 373 + ], + "spans": [ + { + "bbox": [ + 134, + 342, + 480, + 373 + ], + "type": "text", + "content": "104. Wang, Y., Deng, J., Zhou, W., Li, H.: Weakly supervised temporal adjacent network for language grounding. IEEE Transactions on Multimedia 24, 3276-3286 (2021)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 134, + 375, + 480, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 375, + 480, + 407 + ], + "spans": [ + { + "bbox": [ + 134, + 375, + 480, + 407 + ], + "type": "text", + "content": "105. Wang, Z., Chen, J., Jiang, Y.G.: Visual co-occurrence alignment learning for weakly-supervised video moment retrieval. In: Proceedings of the 29th ACM International Conference on Multimedia. pp. 1459–1468 (2021)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 134, + 409, + 480, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 409, + 480, + 453 + ], + "spans": [ + { + "bbox": [ + 134, + 409, + 480, + 453 + ], + "type": "text", + "content": "106. Wen, J., Liu, C., Deng, S., Liu, Y., Fei, L., Yan, K., Xu, Y.: Deep double incomplete multi-view multi-label learning with incomplete labels and missing views. IEEE Transactions on Neural Networks and Learning Systems pp. 1-13 (2023). https://doi.org/10.1109/TNNLS.2023.3260349" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 134, + 453, + 480, + 486 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 453, + 480, + 486 + ], + "spans": [ + { + "bbox": [ + 134, + 453, + 480, + 486 + ], + "type": "text", + "content": "107. Wen, J., Zhang, Z., Li, Z.J.: A survey on incomplete multiview clustering. IEEE transactions on systems, man, and cybernetics. Systems 53(2 Pt.2), 1136-1149 (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 134, + 487, + 480, + 509 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 487, + 480, + 509 + ], + "spans": [ + { + "bbox": [ + 134, + 487, + 480, + 509 + ], + "type": "text", + "content": "108. Winter, E.: The shapley value. Handbook of game theory with economic applications 3, 2025-2054 (2002)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 134, + 510, + 480, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 510, + 480, + 553 + ], + "spans": [ + { + "bbox": [ + 134, + 510, + 480, + 553 + ], + "type": "text", + "content": "109. Wu, H., Lyu, Y., Shen, X., Zhao, X., Wang, M., Zhang, X., Luo, Z.: Atomic-action-based contrastive network for weakly supervised temporal language grounding. In: 2023 IEEE International Conference on Multimedia and Expo (ICME). pp. 1523-1528. IEEE (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 134, + 554, + 480, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 554, + 480, + 586 + ], + "spans": [ + { + "bbox": [ + 134, + 554, + 480, + 586 + ], + "type": "text", + "content": "110. Xiong, Z., Liu, D., Zhou, P.: Gaussian kernel-based cross modal network for spatio-temporal video grounding. In: IEEE International Conference on Image Processing (ICIP). pp. 2481-2485 (2022)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 134, + 587, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 587, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 134, + 587, + 480, + 632 + ], + "type": "text", + "content": "111. Xiong, Z., Liu, D., Zhou, P., Zhu, J.: Tracking objects and activities with attention for temporal sentence grounding. In: ICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). pp. 1-5. IEEE (2023)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 134, + 632, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 632, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 134, + 632, + 480, + 665 + ], + "type": "text", + "content": "112. Yang, W., Zhang, T., Zhang, Y., Wu, F.: Local correspondence network for weakly supervised temporal sentence grounding. IEEE Transactions on Image Processing 30, 3252-3262 (2021)" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 252, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 252, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 252, + 91, + 447, + 102 + ], + "type": "text", + "content": "Rethinking WS-VTG From a Game Perspective" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "text", + "content": "21" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 20 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 658 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 480, + 149 + ], + "type": "text", + "content": "113. Yu, Z., Li, L., Xie, J., Wang, C., Li, W., Ning, X.: Pedestrian 3d shape understanding for person re-identification via multi-view learning. IEEE Transactions on Circuits and Systems for Video Technology (2024)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 149, + 481, + 192 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 149, + 481, + 192 + ], + "spans": [ + { + "bbox": [ + 133, + 149, + 481, + 192 + ], + "type": "text", + "content": "114. Yuan, Y., Ma, L., Wang, J., Liu, W., Zhu, W.: Semantic conditioned dynamic modulation for temporal sentence grounding in videos. In: Proceedings of the 33rd International Conference on Neural Information Processing Systems. pp. 536-546 (2019)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 192, + 480, + 225 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 192, + 480, + 225 + ], + "spans": [ + { + "bbox": [ + 133, + 192, + 480, + 225 + ], + "type": "text", + "content": "115. Zeng, R., Xu, H., Huang, W., Chen, P., Tan, M., Gan, C.: Dense regression network for video grounding. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 10287-10296 (2020)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 225, + 480, + 269 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 225, + 480, + 269 + ], + "spans": [ + { + "bbox": [ + 132, + 225, + 480, + 269 + ], + "type": "text", + "content": "116. Zhang, D., Dai, X., Wang, X., Wang, Y.F., Davis, L.S.: Man: Moment alignment network for natural language moment retrieval via iterative graph adjustment. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 1247-1257 (2019)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 269, + 481, + 301 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 269, + 481, + 301 + ], + "spans": [ + { + "bbox": [ + 132, + 269, + 481, + 301 + ], + "type": "text", + "content": "117. Zhang, H., Sun, A., Jing, W., Zhou, J.T.: Span-based localizing network for natural language video localization. In: Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics. pp. 6543-6554 (2020)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 301, + 480, + 334 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 301, + 480, + 334 + ], + "spans": [ + { + "bbox": [ + 132, + 301, + 480, + 334 + ], + "type": "text", + "content": "118. Zhang, H., Xie, Y., Zheng, L., Zhang, D., Zhang, Q.: Interpreting multivariate shapley interactions in dnns. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 35, pp. 10877-10886 (2021)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 334, + 480, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 334, + 480, + 365 + ], + "spans": [ + { + "bbox": [ + 132, + 334, + 480, + 365 + ], + "type": "text", + "content": "119. Zhang, H., Wang, C., Tian, S., Lu, B., Zhang, L., Ning, X., Bai, X.: Deep learning-based 3d point cloud classification: A systematic survey and outlook. Displays 79, 102456 (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 365, + 480, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 365, + 480, + 397 + ], + "spans": [ + { + "bbox": [ + 132, + 365, + 480, + 397 + ], + "type": "text", + "content": "120. Zhang, H., Wang, C., Yu, L., Tian, S., Ning, X., Rodrigues, J.: Pointgt: A method for point-cloud classification and segmentation based on local geometric transformation. IEEE Transactions on Multimedia (2024)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 397, + 480, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 397, + 480, + 430 + ], + "spans": [ + { + "bbox": [ + 132, + 397, + 480, + 430 + ], + "type": "text", + "content": "121. Zhang, S., Peng, H., Fu, J., Luo, J.: Learning 2d temporal adjacent networks for moment localization with natural language. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 34, pp. 12870-12877 (2020)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 430, + 480, + 474 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 430, + 480, + 474 + ], + "spans": [ + { + "bbox": [ + 132, + 430, + 480, + 474 + ], + "type": "text", + "content": "122. Zhang, Z., Lin, Z., Zhao, Z., Xiao, Z.: Cross-modal interaction networks for query-based moment retrieval in videos. In: Proceedings of the 42nd International ACM SIGIR Conference on Research and Development in Information Retrieval. pp. 655-664 (2019)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 474, + 480, + 506 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 474, + 480, + 506 + ], + "spans": [ + { + "bbox": [ + 132, + 474, + 480, + 506 + ], + "type": "text", + "content": "123. Zhang, Z., Zhao, Z., Lin, Z., He, X., et al.: Counterfactual contrastive learning for weakly-supervised vision-language grounding. Advances in Neural Information Processing Systems 33, 18123-18134 (2020)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 506, + 480, + 538 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 506, + 480, + 538 + ], + "spans": [ + { + "bbox": [ + 132, + 506, + 480, + 538 + ], + "type": "text", + "content": "124. Zheng, M., Huang, Y., Chen, Q., Liu, Y.: Weakly supervised video moment localization with contrastive negative sample mining. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 36, pp. 3517-3525 (2022)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 538, + 480, + 582 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 538, + 480, + 582 + ], + "spans": [ + { + "bbox": [ + 132, + 538, + 480, + 582 + ], + "type": "text", + "content": "125. Zheng, M., Huang, Y., Chen, Q., Peng, Y., Liu, Y.: Weakly supervised temporal sentence grounding with gaussian-based contrastive proposal learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15555-15564 (2022)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 582, + 480, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 582, + 480, + 624 + ], + "spans": [ + { + "bbox": [ + 132, + 582, + 480, + 624 + ], + "type": "text", + "content": "126. Zheng, Q., Dong, J., Qu, X., Yang, X., Wang, Y., Zhou, P., Liu, B., Wang, X.: Progressive localization networks for language-based moment localization. ACM Transactions on Multimedia Computing, Communications and Applications 19(2), 1-21 (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 624, + 480, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 624, + 480, + 658 + ], + "spans": [ + { + "bbox": [ + 132, + 624, + 480, + 658 + ], + "type": "text", + "content": "127. Zhu, J., Liu, D., Zhou, P., Di, X., Cheng, Y., Yang, S., Xu, W., Xu, Z., Wan, Y., Sun, L., et al.: Rethinking the video sampling and reasoning strategies for temporal sentence grounding. arXiv preprint arXiv:2301.00514 (2023)" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "22" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 304, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 304, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 304, + 102 + ], + "type": "text", + "content": "X. Fang, Z. Xiong, W. Fang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 21 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Rethinking and Improving Visual Prompt Selection for In-Context Learning Segmentation Framework/c5a744a9-bd8f-4567-93ec-5cbfe907a7d5_content_list.json b/2024/Rethinking and Improving Visual Prompt Selection for In-Context Learning Segmentation Framework/c5a744a9-bd8f-4567-93ec-5cbfe907a7d5_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..39839c69972cda7db013169f75415ef6ee3fe927 --- /dev/null +++ b/2024/Rethinking and Improving Visual Prompt Selection for In-Context Learning Segmentation Framework/c5a744a9-bd8f-4567-93ec-5cbfe907a7d5_content_list.json @@ -0,0 +1,1637 @@ +[ + { + "type": "text", + "text": "Rethinking and Improving Visual Prompt Selection for In-Context Learning Segmentation", + "text_level": 1, + "bbox": [ + 228, + 140, + 774, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Wei Suo $^{1,3,\\star}$ , Lanqing Lai $^{1,\\star}$ , Mengyang Sun $^{2,3}$ , Hanwang Zhang $^{3}$ , Peng Wang $^{1,\\dagger}$ , and Yanning Zhang $^{1}$", + "bbox": [ + 225, + 217, + 779, + 260 + ], + "page_idx": 0 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1 School of Computer Science and Ningbo Institute, Northwestern Polytechnical University, Xi'an, China", + "$^{2}$ School of Cybersecurity, Northwestern Polytechnical University, Xi'an, China", + "3 School of Computer Science, Nanyang Technological University, Singapore {suwei1994, lailanqing, sunmenmian}@mail.nwpu.edu.cn" + ], + "bbox": [ + 232, + 270, + 767, + 340 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. As a fundamental and extensively studied task in computer vision, image segmentation aims to locate and identify different semantic concepts at the pixel level. Recently, inspired by In-Context Learning (ICL), several generalist segmentation frameworks have been proposed, providing a promising paradigm for segmenting specific objects. However, existing works mostly ignore the value of visual prompts or simply apply similarity sorting to select contextual examples. In this paper, we focus on rethinking and improving the example selection strategy. By comprehensive comparisons, we first demonstrate that ICL-based segmentation models are sensitive to different contexts. Furthermore, empirical evidence indicates that the diversity of contextual prompts plays a crucial role in guiding segmentation. Based on the above insights, we propose a new stepwise context search method. Different from previous works, we construct a small yet rich candidate pool and adaptively search the well-matched contexts. More importantly, this method effectively reduces the annotation cost by compacting the search space. Extensive experiments show that our method is an effective strategy for selecting examples and enhancing segmentation performance. https://github.com/LanqingL/SCS", + "bbox": [ + 259, + 377, + 743, + 627 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: In-Context Learning $\\cdot$ Visual Prompt $\\cdot$ Image Segmentation", + "bbox": [ + 261, + 640, + 740, + 655 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 680, + 375, + 696 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Image segmentation is a fundamental task in computer vision, aiming to locate and identify different semantic concepts at the pixel level. It is essential in multiple applications such as autonomous driving [43,54], video surveillance [20,36,39,41] and complex reasoning [24,34,40]. To achieve precise segmentation, researchers have proposed a variety of segmentation models [5,12,23,29]. Although these methods have made considerable progress, they require training", + "bbox": [ + 212, + 710, + 787, + 803 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Corresponding author", + "bbox": [ + 217, + 810, + 383, + 824 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* These authors contributed equally to this work.", + "bbox": [ + 220, + 825, + 550, + 839 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/30f8de8dc374e8afceee8e3a4c1db12b7251d9cb28be9c7397871693fec642a0.jpg", + "image_caption": [ + "Fig. 1: Comparison of traditional and our method. (a) Existing works rely on dense annotation to build the exempling space. Then, they use the similarity sorting manner to select examples for given queries. (b) Instead, our method significantly alleviates the costs of annotation by searching the typical examples. Moreover, a novel adaptive search module is designed to further select well-matched contexts." + ], + "image_footnote": [], + "bbox": [ + 289, + 143, + 717, + 380 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "the specialist model for each segmentation task. Recently, inspired by In-Context Learning (ICL) [26,30,49], several generalist segmentation frameworks have been proposed and achieved impressive performance such as MAE-VQGAN [6] and SegGPT [45]. Different from the traditional frameworks, ICL-based segmentation methods establish a promising paradigm for localizing vision concepts by feeding one or a few examples as references during inference. Meanwhile, the methodology offers a general solution for various real-world scenarios [44].", + "bbox": [ + 212, + 492, + 787, + 598 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "ICL learning is a relatively new research line in computer vision, while the fundamental ideas can be traced back to the Natural Language Processing (NLP) field [6, 44]. With the development of Large Language Models (LLM), numerous studies have already proved that selecting a suitable context is crucial for following human instruction and improving performance [1, 14, 31]. However, related researches about this topic is still limited in our community. As shown in Fig.1 (a), existing works typically randomly select examples [44] or simply apply similarity sorting to construct contextual prompts [37, 56]. The above insights motivate us to question: 1) Whether different contexts would significantly impact performance? 2) What is the critical factor of visual prompt selection for ICL-based segmentation models? In this paper, we first try to answer the above questions. By comprehensive comparisons, we find that the ICL-based models are sensitive to different examples. With multiple experiments, the performance gap between using different contextual examples even exceeds 5 points. Furthermore, we synthetically investigate the effectiveness of the existing prompt selection strategy. Surprisingly, although the similarity measure (i.e., choosing", + "bbox": [ + 212, + 598, + 787, + 843 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "W. Suo et al.", + "bbox": [ + 271, + 114, + 362, + 126 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "the most similar examples as contexts) provides a relatively effective approach for demonstration search, assembling the most dissimilar contexts results in better performance in $\\sim 40\\%$ of the test instances. To reveal the potential reasons, we empirically demonstrate that the diversity of contexts is a crucial factor in guiding segmentation and enhancing performance (more details in Sec. 3).", + "bbox": [ + 212, + 146, + 782, + 222 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Beyond the above discussions, we rethink the role of contextual examples and propose a new Stepwise Context Search method, which is called SCS for short. As shown in Fig. 1 (b), the SCS focuses on maintaining the diversity of the candidate pool and adaptively searching the well-matched contexts. Meanwhile, our method significantly reduces the annotation cost by searching the typical examples. Specifically, given a large unlabeled database, we first apply a basic cluster algorithm to compact the search space and improve annotation efficiency. Then, we extract the two most dissimilar members in each cluster as candidate examples. By collecting these representative members from all clusters, we construct a small yet rich candidate pool. Furthermore, considering the significant impact of examples on performance, we further introduce a novel adaptive search module based on reinforcement learning. It heuristically utilizes segmentation scores as rewards, guiding the model in determining final demonstrations. Extensive experiments across various datasets have shown the effectiveness of our method. To summarize, our contributions are as follows.", + "bbox": [ + 212, + 222, + 784, + 448 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We take a step forward to systematically analyze the influence of example selection on the ICL-based segmentation frameworks and empirically demonstrate that the diversity of contexts is a key factor in guiding segmentation.", + "- We propose a new SCS method to efficiently select visual prompts by constructing a small yet rich candidate pool. Meanwhile, an adaptive search module is designed to further choose the well-matched demonstrations.", + "- The proposed model significantly improves the segmentation performance of ICL-based frameworks and achieves state-of-the-art on PASCAL-5 $^i$ [32], COCO-20 $^i$ [21] and iSALD-5 $^i$ [50]." + ], + "bbox": [ + 225, + 454, + 784, + 588 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related work", + "text_level": 1, + "bbox": [ + 215, + 608, + 382, + 625 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.1 Visual Segmentation", + "text_level": 1, + "bbox": [ + 215, + 636, + 431, + 652 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As a fundamental task in computer vision, visual segmentation entails locating and identifying different semantic concepts at the pixel level. Numerous works for specific tasks have continued to flourish over the years. For example, CrossoverNet [52] learns the vertical and horizontal crossover relation for medical image segmentation. RFNet [35] fuses RGB and depth information which is designed for real-time road-driving segmentation. [16] proposes DGEN, a densely connected global entropy network in which noise is suppressed by context information losses. MSANet [25] accurately extracts water bodies from remote sensing scenes by utilizing multiscale attention. Recently, inspired by ICL learning, several general segmentation methods have been proposed [18,44,45]. Different from these specialist models, the paradigm can locate visual concepts by feeding a few examples, known as the In-Context Learning (ICL) segmentation frameworks.", + "bbox": [ + 212, + 657, + 787, + 840 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "Visual Prompt Selection for In-Context Learning Segmentation", + "bbox": [ + 310, + 114, + 730, + 130 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2 In-Context Learning", + "text_level": 1, + "bbox": [ + 215, + 146, + 431, + 162 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In-context learning (ICL) is a new learning paradigm that initially emerged from GPT-3 [7]. It presents a text fill-in-the-blank problem with examples, eliminating the necessity for fine-tuning model parameters on downstream NLP tasks. In the vision domain, MAE-VQGAN [6] transforms various visual tasks into grid inpainting problems and proposes the first ICL-based segmentation framework. Different from the previous segmentation methods, this new paradigm provides a promising manner to segment arbitrary objects with a few examples. Painter [44] adopts masked image modeling on continuous pixels, unifying the output form across different visual tasks. As a variant of Painter, SegGPT [45] integrates visual segmentation tasks into in-context coloring problems and demonstrates powerful capabilities in target segmentation. Beyond the previous works, we present comprehensive experiments on example selection for visual ICL models.", + "bbox": [ + 212, + 169, + 787, + 351 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.3 Contextual Example Selection", + "text_level": 1, + "bbox": [ + 215, + 372, + 511, + 387 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "With the development of LLMs, the selection of contextual examples is important for ICL learning [1,30,51]. The emerging studies can be categorized into rule-based unsupervised methods and supervised methods. KATE [22] uses the KNN mechanism to enhance a retriever for selecting in-context examples. EPR [30] adopts a contrastive learning strategy to train a dense retriever. UDR [19] explores the process of retrieval demonstrations for various tasks in a unified manner. However, few studies investigate these effects in the vision domain. Current works typically select examples randomly [44] or simply apply similarity sorting to construct contextual demonstrations [37,56]. Different from the previous rule-based methods, we introduce an adaptive search method with a small yet rich candidate pool. This strategy effectively reduces the annotation cost and boosts segmentation performance.", + "bbox": [ + 212, + 395, + 787, + 578 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3 Preliminary", + "text_level": 1, + "bbox": [ + 215, + 598, + 370, + 616 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this section, we would first present a simple overview of the ICL-based segmentation frameworks. Then, we construct extensive experiments to comprehensively analyze the influence of different examples and the effectiveness of the existing selection strategies. Finally, we experimentally proved that contextual diversity is an important factor for guiding segmentation.", + "bbox": [ + 212, + 628, + 787, + 705 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1 Background", + "text_level": 1, + "bbox": [ + 215, + 724, + 362, + 739 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The design of segmentation models has achieved tremendous development in recent years [20, 41, 52, 57]. However, these solutions require expensive annotation cost and face challenges in adapting to new scenarios. In contrast, the ICL-based segmentation frameworks [6, 45] can implement various specific tasks by leveraging several segmented examples. The convenient and versatile vision models provide a promising perspective for our community.", + "bbox": [ + 212, + 750, + 787, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "W. Suo et al.", + "bbox": [ + 271, + 114, + 362, + 127 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As shown in Fig. 1, the goal of ICL-based segmentation methods is to predict segmentation regions based on exhibited one or a few prompts. By reorganizing the input format, the models can imitate provided contexts to mask specific object regions. Specifically, given a query image $I_q$ , these segmentation frameworks first select $n$ example images $E_d = \\{I_i^d\\}_{i=1}^n$ with corresponding masks $G_d = \\{G_i^d\\}_{i=1}^n$ as contextual demonstrations from a huge database. Here, $I_i^d$ and $G_i^d$ denote the $i$ -th selected example and corresponding mask. By reformating the input, models can perform multiple segmentation tasks across different scenarios. The above operations can be formulated as:", + "bbox": [ + 212, + 146, + 787, + 282 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {I} _ {m} = \\mathbf {S} \\left(I _ {q}, E _ {d}, G _ {d}\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 433, + 297, + 785, + 315 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where the $\\mathbf{S}(\\cdot)$ and $\\hat{I}_m$ are the generalist segmentation model and predicted mask results, respectively. Although the above calculations provide a convenient way to segment objects at arbitrary granularities, the success of the paradigm greatly relies on the given contextual examples. In fact, the basic idea of ICL-based vision models comes from the NLP field, in which it has been proved that contextual prompts play an important role in executing reasoning and boosting performance. However, in the vision community, existing approaches mostly simply choose the demonstrations (i.e., $E_d$ and corresponding $G_d$ ) based on random sampling [44] or rely on similarity sorting [37,56]. To explore the impact of the visual prompts on performance, we construct a series of quantitative analyses on the PASCAL dataset [32] with the state-of-the-art ICL-based segmentation model SegGPT [45]. Note that the SegGPT involves two training stages (i.e., pre-training and in-context tuning). In the tuning process, the SegGPT depends on the whole annotation (including base and new classes) to improve the performance. Since the main advantage of the ICL-based framework is to implement few-shot tasks, we only use the pre-trained weights provided by the SegGPT [3] to construct experiments. Following [6,44], the mean intersection over union (mIoU) is adopted as the evaluation metric.", + "bbox": [ + 212, + 330, + 787, + 604 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2 Influence of Different Examples", + "text_level": 1, + "bbox": [ + 215, + 637, + 522, + 652 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As mentioned above, research on the influence of different examples in segmentation remains insufficient. In this paper, we aim to fill this gap by conducting experiments to quantitatively analyze the impact of different demonstrations. Specifically, we randomly select examples from the entire training set with 5 different seeds. As shown in Fig. 2 (a), we report the best, worst, and average results across the 5 experiments. It reveals that allocating different contexts would dramatically affect the segmentation scores under the 1-shot and 5-shot settings. In particular, the best results can surpass the worst by margins of 5.6 and 4.1 points, respectively. Due to the huge performance discrepancy between different contextual prompts, designing an automated method for searching well-matched examples emerges as a critical challenge.", + "bbox": [ + 212, + 672, + 787, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "Visual Prompt Selection for In-Context Learning Segmentation", + "bbox": [ + 308, + 114, + 730, + 130 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/3ba3f6b6695ae70ac197f798a37dcb2ff00526e74dfd05841a017b0dec74d603.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 215, + 146, + 472, + 292 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/9c0c4efc627eef0ee4845108d2e602c5516ede60697471f5dc06096b8f547f8f.jpg", + "image_caption": [ + "(b)", + "Fig. 2: The influence of different context selection. (a) Randomly sampling contextual examples with 5 runs under 1-shot and 5-shot setting on PASCAL. (b) For each instance, the nearest and farthest examples are retrieved as visual prompts across different similarity-based sorting. Surprisingly, the most dissimilar examples achieved better performance on $\\sim 40\\%$ of the test samples." + ], + "image_footnote": [], + "bbox": [ + 503, + 142, + 751, + 292 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.3 Example Selection with Similarity", + "text_level": 1, + "bbox": [ + 214, + 414, + 542, + 429 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To the best of our knowledge, the related researches to our work are [37, 56], where they simply rely on cosine similarity to select examples with qualitative analysis. In this paper, we try to rethink the strategy through a quantitative discussion across different similarity computation manners. Specifically, we explore two types of image similarity calculation strategies including appearance-level and semantic-level. For the appearance-level comparison, PSNR [47], SSIM [15], FSIM [55] and HaarPSI [28] obtain the similarity between the test instance and the entire training set. On the other hand, for semantic-level comparison, we follow [37, 56] to utilize the off-the-shelf image encoders (i.e., CLIP [27] and ViT [8]) for feature extraction. To verify the effectiveness of these similarity-based approaches, we carefully select the examples based on cosine similarity sorting.", + "bbox": [ + 212, + 435, + 787, + 617 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "As shown in Fig. 2 (b), each metric is represented by two bars including nearest $>$ farthest (yellow bar) and nearest $<$ farthest (green bar). We take yellow as an example, it denotes the proportion of instances where using the most similar examples (nearest) as contexts results in a better IoU than using the most dissimilar ones (farthest), and the green bar indicates the opposite. Surprisingly, we observed that assembling the most dissimilar contexts leads to superior performance for numerous test instances. There is a substantial portion of the test samples ( $\\sim 40\\%$ ) that achieve better IoU scores by utilizing the farthest examples. The above results indicate that using similarity measures is not the only way to search for contextual examples.", + "bbox": [ + 212, + 618, + 789, + 768 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.4 Example Selection with Diversity", + "text_level": 1, + "bbox": [ + 214, + 787, + 539, + 803 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Based on the above analyses, it is imperative to explore a more effective way for searching visual prompts. In Fig. 2 (b), we find that applying the most dissimilar", + "bbox": [ + 212, + 809, + 787, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "W. Suo et al.", + "bbox": [ + 271, + 114, + 362, + 126 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/333d6e77700b90c23dd8c93475e3bc2594dd8fdd9450573fc1981a7317d96300.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 215, + 143, + 442, + 303 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/460ec976bd55dc37574f6d0a6bc629d41d69c6667eb1805ddab4ef5c21bf00c1.jpg", + "image_caption": [ + "(a) Example sampling with SegGPT [45].", + "Fig. 3: Diversity vs Similarity. Based on similarity sorting, the performance of the two Nearest examples (NN), the two Farthest examples (FF), and the Nearest example with the Farthest example (NF) as visual prompts are shown." + ], + "image_footnote": [], + "bbox": [ + 233, + 306, + 393, + 333 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d305ee10253b43082161de2f63eb95fec69e1721d0dee63e99069ce3910ae665.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 496, + 142, + 730, + 303 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a1081596559806d6cb7aa309a775c0670033f39b9a76b5077ac3101ae32a246c.jpg", + "image_caption": [ + "(b) Example sampling with MAE-VQGAN [6]." + ], + "image_footnote": [], + "bbox": [ + 545, + 306, + 681, + 333 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "examples can also yield satisfactory performance for some instances. The results motivate us to investigate whether combining the nearest and farthest contexts can obtain a performance boost. Consequently, we further construct a series of experiments to verify the hypothesis across different ICL-based models.", + "bbox": [ + 212, + 429, + 787, + 491 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "More specifically, based on similarity sorting, we use two Nearest examples (NN), two Farthest examples (FF), and a combination of the Nearest example and the Farthest example (NF) as contextual demonstrations, respectively. As shown in Fig. 3, we report related results across different similarity measure manners. It can be observed that the performance using NF examples significantly surpasses the other two selection strategies in all measures. Regarding ViT, the NF achieves better performance that outperforms equipping the two nearest examples by a margin of 2.4 points.", + "bbox": [ + 212, + 489, + 787, + 612 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.5 Discussion", + "text_level": 1, + "bbox": [ + 214, + 633, + 349, + 647 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this section, we would try to summarize the above findings with a deeper discussion. Based on the insights presented in Sec. 3.2, we find that the ICL-based segmentation framework is sensitive to contextual examples. Given the different visual prompts, the performance gap is even over 5 points on the mIoU. Therefore, building an effective prompt selection method is important for boosting the segmentation results and achieving stable performance.", + "bbox": [ + 212, + 657, + 787, + 748 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To further explore the key factors that influence example selection, we design experiments in Sec. 3.3 and Sec. 3.4. The results demonstrate that combining both similar and dissimilar demonstrations can significantly enhance the IoU scores. We argue that the underlying reason is that this manner provides richer semantic information about the segmented objects by expanding the diversity of visual prompts. Based on the above insights, a natural idea is to select these", + "bbox": [ + 212, + 750, + 787, + 840 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "Visual Prompt Selection for In-Context Learning Segmentation", + "bbox": [ + 308, + 114, + 730, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 114, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "considerable discrepancy examples as contextual prompts for each test instance. However, implementing the straightforward strategy would incur substantial annotation cost and complex computation, since each test sample needs to compare and rerank the candidate list. To mitigate the above problems, we introduce a novel stepwise context search (SCS) method that maintains the diversity of the candidate pool and reduces the cost of annotation. Moreover, considering the necessity of selecting appropriate contextual examples for different test queries, we further design an adaptive search module. It can dynamically select contexts from the candidate pool based on given test samples.", + "bbox": [ + 212, + 146, + 787, + 282 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4 Our Method", + "text_level": 1, + "bbox": [ + 215, + 305, + 375, + 321 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Based on the discussion in Sec. 3, we design a new SCS method to construct a small yet rich candidate pool and adaptively search the contextual demonstrations. Compared to previous works, our method has two potential advantages: 1) Fewer Annotations. The SCS focuses on building a small yet diverse candidate pool. It only needs to annotate a portion of samples while achieving better performance. 2) Adapting Selection. Compared to the previous rule-based mechanism [37,56], our method can adaptively search the well-matched examples for different test instances. Next, we would introduce our method in detail.", + "bbox": [ + 212, + 335, + 787, + 457 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.1 Constructing Candidate Pool", + "text_level": 1, + "bbox": [ + 215, + 479, + 503, + 494 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "As mentioned in Sec. 3, it is important to provide examples for the ICL-based segmentation paradigm. However, previous research mostly depend on huge annotations as contextual candidates, it would inevitably suffers from expensive annotation cost. Motivated by the of Sec. 3.5, we aim to build a small yet diverse candidate pool that can be used for subsequent context searching.", + "bbox": [ + 212, + 503, + 785, + 579 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Given $N$ unlabeled samples, follow in [56], we utilize CLIP to extract image feature into $d$ -dimensional vectors, represented by $F = \\{f_{i}\\}_{i=1}^{N}$ , $f_{i} \\in \\mathbb{R}^{d}$ , where $f_{i}$ is the $i$ -th encoded image feature. To construct a candidate pool comprising a set of representative examples, we employ k-means clustering algorithm [58] to generate $M$ clusters $C = \\{c_{m}\\}_{m=1}^{M}$ , which can be determined by minimizing the sum of squared intra-cluster norms:", + "bbox": [ + 212, + 579, + 787, + 671 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\arg \\min _ {C} \\sum_ {m = 1} ^ {M} \\sum_ {f _ {i} \\in c _ {j}} \\| f _ {i} - \\mu_ {m} \\|, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 401, + 681, + 785, + 724 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $\\mu_{m}$ is centroid of cluster $c_{m}$ . Initially, this algorithm randomly selects $M$ centroids. Each instance is assigned by computing the distance to all selected centroids. The process is iteratively executed until the clusters are stable.", + "bbox": [ + 212, + 734, + 785, + 779 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To construct the candidate pool, a simple method is to select members closest to the centroid, while previous studies have shown that samples closest to the center do not necessarily represent the entire cluster [11, 46]. Moreover, related works have provided several alternative sampling methods such as density", + "bbox": [ + 212, + 780, + 787, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "W. Suo et al.", + "bbox": [ + 271, + 114, + 362, + 127 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/df3994d57e3a7c22667b526a3dc10a846ee3a30739ced14e5ddf894e0176e41b.jpg", + "image_caption": [ + "Fig. 4: Overview of our SCS method. Instead of similarity sorting on a large annotated dataset, we use clustering to select diverse examples from unlabeled data $D$ and construct the candidate pool. Meanwhile, the search agent is used to further select contextual demonstrations for various test samples based on reinforcement learning. During inference, the test sample and candidate pool examples are fed into the model and adaptively search visual prompts." + ], + "image_footnote": [], + "bbox": [ + 217, + 145, + 781, + 315 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "sorting [46] and entropy sorting [11]. In this paper, we design a simpler and more efficient sampling strategy without complex computation. Specifically, for each cluster $c_{m}$ , which includes members $F_{m} = \\{f_{k}\\}_{k=1}^{K}$ , where $K$ denotes the number of members in $c_{m}$ . We sort them into a list $[f_{k}^{(1)}, f_{k}^{(2)}, \\dots, f_{k}^{(K)}]$ in the ascending order of their distance to the cluster's center. Based on the findings of Sec. 3.3, aggregating the nearest and farthest examples can cover more diverse demonstrations. Thus, we consider the $f_{mk}^{(1)}$ and $f_{mk}^{(K)}$ as typical samples for each cluster. By collecting these samples from all clusters, we obtain the final candidate pool $D = \\{f_{1k}^{(1)}, f_{1k}^{(K)}, f_{2k}^{(1)}, \\dots, f_{Mk}^{(1)}, f_{Mk}^{(K)}\\}$ , which contains a total of $2M$ samples. To simplify, we use $f_{d}^{m}$ and $I_{d}^{m}$ to represent $m$ -th image feature and corresponding image, respectively.", + "bbox": [ + 212, + 438, + 787, + 612 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.2 Adaptive Search Module", + "text_level": 1, + "bbox": [ + 215, + 631, + 465, + 645 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "In this module, our goal is to select well-matched examples from the constructed candidate pool. As shown in Fig. 4, we propose a new adaptive search module that can further search and assemble the suitable contexts for given queries.", + "bbox": [ + 212, + 654, + 785, + 699 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Specifically, for a given test instance $I_{q}$ , we first apply the same image encoder (i.e. CLIP) to extract the corresponding feature vector $f_{q} \\in \\mathbb{R}^{d}$ . Then, we concatenate the image feature $f_{q}$ and each example $f_{d}^{m}$ from the candidate pool. These fused features are fed into a search agent that consists of a group of light and simple Multi-Layer Perceptrons (MLP). The search computation can be denoted as follows:", + "bbox": [ + 212, + 699, + 787, + 789 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\ns _ {d} ^ {m} = \\operatorname {M L P} \\left(\\left[ f _ {d} ^ {m}; f _ {q} \\right]\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 434, + 799, + 784, + 815 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\na _ {d} ^ {1}, a _ {d} ^ {2}, \\dots , a _ {d} ^ {2 M} = \\operatorname {s o f t m a x} \\left(s _ {d} ^ {1}, s _ {d} ^ {2}, \\dots , s _ {d} ^ {2 M}\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 357, + 818, + 784, + 835 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "Visual Prompt Selection for In-Context Learning Segmentation", + "bbox": [ + 310, + 114, + 730, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 114, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where $a_{d}^{m}$ represents the probability of $m$ -th example in the candidate pool. To guide the search agent in learning select examples, we apply the IoU scores as rewards for selecting well-matched demonstrations from the candidate pool. Specifically, following the format of the ICL-based segmentation model, we re-organize query $I_{q}$ , the original image of context examples $I_{d}$ and corresponding mask $G_{d}$ into a joint image, as shown in Fig. 4. These images are then fed into the ICL-based segmentation model to obtain the IoU scores $U = \\{u_{d}\\}_{d=1}^{2M}$ . Inspired by [4,38], we use the average IoU score $u_{avg}$ as the base score. By applying reinforcement learning, the gradient is calculated by:", + "bbox": [ + 212, + 146, + 787, + 282 + ], + "page_idx": 9 + }, + { + "type": "equation", + "text": "\n$$\n\\nabla_ {\\theta} L (\\theta) = - \\frac {1}{2 M} \\sum_ {m = 1} ^ {2 M} \\left(u _ {d} ^ {m} - u _ {a v g}\\right) \\nabla_ {\\theta} \\log \\left(a _ {d} ^ {m}\\right), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 344, + 295, + 785, + 335 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "where $a_{d}^{m}$ and $\\theta$ are the probability of $m$ -th example and the parameters of the search agent. Based on the above computation, this gradient would tend to increase the probability of the $m$ -th context when its IoU scores $u_{d}^{m}$ are higher than the average IoU score $u_{avg}$ .", + "bbox": [ + 212, + 347, + 787, + 407 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "During inference, we directly input the test samples and candidate pool examples into the search agent. The contextual examples are then determined by sorting them based on probability scores $a_{d}$ as detailed in Eq. 4.", + "bbox": [ + 212, + 407, + 785, + 454 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "5 Experiment", + "text_level": 1, + "bbox": [ + 215, + 479, + 366, + 497 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "We use different prompt selection strategies and various generalist segmentation frameworks to demonstrate the superiority of our method. Furthermore, we compare our approach with traditional methods and conduct extensive qualitative and quantitative studies on multiple datasets.", + "bbox": [ + 212, + 513, + 785, + 574 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "5.1 Experimental Setting", + "text_level": 1, + "bbox": [ + 215, + 599, + 439, + 616 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Dataset. We follow the previous works [6, 45, 56] and evaluate our approach on PASCAL- $5^{i}$ [32], COCO- $20^{i}$ [21] and iSALD- $5^{i}$ [50]. PASCAL- $5^{i}$ includes 20 semantic categories and is created from PASCAL VOC 2012 [9] with external annotations from SDS [10]. COCO- $20^{i}$ is built upon MSCOCO [21], which is a more challenging dataset and contains 80 categories. For all categories, both COCO- $20^{i}$ and PASCAL- $5^{i}$ are evenly divided into 4 folds for cross-validation, where each fold is selected as the test set and the remaining three folds are used for training. The iSALD- $5^{i}$ contains 15 geospatial categories which derived from iSALD [48] in the remote sensing image semantic segmentation task. Similarly, the iSALD- $5^{i}$ is also partitioned into 3 folds for cross-validation.", + "bbox": [ + 212, + 627, + 787, + 777 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Evaluation Metric. Following [13], mean intersection over union (mIoU) and foreground-background IoU (FBIoU) are adopted as evaluation metrics. The former measures the average overall IoU values for all object classes, while the latter reports the average IoU scores for the foreground and background.", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "W. Suo et al.", + "bbox": [ + 271, + 114, + 359, + 126 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/5897871e0b3fbb511655dc2905a37712928ce00ea5ef58c5cc25368e67883f60.jpg", + "table_caption": [ + "Table 1: The results on COCO- $20^{i}$ and PASCAL- $5^{i}$ based on example-based semantic segmentation. To ensure a fair comparison and follow the few-shot setting, $\\dagger$ indicates that only using the pre-trained weights provided by the SegGPT [3], and we report related results based on randomly sampling (i.e., w/R) and similarity sorting [56]." + ], + "table_footnote": [], + "table_body": "
MethodCOCO-20i [21]PASCAL-5i [32]
1-shot5-shot1-shot5-shot
mIoUFBIoUmIoUFBIoUmIoUFBIoUmIoUFBIoU
Painter [44]32.850.232.650.164.574.264.674.4
MAE-VQGAN [6]12.145.7--27.555.6--
MAE-VQGAN w/ [56]18.046.5--33.657.8--
SegGPT† w/R [45]48.671.352.773.172.183.176.280.3
SegGPT† w/ [56]53.372.256.175.574.484.476.985.9
MAE-VQGAN w/Ours19.147.0--35.061.2--
SegGPT† w/Ours57.275.159.576.876.685.478.286.5
", + "bbox": [ + 243, + 212, + 759, + 383 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Implementation details. For both datasets, the Adam [17] is used as our optimizer and the maximum number of training epochs is 10. We use the CLIPViT [27] image encoder to extract image features. During training, all images are resized to $224 \\times 224$ and fed into the image encoder. The dimension of extracted feature vectors is 1024. We train on one 4080 GPU with an initial learning rate of 1e-3, which is dropped by half every 5 epochs. The number of clusters $M$ and batch size is set to 10 and 8.", + "bbox": [ + 212, + 428, + 787, + 537 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "5.2 Quantitative Evaluation", + "text_level": 1, + "bbox": [ + 215, + 571, + 460, + 588 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "In Table 1, we show a performance comparison of two general ICL-based segmentation models with and without proposed SCS. Following [37,56], all results are reported by repeating the random experiment five times. It can be observed that our method can significantly boost performance in all settings including 1-shot and 5-shot settings. In particular, for the 1-shot setting, the SCS leads to MAE-VQGAN with the mIoU gain of 7.0 and 6.5 on the PASCAL- $5^{i}$ and COCO- $20^{i}$ dataset, respectively. Besides, compared to previous rule-base models [56] in 1-shot and 5-shot semantic segmentation tasks, our method surpasses it by a considerable margin. Moreover, following the few-shot setting and ensuring a fair comparison, we report related results based on the pre-trained weights for SegGPT [3]. It can be observed that our method significantly outperforms these rule-based visual prompt selection methods for SegGPT by 9.7 and 6.8 on the 1-shot and 5-shot settings for COCO- $20^{i}$ . More importantly, the above results demonstrate that the proposed SCS can be considered as a \"plug-and-play\" module to enhance the segmentation ability of ICL-based frameworks.", + "bbox": [ + 212, + 613, + 787, + 840 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "Visual Prompt Selection for In-Context Learning Segmentation", + "bbox": [ + 310, + 114, + 730, + 130 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 782, + 127 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/e6f5507ba1dc6511b390a4f1fdfc97f2919c4427371d8d6ca22c99b46b5c69b3.jpg", + "table_caption": [ + "Table 2: Ablation study. Similarity: Cosine similarity ranking; Cluster: Cluster compacting; Sorting: The nearest-farthest sampling; ASM: Adaptive search module." + ], + "table_footnote": [], + "table_body": "
SimilarityClusterSortingASMmIoUFBIoU
152.773.1
2✓56.174.5
3✓✓55.774.3
4✓✓✓57.675.7
5✓✓✓59.576.8
", + "bbox": [ + 300, + 196, + 702, + 286 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/3ec1bc3870c6af921dceef7628993c1884acb894acb623c7d6be9b49e105963f.jpg", + "table_caption": [ + "Table 3: The effects of different settings for our method." + ], + "table_footnote": [], + "table_body": "
Model1-shot5-shot
mIoUFBIoUmIoUFBIoU
1Entropy53.172.356.175.0
2Foreground53.572.556.779.3
3ViT56.975.058.776.0
4CLIP-resnet56.374.758.576.4
5Density-sort57.675.159.176.3
6Entropy-sort56.974.558.475.9
7Ours57.275.159.576.8
", + "bbox": [ + 354, + 337, + 640, + 468 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "5.3 Ablation Study", + "text_level": 1, + "bbox": [ + 215, + 497, + 388, + 512 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "To clarify the effectiveness of our method, we conduct further ablation studies on the COCO- $20^{i}$ dataset with the 5-shot setting. The SegGPT is used as the basic model due to its outstanding segmentation performance.", + "bbox": [ + 212, + 522, + 785, + 566 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Effects of Cluster. In Table 2, we first build a baseline by randomly selecting examples from the entire training set. The average result is reported across 5 different seeds. In the second row of Table 2, we follow [56] and utilize cosine similarity ranking to select the top-5 examples as contexts. In the third row, we investigate the impact of the cluster on in-context example selection, using the closest examples to each cluster centroid to construct the candidate pool. Compared to the baseline, the results show that compacting the search space not only reduces annotation cost but also improves model performance.", + "bbox": [ + 212, + 568, + 785, + 688 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Effects of Sort. As shown in row 4 of Table 2, we experimentally demonstrate that integrating our sorting mechanism results in $1.9\\mathrm{mIoU}$ improvement compared to directly using the closest examples to each cluster centroid. The inherent reason is that centroid examples describe a limited feature space [33, 46], making it challenging to support a variety of different contexts for test instances. In contrast, our sorting mechanism introduces the nearest and farthest examples, which constructs more diverse demonstrations for ICL learning.", + "bbox": [ + 212, + 688, + 785, + 794 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Effects of Adaptive Search Module. Finally, we further explore the effectiveness of our adaptive search module in the last row of Table 2. Compared to using a simple similarity sorting method, our approach shows an improve", + "bbox": [ + 212, + 794, + 785, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "W. Suo et al.", + "bbox": [ + 271, + 114, + 361, + 126 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/0335616d3ba896a9e0ae905786a2e2d83370134294f1d86fabcd7226f7082d11.jpg", + "table_caption": [ + "Table 4: The results on PASCAL-5 $^i$ [32] Table 5: The results on iSALD-5 $^i$ [48] based on instance segmentation. based on semantic segmentation." + ], + "table_footnote": [], + "table_body": "
Model1-shot5-shot
SegGPT† w/R [45]46.546.9
SegGPT† w/ [56]47.148.5
SegGPT w/Ours50.551.4
", + "bbox": [ + 233, + 184, + 475, + 247 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/3f0290a7ef55b98f56d6925d6144bd07b475c5a67db29807646c38083fe20cf8.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Model1-shot5-shot
SegGPT† w/R [45]29.830.6
SegGPT† w/ [56]34.735.4
SegGPT w/Ours37.939.5
", + "bbox": [ + 509, + 184, + 751, + 246 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "ment of 3.4 points. Meanwhile, benefiting from the diverse candidate pool, our SCS achieves state-of-the-art performance and outperforms the baseline by a considerable margin.", + "bbox": [ + 212, + 273, + 787, + 319 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "5.4 Different Model Settings", + "text_level": 1, + "bbox": [ + 215, + 340, + 465, + 358 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "In this section, we explore several alternative model settings to further discuss the proposed method. All experiments are conducted on the COCO- $20^{i}$ with the SegGPT model. In the first two rows, we first compare the performance using different metrics. Specifically, we use the rule-based methods to select the contextual examples from the complete training set, which including Entropy sorting [11] and Foreground sorting [53]. In particular, \"Entropy\" is executed by calculating the maximum entropy values. The \"Foreground\" refers to a two-stage approach [53]. It uses an object detector [42] to extract the foreground boxes and the foreground similarity is used as the metric. From the 1-shot and 5-shot results, it can be observed that our SCS can dramatically outperform these methods by a large margin. In rows 3-4, we construct the experiments with different feature encoders. We replace the CLIP-ViT with ViT [8] and CLIP-resnet [27] to obtain image feature vectors. These extracted features would be utilized to construct the candidate pool and perform the adaptive searching. The results show that our SCS is robust for different image encoders. Finally, in rows 5-6 of Table 3, we replace our selection mechanism with several alternative manners. However, these approaches mostly involve complex computation. For example, the Density-sort [2] requires additional computations on the order of $O(MN)$ in each iteration to update density lists, where $M$ is the number of clusters and $N$ is the size of the unlabeled data. Instead, our method uses the basic clustering algorithm and obtains a comparable performance.", + "bbox": [ + 212, + 367, + 787, + 686 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "5.5 Further Discussion", + "text_level": 1, + "bbox": [ + 215, + 708, + 418, + 723 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "More Segmentation tasks. In this section, we evaluate the effect of our method on more segmentation tasks and datasets. Because MAE-VQGAN cannot distinguish between individual instances of objects within the same class, we construct more experiments with the SegGPT model. Specifically, we report instance segmentation results on PASCAL-5i in Table 4 and aerial image semantic segmentation results on iSALD-5i in Table 5. The above results indicate that the SCS can enhance the segmentation ability for various real-world scenarios.", + "bbox": [ + 212, + 733, + 787, + 840 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "Visual Prompt Selection for In-Context Learning Segmentation", + "bbox": [ + 308, + 113, + 730, + 130 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 127 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/07ebfd07fd83de029507aa2c73371a7b1fef1abe48c42da646e31893c0e33c0e.jpg", + "image_caption": [ + "Fig.5: Qualitative results under 1-shot setting on COCO- $20^{i}$ . The similarity-based selection method is viewed as our baseline. The green regions, red regions and purple regions are example masks, predicted masks and ground-truth masks." + ], + "image_footnote": [], + "bbox": [ + 264, + 143, + 738, + 407 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Qualitative Results. As shown in Fig. 5, to further analyze and verify the proposed method, we visualized the qualitative segmentation results on COCO $_{20^i}$ with the SegGPT. The first and third columns represent the visual prompts (highlighted in green) based on similarity sorting and our SCS method, respectively. The second and fourth columns show the corresponding segmentation results (highlighted in red). The last column represents the ground truth mask (highlighted in purple). As illustrated in the first row, the baseline model selects the most similar context to the query (i.e., chopping board, knife and carrot), while the segment result is not satisfactory, as the model can not recognize the knife in the image. Instead, Our method selects an image from the candidate pool that accurately segments out the mask. The above results demonstrate that similarity sorting is not the only golden standard and improving the diversity of context prompts can boost the segmentation results.", + "bbox": [ + 212, + 487, + 787, + 684 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "6 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 705, + 349, + 720 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In this paper, we focus on the example selection method for the ICL-based segmentation frameworks. By building comprehensive comparisons, we empirically demonstrate that diversity of contexts is a key factor in guiding segmentation and boosting performance. Beyond the above insights, we propose a new stepwise context search method that improves the diversity of the candidate pool and significantly alleviates the annotation cost. We hope that this work can promote further research to better understand in-context learning in computer vision.", + "bbox": [ + 212, + 733, + 787, + 840 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "W. Suo et al.", + "bbox": [ + 271, + 114, + 361, + 126 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgement. This work was supported by National Natural Science Foundation of China (No.62102323), the Innovation Capability Support Program of Shaanxi(Program No. 2023KJXX-142) and National Natural Science Foundation of China (U23B2013).", + "bbox": [ + 212, + 146, + 787, + 205 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 215, + 232, + 325, + 248 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Agrawal, S., Zhou, C., Lewis, M., Zettlemoyer, L., Ghazvininejad, M.: In-context examples selection for machine translation. arXiv preprint arXiv:2212.02437 (2022)", + "2. et al., W.: \"unsupervised selective labeling for more effective semi-supervised learning\". In: ECCV 2022. pp. 427-445. Springer (2022)", + "3. et al, W.: \"seggspt: Segmenting everything in context\". In ICCV 2023 (2023)", + "4. Anderson, P., He, X., Buehler, C., Teney, D., Johnson, M., Gould, S., Zhang, L.: Bottom-up and top-down attention for image captioning and visual question answering. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 6077-6086 (2018)", + "5. Badrinarayanan, V., Kendall, A., Cipolla, R.: Segnet: A deep convolutional encoder-decoder architecture for image segmentation. IEEE transactions on pattern analysis and machine intelligence 39(12), 2481-2495 (2017)", + "6. Bar, A., Gandelsman, Y., Darrell, T., Globerson, A., Efros, A.: Visual prompting via image inpainting. Advances in Neural Information Processing Systems 35, 25005-25017 (2022)", + "7. Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J.D., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., et al.: Language models are few-shot learners. Advances in neural information processing systems 33, 1877-1901 (2020)", + "8. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929 (2020)", + "9. Everingham, M., Van Gool, L., Williams, C.K., Winn, J., Zisserman, A.: The Pascal visual object classes (voc) challenge. International journal of computer vision 88, 303-338 (2010)", + "0. Hariharan, B., Arbeláez, P., Bourdev, L., Maji, S., Malik, J.: Semantic contours from inverse detectors. In: 2011 international conference on computer vision. pp. 991-998. IEEE (2011)", + "1. Haussmann, E., Fenzi, M., Chitta, K., Ivanecky, J., Xu, H., Roy, D., Mittel, A., Koumchatzky, N., Farabet, C., Alvarez, J.M.: Scalable active learning for object detection. In: 2020 IEEE intelligent vehicles symposium (iv). pp. 1430-1435. IEEE (2020)", + "2. He, K., Gkioxari, G., Dollár, P., Girshick, R.: Mask r-cnn. In: Proceedings of the IEEE international conference on computer vision. pp. 2961-2969 (2017)", + "3. Hong, S., Cho, S., Nam, J., Lin, S., Kim, S.: Cost aggregation with 4d convolutional swim transformer for few-shot segmentation. In: European Conference on Computer Vision. pp. 108-126. Springer (2022)", + "4. Hongjin, S., Kasai, J., Wu, C.H., Shi, W., Wang, T., Xin, J., Zhang, R., Ostendorf, M., Zettlemoyer, L., Smith, N.A., et al.: Selective annotation makes language models better few-shot learners. In: The Eleventh International Conference on Learning Representations (2022)" + ], + "bbox": [ + 225, + 265, + 784, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "Visual Prompt Selection for In-Context Learning Segmentation", + "bbox": [ + 310, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "15. Hore, A., Ziou, D.: Image quality metrics: Psnr vs. sSIM. In: 2010 20th international conference on pattern recognition. pp. 2366-2369. IEEE (2010)", + "16. Hu, H., Li, Z., Li, L., Yang, H., Zhu, H.: Classification of very high-resolution remote sensing imagery using a fully convolutional network with global and local context information enhancements. IEEE Access 8, 14606-14619 (2020)", + "17. Kingma, D.P., Ba, J.: Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980 (2014)", + "18. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., et al.: Segment anything. arXiv preprint arXiv:2304.02643 (2023)", + "19. Li, X., Lv, K., Yan, H., Lin, T., Zhu, W., Ni, Y., Xie, G., Wang, X., Qiu, X.: Unified demonstration retriever for in-context learning. arXiv preprint arXiv:2305.04320 (2023)", + "20. Li, Y., Padmanabhan, A., Zhao, P., Wang, Y., Xu, G.H., Netravali, R.: Reducto: On-camera filtering for resource-efficient real-time video analytics. In: Proceedings of the Annual conference of the ACM Special Interest Group on Data Communication on the applications, technologies, architectures, and protocols for computer communication. pp. 359-376 (2020)", + "21. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollar, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13. pp. 740-755. Springer (2014)", + "22. Liu, J., Shen, D., Zhang, Y., Dolan, B., Carin, L., Chen, W.: What makes good in-context examples for gpt-3? arXiv preprint arXiv:2101.06804 (2021)", + "23. Long, J., Shelhamer, E., Darrell, T.: Fully convolutional networks for semantic segmentation. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 3431-3440 (2015)", + "24. Luo, G., Zhou, Y., Sun, X., Cao, L., Wu, C., Deng, C., Ji, R.: Multi-task collaborative network for joint referring expression comprehension and segmentation. In: Proceedings of the IEEE/CVF Conference on computer vision and pattern recognition. pp. 10034-10043 (2020)", + "25. Lyu, X., Fang, Y., Tong, B., Li, X., Zeng, T.: Multiscale normalization attention network for water body extraction from remote sensing imagery. Remote Sensing 14(19) (2022). https://doi.org/10.3390/rs14194983, https://www.mdpi.com/2072-4292/14/19/4983", + "26. Min, S., Lyu, X., Holtzman, A., Artetxe, M., Lewis, M., Hajishirzi, H., Zettlemoyer, L.: Rethinking the role of demonstrations: What makes in-context learning work? arXiv preprint arXiv:2202.12837 (2022)", + "27. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021)", + "28. Reisenhofer, R., Bosse, S., Kutyniok, G., Wiegand, T.: Aaar wavelet-based perceptual similarity index for image quality assessment. Signal Processing: Image Communication 61, 33-43 (2018)", + "29. Ronneberger, O., Fischer, P., Brox, T.: U-net: Convolutional networks for biomedical image segmentation. In: Medical Image Computing and Computer-Assisted Intervention-MICCAI 2015: 18th International Conference, Munich, Germany, October 5-9, 2015, Proceedings, Part III 18. pp. 234-241. Springer (2015)", + "30. Rubin, O., Herzig, J., Berant, J.: Learning to retrieve prompts for in-context learning. arXiv preprint arXiv:2112.08633 (2021)" + ], + "bbox": [ + 215, + 147, + 784, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "W. Suo et al.", + "bbox": [ + 271, + 114, + 361, + 126 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "31. Rubin, O., Herzig, J., Berant, J.: Learning to retrieve prompts for in-context learning. In: Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies. pp. 2655-2671 (2022)", + "32. Shaban, A., Bansal, S., Liu, Z., Essa, I., Boots, B.: One-shot learning for semantic segmentation. arXiv preprint arXiv:1709.03410 (2017)", + "33. Shen, F., Zhao, L., Du, W., Zhong, W., Qian, F.: Large-scale industrial energy systems optimization under uncertainty: A data-driven robust optimization approach. Applied Energy 259, 114199 (2020)", + "34. Shi, H., Li, H., Meng, F., Wu, Q.: Key-word-aware network for referring expression image segmentation. In: Proceedings of the European Conference on Computer Vision (ECCV). pp. 38-54 (2018)", + "35. Sun, L., Yang, K., Hu, X., Hu, W., Wang, K.: Real-time fusion network forrgb-d semantic segmentation incorporating unexpected obstacle detection for road-driving images. IEEE robotics and automation letters 5(4), 5558-5565 (2020)", + "36. Sun, M., Suo, W., Wang, P., Niu, K., Liu, L., Lin, G., Zhang, Y., Wu, Q.: An adaptive correlation filtering method for text-based person search. International Journal of Computer Vision pp. 1-16 (2024)", + "37. Sun, Y., Chen, Q., Wang, J., Wang, J., Li, Z.: Exploring effective factors for improving visual in-context learning. arXiv preprint arXiv:2304.04748 (2023)", + "38. Suo, W., Sun, M., Liu, W., Gao, Y., Wang, P., Zhang, Y., Wu, Q.: S3c: Semi-supervised vqa natural language explanation via self-critical learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2646-2656 (2023)", + "39. Suo, W., Sun, M., Niu, K., Gao, Y., Wang, P., Zhang, Y., Wu, Q.: A simple and robust correlation filtering method for text-based person search. In: European conference on computer vision. pp. 726-742. Springer (2022)", + "40. Suo, W., Sun, M., Wang, P., Zhang, Y., Wu, Q.: Rethinking and improving feature pyramids for one-stage referring expression comprehension. IEEE Transactions on Image Processing 32, 854-864 (2022)", + "41. Vasudevan, A.B., Dai, D., Van Gool, L.: Semantic object prediction and spatial sound super-resolution with binaural sounds. In: European conference on computer vision. pp. 638-655. Springer (2020)", + "42. Wang, W., Dai, J., Chen, Z., Huang, Z., Li, Z., Zhu, X., Hu, X., Lu, T., Lu, L., Li, H., et al.: Intermimage: Exploring large-scale vision foundation models with deformable convolutions. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14408-14419 (2023)", + "43. Wang, X., Ma, H., You, S.: Deep clustering for weakly-supervised semantic segmentation in autonomous driving scenes. Neurocomputing 381, 20-28 (2020)", + "44. Wang, X., Wang, W., Cao, Y., Shen, C., Huang, T.: Images speak in images: A generalist painter for in-context visual learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6830-6839 (2023)", + "45. Wang, X., Zhang, X., Cao, Y., Wang, W., Shen, C., Huang, T.: Seggpt: Segmenting everything in context. arXiv preprint arXiv:2304.03284 (2023)", + "46. Wang, X., Lian, L., Yu, S.X.: Unsupervised selective labeling for more effective semi-supervised learning. In: European Conference on Computer Vision. pp. 427-445. Springer (2022)", + "47. Wang, Z., Bovik, A.C., Sheikh, H.R., Simoncelli, E.P.: Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing 13(4), 600-612 (2004)" + ], + "bbox": [ + 212, + 146, + 787, + 839 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "Visual Prompt Selection for In-Context Learning Segmentation", + "bbox": [ + 310, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "48. Waqas Zamir, S., Arora, A., Gupta, A., Khan, S., Sun, G., Shahbaz Khan, F., Zhu, F., Shao, L., Xia, G.S., Bai, X.: isaid: A large-scale dataset for instance segmentation in aerial images. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops. pp. 28-37 (2019)", + "49. Wei, J., Wang, X., Schuurmans, D., Bosma, M., Xia, F., Chi, E., Le, Q.V., Zhou, D., et al.: Chain-of-thought prompting elicits reasoning in large language models. Advances in Neural Information Processing Systems 35, 24824-24837 (2022)", + "50. Yao, X., Cao, Q., Feng, X., Cheng, G., Han, J.: Scale-aware detailed matching for few-shot aerial image semantic segmentation. IEEE Transactions on Geoscience and Remote Sensing 60, 1-11 (2021)", + "51. Ye, J., Wu, Z., Feng, J., Yu, T., Kong, L.: Compositional exemplars for in-context learning. arXiv preprint arXiv:2302.05698 (2023)", + "52. Yu, Q., Gao, Y., Zheng, Y., Zhu, J., Dai, Y., Shi, Y.: Crossover-net: Leveraging vertical-horizontal crossover relation for robust medical image segmentation. Pattern Recognition 113, 107756 (2021)", + "53. Yu, Y., Yang, X., Li, Q., Zhou, Y., Da, F., Yan, J.: H2rbox-v2: Incorporating symmetry for boosting horizontal box supervised oriented object detection. In: Thirty-seventh Conference on Neural Information Processing Systems (2023)", + "54. Zendel, O., Schörghuber, M., Rainer, B., Murschitz, M., Beleznai, C.: Unifying panoptic segmentation for autonomous driving. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 21351-21360 (2022)", + "55. Zhang, L., Zhang, L., Mou, X., Zhang, D.: Fsim: A feature similarity index for image quality assessment. IEEE transactions on Image Processing 20(8), 2378-2386 (2011)", + "56. Zhang, Y., Zhou, K., Liu, Z.: What makes good examples for visual in-context learning? arXiv preprint arXiv:2301.13670 (2023)", + "57. Zhang, Y., Liu, H., Hu, Q.: Transfuse: Fusing transformers and cnns for medical image segmentation. In: Medical Image Computing and Computer Assisted Intervention-MICCAI 2021: 24th International Conference, Strasbourg, France, September 27-October 1, 2021, Proceedings, Part I 24. pp. 14-24. Springer (2021)", + "58. Zhang, Z., Zhang, A., Li, M., Smola, A.: Automatic chain of thought prompting in large language models. arXiv preprint arXiv:2210.03493 (2022)" + ], + "bbox": [ + 215, + 146, + 785, + 590 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "W. Suo et al.", + "bbox": [ + 271, + 114, + 361, + 126 + ], + "page_idx": 17 + } +] \ No newline at end of file diff --git a/2024/Rethinking and Improving Visual Prompt Selection for In-Context Learning Segmentation Framework/c5a744a9-bd8f-4567-93ec-5cbfe907a7d5_model.json b/2024/Rethinking and Improving Visual Prompt Selection for In-Context Learning Segmentation Framework/c5a744a9-bd8f-4567-93ec-5cbfe907a7d5_model.json new file mode 100644 index 0000000000000000000000000000000000000000..2b174c5b232902811afd74ced52fbde19630f68a --- /dev/null +++ b/2024/Rethinking and Improving Visual Prompt Selection for In-Context Learning Segmentation Framework/c5a744a9-bd8f-4567-93ec-5cbfe907a7d5_model.json @@ -0,0 +1,2370 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.229, + 0.141, + 0.776, + 0.187 + ], + "angle": 0, + "content": "Rethinking and Improving Visual Prompt Selection for In-Context Learning Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.218, + 0.78, + 0.261 + ], + "angle": 0, + "content": "Wei Suo\\(^{1,3,\\star}\\), Lanqing Lai\\(^{1,\\star}\\), Mengyang Sun\\(^{2,3}\\), Hanwang Zhang\\(^{3}\\), Peng Wang\\(^{1,\\dagger}\\), and Yanning Zhang\\(^{1}\\)" + }, + { + "type": "text", + "bbox": [ + 0.233, + 0.271, + 0.768, + 0.299 + ], + "angle": 0, + "content": "1 School of Computer Science and Ningbo Institute, Northwestern Polytechnical University, Xi'an, China" + }, + { + "type": "text", + "bbox": [ + 0.235, + 0.3, + 0.765, + 0.313 + ], + "angle": 0, + "content": "\\(^{2}\\) School of Cybersecurity, Northwestern Polytechnical University, Xi'an, China" + }, + { + "type": "text", + "bbox": [ + 0.245, + 0.314, + 0.756, + 0.341 + ], + "angle": 0, + "content": "3 School of Computer Science, Nanyang Technological University, Singapore {suwei1994, lailanqing, sunmenmian}@mail.nwpu.edu.cn" + }, + { + "type": "list", + "bbox": [ + 0.233, + 0.271, + 0.768, + 0.341 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.378, + 0.745, + 0.628 + ], + "angle": 0, + "content": "Abstract. As a fundamental and extensively studied task in computer vision, image segmentation aims to locate and identify different semantic concepts at the pixel level. Recently, inspired by In-Context Learning (ICL), several generalist segmentation frameworks have been proposed, providing a promising paradigm for segmenting specific objects. However, existing works mostly ignore the value of visual prompts or simply apply similarity sorting to select contextual examples. In this paper, we focus on rethinking and improving the example selection strategy. By comprehensive comparisons, we first demonstrate that ICL-based segmentation models are sensitive to different contexts. Furthermore, empirical evidence indicates that the diversity of contextual prompts plays a crucial role in guiding segmentation. Based on the above insights, we propose a new stepwise context search method. Different from previous works, we construct a small yet rich candidate pool and adaptively search the well-matched contexts. More importantly, this method effectively reduces the annotation cost by compacting the search space. Extensive experiments show that our method is an effective strategy for selecting examples and enhancing segmentation performance. https://github.com/LanqingL/SCS" + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.641, + 0.741, + 0.656 + ], + "angle": 0, + "content": "Keywords: In-Context Learning \\(\\cdot\\) Visual Prompt \\(\\cdot\\) Image Segmentation" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.681, + 0.377, + 0.697 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.712, + 0.788, + 0.804 + ], + "angle": 0, + "content": "Image segmentation is a fundamental task in computer vision, aiming to locate and identify different semantic concepts at the pixel level. It is essential in multiple applications such as autonomous driving [43,54], video surveillance [20,36,39,41] and complex reasoning [24,34,40]. To achieve precise segmentation, researchers have proposed a variety of segmentation models [5,12,23,29]. Although these methods have made considerable progress, they require training" + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.811, + 0.384, + 0.825 + ], + "angle": 0, + "content": "† Corresponding author" + }, + { + "type": "page_footnote", + "bbox": [ + 0.222, + 0.827, + 0.551, + 0.84 + ], + "angle": 0, + "content": "* These authors contributed equally to this work." + }, + { + "type": "list", + "bbox": [ + 0.218, + 0.811, + 0.551, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.363, + 0.127 + ], + "angle": 0, + "content": "W. Suo et al." + }, + { + "type": "image", + "bbox": [ + 0.29, + 0.144, + 0.718, + 0.381 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.391, + 0.788, + 0.462 + ], + "angle": 0, + "content": "Fig. 1: Comparison of traditional and our method. (a) Existing works rely on dense annotation to build the exempling space. Then, they use the similarity sorting manner to select examples for given queries. (b) Instead, our method significantly alleviates the costs of annotation by searching the typical examples. Moreover, a novel adaptive search module is designed to further select well-matched contexts." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.493, + 0.788, + 0.599 + ], + "angle": 0, + "content": "the specialist model for each segmentation task. Recently, inspired by In-Context Learning (ICL) [26,30,49], several generalist segmentation frameworks have been proposed and achieved impressive performance such as MAE-VQGAN [6] and SegGPT [45]. Different from the traditional frameworks, ICL-based segmentation methods establish a promising paradigm for localizing vision concepts by feeding one or a few examples as references during inference. Meanwhile, the methodology offers a general solution for various real-world scenarios [44]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.599, + 0.789, + 0.844 + ], + "angle": 0, + "content": "ICL learning is a relatively new research line in computer vision, while the fundamental ideas can be traced back to the Natural Language Processing (NLP) field [6, 44]. With the development of Large Language Models (LLM), numerous studies have already proved that selecting a suitable context is crucial for following human instruction and improving performance [1, 14, 31]. However, related researches about this topic is still limited in our community. As shown in Fig.1 (a), existing works typically randomly select examples [44] or simply apply similarity sorting to construct contextual prompts [37, 56]. The above insights motivate us to question: 1) Whether different contexts would significantly impact performance? 2) What is the critical factor of visual prompt selection for ICL-based segmentation models? In this paper, we first try to answer the above questions. By comprehensive comparisons, we find that the ICL-based models are sensitive to different examples. With multiple experiments, the performance gap between using different contextual examples even exceeds 5 points. Furthermore, we synthetically investigate the effectiveness of the existing prompt selection strategy. Surprisingly, although the similarity measure (i.e., choosing" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.311, + 0.115, + 0.732, + 0.131 + ], + "angle": 0, + "content": "Visual Prompt Selection for In-Context Learning Segmentation" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.223 + ], + "angle": 0, + "content": "the most similar examples as contexts) provides a relatively effective approach for demonstration search, assembling the most dissimilar contexts results in better performance in \\(\\sim 40\\%\\) of the test instances. To reveal the potential reasons, we empirically demonstrate that the diversity of contexts is a crucial factor in guiding segmentation and enhancing performance (more details in Sec. 3)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.223, + 0.785, + 0.449 + ], + "angle": 0, + "content": "Beyond the above discussions, we rethink the role of contextual examples and propose a new Stepwise Context Search method, which is called SCS for short. As shown in Fig. 1 (b), the SCS focuses on maintaining the diversity of the candidate pool and adaptively searching the well-matched contexts. Meanwhile, our method significantly reduces the annotation cost by searching the typical examples. Specifically, given a large unlabeled database, we first apply a basic cluster algorithm to compact the search space and improve annotation efficiency. Then, we extract the two most dissimilar members in each cluster as candidate examples. By collecting these representative members from all clusters, we construct a small yet rich candidate pool. Furthermore, considering the significant impact of examples on performance, we further introduce a novel adaptive search module based on reinforcement learning. It heuristically utilizes segmentation scores as rewards, guiding the model in determining final demonstrations. Extensive experiments across various datasets have shown the effectiveness of our method. To summarize, our contributions are as follows." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.455, + 0.784, + 0.499 + ], + "angle": 0, + "content": "- We take a step forward to systematically analyze the influence of example selection on the ICL-based segmentation frameworks and empirically demonstrate that the diversity of contexts is a key factor in guiding segmentation." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.5, + 0.785, + 0.543 + ], + "angle": 0, + "content": "- We propose a new SCS method to efficiently select visual prompts by constructing a small yet rich candidate pool. Meanwhile, an adaptive search module is designed to further choose the well-matched demonstrations." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.544, + 0.785, + 0.589 + ], + "angle": 0, + "content": "- The proposed model significantly improves the segmentation performance of ICL-based frameworks and achieves state-of-the-art on PASCAL-5\\(^i\\) [32], COCO-20\\(^i\\) [21] and iSALD-5\\(^i\\) [50]." + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.455, + 0.785, + 0.589 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.609, + 0.383, + 0.625 + ], + "angle": 0, + "content": "2 Related work" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.637, + 0.433, + 0.653 + ], + "angle": 0, + "content": "2.1 Visual Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.659, + 0.788, + 0.841 + ], + "angle": 0, + "content": "As a fundamental task in computer vision, visual segmentation entails locating and identifying different semantic concepts at the pixel level. Numerous works for specific tasks have continued to flourish over the years. For example, CrossoverNet [52] learns the vertical and horizontal crossover relation for medical image segmentation. RFNet [35] fuses RGB and depth information which is designed for real-time road-driving segmentation. [16] proposes DGEN, a densely connected global entropy network in which noise is suppressed by context information losses. MSANet [25] accurately extracts water bodies from remote sensing scenes by utilizing multiscale attention. Recently, inspired by ICL learning, several general segmentation methods have been proposed [18,44,45]. Different from these specialist models, the paradigm can locate visual concepts by feeding a few examples, known as the In-Context Learning (ICL) segmentation frameworks." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.363, + 0.128 + ], + "angle": 0, + "content": "W. Suo et al." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.147, + 0.433, + 0.163 + ], + "angle": 0, + "content": "2.2 In-Context Learning" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.17, + 0.788, + 0.352 + ], + "angle": 0, + "content": "In-context learning (ICL) is a new learning paradigm that initially emerged from GPT-3 [7]. It presents a text fill-in-the-blank problem with examples, eliminating the necessity for fine-tuning model parameters on downstream NLP tasks. In the vision domain, MAE-VQGAN [6] transforms various visual tasks into grid inpainting problems and proposes the first ICL-based segmentation framework. Different from the previous segmentation methods, this new paradigm provides a promising manner to segment arbitrary objects with a few examples. Painter [44] adopts masked image modeling on continuous pixels, unifying the output form across different visual tasks. As a variant of Painter, SegGPT [45] integrates visual segmentation tasks into in-context coloring problems and demonstrates powerful capabilities in target segmentation. Beyond the previous works, we present comprehensive experiments on example selection for visual ICL models." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.373, + 0.513, + 0.388 + ], + "angle": 0, + "content": "2.3 Contextual Example Selection" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.396, + 0.789, + 0.579 + ], + "angle": 0, + "content": "With the development of LLMs, the selection of contextual examples is important for ICL learning [1,30,51]. The emerging studies can be categorized into rule-based unsupervised methods and supervised methods. KATE [22] uses the KNN mechanism to enhance a retriever for selecting in-context examples. EPR [30] adopts a contrastive learning strategy to train a dense retriever. UDR [19] explores the process of retrieval demonstrations for various tasks in a unified manner. However, few studies investigate these effects in the vision domain. Current works typically select examples randomly [44] or simply apply similarity sorting to construct contextual demonstrations [37,56]. Different from the previous rule-based methods, we introduce an adaptive search method with a small yet rich candidate pool. This strategy effectively reduces the annotation cost and boosts segmentation performance." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.599, + 0.371, + 0.617 + ], + "angle": 0, + "content": "3 Preliminary" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.629, + 0.788, + 0.706 + ], + "angle": 0, + "content": "In this section, we would first present a simple overview of the ICL-based segmentation frameworks. Then, we construct extensive experiments to comprehensively analyze the influence of different examples and the effectiveness of the existing selection strategies. Finally, we experimentally proved that contextual diversity is an important factor for guiding segmentation." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.726, + 0.363, + 0.741 + ], + "angle": 0, + "content": "3.1 Background" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.841 + ], + "angle": 0, + "content": "The design of segmentation models has achieved tremendous development in recent years [20, 41, 52, 57]. However, these solutions require expensive annotation cost and face challenges in adapting to new scenarios. In contrast, the ICL-based segmentation frameworks [6, 45] can implement various specific tasks by leveraging several segmented examples. The convenient and versatile vision models provide a promising perspective for our community." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.31, + 0.115, + 0.732, + 0.131 + ], + "angle": 0, + "content": "Visual Prompt Selection for In-Context Learning Segmentation" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.283 + ], + "angle": 0, + "content": "As shown in Fig. 1, the goal of ICL-based segmentation methods is to predict segmentation regions based on exhibited one or a few prompts. By reorganizing the input format, the models can imitate provided contexts to mask specific object regions. Specifically, given a query image \\(I_q\\), these segmentation frameworks first select \\(n\\) example images \\(E_d = \\{I_i^d\\}_{i=1}^n\\) with corresponding masks \\(G_d = \\{G_i^d\\}_{i=1}^n\\) as contextual demonstrations from a huge database. Here, \\(I_i^d\\) and \\(G_i^d\\) denote the \\(i\\)-th selected example and corresponding mask. By reformating the input, models can perform multiple segmentation tasks across different scenarios. The above operations can be formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.434, + 0.299, + 0.786, + 0.316 + ], + "angle": 0, + "content": "\\[\n\\hat {I} _ {m} = \\mathbf {S} \\left(I _ {q}, E _ {d}, G _ {d}\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.332, + 0.789, + 0.606 + ], + "angle": 0, + "content": "where the \\(\\mathbf{S}(\\cdot)\\) and \\(\\hat{I}_m\\) are the generalist segmentation model and predicted mask results, respectively. Although the above calculations provide a convenient way to segment objects at arbitrary granularities, the success of the paradigm greatly relies on the given contextual examples. In fact, the basic idea of ICL-based vision models comes from the NLP field, in which it has been proved that contextual prompts play an important role in executing reasoning and boosting performance. However, in the vision community, existing approaches mostly simply choose the demonstrations (i.e., \\(E_d\\) and corresponding \\(G_d\\)) based on random sampling [44] or rely on similarity sorting [37,56]. To explore the impact of the visual prompts on performance, we construct a series of quantitative analyses on the PASCAL dataset [32] with the state-of-the-art ICL-based segmentation model SegGPT [45]. Note that the SegGPT involves two training stages (i.e., pre-training and in-context tuning). In the tuning process, the SegGPT depends on the whole annotation (including base and new classes) to improve the performance. Since the main advantage of the ICL-based framework is to implement few-shot tasks, we only use the pre-trained weights provided by the SegGPT [3] to construct experiments. Following [6,44], the mean intersection over union (mIoU) is adopted as the evaluation metric." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.638, + 0.524, + 0.654 + ], + "angle": 0, + "content": "3.2 Influence of Different Examples" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.789, + 0.841 + ], + "angle": 0, + "content": "As mentioned above, research on the influence of different examples in segmentation remains insufficient. In this paper, we aim to fill this gap by conducting experiments to quantitatively analyze the impact of different demonstrations. Specifically, we randomly select examples from the entire training set with 5 different seeds. As shown in Fig. 2 (a), we report the best, worst, and average results across the 5 experiments. It reveals that allocating different contexts would dramatically affect the segmentation scores under the 1-shot and 5-shot settings. In particular, the best results can surpass the worst by margins of 5.6 and 4.1 points, respectively. Due to the huge performance discrepancy between different contextual prompts, designing an automated method for searching well-matched examples emerges as a critical challenge." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.363, + 0.127 + ], + "angle": 0, + "content": "W. Suo et al." + }, + { + "type": "image", + "bbox": [ + 0.216, + 0.147, + 0.473, + 0.294 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.346, + 0.295, + 0.365, + 0.307 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.143, + 0.752, + 0.294 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.633, + 0.295, + 0.654, + 0.307 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.318, + 0.788, + 0.388 + ], + "angle": 0, + "content": "Fig. 2: The influence of different context selection. (a) Randomly sampling contextual examples with 5 runs under 1-shot and 5-shot setting on PASCAL. (b) For each instance, the nearest and farthest examples are retrieved as visual prompts across different similarity-based sorting. Surprisingly, the most dissimilar examples achieved better performance on \\(\\sim 40\\%\\) of the test samples." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.415, + 0.544, + 0.43 + ], + "angle": 0, + "content": "3.3 Example Selection with Similarity" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.436, + 0.789, + 0.618 + ], + "angle": 0, + "content": "To the best of our knowledge, the related researches to our work are [37, 56], where they simply rely on cosine similarity to select examples with qualitative analysis. In this paper, we try to rethink the strategy through a quantitative discussion across different similarity computation manners. Specifically, we explore two types of image similarity calculation strategies including appearance-level and semantic-level. For the appearance-level comparison, PSNR [47], SSIM [15], FSIM [55] and HaarPSI [28] obtain the similarity between the test instance and the entire training set. On the other hand, for semantic-level comparison, we follow [37, 56] to utilize the off-the-shelf image encoders (i.e., CLIP [27] and ViT [8]) for feature extraction. To verify the effectiveness of these similarity-based approaches, we carefully select the examples based on cosine similarity sorting." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.619, + 0.79, + 0.77 + ], + "angle": 0, + "content": "As shown in Fig. 2 (b), each metric is represented by two bars including nearest \\(>\\) farthest (yellow bar) and nearest \\(<\\) farthest (green bar). We take yellow as an example, it denotes the proportion of instances where using the most similar examples (nearest) as contexts results in a better IoU than using the most dissimilar ones (farthest), and the green bar indicates the opposite. Surprisingly, we observed that assembling the most dissimilar contexts leads to superior performance for numerous test instances. There is a substantial portion of the test samples (\\(\\sim 40\\%\\)) that achieve better IoU scores by utilizing the farthest examples. The above results indicate that using similarity measures is not the only way to search for contextual examples." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.788, + 0.54, + 0.804 + ], + "angle": 0, + "content": "3.4 Example Selection with Diversity" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.789, + 0.842 + ], + "angle": 0, + "content": "Based on the above analyses, it is imperative to explore a more effective way for searching visual prompts. In Fig. 2 (b), we find that applying the most dissimilar" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.31, + 0.115, + 0.732, + 0.13 + ], + "angle": 0, + "content": "Visual Prompt Selection for In-Context Learning Segmentation" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "image", + "bbox": [ + 0.216, + 0.145, + 0.443, + 0.304 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.235, + 0.308, + 0.394, + 0.334 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.235, + 0.334, + 0.474, + 0.348 + ], + "angle": 0, + "content": "(a) Example sampling with SegGPT [45]." + }, + { + "type": "image", + "bbox": [ + 0.498, + 0.143, + 0.731, + 0.304 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.547, + 0.308, + 0.683, + 0.334 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.509, + 0.335, + 0.778, + 0.348 + ], + "angle": 0, + "content": "(b) Example sampling with MAE-VQGAN [6]." + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.359, + 0.788, + 0.402 + ], + "angle": 0, + "content": "Fig. 3: Diversity vs Similarity. Based on similarity sorting, the performance of the two Nearest examples (NN), the two Farthest examples (FF), and the Nearest example with the Farthest example (NF) as visual prompts are shown." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.43, + 0.789, + 0.492 + ], + "angle": 0, + "content": "examples can also yield satisfactory performance for some instances. The results motivate us to investigate whether combining the nearest and farthest contexts can obtain a performance boost. Consequently, we further construct a series of experiments to verify the hypothesis across different ICL-based models." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.491, + 0.789, + 0.613 + ], + "angle": 0, + "content": "More specifically, based on similarity sorting, we use two Nearest examples (NN), two Farthest examples (FF), and a combination of the Nearest example and the Farthest example (NF) as contextual demonstrations, respectively. As shown in Fig. 3, we report related results across different similarity measure manners. It can be observed that the performance using NF examples significantly surpasses the other two selection strategies in all measures. Regarding ViT, the NF achieves better performance that outperforms equipping the two nearest examples by a margin of 2.4 points." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.634, + 0.351, + 0.648 + ], + "angle": 0, + "content": "3.5 Discussion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.659, + 0.788, + 0.749 + ], + "angle": 0, + "content": "In this section, we would try to summarize the above findings with a deeper discussion. Based on the insights presented in Sec. 3.2, we find that the ICL-based segmentation framework is sensitive to contextual examples. Given the different visual prompts, the performance gap is even over 5 points on the mIoU. Therefore, building an effective prompt selection method is important for boosting the segmentation results and achieving stable performance." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.789, + 0.841 + ], + "angle": 0, + "content": "To further explore the key factors that influence example selection, we design experiments in Sec. 3.3 and Sec. 3.4. The results demonstrate that combining both similar and dissimilar demonstrations can significantly enhance the IoU scores. We argue that the underlying reason is that this manner provides richer semantic information about the segmented objects by expanding the diversity of visual prompts. Based on the above insights, a natural idea is to select these" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.363, + 0.128 + ], + "angle": 0, + "content": "W. Suo et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.284 + ], + "angle": 0, + "content": "considerable discrepancy examples as contextual prompts for each test instance. However, implementing the straightforward strategy would incur substantial annotation cost and complex computation, since each test sample needs to compare and rerank the candidate list. To mitigate the above problems, we introduce a novel stepwise context search (SCS) method that maintains the diversity of the candidate pool and reduces the cost of annotation. Moreover, considering the necessity of selecting appropriate contextual examples for different test queries, we further design an adaptive search module. It can dynamically select contexts from the candidate pool based on given test samples." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.306, + 0.376, + 0.323 + ], + "angle": 0, + "content": "4 Our Method" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.337, + 0.788, + 0.458 + ], + "angle": 0, + "content": "Based on the discussion in Sec. 3, we design a new SCS method to construct a small yet rich candidate pool and adaptively search the contextual demonstrations. Compared to previous works, our method has two potential advantages: 1) Fewer Annotations. The SCS focuses on building a small yet diverse candidate pool. It only needs to annotate a portion of samples while achieving better performance. 2) Adapting Selection. Compared to the previous rule-based mechanism [37,56], our method can adaptively search the well-matched examples for different test instances. Next, we would introduce our method in detail." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.48, + 0.504, + 0.496 + ], + "angle": 0, + "content": "4.1 Constructing Candidate Pool" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.505, + 0.787, + 0.58 + ], + "angle": 0, + "content": "As mentioned in Sec. 3, it is important to provide examples for the ICL-based segmentation paradigm. However, previous research mostly depend on huge annotations as contextual candidates, it would inevitably suffers from expensive annotation cost. Motivated by the of Sec. 3.5, we aim to build a small yet diverse candidate pool that can be used for subsequent context searching." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.58, + 0.788, + 0.672 + ], + "angle": 0, + "content": "Given \\(N\\) unlabeled samples, follow in [56], we utilize CLIP to extract image feature into \\(d\\)-dimensional vectors, represented by \\(F = \\{f_{i}\\}_{i=1}^{N}\\), \\(f_{i} \\in \\mathbb{R}^{d}\\), where \\(f_{i}\\) is the \\(i\\)-th encoded image feature. To construct a candidate pool comprising a set of representative examples, we employ k-means clustering algorithm [58] to generate \\(M\\) clusters \\(C = \\{c_{m}\\}_{m=1}^{M}\\), which can be determined by minimizing the sum of squared intra-cluster norms:" + }, + { + "type": "equation", + "bbox": [ + 0.403, + 0.682, + 0.787, + 0.726 + ], + "angle": 0, + "content": "\\[\n\\arg \\min _ {C} \\sum_ {m = 1} ^ {M} \\sum_ {f _ {i} \\in c _ {j}} \\| f _ {i} - \\mu_ {m} \\|, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.787, + 0.78 + ], + "angle": 0, + "content": "where \\(\\mu_{m}\\) is centroid of cluster \\(c_{m}\\). Initially, this algorithm randomly selects \\(M\\) centroids. Each instance is assigned by computing the distance to all selected centroids. The process is iteratively executed until the clusters are stable." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.781, + 0.788, + 0.842 + ], + "angle": 0, + "content": "To construct the candidate pool, a simple method is to select members closest to the centroid, while previous studies have shown that samples closest to the center do not necessarily represent the entire cluster [11, 46]. Moreover, related works have provided several alternative sampling methods such as density" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.311, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Visual Prompt Selection for In-Context Learning Segmentation" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.146, + 0.782, + 0.316 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.328, + 0.788, + 0.412 + ], + "angle": 0, + "content": "Fig. 4: Overview of our SCS method. Instead of similarity sorting on a large annotated dataset, we use clustering to select diverse examples from unlabeled data \\(D\\) and construct the candidate pool. Meanwhile, the search agent is used to further select contextual demonstrations for various test samples based on reinforcement learning. During inference, the test sample and candidate pool examples are fed into the model and adaptively search visual prompts." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.439, + 0.788, + 0.613 + ], + "angle": 0, + "content": "sorting [46] and entropy sorting [11]. In this paper, we design a simpler and more efficient sampling strategy without complex computation. Specifically, for each cluster \\( c_{m} \\), which includes members \\( F_{m} = \\{f_{k}\\}_{k=1}^{K} \\), where \\( K \\) denotes the number of members in \\( c_{m} \\). We sort them into a list \\( [f_{k}^{(1)}, f_{k}^{(2)}, \\dots, f_{k}^{(K)}] \\) in the ascending order of their distance to the cluster's center. Based on the findings of Sec. 3.3, aggregating the nearest and farthest examples can cover more diverse demonstrations. Thus, we consider the \\( f_{mk}^{(1)} \\) and \\( f_{mk}^{(K)} \\) as typical samples for each cluster. By collecting these samples from all clusters, we obtain the final candidate pool \\( D = \\{f_{1k}^{(1)}, f_{1k}^{(K)}, f_{2k}^{(1)}, \\dots, f_{Mk}^{(1)}, f_{Mk}^{(K)}\\} \\), which contains a total of \\( 2M \\) samples. To simplify, we use \\( f_{d}^{m} \\) and \\( I_{d}^{m} \\) to represent \\( m \\)-th image feature and corresponding image, respectively." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.632, + 0.467, + 0.646 + ], + "angle": 0, + "content": "4.2 Adaptive Search Module" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.655, + 0.787, + 0.7 + ], + "angle": 0, + "content": "In this module, our goal is to select well-matched examples from the constructed candidate pool. As shown in Fig. 4, we propose a new adaptive search module that can further search and assemble the suitable contexts for given queries." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.7, + 0.788, + 0.79 + ], + "angle": 0, + "content": "Specifically, for a given test instance \\( I_{q} \\), we first apply the same image encoder (i.e. CLIP) to extract the corresponding feature vector \\( f_{q} \\in \\mathbb{R}^{d} \\). Then, we concatenate the image feature \\( f_{q} \\) and each example \\( f_{d}^{m} \\) from the candidate pool. These fused features are fed into a search agent that consists of a group of light and simple Multi-Layer Perceptrons (MLP). The search computation can be denoted as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.436, + 0.8, + 0.785, + 0.816 + ], + "angle": 0, + "content": "\\[\ns _ {d} ^ {m} = \\operatorname {M L P} \\left(\\left[ f _ {d} ^ {m}; f _ {q} \\right]\\right), \\tag {3}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.358, + 0.819, + 0.785, + 0.837 + ], + "angle": 0, + "content": "\\[\na _ {d} ^ {1}, a _ {d} ^ {2}, \\dots , a _ {d} ^ {2 M} = \\operatorname {s o f t m a x} \\left(s _ {d} ^ {1}, s _ {d} ^ {2}, \\dots , s _ {d} ^ {2 M}\\right), \\tag {4}\n\\]" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.361, + 0.127 + ], + "angle": 0, + "content": "W. Suo et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.284 + ], + "angle": 0, + "content": "where \\( a_{d}^{m} \\) represents the probability of \\( m \\)-th example in the candidate pool. To guide the search agent in learning select examples, we apply the IoU scores as rewards for selecting well-matched demonstrations from the candidate pool. Specifically, following the format of the ICL-based segmentation model, we re-organize query \\( I_{q} \\), the original image of context examples \\( I_{d} \\) and corresponding mask \\( G_{d} \\) into a joint image, as shown in Fig. 4. These images are then fed into the ICL-based segmentation model to obtain the IoU scores \\( U = \\{u_{d}\\}_{d=1}^{2M} \\). Inspired by [4,38], we use the average IoU score \\( u_{avg} \\) as the base score. By applying reinforcement learning, the gradient is calculated by:" + }, + { + "type": "equation", + "bbox": [ + 0.345, + 0.296, + 0.786, + 0.337 + ], + "angle": 0, + "content": "\\[\n\\nabla_ {\\theta} L (\\theta) = - \\frac {1}{2 M} \\sum_ {m = 1} ^ {2 M} \\left(u _ {d} ^ {m} - u _ {a v g}\\right) \\nabla_ {\\theta} \\log \\left(a _ {d} ^ {m}\\right), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.348, + 0.788, + 0.409 + ], + "angle": 0, + "content": "where \\( a_{d}^{m} \\) and \\( \\theta \\) are the probability of \\( m \\)-th example and the parameters of the search agent. Based on the above computation, this gradient would tend to increase the probability of the \\( m \\)-th context when its IoU scores \\( u_{d}^{m} \\) are higher than the average IoU score \\( u_{avg} \\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.409, + 0.787, + 0.455 + ], + "angle": 0, + "content": "During inference, we directly input the test samples and candidate pool examples into the search agent. The contextual examples are then determined by sorting them based on probability scores \\( a_{d} \\) as detailed in Eq. 4." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.48, + 0.367, + 0.498 + ], + "angle": 0, + "content": "5 Experiment" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.515, + 0.787, + 0.575 + ], + "angle": 0, + "content": "We use different prompt selection strategies and various generalist segmentation frameworks to demonstrate the superiority of our method. Furthermore, we compare our approach with traditional methods and conduct extensive qualitative and quantitative studies on multiple datasets." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.601, + 0.44, + 0.617 + ], + "angle": 0, + "content": "5.1 Experimental Setting" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.628, + 0.788, + 0.779 + ], + "angle": 0, + "content": "Dataset. We follow the previous works [6, 45, 56] and evaluate our approach on PASCAL- \\(5^{i}\\) [32], COCO- \\(20^{i}\\) [21] and iSALD- \\(5^{i}\\) [50]. PASCAL- \\(5^{i}\\) includes 20 semantic categories and is created from PASCAL VOC 2012 [9] with external annotations from SDS [10]. COCO- \\(20^{i}\\) is built upon MSCOCO [21], which is a more challenging dataset and contains 80 categories. For all categories, both COCO- \\(20^{i}\\) and PASCAL- \\(5^{i}\\) are evenly divided into 4 folds for cross-validation, where each fold is selected as the test set and the remaining three folds are used for training. The iSALD- \\(5^{i}\\) contains 15 geospatial categories which derived from iSALD [48] in the remote sensing image semantic segmentation task. Similarly, the iSALD- \\(5^{i}\\) is also partitioned into 3 folds for cross-validation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Evaluation Metric. Following [13], mean intersection over union (mIoU) and foreground-background IoU (FBIoU) are adopted as evaluation metrics. The former measures the average overall IoU values for all object classes, while the latter reports the average IoU scores for the foreground and background." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.311, + 0.115, + 0.732, + 0.131 + ], + "angle": 0, + "content": "Visual Prompt Selection for In-Context Learning Segmentation" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.784, + 0.128 + ], + "angle": 0, + "content": "11" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.144, + 0.788, + 0.203 + ], + "angle": 0, + "content": "Table 1: The results on COCO- \\(20^{i}\\) and PASCAL- \\(5^{i}\\) based on example-based semantic segmentation. To ensure a fair comparison and follow the few-shot setting, \\(\\dagger\\) indicates that only using the pre-trained weights provided by the SegGPT [3], and we report related results based on randomly sampling (i.e., w/R) and similarity sorting [56]." + }, + { + "type": "table", + "bbox": [ + 0.245, + 0.213, + 0.761, + 0.384 + ], + "angle": 0, + "content": "
MethodCOCO-20i [21]PASCAL-5i [32]
1-shot5-shot1-shot5-shot
mIoUFBIoUmIoUFBIoUmIoUFBIoUmIoUFBIoU
Painter [44]32.850.232.650.164.574.264.674.4
MAE-VQGAN [6]12.145.7--27.555.6--
MAE-VQGAN w/ [56]18.046.5--33.657.8--
SegGPT† w/R [45]48.671.352.773.172.183.176.280.3
SegGPT† w/ [56]53.372.256.175.574.484.476.985.9
MAE-VQGAN w/Ours19.147.0--35.061.2--
SegGPT† w/Ours57.275.159.576.876.685.478.286.5
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.429, + 0.788, + 0.538 + ], + "angle": 0, + "content": "Implementation details. For both datasets, the Adam [17] is used as our optimizer and the maximum number of training epochs is 10. We use the CLIPViT [27] image encoder to extract image features. During training, all images are resized to \\(224 \\times 224\\) and fed into the image encoder. The dimension of extracted feature vectors is 1024. We train on one 4080 GPU with an initial learning rate of 1e-3, which is dropped by half every 5 epochs. The number of clusters \\(M\\) and batch size is set to 10 and 8." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.573, + 0.461, + 0.589 + ], + "angle": 0, + "content": "5.2 Quantitative Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.614, + 0.789, + 0.842 + ], + "angle": 0, + "content": "In Table 1, we show a performance comparison of two general ICL-based segmentation models with and without proposed SCS. Following [37,56], all results are reported by repeating the random experiment five times. It can be observed that our method can significantly boost performance in all settings including 1-shot and 5-shot settings. In particular, for the 1-shot setting, the SCS leads to MAE-VQGAN with the mIoU gain of 7.0 and 6.5 on the PASCAL- \\(5^{i}\\) and COCO- \\(20^{i}\\) dataset, respectively. Besides, compared to previous rule-base models [56] in 1-shot and 5-shot semantic segmentation tasks, our method surpasses it by a considerable margin. Moreover, following the few-shot setting and ensuring a fair comparison, we report related results based on the pre-trained weights for SegGPT [3]. It can be observed that our method significantly outperforms these rule-based visual prompt selection methods for SegGPT by 9.7 and 6.8 on the 1-shot and 5-shot settings for COCO- \\(20^{i}\\). More importantly, the above results demonstrate that the proposed SCS can be considered as a \"plug-and-play\" module to enhance the segmentation ability of ICL-based frameworks." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.362, + 0.127 + ], + "angle": 0, + "content": "W. Suo et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.788, + 0.174 + ], + "angle": 0, + "content": "Table 2: Ablation study. Similarity: Cosine similarity ranking; Cluster: Cluster compacting; Sorting: The nearest-farthest sampling; ASM: Adaptive search module." + }, + { + "type": "table", + "bbox": [ + 0.302, + 0.198, + 0.703, + 0.287 + ], + "angle": 0, + "content": "
SimilarityClusterSortingASMmIoUFBIoU
152.773.1
2✓56.174.5
3✓✓55.774.3
4✓✓✓57.675.7
5✓✓✓59.576.8
" + }, + { + "type": "table_caption", + "bbox": [ + 0.31, + 0.312, + 0.694, + 0.326 + ], + "angle": 0, + "content": "Table 3: The effects of different settings for our method." + }, + { + "type": "table", + "bbox": [ + 0.356, + 0.338, + 0.642, + 0.469 + ], + "angle": 0, + "content": "
Model1-shot5-shot
mIoUFBIoUmIoUFBIoU
1Entropy53.172.356.175.0
2Foreground53.572.556.779.3
3ViT56.975.058.776.0
4CLIP-resnet56.374.758.576.4
5Density-sort57.675.159.176.3
6Entropy-sort56.974.558.475.9
7Ours57.275.159.576.8
" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.498, + 0.389, + 0.513 + ], + "angle": 0, + "content": "5.3 Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.523, + 0.787, + 0.568 + ], + "angle": 0, + "content": "To clarify the effectiveness of our method, we conduct further ablation studies on the COCO- \\(20^{i}\\) dataset with the 5-shot setting. The SegGPT is used as the basic model due to its outstanding segmentation performance." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.569, + 0.787, + 0.689 + ], + "angle": 0, + "content": "Effects of Cluster. In Table 2, we first build a baseline by randomly selecting examples from the entire training set. The average result is reported across 5 different seeds. In the second row of Table 2, we follow [56] and utilize cosine similarity ranking to select the top-5 examples as contexts. In the third row, we investigate the impact of the cluster on in-context example selection, using the closest examples to each cluster centroid to construct the candidate pool. Compared to the baseline, the results show that compacting the search space not only reduces annotation cost but also improves model performance." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.689, + 0.787, + 0.795 + ], + "angle": 0, + "content": "Effects of Sort. As shown in row 4 of Table 2, we experimentally demonstrate that integrating our sorting mechanism results in \\(1.9\\mathrm{mIoU}\\) improvement compared to directly using the closest examples to each cluster centroid. The inherent reason is that centroid examples describe a limited feature space [33, 46], making it challenging to support a variety of different contexts for test instances. In contrast, our sorting mechanism introduces the nearest and farthest examples, which constructs more diverse demonstrations for ICL learning." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.787, + 0.842 + ], + "angle": 0, + "content": "Effects of Adaptive Search Module. Finally, we further explore the effectiveness of our adaptive search module in the last row of Table 2. Compared to using a simple similarity sorting method, our approach shows an improve" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.31, + 0.114, + 0.732, + 0.131 + ], + "angle": 0, + "content": "Visual Prompt Selection for In-Context Learning Segmentation" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.128 + ], + "angle": 0, + "content": "13" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.144, + 0.77, + 0.174 + ], + "angle": 0, + "content": "Table 4: The results on PASCAL-5\\(^i\\) [32] Table 5: The results on iSALD-5\\(^i\\) [48] based on instance segmentation. based on semantic segmentation." + }, + { + "type": "table", + "bbox": [ + 0.235, + 0.185, + 0.477, + 0.248 + ], + "angle": 0, + "content": "
Model1-shot5-shot
SegGPT† w/R [45]46.546.9
SegGPT† w/ [56]47.148.5
SegGPT w/Ours50.551.4
" + }, + { + "type": "table", + "bbox": [ + 0.511, + 0.185, + 0.752, + 0.247 + ], + "angle": 0, + "content": "
Model1-shot5-shot
SegGPT† w/R [45]29.830.6
SegGPT† w/ [56]34.735.4
SegGPT w/Ours37.939.5
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.274, + 0.788, + 0.32 + ], + "angle": 0, + "content": "ment of 3.4 points. Meanwhile, benefiting from the diverse candidate pool, our SCS achieves state-of-the-art performance and outperforms the baseline by a considerable margin." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.342, + 0.467, + 0.359 + ], + "angle": 0, + "content": "5.4 Different Model Settings" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.368, + 0.789, + 0.687 + ], + "angle": 0, + "content": "In this section, we explore several alternative model settings to further discuss the proposed method. All experiments are conducted on the COCO- \\(20^{i}\\) with the SegGPT model. In the first two rows, we first compare the performance using different metrics. Specifically, we use the rule-based methods to select the contextual examples from the complete training set, which including Entropy sorting [11] and Foreground sorting [53]. In particular, \"Entropy\" is executed by calculating the maximum entropy values. The \"Foreground\" refers to a two-stage approach [53]. It uses an object detector [42] to extract the foreground boxes and the foreground similarity is used as the metric. From the 1-shot and 5-shot results, it can be observed that our SCS can dramatically outperform these methods by a large margin. In rows 3-4, we construct the experiments with different feature encoders. We replace the CLIP-ViT with ViT [8] and CLIP-resnet [27] to obtain image feature vectors. These extracted features would be utilized to construct the candidate pool and perform the adaptive searching. The results show that our SCS is robust for different image encoders. Finally, in rows 5-6 of Table 3, we replace our selection mechanism with several alternative manners. However, these approaches mostly involve complex computation. For example, the Density-sort [2] requires additional computations on the order of \\(O(MN)\\) in each iteration to update density lists, where \\(M\\) is the number of clusters and \\(N\\) is the size of the unlabeled data. Instead, our method uses the basic clustering algorithm and obtains a comparable performance." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.709, + 0.419, + 0.724 + ], + "angle": 0, + "content": "5.5 Further Discussion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.734, + 0.788, + 0.841 + ], + "angle": 0, + "content": "More Segmentation tasks. In this section, we evaluate the effect of our method on more segmentation tasks and datasets. Because MAE-VQGAN cannot distinguish between individual instances of objects within the same class, we construct more experiments with the SegGPT model. Specifically, we report instance segmentation results on PASCAL-5i in Table 4 and aerial image semantic segmentation results on iSALD-5i in Table 5. The above results indicate that the SCS can enhance the segmentation ability for various real-world scenarios." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.362, + 0.127 + ], + "angle": 0, + "content": "W. Suo et al." + }, + { + "type": "image", + "bbox": [ + 0.266, + 0.144, + 0.74, + 0.409 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.419, + 0.788, + 0.463 + ], + "angle": 0, + "content": "Fig.5: Qualitative results under 1-shot setting on COCO- \\(20^{i}\\). The similarity-based selection method is viewed as our baseline. The green regions, red regions and purple regions are example masks, predicted masks and ground-truth masks." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.488, + 0.789, + 0.685 + ], + "angle": 0, + "content": "Qualitative Results. As shown in Fig. 5, to further analyze and verify the proposed method, we visualized the qualitative segmentation results on COCO\\(_{20^i}\\) with the SegGPT. The first and third columns represent the visual prompts (highlighted in green) based on similarity sorting and our SCS method, respectively. The second and fourth columns show the corresponding segmentation results (highlighted in red). The last column represents the ground truth mask (highlighted in purple). As illustrated in the first row, the baseline model selects the most similar context to the query (i.e., chopping board, knife and carrot), while the segment result is not satisfactory, as the model can not recognize the knife in the image. Instead, Our method selects an image from the candidate pool that accurately segments out the mask. The above results demonstrate that similarity sorting is not the only golden standard and improving the diversity of context prompts can boost the segmentation results." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.706, + 0.35, + 0.721 + ], + "angle": 0, + "content": "6 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.734, + 0.788, + 0.841 + ], + "angle": 0, + "content": "In this paper, we focus on the example selection method for the ICL-based segmentation frameworks. By building comprehensive comparisons, we empirically demonstrate that diversity of contexts is a key factor in guiding segmentation and boosting performance. Beyond the above insights, we propose a new stepwise context search method that improves the diversity of the candidate pool and significantly alleviates the annotation cost. We hope that this work can promote further research to better understand in-context learning in computer vision." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.311, + 0.115, + 0.732, + 0.13 + ], + "angle": 0, + "content": "Visual Prompt Selection for In-Context Learning Segmentation" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.207 + ], + "angle": 0, + "content": "Acknowledgement. This work was supported by National Natural Science Foundation of China (No.62102323), the Innovation Capability Support Program of Shaanxi(Program No. 2023KJXX-142) and National Natural Science Foundation of China (U23B2013)." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.233, + 0.326, + 0.249 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.266, + 0.785, + 0.295 + ], + "angle": 0, + "content": "1. Agrawal, S., Zhou, C., Lewis, M., Zettlemoyer, L., Ghazvininejad, M.: In-context examples selection for machine translation. arXiv preprint arXiv:2212.02437 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.296, + 0.785, + 0.323 + ], + "angle": 0, + "content": "2. et al., W.: \"unsupervised selective labeling for more effective semi-supervised learning\". In: ECCV 2022. pp. 427-445. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.324, + 0.751, + 0.337 + ], + "angle": 0, + "content": "3. et al, W.: \"seggspt: Segmenting everything in context\". In ICCV 2023 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.338, + 0.785, + 0.393 + ], + "angle": 0, + "content": "4. Anderson, P., He, X., Buehler, C., Teney, D., Johnson, M., Gould, S., Zhang, L.: Bottom-up and top-down attention for image captioning and visual question answering. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 6077-6086 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.394, + 0.785, + 0.435 + ], + "angle": 0, + "content": "5. Badrinarayanan, V., Kendall, A., Cipolla, R.: Segnet: A deep convolutional encoder-decoder architecture for image segmentation. IEEE transactions on pattern analysis and machine intelligence 39(12), 2481-2495 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.436, + 0.785, + 0.476 + ], + "angle": 0, + "content": "6. Bar, A., Gandelsman, Y., Darrell, T., Globerson, A., Efros, A.: Visual prompting via image inpainting. Advances in Neural Information Processing Systems 35, 25005-25017 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.478, + 0.785, + 0.518 + ], + "angle": 0, + "content": "7. Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J.D., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., et al.: Language models are few-shot learners. Advances in neural information processing systems 33, 1877-1901 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.519, + 0.785, + 0.574 + ], + "angle": 0, + "content": "8. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.576, + 0.785, + 0.616 + ], + "angle": 0, + "content": "9. Everingham, M., Van Gool, L., Williams, C.K., Winn, J., Zisserman, A.: The Pascal visual object classes (voc) challenge. International journal of computer vision 88, 303-338 (2010)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.617, + 0.785, + 0.657 + ], + "angle": 0, + "content": "0. Hariharan, B., Arbeláez, P., Bourdev, L., Maji, S., Malik, J.: Semantic contours from inverse detectors. In: 2011 international conference on computer vision. pp. 991-998. IEEE (2011)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.659, + 0.785, + 0.714 + ], + "angle": 0, + "content": "1. Haussmann, E., Fenzi, M., Chitta, K., Ivanecky, J., Xu, H., Roy, D., Mittel, A., Koumchatzky, N., Farabet, C., Alvarez, J.M.: Scalable active learning for object detection. In: 2020 IEEE intelligent vehicles symposium (iv). pp. 1430-1435. IEEE (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.715, + 0.785, + 0.743 + ], + "angle": 0, + "content": "2. He, K., Gkioxari, G., Dollár, P., Girshick, R.: Mask r-cnn. In: Proceedings of the IEEE international conference on computer vision. pp. 2961-2969 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.744, + 0.785, + 0.784 + ], + "angle": 0, + "content": "3. Hong, S., Cho, S., Nam, J., Lin, S., Kim, S.: Cost aggregation with 4d convolutional swim transformer for few-shot segmentation. In: European Conference on Computer Vision. pp. 108-126. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.785, + 0.785, + 0.84 + ], + "angle": 0, + "content": "4. Hongjin, S., Kasai, J., Wu, C.H., Shi, W., Wang, T., Xin, J., Zhang, R., Ostendorf, M., Zettlemoyer, L., Smith, N.A., et al.: Selective annotation makes language models better few-shot learners. In: The Eleventh International Conference on Learning Representations (2022)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.266, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.362, + 0.127 + ], + "angle": 0, + "content": "W. Suo et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.148, + 0.785, + 0.175 + ], + "angle": 0, + "content": "15. Hore, A., Ziou, D.: Image quality metrics: Psnr vs. sSIM. In: 2010 20th international conference on pattern recognition. pp. 2366-2369. IEEE (2010)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.175, + 0.785, + 0.216 + ], + "angle": 0, + "content": "16. Hu, H., Li, Z., Li, L., Yang, H., Zhu, H.: Classification of very high-resolution remote sensing imagery using a fully convolutional network with global and local context information enhancements. IEEE Access 8, 14606-14619 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.216, + 0.785, + 0.242 + ], + "angle": 0, + "content": "17. Kingma, D.P., Ba, J.: Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980 (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.242, + 0.785, + 0.283 + ], + "angle": 0, + "content": "18. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., et al.: Segment anything. arXiv preprint arXiv:2304.02643 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.283, + 0.785, + 0.323 + ], + "angle": 0, + "content": "19. Li, X., Lv, K., Yan, H., Lin, T., Zhu, W., Ni, Y., Xie, G., Wang, X., Qiu, X.: Unified demonstration retriever for in-context learning. arXiv preprint arXiv:2305.04320 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.323, + 0.785, + 0.392 + ], + "angle": 0, + "content": "20. Li, Y., Padmanabhan, A., Zhao, P., Wang, Y., Xu, G.H., Netravali, R.: Reducto: On-camera filtering for resource-efficient real-time video analytics. In: Proceedings of the Annual conference of the ACM Special Interest Group on Data Communication on the applications, technologies, architectures, and protocols for computer communication. pp. 359-376 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.392, + 0.785, + 0.447 + ], + "angle": 0, + "content": "21. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollar, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13. pp. 740-755. Springer (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.447, + 0.785, + 0.473 + ], + "angle": 0, + "content": "22. Liu, J., Shen, D., Zhang, Y., Dolan, B., Carin, L., Chen, W.: What makes good in-context examples for gpt-3? arXiv preprint arXiv:2101.06804 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.473, + 0.785, + 0.514 + ], + "angle": 0, + "content": "23. Long, J., Shelhamer, E., Darrell, T.: Fully convolutional networks for semantic segmentation. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 3431-3440 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.514, + 0.785, + 0.569 + ], + "angle": 0, + "content": "24. Luo, G., Zhou, Y., Sun, X., Cao, L., Wu, C., Deng, C., Ji, R.: Multi-task collaborative network for joint referring expression comprehension and segmentation. In: Proceedings of the IEEE/CVF Conference on computer vision and pattern recognition. pp. 10034-10043 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.569, + 0.785, + 0.622 + ], + "angle": 0, + "content": "25. Lyu, X., Fang, Y., Tong, B., Li, X., Zeng, T.: Multiscale normalization attention network for water body extraction from remote sensing imagery. Remote Sensing 14(19) (2022). https://doi.org/10.3390/rs14194983, https://www.mdpi.com/2072-4292/14/19/4983" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.622, + 0.785, + 0.664 + ], + "angle": 0, + "content": "26. Min, S., Lyu, X., Holtzman, A., Artetxe, M., Lewis, M., Hajishirzi, H., Zettlemoyer, L.: Rethinking the role of demonstrations: What makes in-context learning work? arXiv preprint arXiv:2202.12837 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.664, + 0.785, + 0.718 + ], + "angle": 0, + "content": "27. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.718, + 0.785, + 0.759 + ], + "angle": 0, + "content": "28. Reisenhofer, R., Bosse, S., Kutyniok, G., Wiegand, T.: Aaar wavelet-based perceptual similarity index for image quality assessment. Signal Processing: Image Communication 61, 33-43 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.759, + 0.785, + 0.814 + ], + "angle": 0, + "content": "29. Ronneberger, O., Fischer, P., Brox, T.: U-net: Convolutional networks for biomedical image segmentation. In: Medical Image Computing and Computer-Assisted Intervention-MICCAI 2015: 18th International Conference, Munich, Germany, October 5-9, 2015, Proceedings, Part III 18. pp. 234-241. Springer (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.814, + 0.785, + 0.84 + ], + "angle": 0, + "content": "30. Rubin, O., Herzig, J., Berant, J.: Learning to retrieve prompts for in-context learning. arXiv preprint arXiv:2112.08633 (2021)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.148, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.311, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Visual Prompt Selection for In-Context Learning Segmentation" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.203 + ], + "angle": 0, + "content": "31. Rubin, O., Herzig, J., Berant, J.: Learning to retrieve prompts for in-context learning. In: Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies. pp. 2655-2671 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.204, + 0.787, + 0.231 + ], + "angle": 0, + "content": "32. Shaban, A., Bansal, S., Liu, Z., Essa, I., Boots, B.: One-shot learning for semantic segmentation. arXiv preprint arXiv:1709.03410 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.232, + 0.786, + 0.272 + ], + "angle": 0, + "content": "33. Shen, F., Zhao, L., Du, W., Zhong, W., Qian, F.: Large-scale industrial energy systems optimization under uncertainty: A data-driven robust optimization approach. Applied Energy 259, 114199 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.273, + 0.787, + 0.314 + ], + "angle": 0, + "content": "34. Shi, H., Li, H., Meng, F., Wu, Q.: Key-word-aware network for referring expression image segmentation. In: Proceedings of the European Conference on Computer Vision (ECCV). pp. 38-54 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.315, + 0.787, + 0.355 + ], + "angle": 0, + "content": "35. Sun, L., Yang, K., Hu, X., Hu, W., Wang, K.: Real-time fusion network forrgb-d semantic segmentation incorporating unexpected obstacle detection for road-driving images. IEEE robotics and automation letters 5(4), 5558-5565 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.356, + 0.787, + 0.397 + ], + "angle": 0, + "content": "36. Sun, M., Suo, W., Wang, P., Niu, K., Liu, L., Lin, G., Zhang, Y., Wu, Q.: An adaptive correlation filtering method for text-based person search. International Journal of Computer Vision pp. 1-16 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.398, + 0.787, + 0.424 + ], + "angle": 0, + "content": "37. Sun, Y., Chen, Q., Wang, J., Wang, J., Li, Z.: Exploring effective factors for improving visual in-context learning. arXiv preprint arXiv:2304.04748 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.425, + 0.787, + 0.48 + ], + "angle": 0, + "content": "38. Suo, W., Sun, M., Liu, W., Gao, Y., Wang, P., Zhang, Y., Wu, Q.: S3c: Semi-supervised vqa natural language explanation via self-critical learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2646-2656 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.481, + 0.787, + 0.522 + ], + "angle": 0, + "content": "39. Suo, W., Sun, M., Niu, K., Gao, Y., Wang, P., Zhang, Y., Wu, Q.: A simple and robust correlation filtering method for text-based person search. In: European conference on computer vision. pp. 726-742. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.523, + 0.787, + 0.563 + ], + "angle": 0, + "content": "40. Suo, W., Sun, M., Wang, P., Zhang, Y., Wu, Q.: Rethinking and improving feature pyramids for one-stage referring expression comprehension. IEEE Transactions on Image Processing 32, 854-864 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.564, + 0.787, + 0.605 + ], + "angle": 0, + "content": "41. Vasudevan, A.B., Dai, D., Van Gool, L.: Semantic object prediction and spatial sound super-resolution with binaural sounds. In: European conference on computer vision. pp. 638-655. Springer (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.606, + 0.787, + 0.66 + ], + "angle": 0, + "content": "42. Wang, W., Dai, J., Chen, Z., Huang, Z., Li, Z., Zhu, X., Hu, X., Lu, T., Lu, L., Li, H., et al.: Intermimage: Exploring large-scale vision foundation models with deformable convolutions. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14408-14419 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.661, + 0.787, + 0.688 + ], + "angle": 0, + "content": "43. Wang, X., Ma, H., You, S.: Deep clustering for weakly-supervised semantic segmentation in autonomous driving scenes. Neurocomputing 381, 20-28 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.689, + 0.787, + 0.73 + ], + "angle": 0, + "content": "44. Wang, X., Wang, W., Cao, Y., Shen, C., Huang, T.: Images speak in images: A generalist painter for in-context visual learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6830-6839 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.731, + 0.787, + 0.757 + ], + "angle": 0, + "content": "45. Wang, X., Zhang, X., Cao, Y., Wang, W., Shen, C., Huang, T.: Seggpt: Segmenting everything in context. arXiv preprint arXiv:2304.03284 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.758, + 0.787, + 0.799 + ], + "angle": 0, + "content": "46. Wang, X., Lian, L., Yu, S.X.: Unsupervised selective labeling for more effective semi-supervised learning. In: European Conference on Computer Vision. pp. 427-445. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.799, + 0.787, + 0.84 + ], + "angle": 0, + "content": "47. Wang, Z., Bovik, A.C., Sheikh, H.R., Simoncelli, E.P.: Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing 13(4), 600-612 (2004)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.362, + 0.127 + ], + "angle": 0, + "content": "W. Suo et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.787, + 0.204 + ], + "angle": 0, + "content": "48. Waqas Zamir, S., Arora, A., Gupta, A., Khan, S., Sun, G., Shahbaz Khan, F., Zhu, F., Shao, L., Xia, G.S., Bai, X.: isaid: A large-scale dataset for instance segmentation in aerial images. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops. pp. 28-37 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.205, + 0.786, + 0.245 + ], + "angle": 0, + "content": "49. Wei, J., Wang, X., Schuurmans, D., Bosma, M., Xia, F., Chi, E., Le, Q.V., Zhou, D., et al.: Chain-of-thought prompting elicits reasoning in large language models. Advances in Neural Information Processing Systems 35, 24824-24837 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.246, + 0.787, + 0.286 + ], + "angle": 0, + "content": "50. Yao, X., Cao, Q., Feng, X., Cheng, G., Han, J.: Scale-aware detailed matching for few-shot aerial image semantic segmentation. IEEE Transactions on Geoscience and Remote Sensing 60, 1-11 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.287, + 0.787, + 0.314 + ], + "angle": 0, + "content": "51. Ye, J., Wu, Z., Feng, J., Yu, T., Kong, L.: Compositional exemplars for in-context learning. arXiv preprint arXiv:2302.05698 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.315, + 0.787, + 0.355 + ], + "angle": 0, + "content": "52. Yu, Q., Gao, Y., Zheng, Y., Zhu, J., Dai, Y., Shi, Y.: Crossover-net: Leveraging vertical-horizontal crossover relation for robust medical image segmentation. Pattern Recognition 113, 107756 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.356, + 0.787, + 0.397 + ], + "angle": 0, + "content": "53. Yu, Y., Yang, X., Li, Q., Zhou, Y., Da, F., Yan, J.: H2rbox-v2: Incorporating symmetry for boosting horizontal box supervised oriented object detection. In: Thirty-seventh Conference on Neural Information Processing Systems (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.398, + 0.787, + 0.438 + ], + "angle": 0, + "content": "54. Zendel, O., Schörghuber, M., Rainer, B., Murschitz, M., Beleznai, C.: Unifying panoptic segmentation for autonomous driving. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 21351-21360 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.439, + 0.787, + 0.48 + ], + "angle": 0, + "content": "55. Zhang, L., Zhang, L., Mou, X., Zhang, D.: Fsim: A feature similarity index for image quality assessment. IEEE transactions on Image Processing 20(8), 2378-2386 (2011)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.481, + 0.787, + 0.508 + ], + "angle": 0, + "content": "56. Zhang, Y., Zhou, K., Liu, Z.: What makes good examples for visual in-context learning? arXiv preprint arXiv:2301.13670 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.509, + 0.787, + 0.563 + ], + "angle": 0, + "content": "57. Zhang, Y., Liu, H., Hu, Q.: Transfuse: Fusing transformers and cnns for medical image segmentation. In: Medical Image Computing and Computer Assisted Intervention-MICCAI 2021: 24th International Conference, Strasbourg, France, September 27-October 1, 2021, Proceedings, Part I 24. pp. 14-24. Springer (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.565, + 0.787, + 0.592 + ], + "angle": 0, + "content": "58. Zhang, Z., Zhang, A., Li, M., Smola, A.: Automatic chain of thought prompting in large language models. arXiv preprint arXiv:2210.03493 (2022)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.592 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/Rethinking and Improving Visual Prompt Selection for In-Context Learning Segmentation Framework/c5a744a9-bd8f-4567-93ec-5cbfe907a7d5_origin.pdf b/2024/Rethinking and Improving Visual Prompt Selection for In-Context Learning Segmentation Framework/c5a744a9-bd8f-4567-93ec-5cbfe907a7d5_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..53135a83a033aaf162f692510c5d5a39c0cc5726 --- /dev/null +++ b/2024/Rethinking and Improving Visual Prompt Selection for In-Context Learning Segmentation Framework/c5a744a9-bd8f-4567-93ec-5cbfe907a7d5_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:043ab69b219b5d0812a0ce3348dc2236b2d555a7143e2f8a58d0005e557ce838 +size 5516746 diff --git a/2024/Rethinking and Improving Visual Prompt Selection for In-Context Learning Segmentation Framework/full.md b/2024/Rethinking and Improving Visual Prompt Selection for In-Context Learning Segmentation Framework/full.md new file mode 100644 index 0000000000000000000000000000000000000000..34681f9ba0a2f2a3e4a2a97e0f008e5b5ccb3372 --- /dev/null +++ b/2024/Rethinking and Improving Visual Prompt Selection for In-Context Learning Segmentation Framework/full.md @@ -0,0 +1,280 @@ +# Rethinking and Improving Visual Prompt Selection for In-Context Learning Segmentation + +Wei Suo $^{1,3,\star}$ , Lanqing Lai $^{1,\star}$ , Mengyang Sun $^{2,3}$ , Hanwang Zhang $^{3}$ , Peng Wang $^{1,\dagger}$ , and Yanning Zhang $^{1}$ + +1 School of Computer Science and Ningbo Institute, Northwestern Polytechnical University, Xi'an, China +$^{2}$ School of Cybersecurity, Northwestern Polytechnical University, Xi'an, China +3 School of Computer Science, Nanyang Technological University, Singapore {suwei1994, lailanqing, sunmenmian}@mail.nwpu.edu.cn + +Abstract. As a fundamental and extensively studied task in computer vision, image segmentation aims to locate and identify different semantic concepts at the pixel level. Recently, inspired by In-Context Learning (ICL), several generalist segmentation frameworks have been proposed, providing a promising paradigm for segmenting specific objects. However, existing works mostly ignore the value of visual prompts or simply apply similarity sorting to select contextual examples. In this paper, we focus on rethinking and improving the example selection strategy. By comprehensive comparisons, we first demonstrate that ICL-based segmentation models are sensitive to different contexts. Furthermore, empirical evidence indicates that the diversity of contextual prompts plays a crucial role in guiding segmentation. Based on the above insights, we propose a new stepwise context search method. Different from previous works, we construct a small yet rich candidate pool and adaptively search the well-matched contexts. More importantly, this method effectively reduces the annotation cost by compacting the search space. Extensive experiments show that our method is an effective strategy for selecting examples and enhancing segmentation performance. https://github.com/LanqingL/SCS + +Keywords: In-Context Learning $\cdot$ Visual Prompt $\cdot$ Image Segmentation + +# 1 Introduction + +Image segmentation is a fundamental task in computer vision, aiming to locate and identify different semantic concepts at the pixel level. It is essential in multiple applications such as autonomous driving [43,54], video surveillance [20,36,39,41] and complex reasoning [24,34,40]. To achieve precise segmentation, researchers have proposed a variety of segmentation models [5,12,23,29]. Although these methods have made considerable progress, they require training + +![](images/30f8de8dc374e8afceee8e3a4c1db12b7251d9cb28be9c7397871693fec642a0.jpg) +Fig. 1: Comparison of traditional and our method. (a) Existing works rely on dense annotation to build the exempling space. Then, they use the similarity sorting manner to select examples for given queries. (b) Instead, our method significantly alleviates the costs of annotation by searching the typical examples. Moreover, a novel adaptive search module is designed to further select well-matched contexts. + +the specialist model for each segmentation task. Recently, inspired by In-Context Learning (ICL) [26,30,49], several generalist segmentation frameworks have been proposed and achieved impressive performance such as MAE-VQGAN [6] and SegGPT [45]. Different from the traditional frameworks, ICL-based segmentation methods establish a promising paradigm for localizing vision concepts by feeding one or a few examples as references during inference. Meanwhile, the methodology offers a general solution for various real-world scenarios [44]. + +ICL learning is a relatively new research line in computer vision, while the fundamental ideas can be traced back to the Natural Language Processing (NLP) field [6, 44]. With the development of Large Language Models (LLM), numerous studies have already proved that selecting a suitable context is crucial for following human instruction and improving performance [1, 14, 31]. However, related researches about this topic is still limited in our community. As shown in Fig.1 (a), existing works typically randomly select examples [44] or simply apply similarity sorting to construct contextual prompts [37, 56]. The above insights motivate us to question: 1) Whether different contexts would significantly impact performance? 2) What is the critical factor of visual prompt selection for ICL-based segmentation models? In this paper, we first try to answer the above questions. By comprehensive comparisons, we find that the ICL-based models are sensitive to different examples. With multiple experiments, the performance gap between using different contextual examples even exceeds 5 points. Furthermore, we synthetically investigate the effectiveness of the existing prompt selection strategy. Surprisingly, although the similarity measure (i.e., choosing + +the most similar examples as contexts) provides a relatively effective approach for demonstration search, assembling the most dissimilar contexts results in better performance in $\sim 40\%$ of the test instances. To reveal the potential reasons, we empirically demonstrate that the diversity of contexts is a crucial factor in guiding segmentation and enhancing performance (more details in Sec. 3). + +Beyond the above discussions, we rethink the role of contextual examples and propose a new Stepwise Context Search method, which is called SCS for short. As shown in Fig. 1 (b), the SCS focuses on maintaining the diversity of the candidate pool and adaptively searching the well-matched contexts. Meanwhile, our method significantly reduces the annotation cost by searching the typical examples. Specifically, given a large unlabeled database, we first apply a basic cluster algorithm to compact the search space and improve annotation efficiency. Then, we extract the two most dissimilar members in each cluster as candidate examples. By collecting these representative members from all clusters, we construct a small yet rich candidate pool. Furthermore, considering the significant impact of examples on performance, we further introduce a novel adaptive search module based on reinforcement learning. It heuristically utilizes segmentation scores as rewards, guiding the model in determining final demonstrations. Extensive experiments across various datasets have shown the effectiveness of our method. To summarize, our contributions are as follows. + +- We take a step forward to systematically analyze the influence of example selection on the ICL-based segmentation frameworks and empirically demonstrate that the diversity of contexts is a key factor in guiding segmentation. +- We propose a new SCS method to efficiently select visual prompts by constructing a small yet rich candidate pool. Meanwhile, an adaptive search module is designed to further choose the well-matched demonstrations. +- The proposed model significantly improves the segmentation performance of ICL-based frameworks and achieves state-of-the-art on PASCAL-5 $^i$ [32], COCO-20 $^i$ [21] and iSALD-5 $^i$ [50]. + +# 2 Related work + +# 2.1 Visual Segmentation + +As a fundamental task in computer vision, visual segmentation entails locating and identifying different semantic concepts at the pixel level. Numerous works for specific tasks have continued to flourish over the years. For example, CrossoverNet [52] learns the vertical and horizontal crossover relation for medical image segmentation. RFNet [35] fuses RGB and depth information which is designed for real-time road-driving segmentation. [16] proposes DGEN, a densely connected global entropy network in which noise is suppressed by context information losses. MSANet [25] accurately extracts water bodies from remote sensing scenes by utilizing multiscale attention. Recently, inspired by ICL learning, several general segmentation methods have been proposed [18,44,45]. Different from these specialist models, the paradigm can locate visual concepts by feeding a few examples, known as the In-Context Learning (ICL) segmentation frameworks. + +# 2.2 In-Context Learning + +In-context learning (ICL) is a new learning paradigm that initially emerged from GPT-3 [7]. It presents a text fill-in-the-blank problem with examples, eliminating the necessity for fine-tuning model parameters on downstream NLP tasks. In the vision domain, MAE-VQGAN [6] transforms various visual tasks into grid inpainting problems and proposes the first ICL-based segmentation framework. Different from the previous segmentation methods, this new paradigm provides a promising manner to segment arbitrary objects with a few examples. Painter [44] adopts masked image modeling on continuous pixels, unifying the output form across different visual tasks. As a variant of Painter, SegGPT [45] integrates visual segmentation tasks into in-context coloring problems and demonstrates powerful capabilities in target segmentation. Beyond the previous works, we present comprehensive experiments on example selection for visual ICL models. + +# 2.3 Contextual Example Selection + +With the development of LLMs, the selection of contextual examples is important for ICL learning [1,30,51]. The emerging studies can be categorized into rule-based unsupervised methods and supervised methods. KATE [22] uses the KNN mechanism to enhance a retriever for selecting in-context examples. EPR [30] adopts a contrastive learning strategy to train a dense retriever. UDR [19] explores the process of retrieval demonstrations for various tasks in a unified manner. However, few studies investigate these effects in the vision domain. Current works typically select examples randomly [44] or simply apply similarity sorting to construct contextual demonstrations [37,56]. Different from the previous rule-based methods, we introduce an adaptive search method with a small yet rich candidate pool. This strategy effectively reduces the annotation cost and boosts segmentation performance. + +# 3 Preliminary + +In this section, we would first present a simple overview of the ICL-based segmentation frameworks. Then, we construct extensive experiments to comprehensively analyze the influence of different examples and the effectiveness of the existing selection strategies. Finally, we experimentally proved that contextual diversity is an important factor for guiding segmentation. + +# 3.1 Background + +The design of segmentation models has achieved tremendous development in recent years [20, 41, 52, 57]. However, these solutions require expensive annotation cost and face challenges in adapting to new scenarios. In contrast, the ICL-based segmentation frameworks [6, 45] can implement various specific tasks by leveraging several segmented examples. The convenient and versatile vision models provide a promising perspective for our community. + +As shown in Fig. 1, the goal of ICL-based segmentation methods is to predict segmentation regions based on exhibited one or a few prompts. By reorganizing the input format, the models can imitate provided contexts to mask specific object regions. Specifically, given a query image $I_q$ , these segmentation frameworks first select $n$ example images $E_d = \{I_i^d\}_{i=1}^n$ with corresponding masks $G_d = \{G_i^d\}_{i=1}^n$ as contextual demonstrations from a huge database. Here, $I_i^d$ and $G_i^d$ denote the $i$ -th selected example and corresponding mask. By reformating the input, models can perform multiple segmentation tasks across different scenarios. The above operations can be formulated as: + +$$ +\hat {I} _ {m} = \mathbf {S} \left(I _ {q}, E _ {d}, G _ {d}\right), \tag {1} +$$ + +where the $\mathbf{S}(\cdot)$ and $\hat{I}_m$ are the generalist segmentation model and predicted mask results, respectively. Although the above calculations provide a convenient way to segment objects at arbitrary granularities, the success of the paradigm greatly relies on the given contextual examples. In fact, the basic idea of ICL-based vision models comes from the NLP field, in which it has been proved that contextual prompts play an important role in executing reasoning and boosting performance. However, in the vision community, existing approaches mostly simply choose the demonstrations (i.e., $E_d$ and corresponding $G_d$ ) based on random sampling [44] or rely on similarity sorting [37,56]. To explore the impact of the visual prompts on performance, we construct a series of quantitative analyses on the PASCAL dataset [32] with the state-of-the-art ICL-based segmentation model SegGPT [45]. Note that the SegGPT involves two training stages (i.e., pre-training and in-context tuning). In the tuning process, the SegGPT depends on the whole annotation (including base and new classes) to improve the performance. Since the main advantage of the ICL-based framework is to implement few-shot tasks, we only use the pre-trained weights provided by the SegGPT [3] to construct experiments. Following [6,44], the mean intersection over union (mIoU) is adopted as the evaluation metric. + +# 3.2 Influence of Different Examples + +As mentioned above, research on the influence of different examples in segmentation remains insufficient. In this paper, we aim to fill this gap by conducting experiments to quantitatively analyze the impact of different demonstrations. Specifically, we randomly select examples from the entire training set with 5 different seeds. As shown in Fig. 2 (a), we report the best, worst, and average results across the 5 experiments. It reveals that allocating different contexts would dramatically affect the segmentation scores under the 1-shot and 5-shot settings. In particular, the best results can surpass the worst by margins of 5.6 and 4.1 points, respectively. Due to the huge performance discrepancy between different contextual prompts, designing an automated method for searching well-matched examples emerges as a critical challenge. + +![](images/3ba3f6b6695ae70ac197f798a37dcb2ff00526e74dfd05841a017b0dec74d603.jpg) +(a) + +![](images/9c0c4efc627eef0ee4845108d2e602c5516ede60697471f5dc06096b8f547f8f.jpg) +(b) +Fig. 2: The influence of different context selection. (a) Randomly sampling contextual examples with 5 runs under 1-shot and 5-shot setting on PASCAL. (b) For each instance, the nearest and farthest examples are retrieved as visual prompts across different similarity-based sorting. Surprisingly, the most dissimilar examples achieved better performance on $\sim 40\%$ of the test samples. + +# 3.3 Example Selection with Similarity + +To the best of our knowledge, the related researches to our work are [37, 56], where they simply rely on cosine similarity to select examples with qualitative analysis. In this paper, we try to rethink the strategy through a quantitative discussion across different similarity computation manners. Specifically, we explore two types of image similarity calculation strategies including appearance-level and semantic-level. For the appearance-level comparison, PSNR [47], SSIM [15], FSIM [55] and HaarPSI [28] obtain the similarity between the test instance and the entire training set. On the other hand, for semantic-level comparison, we follow [37, 56] to utilize the off-the-shelf image encoders (i.e., CLIP [27] and ViT [8]) for feature extraction. To verify the effectiveness of these similarity-based approaches, we carefully select the examples based on cosine similarity sorting. + +As shown in Fig. 2 (b), each metric is represented by two bars including nearest $>$ farthest (yellow bar) and nearest $<$ farthest (green bar). We take yellow as an example, it denotes the proportion of instances where using the most similar examples (nearest) as contexts results in a better IoU than using the most dissimilar ones (farthest), and the green bar indicates the opposite. Surprisingly, we observed that assembling the most dissimilar contexts leads to superior performance for numerous test instances. There is a substantial portion of the test samples ( $\sim 40\%$ ) that achieve better IoU scores by utilizing the farthest examples. The above results indicate that using similarity measures is not the only way to search for contextual examples. + +# 3.4 Example Selection with Diversity + +Based on the above analyses, it is imperative to explore a more effective way for searching visual prompts. In Fig. 2 (b), we find that applying the most dissimilar + +![](images/333d6e77700b90c23dd8c93475e3bc2594dd8fdd9450573fc1981a7317d96300.jpg) + +![](images/460ec976bd55dc37574f6d0a6bc629d41d69c6667eb1805ddab4ef5c21bf00c1.jpg) +(a) Example sampling with SegGPT [45]. +Fig. 3: Diversity vs Similarity. Based on similarity sorting, the performance of the two Nearest examples (NN), the two Farthest examples (FF), and the Nearest example with the Farthest example (NF) as visual prompts are shown. + +![](images/d305ee10253b43082161de2f63eb95fec69e1721d0dee63e99069ce3910ae665.jpg) + +![](images/a1081596559806d6cb7aa309a775c0670033f39b9a76b5077ac3101ae32a246c.jpg) +(b) Example sampling with MAE-VQGAN [6]. + +examples can also yield satisfactory performance for some instances. The results motivate us to investigate whether combining the nearest and farthest contexts can obtain a performance boost. Consequently, we further construct a series of experiments to verify the hypothesis across different ICL-based models. + +More specifically, based on similarity sorting, we use two Nearest examples (NN), two Farthest examples (FF), and a combination of the Nearest example and the Farthest example (NF) as contextual demonstrations, respectively. As shown in Fig. 3, we report related results across different similarity measure manners. It can be observed that the performance using NF examples significantly surpasses the other two selection strategies in all measures. Regarding ViT, the NF achieves better performance that outperforms equipping the two nearest examples by a margin of 2.4 points. + +# 3.5 Discussion + +In this section, we would try to summarize the above findings with a deeper discussion. Based on the insights presented in Sec. 3.2, we find that the ICL-based segmentation framework is sensitive to contextual examples. Given the different visual prompts, the performance gap is even over 5 points on the mIoU. Therefore, building an effective prompt selection method is important for boosting the segmentation results and achieving stable performance. + +To further explore the key factors that influence example selection, we design experiments in Sec. 3.3 and Sec. 3.4. The results demonstrate that combining both similar and dissimilar demonstrations can significantly enhance the IoU scores. We argue that the underlying reason is that this manner provides richer semantic information about the segmented objects by expanding the diversity of visual prompts. Based on the above insights, a natural idea is to select these + +considerable discrepancy examples as contextual prompts for each test instance. However, implementing the straightforward strategy would incur substantial annotation cost and complex computation, since each test sample needs to compare and rerank the candidate list. To mitigate the above problems, we introduce a novel stepwise context search (SCS) method that maintains the diversity of the candidate pool and reduces the cost of annotation. Moreover, considering the necessity of selecting appropriate contextual examples for different test queries, we further design an adaptive search module. It can dynamically select contexts from the candidate pool based on given test samples. + +# 4 Our Method + +Based on the discussion in Sec. 3, we design a new SCS method to construct a small yet rich candidate pool and adaptively search the contextual demonstrations. Compared to previous works, our method has two potential advantages: 1) Fewer Annotations. The SCS focuses on building a small yet diverse candidate pool. It only needs to annotate a portion of samples while achieving better performance. 2) Adapting Selection. Compared to the previous rule-based mechanism [37,56], our method can adaptively search the well-matched examples for different test instances. Next, we would introduce our method in detail. + +# 4.1 Constructing Candidate Pool + +As mentioned in Sec. 3, it is important to provide examples for the ICL-based segmentation paradigm. However, previous research mostly depend on huge annotations as contextual candidates, it would inevitably suffers from expensive annotation cost. Motivated by the of Sec. 3.5, we aim to build a small yet diverse candidate pool that can be used for subsequent context searching. + +Given $N$ unlabeled samples, follow in [56], we utilize CLIP to extract image feature into $d$ -dimensional vectors, represented by $F = \{f_{i}\}_{i=1}^{N}$ , $f_{i} \in \mathbb{R}^{d}$ , where $f_{i}$ is the $i$ -th encoded image feature. To construct a candidate pool comprising a set of representative examples, we employ k-means clustering algorithm [58] to generate $M$ clusters $C = \{c_{m}\}_{m=1}^{M}$ , which can be determined by minimizing the sum of squared intra-cluster norms: + +$$ +\arg \min _ {C} \sum_ {m = 1} ^ {M} \sum_ {f _ {i} \in c _ {j}} \| f _ {i} - \mu_ {m} \|, \tag {2} +$$ + +where $\mu_{m}$ is centroid of cluster $c_{m}$ . Initially, this algorithm randomly selects $M$ centroids. Each instance is assigned by computing the distance to all selected centroids. The process is iteratively executed until the clusters are stable. + +To construct the candidate pool, a simple method is to select members closest to the centroid, while previous studies have shown that samples closest to the center do not necessarily represent the entire cluster [11, 46]. Moreover, related works have provided several alternative sampling methods such as density + +![](images/df3994d57e3a7c22667b526a3dc10a846ee3a30739ced14e5ddf894e0176e41b.jpg) +Fig. 4: Overview of our SCS method. Instead of similarity sorting on a large annotated dataset, we use clustering to select diverse examples from unlabeled data $D$ and construct the candidate pool. Meanwhile, the search agent is used to further select contextual demonstrations for various test samples based on reinforcement learning. During inference, the test sample and candidate pool examples are fed into the model and adaptively search visual prompts. + +sorting [46] and entropy sorting [11]. In this paper, we design a simpler and more efficient sampling strategy without complex computation. Specifically, for each cluster $c_{m}$ , which includes members $F_{m} = \{f_{k}\}_{k=1}^{K}$ , where $K$ denotes the number of members in $c_{m}$ . We sort them into a list $[f_{k}^{(1)}, f_{k}^{(2)}, \dots, f_{k}^{(K)}]$ in the ascending order of their distance to the cluster's center. Based on the findings of Sec. 3.3, aggregating the nearest and farthest examples can cover more diverse demonstrations. Thus, we consider the $f_{mk}^{(1)}$ and $f_{mk}^{(K)}$ as typical samples for each cluster. By collecting these samples from all clusters, we obtain the final candidate pool $D = \{f_{1k}^{(1)}, f_{1k}^{(K)}, f_{2k}^{(1)}, \dots, f_{Mk}^{(1)}, f_{Mk}^{(K)}\}$ , which contains a total of $2M$ samples. To simplify, we use $f_{d}^{m}$ and $I_{d}^{m}$ to represent $m$ -th image feature and corresponding image, respectively. + +# 4.2 Adaptive Search Module + +In this module, our goal is to select well-matched examples from the constructed candidate pool. As shown in Fig. 4, we propose a new adaptive search module that can further search and assemble the suitable contexts for given queries. + +Specifically, for a given test instance $I_{q}$ , we first apply the same image encoder (i.e. CLIP) to extract the corresponding feature vector $f_{q} \in \mathbb{R}^{d}$ . Then, we concatenate the image feature $f_{q}$ and each example $f_{d}^{m}$ from the candidate pool. These fused features are fed into a search agent that consists of a group of light and simple Multi-Layer Perceptrons (MLP). The search computation can be denoted as follows: + +$$ +s _ {d} ^ {m} = \operatorname {M L P} \left(\left[ f _ {d} ^ {m}; f _ {q} \right]\right), \tag {3} +$$ + +$$ +a _ {d} ^ {1}, a _ {d} ^ {2}, \dots , a _ {d} ^ {2 M} = \operatorname {s o f t m a x} \left(s _ {d} ^ {1}, s _ {d} ^ {2}, \dots , s _ {d} ^ {2 M}\right), \tag {4} +$$ + +where $a_{d}^{m}$ represents the probability of $m$ -th example in the candidate pool. To guide the search agent in learning select examples, we apply the IoU scores as rewards for selecting well-matched demonstrations from the candidate pool. Specifically, following the format of the ICL-based segmentation model, we re-organize query $I_{q}$ , the original image of context examples $I_{d}$ and corresponding mask $G_{d}$ into a joint image, as shown in Fig. 4. These images are then fed into the ICL-based segmentation model to obtain the IoU scores $U = \{u_{d}\}_{d=1}^{2M}$ . Inspired by [4,38], we use the average IoU score $u_{avg}$ as the base score. By applying reinforcement learning, the gradient is calculated by: + +$$ +\nabla_ {\theta} L (\theta) = - \frac {1}{2 M} \sum_ {m = 1} ^ {2 M} \left(u _ {d} ^ {m} - u _ {a v g}\right) \nabla_ {\theta} \log \left(a _ {d} ^ {m}\right), \tag {5} +$$ + +where $a_{d}^{m}$ and $\theta$ are the probability of $m$ -th example and the parameters of the search agent. Based on the above computation, this gradient would tend to increase the probability of the $m$ -th context when its IoU scores $u_{d}^{m}$ are higher than the average IoU score $u_{avg}$ . + +During inference, we directly input the test samples and candidate pool examples into the search agent. The contextual examples are then determined by sorting them based on probability scores $a_{d}$ as detailed in Eq. 4. + +# 5 Experiment + +We use different prompt selection strategies and various generalist segmentation frameworks to demonstrate the superiority of our method. Furthermore, we compare our approach with traditional methods and conduct extensive qualitative and quantitative studies on multiple datasets. + +# 5.1 Experimental Setting + +Dataset. We follow the previous works [6, 45, 56] and evaluate our approach on PASCAL- $5^{i}$ [32], COCO- $20^{i}$ [21] and iSALD- $5^{i}$ [50]. PASCAL- $5^{i}$ includes 20 semantic categories and is created from PASCAL VOC 2012 [9] with external annotations from SDS [10]. COCO- $20^{i}$ is built upon MSCOCO [21], which is a more challenging dataset and contains 80 categories. For all categories, both COCO- $20^{i}$ and PASCAL- $5^{i}$ are evenly divided into 4 folds for cross-validation, where each fold is selected as the test set and the remaining three folds are used for training. The iSALD- $5^{i}$ contains 15 geospatial categories which derived from iSALD [48] in the remote sensing image semantic segmentation task. Similarly, the iSALD- $5^{i}$ is also partitioned into 3 folds for cross-validation. + +Evaluation Metric. Following [13], mean intersection over union (mIoU) and foreground-background IoU (FBIoU) are adopted as evaluation metrics. The former measures the average overall IoU values for all object classes, while the latter reports the average IoU scores for the foreground and background. + +Table 1: The results on COCO- $20^{i}$ and PASCAL- $5^{i}$ based on example-based semantic segmentation. To ensure a fair comparison and follow the few-shot setting, $\dagger$ indicates that only using the pre-trained weights provided by the SegGPT [3], and we report related results based on randomly sampling (i.e., w/R) and similarity sorting [56]. + +
MethodCOCO-20i [21]PASCAL-5i [32]
1-shot5-shot1-shot5-shot
mIoUFBIoUmIoUFBIoUmIoUFBIoUmIoUFBIoU
Painter [44]32.850.232.650.164.574.264.674.4
MAE-VQGAN [6]12.145.7--27.555.6--
MAE-VQGAN w/ [56]18.046.5--33.657.8--
SegGPT† w/R [45]48.671.352.773.172.183.176.280.3
SegGPT† w/ [56]53.372.256.175.574.484.476.985.9
MAE-VQGAN w/Ours19.147.0--35.061.2--
SegGPT† w/Ours57.275.159.576.876.685.478.286.5
+ +Implementation details. For both datasets, the Adam [17] is used as our optimizer and the maximum number of training epochs is 10. We use the CLIPViT [27] image encoder to extract image features. During training, all images are resized to $224 \times 224$ and fed into the image encoder. The dimension of extracted feature vectors is 1024. We train on one 4080 GPU with an initial learning rate of 1e-3, which is dropped by half every 5 epochs. The number of clusters $M$ and batch size is set to 10 and 8. + +# 5.2 Quantitative Evaluation + +In Table 1, we show a performance comparison of two general ICL-based segmentation models with and without proposed SCS. Following [37,56], all results are reported by repeating the random experiment five times. It can be observed that our method can significantly boost performance in all settings including 1-shot and 5-shot settings. In particular, for the 1-shot setting, the SCS leads to MAE-VQGAN with the mIoU gain of 7.0 and 6.5 on the PASCAL- $5^{i}$ and COCO- $20^{i}$ dataset, respectively. Besides, compared to previous rule-base models [56] in 1-shot and 5-shot semantic segmentation tasks, our method surpasses it by a considerable margin. Moreover, following the few-shot setting and ensuring a fair comparison, we report related results based on the pre-trained weights for SegGPT [3]. It can be observed that our method significantly outperforms these rule-based visual prompt selection methods for SegGPT by 9.7 and 6.8 on the 1-shot and 5-shot settings for COCO- $20^{i}$ . More importantly, the above results demonstrate that the proposed SCS can be considered as a "plug-and-play" module to enhance the segmentation ability of ICL-based frameworks. + +Table 2: Ablation study. Similarity: Cosine similarity ranking; Cluster: Cluster compacting; Sorting: The nearest-farthest sampling; ASM: Adaptive search module. + +
SimilarityClusterSortingASMmIoUFBIoU
152.773.1
2✓56.174.5
3✓✓55.774.3
4✓✓✓57.675.7
5✓✓✓59.576.8
+ +Table 3: The effects of different settings for our method. + +
Model1-shot5-shot
mIoUFBIoUmIoUFBIoU
1Entropy53.172.356.175.0
2Foreground53.572.556.779.3
3ViT56.975.058.776.0
4CLIP-resnet56.374.758.576.4
5Density-sort57.675.159.176.3
6Entropy-sort56.974.558.475.9
7Ours57.275.159.576.8
+ +# 5.3 Ablation Study + +To clarify the effectiveness of our method, we conduct further ablation studies on the COCO- $20^{i}$ dataset with the 5-shot setting. The SegGPT is used as the basic model due to its outstanding segmentation performance. + +Effects of Cluster. In Table 2, we first build a baseline by randomly selecting examples from the entire training set. The average result is reported across 5 different seeds. In the second row of Table 2, we follow [56] and utilize cosine similarity ranking to select the top-5 examples as contexts. In the third row, we investigate the impact of the cluster on in-context example selection, using the closest examples to each cluster centroid to construct the candidate pool. Compared to the baseline, the results show that compacting the search space not only reduces annotation cost but also improves model performance. + +Effects of Sort. As shown in row 4 of Table 2, we experimentally demonstrate that integrating our sorting mechanism results in $1.9\mathrm{mIoU}$ improvement compared to directly using the closest examples to each cluster centroid. The inherent reason is that centroid examples describe a limited feature space [33, 46], making it challenging to support a variety of different contexts for test instances. In contrast, our sorting mechanism introduces the nearest and farthest examples, which constructs more diverse demonstrations for ICL learning. + +Effects of Adaptive Search Module. Finally, we further explore the effectiveness of our adaptive search module in the last row of Table 2. Compared to using a simple similarity sorting method, our approach shows an improve + +Table 4: The results on PASCAL-5 $^i$ [32] Table 5: The results on iSALD-5 $^i$ [48] based on instance segmentation. based on semantic segmentation. + +
Model1-shot5-shot
SegGPT† w/R [45]46.546.9
SegGPT† w/ [56]47.148.5
SegGPT w/Ours50.551.4
+ +
Model1-shot5-shot
SegGPT† w/R [45]29.830.6
SegGPT† w/ [56]34.735.4
SegGPT w/Ours37.939.5
+ +ment of 3.4 points. Meanwhile, benefiting from the diverse candidate pool, our SCS achieves state-of-the-art performance and outperforms the baseline by a considerable margin. + +# 5.4 Different Model Settings + +In this section, we explore several alternative model settings to further discuss the proposed method. All experiments are conducted on the COCO- $20^{i}$ with the SegGPT model. In the first two rows, we first compare the performance using different metrics. Specifically, we use the rule-based methods to select the contextual examples from the complete training set, which including Entropy sorting [11] and Foreground sorting [53]. In particular, "Entropy" is executed by calculating the maximum entropy values. The "Foreground" refers to a two-stage approach [53]. It uses an object detector [42] to extract the foreground boxes and the foreground similarity is used as the metric. From the 1-shot and 5-shot results, it can be observed that our SCS can dramatically outperform these methods by a large margin. In rows 3-4, we construct the experiments with different feature encoders. We replace the CLIP-ViT with ViT [8] and CLIP-resnet [27] to obtain image feature vectors. These extracted features would be utilized to construct the candidate pool and perform the adaptive searching. The results show that our SCS is robust for different image encoders. Finally, in rows 5-6 of Table 3, we replace our selection mechanism with several alternative manners. However, these approaches mostly involve complex computation. For example, the Density-sort [2] requires additional computations on the order of $O(MN)$ in each iteration to update density lists, where $M$ is the number of clusters and $N$ is the size of the unlabeled data. Instead, our method uses the basic clustering algorithm and obtains a comparable performance. + +# 5.5 Further Discussion + +More Segmentation tasks. In this section, we evaluate the effect of our method on more segmentation tasks and datasets. Because MAE-VQGAN cannot distinguish between individual instances of objects within the same class, we construct more experiments with the SegGPT model. Specifically, we report instance segmentation results on PASCAL-5i in Table 4 and aerial image semantic segmentation results on iSALD-5i in Table 5. The above results indicate that the SCS can enhance the segmentation ability for various real-world scenarios. + +![](images/07ebfd07fd83de029507aa2c73371a7b1fef1abe48c42da646e31893c0e33c0e.jpg) +Fig.5: Qualitative results under 1-shot setting on COCO- $20^{i}$ . The similarity-based selection method is viewed as our baseline. The green regions, red regions and purple regions are example masks, predicted masks and ground-truth masks. + +Qualitative Results. As shown in Fig. 5, to further analyze and verify the proposed method, we visualized the qualitative segmentation results on COCO $_{20^i}$ with the SegGPT. The first and third columns represent the visual prompts (highlighted in green) based on similarity sorting and our SCS method, respectively. The second and fourth columns show the corresponding segmentation results (highlighted in red). The last column represents the ground truth mask (highlighted in purple). As illustrated in the first row, the baseline model selects the most similar context to the query (i.e., chopping board, knife and carrot), while the segment result is not satisfactory, as the model can not recognize the knife in the image. Instead, Our method selects an image from the candidate pool that accurately segments out the mask. The above results demonstrate that similarity sorting is not the only golden standard and improving the diversity of context prompts can boost the segmentation results. + +# 6 Conclusion + +In this paper, we focus on the example selection method for the ICL-based segmentation frameworks. By building comprehensive comparisons, we empirically demonstrate that diversity of contexts is a key factor in guiding segmentation and boosting performance. Beyond the above insights, we propose a new stepwise context search method that improves the diversity of the candidate pool and significantly alleviates the annotation cost. We hope that this work can promote further research to better understand in-context learning in computer vision. + +Acknowledgement. This work was supported by National Natural Science Foundation of China (No.62102323), the Innovation Capability Support Program of Shaanxi(Program No. 2023KJXX-142) and National Natural Science Foundation of China (U23B2013). + +# References + +1. Agrawal, S., Zhou, C., Lewis, M., Zettlemoyer, L., Ghazvininejad, M.: In-context examples selection for machine translation. arXiv preprint arXiv:2212.02437 (2022) +2. et al., W.: "unsupervised selective labeling for more effective semi-supervised learning". In: ECCV 2022. pp. 427-445. Springer (2022) +3. et al, W.: "seggspt: Segmenting everything in context". In ICCV 2023 (2023) +4. Anderson, P., He, X., Buehler, C., Teney, D., Johnson, M., Gould, S., Zhang, L.: Bottom-up and top-down attention for image captioning and visual question answering. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 6077-6086 (2018) +5. Badrinarayanan, V., Kendall, A., Cipolla, R.: Segnet: A deep convolutional encoder-decoder architecture for image segmentation. IEEE transactions on pattern analysis and machine intelligence 39(12), 2481-2495 (2017) +6. Bar, A., Gandelsman, Y., Darrell, T., Globerson, A., Efros, A.: Visual prompting via image inpainting. Advances in Neural Information Processing Systems 35, 25005-25017 (2022) +7. Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J.D., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., et al.: Language models are few-shot learners. Advances in neural information processing systems 33, 1877-1901 (2020) +8. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929 (2020) +9. Everingham, M., Van Gool, L., Williams, C.K., Winn, J., Zisserman, A.: The Pascal visual object classes (voc) challenge. International journal of computer vision 88, 303-338 (2010) +0. Hariharan, B., Arbeláez, P., Bourdev, L., Maji, S., Malik, J.: Semantic contours from inverse detectors. In: 2011 international conference on computer vision. pp. 991-998. IEEE (2011) +1. Haussmann, E., Fenzi, M., Chitta, K., Ivanecky, J., Xu, H., Roy, D., Mittel, A., Koumchatzky, N., Farabet, C., Alvarez, J.M.: Scalable active learning for object detection. In: 2020 IEEE intelligent vehicles symposium (iv). pp. 1430-1435. IEEE (2020) +2. He, K., Gkioxari, G., Dollár, P., Girshick, R.: Mask r-cnn. In: Proceedings of the IEEE international conference on computer vision. pp. 2961-2969 (2017) +3. Hong, S., Cho, S., Nam, J., Lin, S., Kim, S.: Cost aggregation with 4d convolutional swim transformer for few-shot segmentation. In: European Conference on Computer Vision. pp. 108-126. Springer (2022) +4. Hongjin, S., Kasai, J., Wu, C.H., Shi, W., Wang, T., Xin, J., Zhang, R., Ostendorf, M., Zettlemoyer, L., Smith, N.A., et al.: Selective annotation makes language models better few-shot learners. In: The Eleventh International Conference on Learning Representations (2022) + +15. Hore, A., Ziou, D.: Image quality metrics: Psnr vs. sSIM. In: 2010 20th international conference on pattern recognition. pp. 2366-2369. IEEE (2010) +16. Hu, H., Li, Z., Li, L., Yang, H., Zhu, H.: Classification of very high-resolution remote sensing imagery using a fully convolutional network with global and local context information enhancements. IEEE Access 8, 14606-14619 (2020) +17. Kingma, D.P., Ba, J.: Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980 (2014) +18. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., et al.: Segment anything. arXiv preprint arXiv:2304.02643 (2023) +19. Li, X., Lv, K., Yan, H., Lin, T., Zhu, W., Ni, Y., Xie, G., Wang, X., Qiu, X.: Unified demonstration retriever for in-context learning. arXiv preprint arXiv:2305.04320 (2023) +20. Li, Y., Padmanabhan, A., Zhao, P., Wang, Y., Xu, G.H., Netravali, R.: Reducto: On-camera filtering for resource-efficient real-time video analytics. In: Proceedings of the Annual conference of the ACM Special Interest Group on Data Communication on the applications, technologies, architectures, and protocols for computer communication. pp. 359-376 (2020) +21. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollar, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13. pp. 740-755. Springer (2014) +22. Liu, J., Shen, D., Zhang, Y., Dolan, B., Carin, L., Chen, W.: What makes good in-context examples for gpt-3? arXiv preprint arXiv:2101.06804 (2021) +23. Long, J., Shelhamer, E., Darrell, T.: Fully convolutional networks for semantic segmentation. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 3431-3440 (2015) +24. Luo, G., Zhou, Y., Sun, X., Cao, L., Wu, C., Deng, C., Ji, R.: Multi-task collaborative network for joint referring expression comprehension and segmentation. In: Proceedings of the IEEE/CVF Conference on computer vision and pattern recognition. pp. 10034-10043 (2020) +25. Lyu, X., Fang, Y., Tong, B., Li, X., Zeng, T.: Multiscale normalization attention network for water body extraction from remote sensing imagery. Remote Sensing 14(19) (2022). https://doi.org/10.3390/rs14194983, https://www.mdpi.com/2072-4292/14/19/4983 +26. Min, S., Lyu, X., Holtzman, A., Artetxe, M., Lewis, M., Hajishirzi, H., Zettlemoyer, L.: Rethinking the role of demonstrations: What makes in-context learning work? arXiv preprint arXiv:2202.12837 (2022) +27. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021) +28. Reisenhofer, R., Bosse, S., Kutyniok, G., Wiegand, T.: Aaar wavelet-based perceptual similarity index for image quality assessment. Signal Processing: Image Communication 61, 33-43 (2018) +29. Ronneberger, O., Fischer, P., Brox, T.: U-net: Convolutional networks for biomedical image segmentation. In: Medical Image Computing and Computer-Assisted Intervention-MICCAI 2015: 18th International Conference, Munich, Germany, October 5-9, 2015, Proceedings, Part III 18. pp. 234-241. Springer (2015) +30. Rubin, O., Herzig, J., Berant, J.: Learning to retrieve prompts for in-context learning. arXiv preprint arXiv:2112.08633 (2021) + +31. Rubin, O., Herzig, J., Berant, J.: Learning to retrieve prompts for in-context learning. In: Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies. pp. 2655-2671 (2022) +32. Shaban, A., Bansal, S., Liu, Z., Essa, I., Boots, B.: One-shot learning for semantic segmentation. arXiv preprint arXiv:1709.03410 (2017) +33. Shen, F., Zhao, L., Du, W., Zhong, W., Qian, F.: Large-scale industrial energy systems optimization under uncertainty: A data-driven robust optimization approach. Applied Energy 259, 114199 (2020) +34. Shi, H., Li, H., Meng, F., Wu, Q.: Key-word-aware network for referring expression image segmentation. In: Proceedings of the European Conference on Computer Vision (ECCV). pp. 38-54 (2018) +35. Sun, L., Yang, K., Hu, X., Hu, W., Wang, K.: Real-time fusion network forrgb-d semantic segmentation incorporating unexpected obstacle detection for road-driving images. IEEE robotics and automation letters 5(4), 5558-5565 (2020) +36. Sun, M., Suo, W., Wang, P., Niu, K., Liu, L., Lin, G., Zhang, Y., Wu, Q.: An adaptive correlation filtering method for text-based person search. International Journal of Computer Vision pp. 1-16 (2024) +37. Sun, Y., Chen, Q., Wang, J., Wang, J., Li, Z.: Exploring effective factors for improving visual in-context learning. arXiv preprint arXiv:2304.04748 (2023) +38. Suo, W., Sun, M., Liu, W., Gao, Y., Wang, P., Zhang, Y., Wu, Q.: S3c: Semi-supervised vqa natural language explanation via self-critical learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2646-2656 (2023) +39. Suo, W., Sun, M., Niu, K., Gao, Y., Wang, P., Zhang, Y., Wu, Q.: A simple and robust correlation filtering method for text-based person search. In: European conference on computer vision. pp. 726-742. Springer (2022) +40. Suo, W., Sun, M., Wang, P., Zhang, Y., Wu, Q.: Rethinking and improving feature pyramids for one-stage referring expression comprehension. IEEE Transactions on Image Processing 32, 854-864 (2022) +41. Vasudevan, A.B., Dai, D., Van Gool, L.: Semantic object prediction and spatial sound super-resolution with binaural sounds. In: European conference on computer vision. pp. 638-655. Springer (2020) +42. Wang, W., Dai, J., Chen, Z., Huang, Z., Li, Z., Zhu, X., Hu, X., Lu, T., Lu, L., Li, H., et al.: Intermimage: Exploring large-scale vision foundation models with deformable convolutions. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14408-14419 (2023) +43. Wang, X., Ma, H., You, S.: Deep clustering for weakly-supervised semantic segmentation in autonomous driving scenes. Neurocomputing 381, 20-28 (2020) +44. Wang, X., Wang, W., Cao, Y., Shen, C., Huang, T.: Images speak in images: A generalist painter for in-context visual learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6830-6839 (2023) +45. Wang, X., Zhang, X., Cao, Y., Wang, W., Shen, C., Huang, T.: Seggpt: Segmenting everything in context. arXiv preprint arXiv:2304.03284 (2023) +46. Wang, X., Lian, L., Yu, S.X.: Unsupervised selective labeling for more effective semi-supervised learning. In: European Conference on Computer Vision. pp. 427-445. Springer (2022) +47. Wang, Z., Bovik, A.C., Sheikh, H.R., Simoncelli, E.P.: Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing 13(4), 600-612 (2004) + +48. Waqas Zamir, S., Arora, A., Gupta, A., Khan, S., Sun, G., Shahbaz Khan, F., Zhu, F., Shao, L., Xia, G.S., Bai, X.: isaid: A large-scale dataset for instance segmentation in aerial images. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops. pp. 28-37 (2019) +49. Wei, J., Wang, X., Schuurmans, D., Bosma, M., Xia, F., Chi, E., Le, Q.V., Zhou, D., et al.: Chain-of-thought prompting elicits reasoning in large language models. Advances in Neural Information Processing Systems 35, 24824-24837 (2022) +50. Yao, X., Cao, Q., Feng, X., Cheng, G., Han, J.: Scale-aware detailed matching for few-shot aerial image semantic segmentation. IEEE Transactions on Geoscience and Remote Sensing 60, 1-11 (2021) +51. Ye, J., Wu, Z., Feng, J., Yu, T., Kong, L.: Compositional exemplars for in-context learning. arXiv preprint arXiv:2302.05698 (2023) +52. Yu, Q., Gao, Y., Zheng, Y., Zhu, J., Dai, Y., Shi, Y.: Crossover-net: Leveraging vertical-horizontal crossover relation for robust medical image segmentation. Pattern Recognition 113, 107756 (2021) +53. Yu, Y., Yang, X., Li, Q., Zhou, Y., Da, F., Yan, J.: H2rbox-v2: Incorporating symmetry for boosting horizontal box supervised oriented object detection. In: Thirty-seventh Conference on Neural Information Processing Systems (2023) +54. Zendel, O., Schörghuber, M., Rainer, B., Murschitz, M., Beleznai, C.: Unifying panoptic segmentation for autonomous driving. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 21351-21360 (2022) +55. Zhang, L., Zhang, L., Mou, X., Zhang, D.: Fsim: A feature similarity index for image quality assessment. IEEE transactions on Image Processing 20(8), 2378-2386 (2011) +56. Zhang, Y., Zhou, K., Liu, Z.: What makes good examples for visual in-context learning? arXiv preprint arXiv:2301.13670 (2023) +57. Zhang, Y., Liu, H., Hu, Q.: Transfuse: Fusing transformers and cnns for medical image segmentation. In: Medical Image Computing and Computer Assisted Intervention-MICCAI 2021: 24th International Conference, Strasbourg, France, September 27-October 1, 2021, Proceedings, Part I 24. pp. 14-24. Springer (2021) +58. Zhang, Z., Zhang, A., Li, M., Smola, A.: Automatic chain of thought prompting in large language models. arXiv preprint arXiv:2210.03493 (2022) \ No newline at end of file diff --git a/2024/Rethinking and Improving Visual Prompt Selection for In-Context Learning Segmentation Framework/images.zip b/2024/Rethinking and Improving Visual Prompt Selection for In-Context Learning Segmentation Framework/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..10a459b20ab8c4a5ba8aef2aabd3feffdd5514d1 --- /dev/null +++ b/2024/Rethinking and Improving Visual Prompt Selection for In-Context Learning Segmentation Framework/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9865f0584b059a0c2a493aa117357b0721c168e9bd6ff8ed876e513a7d05519d +size 465964 diff --git a/2024/Rethinking and Improving Visual Prompt Selection for In-Context Learning Segmentation Framework/layout.json b/2024/Rethinking and Improving Visual Prompt Selection for In-Context Learning Segmentation Framework/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..60afef1db9bca142643fc529fab5871643c6577e --- /dev/null +++ b/2024/Rethinking and Improving Visual Prompt Selection for In-Context Learning Segmentation Framework/layout.json @@ -0,0 +1,9151 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 140, + 111, + 474, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 140, + 111, + 474, + 148 + ], + "spans": [ + { + "bbox": [ + 140, + 111, + 474, + 148 + ], + "type": "text", + "content": "Rethinking and Improving Visual Prompt Selection for In-Context Learning Segmentation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 138, + 172, + 477, + 206 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 172, + 477, + 206 + ], + "spans": [ + { + "bbox": [ + 138, + 172, + 477, + 206 + ], + "type": "text", + "content": "Wei Suo" + }, + { + "bbox": [ + 138, + 172, + 477, + 206 + ], + "type": "inline_equation", + "content": "^{1,3,\\star}" + }, + { + "bbox": [ + 138, + 172, + 477, + 206 + ], + "type": "text", + "content": ", Lanqing Lai" + }, + { + "bbox": [ + 138, + 172, + 477, + 206 + ], + "type": "inline_equation", + "content": "^{1,\\star}" + }, + { + "bbox": [ + 138, + 172, + 477, + 206 + ], + "type": "text", + "content": ", Mengyang Sun" + }, + { + "bbox": [ + 138, + 172, + 477, + 206 + ], + "type": "inline_equation", + "content": "^{2,3}" + }, + { + "bbox": [ + 138, + 172, + 477, + 206 + ], + "type": "text", + "content": ", Hanwang Zhang" + }, + { + "bbox": [ + 138, + 172, + 477, + 206 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 138, + 172, + 477, + 206 + ], + "type": "text", + "content": ", Peng Wang" + }, + { + "bbox": [ + 138, + 172, + 477, + 206 + ], + "type": "inline_equation", + "content": "^{1,\\dagger}" + }, + { + "bbox": [ + 138, + 172, + 477, + 206 + ], + "type": "text", + "content": ", and Yanning Zhang" + }, + { + "bbox": [ + 138, + 172, + 477, + 206 + ], + "type": "inline_equation", + "content": "^{1}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 142, + 214, + 470, + 270 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 142, + 214, + 470, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 142, + 214, + 470, + 236 + ], + "spans": [ + { + "bbox": [ + 142, + 214, + 470, + 236 + ], + "type": "text", + "content": "1 School of Computer Science and Ningbo Institute, Northwestern Polytechnical University, Xi'an, China" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 143, + 237, + 468, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 237, + 468, + 247 + ], + "spans": [ + { + "bbox": [ + 143, + 237, + 468, + 247 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 143, + 237, + 468, + 247 + ], + "type": "text", + "content": " School of Cybersecurity, Northwestern Polytechnical University, Xi'an, China" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 149, + 248, + 462, + 270 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 248, + 462, + 270 + ], + "spans": [ + { + "bbox": [ + 149, + 248, + 462, + 270 + ], + "type": "text", + "content": "3 School of Computer Science, Nanyang Technological University, Singapore {suwei1994, lailanqing, sunmenmian}@mail.nwpu.edu.cn" + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 159, + 299, + 455, + 497 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 299, + 455, + 497 + ], + "spans": [ + { + "bbox": [ + 159, + 299, + 455, + 497 + ], + "type": "text", + "content": "Abstract. As a fundamental and extensively studied task in computer vision, image segmentation aims to locate and identify different semantic concepts at the pixel level. Recently, inspired by In-Context Learning (ICL), several generalist segmentation frameworks have been proposed, providing a promising paradigm for segmenting specific objects. However, existing works mostly ignore the value of visual prompts or simply apply similarity sorting to select contextual examples. In this paper, we focus on rethinking and improving the example selection strategy. By comprehensive comparisons, we first demonstrate that ICL-based segmentation models are sensitive to different contexts. Furthermore, empirical evidence indicates that the diversity of contextual prompts plays a crucial role in guiding segmentation. Based on the above insights, we propose a new stepwise context search method. Different from previous works, we construct a small yet rich candidate pool and adaptively search the well-matched contexts. More importantly, this method effectively reduces the annotation cost by compacting the search space. Extensive experiments show that our method is an effective strategy for selecting examples and enhancing segmentation performance. https://github.com/LanqingL/SCS" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 160, + 507, + 453, + 519 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 507, + 453, + 519 + ], + "spans": [ + { + "bbox": [ + 160, + 507, + 453, + 519 + ], + "type": "text", + "content": "Keywords: In-Context Learning " + }, + { + "bbox": [ + 160, + 507, + 453, + 519 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 507, + 453, + 519 + ], + "type": "text", + "content": " Visual Prompt " + }, + { + "bbox": [ + 160, + 507, + 453, + 519 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 507, + 453, + 519 + ], + "type": "text", + "content": " Image Segmentation" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 539, + 230, + 552 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 539, + 230, + 552 + ], + "spans": [ + { + "bbox": [ + 132, + 539, + 230, + 552 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 563, + 482, + 636 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 563, + 482, + 636 + ], + "spans": [ + { + "bbox": [ + 130, + 563, + 482, + 636 + ], + "type": "text", + "content": "Image segmentation is a fundamental task in computer vision, aiming to locate and identify different semantic concepts at the pixel level. It is essential in multiple applications such as autonomous driving [43,54], video surveillance [20,36,39,41] and complex reasoning [24,34,40]. To achieve precise segmentation, researchers have proposed a variety of segmentation models [5,12,23,29]. Although these methods have made considerable progress, they require training" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 642, + 235, + 653 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 642, + 235, + 653 + ], + "spans": [ + { + "bbox": [ + 133, + 642, + 235, + 653 + ], + "type": "text", + "content": "† Corresponding author" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 135, + 654, + 337, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 654, + 337, + 665 + ], + "spans": [ + { + "bbox": [ + 135, + 654, + 337, + 665 + ], + "type": "text", + "content": "* These authors contributed equally to this work." + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 177, + 114, + 439, + 301 + ], + "blocks": [ + { + "bbox": [ + 177, + 114, + 439, + 301 + ], + "lines": [ + { + "bbox": [ + 177, + 114, + 439, + 301 + ], + "spans": [ + { + "bbox": [ + 177, + 114, + 439, + 301 + ], + "type": "image", + "image_path": "30f8de8dc374e8afceee8e3a4c1db12b7251d9cb28be9c7397871693fec642a0.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 309, + 482, + 365 + ], + "lines": [ + { + "bbox": [ + 130, + 309, + 482, + 365 + ], + "spans": [ + { + "bbox": [ + 130, + 309, + 482, + 365 + ], + "type": "text", + "content": "Fig. 1: Comparison of traditional and our method. (a) Existing works rely on dense annotation to build the exempling space. Then, they use the similarity sorting manner to select examples for given queries. (b) Instead, our method significantly alleviates the costs of annotation by searching the typical examples. Moreover, a novel adaptive search module is designed to further select well-matched contexts." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 390, + 482, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 390, + 482, + 474 + ], + "spans": [ + { + "bbox": [ + 130, + 390, + 482, + 474 + ], + "type": "text", + "content": "the specialist model for each segmentation task. Recently, inspired by In-Context Learning (ICL) [26,30,49], several generalist segmentation frameworks have been proposed and achieved impressive performance such as MAE-VQGAN [6] and SegGPT [45]. Different from the traditional frameworks, ICL-based segmentation methods establish a promising paradigm for localizing vision concepts by feeding one or a few examples as references during inference. Meanwhile, the methodology offers a general solution for various real-world scenarios [44]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 474, + 482, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 474, + 482, + 668 + ], + "spans": [ + { + "bbox": [ + 130, + 474, + 482, + 668 + ], + "type": "text", + "content": "ICL learning is a relatively new research line in computer vision, while the fundamental ideas can be traced back to the Natural Language Processing (NLP) field [6, 44]. With the development of Large Language Models (LLM), numerous studies have already proved that selecting a suitable context is crucial for following human instruction and improving performance [1, 14, 31]. However, related researches about this topic is still limited in our community. As shown in Fig.1 (a), existing works typically randomly select examples [44] or simply apply similarity sorting to construct contextual prompts [37, 56]. The above insights motivate us to question: 1) Whether different contexts would significantly impact performance? 2) What is the critical factor of visual prompt selection for ICL-based segmentation models? In this paper, we first try to answer the above questions. By comprehensive comparisons, we find that the ICL-based models are sensitive to different examples. With multiple experiments, the performance gap between using different contextual examples even exceeds 5 points. Furthermore, we synthetically investigate the effectiveness of the existing prompt selection strategy. Surprisingly, although the similarity measure (i.e., choosing" + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 222, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 222, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 222, + 100 + ], + "type": "text", + "content": "W. Suo et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 176 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 176 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 176 + ], + "type": "text", + "content": "the most similar examples as contexts) provides a relatively effective approach for demonstration search, assembling the most dissimilar contexts results in better performance in " + }, + { + "bbox": [ + 130, + 116, + 479, + 176 + ], + "type": "inline_equation", + "content": "\\sim 40\\%" + }, + { + "bbox": [ + 130, + 116, + 479, + 176 + ], + "type": "text", + "content": " of the test instances. To reveal the potential reasons, we empirically demonstrate that the diversity of contexts is a crucial factor in guiding segmentation and enhancing performance (more details in Sec. 3)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 176, + 480, + 355 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 176, + 480, + 355 + ], + "spans": [ + { + "bbox": [ + 130, + 176, + 480, + 355 + ], + "type": "text", + "content": "Beyond the above discussions, we rethink the role of contextual examples and propose a new Stepwise Context Search method, which is called SCS for short. As shown in Fig. 1 (b), the SCS focuses on maintaining the diversity of the candidate pool and adaptively searching the well-matched contexts. Meanwhile, our method significantly reduces the annotation cost by searching the typical examples. Specifically, given a large unlabeled database, we first apply a basic cluster algorithm to compact the search space and improve annotation efficiency. Then, we extract the two most dissimilar members in each cluster as candidate examples. By collecting these representative members from all clusters, we construct a small yet rich candidate pool. Furthermore, considering the significant impact of examples on performance, we further introduce a novel adaptive search module based on reinforcement learning. It heuristically utilizes segmentation scores as rewards, guiding the model in determining final demonstrations. Extensive experiments across various datasets have shown the effectiveness of our method. To summarize, our contributions are as follows." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 360, + 480, + 466 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 138, + 360, + 479, + 395 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 360, + 479, + 395 + ], + "spans": [ + { + "bbox": [ + 138, + 360, + 479, + 395 + ], + "type": "text", + "content": "- We take a step forward to systematically analyze the influence of example selection on the ICL-based segmentation frameworks and empirically demonstrate that the diversity of contexts is a key factor in guiding segmentation." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 396, + 480, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 396, + 480, + 430 + ], + "spans": [ + { + "bbox": [ + 138, + 396, + 480, + 430 + ], + "type": "text", + "content": "- We propose a new SCS method to efficiently select visual prompts by constructing a small yet rich candidate pool. Meanwhile, an adaptive search module is designed to further choose the well-matched demonstrations." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 430, + 480, + 466 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 430, + 480, + 466 + ], + "spans": [ + { + "bbox": [ + 138, + 430, + 480, + 466 + ], + "type": "text", + "content": "- The proposed model significantly improves the segmentation performance of ICL-based frameworks and achieves state-of-the-art on PASCAL-5" + }, + { + "bbox": [ + 138, + 430, + 480, + 466 + ], + "type": "inline_equation", + "content": "^i" + }, + { + "bbox": [ + 138, + 430, + 480, + 466 + ], + "type": "text", + "content": " [32], COCO-20" + }, + { + "bbox": [ + 138, + 430, + 480, + 466 + ], + "type": "inline_equation", + "content": "^i" + }, + { + "bbox": [ + 138, + 430, + 480, + 466 + ], + "type": "text", + "content": " [21] and iSALD-5" + }, + { + "bbox": [ + 138, + 430, + 480, + 466 + ], + "type": "inline_equation", + "content": "^i" + }, + { + "bbox": [ + 138, + 430, + 480, + 466 + ], + "type": "text", + "content": " [50]." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 132, + 482, + 234, + 495 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 482, + 234, + 495 + ], + "spans": [ + { + "bbox": [ + 132, + 482, + 234, + 495 + ], + "type": "text", + "content": "2 Related work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 504, + 264, + 517 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 504, + 264, + 517 + ], + "spans": [ + { + "bbox": [ + 132, + 504, + 264, + 517 + ], + "type": "text", + "content": "2.1 Visual Segmentation" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 482, + 666 + ], + "type": "text", + "content": "As a fundamental task in computer vision, visual segmentation entails locating and identifying different semantic concepts at the pixel level. Numerous works for specific tasks have continued to flourish over the years. For example, CrossoverNet [52] learns the vertical and horizontal crossover relation for medical image segmentation. RFNet [35] fuses RGB and depth information which is designed for real-time road-driving segmentation. [16] proposes DGEN, a densely connected global entropy network in which noise is suppressed by context information losses. MSANet [25] accurately extracts water bodies from remote sensing scenes by utilizing multiscale attention. Recently, inspired by ICL learning, several general segmentation methods have been proposed [18,44,45]. Different from these specialist models, the paradigm can locate visual concepts by feeding a few examples, known as the In-Context Learning (ICL) segmentation frameworks." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 190, + 91, + 447, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 190, + 91, + 447, + 103 + ], + "spans": [ + { + "bbox": [ + 190, + 91, + 447, + 103 + ], + "type": "text", + "content": "Visual Prompt Selection for In-Context Learning Segmentation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 264, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 264, + 129 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 264, + 129 + ], + "type": "text", + "content": "2.2 In-Context Learning" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 134, + 482, + 278 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 134, + 482, + 278 + ], + "spans": [ + { + "bbox": [ + 130, + 134, + 482, + 278 + ], + "type": "text", + "content": "In-context learning (ICL) is a new learning paradigm that initially emerged from GPT-3 [7]. It presents a text fill-in-the-blank problem with examples, eliminating the necessity for fine-tuning model parameters on downstream NLP tasks. In the vision domain, MAE-VQGAN [6] transforms various visual tasks into grid inpainting problems and proposes the first ICL-based segmentation framework. Different from the previous segmentation methods, this new paradigm provides a promising manner to segment arbitrary objects with a few examples. Painter [44] adopts masked image modeling on continuous pixels, unifying the output form across different visual tasks. As a variant of Painter, SegGPT [45] integrates visual segmentation tasks into in-context coloring problems and demonstrates powerful capabilities in target segmentation. Beyond the previous works, we present comprehensive experiments on example selection for visual ICL models." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 295, + 313, + 307 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 295, + 313, + 307 + ], + "spans": [ + { + "bbox": [ + 132, + 295, + 313, + 307 + ], + "type": "text", + "content": "2.3 Contextual Example Selection" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 313, + 482, + 458 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 313, + 482, + 458 + ], + "spans": [ + { + "bbox": [ + 130, + 313, + 482, + 458 + ], + "type": "text", + "content": "With the development of LLMs, the selection of contextual examples is important for ICL learning [1,30,51]. The emerging studies can be categorized into rule-based unsupervised methods and supervised methods. KATE [22] uses the KNN mechanism to enhance a retriever for selecting in-context examples. EPR [30] adopts a contrastive learning strategy to train a dense retriever. UDR [19] explores the process of retrieval demonstrations for various tasks in a unified manner. However, few studies investigate these effects in the vision domain. Current works typically select examples randomly [44] or simply apply similarity sorting to construct contextual demonstrations [37,56]. Different from the previous rule-based methods, we introduce an adaptive search method with a small yet rich candidate pool. This strategy effectively reduces the annotation cost and boosts segmentation performance." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 474, + 227, + 488 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 474, + 227, + 488 + ], + "spans": [ + { + "bbox": [ + 132, + 474, + 227, + 488 + ], + "type": "text", + "content": "3 Preliminary" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 498, + 482, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 498, + 482, + 559 + ], + "spans": [ + { + "bbox": [ + 130, + 498, + 482, + 559 + ], + "type": "text", + "content": "In this section, we would first present a simple overview of the ICL-based segmentation frameworks. Then, we construct extensive experiments to comprehensively analyze the influence of different examples and the effectiveness of the existing selection strategies. Finally, we experimentally proved that contextual diversity is an important factor for guiding segmentation." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 574, + 222, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 574, + 222, + 586 + ], + "spans": [ + { + "bbox": [ + 132, + 574, + 222, + 586 + ], + "type": "text", + "content": "3.1 Background" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": "The design of segmentation models has achieved tremendous development in recent years [20, 41, 52, 57]. However, these solutions require expensive annotation cost and face challenges in adapting to new scenarios. In contrast, the ICL-based segmentation frameworks [6, 45] can implement various specific tasks by leveraging several segmented examples. The convenient and versatile vision models provide a promising perspective for our community." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 222, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 222, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 222, + 101 + ], + "type": "text", + "content": "W. Suo et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "content": "As shown in Fig. 1, the goal of ICL-based segmentation methods is to predict segmentation regions based on exhibited one or a few prompts. By reorganizing the input format, the models can imitate provided contexts to mask specific object regions. Specifically, given a query image " + }, + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "inline_equation", + "content": "I_q" + }, + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "content": ", these segmentation frameworks first select " + }, + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "content": " example images " + }, + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "inline_equation", + "content": "E_d = \\{I_i^d\\}_{i=1}^n" + }, + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "content": " with corresponding masks " + }, + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "inline_equation", + "content": "G_d = \\{G_i^d\\}_{i=1}^n" + }, + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "content": " as contextual demonstrations from a huge database. Here, " + }, + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "inline_equation", + "content": "I_i^d" + }, + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "inline_equation", + "content": "G_i^d" + }, + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "content": " denote the " + }, + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "content": "-th selected example and corresponding mask. By reformating the input, models can perform multiple segmentation tasks across different scenarios. The above operations can be formulated as:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 265, + 236, + 481, + 250 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 265, + 236, + 481, + 250 + ], + "spans": [ + { + "bbox": [ + 265, + 236, + 481, + 250 + ], + "type": "interline_equation", + "content": "\\hat {I} _ {m} = \\mathbf {S} \\left(I _ {q}, E _ {d}, G _ {d}\\right), \\tag {1}", + "image_path": "bc11176a52767051021baa97b7ee8758208d5e3ffe051c5435041391b8b3cdd8.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 262, + 482, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 262, + 482, + 479 + ], + "spans": [ + { + "bbox": [ + 130, + 262, + 482, + 479 + ], + "type": "text", + "content": "where the " + }, + { + "bbox": [ + 130, + 262, + 482, + 479 + ], + "type": "inline_equation", + "content": "\\mathbf{S}(\\cdot)" + }, + { + "bbox": [ + 130, + 262, + 482, + 479 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 262, + 482, + 479 + ], + "type": "inline_equation", + "content": "\\hat{I}_m" + }, + { + "bbox": [ + 130, + 262, + 482, + 479 + ], + "type": "text", + "content": " are the generalist segmentation model and predicted mask results, respectively. Although the above calculations provide a convenient way to segment objects at arbitrary granularities, the success of the paradigm greatly relies on the given contextual examples. In fact, the basic idea of ICL-based vision models comes from the NLP field, in which it has been proved that contextual prompts play an important role in executing reasoning and boosting performance. However, in the vision community, existing approaches mostly simply choose the demonstrations (i.e., " + }, + { + "bbox": [ + 130, + 262, + 482, + 479 + ], + "type": "inline_equation", + "content": "E_d" + }, + { + "bbox": [ + 130, + 262, + 482, + 479 + ], + "type": "text", + "content": " and corresponding " + }, + { + "bbox": [ + 130, + 262, + 482, + 479 + ], + "type": "inline_equation", + "content": "G_d" + }, + { + "bbox": [ + 130, + 262, + 482, + 479 + ], + "type": "text", + "content": ") based on random sampling [44] or rely on similarity sorting [37,56]. To explore the impact of the visual prompts on performance, we construct a series of quantitative analyses on the PASCAL dataset [32] with the state-of-the-art ICL-based segmentation model SegGPT [45]. Note that the SegGPT involves two training stages (i.e., pre-training and in-context tuning). In the tuning process, the SegGPT depends on the whole annotation (including base and new classes) to improve the performance. Since the main advantage of the ICL-based framework is to implement few-shot tasks, we only use the pre-trained weights provided by the SegGPT [3] to construct experiments. Following [6,44], the mean intersection over union (mIoU) is adopted as the evaluation metric." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 505, + 320, + 517 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 505, + 320, + 517 + ], + "spans": [ + { + "bbox": [ + 132, + 505, + 320, + 517 + ], + "type": "text", + "content": "3.2 Influence of Different Examples" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": "As mentioned above, research on the influence of different examples in segmentation remains insufficient. In this paper, we aim to fill this gap by conducting experiments to quantitatively analyze the impact of different demonstrations. Specifically, we randomly select examples from the entire training set with 5 different seeds. As shown in Fig. 2 (a), we report the best, worst, and average results across the 5 experiments. It reveals that allocating different contexts would dramatically affect the segmentation scores under the 1-shot and 5-shot settings. In particular, the best results can surpass the worst by margins of 5.6 and 4.1 points, respectively. Due to the huge performance discrepancy between different contextual prompts, designing an automated method for searching well-matched examples emerges as a critical challenge." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 189, + 91, + 447, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 189, + 91, + 447, + 103 + ], + "spans": [ + { + "bbox": [ + 189, + 91, + 447, + 103 + ], + "type": "text", + "content": "Visual Prompt Selection for In-Context Learning Segmentation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 132, + 116, + 289, + 232 + ], + "blocks": [ + { + "bbox": [ + 132, + 116, + 289, + 232 + ], + "lines": [ + { + "bbox": [ + 132, + 116, + 289, + 232 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 289, + 232 + ], + "type": "image", + "image_path": "3ba3f6b6695ae70ac197f798a37dcb2ff00526e74dfd05841a017b0dec74d603.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 211, + 233, + 223, + 243 + ], + "lines": [ + { + "bbox": [ + 211, + 233, + 223, + 243 + ], + "spans": [ + { + "bbox": [ + 211, + 233, + 223, + 243 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 308, + 113, + 460, + 232 + ], + "blocks": [ + { + "bbox": [ + 308, + 113, + 460, + 232 + ], + "lines": [ + { + "bbox": [ + 308, + 113, + 460, + 232 + ], + "spans": [ + { + "bbox": [ + 308, + 113, + 460, + 232 + ], + "type": "image", + "image_path": "9c0c4efc627eef0ee4845108d2e602c5516ede60697471f5dc06096b8f547f8f.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 387, + 233, + 400, + 243 + ], + "lines": [ + { + "bbox": [ + 387, + 233, + 400, + 243 + ], + "spans": [ + { + "bbox": [ + 387, + 233, + 400, + 243 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 251, + 482, + 307 + ], + "lines": [ + { + "bbox": [ + 130, + 251, + 482, + 307 + ], + "spans": [ + { + "bbox": [ + 130, + 251, + 482, + 307 + ], + "type": "text", + "content": "Fig. 2: The influence of different context selection. (a) Randomly sampling contextual examples with 5 runs under 1-shot and 5-shot setting on PASCAL. (b) For each instance, the nearest and farthest examples are retrieved as visual prompts across different similarity-based sorting. Surprisingly, the most dissimilar examples achieved better performance on " + }, + { + "bbox": [ + 130, + 251, + 482, + 307 + ], + "type": "inline_equation", + "content": "\\sim 40\\%" + }, + { + "bbox": [ + 130, + 251, + 482, + 307 + ], + "type": "text", + "content": " of the test samples." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 328, + 332, + 340 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 328, + 332, + 340 + ], + "spans": [ + { + "bbox": [ + 131, + 328, + 332, + 340 + ], + "type": "text", + "content": "3.3 Example Selection with Similarity" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 345, + 482, + 489 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 345, + 482, + 489 + ], + "spans": [ + { + "bbox": [ + 130, + 345, + 482, + 489 + ], + "type": "text", + "content": "To the best of our knowledge, the related researches to our work are [37, 56], where they simply rely on cosine similarity to select examples with qualitative analysis. In this paper, we try to rethink the strategy through a quantitative discussion across different similarity computation manners. Specifically, we explore two types of image similarity calculation strategies including appearance-level and semantic-level. For the appearance-level comparison, PSNR [47], SSIM [15], FSIM [55] and HaarPSI [28] obtain the similarity between the test instance and the entire training set. On the other hand, for semantic-level comparison, we follow [37, 56] to utilize the off-the-shelf image encoders (i.e., CLIP [27] and ViT [8]) for feature extraction. To verify the effectiveness of these similarity-based approaches, we carefully select the examples based on cosine similarity sorting." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 490, + 483, + 609 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 490, + 483, + 609 + ], + "spans": [ + { + "bbox": [ + 130, + 490, + 483, + 609 + ], + "type": "text", + "content": "As shown in Fig. 2 (b), each metric is represented by two bars including nearest " + }, + { + "bbox": [ + 130, + 490, + 483, + 609 + ], + "type": "inline_equation", + "content": ">" + }, + { + "bbox": [ + 130, + 490, + 483, + 609 + ], + "type": "text", + "content": " farthest (yellow bar) and nearest " + }, + { + "bbox": [ + 130, + 490, + 483, + 609 + ], + "type": "inline_equation", + "content": "<" + }, + { + "bbox": [ + 130, + 490, + 483, + 609 + ], + "type": "text", + "content": " farthest (green bar). We take yellow as an example, it denotes the proportion of instances where using the most similar examples (nearest) as contexts results in a better IoU than using the most dissimilar ones (farthest), and the green bar indicates the opposite. Surprisingly, we observed that assembling the most dissimilar contexts leads to superior performance for numerous test instances. There is a substantial portion of the test samples (" + }, + { + "bbox": [ + 130, + 490, + 483, + 609 + ], + "type": "inline_equation", + "content": "\\sim 40\\%" + }, + { + "bbox": [ + 130, + 490, + 483, + 609 + ], + "type": "text", + "content": ") that achieve better IoU scores by utilizing the farthest examples. The above results indicate that using similarity measures is not the only way to search for contextual examples." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 624, + 330, + 636 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 624, + 330, + 636 + ], + "spans": [ + { + "bbox": [ + 131, + 624, + 330, + 636 + ], + "type": "text", + "content": "3.4 Example Selection with Diversity" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 482, + 666 + ], + "type": "text", + "content": "Based on the above analyses, it is imperative to explore a more effective way for searching visual prompts. In Fig. 2 (b), we find that applying the most dissimilar" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 222, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 222, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 222, + 100 + ], + "type": "text", + "content": "W. Suo et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 132, + 114, + 271, + 240 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 271, + 240 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 271, + 240 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 271, + 240 + ], + "type": "image", + "image_path": "333d6e77700b90c23dd8c93475e3bc2594dd8fdd9450573fc1981a7317d96300.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 143, + 243, + 241, + 264 + ], + "blocks": [ + { + "bbox": [ + 143, + 243, + 241, + 264 + ], + "lines": [ + { + "bbox": [ + 143, + 243, + 241, + 264 + ], + "spans": [ + { + "bbox": [ + 143, + 243, + 241, + 264 + ], + "type": "image", + "image_path": "460ec976bd55dc37574f6d0a6bc629d41d69c6667eb1805ddab4ef5c21bf00c1.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 143, + 264, + 290, + 275 + ], + "lines": [ + { + "bbox": [ + 143, + 264, + 290, + 275 + ], + "spans": [ + { + "bbox": [ + 143, + 264, + 290, + 275 + ], + "type": "text", + "content": "(a) Example sampling with SegGPT [45]." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 284, + 482, + 318 + ], + "lines": [ + { + "bbox": [ + 130, + 284, + 482, + 318 + ], + "spans": [ + { + "bbox": [ + 130, + 284, + 482, + 318 + ], + "type": "text", + "content": "Fig. 3: Diversity vs Similarity. Based on similarity sorting, the performance of the two Nearest examples (NN), the two Farthest examples (FF), and the Nearest example with the Farthest example (NF) as visual prompts are shown." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 304, + 113, + 447, + 240 + ], + "blocks": [ + { + "bbox": [ + 304, + 113, + 447, + 240 + ], + "lines": [ + { + "bbox": [ + 304, + 113, + 447, + 240 + ], + "spans": [ + { + "bbox": [ + 304, + 113, + 447, + 240 + ], + "type": "image", + "image_path": "d305ee10253b43082161de2f63eb95fec69e1721d0dee63e99069ce3910ae665.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 334, + 243, + 417, + 264 + ], + "blocks": [ + { + "bbox": [ + 334, + 243, + 417, + 264 + ], + "lines": [ + { + "bbox": [ + 334, + 243, + 417, + 264 + ], + "spans": [ + { + "bbox": [ + 334, + 243, + 417, + 264 + ], + "type": "image", + "image_path": "a1081596559806d6cb7aa309a775c0670033f39b9a76b5077ac3101ae32a246c.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 311, + 265, + 476, + 275 + ], + "lines": [ + { + "bbox": [ + 311, + 265, + 476, + 275 + ], + "spans": [ + { + "bbox": [ + 311, + 265, + 476, + 275 + ], + "type": "text", + "content": "(b) Example sampling with MAE-VQGAN [6]." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 340, + 482, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 340, + 482, + 389 + ], + "spans": [ + { + "bbox": [ + 130, + 340, + 482, + 389 + ], + "type": "text", + "content": "examples can also yield satisfactory performance for some instances. The results motivate us to investigate whether combining the nearest and farthest contexts can obtain a performance boost. Consequently, we further construct a series of experiments to verify the hypothesis across different ICL-based models." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 388, + 482, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 388, + 482, + 485 + ], + "spans": [ + { + "bbox": [ + 130, + 388, + 482, + 485 + ], + "type": "text", + "content": "More specifically, based on similarity sorting, we use two Nearest examples (NN), two Farthest examples (FF), and a combination of the Nearest example and the Farthest example (NF) as contextual demonstrations, respectively. As shown in Fig. 3, we report related results across different similarity measure manners. It can be observed that the performance using NF examples significantly surpasses the other two selection strategies in all measures. Regarding ViT, the NF achieves better performance that outperforms equipping the two nearest examples by a margin of 2.4 points." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 131, + 502, + 214, + 513 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 502, + 214, + 513 + ], + "spans": [ + { + "bbox": [ + 131, + 502, + 214, + 513 + ], + "type": "text", + "content": "3.5 Discussion" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 521, + 482, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 482, + 593 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 482, + 593 + ], + "type": "text", + "content": "In this section, we would try to summarize the above findings with a deeper discussion. Based on the insights presented in Sec. 3.2, we find that the ICL-based segmentation framework is sensitive to contextual examples. Given the different visual prompts, the performance gap is even over 5 points on the mIoU. Therefore, building an effective prompt selection method is important for boosting the segmentation results and achieving stable performance." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": "To further explore the key factors that influence example selection, we design experiments in Sec. 3.3 and Sec. 3.4. The results demonstrate that combining both similar and dissimilar demonstrations can significantly enhance the IoU scores. We argue that the underlying reason is that this manner provides richer semantic information about the segmented objects by expanding the diversity of visual prompts. Based on the above insights, a natural idea is to select these" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 189, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 189, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 189, + 91, + 447, + 102 + ], + "type": "text", + "content": "Visual Prompt Selection for In-Context Learning Segmentation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "content": "considerable discrepancy examples as contextual prompts for each test instance. However, implementing the straightforward strategy would incur substantial annotation cost and complex computation, since each test sample needs to compare and rerank the candidate list. To mitigate the above problems, we introduce a novel stepwise context search (SCS) method that maintains the diversity of the candidate pool and reduces the cost of annotation. Moreover, considering the necessity of selecting appropriate contextual examples for different test queries, we further design an adaptive search module. It can dynamically select contexts from the candidate pool based on given test samples." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 242, + 230, + 255 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 242, + 230, + 255 + ], + "spans": [ + { + "bbox": [ + 132, + 242, + 230, + 255 + ], + "type": "text", + "content": "4 Our Method" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 266, + 482, + 362 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 266, + 482, + 362 + ], + "spans": [ + { + "bbox": [ + 130, + 266, + 482, + 362 + ], + "type": "text", + "content": "Based on the discussion in Sec. 3, we design a new SCS method to construct a small yet rich candidate pool and adaptively search the contextual demonstrations. Compared to previous works, our method has two potential advantages: 1) Fewer Annotations. The SCS focuses on building a small yet diverse candidate pool. It only needs to annotate a portion of samples while achieving better performance. 2) Adapting Selection. Compared to the previous rule-based mechanism [37,56], our method can adaptively search the well-matched examples for different test instances. Next, we would introduce our method in detail." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 380, + 308, + 392 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 380, + 308, + 392 + ], + "spans": [ + { + "bbox": [ + 132, + 380, + 308, + 392 + ], + "type": "text", + "content": "4.1 Constructing Candidate Pool" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 399, + 481, + 459 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 399, + 481, + 459 + ], + "spans": [ + { + "bbox": [ + 130, + 399, + 481, + 459 + ], + "type": "text", + "content": "As mentioned in Sec. 3, it is important to provide examples for the ICL-based segmentation paradigm. However, previous research mostly depend on huge annotations as contextual candidates, it would inevitably suffers from expensive annotation cost. Motivated by the of Sec. 3.5, we aim to build a small yet diverse candidate pool that can be used for subsequent context searching." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 459, + 482, + 532 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 459, + 482, + 532 + ], + "spans": [ + { + "bbox": [ + 130, + 459, + 482, + 532 + ], + "type": "text", + "content": "Given " + }, + { + "bbox": [ + 130, + 459, + 482, + 532 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 459, + 482, + 532 + ], + "type": "text", + "content": " unlabeled samples, follow in [56], we utilize CLIP to extract image feature into " + }, + { + "bbox": [ + 130, + 459, + 482, + 532 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 130, + 459, + 482, + 532 + ], + "type": "text", + "content": "-dimensional vectors, represented by " + }, + { + "bbox": [ + 130, + 459, + 482, + 532 + ], + "type": "inline_equation", + "content": "F = \\{f_{i}\\}_{i=1}^{N}" + }, + { + "bbox": [ + 130, + 459, + 482, + 532 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 459, + 482, + 532 + ], + "type": "inline_equation", + "content": "f_{i} \\in \\mathbb{R}^{d}" + }, + { + "bbox": [ + 130, + 459, + 482, + 532 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 459, + 482, + 532 + ], + "type": "inline_equation", + "content": "f_{i}" + }, + { + "bbox": [ + 130, + 459, + 482, + 532 + ], + "type": "text", + "content": " is the " + }, + { + "bbox": [ + 130, + 459, + 482, + 532 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 459, + 482, + 532 + ], + "type": "text", + "content": "-th encoded image feature. To construct a candidate pool comprising a set of representative examples, we employ k-means clustering algorithm [58] to generate " + }, + { + "bbox": [ + 130, + 459, + 482, + 532 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 130, + 459, + 482, + 532 + ], + "type": "text", + "content": " clusters " + }, + { + "bbox": [ + 130, + 459, + 482, + 532 + ], + "type": "inline_equation", + "content": "C = \\{c_{m}\\}_{m=1}^{M}" + }, + { + "bbox": [ + 130, + 459, + 482, + 532 + ], + "type": "text", + "content": ", which can be determined by minimizing the sum of squared intra-cluster norms:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 246, + 540, + 481, + 574 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 246, + 540, + 481, + 574 + ], + "spans": [ + { + "bbox": [ + 246, + 540, + 481, + 574 + ], + "type": "interline_equation", + "content": "\\arg \\min _ {C} \\sum_ {m = 1} ^ {M} \\sum_ {f _ {i} \\in c _ {j}} \\| f _ {i} - \\mu_ {m} \\|, \\tag {2}", + "image_path": "1139b552ccbe1c951d822b21bf20e3a561a378838aba8208bd53fdb4c27e64b8.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 582, + 481, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 481, + 617 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 481, + 617 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 582, + 481, + 617 + ], + "type": "inline_equation", + "content": "\\mu_{m}" + }, + { + "bbox": [ + 130, + 582, + 481, + 617 + ], + "type": "text", + "content": " is centroid of cluster " + }, + { + "bbox": [ + 130, + 582, + 481, + 617 + ], + "type": "inline_equation", + "content": "c_{m}" + }, + { + "bbox": [ + 130, + 582, + 481, + 617 + ], + "type": "text", + "content": ". Initially, this algorithm randomly selects " + }, + { + "bbox": [ + 130, + 582, + 481, + 617 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 130, + 582, + 481, + 617 + ], + "type": "text", + "content": " centroids. Each instance is assigned by computing the distance to all selected centroids. The process is iteratively executed until the clusters are stable." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 618, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 618, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 618, + 482, + 666 + ], + "type": "text", + "content": "To construct the candidate pool, a simple method is to select members closest to the centroid, while previous studies have shown that samples closest to the center do not necessarily represent the entire cluster [11, 46]. Moreover, related works have provided several alternative sampling methods such as density" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 222, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 222, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 222, + 101 + ], + "type": "text", + "content": "W. Suo et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 115, + 478, + 250 + ], + "blocks": [ + { + "bbox": [ + 133, + 115, + 478, + 250 + ], + "lines": [ + { + "bbox": [ + 133, + 115, + 478, + 250 + ], + "spans": [ + { + "bbox": [ + 133, + 115, + 478, + 250 + ], + "type": "image", + "image_path": "df3994d57e3a7c22667b526a3dc10a846ee3a30739ced14e5ddf894e0176e41b.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 259, + 482, + 326 + ], + "lines": [ + { + "bbox": [ + 130, + 259, + 482, + 326 + ], + "spans": [ + { + "bbox": [ + 130, + 259, + 482, + 326 + ], + "type": "text", + "content": "Fig. 4: Overview of our SCS method. Instead of similarity sorting on a large annotated dataset, we use clustering to select diverse examples from unlabeled data " + }, + { + "bbox": [ + 130, + 259, + 482, + 326 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 130, + 259, + 482, + 326 + ], + "type": "text", + "content": " and construct the candidate pool. Meanwhile, the search agent is used to further select contextual demonstrations for various test samples based on reinforcement learning. During inference, the test sample and candidate pool examples are fed into the model and adaptively search visual prompts." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 347, + 482, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 347, + 482, + 485 + ], + "spans": [ + { + "bbox": [ + 130, + 347, + 482, + 485 + ], + "type": "text", + "content": "sorting [46] and entropy sorting [11]. In this paper, we design a simpler and more efficient sampling strategy without complex computation. Specifically, for each cluster " + }, + { + "bbox": [ + 130, + 347, + 482, + 485 + ], + "type": "inline_equation", + "content": "c_{m}" + }, + { + "bbox": [ + 130, + 347, + 482, + 485 + ], + "type": "text", + "content": ", which includes members " + }, + { + "bbox": [ + 130, + 347, + 482, + 485 + ], + "type": "inline_equation", + "content": "F_{m} = \\{f_{k}\\}_{k=1}^{K}" + }, + { + "bbox": [ + 130, + 347, + 482, + 485 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 347, + 482, + 485 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 130, + 347, + 482, + 485 + ], + "type": "text", + "content": " denotes the number of members in " + }, + { + "bbox": [ + 130, + 347, + 482, + 485 + ], + "type": "inline_equation", + "content": "c_{m}" + }, + { + "bbox": [ + 130, + 347, + 482, + 485 + ], + "type": "text", + "content": ". We sort them into a list " + }, + { + "bbox": [ + 130, + 347, + 482, + 485 + ], + "type": "inline_equation", + "content": "[f_{k}^{(1)}, f_{k}^{(2)}, \\dots, f_{k}^{(K)}]" + }, + { + "bbox": [ + 130, + 347, + 482, + 485 + ], + "type": "text", + "content": " in the ascending order of their distance to the cluster's center. Based on the findings of Sec. 3.3, aggregating the nearest and farthest examples can cover more diverse demonstrations. Thus, we consider the " + }, + { + "bbox": [ + 130, + 347, + 482, + 485 + ], + "type": "inline_equation", + "content": "f_{mk}^{(1)}" + }, + { + "bbox": [ + 130, + 347, + 482, + 485 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 347, + 482, + 485 + ], + "type": "inline_equation", + "content": "f_{mk}^{(K)}" + }, + { + "bbox": [ + 130, + 347, + 482, + 485 + ], + "type": "text", + "content": " as typical samples for each cluster. By collecting these samples from all clusters, we obtain the final candidate pool " + }, + { + "bbox": [ + 130, + 347, + 482, + 485 + ], + "type": "inline_equation", + "content": "D = \\{f_{1k}^{(1)}, f_{1k}^{(K)}, f_{2k}^{(1)}, \\dots, f_{Mk}^{(1)}, f_{Mk}^{(K)}\\}" + }, + { + "bbox": [ + 130, + 347, + 482, + 485 + ], + "type": "text", + "content": ", which contains a total of " + }, + { + "bbox": [ + 130, + 347, + 482, + 485 + ], + "type": "inline_equation", + "content": "2M" + }, + { + "bbox": [ + 130, + 347, + 482, + 485 + ], + "type": "text", + "content": " samples. To simplify, we use " + }, + { + "bbox": [ + 130, + 347, + 482, + 485 + ], + "type": "inline_equation", + "content": "f_{d}^{m}" + }, + { + "bbox": [ + 130, + 347, + 482, + 485 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 347, + 482, + 485 + ], + "type": "inline_equation", + "content": "I_{d}^{m}" + }, + { + "bbox": [ + 130, + 347, + 482, + 485 + ], + "type": "text", + "content": " to represent " + }, + { + "bbox": [ + 130, + 347, + 482, + 485 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 130, + 347, + 482, + 485 + ], + "type": "text", + "content": "-th image feature and corresponding image, respectively." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 500, + 285, + 511 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 500, + 285, + 511 + ], + "spans": [ + { + "bbox": [ + 132, + 500, + 285, + 511 + ], + "type": "text", + "content": "4.2 Adaptive Search Module" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 518, + 481, + 554 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 518, + 481, + 554 + ], + "spans": [ + { + "bbox": [ + 130, + 518, + 481, + 554 + ], + "type": "text", + "content": "In this module, our goal is to select well-matched examples from the constructed candidate pool. As shown in Fig. 4, we propose a new adaptive search module that can further search and assemble the suitable contexts for given queries." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 554, + 482, + 625 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 554, + 482, + 625 + ], + "spans": [ + { + "bbox": [ + 130, + 554, + 482, + 625 + ], + "type": "text", + "content": "Specifically, for a given test instance " + }, + { + "bbox": [ + 130, + 554, + 482, + 625 + ], + "type": "inline_equation", + "content": "I_{q}" + }, + { + "bbox": [ + 130, + 554, + 482, + 625 + ], + "type": "text", + "content": ", we first apply the same image encoder (i.e. CLIP) to extract the corresponding feature vector " + }, + { + "bbox": [ + 130, + 554, + 482, + 625 + ], + "type": "inline_equation", + "content": "f_{q} \\in \\mathbb{R}^{d}" + }, + { + "bbox": [ + 130, + 554, + 482, + 625 + ], + "type": "text", + "content": ". Then, we concatenate the image feature " + }, + { + "bbox": [ + 130, + 554, + 482, + 625 + ], + "type": "inline_equation", + "content": "f_{q}" + }, + { + "bbox": [ + 130, + 554, + 482, + 625 + ], + "type": "text", + "content": " and each example " + }, + { + "bbox": [ + 130, + 554, + 482, + 625 + ], + "type": "inline_equation", + "content": "f_{d}^{m}" + }, + { + "bbox": [ + 130, + 554, + 482, + 625 + ], + "type": "text", + "content": " from the candidate pool. These fused features are fed into a search agent that consists of a group of light and simple Multi-Layer Perceptrons (MLP). The search computation can be denoted as follows:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 266, + 633, + 480, + 646 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 266, + 633, + 480, + 646 + ], + "spans": [ + { + "bbox": [ + 266, + 633, + 480, + 646 + ], + "type": "interline_equation", + "content": "s _ {d} ^ {m} = \\operatorname {M L P} \\left(\\left[ f _ {d} ^ {m}; f _ {q} \\right]\\right), \\tag {3}", + "image_path": "7e942f81a9ffe5e07f14aef933639f8c5a379f6c2d8b821fafb7ffb23cb64c0b.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 219, + 648, + 480, + 662 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 219, + 648, + 480, + 662 + ], + "spans": [ + { + "bbox": [ + 219, + 648, + 480, + 662 + ], + "type": "interline_equation", + "content": "a _ {d} ^ {1}, a _ {d} ^ {2}, \\dots , a _ {d} ^ {2 M} = \\operatorname {s o f t m a x} \\left(s _ {d} ^ {1}, s _ {d} ^ {2}, \\dots , s _ {d} ^ {2 M}\\right), \\tag {4}", + "image_path": "c264ef96831d9a7c364312f39453c7e5bdd0288d7a9c2a57ab79e20fdbd4ec51.jpg" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 190, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 190, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 190, + 91, + 447, + 102 + ], + "type": "text", + "content": "Visual Prompt Selection for In-Context Learning Segmentation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "inline_equation", + "content": "a_{d}^{m}" + }, + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "content": " represents the probability of " + }, + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "content": "-th example in the candidate pool. To guide the search agent in learning select examples, we apply the IoU scores as rewards for selecting well-matched demonstrations from the candidate pool. Specifically, following the format of the ICL-based segmentation model, we re-organize query " + }, + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "inline_equation", + "content": "I_{q}" + }, + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "content": ", the original image of context examples " + }, + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "inline_equation", + "content": "I_{d}" + }, + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "content": " and corresponding mask " + }, + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "inline_equation", + "content": "G_{d}" + }, + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "content": " into a joint image, as shown in Fig. 4. These images are then fed into the ICL-based segmentation model to obtain the IoU scores " + }, + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "inline_equation", + "content": "U = \\{u_{d}\\}_{d=1}^{2M}" + }, + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "content": ". Inspired by [4,38], we use the average IoU score " + }, + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "inline_equation", + "content": "u_{avg}" + }, + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "content": " as the base score. By applying reinforcement learning, the gradient is calculated by:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 211, + 234, + 481, + 266 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 211, + 234, + 481, + 266 + ], + "spans": [ + { + "bbox": [ + 211, + 234, + 481, + 266 + ], + "type": "interline_equation", + "content": "\\nabla_ {\\theta} L (\\theta) = - \\frac {1}{2 M} \\sum_ {m = 1} ^ {2 M} \\left(u _ {d} ^ {m} - u _ {a v g}\\right) \\nabla_ {\\theta} \\log \\left(a _ {d} ^ {m}\\right), \\tag {5}", + "image_path": "2ac94dddeffd789eecaa46ef269fe0997cfe923ead1b7a46d2ec3d3e42ecc214.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 275, + 482, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 275, + 482, + 323 + ], + "spans": [ + { + "bbox": [ + 130, + 275, + 482, + 323 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 275, + 482, + 323 + ], + "type": "inline_equation", + "content": "a_{d}^{m}" + }, + { + "bbox": [ + 130, + 275, + 482, + 323 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 275, + 482, + 323 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 130, + 275, + 482, + 323 + ], + "type": "text", + "content": " are the probability of " + }, + { + "bbox": [ + 130, + 275, + 482, + 323 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 130, + 275, + 482, + 323 + ], + "type": "text", + "content": "-th example and the parameters of the search agent. Based on the above computation, this gradient would tend to increase the probability of the " + }, + { + "bbox": [ + 130, + 275, + 482, + 323 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 130, + 275, + 482, + 323 + ], + "type": "text", + "content": "-th context when its IoU scores " + }, + { + "bbox": [ + 130, + 275, + 482, + 323 + ], + "type": "inline_equation", + "content": "u_{d}^{m}" + }, + { + "bbox": [ + 130, + 275, + 482, + 323 + ], + "type": "text", + "content": " are higher than the average IoU score " + }, + { + "bbox": [ + 130, + 275, + 482, + 323 + ], + "type": "inline_equation", + "content": "u_{avg}" + }, + { + "bbox": [ + 130, + 275, + 482, + 323 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 323, + 481, + 360 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 323, + 481, + 360 + ], + "spans": [ + { + "bbox": [ + 130, + 323, + 481, + 360 + ], + "type": "text", + "content": "During inference, we directly input the test samples and candidate pool examples into the search agent. The contextual examples are then determined by sorting them based on probability scores " + }, + { + "bbox": [ + 130, + 323, + 481, + 360 + ], + "type": "inline_equation", + "content": "a_{d}" + }, + { + "bbox": [ + 130, + 323, + 481, + 360 + ], + "type": "text", + "content": " as detailed in Eq. 4." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 380, + 224, + 394 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 380, + 224, + 394 + ], + "spans": [ + { + "bbox": [ + 132, + 380, + 224, + 394 + ], + "type": "text", + "content": "5 Experiment" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 407, + 481, + 455 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 407, + 481, + 455 + ], + "spans": [ + { + "bbox": [ + 130, + 407, + 481, + 455 + ], + "type": "text", + "content": "We use different prompt selection strategies and various generalist segmentation frameworks to demonstrate the superiority of our method. Furthermore, we compare our approach with traditional methods and conduct extensive qualitative and quantitative studies on multiple datasets." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 475, + 269, + 488 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 475, + 269, + 488 + ], + "spans": [ + { + "bbox": [ + 132, + 475, + 269, + 488 + ], + "type": "text", + "content": "5.1 Experimental Setting" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 497, + 482, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 497, + 482, + 616 + ], + "spans": [ + { + "bbox": [ + 130, + 497, + 482, + 616 + ], + "type": "text", + "content": "Dataset. We follow the previous works [6, 45, 56] and evaluate our approach on PASCAL- " + }, + { + "bbox": [ + 130, + 497, + 482, + 616 + ], + "type": "inline_equation", + "content": "5^{i}" + }, + { + "bbox": [ + 130, + 497, + 482, + 616 + ], + "type": "text", + "content": " [32], COCO- " + }, + { + "bbox": [ + 130, + 497, + 482, + 616 + ], + "type": "inline_equation", + "content": "20^{i}" + }, + { + "bbox": [ + 130, + 497, + 482, + 616 + ], + "type": "text", + "content": " [21] and iSALD- " + }, + { + "bbox": [ + 130, + 497, + 482, + 616 + ], + "type": "inline_equation", + "content": "5^{i}" + }, + { + "bbox": [ + 130, + 497, + 482, + 616 + ], + "type": "text", + "content": " [50]. PASCAL- " + }, + { + "bbox": [ + 130, + 497, + 482, + 616 + ], + "type": "inline_equation", + "content": "5^{i}" + }, + { + "bbox": [ + 130, + 497, + 482, + 616 + ], + "type": "text", + "content": " includes 20 semantic categories and is created from PASCAL VOC 2012 [9] with external annotations from SDS [10]. COCO- " + }, + { + "bbox": [ + 130, + 497, + 482, + 616 + ], + "type": "inline_equation", + "content": "20^{i}" + }, + { + "bbox": [ + 130, + 497, + 482, + 616 + ], + "type": "text", + "content": " is built upon MSCOCO [21], which is a more challenging dataset and contains 80 categories. For all categories, both COCO- " + }, + { + "bbox": [ + 130, + 497, + 482, + 616 + ], + "type": "inline_equation", + "content": "20^{i}" + }, + { + "bbox": [ + 130, + 497, + 482, + 616 + ], + "type": "text", + "content": " and PASCAL- " + }, + { + "bbox": [ + 130, + 497, + 482, + 616 + ], + "type": "inline_equation", + "content": "5^{i}" + }, + { + "bbox": [ + 130, + 497, + 482, + 616 + ], + "type": "text", + "content": " are evenly divided into 4 folds for cross-validation, where each fold is selected as the test set and the remaining three folds are used for training. The iSALD- " + }, + { + "bbox": [ + 130, + 497, + 482, + 616 + ], + "type": "inline_equation", + "content": "5^{i}" + }, + { + "bbox": [ + 130, + 497, + 482, + 616 + ], + "type": "text", + "content": " contains 15 geospatial categories which derived from iSALD [48] in the remote sensing image semantic segmentation task. Similarly, the iSALD- " + }, + { + "bbox": [ + 130, + 497, + 482, + 616 + ], + "type": "inline_equation", + "content": "5^{i}" + }, + { + "bbox": [ + 130, + 497, + 482, + 616 + ], + "type": "text", + "content": " is also partitioned into 3 folds for cross-validation." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "Evaluation Metric. Following [13], mean intersection over union (mIoU) and foreground-background IoU (FBIoU) are adopted as evaluation metrics. The former measures the average overall IoU values for all object classes, while the latter reports the average IoU scores for the foreground and background." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 220, + 100 + ], + "type": "text", + "content": "W. Suo et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 149, + 168, + 465, + 304 + ], + "blocks": [ + { + "bbox": [ + 130, + 114, + 482, + 160 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 482, + 160 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 482, + 160 + ], + "type": "text", + "content": "Table 1: The results on COCO- " + }, + { + "bbox": [ + 130, + 114, + 482, + 160 + ], + "type": "inline_equation", + "content": "20^{i}" + }, + { + "bbox": [ + 130, + 114, + 482, + 160 + ], + "type": "text", + "content": " and PASCAL- " + }, + { + "bbox": [ + 130, + 114, + 482, + 160 + ], + "type": "inline_equation", + "content": "5^{i}" + }, + { + "bbox": [ + 130, + 114, + 482, + 160 + ], + "type": "text", + "content": " based on example-based semantic segmentation. To ensure a fair comparison and follow the few-shot setting, " + }, + { + "bbox": [ + 130, + 114, + 482, + 160 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 130, + 114, + 482, + 160 + ], + "type": "text", + "content": " indicates that only using the pre-trained weights provided by the SegGPT [3], and we report related results based on randomly sampling (i.e., w/R) and similarity sorting [56]." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 149, + 168, + 465, + 304 + ], + "lines": [ + { + "bbox": [ + 149, + 168, + 465, + 304 + ], + "spans": [ + { + "bbox": [ + 149, + 168, + 465, + 304 + ], + "type": "table", + "html": "
MethodCOCO-20i [21]PASCAL-5i [32]
1-shot5-shot1-shot5-shot
mIoUFBIoUmIoUFBIoUmIoUFBIoUmIoUFBIoU
Painter [44]32.850.232.650.164.574.264.674.4
MAE-VQGAN [6]12.145.7--27.555.6--
MAE-VQGAN w/ [56]18.046.5--33.657.8--
SegGPT† w/R [45]48.671.352.773.172.183.176.280.3
SegGPT† w/ [56]53.372.256.175.574.484.476.985.9
MAE-VQGAN w/Ours19.147.0--35.061.2--
SegGPT† w/Ours57.275.159.576.876.685.478.286.5
", + "image_path": "5897871e0b3fbb511655dc2905a37712928ce00ea5ef58c5cc25368e67883f60.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 339, + 482, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 339, + 482, + 426 + ], + "spans": [ + { + "bbox": [ + 130, + 339, + 482, + 426 + ], + "type": "text", + "content": "Implementation details. For both datasets, the Adam [17] is used as our optimizer and the maximum number of training epochs is 10. We use the CLIPViT [27] image encoder to extract image features. During training, all images are resized to " + }, + { + "bbox": [ + 130, + 339, + 482, + 426 + ], + "type": "inline_equation", + "content": "224 \\times 224" + }, + { + "bbox": [ + 130, + 339, + 482, + 426 + ], + "type": "text", + "content": " and fed into the image encoder. The dimension of extracted feature vectors is 1024. We train on one 4080 GPU with an initial learning rate of 1e-3, which is dropped by half every 5 epochs. The number of clusters " + }, + { + "bbox": [ + 130, + 339, + 482, + 426 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 130, + 339, + 482, + 426 + ], + "type": "text", + "content": " and batch size is set to 10 and 8." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 453, + 282, + 466 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 453, + 282, + 466 + ], + "spans": [ + { + "bbox": [ + 132, + 453, + 282, + 466 + ], + "type": "text", + "content": "5.2 Quantitative Evaluation" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "type": "text", + "content": "In Table 1, we show a performance comparison of two general ICL-based segmentation models with and without proposed SCS. Following [37,56], all results are reported by repeating the random experiment five times. It can be observed that our method can significantly boost performance in all settings including 1-shot and 5-shot settings. In particular, for the 1-shot setting, the SCS leads to MAE-VQGAN with the mIoU gain of 7.0 and 6.5 on the PASCAL- " + }, + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "type": "inline_equation", + "content": "5^{i}" + }, + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "type": "text", + "content": " and COCO- " + }, + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "type": "inline_equation", + "content": "20^{i}" + }, + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "type": "text", + "content": " dataset, respectively. Besides, compared to previous rule-base models [56] in 1-shot and 5-shot semantic segmentation tasks, our method surpasses it by a considerable margin. Moreover, following the few-shot setting and ensuring a fair comparison, we report related results based on the pre-trained weights for SegGPT [3]. It can be observed that our method significantly outperforms these rule-based visual prompt selection methods for SegGPT by 9.7 and 6.8 on the 1-shot and 5-shot settings for COCO- " + }, + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "type": "inline_equation", + "content": "20^{i}" + }, + { + "bbox": [ + 130, + 486, + 482, + 666 + ], + "type": "text", + "content": ". More importantly, the above results demonstrate that the proposed SCS can be considered as a \"plug-and-play\" module to enhance the segmentation ability of ICL-based frameworks." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 190, + 91, + 447, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 190, + 91, + 447, + 103 + ], + "spans": [ + { + "bbox": [ + 190, + 91, + 447, + 103 + ], + "type": "text", + "content": "Visual Prompt Selection for In-Context Learning Segmentation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 479, + 101 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 479, + 101 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 479, + 101 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 184, + 156, + 430, + 227 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 482, + 137 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 482, + 137 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 482, + 137 + ], + "type": "text", + "content": "Table 2: Ablation study. Similarity: Cosine similarity ranking; Cluster: Cluster compacting; Sorting: The nearest-farthest sampling; ASM: Adaptive search module." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 184, + 156, + 430, + 227 + ], + "lines": [ + { + "bbox": [ + 184, + 156, + 430, + 227 + ], + "spans": [ + { + "bbox": [ + 184, + 156, + 430, + 227 + ], + "type": "table", + "html": "
SimilarityClusterSortingASMmIoUFBIoU
152.773.1
2✓56.174.5
3✓✓55.774.3
4✓✓✓57.675.7
5✓✓✓59.576.8
", + "image_path": "e6f5507ba1dc6511b390a4f1fdfc97f2919c4427371d8d6ca22c99b46b5c69b3.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 217, + 267, + 392, + 371 + ], + "blocks": [ + { + "bbox": [ + 189, + 247, + 424, + 258 + ], + "lines": [ + { + "bbox": [ + 189, + 247, + 424, + 258 + ], + "spans": [ + { + "bbox": [ + 189, + 247, + 424, + 258 + ], + "type": "text", + "content": "Table 3: The effects of different settings for our method." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 217, + 267, + 392, + 371 + ], + "lines": [ + { + "bbox": [ + 217, + 267, + 392, + 371 + ], + "spans": [ + { + "bbox": [ + 217, + 267, + 392, + 371 + ], + "type": "table", + "html": "
Model1-shot5-shot
mIoUFBIoUmIoUFBIoU
1Entropy53.172.356.175.0
2Foreground53.572.556.779.3
3ViT56.975.058.776.0
4CLIP-resnet56.374.758.576.4
5Density-sort57.675.159.176.3
6Entropy-sort56.974.558.475.9
7Ours57.275.159.576.8
", + "image_path": "3ec1bc3870c6af921dceef7628993c1884acb894acb623c7d6be9b49e105963f.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 394, + 238, + 406 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 394, + 238, + 406 + ], + "spans": [ + { + "bbox": [ + 132, + 394, + 238, + 406 + ], + "type": "text", + "content": "5.3 Ablation Study" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 414, + 481, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 414, + 481, + 449 + ], + "spans": [ + { + "bbox": [ + 130, + 414, + 481, + 449 + ], + "type": "text", + "content": "To clarify the effectiveness of our method, we conduct further ablation studies on the COCO- " + }, + { + "bbox": [ + 130, + 414, + 481, + 449 + ], + "type": "inline_equation", + "content": "20^{i}" + }, + { + "bbox": [ + 130, + 414, + 481, + 449 + ], + "type": "text", + "content": " dataset with the 5-shot setting. The SegGPT is used as the basic model due to its outstanding segmentation performance." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 450, + 481, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 450, + 481, + 545 + ], + "spans": [ + { + "bbox": [ + 130, + 450, + 481, + 545 + ], + "type": "text", + "content": "Effects of Cluster. In Table 2, we first build a baseline by randomly selecting examples from the entire training set. The average result is reported across 5 different seeds. In the second row of Table 2, we follow [56] and utilize cosine similarity ranking to select the top-5 examples as contexts. In the third row, we investigate the impact of the cluster on in-context example selection, using the closest examples to each cluster centroid to construct the candidate pool. Compared to the baseline, the results show that compacting the search space not only reduces annotation cost but also improves model performance." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 545, + 481, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 545, + 481, + 629 + ], + "spans": [ + { + "bbox": [ + 130, + 545, + 481, + 629 + ], + "type": "text", + "content": "Effects of Sort. As shown in row 4 of Table 2, we experimentally demonstrate that integrating our sorting mechanism results in " + }, + { + "bbox": [ + 130, + 545, + 481, + 629 + ], + "type": "inline_equation", + "content": "1.9\\mathrm{mIoU}" + }, + { + "bbox": [ + 130, + 545, + 481, + 629 + ], + "type": "text", + "content": " improvement compared to directly using the closest examples to each cluster centroid. The inherent reason is that centroid examples describe a limited feature space [33, 46], making it challenging to support a variety of different contexts for test instances. In contrast, our sorting mechanism introduces the nearest and farthest examples, which constructs more diverse demonstrations for ICL learning." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "type": "text", + "content": "Effects of Adaptive Search Module. Finally, we further explore the effectiveness of our adaptive search module in the last row of Table 2. Compared to using a simple similarity sorting method, our approach shows an improve" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 221, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 221, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 221, + 100 + ], + "type": "text", + "content": "W. Suo et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 143, + 146, + 291, + 196 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 471, + 137 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 471, + 137 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 471, + 137 + ], + "type": "text", + "content": "Table 4: The results on PASCAL-5" + }, + { + "bbox": [ + 132, + 114, + 471, + 137 + ], + "type": "inline_equation", + "content": "^i" + }, + { + "bbox": [ + 132, + 114, + 471, + 137 + ], + "type": "text", + "content": " [32] Table 5: The results on iSALD-5" + }, + { + "bbox": [ + 132, + 114, + 471, + 137 + ], + "type": "inline_equation", + "content": "^i" + }, + { + "bbox": [ + 132, + 114, + 471, + 137 + ], + "type": "text", + "content": " [48] based on instance segmentation. based on semantic segmentation." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 143, + 146, + 291, + 196 + ], + "lines": [ + { + "bbox": [ + 143, + 146, + 291, + 196 + ], + "spans": [ + { + "bbox": [ + 143, + 146, + 291, + 196 + ], + "type": "table", + "html": "
Model1-shot5-shot
SegGPT† w/R [45]46.546.9
SegGPT† w/ [56]47.148.5
SegGPT w/Ours50.551.4
", + "image_path": "0335616d3ba896a9e0ae905786a2e2d83370134294f1d86fabcd7226f7082d11.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 312, + 146, + 460, + 195 + ], + "blocks": [ + { + "bbox": [ + 312, + 146, + 460, + 195 + ], + "lines": [ + { + "bbox": [ + 312, + 146, + 460, + 195 + ], + "spans": [ + { + "bbox": [ + 312, + 146, + 460, + 195 + ], + "type": "table", + "html": "
Model1-shot5-shot
SegGPT† w/R [45]29.830.6
SegGPT† w/ [56]34.735.4
SegGPT w/Ours37.939.5
", + "image_path": "3f0290a7ef55b98f56d6925d6144bd07b475c5a67db29807646c38083fe20cf8.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 217, + 482, + 253 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 217, + 482, + 253 + ], + "spans": [ + { + "bbox": [ + 130, + 217, + 482, + 253 + ], + "type": "text", + "content": "ment of 3.4 points. Meanwhile, benefiting from the diverse candidate pool, our SCS achieves state-of-the-art performance and outperforms the baseline by a considerable margin." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 270, + 285, + 284 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 270, + 285, + 284 + ], + "spans": [ + { + "bbox": [ + 132, + 270, + 285, + 284 + ], + "type": "text", + "content": "5.4 Different Model Settings" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 291, + 482, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 291, + 482, + 544 + ], + "spans": [ + { + "bbox": [ + 130, + 291, + 482, + 544 + ], + "type": "text", + "content": "In this section, we explore several alternative model settings to further discuss the proposed method. All experiments are conducted on the COCO- " + }, + { + "bbox": [ + 130, + 291, + 482, + 544 + ], + "type": "inline_equation", + "content": "20^{i}" + }, + { + "bbox": [ + 130, + 291, + 482, + 544 + ], + "type": "text", + "content": " with the SegGPT model. In the first two rows, we first compare the performance using different metrics. Specifically, we use the rule-based methods to select the contextual examples from the complete training set, which including Entropy sorting [11] and Foreground sorting [53]. In particular, \"Entropy\" is executed by calculating the maximum entropy values. The \"Foreground\" refers to a two-stage approach [53]. It uses an object detector [42] to extract the foreground boxes and the foreground similarity is used as the metric. From the 1-shot and 5-shot results, it can be observed that our SCS can dramatically outperform these methods by a large margin. In rows 3-4, we construct the experiments with different feature encoders. We replace the CLIP-ViT with ViT [8] and CLIP-resnet [27] to obtain image feature vectors. These extracted features would be utilized to construct the candidate pool and perform the adaptive searching. The results show that our SCS is robust for different image encoders. Finally, in rows 5-6 of Table 3, we replace our selection mechanism with several alternative manners. However, these approaches mostly involve complex computation. For example, the Density-sort [2] requires additional computations on the order of " + }, + { + "bbox": [ + 130, + 291, + 482, + 544 + ], + "type": "inline_equation", + "content": "O(MN)" + }, + { + "bbox": [ + 130, + 291, + 482, + 544 + ], + "type": "text", + "content": " in each iteration to update density lists, where " + }, + { + "bbox": [ + 130, + 291, + 482, + 544 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 130, + 291, + 482, + 544 + ], + "type": "text", + "content": " is the number of clusters and " + }, + { + "bbox": [ + 130, + 291, + 482, + 544 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 291, + 482, + 544 + ], + "type": "text", + "content": " is the size of the unlabeled data. Instead, our method uses the basic clustering algorithm and obtains a comparable performance." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 561, + 256, + 573 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 561, + 256, + 573 + ], + "spans": [ + { + "bbox": [ + 132, + 561, + 256, + 573 + ], + "type": "text", + "content": "5.5 Further Discussion" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "content": "More Segmentation tasks. In this section, we evaluate the effect of our method on more segmentation tasks and datasets. Because MAE-VQGAN cannot distinguish between individual instances of objects within the same class, we construct more experiments with the SegGPT model. Specifically, we report instance segmentation results on PASCAL-5i in Table 4 and aerial image semantic segmentation results on iSALD-5i in Table 5. The above results indicate that the SCS can enhance the segmentation ability for various real-world scenarios." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 189, + 90, + 447, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 189, + 90, + 447, + 103 + ], + "spans": [ + { + "bbox": [ + 189, + 90, + 447, + 103 + ], + "type": "text", + "content": "Visual Prompt Selection for In-Context Learning Segmentation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 101 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 101 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 101 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 162, + 114, + 452, + 323 + ], + "blocks": [ + { + "bbox": [ + 162, + 114, + 452, + 323 + ], + "lines": [ + { + "bbox": [ + 162, + 114, + 452, + 323 + ], + "spans": [ + { + "bbox": [ + 162, + 114, + 452, + 323 + ], + "type": "image", + "image_path": "07ebfd07fd83de029507aa2c73371a7b1fef1abe48c42da646e31893c0e33c0e.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 331, + 482, + 366 + ], + "lines": [ + { + "bbox": [ + 130, + 331, + 482, + 366 + ], + "spans": [ + { + "bbox": [ + 130, + 331, + 482, + 366 + ], + "type": "text", + "content": "Fig.5: Qualitative results under 1-shot setting on COCO- " + }, + { + "bbox": [ + 130, + 331, + 482, + 366 + ], + "type": "inline_equation", + "content": "20^{i}" + }, + { + "bbox": [ + 130, + 331, + 482, + 366 + ], + "type": "text", + "content": ". The similarity-based selection method is viewed as our baseline. The green regions, red regions and purple regions are example masks, predicted masks and ground-truth masks." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 386, + 482, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 386, + 482, + 542 + ], + "spans": [ + { + "bbox": [ + 130, + 386, + 482, + 542 + ], + "type": "text", + "content": "Qualitative Results. As shown in Fig. 5, to further analyze and verify the proposed method, we visualized the qualitative segmentation results on COCO" + }, + { + "bbox": [ + 130, + 386, + 482, + 542 + ], + "type": "inline_equation", + "content": "_{20^i}" + }, + { + "bbox": [ + 130, + 386, + 482, + 542 + ], + "type": "text", + "content": " with the SegGPT. The first and third columns represent the visual prompts (highlighted in green) based on similarity sorting and our SCS method, respectively. The second and fourth columns show the corresponding segmentation results (highlighted in red). The last column represents the ground truth mask (highlighted in purple). As illustrated in the first row, the baseline model selects the most similar context to the query (i.e., chopping board, knife and carrot), while the segment result is not satisfactory, as the model can not recognize the knife in the image. Instead, Our method selects an image from the candidate pool that accurately segments out the mask. The above results demonstrate that similarity sorting is not the only golden standard and improving the diversity of context prompts can boost the segmentation results." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 559, + 214, + 571 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 559, + 214, + 571 + ], + "spans": [ + { + "bbox": [ + 132, + 559, + 214, + 571 + ], + "type": "text", + "content": "6 Conclusion" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "content": "In this paper, we focus on the example selection method for the ICL-based segmentation frameworks. By building comprehensive comparisons, we empirically demonstrate that diversity of contexts is a key factor in guiding segmentation and boosting performance. Beyond the above insights, we propose a new stepwise context search method that improves the diversity of the candidate pool and significantly alleviates the annotation cost. We hope that this work can promote further research to better understand in-context learning in computer vision." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 221, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 221, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 221, + 100 + ], + "type": "text", + "content": "W. Suo et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 163 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 163 + ], + "type": "text", + "content": "Acknowledgement. This work was supported by National Natural Science Foundation of China (No.62102323), the Innovation Capability Support Program of Shaanxi(Program No. 2023KJXX-142) and National Natural Science Foundation of China (U23B2013)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 184, + 199, + 197 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 184, + 199, + 197 + ], + "spans": [ + { + "bbox": [ + 132, + 184, + 199, + 197 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 210, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 138, + 210, + 480, + 233 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 210, + 480, + 233 + ], + "spans": [ + { + "bbox": [ + 138, + 210, + 480, + 233 + ], + "type": "text", + "content": "1. Agrawal, S., Zhou, C., Lewis, M., Zettlemoyer, L., Ghazvininejad, M.: In-context examples selection for machine translation. arXiv preprint arXiv:2212.02437 (2022)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 234, + 480, + 255 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 234, + 480, + 255 + ], + "spans": [ + { + "bbox": [ + 138, + 234, + 480, + 255 + ], + "type": "text", + "content": "2. et al., W.: \"unsupervised selective labeling for more effective semi-supervised learning\". In: ECCV 2022. pp. 427-445. Springer (2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 256, + 459, + 266 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 256, + 459, + 266 + ], + "spans": [ + { + "bbox": [ + 138, + 256, + 459, + 266 + ], + "type": "text", + "content": "3. et al, W.: \"seggspt: Segmenting everything in context\". In ICCV 2023 (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 267, + 480, + 311 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 267, + 480, + 311 + ], + "spans": [ + { + "bbox": [ + 138, + 267, + 480, + 311 + ], + "type": "text", + "content": "4. Anderson, P., He, X., Buehler, C., Teney, D., Johnson, M., Gould, S., Zhang, L.: Bottom-up and top-down attention for image captioning and visual question answering. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 6077-6086 (2018)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 312, + 480, + 344 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 312, + 480, + 344 + ], + "spans": [ + { + "bbox": [ + 138, + 312, + 480, + 344 + ], + "type": "text", + "content": "5. Badrinarayanan, V., Kendall, A., Cipolla, R.: Segnet: A deep convolutional encoder-decoder architecture for image segmentation. IEEE transactions on pattern analysis and machine intelligence 39(12), 2481-2495 (2017)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 345, + 480, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 345, + 480, + 376 + ], + "spans": [ + { + "bbox": [ + 138, + 345, + 480, + 376 + ], + "type": "text", + "content": "6. Bar, A., Gandelsman, Y., Darrell, T., Globerson, A., Efros, A.: Visual prompting via image inpainting. Advances in Neural Information Processing Systems 35, 25005-25017 (2022)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 378, + 480, + 410 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 378, + 480, + 410 + ], + "spans": [ + { + "bbox": [ + 138, + 378, + 480, + 410 + ], + "type": "text", + "content": "7. Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J.D., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., et al.: Language models are few-shot learners. Advances in neural information processing systems 33, 1877-1901 (2020)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 411, + 480, + 454 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 411, + 480, + 454 + ], + "spans": [ + { + "bbox": [ + 138, + 411, + 480, + 454 + ], + "type": "text", + "content": "8. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929 (2020)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 456, + 480, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 456, + 480, + 487 + ], + "spans": [ + { + "bbox": [ + 138, + 456, + 480, + 487 + ], + "type": "text", + "content": "9. Everingham, M., Van Gool, L., Williams, C.K., Winn, J., Zisserman, A.: The Pascal visual object classes (voc) challenge. International journal of computer vision 88, 303-338 (2010)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 488, + 480, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 488, + 480, + 520 + ], + "spans": [ + { + "bbox": [ + 138, + 488, + 480, + 520 + ], + "type": "text", + "content": "0. Hariharan, B., Arbeláez, P., Bourdev, L., Maji, S., Malik, J.: Semantic contours from inverse detectors. In: 2011 international conference on computer vision. pp. 991-998. IEEE (2011)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 521, + 480, + 565 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 521, + 480, + 565 + ], + "spans": [ + { + "bbox": [ + 138, + 521, + 480, + 565 + ], + "type": "text", + "content": "1. Haussmann, E., Fenzi, M., Chitta, K., Ivanecky, J., Xu, H., Roy, D., Mittel, A., Koumchatzky, N., Farabet, C., Alvarez, J.M.: Scalable active learning for object detection. In: 2020 IEEE intelligent vehicles symposium (iv). pp. 1430-1435. IEEE (2020)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 566, + 480, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 566, + 480, + 588 + ], + "spans": [ + { + "bbox": [ + 138, + 566, + 480, + 588 + ], + "type": "text", + "content": "2. He, K., Gkioxari, G., Dollár, P., Girshick, R.: Mask r-cnn. In: Proceedings of the IEEE international conference on computer vision. pp. 2961-2969 (2017)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 589, + 480, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 589, + 480, + 620 + ], + "spans": [ + { + "bbox": [ + 138, + 589, + 480, + 620 + ], + "type": "text", + "content": "3. Hong, S., Cho, S., Nam, J., Lin, S., Kim, S.: Cost aggregation with 4d convolutional swim transformer for few-shot segmentation. In: European Conference on Computer Vision. pp. 108-126. Springer (2022)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 621, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 621, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 621, + 480, + 665 + ], + "type": "text", + "content": "4. Hongjin, S., Kasai, J., Wu, C.H., Shi, W., Wang, T., Xin, J., Zhang, R., Ostendorf, M., Zettlemoyer, L., Smith, N.A., et al.: Selective annotation makes language models better few-shot learners. In: The Eleventh International Conference on Learning Representations (2022)" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 190, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 190, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 190, + 91, + 447, + 102 + ], + "type": "text", + "content": "Visual Prompt Selection for In-Context Learning Segmentation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 117, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 133, + 117, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 117, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 133, + 117, + 480, + 138 + ], + "type": "text", + "content": "15. Hore, A., Ziou, D.: Image quality metrics: Psnr vs. sSIM. In: 2010 20th international conference on pattern recognition. pp. 2366-2369. IEEE (2010)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 138, + 480, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 138, + 480, + 171 + ], + "spans": [ + { + "bbox": [ + 133, + 138, + 480, + 171 + ], + "type": "text", + "content": "16. Hu, H., Li, Z., Li, L., Yang, H., Zhu, H.: Classification of very high-resolution remote sensing imagery using a fully convolutional network with global and local context information enhancements. IEEE Access 8, 14606-14619 (2020)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 171, + 480, + 191 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 171, + 480, + 191 + ], + "spans": [ + { + "bbox": [ + 133, + 171, + 480, + 191 + ], + "type": "text", + "content": "17. Kingma, D.P., Ba, J.: Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980 (2014)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 191, + 480, + 224 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 191, + 480, + 224 + ], + "spans": [ + { + "bbox": [ + 133, + 191, + 480, + 224 + ], + "type": "text", + "content": "18. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., et al.: Segment anything. arXiv preprint arXiv:2304.02643 (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 224, + 480, + 255 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 224, + 480, + 255 + ], + "spans": [ + { + "bbox": [ + 132, + 224, + 480, + 255 + ], + "type": "text", + "content": "19. Li, X., Lv, K., Yan, H., Lin, T., Zhu, W., Ni, Y., Xie, G., Wang, X., Qiu, X.: Unified demonstration retriever for in-context learning. arXiv preprint arXiv:2305.04320 (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 255, + 480, + 310 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 255, + 480, + 310 + ], + "spans": [ + { + "bbox": [ + 132, + 255, + 480, + 310 + ], + "type": "text", + "content": "20. Li, Y., Padmanabhan, A., Zhao, P., Wang, Y., Xu, G.H., Netravali, R.: Reducto: On-camera filtering for resource-efficient real-time video analytics. In: Proceedings of the Annual conference of the ACM Special Interest Group on Data Communication on the applications, technologies, architectures, and protocols for computer communication. pp. 359-376 (2020)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 310, + 480, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 310, + 480, + 354 + ], + "spans": [ + { + "bbox": [ + 132, + 310, + 480, + 354 + ], + "type": "text", + "content": "21. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollar, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13. pp. 740-755. Springer (2014)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 354, + 480, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 354, + 480, + 374 + ], + "spans": [ + { + "bbox": [ + 132, + 354, + 480, + 374 + ], + "type": "text", + "content": "22. Liu, J., Shen, D., Zhang, Y., Dolan, B., Carin, L., Chen, W.: What makes good in-context examples for gpt-3? arXiv preprint arXiv:2101.06804 (2021)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 374, + 480, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 374, + 480, + 407 + ], + "spans": [ + { + "bbox": [ + 132, + 374, + 480, + 407 + ], + "type": "text", + "content": "23. Long, J., Shelhamer, E., Darrell, T.: Fully convolutional networks for semantic segmentation. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 3431-3440 (2015)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 407, + 480, + 450 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 407, + 480, + 450 + ], + "spans": [ + { + "bbox": [ + 132, + 407, + 480, + 450 + ], + "type": "text", + "content": "24. Luo, G., Zhou, Y., Sun, X., Cao, L., Wu, C., Deng, C., Ji, R.: Multi-task collaborative network for joint referring expression comprehension and segmentation. In: Proceedings of the IEEE/CVF Conference on computer vision and pattern recognition. pp. 10034-10043 (2020)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 450, + 480, + 492 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 450, + 480, + 492 + ], + "spans": [ + { + "bbox": [ + 132, + 450, + 480, + 492 + ], + "type": "text", + "content": "25. Lyu, X., Fang, Y., Tong, B., Li, X., Zeng, T.: Multiscale normalization attention network for water body extraction from remote sensing imagery. Remote Sensing 14(19) (2022). https://doi.org/10.3390/rs14194983, https://www.mdpi.com/2072-4292/14/19/4983" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 492, + 480, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 492, + 480, + 525 + ], + "spans": [ + { + "bbox": [ + 132, + 492, + 480, + 525 + ], + "type": "text", + "content": "26. Min, S., Lyu, X., Holtzman, A., Artetxe, M., Lewis, M., Hajishirzi, H., Zettlemoyer, L.: Rethinking the role of demonstrations: What makes in-context learning work? arXiv preprint arXiv:2202.12837 (2022)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 525, + 480, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 525, + 480, + 568 + ], + "spans": [ + { + "bbox": [ + 132, + 525, + 480, + 568 + ], + "type": "text", + "content": "27. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 568, + 480, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 568, + 480, + 601 + ], + "spans": [ + { + "bbox": [ + 132, + 568, + 480, + 601 + ], + "type": "text", + "content": "28. Reisenhofer, R., Bosse, S., Kutyniok, G., Wiegand, T.: Aaar wavelet-based perceptual similarity index for image quality assessment. Signal Processing: Image Communication 61, 33-43 (2018)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 601, + 480, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 601, + 480, + 644 + ], + "spans": [ + { + "bbox": [ + 132, + 601, + 480, + 644 + ], + "type": "text", + "content": "29. Ronneberger, O., Fischer, P., Brox, T.: U-net: Convolutional networks for biomedical image segmentation. In: Medical Image Computing and Computer-Assisted Intervention-MICCAI 2015: 18th International Conference, Munich, Germany, October 5-9, 2015, Proceedings, Part III 18. pp. 234-241. Springer (2015)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 644, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 644, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 644, + 480, + 665 + ], + "type": "text", + "content": "30. Rubin, O., Herzig, J., Berant, J.: Learning to retrieve prompts for in-context learning. arXiv preprint arXiv:2112.08633 (2021)" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 221, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 221, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 221, + 100 + ], + "type": "text", + "content": "W. Suo et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 665 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 160 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 160 + ], + "type": "text", + "content": "31. Rubin, O., Herzig, J., Berant, J.: Learning to retrieve prompts for in-context learning. In: Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies. pp. 2655-2671 (2022)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 161, + 481, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 161, + 481, + 182 + ], + "spans": [ + { + "bbox": [ + 130, + 161, + 481, + 182 + ], + "type": "text", + "content": "32. Shaban, A., Bansal, S., Liu, Z., Essa, I., Boots, B.: One-shot learning for semantic segmentation. arXiv preprint arXiv:1709.03410 (2017)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 183, + 481, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 183, + 481, + 215 + ], + "spans": [ + { + "bbox": [ + 132, + 183, + 481, + 215 + ], + "type": "text", + "content": "33. Shen, F., Zhao, L., Du, W., Zhong, W., Qian, F.: Large-scale industrial energy systems optimization under uncertainty: A data-driven robust optimization approach. Applied Energy 259, 114199 (2020)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 216, + 481, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 216, + 481, + 248 + ], + "spans": [ + { + "bbox": [ + 132, + 216, + 481, + 248 + ], + "type": "text", + "content": "34. Shi, H., Li, H., Meng, F., Wu, Q.: Key-word-aware network for referring expression image segmentation. In: Proceedings of the European Conference on Computer Vision (ECCV). pp. 38-54 (2018)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 249, + 481, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 249, + 481, + 281 + ], + "spans": [ + { + "bbox": [ + 132, + 249, + 481, + 281 + ], + "type": "text", + "content": "35. Sun, L., Yang, K., Hu, X., Hu, W., Wang, K.: Real-time fusion network forrgb-d semantic segmentation incorporating unexpected obstacle detection for road-driving images. IEEE robotics and automation letters 5(4), 5558-5565 (2020)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 281, + 481, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 281, + 481, + 314 + ], + "spans": [ + { + "bbox": [ + 132, + 281, + 481, + 314 + ], + "type": "text", + "content": "36. Sun, M., Suo, W., Wang, P., Niu, K., Liu, L., Lin, G., Zhang, Y., Wu, Q.: An adaptive correlation filtering method for text-based person search. International Journal of Computer Vision pp. 1-16 (2024)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 315, + 481, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 315, + 481, + 335 + ], + "spans": [ + { + "bbox": [ + 132, + 315, + 481, + 335 + ], + "type": "text", + "content": "37. Sun, Y., Chen, Q., Wang, J., Wang, J., Li, Z.: Exploring effective factors for improving visual in-context learning. arXiv preprint arXiv:2304.04748 (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 336, + 481, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 336, + 481, + 380 + ], + "spans": [ + { + "bbox": [ + 132, + 336, + 481, + 380 + ], + "type": "text", + "content": "38. Suo, W., Sun, M., Liu, W., Gao, Y., Wang, P., Zhang, Y., Wu, Q.: S3c: Semi-supervised vqa natural language explanation via self-critical learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2646-2656 (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 380, + 481, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 380, + 481, + 413 + ], + "spans": [ + { + "bbox": [ + 132, + 380, + 481, + 413 + ], + "type": "text", + "content": "39. Suo, W., Sun, M., Niu, K., Gao, Y., Wang, P., Zhang, Y., Wu, Q.: A simple and robust correlation filtering method for text-based person search. In: European conference on computer vision. pp. 726-742. Springer (2022)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 414, + 481, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 414, + 481, + 445 + ], + "spans": [ + { + "bbox": [ + 132, + 414, + 481, + 445 + ], + "type": "text", + "content": "40. Suo, W., Sun, M., Wang, P., Zhang, Y., Wu, Q.: Rethinking and improving feature pyramids for one-stage referring expression comprehension. IEEE Transactions on Image Processing 32, 854-864 (2022)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 446, + 481, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 446, + 481, + 479 + ], + "spans": [ + { + "bbox": [ + 132, + 446, + 481, + 479 + ], + "type": "text", + "content": "41. Vasudevan, A.B., Dai, D., Van Gool, L.: Semantic object prediction and spatial sound super-resolution with binaural sounds. In: European conference on computer vision. pp. 638-655. Springer (2020)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 479, + 481, + 522 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 479, + 481, + 522 + ], + "spans": [ + { + "bbox": [ + 132, + 479, + 481, + 522 + ], + "type": "text", + "content": "42. Wang, W., Dai, J., Chen, Z., Huang, Z., Li, Z., Zhu, X., Hu, X., Lu, T., Lu, L., Li, H., et al.: Intermimage: Exploring large-scale vision foundation models with deformable convolutions. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14408-14419 (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 523, + 481, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 523, + 481, + 544 + ], + "spans": [ + { + "bbox": [ + 132, + 523, + 481, + 544 + ], + "type": "text", + "content": "43. Wang, X., Ma, H., You, S.: Deep clustering for weakly-supervised semantic segmentation in autonomous driving scenes. Neurocomputing 381, 20-28 (2020)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 545, + 481, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 545, + 481, + 578 + ], + "spans": [ + { + "bbox": [ + 132, + 545, + 481, + 578 + ], + "type": "text", + "content": "44. Wang, X., Wang, W., Cao, Y., Shen, C., Huang, T.: Images speak in images: A generalist painter for in-context visual learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6830-6839 (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 578, + 481, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 578, + 481, + 599 + ], + "spans": [ + { + "bbox": [ + 132, + 578, + 481, + 599 + ], + "type": "text", + "content": "45. Wang, X., Zhang, X., Cao, Y., Wang, W., Shen, C., Huang, T.: Seggpt: Segmenting everything in context. arXiv preprint arXiv:2304.03284 (2023)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 600, + 481, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 600, + 481, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 600, + 481, + 632 + ], + "type": "text", + "content": "46. Wang, X., Lian, L., Yu, S.X.: Unsupervised selective labeling for more effective semi-supervised learning. In: European Conference on Computer Vision. pp. 427-445. Springer (2022)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 632, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 481, + 665 + ], + "type": "text", + "content": "47. Wang, Z., Bovik, A.C., Sheikh, H.R., Simoncelli, E.P.: Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing 13(4), 600-612 (2004)" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 190, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 190, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 190, + 91, + 447, + 102 + ], + "type": "text", + "content": "Visual Prompt Selection for In-Context Learning Segmentation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 468 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 481, + 161 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 481, + 161 + ], + "type": "text", + "content": "48. Waqas Zamir, S., Arora, A., Gupta, A., Khan, S., Sun, G., Shahbaz Khan, F., Zhu, F., Shao, L., Xia, G.S., Bai, X.: isaid: A large-scale dataset for instance segmentation in aerial images. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops. pp. 28-37 (2019)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 162, + 481, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 162, + 481, + 194 + ], + "spans": [ + { + "bbox": [ + 132, + 162, + 481, + 194 + ], + "type": "text", + "content": "49. Wei, J., Wang, X., Schuurmans, D., Bosma, M., Xia, F., Chi, E., Le, Q.V., Zhou, D., et al.: Chain-of-thought prompting elicits reasoning in large language models. Advances in Neural Information Processing Systems 35, 24824-24837 (2022)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 194, + 481, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 194, + 481, + 226 + ], + "spans": [ + { + "bbox": [ + 132, + 194, + 481, + 226 + ], + "type": "text", + "content": "50. Yao, X., Cao, Q., Feng, X., Cheng, G., Han, J.: Scale-aware detailed matching for few-shot aerial image semantic segmentation. IEEE Transactions on Geoscience and Remote Sensing 60, 1-11 (2021)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 227, + 481, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 227, + 481, + 248 + ], + "spans": [ + { + "bbox": [ + 132, + 227, + 481, + 248 + ], + "type": "text", + "content": "51. Ye, J., Wu, Z., Feng, J., Yu, T., Kong, L.: Compositional exemplars for in-context learning. arXiv preprint arXiv:2302.05698 (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 249, + 481, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 249, + 481, + 281 + ], + "spans": [ + { + "bbox": [ + 132, + 249, + 481, + 281 + ], + "type": "text", + "content": "52. Yu, Q., Gao, Y., Zheng, Y., Zhu, J., Dai, Y., Shi, Y.: Crossover-net: Leveraging vertical-horizontal crossover relation for robust medical image segmentation. Pattern Recognition 113, 107756 (2021)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 281, + 481, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 281, + 481, + 314 + ], + "spans": [ + { + "bbox": [ + 132, + 281, + 481, + 314 + ], + "type": "text", + "content": "53. Yu, Y., Yang, X., Li, Q., Zhou, Y., Da, F., Yan, J.: H2rbox-v2: Incorporating symmetry for boosting horizontal box supervised oriented object detection. In: Thirty-seventh Conference on Neural Information Processing Systems (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 315, + 481, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 315, + 481, + 346 + ], + "spans": [ + { + "bbox": [ + 132, + 315, + 481, + 346 + ], + "type": "text", + "content": "54. Zendel, O., Schörghuber, M., Rainer, B., Murschitz, M., Beleznai, C.: Unifying panoptic segmentation for autonomous driving. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 21351-21360 (2022)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 347, + 481, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 347, + 481, + 380 + ], + "spans": [ + { + "bbox": [ + 132, + 347, + 481, + 380 + ], + "type": "text", + "content": "55. Zhang, L., Zhang, L., Mou, X., Zhang, D.: Fsim: A feature similarity index for image quality assessment. IEEE transactions on Image Processing 20(8), 2378-2386 (2011)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 380, + 481, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 380, + 481, + 402 + ], + "spans": [ + { + "bbox": [ + 132, + 380, + 481, + 402 + ], + "type": "text", + "content": "56. Zhang, Y., Zhou, K., Liu, Z.: What makes good examples for visual in-context learning? arXiv preprint arXiv:2301.13670 (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 403, + 481, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 403, + 481, + 445 + ], + "spans": [ + { + "bbox": [ + 132, + 403, + 481, + 445 + ], + "type": "text", + "content": "57. Zhang, Y., Liu, H., Hu, Q.: Transfuse: Fusing transformers and cnns for medical image segmentation. In: Medical Image Computing and Computer Assisted Intervention-MICCAI 2021: 24th International Conference, Strasbourg, France, September 27-October 1, 2021, Proceedings, Part I 24. pp. 14-24. Springer (2021)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 447, + 481, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 447, + 481, + 468 + ], + "spans": [ + { + "bbox": [ + 132, + 447, + 481, + 468 + ], + "type": "text", + "content": "58. Zhang, Z., Zhang, A., Li, M., Smola, A.: Automatic chain of thought prompting in large language models. arXiv preprint arXiv:2210.03493 (2022)" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 221, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 221, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 221, + 100 + ], + "type": "text", + "content": "W. Suo et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Retrieval Robust to Object Motion Blur/191ce9d4-edfe-414c-a611-8012902a3458_content_list.json b/2024/Retrieval Robust to Object Motion Blur/191ce9d4-edfe-414c-a611-8012902a3458_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..ff97b9b022348308e1bd4b6a6554364c84e23c37 --- /dev/null +++ b/2024/Retrieval Robust to Object Motion Blur/191ce9d4-edfe-414c-a611-8012902a3458_content_list.json @@ -0,0 +1,2966 @@ +[ + { + "type": "text", + "text": "Retrieval Robust to Object Motion Blur", + "text_level": 1, + "bbox": [ + 269, + 141, + 732, + 162 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Rong Zou1, Marc Pollefeys1,2, and Denys Rozumnyi1,3", + "bbox": [ + 285, + 189, + 712, + 205 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ ETH Zürich, $^{2}$ Microsoft, $^{3}$ Czech Technical University in Prague", + "bbox": [ + 284, + 215, + 715, + 231 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Moving objects are frequently seen in daily life and usually appear blurred in images due to their motion. While general object retrieval is a widely explored area in computer vision, it primarily focuses on sharp and static objects, and retrieval of motion-blurred objects in large image collections remains unexplored. We propose a method for object retrieval in images that are affected by motion blur. The proposed method learns a robust representation capable of matching blurred objects to their deblurred versions and vice versa. To evaluate our approach, we present the first large-scale datasets for blurred object retrieval, featuring images with objects exhibiting varying degrees of blur in various poses and scales. We conducted extensive experiments, showing that our method outperforms state-of-the-art retrieval methods on the new blur-retrieval datasets, which validates the effectiveness of the proposed approach. Code, data, and model are available at https://github.com/Rong-Zou/Retrieval-Robust-to-Object-Motion-Blur.", + "bbox": [ + 261, + 250, + 738, + 458 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: object retrieval $\\cdot$ object motion blur $\\cdot$ datasets", + "bbox": [ + 261, + 470, + 660, + 484 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 508, + 375, + 523 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The task of image retrieval has long been a focal point in computer vision research, aiming to efficiently identify images within a large-scale database that share similarities with a given query image. In the realm of image retrieval, methods are often categorized into category-level and instance-level retrieval. We focus on the latter, where the objective is to identify images of a specific object featured in a query within a vast, unordered collection of database images. In this task, image representations play an important role since they capture the essence of the objects in images and enable the measurement of their similarities. Representing images based on hand-crafted features has been extensively studied [3,22]. Owing to the development in deep learning, recent strides have been made in adopting learning-based approaches to encode image content [2,4,20,25,30,41,42,46,47]. However, the current literature focuses exclusively on image representations for sharp objects, assuming that the objects are stationary and clearly visible in images. By only considering sharp and static objects, the existing retrieval methods fall short in addressing the dynamic nature of real-world scenarios, where objects are in movement, resulting in images blurred by object motion — a phenomenon prevalent in environments with fast-moving objects [32,34], low camera frame rates, and extended exposure times.", + "bbox": [ + 212, + 537, + 787, + 809 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To bridge this gap, we introduce a novel task in image retrieval: retrieval with object motion blur. The challenge of this task lies in finding image representations", + "bbox": [ + 215, + 810, + 787, + 840 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/e5851576a0a04cc2b93fcfeace25333c07213c3d62bdcd0f609441a55d80beea.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 225, + 185, + 313, + 239 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/e63d0c979c7d7904bc7a10e2bf78ff5ae5fcbc0cb23c50096e60ea63e030a6fc.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 225, + 292, + 313, + 345 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/b00e63097785baab53a42b295c90ab69424271df59bb8a6d748b9cbec0dace1b.jpg", + "image_caption": [ + "Query" + ], + "image_footnote": [], + "bbox": [ + 316, + 157, + 405, + 210 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/e7208f9d984d549405dec0a8a8e773922a8865fa6f6cc7c6090e7aaed5d18b4c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 318, + 210, + 405, + 263 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/4e705d91bde792b69dc765b0acaace0d17dd8a3577dd91f3ba55780698c0fab5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 318, + 265, + 405, + 316 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/40783217d654850d6ad97c67fbfe2ca0a9f9c84fc2112b4ef8211c7b08f6aaf7.jpg", + "image_caption": [ + "Fig. 1: Retrieval robust to object motion blur. We present the first approach designed for retrieval in the presence of object motion blur. Our method captures blur-invariant features and generates blur-robust representations. We show retrieval results on the newly created datasets for the task of blur-robust retrieval: synthetic (top) and real (bottom). All retrieved images match the query and are marked with green boxes." + ], + "image_footnote": [], + "bbox": [ + 318, + 318, + 405, + 369 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/574e1070ab18bb9ba8340f7146292622aada84baa328454121d807ef710bde1b.jpg", + "image_caption": [ + "Top 10 retrieval results" + ], + "image_footnote": [], + "bbox": [ + 410, + 157, + 496, + 210 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/40d57a3714f9f975241ac1cbc4080aea20390cf1ceefcc80ffb5db007ebc195f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 410, + 210, + 496, + 263 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/52b982c77ccb4aaa5a0f91fef2f26d8b314d53981ae47741b94c37d732391ab9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 410, + 265, + 496, + 316 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/9b44bec1c9c2c3ee2d1f2af2d0ed24223d58c6660f6f1432cf4099f763ed1b19.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 410, + 318, + 496, + 369 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/c586f9365b2165fe63a5b7359f51847adb1c2df5036553a4388306886de0d0a3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 501, + 157, + 588, + 210 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/ebcf9d7f5e9eefa8660749610112ea9f4864dc2d831fb50f008e56289ac71ddc.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 501, + 210, + 588, + 263 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/2490eb1ce51b54e495c33b0da719918fd1116f0afbb8e3f6fe2242d308bc121b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 501, + 265, + 588, + 316 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/67e1059172bf8e1c60c7b7fe3be40d2791ff5831a50eb875fbcf31bc51b7829d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 501, + 318, + 588, + 369 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/3b829b67ef521d45aa31806447fb233ef732f44998ad749e73024fa7b3fbca2a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 591, + 157, + 679, + 210 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/194f5851dc0176ba4dd5548e708e5fe84e344ee68720f4bc9a82831865a0822a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 591, + 210, + 679, + 263 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/9cfcdb3a3294850bc9844cdb76c8c743972dbc7bbadfc4b54e70da314000db36.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 591, + 265, + 679, + 316 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/32201676e556d5003050b8c881a61bfc979b5e35bc03d5b6851ed93a2fd62097.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 591, + 318, + 679, + 369 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/7a513e771ef7cdbf33d3c006c877b76aadf2d581e395aa8abebf722c8430b207.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 683, + 157, + 771, + 210 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/c1b2d150eb6a2a6f4f73a076e08c4f6443ddf9ea793a7e9b0f67bf399b1fd77b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 683, + 210, + 771, + 263 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/ae9b24a76edafa4dc45f4c9a5289659513968840ef374f6456f5c7e38443de9a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 683, + 265, + 771, + 316 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/b52349e974aa6ca3fc8ab09e6276b0b938b93939feba35a9e60e2757f41b8209.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 683, + 318, + 771, + 369 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "robust to object motion blur, in addition to other variations such as background and changes in viewpoint. We also propose the first approach targeting this task that produces representations invariant to motion blur. In response to the unique challenge posed by object motion, we propose loss functions designed to enhance the model's comprehension of blur, thereby ensuring robust retrieval performance even under conditions of extreme blurring, as seen in Fig. 1.", + "bbox": [ + 212, + 458, + 784, + 547 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recognizing the absence of datasets accommodating images with motion-blurred objects in existing retrieval benchmarks, we introduce the first benchmark datasets for this task, including a synthetic dataset for training and evaluation and a real-world dataset created specifically for evaluation. By re-training and evaluating state-of-the-art instance retrieval methods on this comprehensive benchmark, we reveal their shortcomings in handling motion-blurred scenarios.", + "bbox": [ + 212, + 550, + 784, + 638 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "This work marks a key step in advancing practical methods for object retrieval in the challenging presence of motion blur. It has potential applications such as endangered wildlife monitoring, sports analysis, and security surveillance.", + "bbox": [ + 212, + 640, + 784, + 684 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To summarize, our main contributions are as follows:", + "bbox": [ + 238, + 685, + 622, + 699 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "(1) We present the first method specifically designed for the novel retrieval task involving object motion blur. The method is trained with loss functions tailored to improve its understanding of motion blur.", + "(2) We introduce a new benchmark featuring synthetic and real-world datasets specifically constructed for this task. The datasets are carefully processed and are directly applicable for future research in blur retrieval.", + "(3) We conducted extensive experiments and detailed analyses to validate the effectiveness of our approach. Our method outperforms state-of-the-art retrieval methods and demonstrates superior robustness to motion blur." + ], + "bbox": [ + 215, + 704, + 784, + 839 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "R. Zou et al.", + "bbox": [ + 271, + 114, + 357, + 126 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2 Related work", + "text_level": 1, + "bbox": [ + 215, + 143, + 382, + 160 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Standard retrieval. In the area of standard image retrieval, local feature-based methods have evolved from traditional hand-crafted techniques like SIFT [22] and SURF [3] to contemporary deep learning-based approaches such as DELF [25] and HOW [41]. Considering that motion blur can corrupt local information, which may make local descriptors less robust, we adopt a global approach for our task. Classical methods typically derive global features by aggregating hand-crafted local features using techniques such as BoW [38], Fisher vectors [27], VLAD [15], AMSK [40], etc. Modern deep learning-based approaches, on the other hand, generate global descriptors by applying various pooling operations on feature maps encoded by convolutional neural networks (CNNs), such as sum pooling [1,41], weighted sum pooling [16], max pooling, regional max pooling [42], and generalized mean-pooling (GeM) [30], which has gained prominence in recent literature [4,20,47]. Our method integrates insights from recent advancements in global feature-based methods, adopting GeM pooling [30] for feature extraction and incorporating contrastive loss [6] and ArcFace loss [9] during training.", + "bbox": [ + 212, + 172, + 787, + 398 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Among the single-pass global feature-based methods, state-of-the-art standard retrieval approaches such as DELG global branch [4], DOLG [47], and Token [46] face challenges when handling images affected by motion blur, leading to degraded performance. Our work stands out in addressing this critical gap in the literature, as existing retrieval methods overlook the impact of motion blur. To bridge this research gap, we propose novel loss functions aimed at enhancing the model's comprehension of object motion blur, thereby contributing to the advancement of image retrieval in motion-blurred scenarios.", + "bbox": [ + 212, + 400, + 787, + 520 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Blur handling. Handling blur in images is an important topic in computer vision. [43] investigated the effect of different types of blur (defocus, uniform subject motion, camera shake) on the performance of CNNs for image classification and semantic segmentation. [37] studied online object detection under the influence of camera ego-motion-induced blur. Image and video deblurring are long-standing tasks in computer vision, which have been addressed by a variety of methods [10,19,44,49,50]. Joint deblurring and trajectory estimation for motion-blurred objects have also been extensively explored, with notable contributions from methods like [18, 31, 33, 34], some extending their scope to 3D reconstruction [35,36]. These existing works predominantly focus on other types of blur or address distinct tasks, emphasizing the need for specialized efforts in the context of our targeted task of retrieval involving object motion blur.", + "bbox": [ + 212, + 521, + 787, + 703 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Datasets. Several datasets have been proposed that contain generic corruptions including blur, such as ImaneNet-C [13] and Coco-C [23], but they do not focus purely on blurry images, and the types of blur they encompass (Gaussian, defocus, frosted glass, camera shake, and zoom) are not induced by object motion. Many datasets contain blurry-sharp image pairs, such as RealBlur [14] and GoPro [24], but they are created for deblurring tasks and not for retrieval, and the blur in images is mainly caused by camera motion or camera shake, not by object motion. A large variety of standard retrieval datasets have been proposed, including Revisited Oxford5k and Paris6k [28], Google Landmarks V2 [45, 48],", + "bbox": [ + 212, + 704, + 787, + 840 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "Retrieval Robust to Object Motion Blur", + "bbox": [ + 462, + 114, + 730, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/c668d320ad6faf9330c604d900060fd3025451de5b1bf69b7feb94de94df8068.jpg", + "image_caption": [ + "Fig. 2: Overview of the proposed method for blur-robust retrieval. Our model takes a query image as input. Then, it generates a descriptor robust to object motion blur through the Blur Robust Image Descriptor Generator (BRIDGE) module that encodes the input image using a CNN backbone followed by generalized mean-pooling (GeM) [30], extracting a feature vector. This vector is subsequently fed into three heads: the blur estimation head, classification head, and localization head. Each head processes the feature vector to extract relevant information for identifying motion-blurred objects. The resulting three processed feature vectors are concatenated and passed through the final fully connected (FC) layer to produce the blur-robust descriptor. During training, the Blur Level-based Image Sample Selection (BLISS) mechanism is employed to select contrastive samples based on the query blur level (BL, Eq. (6)) $b_{q}$ and a specified blur level range $r$ . Then, these selected image samples are input into the BRIDGE module to extract descriptors for subsequent contrastive learning." + ], + "image_footnote": [], + "bbox": [ + 215, + 145, + 785, + 281 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "SfM120K [29], etc. However, none of them incorporate motion-blurred images. To the best of our knowledge, there is currently no existing dataset tailored for the task of retrieval with object motion blur. In view of this, we created synthetic and real-world datasets specifically designed for the novel task of blur retrieval.", + "bbox": [ + 212, + 491, + 787, + 554 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3 Method", + "text_level": 1, + "bbox": [ + 215, + 593, + 330, + 609 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We perform retrieval via nearest neighbor search in a representation space. Therefore, we require a representation model that is robust to object motion blur. We train a deep learning-based model that maps images to a vector space. The network is trained on our proposed synthetic dataset using multiple losses, including those tailored to object motion blur. As we show, the method generalizes well to real data despite being trained on synthetic data.", + "bbox": [ + 212, + 628, + 785, + 718 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The network takes an RGB image $\\mathbf{I} \\in \\mathbb{R}^{3 \\times H \\times W}$ containing a sharp or blurred object as input and generates a descriptor $\\mathbf{D} \\in \\mathbb{R}^d$ robust to object motion blur through the Blur Robust Image Descriptor Generator module (BRIDGE, Sec. 3.1). For training only, contrastive samples are selected by the Blur Level-based Image Sample Selection mechanism (BLISS, Sec. 3.2) based on the query blur level $b_q$ and the contrast blur level range $r$ . The selected image samples are sent into the BRIDGE module to extract descriptors for contrastive learning. The architecture of the proposed blur retrieval method is shown in Fig. 2.", + "bbox": [ + 212, + 718, + 787, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "R. Zou et al.", + "bbox": [ + 271, + 114, + 357, + 126 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1 BRIDGE module", + "text_level": 1, + "bbox": [ + 215, + 146, + 405, + 159 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The BRIDGE module encodes an image into a descriptor that is robust to object motion blur. The input image is encoded by a ResNet-50 backbone [12] to produce a feature map $\\Omega \\in \\mathbb{R}^{C\\times U\\times V}$ . Then, we use GeM pooling [30] to process the feature map into a feature vector $\\mathbf{f}\\in \\mathbb{R}^C$ , expressed as", + "bbox": [ + 212, + 169, + 787, + 231 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {f} = \\left[ \\left(\\frac {1}{| \\boldsymbol {\\Omega} |} \\sum_ {(u, v) \\in \\boldsymbol {\\Omega}} x _ {(c, u, v)} ^ {p}\\right) ^ {\\frac {1}{p}} \\right] _ {c = 1, \\dots , C}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 357, + 239, + 785, + 297 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $p > 0$ is a parameter, and setting it to values larger than 1 forces the pooled vector to focus on the salient areas. This feature vector then undergoes multi-head processing to generate a motion blur-robust descriptor. Specifically, in the blur estimation head, we obtain blur estimation features $\\mathbf{f}_{be} \\in \\mathbb{R}^{C_b}$ by a fully connected (FC) layer. The same dimension reduction is performed in the localization head, which generates features for location prediction $\\mathbf{f}_{loc} \\in \\mathbb{R}^{C_l}$ . The features generated from the two heads are then concatenated with the whitened [11] features $\\mathbf{f}_{cls} \\in \\mathbb{R}^{C_c}$ from the classification head. We further reduce the dimension of the concatenated features to obtain the final blur-robust descriptor of the image. This can be expressed as:", + "bbox": [ + 214, + 306, + 787, + 457 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {D} = \\mathbf {W} \\cdot \\operatorname {C o n c a t} \\left(\\mathbf {W} _ {b e} \\mathbf {f}, \\mathbf {W} _ {l o c} \\mathbf {f}, \\mathbf {W} _ {c l s} \\mathbf {f}\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 359, + 467, + 785, + 484 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathbf{W}_{be} \\in \\mathbb{R}^{C_b \\times C}$ , $\\mathbf{W}_{loc} \\in \\mathbb{R}^{C_l \\times C}$ and $\\mathbf{W}_{cls} \\in \\mathbb{R}^{C_c \\times C}$ are weights of the FC layers in the blur estimation, localization, and classification heads, respectively, and $\\mathbf{W} \\in \\mathbb{R}^{d \\times (C_b + C_l + C_c)}$ is the weight of the final dimension-reduction layer.", + "bbox": [ + 214, + 491, + 787, + 540 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2 BLISS mechanism", + "text_level": 1, + "bbox": [ + 215, + 559, + 413, + 573 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "While the specific representation of the degree of motion blur for objects in images is yet to be formally defined, it is intuitively perceived as a continuous value. The level of detail exhibited by objects varies with the degree of blur: sharper objects show clear structures, textures, and poses, while more blurred objects primarily reveal color characteristics (see Fig. 3). This continuous nature of blur allows us to indirectly control the distance between samples with significantly different blur conditions during learning image representations. In this context, we manage the distance between samples that exhibit visual features and patterns of the same level of detail due to similar degrees of blur, thus indirectly influencing the distance between samples with significantly different levels of blur. Additionally, we can directly control the distance between samples with large differences in blur conditions by imposing constraints directly between such samples in the descriptor space. There are therefore two key research questions: What is the optimal value for the range of differences in the blur levels of image pairs used in contrastive learning? How does adjusting the range of differences in blur levels affect the model's performance? To answer these questions and identify the optimal blur level range for contrastive learning, we introduce the BLISS", + "bbox": [ + 214, + 583, + 787, + 843 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "Retrieval Robust to Object Motion Blur", + "bbox": [ + 460, + 114, + 730, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "BL 1", + "bbox": [ + 222, + 143, + 259, + 155 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/6642b79676364e8c89cb675831934e6c3fde2d4eb2122b89c7729a17e3cd8fff.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 218, + 157, + 264, + 212 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/6ba913a8085179d9d0c064a8f9e2bf5098f5b0a8728148afd5c225d23b8fed49.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 218, + 212, + 264, + 267 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/7e0e2c791b150efe50efba756ff0351bf0cff18ee89f1e22fafc537f34bfa876.jpg", + "image_caption": [ + "Fig. 3: The introduced synthetic (a) and real-world (b) datasets. Rows 1-2: different trajectories of the same object. Rows 2-3: two different objects from the same category with similar shapes. Rows 4-6: objects from different categories but share similar textures. Columns correspond to different blur levels, from 1 to 6, and they are different for synthetic (BL) and real-world $(\\mathrm{BL}^{\\mathrm{r}})$ datasets." + ], + "image_footnote": [], + "bbox": [ + 220, + 268, + 264, + 323 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "BL2", + "bbox": [ + 271, + 145, + 307, + 155 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/0f633b48db9977093ed4859f535cff3bcf61ee5628e6c223098311ff5824c2b5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 267, + 156, + 312, + 212 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/3978cfa63644f3da896ec7b2d4609c0f41ae7bcb1a1182b80c55f98439b034dd.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 267, + 212, + 312, + 267 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/b0f9afc4191a68846c2b13254c21719342c071fc30bfa92788ee9dc754782e56.jpg", + "image_caption": [ + "(a) Synthetic dataset." + ], + "image_footnote": [], + "bbox": [ + 267, + 267, + 312, + 323 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "BL 3", + "bbox": [ + 321, + 145, + 357, + 155 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/950a91c0f9c25a41324b5c339ebd9375e9dbb4f69d5273528c712f5c043428f9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 313, + 156, + 357, + 212 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/2904c6d27de330c7d66f9e4086e13ba73ff6ce714659eae22fbd98d8bd3a35d4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 313, + 212, + 357, + 267 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/31836addd03294e950c0f7f787519d790f957c029e912c561ea8b1a730214604.jpg", + "image_caption": [ + "Synthetic dataset." + ], + "image_footnote": [], + "bbox": [ + 313, + 267, + 357, + 323 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4 B", + "bbox": [ + 392, + 145, + 426, + 155 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/76b58f60691c164ee6fccfb6f132896910fdb88e7811aa9f965e471627e3de40.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 362, + 156, + 406, + 212 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/c3d999b22e224f92773ff102fe420d91d6598caa7a29a474a43f407ceb2556c0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 361, + 212, + 405, + 267 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/19babffa266ed2d505c33ae2a8230f98671482750f9f3ca3648c16cf7cc2b216.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 361, + 267, + 405, + 323 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "ic datas", + "bbox": [ + 361, + 324, + 405, + 335 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "BL 5", + "bbox": [ + 411, + 145, + 449, + 155 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/561de1022d59e8c05bc3923c61a1f6ea1b4e34d208cf450139b38e2c85e2a0b3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 410, + 156, + 452, + 212 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/937e81c6f721f3aa2092fe093b01de05730f31be1f82ae4bf48f9af4d84908c5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 410, + 212, + 452, + 267 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/c2086dbcc11599aec790972550f698f97794bc37dbcac2ffccf25f85d4d69c5e.jpg", + "image_caption": [ + "et." + ], + "image_footnote": [], + "bbox": [ + 410, + 267, + 452, + 323 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "BL6", + "bbox": [ + 460, + 145, + 496, + 155 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/a7ded884e044cd478ea5912890434bc5bc50b4bc6cff20c995552a87ece343b3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 459, + 156, + 501, + 212 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/f1b8d1f3a0448d1917d874be2a2d1ae350a6e6638450d933c2ee86eb9fa8277b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 459, + 212, + 501, + 267 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/df59a374dc2634ace7b5b87c938fa7bdd9b38b31094bf162828f1522e03672cd.jpg", + "image_caption": [ + "" + ], + "image_footnote": [], + "bbox": [ + 459, + 267, + 501, + 323 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "BLr1", + "bbox": [ + 511, + 145, + 552, + 155 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/797233a28297aa13af5eec7b44a23ab187ebbf6fe7ef11237a3a25a9980890f0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 509, + 156, + 553, + 212 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/6cd73b5fada4d8af40c25e095d416fbd2253e746053b3d1f7fff01f7292399aa.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 509, + 212, + 553, + 267 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/f520f16b08761cdb23e789cce72628bfa883fb95f01130fdd69a5afce426292b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 509, + 267, + 553, + 295 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/f9b0e83ecc240de07685ea5923221a8a5d4496b62daeb1e19ed363af5190a95e.jpg", + "image_caption": [ + "" + ], + "image_footnote": [], + "bbox": [ + 509, + 295, + 553, + 323 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "BL 2", + "bbox": [ + 553, + 145, + 598, + 155 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/1e952ae31f57eed977581ae0316b9a396f14527f97c4ac93e060a513d2d3074d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 553, + 156, + 599, + 212 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/f1212ffd355866d7a77a5cf1446569f46f63b54241e7299ac1a2d0227134a9b5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 553, + 212, + 599, + 267 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/cccb591f91afdfbe3b05a4bb0e6a737e9e82e1b230e286847914ba0ed6c64aeb.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 553, + 267, + 599, + 295 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/713ad930d3dec3548b1bb0c45c216766e38990654989e0bf4f5af281e42602ce.jpg", + "image_caption": [ + "(b) F" + ], + "image_footnote": [], + "bbox": [ + 553, + 295, + 599, + 323 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "BLr3", + "bbox": [ + 599, + 145, + 645, + 155 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/62a5ed96d8abaac7cfe00f8f84ab9d1dd92bd99e9990d6147f1bac85222ae08b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 599, + 156, + 645, + 212 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/f00b4aea26f8e21171959ec9493e2251e4d2ee9d7cb4d3a23d5ee6cde34582e5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 599, + 212, + 645, + 267 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/21506a60ac01a8fff38b4c766eb37cd7d2285ebb2fb39467ee1615debae43d70.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 599, + 267, + 645, + 295 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/56fc5975b524fe2713d15aa2c23f8c9f5c250183b8e81dcb09ea4624897cdb65.jpg", + "image_caption": [ + "Real-wo" + ], + "image_footnote": [], + "bbox": [ + 599, + 295, + 645, + 323 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "BLr 4", + "bbox": [ + 651, + 145, + 696, + 155 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/03859a3d6c8fbf13211382455ac3cdb897e78d337e03169079a0af95993ae3af.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 651, + 156, + 696, + 212 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/c94246543f4209f5b5392d8ff2634f6f57df7cd21cdab4e26b2ab0cd65c6f75c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 651, + 212, + 696, + 267 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/261d641d9962aea66d0965c93b0535df90d5410a5953281c3ae704858f27dd7c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 651, + 267, + 696, + 295 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/f26157c1992fc2dd4e68fbabb09d904aad0d5d08e0ac1251fe42f14c82a4fd48.jpg", + "image_caption": [ + "ld data" + ], + "image_footnote": [], + "bbox": [ + 651, + 295, + 696, + 323 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "BL 5", + "bbox": [ + 699, + 145, + 743, + 155 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/795f32b182797c152ada88fd63301d7201669cbc68003902f4e2c4b9f0bf7f9c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 699, + 156, + 743, + 212 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/2f3c27a1ceb25eba907ad14950ce5b494742c67b633b33d28deeb924419ca33f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 699, + 212, + 743, + 267 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/bacca14b8aae2545db7c510374ba9f24de249da0dc3d5e261adf0a47edf71e93.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 699, + 267, + 743, + 295 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/6a27d4d80f04cf26bd9bde1fdb6af6aa554700f54e39d3e8275f42466a1c51b7.jpg", + "image_caption": [ + "set." + ], + "image_footnote": [], + "bbox": [ + 699, + 295, + 743, + 323 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "mechanism. This mechanism selects contrastive samples based on the blur level in the query image for contrastive learning.", + "bbox": [ + 212, + 422, + 784, + 453 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Blur severity definition. Our definition of blur severity is inspired by the image generation process. The formation of an image $\\mathbf{I}$ containing a motion-blurred object in front of a background can be expressed [32] as", + "bbox": [ + 212, + 454, + 784, + 501 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {I} = \\mathbf {P} * \\mathbf {O} + (1 - \\mathbf {P} * \\mathbf {M}) \\mathbf {B} \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 401, + 508, + 784, + 526 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where the first term denotes the convolution $(\\ast)$ of a point spread function $\\mathbf{P}$ corresponding to a motion trajectory and the object's sharp appearance $\\mathbf{O}$ , and the second term depicts the contribution of the background $\\mathbf{B}$ , in which $\\mathbf{M}$ is the binary mask of $\\mathbf{O}$ . Let", + "bbox": [ + 212, + 534, + 784, + 594 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {\\alpha} = \\mathbf {P} * \\mathbf {M}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 455, + 594, + 784, + 611 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "$\\pmb{\\alpha} \\in \\mathbb{R}^{H \\times W}$ , $\\pmb{\\alpha}(i,j) \\in [0,1], \\forall (i,j) \\in \\mathbf{I}$ . Considering only objects with opaque materials, $\\pmb{\\alpha}$ represents the visibility of the object. In other words, it can characterize the amount of blur. Generally, the sharper the object in the image, the larger the $\\pmb{\\alpha}$ values are. However, at the edge pixels, the object blends with the background, and the alpha values of these edge pixels gradually fade to 0. Combining the alpha mask from Eq. (4) and considering edge effects, we define the Blur Severity (BS) of an image with a motion-blurred object as:", + "bbox": [ + 212, + 614, + 785, + 722 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nB S = 1 - \\frac {\\sum_ {i = 1} ^ {H} \\sum_ {j = 1} ^ {W} \\boldsymbol {\\beta} (i , j) \\cdot \\boldsymbol {\\alpha} (i , j)}{\\sum_ {i = 1} ^ {H} \\sum_ {j = 1} ^ {W} \\boldsymbol {\\beta} (i , j)}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 362, + 728, + 784, + 772 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\beta$ is a thresholded (by 0) and eroded (by 3 pixels) binary mask of $\\alpha$ . Erosion effectively removes the adverse contribution of the edge pixels to the blur severity measure. As a result, $BS \\in (0,1)$ , and the larger the value, the more severe the motion blur in the image.", + "bbox": [ + 212, + 779, + 785, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "R. Zou et al.", + "bbox": [ + 271, + 114, + 357, + 126 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/024088ae819a06654f02c553f1cdd3465b97a58d7d5aa0b0197168f1820eef4e.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Dataset#Images Total#Images by BL (Syn.) or BLr (Real)
123456
Syn. Q20,9954,2883,9324,0784,0892,9301,678
Syn. DB91,62118,87117,50817,88818,02912,54667,79
Syn. Dist.1,091,939214,364177,869222,542235,662149,82891,674
Real Q2,753612620561396315249
Real DB10,3401,9231,8032,0801,7451,3751,414
", + "bbox": [ + 220, + 143, + 787, + 247 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 1: Distribution of images in our synthetic and real datasets, sorted by blur levels. 'Q' stands for 'query', 'DB' denotes database, and 'Dist.' represents distractors.", + "bbox": [ + 215, + 250, + 782, + 275 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Blur level-based sample selection. Since blur severity is a continuous quantity, finding images within a certain blur severity range for each query image is expensive. We thus use a discretized version of blur severity — Blur Level (BL):", + "bbox": [ + 214, + 295, + 782, + 340 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\nB L = \\lceil 1 0 B S \\rceil , \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 447, + 356, + 784, + 372 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $\\lceil \\cdot \\rceil$ denotes ceiling operation. Each image is assigned to a $BL$ according to its $BS$ . During training, we select contrastive samples based on the query blur level $b_{q}$ and the specified contrastive $BL$ range $r$ . The selected $N_{p}$ positive and $N_{n}$ negative samples thus have blur levels within the range of $b_{q} - r$ to $b_{q} + r$ .", + "bbox": [ + 214, + 375, + 782, + 436 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.3 Training objectives", + "text_level": 1, + "bbox": [ + 215, + 454, + 419, + 469 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Blur estimation loss. Blur severity contains information about the visibility of the blurred object in an image (Sec. 3.2). From another perspective, it reflects a signal-to-noise ratio (SNR), since the more severe the motion blur, the greater the ratio of background (noise) to foreground (signal). In view of this, we define blur estimation loss as", + "bbox": [ + 212, + 476, + 782, + 549 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {b e} = | p - (1 - B S) |, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 423, + 551, + 784, + 568 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $p \\in \\mathbb{R}$ is the blur value predicted by the blur estimation head. By constraining the estimate of the continuous $BS$ (Eq. (5)), we implicitly force the network to identify blur and estimate the SNR, which helps it to focus on and extract signal information in the blurred region while ignoring background noises.", + "bbox": [ + 212, + 571, + 782, + 631 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Localization loss. This loss is defined as", + "bbox": [ + 215, + 633, + 526, + 646 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {l o c} = | x - \\hat {x} | + | y - \\hat {y} | + | h - \\hat {h} | + | w - \\hat {w} |, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 346, + 655, + 784, + 671 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $(x,y,w,h)$ and $(\\hat{x},\\hat{y},\\hat{w},\\hat{h})$ represent the normalized ground truth bounding box and the predicted one from the localization head, respectively. Localization loss requires the network to distinguish motion-blurred objects from backgrounds under various blur conditions, forcing it to recognize blur, identify objects at different blur levels, and further, focus more on the foreground objects.", + "bbox": [ + 212, + 678, + 782, + 753 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Classification loss. Following DELG [4], we apply ArcFace margin loss [9] to the descriptor $\\mathbf{D}$ , expressed as", + "bbox": [ + 215, + 755, + 782, + 786 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {c l s} = - \\log \\left(\\frac {\\exp (\\gamma \\times A F (\\hat {\\mathbf {w}} _ {t} ^ {T} \\hat {\\mathbf {D}} , 1))}{\\sum_ {n} \\exp (\\gamma \\times A F (\\hat {\\mathbf {w}} _ {n} ^ {T} \\hat {\\mathbf {D}} , y _ {n}))}\\right), \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 333, + 792, + 784, + 842 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "Retrieval Robust to Object Motion Blur", + "bbox": [ + 460, + 114, + 732, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/3ff9f129d1a0e5210948570e1ecca09b2c92bfeb08c8a8f102792c5b102694d9.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodmAP by query BL (synthetic)mAP by query BL (synthetic + 1M)
All123456All123456
DELG [4]81.6583.5584.9684.8982.6977.3566.2068.1973.6475.4073.3468.0558.2842.46
DOLG [47]83.0385.1086.2685.9384.0478.7868.0769.9775.7577.4775.0170.1060.0142.49
Token [46]84.8486.5387.9987.5085.8180.7071.5770.6575.3277.6675.5170.2461.1948.05
Ours-sharp27.5743.0638.5329.7419.4010.876.1432.6471.9343.8827.1815.417.944.27
Ours91.7893.0593.4893.1492.2590.2082.8684.0988.7489.5687.6884.4176.8962.42
", + "bbox": [ + 218, + 143, + 787, + 218 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 2: Retrieval results on the synthetic dataset with (+1M) and without distractors. Ours-sharp is our model trained with only sharp images. 'All' denotes the overall performance for queries of all blur levels. All methods are re-implemented and re-trained on our synthetic training set. The database contains images of all blur levels.", + "bbox": [ + 215, + 220, + 785, + 275 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $\\gamma$ is a scale factor, $\\hat{\\mathbf{w}}_{\\mathbf{i}}$ refers to the $i$ -th row of the $L_{2}$ normalized $N$ -class classifier $\\hat{\\mathbf{W}} \\in \\mathbb{R}^{d \\times N}$ , $\\hat{\\mathbf{D}}$ denotes the $L_{2}$ normalized $\\mathbf{D}$ . $y$ is the one-hot label, and $t$ represents the ground truth class index ( $y_{t} = 1$ ). $AF$ is the ArcFace-adjusted cosine similarity and it can be expressed as", + "bbox": [ + 212, + 295, + 782, + 354 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\nA F (s, g) = \\left\\{ \\begin{array}{l l} \\cos (\\operatorname {a r c c o s} (s) + m), & \\text {i f} g = 1 \\\\ s, & \\text {i f} g = 0 \\end{array} , \\right. \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 346, + 364, + 785, + 402 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $s$ is the cosine similarity, $m$ is the ArcFace margin, and $g \\in \\{0,1\\}$ . $g = 1$ indicates the ground truth class. The classification loss requires the model to group sharp or motion-blurred images of the same object together, while distinguishing between different objects, which may appear similar, especially when blurred. This implicitly forces the model to focus on blur-invariant features.", + "bbox": [ + 212, + 410, + 782, + 484 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Contrastive loss. We employ contrastive learning [6] to construct a descriptor space, where the distance between similar samples is minimized, and between dissimilar samples maximized. The contrastive loss is applied to (query, positive) and (query, negative) descriptor pairs:", + "bbox": [ + 212, + 487, + 782, + 547 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {c o n}} (i, j) = \\left\\{ \\begin{array}{l l} 0. 5 \\cdot D _ {i, j} ^ {2}, & \\text {i f} y (i, j) = 1 \\\\ 0. 5 \\left(\\max \\{0, \\tau - D _ {i, j} \\}\\right) ^ {2}, & \\text {i f} y (i, j) = 0 \\end{array} , \\right. \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 313, + 556, + 784, + 595 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $D_{i,j} = ||\\hat{\\mathbf{D}}(i) - \\hat{\\mathbf{D}}(j)||$ and $\\hat{\\mathbf{D}}(i)$ is the $L_2$ normalized descriptor of image $i$ . $\\tau$ is a margin parameter, and $y(i,j)$ is binary indicating whether the pair matches $(= 1)$ . The contrastive loss further enhances the focus on blur-invariant features by explicitly forcing descriptors of the same object, whether sharp or blurred, to be close together, and those of different objects to be far apart.", + "bbox": [ + 212, + 604, + 782, + 680 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Joint loss. The joint loss is a weighted sum of all previously defined losses:", + "bbox": [ + 212, + 681, + 766, + 696 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {j o i n t}} = \\mathcal {L} _ {\\text {c o n}} + \\alpha_ {\\text {c l s}} \\mathcal {L} _ {\\text {c l s}} + \\alpha_ {\\text {b e}} \\mathcal {L} _ {\\text {b e}} + \\alpha_ {\\text {l o c}} \\mathcal {L} _ {\\text {l o c}}. \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 341, + 707, + 784, + 722 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4 Dataset", + "text_level": 1, + "bbox": [ + 215, + 765, + 328, + 780 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We create novel synthetic and real datasets for training and evaluating retrieval methods for motion-blurred objects. The synthetic dataset serves for both training and evaluation. We also generate a large-scale synthetic distractor set to", + "bbox": [ + 212, + 794, + 782, + 839 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "R. Zou et al.", + "bbox": [ + 271, + 114, + 357, + 126 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/ccee4c7216ab74a3ae1723748a9b2cc277f04211aaee40d45ab3ecc07616686b.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodmAP AllmAP by Query BLr
123456
DELG [4]54.8249.1363.4357.2555.0153.7742.92
DOLG [47]54.6443.9360.5958.3659.0658.5845.78
Token [46]43.3338.7147.0850.7946.4442.7124.43
Ours-sharp40.2449.5545.0241.3333.2329.4027.91
Ours62.8857.5070.3866.7763.1864.4846.14
", + "bbox": [ + 220, + 143, + 784, + 253 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Table 3: Retrieval results on the real-world dataset (trained on synthetic without fine-tuning). The database includes all blur levels. Our method outperforms all others.", + "bbox": [ + 217, + 253, + 782, + 280 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "evaluate methods in a more challenging setup. The novel real-world dataset is specifically constructed to validate the methods' effectiveness in real scenarios. All datasets are meticulously processed for the blur retrieval task, laying a crucial foundation for future research. Statistics of these datasets are shown in Table 1.", + "bbox": [ + 217, + 297, + 782, + 357 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.1 Synthetic dataset", + "text_level": 1, + "bbox": [ + 217, + 382, + 403, + 396 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "The images in this dataset contain 3D objects moving along different trajectories, in front of different backgrounds, and with different amounts of motion blur. Thus, this dataset poses many challenges to retrieval methods.", + "bbox": [ + 217, + 407, + 782, + 450 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Backgrounds. Background images are sampled from the LHQ dataset [39], cropped to a resolution of $240 \\times 320$ .", + "bbox": [ + 217, + 454, + 782, + 483 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Objects. We select 39 categories with the largest number of textured objects from the ShapeNet dataset [5] and randomly sample 30 objects in each category.", + "bbox": [ + 217, + 484, + 782, + 515 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Trajectories. For each object, we generate 120 random linear trajectories, each divided into 23 sub-trajectories. We then select up to 10 consecutive sub-trajectories to create images featuring objects with diverse levels of motion blur.", + "bbox": [ + 217, + 517, + 782, + 561 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Rendering. We render blurry images with Blender Cycles [7] by capturing moving objects during the camera exposure time. Additionally, we render one sharp image per trajectory, placed at the central position along the trajectory.", + "bbox": [ + 217, + 561, + 782, + 608 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Filtering. We discard renderings where the object occupies less than $1.5\\%$ of the pixels or where the intersection-over-union between silhouette renderings at trajectory endpoints is below $20\\%$ . The former condition suggests the object is too small, while the latter indicates the appearance changes are too drastic.", + "bbox": [ + 217, + 609, + 782, + 670 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Ground truth. For each generated image, we estimate the bounding box coordinates by detecting the non-zero pixels in the alpha channel. We calculate the blur level for each image based on Eq. (6).", + "bbox": [ + 217, + 672, + 782, + 717 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Retrieval dataset. In total, we rendered 1.5 million images for 1138 different instances (examples in Fig. 3a). Additionally, we create a challenging distractor set with over 1 million images (examples in the supplementary material), using the same object categories as before to increase the difficulty of the distractors in terms of intra-class similarity. We randomly select 40 objects per category, excluding the 30 previously chosen objects, and generate 70 trajectories for each. For each trajectory, we render one sharp image and 10 blurry images with varying degrees of blur. Backgrounds are also selected from a different subset of [39].", + "bbox": [ + 217, + 719, + 782, + 839 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "Retrieval Robust to Object Motion Blur", + "bbox": [ + 462, + 114, + 730, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/0171945a1021d33ff40187b7c470e1bfc6e824366c8bd0525a6b9d2733c8cd6b.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
BLQDDELG [4] (range: 57.79; std: 16.75)DOLG [47] (range: 56.81; std: 16.93)
All123456All123456
175.9986.7384.1780.2672.4863.1250.0877.7887.0185.7081.8575.1166.2052.39
277.3582.7486.2282.8776.6866.4650.0078.4484.1887.0083.4377.9167.8851.31
377.1979.6283.4382.9478.4669.4352.8678.4981.8484.7483.8579.6970.5853.12
474.9674.2678.9179.8578.0570.8955.1276.4077.2780.5580.8079.1071.7355.28
562.7157.8661.7264.3765.3668.9656.0564.1961.3163.6865.3866.0569.6855.75
634.8929.5628.9431.7534.0645.3753.8635.5131.2430.2032.2934.1845.0053.41
BLQDToken [46] (range: 55.72; std: 16.11)Ours (range: 53.68; std: 15.65)
All123456All123456
179.6588.2287.3483.5076.8768.5156.6388.5795.0994.2491.5487.1980.9168.15
279.9784.7888.4884.7778.9970.0155.8288.7593.0794.6592.1888.7782.0367.18
379.5081.7985.4684.7580.2172.0958.1888.7691.7093.1892.3289.8383.4168.94
476.8276.5680.4281.2479.3172.6159.6187.1688.1290.1190.0989.3184.3770.35
565.3961.8564.4966.3167.1470.9660.3676.2873.8774.9076.0477.3983.0671.67
637.9833.2332.7634.6136.5047.5657.4347.3143.4341.4142.8944.5558.1369.59
", + "bbox": [ + 218, + 143, + 787, + 345 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Table 4: Retrieval results by query (Q) and database (D) blur level (BL) on our synthetic dataset (with 1M distractors). 'All' denotes the overall performance for queries of all blur levels. The standard deviation and range of mAP values are displayed alongside each method. The best scores under the same settings are marked in bold.", + "bbox": [ + 212, + 347, + 784, + 401 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.2 Real-world dataset", + "text_level": 1, + "bbox": [ + 215, + 417, + 416, + 431 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "We record videos with a GoPro 7 Hero camera at 240 fps with full exposure. We use 35 different objects and move each object 3 to 5 times along different trajectories in front of the camera. None of the real-world objects are in synthetic data and $2/3$ of which are from new categories not present in synthetic data. This better tests the generalization ability of models to new objects and categories. In total, we recorded 139 videos. Different amounts of blur are generated by averaging consecutive frames, ranging from just one up to 200 averaged frames.", + "bbox": [ + 212, + 440, + 784, + 545 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Computing blur level based on Eq. (6) is not feasible for the real-world dataset. Therefore, we manually annotate the data, assigning each image to one of 6 blur levels based on the perceived blur, and we use the ratio of the trajectory length to the average object size as a reference. The annotation enables a more detailed analysis and a more comprehensive evaluation of model performance. Further annotation details can be found in the supplementary material. Since the blur level for real data differs from that in Eq. (6), we denote it as $\\mathrm{BL}^{\\mathrm{r}}$ .", + "bbox": [ + 212, + 546, + 785, + 651 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "The real-world dataset contains 13,093 images, with an average of 374 images for each object (examples in Fig. 3b). Compared to the synthetic dataset, the object scale change is more drastic in different trajectories, e.g. rows 1 & 2. The trajectories are notably different from those in the synthetic dataset. They involve complex curved trajectories with acceleration and other physical forces, which poses an additional challenge for methods trained on our synthetic dataset.", + "bbox": [ + 212, + 652, + 785, + 742 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "5 Experiments", + "text_level": 1, + "bbox": [ + 215, + 763, + 375, + 780 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Setup. For each category of objects in the synthetic dataset, we randomly take $70\\%$ (rounded down, same for validation) for training (792 objects), $15\\%$ for validation (153 objects), and $15\\%$ for testing (193 objects). For the training objects,", + "bbox": [ + 212, + 794, + 785, + 840 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "R. Zou et al.", + "bbox": [ + 271, + 114, + 357, + 126 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/90bf25b235aad8fc46e634961eadf13e9b8574183c30253697f6e8e252ced634.jpg", + "image_caption": [ + "Query", + "(a) Example of retrieval results on our synthetic dataset (with distractors)." + ], + "image_footnote": [], + "bbox": [ + 225, + 157, + 784, + 276 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/31cfcdb2ee0ec2db6a6bda3958324ef169dafb0c823d271a7154a5f1b9f00dab.jpg", + "image_caption": [ + "(b) Example of retrieval results on our real-world dataset.", + "Fig. 4: Comparison of retrieval results. We compare to the best-performing state-of-the-art method for the corresponding query. The retrieved images are sorted from left to right and top to bottom, with the ranking from $1^{\\mathrm{st}}$ to $20^{\\mathrm{th}}$ . The blur level of each image is shown in the bottom left corner. Correct results are in green boxes, while incorrect ones are in red boxes. We provide more results in the supplementary material." + ], + "image_footnote": [], + "bbox": [ + 225, + 290, + 782, + 410 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "we take one image for each trajectory while ensuring a balanced distribution of images across different blur levels. In total, there are 95k images in the training set. For each of the test objects, we sample 20 trajectories to form a query set, and the rest is used for the database. The database images are sampled from all blur levels unless specified otherwise (for per blur level database experiments). We additionally show performance on a subset of queries for a given blur level. In all experiments (except Ours-sharp), only one model is trained for all blur levels, and the evaluation is performed for varying blur levels.", + "bbox": [ + 212, + 517, + 787, + 638 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Implementation details. The proposed method is implemented in PyTorch [26]. We use ResNet-50 [12] up to the last downsampling block as the backbone, deactivate the normalization layers, and initialize the model with ImageNet [8] pre-trained weights. ADAM [17] optimizer is employed to minimize the joint loss with a fixed learning rate $10^{-4}$ , batch size 32, and 30 epochs. We initialize the power parameter to 3 for GeM pooling [30]. The contrastive margin $\\tau$ is set to 0.7. Following [30], we set the number of contrastive samples as $N_{p} = 1$ , $N_{n} = 5$ . The contrastive blur range is experimentally set to 5. For classification, we empirically set the margin $m$ to 0.15 and the scale $\\gamma$ to 30. Loss weights are chosen so that the values of different losses are roughly of the same order of magnitude, i.e. $\\alpha_{cls} = 0.1$ , $\\alpha_{be} = 1.0$ , $\\alpha_{loc} = 10$ . The descriptor dimension $d$ is set to 128.", + "bbox": [ + 212, + 640, + 787, + 806 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Evaluation protocols. We report mean average precision (mAP@all) as commonly used in retrieval. In case 1M distractors are used, we report mAP@100.", + "bbox": [ + 212, + 809, + 785, + 839 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "Retrieval Robust to Object Motion Blur", + "bbox": [ + 462, + 114, + 730, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 116, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/457aefb7d1d7b1857c1c8276bd1352c91cc01fd0016a8b4460b0aa84502993be.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
\\( \\mathcal{L}_{con} \\)\\( \\mathcal{L}_{cls} \\)\\( \\mathcal{L}_{be} \\)\\( \\mathcal{L}_{loc} \\)mAP AllmAP by Query BL
123456
✓✘✘✘78.1380.5181.7081.1679.2073.9361.24
✓✘✓✘81.6683.4985.0184.4382.6977.6467.08
✓✘✘✓85.9487.5488.2587.8386.5283.0875.58
✓✘✓✓87.4888.6989.8989.4088.2484.9776.74
✘✓✘✘78.7381.5382.9782.8379.8672.9359.00
✘✓✓✘83.6785.1987.5687.4085.1079.2265.93
✘✓✘✓88.7489.8990.9190.8889.7586.0777.67
✘✓✓✓91.2392.0293.1693.0991.9789.0082.27
✓✓✘✘85.0687.4288.2987.6685.8581.2069.96
✓✓✓✘87.1789.0390.0389.5588.0783.9173.48
✓✓✘✓90.3991.8592.4592.1491.2088.2079.36
✓✓✓✓91.7893.0593.4893.1492.2590.2082.86
", + "bbox": [ + 232, + 143, + 777, + 351 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Table 5: Ablation study of different loss components on the synthetic dataset. Results are grouped according to the loss function(s) directly applied to the descriptor. The best scores within each split are bolded. The database contains all blur levels.", + "bbox": [ + 215, + 353, + 782, + 393 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Compared methods. Since retrieval with object motion blur has not been addressed before, there are no methods specifically designed for this task. Therefore, we compare to state-of-the-art standard retrieval methods: DELG-global [4] (referred to as DELG), DOLG [47], and Token [46]. All these methods are re-implemented as in the original papers, except for changing the size of the output descriptors to 128 for fair comparison. We re-train the compared methods using the same ArcFace parameters and on the same synthetic data as our approach.", + "bbox": [ + 212, + 411, + 784, + 517 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "5.1 Results", + "text_level": 1, + "bbox": [ + 215, + 539, + 323, + 553 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "All blur levels in the database. Our method outperforms all state-of-the-art retrieval methods for all settings, as shown in Table 2 for synthetic data and Table 3 for real data. Our model demonstrates significant superiority over the state-of-the-art retrieval approaches without introducing complex computational modules, highlighting the efficacy of our proposed loss functions. Results on the real-world dataset show that our method generalizes well to real data even when trained solely on synthetic data. Additionally, we train a version of our method where only sharp images are used (Ours-sharp). When evaluating with mAP@all (no distractors), it fails completely, especially for higher blur levels. When evaluating with mAP@100 (with distractors), it achieves reasonable results on query BL 1, but still worse than all other methods. This underscores the necessity for tailored methods and training data with highly motion-blurred objects.", + "bbox": [ + 212, + 566, + 785, + 747 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Per blur level database. In this setup, we utilize datasets specific to each blur level, considering them as individual subsets of images with the designated blur level, to form the database. We then assess the retrieval performance by using query images with distinct blur levels, resulting in a comprehensive evaluation comprising 36 results (6 database BLs by 6 query BLs) for each method. We also compute the standard deviation (std) and range of mAPs for different settings", + "bbox": [ + 212, + 750, + 787, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "R. Zou et al.", + "bbox": [ + 271, + 114, + 357, + 126 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/677a59dd604b3c8083706bea819b2a288c31762fdec4a56e97846a3a19620edc.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Cont. BL RangemAP AllmAP by query BL
123456
S82.2687.3788.4686.2982.6174.1358.26
M83.3387.9789.1286.9383.8776.0460.61
L84.0988.7489.5687.6884.4176.8962.42
L (Alpha)82.8387.6288.6886.4382.9675.6860.27
", + "bbox": [ + 269, + 143, + 738, + 247 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Table 6: Ablation study of different blur level ranges in contrastive learning and design of blur estimation loss on the synthetic dataset (with 1M distractors). Small: range=1. Medium: range=3. Large: range=5. 'Alpha' corresponds to the alternative design choice for blur estimation loss. The database contains images of all blur levels.", + "bbox": [ + 212, + 248, + 787, + 304 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "per method, reflecting the robustness of each method to different levels of motion blur. Our method outperforms all other state-of-the-art methods for all settings and also achieves the smallest standard deviation and range, showing the highest robustness to motion blur (Table 4). See supplementary material for more results.", + "bbox": [ + 212, + 321, + 785, + 382 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Qualitative results on our proposed synthetic (Fig. 4a) and real-world (Fig. 4b) datasets. Our method is compared to the top-performing state-of-the-art retrieval methods on a per-query basis, which yield numerous false positives resembling the texture of the motion-blurred object in the query image. In contrast, our approach accurately retrieves positive images even under severe motion blur in the query and amidst challenging distractors. Note that standard retrieval methods retrieve positive images primarily with blur levels similar to the query, whereas our method retrieves correct images exhibiting a wide range of blur levels (from 1-6 for synthetic and 2-6 for real). This illustrates that our method effectively learns invariant representations of objects across various blur levels.", + "bbox": [ + 212, + 383, + 787, + 537 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "5.2 Ablation study", + "text_level": 1, + "bbox": [ + 215, + 555, + 387, + 571 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Loss components. Table 5 shows an ablation study of different loss functions (for more details, please see the supplementary material). The results show that all loss functions improve model performance. We highlight that our model outperforms the compared methods across all query blur levels (Table 2), even when utilizing only three out of the four loss terms at any combination. For instance, when using $\\mathcal{L}_{con}$ , $\\mathcal{L}_{cls}$ and $\\mathcal{L}_{be}$ , our method surpasses Token [46] by 2.50 in mAP at blur level 1 and 1.91 at blur level 6, with an overall improvement of 2.33.", + "bbox": [ + 212, + 580, + 787, + 686 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Contrastive range. We investigate the research questions posed in Sec. 3.2 using Table 6, which displays results of using different blur level ranges for contrastive learning. A wider range of contrastive blur levels consistently enhances model performance, with the best results achieved at the maximum range of 5.", + "bbox": [ + 212, + 686, + 787, + 748 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Design of blur estimation loss. We validate the effectiveness of our design of blur estimation loss $\\mathcal{L}_{be}$ (Eq. (7)) in Table 6. We compare with the case where edge effects are ignored in the loss function. This corresponds to directly predicting the average of the entire alpha mask (Eq. (4)) inside the object boundary. The results show that our design of $\\mathcal{L}_{be}$ , which effectively mitigates the detrimental impact of edge blur, improves the model's retrieval performance.", + "bbox": [ + 212, + 750, + 787, + 840 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "Retrieval Robust to Object Motion Blur", + "bbox": [ + 462, + 114, + 730, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Query", + "bbox": [ + 225, + 143, + 269, + 157 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/a494f1580db886cd7639ea060b5a2e60dbde5a8ce3938633c8ad5a762bd24a55.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 225, + 157, + 271, + 200 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/0a6a6e9bd054fdcdee6d0354fefddce2a6a13fe04fedecbcfac38357560e592b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 279, + 157, + 323, + 208 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/030d065373cd3d9d6a3dd21c553c054cbc4fcaa8e562192687e92fb1addae8d6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 330, + 157, + 372, + 208 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/d284e9ca6d7e72a065661071cab88716aa8831fc322cb3d3769f3b3d866f5098.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 380, + 157, + 424, + 214 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Top 20 retrieval results", + "bbox": [ + 429, + 145, + 604, + 157 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/36412bd7feaaac7a3590b476d2c376f50079dd10e3a0276277f66f8316b9cfe4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 429, + 157, + 473, + 214 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/88dbe6e78998eaa391e0519e3f6b89c65ebcda9e1535a705fb20122bd5c1d618.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 482, + 157, + 524, + 214 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/2ace5f59f5ec63f43f9639a097d46e54f4e615382eebc7fff23700b5fe2a8ce1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 529, + 157, + 575, + 214 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/08c33429958f42c562b164d7e66c3a852d0a501db5cb388ba76014198c52b14d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 581, + 157, + 624, + 214 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/0a3143d0ac5601b487fd1a8309dd3be9767bdd97766f14db649d2a1de4a9acde.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 630, + 157, + 674, + 214 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/0830a0141f625ab6230ad0efc105f733c93aa43e9c6df55ddffed236642e9fa9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 681, + 157, + 725, + 214 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/1f295044049f0b5aeacdbb99949d71fbf17f3ac65d18bb98d8e53a6376d73486.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 730, + 157, + 776, + 214 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/d8ab84f06d966f6244be744dd91f6d4a04bd3c04b9deb7c0ee1e19ffad06c4ae.jpg", + "image_caption": [ + "Fig. 5: Retrieval results on a real-world video (available on YouTube). The retrieved images are sorted from left to right and top to bottom, with the ranking from $1^{\\text{st}}$ to $20^{\\text{th}}$ . Correct results are in green boxes, while incorrect ones are in red boxes." + ], + "image_footnote": [], + "bbox": [ + 223, + 238, + 271, + 265 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/b80286f23140b80f8044c7c783af99f3ab5a34ac8de1fe4b8f190f75ce0eb272.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 279, + 223, + 323, + 265 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/0d68963c74d90708e528daafc7c2daa80eccdbfd1a522a2bcc49003f14760864.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 330, + 223, + 372, + 265 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/f35be9a90c52019ffcc7a43e44b2e6b56f2ceea2dd4d185935df010b1aac6747.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 380, + 223, + 424, + 265 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/a1bb25c9044d727f08000882b42c0a000a963eb6d2be208d1d5a7fae5c37047a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 429, + 223, + 473, + 265 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/20c1d50cf4b5b96ce631b016d9042af1576cb104f792ec4a1e134f92a898d5b4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 480, + 223, + 524, + 265 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/a3c1e535c949266694e589f54c4c82fd37e0af5e38de22734fbfc9ee95197f58.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 529, + 223, + 575, + 265 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/b6807b3c17d3e03a0a23fc3153449d104f75c95bd7296cba3231380c3d66ba26.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 581, + 223, + 624, + 265 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/b9373c1d0499af4ea1005997eafa23395ffaef94f215889cb0829d2944ee1302.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 630, + 223, + 674, + 265 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/056cb53277149902596142038df0e148175206702c6ef3de25313ef90c936194.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 681, + 223, + 725, + 265 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/648ddb065bd600ef40956baa2cd1b5e85b65713a61d0bbdc3080282e56bbf3d5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 730, + 223, + 776, + 265 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5.3 Application to real-world video data", + "text_level": 1, + "bbox": [ + 215, + 354, + 568, + 369 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "We conduct an experiment to test our method on real-world video images. Due to the challenge of finding a sufficient number of real images of the same object with a well-distributed range of blur levels for quantitative analysis, we construct a small dataset for qualitative assessment. Specifically, we extract 190 images of the same ball from a YouTube soccer video to serve as both the query and database. We incorporate 4,431 images of sports balls from the MSCOCO dataset [21] as distractors, and additionally, we extract 169 images of a different ball from the same video to act as hard distractors. Both the objects and the backgrounds in these images come from domains entirely different from the synthetic training data. Fig. 5 illustrates examples of retrieval results, demonstrating our method's effectiveness in handling various blur conditions and complex and diverse backgrounds in the real world, despite being trained exclusively on synthetic data.", + "bbox": [ + 212, + 378, + 787, + 561 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "6 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 597, + 359, + 612 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In this paper, we introduce an innovative task: image retrieval in the presence of object motion blur. To tackle this challenge, we propose a global retrieval approach that integrates specialized loss functions tailored for the network to better comprehend motion blur. Our method stands as the first specifically designed solution for this new task. Moreover, we introduce the first benchmark datasets for blur retrieval, comprising both synthetic and real-world data. These datasets are meticulously processed and annotated to ensure their immediate applicability for future research in this emerging area. Through extensive experiments conducted on these benchmark datasets, we demonstrate the limitations of state-of-the-art standard retrieval methods in handling motion-blurred images and showcase the superior performance and robustness of our proposed approach in accurately retrieving images even in challenging motion-blurred scenarios.", + "bbox": [ + 212, + 627, + 787, + 808 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgement. We would like to thank Giorgos Tolias for the highly valuable discussions with him and the support he provided for this work.", + "bbox": [ + 214, + 809, + 785, + 839 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "R. Zou et al.", + "bbox": [ + 271, + 114, + 357, + 126 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 143, + 321, + 159 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Babenko, A., Lempitsky, V.: Aggregating local deep features for image retrieval. In: ICCV. pp. 1269-1277 (2015)", + "2. Babenko, A., Slesarev, A., Chigorin, A., Lempitsky, V.: Neural codes for image retrieval. In: ECCV. pp. 584-599. Springer (2014)", + "3. Bay, H., Ess, A., Tuytelaars, T., Van Gool, L.: Speeded-up robust features (surf). Computer Vision and Image Understanding 110(3), 346-359 (2008)", + "4. Cao, B., Araujo, A., Sim, J.: Unifying deep local and global features for image search. In: Vedaldi, A., Bischof, H., Brox, T., Frahm, J.M. (eds.) ECCV. pp. 726-743. Springer International Publishing, Cham (2020)", + "5. Chang, A.X., Funkhouser, T., Guibas, L., Hanrahan, P., Huang, Q., Li, Z., Savarese, S., Savva, M., Song, S., Su, H., Xiao, J., Yi, L., Yu, F.: ShapeNet: An Information-Rich 3D Model Repository. Tech. Rep. arXiv:1512.03012 [cs.GR], Stanford University — Princeton University — Toyota Technological Institute at Chicago (2015)", + "6. Chopra, S., Hadsell, R., LeCun, Y.: Learning a similarity metric discriminatively, with application to face verification. In: CVPR. vol. 1, pp. 539-546. IEEE (2005)", + "7. Community, B.O.: Blender - a 3D modelling and rendering package. Blender Foundation, Stichting Blender Foundation, Amsterdam (2018), http://www.blender.org", + "8. Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: ImageNet: A LargeScale Hierarchical Image Database. In: CVPR (2009)", + "9. Deng, J., Guo, J., Xue, N., Zafeiriou, S.: Arcface: Additive angular margin loss for deep face recognition. In: 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 4685-4694 (2019). https://doi.org/10.1109/CVPR.2019.00482", + "10. Gong, D., Yang, J., Liu, L., Zhang, Y., Reid, I., Shen, C., Van Den Hengel, A., Shi, Q.: From motion blur to motion flow: A deep learning solution for removing heterogeneous motion blur. In: CVPR. pp. 3806-3815 (2017). https://doi.org/10.1109/CVPR.2017.405", + "11. Gordo, A., Almazan, J., Revaud, J., Larlus, D.: Deep image retrieval: Learning global representations for image search (2016)", + "12. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: CVPR. pp. 770-778 (2016)", + "13. Hendrycks, D., Dietterich, T.: Benchmarking neural network robustness to common corruptions and perturbations. Proceedings of the International Conference on Learning Representations (2019)", + "14. Jaesung, R., Haeyun, L., Jucheol, W., Sunghyun, C.: Real-world blur dataset for learning and benchmarking deblurring algorithms. In: ECCV (2020)", + "15. Jégou, H., Perronnin, F., Douze, M., Sánchez, J., Pérez, P., Schmid, C.: Aggregating local image descriptors into compact codes. IEEE TPAMI 34(9), 1704-1716 (2011)", + "16. Kalantidis, Y., Mellina, C., Osindero, S.: Cross-dimensional weighting for aggregated deep convolutional features. In: ECCVW. pp. 685-701. Springer (2016)", + "17. Kingma, D.P., Ba, J.: Adam: A method for stochastic optimization. In: Bengio, Y., LeCun, Y. (eds.) ICLR (2015), http://arxiv.org/abs/1412.6980", + "18. Kotera, J., Rozumnyi, D., Sroubek, F., Matas, J.: Intra-frame object tracking by deblatting. In: ICCVW (Oct 2019)" + ], + "bbox": [ + 218, + 178, + 784, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "Retrieval Robust to Object Motion Blur", + "bbox": [ + 462, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "19. Kupyn, O., Budzan, V., Mykhailych, M., Mishkin, D., Matas, J.: Deblurgan: Blind motion deblurring using conditional adversarial networks. In: CVPR (June 2018)", + "20. Lee, S., Seong, H., Lee, S., Kim, E.: Correlation verification for image retrieval. In: CVPR. pp. 5374-5384 (2022)", + "21. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dólar, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: ECCV. pp. 740-755. Springer (2014)", + "22. Lowe, D.G.: Distinctive image features from scale-invariant keypoints. IJCV 60, 91-110 (2004)", + "23. Michaelis, C., Mitzkus, B., Geirhos, R., Rusak, E., Bringmann, O., Ecker, A.S., Bethge, M., Brendel, W.: Benchmarking robustness in object detection: Autonomous driving when winter is coming. arXiv preprint arXiv:1907.07484 (2019)", + "24. Nah, S., Kim, T.H., Lee, K.M.: Deep multi-scale convolutional neural network for dynamic scene deblurring. In: CVPR (July 2017)", + "25. Noh, H., Araujo, A., Sim, J., Weyand, T., Han, B.: Large-scale image retrieval with attentive deep local features. In: ICCV. pp. 3456-3465 (2017)", + "26. Paszke, A., Gross, S., Massa, F., Lerer, A., Bradbury, J., Chanan, G., Killeen, T., Lin, Z., Gimelshein, N., Antiga, L., Desmaison, A., Kopf, A., Yang, E., Devito, Z., Raison, M., Tejani, A., Chilamkurthy, S., Steiner, B., Fang, L., Bai, J., Chintala, S.: Pytorch: An imperative style, high-performance deep learning library. In: Wallach, H., Larochelle, H., Beygelzimer, A., Alché-Buc, F., Fox, E., Garnett, R. (eds.) NeurIPS, pp. 8024-8035. Curran Associates, Inc. (2019), http://papers.neurips.cc/paper/9015-pytorch-an-imperative-style-high-performance-deep-learning-library.pdf", + "27. Perronnin, F., Liu, Y., Sánchez, J., Poirier, H.: Large-scale image retrieval with compressed fisher vectors. In: CVPR. pp. 3384-3391. IEEE (2010)", + "28. Radenović, F., Iscen, A., Tolias, G., Avrithis, Y., Chum, O.: Revisiting oxford and paris: Large-scale image retrieval benchmarking. In: CVPR. pp. 5706-5715 (2018)", + "29. Radenović, F., Tobias, G., Chum, O.: Fine-tuning cnn image retrieval with no human annotation. arXiv preprint arXiv:1711.02512 (2017)", + "30. Radenović, F., Tolias, G., Chum, O.: Fine-tuning cnn image retrieval with no human annotation. IEEE TPAMI 41(7), 1655-1668 (2019). https://doi.org/10.1109/TPAMI.2018.2846566", + "31. Rozumnyi, D., Kotera, J., Sroubek, F., Matas, J.: Non-causal tracking by deblatting. In: Fink, G.A., Frintrop, S., Jiang, X. (eds.) GCPR. pp. 122-135. Springer International Publishing, Cham (2019)", + "32. Rozumnyi, D., Kotera, J., Sroubek, F., Novotny, L., Matas, J.: The world of fast moving objects. In: CVPR. pp. 4838-4846 (July 2017). https://doi.org/10.1109/CVPR.2017.514", + "33. Rozumnyi, D., Kotera, J., Sroubek, F., Matas, J.: Sub-frame appearance and 6d pose estimation of fast moving objects. In: CVPR. pp. 6777-6785 (2020)", + "34. Rozumnyi, D., Oswald, M.R., Ferrari, V., Matas, J., Pollefeys, M.: Defmo: Deblurring and shape recovery of fast moving objects. In: CVPR. Nashville, Tennessee, USA (Jun 2021)", + "35. Rozumnyi, D., Oswald, M.R., Ferrari, V., Pollefeys, M.: Shape from blur: Recovering textured 3d shape and motion of fast moving objects. In: NeurIPS (2021)", + "36. Rozumnyi, D., Oswald, M.R., Ferrari, V., Pollefeys, M.: Motion-from-blur: 3d shape and motion estimation of motion-blurred objects in videos. In: CVPR (Jun 2022)", + "37. Sayed, M., Brostow, G.: Improved handling of motion blur in online object detection. In: CVPR. pp. 1706-1716 (June 2021)" + ], + "bbox": [ + 215, + 147, + 784, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "R. Zou et al.", + "bbox": [ + 271, + 114, + 357, + 126 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "38. Sivic, Zisserman: Video google: A text retrieval approach to object matching in videos. In: ICCV. pp. 1470-1477. IEEE (2003)", + "39. Skorokhodov, I., Sotnikov, G., Elhoseiny, M.: Aligning latent and image spaces to connect the unconnectable. In: ICCV. pp. 14144-14153 (2021)", + "40. Tolias, G., Avrithis, Y., Jégou, H.: Image search with selective match kernels: aggregation across single and multiple images. IJCV 116, 247-261 (2016)", + "41. Tolias, G., Jenicek, T., Chum, O.: Learning and aggregating deep local descriptors for instance-level recognition. In: ECCV. pp. 460-477. Springer (2020)", + "42. Tolias, G., Sicre, R., Jégou, H.: Particular object retrieval with integral max-pooling of cnn activations. ICLR (2016)", + "43. Vasiljevic, I., Chakrabarti, A., Shakhnarovich, G.: Examining the impact of blur on recognition by convolutional networks. arXiv preprint arXiv:1611.05760 (2016)", + "44. Wang, Y., Lu, Y., Gao, Y., Wang, L., Zhong, Z., Zheng, Y., Yamashita, A.: Efficient video deblurring guided by motion magnitude. In: ECCV. pp. 413-429. Springer (2022)", + "45. Weyand, T., Araujo, A., Cao, B., Sim, J.: Google landmarks dataset v2-a large-scale benchmark for instance-level recognition and retrieval. In: CVPR. pp. 2574-2583 (2020)", + "46. Wu, H., Wang, M., Zhou, W.T., Hu, Y., Li, H.: Learning token-based representation for image retrieval. In: AAAI Conference on Artificial Intelligence (2021), https://apisemantic scholar.org/CorpusID:245124518", + "47. Yang, M., He, D., Fan, M., Shi, B., Xue, X., Li, F., Ding, E., Huang, J.: Dolg: Single-stage image retrieval with deep orthogonal fusion of local and global features. In: ICCV. pp. 11752-11761. IEEE Computer Society, Los Alamitos, CA, USA (oct 2021). https://doi.org/10.1109/ICCV48922.2021.01156, https://doi.ieeecomputersociety.org/10.1109/ICCV48922.2021.01156", + "48. Yokoo, S., Ozaki, K., Simo-Serra, E., Iizuka, S.: Two-stage discriminative reranking for large-scale landmark retrieval. In: CVPRW. IEEE (2020)", + "49. Zhang, H., Xie, H., Yao, H.: Spatio-temporal deformable attention network for video deblurring. In: ECCV. pp. 581-596. Springer (2022)", + "50. Zhou, S., Zhang, J., Pan, J., Xie, H., Zuo, W., Ren, J.: Spatio-temporal filter adaptive network for video deblurring. In: ICCV (2019)" + ], + "bbox": [ + 212, + 146, + 787, + 590 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "Retrieval Robust to Object Motion Blur", + "bbox": [ + 462, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 16 + } +] \ No newline at end of file diff --git a/2024/Retrieval Robust to Object Motion Blur/191ce9d4-edfe-414c-a611-8012902a3458_model.json b/2024/Retrieval Robust to Object Motion Blur/191ce9d4-edfe-414c-a611-8012902a3458_model.json new file mode 100644 index 0000000000000000000000000000000000000000..04787c38e8b259a9158653f71cc48d2bd1624e62 --- /dev/null +++ b/2024/Retrieval Robust to Object Motion Blur/191ce9d4-edfe-414c-a611-8012902a3458_model.json @@ -0,0 +1,3479 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.271, + 0.142, + 0.733, + 0.163 + ], + "angle": 0, + "content": "Retrieval Robust to Object Motion Blur" + }, + { + "type": "text", + "bbox": [ + 0.287, + 0.19, + 0.714, + 0.206 + ], + "angle": 0, + "content": "Rong Zou1, Marc Pollefeys1,2, and Denys Rozumnyi1,3" + }, + { + "type": "text", + "bbox": [ + 0.285, + 0.217, + 0.717, + 0.232 + ], + "angle": 0, + "content": "\\(^{1}\\)ETH Zürich, \\(^{2}\\)Microsoft, \\(^{3}\\)Czech Technical University in Prague" + }, + { + "type": "text", + "bbox": [ + 0.263, + 0.251, + 0.74, + 0.459 + ], + "angle": 0, + "content": "Abstract. Moving objects are frequently seen in daily life and usually appear blurred in images due to their motion. While general object retrieval is a widely explored area in computer vision, it primarily focuses on sharp and static objects, and retrieval of motion-blurred objects in large image collections remains unexplored. We propose a method for object retrieval in images that are affected by motion blur. The proposed method learns a robust representation capable of matching blurred objects to their deblurred versions and vice versa. To evaluate our approach, we present the first large-scale datasets for blurred object retrieval, featuring images with objects exhibiting varying degrees of blur in various poses and scales. We conducted extensive experiments, showing that our method outperforms state-of-the-art retrieval methods on the new blur-retrieval datasets, which validates the effectiveness of the proposed approach. Code, data, and model are available at https://github.com/Rong-Zou/Retrieval-Robust-to-Object-Motion-Blur." + }, + { + "type": "text", + "bbox": [ + 0.263, + 0.471, + 0.661, + 0.485 + ], + "angle": 0, + "content": "Keywords: object retrieval \\(\\cdot\\) object motion blur \\(\\cdot\\) datasets" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.509, + 0.377, + 0.525 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.538, + 0.788, + 0.81 + ], + "angle": 0, + "content": "The task of image retrieval has long been a focal point in computer vision research, aiming to efficiently identify images within a large-scale database that share similarities with a given query image. In the realm of image retrieval, methods are often categorized into category-level and instance-level retrieval. We focus on the latter, where the objective is to identify images of a specific object featured in a query within a vast, unordered collection of database images. In this task, image representations play an important role since they capture the essence of the objects in images and enable the measurement of their similarities. Representing images based on hand-crafted features has been extensively studied [3,22]. Owing to the development in deep learning, recent strides have been made in adopting learning-based approaches to encode image content [2,4,20,25,30,41,42,46,47]. However, the current literature focuses exclusively on image representations for sharp objects, assuming that the objects are stationary and clearly visible in images. By only considering sharp and static objects, the existing retrieval methods fall short in addressing the dynamic nature of real-world scenarios, where objects are in movement, resulting in images blurred by object motion — a phenomenon prevalent in environments with fast-moving objects [32,34], low camera frame rates, and extended exposure times." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.811, + 0.788, + 0.841 + ], + "angle": 0, + "content": "To bridge this gap, we introduce a novel task in image retrieval: retrieval with object motion blur. The challenge of this task lies in finding image representations" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.358, + 0.127 + ], + "angle": 0, + "content": "R. Zou et al." + }, + { + "type": "image_caption", + "bbox": [ + 0.248, + 0.146, + 0.294, + 0.159 + ], + "angle": 0, + "content": "Query" + }, + { + "type": "image", + "bbox": [ + 0.227, + 0.186, + 0.315, + 0.24 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.227, + 0.294, + 0.315, + 0.346 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.318, + 0.158, + 0.406, + 0.211 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.319, + 0.212, + 0.406, + 0.264 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.319, + 0.266, + 0.406, + 0.318 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.319, + 0.319, + 0.406, + 0.371 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.468, + 0.146, + 0.623, + 0.159 + ], + "angle": 0, + "content": "Top 10 retrieval results" + }, + { + "type": "image", + "bbox": [ + 0.411, + 0.159, + 0.498, + 0.211 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.411, + 0.212, + 0.498, + 0.264 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.411, + 0.266, + 0.498, + 0.318 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.411, + 0.319, + 0.498, + 0.371 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.502, + 0.159, + 0.589, + 0.211 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.502, + 0.212, + 0.589, + 0.264 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.502, + 0.266, + 0.589, + 0.318 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.502, + 0.319, + 0.589, + 0.371 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.592, + 0.159, + 0.68, + 0.211 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.592, + 0.212, + 0.68, + 0.264 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.592, + 0.266, + 0.68, + 0.318 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.592, + 0.319, + 0.68, + 0.371 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.684, + 0.159, + 0.772, + 0.211 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.684, + 0.212, + 0.772, + 0.264 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.684, + 0.266, + 0.772, + 0.318 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.684, + 0.319, + 0.772, + 0.371 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.373, + 0.787, + 0.442 + ], + "angle": 0, + "content": "Fig. 1: Retrieval robust to object motion blur. We present the first approach designed for retrieval in the presence of object motion blur. Our method captures blur-invariant features and generates blur-robust representations. We show retrieval results on the newly created datasets for the task of blur-robust retrieval: synthetic (top) and real (bottom). All retrieved images match the query and are marked with green boxes." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.459, + 0.785, + 0.549 + ], + "angle": 0, + "content": "robust to object motion blur, in addition to other variations such as background and changes in viewpoint. We also propose the first approach targeting this task that produces representations invariant to motion blur. In response to the unique challenge posed by object motion, we propose loss functions designed to enhance the model's comprehension of blur, thereby ensuring robust retrieval performance even under conditions of extreme blurring, as seen in Fig. 1." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.551, + 0.785, + 0.64 + ], + "angle": 0, + "content": "Recognizing the absence of datasets accommodating images with motion-blurred objects in existing retrieval benchmarks, we introduce the first benchmark datasets for this task, including a synthetic dataset for training and evaluation and a real-world dataset created specifically for evaluation. By re-training and evaluating state-of-the-art instance retrieval methods on this comprehensive benchmark, we reveal their shortcomings in handling motion-blurred scenarios." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.641, + 0.785, + 0.685 + ], + "angle": 0, + "content": "This work marks a key step in advancing practical methods for object retrieval in the challenging presence of motion blur. It has potential applications such as endangered wildlife monitoring, sports analysis, and security surveillance." + }, + { + "type": "text", + "bbox": [ + 0.24, + 0.686, + 0.623, + 0.7 + ], + "angle": 0, + "content": "To summarize, our main contributions are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.705, + 0.785, + 0.749 + ], + "angle": 0, + "content": "(1) We present the first method specifically designed for the novel retrieval task involving object motion blur. The method is trained with loss functions tailored to improve its understanding of motion blur." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.75, + 0.785, + 0.794 + ], + "angle": 0, + "content": "(2) We introduce a new benchmark featuring synthetic and real-world datasets specifically constructed for this task. The datasets are carefully processed and are directly applicable for future research in blur retrieval." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.795, + 0.785, + 0.84 + ], + "angle": 0, + "content": "(3) We conducted extensive experiments and detailed analyses to validate the effectiveness of our approach. Our method outperforms state-of-the-art retrieval methods and demonstrates superior robustness to motion blur." + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.705, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.463, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Retrieval Robust to Object Motion Blur" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.145, + 0.383, + 0.161 + ], + "angle": 0, + "content": "2 Related work" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.173, + 0.788, + 0.4 + ], + "angle": 0, + "content": "Standard retrieval. In the area of standard image retrieval, local feature-based methods have evolved from traditional hand-crafted techniques like SIFT [22] and SURF [3] to contemporary deep learning-based approaches such as DELF [25] and HOW [41]. Considering that motion blur can corrupt local information, which may make local descriptors less robust, we adopt a global approach for our task. Classical methods typically derive global features by aggregating hand-crafted local features using techniques such as BoW [38], Fisher vectors [27], VLAD [15], AMSK [40], etc. Modern deep learning-based approaches, on the other hand, generate global descriptors by applying various pooling operations on feature maps encoded by convolutional neural networks (CNNs), such as sum pooling [1,41], weighted sum pooling [16], max pooling, regional max pooling [42], and generalized mean-pooling (GeM) [30], which has gained prominence in recent literature [4,20,47]. Our method integrates insights from recent advancements in global feature-based methods, adopting GeM pooling [30] for feature extraction and incorporating contrastive loss [6] and ArcFace loss [9] during training." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.401, + 0.788, + 0.521 + ], + "angle": 0, + "content": "Among the single-pass global feature-based methods, state-of-the-art standard retrieval approaches such as DELG global branch [4], DOLG [47], and Token [46] face challenges when handling images affected by motion blur, leading to degraded performance. Our work stands out in addressing this critical gap in the literature, as existing retrieval methods overlook the impact of motion blur. To bridge this research gap, we propose novel loss functions aimed at enhancing the model's comprehension of object motion blur, thereby contributing to the advancement of image retrieval in motion-blurred scenarios." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.522, + 0.788, + 0.704 + ], + "angle": 0, + "content": "Blur handling. Handling blur in images is an important topic in computer vision. [43] investigated the effect of different types of blur (defocus, uniform subject motion, camera shake) on the performance of CNNs for image classification and semantic segmentation. [37] studied online object detection under the influence of camera ego-motion-induced blur. Image and video deblurring are long-standing tasks in computer vision, which have been addressed by a variety of methods [10,19,44,49,50]. Joint deblurring and trajectory estimation for motion-blurred objects have also been extensively explored, with notable contributions from methods like [18, 31, 33, 34], some extending their scope to 3D reconstruction [35,36]. These existing works predominantly focus on other types of blur or address distinct tasks, emphasizing the need for specialized efforts in the context of our targeted task of retrieval involving object motion blur." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.705, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Datasets. Several datasets have been proposed that contain generic corruptions including blur, such as ImaneNet-C [13] and Coco-C [23], but they do not focus purely on blurry images, and the types of blur they encompass (Gaussian, defocus, frosted glass, camera shake, and zoom) are not induced by object motion. Many datasets contain blurry-sharp image pairs, such as RealBlur [14] and GoPro [24], but they are created for deblurring tasks and not for retrieval, and the blur in images is mainly caused by camera motion or camera shake, not by object motion. A large variety of standard retrieval datasets have been proposed, including Revisited Oxford5k and Paris6k [28], Google Landmarks V2 [45, 48]," + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.358, + 0.127 + ], + "angle": 0, + "content": "R. Zou et al." + }, + { + "type": "image", + "bbox": [ + 0.216, + 0.146, + 0.787, + 0.282 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.284, + 0.789, + 0.466 + ], + "angle": 0, + "content": "Fig. 2: Overview of the proposed method for blur-robust retrieval. Our model takes a query image as input. Then, it generates a descriptor robust to object motion blur through the Blur Robust Image Descriptor Generator (BRIDGE) module that encodes the input image using a CNN backbone followed by generalized mean-pooling (GeM) [30], extracting a feature vector. This vector is subsequently fed into three heads: the blur estimation head, classification head, and localization head. Each head processes the feature vector to extract relevant information for identifying motion-blurred objects. The resulting three processed feature vectors are concatenated and passed through the final fully connected (FC) layer to produce the blur-robust descriptor. During training, the Blur Level-based Image Sample Selection (BLISS) mechanism is employed to select contrastive samples based on the query blur level (BL, Eq. (6)) \\( b_{q} \\) and a specified blur level range \\( r \\). Then, these selected image samples are input into the BRIDGE module to extract descriptors for subsequent contrastive learning." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.492, + 0.788, + 0.555 + ], + "angle": 0, + "content": "SfM120K [29], etc. However, none of them incorporate motion-blurred images. To the best of our knowledge, there is currently no existing dataset tailored for the task of retrieval with object motion blur. In view of this, we created synthetic and real-world datasets specifically designed for the novel task of blur retrieval." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.594, + 0.331, + 0.61 + ], + "angle": 0, + "content": "3 Method" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.629, + 0.787, + 0.719 + ], + "angle": 0, + "content": "We perform retrieval via nearest neighbor search in a representation space. Therefore, we require a representation model that is robust to object motion blur. We train a deep learning-based model that maps images to a vector space. The network is trained on our proposed synthetic dataset using multiple losses, including those tailored to object motion blur. As we show, the method generalizes well to real data despite being trained on synthetic data." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.719, + 0.788, + 0.841 + ], + "angle": 0, + "content": "The network takes an RGB image \\(\\mathbf{I} \\in \\mathbb{R}^{3 \\times H \\times W}\\) containing a sharp or blurred object as input and generates a descriptor \\(\\mathbf{D} \\in \\mathbb{R}^d\\) robust to object motion blur through the Blur Robust Image Descriptor Generator module (BRIDGE, Sec. 3.1). For training only, contrastive samples are selected by the Blur Level-based Image Sample Selection mechanism (BLISS, Sec. 3.2) based on the query blur level \\(b_q\\) and the contrast blur level range \\(r\\). The selected image samples are sent into the BRIDGE module to extract descriptors for contrastive learning. The architecture of the proposed blur retrieval method is shown in Fig. 2." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.462, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Retrieval Robust to Object Motion Blur" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.147, + 0.406, + 0.16 + ], + "angle": 0, + "content": "3.1 BRIDGE module" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.17, + 0.788, + 0.232 + ], + "angle": 0, + "content": "The BRIDGE module encodes an image into a descriptor that is robust to object motion blur. The input image is encoded by a ResNet-50 backbone [12] to produce a feature map \\(\\Omega \\in \\mathbb{R}^{C\\times U\\times V}\\). Then, we use GeM pooling [30] to process the feature map into a feature vector \\(\\mathbf{f}\\in \\mathbb{R}^C\\), expressed as" + }, + { + "type": "equation", + "bbox": [ + 0.359, + 0.24, + 0.786, + 0.299 + ], + "angle": 0, + "content": "\\[\n\\mathbf {f} = \\left[ \\left(\\frac {1}{| \\boldsymbol {\\Omega} |} \\sum_ {(u, v) \\in \\boldsymbol {\\Omega}} x _ {(c, u, v)} ^ {p}\\right) ^ {\\frac {1}{p}} \\right] _ {c = 1, \\dots , C}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.307, + 0.789, + 0.458 + ], + "angle": 0, + "content": "where \\( p > 0 \\) is a parameter, and setting it to values larger than 1 forces the pooled vector to focus on the salient areas. This feature vector then undergoes multi-head processing to generate a motion blur-robust descriptor. Specifically, in the blur estimation head, we obtain blur estimation features \\( \\mathbf{f}_{be} \\in \\mathbb{R}^{C_b} \\) by a fully connected (FC) layer. The same dimension reduction is performed in the localization head, which generates features for location prediction \\( \\mathbf{f}_{loc} \\in \\mathbb{R}^{C_l} \\). The features generated from the two heads are then concatenated with the whitened [11] features \\( \\mathbf{f}_{cls} \\in \\mathbb{R}^{C_c} \\) from the classification head. We further reduce the dimension of the concatenated features to obtain the final blur-robust descriptor of the image. This can be expressed as:" + }, + { + "type": "equation", + "bbox": [ + 0.361, + 0.468, + 0.786, + 0.485 + ], + "angle": 0, + "content": "\\[\n\\mathbf {D} = \\mathbf {W} \\cdot \\operatorname {C o n c a t} \\left(\\mathbf {W} _ {b e} \\mathbf {f}, \\mathbf {W} _ {l o c} \\mathbf {f}, \\mathbf {W} _ {c l s} \\mathbf {f}\\right), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.492, + 0.788, + 0.541 + ], + "angle": 0, + "content": "where \\(\\mathbf{W}_{be} \\in \\mathbb{R}^{C_b \\times C}\\), \\(\\mathbf{W}_{loc} \\in \\mathbb{R}^{C_l \\times C}\\) and \\(\\mathbf{W}_{cls} \\in \\mathbb{R}^{C_c \\times C}\\) are weights of the FC layers in the blur estimation, localization, and classification heads, respectively, and \\(\\mathbf{W} \\in \\mathbb{R}^{d \\times (C_b + C_l + C_c)}\\) is the weight of the final dimension-reduction layer." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.56, + 0.414, + 0.574 + ], + "angle": 0, + "content": "3.2 BLISS mechanism" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.584, + 0.789, + 0.844 + ], + "angle": 0, + "content": "While the specific representation of the degree of motion blur for objects in images is yet to be formally defined, it is intuitively perceived as a continuous value. The level of detail exhibited by objects varies with the degree of blur: sharper objects show clear structures, textures, and poses, while more blurred objects primarily reveal color characteristics (see Fig. 3). This continuous nature of blur allows us to indirectly control the distance between samples with significantly different blur conditions during learning image representations. In this context, we manage the distance between samples that exhibit visual features and patterns of the same level of detail due to similar degrees of blur, thus indirectly influencing the distance between samples with significantly different levels of blur. Additionally, we can directly control the distance between samples with large differences in blur conditions by imposing constraints directly between such samples in the descriptor space. There are therefore two key research questions: What is the optimal value for the range of differences in the blur levels of image pairs used in contrastive learning? How does adjusting the range of differences in blur levels affect the model's performance? To answer these questions and identify the optimal blur level range for contrastive learning, we introduce the BLISS" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.358, + 0.127 + ], + "angle": 0, + "content": "R. Zou et al." + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.145, + 0.26, + 0.156 + ], + "angle": 0, + "content": "BL 1" + }, + { + "type": "image", + "bbox": [ + 0.22, + 0.158, + 0.266, + 0.213 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.22, + 0.213, + 0.265, + 0.268 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.221, + 0.269, + 0.265, + 0.324 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.272, + 0.146, + 0.308, + 0.156 + ], + "angle": 0, + "content": "BL2" + }, + { + "type": "image", + "bbox": [ + 0.268, + 0.157, + 0.313, + 0.213 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.268, + 0.213, + 0.313, + 0.268 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.268, + 0.268, + 0.313, + 0.324 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.285, + 0.325, + 0.426, + 0.336 + ], + "angle": 0, + "content": "(a) Synthetic dataset." + }, + { + "type": "text", + "bbox": [ + 0.322, + 0.146, + 0.358, + 0.156 + ], + "angle": 0, + "content": "BL 3" + }, + { + "type": "image", + "bbox": [ + 0.315, + 0.157, + 0.359, + 0.213 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.315, + 0.213, + 0.359, + 0.268 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.315, + 0.268, + 0.359, + 0.324 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.315, + 0.325, + 0.432, + 0.336 + ], + "angle": 0, + "content": "Synthetic dataset." + }, + { + "type": "text", + "bbox": [ + 0.393, + 0.146, + 0.428, + 0.156 + ], + "angle": 0, + "content": "4 B" + }, + { + "type": "image", + "bbox": [ + 0.363, + 0.157, + 0.407, + 0.213 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.362, + 0.213, + 0.406, + 0.268 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.362, + 0.268, + 0.406, + 0.324 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.362, + 0.325, + 0.406, + 0.336 + ], + "angle": 0, + "content": "ic datas" + }, + { + "type": "text", + "bbox": [ + 0.413, + 0.146, + 0.45, + 0.156 + ], + "angle": 0, + "content": "BL 5" + }, + { + "type": "image", + "bbox": [ + 0.411, + 0.157, + 0.454, + 0.213 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.411, + 0.213, + 0.454, + 0.268 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.411, + 0.268, + 0.454, + 0.324 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.411, + 0.325, + 0.501, + 0.336 + ], + "angle": 0, + "content": "et." + }, + { + "type": "text", + "bbox": [ + 0.462, + 0.146, + 0.498, + 0.156 + ], + "angle": 0, + "content": "BL6" + }, + { + "type": "image", + "bbox": [ + 0.46, + 0.157, + 0.502, + 0.213 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.46, + 0.213, + 0.502, + 0.268 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.46, + 0.268, + 0.502, + 0.324 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.46, + 0.325, + 0.502, + 0.336 + ], + "angle": 0, + "content": "" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.146, + 0.553, + 0.156 + ], + "angle": 0, + "content": "BLr1" + }, + { + "type": "image", + "bbox": [ + 0.51, + 0.157, + 0.554, + 0.213 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.51, + 0.213, + 0.554, + 0.268 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.51, + 0.268, + 0.554, + 0.296 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.51, + 0.296, + 0.554, + 0.324 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.51, + 0.325, + 0.554, + 0.336 + ], + "angle": 0, + "content": "" + }, + { + "type": "text", + "bbox": [ + 0.555, + 0.146, + 0.599, + 0.156 + ], + "angle": 0, + "content": "BL 2" + }, + { + "type": "image", + "bbox": [ + 0.555, + 0.157, + 0.6, + 0.213 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.555, + 0.213, + 0.6, + 0.268 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.555, + 0.268, + 0.6, + 0.296 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.555, + 0.296, + 0.6, + 0.324 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.555, + 0.325, + 0.6, + 0.336 + ], + "angle": 0, + "content": "(b) F" + }, + { + "type": "text", + "bbox": [ + 0.601, + 0.146, + 0.646, + 0.156 + ], + "angle": 0, + "content": "BLr3" + }, + { + "type": "image", + "bbox": [ + 0.601, + 0.157, + 0.646, + 0.213 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.601, + 0.213, + 0.646, + 0.268 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.601, + 0.268, + 0.646, + 0.296 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.601, + 0.296, + 0.646, + 0.324 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.601, + 0.325, + 0.646, + 0.336 + ], + "angle": 0, + "content": "Real-wo" + }, + { + "type": "text", + "bbox": [ + 0.653, + 0.146, + 0.697, + 0.156 + ], + "angle": 0, + "content": "BLr 4" + }, + { + "type": "image", + "bbox": [ + 0.653, + 0.157, + 0.697, + 0.213 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.653, + 0.213, + 0.697, + 0.268 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.653, + 0.268, + 0.697, + 0.296 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.653, + 0.296, + 0.697, + 0.324 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.653, + 0.325, + 0.697, + 0.336 + ], + "angle": 0, + "content": "ld data" + }, + { + "type": "text", + "bbox": [ + 0.7, + 0.146, + 0.744, + 0.156 + ], + "angle": 0, + "content": "BL 5" + }, + { + "type": "image", + "bbox": [ + 0.7, + 0.157, + 0.744, + 0.213 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.7, + 0.213, + 0.744, + 0.268 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.7, + 0.268, + 0.744, + 0.296 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.7, + 0.296, + 0.744, + 0.324 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.7, + 0.325, + 0.744, + 0.336 + ], + "angle": 0, + "content": "set." + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.339, + 0.787, + 0.409 + ], + "angle": 0, + "content": "Fig. 3: The introduced synthetic (a) and real-world (b) datasets. Rows 1-2: different trajectories of the same object. Rows 2-3: two different objects from the same category with similar shapes. Rows 4-6: objects from different categories but share similar textures. Columns correspond to different blur levels, from 1 to 6, and they are different for synthetic (BL) and real-world \\((\\mathrm{BL}^{\\mathrm{r}})\\) datasets." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.424, + 0.785, + 0.454 + ], + "angle": 0, + "content": "mechanism. This mechanism selects contrastive samples based on the blur level in the query image for contrastive learning." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.455, + 0.785, + 0.502 + ], + "angle": 0, + "content": "Blur severity definition. Our definition of blur severity is inspired by the image generation process. The formation of an image \\(\\mathbf{I}\\) containing a motion-blurred object in front of a background can be expressed [32] as" + }, + { + "type": "equation", + "bbox": [ + 0.402, + 0.51, + 0.785, + 0.527 + ], + "angle": 0, + "content": "\\[\n\\mathbf {I} = \\mathbf {P} * \\mathbf {O} + (1 - \\mathbf {P} * \\mathbf {M}) \\mathbf {B} \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.535, + 0.785, + 0.595 + ], + "angle": 0, + "content": "where the first term denotes the convolution \\((\\ast)\\) of a point spread function \\(\\mathbf{P}\\) corresponding to a motion trajectory and the object's sharp appearance \\(\\mathbf{O}\\), and the second term depicts the contribution of the background \\(\\mathbf{B}\\), in which \\(\\mathbf{M}\\) is the binary mask of \\(\\mathbf{O}\\). Let" + }, + { + "type": "equation", + "bbox": [ + 0.456, + 0.595, + 0.785, + 0.612 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {\\alpha} = \\mathbf {P} * \\mathbf {M}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.615, + 0.787, + 0.723 + ], + "angle": 0, + "content": "\\(\\pmb{\\alpha} \\in \\mathbb{R}^{H \\times W}\\), \\(\\pmb{\\alpha}(i,j) \\in [0,1], \\forall (i,j) \\in \\mathbf{I}\\). Considering only objects with opaque materials, \\(\\pmb{\\alpha}\\) represents the visibility of the object. In other words, it can characterize the amount of blur. Generally, the sharper the object in the image, the larger the \\(\\pmb{\\alpha}\\) values are. However, at the edge pixels, the object blends with the background, and the alpha values of these edge pixels gradually fade to 0. Combining the alpha mask from Eq. (4) and considering edge effects, we define the Blur Severity (BS) of an image with a motion-blurred object as:" + }, + { + "type": "equation", + "bbox": [ + 0.364, + 0.729, + 0.785, + 0.773 + ], + "angle": 0, + "content": "\\[\nB S = 1 - \\frac {\\sum_ {i = 1} ^ {H} \\sum_ {j = 1} ^ {W} \\boldsymbol {\\beta} (i , j) \\cdot \\boldsymbol {\\alpha} (i , j)}{\\sum_ {i = 1} ^ {H} \\sum_ {j = 1} ^ {W} \\boldsymbol {\\beta} (i , j)}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.787, + 0.841 + ], + "angle": 0, + "content": "where \\(\\beta\\) is a thresholded (by 0) and eroded (by 3 pixels) binary mask of \\(\\alpha\\). Erosion effectively removes the adverse contribution of the edge pixels to the blur severity measure. As a result, \\(BS \\in (0,1)\\), and the larger the value, the more severe the motion blur in the image." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.462, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "Retrieval Robust to Object Motion Blur" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "table", + "bbox": [ + 0.221, + 0.144, + 0.788, + 0.248 + ], + "angle": 0, + "content": "
Dataset#Images Total#Images by BL (Syn.) or BLr (Real)
123456
Syn. Q20,9954,2883,9324,0784,0892,9301,678
Syn. DB91,62118,87117,50817,88818,02912,54667,79
Syn. Dist.1,091,939214,364177,869222,542235,662149,82891,674
Real Q2,753612620561396315249
Real DB10,3401,9231,8032,0801,7451,3751,414
" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.25, + 0.784, + 0.276 + ], + "angle": 0, + "content": "Table 1: Distribution of images in our synthetic and real datasets, sorted by blur levels. 'Q' stands for 'query', 'DB' denotes database, and 'Dist.' represents distractors." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.296, + 0.784, + 0.342 + ], + "angle": 0, + "content": "Blur level-based sample selection. Since blur severity is a continuous quantity, finding images within a certain blur severity range for each query image is expensive. We thus use a discretized version of blur severity — Blur Level (BL):" + }, + { + "type": "equation", + "bbox": [ + 0.449, + 0.357, + 0.785, + 0.373 + ], + "angle": 0, + "content": "\\[\nB L = \\lceil 1 0 B S \\rceil , \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.376, + 0.784, + 0.438 + ], + "angle": 0, + "content": "where \\( \\lceil \\cdot \\rceil \\) denotes ceiling operation. Each image is assigned to a \\( BL \\) according to its \\( BS \\). During training, we select contrastive samples based on the query blur level \\( b_{q} \\) and the specified contrastive \\( BL \\) range \\( r \\). The selected \\( N_{p} \\) positive and \\( N_{n} \\) negative samples thus have blur levels within the range of \\( b_{q} - r \\) to \\( b_{q} + r \\)." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.455, + 0.42, + 0.47 + ], + "angle": 0, + "content": "3.3 Training objectives" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.477, + 0.784, + 0.55 + ], + "angle": 0, + "content": "Blur estimation loss. Blur severity contains information about the visibility of the blurred object in an image (Sec. 3.2). From another perspective, it reflects a signal-to-noise ratio (SNR), since the more severe the motion blur, the greater the ratio of background (noise) to foreground (signal). In view of this, we define blur estimation loss as" + }, + { + "type": "equation", + "bbox": [ + 0.424, + 0.553, + 0.785, + 0.569 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {b e} = | p - (1 - B S) |, \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.572, + 0.784, + 0.632 + ], + "angle": 0, + "content": "where \\( p \\in \\mathbb{R} \\) is the blur value predicted by the blur estimation head. By constraining the estimate of the continuous \\( BS \\) (Eq. (5)), we implicitly force the network to identify blur and estimate the SNR, which helps it to focus on and extract signal information in the blurred region while ignoring background noises." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.634, + 0.527, + 0.647 + ], + "angle": 0, + "content": "Localization loss. This loss is defined as" + }, + { + "type": "equation", + "bbox": [ + 0.347, + 0.656, + 0.785, + 0.672 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {l o c} = | x - \\hat {x} | + | y - \\hat {y} | + | h - \\hat {h} | + | w - \\hat {w} |, \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.679, + 0.784, + 0.755 + ], + "angle": 0, + "content": "where \\((x,y,w,h)\\) and \\((\\hat{x},\\hat{y},\\hat{w},\\hat{h})\\) represent the normalized ground truth bounding box and the predicted one from the localization head, respectively. Localization loss requires the network to distinguish motion-blurred objects from backgrounds under various blur conditions, forcing it to recognize blur, identify objects at different blur levels, and further, focus more on the foreground objects." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.756, + 0.784, + 0.787 + ], + "angle": 0, + "content": "Classification loss. Following DELG [4], we apply ArcFace margin loss [9] to the descriptor \\(\\mathbf{D}\\), expressed as" + }, + { + "type": "equation", + "bbox": [ + 0.334, + 0.794, + 0.785, + 0.843 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {c l s} = - \\log \\left(\\frac {\\exp (\\gamma \\times A F (\\hat {\\mathbf {w}} _ {t} ^ {T} \\hat {\\mathbf {D}} , 1))}{\\sum_ {n} \\exp (\\gamma \\times A F (\\hat {\\mathbf {w}} _ {n} ^ {T} \\hat {\\mathbf {D}} , y _ {n}))}\\right), \\tag {9}\n\\]" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.358, + 0.127 + ], + "angle": 0, + "content": "R. Zou et al." + }, + { + "type": "table", + "bbox": [ + 0.22, + 0.145, + 0.788, + 0.219 + ], + "angle": 0, + "content": "
MethodmAP by query BL (synthetic)mAP by query BL (synthetic + 1M)
All123456All123456
DELG [4]81.6583.5584.9684.8982.6977.3566.2068.1973.6475.4073.3468.0558.2842.46
DOLG [47]83.0385.1086.2685.9384.0478.7868.0769.9775.7577.4775.0170.1060.0142.49
Token [46]84.8486.5387.9987.5085.8180.7071.5770.6575.3277.6675.5170.2461.1948.05
Ours-sharp27.5743.0638.5329.7419.4010.876.1432.6471.9343.8827.1815.417.944.27
Ours91.7893.0593.4893.1492.2590.2082.8684.0988.7489.5687.6884.4176.8962.42
" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.221, + 0.787, + 0.276 + ], + "angle": 0, + "content": "Table 2: Retrieval results on the synthetic dataset with (+1M) and without distractors. Ours-sharp is our model trained with only sharp images. 'All' denotes the overall performance for queries of all blur levels. All methods are re-implemented and re-trained on our synthetic training set. The database contains images of all blur levels." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.296, + 0.784, + 0.356 + ], + "angle": 0, + "content": "where \\(\\gamma\\) is a scale factor, \\(\\hat{\\mathbf{w}}_{\\mathbf{i}}\\) refers to the \\(i\\)-th row of the \\(L_{2}\\) normalized \\(N\\)-class classifier \\(\\hat{\\mathbf{W}} \\in \\mathbb{R}^{d \\times N}\\), \\(\\hat{\\mathbf{D}}\\) denotes the \\(L_{2}\\) normalized \\(\\mathbf{D}\\). \\(y\\) is the one-hot label, and \\(t\\) represents the ground truth class index (\\(y_{t} = 1\\)). \\(AF\\) is the ArcFace-adjusted cosine similarity and it can be expressed as" + }, + { + "type": "equation", + "bbox": [ + 0.347, + 0.365, + 0.786, + 0.403 + ], + "angle": 0, + "content": "\\[\nA F (s, g) = \\left\\{ \\begin{array}{l l} \\cos (\\operatorname {a r c c o s} (s) + m), & \\text {i f} g = 1 \\\\ s, & \\text {i f} g = 0 \\end{array} , \\right. \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.411, + 0.784, + 0.485 + ], + "angle": 0, + "content": "where \\( s \\) is the cosine similarity, \\( m \\) is the ArcFace margin, and \\( g \\in \\{0,1\\} \\). \\( g = 1 \\) indicates the ground truth class. The classification loss requires the model to group sharp or motion-blurred images of the same object together, while distinguishing between different objects, which may appear similar, especially when blurred. This implicitly forces the model to focus on blur-invariant features." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.488, + 0.784, + 0.548 + ], + "angle": 0, + "content": "Contrastive loss. We employ contrastive learning [6] to construct a descriptor space, where the distance between similar samples is minimized, and between dissimilar samples maximized. The contrastive loss is applied to (query, positive) and (query, negative) descriptor pairs:" + }, + { + "type": "equation", + "bbox": [ + 0.314, + 0.558, + 0.785, + 0.596 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {c o n}} (i, j) = \\left\\{ \\begin{array}{l l} 0. 5 \\cdot D _ {i, j} ^ {2}, & \\text {i f} y (i, j) = 1 \\\\ 0. 5 \\left(\\max \\{0, \\tau - D _ {i, j} \\}\\right) ^ {2}, & \\text {i f} y (i, j) = 0 \\end{array} , \\right. \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.605, + 0.784, + 0.681 + ], + "angle": 0, + "content": "where \\( D_{i,j} = ||\\hat{\\mathbf{D}}(i) - \\hat{\\mathbf{D}}(j)|| \\) and \\( \\hat{\\mathbf{D}}(i) \\) is the \\( L_2 \\) normalized descriptor of image \\( i \\). \\( \\tau \\) is a margin parameter, and \\( y(i,j) \\) is binary indicating whether the pair matches \\( (= 1) \\). The contrastive loss further enhances the focus on blur-invariant features by explicitly forcing descriptors of the same object, whether sharp or blurred, to be close together, and those of different objects to be far apart." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.683, + 0.767, + 0.698 + ], + "angle": 0, + "content": "Joint loss. The joint loss is a weighted sum of all previously defined losses:" + }, + { + "type": "equation", + "bbox": [ + 0.343, + 0.708, + 0.785, + 0.723 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {j o i n t}} = \\mathcal {L} _ {\\text {c o n}} + \\alpha_ {\\text {c l s}} \\mathcal {L} _ {\\text {c l s}} + \\alpha_ {\\text {b e}} \\mathcal {L} _ {\\text {b e}} + \\alpha_ {\\text {l o c}} \\mathcal {L} _ {\\text {l o c}}. \\tag {12}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.766, + 0.329, + 0.781 + ], + "angle": 0, + "content": "4 Dataset" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.784, + 0.84 + ], + "angle": 0, + "content": "We create novel synthetic and real datasets for training and evaluating retrieval methods for motion-blurred objects. The synthetic dataset serves for both training and evaluation. We also generate a large-scale synthetic distractor set to" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.463, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Retrieval Robust to Object Motion Blur" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "table", + "bbox": [ + 0.222, + 0.144, + 0.785, + 0.254 + ], + "angle": 0, + "content": "
MethodmAP AllmAP by Query BLr
123456
DELG [4]54.8249.1363.4357.2555.0153.7742.92
DOLG [47]54.6443.9360.5958.3659.0658.5845.78
Token [46]43.3338.7147.0850.7946.4442.7124.43
Ours-sharp40.2449.5545.0241.3333.2329.4027.91
Ours62.8857.5070.3866.7763.1864.4846.14
" + }, + { + "type": "table_caption", + "bbox": [ + 0.218, + 0.255, + 0.784, + 0.281 + ], + "angle": 0, + "content": "Table 3: Retrieval results on the real-world dataset (trained on synthetic without fine-tuning). The database includes all blur levels. Our method outperforms all others." + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.299, + 0.784, + 0.358 + ], + "angle": 0, + "content": "evaluate methods in a more challenging setup. The novel real-world dataset is specifically constructed to validate the methods' effectiveness in real scenarios. All datasets are meticulously processed for the blur retrieval task, laying a crucial foundation for future research. Statistics of these datasets are shown in Table 1." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.383, + 0.404, + 0.397 + ], + "angle": 0, + "content": "4.1 Synthetic dataset" + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.409, + 0.784, + 0.452 + ], + "angle": 0, + "content": "The images in this dataset contain 3D objects moving along different trajectories, in front of different backgrounds, and with different amounts of motion blur. Thus, this dataset poses many challenges to retrieval methods." + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.455, + 0.784, + 0.484 + ], + "angle": 0, + "content": "Backgrounds. Background images are sampled from the LHQ dataset [39], cropped to a resolution of \\(240 \\times 320\\)." + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.486, + 0.784, + 0.516 + ], + "angle": 0, + "content": "Objects. We select 39 categories with the largest number of textured objects from the ShapeNet dataset [5] and randomly sample 30 objects in each category." + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.518, + 0.784, + 0.562 + ], + "angle": 0, + "content": "Trajectories. For each object, we generate 120 random linear trajectories, each divided into 23 sub-trajectories. We then select up to 10 consecutive sub-trajectories to create images featuring objects with diverse levels of motion blur." + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.563, + 0.784, + 0.609 + ], + "angle": 0, + "content": "Rendering. We render blurry images with Blender Cycles [7] by capturing moving objects during the camera exposure time. Additionally, we render one sharp image per trajectory, placed at the central position along the trajectory." + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.611, + 0.784, + 0.671 + ], + "angle": 0, + "content": "Filtering. We discard renderings where the object occupies less than \\(1.5\\%\\) of the pixels or where the intersection-over-union between silhouette renderings at trajectory endpoints is below \\(20\\%\\). The former condition suggests the object is too small, while the latter indicates the appearance changes are too drastic." + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.673, + 0.784, + 0.718 + ], + "angle": 0, + "content": "Ground truth. For each generated image, we estimate the bounding box coordinates by detecting the non-zero pixels in the alpha channel. We calculate the blur level for each image based on Eq. (6)." + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.72, + 0.784, + 0.84 + ], + "angle": 0, + "content": "Retrieval dataset. In total, we rendered 1.5 million images for 1138 different instances (examples in Fig. 3a). Additionally, we create a challenging distractor set with over 1 million images (examples in the supplementary material), using the same object categories as before to increase the difficulty of the distractors in terms of intra-class similarity. We randomly select 40 objects per category, excluding the 30 previously chosen objects, and generate 70 trajectories for each. For each trajectory, we render one sharp image and 10 blurry images with varying degrees of blur. Backgrounds are also selected from a different subset of [39]." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.358, + 0.127 + ], + "angle": 0, + "content": "R. Zou et al." + }, + { + "type": "table", + "bbox": [ + 0.22, + 0.144, + 0.788, + 0.346 + ], + "angle": 0, + "content": "
BLQDDELG [4] (range: 57.79; std: 16.75)DOLG [47] (range: 56.81; std: 16.93)
All123456All123456
175.9986.7384.1780.2672.4863.1250.0877.7887.0185.7081.8575.1166.2052.39
277.3582.7486.2282.8776.6866.4650.0078.4484.1887.0083.4377.9167.8851.31
377.1979.6283.4382.9478.4669.4352.8678.4981.8484.7483.8579.6970.5853.12
474.9674.2678.9179.8578.0570.8955.1276.4077.2780.5580.8079.1071.7355.28
562.7157.8661.7264.3765.3668.9656.0564.1961.3163.6865.3866.0569.6855.75
634.8929.5628.9431.7534.0645.3753.8635.5131.2430.2032.2934.1845.0053.41
BLQDToken [46] (range: 55.72; std: 16.11)Ours (range: 53.68; std: 15.65)
All123456All123456
179.6588.2287.3483.5076.8768.5156.6388.5795.0994.2491.5487.1980.9168.15
279.9784.7888.4884.7778.9970.0155.8288.7593.0794.6592.1888.7782.0367.18
379.5081.7985.4684.7580.2172.0958.1888.7691.7093.1892.3289.8383.4168.94
476.8276.5680.4281.2479.3172.6159.6187.1688.1290.1190.0989.3184.3770.35
565.3961.8564.4966.3167.1470.9660.3676.2873.8774.9076.0477.3983.0671.67
637.9833.2332.7634.6136.5047.5657.4347.3143.4341.4142.8944.5558.1369.59
" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.348, + 0.785, + 0.402 + ], + "angle": 0, + "content": "Table 4: Retrieval results by query (Q) and database (D) blur level (BL) on our synthetic dataset (with 1M distractors). 'All' denotes the overall performance for queries of all blur levels. The standard deviation and range of mAP values are displayed alongside each method. The best scores under the same settings are marked in bold." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.418, + 0.418, + 0.432 + ], + "angle": 0, + "content": "4.2 Real-world dataset" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.441, + 0.785, + 0.546 + ], + "angle": 0, + "content": "We record videos with a GoPro 7 Hero camera at 240 fps with full exposure. We use 35 different objects and move each object 3 to 5 times along different trajectories in front of the camera. None of the real-world objects are in synthetic data and \\(2/3\\) of which are from new categories not present in synthetic data. This better tests the generalization ability of models to new objects and categories. In total, we recorded 139 videos. Different amounts of blur are generated by averaging consecutive frames, ranging from just one up to 200 averaged frames." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.547, + 0.787, + 0.652 + ], + "angle": 0, + "content": "Computing blur level based on Eq. (6) is not feasible for the real-world dataset. Therefore, we manually annotate the data, assigning each image to one of 6 blur levels based on the perceived blur, and we use the ratio of the trajectory length to the average object size as a reference. The annotation enables a more detailed analysis and a more comprehensive evaluation of model performance. Further annotation details can be found in the supplementary material. Since the blur level for real data differs from that in Eq. (6), we denote it as \\(\\mathrm{BL}^{\\mathrm{r}}\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.653, + 0.787, + 0.743 + ], + "angle": 0, + "content": "The real-world dataset contains 13,093 images, with an average of 374 images for each object (examples in Fig. 3b). Compared to the synthetic dataset, the object scale change is more drastic in different trajectories, e.g. rows 1 & 2. The trajectories are notably different from those in the synthetic dataset. They involve complex curved trajectories with acceleration and other physical forces, which poses an additional challenge for methods trained on our synthetic dataset." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.764, + 0.376, + 0.781 + ], + "angle": 0, + "content": "5 Experiments" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.787, + 0.842 + ], + "angle": 0, + "content": "Setup. For each category of objects in the synthetic dataset, we randomly take \\(70\\%\\) (rounded down, same for validation) for training (792 objects), \\(15\\%\\) for validation (153 objects), and \\(15\\%\\) for testing (193 objects). For the training objects," + } + ], + [ + { + "type": "header", + "bbox": [ + 0.463, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Retrieval Robust to Object Motion Blur" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "image_caption", + "bbox": [ + 0.227, + 0.145, + 0.273, + 0.159 + ], + "angle": 0, + "content": "Query" + }, + { + "type": "image", + "bbox": [ + 0.226, + 0.159, + 0.785, + 0.277 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.258, + 0.277, + 0.746, + 0.289 + ], + "angle": 0, + "content": "(a) Example of retrieval results on our synthetic dataset (with distractors)." + }, + { + "type": "image", + "bbox": [ + 0.226, + 0.291, + 0.784, + 0.411 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.312, + 0.412, + 0.692, + 0.424 + ], + "angle": 0, + "content": "(b) Example of retrieval results on our real-world dataset." + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.428, + 0.788, + 0.497 + ], + "angle": 0, + "content": "Fig. 4: Comparison of retrieval results. We compare to the best-performing state-of-the-art method for the corresponding query. The retrieved images are sorted from left to right and top to bottom, with the ranking from \\(1^{\\mathrm{st}}\\) to \\(20^{\\mathrm{th}}\\). The blur level of each image is shown in the bottom left corner. Correct results are in green boxes, while incorrect ones are in red boxes. We provide more results in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.518, + 0.788, + 0.639 + ], + "angle": 0, + "content": "we take one image for each trajectory while ensuring a balanced distribution of images across different blur levels. In total, there are 95k images in the training set. For each of the test objects, we sample 20 trajectories to form a query set, and the rest is used for the database. The database images are sampled from all blur levels unless specified otherwise (for per blur level database experiments). We additionally show performance on a subset of queries for a given blur level. In all experiments (except Ours-sharp), only one model is trained for all blur levels, and the evaluation is performed for varying blur levels." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.641, + 0.788, + 0.807 + ], + "angle": 0, + "content": "Implementation details. The proposed method is implemented in PyTorch [26]. We use ResNet-50 [12] up to the last downsampling block as the backbone, deactivate the normalization layers, and initialize the model with ImageNet [8] pre-trained weights. ADAM [17] optimizer is employed to minimize the joint loss with a fixed learning rate \\( 10^{-4} \\), batch size 32, and 30 epochs. We initialize the power parameter to 3 for GeM pooling [30]. The contrastive margin \\( \\tau \\) is set to 0.7. Following [30], we set the number of contrastive samples as \\( N_{p} = 1 \\), \\( N_{n} = 5 \\). The contrastive blur range is experimentally set to 5. For classification, we empirically set the margin \\( m \\) to 0.15 and the scale \\( \\gamma \\) to 30. Loss weights are chosen so that the values of different losses are roughly of the same order of magnitude, i.e. \\( \\alpha_{cls} = 0.1 \\), \\( \\alpha_{be} = 1.0 \\), \\( \\alpha_{loc} = 10 \\). The descriptor dimension \\( d \\) is set to 128." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Evaluation protocols. We report mean average precision (mAP@all) as commonly used in retrieval. In case 1M distractors are used, we report mAP@100." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.358, + 0.127 + ], + "angle": 0, + "content": "R. Zou et al." + }, + { + "type": "table", + "bbox": [ + 0.233, + 0.144, + 0.778, + 0.352 + ], + "angle": 0, + "content": "
\\( \\mathcal{L}_{con} \\)\\( \\mathcal{L}_{cls} \\)\\( \\mathcal{L}_{be} \\)\\( \\mathcal{L}_{loc} \\)mAP AllmAP by Query BL
123456
✓✘✘✘78.1380.5181.7081.1679.2073.9361.24
✓✘✓✘81.6683.4985.0184.4382.6977.6467.08
✓✘✘✓85.9487.5488.2587.8386.5283.0875.58
✓✘✓✓87.4888.6989.8989.4088.2484.9776.74
✘✓✘✘78.7381.5382.9782.8379.8672.9359.00
✘✓✓✘83.6785.1987.5687.4085.1079.2265.93
✘✓✘✓88.7489.8990.9190.8889.7586.0777.67
✘✓✓✓91.2392.0293.1693.0991.9789.0082.27
✓✓✘✘85.0687.4288.2987.6685.8581.2069.96
✓✓✓✘87.1789.0390.0389.5588.0783.9173.48
✓✓✘✓90.3991.8592.4592.1491.2088.2079.36
✓✓✓✓91.7893.0593.4893.1492.2590.2082.86
" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.354, + 0.784, + 0.395 + ], + "angle": 0, + "content": "Table 5: Ablation study of different loss components on the synthetic dataset. Results are grouped according to the loss function(s) directly applied to the descriptor. The best scores within each split are bolded. The database contains all blur levels." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.412, + 0.785, + 0.518 + ], + "angle": 0, + "content": "Compared methods. Since retrieval with object motion blur has not been addressed before, there are no methods specifically designed for this task. Therefore, we compare to state-of-the-art standard retrieval methods: DELG-global [4] (referred to as DELG), DOLG [47], and Token [46]. All these methods are re-implemented as in the original papers, except for changing the size of the output descriptors to 128 for fair comparison. We re-train the compared methods using the same ArcFace parameters and on the same synthetic data as our approach." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.54, + 0.324, + 0.554 + ], + "angle": 0, + "content": "5.1 Results" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.567, + 0.787, + 0.748 + ], + "angle": 0, + "content": "All blur levels in the database. Our method outperforms all state-of-the-art retrieval methods for all settings, as shown in Table 2 for synthetic data and Table 3 for real data. Our model demonstrates significant superiority over the state-of-the-art retrieval approaches without introducing complex computational modules, highlighting the efficacy of our proposed loss functions. Results on the real-world dataset show that our method generalizes well to real data even when trained solely on synthetic data. Additionally, we train a version of our method where only sharp images are used (Ours-sharp). When evaluating with mAP@all (no distractors), it fails completely, especially for higher blur levels. When evaluating with mAP@100 (with distractors), it achieves reasonable results on query BL 1, but still worse than all other methods. This underscores the necessity for tailored methods and training data with highly motion-blurred objects." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Per blur level database. In this setup, we utilize datasets specific to each blur level, considering them as individual subsets of images with the designated blur level, to form the database. We then assess the retrieval performance by using query images with distinct blur levels, resulting in a comprehensive evaluation comprising 36 results (6 database BLs by 6 query BLs) for each method. We also compute the standard deviation (std) and range of mAPs for different settings" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.463, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Retrieval Robust to Object Motion Blur" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "table", + "bbox": [ + 0.271, + 0.144, + 0.739, + 0.248 + ], + "angle": 0, + "content": "
Cont. BL RangemAP AllmAP by query BL
123456
S82.2687.3788.4686.2982.6174.1358.26
M83.3387.9789.1286.9383.8776.0460.61
L84.0988.7489.5687.6884.4176.8962.42
L (Alpha)82.8387.6288.6886.4382.9675.6860.27
" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.249, + 0.788, + 0.305 + ], + "angle": 0, + "content": "Table 6: Ablation study of different blur level ranges in contrastive learning and design of blur estimation loss on the synthetic dataset (with 1M distractors). Small: range=1. Medium: range=3. Large: range=5. 'Alpha' corresponds to the alternative design choice for blur estimation loss. The database contains images of all blur levels." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.323, + 0.786, + 0.383 + ], + "angle": 0, + "content": "per method, reflecting the robustness of each method to different levels of motion blur. Our method outperforms all other state-of-the-art methods for all settings and also achieves the smallest standard deviation and range, showing the highest robustness to motion blur (Table 4). See supplementary material for more results." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.385, + 0.788, + 0.538 + ], + "angle": 0, + "content": "Qualitative results on our proposed synthetic (Fig. 4a) and real-world (Fig. 4b) datasets. Our method is compared to the top-performing state-of-the-art retrieval methods on a per-query basis, which yield numerous false positives resembling the texture of the motion-blurred object in the query image. In contrast, our approach accurately retrieves positive images even under severe motion blur in the query and amidst challenging distractors. Note that standard retrieval methods retrieve positive images primarily with blur levels similar to the query, whereas our method retrieves correct images exhibiting a wide range of blur levels (from 1-6 for synthetic and 2-6 for real). This illustrates that our method effectively learns invariant representations of objects across various blur levels." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.556, + 0.388, + 0.572 + ], + "angle": 0, + "content": "5.2 Ablation study" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.581, + 0.788, + 0.687 + ], + "angle": 0, + "content": "Loss components. Table 5 shows an ablation study of different loss functions (for more details, please see the supplementary material). The results show that all loss functions improve model performance. We highlight that our model outperforms the compared methods across all query blur levels (Table 2), even when utilizing only three out of the four loss terms at any combination. For instance, when using \\(\\mathcal{L}_{con}\\), \\(\\mathcal{L}_{cls}\\) and \\(\\mathcal{L}_{be}\\), our method surpasses Token [46] by 2.50 in mAP at blur level 1 and 1.91 at blur level 6, with an overall improvement of 2.33." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.688, + 0.788, + 0.749 + ], + "angle": 0, + "content": "Contrastive range. We investigate the research questions posed in Sec. 3.2 using Table 6, which displays results of using different blur level ranges for contrastive learning. A wider range of contrastive blur levels consistently enhances model performance, with the best results achieved at the maximum range of 5." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Design of blur estimation loss. We validate the effectiveness of our design of blur estimation loss \\(\\mathcal{L}_{be}\\) (Eq. (7)) in Table 6. We compare with the case where edge effects are ignored in the loss function. This corresponds to directly predicting the average of the entire alpha mask (Eq. (4)) inside the object boundary. The results show that our design of \\(\\mathcal{L}_{be}\\), which effectively mitigates the detrimental impact of edge blur, improves the model's retrieval performance." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.358, + 0.127 + ], + "angle": 0, + "content": "R. Zou et al." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.145, + 0.271, + 0.159 + ], + "angle": 0, + "content": "Query" + }, + { + "type": "image", + "bbox": [ + 0.226, + 0.159, + 0.272, + 0.202 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.28, + 0.159, + 0.325, + 0.209 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.331, + 0.159, + 0.374, + 0.209 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.381, + 0.159, + 0.425, + 0.215 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.431, + 0.146, + 0.606, + 0.159 + ], + "angle": 0, + "content": "Top 20 retrieval results" + }, + { + "type": "image", + "bbox": [ + 0.431, + 0.159, + 0.475, + 0.215 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.483, + 0.159, + 0.526, + 0.215 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.531, + 0.159, + 0.576, + 0.215 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.582, + 0.159, + 0.625, + 0.215 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.631, + 0.159, + 0.676, + 0.215 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.682, + 0.159, + 0.726, + 0.215 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.731, + 0.159, + 0.777, + 0.215 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.225, + 0.239, + 0.272, + 0.266 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.28, + 0.224, + 0.325, + 0.266 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.331, + 0.224, + 0.374, + 0.266 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.381, + 0.224, + 0.425, + 0.266 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.431, + 0.224, + 0.475, + 0.266 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.481, + 0.224, + 0.526, + 0.266 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.531, + 0.224, + 0.576, + 0.266 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.582, + 0.224, + 0.625, + 0.266 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.631, + 0.224, + 0.676, + 0.266 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.682, + 0.224, + 0.726, + 0.266 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.731, + 0.224, + 0.777, + 0.266 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.284, + 0.787, + 0.326 + ], + "angle": 0, + "content": "Fig. 5: Retrieval results on a real-world video (available on YouTube). The retrieved images are sorted from left to right and top to bottom, with the ranking from \\(1^{\\text{st}}\\) to \\(20^{\\text{th}}\\). Correct results are in green boxes, while incorrect ones are in red boxes." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.355, + 0.569, + 0.371 + ], + "angle": 0, + "content": "5.3 Application to real-world video data" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.379, + 0.789, + 0.562 + ], + "angle": 0, + "content": "We conduct an experiment to test our method on real-world video images. Due to the challenge of finding a sufficient number of real images of the same object with a well-distributed range of blur levels for quantitative analysis, we construct a small dataset for qualitative assessment. Specifically, we extract 190 images of the same ball from a YouTube soccer video to serve as both the query and database. We incorporate 4,431 images of sports balls from the MSCOCO dataset [21] as distractors, and additionally, we extract 169 images of a different ball from the same video to act as hard distractors. Both the objects and the backgrounds in these images come from domains entirely different from the synthetic training data. Fig. 5 illustrates examples of retrieval results, demonstrating our method's effectiveness in handling various blur conditions and complex and diverse backgrounds in the real world, despite being trained exclusively on synthetic data." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.598, + 0.36, + 0.613 + ], + "angle": 0, + "content": "6 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.628, + 0.788, + 0.809 + ], + "angle": 0, + "content": "In this paper, we introduce an innovative task: image retrieval in the presence of object motion blur. To tackle this challenge, we propose a global retrieval approach that integrates specialized loss functions tailored for the network to better comprehend motion blur. Our method stands as the first specifically designed solution for this new task. Moreover, we introduce the first benchmark datasets for blur retrieval, comprising both synthetic and real-world data. These datasets are meticulously processed and annotated to ensure their immediate applicability for future research in this emerging area. Through extensive experiments conducted on these benchmark datasets, we demonstrate the limitations of state-of-the-art standard retrieval methods in handling motion-blurred images and showcase the superior performance and robustness of our proposed approach in accurately retrieving images even in challenging motion-blurred scenarios." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.81, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Acknowledgement. We would like to thank Giorgos Tolias for the highly valuable discussions with him and the support he provided for this work." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.463, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Retrieval Robust to Object Motion Blur" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.323, + 0.16 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.179, + 0.784, + 0.208 + ], + "angle": 0, + "content": "1. Babenko, A., Lempitsky, V.: Aggregating local deep features for image retrieval. In: ICCV. pp. 1269-1277 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.209, + 0.785, + 0.236 + ], + "angle": 0, + "content": "2. Babenko, A., Slesarev, A., Chigorin, A., Lempitsky, V.: Neural codes for image retrieval. In: ECCV. pp. 584-599. Springer (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.237, + 0.785, + 0.264 + ], + "angle": 0, + "content": "3. Bay, H., Ess, A., Tuytelaars, T., Van Gool, L.: Speeded-up robust features (surf). Computer Vision and Image Understanding 110(3), 346-359 (2008)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.265, + 0.785, + 0.307 + ], + "angle": 0, + "content": "4. Cao, B., Araujo, A., Sim, J.: Unifying deep local and global features for image search. In: Vedaldi, A., Bischof, H., Brox, T., Frahm, J.M. (eds.) ECCV. pp. 726-743. Springer International Publishing, Cham (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.308, + 0.785, + 0.376 + ], + "angle": 0, + "content": "5. Chang, A.X., Funkhouser, T., Guibas, L., Hanrahan, P., Huang, Q., Li, Z., Savarese, S., Savva, M., Song, S., Su, H., Xiao, J., Yi, L., Yu, F.: ShapeNet: An Information-Rich 3D Model Repository. Tech. Rep. arXiv:1512.03012 [cs.GR], Stanford University — Princeton University — Toyota Technological Institute at Chicago (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.377, + 0.785, + 0.404 + ], + "angle": 0, + "content": "6. Chopra, S., Hadsell, R., LeCun, Y.: Learning a similarity metric discriminatively, with application to face verification. In: CVPR. vol. 1, pp. 539-546. IEEE (2005)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.406, + 0.785, + 0.447 + ], + "angle": 0, + "content": "7. Community, B.O.: Blender - a 3D modelling and rendering package. Blender Foundation, Stichting Blender Foundation, Amsterdam (2018), http://www.blender.org" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.448, + 0.785, + 0.474 + ], + "angle": 0, + "content": "8. Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: ImageNet: A LargeScale Hierarchical Image Database. In: CVPR (2009)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.476, + 0.785, + 0.529 + ], + "angle": 0, + "content": "9. Deng, J., Guo, J., Xue, N., Zafeiriou, S.: Arcface: Additive angular margin loss for deep face recognition. In: 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 4685-4694 (2019). https://doi.org/10.1109/CVPR.2019.00482" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.531, + 0.785, + 0.585 + ], + "angle": 0, + "content": "10. Gong, D., Yang, J., Liu, L., Zhang, Y., Reid, I., Shen, C., Van Den Hengel, A., Shi, Q.: From motion blur to motion flow: A deep learning solution for removing heterogeneous motion blur. In: CVPR. pp. 3806-3815 (2017). https://doi.org/10.1109/CVPR.2017.405" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.587, + 0.785, + 0.614 + ], + "angle": 0, + "content": "11. Gordo, A., Almazan, J., Revaud, J., Larlus, D.: Deep image retrieval: Learning global representations for image search (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.615, + 0.785, + 0.643 + ], + "angle": 0, + "content": "12. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: CVPR. pp. 770-778 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.644, + 0.785, + 0.685 + ], + "angle": 0, + "content": "13. Hendrycks, D., Dietterich, T.: Benchmarking neural network robustness to common corruptions and perturbations. Proceedings of the International Conference on Learning Representations (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.686, + 0.785, + 0.713 + ], + "angle": 0, + "content": "14. Jaesung, R., Haeyun, L., Jucheol, W., Sunghyun, C.: Real-world blur dataset for learning and benchmarking deblurring algorithms. In: ECCV (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.714, + 0.785, + 0.755 + ], + "angle": 0, + "content": "15. Jégou, H., Perronnin, F., Douze, M., Sánchez, J., Pérez, P., Schmid, C.: Aggregating local image descriptors into compact codes. IEEE TPAMI 34(9), 1704-1716 (2011)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.757, + 0.785, + 0.784 + ], + "angle": 0, + "content": "16. Kalantidis, Y., Mellina, C., Osindero, S.: Cross-dimensional weighting for aggregated deep convolutional features. In: ECCVW. pp. 685-701. Springer (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.785, + 0.785, + 0.812 + ], + "angle": 0, + "content": "17. Kingma, D.P., Ba, J.: Adam: A method for stochastic optimization. In: Bengio, Y., LeCun, Y. (eds.) ICLR (2015), http://arxiv.org/abs/1412.6980" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.813, + 0.785, + 0.84 + ], + "angle": 0, + "content": "18. Kotera, J., Rozumnyi, D., Sroubek, F., Matas, J.: Intra-frame object tracking by deblatting. In: ICCVW (Oct 2019)" + }, + { + "type": "list", + "bbox": [ + 0.22, + 0.179, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.358, + 0.127 + ], + "angle": 0, + "content": "R. Zou et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.148, + 0.785, + 0.175 + ], + "angle": 0, + "content": "19. Kupyn, O., Budzan, V., Mykhailych, M., Mishkin, D., Matas, J.: Deblurgan: Blind motion deblurring using conditional adversarial networks. In: CVPR (June 2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.177, + 0.785, + 0.202 + ], + "angle": 0, + "content": "20. Lee, S., Seong, H., Lee, S., Kim, E.: Correlation verification for image retrieval. In: CVPR. pp. 5374-5384 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.203, + 0.785, + 0.243 + ], + "angle": 0, + "content": "21. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dólar, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: ECCV. pp. 740-755. Springer (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.244, + 0.785, + 0.269 + ], + "angle": 0, + "content": "22. Lowe, D.G.: Distinctive image features from scale-invariant keypoints. IJCV 60, 91-110 (2004)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.27, + 0.785, + 0.311 + ], + "angle": 0, + "content": "23. Michaelis, C., Mitzkus, B., Geirhos, R., Rusak, E., Bringmann, O., Ecker, A.S., Bethge, M., Brendel, W.: Benchmarking robustness in object detection: Autonomous driving when winter is coming. arXiv preprint arXiv:1907.07484 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.312, + 0.785, + 0.338 + ], + "angle": 0, + "content": "24. Nah, S., Kim, T.H., Lee, K.M.: Deep multi-scale convolutional neural network for dynamic scene deblurring. In: CVPR (July 2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.339, + 0.785, + 0.365 + ], + "angle": 0, + "content": "25. Noh, H., Araujo, A., Sim, J., Weyand, T., Han, B.: Large-scale image retrieval with attentive deep local features. In: ICCV. pp. 3456-3465 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.365, + 0.785, + 0.475 + ], + "angle": 0, + "content": "26. Paszke, A., Gross, S., Massa, F., Lerer, A., Bradbury, J., Chanan, G., Killeen, T., Lin, Z., Gimelshein, N., Antiga, L., Desmaison, A., Kopf, A., Yang, E., Devito, Z., Raison, M., Tejani, A., Chilamkurthy, S., Steiner, B., Fang, L., Bai, J., Chintala, S.: Pytorch: An imperative style, high-performance deep learning library. In: Wallach, H., Larochelle, H., Beygelzimer, A., Alché-Buc, F., Fox, E., Garnett, R. (eds.) NeurIPS, pp. 8024-8035. Curran Associates, Inc. (2019), http://papers.neurips.cc/paper/9015-pytorch-an-imperative-style-high-performance-deep-learning-library.pdf" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.476, + 0.785, + 0.502 + ], + "angle": 0, + "content": "27. Perronnin, F., Liu, Y., Sánchez, J., Poirier, H.: Large-scale image retrieval with compressed fisher vectors. In: CVPR. pp. 3384-3391. IEEE (2010)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.503, + 0.785, + 0.529 + ], + "angle": 0, + "content": "28. Radenović, F., Iscen, A., Tolias, G., Avrithis, Y., Chum, O.: Revisiting oxford and paris: Large-scale image retrieval benchmarking. In: CVPR. pp. 5706-5715 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.53, + 0.785, + 0.555 + ], + "angle": 0, + "content": "29. Radenović, F., Tobias, G., Chum, O.: Fine-tuning cnn image retrieval with no human annotation. arXiv preprint arXiv:1711.02512 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.556, + 0.785, + 0.595 + ], + "angle": 0, + "content": "30. Radenović, F., Tolias, G., Chum, O.: Fine-tuning cnn image retrieval with no human annotation. IEEE TPAMI 41(7), 1655-1668 (2019). https://doi.org/10.1109/TPAMI.2018.2846566" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.597, + 0.785, + 0.637 + ], + "angle": 0, + "content": "31. Rozumnyi, D., Kotera, J., Sroubek, F., Matas, J.: Non-causal tracking by deblatting. In: Fink, G.A., Frintrop, S., Jiang, X. (eds.) GCPR. pp. 122-135. Springer International Publishing, Cham (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.638, + 0.785, + 0.677 + ], + "angle": 0, + "content": "32. Rozumnyi, D., Kotera, J., Sroubek, F., Novotny, L., Matas, J.: The world of fast moving objects. In: CVPR. pp. 4838-4846 (July 2017). https://doi.org/10.1109/CVPR.2017.514" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.678, + 0.785, + 0.704 + ], + "angle": 0, + "content": "33. Rozumnyi, D., Kotera, J., Sroubek, F., Matas, J.: Sub-frame appearance and 6d pose estimation of fast moving objects. In: CVPR. pp. 6777-6785 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.705, + 0.785, + 0.745 + ], + "angle": 0, + "content": "34. Rozumnyi, D., Oswald, M.R., Ferrari, V., Matas, J., Pollefeys, M.: Defmo: Deblurring and shape recovery of fast moving objects. In: CVPR. Nashville, Tennessee, USA (Jun 2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.746, + 0.785, + 0.772 + ], + "angle": 0, + "content": "35. Rozumnyi, D., Oswald, M.R., Ferrari, V., Pollefeys, M.: Shape from blur: Recovering textured 3d shape and motion of fast moving objects. In: NeurIPS (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.773, + 0.785, + 0.812 + ], + "angle": 0, + "content": "36. Rozumnyi, D., Oswald, M.R., Ferrari, V., Pollefeys, M.: Motion-from-blur: 3d shape and motion estimation of motion-blurred objects in videos. In: CVPR (Jun 2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.813, + 0.785, + 0.84 + ], + "angle": 0, + "content": "37. Sayed, M., Brostow, G.: Improved handling of motion blur in online object detection. In: CVPR. pp. 1706-1716 (June 2021)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.148, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.463, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Retrieval Robust to Object Motion Blur" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.175 + ], + "angle": 0, + "content": "38. Sivic, Zisserman: Video google: A text retrieval approach to object matching in videos. In: ICCV. pp. 1470-1477. IEEE (2003)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.175, + 0.787, + 0.203 + ], + "angle": 0, + "content": "39. Skorokhodov, I., Sotnikov, G., Elhoseiny, M.: Aligning latent and image spaces to connect the unconnectable. In: ICCV. pp. 14144-14153 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.203, + 0.787, + 0.231 + ], + "angle": 0, + "content": "40. Tolias, G., Avrithis, Y., Jégou, H.: Image search with selective match kernels: aggregation across single and multiple images. IJCV 116, 247-261 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.231, + 0.787, + 0.259 + ], + "angle": 0, + "content": "41. Tolias, G., Jenicek, T., Chum, O.: Learning and aggregating deep local descriptors for instance-level recognition. In: ECCV. pp. 460-477. Springer (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.259, + 0.787, + 0.286 + ], + "angle": 0, + "content": "42. Tolias, G., Sicre, R., Jégou, H.: Particular object retrieval with integral max-pooling of cnn activations. ICLR (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.286, + 0.787, + 0.314 + ], + "angle": 0, + "content": "43. Vasiljevic, I., Chakrabarti, A., Shakhnarovich, G.: Examining the impact of blur on recognition by convolutional networks. arXiv preprint arXiv:1611.05760 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.314, + 0.787, + 0.354 + ], + "angle": 0, + "content": "44. Wang, Y., Lu, Y., Gao, Y., Wang, L., Zhong, Z., Zheng, Y., Yamashita, A.: Efficient video deblurring guided by motion magnitude. In: ECCV. pp. 413-429. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.354, + 0.787, + 0.396 + ], + "angle": 0, + "content": "45. Weyand, T., Araujo, A., Cao, B., Sim, J.: Google landmarks dataset v2-a large-scale benchmark for instance-level recognition and retrieval. In: CVPR. pp. 2574-2583 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.396, + 0.787, + 0.438 + ], + "angle": 0, + "content": "46. Wu, H., Wang, M., Zhou, W.T., Hu, Y., Li, H.: Learning token-based representation for image retrieval. In: AAAI Conference on Artificial Intelligence (2021), https://apisemantic scholar.org/CorpusID:245124518" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.438, + 0.787, + 0.507 + ], + "angle": 0, + "content": "47. Yang, M., He, D., Fan, M., Shi, B., Xue, X., Li, F., Ding, E., Huang, J.: Dolg: Single-stage image retrieval with deep orthogonal fusion of local and global features. In: ICCV. pp. 11752-11761. IEEE Computer Society, Los Alamitos, CA, USA (oct 2021). https://doi.org/10.1109/ICCV48922.2021.01156, https://doi.ieeecomputersociety.org/10.1109/ICCV48922.2021.01156" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.507, + 0.787, + 0.536 + ], + "angle": 0, + "content": "48. Yokoo, S., Ozaki, K., Simo-Serra, E., Iizuka, S.: Two-stage discriminative reranking for large-scale landmark retrieval. In: CVPRW. IEEE (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.536, + 0.787, + 0.563 + ], + "angle": 0, + "content": "49. Zhang, H., Xie, H., Yao, H.: Spatio-temporal deformable attention network for video deblurring. In: ECCV. pp. 581-596. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.563, + 0.787, + 0.591 + ], + "angle": 0, + "content": "50. Zhou, S., Zhang, J., Pan, J., Xie, H., Zuo, W., Ren, J.: Spatio-temporal filter adaptive network for video deblurring. In: ICCV (2019)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.591 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/Retrieval Robust to Object Motion Blur/191ce9d4-edfe-414c-a611-8012902a3458_origin.pdf b/2024/Retrieval Robust to Object Motion Blur/191ce9d4-edfe-414c-a611-8012902a3458_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..95e9aebc70f47dbdf9a98bdd00855b255b144050 --- /dev/null +++ b/2024/Retrieval Robust to Object Motion Blur/191ce9d4-edfe-414c-a611-8012902a3458_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:81048d9cbde5d0d0a5e5516022f5b80fee1100d19ddb039ec468c9e25cd9f071 +size 9831849 diff --git a/2024/Retrieval Robust to Object Motion Blur/full.md b/2024/Retrieval Robust to Object Motion Blur/full.md new file mode 100644 index 0000000000000000000000000000000000000000..677204e66052de34506f8d22b095979d3df9270f --- /dev/null +++ b/2024/Retrieval Robust to Object Motion Blur/full.md @@ -0,0 +1,507 @@ +# Retrieval Robust to Object Motion Blur + +Rong Zou1, Marc Pollefeys1,2, and Denys Rozumnyi1,3 + +$^{1}$ ETH Zürich, $^{2}$ Microsoft, $^{3}$ Czech Technical University in Prague + +Abstract. Moving objects are frequently seen in daily life and usually appear blurred in images due to their motion. While general object retrieval is a widely explored area in computer vision, it primarily focuses on sharp and static objects, and retrieval of motion-blurred objects in large image collections remains unexplored. We propose a method for object retrieval in images that are affected by motion blur. The proposed method learns a robust representation capable of matching blurred objects to their deblurred versions and vice versa. To evaluate our approach, we present the first large-scale datasets for blurred object retrieval, featuring images with objects exhibiting varying degrees of blur in various poses and scales. We conducted extensive experiments, showing that our method outperforms state-of-the-art retrieval methods on the new blur-retrieval datasets, which validates the effectiveness of the proposed approach. Code, data, and model are available at https://github.com/Rong-Zou/Retrieval-Robust-to-Object-Motion-Blur. + +Keywords: object retrieval $\cdot$ object motion blur $\cdot$ datasets + +# 1 Introduction + +The task of image retrieval has long been a focal point in computer vision research, aiming to efficiently identify images within a large-scale database that share similarities with a given query image. In the realm of image retrieval, methods are often categorized into category-level and instance-level retrieval. We focus on the latter, where the objective is to identify images of a specific object featured in a query within a vast, unordered collection of database images. In this task, image representations play an important role since they capture the essence of the objects in images and enable the measurement of their similarities. Representing images based on hand-crafted features has been extensively studied [3,22]. Owing to the development in deep learning, recent strides have been made in adopting learning-based approaches to encode image content [2,4,20,25,30,41,42,46,47]. However, the current literature focuses exclusively on image representations for sharp objects, assuming that the objects are stationary and clearly visible in images. By only considering sharp and static objects, the existing retrieval methods fall short in addressing the dynamic nature of real-world scenarios, where objects are in movement, resulting in images blurred by object motion — a phenomenon prevalent in environments with fast-moving objects [32,34], low camera frame rates, and extended exposure times. + +To bridge this gap, we introduce a novel task in image retrieval: retrieval with object motion blur. The challenge of this task lies in finding image representations + +![](images/e5851576a0a04cc2b93fcfeace25333c07213c3d62bdcd0f609441a55d80beea.jpg) + +![](images/e63d0c979c7d7904bc7a10e2bf78ff5ae5fcbc0cb23c50096e60ea63e030a6fc.jpg) + +![](images/b00e63097785baab53a42b295c90ab69424271df59bb8a6d748b9cbec0dace1b.jpg) +Query + +![](images/e7208f9d984d549405dec0a8a8e773922a8865fa6f6cc7c6090e7aaed5d18b4c.jpg) + +![](images/4e705d91bde792b69dc765b0acaace0d17dd8a3577dd91f3ba55780698c0fab5.jpg) + +![](images/40783217d654850d6ad97c67fbfe2ca0a9f9c84fc2112b4ef8211c7b08f6aaf7.jpg) +Fig. 1: Retrieval robust to object motion blur. We present the first approach designed for retrieval in the presence of object motion blur. Our method captures blur-invariant features and generates blur-robust representations. We show retrieval results on the newly created datasets for the task of blur-robust retrieval: synthetic (top) and real (bottom). All retrieved images match the query and are marked with green boxes. + +![](images/574e1070ab18bb9ba8340f7146292622aada84baa328454121d807ef710bde1b.jpg) +Top 10 retrieval results + +![](images/40d57a3714f9f975241ac1cbc4080aea20390cf1ceefcc80ffb5db007ebc195f.jpg) + +![](images/52b982c77ccb4aaa5a0f91fef2f26d8b314d53981ae47741b94c37d732391ab9.jpg) + +![](images/9b44bec1c9c2c3ee2d1f2af2d0ed24223d58c6660f6f1432cf4099f763ed1b19.jpg) + +![](images/c586f9365b2165fe63a5b7359f51847adb1c2df5036553a4388306886de0d0a3.jpg) + +![](images/ebcf9d7f5e9eefa8660749610112ea9f4864dc2d831fb50f008e56289ac71ddc.jpg) + +![](images/2490eb1ce51b54e495c33b0da719918fd1116f0afbb8e3f6fe2242d308bc121b.jpg) + +![](images/67e1059172bf8e1c60c7b7fe3be40d2791ff5831a50eb875fbcf31bc51b7829d.jpg) + +![](images/3b829b67ef521d45aa31806447fb233ef732f44998ad749e73024fa7b3fbca2a.jpg) + +![](images/194f5851dc0176ba4dd5548e708e5fe84e344ee68720f4bc9a82831865a0822a.jpg) + +![](images/9cfcdb3a3294850bc9844cdb76c8c743972dbc7bbadfc4b54e70da314000db36.jpg) + +![](images/32201676e556d5003050b8c881a61bfc979b5e35bc03d5b6851ed93a2fd62097.jpg) + +![](images/7a513e771ef7cdbf33d3c006c877b76aadf2d581e395aa8abebf722c8430b207.jpg) + +![](images/c1b2d150eb6a2a6f4f73a076e08c4f6443ddf9ea793a7e9b0f67bf399b1fd77b.jpg) + +![](images/ae9b24a76edafa4dc45f4c9a5289659513968840ef374f6456f5c7e38443de9a.jpg) + +![](images/b52349e974aa6ca3fc8ab09e6276b0b938b93939feba35a9e60e2757f41b8209.jpg) + +robust to object motion blur, in addition to other variations such as background and changes in viewpoint. We also propose the first approach targeting this task that produces representations invariant to motion blur. In response to the unique challenge posed by object motion, we propose loss functions designed to enhance the model's comprehension of blur, thereby ensuring robust retrieval performance even under conditions of extreme blurring, as seen in Fig. 1. + +Recognizing the absence of datasets accommodating images with motion-blurred objects in existing retrieval benchmarks, we introduce the first benchmark datasets for this task, including a synthetic dataset for training and evaluation and a real-world dataset created specifically for evaluation. By re-training and evaluating state-of-the-art instance retrieval methods on this comprehensive benchmark, we reveal their shortcomings in handling motion-blurred scenarios. + +This work marks a key step in advancing practical methods for object retrieval in the challenging presence of motion blur. It has potential applications such as endangered wildlife monitoring, sports analysis, and security surveillance. + +To summarize, our main contributions are as follows: + +(1) We present the first method specifically designed for the novel retrieval task involving object motion blur. The method is trained with loss functions tailored to improve its understanding of motion blur. +(2) We introduce a new benchmark featuring synthetic and real-world datasets specifically constructed for this task. The datasets are carefully processed and are directly applicable for future research in blur retrieval. +(3) We conducted extensive experiments and detailed analyses to validate the effectiveness of our approach. Our method outperforms state-of-the-art retrieval methods and demonstrates superior robustness to motion blur. + +# 2 Related work + +Standard retrieval. In the area of standard image retrieval, local feature-based methods have evolved from traditional hand-crafted techniques like SIFT [22] and SURF [3] to contemporary deep learning-based approaches such as DELF [25] and HOW [41]. Considering that motion blur can corrupt local information, which may make local descriptors less robust, we adopt a global approach for our task. Classical methods typically derive global features by aggregating hand-crafted local features using techniques such as BoW [38], Fisher vectors [27], VLAD [15], AMSK [40], etc. Modern deep learning-based approaches, on the other hand, generate global descriptors by applying various pooling operations on feature maps encoded by convolutional neural networks (CNNs), such as sum pooling [1,41], weighted sum pooling [16], max pooling, regional max pooling [42], and generalized mean-pooling (GeM) [30], which has gained prominence in recent literature [4,20,47]. Our method integrates insights from recent advancements in global feature-based methods, adopting GeM pooling [30] for feature extraction and incorporating contrastive loss [6] and ArcFace loss [9] during training. + +Among the single-pass global feature-based methods, state-of-the-art standard retrieval approaches such as DELG global branch [4], DOLG [47], and Token [46] face challenges when handling images affected by motion blur, leading to degraded performance. Our work stands out in addressing this critical gap in the literature, as existing retrieval methods overlook the impact of motion blur. To bridge this research gap, we propose novel loss functions aimed at enhancing the model's comprehension of object motion blur, thereby contributing to the advancement of image retrieval in motion-blurred scenarios. + +Blur handling. Handling blur in images is an important topic in computer vision. [43] investigated the effect of different types of blur (defocus, uniform subject motion, camera shake) on the performance of CNNs for image classification and semantic segmentation. [37] studied online object detection under the influence of camera ego-motion-induced blur. Image and video deblurring are long-standing tasks in computer vision, which have been addressed by a variety of methods [10,19,44,49,50]. Joint deblurring and trajectory estimation for motion-blurred objects have also been extensively explored, with notable contributions from methods like [18, 31, 33, 34], some extending their scope to 3D reconstruction [35,36]. These existing works predominantly focus on other types of blur or address distinct tasks, emphasizing the need for specialized efforts in the context of our targeted task of retrieval involving object motion blur. + +Datasets. Several datasets have been proposed that contain generic corruptions including blur, such as ImaneNet-C [13] and Coco-C [23], but they do not focus purely on blurry images, and the types of blur they encompass (Gaussian, defocus, frosted glass, camera shake, and zoom) are not induced by object motion. Many datasets contain blurry-sharp image pairs, such as RealBlur [14] and GoPro [24], but they are created for deblurring tasks and not for retrieval, and the blur in images is mainly caused by camera motion or camera shake, not by object motion. A large variety of standard retrieval datasets have been proposed, including Revisited Oxford5k and Paris6k [28], Google Landmarks V2 [45, 48], + +![](images/c668d320ad6faf9330c604d900060fd3025451de5b1bf69b7feb94de94df8068.jpg) +Fig. 2: Overview of the proposed method for blur-robust retrieval. Our model takes a query image as input. Then, it generates a descriptor robust to object motion blur through the Blur Robust Image Descriptor Generator (BRIDGE) module that encodes the input image using a CNN backbone followed by generalized mean-pooling (GeM) [30], extracting a feature vector. This vector is subsequently fed into three heads: the blur estimation head, classification head, and localization head. Each head processes the feature vector to extract relevant information for identifying motion-blurred objects. The resulting three processed feature vectors are concatenated and passed through the final fully connected (FC) layer to produce the blur-robust descriptor. During training, the Blur Level-based Image Sample Selection (BLISS) mechanism is employed to select contrastive samples based on the query blur level (BL, Eq. (6)) $b_{q}$ and a specified blur level range $r$ . Then, these selected image samples are input into the BRIDGE module to extract descriptors for subsequent contrastive learning. + +SfM120K [29], etc. However, none of them incorporate motion-blurred images. To the best of our knowledge, there is currently no existing dataset tailored for the task of retrieval with object motion blur. In view of this, we created synthetic and real-world datasets specifically designed for the novel task of blur retrieval. + +# 3 Method + +We perform retrieval via nearest neighbor search in a representation space. Therefore, we require a representation model that is robust to object motion blur. We train a deep learning-based model that maps images to a vector space. The network is trained on our proposed synthetic dataset using multiple losses, including those tailored to object motion blur. As we show, the method generalizes well to real data despite being trained on synthetic data. + +The network takes an RGB image $\mathbf{I} \in \mathbb{R}^{3 \times H \times W}$ containing a sharp or blurred object as input and generates a descriptor $\mathbf{D} \in \mathbb{R}^d$ robust to object motion blur through the Blur Robust Image Descriptor Generator module (BRIDGE, Sec. 3.1). For training only, contrastive samples are selected by the Blur Level-based Image Sample Selection mechanism (BLISS, Sec. 3.2) based on the query blur level $b_q$ and the contrast blur level range $r$ . The selected image samples are sent into the BRIDGE module to extract descriptors for contrastive learning. The architecture of the proposed blur retrieval method is shown in Fig. 2. + +# 3.1 BRIDGE module + +The BRIDGE module encodes an image into a descriptor that is robust to object motion blur. The input image is encoded by a ResNet-50 backbone [12] to produce a feature map $\Omega \in \mathbb{R}^{C\times U\times V}$ . Then, we use GeM pooling [30] to process the feature map into a feature vector $\mathbf{f}\in \mathbb{R}^C$ , expressed as + +$$ +\mathbf {f} = \left[ \left(\frac {1}{| \boldsymbol {\Omega} |} \sum_ {(u, v) \in \boldsymbol {\Omega}} x _ {(c, u, v)} ^ {p}\right) ^ {\frac {1}{p}} \right] _ {c = 1, \dots , C}, \tag {1} +$$ + +where $p > 0$ is a parameter, and setting it to values larger than 1 forces the pooled vector to focus on the salient areas. This feature vector then undergoes multi-head processing to generate a motion blur-robust descriptor. Specifically, in the blur estimation head, we obtain blur estimation features $\mathbf{f}_{be} \in \mathbb{R}^{C_b}$ by a fully connected (FC) layer. The same dimension reduction is performed in the localization head, which generates features for location prediction $\mathbf{f}_{loc} \in \mathbb{R}^{C_l}$ . The features generated from the two heads are then concatenated with the whitened [11] features $\mathbf{f}_{cls} \in \mathbb{R}^{C_c}$ from the classification head. We further reduce the dimension of the concatenated features to obtain the final blur-robust descriptor of the image. This can be expressed as: + +$$ +\mathbf {D} = \mathbf {W} \cdot \operatorname {C o n c a t} \left(\mathbf {W} _ {b e} \mathbf {f}, \mathbf {W} _ {l o c} \mathbf {f}, \mathbf {W} _ {c l s} \mathbf {f}\right), \tag {2} +$$ + +where $\mathbf{W}_{be} \in \mathbb{R}^{C_b \times C}$ , $\mathbf{W}_{loc} \in \mathbb{R}^{C_l \times C}$ and $\mathbf{W}_{cls} \in \mathbb{R}^{C_c \times C}$ are weights of the FC layers in the blur estimation, localization, and classification heads, respectively, and $\mathbf{W} \in \mathbb{R}^{d \times (C_b + C_l + C_c)}$ is the weight of the final dimension-reduction layer. + +# 3.2 BLISS mechanism + +While the specific representation of the degree of motion blur for objects in images is yet to be formally defined, it is intuitively perceived as a continuous value. The level of detail exhibited by objects varies with the degree of blur: sharper objects show clear structures, textures, and poses, while more blurred objects primarily reveal color characteristics (see Fig. 3). This continuous nature of blur allows us to indirectly control the distance between samples with significantly different blur conditions during learning image representations. In this context, we manage the distance between samples that exhibit visual features and patterns of the same level of detail due to similar degrees of blur, thus indirectly influencing the distance between samples with significantly different levels of blur. Additionally, we can directly control the distance between samples with large differences in blur conditions by imposing constraints directly between such samples in the descriptor space. There are therefore two key research questions: What is the optimal value for the range of differences in the blur levels of image pairs used in contrastive learning? How does adjusting the range of differences in blur levels affect the model's performance? To answer these questions and identify the optimal blur level range for contrastive learning, we introduce the BLISS + +BL 1 + +![](images/6642b79676364e8c89cb675831934e6c3fde2d4eb2122b89c7729a17e3cd8fff.jpg) + +![](images/6ba913a8085179d9d0c064a8f9e2bf5098f5b0a8728148afd5c225d23b8fed49.jpg) + +![](images/7e0e2c791b150efe50efba756ff0351bf0cff18ee89f1e22fafc537f34bfa876.jpg) +Fig. 3: The introduced synthetic (a) and real-world (b) datasets. Rows 1-2: different trajectories of the same object. Rows 2-3: two different objects from the same category with similar shapes. Rows 4-6: objects from different categories but share similar textures. Columns correspond to different blur levels, from 1 to 6, and they are different for synthetic (BL) and real-world $(\mathrm{BL}^{\mathrm{r}})$ datasets. + +BL2 + +![](images/0f633b48db9977093ed4859f535cff3bcf61ee5628e6c223098311ff5824c2b5.jpg) + +![](images/3978cfa63644f3da896ec7b2d4609c0f41ae7bcb1a1182b80c55f98439b034dd.jpg) + +![](images/b0f9afc4191a68846c2b13254c21719342c071fc30bfa92788ee9dc754782e56.jpg) +(a) Synthetic dataset. + +BL 3 + +![](images/950a91c0f9c25a41324b5c339ebd9375e9dbb4f69d5273528c712f5c043428f9.jpg) + +![](images/2904c6d27de330c7d66f9e4086e13ba73ff6ce714659eae22fbd98d8bd3a35d4.jpg) + +![](images/31836addd03294e950c0f7f787519d790f957c029e912c561ea8b1a730214604.jpg) +Synthetic dataset. + +4 B + +![](images/76b58f60691c164ee6fccfb6f132896910fdb88e7811aa9f965e471627e3de40.jpg) + +![](images/c3d999b22e224f92773ff102fe420d91d6598caa7a29a474a43f407ceb2556c0.jpg) + +![](images/19babffa266ed2d505c33ae2a8230f98671482750f9f3ca3648c16cf7cc2b216.jpg) + +ic datas + +BL 5 + +![](images/561de1022d59e8c05bc3923c61a1f6ea1b4e34d208cf450139b38e2c85e2a0b3.jpg) + +![](images/937e81c6f721f3aa2092fe093b01de05730f31be1f82ae4bf48f9af4d84908c5.jpg) + +![](images/c2086dbcc11599aec790972550f698f97794bc37dbcac2ffccf25f85d4d69c5e.jpg) +et. + +BL6 + +![](images/a7ded884e044cd478ea5912890434bc5bc50b4bc6cff20c995552a87ece343b3.jpg) + +![](images/f1b8d1f3a0448d1917d874be2a2d1ae350a6e6638450d933c2ee86eb9fa8277b.jpg) + +![](images/df59a374dc2634ace7b5b87c938fa7bdd9b38b31094bf162828f1522e03672cd.jpg) + +BLr1 + +![](images/797233a28297aa13af5eec7b44a23ab187ebbf6fe7ef11237a3a25a9980890f0.jpg) + +![](images/6cd73b5fada4d8af40c25e095d416fbd2253e746053b3d1f7fff01f7292399aa.jpg) + +![](images/f520f16b08761cdb23e789cce72628bfa883fb95f01130fdd69a5afce426292b.jpg) + +![](images/f9b0e83ecc240de07685ea5923221a8a5d4496b62daeb1e19ed363af5190a95e.jpg) + +BL 2 + +![](images/1e952ae31f57eed977581ae0316b9a396f14527f97c4ac93e060a513d2d3074d.jpg) + +![](images/f1212ffd355866d7a77a5cf1446569f46f63b54241e7299ac1a2d0227134a9b5.jpg) + +![](images/cccb591f91afdfbe3b05a4bb0e6a737e9e82e1b230e286847914ba0ed6c64aeb.jpg) + +![](images/713ad930d3dec3548b1bb0c45c216766e38990654989e0bf4f5af281e42602ce.jpg) +(b) F + +BLr3 + +![](images/62a5ed96d8abaac7cfe00f8f84ab9d1dd92bd99e9990d6147f1bac85222ae08b.jpg) + +![](images/f00b4aea26f8e21171959ec9493e2251e4d2ee9d7cb4d3a23d5ee6cde34582e5.jpg) + +![](images/21506a60ac01a8fff38b4c766eb37cd7d2285ebb2fb39467ee1615debae43d70.jpg) + +![](images/56fc5975b524fe2713d15aa2c23f8c9f5c250183b8e81dcb09ea4624897cdb65.jpg) +Real-wo + +BLr 4 + +![](images/03859a3d6c8fbf13211382455ac3cdb897e78d337e03169079a0af95993ae3af.jpg) + +![](images/c94246543f4209f5b5392d8ff2634f6f57df7cd21cdab4e26b2ab0cd65c6f75c.jpg) + +![](images/261d641d9962aea66d0965c93b0535df90d5410a5953281c3ae704858f27dd7c.jpg) + +![](images/f26157c1992fc2dd4e68fbabb09d904aad0d5d08e0ac1251fe42f14c82a4fd48.jpg) +ld data + +BL 5 + +![](images/795f32b182797c152ada88fd63301d7201669cbc68003902f4e2c4b9f0bf7f9c.jpg) + +![](images/2f3c27a1ceb25eba907ad14950ce5b494742c67b633b33d28deeb924419ca33f.jpg) + +![](images/bacca14b8aae2545db7c510374ba9f24de249da0dc3d5e261adf0a47edf71e93.jpg) + +![](images/6a27d4d80f04cf26bd9bde1fdb6af6aa554700f54e39d3e8275f42466a1c51b7.jpg) +set. + +mechanism. This mechanism selects contrastive samples based on the blur level in the query image for contrastive learning. + +Blur severity definition. Our definition of blur severity is inspired by the image generation process. The formation of an image $\mathbf{I}$ containing a motion-blurred object in front of a background can be expressed [32] as + +$$ +\mathbf {I} = \mathbf {P} * \mathbf {O} + (1 - \mathbf {P} * \mathbf {M}) \mathbf {B} \tag {3} +$$ + +where the first term denotes the convolution $(\ast)$ of a point spread function $\mathbf{P}$ corresponding to a motion trajectory and the object's sharp appearance $\mathbf{O}$ , and the second term depicts the contribution of the background $\mathbf{B}$ , in which $\mathbf{M}$ is the binary mask of $\mathbf{O}$ . Let + +$$ +\boldsymbol {\alpha} = \mathbf {P} * \mathbf {M}, \tag {4} +$$ + +$\pmb{\alpha} \in \mathbb{R}^{H \times W}$ , $\pmb{\alpha}(i,j) \in [0,1], \forall (i,j) \in \mathbf{I}$ . Considering only objects with opaque materials, $\pmb{\alpha}$ represents the visibility of the object. In other words, it can characterize the amount of blur. Generally, the sharper the object in the image, the larger the $\pmb{\alpha}$ values are. However, at the edge pixels, the object blends with the background, and the alpha values of these edge pixels gradually fade to 0. Combining the alpha mask from Eq. (4) and considering edge effects, we define the Blur Severity (BS) of an image with a motion-blurred object as: + +$$ +B S = 1 - \frac {\sum_ {i = 1} ^ {H} \sum_ {j = 1} ^ {W} \boldsymbol {\beta} (i , j) \cdot \boldsymbol {\alpha} (i , j)}{\sum_ {i = 1} ^ {H} \sum_ {j = 1} ^ {W} \boldsymbol {\beta} (i , j)}, \tag {5} +$$ + +where $\beta$ is a thresholded (by 0) and eroded (by 3 pixels) binary mask of $\alpha$ . Erosion effectively removes the adverse contribution of the edge pixels to the blur severity measure. As a result, $BS \in (0,1)$ , and the larger the value, the more severe the motion blur in the image. + +
Dataset#Images Total#Images by BL (Syn.) or BLr (Real)
123456
Syn. Q20,9954,2883,9324,0784,0892,9301,678
Syn. DB91,62118,87117,50817,88818,02912,54667,79
Syn. Dist.1,091,939214,364177,869222,542235,662149,82891,674
Real Q2,753612620561396315249
Real DB10,3401,9231,8032,0801,7451,3751,414
+ +Table 1: Distribution of images in our synthetic and real datasets, sorted by blur levels. 'Q' stands for 'query', 'DB' denotes database, and 'Dist.' represents distractors. + +Blur level-based sample selection. Since blur severity is a continuous quantity, finding images within a certain blur severity range for each query image is expensive. We thus use a discretized version of blur severity — Blur Level (BL): + +$$ +B L = \lceil 1 0 B S \rceil , \tag {6} +$$ + +where $\lceil \cdot \rceil$ denotes ceiling operation. Each image is assigned to a $BL$ according to its $BS$ . During training, we select contrastive samples based on the query blur level $b_{q}$ and the specified contrastive $BL$ range $r$ . The selected $N_{p}$ positive and $N_{n}$ negative samples thus have blur levels within the range of $b_{q} - r$ to $b_{q} + r$ . + +# 3.3 Training objectives + +Blur estimation loss. Blur severity contains information about the visibility of the blurred object in an image (Sec. 3.2). From another perspective, it reflects a signal-to-noise ratio (SNR), since the more severe the motion blur, the greater the ratio of background (noise) to foreground (signal). In view of this, we define blur estimation loss as + +$$ +\mathcal {L} _ {b e} = | p - (1 - B S) |, \tag {7} +$$ + +where $p \in \mathbb{R}$ is the blur value predicted by the blur estimation head. By constraining the estimate of the continuous $BS$ (Eq. (5)), we implicitly force the network to identify blur and estimate the SNR, which helps it to focus on and extract signal information in the blurred region while ignoring background noises. + +Localization loss. This loss is defined as + +$$ +\mathcal {L} _ {l o c} = | x - \hat {x} | + | y - \hat {y} | + | h - \hat {h} | + | w - \hat {w} |, \tag {8} +$$ + +where $(x,y,w,h)$ and $(\hat{x},\hat{y},\hat{w},\hat{h})$ represent the normalized ground truth bounding box and the predicted one from the localization head, respectively. Localization loss requires the network to distinguish motion-blurred objects from backgrounds under various blur conditions, forcing it to recognize blur, identify objects at different blur levels, and further, focus more on the foreground objects. + +Classification loss. Following DELG [4], we apply ArcFace margin loss [9] to the descriptor $\mathbf{D}$ , expressed as + +$$ +\mathcal {L} _ {c l s} = - \log \left(\frac {\exp (\gamma \times A F (\hat {\mathbf {w}} _ {t} ^ {T} \hat {\mathbf {D}} , 1))}{\sum_ {n} \exp (\gamma \times A F (\hat {\mathbf {w}} _ {n} ^ {T} \hat {\mathbf {D}} , y _ {n}))}\right), \tag {9} +$$ + +
MethodmAP by query BL (synthetic)mAP by query BL (synthetic + 1M)
All123456All123456
DELG [4]81.6583.5584.9684.8982.6977.3566.2068.1973.6475.4073.3468.0558.2842.46
DOLG [47]83.0385.1086.2685.9384.0478.7868.0769.9775.7577.4775.0170.1060.0142.49
Token [46]84.8486.5387.9987.5085.8180.7071.5770.6575.3277.6675.5170.2461.1948.05
Ours-sharp27.5743.0638.5329.7419.4010.876.1432.6471.9343.8827.1815.417.944.27
Ours91.7893.0593.4893.1492.2590.2082.8684.0988.7489.5687.6884.4176.8962.42
+ +Table 2: Retrieval results on the synthetic dataset with (+1M) and without distractors. Ours-sharp is our model trained with only sharp images. 'All' denotes the overall performance for queries of all blur levels. All methods are re-implemented and re-trained on our synthetic training set. The database contains images of all blur levels. + +where $\gamma$ is a scale factor, $\hat{\mathbf{w}}_{\mathbf{i}}$ refers to the $i$ -th row of the $L_{2}$ normalized $N$ -class classifier $\hat{\mathbf{W}} \in \mathbb{R}^{d \times N}$ , $\hat{\mathbf{D}}$ denotes the $L_{2}$ normalized $\mathbf{D}$ . $y$ is the one-hot label, and $t$ represents the ground truth class index ( $y_{t} = 1$ ). $AF$ is the ArcFace-adjusted cosine similarity and it can be expressed as + +$$ +A F (s, g) = \left\{ \begin{array}{l l} \cos (\operatorname {a r c c o s} (s) + m), & \text {i f} g = 1 \\ s, & \text {i f} g = 0 \end{array} , \right. \tag {10} +$$ + +where $s$ is the cosine similarity, $m$ is the ArcFace margin, and $g \in \{0,1\}$ . $g = 1$ indicates the ground truth class. The classification loss requires the model to group sharp or motion-blurred images of the same object together, while distinguishing between different objects, which may appear similar, especially when blurred. This implicitly forces the model to focus on blur-invariant features. + +Contrastive loss. We employ contrastive learning [6] to construct a descriptor space, where the distance between similar samples is minimized, and between dissimilar samples maximized. The contrastive loss is applied to (query, positive) and (query, negative) descriptor pairs: + +$$ +\mathcal {L} _ {\text {c o n}} (i, j) = \left\{ \begin{array}{l l} 0. 5 \cdot D _ {i, j} ^ {2}, & \text {i f} y (i, j) = 1 \\ 0. 5 \left(\max \{0, \tau - D _ {i, j} \}\right) ^ {2}, & \text {i f} y (i, j) = 0 \end{array} , \right. \tag {11} +$$ + +where $D_{i,j} = ||\hat{\mathbf{D}}(i) - \hat{\mathbf{D}}(j)||$ and $\hat{\mathbf{D}}(i)$ is the $L_2$ normalized descriptor of image $i$ . $\tau$ is a margin parameter, and $y(i,j)$ is binary indicating whether the pair matches $(= 1)$ . The contrastive loss further enhances the focus on blur-invariant features by explicitly forcing descriptors of the same object, whether sharp or blurred, to be close together, and those of different objects to be far apart. + +Joint loss. The joint loss is a weighted sum of all previously defined losses: + +$$ +\mathcal {L} _ {\text {j o i n t}} = \mathcal {L} _ {\text {c o n}} + \alpha_ {\text {c l s}} \mathcal {L} _ {\text {c l s}} + \alpha_ {\text {b e}} \mathcal {L} _ {\text {b e}} + \alpha_ {\text {l o c}} \mathcal {L} _ {\text {l o c}}. \tag {12} +$$ + +# 4 Dataset + +We create novel synthetic and real datasets for training and evaluating retrieval methods for motion-blurred objects. The synthetic dataset serves for both training and evaluation. We also generate a large-scale synthetic distractor set to + +
MethodmAP AllmAP by Query BLr
123456
DELG [4]54.8249.1363.4357.2555.0153.7742.92
DOLG [47]54.6443.9360.5958.3659.0658.5845.78
Token [46]43.3338.7147.0850.7946.4442.7124.43
Ours-sharp40.2449.5545.0241.3333.2329.4027.91
Ours62.8857.5070.3866.7763.1864.4846.14
+ +Table 3: Retrieval results on the real-world dataset (trained on synthetic without fine-tuning). The database includes all blur levels. Our method outperforms all others. + +evaluate methods in a more challenging setup. The novel real-world dataset is specifically constructed to validate the methods' effectiveness in real scenarios. All datasets are meticulously processed for the blur retrieval task, laying a crucial foundation for future research. Statistics of these datasets are shown in Table 1. + +# 4.1 Synthetic dataset + +The images in this dataset contain 3D objects moving along different trajectories, in front of different backgrounds, and with different amounts of motion blur. Thus, this dataset poses many challenges to retrieval methods. + +Backgrounds. Background images are sampled from the LHQ dataset [39], cropped to a resolution of $240 \times 320$ . + +Objects. We select 39 categories with the largest number of textured objects from the ShapeNet dataset [5] and randomly sample 30 objects in each category. + +Trajectories. For each object, we generate 120 random linear trajectories, each divided into 23 sub-trajectories. We then select up to 10 consecutive sub-trajectories to create images featuring objects with diverse levels of motion blur. + +Rendering. We render blurry images with Blender Cycles [7] by capturing moving objects during the camera exposure time. Additionally, we render one sharp image per trajectory, placed at the central position along the trajectory. + +Filtering. We discard renderings where the object occupies less than $1.5\%$ of the pixels or where the intersection-over-union between silhouette renderings at trajectory endpoints is below $20\%$ . The former condition suggests the object is too small, while the latter indicates the appearance changes are too drastic. + +Ground truth. For each generated image, we estimate the bounding box coordinates by detecting the non-zero pixels in the alpha channel. We calculate the blur level for each image based on Eq. (6). + +Retrieval dataset. In total, we rendered 1.5 million images for 1138 different instances (examples in Fig. 3a). Additionally, we create a challenging distractor set with over 1 million images (examples in the supplementary material), using the same object categories as before to increase the difficulty of the distractors in terms of intra-class similarity. We randomly select 40 objects per category, excluding the 30 previously chosen objects, and generate 70 trajectories for each. For each trajectory, we render one sharp image and 10 blurry images with varying degrees of blur. Backgrounds are also selected from a different subset of [39]. + +
BLQDDELG [4] (range: 57.79; std: 16.75)DOLG [47] (range: 56.81; std: 16.93)
All123456All123456
175.9986.7384.1780.2672.4863.1250.0877.7887.0185.7081.8575.1166.2052.39
277.3582.7486.2282.8776.6866.4650.0078.4484.1887.0083.4377.9167.8851.31
377.1979.6283.4382.9478.4669.4352.8678.4981.8484.7483.8579.6970.5853.12
474.9674.2678.9179.8578.0570.8955.1276.4077.2780.5580.8079.1071.7355.28
562.7157.8661.7264.3765.3668.9656.0564.1961.3163.6865.3866.0569.6855.75
634.8929.5628.9431.7534.0645.3753.8635.5131.2430.2032.2934.1845.0053.41
BLQDToken [46] (range: 55.72; std: 16.11)Ours (range: 53.68; std: 15.65)
All123456All123456
179.6588.2287.3483.5076.8768.5156.6388.5795.0994.2491.5487.1980.9168.15
279.9784.7888.4884.7778.9970.0155.8288.7593.0794.6592.1888.7782.0367.18
379.5081.7985.4684.7580.2172.0958.1888.7691.7093.1892.3289.8383.4168.94
476.8276.5680.4281.2479.3172.6159.6187.1688.1290.1190.0989.3184.3770.35
565.3961.8564.4966.3167.1470.9660.3676.2873.8774.9076.0477.3983.0671.67
637.9833.2332.7634.6136.5047.5657.4347.3143.4341.4142.8944.5558.1369.59
+ +Table 4: Retrieval results by query (Q) and database (D) blur level (BL) on our synthetic dataset (with 1M distractors). 'All' denotes the overall performance for queries of all blur levels. The standard deviation and range of mAP values are displayed alongside each method. The best scores under the same settings are marked in bold. + +# 4.2 Real-world dataset + +We record videos with a GoPro 7 Hero camera at 240 fps with full exposure. We use 35 different objects and move each object 3 to 5 times along different trajectories in front of the camera. None of the real-world objects are in synthetic data and $2/3$ of which are from new categories not present in synthetic data. This better tests the generalization ability of models to new objects and categories. In total, we recorded 139 videos. Different amounts of blur are generated by averaging consecutive frames, ranging from just one up to 200 averaged frames. + +Computing blur level based on Eq. (6) is not feasible for the real-world dataset. Therefore, we manually annotate the data, assigning each image to one of 6 blur levels based on the perceived blur, and we use the ratio of the trajectory length to the average object size as a reference. The annotation enables a more detailed analysis and a more comprehensive evaluation of model performance. Further annotation details can be found in the supplementary material. Since the blur level for real data differs from that in Eq. (6), we denote it as $\mathrm{BL}^{\mathrm{r}}$ . + +The real-world dataset contains 13,093 images, with an average of 374 images for each object (examples in Fig. 3b). Compared to the synthetic dataset, the object scale change is more drastic in different trajectories, e.g. rows 1 & 2. The trajectories are notably different from those in the synthetic dataset. They involve complex curved trajectories with acceleration and other physical forces, which poses an additional challenge for methods trained on our synthetic dataset. + +# 5 Experiments + +Setup. For each category of objects in the synthetic dataset, we randomly take $70\%$ (rounded down, same for validation) for training (792 objects), $15\%$ for validation (153 objects), and $15\%$ for testing (193 objects). For the training objects, + +![](images/90bf25b235aad8fc46e634961eadf13e9b8574183c30253697f6e8e252ced634.jpg) +Query +(a) Example of retrieval results on our synthetic dataset (with distractors). + +![](images/31cfcdb2ee0ec2db6a6bda3958324ef169dafb0c823d271a7154a5f1b9f00dab.jpg) +(b) Example of retrieval results on our real-world dataset. +Fig. 4: Comparison of retrieval results. We compare to the best-performing state-of-the-art method for the corresponding query. The retrieved images are sorted from left to right and top to bottom, with the ranking from $1^{\mathrm{st}}$ to $20^{\mathrm{th}}$ . The blur level of each image is shown in the bottom left corner. Correct results are in green boxes, while incorrect ones are in red boxes. We provide more results in the supplementary material. + +we take one image for each trajectory while ensuring a balanced distribution of images across different blur levels. In total, there are 95k images in the training set. For each of the test objects, we sample 20 trajectories to form a query set, and the rest is used for the database. The database images are sampled from all blur levels unless specified otherwise (for per blur level database experiments). We additionally show performance on a subset of queries for a given blur level. In all experiments (except Ours-sharp), only one model is trained for all blur levels, and the evaluation is performed for varying blur levels. + +Implementation details. The proposed method is implemented in PyTorch [26]. We use ResNet-50 [12] up to the last downsampling block as the backbone, deactivate the normalization layers, and initialize the model with ImageNet [8] pre-trained weights. ADAM [17] optimizer is employed to minimize the joint loss with a fixed learning rate $10^{-4}$ , batch size 32, and 30 epochs. We initialize the power parameter to 3 for GeM pooling [30]. The contrastive margin $\tau$ is set to 0.7. Following [30], we set the number of contrastive samples as $N_{p} = 1$ , $N_{n} = 5$ . The contrastive blur range is experimentally set to 5. For classification, we empirically set the margin $m$ to 0.15 and the scale $\gamma$ to 30. Loss weights are chosen so that the values of different losses are roughly of the same order of magnitude, i.e. $\alpha_{cls} = 0.1$ , $\alpha_{be} = 1.0$ , $\alpha_{loc} = 10$ . The descriptor dimension $d$ is set to 128. + +Evaluation protocols. We report mean average precision (mAP@all) as commonly used in retrieval. In case 1M distractors are used, we report mAP@100. + +
\( \mathcal{L}_{con} \)\( \mathcal{L}_{cls} \)\( \mathcal{L}_{be} \)\( \mathcal{L}_{loc} \)mAP AllmAP by Query BL
123456
✓✘✘✘78.1380.5181.7081.1679.2073.9361.24
✓✘✓✘81.6683.4985.0184.4382.6977.6467.08
✓✘✘✓85.9487.5488.2587.8386.5283.0875.58
✓✘✓✓87.4888.6989.8989.4088.2484.9776.74
✘✓✘✘78.7381.5382.9782.8379.8672.9359.00
✘✓✓✘83.6785.1987.5687.4085.1079.2265.93
✘✓✘✓88.7489.8990.9190.8889.7586.0777.67
✘✓✓✓91.2392.0293.1693.0991.9789.0082.27
✓✓✘✘85.0687.4288.2987.6685.8581.2069.96
✓✓✓✘87.1789.0390.0389.5588.0783.9173.48
✓✓✘✓90.3991.8592.4592.1491.2088.2079.36
✓✓✓✓91.7893.0593.4893.1492.2590.2082.86
+ +Table 5: Ablation study of different loss components on the synthetic dataset. Results are grouped according to the loss function(s) directly applied to the descriptor. The best scores within each split are bolded. The database contains all blur levels. + +Compared methods. Since retrieval with object motion blur has not been addressed before, there are no methods specifically designed for this task. Therefore, we compare to state-of-the-art standard retrieval methods: DELG-global [4] (referred to as DELG), DOLG [47], and Token [46]. All these methods are re-implemented as in the original papers, except for changing the size of the output descriptors to 128 for fair comparison. We re-train the compared methods using the same ArcFace parameters and on the same synthetic data as our approach. + +# 5.1 Results + +All blur levels in the database. Our method outperforms all state-of-the-art retrieval methods for all settings, as shown in Table 2 for synthetic data and Table 3 for real data. Our model demonstrates significant superiority over the state-of-the-art retrieval approaches without introducing complex computational modules, highlighting the efficacy of our proposed loss functions. Results on the real-world dataset show that our method generalizes well to real data even when trained solely on synthetic data. Additionally, we train a version of our method where only sharp images are used (Ours-sharp). When evaluating with mAP@all (no distractors), it fails completely, especially for higher blur levels. When evaluating with mAP@100 (with distractors), it achieves reasonable results on query BL 1, but still worse than all other methods. This underscores the necessity for tailored methods and training data with highly motion-blurred objects. + +Per blur level database. In this setup, we utilize datasets specific to each blur level, considering them as individual subsets of images with the designated blur level, to form the database. We then assess the retrieval performance by using query images with distinct blur levels, resulting in a comprehensive evaluation comprising 36 results (6 database BLs by 6 query BLs) for each method. We also compute the standard deviation (std) and range of mAPs for different settings + +
Cont. BL RangemAP AllmAP by query BL
123456
S82.2687.3788.4686.2982.6174.1358.26
M83.3387.9789.1286.9383.8776.0460.61
L84.0988.7489.5687.6884.4176.8962.42
L (Alpha)82.8387.6288.6886.4382.9675.6860.27
+ +Table 6: Ablation study of different blur level ranges in contrastive learning and design of blur estimation loss on the synthetic dataset (with 1M distractors). Small: range=1. Medium: range=3. Large: range=5. 'Alpha' corresponds to the alternative design choice for blur estimation loss. The database contains images of all blur levels. + +per method, reflecting the robustness of each method to different levels of motion blur. Our method outperforms all other state-of-the-art methods for all settings and also achieves the smallest standard deviation and range, showing the highest robustness to motion blur (Table 4). See supplementary material for more results. + +Qualitative results on our proposed synthetic (Fig. 4a) and real-world (Fig. 4b) datasets. Our method is compared to the top-performing state-of-the-art retrieval methods on a per-query basis, which yield numerous false positives resembling the texture of the motion-blurred object in the query image. In contrast, our approach accurately retrieves positive images even under severe motion blur in the query and amidst challenging distractors. Note that standard retrieval methods retrieve positive images primarily with blur levels similar to the query, whereas our method retrieves correct images exhibiting a wide range of blur levels (from 1-6 for synthetic and 2-6 for real). This illustrates that our method effectively learns invariant representations of objects across various blur levels. + +# 5.2 Ablation study + +Loss components. Table 5 shows an ablation study of different loss functions (for more details, please see the supplementary material). The results show that all loss functions improve model performance. We highlight that our model outperforms the compared methods across all query blur levels (Table 2), even when utilizing only three out of the four loss terms at any combination. For instance, when using $\mathcal{L}_{con}$ , $\mathcal{L}_{cls}$ and $\mathcal{L}_{be}$ , our method surpasses Token [46] by 2.50 in mAP at blur level 1 and 1.91 at blur level 6, with an overall improvement of 2.33. + +Contrastive range. We investigate the research questions posed in Sec. 3.2 using Table 6, which displays results of using different blur level ranges for contrastive learning. A wider range of contrastive blur levels consistently enhances model performance, with the best results achieved at the maximum range of 5. + +Design of blur estimation loss. We validate the effectiveness of our design of blur estimation loss $\mathcal{L}_{be}$ (Eq. (7)) in Table 6. We compare with the case where edge effects are ignored in the loss function. This corresponds to directly predicting the average of the entire alpha mask (Eq. (4)) inside the object boundary. The results show that our design of $\mathcal{L}_{be}$ , which effectively mitigates the detrimental impact of edge blur, improves the model's retrieval performance. + +Query + +![](images/a494f1580db886cd7639ea060b5a2e60dbde5a8ce3938633c8ad5a762bd24a55.jpg) + +![](images/0a6a6e9bd054fdcdee6d0354fefddce2a6a13fe04fedecbcfac38357560e592b.jpg) + +![](images/030d065373cd3d9d6a3dd21c553c054cbc4fcaa8e562192687e92fb1addae8d6.jpg) + +![](images/d284e9ca6d7e72a065661071cab88716aa8831fc322cb3d3769f3b3d866f5098.jpg) + +Top 20 retrieval results + +![](images/36412bd7feaaac7a3590b476d2c376f50079dd10e3a0276277f66f8316b9cfe4.jpg) + +![](images/88dbe6e78998eaa391e0519e3f6b89c65ebcda9e1535a705fb20122bd5c1d618.jpg) + +![](images/2ace5f59f5ec63f43f9639a097d46e54f4e615382eebc7fff23700b5fe2a8ce1.jpg) + +![](images/08c33429958f42c562b164d7e66c3a852d0a501db5cb388ba76014198c52b14d.jpg) + +![](images/0a3143d0ac5601b487fd1a8309dd3be9767bdd97766f14db649d2a1de4a9acde.jpg) + +![](images/0830a0141f625ab6230ad0efc105f733c93aa43e9c6df55ddffed236642e9fa9.jpg) + +![](images/1f295044049f0b5aeacdbb99949d71fbf17f3ac65d18bb98d8e53a6376d73486.jpg) + +![](images/d8ab84f06d966f6244be744dd91f6d4a04bd3c04b9deb7c0ee1e19ffad06c4ae.jpg) +Fig. 5: Retrieval results on a real-world video (available on YouTube). The retrieved images are sorted from left to right and top to bottom, with the ranking from $1^{\text{st}}$ to $20^{\text{th}}$ . Correct results are in green boxes, while incorrect ones are in red boxes. + +![](images/b80286f23140b80f8044c7c783af99f3ab5a34ac8de1fe4b8f190f75ce0eb272.jpg) + +![](images/0d68963c74d90708e528daafc7c2daa80eccdbfd1a522a2bcc49003f14760864.jpg) + +![](images/f35be9a90c52019ffcc7a43e44b2e6b56f2ceea2dd4d185935df010b1aac6747.jpg) + +![](images/a1bb25c9044d727f08000882b42c0a000a963eb6d2be208d1d5a7fae5c37047a.jpg) + +![](images/20c1d50cf4b5b96ce631b016d9042af1576cb104f792ec4a1e134f92a898d5b4.jpg) + +![](images/a3c1e535c949266694e589f54c4c82fd37e0af5e38de22734fbfc9ee95197f58.jpg) + +![](images/b6807b3c17d3e03a0a23fc3153449d104f75c95bd7296cba3231380c3d66ba26.jpg) + +![](images/b9373c1d0499af4ea1005997eafa23395ffaef94f215889cb0829d2944ee1302.jpg) + +![](images/056cb53277149902596142038df0e148175206702c6ef3de25313ef90c936194.jpg) + +![](images/648ddb065bd600ef40956baa2cd1b5e85b65713a61d0bbdc3080282e56bbf3d5.jpg) + +# 5.3 Application to real-world video data + +We conduct an experiment to test our method on real-world video images. Due to the challenge of finding a sufficient number of real images of the same object with a well-distributed range of blur levels for quantitative analysis, we construct a small dataset for qualitative assessment. Specifically, we extract 190 images of the same ball from a YouTube soccer video to serve as both the query and database. We incorporate 4,431 images of sports balls from the MSCOCO dataset [21] as distractors, and additionally, we extract 169 images of a different ball from the same video to act as hard distractors. Both the objects and the backgrounds in these images come from domains entirely different from the synthetic training data. Fig. 5 illustrates examples of retrieval results, demonstrating our method's effectiveness in handling various blur conditions and complex and diverse backgrounds in the real world, despite being trained exclusively on synthetic data. + +# 6 Conclusion + +In this paper, we introduce an innovative task: image retrieval in the presence of object motion blur. To tackle this challenge, we propose a global retrieval approach that integrates specialized loss functions tailored for the network to better comprehend motion blur. Our method stands as the first specifically designed solution for this new task. Moreover, we introduce the first benchmark datasets for blur retrieval, comprising both synthetic and real-world data. These datasets are meticulously processed and annotated to ensure their immediate applicability for future research in this emerging area. Through extensive experiments conducted on these benchmark datasets, we demonstrate the limitations of state-of-the-art standard retrieval methods in handling motion-blurred images and showcase the superior performance and robustness of our proposed approach in accurately retrieving images even in challenging motion-blurred scenarios. + +Acknowledgement. We would like to thank Giorgos Tolias for the highly valuable discussions with him and the support he provided for this work. + +# References + +1. Babenko, A., Lempitsky, V.: Aggregating local deep features for image retrieval. In: ICCV. pp. 1269-1277 (2015) +2. Babenko, A., Slesarev, A., Chigorin, A., Lempitsky, V.: Neural codes for image retrieval. In: ECCV. pp. 584-599. Springer (2014) +3. Bay, H., Ess, A., Tuytelaars, T., Van Gool, L.: Speeded-up robust features (surf). Computer Vision and Image Understanding 110(3), 346-359 (2008) +4. Cao, B., Araujo, A., Sim, J.: Unifying deep local and global features for image search. In: Vedaldi, A., Bischof, H., Brox, T., Frahm, J.M. (eds.) ECCV. pp. 726-743. Springer International Publishing, Cham (2020) +5. Chang, A.X., Funkhouser, T., Guibas, L., Hanrahan, P., Huang, Q., Li, Z., Savarese, S., Savva, M., Song, S., Su, H., Xiao, J., Yi, L., Yu, F.: ShapeNet: An Information-Rich 3D Model Repository. Tech. Rep. arXiv:1512.03012 [cs.GR], Stanford University — Princeton University — Toyota Technological Institute at Chicago (2015) +6. Chopra, S., Hadsell, R., LeCun, Y.: Learning a similarity metric discriminatively, with application to face verification. In: CVPR. vol. 1, pp. 539-546. IEEE (2005) +7. Community, B.O.: Blender - a 3D modelling and rendering package. Blender Foundation, Stichting Blender Foundation, Amsterdam (2018), http://www.blender.org +8. Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: ImageNet: A LargeScale Hierarchical Image Database. In: CVPR (2009) +9. Deng, J., Guo, J., Xue, N., Zafeiriou, S.: Arcface: Additive angular margin loss for deep face recognition. In: 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 4685-4694 (2019). https://doi.org/10.1109/CVPR.2019.00482 +10. Gong, D., Yang, J., Liu, L., Zhang, Y., Reid, I., Shen, C., Van Den Hengel, A., Shi, Q.: From motion blur to motion flow: A deep learning solution for removing heterogeneous motion blur. In: CVPR. pp. 3806-3815 (2017). https://doi.org/10.1109/CVPR.2017.405 +11. Gordo, A., Almazan, J., Revaud, J., Larlus, D.: Deep image retrieval: Learning global representations for image search (2016) +12. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: CVPR. pp. 770-778 (2016) +13. Hendrycks, D., Dietterich, T.: Benchmarking neural network robustness to common corruptions and perturbations. Proceedings of the International Conference on Learning Representations (2019) +14. Jaesung, R., Haeyun, L., Jucheol, W., Sunghyun, C.: Real-world blur dataset for learning and benchmarking deblurring algorithms. In: ECCV (2020) +15. Jégou, H., Perronnin, F., Douze, M., Sánchez, J., Pérez, P., Schmid, C.: Aggregating local image descriptors into compact codes. IEEE TPAMI 34(9), 1704-1716 (2011) +16. Kalantidis, Y., Mellina, C., Osindero, S.: Cross-dimensional weighting for aggregated deep convolutional features. In: ECCVW. pp. 685-701. Springer (2016) +17. Kingma, D.P., Ba, J.: Adam: A method for stochastic optimization. In: Bengio, Y., LeCun, Y. (eds.) ICLR (2015), http://arxiv.org/abs/1412.6980 +18. Kotera, J., Rozumnyi, D., Sroubek, F., Matas, J.: Intra-frame object tracking by deblatting. In: ICCVW (Oct 2019) + +19. Kupyn, O., Budzan, V., Mykhailych, M., Mishkin, D., Matas, J.: Deblurgan: Blind motion deblurring using conditional adversarial networks. In: CVPR (June 2018) +20. Lee, S., Seong, H., Lee, S., Kim, E.: Correlation verification for image retrieval. In: CVPR. pp. 5374-5384 (2022) +21. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dólar, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: ECCV. pp. 740-755. Springer (2014) +22. Lowe, D.G.: Distinctive image features from scale-invariant keypoints. IJCV 60, 91-110 (2004) +23. Michaelis, C., Mitzkus, B., Geirhos, R., Rusak, E., Bringmann, O., Ecker, A.S., Bethge, M., Brendel, W.: Benchmarking robustness in object detection: Autonomous driving when winter is coming. arXiv preprint arXiv:1907.07484 (2019) +24. Nah, S., Kim, T.H., Lee, K.M.: Deep multi-scale convolutional neural network for dynamic scene deblurring. In: CVPR (July 2017) +25. Noh, H., Araujo, A., Sim, J., Weyand, T., Han, B.: Large-scale image retrieval with attentive deep local features. In: ICCV. pp. 3456-3465 (2017) +26. Paszke, A., Gross, S., Massa, F., Lerer, A., Bradbury, J., Chanan, G., Killeen, T., Lin, Z., Gimelshein, N., Antiga, L., Desmaison, A., Kopf, A., Yang, E., Devito, Z., Raison, M., Tejani, A., Chilamkurthy, S., Steiner, B., Fang, L., Bai, J., Chintala, S.: Pytorch: An imperative style, high-performance deep learning library. In: Wallach, H., Larochelle, H., Beygelzimer, A., Alché-Buc, F., Fox, E., Garnett, R. (eds.) NeurIPS, pp. 8024-8035. Curran Associates, Inc. (2019), http://papers.neurips.cc/paper/9015-pytorch-an-imperative-style-high-performance-deep-learning-library.pdf +27. Perronnin, F., Liu, Y., Sánchez, J., Poirier, H.: Large-scale image retrieval with compressed fisher vectors. In: CVPR. pp. 3384-3391. IEEE (2010) +28. Radenović, F., Iscen, A., Tolias, G., Avrithis, Y., Chum, O.: Revisiting oxford and paris: Large-scale image retrieval benchmarking. In: CVPR. pp. 5706-5715 (2018) +29. Radenović, F., Tobias, G., Chum, O.: Fine-tuning cnn image retrieval with no human annotation. arXiv preprint arXiv:1711.02512 (2017) +30. Radenović, F., Tolias, G., Chum, O.: Fine-tuning cnn image retrieval with no human annotation. IEEE TPAMI 41(7), 1655-1668 (2019). https://doi.org/10.1109/TPAMI.2018.2846566 +31. Rozumnyi, D., Kotera, J., Sroubek, F., Matas, J.: Non-causal tracking by deblatting. In: Fink, G.A., Frintrop, S., Jiang, X. (eds.) GCPR. pp. 122-135. Springer International Publishing, Cham (2019) +32. Rozumnyi, D., Kotera, J., Sroubek, F., Novotny, L., Matas, J.: The world of fast moving objects. In: CVPR. pp. 4838-4846 (July 2017). https://doi.org/10.1109/CVPR.2017.514 +33. Rozumnyi, D., Kotera, J., Sroubek, F., Matas, J.: Sub-frame appearance and 6d pose estimation of fast moving objects. In: CVPR. pp. 6777-6785 (2020) +34. Rozumnyi, D., Oswald, M.R., Ferrari, V., Matas, J., Pollefeys, M.: Defmo: Deblurring and shape recovery of fast moving objects. In: CVPR. Nashville, Tennessee, USA (Jun 2021) +35. Rozumnyi, D., Oswald, M.R., Ferrari, V., Pollefeys, M.: Shape from blur: Recovering textured 3d shape and motion of fast moving objects. In: NeurIPS (2021) +36. Rozumnyi, D., Oswald, M.R., Ferrari, V., Pollefeys, M.: Motion-from-blur: 3d shape and motion estimation of motion-blurred objects in videos. In: CVPR (Jun 2022) +37. Sayed, M., Brostow, G.: Improved handling of motion blur in online object detection. In: CVPR. pp. 1706-1716 (June 2021) + +38. Sivic, Zisserman: Video google: A text retrieval approach to object matching in videos. In: ICCV. pp. 1470-1477. IEEE (2003) +39. Skorokhodov, I., Sotnikov, G., Elhoseiny, M.: Aligning latent and image spaces to connect the unconnectable. In: ICCV. pp. 14144-14153 (2021) +40. Tolias, G., Avrithis, Y., Jégou, H.: Image search with selective match kernels: aggregation across single and multiple images. IJCV 116, 247-261 (2016) +41. Tolias, G., Jenicek, T., Chum, O.: Learning and aggregating deep local descriptors for instance-level recognition. In: ECCV. pp. 460-477. Springer (2020) +42. Tolias, G., Sicre, R., Jégou, H.: Particular object retrieval with integral max-pooling of cnn activations. ICLR (2016) +43. Vasiljevic, I., Chakrabarti, A., Shakhnarovich, G.: Examining the impact of blur on recognition by convolutional networks. arXiv preprint arXiv:1611.05760 (2016) +44. Wang, Y., Lu, Y., Gao, Y., Wang, L., Zhong, Z., Zheng, Y., Yamashita, A.: Efficient video deblurring guided by motion magnitude. In: ECCV. pp. 413-429. Springer (2022) +45. Weyand, T., Araujo, A., Cao, B., Sim, J.: Google landmarks dataset v2-a large-scale benchmark for instance-level recognition and retrieval. In: CVPR. pp. 2574-2583 (2020) +46. Wu, H., Wang, M., Zhou, W.T., Hu, Y., Li, H.: Learning token-based representation for image retrieval. In: AAAI Conference on Artificial Intelligence (2021), https://apisemantic scholar.org/CorpusID:245124518 +47. Yang, M., He, D., Fan, M., Shi, B., Xue, X., Li, F., Ding, E., Huang, J.: Dolg: Single-stage image retrieval with deep orthogonal fusion of local and global features. In: ICCV. pp. 11752-11761. IEEE Computer Society, Los Alamitos, CA, USA (oct 2021). https://doi.org/10.1109/ICCV48922.2021.01156, https://doi.ieeecomputersociety.org/10.1109/ICCV48922.2021.01156 +48. Yokoo, S., Ozaki, K., Simo-Serra, E., Iizuka, S.: Two-stage discriminative reranking for large-scale landmark retrieval. In: CVPRW. IEEE (2020) +49. Zhang, H., Xie, H., Yao, H.: Spatio-temporal deformable attention network for video deblurring. In: ECCV. pp. 581-596. Springer (2022) +50. Zhou, S., Zhang, J., Pan, J., Xie, H., Zuo, W., Ren, J.: Spatio-temporal filter adaptive network for video deblurring. In: ICCV (2019) \ No newline at end of file diff --git a/2024/Retrieval Robust to Object Motion Blur/images.zip b/2024/Retrieval Robust to Object Motion Blur/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..8f1233334a010b6cf601fd79ccc2c89dc8156944 --- /dev/null +++ b/2024/Retrieval Robust to Object Motion Blur/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:add1cf5558c23b45ad101482b15e46c5c0ffb7920b17d85fa5289b78849989d5 +size 803073 diff --git a/2024/Retrieval Robust to Object Motion Blur/layout.json b/2024/Retrieval Robust to Object Motion Blur/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..dac86b6d056dd81245e9641608564c450fe698dc --- /dev/null +++ b/2024/Retrieval Robust to Object Motion Blur/layout.json @@ -0,0 +1,13575 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 165, + 112, + 448, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 165, + 112, + 448, + 129 + ], + "spans": [ + { + "bbox": [ + 165, + 112, + 448, + 129 + ], + "type": "text", + "content": "Retrieval Robust to Object Motion Blur" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 175, + 150, + 436, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 175, + 150, + 436, + 163 + ], + "spans": [ + { + "bbox": [ + 175, + 150, + 436, + 163 + ], + "type": "text", + "content": "Rong Zou1, Marc Pollefeys1,2, and Denys Rozumnyi1,3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 174, + 171, + 438, + 183 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 174, + 171, + 438, + 183 + ], + "spans": [ + { + "bbox": [ + 174, + 171, + 438, + 183 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 174, + 171, + 438, + 183 + ], + "type": "text", + "content": "ETH Zürich, " + }, + { + "bbox": [ + 174, + 171, + 438, + 183 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 174, + 171, + 438, + 183 + ], + "type": "text", + "content": "Microsoft, " + }, + { + "bbox": [ + 174, + 171, + 438, + 183 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 174, + 171, + 438, + 183 + ], + "type": "text", + "content": "Czech Technical University in Prague" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 160, + 198, + 452, + 363 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 198, + 452, + 363 + ], + "spans": [ + { + "bbox": [ + 160, + 198, + 452, + 363 + ], + "type": "text", + "content": "Abstract. Moving objects are frequently seen in daily life and usually appear blurred in images due to their motion. While general object retrieval is a widely explored area in computer vision, it primarily focuses on sharp and static objects, and retrieval of motion-blurred objects in large image collections remains unexplored. We propose a method for object retrieval in images that are affected by motion blur. The proposed method learns a robust representation capable of matching blurred objects to their deblurred versions and vice versa. To evaluate our approach, we present the first large-scale datasets for blurred object retrieval, featuring images with objects exhibiting varying degrees of blur in various poses and scales. We conducted extensive experiments, showing that our method outperforms state-of-the-art retrieval methods on the new blur-retrieval datasets, which validates the effectiveness of the proposed approach. Code, data, and model are available at https://github.com/Rong-Zou/Retrieval-Robust-to-Object-Motion-Blur." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 160, + 373, + 404, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 373, + 404, + 384 + ], + "spans": [ + { + "bbox": [ + 160, + 373, + 404, + 384 + ], + "type": "text", + "content": "Keywords: object retrieval " + }, + { + "bbox": [ + 160, + 373, + 404, + 384 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 373, + 404, + 384 + ], + "type": "text", + "content": " object motion blur " + }, + { + "bbox": [ + 160, + 373, + 404, + 384 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 373, + 404, + 384 + ], + "type": "text", + "content": " datasets" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 403, + 230, + 415 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 403, + 230, + 415 + ], + "spans": [ + { + "bbox": [ + 132, + 403, + 230, + 415 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 426, + 482, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 426, + 482, + 641 + ], + "spans": [ + { + "bbox": [ + 130, + 426, + 482, + 641 + ], + "type": "text", + "content": "The task of image retrieval has long been a focal point in computer vision research, aiming to efficiently identify images within a large-scale database that share similarities with a given query image. In the realm of image retrieval, methods are often categorized into category-level and instance-level retrieval. We focus on the latter, where the objective is to identify images of a specific object featured in a query within a vast, unordered collection of database images. In this task, image representations play an important role since they capture the essence of the objects in images and enable the measurement of their similarities. Representing images based on hand-crafted features has been extensively studied [3,22]. Owing to the development in deep learning, recent strides have been made in adopting learning-based approaches to encode image content [2,4,20,25,30,41,42,46,47]. However, the current literature focuses exclusively on image representations for sharp objects, assuming that the objects are stationary and clearly visible in images. By only considering sharp and static objects, the existing retrieval methods fall short in addressing the dynamic nature of real-world scenarios, where objects are in movement, resulting in images blurred by object motion — a phenomenon prevalent in environments with fast-moving objects [32,34], low camera frame rates, and extended exposure times." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 642, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 642, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 642, + 482, + 666 + ], + "type": "text", + "content": "To bridge this gap, we introduce a novel task in image retrieval: retrieval with object motion blur. The challenge of this task lies in finding image representations" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 138, + 147, + 192, + 190 + ], + "blocks": [ + { + "bbox": [ + 138, + 147, + 192, + 190 + ], + "lines": [ + { + "bbox": [ + 138, + 147, + 192, + 190 + ], + "spans": [ + { + "bbox": [ + 138, + 147, + 192, + 190 + ], + "type": "image", + "image_path": "e5851576a0a04cc2b93fcfeace25333c07213c3d62bdcd0f609441a55d80beea.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 138, + 232, + 192, + 274 + ], + "blocks": [ + { + "bbox": [ + 138, + 232, + 192, + 274 + ], + "lines": [ + { + "bbox": [ + 138, + 232, + 192, + 274 + ], + "spans": [ + { + "bbox": [ + 138, + 232, + 192, + 274 + ], + "type": "image", + "image_path": "e63d0c979c7d7904bc7a10e2bf78ff5ae5fcbc0cb23c50096e60ea63e030a6fc.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 194, + 125, + 248, + 167 + ], + "blocks": [ + { + "bbox": [ + 151, + 115, + 179, + 125 + ], + "lines": [ + { + "bbox": [ + 151, + 115, + 179, + 125 + ], + "spans": [ + { + "bbox": [ + 151, + 115, + 179, + 125 + ], + "type": "text", + "content": "Query" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 194, + 125, + 248, + 167 + ], + "lines": [ + { + "bbox": [ + 194, + 125, + 248, + 167 + ], + "spans": [ + { + "bbox": [ + 194, + 125, + 248, + 167 + ], + "type": "image", + "image_path": "b00e63097785baab53a42b295c90ab69424271df59bb8a6d748b9cbec0dace1b.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 195, + 167, + 248, + 209 + ], + "blocks": [ + { + "bbox": [ + 195, + 167, + 248, + 209 + ], + "lines": [ + { + "bbox": [ + 195, + 167, + 248, + 209 + ], + "spans": [ + { + "bbox": [ + 195, + 167, + 248, + 209 + ], + "type": "image", + "image_path": "e7208f9d984d549405dec0a8a8e773922a8865fa6f6cc7c6090e7aaed5d18b4c.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 195, + 210, + 248, + 251 + ], + "blocks": [ + { + "bbox": [ + 195, + 210, + 248, + 251 + ], + "lines": [ + { + "bbox": [ + 195, + 210, + 248, + 251 + ], + "spans": [ + { + "bbox": [ + 195, + 210, + 248, + 251 + ], + "type": "image", + "image_path": "4e705d91bde792b69dc765b0acaace0d17dd8a3577dd91f3ba55780698c0fab5.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 195, + 252, + 248, + 293 + ], + "blocks": [ + { + "bbox": [ + 195, + 252, + 248, + 293 + ], + "lines": [ + { + "bbox": [ + 195, + 252, + 248, + 293 + ], + "spans": [ + { + "bbox": [ + 195, + 252, + 248, + 293 + ], + "type": "image", + "image_path": "40783217d654850d6ad97c67fbfe2ca0a9f9c84fc2112b4ef8211c7b08f6aaf7.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 295, + 481, + 350 + ], + "lines": [ + { + "bbox": [ + 132, + 295, + 481, + 350 + ], + "spans": [ + { + "bbox": [ + 132, + 295, + 481, + 350 + ], + "type": "text", + "content": "Fig. 1: Retrieval robust to object motion blur. We present the first approach designed for retrieval in the presence of object motion blur. Our method captures blur-invariant features and generates blur-robust representations. We show retrieval results on the newly created datasets for the task of blur-robust retrieval: synthetic (top) and real (bottom). All retrieved images match the query and are marked with green boxes." + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 251, + 125, + 304, + 167 + ], + "blocks": [ + { + "bbox": [ + 286, + 115, + 381, + 125 + ], + "lines": [ + { + "bbox": [ + 286, + 115, + 381, + 125 + ], + "spans": [ + { + "bbox": [ + 286, + 115, + 381, + 125 + ], + "type": "text", + "content": "Top 10 retrieval results" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 251, + 125, + 304, + 167 + ], + "lines": [ + { + "bbox": [ + 251, + 125, + 304, + 167 + ], + "spans": [ + { + "bbox": [ + 251, + 125, + 304, + 167 + ], + "type": "image", + "image_path": "574e1070ab18bb9ba8340f7146292622aada84baa328454121d807ef710bde1b.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 251, + 167, + 304, + 209 + ], + "blocks": [ + { + "bbox": [ + 251, + 167, + 304, + 209 + ], + "lines": [ + { + "bbox": [ + 251, + 167, + 304, + 209 + ], + "spans": [ + { + "bbox": [ + 251, + 167, + 304, + 209 + ], + "type": "image", + "image_path": "40d57a3714f9f975241ac1cbc4080aea20390cf1ceefcc80ffb5db007ebc195f.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 251, + 210, + 304, + 251 + ], + "blocks": [ + { + "bbox": [ + 251, + 210, + 304, + 251 + ], + "lines": [ + { + "bbox": [ + 251, + 210, + 304, + 251 + ], + "spans": [ + { + "bbox": [ + 251, + 210, + 304, + 251 + ], + "type": "image", + "image_path": "52b982c77ccb4aaa5a0f91fef2f26d8b314d53981ae47741b94c37d732391ab9.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 251, + 252, + 304, + 293 + ], + "blocks": [ + { + "bbox": [ + 251, + 252, + 304, + 293 + ], + "lines": [ + { + "bbox": [ + 251, + 252, + 304, + 293 + ], + "spans": [ + { + "bbox": [ + 251, + 252, + 304, + 293 + ], + "type": "image", + "image_path": "9b44bec1c9c2c3ee2d1f2af2d0ed24223d58c6660f6f1432cf4099f763ed1b19.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 307, + 125, + 360, + 167 + ], + "blocks": [ + { + "bbox": [ + 307, + 125, + 360, + 167 + ], + "lines": [ + { + "bbox": [ + 307, + 125, + 360, + 167 + ], + "spans": [ + { + "bbox": [ + 307, + 125, + 360, + 167 + ], + "type": "image", + "image_path": "c586f9365b2165fe63a5b7359f51847adb1c2df5036553a4388306886de0d0a3.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 307, + 167, + 360, + 209 + ], + "blocks": [ + { + "bbox": [ + 307, + 167, + 360, + 209 + ], + "lines": [ + { + "bbox": [ + 307, + 167, + 360, + 209 + ], + "spans": [ + { + "bbox": [ + 307, + 167, + 360, + 209 + ], + "type": "image", + "image_path": "ebcf9d7f5e9eefa8660749610112ea9f4864dc2d831fb50f008e56289ac71ddc.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 307, + 210, + 360, + 251 + ], + "blocks": [ + { + "bbox": [ + 307, + 210, + 360, + 251 + ], + "lines": [ + { + "bbox": [ + 307, + 210, + 360, + 251 + ], + "spans": [ + { + "bbox": [ + 307, + 210, + 360, + 251 + ], + "type": "image", + "image_path": "2490eb1ce51b54e495c33b0da719918fd1116f0afbb8e3f6fe2242d308bc121b.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 307, + 252, + 360, + 293 + ], + "blocks": [ + { + "bbox": [ + 307, + 252, + 360, + 293 + ], + "lines": [ + { + "bbox": [ + 307, + 252, + 360, + 293 + ], + "spans": [ + { + "bbox": [ + 307, + 252, + 360, + 293 + ], + "type": "image", + "image_path": "67e1059172bf8e1c60c7b7fe3be40d2791ff5831a50eb875fbcf31bc51b7829d.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 362, + 125, + 416, + 167 + ], + "blocks": [ + { + "bbox": [ + 362, + 125, + 416, + 167 + ], + "lines": [ + { + "bbox": [ + 362, + 125, + 416, + 167 + ], + "spans": [ + { + "bbox": [ + 362, + 125, + 416, + 167 + ], + "type": "image", + "image_path": "3b829b67ef521d45aa31806447fb233ef732f44998ad749e73024fa7b3fbca2a.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 362, + 167, + 416, + 209 + ], + "blocks": [ + { + "bbox": [ + 362, + 167, + 416, + 209 + ], + "lines": [ + { + "bbox": [ + 362, + 167, + 416, + 209 + ], + "spans": [ + { + "bbox": [ + 362, + 167, + 416, + 209 + ], + "type": "image", + "image_path": "194f5851dc0176ba4dd5548e708e5fe84e344ee68720f4bc9a82831865a0822a.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 362, + 210, + 416, + 251 + ], + "blocks": [ + { + "bbox": [ + 362, + 210, + 416, + 251 + ], + "lines": [ + { + "bbox": [ + 362, + 210, + 416, + 251 + ], + "spans": [ + { + "bbox": [ + 362, + 210, + 416, + 251 + ], + "type": "image", + "image_path": "9cfcdb3a3294850bc9844cdb76c8c743972dbc7bbadfc4b54e70da314000db36.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 362, + 252, + 416, + 293 + ], + "blocks": [ + { + "bbox": [ + 362, + 252, + 416, + 293 + ], + "lines": [ + { + "bbox": [ + 362, + 252, + 416, + 293 + ], + "spans": [ + { + "bbox": [ + 362, + 252, + 416, + 293 + ], + "type": "image", + "image_path": "32201676e556d5003050b8c881a61bfc979b5e35bc03d5b6851ed93a2fd62097.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 418, + 125, + 472, + 167 + ], + "blocks": [ + { + "bbox": [ + 418, + 125, + 472, + 167 + ], + "lines": [ + { + "bbox": [ + 418, + 125, + 472, + 167 + ], + "spans": [ + { + "bbox": [ + 418, + 125, + 472, + 167 + ], + "type": "image", + "image_path": "7a513e771ef7cdbf33d3c006c877b76aadf2d581e395aa8abebf722c8430b207.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 418, + 167, + 472, + 209 + ], + "blocks": [ + { + "bbox": [ + 418, + 167, + 472, + 209 + ], + "lines": [ + { + "bbox": [ + 418, + 167, + 472, + 209 + ], + "spans": [ + { + "bbox": [ + 418, + 167, + 472, + 209 + ], + "type": "image", + "image_path": "c1b2d150eb6a2a6f4f73a076e08c4f6443ddf9ea793a7e9b0f67bf399b1fd77b.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 418, + 210, + 472, + 251 + ], + "blocks": [ + { + "bbox": [ + 418, + 210, + 472, + 251 + ], + "lines": [ + { + "bbox": [ + 418, + 210, + 472, + 251 + ], + "spans": [ + { + "bbox": [ + 418, + 210, + 472, + 251 + ], + "type": "image", + "image_path": "ae9b24a76edafa4dc45f4c9a5289659513968840ef374f6456f5c7e38443de9a.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 418, + 252, + 472, + 293 + ], + "blocks": [ + { + "bbox": [ + 418, + 252, + 472, + 293 + ], + "lines": [ + { + "bbox": [ + 418, + 252, + 472, + 293 + ], + "spans": [ + { + "bbox": [ + 418, + 252, + 472, + 293 + ], + "type": "image", + "image_path": "b52349e974aa6ca3fc8ab09e6276b0b938b93939feba35a9e60e2757f41b8209.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + } + ], + "index": 25 + }, + { + "bbox": [ + 130, + 363, + 480, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 363, + 480, + 434 + ], + "spans": [ + { + "bbox": [ + 130, + 363, + 480, + 434 + ], + "type": "text", + "content": "robust to object motion blur, in addition to other variations such as background and changes in viewpoint. We also propose the first approach targeting this task that produces representations invariant to motion blur. In response to the unique challenge posed by object motion, we propose loss functions designed to enhance the model's comprehension of blur, thereby ensuring robust retrieval performance even under conditions of extreme blurring, as seen in Fig. 1." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 130, + 436, + 480, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 436, + 480, + 506 + ], + "spans": [ + { + "bbox": [ + 130, + 436, + 480, + 506 + ], + "type": "text", + "content": "Recognizing the absence of datasets accommodating images with motion-blurred objects in existing retrieval benchmarks, we introduce the first benchmark datasets for this task, including a synthetic dataset for training and evaluation and a real-world dataset created specifically for evaluation. By re-training and evaluating state-of-the-art instance retrieval methods on this comprehensive benchmark, we reveal their shortcomings in handling motion-blurred scenarios." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 130, + 507, + 480, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 507, + 480, + 542 + ], + "spans": [ + { + "bbox": [ + 130, + 507, + 480, + 542 + ], + "type": "text", + "content": "This work marks a key step in advancing practical methods for object retrieval in the challenging presence of motion blur. It has potential applications such as endangered wildlife monitoring, sports analysis, and security surveillance." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 146, + 543, + 381, + 554 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 543, + 381, + 554 + ], + "spans": [ + { + "bbox": [ + 146, + 543, + 381, + 554 + ], + "type": "text", + "content": "To summarize, our main contributions are as follows:" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 132, + 558, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 34, + "blocks": [ + { + "bbox": [ + 132, + 558, + 480, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 558, + 480, + 593 + ], + "spans": [ + { + "bbox": [ + 132, + 558, + 480, + 593 + ], + "type": "text", + "content": "(1) We present the first method specifically designed for the novel retrieval task involving object motion blur. The method is trained with loss functions tailored to improve its understanding of motion blur." + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 132, + 594, + 480, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 594, + 480, + 628 + ], + "spans": [ + { + "bbox": [ + 132, + 594, + 480, + 628 + ], + "type": "text", + "content": "(2) We introduce a new benchmark featuring synthetic and real-world datasets specifically constructed for this task. The datasets are carefully processed and are directly applicable for future research in blur retrieval." + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 132, + 629, + 480, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 629, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 629, + 480, + 665 + ], + "type": "text", + "content": "(3) We conducted extensive experiments and detailed analyses to validate the effectiveness of our approach. Our method outperforms state-of-the-art retrieval methods and demonstrates superior robustness to motion blur." + } + ] + } + ], + "index": 33 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "text", + "content": "R. Zou et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 234, + 127 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 234, + 127 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 234, + 127 + ], + "type": "text", + "content": "2 Related work" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 137, + 482, + 316 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 137, + 482, + 316 + ], + "spans": [ + { + "bbox": [ + 130, + 137, + 482, + 316 + ], + "type": "text", + "content": "Standard retrieval. In the area of standard image retrieval, local feature-based methods have evolved from traditional hand-crafted techniques like SIFT [22] and SURF [3] to contemporary deep learning-based approaches such as DELF [25] and HOW [41]. Considering that motion blur can corrupt local information, which may make local descriptors less robust, we adopt a global approach for our task. Classical methods typically derive global features by aggregating hand-crafted local features using techniques such as BoW [38], Fisher vectors [27], VLAD [15], AMSK [40], etc. Modern deep learning-based approaches, on the other hand, generate global descriptors by applying various pooling operations on feature maps encoded by convolutional neural networks (CNNs), such as sum pooling [1,41], weighted sum pooling [16], max pooling, regional max pooling [42], and generalized mean-pooling (GeM) [30], which has gained prominence in recent literature [4,20,47]. Our method integrates insights from recent advancements in global feature-based methods, adopting GeM pooling [30] for feature extraction and incorporating contrastive loss [6] and ArcFace loss [9] during training." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 317, + 482, + 412 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 317, + 482, + 412 + ], + "spans": [ + { + "bbox": [ + 130, + 317, + 482, + 412 + ], + "type": "text", + "content": "Among the single-pass global feature-based methods, state-of-the-art standard retrieval approaches such as DELG global branch [4], DOLG [47], and Token [46] face challenges when handling images affected by motion blur, leading to degraded performance. Our work stands out in addressing this critical gap in the literature, as existing retrieval methods overlook the impact of motion blur. To bridge this research gap, we propose novel loss functions aimed at enhancing the model's comprehension of object motion blur, thereby contributing to the advancement of image retrieval in motion-blurred scenarios." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 413, + 482, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 413, + 482, + 557 + ], + "spans": [ + { + "bbox": [ + 130, + 413, + 482, + 557 + ], + "type": "text", + "content": "Blur handling. Handling blur in images is an important topic in computer vision. [43] investigated the effect of different types of blur (defocus, uniform subject motion, camera shake) on the performance of CNNs for image classification and semantic segmentation. [37] studied online object detection under the influence of camera ego-motion-induced blur. Image and video deblurring are long-standing tasks in computer vision, which have been addressed by a variety of methods [10,19,44,49,50]. Joint deblurring and trajectory estimation for motion-blurred objects have also been extensively explored, with notable contributions from methods like [18, 31, 33, 34], some extending their scope to 3D reconstruction [35,36]. These existing works predominantly focus on other types of blur or address distinct tasks, emphasizing the need for specialized efforts in the context of our targeted task of retrieval involving object motion blur." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "text", + "content": "Datasets. Several datasets have been proposed that contain generic corruptions including blur, such as ImaneNet-C [13] and Coco-C [23], but they do not focus purely on blurry images, and the types of blur they encompass (Gaussian, defocus, frosted glass, camera shake, and zoom) are not induced by object motion. Many datasets contain blurry-sharp image pairs, such as RealBlur [14] and GoPro [24], but they are created for deblurring tasks and not for retrieval, and the blur in images is mainly caused by camera motion or camera shake, not by object motion. A large variety of standard retrieval datasets have been proposed, including Revisited Oxford5k and Paris6k [28], Google Landmarks V2 [45, 48]," + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 283, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 283, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 283, + 91, + 447, + 102 + ], + "type": "text", + "content": "Retrieval Robust to Object Motion Blur" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 132, + 115, + 481, + 223 + ], + "blocks": [ + { + "bbox": [ + 132, + 115, + 481, + 223 + ], + "lines": [ + { + "bbox": [ + 132, + 115, + 481, + 223 + ], + "spans": [ + { + "bbox": [ + 132, + 115, + 481, + 223 + ], + "type": "image", + "image_path": "c668d320ad6faf9330c604d900060fd3025451de5b1bf69b7feb94de94df8068.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 224, + 482, + 369 + ], + "lines": [ + { + "bbox": [ + 130, + 224, + 482, + 369 + ], + "spans": [ + { + "bbox": [ + 130, + 224, + 482, + 369 + ], + "type": "text", + "content": "Fig. 2: Overview of the proposed method for blur-robust retrieval. Our model takes a query image as input. Then, it generates a descriptor robust to object motion blur through the Blur Robust Image Descriptor Generator (BRIDGE) module that encodes the input image using a CNN backbone followed by generalized mean-pooling (GeM) [30], extracting a feature vector. This vector is subsequently fed into three heads: the blur estimation head, classification head, and localization head. Each head processes the feature vector to extract relevant information for identifying motion-blurred objects. The resulting three processed feature vectors are concatenated and passed through the final fully connected (FC) layer to produce the blur-robust descriptor. During training, the Blur Level-based Image Sample Selection (BLISS) mechanism is employed to select contrastive samples based on the query blur level (BL, Eq. (6)) " + }, + { + "bbox": [ + 130, + 224, + 482, + 369 + ], + "type": "inline_equation", + "content": "b_{q}" + }, + { + "bbox": [ + 130, + 224, + 482, + 369 + ], + "type": "text", + "content": " and a specified blur level range " + }, + { + "bbox": [ + 130, + 224, + 482, + 369 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 130, + 224, + 482, + 369 + ], + "type": "text", + "content": ". Then, these selected image samples are input into the BRIDGE module to extract descriptors for subsequent contrastive learning." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 389, + 482, + 439 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 389, + 482, + 439 + ], + "spans": [ + { + "bbox": [ + 130, + 389, + 482, + 439 + ], + "type": "text", + "content": "SfM120K [29], etc. However, none of them incorporate motion-blurred images. To the best of our knowledge, there is currently no existing dataset tailored for the task of retrieval with object motion blur. In view of this, we created synthetic and real-world datasets specifically designed for the novel task of blur retrieval." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 470, + 202, + 483 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 470, + 202, + 483 + ], + "spans": [ + { + "bbox": [ + 132, + 470, + 202, + 483 + ], + "type": "text", + "content": "3 Method" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 498, + 481, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 498, + 481, + 569 + ], + "spans": [ + { + "bbox": [ + 130, + 498, + 481, + 569 + ], + "type": "text", + "content": "We perform retrieval via nearest neighbor search in a representation space. Therefore, we require a representation model that is robust to object motion blur. We train a deep learning-based model that maps images to a vector space. The network is trained on our proposed synthetic dataset using multiple losses, including those tailored to object motion blur. As we show, the method generalizes well to real data despite being trained on synthetic data." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "type": "text", + "content": "The network takes an RGB image " + }, + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{I} \\in \\mathbb{R}^{3 \\times H \\times W}" + }, + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "type": "text", + "content": " containing a sharp or blurred object as input and generates a descriptor " + }, + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{D} \\in \\mathbb{R}^d" + }, + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "type": "text", + "content": " robust to object motion blur through the Blur Robust Image Descriptor Generator module (BRIDGE, Sec. 3.1). For training only, contrastive samples are selected by the Blur Level-based Image Sample Selection mechanism (BLISS, Sec. 3.2) based on the query blur level " + }, + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "type": "inline_equation", + "content": "b_q" + }, + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "type": "text", + "content": " and the contrast blur level range " + }, + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 130, + 569, + 482, + 666 + ], + "type": "text", + "content": ". The selected image samples are sent into the BRIDGE module to extract descriptors for contrastive learning. The architecture of the proposed blur retrieval method is shown in Fig. 2." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "text", + "content": "R. Zou et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 248, + 126 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 248, + 126 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 248, + 126 + ], + "type": "text", + "content": "3.1 BRIDGE module" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 134, + 482, + 183 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 134, + 482, + 183 + ], + "spans": [ + { + "bbox": [ + 130, + 134, + 482, + 183 + ], + "type": "text", + "content": "The BRIDGE module encodes an image into a descriptor that is robust to object motion blur. The input image is encoded by a ResNet-50 backbone [12] to produce a feature map " + }, + { + "bbox": [ + 130, + 134, + 482, + 183 + ], + "type": "inline_equation", + "content": "\\Omega \\in \\mathbb{R}^{C\\times U\\times V}" + }, + { + "bbox": [ + 130, + 134, + 482, + 183 + ], + "type": "text", + "content": ". Then, we use GeM pooling [30] to process the feature map into a feature vector " + }, + { + "bbox": [ + 130, + 134, + 482, + 183 + ], + "type": "inline_equation", + "content": "\\mathbf{f}\\in \\mathbb{R}^C" + }, + { + "bbox": [ + 130, + 134, + 482, + 183 + ], + "type": "text", + "content": ", expressed as" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 219, + 190, + 481, + 236 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 219, + 190, + 481, + 236 + ], + "spans": [ + { + "bbox": [ + 219, + 190, + 481, + 236 + ], + "type": "interline_equation", + "content": "\\mathbf {f} = \\left[ \\left(\\frac {1}{| \\boldsymbol {\\Omega} |} \\sum_ {(u, v) \\in \\boldsymbol {\\Omega}} x _ {(c, u, v)} ^ {p}\\right) ^ {\\frac {1}{p}} \\right] _ {c = 1, \\dots , C}, \\tag {1}", + "image_path": "9ddc71ef521c88dd31ce0ecb5db86ab787e90df0d5f0f433c792e5da527a19fd.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 243, + 482, + 362 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 243, + 482, + 362 + ], + "spans": [ + { + "bbox": [ + 131, + 243, + 482, + 362 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 243, + 482, + 362 + ], + "type": "inline_equation", + "content": "p > 0" + }, + { + "bbox": [ + 131, + 243, + 482, + 362 + ], + "type": "text", + "content": " is a parameter, and setting it to values larger than 1 forces the pooled vector to focus on the salient areas. This feature vector then undergoes multi-head processing to generate a motion blur-robust descriptor. Specifically, in the blur estimation head, we obtain blur estimation features " + }, + { + "bbox": [ + 131, + 243, + 482, + 362 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_{be} \\in \\mathbb{R}^{C_b}" + }, + { + "bbox": [ + 131, + 243, + 482, + 362 + ], + "type": "text", + "content": " by a fully connected (FC) layer. The same dimension reduction is performed in the localization head, which generates features for location prediction " + }, + { + "bbox": [ + 131, + 243, + 482, + 362 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_{loc} \\in \\mathbb{R}^{C_l}" + }, + { + "bbox": [ + 131, + 243, + 482, + 362 + ], + "type": "text", + "content": ". The features generated from the two heads are then concatenated with the whitened [11] features " + }, + { + "bbox": [ + 131, + 243, + 482, + 362 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_{cls} \\in \\mathbb{R}^{C_c}" + }, + { + "bbox": [ + 131, + 243, + 482, + 362 + ], + "type": "text", + "content": " from the classification head. We further reduce the dimension of the concatenated features to obtain the final blur-robust descriptor of the image. This can be expressed as:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 220, + 370, + 481, + 384 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 220, + 370, + 481, + 384 + ], + "spans": [ + { + "bbox": [ + 220, + 370, + 481, + 384 + ], + "type": "interline_equation", + "content": "\\mathbf {D} = \\mathbf {W} \\cdot \\operatorname {C o n c a t} \\left(\\mathbf {W} _ {b e} \\mathbf {f}, \\mathbf {W} _ {l o c} \\mathbf {f}, \\mathbf {W} _ {c l s} \\mathbf {f}\\right), \\tag {2}", + "image_path": "d161b98850f2201e4b97112294c00ce108233a1bcb4a834d70d8b3e0c94ef04f.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 389, + 482, + 428 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 389, + 482, + 428 + ], + "spans": [ + { + "bbox": [ + 131, + 389, + 482, + 428 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 389, + 482, + 428 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_{be} \\in \\mathbb{R}^{C_b \\times C}" + }, + { + "bbox": [ + 131, + 389, + 482, + 428 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 131, + 389, + 482, + 428 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_{loc} \\in \\mathbb{R}^{C_l \\times C}" + }, + { + "bbox": [ + 131, + 389, + 482, + 428 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 389, + 482, + 428 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_{cls} \\in \\mathbb{R}^{C_c \\times C}" + }, + { + "bbox": [ + 131, + 389, + 482, + 428 + ], + "type": "text", + "content": " are weights of the FC layers in the blur estimation, localization, and classification heads, respectively, and " + }, + { + "bbox": [ + 131, + 389, + 482, + 428 + ], + "type": "inline_equation", + "content": "\\mathbf{W} \\in \\mathbb{R}^{d \\times (C_b + C_l + C_c)}" + }, + { + "bbox": [ + 131, + 389, + 482, + 428 + ], + "type": "text", + "content": " is the weight of the final dimension-reduction layer." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 443, + 253, + 454 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 443, + 253, + 454 + ], + "spans": [ + { + "bbox": [ + 132, + 443, + 253, + 454 + ], + "type": "text", + "content": "3.2 BLISS mechanism" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 462, + 482, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 462, + 482, + 668 + ], + "spans": [ + { + "bbox": [ + 131, + 462, + 482, + 668 + ], + "type": "text", + "content": "While the specific representation of the degree of motion blur for objects in images is yet to be formally defined, it is intuitively perceived as a continuous value. The level of detail exhibited by objects varies with the degree of blur: sharper objects show clear structures, textures, and poses, while more blurred objects primarily reveal color characteristics (see Fig. 3). This continuous nature of blur allows us to indirectly control the distance between samples with significantly different blur conditions during learning image representations. In this context, we manage the distance between samples that exhibit visual features and patterns of the same level of detail due to similar degrees of blur, thus indirectly influencing the distance between samples with significantly different levels of blur. Additionally, we can directly control the distance between samples with large differences in blur conditions by imposing constraints directly between such samples in the descriptor space. There are therefore two key research questions: What is the optimal value for the range of differences in the blur levels of image pairs used in contrastive learning? How does adjusting the range of differences in blur levels affect the model's performance? To answer these questions and identify the optimal blur level range for contrastive learning, we introduce the BLISS" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 282, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 282, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 282, + 91, + 447, + 102 + ], + "type": "text", + "content": "Retrieval Robust to Object Motion Blur" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 136, + 114, + 159, + 123 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 114, + 159, + 123 + ], + "spans": [ + { + "bbox": [ + 136, + 114, + 159, + 123 + ], + "type": "text", + "content": "BL 1" + } + ] + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 134, + 125, + 162, + 168 + ], + "blocks": [ + { + "bbox": [ + 134, + 125, + 162, + 168 + ], + "lines": [ + { + "bbox": [ + 134, + 125, + 162, + 168 + ], + "spans": [ + { + "bbox": [ + 134, + 125, + 162, + 168 + ], + "type": "image", + "image_path": "6642b79676364e8c89cb675831934e6c3fde2d4eb2122b89c7729a17e3cd8fff.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 134, + 168, + 162, + 212 + ], + "blocks": [ + { + "bbox": [ + 134, + 168, + 162, + 212 + ], + "lines": [ + { + "bbox": [ + 134, + 168, + 162, + 212 + ], + "spans": [ + { + "bbox": [ + 134, + 168, + 162, + 212 + ], + "type": "image", + "image_path": "6ba913a8085179d9d0c064a8f9e2bf5098f5b0a8728148afd5c225d23b8fed49.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 135, + 213, + 162, + 256 + ], + "blocks": [ + { + "bbox": [ + 135, + 213, + 162, + 256 + ], + "lines": [ + { + "bbox": [ + 135, + 213, + 162, + 256 + ], + "spans": [ + { + "bbox": [ + 135, + 213, + 162, + 256 + ], + "type": "image", + "image_path": "7e0e2c791b150efe50efba756ff0351bf0cff18ee89f1e22fafc537f34bfa876.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 268, + 481, + 323 + ], + "lines": [ + { + "bbox": [ + 132, + 268, + 481, + 323 + ], + "spans": [ + { + "bbox": [ + 132, + 268, + 481, + 323 + ], + "type": "text", + "content": "Fig. 3: The introduced synthetic (a) and real-world (b) datasets. Rows 1-2: different trajectories of the same object. Rows 2-3: two different objects from the same category with similar shapes. Rows 4-6: objects from different categories but share similar textures. Columns correspond to different blur levels, from 1 to 6, and they are different for synthetic (BL) and real-world " + }, + { + "bbox": [ + 132, + 268, + 481, + 323 + ], + "type": "inline_equation", + "content": "(\\mathrm{BL}^{\\mathrm{r}})" + }, + { + "bbox": [ + 132, + 268, + 481, + 323 + ], + "type": "text", + "content": " datasets." + } + ] + } + ], + "index": 61, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 166, + 115, + 188, + 123 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 115, + 188, + 123 + ], + "spans": [ + { + "bbox": [ + 166, + 115, + 188, + 123 + ], + "type": "text", + "content": "BL2" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 164, + 124, + 191, + 168 + ], + "blocks": [ + { + "bbox": [ + 164, + 124, + 191, + 168 + ], + "lines": [ + { + "bbox": [ + 164, + 124, + 191, + 168 + ], + "spans": [ + { + "bbox": [ + 164, + 124, + 191, + 168 + ], + "type": "image", + "image_path": "0f633b48db9977093ed4859f535cff3bcf61ee5628e6c223098311ff5824c2b5.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 164, + 168, + 191, + 212 + ], + "blocks": [ + { + "bbox": [ + 164, + 168, + 191, + 212 + ], + "lines": [ + { + "bbox": [ + 164, + 168, + 191, + 212 + ], + "spans": [ + { + "bbox": [ + 164, + 168, + 191, + 212 + ], + "type": "image", + "image_path": "3978cfa63644f3da896ec7b2d4609c0f41ae7bcb1a1182b80c55f98439b034dd.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 164, + 212, + 191, + 256 + ], + "blocks": [ + { + "bbox": [ + 164, + 212, + 191, + 256 + ], + "lines": [ + { + "bbox": [ + 164, + 212, + 191, + 256 + ], + "spans": [ + { + "bbox": [ + 164, + 212, + 191, + 256 + ], + "type": "image", + "image_path": "b0f9afc4191a68846c2b13254c21719342c071fc30bfa92788ee9dc754782e56.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 174, + 257, + 260, + 266 + ], + "lines": [ + { + "bbox": [ + 174, + 257, + 260, + 266 + ], + "spans": [ + { + "bbox": [ + 174, + 257, + 260, + 266 + ], + "type": "text", + "content": "(a) Synthetic dataset." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 197, + 115, + 219, + 123 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 197, + 115, + 219, + 123 + ], + "spans": [ + { + "bbox": [ + 197, + 115, + 219, + 123 + ], + "type": "text", + "content": "BL 3" + } + ] + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 192, + 124, + 219, + 168 + ], + "blocks": [ + { + "bbox": [ + 192, + 124, + 219, + 168 + ], + "lines": [ + { + "bbox": [ + 192, + 124, + 219, + 168 + ], + "spans": [ + { + "bbox": [ + 192, + 124, + 219, + 168 + ], + "type": "image", + "image_path": "950a91c0f9c25a41324b5c339ebd9375e9dbb4f69d5273528c712f5c043428f9.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 192, + 168, + 219, + 212 + ], + "blocks": [ + { + "bbox": [ + 192, + 168, + 219, + 212 + ], + "lines": [ + { + "bbox": [ + 192, + 168, + 219, + 212 + ], + "spans": [ + { + "bbox": [ + 192, + 168, + 219, + 212 + ], + "type": "image", + "image_path": "2904c6d27de330c7d66f9e4086e13ba73ff6ce714659eae22fbd98d8bd3a35d4.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 192, + 212, + 219, + 256 + ], + "blocks": [ + { + "bbox": [ + 192, + 212, + 219, + 256 + ], + "lines": [ + { + "bbox": [ + 192, + 212, + 219, + 256 + ], + "spans": [ + { + "bbox": [ + 192, + 212, + 219, + 256 + ], + "type": "image", + "image_path": "31836addd03294e950c0f7f787519d790f957c029e912c561ea8b1a730214604.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 192, + 257, + 264, + 266 + ], + "lines": [ + { + "bbox": [ + 192, + 257, + 264, + 266 + ], + "spans": [ + { + "bbox": [ + 192, + 257, + 264, + 266 + ], + "type": "text", + "content": "Synthetic dataset." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "bbox": [ + 240, + 115, + 261, + 123 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 240, + 115, + 261, + 123 + ], + "spans": [ + { + "bbox": [ + 240, + 115, + 261, + 123 + ], + "type": "text", + "content": "4 B" + } + ] + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 222, + 124, + 249, + 168 + ], + "blocks": [ + { + "bbox": [ + 222, + 124, + 249, + 168 + ], + "lines": [ + { + "bbox": [ + 222, + 124, + 249, + 168 + ], + "spans": [ + { + "bbox": [ + 222, + 124, + 249, + 168 + ], + "type": "image", + "image_path": "76b58f60691c164ee6fccfb6f132896910fdb88e7811aa9f965e471627e3de40.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 221, + 168, + 248, + 212 + ], + "blocks": [ + { + "bbox": [ + 221, + 168, + 248, + 212 + ], + "lines": [ + { + "bbox": [ + 221, + 168, + 248, + 212 + ], + "spans": [ + { + "bbox": [ + 221, + 168, + 248, + 212 + ], + "type": "image", + "image_path": "c3d999b22e224f92773ff102fe420d91d6598caa7a29a474a43f407ceb2556c0.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 221, + 212, + 248, + 256 + ], + "blocks": [ + { + "bbox": [ + 221, + 212, + 248, + 256 + ], + "lines": [ + { + "bbox": [ + 221, + 212, + 248, + 256 + ], + "spans": [ + { + "bbox": [ + 221, + 212, + 248, + 256 + ], + "type": "image", + "image_path": "19babffa266ed2d505c33ae2a8230f98671482750f9f3ca3648c16cf7cc2b216.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "bbox": [ + 221, + 257, + 248, + 266 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 221, + 257, + 248, + 266 + ], + "spans": [ + { + "bbox": [ + 221, + 257, + 248, + 266 + ], + "type": "text", + "content": "ic datas" + } + ] + } + ], + "index": 20, + "type": "text" + }, + { + "bbox": [ + 252, + 115, + 275, + 123 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 252, + 115, + 275, + 123 + ], + "spans": [ + { + "bbox": [ + 252, + 115, + 275, + 123 + ], + "type": "text", + "content": "BL 5" + } + ] + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 251, + 124, + 277, + 168 + ], + "blocks": [ + { + "bbox": [ + 251, + 124, + 277, + 168 + ], + "lines": [ + { + "bbox": [ + 251, + 124, + 277, + 168 + ], + "spans": [ + { + "bbox": [ + 251, + 124, + 277, + 168 + ], + "type": "image", + "image_path": "561de1022d59e8c05bc3923c61a1f6ea1b4e34d208cf450139b38e2c85e2a0b3.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 251, + 168, + 277, + 212 + ], + "blocks": [ + { + "bbox": [ + 251, + 168, + 277, + 212 + ], + "lines": [ + { + "bbox": [ + 251, + 168, + 277, + 212 + ], + "spans": [ + { + "bbox": [ + 251, + 168, + 277, + 212 + ], + "type": "image", + "image_path": "937e81c6f721f3aa2092fe093b01de05730f31be1f82ae4bf48f9af4d84908c5.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 251, + 212, + 277, + 256 + ], + "blocks": [ + { + "bbox": [ + 251, + 212, + 277, + 256 + ], + "lines": [ + { + "bbox": [ + 251, + 212, + 277, + 256 + ], + "spans": [ + { + "bbox": [ + 251, + 212, + 277, + 256 + ], + "type": "image", + "image_path": "c2086dbcc11599aec790972550f698f97794bc37dbcac2ffccf25f85d4d69c5e.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 251, + 257, + 306, + 266 + ], + "lines": [ + { + "bbox": [ + 251, + 257, + 306, + 266 + ], + "spans": [ + { + "bbox": [ + 251, + 257, + 306, + 266 + ], + "type": "text", + "content": "et." + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + } + ], + "index": 24 + }, + { + "bbox": [ + 282, + 115, + 304, + 123 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 282, + 115, + 304, + 123 + ], + "spans": [ + { + "bbox": [ + 282, + 115, + 304, + 123 + ], + "type": "text", + "content": "BL6" + } + ] + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 281, + 124, + 307, + 168 + ], + "blocks": [ + { + "bbox": [ + 281, + 124, + 307, + 168 + ], + "lines": [ + { + "bbox": [ + 281, + 124, + 307, + 168 + ], + "spans": [ + { + "bbox": [ + 281, + 124, + 307, + 168 + ], + "type": "image", + "image_path": "a7ded884e044cd478ea5912890434bc5bc50b4bc6cff20c995552a87ece343b3.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 281, + 168, + 307, + 212 + ], + "blocks": [ + { + "bbox": [ + 281, + 168, + 307, + 212 + ], + "lines": [ + { + "bbox": [ + 281, + 168, + 307, + 212 + ], + "spans": [ + { + "bbox": [ + 281, + 168, + 307, + 212 + ], + "type": "image", + "image_path": "f1b8d1f3a0448d1917d874be2a2d1ae350a6e6638450d933c2ee86eb9fa8277b.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 281, + 212, + 307, + 256 + ], + "blocks": [ + { + "bbox": [ + 281, + 212, + 307, + 256 + ], + "lines": [ + { + "bbox": [ + 281, + 212, + 307, + 256 + ], + "spans": [ + { + "bbox": [ + 281, + 212, + 307, + 256 + ], + "type": "image", + "image_path": "df59a374dc2634ace7b5b87c938fa7bdd9b38b31094bf162828f1522e03672cd.jpg" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 281, + 257, + 307, + 266 + ], + "lines": [ + { + "bbox": [ + 281, + 257, + 307, + 266 + ], + "spans": [ + { + "bbox": [ + 281, + 257, + 307, + 266 + ], + "type": "text", + "content": "" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_caption" + } + ], + "index": 29 + }, + { + "bbox": [ + 313, + 115, + 338, + 123 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 115, + 338, + 123 + ], + "spans": [ + { + "bbox": [ + 313, + 115, + 338, + 123 + ], + "type": "text", + "content": "BLr1" + } + ] + } + ], + "index": 31 + }, + { + "type": "image", + "bbox": [ + 312, + 124, + 339, + 168 + ], + "blocks": [ + { + "bbox": [ + 312, + 124, + 339, + 168 + ], + "lines": [ + { + "bbox": [ + 312, + 124, + 339, + 168 + ], + "spans": [ + { + "bbox": [ + 312, + 124, + 339, + 168 + ], + "type": "image", + "image_path": "797233a28297aa13af5eec7b44a23ab187ebbf6fe7ef11237a3a25a9980890f0.jpg" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_body" + } + ], + "index": 32 + }, + { + "type": "image", + "bbox": [ + 312, + 168, + 339, + 212 + ], + "blocks": [ + { + "bbox": [ + 312, + 168, + 339, + 212 + ], + "lines": [ + { + "bbox": [ + 312, + 168, + 339, + 212 + ], + "spans": [ + { + "bbox": [ + 312, + 168, + 339, + 212 + ], + "type": "image", + "image_path": "6cd73b5fada4d8af40c25e095d416fbd2253e746053b3d1f7fff01f7292399aa.jpg" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_body" + } + ], + "index": 33 + }, + { + "type": "image", + "bbox": [ + 312, + 212, + 339, + 234 + ], + "blocks": [ + { + "bbox": [ + 312, + 212, + 339, + 234 + ], + "lines": [ + { + "bbox": [ + 312, + 212, + 339, + 234 + ], + "spans": [ + { + "bbox": [ + 312, + 212, + 339, + 234 + ], + "type": "image", + "image_path": "f520f16b08761cdb23e789cce72628bfa883fb95f01130fdd69a5afce426292b.jpg" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_body" + } + ], + "index": 34 + }, + { + "type": "image", + "bbox": [ + 312, + 234, + 339, + 256 + ], + "blocks": [ + { + "bbox": [ + 312, + 234, + 339, + 256 + ], + "lines": [ + { + "bbox": [ + 312, + 234, + 339, + 256 + ], + "spans": [ + { + "bbox": [ + 312, + 234, + 339, + 256 + ], + "type": "image", + "image_path": "f9b0e83ecc240de07685ea5923221a8a5d4496b62daeb1e19ed363af5190a95e.jpg" + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 312, + 257, + 339, + 266 + ], + "lines": [ + { + "bbox": [ + 312, + 257, + 339, + 266 + ], + "spans": [ + { + "bbox": [ + 312, + 257, + 339, + 266 + ], + "type": "text", + "content": "" + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_caption" + } + ], + "index": 35 + }, + { + "bbox": [ + 339, + 115, + 366, + 123 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 339, + 115, + 366, + 123 + ], + "spans": [ + { + "bbox": [ + 339, + 115, + 366, + 123 + ], + "type": "text", + "content": "BL 2" + } + ] + } + ], + "index": 37 + }, + { + "type": "image", + "bbox": [ + 339, + 124, + 367, + 168 + ], + "blocks": [ + { + "bbox": [ + 339, + 124, + 367, + 168 + ], + "lines": [ + { + "bbox": [ + 339, + 124, + 367, + 168 + ], + "spans": [ + { + "bbox": [ + 339, + 124, + 367, + 168 + ], + "type": "image", + "image_path": "1e952ae31f57eed977581ae0316b9a396f14527f97c4ac93e060a513d2d3074d.jpg" + } + ] + } + ], + "index": 38, + "angle": 0, + "type": "image_body" + } + ], + "index": 38 + }, + { + "type": "image", + "bbox": [ + 339, + 168, + 367, + 212 + ], + "blocks": [ + { + "bbox": [ + 339, + 168, + 367, + 212 + ], + "lines": [ + { + "bbox": [ + 339, + 168, + 367, + 212 + ], + "spans": [ + { + "bbox": [ + 339, + 168, + 367, + 212 + ], + "type": "image", + "image_path": "f1212ffd355866d7a77a5cf1446569f46f63b54241e7299ac1a2d0227134a9b5.jpg" + } + ] + } + ], + "index": 39, + "angle": 0, + "type": "image_body" + } + ], + "index": 39 + }, + { + "type": "image", + "bbox": [ + 339, + 212, + 367, + 234 + ], + "blocks": [ + { + "bbox": [ + 339, + 212, + 367, + 234 + ], + "lines": [ + { + "bbox": [ + 339, + 212, + 367, + 234 + ], + "spans": [ + { + "bbox": [ + 339, + 212, + 367, + 234 + ], + "type": "image", + "image_path": "cccb591f91afdfbe3b05a4bb0e6a737e9e82e1b230e286847914ba0ed6c64aeb.jpg" + } + ] + } + ], + "index": 40, + "angle": 0, + "type": "image_body" + } + ], + "index": 40 + }, + { + "type": "image", + "bbox": [ + 339, + 234, + 367, + 256 + ], + "blocks": [ + { + "bbox": [ + 339, + 234, + 367, + 256 + ], + "lines": [ + { + "bbox": [ + 339, + 234, + 367, + 256 + ], + "spans": [ + { + "bbox": [ + 339, + 234, + 367, + 256 + ], + "type": "image", + "image_path": "713ad930d3dec3548b1bb0c45c216766e38990654989e0bf4f5af281e42602ce.jpg" + } + ] + } + ], + "index": 41, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 339, + 257, + 367, + 266 + ], + "lines": [ + { + "bbox": [ + 339, + 257, + 367, + 266 + ], + "spans": [ + { + "bbox": [ + 339, + 257, + 367, + 266 + ], + "type": "text", + "content": "(b) F" + } + ] + } + ], + "index": 42, + "angle": 0, + "type": "image_caption" + } + ], + "index": 41 + }, + { + "bbox": [ + 367, + 115, + 395, + 123 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 367, + 115, + 395, + 123 + ], + "spans": [ + { + "bbox": [ + 367, + 115, + 395, + 123 + ], + "type": "text", + "content": "BLr3" + } + ] + } + ], + "index": 43 + }, + { + "type": "image", + "bbox": [ + 367, + 124, + 395, + 168 + ], + "blocks": [ + { + "bbox": [ + 367, + 124, + 395, + 168 + ], + "lines": [ + { + "bbox": [ + 367, + 124, + 395, + 168 + ], + "spans": [ + { + "bbox": [ + 367, + 124, + 395, + 168 + ], + "type": "image", + "image_path": "62a5ed96d8abaac7cfe00f8f84ab9d1dd92bd99e9990d6147f1bac85222ae08b.jpg" + } + ] + } + ], + "index": 44, + "angle": 0, + "type": "image_body" + } + ], + "index": 44 + }, + { + "type": "image", + "bbox": [ + 367, + 168, + 395, + 212 + ], + "blocks": [ + { + "bbox": [ + 367, + 168, + 395, + 212 + ], + "lines": [ + { + "bbox": [ + 367, + 168, + 395, + 212 + ], + "spans": [ + { + "bbox": [ + 367, + 168, + 395, + 212 + ], + "type": "image", + "image_path": "f00b4aea26f8e21171959ec9493e2251e4d2ee9d7cb4d3a23d5ee6cde34582e5.jpg" + } + ] + } + ], + "index": 45, + "angle": 0, + "type": "image_body" + } + ], + "index": 45 + }, + { + "type": "image", + "bbox": [ + 367, + 212, + 395, + 234 + ], + "blocks": [ + { + "bbox": [ + 367, + 212, + 395, + 234 + ], + "lines": [ + { + "bbox": [ + 367, + 212, + 395, + 234 + ], + "spans": [ + { + "bbox": [ + 367, + 212, + 395, + 234 + ], + "type": "image", + "image_path": "21506a60ac01a8fff38b4c766eb37cd7d2285ebb2fb39467ee1615debae43d70.jpg" + } + ] + } + ], + "index": 46, + "angle": 0, + "type": "image_body" + } + ], + "index": 46 + }, + { + "type": "image", + "bbox": [ + 367, + 234, + 395, + 256 + ], + "blocks": [ + { + "bbox": [ + 367, + 234, + 395, + 256 + ], + "lines": [ + { + "bbox": [ + 367, + 234, + 395, + 256 + ], + "spans": [ + { + "bbox": [ + 367, + 234, + 395, + 256 + ], + "type": "image", + "image_path": "56fc5975b524fe2713d15aa2c23f8c9f5c250183b8e81dcb09ea4624897cdb65.jpg" + } + ] + } + ], + "index": 47, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 367, + 257, + 395, + 266 + ], + "lines": [ + { + "bbox": [ + 367, + 257, + 395, + 266 + ], + "spans": [ + { + "bbox": [ + 367, + 257, + 395, + 266 + ], + "type": "text", + "content": "Real-wo" + } + ] + } + ], + "index": 48, + "angle": 0, + "type": "image_caption" + } + ], + "index": 47 + }, + { + "bbox": [ + 399, + 115, + 426, + 123 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 399, + 115, + 426, + 123 + ], + "spans": [ + { + "bbox": [ + 399, + 115, + 426, + 123 + ], + "type": "text", + "content": "BLr 4" + } + ] + } + ], + "index": 49 + }, + { + "type": "image", + "bbox": [ + 399, + 124, + 426, + 168 + ], + "blocks": [ + { + "bbox": [ + 399, + 124, + 426, + 168 + ], + "lines": [ + { + "bbox": [ + 399, + 124, + 426, + 168 + ], + "spans": [ + { + "bbox": [ + 399, + 124, + 426, + 168 + ], + "type": "image", + "image_path": "03859a3d6c8fbf13211382455ac3cdb897e78d337e03169079a0af95993ae3af.jpg" + } + ] + } + ], + "index": 50, + "angle": 0, + "type": "image_body" + } + ], + "index": 50 + }, + { + "type": "image", + "bbox": [ + 399, + 168, + 426, + 212 + ], + "blocks": [ + { + "bbox": [ + 399, + 168, + 426, + 212 + ], + "lines": [ + { + "bbox": [ + 399, + 168, + 426, + 212 + ], + "spans": [ + { + "bbox": [ + 399, + 168, + 426, + 212 + ], + "type": "image", + "image_path": "c94246543f4209f5b5392d8ff2634f6f57df7cd21cdab4e26b2ab0cd65c6f75c.jpg" + } + ] + } + ], + "index": 51, + "angle": 0, + "type": "image_body" + } + ], + "index": 51 + }, + { + "type": "image", + "bbox": [ + 399, + 212, + 426, + 234 + ], + "blocks": [ + { + "bbox": [ + 399, + 212, + 426, + 234 + ], + "lines": [ + { + "bbox": [ + 399, + 212, + 426, + 234 + ], + "spans": [ + { + "bbox": [ + 399, + 212, + 426, + 234 + ], + "type": "image", + "image_path": "261d641d9962aea66d0965c93b0535df90d5410a5953281c3ae704858f27dd7c.jpg" + } + ] + } + ], + "index": 52, + "angle": 0, + "type": "image_body" + } + ], + "index": 52 + }, + { + "type": "image", + "bbox": [ + 399, + 234, + 426, + 256 + ], + "blocks": [ + { + "bbox": [ + 399, + 234, + 426, + 256 + ], + "lines": [ + { + "bbox": [ + 399, + 234, + 426, + 256 + ], + "spans": [ + { + "bbox": [ + 399, + 234, + 426, + 256 + ], + "type": "image", + "image_path": "f26157c1992fc2dd4e68fbabb09d904aad0d5d08e0ac1251fe42f14c82a4fd48.jpg" + } + ] + } + ], + "index": 53, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 399, + 257, + 426, + 266 + ], + "lines": [ + { + "bbox": [ + 399, + 257, + 426, + 266 + ], + "spans": [ + { + "bbox": [ + 399, + 257, + 426, + 266 + ], + "type": "text", + "content": "ld data" + } + ] + } + ], + "index": 54, + "angle": 0, + "type": "image_caption" + } + ], + "index": 53 + }, + { + "bbox": [ + 428, + 115, + 455, + 123 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 428, + 115, + 455, + 123 + ], + "spans": [ + { + "bbox": [ + 428, + 115, + 455, + 123 + ], + "type": "text", + "content": "BL 5" + } + ] + } + ], + "index": 55 + }, + { + "type": "image", + "bbox": [ + 428, + 124, + 455, + 168 + ], + "blocks": [ + { + "bbox": [ + 428, + 124, + 455, + 168 + ], + "lines": [ + { + "bbox": [ + 428, + 124, + 455, + 168 + ], + "spans": [ + { + "bbox": [ + 428, + 124, + 455, + 168 + ], + "type": "image", + "image_path": "795f32b182797c152ada88fd63301d7201669cbc68003902f4e2c4b9f0bf7f9c.jpg" + } + ] + } + ], + "index": 56, + "angle": 0, + "type": "image_body" + } + ], + "index": 56 + }, + { + "type": "image", + "bbox": [ + 428, + 168, + 455, + 212 + ], + "blocks": [ + { + "bbox": [ + 428, + 168, + 455, + 212 + ], + "lines": [ + { + "bbox": [ + 428, + 168, + 455, + 212 + ], + "spans": [ + { + "bbox": [ + 428, + 168, + 455, + 212 + ], + "type": "image", + "image_path": "2f3c27a1ceb25eba907ad14950ce5b494742c67b633b33d28deeb924419ca33f.jpg" + } + ] + } + ], + "index": 57, + "angle": 0, + "type": "image_body" + } + ], + "index": 57 + }, + { + "type": "image", + "bbox": [ + 428, + 212, + 455, + 234 + ], + "blocks": [ + { + "bbox": [ + 428, + 212, + 455, + 234 + ], + "lines": [ + { + "bbox": [ + 428, + 212, + 455, + 234 + ], + "spans": [ + { + "bbox": [ + 428, + 212, + 455, + 234 + ], + "type": "image", + "image_path": "bacca14b8aae2545db7c510374ba9f24de249da0dc3d5e261adf0a47edf71e93.jpg" + } + ] + } + ], + "index": 58, + "angle": 0, + "type": "image_body" + } + ], + "index": 58 + }, + { + "type": "image", + "bbox": [ + 428, + 234, + 455, + 256 + ], + "blocks": [ + { + "bbox": [ + 428, + 234, + 455, + 256 + ], + "lines": [ + { + "bbox": [ + 428, + 234, + 455, + 256 + ], + "spans": [ + { + "bbox": [ + 428, + 234, + 455, + 256 + ], + "type": "image", + "image_path": "6a27d4d80f04cf26bd9bde1fdb6af6aa554700f54e39d3e8275f42466a1c51b7.jpg" + } + ] + } + ], + "index": 59, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 428, + 257, + 455, + 266 + ], + "lines": [ + { + "bbox": [ + 428, + 257, + 455, + 266 + ], + "spans": [ + { + "bbox": [ + 428, + 257, + 455, + 266 + ], + "type": "text", + "content": "set." + } + ] + } + ], + "index": 60, + "angle": 0, + "type": "image_caption" + } + ], + "index": 59 + }, + { + "bbox": [ + 130, + 335, + 480, + 359 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 335, + 480, + 359 + ], + "spans": [ + { + "bbox": [ + 130, + 335, + 480, + 359 + ], + "type": "text", + "content": "mechanism. This mechanism selects contrastive samples based on the blur level in the query image for contrastive learning." + } + ] + } + ], + "index": 62 + }, + { + "bbox": [ + 130, + 360, + 480, + 397 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 360, + 480, + 397 + ], + "spans": [ + { + "bbox": [ + 130, + 360, + 480, + 397 + ], + "type": "text", + "content": "Blur severity definition. Our definition of blur severity is inspired by the image generation process. The formation of an image " + }, + { + "bbox": [ + 130, + 360, + 480, + 397 + ], + "type": "inline_equation", + "content": "\\mathbf{I}" + }, + { + "bbox": [ + 130, + 360, + 480, + 397 + ], + "type": "text", + "content": " containing a motion-blurred object in front of a background can be expressed [32] as" + } + ] + } + ], + "index": 63 + }, + { + "bbox": [ + 246, + 403, + 480, + 417 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 246, + 403, + 480, + 417 + ], + "spans": [ + { + "bbox": [ + 246, + 403, + 480, + 417 + ], + "type": "interline_equation", + "content": "\\mathbf {I} = \\mathbf {P} * \\mathbf {O} + (1 - \\mathbf {P} * \\mathbf {M}) \\mathbf {B} \\tag {3}", + "image_path": "dd6b77967306ca6ab6644d8b04a3cf3931eb1f9c14edc0add11c5cd7cd2f68a8.jpg" + } + ] + } + ], + "index": 64 + }, + { + "bbox": [ + 130, + 423, + 480, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 423, + 480, + 471 + ], + "spans": [ + { + "bbox": [ + 130, + 423, + 480, + 471 + ], + "type": "text", + "content": "where the first term denotes the convolution " + }, + { + "bbox": [ + 130, + 423, + 480, + 471 + ], + "type": "inline_equation", + "content": "(\\ast)" + }, + { + "bbox": [ + 130, + 423, + 480, + 471 + ], + "type": "text", + "content": " of a point spread function " + }, + { + "bbox": [ + 130, + 423, + 480, + 471 + ], + "type": "inline_equation", + "content": "\\mathbf{P}" + }, + { + "bbox": [ + 130, + 423, + 480, + 471 + ], + "type": "text", + "content": " corresponding to a motion trajectory and the object's sharp appearance " + }, + { + "bbox": [ + 130, + 423, + 480, + 471 + ], + "type": "inline_equation", + "content": "\\mathbf{O}" + }, + { + "bbox": [ + 130, + 423, + 480, + 471 + ], + "type": "text", + "content": ", and the second term depicts the contribution of the background " + }, + { + "bbox": [ + 130, + 423, + 480, + 471 + ], + "type": "inline_equation", + "content": "\\mathbf{B}" + }, + { + "bbox": [ + 130, + 423, + 480, + 471 + ], + "type": "text", + "content": ", in which " + }, + { + "bbox": [ + 130, + 423, + 480, + 471 + ], + "type": "inline_equation", + "content": "\\mathbf{M}" + }, + { + "bbox": [ + 130, + 423, + 480, + 471 + ], + "type": "text", + "content": " is the binary mask of " + }, + { + "bbox": [ + 130, + 423, + 480, + 471 + ], + "type": "inline_equation", + "content": "\\mathbf{O}" + }, + { + "bbox": [ + 130, + 423, + 480, + 471 + ], + "type": "text", + "content": ". Let" + } + ] + } + ], + "index": 65 + }, + { + "bbox": [ + 279, + 471, + 480, + 484 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 279, + 471, + 480, + 484 + ], + "spans": [ + { + "bbox": [ + 279, + 471, + 480, + 484 + ], + "type": "interline_equation", + "content": "\\boldsymbol {\\alpha} = \\mathbf {P} * \\mathbf {M}, \\tag {4}", + "image_path": "c76cb06bac488d64ba0527ddbfe1ae7e8b73257eb48c9765062c07de12a6df3f.jpg" + } + ] + } + ], + "index": 66 + }, + { + "bbox": [ + 130, + 487, + 481, + 572 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 487, + 481, + 572 + ], + "spans": [ + { + "bbox": [ + 130, + 487, + 481, + 572 + ], + "type": "inline_equation", + "content": "\\pmb{\\alpha} \\in \\mathbb{R}^{H \\times W}" + }, + { + "bbox": [ + 130, + 487, + 481, + 572 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 487, + 481, + 572 + ], + "type": "inline_equation", + "content": "\\pmb{\\alpha}(i,j) \\in [0,1], \\forall (i,j) \\in \\mathbf{I}" + }, + { + "bbox": [ + 130, + 487, + 481, + 572 + ], + "type": "text", + "content": ". Considering only objects with opaque materials, " + }, + { + "bbox": [ + 130, + 487, + 481, + 572 + ], + "type": "inline_equation", + "content": "\\pmb{\\alpha}" + }, + { + "bbox": [ + 130, + 487, + 481, + 572 + ], + "type": "text", + "content": " represents the visibility of the object. In other words, it can characterize the amount of blur. Generally, the sharper the object in the image, the larger the " + }, + { + "bbox": [ + 130, + 487, + 481, + 572 + ], + "type": "inline_equation", + "content": "\\pmb{\\alpha}" + }, + { + "bbox": [ + 130, + 487, + 481, + 572 + ], + "type": "text", + "content": " values are. However, at the edge pixels, the object blends with the background, and the alpha values of these edge pixels gradually fade to 0. Combining the alpha mask from Eq. (4) and considering edge effects, we define the Blur Severity (BS) of an image with a motion-blurred object as:" + } + ] + } + ], + "index": 67 + }, + { + "bbox": [ + 222, + 577, + 480, + 612 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 222, + 577, + 480, + 612 + ], + "spans": [ + { + "bbox": [ + 222, + 577, + 480, + 612 + ], + "type": "interline_equation", + "content": "B S = 1 - \\frac {\\sum_ {i = 1} ^ {H} \\sum_ {j = 1} ^ {W} \\boldsymbol {\\beta} (i , j) \\cdot \\boldsymbol {\\alpha} (i , j)}{\\sum_ {i = 1} ^ {H} \\sum_ {j = 1} ^ {W} \\boldsymbol {\\beta} (i , j)}, \\tag {5}", + "image_path": "202f2af45df83ba499c3a4662a1fe65839f588832e1032d9f646a8c65f7ec6ec.jpg" + } + ] + } + ], + "index": 68 + }, + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "type": "text", + "content": " is a thresholded (by 0) and eroded (by 3 pixels) binary mask of " + }, + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "type": "text", + "content": ". Erosion effectively removes the adverse contribution of the edge pixels to the blur severity measure. As a result, " + }, + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "type": "inline_equation", + "content": "BS \\in (0,1)" + }, + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "type": "text", + "content": ", and the larger the value, the more severe the motion blur in the image." + } + ] + } + ], + "index": 69 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "text", + "content": "R. Zou et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 135, + 114, + 482, + 196 + ], + "blocks": [ + { + "bbox": [ + 135, + 114, + 482, + 196 + ], + "lines": [ + { + "bbox": [ + 135, + 114, + 482, + 196 + ], + "spans": [ + { + "bbox": [ + 135, + 114, + 482, + 196 + ], + "type": "table", + "html": "
Dataset#Images Total#Images by BL (Syn.) or BLr (Real)
123456
Syn. Q20,9954,2883,9324,0784,0892,9301,678
Syn. DB91,62118,87117,50817,88818,02912,54667,79
Syn. Dist.1,091,939214,364177,869222,542235,662149,82891,674
Real Q2,753612620561396315249
Real DB10,3401,9231,8032,0801,7451,3751,414
", + "image_path": "024088ae819a06654f02c553f1cdd3465b97a58d7d5aa0b0197168f1820eef4e.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 198, + 479, + 218 + ], + "lines": [ + { + "bbox": [ + 132, + 198, + 479, + 218 + ], + "spans": [ + { + "bbox": [ + 132, + 198, + 479, + 218 + ], + "type": "text", + "content": "Table 1: Distribution of images in our synthetic and real datasets, sorted by blur levels. 'Q' stands for 'query', 'DB' denotes database, and 'Dist.' represents distractors." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 131, + 234, + 479, + 270 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 234, + 479, + 270 + ], + "spans": [ + { + "bbox": [ + 131, + 234, + 479, + 270 + ], + "type": "text", + "content": "Blur level-based sample selection. Since blur severity is a continuous quantity, finding images within a certain blur severity range for each query image is expensive. We thus use a discretized version of blur severity — Blur Level (BL):" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 274, + 282, + 480, + 295 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 274, + 282, + 480, + 295 + ], + "spans": [ + { + "bbox": [ + 274, + 282, + 480, + 295 + ], + "type": "interline_equation", + "content": "B L = \\lceil 1 0 B S \\rceil , \\tag {6}", + "image_path": "00692438352e74cb49aaf5e32c728109fe276fb714ae216f54c5b827b6fc4f5c.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 297, + 479, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 297, + 479, + 346 + ], + "spans": [ + { + "bbox": [ + 131, + 297, + 479, + 346 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 297, + 479, + 346 + ], + "type": "inline_equation", + "content": "\\lceil \\cdot \\rceil" + }, + { + "bbox": [ + 131, + 297, + 479, + 346 + ], + "type": "text", + "content": " denotes ceiling operation. Each image is assigned to a " + }, + { + "bbox": [ + 131, + 297, + 479, + 346 + ], + "type": "inline_equation", + "content": "BL" + }, + { + "bbox": [ + 131, + 297, + 479, + 346 + ], + "type": "text", + "content": " according to its " + }, + { + "bbox": [ + 131, + 297, + 479, + 346 + ], + "type": "inline_equation", + "content": "BS" + }, + { + "bbox": [ + 131, + 297, + 479, + 346 + ], + "type": "text", + "content": ". During training, we select contrastive samples based on the query blur level " + }, + { + "bbox": [ + 131, + 297, + 479, + 346 + ], + "type": "inline_equation", + "content": "b_{q}" + }, + { + "bbox": [ + 131, + 297, + 479, + 346 + ], + "type": "text", + "content": " and the specified contrastive " + }, + { + "bbox": [ + 131, + 297, + 479, + 346 + ], + "type": "inline_equation", + "content": "BL" + }, + { + "bbox": [ + 131, + 297, + 479, + 346 + ], + "type": "text", + "content": " range " + }, + { + "bbox": [ + 131, + 297, + 479, + 346 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 131, + 297, + 479, + 346 + ], + "type": "text", + "content": ". The selected " + }, + { + "bbox": [ + 131, + 297, + 479, + 346 + ], + "type": "inline_equation", + "content": "N_{p}" + }, + { + "bbox": [ + 131, + 297, + 479, + 346 + ], + "type": "text", + "content": " positive and " + }, + { + "bbox": [ + 131, + 297, + 479, + 346 + ], + "type": "inline_equation", + "content": "N_{n}" + }, + { + "bbox": [ + 131, + 297, + 479, + 346 + ], + "type": "text", + "content": " negative samples thus have blur levels within the range of " + }, + { + "bbox": [ + 131, + 297, + 479, + 346 + ], + "type": "inline_equation", + "content": "b_{q} - r" + }, + { + "bbox": [ + 131, + 297, + 479, + 346 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 131, + 297, + 479, + 346 + ], + "type": "inline_equation", + "content": "b_{q} + r" + }, + { + "bbox": [ + 131, + 297, + 479, + 346 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 360, + 257, + 372 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 360, + 257, + 372 + ], + "spans": [ + { + "bbox": [ + 132, + 360, + 257, + 372 + ], + "type": "text", + "content": "3.3 Training objectives" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 377, + 479, + 435 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 377, + 479, + 435 + ], + "spans": [ + { + "bbox": [ + 130, + 377, + 479, + 435 + ], + "type": "text", + "content": "Blur estimation loss. Blur severity contains information about the visibility of the blurred object in an image (Sec. 3.2). From another perspective, it reflects a signal-to-noise ratio (SNR), since the more severe the motion blur, the greater the ratio of background (noise) to foreground (signal). In view of this, we define blur estimation loss as" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 259, + 437, + 480, + 450 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 259, + 437, + 480, + 450 + ], + "spans": [ + { + "bbox": [ + 259, + 437, + 480, + 450 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {b e} = | p - (1 - B S) |, \\tag {7}", + "image_path": "52a7b20e8c0b2fc17c6d0fdd43e2ae6396df00cb9bb57ecb60c8eb03bd186db7.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 453, + 479, + 500 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 453, + 479, + 500 + ], + "spans": [ + { + "bbox": [ + 130, + 453, + 479, + 500 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 453, + 479, + 500 + ], + "type": "inline_equation", + "content": "p \\in \\mathbb{R}" + }, + { + "bbox": [ + 130, + 453, + 479, + 500 + ], + "type": "text", + "content": " is the blur value predicted by the blur estimation head. By constraining the estimate of the continuous " + }, + { + "bbox": [ + 130, + 453, + 479, + 500 + ], + "type": "inline_equation", + "content": "BS" + }, + { + "bbox": [ + 130, + 453, + 479, + 500 + ], + "type": "text", + "content": " (Eq. (5)), we implicitly force the network to identify blur and estimate the SNR, which helps it to focus on and extract signal information in the blurred region while ignoring background noises." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 502, + 322, + 512 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 502, + 322, + 512 + ], + "spans": [ + { + "bbox": [ + 132, + 502, + 322, + 512 + ], + "type": "text", + "content": "Localization loss. This loss is defined as" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 212, + 519, + 480, + 532 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 212, + 519, + 480, + 532 + ], + "spans": [ + { + "bbox": [ + 212, + 519, + 480, + 532 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {l o c} = | x - \\hat {x} | + | y - \\hat {y} | + | h - \\hat {h} | + | w - \\hat {w} |, \\tag {8}", + "image_path": "d5d736a12e22ff3a5e23489db4a125e3f1d82392d6be2141012bd66a6a3e30ba.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 537, + 479, + 597 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 537, + 479, + 597 + ], + "spans": [ + { + "bbox": [ + 130, + 537, + 479, + 597 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 537, + 479, + 597 + ], + "type": "inline_equation", + "content": "(x,y,w,h)" + }, + { + "bbox": [ + 130, + 537, + 479, + 597 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 537, + 479, + 597 + ], + "type": "inline_equation", + "content": "(\\hat{x},\\hat{y},\\hat{w},\\hat{h})" + }, + { + "bbox": [ + 130, + 537, + 479, + 597 + ], + "type": "text", + "content": " represent the normalized ground truth bounding box and the predicted one from the localization head, respectively. Localization loss requires the network to distinguish motion-blurred objects from backgrounds under various blur conditions, forcing it to recognize blur, identify objects at different blur levels, and further, focus more on the foreground objects." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 598, + 479, + 623 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 598, + 479, + 623 + ], + "spans": [ + { + "bbox": [ + 132, + 598, + 479, + 623 + ], + "type": "text", + "content": "Classification loss. Following DELG [4], we apply ArcFace margin loss [9] to the descriptor " + }, + { + "bbox": [ + 132, + 598, + 479, + 623 + ], + "type": "inline_equation", + "content": "\\mathbf{D}" + }, + { + "bbox": [ + 132, + 598, + 479, + 623 + ], + "type": "text", + "content": ", expressed as" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 204, + 628, + 480, + 667 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 204, + 628, + 480, + 667 + ], + "spans": [ + { + "bbox": [ + 204, + 628, + 480, + 667 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {c l s} = - \\log \\left(\\frac {\\exp (\\gamma \\times A F (\\hat {\\mathbf {w}} _ {t} ^ {T} \\hat {\\mathbf {D}} , 1))}{\\sum_ {n} \\exp (\\gamma \\times A F (\\hat {\\mathbf {w}} _ {n} ^ {T} \\hat {\\mathbf {D}} , y _ {n}))}\\right), \\tag {9}", + "image_path": "257194d940221f9fceaba95426866f9ba19dbc6d75428e75c990a1aa863dfb15.jpg" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 282, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 282, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 282, + 91, + 448, + 102 + ], + "type": "text", + "content": "Retrieval Robust to Object Motion Blur" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 134, + 114, + 482, + 173 + ], + "blocks": [ + { + "bbox": [ + 134, + 114, + 482, + 173 + ], + "lines": [ + { + "bbox": [ + 134, + 114, + 482, + 173 + ], + "spans": [ + { + "bbox": [ + 134, + 114, + 482, + 173 + ], + "type": "table", + "html": "
MethodmAP by query BL (synthetic)mAP by query BL (synthetic + 1M)
All123456All123456
DELG [4]81.6583.5584.9684.8982.6977.3566.2068.1973.6475.4073.3468.0558.2842.46
DOLG [47]83.0385.1086.2685.9384.0478.7868.0769.9775.7577.4775.0170.1060.0142.49
Token [46]84.8486.5387.9987.5085.8180.7071.5770.6575.3277.6675.5170.2461.1948.05
Ours-sharp27.5743.0638.5329.7419.4010.876.1432.6471.9343.8827.1815.417.944.27
Ours91.7893.0593.4893.1492.2590.2082.8684.0988.7489.5687.6884.4176.8962.42
", + "image_path": "3ff9f129d1a0e5210948570e1ecca09b2c92bfeb08c8a8f102792c5b102694d9.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 175, + 481, + 218 + ], + "lines": [ + { + "bbox": [ + 132, + 175, + 481, + 218 + ], + "spans": [ + { + "bbox": [ + 132, + 175, + 481, + 218 + ], + "type": "text", + "content": "Table 2: Retrieval results on the synthetic dataset with (+1M) and without distractors. Ours-sharp is our model trained with only sharp images. 'All' denotes the overall performance for queries of all blur levels. All methods are re-implemented and re-trained on our synthetic training set. The database contains images of all blur levels." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 130, + 234, + 479, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 234, + 479, + 281 + ], + "spans": [ + { + "bbox": [ + 130, + 234, + 479, + 281 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 234, + 479, + 281 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 130, + 234, + 479, + 281 + ], + "type": "text", + "content": " is a scale factor, " + }, + { + "bbox": [ + 130, + 234, + 479, + 281 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{w}}_{\\mathbf{i}}" + }, + { + "bbox": [ + 130, + 234, + 479, + 281 + ], + "type": "text", + "content": " refers to the " + }, + { + "bbox": [ + 130, + 234, + 479, + 281 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 234, + 479, + 281 + ], + "type": "text", + "content": "-th row of the " + }, + { + "bbox": [ + 130, + 234, + 479, + 281 + ], + "type": "inline_equation", + "content": "L_{2}" + }, + { + "bbox": [ + 130, + 234, + 479, + 281 + ], + "type": "text", + "content": " normalized " + }, + { + "bbox": [ + 130, + 234, + 479, + 281 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 234, + 479, + 281 + ], + "type": "text", + "content": "-class classifier " + }, + { + "bbox": [ + 130, + 234, + 479, + 281 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{W}} \\in \\mathbb{R}^{d \\times N}" + }, + { + "bbox": [ + 130, + 234, + 479, + 281 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 234, + 479, + 281 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{D}}" + }, + { + "bbox": [ + 130, + 234, + 479, + 281 + ], + "type": "text", + "content": " denotes the " + }, + { + "bbox": [ + 130, + 234, + 479, + 281 + ], + "type": "inline_equation", + "content": "L_{2}" + }, + { + "bbox": [ + 130, + 234, + 479, + 281 + ], + "type": "text", + "content": " normalized " + }, + { + "bbox": [ + 130, + 234, + 479, + 281 + ], + "type": "inline_equation", + "content": "\\mathbf{D}" + }, + { + "bbox": [ + 130, + 234, + 479, + 281 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 130, + 234, + 479, + 281 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 130, + 234, + 479, + 281 + ], + "type": "text", + "content": " is the one-hot label, and " + }, + { + "bbox": [ + 130, + 234, + 479, + 281 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 234, + 479, + 281 + ], + "type": "text", + "content": " represents the ground truth class index (" + }, + { + "bbox": [ + 130, + 234, + 479, + 281 + ], + "type": "inline_equation", + "content": "y_{t} = 1" + }, + { + "bbox": [ + 130, + 234, + 479, + 281 + ], + "type": "text", + "content": "). " + }, + { + "bbox": [ + 130, + 234, + 479, + 281 + ], + "type": "inline_equation", + "content": "AF" + }, + { + "bbox": [ + 130, + 234, + 479, + 281 + ], + "type": "text", + "content": " is the ArcFace-adjusted cosine similarity and it can be expressed as" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 212, + 289, + 481, + 319 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 212, + 289, + 481, + 319 + ], + "spans": [ + { + "bbox": [ + 212, + 289, + 481, + 319 + ], + "type": "interline_equation", + "content": "A F (s, g) = \\left\\{ \\begin{array}{l l} \\cos (\\operatorname {a r c c o s} (s) + m), & \\text {i f} g = 1 \\\\ s, & \\text {i f} g = 0 \\end{array} , \\right. \\tag {10}", + "image_path": "d179b76a346e8d2fea9803a4f19571b2a4ae110681df0cca7ff3983d0088643d.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 325, + 479, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 325, + 479, + 384 + ], + "spans": [ + { + "bbox": [ + 130, + 325, + 479, + 384 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 325, + 479, + 384 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 130, + 325, + 479, + 384 + ], + "type": "text", + "content": " is the cosine similarity, " + }, + { + "bbox": [ + 130, + 325, + 479, + 384 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 130, + 325, + 479, + 384 + ], + "type": "text", + "content": " is the ArcFace margin, and " + }, + { + "bbox": [ + 130, + 325, + 479, + 384 + ], + "type": "inline_equation", + "content": "g \\in \\{0,1\\}" + }, + { + "bbox": [ + 130, + 325, + 479, + 384 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 130, + 325, + 479, + 384 + ], + "type": "inline_equation", + "content": "g = 1" + }, + { + "bbox": [ + 130, + 325, + 479, + 384 + ], + "type": "text", + "content": " indicates the ground truth class. The classification loss requires the model to group sharp or motion-blurred images of the same object together, while distinguishing between different objects, which may appear similar, especially when blurred. This implicitly forces the model to focus on blur-invariant features." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 386, + 479, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 386, + 479, + 434 + ], + "spans": [ + { + "bbox": [ + 130, + 386, + 479, + 434 + ], + "type": "text", + "content": "Contrastive loss. We employ contrastive learning [6] to construct a descriptor space, where the distance between similar samples is minimized, and between dissimilar samples maximized. The contrastive loss is applied to (query, positive) and (query, negative) descriptor pairs:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 192, + 441, + 480, + 472 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 192, + 441, + 480, + 472 + ], + "spans": [ + { + "bbox": [ + 192, + 441, + 480, + 472 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {c o n}} (i, j) = \\left\\{ \\begin{array}{l l} 0. 5 \\cdot D _ {i, j} ^ {2}, & \\text {i f} y (i, j) = 1 \\\\ 0. 5 \\left(\\max \\{0, \\tau - D _ {i, j} \\}\\right) ^ {2}, & \\text {i f} y (i, j) = 0 \\end{array} , \\right. \\tag {11}", + "image_path": "f55807e37301f2597379c894506a3a309f81fd3baaf9f01bad00920558498e42.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 479, + 479, + 539 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 479, + 479, + 539 + ], + "spans": [ + { + "bbox": [ + 130, + 479, + 479, + 539 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 479, + 479, + 539 + ], + "type": "inline_equation", + "content": "D_{i,j} = ||\\hat{\\mathbf{D}}(i) - \\hat{\\mathbf{D}}(j)||" + }, + { + "bbox": [ + 130, + 479, + 479, + 539 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 479, + 479, + 539 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{D}}(i)" + }, + { + "bbox": [ + 130, + 479, + 479, + 539 + ], + "type": "text", + "content": " is the " + }, + { + "bbox": [ + 130, + 479, + 479, + 539 + ], + "type": "inline_equation", + "content": "L_2" + }, + { + "bbox": [ + 130, + 479, + 479, + 539 + ], + "type": "text", + "content": " normalized descriptor of image " + }, + { + "bbox": [ + 130, + 479, + 479, + 539 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 479, + 479, + 539 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 130, + 479, + 479, + 539 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 130, + 479, + 479, + 539 + ], + "type": "text", + "content": " is a margin parameter, and " + }, + { + "bbox": [ + 130, + 479, + 479, + 539 + ], + "type": "inline_equation", + "content": "y(i,j)" + }, + { + "bbox": [ + 130, + 479, + 479, + 539 + ], + "type": "text", + "content": " is binary indicating whether the pair matches " + }, + { + "bbox": [ + 130, + 479, + 479, + 539 + ], + "type": "inline_equation", + "content": "(= 1)" + }, + { + "bbox": [ + 130, + 479, + 479, + 539 + ], + "type": "text", + "content": ". The contrastive loss further enhances the focus on blur-invariant features by explicitly forcing descriptors of the same object, whether sharp or blurred, to be close together, and those of different objects to be far apart." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 540, + 469, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 540, + 469, + 552 + ], + "spans": [ + { + "bbox": [ + 130, + 540, + 469, + 552 + ], + "type": "text", + "content": "Joint loss. The joint loss is a weighted sum of all previously defined losses:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 209, + 560, + 480, + 572 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 209, + 560, + 480, + 572 + ], + "spans": [ + { + "bbox": [ + 209, + 560, + 480, + 572 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {j o i n t}} = \\mathcal {L} _ {\\text {c o n}} + \\alpha_ {\\text {c l s}} \\mathcal {L} _ {\\text {c l s}} + \\alpha_ {\\text {b e}} \\mathcal {L} _ {\\text {b e}} + \\alpha_ {\\text {l o c}} \\mathcal {L} _ {\\text {l o c}}. \\tag {12}", + "image_path": "2f9cdaba27c154c92bd192e081e59d91ab33b460d8badba94bd5ba651f506736.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 606, + 201, + 618 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 606, + 201, + 618 + ], + "spans": [ + { + "bbox": [ + 132, + 606, + 201, + 618 + ], + "type": "text", + "content": "4 Dataset" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 629, + 479, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 479, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 479, + 665 + ], + "type": "text", + "content": "We create novel synthetic and real datasets for training and evaluating retrieval methods for motion-blurred objects. The synthetic dataset serves for both training and evaluation. We also generate a large-scale synthetic distractor set to" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "text", + "content": "R. Zou et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 135, + 114, + 480, + 201 + ], + "blocks": [ + { + "bbox": [ + 135, + 114, + 480, + 201 + ], + "lines": [ + { + "bbox": [ + 135, + 114, + 480, + 201 + ], + "spans": [ + { + "bbox": [ + 135, + 114, + 480, + 201 + ], + "type": "table", + "html": "
MethodmAP AllmAP by Query BLr
123456
DELG [4]54.8249.1363.4357.2555.0153.7742.92
DOLG [47]54.6443.9360.5958.3659.0658.5845.78
Token [46]43.3338.7147.0850.7946.4442.7124.43
Ours-sharp40.2449.5545.0241.3333.2329.4027.91
Ours62.8857.5070.3866.7763.1864.4846.14
", + "image_path": "ccee4c7216ab74a3ae1723748a9b2cc277f04211aaee40d45ab3ecc07616686b.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 201, + 479, + 222 + ], + "lines": [ + { + "bbox": [ + 133, + 201, + 479, + 222 + ], + "spans": [ + { + "bbox": [ + 133, + 201, + 479, + 222 + ], + "type": "text", + "content": "Table 3: Retrieval results on the real-world dataset (trained on synthetic without fine-tuning). The database includes all blur levels. Our method outperforms all others." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 133, + 236, + 479, + 283 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 236, + 479, + 283 + ], + "spans": [ + { + "bbox": [ + 133, + 236, + 479, + 283 + ], + "type": "text", + "content": "evaluate methods in a more challenging setup. The novel real-world dataset is specifically constructed to validate the methods' effectiveness in real scenarios. All datasets are meticulously processed for the blur retrieval task, laying a crucial foundation for future research. Statistics of these datasets are shown in Table 1." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 303, + 247, + 314 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 303, + 247, + 314 + ], + "spans": [ + { + "bbox": [ + 133, + 303, + 247, + 314 + ], + "type": "text", + "content": "4.1 Synthetic dataset" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 133, + 323, + 479, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 323, + 479, + 357 + ], + "spans": [ + { + "bbox": [ + 133, + 323, + 479, + 357 + ], + "type": "text", + "content": "The images in this dataset contain 3D objects moving along different trajectories, in front of different backgrounds, and with different amounts of motion blur. Thus, this dataset poses many challenges to retrieval methods." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 133, + 360, + 479, + 383 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 360, + 479, + 383 + ], + "spans": [ + { + "bbox": [ + 133, + 360, + 479, + 383 + ], + "type": "text", + "content": "Backgrounds. Background images are sampled from the LHQ dataset [39], cropped to a resolution of " + }, + { + "bbox": [ + 133, + 360, + 479, + 383 + ], + "type": "inline_equation", + "content": "240 \\times 320" + }, + { + "bbox": [ + 133, + 360, + 479, + 383 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 133, + 384, + 479, + 408 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 384, + 479, + 408 + ], + "spans": [ + { + "bbox": [ + 133, + 384, + 479, + 408 + ], + "type": "text", + "content": "Objects. We select 39 categories with the largest number of textured objects from the ShapeNet dataset [5] and randomly sample 30 objects in each category." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 133, + 410, + 479, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 410, + 479, + 445 + ], + "spans": [ + { + "bbox": [ + 133, + 410, + 479, + 445 + ], + "type": "text", + "content": "Trajectories. For each object, we generate 120 random linear trajectories, each divided into 23 sub-trajectories. We then select up to 10 consecutive sub-trajectories to create images featuring objects with diverse levels of motion blur." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 133, + 445, + 479, + 482 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 445, + 479, + 482 + ], + "spans": [ + { + "bbox": [ + 133, + 445, + 479, + 482 + ], + "type": "text", + "content": "Rendering. We render blurry images with Blender Cycles [7] by capturing moving objects during the camera exposure time. Additionally, we render one sharp image per trajectory, placed at the central position along the trajectory." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 133, + 483, + 479, + 531 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 483, + 479, + 531 + ], + "spans": [ + { + "bbox": [ + 133, + 483, + 479, + 531 + ], + "type": "text", + "content": "Filtering. We discard renderings where the object occupies less than " + }, + { + "bbox": [ + 133, + 483, + 479, + 531 + ], + "type": "inline_equation", + "content": "1.5\\%" + }, + { + "bbox": [ + 133, + 483, + 479, + 531 + ], + "type": "text", + "content": " of the pixels or where the intersection-over-union between silhouette renderings at trajectory endpoints is below " + }, + { + "bbox": [ + 133, + 483, + 479, + 531 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 133, + 483, + 479, + 531 + ], + "type": "text", + "content": ". The former condition suggests the object is too small, while the latter indicates the appearance changes are too drastic." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 133, + 533, + 479, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 533, + 479, + 568 + ], + "spans": [ + { + "bbox": [ + 133, + 533, + 479, + 568 + ], + "type": "text", + "content": "Ground truth. For each generated image, we estimate the bounding box coordinates by detecting the non-zero pixels in the alpha channel. We calculate the blur level for each image based on Eq. (6)." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 133, + 570, + 479, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 570, + 479, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 570, + 479, + 665 + ], + "type": "text", + "content": "Retrieval dataset. In total, we rendered 1.5 million images for 1138 different instances (examples in Fig. 3a). Additionally, we create a challenging distractor set with over 1 million images (examples in the supplementary material), using the same object categories as before to increase the difficulty of the distractors in terms of intra-class similarity. We randomly select 40 objects per category, excluding the 30 previously chosen objects, and generate 70 trajectories for each. For each trajectory, we render one sharp image and 10 blurry images with varying degrees of blur. Backgrounds are also selected from a different subset of [39]." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 283, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 283, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 283, + 91, + 447, + 102 + ], + "type": "text", + "content": "Retrieval Robust to Object Motion Blur" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 134, + 114, + 482, + 274 + ], + "blocks": [ + { + "bbox": [ + 134, + 114, + 482, + 274 + ], + "lines": [ + { + "bbox": [ + 134, + 114, + 482, + 274 + ], + "spans": [ + { + "bbox": [ + 134, + 114, + 482, + 274 + ], + "type": "table", + "html": "
BLQDDELG [4] (range: 57.79; std: 16.75)DOLG [47] (range: 56.81; std: 16.93)
All123456All123456
175.9986.7384.1780.2672.4863.1250.0877.7887.0185.7081.8575.1166.2052.39
277.3582.7486.2282.8776.6866.4650.0078.4484.1887.0083.4377.9167.8851.31
377.1979.6283.4382.9478.4669.4352.8678.4981.8484.7483.8579.6970.5853.12
474.9674.2678.9179.8578.0570.8955.1276.4077.2780.5580.8079.1071.7355.28
562.7157.8661.7264.3765.3668.9656.0564.1961.3163.6865.3866.0569.6855.75
634.8929.5628.9431.7534.0645.3753.8635.5131.2430.2032.2934.1845.0053.41
BLQDToken [46] (range: 55.72; std: 16.11)Ours (range: 53.68; std: 15.65)
All123456All123456
179.6588.2287.3483.5076.8768.5156.6388.5795.0994.2491.5487.1980.9168.15
279.9784.7888.4884.7778.9970.0155.8288.7593.0794.6592.1888.7782.0367.18
379.5081.7985.4684.7580.2172.0958.1888.7691.7093.1892.3289.8383.4168.94
476.8276.5680.4281.2479.3172.6159.6187.1688.1290.1190.0989.3184.3770.35
565.3961.8564.4966.3167.1470.9660.3676.2873.8774.9076.0477.3983.0671.67
637.9833.2332.7634.6136.5047.5657.4347.3143.4341.4142.8944.5558.1369.59
", + "image_path": "0171945a1021d33ff40187b7c470e1bfc6e824366c8bd0525a6b9d2733c8cd6b.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 275, + 480, + 318 + ], + "lines": [ + { + "bbox": [ + 130, + 275, + 480, + 318 + ], + "spans": [ + { + "bbox": [ + 130, + 275, + 480, + 318 + ], + "type": "text", + "content": "Table 4: Retrieval results by query (Q) and database (D) blur level (BL) on our synthetic dataset (with 1M distractors). 'All' denotes the overall performance for queries of all blur levels. The standard deviation and range of mAP values are displayed alongside each method. The best scores under the same settings are marked in bold." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 132, + 331, + 255, + 342 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 331, + 255, + 342 + ], + "spans": [ + { + "bbox": [ + 132, + 331, + 255, + 342 + ], + "type": "text", + "content": "4.2 Real-world dataset" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 349, + 480, + 432 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 349, + 480, + 432 + ], + "spans": [ + { + "bbox": [ + 130, + 349, + 480, + 432 + ], + "type": "text", + "content": "We record videos with a GoPro 7 Hero camera at 240 fps with full exposure. We use 35 different objects and move each object 3 to 5 times along different trajectories in front of the camera. None of the real-world objects are in synthetic data and " + }, + { + "bbox": [ + 130, + 349, + 480, + 432 + ], + "type": "inline_equation", + "content": "2/3" + }, + { + "bbox": [ + 130, + 349, + 480, + 432 + ], + "type": "text", + "content": " of which are from new categories not present in synthetic data. This better tests the generalization ability of models to new objects and categories. In total, we recorded 139 videos. Different amounts of blur are generated by averaging consecutive frames, ranging from just one up to 200 averaged frames." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 433, + 481, + 516 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 433, + 481, + 516 + ], + "spans": [ + { + "bbox": [ + 130, + 433, + 481, + 516 + ], + "type": "text", + "content": "Computing blur level based on Eq. (6) is not feasible for the real-world dataset. Therefore, we manually annotate the data, assigning each image to one of 6 blur levels based on the perceived blur, and we use the ratio of the trajectory length to the average object size as a reference. The annotation enables a more detailed analysis and a more comprehensive evaluation of model performance. Further annotation details can be found in the supplementary material. Since the blur level for real data differs from that in Eq. (6), we denote it as " + }, + { + "bbox": [ + 130, + 433, + 481, + 516 + ], + "type": "inline_equation", + "content": "\\mathrm{BL}^{\\mathrm{r}}" + }, + { + "bbox": [ + 130, + 433, + 481, + 516 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 517, + 481, + 588 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 517, + 481, + 588 + ], + "spans": [ + { + "bbox": [ + 130, + 517, + 481, + 588 + ], + "type": "text", + "content": "The real-world dataset contains 13,093 images, with an average of 374 images for each object (examples in Fig. 3b). Compared to the synthetic dataset, the object scale change is more drastic in different trajectories, e.g. rows 1 & 2. The trajectories are notably different from those in the synthetic dataset. They involve complex curved trajectories with acceleration and other physical forces, which poses an additional challenge for methods trained on our synthetic dataset." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 605, + 230, + 618 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 605, + 230, + 618 + ], + "spans": [ + { + "bbox": [ + 132, + 605, + 230, + 618 + ], + "type": "text", + "content": "5 Experiments" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "type": "text", + "content": "Setup. For each category of objects in the synthetic dataset, we randomly take " + }, + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "type": "inline_equation", + "content": "70\\%" + }, + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "type": "text", + "content": " (rounded down, same for validation) for training (792 objects), " + }, + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "type": "inline_equation", + "content": "15\\%" + }, + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "type": "text", + "content": " for validation (153 objects), and " + }, + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "type": "inline_equation", + "content": "15\\%" + }, + { + "bbox": [ + 130, + 629, + 481, + 666 + ], + "type": "text", + "content": " for testing (193 objects). For the training objects," + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "text", + "content": "R. Zou et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 138, + 125, + 480, + 219 + ], + "blocks": [ + { + "bbox": [ + 138, + 114, + 167, + 125 + ], + "lines": [ + { + "bbox": [ + 138, + 114, + 167, + 125 + ], + "spans": [ + { + "bbox": [ + 138, + 114, + 167, + 125 + ], + "type": "text", + "content": "Query" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 138, + 125, + 480, + 219 + ], + "lines": [ + { + "bbox": [ + 138, + 125, + 480, + 219 + ], + "spans": [ + { + "bbox": [ + 138, + 125, + 480, + 219 + ], + "type": "image", + "image_path": "90bf25b235aad8fc46e634961eadf13e9b8574183c30253697f6e8e252ced634.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 157, + 219, + 456, + 228 + ], + "lines": [ + { + "bbox": [ + 157, + 219, + 456, + 228 + ], + "spans": [ + { + "bbox": [ + 157, + 219, + 456, + 228 + ], + "type": "text", + "content": "(a) Example of retrieval results on our synthetic dataset (with distractors)." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 138, + 230, + 479, + 325 + ], + "blocks": [ + { + "bbox": [ + 138, + 230, + 479, + 325 + ], + "lines": [ + { + "bbox": [ + 138, + 230, + 479, + 325 + ], + "spans": [ + { + "bbox": [ + 138, + 230, + 479, + 325 + ], + "type": "image", + "image_path": "31cfcdb2ee0ec2db6a6bda3958324ef169dafb0c823d271a7154a5f1b9f00dab.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 190, + 326, + 423, + 335 + ], + "lines": [ + { + "bbox": [ + 190, + 326, + 423, + 335 + ], + "spans": [ + { + "bbox": [ + 190, + 326, + 423, + 335 + ], + "type": "text", + "content": "(b) Example of retrieval results on our real-world dataset." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 338, + 482, + 393 + ], + "lines": [ + { + "bbox": [ + 130, + 338, + 482, + 393 + ], + "spans": [ + { + "bbox": [ + 130, + 338, + 482, + 393 + ], + "type": "text", + "content": "Fig. 4: Comparison of retrieval results. We compare to the best-performing state-of-the-art method for the corresponding query. The retrieved images are sorted from left to right and top to bottom, with the ranking from " + }, + { + "bbox": [ + 130, + 338, + 482, + 393 + ], + "type": "inline_equation", + "content": "1^{\\mathrm{st}}" + }, + { + "bbox": [ + 130, + 338, + 482, + 393 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 130, + 338, + 482, + 393 + ], + "type": "inline_equation", + "content": "20^{\\mathrm{th}}" + }, + { + "bbox": [ + 130, + 338, + 482, + 393 + ], + "type": "text", + "content": ". The blur level of each image is shown in the bottom left corner. Correct results are in green boxes, while incorrect ones are in red boxes. We provide more results in the supplementary material." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 410, + 482, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 410, + 482, + 506 + ], + "spans": [ + { + "bbox": [ + 130, + 410, + 482, + 506 + ], + "type": "text", + "content": "we take one image for each trajectory while ensuring a balanced distribution of images across different blur levels. In total, there are 95k images in the training set. For each of the test objects, we sample 20 trajectories to form a query set, and the rest is used for the database. The database images are sampled from all blur levels unless specified otherwise (for per blur level database experiments). We additionally show performance on a subset of queries for a given blur level. In all experiments (except Ours-sharp), only one model is trained for all blur levels, and the evaluation is performed for varying blur levels." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 507, + 482, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 507, + 482, + 639 + ], + "spans": [ + { + "bbox": [ + 130, + 507, + 482, + 639 + ], + "type": "text", + "content": "Implementation details. The proposed method is implemented in PyTorch [26]. We use ResNet-50 [12] up to the last downsampling block as the backbone, deactivate the normalization layers, and initialize the model with ImageNet [8] pre-trained weights. ADAM [17] optimizer is employed to minimize the joint loss with a fixed learning rate " + }, + { + "bbox": [ + 130, + 507, + 482, + 639 + ], + "type": "inline_equation", + "content": "10^{-4}" + }, + { + "bbox": [ + 130, + 507, + 482, + 639 + ], + "type": "text", + "content": ", batch size 32, and 30 epochs. We initialize the power parameter to 3 for GeM pooling [30]. The contrastive margin " + }, + { + "bbox": [ + 130, + 507, + 482, + 639 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 130, + 507, + 482, + 639 + ], + "type": "text", + "content": " is set to 0.7. Following [30], we set the number of contrastive samples as " + }, + { + "bbox": [ + 130, + 507, + 482, + 639 + ], + "type": "inline_equation", + "content": "N_{p} = 1" + }, + { + "bbox": [ + 130, + 507, + 482, + 639 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 507, + 482, + 639 + ], + "type": "inline_equation", + "content": "N_{n} = 5" + }, + { + "bbox": [ + 130, + 507, + 482, + 639 + ], + "type": "text", + "content": ". The contrastive blur range is experimentally set to 5. For classification, we empirically set the margin " + }, + { + "bbox": [ + 130, + 507, + 482, + 639 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 130, + 507, + 482, + 639 + ], + "type": "text", + "content": " to 0.15 and the scale " + }, + { + "bbox": [ + 130, + 507, + 482, + 639 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 130, + 507, + 482, + 639 + ], + "type": "text", + "content": " to 30. Loss weights are chosen so that the values of different losses are roughly of the same order of magnitude, i.e. " + }, + { + "bbox": [ + 130, + 507, + 482, + 639 + ], + "type": "inline_equation", + "content": "\\alpha_{cls} = 0.1" + }, + { + "bbox": [ + 130, + 507, + 482, + 639 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 507, + 482, + 639 + ], + "type": "inline_equation", + "content": "\\alpha_{be} = 1.0" + }, + { + "bbox": [ + 130, + 507, + 482, + 639 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 507, + 482, + 639 + ], + "type": "inline_equation", + "content": "\\alpha_{loc} = 10" + }, + { + "bbox": [ + 130, + 507, + 482, + 639 + ], + "type": "text", + "content": ". The descriptor dimension " + }, + { + "bbox": [ + 130, + 507, + 482, + 639 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 130, + 507, + 482, + 639 + ], + "type": "text", + "content": " is set to 128." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "content": "Evaluation protocols. We report mean average precision (mAP@all) as commonly used in retrieval. In case 1M distractors are used, we report mAP@100." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 283, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 283, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 283, + 91, + 447, + 102 + ], + "type": "text", + "content": "Retrieval Robust to Object Motion Blur" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 142, + 114, + 476, + 278 + ], + "blocks": [ + { + "bbox": [ + 142, + 114, + 476, + 278 + ], + "lines": [ + { + "bbox": [ + 142, + 114, + 476, + 278 + ], + "spans": [ + { + "bbox": [ + 142, + 114, + 476, + 278 + ], + "type": "table", + "html": "
\\( \\mathcal{L}_{con} \\)\\( \\mathcal{L}_{cls} \\)\\( \\mathcal{L}_{be} \\)\\( \\mathcal{L}_{loc} \\)mAP AllmAP by Query BL
123456
✓✘✘✘78.1380.5181.7081.1679.2073.9361.24
✓✘✓✘81.6683.4985.0184.4382.6977.6467.08
✓✘✘✓85.9487.5488.2587.8386.5283.0875.58
✓✘✓✓87.4888.6989.8989.4088.2484.9776.74
✘✓✘✘78.7381.5382.9782.8379.8672.9359.00
✘✓✓✘83.6785.1987.5687.4085.1079.2265.93
✘✓✘✓88.7489.8990.9190.8889.7586.0777.67
✘✓✓✓91.2392.0293.1693.0991.9789.0082.27
✓✓✘✘85.0687.4288.2987.6685.8581.2069.96
✓✓✓✘87.1789.0390.0389.5588.0783.9173.48
✓✓✘✓90.3991.8592.4592.1491.2088.2079.36
✓✓✓✓91.7893.0593.4893.1492.2590.2082.86
", + "image_path": "457aefb7d1d7b1857c1c8276bd1352c91cc01fd0016a8b4460b0aa84502993be.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 280, + 479, + 312 + ], + "lines": [ + { + "bbox": [ + 132, + 280, + 479, + 312 + ], + "spans": [ + { + "bbox": [ + 132, + 280, + 479, + 312 + ], + "type": "text", + "content": "Table 5: Ablation study of different loss components on the synthetic dataset. Results are grouped according to the loss function(s) directly applied to the descriptor. The best scores within each split are bolded. The database contains all blur levels." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 130, + 326, + 480, + 410 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 326, + 480, + 410 + ], + "spans": [ + { + "bbox": [ + 130, + 326, + 480, + 410 + ], + "type": "text", + "content": "Compared methods. Since retrieval with object motion blur has not been addressed before, there are no methods specifically designed for this task. Therefore, we compare to state-of-the-art standard retrieval methods: DELG-global [4] (referred to as DELG), DOLG [47], and Token [46]. All these methods are re-implemented as in the original papers, except for changing the size of the output descriptors to 128 for fair comparison. We re-train the compared methods using the same ArcFace parameters and on the same synthetic data as our approach." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 427, + 198, + 438 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 427, + 198, + 438 + ], + "spans": [ + { + "bbox": [ + 132, + 427, + 198, + 438 + ], + "type": "text", + "content": "5.1 Results" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 449, + 481, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 449, + 481, + 592 + ], + "spans": [ + { + "bbox": [ + 130, + 449, + 481, + 592 + ], + "type": "text", + "content": "All blur levels in the database. Our method outperforms all state-of-the-art retrieval methods for all settings, as shown in Table 2 for synthetic data and Table 3 for real data. Our model demonstrates significant superiority over the state-of-the-art retrieval approaches without introducing complex computational modules, highlighting the efficacy of our proposed loss functions. Results on the real-world dataset show that our method generalizes well to real data even when trained solely on synthetic data. Additionally, we train a version of our method where only sharp images are used (Ours-sharp). When evaluating with mAP@all (no distractors), it fails completely, especially for higher blur levels. When evaluating with mAP@100 (with distractors), it achieves reasonable results on query BL 1, but still worse than all other methods. This underscores the necessity for tailored methods and training data with highly motion-blurred objects." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": "Per blur level database. In this setup, we utilize datasets specific to each blur level, considering them as individual subsets of images with the designated blur level, to form the database. We then assess the retrieval performance by using query images with distinct blur levels, resulting in a comprehensive evaluation comprising 36 results (6 database BLs by 6 query BLs) for each method. We also compute the standard deviation (std) and range of mAPs for different settings" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "text", + "content": "R. Zou et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 165, + 114, + 452, + 196 + ], + "blocks": [ + { + "bbox": [ + 165, + 114, + 452, + 196 + ], + "lines": [ + { + "bbox": [ + 165, + 114, + 452, + 196 + ], + "spans": [ + { + "bbox": [ + 165, + 114, + 452, + 196 + ], + "type": "table", + "html": "
Cont. BL RangemAP AllmAP by query BL
123456
S82.2687.3788.4686.2982.6174.1358.26
M83.3387.9789.1286.9383.8776.0460.61
L84.0988.7489.5687.6884.4176.8962.42
L (Alpha)82.8387.6288.6886.4382.9675.6860.27
", + "image_path": "677a59dd604b3c8083706bea819b2a288c31762fdec4a56e97846a3a19620edc.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 197, + 482, + 241 + ], + "lines": [ + { + "bbox": [ + 130, + 197, + 482, + 241 + ], + "spans": [ + { + "bbox": [ + 130, + 197, + 482, + 241 + ], + "type": "text", + "content": "Table 6: Ablation study of different blur level ranges in contrastive learning and design of blur estimation loss on the synthetic dataset (with 1M distractors). Small: range=1. Medium: range=3. Large: range=5. 'Alpha' corresponds to the alternative design choice for blur estimation loss. The database contains images of all blur levels." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 130, + 255, + 481, + 303 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 255, + 481, + 303 + ], + "spans": [ + { + "bbox": [ + 130, + 255, + 481, + 303 + ], + "type": "text", + "content": "per method, reflecting the robustness of each method to different levels of motion blur. Our method outperforms all other state-of-the-art methods for all settings and also achieves the smallest standard deviation and range, showing the highest robustness to motion blur (Table 4). See supplementary material for more results." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 304, + 482, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 304, + 482, + 426 + ], + "spans": [ + { + "bbox": [ + 130, + 304, + 482, + 426 + ], + "type": "text", + "content": "Qualitative results on our proposed synthetic (Fig. 4a) and real-world (Fig. 4b) datasets. Our method is compared to the top-performing state-of-the-art retrieval methods on a per-query basis, which yield numerous false positives resembling the texture of the motion-blurred object in the query image. In contrast, our approach accurately retrieves positive images even under severe motion blur in the query and amidst challenging distractors. Note that standard retrieval methods retrieve positive images primarily with blur levels similar to the query, whereas our method retrieves correct images exhibiting a wide range of blur levels (from 1-6 for synthetic and 2-6 for real). This illustrates that our method effectively learns invariant representations of objects across various blur levels." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 440, + 237, + 453 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 440, + 237, + 453 + ], + "spans": [ + { + "bbox": [ + 132, + 440, + 237, + 453 + ], + "type": "text", + "content": "5.2 Ablation study" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 460, + 482, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 460, + 482, + 544 + ], + "spans": [ + { + "bbox": [ + 130, + 460, + 482, + 544 + ], + "type": "text", + "content": "Loss components. Table 5 shows an ablation study of different loss functions (for more details, please see the supplementary material). The results show that all loss functions improve model performance. We highlight that our model outperforms the compared methods across all query blur levels (Table 2), even when utilizing only three out of the four loss terms at any combination. For instance, when using " + }, + { + "bbox": [ + 130, + 460, + 482, + 544 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{con}" + }, + { + "bbox": [ + 130, + 460, + 482, + 544 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 460, + 482, + 544 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{cls}" + }, + { + "bbox": [ + 130, + 460, + 482, + 544 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 460, + 482, + 544 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{be}" + }, + { + "bbox": [ + 130, + 460, + 482, + 544 + ], + "type": "text", + "content": ", our method surpasses Token [46] by 2.50 in mAP at blur level 1 and 1.91 at blur level 6, with an overall improvement of 2.33." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 544, + 482, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 544, + 482, + 593 + ], + "spans": [ + { + "bbox": [ + 130, + 544, + 482, + 593 + ], + "type": "text", + "content": "Contrastive range. We investigate the research questions posed in Sec. 3.2 using Table 6, which displays results of using different blur level ranges for contrastive learning. A wider range of contrastive blur levels consistently enhances model performance, with the best results achieved at the maximum range of 5." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": "Design of blur estimation loss. We validate the effectiveness of our design of blur estimation loss " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{be}" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": " (Eq. (7)) in Table 6. We compare with the case where edge effects are ignored in the loss function. This corresponds to directly predicting the average of the entire alpha mask (Eq. (4)) inside the object boundary. The results show that our design of " + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{be}" + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": ", which effectively mitigates the detrimental impact of edge blur, improves the model's retrieval performance." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 283, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 283, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 283, + 91, + 447, + 102 + ], + "type": "text", + "content": "Retrieval Robust to Object Motion Blur" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "bbox": [ + 138, + 114, + 165, + 125 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 114, + 165, + 125 + ], + "spans": [ + { + "bbox": [ + 138, + 114, + 165, + 125 + ], + "type": "text", + "content": "Query" + } + ] + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 138, + 125, + 166, + 159 + ], + "blocks": [ + { + "bbox": [ + 138, + 125, + 166, + 159 + ], + "lines": [ + { + "bbox": [ + 138, + 125, + 166, + 159 + ], + "spans": [ + { + "bbox": [ + 138, + 125, + 166, + 159 + ], + "type": "image", + "image_path": "a494f1580db886cd7639ea060b5a2e60dbde5a8ce3938633c8ad5a762bd24a55.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 171, + 125, + 198, + 165 + ], + "blocks": [ + { + "bbox": [ + 171, + 125, + 198, + 165 + ], + "lines": [ + { + "bbox": [ + 171, + 125, + 198, + 165 + ], + "spans": [ + { + "bbox": [ + 171, + 125, + 198, + 165 + ], + "type": "image", + "image_path": "0a6a6e9bd054fdcdee6d0354fefddce2a6a13fe04fedecbcfac38357560e592b.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 202, + 125, + 228, + 165 + ], + "blocks": [ + { + "bbox": [ + 202, + 125, + 228, + 165 + ], + "lines": [ + { + "bbox": [ + 202, + 125, + 228, + 165 + ], + "spans": [ + { + "bbox": [ + 202, + 125, + 228, + 165 + ], + "type": "image", + "image_path": "030d065373cd3d9d6a3dd21c553c054cbc4fcaa8e562192687e92fb1addae8d6.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 233, + 125, + 260, + 170 + ], + "blocks": [ + { + "bbox": [ + 233, + 125, + 260, + 170 + ], + "lines": [ + { + "bbox": [ + 233, + 125, + 260, + 170 + ], + "spans": [ + { + "bbox": [ + 233, + 125, + 260, + 170 + ], + "type": "image", + "image_path": "d284e9ca6d7e72a065661071cab88716aa8831fc322cb3d3769f3b3d866f5098.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 263, + 115, + 370, + 125 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 263, + 115, + 370, + 125 + ], + "spans": [ + { + "bbox": [ + 263, + 115, + 370, + 125 + ], + "type": "text", + "content": "Top 20 retrieval results" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 263, + 125, + 290, + 170 + ], + "blocks": [ + { + "bbox": [ + 263, + 125, + 290, + 170 + ], + "lines": [ + { + "bbox": [ + 263, + 125, + 290, + 170 + ], + "spans": [ + { + "bbox": [ + 263, + 125, + 290, + 170 + ], + "type": "image", + "image_path": "36412bd7feaaac7a3590b476d2c376f50079dd10e3a0276277f66f8316b9cfe4.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 295, + 125, + 321, + 170 + ], + "blocks": [ + { + "bbox": [ + 295, + 125, + 321, + 170 + ], + "lines": [ + { + "bbox": [ + 295, + 125, + 321, + 170 + ], + "spans": [ + { + "bbox": [ + 295, + 125, + 321, + 170 + ], + "type": "image", + "image_path": "88dbe6e78998eaa391e0519e3f6b89c65ebcda9e1535a705fb20122bd5c1d618.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 324, + 125, + 352, + 170 + ], + "blocks": [ + { + "bbox": [ + 324, + 125, + 352, + 170 + ], + "lines": [ + { + "bbox": [ + 324, + 125, + 352, + 170 + ], + "spans": [ + { + "bbox": [ + 324, + 125, + 352, + 170 + ], + "type": "image", + "image_path": "2ace5f59f5ec63f43f9639a097d46e54f4e615382eebc7fff23700b5fe2a8ce1.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 356, + 125, + 382, + 170 + ], + "blocks": [ + { + "bbox": [ + 356, + 125, + 382, + 170 + ], + "lines": [ + { + "bbox": [ + 356, + 125, + 382, + 170 + ], + "spans": [ + { + "bbox": [ + 356, + 125, + 382, + 170 + ], + "type": "image", + "image_path": "08c33429958f42c562b164d7e66c3a852d0a501db5cb388ba76014198c52b14d.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 386, + 125, + 413, + 170 + ], + "blocks": [ + { + "bbox": [ + 386, + 125, + 413, + 170 + ], + "lines": [ + { + "bbox": [ + 386, + 125, + 413, + 170 + ], + "spans": [ + { + "bbox": [ + 386, + 125, + 413, + 170 + ], + "type": "image", + "image_path": "0a3143d0ac5601b487fd1a8309dd3be9767bdd97766f14db649d2a1de4a9acde.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 417, + 125, + 444, + 170 + ], + "blocks": [ + { + "bbox": [ + 417, + 125, + 444, + 170 + ], + "lines": [ + { + "bbox": [ + 417, + 125, + 444, + 170 + ], + "spans": [ + { + "bbox": [ + 417, + 125, + 444, + 170 + ], + "type": "image", + "image_path": "0830a0141f625ab6230ad0efc105f733c93aa43e9c6df55ddffed236642e9fa9.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 447, + 125, + 475, + 170 + ], + "blocks": [ + { + "bbox": [ + 447, + 125, + 475, + 170 + ], + "lines": [ + { + "bbox": [ + 447, + 125, + 475, + 170 + ], + "spans": [ + { + "bbox": [ + 447, + 125, + 475, + 170 + ], + "type": "image", + "image_path": "1f295044049f0b5aeacdbb99949d71fbf17f3ac65d18bb98d8e53a6376d73486.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 137, + 189, + 166, + 210 + ], + "blocks": [ + { + "bbox": [ + 137, + 189, + 166, + 210 + ], + "lines": [ + { + "bbox": [ + 137, + 189, + 166, + 210 + ], + "spans": [ + { + "bbox": [ + 137, + 189, + 166, + 210 + ], + "type": "image", + "image_path": "d8ab84f06d966f6244be744dd91f6d4a04bd3c04b9deb7c0ee1e19ffad06c4ae.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 224, + 481, + 258 + ], + "lines": [ + { + "bbox": [ + 132, + 224, + 481, + 258 + ], + "spans": [ + { + "bbox": [ + 132, + 224, + 481, + 258 + ], + "type": "text", + "content": "Fig. 5: Retrieval results on a real-world video (available on YouTube). The retrieved images are sorted from left to right and top to bottom, with the ranking from " + }, + { + "bbox": [ + 132, + 224, + 481, + 258 + ], + "type": "inline_equation", + "content": "1^{\\text{st}}" + }, + { + "bbox": [ + 132, + 224, + 481, + 258 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 132, + 224, + 481, + 258 + ], + "type": "inline_equation", + "content": "20^{\\text{th}}" + }, + { + "bbox": [ + 132, + 224, + 481, + 258 + ], + "type": "text", + "content": ". Correct results are in green boxes, while incorrect ones are in red boxes." + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 171, + 177, + 198, + 210 + ], + "blocks": [ + { + "bbox": [ + 171, + 177, + 198, + 210 + ], + "lines": [ + { + "bbox": [ + 171, + 177, + 198, + 210 + ], + "spans": [ + { + "bbox": [ + 171, + 177, + 198, + 210 + ], + "type": "image", + "image_path": "b80286f23140b80f8044c7c783af99f3ab5a34ac8de1fe4b8f190f75ce0eb272.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 202, + 177, + 228, + 210 + ], + "blocks": [ + { + "bbox": [ + 202, + 177, + 228, + 210 + ], + "lines": [ + { + "bbox": [ + 202, + 177, + 228, + 210 + ], + "spans": [ + { + "bbox": [ + 202, + 177, + 228, + 210 + ], + "type": "image", + "image_path": "0d68963c74d90708e528daafc7c2daa80eccdbfd1a522a2bcc49003f14760864.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 233, + 177, + 260, + 210 + ], + "blocks": [ + { + "bbox": [ + 233, + 177, + 260, + 210 + ], + "lines": [ + { + "bbox": [ + 233, + 177, + 260, + 210 + ], + "spans": [ + { + "bbox": [ + 233, + 177, + 260, + 210 + ], + "type": "image", + "image_path": "f35be9a90c52019ffcc7a43e44b2e6b56f2ceea2dd4d185935df010b1aac6747.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 263, + 177, + 290, + 210 + ], + "blocks": [ + { + "bbox": [ + 263, + 177, + 290, + 210 + ], + "lines": [ + { + "bbox": [ + 263, + 177, + 290, + 210 + ], + "spans": [ + { + "bbox": [ + 263, + 177, + 290, + 210 + ], + "type": "image", + "image_path": "a1bb25c9044d727f08000882b42c0a000a963eb6d2be208d1d5a7fae5c37047a.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 294, + 177, + 321, + 210 + ], + "blocks": [ + { + "bbox": [ + 294, + 177, + 321, + 210 + ], + "lines": [ + { + "bbox": [ + 294, + 177, + 321, + 210 + ], + "spans": [ + { + "bbox": [ + 294, + 177, + 321, + 210 + ], + "type": "image", + "image_path": "20c1d50cf4b5b96ce631b016d9042af1576cb104f792ec4a1e134f92a898d5b4.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 324, + 177, + 352, + 210 + ], + "blocks": [ + { + "bbox": [ + 324, + 177, + 352, + 210 + ], + "lines": [ + { + "bbox": [ + 324, + 177, + 352, + 210 + ], + "spans": [ + { + "bbox": [ + 324, + 177, + 352, + 210 + ], + "type": "image", + "image_path": "a3c1e535c949266694e589f54c4c82fd37e0af5e38de22734fbfc9ee95197f58.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 356, + 177, + 382, + 210 + ], + "blocks": [ + { + "bbox": [ + 356, + 177, + 382, + 210 + ], + "lines": [ + { + "bbox": [ + 356, + 177, + 382, + 210 + ], + "spans": [ + { + "bbox": [ + 356, + 177, + 382, + 210 + ], + "type": "image", + "image_path": "b6807b3c17d3e03a0a23fc3153449d104f75c95bd7296cba3231380c3d66ba26.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 386, + 177, + 413, + 210 + ], + "blocks": [ + { + "bbox": [ + 386, + 177, + 413, + 210 + ], + "lines": [ + { + "bbox": [ + 386, + 177, + 413, + 210 + ], + "spans": [ + { + "bbox": [ + 386, + 177, + 413, + 210 + ], + "type": "image", + "image_path": "b9373c1d0499af4ea1005997eafa23395ffaef94f215889cb0829d2944ee1302.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 417, + 177, + 444, + 210 + ], + "blocks": [ + { + "bbox": [ + 417, + 177, + 444, + 210 + ], + "lines": [ + { + "bbox": [ + 417, + 177, + 444, + 210 + ], + "spans": [ + { + "bbox": [ + 417, + 177, + 444, + 210 + ], + "type": "image", + "image_path": "056cb53277149902596142038df0e148175206702c6ef3de25313ef90c936194.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 447, + 177, + 475, + 210 + ], + "blocks": [ + { + "bbox": [ + 447, + 177, + 475, + 210 + ], + "lines": [ + { + "bbox": [ + 447, + 177, + 475, + 210 + ], + "spans": [ + { + "bbox": [ + 447, + 177, + 475, + 210 + ], + "type": "image", + "image_path": "648ddb065bd600ef40956baa2cd1b5e85b65713a61d0bbdc3080282e56bbf3d5.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + } + ], + "index": 25 + }, + { + "bbox": [ + 132, + 281, + 348, + 293 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 281, + 348, + 293 + ], + "spans": [ + { + "bbox": [ + 132, + 281, + 348, + 293 + ], + "type": "text", + "content": "5.3 Application to real-world video data" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 130, + 300, + 482, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 300, + 482, + 445 + ], + "spans": [ + { + "bbox": [ + 130, + 300, + 482, + 445 + ], + "type": "text", + "content": "We conduct an experiment to test our method on real-world video images. Due to the challenge of finding a sufficient number of real images of the same object with a well-distributed range of blur levels for quantitative analysis, we construct a small dataset for qualitative assessment. Specifically, we extract 190 images of the same ball from a YouTube soccer video to serve as both the query and database. We incorporate 4,431 images of sports balls from the MSCOCO dataset [21] as distractors, and additionally, we extract 169 images of a different ball from the same video to act as hard distractors. Both the objects and the backgrounds in these images come from domains entirely different from the synthetic training data. Fig. 5 illustrates examples of retrieval results, demonstrating our method's effectiveness in handling various blur conditions and complex and diverse backgrounds in the real world, despite being trained exclusively on synthetic data." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 132, + 473, + 220, + 485 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 473, + 220, + 485 + ], + "spans": [ + { + "bbox": [ + 132, + 473, + 220, + 485 + ], + "type": "text", + "content": "6 Conclusion" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 130, + 497, + 482, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 497, + 482, + 640 + ], + "spans": [ + { + "bbox": [ + 130, + 497, + 482, + 640 + ], + "type": "text", + "content": "In this paper, we introduce an innovative task: image retrieval in the presence of object motion blur. To tackle this challenge, we propose a global retrieval approach that integrates specialized loss functions tailored for the network to better comprehend motion blur. Our method stands as the first specifically designed solution for this new task. Moreover, we introduce the first benchmark datasets for blur retrieval, comprising both synthetic and real-world data. These datasets are meticulously processed and annotated to ensure their immediate applicability for future research in this emerging area. Through extensive experiments conducted on these benchmark datasets, we demonstrate the limitations of state-of-the-art standard retrieval methods in handling motion-blurred images and showcase the superior performance and robustness of our proposed approach in accurately retrieving images even in challenging motion-blurred scenarios." + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 131, + 641, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 641, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 131, + 641, + 481, + 665 + ], + "type": "text", + "content": "Acknowledgement. We would like to thank Giorgos Tolias for the highly valuable discussions with him and the support he provided for this work." + } + ] + } + ], + "index": 31 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "text", + "content": "R. Zou et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 197, + 126 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 197, + 126 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 197, + 126 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 134, + 141, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 138, + 141, + 479, + 164 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 141, + 479, + 164 + ], + "spans": [ + { + "bbox": [ + 138, + 141, + 479, + 164 + ], + "type": "text", + "content": "1. Babenko, A., Lempitsky, V.: Aggregating local deep features for image retrieval. In: ICCV. pp. 1269-1277 (2015)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 165, + 480, + 186 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 165, + 480, + 186 + ], + "spans": [ + { + "bbox": [ + 138, + 165, + 480, + 186 + ], + "type": "text", + "content": "2. Babenko, A., Slesarev, A., Chigorin, A., Lempitsky, V.: Neural codes for image retrieval. In: ECCV. pp. 584-599. Springer (2014)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 187, + 480, + 209 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 187, + 480, + 209 + ], + "spans": [ + { + "bbox": [ + 138, + 187, + 480, + 209 + ], + "type": "text", + "content": "3. Bay, H., Ess, A., Tuytelaars, T., Van Gool, L.: Speeded-up robust features (surf). Computer Vision and Image Understanding 110(3), 346-359 (2008)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 209, + 480, + 243 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 209, + 480, + 243 + ], + "spans": [ + { + "bbox": [ + 138, + 209, + 480, + 243 + ], + "type": "text", + "content": "4. Cao, B., Araujo, A., Sim, J.: Unifying deep local and global features for image search. In: Vedaldi, A., Bischof, H., Brox, T., Frahm, J.M. (eds.) ECCV. pp. 726-743. Springer International Publishing, Cham (2020)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 243, + 480, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 243, + 480, + 297 + ], + "spans": [ + { + "bbox": [ + 138, + 243, + 480, + 297 + ], + "type": "text", + "content": "5. Chang, A.X., Funkhouser, T., Guibas, L., Hanrahan, P., Huang, Q., Li, Z., Savarese, S., Savva, M., Song, S., Su, H., Xiao, J., Yi, L., Yu, F.: ShapeNet: An Information-Rich 3D Model Repository. Tech. Rep. arXiv:1512.03012 [cs.GR], Stanford University — Princeton University — Toyota Technological Institute at Chicago (2015)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 298, + 480, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 298, + 480, + 319 + ], + "spans": [ + { + "bbox": [ + 138, + 298, + 480, + 319 + ], + "type": "text", + "content": "6. Chopra, S., Hadsell, R., LeCun, Y.: Learning a similarity metric discriminatively, with application to face verification. In: CVPR. vol. 1, pp. 539-546. IEEE (2005)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 321, + 480, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 321, + 480, + 354 + ], + "spans": [ + { + "bbox": [ + 138, + 321, + 480, + 354 + ], + "type": "text", + "content": "7. Community, B.O.: Blender - a 3D modelling and rendering package. Blender Foundation, Stichting Blender Foundation, Amsterdam (2018), http://www.blender.org" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 354, + 480, + 375 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 354, + 480, + 375 + ], + "spans": [ + { + "bbox": [ + 138, + 354, + 480, + 375 + ], + "type": "text", + "content": "8. Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: ImageNet: A LargeScale Hierarchical Image Database. In: CVPR (2009)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 376, + 480, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 376, + 480, + 418 + ], + "spans": [ + { + "bbox": [ + 138, + 376, + 480, + 418 + ], + "type": "text", + "content": "9. Deng, J., Guo, J., Xue, N., Zafeiriou, S.: Arcface: Additive angular margin loss for deep face recognition. In: 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 4685-4694 (2019). https://doi.org/10.1109/CVPR.2019.00482" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 134, + 420, + 480, + 463 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 420, + 480, + 463 + ], + "spans": [ + { + "bbox": [ + 134, + 420, + 480, + 463 + ], + "type": "text", + "content": "10. Gong, D., Yang, J., Liu, L., Zhang, Y., Reid, I., Shen, C., Van Den Hengel, A., Shi, Q.: From motion blur to motion flow: A deep learning solution for removing heterogeneous motion blur. In: CVPR. pp. 3806-3815 (2017). https://doi.org/10.1109/CVPR.2017.405" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 134, + 464, + 480, + 486 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 464, + 480, + 486 + ], + "spans": [ + { + "bbox": [ + 134, + 464, + 480, + 486 + ], + "type": "text", + "content": "11. Gordo, A., Almazan, J., Revaud, J., Larlus, D.: Deep image retrieval: Learning global representations for image search (2016)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 134, + 487, + 480, + 509 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 487, + 480, + 509 + ], + "spans": [ + { + "bbox": [ + 134, + 487, + 480, + 509 + ], + "type": "text", + "content": "12. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: CVPR. pp. 770-778 (2016)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 134, + 510, + 480, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 510, + 480, + 542 + ], + "spans": [ + { + "bbox": [ + 134, + 510, + 480, + 542 + ], + "type": "text", + "content": "13. Hendrycks, D., Dietterich, T.: Benchmarking neural network robustness to common corruptions and perturbations. Proceedings of the International Conference on Learning Representations (2019)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 134, + 543, + 480, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 543, + 480, + 564 + ], + "spans": [ + { + "bbox": [ + 134, + 543, + 480, + 564 + ], + "type": "text", + "content": "14. Jaesung, R., Haeyun, L., Jucheol, W., Sunghyun, C.: Real-world blur dataset for learning and benchmarking deblurring algorithms. In: ECCV (2020)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 134, + 565, + 480, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 565, + 480, + 597 + ], + "spans": [ + { + "bbox": [ + 134, + 565, + 480, + 597 + ], + "type": "text", + "content": "15. Jégou, H., Perronnin, F., Douze, M., Sánchez, J., Pérez, P., Schmid, C.: Aggregating local image descriptors into compact codes. IEEE TPAMI 34(9), 1704-1716 (2011)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 134, + 599, + 480, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 599, + 480, + 620 + ], + "spans": [ + { + "bbox": [ + 134, + 599, + 480, + 620 + ], + "type": "text", + "content": "16. Kalantidis, Y., Mellina, C., Osindero, S.: Cross-dimensional weighting for aggregated deep convolutional features. In: ECCVW. pp. 685-701. Springer (2016)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 134, + 621, + 480, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 621, + 480, + 643 + ], + "spans": [ + { + "bbox": [ + 134, + 621, + 480, + 643 + ], + "type": "text", + "content": "17. Kingma, D.P., Ba, J.: Adam: A method for stochastic optimization. In: Bengio, Y., LeCun, Y. (eds.) ICLR (2015), http://arxiv.org/abs/1412.6980" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 134, + 643, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 643, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 134, + 643, + 480, + 665 + ], + "type": "text", + "content": "18. Kotera, J., Rozumnyi, D., Sroubek, F., Matas, J.: Intra-frame object tracking by deblatting. In: ICCVW (Oct 2019)" + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 283, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 283, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 283, + 91, + 447, + 102 + ], + "type": "text", + "content": "Retrieval Robust to Object Motion Blur" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 117, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 133, + 117, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 117, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 133, + 117, + 480, + 138 + ], + "type": "text", + "content": "19. Kupyn, O., Budzan, V., Mykhailych, M., Mishkin, D., Matas, J.: Deblurgan: Blind motion deblurring using conditional adversarial networks. In: CVPR (June 2018)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 140, + 480, + 159 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 140, + 480, + 159 + ], + "spans": [ + { + "bbox": [ + 133, + 140, + 480, + 159 + ], + "type": "text", + "content": "20. Lee, S., Seong, H., Lee, S., Kim, E.: Correlation verification for image retrieval. In: CVPR. pp. 5374-5384 (2022)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 160, + 480, + 192 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 160, + 480, + 192 + ], + "spans": [ + { + "bbox": [ + 133, + 160, + 480, + 192 + ], + "type": "text", + "content": "21. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dólar, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: ECCV. pp. 740-755. Springer (2014)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 193, + 480, + 213 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 193, + 480, + 213 + ], + "spans": [ + { + "bbox": [ + 133, + 193, + 480, + 213 + ], + "type": "text", + "content": "22. Lowe, D.G.: Distinctive image features from scale-invariant keypoints. IJCV 60, 91-110 (2004)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 213, + 480, + 246 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 213, + 480, + 246 + ], + "spans": [ + { + "bbox": [ + 132, + 213, + 480, + 246 + ], + "type": "text", + "content": "23. Michaelis, C., Mitzkus, B., Geirhos, R., Rusak, E., Bringmann, O., Ecker, A.S., Bethge, M., Brendel, W.: Benchmarking robustness in object detection: Autonomous driving when winter is coming. arXiv preprint arXiv:1907.07484 (2019)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 247, + 480, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 247, + 480, + 267 + ], + "spans": [ + { + "bbox": [ + 132, + 247, + 480, + 267 + ], + "type": "text", + "content": "24. Nah, S., Kim, T.H., Lee, K.M.: Deep multi-scale convolutional neural network for dynamic scene deblurring. In: CVPR (July 2017)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 268, + 480, + 289 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 268, + 480, + 289 + ], + "spans": [ + { + "bbox": [ + 132, + 268, + 480, + 289 + ], + "type": "text", + "content": "25. Noh, H., Araujo, A., Sim, J., Weyand, T., Han, B.: Large-scale image retrieval with attentive deep local features. In: ICCV. pp. 3456-3465 (2017)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 289, + 480, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 289, + 480, + 376 + ], + "spans": [ + { + "bbox": [ + 132, + 289, + 480, + 376 + ], + "type": "text", + "content": "26. Paszke, A., Gross, S., Massa, F., Lerer, A., Bradbury, J., Chanan, G., Killeen, T., Lin, Z., Gimelshein, N., Antiga, L., Desmaison, A., Kopf, A., Yang, E., Devito, Z., Raison, M., Tejani, A., Chilamkurthy, S., Steiner, B., Fang, L., Bai, J., Chintala, S.: Pytorch: An imperative style, high-performance deep learning library. In: Wallach, H., Larochelle, H., Beygelzimer, A., Alché-Buc, F., Fox, E., Garnett, R. (eds.) NeurIPS, pp. 8024-8035. Curran Associates, Inc. (2019), http://papers.neurips.cc/paper/9015-pytorch-an-imperative-style-high-performance-deep-learning-library.pdf" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 376, + 480, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 376, + 480, + 397 + ], + "spans": [ + { + "bbox": [ + 132, + 376, + 480, + 397 + ], + "type": "text", + "content": "27. Perronnin, F., Liu, Y., Sánchez, J., Poirier, H.: Large-scale image retrieval with compressed fisher vectors. In: CVPR. pp. 3384-3391. IEEE (2010)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 398, + 480, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 398, + 480, + 418 + ], + "spans": [ + { + "bbox": [ + 132, + 398, + 480, + 418 + ], + "type": "text", + "content": "28. Radenović, F., Iscen, A., Tolias, G., Avrithis, Y., Chum, O.: Revisiting oxford and paris: Large-scale image retrieval benchmarking. In: CVPR. pp. 5706-5715 (2018)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 419, + 480, + 439 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 419, + 480, + 439 + ], + "spans": [ + { + "bbox": [ + 132, + 419, + 480, + 439 + ], + "type": "text", + "content": "29. Radenović, F., Tobias, G., Chum, O.: Fine-tuning cnn image retrieval with no human annotation. arXiv preprint arXiv:1711.02512 (2017)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 133, + 440, + 480, + 471 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 440, + 480, + 471 + ], + "spans": [ + { + "bbox": [ + 133, + 440, + 480, + 471 + ], + "type": "text", + "content": "30. Radenović, F., Tolias, G., Chum, O.: Fine-tuning cnn image retrieval with no human annotation. IEEE TPAMI 41(7), 1655-1668 (2019). https://doi.org/10.1109/TPAMI.2018.2846566" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 133, + 472, + 480, + 504 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 472, + 480, + 504 + ], + "spans": [ + { + "bbox": [ + 133, + 472, + 480, + 504 + ], + "type": "text", + "content": "31. Rozumnyi, D., Kotera, J., Sroubek, F., Matas, J.: Non-causal tracking by deblatting. In: Fink, G.A., Frintrop, S., Jiang, X. (eds.) GCPR. pp. 122-135. Springer International Publishing, Cham (2019)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 133, + 505, + 480, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 505, + 480, + 536 + ], + "spans": [ + { + "bbox": [ + 133, + 505, + 480, + 536 + ], + "type": "text", + "content": "32. Rozumnyi, D., Kotera, J., Sroubek, F., Novotny, L., Matas, J.: The world of fast moving objects. In: CVPR. pp. 4838-4846 (July 2017). https://doi.org/10.1109/CVPR.2017.514" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 133, + 536, + 480, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 536, + 480, + 557 + ], + "spans": [ + { + "bbox": [ + 133, + 536, + 480, + 557 + ], + "type": "text", + "content": "33. Rozumnyi, D., Kotera, J., Sroubek, F., Matas, J.: Sub-frame appearance and 6d pose estimation of fast moving objects. In: CVPR. pp. 6777-6785 (2020)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 133, + 558, + 480, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 558, + 480, + 590 + ], + "spans": [ + { + "bbox": [ + 133, + 558, + 480, + 590 + ], + "type": "text", + "content": "34. Rozumnyi, D., Oswald, M.R., Ferrari, V., Matas, J., Pollefeys, M.: Defmo: Deblurring and shape recovery of fast moving objects. In: CVPR. Nashville, Tennessee, USA (Jun 2021)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 590, + 480, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 590, + 480, + 611 + ], + "spans": [ + { + "bbox": [ + 132, + 590, + 480, + 611 + ], + "type": "text", + "content": "35. Rozumnyi, D., Oswald, M.R., Ferrari, V., Pollefeys, M.: Shape from blur: Recovering textured 3d shape and motion of fast moving objects. In: NeurIPS (2021)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 133, + 612, + 480, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 612, + 480, + 643 + ], + "spans": [ + { + "bbox": [ + 133, + 612, + 480, + 643 + ], + "type": "text", + "content": "36. Rozumnyi, D., Oswald, M.R., Ferrari, V., Pollefeys, M.: Motion-from-blur: 3d shape and motion estimation of motion-blurred objects in videos. In: CVPR (Jun 2022)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 132, + 643, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 643, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 643, + 480, + 665 + ], + "type": "text", + "content": "37. Sayed, M., Brostow, G.: Improved handling of motion blur in online object detection. In: CVPR. pp. 1706-1716 (June 2021)" + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 219, + 100 + ], + "type": "text", + "content": "R. Zou et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 468 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 138 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 138 + ], + "type": "text", + "content": "38. Sivic, Zisserman: Video google: A text retrieval approach to object matching in videos. In: ICCV. pp. 1470-1477. IEEE (2003)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 138, + 481, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 138, + 481, + 160 + ], + "spans": [ + { + "bbox": [ + 130, + 138, + 481, + 160 + ], + "type": "text", + "content": "39. Skorokhodov, I., Sotnikov, G., Elhoseiny, M.: Aligning latent and image spaces to connect the unconnectable. In: ICCV. pp. 14144-14153 (2021)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 160, + 481, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 160, + 481, + 182 + ], + "spans": [ + { + "bbox": [ + 130, + 160, + 481, + 182 + ], + "type": "text", + "content": "40. Tolias, G., Avrithis, Y., Jégou, H.: Image search with selective match kernels: aggregation across single and multiple images. IJCV 116, 247-261 (2016)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 182, + 481, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 182, + 481, + 205 + ], + "spans": [ + { + "bbox": [ + 130, + 182, + 481, + 205 + ], + "type": "text", + "content": "41. Tolias, G., Jenicek, T., Chum, O.: Learning and aggregating deep local descriptors for instance-level recognition. In: ECCV. pp. 460-477. Springer (2020)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 205, + 481, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 205, + 481, + 226 + ], + "spans": [ + { + "bbox": [ + 130, + 205, + 481, + 226 + ], + "type": "text", + "content": "42. Tolias, G., Sicre, R., Jégou, H.: Particular object retrieval with integral max-pooling of cnn activations. ICLR (2016)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 226, + 481, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 226, + 481, + 248 + ], + "spans": [ + { + "bbox": [ + 130, + 226, + 481, + 248 + ], + "type": "text", + "content": "43. Vasiljevic, I., Chakrabarti, A., Shakhnarovich, G.: Examining the impact of blur on recognition by convolutional networks. arXiv preprint arXiv:1611.05760 (2016)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 248, + 481, + 280 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 248, + 481, + 280 + ], + "spans": [ + { + "bbox": [ + 130, + 248, + 481, + 280 + ], + "type": "text", + "content": "44. Wang, Y., Lu, Y., Gao, Y., Wang, L., Zhong, Z., Zheng, Y., Yamashita, A.: Efficient video deblurring guided by motion magnitude. In: ECCV. pp. 413-429. Springer (2022)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 280, + 481, + 313 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 280, + 481, + 313 + ], + "spans": [ + { + "bbox": [ + 130, + 280, + 481, + 313 + ], + "type": "text", + "content": "45. Weyand, T., Araujo, A., Cao, B., Sim, J.: Google landmarks dataset v2-a large-scale benchmark for instance-level recognition and retrieval. In: CVPR. pp. 2574-2583 (2020)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 313, + 481, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 313, + 481, + 346 + ], + "spans": [ + { + "bbox": [ + 130, + 313, + 481, + 346 + ], + "type": "text", + "content": "46. Wu, H., Wang, M., Zhou, W.T., Hu, Y., Li, H.: Learning token-based representation for image retrieval. In: AAAI Conference on Artificial Intelligence (2021), https://apisemantic scholar.org/CorpusID:245124518" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 346, + 481, + 401 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 346, + 481, + 401 + ], + "spans": [ + { + "bbox": [ + 130, + 346, + 481, + 401 + ], + "type": "text", + "content": "47. Yang, M., He, D., Fan, M., Shi, B., Xue, X., Li, F., Ding, E., Huang, J.: Dolg: Single-stage image retrieval with deep orthogonal fusion of local and global features. In: ICCV. pp. 11752-11761. IEEE Computer Society, Los Alamitos, CA, USA (oct 2021). https://doi.org/10.1109/ICCV48922.2021.01156, https://doi.ieeecomputersociety.org/10.1109/ICCV48922.2021.01156" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 401, + 481, + 424 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 401, + 481, + 424 + ], + "spans": [ + { + "bbox": [ + 130, + 401, + 481, + 424 + ], + "type": "text", + "content": "48. Yokoo, S., Ozaki, K., Simo-Serra, E., Iizuka, S.: Two-stage discriminative reranking for large-scale landmark retrieval. In: CVPRW. IEEE (2020)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 424, + 481, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 424, + 481, + 445 + ], + "spans": [ + { + "bbox": [ + 130, + 424, + 481, + 445 + ], + "type": "text", + "content": "49. Zhang, H., Xie, H., Yao, H.: Spatio-temporal deformable attention network for video deblurring. In: ECCV. pp. 581-596. Springer (2022)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 445, + 481, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 445, + 481, + 468 + ], + "spans": [ + { + "bbox": [ + 130, + 445, + 481, + 468 + ], + "type": "text", + "content": "50. Zhou, S., Zhang, J., Pan, J., Xie, H., Zuo, W., Ren, J.: Spatio-temporal filter adaptive network for video deblurring. In: ICCV (2019)" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 283, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 283, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 283, + 91, + 447, + 102 + ], + "type": "text", + "content": "Retrieval Robust to Object Motion Blur" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Revising Densification in Gaussian Splatting/918310b7-b91b-411d-ade7-18df636ce501_content_list.json b/2024/Revising Densification in Gaussian Splatting/918310b7-b91b-411d-ade7-18df636ce501_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..ba4a10ebb3f364069ecb9c9b234b52907f64cc92 --- /dev/null +++ b/2024/Revising Densification in Gaussian Splatting/918310b7-b91b-411d-ade7-18df636ce501_content_list.json @@ -0,0 +1,1893 @@ +[ + { + "type": "text", + "text": "Revising Densification in Gaussian Splitting", + "text_level": 1, + "bbox": [ + 246, + 141, + 754, + 162 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Samuel Rota Bulò, Lorenzo Porzi, and Peter Kontschieder", + "bbox": [ + 264, + 189, + 736, + 205 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Meta Reality Labs Zurich", + "bbox": [ + 411, + 217, + 584, + 231 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/01da4aa633a1585e8b5f170e748342671bc84b2736e2a61b43c97120485ef712.jpg", + "image_caption": [ + "Fig. 1: Densification is a critical component of 3D Gaussian Splatting (3DGS), and a common failure point. In this example (ground truth on the left) we show how 3DGS can fail (center) to add primitives to high-texture areas, like the grass in the bottom part of the pictures, producing large and blurry artifacts. Our approach (right) solves this issue by comprehensively revising densification in 3DGS." + ], + "image_footnote": [], + "bbox": [ + 217, + 256, + 406, + 354 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/0aa0650b8e46425032dbe6a5487b289f066b1aa6f81b3dbb18502e95b264f3cb.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 406, + 256, + 596, + 354 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/3d4d96cac9f0c8b618afe30ec5fa20aa673d537d5373f7e503097eb836cffc54.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 596, + 256, + 785, + 354 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. In this paper, we address the limitations of Adaptive Density Control (ADC) in 3D Gaussian Splatting (3DGS), a scene representation method achieving high-quality, photorealistic results for novel view synthesis. ADC has been introduced for automatic 3D point primitive management, controlling densification and pruning, however, with certain limitations in the densification logic. Our main contribution is a more principled, pixel-error driven formulation for density control in 3DGS, leveraging an auxiliary, per-pixel error function as the criterion for densification. We further introduce a mechanism to control the total number of primitives generated per scene and correct a bias in the current opacity handling strategy of ADC during cloning operations. Our approach leads to consistent quality improvements across a variety of benchmark scenes, without sacrificing the method's efficiency.", + "bbox": [ + 259, + 455, + 740, + 635 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Gaussian Splatting $\\cdot$ 3D reconstruction $\\cdot$ Novel View Synthesis", + "bbox": [ + 259, + 648, + 740, + 675 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 702, + 375, + 718 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "High-quality, photorealistic scene modelling from images has been an important research area in computer vision and graphics, with plentiful applications in AR/VR/MR, robotics, etc. In the last years, this field has gained a lot of attention due to advances in Neural 3D scene representations, particularly Neural Radiance Fields (NeRFs) [17]. NeRFs take a new approach to 3D scene representation and rendering, by leveraging a combination of deep learning and volumetric rendering techniques for generating photorealistic images from novel", + "bbox": [ + 212, + 733, + 787, + 840 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "viewpoints. By optimizing MLPs to map from spatial coordinates and viewing directions to density and colour fields, these models have demonstrated astonishing capabilities for capturing the complex interplay of light and geometry in a data-driven way. While highly efficient in terms of representation quality, the original NeRF representation relies on time-consuming sampling strategies and thus excludes applications with fast rendering requirements. With many advances in terms of the underlying representation, these models have been significantly optimized towards improved training time and scene representation fidelity. However, inference speed for high-resolution, novel view synthesis remains an ongoing limitation.", + "bbox": [ + 212, + 146, + 787, + 297 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "More recently, 3D Gaussian Splatting (3DGS) [9] has been proposed as an alternative and expressive scene representation, enabling both high-speed, high-fidelity training of models and high-resolution, GPU rasterization-friendly rendering of novel views. Their core representation is an optimized set of (anisotropic) 3D Gaussians, after being randomly distributed in 3D space, or systematically initialized at points obtained by Structure-from-Motion [20]. For obtaining a 2D image, all relevant 3D primitives are efficiently rendered via splatting-based rasterization with low-pass filtering.", + "bbox": [ + 212, + 301, + 789, + 422 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In 3DGS, each 3D primitive is parameterized as a 3D Gaussian distribution (i.e., with position and covariance), together with parameters controlling its opacity and describing its directional appearance (typically spherical harmonics). The parameter optimization procedure is guided by a multi-view, photometric loss, and is interleaved with Adaptive Density Control (ADC), a mechanism controlling density management for 3D points by means of introducing or deleting 3D primitives. ADC plays a critical role as it determines where to expand/shrink the scene representation budget for empty or over-reconstructed regions, respectively. Both growing and pruning operations are activated based on user-defined thresholds: Growing depends on the accumulated positional gradients of existing primitives and is, conditioned on the size of the Gaussians, executed by either splitting large primitives or by cloning smaller ones. Pruning is activated once the opacity falls below a provided threshold. While quite effective in practice, such density management strategies have several limitations. First, estimating a gradient magnitude-based threshold is rather non-intuitive and not robust to potential changes in the model, loss terms, etc. Second, there are cases where only few and large Gaussians are modeling high-frequency patterns like grass as shown in the middle of Fig. 1. Here, changes accumulated from positional gradients might remain very low and thus fail to trigger the densification mechanism, which in turn leads to substantial scene underfitting. Finally, ADC lacks explicit control of the maximum number of Gaussians generated per scene. This has important, practical implications as uncontrolled growth might easily lead to out-of-memory errors during training.", + "bbox": [ + 212, + 426, + 787, + 775 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this work we address the shortcomings of Adaptive Density Control proposed in the original 3D Gaussian splatting method. Our core contribution is a more principled, pixel-error driven formulation for density control in 3DGS. We describe how 2D, per-pixel errors as e.g. derived from Structural Similarity", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "Rota Bulò et al.", + "bbox": [ + 271, + 114, + 380, + 126 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "(or any other informative objective function) can be propagated back as errors to contributing Gaussian primitives. In our solution, we first break down the per-pixel errors according to each Gaussian's contribution, and in a camera-specific way. This allows us to track the maximum error per primitive for all views and across two subsequent ADC runs, yielding our novel, error-specific, and thus more intuitive decision criterion for densification. Our second contribution is correcting a bias introduced with the current form of opacity handling in ADC when conducting a primitive cloning operation. The original approach suggests to keep the same opacity for the cloned Gaussian, which however biases the alpha-compositing logic applied for rendering the pixel colors. Indeed, this procedure leads to an overall increase of opacity in the cloned region, preventing the model to correctly account for contributions of other primitives and thus negatively affecting the densification process. Our third contribution is a mechanism for controlling the total number of primitives generated per scene and the maximum amount of novel primitives introduced per densification run. With this functionality, we can avoid undesired out-of-memory errors and better tune the method's behaviour w.r.t. given hardware constraints. We extensively validate our contributions on standard benchmark datasets like Mip-NeRF 360 [1], Tanks and Temples [10], and Deep Blending [6]. Our experiments show consistent improvements over different baselines including 3DGS [9] and Mip-Splatting [29]. To summarize, our contributions are improving methodological shortcomings in 3DGS' Adaptive Density Control mechanism as follows:", + "bbox": [ + 217, + 146, + 785, + 478 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose a principled approach that enables the guidance of the densification process according to an auxiliary, per-pixel error function, rather than relying on positional gradients.", + "- We correct an existing, systematic bias from the primitive growing procedure when cloning Gaussians, negatively impacting the overall densification.", + "- We present ablations and experimental evaluations on different, real-world benchmarks, confirming quantitative and qualitative improvements." + ], + "bbox": [ + 225, + 484, + 784, + 587 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "1.1 Related works", + "text_level": 1, + "bbox": [ + 217, + 607, + 379, + 619 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Since it was presented in [9], 3DGS has been used in a remarkably wide set of downstream applications, including Simultaneous Localization and Mapping [8, 16,24,30], text-to-3D generation [2,21,28], photo-realistic human avatars [11,13, 19,32], dynamic scene modeling [15,22,25] and more [5,23,27]. However, only a handful of works like ours have focused on advancing 3DGS itself, by improving its quality or overcoming some of its limitations.", + "bbox": [ + 217, + 628, + 785, + 718 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In $\\mathrm{GS} + +$ [7], Huang et al. present an improved approximation of the 3D-to-2D splatting operation at the core of 3DGS, which achieves better accuracy near image edges and solves some common visual artifacts. Spec-Gaussian [26] and Scaffold-gs [14] focus on improving view-dependent appearance modeling: the former by replacing spherical harmonics with an anisotropic spherical Gaussian appearance field; the latter by making all 3D Gaussian parameters, including whether specific primitives should be rendered or not, dependent on view direction through a small MLP. Mip-Splatting [29] tackles the strong artifacts that", + "bbox": [ + 217, + 719, + 785, + 839 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "Revising Densification in Gaussian Splitting", + "bbox": [ + 434, + 114, + 730, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "appear in 3DGS models when they are rendered at widely different resolutions (or viewing distances) compared to the images they were trained on. To do this, Yu et al. propose to incorporate a 3D filter to constrain the size of the 3D primitives depending on their maximal sampling rate on the training views, and a 2D Mip filter to mitigate aliasing issues. All these works adopt the original ADC strategy proposed in [9], and can potentially benefit from our improved approach, as we show for Mip-Splatting in Sec. 4.", + "bbox": [ + 212, + 146, + 782, + 251 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Only few concurrent works have touched on densification, while putting most of their focus on other aspects of 3DGS. Lee et al. [12] propose a quantization-based approach to produce more compact 3DGS representations, which includes a continuous sparsification strategy that takes both primitive size and opacity into account. GaussianPro [3] directly tackles related densification limitations as we explore in our work, filling the gaps from SfM-based initialization. They propose a rather complex procedure based on the progressive propagation of primitives along estimated planes, using patch-matching and geometric consistency as guidance. In contrast to our method, [3] focuses on fixing the quality of planar regions, instead of holistically improving densification. We also note that a fair comparison with their method on the standard Mip-NeRF 360 benchmark is not feasible at the time of submission, as the authors did not publicly share the improved SfM point cloud used in their experiments (see §5.2 of [3]).", + "bbox": [ + 212, + 252, + 784, + 448 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2 Preliminaries: Gaussian Splitting", + "text_level": 1, + "bbox": [ + 215, + 469, + 576, + 488 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Gaussian Splatting [9] revisits ideas from EWA splatting [33] and proposes to fit a 3D scene as a collection of 3D Gaussian primitives $\\Gamma \\coloneqq \\{\\gamma_1,\\dots ,\\gamma_K\\}$ that can be rendered by leveraging volume splatting.", + "bbox": [ + 212, + 500, + 782, + 546 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Gaussian primitive. A Gaussian primitive $\\gamma_{k} := (\\pmb{\\mu}_{k}, \\pmb{\\Sigma}_{k}, \\alpha_{k}, \\pmb{f}_{k})$ geometrically resembles a 3D Gaussian kernel", + "bbox": [ + 214, + 551, + 784, + 582 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {G} _ {k} (\\boldsymbol {x}) := \\exp \\left(- \\frac {1}{2} \\left(\\boldsymbol {x} - \\boldsymbol {\\mu} _ {k}\\right) ^ {\\top} \\Sigma_ {k} ^ {- 1} \\left(\\boldsymbol {x} - \\boldsymbol {\\mu} _ {k}\\right)\\right) \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 346, + 588, + 784, + 621 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "centered in $\\pmb{\\mu}_k\\in \\mathbb{R}^3$ and having $\\Sigma_{k}$ as its $3\\times 3$ covariance matrix. Each primitive additionally entails an opacity factor $\\alpha_{k}\\in [0,1]$ and a feature vector $\\pmb {f}_k\\in \\mathbb{R}^d$ (e.g. RGB color or spherical harmonics coefficients).", + "bbox": [ + 214, + 630, + 782, + 678 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Splatting. This is the operation of projecting a Gaussian primitive $\\gamma_{k}$ to a camera pixel space via its world-to-image transformation $\\pi : \\mathbb{R}^3 \\to \\mathbb{R}^2$ , which we refer directly to as the camera for simplicity. The projection $\\pi$ is approximated to the first order at the primitive's center $\\pmb{\\mu}_{k}$ so that the projected primitive is geometrically equivalent to a 2D Gaussian kernel $\\mathcal{G}_k^\\pi$ with mean $\\pi (\\pmb{\\mu}_k)\\in \\mathbb{R}^2$ and 2D covariance $\\mathrm{J}_k^\\pi \\Sigma_k\\mathrm{J}_{\\mathbf{k}}^\\pi {}^\\top$ with $\\mathrm{J}_k^\\pi$ being the Jacobian of $\\pi$ evaluated at $\\pmb{\\mu}_k$ .", + "bbox": [ + 214, + 681, + 784, + 775 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Rendering. To render the primitives $\\varGamma$ representing a scene from camera $\\pi$ , we require a decoder $\\varPhi$ to be specified, which provides the feature we want to render as $\\varPhi(\\gamma_k,\\pmb{u})\\in\\mathbb{R}^m$ for each Gaussian primitive $\\gamma_{k}$ and pixel $\\pmb{u}$ . Moreover, we assume Gaussian primitives $\\varGamma$ to be ordered with respect to their center's depth,", + "bbox": [ + 214, + 779, + 785, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "Rota Bulò et al.", + "bbox": [ + 271, + 114, + 380, + 126 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "when seen from the camera's reference frame. Then, the rendering equation takes the following form (with $\\varGamma$ being omitted from the notation)", + "bbox": [ + 214, + 146, + 782, + 176 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {R} [ \\pi , \\Phi ] (\\boldsymbol {u}) := \\sum_ {k = 1} ^ {K} \\Phi \\left(\\gamma_ {k}, \\boldsymbol {u}\\right) \\omega_ {k} ^ {\\pi} (\\boldsymbol {u}), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 385, + 191, + 785, + 231 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\omega_{k}^{\\pi}(\\pmb{u})$ are alpha-compositing coefficients given by", + "bbox": [ + 214, + 241, + 620, + 258 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\omega_ {k} ^ {\\pi} (\\boldsymbol {u}) := \\alpha_ {k} \\mathcal {G} _ {k} ^ {\\pi} (\\boldsymbol {u}) \\prod_ {j = 1} ^ {k - 1} \\left(1 - \\alpha_ {j} \\mathcal {G} _ {j} ^ {\\pi} (\\boldsymbol {u})\\right). \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 366, + 271, + 785, + 311 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "If we assume the feature vectors $\\pmb{f}_k$ to be spherical harmonics coefficients encoding an RGB function on the sphere, we can regard $\\varPhi_{\\mathrm{RGB}}(\\pmb{u})$ as the decoded RGB color for the given view direction associated to pixel $\\pmb{u}$ . If we use $\\varPhi_{\\mathrm{RGB}}$ as the decoder in the rendering equation, we obtain a rendered color image $C_{\\pi}(\\pmb{u}) := \\mathcal{R}[\\pi, \\varPhi_{\\mathrm{RGB}}](\\pmb{u})$ for each camera $\\pi$ . Similarly, one can pick different $\\varPhi$ 's to enable the rendering of depth, normals, or other quantities of interest as we will show later.", + "bbox": [ + 214, + 323, + 782, + 426 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Mip-splatting. In [29], the authors introduce a variation of standard Gaussian splatting that focuses on solving aliasing issues. We refer the reader to the original paper for details, but the idea is to track the maximum sampling rate for each Gaussian primitive and use it to reduce aliasing effects by attenuating the Gaussian primitives' opacity.", + "bbox": [ + 214, + 434, + 782, + 510 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3 Revising Densification", + "text_level": 1, + "bbox": [ + 214, + 534, + 465, + 551 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We first review the Adaptive Density Control module proposed in the original Gaussian splatting work [9], highlight some of its limitations, and then introduce our novel and improved densification procedure.", + "bbox": [ + 214, + 566, + 782, + 612 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1 Adaptive Density Control and its limitations", + "text_level": 1, + "bbox": [ + 214, + 635, + 630, + 650 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3DGS [9] and follow-up extensions (e.g. Mip-splatting [29]) rely on the Adaptive Density Control (ADC) module to grow or prune Gaussian primitives. This module is run according to a predetermined schedule and densification decisions are based on gradient statistics collected across the ADC runs. Specifically, for each Gaussian primitive $\\gamma_{k}$ the positional gradient magnitude $\\left\\| \\frac{\\partial L_{\\pi}}{\\partial \\mu_k} \\right\\|$ is tracked and averaged over all rendered views $\\pi \\in \\Pi$ within the collection period, where $L_{\\pi}$ denotes the loss that is optimized for camera $\\pi$ . The resulting quantity is denoted by $\\tau_{k}$ .", + "bbox": [ + 214, + 660, + 782, + 787 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Growing. ADC grows new Gaussian primitives via a clone or a split operation. A primitive $\\gamma_{k}$ will be considered for a growing operation only if $\\tau_{k}$ exceeds a user-defined threshold. The decision about which operation to apply depends", + "bbox": [ + 214, + 794, + 782, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "Revising Densification in Gaussian Splitting", + "bbox": [ + 434, + 114, + 732, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "on the size of the primitive measured in terms of the largest eigenvalue of the covariance matrix $\\Sigma_{k}$ . Specifically, primitives larger than a threshold are split, otherwise cloned. When a primitive $\\gamma_{k}$ is split, two new primitives are generated with their position being sampled from $\\mathcal{G}_k$ and their covariance being a scaled down version of $\\Sigma_{k}$ , while preserving the same opacity and feature vector. When a clone operation takes place, a simple clone of $\\gamma_{k}$ is instantiated.", + "bbox": [ + 212, + 146, + 787, + 238 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Pruning. ADC prunes a Gaussian primitive $\\gamma_{k}$ if its opacity $\\alpha_{k}$ is below a user-defined threshold, typically 0.005. To ensure that an unused primitive is eventually pruned, a hard-reset of the opacity to a minimum value (usually 0.01) is enforced according to a predefined schedule.", + "bbox": [ + 212, + 243, + 787, + 304 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Limitations. Deciding which Gaussian primitives to split/clone based on the magnitude of the positional gradient suffers from a number of limitations:", + "bbox": [ + 214, + 309, + 785, + 340 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Determining a threshold for a gradient magnitude is not intuitive and very sensitive to modifications to the model, losses and hyperparameters,", + "- There are cases of scene underfitting also when the value of $\\tau_{k}$ is below the threshold that triggers densification (see Fig. 1).", + "- It is not possible to directly control the number of Gaussian primitives that are generated for a given scene, resulting in possible out-of-memory errors if their number grows abnormally." + ], + "bbox": [ + 225, + 347, + 785, + 450 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In addition, we found that the ADC's logic of growing primitives suffers from a bias that weights more the contribution of freshly cloned primitives. More details will follow in Sec. 3.3", + "bbox": [ + 212, + 457, + 787, + 503 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.2 Error-based densification", + "text_level": 1, + "bbox": [ + 214, + 523, + 470, + 537 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Assume we have an image with an area characterized by a high-frequency pattern and covered by few large splatted Gaussian primitives (e.g. the grass in Fig. 1). Under this scenario, an infinitesimal change in the 3D location $\\pmb{\\mu}_k$ of one of the corresponding Gaussian primitives $\\gamma_k$ will leave the error almost unchanged and, hence, the collected magnitude of the positional gradient $\\tau_k$ remains close to zero. In fact, $\\tau_k$ is sensitive to error-changes, but is blind to the absolute value of the error. This becomes a problem, for we expect to increase the number of Gaussian primitives in areas exhibiting a larger error.", + "bbox": [ + 212, + 545, + 787, + 666 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Given the above considerations, we propose to steer the densification decisions directly based on an auxiliary per-pixel error function $\\mathcal{E}_{\\pi}$ (e.g. Structural Similarity) that we measure when rendering on a camera $\\pi$ with available ground-truth. One problem to address is how to turn per-pixel errors into per-Gaussian-primitive errors in light of the fact that each pixel error entangles the contribution of multiple Gaussian primitives. Our solution consists of first re-distributing the per-pixel errors $\\mathcal{E}_{\\pi}(\\pmb{u})$ to each Gaussian primitive $\\gamma_{k}$ proportionally to their contribution to the rendered pixel color, i.e. proportionally to $w_{k}^{\\pi}(\\pmb{u})$ . This yields the following error for each primitive $\\gamma_{k}$ and camera $\\pi$ :", + "bbox": [ + 212, + 667, + 787, + 803 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nE _ {k} ^ {\\pi} := \\sum_ {\\boldsymbol {u} \\in \\operatorname {P i x}} \\mathcal {E} _ {\\pi} (\\boldsymbol {u}) w _ {k} ^ {\\pi} (\\boldsymbol {u}), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 411, + 811, + 785, + 843 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "Rota Bulò et al.", + "bbox": [ + 271, + 114, + 380, + 127 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where the sum runs over the image pixels. Then, for each primitive $\\gamma_{k}$ we track the maximum value of the error $E_{k}^{\\pi}$ across all views $\\pi \\in \\Pi$ seen between two runs of the ADC module, i.e.", + "bbox": [ + 212, + 146, + 782, + 191 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\nE _ {k} := \\max _ {\\pi \\in \\Pi} E _ {k} ^ {\\pi}. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 449, + 202, + 785, + 223 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "This is the score that we use to prioritize the growing of Gaussian primitives. As opposed to $\\tau_{k}$ , it is easier to set a threshold for our new densification score, for it is typically expressed in terms of a known error metric.", + "bbox": [ + 212, + 229, + 782, + 276 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Implementation details. In order to compute $E_{k}^{\\pi}$ we assign an additional scalar $e_{k}$ to each Gaussian primitive $\\gamma_{k}$ , and enable the possibility of rendering it via the decoder $\\varPhi_{\\mathrm{ERR}}(\\gamma_k,\\pmb {u})\\coloneqq e_k$ . Then, we add the following auxiliary loss to the standard Gaussian splatting training objective:", + "bbox": [ + 212, + 282, + 784, + 342 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\nL _ {\\pi} ^ {\\text {a u x}} := \\sum_ {\\boldsymbol {u} \\in \\operatorname {P i x}} \\mathcal {N} [ \\mathcal {E} _ {\\pi} (\\boldsymbol {u}) ] \\underbrace {\\mathcal {R} [ \\pi , \\Phi_ {\\mathrm {E R R}} ] (\\boldsymbol {u})} _ {= \\sum_ {k = 1} ^ {K} e _ {k} \\omega_ {k} ^ {\\pi} (\\boldsymbol {u})}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 369, + 352, + 785, + 391 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "which is basically the dot product of the per-pixel error with gradient detached (denoted by $\\mathcal{X}$ ) and the rendering of the newly-added scalar. We initialize $e_k$ to 0 for each Gaussian primitive $\\gamma_k$ and never update it during training. In this way, $L_{\\pi}^{\\mathrm{aux}} = 0$ and all Gaussian primitives' parameters, excepting $e_k$ , are left invariant by this loss. The gradient with respect to $e_k$ instead yields", + "bbox": [ + 212, + 398, + 782, + 474 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\frac {\\partial L _ {\\pi} ^ {\\mathrm {a u x}}}{\\partial e _ {k}} = \\sum_ {\\boldsymbol {u} \\in \\operatorname {P i x}} \\mathcal {E} _ {\\pi} (\\boldsymbol {u}) \\omega_ {k} ^ {\\pi} (\\boldsymbol {u}) = E _ {k} ^ {\\pi}, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 382, + 481, + 785, + 517 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "which is the per-Gaussian-primitive error for camera $\\pi$ we wanted to compute.", + "bbox": [ + 212, + 525, + 779, + 541 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.3 Opacity correction after cloning", + "text_level": 1, + "bbox": [ + 212, + 560, + 526, + 575 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In the original ADC module, when a Gaussian primitive is split or cloned, the opacity value is preserved. This choice introduces a bias in the case of the clone operation by implicitly increasing the impact of the densified primitive on the final rendered color. To see why this is the case we can follow the example in Fig. 2, where we consider what happens if we render a splatted Gaussian in its center pixel assuming an opacity value $\\alpha$ . Before a cloning operation happens, the rendered color depends on primitives that come next in the ordering with weight $1 - \\alpha$ . But after we clone, due to the alpha-compositing logic, we have that primitives that come next weight $(1 - \\alpha)^{2}$ , which is lower than $1 - \\alpha$ for all opacity values in $(0,1)$ . Accordingly, by applying the standard logic of preserving the opacity after cloning we have a bias to weight more the cloned primitives. The solution we suggest consists in reducing the opacity of the primitives after cloning so that the bias is removed. The new opacity value $\\hat{\\alpha}$ can be found by solving the equation $(1 - \\alpha) = (1 - \\hat{\\alpha})^{2}$ , which yields $\\hat{\\alpha} := 1 - \\sqrt{1 - \\alpha}$ .", + "bbox": [ + 212, + 583, + 785, + 794 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "If we depart from the simplified setting of considering only the center pixel and rather consider all pixels, it is unfortunately not possible to remove completely the bias. Nonetheless, the correction factor we introduce reduces the bias", + "bbox": [ + 212, + 795, + 785, + 839 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "Revising Densification in Gaussian Splitting", + "bbox": [ + 433, + 114, + 732, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 114, + 784, + 126 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/552cacd468ecb90ada953eabc7d71a84878e5040e495d5cfeedd3a3e18f98f3c.jpg", + "image_caption": [ + "Fig. 2: Consider rendering a single splatted Gaussian in its center pixel with opacity $\\alpha$ before and after cloning. Before we clone, the rendered color depends with weight $1 - \\alpha$ on what comes next. After we clone, since we preserve the opacity, the rendered color depends with weight $(1 - \\alpha)^{2}$ on what comes next. Since $(1 - \\alpha) \\geq (1 - \\alpha)^{2}$ we have a bias towards weighting more Gaussian primitives that get cloned. The proposed correction changes the opacity post clone to $\\hat{\\alpha}$ so that the bias is removed." + ], + "image_footnote": [], + "bbox": [ + 215, + 143, + 333, + 219 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/0e7914399d185eca8f4c05ee299107cafa9d7ef7b0741be5da1c520276602092.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 367, + 142, + 531, + 217 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/6f166a9d41021277a371fa7e88e62f4a89a03410b2de69a09502ca29d5a2c5d6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 555, + 142, + 777, + 215 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "for all pixels compared to keeping the opacity of the cloned primitive. Indeed, the following relation holds for all $\\alpha_{k}\\in (0,1)$ and all pixels $\\pmb{u}$ :", + "bbox": [ + 212, + 335, + 782, + 367 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n1 - \\alpha_ {k} \\mathcal {G} _ {k} ^ {\\pi} (\\boldsymbol {u}) \\geq \\left(1 - \\hat {\\alpha} _ {k} \\mathcal {G} _ {k} ^ {\\pi} (\\boldsymbol {u})\\right) ^ {2} > \\left(1 - \\alpha_ {k} \\mathcal {G} _ {k} ^ {\\pi} (\\boldsymbol {u})\\right) ^ {2}, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 316, + 372, + 785, + 390 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $\\hat{\\alpha}_k \\coloneqq 1 - \\sqrt{1 - \\alpha_k}$ is our corrected opacity. The proof of the relation follows by noting that $\\hat{\\alpha}_k$ can be rewritten as $\\frac{\\alpha_k}{1 + \\sqrt{1 - \\alpha_k}}$ which is strictly smaller than $\\alpha_k$ for $\\alpha_k \\in (0,1)$ .", + "bbox": [ + 212, + 396, + 782, + 444 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Finally, the correction of the opacity as shown above is derived assuming we clone a Gaussian primitive, but does not strictly match the case of a split operation, for when we split we move the two offspring randomly away from the previous center and we change the covariance scale. For this reason, we stick to the standard rule of preserving the opacity of a primitive we split.", + "bbox": [ + 212, + 444, + 784, + 521 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "3.4 Primitives growth control", + "text_level": 1, + "bbox": [ + 214, + 537, + 473, + 553 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The ADC module grows a Gaussian primitive if $\\tau_{k}$ is larger than a threshold. This mechanism can lead to unpredictable growth of the number of primitives, eventually resulting in out-of-memory issues. To avoid this problem, we introduce a global limit to the maximum number of Gaussian primitives, and a mechanism to control the maximum number of primitives that can be created each time densification is run. Among the many possible options, we explore a logic that limits new primitive offspring to a fixed fraction of the primitives that already exist. In case the number of primitives that are entitled to be densified exceeds the available budget, we retain only the ones that exhibit the highest densification score. An example of this process is shown in Fig. 3, compared to the one from 3DGS: For Ours, the number of primitives grows smoothly until it reaches the allotted maximum, without the discontinuities induced by opacity reset (see Sec. 3.5). The way we control the number of primitives is not limited to our error-based densification logic, but can be applied equally to the original gradient-based one.", + "bbox": [ + 212, + 559, + 787, + 772 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "3.5 Alternative to opacity reset", + "text_level": 1, + "bbox": [ + 214, + 787, + 491, + 804 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The strategy introduced in [9] to favour the sparsification of Gaussian primitives consists in periodical hard resets of the opacity for all primitives to a low", + "bbox": [ + 212, + 809, + 785, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "Rota Bulò et al.", + "bbox": [ + 271, + 114, + 380, + 126 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/3e11b4fbe66ba682b8f1acde909f2031ca4f58821e9935e0cef6f6452daf91cc.jpg", + "image_caption": [ + "Number of Gaussians by training iteration", + "Fig. 3: Evolution of the number of Gaussians in 3DGS, and in our method with upper limit set to the number reached by 3DGS (on the garden scene from the Mip-NeRF 360 dataset). Note that, while 3DGS' ADC process stops after $15\\mathrm{k}$ iterations, ours remains active for $27\\mathrm{k}$ . This is not immediately visible from the plot, since pruned primitives are immediately replaced by newly spawned ones, keeping the overall number stable once the maximum is reached." + ], + "image_footnote": [], + "bbox": [ + 299, + 162, + 702, + 305 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "value, so that primitives whose opacity is not increased again by the optimization will eventually be pruned. This introduces a small shock in the training trajectory, which is suboptimal for the sake of having stable and predictable training dynamics. Moreover, resetting the opacity is particularly harmful for our error-based densification method, for it will lead to misleading error statistics right after the hard-reset, potentially triggering wrong densification decisions. For this reason, we propose a different logic to favour primitives pruning in a smoother way. Specifically, we decrease the opacity of each primitive by a fixed amount (we use 0.001) after each densification run, so that the opacity will gradually move towards the pruning range. In this way, we avoid sudden changes in the densification metric, while preserving the desired sparsification properties.", + "bbox": [ + 212, + 434, + 787, + 601 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "One downside of the new opacity regularization logic is that the constant push towards lowering the opacity of the primitives implicitly invites the model to make more use of the background where possible. This is also harmful, for it could generate more holes in the scene that will be visible from novel views. To counteract this dynamics, we also regularize the residual probabilities of the alpha-compositing (a.k.a. residual transmittance) to be zero for every pixel, by simply minimizing their average value, weighted by a hyperparameter (here 0.1).", + "bbox": [ + 212, + 604, + 787, + 709 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4 Experimental Evaluation", + "text_level": 1, + "bbox": [ + 214, + 748, + 493, + 763 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "In the following we show how our improved ADC mechanism can equally be applied both to standard 3DGS [9] and its Mip-Splatting extension [29], providing benefits to both.", + "bbox": [ + 212, + 794, + 785, + 839 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "Revising Densification in Gaussian Splitting", + "bbox": [ + 434, + 114, + 730, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.1 Datasets and metrics", + "text_level": 1, + "bbox": [ + 215, + 146, + 434, + 159 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "We follow the experimental setup from the 3DGS [9] paper, focusing on the real-world scenes from the Mip-NeRF 360 [1], Tanks and Temples [10] and Deep Blending [6] datasets. Mip-NeRF 360 comprises nine scenes (5 outdoor, 4 indoor) captured in a circular pattern which focuses on a central area of a few meters, with a potentially unbounded background. For Tanks and Temples, we focus on the \"Truck\" and \"Train\" scenes, while for Deep Blending we focus on the \"Dr Johnson\" and \"Playroom\" scenes, using the images and SfM reconstructions shared by the Gaussian Splatting authors. In each experiment we set aside each 8th image as a validation set, and report peak signal-to-noise ratio (PSNR), structural similarity (SSIM) and the perceptual metric from [31] (LPIPS).", + "bbox": [ + 212, + 174, + 787, + 325 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Additional experiments on NeRF-Synthetic [17] are provided in the supplementary material.", + "bbox": [ + 212, + 325, + 785, + 354 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.2 Experimental setup", + "text_level": 1, + "bbox": [ + 215, + 381, + 423, + 396 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "We evaluate based on our re-implementations of 3DGS, which allows us to easily switch between standard 3DGS, Mip-Splatting, the original ADC of [9], our contributions or any combination thereof. We reproduce the training settings proposed in [9,29] and the respective public code-bases $^{12}$ , including number of training iterations, batch size, input resolution, learning rates etc. When training with our contributions, we grow Gaussians with $E_{k} > 0.1$ , adding up to $5\\%$ of the current number of primitives at each densification step. Differently from 3DGS, we keep our ADC process active for 27k iterations (i.e. $90\\%$ of the training process), instead of stopping it after 15k. Other relevant hyper-parameters are left to the default values used in 3DGS, and shared across all datasets and scenes. In all our experiments, we set the maximum primitives budget to the number of primitives (or its median, for experiments with multiple runs) generated by the corresponding baseline, in order to obtain perfectly comparable models.", + "bbox": [ + 212, + 409, + 787, + 604 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "A note about LPIPS. Investigating the 3DGS and Mip-Splating baselines, we discovered a discrepancy in the way LPIPS is calculated in both public code-bases, which resulted in under-estimated values being reported in the original papers. This was confirmed in private correspondence with the authors. In order to simplify comparisons with future works that don't rely on these code-bases, and might be unaware of this issue, we report correct LPIPS values here.", + "bbox": [ + 212, + 611, + 787, + 702 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "A note about evaluations on MipNeRF360. The standard protocol for the dataset trains and evaluates the models at the same image scale, with indoor scenes down-sampled by a factor of two and outdoor scenes by four. The authors of the dataset provide down-scaled versions of the images. However, several recent works like [29] do not use the original images, which suffer from compression artifacts, but down-sample the images from the higher resolution counterparts.", + "bbox": [ + 212, + 709, + 787, + 800 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "Rota Bulò et al.", + "bbox": [ + 271, + 114, + 380, + 126 + ], + "page_idx": 9 + }, + { + "type": "page_footnote", + "text": "1 https://github.com/graphdeco-inria/gaussian-splatting", + "bbox": [ + 217, + 810, + 645, + 825 + ], + "page_idx": 9 + }, + { + "type": "page_footnote", + "text": "2 https://github.com/autonomousvision/mip-splatting", + "bbox": [ + 220, + 825, + 614, + 839 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/ef286a968d3437a4116d7053db1578b08a3f1c0d01c227aa786467b3fad2272e.jpg", + "table_caption": [ + "Table 1: Results on the Mip-NeRF 360 dataset. Top section of the table: results from the Gaussian Splitting paper; bottom section: results from our re-implementation averaged over 5 runs." + ], + "table_footnote": [], + "table_body": "
MethodPSNR ↑SSIM ↑LPIPS ↓
MeanStdevMeanStdevMeanStdev
Plenoxels [4]23.08-0.626-0.436-
INGP [18]25.59-0.699-0.331-
Mip-NeRF 360 [1]27.69-0.792-0.237-
3DGS [9]27.450.0810.8170.0010.2500.001
Mip-Splatting [29]27.510.0740.8180.0010.2510.001
Ours, 3DGS27.610.0670.8220.0010.2230.001
Ours, Mip-Splatting27.700.0650.8230.0010.2230.001
", + "bbox": [ + 250, + 198, + 753, + 347 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/5cabc75502c9ce24c7f2468c1e46491cafd25efefeca28c2d292bd70836eb1aa.jpg", + "table_caption": [ + "Table 2: Results on the Tanks and Temples dataset. Top section of the table: results from the Gaussian Splitting paper; bottom section: results from our re-implementation averaged over 5 runs." + ], + "table_footnote": [], + "table_body": "
MethodPSNR ↑SSIM ↑LPIPS ↓
MeanStdevMeanStdevMeanStdev
Plenoxels [4]21.08-0.719-0.379-
INGP [18]21.92-0.745-0.305-
Mip-NeRF 360 [1]22.22-0.759-0.257-
3DGS [9]23.440.0890.8450.0020.2120.001
Mip-Splatting [29]23.650.0660.8490.0010.2110.001
Ours, 3DGS23.930.0700.8530.0010.1870.001
Ours, Mip-Splatting24.100.1090.8570.0020.1830.001
", + "bbox": [ + 250, + 414, + 753, + 563 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "As a result, the scores they report are significantly higher. We use the downscaled images from the dataset for training/evaluation to stay consistent with older NeRF works and report scores of competitors under the same setting.", + "bbox": [ + 215, + 592, + 784, + 637 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "4.3 Main results", + "text_level": 1, + "bbox": [ + 217, + 661, + 364, + 674 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "In a first set of experiments, we evaluate the effectiveness of our improved ADC strategy (Ours) when applied to 3DGS and Mip-Splatting. Results, collected over 5 training runs to average out the randomness induced by stochastic primitive splitting, are reported in Tab. 1, 2 and 3. For the sake of completeness, we also include scores obtained with three NeRF baselines, i.e. Plenoxels [4], InstantNGP (INGP) [18] and Mip-NeRF 360 [1], as originally reported in [9]. Our approach consistently outperforms the corresponding baselines (i.e. Ours, 3DGS vs. 3DGS; Ours, Mip-Splatting vs. Mip-Splatting), particularly on SSIM and LPIPS. This is in line with what we discussed in Sec. 3.1 and 3.2: standard ADC often leads to localized under-fitting, as it fails to split large gaussians", + "bbox": [ + 215, + 688, + 785, + 839 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "Revising Densification in Gaussian Splitting", + "bbox": [ + 434, + 114, + 730, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 116, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/858307533f28b0a273f85513a87d74920db2378a348ac62ef90c84b155f0b6e0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 217, + 199, + 405, + 281 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/ce614e38d782800e44c355e081982514ac07258efeba0e213ab19abea58ded83.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 245, + 282, + 377, + 318 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/21409c2181685c0f81b46d78b8c09c6d84087863a5f7ab6f4e3e4cb9656fc2e4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 406, + 200, + 594, + 282 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/3770316d0329bcaece8805ac389529ae426c2c9e8759912a6de143cad65e0a9f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 434, + 282, + 566, + 318 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/2153b1b9400cd93249e1f2dc103206d8290523984cff59ac5d67df104a227b3b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 596, + 200, + 785, + 282 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/b6efd124893b2007e6a08a83a91e6654d01db4332f0b786f514bf6557ee3301e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 624, + 282, + 756, + 316 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/bb3add9c6dc83d1e7bacc8eedf7e39f62643ca5b9876469c14f82b301c5c7ca3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 215, + 327, + 405, + 410 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/8c4abd8c09beefacf463fa8991016ff44b63dcf084a31bbe9cb25ce081686548.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 245, + 410, + 377, + 445 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/d02658d9ad8b4c5a431e5239d482520cb2647dc02dac41beaa164ea3d5a769c1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 406, + 327, + 594, + 410 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/815814d264d5563d9098d5391722c69d4c73f601d77b044d731d19c9898c9c0c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 434, + 410, + 566, + 445 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/0ceaccbae16322cd5b24a0d91ff0fcaee5bf3ad40088ee1e2d986a284161a0a0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 596, + 327, + 784, + 410 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/195593b97f31df17f4d393da409b166956ce968d495b03349727ab9fbcbf7689.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 624, + 410, + 756, + 445 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/b25904c7913bab6dbf6ae598b55fa623e56ef1bcfa3feeec360cc593aee43395.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 215, + 450, + 403, + 547 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/d8062f087f92f6162452b56cda07b84884c721551fb7050d8963e18cc5a7833b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 245, + 547, + 377, + 583 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/e73b6a6de4f855f05c46a39c2048cb36c6152d5cfe6ea1c509c7821c5459e555.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 406, + 450, + 594, + 547 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/d303d692b14d6a1edca90b44b324c70ebd5afe5a60a338dd4a7ec270e6d4f1a3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 434, + 547, + 566, + 583 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/eb17d26e4871dd0b1e7aaaa420270651d57fda735efcabd4c5a3400fb62419bc.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 596, + 450, + 784, + 547 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/751134d20b393733d0ba796c5734f3579fd460728cccac4b17eae3148eefa104.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 624, + 547, + 756, + 583 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/35c59216cbc1fb86523647ce1f2dcceafa21928fd7721278b8429398c8b1b56c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 217, + 593, + 405, + 690 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/3893412dd7c444e7017fbf694484fef839a9ce5c1c2886fe92484ebd1ca7121e.jpg", + "image_caption": [ + "Fig. 4: Qualitative results on the Mip-NeRF 360 and Tanks and Temples validation sets. Both 3DGS and Ours use the same number of primitives. Best viewed on screen at high magnification." + ], + "image_footnote": [], + "bbox": [ + 245, + 690, + 377, + 726 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/78c0e9458d6af011f9f984c0ff83cdcc93f633ae5697fb0ed67daf68ff3ff713.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 406, + 593, + 594, + 690 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/b055c6bbd040714ad14f987e807c1607a5fee66d305d2e0c52ff6f4f37c3a47a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 434, + 690, + 566, + 726 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/60cf8f0109a37415d4361fc5d43ed399d5dc85ec98bd5c563bc0bf1470e675f6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 596, + 593, + 784, + 690 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/4ec7a106b865c0af9fae1265483ad5f8195d5b1c0335a173ced0447833bac32d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 624, + 690, + 756, + 726 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "Rota Bulò et al.", + "bbox": [ + 271, + 114, + 380, + 126 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/cac36875a267502624f34ee5e1219a0a82c0476b881b9a9db09dd3076107227e.jpg", + "table_caption": [ + "Table 3: Results on the Deep Blending dataset. Top section of the table: results from the Gaussian Splatting paper; bottom section: results from our re-implementation averaged over 5 runs." + ], + "table_footnote": [], + "table_body": "
MethodPSNR ↑SSIM ↑LPIPS ↓
MeanStdevMeanStdevMeanStdev
Plenoxels [4]23.06-0.795-0.510-
INGP [18]24.96-0.817-0.390-
Mip-NeRF 360 [1]29.40-0.901-0.245-
3DGS [9]29.540.0960.9020.0000.3110.001
Mip-Splatting [29]29.680.0680.9030.0000.3090.000
Ours, 3DGS29.500.1100.9040.0010.3050.001
Ours, Mip-Splatting29.640.1260.9050.0010.3030.001
", + "bbox": [ + 250, + 198, + 753, + 347 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/e214c4058508a85035e036ace0bf374b32e71eea82e4a89f589baf42612f8f48.jpg", + "table_caption": [ + "Table 4: Ablation experiments on the Mip-NeRF 360 dataset, adding individual contributions to 3DGS or removing them from Ours. OC: Opacity Correction, Sec. 3.3; GC: Growth Control, Sec. 3.4; OR: Opacity Regularization, Sec. 3.5." + ], + "table_footnote": [], + "table_body": "
Baseline3DGS+OROurs
+OC+GC+ORFull-OC-GC-OR
PSNR ↑27.4527.6527.3527.4827.6127.0427.5427.28
SSIM ↑0.8170.8220.8100.8180.8220.8120.8180.810
LPIPS ↓0.2500.2390.2560.2430.2230.2350.2230.234
", + "bbox": [ + 235, + 412, + 767, + 500 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "that cover highly-textured regions of the scene. This kind of errors is poorly reflected by PSNR, which measures the \"average fit\" over image pixels, but are promptly detected by perceptual metrics like LPIPS. On the Deep Blending dataset, we observe thinner gaps, PSNR actually showing a small regression w.r.t. the baselines, although with low confidence margins. We suspect this might be related to the fact that Deep Blending contains many flat, untextured surfaces (see Fig. 4), that are particularly challenging to reconstruct accurately with 3DGS-like methods, independently of the ADC strategy being adopted. In the supplementary material, we provide an additional discussion about why some metrics do not reflect the actual perceived quality of the generated images.", + "bbox": [ + 212, + 525, + 784, + 676 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Figure 4 contains a qualitative comparison between standard 3DGS and 3DGS augmented with our contributions (Ours). Areas with under-fitting artifacts are highlighted, showing how these are notably ameliorated by our approach. It is also worth noting that Ours effectively maintains the same quality as 3DGS in non-problematic areas, producing a more perceptually accurate reconstruction while using the same number of primitives (see Sec. 4.2).", + "bbox": [ + 212, + 676, + 784, + 767 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "4.4 Ablation experiments", + "text_level": 1, + "bbox": [ + 214, + 786, + 439, + 801 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "In Tab. 4 we ablate the effects of Opacity Correction (OC, Sec. 3.3), Growth Control (GC, Sec. 3.4) and Opacity Regularization (OR, Sec. 3.5) on the Mip-NeRF", + "bbox": [ + 212, + 809, + 782, + 839 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "Revising Densification in Gaussian Splitting", + "bbox": [ + 434, + 114, + 730, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 784, + 126 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "360 dataset. In particular, we evaluate 3DGS augmented with each of these components (left side of the table), and our method with the components replaced by the corresponding baseline mechanism in 3DGS' standard ADC (right side of the table). First, we observe that OC, GC and OR all contribute to our method, as the Full version of Ours achieves the overall best results on all metrics, and removing them consistently degrades performance. Interestingly, Opacity Correction seems to have the largest impact here, as it produces both the largest increase in the scores when added to 3DGS, and the largest decrease when removed from Ours. Finally, Growth Control has a negative impact on 3DGS when utilized in isolation, while only slightly degrading the results when removed from Ours. Note that this observation doesn't detract from GC's usefulness as a strategy to control and limit the capacity of the model. We hypothesize that GC's negative effect on 3DGS might be a consequence of the fact that the standard, gradient-based densification score is actually a poor choice for comparing gaussians in terms of how soon they should be split or cloned (remember that GC ranks Gaussians based on their score).", + "bbox": [ + 212, + 146, + 787, + 388 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "4.5 Limitations", + "text_level": 1, + "bbox": [ + 215, + 417, + 356, + 431 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "While our method appears to be quite effective at solving under-fitting issues, these can still be present in especially difficult scenes (e.g. treehill in the Mip-NeRF 360 dataset, both scenes from the Deep Blending dataset). Focusing on the problematic areas that our ADC approach handles successfully, we observe that, while perceptually more \"correct\", the reconstruction there can still be quite inaccurate when closely compared to the ground truth (see e.g. the flowers scene in Fig. 4). We suspect both these issues might be related to 3DGS' intrinsic limits in handling i) strong view-dependent effects; ii) appearance variations across images; and iii) errors induced by the linear approximation in the Splitting operation (see Sec. 2). An interesting future direction could be to combine our approach with works that address these issues, e.g. Spec-Gaussian [26] for (i) and GS++ [7] for (iii).", + "bbox": [ + 212, + 450, + 787, + 633 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Conclusion", + "text_level": 1, + "bbox": [ + 215, + 664, + 359, + 680 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In this paper, we addressed the limitations of the Adaptive Density Control (ADC) mechanism in 3D Gaussian Splatting (3DGS), a scene representation method for high-quality, photorealistic rendering. Our main contribution is a more principled, pixel-error driven formulation for density control in 3DGS. We propose how to leverage a novel decision criterion for densification based on per-pixel errors and introduce a mechanism to control the total number of primitives generated per scene. We also correct a bias in the current opacity handling in ADC during cloning. Our approach leads to consistent and systematic improvements over previous methods, particularly in perceptual metrics like LPIPS.", + "bbox": [ + 212, + 703, + 787, + 840 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "Rota Bulò et al.", + "bbox": [ + 271, + 114, + 380, + 126 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 143, + 321, + 159 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Barron, J.T., Mildenhall, B., Verbin, D., Srinivasan, P.P., Hedman, P.: Mipnerf 360: Unbounded anti-aliased neural radiance fields. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5470-5479 (2022)", + "2. Chen, Z., Wang, F., Liu, H.: Text-to-3d using gaussian splatting. arXiv preprint arXiv:2309.16585 (2023)", + "3. Cheng, K., Long, X., Yang, K., Yao, Y., Yin, W., Ma, Y., Wang, W., Chen, X.: Gaussianpro: 3d gaussian splatting with progressive propagation. arXiv preprint arXiv:2402.14650 (2024)", + "4. Fridovich-Keil, S., Yu, A., Tancik, M., Chen, Q., Recht, B., Kanazawa, A.: Plenoxels: Radiance fields without neural networks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5501-5510 (2022)", + "5. Guédon, A., Lepetit, V.: Sugar: Surface-aligned gaussian splatting for efficient 3d mesh reconstruction and high-quality mesh rendering. arXiv preprint arXiv:2311.12775 (2023)", + "6. Hedman, P., Philip, J., Price, T., Frahm, J.M., Drettakis, G., Brostow, G.: Deep blending for free-viewpoint image-based rendering. ACM Transactions on Graphics (ToG) 37(6), 1-15 (2018)", + "7. Huang, L., Bai, J., Guo, J., Guo, Y.: $\\mathrm{Gs} + +$ : Error analyzing and optimal gaussian splatting. arXiv preprint arXiv:2402.00752 (2024)", + "8. Keetha, N., Karhade, J., Jatavallabhula, K.M., Yang, G., Scherer, S., Ramanan, D., Luiten, J.: Splatam: Splat, track & map 3d gaussians for dense rgb-d slam. arXiv preprint arXiv:2312.02126 (2023)", + "9. Kerbl, B., Kopanas, G., Leimkuhler, T., Drettakis, G.: 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics 42(4) (July 2023), https://repo-sam.inria.fr/fungraph/3d-gaussian-splatting/", + "0. Knapitsch, A., Park, J., Zhou, Q.Y., Koltun, V.: Tanks and temples: Benchmarking large-scale scene reconstruction. ACM Transactions on Graphics (ToG) 36(4), 1-13 (2017)", + "1. Kocabas, M., Chang, J.H.R., Gabriel, J., Tuzel, O., Ranjan, A.: Hugs: Human gaussian splats. arXiv preprint arXiv:2311.17910 (2023)", + "2. Lee, J.C., Rho, D., Sun, X., Ko, J.H., Park, E.: Compact 3d gaussian representation for radiance field. arXiv preprint arXiv:2311.13681 (2023)", + "3. Lei, J., Wang, Y., Pavlakos, G., Liu, L., Daniilidis, K.: Gart: Gaussian articulated template models. arXiv preprint arXiv:2311.16099 (2023)", + "4. Lu, T., Yu, M., Xu, L., Xiangli, Y., Wang, L., Lin, D., Dai, B.: Scaffold-gs: Structured 3d gaussians for view-adaptive rendering. arXiv preprint arXiv:2312.00109 (2023)", + "5. Luiten, J., Kopanas, G., Leibe, B., Ramanan, D.: Dynamic 3d gaussians: Tracking by persistent dynamic view synthesis. arXiv preprint arXiv:2308.09713 (2023)", + "6. Matsuki, H., Murai, R., Kelly, P.H., Davison, A.J.: Gaussian splatting slam. arXiv preprint arXiv:2312.06741 (2023)", + "7. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. In: ECCV. p. 405-421 (2020)", + "8. Müller, T., Evans, A., Schied, C., Keller, A.: Instant neural graphics primitives with a multiresolution hash encoding. ACM Trans. Graph. 41(4), 102:1-102:15 (Jul 2022). https://doi.org/10.1145/3528223.3530127, https://doi.org/10.1145/3528223.3530127" + ], + "bbox": [ + 225, + 172, + 785, + 838 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "Revising Densification in Gaussian Splitting", + "bbox": [ + 434, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "19. Saito, S., Schwartz, G., Simon, T., Li, J., Nam, G.: Relightable gaussian codec avatars. arXiv preprint arXiv:2312.03704 (2023)", + "20. Snavely, N., Seitz, S.M., Szeliski, R.: Photo tourism: exploring photo collections in 3d. In: ACM siggraph 2006 papers, pp. 835-846 (2006)", + "21. Tang, J., Ren, J., Zhou, H., Liu, Z., Zeng, G.: Dreamgaussian: Generative gaussian splatting for efficient 3d content creation. arXiv preprint arXiv:2309.16653 (2023)", + "22. Wu, G., Yi, T., Fang, J., Xie, L., Zhang, X., Wei, W., Liu, W., Tian, Q., Wang, X.: 4d gaussian splatting for real-time dynamic scene rendering. arXiv preprint arXiv:2310.08528 (2023)", + "23. Xie, T., Zong, Z., Qiu, Y., Li, X., Feng, Y., Yang, Y., Jiang, C.: Physgaus-sian: Physics-integrated 3d gaussians for generative dynamics. arXiv preprint arXiv:2311.12198 (2023)", + "24. Yan, C., Qu, D., Wang, D., Xu, D., Wang, Z., Zhao, B., Li, X.: Gs-slam: Dense visual slam with 3d gaussian splatting. arXiv preprint arXiv:2311.11700 (2023)", + "25. Yang, Z., Yang, H., Pan, Z., Zhu, X., Zhang, L.: Real-time photorealistic dynamic scene representation and rendering with 4d gaussian splatting. arXiv preprint arXiv:2310.10642 (2023)", + "26. Yang, Z., Gao, X., Sun, Y., Huang, Y., Lyu, X., Zhou, W., Jiao, S., Qi, X., Jin, X.: Spec-gaussian: Anisotropic view-dependent appearance for 3d gaussian splatting. arXiv preprint arXiv:2402.15870 (2024)", + "27. Ye, M., Danelljan, M., Yu, F., Ke, L.: Gaussian grouping: Segment and edit anything in 3d scenes. arXiv preprint arXiv:2312.00732 (2023)", + "28. Yi, T., Fang, J., Wu, G., Xie, L., Zhang, X., Liu, W., Tian, Q., Wang, X.: Gaussian dreamer: Fast generation from text to 3d gaussian splatting with point cloud priors. arXiv preprint arXiv:2310.08529 (2023)", + "29. Yu, Z., Chen, A., Huang, B., Sattler, T., Geiger, A.: Mip-splatting: Alias-free 3d gaussian splatting. arXiv preprint arXiv:2311.16493 (2023)", + "30. Yugay, V., Li, Y., Gevers, T., Oswald, M.R.: Gaussian-slam: Photo-realistic dense slam with gaussian splatting. arXiv preprint arXiv:2312.10070 (2023)", + "31. Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual metric. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 586-595 (2018)", + "32. Zielonka, W., Bagautdinov, T., Saito, S., Zollhöfer, M., Thies, J., Romero, J.: Drivable 3d gaussian avatars. arXiv preprint arXiv:2311.08581 (2023)", + "33. Zwicker, M., Pfister, H., van Baar, J., Gross, M.: Ewa volume splatting. In: Proceedings Visualization, 2001. VIS '01. pp. 29-538 (2001). https://doi.org/10.1109/VISUAL.2001.964490" + ], + "bbox": [ + 215, + 147, + 784, + 657 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "Rota Bulò et al.", + "bbox": [ + 271, + 114, + 380, + 126 + ], + "page_idx": 15 + } +] \ No newline at end of file diff --git a/2024/Revising Densification in Gaussian Splatting/918310b7-b91b-411d-ade7-18df636ce501_model.json b/2024/Revising Densification in Gaussian Splatting/918310b7-b91b-411d-ade7-18df636ce501_model.json new file mode 100644 index 0000000000000000000000000000000000000000..8f0dcb2c76c8d4da774e08b7e14c6d2bcfc726a5 --- /dev/null +++ b/2024/Revising Densification in Gaussian Splatting/918310b7-b91b-411d-ade7-18df636ce501_model.json @@ -0,0 +1,2300 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.248, + 0.142, + 0.756, + 0.164 + ], + "angle": 0, + "content": "Revising Densification in Gaussian Splitting" + }, + { + "type": "text", + "bbox": [ + 0.266, + 0.19, + 0.737, + 0.206 + ], + "angle": 0, + "content": "Samuel Rota Bulò, Lorenzo Porzi, and Peter Kontschieder" + }, + { + "type": "text", + "bbox": [ + 0.412, + 0.218, + 0.586, + 0.232 + ], + "angle": 0, + "content": "Meta Reality Labs Zurich" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.257, + 0.408, + 0.355 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.408, + 0.257, + 0.597, + 0.355 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.597, + 0.257, + 0.786, + 0.355 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.366, + 0.788, + 0.437 + ], + "angle": 0, + "content": "Fig. 1: Densification is a critical component of 3D Gaussian Splatting (3DGS), and a common failure point. In this example (ground truth on the left) we show how 3DGS can fail (center) to add primitives to high-texture areas, like the grass in the bottom part of the pictures, producing large and blurry artifacts. Our approach (right) solves this issue by comprehensively revising densification in 3DGS." + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.456, + 0.741, + 0.636 + ], + "angle": 0, + "content": "Abstract. In this paper, we address the limitations of Adaptive Density Control (ADC) in 3D Gaussian Splatting (3DGS), a scene representation method achieving high-quality, photorealistic results for novel view synthesis. ADC has been introduced for automatic 3D point primitive management, controlling densification and pruning, however, with certain limitations in the densification logic. Our main contribution is a more principled, pixel-error driven formulation for density control in 3DGS, leveraging an auxiliary, per-pixel error function as the criterion for densification. We further introduce a mechanism to control the total number of primitives generated per scene and correct a bias in the current opacity handling strategy of ADC during cloning operations. Our approach leads to consistent quality improvements across a variety of benchmark scenes, without sacrificing the method's efficiency." + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.649, + 0.741, + 0.676 + ], + "angle": 0, + "content": "Keywords: Gaussian Splatting \\(\\cdot\\) 3D reconstruction \\(\\cdot\\) Novel View Synthesis" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.703, + 0.376, + 0.719 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.734, + 0.788, + 0.841 + ], + "angle": 0, + "content": "High-quality, photorealistic scene modelling from images has been an important research area in computer vision and graphics, with plentiful applications in AR/VR/MR, robotics, etc. In the last years, this field has gained a lot of attention due to advances in Neural 3D scene representations, particularly Neural Radiance Fields (NeRFs) [17]. NeRFs take a new approach to 3D scene representation and rendering, by leveraging a combination of deep learning and volumetric rendering techniques for generating photorealistic images from novel" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.381, + 0.127 + ], + "angle": 0, + "content": "Rota Bulò et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.298 + ], + "angle": 0, + "content": "viewpoints. By optimizing MLPs to map from spatial coordinates and viewing directions to density and colour fields, these models have demonstrated astonishing capabilities for capturing the complex interplay of light and geometry in a data-driven way. While highly efficient in terms of representation quality, the original NeRF representation relies on time-consuming sampling strategies and thus excludes applications with fast rendering requirements. With many advances in terms of the underlying representation, these models have been significantly optimized towards improved training time and scene representation fidelity. However, inference speed for high-resolution, novel view synthesis remains an ongoing limitation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.302, + 0.79, + 0.424 + ], + "angle": 0, + "content": "More recently, 3D Gaussian Splatting (3DGS) [9] has been proposed as an alternative and expressive scene representation, enabling both high-speed, high-fidelity training of models and high-resolution, GPU rasterization-friendly rendering of novel views. Their core representation is an optimized set of (anisotropic) 3D Gaussians, after being randomly distributed in 3D space, or systematically initialized at points obtained by Structure-from-Motion [20]. For obtaining a 2D image, all relevant 3D primitives are efficiently rendered via splatting-based rasterization with low-pass filtering." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.428, + 0.789, + 0.776 + ], + "angle": 0, + "content": "In 3DGS, each 3D primitive is parameterized as a 3D Gaussian distribution (i.e., with position and covariance), together with parameters controlling its opacity and describing its directional appearance (typically spherical harmonics). The parameter optimization procedure is guided by a multi-view, photometric loss, and is interleaved with Adaptive Density Control (ADC), a mechanism controlling density management for 3D points by means of introducing or deleting 3D primitives. ADC plays a critical role as it determines where to expand/shrink the scene representation budget for empty or over-reconstructed regions, respectively. Both growing and pruning operations are activated based on user-defined thresholds: Growing depends on the accumulated positional gradients of existing primitives and is, conditioned on the size of the Gaussians, executed by either splitting large primitives or by cloning smaller ones. Pruning is activated once the opacity falls below a provided threshold. While quite effective in practice, such density management strategies have several limitations. First, estimating a gradient magnitude-based threshold is rather non-intuitive and not robust to potential changes in the model, loss terms, etc. Second, there are cases where only few and large Gaussians are modeling high-frequency patterns like grass as shown in the middle of Fig. 1. Here, changes accumulated from positional gradients might remain very low and thus fail to trigger the densification mechanism, which in turn leads to substantial scene underfitting. Finally, ADC lacks explicit control of the maximum number of Gaussians generated per scene. This has important, practical implications as uncontrolled growth might easily lead to out-of-memory errors during training." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.789, + 0.842 + ], + "angle": 0, + "content": "In this work we address the shortcomings of Adaptive Density Control proposed in the original 3D Gaussian splatting method. Our core contribution is a more principled, pixel-error driven formulation for density control in 3DGS. We describe how 2D, per-pixel errors as e.g. derived from Structural Similarity" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.435, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Revising Densification in Gaussian Splitting" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.147, + 0.786, + 0.479 + ], + "angle": 0, + "content": "(or any other informative objective function) can be propagated back as errors to contributing Gaussian primitives. In our solution, we first break down the per-pixel errors according to each Gaussian's contribution, and in a camera-specific way. This allows us to track the maximum error per primitive for all views and across two subsequent ADC runs, yielding our novel, error-specific, and thus more intuitive decision criterion for densification. Our second contribution is correcting a bias introduced with the current form of opacity handling in ADC when conducting a primitive cloning operation. The original approach suggests to keep the same opacity for the cloned Gaussian, which however biases the alpha-compositing logic applied for rendering the pixel colors. Indeed, this procedure leads to an overall increase of opacity in the cloned region, preventing the model to correctly account for contributions of other primitives and thus negatively affecting the densification process. Our third contribution is a mechanism for controlling the total number of primitives generated per scene and the maximum amount of novel primitives introduced per densification run. With this functionality, we can avoid undesired out-of-memory errors and better tune the method's behaviour w.r.t. given hardware constraints. We extensively validate our contributions on standard benchmark datasets like Mip-NeRF 360 [1], Tanks and Temples [10], and Deep Blending [6]. Our experiments show consistent improvements over different baselines including 3DGS [9] and Mip-Splatting [29]. To summarize, our contributions are improving methodological shortcomings in 3DGS' Adaptive Density Control mechanism as follows:" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.485, + 0.785, + 0.529 + ], + "angle": 0, + "content": "- We propose a principled approach that enables the guidance of the densification process according to an auxiliary, per-pixel error function, rather than relying on positional gradients." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.53, + 0.785, + 0.559 + ], + "angle": 0, + "content": "- We correct an existing, systematic bias from the primitive growing procedure when cloning Gaussians, negatively impacting the overall densification." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.559, + 0.785, + 0.588 + ], + "angle": 0, + "content": "- We present ablations and experimental evaluations on different, real-world benchmarks, confirming quantitative and qualitative improvements." + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.485, + 0.785, + 0.588 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.608, + 0.38, + 0.621 + ], + "angle": 0, + "content": "1.1 Related works" + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.629, + 0.786, + 0.719 + ], + "angle": 0, + "content": "Since it was presented in [9], 3DGS has been used in a remarkably wide set of downstream applications, including Simultaneous Localization and Mapping [8, 16,24,30], text-to-3D generation [2,21,28], photo-realistic human avatars [11,13, 19,32], dynamic scene modeling [15,22,25] and more [5,23,27]. However, only a handful of works like ours have focused on advancing 3DGS itself, by improving its quality or overcoming some of its limitations." + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.72, + 0.786, + 0.84 + ], + "angle": 0, + "content": "In \\(\\mathrm{GS} + +\\) [7], Huang et al. present an improved approximation of the 3D-to-2D splatting operation at the core of 3DGS, which achieves better accuracy near image edges and solves some common visual artifacts. Spec-Gaussian [26] and Scaffold-gs [14] focus on improving view-dependent appearance modeling: the former by replacing spherical harmonics with an anisotropic spherical Gaussian appearance field; the latter by making all 3D Gaussian parameters, including whether specific primitives should be rendered or not, dependent on view direction through a small MLP. Mip-Splatting [29] tackles the strong artifacts that" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.381, + 0.127 + ], + "angle": 0, + "content": "Rota Bulò et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.252 + ], + "angle": 0, + "content": "appear in 3DGS models when they are rendered at widely different resolutions (or viewing distances) compared to the images they were trained on. To do this, Yu et al. propose to incorporate a 3D filter to constrain the size of the 3D primitives depending on their maximal sampling rate on the training views, and a 2D Mip filter to mitigate aliasing issues. All these works adopt the original ADC strategy proposed in [9], and can potentially benefit from our improved approach, as we show for Mip-Splatting in Sec. 4." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.253, + 0.785, + 0.449 + ], + "angle": 0, + "content": "Only few concurrent works have touched on densification, while putting most of their focus on other aspects of 3DGS. Lee et al. [12] propose a quantization-based approach to produce more compact 3DGS representations, which includes a continuous sparsification strategy that takes both primitive size and opacity into account. GaussianPro [3] directly tackles related densification limitations as we explore in our work, filling the gaps from SfM-based initialization. They propose a rather complex procedure based on the progressive propagation of primitives along estimated planes, using patch-matching and geometric consistency as guidance. In contrast to our method, [3] focuses on fixing the quality of planar regions, instead of holistically improving densification. We also note that a fair comparison with their method on the standard Mip-NeRF 360 benchmark is not feasible at the time of submission, as the authors did not publicly share the improved SfM point cloud used in their experiments (see §5.2 of [3])." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.47, + 0.578, + 0.489 + ], + "angle": 0, + "content": "2 Preliminaries: Gaussian Splitting" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.501, + 0.784, + 0.547 + ], + "angle": 0, + "content": "Gaussian Splatting [9] revisits ideas from EWA splatting [33] and proposes to fit a 3D scene as a collection of 3D Gaussian primitives \\(\\Gamma \\coloneqq \\{\\gamma_1,\\dots ,\\gamma_K\\}\\) that can be rendered by leveraging volume splatting." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.552, + 0.785, + 0.583 + ], + "angle": 0, + "content": "Gaussian primitive. A Gaussian primitive \\(\\gamma_{k} := (\\pmb{\\mu}_{k}, \\pmb{\\Sigma}_{k}, \\alpha_{k}, \\pmb{f}_{k})\\) geometrically resembles a 3D Gaussian kernel" + }, + { + "type": "equation", + "bbox": [ + 0.348, + 0.589, + 0.785, + 0.622 + ], + "angle": 0, + "content": "\\[\n\\mathcal {G} _ {k} (\\boldsymbol {x}) := \\exp \\left(- \\frac {1}{2} \\left(\\boldsymbol {x} - \\boldsymbol {\\mu} _ {k}\\right) ^ {\\top} \\Sigma_ {k} ^ {- 1} \\left(\\boldsymbol {x} - \\boldsymbol {\\mu} _ {k}\\right)\\right) \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.631, + 0.784, + 0.679 + ], + "angle": 0, + "content": "centered in \\(\\pmb{\\mu}_k\\in \\mathbb{R}^3\\) and having \\(\\Sigma_{k}\\) as its \\(3\\times 3\\) covariance matrix. Each primitive additionally entails an opacity factor \\(\\alpha_{k}\\in [0,1]\\) and a feature vector \\(\\pmb {f}_k\\in \\mathbb{R}^d\\) (e.g. RGB color or spherical harmonics coefficients)." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.683, + 0.785, + 0.776 + ], + "angle": 0, + "content": "Splatting. This is the operation of projecting a Gaussian primitive \\(\\gamma_{k}\\) to a camera pixel space via its world-to-image transformation \\(\\pi : \\mathbb{R}^3 \\to \\mathbb{R}^2\\), which we refer directly to as the camera for simplicity. The projection \\(\\pi\\) is approximated to the first order at the primitive's center \\(\\pmb{\\mu}_{k}\\) so that the projected primitive is geometrically equivalent to a 2D Gaussian kernel \\(\\mathcal{G}_k^\\pi\\) with mean \\(\\pi (\\pmb{\\mu}_k)\\in \\mathbb{R}^2\\) and 2D covariance \\(\\mathrm{J}_k^\\pi \\Sigma_k\\mathrm{J}_{\\mathbf{k}}^\\pi {}^\\top\\) with \\(\\mathrm{J}_k^\\pi\\) being the Jacobian of \\(\\pi\\) evaluated at \\(\\pmb{\\mu}_k\\)." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.78, + 0.786, + 0.841 + ], + "angle": 0, + "content": "Rendering. To render the primitives \\(\\varGamma\\) representing a scene from camera \\(\\pi\\), we require a decoder \\(\\varPhi\\) to be specified, which provides the feature we want to render as \\(\\varPhi(\\gamma_k,\\pmb{u})\\in\\mathbb{R}^m\\) for each Gaussian primitive \\(\\gamma_{k}\\) and pixel \\(\\pmb{u}\\). Moreover, we assume Gaussian primitives \\(\\varGamma\\) to be ordered with respect to their center's depth," + } + ], + [ + { + "type": "header", + "bbox": [ + 0.435, + 0.115, + 0.733, + 0.129 + ], + "angle": 0, + "content": "Revising Densification in Gaussian Splitting" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.147, + 0.784, + 0.178 + ], + "angle": 0, + "content": "when seen from the camera's reference frame. Then, the rendering equation takes the following form (with \\(\\varGamma\\) being omitted from the notation)" + }, + { + "type": "equation", + "bbox": [ + 0.386, + 0.192, + 0.786, + 0.232 + ], + "angle": 0, + "content": "\\[\n\\mathcal {R} [ \\pi , \\Phi ] (\\boldsymbol {u}) := \\sum_ {k = 1} ^ {K} \\Phi \\left(\\gamma_ {k}, \\boldsymbol {u}\\right) \\omega_ {k} ^ {\\pi} (\\boldsymbol {u}), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.242, + 0.621, + 0.259 + ], + "angle": 0, + "content": "where \\(\\omega_{k}^{\\pi}(\\pmb{u})\\) are alpha-compositing coefficients given by" + }, + { + "type": "equation", + "bbox": [ + 0.367, + 0.272, + 0.786, + 0.313 + ], + "angle": 0, + "content": "\\[\n\\omega_ {k} ^ {\\pi} (\\boldsymbol {u}) := \\alpha_ {k} \\mathcal {G} _ {k} ^ {\\pi} (\\boldsymbol {u}) \\prod_ {j = 1} ^ {k - 1} \\left(1 - \\alpha_ {j} \\mathcal {G} _ {j} ^ {\\pi} (\\boldsymbol {u})\\right). \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.324, + 0.784, + 0.428 + ], + "angle": 0, + "content": "If we assume the feature vectors \\( \\pmb{f}_k \\) to be spherical harmonics coefficients encoding an RGB function on the sphere, we can regard \\( \\varPhi_{\\mathrm{RGB}}(\\pmb{u}) \\) as the decoded RGB color for the given view direction associated to pixel \\( \\pmb{u} \\). If we use \\( \\varPhi_{\\mathrm{RGB}} \\) as the decoder in the rendering equation, we obtain a rendered color image \\( C_{\\pi}(\\pmb{u}) := \\mathcal{R}[\\pi, \\varPhi_{\\mathrm{RGB}}](\\pmb{u}) \\) for each camera \\( \\pi \\). Similarly, one can pick different \\( \\varPhi \\)'s to enable the rendering of depth, normals, or other quantities of interest as we will show later." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.435, + 0.784, + 0.511 + ], + "angle": 0, + "content": "Mip-splatting. In [29], the authors introduce a variation of standard Gaussian splatting that focuses on solving aliasing issues. We refer the reader to the original paper for details, but the idea is to track the maximum sampling rate for each Gaussian primitive and use it to reduce aliasing effects by attenuating the Gaussian primitives' opacity." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.535, + 0.467, + 0.552 + ], + "angle": 0, + "content": "3 Revising Densification" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.567, + 0.784, + 0.613 + ], + "angle": 0, + "content": "We first review the Adaptive Density Control module proposed in the original Gaussian splatting work [9], highlight some of its limitations, and then introduce our novel and improved densification procedure." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.636, + 0.631, + 0.651 + ], + "angle": 0, + "content": "3.1 Adaptive Density Control and its limitations" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.661, + 0.784, + 0.789 + ], + "angle": 0, + "content": "3DGS [9] and follow-up extensions (e.g. Mip-splatting [29]) rely on the Adaptive Density Control (ADC) module to grow or prune Gaussian primitives. This module is run according to a predetermined schedule and densification decisions are based on gradient statistics collected across the ADC runs. Specifically, for each Gaussian primitive \\(\\gamma_{k}\\) the positional gradient magnitude \\(\\left\\| \\frac{\\partial L_{\\pi}}{\\partial \\mu_k} \\right\\|\\) is tracked and averaged over all rendered views \\(\\pi \\in \\Pi\\) within the collection period, where \\(L_{\\pi}\\) denotes the loss that is optimized for camera \\(\\pi\\). The resulting quantity is denoted by \\(\\tau_{k}\\)." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.795, + 0.784, + 0.841 + ], + "angle": 0, + "content": "Growing. ADC grows new Gaussian primitives via a clone or a split operation. A primitive \\(\\gamma_{k}\\) will be considered for a growing operation only if \\(\\tau_{k}\\) exceeds a user-defined threshold. The decision about which operation to apply depends" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.381, + 0.128 + ], + "angle": 0, + "content": "Rota Bulò et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.239 + ], + "angle": 0, + "content": "on the size of the primitive measured in terms of the largest eigenvalue of the covariance matrix \\(\\Sigma_{k}\\). Specifically, primitives larger than a threshold are split, otherwise cloned. When a primitive \\(\\gamma_{k}\\) is split, two new primitives are generated with their position being sampled from \\(\\mathcal{G}_k\\) and their covariance being a scaled down version of \\(\\Sigma_{k}\\), while preserving the same opacity and feature vector. When a clone operation takes place, a simple clone of \\(\\gamma_{k}\\) is instantiated." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.244, + 0.788, + 0.305 + ], + "angle": 0, + "content": "Pruning. ADC prunes a Gaussian primitive \\(\\gamma_{k}\\) if its opacity \\(\\alpha_{k}\\) is below a user-defined threshold, typically 0.005. To ensure that an unused primitive is eventually pruned, a hard-reset of the opacity to a minimum value (usually 0.01) is enforced according to a predefined schedule." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.31, + 0.787, + 0.341 + ], + "angle": 0, + "content": "Limitations. Deciding which Gaussian primitives to split/clone based on the magnitude of the positional gradient suffers from a number of limitations:" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.348, + 0.786, + 0.377 + ], + "angle": 0, + "content": "- Determining a threshold for a gradient magnitude is not intuitive and very sensitive to modifications to the model, losses and hyperparameters," + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.377, + 0.786, + 0.407 + ], + "angle": 0, + "content": "- There are cases of scene underfitting also when the value of \\(\\tau_{k}\\) is below the threshold that triggers densification (see Fig. 1)." + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.407, + 0.787, + 0.452 + ], + "angle": 0, + "content": "- It is not possible to directly control the number of Gaussian primitives that are generated for a given scene, resulting in possible out-of-memory errors if their number grows abnormally." + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.348, + 0.787, + 0.452 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.458, + 0.788, + 0.504 + ], + "angle": 0, + "content": "In addition, we found that the ADC's logic of growing primitives suffers from a bias that weights more the contribution of freshly cloned primitives. More details will follow in Sec. 3.3" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.524, + 0.471, + 0.538 + ], + "angle": 0, + "content": "3.2 Error-based densification" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.546, + 0.788, + 0.667 + ], + "angle": 0, + "content": "Assume we have an image with an area characterized by a high-frequency pattern and covered by few large splatted Gaussian primitives (e.g. the grass in Fig. 1). Under this scenario, an infinitesimal change in the 3D location \\(\\pmb{\\mu}_k\\) of one of the corresponding Gaussian primitives \\(\\gamma_k\\) will leave the error almost unchanged and, hence, the collected magnitude of the positional gradient \\(\\tau_k\\) remains close to zero. In fact, \\(\\tau_k\\) is sensitive to error-changes, but is blind to the absolute value of the error. This becomes a problem, for we expect to increase the number of Gaussian primitives in areas exhibiting a larger error." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.668, + 0.788, + 0.804 + ], + "angle": 0, + "content": "Given the above considerations, we propose to steer the densification decisions directly based on an auxiliary per-pixel error function \\(\\mathcal{E}_{\\pi}\\) (e.g. Structural Similarity) that we measure when rendering on a camera \\(\\pi\\) with available ground-truth. One problem to address is how to turn per-pixel errors into per-Gaussian-primitive errors in light of the fact that each pixel error entangles the contribution of multiple Gaussian primitives. Our solution consists of first re-distributing the per-pixel errors \\(\\mathcal{E}_{\\pi}(\\pmb{u})\\) to each Gaussian primitive \\(\\gamma_{k}\\) proportionally to their contribution to the rendered pixel color, i.e. proportionally to \\(w_{k}^{\\pi}(\\pmb{u})\\). This yields the following error for each primitive \\(\\gamma_{k}\\) and camera \\(\\pi\\):" + }, + { + "type": "equation", + "bbox": [ + 0.412, + 0.813, + 0.786, + 0.844 + ], + "angle": 0, + "content": "\\[\nE _ {k} ^ {\\pi} := \\sum_ {\\boldsymbol {u} \\in \\operatorname {P i x}} \\mathcal {E} _ {\\pi} (\\boldsymbol {u}) w _ {k} ^ {\\pi} (\\boldsymbol {u}), \\tag {4}\n\\]" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.434, + 0.115, + 0.733, + 0.13 + ], + "angle": 0, + "content": "Revising Densification in Gaussian Splitting" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.193 + ], + "angle": 0, + "content": "where the sum runs over the image pixels. Then, for each primitive \\(\\gamma_{k}\\) we track the maximum value of the error \\(E_{k}^{\\pi}\\) across all views \\(\\pi \\in \\Pi\\) seen between two runs of the ADC module, i.e." + }, + { + "type": "equation", + "bbox": [ + 0.45, + 0.203, + 0.786, + 0.224 + ], + "angle": 0, + "content": "\\[\nE _ {k} := \\max _ {\\pi \\in \\Pi} E _ {k} ^ {\\pi}. \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.231, + 0.784, + 0.277 + ], + "angle": 0, + "content": "This is the score that we use to prioritize the growing of Gaussian primitives. As opposed to \\(\\tau_{k}\\), it is easier to set a threshold for our new densification score, for it is typically expressed in terms of a known error metric." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.283, + 0.785, + 0.343 + ], + "angle": 0, + "content": "Implementation details. In order to compute \\( E_{k}^{\\pi} \\) we assign an additional scalar \\( e_{k} \\) to each Gaussian primitive \\( \\gamma_{k} \\), and enable the possibility of rendering it via the decoder \\( \\varPhi_{\\mathrm{ERR}}(\\gamma_k,\\pmb {u})\\coloneqq e_k \\). Then, we add the following auxiliary loss to the standard Gaussian splatting training objective:" + }, + { + "type": "equation", + "bbox": [ + 0.37, + 0.353, + 0.786, + 0.392 + ], + "angle": 0, + "content": "\\[\nL _ {\\pi} ^ {\\text {a u x}} := \\sum_ {\\boldsymbol {u} \\in \\operatorname {P i x}} \\mathcal {N} [ \\mathcal {E} _ {\\pi} (\\boldsymbol {u}) ] \\underbrace {\\mathcal {R} [ \\pi , \\Phi_ {\\mathrm {E R R}} ] (\\boldsymbol {u})} _ {= \\sum_ {k = 1} ^ {K} e _ {k} \\omega_ {k} ^ {\\pi} (\\boldsymbol {u})}, \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.399, + 0.784, + 0.475 + ], + "angle": 0, + "content": "which is basically the dot product of the per-pixel error with gradient detached (denoted by \\(\\mathcal{X}\\)) and the rendering of the newly-added scalar. We initialize \\(e_k\\) to 0 for each Gaussian primitive \\(\\gamma_k\\) and never update it during training. In this way, \\(L_{\\pi}^{\\mathrm{aux}} = 0\\) and all Gaussian primitives' parameters, excepting \\(e_k\\), are left invariant by this loss. The gradient with respect to \\(e_k\\) instead yields" + }, + { + "type": "equation", + "bbox": [ + 0.383, + 0.482, + 0.786, + 0.518 + ], + "angle": 0, + "content": "\\[\n\\frac {\\partial L _ {\\pi} ^ {\\mathrm {a u x}}}{\\partial e _ {k}} = \\sum_ {\\boldsymbol {u} \\in \\operatorname {P i x}} \\mathcal {E} _ {\\pi} (\\boldsymbol {u}) \\omega_ {k} ^ {\\pi} (\\boldsymbol {u}) = E _ {k} ^ {\\pi}, \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.526, + 0.78, + 0.542 + ], + "angle": 0, + "content": "which is the per-Gaussian-primitive error for camera \\(\\pi\\) we wanted to compute." + }, + { + "type": "title", + "bbox": [ + 0.214, + 0.561, + 0.527, + 0.577 + ], + "angle": 0, + "content": "3.3 Opacity correction after cloning" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.584, + 0.786, + 0.795 + ], + "angle": 0, + "content": "In the original ADC module, when a Gaussian primitive is split or cloned, the opacity value is preserved. This choice introduces a bias in the case of the clone operation by implicitly increasing the impact of the densified primitive on the final rendered color. To see why this is the case we can follow the example in Fig. 2, where we consider what happens if we render a splatted Gaussian in its center pixel assuming an opacity value \\(\\alpha\\). Before a cloning operation happens, the rendered color depends on primitives that come next in the ordering with weight \\(1 - \\alpha\\). But after we clone, due to the alpha-compositing logic, we have that primitives that come next weight \\((1 - \\alpha)^{2}\\), which is lower than \\(1 - \\alpha\\) for all opacity values in \\((0,1)\\). Accordingly, by applying the standard logic of preserving the opacity after cloning we have a bias to weight more the cloned primitives. The solution we suggest consists in reducing the opacity of the primitives after cloning so that the bias is removed. The new opacity value \\(\\hat{\\alpha}\\) can be found by solving the equation \\((1 - \\alpha) = (1 - \\hat{\\alpha})^{2}\\), which yields \\(\\hat{\\alpha} := 1 - \\sqrt{1 - \\alpha}\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.796, + 0.786, + 0.84 + ], + "angle": 0, + "content": "If we depart from the simplified setting of considering only the center pixel and rather consider all pixels, it is unfortunately not possible to remove completely the bias. Nonetheless, the correction factor we introduce reduces the bias" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.381, + 0.127 + ], + "angle": 0, + "content": "Rota Bulò et al." + }, + { + "type": "image", + "bbox": [ + 0.216, + 0.145, + 0.334, + 0.22 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.368, + 0.143, + 0.532, + 0.218 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.556, + 0.143, + 0.779, + 0.217 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.227, + 0.788, + 0.312 + ], + "angle": 0, + "content": "Fig. 2: Consider rendering a single splatted Gaussian in its center pixel with opacity \\(\\alpha\\) before and after cloning. Before we clone, the rendered color depends with weight \\(1 - \\alpha\\) on what comes next. After we clone, since we preserve the opacity, the rendered color depends with weight \\((1 - \\alpha)^{2}\\) on what comes next. Since \\((1 - \\alpha) \\geq (1 - \\alpha)^{2}\\) we have a bias towards weighting more Gaussian primitives that get cloned. The proposed correction changes the opacity post clone to \\(\\hat{\\alpha}\\) so that the bias is removed." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.337, + 0.784, + 0.368 + ], + "angle": 0, + "content": "for all pixels compared to keeping the opacity of the cloned primitive. Indeed, the following relation holds for all \\(\\alpha_{k}\\in (0,1)\\) and all pixels \\(\\pmb{u}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.318, + 0.373, + 0.786, + 0.391 + ], + "angle": 0, + "content": "\\[\n1 - \\alpha_ {k} \\mathcal {G} _ {k} ^ {\\pi} (\\boldsymbol {u}) \\geq \\left(1 - \\hat {\\alpha} _ {k} \\mathcal {G} _ {k} ^ {\\pi} (\\boldsymbol {u})\\right) ^ {2} > \\left(1 - \\alpha_ {k} \\mathcal {G} _ {k} ^ {\\pi} (\\boldsymbol {u})\\right) ^ {2}, \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.397, + 0.784, + 0.445 + ], + "angle": 0, + "content": "where \\(\\hat{\\alpha}_k \\coloneqq 1 - \\sqrt{1 - \\alpha_k}\\) is our corrected opacity. The proof of the relation follows by noting that \\(\\hat{\\alpha}_k\\) can be rewritten as \\(\\frac{\\alpha_k}{1 + \\sqrt{1 - \\alpha_k}}\\) which is strictly smaller than \\(\\alpha_k\\) for \\(\\alpha_k \\in (0,1)\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.445, + 0.785, + 0.522 + ], + "angle": 0, + "content": "Finally, the correction of the opacity as shown above is derived assuming we clone a Gaussian primitive, but does not strictly match the case of a split operation, for when we split we move the two offspring randomly away from the previous center and we change the covariance scale. For this reason, we stick to the standard rule of preserving the opacity of a primitive we split." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.539, + 0.475, + 0.554 + ], + "angle": 0, + "content": "3.4 Primitives growth control" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.56, + 0.788, + 0.773 + ], + "angle": 0, + "content": "The ADC module grows a Gaussian primitive if \\(\\tau_{k}\\) is larger than a threshold. This mechanism can lead to unpredictable growth of the number of primitives, eventually resulting in out-of-memory issues. To avoid this problem, we introduce a global limit to the maximum number of Gaussian primitives, and a mechanism to control the maximum number of primitives that can be created each time densification is run. Among the many possible options, we explore a logic that limits new primitive offspring to a fixed fraction of the primitives that already exist. In case the number of primitives that are entitled to be densified exceeds the available budget, we retain only the ones that exhibit the highest densification score. An example of this process is shown in Fig. 3, compared to the one from 3DGS: For Ours, the number of primitives grows smoothly until it reaches the allotted maximum, without the discontinuities induced by opacity reset (see Sec. 3.5). The way we control the number of primitives is not limited to our error-based densification logic, but can be applied equally to the original gradient-based one." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.789, + 0.492, + 0.805 + ], + "angle": 0, + "content": "3.5 Alternative to opacity reset" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.842 + ], + "angle": 0, + "content": "The strategy introduced in [9] to favour the sparsification of Gaussian primitives consists in periodical hard resets of the opacity for all primitives to a low" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.435, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Revising Densification in Gaussian Splitting" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "image_caption", + "bbox": [ + 0.352, + 0.145, + 0.669, + 0.161 + ], + "angle": 0, + "content": "Number of Gaussians by training iteration" + }, + { + "type": "image", + "bbox": [ + 0.3, + 0.164, + 0.703, + 0.306 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.315, + 0.788, + 0.399 + ], + "angle": 0, + "content": "Fig. 3: Evolution of the number of Gaussians in 3DGS, and in our method with upper limit set to the number reached by 3DGS (on the garden scene from the Mip-NeRF 360 dataset). Note that, while 3DGS' ADC process stops after \\(15\\mathrm{k}\\) iterations, ours remains active for \\(27\\mathrm{k}\\). This is not immediately visible from the plot, since pruned primitives are immediately replaced by newly spawned ones, keeping the overall number stable once the maximum is reached." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.435, + 0.788, + 0.602 + ], + "angle": 0, + "content": "value, so that primitives whose opacity is not increased again by the optimization will eventually be pruned. This introduces a small shock in the training trajectory, which is suboptimal for the sake of having stable and predictable training dynamics. Moreover, resetting the opacity is particularly harmful for our error-based densification method, for it will lead to misleading error statistics right after the hard-reset, potentially triggering wrong densification decisions. For this reason, we propose a different logic to favour primitives pruning in a smoother way. Specifically, we decrease the opacity of each primitive by a fixed amount (we use 0.001) after each densification run, so that the opacity will gradually move towards the pruning range. In this way, we avoid sudden changes in the densification metric, while preserving the desired sparsification properties." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.605, + 0.788, + 0.71 + ], + "angle": 0, + "content": "One downside of the new opacity regularization logic is that the constant push towards lowering the opacity of the primitives implicitly invites the model to make more use of the background where possible. This is also harmful, for it could generate more holes in the scene that will be visible from novel views. To counteract this dynamics, we also regularize the residual probabilities of the alpha-compositing (a.k.a. residual transmittance) to be zero for every pixel, by simply minimizing their average value, weighted by a hyperparameter (here 0.1)." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.749, + 0.494, + 0.765 + ], + "angle": 0, + "content": "4 Experimental Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.787, + 0.84 + ], + "angle": 0, + "content": "In the following we show how our improved ADC mechanism can equally be applied both to standard 3DGS [9] and its Mip-Splatting extension [29], providing benefits to both." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.381, + 0.127 + ], + "angle": 0, + "content": "Rota Bulò et al." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.147, + 0.436, + 0.16 + ], + "angle": 0, + "content": "4.1 Datasets and metrics" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.175, + 0.788, + 0.326 + ], + "angle": 0, + "content": "We follow the experimental setup from the 3DGS [9] paper, focusing on the real-world scenes from the Mip-NeRF 360 [1], Tanks and Temples [10] and Deep Blending [6] datasets. Mip-NeRF 360 comprises nine scenes (5 outdoor, 4 indoor) captured in a circular pattern which focuses on a central area of a few meters, with a potentially unbounded background. For Tanks and Temples, we focus on the \"Truck\" and \"Train\" scenes, while for Deep Blending we focus on the \"Dr Johnson\" and \"Playroom\" scenes, using the images and SfM reconstructions shared by the Gaussian Splatting authors. In each experiment we set aside each 8th image as a validation set, and report peak signal-to-noise ratio (PSNR), structural similarity (SSIM) and the perceptual metric from [31] (LPIPS)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.326, + 0.787, + 0.356 + ], + "angle": 0, + "content": "Additional experiments on NeRF-Synthetic [17] are provided in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.382, + 0.424, + 0.397 + ], + "angle": 0, + "content": "4.2 Experimental setup" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.41, + 0.789, + 0.606 + ], + "angle": 0, + "content": "We evaluate based on our re-implementations of 3DGS, which allows us to easily switch between standard 3DGS, Mip-Splatting, the original ADC of [9], our contributions or any combination thereof. We reproduce the training settings proposed in [9,29] and the respective public code-bases\\(^{12}\\), including number of training iterations, batch size, input resolution, learning rates etc. When training with our contributions, we grow Gaussians with \\(E_{k} > 0.1\\), adding up to \\(5\\%\\) of the current number of primitives at each densification step. Differently from 3DGS, we keep our ADC process active for 27k iterations (i.e. \\(90\\%\\) of the training process), instead of stopping it after 15k. Other relevant hyper-parameters are left to the default values used in 3DGS, and shared across all datasets and scenes. In all our experiments, we set the maximum primitives budget to the number of primitives (or its median, for experiments with multiple runs) generated by the corresponding baseline, in order to obtain perfectly comparable models." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.612, + 0.788, + 0.703 + ], + "angle": 0, + "content": "A note about LPIPS. Investigating the 3DGS and Mip-Splating baselines, we discovered a discrepancy in the way LPIPS is calculated in both public code-bases, which resulted in under-estimated values being reported in the original papers. This was confirmed in private correspondence with the authors. In order to simplify comparisons with future works that don't rely on these code-bases, and might be unaware of this issue, we report correct LPIPS values here." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.71, + 0.788, + 0.801 + ], + "angle": 0, + "content": "A note about evaluations on MipNeRF360. The standard protocol for the dataset trains and evaluates the models at the same image scale, with indoor scenes down-sampled by a factor of two and outdoor scenes by four. The authors of the dataset provide down-scaled versions of the images. However, several recent works like [29] do not use the original images, which suffer from compression artifacts, but down-sample the images from the higher resolution counterparts." + }, + { + "type": "page_footnote", + "bbox": [ + 0.218, + 0.811, + 0.647, + 0.826 + ], + "angle": 0, + "content": "1 https://github.com/graphdeco-inria/gaussian-splatting" + }, + { + "type": "page_footnote", + "bbox": [ + 0.222, + 0.826, + 0.615, + 0.84 + ], + "angle": 0, + "content": "2 https://github.com/autonomousvision/mip-splatting" + }, + { + "type": "list", + "bbox": [ + 0.218, + 0.811, + 0.647, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.435, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Revising Densification in Gaussian Splitting" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "table_caption", + "bbox": [ + 0.217, + 0.145, + 0.785, + 0.186 + ], + "angle": 0, + "content": "Table 1: Results on the Mip-NeRF 360 dataset. Top section of the table: results from the Gaussian Splitting paper; bottom section: results from our re-implementation averaged over 5 runs." + }, + { + "type": "table", + "bbox": [ + 0.25, + 0.199, + 0.754, + 0.348 + ], + "angle": 0, + "content": "
MethodPSNR ↑SSIM ↑LPIPS ↓
MeanStdevMeanStdevMeanStdev
Plenoxels [4]23.08-0.626-0.436-
INGP [18]25.59-0.699-0.331-
Mip-NeRF 360 [1]27.69-0.792-0.237-
3DGS [9]27.450.0810.8170.0010.2500.001
Mip-Splatting [29]27.510.0740.8180.0010.2510.001
Ours, 3DGS27.610.0670.8220.0010.2230.001
Ours, Mip-Splatting27.700.0650.8230.0010.2230.001
" + }, + { + "type": "table_caption", + "bbox": [ + 0.217, + 0.361, + 0.785, + 0.402 + ], + "angle": 0, + "content": "Table 2: Results on the Tanks and Temples dataset. Top section of the table: results from the Gaussian Splitting paper; bottom section: results from our re-implementation averaged over 5 runs." + }, + { + "type": "table", + "bbox": [ + 0.25, + 0.415, + 0.754, + 0.564 + ], + "angle": 0, + "content": "
MethodPSNR ↑SSIM ↑LPIPS ↓
MeanStdevMeanStdevMeanStdev
Plenoxels [4]21.08-0.719-0.379-
INGP [18]21.92-0.745-0.305-
Mip-NeRF 360 [1]22.22-0.759-0.257-
3DGS [9]23.440.0890.8450.0020.2120.001
Mip-Splatting [29]23.650.0660.8490.0010.2110.001
Ours, 3DGS23.930.0700.8530.0010.1870.001
Ours, Mip-Splatting24.100.1090.8570.0020.1830.001
" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.593, + 0.785, + 0.638 + ], + "angle": 0, + "content": "As a result, the scores they report are significantly higher. We use the downscaled images from the dataset for training/evaluation to stay consistent with older NeRF works and report scores of competitors under the same setting." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.662, + 0.365, + 0.675 + ], + "angle": 0, + "content": "4.3 Main results" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.689, + 0.787, + 0.84 + ], + "angle": 0, + "content": "In a first set of experiments, we evaluate the effectiveness of our improved ADC strategy (Ours) when applied to 3DGS and Mip-Splatting. Results, collected over 5 training runs to average out the randomness induced by stochastic primitive splitting, are reported in Tab. 1, 2 and 3. For the sake of completeness, we also include scores obtained with three NeRF baselines, i.e. Plenoxels [4], InstantNGP (INGP) [18] and Mip-NeRF 360 [1], as originally reported in [9]. Our approach consistently outperforms the corresponding baselines (i.e. Ours, 3DGS vs. 3DGS; Ours, Mip-Splatting vs. Mip-Splatting), particularly on SSIM and LPIPS. This is in line with what we discussed in Sec. 3.1 and 3.2: standard ADC often leads to localized under-fitting, as it fails to split large gaussians" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.381, + 0.127 + ], + "angle": 0, + "content": "Rota Bulò et al." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.2, + 0.406, + 0.282 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.246, + 0.284, + 0.379, + 0.319 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.408, + 0.201, + 0.596, + 0.283 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.436, + 0.284, + 0.567, + 0.319 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.597, + 0.201, + 0.786, + 0.283 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.625, + 0.284, + 0.757, + 0.318 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.217, + 0.328, + 0.406, + 0.411 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.246, + 0.411, + 0.378, + 0.446 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.408, + 0.328, + 0.596, + 0.411 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.436, + 0.411, + 0.567, + 0.446 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.597, + 0.328, + 0.785, + 0.411 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.625, + 0.411, + 0.757, + 0.446 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.217, + 0.451, + 0.405, + 0.549 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.246, + 0.549, + 0.378, + 0.584 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.408, + 0.451, + 0.596, + 0.549 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.436, + 0.549, + 0.567, + 0.584 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.597, + 0.452, + 0.785, + 0.549 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.625, + 0.549, + 0.757, + 0.584 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.594, + 0.406, + 0.691 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.246, + 0.691, + 0.378, + 0.727 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.408, + 0.594, + 0.596, + 0.691 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.436, + 0.691, + 0.567, + 0.727 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.597, + 0.594, + 0.785, + 0.691 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.625, + 0.691, + 0.757, + 0.727 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.738, + 0.788, + 0.78 + ], + "angle": 0, + "content": "Fig. 4: Qualitative results on the Mip-NeRF 360 and Tanks and Temples validation sets. Both 3DGS and Ours use the same number of primitives. Best viewed on screen at high magnification." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.435, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Revising Densification in Gaussian Splitting" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.145, + 0.784, + 0.186 + ], + "angle": 0, + "content": "Table 3: Results on the Deep Blending dataset. Top section of the table: results from the Gaussian Splatting paper; bottom section: results from our re-implementation averaged over 5 runs." + }, + { + "type": "table", + "bbox": [ + 0.25, + 0.199, + 0.754, + 0.348 + ], + "angle": 0, + "content": "
MethodPSNR ↑SSIM ↑LPIPS ↓
MeanStdevMeanStdevMeanStdev
Plenoxels [4]23.06-0.795-0.510-
INGP [18]24.96-0.817-0.390-
Mip-NeRF 360 [1]29.40-0.901-0.245-
3DGS [9]29.540.0960.9020.0000.3110.001
Mip-Splatting [29]29.680.0680.9030.0000.3090.000
Ours, 3DGS29.500.1100.9040.0010.3050.001
Ours, Mip-Splatting29.640.1260.9050.0010.3030.001
" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.359, + 0.784, + 0.401 + ], + "angle": 0, + "content": "Table 4: Ablation experiments on the Mip-NeRF 360 dataset, adding individual contributions to 3DGS or removing them from Ours. OC: Opacity Correction, Sec. 3.3; GC: Growth Control, Sec. 3.4; OR: Opacity Regularization, Sec. 3.5." + }, + { + "type": "table", + "bbox": [ + 0.236, + 0.414, + 0.769, + 0.501 + ], + "angle": 0, + "content": "
Baseline3DGS+OROurs
+OC+GC+ORFull-OC-GC-OR
PSNR ↑27.4527.6527.3527.4827.6127.0427.5427.28
SSIM ↑0.8170.8220.8100.8180.8220.8120.8180.810
LPIPS ↓0.2500.2390.2560.2430.2230.2350.2230.234
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.526, + 0.785, + 0.677 + ], + "angle": 0, + "content": "that cover highly-textured regions of the scene. This kind of errors is poorly reflected by PSNR, which measures the \"average fit\" over image pixels, but are promptly detected by perceptual metrics like LPIPS. On the Deep Blending dataset, we observe thinner gaps, PSNR actually showing a small regression w.r.t. the baselines, although with low confidence margins. We suspect this might be related to the fact that Deep Blending contains many flat, untextured surfaces (see Fig. 4), that are particularly challenging to reconstruct accurately with 3DGS-like methods, independently of the ADC strategy being adopted. In the supplementary material, we provide an additional discussion about why some metrics do not reflect the actual perceived quality of the generated images." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.678, + 0.785, + 0.768 + ], + "angle": 0, + "content": "Figure 4 contains a qualitative comparison between standard 3DGS and 3DGS augmented with our contributions (Ours). Areas with under-fitting artifacts are highlighted, showing how these are notably ameliorated by our approach. It is also worth noting that Ours effectively maintains the same quality as 3DGS in non-problematic areas, producing a more perceptually accurate reconstruction while using the same number of primitives (see Sec. 4.2)." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.787, + 0.441, + 0.803 + ], + "angle": 0, + "content": "4.4 Ablation experiments" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.784, + 0.84 + ], + "angle": 0, + "content": "In Tab. 4 we ablate the effects of Opacity Correction (OC, Sec. 3.3), Growth Control (GC, Sec. 3.4) and Opacity Regularization (OR, Sec. 3.5) on the Mip-NeRF" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.381, + 0.127 + ], + "angle": 0, + "content": "Rota Bulò et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.39 + ], + "angle": 0, + "content": "360 dataset. In particular, we evaluate 3DGS augmented with each of these components (left side of the table), and our method with the components replaced by the corresponding baseline mechanism in 3DGS' standard ADC (right side of the table). First, we observe that OC, GC and OR all contribute to our method, as the Full version of Ours achieves the overall best results on all metrics, and removing them consistently degrades performance. Interestingly, Opacity Correction seems to have the largest impact here, as it produces both the largest increase in the scores when added to 3DGS, and the largest decrease when removed from Ours. Finally, Growth Control has a negative impact on 3DGS when utilized in isolation, while only slightly degrading the results when removed from Ours. Note that this observation doesn't detract from GC's usefulness as a strategy to control and limit the capacity of the model. We hypothesize that GC's negative effect on 3DGS might be a consequence of the fact that the standard, gradient-based densification score is actually a poor choice for comparing gaussians in terms of how soon they should be split or cloned (remember that GC ranks Gaussians based on their score)." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.419, + 0.357, + 0.433 + ], + "angle": 0, + "content": "4.5 Limitations" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.452, + 0.788, + 0.634 + ], + "angle": 0, + "content": "While our method appears to be quite effective at solving under-fitting issues, these can still be present in especially difficult scenes (e.g. treehill in the Mip-NeRF 360 dataset, both scenes from the Deep Blending dataset). Focusing on the problematic areas that our ADC approach handles successfully, we observe that, while perceptually more \"correct\", the reconstruction there can still be quite inaccurate when closely compared to the ground truth (see e.g. the flowers scene in Fig. 4). We suspect both these issues might be related to 3DGS' intrinsic limits in handling i) strong view-dependent effects; ii) appearance variations across images; and iii) errors induced by the linear approximation in the Splitting operation (see Sec. 2). An interesting future direction could be to combine our approach with works that address these issues, e.g. Spec-Gaussian [26] for (i) and GS++ [7] for (iii)." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.665, + 0.36, + 0.681 + ], + "angle": 0, + "content": "5 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.704, + 0.788, + 0.841 + ], + "angle": 0, + "content": "In this paper, we addressed the limitations of the Adaptive Density Control (ADC) mechanism in 3D Gaussian Splatting (3DGS), a scene representation method for high-quality, photorealistic rendering. Our main contribution is a more principled, pixel-error driven formulation for density control in 3DGS. We propose how to leverage a novel decision criterion for densification based on per-pixel errors and introduce a mechanism to control the total number of primitives generated per scene. We also correct a bias in the current opacity handling in ADC during cloning. Our approach leads to consistent and systematic improvements over previous methods, particularly in perceptual metrics like LPIPS." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.435, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Revising Densification in Gaussian Splitting" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.323, + 0.16 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.173, + 0.787, + 0.228 + ], + "angle": 0, + "content": "1. Barron, J.T., Mildenhall, B., Verbin, D., Srinivasan, P.P., Hedman, P.: Mipnerf 360: Unbounded anti-aliased neural radiance fields. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5470-5479 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.229, + 0.786, + 0.255 + ], + "angle": 0, + "content": "2. Chen, Z., Wang, F., Liu, H.: Text-to-3d using gaussian splatting. arXiv preprint arXiv:2309.16585 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.256, + 0.787, + 0.296 + ], + "angle": 0, + "content": "3. Cheng, K., Long, X., Yang, K., Yao, Y., Yin, W., Ma, Y., Wang, W., Chen, X.: Gaussianpro: 3d gaussian splatting with progressive propagation. arXiv preprint arXiv:2402.14650 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.297, + 0.787, + 0.338 + ], + "angle": 0, + "content": "4. Fridovich-Keil, S., Yu, A., Tancik, M., Chen, Q., Recht, B., Kanazawa, A.: Plenoxels: Radiance fields without neural networks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5501-5510 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.338, + 0.787, + 0.378 + ], + "angle": 0, + "content": "5. Guédon, A., Lepetit, V.: Sugar: Surface-aligned gaussian splatting for efficient 3d mesh reconstruction and high-quality mesh rendering. arXiv preprint arXiv:2311.12775 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.379, + 0.787, + 0.419 + ], + "angle": 0, + "content": "6. Hedman, P., Philip, J., Price, T., Frahm, J.M., Drettakis, G., Brostow, G.: Deep blending for free-viewpoint image-based rendering. ACM Transactions on Graphics (ToG) 37(6), 1-15 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.42, + 0.787, + 0.446 + ], + "angle": 0, + "content": "7. Huang, L., Bai, J., Guo, J., Guo, Y.: \\(\\mathrm{Gs} + +\\): Error analyzing and optimal gaussian splatting. arXiv preprint arXiv:2402.00752 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.447, + 0.787, + 0.488 + ], + "angle": 0, + "content": "8. Keetha, N., Karhade, J., Jatavallabhula, K.M., Yang, G., Scherer, S., Ramanan, D., Luiten, J.: Splatam: Splat, track & map 3d gaussians for dense rgb-d slam. arXiv preprint arXiv:2312.02126 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.488, + 0.787, + 0.528 + ], + "angle": 0, + "content": "9. Kerbl, B., Kopanas, G., Leimkuhler, T., Drettakis, G.: 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics 42(4) (July 2023), https://repo-sam.inria.fr/fungraph/3d-gaussian-splatting/" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.529, + 0.787, + 0.569 + ], + "angle": 0, + "content": "0. Knapitsch, A., Park, J., Zhou, Q.Y., Koltun, V.: Tanks and temples: Benchmarking large-scale scene reconstruction. ACM Transactions on Graphics (ToG) 36(4), 1-13 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.569, + 0.787, + 0.596 + ], + "angle": 0, + "content": "1. Kocabas, M., Chang, J.H.R., Gabriel, J., Tuzel, O., Ranjan, A.: Hugs: Human gaussian splats. arXiv preprint arXiv:2311.17910 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.597, + 0.787, + 0.623 + ], + "angle": 0, + "content": "2. Lee, J.C., Rho, D., Sun, X., Ko, J.H., Park, E.: Compact 3d gaussian representation for radiance field. arXiv preprint arXiv:2311.13681 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.623, + 0.787, + 0.65 + ], + "angle": 0, + "content": "3. Lei, J., Wang, Y., Pavlakos, G., Liu, L., Daniilidis, K.: Gart: Gaussian articulated template models. arXiv preprint arXiv:2311.16099 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.65, + 0.787, + 0.691 + ], + "angle": 0, + "content": "4. Lu, T., Yu, M., Xu, L., Xiangli, Y., Wang, L., Lin, D., Dai, B.: Scaffold-gs: Structured 3d gaussians for view-adaptive rendering. arXiv preprint arXiv:2312.00109 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.691, + 0.787, + 0.718 + ], + "angle": 0, + "content": "5. Luiten, J., Kopanas, G., Leibe, B., Ramanan, D.: Dynamic 3d gaussians: Tracking by persistent dynamic view synthesis. arXiv preprint arXiv:2308.09713 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.718, + 0.787, + 0.744 + ], + "angle": 0, + "content": "6. Matsuki, H., Murai, R., Kelly, P.H., Davison, A.J.: Gaussian splatting slam. arXiv preprint arXiv:2312.06741 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.745, + 0.787, + 0.785 + ], + "angle": 0, + "content": "7. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. In: ECCV. p. 405-421 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.785, + 0.787, + 0.839 + ], + "angle": 0, + "content": "8. Müller, T., Evans, A., Schied, C., Keller, A.: Instant neural graphics primitives with a multiresolution hash encoding. ACM Trans. Graph. 41(4), 102:1-102:15 (Jul 2022). https://doi.org/10.1145/3528223.3530127, https://doi.org/10.1145/3528223.3530127" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.173, + 0.787, + 0.839 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.381, + 0.127 + ], + "angle": 0, + "content": "Rota Bulò et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.148, + 0.785, + 0.175 + ], + "angle": 0, + "content": "19. Saito, S., Schwartz, G., Simon, T., Li, J., Nam, G.: Relightable gaussian codec avatars. arXiv preprint arXiv:2312.03704 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.177, + 0.785, + 0.203 + ], + "angle": 0, + "content": "20. Snavely, N., Seitz, S.M., Szeliski, R.: Photo tourism: exploring photo collections in 3d. In: ACM siggraph 2006 papers, pp. 835-846 (2006)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.204, + 0.785, + 0.231 + ], + "angle": 0, + "content": "21. Tang, J., Ren, J., Zhou, H., Liu, Z., Zeng, G.: Dreamgaussian: Generative gaussian splatting for efficient 3d content creation. arXiv preprint arXiv:2309.16653 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.232, + 0.785, + 0.272 + ], + "angle": 0, + "content": "22. Wu, G., Yi, T., Fang, J., Xie, L., Zhang, X., Wei, W., Liu, W., Tian, Q., Wang, X.: 4d gaussian splatting for real-time dynamic scene rendering. arXiv preprint arXiv:2310.08528 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.273, + 0.785, + 0.313 + ], + "angle": 0, + "content": "23. Xie, T., Zong, Z., Qiu, Y., Li, X., Feng, Y., Yang, Y., Jiang, C.: Physgaus-sian: Physics-integrated 3d gaussians for generative dynamics. arXiv preprint arXiv:2311.12198 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.315, + 0.785, + 0.342 + ], + "angle": 0, + "content": "24. Yan, C., Qu, D., Wang, D., Xu, D., Wang, Z., Zhao, B., Li, X.: Gs-slam: Dense visual slam with 3d gaussian splatting. arXiv preprint arXiv:2311.11700 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.343, + 0.785, + 0.383 + ], + "angle": 0, + "content": "25. Yang, Z., Yang, H., Pan, Z., Zhu, X., Zhang, L.: Real-time photorealistic dynamic scene representation and rendering with 4d gaussian splatting. arXiv preprint arXiv:2310.10642 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.384, + 0.785, + 0.424 + ], + "angle": 0, + "content": "26. Yang, Z., Gao, X., Sun, Y., Huang, Y., Lyu, X., Zhou, W., Jiao, S., Qi, X., Jin, X.: Spec-gaussian: Anisotropic view-dependent appearance for 3d gaussian splatting. arXiv preprint arXiv:2402.15870 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.426, + 0.785, + 0.452 + ], + "angle": 0, + "content": "27. Ye, M., Danelljan, M., Yu, F., Ke, L.: Gaussian grouping: Segment and edit anything in 3d scenes. arXiv preprint arXiv:2312.00732 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.453, + 0.785, + 0.493 + ], + "angle": 0, + "content": "28. Yi, T., Fang, J., Wu, G., Xie, L., Zhang, X., Liu, W., Tian, Q., Wang, X.: Gaussian dreamer: Fast generation from text to 3d gaussian splatting with point cloud priors. arXiv preprint arXiv:2310.08529 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.494, + 0.785, + 0.522 + ], + "angle": 0, + "content": "29. Yu, Z., Chen, A., Huang, B., Sattler, T., Geiger, A.: Mip-splatting: Alias-free 3d gaussian splatting. arXiv preprint arXiv:2311.16493 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.523, + 0.785, + 0.55 + ], + "angle": 0, + "content": "30. Yugay, V., Li, Y., Gevers, T., Oswald, M.R.: Gaussian-slam: Photo-realistic dense slam with gaussian splatting. arXiv preprint arXiv:2312.10070 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.551, + 0.785, + 0.591 + ], + "angle": 0, + "content": "31. Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual metric. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 586-595 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.592, + 0.785, + 0.619 + ], + "angle": 0, + "content": "32. Zielonka, W., Bagautdinov, T., Saito, S., Zollhöfer, M., Thies, J., Romero, J.: Drivable 3d gaussian avatars. arXiv preprint arXiv:2311.08581 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.62, + 0.785, + 0.659 + ], + "angle": 0, + "content": "33. Zwicker, M., Pfister, H., van Baar, J., Gross, M.: Ewa volume splatting. In: Proceedings Visualization, 2001. VIS '01. pp. 29-538 (2001). https://doi.org/10.1109/VISUAL.2001.964490" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.148, + 0.785, + 0.659 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/Revising Densification in Gaussian Splatting/918310b7-b91b-411d-ade7-18df636ce501_origin.pdf b/2024/Revising Densification in Gaussian Splatting/918310b7-b91b-411d-ade7-18df636ce501_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..50f56a0542ac2920939558321a583c4daa2d55d9 --- /dev/null +++ b/2024/Revising Densification in Gaussian Splatting/918310b7-b91b-411d-ade7-18df636ce501_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2562ec8e3d13bc10aaa3377ca87991986e98c485be963dd6a69bb9541e7162fe +size 8965026 diff --git a/2024/Revising Densification in Gaussian Splatting/full.md b/2024/Revising Densification in Gaussian Splatting/full.md new file mode 100644 index 0000000000000000000000000000000000000000..8a3270d0f02bc5c3f61fd12e01221c9a68d55040 --- /dev/null +++ b/2024/Revising Densification in Gaussian Splatting/full.md @@ -0,0 +1,313 @@ +# Revising Densification in Gaussian Splitting + +Samuel Rota Bulò, Lorenzo Porzi, and Peter Kontschieder + +Meta Reality Labs Zurich + +![](images/01da4aa633a1585e8b5f170e748342671bc84b2736e2a61b43c97120485ef712.jpg) +Fig. 1: Densification is a critical component of 3D Gaussian Splatting (3DGS), and a common failure point. In this example (ground truth on the left) we show how 3DGS can fail (center) to add primitives to high-texture areas, like the grass in the bottom part of the pictures, producing large and blurry artifacts. Our approach (right) solves this issue by comprehensively revising densification in 3DGS. + +![](images/0aa0650b8e46425032dbe6a5487b289f066b1aa6f81b3dbb18502e95b264f3cb.jpg) + +![](images/3d4d96cac9f0c8b618afe30ec5fa20aa673d537d5373f7e503097eb836cffc54.jpg) + +Abstract. In this paper, we address the limitations of Adaptive Density Control (ADC) in 3D Gaussian Splatting (3DGS), a scene representation method achieving high-quality, photorealistic results for novel view synthesis. ADC has been introduced for automatic 3D point primitive management, controlling densification and pruning, however, with certain limitations in the densification logic. Our main contribution is a more principled, pixel-error driven formulation for density control in 3DGS, leveraging an auxiliary, per-pixel error function as the criterion for densification. We further introduce a mechanism to control the total number of primitives generated per scene and correct a bias in the current opacity handling strategy of ADC during cloning operations. Our approach leads to consistent quality improvements across a variety of benchmark scenes, without sacrificing the method's efficiency. + +Keywords: Gaussian Splatting $\cdot$ 3D reconstruction $\cdot$ Novel View Synthesis + +# 1 Introduction + +High-quality, photorealistic scene modelling from images has been an important research area in computer vision and graphics, with plentiful applications in AR/VR/MR, robotics, etc. In the last years, this field has gained a lot of attention due to advances in Neural 3D scene representations, particularly Neural Radiance Fields (NeRFs) [17]. NeRFs take a new approach to 3D scene representation and rendering, by leveraging a combination of deep learning and volumetric rendering techniques for generating photorealistic images from novel + +viewpoints. By optimizing MLPs to map from spatial coordinates and viewing directions to density and colour fields, these models have demonstrated astonishing capabilities for capturing the complex interplay of light and geometry in a data-driven way. While highly efficient in terms of representation quality, the original NeRF representation relies on time-consuming sampling strategies and thus excludes applications with fast rendering requirements. With many advances in terms of the underlying representation, these models have been significantly optimized towards improved training time and scene representation fidelity. However, inference speed for high-resolution, novel view synthesis remains an ongoing limitation. + +More recently, 3D Gaussian Splatting (3DGS) [9] has been proposed as an alternative and expressive scene representation, enabling both high-speed, high-fidelity training of models and high-resolution, GPU rasterization-friendly rendering of novel views. Their core representation is an optimized set of (anisotropic) 3D Gaussians, after being randomly distributed in 3D space, or systematically initialized at points obtained by Structure-from-Motion [20]. For obtaining a 2D image, all relevant 3D primitives are efficiently rendered via splatting-based rasterization with low-pass filtering. + +In 3DGS, each 3D primitive is parameterized as a 3D Gaussian distribution (i.e., with position and covariance), together with parameters controlling its opacity and describing its directional appearance (typically spherical harmonics). The parameter optimization procedure is guided by a multi-view, photometric loss, and is interleaved with Adaptive Density Control (ADC), a mechanism controlling density management for 3D points by means of introducing or deleting 3D primitives. ADC plays a critical role as it determines where to expand/shrink the scene representation budget for empty or over-reconstructed regions, respectively. Both growing and pruning operations are activated based on user-defined thresholds: Growing depends on the accumulated positional gradients of existing primitives and is, conditioned on the size of the Gaussians, executed by either splitting large primitives or by cloning smaller ones. Pruning is activated once the opacity falls below a provided threshold. While quite effective in practice, such density management strategies have several limitations. First, estimating a gradient magnitude-based threshold is rather non-intuitive and not robust to potential changes in the model, loss terms, etc. Second, there are cases where only few and large Gaussians are modeling high-frequency patterns like grass as shown in the middle of Fig. 1. Here, changes accumulated from positional gradients might remain very low and thus fail to trigger the densification mechanism, which in turn leads to substantial scene underfitting. Finally, ADC lacks explicit control of the maximum number of Gaussians generated per scene. This has important, practical implications as uncontrolled growth might easily lead to out-of-memory errors during training. + +In this work we address the shortcomings of Adaptive Density Control proposed in the original 3D Gaussian splatting method. Our core contribution is a more principled, pixel-error driven formulation for density control in 3DGS. We describe how 2D, per-pixel errors as e.g. derived from Structural Similarity + +(or any other informative objective function) can be propagated back as errors to contributing Gaussian primitives. In our solution, we first break down the per-pixel errors according to each Gaussian's contribution, and in a camera-specific way. This allows us to track the maximum error per primitive for all views and across two subsequent ADC runs, yielding our novel, error-specific, and thus more intuitive decision criterion for densification. Our second contribution is correcting a bias introduced with the current form of opacity handling in ADC when conducting a primitive cloning operation. The original approach suggests to keep the same opacity for the cloned Gaussian, which however biases the alpha-compositing logic applied for rendering the pixel colors. Indeed, this procedure leads to an overall increase of opacity in the cloned region, preventing the model to correctly account for contributions of other primitives and thus negatively affecting the densification process. Our third contribution is a mechanism for controlling the total number of primitives generated per scene and the maximum amount of novel primitives introduced per densification run. With this functionality, we can avoid undesired out-of-memory errors and better tune the method's behaviour w.r.t. given hardware constraints. We extensively validate our contributions on standard benchmark datasets like Mip-NeRF 360 [1], Tanks and Temples [10], and Deep Blending [6]. Our experiments show consistent improvements over different baselines including 3DGS [9] and Mip-Splatting [29]. To summarize, our contributions are improving methodological shortcomings in 3DGS' Adaptive Density Control mechanism as follows: + +- We propose a principled approach that enables the guidance of the densification process according to an auxiliary, per-pixel error function, rather than relying on positional gradients. +- We correct an existing, systematic bias from the primitive growing procedure when cloning Gaussians, negatively impacting the overall densification. +- We present ablations and experimental evaluations on different, real-world benchmarks, confirming quantitative and qualitative improvements. + +# 1.1 Related works + +Since it was presented in [9], 3DGS has been used in a remarkably wide set of downstream applications, including Simultaneous Localization and Mapping [8, 16,24,30], text-to-3D generation [2,21,28], photo-realistic human avatars [11,13, 19,32], dynamic scene modeling [15,22,25] and more [5,23,27]. However, only a handful of works like ours have focused on advancing 3DGS itself, by improving its quality or overcoming some of its limitations. + +In $\mathrm{GS} + +$ [7], Huang et al. present an improved approximation of the 3D-to-2D splatting operation at the core of 3DGS, which achieves better accuracy near image edges and solves some common visual artifacts. Spec-Gaussian [26] and Scaffold-gs [14] focus on improving view-dependent appearance modeling: the former by replacing spherical harmonics with an anisotropic spherical Gaussian appearance field; the latter by making all 3D Gaussian parameters, including whether specific primitives should be rendered or not, dependent on view direction through a small MLP. Mip-Splatting [29] tackles the strong artifacts that + +appear in 3DGS models when they are rendered at widely different resolutions (or viewing distances) compared to the images they were trained on. To do this, Yu et al. propose to incorporate a 3D filter to constrain the size of the 3D primitives depending on their maximal sampling rate on the training views, and a 2D Mip filter to mitigate aliasing issues. All these works adopt the original ADC strategy proposed in [9], and can potentially benefit from our improved approach, as we show for Mip-Splatting in Sec. 4. + +Only few concurrent works have touched on densification, while putting most of their focus on other aspects of 3DGS. Lee et al. [12] propose a quantization-based approach to produce more compact 3DGS representations, which includes a continuous sparsification strategy that takes both primitive size and opacity into account. GaussianPro [3] directly tackles related densification limitations as we explore in our work, filling the gaps from SfM-based initialization. They propose a rather complex procedure based on the progressive propagation of primitives along estimated planes, using patch-matching and geometric consistency as guidance. In contrast to our method, [3] focuses on fixing the quality of planar regions, instead of holistically improving densification. We also note that a fair comparison with their method on the standard Mip-NeRF 360 benchmark is not feasible at the time of submission, as the authors did not publicly share the improved SfM point cloud used in their experiments (see §5.2 of [3]). + +# 2 Preliminaries: Gaussian Splitting + +Gaussian Splatting [9] revisits ideas from EWA splatting [33] and proposes to fit a 3D scene as a collection of 3D Gaussian primitives $\Gamma \coloneqq \{\gamma_1,\dots ,\gamma_K\}$ that can be rendered by leveraging volume splatting. + +Gaussian primitive. A Gaussian primitive $\gamma_{k} := (\pmb{\mu}_{k}, \pmb{\Sigma}_{k}, \alpha_{k}, \pmb{f}_{k})$ geometrically resembles a 3D Gaussian kernel + +$$ +\mathcal {G} _ {k} (\boldsymbol {x}) := \exp \left(- \frac {1}{2} \left(\boldsymbol {x} - \boldsymbol {\mu} _ {k}\right) ^ {\top} \Sigma_ {k} ^ {- 1} \left(\boldsymbol {x} - \boldsymbol {\mu} _ {k}\right)\right) \tag {1} +$$ + +centered in $\pmb{\mu}_k\in \mathbb{R}^3$ and having $\Sigma_{k}$ as its $3\times 3$ covariance matrix. Each primitive additionally entails an opacity factor $\alpha_{k}\in [0,1]$ and a feature vector $\pmb {f}_k\in \mathbb{R}^d$ (e.g. RGB color or spherical harmonics coefficients). + +Splatting. This is the operation of projecting a Gaussian primitive $\gamma_{k}$ to a camera pixel space via its world-to-image transformation $\pi : \mathbb{R}^3 \to \mathbb{R}^2$ , which we refer directly to as the camera for simplicity. The projection $\pi$ is approximated to the first order at the primitive's center $\pmb{\mu}_{k}$ so that the projected primitive is geometrically equivalent to a 2D Gaussian kernel $\mathcal{G}_k^\pi$ with mean $\pi (\pmb{\mu}_k)\in \mathbb{R}^2$ and 2D covariance $\mathrm{J}_k^\pi \Sigma_k\mathrm{J}_{\mathbf{k}}^\pi {}^\top$ with $\mathrm{J}_k^\pi$ being the Jacobian of $\pi$ evaluated at $\pmb{\mu}_k$ . + +Rendering. To render the primitives $\varGamma$ representing a scene from camera $\pi$ , we require a decoder $\varPhi$ to be specified, which provides the feature we want to render as $\varPhi(\gamma_k,\pmb{u})\in\mathbb{R}^m$ for each Gaussian primitive $\gamma_{k}$ and pixel $\pmb{u}$ . Moreover, we assume Gaussian primitives $\varGamma$ to be ordered with respect to their center's depth, + +when seen from the camera's reference frame. Then, the rendering equation takes the following form (with $\varGamma$ being omitted from the notation) + +$$ +\mathcal {R} [ \pi , \Phi ] (\boldsymbol {u}) := \sum_ {k = 1} ^ {K} \Phi \left(\gamma_ {k}, \boldsymbol {u}\right) \omega_ {k} ^ {\pi} (\boldsymbol {u}), \tag {2} +$$ + +where $\omega_{k}^{\pi}(\pmb{u})$ are alpha-compositing coefficients given by + +$$ +\omega_ {k} ^ {\pi} (\boldsymbol {u}) := \alpha_ {k} \mathcal {G} _ {k} ^ {\pi} (\boldsymbol {u}) \prod_ {j = 1} ^ {k - 1} \left(1 - \alpha_ {j} \mathcal {G} _ {j} ^ {\pi} (\boldsymbol {u})\right). \tag {3} +$$ + +If we assume the feature vectors $\pmb{f}_k$ to be spherical harmonics coefficients encoding an RGB function on the sphere, we can regard $\varPhi_{\mathrm{RGB}}(\pmb{u})$ as the decoded RGB color for the given view direction associated to pixel $\pmb{u}$ . If we use $\varPhi_{\mathrm{RGB}}$ as the decoder in the rendering equation, we obtain a rendered color image $C_{\pi}(\pmb{u}) := \mathcal{R}[\pi, \varPhi_{\mathrm{RGB}}](\pmb{u})$ for each camera $\pi$ . Similarly, one can pick different $\varPhi$ 's to enable the rendering of depth, normals, or other quantities of interest as we will show later. + +Mip-splatting. In [29], the authors introduce a variation of standard Gaussian splatting that focuses on solving aliasing issues. We refer the reader to the original paper for details, but the idea is to track the maximum sampling rate for each Gaussian primitive and use it to reduce aliasing effects by attenuating the Gaussian primitives' opacity. + +# 3 Revising Densification + +We first review the Adaptive Density Control module proposed in the original Gaussian splatting work [9], highlight some of its limitations, and then introduce our novel and improved densification procedure. + +# 3.1 Adaptive Density Control and its limitations + +3DGS [9] and follow-up extensions (e.g. Mip-splatting [29]) rely on the Adaptive Density Control (ADC) module to grow or prune Gaussian primitives. This module is run according to a predetermined schedule and densification decisions are based on gradient statistics collected across the ADC runs. Specifically, for each Gaussian primitive $\gamma_{k}$ the positional gradient magnitude $\left\| \frac{\partial L_{\pi}}{\partial \mu_k} \right\|$ is tracked and averaged over all rendered views $\pi \in \Pi$ within the collection period, where $L_{\pi}$ denotes the loss that is optimized for camera $\pi$ . The resulting quantity is denoted by $\tau_{k}$ . + +Growing. ADC grows new Gaussian primitives via a clone or a split operation. A primitive $\gamma_{k}$ will be considered for a growing operation only if $\tau_{k}$ exceeds a user-defined threshold. The decision about which operation to apply depends + +on the size of the primitive measured in terms of the largest eigenvalue of the covariance matrix $\Sigma_{k}$ . Specifically, primitives larger than a threshold are split, otherwise cloned. When a primitive $\gamma_{k}$ is split, two new primitives are generated with their position being sampled from $\mathcal{G}_k$ and their covariance being a scaled down version of $\Sigma_{k}$ , while preserving the same opacity and feature vector. When a clone operation takes place, a simple clone of $\gamma_{k}$ is instantiated. + +Pruning. ADC prunes a Gaussian primitive $\gamma_{k}$ if its opacity $\alpha_{k}$ is below a user-defined threshold, typically 0.005. To ensure that an unused primitive is eventually pruned, a hard-reset of the opacity to a minimum value (usually 0.01) is enforced according to a predefined schedule. + +Limitations. Deciding which Gaussian primitives to split/clone based on the magnitude of the positional gradient suffers from a number of limitations: + +- Determining a threshold for a gradient magnitude is not intuitive and very sensitive to modifications to the model, losses and hyperparameters, +- There are cases of scene underfitting also when the value of $\tau_{k}$ is below the threshold that triggers densification (see Fig. 1). +- It is not possible to directly control the number of Gaussian primitives that are generated for a given scene, resulting in possible out-of-memory errors if their number grows abnormally. + +In addition, we found that the ADC's logic of growing primitives suffers from a bias that weights more the contribution of freshly cloned primitives. More details will follow in Sec. 3.3 + +# 3.2 Error-based densification + +Assume we have an image with an area characterized by a high-frequency pattern and covered by few large splatted Gaussian primitives (e.g. the grass in Fig. 1). Under this scenario, an infinitesimal change in the 3D location $\pmb{\mu}_k$ of one of the corresponding Gaussian primitives $\gamma_k$ will leave the error almost unchanged and, hence, the collected magnitude of the positional gradient $\tau_k$ remains close to zero. In fact, $\tau_k$ is sensitive to error-changes, but is blind to the absolute value of the error. This becomes a problem, for we expect to increase the number of Gaussian primitives in areas exhibiting a larger error. + +Given the above considerations, we propose to steer the densification decisions directly based on an auxiliary per-pixel error function $\mathcal{E}_{\pi}$ (e.g. Structural Similarity) that we measure when rendering on a camera $\pi$ with available ground-truth. One problem to address is how to turn per-pixel errors into per-Gaussian-primitive errors in light of the fact that each pixel error entangles the contribution of multiple Gaussian primitives. Our solution consists of first re-distributing the per-pixel errors $\mathcal{E}_{\pi}(\pmb{u})$ to each Gaussian primitive $\gamma_{k}$ proportionally to their contribution to the rendered pixel color, i.e. proportionally to $w_{k}^{\pi}(\pmb{u})$ . This yields the following error for each primitive $\gamma_{k}$ and camera $\pi$ : + +$$ +E _ {k} ^ {\pi} := \sum_ {\boldsymbol {u} \in \operatorname {P i x}} \mathcal {E} _ {\pi} (\boldsymbol {u}) w _ {k} ^ {\pi} (\boldsymbol {u}), \tag {4} +$$ + +where the sum runs over the image pixels. Then, for each primitive $\gamma_{k}$ we track the maximum value of the error $E_{k}^{\pi}$ across all views $\pi \in \Pi$ seen between two runs of the ADC module, i.e. + +$$ +E _ {k} := \max _ {\pi \in \Pi} E _ {k} ^ {\pi}. \tag {5} +$$ + +This is the score that we use to prioritize the growing of Gaussian primitives. As opposed to $\tau_{k}$ , it is easier to set a threshold for our new densification score, for it is typically expressed in terms of a known error metric. + +Implementation details. In order to compute $E_{k}^{\pi}$ we assign an additional scalar $e_{k}$ to each Gaussian primitive $\gamma_{k}$ , and enable the possibility of rendering it via the decoder $\varPhi_{\mathrm{ERR}}(\gamma_k,\pmb {u})\coloneqq e_k$ . Then, we add the following auxiliary loss to the standard Gaussian splatting training objective: + +$$ +L _ {\pi} ^ {\text {a u x}} := \sum_ {\boldsymbol {u} \in \operatorname {P i x}} \mathcal {N} [ \mathcal {E} _ {\pi} (\boldsymbol {u}) ] \underbrace {\mathcal {R} [ \pi , \Phi_ {\mathrm {E R R}} ] (\boldsymbol {u})} _ {= \sum_ {k = 1} ^ {K} e _ {k} \omega_ {k} ^ {\pi} (\boldsymbol {u})}, \tag {6} +$$ + +which is basically the dot product of the per-pixel error with gradient detached (denoted by $\mathcal{X}$ ) and the rendering of the newly-added scalar. We initialize $e_k$ to 0 for each Gaussian primitive $\gamma_k$ and never update it during training. In this way, $L_{\pi}^{\mathrm{aux}} = 0$ and all Gaussian primitives' parameters, excepting $e_k$ , are left invariant by this loss. The gradient with respect to $e_k$ instead yields + +$$ +\frac {\partial L _ {\pi} ^ {\mathrm {a u x}}}{\partial e _ {k}} = \sum_ {\boldsymbol {u} \in \operatorname {P i x}} \mathcal {E} _ {\pi} (\boldsymbol {u}) \omega_ {k} ^ {\pi} (\boldsymbol {u}) = E _ {k} ^ {\pi}, \tag {7} +$$ + +which is the per-Gaussian-primitive error for camera $\pi$ we wanted to compute. + +# 3.3 Opacity correction after cloning + +In the original ADC module, when a Gaussian primitive is split or cloned, the opacity value is preserved. This choice introduces a bias in the case of the clone operation by implicitly increasing the impact of the densified primitive on the final rendered color. To see why this is the case we can follow the example in Fig. 2, where we consider what happens if we render a splatted Gaussian in its center pixel assuming an opacity value $\alpha$ . Before a cloning operation happens, the rendered color depends on primitives that come next in the ordering with weight $1 - \alpha$ . But after we clone, due to the alpha-compositing logic, we have that primitives that come next weight $(1 - \alpha)^{2}$ , which is lower than $1 - \alpha$ for all opacity values in $(0,1)$ . Accordingly, by applying the standard logic of preserving the opacity after cloning we have a bias to weight more the cloned primitives. The solution we suggest consists in reducing the opacity of the primitives after cloning so that the bias is removed. The new opacity value $\hat{\alpha}$ can be found by solving the equation $(1 - \alpha) = (1 - \hat{\alpha})^{2}$ , which yields $\hat{\alpha} := 1 - \sqrt{1 - \alpha}$ . + +If we depart from the simplified setting of considering only the center pixel and rather consider all pixels, it is unfortunately not possible to remove completely the bias. Nonetheless, the correction factor we introduce reduces the bias + +![](images/552cacd468ecb90ada953eabc7d71a84878e5040e495d5cfeedd3a3e18f98f3c.jpg) +Fig. 2: Consider rendering a single splatted Gaussian in its center pixel with opacity $\alpha$ before and after cloning. Before we clone, the rendered color depends with weight $1 - \alpha$ on what comes next. After we clone, since we preserve the opacity, the rendered color depends with weight $(1 - \alpha)^{2}$ on what comes next. Since $(1 - \alpha) \geq (1 - \alpha)^{2}$ we have a bias towards weighting more Gaussian primitives that get cloned. The proposed correction changes the opacity post clone to $\hat{\alpha}$ so that the bias is removed. + +![](images/0e7914399d185eca8f4c05ee299107cafa9d7ef7b0741be5da1c520276602092.jpg) + +![](images/6f166a9d41021277a371fa7e88e62f4a89a03410b2de69a09502ca29d5a2c5d6.jpg) + +for all pixels compared to keeping the opacity of the cloned primitive. Indeed, the following relation holds for all $\alpha_{k}\in (0,1)$ and all pixels $\pmb{u}$ : + +$$ +1 - \alpha_ {k} \mathcal {G} _ {k} ^ {\pi} (\boldsymbol {u}) \geq \left(1 - \hat {\alpha} _ {k} \mathcal {G} _ {k} ^ {\pi} (\boldsymbol {u})\right) ^ {2} > \left(1 - \alpha_ {k} \mathcal {G} _ {k} ^ {\pi} (\boldsymbol {u})\right) ^ {2}, \tag {8} +$$ + +where $\hat{\alpha}_k \coloneqq 1 - \sqrt{1 - \alpha_k}$ is our corrected opacity. The proof of the relation follows by noting that $\hat{\alpha}_k$ can be rewritten as $\frac{\alpha_k}{1 + \sqrt{1 - \alpha_k}}$ which is strictly smaller than $\alpha_k$ for $\alpha_k \in (0,1)$ . + +Finally, the correction of the opacity as shown above is derived assuming we clone a Gaussian primitive, but does not strictly match the case of a split operation, for when we split we move the two offspring randomly away from the previous center and we change the covariance scale. For this reason, we stick to the standard rule of preserving the opacity of a primitive we split. + +# 3.4 Primitives growth control + +The ADC module grows a Gaussian primitive if $\tau_{k}$ is larger than a threshold. This mechanism can lead to unpredictable growth of the number of primitives, eventually resulting in out-of-memory issues. To avoid this problem, we introduce a global limit to the maximum number of Gaussian primitives, and a mechanism to control the maximum number of primitives that can be created each time densification is run. Among the many possible options, we explore a logic that limits new primitive offspring to a fixed fraction of the primitives that already exist. In case the number of primitives that are entitled to be densified exceeds the available budget, we retain only the ones that exhibit the highest densification score. An example of this process is shown in Fig. 3, compared to the one from 3DGS: For Ours, the number of primitives grows smoothly until it reaches the allotted maximum, without the discontinuities induced by opacity reset (see Sec. 3.5). The way we control the number of primitives is not limited to our error-based densification logic, but can be applied equally to the original gradient-based one. + +# 3.5 Alternative to opacity reset + +The strategy introduced in [9] to favour the sparsification of Gaussian primitives consists in periodical hard resets of the opacity for all primitives to a low + +![](images/3e11b4fbe66ba682b8f1acde909f2031ca4f58821e9935e0cef6f6452daf91cc.jpg) +Number of Gaussians by training iteration +Fig. 3: Evolution of the number of Gaussians in 3DGS, and in our method with upper limit set to the number reached by 3DGS (on the garden scene from the Mip-NeRF 360 dataset). Note that, while 3DGS' ADC process stops after $15\mathrm{k}$ iterations, ours remains active for $27\mathrm{k}$ . This is not immediately visible from the plot, since pruned primitives are immediately replaced by newly spawned ones, keeping the overall number stable once the maximum is reached. + +value, so that primitives whose opacity is not increased again by the optimization will eventually be pruned. This introduces a small shock in the training trajectory, which is suboptimal for the sake of having stable and predictable training dynamics. Moreover, resetting the opacity is particularly harmful for our error-based densification method, for it will lead to misleading error statistics right after the hard-reset, potentially triggering wrong densification decisions. For this reason, we propose a different logic to favour primitives pruning in a smoother way. Specifically, we decrease the opacity of each primitive by a fixed amount (we use 0.001) after each densification run, so that the opacity will gradually move towards the pruning range. In this way, we avoid sudden changes in the densification metric, while preserving the desired sparsification properties. + +One downside of the new opacity regularization logic is that the constant push towards lowering the opacity of the primitives implicitly invites the model to make more use of the background where possible. This is also harmful, for it could generate more holes in the scene that will be visible from novel views. To counteract this dynamics, we also regularize the residual probabilities of the alpha-compositing (a.k.a. residual transmittance) to be zero for every pixel, by simply minimizing their average value, weighted by a hyperparameter (here 0.1). + +# 4 Experimental Evaluation + +In the following we show how our improved ADC mechanism can equally be applied both to standard 3DGS [9] and its Mip-Splatting extension [29], providing benefits to both. + +# 4.1 Datasets and metrics + +We follow the experimental setup from the 3DGS [9] paper, focusing on the real-world scenes from the Mip-NeRF 360 [1], Tanks and Temples [10] and Deep Blending [6] datasets. Mip-NeRF 360 comprises nine scenes (5 outdoor, 4 indoor) captured in a circular pattern which focuses on a central area of a few meters, with a potentially unbounded background. For Tanks and Temples, we focus on the "Truck" and "Train" scenes, while for Deep Blending we focus on the "Dr Johnson" and "Playroom" scenes, using the images and SfM reconstructions shared by the Gaussian Splatting authors. In each experiment we set aside each 8th image as a validation set, and report peak signal-to-noise ratio (PSNR), structural similarity (SSIM) and the perceptual metric from [31] (LPIPS). + +Additional experiments on NeRF-Synthetic [17] are provided in the supplementary material. + +# 4.2 Experimental setup + +We evaluate based on our re-implementations of 3DGS, which allows us to easily switch between standard 3DGS, Mip-Splatting, the original ADC of [9], our contributions or any combination thereof. We reproduce the training settings proposed in [9,29] and the respective public code-bases $^{12}$ , including number of training iterations, batch size, input resolution, learning rates etc. When training with our contributions, we grow Gaussians with $E_{k} > 0.1$ , adding up to $5\%$ of the current number of primitives at each densification step. Differently from 3DGS, we keep our ADC process active for 27k iterations (i.e. $90\%$ of the training process), instead of stopping it after 15k. Other relevant hyper-parameters are left to the default values used in 3DGS, and shared across all datasets and scenes. In all our experiments, we set the maximum primitives budget to the number of primitives (or its median, for experiments with multiple runs) generated by the corresponding baseline, in order to obtain perfectly comparable models. + +A note about LPIPS. Investigating the 3DGS and Mip-Splating baselines, we discovered a discrepancy in the way LPIPS is calculated in both public code-bases, which resulted in under-estimated values being reported in the original papers. This was confirmed in private correspondence with the authors. In order to simplify comparisons with future works that don't rely on these code-bases, and might be unaware of this issue, we report correct LPIPS values here. + +A note about evaluations on MipNeRF360. The standard protocol for the dataset trains and evaluates the models at the same image scale, with indoor scenes down-sampled by a factor of two and outdoor scenes by four. The authors of the dataset provide down-scaled versions of the images. However, several recent works like [29] do not use the original images, which suffer from compression artifacts, but down-sample the images from the higher resolution counterparts. + +Table 1: Results on the Mip-NeRF 360 dataset. Top section of the table: results from the Gaussian Splitting paper; bottom section: results from our re-implementation averaged over 5 runs. + +
MethodPSNR ↑SSIM ↑LPIPS ↓
MeanStdevMeanStdevMeanStdev
Plenoxels [4]23.08-0.626-0.436-
INGP [18]25.59-0.699-0.331-
Mip-NeRF 360 [1]27.69-0.792-0.237-
3DGS [9]27.450.0810.8170.0010.2500.001
Mip-Splatting [29]27.510.0740.8180.0010.2510.001
Ours, 3DGS27.610.0670.8220.0010.2230.001
Ours, Mip-Splatting27.700.0650.8230.0010.2230.001
+ +Table 2: Results on the Tanks and Temples dataset. Top section of the table: results from the Gaussian Splitting paper; bottom section: results from our re-implementation averaged over 5 runs. + +
MethodPSNR ↑SSIM ↑LPIPS ↓
MeanStdevMeanStdevMeanStdev
Plenoxels [4]21.08-0.719-0.379-
INGP [18]21.92-0.745-0.305-
Mip-NeRF 360 [1]22.22-0.759-0.257-
3DGS [9]23.440.0890.8450.0020.2120.001
Mip-Splatting [29]23.650.0660.8490.0010.2110.001
Ours, 3DGS23.930.0700.8530.0010.1870.001
Ours, Mip-Splatting24.100.1090.8570.0020.1830.001
+ +As a result, the scores they report are significantly higher. We use the downscaled images from the dataset for training/evaluation to stay consistent with older NeRF works and report scores of competitors under the same setting. + +# 4.3 Main results + +In a first set of experiments, we evaluate the effectiveness of our improved ADC strategy (Ours) when applied to 3DGS and Mip-Splatting. Results, collected over 5 training runs to average out the randomness induced by stochastic primitive splitting, are reported in Tab. 1, 2 and 3. For the sake of completeness, we also include scores obtained with three NeRF baselines, i.e. Plenoxels [4], InstantNGP (INGP) [18] and Mip-NeRF 360 [1], as originally reported in [9]. Our approach consistently outperforms the corresponding baselines (i.e. Ours, 3DGS vs. 3DGS; Ours, Mip-Splatting vs. Mip-Splatting), particularly on SSIM and LPIPS. This is in line with what we discussed in Sec. 3.1 and 3.2: standard ADC often leads to localized under-fitting, as it fails to split large gaussians + +![](images/858307533f28b0a273f85513a87d74920db2378a348ac62ef90c84b155f0b6e0.jpg) + +![](images/ce614e38d782800e44c355e081982514ac07258efeba0e213ab19abea58ded83.jpg) + +![](images/21409c2181685c0f81b46d78b8c09c6d84087863a5f7ab6f4e3e4cb9656fc2e4.jpg) + +![](images/3770316d0329bcaece8805ac389529ae426c2c9e8759912a6de143cad65e0a9f.jpg) + +![](images/2153b1b9400cd93249e1f2dc103206d8290523984cff59ac5d67df104a227b3b.jpg) + +![](images/b6efd124893b2007e6a08a83a91e6654d01db4332f0b786f514bf6557ee3301e.jpg) + +![](images/bb3add9c6dc83d1e7bacc8eedf7e39f62643ca5b9876469c14f82b301c5c7ca3.jpg) + +![](images/8c4abd8c09beefacf463fa8991016ff44b63dcf084a31bbe9cb25ce081686548.jpg) + +![](images/d02658d9ad8b4c5a431e5239d482520cb2647dc02dac41beaa164ea3d5a769c1.jpg) + +![](images/815814d264d5563d9098d5391722c69d4c73f601d77b044d731d19c9898c9c0c.jpg) + +![](images/0ceaccbae16322cd5b24a0d91ff0fcaee5bf3ad40088ee1e2d986a284161a0a0.jpg) + +![](images/195593b97f31df17f4d393da409b166956ce968d495b03349727ab9fbcbf7689.jpg) + +![](images/b25904c7913bab6dbf6ae598b55fa623e56ef1bcfa3feeec360cc593aee43395.jpg) + +![](images/d8062f087f92f6162452b56cda07b84884c721551fb7050d8963e18cc5a7833b.jpg) + +![](images/e73b6a6de4f855f05c46a39c2048cb36c6152d5cfe6ea1c509c7821c5459e555.jpg) + +![](images/d303d692b14d6a1edca90b44b324c70ebd5afe5a60a338dd4a7ec270e6d4f1a3.jpg) + +![](images/eb17d26e4871dd0b1e7aaaa420270651d57fda735efcabd4c5a3400fb62419bc.jpg) + +![](images/751134d20b393733d0ba796c5734f3579fd460728cccac4b17eae3148eefa104.jpg) + +![](images/35c59216cbc1fb86523647ce1f2dcceafa21928fd7721278b8429398c8b1b56c.jpg) + +![](images/3893412dd7c444e7017fbf694484fef839a9ce5c1c2886fe92484ebd1ca7121e.jpg) +Fig. 4: Qualitative results on the Mip-NeRF 360 and Tanks and Temples validation sets. Both 3DGS and Ours use the same number of primitives. Best viewed on screen at high magnification. + +![](images/78c0e9458d6af011f9f984c0ff83cdcc93f633ae5697fb0ed67daf68ff3ff713.jpg) + +![](images/b055c6bbd040714ad14f987e807c1607a5fee66d305d2e0c52ff6f4f37c3a47a.jpg) + +![](images/60cf8f0109a37415d4361fc5d43ed399d5dc85ec98bd5c563bc0bf1470e675f6.jpg) + +![](images/4ec7a106b865c0af9fae1265483ad5f8195d5b1c0335a173ced0447833bac32d.jpg) + +Table 3: Results on the Deep Blending dataset. Top section of the table: results from the Gaussian Splatting paper; bottom section: results from our re-implementation averaged over 5 runs. + +
MethodPSNR ↑SSIM ↑LPIPS ↓
MeanStdevMeanStdevMeanStdev
Plenoxels [4]23.06-0.795-0.510-
INGP [18]24.96-0.817-0.390-
Mip-NeRF 360 [1]29.40-0.901-0.245-
3DGS [9]29.540.0960.9020.0000.3110.001
Mip-Splatting [29]29.680.0680.9030.0000.3090.000
Ours, 3DGS29.500.1100.9040.0010.3050.001
Ours, Mip-Splatting29.640.1260.9050.0010.3030.001
+ +Table 4: Ablation experiments on the Mip-NeRF 360 dataset, adding individual contributions to 3DGS or removing them from Ours. OC: Opacity Correction, Sec. 3.3; GC: Growth Control, Sec. 3.4; OR: Opacity Regularization, Sec. 3.5. + +
Baseline3DGS+OROurs
+OC+GC+ORFull-OC-GC-OR
PSNR ↑27.4527.6527.3527.4827.6127.0427.5427.28
SSIM ↑0.8170.8220.8100.8180.8220.8120.8180.810
LPIPS ↓0.2500.2390.2560.2430.2230.2350.2230.234
+ +that cover highly-textured regions of the scene. This kind of errors is poorly reflected by PSNR, which measures the "average fit" over image pixels, but are promptly detected by perceptual metrics like LPIPS. On the Deep Blending dataset, we observe thinner gaps, PSNR actually showing a small regression w.r.t. the baselines, although with low confidence margins. We suspect this might be related to the fact that Deep Blending contains many flat, untextured surfaces (see Fig. 4), that are particularly challenging to reconstruct accurately with 3DGS-like methods, independently of the ADC strategy being adopted. In the supplementary material, we provide an additional discussion about why some metrics do not reflect the actual perceived quality of the generated images. + +Figure 4 contains a qualitative comparison between standard 3DGS and 3DGS augmented with our contributions (Ours). Areas with under-fitting artifacts are highlighted, showing how these are notably ameliorated by our approach. It is also worth noting that Ours effectively maintains the same quality as 3DGS in non-problematic areas, producing a more perceptually accurate reconstruction while using the same number of primitives (see Sec. 4.2). + +# 4.4 Ablation experiments + +In Tab. 4 we ablate the effects of Opacity Correction (OC, Sec. 3.3), Growth Control (GC, Sec. 3.4) and Opacity Regularization (OR, Sec. 3.5) on the Mip-NeRF + +360 dataset. In particular, we evaluate 3DGS augmented with each of these components (left side of the table), and our method with the components replaced by the corresponding baseline mechanism in 3DGS' standard ADC (right side of the table). First, we observe that OC, GC and OR all contribute to our method, as the Full version of Ours achieves the overall best results on all metrics, and removing them consistently degrades performance. Interestingly, Opacity Correction seems to have the largest impact here, as it produces both the largest increase in the scores when added to 3DGS, and the largest decrease when removed from Ours. Finally, Growth Control has a negative impact on 3DGS when utilized in isolation, while only slightly degrading the results when removed from Ours. Note that this observation doesn't detract from GC's usefulness as a strategy to control and limit the capacity of the model. We hypothesize that GC's negative effect on 3DGS might be a consequence of the fact that the standard, gradient-based densification score is actually a poor choice for comparing gaussians in terms of how soon they should be split or cloned (remember that GC ranks Gaussians based on their score). + +# 4.5 Limitations + +While our method appears to be quite effective at solving under-fitting issues, these can still be present in especially difficult scenes (e.g. treehill in the Mip-NeRF 360 dataset, both scenes from the Deep Blending dataset). Focusing on the problematic areas that our ADC approach handles successfully, we observe that, while perceptually more "correct", the reconstruction there can still be quite inaccurate when closely compared to the ground truth (see e.g. the flowers scene in Fig. 4). We suspect both these issues might be related to 3DGS' intrinsic limits in handling i) strong view-dependent effects; ii) appearance variations across images; and iii) errors induced by the linear approximation in the Splitting operation (see Sec. 2). An interesting future direction could be to combine our approach with works that address these issues, e.g. Spec-Gaussian [26] for (i) and GS++ [7] for (iii). + +# 5 Conclusion + +In this paper, we addressed the limitations of the Adaptive Density Control (ADC) mechanism in 3D Gaussian Splatting (3DGS), a scene representation method for high-quality, photorealistic rendering. Our main contribution is a more principled, pixel-error driven formulation for density control in 3DGS. We propose how to leverage a novel decision criterion for densification based on per-pixel errors and introduce a mechanism to control the total number of primitives generated per scene. We also correct a bias in the current opacity handling in ADC during cloning. Our approach leads to consistent and systematic improvements over previous methods, particularly in perceptual metrics like LPIPS. + +# References + +1. Barron, J.T., Mildenhall, B., Verbin, D., Srinivasan, P.P., Hedman, P.: Mipnerf 360: Unbounded anti-aliased neural radiance fields. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5470-5479 (2022) +2. Chen, Z., Wang, F., Liu, H.: Text-to-3d using gaussian splatting. arXiv preprint arXiv:2309.16585 (2023) +3. Cheng, K., Long, X., Yang, K., Yao, Y., Yin, W., Ma, Y., Wang, W., Chen, X.: Gaussianpro: 3d gaussian splatting with progressive propagation. arXiv preprint arXiv:2402.14650 (2024) +4. Fridovich-Keil, S., Yu, A., Tancik, M., Chen, Q., Recht, B., Kanazawa, A.: Plenoxels: Radiance fields without neural networks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5501-5510 (2022) +5. Guédon, A., Lepetit, V.: Sugar: Surface-aligned gaussian splatting for efficient 3d mesh reconstruction and high-quality mesh rendering. arXiv preprint arXiv:2311.12775 (2023) +6. Hedman, P., Philip, J., Price, T., Frahm, J.M., Drettakis, G., Brostow, G.: Deep blending for free-viewpoint image-based rendering. ACM Transactions on Graphics (ToG) 37(6), 1-15 (2018) +7. Huang, L., Bai, J., Guo, J., Guo, Y.: $\mathrm{Gs} + +$ : Error analyzing and optimal gaussian splatting. arXiv preprint arXiv:2402.00752 (2024) +8. Keetha, N., Karhade, J., Jatavallabhula, K.M., Yang, G., Scherer, S., Ramanan, D., Luiten, J.: Splatam: Splat, track & map 3d gaussians for dense rgb-d slam. arXiv preprint arXiv:2312.02126 (2023) +9. Kerbl, B., Kopanas, G., Leimkuhler, T., Drettakis, G.: 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics 42(4) (July 2023), https://repo-sam.inria.fr/fungraph/3d-gaussian-splatting/ +0. Knapitsch, A., Park, J., Zhou, Q.Y., Koltun, V.: Tanks and temples: Benchmarking large-scale scene reconstruction. ACM Transactions on Graphics (ToG) 36(4), 1-13 (2017) +1. Kocabas, M., Chang, J.H.R., Gabriel, J., Tuzel, O., Ranjan, A.: Hugs: Human gaussian splats. arXiv preprint arXiv:2311.17910 (2023) +2. Lee, J.C., Rho, D., Sun, X., Ko, J.H., Park, E.: Compact 3d gaussian representation for radiance field. arXiv preprint arXiv:2311.13681 (2023) +3. Lei, J., Wang, Y., Pavlakos, G., Liu, L., Daniilidis, K.: Gart: Gaussian articulated template models. arXiv preprint arXiv:2311.16099 (2023) +4. Lu, T., Yu, M., Xu, L., Xiangli, Y., Wang, L., Lin, D., Dai, B.: Scaffold-gs: Structured 3d gaussians for view-adaptive rendering. arXiv preprint arXiv:2312.00109 (2023) +5. Luiten, J., Kopanas, G., Leibe, B., Ramanan, D.: Dynamic 3d gaussians: Tracking by persistent dynamic view synthesis. arXiv preprint arXiv:2308.09713 (2023) +6. Matsuki, H., Murai, R., Kelly, P.H., Davison, A.J.: Gaussian splatting slam. arXiv preprint arXiv:2312.06741 (2023) +7. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. In: ECCV. p. 405-421 (2020) +8. Müller, T., Evans, A., Schied, C., Keller, A.: Instant neural graphics primitives with a multiresolution hash encoding. ACM Trans. Graph. 41(4), 102:1-102:15 (Jul 2022). https://doi.org/10.1145/3528223.3530127, https://doi.org/10.1145/3528223.3530127 + +19. Saito, S., Schwartz, G., Simon, T., Li, J., Nam, G.: Relightable gaussian codec avatars. arXiv preprint arXiv:2312.03704 (2023) +20. Snavely, N., Seitz, S.M., Szeliski, R.: Photo tourism: exploring photo collections in 3d. In: ACM siggraph 2006 papers, pp. 835-846 (2006) +21. Tang, J., Ren, J., Zhou, H., Liu, Z., Zeng, G.: Dreamgaussian: Generative gaussian splatting for efficient 3d content creation. arXiv preprint arXiv:2309.16653 (2023) +22. Wu, G., Yi, T., Fang, J., Xie, L., Zhang, X., Wei, W., Liu, W., Tian, Q., Wang, X.: 4d gaussian splatting for real-time dynamic scene rendering. arXiv preprint arXiv:2310.08528 (2023) +23. Xie, T., Zong, Z., Qiu, Y., Li, X., Feng, Y., Yang, Y., Jiang, C.: Physgaus-sian: Physics-integrated 3d gaussians for generative dynamics. arXiv preprint arXiv:2311.12198 (2023) +24. Yan, C., Qu, D., Wang, D., Xu, D., Wang, Z., Zhao, B., Li, X.: Gs-slam: Dense visual slam with 3d gaussian splatting. arXiv preprint arXiv:2311.11700 (2023) +25. Yang, Z., Yang, H., Pan, Z., Zhu, X., Zhang, L.: Real-time photorealistic dynamic scene representation and rendering with 4d gaussian splatting. arXiv preprint arXiv:2310.10642 (2023) +26. Yang, Z., Gao, X., Sun, Y., Huang, Y., Lyu, X., Zhou, W., Jiao, S., Qi, X., Jin, X.: Spec-gaussian: Anisotropic view-dependent appearance for 3d gaussian splatting. arXiv preprint arXiv:2402.15870 (2024) +27. Ye, M., Danelljan, M., Yu, F., Ke, L.: Gaussian grouping: Segment and edit anything in 3d scenes. arXiv preprint arXiv:2312.00732 (2023) +28. Yi, T., Fang, J., Wu, G., Xie, L., Zhang, X., Liu, W., Tian, Q., Wang, X.: Gaussian dreamer: Fast generation from text to 3d gaussian splatting with point cloud priors. arXiv preprint arXiv:2310.08529 (2023) +29. Yu, Z., Chen, A., Huang, B., Sattler, T., Geiger, A.: Mip-splatting: Alias-free 3d gaussian splatting. arXiv preprint arXiv:2311.16493 (2023) +30. Yugay, V., Li, Y., Gevers, T., Oswald, M.R.: Gaussian-slam: Photo-realistic dense slam with gaussian splatting. arXiv preprint arXiv:2312.10070 (2023) +31. Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual metric. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 586-595 (2018) +32. Zielonka, W., Bagautdinov, T., Saito, S., Zollhöfer, M., Thies, J., Romero, J.: Drivable 3d gaussian avatars. arXiv preprint arXiv:2311.08581 (2023) +33. Zwicker, M., Pfister, H., van Baar, J., Gross, M.: Ewa volume splatting. In: Proceedings Visualization, 2001. VIS '01. pp. 29-538 (2001). https://doi.org/10.1109/VISUAL.2001.964490 \ No newline at end of file diff --git a/2024/Revising Densification in Gaussian Splatting/images.zip b/2024/Revising Densification in Gaussian Splatting/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..35b6a8bdc9b60884aa1fc379535191da45ef2954 --- /dev/null +++ b/2024/Revising Densification in Gaussian Splatting/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f3b97e1a4583b617fba0a7a5703623637fb1e04ea4cea24b95bd97175c4ac168 +size 536085 diff --git a/2024/Revising Densification in Gaussian Splatting/layout.json b/2024/Revising Densification in Gaussian Splatting/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..8f91dfc38b8163e5c5d2ce51849c7478ae6dd0d8 --- /dev/null +++ b/2024/Revising Densification in Gaussian Splatting/layout.json @@ -0,0 +1,9422 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 151, + 112, + 462, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 112, + 462, + 129 + ], + "spans": [ + { + "bbox": [ + 151, + 112, + 462, + 129 + ], + "type": "text", + "content": "Revising Densification in Gaussian Splitting" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 162, + 150, + 451, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 162, + 150, + 451, + 163 + ], + "spans": [ + { + "bbox": [ + 162, + 150, + 451, + 163 + ], + "type": "text", + "content": "Samuel Rota Bulò, Lorenzo Porzi, and Peter Kontschieder" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 252, + 172, + 358, + 183 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 252, + 172, + 358, + 183 + ], + "spans": [ + { + "bbox": [ + 252, + 172, + 358, + 183 + ], + "type": "text", + "content": "Meta Reality Labs Zurich" + } + ] + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 133, + 203, + 249, + 281 + ], + "blocks": [ + { + "bbox": [ + 133, + 203, + 249, + 281 + ], + "lines": [ + { + "bbox": [ + 133, + 203, + 249, + 281 + ], + "spans": [ + { + "bbox": [ + 133, + 203, + 249, + 281 + ], + "type": "image", + "image_path": "01da4aa633a1585e8b5f170e748342671bc84b2736e2a61b43c97120485ef712.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 289, + 482, + 346 + ], + "lines": [ + { + "bbox": [ + 130, + 289, + 482, + 346 + ], + "spans": [ + { + "bbox": [ + 130, + 289, + 482, + 346 + ], + "type": "text", + "content": "Fig. 1: Densification is a critical component of 3D Gaussian Splatting (3DGS), and a common failure point. In this example (ground truth on the left) we show how 3DGS can fail (center) to add primitives to high-texture areas, like the grass in the bottom part of the pictures, producing large and blurry artifacts. Our approach (right) solves this issue by comprehensively revising densification in 3DGS." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 249, + 203, + 365, + 281 + ], + "blocks": [ + { + "bbox": [ + 249, + 203, + 365, + 281 + ], + "lines": [ + { + "bbox": [ + 249, + 203, + 365, + 281 + ], + "spans": [ + { + "bbox": [ + 249, + 203, + 365, + 281 + ], + "type": "image", + "image_path": "0aa0650b8e46425032dbe6a5487b289f066b1aa6f81b3dbb18502e95b264f3cb.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 365, + 203, + 481, + 281 + ], + "blocks": [ + { + "bbox": [ + 365, + 203, + 481, + 281 + ], + "lines": [ + { + "bbox": [ + 365, + 203, + 481, + 281 + ], + "spans": [ + { + "bbox": [ + 365, + 203, + 481, + 281 + ], + "type": "image", + "image_path": "3d4d96cac9f0c8b618afe30ec5fa20aa673d537d5373f7e503097eb836cffc54.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 159, + 361, + 453, + 503 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 361, + 453, + 503 + ], + "spans": [ + { + "bbox": [ + 159, + 361, + 453, + 503 + ], + "type": "text", + "content": "Abstract. In this paper, we address the limitations of Adaptive Density Control (ADC) in 3D Gaussian Splatting (3DGS), a scene representation method achieving high-quality, photorealistic results for novel view synthesis. ADC has been introduced for automatic 3D point primitive management, controlling densification and pruning, however, with certain limitations in the densification logic. Our main contribution is a more principled, pixel-error driven formulation for density control in 3DGS, leveraging an auxiliary, per-pixel error function as the criterion for densification. We further introduce a mechanism to control the total number of primitives generated per scene and correct a bias in the current opacity handling strategy of ADC during cloning operations. Our approach leads to consistent quality improvements across a variety of benchmark scenes, without sacrificing the method's efficiency." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 159, + 514, + 453, + 535 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 514, + 453, + 535 + ], + "spans": [ + { + "bbox": [ + 159, + 514, + 453, + 535 + ], + "type": "text", + "content": "Keywords: Gaussian Splatting " + }, + { + "bbox": [ + 159, + 514, + 453, + 535 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 514, + 453, + 535 + ], + "type": "text", + "content": " 3D reconstruction " + }, + { + "bbox": [ + 159, + 514, + 453, + 535 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 159, + 514, + 453, + 535 + ], + "type": "text", + "content": " Novel View Synthesis" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 556, + 230, + 569 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 556, + 230, + 569 + ], + "spans": [ + { + "bbox": [ + 132, + 556, + 230, + 569 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 581, + 482, + 666 + ], + "type": "text", + "content": "High-quality, photorealistic scene modelling from images has been an important research area in computer vision and graphics, with plentiful applications in AR/VR/MR, robotics, etc. In the last years, this field has gained a lot of attention due to advances in Neural 3D scene representations, particularly Neural Radiance Fields (NeRFs) [17]. NeRFs take a new approach to 3D scene representation and rendering, by leveraging a combination of deep learning and volumetric rendering techniques for generating photorealistic images from novel" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 236 + ], + "type": "text", + "content": "viewpoints. By optimizing MLPs to map from spatial coordinates and viewing directions to density and colour fields, these models have demonstrated astonishing capabilities for capturing the complex interplay of light and geometry in a data-driven way. While highly efficient in terms of representation quality, the original NeRF representation relies on time-consuming sampling strategies and thus excludes applications with fast rendering requirements. With many advances in terms of the underlying representation, these models have been significantly optimized towards improved training time and scene representation fidelity. However, inference speed for high-resolution, novel view synthesis remains an ongoing limitation." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 239, + 483, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 239, + 483, + 335 + ], + "spans": [ + { + "bbox": [ + 130, + 239, + 483, + 335 + ], + "type": "text", + "content": "More recently, 3D Gaussian Splatting (3DGS) [9] has been proposed as an alternative and expressive scene representation, enabling both high-speed, high-fidelity training of models and high-resolution, GPU rasterization-friendly rendering of novel views. Their core representation is an optimized set of (anisotropic) 3D Gaussians, after being randomly distributed in 3D space, or systematically initialized at points obtained by Structure-from-Motion [20]. For obtaining a 2D image, all relevant 3D primitives are efficiently rendered via splatting-based rasterization with low-pass filtering." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 338, + 482, + 614 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 338, + 482, + 614 + ], + "spans": [ + { + "bbox": [ + 130, + 338, + 482, + 614 + ], + "type": "text", + "content": "In 3DGS, each 3D primitive is parameterized as a 3D Gaussian distribution (i.e., with position and covariance), together with parameters controlling its opacity and describing its directional appearance (typically spherical harmonics). The parameter optimization procedure is guided by a multi-view, photometric loss, and is interleaved with Adaptive Density Control (ADC), a mechanism controlling density management for 3D points by means of introducing or deleting 3D primitives. ADC plays a critical role as it determines where to expand/shrink the scene representation budget for empty or over-reconstructed regions, respectively. Both growing and pruning operations are activated based on user-defined thresholds: Growing depends on the accumulated positional gradients of existing primitives and is, conditioned on the size of the Gaussians, executed by either splitting large primitives or by cloning smaller ones. Pruning is activated once the opacity falls below a provided threshold. While quite effective in practice, such density management strategies have several limitations. First, estimating a gradient magnitude-based threshold is rather non-intuitive and not robust to potential changes in the model, loss terms, etc. Second, there are cases where only few and large Gaussians are modeling high-frequency patterns like grass as shown in the middle of Fig. 1. Here, changes accumulated from positional gradients might remain very low and thus fail to trigger the densification mechanism, which in turn leads to substantial scene underfitting. Finally, ADC lacks explicit control of the maximum number of Gaussians generated per scene. This has important, practical implications as uncontrolled growth might easily lead to out-of-memory errors during training." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "In this work we address the shortcomings of Adaptive Density Control proposed in the original 3D Gaussian splatting method. Our core contribution is a more principled, pixel-error driven formulation for density control in 3DGS. We describe how 2D, per-pixel errors as e.g. derived from Structural Similarity" + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 233, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 233, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 233, + 100 + ], + "type": "text", + "content": "Rota Bulò et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 379 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 481, + 379 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 481, + 379 + ], + "type": "text", + "content": "(or any other informative objective function) can be propagated back as errors to contributing Gaussian primitives. In our solution, we first break down the per-pixel errors according to each Gaussian's contribution, and in a camera-specific way. This allows us to track the maximum error per primitive for all views and across two subsequent ADC runs, yielding our novel, error-specific, and thus more intuitive decision criterion for densification. Our second contribution is correcting a bias introduced with the current form of opacity handling in ADC when conducting a primitive cloning operation. The original approach suggests to keep the same opacity for the cloned Gaussian, which however biases the alpha-compositing logic applied for rendering the pixel colors. Indeed, this procedure leads to an overall increase of opacity in the cloned region, preventing the model to correctly account for contributions of other primitives and thus negatively affecting the densification process. Our third contribution is a mechanism for controlling the total number of primitives generated per scene and the maximum amount of novel primitives introduced per densification run. With this functionality, we can avoid undesired out-of-memory errors and better tune the method's behaviour w.r.t. given hardware constraints. We extensively validate our contributions on standard benchmark datasets like Mip-NeRF 360 [1], Tanks and Temples [10], and Deep Blending [6]. Our experiments show consistent improvements over different baselines including 3DGS [9] and Mip-Splatting [29]. To summarize, our contributions are improving methodological shortcomings in 3DGS' Adaptive Density Control mechanism as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 384, + 480, + 465 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 138, + 384, + 480, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 384, + 480, + 418 + ], + "spans": [ + { + "bbox": [ + 138, + 384, + 480, + 418 + ], + "type": "text", + "content": "- We propose a principled approach that enables the guidance of the densification process according to an auxiliary, per-pixel error function, rather than relying on positional gradients." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 419, + 480, + 442 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 419, + 480, + 442 + ], + "spans": [ + { + "bbox": [ + 138, + 419, + 480, + 442 + ], + "type": "text", + "content": "- We correct an existing, systematic bias from the primitive growing procedure when cloning Gaussians, negatively impacting the overall densification." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 442, + 480, + 465 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 442, + 480, + 465 + ], + "spans": [ + { + "bbox": [ + 138, + 442, + 480, + 465 + ], + "type": "text", + "content": "- We present ablations and experimental evaluations on different, real-world benchmarks, confirming quantitative and qualitative improvements." + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 133, + 481, + 232, + 491 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 481, + 232, + 491 + ], + "spans": [ + { + "bbox": [ + 133, + 481, + 232, + 491 + ], + "type": "text", + "content": "1.1 Related works" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 133, + 498, + 481, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 498, + 481, + 569 + ], + "spans": [ + { + "bbox": [ + 133, + 498, + 481, + 569 + ], + "type": "text", + "content": "Since it was presented in [9], 3DGS has been used in a remarkably wide set of downstream applications, including Simultaneous Localization and Mapping [8, 16,24,30], text-to-3D generation [2,21,28], photo-realistic human avatars [11,13, 19,32], dynamic scene modeling [15,22,25] and more [5,23,27]. However, only a handful of works like ours have focused on advancing 3DGS itself, by improving its quality or overcoming some of its limitations." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 133, + 570, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 570, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 570, + 481, + 665 + ], + "type": "text", + "content": "In " + }, + { + "bbox": [ + 133, + 570, + 481, + 665 + ], + "type": "inline_equation", + "content": "\\mathrm{GS} + +" + }, + { + "bbox": [ + 133, + 570, + 481, + 665 + ], + "type": "text", + "content": " [7], Huang et al. present an improved approximation of the 3D-to-2D splatting operation at the core of 3DGS, which achieves better accuracy near image edges and solves some common visual artifacts. Spec-Gaussian [26] and Scaffold-gs [14] focus on improving view-dependent appearance modeling: the former by replacing spherical harmonics with an anisotropic spherical Gaussian appearance field; the latter by making all 3D Gaussian parameters, including whether specific primitives should be rendered or not, dependent on view direction through a small MLP. Mip-Splatting [29] tackles the strong artifacts that" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 266, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 266, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 266, + 91, + 447, + 102 + ], + "type": "text", + "content": "Revising Densification in Gaussian Splitting" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 199 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 199 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 199 + ], + "type": "text", + "content": "appear in 3DGS models when they are rendered at widely different resolutions (or viewing distances) compared to the images they were trained on. To do this, Yu et al. propose to incorporate a 3D filter to constrain the size of the 3D primitives depending on their maximal sampling rate on the training views, and a 2D Mip filter to mitigate aliasing issues. All these works adopt the original ADC strategy proposed in [9], and can potentially benefit from our improved approach, as we show for Mip-Splatting in Sec. 4." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 200, + 480, + 355 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 200, + 480, + 355 + ], + "spans": [ + { + "bbox": [ + 130, + 200, + 480, + 355 + ], + "type": "text", + "content": "Only few concurrent works have touched on densification, while putting most of their focus on other aspects of 3DGS. Lee et al. [12] propose a quantization-based approach to produce more compact 3DGS representations, which includes a continuous sparsification strategy that takes both primitive size and opacity into account. GaussianPro [3] directly tackles related densification limitations as we explore in our work, filling the gaps from SfM-based initialization. They propose a rather complex procedure based on the progressive propagation of primitives along estimated planes, using patch-matching and geometric consistency as guidance. In contrast to our method, [3] focuses on fixing the quality of planar regions, instead of holistically improving densification. We also note that a fair comparison with their method on the standard Mip-NeRF 360 benchmark is not feasible at the time of submission, as the authors did not publicly share the improved SfM point cloud used in their experiments (see §5.2 of [3])." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 372, + 353, + 387 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 372, + 353, + 387 + ], + "spans": [ + { + "bbox": [ + 132, + 372, + 353, + 387 + ], + "type": "text", + "content": "2 Preliminaries: Gaussian Splitting" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 396, + 479, + 433 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 396, + 479, + 433 + ], + "spans": [ + { + "bbox": [ + 130, + 396, + 479, + 433 + ], + "type": "text", + "content": "Gaussian Splatting [9] revisits ideas from EWA splatting [33] and proposes to fit a 3D scene as a collection of 3D Gaussian primitives " + }, + { + "bbox": [ + 130, + 396, + 479, + 433 + ], + "type": "inline_equation", + "content": "\\Gamma \\coloneqq \\{\\gamma_1,\\dots ,\\gamma_K\\}" + }, + { + "bbox": [ + 130, + 396, + 479, + 433 + ], + "type": "text", + "content": " that can be rendered by leveraging volume splatting." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 437, + 480, + 461 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 437, + 480, + 461 + ], + "spans": [ + { + "bbox": [ + 131, + 437, + 480, + 461 + ], + "type": "text", + "content": "Gaussian primitive. A Gaussian primitive " + }, + { + "bbox": [ + 131, + 437, + 480, + 461 + ], + "type": "inline_equation", + "content": "\\gamma_{k} := (\\pmb{\\mu}_{k}, \\pmb{\\Sigma}_{k}, \\alpha_{k}, \\pmb{f}_{k})" + }, + { + "bbox": [ + 131, + 437, + 480, + 461 + ], + "type": "text", + "content": " geometrically resembles a 3D Gaussian kernel" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 212, + 466, + 480, + 492 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 212, + 466, + 480, + 492 + ], + "spans": [ + { + "bbox": [ + 212, + 466, + 480, + 492 + ], + "type": "interline_equation", + "content": "\\mathcal {G} _ {k} (\\boldsymbol {x}) := \\exp \\left(- \\frac {1}{2} \\left(\\boldsymbol {x} - \\boldsymbol {\\mu} _ {k}\\right) ^ {\\top} \\Sigma_ {k} ^ {- 1} \\left(\\boldsymbol {x} - \\boldsymbol {\\mu} _ {k}\\right)\\right) \\tag {1}", + "image_path": "ca82cc0300cd2e9ebe9b570d470f20a2373525cb879bfa307998ace565a6ed3d.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 499, + 479, + 537 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 499, + 479, + 537 + ], + "spans": [ + { + "bbox": [ + 131, + 499, + 479, + 537 + ], + "type": "text", + "content": "centered in " + }, + { + "bbox": [ + 131, + 499, + 479, + 537 + ], + "type": "inline_equation", + "content": "\\pmb{\\mu}_k\\in \\mathbb{R}^3" + }, + { + "bbox": [ + 131, + 499, + 479, + 537 + ], + "type": "text", + "content": " and having " + }, + { + "bbox": [ + 131, + 499, + 479, + 537 + ], + "type": "inline_equation", + "content": "\\Sigma_{k}" + }, + { + "bbox": [ + 131, + 499, + 479, + 537 + ], + "type": "text", + "content": " as its " + }, + { + "bbox": [ + 131, + 499, + 479, + 537 + ], + "type": "inline_equation", + "content": "3\\times 3" + }, + { + "bbox": [ + 131, + 499, + 479, + 537 + ], + "type": "text", + "content": " covariance matrix. Each primitive additionally entails an opacity factor " + }, + { + "bbox": [ + 131, + 499, + 479, + 537 + ], + "type": "inline_equation", + "content": "\\alpha_{k}\\in [0,1]" + }, + { + "bbox": [ + 131, + 499, + 479, + 537 + ], + "type": "text", + "content": " and a feature vector " + }, + { + "bbox": [ + 131, + 499, + 479, + 537 + ], + "type": "inline_equation", + "content": "\\pmb {f}_k\\in \\mathbb{R}^d" + }, + { + "bbox": [ + 131, + 499, + 479, + 537 + ], + "type": "text", + "content": " (e.g. RGB color or spherical harmonics coefficients)." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 540, + 480, + 614 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 540, + 480, + 614 + ], + "spans": [ + { + "bbox": [ + 131, + 540, + 480, + 614 + ], + "type": "text", + "content": "Splatting. This is the operation of projecting a Gaussian primitive " + }, + { + "bbox": [ + 131, + 540, + 480, + 614 + ], + "type": "inline_equation", + "content": "\\gamma_{k}" + }, + { + "bbox": [ + 131, + 540, + 480, + 614 + ], + "type": "text", + "content": " to a camera pixel space via its world-to-image transformation " + }, + { + "bbox": [ + 131, + 540, + 480, + 614 + ], + "type": "inline_equation", + "content": "\\pi : \\mathbb{R}^3 \\to \\mathbb{R}^2" + }, + { + "bbox": [ + 131, + 540, + 480, + 614 + ], + "type": "text", + "content": ", which we refer directly to as the camera for simplicity. The projection " + }, + { + "bbox": [ + 131, + 540, + 480, + 614 + ], + "type": "inline_equation", + "content": "\\pi" + }, + { + "bbox": [ + 131, + 540, + 480, + 614 + ], + "type": "text", + "content": " is approximated to the first order at the primitive's center " + }, + { + "bbox": [ + 131, + 540, + 480, + 614 + ], + "type": "inline_equation", + "content": "\\pmb{\\mu}_{k}" + }, + { + "bbox": [ + 131, + 540, + 480, + 614 + ], + "type": "text", + "content": " so that the projected primitive is geometrically equivalent to a 2D Gaussian kernel " + }, + { + "bbox": [ + 131, + 540, + 480, + 614 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_k^\\pi" + }, + { + "bbox": [ + 131, + 540, + 480, + 614 + ], + "type": "text", + "content": " with mean " + }, + { + "bbox": [ + 131, + 540, + 480, + 614 + ], + "type": "inline_equation", + "content": "\\pi (\\pmb{\\mu}_k)\\in \\mathbb{R}^2" + }, + { + "bbox": [ + 131, + 540, + 480, + 614 + ], + "type": "text", + "content": " and 2D covariance " + }, + { + "bbox": [ + 131, + 540, + 480, + 614 + ], + "type": "inline_equation", + "content": "\\mathrm{J}_k^\\pi \\Sigma_k\\mathrm{J}_{\\mathbf{k}}^\\pi {}^\\top" + }, + { + "bbox": [ + 131, + 540, + 480, + 614 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 131, + 540, + 480, + 614 + ], + "type": "inline_equation", + "content": "\\mathrm{J}_k^\\pi" + }, + { + "bbox": [ + 131, + 540, + 480, + 614 + ], + "type": "text", + "content": " being the Jacobian of " + }, + { + "bbox": [ + 131, + 540, + 480, + 614 + ], + "type": "inline_equation", + "content": "\\pi" + }, + { + "bbox": [ + 131, + 540, + 480, + 614 + ], + "type": "text", + "content": " evaluated at " + }, + { + "bbox": [ + 131, + 540, + 480, + 614 + ], + "type": "inline_equation", + "content": "\\pmb{\\mu}_k" + }, + { + "bbox": [ + 131, + 540, + 480, + 614 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 617, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 617, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 131, + 617, + 481, + 666 + ], + "type": "text", + "content": "Rendering. To render the primitives " + }, + { + "bbox": [ + 131, + 617, + 481, + 666 + ], + "type": "inline_equation", + "content": "\\varGamma" + }, + { + "bbox": [ + 131, + 617, + 481, + 666 + ], + "type": "text", + "content": " representing a scene from camera " + }, + { + "bbox": [ + 131, + 617, + 481, + 666 + ], + "type": "inline_equation", + "content": "\\pi" + }, + { + "bbox": [ + 131, + 617, + 481, + 666 + ], + "type": "text", + "content": ", we require a decoder " + }, + { + "bbox": [ + 131, + 617, + 481, + 666 + ], + "type": "inline_equation", + "content": "\\varPhi" + }, + { + "bbox": [ + 131, + 617, + 481, + 666 + ], + "type": "text", + "content": " to be specified, which provides the feature we want to render as " + }, + { + "bbox": [ + 131, + 617, + 481, + 666 + ], + "type": "inline_equation", + "content": "\\varPhi(\\gamma_k,\\pmb{u})\\in\\mathbb{R}^m" + }, + { + "bbox": [ + 131, + 617, + 481, + 666 + ], + "type": "text", + "content": " for each Gaussian primitive " + }, + { + "bbox": [ + 131, + 617, + 481, + 666 + ], + "type": "inline_equation", + "content": "\\gamma_{k}" + }, + { + "bbox": [ + 131, + 617, + 481, + 666 + ], + "type": "text", + "content": " and pixel " + }, + { + "bbox": [ + 131, + 617, + 481, + 666 + ], + "type": "inline_equation", + "content": "\\pmb{u}" + }, + { + "bbox": [ + 131, + 617, + 481, + 666 + ], + "type": "text", + "content": ". Moreover, we assume Gaussian primitives " + }, + { + "bbox": [ + 131, + 617, + 481, + 666 + ], + "type": "inline_equation", + "content": "\\varGamma" + }, + { + "bbox": [ + 131, + 617, + 481, + 666 + ], + "type": "text", + "content": " to be ordered with respect to their center's depth," + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 233, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 233, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 233, + 100 + ], + "type": "text", + "content": "Rota Bulò et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 131, + 116, + 479, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 116, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 131, + 116, + 479, + 140 + ], + "type": "text", + "content": "when seen from the camera's reference frame. Then, the rendering equation takes the following form (with " + }, + { + "bbox": [ + 131, + 116, + 479, + 140 + ], + "type": "inline_equation", + "content": "\\varGamma" + }, + { + "bbox": [ + 131, + 116, + 479, + 140 + ], + "type": "text", + "content": " being omitted from the notation)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 236, + 152, + 481, + 183 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 236, + 152, + 481, + 183 + ], + "spans": [ + { + "bbox": [ + 236, + 152, + 481, + 183 + ], + "type": "interline_equation", + "content": "\\mathcal {R} [ \\pi , \\Phi ] (\\boldsymbol {u}) := \\sum_ {k = 1} ^ {K} \\Phi \\left(\\gamma_ {k}, \\boldsymbol {u}\\right) \\omega_ {k} ^ {\\pi} (\\boldsymbol {u}), \\tag {2}", + "image_path": "0d9e207a7003ef1d4b96ac06c5eacd4222a5f7a71a4fd46b7d9138ddf249f1f8.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 191, + 380, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 191, + 380, + 205 + ], + "spans": [ + { + "bbox": [ + 131, + 191, + 380, + 205 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 191, + 380, + 205 + ], + "type": "inline_equation", + "content": "\\omega_{k}^{\\pi}(\\pmb{u})" + }, + { + "bbox": [ + 131, + 191, + 380, + 205 + ], + "type": "text", + "content": " are alpha-compositing coefficients given by" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 224, + 215, + 481, + 247 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 224, + 215, + 481, + 247 + ], + "spans": [ + { + "bbox": [ + 224, + 215, + 481, + 247 + ], + "type": "interline_equation", + "content": "\\omega_ {k} ^ {\\pi} (\\boldsymbol {u}) := \\alpha_ {k} \\mathcal {G} _ {k} ^ {\\pi} (\\boldsymbol {u}) \\prod_ {j = 1} ^ {k - 1} \\left(1 - \\alpha_ {j} \\mathcal {G} _ {j} ^ {\\pi} (\\boldsymbol {u})\\right). \\tag {3}", + "image_path": "0658cfa71067fd939ed6421537b59582d0201fdbdf8493cb8cad4032f9d6f4d2.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 256, + 479, + 338 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 256, + 479, + 338 + ], + "spans": [ + { + "bbox": [ + 131, + 256, + 479, + 338 + ], + "type": "text", + "content": "If we assume the feature vectors " + }, + { + "bbox": [ + 131, + 256, + 479, + 338 + ], + "type": "inline_equation", + "content": "\\pmb{f}_k" + }, + { + "bbox": [ + 131, + 256, + 479, + 338 + ], + "type": "text", + "content": " to be spherical harmonics coefficients encoding an RGB function on the sphere, we can regard " + }, + { + "bbox": [ + 131, + 256, + 479, + 338 + ], + "type": "inline_equation", + "content": "\\varPhi_{\\mathrm{RGB}}(\\pmb{u})" + }, + { + "bbox": [ + 131, + 256, + 479, + 338 + ], + "type": "text", + "content": " as the decoded RGB color for the given view direction associated to pixel " + }, + { + "bbox": [ + 131, + 256, + 479, + 338 + ], + "type": "inline_equation", + "content": "\\pmb{u}" + }, + { + "bbox": [ + 131, + 256, + 479, + 338 + ], + "type": "text", + "content": ". If we use " + }, + { + "bbox": [ + 131, + 256, + 479, + 338 + ], + "type": "inline_equation", + "content": "\\varPhi_{\\mathrm{RGB}}" + }, + { + "bbox": [ + 131, + 256, + 479, + 338 + ], + "type": "text", + "content": " as the decoder in the rendering equation, we obtain a rendered color image " + }, + { + "bbox": [ + 131, + 256, + 479, + 338 + ], + "type": "inline_equation", + "content": "C_{\\pi}(\\pmb{u}) := \\mathcal{R}[\\pi, \\varPhi_{\\mathrm{RGB}}](\\pmb{u})" + }, + { + "bbox": [ + 131, + 256, + 479, + 338 + ], + "type": "text", + "content": " for each camera " + }, + { + "bbox": [ + 131, + 256, + 479, + 338 + ], + "type": "inline_equation", + "content": "\\pi" + }, + { + "bbox": [ + 131, + 256, + 479, + 338 + ], + "type": "text", + "content": ". Similarly, one can pick different " + }, + { + "bbox": [ + 131, + 256, + 479, + 338 + ], + "type": "inline_equation", + "content": "\\varPhi" + }, + { + "bbox": [ + 131, + 256, + 479, + 338 + ], + "type": "text", + "content": "'s to enable the rendering of depth, normals, or other quantities of interest as we will show later." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 344, + 479, + 404 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 344, + 479, + 404 + ], + "spans": [ + { + "bbox": [ + 131, + 344, + 479, + 404 + ], + "type": "text", + "content": "Mip-splatting. In [29], the authors introduce a variation of standard Gaussian splatting that focuses on solving aliasing issues. We refer the reader to the original paper for details, but the idea is to track the maximum sampling rate for each Gaussian primitive and use it to reduce aliasing effects by attenuating the Gaussian primitives' opacity." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 423, + 285, + 437 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 423, + 285, + 437 + ], + "spans": [ + { + "bbox": [ + 131, + 423, + 285, + 437 + ], + "type": "text", + "content": "3 Revising Densification" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 449, + 479, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 449, + 479, + 485 + ], + "spans": [ + { + "bbox": [ + 131, + 449, + 479, + 485 + ], + "type": "text", + "content": "We first review the Adaptive Density Control module proposed in the original Gaussian splatting work [9], highlight some of its limitations, and then introduce our novel and improved densification procedure." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 503, + 386, + 515 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 503, + 386, + 515 + ], + "spans": [ + { + "bbox": [ + 131, + 503, + 386, + 515 + ], + "type": "text", + "content": "3.1 Adaptive Density Control and its limitations" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 131, + 523, + 479, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 523, + 479, + 624 + ], + "spans": [ + { + "bbox": [ + 131, + 523, + 479, + 624 + ], + "type": "text", + "content": "3DGS [9] and follow-up extensions (e.g. Mip-splatting [29]) rely on the Adaptive Density Control (ADC) module to grow or prune Gaussian primitives. This module is run according to a predetermined schedule and densification decisions are based on gradient statistics collected across the ADC runs. Specifically, for each Gaussian primitive " + }, + { + "bbox": [ + 131, + 523, + 479, + 624 + ], + "type": "inline_equation", + "content": "\\gamma_{k}" + }, + { + "bbox": [ + 131, + 523, + 479, + 624 + ], + "type": "text", + "content": " the positional gradient magnitude " + }, + { + "bbox": [ + 131, + 523, + 479, + 624 + ], + "type": "inline_equation", + "content": "\\left\\| \\frac{\\partial L_{\\pi}}{\\partial \\mu_k} \\right\\|" + }, + { + "bbox": [ + 131, + 523, + 479, + 624 + ], + "type": "text", + "content": " is tracked and averaged over all rendered views " + }, + { + "bbox": [ + 131, + 523, + 479, + 624 + ], + "type": "inline_equation", + "content": "\\pi \\in \\Pi" + }, + { + "bbox": [ + 131, + 523, + 479, + 624 + ], + "type": "text", + "content": " within the collection period, where " + }, + { + "bbox": [ + 131, + 523, + 479, + 624 + ], + "type": "inline_equation", + "content": "L_{\\pi}" + }, + { + "bbox": [ + 131, + 523, + 479, + 624 + ], + "type": "text", + "content": " denotes the loss that is optimized for camera " + }, + { + "bbox": [ + 131, + 523, + 479, + 624 + ], + "type": "inline_equation", + "content": "\\pi" + }, + { + "bbox": [ + 131, + 523, + 479, + 624 + ], + "type": "text", + "content": ". The resulting quantity is denoted by " + }, + { + "bbox": [ + 131, + 523, + 479, + 624 + ], + "type": "inline_equation", + "content": "\\tau_{k}" + }, + { + "bbox": [ + 131, + 523, + 479, + 624 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 131, + 629, + 479, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 629, + 479, + 666 + ], + "spans": [ + { + "bbox": [ + 131, + 629, + 479, + 666 + ], + "type": "text", + "content": "Growing. ADC grows new Gaussian primitives via a clone or a split operation. A primitive " + }, + { + "bbox": [ + 131, + 629, + 479, + 666 + ], + "type": "inline_equation", + "content": "\\gamma_{k}" + }, + { + "bbox": [ + 131, + 629, + 479, + 666 + ], + "type": "text", + "content": " will be considered for a growing operation only if " + }, + { + "bbox": [ + 131, + 629, + 479, + 666 + ], + "type": "inline_equation", + "content": "\\tau_{k}" + }, + { + "bbox": [ + 131, + 629, + 479, + 666 + ], + "type": "text", + "content": " exceeds a user-defined threshold. The decision about which operation to apply depends" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 266, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 266, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 266, + 91, + 448, + 102 + ], + "type": "text", + "content": "Revising Densification in Gaussian Splitting" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "text", + "content": "on the size of the primitive measured in terms of the largest eigenvalue of the covariance matrix " + }, + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "inline_equation", + "content": "\\Sigma_{k}" + }, + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "text", + "content": ". Specifically, primitives larger than a threshold are split, otherwise cloned. When a primitive " + }, + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "inline_equation", + "content": "\\gamma_{k}" + }, + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "text", + "content": " is split, two new primitives are generated with their position being sampled from " + }, + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_k" + }, + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "text", + "content": " and their covariance being a scaled down version of " + }, + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "inline_equation", + "content": "\\Sigma_{k}" + }, + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "text", + "content": ", while preserving the same opacity and feature vector. When a clone operation takes place, a simple clone of " + }, + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "inline_equation", + "content": "\\gamma_{k}" + }, + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "text", + "content": " is instantiated." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 193, + 482, + 241 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 193, + 482, + 241 + ], + "spans": [ + { + "bbox": [ + 130, + 193, + 482, + 241 + ], + "type": "text", + "content": "Pruning. ADC prunes a Gaussian primitive " + }, + { + "bbox": [ + 130, + 193, + 482, + 241 + ], + "type": "inline_equation", + "content": "\\gamma_{k}" + }, + { + "bbox": [ + 130, + 193, + 482, + 241 + ], + "type": "text", + "content": " if its opacity " + }, + { + "bbox": [ + 130, + 193, + 482, + 241 + ], + "type": "inline_equation", + "content": "\\alpha_{k}" + }, + { + "bbox": [ + 130, + 193, + 482, + 241 + ], + "type": "text", + "content": " is below a user-defined threshold, typically 0.005. To ensure that an unused primitive is eventually pruned, a hard-reset of the opacity to a minimum value (usually 0.01) is enforced according to a predefined schedule." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 245, + 481, + 270 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 245, + 481, + 270 + ], + "spans": [ + { + "bbox": [ + 131, + 245, + 481, + 270 + ], + "type": "text", + "content": "Limitations. Deciding which Gaussian primitives to split/clone based on the magnitude of the positional gradient suffers from a number of limitations:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 275, + 481, + 357 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 138, + 275, + 481, + 298 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 275, + 481, + 298 + ], + "spans": [ + { + "bbox": [ + 138, + 275, + 481, + 298 + ], + "type": "text", + "content": "- Determining a threshold for a gradient magnitude is not intuitive and very sensitive to modifications to the model, losses and hyperparameters," + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 298, + 481, + 322 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 298, + 481, + 322 + ], + "spans": [ + { + "bbox": [ + 138, + 298, + 481, + 322 + ], + "type": "text", + "content": "- There are cases of scene underfitting also when the value of " + }, + { + "bbox": [ + 138, + 298, + 481, + 322 + ], + "type": "inline_equation", + "content": "\\tau_{k}" + }, + { + "bbox": [ + 138, + 298, + 481, + 322 + ], + "type": "text", + "content": " is below the threshold that triggers densification (see Fig. 1)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 322, + 481, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 322, + 481, + 357 + ], + "spans": [ + { + "bbox": [ + 138, + 322, + 481, + 357 + ], + "type": "text", + "content": "- It is not possible to directly control the number of Gaussian primitives that are generated for a given scene, resulting in possible out-of-memory errors if their number grows abnormally." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 130, + 362, + 482, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 362, + 482, + 399 + ], + "spans": [ + { + "bbox": [ + 130, + 362, + 482, + 399 + ], + "type": "text", + "content": "In addition, we found that the ADC's logic of growing primitives suffers from a bias that weights more the contribution of freshly cloned primitives. More details will follow in Sec. 3.3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 415, + 288, + 426 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 415, + 288, + 426 + ], + "spans": [ + { + "bbox": [ + 131, + 415, + 288, + 426 + ], + "type": "text", + "content": "3.2 Error-based densification" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 432, + 482, + 528 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 432, + 482, + 528 + ], + "spans": [ + { + "bbox": [ + 130, + 432, + 482, + 528 + ], + "type": "text", + "content": "Assume we have an image with an area characterized by a high-frequency pattern and covered by few large splatted Gaussian primitives (e.g. the grass in Fig. 1). Under this scenario, an infinitesimal change in the 3D location " + }, + { + "bbox": [ + 130, + 432, + 482, + 528 + ], + "type": "inline_equation", + "content": "\\pmb{\\mu}_k" + }, + { + "bbox": [ + 130, + 432, + 482, + 528 + ], + "type": "text", + "content": " of one of the corresponding Gaussian primitives " + }, + { + "bbox": [ + 130, + 432, + 482, + 528 + ], + "type": "inline_equation", + "content": "\\gamma_k" + }, + { + "bbox": [ + 130, + 432, + 482, + 528 + ], + "type": "text", + "content": " will leave the error almost unchanged and, hence, the collected magnitude of the positional gradient " + }, + { + "bbox": [ + 130, + 432, + 482, + 528 + ], + "type": "inline_equation", + "content": "\\tau_k" + }, + { + "bbox": [ + 130, + 432, + 482, + 528 + ], + "type": "text", + "content": " remains close to zero. In fact, " + }, + { + "bbox": [ + 130, + 432, + 482, + 528 + ], + "type": "inline_equation", + "content": "\\tau_k" + }, + { + "bbox": [ + 130, + 432, + 482, + 528 + ], + "type": "text", + "content": " is sensitive to error-changes, but is blind to the absolute value of the error. This becomes a problem, for we expect to increase the number of Gaussian primitives in areas exhibiting a larger error." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 529, + 482, + 636 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 529, + 482, + 636 + ], + "spans": [ + { + "bbox": [ + 130, + 529, + 482, + 636 + ], + "type": "text", + "content": "Given the above considerations, we propose to steer the densification decisions directly based on an auxiliary per-pixel error function " + }, + { + "bbox": [ + 130, + 529, + 482, + 636 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_{\\pi}" + }, + { + "bbox": [ + 130, + 529, + 482, + 636 + ], + "type": "text", + "content": " (e.g. Structural Similarity) that we measure when rendering on a camera " + }, + { + "bbox": [ + 130, + 529, + 482, + 636 + ], + "type": "inline_equation", + "content": "\\pi" + }, + { + "bbox": [ + 130, + 529, + 482, + 636 + ], + "type": "text", + "content": " with available ground-truth. One problem to address is how to turn per-pixel errors into per-Gaussian-primitive errors in light of the fact that each pixel error entangles the contribution of multiple Gaussian primitives. Our solution consists of first re-distributing the per-pixel errors " + }, + { + "bbox": [ + 130, + 529, + 482, + 636 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_{\\pi}(\\pmb{u})" + }, + { + "bbox": [ + 130, + 529, + 482, + 636 + ], + "type": "text", + "content": " to each Gaussian primitive " + }, + { + "bbox": [ + 130, + 529, + 482, + 636 + ], + "type": "inline_equation", + "content": "\\gamma_{k}" + }, + { + "bbox": [ + 130, + 529, + 482, + 636 + ], + "type": "text", + "content": " proportionally to their contribution to the rendered pixel color, i.e. proportionally to " + }, + { + "bbox": [ + 130, + 529, + 482, + 636 + ], + "type": "inline_equation", + "content": "w_{k}^{\\pi}(\\pmb{u})" + }, + { + "bbox": [ + 130, + 529, + 482, + 636 + ], + "type": "text", + "content": ". This yields the following error for each primitive " + }, + { + "bbox": [ + 130, + 529, + 482, + 636 + ], + "type": "inline_equation", + "content": "\\gamma_{k}" + }, + { + "bbox": [ + 130, + 529, + 482, + 636 + ], + "type": "text", + "content": " and camera " + }, + { + "bbox": [ + 130, + 529, + 482, + 636 + ], + "type": "inline_equation", + "content": "\\pi" + }, + { + "bbox": [ + 130, + 529, + 482, + 636 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 252, + 643, + 481, + 668 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 252, + 643, + 481, + 668 + ], + "spans": [ + { + "bbox": [ + 252, + 643, + 481, + 668 + ], + "type": "interline_equation", + "content": "E _ {k} ^ {\\pi} := \\sum_ {\\boldsymbol {u} \\in \\operatorname {P i x}} \\mathcal {E} _ {\\pi} (\\boldsymbol {u}) w _ {k} ^ {\\pi} (\\boldsymbol {u}), \\tag {4}", + "image_path": "0c7c3bd7e648500f8e1113d1115a82829d8355ab79729e000c81ea9e96cbcbf9.jpg" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 233, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 233, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 233, + 101 + ], + "type": "text", + "content": "Rota Bulò et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "content": "where the sum runs over the image pixels. Then, for each primitive " + }, + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "inline_equation", + "content": "\\gamma_{k}" + }, + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "content": " we track the maximum value of the error " + }, + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "inline_equation", + "content": "E_{k}^{\\pi}" + }, + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "content": " across all views " + }, + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "inline_equation", + "content": "\\pi \\in \\Pi" + }, + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "content": " seen between two runs of the ADC module, i.e." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 275, + 160, + 481, + 177 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 275, + 160, + 481, + 177 + ], + "spans": [ + { + "bbox": [ + 275, + 160, + 481, + 177 + ], + "type": "interline_equation", + "content": "E _ {k} := \\max _ {\\pi \\in \\Pi} E _ {k} ^ {\\pi}. \\tag {5}", + "image_path": "b470c4e0c1d37fc1a3569d33915f904e39f7f94c2ad333983339aea104430bad.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 182, + 479, + 219 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 182, + 479, + 219 + ], + "spans": [ + { + "bbox": [ + 130, + 182, + 479, + 219 + ], + "type": "text", + "content": "This is the score that we use to prioritize the growing of Gaussian primitives. As opposed to " + }, + { + "bbox": [ + 130, + 182, + 479, + 219 + ], + "type": "inline_equation", + "content": "\\tau_{k}" + }, + { + "bbox": [ + 130, + 182, + 479, + 219 + ], + "type": "text", + "content": ", it is easier to set a threshold for our new densification score, for it is typically expressed in terms of a known error metric." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 224, + 480, + 271 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 224, + 480, + 271 + ], + "spans": [ + { + "bbox": [ + 130, + 224, + 480, + 271 + ], + "type": "text", + "content": "Implementation details. In order to compute " + }, + { + "bbox": [ + 130, + 224, + 480, + 271 + ], + "type": "inline_equation", + "content": "E_{k}^{\\pi}" + }, + { + "bbox": [ + 130, + 224, + 480, + 271 + ], + "type": "text", + "content": " we assign an additional scalar " + }, + { + "bbox": [ + 130, + 224, + 480, + 271 + ], + "type": "inline_equation", + "content": "e_{k}" + }, + { + "bbox": [ + 130, + 224, + 480, + 271 + ], + "type": "text", + "content": " to each Gaussian primitive " + }, + { + "bbox": [ + 130, + 224, + 480, + 271 + ], + "type": "inline_equation", + "content": "\\gamma_{k}" + }, + { + "bbox": [ + 130, + 224, + 480, + 271 + ], + "type": "text", + "content": ", and enable the possibility of rendering it via the decoder " + }, + { + "bbox": [ + 130, + 224, + 480, + 271 + ], + "type": "inline_equation", + "content": "\\varPhi_{\\mathrm{ERR}}(\\gamma_k,\\pmb {u})\\coloneqq e_k" + }, + { + "bbox": [ + 130, + 224, + 480, + 271 + ], + "type": "text", + "content": ". Then, we add the following auxiliary loss to the standard Gaussian splatting training objective:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 226, + 279, + 481, + 310 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 226, + 279, + 481, + 310 + ], + "spans": [ + { + "bbox": [ + 226, + 279, + 481, + 310 + ], + "type": "interline_equation", + "content": "L _ {\\pi} ^ {\\text {a u x}} := \\sum_ {\\boldsymbol {u} \\in \\operatorname {P i x}} \\mathcal {N} [ \\mathcal {E} _ {\\pi} (\\boldsymbol {u}) ] \\underbrace {\\mathcal {R} [ \\pi , \\Phi_ {\\mathrm {E R R}} ] (\\boldsymbol {u})} _ {= \\sum_ {k = 1} ^ {K} e _ {k} \\omega_ {k} ^ {\\pi} (\\boldsymbol {u})}, \\tag {6}", + "image_path": "b2906ed76dba472f9018c408f2c4c1e1a082a71e772543318ae908be6a68aee5.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 316, + 479, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 316, + 479, + 376 + ], + "spans": [ + { + "bbox": [ + 130, + 316, + 479, + 376 + ], + "type": "text", + "content": "which is basically the dot product of the per-pixel error with gradient detached (denoted by " + }, + { + "bbox": [ + 130, + 316, + 479, + 376 + ], + "type": "inline_equation", + "content": "\\mathcal{X}" + }, + { + "bbox": [ + 130, + 316, + 479, + 376 + ], + "type": "text", + "content": ") and the rendering of the newly-added scalar. We initialize " + }, + { + "bbox": [ + 130, + 316, + 479, + 376 + ], + "type": "inline_equation", + "content": "e_k" + }, + { + "bbox": [ + 130, + 316, + 479, + 376 + ], + "type": "text", + "content": " to 0 for each Gaussian primitive " + }, + { + "bbox": [ + 130, + 316, + 479, + 376 + ], + "type": "inline_equation", + "content": "\\gamma_k" + }, + { + "bbox": [ + 130, + 316, + 479, + 376 + ], + "type": "text", + "content": " and never update it during training. In this way, " + }, + { + "bbox": [ + 130, + 316, + 479, + 376 + ], + "type": "inline_equation", + "content": "L_{\\pi}^{\\mathrm{aux}} = 0" + }, + { + "bbox": [ + 130, + 316, + 479, + 376 + ], + "type": "text", + "content": " and all Gaussian primitives' parameters, excepting " + }, + { + "bbox": [ + 130, + 316, + 479, + 376 + ], + "type": "inline_equation", + "content": "e_k" + }, + { + "bbox": [ + 130, + 316, + 479, + 376 + ], + "type": "text", + "content": ", are left invariant by this loss. The gradient with respect to " + }, + { + "bbox": [ + 130, + 316, + 479, + 376 + ], + "type": "inline_equation", + "content": "e_k" + }, + { + "bbox": [ + 130, + 316, + 479, + 376 + ], + "type": "text", + "content": " instead yields" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 234, + 381, + 481, + 410 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 234, + 381, + 481, + 410 + ], + "spans": [ + { + "bbox": [ + 234, + 381, + 481, + 410 + ], + "type": "interline_equation", + "content": "\\frac {\\partial L _ {\\pi} ^ {\\mathrm {a u x}}}{\\partial e _ {k}} = \\sum_ {\\boldsymbol {u} \\in \\operatorname {P i x}} \\mathcal {E} _ {\\pi} (\\boldsymbol {u}) \\omega_ {k} ^ {\\pi} (\\boldsymbol {u}) = E _ {k} ^ {\\pi}, \\tag {7}", + "image_path": "919bbc3293802d8403a1d1cff77ccec9a267e85744355b5cd63f62586846e472.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 416, + 477, + 429 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 416, + 477, + 429 + ], + "spans": [ + { + "bbox": [ + 130, + 416, + 477, + 429 + ], + "type": "text", + "content": "which is the per-Gaussian-primitive error for camera " + }, + { + "bbox": [ + 130, + 416, + 477, + 429 + ], + "type": "inline_equation", + "content": "\\pi" + }, + { + "bbox": [ + 130, + 416, + 477, + 429 + ], + "type": "text", + "content": " we wanted to compute." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 444, + 322, + 456 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 444, + 322, + 456 + ], + "spans": [ + { + "bbox": [ + 130, + 444, + 322, + 456 + ], + "type": "text", + "content": "3.3 Opacity correction after cloning" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 462, + 481, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 462, + 481, + 629 + ], + "spans": [ + { + "bbox": [ + 130, + 462, + 481, + 629 + ], + "type": "text", + "content": "In the original ADC module, when a Gaussian primitive is split or cloned, the opacity value is preserved. This choice introduces a bias in the case of the clone operation by implicitly increasing the impact of the densified primitive on the final rendered color. To see why this is the case we can follow the example in Fig. 2, where we consider what happens if we render a splatted Gaussian in its center pixel assuming an opacity value " + }, + { + "bbox": [ + 130, + 462, + 481, + 629 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 130, + 462, + 481, + 629 + ], + "type": "text", + "content": ". Before a cloning operation happens, the rendered color depends on primitives that come next in the ordering with weight " + }, + { + "bbox": [ + 130, + 462, + 481, + 629 + ], + "type": "inline_equation", + "content": "1 - \\alpha" + }, + { + "bbox": [ + 130, + 462, + 481, + 629 + ], + "type": "text", + "content": ". But after we clone, due to the alpha-compositing logic, we have that primitives that come next weight " + }, + { + "bbox": [ + 130, + 462, + 481, + 629 + ], + "type": "inline_equation", + "content": "(1 - \\alpha)^{2}" + }, + { + "bbox": [ + 130, + 462, + 481, + 629 + ], + "type": "text", + "content": ", which is lower than " + }, + { + "bbox": [ + 130, + 462, + 481, + 629 + ], + "type": "inline_equation", + "content": "1 - \\alpha" + }, + { + "bbox": [ + 130, + 462, + 481, + 629 + ], + "type": "text", + "content": " for all opacity values in " + }, + { + "bbox": [ + 130, + 462, + 481, + 629 + ], + "type": "inline_equation", + "content": "(0,1)" + }, + { + "bbox": [ + 130, + 462, + 481, + 629 + ], + "type": "text", + "content": ". Accordingly, by applying the standard logic of preserving the opacity after cloning we have a bias to weight more the cloned primitives. The solution we suggest consists in reducing the opacity of the primitives after cloning so that the bias is removed. The new opacity value " + }, + { + "bbox": [ + 130, + 462, + 481, + 629 + ], + "type": "inline_equation", + "content": "\\hat{\\alpha}" + }, + { + "bbox": [ + 130, + 462, + 481, + 629 + ], + "type": "text", + "content": " can be found by solving the equation " + }, + { + "bbox": [ + 130, + 462, + 481, + 629 + ], + "type": "inline_equation", + "content": "(1 - \\alpha) = (1 - \\hat{\\alpha})^{2}" + }, + { + "bbox": [ + 130, + 462, + 481, + 629 + ], + "type": "text", + "content": ", which yields " + }, + { + "bbox": [ + 130, + 462, + 481, + 629 + ], + "type": "inline_equation", + "content": "\\hat{\\alpha} := 1 - \\sqrt{1 - \\alpha}" + }, + { + "bbox": [ + 130, + 462, + 481, + 629 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 630, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 630, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 630, + 481, + 665 + ], + "type": "text", + "content": "If we depart from the simplified setting of considering only the center pixel and rather consider all pixels, it is unfortunately not possible to remove completely the bias. Nonetheless, the correction factor we introduce reduces the bias" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 265, + 91, + 448, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 265, + 91, + 448, + 102 + ], + "spans": [ + { + "bbox": [ + 265, + 91, + 448, + 102 + ], + "type": "text", + "content": "Revising Densification in Gaussian Splitting" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 480, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 132, + 114, + 204, + 174 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 204, + 174 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 204, + 174 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 204, + 174 + ], + "type": "image", + "image_path": "552cacd468ecb90ada953eabc7d71a84878e5040e495d5cfeedd3a3e18f98f3c.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 179, + 482, + 247 + ], + "lines": [ + { + "bbox": [ + 130, + 179, + 482, + 247 + ], + "spans": [ + { + "bbox": [ + 130, + 179, + 482, + 247 + ], + "type": "text", + "content": "Fig. 2: Consider rendering a single splatted Gaussian in its center pixel with opacity " + }, + { + "bbox": [ + 130, + 179, + 482, + 247 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 130, + 179, + 482, + 247 + ], + "type": "text", + "content": " before and after cloning. Before we clone, the rendered color depends with weight " + }, + { + "bbox": [ + 130, + 179, + 482, + 247 + ], + "type": "inline_equation", + "content": "1 - \\alpha" + }, + { + "bbox": [ + 130, + 179, + 482, + 247 + ], + "type": "text", + "content": " on what comes next. After we clone, since we preserve the opacity, the rendered color depends with weight " + }, + { + "bbox": [ + 130, + 179, + 482, + 247 + ], + "type": "inline_equation", + "content": "(1 - \\alpha)^{2}" + }, + { + "bbox": [ + 130, + 179, + 482, + 247 + ], + "type": "text", + "content": " on what comes next. Since " + }, + { + "bbox": [ + 130, + 179, + 482, + 247 + ], + "type": "inline_equation", + "content": "(1 - \\alpha) \\geq (1 - \\alpha)^{2}" + }, + { + "bbox": [ + 130, + 179, + 482, + 247 + ], + "type": "text", + "content": " we have a bias towards weighting more Gaussian primitives that get cloned. The proposed correction changes the opacity post clone to " + }, + { + "bbox": [ + 130, + 179, + 482, + 247 + ], + "type": "inline_equation", + "content": "\\hat{\\alpha}" + }, + { + "bbox": [ + 130, + 179, + 482, + 247 + ], + "type": "text", + "content": " so that the bias is removed." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 225, + 113, + 325, + 172 + ], + "blocks": [ + { + "bbox": [ + 225, + 113, + 325, + 172 + ], + "lines": [ + { + "bbox": [ + 225, + 113, + 325, + 172 + ], + "spans": [ + { + "bbox": [ + 225, + 113, + 325, + 172 + ], + "type": "image", + "image_path": "0e7914399d185eca8f4c05ee299107cafa9d7ef7b0741be5da1c520276602092.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 340, + 113, + 476, + 171 + ], + "blocks": [ + { + "bbox": [ + 340, + 113, + 476, + 171 + ], + "lines": [ + { + "bbox": [ + 340, + 113, + 476, + 171 + ], + "spans": [ + { + "bbox": [ + 340, + 113, + 476, + 171 + ], + "type": "image", + "image_path": "6f166a9d41021277a371fa7e88e62f4a89a03410b2de69a09502ca29d5a2c5d6.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 266, + 479, + 291 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 266, + 479, + 291 + ], + "spans": [ + { + "bbox": [ + 130, + 266, + 479, + 291 + ], + "type": "text", + "content": "for all pixels compared to keeping the opacity of the cloned primitive. Indeed, the following relation holds for all " + }, + { + "bbox": [ + 130, + 266, + 479, + 291 + ], + "type": "inline_equation", + "content": "\\alpha_{k}\\in (0,1)" + }, + { + "bbox": [ + 130, + 266, + 479, + 291 + ], + "type": "text", + "content": " and all pixels " + }, + { + "bbox": [ + 130, + 266, + 479, + 291 + ], + "type": "inline_equation", + "content": "\\pmb{u}" + }, + { + "bbox": [ + 130, + 266, + 479, + 291 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 194, + 295, + 481, + 309 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 194, + 295, + 481, + 309 + ], + "spans": [ + { + "bbox": [ + 194, + 295, + 481, + 309 + ], + "type": "interline_equation", + "content": "1 - \\alpha_ {k} \\mathcal {G} _ {k} ^ {\\pi} (\\boldsymbol {u}) \\geq \\left(1 - \\hat {\\alpha} _ {k} \\mathcal {G} _ {k} ^ {\\pi} (\\boldsymbol {u})\\right) ^ {2} > \\left(1 - \\alpha_ {k} \\mathcal {G} _ {k} ^ {\\pi} (\\boldsymbol {u})\\right) ^ {2}, \\tag {8}", + "image_path": "f92c908b6f0b248f24538b8e07068e02efe63e83448aaa3a0ec22259efd82cfc.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 314, + 479, + 352 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 314, + 479, + 352 + ], + "spans": [ + { + "bbox": [ + 130, + 314, + 479, + 352 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 314, + 479, + 352 + ], + "type": "inline_equation", + "content": "\\hat{\\alpha}_k \\coloneqq 1 - \\sqrt{1 - \\alpha_k}" + }, + { + "bbox": [ + 130, + 314, + 479, + 352 + ], + "type": "text", + "content": " is our corrected opacity. The proof of the relation follows by noting that " + }, + { + "bbox": [ + 130, + 314, + 479, + 352 + ], + "type": "inline_equation", + "content": "\\hat{\\alpha}_k" + }, + { + "bbox": [ + 130, + 314, + 479, + 352 + ], + "type": "text", + "content": " can be rewritten as " + }, + { + "bbox": [ + 130, + 314, + 479, + 352 + ], + "type": "inline_equation", + "content": "\\frac{\\alpha_k}{1 + \\sqrt{1 - \\alpha_k}}" + }, + { + "bbox": [ + 130, + 314, + 479, + 352 + ], + "type": "text", + "content": " which is strictly smaller than " + }, + { + "bbox": [ + 130, + 314, + 479, + 352 + ], + "type": "inline_equation", + "content": "\\alpha_k" + }, + { + "bbox": [ + 130, + 314, + 479, + 352 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 130, + 314, + 479, + 352 + ], + "type": "inline_equation", + "content": "\\alpha_k \\in (0,1)" + }, + { + "bbox": [ + 130, + 314, + 479, + 352 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 352, + 480, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 352, + 480, + 413 + ], + "spans": [ + { + "bbox": [ + 130, + 352, + 480, + 413 + ], + "type": "text", + "content": "Finally, the correction of the opacity as shown above is derived assuming we clone a Gaussian primitive, but does not strictly match the case of a split operation, for when we split we move the two offspring randomly away from the previous center and we change the covariance scale. For this reason, we stick to the standard rule of preserving the opacity of a primitive we split." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 426, + 290, + 438 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 426, + 290, + 438 + ], + "spans": [ + { + "bbox": [ + 131, + 426, + 290, + 438 + ], + "type": "text", + "content": "3.4 Primitives growth control" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 443, + 482, + 612 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 443, + 482, + 612 + ], + "spans": [ + { + "bbox": [ + 130, + 443, + 482, + 612 + ], + "type": "text", + "content": "The ADC module grows a Gaussian primitive if " + }, + { + "bbox": [ + 130, + 443, + 482, + 612 + ], + "type": "inline_equation", + "content": "\\tau_{k}" + }, + { + "bbox": [ + 130, + 443, + 482, + 612 + ], + "type": "text", + "content": " is larger than a threshold. This mechanism can lead to unpredictable growth of the number of primitives, eventually resulting in out-of-memory issues. To avoid this problem, we introduce a global limit to the maximum number of Gaussian primitives, and a mechanism to control the maximum number of primitives that can be created each time densification is run. Among the many possible options, we explore a logic that limits new primitive offspring to a fixed fraction of the primitives that already exist. In case the number of primitives that are entitled to be densified exceeds the available budget, we retain only the ones that exhibit the highest densification score. An example of this process is shown in Fig. 3, compared to the one from 3DGS: For Ours, the number of primitives grows smoothly until it reaches the allotted maximum, without the discontinuities induced by opacity reset (see Sec. 3.5). The way we control the number of primitives is not limited to our error-based densification logic, but can be applied equally to the original gradient-based one." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 131, + 624, + 301, + 637 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 624, + 301, + 637 + ], + "spans": [ + { + "bbox": [ + 131, + 624, + 301, + 637 + ], + "type": "text", + "content": "3.5 Alternative to opacity reset" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "content": "The strategy introduced in [9] to favour the sparsification of Gaussian primitives consists in periodical hard resets of the opacity for all primitives to a low" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 233, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 233, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 233, + 100 + ], + "type": "text", + "content": "Rota Bulò et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 183, + 129, + 430, + 242 + ], + "blocks": [ + { + "bbox": [ + 215, + 114, + 409, + 127 + ], + "lines": [ + { + "bbox": [ + 215, + 114, + 409, + 127 + ], + "spans": [ + { + "bbox": [ + 215, + 114, + 409, + 127 + ], + "type": "text", + "content": "Number of Gaussians by training iteration" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 183, + 129, + 430, + 242 + ], + "lines": [ + { + "bbox": [ + 183, + 129, + 430, + 242 + ], + "spans": [ + { + "bbox": [ + 183, + 129, + 430, + 242 + ], + "type": "image", + "image_path": "3e11b4fbe66ba682b8f1acde909f2031ca4f58821e9935e0cef6f6452daf91cc.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 249, + 482, + 316 + ], + "lines": [ + { + "bbox": [ + 130, + 249, + 482, + 316 + ], + "spans": [ + { + "bbox": [ + 130, + 249, + 482, + 316 + ], + "type": "text", + "content": "Fig. 3: Evolution of the number of Gaussians in 3DGS, and in our method with upper limit set to the number reached by 3DGS (on the garden scene from the Mip-NeRF 360 dataset). Note that, while 3DGS' ADC process stops after " + }, + { + "bbox": [ + 130, + 249, + 482, + 316 + ], + "type": "inline_equation", + "content": "15\\mathrm{k}" + }, + { + "bbox": [ + 130, + 249, + 482, + 316 + ], + "type": "text", + "content": " iterations, ours remains active for " + }, + { + "bbox": [ + 130, + 249, + 482, + 316 + ], + "type": "inline_equation", + "content": "27\\mathrm{k}" + }, + { + "bbox": [ + 130, + 249, + 482, + 316 + ], + "type": "text", + "content": ". This is not immediately visible from the plot, since pruned primitives are immediately replaced by newly spawned ones, keeping the overall number stable once the maximum is reached." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 344, + 482, + 476 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 344, + 482, + 476 + ], + "spans": [ + { + "bbox": [ + 130, + 344, + 482, + 476 + ], + "type": "text", + "content": "value, so that primitives whose opacity is not increased again by the optimization will eventually be pruned. This introduces a small shock in the training trajectory, which is suboptimal for the sake of having stable and predictable training dynamics. Moreover, resetting the opacity is particularly harmful for our error-based densification method, for it will lead to misleading error statistics right after the hard-reset, potentially triggering wrong densification decisions. For this reason, we propose a different logic to favour primitives pruning in a smoother way. Specifically, we decrease the opacity of each primitive by a fixed amount (we use 0.001) after each densification run, so that the opacity will gradually move towards the pruning range. In this way, we avoid sudden changes in the densification metric, while preserving the desired sparsification properties." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 479, + 482, + 562 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 479, + 482, + 562 + ], + "spans": [ + { + "bbox": [ + 130, + 479, + 482, + 562 + ], + "type": "text", + "content": "One downside of the new opacity regularization logic is that the constant push towards lowering the opacity of the primitives implicitly invites the model to make more use of the background where possible. This is also harmful, for it could generate more holes in the scene that will be visible from novel views. To counteract this dynamics, we also regularize the residual probabilities of the alpha-compositing (a.k.a. residual transmittance) to be zero for every pixel, by simply minimizing their average value, weighted by a hyperparameter (here 0.1)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 593, + 302, + 605 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 593, + 302, + 605 + ], + "spans": [ + { + "bbox": [ + 131, + 593, + 302, + 605 + ], + "type": "text", + "content": "4 Experimental Evaluation" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 629, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 481, + 665 + ], + "type": "text", + "content": "In the following we show how our improved ADC mechanism can equally be applied both to standard 3DGS [9] and its Mip-Splatting extension [29], providing benefits to both." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 266, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 266, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 266, + 91, + 447, + 102 + ], + "type": "text", + "content": "Revising Densification in Gaussian Splitting" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 266, + 126 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 266, + 126 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 266, + 126 + ], + "type": "text", + "content": "4.1 Datasets and metrics" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 138, + 482, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 138, + 482, + 258 + ], + "spans": [ + { + "bbox": [ + 130, + 138, + 482, + 258 + ], + "type": "text", + "content": "We follow the experimental setup from the 3DGS [9] paper, focusing on the real-world scenes from the Mip-NeRF 360 [1], Tanks and Temples [10] and Deep Blending [6] datasets. Mip-NeRF 360 comprises nine scenes (5 outdoor, 4 indoor) captured in a circular pattern which focuses on a central area of a few meters, with a potentially unbounded background. For Tanks and Temples, we focus on the \"Truck\" and \"Train\" scenes, while for Deep Blending we focus on the \"Dr Johnson\" and \"Playroom\" scenes, using the images and SfM reconstructions shared by the Gaussian Splatting authors. In each experiment we set aside each 8th image as a validation set, and report peak signal-to-noise ratio (PSNR), structural similarity (SSIM) and the perceptual metric from [31] (LPIPS)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 258, + 481, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 258, + 481, + 281 + ], + "spans": [ + { + "bbox": [ + 130, + 258, + 481, + 281 + ], + "type": "text", + "content": "Additional experiments on NeRF-Synthetic [17] are provided in the supplementary material." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 302, + 259, + 314 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 302, + 259, + 314 + ], + "spans": [ + { + "bbox": [ + 132, + 302, + 259, + 314 + ], + "type": "text", + "content": "4.2 Experimental setup" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 324, + 482, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 324, + 482, + 479 + ], + "spans": [ + { + "bbox": [ + 130, + 324, + 482, + 479 + ], + "type": "text", + "content": "We evaluate based on our re-implementations of 3DGS, which allows us to easily switch between standard 3DGS, Mip-Splatting, the original ADC of [9], our contributions or any combination thereof. We reproduce the training settings proposed in [9,29] and the respective public code-bases" + }, + { + "bbox": [ + 130, + 324, + 482, + 479 + ], + "type": "inline_equation", + "content": "^{12}" + }, + { + "bbox": [ + 130, + 324, + 482, + 479 + ], + "type": "text", + "content": ", including number of training iterations, batch size, input resolution, learning rates etc. When training with our contributions, we grow Gaussians with " + }, + { + "bbox": [ + 130, + 324, + 482, + 479 + ], + "type": "inline_equation", + "content": "E_{k} > 0.1" + }, + { + "bbox": [ + 130, + 324, + 482, + 479 + ], + "type": "text", + "content": ", adding up to " + }, + { + "bbox": [ + 130, + 324, + 482, + 479 + ], + "type": "inline_equation", + "content": "5\\%" + }, + { + "bbox": [ + 130, + 324, + 482, + 479 + ], + "type": "text", + "content": " of the current number of primitives at each densification step. Differently from 3DGS, we keep our ADC process active for 27k iterations (i.e. " + }, + { + "bbox": [ + 130, + 324, + 482, + 479 + ], + "type": "inline_equation", + "content": "90\\%" + }, + { + "bbox": [ + 130, + 324, + 482, + 479 + ], + "type": "text", + "content": " of the training process), instead of stopping it after 15k. Other relevant hyper-parameters are left to the default values used in 3DGS, and shared across all datasets and scenes. In all our experiments, we set the maximum primitives budget to the number of primitives (or its median, for experiments with multiple runs) generated by the corresponding baseline, in order to obtain perfectly comparable models." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 484, + 482, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 484, + 482, + 556 + ], + "spans": [ + { + "bbox": [ + 130, + 484, + 482, + 556 + ], + "type": "text", + "content": "A note about LPIPS. Investigating the 3DGS and Mip-Splating baselines, we discovered a discrepancy in the way LPIPS is calculated in both public code-bases, which resulted in under-estimated values being reported in the original papers. This was confirmed in private correspondence with the authors. In order to simplify comparisons with future works that don't rely on these code-bases, and might be unaware of this issue, we report correct LPIPS values here." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 562, + 482, + 634 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 562, + 482, + 634 + ], + "spans": [ + { + "bbox": [ + 130, + 562, + 482, + 634 + ], + "type": "text", + "content": "A note about evaluations on MipNeRF360. The standard protocol for the dataset trains and evaluates the models at the same image scale, with indoor scenes down-sampled by a factor of two and outdoor scenes by four. The authors of the dataset provide down-scaled versions of the images. However, several recent works like [29] do not use the original images, which suffer from compression artifacts, but down-sample the images from the higher resolution counterparts." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 233, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 233, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 233, + 100 + ], + "type": "text", + "content": "Rota Bulò et al." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 133, + 642, + 395, + 654 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 642, + 395, + 654 + ], + "spans": [ + { + "bbox": [ + 133, + 642, + 395, + 654 + ], + "type": "text", + "content": "1 https://github.com/graphdeco-inria/gaussian-splatting" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 135, + 654, + 376, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 654, + 376, + 665 + ], + "spans": [ + { + "bbox": [ + 135, + 654, + 376, + 665 + ], + "type": "text", + "content": "2 https://github.com/autonomousvision/mip-splatting" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 153, + 157, + 461, + 275 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 480, + 147 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 480, + 147 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 480, + 147 + ], + "type": "text", + "content": "Table 1: Results on the Mip-NeRF 360 dataset. Top section of the table: results from the Gaussian Splitting paper; bottom section: results from our re-implementation averaged over 5 runs." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 153, + 157, + 461, + 275 + ], + "lines": [ + { + "bbox": [ + 153, + 157, + 461, + 275 + ], + "spans": [ + { + "bbox": [ + 153, + 157, + 461, + 275 + ], + "type": "table", + "html": "
MethodPSNR ↑SSIM ↑LPIPS ↓
MeanStdevMeanStdevMeanStdev
Plenoxels [4]23.08-0.626-0.436-
INGP [18]25.59-0.699-0.331-
Mip-NeRF 360 [1]27.69-0.792-0.237-
3DGS [9]27.450.0810.8170.0010.2500.001
Mip-Splatting [29]27.510.0740.8180.0010.2510.001
Ours, 3DGS27.610.0670.8220.0010.2230.001
Ours, Mip-Splatting27.700.0650.8230.0010.2230.001
", + "image_path": "ef286a968d3437a4116d7053db1578b08a3f1c0d01c227aa786467b3fad2272e.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 153, + 328, + 461, + 446 + ], + "blocks": [ + { + "bbox": [ + 132, + 285, + 480, + 318 + ], + "lines": [ + { + "bbox": [ + 132, + 285, + 480, + 318 + ], + "spans": [ + { + "bbox": [ + 132, + 285, + 480, + 318 + ], + "type": "text", + "content": "Table 2: Results on the Tanks and Temples dataset. Top section of the table: results from the Gaussian Splitting paper; bottom section: results from our re-implementation averaged over 5 runs." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 153, + 328, + 461, + 446 + ], + "lines": [ + { + "bbox": [ + 153, + 328, + 461, + 446 + ], + "spans": [ + { + "bbox": [ + 153, + 328, + 461, + 446 + ], + "type": "table", + "html": "
MethodPSNR ↑SSIM ↑LPIPS ↓
MeanStdevMeanStdevMeanStdev
Plenoxels [4]21.08-0.719-0.379-
INGP [18]21.92-0.745-0.305-
Mip-NeRF 360 [1]22.22-0.759-0.257-
3DGS [9]23.440.0890.8450.0020.2120.001
Mip-Splatting [29]23.650.0660.8490.0010.2110.001
Ours, 3DGS23.930.0700.8530.0010.1870.001
Ours, Mip-Splatting24.100.1090.8570.0020.1830.001
", + "image_path": "5cabc75502c9ce24c7f2468c1e46491cafd25efefeca28c2d292bd70836eb1aa.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 469, + 480, + 505 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 469, + 480, + 505 + ], + "spans": [ + { + "bbox": [ + 132, + 469, + 480, + 505 + ], + "type": "text", + "content": "As a result, the scores they report are significantly higher. We use the downscaled images from the dataset for training/evaluation to stay consistent with older NeRF works and report scores of competitors under the same setting." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 133, + 524, + 223, + 534 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 524, + 223, + 534 + ], + "spans": [ + { + "bbox": [ + 133, + 524, + 223, + 534 + ], + "type": "text", + "content": "4.3 Main results" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 545, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 545, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 545, + 481, + 665 + ], + "type": "text", + "content": "In a first set of experiments, we evaluate the effectiveness of our improved ADC strategy (Ours) when applied to 3DGS and Mip-Splatting. Results, collected over 5 training runs to average out the randomness induced by stochastic primitive splitting, are reported in Tab. 1, 2 and 3. For the sake of completeness, we also include scores obtained with three NeRF baselines, i.e. Plenoxels [4], InstantNGP (INGP) [18] and Mip-NeRF 360 [1], as originally reported in [9]. Our approach consistently outperforms the corresponding baselines (i.e. Ours, 3DGS vs. 3DGS; Ours, Mip-Splatting vs. Mip-Splatting), particularly on SSIM and LPIPS. This is in line with what we discussed in Sec. 3.1 and 3.2: standard ADC often leads to localized under-fitting, as it fails to split large gaussians" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 266, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 266, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 266, + 91, + 447, + 102 + ], + "type": "text", + "content": "Revising Densification in Gaussian Splitting" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 158, + 248, + 223 + ], + "blocks": [ + { + "bbox": [ + 133, + 158, + 248, + 223 + ], + "lines": [ + { + "bbox": [ + 133, + 158, + 248, + 223 + ], + "spans": [ + { + "bbox": [ + 133, + 158, + 248, + 223 + ], + "type": "image", + "image_path": "858307533f28b0a273f85513a87d74920db2378a348ac62ef90c84b155f0b6e0.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 150, + 224, + 231, + 252 + ], + "blocks": [ + { + "bbox": [ + 150, + 224, + 231, + 252 + ], + "lines": [ + { + "bbox": [ + 150, + 224, + 231, + 252 + ], + "spans": [ + { + "bbox": [ + 150, + 224, + 231, + 252 + ], + "type": "image", + "image_path": "ce614e38d782800e44c355e081982514ac07258efeba0e213ab19abea58ded83.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 249, + 159, + 364, + 224 + ], + "blocks": [ + { + "bbox": [ + 249, + 159, + 364, + 224 + ], + "lines": [ + { + "bbox": [ + 249, + 159, + 364, + 224 + ], + "spans": [ + { + "bbox": [ + 249, + 159, + 364, + 224 + ], + "type": "image", + "image_path": "21409c2181685c0f81b46d78b8c09c6d84087863a5f7ab6f4e3e4cb9656fc2e4.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 266, + 224, + 347, + 252 + ], + "blocks": [ + { + "bbox": [ + 266, + 224, + 347, + 252 + ], + "lines": [ + { + "bbox": [ + 266, + 224, + 347, + 252 + ], + "spans": [ + { + "bbox": [ + 266, + 224, + 347, + 252 + ], + "type": "image", + "image_path": "3770316d0329bcaece8805ac389529ae426c2c9e8759912a6de143cad65e0a9f.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 365, + 159, + 481, + 224 + ], + "blocks": [ + { + "bbox": [ + 365, + 159, + 481, + 224 + ], + "lines": [ + { + "bbox": [ + 365, + 159, + 481, + 224 + ], + "spans": [ + { + "bbox": [ + 365, + 159, + 481, + 224 + ], + "type": "image", + "image_path": "2153b1b9400cd93249e1f2dc103206d8290523984cff59ac5d67df104a227b3b.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 382, + 224, + 463, + 251 + ], + "blocks": [ + { + "bbox": [ + 382, + 224, + 463, + 251 + ], + "lines": [ + { + "bbox": [ + 382, + 224, + 463, + 251 + ], + "spans": [ + { + "bbox": [ + 382, + 224, + 463, + 251 + ], + "type": "image", + "image_path": "b6efd124893b2007e6a08a83a91e6654d01db4332f0b786f514bf6557ee3301e.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 132, + 259, + 248, + 325 + ], + "blocks": [ + { + "bbox": [ + 132, + 259, + 248, + 325 + ], + "lines": [ + { + "bbox": [ + 132, + 259, + 248, + 325 + ], + "spans": [ + { + "bbox": [ + 132, + 259, + 248, + 325 + ], + "type": "image", + "image_path": "bb3add9c6dc83d1e7bacc8eedf7e39f62643ca5b9876469c14f82b301c5c7ca3.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 150, + 325, + 231, + 353 + ], + "blocks": [ + { + "bbox": [ + 150, + 325, + 231, + 353 + ], + "lines": [ + { + "bbox": [ + 150, + 325, + 231, + 353 + ], + "spans": [ + { + "bbox": [ + 150, + 325, + 231, + 353 + ], + "type": "image", + "image_path": "8c4abd8c09beefacf463fa8991016ff44b63dcf084a31bbe9cb25ce081686548.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 249, + 259, + 364, + 325 + ], + "blocks": [ + { + "bbox": [ + 249, + 259, + 364, + 325 + ], + "lines": [ + { + "bbox": [ + 249, + 259, + 364, + 325 + ], + "spans": [ + { + "bbox": [ + 249, + 259, + 364, + 325 + ], + "type": "image", + "image_path": "d02658d9ad8b4c5a431e5239d482520cb2647dc02dac41beaa164ea3d5a769c1.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 266, + 325, + 347, + 353 + ], + "blocks": [ + { + "bbox": [ + 266, + 325, + 347, + 353 + ], + "lines": [ + { + "bbox": [ + 266, + 325, + 347, + 353 + ], + "spans": [ + { + "bbox": [ + 266, + 325, + 347, + 353 + ], + "type": "image", + "image_path": "815814d264d5563d9098d5391722c69d4c73f601d77b044d731d19c9898c9c0c.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 365, + 259, + 480, + 325 + ], + "blocks": [ + { + "bbox": [ + 365, + 259, + 480, + 325 + ], + "lines": [ + { + "bbox": [ + 365, + 259, + 480, + 325 + ], + "spans": [ + { + "bbox": [ + 365, + 259, + 480, + 325 + ], + "type": "image", + "image_path": "0ceaccbae16322cd5b24a0d91ff0fcaee5bf3ad40088ee1e2d986a284161a0a0.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 382, + 325, + 463, + 353 + ], + "blocks": [ + { + "bbox": [ + 382, + 325, + 463, + 353 + ], + "lines": [ + { + "bbox": [ + 382, + 325, + 463, + 353 + ], + "spans": [ + { + "bbox": [ + 382, + 325, + 463, + 353 + ], + "type": "image", + "image_path": "195593b97f31df17f4d393da409b166956ce968d495b03349727ab9fbcbf7689.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 132, + 357, + 247, + 434 + ], + "blocks": [ + { + "bbox": [ + 132, + 357, + 247, + 434 + ], + "lines": [ + { + "bbox": [ + 132, + 357, + 247, + 434 + ], + "spans": [ + { + "bbox": [ + 132, + 357, + 247, + 434 + ], + "type": "image", + "image_path": "b25904c7913bab6dbf6ae598b55fa623e56ef1bcfa3feeec360cc593aee43395.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 150, + 434, + 231, + 462 + ], + "blocks": [ + { + "bbox": [ + 150, + 434, + 231, + 462 + ], + "lines": [ + { + "bbox": [ + 150, + 434, + 231, + 462 + ], + "spans": [ + { + "bbox": [ + 150, + 434, + 231, + 462 + ], + "type": "image", + "image_path": "d8062f087f92f6162452b56cda07b84884c721551fb7050d8963e18cc5a7833b.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 249, + 357, + 364, + 434 + ], + "blocks": [ + { + "bbox": [ + 249, + 357, + 364, + 434 + ], + "lines": [ + { + "bbox": [ + 249, + 357, + 364, + 434 + ], + "spans": [ + { + "bbox": [ + 249, + 357, + 364, + 434 + ], + "type": "image", + "image_path": "e73b6a6de4f855f05c46a39c2048cb36c6152d5cfe6ea1c509c7821c5459e555.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 266, + 434, + 347, + 462 + ], + "blocks": [ + { + "bbox": [ + 266, + 434, + 347, + 462 + ], + "lines": [ + { + "bbox": [ + 266, + 434, + 347, + 462 + ], + "spans": [ + { + "bbox": [ + 266, + 434, + 347, + 462 + ], + "type": "image", + "image_path": "d303d692b14d6a1edca90b44b324c70ebd5afe5a60a338dd4a7ec270e6d4f1a3.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 365, + 357, + 480, + 434 + ], + "blocks": [ + { + "bbox": [ + 365, + 357, + 480, + 434 + ], + "lines": [ + { + "bbox": [ + 365, + 357, + 480, + 434 + ], + "spans": [ + { + "bbox": [ + 365, + 357, + 480, + 434 + ], + "type": "image", + "image_path": "eb17d26e4871dd0b1e7aaaa420270651d57fda735efcabd4c5a3400fb62419bc.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 382, + 434, + 463, + 462 + ], + "blocks": [ + { + "bbox": [ + 382, + 434, + 463, + 462 + ], + "lines": [ + { + "bbox": [ + 382, + 434, + 463, + 462 + ], + "spans": [ + { + "bbox": [ + 382, + 434, + 463, + 462 + ], + "type": "image", + "image_path": "751134d20b393733d0ba796c5734f3579fd460728cccac4b17eae3148eefa104.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 133, + 470, + 248, + 547 + ], + "blocks": [ + { + "bbox": [ + 133, + 470, + 248, + 547 + ], + "lines": [ + { + "bbox": [ + 133, + 470, + 248, + 547 + ], + "spans": [ + { + "bbox": [ + 133, + 470, + 248, + 547 + ], + "type": "image", + "image_path": "35c59216cbc1fb86523647ce1f2dcceafa21928fd7721278b8429398c8b1b56c.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 150, + 547, + 231, + 575 + ], + "blocks": [ + { + "bbox": [ + 150, + 547, + 231, + 575 + ], + "lines": [ + { + "bbox": [ + 150, + 547, + 231, + 575 + ], + "spans": [ + { + "bbox": [ + 150, + 547, + 231, + 575 + ], + "type": "image", + "image_path": "3893412dd7c444e7017fbf694484fef839a9ce5c1c2886fe92484ebd1ca7121e.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 584, + 482, + 617 + ], + "lines": [ + { + "bbox": [ + 130, + 584, + 482, + 617 + ], + "spans": [ + { + "bbox": [ + 130, + 584, + 482, + 617 + ], + "type": "text", + "content": "Fig. 4: Qualitative results on the Mip-NeRF 360 and Tanks and Temples validation sets. Both 3DGS and Ours use the same number of primitives. Best viewed on screen at high magnification." + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_caption" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 249, + 470, + 364, + 547 + ], + "blocks": [ + { + "bbox": [ + 249, + 470, + 364, + 547 + ], + "lines": [ + { + "bbox": [ + 249, + 470, + 364, + 547 + ], + "spans": [ + { + "bbox": [ + 249, + 470, + 364, + 547 + ], + "type": "image", + "image_path": "78c0e9458d6af011f9f984c0ff83cdcc93f633ae5697fb0ed67daf68ff3ff713.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 266, + 547, + 347, + 575 + ], + "blocks": [ + { + "bbox": [ + 266, + 547, + 347, + 575 + ], + "lines": [ + { + "bbox": [ + 266, + 547, + 347, + 575 + ], + "spans": [ + { + "bbox": [ + 266, + 547, + 347, + 575 + ], + "type": "image", + "image_path": "b055c6bbd040714ad14f987e807c1607a5fee66d305d2e0c52ff6f4f37c3a47a.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 365, + 470, + 480, + 547 + ], + "blocks": [ + { + "bbox": [ + 365, + 470, + 480, + 547 + ], + "lines": [ + { + "bbox": [ + 365, + 470, + 480, + 547 + ], + "spans": [ + { + "bbox": [ + 365, + 470, + 480, + 547 + ], + "type": "image", + "image_path": "60cf8f0109a37415d4361fc5d43ed399d5dc85ec98bd5c563bc0bf1470e675f6.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 382, + 547, + 463, + 575 + ], + "blocks": [ + { + "bbox": [ + 382, + 547, + 463, + 575 + ], + "lines": [ + { + "bbox": [ + 382, + 547, + 463, + 575 + ], + "spans": [ + { + "bbox": [ + 382, + 547, + 463, + 575 + ], + "type": "image", + "image_path": "4ec7a106b865c0af9fae1265483ad5f8195d5b1c0335a173ced0447833bac32d.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + } + ], + "index": 25 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 233, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 233, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 233, + 100 + ], + "type": "text", + "content": "Rota Bulò et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 153, + 157, + 461, + 275 + ], + "blocks": [ + { + "bbox": [ + 130, + 114, + 479, + 147 + ], + "lines": [ + { + "bbox": [ + 130, + 114, + 479, + 147 + ], + "spans": [ + { + "bbox": [ + 130, + 114, + 479, + 147 + ], + "type": "text", + "content": "Table 3: Results on the Deep Blending dataset. Top section of the table: results from the Gaussian Splatting paper; bottom section: results from our re-implementation averaged over 5 runs." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 153, + 157, + 461, + 275 + ], + "lines": [ + { + "bbox": [ + 153, + 157, + 461, + 275 + ], + "spans": [ + { + "bbox": [ + 153, + 157, + 461, + 275 + ], + "type": "table", + "html": "
MethodPSNR ↑SSIM ↑LPIPS ↓
MeanStdevMeanStdevMeanStdev
Plenoxels [4]23.06-0.795-0.510-
INGP [18]24.96-0.817-0.390-
Mip-NeRF 360 [1]29.40-0.901-0.245-
3DGS [9]29.540.0960.9020.0000.3110.001
Mip-Splatting [29]29.680.0680.9030.0000.3090.000
Ours, 3DGS29.500.1100.9040.0010.3050.001
Ours, Mip-Splatting29.640.1260.9050.0010.3030.001
", + "image_path": "cac36875a267502624f34ee5e1219a0a82c0476b881b9a9db09dd3076107227e.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 144, + 327, + 470, + 396 + ], + "blocks": [ + { + "bbox": [ + 130, + 284, + 479, + 317 + ], + "lines": [ + { + "bbox": [ + 130, + 284, + 479, + 317 + ], + "spans": [ + { + "bbox": [ + 130, + 284, + 479, + 317 + ], + "type": "text", + "content": "Table 4: Ablation experiments on the Mip-NeRF 360 dataset, adding individual contributions to 3DGS or removing them from Ours. OC: Opacity Correction, Sec. 3.3; GC: Growth Control, Sec. 3.4; OR: Opacity Regularization, Sec. 3.5." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 144, + 327, + 470, + 396 + ], + "lines": [ + { + "bbox": [ + 144, + 327, + 470, + 396 + ], + "spans": [ + { + "bbox": [ + 144, + 327, + 470, + 396 + ], + "type": "table", + "html": "
Baseline3DGS+OROurs
+OC+GC+ORFull-OC-GC-OR
PSNR ↑27.4527.6527.3527.4827.6127.0427.5427.28
SSIM ↑0.8170.8220.8100.8180.8220.8120.8180.810
LPIPS ↓0.2500.2390.2560.2430.2230.2350.2230.234
", + "image_path": "e214c4058508a85035e036ace0bf374b32e71eea82e4a89f589baf42612f8f48.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 416, + 480, + 536 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 416, + 480, + 536 + ], + "spans": [ + { + "bbox": [ + 130, + 416, + 480, + 536 + ], + "type": "text", + "content": "that cover highly-textured regions of the scene. This kind of errors is poorly reflected by PSNR, which measures the \"average fit\" over image pixels, but are promptly detected by perceptual metrics like LPIPS. On the Deep Blending dataset, we observe thinner gaps, PSNR actually showing a small regression w.r.t. the baselines, although with low confidence margins. We suspect this might be related to the fact that Deep Blending contains many flat, untextured surfaces (see Fig. 4), that are particularly challenging to reconstruct accurately with 3DGS-like methods, independently of the ADC strategy being adopted. In the supplementary material, we provide an additional discussion about why some metrics do not reflect the actual perceived quality of the generated images." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 536, + 480, + 608 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 536, + 480, + 608 + ], + "spans": [ + { + "bbox": [ + 130, + 536, + 480, + 608 + ], + "type": "text", + "content": "Figure 4 contains a qualitative comparison between standard 3DGS and 3DGS augmented with our contributions (Ours). Areas with under-fitting artifacts are highlighted, showing how these are notably ameliorated by our approach. It is also worth noting that Ours effectively maintains the same quality as 3DGS in non-problematic areas, producing a more perceptually accurate reconstruction while using the same number of primitives (see Sec. 4.2)." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 623, + 269, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 623, + 269, + 635 + ], + "spans": [ + { + "bbox": [ + 131, + 623, + 269, + 635 + ], + "type": "text", + "content": "4.4 Ablation experiments" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 641, + 479, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 479, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 479, + 665 + ], + "type": "text", + "content": "In Tab. 4 we ablate the effects of Opacity Correction (OC, Sec. 3.3), Growth Control (GC, Sec. 3.4) and Opacity Regularization (OR, Sec. 3.5) on the Mip-NeRF" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 266, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 266, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 266, + 91, + 447, + 102 + ], + "type": "text", + "content": "Revising Densification in Gaussian Splitting" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 308 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 308 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 308 + ], + "type": "text", + "content": "360 dataset. In particular, we evaluate 3DGS augmented with each of these components (left side of the table), and our method with the components replaced by the corresponding baseline mechanism in 3DGS' standard ADC (right side of the table). First, we observe that OC, GC and OR all contribute to our method, as the Full version of Ours achieves the overall best results on all metrics, and removing them consistently degrades performance. Interestingly, Opacity Correction seems to have the largest impact here, as it produces both the largest increase in the scores when added to 3DGS, and the largest decrease when removed from Ours. Finally, Growth Control has a negative impact on 3DGS when utilized in isolation, while only slightly degrading the results when removed from Ours. Note that this observation doesn't detract from GC's usefulness as a strategy to control and limit the capacity of the model. We hypothesize that GC's negative effect on 3DGS might be a consequence of the fact that the standard, gradient-based densification score is actually a poor choice for comparing gaussians in terms of how soon they should be split or cloned (remember that GC ranks Gaussians based on their score)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 331, + 218, + 342 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 331, + 218, + 342 + ], + "spans": [ + { + "bbox": [ + 132, + 331, + 218, + 342 + ], + "type": "text", + "content": "4.5 Limitations" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 357, + 482, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 357, + 482, + 502 + ], + "spans": [ + { + "bbox": [ + 130, + 357, + 482, + 502 + ], + "type": "text", + "content": "While our method appears to be quite effective at solving under-fitting issues, these can still be present in especially difficult scenes (e.g. treehill in the Mip-NeRF 360 dataset, both scenes from the Deep Blending dataset). Focusing on the problematic areas that our ADC approach handles successfully, we observe that, while perceptually more \"correct\", the reconstruction there can still be quite inaccurate when closely compared to the ground truth (see e.g. the flowers scene in Fig. 4). We suspect both these issues might be related to 3DGS' intrinsic limits in handling i) strong view-dependent effects; ii) appearance variations across images; and iii) errors induced by the linear approximation in the Splitting operation (see Sec. 2). An interesting future direction could be to combine our approach with works that address these issues, e.g. Spec-Gaussian [26] for (i) and GS++ [7] for (iii)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 526, + 220, + 539 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 526, + 220, + 539 + ], + "spans": [ + { + "bbox": [ + 132, + 526, + 220, + 539 + ], + "type": "text", + "content": "5 Conclusion" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 557, + 482, + 666 + ], + "type": "text", + "content": "In this paper, we addressed the limitations of the Adaptive Density Control (ADC) mechanism in 3D Gaussian Splatting (3DGS), a scene representation method for high-quality, photorealistic rendering. Our main contribution is a more principled, pixel-error driven formulation for density control in 3DGS. We propose how to leverage a novel decision criterion for densification based on per-pixel errors and introduce a mechanism to control the total number of primitives generated per scene. We also correct a bias in the current opacity handling in ADC during cloning. Our approach leads to consistent and systematic improvements over previous methods, particularly in perceptual metrics like LPIPS." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 233, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 233, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 233, + 100 + ], + "type": "text", + "content": "Rota Bulò et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 197, + 126 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 197, + 126 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 197, + 126 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 137, + 481, + 664 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 138, + 137, + 481, + 180 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 137, + 481, + 180 + ], + "spans": [ + { + "bbox": [ + 138, + 137, + 481, + 180 + ], + "type": "text", + "content": "1. Barron, J.T., Mildenhall, B., Verbin, D., Srinivasan, P.P., Hedman, P.: Mipnerf 360: Unbounded anti-aliased neural radiance fields. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5470-5479 (2022)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 181, + 481, + 201 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 181, + 481, + 201 + ], + "spans": [ + { + "bbox": [ + 138, + 181, + 481, + 201 + ], + "type": "text", + "content": "2. Chen, Z., Wang, F., Liu, H.: Text-to-3d using gaussian splatting. arXiv preprint arXiv:2309.16585 (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 202, + 481, + 234 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 202, + 481, + 234 + ], + "spans": [ + { + "bbox": [ + 138, + 202, + 481, + 234 + ], + "type": "text", + "content": "3. Cheng, K., Long, X., Yang, K., Yao, Y., Yin, W., Ma, Y., Wang, W., Chen, X.: Gaussianpro: 3d gaussian splatting with progressive propagation. arXiv preprint arXiv:2402.14650 (2024)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 235, + 481, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 235, + 481, + 267 + ], + "spans": [ + { + "bbox": [ + 138, + 235, + 481, + 267 + ], + "type": "text", + "content": "4. Fridovich-Keil, S., Yu, A., Tancik, M., Chen, Q., Recht, B., Kanazawa, A.: Plenoxels: Radiance fields without neural networks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5501-5510 (2022)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 267, + 481, + 299 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 267, + 481, + 299 + ], + "spans": [ + { + "bbox": [ + 138, + 267, + 481, + 299 + ], + "type": "text", + "content": "5. Guédon, A., Lepetit, V.: Sugar: Surface-aligned gaussian splatting for efficient 3d mesh reconstruction and high-quality mesh rendering. arXiv preprint arXiv:2311.12775 (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 300, + 481, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 300, + 481, + 331 + ], + "spans": [ + { + "bbox": [ + 138, + 300, + 481, + 331 + ], + "type": "text", + "content": "6. Hedman, P., Philip, J., Price, T., Frahm, J.M., Drettakis, G., Brostow, G.: Deep blending for free-viewpoint image-based rendering. ACM Transactions on Graphics (ToG) 37(6), 1-15 (2018)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 332, + 481, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 332, + 481, + 353 + ], + "spans": [ + { + "bbox": [ + 138, + 332, + 481, + 353 + ], + "type": "text", + "content": "7. Huang, L., Bai, J., Guo, J., Guo, Y.: " + }, + { + "bbox": [ + 138, + 332, + 481, + 353 + ], + "type": "inline_equation", + "content": "\\mathrm{Gs} + +" + }, + { + "bbox": [ + 138, + 332, + 481, + 353 + ], + "type": "text", + "content": ": Error analyzing and optimal gaussian splatting. arXiv preprint arXiv:2402.00752 (2024)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 354, + 481, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 354, + 481, + 386 + ], + "spans": [ + { + "bbox": [ + 138, + 354, + 481, + 386 + ], + "type": "text", + "content": "8. Keetha, N., Karhade, J., Jatavallabhula, K.M., Yang, G., Scherer, S., Ramanan, D., Luiten, J.: Splatam: Splat, track & map 3d gaussians for dense rgb-d slam. arXiv preprint arXiv:2312.02126 (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 386, + 481, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 386, + 481, + 418 + ], + "spans": [ + { + "bbox": [ + 138, + 386, + 481, + 418 + ], + "type": "text", + "content": "9. Kerbl, B., Kopanas, G., Leimkuhler, T., Drettakis, G.: 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics 42(4) (July 2023), https://repo-sam.inria.fr/fungraph/3d-gaussian-splatting/" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 418, + 481, + 450 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 418, + 481, + 450 + ], + "spans": [ + { + "bbox": [ + 138, + 418, + 481, + 450 + ], + "type": "text", + "content": "0. Knapitsch, A., Park, J., Zhou, Q.Y., Koltun, V.: Tanks and temples: Benchmarking large-scale scene reconstruction. ACM Transactions on Graphics (ToG) 36(4), 1-13 (2017)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 450, + 481, + 472 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 450, + 481, + 472 + ], + "spans": [ + { + "bbox": [ + 138, + 450, + 481, + 472 + ], + "type": "text", + "content": "1. Kocabas, M., Chang, J.H.R., Gabriel, J., Tuzel, O., Ranjan, A.: Hugs: Human gaussian splats. arXiv preprint arXiv:2311.17910 (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 472, + 481, + 493 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 472, + 481, + 493 + ], + "spans": [ + { + "bbox": [ + 138, + 472, + 481, + 493 + ], + "type": "text", + "content": "2. Lee, J.C., Rho, D., Sun, X., Ko, J.H., Park, E.: Compact 3d gaussian representation for radiance field. arXiv preprint arXiv:2311.13681 (2023)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 493, + 481, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 493, + 481, + 514 + ], + "spans": [ + { + "bbox": [ + 138, + 493, + 481, + 514 + ], + "type": "text", + "content": "3. Lei, J., Wang, Y., Pavlakos, G., Liu, L., Daniilidis, K.: Gart: Gaussian articulated template models. arXiv preprint arXiv:2311.16099 (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 514, + 481, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 514, + 481, + 547 + ], + "spans": [ + { + "bbox": [ + 138, + 514, + 481, + 547 + ], + "type": "text", + "content": "4. Lu, T., Yu, M., Xu, L., Xiangli, Y., Wang, L., Lin, D., Dai, B.: Scaffold-gs: Structured 3d gaussians for view-adaptive rendering. arXiv preprint arXiv:2312.00109 (2023)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 547, + 481, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 547, + 481, + 568 + ], + "spans": [ + { + "bbox": [ + 138, + 547, + 481, + 568 + ], + "type": "text", + "content": "5. Luiten, J., Kopanas, G., Leibe, B., Ramanan, D.: Dynamic 3d gaussians: Tracking by persistent dynamic view synthesis. arXiv preprint arXiv:2308.09713 (2023)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 568, + 481, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 568, + 481, + 589 + ], + "spans": [ + { + "bbox": [ + 138, + 568, + 481, + 589 + ], + "type": "text", + "content": "6. Matsuki, H., Murai, R., Kelly, P.H., Davison, A.J.: Gaussian splatting slam. arXiv preprint arXiv:2312.06741 (2023)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 138, + 590, + 481, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 590, + 481, + 621 + ], + "spans": [ + { + "bbox": [ + 138, + 590, + 481, + 621 + ], + "type": "text", + "content": "7. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. In: ECCV. p. 405-421 (2020)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 138, + 621, + 481, + 664 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 621, + 481, + 664 + ], + "spans": [ + { + "bbox": [ + 138, + 621, + 481, + 664 + ], + "type": "text", + "content": "8. Müller, T., Evans, A., Schied, C., Keller, A.: Instant neural graphics primitives with a multiresolution hash encoding. ACM Trans. Graph. 41(4), 102:1-102:15 (Jul 2022). https://doi.org/10.1145/3528223.3530127, https://doi.org/10.1145/3528223.3530127" + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 266, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 266, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 266, + 91, + 447, + 102 + ], + "type": "text", + "content": "Revising Densification in Gaussian Splitting" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 117, + 480, + 521 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 133, + 117, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 117, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 133, + 117, + 480, + 138 + ], + "type": "text", + "content": "19. Saito, S., Schwartz, G., Simon, T., Li, J., Nam, G.: Relightable gaussian codec avatars. arXiv preprint arXiv:2312.03704 (2023)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 140, + 480, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 140, + 480, + 160 + ], + "spans": [ + { + "bbox": [ + 132, + 140, + 480, + 160 + ], + "type": "text", + "content": "20. Snavely, N., Seitz, S.M., Szeliski, R.: Photo tourism: exploring photo collections in 3d. In: ACM siggraph 2006 papers, pp. 835-846 (2006)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 161, + 480, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 161, + 480, + 182 + ], + "spans": [ + { + "bbox": [ + 133, + 161, + 480, + 182 + ], + "type": "text", + "content": "21. Tang, J., Ren, J., Zhou, H., Liu, Z., Zeng, G.: Dreamgaussian: Generative gaussian splatting for efficient 3d content creation. arXiv preprint arXiv:2309.16653 (2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 183, + 480, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 183, + 480, + 215 + ], + "spans": [ + { + "bbox": [ + 133, + 183, + 480, + 215 + ], + "type": "text", + "content": "22. Wu, G., Yi, T., Fang, J., Xie, L., Zhang, X., Wei, W., Liu, W., Tian, Q., Wang, X.: 4d gaussian splatting for real-time dynamic scene rendering. arXiv preprint arXiv:2310.08528 (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 216, + 480, + 247 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 216, + 480, + 247 + ], + "spans": [ + { + "bbox": [ + 132, + 216, + 480, + 247 + ], + "type": "text", + "content": "23. Xie, T., Zong, Z., Qiu, Y., Li, X., Feng, Y., Yang, Y., Jiang, C.: Physgaus-sian: Physics-integrated 3d gaussians for generative dynamics. arXiv preprint arXiv:2311.12198 (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 249, + 480, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 249, + 480, + 270 + ], + "spans": [ + { + "bbox": [ + 132, + 249, + 480, + 270 + ], + "type": "text", + "content": "24. Yan, C., Qu, D., Wang, D., Xu, D., Wang, Z., Zhao, B., Li, X.: Gs-slam: Dense visual slam with 3d gaussian splatting. arXiv preprint arXiv:2311.11700 (2023)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 271, + 480, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 271, + 480, + 303 + ], + "spans": [ + { + "bbox": [ + 132, + 271, + 480, + 303 + ], + "type": "text", + "content": "25. Yang, Z., Yang, H., Pan, Z., Zhu, X., Zhang, L.: Real-time photorealistic dynamic scene representation and rendering with 4d gaussian splatting. arXiv preprint arXiv:2310.10642 (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 304, + 480, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 304, + 480, + 335 + ], + "spans": [ + { + "bbox": [ + 132, + 304, + 480, + 335 + ], + "type": "text", + "content": "26. Yang, Z., Gao, X., Sun, Y., Huang, Y., Lyu, X., Zhou, W., Jiao, S., Qi, X., Jin, X.: Spec-gaussian: Anisotropic view-dependent appearance for 3d gaussian splatting. arXiv preprint arXiv:2402.15870 (2024)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 337, + 480, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 337, + 480, + 357 + ], + "spans": [ + { + "bbox": [ + 132, + 337, + 480, + 357 + ], + "type": "text", + "content": "27. Ye, M., Danelljan, M., Yu, F., Ke, L.: Gaussian grouping: Segment and edit anything in 3d scenes. arXiv preprint arXiv:2312.00732 (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 358, + 480, + 390 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 358, + 480, + 390 + ], + "spans": [ + { + "bbox": [ + 132, + 358, + 480, + 390 + ], + "type": "text", + "content": "28. Yi, T., Fang, J., Wu, G., Xie, L., Zhang, X., Liu, W., Tian, Q., Wang, X.: Gaussian dreamer: Fast generation from text to 3d gaussian splatting with point cloud priors. arXiv preprint arXiv:2310.08529 (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 391, + 480, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 391, + 480, + 413 + ], + "spans": [ + { + "bbox": [ + 132, + 391, + 480, + 413 + ], + "type": "text", + "content": "29. Yu, Z., Chen, A., Huang, B., Sattler, T., Geiger, A.: Mip-splatting: Alias-free 3d gaussian splatting. arXiv preprint arXiv:2311.16493 (2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 414, + 480, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 414, + 480, + 435 + ], + "spans": [ + { + "bbox": [ + 132, + 414, + 480, + 435 + ], + "type": "text", + "content": "30. Yugay, V., Li, Y., Gevers, T., Oswald, M.R.: Gaussian-slam: Photo-realistic dense slam with gaussian splatting. arXiv preprint arXiv:2312.10070 (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 436, + 480, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 436, + 480, + 468 + ], + "spans": [ + { + "bbox": [ + 132, + 436, + 480, + 468 + ], + "type": "text", + "content": "31. Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual metric. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 586-595 (2018)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 468, + 480, + 490 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 468, + 480, + 490 + ], + "spans": [ + { + "bbox": [ + 132, + 468, + 480, + 490 + ], + "type": "text", + "content": "32. Zielonka, W., Bagautdinov, T., Saito, S., Zollhöfer, M., Thies, J., Romero, J.: Drivable 3d gaussian avatars. arXiv preprint arXiv:2311.08581 (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 491, + 480, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 491, + 480, + 521 + ], + "spans": [ + { + "bbox": [ + 132, + 491, + 480, + 521 + ], + "type": "text", + "content": "33. Zwicker, M., Pfister, H., van Baar, J., Gross, M.: Ewa volume splatting. In: Proceedings Visualization, 2001. VIS '01. pp. 29-538 (2001). https://doi.org/10.1109/VISUAL.2001.964490" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 233, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 233, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 233, + 100 + ], + "type": "text", + "content": "Rota Bulò et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Revisit Anything_ Visual Place Recognition via Image Segment Retrieval/32b8503b-ae81-47c5-bdc0-0efb4201113d_content_list.json b/2024/Revisit Anything_ Visual Place Recognition via Image Segment Retrieval/32b8503b-ae81-47c5-bdc0-0efb4201113d_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..cbdfd2f35ac5a85c8f5f2e10c4a3156562992c35 --- /dev/null +++ b/2024/Revisit Anything_ Visual Place Recognition via Image Segment Retrieval/32b8503b-ae81-47c5-bdc0-0efb4201113d_content_list.json @@ -0,0 +1,1584 @@ +[ + { + "type": "text", + "text": "Revisit Anything: Visual Place Recognition via Image Segment Retrieval", + "text_level": 1, + "bbox": [ + 232, + 140, + 772, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Kartik Garg $^{1}$ , Sai Shubodh Pulgilla $^{2}$ , Shishir Kolathaya $^{1}$ , Madhava Krishna $^{2}$ , and Sourav Garg $^{3}$", + "bbox": [ + 282, + 212, + 718, + 244 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Indian Institute of Science (IISc), Bengaluru, India \n $^{2}$ International Institute of Information Technology, Hyderabad, India \n $^{3}$ University of Adelaide, Australia", + "bbox": [ + 215, + 253, + 759, + 297 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "kartikgarg@iisc.ac.in, p.saishubodh@gmail.com, shishirk@iisc.ac.in, mkrishna@iiit.ac.in, sourav.garg@adelaide.edu.au", + "bbox": [ + 217, + 297, + 785, + 325 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Accurately recognizing a revisited place is crucial for embodied agents to localize and navigate. This requires visual representations to be distinct, despite strong variations in camera viewpoint and scene appearance. Existing visual place recognition pipelines encode the whole image and search for matches. This poses a fundamental challenge in matching two images of the same place captured from different camera viewpoints: the similarity of what overlaps can be dominated by the dissimilarity of what does not overlap. We address this by encoding and searching for image segments instead of the whole images. We propose to use open-set image segmentation to decompose an image into 'meaningful' entities (i.e., things and stuff). This enables us to create a novel image representation as a collection of multiple overlapping subgraphs connecting a segment with its neighboring segments, dubbed SuperSegment. Furthermore, to efficiently encode these SuperSegments into compact vector representations, we propose a novel factorized representation of feature aggregation. We show that retrieving these partial representations leads to significantly higher recognition recall than the typical whole image based retrieval. Our segments-based approach, dubbed SegVLAD, sets a new state-of-the-art in place recognition on a diverse selection of benchmark datasets, while being applicable to both generic and task-specialized image encoders. Finally, we demonstrate the potential of our method to \"revisit anything\" by evaluating our method on an object instance retrieval task, which bridges the two disparate areas of research: visual place recognition and object-goal navigation, through their common aim of recognizing goal objects specific to a place. Source code: https://github.com/AnyLoc/Revisit-Anything.", + "bbox": [ + 261, + 364, + 743, + 696 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Visual Place Recognition $\\cdot$ Image Segmentation $\\cdot$ Robotics", + "bbox": [ + 261, + 710, + 686, + 724 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 752, + 357, + 767 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Visual Place Recognition (VPR) is an important capability for embodied agents to localize and navigate autonomously. A predominant solution for VPR is to encode", + "bbox": [ + 212, + 784, + 787, + 816 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Equal contribution", + "bbox": [ + 243, + 825, + 370, + 840 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "an image into a global vector and retrieve similar vectors as coarse localization hypotheses [19, 44, 60, 67]. Thus, for almost a decade, researchers have focused on learning/finetuning image encoders so that global descriptors are induced with invariance to appearance [3,56,69], viewpoint [3,8], and clutter [27]. On the other hand, there is a vast literature on local descriptors (point/pixel-level), mainly relevant for geometric reranking in hierarchical VPR [11, 16, 25, 57, 68]. In the middle of local and global descriptors exists a variety of methods that use regions/patches [4, 25], lines/planes [17], objects (things/stuff) [21, 31, 45, 65], and segments [26, 33, 50] to represent images. However, these methods are still only aimed at either improving global descriptors based coarse retrieval or local feature matching based reranking. In this work, in contrast to conventional retrieval-based VPR, we explore an alternative: retrieval via encoding segments instead of the whole image. This is particularly enabled by recent advances in open-set image segmentation [37] which can meaningfully deconstruct a place into 'things' (and/or 'stuff') [10]. Thus, we reformulate the VPR problem of revisiting places as that to revisiting things by enabling recognition of these specific things within the context of their place. While such a segment-level place recognition approach provides a direct link to higher-level semantic tasks, such as object-goal navigation [12, 20, 24, 43], it also addresses a fundamental issue in matching partially-overlapping images from across significant viewpoint change. Segments-based partial image representation avoids the mismatches caused by the whole-image representation when the similarity of what overlaps is dominated by the dissimilarity of what does not overlap. Our novel segments-based VPR method, dubbed SegVLAD (Segment based Vector of Locally Aggregated Descriptors), is illustrated in Figure 1, which makes the following novel contributions:", + "bbox": [ + 215, + 147, + 787, + 494 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. an image representation as a collection of multiple overlapping subgraphs of segments, dubbed SuperSegments, which enables accurate recognition across partially-overlapping images;", + "2. a factorized representation of feature aggregation to effectively accommodate both segment-level information as well as segment neighborhood information; and", + "3. a similarity-weighted ranking method to convert segment-level retrieval into image-level retrieval." + ], + "bbox": [ + 222, + 507, + 782, + 609 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Using a diverse set of data sources, we demonstrate that our proposed segments-based retrieval enables place recognition under wide viewpoint variations, where global descriptor based retrieval suffers. SegVLAD achieves a new state-of-the-art on multiple challenging datasets. We also introduce an evaluation of our method on an instance-level object retrieval task – a novel capability of our pipeline unlike conventional VPR methods. We conduct several ablations and parameter studies to justify the design choices and emphasize the effectiveness of our method as an open-set segments-based coarse retriever.", + "bbox": [ + 215, + 627, + 787, + 746 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2 Related Works", + "text_level": 1, + "bbox": [ + 217, + 773, + 375, + 791 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Image retrieval-based Visual Place Recognition (VPR) is a well-established area of research in visual localization [19,44,60,67]. It is important both during mapping for loop", + "bbox": [ + 215, + 809, + 785, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "K. Garg et al.", + "bbox": [ + 271, + 114, + 354, + 128 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "closures [67] as well as for relocalization [51, 58]. The underlying task in both the scenarios remains the same: how to recognize a previously seen place. The state-of-the-art methods in VPR use a global descriptor-based approach which converts an image into a compact vector to enable fast retrieval [2, 3, 6, 29, 32, 56]. The top retrieved hypotheses are often then re-ranked through compute-intensive local feature matching using geometric information [11, 16, 25, 33, 57, 68]. In contrast to previous approaches, we aim to explore image segment level descriptors in this work. This representation falls between point-based local descriptors and the whole-image based global descriptors. Our approach can be considered as 'semi-global', with the proposed segment (and SuperSegment) based descriptors being a 'spatially-reduced' form of whole-image global representation. This is motivated by our hypothesis that to deal with viewpoint variations in VPR with partially-overlapping images, we need a way to partially represent and match them.", + "bbox": [ + 218, + 145, + 787, + 340 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.1 Whole Image Encoders", + "text_level": 1, + "bbox": [ + 217, + 363, + 419, + 378 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Earlier works in whole-image representation used methods like Gist [48], BoW (Bag of Word) [63], and VLAD (Vector of Locally Aggregated Descriptors) [30], often defined using hand-crafted features such as SIFT [41]. In recent years, deep learning based methods have demonstrated remarkable performance, with initial successful methods like NetVLAD [3] now rapidly outperformed by better alternatives such as CosPlace [6], MixVPR [2], EigenPlaces [8], TransVPR [68], and more recently AnyLoc [32], SALAD [29] and VLAD-BuFF [35]. All these learning-based methods improve different aspects of representation learning: training datasets [1, 6, 69], objective/loss functions [2, 8], aggregation methods [29, 54, 56, 68], and generalization [32]. Our approach complements these existing methods as we mainly focus on the use of segment-based information, where the segments can be described by any of the image encoders from the aforementioned techniques. In particular, we demonstrate that both - an off-the-shelf encoder, e.g., DINOv2-AnyLoc [32, 49] or that finetuned specifically for the VPR task, e.g., DINOv2-NetVLAD [35] - can be used in conjunction with our segment-based approach to further elevate place recognition capability.", + "bbox": [ + 218, + 387, + 787, + 614 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2 Region/Patch Based Methods", + "text_level": 1, + "bbox": [ + 217, + 633, + 460, + 648 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "There exist several methods that employ region or patch level information to enhance representational power [13, 14, 34, 40, 52, 71, 73]. However, most of these methods only use this additional information to generate a single (or concatenated) compact vector representation of an image. Other methods such as Patch-NetVLAD [25] create multiple features per image but their primary purpose is to perform local matching based reranking. In contrast to these methods, we aim to use multiple segment descriptors per image to directly retrieve from a database of segments, without using any geometric information or reranking. The motivation behind this stems from the very nature of hierarchical VPR pipelines: reranking recall is upper bounded by the coarse retriever. A better coarse retriever can improve reranking performance without needing to rerank from a longer list of top hypotheses. MultiVLAD [4] is similar to our method in the spirit of retrieving multiple features per query image. However, like aforementioned", + "bbox": [ + 218, + 657, + 787, + 839 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "Revisit Anything", + "bbox": [ + 627, + 114, + 730, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "methods, MultiVLAD defines regions arbitrarily, whereas we use image segments obtained from Segment Anything Model (SAM) [37] which are semantically meaningful.", + "bbox": [ + 215, + 146, + 785, + 176 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.3 Segments-Enhanced Methods", + "text_level": 1, + "bbox": [ + 215, + 196, + 464, + 212 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "There exist several methods that use semantic segmentation information to improve VPR, as also surveyed in [22]. These methods vary in terms of type of segmentation used and the specific ways in which it is integrated in the VPR pipeline, e.g., planes [17], objects [15,31,45], landmarks [65], outdoor semantics [21,23,46,47], utility/confusion based [33, 38], domain adaptation [26] and even learning to segment for VPR [50]. However, neither these methods aim to perform segment-level retrieval nor do they use open-set segmentation. We also review two concurrent works: MESA [75] and Region-Revisited [61]. Similar to our method, they both use SAM to segment images but for different specific tasks. MESA [75] proposes a graph-based local feature/area matching method to obtain point correspondences. Our method complements this, as we perform coarse retrieval for VPR, which could potentially use MESA for reranking. Regions Revisited [61] delves into the advantages of using SAM masks in conjunction with SLIC [36] to improve semantic segmentation, activity recognition and object category retrieval. In contrast, we aim to improve instance-level recognition by recognizing specific things belonging to specific places that a robot encounters during a revisit. Similar to our work, [20] creates an image sequence-based topological graph of segments where its segment neighbourhood aggregation is based on average pooling, similar to [61]. In Section 5.3, we show that such segment average pooling deteriorates recognition performance for the VPR task.", + "bbox": [ + 215, + 219, + 785, + 503 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.4 Open-set VPR", + "text_level": 1, + "bbox": [ + 215, + 525, + 356, + 540 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Researchers have recently started to shift their focus to open-set, generally-applicable techniques, including that for VPR [31, 32, 45, 62]. FM-Loc [45] uses GPT [9] to recognize object and place labels, whereas [31] uses CLIP [55] for open-set place recognition. LIP-Loc [62] proposes pretraining for cross-modal VPR, but is limited in its zero-shot capabilities. AnyLoc [32] proposes to use DINOv2 with domain-level vocabularies and hard-assignment based VLAD. It achieves state-of-the-art performance particularly on non-streetview datasets, where current VPR-trained methods tend to fail. In this work, we propose a generally-applicable approach which is built on top of models like SAM [37] and DINOv2 [49], and works with both VPR-agnostic [32] and VPR-specific [35] backbone models. We particularly aim for a new paradigm in retrieval based VPR, where we move away from the conventional whole image global descriptors to segments based descriptors and retrieval, which achieves a new state-of-the-art on diverse domains under wide viewpoint variations.", + "bbox": [ + 215, + 547, + 785, + 744 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3 Proposed Approach", + "text_level": 1, + "bbox": [ + 215, + 765, + 415, + 782 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Despite recent advances in place recognition, viewpoint variations continue to be an open challenge for an embodied agent to recognize the same specific things in its environment. Current methods in visual place recognition tackle this problem by converting", + "bbox": [ + 215, + 794, + 785, + 839 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 116, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "K. Garg et al.", + "bbox": [ + 271, + 114, + 354, + 128 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/43e8379a038ad16d7f713e22f77c5f5c6bf8270f93a5ce55510a261c89221503.jpg", + "image_caption": [ + "Fig. 1: Overview of our segment-retrieval based VPR pipeline, dubbed SegVLAD: We use open-set segmentor (SAM) to extract segment masks, which are converted into SuperSegments using the neighbouring image segments. Using pixel-level DINOv2 descriptors with VLAD-based aggregation, we obtain SuperSegment descriptors, which are matched against a flat index of SuperSegment descriptors obtained from all the images of the entire reference database." + ], + "image_footnote": [], + "bbox": [ + 215, + 148, + 787, + 335 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "an image as a whole into a global descriptor, which does not explicitly deal with the problem of partial visual overlap caused by viewpoint variations. We propose an alternative solution by representing images partially with the help of image segments. In the following subsections, we describe our representation and retrieval method, which deviates from the conventional VPR techniques but creates a new capability in terms of recognizing objects/things that constitute a place.", + "bbox": [ + 212, + 450, + 787, + 542 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1 Problem Formulation", + "text_level": 1, + "bbox": [ + 215, + 566, + 408, + 579 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We represent an image as a set of segment descriptors instead of a single global descriptor. For an image $I$ , we obtain binary image segment masks $M \\in \\{0,1\\}^{S \\times N}$ and dense pixel-level descriptors $f_{p} \\in \\mathcal{R}^{D}$ , where $S$ represents the number of segments per image, $D$ is the descriptor dimension, and $p \\in [1,N]$ represents spatial elements across the width $(W)$ and height $(H)$ of the image encoder's output, flattened into $N = W \\times H$ for convenience. Figure 1 shows an illustration of our proposed pipeline, as explained in the following subsections.", + "bbox": [ + 212, + 592, + 787, + 699 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2 Super Segments", + "text_level": 1, + "bbox": [ + 215, + 722, + 370, + 738 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Humans are remarkable at visual recognition, where existing studies suggest that we often leverage spatial associations among objects in an environment to represent it internally [5, 28]. This enables us to distinguish between two different scenes through the surrounding context of the objects of interest. In this work, we imbibe this context through the spatial neighbourhood of the image segments. For each image, we construct a graph of segments through their pixel centers using Delaunay Triangulation.", + "bbox": [ + 212, + 750, + 787, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "Revisit Anything", + "bbox": [ + 627, + 114, + 730, + 128 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 114, + 785, + 126 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/dced319d20016bdc56ce97d6c7c1379b3df2ff98a3f48ddbd2f9a35a4e04a343.jpg", + "image_caption": [ + "Fig. 2: Neighborhood expansion (Eq. 1) of a window in the leftmost image to the whole building in the rightmost image, progressing from no neighborhood aggregation to a third-order aggregation. This neighborhood expansion is in stark contrast with a typical regular grid- or patch-based approach which may not capture semantically-meaningful SuperSegments." + ], + "image_footnote": [], + "bbox": [ + 217, + 143, + 785, + 227 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/27792164f535fc6e55aef41b060198439100414bd496b51ad08fbb48df3dba01.jpg", + "image_caption": [ + "Fig. 3: Illustration of four SuperSegments obtained from the same image. All four of these spatially overlap with each other, which is different from coarse segmentation methods that do not typically allow overlap across segments." + ], + "image_footnote": [], + "bbox": [ + 217, + 306, + 785, + 371 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "This provides us with a binary adjacency matrix $A \\in \\{0,1\\}^{S \\times S}$ to define the neighborhood for individual segments. We use this adjacency information to expand the context of individual segments to generate new SuperSegment masks $(\\mathcal{M})$ as below:", + "bbox": [ + 212, + 450, + 787, + 496 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {M} _ {S \\times N} = ^ {1} \\left(A _ {S \\times S} ^ {o} \\cdot M _ {S \\times N}\\right) \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 398, + 502, + 785, + 520 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $o \\geq 0$ refers to the order for expanding the neighborhood by multiplying the adjacency matrix by itself as $A^{o + 1} = A^o \\cdot A$ . This is matrix-multiplied with the original segmentation masks $M$ to expand the neighborhood at pixel level. $\\mathcal{M}$ is obtained after element-wise binarization (denoted with $\\mathbb{1}()$ ) so that all pixels in the SuperSegment mask may only contribute once to the subsequent feature aggregation. In Figure 2, we illustrate the extent of image area covered with different orders of mask expansion. Unlike, a patch or regular grid-based approach, the expanded mask of the window in the leftmost image covers a meaningful entity (building) in the rightmost image. Our approach to creating SuperSegments differs from coarse segmentation methods or superpixels in terms of the 'self-overlap'. By expanding neighborhood of each individual segment, we obtain several partially overlapping SuperSegments. A coarse segmentor will need to make assumptions about the right sub-segments to be coalesced so that it can enable accurate recognition from a different viewpoint, which could otherwise lead to the same limitation as that of the whole-image descriptors. Figure 3 presents examples of multiple overlapping SuperSegments from the same image.", + "bbox": [ + 212, + 527, + 787, + 755 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.3 SuperSegment Descriptors", + "text_level": 1, + "bbox": [ + 214, + 773, + 444, + 787 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this section, we describe our feature aggregation method to obtain SuperSegment descriptors. Recent state-of-the-art method AnyLoc [32] demonstrated that using off-the-shelf powerful image encoders such as DINOv2 with hard assignment based VLAD", + "bbox": [ + 212, + 794, + 787, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 116, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "K. Garg et al.", + "bbox": [ + 271, + 114, + 354, + 128 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "aggregation achieves superior recognition performance. However, AnyLoc does not use segmentation information and only operates at the whole-image level. More recently, [61] showed that average pooling works well for segment-level descriptors, but it didn't consider segment neighborhood information. In this work, we propose a unified formulation for feature aggregation that can easily switch across segments, segment neighborhood and the whole image as well as different aggregation types (see supplementary for details). This simply extends Equation 1 as below:", + "bbox": [ + 214, + 146, + 787, + 252 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\nF _ {S \\times D} = ^ {1} \\left(A _ {S \\times S} ^ {o} \\cdot M _ {S \\times N}\\right) \\cdot T _ {N \\times D} \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 377, + 266, + 785, + 284 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $T$ represents the features to be aggregated. By replacing $A$ and $M$ with ones matrices, one can obtain the whole-image global descriptor. For methods like Global Average Pooling (GAP), $T$ can be directly used as the output of the image encoder. In our work, we use Hard-VLAD, for which $T$ is the residual feature matrix per cluster and is obtained as below with respect to each of the cluster centers $c_k$ :", + "bbox": [ + 214, + 300, + 787, + 376 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\nT _ {N _ {k} \\times D} ^ {k} = \\left\\{\\alpha_ {k} \\left(f _ {p}\\right) \\left(f _ {p} - c _ {k}\\right) \\mid \\alpha_ {k} \\left(f _ {p}\\right) = 1 \\right\\} \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 352, + 390, + 785, + 409 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $\\alpha_{k}(f_{p})\\in \\{0,1\\}$ is 1 if $f_{p}$ belongs to $c_{k}$ . The cluster centers (vocabulary) can be constructed using the map or the domain [32]. The SuperSegment VLAD descriptors obtained from Eq. 2 are intra-normalized, concatenated across clusters and then finally l2-normalized, following existing works [3, 32].", + "bbox": [ + 214, + 422, + 787, + 484 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.4 Image Retrieval via Segments", + "text_level": 1, + "bbox": [ + 215, + 516, + 465, + 532 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Existing global descriptor based VPR techniques produce a single vector per image to search against a database of reference image vectors. In our method, we obtain multiple SuperSegment descriptors per image. We perform retrieval at segment-level, that is, we search for the top matches for each query segment against a flat index of all segments from all the images of the reference database/map. To evaluate in the form of image retrieval-based VPR, we convert the top retrieved segment indices across all segments of a query image into top reference image indices. This is achieved through a weighted frequency measure (i.e., weighted bin/word counting). We first map the top $K'$ retrieved segment indices for each of the query segments $s \\in [1, S]$ to their respective reference image indices, denoted with $r$ . Then, for each of the unique retrieved image indices $r_j$ , we accumulate its segment similarity $\\theta$ and then use the cumulative similarity score $\\hat{\\theta}$ to rank the image indices to obtain the top image match $r_j^*$ :", + "bbox": [ + 214, + 551, + 787, + 737 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\nr _ {j} ^ {*} = \\underset {r _ {j}} {\\operatorname {a r g m a x}} \\hat {\\theta} \\left(r _ {j}\\right); \\quad \\hat {\\theta} \\left(r _ {j}\\right) = \\sum_ {s = 1} ^ {S} \\sum_ {k = 1} ^ {K ^ {\\prime}} \\theta_ {s k} \\cdot \\mathbb {1} _ {\\left\\{r _ {s k} = r _ {j} \\right\\}} \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 318, + 752, + 785, + 796 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In Section 5.3, we compare our similarity-weighted ranking with other alternatives based on frequency or similarity alone.", + "bbox": [ + 214, + 809, + 787, + 840 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "Revisit Anything", + "bbox": [ + 627, + 114, + 730, + 128 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 114, + 785, + 126 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4 Experimental Setup", + "text_level": 1, + "bbox": [ + 217, + 143, + 415, + 162 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Datasets: VPR datasets are in abundance, as can be found in several benchmarks including VPR-Bench [74], Deep Visual GeoLocalization Benchmark [7], and AnyLoc [32]. In this work, we used a variety of datasets covering both outdoor and indoor environments. Outdoor datasets include Pitts30k [66], AmsterTime [72], Mapillary Street Level Sequences (MSLS) [69], SF-XL [6], VPAir [59], Revisted Oxford5K and Revisited Paris6k [53]. Indoor datasets include Baidu Mall [64], 17Places [76] and InsideOut [27]. Additional datasets-related details are provided in the supplementary.", + "bbox": [ + 217, + 176, + 785, + 282 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Evaluation and Benchmarking: We evaluate our method as an image retrieval task using Recall@K metric, where top K' (= 50) retrieved segments per query segment are used to obtain top K (= 5) images (see Eq. 4). We compare against the most recent and high-performing VPR baseline methods. This includes CosPlace [6], MixVPR [2] and EigenPlaces [8], which are trained on large-scale urban datasets for VPR tasks. We further include two very recent state-of-the-art methods that use DINOv2 as the backbone. These include AnyLoc [32] which uses an off-the-shelf DINOv2 model and SALAD [29] which uses a finetuned DINOv2 backbone. Given the dichotomy between general-purpose VPR benchmarking of AnyLoc and the typical outdoor-focused benchmarking in SALAD and other methods, we evaluated our method using two different backbones. a) SegVLAD-PreT: we use the same backbone and aggregation as AnyLoc, i.e., off-the-shelf pretrained DINOv2 (ViT-G) backbone with hard VLAD assignment, but the key difference is in the use of SuperSegments for our method as opposed to whole-image description of AnyLoc. b) SegVLAD-FineT: since our default aggregation method is VLAD, we use a finetuned DINOv2 (ViT-B) backbone which is similar to SALAD but replaces its aggregation layer with the original NetVLAD aggregation [3] using 64 clusters, as described in [35]. We use this finetuned backbone with hard VLAD based assignment, similar to AnyLoc. For both these models, we reduce the descriptor dimensions of the VLAD descriptor to 1024 using PCA, as commonly done in previous works [3, 32]. We train PCA transform in a map-specific manner using the database images of the dataset.", + "bbox": [ + 217, + 297, + 785, + 614 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5 Results", + "text_level": 1, + "bbox": [ + 217, + 637, + 310, + 654 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We first present benchmark comparison of our method against state-of-the-art VPR methods. This is followed by detailed analysis of our proposed aggregation technique. Lastly, we demonstrate results on a downstream task of Object of Interest (OOI) retrieval, showcasing the versatility of our method.", + "bbox": [ + 217, + 670, + 785, + 729 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.1 State-of-the-art comparisons", + "text_level": 1, + "bbox": [ + 217, + 752, + 455, + 768 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 1 presents Recall@1/5 comparison against state-of-the-art VPR methods on standard outdoor street-view datasets, which are similar to the typical training datasets used for VPR [2,3,6]. Table 2 covers 'out-of-distribution' datasets, inspired by AnyLoc [32], covering indoor environments (Baidu Mall and 17 Places), aerial imagery (VPAir),", + "bbox": [ + 217, + 777, + 785, + 839 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "K. Garg et al.", + "bbox": [ + 271, + 114, + 354, + 128 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/770410ddd0e8122f23a2f3949c8b0138f3dd30b9246759bdb6c73adcc1061500.jpg", + "table_caption": [ + "Table 1: Recall@ 1/5 benchmark comparison on outdoor street-view datasets." + ], + "table_footnote": [], + "table_body": "
MethodPitts-30KMSLS SFMSLS CPHSF-XL ValRO5k MedRO5k HardRP6k MedRP6k Hard
CosPlace90.4/95.793.4/97.584.9/92.094.6/97.685.7/87.127.1/45.794.3/95.77.1/15.7
MixVPR91.5/95.591.3/95.987.1/92.487.8/93.868.6/80.032.9/54.394.3/10010.0/32.9
EigenPlaces92.6/96.792.6/97.187.1/92.896.4/98.285.7/88.642.8/57.195.7/98.64.3/11.4
AnyLoc87.7/94.783.4/94.679.9/89.184.4/91.988.6/92.940.0/58.697.1/10011.4/44.3
SALAD92.6/96.591.7/97.192.3/96.193.6/97.382.9/90.037.1/54.395.7/98.614.3/58.6
SegVLAD-PreT86.7/94.288.4/94.281.7/90.790.9/96.491.4/95.760.0/81.494.3/1008.6/48.6
SegVLAD-FineT93.2/96.894.6/97.190.9/95.794.9/98.187.1/95.751.4/70.095.7/10010.0/48.6
", + "bbox": [ + 246, + 170, + 754, + 282 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/8170fe01fbfa82cace38a56ea5e323da47967078504fe75c7119809c2d98cb35.jpg", + "table_caption": [ + "Table 2: Recall@ 1/5 benchmark comparison on 'out-of-distribution' datasets." + ], + "table_footnote": [], + "table_body": "
MethodBaiduAmstTimeInsideOut17PlacesVPAir
CosPlace41.6/55.047.7/69.80.2/2.081.3/88.24.6/13.7
MixVPR64.4/80.340.2/59.10.0/1.885.2/90.16.8/16.1
EigenPlaces56.5/72.848.9/69.50.4/1.483.0/90.16.5/17.9
AnyLoc75.2/87.650.3/73.02.4/8.095.3/97.366.7/79.2
SALAD74.8/86.555.4/75.60.6/1.882.5/88.225.8/38.7
SegVLAD-PreT78.5/93.856.8/77.74.2/9.495.3/98.069.8/83.7
SegVLAD-FineT68.1/89.058.9/79.37.4/15.695.1/97.535.4/55.3
", + "bbox": [ + 318, + 323, + 687, + 429 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "indoor-to-outdoor viewing (InsideOut), and historical image matching (AmsterTime). Below, we discuss two key aspects of this comparative analysis: $i$ ) how our segment-based approach compares against whole-image global descriptor based methods and ii) how performance trends vary depending on the choices of feature backbone with regards to task-specific (VPR) training.", + "bbox": [ + 217, + 458, + 787, + 535 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Aggregating Segments vs Whole Images Table 1 and Table 2 show that our proposed method SegVLAD achieves a new state-of-the-art on the majority of datasets, considering both the backbone variants: PreT and FineT. AnyLoc and SALAD respectively differ from SegVLAD-PreT and SegVLAD-FineT in terms of the aggregation scope (global vs segments). Thus, the superior performance of SegVLAD clearly highlights the role of segments based retrieval over whole-image based approach. On the Baidu Mall dataset – highly-aliased indoor environment – our method improves over AnyLoc by $3\\%$ for R@1 and $6\\%$ for R@5 in absolute gains. On the InsideOut dataset – matching outdoor images viewed from within indoors – our method leads to a ‘meaningful’ recall, unlike all other baselines. Overall, these results highlight that even with the use of powerful image encoders (DINOv2), global aggregation struggles to deal with the challenges of matching images across major viewpoint shifts – it is thus the partial image representation and matching (using the same encoder) which is needed to obtain superior recognition performance.", + "bbox": [ + 217, + 558, + 787, + 770 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "VPR Fine-tuned Encoders + Segments For SegVLAD-FineT, we used a DINOv2 backbone finetuned for the purpose of VPR, mainly to observe the benefit of segments over global descriptor based approach in a task-specific manner. Table 1 and Table 2", + "bbox": [ + 217, + 794, + 787, + 840 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "Revisit Anything", + "bbox": [ + 627, + 114, + 730, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 114, + 785, + 126 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/0d8ce0d71710572182472082fe89623d17735ef79ec4364dd6ec6fb8dc5ae634.jpg", + "table_caption": [ + "Table 3: Recall@1 results for various approaches on Object-Instance Retrieval Task" + ], + "table_footnote": [], + "table_body": "
MethodSegVLAD NoNbrAggSegVLADSegment-to-GlobalGlobal-to-Global
R@164.192.730.086.4
", + "bbox": [ + 223, + 170, + 777, + 203 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "show that, on the outdoor street-view datasets, SALAD (finetuned DINOv2) generally performs better than AnyLoc (its pretrained counterpart), whereas the latter generally outperforms the former on 'out-of-distribution' datasets. It can be clearly observed that these performance patterns translate well from global- to segment-level results.", + "bbox": [ + 212, + 238, + 784, + 297 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "5.2 Revisiting Objects of Interest (OOI): Object Instance Retrieval", + "text_level": 1, + "bbox": [ + 214, + 325, + 694, + 340 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "A typical requirement of an embodied agent is to understand the context of its task through its memory/map information, which is composed of visual and/or semantic cues. For example, navigating to a given object goal requires a robot to visually recognize the goal and not be confused by perceptually-similar items. In this section, we demonstrate our method's ability to retrieve the correct image given just an Object Of Interest (OOI) as a query segment. For this purpose, we use an extended version of the Baidu dataset [70] which annotates OOI as various discriminative areas that can be reliably detected under variable viewpoint and lighting conditions. In total, there are 220 OOI, which cover various things such as logos, brand names, posters, etc., in a highly cluttered mall environment. To cast this dataset in terms of revisiting things, we use the original query images of the Baidu Mall [64] dataset as the database and the images with OOI as the queries. This allows us to evaluate the OOIs directly. This is similar to VPR evaluation of recall in terms of image retrieval but with querying of a specific segment instead of using all the segments of the query image.", + "bbox": [ + 212, + 354, + 787, + 566 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "We consider four different methods of recognizing known objects in this study. i) Global-to-Global: as a baseline method, we use whole images to represent and retrieve, i.e., without using the OOI mask; this resembles object-goal recognition problem for an InstanceImageNavigation task [39]. ii) Segment-to-Global: this is the same as the previous setting except that the query image descriptor is aggregated only using the OOI mask; this tests the ability of the image encoder/aggregator to match segment-level descriptor against global descriptors. iii) SegVLAD and iv) SegVLAD NoNbrAgg, which are our proposed methods but the latter does not use any neighborhood information; this highlights the relevance of spatial context around the OOI for recognition. For SegVLAD, we create a virtual segment mask for the OOI, append it to the other masks of the image, and then perform our neighborhood expansion and feature aggregation, as described in Section 3.", + "bbox": [ + 212, + 566, + 787, + 747 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Table 3 reports Recall@1 for different recognition methods. It can be observed that SegVLAD outperforms Global-to-Global matching by a large margin, which shows that recognizing specific object instances through their images (as in InstanceImageNav) is more prone to failures. It can further be observed from low recall of SegVLAD NoNbrAgg that neighborhood aggregation around segments is crucial to capture the required context. Finally, poor recall of Segment-to-Global highlights", + "bbox": [ + 212, + 750, + 787, + 840 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "K. Garg et al.", + "bbox": [ + 271, + 114, + 354, + 128 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/c951be63c9fefc81f88b898c8b2e7ca8c0398aaaebe1bed8b47e15e953d224e6.jpg", + "table_caption": [ + "Table 4: Recall@1/5 for Baidu mall dataset Table 5: Recall@1/5 comparison between for different aggregation methods and dif- different methods for ranking images based. ferent orders of neighborhood expansion. on segment-level retrieval." + ], + "table_footnote": [], + "table_body": "
OrderSegVLADSAP
073.1/89.974.6/91.1
177.4/91.765.6/87.2
276.3/92.453.2/81.3
377.7/92.649.8/78.0
Max Seg78.5/93.953.9/70.4
Max Sim65.2/92.734.4/62.4
Ours78.5/93.854.4/76.3
", + "bbox": [ + 245, + 212, + 754, + 292 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "that matching a part of an image (OOI) with the whole image is not a viable solution for object instance recognition.", + "bbox": [ + 212, + 323, + 782, + 354 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "5.3 Ablation Studies", + "text_level": 1, + "bbox": [ + 215, + 378, + 372, + 393 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Aggregation method & Order of Neighborhood Expansion Previous studies on VPR such as AnyLoc [32] have shown VLAD to be better than other aggregation methods for whole-image based global descriptors. However, in an increasing number of segment-based approaches [12, 20, 24, 61], segment average pooling (SAP) is used more commonly. Thus, we compare hard-assignment VLAD against SAP on Baidu dataset. For SAP, we upsample the DINOv2 features to match the resolution of our SAM masks – this upsampling is shown to enhance performance in [61]. For VLAD aggregation, we use our proposed method SegVLAD, where we downsample masks to match with the low resolution of DINOv2. Table 4 shows that SAP performs well for order 0 aggregation (i.e., no neighborhood aggregation) but its performance reduces as the neighborhood expands. On the other hand, SegVLAD has low recall when no neighborhood is considered but benefits significantly even with its immediate neighborhood (order 1). We attribute these inverted trends of SegVLAD and SAP to the very nature of these aggregation methods: as more information becomes available SAP smoothes out the overall information content whereas SegVLAD benefits from additional information which gets distributed across its clusters, thus minimizing any possible smoothing effect. It can be observed that R@5 increases for SegVLAD with an increasing order of neighborhood expansion but margins diminish for higher orders. Overall, SegVLAD (order 3) achieves the best results, despite aggregating at a $14 \\times$ lower resolution than SAP's upsampling based aggregation.", + "bbox": [ + 212, + 406, + 787, + 709 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Segment to Image Retrieval Unlike conventional global descriptor retrieval based VPR, we perform retrieval for multiple SuperSegments of the query image. To obtain retrieval output in terms of images (as that is what VPR is typically evaluated on), there exist multiple ways to combine the top segment-level matches across all the query segments. We consider the following alternative options. i) MaxSeg: we obtain the best matching segment for each query segment, associate the matched segments to their respective reference image indices, and then rank these image indices based on their", + "bbox": [ + 212, + 734, + 787, + 839 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "Revisit Anything", + "bbox": [ + 627, + 114, + 730, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 784, + 126 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/a580aa5b9234506efb32d3196b7efd79d541a96dc84361086e7e6e6579d67123.jpg", + "image_caption": [ + "Fig. 4: Qualitative results: Columns respectively represent the query image, correct match of SegVLAD and incorrect match of AnyLoc. Examples from different datasets: AmsterTime, Baidu Mall, Pitts-30K are presented across the rows." + ], + "image_footnote": [], + "bbox": [ + 233, + 142, + 769, + 450 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "frequency; this method weakly resembles an inverted index list based counting of common segments between the query and the reference. ii) MaxSim: across the best matching reference segments, we order their image indices based on the segment similarity; this method is similar to that used in MultiVLAD [4]. iii) Similarity-Weighted Frequency: this is our proposed method as defined in Section 3.4. Table 5 shows that our proposed method for combining segment-level hypotheses consistently achieves superior results for both the datasets. While MaxSeg achieves a similar performance on Baidu, it suffers a drop in recall for AmsterTime. Both the methods outperform MaxSim at R@1 by a large margin.", + "bbox": [ + 212, + 547, + 787, + 684 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Patches vs Segment While open-set segmentation [37] is aimed at a meaningful segregation of visual entities, a simple alternative to our segment-based retrieval is to use uniformly defined regions/patches. Table 6 compares SegVLAD with a patch-based approach, where we consider arbitrary square patch sizes to segment an image. It can be observed that SegVLAD outperforms its patch-based counterparts, where smaller patches perform the worst, and for larger patches, R@1 saturates while R@5 reduces. These results are also in line with similar findings of a recent work [61].", + "bbox": [ + 212, + 734, + 787, + 839 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "K. Garg et al.", + "bbox": [ + 271, + 114, + 354, + 128 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/f4b8ec815e3f5b6f1e792ebaf19ae217e83292ed60fe90566e92a0bf58057df9.jpg", + "image_caption": [ + "Fig.5: Recall vs storage/retrieval time on AmsterTime." + ], + "image_footnote": [], + "bbox": [ + 236, + 147, + 470, + 255 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/5fd7cde689077b3cf64885e0e38a8912180d6e9a4d1e540a7397b60f9433e982.jpg", + "table_caption": [ + "Table 6: Patch vs Segments on AmsterTime (Reso. $256\\times 256$" + ], + "table_footnote": [], + "table_body": "
Patch SizeNumSegDbNumSegImR@1/5
16×1631513625635.7/62.1
32×32787846445.9/72.5
64×64196961652.0/75.1
128×1284924453.5/65.6
SegVLAD12963710556.8/77.7
", + "bbox": [ + 521, + 196, + 769, + 289 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "5.4 Limitations, Storage, Compute Time & IOU Based Filtering", + "text_level": 1, + "bbox": [ + 215, + 328, + 674, + 345 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "A key limitation of our method is its large map size, i.e., large storage requirement for segment-level descriptors (see supplementary for further details). In this section, we analyze the resource requirements in terms of database (index) storage and query retrieval time for our method, along with a preliminary study on IOU (Intersection over Union) based filtering of SuperSegments to reduce such costs. We compute IOU between all pairs of SuperSegments in a given image, and remove segments with $IOU(s_i, s_j) > \\psi \\forall i \\in [1, S]$ , $j \\in [i, S]$ , where $\\psi \\in [0, 1]$ is a threshold on IOU and $s_i$ refers to the list of SuperSegments sorted by their pixel area in a descending order. We only perform this culling on the database segments. We use the outdoor-finetuned DINOv2 backbone for this purpose and compare SegVLAD with SALAD on AmsterTime (see supplementary for additional results on Pitts30K). Figure 5 shows that SegVLAD outperforms SALAD while requiring less storage (annotated on points) and comparable retrieval time (excludes extraction time), using IoU-based filtering threshold ranging from 0.2 to 0.8 (left to right) with a step size of 0.2. In particular, at 0.4 IOU threshold, only 20% of SuperSegments are retained (0.05 GB) while still outperforming the baseline.", + "bbox": [ + 212, + 354, + 787, + 580 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "5.5 Qualitative Analyses", + "text_level": 1, + "bbox": [ + 215, + 603, + 401, + 618 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "In this section, we further demonstrate the capabilities of our method through qualitative visualizations. We compare our method against AnyLoc, where the only difference between the two methods is Global aggregation/retrieval vs SuperSegment aggregation/retrieval. We particularly consider the queries for which our approach successfully retrieved the correct match but AnyLoc failed to do so (additional examples can be found in the supplementary). Figure 4 shows triplets of images in the order of query, correct match (ours), and incorrect match (Anyloc). The segmented part shows one of the correctly matched SuperSegments, displayed as a subgraph in white color overlaid on the corresponding segment masks.", + "bbox": [ + 212, + 628, + 787, + 763 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "The first row shows a triplet from AmsterTime where our proposed method is able to correctly recognize a subgraph of building across the image pair, whereas Anyloc retrieves an incorrect image of a street-view with buildings, cars, and road, laid out similarly across the image pair. This highlights that a global descriptor can get confused with perceptually-aliased global context. In contrast, our SuperSegment based", + "bbox": [ + 212, + 765, + 787, + 840 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "Revisit Anything", + "bbox": [ + 627, + 114, + 730, + 128 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 769, + 114, + 784, + 126 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "SegVLAD is not only able to retrieve the correct image but it also correctly finds the mutually-overlapping area. A similar trend follows for the Baidu Mall (middle row) and Pitts30K (last row). In the Baidu Mall example, our approach identifies the piano and the region around it in the query image. It correctly retrieves an image having similar spatial context with the piano. This is akin to how humans use spatial context to identify places. AnyLoc, on the other hand, retrieves an image with similar floor tiles and railings. This example particularly highlights our hypothesis that dissimilarity of nonoverlapping regions can dominate the similarity of overlapping regions in global whole-image descriptors. Finally, the Pitts30K example (last row) shows a case of strong viewpoint change. While SegVLAD correctly matches the traffic signal and signboards to retrieve the correct match, AnyLoc retrieves a similar looking image while missing the finer details. This example particularly reinforces the idea of 'revisiting things', as even though the background mountain is common across the triplet, it is the context of the things near the camera/robot which helps in uniquely recognizing a place.", + "bbox": [ + 215, + 146, + 787, + 358 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "6 Conclusion and Future Work", + "text_level": 1, + "bbox": [ + 215, + 381, + 493, + 398 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In this paper, we presented a novel visual place recognition pipeline SegVLAD based on image segments-based description and retrieval, which is akin to 'revisting things' as a means to recognize specific instances of what constitute a place. Our proposed SuperSegments based image representation and a novel factorization based feature aggregation enables us to effectively represent and retrieve images using our segment similarity-weighted image ranking. Our results show that despite using powerful image encoders such as DINOv2 (pretrained or VPR-finetuned), existing global descriptor based techniques are unable to deal with the challenges of viewpoint variations. In contrast, SegVLAD is able to correctly retrieve images through its ability to match partially-overlapping images with its partial image representation in the form of semiglobal subgraphs of segments, i.e., SuperSegments. Thus, our method achieves state-of-the-art results on three diverse datasets (indoor and outdoor) that exhibit strong viewpoint variations on top of other challenges of appearance shift and high perceptual aliasing. Through an additional object instance retrieval study, we demonstrate the unique ability of our method to recognize objects instances within their specific place contexts - an open-set recognition capability that existing VPR methods lack.", + "bbox": [ + 215, + 412, + 787, + 654 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Our approach shifts the paradigm in retrieval based VPR research, as the conventional methods predominantly classify into either whole-image global descriptor based coarse retrieval or local feature based geometric reranking. Our approach complements recent concurrent works like MESA [75]; future work can explore a hierarchical VPR pipeline that closely integrates a segment-based coarse retriever with segments-based reranker such as MESA, thus doing away entirely with global whole-image descriptors. Furthermore, segments-based representation with implicitly baked semantics provide a natural way for creating text-based interfaces through CLIP [55] and LLMs (Large Language Models) [9], which can be easily integrated with recent efforts in this direction [12, 18, 20, 24, 42, 43, 77].", + "bbox": [ + 215, + 655, + 787, + 806 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "K. Garg et al.", + "bbox": [ + 271, + 114, + 354, + 128 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 215, + 143, + 382, + 162 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "This work was supported by the Centre for Augmented Reasoning (CAR) at the Australian Institute for Machine Learning (AIML), University of Adelaide, Australia. The authors acknowledge the computational support provided by the Indian Institute of Science (IISc), Bengaluru, India, and the International Institute of Information Technology, Hyderabad (IIITH), India. The authors thank Ahmad Khaliq for technical support, Sarah Ibrahimi for sharing the InsideOut dataset, and Martin Humenberger for sharing the OOI annotations for the Baidu Mall dataset.", + "bbox": [ + 212, + 176, + 787, + 282 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 215, + 306, + 310, + 321 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Ali-bey, A., Chaib-draa, B., Giguère, P.: Gsv-cities: Toward appropriate supervised visual place recognition. Neurocomputing 513, 194-203 (2022)", + "2. Ali-bey, A., Chaib-draa, B., Giguère, P.: Mixvpr: Feature mixing for visual place recognition. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 2998-3007 (2023)", + "3. Arandjelovic, R., Gronat, P., Torii, A., Pajdla, T., Sivic, J.: Netvlad: Cnn architecture for weakly supervised place recognition. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 5297-5307 (2016)", + "4. Arandjelovic, R., Zisserman, A.: All about vlad. In: Proceedings of the IEEE conference on Computer Vision and Pattern Recognition. pp. 1578-1585 (2013)", + "5. Bar, M.: Visual objects in context. Nature Reviews Neuroscience 5(8), 617-629 (2004)", + "6. Berton, G., Masone, C., Caputo, B.: Rethinking visual geo-localization for large-scale applications. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4878-4888 (2022)", + "7. Berton, G., Mereu, R., Trivigno, G., Masone, C., Csurka, G., Sattler, T., Caputo, B.: Deep visual geo-localization benchmark. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5396-5407 (2022)", + "8. Berton, G., Trivigno, G., Caputo, B., Masone, C.: Eigenplaces: Training viewpoint robust models for visual place recognition. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 11080-11090 (October 2023)", + "9. Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J.D., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., et al.: Language models are few-shot learners. Advances in neural information processing systems 33, 1877-1901 (2020)", + "0. Caesar, H., Uijlings, J., Ferrari, V.: Coco-stuff: Thing and stuff classes in context. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1209-1218 (2018)", + "1. Cao, B., Araujo, A., Sim, J.: Unifying deep local and global features for image search. In: European Conference on Computer Vision. pp. 726-743. Springer (2020)", + "2. Chang, M., Gervet, T., Khanna, M., Yenamandra, S., Shah, D., Min, S.Y., Shah, K., Paxton, C., Gupta, S., Batra, D., et al.: Goat: Go to any thing. arXiv preprint arXiv:2311.06430 (2023)", + "3. Chen, L.C., Yang, Y., Wang, J., Xu, W., Yuille, A.L.: Attention to scale: Scale-aware semantic image segmentation. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 3640-3649 (2016)", + "4. Chen, Z., Maffra, F., Sa, I., Chli, M.: Only look once, mining distinctive landmarks from convnet for visual place recognition. In: Intelligent Robots and Systems (IROS), 2017 IEEE/RSJ International Conference on. pp. 9-16. IEEE (2017)" + ], + "bbox": [ + 225, + 339, + 784, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "Revisit Anything", + "bbox": [ + 627, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 769, + 114, + 784, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "15. Cheng, C., Page, D.L., Abidi, M.A.: Object-based place recognition and loop closing with jigsaw puzzle image segmentation algorithm. In: 2008 IEEE International Conference on Robotics and Automation. pp. 557-562. IEEE (2008)", + "16. Cummins, M., Newman, P.: Appearance-only slam at large scale with fab-map 2.0. The International Journal of Robotics Research 30(9), 1100-1123 (2011)", + "17. Cupec, R., Nyarko, E.K., Filko, D., Kitanov, A., Petrovic, I.: Place recognition based on matching of planar surfaces and line segments. The International Journal of Robotics Research 34(4-5), 674-704 (2015)", + "18. Folorunsho, S.O.: Semantic segmentation-based approach for autonomous navigation in challenging farm terrains. algorithms 15, 3 (2024)", + "19. Garg, S., Fischer, T., Milford, M.: Where is your place, visual place recognition? In: IJCAI (2021)", + "20. Garg, S., Rana, K., Hosseinzadeh, M., Mares, L., Suenderhauf, N., Dayoub, F., Reid, I.: Robohop: Segment-based topological map representation for open-world visual navigation. In: 2024 IEEE International Conference on Robotics and Automation (ICRA) (2024)", + "21. Garg, S., Suenderhauf, N., Milford, M.: Lost? appearance-invariant place recognition for opposite viewpoints using visual semantics. In: Proceedings of Robotics: Science and Systems XIV (2018)", + "22. Garg, S., Sunderhauf, N., Dayoub, F., Morrison, D., Cosgun, A., Carneiro, G., Wu, Q., Chin, T.J., Reid, I., Gould, S., Corke, P., Milford, M.: Semantics for robotic mapping, perception and interaction: A survey. Foundations and Trends® in Robotics 8(1-2), 1-224 (2020). https://doi.org/10.1561/2300000059, http://dx.doi.org/10.1561/2300000059", + "23. Gawel, A., Del Don, C., Siegwart, R., Nieto, J., Cadena, C.: X-view: Graph-based semantic multi-view localization. IEEE Robotics and Automation Letters 3(3), 1687-1694 (2018)", + "24. Gu, Q., Kuwajerwala, A., Morin, S., Jatavallabhula, K.M., Sen, B., Agarwal, A., Rivera, C., Paul, W., Ellis, K., Chellappa, R., et al.: Conceptgraphs: Open-vocabulary 3d scene graphs for perception and planning. arXiv preprint arXiv:2309.16650 (2023)", + "25. Hausler, S., Garg, S., Xu, M., Milford, M., Fischer, T.: Patch-netvlad: Multi-scale fusion of locally-global descriptors for place recognition. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14141-14152 (2021)", + "26. Hu, H., Qiao, Z., Cheng, M., Liu, Z., Wang, H.: Dasgil: Domain adaptation for semantic and geometric-aware image-based localization. IEEE Transactions on Image Processing 30, 1342-1353 (2020)", + "27. Ibrahimi, S., Van Noord, N., Alpherts, T., Worring, M.: Inside out visual place recognition. arXiv preprint arXiv:2111.13546 (2021)", + "28. Intraub, H.: The representation of visual scenes. Trends in cognitive sciences 1(6), 217-222 (1997)", + "29. Izquierdo, S., Civera, J.: Optimal transport aggregation for visual place recognition (2023)", + "30. Jégou, H., Douze, M., Schmid, C., Pérez, P.: Aggregating local descriptors into a compact image representation. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 3304-3311 (2010)", + "31. Kassab, C., Mattamala, M., Fallon, M.: Clip-based features achieve competitive zero-shot visual localization. OpenReview preprint arXiv:2306.14846 (2023)", + "32. Keetha, N., Mishra, A., Karhade, J., Jatavallabhula, K.M., Scherer, S., Krishna, M., Garg, S.: Anyloc: Towards universal visual place recognition. IEEE Robotics and Automation Letters (2023)", + "33. Keetha, N.V., Milford, M., Garg, S.: A hierarchical dual model of environment-and place-specific utility for visual place recognition. IEEE Robotics and Automation Letters 6(4), 6969-6976 (2021)" + ], + "bbox": [ + 215, + 147, + 785, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "K. Garg et al.", + "bbox": [ + 271, + 114, + 354, + 128 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "34. Khaliq, A., Ehsan, S., Milford, M., McDonald-Maier, K.: A holistic visual place recognition approach using lightweight cnns for severe viewpoint and appearance changes. arXiv preprint arXiv:1811.03032 (2018)", + "35. Khaliq, A., Xu, M., Hausler, S., Milford, M., Garg, S.: Vlad-buff: Burst-aware fast feature aggregation for visual place recognition. In: European Conference on Computer Vision. Springer (2024)", + "36. Khorasgani, S.H., Chen, Y., Shkurti, F.: Slic: Self-supervised learning with iterative clustering for human action videos. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16091-16101 (2022)", + "37. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., et al.: Segment anything. arXiv preprint arXiv:2304.02643 (2023)", + "38. Knopp, J., Sivic, J., Pajdla, T.: Avoiding confusing features in place recognition. In: European Conference on Computer Vision. pp. 748-761. Springer (2010)", + "39. Krantz, J., Lee, S., Malik, J., Batra, D., Chaplot, D.S.: Instance-specific image goal navigation: Training embodied agents to find object instances. arXiv preprint arXiv:2211.15876 (2022)", + "40. Le, D.C., Youn, C.H.: City-scale visual place recognition with deep local features based on multi-scale ordered VLAD pooling. arXiv preprint arXiv:2009.09255 (2020)", + "41. Lowe, D.G.: Distinctive image features from scale-invariant keypoints. International journal of computer vision 60(2), 91-110 (2004)", + "42. Maalouf, A., Jadhav, N., Jatavallabhula, K.M., Chahine, M., Vogt, D.M., Wood, R.J., Torralba, A., Rus, D.: Follow anything: Open-set detection, tracking, and following in real-time. IEEE Robotics and Automation Letters 9(4), 3283-3290 (2024)", + "43. Manglani, S.: Real-time vision-based navigation for a robot in an indoor environment. arXiv preprint arXiv:2307.00666 (2023)", + "44. Masone, C., Caputo, B.: A survey on deep visual place recognition. IEEE Access 9, 19516-19547 (2021)", + "45. Mirjalili, R., Krawez, M., Burgard, W.: Fm-loc: Using foundation models for improved vision-based localization. In: Robotics and Automation (ICRA), 2023 IEEE International Conference on. IEEE (2023)", + "46. Mousavian, A., Kosecka, J.: Semantic image based geolocation given a map (author's initial manuscript). Tech. rep., George Mason University Fairfax United States (2016)", + "47. Naseer, T., Oliveira, G.L., Brox, T., Burgard, W.: Semantics-aware visual localization under challenging perceptual conditions. In: IEEE International Conference on Robotics and Automation (ICRA) (2017)", + "48. Oliva, A.: Gist of the scene. Neurobiology of attention 696(64), 251-258 (2005)", + "49. Oquab, M., Darcet, T., Moutakanni, T., Vo, H.V., Szafraniec, M., Khalidov, V., Fernandez, P., Haziza, D., Massa, F., El-Nouby, A., Howes, R., Huang, P.Y., Xu, H., Sharma, V., Li, S.W., Galuba, W., Rabbat, M., Assran, M., Ballas, N., Synnaeve, G., Misra, I., Jegou, H., Mairal, J., Labatut, P., Joulin, A., Bojanowski, P.: Dinov2: Learning robust visual features without supervision (2023)", + "50. Paolicelli, V., Tavera, A., Masone, C., Berton, G., Caputo, B.: Learning semantics for visual place recognition through multi-scale attention. In: Image Analysis and Processing-ICIAP 2022: 21st International Conference, Lecce, Italy, May 23-27, 2022, Proceedings, Part II. pp. 454-466. Springer (2022)", + "51. Pion, N., Humenberger, M., Csurka, G., Cabon, Y., Sattler, T.: Benchmarking image retrieval for visual localization. In: 2020 International Conference on 3D Vision (3DV). pp. 483-494. IEEE (2020)", + "52. Puligilla, S.S., Tourani, S., Vaidya, T., Parihar, U.S., Sarvadevabhatla, R.K., Krishna, K.M.: Topological mapping for manhattan-like repetitive environments. In: 2020 IEEE International Conference on Robotics and Automation (ICRA). pp. 6268-6274. IEEE (2020)" + ], + "bbox": [ + 212, + 146, + 787, + 839 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "Revisit Anything", + "bbox": [ + 627, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "53. Radenović, F., Iscen, A., Tobias, G., Avrithis, Y., Chum, O.: Revisiting oxford and paris: Large-scale image retrieval benchmarking. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 5706-5715 (2018)", + "54. Radenovic, F., Tobias, G., Chum, O.: Fine-tuning cnn image retrieval with no human annotation. IEEE Trans. Pattern Anal. Mach. Intell. 41(7), 1655-1668 (2018)", + "55. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021)", + "56. Revaud, J., Almazán, J., Rezende, R.S., Souza, C.R.d.: Learning with average precision: Training image retrieval with a listwise loss. In: Proceedings of the IEEE International Conference on Computer Vision. pp. 5107-5116 (2019)", + "57. Sarlin, P.E., Cadena, C., Siegwart, R., Dymczyk, M.: From coarse to fine: Robust hierarchical localization at large scale. arXiv preprint arXiv:1812.03506 (2018)", + "58. Sattler, T., Maddern, W., Toft, C., Torii, A., Hammarstrand, L., Stenberg, E., Safari, D., Okutomi, M., M. P., J. S., F. K., T. P.: Benchmarking 6dof outdoor visual localization in changing conditions. In: Proc. CVPR (2018)", + "59. Schleiss, M., Rouatbi, F., Cremers, D.: Vpair - aerial visual place recognition and localization in large-scale outdoor environments (2022)", + "60. Schubert, S., Neubert, P., Garg, S., Milford, M., Fischer, T.: Visual place recognition: A tutorial. RAM (2023)", + "61. Shlapentokh-Rothman, M., Blume, A., Xiao, Y., Wu, Y., TV, S., Tao, H., Lee, J.Y., Torres, W., Wang, Y.X., Hoiem, D.: Region-based representations revisited. arXiv preprint arXiv:2402.02352 (2024)", + "62. Shubodh, S., Omama, M., Zaidi, H., Parihar, U.S., Krishna, M.: Lip-loc: Lidar image pretraining for cross-modal localization. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 948-957 (2024)", + "63. Sivic, J., Zisserman, A.: Video google: A text retrieval approach to object matching in videos. In: Proceedings of International Conference on Computer Vision (ICCV). p. 1470. IEEE (2003)", + "64. Sun, X., Xie, Y., Luo, P., Wang, L.: A dataset for benchmarking image-based localization. 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR) pp. 5641-5649 (2017), https://apisemantic scholar.org/CorpusID:20531893", + "65. Sünderhauf, N., Shirazi, S., Jacobson, A., Dayoub, F., Pepperell, E., Upcroft, B., Milford, M.: Place recognition with convnet landmarks: Viewpoint-robust, condition-robust, training-free. Proceedings of Robotics: Science and Systems XII (2015)", + "66. Torii, A., Sivic, J., Pajdla, T., Okutomi, M.: Visual place recognition with repetitive structures. In: 2013 IEEE Conference on Computer Vision and Pattern Recognition. pp. 883-890 (2013). https://doi.org/10.1109/CVPR.2013.119", + "67. Tsintotas, K.A., Bampis, L., Gasteratos, A.: The revisiting problem in simultaneous localization and mapping: A survey on visual loop closure detection. IEEE Transactions on Intelligent Transportation Systems 23(11), 19929-19953 (2022)", + "68. Wang, R., Shen, Y., Zuo, W., Zhou, S., Zheng, N.: Transvpr: Transformer-based place recognition with multi-level attention aggregation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13648-13657 (2022)", + "69. Warburg, F., Hauberg, S., López-Antequera, M., Gargallo, P., Kuang, Y., Civera, J.: Mapillary street-level sequences: A dataset for lifelong place recognition. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2626-2635 (2020)", + "70. Weinzaepfel, P., Csurka, G., Cabon, Y., Humenberger, M.: Visual localization by learning objects-of-interest dense match regression. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5634-5643 (2019)" + ], + "bbox": [ + 215, + 147, + 784, + 839 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "K. Garg et al.", + "bbox": [ + 271, + 114, + 354, + 128 + ], + "page_idx": 17 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "71. Xin, Z., Cai, Y., Lu, T., Xing, X., Cai, S., Zhang, J., Yang, Y., Wang, Y.: Localizing discriminative visual landmarks for place recognition. In: 2019 International conference on robotics and automation (ICRA). pp. 5979-5985. IEEE (2019)", + "72. Yildiz, B., Khademi, S., Siebes, R.M., Van Gemert, J.: Amstertime: A visual place recognition benchmark dataset for severe domain shift. In: 2022 26th International Conference on Pattern Recognition (ICPR). IEEE (Aug 2022). https://doi.org/10.1109/icpr56361.2022.9956049, http://dx.doi.org/10.1109/ICPR56361.2022.9956049", + "73. Yu, J., Zhu, C., Zhang, J., Huang, Q., Tao, D.: Spatial pyramid-enhanced netvlad with weighted triplet loss for place recognition. IEEE Trans. Neural Netw. Learn. Syst. 31(2), 661-674 (2019)", + "74. Zaffar, M., Garg, S., Milford, M., Kooij, J., Flynn, D., McDonald-Maier, K., Ehsan, S.: Vpr-bench: An open-source visual place recognition evaluation framework with quantifiable viewpoint and appearance change. International Journal of Computer Vision pp. 1-39 (2021)", + "75. Zhang, Y., Zhao, X.: Mesa: Matching everything by segmenting anything. arXiv preprint arXiv:2401.16741 (2024)", + "76. Zhou, B., Lapedriza, A., Khosla, A., Oliva, A., Torralba, A.: Places: A 10 million image database for scene recognition. IEEE Transactions on Pattern Analysis and Machine Intelligence (2017)", + "77. Zou, X., Yang, J., Zhang, H., Li, F., Li, L., Wang, J., Wang, L., Gao, J., Lee, Y.J.: Segment everything everywhere all at once. Advances in Neural Information Processing Systems 36 (2024)" + ], + "bbox": [ + 212, + 146, + 787, + 465 + ], + "page_idx": 18 + }, + { + "type": "header", + "text": "Revisit Anything", + "bbox": [ + 627, + 114, + 730, + 128 + ], + "page_idx": 18 + }, + { + "type": "page_number", + "text": "19", + "bbox": [ + 769, + 114, + 784, + 126 + ], + "page_idx": 18 + } +] \ No newline at end of file diff --git a/2024/Revisit Anything_ Visual Place Recognition via Image Segment Retrieval/32b8503b-ae81-47c5-bdc0-0efb4201113d_model.json b/2024/Revisit Anything_ Visual Place Recognition via Image Segment Retrieval/32b8503b-ae81-47c5-bdc0-0efb4201113d_model.json new file mode 100644 index 0000000000000000000000000000000000000000..5907cdaeeefc16ca9a044b9323f82386b57f33b0 --- /dev/null +++ b/2024/Revisit Anything_ Visual Place Recognition via Image Segment Retrieval/32b8503b-ae81-47c5-bdc0-0efb4201113d_model.json @@ -0,0 +1,2449 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.233, + 0.141, + 0.773, + 0.187 + ], + "angle": 0, + "content": "Revisit Anything: Visual Place Recognition via Image Segment Retrieval" + }, + { + "type": "text", + "bbox": [ + 0.284, + 0.213, + 0.72, + 0.245 + ], + "angle": 0, + "content": "Kartik Garg\\(^{1}\\), Sai Shubodh Pulgilla\\(^{2}\\), Shishir Kolathaya\\(^{1}\\), Madhava Krishna\\(^{2}\\), and Sourav Garg\\(^{3}\\)" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.255, + 0.761, + 0.298 + ], + "angle": 0, + "content": "\\(^{1}\\) Indian Institute of Science (IISc), Bengaluru, India \n\\(^{2}\\) International Institute of Information Technology, Hyderabad, India \n\\(^{3}\\) University of Adelaide, Australia" + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.299, + 0.786, + 0.326 + ], + "angle": 0, + "content": "kartikgarg@iisc.ac.in, p.saishubodh@gmail.com, shishirk@iisc.ac.in, mkrishna@iiit.ac.in, sourav.garg@adelaide.edu.au" + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.365, + 0.744, + 0.697 + ], + "angle": 0, + "content": "Abstract. Accurately recognizing a revisited place is crucial for embodied agents to localize and navigate. This requires visual representations to be distinct, despite strong variations in camera viewpoint and scene appearance. Existing visual place recognition pipelines encode the whole image and search for matches. This poses a fundamental challenge in matching two images of the same place captured from different camera viewpoints: the similarity of what overlaps can be dominated by the dissimilarity of what does not overlap. We address this by encoding and searching for image segments instead of the whole images. We propose to use open-set image segmentation to decompose an image into 'meaningful' entities (i.e., things and stuff). This enables us to create a novel image representation as a collection of multiple overlapping subgraphs connecting a segment with its neighboring segments, dubbed SuperSegment. Furthermore, to efficiently encode these SuperSegments into compact vector representations, we propose a novel factorized representation of feature aggregation. We show that retrieving these partial representations leads to significantly higher recognition recall than the typical whole image based retrieval. Our segments-based approach, dubbed SegVLAD, sets a new state-of-the-art in place recognition on a diverse selection of benchmark datasets, while being applicable to both generic and task-specialized image encoders. Finally, we demonstrate the potential of our method to \"revisit anything\" by evaluating our method on an object instance retrieval task, which bridges the two disparate areas of research: visual place recognition and object-goal navigation, through their common aim of recognizing goal objects specific to a place. Source code: https://github.com/AnyLoc/Revisit-Anything." + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.711, + 0.687, + 0.726 + ], + "angle": 0, + "content": "Keywords: Visual Place Recognition \\(\\cdot\\) Image Segmentation \\(\\cdot\\) Robotics" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.753, + 0.359, + 0.768 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.785, + 0.789, + 0.817 + ], + "angle": 0, + "content": "Visual Place Recognition (VPR) is an important capability for embodied agents to localize and navigate autonomously. A predominant solution for VPR is to encode" + }, + { + "type": "page_footnote", + "bbox": [ + 0.245, + 0.826, + 0.372, + 0.841 + ], + "angle": 0, + "content": "* Equal contribution" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.116, + 0.355, + 0.129 + ], + "angle": 0, + "content": "K. Garg et al." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.148, + 0.788, + 0.495 + ], + "angle": 0, + "content": "an image into a global vector and retrieve similar vectors as coarse localization hypotheses [19, 44, 60, 67]. Thus, for almost a decade, researchers have focused on learning/finetuning image encoders so that global descriptors are induced with invariance to appearance [3,56,69], viewpoint [3,8], and clutter [27]. On the other hand, there is a vast literature on local descriptors (point/pixel-level), mainly relevant for geometric reranking in hierarchical VPR [11, 16, 25, 57, 68]. In the middle of local and global descriptors exists a variety of methods that use regions/patches [4, 25], lines/planes [17], objects (things/stuff) [21, 31, 45, 65], and segments [26, 33, 50] to represent images. However, these methods are still only aimed at either improving global descriptors based coarse retrieval or local feature matching based reranking. In this work, in contrast to conventional retrieval-based VPR, we explore an alternative: retrieval via encoding segments instead of the whole image. This is particularly enabled by recent advances in open-set image segmentation [37] which can meaningfully deconstruct a place into 'things' (and/or 'stuff') [10]. Thus, we reformulate the VPR problem of revisiting places as that to revisiting things by enabling recognition of these specific things within the context of their place. While such a segment-level place recognition approach provides a direct link to higher-level semantic tasks, such as object-goal navigation [12, 20, 24, 43], it also addresses a fundamental issue in matching partially-overlapping images from across significant viewpoint change. Segments-based partial image representation avoids the mismatches caused by the whole-image representation when the similarity of what overlaps is dominated by the dissimilarity of what does not overlap. Our novel segments-based VPR method, dubbed SegVLAD (Segment based Vector of Locally Aggregated Descriptors), is illustrated in Figure 1, which makes the following novel contributions:" + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.508, + 0.784, + 0.55 + ], + "angle": 0, + "content": "1. an image representation as a collection of multiple overlapping subgraphs of segments, dubbed SuperSegments, which enables accurate recognition across partially-overlapping images;" + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.553, + 0.784, + 0.58 + ], + "angle": 0, + "content": "2. a factorized representation of feature aggregation to effectively accommodate both segment-level information as well as segment neighborhood information; and" + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.584, + 0.784, + 0.61 + ], + "angle": 0, + "content": "3. a similarity-weighted ranking method to convert segment-level retrieval into image-level retrieval." + }, + { + "type": "list", + "bbox": [ + 0.223, + 0.508, + 0.784, + 0.61 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.628, + 0.788, + 0.747 + ], + "angle": 0, + "content": "Using a diverse set of data sources, we demonstrate that our proposed segments-based retrieval enables place recognition under wide viewpoint variations, where global descriptor based retrieval suffers. SegVLAD achieves a new state-of-the-art on multiple challenging datasets. We also introduce an evaluation of our method on an instance-level object retrieval task – a novel capability of our pipeline unlike conventional VPR methods. We conduct several ablations and parameter studies to justify the design choices and emphasize the effectiveness of our method as an open-set segments-based coarse retriever." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.775, + 0.377, + 0.792 + ], + "angle": 0, + "content": "2 Related Works" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.81, + 0.787, + 0.841 + ], + "angle": 0, + "content": "Image retrieval-based Visual Place Recognition (VPR) is a well-established area of research in visual localization [19,44,60,67]. It is important both during mapping for loop" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.628, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Revisit Anything" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.219, + 0.146, + 0.788, + 0.342 + ], + "angle": 0, + "content": "closures [67] as well as for relocalization [51, 58]. The underlying task in both the scenarios remains the same: how to recognize a previously seen place. The state-of-the-art methods in VPR use a global descriptor-based approach which converts an image into a compact vector to enable fast retrieval [2, 3, 6, 29, 32, 56]. The top retrieved hypotheses are often then re-ranked through compute-intensive local feature matching using geometric information [11, 16, 25, 33, 57, 68]. In contrast to previous approaches, we aim to explore image segment level descriptors in this work. This representation falls between point-based local descriptors and the whole-image based global descriptors. Our approach can be considered as 'semi-global', with the proposed segment (and SuperSegment) based descriptors being a 'spatially-reduced' form of whole-image global representation. This is motivated by our hypothesis that to deal with viewpoint variations in VPR with partially-overlapping images, we need a way to partially represent and match them." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.364, + 0.42, + 0.379 + ], + "angle": 0, + "content": "2.1 Whole Image Encoders" + }, + { + "type": "text", + "bbox": [ + 0.219, + 0.388, + 0.788, + 0.615 + ], + "angle": 0, + "content": "Earlier works in whole-image representation used methods like Gist [48], BoW (Bag of Word) [63], and VLAD (Vector of Locally Aggregated Descriptors) [30], often defined using hand-crafted features such as SIFT [41]. In recent years, deep learning based methods have demonstrated remarkable performance, with initial successful methods like NetVLAD [3] now rapidly outperformed by better alternatives such as CosPlace [6], MixVPR [2], EigenPlaces [8], TransVPR [68], and more recently AnyLoc [32], SALAD [29] and VLAD-BuFF [35]. All these learning-based methods improve different aspects of representation learning: training datasets [1, 6, 69], objective/loss functions [2, 8], aggregation methods [29, 54, 56, 68], and generalization [32]. Our approach complements these existing methods as we mainly focus on the use of segment-based information, where the segments can be described by any of the image encoders from the aforementioned techniques. In particular, we demonstrate that both - an off-the-shelf encoder, e.g., DINOv2-AnyLoc [32, 49] or that finetuned specifically for the VPR task, e.g., DINOv2-NetVLAD [35] - can be used in conjunction with our segment-based approach to further elevate place recognition capability." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.635, + 0.461, + 0.65 + ], + "angle": 0, + "content": "2.2 Region/Patch Based Methods" + }, + { + "type": "text", + "bbox": [ + 0.219, + 0.659, + 0.788, + 0.84 + ], + "angle": 0, + "content": "There exist several methods that employ region or patch level information to enhance representational power [13, 14, 34, 40, 52, 71, 73]. However, most of these methods only use this additional information to generate a single (or concatenated) compact vector representation of an image. Other methods such as Patch-NetVLAD [25] create multiple features per image but their primary purpose is to perform local matching based reranking. In contrast to these methods, we aim to use multiple segment descriptors per image to directly retrieve from a database of segments, without using any geometric information or reranking. The motivation behind this stems from the very nature of hierarchical VPR pipelines: reranking recall is upper bounded by the coarse retriever. A better coarse retriever can improve reranking performance without needing to rerank from a longer list of top hypotheses. MultiVLAD [4] is similar to our method in the spirit of retrieving multiple features per query image. However, like aforementioned" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.117, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.116, + 0.355, + 0.129 + ], + "angle": 0, + "content": "K. Garg et al." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.147, + 0.786, + 0.178 + ], + "angle": 0, + "content": "methods, MultiVLAD defines regions arbitrarily, whereas we use image segments obtained from Segment Anything Model (SAM) [37] which are semantically meaningful." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.197, + 0.465, + 0.213 + ], + "angle": 0, + "content": "2.3 Segments-Enhanced Methods" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.22, + 0.787, + 0.505 + ], + "angle": 0, + "content": "There exist several methods that use semantic segmentation information to improve VPR, as also surveyed in [22]. These methods vary in terms of type of segmentation used and the specific ways in which it is integrated in the VPR pipeline, e.g., planes [17], objects [15,31,45], landmarks [65], outdoor semantics [21,23,46,47], utility/confusion based [33, 38], domain adaptation [26] and even learning to segment for VPR [50]. However, neither these methods aim to perform segment-level retrieval nor do they use open-set segmentation. We also review two concurrent works: MESA [75] and Region-Revisited [61]. Similar to our method, they both use SAM to segment images but for different specific tasks. MESA [75] proposes a graph-based local feature/area matching method to obtain point correspondences. Our method complements this, as we perform coarse retrieval for VPR, which could potentially use MESA for reranking. Regions Revisited [61] delves into the advantages of using SAM masks in conjunction with SLIC [36] to improve semantic segmentation, activity recognition and object category retrieval. In contrast, we aim to improve instance-level recognition by recognizing specific things belonging to specific places that a robot encounters during a revisit. Similar to our work, [20] creates an image sequence-based topological graph of segments where its segment neighbourhood aggregation is based on average pooling, similar to [61]. In Section 5.3, we show that such segment average pooling deteriorates recognition performance for the VPR task." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.526, + 0.357, + 0.541 + ], + "angle": 0, + "content": "2.4 Open-set VPR" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.549, + 0.787, + 0.745 + ], + "angle": 0, + "content": "Researchers have recently started to shift their focus to open-set, generally-applicable techniques, including that for VPR [31, 32, 45, 62]. FM-Loc [45] uses GPT [9] to recognize object and place labels, whereas [31] uses CLIP [55] for open-set place recognition. LIP-Loc [62] proposes pretraining for cross-modal VPR, but is limited in its zero-shot capabilities. AnyLoc [32] proposes to use DINOv2 with domain-level vocabularies and hard-assignment based VLAD. It achieves state-of-the-art performance particularly on non-streetview datasets, where current VPR-trained methods tend to fail. In this work, we propose a generally-applicable approach which is built on top of models like SAM [37] and DINOv2 [49], and works with both VPR-agnostic [32] and VPR-specific [35] backbone models. We particularly aim for a new paradigm in retrieval based VPR, where we move away from the conventional whole image global descriptors to segments based descriptors and retrieval, which achieves a new state-of-the-art on diverse domains under wide viewpoint variations." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.766, + 0.416, + 0.784 + ], + "angle": 0, + "content": "3 Proposed Approach" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.795, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Despite recent advances in place recognition, viewpoint variations continue to be an open challenge for an embodied agent to recognize the same specific things in its environment. Current methods in visual place recognition tackle this problem by converting" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.628, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Revisit Anything" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "image", + "bbox": [ + 0.216, + 0.15, + 0.788, + 0.337 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.352, + 0.788, + 0.422 + ], + "angle": 0, + "content": "Fig. 1: Overview of our segment-retrieval based VPR pipeline, dubbed SegVLAD: We use open-set segmentor (SAM) to extract segment masks, which are converted into SuperSegments using the neighbouring image segments. Using pixel-level DINOv2 descriptors with VLAD-based aggregation, we obtain SuperSegment descriptors, which are matched against a flat index of SuperSegment descriptors obtained from all the images of the entire reference database." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.452, + 0.788, + 0.543 + ], + "angle": 0, + "content": "an image as a whole into a global descriptor, which does not explicitly deal with the problem of partial visual overlap caused by viewpoint variations. We propose an alternative solution by representing images partially with the help of image segments. In the following subsections, we describe our representation and retrieval method, which deviates from the conventional VPR techniques but creates a new capability in terms of recognizing objects/things that constitute a place." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.567, + 0.409, + 0.58 + ], + "angle": 0, + "content": "3.1 Problem Formulation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.593, + 0.788, + 0.7 + ], + "angle": 0, + "content": "We represent an image as a set of segment descriptors instead of a single global descriptor. For an image \\(I\\), we obtain binary image segment masks \\(M \\in \\{0,1\\}^{S \\times N}\\) and dense pixel-level descriptors \\(f_{p} \\in \\mathcal{R}^{D}\\), where \\(S\\) represents the number of segments per image, \\(D\\) is the descriptor dimension, and \\(p \\in [1,N]\\) represents spatial elements across the width \\((W)\\) and height \\((H)\\) of the image encoder's output, flattened into \\(N = W \\times H\\) for convenience. Figure 1 shows an illustration of our proposed pipeline, as explained in the following subsections." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.723, + 0.371, + 0.739 + ], + "angle": 0, + "content": "3.2 Super Segments" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Humans are remarkable at visual recognition, where existing studies suggest that we often leverage spatial associations among objects in an environment to represent it internally [5, 28]. This enables us to distinguish between two different scenes through the surrounding context of the objects of interest. In this work, we imbibe this context through the spatial neighbourhood of the image segments. For each image, we construct a graph of segments through their pixel centers using Delaunay Triangulation." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.117, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.116, + 0.355, + 0.129 + ], + "angle": 0, + "content": "K. Garg et al." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.144, + 0.787, + 0.228 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.239, + 0.788, + 0.296 + ], + "angle": 0, + "content": "Fig. 2: Neighborhood expansion (Eq. 1) of a window in the leftmost image to the whole building in the rightmost image, progressing from no neighborhood aggregation to a third-order aggregation. This neighborhood expansion is in stark contrast with a typical regular grid- or patch-based approach which may not capture semantically-meaningful SuperSegments." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.307, + 0.787, + 0.372 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.383, + 0.788, + 0.426 + ], + "angle": 0, + "content": "Fig. 3: Illustration of four SuperSegments obtained from the same image. All four of these spatially overlap with each other, which is different from coarse segmentation methods that do not typically allow overlap across segments." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.451, + 0.788, + 0.497 + ], + "angle": 0, + "content": "This provides us with a binary adjacency matrix \\( A \\in \\{0,1\\}^{S \\times S} \\) to define the neighborhood for individual segments. We use this adjacency information to expand the context of individual segments to generate new SuperSegment masks \\( (\\mathcal{M}) \\) as below:" + }, + { + "type": "equation", + "bbox": [ + 0.4, + 0.503, + 0.787, + 0.521 + ], + "angle": 0, + "content": "\\[\n\\mathcal {M} _ {S \\times N} = ^ {1} \\left(A _ {S \\times S} ^ {o} \\cdot M _ {S \\times N}\\right) \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.528, + 0.789, + 0.756 + ], + "angle": 0, + "content": "where \\( o \\geq 0 \\) refers to the order for expanding the neighborhood by multiplying the adjacency matrix by itself as \\( A^{o + 1} = A^o \\cdot A \\). This is matrix-multiplied with the original segmentation masks \\( M \\) to expand the neighborhood at pixel level. \\( \\mathcal{M} \\) is obtained after element-wise binarization (denoted with \\( \\mathbb{1}() \\)) so that all pixels in the SuperSegment mask may only contribute once to the subsequent feature aggregation. In Figure 2, we illustrate the extent of image area covered with different orders of mask expansion. Unlike, a patch or regular grid-based approach, the expanded mask of the window in the leftmost image covers a meaningful entity (building) in the rightmost image. Our approach to creating SuperSegments differs from coarse segmentation methods or superpixels in terms of the 'self-overlap'. By expanding neighborhood of each individual segment, we obtain several partially overlapping SuperSegments. A coarse segmentor will need to make assumptions about the right sub-segments to be coalesced so that it can enable accurate recognition from a different viewpoint, which could otherwise lead to the same limitation as that of the whole-image descriptors. Figure 3 presents examples of multiple overlapping SuperSegments from the same image." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.774, + 0.445, + 0.789 + ], + "angle": 0, + "content": "3.3 SuperSegment Descriptors" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.788, + 0.841 + ], + "angle": 0, + "content": "In this section, we describe our feature aggregation method to obtain SuperSegment descriptors. Recent state-of-the-art method AnyLoc [32] demonstrated that using off-the-shelf powerful image encoders such as DINOv2 with hard assignment based VLAD" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.628, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Revisit Anything" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.147, + 0.788, + 0.253 + ], + "angle": 0, + "content": "aggregation achieves superior recognition performance. However, AnyLoc does not use segmentation information and only operates at the whole-image level. More recently, [61] showed that average pooling works well for segment-level descriptors, but it didn't consider segment neighborhood information. In this work, we propose a unified formulation for feature aggregation that can easily switch across segments, segment neighborhood and the whole image as well as different aggregation types (see supplementary for details). This simply extends Equation 1 as below:" + }, + { + "type": "equation", + "bbox": [ + 0.378, + 0.267, + 0.786, + 0.285 + ], + "angle": 0, + "content": "\\[\nF _ {S \\times D} = ^ {1} \\left(A _ {S \\times S} ^ {o} \\cdot M _ {S \\times N}\\right) \\cdot T _ {N \\times D} \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.301, + 0.788, + 0.377 + ], + "angle": 0, + "content": "where \\( T \\) represents the features to be aggregated. By replacing \\( A \\) and \\( M \\) with ones matrices, one can obtain the whole-image global descriptor. For methods like Global Average Pooling (GAP), \\( T \\) can be directly used as the output of the image encoder. In our work, we use Hard-VLAD, for which \\( T \\) is the residual feature matrix per cluster and is obtained as below with respect to each of the cluster centers \\( c_k \\):" + }, + { + "type": "equation", + "bbox": [ + 0.354, + 0.391, + 0.787, + 0.41 + ], + "angle": 0, + "content": "\\[\nT _ {N _ {k} \\times D} ^ {k} = \\left\\{\\alpha_ {k} \\left(f _ {p}\\right) \\left(f _ {p} - c _ {k}\\right) \\mid \\alpha_ {k} \\left(f _ {p}\\right) = 1 \\right\\} \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.424, + 0.788, + 0.486 + ], + "angle": 0, + "content": "where \\(\\alpha_{k}(f_{p})\\in \\{0,1\\}\\) is 1 if \\(f_{p}\\) belongs to \\(c_{k}\\). The cluster centers (vocabulary) can be constructed using the map or the domain [32]. The SuperSegment VLAD descriptors obtained from Eq. 2 are intra-normalized, concatenated across clusters and then finally l2-normalized, following existing works [3, 32]." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.517, + 0.466, + 0.533 + ], + "angle": 0, + "content": "3.4 Image Retrieval via Segments" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.552, + 0.789, + 0.738 + ], + "angle": 0, + "content": "Existing global descriptor based VPR techniques produce a single vector per image to search against a database of reference image vectors. In our method, we obtain multiple SuperSegment descriptors per image. We perform retrieval at segment-level, that is, we search for the top matches for each query segment against a flat index of all segments from all the images of the reference database/map. To evaluate in the form of image retrieval-based VPR, we convert the top retrieved segment indices across all segments of a query image into top reference image indices. This is achieved through a weighted frequency measure (i.e., weighted bin/word counting). We first map the top \\( K' \\) retrieved segment indices for each of the query segments \\( s \\in [1, S] \\) to their respective reference image indices, denoted with \\( r \\). Then, for each of the unique retrieved image indices \\( r_j \\), we accumulate its segment similarity \\( \\theta \\) and then use the cumulative similarity score \\( \\hat{\\theta} \\) to rank the image indices to obtain the top image match \\( r_j^* \\):" + }, + { + "type": "equation", + "bbox": [ + 0.319, + 0.753, + 0.787, + 0.797 + ], + "angle": 0, + "content": "\\[\nr _ {j} ^ {*} = \\underset {r _ {j}} {\\operatorname {a r g m a x}} \\hat {\\theta} \\left(r _ {j}\\right); \\quad \\hat {\\theta} \\left(r _ {j}\\right) = \\sum_ {s = 1} ^ {S} \\sum_ {k = 1} ^ {K ^ {\\prime}} \\theta_ {s k} \\cdot \\mathbb {1} _ {\\left\\{r _ {s k} = r _ {j} \\right\\}} \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.81, + 0.788, + 0.841 + ], + "angle": 0, + "content": "In Section 5.3, we compare our similarity-weighted ranking with other alternatives based on frequency or similarity alone." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.116, + 0.355, + 0.129 + ], + "angle": 0, + "content": "K. Garg et al." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.416, + 0.163 + ], + "angle": 0, + "content": "4 Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.177, + 0.787, + 0.283 + ], + "angle": 0, + "content": "Datasets: VPR datasets are in abundance, as can be found in several benchmarks including VPR-Bench [74], Deep Visual GeoLocalization Benchmark [7], and AnyLoc [32]. In this work, we used a variety of datasets covering both outdoor and indoor environments. Outdoor datasets include Pitts30k [66], AmsterTime [72], Mapillary Street Level Sequences (MSLS) [69], SF-XL [6], VPAir [59], Revisted Oxford5K and Revisited Paris6k [53]. Indoor datasets include Baidu Mall [64], 17Places [76] and InsideOut [27]. Additional datasets-related details are provided in the supplementary." + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.298, + 0.787, + 0.615 + ], + "angle": 0, + "content": "Evaluation and Benchmarking: We evaluate our method as an image retrieval task using Recall@K metric, where top K' (= 50) retrieved segments per query segment are used to obtain top K (= 5) images (see Eq. 4). We compare against the most recent and high-performing VPR baseline methods. This includes CosPlace [6], MixVPR [2] and EigenPlaces [8], which are trained on large-scale urban datasets for VPR tasks. We further include two very recent state-of-the-art methods that use DINOv2 as the backbone. These include AnyLoc [32] which uses an off-the-shelf DINOv2 model and SALAD [29] which uses a finetuned DINOv2 backbone. Given the dichotomy between general-purpose VPR benchmarking of AnyLoc and the typical outdoor-focused benchmarking in SALAD and other methods, we evaluated our method using two different backbones. a) SegVLAD-PreT: we use the same backbone and aggregation as AnyLoc, i.e., off-the-shelf pretrained DINOv2 (ViT-G) backbone with hard VLAD assignment, but the key difference is in the use of SuperSegments for our method as opposed to whole-image description of AnyLoc. b) SegVLAD-FineT: since our default aggregation method is VLAD, we use a finetuned DINOv2 (ViT-B) backbone which is similar to SALAD but replaces its aggregation layer with the original NetVLAD aggregation [3] using 64 clusters, as described in [35]. We use this finetuned backbone with hard VLAD based assignment, similar to AnyLoc. For both these models, we reduce the descriptor dimensions of the VLAD descriptor to 1024 using PCA, as commonly done in previous works [3, 32]. We train PCA transform in a map-specific manner using the database images of the dataset." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.638, + 0.311, + 0.655 + ], + "angle": 0, + "content": "5 Results" + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.671, + 0.787, + 0.731 + ], + "angle": 0, + "content": "We first present benchmark comparison of our method against state-of-the-art VPR methods. This is followed by detailed analysis of our proposed aggregation technique. Lastly, we demonstrate results on a downstream task of Object of Interest (OOI) retrieval, showcasing the versatility of our method." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.753, + 0.456, + 0.77 + ], + "angle": 0, + "content": "5.1 State-of-the-art comparisons" + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.779, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Table 1 presents Recall@1/5 comparison against state-of-the-art VPR methods on standard outdoor street-view datasets, which are similar to the typical training datasets used for VPR [2,3,6]. Table 2 covers 'out-of-distribution' datasets, inspired by AnyLoc [32], covering indoor environments (Baidu Mall and 17 Places), aerial imagery (VPAir)," + } + ], + [ + { + "type": "header", + "bbox": [ + 0.628, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Revisit Anything" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "table_caption", + "bbox": [ + 0.272, + 0.145, + 0.731, + 0.159 + ], + "angle": 0, + "content": "Table 1: Recall@ 1/5 benchmark comparison on outdoor street-view datasets." + }, + { + "type": "table", + "bbox": [ + 0.247, + 0.171, + 0.756, + 0.284 + ], + "angle": 0, + "content": "
MethodPitts-30KMSLS SFMSLS CPHSF-XL ValRO5k MedRO5k HardRP6k MedRP6k Hard
CosPlace90.4/95.793.4/97.584.9/92.094.6/97.685.7/87.127.1/45.794.3/95.77.1/15.7
MixVPR91.5/95.591.3/95.987.1/92.487.8/93.868.6/80.032.9/54.394.3/10010.0/32.9
EigenPlaces92.6/96.792.6/97.187.1/92.896.4/98.285.7/88.642.8/57.195.7/98.64.3/11.4
AnyLoc87.7/94.783.4/94.679.9/89.184.4/91.988.6/92.940.0/58.697.1/10011.4/44.3
SALAD92.6/96.591.7/97.192.3/96.193.6/97.382.9/90.037.1/54.395.7/98.614.3/58.6
SegVLAD-PreT86.7/94.288.4/94.281.7/90.790.9/96.491.4/95.760.0/81.494.3/1008.6/48.6
SegVLAD-FineT93.2/96.894.6/97.190.9/95.794.9/98.187.1/95.751.4/70.095.7/10010.0/48.6
" + }, + { + "type": "table_caption", + "bbox": [ + 0.271, + 0.298, + 0.733, + 0.312 + ], + "angle": 0, + "content": "Table 2: Recall@ 1/5 benchmark comparison on 'out-of-distribution' datasets." + }, + { + "type": "table", + "bbox": [ + 0.319, + 0.324, + 0.688, + 0.43 + ], + "angle": 0, + "content": "
MethodBaiduAmstTimeInsideOut17PlacesVPAir
CosPlace41.6/55.047.7/69.80.2/2.081.3/88.24.6/13.7
MixVPR64.4/80.340.2/59.10.0/1.885.2/90.16.8/16.1
EigenPlaces56.5/72.848.9/69.50.4/1.483.0/90.16.5/17.9
AnyLoc75.2/87.650.3/73.02.4/8.095.3/97.366.7/79.2
SALAD74.8/86.555.4/75.60.6/1.882.5/88.225.8/38.7
SegVLAD-PreT78.5/93.856.8/77.74.2/9.495.3/98.069.8/83.7
SegVLAD-FineT68.1/89.058.9/79.37.4/15.695.1/97.535.4/55.3
" + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.459, + 0.788, + 0.536 + ], + "angle": 0, + "content": "indoor-to-outdoor viewing (InsideOut), and historical image matching (AmsterTime). Below, we discuss two key aspects of this comparative analysis: \\(i\\) ) how our segment-based approach compares against whole-image global descriptor based methods and ii) how performance trends vary depending on the choices of feature backbone with regards to task-specific (VPR) training." + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.559, + 0.788, + 0.771 + ], + "angle": 0, + "content": "Aggregating Segments vs Whole Images Table 1 and Table 2 show that our proposed method SegVLAD achieves a new state-of-the-art on the majority of datasets, considering both the backbone variants: PreT and FineT. AnyLoc and SALAD respectively differ from SegVLAD-PreT and SegVLAD-FineT in terms of the aggregation scope (global vs segments). Thus, the superior performance of SegVLAD clearly highlights the role of segments based retrieval over whole-image based approach. On the Baidu Mall dataset – highly-aliased indoor environment – our method improves over AnyLoc by \\(3\\%\\) for R@1 and \\(6\\%\\) for R@5 in absolute gains. On the InsideOut dataset – matching outdoor images viewed from within indoors – our method leads to a ‘meaningful’ recall, unlike all other baselines. Overall, these results highlight that even with the use of powerful image encoders (DINOv2), global aggregation struggles to deal with the challenges of matching images across major viewpoint shifts – it is thus the partial image representation and matching (using the same encoder) which is needed to obtain superior recognition performance." + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.795, + 0.788, + 0.841 + ], + "angle": 0, + "content": "VPR Fine-tuned Encoders + Segments For SegVLAD-FineT, we used a DINOv2 backbone finetuned for the purpose of VPR, mainly to observe the benefit of segments over global descriptor based approach in a task-specific manner. Table 1 and Table 2" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.355, + 0.129 + ], + "angle": 0, + "content": "K. Garg et al." + }, + { + "type": "table_caption", + "bbox": [ + 0.252, + 0.145, + 0.751, + 0.16 + ], + "angle": 0, + "content": "Table 3: Recall@1 results for various approaches on Object-Instance Retrieval Task" + }, + { + "type": "table", + "bbox": [ + 0.224, + 0.171, + 0.779, + 0.204 + ], + "angle": 0, + "content": "
MethodSegVLAD NoNbrAggSegVLADSegment-to-GlobalGlobal-to-Global
R@164.192.730.086.4
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.239, + 0.785, + 0.299 + ], + "angle": 0, + "content": "show that, on the outdoor street-view datasets, SALAD (finetuned DINOv2) generally performs better than AnyLoc (its pretrained counterpart), whereas the latter generally outperforms the former on 'out-of-distribution' datasets. It can be clearly observed that these performance patterns translate well from global- to segment-level results." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.326, + 0.695, + 0.342 + ], + "angle": 0, + "content": "5.2 Revisiting Objects of Interest (OOI): Object Instance Retrieval" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.356, + 0.788, + 0.567 + ], + "angle": 0, + "content": "A typical requirement of an embodied agent is to understand the context of its task through its memory/map information, which is composed of visual and/or semantic cues. For example, navigating to a given object goal requires a robot to visually recognize the goal and not be confused by perceptually-similar items. In this section, we demonstrate our method's ability to retrieve the correct image given just an Object Of Interest (OOI) as a query segment. For this purpose, we use an extended version of the Baidu dataset [70] which annotates OOI as various discriminative areas that can be reliably detected under variable viewpoint and lighting conditions. In total, there are 220 OOI, which cover various things such as logos, brand names, posters, etc., in a highly cluttered mall environment. To cast this dataset in terms of revisiting things, we use the original query images of the Baidu Mall [64] dataset as the database and the images with OOI as the queries. This allows us to evaluate the OOIs directly. This is similar to VPR evaluation of recall in terms of image retrieval but with querying of a specific segment instead of using all the segments of the query image." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.568, + 0.788, + 0.748 + ], + "angle": 0, + "content": "We consider four different methods of recognizing known objects in this study. i) Global-to-Global: as a baseline method, we use whole images to represent and retrieve, i.e., without using the OOI mask; this resembles object-goal recognition problem for an InstanceImageNavigation task [39]. ii) Segment-to-Global: this is the same as the previous setting except that the query image descriptor is aggregated only using the OOI mask; this tests the ability of the image encoder/aggregator to match segment-level descriptor against global descriptors. iii) SegVLAD and iv) SegVLAD NoNbrAgg, which are our proposed methods but the latter does not use any neighborhood information; this highlights the relevance of spatial context around the OOI for recognition. For SegVLAD, we create a virtual segment mask for the OOI, append it to the other masks of the image, and then perform our neighborhood expansion and feature aggregation, as described in Section 3." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Table 3 reports Recall@1 for different recognition methods. It can be observed that SegVLAD outperforms Global-to-Global matching by a large margin, which shows that recognizing specific object instances through their images (as in InstanceImageNav) is more prone to failures. It can further be observed from low recall of SegVLAD NoNbrAgg that neighborhood aggregation around segments is crucial to capture the required context. Finally, poor recall of Segment-to-Global highlights" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.628, + 0.115, + 0.732, + 0.13 + ], + "angle": 0, + "content": "Revisit Anything" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.145, + 0.784, + 0.19 + ], + "angle": 0, + "content": "Table 4: Recall@1/5 for Baidu mall dataset Table 5: Recall@1/5 comparison between for different aggregation methods and dif- different methods for ranking images based. ferent orders of neighborhood expansion. on segment-level retrieval." + }, + { + "type": "table", + "bbox": [ + 0.246, + 0.213, + 0.756, + 0.294 + ], + "angle": 0, + "content": "
OrderSegVLADSAP
073.1/89.974.6/91.1
177.4/91.765.6/87.2
276.3/92.453.2/81.3
377.7/92.649.8/78.0
Max Seg78.5/93.953.9/70.4
Max Sim65.2/92.734.4/62.4
Ours78.5/93.854.4/76.3
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.324, + 0.784, + 0.355 + ], + "angle": 0, + "content": "that matching a part of an image (OOI) with the whole image is not a viable solution for object instance recognition." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.379, + 0.373, + 0.394 + ], + "angle": 0, + "content": "5.3 Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.407, + 0.788, + 0.71 + ], + "angle": 0, + "content": "Aggregation method & Order of Neighborhood Expansion Previous studies on VPR such as AnyLoc [32] have shown VLAD to be better than other aggregation methods for whole-image based global descriptors. However, in an increasing number of segment-based approaches [12, 20, 24, 61], segment average pooling (SAP) is used more commonly. Thus, we compare hard-assignment VLAD against SAP on Baidu dataset. For SAP, we upsample the DINOv2 features to match the resolution of our SAM masks – this upsampling is shown to enhance performance in [61]. For VLAD aggregation, we use our proposed method SegVLAD, where we downsample masks to match with the low resolution of DINOv2. Table 4 shows that SAP performs well for order 0 aggregation (i.e., no neighborhood aggregation) but its performance reduces as the neighborhood expands. On the other hand, SegVLAD has low recall when no neighborhood is considered but benefits significantly even with its immediate neighborhood (order 1). We attribute these inverted trends of SegVLAD and SAP to the very nature of these aggregation methods: as more information becomes available SAP smoothes out the overall information content whereas SegVLAD benefits from additional information which gets distributed across its clusters, thus minimizing any possible smoothing effect. It can be observed that R@5 increases for SegVLAD with an increasing order of neighborhood expansion but margins diminish for higher orders. Overall, SegVLAD (order 3) achieves the best results, despite aggregating at a \\(14 \\times\\) lower resolution than SAP's upsampling based aggregation." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Segment to Image Retrieval Unlike conventional global descriptor retrieval based VPR, we perform retrieval for multiple SuperSegments of the query image. To obtain retrieval output in terms of images (as that is what VPR is typically evaluated on), there exist multiple ways to combine the top segment-level matches across all the query segments. We consider the following alternative options. i) MaxSeg: we obtain the best matching segment for each query segment, associate the matched segments to their respective reference image indices, and then rank these image indices based on their" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.355, + 0.129 + ], + "angle": 0, + "content": "K. Garg et al." + }, + { + "type": "image", + "bbox": [ + 0.235, + 0.143, + 0.77, + 0.452 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.465, + 0.788, + 0.507 + ], + "angle": 0, + "content": "Fig. 4: Qualitative results: Columns respectively represent the query image, correct match of SegVLAD and incorrect match of AnyLoc. Examples from different datasets: AmsterTime, Baidu Mall, Pitts-30K are presented across the rows." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.548, + 0.788, + 0.685 + ], + "angle": 0, + "content": "frequency; this method weakly resembles an inverted index list based counting of common segments between the query and the reference. ii) MaxSim: across the best matching reference segments, we order their image indices based on the segment similarity; this method is similar to that used in MultiVLAD [4]. iii) Similarity-Weighted Frequency: this is our proposed method as defined in Section 3.4. Table 5 shows that our proposed method for combining segment-level hypotheses consistently achieves superior results for both the datasets. While MaxSeg achieves a similar performance on Baidu, it suffers a drop in recall for AmsterTime. Both the methods outperform MaxSim at R@1 by a large margin." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.788, + 0.84 + ], + "angle": 0, + "content": "Patches vs Segment While open-set segmentation [37] is aimed at a meaningful segregation of visual entities, a simple alternative to our segment-based retrieval is to use uniformly defined regions/patches. Table 6 compares SegVLAD with a patch-based approach, where we consider arbitrary square patch sizes to segment an image. It can be observed that SegVLAD outperforms its patch-based counterparts, where smaller patches perform the worst, and for larger patches, R@1 saturates while R@5 reduces. These results are also in line with similar findings of a recent work [61]." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.628, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Revisit Anything" + }, + { + "type": "page_number", + "bbox": [ + 0.77, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "image", + "bbox": [ + 0.237, + 0.148, + 0.472, + 0.256 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.271, + 0.493, + 0.298 + ], + "angle": 0, + "content": "Fig.5: Recall vs storage/retrieval time on AmsterTime." + }, + { + "type": "table_caption", + "bbox": [ + 0.51, + 0.169, + 0.787, + 0.196 + ], + "angle": 0, + "content": "Table 6: Patch vs Segments on AmsterTime (Reso. \\(256\\times 256\\)" + }, + { + "type": "table", + "bbox": [ + 0.522, + 0.198, + 0.77, + 0.29 + ], + "angle": 0, + "content": "
Patch SizeNumSegDbNumSegImR@1/5
16×1631513625635.7/62.1
32×32787846445.9/72.5
64×64196961652.0/75.1
128×1284924453.5/65.6
SegVLAD12963710556.8/77.7
" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.329, + 0.676, + 0.346 + ], + "angle": 0, + "content": "5.4 Limitations, Storage, Compute Time & IOU Based Filtering" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.355, + 0.788, + 0.581 + ], + "angle": 0, + "content": "A key limitation of our method is its large map size, i.e., large storage requirement for segment-level descriptors (see supplementary for further details). In this section, we analyze the resource requirements in terms of database (index) storage and query retrieval time for our method, along with a preliminary study on IOU (Intersection over Union) based filtering of SuperSegments to reduce such costs. We compute IOU between all pairs of SuperSegments in a given image, and remove segments with \\( IOU(s_i, s_j) > \\psi \\forall i \\in [1, S] \\), \\( j \\in [i, S] \\), where \\( \\psi \\in [0, 1] \\) is a threshold on IOU and \\( s_i \\) refers to the list of SuperSegments sorted by their pixel area in a descending order. We only perform this culling on the database segments. We use the outdoor-finetuned DINOv2 backbone for this purpose and compare SegVLAD with SALAD on AmsterTime (see supplementary for additional results on Pitts30K). Figure 5 shows that SegVLAD outperforms SALAD while requiring less storage (annotated on points) and comparable retrieval time (excludes extraction time), using IoU-based filtering threshold ranging from 0.2 to 0.8 (left to right) with a step size of 0.2. In particular, at 0.4 IOU threshold, only 20% of SuperSegments are retained (0.05 GB) while still outperforming the baseline." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.604, + 0.403, + 0.619 + ], + "angle": 0, + "content": "5.5 Qualitative Analyses" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.629, + 0.788, + 0.765 + ], + "angle": 0, + "content": "In this section, we further demonstrate the capabilities of our method through qualitative visualizations. We compare our method against AnyLoc, where the only difference between the two methods is Global aggregation/retrieval vs SuperSegment aggregation/retrieval. We particularly consider the queries for which our approach successfully retrieved the correct match but AnyLoc failed to do so (additional examples can be found in the supplementary). Figure 4 shows triplets of images in the order of query, correct match (ours), and incorrect match (Anyloc). The segmented part shows one of the correctly matched SuperSegments, displayed as a subgraph in white color overlaid on the corresponding segment masks." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.766, + 0.788, + 0.841 + ], + "angle": 0, + "content": "The first row shows a triplet from AmsterTime where our proposed method is able to correctly recognize a subgraph of building across the image pair, whereas Anyloc retrieves an incorrect image of a street-view with buildings, cars, and road, laid out similarly across the image pair. This highlights that a global descriptor can get confused with perceptually-aliased global context. In contrast, our SuperSegment based" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.355, + 0.129 + ], + "angle": 0, + "content": "K. Garg et al." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.147, + 0.788, + 0.359 + ], + "angle": 0, + "content": "SegVLAD is not only able to retrieve the correct image but it also correctly finds the mutually-overlapping area. A similar trend follows for the Baidu Mall (middle row) and Pitts30K (last row). In the Baidu Mall example, our approach identifies the piano and the region around it in the query image. It correctly retrieves an image having similar spatial context with the piano. This is akin to how humans use spatial context to identify places. AnyLoc, on the other hand, retrieves an image with similar floor tiles and railings. This example particularly highlights our hypothesis that dissimilarity of nonoverlapping regions can dominate the similarity of overlapping regions in global whole-image descriptors. Finally, the Pitts30K example (last row) shows a case of strong viewpoint change. While SegVLAD correctly matches the traffic signal and signboards to retrieve the correct match, AnyLoc retrieves a similar looking image while missing the finer details. This example particularly reinforces the idea of 'revisiting things', as even though the background mountain is common across the triplet, it is the context of the things near the camera/robot which helps in uniquely recognizing a place." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.382, + 0.495, + 0.399 + ], + "angle": 0, + "content": "6 Conclusion and Future Work" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.414, + 0.788, + 0.655 + ], + "angle": 0, + "content": "In this paper, we presented a novel visual place recognition pipeline SegVLAD based on image segments-based description and retrieval, which is akin to 'revisting things' as a means to recognize specific instances of what constitute a place. Our proposed SuperSegments based image representation and a novel factorization based feature aggregation enables us to effectively represent and retrieve images using our segment similarity-weighted image ranking. Our results show that despite using powerful image encoders such as DINOv2 (pretrained or VPR-finetuned), existing global descriptor based techniques are unable to deal with the challenges of viewpoint variations. In contrast, SegVLAD is able to correctly retrieve images through its ability to match partially-overlapping images with its partial image representation in the form of semiglobal subgraphs of segments, i.e., SuperSegments. Thus, our method achieves state-of-the-art results on three diverse datasets (indoor and outdoor) that exhibit strong viewpoint variations on top of other challenges of appearance shift and high perceptual aliasing. Through an additional object instance retrieval study, we demonstrate the unique ability of our method to recognize objects instances within their specific place contexts - an open-set recognition capability that existing VPR methods lack." + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.656, + 0.788, + 0.807 + ], + "angle": 0, + "content": "Our approach shifts the paradigm in retrieval based VPR research, as the conventional methods predominantly classify into either whole-image global descriptor based coarse retrieval or local feature based geometric reranking. Our approach complements recent concurrent works like MESA [75]; future work can explore a hierarchical VPR pipeline that closely integrates a segment-based coarse retriever with segments-based reranker such as MESA, thus doing away entirely with global whole-image descriptors. Furthermore, segments-based representation with implicitly baked semantics provide a natural way for creating text-based interfaces through CLIP [55] and LLMs (Large Language Models) [9], which can be easily integrated with recent efforts in this direction [12, 18, 20, 24, 42, 43, 77]." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.628, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Revisit Anything" + }, + { + "type": "page_number", + "bbox": [ + 0.77, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.145, + 0.383, + 0.163 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.178, + 0.788, + 0.284 + ], + "angle": 0, + "content": "This work was supported by the Centre for Augmented Reasoning (CAR) at the Australian Institute for Machine Learning (AIML), University of Adelaide, Australia. The authors acknowledge the computational support provided by the Indian Institute of Science (IISc), Bengaluru, India, and the International Institute of Information Technology, Hyderabad (IIITH), India. The authors thank Ahmad Khaliq for technical support, Sarah Ibrahimi for sharing the InsideOut dataset, and Martin Humenberger for sharing the OOI annotations for the Baidu Mall dataset." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.308, + 0.312, + 0.323 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.34, + 0.785, + 0.368 + ], + "angle": 0, + "content": "1. Ali-bey, A., Chaib-draa, B., Giguère, P.: Gsv-cities: Toward appropriate supervised visual place recognition. Neurocomputing 513, 194-203 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.369, + 0.785, + 0.41 + ], + "angle": 0, + "content": "2. Ali-bey, A., Chaib-draa, B., Giguère, P.: Mixvpr: Feature mixing for visual place recognition. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 2998-3007 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.411, + 0.785, + 0.451 + ], + "angle": 0, + "content": "3. Arandjelovic, R., Gronat, P., Torii, A., Pajdla, T., Sivic, J.: Netvlad: Cnn architecture for weakly supervised place recognition. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 5297-5307 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.452, + 0.785, + 0.479 + ], + "angle": 0, + "content": "4. Arandjelovic, R., Zisserman, A.: All about vlad. In: Proceedings of the IEEE conference on Computer Vision and Pattern Recognition. pp. 1578-1585 (2013)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.48, + 0.758, + 0.493 + ], + "angle": 0, + "content": "5. Bar, M.: Visual objects in context. Nature Reviews Neuroscience 5(8), 617-629 (2004)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.494, + 0.785, + 0.535 + ], + "angle": 0, + "content": "6. Berton, G., Masone, C., Caputo, B.: Rethinking visual geo-localization for large-scale applications. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4878-4888 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.536, + 0.785, + 0.577 + ], + "angle": 0, + "content": "7. Berton, G., Mereu, R., Trivigno, G., Masone, C., Csurka, G., Sattler, T., Caputo, B.: Deep visual geo-localization benchmark. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5396-5407 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.578, + 0.785, + 0.618 + ], + "angle": 0, + "content": "8. Berton, G., Trivigno, G., Caputo, B., Masone, C.: Eigenplaces: Training viewpoint robust models for visual place recognition. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 11080-11090 (October 2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.619, + 0.785, + 0.659 + ], + "angle": 0, + "content": "9. Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J.D., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., et al.: Language models are few-shot learners. Advances in neural information processing systems 33, 1877-1901 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.66, + 0.785, + 0.701 + ], + "angle": 0, + "content": "0. Caesar, H., Uijlings, J., Ferrari, V.: Coco-stuff: Thing and stuff classes in context. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1209-1218 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.702, + 0.785, + 0.73 + ], + "angle": 0, + "content": "1. Cao, B., Araujo, A., Sim, J.: Unifying deep local and global features for image search. In: European Conference on Computer Vision. pp. 726-743. Springer (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.731, + 0.785, + 0.757 + ], + "angle": 0, + "content": "2. Chang, M., Gervet, T., Khanna, M., Yenamandra, S., Shah, D., Min, S.Y., Shah, K., Paxton, C., Gupta, S., Batra, D., et al.: Goat: Go to any thing. arXiv preprint arXiv:2311.06430 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.758, + 0.785, + 0.799 + ], + "angle": 0, + "content": "3. Chen, L.C., Yang, Y., Wang, J., Xu, W., Yuille, A.L.: Attention to scale: Scale-aware semantic image segmentation. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 3640-3649 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.8, + 0.785, + 0.84 + ], + "angle": 0, + "content": "4. Chen, Z., Maffra, F., Sa, I., Chli, M.: Only look once, mining distinctive landmarks from convnet for visual place recognition. In: Intelligent Robots and Systems (IROS), 2017 IEEE/RSJ International Conference on. pp. 9-16. IEEE (2017)" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.34, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.355, + 0.129 + ], + "angle": 0, + "content": "K. Garg et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.148, + 0.787, + 0.189 + ], + "angle": 0, + "content": "15. Cheng, C., Page, D.L., Abidi, M.A.: Object-based place recognition and loop closing with jigsaw puzzle image segmentation algorithm. In: 2008 IEEE International Conference on Robotics and Automation. pp. 557-562. IEEE (2008)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.19, + 0.787, + 0.218 + ], + "angle": 0, + "content": "16. Cummins, M., Newman, P.: Appearance-only slam at large scale with fab-map 2.0. The International Journal of Robotics Research 30(9), 1100-1123 (2011)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.219, + 0.787, + 0.26 + ], + "angle": 0, + "content": "17. Cupec, R., Nyarko, E.K., Filko, D., Kitanov, A., Petrovic, I.: Place recognition based on matching of planar surfaces and line segments. The International Journal of Robotics Research 34(4-5), 674-704 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.261, + 0.787, + 0.289 + ], + "angle": 0, + "content": "18. Folorunsho, S.O.: Semantic segmentation-based approach for autonomous navigation in challenging farm terrains. algorithms 15, 3 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.29, + 0.787, + 0.317 + ], + "angle": 0, + "content": "19. Garg, S., Fischer, T., Milford, M.: Where is your place, visual place recognition? In: IJCAI (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.318, + 0.787, + 0.359 + ], + "angle": 0, + "content": "20. Garg, S., Rana, K., Hosseinzadeh, M., Mares, L., Suenderhauf, N., Dayoub, F., Reid, I.: Robohop: Segment-based topological map representation for open-world visual navigation. In: 2024 IEEE International Conference on Robotics and Automation (ICRA) (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.36, + 0.787, + 0.401 + ], + "angle": 0, + "content": "21. Garg, S., Suenderhauf, N., Milford, M.: Lost? appearance-invariant place recognition for opposite viewpoints using visual semantics. In: Proceedings of Robotics: Science and Systems XIV (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.402, + 0.787, + 0.47 + ], + "angle": 0, + "content": "22. Garg, S., Sunderhauf, N., Dayoub, F., Morrison, D., Cosgun, A., Carneiro, G., Wu, Q., Chin, T.J., Reid, I., Gould, S., Corke, P., Milford, M.: Semantics for robotic mapping, perception and interaction: A survey. Foundations and Trends® in Robotics 8(1-2), 1-224 (2020). https://doi.org/10.1561/2300000059, http://dx.doi.org/10.1561/2300000059" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.472, + 0.787, + 0.5 + ], + "angle": 0, + "content": "23. Gawel, A., Del Don, C., Siegwart, R., Nieto, J., Cadena, C.: X-view: Graph-based semantic multi-view localization. IEEE Robotics and Automation Letters 3(3), 1687-1694 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.501, + 0.787, + 0.543 + ], + "angle": 0, + "content": "24. Gu, Q., Kuwajerwala, A., Morin, S., Jatavallabhula, K.M., Sen, B., Agarwal, A., Rivera, C., Paul, W., Ellis, K., Chellappa, R., et al.: Conceptgraphs: Open-vocabulary 3d scene graphs for perception and planning. arXiv preprint arXiv:2309.16650 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.544, + 0.787, + 0.585 + ], + "angle": 0, + "content": "25. Hausler, S., Garg, S., Xu, M., Milford, M., Fischer, T.: Patch-netvlad: Multi-scale fusion of locally-global descriptors for place recognition. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14141-14152 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.586, + 0.787, + 0.627 + ], + "angle": 0, + "content": "26. Hu, H., Qiao, Z., Cheng, M., Liu, Z., Wang, H.: Dasgil: Domain adaptation for semantic and geometric-aware image-based localization. IEEE Transactions on Image Processing 30, 1342-1353 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.628, + 0.787, + 0.655 + ], + "angle": 0, + "content": "27. Ibrahimi, S., Van Noord, N., Alpherts, T., Worring, M.: Inside out visual place recognition. arXiv preprint arXiv:2111.13546 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.657, + 0.787, + 0.683 + ], + "angle": 0, + "content": "28. Intraub, H.: The representation of visual scenes. Trends in cognitive sciences 1(6), 217-222 (1997)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.685, + 0.776, + 0.699 + ], + "angle": 0, + "content": "29. Izquierdo, S., Civera, J.: Optimal transport aggregation for visual place recognition (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.7, + 0.787, + 0.727 + ], + "angle": 0, + "content": "30. Jégou, H., Douze, M., Schmid, C., Pérez, P.: Aggregating local descriptors into a compact image representation. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 3304-3311 (2010)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.728, + 0.787, + 0.755 + ], + "angle": 0, + "content": "31. Kassab, C., Mattamala, M., Fallon, M.: Clip-based features achieve competitive zero-shot visual localization. OpenReview preprint arXiv:2306.14846 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.757, + 0.787, + 0.797 + ], + "angle": 0, + "content": "32. Keetha, N., Mishra, A., Karhade, J., Jatavallabhula, K.M., Scherer, S., Krishna, M., Garg, S.: Anyloc: Towards universal visual place recognition. IEEE Robotics and Automation Letters (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.799, + 0.787, + 0.84 + ], + "angle": 0, + "content": "33. Keetha, N.V., Milford, M., Garg, S.: A hierarchical dual model of environment-and place-specific utility for visual place recognition. IEEE Robotics and Automation Letters 6(4), 6969-6976 (2021)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.148, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.628, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Revisit Anything" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.189 + ], + "angle": 0, + "content": "34. Khaliq, A., Ehsan, S., Milford, M., McDonald-Maier, K.: A holistic visual place recognition approach using lightweight cnns for severe viewpoint and appearance changes. arXiv preprint arXiv:1811.03032 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.189, + 0.788, + 0.23 + ], + "angle": 0, + "content": "35. Khaliq, A., Xu, M., Hausler, S., Milford, M., Garg, S.: Vlad-buff: Burst-aware fast feature aggregation for visual place recognition. In: European Conference on Computer Vision. Springer (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.23, + 0.788, + 0.271 + ], + "angle": 0, + "content": "36. Khorasgani, S.H., Chen, Y., Shkurti, F.: Slic: Self-supervised learning with iterative clustering for human action videos. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16091-16101 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.271, + 0.788, + 0.298 + ], + "angle": 0, + "content": "37. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., et al.: Segment anything. arXiv preprint arXiv:2304.02643 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.298, + 0.788, + 0.325 + ], + "angle": 0, + "content": "38. Knopp, J., Sivic, J., Pajdla, T.: Avoiding confusing features in place recognition. In: European Conference on Computer Vision. pp. 748-761. Springer (2010)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.325, + 0.788, + 0.365 + ], + "angle": 0, + "content": "39. Krantz, J., Lee, S., Malik, J., Batra, D., Chaplot, D.S.: Instance-specific image goal navigation: Training embodied agents to find object instances. arXiv preprint arXiv:2211.15876 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.365, + 0.788, + 0.393 + ], + "angle": 0, + "content": "40. Le, D.C., Youn, C.H.: City-scale visual place recognition with deep local features based on multi-scale ordered VLAD pooling. arXiv preprint arXiv:2009.09255 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.393, + 0.788, + 0.419 + ], + "angle": 0, + "content": "41. Lowe, D.G.: Distinctive image features from scale-invariant keypoints. International journal of computer vision 60(2), 91-110 (2004)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.42, + 0.788, + 0.46 + ], + "angle": 0, + "content": "42. Maalouf, A., Jadhav, N., Jatavallabhula, K.M., Chahine, M., Vogt, D.M., Wood, R.J., Torralba, A., Rus, D.: Follow anything: Open-set detection, tracking, and following in real-time. IEEE Robotics and Automation Letters 9(4), 3283-3290 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.46, + 0.788, + 0.487 + ], + "angle": 0, + "content": "43. Manglani, S.: Real-time vision-based navigation for a robot in an indoor environment. arXiv preprint arXiv:2307.00666 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.487, + 0.788, + 0.514 + ], + "angle": 0, + "content": "44. Masone, C., Caputo, B.: A survey on deep visual place recognition. IEEE Access 9, 19516-19547 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.514, + 0.788, + 0.554 + ], + "angle": 0, + "content": "45. Mirjalili, R., Krawez, M., Burgard, W.: Fm-loc: Using foundation models for improved vision-based localization. In: Robotics and Automation (ICRA), 2023 IEEE International Conference on. IEEE (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.554, + 0.788, + 0.582 + ], + "angle": 0, + "content": "46. Mousavian, A., Kosecka, J.: Semantic image based geolocation given a map (author's initial manuscript). Tech. rep., George Mason University Fairfax United States (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.582, + 0.788, + 0.622 + ], + "angle": 0, + "content": "47. Naseer, T., Oliveira, G.L., Brox, T., Burgard, W.: Semantics-aware visual localization under challenging perceptual conditions. In: IEEE International Conference on Robotics and Automation (ICRA) (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.622, + 0.719, + 0.636 + ], + "angle": 0, + "content": "48. Oliva, A.: Gist of the scene. Neurobiology of attention 696(64), 251-258 (2005)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.636, + 0.788, + 0.704 + ], + "angle": 0, + "content": "49. Oquab, M., Darcet, T., Moutakanni, T., Vo, H.V., Szafraniec, M., Khalidov, V., Fernandez, P., Haziza, D., Massa, F., El-Nouby, A., Howes, R., Huang, P.Y., Xu, H., Sharma, V., Li, S.W., Galuba, W., Rabbat, M., Assran, M., Ballas, N., Synnaeve, G., Misra, I., Jegou, H., Mairal, J., Labatut, P., Joulin, A., Bojanowski, P.: Dinov2: Learning robust visual features without supervision (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.704, + 0.788, + 0.759 + ], + "angle": 0, + "content": "50. Paolicelli, V., Tavera, A., Masone, C., Berton, G., Caputo, B.: Learning semantics for visual place recognition through multi-scale attention. In: Image Analysis and Processing-ICIAP 2022: 21st International Conference, Lecce, Italy, May 23-27, 2022, Proceedings, Part II. pp. 454-466. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.759, + 0.788, + 0.799 + ], + "angle": 0, + "content": "51. Pion, N., Humenberger, M., Csurka, G., Cabon, Y., Sattler, T.: Benchmarking image retrieval for visual localization. In: 2020 International Conference on 3D Vision (3DV). pp. 483-494. IEEE (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.799, + 0.788, + 0.84 + ], + "angle": 0, + "content": "52. Puligilla, S.S., Tourani, S., Vaidya, T., Parihar, U.S., Sarvadevabhatla, R.K., Krishna, K.M.: Topological mapping for manhattan-like repetitive environments. In: 2020 IEEE International Conference on Robotics and Automation (ICRA). pp. 6268-6274. IEEE (2020)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.355, + 0.129 + ], + "angle": 0, + "content": "K. Garg et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.148, + 0.785, + 0.189 + ], + "angle": 0, + "content": "53. Radenović, F., Iscen, A., Tobias, G., Avrithis, Y., Chum, O.: Revisiting oxford and paris: Large-scale image retrieval benchmarking. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 5706-5715 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.19, + 0.785, + 0.217 + ], + "angle": 0, + "content": "54. Radenovic, F., Tobias, G., Chum, O.: Fine-tuning cnn image retrieval with no human annotation. IEEE Trans. Pattern Anal. Mach. Intell. 41(7), 1655-1668 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.218, + 0.785, + 0.259 + ], + "angle": 0, + "content": "55. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.26, + 0.785, + 0.3 + ], + "angle": 0, + "content": "56. Revaud, J., Almazán, J., Rezende, R.S., Souza, C.R.d.: Learning with average precision: Training image retrieval with a listwise loss. In: Proceedings of the IEEE International Conference on Computer Vision. pp. 5107-5116 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.301, + 0.785, + 0.327 + ], + "angle": 0, + "content": "57. Sarlin, P.E., Cadena, C., Siegwart, R., Dymczyk, M.: From coarse to fine: Robust hierarchical localization at large scale. arXiv preprint arXiv:1812.03506 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.328, + 0.785, + 0.37 + ], + "angle": 0, + "content": "58. Sattler, T., Maddern, W., Toft, C., Torii, A., Hammarstrand, L., Stenberg, E., Safari, D., Okutomi, M., M. P., J. S., F. K., T. P.: Benchmarking 6dof outdoor visual localization in changing conditions. In: Proc. CVPR (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.371, + 0.785, + 0.397 + ], + "angle": 0, + "content": "59. Schleiss, M., Rouatbi, F., Cremers, D.: Vpair - aerial visual place recognition and localization in large-scale outdoor environments (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.398, + 0.785, + 0.424 + ], + "angle": 0, + "content": "60. Schubert, S., Neubert, P., Garg, S., Milford, M., Fischer, T.: Visual place recognition: A tutorial. RAM (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.425, + 0.785, + 0.465 + ], + "angle": 0, + "content": "61. Shlapentokh-Rothman, M., Blume, A., Xiao, Y., Wu, Y., TV, S., Tao, H., Lee, J.Y., Torres, W., Wang, Y.X., Hoiem, D.: Region-based representations revisited. arXiv preprint arXiv:2402.02352 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.467, + 0.785, + 0.508 + ], + "angle": 0, + "content": "62. Shubodh, S., Omama, M., Zaidi, H., Parihar, U.S., Krishna, M.: Lip-loc: Lidar image pretraining for cross-modal localization. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 948-957 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.509, + 0.785, + 0.549 + ], + "angle": 0, + "content": "63. Sivic, J., Zisserman, A.: Video google: A text retrieval approach to object matching in videos. In: Proceedings of International Conference on Computer Vision (ICCV). p. 1470. IEEE (2003)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.55, + 0.785, + 0.591 + ], + "angle": 0, + "content": "64. Sun, X., Xie, Y., Luo, P., Wang, L.: A dataset for benchmarking image-based localization. 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR) pp. 5641-5649 (2017), https://apisemantic scholar.org/CorpusID:20531893" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.592, + 0.785, + 0.632 + ], + "angle": 0, + "content": "65. Sünderhauf, N., Shirazi, S., Jacobson, A., Dayoub, F., Pepperell, E., Upcroft, B., Milford, M.: Place recognition with convnet landmarks: Viewpoint-robust, condition-robust, training-free. Proceedings of Robotics: Science and Systems XII (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.633, + 0.785, + 0.674 + ], + "angle": 0, + "content": "66. Torii, A., Sivic, J., Pajdla, T., Okutomi, M.: Visual place recognition with repetitive structures. In: 2013 IEEE Conference on Computer Vision and Pattern Recognition. pp. 883-890 (2013). https://doi.org/10.1109/CVPR.2013.119" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.675, + 0.785, + 0.715 + ], + "angle": 0, + "content": "67. Tsintotas, K.A., Bampis, L., Gasteratos, A.: The revisiting problem in simultaneous localization and mapping: A survey on visual loop closure detection. IEEE Transactions on Intelligent Transportation Systems 23(11), 19929-19953 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.716, + 0.785, + 0.757 + ], + "angle": 0, + "content": "68. Wang, R., Shen, Y., Zuo, W., Zhou, S., Zheng, N.: Transvpr: Transformer-based place recognition with multi-level attention aggregation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13648-13657 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.758, + 0.785, + 0.799 + ], + "angle": 0, + "content": "69. Warburg, F., Hauberg, S., López-Antequera, M., Gargallo, P., Kuang, Y., Civera, J.: Mapillary street-level sequences: A dataset for lifelong place recognition. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2626-2635 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.799, + 0.785, + 0.84 + ], + "angle": 0, + "content": "70. Weinzaepfel, P., Csurka, G., Cabon, Y., Humenberger, M.: Visual localization by learning objects-of-interest dense match regression. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5634-5643 (2019)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.148, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.628, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Revisit Anything" + }, + { + "type": "page_number", + "bbox": [ + 0.77, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "19" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.189 + ], + "angle": 0, + "content": "71. Xin, Z., Cai, Y., Lu, T., Xing, X., Cai, S., Zhang, J., Yang, Y., Wang, Y.: Localizing discriminative visual landmarks for place recognition. In: 2019 International conference on robotics and automation (ICRA). pp. 5979-5985. IEEE (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.19, + 0.788, + 0.258 + ], + "angle": 0, + "content": "72. Yildiz, B., Khademi, S., Siebes, R.M., Van Gemert, J.: Amstertime: A visual place recognition benchmark dataset for severe domain shift. In: 2022 26th International Conference on Pattern Recognition (ICPR). IEEE (Aug 2022). https://doi.org/10.1109/icpr56361.2022.9956049, http://dx.doi.org/10.1109/ICPR56361.2022.9956049" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.259, + 0.788, + 0.3 + ], + "angle": 0, + "content": "73. Yu, J., Zhu, C., Zhang, J., Huang, Q., Tao, D.: Spatial pyramid-enhanced netvlad with weighted triplet loss for place recognition. IEEE Trans. Neural Netw. Learn. Syst. 31(2), 661-674 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.3, + 0.788, + 0.355 + ], + "angle": 0, + "content": "74. Zaffar, M., Garg, S., Milford, M., Kooij, J., Flynn, D., McDonald-Maier, K., Ehsan, S.: Vpr-bench: An open-source visual place recognition evaluation framework with quantifiable viewpoint and appearance change. International Journal of Computer Vision pp. 1-39 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.356, + 0.788, + 0.383 + ], + "angle": 0, + "content": "75. Zhang, Y., Zhao, X.: Mesa: Matching everything by segmenting anything. arXiv preprint arXiv:2401.16741 (2024)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.384, + 0.788, + 0.424 + ], + "angle": 0, + "content": "76. Zhou, B., Lapedriza, A., Khosla, A., Oliva, A., Torralba, A.: Places: A 10 million image database for scene recognition. IEEE Transactions on Pattern Analysis and Machine Intelligence (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.424, + 0.788, + 0.466 + ], + "angle": 0, + "content": "77. Zou, X., Yang, J., Zhang, H., Li, F., Li, L., Wang, J., Wang, L., Gao, J., Lee, Y.J.: Segment everything everywhere all at once. Advances in Neural Information Processing Systems 36 (2024)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.466 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/Revisit Anything_ Visual Place Recognition via Image Segment Retrieval/32b8503b-ae81-47c5-bdc0-0efb4201113d_origin.pdf b/2024/Revisit Anything_ Visual Place Recognition via Image Segment Retrieval/32b8503b-ae81-47c5-bdc0-0efb4201113d_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..cd4a4197fff99a860c75a3008ef8957048bd0e70 --- /dev/null +++ b/2024/Revisit Anything_ Visual Place Recognition via Image Segment Retrieval/32b8503b-ae81-47c5-bdc0-0efb4201113d_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7c207580572499cb69a4ef228f2117137981ecd6bec0e9fd2b888ac0d681dbe0 +size 1543790 diff --git a/2024/Revisit Anything_ Visual Place Recognition via Image Segment Retrieval/full.md b/2024/Revisit Anything_ Visual Place Recognition via Image Segment Retrieval/full.md new file mode 100644 index 0000000000000000000000000000000000000000..90d1070af5e5d039e207ea8939522bd85bcb9687 --- /dev/null +++ b/2024/Revisit Anything_ Visual Place Recognition via Image Segment Retrieval/full.md @@ -0,0 +1,282 @@ +# Revisit Anything: Visual Place Recognition via Image Segment Retrieval + +Kartik Garg $^{1}$ , Sai Shubodh Pulgilla $^{2}$ , Shishir Kolathaya $^{1}$ , Madhava Krishna $^{2}$ , and Sourav Garg $^{3}$ + +$^{1}$ Indian Institute of Science (IISc), Bengaluru, India + $^{2}$ International Institute of Information Technology, Hyderabad, India + $^{3}$ University of Adelaide, Australia + +kartikgarg@iisc.ac.in, p.saishubodh@gmail.com, shishirk@iisc.ac.in, mkrishna@iiit.ac.in, sourav.garg@adelaide.edu.au + +Abstract. Accurately recognizing a revisited place is crucial for embodied agents to localize and navigate. This requires visual representations to be distinct, despite strong variations in camera viewpoint and scene appearance. Existing visual place recognition pipelines encode the whole image and search for matches. This poses a fundamental challenge in matching two images of the same place captured from different camera viewpoints: the similarity of what overlaps can be dominated by the dissimilarity of what does not overlap. We address this by encoding and searching for image segments instead of the whole images. We propose to use open-set image segmentation to decompose an image into 'meaningful' entities (i.e., things and stuff). This enables us to create a novel image representation as a collection of multiple overlapping subgraphs connecting a segment with its neighboring segments, dubbed SuperSegment. Furthermore, to efficiently encode these SuperSegments into compact vector representations, we propose a novel factorized representation of feature aggregation. We show that retrieving these partial representations leads to significantly higher recognition recall than the typical whole image based retrieval. Our segments-based approach, dubbed SegVLAD, sets a new state-of-the-art in place recognition on a diverse selection of benchmark datasets, while being applicable to both generic and task-specialized image encoders. Finally, we demonstrate the potential of our method to "revisit anything" by evaluating our method on an object instance retrieval task, which bridges the two disparate areas of research: visual place recognition and object-goal navigation, through their common aim of recognizing goal objects specific to a place. Source code: https://github.com/AnyLoc/Revisit-Anything. + +Keywords: Visual Place Recognition $\cdot$ Image Segmentation $\cdot$ Robotics + +# 1 Introduction + +Visual Place Recognition (VPR) is an important capability for embodied agents to localize and navigate autonomously. A predominant solution for VPR is to encode + +an image into a global vector and retrieve similar vectors as coarse localization hypotheses [19, 44, 60, 67]. Thus, for almost a decade, researchers have focused on learning/finetuning image encoders so that global descriptors are induced with invariance to appearance [3,56,69], viewpoint [3,8], and clutter [27]. On the other hand, there is a vast literature on local descriptors (point/pixel-level), mainly relevant for geometric reranking in hierarchical VPR [11, 16, 25, 57, 68]. In the middle of local and global descriptors exists a variety of methods that use regions/patches [4, 25], lines/planes [17], objects (things/stuff) [21, 31, 45, 65], and segments [26, 33, 50] to represent images. However, these methods are still only aimed at either improving global descriptors based coarse retrieval or local feature matching based reranking. In this work, in contrast to conventional retrieval-based VPR, we explore an alternative: retrieval via encoding segments instead of the whole image. This is particularly enabled by recent advances in open-set image segmentation [37] which can meaningfully deconstruct a place into 'things' (and/or 'stuff') [10]. Thus, we reformulate the VPR problem of revisiting places as that to revisiting things by enabling recognition of these specific things within the context of their place. While such a segment-level place recognition approach provides a direct link to higher-level semantic tasks, such as object-goal navigation [12, 20, 24, 43], it also addresses a fundamental issue in matching partially-overlapping images from across significant viewpoint change. Segments-based partial image representation avoids the mismatches caused by the whole-image representation when the similarity of what overlaps is dominated by the dissimilarity of what does not overlap. Our novel segments-based VPR method, dubbed SegVLAD (Segment based Vector of Locally Aggregated Descriptors), is illustrated in Figure 1, which makes the following novel contributions: + +1. an image representation as a collection of multiple overlapping subgraphs of segments, dubbed SuperSegments, which enables accurate recognition across partially-overlapping images; +2. a factorized representation of feature aggregation to effectively accommodate both segment-level information as well as segment neighborhood information; and +3. a similarity-weighted ranking method to convert segment-level retrieval into image-level retrieval. + +Using a diverse set of data sources, we demonstrate that our proposed segments-based retrieval enables place recognition under wide viewpoint variations, where global descriptor based retrieval suffers. SegVLAD achieves a new state-of-the-art on multiple challenging datasets. We also introduce an evaluation of our method on an instance-level object retrieval task – a novel capability of our pipeline unlike conventional VPR methods. We conduct several ablations and parameter studies to justify the design choices and emphasize the effectiveness of our method as an open-set segments-based coarse retriever. + +# 2 Related Works + +Image retrieval-based Visual Place Recognition (VPR) is a well-established area of research in visual localization [19,44,60,67]. It is important both during mapping for loop + +closures [67] as well as for relocalization [51, 58]. The underlying task in both the scenarios remains the same: how to recognize a previously seen place. The state-of-the-art methods in VPR use a global descriptor-based approach which converts an image into a compact vector to enable fast retrieval [2, 3, 6, 29, 32, 56]. The top retrieved hypotheses are often then re-ranked through compute-intensive local feature matching using geometric information [11, 16, 25, 33, 57, 68]. In contrast to previous approaches, we aim to explore image segment level descriptors in this work. This representation falls between point-based local descriptors and the whole-image based global descriptors. Our approach can be considered as 'semi-global', with the proposed segment (and SuperSegment) based descriptors being a 'spatially-reduced' form of whole-image global representation. This is motivated by our hypothesis that to deal with viewpoint variations in VPR with partially-overlapping images, we need a way to partially represent and match them. + +# 2.1 Whole Image Encoders + +Earlier works in whole-image representation used methods like Gist [48], BoW (Bag of Word) [63], and VLAD (Vector of Locally Aggregated Descriptors) [30], often defined using hand-crafted features such as SIFT [41]. In recent years, deep learning based methods have demonstrated remarkable performance, with initial successful methods like NetVLAD [3] now rapidly outperformed by better alternatives such as CosPlace [6], MixVPR [2], EigenPlaces [8], TransVPR [68], and more recently AnyLoc [32], SALAD [29] and VLAD-BuFF [35]. All these learning-based methods improve different aspects of representation learning: training datasets [1, 6, 69], objective/loss functions [2, 8], aggregation methods [29, 54, 56, 68], and generalization [32]. Our approach complements these existing methods as we mainly focus on the use of segment-based information, where the segments can be described by any of the image encoders from the aforementioned techniques. In particular, we demonstrate that both - an off-the-shelf encoder, e.g., DINOv2-AnyLoc [32, 49] or that finetuned specifically for the VPR task, e.g., DINOv2-NetVLAD [35] - can be used in conjunction with our segment-based approach to further elevate place recognition capability. + +# 2.2 Region/Patch Based Methods + +There exist several methods that employ region or patch level information to enhance representational power [13, 14, 34, 40, 52, 71, 73]. However, most of these methods only use this additional information to generate a single (or concatenated) compact vector representation of an image. Other methods such as Patch-NetVLAD [25] create multiple features per image but their primary purpose is to perform local matching based reranking. In contrast to these methods, we aim to use multiple segment descriptors per image to directly retrieve from a database of segments, without using any geometric information or reranking. The motivation behind this stems from the very nature of hierarchical VPR pipelines: reranking recall is upper bounded by the coarse retriever. A better coarse retriever can improve reranking performance without needing to rerank from a longer list of top hypotheses. MultiVLAD [4] is similar to our method in the spirit of retrieving multiple features per query image. However, like aforementioned + +methods, MultiVLAD defines regions arbitrarily, whereas we use image segments obtained from Segment Anything Model (SAM) [37] which are semantically meaningful. + +# 2.3 Segments-Enhanced Methods + +There exist several methods that use semantic segmentation information to improve VPR, as also surveyed in [22]. These methods vary in terms of type of segmentation used and the specific ways in which it is integrated in the VPR pipeline, e.g., planes [17], objects [15,31,45], landmarks [65], outdoor semantics [21,23,46,47], utility/confusion based [33, 38], domain adaptation [26] and even learning to segment for VPR [50]. However, neither these methods aim to perform segment-level retrieval nor do they use open-set segmentation. We also review two concurrent works: MESA [75] and Region-Revisited [61]. Similar to our method, they both use SAM to segment images but for different specific tasks. MESA [75] proposes a graph-based local feature/area matching method to obtain point correspondences. Our method complements this, as we perform coarse retrieval for VPR, which could potentially use MESA for reranking. Regions Revisited [61] delves into the advantages of using SAM masks in conjunction with SLIC [36] to improve semantic segmentation, activity recognition and object category retrieval. In contrast, we aim to improve instance-level recognition by recognizing specific things belonging to specific places that a robot encounters during a revisit. Similar to our work, [20] creates an image sequence-based topological graph of segments where its segment neighbourhood aggregation is based on average pooling, similar to [61]. In Section 5.3, we show that such segment average pooling deteriorates recognition performance for the VPR task. + +# 2.4 Open-set VPR + +Researchers have recently started to shift their focus to open-set, generally-applicable techniques, including that for VPR [31, 32, 45, 62]. FM-Loc [45] uses GPT [9] to recognize object and place labels, whereas [31] uses CLIP [55] for open-set place recognition. LIP-Loc [62] proposes pretraining for cross-modal VPR, but is limited in its zero-shot capabilities. AnyLoc [32] proposes to use DINOv2 with domain-level vocabularies and hard-assignment based VLAD. It achieves state-of-the-art performance particularly on non-streetview datasets, where current VPR-trained methods tend to fail. In this work, we propose a generally-applicable approach which is built on top of models like SAM [37] and DINOv2 [49], and works with both VPR-agnostic [32] and VPR-specific [35] backbone models. We particularly aim for a new paradigm in retrieval based VPR, where we move away from the conventional whole image global descriptors to segments based descriptors and retrieval, which achieves a new state-of-the-art on diverse domains under wide viewpoint variations. + +# 3 Proposed Approach + +Despite recent advances in place recognition, viewpoint variations continue to be an open challenge for an embodied agent to recognize the same specific things in its environment. Current methods in visual place recognition tackle this problem by converting + +![](images/43e8379a038ad16d7f713e22f77c5f5c6bf8270f93a5ce55510a261c89221503.jpg) +Fig. 1: Overview of our segment-retrieval based VPR pipeline, dubbed SegVLAD: We use open-set segmentor (SAM) to extract segment masks, which are converted into SuperSegments using the neighbouring image segments. Using pixel-level DINOv2 descriptors with VLAD-based aggregation, we obtain SuperSegment descriptors, which are matched against a flat index of SuperSegment descriptors obtained from all the images of the entire reference database. + +an image as a whole into a global descriptor, which does not explicitly deal with the problem of partial visual overlap caused by viewpoint variations. We propose an alternative solution by representing images partially with the help of image segments. In the following subsections, we describe our representation and retrieval method, which deviates from the conventional VPR techniques but creates a new capability in terms of recognizing objects/things that constitute a place. + +# 3.1 Problem Formulation + +We represent an image as a set of segment descriptors instead of a single global descriptor. For an image $I$ , we obtain binary image segment masks $M \in \{0,1\}^{S \times N}$ and dense pixel-level descriptors $f_{p} \in \mathcal{R}^{D}$ , where $S$ represents the number of segments per image, $D$ is the descriptor dimension, and $p \in [1,N]$ represents spatial elements across the width $(W)$ and height $(H)$ of the image encoder's output, flattened into $N = W \times H$ for convenience. Figure 1 shows an illustration of our proposed pipeline, as explained in the following subsections. + +# 3.2 Super Segments + +Humans are remarkable at visual recognition, where existing studies suggest that we often leverage spatial associations among objects in an environment to represent it internally [5, 28]. This enables us to distinguish between two different scenes through the surrounding context of the objects of interest. In this work, we imbibe this context through the spatial neighbourhood of the image segments. For each image, we construct a graph of segments through their pixel centers using Delaunay Triangulation. + +![](images/dced319d20016bdc56ce97d6c7c1379b3df2ff98a3f48ddbd2f9a35a4e04a343.jpg) +Fig. 2: Neighborhood expansion (Eq. 1) of a window in the leftmost image to the whole building in the rightmost image, progressing from no neighborhood aggregation to a third-order aggregation. This neighborhood expansion is in stark contrast with a typical regular grid- or patch-based approach which may not capture semantically-meaningful SuperSegments. + +![](images/27792164f535fc6e55aef41b060198439100414bd496b51ad08fbb48df3dba01.jpg) +Fig. 3: Illustration of four SuperSegments obtained from the same image. All four of these spatially overlap with each other, which is different from coarse segmentation methods that do not typically allow overlap across segments. + +This provides us with a binary adjacency matrix $A \in \{0,1\}^{S \times S}$ to define the neighborhood for individual segments. We use this adjacency information to expand the context of individual segments to generate new SuperSegment masks $(\mathcal{M})$ as below: + +$$ +\mathcal {M} _ {S \times N} = ^ {1} \left(A _ {S \times S} ^ {o} \cdot M _ {S \times N}\right) \tag {1} +$$ + +where $o \geq 0$ refers to the order for expanding the neighborhood by multiplying the adjacency matrix by itself as $A^{o + 1} = A^o \cdot A$ . This is matrix-multiplied with the original segmentation masks $M$ to expand the neighborhood at pixel level. $\mathcal{M}$ is obtained after element-wise binarization (denoted with $\mathbb{1}()$ ) so that all pixels in the SuperSegment mask may only contribute once to the subsequent feature aggregation. In Figure 2, we illustrate the extent of image area covered with different orders of mask expansion. Unlike, a patch or regular grid-based approach, the expanded mask of the window in the leftmost image covers a meaningful entity (building) in the rightmost image. Our approach to creating SuperSegments differs from coarse segmentation methods or superpixels in terms of the 'self-overlap'. By expanding neighborhood of each individual segment, we obtain several partially overlapping SuperSegments. A coarse segmentor will need to make assumptions about the right sub-segments to be coalesced so that it can enable accurate recognition from a different viewpoint, which could otherwise lead to the same limitation as that of the whole-image descriptors. Figure 3 presents examples of multiple overlapping SuperSegments from the same image. + +# 3.3 SuperSegment Descriptors + +In this section, we describe our feature aggregation method to obtain SuperSegment descriptors. Recent state-of-the-art method AnyLoc [32] demonstrated that using off-the-shelf powerful image encoders such as DINOv2 with hard assignment based VLAD + +aggregation achieves superior recognition performance. However, AnyLoc does not use segmentation information and only operates at the whole-image level. More recently, [61] showed that average pooling works well for segment-level descriptors, but it didn't consider segment neighborhood information. In this work, we propose a unified formulation for feature aggregation that can easily switch across segments, segment neighborhood and the whole image as well as different aggregation types (see supplementary for details). This simply extends Equation 1 as below: + +$$ +F _ {S \times D} = ^ {1} \left(A _ {S \times S} ^ {o} \cdot M _ {S \times N}\right) \cdot T _ {N \times D} \tag {2} +$$ + +where $T$ represents the features to be aggregated. By replacing $A$ and $M$ with ones matrices, one can obtain the whole-image global descriptor. For methods like Global Average Pooling (GAP), $T$ can be directly used as the output of the image encoder. In our work, we use Hard-VLAD, for which $T$ is the residual feature matrix per cluster and is obtained as below with respect to each of the cluster centers $c_k$ : + +$$ +T _ {N _ {k} \times D} ^ {k} = \left\{\alpha_ {k} \left(f _ {p}\right) \left(f _ {p} - c _ {k}\right) \mid \alpha_ {k} \left(f _ {p}\right) = 1 \right\} \tag {3} +$$ + +where $\alpha_{k}(f_{p})\in \{0,1\}$ is 1 if $f_{p}$ belongs to $c_{k}$ . The cluster centers (vocabulary) can be constructed using the map or the domain [32]. The SuperSegment VLAD descriptors obtained from Eq. 2 are intra-normalized, concatenated across clusters and then finally l2-normalized, following existing works [3, 32]. + +# 3.4 Image Retrieval via Segments + +Existing global descriptor based VPR techniques produce a single vector per image to search against a database of reference image vectors. In our method, we obtain multiple SuperSegment descriptors per image. We perform retrieval at segment-level, that is, we search for the top matches for each query segment against a flat index of all segments from all the images of the reference database/map. To evaluate in the form of image retrieval-based VPR, we convert the top retrieved segment indices across all segments of a query image into top reference image indices. This is achieved through a weighted frequency measure (i.e., weighted bin/word counting). We first map the top $K'$ retrieved segment indices for each of the query segments $s \in [1, S]$ to their respective reference image indices, denoted with $r$ . Then, for each of the unique retrieved image indices $r_j$ , we accumulate its segment similarity $\theta$ and then use the cumulative similarity score $\hat{\theta}$ to rank the image indices to obtain the top image match $r_j^*$ : + +$$ +r _ {j} ^ {*} = \underset {r _ {j}} {\operatorname {a r g m a x}} \hat {\theta} \left(r _ {j}\right); \quad \hat {\theta} \left(r _ {j}\right) = \sum_ {s = 1} ^ {S} \sum_ {k = 1} ^ {K ^ {\prime}} \theta_ {s k} \cdot \mathbb {1} _ {\left\{r _ {s k} = r _ {j} \right\}} \tag {4} +$$ + +In Section 5.3, we compare our similarity-weighted ranking with other alternatives based on frequency or similarity alone. + +# 4 Experimental Setup + +Datasets: VPR datasets are in abundance, as can be found in several benchmarks including VPR-Bench [74], Deep Visual GeoLocalization Benchmark [7], and AnyLoc [32]. In this work, we used a variety of datasets covering both outdoor and indoor environments. Outdoor datasets include Pitts30k [66], AmsterTime [72], Mapillary Street Level Sequences (MSLS) [69], SF-XL [6], VPAir [59], Revisted Oxford5K and Revisited Paris6k [53]. Indoor datasets include Baidu Mall [64], 17Places [76] and InsideOut [27]. Additional datasets-related details are provided in the supplementary. + +Evaluation and Benchmarking: We evaluate our method as an image retrieval task using Recall@K metric, where top K' (= 50) retrieved segments per query segment are used to obtain top K (= 5) images (see Eq. 4). We compare against the most recent and high-performing VPR baseline methods. This includes CosPlace [6], MixVPR [2] and EigenPlaces [8], which are trained on large-scale urban datasets for VPR tasks. We further include two very recent state-of-the-art methods that use DINOv2 as the backbone. These include AnyLoc [32] which uses an off-the-shelf DINOv2 model and SALAD [29] which uses a finetuned DINOv2 backbone. Given the dichotomy between general-purpose VPR benchmarking of AnyLoc and the typical outdoor-focused benchmarking in SALAD and other methods, we evaluated our method using two different backbones. a) SegVLAD-PreT: we use the same backbone and aggregation as AnyLoc, i.e., off-the-shelf pretrained DINOv2 (ViT-G) backbone with hard VLAD assignment, but the key difference is in the use of SuperSegments for our method as opposed to whole-image description of AnyLoc. b) SegVLAD-FineT: since our default aggregation method is VLAD, we use a finetuned DINOv2 (ViT-B) backbone which is similar to SALAD but replaces its aggregation layer with the original NetVLAD aggregation [3] using 64 clusters, as described in [35]. We use this finetuned backbone with hard VLAD based assignment, similar to AnyLoc. For both these models, we reduce the descriptor dimensions of the VLAD descriptor to 1024 using PCA, as commonly done in previous works [3, 32]. We train PCA transform in a map-specific manner using the database images of the dataset. + +# 5 Results + +We first present benchmark comparison of our method against state-of-the-art VPR methods. This is followed by detailed analysis of our proposed aggregation technique. Lastly, we demonstrate results on a downstream task of Object of Interest (OOI) retrieval, showcasing the versatility of our method. + +# 5.1 State-of-the-art comparisons + +Table 1 presents Recall@1/5 comparison against state-of-the-art VPR methods on standard outdoor street-view datasets, which are similar to the typical training datasets used for VPR [2,3,6]. Table 2 covers 'out-of-distribution' datasets, inspired by AnyLoc [32], covering indoor environments (Baidu Mall and 17 Places), aerial imagery (VPAir), + +Table 1: Recall@ 1/5 benchmark comparison on outdoor street-view datasets. + +
MethodPitts-30KMSLS SFMSLS CPHSF-XL ValRO5k MedRO5k HardRP6k MedRP6k Hard
CosPlace90.4/95.793.4/97.584.9/92.094.6/97.685.7/87.127.1/45.794.3/95.77.1/15.7
MixVPR91.5/95.591.3/95.987.1/92.487.8/93.868.6/80.032.9/54.394.3/10010.0/32.9
EigenPlaces92.6/96.792.6/97.187.1/92.896.4/98.285.7/88.642.8/57.195.7/98.64.3/11.4
AnyLoc87.7/94.783.4/94.679.9/89.184.4/91.988.6/92.940.0/58.697.1/10011.4/44.3
SALAD92.6/96.591.7/97.192.3/96.193.6/97.382.9/90.037.1/54.395.7/98.614.3/58.6
SegVLAD-PreT86.7/94.288.4/94.281.7/90.790.9/96.491.4/95.760.0/81.494.3/1008.6/48.6
SegVLAD-FineT93.2/96.894.6/97.190.9/95.794.9/98.187.1/95.751.4/70.095.7/10010.0/48.6
+ +Table 2: Recall@ 1/5 benchmark comparison on 'out-of-distribution' datasets. + +
MethodBaiduAmstTimeInsideOut17PlacesVPAir
CosPlace41.6/55.047.7/69.80.2/2.081.3/88.24.6/13.7
MixVPR64.4/80.340.2/59.10.0/1.885.2/90.16.8/16.1
EigenPlaces56.5/72.848.9/69.50.4/1.483.0/90.16.5/17.9
AnyLoc75.2/87.650.3/73.02.4/8.095.3/97.366.7/79.2
SALAD74.8/86.555.4/75.60.6/1.882.5/88.225.8/38.7
SegVLAD-PreT78.5/93.856.8/77.74.2/9.495.3/98.069.8/83.7
SegVLAD-FineT68.1/89.058.9/79.37.4/15.695.1/97.535.4/55.3
+ +indoor-to-outdoor viewing (InsideOut), and historical image matching (AmsterTime). Below, we discuss two key aspects of this comparative analysis: $i$ ) how our segment-based approach compares against whole-image global descriptor based methods and ii) how performance trends vary depending on the choices of feature backbone with regards to task-specific (VPR) training. + +Aggregating Segments vs Whole Images Table 1 and Table 2 show that our proposed method SegVLAD achieves a new state-of-the-art on the majority of datasets, considering both the backbone variants: PreT and FineT. AnyLoc and SALAD respectively differ from SegVLAD-PreT and SegVLAD-FineT in terms of the aggregation scope (global vs segments). Thus, the superior performance of SegVLAD clearly highlights the role of segments based retrieval over whole-image based approach. On the Baidu Mall dataset – highly-aliased indoor environment – our method improves over AnyLoc by $3\%$ for R@1 and $6\%$ for R@5 in absolute gains. On the InsideOut dataset – matching outdoor images viewed from within indoors – our method leads to a ‘meaningful’ recall, unlike all other baselines. Overall, these results highlight that even with the use of powerful image encoders (DINOv2), global aggregation struggles to deal with the challenges of matching images across major viewpoint shifts – it is thus the partial image representation and matching (using the same encoder) which is needed to obtain superior recognition performance. + +VPR Fine-tuned Encoders + Segments For SegVLAD-FineT, we used a DINOv2 backbone finetuned for the purpose of VPR, mainly to observe the benefit of segments over global descriptor based approach in a task-specific manner. Table 1 and Table 2 + +Table 3: Recall@1 results for various approaches on Object-Instance Retrieval Task + +
MethodSegVLAD NoNbrAggSegVLADSegment-to-GlobalGlobal-to-Global
R@164.192.730.086.4
+ +show that, on the outdoor street-view datasets, SALAD (finetuned DINOv2) generally performs better than AnyLoc (its pretrained counterpart), whereas the latter generally outperforms the former on 'out-of-distribution' datasets. It can be clearly observed that these performance patterns translate well from global- to segment-level results. + +# 5.2 Revisiting Objects of Interest (OOI): Object Instance Retrieval + +A typical requirement of an embodied agent is to understand the context of its task through its memory/map information, which is composed of visual and/or semantic cues. For example, navigating to a given object goal requires a robot to visually recognize the goal and not be confused by perceptually-similar items. In this section, we demonstrate our method's ability to retrieve the correct image given just an Object Of Interest (OOI) as a query segment. For this purpose, we use an extended version of the Baidu dataset [70] which annotates OOI as various discriminative areas that can be reliably detected under variable viewpoint and lighting conditions. In total, there are 220 OOI, which cover various things such as logos, brand names, posters, etc., in a highly cluttered mall environment. To cast this dataset in terms of revisiting things, we use the original query images of the Baidu Mall [64] dataset as the database and the images with OOI as the queries. This allows us to evaluate the OOIs directly. This is similar to VPR evaluation of recall in terms of image retrieval but with querying of a specific segment instead of using all the segments of the query image. + +We consider four different methods of recognizing known objects in this study. i) Global-to-Global: as a baseline method, we use whole images to represent and retrieve, i.e., without using the OOI mask; this resembles object-goal recognition problem for an InstanceImageNavigation task [39]. ii) Segment-to-Global: this is the same as the previous setting except that the query image descriptor is aggregated only using the OOI mask; this tests the ability of the image encoder/aggregator to match segment-level descriptor against global descriptors. iii) SegVLAD and iv) SegVLAD NoNbrAgg, which are our proposed methods but the latter does not use any neighborhood information; this highlights the relevance of spatial context around the OOI for recognition. For SegVLAD, we create a virtual segment mask for the OOI, append it to the other masks of the image, and then perform our neighborhood expansion and feature aggregation, as described in Section 3. + +Table 3 reports Recall@1 for different recognition methods. It can be observed that SegVLAD outperforms Global-to-Global matching by a large margin, which shows that recognizing specific object instances through their images (as in InstanceImageNav) is more prone to failures. It can further be observed from low recall of SegVLAD NoNbrAgg that neighborhood aggregation around segments is crucial to capture the required context. Finally, poor recall of Segment-to-Global highlights + +Table 4: Recall@1/5 for Baidu mall dataset Table 5: Recall@1/5 comparison between for different aggregation methods and dif- different methods for ranking images based. ferent orders of neighborhood expansion. on segment-level retrieval. + +
OrderSegVLADSAP
073.1/89.974.6/91.1
177.4/91.765.6/87.2
276.3/92.453.2/81.3
377.7/92.649.8/78.0
Max Seg78.5/93.953.9/70.4
Max Sim65.2/92.734.4/62.4
Ours78.5/93.854.4/76.3
+ +that matching a part of an image (OOI) with the whole image is not a viable solution for object instance recognition. + +# 5.3 Ablation Studies + +Aggregation method & Order of Neighborhood Expansion Previous studies on VPR such as AnyLoc [32] have shown VLAD to be better than other aggregation methods for whole-image based global descriptors. However, in an increasing number of segment-based approaches [12, 20, 24, 61], segment average pooling (SAP) is used more commonly. Thus, we compare hard-assignment VLAD against SAP on Baidu dataset. For SAP, we upsample the DINOv2 features to match the resolution of our SAM masks – this upsampling is shown to enhance performance in [61]. For VLAD aggregation, we use our proposed method SegVLAD, where we downsample masks to match with the low resolution of DINOv2. Table 4 shows that SAP performs well for order 0 aggregation (i.e., no neighborhood aggregation) but its performance reduces as the neighborhood expands. On the other hand, SegVLAD has low recall when no neighborhood is considered but benefits significantly even with its immediate neighborhood (order 1). We attribute these inverted trends of SegVLAD and SAP to the very nature of these aggregation methods: as more information becomes available SAP smoothes out the overall information content whereas SegVLAD benefits from additional information which gets distributed across its clusters, thus minimizing any possible smoothing effect. It can be observed that R@5 increases for SegVLAD with an increasing order of neighborhood expansion but margins diminish for higher orders. Overall, SegVLAD (order 3) achieves the best results, despite aggregating at a $14 \times$ lower resolution than SAP's upsampling based aggregation. + +Segment to Image Retrieval Unlike conventional global descriptor retrieval based VPR, we perform retrieval for multiple SuperSegments of the query image. To obtain retrieval output in terms of images (as that is what VPR is typically evaluated on), there exist multiple ways to combine the top segment-level matches across all the query segments. We consider the following alternative options. i) MaxSeg: we obtain the best matching segment for each query segment, associate the matched segments to their respective reference image indices, and then rank these image indices based on their + +![](images/a580aa5b9234506efb32d3196b7efd79d541a96dc84361086e7e6e6579d67123.jpg) +Fig. 4: Qualitative results: Columns respectively represent the query image, correct match of SegVLAD and incorrect match of AnyLoc. Examples from different datasets: AmsterTime, Baidu Mall, Pitts-30K are presented across the rows. + +frequency; this method weakly resembles an inverted index list based counting of common segments between the query and the reference. ii) MaxSim: across the best matching reference segments, we order their image indices based on the segment similarity; this method is similar to that used in MultiVLAD [4]. iii) Similarity-Weighted Frequency: this is our proposed method as defined in Section 3.4. Table 5 shows that our proposed method for combining segment-level hypotheses consistently achieves superior results for both the datasets. While MaxSeg achieves a similar performance on Baidu, it suffers a drop in recall for AmsterTime. Both the methods outperform MaxSim at R@1 by a large margin. + +Patches vs Segment While open-set segmentation [37] is aimed at a meaningful segregation of visual entities, a simple alternative to our segment-based retrieval is to use uniformly defined regions/patches. Table 6 compares SegVLAD with a patch-based approach, where we consider arbitrary square patch sizes to segment an image. It can be observed that SegVLAD outperforms its patch-based counterparts, where smaller patches perform the worst, and for larger patches, R@1 saturates while R@5 reduces. These results are also in line with similar findings of a recent work [61]. + +![](images/f4b8ec815e3f5b6f1e792ebaf19ae217e83292ed60fe90566e92a0bf58057df9.jpg) +Fig.5: Recall vs storage/retrieval time on AmsterTime. + +Table 6: Patch vs Segments on AmsterTime (Reso. $256\times 256$ + +
Patch SizeNumSegDbNumSegImR@1/5
16×1631513625635.7/62.1
32×32787846445.9/72.5
64×64196961652.0/75.1
128×1284924453.5/65.6
SegVLAD12963710556.8/77.7
+ +# 5.4 Limitations, Storage, Compute Time & IOU Based Filtering + +A key limitation of our method is its large map size, i.e., large storage requirement for segment-level descriptors (see supplementary for further details). In this section, we analyze the resource requirements in terms of database (index) storage and query retrieval time for our method, along with a preliminary study on IOU (Intersection over Union) based filtering of SuperSegments to reduce such costs. We compute IOU between all pairs of SuperSegments in a given image, and remove segments with $IOU(s_i, s_j) > \psi \forall i \in [1, S]$ , $j \in [i, S]$ , where $\psi \in [0, 1]$ is a threshold on IOU and $s_i$ refers to the list of SuperSegments sorted by their pixel area in a descending order. We only perform this culling on the database segments. We use the outdoor-finetuned DINOv2 backbone for this purpose and compare SegVLAD with SALAD on AmsterTime (see supplementary for additional results on Pitts30K). Figure 5 shows that SegVLAD outperforms SALAD while requiring less storage (annotated on points) and comparable retrieval time (excludes extraction time), using IoU-based filtering threshold ranging from 0.2 to 0.8 (left to right) with a step size of 0.2. In particular, at 0.4 IOU threshold, only 20% of SuperSegments are retained (0.05 GB) while still outperforming the baseline. + +# 5.5 Qualitative Analyses + +In this section, we further demonstrate the capabilities of our method through qualitative visualizations. We compare our method against AnyLoc, where the only difference between the two methods is Global aggregation/retrieval vs SuperSegment aggregation/retrieval. We particularly consider the queries for which our approach successfully retrieved the correct match but AnyLoc failed to do so (additional examples can be found in the supplementary). Figure 4 shows triplets of images in the order of query, correct match (ours), and incorrect match (Anyloc). The segmented part shows one of the correctly matched SuperSegments, displayed as a subgraph in white color overlaid on the corresponding segment masks. + +The first row shows a triplet from AmsterTime where our proposed method is able to correctly recognize a subgraph of building across the image pair, whereas Anyloc retrieves an incorrect image of a street-view with buildings, cars, and road, laid out similarly across the image pair. This highlights that a global descriptor can get confused with perceptually-aliased global context. In contrast, our SuperSegment based + +SegVLAD is not only able to retrieve the correct image but it also correctly finds the mutually-overlapping area. A similar trend follows for the Baidu Mall (middle row) and Pitts30K (last row). In the Baidu Mall example, our approach identifies the piano and the region around it in the query image. It correctly retrieves an image having similar spatial context with the piano. This is akin to how humans use spatial context to identify places. AnyLoc, on the other hand, retrieves an image with similar floor tiles and railings. This example particularly highlights our hypothesis that dissimilarity of nonoverlapping regions can dominate the similarity of overlapping regions in global whole-image descriptors. Finally, the Pitts30K example (last row) shows a case of strong viewpoint change. While SegVLAD correctly matches the traffic signal and signboards to retrieve the correct match, AnyLoc retrieves a similar looking image while missing the finer details. This example particularly reinforces the idea of 'revisiting things', as even though the background mountain is common across the triplet, it is the context of the things near the camera/robot which helps in uniquely recognizing a place. + +# 6 Conclusion and Future Work + +In this paper, we presented a novel visual place recognition pipeline SegVLAD based on image segments-based description and retrieval, which is akin to 'revisting things' as a means to recognize specific instances of what constitute a place. Our proposed SuperSegments based image representation and a novel factorization based feature aggregation enables us to effectively represent and retrieve images using our segment similarity-weighted image ranking. Our results show that despite using powerful image encoders such as DINOv2 (pretrained or VPR-finetuned), existing global descriptor based techniques are unable to deal with the challenges of viewpoint variations. In contrast, SegVLAD is able to correctly retrieve images through its ability to match partially-overlapping images with its partial image representation in the form of semiglobal subgraphs of segments, i.e., SuperSegments. Thus, our method achieves state-of-the-art results on three diverse datasets (indoor and outdoor) that exhibit strong viewpoint variations on top of other challenges of appearance shift and high perceptual aliasing. Through an additional object instance retrieval study, we demonstrate the unique ability of our method to recognize objects instances within their specific place contexts - an open-set recognition capability that existing VPR methods lack. + +Our approach shifts the paradigm in retrieval based VPR research, as the conventional methods predominantly classify into either whole-image global descriptor based coarse retrieval or local feature based geometric reranking. Our approach complements recent concurrent works like MESA [75]; future work can explore a hierarchical VPR pipeline that closely integrates a segment-based coarse retriever with segments-based reranker such as MESA, thus doing away entirely with global whole-image descriptors. Furthermore, segments-based representation with implicitly baked semantics provide a natural way for creating text-based interfaces through CLIP [55] and LLMs (Large Language Models) [9], which can be easily integrated with recent efforts in this direction [12, 18, 20, 24, 42, 43, 77]. + +# Acknowledgements + +This work was supported by the Centre for Augmented Reasoning (CAR) at the Australian Institute for Machine Learning (AIML), University of Adelaide, Australia. The authors acknowledge the computational support provided by the Indian Institute of Science (IISc), Bengaluru, India, and the International Institute of Information Technology, Hyderabad (IIITH), India. The authors thank Ahmad Khaliq for technical support, Sarah Ibrahimi for sharing the InsideOut dataset, and Martin Humenberger for sharing the OOI annotations for the Baidu Mall dataset. + +# References + +1. Ali-bey, A., Chaib-draa, B., Giguère, P.: Gsv-cities: Toward appropriate supervised visual place recognition. Neurocomputing 513, 194-203 (2022) +2. Ali-bey, A., Chaib-draa, B., Giguère, P.: Mixvpr: Feature mixing for visual place recognition. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 2998-3007 (2023) +3. Arandjelovic, R., Gronat, P., Torii, A., Pajdla, T., Sivic, J.: Netvlad: Cnn architecture for weakly supervised place recognition. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 5297-5307 (2016) +4. Arandjelovic, R., Zisserman, A.: All about vlad. In: Proceedings of the IEEE conference on Computer Vision and Pattern Recognition. pp. 1578-1585 (2013) +5. Bar, M.: Visual objects in context. Nature Reviews Neuroscience 5(8), 617-629 (2004) +6. Berton, G., Masone, C., Caputo, B.: Rethinking visual geo-localization for large-scale applications. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4878-4888 (2022) +7. Berton, G., Mereu, R., Trivigno, G., Masone, C., Csurka, G., Sattler, T., Caputo, B.: Deep visual geo-localization benchmark. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5396-5407 (2022) +8. Berton, G., Trivigno, G., Caputo, B., Masone, C.: Eigenplaces: Training viewpoint robust models for visual place recognition. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 11080-11090 (October 2023) +9. Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J.D., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., et al.: Language models are few-shot learners. Advances in neural information processing systems 33, 1877-1901 (2020) +0. Caesar, H., Uijlings, J., Ferrari, V.: Coco-stuff: Thing and stuff classes in context. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1209-1218 (2018) +1. Cao, B., Araujo, A., Sim, J.: Unifying deep local and global features for image search. In: European Conference on Computer Vision. pp. 726-743. Springer (2020) +2. Chang, M., Gervet, T., Khanna, M., Yenamandra, S., Shah, D., Min, S.Y., Shah, K., Paxton, C., Gupta, S., Batra, D., et al.: Goat: Go to any thing. arXiv preprint arXiv:2311.06430 (2023) +3. Chen, L.C., Yang, Y., Wang, J., Xu, W., Yuille, A.L.: Attention to scale: Scale-aware semantic image segmentation. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 3640-3649 (2016) +4. Chen, Z., Maffra, F., Sa, I., Chli, M.: Only look once, mining distinctive landmarks from convnet for visual place recognition. In: Intelligent Robots and Systems (IROS), 2017 IEEE/RSJ International Conference on. pp. 9-16. IEEE (2017) + +15. Cheng, C., Page, D.L., Abidi, M.A.: Object-based place recognition and loop closing with jigsaw puzzle image segmentation algorithm. In: 2008 IEEE International Conference on Robotics and Automation. pp. 557-562. IEEE (2008) +16. Cummins, M., Newman, P.: Appearance-only slam at large scale with fab-map 2.0. The International Journal of Robotics Research 30(9), 1100-1123 (2011) +17. Cupec, R., Nyarko, E.K., Filko, D., Kitanov, A., Petrovic, I.: Place recognition based on matching of planar surfaces and line segments. The International Journal of Robotics Research 34(4-5), 674-704 (2015) +18. Folorunsho, S.O.: Semantic segmentation-based approach for autonomous navigation in challenging farm terrains. algorithms 15, 3 (2024) +19. Garg, S., Fischer, T., Milford, M.: Where is your place, visual place recognition? In: IJCAI (2021) +20. Garg, S., Rana, K., Hosseinzadeh, M., Mares, L., Suenderhauf, N., Dayoub, F., Reid, I.: Robohop: Segment-based topological map representation for open-world visual navigation. In: 2024 IEEE International Conference on Robotics and Automation (ICRA) (2024) +21. Garg, S., Suenderhauf, N., Milford, M.: Lost? appearance-invariant place recognition for opposite viewpoints using visual semantics. In: Proceedings of Robotics: Science and Systems XIV (2018) +22. Garg, S., Sunderhauf, N., Dayoub, F., Morrison, D., Cosgun, A., Carneiro, G., Wu, Q., Chin, T.J., Reid, I., Gould, S., Corke, P., Milford, M.: Semantics for robotic mapping, perception and interaction: A survey. Foundations and Trends® in Robotics 8(1-2), 1-224 (2020). https://doi.org/10.1561/2300000059, http://dx.doi.org/10.1561/2300000059 +23. Gawel, A., Del Don, C., Siegwart, R., Nieto, J., Cadena, C.: X-view: Graph-based semantic multi-view localization. IEEE Robotics and Automation Letters 3(3), 1687-1694 (2018) +24. Gu, Q., Kuwajerwala, A., Morin, S., Jatavallabhula, K.M., Sen, B., Agarwal, A., Rivera, C., Paul, W., Ellis, K., Chellappa, R., et al.: Conceptgraphs: Open-vocabulary 3d scene graphs for perception and planning. arXiv preprint arXiv:2309.16650 (2023) +25. Hausler, S., Garg, S., Xu, M., Milford, M., Fischer, T.: Patch-netvlad: Multi-scale fusion of locally-global descriptors for place recognition. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14141-14152 (2021) +26. Hu, H., Qiao, Z., Cheng, M., Liu, Z., Wang, H.: Dasgil: Domain adaptation for semantic and geometric-aware image-based localization. IEEE Transactions on Image Processing 30, 1342-1353 (2020) +27. Ibrahimi, S., Van Noord, N., Alpherts, T., Worring, M.: Inside out visual place recognition. arXiv preprint arXiv:2111.13546 (2021) +28. Intraub, H.: The representation of visual scenes. Trends in cognitive sciences 1(6), 217-222 (1997) +29. Izquierdo, S., Civera, J.: Optimal transport aggregation for visual place recognition (2023) +30. Jégou, H., Douze, M., Schmid, C., Pérez, P.: Aggregating local descriptors into a compact image representation. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 3304-3311 (2010) +31. Kassab, C., Mattamala, M., Fallon, M.: Clip-based features achieve competitive zero-shot visual localization. OpenReview preprint arXiv:2306.14846 (2023) +32. Keetha, N., Mishra, A., Karhade, J., Jatavallabhula, K.M., Scherer, S., Krishna, M., Garg, S.: Anyloc: Towards universal visual place recognition. IEEE Robotics and Automation Letters (2023) +33. Keetha, N.V., Milford, M., Garg, S.: A hierarchical dual model of environment-and place-specific utility for visual place recognition. IEEE Robotics and Automation Letters 6(4), 6969-6976 (2021) + +34. Khaliq, A., Ehsan, S., Milford, M., McDonald-Maier, K.: A holistic visual place recognition approach using lightweight cnns for severe viewpoint and appearance changes. arXiv preprint arXiv:1811.03032 (2018) +35. Khaliq, A., Xu, M., Hausler, S., Milford, M., Garg, S.: Vlad-buff: Burst-aware fast feature aggregation for visual place recognition. In: European Conference on Computer Vision. Springer (2024) +36. Khorasgani, S.H., Chen, Y., Shkurti, F.: Slic: Self-supervised learning with iterative clustering for human action videos. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16091-16101 (2022) +37. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., et al.: Segment anything. arXiv preprint arXiv:2304.02643 (2023) +38. Knopp, J., Sivic, J., Pajdla, T.: Avoiding confusing features in place recognition. In: European Conference on Computer Vision. pp. 748-761. Springer (2010) +39. Krantz, J., Lee, S., Malik, J., Batra, D., Chaplot, D.S.: Instance-specific image goal navigation: Training embodied agents to find object instances. arXiv preprint arXiv:2211.15876 (2022) +40. Le, D.C., Youn, C.H.: City-scale visual place recognition with deep local features based on multi-scale ordered VLAD pooling. arXiv preprint arXiv:2009.09255 (2020) +41. Lowe, D.G.: Distinctive image features from scale-invariant keypoints. International journal of computer vision 60(2), 91-110 (2004) +42. Maalouf, A., Jadhav, N., Jatavallabhula, K.M., Chahine, M., Vogt, D.M., Wood, R.J., Torralba, A., Rus, D.: Follow anything: Open-set detection, tracking, and following in real-time. IEEE Robotics and Automation Letters 9(4), 3283-3290 (2024) +43. Manglani, S.: Real-time vision-based navigation for a robot in an indoor environment. arXiv preprint arXiv:2307.00666 (2023) +44. Masone, C., Caputo, B.: A survey on deep visual place recognition. IEEE Access 9, 19516-19547 (2021) +45. Mirjalili, R., Krawez, M., Burgard, W.: Fm-loc: Using foundation models for improved vision-based localization. In: Robotics and Automation (ICRA), 2023 IEEE International Conference on. IEEE (2023) +46. Mousavian, A., Kosecka, J.: Semantic image based geolocation given a map (author's initial manuscript). Tech. rep., George Mason University Fairfax United States (2016) +47. Naseer, T., Oliveira, G.L., Brox, T., Burgard, W.: Semantics-aware visual localization under challenging perceptual conditions. In: IEEE International Conference on Robotics and Automation (ICRA) (2017) +48. Oliva, A.: Gist of the scene. Neurobiology of attention 696(64), 251-258 (2005) +49. Oquab, M., Darcet, T., Moutakanni, T., Vo, H.V., Szafraniec, M., Khalidov, V., Fernandez, P., Haziza, D., Massa, F., El-Nouby, A., Howes, R., Huang, P.Y., Xu, H., Sharma, V., Li, S.W., Galuba, W., Rabbat, M., Assran, M., Ballas, N., Synnaeve, G., Misra, I., Jegou, H., Mairal, J., Labatut, P., Joulin, A., Bojanowski, P.: Dinov2: Learning robust visual features without supervision (2023) +50. Paolicelli, V., Tavera, A., Masone, C., Berton, G., Caputo, B.: Learning semantics for visual place recognition through multi-scale attention. In: Image Analysis and Processing-ICIAP 2022: 21st International Conference, Lecce, Italy, May 23-27, 2022, Proceedings, Part II. pp. 454-466. Springer (2022) +51. Pion, N., Humenberger, M., Csurka, G., Cabon, Y., Sattler, T.: Benchmarking image retrieval for visual localization. In: 2020 International Conference on 3D Vision (3DV). pp. 483-494. IEEE (2020) +52. Puligilla, S.S., Tourani, S., Vaidya, T., Parihar, U.S., Sarvadevabhatla, R.K., Krishna, K.M.: Topological mapping for manhattan-like repetitive environments. In: 2020 IEEE International Conference on Robotics and Automation (ICRA). pp. 6268-6274. IEEE (2020) + +53. Radenović, F., Iscen, A., Tobias, G., Avrithis, Y., Chum, O.: Revisiting oxford and paris: Large-scale image retrieval benchmarking. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 5706-5715 (2018) +54. Radenovic, F., Tobias, G., Chum, O.: Fine-tuning cnn image retrieval with no human annotation. IEEE Trans. Pattern Anal. Mach. Intell. 41(7), 1655-1668 (2018) +55. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021) +56. Revaud, J., Almazán, J., Rezende, R.S., Souza, C.R.d.: Learning with average precision: Training image retrieval with a listwise loss. In: Proceedings of the IEEE International Conference on Computer Vision. pp. 5107-5116 (2019) +57. Sarlin, P.E., Cadena, C., Siegwart, R., Dymczyk, M.: From coarse to fine: Robust hierarchical localization at large scale. arXiv preprint arXiv:1812.03506 (2018) +58. Sattler, T., Maddern, W., Toft, C., Torii, A., Hammarstrand, L., Stenberg, E., Safari, D., Okutomi, M., M. P., J. S., F. K., T. P.: Benchmarking 6dof outdoor visual localization in changing conditions. In: Proc. CVPR (2018) +59. Schleiss, M., Rouatbi, F., Cremers, D.: Vpair - aerial visual place recognition and localization in large-scale outdoor environments (2022) +60. Schubert, S., Neubert, P., Garg, S., Milford, M., Fischer, T.: Visual place recognition: A tutorial. RAM (2023) +61. Shlapentokh-Rothman, M., Blume, A., Xiao, Y., Wu, Y., TV, S., Tao, H., Lee, J.Y., Torres, W., Wang, Y.X., Hoiem, D.: Region-based representations revisited. arXiv preprint arXiv:2402.02352 (2024) +62. Shubodh, S., Omama, M., Zaidi, H., Parihar, U.S., Krishna, M.: Lip-loc: Lidar image pretraining for cross-modal localization. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 948-957 (2024) +63. Sivic, J., Zisserman, A.: Video google: A text retrieval approach to object matching in videos. In: Proceedings of International Conference on Computer Vision (ICCV). p. 1470. IEEE (2003) +64. Sun, X., Xie, Y., Luo, P., Wang, L.: A dataset for benchmarking image-based localization. 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR) pp. 5641-5649 (2017), https://apisemantic scholar.org/CorpusID:20531893 +65. Sünderhauf, N., Shirazi, S., Jacobson, A., Dayoub, F., Pepperell, E., Upcroft, B., Milford, M.: Place recognition with convnet landmarks: Viewpoint-robust, condition-robust, training-free. Proceedings of Robotics: Science and Systems XII (2015) +66. Torii, A., Sivic, J., Pajdla, T., Okutomi, M.: Visual place recognition with repetitive structures. In: 2013 IEEE Conference on Computer Vision and Pattern Recognition. pp. 883-890 (2013). https://doi.org/10.1109/CVPR.2013.119 +67. Tsintotas, K.A., Bampis, L., Gasteratos, A.: The revisiting problem in simultaneous localization and mapping: A survey on visual loop closure detection. IEEE Transactions on Intelligent Transportation Systems 23(11), 19929-19953 (2022) +68. Wang, R., Shen, Y., Zuo, W., Zhou, S., Zheng, N.: Transvpr: Transformer-based place recognition with multi-level attention aggregation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13648-13657 (2022) +69. Warburg, F., Hauberg, S., López-Antequera, M., Gargallo, P., Kuang, Y., Civera, J.: Mapillary street-level sequences: A dataset for lifelong place recognition. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2626-2635 (2020) +70. Weinzaepfel, P., Csurka, G., Cabon, Y., Humenberger, M.: Visual localization by learning objects-of-interest dense match regression. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5634-5643 (2019) + +71. Xin, Z., Cai, Y., Lu, T., Xing, X., Cai, S., Zhang, J., Yang, Y., Wang, Y.: Localizing discriminative visual landmarks for place recognition. In: 2019 International conference on robotics and automation (ICRA). pp. 5979-5985. IEEE (2019) +72. Yildiz, B., Khademi, S., Siebes, R.M., Van Gemert, J.: Amstertime: A visual place recognition benchmark dataset for severe domain shift. In: 2022 26th International Conference on Pattern Recognition (ICPR). IEEE (Aug 2022). https://doi.org/10.1109/icpr56361.2022.9956049, http://dx.doi.org/10.1109/ICPR56361.2022.9956049 +73. Yu, J., Zhu, C., Zhang, J., Huang, Q., Tao, D.: Spatial pyramid-enhanced netvlad with weighted triplet loss for place recognition. IEEE Trans. Neural Netw. Learn. Syst. 31(2), 661-674 (2019) +74. Zaffar, M., Garg, S., Milford, M., Kooij, J., Flynn, D., McDonald-Maier, K., Ehsan, S.: Vpr-bench: An open-source visual place recognition evaluation framework with quantifiable viewpoint and appearance change. International Journal of Computer Vision pp. 1-39 (2021) +75. Zhang, Y., Zhao, X.: Mesa: Matching everything by segmenting anything. arXiv preprint arXiv:2401.16741 (2024) +76. Zhou, B., Lapedriza, A., Khosla, A., Oliva, A., Torralba, A.: Places: A 10 million image database for scene recognition. IEEE Transactions on Pattern Analysis and Machine Intelligence (2017) +77. Zou, X., Yang, J., Zhang, H., Li, F., Li, L., Wang, J., Wang, L., Gao, J., Lee, Y.J.: Segment everything everywhere all at once. Advances in Neural Information Processing Systems 36 (2024) \ No newline at end of file diff --git a/2024/Revisit Anything_ Visual Place Recognition via Image Segment Retrieval/images.zip b/2024/Revisit Anything_ Visual Place Recognition via Image Segment Retrieval/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..c7df62cdb7f481b2eae4045b69ccc866abacaac3 --- /dev/null +++ b/2024/Revisit Anything_ Visual Place Recognition via Image Segment Retrieval/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:dfeae20b298ec0f920e155818e47227e1055112219a0b94f9f8de39ade655bf6 +size 422000 diff --git a/2024/Revisit Anything_ Visual Place Recognition via Image Segment Retrieval/layout.json b/2024/Revisit Anything_ Visual Place Recognition via Image Segment Retrieval/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..f17231a9f8908f879e00a320615e86331d4b6447 --- /dev/null +++ b/2024/Revisit Anything_ Visual Place Recognition via Image Segment Retrieval/layout.json @@ -0,0 +1,8447 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 142, + 111, + 473, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 142, + 111, + 473, + 148 + ], + "spans": [ + { + "bbox": [ + 142, + 111, + 473, + 148 + ], + "type": "text", + "content": "Revisit Anything: Visual Place Recognition via Image Segment Retrieval" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 173, + 168, + 440, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 173, + 168, + 440, + 194 + ], + "spans": [ + { + "bbox": [ + 173, + 168, + 440, + 194 + ], + "type": "text", + "content": "Kartik Garg" + }, + { + "bbox": [ + 173, + 168, + 440, + 194 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 173, + 168, + 440, + 194 + ], + "type": "text", + "content": ", Sai Shubodh Pulgilla" + }, + { + "bbox": [ + 173, + 168, + 440, + 194 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 173, + 168, + 440, + 194 + ], + "type": "text", + "content": ", Shishir Kolathaya" + }, + { + "bbox": [ + 173, + 168, + 440, + 194 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 173, + 168, + 440, + 194 + ], + "type": "text", + "content": ", Madhava Krishna" + }, + { + "bbox": [ + 173, + 168, + 440, + 194 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 173, + 168, + 440, + 194 + ], + "type": "text", + "content": ", and Sourav Garg" + }, + { + "bbox": [ + 173, + 168, + 440, + 194 + ], + "type": "inline_equation", + "content": "^{3}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 132, + 201, + 465, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 201, + 465, + 236 + ], + "spans": [ + { + "bbox": [ + 132, + 201, + 465, + 236 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 132, + 201, + 465, + 236 + ], + "type": "text", + "content": " Indian Institute of Science (IISc), Bengaluru, India \n" + }, + { + "bbox": [ + 132, + 201, + 465, + 236 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 132, + 201, + 465, + 236 + ], + "type": "text", + "content": " International Institute of Information Technology, Hyderabad, India \n" + }, + { + "bbox": [ + 132, + 201, + 465, + 236 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 132, + 201, + 465, + 236 + ], + "type": "text", + "content": " University of Adelaide, Australia" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 236, + 481, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 236, + 481, + 258 + ], + "spans": [ + { + "bbox": [ + 133, + 236, + 481, + 258 + ], + "type": "text", + "content": "kartikgarg@iisc.ac.in, p.saishubodh@gmail.com, shishirk@iisc.ac.in, mkrishna@iiit.ac.in, sourav.garg@adelaide.edu.au" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 160, + 289, + 455, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 289, + 455, + 552 + ], + "spans": [ + { + "bbox": [ + 160, + 289, + 455, + 552 + ], + "type": "text", + "content": "Abstract. Accurately recognizing a revisited place is crucial for embodied agents to localize and navigate. This requires visual representations to be distinct, despite strong variations in camera viewpoint and scene appearance. Existing visual place recognition pipelines encode the whole image and search for matches. This poses a fundamental challenge in matching two images of the same place captured from different camera viewpoints: the similarity of what overlaps can be dominated by the dissimilarity of what does not overlap. We address this by encoding and searching for image segments instead of the whole images. We propose to use open-set image segmentation to decompose an image into 'meaningful' entities (i.e., things and stuff). This enables us to create a novel image representation as a collection of multiple overlapping subgraphs connecting a segment with its neighboring segments, dubbed SuperSegment. Furthermore, to efficiently encode these SuperSegments into compact vector representations, we propose a novel factorized representation of feature aggregation. We show that retrieving these partial representations leads to significantly higher recognition recall than the typical whole image based retrieval. Our segments-based approach, dubbed SegVLAD, sets a new state-of-the-art in place recognition on a diverse selection of benchmark datasets, while being applicable to both generic and task-specialized image encoders. Finally, we demonstrate the potential of our method to \"revisit anything\" by evaluating our method on an object instance retrieval task, which bridges the two disparate areas of research: visual place recognition and object-goal navigation, through their common aim of recognizing goal objects specific to a place. Source code: https://github.com/AnyLoc/Revisit-Anything." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 160, + 563, + 420, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 563, + 420, + 574 + ], + "spans": [ + { + "bbox": [ + 160, + 563, + 420, + 574 + ], + "type": "text", + "content": "Keywords: Visual Place Recognition " + }, + { + "bbox": [ + 160, + 563, + 420, + 574 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 563, + 420, + 574 + ], + "type": "text", + "content": " Image Segmentation " + }, + { + "bbox": [ + 160, + 563, + 420, + 574 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 563, + 420, + 574 + ], + "type": "text", + "content": " Robotics" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 596, + 219, + 608 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 596, + 219, + 608 + ], + "spans": [ + { + "bbox": [ + 132, + 596, + 219, + 608 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 621, + 482, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 621, + 482, + 647 + ], + "spans": [ + { + "bbox": [ + 130, + 621, + 482, + 647 + ], + "type": "text", + "content": "Visual Place Recognition (VPR) is an important capability for embodied agents to localize and navigate autonomously. A predominant solution for VPR is to encode" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 149, + 654, + 227, + 666 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 654, + 227, + 666 + ], + "spans": [ + { + "bbox": [ + 149, + 654, + 227, + 666 + ], + "type": "text", + "content": "* Equal contribution" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 117, + 482, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 117, + 482, + 392 + ], + "spans": [ + { + "bbox": [ + 132, + 117, + 482, + 392 + ], + "type": "text", + "content": "an image into a global vector and retrieve similar vectors as coarse localization hypotheses [19, 44, 60, 67]. Thus, for almost a decade, researchers have focused on learning/finetuning image encoders so that global descriptors are induced with invariance to appearance [3,56,69], viewpoint [3,8], and clutter [27]. On the other hand, there is a vast literature on local descriptors (point/pixel-level), mainly relevant for geometric reranking in hierarchical VPR [11, 16, 25, 57, 68]. In the middle of local and global descriptors exists a variety of methods that use regions/patches [4, 25], lines/planes [17], objects (things/stuff) [21, 31, 45, 65], and segments [26, 33, 50] to represent images. However, these methods are still only aimed at either improving global descriptors based coarse retrieval or local feature matching based reranking. In this work, in contrast to conventional retrieval-based VPR, we explore an alternative: retrieval via encoding segments instead of the whole image. This is particularly enabled by recent advances in open-set image segmentation [37] which can meaningfully deconstruct a place into 'things' (and/or 'stuff') [10]. Thus, we reformulate the VPR problem of revisiting places as that to revisiting things by enabling recognition of these specific things within the context of their place. While such a segment-level place recognition approach provides a direct link to higher-level semantic tasks, such as object-goal navigation [12, 20, 24, 43], it also addresses a fundamental issue in matching partially-overlapping images from across significant viewpoint change. Segments-based partial image representation avoids the mismatches caused by the whole-image representation when the similarity of what overlaps is dominated by the dissimilarity of what does not overlap. Our novel segments-based VPR method, dubbed SegVLAD (Segment based Vector of Locally Aggregated Descriptors), is illustrated in Figure 1, which makes the following novel contributions:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 136, + 402, + 479, + 483 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 136, + 402, + 479, + 435 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 402, + 479, + 435 + ], + "spans": [ + { + "bbox": [ + 136, + 402, + 479, + 435 + ], + "type": "text", + "content": "1. an image representation as a collection of multiple overlapping subgraphs of segments, dubbed SuperSegments, which enables accurate recognition across partially-overlapping images;" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 136, + 437, + 479, + 459 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 437, + 479, + 459 + ], + "spans": [ + { + "bbox": [ + 136, + 437, + 479, + 459 + ], + "type": "text", + "content": "2. a factorized representation of feature aggregation to effectively accommodate both segment-level information as well as segment neighborhood information; and" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 136, + 462, + 479, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 462, + 479, + 483 + ], + "spans": [ + { + "bbox": [ + 136, + 462, + 479, + 483 + ], + "type": "text", + "content": "3. a similarity-weighted ranking method to convert segment-level retrieval into image-level retrieval." + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 132, + 497, + 482, + 591 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 497, + 482, + 591 + ], + "spans": [ + { + "bbox": [ + 132, + 497, + 482, + 591 + ], + "type": "text", + "content": "Using a diverse set of data sources, we demonstrate that our proposed segments-based retrieval enables place recognition under wide viewpoint variations, where global descriptor based retrieval suffers. SegVLAD achieves a new state-of-the-art on multiple challenging datasets. We also introduce an evaluation of our method on an instance-level object retrieval task – a novel capability of our pipeline unlike conventional VPR methods. We conduct several ablations and parameter studies to justify the design choices and emphasize the effectiveness of our method as an open-set segments-based coarse retriever." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 133, + 613, + 230, + 627 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 613, + 230, + 627 + ], + "spans": [ + { + "bbox": [ + 133, + 613, + 230, + 627 + ], + "type": "text", + "content": "2 Related Works" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 641, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 641, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 641, + 481, + 666 + ], + "type": "text", + "content": "Image retrieval-based Visual Place Recognition (VPR) is a well-established area of research in visual localization [19,44,60,67]. It is important both during mapping for loop" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "type": "text", + "content": "K. Garg et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 134, + 115, + 482, + 270 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 115, + 482, + 270 + ], + "spans": [ + { + "bbox": [ + 134, + 115, + 482, + 270 + ], + "type": "text", + "content": "closures [67] as well as for relocalization [51, 58]. The underlying task in both the scenarios remains the same: how to recognize a previously seen place. The state-of-the-art methods in VPR use a global descriptor-based approach which converts an image into a compact vector to enable fast retrieval [2, 3, 6, 29, 32, 56]. The top retrieved hypotheses are often then re-ranked through compute-intensive local feature matching using geometric information [11, 16, 25, 33, 57, 68]. In contrast to previous approaches, we aim to explore image segment level descriptors in this work. This representation falls between point-based local descriptors and the whole-image based global descriptors. Our approach can be considered as 'semi-global', with the proposed segment (and SuperSegment) based descriptors being a 'spatially-reduced' form of whole-image global representation. This is motivated by our hypothesis that to deal with viewpoint variations in VPR with partially-overlapping images, we need a way to partially represent and match them." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 288, + 257, + 300 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 288, + 257, + 300 + ], + "spans": [ + { + "bbox": [ + 133, + 288, + 257, + 300 + ], + "type": "text", + "content": "2.1 Whole Image Encoders" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 134, + 307, + 482, + 487 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 307, + 482, + 487 + ], + "spans": [ + { + "bbox": [ + 134, + 307, + 482, + 487 + ], + "type": "text", + "content": "Earlier works in whole-image representation used methods like Gist [48], BoW (Bag of Word) [63], and VLAD (Vector of Locally Aggregated Descriptors) [30], often defined using hand-crafted features such as SIFT [41]. In recent years, deep learning based methods have demonstrated remarkable performance, with initial successful methods like NetVLAD [3] now rapidly outperformed by better alternatives such as CosPlace [6], MixVPR [2], EigenPlaces [8], TransVPR [68], and more recently AnyLoc [32], SALAD [29] and VLAD-BuFF [35]. All these learning-based methods improve different aspects of representation learning: training datasets [1, 6, 69], objective/loss functions [2, 8], aggregation methods [29, 54, 56, 68], and generalization [32]. Our approach complements these existing methods as we mainly focus on the use of segment-based information, where the segments can be described by any of the image encoders from the aforementioned techniques. In particular, we demonstrate that both - an off-the-shelf encoder, e.g., DINOv2-AnyLoc [32, 49] or that finetuned specifically for the VPR task, e.g., DINOv2-NetVLAD [35] - can be used in conjunction with our segment-based approach to further elevate place recognition capability." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 502, + 282, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 502, + 282, + 514 + ], + "spans": [ + { + "bbox": [ + 133, + 502, + 282, + 514 + ], + "type": "text", + "content": "2.2 Region/Patch Based Methods" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 134, + 521, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 521, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 134, + 521, + 482, + 665 + ], + "type": "text", + "content": "There exist several methods that employ region or patch level information to enhance representational power [13, 14, 34, 40, 52, 71, 73]. However, most of these methods only use this additional information to generate a single (or concatenated) compact vector representation of an image. Other methods such as Patch-NetVLAD [25] create multiple features per image but their primary purpose is to perform local matching based reranking. In contrast to these methods, we aim to use multiple segment descriptors per image to directly retrieve from a database of segments, without using any geometric information or reranking. The motivation behind this stems from the very nature of hierarchical VPR pipelines: reranking recall is upper bounded by the coarse retriever. A better coarse retriever can improve reranking performance without needing to rerank from a longer list of top hypotheses. MultiVLAD [4] is similar to our method in the spirit of retrieving multiple features per query image. However, like aforementioned" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 384, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 384, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 384, + 91, + 447, + 102 + ], + "type": "text", + "content": "Revisit Anything" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 481, + 140 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 481, + 140 + ], + "type": "text", + "content": "methods, MultiVLAD defines regions arbitrarily, whereas we use image segments obtained from Segment Anything Model (SAM) [37] which are semantically meaningful." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 156, + 284, + 168 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 156, + 284, + 168 + ], + "spans": [ + { + "bbox": [ + 132, + 156, + 284, + 168 + ], + "type": "text", + "content": "2.3 Segments-Enhanced Methods" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 174, + 481, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 174, + 481, + 399 + ], + "spans": [ + { + "bbox": [ + 132, + 174, + 481, + 399 + ], + "type": "text", + "content": "There exist several methods that use semantic segmentation information to improve VPR, as also surveyed in [22]. These methods vary in terms of type of segmentation used and the specific ways in which it is integrated in the VPR pipeline, e.g., planes [17], objects [15,31,45], landmarks [65], outdoor semantics [21,23,46,47], utility/confusion based [33, 38], domain adaptation [26] and even learning to segment for VPR [50]. However, neither these methods aim to perform segment-level retrieval nor do they use open-set segmentation. We also review two concurrent works: MESA [75] and Region-Revisited [61]. Similar to our method, they both use SAM to segment images but for different specific tasks. MESA [75] proposes a graph-based local feature/area matching method to obtain point correspondences. Our method complements this, as we perform coarse retrieval for VPR, which could potentially use MESA for reranking. Regions Revisited [61] delves into the advantages of using SAM masks in conjunction with SLIC [36] to improve semantic segmentation, activity recognition and object category retrieval. In contrast, we aim to improve instance-level recognition by recognizing specific things belonging to specific places that a robot encounters during a revisit. Similar to our work, [20] creates an image sequence-based topological graph of segments where its segment neighbourhood aggregation is based on average pooling, similar to [61]. In Section 5.3, we show that such segment average pooling deteriorates recognition performance for the VPR task." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 416, + 218, + 428 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 416, + 218, + 428 + ], + "spans": [ + { + "bbox": [ + 132, + 416, + 218, + 428 + ], + "type": "text", + "content": "2.4 Open-set VPR" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 434, + 481, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 434, + 481, + 590 + ], + "spans": [ + { + "bbox": [ + 132, + 434, + 481, + 590 + ], + "type": "text", + "content": "Researchers have recently started to shift their focus to open-set, generally-applicable techniques, including that for VPR [31, 32, 45, 62]. FM-Loc [45] uses GPT [9] to recognize object and place labels, whereas [31] uses CLIP [55] for open-set place recognition. LIP-Loc [62] proposes pretraining for cross-modal VPR, but is limited in its zero-shot capabilities. AnyLoc [32] proposes to use DINOv2 with domain-level vocabularies and hard-assignment based VLAD. It achieves state-of-the-art performance particularly on non-streetview datasets, where current VPR-trained methods tend to fail. In this work, we propose a generally-applicable approach which is built on top of models like SAM [37] and DINOv2 [49], and works with both VPR-agnostic [32] and VPR-specific [35] backbone models. We particularly aim for a new paradigm in retrieval based VPR, where we move away from the conventional whole image global descriptors to segments based descriptors and retrieval, which achieves a new state-of-the-art on diverse domains under wide viewpoint variations." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 606, + 254, + 620 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 606, + 254, + 620 + ], + "spans": [ + { + "bbox": [ + 132, + 606, + 254, + 620 + ], + "type": "text", + "content": "3 Proposed Approach" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 629, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 629, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 629, + 481, + 665 + ], + "type": "text", + "content": "Despite recent advances in place recognition, viewpoint variations continue to be an open challenge for an embodied agent to recognize the same specific things in its environment. Current methods in visual place recognition tackle this problem by converting" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 92, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 92, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 92, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "type": "text", + "content": "K. Garg et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 132, + 118, + 482, + 266 + ], + "blocks": [ + { + "bbox": [ + 132, + 118, + 482, + 266 + ], + "lines": [ + { + "bbox": [ + 132, + 118, + 482, + 266 + ], + "spans": [ + { + "bbox": [ + 132, + 118, + 482, + 266 + ], + "type": "image", + "image_path": "43e8379a038ad16d7f713e22f77c5f5c6bf8270f93a5ce55510a261c89221503.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 278, + 482, + 334 + ], + "lines": [ + { + "bbox": [ + 130, + 278, + 482, + 334 + ], + "spans": [ + { + "bbox": [ + 130, + 278, + 482, + 334 + ], + "type": "text", + "content": "Fig. 1: Overview of our segment-retrieval based VPR pipeline, dubbed SegVLAD: We use open-set segmentor (SAM) to extract segment masks, which are converted into SuperSegments using the neighbouring image segments. Using pixel-level DINOv2 descriptors with VLAD-based aggregation, we obtain SuperSegment descriptors, which are matched against a flat index of SuperSegment descriptors obtained from all the images of the entire reference database." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 357, + 482, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 357, + 482, + 430 + ], + "spans": [ + { + "bbox": [ + 130, + 357, + 482, + 430 + ], + "type": "text", + "content": "an image as a whole into a global descriptor, which does not explicitly deal with the problem of partial visual overlap caused by viewpoint variations. We propose an alternative solution by representing images partially with the help of image segments. In the following subsections, we describe our representation and retrieval method, which deviates from the conventional VPR techniques but creates a new capability in terms of recognizing objects/things that constitute a place." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 449, + 250, + 459 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 449, + 250, + 459 + ], + "spans": [ + { + "bbox": [ + 132, + 449, + 250, + 459 + ], + "type": "text", + "content": "3.1 Problem Formulation" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 469, + 482, + 554 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 469, + 482, + 554 + ], + "spans": [ + { + "bbox": [ + 130, + 469, + 482, + 554 + ], + "type": "text", + "content": "We represent an image as a set of segment descriptors instead of a single global descriptor. For an image " + }, + { + "bbox": [ + 130, + 469, + 482, + 554 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 130, + 469, + 482, + 554 + ], + "type": "text", + "content": ", we obtain binary image segment masks " + }, + { + "bbox": [ + 130, + 469, + 482, + 554 + ], + "type": "inline_equation", + "content": "M \\in \\{0,1\\}^{S \\times N}" + }, + { + "bbox": [ + 130, + 469, + 482, + 554 + ], + "type": "text", + "content": " and dense pixel-level descriptors " + }, + { + "bbox": [ + 130, + 469, + 482, + 554 + ], + "type": "inline_equation", + "content": "f_{p} \\in \\mathcal{R}^{D}" + }, + { + "bbox": [ + 130, + 469, + 482, + 554 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 469, + 482, + 554 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 130, + 469, + 482, + 554 + ], + "type": "text", + "content": " represents the number of segments per image, " + }, + { + "bbox": [ + 130, + 469, + 482, + 554 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 130, + 469, + 482, + 554 + ], + "type": "text", + "content": " is the descriptor dimension, and " + }, + { + "bbox": [ + 130, + 469, + 482, + 554 + ], + "type": "inline_equation", + "content": "p \\in [1,N]" + }, + { + "bbox": [ + 130, + 469, + 482, + 554 + ], + "type": "text", + "content": " represents spatial elements across the width " + }, + { + "bbox": [ + 130, + 469, + 482, + 554 + ], + "type": "inline_equation", + "content": "(W)" + }, + { + "bbox": [ + 130, + 469, + 482, + 554 + ], + "type": "text", + "content": " and height " + }, + { + "bbox": [ + 130, + 469, + 482, + 554 + ], + "type": "inline_equation", + "content": "(H)" + }, + { + "bbox": [ + 130, + 469, + 482, + 554 + ], + "type": "text", + "content": " of the image encoder's output, flattened into " + }, + { + "bbox": [ + 130, + 469, + 482, + 554 + ], + "type": "inline_equation", + "content": "N = W \\times H" + }, + { + "bbox": [ + 130, + 469, + 482, + 554 + ], + "type": "text", + "content": " for convenience. Figure 1 shows an illustration of our proposed pipeline, as explained in the following subsections." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 572, + 227, + 585 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 572, + 227, + 585 + ], + "spans": [ + { + "bbox": [ + 132, + 572, + 227, + 585 + ], + "type": "text", + "content": "3.2 Super Segments" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": "Humans are remarkable at visual recognition, where existing studies suggest that we often leverage spatial associations among objects in an environment to represent it internally [5, 28]. This enables us to distinguish between two different scenes through the surrounding context of the objects of interest. In this work, we imbibe this context through the spatial neighbourhood of the image segments. For each image, we construct a graph of segments through their pixel centers using Delaunay Triangulation." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 384, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 384, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 384, + 91, + 447, + 102 + ], + "type": "text", + "content": "Revisit Anything" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 114, + 481, + 180 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 481, + 180 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 481, + 180 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 481, + 180 + ], + "type": "image", + "image_path": "dced319d20016bdc56ce97d6c7c1379b3df2ff98a3f48ddbd2f9a35a4e04a343.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 189, + 482, + 234 + ], + "lines": [ + { + "bbox": [ + 130, + 189, + 482, + 234 + ], + "spans": [ + { + "bbox": [ + 130, + 189, + 482, + 234 + ], + "type": "text", + "content": "Fig. 2: Neighborhood expansion (Eq. 1) of a window in the leftmost image to the whole building in the rightmost image, progressing from no neighborhood aggregation to a third-order aggregation. This neighborhood expansion is in stark contrast with a typical regular grid- or patch-based approach which may not capture semantically-meaningful SuperSegments." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 133, + 243, + 481, + 294 + ], + "blocks": [ + { + "bbox": [ + 133, + 243, + 481, + 294 + ], + "lines": [ + { + "bbox": [ + 133, + 243, + 481, + 294 + ], + "spans": [ + { + "bbox": [ + 133, + 243, + 481, + 294 + ], + "type": "image", + "image_path": "27792164f535fc6e55aef41b060198439100414bd496b51ad08fbb48df3dba01.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 303, + 482, + 337 + ], + "lines": [ + { + "bbox": [ + 130, + 303, + 482, + 337 + ], + "spans": [ + { + "bbox": [ + 130, + 303, + 482, + 337 + ], + "type": "text", + "content": "Fig. 3: Illustration of four SuperSegments obtained from the same image. All four of these spatially overlap with each other, which is different from coarse segmentation methods that do not typically allow overlap across segments." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 357, + 482, + 393 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 357, + 482, + 393 + ], + "spans": [ + { + "bbox": [ + 130, + 357, + 482, + 393 + ], + "type": "text", + "content": "This provides us with a binary adjacency matrix " + }, + { + "bbox": [ + 130, + 357, + 482, + 393 + ], + "type": "inline_equation", + "content": "A \\in \\{0,1\\}^{S \\times S}" + }, + { + "bbox": [ + 130, + 357, + 482, + 393 + ], + "type": "text", + "content": " to define the neighborhood for individual segments. We use this adjacency information to expand the context of individual segments to generate new SuperSegment masks " + }, + { + "bbox": [ + 130, + 357, + 482, + 393 + ], + "type": "inline_equation", + "content": "(\\mathcal{M})" + }, + { + "bbox": [ + 130, + 357, + 482, + 393 + ], + "type": "text", + "content": " as below:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 244, + 398, + 481, + 412 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 244, + 398, + 481, + 412 + ], + "spans": [ + { + "bbox": [ + 244, + 398, + 481, + 412 + ], + "type": "interline_equation", + "content": "\\mathcal {M} _ {S \\times N} = ^ {1} \\left(A _ {S \\times S} ^ {o} \\cdot M _ {S \\times N}\\right) \\tag {1}", + "image_path": "9f2947f376dc2ac7cf389ec7b5dbd03423dd48c81ad298a471a623676fd8ebcd.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 418, + 482, + 598 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 418, + 482, + 598 + ], + "spans": [ + { + "bbox": [ + 130, + 418, + 482, + 598 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 418, + 482, + 598 + ], + "type": "inline_equation", + "content": "o \\geq 0" + }, + { + "bbox": [ + 130, + 418, + 482, + 598 + ], + "type": "text", + "content": " refers to the order for expanding the neighborhood by multiplying the adjacency matrix by itself as " + }, + { + "bbox": [ + 130, + 418, + 482, + 598 + ], + "type": "inline_equation", + "content": "A^{o + 1} = A^o \\cdot A" + }, + { + "bbox": [ + 130, + 418, + 482, + 598 + ], + "type": "text", + "content": ". This is matrix-multiplied with the original segmentation masks " + }, + { + "bbox": [ + 130, + 418, + 482, + 598 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 130, + 418, + 482, + 598 + ], + "type": "text", + "content": " to expand the neighborhood at pixel level. " + }, + { + "bbox": [ + 130, + 418, + 482, + 598 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 130, + 418, + 482, + 598 + ], + "type": "text", + "content": " is obtained after element-wise binarization (denoted with " + }, + { + "bbox": [ + 130, + 418, + 482, + 598 + ], + "type": "inline_equation", + "content": "\\mathbb{1}()" + }, + { + "bbox": [ + 130, + 418, + 482, + 598 + ], + "type": "text", + "content": ") so that all pixels in the SuperSegment mask may only contribute once to the subsequent feature aggregation. In Figure 2, we illustrate the extent of image area covered with different orders of mask expansion. Unlike, a patch or regular grid-based approach, the expanded mask of the window in the leftmost image covers a meaningful entity (building) in the rightmost image. Our approach to creating SuperSegments differs from coarse segmentation methods or superpixels in terms of the 'self-overlap'. By expanding neighborhood of each individual segment, we obtain several partially overlapping SuperSegments. A coarse segmentor will need to make assumptions about the right sub-segments to be coalesced so that it can enable accurate recognition from a different viewpoint, which could otherwise lead to the same limitation as that of the whole-image descriptors. Figure 3 presents examples of multiple overlapping SuperSegments from the same image." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 613, + 272, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 613, + 272, + 624 + ], + "spans": [ + { + "bbox": [ + 131, + 613, + 272, + 624 + ], + "type": "text", + "content": "3.3 SuperSegment Descriptors" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "text", + "content": "In this section, we describe our feature aggregation method to obtain SuperSegment descriptors. Recent state-of-the-art method AnyLoc [32] demonstrated that using off-the-shelf powerful image encoders such as DINOv2 with hard assignment based VLAD" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 92, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 92, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 92, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "type": "text", + "content": "K. Garg et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 131, + 116, + 482, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 116, + 482, + 200 + ], + "spans": [ + { + "bbox": [ + 131, + 116, + 482, + 200 + ], + "type": "text", + "content": "aggregation achieves superior recognition performance. However, AnyLoc does not use segmentation information and only operates at the whole-image level. More recently, [61] showed that average pooling works well for segment-level descriptors, but it didn't consider segment neighborhood information. In this work, we propose a unified formulation for feature aggregation that can easily switch across segments, segment neighborhood and the whole image as well as different aggregation types (see supplementary for details). This simply extends Equation 1 as below:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 231, + 211, + 481, + 225 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 231, + 211, + 481, + 225 + ], + "spans": [ + { + "bbox": [ + 231, + 211, + 481, + 225 + ], + "type": "interline_equation", + "content": "F _ {S \\times D} = ^ {1} \\left(A _ {S \\times S} ^ {o} \\cdot M _ {S \\times N}\\right) \\cdot T _ {N \\times D} \\tag {2}", + "image_path": "077ea0bfe4b1eaaadae71fff5212d5aae35473ee456e8fa24fbd7c37bd517f0d.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 238, + 482, + 298 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 238, + 482, + 298 + ], + "spans": [ + { + "bbox": [ + 131, + 238, + 482, + 298 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 238, + 482, + 298 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 131, + 238, + 482, + 298 + ], + "type": "text", + "content": " represents the features to be aggregated. By replacing " + }, + { + "bbox": [ + 131, + 238, + 482, + 298 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 131, + 238, + 482, + 298 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 238, + 482, + 298 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 131, + 238, + 482, + 298 + ], + "type": "text", + "content": " with ones matrices, one can obtain the whole-image global descriptor. For methods like Global Average Pooling (GAP), " + }, + { + "bbox": [ + 131, + 238, + 482, + 298 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 131, + 238, + 482, + 298 + ], + "type": "text", + "content": " can be directly used as the output of the image encoder. In our work, we use Hard-VLAD, for which " + }, + { + "bbox": [ + 131, + 238, + 482, + 298 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 131, + 238, + 482, + 298 + ], + "type": "text", + "content": " is the residual feature matrix per cluster and is obtained as below with respect to each of the cluster centers " + }, + { + "bbox": [ + 131, + 238, + 482, + 298 + ], + "type": "inline_equation", + "content": "c_k" + }, + { + "bbox": [ + 131, + 238, + 482, + 298 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 216, + 309, + 481, + 324 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 216, + 309, + 481, + 324 + ], + "spans": [ + { + "bbox": [ + 216, + 309, + 481, + 324 + ], + "type": "interline_equation", + "content": "T _ {N _ {k} \\times D} ^ {k} = \\left\\{\\alpha_ {k} \\left(f _ {p}\\right) \\left(f _ {p} - c _ {k}\\right) \\mid \\alpha_ {k} \\left(f _ {p}\\right) = 1 \\right\\} \\tag {3}", + "image_path": "759bce08b5e813c4a56f3f028f48c72c6c4ced351dcd0bc33724423b7643266f.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 335, + 482, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 335, + 482, + 384 + ], + "spans": [ + { + "bbox": [ + 131, + 335, + 482, + 384 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 335, + 482, + 384 + ], + "type": "inline_equation", + "content": "\\alpha_{k}(f_{p})\\in \\{0,1\\}" + }, + { + "bbox": [ + 131, + 335, + 482, + 384 + ], + "type": "text", + "content": " is 1 if " + }, + { + "bbox": [ + 131, + 335, + 482, + 384 + ], + "type": "inline_equation", + "content": "f_{p}" + }, + { + "bbox": [ + 131, + 335, + 482, + 384 + ], + "type": "text", + "content": " belongs to " + }, + { + "bbox": [ + 131, + 335, + 482, + 384 + ], + "type": "inline_equation", + "content": "c_{k}" + }, + { + "bbox": [ + 131, + 335, + 482, + 384 + ], + "type": "text", + "content": ". The cluster centers (vocabulary) can be constructed using the map or the domain [32]. The SuperSegment VLAD descriptors obtained from Eq. 2 are intra-normalized, concatenated across clusters and then finally l2-normalized, following existing works [3, 32]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 409, + 285, + 422 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 409, + 285, + 422 + ], + "spans": [ + { + "bbox": [ + 132, + 409, + 285, + 422 + ], + "type": "text", + "content": "3.4 Image Retrieval via Segments" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 437, + 482, + 584 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 437, + 482, + 584 + ], + "spans": [ + { + "bbox": [ + 131, + 437, + 482, + 584 + ], + "type": "text", + "content": "Existing global descriptor based VPR techniques produce a single vector per image to search against a database of reference image vectors. In our method, we obtain multiple SuperSegment descriptors per image. We perform retrieval at segment-level, that is, we search for the top matches for each query segment against a flat index of all segments from all the images of the reference database/map. To evaluate in the form of image retrieval-based VPR, we convert the top retrieved segment indices across all segments of a query image into top reference image indices. This is achieved through a weighted frequency measure (i.e., weighted bin/word counting). We first map the top " + }, + { + "bbox": [ + 131, + 437, + 482, + 584 + ], + "type": "inline_equation", + "content": "K'" + }, + { + "bbox": [ + 131, + 437, + 482, + 584 + ], + "type": "text", + "content": " retrieved segment indices for each of the query segments " + }, + { + "bbox": [ + 131, + 437, + 482, + 584 + ], + "type": "inline_equation", + "content": "s \\in [1, S]" + }, + { + "bbox": [ + 131, + 437, + 482, + 584 + ], + "type": "text", + "content": " to their respective reference image indices, denoted with " + }, + { + "bbox": [ + 131, + 437, + 482, + 584 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 131, + 437, + 482, + 584 + ], + "type": "text", + "content": ". Then, for each of the unique retrieved image indices " + }, + { + "bbox": [ + 131, + 437, + 482, + 584 + ], + "type": "inline_equation", + "content": "r_j" + }, + { + "bbox": [ + 131, + 437, + 482, + 584 + ], + "type": "text", + "content": ", we accumulate its segment similarity " + }, + { + "bbox": [ + 131, + 437, + 482, + 584 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 131, + 437, + 482, + 584 + ], + "type": "text", + "content": " and then use the cumulative similarity score " + }, + { + "bbox": [ + 131, + 437, + 482, + 584 + ], + "type": "inline_equation", + "content": "\\hat{\\theta}" + }, + { + "bbox": [ + 131, + 437, + 482, + 584 + ], + "type": "text", + "content": " to rank the image indices to obtain the top image match " + }, + { + "bbox": [ + 131, + 437, + 482, + 584 + ], + "type": "inline_equation", + "content": "r_j^*" + }, + { + "bbox": [ + 131, + 437, + 482, + 584 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 195, + 596, + 481, + 631 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 195, + 596, + 481, + 631 + ], + "spans": [ + { + "bbox": [ + 195, + 596, + 481, + 631 + ], + "type": "interline_equation", + "content": "r _ {j} ^ {*} = \\underset {r _ {j}} {\\operatorname {a r g m a x}} \\hat {\\theta} \\left(r _ {j}\\right); \\quad \\hat {\\theta} \\left(r _ {j}\\right) = \\sum_ {s = 1} ^ {S} \\sum_ {k = 1} ^ {K ^ {\\prime}} \\theta_ {s k} \\cdot \\mathbb {1} _ {\\left\\{r _ {s k} = r _ {j} \\right\\}} \\tag {4}", + "image_path": "b92991ff1e978fa5c5bc1f927a6309ebd6a798daad980160045a0dec0908cff0.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 641, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 641, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 131, + 641, + 482, + 666 + ], + "type": "text", + "content": "In Section 5.3, we compare our similarity-weighted ranking with other alternatives based on frequency or similarity alone." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 384, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 384, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 384, + 91, + 447, + 102 + ], + "type": "text", + "content": "Revisit Anything" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 254, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 254, + 129 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 254, + 129 + ], + "type": "text", + "content": "4 Experimental Setup" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 140, + 481, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 140, + 481, + 224 + ], + "spans": [ + { + "bbox": [ + 133, + 140, + 481, + 224 + ], + "type": "text", + "content": "Datasets: VPR datasets are in abundance, as can be found in several benchmarks including VPR-Bench [74], Deep Visual GeoLocalization Benchmark [7], and AnyLoc [32]. In this work, we used a variety of datasets covering both outdoor and indoor environments. Outdoor datasets include Pitts30k [66], AmsterTime [72], Mapillary Street Level Sequences (MSLS) [69], SF-XL [6], VPAir [59], Revisted Oxford5K and Revisited Paris6k [53]. Indoor datasets include Baidu Mall [64], 17Places [76] and InsideOut [27]. Additional datasets-related details are provided in the supplementary." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 236, + 481, + 487 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 236, + 481, + 487 + ], + "spans": [ + { + "bbox": [ + 133, + 236, + 481, + 487 + ], + "type": "text", + "content": "Evaluation and Benchmarking: We evaluate our method as an image retrieval task using Recall@K metric, where top K' (= 50) retrieved segments per query segment are used to obtain top K (= 5) images (see Eq. 4). We compare against the most recent and high-performing VPR baseline methods. This includes CosPlace [6], MixVPR [2] and EigenPlaces [8], which are trained on large-scale urban datasets for VPR tasks. We further include two very recent state-of-the-art methods that use DINOv2 as the backbone. These include AnyLoc [32] which uses an off-the-shelf DINOv2 model and SALAD [29] which uses a finetuned DINOv2 backbone. Given the dichotomy between general-purpose VPR benchmarking of AnyLoc and the typical outdoor-focused benchmarking in SALAD and other methods, we evaluated our method using two different backbones. a) SegVLAD-PreT: we use the same backbone and aggregation as AnyLoc, i.e., off-the-shelf pretrained DINOv2 (ViT-G) backbone with hard VLAD assignment, but the key difference is in the use of SuperSegments for our method as opposed to whole-image description of AnyLoc. b) SegVLAD-FineT: since our default aggregation method is VLAD, we use a finetuned DINOv2 (ViT-B) backbone which is similar to SALAD but replaces its aggregation layer with the original NetVLAD aggregation [3] using 64 clusters, as described in [35]. We use this finetuned backbone with hard VLAD based assignment, similar to AnyLoc. For both these models, we reduce the descriptor dimensions of the VLAD descriptor to 1024 using PCA, as commonly done in previous works [3, 32]. We train PCA transform in a map-specific manner using the database images of the dataset." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 505, + 190, + 518 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 505, + 190, + 518 + ], + "spans": [ + { + "bbox": [ + 133, + 505, + 190, + 518 + ], + "type": "text", + "content": "5 Results" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 133, + 531, + 481, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 531, + 481, + 578 + ], + "spans": [ + { + "bbox": [ + 133, + 531, + 481, + 578 + ], + "type": "text", + "content": "We first present benchmark comparison of our method against state-of-the-art VPR methods. This is followed by detailed analysis of our proposed aggregation technique. Lastly, we demonstrate results on a downstream task of Object of Interest (OOI) retrieval, showcasing the versatility of our method." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 133, + 596, + 279, + 609 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 596, + 279, + 609 + ], + "spans": [ + { + "bbox": [ + 133, + 596, + 279, + 609 + ], + "type": "text", + "content": "5.1 State-of-the-art comparisons" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 133, + 616, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 616, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 133, + 616, + 481, + 665 + ], + "type": "text", + "content": "Table 1 presents Recall@1/5 comparison against state-of-the-art VPR methods on standard outdoor street-view datasets, which are similar to the typical training datasets used for VPR [2,3,6]. Table 2 covers 'out-of-distribution' datasets, inspired by AnyLoc [32], covering indoor environments (Baidu Mall and 17 Places), aerial imagery (VPAir)," + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "type": "text", + "content": "K. Garg et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 151, + 135, + 462, + 224 + ], + "blocks": [ + { + "bbox": [ + 166, + 114, + 447, + 125 + ], + "lines": [ + { + "bbox": [ + 166, + 114, + 447, + 125 + ], + "spans": [ + { + "bbox": [ + 166, + 114, + 447, + 125 + ], + "type": "text", + "content": "Table 1: Recall@ 1/5 benchmark comparison on outdoor street-view datasets." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 151, + 135, + 462, + 224 + ], + "lines": [ + { + "bbox": [ + 151, + 135, + 462, + 224 + ], + "spans": [ + { + "bbox": [ + 151, + 135, + 462, + 224 + ], + "type": "table", + "html": "
MethodPitts-30KMSLS SFMSLS CPHSF-XL ValRO5k MedRO5k HardRP6k MedRP6k Hard
CosPlace90.4/95.793.4/97.584.9/92.094.6/97.685.7/87.127.1/45.794.3/95.77.1/15.7
MixVPR91.5/95.591.3/95.987.1/92.487.8/93.868.6/80.032.9/54.394.3/10010.0/32.9
EigenPlaces92.6/96.792.6/97.187.1/92.896.4/98.285.7/88.642.8/57.195.7/98.64.3/11.4
AnyLoc87.7/94.783.4/94.679.9/89.184.4/91.988.6/92.940.0/58.697.1/10011.4/44.3
SALAD92.6/96.591.7/97.192.3/96.193.6/97.382.9/90.037.1/54.395.7/98.614.3/58.6
SegVLAD-PreT86.7/94.288.4/94.281.7/90.790.9/96.491.4/95.760.0/81.494.3/1008.6/48.6
SegVLAD-FineT93.2/96.894.6/97.190.9/95.794.9/98.187.1/95.751.4/70.095.7/10010.0/48.6
", + "image_path": "770410ddd0e8122f23a2f3949c8b0138f3dd30b9246759bdb6c73adcc1061500.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 195, + 256, + 421, + 340 + ], + "blocks": [ + { + "bbox": [ + 165, + 236, + 448, + 247 + ], + "lines": [ + { + "bbox": [ + 165, + 236, + 448, + 247 + ], + "spans": [ + { + "bbox": [ + 165, + 236, + 448, + 247 + ], + "type": "text", + "content": "Table 2: Recall@ 1/5 benchmark comparison on 'out-of-distribution' datasets." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 195, + 256, + 421, + 340 + ], + "lines": [ + { + "bbox": [ + 195, + 256, + 421, + 340 + ], + "spans": [ + { + "bbox": [ + 195, + 256, + 421, + 340 + ], + "type": "table", + "html": "
MethodBaiduAmstTimeInsideOut17PlacesVPAir
CosPlace41.6/55.047.7/69.80.2/2.081.3/88.24.6/13.7
MixVPR64.4/80.340.2/59.10.0/1.885.2/90.16.8/16.1
EigenPlaces56.5/72.848.9/69.50.4/1.483.0/90.16.5/17.9
AnyLoc75.2/87.650.3/73.02.4/8.095.3/97.366.7/79.2
SALAD74.8/86.555.4/75.60.6/1.882.5/88.225.8/38.7
SegVLAD-PreT78.5/93.856.8/77.74.2/9.495.3/98.069.8/83.7
SegVLAD-FineT68.1/89.058.9/79.37.4/15.695.1/97.535.4/55.3
", + "image_path": "8170fe01fbfa82cace38a56ea5e323da47967078504fe75c7119809c2d98cb35.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 133, + 363, + 482, + 424 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 363, + 482, + 424 + ], + "spans": [ + { + "bbox": [ + 133, + 363, + 482, + 424 + ], + "type": "text", + "content": "indoor-to-outdoor viewing (InsideOut), and historical image matching (AmsterTime). Below, we discuss two key aspects of this comparative analysis: " + }, + { + "bbox": [ + 133, + 363, + 482, + 424 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 133, + 363, + 482, + 424 + ], + "type": "text", + "content": " ) how our segment-based approach compares against whole-image global descriptor based methods and ii) how performance trends vary depending on the choices of feature backbone with regards to task-specific (VPR) training." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 133, + 442, + 482, + 610 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 442, + 482, + 610 + ], + "spans": [ + { + "bbox": [ + 133, + 442, + 482, + 610 + ], + "type": "text", + "content": "Aggregating Segments vs Whole Images Table 1 and Table 2 show that our proposed method SegVLAD achieves a new state-of-the-art on the majority of datasets, considering both the backbone variants: PreT and FineT. AnyLoc and SALAD respectively differ from SegVLAD-PreT and SegVLAD-FineT in terms of the aggregation scope (global vs segments). Thus, the superior performance of SegVLAD clearly highlights the role of segments based retrieval over whole-image based approach. On the Baidu Mall dataset – highly-aliased indoor environment – our method improves over AnyLoc by " + }, + { + "bbox": [ + 133, + 442, + 482, + 610 + ], + "type": "inline_equation", + "content": "3\\%" + }, + { + "bbox": [ + 133, + 442, + 482, + 610 + ], + "type": "text", + "content": " for R@1 and " + }, + { + "bbox": [ + 133, + 442, + 482, + 610 + ], + "type": "inline_equation", + "content": "6\\%" + }, + { + "bbox": [ + 133, + 442, + 482, + 610 + ], + "type": "text", + "content": " for R@5 in absolute gains. On the InsideOut dataset – matching outdoor images viewed from within indoors – our method leads to a ‘meaningful’ recall, unlike all other baselines. Overall, these results highlight that even with the use of powerful image encoders (DINOv2), global aggregation struggles to deal with the challenges of matching images across major viewpoint shifts – it is thus the partial image representation and matching (using the same encoder) which is needed to obtain superior recognition performance." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 133, + 629, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 629, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 133, + 629, + 482, + 666 + ], + "type": "text", + "content": "VPR Fine-tuned Encoders + Segments For SegVLAD-FineT, we used a DINOv2 backbone finetuned for the purpose of VPR, mainly to observe the benefit of segments over global descriptor based approach in a task-specific manner. Table 1 and Table 2" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 384, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 384, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 384, + 91, + 447, + 102 + ], + "type": "text", + "content": "Revisit Anything" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 137, + 135, + 476, + 161 + ], + "blocks": [ + { + "bbox": [ + 154, + 114, + 459, + 126 + ], + "lines": [ + { + "bbox": [ + 154, + 114, + 459, + 126 + ], + "spans": [ + { + "bbox": [ + 154, + 114, + 459, + 126 + ], + "type": "text", + "content": "Table 3: Recall@1 results for various approaches on Object-Instance Retrieval Task" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 137, + 135, + 476, + 161 + ], + "lines": [ + { + "bbox": [ + 137, + 135, + 476, + 161 + ], + "spans": [ + { + "bbox": [ + 137, + 135, + 476, + 161 + ], + "type": "table", + "html": "
MethodSegVLAD NoNbrAggSegVLADSegment-to-GlobalGlobal-to-Global
R@164.192.730.086.4
", + "image_path": "0d8ce0d71710572182472082fe89623d17735ef79ec4364dd6ec6fb8dc5ae634.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 189, + 480, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 189, + 480, + 236 + ], + "spans": [ + { + "bbox": [ + 130, + 189, + 480, + 236 + ], + "type": "text", + "content": "show that, on the outdoor street-view datasets, SALAD (finetuned DINOv2) generally performs better than AnyLoc (its pretrained counterpart), whereas the latter generally outperforms the former on 'out-of-distribution' datasets. It can be clearly observed that these performance patterns translate well from global- to segment-level results." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 258, + 425, + 270 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 258, + 425, + 270 + ], + "spans": [ + { + "bbox": [ + 131, + 258, + 425, + 270 + ], + "type": "text", + "content": "5.2 Revisiting Objects of Interest (OOI): Object Instance Retrieval" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 281, + 482, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 281, + 482, + 449 + ], + "spans": [ + { + "bbox": [ + 130, + 281, + 482, + 449 + ], + "type": "text", + "content": "A typical requirement of an embodied agent is to understand the context of its task through its memory/map information, which is composed of visual and/or semantic cues. For example, navigating to a given object goal requires a robot to visually recognize the goal and not be confused by perceptually-similar items. In this section, we demonstrate our method's ability to retrieve the correct image given just an Object Of Interest (OOI) as a query segment. For this purpose, we use an extended version of the Baidu dataset [70] which annotates OOI as various discriminative areas that can be reliably detected under variable viewpoint and lighting conditions. In total, there are 220 OOI, which cover various things such as logos, brand names, posters, etc., in a highly cluttered mall environment. To cast this dataset in terms of revisiting things, we use the original query images of the Baidu Mall [64] dataset as the database and the images with OOI as the queries. This allows us to evaluate the OOIs directly. This is similar to VPR evaluation of recall in terms of image retrieval but with querying of a specific segment instead of using all the segments of the query image." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 449, + 482, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 449, + 482, + 592 + ], + "spans": [ + { + "bbox": [ + 130, + 449, + 482, + 592 + ], + "type": "text", + "content": "We consider four different methods of recognizing known objects in this study. i) Global-to-Global: as a baseline method, we use whole images to represent and retrieve, i.e., without using the OOI mask; this resembles object-goal recognition problem for an InstanceImageNavigation task [39]. ii) Segment-to-Global: this is the same as the previous setting except that the query image descriptor is aggregated only using the OOI mask; this tests the ability of the image encoder/aggregator to match segment-level descriptor against global descriptors. iii) SegVLAD and iv) SegVLAD NoNbrAgg, which are our proposed methods but the latter does not use any neighborhood information; this highlights the relevance of spatial context around the OOI for recognition. For SegVLAD, we create a virtual segment mask for the OOI, append it to the other masks of the image, and then perform our neighborhood expansion and feature aggregation, as described in Section 3." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": "Table 3 reports Recall@1 for different recognition methods. It can be observed that SegVLAD outperforms Global-to-Global matching by a large margin, which shows that recognizing specific object instances through their images (as in InstanceImageNav) is more prone to failures. It can further be observed from low recall of SegVLAD NoNbrAgg that neighborhood aggregation around segments is crucial to capture the required context. Finally, poor recall of Segment-to-Global highlights" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "type": "text", + "content": "K. Garg et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 150, + 168, + 462, + 232 + ], + "blocks": [ + { + "bbox": [ + 132, + 114, + 479, + 150 + ], + "lines": [ + { + "bbox": [ + 132, + 114, + 479, + 150 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 479, + 150 + ], + "type": "text", + "content": "Table 4: Recall@1/5 for Baidu mall dataset Table 5: Recall@1/5 comparison between for different aggregation methods and dif- different methods for ranking images based. ferent orders of neighborhood expansion. on segment-level retrieval." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 150, + 168, + 462, + 232 + ], + "lines": [ + { + "bbox": [ + 150, + 168, + 462, + 232 + ], + "spans": [ + { + "bbox": [ + 150, + 168, + 462, + 232 + ], + "type": "table", + "html": "
OrderSegVLADSAP
073.1/89.974.6/91.1
177.4/91.765.6/87.2
276.3/92.453.2/81.3
377.7/92.649.8/78.0
Max Seg78.5/93.953.9/70.4
Max Sim65.2/92.734.4/62.4
Ours78.5/93.854.4/76.3
", + "image_path": "c951be63c9fefc81f88b898c8b2e7ca8c0398aaaebe1bed8b47e15e953d224e6.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 256, + 479, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 256, + 479, + 281 + ], + "spans": [ + { + "bbox": [ + 130, + 256, + 479, + 281 + ], + "type": "text", + "content": "that matching a part of an image (OOI) with the whole image is not a viable solution for object instance recognition." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 300, + 228, + 312 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 300, + 228, + 312 + ], + "spans": [ + { + "bbox": [ + 132, + 300, + 228, + 312 + ], + "type": "text", + "content": "5.3 Ablation Studies" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 322, + 482, + 562 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 322, + 482, + 562 + ], + "spans": [ + { + "bbox": [ + 130, + 322, + 482, + 562 + ], + "type": "text", + "content": "Aggregation method & Order of Neighborhood Expansion Previous studies on VPR such as AnyLoc [32] have shown VLAD to be better than other aggregation methods for whole-image based global descriptors. However, in an increasing number of segment-based approaches [12, 20, 24, 61], segment average pooling (SAP) is used more commonly. Thus, we compare hard-assignment VLAD against SAP on Baidu dataset. For SAP, we upsample the DINOv2 features to match the resolution of our SAM masks – this upsampling is shown to enhance performance in [61]. For VLAD aggregation, we use our proposed method SegVLAD, where we downsample masks to match with the low resolution of DINOv2. Table 4 shows that SAP performs well for order 0 aggregation (i.e., no neighborhood aggregation) but its performance reduces as the neighborhood expands. On the other hand, SegVLAD has low recall when no neighborhood is considered but benefits significantly even with its immediate neighborhood (order 1). We attribute these inverted trends of SegVLAD and SAP to the very nature of these aggregation methods: as more information becomes available SAP smoothes out the overall information content whereas SegVLAD benefits from additional information which gets distributed across its clusters, thus minimizing any possible smoothing effect. It can be observed that R@5 increases for SegVLAD with an increasing order of neighborhood expansion but margins diminish for higher orders. Overall, SegVLAD (order 3) achieves the best results, despite aggregating at a " + }, + { + "bbox": [ + 130, + 322, + 482, + 562 + ], + "type": "inline_equation", + "content": "14 \\times" + }, + { + "bbox": [ + 130, + 322, + 482, + 562 + ], + "type": "text", + "content": " lower resolution than SAP's upsampling based aggregation." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 582, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 482, + 665 + ], + "type": "text", + "content": "Segment to Image Retrieval Unlike conventional global descriptor retrieval based VPR, we perform retrieval for multiple SuperSegments of the query image. To obtain retrieval output in terms of images (as that is what VPR is typically evaluated on), there exist multiple ways to combine the top segment-level matches across all the query segments. We consider the following alternative options. i) MaxSeg: we obtain the best matching segment for each query segment, associate the matched segments to their respective reference image indices, and then rank these image indices based on their" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 384, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 384, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 384, + 91, + 447, + 102 + ], + "type": "text", + "content": "Revisit Anything" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 143, + 113, + 471, + 357 + ], + "blocks": [ + { + "bbox": [ + 143, + 113, + 471, + 357 + ], + "lines": [ + { + "bbox": [ + 143, + 113, + 471, + 357 + ], + "spans": [ + { + "bbox": [ + 143, + 113, + 471, + 357 + ], + "type": "image", + "image_path": "a580aa5b9234506efb32d3196b7efd79d541a96dc84361086e7e6e6579d67123.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 368, + 482, + 401 + ], + "lines": [ + { + "bbox": [ + 130, + 368, + 482, + 401 + ], + "spans": [ + { + "bbox": [ + 130, + 368, + 482, + 401 + ], + "type": "text", + "content": "Fig. 4: Qualitative results: Columns respectively represent the query image, correct match of SegVLAD and incorrect match of AnyLoc. Examples from different datasets: AmsterTime, Baidu Mall, Pitts-30K are presented across the rows." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 434, + 482, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 434, + 482, + 542 + ], + "spans": [ + { + "bbox": [ + 130, + 434, + 482, + 542 + ], + "type": "text", + "content": "frequency; this method weakly resembles an inverted index list based counting of common segments between the query and the reference. ii) MaxSim: across the best matching reference segments, we order their image indices based on the segment similarity; this method is similar to that used in MultiVLAD [4]. iii) Similarity-Weighted Frequency: this is our proposed method as defined in Section 3.4. Table 5 shows that our proposed method for combining segment-level hypotheses consistently achieves superior results for both the datasets. While MaxSeg achieves a similar performance on Baidu, it suffers a drop in recall for AmsterTime. Both the methods outperform MaxSim at R@1 by a large margin." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 582, + 482, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 482, + 665 + ], + "type": "text", + "content": "Patches vs Segment While open-set segmentation [37] is aimed at a meaningful segregation of visual entities, a simple alternative to our segment-based retrieval is to use uniformly defined regions/patches. Table 6 compares SegVLAD with a patch-based approach, where we consider arbitrary square patch sizes to segment an image. It can be observed that SegVLAD outperforms its patch-based counterparts, where smaller patches perform the worst, and for larger patches, R@1 saturates while R@5 reduces. These results are also in line with similar findings of a recent work [61]." + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "type": "text", + "content": "K. Garg et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 145, + 117, + 288, + 202 + ], + "blocks": [ + { + "bbox": [ + 145, + 117, + 288, + 202 + ], + "lines": [ + { + "bbox": [ + 145, + 117, + 288, + 202 + ], + "spans": [ + { + "bbox": [ + 145, + 117, + 288, + 202 + ], + "type": "image", + "image_path": "f4b8ec815e3f5b6f1e792ebaf19ae217e83292ed60fe90566e92a0bf58057df9.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 214, + 301, + 236 + ], + "lines": [ + { + "bbox": [ + 132, + 214, + 301, + 236 + ], + "spans": [ + { + "bbox": [ + 132, + 214, + 301, + 236 + ], + "type": "text", + "content": "Fig.5: Recall vs storage/retrieval time on AmsterTime." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 319, + 156, + 471, + 229 + ], + "blocks": [ + { + "bbox": [ + 312, + 133, + 481, + 155 + ], + "lines": [ + { + "bbox": [ + 312, + 133, + 481, + 155 + ], + "spans": [ + { + "bbox": [ + 312, + 133, + 481, + 155 + ], + "type": "text", + "content": "Table 6: Patch vs Segments on AmsterTime (Reso. " + }, + { + "bbox": [ + 312, + 133, + 481, + 155 + ], + "type": "inline_equation", + "content": "256\\times 256" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 319, + 156, + 471, + 229 + ], + "lines": [ + { + "bbox": [ + 319, + 156, + 471, + 229 + ], + "spans": [ + { + "bbox": [ + 319, + 156, + 471, + 229 + ], + "type": "table", + "html": "
Patch SizeNumSegDbNumSegImR@1/5
16×1631513625635.7/62.1
32×32787846445.9/72.5
64×64196961652.0/75.1
128×1284924453.5/65.6
SegVLAD12963710556.8/77.7
", + "image_path": "5fd7cde689077b3cf64885e0e38a8912180d6e9a4d1e540a7397b60f9433e982.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 260, + 413, + 274 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 260, + 413, + 274 + ], + "spans": [ + { + "bbox": [ + 132, + 260, + 413, + 274 + ], + "type": "text", + "content": "5.4 Limitations, Storage, Compute Time & IOU Based Filtering" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 281, + 482, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 281, + 482, + 460 + ], + "spans": [ + { + "bbox": [ + 130, + 281, + 482, + 460 + ], + "type": "text", + "content": "A key limitation of our method is its large map size, i.e., large storage requirement for segment-level descriptors (see supplementary for further details). In this section, we analyze the resource requirements in terms of database (index) storage and query retrieval time for our method, along with a preliminary study on IOU (Intersection over Union) based filtering of SuperSegments to reduce such costs. We compute IOU between all pairs of SuperSegments in a given image, and remove segments with " + }, + { + "bbox": [ + 130, + 281, + 482, + 460 + ], + "type": "inline_equation", + "content": "IOU(s_i, s_j) > \\psi \\forall i \\in [1, S]" + }, + { + "bbox": [ + 130, + 281, + 482, + 460 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 281, + 482, + 460 + ], + "type": "inline_equation", + "content": "j \\in [i, S]" + }, + { + "bbox": [ + 130, + 281, + 482, + 460 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 281, + 482, + 460 + ], + "type": "inline_equation", + "content": "\\psi \\in [0, 1]" + }, + { + "bbox": [ + 130, + 281, + 482, + 460 + ], + "type": "text", + "content": " is a threshold on IOU and " + }, + { + "bbox": [ + 130, + 281, + 482, + 460 + ], + "type": "inline_equation", + "content": "s_i" + }, + { + "bbox": [ + 130, + 281, + 482, + 460 + ], + "type": "text", + "content": " refers to the list of SuperSegments sorted by their pixel area in a descending order. We only perform this culling on the database segments. We use the outdoor-finetuned DINOv2 backbone for this purpose and compare SegVLAD with SALAD on AmsterTime (see supplementary for additional results on Pitts30K). Figure 5 shows that SegVLAD outperforms SALAD while requiring less storage (annotated on points) and comparable retrieval time (excludes extraction time), using IoU-based filtering threshold ranging from 0.2 to 0.8 (left to right) with a step size of 0.2. In particular, at 0.4 IOU threshold, only 20% of SuperSegments are retained (0.05 GB) while still outperforming the baseline." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 478, + 246, + 490 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 478, + 246, + 490 + ], + "spans": [ + { + "bbox": [ + 132, + 478, + 246, + 490 + ], + "type": "text", + "content": "5.5 Qualitative Analyses" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 498, + 482, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 498, + 482, + 605 + ], + "spans": [ + { + "bbox": [ + 130, + 498, + 482, + 605 + ], + "type": "text", + "content": "In this section, we further demonstrate the capabilities of our method through qualitative visualizations. We compare our method against AnyLoc, where the only difference between the two methods is Global aggregation/retrieval vs SuperSegment aggregation/retrieval. We particularly consider the queries for which our approach successfully retrieved the correct match but AnyLoc failed to do so (additional examples can be found in the supplementary). Figure 4 shows triplets of images in the order of query, correct match (ours), and incorrect match (Anyloc). The segmented part shows one of the correctly matched SuperSegments, displayed as a subgraph in white color overlaid on the corresponding segment masks." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 606, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 606, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 606, + 482, + 666 + ], + "type": "text", + "content": "The first row shows a triplet from AmsterTime where our proposed method is able to correctly recognize a subgraph of building across the image pair, whereas Anyloc retrieves an incorrect image of a street-view with buildings, cars, and road, laid out similarly across the image pair. This highlights that a global descriptor can get confused with perceptually-aliased global context. In contrast, our SuperSegment based" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 384, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 384, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 384, + 91, + 447, + 102 + ], + "type": "text", + "content": "Revisit Anything" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 471, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 471, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 471, + 91, + 480, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 482, + 284 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 482, + 284 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 482, + 284 + ], + "type": "text", + "content": "SegVLAD is not only able to retrieve the correct image but it also correctly finds the mutually-overlapping area. A similar trend follows for the Baidu Mall (middle row) and Pitts30K (last row). In the Baidu Mall example, our approach identifies the piano and the region around it in the query image. It correctly retrieves an image having similar spatial context with the piano. This is akin to how humans use spatial context to identify places. AnyLoc, on the other hand, retrieves an image with similar floor tiles and railings. This example particularly highlights our hypothesis that dissimilarity of nonoverlapping regions can dominate the similarity of overlapping regions in global whole-image descriptors. Finally, the Pitts30K example (last row) shows a case of strong viewpoint change. While SegVLAD correctly matches the traffic signal and signboards to retrieve the correct match, AnyLoc retrieves a similar looking image while missing the finer details. This example particularly reinforces the idea of 'revisiting things', as even though the background mountain is common across the triplet, it is the context of the things near the camera/robot which helps in uniquely recognizing a place." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 302, + 302, + 316 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 302, + 302, + 316 + ], + "spans": [ + { + "bbox": [ + 132, + 302, + 302, + 316 + ], + "type": "text", + "content": "6 Conclusion and Future Work" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 327, + 482, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 327, + 482, + 518 + ], + "spans": [ + { + "bbox": [ + 132, + 327, + 482, + 518 + ], + "type": "text", + "content": "In this paper, we presented a novel visual place recognition pipeline SegVLAD based on image segments-based description and retrieval, which is akin to 'revisting things' as a means to recognize specific instances of what constitute a place. Our proposed SuperSegments based image representation and a novel factorization based feature aggregation enables us to effectively represent and retrieve images using our segment similarity-weighted image ranking. Our results show that despite using powerful image encoders such as DINOv2 (pretrained or VPR-finetuned), existing global descriptor based techniques are unable to deal with the challenges of viewpoint variations. In contrast, SegVLAD is able to correctly retrieve images through its ability to match partially-overlapping images with its partial image representation in the form of semiglobal subgraphs of segments, i.e., SuperSegments. Thus, our method achieves state-of-the-art results on three diverse datasets (indoor and outdoor) that exhibit strong viewpoint variations on top of other challenges of appearance shift and high perceptual aliasing. Through an additional object instance retrieval study, we demonstrate the unique ability of our method to recognize objects instances within their specific place contexts - an open-set recognition capability that existing VPR methods lack." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 519, + 482, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 519, + 482, + 639 + ], + "spans": [ + { + "bbox": [ + 132, + 519, + 482, + 639 + ], + "type": "text", + "content": "Our approach shifts the paradigm in retrieval based VPR research, as the conventional methods predominantly classify into either whole-image global descriptor based coarse retrieval or local feature based geometric reranking. Our approach complements recent concurrent works like MESA [75]; future work can explore a hierarchical VPR pipeline that closely integrates a segment-based coarse retriever with segments-based reranker such as MESA, thus doing away entirely with global whole-image descriptors. Furthermore, segments-based representation with implicitly baked semantics provide a natural way for creating text-based interfaces through CLIP [55] and LLMs (Large Language Models) [9], which can be easily integrated with recent efforts in this direction [12, 18, 20, 24, 42, 43, 77]." + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "type": "text", + "content": "K. Garg et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 114, + 234, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 114, + 234, + 129 + ], + "spans": [ + { + "bbox": [ + 132, + 114, + 234, + 129 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 140, + 482, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 140, + 482, + 224 + ], + "spans": [ + { + "bbox": [ + 130, + 140, + 482, + 224 + ], + "type": "text", + "content": "This work was supported by the Centre for Augmented Reasoning (CAR) at the Australian Institute for Machine Learning (AIML), University of Adelaide, Australia. The authors acknowledge the computational support provided by the Indian Institute of Science (IISc), Bengaluru, India, and the International Institute of Information Technology, Hyderabad (IIITH), India. The authors thank Ahmad Khaliq for technical support, Sarah Ibrahimi for sharing the InsideOut dataset, and Martin Humenberger for sharing the OOI annotations for the Baidu Mall dataset." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 243, + 190, + 255 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 243, + 190, + 255 + ], + "spans": [ + { + "bbox": [ + 132, + 243, + 190, + 255 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 269, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 138, + 269, + 480, + 291 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 269, + 480, + 291 + ], + "spans": [ + { + "bbox": [ + 138, + 269, + 480, + 291 + ], + "type": "text", + "content": "1. Ali-bey, A., Chaib-draa, B., Giguère, P.: Gsv-cities: Toward appropriate supervised visual place recognition. Neurocomputing 513, 194-203 (2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 292, + 480, + 324 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 292, + 480, + 324 + ], + "spans": [ + { + "bbox": [ + 138, + 292, + 480, + 324 + ], + "type": "text", + "content": "2. Ali-bey, A., Chaib-draa, B., Giguère, P.: Mixvpr: Feature mixing for visual place recognition. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 2998-3007 (2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 325, + 480, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 325, + 480, + 357 + ], + "spans": [ + { + "bbox": [ + 138, + 325, + 480, + 357 + ], + "type": "text", + "content": "3. Arandjelovic, R., Gronat, P., Torii, A., Pajdla, T., Sivic, J.: Netvlad: Cnn architecture for weakly supervised place recognition. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 5297-5307 (2016)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 357, + 480, + 379 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 357, + 480, + 379 + ], + "spans": [ + { + "bbox": [ + 138, + 357, + 480, + 379 + ], + "type": "text", + "content": "4. Arandjelovic, R., Zisserman, A.: All about vlad. In: Proceedings of the IEEE conference on Computer Vision and Pattern Recognition. pp. 1578-1585 (2013)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 380, + 463, + 390 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 380, + 463, + 390 + ], + "spans": [ + { + "bbox": [ + 138, + 380, + 463, + 390 + ], + "type": "text", + "content": "5. Bar, M.: Visual objects in context. Nature Reviews Neuroscience 5(8), 617-629 (2004)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 391, + 480, + 423 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 391, + 480, + 423 + ], + "spans": [ + { + "bbox": [ + 138, + 391, + 480, + 423 + ], + "type": "text", + "content": "6. Berton, G., Masone, C., Caputo, B.: Rethinking visual geo-localization for large-scale applications. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4878-4888 (2022)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 424, + 480, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 424, + 480, + 456 + ], + "spans": [ + { + "bbox": [ + 138, + 424, + 480, + 456 + ], + "type": "text", + "content": "7. Berton, G., Mereu, R., Trivigno, G., Masone, C., Csurka, G., Sattler, T., Caputo, B.: Deep visual geo-localization benchmark. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5396-5407 (2022)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 457, + 480, + 489 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 457, + 480, + 489 + ], + "spans": [ + { + "bbox": [ + 138, + 457, + 480, + 489 + ], + "type": "text", + "content": "8. Berton, G., Trivigno, G., Caputo, B., Masone, C.: Eigenplaces: Training viewpoint robust models for visual place recognition. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 11080-11090 (October 2023)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 490, + 480, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 490, + 480, + 521 + ], + "spans": [ + { + "bbox": [ + 138, + 490, + 480, + 521 + ], + "type": "text", + "content": "9. Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J.D., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., et al.: Language models are few-shot learners. Advances in neural information processing systems 33, 1877-1901 (2020)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 522, + 480, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 522, + 480, + 555 + ], + "spans": [ + { + "bbox": [ + 138, + 522, + 480, + 555 + ], + "type": "text", + "content": "0. Caesar, H., Uijlings, J., Ferrari, V.: Coco-stuff: Thing and stuff classes in context. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 1209-1218 (2018)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 555, + 480, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 555, + 480, + 578 + ], + "spans": [ + { + "bbox": [ + 138, + 555, + 480, + 578 + ], + "type": "text", + "content": "1. Cao, B., Araujo, A., Sim, J.: Unifying deep local and global features for image search. In: European Conference on Computer Vision. pp. 726-743. Springer (2020)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 578, + 480, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 578, + 480, + 599 + ], + "spans": [ + { + "bbox": [ + 138, + 578, + 480, + 599 + ], + "type": "text", + "content": "2. Chang, M., Gervet, T., Khanna, M., Yenamandra, S., Shah, D., Min, S.Y., Shah, K., Paxton, C., Gupta, S., Batra, D., et al.: Goat: Go to any thing. arXiv preprint arXiv:2311.06430 (2023)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 138, + 600, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 600, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 138, + 600, + 480, + 632 + ], + "type": "text", + "content": "3. Chen, L.C., Yang, Y., Wang, J., Xu, W., Yuille, A.L.: Attention to scale: Scale-aware semantic image segmentation. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 3640-3649 (2016)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 138, + 633, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 633, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 633, + 480, + 665 + ], + "type": "text", + "content": "4. Chen, Z., Maffra, F., Sa, I., Chli, M.: Only look once, mining distinctive landmarks from convnet for visual place recognition. In: Intelligent Robots and Systems (IROS), 2017 IEEE/RSJ International Conference on. pp. 9-16. IEEE (2017)" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 384, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 384, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 384, + 91, + 447, + 102 + ], + "type": "text", + "content": "Revisit Anything" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 471, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 471, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 471, + 91, + 480, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 117, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 133, + 117, + 481, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 117, + 481, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 117, + 481, + 149 + ], + "type": "text", + "content": "15. Cheng, C., Page, D.L., Abidi, M.A.: Object-based place recognition and loop closing with jigsaw puzzle image segmentation algorithm. In: 2008 IEEE International Conference on Robotics and Automation. pp. 557-562. IEEE (2008)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 150, + 481, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 150, + 481, + 172 + ], + "spans": [ + { + "bbox": [ + 133, + 150, + 481, + 172 + ], + "type": "text", + "content": "16. Cummins, M., Newman, P.: Appearance-only slam at large scale with fab-map 2.0. The International Journal of Robotics Research 30(9), 1100-1123 (2011)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 173, + 481, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 173, + 481, + 205 + ], + "spans": [ + { + "bbox": [ + 133, + 173, + 481, + 205 + ], + "type": "text", + "content": "17. Cupec, R., Nyarko, E.K., Filko, D., Kitanov, A., Petrovic, I.: Place recognition based on matching of planar surfaces and line segments. The International Journal of Robotics Research 34(4-5), 674-704 (2015)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 206, + 481, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 206, + 481, + 228 + ], + "spans": [ + { + "bbox": [ + 132, + 206, + 481, + 228 + ], + "type": "text", + "content": "18. Folorunsho, S.O.: Semantic segmentation-based approach for autonomous navigation in challenging farm terrains. algorithms 15, 3 (2024)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 229, + 481, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 229, + 481, + 251 + ], + "spans": [ + { + "bbox": [ + 132, + 229, + 481, + 251 + ], + "type": "text", + "content": "19. Garg, S., Fischer, T., Milford, M.: Where is your place, visual place recognition? In: IJCAI (2021)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 251, + 481, + 284 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 251, + 481, + 284 + ], + "spans": [ + { + "bbox": [ + 132, + 251, + 481, + 284 + ], + "type": "text", + "content": "20. Garg, S., Rana, K., Hosseinzadeh, M., Mares, L., Suenderhauf, N., Dayoub, F., Reid, I.: Robohop: Segment-based topological map representation for open-world visual navigation. In: 2024 IEEE International Conference on Robotics and Automation (ICRA) (2024)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 285, + 481, + 317 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 285, + 481, + 317 + ], + "spans": [ + { + "bbox": [ + 132, + 285, + 481, + 317 + ], + "type": "text", + "content": "21. Garg, S., Suenderhauf, N., Milford, M.: Lost? appearance-invariant place recognition for opposite viewpoints using visual semantics. In: Proceedings of Robotics: Science and Systems XIV (2018)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 318, + 481, + 372 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 318, + 481, + 372 + ], + "spans": [ + { + "bbox": [ + 132, + 318, + 481, + 372 + ], + "type": "text", + "content": "22. Garg, S., Sunderhauf, N., Dayoub, F., Morrison, D., Cosgun, A., Carneiro, G., Wu, Q., Chin, T.J., Reid, I., Gould, S., Corke, P., Milford, M.: Semantics for robotic mapping, perception and interaction: A survey. Foundations and Trends® in Robotics 8(1-2), 1-224 (2020). https://doi.org/10.1561/2300000059, http://dx.doi.org/10.1561/2300000059" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 373, + 481, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 373, + 481, + 396 + ], + "spans": [ + { + "bbox": [ + 132, + 373, + 481, + 396 + ], + "type": "text", + "content": "23. Gawel, A., Del Don, C., Siegwart, R., Nieto, J., Cadena, C.: X-view: Graph-based semantic multi-view localization. IEEE Robotics and Automation Letters 3(3), 1687-1694 (2018)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 396, + 481, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 396, + 481, + 430 + ], + "spans": [ + { + "bbox": [ + 132, + 396, + 481, + 430 + ], + "type": "text", + "content": "24. Gu, Q., Kuwajerwala, A., Morin, S., Jatavallabhula, K.M., Sen, B., Agarwal, A., Rivera, C., Paul, W., Ellis, K., Chellappa, R., et al.: Conceptgraphs: Open-vocabulary 3d scene graphs for perception and planning. arXiv preprint arXiv:2309.16650 (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 430, + 481, + 463 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 430, + 481, + 463 + ], + "spans": [ + { + "bbox": [ + 132, + 430, + 481, + 463 + ], + "type": "text", + "content": "25. Hausler, S., Garg, S., Xu, M., Milford, M., Fischer, T.: Patch-netvlad: Multi-scale fusion of locally-global descriptors for place recognition. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14141-14152 (2021)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 464, + 481, + 496 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 464, + 481, + 496 + ], + "spans": [ + { + "bbox": [ + 132, + 464, + 481, + 496 + ], + "type": "text", + "content": "26. Hu, H., Qiao, Z., Cheng, M., Liu, Z., Wang, H.: Dasgil: Domain adaptation for semantic and geometric-aware image-based localization. IEEE Transactions on Image Processing 30, 1342-1353 (2020)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 497, + 481, + 518 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 497, + 481, + 518 + ], + "spans": [ + { + "bbox": [ + 132, + 497, + 481, + 518 + ], + "type": "text", + "content": "27. Ibrahimi, S., Van Noord, N., Alpherts, T., Worring, M.: Inside out visual place recognition. arXiv preprint arXiv:2111.13546 (2021)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 520, + 481, + 540 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 520, + 481, + 540 + ], + "spans": [ + { + "bbox": [ + 132, + 520, + 481, + 540 + ], + "type": "text", + "content": "28. Intraub, H.: The representation of visual scenes. Trends in cognitive sciences 1(6), 217-222 (1997)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 542, + 474, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 542, + 474, + 553 + ], + "spans": [ + { + "bbox": [ + 132, + 542, + 474, + 553 + ], + "type": "text", + "content": "29. Izquierdo, S., Civera, J.: Optimal transport aggregation for visual place recognition (2023)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 554, + 481, + 575 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 554, + 481, + 575 + ], + "spans": [ + { + "bbox": [ + 132, + 554, + 481, + 575 + ], + "type": "text", + "content": "30. Jégou, H., Douze, M., Schmid, C., Pérez, P.: Aggregating local descriptors into a compact image representation. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 3304-3311 (2010)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 576, + 481, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 576, + 481, + 597 + ], + "spans": [ + { + "bbox": [ + 132, + 576, + 481, + 597 + ], + "type": "text", + "content": "31. Kassab, C., Mattamala, M., Fallon, M.: Clip-based features achieve competitive zero-shot visual localization. OpenReview preprint arXiv:2306.14846 (2023)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 599, + 481, + 631 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 599, + 481, + 631 + ], + "spans": [ + { + "bbox": [ + 132, + 599, + 481, + 631 + ], + "type": "text", + "content": "32. Keetha, N., Mishra, A., Karhade, J., Jatavallabhula, K.M., Scherer, S., Krishna, M., Garg, S.: Anyloc: Towards universal visual place recognition. IEEE Robotics and Automation Letters (2023)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 132, + 632, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 481, + 665 + ], + "type": "text", + "content": "33. Keetha, N.V., Milford, M., Garg, S.: A hierarchical dual model of environment-and place-specific utility for visual place recognition. IEEE Robotics and Automation Letters 6(4), 6969-6976 (2021)" + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "type": "text", + "content": "K. Garg et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 665 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "text", + "content": "34. Khaliq, A., Ehsan, S., Milford, M., McDonald-Maier, K.: A holistic visual place recognition approach using lightweight cnns for severe viewpoint and appearance changes. arXiv preprint arXiv:1811.03032 (2018)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 149, + 482, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 149, + 482, + 182 + ], + "spans": [ + { + "bbox": [ + 130, + 149, + 482, + 182 + ], + "type": "text", + "content": "35. Khaliq, A., Xu, M., Hausler, S., Milford, M., Garg, S.: Vlad-buff: Burst-aware fast feature aggregation for visual place recognition. In: European Conference on Computer Vision. Springer (2024)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 182, + 482, + 214 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 182, + 482, + 214 + ], + "spans": [ + { + "bbox": [ + 130, + 182, + 482, + 214 + ], + "type": "text", + "content": "36. Khorasgani, S.H., Chen, Y., Shkurti, F.: Slic: Self-supervised learning with iterative clustering for human action videos. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16091-16101 (2022)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 214, + 482, + 236 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 214, + 482, + 236 + ], + "spans": [ + { + "bbox": [ + 130, + 214, + 482, + 236 + ], + "type": "text", + "content": "37. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., et al.: Segment anything. arXiv preprint arXiv:2304.02643 (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 236, + 482, + 257 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 236, + 482, + 257 + ], + "spans": [ + { + "bbox": [ + 130, + 236, + 482, + 257 + ], + "type": "text", + "content": "38. Knopp, J., Sivic, J., Pajdla, T.: Avoiding confusing features in place recognition. In: European Conference on Computer Vision. pp. 748-761. Springer (2010)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 257, + 482, + 289 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 257, + 482, + 289 + ], + "spans": [ + { + "bbox": [ + 130, + 257, + 482, + 289 + ], + "type": "text", + "content": "39. Krantz, J., Lee, S., Malik, J., Batra, D., Chaplot, D.S.: Instance-specific image goal navigation: Training embodied agents to find object instances. arXiv preprint arXiv:2211.15876 (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 289, + 482, + 311 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 289, + 482, + 311 + ], + "spans": [ + { + "bbox": [ + 130, + 289, + 482, + 311 + ], + "type": "text", + "content": "40. Le, D.C., Youn, C.H.: City-scale visual place recognition with deep local features based on multi-scale ordered VLAD pooling. arXiv preprint arXiv:2009.09255 (2020)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 311, + 482, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 311, + 482, + 331 + ], + "spans": [ + { + "bbox": [ + 130, + 311, + 482, + 331 + ], + "type": "text", + "content": "41. Lowe, D.G.: Distinctive image features from scale-invariant keypoints. International journal of computer vision 60(2), 91-110 (2004)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 332, + 482, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 332, + 482, + 364 + ], + "spans": [ + { + "bbox": [ + 130, + 332, + 482, + 364 + ], + "type": "text", + "content": "42. Maalouf, A., Jadhav, N., Jatavallabhula, K.M., Chahine, M., Vogt, D.M., Wood, R.J., Torralba, A., Rus, D.: Follow anything: Open-set detection, tracking, and following in real-time. IEEE Robotics and Automation Letters 9(4), 3283-3290 (2024)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 364, + 482, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 364, + 482, + 385 + ], + "spans": [ + { + "bbox": [ + 130, + 364, + 482, + 385 + ], + "type": "text", + "content": "43. Manglani, S.: Real-time vision-based navigation for a robot in an indoor environment. arXiv preprint arXiv:2307.00666 (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 385, + 482, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 385, + 482, + 407 + ], + "spans": [ + { + "bbox": [ + 130, + 385, + 482, + 407 + ], + "type": "text", + "content": "44. Masone, C., Caputo, B.: A survey on deep visual place recognition. IEEE Access 9, 19516-19547 (2021)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 407, + 482, + 438 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 407, + 482, + 438 + ], + "spans": [ + { + "bbox": [ + 130, + 407, + 482, + 438 + ], + "type": "text", + "content": "45. Mirjalili, R., Krawez, M., Burgard, W.: Fm-loc: Using foundation models for improved vision-based localization. In: Robotics and Automation (ICRA), 2023 IEEE International Conference on. IEEE (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 438, + 482, + 460 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 438, + 482, + 460 + ], + "spans": [ + { + "bbox": [ + 130, + 438, + 482, + 460 + ], + "type": "text", + "content": "46. Mousavian, A., Kosecka, J.: Semantic image based geolocation given a map (author's initial manuscript). Tech. rep., George Mason University Fairfax United States (2016)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 460, + 482, + 492 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 460, + 482, + 492 + ], + "spans": [ + { + "bbox": [ + 130, + 460, + 482, + 492 + ], + "type": "text", + "content": "47. Naseer, T., Oliveira, G.L., Brox, T., Burgard, W.: Semantics-aware visual localization under challenging perceptual conditions. In: IEEE International Conference on Robotics and Automation (ICRA) (2017)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 492, + 440, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 492, + 440, + 503 + ], + "spans": [ + { + "bbox": [ + 130, + 492, + 440, + 503 + ], + "type": "text", + "content": "48. Oliva, A.: Gist of the scene. Neurobiology of attention 696(64), 251-258 (2005)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 130, + 503, + 482, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 503, + 482, + 557 + ], + "spans": [ + { + "bbox": [ + 130, + 503, + 482, + 557 + ], + "type": "text", + "content": "49. Oquab, M., Darcet, T., Moutakanni, T., Vo, H.V., Szafraniec, M., Khalidov, V., Fernandez, P., Haziza, D., Massa, F., El-Nouby, A., Howes, R., Huang, P.Y., Xu, H., Sharma, V., Li, S.W., Galuba, W., Rabbat, M., Assran, M., Ballas, N., Synnaeve, G., Misra, I., Jegou, H., Mairal, J., Labatut, P., Joulin, A., Bojanowski, P.: Dinov2: Learning robust visual features without supervision (2023)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 130, + 557, + 482, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 557, + 482, + 601 + ], + "spans": [ + { + "bbox": [ + 130, + 557, + 482, + 601 + ], + "type": "text", + "content": "50. Paolicelli, V., Tavera, A., Masone, C., Berton, G., Caputo, B.: Learning semantics for visual place recognition through multi-scale attention. In: Image Analysis and Processing-ICIAP 2022: 21st International Conference, Lecce, Italy, May 23-27, 2022, Proceedings, Part II. pp. 454-466. Springer (2022)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 130, + 601, + 482, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 601, + 482, + 632 + ], + "spans": [ + { + "bbox": [ + 130, + 601, + 482, + 632 + ], + "type": "text", + "content": "51. Pion, N., Humenberger, M., Csurka, G., Cabon, Y., Sattler, T.: Benchmarking image retrieval for visual localization. In: 2020 International Conference on 3D Vision (3DV). pp. 483-494. IEEE (2020)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 130, + 632, + 482, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 632, + 482, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 632, + 482, + 665 + ], + "type": "text", + "content": "52. Puligilla, S.S., Tourani, S., Vaidya, T., Parihar, U.S., Sarvadevabhatla, R.K., Krishna, K.M.: Topological mapping for manhattan-like repetitive environments. In: 2020 IEEE International Conference on Robotics and Automation (ICRA). pp. 6268-6274. IEEE (2020)" + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 384, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 384, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 384, + 91, + 447, + 102 + ], + "type": "text", + "content": "Revisit Anything" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 117, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 133, + 117, + 480, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 117, + 480, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 117, + 480, + 149 + ], + "type": "text", + "content": "53. Radenović, F., Iscen, A., Tobias, G., Avrithis, Y., Chum, O.: Revisiting oxford and paris: Large-scale image retrieval benchmarking. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 5706-5715 (2018)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 150, + 480, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 150, + 480, + 171 + ], + "spans": [ + { + "bbox": [ + 132, + 150, + 480, + 171 + ], + "type": "text", + "content": "54. Radenovic, F., Tobias, G., Chum, O.: Fine-tuning cnn image retrieval with no human annotation. IEEE Trans. Pattern Anal. Mach. Intell. 41(7), 1655-1668 (2018)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 172, + 480, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 172, + 480, + 205 + ], + "spans": [ + { + "bbox": [ + 132, + 172, + 480, + 205 + ], + "type": "text", + "content": "55. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748-8763. PMLR (2021)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 205, + 480, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 205, + 480, + 237 + ], + "spans": [ + { + "bbox": [ + 132, + 205, + 480, + 237 + ], + "type": "text", + "content": "56. Revaud, J., Almazán, J., Rezende, R.S., Souza, C.R.d.: Learning with average precision: Training image retrieval with a listwise loss. In: Proceedings of the IEEE International Conference on Computer Vision. pp. 5107-5116 (2019)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 238, + 480, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 238, + 480, + 258 + ], + "spans": [ + { + "bbox": [ + 132, + 238, + 480, + 258 + ], + "type": "text", + "content": "57. Sarlin, P.E., Cadena, C., Siegwart, R., Dymczyk, M.: From coarse to fine: Robust hierarchical localization at large scale. arXiv preprint arXiv:1812.03506 (2018)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 259, + 480, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 259, + 480, + 293 + ], + "spans": [ + { + "bbox": [ + 132, + 259, + 480, + 293 + ], + "type": "text", + "content": "58. Sattler, T., Maddern, W., Toft, C., Torii, A., Hammarstrand, L., Stenberg, E., Safari, D., Okutomi, M., M. P., J. S., F. K., T. P.: Benchmarking 6dof outdoor visual localization in changing conditions. In: Proc. CVPR (2018)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 293, + 480, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 293, + 480, + 314 + ], + "spans": [ + { + "bbox": [ + 132, + 293, + 480, + 314 + ], + "type": "text", + "content": "59. Schleiss, M., Rouatbi, F., Cremers, D.: Vpair - aerial visual place recognition and localization in large-scale outdoor environments (2022)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 315, + 480, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 315, + 480, + 335 + ], + "spans": [ + { + "bbox": [ + 132, + 315, + 480, + 335 + ], + "type": "text", + "content": "60. Schubert, S., Neubert, P., Garg, S., Milford, M., Fischer, T.: Visual place recognition: A tutorial. RAM (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 336, + 480, + 368 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 336, + 480, + 368 + ], + "spans": [ + { + "bbox": [ + 132, + 336, + 480, + 368 + ], + "type": "text", + "content": "61. Shlapentokh-Rothman, M., Blume, A., Xiao, Y., Wu, Y., TV, S., Tao, H., Lee, J.Y., Torres, W., Wang, Y.X., Hoiem, D.: Region-based representations revisited. arXiv preprint arXiv:2402.02352 (2024)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 369, + 480, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 369, + 480, + 402 + ], + "spans": [ + { + "bbox": [ + 132, + 369, + 480, + 402 + ], + "type": "text", + "content": "62. Shubodh, S., Omama, M., Zaidi, H., Parihar, U.S., Krishna, M.: Lip-loc: Lidar image pretraining for cross-modal localization. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 948-957 (2024)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 403, + 480, + 434 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 403, + 480, + 434 + ], + "spans": [ + { + "bbox": [ + 132, + 403, + 480, + 434 + ], + "type": "text", + "content": "63. Sivic, J., Zisserman, A.: Video google: A text retrieval approach to object matching in videos. In: Proceedings of International Conference on Computer Vision (ICCV). p. 1470. IEEE (2003)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 435, + 480, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 435, + 480, + 468 + ], + "spans": [ + { + "bbox": [ + 132, + 435, + 480, + 468 + ], + "type": "text", + "content": "64. Sun, X., Xie, Y., Luo, P., Wang, L.: A dataset for benchmarking image-based localization. 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR) pp. 5641-5649 (2017), https://apisemantic scholar.org/CorpusID:20531893" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 468, + 480, + 500 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 468, + 480, + 500 + ], + "spans": [ + { + "bbox": [ + 132, + 468, + 480, + 500 + ], + "type": "text", + "content": "65. Sünderhauf, N., Shirazi, S., Jacobson, A., Dayoub, F., Pepperell, E., Upcroft, B., Milford, M.: Place recognition with convnet landmarks: Viewpoint-robust, condition-robust, training-free. Proceedings of Robotics: Science and Systems XII (2015)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 501, + 480, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 501, + 480, + 533 + ], + "spans": [ + { + "bbox": [ + 132, + 501, + 480, + 533 + ], + "type": "text", + "content": "66. Torii, A., Sivic, J., Pajdla, T., Okutomi, M.: Visual place recognition with repetitive structures. In: 2013 IEEE Conference on Computer Vision and Pattern Recognition. pp. 883-890 (2013). https://doi.org/10.1109/CVPR.2013.119" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 534, + 480, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 534, + 480, + 566 + ], + "spans": [ + { + "bbox": [ + 132, + 534, + 480, + 566 + ], + "type": "text", + "content": "67. Tsintotas, K.A., Bampis, L., Gasteratos, A.: The revisiting problem in simultaneous localization and mapping: A survey on visual loop closure detection. IEEE Transactions on Intelligent Transportation Systems 23(11), 19929-19953 (2022)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 567, + 480, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 567, + 480, + 599 + ], + "spans": [ + { + "bbox": [ + 132, + 567, + 480, + 599 + ], + "type": "text", + "content": "68. Wang, R., Shen, Y., Zuo, W., Zhou, S., Zheng, N.: Transvpr: Transformer-based place recognition with multi-level attention aggregation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13648-13657 (2022)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 600, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 600, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 600, + 480, + 632 + ], + "type": "text", + "content": "69. Warburg, F., Hauberg, S., López-Antequera, M., Gargallo, P., Kuang, Y., Civera, J.: Mapillary street-level sequences: A dataset for lifelong place recognition. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2626-2635 (2020)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 480, + 665 + ], + "type": "text", + "content": "70. Weinzaepfel, P., Csurka, G., Cabon, Y., Humenberger, M.: Visual localization by learning objects-of-interest dense match regression. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5634-5643 (2019)" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 217, + 102 + ], + "type": "text", + "content": "K. Garg et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 369 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "text", + "content": "71. Xin, Z., Cai, Y., Lu, T., Xing, X., Cai, S., Zhang, J., Yang, Y., Wang, Y.: Localizing discriminative visual landmarks for place recognition. In: 2019 International conference on robotics and automation (ICRA). pp. 5979-5985. IEEE (2019)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 150, + 482, + 204 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 150, + 482, + 204 + ], + "spans": [ + { + "bbox": [ + 130, + 150, + 482, + 204 + ], + "type": "text", + "content": "72. Yildiz, B., Khademi, S., Siebes, R.M., Van Gemert, J.: Amstertime: A visual place recognition benchmark dataset for severe domain shift. In: 2022 26th International Conference on Pattern Recognition (ICPR). IEEE (Aug 2022). https://doi.org/10.1109/icpr56361.2022.9956049, http://dx.doi.org/10.1109/ICPR56361.2022.9956049" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 205, + 482, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 205, + 482, + 237 + ], + "spans": [ + { + "bbox": [ + 130, + 205, + 482, + 237 + ], + "type": "text", + "content": "73. Yu, J., Zhu, C., Zhang, J., Huang, Q., Tao, D.: Spatial pyramid-enhanced netvlad with weighted triplet loss for place recognition. IEEE Trans. Neural Netw. Learn. Syst. 31(2), 661-674 (2019)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 237, + 482, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 237, + 482, + 281 + ], + "spans": [ + { + "bbox": [ + 130, + 237, + 482, + 281 + ], + "type": "text", + "content": "74. Zaffar, M., Garg, S., Milford, M., Kooij, J., Flynn, D., McDonald-Maier, K., Ehsan, S.: Vpr-bench: An open-source visual place recognition evaluation framework with quantifiable viewpoint and appearance change. International Journal of Computer Vision pp. 1-39 (2021)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 281, + 482, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 281, + 482, + 303 + ], + "spans": [ + { + "bbox": [ + 130, + 281, + 482, + 303 + ], + "type": "text", + "content": "75. Zhang, Y., Zhao, X.: Mesa: Matching everything by segmenting anything. arXiv preprint arXiv:2401.16741 (2024)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 304, + 482, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 304, + 482, + 335 + ], + "spans": [ + { + "bbox": [ + 130, + 304, + 482, + 335 + ], + "type": "text", + "content": "76. Zhou, B., Lapedriza, A., Khosla, A., Oliva, A., Torralba, A.: Places: A 10 million image database for scene recognition. IEEE Transactions on Pattern Analysis and Machine Intelligence (2017)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 335, + 482, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 335, + 482, + 369 + ], + "spans": [ + { + "bbox": [ + 130, + 335, + 482, + 369 + ], + "type": "text", + "content": "77. Zou, X., Yang, J., Zhang, H., Li, F., Li, L., Wang, J., Wang, L., Gao, J., Lee, Y.J.: Segment everything everywhere all at once. Advances in Neural Information Processing Systems 36 (2024)" + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 384, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 384, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 384, + 91, + 447, + 102 + ], + "type": "text", + "content": "Revisit Anything" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 471, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 471, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 471, + 91, + 480, + 100 + ], + "type": "text", + "content": "19" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 18 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Revisit Event Generation Model_ Self-Supervised Learning of Event-to-Video Reconstruction with Implicit Neural Representations/a41e408e-cbb3-4c38-85c3-298d71f3cf9d_content_list.json b/2024/Revisit Event Generation Model_ Self-Supervised Learning of Event-to-Video Reconstruction with Implicit Neural Representations/a41e408e-cbb3-4c38-85c3-298d71f3cf9d_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..a7f70f452016d73c636f057f01f62ebbcbd2431c --- /dev/null +++ b/2024/Revisit Event Generation Model_ Self-Supervised Learning of Event-to-Video Reconstruction with Implicit Neural Representations/a41e408e-cbb3-4c38-85c3-298d71f3cf9d_content_list.json @@ -0,0 +1,1910 @@ +[ + { + "type": "text", + "text": "Revisit Event Generation Model: Self-Supervised Learning of Event-to-Video Reconstruction with Implicit Neural Representations", + "text_level": 1, + "bbox": [ + 222, + 140, + 781, + 209 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zipeng Wang1, Yunfan Lu1, and Lin Wang1,2", + "bbox": [ + 315, + 234, + 686, + 251 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Artificial Intelligence Thrust, HKUST(GZ)", + "bbox": [ + 349, + 261, + 651, + 275 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ Dept. of Computer Science and Engineering, HKUST", + "bbox": [ + 315, + 275, + 686, + 290 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{zwang253,ylu066}@connect.hkust-gz.edu.cn, linwang@ust.hk", + "bbox": [ + 281, + 291, + 720, + 304 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Reconstructing intensity frames from event data while maintaining high temporal resolution and dynamic range is crucial for bridging the gap between event-based and frame-based computer vision. Previous approaches have depended on supervised learning on synthetic data, which lacks interpretability and risk over-fitting to the setting of the event simulator. Recently, self-supervised learning (SSL) based methods, which primarily utilize per-frame optical flow to estimate intensity via photometric constancy, has been actively investigated. However, they are vulnerable to errors in the case of inaccurate optical flow. This paper proposes a novel SSL event-to-video reconstruction approach, dubbed EvINR, which eliminates the need for labeled data or optical flow estimation. Our core idea is to reconstruct intensity frames by directly addressing the event generation model, essentially a partial differential equation (PDE) that describes how events are generated based on the time-varying brightness signals. Specifically, we utilize an implicit neural representation (INR), which takes in spatiotemporal coordinate $(x,y,t)$ and predicts intensity values, to represent the solution of the event generation equation. The INR, parameterized as a fully-connected Multi-layer Perceptron (MLP), can be optimized with its temporal derivatives supervised by events. To make EvINR feasible for online requisites, we propose several acceleration techniques that substantially expedite the training process. Comprehensive experiments demonstrate that our EvINR surpasses previous SSL methods by $38\\%$ w.r.t. Mean Squared Error (MSE) and is comparable or superior to SoTA supervised methods. Project page: https://vlislab22.github.io/EvINR/.", + "bbox": [ + 261, + 339, + 743, + 686 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 710, + 375, + 726 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Event cameras [13,55] are novel sensors that offer numerous advantages over traditional frame-based cameras, including low power consumption, high dynamic range (HDR), and high temporal resolution [34]. However, their unique imaging paradigm presents a challenge when applying vision algorithms designed for frame-based cameras. To address this challenge and bridge the gap between", + "bbox": [ + 212, + 741, + 787, + 818 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Corresponding author", + "bbox": [ + 218, + 824, + 382, + 839 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/9271eebe877c259690f453580b09094e8509cd9bbac6be9b1733ab51887f8a91.jpg", + "image_caption": [ + "Event Generation Model", + "Fig. 1: Connection between event generation model and EvINR: Event generation model reveals the relation between discrete events and continuous temporal intensity changes, described as the event generation equation (Eq. 4). EvINR utilizes an INR to solve Eq. 4 and recovering a continuous function of intensity w.r.t. time, implicitly parameterized with a fully connected MLP." + ], + "image_footnote": [], + "bbox": [ + 256, + 164, + 743, + 335 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "event-based and standard computer vision [10, 15, 17, 18, 36], many methods have been proposed to reconstruct intensity frames from events.", + "bbox": [ + 212, + 455, + 784, + 484 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Early methods primarily rely on hand-crafted integrators or filters, resulting in significant loss of detail in reconstructed results [3,22,28,40,59]. More recently, deep learning-based methods [6,35,41,49,53,64] have emerged, typically supervised by large-scale synthetic datasets generated using an event simulator [14,19]. However, the interpretability of these methods is limited due to the 'black-box' nature of the neural networks. Furthermore, the performance of these methods is constrained by the domain gap between synthetic and real-world data [49], leading to suboptimal results if the event simulator's settings do not accurately match those of the inference data. To address these issues, some researchers have explored self-supervised learning (SSL) frameworks [31,60], aiming to eliminate the dependency on labeled synthetic data. Nonetheless, these approaches still depend on event-based optical flow estimation, which is prone to over-fitting, occlusions, and non-convergence issues [44, 45]. As a result, the reconstructed frames often suffer from the loss of textural details and various artifacts.", + "bbox": [ + 212, + 489, + 787, + 700 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address these challenges, we revisit the event generation model [13], which forms a fundamental link between events and intensity. The event generation model can be expressed as a partial differential equation (PDE), describing how events are triggered by logarithmic intensity changes that exceed a certain threshold. This PDE, known as the event generation equation, establishes a direct connection between discrete events and the partial derivative of the continuous intensity function w.r.t. time. Our key insight is that, solving the event generation equation offers an ideal self-supervised solution for event-to-video reconstruction, thereby eliminating the need for synthetic data or optical", + "bbox": [ + 212, + 704, + 787, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "Z. Wang et al.", + "bbox": [ + 271, + 114, + 369, + 128 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "flow estimation. However, several properties of event data make solving the equation nontrivial: 1) events can be triggered at extremely high frequencies (up to $10^{8} / s$ ), which results in a large volume of data and a heavy computational burden. 2) events are inherently noisy, particularly in extreme visual conditions [51], which pose challenge to the robustness of solvers. 3) events do not capture the initial intensity, which makes determining the boundary values challenging.", + "bbox": [ + 212, + 146, + 787, + 236 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Recently, implicit neural representations (INRs) have gained popularity for solving the inverse problems in 3D reconstruction or image super-resolution [8, 26,46] by parameterizing complex signals via deep neural networks. In this work, we find that INRs possess several key advantages that render them particularly suitable for solving the event generation equation: they inherently accommodate a large volume of event data, exhibit high noise tolerance, and are flexible to add additional loss terms to regulate initial value.", + "bbox": [ + 212, + 237, + 787, + 343 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In light of this, we propose a novel SSL framework, termed EvINR, that employs an INR to represent the solution of the event generation equation. Our EvINR can be directly optimized by minimizing the residual between its temporal derivatives and local intensity changes estimated from event data based on the event generation model (Sec. 3.2). Moreover, we incorporate a spatial regularization term that regularizes the relative values of adjacent pixels by constraining the magnitude of spatial gradients, which effectively reduces noise in the reconstruction process (Sec. 3.2). Although the basic implementation of EvINR yields acceptable results, its convergence on seconds-long event sequences takes minutes, limiting its real-world applicability, especially in online scenarios. To expedite EvINR's training process, we introduce several acceleration techniques, including frame-based optimization, coarse-to-fine training, and model ensembling. These approaches reduce the training time from minutes to seconds while not compromising the reconstruction quality (Sec. 3.4).", + "bbox": [ + 212, + 343, + 787, + 554 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Moreover, most approaches are typically evaluated on event datasets captured by DAVIS sensors [27, 42, 49, 61], making it difficult to evaluate their stability and robustness to other types of event camera [1, 2, 39]. For this reason, we collected a new event dataset using an ALPIX-Eiger event camera [1], with well-aligned events and intensity frames.", + "bbox": [ + 212, + 555, + 787, + 630 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In summary, our paper makes three key contributions: (I) We propose EvINR, a concise SSL framework that solves the event generation equation via implicit neural representations for event-to-video reconstruction. (II) Our EvINR substantially outperforms previous SSL methods [31, 40, 59] and attains comparable, or even superior, performance compared to state-of-the-art supervised methods [35, 49, 53]. (III) We collect a real-world dataset with an ALPIX-Eiger event camera, complementary to the datasets captured by DAVIS cameras.", + "bbox": [ + 212, + 631, + 787, + 736 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Works", + "text_level": 1, + "bbox": [ + 214, + 761, + 397, + 777 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Implicit Neural Representations (INRs) have emerged as a powerful tool for parameterizing signals, such as images, videos, and audio, in a continuous manner using neural networks [46]. Compared to traditional discrete sig-", + "bbox": [ + 212, + 794, + 787, + 840 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "EvINR", + "bbox": [ + 681, + 114, + 730, + 126 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 114, + 785, + 126 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/8b29df90046efc81ca4d72191edd630d2ab8ec1e1cd441d98a91e34a3d772a8e.jpg", + "image_caption": [ + "Fig. 2: Overview of EvINR. A fully connected MLP is used to implicitly solve the event generation equation. The temporal gradient of the MLP is supervised by temporal intensity changes of events, and the spatial gradient is penalized to reduce noise." + ], + "image_footnote": [], + "bbox": [ + 223, + 148, + 781, + 292 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "nal representations, INRs offer the advantage of being able to be sampled at arbitrary resolutions with fixed memory requirements. As a result, INRs have found widespread applications in various fields, including 3D scene representation [25, 26, 32], video representation [7], generative models [29, 43, 47], and model compression [50]. A unique property of INRs is that they can be effectively learned from the derivatives of signals, such as the normals of 3D shapes [30,46,58] and the gradients of images [46]. Such a characteristic has motivated us to approach the task of event-to-video reconstruction by optimizing the INR of video from its temporal derivatives. Previous research has investigated the potential of INR for novel view synthesis using event data [20,24,38]. These approaches reconstruct 3D neural radiance fields using multiple event sequences with known camera poses from a stationary scene. Most similar to our work, E-CIR [48] uses polynomial to regress the intensity function, which may only represent a short time interval (e.g., the exposure time of an image). In contrast, our method employs INR, which can represent longer time of intensity change.", + "bbox": [ + 212, + 382, + 787, + 609 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Event-based Video Reconstruction has been a hot topic in the literature. Early attempts [3, 9, 21] address this problem based on the photometric constancy, which describes the relationship between intensity gradients and optical flow. Other approaches [28, 40] are based on direct event integration without estimating optical flow. Rebecq et al. [35] developed the first DL-based framework, called E2VID, to reconstruct intensity frames from events in an end-to-end manner, outperforming earlier techniques by a significant margin. E2VID was updated by some following research, trying to tackle problems of inference speed [41], the cold start problem [6], and training strategy [49]. These methods are supervised and learned using the synthetic dataset obtained via the event simulators [14, 19] due to the lack of real-world datasets with well-aligned events and intensity frames. Therefore, the generalization capability of these methods is limited by the simulator-to-real gap [49].", + "bbox": [ + 212, + 611, + 787, + 808 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To address this problem, Federico et al. [31] proposed an SSL framework with a network to estimate optical flow and another network to reconstruct intensity", + "bbox": [ + 212, + 809, + 785, + 839 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "Z. Wang et al.", + "bbox": [ + 271, + 114, + 367, + 128 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "frames based on the photometric constancy. Zhang et al. [60] updated this idea into a linear inverse problem that can be solved using modern linear solvers without using deep learning. However, these methods either assume optical flow is known or estimate the optical flow using contrast maximization [62,63]. Consequently, their performance cannot be guaranteed unless high-quality optical flow can be obtained. In contrast, our approach is based solely on the physical event generation model, leading to a more straightforward solution that also demonstrates significantly improved performance and greater flexibility.", + "bbox": [ + 214, + 146, + 787, + 268 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3 Method", + "text_level": 1, + "bbox": [ + 215, + 287, + 330, + 305 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Overview: Our objective is to reconstruct intensity frames from events in a self-supervised manner, without the need for end-to-end training or optical flow estimation. To accomplish this, we have reformulated event-based video reconstruction as solving the event generation equation. We employ an INR that is supervised solely by event data to represent the intensity function. An overview of our approach can be seen in Fig.2. In Sec. 3.1, we explain the event generation model and its connection to event-based video reconstruction. In Sec. 3.2, we detail how to train an EvINR by supervising its spatial and temporal gradients. Techniques to speed up the training of EvINR for online applications are discussed in Sec. 3.3. Finally, in Sec. 3.4, we describe our collected dataset using the ALPIX event camera [1].", + "bbox": [ + 212, + 319, + 787, + 486 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1 Preliminary: Event Generation Model", + "text_level": 1, + "bbox": [ + 214, + 506, + 576, + 520 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We begin by providing a brief overview of the event generation model [13], which forms the theoretical foundation for our approach. Let $I(x,y,t)$ denote the intensity of the spatial location $(x,y)$ at time $t$ in a video. Since event cameras operate with logarithmic intensity, we denote $L(x,y,t) = \\log I(x,y,t)$ .", + "bbox": [ + 212, + 527, + 785, + 589 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "An event camera comprises a frame of independent pixels that respond to changes in the logarithmic intensity signal and produce sequences of sparse and asynchronous events. An event $(x_{i},y_{i},t_{i},p_{i})$ is triggered when the logarithmic intensity change surpasses a threshold $C$ since the previous event was triggered at the same pixel. where $(x,y)$ is the spatial location of the pixel, $t$ is the timestamp, and $p\\in \\{-1,1\\}$ is the polarity of the logarithmic intensity change.", + "bbox": [ + 214, + 590, + 787, + 679 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For simplicity, let us consider the temporal changes in logarithmic intensity of a single pixel with fixed spatial coordinates and disregard the spatial terms $(x,y)$ . We can describe an event $e_i$ on that pixel using the Dirac delta function as follows:", + "bbox": [ + 214, + 680, + 787, + 739 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\ne _ {i} (t) = p _ {i} \\cdot C \\cdot \\delta \\left(t - t _ {i}\\right). \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 418, + 741, + 785, + 756 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Therefore, the logarithmic intensity increment $\\Delta L = L(t_2) - L(t_1)$ in a time interval $[t_1,t_2]$ can be represented by the accumulation of events, which can be expressed as:", + "bbox": [ + 214, + 761, + 787, + 806 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\Delta L = \\int_ {t _ {1}} ^ {t _ {2}} \\sum_ {i} e _ {i} (t) \\mathrm {d} t. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 424, + 806, + 785, + 842 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "EvINR", + "bbox": [ + 681, + 114, + 730, + 126 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Assuming a short time interval and ignoring noise, we can approximate the logarithmic intensity increment by its first-order temporal derivative using Taylor expansion:", + "bbox": [ + 212, + 146, + 782, + 191 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\frac {\\Delta L}{t _ {2} - t _ {1}} = \\frac {\\partial L ((t _ {1} + t _ {2}) / 2)}{\\partial t}. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 406, + 191, + 785, + 224 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "By substituting Eq. 2 into Eq. 3, we derive the event generation equation (Eq. 4), which bridges the discrete event data with the continuous temporal derivatives of logarithmic intensity:", + "bbox": [ + 214, + 231, + 782, + 277 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\frac {\\partial L \\left(\\left(t _ {1} + t _ {2}\\right) / 2\\right)}{\\partial t} = \\frac {1}{t _ {2} - t _ {1}} \\int_ {t _ {1}} ^ {t _ {2}} \\sum_ {i} e _ {i} (t) \\mathrm {d} t. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 354, + 290, + 785, + 327 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The left-hand side of Eq. 4 represents the partial derivative of logarithmic intensity with respect to time, while the right-hand side can be calculated by accumulating events. Intuitively, event-based video reconstruction can be formulated as solving Eq. 4, i.e., finding a logarithmic intensity function that satisfies the equation. However, as stated earlier, solving Eq. 4 can be non-trivial due to the large amount, high noise rate of event data, and unknown boundary intensity values. To address these challenges, we employ INRs to solve Eq. 4 as they naturally scale up to large data and have high noise tolerance [20]. Boundary intensity values can also easily be regularized by injecting natural image priors as additional loss functions.", + "bbox": [ + 212, + 339, + 784, + 489 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.2 Learning INRs from Events", + "text_level": 1, + "bbox": [ + 215, + 517, + 490, + 532 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We aim to learn an INR $F_{\\Theta}$ given an event stream $\\{e_i\\}_{i=0}^N$ . Here, $\\Theta$ represents the parameters of a fully connected MLP that predicts the logarithmic intensity value $\\hat{L}$ at any spatiotemporal coordinate $(x,y,t)$ . We adapt the event frame representation [35] and stack the given events uniformly into $T$ event frames.", + "bbox": [ + 212, + 545, + 782, + 608 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Temporal Supervision The INR is optimized by directly minimizing the temporal loss between the predicted logarithmic intensity change $\\Delta \\hat{L}$ and the logarithmic intensity change estimated by event accumulation $\\Delta L$ . Here $\\Delta \\hat{L}$ can be derived by accumulating events as Eq. 2 and $\\Delta \\hat{L}$ is the change of $\\hat{L}$ with respect to $t$ obtained by performing double back-propagation [33] of $F_{\\Theta}$ :", + "bbox": [ + 212, + 633, + 782, + 709 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\Delta \\hat {L} = \\frac {\\partial F _ {\\Theta} \\left(\\left(t _ {1} + t _ {2}\\right) / 2\\right)}{\\partial t} \\cdot \\left(t _ {2} - t _ {1}\\right), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 380, + 723, + 785, + 752 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Here we adopt the mean squared error (MSE) loss for temporal supervision:", + "bbox": [ + 214, + 763, + 761, + 780 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {t e m p}} = \\left(\\Delta L - \\Delta \\hat {L}\\right) ^ {2}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 423, + 792, + 785, + 811 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\Delta L$ and $\\Delta \\hat{L}$ are given by Eq. 2 and Eq. 5, respectively.", + "bbox": [ + 214, + 823, + 658, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "Z. Wang et al.", + "bbox": [ + 271, + 114, + 367, + 128 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Spatial Regularization Although the temporal supervision can estimate the intensity function, the results may still contain unnatural artifacts as the INR network $F_{\\Theta}$ has no prior knowledge of the initial intensity values of each pixel. To address this issue, we introduce a spatial regularization term that encourages the solution to be in the space of natural-looking images. We adopt Tikhonov regularization [4] that penalizes the spatial gradients of logarithmic intensity:", + "bbox": [ + 214, + 146, + 782, + 237 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {r e g}} = \\left(\\frac {\\partial F _ {\\Theta}}{\\partial x}\\right) ^ {2} + \\left(\\frac {\\partial F _ {\\Theta}}{\\partial y}\\right) ^ {2}. \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 408, + 250, + 782, + 280 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "It is worth noting that we do not employ more complex regularization methods, such as CNN denoisers [56] as used in [60], because they typically utilize large network models that can significantly slow down the training process.", + "bbox": [ + 214, + 287, + 782, + 334 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Optimizing EvINR The overall objective is given by the equation:", + "bbox": [ + 215, + 356, + 712, + 371 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathcal {L} _ {\\text {t e m p}} + \\lambda \\mathcal {L} _ {\\text {r e g}}, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 433, + 385, + 782, + 398 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $\\mathcal{L}_{\\mathrm{temp}}$ and $\\mathcal{L}_{\\mathrm{reg}}$ were introduced in Sec. 3.2 and Sec. 3.2, respectively, as the temporal supervision and spatial regularization term. $\\lambda$ is a hyper-parameter used to adjust the weight of the spatial regularization term.", + "bbox": [ + 214, + 410, + 782, + 455 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Tone-mapping The output of the EvINR is the logarithmic intensity of the reconstructed frames. We first use the exponential function to convert the predicted logarithmic intensity values to high dynamic range (HDR) intensity values $I \\in [0, \\infty)$ :", + "bbox": [ + 214, + 478, + 782, + 537 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\nI (x, y, t) = \\exp \\left(F _ {\\Theta} (x, y, t)\\right). \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 403, + 539, + 782, + 553 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Then, we adopt the Reinhard function [37] to map the HDR intensity values into the low dynamic range $[0,1]$ :", + "bbox": [ + 214, + 561, + 782, + 590 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\Gamma (I) = \\left(\\frac {I}{I + 1}\\right) ^ {\\gamma}. \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 441, + 602, + 782, + 630 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $\\gamma$ is a hyper-parameter used to control the contrast of $\\Gamma(I)$ . We assess our reconstruction performance by applying $\\Gamma(I)$ in all experiments.", + "bbox": [ + 214, + 641, + 782, + 671 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.3 Accelerating EvINR", + "text_level": 1, + "bbox": [ + 215, + 694, + 428, + 709 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "While solving the event generation equation using the basic implementation of EvINR, as described in Sec. 3.3, yields satisfactory reconstruction results, it is not efficient enough for online tasks. Optimizing one INR network takes about 20 minutes, and a network can only represent approximately 1 second of a sequence. Reducing the training time or increasing the sequence time leads to severe performance degradation. To enable online usage for EvINR, we propose several techniques to reduce the training time and increase the representation capacity for EvINR, as illustrated in Fig. 3.", + "bbox": [ + 212, + 719, + 784, + 839 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "EvINR", + "bbox": [ + 681, + 114, + 730, + 126 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/0bce85b79b74465683d976941653301334345c5a6e7cba7e8b8909d6a51a5a9d.jpg", + "image_caption": [ + "Fig. 3: Overview of acceleration techniques. (a) and (b) illustrate the difference between the basic coordinate-based and our frame-based optimization, respectively. (c) depicts our proposed coarse-to-fine training scheme and network ensembling technique." + ], + "image_footnote": [], + "bbox": [ + 263, + 147, + 450, + 292 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/1657c628445946bdf53a0eef143f623a836bf6a1b9764428b9c3c28a195e7e2a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 455, + 148, + 743, + 292 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Coordinate-based to frame-based optimization: The basic EvINR algorithm is designed to learn a mapping from 3D coordinates $(x,y,t)$ to the logarithmic intensity $L$ . However, this coordinate-based optimization requires the network to 'remember' $H\\times W\\times T$ coordinate-to-logarithmic intensity mappings, which can be highly complex and lead to slow convergence. To address this issue, we propose a frame-based optimization scheme that learns a mapping from timestamps $t$ to the logarithmic intensity of all pixels at that timestamp. The frame-based optimization scheme can be represented as:", + "bbox": [ + 214, + 378, + 787, + 500 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\nF _ {\\Theta} (t) = \\left[ L (i, j, t) \\right] _ {1 \\leq i \\leq H, 1 \\leq j \\leq W}, \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 383, + 510, + 785, + 529 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where the right hand-side represents a matrix that contains all logarithmic intensity of all pixels at $t$ . The frame-based optimization approach requires the network to remember only $T$ coordinate-to-logarithmic intensity mappings, resulting in a significant reduction in complexity and faster convergence. Empirical results show that this approach reduces the convergence time by two orders of magnitude, from minutes to seconds. The comparison between coordinate-based and frame-based optimizations is illustrated in Fig. 3 (a) and (b).", + "bbox": [ + 214, + 537, + 787, + 643 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Coarse-to-fine training: We adopt a coarse-to-fine training scheme, which enables the network to learn overall logarithmic intensity changes before focusing on finer details. We structure the training process into $s$ distinct stages, at each of which we increase the temporal resolution by a factor of 2. Initially, we divide the event sequence into $N$ segments of equal length and optimize the EvINR in accordance with Eq. 6 over a specified number of iterations. Subsequently, we proceed to bisect each of these $N$ segments into two smaller ones, ensuring that they contain an equal number of events. For all the experiments, we set the number of stages $s$ to 3 and scheduled the upsampling to occur after 100 and 200 iterations, respectively. This approach accelerates training by reducing the number of event frames needed in the early stages by approximately 2 times.", + "bbox": [ + 214, + 643, + 787, + 809 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Network ensembling: As each EvINR network requires only about 1GB of GPU memory, we further leverage network ensemble techniques [16] to train $N$", + "bbox": [ + 214, + 809, + 787, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "Z. Wang et al.", + "bbox": [ + 271, + 114, + 369, + 128 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "EvINR networks simultaneously on a single GPU to achieve higher parallelism. This approach enables us to exploit the computational resources of modern GPUs more effectively and speed up the training process. We keep overlap periods between nearby networks to keep logarithmic intensity predictions constant among all networks. Fig. 3(c) depicts the coarse-to-fine training process and network ensembling techniques employed in our approach.", + "bbox": [ + 212, + 146, + 787, + 238 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "3.4 Dataset Collection", + "text_level": 1, + "bbox": [ + 214, + 258, + 415, + 272 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Most event-to-video reconstruction approaches are typically evaluated on event datasets collected using DAVIS sensors [5], such as IJRR [27], HQF [49], MVSEC [61], and CED [42]. Recently, other types of event cameras [1, 2, 39] have been developed, which share the same event generation model with DAVIS sensors but may differ in detailed configurations and settings. Therefore, it is essential to verify the generalization capability of existing methods and our EvINR on those new types of event sensors.", + "bbox": [ + 212, + 281, + 787, + 385 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "To fill the gap, we introduce a new real-world dataset, called the ALPIX Event Dataset (AED), which is collected using an ALPIX-Eiger event camera [1], featuring static scenes accompanied by gradual camera motions. The camera provides well-aligned RGB frames and color events. The RGB frames have a resolution of $3264 \\times 2448$ and the events have a resolution of $1632 \\times 1224$ . The AED dataset includes seven video sequences with diverse scenes, such as streets, buildings, indoor scenes, textures, and tools, and each approximately lasts for ten seconds. Note that, in this paper, we only focus on reconstructing grayscale frames to keep consistency with previous works [31, 60] and also for a fair comparison. Therefore, we first demosaic the RGB frames and events according to the Quad Bayer pattern, resulting in intensity frames with a resolution of $816 \\times 612$ and event data with a resolution of $408 \\times 306$ . Details on the post-processing of AED dataset can be found in the supplementary material.", + "bbox": [ + 212, + 387, + 787, + 583 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 214, + 604, + 375, + 622 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.1 Experiments Settings", + "text_level": 1, + "bbox": [ + 214, + 635, + 439, + 651 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Datasets: We conduct experiments to evaluate the effectiveness of our proposed method using three datasets: IJRR [27], HQF [49], and our AED dataset. The IJRR dataset consists of intensity frames and events in 25 real scenes and 2 synthetic scenes, captured by a DAVIS240C camera [5]. HQF dataset provides 14 event data sequences captured with two DAVIS240C cameras, delivering well-exposed and clear intensity frames. The spatial resolution of both the IJRR and HQF datasets is $240 \\times 180$ . The AED dataset contains 7 event sequences with a resolution of $408 \\times 306$ . More details of the exact time split can be found in the supplementary material.", + "bbox": [ + 212, + 657, + 787, + 792 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Evaluation Metrics: We evaluate the efficacy of our method using several image quality metrics, including mean squared error (MSE), structural similarity (SSIM) [52], and learned perceptual image patch similarity (LPIPS) [57].", + "bbox": [ + 214, + 795, + 787, + 842 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "EvINR", + "bbox": [ + 681, + 114, + 730, + 126 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 785, + 126 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/b88b076fc57a82297a1633d086d37ebbae2326e362e773ae4316c839a17975f9.jpg", + "image_caption": [ + "Fig. 4: Qualitative comparison with baseline methods on IJRR(Row 1&2), HQF (Row 3&4) and AED(Row 5&6)." + ], + "image_footnote": [], + "bbox": [ + 225, + 148, + 777, + 483 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Implementation Details: We use a SIREN network [46] with three hidden layers and 512 neurons per layer as our INR network. We partition each event sequence into sub-sequences lasting 5 seconds each and concurrently train all subsequences using model ensembling for 300 iterations. We first stack the events within $\\frac{1}{32}$ second into an event frame and temporal upsample the event frames in 100 and 200 iterations, as described in Sec. 3.3. We adopt the Adam optimizer [23] with a learning rate of $1\\mathrm{e} - 4$ and exponentially reduce the learning rate every 10 iterations with a decay rate of 0.95. The weight of spatial regularization $\\lambda$ is set to 0.05, and $\\gamma$ in Eq. 10 is set to 0.6. The activated threshold $C$ is set to 1 for the IJRR and HQF datasets and 0.25 for the AED dataset. The training process takes approximately 8 seconds on a single RTX3090 GPU.", + "bbox": [ + 212, + 551, + 787, + 719 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "4.2 Evaluation of Video Reconstruction", + "text_level": 1, + "bbox": [ + 215, + 739, + 555, + 753 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "We assess the effectiveness of our approach by comparing it against eight state-of-the-art (SoTA) methods, classified based on the amount of data required for training. The methods are categorized into supervised learning (SL) methods that utilize synthetic ground-truth intensity frames for supervision and self-supervised learning (SSL) methods that rely solely on event data. Specifically, we", + "bbox": [ + 212, + 763, + 787, + 840 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "Z. Wang et al.", + "bbox": [ + 271, + 114, + 367, + 128 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "compare against: 1) Five SL methods: FireNet [41], E2VID [35], FireNet+ [49], E2VID+ [49], and ET-Net [53]. 2) Three SSL method: SSL-E2VID [31], HF [40] and ELRP [60]. Reconstructed results for all methods were generated at each timestamp of the intensity frame. We use optical flow estimation from the FlowNet of SSL-E2VID [31] for ELRP [60]. We apply Contrast Limited Adaptive Histogram Equalization (CLAHE) [54] to both ground truth and synthesized frames before evaluation, following [31]. Note that the quantitative results for FireNet [41], E2VID [35], FireNet+ [49], E2VID+ [49], and ET-Net [53] are referred from a recent benchmark paper [11]. For the visual comparison, we obtained the visualization results by downloading the publicly available checkpoints and test them in the local environment.", + "bbox": [ + 212, + 145, + 787, + 313 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/7d7d49329c23b134ec783b24ffc0eff2cc33e0395c731e60816f58ece0b7c846.jpg", + "table_caption": [ + "Table 1: Comparison of quantitative results on the IJRR, HQF, and AED datasets. Bold values indicate the best results among all methods, while underlined values indicate the best results among SSL methods." + ], + "table_footnote": [], + "table_body": "
MethodsIJRRHQFAED
MSESSIMLPIPSMSESSIMLPIPSMSESSIMLPIPS
SLFireNet [41]0.1310.5020.3200.0940.5330.4410.0740.2980.579
E2VID [35]0.2120.4240.3500.1270.5400.3820.0560.4240.500
FireNet+ [49]0.0630.5550.2900.0400.6140.3140.0940.2320.566
E2VID+ [49]0.0700.5600.2360.0360.6430.2520.0740.3450.462
ET-Net [53]0.0470.6170.2240.0320.6580.2600.0840.3120.482
SSLSSL_E2VID [31]0.0970.4730.4090.0700.4800.4640.0940.3160.453
HF [40]0.1640.3340.6580.1330.2320.6700.0800.2400.943
ELRP [60]0.0800.4370.4850.0740.4500.4740.0840.3050.473
Ours0.0470.6280.2510.0480.5310.3330.0670.4580.366
", + "bbox": [ + 241, + 393, + 754, + 545 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "The quantitative results are presented in Table 1. Our method demonstrates superior performance compared to the best SL methods, with improvements of $7\\%$ , $13\\%$ , and $4\\%$ in terms of MSE, SSIM, and LPIPS respectively on the IJRR dataset [27]. Although the gap between our method and SL methods widens on the HQF dataset [49] due to the lower event density, our method remains comparable. On the AED dataset, our method outperforms the state-of-the-art SL methods, E2VID+ and ET-Net, by a clear margin for all three metrics. Notably, our method shows a significant improvement over previous SSL methods. In particular, compared with SSL-E2VID, it improves MSE, SSIM, and LPIPS by $35\\%$ , $25\\%$ , and $21\\%$ , respectively.", + "bbox": [ + 212, + 580, + 787, + 733 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "A qualitative comparison of our method with the baseline methods is depicted in Fig. 4. Our method produces intensity frames with better contrast and overall visual quality, compared with other methods that suffer from issues such as foggy effects, artifacts, and loss of detailed structures. It is worth noting that the performance of E2VID+ and ET-Net degrades significantly on the AED dataset, which suggests that their training strategy may overfit the DAVIS sensor setting. Additional results can be found in the supplementary material.", + "bbox": [ + 212, + 734, + 787, + 840 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "EvINR", + "bbox": [ + 681, + 114, + 730, + 126 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 114, + 784, + 126 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/f7a9257782005390f2966a919666006df291eb6866ff46bc08560ada41cdefdf.jpg", + "image_caption": [ + "Ground Truth" + ], + "image_footnote": [], + "bbox": [ + 264, + 148, + 382, + 223 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/0ba6338b8a4ac641f5742e6d3e509e7169edf9069876090374e97d3dfd0be88c.jpg", + "image_caption": [ + "No Reg.", + "Fig. 5: The impact of removing the spatial regularization and coarse-to-fine training." + ], + "image_footnote": [], + "bbox": [ + 383, + 148, + 500, + 223 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/c940bc1649a90698a9eac470e635f639ce71cc57ba404872ea2f51d56821a8d3.jpg", + "image_caption": [ + "No $C2T$" + ], + "image_footnote": [], + "bbox": [ + 500, + 148, + 616, + 223 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/4b3f6d33b3d4a79d16b6af3d999865c30834d84df3a12cfecb774900ea55ec5e.jpg", + "image_caption": [ + "Full Model" + ], + "image_footnote": [], + "bbox": [ + 616, + 148, + 733, + 223 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/5a41ba43443cfceab2f1e8187a11f7f91e9bd95d10d1a6a8e6908181ab66c90b.jpg", + "table_caption": [ + "Table 2: Ablation of the spatial regularization term (Reg.) and coarse-to-fine training $(C2F_{\\cdot})$ on the IJRR [27] dataset." + ], + "table_footnote": [], + "table_body": "
MSESSIMLPIPSTime(s)FPS
Base0.0630.5420.35717.2913.77
Base+Reg.0.0440.6120.27517.4413.64
Base+C2F.0.0610.5660.3437.9431.33
Full Model0.0470.6580.2518.0829.45
", + "bbox": [ + 372, + 316, + 624, + 380 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "4.3 Ablation Study", + "text_level": 1, + "bbox": [ + 215, + 407, + 388, + 422 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "We conduct ablation experiments on the IJRR dataset to assess the significance of each individual component of our method, and the results are presented in Tab. 2. Our method's basic implementation, which doesn't include spatial regularization term as described in Sec. 3.2 or use coarse-to-fine training as described in Sec. 3.3, is called Base. We also evaluate the performance of two modified versions of the full model: $\\text{Base} + \\text{Reg}$ , which adds a spatial regularization term (Eq. 7) with $\\lambda = 0.05$ , and $\\text{Base} + C2F$ , which uses coarse-to-fine training. Experimental results show that both spatial regularization and coarse-to-fine training contribute to the improved performance of our full model. Fig. 5 confirms the effectiveness of these two modules, showing that removing the spatial regularization term introduces noticeable noise (e.g., in the background), while removing the coarse-to-fine training scheme leads to an increase in training time and missing details (e.g., in the people's clothes and the book).", + "bbox": [ + 212, + 431, + 787, + 628 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "4.4 Discussions", + "text_level": 1, + "bbox": [ + 215, + 648, + 356, + 662 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Training speed: To assess the training speed of EvINR, we conducted an analysis using various model ensembling configurations on the whole Calibration sequence from the IJRR dataset [27]. We divide the event sequence, which had a total duration of 50 seconds, into $N$ partitions, with each partition having a duration of $\\tau$ seconds. Subsequently, we trained $N$ EvINR models in parallel, incorporating the model ensembling techniques outlined in Sec. 3.4. The experimental findings are presented in Table 3. By training 10 EvINR models, each corresponding to 5 seconds of the event sequence, we achieved a performance gain of over $26\\%$ , while only requiring an additional $14\\%$ of training time compared to training a single EvINR model on the entire sequence. However, we observed a significant drop in performance when increasing the partition size to", + "bbox": [ + 212, + 672, + 787, + 840 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "Z. Wang et al.", + "bbox": [ + 271, + 114, + 367, + 128 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/381a6eb941d4682e436184331fcc3c8c50d210e2348003a6740e6488de7fbecc.jpg", + "table_caption": [ + "Table 3: Impact of hyper-parameters of model ensembling on the training speed and reconstruction performance." + ], + "table_footnote": [], + "table_body": "
NτMSESSIMLPIPSTime(s)FPS
1500.0840.460.38015.7175.66
5100.06880.6000.25316.9170.31
1050.06250.6120.24418.3264.90
5010.08750.4880.34329.4340.39
", + "bbox": [ + 383, + 185, + 614, + 250 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "50, suggesting that a 1-second sequence is insufficient for EvINR to converge to a stable solution of the event generation equation. We also note that the optimal choice of hyper-parameters may differ for different event sequences due to the per-scene optimization nature of our INR approach.", + "bbox": [ + 212, + 277, + 782, + 338 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Event enhancement: Our approach provides a smooth and continuous representation of event data by representing events triggered within a small time window $\\Delta t$ as $\\frac{\\partial F_{\\Theta}}{\\partial t}\\Delta t$ . This event representation automatically reduces noise and preserves critical information through INR optimization, making it highly robust to noise. Fig. 6 compares the denoising results of our INR representation with several SoTA denoising methods [12, 51].", + "bbox": [ + 212, + 339, + 784, + 430 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/8032d3c12d165de605fa1625b9fa5542f9857961e8e68a34622950fe860b7716.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 277, + 460, + 387, + 523 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/7d1ac315e7e67e7471f35e452d773201290c823b840a7b64ecf2033f15067ebb.jpg", + "image_caption": [ + "Event Frame" + ], + "image_footnote": [], + "bbox": [ + 287, + 523, + 388, + 587 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/4b506b332bff3bcdee37dfb10d1054f532c888f8b80af291913e10bb45257318.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 387, + 460, + 493, + 523 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/39a772ba0b3e40f480c2aaea9a8513d99950c45bb8ef43c37455a23c3e7bb1b0.jpg", + "image_caption": [ + "DF" + ], + "image_footnote": [], + "bbox": [ + 388, + 523, + 495, + 587 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/1619bce478bbddd37d6b8151767cf916315f343e5f9d286d3f61a5aeea8591cb.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 495, + 460, + 602, + 523 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/a4710b8d3db01e9bf2f8041c897cbe2cac524d9278a7a4ee6a346d78a05f7848.jpg", + "image_caption": [ + "EvGait" + ], + "image_footnote": [], + "bbox": [ + 496, + 523, + 602, + 587 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/9d82d6ca8bbd68704ad5e127b71bf7764d211f739eadb2178620b7bd99c94080.jpg", + "image_caption": [ + "EvINR", + "Fig. 6: Qualitative comparison with baseline methods of event stream denoising." + ], + "image_footnote": [], + "bbox": [ + 611, + 460, + 722, + 588 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Early frames reconstruction: [6] suggested that RNN-based techniques for video reconstruction, e.g. E2VID [35], require an initialization period to achieve satisfactory results. Consequently, the initial frames generated by these methods may be of poor quality, restricting their usefulness to short event sequences. Conversely, our proposed approach is capable of producing realistic outcomes using a minimal number of events. Fig. 7 illustrates a comparison of the first frame generated by our method and RNN-based techniques, demonstrating our ability to rapidly create high-quality outputs with minimal input.", + "bbox": [ + 212, + 657, + 784, + 777 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Inference speed: Compared to other SoTA approaches, our method offers a significant improvement in terms of inference speed, as demonstrated in Tab. 4. The main reason for this improvement is that other methods require converting event data into event frames or voxel grids during the inference period, which", + "bbox": [ + 212, + 779, + 784, + 839 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "EvINR", + "bbox": [ + 681, + 114, + 730, + 126 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/c35390e0359931825e3b4039a2fd20f9b91363e11be62017c32f36dd25f9aefd.jpg", + "image_caption": [ + "E2VID[34]" + ], + "image_footnote": [], + "bbox": [ + 303, + 151, + 433, + 210 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/2dfe2d1cc82bb52f4b5921891797685b286cfcba654471c19da9a619e1c6f688.jpg", + "image_caption": [ + "E2VID+[47]" + ], + "image_footnote": [], + "bbox": [ + 434, + 151, + 562, + 210 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/3730b3787cf9da2d3280bdb17a773f1d521a9cf1c60450669f2456005a04886a.jpg", + "image_caption": [ + "ET-Net[51]" + ], + "image_footnote": [], + "bbox": [ + 563, + 151, + 691, + 210 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/42df6c93f0106a00743e6713913223fe206d95d2a24985799ef5631daaf62778.jpg", + "image_caption": [ + "SSL-E2VID[30]" + ], + "image_footnote": [], + "bbox": [ + 303, + 231, + 433, + 291 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/b1fb2954e7b8627e18580774eae86cbc05ee4ed39006f3f28b188cba67cb09ff.jpg", + "image_caption": [ + "EvINR" + ], + "image_footnote": [], + "bbox": [ + 434, + 231, + 562, + 291 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/4591044b040d257a1814d11a0bea96ca0867343a5f7f537aef2b2e855115169c.jpg", + "image_caption": [ + "GT", + "Fig. 7: The reconstructed initial frame from different compared methods." + ], + "image_footnote": [], + "bbox": [ + 563, + 231, + 691, + 291 + ], + "page_idx": 13 + }, + { + "type": "table", + "img_path": "images/5129d4987ce996e606c1c1cf0475cf098116a247a215a2e447d24b88f0193828.jpg", + "table_caption": [ + "Table 4: Comparison of inference time in terms of frames-per-second (fps) at two resolutions from IJRR [27] and AED dataset respectively." + ], + "table_footnote": [], + "table_body": "
ResolutionFireNetE2VIDET-NetOurs
(240, 180)118.24100.9636.1178.19
(408, 306)28.8328.5618.2755.37
", + "bbox": [ + 354, + 387, + 643, + 441 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "adds significant overhead. However, our method only requires the spatiotemporal coordinates as input, thus avoiding this issue.", + "bbox": [ + 212, + 472, + 784, + 502 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "5 Conclusion", + "text_level": 1, + "bbox": [ + 214, + 532, + 359, + 547 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "This paper introduced EvINR, a SSL method for event-to-video reconstruction that relieves the need for synthetic data or optical flow estimation. We, for the first time, show that high-quality videos can be reconstructed in a self-supervised and interpretable way without time-consuming end-to-end training. Our method is based on directly solving the event generation model via optimizing an INR whose temporal derivative is self-supervised by events and spatial derivative is regularized to reduce artifacts. Additionally, we propose several acceleration techniques that significantly reduce the training time of EvINR, making it applicable for online tasks. Experiments show that our approach significantly outperforms the previous SSL methods, and is competitive with the SoTA supervised methods. Our approach also demonstrates superior interpretability and robustness to various event dataset. Overall, our work contributes to the advancement of event-to-video reconstruction and offers a promising direction for future research that combines INRs with event data.", + "bbox": [ + 212, + 566, + 787, + 777 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Limitations and Future Work: The current parameter size of EvINR takes up approximately the same amount of storage as the original event data. In future work, we plan to explore network pruning and quantization techniques to further reduce the parameter size.", + "bbox": [ + 212, + 779, + 787, + 840 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "Z. Wang et al.", + "bbox": [ + 271, + 114, + 367, + 128 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgments", + "text_level": 1, + "bbox": [ + 217, + 143, + 392, + 162 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "This paper is supported by the National Natural Science Foundation of China (NSF) under Grant No. NSFC22FYT45, the Guangzhou City, University and Enterprise Joint Fund under Grant No.SL2022A03J01278, and Guangzhou Fundamental and Applied Basic Research (Grant Number: 2024A04J4072)", + "bbox": [ + 215, + 175, + 785, + 236 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 215, + 258, + 321, + 273 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Alpsentek products. https://www.alpsentek.com/product, [Accessed: Feb. 17, 2023]", + "2. Pepperl fuchs event camera. https://www.pepperl-fuchs.com/global/en/classid_11544.htm?view=productgroupliterature, accessed on 17 Feb 2023", + "3. Bardow, P., Davison, A.J., Leutenegger, S.: Simultaneous optical flow and intensity estimation from an event camera. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 884-892 (2016)", + "4. Bishop, C.M.: Pattern Recognition and Machine Learning. Springer (2006)", + "5. Brandli, C., Berner, R., Yang, M., Liu, S.C., Delbruck, T.: A $240 \\times 180$ 130 db $3\\mu s$ latency global shutter spatiotemporal vision sensor. IEEE Journal of Solid-State Circuits 49(10), 2333-2341 (2014)", + "6. Cadena, P.R.G., Qian, Y., Wang, C., Yang, M.: Spade-e2vid: Spatially-adaptive denormalization for event-based video reconstruction. IEEE Transactions on Image Processing 30, 2488-2500 (2021)", + "7. Chen, H., He, B., Wang, H., Ren, Y., Lim, S.N., Shrivastava, A.: Nerv: Neural representations for videos. Advances in Neural Information Processing Systems 34, 21557-21568 (2021)", + "8. Chen, Y., Liu, S., Wang, X.: Learning continuous image representation with local implicit image function. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 8628-8638 (2021)", + "9. Cook, M., Gugelmann, L., Jug, F., Krautz, C., Steger, A.: Interacting maps for fast visual interpretation. In: The 2011 International Joint Conference on Neural Networks. pp. 770-776. IEEE (2011)", + "10. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929 (2020)", + "11. Ercan, B., Eker, O., Erdem, A., Erdem, E.: Evreal: Towards a comprehensive benchmark and analysis suite for event-based video reconstruction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3942-3951 (2023)", + "12. Feng, Y., Lv, H., Liu, H., Zhang, Y., Xiao, Y., Han, C.: Event density based denoising method for dynamic vision sensor. Applied Sciences 10(6), 2024 (2020)", + "13. Gallego, G., Delbrück, T., Orchard, G., Bartolozzi, C., Taba, B., Censi, A., Leutenegger, S., Davison, A.J., Conradt, J., Daniilidis, K., et al.: Event-based vision: A survey. IEEE transactions on pattern analysis and machine intelligence 44(1), 154-180 (2020)", + "14. Gehrig, D., Gehrig, M., Hidalgo-Carrio, J., Scaramuzza, D.: Video to events: Recycling video datasets for event cameras. In: IEEE Conf. Comput. Vis. Pattern Recog. (CVPR) (June 2020)" + ], + "bbox": [ + 218, + 287, + 784, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "EvINR", + "bbox": [ + 681, + 114, + 730, + 126 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "15. Girshick, R.: Fast r-cnn. In: Proceedings of the IEEE international conference on computer vision. pp. 1440-1448 (2015)", + "16. Hansen, L.K., Salamon, P.: Neural network ensembles. IEEE transactions on pattern analysis and machine intelligence 12(10), 993-1001 (1990)", + "17. He, K., Gkioxari, G., Dollár, P., Girshick, R.: Mask r-cnn. In: Proceedings of the IEEE international conference on computer vision. pp. 2961-2969 (2017)", + "18. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 770-778 (2016)", + "19. Hu, Y., Liu, S.C., Delbruck, T.: v2e: From video frames to realistic dvs events. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1312-1321 (2021)", + "20. Hwang, I., Kim, J., Kim, Y.M.: Ev-nerf: Event based neural radiance field. arXiv preprint arXiv:2206.12455 (2022)", + "21. Kim, H., Handa, A., Benosman, R., Ieng, S.H., Davison, A.J.: Simultaneous mosaicing and tracking with an event camera. J. Solid State Circ 43, 566-576 (2008)", + "22. Kim, H., Leutenegger, S., Davison, A.J.: Real-time 3d reconstruction and 6-dof tracking with an event camera. In: European conference on computer vision. pp. 349-364. Springer (2016)", + "23. Kingma, D.P., Ba, J.: Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980 (2014)", + "24. Klenk, S., Koestler, L., Scaramuzza, D., Cremers, D.: E-nerf: Neural radiance fields from a moving event camera. IEEE Robotics and Automation Letters (2023)", + "25. Mescheder, L., Oechsle, M., Niemeyer, M., Nowozin, S., Geiger, A.: Occupancy networks: Learning 3d reconstruction in function space. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 4460-4470 (2019)", + "26. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM 65(1), 99-106 (2021)", + "27. Mueggler, E., Rebecq, H., Gallego, G., Delbruck, T., Scaramuzza, D.: The event-camera dataset and simulator: Event-based data for pose estimation, visual odometry, and slam. The International Journal of Robotics Research 36(2), 142-149 (2017)", + "28. Munda, G., Reinbacher, C., Pock, T.: Real-time intensity-image reconstruction for event cameras using manifold regularisation. International Journal of Computer Vision 126(12), 1381-1393 (2018)", + "29. Niemeyer, M., Geiger, A.: Giraffe: Representing scenes as compositional generative neural feature fields. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11453-11464 (2021)", + "30. Oechsle, M., Peng, S., Geiger, A.: Unisurf: Unifying neural implicit surfaces and radiance fields for multi-view reconstruction. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5589-5599 (2021)", + "31. Paredes-Vallés, F., de Croon, G.C.: Back to event basics: Self-supervised learning of image reconstruction for event cameras via photometric constancy. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3446-3455 (2021)", + "32. Park, J.J., Florence, P., Straub, J., Newcombe, R., Lovegrove, S.: Deepsdf: Learning continuous signed distance functions for shape representation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 165-174 (2019)" + ], + "bbox": [ + 215, + 147, + 784, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "Z. Wang et al.", + "bbox": [ + 271, + 114, + 367, + 128 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "33. Paszke, A., Gross, S., Chintala, S., Chanan, G., Yang, E., DeVito, Z., Lin, Z., Desmaison, A., Antiga, L., Lerer, A.: Automatic differentiation in pytorch (2017)", + "34. Pini, S., Borghi, G., Vezzani, R.: Learn to see by events: Color frame synthesis from event and rgb cameras. arXiv preprint arXiv:1812.02041 (2018)", + "35. Rebecq, H., Ranftl, R., Koltun, V., Scaramuzza, D.: Events-to-video: Bringing modern computer vision to event cameras. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3857-3866 (2019)", + "36. Redmon, J., Divvala, S., Girshick, R., Farhadi, A.: You only look once: Unified, real-time object detection. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 779-788 (2016)", + "37. Reinhard, E., Stark, M., Shirley, P., Ferwerda, J.: Photographic tone reproduction for digital images. ACM Transactions on Graphics 21(3), 267-276 (2002). https://doi.org/10.1145/566654.566576", + "38. Rudnev, V., Elgharib, M., Theobalt, C., Golyanik, V.: Eventnerf: Neural radiance fields from a single colour event camera. arXiv preprint arXiv:2206.11896 (2022)", + "39. Ryu, H.E.: Industrial dvs design; key features and applications. In: Conf. on Computer Vision and Pattern Recognition. vol. 3 (2019)", + "40. Scheerlinck, C., Barnes, N., Mahony, R.: Continuous-time intensity estimation using event cameras. In: Asian Conference on Computer Vision. pp. 308-324. Springer (2018)", + "41. Scheerlinck, C., Rebecq, H., Gehrig, D., Barnes, N., Mahony, R., Scaramuzza, D.: Fast image reconstruction with an event camera. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 156-163 (2020)", + "42. Scheerlinck, C., Rebecq, H., Stoffregen, T., Barnes, N., Mahony, R., Scaramuzza, D.: Ced: Color event camera dataset. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops. pp. 0-0 (2019)", + "43. Schwarz, K., Liao, Y., Niemeyer, M., Geiger, A.: Graf: Generative radiance fields for 3d-aware image synthesis. In: Advances in Neural Information Processing Systems (NeurIPS) (2020)", + "44. Shiba, S., Aoki, Y., Gallego, G.: Event collapse in contrast maximization frameworks. Sensors 22(14), 5190 (2022)", + "45. Shiba, S., Aoki, Y., Gallego, G.: Secrets of event-based optical flow. In: European Conference on Computer Vision. pp. 628-645. Springer (2022)", + "46. Sitzmann, V., Martel, J., Bergman, A., Lindell, D., Wetzstein, G.: Implicit neural representations with periodic activation functions. Advances in Neural Information Processing Systems 33, 7462-7473 (2020)", + "47. Skorokhodov, I., Ignatyev, S., Elhoseiny, M.: Adversarial generation of continuous images. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10753-10764 (2021)", + "48. Song, C., Huang, Q., Bajaj, C.: E-cir: Event-enhanced continuous intensity recovery. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7803-7812 (2022)", + "49. Stoffregen, T., Scheerlinck, C., Scaramuzza, D., Drummond, T., Barnes, N., Kleeman, L., Mahony, R.: Reducing the sim-to-real gap for event cameras. In: European Conference on Computer Vision. pp. 534-549. Springer (2020)", + "50. Strümpler, Y., Postels, J., Yang, R., Gool, L.V., Tombari, F.: Implicit neural representations for image compression. In: Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXVI. pp. 74-91. Springer (2022)" + ], + "bbox": [ + 212, + 146, + 784, + 839 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "EvINR", + "bbox": [ + 681, + 114, + 730, + 126 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "51. Wang, Y., Du, B., Shen, Y., Wu, K., Zhao, G., Sun, J., Wen, H.: Ev-gait: Event-based robust gait recognition using dynamic vision sensors. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6358-6367 (2019)", + "52. Wang, Z., Bovik, A.C., Sheikh, H.R., Simoncelli, E.P.: Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing 13(4), 600-612 (2004)", + "53. Weng, W., Zhang, Y., Xiong, Z.: Event-based video reconstruction using transformer. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 2563-2572 (2021)", + "54. Yadav, G., Maheshwari, S., Agarwal, A.: Contrast limited adaptive histogram equalization based enhancement for real time video system. In: 2014 international conference on advances in computing, communications and informatics (ICACCI). pp. 2392-2397. IEEE (2014)", + "55. Yang, M., Liu, S.C., Delbruck, T.: A dynamic vision sensor with $1\\%$ temporal contrast sensitivity and in-pixel asynchronous delta modulator for event encoding. IEEE Journal of Solid-State Circuits 50(9), 2149-2160 (2015)", + "56. Zhang, K., Li, Y., Zuo, W., Zhang, L., Van Gool, L., Timofte, R.: Plug-and-play image restoration with deep denoiser prior. IEEE Transactions on Pattern Analysis and Machine Intelligence 44(10), 6360-6376 (2021)", + "57. Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual metric. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 586-595 (2018)", + "58. Zhang, X., Srinivasan, P.P., Deng, B., Debevec, P., Freeman, W.T., Barron, J.T.: Nerfactor: Neural factorization of shape and reflectance under an unknown illumination. ACM Transactions on Graphics (TOG) 40(6), 1-18 (2021)", + "59. Zhang, Z., Yezzi, A., Gallego, G.: Image reconstruction from events. why learn it? arXiv preprint arXiv:2112.06242 (2021)", + "60. Zhang, Z., Yezzi, A.J., Gallego, G.: Formulating event-based image reconstruction as a linear inverse problem with deep regularization using optical flow. IEEE Transactions on Pattern Analysis & Machine Intelligence 45(07), 8372-8389 (2023)", + "61. Zhu, A.Z., Thakur, D., Ozaslan, T., Pfrommer, B., Kumar, V., Daniilidis, K.: The multivehicle stereo event camera dataset: An event camera dataset for 3d perception. IEEE Robotics and Automation Letters 3(3), 2032-2039 (2018)", + "62. Zhu, A.Z., Yuan, L., Chaney, K., Daniilidis, K.: Ev-flownet: Self-supervised optical flow estimation for event-based cameras. arXiv preprint arXiv:1802.06898 (2018)", + "63. Zhu, A.Z., Yuan, L., Chaney, K., Daniilidis, K.: Unsupervised event-based learning of optical flow, depth, and egomotion. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 989-997 (2019)", + "64. Zhu, L., Wang, X., Chang, Y., Li, J., Huang, T., Tian, Y.: Event-based video reconstruction via potential-assisted spiking neural network. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3594-3604 (2022)" + ], + "bbox": [ + 215, + 146, + 787, + 742 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "Z. Wang et al.", + "bbox": [ + 271, + 114, + 367, + 128 + ], + "page_idx": 17 + } +] \ No newline at end of file diff --git a/2024/Revisit Event Generation Model_ Self-Supervised Learning of Event-to-Video Reconstruction with Implicit Neural Representations/a41e408e-cbb3-4c38-85c3-298d71f3cf9d_model.json b/2024/Revisit Event Generation Model_ Self-Supervised Learning of Event-to-Video Reconstruction with Implicit Neural Representations/a41e408e-cbb3-4c38-85c3-298d71f3cf9d_model.json new file mode 100644 index 0000000000000000000000000000000000000000..1a708981f95691eb65ab5ac1d55b5bf3ae922f12 --- /dev/null +++ b/2024/Revisit Event Generation Model_ Self-Supervised Learning of Event-to-Video Reconstruction with Implicit Neural Representations/a41e408e-cbb3-4c38-85c3-298d71f3cf9d_model.json @@ -0,0 +1,2733 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.223, + 0.141, + 0.782, + 0.21 + ], + "angle": 0, + "content": "Revisit Event Generation Model: Self-Supervised Learning of Event-to-Video Reconstruction with Implicit Neural Representations" + }, + { + "type": "text", + "bbox": [ + 0.316, + 0.235, + 0.687, + 0.252 + ], + "angle": 0, + "content": "Zipeng Wang1, Yunfan Lu1, and Lin Wang1,2" + }, + { + "type": "text", + "bbox": [ + 0.35, + 0.262, + 0.653, + 0.276 + ], + "angle": 0, + "content": "1 Artificial Intelligence Thrust, HKUST(GZ)" + }, + { + "type": "text", + "bbox": [ + 0.316, + 0.276, + 0.687, + 0.291 + ], + "angle": 0, + "content": "\\(^{2}\\) Dept. of Computer Science and Engineering, HKUST" + }, + { + "type": "text", + "bbox": [ + 0.282, + 0.292, + 0.722, + 0.305 + ], + "angle": 0, + "content": "{zwang253,ylu066}@connect.hkust-gz.edu.cn, linwang@ust.hk" + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.34, + 0.744, + 0.687 + ], + "angle": 0, + "content": "Abstract. Reconstructing intensity frames from event data while maintaining high temporal resolution and dynamic range is crucial for bridging the gap between event-based and frame-based computer vision. Previous approaches have depended on supervised learning on synthetic data, which lacks interpretability and risk over-fitting to the setting of the event simulator. Recently, self-supervised learning (SSL) based methods, which primarily utilize per-frame optical flow to estimate intensity via photometric constancy, has been actively investigated. However, they are vulnerable to errors in the case of inaccurate optical flow. This paper proposes a novel SSL event-to-video reconstruction approach, dubbed EvINR, which eliminates the need for labeled data or optical flow estimation. Our core idea is to reconstruct intensity frames by directly addressing the event generation model, essentially a partial differential equation (PDE) that describes how events are generated based on the time-varying brightness signals. Specifically, we utilize an implicit neural representation (INR), which takes in spatiotemporal coordinate \\((x,y,t)\\) and predicts intensity values, to represent the solution of the event generation equation. The INR, parameterized as a fully-connected Multi-layer Perceptron (MLP), can be optimized with its temporal derivatives supervised by events. To make EvINR feasible for online requisites, we propose several acceleration techniques that substantially expedite the training process. Comprehensive experiments demonstrate that our EvINR surpasses previous SSL methods by \\(38\\%\\) w.r.t. Mean Squared Error (MSE) and is comparable or superior to SoTA supervised methods. Project page: https://vlislab22.github.io/EvINR/." + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.711, + 0.377, + 0.727 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.742, + 0.788, + 0.819 + ], + "angle": 0, + "content": "Event cameras [13,55] are novel sensors that offer numerous advantages over traditional frame-based cameras, including low power consumption, high dynamic range (HDR), and high temporal resolution [34]. However, their unique imaging paradigm presents a challenge when applying vision algorithms designed for frame-based cameras. To address this challenge and bridge the gap between" + }, + { + "type": "page_footnote", + "bbox": [ + 0.22, + 0.825, + 0.383, + 0.84 + ], + "angle": 0, + "content": "* Corresponding author" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.37, + 0.129 + ], + "angle": 0, + "content": "Z. Wang et al." + }, + { + "type": "image_caption", + "bbox": [ + 0.26, + 0.147, + 0.42, + 0.162 + ], + "angle": 0, + "content": "Event Generation Model" + }, + { + "type": "image", + "bbox": [ + 0.258, + 0.165, + 0.744, + 0.337 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.349, + 0.788, + 0.419 + ], + "angle": 0, + "content": "Fig. 1: Connection between event generation model and EvINR: Event generation model reveals the relation between discrete events and continuous temporal intensity changes, described as the event generation equation (Eq. 4). EvINR utilizes an INR to solve Eq. 4 and recovering a continuous function of intensity w.r.t. time, implicitly parameterized with a fully connected MLP." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.456, + 0.785, + 0.486 + ], + "angle": 0, + "content": "event-based and standard computer vision [10, 15, 17, 18, 36], many methods have been proposed to reconstruct intensity frames from events." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.49, + 0.788, + 0.701 + ], + "angle": 0, + "content": "Early methods primarily rely on hand-crafted integrators or filters, resulting in significant loss of detail in reconstructed results [3,22,28,40,59]. More recently, deep learning-based methods [6,35,41,49,53,64] have emerged, typically supervised by large-scale synthetic datasets generated using an event simulator [14,19]. However, the interpretability of these methods is limited due to the 'black-box' nature of the neural networks. Furthermore, the performance of these methods is constrained by the domain gap between synthetic and real-world data [49], leading to suboptimal results if the event simulator's settings do not accurately match those of the inference data. To address these issues, some researchers have explored self-supervised learning (SSL) frameworks [31,60], aiming to eliminate the dependency on labeled synthetic data. Nonetheless, these approaches still depend on event-based optical flow estimation, which is prone to over-fitting, occlusions, and non-convergence issues [44, 45]. As a result, the reconstructed frames often suffer from the loss of textural details and various artifacts." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.705, + 0.789, + 0.842 + ], + "angle": 0, + "content": "To address these challenges, we revisit the event generation model [13], which forms a fundamental link between events and intensity. The event generation model can be expressed as a partial differential equation (PDE), describing how events are triggered by logarithmic intensity changes that exceed a certain threshold. This PDE, known as the event generation equation, establishes a direct connection between discrete events and the partial derivative of the continuous intensity function w.r.t. time. Our key insight is that, solving the event generation equation offers an ideal self-supervised solution for event-to-video reconstruction, thereby eliminating the need for synthetic data or optical" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.682, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "EvINR" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.787, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.237 + ], + "angle": 0, + "content": "flow estimation. However, several properties of event data make solving the equation nontrivial: 1) events can be triggered at extremely high frequencies (up to \\(10^{8} / s\\)), which results in a large volume of data and a heavy computational burden. 2) events are inherently noisy, particularly in extreme visual conditions [51], which pose challenge to the robustness of solvers. 3) events do not capture the initial intensity, which makes determining the boundary values challenging." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.238, + 0.788, + 0.344 + ], + "angle": 0, + "content": "Recently, implicit neural representations (INRs) have gained popularity for solving the inverse problems in 3D reconstruction or image super-resolution [8, 26,46] by parameterizing complex signals via deep neural networks. In this work, we find that INRs possess several key advantages that render them particularly suitable for solving the event generation equation: they inherently accommodate a large volume of event data, exhibit high noise tolerance, and are flexible to add additional loss terms to regulate initial value." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.344, + 0.788, + 0.555 + ], + "angle": 0, + "content": "In light of this, we propose a novel SSL framework, termed EvINR, that employs an INR to represent the solution of the event generation equation. Our EvINR can be directly optimized by minimizing the residual between its temporal derivatives and local intensity changes estimated from event data based on the event generation model (Sec. 3.2). Moreover, we incorporate a spatial regularization term that regularizes the relative values of adjacent pixels by constraining the magnitude of spatial gradients, which effectively reduces noise in the reconstruction process (Sec. 3.2). Although the basic implementation of EvINR yields acceptable results, its convergence on seconds-long event sequences takes minutes, limiting its real-world applicability, especially in online scenarios. To expedite EvINR's training process, we introduce several acceleration techniques, including frame-based optimization, coarse-to-fine training, and model ensembling. These approaches reduce the training time from minutes to seconds while not compromising the reconstruction quality (Sec. 3.4)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.556, + 0.788, + 0.631 + ], + "angle": 0, + "content": "Moreover, most approaches are typically evaluated on event datasets captured by DAVIS sensors [27, 42, 49, 61], making it difficult to evaluate their stability and robustness to other types of event camera [1, 2, 39]. For this reason, we collected a new event dataset using an ALPIX-Eiger event camera [1], with well-aligned events and intensity frames." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.632, + 0.788, + 0.737 + ], + "angle": 0, + "content": "In summary, our paper makes three key contributions: (I) We propose EvINR, a concise SSL framework that solves the event generation equation via implicit neural representations for event-to-video reconstruction. (II) Our EvINR substantially outperforms previous SSL methods [31, 40, 59] and attains comparable, or even superior, performance compared to state-of-the-art supervised methods [35, 49, 53]. (III) We collect a real-world dataset with an ALPIX-Eiger event camera, complementary to the datasets captured by DAVIS cameras." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.762, + 0.398, + 0.779 + ], + "angle": 0, + "content": "2 Related Works" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Implicit Neural Representations (INRs) have emerged as a powerful tool for parameterizing signals, such as images, videos, and audio, in a continuous manner using neural networks [46]. Compared to traditional discrete sig-" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.369, + 0.129 + ], + "angle": 0, + "content": "Z. Wang et al." + }, + { + "type": "image", + "bbox": [ + 0.224, + 0.149, + 0.782, + 0.293 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.305, + 0.788, + 0.348 + ], + "angle": 0, + "content": "Fig. 2: Overview of EvINR. A fully connected MLP is used to implicitly solve the event generation equation. The temporal gradient of the MLP is supervised by temporal intensity changes of events, and the spatial gradient is penalized to reduce noise." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.383, + 0.788, + 0.61 + ], + "angle": 0, + "content": "nal representations, INRs offer the advantage of being able to be sampled at arbitrary resolutions with fixed memory requirements. As a result, INRs have found widespread applications in various fields, including 3D scene representation [25, 26, 32], video representation [7], generative models [29, 43, 47], and model compression [50]. A unique property of INRs is that they can be effectively learned from the derivatives of signals, such as the normals of 3D shapes [30,46,58] and the gradients of images [46]. Such a characteristic has motivated us to approach the task of event-to-video reconstruction by optimizing the INR of video from its temporal derivatives. Previous research has investigated the potential of INR for novel view synthesis using event data [20,24,38]. These approaches reconstruct 3D neural radiance fields using multiple event sequences with known camera poses from a stationary scene. Most similar to our work, E-CIR [48] uses polynomial to regress the intensity function, which may only represent a short time interval (e.g., the exposure time of an image). In contrast, our method employs INR, which can represent longer time of intensity change." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.612, + 0.788, + 0.809 + ], + "angle": 0, + "content": "Event-based Video Reconstruction has been a hot topic in the literature. Early attempts [3, 9, 21] address this problem based on the photometric constancy, which describes the relationship between intensity gradients and optical flow. Other approaches [28, 40] are based on direct event integration without estimating optical flow. Rebecq et al. [35] developed the first DL-based framework, called E2VID, to reconstruct intensity frames from events in an end-to-end manner, outperforming earlier techniques by a significant margin. E2VID was updated by some following research, trying to tackle problems of inference speed [41], the cold start problem [6], and training strategy [49]. These methods are supervised and learned using the synthetic dataset obtained via the event simulators [14, 19] due to the lack of real-world datasets with well-aligned events and intensity frames. Therefore, the generalization capability of these methods is limited by the simulator-to-real gap [49]." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.84 + ], + "angle": 0, + "content": "To address this problem, Federico et al. [31] proposed an SSL framework with a network to estimate optical flow and another network to reconstruct intensity" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.682, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "EvINR" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.147, + 0.788, + 0.269 + ], + "angle": 0, + "content": "frames based on the photometric constancy. Zhang et al. [60] updated this idea into a linear inverse problem that can be solved using modern linear solvers without using deep learning. However, these methods either assume optical flow is known or estimate the optical flow using contrast maximization [62,63]. Consequently, their performance cannot be guaranteed unless high-quality optical flow can be obtained. In contrast, our approach is based solely on the physical event generation model, leading to a more straightforward solution that also demonstrates significantly improved performance and greater flexibility." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.289, + 0.331, + 0.306 + ], + "angle": 0, + "content": "3 Method" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.32, + 0.789, + 0.487 + ], + "angle": 0, + "content": "Overview: Our objective is to reconstruct intensity frames from events in a self-supervised manner, without the need for end-to-end training or optical flow estimation. To accomplish this, we have reformulated event-based video reconstruction as solving the event generation equation. We employ an INR that is supervised solely by event data to represent the intensity function. An overview of our approach can be seen in Fig.2. In Sec. 3.1, we explain the event generation model and its connection to event-based video reconstruction. In Sec. 3.2, we detail how to train an EvINR by supervising its spatial and temporal gradients. Techniques to speed up the training of EvINR for online applications are discussed in Sec. 3.3. Finally, in Sec. 3.4, we describe our collected dataset using the ALPIX event camera [1]." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.507, + 0.578, + 0.521 + ], + "angle": 0, + "content": "3.1 Preliminary: Event Generation Model" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.529, + 0.787, + 0.59 + ], + "angle": 0, + "content": "We begin by providing a brief overview of the event generation model [13], which forms the theoretical foundation for our approach. Let \\( I(x,y,t) \\) denote the intensity of the spatial location \\( (x,y) \\) at time \\( t \\) in a video. Since event cameras operate with logarithmic intensity, we denote \\( L(x,y,t) = \\log I(x,y,t) \\)." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.591, + 0.788, + 0.68 + ], + "angle": 0, + "content": "An event camera comprises a frame of independent pixels that respond to changes in the logarithmic intensity signal and produce sequences of sparse and asynchronous events. An event \\((x_{i},y_{i},t_{i},p_{i})\\) is triggered when the logarithmic intensity change surpasses a threshold \\(C\\) since the previous event was triggered at the same pixel. where \\((x,y)\\) is the spatial location of the pixel, \\(t\\) is the timestamp, and \\(p\\in \\{-1,1\\}\\) is the polarity of the logarithmic intensity change." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.681, + 0.788, + 0.74 + ], + "angle": 0, + "content": "For simplicity, let us consider the temporal changes in logarithmic intensity of a single pixel with fixed spatial coordinates and disregard the spatial terms \\((x,y)\\). We can describe an event \\(e_i\\) on that pixel using the Dirac delta function as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.419, + 0.742, + 0.786, + 0.757 + ], + "angle": 0, + "content": "\\[\ne _ {i} (t) = p _ {i} \\cdot C \\cdot \\delta \\left(t - t _ {i}\\right). \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.762, + 0.788, + 0.807 + ], + "angle": 0, + "content": "Therefore, the logarithmic intensity increment \\(\\Delta L = L(t_2) - L(t_1)\\) in a time interval \\([t_1,t_2]\\) can be represented by the accumulation of events, which can be expressed as:" + }, + { + "type": "equation", + "bbox": [ + 0.425, + 0.807, + 0.786, + 0.843 + ], + "angle": 0, + "content": "\\[\n\\Delta L = \\int_ {t _ {1}} ^ {t _ {2}} \\sum_ {i} e _ {i} (t) \\mathrm {d} t. \\tag {2}\n\\]" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.369, + 0.129 + ], + "angle": 0, + "content": "Z. Wang et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.192 + ], + "angle": 0, + "content": "Assuming a short time interval and ignoring noise, we can approximate the logarithmic intensity increment by its first-order temporal derivative using Taylor expansion:" + }, + { + "type": "equation", + "bbox": [ + 0.408, + 0.193, + 0.786, + 0.226 + ], + "angle": 0, + "content": "\\[\n\\frac {\\Delta L}{t _ {2} - t _ {1}} = \\frac {\\partial L ((t _ {1} + t _ {2}) / 2)}{\\partial t}. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.232, + 0.784, + 0.279 + ], + "angle": 0, + "content": "By substituting Eq. 2 into Eq. 3, we derive the event generation equation (Eq. 4), which bridges the discrete event data with the continuous temporal derivatives of logarithmic intensity:" + }, + { + "type": "equation", + "bbox": [ + 0.355, + 0.291, + 0.786, + 0.328 + ], + "angle": 0, + "content": "\\[\n\\frac {\\partial L \\left(\\left(t _ {1} + t _ {2}\\right) / 2\\right)}{\\partial t} = \\frac {1}{t _ {2} - t _ {1}} \\int_ {t _ {1}} ^ {t _ {2}} \\sum_ {i} e _ {i} (t) \\mathrm {d} t. \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.34, + 0.785, + 0.491 + ], + "angle": 0, + "content": "The left-hand side of Eq. 4 represents the partial derivative of logarithmic intensity with respect to time, while the right-hand side can be calculated by accumulating events. Intuitively, event-based video reconstruction can be formulated as solving Eq. 4, i.e., finding a logarithmic intensity function that satisfies the equation. However, as stated earlier, solving Eq. 4 can be non-trivial due to the large amount, high noise rate of event data, and unknown boundary intensity values. To address these challenges, we employ INRs to solve Eq. 4 as they naturally scale up to large data and have high noise tolerance [20]. Boundary intensity values can also easily be regularized by injecting natural image priors as additional loss functions." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.518, + 0.491, + 0.533 + ], + "angle": 0, + "content": "3.2 Learning INRs from Events" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.546, + 0.784, + 0.609 + ], + "angle": 0, + "content": "We aim to learn an INR \\( F_{\\Theta} \\) given an event stream \\( \\{e_i\\}_{i=0}^N \\). Here, \\( \\Theta \\) represents the parameters of a fully connected MLP that predicts the logarithmic intensity value \\( \\hat{L} \\) at any spatiotemporal coordinate \\( (x,y,t) \\). We adapt the event frame representation [35] and stack the given events uniformly into \\( T \\) event frames." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.634, + 0.784, + 0.71 + ], + "angle": 0, + "content": "Temporal Supervision The INR is optimized by directly minimizing the temporal loss between the predicted logarithmic intensity change \\(\\Delta \\hat{L}\\) and the logarithmic intensity change estimated by event accumulation \\(\\Delta L\\). Here \\(\\Delta \\hat{L}\\) can be derived by accumulating events as Eq. 2 and \\(\\Delta \\hat{L}\\) is the change of \\(\\hat{L}\\) with respect to \\(t\\) obtained by performing double back-propagation [33] of \\(F_{\\Theta}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.381, + 0.724, + 0.786, + 0.753 + ], + "angle": 0, + "content": "\\[\n\\Delta \\hat {L} = \\frac {\\partial F _ {\\Theta} \\left(\\left(t _ {1} + t _ {2}\\right) / 2\\right)}{\\partial t} \\cdot \\left(t _ {2} - t _ {1}\\right), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.765, + 0.763, + 0.781 + ], + "angle": 0, + "content": "Here we adopt the mean squared error (MSE) loss for temporal supervision:" + }, + { + "type": "equation", + "bbox": [ + 0.424, + 0.793, + 0.786, + 0.812 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {t e m p}} = \\left(\\Delta L - \\Delta \\hat {L}\\right) ^ {2}, \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.824, + 0.66, + 0.841 + ], + "angle": 0, + "content": "where \\(\\Delta L\\) and \\(\\Delta \\hat{L}\\) are given by Eq. 2 and Eq. 5, respectively." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.682, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "EvINR" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.147, + 0.784, + 0.238 + ], + "angle": 0, + "content": "Spatial Regularization Although the temporal supervision can estimate the intensity function, the results may still contain unnatural artifacts as the INR network \\( F_{\\Theta} \\) has no prior knowledge of the initial intensity values of each pixel. To address this issue, we introduce a spatial regularization term that encourages the solution to be in the space of natural-looking images. We adopt Tikhonov regularization [4] that penalizes the spatial gradients of logarithmic intensity:" + }, + { + "type": "equation", + "bbox": [ + 0.41, + 0.25, + 0.784, + 0.281 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {r e g}} = \\left(\\frac {\\partial F _ {\\Theta}}{\\partial x}\\right) ^ {2} + \\left(\\frac {\\partial F _ {\\Theta}}{\\partial y}\\right) ^ {2}. \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.289, + 0.784, + 0.335 + ], + "angle": 0, + "content": "It is worth noting that we do not employ more complex regularization methods, such as CNN denoisers [56] as used in [60], because they typically utilize large network models that can significantly slow down the training process." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.357, + 0.713, + 0.372 + ], + "angle": 0, + "content": "Optimizing EvINR The overall objective is given by the equation:" + }, + { + "type": "equation", + "bbox": [ + 0.434, + 0.386, + 0.784, + 0.4 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathcal {L} _ {\\text {t e m p}} + \\lambda \\mathcal {L} _ {\\text {r e g}}, \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.411, + 0.784, + 0.456 + ], + "angle": 0, + "content": "where \\(\\mathcal{L}_{\\mathrm{temp}}\\) and \\(\\mathcal{L}_{\\mathrm{reg}}\\) were introduced in Sec. 3.2 and Sec. 3.2, respectively, as the temporal supervision and spatial regularization term. \\(\\lambda\\) is a hyper-parameter used to adjust the weight of the spatial regularization term." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.479, + 0.784, + 0.538 + ], + "angle": 0, + "content": "Tone-mapping The output of the EvINR is the logarithmic intensity of the reconstructed frames. We first use the exponential function to convert the predicted logarithmic intensity values to high dynamic range (HDR) intensity values \\( I \\in [0, \\infty) \\):" + }, + { + "type": "equation", + "bbox": [ + 0.405, + 0.54, + 0.784, + 0.554 + ], + "angle": 0, + "content": "\\[\nI (x, y, t) = \\exp \\left(F _ {\\Theta} (x, y, t)\\right). \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.562, + 0.784, + 0.592 + ], + "angle": 0, + "content": "Then, we adopt the Reinhard function [37] to map the HDR intensity values into the low dynamic range \\([0,1]\\):" + }, + { + "type": "equation", + "bbox": [ + 0.442, + 0.603, + 0.784, + 0.631 + ], + "angle": 0, + "content": "\\[\n\\Gamma (I) = \\left(\\frac {I}{I + 1}\\right) ^ {\\gamma}. \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.642, + 0.784, + 0.672 + ], + "angle": 0, + "content": "where \\(\\gamma\\) is a hyper-parameter used to control the contrast of \\(\\Gamma(I)\\). We assess our reconstruction performance by applying \\(\\Gamma(I)\\) in all experiments." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.695, + 0.429, + 0.71 + ], + "angle": 0, + "content": "3.3 Accelerating EvINR" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.785, + 0.84 + ], + "angle": 0, + "content": "While solving the event generation equation using the basic implementation of EvINR, as described in Sec. 3.3, yields satisfactory reconstruction results, it is not efficient enough for online tasks. Optimizing one INR network takes about 20 minutes, and a network can only represent approximately 1 second of a sequence. Reducing the training time or increasing the sequence time leads to severe performance degradation. To enable online usage for EvINR, we propose several techniques to reduce the training time and increase the representation capacity for EvINR, as illustrated in Fig. 3." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.37, + 0.129 + ], + "angle": 0, + "content": "Z. Wang et al." + }, + { + "type": "image", + "bbox": [ + 0.264, + 0.148, + 0.451, + 0.294 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.456, + 0.149, + 0.744, + 0.293 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.308, + 0.789, + 0.352 + ], + "angle": 0, + "content": "Fig. 3: Overview of acceleration techniques. (a) and (b) illustrate the difference between the basic coordinate-based and our frame-based optimization, respectively. (c) depicts our proposed coarse-to-fine training scheme and network ensembling technique." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.38, + 0.789, + 0.501 + ], + "angle": 0, + "content": "Coordinate-based to frame-based optimization: The basic EvINR algorithm is designed to learn a mapping from 3D coordinates \\((x,y,t)\\) to the logarithmic intensity \\(L\\). However, this coordinate-based optimization requires the network to 'remember' \\(H\\times W\\times T\\) coordinate-to-logarithmic intensity mappings, which can be highly complex and lead to slow convergence. To address this issue, we propose a frame-based optimization scheme that learns a mapping from timestamps \\(t\\) to the logarithmic intensity of all pixels at that timestamp. The frame-based optimization scheme can be represented as:" + }, + { + "type": "equation", + "bbox": [ + 0.385, + 0.511, + 0.787, + 0.53 + ], + "angle": 0, + "content": "\\[\nF _ {\\Theta} (t) = \\left[ L (i, j, t) \\right] _ {1 \\leq i \\leq H, 1 \\leq j \\leq W}, \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.538, + 0.789, + 0.644 + ], + "angle": 0, + "content": "where the right hand-side represents a matrix that contains all logarithmic intensity of all pixels at \\( t \\). The frame-based optimization approach requires the network to remember only \\( T \\) coordinate-to-logarithmic intensity mappings, resulting in a significant reduction in complexity and faster convergence. Empirical results show that this approach reduces the convergence time by two orders of magnitude, from minutes to seconds. The comparison between coordinate-based and frame-based optimizations is illustrated in Fig. 3 (a) and (b)." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.645, + 0.789, + 0.81 + ], + "angle": 0, + "content": "Coarse-to-fine training: We adopt a coarse-to-fine training scheme, which enables the network to learn overall logarithmic intensity changes before focusing on finer details. We structure the training process into \\(s\\) distinct stages, at each of which we increase the temporal resolution by a factor of 2. Initially, we divide the event sequence into \\(N\\) segments of equal length and optimize the EvINR in accordance with Eq. 6 over a specified number of iterations. Subsequently, we proceed to bisect each of these \\(N\\) segments into two smaller ones, ensuring that they contain an equal number of events. For all the experiments, we set the number of stages \\(s\\) to 3 and scheduled the upsampling to occur after 100 and 200 iterations, respectively. This approach accelerates training by reducing the number of event frames needed in the early stages by approximately 2 times." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.81, + 0.789, + 0.842 + ], + "angle": 0, + "content": "Network ensembling: As each EvINR network requires only about 1GB of GPU memory, we further leverage network ensemble techniques [16] to train \\(N\\)" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.682, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "EvINR" + }, + { + "type": "page_number", + "bbox": [ + 0.776, + 0.117, + 0.786, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.239 + ], + "angle": 0, + "content": "EvINR networks simultaneously on a single GPU to achieve higher parallelism. This approach enables us to exploit the computational resources of modern GPUs more effectively and speed up the training process. We keep overlap periods between nearby networks to keep logarithmic intensity predictions constant among all networks. Fig. 3(c) depicts the coarse-to-fine training process and network ensembling techniques employed in our approach." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.259, + 0.416, + 0.273 + ], + "angle": 0, + "content": "3.4 Dataset Collection" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.282, + 0.788, + 0.386 + ], + "angle": 0, + "content": "Most event-to-video reconstruction approaches are typically evaluated on event datasets collected using DAVIS sensors [5], such as IJRR [27], HQF [49], MVSEC [61], and CED [42]. Recently, other types of event cameras [1, 2, 39] have been developed, which share the same event generation model with DAVIS sensors but may differ in detailed configurations and settings. Therefore, it is essential to verify the generalization capability of existing methods and our EvINR on those new types of event sensors." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.388, + 0.789, + 0.584 + ], + "angle": 0, + "content": "To fill the gap, we introduce a new real-world dataset, called the ALPIX Event Dataset (AED), which is collected using an ALPIX-Eiger event camera [1], featuring static scenes accompanied by gradual camera motions. The camera provides well-aligned RGB frames and color events. The RGB frames have a resolution of \\(3264 \\times 2448\\) and the events have a resolution of \\(1632 \\times 1224\\). The AED dataset includes seven video sequences with diverse scenes, such as streets, buildings, indoor scenes, textures, and tools, and each approximately lasts for ten seconds. Note that, in this paper, we only focus on reconstructing grayscale frames to keep consistency with previous works [31, 60] and also for a fair comparison. Therefore, we first demosaic the RGB frames and events according to the Quad Bayer pattern, resulting in intensity frames with a resolution of \\(816 \\times 612\\) and event data with a resolution of \\(408 \\times 306\\). Details on the post-processing of AED dataset can be found in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.606, + 0.377, + 0.623 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.636, + 0.441, + 0.652 + ], + "angle": 0, + "content": "4.1 Experiments Settings" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.659, + 0.788, + 0.794 + ], + "angle": 0, + "content": "Datasets: We conduct experiments to evaluate the effectiveness of our proposed method using three datasets: IJRR [27], HQF [49], and our AED dataset. The IJRR dataset consists of intensity frames and events in 25 real scenes and 2 synthetic scenes, captured by a DAVIS240C camera [5]. HQF dataset provides 14 event data sequences captured with two DAVIS240C cameras, delivering well-exposed and clear intensity frames. The spatial resolution of both the IJRR and HQF datasets is \\(240 \\times 180\\). The AED dataset contains 7 event sequences with a resolution of \\(408 \\times 306\\). More details of the exact time split can be found in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.796, + 0.788, + 0.843 + ], + "angle": 0, + "content": "Evaluation Metrics: We evaluate the efficacy of our method using several image quality metrics, including mean squared error (MSE), structural similarity (SSIM) [52], and learned perceptual image patch similarity (LPIPS) [57]." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.369, + 0.129 + ], + "angle": 0, + "content": "Z. Wang et al." + }, + { + "type": "image", + "bbox": [ + 0.227, + 0.15, + 0.779, + 0.484 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.495, + 0.788, + 0.525 + ], + "angle": 0, + "content": "Fig. 4: Qualitative comparison with baseline methods on IJRR(Row 1&2), HQF (Row 3&4) and AED(Row 5&6)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.552, + 0.788, + 0.72 + ], + "angle": 0, + "content": "Implementation Details: We use a SIREN network [46] with three hidden layers and 512 neurons per layer as our INR network. We partition each event sequence into sub-sequences lasting 5 seconds each and concurrently train all subsequences using model ensembling for 300 iterations. We first stack the events within \\(\\frac{1}{32}\\) second into an event frame and temporal upsample the event frames in 100 and 200 iterations, as described in Sec. 3.3. We adopt the Adam optimizer [23] with a learning rate of \\(1\\mathrm{e} - 4\\) and exponentially reduce the learning rate every 10 iterations with a decay rate of 0.95. The weight of spatial regularization \\(\\lambda\\) is set to 0.05, and \\(\\gamma\\) in Eq. 10 is set to 0.6. The activated threshold \\(C\\) is set to 1 for the IJRR and HQF datasets and 0.25 for the AED dataset. The training process takes approximately 8 seconds on a single RTX3090 GPU." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.74, + 0.557, + 0.755 + ], + "angle": 0, + "content": "4.2 Evaluation of Video Reconstruction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.765, + 0.788, + 0.842 + ], + "angle": 0, + "content": "We assess the effectiveness of our approach by comparing it against eight state-of-the-art (SoTA) methods, classified based on the amount of data required for training. The methods are categorized into supervised learning (SL) methods that utilize synthetic ground-truth intensity frames for supervision and self-supervised learning (SSL) methods that rely solely on event data. Specifically, we" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.682, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "EvINR" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.785, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.146, + 0.788, + 0.314 + ], + "angle": 0, + "content": "compare against: 1) Five SL methods: FireNet [41], E2VID [35], FireNet+ [49], E2VID+ [49], and ET-Net [53]. 2) Three SSL method: SSL-E2VID [31], HF [40] and ELRP [60]. Reconstructed results for all methods were generated at each timestamp of the intensity frame. We use optical flow estimation from the FlowNet of SSL-E2VID [31] for ELRP [60]. We apply Contrast Limited Adaptive Histogram Equalization (CLAHE) [54] to both ground truth and synthesized frames before evaluation, following [31]. Note that the quantitative results for FireNet [41], E2VID [35], FireNet+ [49], E2VID+ [49], and ET-Net [53] are referred from a recent benchmark paper [11]. For the visual comparison, we obtained the visualization results by downloading the publicly available checkpoints and test them in the local environment." + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.34, + 0.788, + 0.385 + ], + "angle": 0, + "content": "Table 1: Comparison of quantitative results on the IJRR, HQF, and AED datasets. Bold values indicate the best results among all methods, while underlined values indicate the best results among SSL methods." + }, + { + "type": "table", + "bbox": [ + 0.243, + 0.395, + 0.756, + 0.546 + ], + "angle": 0, + "content": "
MethodsIJRRHQFAED
MSESSIMLPIPSMSESSIMLPIPSMSESSIMLPIPS
SLFireNet [41]0.1310.5020.3200.0940.5330.4410.0740.2980.579
E2VID [35]0.2120.4240.3500.1270.5400.3820.0560.4240.500
FireNet+ [49]0.0630.5550.2900.0400.6140.3140.0940.2320.566
E2VID+ [49]0.0700.5600.2360.0360.6430.2520.0740.3450.462
ET-Net [53]0.0470.6170.2240.0320.6580.2600.0840.3120.482
SSLSSL_E2VID [31]0.0970.4730.4090.0700.4800.4640.0940.3160.453
HF [40]0.1640.3340.6580.1330.2320.6700.0800.2400.943
ELRP [60]0.0800.4370.4850.0740.4500.4740.0840.3050.473
Ours0.0470.6280.2510.0480.5310.3330.0670.4580.366
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.581, + 0.788, + 0.734 + ], + "angle": 0, + "content": "The quantitative results are presented in Table 1. Our method demonstrates superior performance compared to the best SL methods, with improvements of \\(7\\%\\), \\(13\\%\\), and \\(4\\%\\) in terms of MSE, SSIM, and LPIPS respectively on the IJRR dataset [27]. Although the gap between our method and SL methods widens on the HQF dataset [49] due to the lower event density, our method remains comparable. On the AED dataset, our method outperforms the state-of-the-art SL methods, E2VID+ and ET-Net, by a clear margin for all three metrics. Notably, our method shows a significant improvement over previous SSL methods. In particular, compared with SSL-E2VID, it improves MSE, SSIM, and LPIPS by \\(35\\%\\), \\(25\\%\\), and \\(21\\%\\), respectively." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.735, + 0.788, + 0.842 + ], + "angle": 0, + "content": "A qualitative comparison of our method with the baseline methods is depicted in Fig. 4. Our method produces intensity frames with better contrast and overall visual quality, compared with other methods that suffer from issues such as foggy effects, artifacts, and loss of detailed structures. It is worth noting that the performance of E2VID+ and ET-Net degrades significantly on the AED dataset, which suggests that their training strategy may overfit the DAVIS sensor setting. Additional results can be found in the supplementary material." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.369, + 0.129 + ], + "angle": 0, + "content": "Z. Wang et al." + }, + { + "type": "image", + "bbox": [ + 0.266, + 0.149, + 0.383, + 0.224 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.295, + 0.225, + 0.356, + 0.235 + ], + "angle": 0, + "content": "Ground Truth" + }, + { + "type": "image", + "bbox": [ + 0.384, + 0.149, + 0.501, + 0.224 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.428, + 0.225, + 0.464, + 0.235 + ], + "angle": 0, + "content": "No Reg." + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.149, + 0.617, + 0.224 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.536, + 0.225, + 0.575, + 0.235 + ], + "angle": 0, + "content": "No \\(C2T\\)" + }, + { + "type": "image", + "bbox": [ + 0.617, + 0.149, + 0.734, + 0.224 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.652, + 0.225, + 0.701, + 0.235 + ], + "angle": 0, + "content": "Full Model" + }, + { + "type": "image_caption", + "bbox": [ + 0.217, + 0.248, + 0.784, + 0.264 + ], + "angle": 0, + "content": "Fig. 5: The impact of removing the spatial regularization and coarse-to-fine training." + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.276, + 0.787, + 0.304 + ], + "angle": 0, + "content": "Table 2: Ablation of the spatial regularization term (Reg.) and coarse-to-fine training \\((C2F_{\\cdot})\\) on the IJRR [27] dataset." + }, + { + "type": "table", + "bbox": [ + 0.374, + 0.318, + 0.625, + 0.381 + ], + "angle": 0, + "content": "
MSESSIMLPIPSTime(s)FPS
Base0.0630.5420.35717.2913.77
Base+Reg.0.0440.6120.27517.4413.64
Base+C2F.0.0610.5660.3437.9431.33
Full Model0.0470.6580.2518.0829.45
" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.408, + 0.389, + 0.423 + ], + "angle": 0, + "content": "4.3 Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.432, + 0.788, + 0.63 + ], + "angle": 0, + "content": "We conduct ablation experiments on the IJRR dataset to assess the significance of each individual component of our method, and the results are presented in Tab. 2. Our method's basic implementation, which doesn't include spatial regularization term as described in Sec. 3.2 or use coarse-to-fine training as described in Sec. 3.3, is called Base. We also evaluate the performance of two modified versions of the full model: \\( \\text{Base} + \\text{Reg} \\), which adds a spatial regularization term (Eq. 7) with \\( \\lambda = 0.05 \\), and \\( \\text{Base} + C2F \\), which uses coarse-to-fine training. Experimental results show that both spatial regularization and coarse-to-fine training contribute to the improved performance of our full model. Fig. 5 confirms the effectiveness of these two modules, showing that removing the spatial regularization term introduces noticeable noise (e.g., in the background), while removing the coarse-to-fine training scheme leads to an increase in training time and missing details (e.g., in the people's clothes and the book)." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.65, + 0.357, + 0.663 + ], + "angle": 0, + "content": "4.4 Discussions" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.674, + 0.789, + 0.842 + ], + "angle": 0, + "content": "Training speed: To assess the training speed of EvINR, we conducted an analysis using various model ensembling configurations on the whole Calibration sequence from the IJRR dataset [27]. We divide the event sequence, which had a total duration of 50 seconds, into \\( N \\) partitions, with each partition having a duration of \\( \\tau \\) seconds. Subsequently, we trained \\( N \\) EvINR models in parallel, incorporating the model ensembling techniques outlined in Sec. 3.4. The experimental findings are presented in Table 3. By training 10 EvINR models, each corresponding to 5 seconds of the event sequence, we achieved a performance gain of over \\( 26\\% \\), while only requiring an additional \\( 14\\% \\) of training time compared to training a single EvINR model on the entire sequence. However, we observed a significant drop in performance when increasing the partition size to" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.682, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "EvINR" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "table_caption", + "bbox": [ + 0.215, + 0.145, + 0.784, + 0.173 + ], + "angle": 0, + "content": "Table 3: Impact of hyper-parameters of model ensembling on the training speed and reconstruction performance." + }, + { + "type": "table", + "bbox": [ + 0.384, + 0.186, + 0.615, + 0.251 + ], + "angle": 0, + "content": "
NτMSESSIMLPIPSTime(s)FPS
1500.0840.460.38015.7175.66
5100.06880.6000.25316.9170.31
1050.06250.6120.24418.3264.90
5010.08750.4880.34329.4340.39
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.279, + 0.784, + 0.339 + ], + "angle": 0, + "content": "50, suggesting that a 1-second sequence is insufficient for EvINR to converge to a stable solution of the event generation equation. We also note that the optimal choice of hyper-parameters may differ for different event sequences due to the per-scene optimization nature of our INR approach." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.34, + 0.785, + 0.431 + ], + "angle": 0, + "content": "Event enhancement: Our approach provides a smooth and continuous representation of event data by representing events triggered within a small time window \\(\\Delta t\\) as \\(\\frac{\\partial F_{\\Theta}}{\\partial t}\\Delta t\\). This event representation automatically reduces noise and preserves critical information through INR optimization, making it highly robust to noise. Fig. 6 compares the denoising results of our INR representation with several SoTA denoising methods [12, 51]." + }, + { + "type": "image", + "bbox": [ + 0.278, + 0.462, + 0.388, + 0.525 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.288, + 0.525, + 0.39, + 0.588 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.305, + 0.59, + 0.366, + 0.6 + ], + "angle": 0, + "content": "Event Frame" + }, + { + "type": "image", + "bbox": [ + 0.388, + 0.462, + 0.495, + 0.525 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.39, + 0.525, + 0.496, + 0.588 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.438, + 0.591, + 0.455, + 0.6 + ], + "angle": 0, + "content": "DF" + }, + { + "type": "image", + "bbox": [ + 0.496, + 0.462, + 0.604, + 0.525 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.498, + 0.525, + 0.604, + 0.588 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.532, + 0.591, + 0.566, + 0.6 + ], + "angle": 0, + "content": "EvGait" + }, + { + "type": "image", + "bbox": [ + 0.612, + 0.462, + 0.723, + 0.589 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.651, + 0.591, + 0.684, + 0.6 + ], + "angle": 0, + "content": "EvINR" + }, + { + "type": "image_caption", + "bbox": [ + 0.232, + 0.615, + 0.768, + 0.63 + ], + "angle": 0, + "content": "Fig. 6: Qualitative comparison with baseline methods of event stream denoising." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.659, + 0.785, + 0.779 + ], + "angle": 0, + "content": "Early frames reconstruction: [6] suggested that RNN-based techniques for video reconstruction, e.g. E2VID [35], require an initialization period to achieve satisfactory results. Consequently, the initial frames generated by these methods may be of poor quality, restricting their usefulness to short event sequences. Conversely, our proposed approach is capable of producing realistic outcomes using a minimal number of events. Fig. 7 illustrates a comparison of the first frame generated by our method and RNN-based techniques, demonstrating our ability to rapidly create high-quality outputs with minimal input." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.785, + 0.84 + ], + "angle": 0, + "content": "Inference speed: Compared to other SoTA approaches, our method offers a significant improvement in terms of inference speed, as demonstrated in Tab. 4. The main reason for this improvement is that other methods require converting event data into event frames or voxel grids during the inference period, which" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.369, + 0.129 + ], + "angle": 0, + "content": "Z. Wang et al." + }, + { + "type": "image", + "bbox": [ + 0.304, + 0.152, + 0.434, + 0.212 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.342, + 0.214, + 0.398, + 0.225 + ], + "angle": 0, + "content": "E2VID[34]" + }, + { + "type": "image", + "bbox": [ + 0.435, + 0.152, + 0.563, + 0.212 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.471, + 0.214, + 0.534, + 0.225 + ], + "angle": 0, + "content": "E2VID+[47]" + }, + { + "type": "image", + "bbox": [ + 0.565, + 0.152, + 0.692, + 0.212 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.607, + 0.214, + 0.664, + 0.225 + ], + "angle": 0, + "content": "ET-Net[51]" + }, + { + "type": "image", + "bbox": [ + 0.304, + 0.232, + 0.434, + 0.292 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.328, + 0.293, + 0.408, + 0.304 + ], + "angle": 0, + "content": "SSL-E2VID[30]" + }, + { + "type": "image", + "bbox": [ + 0.435, + 0.232, + 0.563, + 0.292 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.48, + 0.293, + 0.516, + 0.304 + ], + "angle": 0, + "content": "EvINR" + }, + { + "type": "image", + "bbox": [ + 0.565, + 0.232, + 0.692, + 0.292 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.617, + 0.293, + 0.635, + 0.304 + ], + "angle": 0, + "content": "GT" + }, + { + "type": "image_caption", + "bbox": [ + 0.255, + 0.316, + 0.746, + 0.331 + ], + "angle": 0, + "content": "Fig. 7: The reconstructed initial frame from different compared methods." + }, + { + "type": "table_caption", + "bbox": [ + 0.215, + 0.347, + 0.787, + 0.377 + ], + "angle": 0, + "content": "Table 4: Comparison of inference time in terms of frames-per-second (fps) at two resolutions from IJRR [27] and AED dataset respectively." + }, + { + "type": "table", + "bbox": [ + 0.356, + 0.388, + 0.645, + 0.443 + ], + "angle": 0, + "content": "
ResolutionFireNetE2VIDET-NetOurs
(240, 180)118.24100.9636.1178.19
(408, 306)28.8328.5618.2755.37
" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.473, + 0.785, + 0.503 + ], + "angle": 0, + "content": "adds significant overhead. However, our method only requires the spatiotemporal coordinates as input, thus avoiding this issue." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.533, + 0.36, + 0.549 + ], + "angle": 0, + "content": "5 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.567, + 0.788, + 0.779 + ], + "angle": 0, + "content": "This paper introduced EvINR, a SSL method for event-to-video reconstruction that relieves the need for synthetic data or optical flow estimation. We, for the first time, show that high-quality videos can be reconstructed in a self-supervised and interpretable way without time-consuming end-to-end training. Our method is based on directly solving the event generation model via optimizing an INR whose temporal derivative is self-supervised by events and spatial derivative is regularized to reduce artifacts. Additionally, we propose several acceleration techniques that significantly reduce the training time of EvINR, making it applicable for online tasks. Experiments show that our approach significantly outperforms the previous SSL methods, and is competitive with the SoTA supervised methods. Our approach also demonstrates superior interpretability and robustness to various event dataset. Overall, our work contributes to the advancement of event-to-video reconstruction and offers a promising direction for future research that combines INRs with event data." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.788, + 0.841 + ], + "angle": 0, + "content": "Limitations and Future Work: The current parameter size of EvINR takes up approximately the same amount of storage as the original event data. In future work, we plan to explore network pruning and quantization techniques to further reduce the parameter size." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.682, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "EvINR" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.393, + 0.163 + ], + "angle": 0, + "content": "Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.176, + 0.787, + 0.237 + ], + "angle": 0, + "content": "This paper is supported by the National Natural Science Foundation of China (NSF) under Grant No. NSFC22FYT45, the Guangzhou City, University and Enterprise Joint Fund under Grant No.SL2022A03J01278, and Guangzhou Fundamental and Applied Basic Research (Grant Number: 2024A04J4072)" + }, + { + "type": "title", + "bbox": [ + 0.217, + 0.259, + 0.323, + 0.274 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.289, + 0.785, + 0.317 + ], + "angle": 0, + "content": "1. Alpsentek products. https://www.alpsentek.com/product, [Accessed: Feb. 17, 2023]" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.318, + 0.785, + 0.344 + ], + "angle": 0, + "content": "2. Pepperl fuchs event camera. https://www.pepperl-fuchs.com/global/en/classid_11544.htm?view=productgroupliterature, accessed on 17 Feb 2023" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.346, + 0.785, + 0.386 + ], + "angle": 0, + "content": "3. Bardow, P., Davison, A.J., Leutenegger, S.: Simultaneous optical flow and intensity estimation from an event camera. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 884-892 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.387, + 0.74, + 0.399 + ], + "angle": 0, + "content": "4. Bishop, C.M.: Pattern Recognition and Machine Learning. Springer (2006)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.401, + 0.785, + 0.44 + ], + "angle": 0, + "content": "5. Brandli, C., Berner, R., Yang, M., Liu, S.C., Delbruck, T.: A \\(240 \\times 180\\) 130 db \\(3\\mu s\\) latency global shutter spatiotemporal vision sensor. IEEE Journal of Solid-State Circuits 49(10), 2333-2341 (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.442, + 0.785, + 0.481 + ], + "angle": 0, + "content": "6. Cadena, P.R.G., Qian, Y., Wang, C., Yang, M.: Spade-e2vid: Spatially-adaptive denormalization for event-based video reconstruction. IEEE Transactions on Image Processing 30, 2488-2500 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.483, + 0.785, + 0.523 + ], + "angle": 0, + "content": "7. Chen, H., He, B., Wang, H., Ren, Y., Lim, S.N., Shrivastava, A.: Nerv: Neural representations for videos. Advances in Neural Information Processing Systems 34, 21557-21568 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.525, + 0.785, + 0.564 + ], + "angle": 0, + "content": "8. Chen, Y., Liu, S., Wang, X.: Learning continuous image representation with local implicit image function. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 8628-8638 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.566, + 0.785, + 0.606 + ], + "angle": 0, + "content": "9. Cook, M., Gugelmann, L., Jug, F., Krautz, C., Steger, A.: Interacting maps for fast visual interpretation. In: The 2011 International Joint Conference on Neural Networks. pp. 770-776. IEEE (2011)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.607, + 0.785, + 0.661 + ], + "angle": 0, + "content": "10. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.662, + 0.785, + 0.716 + ], + "angle": 0, + "content": "11. Ercan, B., Eker, O., Erdem, A., Erdem, E.: Evreal: Towards a comprehensive benchmark and analysis suite for event-based video reconstruction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3942-3951 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.717, + 0.785, + 0.744 + ], + "angle": 0, + "content": "12. Feng, Y., Lv, H., Liu, H., Zhang, Y., Xiao, Y., Han, C.: Event density based denoising method for dynamic vision sensor. Applied Sciences 10(6), 2024 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.745, + 0.785, + 0.799 + ], + "angle": 0, + "content": "13. Gallego, G., Delbrück, T., Orchard, G., Bartolozzi, C., Taba, B., Censi, A., Leutenegger, S., Davison, A.J., Conradt, J., Daniilidis, K., et al.: Event-based vision: A survey. IEEE transactions on pattern analysis and machine intelligence 44(1), 154-180 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.22, + 0.8, + 0.785, + 0.84 + ], + "angle": 0, + "content": "14. Gehrig, D., Gehrig, M., Hidalgo-Carrio, J., Scaramuzza, D.: Video to events: Recycling video datasets for event cameras. In: IEEE Conf. Comput. Vis. Pattern Recog. (CVPR) (June 2020)" + }, + { + "type": "list", + "bbox": [ + 0.22, + 0.289, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.369, + 0.129 + ], + "angle": 0, + "content": "Z. Wang et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.148, + 0.785, + 0.175 + ], + "angle": 0, + "content": "15. Girshick, R.: Fast r-cnn. In: Proceedings of the IEEE international conference on computer vision. pp. 1440-1448 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.176, + 0.785, + 0.202 + ], + "angle": 0, + "content": "16. Hansen, L.K., Salamon, P.: Neural network ensembles. IEEE transactions on pattern analysis and machine intelligence 12(10), 993-1001 (1990)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.203, + 0.785, + 0.229 + ], + "angle": 0, + "content": "17. He, K., Gkioxari, G., Dollár, P., Girshick, R.: Mask r-cnn. In: Proceedings of the IEEE international conference on computer vision. pp. 2961-2969 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.23, + 0.785, + 0.27 + ], + "angle": 0, + "content": "18. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 770-778 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.271, + 0.785, + 0.311 + ], + "angle": 0, + "content": "19. Hu, Y., Liu, S.C., Delbruck, T.: v2e: From video frames to realistic dvs events. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1312-1321 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.311, + 0.785, + 0.338 + ], + "angle": 0, + "content": "20. Hwang, I., Kim, J., Kim, Y.M.: Ev-nerf: Event based neural radiance field. arXiv preprint arXiv:2206.12455 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.338, + 0.785, + 0.365 + ], + "angle": 0, + "content": "21. Kim, H., Handa, A., Benosman, R., Ieng, S.H., Davison, A.J.: Simultaneous mosaicing and tracking with an event camera. J. Solid State Circ 43, 566-576 (2008)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.365, + 0.785, + 0.405 + ], + "angle": 0, + "content": "22. Kim, H., Leutenegger, S., Davison, A.J.: Real-time 3d reconstruction and 6-dof tracking with an event camera. In: European conference on computer vision. pp. 349-364. Springer (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.406, + 0.785, + 0.432 + ], + "angle": 0, + "content": "23. Kingma, D.P., Ba, J.: Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980 (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.433, + 0.785, + 0.459 + ], + "angle": 0, + "content": "24. Klenk, S., Koestler, L., Scaramuzza, D., Cremers, D.: E-nerf: Neural radiance fields from a moving event camera. IEEE Robotics and Automation Letters (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.459, + 0.785, + 0.513 + ], + "angle": 0, + "content": "25. Mescheder, L., Oechsle, M., Niemeyer, M., Nowozin, S., Geiger, A.: Occupancy networks: Learning 3d reconstruction in function space. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 4460-4470 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.514, + 0.785, + 0.554 + ], + "angle": 0, + "content": "26. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM 65(1), 99-106 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.555, + 0.785, + 0.609 + ], + "angle": 0, + "content": "27. Mueggler, E., Rebecq, H., Gallego, G., Delbruck, T., Scaramuzza, D.: The event-camera dataset and simulator: Event-based data for pose estimation, visual odometry, and slam. The International Journal of Robotics Research 36(2), 142-149 (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.61, + 0.785, + 0.649 + ], + "angle": 0, + "content": "28. Munda, G., Reinbacher, C., Pock, T.: Real-time intensity-image reconstruction for event cameras using manifold regularisation. International Journal of Computer Vision 126(12), 1381-1393 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.65, + 0.785, + 0.69 + ], + "angle": 0, + "content": "29. Niemeyer, M., Geiger, A.: Giraffe: Representing scenes as compositional generative neural feature fields. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11453-11464 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.691, + 0.785, + 0.731 + ], + "angle": 0, + "content": "30. Oechsle, M., Peng, S., Geiger, A.: Unisurf: Unifying neural implicit surfaces and radiance fields for multi-view reconstruction. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5589-5599 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.731, + 0.785, + 0.785 + ], + "angle": 0, + "content": "31. Paredes-Vallés, F., de Croon, G.C.: Back to event basics: Self-supervised learning of image reconstruction for event cameras via photometric constancy. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3446-3455 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.786, + 0.785, + 0.84 + ], + "angle": 0, + "content": "32. Park, J.J., Florence, P., Straub, J., Newcombe, R., Lovegrove, S.: Deepsdf: Learning continuous signed distance functions for shape representation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 165-174 (2019)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.148, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.683, + 0.115, + 0.732, + 0.127 + ], + "angle": 0, + "content": "EvINR" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.785, + 0.175 + ], + "angle": 0, + "content": "33. Paszke, A., Gross, S., Chintala, S., Chanan, G., Yang, E., DeVito, Z., Lin, Z., Desmaison, A., Antiga, L., Lerer, A.: Automatic differentiation in pytorch (2017)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.177, + 0.785, + 0.204 + ], + "angle": 0, + "content": "34. Pini, S., Borghi, G., Vezzani, R.: Learn to see by events: Color frame synthesis from event and rgb cameras. arXiv preprint arXiv:1812.02041 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.206, + 0.785, + 0.247 + ], + "angle": 0, + "content": "35. Rebecq, H., Ranftl, R., Koltun, V., Scaramuzza, D.: Events-to-video: Bringing modern computer vision to event cameras. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3857-3866 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.249, + 0.785, + 0.289 + ], + "angle": 0, + "content": "36. Redmon, J., Divvala, S., Girshick, R., Farhadi, A.: You only look once: Unified, real-time object detection. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 779-788 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.291, + 0.785, + 0.33 + ], + "angle": 0, + "content": "37. Reinhard, E., Stark, M., Shirley, P., Ferwerda, J.: Photographic tone reproduction for digital images. ACM Transactions on Graphics 21(3), 267-276 (2002). https://doi.org/10.1145/566654.566576" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.332, + 0.785, + 0.36 + ], + "angle": 0, + "content": "38. Rudnev, V., Elgharib, M., Theobalt, C., Golyanik, V.: Eventnerf: Neural radiance fields from a single colour event camera. arXiv preprint arXiv:2206.11896 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.361, + 0.785, + 0.388 + ], + "angle": 0, + "content": "39. Ryu, H.E.: Industrial dvs design; key features and applications. In: Conf. on Computer Vision and Pattern Recognition. vol. 3 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.39, + 0.785, + 0.431 + ], + "angle": 0, + "content": "40. Scheerlinck, C., Barnes, N., Mahony, R.: Continuous-time intensity estimation using event cameras. In: Asian Conference on Computer Vision. pp. 308-324. Springer (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.432, + 0.785, + 0.473 + ], + "angle": 0, + "content": "41. Scheerlinck, C., Rebecq, H., Gehrig, D., Barnes, N., Mahony, R., Scaramuzza, D.: Fast image reconstruction with an event camera. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 156-163 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.475, + 0.785, + 0.515 + ], + "angle": 0, + "content": "42. Scheerlinck, C., Rebecq, H., Stoffregen, T., Barnes, N., Mahony, R., Scaramuzza, D.: Ced: Color event camera dataset. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops. pp. 0-0 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.517, + 0.785, + 0.557 + ], + "angle": 0, + "content": "43. Schwarz, K., Liao, Y., Niemeyer, M., Geiger, A.: Graf: Generative radiance fields for 3d-aware image synthesis. In: Advances in Neural Information Processing Systems (NeurIPS) (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.559, + 0.785, + 0.586 + ], + "angle": 0, + "content": "44. Shiba, S., Aoki, Y., Gallego, G.: Event collapse in contrast maximization frameworks. Sensors 22(14), 5190 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.588, + 0.785, + 0.614 + ], + "angle": 0, + "content": "45. Shiba, S., Aoki, Y., Gallego, G.: Secrets of event-based optical flow. In: European Conference on Computer Vision. pp. 628-645. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.616, + 0.785, + 0.657 + ], + "angle": 0, + "content": "46. Sitzmann, V., Martel, J., Bergman, A., Lindell, D., Wetzstein, G.: Implicit neural representations with periodic activation functions. Advances in Neural Information Processing Systems 33, 7462-7473 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.659, + 0.785, + 0.699 + ], + "angle": 0, + "content": "47. Skorokhodov, I., Ignatyev, S., Elhoseiny, M.: Adversarial generation of continuous images. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10753-10764 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.701, + 0.785, + 0.741 + ], + "angle": 0, + "content": "48. Song, C., Huang, Q., Bajaj, C.: E-cir: Event-enhanced continuous intensity recovery. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7803-7812 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.743, + 0.785, + 0.784 + ], + "angle": 0, + "content": "49. Stoffregen, T., Scheerlinck, C., Scaramuzza, D., Drummond, T., Barnes, N., Kleeman, L., Mahony, R.: Reducing the sim-to-real gap for event cameras. In: European Conference on Computer Vision. pp. 534-549. Springer (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.216, + 0.786, + 0.785, + 0.84 + ], + "angle": 0, + "content": "50. Strümpler, Y., Postels, J., Yang, R., Gool, L.V., Tombari, F.: Implicit neural representations for image compression. In: Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXVI. pp. 74-91. Springer (2022)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.369, + 0.129 + ], + "angle": 0, + "content": "Z. Wang et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.788, + 0.203 + ], + "angle": 0, + "content": "51. Wang, Y., Du, B., Shen, Y., Wu, K., Zhao, G., Sun, J., Wen, H.: Ev-gait: Event-based robust gait recognition using dynamic vision sensors. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6358-6367 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.204, + 0.788, + 0.245 + ], + "angle": 0, + "content": "52. Wang, Z., Bovik, A.C., Sheikh, H.R., Simoncelli, E.P.: Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing 13(4), 600-612 (2004)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.246, + 0.788, + 0.286 + ], + "angle": 0, + "content": "53. Weng, W., Zhang, Y., Xiong, Z.: Event-based video reconstruction using transformer. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 2563-2572 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.287, + 0.788, + 0.342 + ], + "angle": 0, + "content": "54. Yadav, G., Maheshwari, S., Agarwal, A.: Contrast limited adaptive histogram equalization based enhancement for real time video system. In: 2014 international conference on advances in computing, communications and informatics (ICACCI). pp. 2392-2397. IEEE (2014)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.343, + 0.788, + 0.383 + ], + "angle": 0, + "content": "55. Yang, M., Liu, S.C., Delbruck, T.: A dynamic vision sensor with \\(1\\%\\) temporal contrast sensitivity and in-pixel asynchronous delta modulator for event encoding. IEEE Journal of Solid-State Circuits 50(9), 2149-2160 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.384, + 0.788, + 0.424 + ], + "angle": 0, + "content": "56. Zhang, K., Li, Y., Zuo, W., Zhang, L., Van Gool, L., Timofte, R.: Plug-and-play image restoration with deep denoiser prior. IEEE Transactions on Pattern Analysis and Machine Intelligence 44(10), 6360-6376 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.425, + 0.788, + 0.467 + ], + "angle": 0, + "content": "57. Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual metric. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 586-595 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.468, + 0.788, + 0.508 + ], + "angle": 0, + "content": "58. Zhang, X., Srinivasan, P.P., Deng, B., Debevec, P., Freeman, W.T., Barron, J.T.: Nerfactor: Neural factorization of shape and reflectance under an unknown illumination. ACM Transactions on Graphics (TOG) 40(6), 1-18 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.509, + 0.788, + 0.536 + ], + "angle": 0, + "content": "59. Zhang, Z., Yezzi, A., Gallego, G.: Image reconstruction from events. why learn it? arXiv preprint arXiv:2112.06242 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.537, + 0.788, + 0.578 + ], + "angle": 0, + "content": "60. Zhang, Z., Yezzi, A.J., Gallego, G.: Formulating event-based image reconstruction as a linear inverse problem with deep regularization using optical flow. IEEE Transactions on Pattern Analysis & Machine Intelligence 45(07), 8372-8389 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.579, + 0.788, + 0.619 + ], + "angle": 0, + "content": "61. Zhu, A.Z., Thakur, D., Ozaslan, T., Pfrommer, B., Kumar, V., Daniilidis, K.: The multivehicle stereo event camera dataset: An event camera dataset for 3d perception. IEEE Robotics and Automation Letters 3(3), 2032-2039 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.62, + 0.788, + 0.646 + ], + "angle": 0, + "content": "62. Zhu, A.Z., Yuan, L., Chaney, K., Daniilidis, K.: Ev-flownet: Self-supervised optical flow estimation for event-based cameras. arXiv preprint arXiv:1802.06898 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.647, + 0.788, + 0.688 + ], + "angle": 0, + "content": "63. Zhu, A.Z., Yuan, L., Chaney, K., Daniilidis, K.: Unsupervised event-based learning of optical flow, depth, and egomotion. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 989-997 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.689, + 0.788, + 0.743 + ], + "angle": 0, + "content": "64. Zhu, L., Wang, X., Chang, Y., Li, J., Huang, T., Tian, Y.: Event-based video reconstruction via potential-assisted spiking neural network. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3594-3604 (2022)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.788, + 0.743 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/Revisit Event Generation Model_ Self-Supervised Learning of Event-to-Video Reconstruction with Implicit Neural Representations/a41e408e-cbb3-4c38-85c3-298d71f3cf9d_origin.pdf b/2024/Revisit Event Generation Model_ Self-Supervised Learning of Event-to-Video Reconstruction with Implicit Neural Representations/a41e408e-cbb3-4c38-85c3-298d71f3cf9d_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..7984a7453a43e6f8c7a3c6f10a1719d2538a93f5 --- /dev/null +++ b/2024/Revisit Event Generation Model_ Self-Supervised Learning of Event-to-Video Reconstruction with Implicit Neural Representations/a41e408e-cbb3-4c38-85c3-298d71f3cf9d_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:aeee88a68d15edac99f36ef5f9dc86ef7c3f7d671a18bfcfbd0d3994057fa948 +size 1459840 diff --git a/2024/Revisit Event Generation Model_ Self-Supervised Learning of Event-to-Video Reconstruction with Implicit Neural Representations/full.md b/2024/Revisit Event Generation Model_ Self-Supervised Learning of Event-to-Video Reconstruction with Implicit Neural Representations/full.md new file mode 100644 index 0000000000000000000000000000000000000000..8511aee3c9a46b6bd29d9e739d91c57fc835d9db --- /dev/null +++ b/2024/Revisit Event Generation Model_ Self-Supervised Learning of Event-to-Video Reconstruction with Implicit Neural Representations/full.md @@ -0,0 +1,351 @@ +# Revisit Event Generation Model: Self-Supervised Learning of Event-to-Video Reconstruction with Implicit Neural Representations + +Zipeng Wang1, Yunfan Lu1, and Lin Wang1,2 + +1 Artificial Intelligence Thrust, HKUST(GZ) + +$^{2}$ Dept. of Computer Science and Engineering, HKUST + +{zwang253,ylu066}@connect.hkust-gz.edu.cn, linwang@ust.hk + +Abstract. Reconstructing intensity frames from event data while maintaining high temporal resolution and dynamic range is crucial for bridging the gap between event-based and frame-based computer vision. Previous approaches have depended on supervised learning on synthetic data, which lacks interpretability and risk over-fitting to the setting of the event simulator. Recently, self-supervised learning (SSL) based methods, which primarily utilize per-frame optical flow to estimate intensity via photometric constancy, has been actively investigated. However, they are vulnerable to errors in the case of inaccurate optical flow. This paper proposes a novel SSL event-to-video reconstruction approach, dubbed EvINR, which eliminates the need for labeled data or optical flow estimation. Our core idea is to reconstruct intensity frames by directly addressing the event generation model, essentially a partial differential equation (PDE) that describes how events are generated based on the time-varying brightness signals. Specifically, we utilize an implicit neural representation (INR), which takes in spatiotemporal coordinate $(x,y,t)$ and predicts intensity values, to represent the solution of the event generation equation. The INR, parameterized as a fully-connected Multi-layer Perceptron (MLP), can be optimized with its temporal derivatives supervised by events. To make EvINR feasible for online requisites, we propose several acceleration techniques that substantially expedite the training process. Comprehensive experiments demonstrate that our EvINR surpasses previous SSL methods by $38\%$ w.r.t. Mean Squared Error (MSE) and is comparable or superior to SoTA supervised methods. Project page: https://vlislab22.github.io/EvINR/. + +# 1 Introduction + +Event cameras [13,55] are novel sensors that offer numerous advantages over traditional frame-based cameras, including low power consumption, high dynamic range (HDR), and high temporal resolution [34]. However, their unique imaging paradigm presents a challenge when applying vision algorithms designed for frame-based cameras. To address this challenge and bridge the gap between + +![](images/9271eebe877c259690f453580b09094e8509cd9bbac6be9b1733ab51887f8a91.jpg) +Event Generation Model +Fig. 1: Connection between event generation model and EvINR: Event generation model reveals the relation between discrete events and continuous temporal intensity changes, described as the event generation equation (Eq. 4). EvINR utilizes an INR to solve Eq. 4 and recovering a continuous function of intensity w.r.t. time, implicitly parameterized with a fully connected MLP. + +event-based and standard computer vision [10, 15, 17, 18, 36], many methods have been proposed to reconstruct intensity frames from events. + +Early methods primarily rely on hand-crafted integrators or filters, resulting in significant loss of detail in reconstructed results [3,22,28,40,59]. More recently, deep learning-based methods [6,35,41,49,53,64] have emerged, typically supervised by large-scale synthetic datasets generated using an event simulator [14,19]. However, the interpretability of these methods is limited due to the 'black-box' nature of the neural networks. Furthermore, the performance of these methods is constrained by the domain gap between synthetic and real-world data [49], leading to suboptimal results if the event simulator's settings do not accurately match those of the inference data. To address these issues, some researchers have explored self-supervised learning (SSL) frameworks [31,60], aiming to eliminate the dependency on labeled synthetic data. Nonetheless, these approaches still depend on event-based optical flow estimation, which is prone to over-fitting, occlusions, and non-convergence issues [44, 45]. As a result, the reconstructed frames often suffer from the loss of textural details and various artifacts. + +To address these challenges, we revisit the event generation model [13], which forms a fundamental link between events and intensity. The event generation model can be expressed as a partial differential equation (PDE), describing how events are triggered by logarithmic intensity changes that exceed a certain threshold. This PDE, known as the event generation equation, establishes a direct connection between discrete events and the partial derivative of the continuous intensity function w.r.t. time. Our key insight is that, solving the event generation equation offers an ideal self-supervised solution for event-to-video reconstruction, thereby eliminating the need for synthetic data or optical + +flow estimation. However, several properties of event data make solving the equation nontrivial: 1) events can be triggered at extremely high frequencies (up to $10^{8} / s$ ), which results in a large volume of data and a heavy computational burden. 2) events are inherently noisy, particularly in extreme visual conditions [51], which pose challenge to the robustness of solvers. 3) events do not capture the initial intensity, which makes determining the boundary values challenging. + +Recently, implicit neural representations (INRs) have gained popularity for solving the inverse problems in 3D reconstruction or image super-resolution [8, 26,46] by parameterizing complex signals via deep neural networks. In this work, we find that INRs possess several key advantages that render them particularly suitable for solving the event generation equation: they inherently accommodate a large volume of event data, exhibit high noise tolerance, and are flexible to add additional loss terms to regulate initial value. + +In light of this, we propose a novel SSL framework, termed EvINR, that employs an INR to represent the solution of the event generation equation. Our EvINR can be directly optimized by minimizing the residual between its temporal derivatives and local intensity changes estimated from event data based on the event generation model (Sec. 3.2). Moreover, we incorporate a spatial regularization term that regularizes the relative values of adjacent pixels by constraining the magnitude of spatial gradients, which effectively reduces noise in the reconstruction process (Sec. 3.2). Although the basic implementation of EvINR yields acceptable results, its convergence on seconds-long event sequences takes minutes, limiting its real-world applicability, especially in online scenarios. To expedite EvINR's training process, we introduce several acceleration techniques, including frame-based optimization, coarse-to-fine training, and model ensembling. These approaches reduce the training time from minutes to seconds while not compromising the reconstruction quality (Sec. 3.4). + +Moreover, most approaches are typically evaluated on event datasets captured by DAVIS sensors [27, 42, 49, 61], making it difficult to evaluate their stability and robustness to other types of event camera [1, 2, 39]. For this reason, we collected a new event dataset using an ALPIX-Eiger event camera [1], with well-aligned events and intensity frames. + +In summary, our paper makes three key contributions: (I) We propose EvINR, a concise SSL framework that solves the event generation equation via implicit neural representations for event-to-video reconstruction. (II) Our EvINR substantially outperforms previous SSL methods [31, 40, 59] and attains comparable, or even superior, performance compared to state-of-the-art supervised methods [35, 49, 53]. (III) We collect a real-world dataset with an ALPIX-Eiger event camera, complementary to the datasets captured by DAVIS cameras. + +# 2 Related Works + +Implicit Neural Representations (INRs) have emerged as a powerful tool for parameterizing signals, such as images, videos, and audio, in a continuous manner using neural networks [46]. Compared to traditional discrete sig- + +![](images/8b29df90046efc81ca4d72191edd630d2ab8ec1e1cd441d98a91e34a3d772a8e.jpg) +Fig. 2: Overview of EvINR. A fully connected MLP is used to implicitly solve the event generation equation. The temporal gradient of the MLP is supervised by temporal intensity changes of events, and the spatial gradient is penalized to reduce noise. + +nal representations, INRs offer the advantage of being able to be sampled at arbitrary resolutions with fixed memory requirements. As a result, INRs have found widespread applications in various fields, including 3D scene representation [25, 26, 32], video representation [7], generative models [29, 43, 47], and model compression [50]. A unique property of INRs is that they can be effectively learned from the derivatives of signals, such as the normals of 3D shapes [30,46,58] and the gradients of images [46]. Such a characteristic has motivated us to approach the task of event-to-video reconstruction by optimizing the INR of video from its temporal derivatives. Previous research has investigated the potential of INR for novel view synthesis using event data [20,24,38]. These approaches reconstruct 3D neural radiance fields using multiple event sequences with known camera poses from a stationary scene. Most similar to our work, E-CIR [48] uses polynomial to regress the intensity function, which may only represent a short time interval (e.g., the exposure time of an image). In contrast, our method employs INR, which can represent longer time of intensity change. + +Event-based Video Reconstruction has been a hot topic in the literature. Early attempts [3, 9, 21] address this problem based on the photometric constancy, which describes the relationship between intensity gradients and optical flow. Other approaches [28, 40] are based on direct event integration without estimating optical flow. Rebecq et al. [35] developed the first DL-based framework, called E2VID, to reconstruct intensity frames from events in an end-to-end manner, outperforming earlier techniques by a significant margin. E2VID was updated by some following research, trying to tackle problems of inference speed [41], the cold start problem [6], and training strategy [49]. These methods are supervised and learned using the synthetic dataset obtained via the event simulators [14, 19] due to the lack of real-world datasets with well-aligned events and intensity frames. Therefore, the generalization capability of these methods is limited by the simulator-to-real gap [49]. + +To address this problem, Federico et al. [31] proposed an SSL framework with a network to estimate optical flow and another network to reconstruct intensity + +frames based on the photometric constancy. Zhang et al. [60] updated this idea into a linear inverse problem that can be solved using modern linear solvers without using deep learning. However, these methods either assume optical flow is known or estimate the optical flow using contrast maximization [62,63]. Consequently, their performance cannot be guaranteed unless high-quality optical flow can be obtained. In contrast, our approach is based solely on the physical event generation model, leading to a more straightforward solution that also demonstrates significantly improved performance and greater flexibility. + +# 3 Method + +Overview: Our objective is to reconstruct intensity frames from events in a self-supervised manner, without the need for end-to-end training or optical flow estimation. To accomplish this, we have reformulated event-based video reconstruction as solving the event generation equation. We employ an INR that is supervised solely by event data to represent the intensity function. An overview of our approach can be seen in Fig.2. In Sec. 3.1, we explain the event generation model and its connection to event-based video reconstruction. In Sec. 3.2, we detail how to train an EvINR by supervising its spatial and temporal gradients. Techniques to speed up the training of EvINR for online applications are discussed in Sec. 3.3. Finally, in Sec. 3.4, we describe our collected dataset using the ALPIX event camera [1]. + +# 3.1 Preliminary: Event Generation Model + +We begin by providing a brief overview of the event generation model [13], which forms the theoretical foundation for our approach. Let $I(x,y,t)$ denote the intensity of the spatial location $(x,y)$ at time $t$ in a video. Since event cameras operate with logarithmic intensity, we denote $L(x,y,t) = \log I(x,y,t)$ . + +An event camera comprises a frame of independent pixels that respond to changes in the logarithmic intensity signal and produce sequences of sparse and asynchronous events. An event $(x_{i},y_{i},t_{i},p_{i})$ is triggered when the logarithmic intensity change surpasses a threshold $C$ since the previous event was triggered at the same pixel. where $(x,y)$ is the spatial location of the pixel, $t$ is the timestamp, and $p\in \{-1,1\}$ is the polarity of the logarithmic intensity change. + +For simplicity, let us consider the temporal changes in logarithmic intensity of a single pixel with fixed spatial coordinates and disregard the spatial terms $(x,y)$ . We can describe an event $e_i$ on that pixel using the Dirac delta function as follows: + +$$ +e _ {i} (t) = p _ {i} \cdot C \cdot \delta \left(t - t _ {i}\right). \tag {1} +$$ + +Therefore, the logarithmic intensity increment $\Delta L = L(t_2) - L(t_1)$ in a time interval $[t_1,t_2]$ can be represented by the accumulation of events, which can be expressed as: + +$$ +\Delta L = \int_ {t _ {1}} ^ {t _ {2}} \sum_ {i} e _ {i} (t) \mathrm {d} t. \tag {2} +$$ + +Assuming a short time interval and ignoring noise, we can approximate the logarithmic intensity increment by its first-order temporal derivative using Taylor expansion: + +$$ +\frac {\Delta L}{t _ {2} - t _ {1}} = \frac {\partial L ((t _ {1} + t _ {2}) / 2)}{\partial t}. \tag {3} +$$ + +By substituting Eq. 2 into Eq. 3, we derive the event generation equation (Eq. 4), which bridges the discrete event data with the continuous temporal derivatives of logarithmic intensity: + +$$ +\frac {\partial L \left(\left(t _ {1} + t _ {2}\right) / 2\right)}{\partial t} = \frac {1}{t _ {2} - t _ {1}} \int_ {t _ {1}} ^ {t _ {2}} \sum_ {i} e _ {i} (t) \mathrm {d} t. \tag {4} +$$ + +The left-hand side of Eq. 4 represents the partial derivative of logarithmic intensity with respect to time, while the right-hand side can be calculated by accumulating events. Intuitively, event-based video reconstruction can be formulated as solving Eq. 4, i.e., finding a logarithmic intensity function that satisfies the equation. However, as stated earlier, solving Eq. 4 can be non-trivial due to the large amount, high noise rate of event data, and unknown boundary intensity values. To address these challenges, we employ INRs to solve Eq. 4 as they naturally scale up to large data and have high noise tolerance [20]. Boundary intensity values can also easily be regularized by injecting natural image priors as additional loss functions. + +# 3.2 Learning INRs from Events + +We aim to learn an INR $F_{\Theta}$ given an event stream $\{e_i\}_{i=0}^N$ . Here, $\Theta$ represents the parameters of a fully connected MLP that predicts the logarithmic intensity value $\hat{L}$ at any spatiotemporal coordinate $(x,y,t)$ . We adapt the event frame representation [35] and stack the given events uniformly into $T$ event frames. + +Temporal Supervision The INR is optimized by directly minimizing the temporal loss between the predicted logarithmic intensity change $\Delta \hat{L}$ and the logarithmic intensity change estimated by event accumulation $\Delta L$ . Here $\Delta \hat{L}$ can be derived by accumulating events as Eq. 2 and $\Delta \hat{L}$ is the change of $\hat{L}$ with respect to $t$ obtained by performing double back-propagation [33] of $F_{\Theta}$ : + +$$ +\Delta \hat {L} = \frac {\partial F _ {\Theta} \left(\left(t _ {1} + t _ {2}\right) / 2\right)}{\partial t} \cdot \left(t _ {2} - t _ {1}\right), \tag {5} +$$ + +Here we adopt the mean squared error (MSE) loss for temporal supervision: + +$$ +\mathcal {L} _ {\text {t e m p}} = \left(\Delta L - \Delta \hat {L}\right) ^ {2}, \tag {6} +$$ + +where $\Delta L$ and $\Delta \hat{L}$ are given by Eq. 2 and Eq. 5, respectively. + +Spatial Regularization Although the temporal supervision can estimate the intensity function, the results may still contain unnatural artifacts as the INR network $F_{\Theta}$ has no prior knowledge of the initial intensity values of each pixel. To address this issue, we introduce a spatial regularization term that encourages the solution to be in the space of natural-looking images. We adopt Tikhonov regularization [4] that penalizes the spatial gradients of logarithmic intensity: + +$$ +\mathcal {L} _ {\text {r e g}} = \left(\frac {\partial F _ {\Theta}}{\partial x}\right) ^ {2} + \left(\frac {\partial F _ {\Theta}}{\partial y}\right) ^ {2}. \tag {7} +$$ + +It is worth noting that we do not employ more complex regularization methods, such as CNN denoisers [56] as used in [60], because they typically utilize large network models that can significantly slow down the training process. + +Optimizing EvINR The overall objective is given by the equation: + +$$ +\mathcal {L} = \mathcal {L} _ {\text {t e m p}} + \lambda \mathcal {L} _ {\text {r e g}}, \tag {8} +$$ + +where $\mathcal{L}_{\mathrm{temp}}$ and $\mathcal{L}_{\mathrm{reg}}$ were introduced in Sec. 3.2 and Sec. 3.2, respectively, as the temporal supervision and spatial regularization term. $\lambda$ is a hyper-parameter used to adjust the weight of the spatial regularization term. + +Tone-mapping The output of the EvINR is the logarithmic intensity of the reconstructed frames. We first use the exponential function to convert the predicted logarithmic intensity values to high dynamic range (HDR) intensity values $I \in [0, \infty)$ : + +$$ +I (x, y, t) = \exp \left(F _ {\Theta} (x, y, t)\right). \tag {9} +$$ + +Then, we adopt the Reinhard function [37] to map the HDR intensity values into the low dynamic range $[0,1]$ : + +$$ +\Gamma (I) = \left(\frac {I}{I + 1}\right) ^ {\gamma}. \tag {10} +$$ + +where $\gamma$ is a hyper-parameter used to control the contrast of $\Gamma(I)$ . We assess our reconstruction performance by applying $\Gamma(I)$ in all experiments. + +# 3.3 Accelerating EvINR + +While solving the event generation equation using the basic implementation of EvINR, as described in Sec. 3.3, yields satisfactory reconstruction results, it is not efficient enough for online tasks. Optimizing one INR network takes about 20 minutes, and a network can only represent approximately 1 second of a sequence. Reducing the training time or increasing the sequence time leads to severe performance degradation. To enable online usage for EvINR, we propose several techniques to reduce the training time and increase the representation capacity for EvINR, as illustrated in Fig. 3. + +![](images/0bce85b79b74465683d976941653301334345c5a6e7cba7e8b8909d6a51a5a9d.jpg) +Fig. 3: Overview of acceleration techniques. (a) and (b) illustrate the difference between the basic coordinate-based and our frame-based optimization, respectively. (c) depicts our proposed coarse-to-fine training scheme and network ensembling technique. + +![](images/1657c628445946bdf53a0eef143f623a836bf6a1b9764428b9c3c28a195e7e2a.jpg) + +Coordinate-based to frame-based optimization: The basic EvINR algorithm is designed to learn a mapping from 3D coordinates $(x,y,t)$ to the logarithmic intensity $L$ . However, this coordinate-based optimization requires the network to 'remember' $H\times W\times T$ coordinate-to-logarithmic intensity mappings, which can be highly complex and lead to slow convergence. To address this issue, we propose a frame-based optimization scheme that learns a mapping from timestamps $t$ to the logarithmic intensity of all pixels at that timestamp. The frame-based optimization scheme can be represented as: + +$$ +F _ {\Theta} (t) = \left[ L (i, j, t) \right] _ {1 \leq i \leq H, 1 \leq j \leq W}, \tag {11} +$$ + +where the right hand-side represents a matrix that contains all logarithmic intensity of all pixels at $t$ . The frame-based optimization approach requires the network to remember only $T$ coordinate-to-logarithmic intensity mappings, resulting in a significant reduction in complexity and faster convergence. Empirical results show that this approach reduces the convergence time by two orders of magnitude, from minutes to seconds. The comparison between coordinate-based and frame-based optimizations is illustrated in Fig. 3 (a) and (b). + +Coarse-to-fine training: We adopt a coarse-to-fine training scheme, which enables the network to learn overall logarithmic intensity changes before focusing on finer details. We structure the training process into $s$ distinct stages, at each of which we increase the temporal resolution by a factor of 2. Initially, we divide the event sequence into $N$ segments of equal length and optimize the EvINR in accordance with Eq. 6 over a specified number of iterations. Subsequently, we proceed to bisect each of these $N$ segments into two smaller ones, ensuring that they contain an equal number of events. For all the experiments, we set the number of stages $s$ to 3 and scheduled the upsampling to occur after 100 and 200 iterations, respectively. This approach accelerates training by reducing the number of event frames needed in the early stages by approximately 2 times. + +Network ensembling: As each EvINR network requires only about 1GB of GPU memory, we further leverage network ensemble techniques [16] to train $N$ + +EvINR networks simultaneously on a single GPU to achieve higher parallelism. This approach enables us to exploit the computational resources of modern GPUs more effectively and speed up the training process. We keep overlap periods between nearby networks to keep logarithmic intensity predictions constant among all networks. Fig. 3(c) depicts the coarse-to-fine training process and network ensembling techniques employed in our approach. + +# 3.4 Dataset Collection + +Most event-to-video reconstruction approaches are typically evaluated on event datasets collected using DAVIS sensors [5], such as IJRR [27], HQF [49], MVSEC [61], and CED [42]. Recently, other types of event cameras [1, 2, 39] have been developed, which share the same event generation model with DAVIS sensors but may differ in detailed configurations and settings. Therefore, it is essential to verify the generalization capability of existing methods and our EvINR on those new types of event sensors. + +To fill the gap, we introduce a new real-world dataset, called the ALPIX Event Dataset (AED), which is collected using an ALPIX-Eiger event camera [1], featuring static scenes accompanied by gradual camera motions. The camera provides well-aligned RGB frames and color events. The RGB frames have a resolution of $3264 \times 2448$ and the events have a resolution of $1632 \times 1224$ . The AED dataset includes seven video sequences with diverse scenes, such as streets, buildings, indoor scenes, textures, and tools, and each approximately lasts for ten seconds. Note that, in this paper, we only focus on reconstructing grayscale frames to keep consistency with previous works [31, 60] and also for a fair comparison. Therefore, we first demosaic the RGB frames and events according to the Quad Bayer pattern, resulting in intensity frames with a resolution of $816 \times 612$ and event data with a resolution of $408 \times 306$ . Details on the post-processing of AED dataset can be found in the supplementary material. + +# 4 Experiments + +# 4.1 Experiments Settings + +Datasets: We conduct experiments to evaluate the effectiveness of our proposed method using three datasets: IJRR [27], HQF [49], and our AED dataset. The IJRR dataset consists of intensity frames and events in 25 real scenes and 2 synthetic scenes, captured by a DAVIS240C camera [5]. HQF dataset provides 14 event data sequences captured with two DAVIS240C cameras, delivering well-exposed and clear intensity frames. The spatial resolution of both the IJRR and HQF datasets is $240 \times 180$ . The AED dataset contains 7 event sequences with a resolution of $408 \times 306$ . More details of the exact time split can be found in the supplementary material. + +Evaluation Metrics: We evaluate the efficacy of our method using several image quality metrics, including mean squared error (MSE), structural similarity (SSIM) [52], and learned perceptual image patch similarity (LPIPS) [57]. + +![](images/b88b076fc57a82297a1633d086d37ebbae2326e362e773ae4316c839a17975f9.jpg) +Fig. 4: Qualitative comparison with baseline methods on IJRR(Row 1&2), HQF (Row 3&4) and AED(Row 5&6). + +Implementation Details: We use a SIREN network [46] with three hidden layers and 512 neurons per layer as our INR network. We partition each event sequence into sub-sequences lasting 5 seconds each and concurrently train all subsequences using model ensembling for 300 iterations. We first stack the events within $\frac{1}{32}$ second into an event frame and temporal upsample the event frames in 100 and 200 iterations, as described in Sec. 3.3. We adopt the Adam optimizer [23] with a learning rate of $1\mathrm{e} - 4$ and exponentially reduce the learning rate every 10 iterations with a decay rate of 0.95. The weight of spatial regularization $\lambda$ is set to 0.05, and $\gamma$ in Eq. 10 is set to 0.6. The activated threshold $C$ is set to 1 for the IJRR and HQF datasets and 0.25 for the AED dataset. The training process takes approximately 8 seconds on a single RTX3090 GPU. + +# 4.2 Evaluation of Video Reconstruction + +We assess the effectiveness of our approach by comparing it against eight state-of-the-art (SoTA) methods, classified based on the amount of data required for training. The methods are categorized into supervised learning (SL) methods that utilize synthetic ground-truth intensity frames for supervision and self-supervised learning (SSL) methods that rely solely on event data. Specifically, we + +compare against: 1) Five SL methods: FireNet [41], E2VID [35], FireNet+ [49], E2VID+ [49], and ET-Net [53]. 2) Three SSL method: SSL-E2VID [31], HF [40] and ELRP [60]. Reconstructed results for all methods were generated at each timestamp of the intensity frame. We use optical flow estimation from the FlowNet of SSL-E2VID [31] for ELRP [60]. We apply Contrast Limited Adaptive Histogram Equalization (CLAHE) [54] to both ground truth and synthesized frames before evaluation, following [31]. Note that the quantitative results for FireNet [41], E2VID [35], FireNet+ [49], E2VID+ [49], and ET-Net [53] are referred from a recent benchmark paper [11]. For the visual comparison, we obtained the visualization results by downloading the publicly available checkpoints and test them in the local environment. + +Table 1: Comparison of quantitative results on the IJRR, HQF, and AED datasets. Bold values indicate the best results among all methods, while underlined values indicate the best results among SSL methods. + +
MethodsIJRRHQFAED
MSESSIMLPIPSMSESSIMLPIPSMSESSIMLPIPS
SLFireNet [41]0.1310.5020.3200.0940.5330.4410.0740.2980.579
E2VID [35]0.2120.4240.3500.1270.5400.3820.0560.4240.500
FireNet+ [49]0.0630.5550.2900.0400.6140.3140.0940.2320.566
E2VID+ [49]0.0700.5600.2360.0360.6430.2520.0740.3450.462
ET-Net [53]0.0470.6170.2240.0320.6580.2600.0840.3120.482
SSLSSL_E2VID [31]0.0970.4730.4090.0700.4800.4640.0940.3160.453
HF [40]0.1640.3340.6580.1330.2320.6700.0800.2400.943
ELRP [60]0.0800.4370.4850.0740.4500.4740.0840.3050.473
Ours0.0470.6280.2510.0480.5310.3330.0670.4580.366
+ +The quantitative results are presented in Table 1. Our method demonstrates superior performance compared to the best SL methods, with improvements of $7\%$ , $13\%$ , and $4\%$ in terms of MSE, SSIM, and LPIPS respectively on the IJRR dataset [27]. Although the gap between our method and SL methods widens on the HQF dataset [49] due to the lower event density, our method remains comparable. On the AED dataset, our method outperforms the state-of-the-art SL methods, E2VID+ and ET-Net, by a clear margin for all three metrics. Notably, our method shows a significant improvement over previous SSL methods. In particular, compared with SSL-E2VID, it improves MSE, SSIM, and LPIPS by $35\%$ , $25\%$ , and $21\%$ , respectively. + +A qualitative comparison of our method with the baseline methods is depicted in Fig. 4. Our method produces intensity frames with better contrast and overall visual quality, compared with other methods that suffer from issues such as foggy effects, artifacts, and loss of detailed structures. It is worth noting that the performance of E2VID+ and ET-Net degrades significantly on the AED dataset, which suggests that their training strategy may overfit the DAVIS sensor setting. Additional results can be found in the supplementary material. + +![](images/f7a9257782005390f2966a919666006df291eb6866ff46bc08560ada41cdefdf.jpg) +Ground Truth + +![](images/0ba6338b8a4ac641f5742e6d3e509e7169edf9069876090374e97d3dfd0be88c.jpg) +No Reg. +Fig. 5: The impact of removing the spatial regularization and coarse-to-fine training. + +![](images/c940bc1649a90698a9eac470e635f639ce71cc57ba404872ea2f51d56821a8d3.jpg) +No $C2T$ + +![](images/4b3f6d33b3d4a79d16b6af3d999865c30834d84df3a12cfecb774900ea55ec5e.jpg) +Full Model + +Table 2: Ablation of the spatial regularization term (Reg.) and coarse-to-fine training $(C2F_{\cdot})$ on the IJRR [27] dataset. + +
MSESSIMLPIPSTime(s)FPS
Base0.0630.5420.35717.2913.77
Base+Reg.0.0440.6120.27517.4413.64
Base+C2F.0.0610.5660.3437.9431.33
Full Model0.0470.6580.2518.0829.45
+ +# 4.3 Ablation Study + +We conduct ablation experiments on the IJRR dataset to assess the significance of each individual component of our method, and the results are presented in Tab. 2. Our method's basic implementation, which doesn't include spatial regularization term as described in Sec. 3.2 or use coarse-to-fine training as described in Sec. 3.3, is called Base. We also evaluate the performance of two modified versions of the full model: $\text{Base} + \text{Reg}$ , which adds a spatial regularization term (Eq. 7) with $\lambda = 0.05$ , and $\text{Base} + C2F$ , which uses coarse-to-fine training. Experimental results show that both spatial regularization and coarse-to-fine training contribute to the improved performance of our full model. Fig. 5 confirms the effectiveness of these two modules, showing that removing the spatial regularization term introduces noticeable noise (e.g., in the background), while removing the coarse-to-fine training scheme leads to an increase in training time and missing details (e.g., in the people's clothes and the book). + +# 4.4 Discussions + +Training speed: To assess the training speed of EvINR, we conducted an analysis using various model ensembling configurations on the whole Calibration sequence from the IJRR dataset [27]. We divide the event sequence, which had a total duration of 50 seconds, into $N$ partitions, with each partition having a duration of $\tau$ seconds. Subsequently, we trained $N$ EvINR models in parallel, incorporating the model ensembling techniques outlined in Sec. 3.4. The experimental findings are presented in Table 3. By training 10 EvINR models, each corresponding to 5 seconds of the event sequence, we achieved a performance gain of over $26\%$ , while only requiring an additional $14\%$ of training time compared to training a single EvINR model on the entire sequence. However, we observed a significant drop in performance when increasing the partition size to + +Table 3: Impact of hyper-parameters of model ensembling on the training speed and reconstruction performance. + +
NτMSESSIMLPIPSTime(s)FPS
1500.0840.460.38015.7175.66
5100.06880.6000.25316.9170.31
1050.06250.6120.24418.3264.90
5010.08750.4880.34329.4340.39
+ +50, suggesting that a 1-second sequence is insufficient for EvINR to converge to a stable solution of the event generation equation. We also note that the optimal choice of hyper-parameters may differ for different event sequences due to the per-scene optimization nature of our INR approach. + +Event enhancement: Our approach provides a smooth and continuous representation of event data by representing events triggered within a small time window $\Delta t$ as $\frac{\partial F_{\Theta}}{\partial t}\Delta t$ . This event representation automatically reduces noise and preserves critical information through INR optimization, making it highly robust to noise. Fig. 6 compares the denoising results of our INR representation with several SoTA denoising methods [12, 51]. + +![](images/8032d3c12d165de605fa1625b9fa5542f9857961e8e68a34622950fe860b7716.jpg) + +![](images/7d1ac315e7e67e7471f35e452d773201290c823b840a7b64ecf2033f15067ebb.jpg) +Event Frame + +![](images/4b506b332bff3bcdee37dfb10d1054f532c888f8b80af291913e10bb45257318.jpg) + +![](images/39a772ba0b3e40f480c2aaea9a8513d99950c45bb8ef43c37455a23c3e7bb1b0.jpg) +DF + +![](images/1619bce478bbddd37d6b8151767cf916315f343e5f9d286d3f61a5aeea8591cb.jpg) + +![](images/a4710b8d3db01e9bf2f8041c897cbe2cac524d9278a7a4ee6a346d78a05f7848.jpg) +EvGait + +![](images/9d82d6ca8bbd68704ad5e127b71bf7764d211f739eadb2178620b7bd99c94080.jpg) +EvINR +Fig. 6: Qualitative comparison with baseline methods of event stream denoising. + +Early frames reconstruction: [6] suggested that RNN-based techniques for video reconstruction, e.g. E2VID [35], require an initialization period to achieve satisfactory results. Consequently, the initial frames generated by these methods may be of poor quality, restricting their usefulness to short event sequences. Conversely, our proposed approach is capable of producing realistic outcomes using a minimal number of events. Fig. 7 illustrates a comparison of the first frame generated by our method and RNN-based techniques, demonstrating our ability to rapidly create high-quality outputs with minimal input. + +Inference speed: Compared to other SoTA approaches, our method offers a significant improvement in terms of inference speed, as demonstrated in Tab. 4. The main reason for this improvement is that other methods require converting event data into event frames or voxel grids during the inference period, which + +![](images/c35390e0359931825e3b4039a2fd20f9b91363e11be62017c32f36dd25f9aefd.jpg) +E2VID[34] + +![](images/2dfe2d1cc82bb52f4b5921891797685b286cfcba654471c19da9a619e1c6f688.jpg) +E2VID+[47] + +![](images/3730b3787cf9da2d3280bdb17a773f1d521a9cf1c60450669f2456005a04886a.jpg) +ET-Net[51] + +![](images/42df6c93f0106a00743e6713913223fe206d95d2a24985799ef5631daaf62778.jpg) +SSL-E2VID[30] + +![](images/b1fb2954e7b8627e18580774eae86cbc05ee4ed39006f3f28b188cba67cb09ff.jpg) +EvINR + +![](images/4591044b040d257a1814d11a0bea96ca0867343a5f7f537aef2b2e855115169c.jpg) +GT +Fig. 7: The reconstructed initial frame from different compared methods. + +Table 4: Comparison of inference time in terms of frames-per-second (fps) at two resolutions from IJRR [27] and AED dataset respectively. + +
ResolutionFireNetE2VIDET-NetOurs
(240, 180)118.24100.9636.1178.19
(408, 306)28.8328.5618.2755.37
+ +adds significant overhead. However, our method only requires the spatiotemporal coordinates as input, thus avoiding this issue. + +# 5 Conclusion + +This paper introduced EvINR, a SSL method for event-to-video reconstruction that relieves the need for synthetic data or optical flow estimation. We, for the first time, show that high-quality videos can be reconstructed in a self-supervised and interpretable way without time-consuming end-to-end training. Our method is based on directly solving the event generation model via optimizing an INR whose temporal derivative is self-supervised by events and spatial derivative is regularized to reduce artifacts. Additionally, we propose several acceleration techniques that significantly reduce the training time of EvINR, making it applicable for online tasks. Experiments show that our approach significantly outperforms the previous SSL methods, and is competitive with the SoTA supervised methods. Our approach also demonstrates superior interpretability and robustness to various event dataset. Overall, our work contributes to the advancement of event-to-video reconstruction and offers a promising direction for future research that combines INRs with event data. + +Limitations and Future Work: The current parameter size of EvINR takes up approximately the same amount of storage as the original event data. In future work, we plan to explore network pruning and quantization techniques to further reduce the parameter size. + +# Acknowledgments + +This paper is supported by the National Natural Science Foundation of China (NSF) under Grant No. NSFC22FYT45, the Guangzhou City, University and Enterprise Joint Fund under Grant No.SL2022A03J01278, and Guangzhou Fundamental and Applied Basic Research (Grant Number: 2024A04J4072) + +# References + +1. Alpsentek products. https://www.alpsentek.com/product, [Accessed: Feb. 17, 2023] +2. Pepperl fuchs event camera. https://www.pepperl-fuchs.com/global/en/classid_11544.htm?view=productgroupliterature, accessed on 17 Feb 2023 +3. Bardow, P., Davison, A.J., Leutenegger, S.: Simultaneous optical flow and intensity estimation from an event camera. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 884-892 (2016) +4. Bishop, C.M.: Pattern Recognition and Machine Learning. Springer (2006) +5. Brandli, C., Berner, R., Yang, M., Liu, S.C., Delbruck, T.: A $240 \times 180$ 130 db $3\mu s$ latency global shutter spatiotemporal vision sensor. IEEE Journal of Solid-State Circuits 49(10), 2333-2341 (2014) +6. Cadena, P.R.G., Qian, Y., Wang, C., Yang, M.: Spade-e2vid: Spatially-adaptive denormalization for event-based video reconstruction. IEEE Transactions on Image Processing 30, 2488-2500 (2021) +7. Chen, H., He, B., Wang, H., Ren, Y., Lim, S.N., Shrivastava, A.: Nerv: Neural representations for videos. Advances in Neural Information Processing Systems 34, 21557-21568 (2021) +8. Chen, Y., Liu, S., Wang, X.: Learning continuous image representation with local implicit image function. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 8628-8638 (2021) +9. Cook, M., Gugelmann, L., Jug, F., Krautz, C., Steger, A.: Interacting maps for fast visual interpretation. In: The 2011 International Joint Conference on Neural Networks. pp. 770-776. IEEE (2011) +10. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929 (2020) +11. Ercan, B., Eker, O., Erdem, A., Erdem, E.: Evreal: Towards a comprehensive benchmark and analysis suite for event-based video reconstruction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3942-3951 (2023) +12. Feng, Y., Lv, H., Liu, H., Zhang, Y., Xiao, Y., Han, C.: Event density based denoising method for dynamic vision sensor. Applied Sciences 10(6), 2024 (2020) +13. Gallego, G., Delbrück, T., Orchard, G., Bartolozzi, C., Taba, B., Censi, A., Leutenegger, S., Davison, A.J., Conradt, J., Daniilidis, K., et al.: Event-based vision: A survey. IEEE transactions on pattern analysis and machine intelligence 44(1), 154-180 (2020) +14. Gehrig, D., Gehrig, M., Hidalgo-Carrio, J., Scaramuzza, D.: Video to events: Recycling video datasets for event cameras. In: IEEE Conf. Comput. Vis. Pattern Recog. (CVPR) (June 2020) + +15. Girshick, R.: Fast r-cnn. In: Proceedings of the IEEE international conference on computer vision. pp. 1440-1448 (2015) +16. Hansen, L.K., Salamon, P.: Neural network ensembles. IEEE transactions on pattern analysis and machine intelligence 12(10), 993-1001 (1990) +17. He, K., Gkioxari, G., Dollár, P., Girshick, R.: Mask r-cnn. In: Proceedings of the IEEE international conference on computer vision. pp. 2961-2969 (2017) +18. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 770-778 (2016) +19. Hu, Y., Liu, S.C., Delbruck, T.: v2e: From video frames to realistic dvs events. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1312-1321 (2021) +20. Hwang, I., Kim, J., Kim, Y.M.: Ev-nerf: Event based neural radiance field. arXiv preprint arXiv:2206.12455 (2022) +21. Kim, H., Handa, A., Benosman, R., Ieng, S.H., Davison, A.J.: Simultaneous mosaicing and tracking with an event camera. J. Solid State Circ 43, 566-576 (2008) +22. Kim, H., Leutenegger, S., Davison, A.J.: Real-time 3d reconstruction and 6-dof tracking with an event camera. In: European conference on computer vision. pp. 349-364. Springer (2016) +23. Kingma, D.P., Ba, J.: Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980 (2014) +24. Klenk, S., Koestler, L., Scaramuzza, D., Cremers, D.: E-nerf: Neural radiance fields from a moving event camera. IEEE Robotics and Automation Letters (2023) +25. Mescheder, L., Oechsle, M., Niemeyer, M., Nowozin, S., Geiger, A.: Occupancy networks: Learning 3d reconstruction in function space. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 4460-4470 (2019) +26. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM 65(1), 99-106 (2021) +27. Mueggler, E., Rebecq, H., Gallego, G., Delbruck, T., Scaramuzza, D.: The event-camera dataset and simulator: Event-based data for pose estimation, visual odometry, and slam. The International Journal of Robotics Research 36(2), 142-149 (2017) +28. Munda, G., Reinbacher, C., Pock, T.: Real-time intensity-image reconstruction for event cameras using manifold regularisation. International Journal of Computer Vision 126(12), 1381-1393 (2018) +29. Niemeyer, M., Geiger, A.: Giraffe: Representing scenes as compositional generative neural feature fields. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11453-11464 (2021) +30. Oechsle, M., Peng, S., Geiger, A.: Unisurf: Unifying neural implicit surfaces and radiance fields for multi-view reconstruction. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5589-5599 (2021) +31. Paredes-Vallés, F., de Croon, G.C.: Back to event basics: Self-supervised learning of image reconstruction for event cameras via photometric constancy. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3446-3455 (2021) +32. Park, J.J., Florence, P., Straub, J., Newcombe, R., Lovegrove, S.: Deepsdf: Learning continuous signed distance functions for shape representation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 165-174 (2019) + +33. Paszke, A., Gross, S., Chintala, S., Chanan, G., Yang, E., DeVito, Z., Lin, Z., Desmaison, A., Antiga, L., Lerer, A.: Automatic differentiation in pytorch (2017) +34. Pini, S., Borghi, G., Vezzani, R.: Learn to see by events: Color frame synthesis from event and rgb cameras. arXiv preprint arXiv:1812.02041 (2018) +35. Rebecq, H., Ranftl, R., Koltun, V., Scaramuzza, D.: Events-to-video: Bringing modern computer vision to event cameras. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3857-3866 (2019) +36. Redmon, J., Divvala, S., Girshick, R., Farhadi, A.: You only look once: Unified, real-time object detection. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 779-788 (2016) +37. Reinhard, E., Stark, M., Shirley, P., Ferwerda, J.: Photographic tone reproduction for digital images. ACM Transactions on Graphics 21(3), 267-276 (2002). https://doi.org/10.1145/566654.566576 +38. Rudnev, V., Elgharib, M., Theobalt, C., Golyanik, V.: Eventnerf: Neural radiance fields from a single colour event camera. arXiv preprint arXiv:2206.11896 (2022) +39. Ryu, H.E.: Industrial dvs design; key features and applications. In: Conf. on Computer Vision and Pattern Recognition. vol. 3 (2019) +40. Scheerlinck, C., Barnes, N., Mahony, R.: Continuous-time intensity estimation using event cameras. In: Asian Conference on Computer Vision. pp. 308-324. Springer (2018) +41. Scheerlinck, C., Rebecq, H., Gehrig, D., Barnes, N., Mahony, R., Scaramuzza, D.: Fast image reconstruction with an event camera. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 156-163 (2020) +42. Scheerlinck, C., Rebecq, H., Stoffregen, T., Barnes, N., Mahony, R., Scaramuzza, D.: Ced: Color event camera dataset. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops. pp. 0-0 (2019) +43. Schwarz, K., Liao, Y., Niemeyer, M., Geiger, A.: Graf: Generative radiance fields for 3d-aware image synthesis. In: Advances in Neural Information Processing Systems (NeurIPS) (2020) +44. Shiba, S., Aoki, Y., Gallego, G.: Event collapse in contrast maximization frameworks. Sensors 22(14), 5190 (2022) +45. Shiba, S., Aoki, Y., Gallego, G.: Secrets of event-based optical flow. In: European Conference on Computer Vision. pp. 628-645. Springer (2022) +46. Sitzmann, V., Martel, J., Bergman, A., Lindell, D., Wetzstein, G.: Implicit neural representations with periodic activation functions. Advances in Neural Information Processing Systems 33, 7462-7473 (2020) +47. Skorokhodov, I., Ignatyev, S., Elhoseiny, M.: Adversarial generation of continuous images. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10753-10764 (2021) +48. Song, C., Huang, Q., Bajaj, C.: E-cir: Event-enhanced continuous intensity recovery. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7803-7812 (2022) +49. Stoffregen, T., Scheerlinck, C., Scaramuzza, D., Drummond, T., Barnes, N., Kleeman, L., Mahony, R.: Reducing the sim-to-real gap for event cameras. In: European Conference on Computer Vision. pp. 534-549. Springer (2020) +50. Strümpler, Y., Postels, J., Yang, R., Gool, L.V., Tombari, F.: Implicit neural representations for image compression. In: Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXVI. pp. 74-91. Springer (2022) + +51. Wang, Y., Du, B., Shen, Y., Wu, K., Zhao, G., Sun, J., Wen, H.: Ev-gait: Event-based robust gait recognition using dynamic vision sensors. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6358-6367 (2019) +52. Wang, Z., Bovik, A.C., Sheikh, H.R., Simoncelli, E.P.: Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing 13(4), 600-612 (2004) +53. Weng, W., Zhang, Y., Xiong, Z.: Event-based video reconstruction using transformer. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 2563-2572 (2021) +54. Yadav, G., Maheshwari, S., Agarwal, A.: Contrast limited adaptive histogram equalization based enhancement for real time video system. In: 2014 international conference on advances in computing, communications and informatics (ICACCI). pp. 2392-2397. IEEE (2014) +55. Yang, M., Liu, S.C., Delbruck, T.: A dynamic vision sensor with $1\%$ temporal contrast sensitivity and in-pixel asynchronous delta modulator for event encoding. IEEE Journal of Solid-State Circuits 50(9), 2149-2160 (2015) +56. Zhang, K., Li, Y., Zuo, W., Zhang, L., Van Gool, L., Timofte, R.: Plug-and-play image restoration with deep denoiser prior. IEEE Transactions on Pattern Analysis and Machine Intelligence 44(10), 6360-6376 (2021) +57. Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual metric. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 586-595 (2018) +58. Zhang, X., Srinivasan, P.P., Deng, B., Debevec, P., Freeman, W.T., Barron, J.T.: Nerfactor: Neural factorization of shape and reflectance under an unknown illumination. ACM Transactions on Graphics (TOG) 40(6), 1-18 (2021) +59. Zhang, Z., Yezzi, A., Gallego, G.: Image reconstruction from events. why learn it? arXiv preprint arXiv:2112.06242 (2021) +60. Zhang, Z., Yezzi, A.J., Gallego, G.: Formulating event-based image reconstruction as a linear inverse problem with deep regularization using optical flow. IEEE Transactions on Pattern Analysis & Machine Intelligence 45(07), 8372-8389 (2023) +61. Zhu, A.Z., Thakur, D., Ozaslan, T., Pfrommer, B., Kumar, V., Daniilidis, K.: The multivehicle stereo event camera dataset: An event camera dataset for 3d perception. IEEE Robotics and Automation Letters 3(3), 2032-2039 (2018) +62. Zhu, A.Z., Yuan, L., Chaney, K., Daniilidis, K.: Ev-flownet: Self-supervised optical flow estimation for event-based cameras. arXiv preprint arXiv:1802.06898 (2018) +63. Zhu, A.Z., Yuan, L., Chaney, K., Daniilidis, K.: Unsupervised event-based learning of optical flow, depth, and egomotion. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 989-997 (2019) +64. Zhu, L., Wang, X., Chang, Y., Li, J., Huang, T., Tian, Y.: Event-based video reconstruction via potential-assisted spiking neural network. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3594-3604 (2022) \ No newline at end of file diff --git a/2024/Revisit Event Generation Model_ Self-Supervised Learning of Event-to-Video Reconstruction with Implicit Neural Representations/images.zip b/2024/Revisit Event Generation Model_ Self-Supervised Learning of Event-to-Video Reconstruction with Implicit Neural Representations/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..567b2f927b1d7372f13815ee22441b840a2f016e --- /dev/null +++ b/2024/Revisit Event Generation Model_ Self-Supervised Learning of Event-to-Video Reconstruction with Implicit Neural Representations/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7cb72d1da2aed896c227a8395511b4519c99be08c7998753580dc92394d94eed +size 598372 diff --git a/2024/Revisit Event Generation Model_ Self-Supervised Learning of Event-to-Video Reconstruction with Implicit Neural Representations/layout.json b/2024/Revisit Event Generation Model_ Self-Supervised Learning of Event-to-Video Reconstruction with Implicit Neural Representations/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..c802d036d75f1f79d69187b715e77aa1a3c0a6c3 --- /dev/null +++ b/2024/Revisit Event Generation Model_ Self-Supervised Learning of Event-to-Video Reconstruction with Implicit Neural Representations/layout.json @@ -0,0 +1,10190 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 136, + 111, + 478, + 166 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 111, + 478, + 166 + ], + "spans": [ + { + "bbox": [ + 136, + 111, + 478, + 166 + ], + "type": "text", + "content": "Revisit Event Generation Model: Self-Supervised Learning of Event-to-Video Reconstruction with Implicit Neural Representations" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 193, + 186, + 420, + 199 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 193, + 186, + 420, + 199 + ], + "spans": [ + { + "bbox": [ + 193, + 186, + 420, + 199 + ], + "type": "text", + "content": "Zipeng Wang1, Yunfan Lu1, and Lin Wang1,2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 214, + 207, + 399, + 218 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 214, + 207, + 399, + 218 + ], + "spans": [ + { + "bbox": [ + 214, + 207, + 399, + 218 + ], + "type": "text", + "content": "1 Artificial Intelligence Thrust, HKUST(GZ)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 193, + 218, + 420, + 230 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 193, + 218, + 420, + 230 + ], + "spans": [ + { + "bbox": [ + 193, + 218, + 420, + 230 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 193, + 218, + 420, + 230 + ], + "type": "text", + "content": " Dept. of Computer Science and Engineering, HKUST" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 172, + 231, + 441, + 241 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 172, + 231, + 441, + 241 + ], + "spans": [ + { + "bbox": [ + 172, + 231, + 441, + 241 + ], + "type": "text", + "content": "{zwang253,ylu066}@connect.hkust-gz.edu.cn, linwang@ust.hk" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 160, + 269, + 455, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 269, + 455, + 544 + ], + "spans": [ + { + "bbox": [ + 160, + 269, + 455, + 544 + ], + "type": "text", + "content": "Abstract. Reconstructing intensity frames from event data while maintaining high temporal resolution and dynamic range is crucial for bridging the gap between event-based and frame-based computer vision. Previous approaches have depended on supervised learning on synthetic data, which lacks interpretability and risk over-fitting to the setting of the event simulator. Recently, self-supervised learning (SSL) based methods, which primarily utilize per-frame optical flow to estimate intensity via photometric constancy, has been actively investigated. However, they are vulnerable to errors in the case of inaccurate optical flow. This paper proposes a novel SSL event-to-video reconstruction approach, dubbed EvINR, which eliminates the need for labeled data or optical flow estimation. Our core idea is to reconstruct intensity frames by directly addressing the event generation model, essentially a partial differential equation (PDE) that describes how events are generated based on the time-varying brightness signals. Specifically, we utilize an implicit neural representation (INR), which takes in spatiotemporal coordinate " + }, + { + "bbox": [ + 160, + 269, + 455, + 544 + ], + "type": "inline_equation", + "content": "(x,y,t)" + }, + { + "bbox": [ + 160, + 269, + 455, + 544 + ], + "type": "text", + "content": " and predicts intensity values, to represent the solution of the event generation equation. The INR, parameterized as a fully-connected Multi-layer Perceptron (MLP), can be optimized with its temporal derivatives supervised by events. To make EvINR feasible for online requisites, we propose several acceleration techniques that substantially expedite the training process. Comprehensive experiments demonstrate that our EvINR surpasses previous SSL methods by " + }, + { + "bbox": [ + 160, + 269, + 455, + 544 + ], + "type": "inline_equation", + "content": "38\\%" + }, + { + "bbox": [ + 160, + 269, + 455, + 544 + ], + "type": "text", + "content": " w.r.t. Mean Squared Error (MSE) and is comparable or superior to SoTA supervised methods. Project page: https://vlislab22.github.io/EvINR/." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 563, + 230, + 575 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 563, + 230, + 575 + ], + "spans": [ + { + "bbox": [ + 132, + 563, + 230, + 575 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 587, + 482, + 648 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 587, + 482, + 648 + ], + "spans": [ + { + "bbox": [ + 130, + 587, + 482, + 648 + ], + "type": "text", + "content": "Event cameras [13,55] are novel sensors that offer numerous advantages over traditional frame-based cameras, including low power consumption, high dynamic range (HDR), and high temporal resolution [34]. However, their unique imaging paradigm presents a challenge when applying vision algorithms designed for frame-based cameras. To address this challenge and bridge the gap between" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 134, + 653, + 234, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 653, + 234, + 665 + ], + "spans": [ + { + "bbox": [ + 134, + 653, + 234, + 665 + ], + "type": "text", + "content": "* Corresponding author" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 157, + 130, + 455, + 266 + ], + "blocks": [ + { + "bbox": [ + 159, + 116, + 257, + 128 + ], + "lines": [ + { + "bbox": [ + 159, + 116, + 257, + 128 + ], + "spans": [ + { + "bbox": [ + 159, + 116, + 257, + 128 + ], + "type": "text", + "content": "Event Generation Model" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 157, + 130, + 455, + 266 + ], + "lines": [ + { + "bbox": [ + 157, + 130, + 455, + 266 + ], + "spans": [ + { + "bbox": [ + 157, + 130, + 455, + 266 + ], + "type": "image", + "image_path": "9271eebe877c259690f453580b09094e8509cd9bbac6be9b1733ab51887f8a91.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 276, + 482, + 331 + ], + "lines": [ + { + "bbox": [ + 130, + 276, + 482, + 331 + ], + "spans": [ + { + "bbox": [ + 130, + 276, + 482, + 331 + ], + "type": "text", + "content": "Fig. 1: Connection between event generation model and EvINR: Event generation model reveals the relation between discrete events and continuous temporal intensity changes, described as the event generation equation (Eq. 4). EvINR utilizes an INR to solve Eq. 4 and recovering a continuous function of intensity w.r.t. time, implicitly parameterized with a fully connected MLP." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 361, + 480, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 361, + 480, + 384 + ], + "spans": [ + { + "bbox": [ + 130, + 361, + 480, + 384 + ], + "type": "text", + "content": "event-based and standard computer vision [10, 15, 17, 18, 36], many methods have been proposed to reconstruct intensity frames from events." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 388, + 482, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 388, + 482, + 555 + ], + "spans": [ + { + "bbox": [ + 130, + 388, + 482, + 555 + ], + "type": "text", + "content": "Early methods primarily rely on hand-crafted integrators or filters, resulting in significant loss of detail in reconstructed results [3,22,28,40,59]. More recently, deep learning-based methods [6,35,41,49,53,64] have emerged, typically supervised by large-scale synthetic datasets generated using an event simulator [14,19]. However, the interpretability of these methods is limited due to the 'black-box' nature of the neural networks. Furthermore, the performance of these methods is constrained by the domain gap between synthetic and real-world data [49], leading to suboptimal results if the event simulator's settings do not accurately match those of the inference data. To address these issues, some researchers have explored self-supervised learning (SSL) frameworks [31,60], aiming to eliminate the dependency on labeled synthetic data. Nonetheless, these approaches still depend on event-based optical flow estimation, which is prone to over-fitting, occlusions, and non-convergence issues [44, 45]. As a result, the reconstructed frames often suffer from the loss of textural details and various artifacts." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 558, + 482, + 666 + ], + "type": "text", + "content": "To address these challenges, we revisit the event generation model [13], which forms a fundamental link between events and intensity. The event generation model can be expressed as a partial differential equation (PDE), describing how events are triggered by logarithmic intensity changes that exceed a certain threshold. This PDE, known as the event generation equation, establishes a direct connection between discrete events and the partial derivative of the continuous intensity function w.r.t. time. Our key insight is that, solving the event generation equation offers an ideal self-supervised solution for event-to-video reconstruction, thereby eliminating the need for synthetic data or optical" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 226, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 226, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 226, + 102 + ], + "type": "text", + "content": "Z. Wang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 187 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 187 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 187 + ], + "type": "text", + "content": "flow estimation. However, several properties of event data make solving the equation nontrivial: 1) events can be triggered at extremely high frequencies (up to " + }, + { + "bbox": [ + 130, + 116, + 482, + 187 + ], + "type": "inline_equation", + "content": "10^{8} / s" + }, + { + "bbox": [ + 130, + 116, + 482, + 187 + ], + "type": "text", + "content": "), which results in a large volume of data and a heavy computational burden. 2) events are inherently noisy, particularly in extreme visual conditions [51], which pose challenge to the robustness of solvers. 3) events do not capture the initial intensity, which makes determining the boundary values challenging." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 188, + 482, + 272 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 188, + 482, + 272 + ], + "spans": [ + { + "bbox": [ + 130, + 188, + 482, + 272 + ], + "type": "text", + "content": "Recently, implicit neural representations (INRs) have gained popularity for solving the inverse problems in 3D reconstruction or image super-resolution [8, 26,46] by parameterizing complex signals via deep neural networks. In this work, we find that INRs possess several key advantages that render them particularly suitable for solving the event generation equation: they inherently accommodate a large volume of event data, exhibit high noise tolerance, and are flexible to add additional loss terms to regulate initial value." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 272, + 482, + 439 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 272, + 482, + 439 + ], + "spans": [ + { + "bbox": [ + 130, + 272, + 482, + 439 + ], + "type": "text", + "content": "In light of this, we propose a novel SSL framework, termed EvINR, that employs an INR to represent the solution of the event generation equation. Our EvINR can be directly optimized by minimizing the residual between its temporal derivatives and local intensity changes estimated from event data based on the event generation model (Sec. 3.2). Moreover, we incorporate a spatial regularization term that regularizes the relative values of adjacent pixels by constraining the magnitude of spatial gradients, which effectively reduces noise in the reconstruction process (Sec. 3.2). Although the basic implementation of EvINR yields acceptable results, its convergence on seconds-long event sequences takes minutes, limiting its real-world applicability, especially in online scenarios. To expedite EvINR's training process, we introduce several acceleration techniques, including frame-based optimization, coarse-to-fine training, and model ensembling. These approaches reduce the training time from minutes to seconds while not compromising the reconstruction quality (Sec. 3.4)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 440, + 482, + 499 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 440, + 482, + 499 + ], + "spans": [ + { + "bbox": [ + 130, + 440, + 482, + 499 + ], + "type": "text", + "content": "Moreover, most approaches are typically evaluated on event datasets captured by DAVIS sensors [27, 42, 49, 61], making it difficult to evaluate their stability and robustness to other types of event camera [1, 2, 39]. For this reason, we collected a new event dataset using an ALPIX-Eiger event camera [1], with well-aligned events and intensity frames." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 500, + 482, + 583 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 500, + 482, + 583 + ], + "spans": [ + { + "bbox": [ + 130, + 500, + 482, + 583 + ], + "type": "text", + "content": "In summary, our paper makes three key contributions: (I) We propose EvINR, a concise SSL framework that solves the event generation equation via implicit neural representations for event-to-video reconstruction. (II) Our EvINR substantially outperforms previous SSL methods [31, 40, 59] and attains comparable, or even superior, performance compared to state-of-the-art supervised methods [35, 49, 53]. (III) We collect a real-world dataset with an ALPIX-Eiger event camera, complementary to the datasets captured by DAVIS cameras." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 603, + 243, + 616 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 603, + 243, + 616 + ], + "spans": [ + { + "bbox": [ + 131, + 603, + 243, + 616 + ], + "type": "text", + "content": "2 Related Works" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "text", + "content": "Implicit Neural Representations (INRs) have emerged as a powerful tool for parameterizing signals, such as images, videos, and audio, in a continuous manner using neural networks [46]. Compared to traditional discrete sig-" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 417, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 417, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 417, + 91, + 447, + 100 + ], + "type": "text", + "content": "EvINR" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 137, + 118, + 478, + 232 + ], + "blocks": [ + { + "bbox": [ + 137, + 118, + 478, + 232 + ], + "lines": [ + { + "bbox": [ + 137, + 118, + 478, + 232 + ], + "spans": [ + { + "bbox": [ + 137, + 118, + 478, + 232 + ], + "type": "image", + "image_path": "8b29df90046efc81ca4d72191edd630d2ab8ec1e1cd441d98a91e34a3d772a8e.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 241, + 482, + 275 + ], + "lines": [ + { + "bbox": [ + 130, + 241, + 482, + 275 + ], + "spans": [ + { + "bbox": [ + 130, + 241, + 482, + 275 + ], + "type": "text", + "content": "Fig. 2: Overview of EvINR. A fully connected MLP is used to implicitly solve the event generation equation. The temporal gradient of the MLP is supervised by temporal intensity changes of events, and the spatial gradient is penalized to reduce noise." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 303, + 482, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 303, + 482, + 483 + ], + "spans": [ + { + "bbox": [ + 130, + 303, + 482, + 483 + ], + "type": "text", + "content": "nal representations, INRs offer the advantage of being able to be sampled at arbitrary resolutions with fixed memory requirements. As a result, INRs have found widespread applications in various fields, including 3D scene representation [25, 26, 32], video representation [7], generative models [29, 43, 47], and model compression [50]. A unique property of INRs is that they can be effectively learned from the derivatives of signals, such as the normals of 3D shapes [30,46,58] and the gradients of images [46]. Such a characteristic has motivated us to approach the task of event-to-video reconstruction by optimizing the INR of video from its temporal derivatives. Previous research has investigated the potential of INR for novel view synthesis using event data [20,24,38]. These approaches reconstruct 3D neural radiance fields using multiple event sequences with known camera poses from a stationary scene. Most similar to our work, E-CIR [48] uses polynomial to regress the intensity function, which may only represent a short time interval (e.g., the exposure time of an image). In contrast, our method employs INR, which can represent longer time of intensity change." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 484, + 482, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 484, + 482, + 640 + ], + "spans": [ + { + "bbox": [ + 130, + 484, + 482, + 640 + ], + "type": "text", + "content": "Event-based Video Reconstruction has been a hot topic in the literature. Early attempts [3, 9, 21] address this problem based on the photometric constancy, which describes the relationship between intensity gradients and optical flow. Other approaches [28, 40] are based on direct event integration without estimating optical flow. Rebecq et al. [35] developed the first DL-based framework, called E2VID, to reconstruct intensity frames from events in an end-to-end manner, outperforming earlier techniques by a significant margin. E2VID was updated by some following research, trying to tackle problems of inference speed [41], the cold start problem [6], and training strategy [49]. These methods are supervised and learned using the synthetic dataset obtained via the event simulators [14, 19] due to the lack of real-world datasets with well-aligned events and intensity frames. Therefore, the generalization capability of these methods is limited by the simulator-to-real gap [49]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 665 + ], + "type": "text", + "content": "To address this problem, Federico et al. [31] proposed an SSL framework with a network to estimate optical flow and another network to reconstruct intensity" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "type": "text", + "content": "Z. Wang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 131, + 116, + 482, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 116, + 482, + 213 + ], + "spans": [ + { + "bbox": [ + 131, + 116, + 482, + 213 + ], + "type": "text", + "content": "frames based on the photometric constancy. Zhang et al. [60] updated this idea into a linear inverse problem that can be solved using modern linear solvers without using deep learning. However, these methods either assume optical flow is known or estimate the optical flow using contrast maximization [62,63]. Consequently, their performance cannot be guaranteed unless high-quality optical flow can be obtained. In contrast, our approach is based solely on the physical event generation model, leading to a more straightforward solution that also demonstrates significantly improved performance and greater flexibility." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 228, + 202, + 242 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 228, + 202, + 242 + ], + "spans": [ + { + "bbox": [ + 132, + 228, + 202, + 242 + ], + "type": "text", + "content": "3 Method" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 253, + 482, + 385 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 253, + 482, + 385 + ], + "spans": [ + { + "bbox": [ + 130, + 253, + 482, + 385 + ], + "type": "text", + "content": "Overview: Our objective is to reconstruct intensity frames from events in a self-supervised manner, without the need for end-to-end training or optical flow estimation. To accomplish this, we have reformulated event-based video reconstruction as solving the event generation equation. We employ an INR that is supervised solely by event data to represent the intensity function. An overview of our approach can be seen in Fig.2. In Sec. 3.1, we explain the event generation model and its connection to event-based video reconstruction. In Sec. 3.2, we detail how to train an EvINR by supervising its spatial and temporal gradients. Techniques to speed up the training of EvINR for online applications are discussed in Sec. 3.3. Finally, in Sec. 3.4, we describe our collected dataset using the ALPIX event camera [1]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 401, + 353, + 412 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 401, + 353, + 412 + ], + "spans": [ + { + "bbox": [ + 131, + 401, + 353, + 412 + ], + "type": "text", + "content": "3.1 Preliminary: Event Generation Model" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 418, + 481, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 418, + 481, + 467 + ], + "spans": [ + { + "bbox": [ + 130, + 418, + 481, + 467 + ], + "type": "text", + "content": "We begin by providing a brief overview of the event generation model [13], which forms the theoretical foundation for our approach. Let " + }, + { + "bbox": [ + 130, + 418, + 481, + 467 + ], + "type": "inline_equation", + "content": "I(x,y,t)" + }, + { + "bbox": [ + 130, + 418, + 481, + 467 + ], + "type": "text", + "content": " denote the intensity of the spatial location " + }, + { + "bbox": [ + 130, + 418, + 481, + 467 + ], + "type": "inline_equation", + "content": "(x,y)" + }, + { + "bbox": [ + 130, + 418, + 481, + 467 + ], + "type": "text", + "content": " at time " + }, + { + "bbox": [ + 130, + 418, + 481, + 467 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 418, + 481, + 467 + ], + "type": "text", + "content": " in a video. Since event cameras operate with logarithmic intensity, we denote " + }, + { + "bbox": [ + 130, + 418, + 481, + 467 + ], + "type": "inline_equation", + "content": "L(x,y,t) = \\log I(x,y,t)" + }, + { + "bbox": [ + 130, + 418, + 481, + 467 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 468, + 482, + 538 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 468, + 482, + 538 + ], + "spans": [ + { + "bbox": [ + 131, + 468, + 482, + 538 + ], + "type": "text", + "content": "An event camera comprises a frame of independent pixels that respond to changes in the logarithmic intensity signal and produce sequences of sparse and asynchronous events. An event " + }, + { + "bbox": [ + 131, + 468, + 482, + 538 + ], + "type": "inline_equation", + "content": "(x_{i},y_{i},t_{i},p_{i})" + }, + { + "bbox": [ + 131, + 468, + 482, + 538 + ], + "type": "text", + "content": " is triggered when the logarithmic intensity change surpasses a threshold " + }, + { + "bbox": [ + 131, + 468, + 482, + 538 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 131, + 468, + 482, + 538 + ], + "type": "text", + "content": " since the previous event was triggered at the same pixel. where " + }, + { + "bbox": [ + 131, + 468, + 482, + 538 + ], + "type": "inline_equation", + "content": "(x,y)" + }, + { + "bbox": [ + 131, + 468, + 482, + 538 + ], + "type": "text", + "content": " is the spatial location of the pixel, " + }, + { + "bbox": [ + 131, + 468, + 482, + 538 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 131, + 468, + 482, + 538 + ], + "type": "text", + "content": " is the timestamp, and " + }, + { + "bbox": [ + 131, + 468, + 482, + 538 + ], + "type": "inline_equation", + "content": "p\\in \\{-1,1\\}" + }, + { + "bbox": [ + 131, + 468, + 482, + 538 + ], + "type": "text", + "content": " is the polarity of the logarithmic intensity change." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 539, + 482, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 539, + 482, + 586 + ], + "spans": [ + { + "bbox": [ + 131, + 539, + 482, + 586 + ], + "type": "text", + "content": "For simplicity, let us consider the temporal changes in logarithmic intensity of a single pixel with fixed spatial coordinates and disregard the spatial terms " + }, + { + "bbox": [ + 131, + 539, + 482, + 586 + ], + "type": "inline_equation", + "content": "(x,y)" + }, + { + "bbox": [ + 131, + 539, + 482, + 586 + ], + "type": "text", + "content": ". We can describe an event " + }, + { + "bbox": [ + 131, + 539, + 482, + 586 + ], + "type": "inline_equation", + "content": "e_i" + }, + { + "bbox": [ + 131, + 539, + 482, + 586 + ], + "type": "text", + "content": " on that pixel using the Dirac delta function as follows:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 256, + 587, + 481, + 599 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 256, + 587, + 481, + 599 + ], + "spans": [ + { + "bbox": [ + 256, + 587, + 481, + 599 + ], + "type": "interline_equation", + "content": "e _ {i} (t) = p _ {i} \\cdot C \\cdot \\delta \\left(t - t _ {i}\\right). \\tag {1}", + "image_path": "4ac02d6a073e28043cc1294ce44d6118524179a00a8cd81d12e1d2029a7ab646.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 603, + 482, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 603, + 482, + 639 + ], + "spans": [ + { + "bbox": [ + 131, + 603, + 482, + 639 + ], + "type": "text", + "content": "Therefore, the logarithmic intensity increment " + }, + { + "bbox": [ + 131, + 603, + 482, + 639 + ], + "type": "inline_equation", + "content": "\\Delta L = L(t_2) - L(t_1)" + }, + { + "bbox": [ + 131, + 603, + 482, + 639 + ], + "type": "text", + "content": " in a time interval " + }, + { + "bbox": [ + 131, + 603, + 482, + 639 + ], + "type": "inline_equation", + "content": "[t_1,t_2]" + }, + { + "bbox": [ + 131, + 603, + 482, + 639 + ], + "type": "text", + "content": " can be represented by the accumulation of events, which can be expressed as:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 260, + 639, + 481, + 667 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 260, + 639, + 481, + 667 + ], + "spans": [ + { + "bbox": [ + 260, + 639, + 481, + 667 + ], + "type": "interline_equation", + "content": "\\Delta L = \\int_ {t _ {1}} ^ {t _ {2}} \\sum_ {i} e _ {i} (t) \\mathrm {d} t. \\tag {2}", + "image_path": "3ca216b68ca6505ef11b9c06353a5e4ae25685a37bcaa32d182b50cf5c07df61.jpg" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 417, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 417, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 417, + 91, + 447, + 100 + ], + "type": "text", + "content": "EvINR" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "content": "Assuming a short time interval and ignoring noise, we can approximate the logarithmic intensity increment by its first-order temporal derivative using Taylor expansion:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 249, + 152, + 481, + 178 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 249, + 152, + 481, + 178 + ], + "spans": [ + { + "bbox": [ + 249, + 152, + 481, + 178 + ], + "type": "interline_equation", + "content": "\\frac {\\Delta L}{t _ {2} - t _ {1}} = \\frac {\\partial L ((t _ {1} + t _ {2}) / 2)}{\\partial t}. \\tag {3}", + "image_path": "c3ddce08163cba6b02594588bb0310b726c6c10d93de22a7adfd72502ac5e08b.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 183, + 479, + 220 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 183, + 479, + 220 + ], + "spans": [ + { + "bbox": [ + 131, + 183, + 479, + 220 + ], + "type": "text", + "content": "By substituting Eq. 2 into Eq. 3, we derive the event generation equation (Eq. 4), which bridges the discrete event data with the continuous temporal derivatives of logarithmic intensity:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 217, + 230, + 481, + 259 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 217, + 230, + 481, + 259 + ], + "spans": [ + { + "bbox": [ + 217, + 230, + 481, + 259 + ], + "type": "interline_equation", + "content": "\\frac {\\partial L \\left(\\left(t _ {1} + t _ {2}\\right) / 2\\right)}{\\partial t} = \\frac {1}{t _ {2} - t _ {1}} \\int_ {t _ {1}} ^ {t _ {2}} \\sum_ {i} e _ {i} (t) \\mathrm {d} t. \\tag {4}", + "image_path": "6c418f5f6e7598c051641e7c0fd5a9ec2b5b4cb15761c953c1982f3387832970.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 269, + 480, + 388 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 269, + 480, + 388 + ], + "spans": [ + { + "bbox": [ + 130, + 269, + 480, + 388 + ], + "type": "text", + "content": "The left-hand side of Eq. 4 represents the partial derivative of logarithmic intensity with respect to time, while the right-hand side can be calculated by accumulating events. Intuitively, event-based video reconstruction can be formulated as solving Eq. 4, i.e., finding a logarithmic intensity function that satisfies the equation. However, as stated earlier, solving Eq. 4 can be non-trivial due to the large amount, high noise rate of event data, and unknown boundary intensity values. To address these challenges, we employ INRs to solve Eq. 4 as they naturally scale up to large data and have high noise tolerance [20]. Boundary intensity values can also easily be regularized by injecting natural image priors as additional loss functions." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 410, + 300, + 422 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 410, + 300, + 422 + ], + "spans": [ + { + "bbox": [ + 132, + 410, + 300, + 422 + ], + "type": "text", + "content": "3.2 Learning INRs from Events" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 432, + 479, + 482 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 432, + 479, + 482 + ], + "spans": [ + { + "bbox": [ + 130, + 432, + 479, + 482 + ], + "type": "text", + "content": "We aim to learn an INR " + }, + { + "bbox": [ + 130, + 432, + 479, + 482 + ], + "type": "inline_equation", + "content": "F_{\\Theta}" + }, + { + "bbox": [ + 130, + 432, + 479, + 482 + ], + "type": "text", + "content": " given an event stream " + }, + { + "bbox": [ + 130, + 432, + 479, + 482 + ], + "type": "inline_equation", + "content": "\\{e_i\\}_{i=0}^N" + }, + { + "bbox": [ + 130, + 432, + 479, + 482 + ], + "type": "text", + "content": ". Here, " + }, + { + "bbox": [ + 130, + 432, + 479, + 482 + ], + "type": "inline_equation", + "content": "\\Theta" + }, + { + "bbox": [ + 130, + 432, + 479, + 482 + ], + "type": "text", + "content": " represents the parameters of a fully connected MLP that predicts the logarithmic intensity value " + }, + { + "bbox": [ + 130, + 432, + 479, + 482 + ], + "type": "inline_equation", + "content": "\\hat{L}" + }, + { + "bbox": [ + 130, + 432, + 479, + 482 + ], + "type": "text", + "content": " at any spatiotemporal coordinate " + }, + { + "bbox": [ + 130, + 432, + 479, + 482 + ], + "type": "inline_equation", + "content": "(x,y,t)" + }, + { + "bbox": [ + 130, + 432, + 479, + 482 + ], + "type": "text", + "content": ". We adapt the event frame representation [35] and stack the given events uniformly into " + }, + { + "bbox": [ + 130, + 432, + 479, + 482 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 130, + 432, + 479, + 482 + ], + "type": "text", + "content": " event frames." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 502, + 479, + 562 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 502, + 479, + 562 + ], + "spans": [ + { + "bbox": [ + 130, + 502, + 479, + 562 + ], + "type": "text", + "content": "Temporal Supervision The INR is optimized by directly minimizing the temporal loss between the predicted logarithmic intensity change " + }, + { + "bbox": [ + 130, + 502, + 479, + 562 + ], + "type": "inline_equation", + "content": "\\Delta \\hat{L}" + }, + { + "bbox": [ + 130, + 502, + 479, + 562 + ], + "type": "text", + "content": " and the logarithmic intensity change estimated by event accumulation " + }, + { + "bbox": [ + 130, + 502, + 479, + 562 + ], + "type": "inline_equation", + "content": "\\Delta L" + }, + { + "bbox": [ + 130, + 502, + 479, + 562 + ], + "type": "text", + "content": ". Here " + }, + { + "bbox": [ + 130, + 502, + 479, + 562 + ], + "type": "inline_equation", + "content": "\\Delta \\hat{L}" + }, + { + "bbox": [ + 130, + 502, + 479, + 562 + ], + "type": "text", + "content": " can be derived by accumulating events as Eq. 2 and " + }, + { + "bbox": [ + 130, + 502, + 479, + 562 + ], + "type": "inline_equation", + "content": "\\Delta \\hat{L}" + }, + { + "bbox": [ + 130, + 502, + 479, + 562 + ], + "type": "text", + "content": " is the change of " + }, + { + "bbox": [ + 130, + 502, + 479, + 562 + ], + "type": "inline_equation", + "content": "\\hat{L}" + }, + { + "bbox": [ + 130, + 502, + 479, + 562 + ], + "type": "text", + "content": " with respect to " + }, + { + "bbox": [ + 130, + 502, + 479, + 562 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 502, + 479, + 562 + ], + "type": "text", + "content": " obtained by performing double back-propagation [33] of " + }, + { + "bbox": [ + 130, + 502, + 479, + 562 + ], + "type": "inline_equation", + "content": "F_{\\Theta}" + }, + { + "bbox": [ + 130, + 502, + 479, + 562 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 233, + 573, + 481, + 596 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 233, + 573, + 481, + 596 + ], + "spans": [ + { + "bbox": [ + 233, + 573, + 481, + 596 + ], + "type": "interline_equation", + "content": "\\Delta \\hat {L} = \\frac {\\partial F _ {\\Theta} \\left(\\left(t _ {1} + t _ {2}\\right) / 2\\right)}{\\partial t} \\cdot \\left(t _ {2} - t _ {1}\\right), \\tag {5}", + "image_path": "44ada1fefaeb34fc44783c7ee22bcceb5f9064edcc18baefee962751f2c3c3b9.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 131, + 605, + 466, + 618 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 605, + 466, + 618 + ], + "spans": [ + { + "bbox": [ + 131, + 605, + 466, + 618 + ], + "type": "text", + "content": "Here we adopt the mean squared error (MSE) loss for temporal supervision:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 259, + 628, + 481, + 643 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 259, + 628, + 481, + 643 + ], + "spans": [ + { + "bbox": [ + 259, + 628, + 481, + 643 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {t e m p}} = \\left(\\Delta L - \\Delta \\hat {L}\\right) ^ {2}, \\tag {6}", + "image_path": "8da5ac740fb13ff901762c90a8aaf027692f21aa15fb03fcd2f26da35a78994d.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 131, + 652, + 403, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 652, + 403, + 666 + ], + "spans": [ + { + "bbox": [ + 131, + 652, + 403, + 666 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 652, + 403, + 666 + ], + "type": "inline_equation", + "content": "\\Delta L" + }, + { + "bbox": [ + 131, + 652, + 403, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 652, + 403, + 666 + ], + "type": "inline_equation", + "content": "\\Delta \\hat{L}" + }, + { + "bbox": [ + 131, + 652, + 403, + 666 + ], + "type": "text", + "content": " are given by Eq. 2 and Eq. 5, respectively." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "type": "text", + "content": "Z. Wang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 131, + 116, + 479, + 188 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 116, + 479, + 188 + ], + "spans": [ + { + "bbox": [ + 131, + 116, + 479, + 188 + ], + "type": "text", + "content": "Spatial Regularization Although the temporal supervision can estimate the intensity function, the results may still contain unnatural artifacts as the INR network " + }, + { + "bbox": [ + 131, + 116, + 479, + 188 + ], + "type": "inline_equation", + "content": "F_{\\Theta}" + }, + { + "bbox": [ + 131, + 116, + 479, + 188 + ], + "type": "text", + "content": " has no prior knowledge of the initial intensity values of each pixel. To address this issue, we introduce a spatial regularization term that encourages the solution to be in the space of natural-looking images. We adopt Tikhonov regularization [4] that penalizes the spatial gradients of logarithmic intensity:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 250, + 198, + 479, + 222 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 250, + 198, + 479, + 222 + ], + "spans": [ + { + "bbox": [ + 250, + 198, + 479, + 222 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {r e g}} = \\left(\\frac {\\partial F _ {\\Theta}}{\\partial x}\\right) ^ {2} + \\left(\\frac {\\partial F _ {\\Theta}}{\\partial y}\\right) ^ {2}. \\tag {7}", + "image_path": "d8bb1ca8ef8c38fb7c89f15663ea3be405b862b1f4f60af17018593931828d39.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 228, + 479, + 265 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 228, + 479, + 265 + ], + "spans": [ + { + "bbox": [ + 131, + 228, + 479, + 265 + ], + "type": "text", + "content": "It is worth noting that we do not employ more complex regularization methods, such as CNN denoisers [56] as used in [60], because they typically utilize large network models that can significantly slow down the training process." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 282, + 436, + 294 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 282, + 436, + 294 + ], + "spans": [ + { + "bbox": [ + 132, + 282, + 436, + 294 + ], + "type": "text", + "content": "Optimizing EvINR The overall objective is given by the equation:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 265, + 305, + 479, + 316 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 265, + 305, + 479, + 316 + ], + "spans": [ + { + "bbox": [ + 265, + 305, + 479, + 316 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathcal {L} _ {\\text {t e m p}} + \\lambda \\mathcal {L} _ {\\text {r e g}}, \\tag {8}", + "image_path": "3cea7fdb74b483b62e1eb7e7d8aa6f111f8ee376d1befce556f04336d3e35f92.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 325, + 479, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 325, + 479, + 361 + ], + "spans": [ + { + "bbox": [ + 131, + 325, + 479, + 361 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 325, + 479, + 361 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{temp}}" + }, + { + "bbox": [ + 131, + 325, + 479, + 361 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 131, + 325, + 479, + 361 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{reg}}" + }, + { + "bbox": [ + 131, + 325, + 479, + 361 + ], + "type": "text", + "content": " were introduced in Sec. 3.2 and Sec. 3.2, respectively, as the temporal supervision and spatial regularization term. " + }, + { + "bbox": [ + 131, + 325, + 479, + 361 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 131, + 325, + 479, + 361 + ], + "type": "text", + "content": " is a hyper-parameter used to adjust the weight of the spatial regularization term." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 379, + 479, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 379, + 479, + 426 + ], + "spans": [ + { + "bbox": [ + 131, + 379, + 479, + 426 + ], + "type": "text", + "content": "Tone-mapping The output of the EvINR is the logarithmic intensity of the reconstructed frames. We first use the exponential function to convert the predicted logarithmic intensity values to high dynamic range (HDR) intensity values " + }, + { + "bbox": [ + 131, + 379, + 479, + 426 + ], + "type": "inline_equation", + "content": "I \\in [0, \\infty)" + }, + { + "bbox": [ + 131, + 379, + 479, + 426 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 247, + 427, + 479, + 438 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 247, + 427, + 479, + 438 + ], + "spans": [ + { + "bbox": [ + 247, + 427, + 479, + 438 + ], + "type": "interline_equation", + "content": "I (x, y, t) = \\exp \\left(F _ {\\Theta} (x, y, t)\\right). \\tag {9}", + "image_path": "c133a2d95775c6d6054aee8af029805fc0c1bee0f7b2cc68e65cac26e8ddb52e.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 445, + 479, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 445, + 479, + 468 + ], + "spans": [ + { + "bbox": [ + 131, + 445, + 479, + 468 + ], + "type": "text", + "content": "Then, we adopt the Reinhard function [37] to map the HDR intensity values into the low dynamic range " + }, + { + "bbox": [ + 131, + 445, + 479, + 468 + ], + "type": "inline_equation", + "content": "[0,1]" + }, + { + "bbox": [ + 131, + 445, + 479, + 468 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 270, + 477, + 479, + 499 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 270, + 477, + 479, + 499 + ], + "spans": [ + { + "bbox": [ + 270, + 477, + 479, + 499 + ], + "type": "interline_equation", + "content": "\\Gamma (I) = \\left(\\frac {I}{I + 1}\\right) ^ {\\gamma}. \\tag {10}", + "image_path": "270bced0ee255118f9794b9d1f435c56adaac259e66110cb7868cc8a56807fef.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 131, + 508, + 479, + 532 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 508, + 479, + 532 + ], + "spans": [ + { + "bbox": [ + 131, + 508, + 479, + 532 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 131, + 508, + 479, + 532 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 131, + 508, + 479, + 532 + ], + "type": "text", + "content": " is a hyper-parameter used to control the contrast of " + }, + { + "bbox": [ + 131, + 508, + 479, + 532 + ], + "type": "inline_equation", + "content": "\\Gamma(I)" + }, + { + "bbox": [ + 131, + 508, + 479, + 532 + ], + "type": "text", + "content": ". We assess our reconstruction performance by applying " + }, + { + "bbox": [ + 131, + 508, + 479, + 532 + ], + "type": "inline_equation", + "content": "\\Gamma(I)" + }, + { + "bbox": [ + 131, + 508, + 479, + 532 + ], + "type": "text", + "content": " in all experiments." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 550, + 262, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 550, + 262, + 562 + ], + "spans": [ + { + "bbox": [ + 132, + 550, + 262, + 562 + ], + "type": "text", + "content": "3.3 Accelerating EvINR" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 570, + 480, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 480, + 665 + ], + "type": "text", + "content": "While solving the event generation equation using the basic implementation of EvINR, as described in Sec. 3.3, yields satisfactory reconstruction results, it is not efficient enough for online tasks. Optimizing one INR network takes about 20 minutes, and a network can only represent approximately 1 second of a sequence. Reducing the training time or increasing the sequence time leads to severe performance degradation. To enable online usage for EvINR, we propose several techniques to reduce the training time and increase the representation capacity for EvINR, as illustrated in Fig. 3." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 417, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 417, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 417, + 91, + 447, + 100 + ], + "type": "text", + "content": "EvINR" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 161, + 117, + 276, + 232 + ], + "blocks": [ + { + "bbox": [ + 161, + 117, + 276, + 232 + ], + "lines": [ + { + "bbox": [ + 161, + 117, + 276, + 232 + ], + "spans": [ + { + "bbox": [ + 161, + 117, + 276, + 232 + ], + "type": "image", + "image_path": "0bce85b79b74465683d976941653301334345c5a6e7cba7e8b8909d6a51a5a9d.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 243, + 482, + 278 + ], + "lines": [ + { + "bbox": [ + 131, + 243, + 482, + 278 + ], + "spans": [ + { + "bbox": [ + 131, + 243, + 482, + 278 + ], + "type": "text", + "content": "Fig. 3: Overview of acceleration techniques. (a) and (b) illustrate the difference between the basic coordinate-based and our frame-based optimization, respectively. (c) depicts our proposed coarse-to-fine training scheme and network ensembling technique." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 279, + 118, + 455, + 232 + ], + "blocks": [ + { + "bbox": [ + 279, + 118, + 455, + 232 + ], + "lines": [ + { + "bbox": [ + 279, + 118, + 455, + 232 + ], + "spans": [ + { + "bbox": [ + 279, + 118, + 455, + 232 + ], + "type": "image", + "image_path": "1657c628445946bdf53a0eef143f623a836bf6a1b9764428b9c3c28a195e7e2a.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 300, + 482, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 300, + 482, + 396 + ], + "spans": [ + { + "bbox": [ + 131, + 300, + 482, + 396 + ], + "type": "text", + "content": "Coordinate-based to frame-based optimization: The basic EvINR algorithm is designed to learn a mapping from 3D coordinates " + }, + { + "bbox": [ + 131, + 300, + 482, + 396 + ], + "type": "inline_equation", + "content": "(x,y,t)" + }, + { + "bbox": [ + 131, + 300, + 482, + 396 + ], + "type": "text", + "content": " to the logarithmic intensity " + }, + { + "bbox": [ + 131, + 300, + 482, + 396 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 131, + 300, + 482, + 396 + ], + "type": "text", + "content": ". However, this coordinate-based optimization requires the network to 'remember' " + }, + { + "bbox": [ + 131, + 300, + 482, + 396 + ], + "type": "inline_equation", + "content": "H\\times W\\times T" + }, + { + "bbox": [ + 131, + 300, + 482, + 396 + ], + "type": "text", + "content": " coordinate-to-logarithmic intensity mappings, which can be highly complex and lead to slow convergence. To address this issue, we propose a frame-based optimization scheme that learns a mapping from timestamps " + }, + { + "bbox": [ + 131, + 300, + 482, + 396 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 131, + 300, + 482, + 396 + ], + "type": "text", + "content": " to the logarithmic intensity of all pixels at that timestamp. The frame-based optimization scheme can be represented as:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 235, + 404, + 481, + 419 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 235, + 404, + 481, + 419 + ], + "spans": [ + { + "bbox": [ + 235, + 404, + 481, + 419 + ], + "type": "interline_equation", + "content": "F _ {\\Theta} (t) = \\left[ L (i, j, t) \\right] _ {1 \\leq i \\leq H, 1 \\leq j \\leq W}, \\tag {11}", + "image_path": "5b4d82c8b9fb6ec279d4041aadc6cf6d3c65100d58f0865b0466607e50a5dfa6.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 426, + 482, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 426, + 482, + 510 + ], + "spans": [ + { + "bbox": [ + 131, + 426, + 482, + 510 + ], + "type": "text", + "content": "where the right hand-side represents a matrix that contains all logarithmic intensity of all pixels at " + }, + { + "bbox": [ + 131, + 426, + 482, + 510 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 131, + 426, + 482, + 510 + ], + "type": "text", + "content": ". The frame-based optimization approach requires the network to remember only " + }, + { + "bbox": [ + 131, + 426, + 482, + 510 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 131, + 426, + 482, + 510 + ], + "type": "text", + "content": " coordinate-to-logarithmic intensity mappings, resulting in a significant reduction in complexity and faster convergence. Empirical results show that this approach reduces the convergence time by two orders of magnitude, from minutes to seconds. The comparison between coordinate-based and frame-based optimizations is illustrated in Fig. 3 (a) and (b)." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 510, + 482, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 510, + 482, + 641 + ], + "spans": [ + { + "bbox": [ + 131, + 510, + 482, + 641 + ], + "type": "text", + "content": "Coarse-to-fine training: We adopt a coarse-to-fine training scheme, which enables the network to learn overall logarithmic intensity changes before focusing on finer details. We structure the training process into " + }, + { + "bbox": [ + 131, + 510, + 482, + 641 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 131, + 510, + 482, + 641 + ], + "type": "text", + "content": " distinct stages, at each of which we increase the temporal resolution by a factor of 2. Initially, we divide the event sequence into " + }, + { + "bbox": [ + 131, + 510, + 482, + 641 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 131, + 510, + 482, + 641 + ], + "type": "text", + "content": " segments of equal length and optimize the EvINR in accordance with Eq. 6 over a specified number of iterations. Subsequently, we proceed to bisect each of these " + }, + { + "bbox": [ + 131, + 510, + 482, + 641 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 131, + 510, + 482, + 641 + ], + "type": "text", + "content": " segments into two smaller ones, ensuring that they contain an equal number of events. For all the experiments, we set the number of stages " + }, + { + "bbox": [ + 131, + 510, + 482, + 641 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 131, + 510, + 482, + 641 + ], + "type": "text", + "content": " to 3 and scheduled the upsampling to occur after 100 and 200 iterations, respectively. This approach accelerates training by reducing the number of event frames needed in the early stages by approximately 2 times." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 641, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 641, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 131, + 641, + 482, + 666 + ], + "type": "text", + "content": "Network ensembling: As each EvINR network requires only about 1GB of GPU memory, we further leverage network ensemble techniques [16] to train " + }, + { + "bbox": [ + 131, + 641, + 482, + 666 + ], + "type": "inline_equation", + "content": "N" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 226, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 226, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 226, + 102 + ], + "type": "text", + "content": "Z. Wang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 189 + ], + "type": "text", + "content": "EvINR networks simultaneously on a single GPU to achieve higher parallelism. This approach enables us to exploit the computational resources of modern GPUs more effectively and speed up the training process. We keep overlap periods between nearby networks to keep logarithmic intensity predictions constant among all networks. Fig. 3(c) depicts the coarse-to-fine training process and network ensembling techniques employed in our approach." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 205, + 254, + 216 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 205, + 254, + 216 + ], + "spans": [ + { + "bbox": [ + 131, + 205, + 254, + 216 + ], + "type": "text", + "content": "3.4 Dataset Collection" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 223, + 482, + 305 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 223, + 482, + 305 + ], + "spans": [ + { + "bbox": [ + 130, + 223, + 482, + 305 + ], + "type": "text", + "content": "Most event-to-video reconstruction approaches are typically evaluated on event datasets collected using DAVIS sensors [5], such as IJRR [27], HQF [49], MVSEC [61], and CED [42]. Recently, other types of event cameras [1, 2, 39] have been developed, which share the same event generation model with DAVIS sensors but may differ in detailed configurations and settings. Therefore, it is essential to verify the generalization capability of existing methods and our EvINR on those new types of event sensors." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 307, + 482, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 307, + 482, + 462 + ], + "spans": [ + { + "bbox": [ + 130, + 307, + 482, + 462 + ], + "type": "text", + "content": "To fill the gap, we introduce a new real-world dataset, called the ALPIX Event Dataset (AED), which is collected using an ALPIX-Eiger event camera [1], featuring static scenes accompanied by gradual camera motions. The camera provides well-aligned RGB frames and color events. The RGB frames have a resolution of " + }, + { + "bbox": [ + 130, + 307, + 482, + 462 + ], + "type": "inline_equation", + "content": "3264 \\times 2448" + }, + { + "bbox": [ + 130, + 307, + 482, + 462 + ], + "type": "text", + "content": " and the events have a resolution of " + }, + { + "bbox": [ + 130, + 307, + 482, + 462 + ], + "type": "inline_equation", + "content": "1632 \\times 1224" + }, + { + "bbox": [ + 130, + 307, + 482, + 462 + ], + "type": "text", + "content": ". The AED dataset includes seven video sequences with diverse scenes, such as streets, buildings, indoor scenes, textures, and tools, and each approximately lasts for ten seconds. Note that, in this paper, we only focus on reconstructing grayscale frames to keep consistency with previous works [31, 60] and also for a fair comparison. Therefore, we first demosaic the RGB frames and events according to the Quad Bayer pattern, resulting in intensity frames with a resolution of " + }, + { + "bbox": [ + 130, + 307, + 482, + 462 + ], + "type": "inline_equation", + "content": "816 \\times 612" + }, + { + "bbox": [ + 130, + 307, + 482, + 462 + ], + "type": "text", + "content": " and event data with a resolution of " + }, + { + "bbox": [ + 130, + 307, + 482, + 462 + ], + "type": "inline_equation", + "content": "408 \\times 306" + }, + { + "bbox": [ + 130, + 307, + 482, + 462 + ], + "type": "text", + "content": ". Details on the post-processing of AED dataset can be found in the supplementary material." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 479, + 230, + 493 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 479, + 230, + 493 + ], + "spans": [ + { + "bbox": [ + 131, + 479, + 230, + 493 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 503, + 269, + 516 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 503, + 269, + 516 + ], + "spans": [ + { + "bbox": [ + 131, + 503, + 269, + 516 + ], + "type": "text", + "content": "4.1 Experiments Settings" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 521, + 482, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 482, + 628 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 482, + 628 + ], + "type": "text", + "content": "Datasets: We conduct experiments to evaluate the effectiveness of our proposed method using three datasets: IJRR [27], HQF [49], and our AED dataset. The IJRR dataset consists of intensity frames and events in 25 real scenes and 2 synthetic scenes, captured by a DAVIS240C camera [5]. HQF dataset provides 14 event data sequences captured with two DAVIS240C cameras, delivering well-exposed and clear intensity frames. The spatial resolution of both the IJRR and HQF datasets is " + }, + { + "bbox": [ + 130, + 521, + 482, + 628 + ], + "type": "inline_equation", + "content": "240 \\times 180" + }, + { + "bbox": [ + 130, + 521, + 482, + 628 + ], + "type": "text", + "content": ". The AED dataset contains 7 event sequences with a resolution of " + }, + { + "bbox": [ + 130, + 521, + 482, + 628 + ], + "type": "inline_equation", + "content": "408 \\times 306" + }, + { + "bbox": [ + 130, + 521, + 482, + 628 + ], + "type": "text", + "content": ". More details of the exact time split can be found in the supplementary material." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 630, + 482, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 630, + 482, + 667 + ], + "spans": [ + { + "bbox": [ + 131, + 630, + 482, + 667 + ], + "type": "text", + "content": "Evaluation Metrics: We evaluate the efficacy of our method using several image quality metrics, including mean squared error (MSE), structural similarity (SSIM) [52], and learned perceptual image patch similarity (LPIPS) [57]." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 417, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 417, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 417, + 91, + 447, + 100 + ], + "type": "text", + "content": "EvINR" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 481, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 138, + 118, + 476, + 383 + ], + "blocks": [ + { + "bbox": [ + 138, + 118, + 476, + 383 + ], + "lines": [ + { + "bbox": [ + 138, + 118, + 476, + 383 + ], + "spans": [ + { + "bbox": [ + 138, + 118, + 476, + 383 + ], + "type": "image", + "image_path": "b88b076fc57a82297a1633d086d37ebbae2326e362e773ae4316c839a17975f9.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 392, + 482, + 415 + ], + "lines": [ + { + "bbox": [ + 130, + 392, + 482, + 415 + ], + "spans": [ + { + "bbox": [ + 130, + 392, + 482, + 415 + ], + "type": "text", + "content": "Fig. 4: Qualitative comparison with baseline methods on IJRR(Row 1&2), HQF (Row 3&4) and AED(Row 5&6)." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 437, + 482, + 570 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 437, + 482, + 570 + ], + "spans": [ + { + "bbox": [ + 130, + 437, + 482, + 570 + ], + "type": "text", + "content": "Implementation Details: We use a SIREN network [46] with three hidden layers and 512 neurons per layer as our INR network. We partition each event sequence into sub-sequences lasting 5 seconds each and concurrently train all subsequences using model ensembling for 300 iterations. We first stack the events within " + }, + { + "bbox": [ + 130, + 437, + 482, + 570 + ], + "type": "inline_equation", + "content": "\\frac{1}{32}" + }, + { + "bbox": [ + 130, + 437, + 482, + 570 + ], + "type": "text", + "content": " second into an event frame and temporal upsample the event frames in 100 and 200 iterations, as described in Sec. 3.3. We adopt the Adam optimizer [23] with a learning rate of " + }, + { + "bbox": [ + 130, + 437, + 482, + 570 + ], + "type": "inline_equation", + "content": "1\\mathrm{e} - 4" + }, + { + "bbox": [ + 130, + 437, + 482, + 570 + ], + "type": "text", + "content": " and exponentially reduce the learning rate every 10 iterations with a decay rate of 0.95. The weight of spatial regularization " + }, + { + "bbox": [ + 130, + 437, + 482, + 570 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 130, + 437, + 482, + 570 + ], + "type": "text", + "content": " is set to 0.05, and " + }, + { + "bbox": [ + 130, + 437, + 482, + 570 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 130, + 437, + 482, + 570 + ], + "type": "text", + "content": " in Eq. 10 is set to 0.6. The activated threshold " + }, + { + "bbox": [ + 130, + 437, + 482, + 570 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 130, + 437, + 482, + 570 + ], + "type": "text", + "content": " is set to 1 for the IJRR and HQF datasets and 0.25 for the AED dataset. The training process takes approximately 8 seconds on a single RTX3090 GPU." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 586, + 340, + 597 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 586, + 340, + 597 + ], + "spans": [ + { + "bbox": [ + 132, + 586, + 340, + 597 + ], + "type": "text", + "content": "4.2 Evaluation of Video Reconstruction" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 482, + 666 + ], + "type": "text", + "content": "We assess the effectiveness of our approach by comparing it against eight state-of-the-art (SoTA) methods, classified based on the amount of data required for training. The methods are categorized into supervised learning (SL) methods that utilize synthetic ground-truth intensity frames for supervision and self-supervised learning (SSL) methods that rely solely on event data. Specifically, we" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "type": "text", + "content": "Z. Wang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 115, + 482, + 248 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 115, + 482, + 248 + ], + "spans": [ + { + "bbox": [ + 130, + 115, + 482, + 248 + ], + "type": "text", + "content": "compare against: 1) Five SL methods: FireNet [41], E2VID [35], FireNet+ [49], E2VID+ [49], and ET-Net [53]. 2) Three SSL method: SSL-E2VID [31], HF [40] and ELRP [60]. Reconstructed results for all methods were generated at each timestamp of the intensity frame. We use optical flow estimation from the FlowNet of SSL-E2VID [31] for ELRP [60]. We apply Contrast Limited Adaptive Histogram Equalization (CLAHE) [54] to both ground truth and synthesized frames before evaluation, following [31]. Note that the quantitative results for FireNet [41], E2VID [35], FireNet+ [49], E2VID+ [49], and ET-Net [53] are referred from a recent benchmark paper [11]. For the visual comparison, we obtained the visualization results by downloading the publicly available checkpoints and test them in the local environment." + } + ] + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 148, + 312, + 462, + 432 + ], + "blocks": [ + { + "bbox": [ + 130, + 269, + 482, + 304 + ], + "lines": [ + { + "bbox": [ + 130, + 269, + 482, + 304 + ], + "spans": [ + { + "bbox": [ + 130, + 269, + 482, + 304 + ], + "type": "text", + "content": "Table 1: Comparison of quantitative results on the IJRR, HQF, and AED datasets. Bold values indicate the best results among all methods, while underlined values indicate the best results among SSL methods." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 148, + 312, + 462, + 432 + ], + "lines": [ + { + "bbox": [ + 148, + 312, + 462, + 432 + ], + "spans": [ + { + "bbox": [ + 148, + 312, + 462, + 432 + ], + "type": "table", + "html": "
MethodsIJRRHQFAED
MSESSIMLPIPSMSESSIMLPIPSMSESSIMLPIPS
SLFireNet [41]0.1310.5020.3200.0940.5330.4410.0740.2980.579
E2VID [35]0.2120.4240.3500.1270.5400.3820.0560.4240.500
FireNet+ [49]0.0630.5550.2900.0400.6140.3140.0940.2320.566
E2VID+ [49]0.0700.5600.2360.0360.6430.2520.0740.3450.462
ET-Net [53]0.0470.6170.2240.0320.6580.2600.0840.3120.482
SSLSSL_E2VID [31]0.0970.4730.4090.0700.4800.4640.0940.3160.453
HF [40]0.1640.3340.6580.1330.2320.6700.0800.2400.943
ELRP [60]0.0800.4370.4850.0740.4500.4740.0840.3050.473
Ours0.0470.6280.2510.0480.5310.3330.0670.4580.366
", + "image_path": "7d7d49329c23b134ec783b24ffc0eff2cc33e0395c731e60816f58ece0b7c846.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 460, + 482, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 460, + 482, + 581 + ], + "spans": [ + { + "bbox": [ + 130, + 460, + 482, + 581 + ], + "type": "text", + "content": "The quantitative results are presented in Table 1. Our method demonstrates superior performance compared to the best SL methods, with improvements of " + }, + { + "bbox": [ + 130, + 460, + 482, + 581 + ], + "type": "inline_equation", + "content": "7\\%" + }, + { + "bbox": [ + 130, + 460, + 482, + 581 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 460, + 482, + 581 + ], + "type": "inline_equation", + "content": "13\\%" + }, + { + "bbox": [ + 130, + 460, + 482, + 581 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 460, + 482, + 581 + ], + "type": "inline_equation", + "content": "4\\%" + }, + { + "bbox": [ + 130, + 460, + 482, + 581 + ], + "type": "text", + "content": " in terms of MSE, SSIM, and LPIPS respectively on the IJRR dataset [27]. Although the gap between our method and SL methods widens on the HQF dataset [49] due to the lower event density, our method remains comparable. On the AED dataset, our method outperforms the state-of-the-art SL methods, E2VID+ and ET-Net, by a clear margin for all three metrics. Notably, our method shows a significant improvement over previous SSL methods. In particular, compared with SSL-E2VID, it improves MSE, SSIM, and LPIPS by " + }, + { + "bbox": [ + 130, + 460, + 482, + 581 + ], + "type": "inline_equation", + "content": "35\\%" + }, + { + "bbox": [ + 130, + 460, + 482, + 581 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 130, + 460, + 482, + 581 + ], + "type": "inline_equation", + "content": "25\\%" + }, + { + "bbox": [ + 130, + 460, + 482, + 581 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 460, + 482, + 581 + ], + "type": "inline_equation", + "content": "21\\%" + }, + { + "bbox": [ + 130, + 460, + 482, + 581 + ], + "type": "text", + "content": ", respectively." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 582, + 482, + 666 + ], + "type": "text", + "content": "A qualitative comparison of our method with the baseline methods is depicted in Fig. 4. Our method produces intensity frames with better contrast and overall visual quality, compared with other methods that suffer from issues such as foggy effects, artifacts, and loss of detailed structures. It is worth noting that the performance of E2VID+ and ET-Net degrades significantly on the AED dataset, which suggests that their training strategy may overfit the DAVIS sensor setting. Additional results can be found in the supplementary material." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 417, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 417, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 417, + 91, + 447, + 100 + ], + "type": "text", + "content": "EvINR" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 480, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 162, + 118, + 234, + 177 + ], + "blocks": [ + { + "bbox": [ + 162, + 118, + 234, + 177 + ], + "lines": [ + { + "bbox": [ + 162, + 118, + 234, + 177 + ], + "spans": [ + { + "bbox": [ + 162, + 118, + 234, + 177 + ], + "type": "image", + "image_path": "f7a9257782005390f2966a919666006df291eb6866ff46bc08560ada41cdefdf.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 180, + 178, + 217, + 186 + ], + "lines": [ + { + "bbox": [ + 180, + 178, + 217, + 186 + ], + "spans": [ + { + "bbox": [ + 180, + 178, + 217, + 186 + ], + "type": "text", + "content": "Ground Truth" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 235, + 118, + 306, + 177 + ], + "blocks": [ + { + "bbox": [ + 235, + 118, + 306, + 177 + ], + "lines": [ + { + "bbox": [ + 235, + 118, + 306, + 177 + ], + "spans": [ + { + "bbox": [ + 235, + 118, + 306, + 177 + ], + "type": "image", + "image_path": "0ba6338b8a4ac641f5742e6d3e509e7169edf9069876090374e97d3dfd0be88c.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 261, + 178, + 283, + 186 + ], + "lines": [ + { + "bbox": [ + 261, + 178, + 283, + 186 + ], + "spans": [ + { + "bbox": [ + 261, + 178, + 283, + 186 + ], + "type": "text", + "content": "No Reg." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 132, + 196, + 479, + 209 + ], + "lines": [ + { + "bbox": [ + 132, + 196, + 479, + 209 + ], + "spans": [ + { + "bbox": [ + 132, + 196, + 479, + 209 + ], + "type": "text", + "content": "Fig. 5: The impact of removing the spatial regularization and coarse-to-fine training." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 306, + 118, + 377, + 177 + ], + "blocks": [ + { + "bbox": [ + 306, + 118, + 377, + 177 + ], + "lines": [ + { + "bbox": [ + 306, + 118, + 377, + 177 + ], + "spans": [ + { + "bbox": [ + 306, + 118, + 377, + 177 + ], + "type": "image", + "image_path": "c940bc1649a90698a9eac470e635f639ce71cc57ba404872ea2f51d56821a8d3.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 328, + 178, + 351, + 186 + ], + "lines": [ + { + "bbox": [ + 328, + 178, + 351, + 186 + ], + "spans": [ + { + "bbox": [ + 328, + 178, + 351, + 186 + ], + "type": "text", + "content": "No " + }, + { + "bbox": [ + 328, + 178, + 351, + 186 + ], + "type": "inline_equation", + "content": "C2T" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 377, + 118, + 449, + 177 + ], + "blocks": [ + { + "bbox": [ + 377, + 118, + 449, + 177 + ], + "lines": [ + { + "bbox": [ + 377, + 118, + 449, + 177 + ], + "spans": [ + { + "bbox": [ + 377, + 118, + 449, + 177 + ], + "type": "image", + "image_path": "4b3f6d33b3d4a79d16b6af3d999865c30834d84df3a12cfecb774900ea55ec5e.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 399, + 178, + 429, + 186 + ], + "lines": [ + { + "bbox": [ + 399, + 178, + 429, + 186 + ], + "spans": [ + { + "bbox": [ + 399, + 178, + 429, + 186 + ], + "type": "text", + "content": "Full Model" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 228, + 251, + 382, + 301 + ], + "blocks": [ + { + "bbox": [ + 132, + 218, + 481, + 240 + ], + "lines": [ + { + "bbox": [ + 132, + 218, + 481, + 240 + ], + "spans": [ + { + "bbox": [ + 132, + 218, + 481, + 240 + ], + "type": "text", + "content": "Table 2: Ablation of the spatial regularization term (Reg.) and coarse-to-fine training " + }, + { + "bbox": [ + 132, + 218, + 481, + 240 + ], + "type": "inline_equation", + "content": "(C2F_{\\cdot})" + }, + { + "bbox": [ + 132, + 218, + 481, + 240 + ], + "type": "text", + "content": " on the IJRR [27] dataset." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 228, + 251, + 382, + 301 + ], + "lines": [ + { + "bbox": [ + 228, + 251, + 382, + 301 + ], + "spans": [ + { + "bbox": [ + 228, + 251, + 382, + 301 + ], + "type": "table", + "html": "
MSESSIMLPIPSTime(s)FPS
Base0.0630.5420.35717.2913.77
Base+Reg.0.0440.6120.27517.4413.64
Base+C2F.0.0610.5660.3437.9431.33
Full Model0.0470.6580.2518.0829.45
", + "image_path": "5a41ba43443cfceab2f1e8187a11f7f91e9bd95d10d1a6a8e6908181ab66c90b.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 323, + 238, + 335 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 323, + 238, + 335 + ], + "spans": [ + { + "bbox": [ + 132, + 323, + 238, + 335 + ], + "type": "text", + "content": "4.3 Ablation Study" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 342, + 482, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 342, + 482, + 498 + ], + "spans": [ + { + "bbox": [ + 130, + 342, + 482, + 498 + ], + "type": "text", + "content": "We conduct ablation experiments on the IJRR dataset to assess the significance of each individual component of our method, and the results are presented in Tab. 2. Our method's basic implementation, which doesn't include spatial regularization term as described in Sec. 3.2 or use coarse-to-fine training as described in Sec. 3.3, is called Base. We also evaluate the performance of two modified versions of the full model: " + }, + { + "bbox": [ + 130, + 342, + 482, + 498 + ], + "type": "inline_equation", + "content": "\\text{Base} + \\text{Reg}" + }, + { + "bbox": [ + 130, + 342, + 482, + 498 + ], + "type": "text", + "content": ", which adds a spatial regularization term (Eq. 7) with " + }, + { + "bbox": [ + 130, + 342, + 482, + 498 + ], + "type": "inline_equation", + "content": "\\lambda = 0.05" + }, + { + "bbox": [ + 130, + 342, + 482, + 498 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 342, + 482, + 498 + ], + "type": "inline_equation", + "content": "\\text{Base} + C2F" + }, + { + "bbox": [ + 130, + 342, + 482, + 498 + ], + "type": "text", + "content": ", which uses coarse-to-fine training. Experimental results show that both spatial regularization and coarse-to-fine training contribute to the improved performance of our full model. Fig. 5 confirms the effectiveness of these two modules, showing that removing the spatial regularization term introduces noticeable noise (e.g., in the background), while removing the coarse-to-fine training scheme leads to an increase in training time and missing details (e.g., in the people's clothes and the book)." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 514, + 218, + 525 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 514, + 218, + 525 + ], + "spans": [ + { + "bbox": [ + 132, + 514, + 218, + 525 + ], + "type": "text", + "content": "4.4 Discussions" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": "Training speed: To assess the training speed of EvINR, we conducted an analysis using various model ensembling configurations on the whole Calibration sequence from the IJRR dataset [27]. We divide the event sequence, which had a total duration of 50 seconds, into " + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": " partitions, with each partition having a duration of " + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": " seconds. Subsequently, we trained " + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": " EvINR models in parallel, incorporating the model ensembling techniques outlined in Sec. 3.4. The experimental findings are presented in Table 3. By training 10 EvINR models, each corresponding to 5 seconds of the event sequence, we achieved a performance gain of over " + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "inline_equation", + "content": "26\\%" + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": ", while only requiring an additional " + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "inline_equation", + "content": "14\\%" + }, + { + "bbox": [ + 130, + 533, + 482, + 666 + ], + "type": "text", + "content": " of training time compared to training a single EvINR model on the entire sequence. However, we observed a significant drop in performance when increasing the partition size to" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "type": "text", + "content": "Z. Wang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 235, + 147, + 376, + 198 + ], + "blocks": [ + { + "bbox": [ + 131, + 114, + 479, + 137 + ], + "lines": [ + { + "bbox": [ + 131, + 114, + 479, + 137 + ], + "spans": [ + { + "bbox": [ + 131, + 114, + 479, + 137 + ], + "type": "text", + "content": "Table 3: Impact of hyper-parameters of model ensembling on the training speed and reconstruction performance." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 235, + 147, + 376, + 198 + ], + "lines": [ + { + "bbox": [ + 235, + 147, + 376, + 198 + ], + "spans": [ + { + "bbox": [ + 235, + 147, + 376, + 198 + ], + "type": "table", + "html": "
NτMSESSIMLPIPSTime(s)FPS
1500.0840.460.38015.7175.66
5100.06880.6000.25316.9170.31
1050.06250.6120.24418.3264.90
5010.08750.4880.34329.4340.39
", + "image_path": "381a6eb941d4682e436184331fcc3c8c50d210e2348003a6740e6488de7fbecc.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 220, + 479, + 268 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 220, + 479, + 268 + ], + "spans": [ + { + "bbox": [ + 130, + 220, + 479, + 268 + ], + "type": "text", + "content": "50, suggesting that a 1-second sequence is insufficient for EvINR to converge to a stable solution of the event generation equation. We also note that the optimal choice of hyper-parameters may differ for different event sequences due to the per-scene optimization nature of our INR approach." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 269, + 480, + 341 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 269, + 480, + 341 + ], + "spans": [ + { + "bbox": [ + 130, + 269, + 480, + 341 + ], + "type": "text", + "content": "Event enhancement: Our approach provides a smooth and continuous representation of event data by representing events triggered within a small time window " + }, + { + "bbox": [ + 130, + 269, + 480, + 341 + ], + "type": "inline_equation", + "content": "\\Delta t" + }, + { + "bbox": [ + 130, + 269, + 480, + 341 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 130, + 269, + 480, + 341 + ], + "type": "inline_equation", + "content": "\\frac{\\partial F_{\\Theta}}{\\partial t}\\Delta t" + }, + { + "bbox": [ + 130, + 269, + 480, + 341 + ], + "type": "text", + "content": ". This event representation automatically reduces noise and preserves critical information through INR optimization, making it highly robust to noise. Fig. 6 compares the denoising results of our INR representation with several SoTA denoising methods [12, 51]." + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 170, + 365, + 237, + 415 + ], + "blocks": [ + { + "bbox": [ + 170, + 365, + 237, + 415 + ], + "lines": [ + { + "bbox": [ + 170, + 365, + 237, + 415 + ], + "spans": [ + { + "bbox": [ + 170, + 365, + 237, + 415 + ], + "type": "image", + "image_path": "8032d3c12d165de605fa1625b9fa5542f9857961e8e68a34622950fe860b7716.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 176, + 415, + 238, + 465 + ], + "blocks": [ + { + "bbox": [ + 176, + 415, + 238, + 465 + ], + "lines": [ + { + "bbox": [ + 176, + 415, + 238, + 465 + ], + "spans": [ + { + "bbox": [ + 176, + 415, + 238, + 465 + ], + "type": "image", + "image_path": "7d1ac315e7e67e7471f35e452d773201290c823b840a7b64ecf2033f15067ebb.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 186, + 467, + 223, + 475 + ], + "lines": [ + { + "bbox": [ + 186, + 467, + 223, + 475 + ], + "spans": [ + { + "bbox": [ + 186, + 467, + 223, + 475 + ], + "type": "text", + "content": "Event Frame" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 237, + 365, + 302, + 415 + ], + "blocks": [ + { + "bbox": [ + 237, + 365, + 302, + 415 + ], + "lines": [ + { + "bbox": [ + 237, + 365, + 302, + 415 + ], + "spans": [ + { + "bbox": [ + 237, + 365, + 302, + 415 + ], + "type": "image", + "image_path": "4b506b332bff3bcdee37dfb10d1054f532c888f8b80af291913e10bb45257318.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 238, + 415, + 303, + 465 + ], + "blocks": [ + { + "bbox": [ + 238, + 415, + 303, + 465 + ], + "lines": [ + { + "bbox": [ + 238, + 415, + 303, + 465 + ], + "spans": [ + { + "bbox": [ + 238, + 415, + 303, + 465 + ], + "type": "image", + "image_path": "39a772ba0b3e40f480c2aaea9a8513d99950c45bb8ef43c37455a23c3e7bb1b0.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 268, + 468, + 278, + 475 + ], + "lines": [ + { + "bbox": [ + 268, + 468, + 278, + 475 + ], + "spans": [ + { + "bbox": [ + 268, + 468, + 278, + 475 + ], + "type": "text", + "content": "DF" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 303, + 365, + 369, + 415 + ], + "blocks": [ + { + "bbox": [ + 303, + 365, + 369, + 415 + ], + "lines": [ + { + "bbox": [ + 303, + 365, + 369, + 415 + ], + "spans": [ + { + "bbox": [ + 303, + 365, + 369, + 415 + ], + "type": "image", + "image_path": "1619bce478bbddd37d6b8151767cf916315f343e5f9d286d3f61a5aeea8591cb.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 304, + 415, + 369, + 465 + ], + "blocks": [ + { + "bbox": [ + 304, + 415, + 369, + 465 + ], + "lines": [ + { + "bbox": [ + 304, + 415, + 369, + 465 + ], + "spans": [ + { + "bbox": [ + 304, + 415, + 369, + 465 + ], + "type": "image", + "image_path": "a4710b8d3db01e9bf2f8041c897cbe2cac524d9278a7a4ee6a346d78a05f7848.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 325, + 468, + 346, + 475 + ], + "lines": [ + { + "bbox": [ + 325, + 468, + 346, + 475 + ], + "spans": [ + { + "bbox": [ + 325, + 468, + 346, + 475 + ], + "type": "text", + "content": "EvGait" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 374, + 365, + 442, + 466 + ], + "blocks": [ + { + "bbox": [ + 374, + 365, + 442, + 466 + ], + "lines": [ + { + "bbox": [ + 374, + 365, + 442, + 466 + ], + "spans": [ + { + "bbox": [ + 374, + 365, + 442, + 466 + ], + "type": "image", + "image_path": "9d82d6ca8bbd68704ad5e127b71bf7764d211f739eadb2178620b7bd99c94080.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 398, + 468, + 418, + 475 + ], + "lines": [ + { + "bbox": [ + 398, + 468, + 418, + 475 + ], + "spans": [ + { + "bbox": [ + 398, + 468, + 418, + 475 + ], + "type": "text", + "content": "EvINR" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 141, + 487, + 470, + 498 + ], + "lines": [ + { + "bbox": [ + 141, + 487, + 470, + 498 + ], + "spans": [ + { + "bbox": [ + 141, + 487, + 470, + 498 + ], + "type": "text", + "content": "Fig. 6: Qualitative comparison with baseline methods of event stream denoising." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 521, + 480, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 521, + 480, + 616 + ], + "spans": [ + { + "bbox": [ + 130, + 521, + 480, + 616 + ], + "type": "text", + "content": "Early frames reconstruction: [6] suggested that RNN-based techniques for video reconstruction, e.g. E2VID [35], require an initialization period to achieve satisfactory results. Consequently, the initial frames generated by these methods may be of poor quality, restricting their usefulness to short event sequences. Conversely, our proposed approach is capable of producing realistic outcomes using a minimal number of events. Fig. 7 illustrates a comparison of the first frame generated by our method and RNN-based techniques, demonstrating our ability to rapidly create high-quality outputs with minimal input." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 130, + 617, + 480, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 480, + 665 + ], + "type": "text", + "content": "Inference speed: Compared to other SoTA approaches, our method offers a significant improvement in terms of inference speed, as demonstrated in Tab. 4. The main reason for this improvement is that other methods require converting event data into event frames or voxel grids during the inference period, which" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 417, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 417, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 417, + 91, + 447, + 100 + ], + "type": "text", + "content": "EvINR" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 186, + 120, + 265, + 167 + ], + "blocks": [ + { + "bbox": [ + 186, + 120, + 265, + 167 + ], + "lines": [ + { + "bbox": [ + 186, + 120, + 265, + 167 + ], + "spans": [ + { + "bbox": [ + 186, + 120, + 265, + 167 + ], + "type": "image", + "image_path": "c35390e0359931825e3b4039a2fd20f9b91363e11be62017c32f36dd25f9aefd.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 209, + 169, + 243, + 178 + ], + "lines": [ + { + "bbox": [ + 209, + 169, + 243, + 178 + ], + "spans": [ + { + "bbox": [ + 209, + 169, + 243, + 178 + ], + "type": "text", + "content": "E2VID[34]" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 266, + 120, + 344, + 167 + ], + "blocks": [ + { + "bbox": [ + 266, + 120, + 344, + 167 + ], + "lines": [ + { + "bbox": [ + 266, + 120, + 344, + 167 + ], + "spans": [ + { + "bbox": [ + 266, + 120, + 344, + 167 + ], + "type": "image", + "image_path": "2dfe2d1cc82bb52f4b5921891797685b286cfcba654471c19da9a619e1c6f688.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 288, + 169, + 326, + 178 + ], + "lines": [ + { + "bbox": [ + 288, + 169, + 326, + 178 + ], + "spans": [ + { + "bbox": [ + 288, + 169, + 326, + 178 + ], + "type": "text", + "content": "E2VID+[47]" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 345, + 120, + 423, + 167 + ], + "blocks": [ + { + "bbox": [ + 345, + 120, + 423, + 167 + ], + "lines": [ + { + "bbox": [ + 345, + 120, + 423, + 167 + ], + "spans": [ + { + "bbox": [ + 345, + 120, + 423, + 167 + ], + "type": "image", + "image_path": "3730b3787cf9da2d3280bdb17a773f1d521a9cf1c60450669f2456005a04886a.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 371, + 169, + 406, + 178 + ], + "lines": [ + { + "bbox": [ + 371, + 169, + 406, + 178 + ], + "spans": [ + { + "bbox": [ + 371, + 169, + 406, + 178 + ], + "type": "text", + "content": "ET-Net[51]" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 186, + 183, + 265, + 231 + ], + "blocks": [ + { + "bbox": [ + 186, + 183, + 265, + 231 + ], + "lines": [ + { + "bbox": [ + 186, + 183, + 265, + 231 + ], + "spans": [ + { + "bbox": [ + 186, + 183, + 265, + 231 + ], + "type": "image", + "image_path": "42df6c93f0106a00743e6713913223fe206d95d2a24985799ef5631daaf62778.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 200, + 232, + 249, + 240 + ], + "lines": [ + { + "bbox": [ + 200, + 232, + 249, + 240 + ], + "spans": [ + { + "bbox": [ + 200, + 232, + 249, + 240 + ], + "type": "text", + "content": "SSL-E2VID[30]" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 266, + 183, + 344, + 231 + ], + "blocks": [ + { + "bbox": [ + 266, + 183, + 344, + 231 + ], + "lines": [ + { + "bbox": [ + 266, + 183, + 344, + 231 + ], + "spans": [ + { + "bbox": [ + 266, + 183, + 344, + 231 + ], + "type": "image", + "image_path": "b1fb2954e7b8627e18580774eae86cbc05ee4ed39006f3f28b188cba67cb09ff.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 293, + 232, + 315, + 240 + ], + "lines": [ + { + "bbox": [ + 293, + 232, + 315, + 240 + ], + "spans": [ + { + "bbox": [ + 293, + 232, + 315, + 240 + ], + "type": "text", + "content": "EvINR" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 345, + 183, + 423, + 231 + ], + "blocks": [ + { + "bbox": [ + 345, + 183, + 423, + 231 + ], + "lines": [ + { + "bbox": [ + 345, + 183, + 423, + 231 + ], + "spans": [ + { + "bbox": [ + 345, + 183, + 423, + 231 + ], + "type": "image", + "image_path": "4591044b040d257a1814d11a0bea96ca0867343a5f7f537aef2b2e855115169c.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 377, + 232, + 388, + 240 + ], + "lines": [ + { + "bbox": [ + 377, + 232, + 388, + 240 + ], + "spans": [ + { + "bbox": [ + 377, + 232, + 388, + 240 + ], + "type": "text", + "content": "GT" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 156, + 250, + 456, + 262 + ], + "lines": [ + { + "bbox": [ + 156, + 250, + 456, + 262 + ], + "spans": [ + { + "bbox": [ + 156, + 250, + 456, + 262 + ], + "type": "text", + "content": "Fig. 7: The reconstructed initial frame from different compared methods." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "table", + "bbox": [ + 217, + 307, + 394, + 350 + ], + "blocks": [ + { + "bbox": [ + 131, + 274, + 481, + 298 + ], + "lines": [ + { + "bbox": [ + 131, + 274, + 481, + 298 + ], + "spans": [ + { + "bbox": [ + 131, + 274, + 481, + 298 + ], + "type": "text", + "content": "Table 4: Comparison of inference time in terms of frames-per-second (fps) at two resolutions from IJRR [27] and AED dataset respectively." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 217, + 307, + 394, + 350 + ], + "lines": [ + { + "bbox": [ + 217, + 307, + 394, + 350 + ], + "spans": [ + { + "bbox": [ + 217, + 307, + 394, + 350 + ], + "type": "table", + "html": "
ResolutionFireNetE2VIDET-NetOurs
(240, 180)118.24100.9636.1178.19
(408, 306)28.8328.5618.2755.37
", + "image_path": "5129d4987ce996e606c1c1cf0475cf098116a247a215a2e447d24b88f0193828.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "table_body" + } + ], + "index": 16 + }, + { + "bbox": [ + 130, + 374, + 480, + 398 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 374, + 480, + 398 + ], + "spans": [ + { + "bbox": [ + 130, + 374, + 480, + 398 + ], + "type": "text", + "content": "adds significant overhead. However, our method only requires the spatiotemporal coordinates as input, thus avoiding this issue." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 131, + 422, + 220, + 434 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 422, + 220, + 434 + ], + "spans": [ + { + "bbox": [ + 131, + 422, + 220, + 434 + ], + "type": "text", + "content": "5 Conclusion" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 130, + 449, + 482, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 449, + 482, + 616 + ], + "spans": [ + { + "bbox": [ + 130, + 449, + 482, + 616 + ], + "type": "text", + "content": "This paper introduced EvINR, a SSL method for event-to-video reconstruction that relieves the need for synthetic data or optical flow estimation. We, for the first time, show that high-quality videos can be reconstructed in a self-supervised and interpretable way without time-consuming end-to-end training. Our method is based on directly solving the event generation model via optimizing an INR whose temporal derivative is self-supervised by events and spatial derivative is regularized to reduce artifacts. Additionally, we propose several acceleration techniques that significantly reduce the training time of EvINR, making it applicable for online tasks. Experiments show that our approach significantly outperforms the previous SSL methods, and is competitive with the SoTA supervised methods. Our approach also demonstrates superior interpretability and robustness to various event dataset. Overall, our work contributes to the advancement of event-to-video reconstruction and offers a promising direction for future research that combines INRs with event data." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 482, + 666 + ], + "type": "text", + "content": "Limitations and Future Work: The current parameter size of EvINR takes up approximately the same amount of storage as the original event data. In future work, we plan to explore network pruning and quantization techniques to further reduce the parameter size." + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "type": "text", + "content": "Z. Wang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 240, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 240, + 129 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 240, + 129 + ], + "type": "text", + "content": "Acknowledgments" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 139, + 481, + 187 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 139, + 481, + 187 + ], + "spans": [ + { + "bbox": [ + 132, + 139, + 481, + 187 + ], + "type": "text", + "content": "This paper is supported by the National Natural Science Foundation of China (NSF) under Grant No. NSFC22FYT45, the Guangzhou City, University and Enterprise Joint Fund under Grant No.SL2022A03J01278, and Guangzhou Fundamental and Applied Basic Research (Grant Number: 2024A04J4072)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 205, + 197, + 217 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 205, + 197, + 217 + ], + "spans": [ + { + "bbox": [ + 132, + 205, + 197, + 217 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 134, + 228, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 138, + 228, + 480, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 228, + 480, + 251 + ], + "spans": [ + { + "bbox": [ + 138, + 228, + 480, + 251 + ], + "type": "text", + "content": "1. Alpsentek products. https://www.alpsentek.com/product, [Accessed: Feb. 17, 2023]" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 251, + 480, + 272 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 251, + 480, + 272 + ], + "spans": [ + { + "bbox": [ + 138, + 251, + 480, + 272 + ], + "type": "text", + "content": "2. Pepperl fuchs event camera. https://www.pepperl-fuchs.com/global/en/classid_11544.htm?view=productgroupliterature, accessed on 17 Feb 2023" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 274, + 480, + 305 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 274, + 480, + 305 + ], + "spans": [ + { + "bbox": [ + 138, + 274, + 480, + 305 + ], + "type": "text", + "content": "3. Bardow, P., Davison, A.J., Leutenegger, S.: Simultaneous optical flow and intensity estimation from an event camera. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 884-892 (2016)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 306, + 452, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 306, + 452, + 316 + ], + "spans": [ + { + "bbox": [ + 138, + 306, + 452, + 316 + ], + "type": "text", + "content": "4. Bishop, C.M.: Pattern Recognition and Machine Learning. Springer (2006)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 317, + 480, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 317, + 480, + 348 + ], + "spans": [ + { + "bbox": [ + 138, + 317, + 480, + 348 + ], + "type": "text", + "content": "5. Brandli, C., Berner, R., Yang, M., Liu, S.C., Delbruck, T.: A " + }, + { + "bbox": [ + 138, + 317, + 480, + 348 + ], + "type": "inline_equation", + "content": "240 \\times 180" + }, + { + "bbox": [ + 138, + 317, + 480, + 348 + ], + "type": "text", + "content": " 130 db " + }, + { + "bbox": [ + 138, + 317, + 480, + 348 + ], + "type": "inline_equation", + "content": "3\\mu s" + }, + { + "bbox": [ + 138, + 317, + 480, + 348 + ], + "type": "text", + "content": " latency global shutter spatiotemporal vision sensor. IEEE Journal of Solid-State Circuits 49(10), 2333-2341 (2014)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 350, + 480, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 350, + 480, + 380 + ], + "spans": [ + { + "bbox": [ + 138, + 350, + 480, + 380 + ], + "type": "text", + "content": "6. Cadena, P.R.G., Qian, Y., Wang, C., Yang, M.: Spade-e2vid: Spatially-adaptive denormalization for event-based video reconstruction. IEEE Transactions on Image Processing 30, 2488-2500 (2021)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 382, + 480, + 414 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 382, + 480, + 414 + ], + "spans": [ + { + "bbox": [ + 138, + 382, + 480, + 414 + ], + "type": "text", + "content": "7. Chen, H., He, B., Wang, H., Ren, Y., Lim, S.N., Shrivastava, A.: Nerv: Neural representations for videos. Advances in Neural Information Processing Systems 34, 21557-21568 (2021)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 415, + 480, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 415, + 480, + 446 + ], + "spans": [ + { + "bbox": [ + 138, + 415, + 480, + 446 + ], + "type": "text", + "content": "8. Chen, Y., Liu, S., Wang, X.: Learning continuous image representation with local implicit image function. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 8628-8638 (2021)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 448, + 480, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 448, + 480, + 479 + ], + "spans": [ + { + "bbox": [ + 138, + 448, + 480, + 479 + ], + "type": "text", + "content": "9. Cook, M., Gugelmann, L., Jug, F., Krautz, C., Steger, A.: Interacting maps for fast visual interpretation. In: The 2011 International Joint Conference on Neural Networks. pp. 770-776. IEEE (2011)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 134, + 480, + 480, + 523 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 480, + 480, + 523 + ], + "spans": [ + { + "bbox": [ + 134, + 480, + 480, + 523 + ], + "type": "text", + "content": "10. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929 (2020)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 134, + 524, + 480, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 524, + 480, + 567 + ], + "spans": [ + { + "bbox": [ + 134, + 524, + 480, + 567 + ], + "type": "text", + "content": "11. Ercan, B., Eker, O., Erdem, A., Erdem, E.: Evreal: Towards a comprehensive benchmark and analysis suite for event-based video reconstruction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3942-3951 (2023)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 134, + 567, + 480, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 567, + 480, + 589 + ], + "spans": [ + { + "bbox": [ + 134, + 567, + 480, + 589 + ], + "type": "text", + "content": "12. Feng, Y., Lv, H., Liu, H., Zhang, Y., Xiao, Y., Han, C.: Event density based denoising method for dynamic vision sensor. Applied Sciences 10(6), 2024 (2020)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 134, + 590, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 590, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 134, + 590, + 480, + 632 + ], + "type": "text", + "content": "13. Gallego, G., Delbrück, T., Orchard, G., Bartolozzi, C., Taba, B., Censi, A., Leutenegger, S., Davison, A.J., Conradt, J., Daniilidis, K., et al.: Event-based vision: A survey. IEEE transactions on pattern analysis and machine intelligence 44(1), 154-180 (2020)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 134, + 633, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 633, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 134, + 633, + 480, + 665 + ], + "type": "text", + "content": "14. Gehrig, D., Gehrig, M., Hidalgo-Carrio, J., Scaramuzza, D.: Video to events: Recycling video datasets for event cameras. In: IEEE Conf. Comput. Vis. Pattern Recog. (CVPR) (June 2020)" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 417, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 417, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 417, + 91, + 447, + 100 + ], + "type": "text", + "content": "EvINR" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 117, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 133, + 117, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 117, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 133, + 117, + 480, + 138 + ], + "type": "text", + "content": "15. Girshick, R.: Fast r-cnn. In: Proceedings of the IEEE international conference on computer vision. pp. 1440-1448 (2015)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 139, + 480, + 159 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 139, + 480, + 159 + ], + "spans": [ + { + "bbox": [ + 133, + 139, + 480, + 159 + ], + "type": "text", + "content": "16. Hansen, L.K., Salamon, P.: Neural network ensembles. IEEE transactions on pattern analysis and machine intelligence 12(10), 993-1001 (1990)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 160, + 480, + 181 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 160, + 480, + 181 + ], + "spans": [ + { + "bbox": [ + 133, + 160, + 480, + 181 + ], + "type": "text", + "content": "17. He, K., Gkioxari, G., Dollár, P., Girshick, R.: Mask r-cnn. In: Proceedings of the IEEE international conference on computer vision. pp. 2961-2969 (2017)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 182, + 480, + 213 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 182, + 480, + 213 + ], + "spans": [ + { + "bbox": [ + 133, + 182, + 480, + 213 + ], + "type": "text", + "content": "18. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 770-778 (2016)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 133, + 214, + 480, + 246 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 214, + 480, + 246 + ], + "spans": [ + { + "bbox": [ + 133, + 214, + 480, + 246 + ], + "type": "text", + "content": "19. Hu, Y., Liu, S.C., Delbruck, T.: v2e: From video frames to realistic dvs events. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1312-1321 (2021)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 246, + 480, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 246, + 480, + 267 + ], + "spans": [ + { + "bbox": [ + 132, + 246, + 480, + 267 + ], + "type": "text", + "content": "20. Hwang, I., Kim, J., Kim, Y.M.: Ev-nerf: Event based neural radiance field. arXiv preprint arXiv:2206.12455 (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 267, + 480, + 289 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 267, + 480, + 289 + ], + "spans": [ + { + "bbox": [ + 132, + 267, + 480, + 289 + ], + "type": "text", + "content": "21. Kim, H., Handa, A., Benosman, R., Ieng, S.H., Davison, A.J.: Simultaneous mosaicing and tracking with an event camera. J. Solid State Circ 43, 566-576 (2008)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 289, + 480, + 320 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 289, + 480, + 320 + ], + "spans": [ + { + "bbox": [ + 132, + 289, + 480, + 320 + ], + "type": "text", + "content": "22. Kim, H., Leutenegger, S., Davison, A.J.: Real-time 3d reconstruction and 6-dof tracking with an event camera. In: European conference on computer vision. pp. 349-364. Springer (2016)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 321, + 480, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 321, + 480, + 342 + ], + "spans": [ + { + "bbox": [ + 132, + 321, + 480, + 342 + ], + "type": "text", + "content": "23. Kingma, D.P., Ba, J.: Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980 (2014)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 342, + 480, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 342, + 480, + 363 + ], + "spans": [ + { + "bbox": [ + 132, + 342, + 480, + 363 + ], + "type": "text", + "content": "24. Klenk, S., Koestler, L., Scaramuzza, D., Cremers, D.: E-nerf: Neural radiance fields from a moving event camera. IEEE Robotics and Automation Letters (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 363, + 480, + 406 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 363, + 480, + 406 + ], + "spans": [ + { + "bbox": [ + 132, + 363, + 480, + 406 + ], + "type": "text", + "content": "25. Mescheder, L., Oechsle, M., Niemeyer, M., Nowozin, S., Geiger, A.: Occupancy networks: Learning 3d reconstruction in function space. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 4460-4470 (2019)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 407, + 480, + 438 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 407, + 480, + 438 + ], + "spans": [ + { + "bbox": [ + 132, + 407, + 480, + 438 + ], + "type": "text", + "content": "26. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM 65(1), 99-106 (2021)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 439, + 480, + 482 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 439, + 480, + 482 + ], + "spans": [ + { + "bbox": [ + 132, + 439, + 480, + 482 + ], + "type": "text", + "content": "27. Mueggler, E., Rebecq, H., Gallego, G., Delbruck, T., Scaramuzza, D.: The event-camera dataset and simulator: Event-based data for pose estimation, visual odometry, and slam. The International Journal of Robotics Research 36(2), 142-149 (2017)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 483, + 480, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 483, + 480, + 514 + ], + "spans": [ + { + "bbox": [ + 132, + 483, + 480, + 514 + ], + "type": "text", + "content": "28. Munda, G., Reinbacher, C., Pock, T.: Real-time intensity-image reconstruction for event cameras using manifold regularisation. International Journal of Computer Vision 126(12), 1381-1393 (2018)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 514, + 480, + 546 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 514, + 480, + 546 + ], + "spans": [ + { + "bbox": [ + 132, + 514, + 480, + 546 + ], + "type": "text", + "content": "29. Niemeyer, M., Geiger, A.: Giraffe: Representing scenes as compositional generative neural feature fields. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11453-11464 (2021)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 547, + 480, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 547, + 480, + 578 + ], + "spans": [ + { + "bbox": [ + 132, + 547, + 480, + 578 + ], + "type": "text", + "content": "30. Oechsle, M., Peng, S., Geiger, A.: Unisurf: Unifying neural implicit surfaces and radiance fields for multi-view reconstruction. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5589-5599 (2021)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 578, + 480, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 578, + 480, + 621 + ], + "spans": [ + { + "bbox": [ + 132, + 578, + 480, + 621 + ], + "type": "text", + "content": "31. Paredes-Vallés, F., de Croon, G.C.: Back to event basics: Self-supervised learning of image reconstruction for event cameras via photometric constancy. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3446-3455 (2021)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 622, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 622, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 622, + 480, + 665 + ], + "type": "text", + "content": "32. Park, J.J., Florence, P., Straub, J., Newcombe, R., Lovegrove, S.: Deepsdf: Learning continuous signed distance functions for shape representation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 165-174 (2019)" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "type": "text", + "content": "Z. Wang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 130, + 116, + 480, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 480, + 138 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 480, + 138 + ], + "type": "text", + "content": "33. Paszke, A., Gross, S., Chintala, S., Chanan, G., Yang, E., DeVito, Z., Lin, Z., Desmaison, A., Antiga, L., Lerer, A.: Automatic differentiation in pytorch (2017)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 140, + 480, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 140, + 480, + 161 + ], + "spans": [ + { + "bbox": [ + 130, + 140, + 480, + 161 + ], + "type": "text", + "content": "34. Pini, S., Borghi, G., Vezzani, R.: Learn to see by events: Color frame synthesis from event and rgb cameras. arXiv preprint arXiv:1812.02041 (2018)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 163, + 480, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 163, + 480, + 195 + ], + "spans": [ + { + "bbox": [ + 132, + 163, + 480, + 195 + ], + "type": "text", + "content": "35. Rebecq, H., Ranftl, R., Koltun, V., Scaramuzza, D.: Events-to-video: Bringing modern computer vision to event cameras. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3857-3866 (2019)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 197, + 480, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 197, + 480, + 228 + ], + "spans": [ + { + "bbox": [ + 132, + 197, + 480, + 228 + ], + "type": "text", + "content": "36. Redmon, J., Divvala, S., Girshick, R., Farhadi, A.: You only look once: Unified, real-time object detection. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 779-788 (2016)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 230, + 480, + 261 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 230, + 480, + 261 + ], + "spans": [ + { + "bbox": [ + 132, + 230, + 480, + 261 + ], + "type": "text", + "content": "37. Reinhard, E., Stark, M., Shirley, P., Ferwerda, J.: Photographic tone reproduction for digital images. ACM Transactions on Graphics 21(3), 267-276 (2002). https://doi.org/10.1145/566654.566576" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 262, + 480, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 262, + 480, + 285 + ], + "spans": [ + { + "bbox": [ + 132, + 262, + 480, + 285 + ], + "type": "text", + "content": "38. Rudnev, V., Elgharib, M., Theobalt, C., Golyanik, V.: Eventnerf: Neural radiance fields from a single colour event camera. arXiv preprint arXiv:2206.11896 (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 285, + 480, + 307 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 285, + 480, + 307 + ], + "spans": [ + { + "bbox": [ + 132, + 285, + 480, + 307 + ], + "type": "text", + "content": "39. Ryu, H.E.: Industrial dvs design; key features and applications. In: Conf. on Computer Vision and Pattern Recognition. vol. 3 (2019)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 308, + 480, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 308, + 480, + 341 + ], + "spans": [ + { + "bbox": [ + 132, + 308, + 480, + 341 + ], + "type": "text", + "content": "40. Scheerlinck, C., Barnes, N., Mahony, R.: Continuous-time intensity estimation using event cameras. In: Asian Conference on Computer Vision. pp. 308-324. Springer (2018)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 342, + 480, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 342, + 480, + 374 + ], + "spans": [ + { + "bbox": [ + 132, + 342, + 480, + 374 + ], + "type": "text", + "content": "41. Scheerlinck, C., Rebecq, H., Gehrig, D., Barnes, N., Mahony, R., Scaramuzza, D.: Fast image reconstruction with an event camera. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 156-163 (2020)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 376, + 480, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 376, + 480, + 407 + ], + "spans": [ + { + "bbox": [ + 132, + 376, + 480, + 407 + ], + "type": "text", + "content": "42. Scheerlinck, C., Rebecq, H., Stoffregen, T., Barnes, N., Mahony, R., Scaramuzza, D.: Ced: Color event camera dataset. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops. pp. 0-0 (2019)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 409, + 480, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 409, + 480, + 441 + ], + "spans": [ + { + "bbox": [ + 132, + 409, + 480, + 441 + ], + "type": "text", + "content": "43. Schwarz, K., Liao, Y., Niemeyer, M., Geiger, A.: Graf: Generative radiance fields for 3d-aware image synthesis. In: Advances in Neural Information Processing Systems (NeurIPS) (2020)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 442, + 480, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 442, + 480, + 464 + ], + "spans": [ + { + "bbox": [ + 132, + 442, + 480, + 464 + ], + "type": "text", + "content": "44. Shiba, S., Aoki, Y., Gallego, G.: Event collapse in contrast maximization frameworks. Sensors 22(14), 5190 (2022)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 465, + 480, + 486 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 465, + 480, + 486 + ], + "spans": [ + { + "bbox": [ + 132, + 465, + 480, + 486 + ], + "type": "text", + "content": "45. Shiba, S., Aoki, Y., Gallego, G.: Secrets of event-based optical flow. In: European Conference on Computer Vision. pp. 628-645. Springer (2022)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 487, + 480, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 487, + 480, + 520 + ], + "spans": [ + { + "bbox": [ + 132, + 487, + 480, + 520 + ], + "type": "text", + "content": "46. Sitzmann, V., Martel, J., Bergman, A., Lindell, D., Wetzstein, G.: Implicit neural representations with periodic activation functions. Advances in Neural Information Processing Systems 33, 7462-7473 (2020)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 521, + 480, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 521, + 480, + 553 + ], + "spans": [ + { + "bbox": [ + 132, + 521, + 480, + 553 + ], + "type": "text", + "content": "47. Skorokhodov, I., Ignatyev, S., Elhoseiny, M.: Adversarial generation of continuous images. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10753-10764 (2021)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 555, + 480, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 555, + 480, + 586 + ], + "spans": [ + { + "bbox": [ + 132, + 555, + 480, + 586 + ], + "type": "text", + "content": "48. Song, C., Huang, Q., Bajaj, C.: E-cir: Event-enhanced continuous intensity recovery. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7803-7812 (2022)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 588, + 480, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 588, + 480, + 620 + ], + "spans": [ + { + "bbox": [ + 132, + 588, + 480, + 620 + ], + "type": "text", + "content": "49. Stoffregen, T., Scheerlinck, C., Scaramuzza, D., Drummond, T., Barnes, N., Kleeman, L., Mahony, R.: Reducing the sim-to-real gap for event cameras. In: European Conference on Computer Vision. pp. 534-549. Springer (2020)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 622, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 622, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 622, + 480, + 665 + ], + "type": "text", + "content": "50. Strümpler, Y., Postels, J., Yang, R., Gool, L.V., Tombari, F.: Implicit neural representations for image compression. In: Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXVI. pp. 74-91. Springer (2022)" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 417, + 91, + 447, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 417, + 91, + 447, + 100 + ], + "spans": [ + { + "bbox": [ + 417, + 91, + 447, + 100 + ], + "type": "text", + "content": "EvINR" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 482, + 588 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 132, + 116, + 482, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 482, + 160 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 482, + 160 + ], + "type": "text", + "content": "51. Wang, Y., Du, B., Shen, Y., Wu, K., Zhao, G., Sun, J., Wen, H.: Ev-gait: Event-based robust gait recognition using dynamic vision sensors. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6358-6367 (2019)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 161, + 482, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 161, + 482, + 194 + ], + "spans": [ + { + "bbox": [ + 132, + 161, + 482, + 194 + ], + "type": "text", + "content": "52. Wang, Z., Bovik, A.C., Sheikh, H.R., Simoncelli, E.P.: Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing 13(4), 600-612 (2004)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 194, + 482, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 194, + 482, + 226 + ], + "spans": [ + { + "bbox": [ + 132, + 194, + 482, + 226 + ], + "type": "text", + "content": "53. Weng, W., Zhang, Y., Xiong, Z.: Event-based video reconstruction using transformer. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 2563-2572 (2021)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 227, + 482, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 227, + 482, + 270 + ], + "spans": [ + { + "bbox": [ + 132, + 227, + 482, + 270 + ], + "type": "text", + "content": "54. Yadav, G., Maheshwari, S., Agarwal, A.: Contrast limited adaptive histogram equalization based enhancement for real time video system. In: 2014 international conference on advances in computing, communications and informatics (ICACCI). pp. 2392-2397. IEEE (2014)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 271, + 482, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 271, + 482, + 303 + ], + "spans": [ + { + "bbox": [ + 132, + 271, + 482, + 303 + ], + "type": "text", + "content": "55. Yang, M., Liu, S.C., Delbruck, T.: A dynamic vision sensor with " + }, + { + "bbox": [ + 132, + 271, + 482, + 303 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 132, + 271, + 482, + 303 + ], + "type": "text", + "content": " temporal contrast sensitivity and in-pixel asynchronous delta modulator for event encoding. IEEE Journal of Solid-State Circuits 50(9), 2149-2160 (2015)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 304, + 482, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 304, + 482, + 335 + ], + "spans": [ + { + "bbox": [ + 132, + 304, + 482, + 335 + ], + "type": "text", + "content": "56. Zhang, K., Li, Y., Zuo, W., Zhang, L., Van Gool, L., Timofte, R.: Plug-and-play image restoration with deep denoiser prior. IEEE Transactions on Pattern Analysis and Machine Intelligence 44(10), 6360-6376 (2021)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 336, + 482, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 336, + 482, + 369 + ], + "spans": [ + { + "bbox": [ + 132, + 336, + 482, + 369 + ], + "type": "text", + "content": "57. Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual metric. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 586-595 (2018)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 370, + 482, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 370, + 482, + 402 + ], + "spans": [ + { + "bbox": [ + 132, + 370, + 482, + 402 + ], + "type": "text", + "content": "58. Zhang, X., Srinivasan, P.P., Deng, B., Debevec, P., Freeman, W.T., Barron, J.T.: Nerfactor: Neural factorization of shape and reflectance under an unknown illumination. ACM Transactions on Graphics (TOG) 40(6), 1-18 (2021)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 403, + 482, + 424 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 403, + 482, + 424 + ], + "spans": [ + { + "bbox": [ + 132, + 403, + 482, + 424 + ], + "type": "text", + "content": "59. Zhang, Z., Yezzi, A., Gallego, G.: Image reconstruction from events. why learn it? arXiv preprint arXiv:2112.06242 (2021)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 425, + 482, + 457 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 425, + 482, + 457 + ], + "spans": [ + { + "bbox": [ + 132, + 425, + 482, + 457 + ], + "type": "text", + "content": "60. Zhang, Z., Yezzi, A.J., Gallego, G.: Formulating event-based image reconstruction as a linear inverse problem with deep regularization using optical flow. IEEE Transactions on Pattern Analysis & Machine Intelligence 45(07), 8372-8389 (2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 458, + 482, + 490 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 458, + 482, + 490 + ], + "spans": [ + { + "bbox": [ + 132, + 458, + 482, + 490 + ], + "type": "text", + "content": "61. Zhu, A.Z., Thakur, D., Ozaslan, T., Pfrommer, B., Kumar, V., Daniilidis, K.: The multivehicle stereo event camera dataset: An event camera dataset for 3d perception. IEEE Robotics and Automation Letters 3(3), 2032-2039 (2018)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 491, + 482, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 491, + 482, + 511 + ], + "spans": [ + { + "bbox": [ + 132, + 491, + 482, + 511 + ], + "type": "text", + "content": "62. Zhu, A.Z., Yuan, L., Chaney, K., Daniilidis, K.: Ev-flownet: Self-supervised optical flow estimation for event-based cameras. arXiv preprint arXiv:1802.06898 (2018)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 512, + 482, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 512, + 482, + 544 + ], + "spans": [ + { + "bbox": [ + 132, + 512, + 482, + 544 + ], + "type": "text", + "content": "63. Zhu, A.Z., Yuan, L., Chaney, K., Daniilidis, K.: Unsupervised event-based learning of optical flow, depth, and egomotion. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 989-997 (2019)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 545, + 482, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 545, + 482, + 588 + ], + "spans": [ + { + "bbox": [ + 132, + 545, + 482, + 588 + ], + "type": "text", + "content": "64. Zhu, L., Wang, X., Chang, Y., Li, J., Huang, T., Tian, Y.: Event-based video reconstruction via potential-assisted spiking neural network. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3594-3604 (2022)" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 225, + 102 + ], + "type": "text", + "content": "Z. Wang et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Revisit Human-Scene Interaction via Space Occupancy/f5977317-ef8e-4afa-8ef9-88b9ecbe0b04_content_list.json b/2024/Revisit Human-Scene Interaction via Space Occupancy/f5977317-ef8e-4afa-8ef9-88b9ecbe0b04_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..ec7e7040858bad908a657638d3613be2a25a50fa --- /dev/null +++ b/2024/Revisit Human-Scene Interaction via Space Occupancy/f5977317-ef8e-4afa-8ef9-88b9ecbe0b04_content_list.json @@ -0,0 +1,1880 @@ +[ + { + "type": "text", + "text": "Revisit Human-Scene Interaction via Space Occupancy", + "text_level": 1, + "bbox": [ + 254, + 140, + 750, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Xinpeng Liu1, Haowen Hou1,2, Yanchao Yang3, Yong-Lu Li1, and Cewu Lu1", + "bbox": [ + 228, + 210, + 774, + 242 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Shanghai Jiao Tong University", + "bbox": [ + 388, + 253, + 611, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ Soochow University", + "bbox": [ + 426, + 268, + 573, + 282 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3 The University of Hong Kong", + "bbox": [ + 395, + 282, + 607, + 297 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "xinpengliu0907@gmail.com, haowenhou@outlook.com, yanchaoy@hku.hk, {yonglu_li, lucewu}@sjtu.edu.cn", + "bbox": [ + 250, + 297, + 750, + 325 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract. Human-scene Interaction (HSI) generation is a challenging task and crucial for various downstream tasks. However, one of the major obstacles is its limited data scale. High-quality data with simultaneously captured human and 3D environments is hard to acquire, resulting in limited data diversity and complexity. In this work, we argue that interaction with a scene is essentially interacting with the space occupancy of the scene from an abstract physical perspective, leading us to a unified novel view of Human-Occupancy Interaction. By treating pure motion sequences as records of humans interacting with invisible scene occupancy, we can aggregate motion-only data into a large-scale paired human-occupancy interaction database: Motion Occupancy Base (MOB). Thus, the need for costly paired motion-scene datasets with high-quality scene scans can be substantially alleviated. With this new unified view of Human-Occupancy interaction, a single motion controller is proposed to reach the target state given the surrounding occupancy. Once trained on MOB with complex occupancy layout, which is stringent to human movements, the controller could handle cramped scenes and generalize well to general scenes with limited complexity like regular living rooms. With no GT 3D scenes for training, our method can generate realistic and stable HSI motions in diverse scenarios, including both static and dynamic scenes. The project is available at https://foruck.github.io/occu-page/.", + "bbox": [ + 259, + 363, + 743, + 670 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: Human-Scene Interaction $\\cdot$ Motion Generation", + "bbox": [ + 261, + 681, + 661, + 696 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 215, + 723, + 375, + 739 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Human-Scene Interaction (HSI) generation has been an active field recently with its application in animation [36] and embodied AI [35, 38]. For ideal HSI generation, models should be able to control the flexible articulated human body", + "bbox": [ + 212, + 756, + 787, + 801 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Equal contribution.", + "bbox": [ + 215, + 811, + 364, + 824 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "\\*\\* Corresponding author.", + "bbox": [ + 215, + 825, + 383, + 839 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/e5f11cd6ab03abe5894adc5853a9ba71a584ea6e0172351860ef03795388e774.jpg", + "image_caption": [ + "Fig. 1: We propose that interacting with scenes is essentially interacting with its space occupancy for static HSI. In this view, we can unify motion-only data into a unified human-occupancy knowledge base and train a versatile Human-Occupancy Interaction controller upon it, achieving stable generation under various scenarios." + ], + "image_footnote": [], + "bbox": [ + 217, + 143, + 785, + 296 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "to conduct various motions in highly complex 3D environments, considering geometry constraints, motion feasibility, and diversity. In this paper, we focus on interactions without changing the scene, which we refer to as static HSI.", + "bbox": [ + 212, + 398, + 784, + 443 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The limited data acquisition is a major obstacle for static HSI generation. Simultaneous capture of human and 3D scenes is expensive. Some [11,56] collected data only for a narrow range of target objects. Widely adopted PROX [12] contains only 60 clips with limited diversity and complexity. HUMANISE [45] synthesized data by fitting motion from large-scale motion base AMASS [27] into various scenes. CIRCLE [1] put the scenes in VR applications to reduce the cost. Most efforts take explicit scenes as necessary assets. Instead, we dig into the essence of static HSI and revisit it from a space occupancy view.", + "bbox": [ + 212, + 444, + 785, + 566 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Space occupancy is usually identified as an elementary constraint for HSI: humans and scenes cannot occupy the same space simultaneously, which prevents humans from penetrating scenes. We argue that occupancy is inherently the primary factor, especially for static HSI, which is essentially interaction with the scene occupancy. As in Fig. 1, sitting on a chair is practically sitting on a chair-shaped solid being. The chair's occupancy allows us to rest our weight upon it, offering options to sit or stand, i.e., affordances, while its height and shape naturally invite sitting. Other elements like category labels (e.g., identifying it as a \"chair\" semantically) influence our likelihood of choosing to sit. The scene occupancy, therefore, offers a range of affordances and delineates potential interaction types, establishing an initial probability distribution for interactions. This is primarily accessible to the model and should be learned well at first.", + "bbox": [ + 212, + 566, + 787, + 747 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In light of this, the requirements of realistic scenes for HSI learning could be loosened, inspiring us to revisit large-scale motion datasets. Instead of treating motion-only data as locomotion knowledge bases like existing efforts [58,60], we transform them into a comprehensive Human-Occupancy Interaction database, named Motion Occupancy Base (MOB). For any given motion sequence, we first identify the space occupied by the human at any time in the sequence. The", + "bbox": [ + 212, + 750, + 787, + 840 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "Liu et al.", + "bbox": [ + 271, + 114, + 334, + 126 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "remaining space is conceptualized as the \"scene\". In this framework, the motion becomes a record of interaction with these defined space occupancies. As a result, MOB aggregates enormous paired human-occupancy data from existing large-scale motion-only data. Compared to current datasets [1, 12, 45], MOB presents over-complex and stringent geometric constraints, pushing the model to the limit in interacting with space occupancy.", + "bbox": [ + 212, + 146, + 784, + 236 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "How do we exploit the aggregated MOB for static HSI? Previous approaches generally provided two different answers. Some [1,16,45,61] adopt an end-to-end pipeline from scene to motion, which is straightforward but inflexible without the ability to interact with dynamic scenes. Some [43, 44, 60] typically decouple HSI into interaction and locomotion, adopting a two-stage paradigm: first generating key interaction frames and then synthesizing locomotion between them, with the help of methods like path-finding algorithms. However, this decomposition overlooks the integrated nature of locomotion and interaction, like simultaneous arm movement and walking to navigate obstacles. Our approach, rooted in space occupancy, reconceptualizes locomotion and interaction as a unified Human-Occupancy Interaction. We develop an auto-regressive HumanOccupancy Interaction controller for motion generation. The controller processes various inputs, including surrounding occupancy and target locations, as control signals at each time step, producing future motion predictions. Additionally, we introduce a field regulation module for enhanced collision avoidance. There are three major advantages of our controller. First, by learning from the challenging cases in MOB, it could handle complex scenes but also generalize well to regular simpler scenarios. Second, its versatility accommodates any static HSI without the need to switch between models for locomotion and interactions, even for HOI motion synthesis. Third, its auto-regressive nature allows it to adapt to dynamic environments such as an automatic door, a capability lacked in previous methods. In extensive experiments, our controller performs effectively across various settings, even without training on real 3D scene data.", + "bbox": [ + 212, + 237, + 785, + 583 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Overall, our contributions are three-fold: 1) We exhume Human-Occupancy Interaction as a major component of static HSI. 2) By digging out the inherent Human-Occupancy relationships of pure motions, we convert motion-only data into a paired Human-Occupancy database MOB, expanding the data scale. 3) We develop a versatile controller for flexible and stable HSI motion generation.", + "bbox": [ + 212, + 585, + 785, + 660 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Related Works", + "text_level": 1, + "bbox": [ + 214, + 685, + 395, + 700 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Human-Scene Interaction Generation. Generating natural human motion sequences has received increasing attention with the emerging MoCap datasets [4, 7, 18, 24, 27, 33]. Early efforts were paid to prefix/suffix-conditioned motion generation [2, 9, 10, 15], known as motion prediction and interpolation. There also have been efforts to pursue generating motion with respect to specific action semantics, either represented in action categories [8, 30, 47] or natural language [6, 7, 26, 31, 32, 41, 42, 52–54]. Audio-conditioned motion generation was also explored, like music for dance generation [21, 22]. However, these approaches tend", + "bbox": [ + 212, + 719, + 785, + 840 + ], + "page_idx": 2 + }, + { + "type": "header", + "text": "Revisit Human-Scene Interaction via Space Occupancy", + "bbox": [ + 364, + 114, + 730, + 128 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "to focus on human motion only, without placing the generated motion in a grounded 3D environment. Recently, progress has been made to advance HSI generation. Early efforts [23,51] first reconstructed human-object interaction in 3D. Some [13,55,57] proceeded by placing SMPL-X [29] meshes in scanned 3D scenes. Semantic controls were also introduced [59]. For HSI motion generation, some [11,36,56] focused on interacting with a single target object, mostly chairs. GOAL [39] and SAGA [46] generated full-body pre-grasp motions. There have also been efforts [14,40] for the physical control of specific interactions. With paired motion-scene datasets [1,12], approaches that synthesize HSI motion in cluttered scenes appeared. Two-stage based algorithms [43,44] were proposed to first generate static key-frame poses, and then interpolate them for a coherent motion. GAMMA [58] proposed an RL framework for locomotion on flat grounds with the help of off-the-shelf path-finding algorithms. DIMOS [60] extended GAMMA with extra interaction policies. In contrast, the follow-ups [1,16,45,61] adopted a simple end-to-end one-stage paradigm, while focusing more on boosting data. GIMO [61] collected data with gaze information. HUMANISE [45] synthesized pseudo-data from large-scale motion bases and 3D scan bases. CIRCLE [1] collected data with a VR application. LAMA [19] coupled the RL framework with a motion matching algorithm, generated HSI motions with only a unified policy, and leveraged motion editing for optimization. Our major difference from previous efforts is the Human-Occupancy Interaction view, which alleviates the data hunger, and also enables a unified controller. Moreover, we are the first to generate motion in a dynamic scene, which previous efforts failed to handle.", + "bbox": [ + 212, + 146, + 787, + 492 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Scene Creation from Motion. Pose2Room [28] predicted 3D object bounding boxes from human motion. MOVER [50] reconstructed 3D scenes from videos that record human-scene interaction. SUMMON [48] and MIME [49] placed 3D object models concerning human motion. These methods could be adopted to boost HSI data, though usually limited to a small range of indoor furniture.", + "bbox": [ + 212, + 493, + 787, + 570 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3 Motion Occupancy Base", + "text_level": 1, + "bbox": [ + 215, + 594, + 488, + 612 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "HSI generation could be formulated as $M = \\mathcal{G}(S)$ , where $M$ is the generated motion, $\\mathcal{G}$ is the motion generator, and $S$ is a 3D scene. Conventionally, the scene is expected to be as realistic as possible, requiring expensive acquisition. We find that these scenes could function almost equally to their occupancy for static HSI. Thus, high-quality scenes might become less necessary here. Given this, we revisit the large-scale motion-only datasets. Intuitively, motion only contains body movement information. But if we switch our sight to the vacant space around the moving person, considering the space as another form of occupancy, we could find that these motions possess information on how humans interact with certain space occupancies. To this end, we aggregate motion-only datasets into the human-occupancy interaction database: Motion Occupancy Base (MOB).", + "bbox": [ + 212, + 627, + 787, + 792 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The overall pipeline is shown in Fig. 2. Motion with SMPL(-X) [25,29] representation are collected, re-sampled to 30 FPS, and rotated so that the gravity direction points to the negative z-axis. For motion sequence $X = \\{x^i\\}_{i=1}^t$ , where", + "bbox": [ + 212, + 795, + 787, + 842 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "Liu et al.", + "bbox": [ + 271, + 114, + 334, + 126 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/c49f40fd5050f99401eb9c13f41457627a671cb695e7cc0f89bc1cece52ed7a0.jpg", + "image_caption": [ + "Fig. 3: The architecture of our versatile motion controller. Given motion state and histories $x^{t}, h^{t}$ , and control signals $c_{o}^{t}, c_{t}^{t}$ , the controller auto-regressively generates the next motion frame w.r.t. the target pose and the canonical occupancy." + ], + "image_footnote": [], + "bbox": [ + 245, + 143, + 759, + 280 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "$x_{i}$ is the SMPL(-X) representation in the $i$ -th frame, $t$ is the number of frames, we first obtain mesh sequence $X_{m} = \\{x_{m}^{i}\\}_{i = 1}^{t}$ via SMPL(-X) model, where $x_{m}^{i} = (f^{i},v^{i})$ is the mesh for the $i$ -th frame, $f^{i},v^{i}$ are the mesh faces and surface", + "bbox": [ + 212, + 358, + 787, + 405 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "vertices. A bounding box of $X_{m}$ is obtained based on vertices $V = \\{v_{i}\\}_{i=1}^{t}$ . Given the predefined grid unit size $u$ , the box region is voxelized into motion occupancy volume $\\hat{O} \\in \\{0,1\\}^{[H \\times W \\times D]}$ , computed in two steps: 1) $V$ occupy the corresponding voxels in $\\hat{O}$ . 2) Given the centers of the un-occupied voxels as $C = \\{c_{i}\\}_{i=1}^{|C|}$ , if $c_{i}$ is inside any human mesh $x_{m}^{i}$ , the corresponding voxel is occupied. The pseudo-scene occupancy $O$ is then obtained by negating the motion occupancy, formulated as $O = 1 - \\hat{O}$ . Thus, we convert motion-only sample $X$ into a motion-occupancy pair $(X,O)$ in MOB. MOB", + "bbox": [ + 212, + 404, + 488, + 662 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/9f04fe83a768594cd8978f1701a9470940a2412e939bee049ba6b5b61721b096.jpg", + "image_caption": [ + "Fig. 2: The construction process of Motion Occupancy Base. Note that the ceilings of the occupancy are hidden for clarity." + ], + "image_footnote": [], + "bbox": [ + 500, + 434, + 787, + 579 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "aggregates 13 datasets with 98k instances. More details are in the supplementary.", + "bbox": [ + 212, + 662, + 785, + 676 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As shown in Fig. 2, humans in MOB are acting with rather strict geometric constraints, which is uncommon in regular scenes covered by existing datasets [1, 12, 45]. However, we show that by learning from hard cases in MOB, our model could grow its ability to handle both complex and regular scenarios well.", + "bbox": [ + 212, + 676, + 787, + 738 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4 Versatile Motion Controller", + "text_level": 1, + "bbox": [ + 214, + 758, + 519, + 773 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1 Preliminaries", + "text_level": 1, + "bbox": [ + 214, + 787, + 372, + 801 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To instantiate $M = \\mathcal{G}(S)$ for HSI generation, previous efforts either adopt a one-shot transformer generator or decompose $\\mathcal{G}$ into two separate generators, one for", + "bbox": [ + 212, + 809, + 785, + 840 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "Revisit Human-Scene Interaction via Space Occupancy", + "bbox": [ + 364, + 114, + 732, + 130 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "interaction key-frame generation, the other for key-frame interpolation. Instead, we formulate our generator as an auto-regressive controller. At time step $t$ , given historical motion state $h^t$ , current pose state $x^t$ , we expect a controller as", + "bbox": [ + 212, + 146, + 782, + 191 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {y} ^ {t} = \\mathbf {G} \\left(h ^ {t}, x ^ {t}, c ^ {t}\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 437, + 202, + 785, + 219 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\hat{y}^t$ is the predicted future motion, and $c^t$ is the control signals. An overview is in Fig. 3. Sec. 4.2 introduces motion state representation $h^t, x^t, y^t$ . The control signals are covered in Sec. 4.3. The controller architecture is detailed in Sec. 4.4. Sec. 4.5 demonstrates the training and inference procedure.", + "bbox": [ + 212, + 229, + 782, + 290 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2 Motion State Representation", + "text_level": 1, + "bbox": [ + 215, + 311, + 501, + 325 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Notations. In the global coordinate, we represent motion sequence $X$ with $T$ frames as $X = \\{x^{t}\\}_{t = 1}^{T}$ , where $x^{t} = (p_{r}^{t},\\theta_{r}^{t},r^{t})$ is SMPL [25] parameters, with root location $p_r^t\\in \\mathcal{R}^3$ , root orientation $\\theta_r^t\\in \\mathcal{R}^6$ , and the orientation of $j$ limbs as $\\theta^t\\in \\mathcal{R}^{j\\times 6}$ . Based on this, the location of $j$ joints $p^t\\in \\mathcal{R}^{j\\times 3}$ could also be computed via SMPL. For simplicity, we use $\\cdot^{-w}$ to represent the value from frame $t - w$ to $t$ , and $\\cdot^{+w}$ for $t$ to $t + w$ . The 6D form from [62] is adopted for all rotation and orientation representations.", + "bbox": [ + 212, + 335, + 784, + 440 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Current Pose State could be formulated as", + "bbox": [ + 238, + 441, + 571, + 455 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nx ^ {t} = \\operatorname {C a n o} \\left(\\left\\{p _ {r} ^ {t}, p ^ {t}, \\theta_ {r} ^ {t}, \\theta^ {t}, \\dot {p} _ {r} ^ {t}, \\dot {p} ^ {t}, \\dot {\\theta} _ {r} ^ {t} \\right\\}, t\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 367, + 467, + 784, + 484 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\cdot^t$ indicates the first derivative at $t$ . Cano $(\\cdot, t)$ canonicalizes the input w.r.t. the root location and facing direction at $t$ , ensuring consistency across frames.", + "bbox": [ + 212, + 494, + 782, + 523 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Historical Motion State $h^t$ captures motion from frame $t - w$ to $t$ as", + "bbox": [ + 238, + 525, + 759, + 539 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nh ^ {t} = \\operatorname {C a n o} \\left(\\left\\{p _ {r} ^ {- w}, p ^ {- w}, \\theta_ {r} ^ {- w}, \\dot {p} _ {r} ^ {- w}, \\dot {p} ^ {- w}, \\dot {\\theta} _ {r} ^ {- w} \\right\\}, t\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 334, + 550, + 784, + 566 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "with trajectories $p_r^{-w}, p^{-w}$ , root orientation $\\theta_r^{-w}$ , velocities $\\dot{p}_r^{-w}, \\dot{p}^{-w}$ , and root angular velocity $\\dot{\\theta}_r^{-w}$ . For conciseness, if not specified, only the $p^{-w}$ of the hands and feet is adopted in practice.", + "bbox": [ + 212, + 578, + 784, + 622 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Future Prediction $\\hat{y}^t$ is composed of two parts as $\\hat{y}^t = \\{\\hat{x}^{t + 1},\\hat{u}^t\\}$ . Note that $\\hat{x}^{t + 1}$ is in the egocentric canonical of frame $t$ , and $\\hat{p}_r,\\hat{\\dot{\\theta}}\\subset \\hat{x}^{t + 1}$ are adopted to transit from $t$ to $t + 1$ . Mirroring the historical motion state $h^t$ , we also incorporate a comprehensive set of future trajectories, facing directions and their changing rates as $\\hat{u}^t = \\{p_r^{+f},p^{+f},\\theta^{+f},\\dot{p}_r^{+f},\\dot{p}^{+f},\\dot{\\theta}^{+f}\\}$ for a window size of $f$ , all in the canonical view of $t + 1$ . Empirically, integrating $\\hat{u}^t$ significantly enhances the model's precision in navigating toward the target.", + "bbox": [ + 212, + 623, + 784, + 733 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3 Control Signals", + "text_level": 1, + "bbox": [ + 215, + 753, + 388, + 768 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this section, we introduce how the different control signals are encoded. Note that our models can operate without these optional signals.", + "bbox": [ + 212, + 779, + 782, + 808 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Canonical Occupancy. A simple voxel-based representation is adopted to encode how the canonical space is occupied by the scenes. We utilize a cubic", + "bbox": [ + 212, + 810, + 784, + 839 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 5 + }, + { + "type": "header", + "text": "Liu et al.", + "bbox": [ + 271, + 114, + 334, + 127 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "point grid $o^t \\in \\{0,1\\}^{s \\times s \\times s}$ characterized by the grid size $s$ and the grid unit size $u$ at frame $t$ . The grid is positioned in the human's canonical space, with its center offset by $s/4$ in the human-facing direction. Each cell contains a binary value, indicating whether the cell is intersecting with the global scene occupancy $O$ . The canonical occupancy is thus represented as $c_o^t = \\text{Flatten}(o^t) \\in \\{0,1\\}^{s^3}$ . Note that because the occupancy is calculated in real-time, this control signal can naturally adapt to dynamic scenes where each timestep is treated statically. We will show in Section 5 that even though our training data is all static, our model exhibits a smooth generalization to dynamic scenes.", + "bbox": [ + 212, + 146, + 787, + 282 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Target Pose is also considered a flexible control signal, which can be given in two ways. The first is the positions of the root, hands, and feet, each defined in relation to the canonical space of the current pose, represented as $c_t^t \\in \\mathcal{R}^{5 \\times 3}$ . In practice, these could be assigned manually or generated from off-the-shelf models [59]. Also, partial or blank targets are acceptable. The second is the future trajectory of end-effectors, denoted as $c_t^{+f}$ for a future window size of $f$ . This paves the way for different applications in Section 5.", + "bbox": [ + 212, + 282, + 789, + 391 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4 Controller Architecture", + "text_level": 1, + "bbox": [ + 215, + 417, + 457, + 430 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We adopt a lightweight transformer-based architecture. $n$ fully connected layers first project the $n$ input features from $\\{h^t,x^t,c^t\\}$ to the latent $z\\in \\mathcal{R}^{n\\times d}$ . $z$ is then fed into encoder $E$ to distill dynamic information from the motion states and control signals. We devise $E$ as a sequence of Transformer encoders, which takes $z$ as input and outputs refined dynamic information $\\hat{z} = E(z)\\in \\mathcal{R}^{n\\times d}$ . Finally, we adopt a transformer decoder $D$ followed by a linear layer to condense the dynamic information, resulting in the final predictions. With learnable query $q\\in \\mathcal{R}^d$ and the dynamic information $\\hat{z}$ serving as K and V, respectively, the output of the transformer is further concatenated with a random noise $\\epsilon \\in \\mathcal{R}^d$ for randomness. The linear layer then outputs the final prediction as $\\hat{y}^t = D(\\hat{z},q,\\epsilon)$ .", + "bbox": [ + 212, + 446, + 599, + 672 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Furthermore, to enhance collision avoidance, we introduce a novel module called Field Regulation. The basic idea is that humans tend to move faster in obstacle-sparse areas and decelerate in complex environments or when an obstacle is nearby. At very short distances, especially with an imminent collision happening, a complete halt is necessary. Given this and taking inspiration from the potential field in robot motion planning [17], we propose Occupancy Field, which exerts negligible influence on humans at a distance but", + "bbox": [ + 212, + 674, + 602, + 823 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "escalates to a sturdy repulsion as proximity increases, thereby compelling the", + "bbox": [ + 212, + 824, + 785, + 840 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/30e2668dbbe56429deaffa35f1c05c951373bdb753827e069578e8503f4ffb3c.jpg", + "image_caption": [ + "Fig.4: Occupied voxel centers $\\{P_i\\}$ are located at the top left of the figure. The initial velocity vector, $\\dot{p}_j$ , is directed upwards, posing a risk of collision. To mitigate this, the occupancy field introduces a corrective velocity component $\\Delta \\dot{p}_j$ . This redirects the final velocity $\\dot{p}_j^{out}$ to avoid collision." + ], + "image_footnote": [], + "bbox": [ + 612, + 477, + 785, + 609 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "Revisit Human-Scene Interaction via Space Occupancy", + "bbox": [ + 364, + 114, + 732, + 130 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7", + "bbox": [ + 774, + 114, + 785, + 126 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "human to decelerate and avoid collision. The occupancy field influences human motion only when a human is in motion toward it. The field functions as", + "bbox": [ + 215, + 146, + 784, + 176 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\Delta \\dot {p} _ {j} ^ {t + 1} = \\mathcal {F} \\left(p _ {j} ^ {t + 1}, c _ {o} ^ {t}\\right). \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 426, + 184, + 785, + 204 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The field $\\mathcal{F}$ calculates a correction $\\Delta \\dot{p}^{t + 1}$ w.r.t. predicted joint velocity $\\dot{p}_j^{t + 1} \\in \\{\\dot{p}_r^{t + 1}, \\dot{p}^{t + 1}\\}$ and canonical occupancy $c_o^t$ , which is then added to $\\dot{p}^{t + 1}$ . For simplicity, the superscripts indicating the time steps are omitted in the following equations. In detail, given the center points of all the occupied voxels in $c_o$ as $P \\subset \\mathcal{R}^{n_p \\times 3}$ , the correction is calculated as", + "bbox": [ + 215, + 210, + 782, + 287 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\Delta \\dot {p} _ {j} = \\sum_ {P _ {i} \\in P} - k r _ {i} \\max \\left(0, \\frac {1}{\\| v _ {i j} \\| _ {a} - \\gamma} - b\\right), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 344, + 297, + 785, + 333 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\nr _ {i} = \\dot {p} _ {j} \\max (0, \\langle \\dot {p} _ {j}, v _ {i j} \\rangle), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 357, + 338, + 784, + 354 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $v_{ij} = P_i - p_j$ is the vector from the joint $j$ to the voxel center, $\\langle \\cdot, \\cdot \\rangle$ indicates the cosine between the two vectors, $\\| \\|_a$ denotes $a$ -norm, and $k$ is the stiffness factor. $\\gamma$ and $b$ are thresholds to constrain the distance range of effective voxels, where $\\gamma$ reflects the human body's width, and voxels within this range are discounted for stability. An illustration of Field Regulation is shown in Fig. 4. The field regulation is incorporated in the training of the controller to adapt to it, avoiding potential foot sliding caused due to direct modification on velocities.", + "bbox": [ + 215, + 362, + 785, + 467 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.5 Training and Inference", + "text_level": 1, + "bbox": [ + 215, + 487, + 450, + 502 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Training Objectives The controller is trained in a fully-supervised manner. We employ a mixture of L1 and L2 losses as basic objectives. Besides, penetration loss and occupancy field loss are designed to hinder undesired collisions. Specifically, L1 loss is utilized for rotation-related output values $\\hat{y}_{\\theta}^{t}\\in \\hat{y}^{t}$ , while L2 loss is applied to the rest outputs $\\hat{y}_p^t$ . The mixture loss is formulated as", + "bbox": [ + 215, + 508, + 782, + 585 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {m i x} = \\left\\| y _ {\\theta} ^ {t} - \\hat {y} _ {\\theta} ^ {t} \\right\\| _ {1} + \\left\\| y _ {p} ^ {t} - \\hat {y} _ {p} ^ {t} \\right\\| _ {2}. \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 387, + 595, + 784, + 613 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Given the predicted joint location $\\hat{p}^{t + 1}$ , penetration loss is computed as", + "bbox": [ + 215, + 619, + 732, + 635 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {p e n} = \\sum_ {i = 0} ^ {j} O \\left(\\hat {p} _ {j} ^ {t + 1}\\right) \\| r e f \\left(\\hat {p} _ {j} ^ {t + 1}, O\\right) \\| _ {2}, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 375, + 643, + 784, + 683 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $O(\\cdot)$ returns 1 if the point is occupied in $O$ else 0, and $ref(\\hat{p}_j^{t + 1},O)$ calculates the closest unoccupied voxel center in $O$ for joint $\\hat{p}_j^{t + 1}$ . Occupancy field loss has two components. First, to deter undesired interactions, we penalize the proportion of the occupancy field effect relative to the output root velocity. Second, observing occasional unnatural spikes in velocity, where the human figure seemingly attempts to navigate through an obstacle, we punish excessive velocities. The occupancy field loss is formulated as", + "bbox": [ + 215, + 691, + 785, + 800 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {f i e l d}} = \\left(\\frac {| \\Delta \\dot {p} |}{| \\dot {p} |}\\right) ^ {2} + | \\dot {p} | ^ {2}. \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 406, + 808, + 784, + 844 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8", + "bbox": [ + 217, + 114, + 228, + 126 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "Liu et al.", + "bbox": [ + 271, + 114, + 334, + 127 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/7708eee9e0cc8944ecac7348858a1abce4a4dda3f12efb60052da42dea5a2d81.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodSuc. (%) ↑DT (cm) ↓Time (s)FS (%)PEN ↓ERP
GT100.000.003.003.360.00-
NSM [36]11.7632.795.3311.04134.497.26
SAMP [11]13.9452.496.549.87145.746.98
DIMOS [60]15.7348.976.8210.52123.126.82
Ours w/o Occupancy71.2214.662.256.8946.054.13
Ours77.9115.303.628.1315.214.24
Ours w/o Lpen74.1414.893.547.8931.934.96
Ours w/o Lfield75.3415.692.8611.6525.835.70
Ours w/o FR77.3814.852.346.5733.494.18
Ours w/ BPS73.6416.833.359.3436.724.72
", + "bbox": [ + 305, + 143, + 702, + 255 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Table 1: Quantitative results on MOB.", + "bbox": [ + 367, + 255, + 633, + 268 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "The final loss is computed with loss coefficients $\\alpha, \\beta$ as", + "bbox": [ + 215, + 310, + 609, + 325 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathcal {L} _ {m i x} + \\alpha \\mathcal {L} _ {p e n} + \\beta \\mathcal {L} _ {\\text {f i e l d}}. \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 398, + 338, + 784, + 352 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Auto-regressive Inference. We perform motion generation in an auto-regressive manner, enabling the model to adapt to dynamic environmental changes. Since the model solely manipulates canonical information, we track the global root position $p_{g}$ and the facing direction $\\theta_{g}$ of the human in the world coordinate system. Initially, we set past and current velocities $\\dot{p}^{0}$ to zero. Before each prediction at frame $t$ , we compute the control signals (if there are any), i.e., the relative target position $c_{t}^{t}$ and the canonical space occupancy $c_{o}^{t}$ , based on the current values of $p_{g}^{t}$ and $\\theta_{g}^{t}$ . After the prediction, we update $p_{g}^{t}$ and $\\theta_{g}^{t}$ with the model's output, along with the current pose state $c^{t + 1}$ . Concurrently, we refresh the historical motion state $h^{t + 1}$ to prepare for the next input.", + "bbox": [ + 212, + 363, + 787, + 513 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "5 Experiments", + "text_level": 1, + "bbox": [ + 215, + 537, + 375, + 555 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Implementation Details. The controller consists of a two-layer transformer encoder and a single-layer transformer decoder, both with a latent dimension of 512 and 8 heads. The canonical occupancy grid size is set as (25, 25, 25), with a unit size of $8\\mathrm{cm}$ . The controller operates at 10 FPS with the history and future window sizes both set as 1, which empirically brings more stable outcomes. The controller is trained on the MOB, with 1,393 sequences excluded from training as the test set. We train the controller for 75k iterations, with a cosine learning rate decay restart strategy. The initial learning rate is set as 1e-4. The control signals are randomly masked for augmentation during training. Scheduled Sampling [3] is adopted. The whole training takes approximately 6 hours on a single 12G NVIDIA Titan Xp. More details are in the supplementary.", + "bbox": [ + 212, + 566, + 787, + 734 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "5.1 Interacting with Occupancy from MOB", + "text_level": 1, + "bbox": [ + 215, + 755, + 586, + 771 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We first evaluate the ability to interact with occupancy volumes with a goal-reaching task on MOB. Given an occupancy volume, a starting pose, and a target end-effector pose, the model is required to generate transition motion from the starting pose to the target pose. We split the test set into 30k 3-second", + "bbox": [ + 212, + 779, + 785, + 840 + ], + "page_idx": 8 + }, + { + "type": "header", + "text": "Revisit Human-Scene Interaction via Space Occupancy", + "bbox": [ + 364, + 114, + 730, + 128 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9", + "bbox": [ + 774, + 116, + 784, + 126 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/e9616a1ae9e7663d319be7e5a38fcc42cffcfdf67a603753a4251b8e942e315f.jpg", + "image_caption": [ + "Fig. 5: Our controller can naturally reach the target (white to blue) given the complex occupancy. Furthermore, it can stabilize the motions (red to white) around the targets after reaching them. Previous SOTA [60] suffers from severe penetration and fails to reach the target. For simplicity, the ceilings of the occupancy are hidden." + ], + "image_footnote": [], + "bbox": [ + 232, + 143, + 769, + 392 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "clips, adopting the first frame as the starting pose and the final frame as the target pose. The models are required to generate motion for 10s towards the target. There are two reasons for this setting. First, we would like to provide enough time for the models to accomplish the tasks. Second, we are also curious about how the models would behave after the target is reached and no new target is given, or when a very close target is given. The results are evaluated from three aspects. Task Completion is determined by whether the target pose is reached. The target is identified as completed if the average Euclidean distance between the generated and target pose is lower than 20 centimeters and less than 50 voxels are penetrated. We report the success rate (Suc.). The average minimum distance to the target pose (DT) and the time taken to reach the target (Time) are reported for reference. Note that Time is only calculated for successful samples. Motion Quality is evaluated via foot sliding (FS) and penetration (PEN) detection. A frame is identified as FS with its foot velocity $>7.5\\mathrm{cm / s}$ following GAMMA [58]. We report the percentage of frames with FS. For PEN, we report the number of voxels penetrated per frame. Similarity to Dataset is measured by the similarity between the predicted and GT target end-effector trajectory. We report Edit distance for the real penalty (ERP) [5] since the trajectories might have different lengths. It is only given as a reference, which is not the lower the better.", + "bbox": [ + 212, + 491, + 787, + 792 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Quantitative results are shown in Tab. 1. For DIMOS [60], we convert the target pose into corresponding marker sets and report the best results from different policies for each test sample. For w/o Occ., we remove the canonical", + "bbox": [ + 212, + 794, + 787, + 840 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 9 + }, + { + "type": "header", + "text": "Liu et al.", + "bbox": [ + 271, + 114, + 334, + 126 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/87245a5bd86dad665600291e9b5485cc67104ca506032de232652b4493d7de8b.jpg", + "image_caption": [ + "Space Occupancy overlay DIMOS Walkable Area from NavMesh" + ], + "image_footnote": [], + "bbox": [ + 223, + 143, + 357, + 234 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/d334ea9aa2a44a4d217eb82acbab90ea5f056d02f16ea15cb93a3c936e47bd0a.jpg", + "image_caption": [ + "DIMOS path planning results", + "Sit Down" + ], + "image_footnote": [], + "bbox": [ + 369, + 143, + 500, + 234 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/5e84eab1357ea931c3eeaa4888955163dffd3b6d6a527245a9c42e09ed722657.jpg", + "image_caption": [ + "Fig. 6: A typical DIMOS [60] failure case. Fig. 7: With the unified view of human-The NavMesh algorithm produces sub- occupancy interaction, the same controller optimal walkable areas, resulting in path- in Fig. 5 produces diverse motions, includ-finding failures for $69.43\\%$ MOB samples. ing sitting down, crawling, and getting up." + ], + "image_footnote": [], + "bbox": [ + 553, + 143, + 785, + 252 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "occupancy-related components. Previous efforts frequently fail to reach the goal successfully in complex scenarios, which we will elucidate later with qualitative results. Even when the goal is reached, severe penetration emerges, resulting in low success rates in Tab. 1. Our models outperform DIMOS by a considerable margin with a single controller compared to multiple policies. This reveals that the straightforward voxel-based representation is capable of encoding holistic information for human-occupancy interaction. The model without occupancy manages to approach the target as closely and quickly as possible. However, it suffers from more severe interpenetration with the occupancy, resulting in a lower success rate. Instead, our model with consideration of occupancy behaves more conservatively, spending more time to accomplish the tasks with less penetration. A slightly higher foot sliding rate is also observed, which is a trade-off for the penetration-avoiding ability that field regulation brings.", + "bbox": [ + 212, + 354, + 787, + 551 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Qualitative results are shown in Fig. 5, 7. Given the complex occupancy, previous state-of-the-art DIMOS [60] frequently fails to reach the goal. Even when it reaches the goal, severe penetration is usually observed. Instead, our controller generates feasible motions to reach the targets. Previous methods' typical locomotion-interaction decomposition [44, 58, 60] tends to over-simplify both. For locomotion, most methods assume a rigid character (usually simplified as a rigid cylinder in $\\mathrm{A}^*$ [11, 44] and NavMesh algorithms [60]), overseeing the flexibility of humans. It might be okay for spacious rooms, however, when posed in a cramped scene or scene with complex geometric structures like in Fig. 5, models could fail in path-finding and thus degenerate. As shown in Fig. 6, for $69.43\\%$ MOB samples, the NavMesh algorithm adopted by DIMOS failed to generate a feasible path from the start position to the target position. The statistic for $\\mathrm{A}^*$ from SAMP [11] is $70.29\\%$ . This explains the unsatisfying quantitative results of DIMOS. For interaction, the categories are typically limited to predefined ones. In contrast, with our unified view of Human-Occupancy interaction, our controller applies to various complex scenarios beyond simplified locomotion and interaction. Furthermore, diverse motions could be performed by the same controller (Fig. 7). By comparing our controllers with and without occupancy, we find that training with occupancy helps to avoid penetration significantly.", + "bbox": [ + 212, + 553, + 787, + 840 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "Revisit Human-Scene Interaction via Space Occupancy", + "bbox": [ + 364, + 114, + 730, + 128 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11", + "bbox": [ + 767, + 116, + 782, + 126 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/bcaacbbd69fa041030a9d5a6713e2485d6997e79e69fa17b2e2181fe2acff799.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelSuc. (%)Dist. to goal (cm)Collision depth (cm)FS (%)
GT100.000.008.742.45
Ours (w/ MOB only)73.017.199.5613.83
Ours (w/ CIRCLE only)75.526.8813.7411.76
Ours (w/ MOB + CIRCLE)78.976.7311.8512.87
", + "bbox": [ + 277, + 143, + 730, + 196 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Table 2: Results on CIRCLE [1].", + "bbox": [ + 387, + 198, + 612, + 210 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/ad6c5920472bc89dea76c37b86bcb6c6374941503bec0779fe3945dca69b5970.jpg", + "image_caption": [ + "Fig. 8: We can generate long-term motion. Fig. 9: We synthesize Human-Object Interactions in realistic rooms: A human circulation action, with hand trajectory as the control navigates a cabinet, sits on a sofa, perches signal. The human successfully moves a floor on a bed, and then resumes walking. lamp to a new place." + ], + "image_footnote": [], + "bbox": [ + 256, + 237, + 454, + 358 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/fc65b660786abf47cdc3ce3fdb42a77038104bc6c1a9346444a5bf7a395ba770.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 537, + 237, + 741, + 362 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Also, the controller with occupancy could stabilize the motion around the target after reaching it, while its counterpart tends to wander further inertially.", + "bbox": [ + 212, + 459, + 784, + 489 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "5.2 Interacting with Static Scenes", + "text_level": 1, + "bbox": [ + 214, + 512, + 509, + 527 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Once our controller is trained on the more complex MOB, it could be effortlessly applied to realistic synthesized rooms and scene scans.", + "bbox": [ + 212, + 537, + 784, + 566 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Quantitative results is reported on CIRCLE [1]. Given the initial full-body pose and the target right wrist position, the model is required to generate a right-hand reaching motion. Following [1], success rate, distance to the goal, cumulative collision depth, and foot sliding are reported as metrics. Besides the controller trained on MOB (excluding CIRCLE), we also train our controller on CIRCLE-only and MOB+CIRCLE. We randomly split CIRCLE into a 2,565-sample train set (for the ablative baseline only) and a 453-sample test set since the original split has not been made public yet. The results in Tab. 2 reveal that our controller could handle various tasks in complex natural scenes well. Also, its potential to advance HOI-related motion, like hand-reaching before grasping, is showcased. It is noticeable that the MOB-only model demonstrates a competitive success rate and better collision avoidance, revealing the efficacy of MOB with its advanced complexity. When incorporating MOB with CIRCLE, superior task completion is observed, showing that MOB could function as an effective auxiliary resource for the conventional HSI datasets.", + "bbox": [ + 212, + 568, + 785, + 792 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Qualitative results. We test our model in a synthetic room from DIMOS [60] in Fig. 8 and a realistic scan from Replica [37] in Fig. 1. The occupancy is transferred from the scene SDFs. Our model produces convincing motions, nav-", + "bbox": [ + 212, + 795, + 785, + 839 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 11 + }, + { + "type": "header", + "text": "Liu et al.", + "bbox": [ + 271, + 114, + 334, + 126 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/095890d1c8cf55261fd0b148674fa8869e946b23356fefdb98c2c43f721a70c0.jpg", + "image_caption": [ + "Fig. 10: Our model can avoid dynamic obstacles and reasonably reach the target circle." + ], + "image_footnote": [], + "bbox": [ + 253, + 143, + 750, + 220 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "igates around obstacles, and transitions seamlessly between locomotion and interaction. Remarkably, it adapts to sitting on various objects, such as sofas, beds, and stools, revealing its versatility in recognizing different occupancies as potential seats. The test room features sparse furnishings, highlighting the model's ability to operate in open spaces diverging from MOB samples. Conversely, the Replica scene presents a more cluttered setup. Here, the model adeptly maneuvers through tight spaces and produces plausible movements. Furthermore, our model can generate long-term motions. The motions in Fig. 1 and Fig. 8 are generated in a single attempt by merely altering the target, both spanning around 30 seconds, which are clipped from longer motions for clear illustration in the limited space. More results are in the supplementary materials.", + "bbox": [ + 212, + 277, + 787, + 445 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "5.3 Interacting with Objects", + "text_level": 1, + "bbox": [ + 215, + 468, + 465, + 484 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Our flexible control signal for target pose enables HOI motion synthesis. By fine-tuning on OMOMO [20], given the initial pose and hand trajectories, we synthesize plausible motions to interact with objects as in Figure 9, which exhibits rational full-body motion w.r.t. hand and object trajectories. More results are in the supplementary.", + "bbox": [ + 212, + 494, + 785, + 571 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "5.4 Interacting with Dynamic Scenes", + "text_level": 1, + "bbox": [ + 215, + 595, + 535, + 611 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "During training, the canonical occupancy perceived by our controller is dynamically changing w.r.t. to the human movement, enabling it to handle dynamic scenarios though trained on essentially static MOB. We demonstrate the interaction in dynamic scenes with a custom automatic door scenario in Fig. 10. With the target (the circle) in front of the human (a) blocked by a revolving door, we control the human to adapt its speed w.r.t. the door till enough clearance and halt precisely at the target. This sequence exemplifies the model's ability to produce realistic and contextually appropriate motion in dynamic environments. More details are in the supplementary.", + "bbox": [ + 212, + 622, + 787, + 758 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "5.5 Ablation Studies", + "text_level": 1, + "bbox": [ + 215, + 782, + 401, + 796 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "With our controller as the baseline, selected ablation studies on MOB are reported in Tab. 1. Complete results and details are included in the supplementary.", + "bbox": [ + 212, + 809, + 785, + 840 + ], + "page_idx": 12 + }, + { + "type": "header", + "text": "Revisit Human-Scene Interaction via Space Occupancy", + "bbox": [ + 364, + 114, + 730, + 130 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "13", + "bbox": [ + 767, + 114, + 785, + 126 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Different loss terms. By removing penetration loss $L_{pen}$ , a major degeneration in Success rate and PEN is observed. This shows that explicitly punishing penetration is necessary. The removal of $L_{field}$ also hurts the success rate by $2.57\\%$ , with PEN increasing by 10.62. FS also increases notably. We identify that without $L_{field}$ , the field regulation would generate unnatural spikes.", + "bbox": [ + 212, + 146, + 782, + 220 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Field Regulation (FR) is also evaluated. Eliminating the field regulation decreases the success rate slightly, with a notable PEN increase (+18.28) and FS decrease (-1.56%), revealing the tradeoff it conducts between PEN and FS.", + "bbox": [ + 212, + 220, + 782, + 263 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Occupancy representation. We replaced the voxel-based representation of canonical occupancy with the more compact and efficient BPS [34]. The field regulation is also modified accordingly. Degradation is observed in most metrics. Despite its efficiency, BPS might ignore too much information and bring instability to the field regulation.", + "bbox": [ + 212, + 267, + 784, + 340 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "6 Discussion", + "text_level": 1, + "bbox": [ + 214, + 363, + 354, + 378 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Limitations. First, when interacting with dynamic scenes, our controller could perform sub-optimally with slight collisions. This could be ascribed to the space occupancy changing rate gap between training data and testing scenes. During training, the changing rate is determined by the human locomotion speed only. While for dynamic scenes, there exists additional space occupancy speed. Mitigating this gap is worth exploring. Second, foot sliding is another artifact that could happen. However, the metrics are based on simple heuristics with gaps compared to real foot sliding. A reasonable and effective tool to address this would be promising. Third, the current MOB poses its main focus on static HSI. Despite the proposed pipeline that could apply to HOI motion synthesis, further exploration of the space-occupancy view of HOI would be an interesting future direction. Finally, the proposed versatile controller only accepts target poses, space occupancy, and joint trajectories as control signals. Extending the controller for more signals would be promising.", + "bbox": [ + 212, + 392, + 784, + 603 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Broader Impacts. The exploration of HSI is beneficial for navigation in complex environments, with significant impacts on Embodied AI and Robotics. Also, multiple fields like animation production and interior design could be advanced. However, this might also be misused for false information.", + "bbox": [ + 212, + 623, + 782, + 683 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "7 Conclusion", + "text_level": 1, + "bbox": [ + 214, + 705, + 357, + 720 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "We revisited static HSI from a unified space occupancy view by identifying the major component of static HSI as Human-Occupancy Interaction. Large-scale Human Occupancy Interaction database MOB was aggregated from motion-only data, substantially improving diversity and complexity. Moreover, a versatile auto-regressive controller was trained on MOB for Human-Occupancy Interaction generation, which supports flexible and stable generation for static HSI under a spectrum of scenarios without training with real 3D scene data.", + "bbox": [ + 212, + 733, + 782, + 839 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 13 + }, + { + "type": "header", + "text": "Liu et al.", + "bbox": [ + 271, + 114, + 334, + 126 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Acknowledgments", + "text_level": 1, + "bbox": [ + 217, + 143, + 392, + 162 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "This work is supported in part by the National Natural Science Foundation of China under Grants No.62306175, No.62302296, National Key Research and Development Project of China (No.2022ZD0160102, No.2021ZD0110704), Shanghai Artificial Intelligence Laboratory, XPLORER PRIZE grants.", + "bbox": [ + 215, + 175, + 785, + 234 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 217, + 258, + 321, + 273 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Araújo, J.P., Li, J., Vetrivel, K., Agarwal, R., Wu, J., Gopinath, D., Clegg, A.W., Liu, K.: Circle: Capture in rich contextual environments. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 21211-21221 (2023)", + "2. Athanasiou, N., Petrovich, M., Black, M.J., Varol, G.: Teach: Temporal action composition for 3d humans. In: 2022 International Conference on 3D Vision (3DV). pp. 414-423. IEEE (2022)", + "3. Bengio, S., Vinyals, O., Jaitly, N., Shazeer, N.: Scheduled sampling for sequence prediction with recurrent neural networks. Advances in neural information processing systems 28 (2015)", + "4. Cai, Z., Ren, D., Zeng, A., Lin, Z., Yu, T., Wang, W., Fan, X., Gao, Y., Yu, Y., Pan, L., Hong, F., Zhang, M., Loy, C.C., Yang, L., Liu, Z.: Humman: Multi-modal 4d human dataset for versatile sensing and modeling. In: Avidan, S., Brostow, G., Cisse, M., Farinella, G.M., Hassner, T. (eds.) Computer Vision - ECCV 2022. pp. 557-577. Springer Nature Switzerland, Cham (2022)", + "5. Chen, L., Ng, R.: On the marriage of lp-norms and edit distance. In: Proceedings of the Thirtieth international conference on Very large data bases-Volume 30. pp. 792-803 (2004)", + "6. Chen, X., Jiang, B., Liu, W., Huang, Z., Fu, B., Chen, T., Yu, G.: Executing your commands via motion diffusion in latent space. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18000-18010 (2023)", + "7. Guo, C., Zou, S., Zuo, X., Wang, S., Ji, W., Li, X., Cheng, L.: Generating diverse and natural 3d human motions from text. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 5152-5161 (June 2022)", + "8. Guo, C., Zuo, X., Wang, S., Zou, S., Sun, Q., Deng, A., Gong, M., Cheng, L.: Action2motion: Conditioned generation of 3d human motions. In: Proceedings of the 28th ACM International Conference on Multimedia. pp. 2021-2029 (2020)", + "9. Guo, W., Du, Y., Shen, X., Lepetit, V., Alameda-Pineda, X., Moreno-Noguer, F.: Back to mlp: A simple baseline for human motion prediction. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 4809-4819 (2023)", + "0. Harvey, F.G., Yurick, M., Nowrouzezahrai, D., Pal, C.: Robust motion in-between. ACM Transactions on Graphics (TOG) 39(4), 60-1 (2020)", + "1. Hassan, M., Ceylan, D., Villegas, R., Saito, J., Yang, J., Zhou, Y., Black, M.J.: Stochastic scene-aware motion prediction. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 11374-11384 (2021)", + "2. Hassan, M., Choutas, V., Tzionas, D., Black, M.J.: Resolving 3D human pose ambiguities with 3D scene constraints. In: International Conference on Computer Vision. pp. 2282-2292 (Oct 2019), https://prox.is.tue.mpg.de" + ], + "bbox": [ + 225, + 287, + 784, + 839 + ], + "page_idx": 14 + }, + { + "type": "header", + "text": "Revisit Human-Scene Interaction via Space Occupancy", + "bbox": [ + 364, + 114, + 730, + 128 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "13. Hassan, M., Ghosh, P., Tesch, J., Tzionas, D., Black, M.J.: Populating 3d scenes by learning human-scene interaction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14708-14718 (2021)", + "14. Hassan, M., Guo, Y., Wang, T., Black, M., Fidler, S., Peng, X.B.: Synthesizing physical character-scene interactions. In: ACM SIGGRAPH 2023 Conference Proceedings. SIGGRAPH '23, Association for Computing Machinery, New York, NY, USA (2023). https://doi.org/10.1145/3588432.3591525, https://doi.org/10.1145/3588432.3591525", + "15. Hernandez, A., Gall, J., Moreno-Noguer, F.: Human motion prediction via spatiotemporal inpainting. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 7134-7143 (2019)", + "16. Huang, S., Wang, Z., Li, P., Jia, B., Liu, T., Zhu, Y., Liang, W., Zhu, S.C.: Diffusion-based generation, optimization, and planning in 3d scenes. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16750-16761 (2023)", + "17. Hwang, Y.K., Ahuja, N., et al.: A potential field approach to path planning. IEEE transactions on robotics and automation 8(1), 23-32 (1992)", + "18. Ji, Y., Xu, F., Yang, Y., Shen, F., Shen, H.T., Zheng, W.S.: A large-scale rgb-d database for arbitrary-view human action recognition. In: Proceedings of the 26th ACM international Conference on Multimedia. pp. 1510-1518 (2018)", + "19. Lee, J., Joo, H.: Locomotion-action-manipulation: Synthesizing human-scene interactions in complex 3d environments. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 9663-9674 (October 2023)", + "20. Li, J., Wu, J., Liu, C.K.: Object motion guided human motion synthesis. ACM Transactions on Graphics (TOG) 42(6), 1-11 (2023)", + "21. Li, R., Yang, S., Ross, D.A., Kanazawa, A.: Ai choreographer: Music conditioned 3d dance generation with aist++. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 13401-13412 (2021)", + "22. Li, R., Yang, S., Ross, D.A., Kanazawa, A.: Learn to dance with aist++: Music conditioned 3d dance generation. arXiv preprint arXiv:2101.08779 2(3) (2021)", + "23. Li, Y.L., Liu, X., Lu, H., Wang, S., Liu, J., Li, J., Lu, C.: Detailed 2d-3d joint representation for human-object interaction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10166–10175 (2020)", + "24. Lin, J., Zeng, A., Lu, S., Cai, Y., Zhang, R., Wang, H., Zhang, L.: Motion-x: A large-scale 3d expressive whole-body human motion dataset. arXiv preprint arXiv:2307.00818 (2023)", + "25. Loper, M., Mahmood, N., Romero, J., Pons-Moll, G., Black, M.J.: Smpl: A skinned multi-person linear model. ACM transactions on graphics (TOG) 34(6), 1-16 (2015)", + "26. Lucas, T., Baradel, F., Weinzaepfel, P., Rogez, G.: Posegpt: quantization-based 3d human motion generation and forecasting. In: Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part VI. pp. 417-435. Springer (2022)", + "27. Mahmood, N., Ghorbani, N., Troje, N.F., Pons-Moll, G., Black, M.J.: Amass: Archive of motion capture as surface shapes. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 5442-5451 (2019)", + "28. Nie, Y., Dai, A., Han, X., Nießner, M.: Pose2room: understanding 3d scenes from human activities. In: European Conference on Computer Vision. pp. 425-443. Springer (2022)" + ], + "bbox": [ + 215, + 146, + 785, + 839 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "16", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 15 + }, + { + "type": "header", + "text": "Liu et al.", + "bbox": [ + 271, + 114, + 334, + 126 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "29. Pavlakos, G., Choutas, V., Ghorbani, N., Bolkart, T., Osman, A.A.A., Tzionas, D., Black, M.J.: Expressive body capture: 3D hands, face, and body from a single image. In: Proceedings IEEE Conf. on Computer Vision and Pattern Recognition (CVPR). pp. 10975-10985 (2019)", + "30. Petrovich, M., Black, M.J., Varol, G.: Action-conditioned 3d human motion synthesis with transformer vae. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 10985-10995 (2021)", + "31. Petrovich, M., Black, M.J., Varol, G.: Temos: Generating diverse human motions from textual descriptions. In: Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXII. pp. 480-497. Springer (2022)", + "32. Petrovich, M., Black, M.J., Varol, G.: Tmr: Text-to-motion retrieval using contrastive 3d human motion synthesis. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 9488-9497 (October 2023)", + "33. Plappert, M., Mandery, C., Asfour, T.: The kit motion-language dataset. Big data 4(4), 236-252 (2016)", + "34. Prokudin, S., Lassner, C., Romero, J.: Efficient learning on point clouds with basis point sets. In: Proceedings of the IEEE International Conference on Computer Vision. pp. 4332-4341 (2019)", + "35. Puig, X., Undersander, E., Szot, A., Cote, M.D., Partsey, R., Yang, J., Desai, R., Clegg, A.W., Hlavac, M., Min, T., Gervet, T., Vondruš, V., Berges, V.P., Turner, J., Maksymets, O., Kira, Z., Kalakrishnan, M., Malik, J., Chaplot, D.S., Jain, U., Batra, D., Rai, A., Mottaghi, R.: Habitat 3.0: A co-habitat for humans, avatars and robots (2023)", + "36. Starke, S., Zhang, H., Komura, T., Saito, J.: Neural state machine for character-scene interactions. ACM Trans. Graph. 38(6), 209-1 (2019)", + "37. Straub, J., Whelan, T., Ma, L., Chen, Y., Wijmans, E., Green, S., Engel, J.J., Mur-Artal, R., Ren, C., Verma, S., et al.: The replica dataset: A digital replica of indoor spaces. arXiv preprint arXiv:1906.05797 (2019)", + "38. Szot, A., Clegg, A., Undersander, E., Wijmans, E., Zhao, Y., Turner, J., Maestre, N., Mukadam, M., Chaplot, D., Maksymets, O., Gokaslan, A., Vondrus, V., Dharur, S., Meier, F., Galuba, W., Chang, A., Kira, Z., Koltun, V., Malik, J., Savva, M., Batra, D.: Habitat 2.0: Training home assistants to rearrange their habitat. In: Advances in Neural Information Processing Systems (NeurIPS) (2021)", + "39. Taheri, O., Choutas, V., Black, M.J., Tzionas, D.: Goal: Generating 4d whole-body motion for hand-object grasping. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13263-13273 (2022)", + "40. Tessler, C., Kasten, Y., Guo, Y., Mannor, S., Chechik, G., Peng, X.B.: Calm: Conditional adversarial latent models for directable virtual characters. In: ACM SIGGRAPH 2023 Conference Proceedings. SIGGRAPH '23, Association for Computing Machinery, New York, NY, USA (2023). https://doi.org/10.1145/3588432.3591541, https://doi.org/10.1145/3588432.3591541", + "41. Tevet, G., Gordon, B., Hertz, A., Bermano, A.H., Cohen-Or, D.: Motionclip: Exposing human motion generation to clip space. In: Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXII. pp. 358-374. Springer (2022)", + "42. Tevet, G., Raab, S., Gordon, B., Shafir, Y., Cohen-or, D., Bermano, A.H.: Human motion diffusion model. In: The Eleventh International Conference on Learning Representations (2022)" + ], + "bbox": [ + 215, + 146, + 787, + 840 + ], + "page_idx": 16 + }, + { + "type": "header", + "text": "Revisit Human-Scene Interaction via Space Occupancy", + "bbox": [ + 364, + 114, + 730, + 128 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "43. Wang, J., Xu, H., Xu, J., Liu, S., Wang, X.: Synthesizing long-term 3d human motion and interaction in 3d scenes. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9401-9411 (2021)", + "44. Wang, J., Rong, Y., Liu, J., Yan, S., Lin, D., Dai, B.: Towards diverse and natural scene-aware 3d human motion synthesis. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 20460-20469 (2022)", + "45. Wang, Z., Chen, Y., Liu, T., Zhu, Y., Liang, W., Huang, S.: Humanise: Language-conditioned human motion generation in 3d scenes. Advances in Neural Information Processing Systems 35, 14959-14971 (2022)", + "46. Wu, Y., Wang, J., Zhang, Y., Zhang, S., Hilliges, O., Yu, F., Tang, S.: Saga: Stochastic whole-body grasping with contact. In: European Conference on Computer Vision. pp. 257-274. Springer (2022)", + "47. Xu, L., Song, Z., Wang, D., Su, J., Fang, Z., Ding, C., Gan, W., Yan, Y., Jin, X., Yang, X., Zeng, W., Wu, W.: Actformer: A gan-based transformer towards general action-conditioned 3d human motion generation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 2228-2238 (October 2023)", + "48. Ye, S., Wang, Y., Li, J., Park, D., Liu, C.K., Xu, H., Wu, J.: Scene synthesis from human motion. In: SIGGRAPH Asia 2022 Conference Papers (2022)", + "49. Yi, H., Huang, C.H.P., Tripathi, S., Hering, L., Thies, J., Black, M.J.: Mime: Human-aware 3d scene generation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 12965-12976 (June 2023)", + "50. Yi, H., Huang, C.H.P., Tzionas, D., Kocabas, M., Hassan, M., Tang, S., Thies, J., Black, M.J.: Human-aware object placement for visual environment reconstruction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3959-3970 (2022)", + "51. Zhang, J.Y., Pepose, S., Joo, H., Ramanan, D., Malik, J., Kanazawa, A.: Perceiving 3d human-object spatial arrangements from a single image in the wild. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XII 16. pp. 34-51. Springer (2020)", + "52. Zhang, J., Zhang, Y., Cun, X., Zhang, Y., Zhao, H., Lu, H., Shen, X., Shan, Y.: Generating human motion from textual descriptions with discrete representations. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 14730-14740 (June 2023)", + "53. Zhang, M., Cai, Z., Pan, L., Hong, F., Guo, X., Yang, L., Liu, Z.: Motiondiffuse: Text-driven human motion generation with diffusion model. arXiv preprint arXiv:2208.15001 (2022)", + "54. Zhang, M., Guo, X., Pan, L., Cai, Z., Hong, F., Li, H., Yang, L., Liu, Z.: Remodiffuse: Retrieval-augmented motion diffusion model. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 364-373 (October 2023)", + "55. Zhang, S., Zhang, Y., Ma, Q., Black, M.J., Tang, S.: Place: Proximity learning of articulation and contact in 3d environments. In: 2020 International Conference on 3D Vision (3DV). pp. 642-651. IEEE (2020)", + "56. Zhang, X., Bhatnagar, B.L., Starke, S., Guzov, V., Pons-Moll, G.: Couch: Towards controllable human-chair interactions. In: European Conference on Computer Vision. pp. 518-535. Springer (2022)", + "57. Zhang, Y., Hassan, M., Neumann, H., Black, M.J., Tang, S.: Generating 3d people in scenes without people. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 6194-6204 (2020)" + ], + "bbox": [ + 215, + 146, + 785, + 839 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "18", + "bbox": [ + 217, + 114, + 235, + 126 + ], + "page_idx": 17 + }, + { + "type": "header", + "text": "Liu et al.", + "bbox": [ + 271, + 114, + 334, + 126 + ], + "page_idx": 17 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "58. Zhang, Y., Tang, S.: The wanderings of odysseus in 3d scenes. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 20481-20491 (2022)", + "59. Zhao, K., Wang, S., Zhang, Y., Beeler, T., Tang, S.: Compositional human-scene interaction synthesis with semantic control. In: European Conference on Computer Vision. pp. 311-327. Springer (2022)", + "60. Zhao, K., Zhang, Y., Wang, S., Beeler, T., Tang, S.: Synthesizing diverse human motions in 3d indoor scenes. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 14738-14749 (October 2023)", + "61. Zheng, Y., Yang, Y., Mo, K., Li, J., Yu, T., Liu, Y., Liu, C.K., Guibas, L.J.: Gimo: Gaze-informed human motion prediction in context. In: European Conference on Computer Vision. pp. 676-694. Springer (2022)", + "62. Zhou, Y., Barnes, C., Lu, J., Yang, J., Li, H.: On the continuity of rotation representations in neural networks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5745-5753 (2019)" + ], + "bbox": [ + 212, + 146, + 787, + 354 + ], + "page_idx": 18 + }, + { + "type": "header", + "text": "Revisit Human-Scene Interaction via Space Occupancy", + "bbox": [ + 364, + 114, + 730, + 128 + ], + "page_idx": 18 + }, + { + "type": "page_number", + "text": "19", + "bbox": [ + 767, + 116, + 784, + 126 + ], + "page_idx": 18 + } +] \ No newline at end of file diff --git a/2024/Revisit Human-Scene Interaction via Space Occupancy/f5977317-ef8e-4afa-8ef9-88b9ecbe0b04_model.json b/2024/Revisit Human-Scene Interaction via Space Occupancy/f5977317-ef8e-4afa-8ef9-88b9ecbe0b04_model.json new file mode 100644 index 0000000000000000000000000000000000000000..621728851072509b0541ce86e5c524286eb1b2fb --- /dev/null +++ b/2024/Revisit Human-Scene Interaction via Space Occupancy/f5977317-ef8e-4afa-8ef9-88b9ecbe0b04_model.json @@ -0,0 +1,2581 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.255, + 0.141, + 0.75, + 0.187 + ], + "angle": 0, + "content": "Revisit Human-Scene Interaction via Space Occupancy" + }, + { + "type": "text", + "bbox": [ + 0.229, + 0.212, + 0.776, + 0.243 + ], + "angle": 0, + "content": "Xinpeng Liu1, Haowen Hou1,2, Yanchao Yang3, Yong-Lu Li1, and Cewu Lu1" + }, + { + "type": "text", + "bbox": [ + 0.39, + 0.255, + 0.612, + 0.269 + ], + "angle": 0, + "content": "1 Shanghai Jiao Tong University" + }, + { + "type": "text", + "bbox": [ + 0.428, + 0.269, + 0.575, + 0.283 + ], + "angle": 0, + "content": "\\(^{2}\\) Soochow University" + }, + { + "type": "text", + "bbox": [ + 0.396, + 0.283, + 0.609, + 0.298 + ], + "angle": 0, + "content": "3 The University of Hong Kong" + }, + { + "type": "text", + "bbox": [ + 0.25, + 0.298, + 0.751, + 0.326 + ], + "angle": 0, + "content": "xinpengliu0907@gmail.com, haowenhou@outlook.com, yanchaoy@hku.hk, {yonglu_li, lucewu}@sjtu.edu.cn" + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.364, + 0.744, + 0.671 + ], + "angle": 0, + "content": "Abstract. Human-scene Interaction (HSI) generation is a challenging task and crucial for various downstream tasks. However, one of the major obstacles is its limited data scale. High-quality data with simultaneously captured human and 3D environments is hard to acquire, resulting in limited data diversity and complexity. In this work, we argue that interaction with a scene is essentially interacting with the space occupancy of the scene from an abstract physical perspective, leading us to a unified novel view of Human-Occupancy Interaction. By treating pure motion sequences as records of humans interacting with invisible scene occupancy, we can aggregate motion-only data into a large-scale paired human-occupancy interaction database: Motion Occupancy Base (MOB). Thus, the need for costly paired motion-scene datasets with high-quality scene scans can be substantially alleviated. With this new unified view of Human-Occupancy interaction, a single motion controller is proposed to reach the target state given the surrounding occupancy. Once trained on MOB with complex occupancy layout, which is stringent to human movements, the controller could handle cramped scenes and generalize well to general scenes with limited complexity like regular living rooms. With no GT 3D scenes for training, our method can generate realistic and stable HSI motions in diverse scenarios, including both static and dynamic scenes. The project is available at https://foruck.github.io/occu-page/." + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.683, + 0.663, + 0.698 + ], + "angle": 0, + "content": "Keywords: Human-Scene Interaction \\(\\cdot\\) Motion Generation" + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.724, + 0.377, + 0.74 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.757, + 0.788, + 0.803 + ], + "angle": 0, + "content": "Human-Scene Interaction (HSI) generation has been an active field recently with its application in animation [36] and embodied AI [35, 38]. For ideal HSI generation, models should be able to control the flexible articulated human body" + }, + { + "type": "page_footnote", + "bbox": [ + 0.216, + 0.812, + 0.365, + 0.825 + ], + "angle": 0, + "content": "* Equal contribution." + }, + { + "type": "page_footnote", + "bbox": [ + 0.216, + 0.826, + 0.384, + 0.84 + ], + "angle": 0, + "content": "\\*\\* Corresponding author." + }, + { + "type": "list", + "bbox": [ + 0.216, + 0.812, + 0.384, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "2" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.335, + 0.127 + ], + "angle": 0, + "content": "Liu et al." + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.145, + 0.787, + 0.297 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.309, + 0.788, + 0.366 + ], + "angle": 0, + "content": "Fig. 1: We propose that interacting with scenes is essentially interacting with its space occupancy for static HSI. In this view, we can unify motion-only data into a unified human-occupancy knowledge base and train a versatile Human-Occupancy Interaction controller upon it, achieving stable generation under various scenarios." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.399, + 0.785, + 0.444 + ], + "angle": 0, + "content": "to conduct various motions in highly complex 3D environments, considering geometry constraints, motion feasibility, and diversity. In this paper, we focus on interactions without changing the scene, which we refer to as static HSI." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.445, + 0.787, + 0.567 + ], + "angle": 0, + "content": "The limited data acquisition is a major obstacle for static HSI generation. Simultaneous capture of human and 3D scenes is expensive. Some [11,56] collected data only for a narrow range of target objects. Widely adopted PROX [12] contains only 60 clips with limited diversity and complexity. HUMANISE [45] synthesized data by fitting motion from large-scale motion base AMASS [27] into various scenes. CIRCLE [1] put the scenes in VR applications to reduce the cost. Most efforts take explicit scenes as necessary assets. Instead, we dig into the essence of static HSI and revisit it from a space occupancy view." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.568, + 0.788, + 0.748 + ], + "angle": 0, + "content": "Space occupancy is usually identified as an elementary constraint for HSI: humans and scenes cannot occupy the same space simultaneously, which prevents humans from penetrating scenes. We argue that occupancy is inherently the primary factor, especially for static HSI, which is essentially interaction with the scene occupancy. As in Fig. 1, sitting on a chair is practically sitting on a chair-shaped solid being. The chair's occupancy allows us to rest our weight upon it, offering options to sit or stand, i.e., affordances, while its height and shape naturally invite sitting. Other elements like category labels (e.g., identifying it as a \"chair\" semantically) influence our likelihood of choosing to sit. The scene occupancy, therefore, offers a range of affordances and delineates potential interaction types, establishing an initial probability distribution for interactions. This is primarily accessible to the model and should be learned well at first." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.75, + 0.788, + 0.841 + ], + "angle": 0, + "content": "In light of this, the requirements of realistic scenes for HSI learning could be loosened, inspiring us to revisit large-scale motion datasets. Instead of treating motion-only data as locomotion knowledge bases like existing efforts [58,60], we transform them into a comprehensive Human-Occupancy Interaction database, named Motion Occupancy Base (MOB). For any given motion sequence, we first identify the space occupied by the human at any time in the sequence. The" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.366, + 0.115, + 0.732, + 0.13 + ], + "angle": 0, + "content": "Revisit Human-Scene Interaction via Space Occupancy" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "3" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.785, + 0.237 + ], + "angle": 0, + "content": "remaining space is conceptualized as the \"scene\". In this framework, the motion becomes a record of interaction with these defined space occupancies. As a result, MOB aggregates enormous paired human-occupancy data from existing large-scale motion-only data. Compared to current datasets [1, 12, 45], MOB presents over-complex and stringent geometric constraints, pushing the model to the limit in interacting with space occupancy." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.238, + 0.786, + 0.584 + ], + "angle": 0, + "content": "How do we exploit the aggregated MOB for static HSI? Previous approaches generally provided two different answers. Some [1,16,45,61] adopt an end-to-end pipeline from scene to motion, which is straightforward but inflexible without the ability to interact with dynamic scenes. Some [43, 44, 60] typically decouple HSI into interaction and locomotion, adopting a two-stage paradigm: first generating key interaction frames and then synthesizing locomotion between them, with the help of methods like path-finding algorithms. However, this decomposition overlooks the integrated nature of locomotion and interaction, like simultaneous arm movement and walking to navigate obstacles. Our approach, rooted in space occupancy, reconceptualizes locomotion and interaction as a unified Human-Occupancy Interaction. We develop an auto-regressive HumanOccupancy Interaction controller for motion generation. The controller processes various inputs, including surrounding occupancy and target locations, as control signals at each time step, producing future motion predictions. Additionally, we introduce a field regulation module for enhanced collision avoidance. There are three major advantages of our controller. First, by learning from the challenging cases in MOB, it could handle complex scenes but also generalize well to regular simpler scenarios. Second, its versatility accommodates any static HSI without the need to switch between models for locomotion and interactions, even for HOI motion synthesis. Third, its auto-regressive nature allows it to adapt to dynamic environments such as an automatic door, a capability lacked in previous methods. In extensive experiments, our controller performs effectively across various settings, even without training on real 3D scene data." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.586, + 0.787, + 0.661 + ], + "angle": 0, + "content": "Overall, our contributions are three-fold: 1) We exhume Human-Occupancy Interaction as a major component of static HSI. 2) By digging out the inherent Human-Occupancy relationships of pure motions, we convert motion-only data into a paired Human-Occupancy database MOB, expanding the data scale. 3) We develop a versatile controller for flexible and stable HSI motion generation." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.686, + 0.396, + 0.701 + ], + "angle": 0, + "content": "2 Related Works" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.72, + 0.787, + 0.841 + ], + "angle": 0, + "content": "Human-Scene Interaction Generation. Generating natural human motion sequences has received increasing attention with the emerging MoCap datasets [4, 7, 18, 24, 27, 33]. Early efforts were paid to prefix/suffix-conditioned motion generation [2, 9, 10, 15], known as motion prediction and interpolation. There also have been efforts to pursue generating motion with respect to specific action semantics, either represented in action categories [8, 30, 47] or natural language [6, 7, 26, 31, 32, 41, 42, 52–54]. Audio-conditioned motion generation was also explored, like music for dance generation [21, 22]. However, these approaches tend" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "4" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.335, + 0.127 + ], + "angle": 0, + "content": "Liu et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.493 + ], + "angle": 0, + "content": "to focus on human motion only, without placing the generated motion in a grounded 3D environment. Recently, progress has been made to advance HSI generation. Early efforts [23,51] first reconstructed human-object interaction in 3D. Some [13,55,57] proceeded by placing SMPL-X [29] meshes in scanned 3D scenes. Semantic controls were also introduced [59]. For HSI motion generation, some [11,36,56] focused on interacting with a single target object, mostly chairs. GOAL [39] and SAGA [46] generated full-body pre-grasp motions. There have also been efforts [14,40] for the physical control of specific interactions. With paired motion-scene datasets [1,12], approaches that synthesize HSI motion in cluttered scenes appeared. Two-stage based algorithms [43,44] were proposed to first generate static key-frame poses, and then interpolate them for a coherent motion. GAMMA [58] proposed an RL framework for locomotion on flat grounds with the help of off-the-shelf path-finding algorithms. DIMOS [60] extended GAMMA with extra interaction policies. In contrast, the follow-ups [1,16,45,61] adopted a simple end-to-end one-stage paradigm, while focusing more on boosting data. GIMO [61] collected data with gaze information. HUMANISE [45] synthesized pseudo-data from large-scale motion bases and 3D scan bases. CIRCLE [1] collected data with a VR application. LAMA [19] coupled the RL framework with a motion matching algorithm, generated HSI motions with only a unified policy, and leveraged motion editing for optimization. Our major difference from previous efforts is the Human-Occupancy Interaction view, which alleviates the data hunger, and also enables a unified controller. Moreover, we are the first to generate motion in a dynamic scene, which previous efforts failed to handle." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.494, + 0.789, + 0.571 + ], + "angle": 0, + "content": "Scene Creation from Motion. Pose2Room [28] predicted 3D object bounding boxes from human motion. MOVER [50] reconstructed 3D scenes from videos that record human-scene interaction. SUMMON [48] and MIME [49] placed 3D object models concerning human motion. These methods could be adopted to boost HSI data, though usually limited to a small range of indoor furniture." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.595, + 0.489, + 0.613 + ], + "angle": 0, + "content": "3 Motion Occupancy Base" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.628, + 0.788, + 0.794 + ], + "angle": 0, + "content": "HSI generation could be formulated as \\( M = \\mathcal{G}(S) \\), where \\( M \\) is the generated motion, \\( \\mathcal{G} \\) is the motion generator, and \\( S \\) is a 3D scene. Conventionally, the scene is expected to be as realistic as possible, requiring expensive acquisition. We find that these scenes could function almost equally to their occupancy for static HSI. Thus, high-quality scenes might become less necessary here. Given this, we revisit the large-scale motion-only datasets. Intuitively, motion only contains body movement information. But if we switch our sight to the vacant space around the moving person, considering the space as another form of occupancy, we could find that these motions possess information on how humans interact with certain space occupancies. To this end, we aggregate motion-only datasets into the human-occupancy interaction database: Motion Occupancy Base (MOB)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.796, + 0.788, + 0.843 + ], + "angle": 0, + "content": "The overall pipeline is shown in Fig. 2. Motion with SMPL(-X) [25,29] representation are collected, re-sampled to 30 FPS, and rotated so that the gravity direction points to the negative z-axis. For motion sequence \\( X = \\{x^i\\}_{i=1}^t \\), where" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.366, + 0.115, + 0.733, + 0.131 + ], + "angle": 0, + "content": "Revisit Human-Scene Interaction via Space Occupancy" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "5" + }, + { + "type": "image", + "bbox": [ + 0.246, + 0.144, + 0.761, + 0.281 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.29, + 0.788, + 0.334 + ], + "angle": 0, + "content": "Fig. 3: The architecture of our versatile motion controller. Given motion state and histories \\( x^{t}, h^{t} \\), and control signals \\( c_{o}^{t}, c_{t}^{t} \\), the controller auto-regressively generates the next motion frame w.r.t. the target pose and the canonical occupancy." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.359, + 0.788, + 0.406 + ], + "angle": 0, + "content": "\\(x_{i}\\) is the SMPL(-X) representation in the \\(i\\)-th frame, \\(t\\) is the number of frames, we first obtain mesh sequence \\(X_{m} = \\{x_{m}^{i}\\}_{i = 1}^{t}\\) via SMPL(-X) model, where \\(x_{m}^{i} = (f^{i},v^{i})\\) is the mesh for the \\(i\\)-th frame, \\(f^{i},v^{i}\\) are the mesh faces and surface" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.405, + 0.489, + 0.663 + ], + "angle": 0, + "content": "vertices. A bounding box of \\(X_{m}\\) is obtained based on vertices \\(V = \\{v_{i}\\}_{i=1}^{t}\\). Given the predefined grid unit size \\(u\\), the box region is voxelized into motion occupancy volume \\(\\hat{O} \\in \\{0,1\\}^{[H \\times W \\times D]}\\), computed in two steps: 1) \\(V\\) occupy the corresponding voxels in \\(\\hat{O}\\). 2) Given the centers of the un-occupied voxels as \\(C = \\{c_{i}\\}_{i=1}^{|C|}\\), if \\(c_{i}\\) is inside any human mesh \\(x_{m}^{i}\\), the corresponding voxel is occupied. The pseudo-scene occupancy \\(O\\) is then obtained by negating the motion occupancy, formulated as \\(O = 1 - \\hat{O}\\). Thus, we convert motion-only sample \\(X\\) into a motion-occupancy pair \\((X,O)\\) in MOB. MOB" + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.435, + 0.788, + 0.58 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.59, + 0.789, + 0.634 + ], + "angle": 0, + "content": "Fig. 2: The construction process of Motion Occupancy Base. Note that the ceilings of the occupancy are hidden for clarity." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.663, + 0.786, + 0.678 + ], + "angle": 0, + "content": "aggregates 13 datasets with 98k instances. More details are in the supplementary." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.678, + 0.788, + 0.739 + ], + "angle": 0, + "content": "As shown in Fig. 2, humans in MOB are acting with rather strict geometric constraints, which is uncommon in regular scenes covered by existing datasets [1, 12, 45]. However, we show that by learning from hard cases in MOB, our model could grow its ability to handle both complex and regular scenarios well." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.759, + 0.521, + 0.775 + ], + "angle": 0, + "content": "4 Versatile Motion Controller" + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.788, + 0.373, + 0.802 + ], + "angle": 0, + "content": "4.1 Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.842 + ], + "angle": 0, + "content": "To instantiate \\( M = \\mathcal{G}(S) \\) for HSI generation, previous efforts either adopt a one-shot transformer generator or decompose \\( \\mathcal{G} \\) into two separate generators, one for" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "6" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.335, + 0.128 + ], + "angle": 0, + "content": "Liu et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.784, + 0.193 + ], + "angle": 0, + "content": "interaction key-frame generation, the other for key-frame interpolation. Instead, we formulate our generator as an auto-regressive controller. At time step \\( t \\), given historical motion state \\( h^t \\), current pose state \\( x^t \\), we expect a controller as" + }, + { + "type": "equation", + "bbox": [ + 0.439, + 0.203, + 0.786, + 0.22 + ], + "angle": 0, + "content": "\\[\n\\hat {y} ^ {t} = \\mathbf {G} \\left(h ^ {t}, x ^ {t}, c ^ {t}\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.23, + 0.784, + 0.291 + ], + "angle": 0, + "content": "where \\(\\hat{y}^t\\) is the predicted future motion, and \\(c^t\\) is the control signals. An overview is in Fig. 3. Sec. 4.2 introduces motion state representation \\(h^t, x^t, y^t\\). The control signals are covered in Sec. 4.3. The controller architecture is detailed in Sec. 4.4. Sec. 4.5 demonstrates the training and inference procedure." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.313, + 0.502, + 0.327 + ], + "angle": 0, + "content": "4.2 Motion State Representation" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.337, + 0.785, + 0.441 + ], + "angle": 0, + "content": "Notations. In the global coordinate, we represent motion sequence \\(X\\) with \\(T\\) frames as \\(X = \\{x^{t}\\}_{t = 1}^{T}\\), where \\(x^{t} = (p_{r}^{t},\\theta_{r}^{t},r^{t})\\) is SMPL [25] parameters, with root location \\(p_r^t\\in \\mathcal{R}^3\\), root orientation \\(\\theta_r^t\\in \\mathcal{R}^6\\), and the orientation of \\(j\\) limbs as \\(\\theta^t\\in \\mathcal{R}^{j\\times 6}\\). Based on this, the location of \\(j\\) joints \\(p^t\\in \\mathcal{R}^{j\\times 3}\\) could also be computed via SMPL. For simplicity, we use \\(\\cdot^{-w}\\) to represent the value from frame \\(t - w\\) to \\(t\\), and \\(\\cdot^{+w}\\) for \\(t\\) to \\(t + w\\). The 6D form from [62] is adopted for all rotation and orientation representations." + }, + { + "type": "text", + "bbox": [ + 0.24, + 0.443, + 0.573, + 0.456 + ], + "angle": 0, + "content": "Current Pose State could be formulated as" + }, + { + "type": "equation", + "bbox": [ + 0.368, + 0.468, + 0.785, + 0.485 + ], + "angle": 0, + "content": "\\[\nx ^ {t} = \\operatorname {C a n o} \\left(\\left\\{p _ {r} ^ {t}, p ^ {t}, \\theta_ {r} ^ {t}, \\theta^ {t}, \\dot {p} _ {r} ^ {t}, \\dot {p} ^ {t}, \\dot {\\theta} _ {r} ^ {t} \\right\\}, t\\right), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.495, + 0.784, + 0.525 + ], + "angle": 0, + "content": "where \\(\\cdot^t\\) indicates the first derivative at \\(t\\). Cano \\((\\cdot, t)\\) canonicalizes the input w.r.t. the root location and facing direction at \\(t\\), ensuring consistency across frames." + }, + { + "type": "text", + "bbox": [ + 0.24, + 0.526, + 0.761, + 0.54 + ], + "angle": 0, + "content": "Historical Motion State \\( h^t \\) captures motion from frame \\( t - w \\) to \\( t \\) as" + }, + { + "type": "equation", + "bbox": [ + 0.336, + 0.551, + 0.785, + 0.568 + ], + "angle": 0, + "content": "\\[\nh ^ {t} = \\operatorname {C a n o} \\left(\\left\\{p _ {r} ^ {- w}, p ^ {- w}, \\theta_ {r} ^ {- w}, \\dot {p} _ {r} ^ {- w}, \\dot {p} ^ {- w}, \\dot {\\theta} _ {r} ^ {- w} \\right\\}, t\\right), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.579, + 0.785, + 0.623 + ], + "angle": 0, + "content": "with trajectories \\( p_r^{-w}, p^{-w} \\), root orientation \\( \\theta_r^{-w} \\), velocities \\( \\dot{p}_r^{-w}, \\dot{p}^{-w} \\), and root angular velocity \\( \\dot{\\theta}_r^{-w} \\). For conciseness, if not specified, only the \\( p^{-w} \\) of the hands and feet is adopted in practice." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.624, + 0.785, + 0.734 + ], + "angle": 0, + "content": "Future Prediction \\(\\hat{y}^t\\) is composed of two parts as \\(\\hat{y}^t = \\{\\hat{x}^{t + 1},\\hat{u}^t\\}\\). Note that \\(\\hat{x}^{t + 1}\\) is in the egocentric canonical of frame \\(t\\), and \\(\\hat{p}_r,\\hat{\\dot{\\theta}}\\subset \\hat{x}^{t + 1}\\) are adopted to transit from \\(t\\) to \\(t + 1\\). Mirroring the historical motion state \\(h^t\\), we also incorporate a comprehensive set of future trajectories, facing directions and their changing rates as \\(\\hat{u}^t = \\{p_r^{+f},p^{+f},\\theta^{+f},\\dot{p}_r^{+f},\\dot{p}^{+f},\\dot{\\theta}^{+f}\\}\\) for a window size of \\(f\\), all in the canonical view of \\(t + 1\\). Empirically, integrating \\(\\hat{u}^t\\) significantly enhances the model's precision in navigating toward the target." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.755, + 0.389, + 0.77 + ], + "angle": 0, + "content": "4.3 Control Signals" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.784, + 0.809 + ], + "angle": 0, + "content": "In this section, we introduce how the different control signals are encoded. Note that our models can operate without these optional signals." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.811, + 0.785, + 0.84 + ], + "angle": 0, + "content": "Canonical Occupancy. A simple voxel-based representation is adopted to encode how the canonical space is occupied by the scenes. We utilize a cubic" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.365, + 0.115, + 0.733, + 0.131 + ], + "angle": 0, + "content": "Revisit Human-Scene Interaction via Space Occupancy" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.116, + 0.787, + 0.127 + ], + "angle": 0, + "content": "7" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.283 + ], + "angle": 0, + "content": "point grid \\( o^t \\in \\{0,1\\}^{s \\times s \\times s} \\) characterized by the grid size \\( s \\) and the grid unit size \\( u \\) at frame \\( t \\). The grid is positioned in the human's canonical space, with its center offset by \\( s/4 \\) in the human-facing direction. Each cell contains a binary value, indicating whether the cell is intersecting with the global scene occupancy \\( O \\). The canonical occupancy is thus represented as \\( c_o^t = \\text{Flatten}(o^t) \\in \\{0,1\\}^{s^3} \\). Note that because the occupancy is calculated in real-time, this control signal can naturally adapt to dynamic scenes where each timestep is treated statically. We will show in Section 5 that even though our training data is all static, our model exhibits a smooth generalization to dynamic scenes." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.284, + 0.79, + 0.392 + ], + "angle": 0, + "content": "Target Pose is also considered a flexible control signal, which can be given in two ways. The first is the positions of the root, hands, and feet, each defined in relation to the canonical space of the current pose, represented as \\( c_t^t \\in \\mathcal{R}^{5 \\times 3} \\). In practice, these could be assigned manually or generated from off-the-shelf models [59]. Also, partial or blank targets are acceptable. The second is the future trajectory of end-effectors, denoted as \\( c_t^{+f} \\) for a future window size of \\( f \\). This paves the way for different applications in Section 5." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.418, + 0.458, + 0.431 + ], + "angle": 0, + "content": "4.4 Controller Architecture" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.447, + 0.601, + 0.674 + ], + "angle": 0, + "content": "We adopt a lightweight transformer-based architecture. \\(n\\) fully connected layers first project the \\(n\\) input features from \\(\\{h^t,x^t,c^t\\}\\) to the latent \\(z\\in \\mathcal{R}^{n\\times d}\\). \\(z\\) is then fed into encoder \\(E\\) to distill dynamic information from the motion states and control signals. We devise \\(E\\) as a sequence of Transformer encoders, which takes \\(z\\) as input and outputs refined dynamic information \\(\\hat{z} = E(z)\\in \\mathcal{R}^{n\\times d}\\). Finally, we adopt a transformer decoder \\(D\\) followed by a linear layer to condense the dynamic information, resulting in the final predictions. With learnable query \\(q\\in \\mathcal{R}^d\\) and the dynamic information \\(\\hat{z}\\) serving as K and V, respectively, the output of the transformer is further concatenated with a random noise \\(\\epsilon \\in \\mathcal{R}^d\\) for randomness. The linear layer then outputs the final prediction as \\(\\hat{y}^t = D(\\hat{z},q,\\epsilon)\\)." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.675, + 0.603, + 0.824 + ], + "angle": 0, + "content": "Furthermore, to enhance collision avoidance, we introduce a novel module called Field Regulation. The basic idea is that humans tend to move faster in obstacle-sparse areas and decelerate in complex environments or when an obstacle is nearby. At very short distances, especially with an imminent collision happening, a complete halt is necessary. Given this and taking inspiration from the potential field in robot motion planning [17], we propose Occupancy Field, which exerts negligible influence on humans at a distance but" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.825, + 0.787, + 0.841 + ], + "angle": 0, + "content": "escalates to a sturdy repulsion as proximity increases, thereby compelling the" + }, + { + "type": "image", + "bbox": [ + 0.614, + 0.478, + 0.787, + 0.611 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.61, + 0.621, + 0.789, + 0.79 + ], + "angle": 0, + "content": "Fig.4: Occupied voxel centers \\(\\{P_i\\}\\) are located at the top left of the figure. The initial velocity vector, \\(\\dot{p}_j\\), is directed upwards, posing a risk of collision. To mitigate this, the occupancy field introduces a corrective velocity component \\(\\Delta \\dot{p}_j\\). This redirects the final velocity \\(\\dot{p}_j^{out}\\) to avoid collision." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.23, + 0.127 + ], + "angle": 0, + "content": "8" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.335, + 0.128 + ], + "angle": 0, + "content": "Liu et al." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.147, + 0.785, + 0.177 + ], + "angle": 0, + "content": "human to decelerate and avoid collision. The occupancy field influences human motion only when a human is in motion toward it. The field functions as" + }, + { + "type": "equation", + "bbox": [ + 0.428, + 0.185, + 0.786, + 0.205 + ], + "angle": 0, + "content": "\\[\n\\Delta \\dot {p} _ {j} ^ {t + 1} = \\mathcal {F} \\left(p _ {j} ^ {t + 1}, c _ {o} ^ {t}\\right). \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.211, + 0.784, + 0.289 + ], + "angle": 0, + "content": "The field \\(\\mathcal{F}\\) calculates a correction \\(\\Delta \\dot{p}^{t + 1}\\) w.r.t. predicted joint velocity \\(\\dot{p}_j^{t + 1} \\in \\{\\dot{p}_r^{t + 1}, \\dot{p}^{t + 1}\\}\\) and canonical occupancy \\(c_o^t\\), which is then added to \\(\\dot{p}^{t + 1}\\). For simplicity, the superscripts indicating the time steps are omitted in the following equations. In detail, given the center points of all the occupied voxels in \\(c_o\\) as \\(P \\subset \\mathcal{R}^{n_p \\times 3}\\), the correction is calculated as" + }, + { + "type": "equation", + "bbox": [ + 0.345, + 0.298, + 0.786, + 0.334 + ], + "angle": 0, + "content": "\\[\n\\Delta \\dot {p} _ {j} = \\sum_ {P _ {i} \\in P} - k r _ {i} \\max \\left(0, \\frac {1}{\\| v _ {i j} \\| _ {a} - \\gamma} - b\\right), \\tag {5}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.359, + 0.339, + 0.785, + 0.356 + ], + "angle": 0, + "content": "\\[\nr _ {i} = \\dot {p} _ {j} \\max (0, \\langle \\dot {p} _ {j}, v _ {i j} \\rangle), \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.363, + 0.786, + 0.468 + ], + "angle": 0, + "content": "where \\( v_{ij} = P_i - p_j \\) is the vector from the joint \\( j \\) to the voxel center, \\( \\langle \\cdot, \\cdot \\rangle \\) indicates the cosine between the two vectors, \\( \\| \\|_a \\) denotes \\( a \\)-norm, and \\( k \\) is the stiffness factor. \\( \\gamma \\) and \\( b \\) are thresholds to constrain the distance range of effective voxels, where \\( \\gamma \\) reflects the human body's width, and voxels within this range are discounted for stability. An illustration of Field Regulation is shown in Fig. 4. The field regulation is incorporated in the training of the controller to adapt to it, avoiding potential foot sliding caused due to direct modification on velocities." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.488, + 0.451, + 0.503 + ], + "angle": 0, + "content": "4.5 Training and Inference" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.51, + 0.784, + 0.587 + ], + "angle": 0, + "content": "Training Objectives The controller is trained in a fully-supervised manner. We employ a mixture of L1 and L2 losses as basic objectives. Besides, penetration loss and occupancy field loss are designed to hinder undesired collisions. Specifically, L1 loss is utilized for rotation-related output values \\(\\hat{y}_{\\theta}^{t}\\in \\hat{y}^{t}\\), while L2 loss is applied to the rest outputs \\(\\hat{y}_p^t\\). The mixture loss is formulated as" + }, + { + "type": "equation", + "bbox": [ + 0.388, + 0.596, + 0.785, + 0.614 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {m i x} = \\left\\| y _ {\\theta} ^ {t} - \\hat {y} _ {\\theta} ^ {t} \\right\\| _ {1} + \\left\\| y _ {p} ^ {t} - \\hat {y} _ {p} ^ {t} \\right\\| _ {2}. \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.62, + 0.733, + 0.636 + ], + "angle": 0, + "content": "Given the predicted joint location \\(\\hat{p}^{t + 1}\\), penetration loss is computed as" + }, + { + "type": "equation", + "bbox": [ + 0.376, + 0.645, + 0.785, + 0.684 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {p e n} = \\sum_ {i = 0} ^ {j} O \\left(\\hat {p} _ {j} ^ {t + 1}\\right) \\| r e f \\left(\\hat {p} _ {j} ^ {t + 1}, O\\right) \\| _ {2}, \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.692, + 0.786, + 0.801 + ], + "angle": 0, + "content": "where \\( O(\\cdot) \\) returns 1 if the point is occupied in \\( O \\) else 0, and \\( ref(\\hat{p}_j^{t + 1},O) \\) calculates the closest unoccupied voxel center in \\( O \\) for joint \\( \\hat{p}_j^{t + 1} \\). Occupancy field loss has two components. First, to deter undesired interactions, we penalize the proportion of the occupancy field effect relative to the output root velocity. Second, observing occasional unnatural spikes in velocity, where the human figure seemingly attempts to navigate through an obstacle, we punish excessive velocities. The occupancy field loss is formulated as" + }, + { + "type": "equation", + "bbox": [ + 0.407, + 0.809, + 0.785, + 0.845 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {f i e l d}} = \\left(\\frac {| \\Delta \\dot {p} |}{| \\dot {p} |}\\right) ^ {2} + | \\dot {p} | ^ {2}. \\tag {9}\n\\]" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.366, + 0.115, + 0.731, + 0.129 + ], + "angle": 0, + "content": "Revisit Human-Scene Interaction via Space Occupancy" + }, + { + "type": "page_number", + "bbox": [ + 0.775, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "9" + }, + { + "type": "table", + "bbox": [ + 0.306, + 0.144, + 0.703, + 0.256 + ], + "angle": 0, + "content": "
MethodSuc. (%) ↑DT (cm) ↓Time (s)FS (%)PEN ↓ERP
GT100.000.003.003.360.00-
NSM [36]11.7632.795.3311.04134.497.26
SAMP [11]13.9452.496.549.87145.746.98
DIMOS [60]15.7348.976.8210.52123.126.82
Ours w/o Occupancy71.2214.662.256.8946.054.13
Ours77.9115.303.628.1315.214.24
Ours w/o Lpen74.1414.893.547.8931.934.96
Ours w/o Lfield75.3415.692.8611.6525.835.70
Ours w/o FR77.3814.852.346.5733.494.18
Ours w/ BPS73.6416.833.359.3436.724.72
" + }, + { + "type": "table_caption", + "bbox": [ + 0.368, + 0.256, + 0.634, + 0.269 + ], + "angle": 0, + "content": "Table 1: Quantitative results on MOB." + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.311, + 0.611, + 0.326 + ], + "angle": 0, + "content": "The final loss is computed with loss coefficients \\(\\alpha, \\beta\\) as" + }, + { + "type": "equation", + "bbox": [ + 0.399, + 0.339, + 0.785, + 0.353 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathcal {L} _ {m i x} + \\alpha \\mathcal {L} _ {p e n} + \\beta \\mathcal {L} _ {\\text {f i e l d}}. \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.364, + 0.788, + 0.515 + ], + "angle": 0, + "content": "Auto-regressive Inference. We perform motion generation in an auto-regressive manner, enabling the model to adapt to dynamic environmental changes. Since the model solely manipulates canonical information, we track the global root position \\( p_{g} \\) and the facing direction \\( \\theta_{g} \\) of the human in the world coordinate system. Initially, we set past and current velocities \\( \\dot{p}^{0} \\) to zero. Before each prediction at frame \\( t \\), we compute the control signals (if there are any), i.e., the relative target position \\( c_{t}^{t} \\) and the canonical space occupancy \\( c_{o}^{t} \\), based on the current values of \\( p_{g}^{t} \\) and \\( \\theta_{g}^{t} \\). After the prediction, we update \\( p_{g}^{t} \\) and \\( \\theta_{g}^{t} \\) with the model's output, along with the current pose state \\( c^{t + 1} \\). Concurrently, we refresh the historical motion state \\( h^{t + 1} \\) to prepare for the next input." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.538, + 0.376, + 0.556 + ], + "angle": 0, + "content": "5 Experiments" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.568, + 0.788, + 0.735 + ], + "angle": 0, + "content": "Implementation Details. The controller consists of a two-layer transformer encoder and a single-layer transformer decoder, both with a latent dimension of 512 and 8 heads. The canonical occupancy grid size is set as (25, 25, 25), with a unit size of \\(8\\mathrm{cm}\\). The controller operates at 10 FPS with the history and future window sizes both set as 1, which empirically brings more stable outcomes. The controller is trained on the MOB, with 1,393 sequences excluded from training as the test set. We train the controller for 75k iterations, with a cosine learning rate decay restart strategy. The initial learning rate is set as 1e-4. The control signals are randomly masked for augmentation during training. Scheduled Sampling [3] is adopted. The whole training takes approximately 6 hours on a single 12G NVIDIA Titan Xp. More details are in the supplementary." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.756, + 0.588, + 0.772 + ], + "angle": 0, + "content": "5.1 Interacting with Occupancy from MOB" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.78, + 0.787, + 0.841 + ], + "angle": 0, + "content": "We first evaluate the ability to interact with occupancy volumes with a goal-reaching task on MOB. Given an occupancy volume, a starting pose, and a target end-effector pose, the model is required to generate transition motion from the starting pose to the target pose. We split the test set into 30k 3-second" + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "10" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.335, + 0.127 + ], + "angle": 0, + "content": "Liu et al." + }, + { + "type": "image", + "bbox": [ + 0.233, + 0.145, + 0.77, + 0.393 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.404, + 0.788, + 0.462 + ], + "angle": 0, + "content": "Fig. 5: Our controller can naturally reach the target (white to blue) given the complex occupancy. Furthermore, it can stabilize the motions (red to white) around the targets after reaching them. Previous SOTA [60] suffers from severe penetration and fails to reach the target. For simplicity, the ceilings of the occupancy are hidden." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.492, + 0.788, + 0.793 + ], + "angle": 0, + "content": "clips, adopting the first frame as the starting pose and the final frame as the target pose. The models are required to generate motion for 10s towards the target. There are two reasons for this setting. First, we would like to provide enough time for the models to accomplish the tasks. Second, we are also curious about how the models would behave after the target is reached and no new target is given, or when a very close target is given. The results are evaluated from three aspects. Task Completion is determined by whether the target pose is reached. The target is identified as completed if the average Euclidean distance between the generated and target pose is lower than 20 centimeters and less than 50 voxels are penetrated. We report the success rate (Suc.). The average minimum distance to the target pose (DT) and the time taken to reach the target (Time) are reported for reference. Note that Time is only calculated for successful samples. Motion Quality is evaluated via foot sliding (FS) and penetration (PEN) detection. A frame is identified as FS with its foot velocity \\(>7.5\\mathrm{cm / s}\\) following GAMMA [58]. We report the percentage of frames with FS. For PEN, we report the number of voxels penetrated per frame. Similarity to Dataset is measured by the similarity between the predicted and GT target end-effector trajectory. We report Edit distance for the real penalty (ERP) [5] since the trajectories might have different lengths. It is only given as a reference, which is not the lower the better." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.795, + 0.788, + 0.842 + ], + "angle": 0, + "content": "Quantitative results are shown in Tab. 1. For DIMOS [60], we convert the target pose into corresponding marker sets and report the best results from different policies for each test sample. For w/o Occ., we remove the canonical" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.366, + 0.115, + 0.732, + 0.13 + ], + "angle": 0, + "content": "Revisit Human-Scene Interaction via Space Occupancy" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.784, + 0.127 + ], + "angle": 0, + "content": "11" + }, + { + "type": "image", + "bbox": [ + 0.225, + 0.145, + 0.359, + 0.235 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.218, + 0.236, + 0.366, + 0.252 + ], + "angle": 0, + "content": "Space Occupancy overlay DIMOS Walkable Area from NavMesh" + }, + { + "type": "image", + "bbox": [ + 0.37, + 0.144, + 0.501, + 0.235 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.379, + 0.239, + 0.492, + 0.248 + ], + "angle": 0, + "content": "DIMOS path planning results" + }, + { + "type": "image_caption", + "bbox": [ + 0.503, + 0.159, + 0.549, + 0.17 + ], + "angle": 0, + "content": "Sit Down" + }, + { + "type": "image", + "bbox": [ + 0.555, + 0.145, + 0.787, + 0.253 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.263, + 0.789, + 0.32 + ], + "angle": 0, + "content": "Fig. 6: A typical DIMOS [60] failure case. Fig. 7: With the unified view of human-The NavMesh algorithm produces sub- occupancy interaction, the same controller optimal walkable areas, resulting in path- in Fig. 5 produces diverse motions, includ-finding failures for \\(69.43\\%\\) MOB samples. ing sitting down, crawling, and getting up." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.355, + 0.789, + 0.552 + ], + "angle": 0, + "content": "occupancy-related components. Previous efforts frequently fail to reach the goal successfully in complex scenarios, which we will elucidate later with qualitative results. Even when the goal is reached, severe penetration emerges, resulting in low success rates in Tab. 1. Our models outperform DIMOS by a considerable margin with a single controller compared to multiple policies. This reveals that the straightforward voxel-based representation is capable of encoding holistic information for human-occupancy interaction. The model without occupancy manages to approach the target as closely and quickly as possible. However, it suffers from more severe interpenetration with the occupancy, resulting in a lower success rate. Instead, our model with consideration of occupancy behaves more conservatively, spending more time to accomplish the tasks with less penetration. A slightly higher foot sliding rate is also observed, which is a trade-off for the penetration-avoiding ability that field regulation brings." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.554, + 0.789, + 0.842 + ], + "angle": 0, + "content": "Qualitative results are shown in Fig. 5, 7. Given the complex occupancy, previous state-of-the-art DIMOS [60] frequently fails to reach the goal. Even when it reaches the goal, severe penetration is usually observed. Instead, our controller generates feasible motions to reach the targets. Previous methods' typical locomotion-interaction decomposition [44, 58, 60] tends to over-simplify both. For locomotion, most methods assume a rigid character (usually simplified as a rigid cylinder in \\(\\mathrm{A}^*\\) [11, 44] and NavMesh algorithms [60]), overseeing the flexibility of humans. It might be okay for spacious rooms, however, when posed in a cramped scene or scene with complex geometric structures like in Fig. 5, models could fail in path-finding and thus degenerate. As shown in Fig. 6, for \\(69.43\\%\\) MOB samples, the NavMesh algorithm adopted by DIMOS failed to generate a feasible path from the start position to the target position. The statistic for \\(\\mathrm{A}^*\\) from SAMP [11] is \\(70.29\\%\\). This explains the unsatisfying quantitative results of DIMOS. For interaction, the categories are typically limited to predefined ones. In contrast, with our unified view of Human-Occupancy interaction, our controller applies to various complex scenarios beyond simplified locomotion and interaction. Furthermore, diverse motions could be performed by the same controller (Fig. 7). By comparing our controllers with and without occupancy, we find that training with occupancy helps to avoid penetration significantly." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "12" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.335, + 0.127 + ], + "angle": 0, + "content": "Liu et al." + }, + { + "type": "table", + "bbox": [ + 0.278, + 0.144, + 0.731, + 0.198 + ], + "angle": 0, + "content": "
ModelSuc. (%)Dist. to goal (cm)Collision depth (cm)FS (%)
GT100.000.008.742.45
Ours (w/ MOB only)73.017.199.5613.83
Ours (w/ CIRCLE only)75.526.8813.7411.76
Ours (w/ MOB + CIRCLE)78.976.7311.8512.87
" + }, + { + "type": "table_caption", + "bbox": [ + 0.388, + 0.199, + 0.614, + 0.212 + ], + "angle": 0, + "content": "Table 2: Results on CIRCLE [1]." + }, + { + "type": "image", + "bbox": [ + 0.258, + 0.238, + 0.455, + 0.359 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.539, + 0.238, + 0.743, + 0.363 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.373, + 0.79, + 0.43 + ], + "angle": 0, + "content": "Fig. 8: We can generate long-term motion. Fig. 9: We synthesize Human-Object Interactions in realistic rooms: A human circulation action, with hand trajectory as the control navigates a cabinet, sits on a sofa, perches signal. The human successfully moves a floor on a bed, and then resumes walking. lamp to a new place." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.46, + 0.785, + 0.49 + ], + "angle": 0, + "content": "Also, the controller with occupancy could stabilize the motion around the target after reaching it, while its counterpart tends to wander further inertially." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.513, + 0.51, + 0.528 + ], + "angle": 0, + "content": "5.2 Interacting with Static Scenes" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.538, + 0.785, + 0.567 + ], + "angle": 0, + "content": "Once our controller is trained on the more complex MOB, it could be effortlessly applied to realistic synthesized rooms and scene scans." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.569, + 0.787, + 0.794 + ], + "angle": 0, + "content": "Quantitative results is reported on CIRCLE [1]. Given the initial full-body pose and the target right wrist position, the model is required to generate a right-hand reaching motion. Following [1], success rate, distance to the goal, cumulative collision depth, and foot sliding are reported as metrics. Besides the controller trained on MOB (excluding CIRCLE), we also train our controller on CIRCLE-only and MOB+CIRCLE. We randomly split CIRCLE into a 2,565-sample train set (for the ablative baseline only) and a 453-sample test set since the original split has not been made public yet. The results in Tab. 2 reveal that our controller could handle various tasks in complex natural scenes well. Also, its potential to advance HOI-related motion, like hand-reaching before grasping, is showcased. It is noticeable that the MOB-only model demonstrates a competitive success rate and better collision avoidance, revealing the efficacy of MOB with its advanced complexity. When incorporating MOB with CIRCLE, superior task completion is observed, showing that MOB could function as an effective auxiliary resource for the conventional HSI datasets." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.796, + 0.787, + 0.84 + ], + "angle": 0, + "content": "Qualitative results. We test our model in a synthetic room from DIMOS [60] in Fig. 8 and a realistic scan from Replica [37] in Fig. 1. The occupancy is transferred from the scene SDFs. Our model produces convincing motions, nav-" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.365, + 0.115, + 0.732, + 0.131 + ], + "angle": 0, + "content": "Revisit Human-Scene Interaction via Space Occupancy" + }, + { + "type": "page_number", + "bbox": [ + 0.768, + 0.116, + 0.786, + 0.127 + ], + "angle": 0, + "content": "13" + }, + { + "type": "image", + "bbox": [ + 0.254, + 0.144, + 0.75, + 0.221 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.234, + 0.785, + 0.249 + ], + "angle": 0, + "content": "Fig. 10: Our model can avoid dynamic obstacles and reasonably reach the target circle." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.279, + 0.788, + 0.446 + ], + "angle": 0, + "content": "igates around obstacles, and transitions seamlessly between locomotion and interaction. Remarkably, it adapts to sitting on various objects, such as sofas, beds, and stools, revealing its versatility in recognizing different occupancies as potential seats. The test room features sparse furnishings, highlighting the model's ability to operate in open spaces diverging from MOB samples. Conversely, the Replica scene presents a more cluttered setup. Here, the model adeptly maneuvers through tight spaces and produces plausible movements. Furthermore, our model can generate long-term motions. The motions in Fig. 1 and Fig. 8 are generated in a single attempt by merely altering the target, both spanning around 30 seconds, which are clipped from longer motions for clear illustration in the limited space. More results are in the supplementary materials." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.469, + 0.466, + 0.485 + ], + "angle": 0, + "content": "5.3 Interacting with Objects" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.496, + 0.787, + 0.572 + ], + "angle": 0, + "content": "Our flexible control signal for target pose enables HOI motion synthesis. By fine-tuning on OMOMO [20], given the initial pose and hand trajectories, we synthesize plausible motions to interact with objects as in Figure 9, which exhibits rational full-body motion w.r.t. hand and object trajectories. More results are in the supplementary." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.596, + 0.536, + 0.612 + ], + "angle": 0, + "content": "5.4 Interacting with Dynamic Scenes" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.623, + 0.788, + 0.76 + ], + "angle": 0, + "content": "During training, the canonical occupancy perceived by our controller is dynamically changing w.r.t. to the human movement, enabling it to handle dynamic scenarios though trained on essentially static MOB. We demonstrate the interaction in dynamic scenes with a custom automatic door scenario in Fig. 10. With the target (the circle) in front of the human (a) blocked by a revolving door, we control the human to adapt its speed w.r.t. the door till enough clearance and halt precisely at the target. This sequence exemplifies the model's ability to produce realistic and contextually appropriate motion in dynamic environments. More details are in the supplementary." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.783, + 0.402, + 0.797 + ], + "angle": 0, + "content": "5.5 Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.81, + 0.787, + 0.842 + ], + "angle": 0, + "content": "With our controller as the baseline, selected ablation studies on MOB are reported in Tab. 1. Complete results and details are included in the supplementary." + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "14" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.335, + 0.127 + ], + "angle": 0, + "content": "Liu et al." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.147, + 0.783, + 0.221 + ], + "angle": 0, + "content": "Different loss terms. By removing penetration loss \\( L_{pen} \\), a major degeneration in Success rate and PEN is observed. This shows that explicitly punishing penetration is necessary. The removal of \\( L_{field} \\) also hurts the success rate by \\( 2.57\\% \\), with PEN increasing by 10.62. FS also increases notably. We identify that without \\( L_{field} \\), the field regulation would generate unnatural spikes." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.222, + 0.784, + 0.265 + ], + "angle": 0, + "content": "Field Regulation (FR) is also evaluated. Eliminating the field regulation decreases the success rate slightly, with a notable PEN increase (+18.28) and FS decrease (-1.56%), revealing the tradeoff it conducts between PEN and FS." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.268, + 0.785, + 0.342 + ], + "angle": 0, + "content": "Occupancy representation. We replaced the voxel-based representation of canonical occupancy with the more compact and efficient BPS [34]. The field regulation is also modified accordingly. Degradation is observed in most metrics. Despite its efficiency, BPS might ignore too much information and bring instability to the field regulation." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.364, + 0.355, + 0.38 + ], + "angle": 0, + "content": "6 Discussion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.393, + 0.785, + 0.604 + ], + "angle": 0, + "content": "Limitations. First, when interacting with dynamic scenes, our controller could perform sub-optimally with slight collisions. This could be ascribed to the space occupancy changing rate gap between training data and testing scenes. During training, the changing rate is determined by the human locomotion speed only. While for dynamic scenes, there exists additional space occupancy speed. Mitigating this gap is worth exploring. Second, foot sliding is another artifact that could happen. However, the metrics are based on simple heuristics with gaps compared to real foot sliding. A reasonable and effective tool to address this would be promising. Third, the current MOB poses its main focus on static HSI. Despite the proposed pipeline that could apply to HOI motion synthesis, further exploration of the space-occupancy view of HOI would be an interesting future direction. Finally, the proposed versatile controller only accepts target poses, space occupancy, and joint trajectories as control signals. Extending the controller for more signals would be promising." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.624, + 0.784, + 0.684 + ], + "angle": 0, + "content": "Broader Impacts. The exploration of HSI is beneficial for navigation in complex environments, with significant impacts on Embodied AI and Robotics. Also, multiple fields like animation production and interior design could be advanced. However, this might also be misused for false information." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.706, + 0.358, + 0.721 + ], + "angle": 0, + "content": "7 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.734, + 0.784, + 0.84 + ], + "angle": 0, + "content": "We revisited static HSI from a unified space occupancy view by identifying the major component of static HSI as Human-Occupancy Interaction. Large-scale Human Occupancy Interaction database MOB was aggregated from motion-only data, substantially improving diversity and complexity. Moreover, a versatile auto-regressive controller was trained on MOB for Human-Occupancy Interaction generation, which supports flexible and stable generation for static HSI under a spectrum of scenarios without training with real 3D scene data." + } + ], + [ + { + "type": "header", + "bbox": [ + 0.366, + 0.115, + 0.732, + 0.13 + ], + "angle": 0, + "content": "Revisit Human-Scene Interaction via Space Occupancy" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "15" + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.145, + 0.393, + 0.163 + ], + "angle": 0, + "content": "Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.176, + 0.787, + 0.236 + ], + "angle": 0, + "content": "This work is supported in part by the National Natural Science Foundation of China under Grants No.62306175, No.62302296, National Key Research and Development Project of China (No.2022ZD0160102, No.2021ZD0110704), Shanghai Artificial Intelligence Laboratory, XPLORER PRIZE grants." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.259, + 0.323, + 0.275 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.288, + 0.785, + 0.344 + ], + "angle": 0, + "content": "1. Araújo, J.P., Li, J., Vetrivel, K., Agarwal, R., Wu, J., Gopinath, D., Clegg, A.W., Liu, K.: Circle: Capture in rich contextual environments. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 21211-21221 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.346, + 0.785, + 0.385 + ], + "angle": 0, + "content": "2. Athanasiou, N., Petrovich, M., Black, M.J., Varol, G.: Teach: Temporal action composition for 3d humans. In: 2022 International Conference on 3D Vision (3DV). pp. 414-423. IEEE (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.387, + 0.785, + 0.426 + ], + "angle": 0, + "content": "3. Bengio, S., Vinyals, O., Jaitly, N., Shazeer, N.: Scheduled sampling for sequence prediction with recurrent neural networks. Advances in neural information processing systems 28 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.428, + 0.785, + 0.496 + ], + "angle": 0, + "content": "4. Cai, Z., Ren, D., Zeng, A., Lin, Z., Yu, T., Wang, W., Fan, X., Gao, Y., Yu, Y., Pan, L., Hong, F., Zhang, M., Loy, C.C., Yang, L., Liu, Z.: Humman: Multi-modal 4d human dataset for versatile sensing and modeling. In: Avidan, S., Brostow, G., Cisse, M., Farinella, G.M., Hassner, T. (eds.) Computer Vision - ECCV 2022. pp. 557-577. Springer Nature Switzerland, Cham (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.498, + 0.785, + 0.536 + ], + "angle": 0, + "content": "5. Chen, L., Ng, R.: On the marriage of lp-norms and edit distance. In: Proceedings of the Thirtieth international conference on Very large data bases-Volume 30. pp. 792-803 (2004)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.539, + 0.785, + 0.578 + ], + "angle": 0, + "content": "6. Chen, X., Jiang, B., Liu, W., Huang, Z., Fu, B., Chen, T., Yu, G.: Executing your commands via motion diffusion in latent space. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18000-18010 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.58, + 0.785, + 0.633 + ], + "angle": 0, + "content": "7. Guo, C., Zou, S., Zuo, X., Wang, S., Ji, W., Li, X., Cheng, L.: Generating diverse and natural 3d human motions from text. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 5152-5161 (June 2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.635, + 0.785, + 0.674 + ], + "angle": 0, + "content": "8. Guo, C., Zuo, X., Wang, S., Zou, S., Sun, Q., Deng, A., Gong, M., Cheng, L.: Action2motion: Conditioned generation of 3d human motions. In: Proceedings of the 28th ACM International Conference on Multimedia. pp. 2021-2029 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.676, + 0.785, + 0.731 + ], + "angle": 0, + "content": "9. Guo, W., Du, Y., Shen, X., Lepetit, V., Alameda-Pineda, X., Moreno-Noguer, F.: Back to mlp: A simple baseline for human motion prediction. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 4809-4819 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.732, + 0.785, + 0.757 + ], + "angle": 0, + "content": "0. Harvey, F.G., Yurick, M., Nowrouzezahrai, D., Pal, C.: Robust motion in-between. ACM Transactions on Graphics (TOG) 39(4), 60-1 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.759, + 0.785, + 0.798 + ], + "angle": 0, + "content": "1. Hassan, M., Ceylan, D., Villegas, R., Saito, J., Yang, J., Zhou, Y., Black, M.J.: Stochastic scene-aware motion prediction. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 11374-11384 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.226, + 0.8, + 0.785, + 0.84 + ], + "angle": 0, + "content": "2. Hassan, M., Choutas, V., Tzionas, D., Black, M.J.: Resolving 3D human pose ambiguities with 3D scene constraints. In: International Conference on Computer Vision. pp. 2282-2292 (Oct 2019), https://prox.is.tue.mpg.de" + }, + { + "type": "list", + "bbox": [ + 0.226, + 0.288, + 0.785, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "16" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.335, + 0.127 + ], + "angle": 0, + "content": "Liu et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.787, + 0.189 + ], + "angle": 0, + "content": "13. Hassan, M., Ghosh, P., Tesch, J., Tzionas, D., Black, M.J.: Populating 3d scenes by learning human-scene interaction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14708-14718 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.19, + 0.787, + 0.259 + ], + "angle": 0, + "content": "14. Hassan, M., Guo, Y., Wang, T., Black, M., Fidler, S., Peng, X.B.: Synthesizing physical character-scene interactions. In: ACM SIGGRAPH 2023 Conference Proceedings. SIGGRAPH '23, Association for Computing Machinery, New York, NY, USA (2023). https://doi.org/10.1145/3588432.3591525, https://doi.org/10.1145/3588432.3591525" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.261, + 0.787, + 0.302 + ], + "angle": 0, + "content": "15. Hernandez, A., Gall, J., Moreno-Noguer, F.: Human motion prediction via spatiotemporal inpainting. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 7134-7143 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.303, + 0.787, + 0.358 + ], + "angle": 0, + "content": "16. Huang, S., Wang, Z., Li, P., Jia, B., Liu, T., Zhu, Y., Liang, W., Zhu, S.C.: Diffusion-based generation, optimization, and planning in 3d scenes. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16750-16761 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.36, + 0.787, + 0.387 + ], + "angle": 0, + "content": "17. Hwang, Y.K., Ahuja, N., et al.: A potential field approach to path planning. IEEE transactions on robotics and automation 8(1), 23-32 (1992)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.388, + 0.787, + 0.43 + ], + "angle": 0, + "content": "18. Ji, Y., Xu, F., Yang, Y., Shen, F., Shen, H.T., Zheng, W.S.: A large-scale rgb-d database for arbitrary-view human action recognition. In: Proceedings of the 26th ACM international Conference on Multimedia. pp. 1510-1518 (2018)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.431, + 0.787, + 0.472 + ], + "angle": 0, + "content": "19. Lee, J., Joo, H.: Locomotion-action-manipulation: Synthesizing human-scene interactions in complex 3d environments. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 9663-9674 (October 2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.473, + 0.787, + 0.5 + ], + "angle": 0, + "content": "20. Li, J., Wu, J., Liu, C.K.: Object motion guided human motion synthesis. ACM Transactions on Graphics (TOG) 42(6), 1-11 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.502, + 0.787, + 0.543 + ], + "angle": 0, + "content": "21. Li, R., Yang, S., Ross, D.A., Kanazawa, A.: Ai choreographer: Music conditioned 3d dance generation with aist++. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 13401-13412 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.544, + 0.787, + 0.572 + ], + "angle": 0, + "content": "22. Li, R., Yang, S., Ross, D.A., Kanazawa, A.: Learn to dance with aist++: Music conditioned 3d dance generation. arXiv preprint arXiv:2101.08779 2(3) (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.573, + 0.787, + 0.614 + ], + "angle": 0, + "content": "23. Li, Y.L., Liu, X., Lu, H., Wang, S., Liu, J., Li, J., Lu, C.: Detailed 2d-3d joint representation for human-object interaction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10166–10175 (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.615, + 0.787, + 0.655 + ], + "angle": 0, + "content": "24. Lin, J., Zeng, A., Lu, S., Cai, Y., Zhang, R., Wang, H., Zhang, L.: Motion-x: A large-scale 3d expressive whole-body human motion dataset. arXiv preprint arXiv:2307.00818 (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.657, + 0.787, + 0.698 + ], + "angle": 0, + "content": "25. Loper, M., Mahmood, N., Romero, J., Pons-Moll, G., Black, M.J.: Smpl: A skinned multi-person linear model. ACM transactions on graphics (TOG) 34(6), 1-16 (2015)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.7, + 0.787, + 0.755 + ], + "angle": 0, + "content": "26. Lucas, T., Baradel, F., Weinzaepfel, P., Rogez, G.: Posegpt: quantization-based 3d human motion generation and forecasting. In: Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part VI. pp. 417-435. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.757, + 0.787, + 0.798 + ], + "angle": 0, + "content": "27. Mahmood, N., Ghorbani, N., Troje, N.F., Pons-Moll, G., Black, M.J.: Amass: Archive of motion capture as surface shapes. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 5442-5451 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.799, + 0.787, + 0.84 + ], + "angle": 0, + "content": "28. Nie, Y., Dai, A., Han, X., Nießner, M.: Pose2room: understanding 3d scenes from human activities. In: European Conference on Computer Vision. pp. 425-443. Springer (2022)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.366, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Revisit Human-Scene Interaction via Space Occupancy" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "17" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.147, + 0.788, + 0.205 + ], + "angle": 0, + "content": "29. Pavlakos, G., Choutas, V., Ghorbani, N., Bolkart, T., Osman, A.A.A., Tzionas, D., Black, M.J.: Expressive body capture: 3D hands, face, and body from a single image. In: Proceedings IEEE Conf. on Computer Vision and Pattern Recognition (CVPR). pp. 10975-10985 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.206, + 0.787, + 0.247 + ], + "angle": 0, + "content": "30. Petrovich, M., Black, M.J., Varol, G.: Action-conditioned 3d human motion synthesis with transformer vae. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 10985-10995 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.248, + 0.787, + 0.303 + ], + "angle": 0, + "content": "31. Petrovich, M., Black, M.J., Varol, G.: Temos: Generating diverse human motions from textual descriptions. In: Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXII. pp. 480-497. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.304, + 0.787, + 0.346 + ], + "angle": 0, + "content": "32. Petrovich, M., Black, M.J., Varol, G.: Tmr: Text-to-motion retrieval using contrastive 3d human motion synthesis. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 9488-9497 (October 2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.346, + 0.787, + 0.375 + ], + "angle": 0, + "content": "33. Plappert, M., Mandery, C., Asfour, T.: The kit motion-language dataset. Big data 4(4), 236-252 (2016)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.376, + 0.787, + 0.417 + ], + "angle": 0, + "content": "34. Prokudin, S., Lassner, C., Romero, J.: Efficient learning on point clouds with basis point sets. In: Proceedings of the IEEE International Conference on Computer Vision. pp. 4332-4341 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.418, + 0.787, + 0.487 + ], + "angle": 0, + "content": "35. Puig, X., Undersander, E., Szot, A., Cote, M.D., Partsey, R., Yang, J., Desai, R., Clegg, A.W., Hlavac, M., Min, T., Gervet, T., Vondruš, V., Berges, V.P., Turner, J., Maksymets, O., Kira, Z., Kalakrishnan, M., Malik, J., Chaplot, D.S., Jain, U., Batra, D., Rai, A., Mottaghi, R.: Habitat 3.0: A co-habitat for humans, avatars and robots (2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.488, + 0.787, + 0.516 + ], + "angle": 0, + "content": "36. Starke, S., Zhang, H., Komura, T., Saito, J.: Neural state machine for character-scene interactions. ACM Trans. Graph. 38(6), 209-1 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.517, + 0.787, + 0.559 + ], + "angle": 0, + "content": "37. Straub, J., Whelan, T., Ma, L., Chen, Y., Wijmans, E., Green, S., Engel, J.J., Mur-Artal, R., Ren, C., Verma, S., et al.: The replica dataset: A digital replica of indoor spaces. arXiv preprint arXiv:1906.05797 (2019)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.56, + 0.787, + 0.629 + ], + "angle": 0, + "content": "38. Szot, A., Clegg, A., Undersander, E., Wijmans, E., Zhao, Y., Turner, J., Maestre, N., Mukadam, M., Chaplot, D., Maksymets, O., Gokaslan, A., Vondrus, V., Dharur, S., Meier, F., Galuba, W., Chang, A., Kira, Z., Koltun, V., Malik, J., Savva, M., Batra, D.: Habitat 2.0: Training home assistants to rearrange their habitat. In: Advances in Neural Information Processing Systems (NeurIPS) (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.63, + 0.787, + 0.672 + ], + "angle": 0, + "content": "39. Taheri, O., Choutas, V., Black, M.J., Tzionas, D.: Goal: Generating 4d whole-body motion for hand-object grasping. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13263-13273 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.673, + 0.787, + 0.741 + ], + "angle": 0, + "content": "40. Tessler, C., Kasten, Y., Guo, Y., Mannor, S., Chechik, G., Peng, X.B.: Calm: Conditional adversarial latent models for directable virtual characters. In: ACM SIGGRAPH 2023 Conference Proceedings. SIGGRAPH '23, Association for Computing Machinery, New York, NY, USA (2023). https://doi.org/10.1145/3588432.3591541, https://doi.org/10.1145/3588432.3591541" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.742, + 0.787, + 0.798 + ], + "angle": 0, + "content": "41. Tevet, G., Gordon, B., Hertz, A., Bermano, A.H., Cohen-Or, D.: Motionclip: Exposing human motion generation to clip space. In: Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXII. pp. 358-374. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.799, + 0.787, + 0.841 + ], + "angle": 0, + "content": "42. Tevet, G., Raab, S., Gordon, B., Shafir, Y., Cohen-or, D., Bermano, A.H.: Human motion diffusion model. In: The Eleventh International Conference on Learning Representations (2022)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.788, + 0.841 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "page_number", + "bbox": [ + 0.218, + 0.116, + 0.236, + 0.127 + ], + "angle": 0, + "content": "18" + }, + { + "type": "header", + "bbox": [ + 0.272, + 0.115, + 0.335, + 0.127 + ], + "angle": 0, + "content": "Liu et al." + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.147, + 0.787, + 0.189 + ], + "angle": 0, + "content": "43. Wang, J., Xu, H., Xu, J., Liu, S., Wang, X.: Synthesizing long-term 3d human motion and interaction in 3d scenes. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9401-9411 (2021)" + }, + { + "type": "ref_text", + "bbox": [ + 0.218, + 0.189, + 0.787, + 0.23 + ], + "angle": 0, + "content": "44. Wang, J., Rong, Y., Liu, J., Yan, S., Lin, D., Dai, B.: Towards diverse and natural scene-aware 3d human motion synthesis. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 20460-20469 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.23, + 0.787, + 0.27 + ], + "angle": 0, + "content": "45. Wang, Z., Chen, Y., Liu, T., Zhu, Y., Liang, W., Huang, S.: Humanise: Language-conditioned human motion generation in 3d scenes. Advances in Neural Information Processing Systems 35, 14959-14971 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.27, + 0.787, + 0.311 + ], + "angle": 0, + "content": "46. Wu, Y., Wang, J., Zhang, Y., Zhang, S., Hilliges, O., Yu, F., Tang, S.: Saga: Stochastic whole-body grasping with contact. In: European Conference on Computer Vision. pp. 257-274. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.311, + 0.787, + 0.379 + ], + "angle": 0, + "content": "47. Xu, L., Song, Z., Wang, D., Su, J., Fang, Z., Ding, C., Gan, W., Yan, Y., Jin, X., Yang, X., Zeng, W., Wu, W.: Actformer: A gan-based transformer towards general action-conditioned 3d human motion generation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 2228-2238 (October 2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.379, + 0.787, + 0.405 + ], + "angle": 0, + "content": "48. Ye, S., Wang, Y., Li, J., Park, D., Liu, C.K., Xu, H., Wu, J.: Scene synthesis from human motion. In: SIGGRAPH Asia 2022 Conference Papers (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.405, + 0.787, + 0.461 + ], + "angle": 0, + "content": "49. Yi, H., Huang, C.H.P., Tripathi, S., Hering, L., Thies, J., Black, M.J.: Mime: Human-aware 3d scene generation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 12965-12976 (June 2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.461, + 0.787, + 0.514 + ], + "angle": 0, + "content": "50. Yi, H., Huang, C.H.P., Tzionas, D., Kocabas, M., Hassan, M., Tang, S., Thies, J., Black, M.J.: Human-aware object placement for visual environment reconstruction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3959-3970 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.514, + 0.787, + 0.57 + ], + "angle": 0, + "content": "51. Zhang, J.Y., Pepose, S., Joo, H., Ramanan, D., Malik, J., Kanazawa, A.: Perceiving 3d human-object spatial arrangements from a single image in the wild. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XII 16. pp. 34-51. Springer (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.569, + 0.787, + 0.624 + ], + "angle": 0, + "content": "52. Zhang, J., Zhang, Y., Cun, X., Zhang, Y., Zhao, H., Lu, H., Shen, X., Shan, Y.: Generating human motion from textual descriptions with discrete representations. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 14730-14740 (June 2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.623, + 0.787, + 0.664 + ], + "angle": 0, + "content": "53. Zhang, M., Cai, Z., Pan, L., Hong, F., Guo, X., Yang, L., Liu, Z.: Motiondiffuse: Text-driven human motion generation with diffusion model. arXiv preprint arXiv:2208.15001 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.664, + 0.787, + 0.718 + ], + "angle": 0, + "content": "54. Zhang, M., Guo, X., Pan, L., Cai, Z., Hong, F., Li, H., Yang, L., Liu, Z.: Remodiffuse: Retrieval-augmented motion diffusion model. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 364-373 (October 2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.718, + 0.787, + 0.759 + ], + "angle": 0, + "content": "55. Zhang, S., Zhang, Y., Ma, Q., Black, M.J., Tang, S.: Place: Proximity learning of articulation and contact in 3d environments. In: 2020 International Conference on 3D Vision (3DV). pp. 642-651. IEEE (2020)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.759, + 0.787, + 0.8 + ], + "angle": 0, + "content": "56. Zhang, X., Bhatnagar, B.L., Starke, S., Guzov, V., Pons-Moll, G.: Couch: Towards controllable human-chair interactions. In: European Conference on Computer Vision. pp. 518-535. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.217, + 0.8, + 0.787, + 0.84 + ], + "angle": 0, + "content": "57. Zhang, Y., Hassan, M., Neumann, H., Black, M.J., Tang, S.: Generating 3d people in scenes without people. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 6194-6204 (2020)" + }, + { + "type": "list", + "bbox": [ + 0.217, + 0.147, + 0.787, + 0.84 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "header", + "bbox": [ + 0.366, + 0.115, + 0.732, + 0.129 + ], + "angle": 0, + "content": "Revisit Human-Scene Interaction via Space Occupancy" + }, + { + "type": "page_number", + "bbox": [ + 0.769, + 0.117, + 0.785, + 0.127 + ], + "angle": 0, + "content": "19" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.189 + ], + "angle": 0, + "content": "58. Zhang, Y., Tang, S.: The wanderings of odysseus in 3d scenes. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 20481-20491 (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.189, + 0.788, + 0.231 + ], + "angle": 0, + "content": "59. Zhao, K., Wang, S., Zhang, Y., Beeler, T., Tang, S.: Compositional human-scene interaction synthesis with semantic control. In: European Conference on Computer Vision. pp. 311-327. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.231, + 0.788, + 0.272 + ], + "angle": 0, + "content": "60. Zhao, K., Zhang, Y., Wang, S., Beeler, T., Tang, S.: Synthesizing diverse human motions in 3d indoor scenes. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 14738-14749 (October 2023)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.272, + 0.788, + 0.314 + ], + "angle": 0, + "content": "61. Zheng, Y., Yang, Y., Mo, K., Li, J., Yu, T., Liu, Y., Liu, C.K., Guibas, L.J.: Gimo: Gaze-informed human motion prediction in context. In: European Conference on Computer Vision. pp. 676-694. Springer (2022)" + }, + { + "type": "ref_text", + "bbox": [ + 0.214, + 0.314, + 0.788, + 0.356 + ], + "angle": 0, + "content": "62. Zhou, Y., Barnes, C., Lu, J., Yang, J., Li, H.: On the continuity of rotation representations in neural networks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5745-5753 (2019)" + }, + { + "type": "list", + "bbox": [ + 0.214, + 0.147, + 0.788, + 0.356 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2024/Revisit Human-Scene Interaction via Space Occupancy/f5977317-ef8e-4afa-8ef9-88b9ecbe0b04_origin.pdf b/2024/Revisit Human-Scene Interaction via Space Occupancy/f5977317-ef8e-4afa-8ef9-88b9ecbe0b04_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..b19e3d89a4a3f13b037da7ae3dede1d4ef26b698 --- /dev/null +++ b/2024/Revisit Human-Scene Interaction via Space Occupancy/f5977317-ef8e-4afa-8ef9-88b9ecbe0b04_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a9f774d836dbb84609f84b192ed356298843cb6e1be47ef25e244bb6b213ce28 +size 9866638 diff --git a/2024/Revisit Human-Scene Interaction via Space Occupancy/full.md b/2024/Revisit Human-Scene Interaction via Space Occupancy/full.md new file mode 100644 index 0000000000000000000000000000000000000000..87284531e135f180b2aff9ecb45fad48ff021c5a --- /dev/null +++ b/2024/Revisit Human-Scene Interaction via Space Occupancy/full.md @@ -0,0 +1,325 @@ +# Revisit Human-Scene Interaction via Space Occupancy + +Xinpeng Liu1, Haowen Hou1,2, Yanchao Yang3, Yong-Lu Li1, and Cewu Lu1 + +1 Shanghai Jiao Tong University + +$^{2}$ Soochow University + +3 The University of Hong Kong + +xinpengliu0907@gmail.com, haowenhou@outlook.com, yanchaoy@hku.hk, {yonglu_li, lucewu}@sjtu.edu.cn + +Abstract. Human-scene Interaction (HSI) generation is a challenging task and crucial for various downstream tasks. However, one of the major obstacles is its limited data scale. High-quality data with simultaneously captured human and 3D environments is hard to acquire, resulting in limited data diversity and complexity. In this work, we argue that interaction with a scene is essentially interacting with the space occupancy of the scene from an abstract physical perspective, leading us to a unified novel view of Human-Occupancy Interaction. By treating pure motion sequences as records of humans interacting with invisible scene occupancy, we can aggregate motion-only data into a large-scale paired human-occupancy interaction database: Motion Occupancy Base (MOB). Thus, the need for costly paired motion-scene datasets with high-quality scene scans can be substantially alleviated. With this new unified view of Human-Occupancy interaction, a single motion controller is proposed to reach the target state given the surrounding occupancy. Once trained on MOB with complex occupancy layout, which is stringent to human movements, the controller could handle cramped scenes and generalize well to general scenes with limited complexity like regular living rooms. With no GT 3D scenes for training, our method can generate realistic and stable HSI motions in diverse scenarios, including both static and dynamic scenes. The project is available at https://foruck.github.io/occu-page/. + +Keywords: Human-Scene Interaction $\cdot$ Motion Generation + +# 1 Introduction + +Human-Scene Interaction (HSI) generation has been an active field recently with its application in animation [36] and embodied AI [35, 38]. For ideal HSI generation, models should be able to control the flexible articulated human body + +![](images/e5f11cd6ab03abe5894adc5853a9ba71a584ea6e0172351860ef03795388e774.jpg) +Fig. 1: We propose that interacting with scenes is essentially interacting with its space occupancy for static HSI. In this view, we can unify motion-only data into a unified human-occupancy knowledge base and train a versatile Human-Occupancy Interaction controller upon it, achieving stable generation under various scenarios. + +to conduct various motions in highly complex 3D environments, considering geometry constraints, motion feasibility, and diversity. In this paper, we focus on interactions without changing the scene, which we refer to as static HSI. + +The limited data acquisition is a major obstacle for static HSI generation. Simultaneous capture of human and 3D scenes is expensive. Some [11,56] collected data only for a narrow range of target objects. Widely adopted PROX [12] contains only 60 clips with limited diversity and complexity. HUMANISE [45] synthesized data by fitting motion from large-scale motion base AMASS [27] into various scenes. CIRCLE [1] put the scenes in VR applications to reduce the cost. Most efforts take explicit scenes as necessary assets. Instead, we dig into the essence of static HSI and revisit it from a space occupancy view. + +Space occupancy is usually identified as an elementary constraint for HSI: humans and scenes cannot occupy the same space simultaneously, which prevents humans from penetrating scenes. We argue that occupancy is inherently the primary factor, especially for static HSI, which is essentially interaction with the scene occupancy. As in Fig. 1, sitting on a chair is practically sitting on a chair-shaped solid being. The chair's occupancy allows us to rest our weight upon it, offering options to sit or stand, i.e., affordances, while its height and shape naturally invite sitting. Other elements like category labels (e.g., identifying it as a "chair" semantically) influence our likelihood of choosing to sit. The scene occupancy, therefore, offers a range of affordances and delineates potential interaction types, establishing an initial probability distribution for interactions. This is primarily accessible to the model and should be learned well at first. + +In light of this, the requirements of realistic scenes for HSI learning could be loosened, inspiring us to revisit large-scale motion datasets. Instead of treating motion-only data as locomotion knowledge bases like existing efforts [58,60], we transform them into a comprehensive Human-Occupancy Interaction database, named Motion Occupancy Base (MOB). For any given motion sequence, we first identify the space occupied by the human at any time in the sequence. The + +remaining space is conceptualized as the "scene". In this framework, the motion becomes a record of interaction with these defined space occupancies. As a result, MOB aggregates enormous paired human-occupancy data from existing large-scale motion-only data. Compared to current datasets [1, 12, 45], MOB presents over-complex and stringent geometric constraints, pushing the model to the limit in interacting with space occupancy. + +How do we exploit the aggregated MOB for static HSI? Previous approaches generally provided two different answers. Some [1,16,45,61] adopt an end-to-end pipeline from scene to motion, which is straightforward but inflexible without the ability to interact with dynamic scenes. Some [43, 44, 60] typically decouple HSI into interaction and locomotion, adopting a two-stage paradigm: first generating key interaction frames and then synthesizing locomotion between them, with the help of methods like path-finding algorithms. However, this decomposition overlooks the integrated nature of locomotion and interaction, like simultaneous arm movement and walking to navigate obstacles. Our approach, rooted in space occupancy, reconceptualizes locomotion and interaction as a unified Human-Occupancy Interaction. We develop an auto-regressive HumanOccupancy Interaction controller for motion generation. The controller processes various inputs, including surrounding occupancy and target locations, as control signals at each time step, producing future motion predictions. Additionally, we introduce a field regulation module for enhanced collision avoidance. There are three major advantages of our controller. First, by learning from the challenging cases in MOB, it could handle complex scenes but also generalize well to regular simpler scenarios. Second, its versatility accommodates any static HSI without the need to switch between models for locomotion and interactions, even for HOI motion synthesis. Third, its auto-regressive nature allows it to adapt to dynamic environments such as an automatic door, a capability lacked in previous methods. In extensive experiments, our controller performs effectively across various settings, even without training on real 3D scene data. + +Overall, our contributions are three-fold: 1) We exhume Human-Occupancy Interaction as a major component of static HSI. 2) By digging out the inherent Human-Occupancy relationships of pure motions, we convert motion-only data into a paired Human-Occupancy database MOB, expanding the data scale. 3) We develop a versatile controller for flexible and stable HSI motion generation. + +# 2 Related Works + +Human-Scene Interaction Generation. Generating natural human motion sequences has received increasing attention with the emerging MoCap datasets [4, 7, 18, 24, 27, 33]. Early efforts were paid to prefix/suffix-conditioned motion generation [2, 9, 10, 15], known as motion prediction and interpolation. There also have been efforts to pursue generating motion with respect to specific action semantics, either represented in action categories [8, 30, 47] or natural language [6, 7, 26, 31, 32, 41, 42, 52–54]. Audio-conditioned motion generation was also explored, like music for dance generation [21, 22]. However, these approaches tend + +to focus on human motion only, without placing the generated motion in a grounded 3D environment. Recently, progress has been made to advance HSI generation. Early efforts [23,51] first reconstructed human-object interaction in 3D. Some [13,55,57] proceeded by placing SMPL-X [29] meshes in scanned 3D scenes. Semantic controls were also introduced [59]. For HSI motion generation, some [11,36,56] focused on interacting with a single target object, mostly chairs. GOAL [39] and SAGA [46] generated full-body pre-grasp motions. There have also been efforts [14,40] for the physical control of specific interactions. With paired motion-scene datasets [1,12], approaches that synthesize HSI motion in cluttered scenes appeared. Two-stage based algorithms [43,44] were proposed to first generate static key-frame poses, and then interpolate them for a coherent motion. GAMMA [58] proposed an RL framework for locomotion on flat grounds with the help of off-the-shelf path-finding algorithms. DIMOS [60] extended GAMMA with extra interaction policies. In contrast, the follow-ups [1,16,45,61] adopted a simple end-to-end one-stage paradigm, while focusing more on boosting data. GIMO [61] collected data with gaze information. HUMANISE [45] synthesized pseudo-data from large-scale motion bases and 3D scan bases. CIRCLE [1] collected data with a VR application. LAMA [19] coupled the RL framework with a motion matching algorithm, generated HSI motions with only a unified policy, and leveraged motion editing for optimization. Our major difference from previous efforts is the Human-Occupancy Interaction view, which alleviates the data hunger, and also enables a unified controller. Moreover, we are the first to generate motion in a dynamic scene, which previous efforts failed to handle. + +Scene Creation from Motion. Pose2Room [28] predicted 3D object bounding boxes from human motion. MOVER [50] reconstructed 3D scenes from videos that record human-scene interaction. SUMMON [48] and MIME [49] placed 3D object models concerning human motion. These methods could be adopted to boost HSI data, though usually limited to a small range of indoor furniture. + +# 3 Motion Occupancy Base + +HSI generation could be formulated as $M = \mathcal{G}(S)$ , where $M$ is the generated motion, $\mathcal{G}$ is the motion generator, and $S$ is a 3D scene. Conventionally, the scene is expected to be as realistic as possible, requiring expensive acquisition. We find that these scenes could function almost equally to their occupancy for static HSI. Thus, high-quality scenes might become less necessary here. Given this, we revisit the large-scale motion-only datasets. Intuitively, motion only contains body movement information. But if we switch our sight to the vacant space around the moving person, considering the space as another form of occupancy, we could find that these motions possess information on how humans interact with certain space occupancies. To this end, we aggregate motion-only datasets into the human-occupancy interaction database: Motion Occupancy Base (MOB). + +The overall pipeline is shown in Fig. 2. Motion with SMPL(-X) [25,29] representation are collected, re-sampled to 30 FPS, and rotated so that the gravity direction points to the negative z-axis. For motion sequence $X = \{x^i\}_{i=1}^t$ , where + +![](images/c49f40fd5050f99401eb9c13f41457627a671cb695e7cc0f89bc1cece52ed7a0.jpg) +Fig. 3: The architecture of our versatile motion controller. Given motion state and histories $x^{t}, h^{t}$ , and control signals $c_{o}^{t}, c_{t}^{t}$ , the controller auto-regressively generates the next motion frame w.r.t. the target pose and the canonical occupancy. + +$x_{i}$ is the SMPL(-X) representation in the $i$ -th frame, $t$ is the number of frames, we first obtain mesh sequence $X_{m} = \{x_{m}^{i}\}_{i = 1}^{t}$ via SMPL(-X) model, where $x_{m}^{i} = (f^{i},v^{i})$ is the mesh for the $i$ -th frame, $f^{i},v^{i}$ are the mesh faces and surface + +vertices. A bounding box of $X_{m}$ is obtained based on vertices $V = \{v_{i}\}_{i=1}^{t}$ . Given the predefined grid unit size $u$ , the box region is voxelized into motion occupancy volume $\hat{O} \in \{0,1\}^{[H \times W \times D]}$ , computed in two steps: 1) $V$ occupy the corresponding voxels in $\hat{O}$ . 2) Given the centers of the un-occupied voxels as $C = \{c_{i}\}_{i=1}^{|C|}$ , if $c_{i}$ is inside any human mesh $x_{m}^{i}$ , the corresponding voxel is occupied. The pseudo-scene occupancy $O$ is then obtained by negating the motion occupancy, formulated as $O = 1 - \hat{O}$ . Thus, we convert motion-only sample $X$ into a motion-occupancy pair $(X,O)$ in MOB. MOB + +![](images/9f04fe83a768594cd8978f1701a9470940a2412e939bee049ba6b5b61721b096.jpg) +Fig. 2: The construction process of Motion Occupancy Base. Note that the ceilings of the occupancy are hidden for clarity. + +aggregates 13 datasets with 98k instances. More details are in the supplementary. + +As shown in Fig. 2, humans in MOB are acting with rather strict geometric constraints, which is uncommon in regular scenes covered by existing datasets [1, 12, 45]. However, we show that by learning from hard cases in MOB, our model could grow its ability to handle both complex and regular scenarios well. + +# 4 Versatile Motion Controller + +# 4.1 Preliminaries + +To instantiate $M = \mathcal{G}(S)$ for HSI generation, previous efforts either adopt a one-shot transformer generator or decompose $\mathcal{G}$ into two separate generators, one for + +interaction key-frame generation, the other for key-frame interpolation. Instead, we formulate our generator as an auto-regressive controller. At time step $t$ , given historical motion state $h^t$ , current pose state $x^t$ , we expect a controller as + +$$ +\hat {y} ^ {t} = \mathbf {G} \left(h ^ {t}, x ^ {t}, c ^ {t}\right), \tag {1} +$$ + +where $\hat{y}^t$ is the predicted future motion, and $c^t$ is the control signals. An overview is in Fig. 3. Sec. 4.2 introduces motion state representation $h^t, x^t, y^t$ . The control signals are covered in Sec. 4.3. The controller architecture is detailed in Sec. 4.4. Sec. 4.5 demonstrates the training and inference procedure. + +# 4.2 Motion State Representation + +Notations. In the global coordinate, we represent motion sequence $X$ with $T$ frames as $X = \{x^{t}\}_{t = 1}^{T}$ , where $x^{t} = (p_{r}^{t},\theta_{r}^{t},r^{t})$ is SMPL [25] parameters, with root location $p_r^t\in \mathcal{R}^3$ , root orientation $\theta_r^t\in \mathcal{R}^6$ , and the orientation of $j$ limbs as $\theta^t\in \mathcal{R}^{j\times 6}$ . Based on this, the location of $j$ joints $p^t\in \mathcal{R}^{j\times 3}$ could also be computed via SMPL. For simplicity, we use $\cdot^{-w}$ to represent the value from frame $t - w$ to $t$ , and $\cdot^{+w}$ for $t$ to $t + w$ . The 6D form from [62] is adopted for all rotation and orientation representations. + +Current Pose State could be formulated as + +$$ +x ^ {t} = \operatorname {C a n o} \left(\left\{p _ {r} ^ {t}, p ^ {t}, \theta_ {r} ^ {t}, \theta^ {t}, \dot {p} _ {r} ^ {t}, \dot {p} ^ {t}, \dot {\theta} _ {r} ^ {t} \right\}, t\right), \tag {2} +$$ + +where $\cdot^t$ indicates the first derivative at $t$ . Cano $(\cdot, t)$ canonicalizes the input w.r.t. the root location and facing direction at $t$ , ensuring consistency across frames. + +Historical Motion State $h^t$ captures motion from frame $t - w$ to $t$ as + +$$ +h ^ {t} = \operatorname {C a n o} \left(\left\{p _ {r} ^ {- w}, p ^ {- w}, \theta_ {r} ^ {- w}, \dot {p} _ {r} ^ {- w}, \dot {p} ^ {- w}, \dot {\theta} _ {r} ^ {- w} \right\}, t\right), \tag {3} +$$ + +with trajectories $p_r^{-w}, p^{-w}$ , root orientation $\theta_r^{-w}$ , velocities $\dot{p}_r^{-w}, \dot{p}^{-w}$ , and root angular velocity $\dot{\theta}_r^{-w}$ . For conciseness, if not specified, only the $p^{-w}$ of the hands and feet is adopted in practice. + +Future Prediction $\hat{y}^t$ is composed of two parts as $\hat{y}^t = \{\hat{x}^{t + 1},\hat{u}^t\}$ . Note that $\hat{x}^{t + 1}$ is in the egocentric canonical of frame $t$ , and $\hat{p}_r,\hat{\dot{\theta}}\subset \hat{x}^{t + 1}$ are adopted to transit from $t$ to $t + 1$ . Mirroring the historical motion state $h^t$ , we also incorporate a comprehensive set of future trajectories, facing directions and their changing rates as $\hat{u}^t = \{p_r^{+f},p^{+f},\theta^{+f},\dot{p}_r^{+f},\dot{p}^{+f},\dot{\theta}^{+f}\}$ for a window size of $f$ , all in the canonical view of $t + 1$ . Empirically, integrating $\hat{u}^t$ significantly enhances the model's precision in navigating toward the target. + +# 4.3 Control Signals + +In this section, we introduce how the different control signals are encoded. Note that our models can operate without these optional signals. + +Canonical Occupancy. A simple voxel-based representation is adopted to encode how the canonical space is occupied by the scenes. We utilize a cubic + +point grid $o^t \in \{0,1\}^{s \times s \times s}$ characterized by the grid size $s$ and the grid unit size $u$ at frame $t$ . The grid is positioned in the human's canonical space, with its center offset by $s/4$ in the human-facing direction. Each cell contains a binary value, indicating whether the cell is intersecting with the global scene occupancy $O$ . The canonical occupancy is thus represented as $c_o^t = \text{Flatten}(o^t) \in \{0,1\}^{s^3}$ . Note that because the occupancy is calculated in real-time, this control signal can naturally adapt to dynamic scenes where each timestep is treated statically. We will show in Section 5 that even though our training data is all static, our model exhibits a smooth generalization to dynamic scenes. + +Target Pose is also considered a flexible control signal, which can be given in two ways. The first is the positions of the root, hands, and feet, each defined in relation to the canonical space of the current pose, represented as $c_t^t \in \mathcal{R}^{5 \times 3}$ . In practice, these could be assigned manually or generated from off-the-shelf models [59]. Also, partial or blank targets are acceptable. The second is the future trajectory of end-effectors, denoted as $c_t^{+f}$ for a future window size of $f$ . This paves the way for different applications in Section 5. + +# 4.4 Controller Architecture + +We adopt a lightweight transformer-based architecture. $n$ fully connected layers first project the $n$ input features from $\{h^t,x^t,c^t\}$ to the latent $z\in \mathcal{R}^{n\times d}$ . $z$ is then fed into encoder $E$ to distill dynamic information from the motion states and control signals. We devise $E$ as a sequence of Transformer encoders, which takes $z$ as input and outputs refined dynamic information $\hat{z} = E(z)\in \mathcal{R}^{n\times d}$ . Finally, we adopt a transformer decoder $D$ followed by a linear layer to condense the dynamic information, resulting in the final predictions. With learnable query $q\in \mathcal{R}^d$ and the dynamic information $\hat{z}$ serving as K and V, respectively, the output of the transformer is further concatenated with a random noise $\epsilon \in \mathcal{R}^d$ for randomness. The linear layer then outputs the final prediction as $\hat{y}^t = D(\hat{z},q,\epsilon)$ . + +Furthermore, to enhance collision avoidance, we introduce a novel module called Field Regulation. The basic idea is that humans tend to move faster in obstacle-sparse areas and decelerate in complex environments or when an obstacle is nearby. At very short distances, especially with an imminent collision happening, a complete halt is necessary. Given this and taking inspiration from the potential field in robot motion planning [17], we propose Occupancy Field, which exerts negligible influence on humans at a distance but + +escalates to a sturdy repulsion as proximity increases, thereby compelling the + +![](images/30e2668dbbe56429deaffa35f1c05c951373bdb753827e069578e8503f4ffb3c.jpg) +Fig.4: Occupied voxel centers $\{P_i\}$ are located at the top left of the figure. The initial velocity vector, $\dot{p}_j$ , is directed upwards, posing a risk of collision. To mitigate this, the occupancy field introduces a corrective velocity component $\Delta \dot{p}_j$ . This redirects the final velocity $\dot{p}_j^{out}$ to avoid collision. + +human to decelerate and avoid collision. The occupancy field influences human motion only when a human is in motion toward it. The field functions as + +$$ +\Delta \dot {p} _ {j} ^ {t + 1} = \mathcal {F} \left(p _ {j} ^ {t + 1}, c _ {o} ^ {t}\right). \tag {4} +$$ + +The field $\mathcal{F}$ calculates a correction $\Delta \dot{p}^{t + 1}$ w.r.t. predicted joint velocity $\dot{p}_j^{t + 1} \in \{\dot{p}_r^{t + 1}, \dot{p}^{t + 1}\}$ and canonical occupancy $c_o^t$ , which is then added to $\dot{p}^{t + 1}$ . For simplicity, the superscripts indicating the time steps are omitted in the following equations. In detail, given the center points of all the occupied voxels in $c_o$ as $P \subset \mathcal{R}^{n_p \times 3}$ , the correction is calculated as + +$$ +\Delta \dot {p} _ {j} = \sum_ {P _ {i} \in P} - k r _ {i} \max \left(0, \frac {1}{\| v _ {i j} \| _ {a} - \gamma} - b\right), \tag {5} +$$ + +$$ +r _ {i} = \dot {p} _ {j} \max (0, \langle \dot {p} _ {j}, v _ {i j} \rangle), \tag {6} +$$ + +where $v_{ij} = P_i - p_j$ is the vector from the joint $j$ to the voxel center, $\langle \cdot, \cdot \rangle$ indicates the cosine between the two vectors, $\| \|_a$ denotes $a$ -norm, and $k$ is the stiffness factor. $\gamma$ and $b$ are thresholds to constrain the distance range of effective voxels, where $\gamma$ reflects the human body's width, and voxels within this range are discounted for stability. An illustration of Field Regulation is shown in Fig. 4. The field regulation is incorporated in the training of the controller to adapt to it, avoiding potential foot sliding caused due to direct modification on velocities. + +# 4.5 Training and Inference + +Training Objectives The controller is trained in a fully-supervised manner. We employ a mixture of L1 and L2 losses as basic objectives. Besides, penetration loss and occupancy field loss are designed to hinder undesired collisions. Specifically, L1 loss is utilized for rotation-related output values $\hat{y}_{\theta}^{t}\in \hat{y}^{t}$ , while L2 loss is applied to the rest outputs $\hat{y}_p^t$ . The mixture loss is formulated as + +$$ +\mathcal {L} _ {m i x} = \left\| y _ {\theta} ^ {t} - \hat {y} _ {\theta} ^ {t} \right\| _ {1} + \left\| y _ {p} ^ {t} - \hat {y} _ {p} ^ {t} \right\| _ {2}. \tag {7} +$$ + +Given the predicted joint location $\hat{p}^{t + 1}$ , penetration loss is computed as + +$$ +\mathcal {L} _ {p e n} = \sum_ {i = 0} ^ {j} O \left(\hat {p} _ {j} ^ {t + 1}\right) \| r e f \left(\hat {p} _ {j} ^ {t + 1}, O\right) \| _ {2}, \tag {8} +$$ + +where $O(\cdot)$ returns 1 if the point is occupied in $O$ else 0, and $ref(\hat{p}_j^{t + 1},O)$ calculates the closest unoccupied voxel center in $O$ for joint $\hat{p}_j^{t + 1}$ . Occupancy field loss has two components. First, to deter undesired interactions, we penalize the proportion of the occupancy field effect relative to the output root velocity. Second, observing occasional unnatural spikes in velocity, where the human figure seemingly attempts to navigate through an obstacle, we punish excessive velocities. The occupancy field loss is formulated as + +$$ +\mathcal {L} _ {\text {f i e l d}} = \left(\frac {| \Delta \dot {p} |}{| \dot {p} |}\right) ^ {2} + | \dot {p} | ^ {2}. \tag {9} +$$ + +
MethodSuc. (%) ↑DT (cm) ↓Time (s)FS (%)PEN ↓ERP
GT100.000.003.003.360.00-
NSM [36]11.7632.795.3311.04134.497.26
SAMP [11]13.9452.496.549.87145.746.98
DIMOS [60]15.7348.976.8210.52123.126.82
Ours w/o Occupancy71.2214.662.256.8946.054.13
Ours77.9115.303.628.1315.214.24
Ours w/o Lpen74.1414.893.547.8931.934.96
Ours w/o Lfield75.3415.692.8611.6525.835.70
Ours w/o FR77.3814.852.346.5733.494.18
Ours w/ BPS73.6416.833.359.3436.724.72
+ +Table 1: Quantitative results on MOB. + +The final loss is computed with loss coefficients $\alpha, \beta$ as + +$$ +\mathcal {L} = \mathcal {L} _ {m i x} + \alpha \mathcal {L} _ {p e n} + \beta \mathcal {L} _ {\text {f i e l d}}. \tag {10} +$$ + +Auto-regressive Inference. We perform motion generation in an auto-regressive manner, enabling the model to adapt to dynamic environmental changes. Since the model solely manipulates canonical information, we track the global root position $p_{g}$ and the facing direction $\theta_{g}$ of the human in the world coordinate system. Initially, we set past and current velocities $\dot{p}^{0}$ to zero. Before each prediction at frame $t$ , we compute the control signals (if there are any), i.e., the relative target position $c_{t}^{t}$ and the canonical space occupancy $c_{o}^{t}$ , based on the current values of $p_{g}^{t}$ and $\theta_{g}^{t}$ . After the prediction, we update $p_{g}^{t}$ and $\theta_{g}^{t}$ with the model's output, along with the current pose state $c^{t + 1}$ . Concurrently, we refresh the historical motion state $h^{t + 1}$ to prepare for the next input. + +# 5 Experiments + +Implementation Details. The controller consists of a two-layer transformer encoder and a single-layer transformer decoder, both with a latent dimension of 512 and 8 heads. The canonical occupancy grid size is set as (25, 25, 25), with a unit size of $8\mathrm{cm}$ . The controller operates at 10 FPS with the history and future window sizes both set as 1, which empirically brings more stable outcomes. The controller is trained on the MOB, with 1,393 sequences excluded from training as the test set. We train the controller for 75k iterations, with a cosine learning rate decay restart strategy. The initial learning rate is set as 1e-4. The control signals are randomly masked for augmentation during training. Scheduled Sampling [3] is adopted. The whole training takes approximately 6 hours on a single 12G NVIDIA Titan Xp. More details are in the supplementary. + +# 5.1 Interacting with Occupancy from MOB + +We first evaluate the ability to interact with occupancy volumes with a goal-reaching task on MOB. Given an occupancy volume, a starting pose, and a target end-effector pose, the model is required to generate transition motion from the starting pose to the target pose. We split the test set into 30k 3-second + +![](images/e9616a1ae9e7663d319be7e5a38fcc42cffcfdf67a603753a4251b8e942e315f.jpg) +Fig. 5: Our controller can naturally reach the target (white to blue) given the complex occupancy. Furthermore, it can stabilize the motions (red to white) around the targets after reaching them. Previous SOTA [60] suffers from severe penetration and fails to reach the target. For simplicity, the ceilings of the occupancy are hidden. + +clips, adopting the first frame as the starting pose and the final frame as the target pose. The models are required to generate motion for 10s towards the target. There are two reasons for this setting. First, we would like to provide enough time for the models to accomplish the tasks. Second, we are also curious about how the models would behave after the target is reached and no new target is given, or when a very close target is given. The results are evaluated from three aspects. Task Completion is determined by whether the target pose is reached. The target is identified as completed if the average Euclidean distance between the generated and target pose is lower than 20 centimeters and less than 50 voxels are penetrated. We report the success rate (Suc.). The average minimum distance to the target pose (DT) and the time taken to reach the target (Time) are reported for reference. Note that Time is only calculated for successful samples. Motion Quality is evaluated via foot sliding (FS) and penetration (PEN) detection. A frame is identified as FS with its foot velocity $>7.5\mathrm{cm / s}$ following GAMMA [58]. We report the percentage of frames with FS. For PEN, we report the number of voxels penetrated per frame. Similarity to Dataset is measured by the similarity between the predicted and GT target end-effector trajectory. We report Edit distance for the real penalty (ERP) [5] since the trajectories might have different lengths. It is only given as a reference, which is not the lower the better. + +Quantitative results are shown in Tab. 1. For DIMOS [60], we convert the target pose into corresponding marker sets and report the best results from different policies for each test sample. For w/o Occ., we remove the canonical + +![](images/87245a5bd86dad665600291e9b5485cc67104ca506032de232652b4493d7de8b.jpg) +Space Occupancy overlay DIMOS Walkable Area from NavMesh + +![](images/d334ea9aa2a44a4d217eb82acbab90ea5f056d02f16ea15cb93a3c936e47bd0a.jpg) +DIMOS path planning results +Sit Down + +![](images/5e84eab1357ea931c3eeaa4888955163dffd3b6d6a527245a9c42e09ed722657.jpg) +Fig. 6: A typical DIMOS [60] failure case. Fig. 7: With the unified view of human-The NavMesh algorithm produces sub- occupancy interaction, the same controller optimal walkable areas, resulting in path- in Fig. 5 produces diverse motions, includ-finding failures for $69.43\%$ MOB samples. ing sitting down, crawling, and getting up. + +occupancy-related components. Previous efforts frequently fail to reach the goal successfully in complex scenarios, which we will elucidate later with qualitative results. Even when the goal is reached, severe penetration emerges, resulting in low success rates in Tab. 1. Our models outperform DIMOS by a considerable margin with a single controller compared to multiple policies. This reveals that the straightforward voxel-based representation is capable of encoding holistic information for human-occupancy interaction. The model without occupancy manages to approach the target as closely and quickly as possible. However, it suffers from more severe interpenetration with the occupancy, resulting in a lower success rate. Instead, our model with consideration of occupancy behaves more conservatively, spending more time to accomplish the tasks with less penetration. A slightly higher foot sliding rate is also observed, which is a trade-off for the penetration-avoiding ability that field regulation brings. + +Qualitative results are shown in Fig. 5, 7. Given the complex occupancy, previous state-of-the-art DIMOS [60] frequently fails to reach the goal. Even when it reaches the goal, severe penetration is usually observed. Instead, our controller generates feasible motions to reach the targets. Previous methods' typical locomotion-interaction decomposition [44, 58, 60] tends to over-simplify both. For locomotion, most methods assume a rigid character (usually simplified as a rigid cylinder in $\mathrm{A}^*$ [11, 44] and NavMesh algorithms [60]), overseeing the flexibility of humans. It might be okay for spacious rooms, however, when posed in a cramped scene or scene with complex geometric structures like in Fig. 5, models could fail in path-finding and thus degenerate. As shown in Fig. 6, for $69.43\%$ MOB samples, the NavMesh algorithm adopted by DIMOS failed to generate a feasible path from the start position to the target position. The statistic for $\mathrm{A}^*$ from SAMP [11] is $70.29\%$ . This explains the unsatisfying quantitative results of DIMOS. For interaction, the categories are typically limited to predefined ones. In contrast, with our unified view of Human-Occupancy interaction, our controller applies to various complex scenarios beyond simplified locomotion and interaction. Furthermore, diverse motions could be performed by the same controller (Fig. 7). By comparing our controllers with and without occupancy, we find that training with occupancy helps to avoid penetration significantly. + +
ModelSuc. (%)Dist. to goal (cm)Collision depth (cm)FS (%)
GT100.000.008.742.45
Ours (w/ MOB only)73.017.199.5613.83
Ours (w/ CIRCLE only)75.526.8813.7411.76
Ours (w/ MOB + CIRCLE)78.976.7311.8512.87
+ +Table 2: Results on CIRCLE [1]. + +![](images/ad6c5920472bc89dea76c37b86bcb6c6374941503bec0779fe3945dca69b5970.jpg) +Fig. 8: We can generate long-term motion. Fig. 9: We synthesize Human-Object Interactions in realistic rooms: A human circulation action, with hand trajectory as the control navigates a cabinet, sits on a sofa, perches signal. The human successfully moves a floor on a bed, and then resumes walking. lamp to a new place. + +![](images/fc65b660786abf47cdc3ce3fdb42a77038104bc6c1a9346444a5bf7a395ba770.jpg) + +Also, the controller with occupancy could stabilize the motion around the target after reaching it, while its counterpart tends to wander further inertially. + +# 5.2 Interacting with Static Scenes + +Once our controller is trained on the more complex MOB, it could be effortlessly applied to realistic synthesized rooms and scene scans. + +Quantitative results is reported on CIRCLE [1]. Given the initial full-body pose and the target right wrist position, the model is required to generate a right-hand reaching motion. Following [1], success rate, distance to the goal, cumulative collision depth, and foot sliding are reported as metrics. Besides the controller trained on MOB (excluding CIRCLE), we also train our controller on CIRCLE-only and MOB+CIRCLE. We randomly split CIRCLE into a 2,565-sample train set (for the ablative baseline only) and a 453-sample test set since the original split has not been made public yet. The results in Tab. 2 reveal that our controller could handle various tasks in complex natural scenes well. Also, its potential to advance HOI-related motion, like hand-reaching before grasping, is showcased. It is noticeable that the MOB-only model demonstrates a competitive success rate and better collision avoidance, revealing the efficacy of MOB with its advanced complexity. When incorporating MOB with CIRCLE, superior task completion is observed, showing that MOB could function as an effective auxiliary resource for the conventional HSI datasets. + +Qualitative results. We test our model in a synthetic room from DIMOS [60] in Fig. 8 and a realistic scan from Replica [37] in Fig. 1. The occupancy is transferred from the scene SDFs. Our model produces convincing motions, nav- + +![](images/095890d1c8cf55261fd0b148674fa8869e946b23356fefdb98c2c43f721a70c0.jpg) +Fig. 10: Our model can avoid dynamic obstacles and reasonably reach the target circle. + +igates around obstacles, and transitions seamlessly between locomotion and interaction. Remarkably, it adapts to sitting on various objects, such as sofas, beds, and stools, revealing its versatility in recognizing different occupancies as potential seats. The test room features sparse furnishings, highlighting the model's ability to operate in open spaces diverging from MOB samples. Conversely, the Replica scene presents a more cluttered setup. Here, the model adeptly maneuvers through tight spaces and produces plausible movements. Furthermore, our model can generate long-term motions. The motions in Fig. 1 and Fig. 8 are generated in a single attempt by merely altering the target, both spanning around 30 seconds, which are clipped from longer motions for clear illustration in the limited space. More results are in the supplementary materials. + +# 5.3 Interacting with Objects + +Our flexible control signal for target pose enables HOI motion synthesis. By fine-tuning on OMOMO [20], given the initial pose and hand trajectories, we synthesize plausible motions to interact with objects as in Figure 9, which exhibits rational full-body motion w.r.t. hand and object trajectories. More results are in the supplementary. + +# 5.4 Interacting with Dynamic Scenes + +During training, the canonical occupancy perceived by our controller is dynamically changing w.r.t. to the human movement, enabling it to handle dynamic scenarios though trained on essentially static MOB. We demonstrate the interaction in dynamic scenes with a custom automatic door scenario in Fig. 10. With the target (the circle) in front of the human (a) blocked by a revolving door, we control the human to adapt its speed w.r.t. the door till enough clearance and halt precisely at the target. This sequence exemplifies the model's ability to produce realistic and contextually appropriate motion in dynamic environments. More details are in the supplementary. + +# 5.5 Ablation Studies + +With our controller as the baseline, selected ablation studies on MOB are reported in Tab. 1. Complete results and details are included in the supplementary. + +Different loss terms. By removing penetration loss $L_{pen}$ , a major degeneration in Success rate and PEN is observed. This shows that explicitly punishing penetration is necessary. The removal of $L_{field}$ also hurts the success rate by $2.57\%$ , with PEN increasing by 10.62. FS also increases notably. We identify that without $L_{field}$ , the field regulation would generate unnatural spikes. + +Field Regulation (FR) is also evaluated. Eliminating the field regulation decreases the success rate slightly, with a notable PEN increase (+18.28) and FS decrease (-1.56%), revealing the tradeoff it conducts between PEN and FS. + +Occupancy representation. We replaced the voxel-based representation of canonical occupancy with the more compact and efficient BPS [34]. The field regulation is also modified accordingly. Degradation is observed in most metrics. Despite its efficiency, BPS might ignore too much information and bring instability to the field regulation. + +# 6 Discussion + +Limitations. First, when interacting with dynamic scenes, our controller could perform sub-optimally with slight collisions. This could be ascribed to the space occupancy changing rate gap between training data and testing scenes. During training, the changing rate is determined by the human locomotion speed only. While for dynamic scenes, there exists additional space occupancy speed. Mitigating this gap is worth exploring. Second, foot sliding is another artifact that could happen. However, the metrics are based on simple heuristics with gaps compared to real foot sliding. A reasonable and effective tool to address this would be promising. Third, the current MOB poses its main focus on static HSI. Despite the proposed pipeline that could apply to HOI motion synthesis, further exploration of the space-occupancy view of HOI would be an interesting future direction. Finally, the proposed versatile controller only accepts target poses, space occupancy, and joint trajectories as control signals. Extending the controller for more signals would be promising. + +Broader Impacts. The exploration of HSI is beneficial for navigation in complex environments, with significant impacts on Embodied AI and Robotics. Also, multiple fields like animation production and interior design could be advanced. However, this might also be misused for false information. + +# 7 Conclusion + +We revisited static HSI from a unified space occupancy view by identifying the major component of static HSI as Human-Occupancy Interaction. Large-scale Human Occupancy Interaction database MOB was aggregated from motion-only data, substantially improving diversity and complexity. Moreover, a versatile auto-regressive controller was trained on MOB for Human-Occupancy Interaction generation, which supports flexible and stable generation for static HSI under a spectrum of scenarios without training with real 3D scene data. + +# Acknowledgments + +This work is supported in part by the National Natural Science Foundation of China under Grants No.62306175, No.62302296, National Key Research and Development Project of China (No.2022ZD0160102, No.2021ZD0110704), Shanghai Artificial Intelligence Laboratory, XPLORER PRIZE grants. + +# References + +1. Araújo, J.P., Li, J., Vetrivel, K., Agarwal, R., Wu, J., Gopinath, D., Clegg, A.W., Liu, K.: Circle: Capture in rich contextual environments. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 21211-21221 (2023) +2. Athanasiou, N., Petrovich, M., Black, M.J., Varol, G.: Teach: Temporal action composition for 3d humans. In: 2022 International Conference on 3D Vision (3DV). pp. 414-423. IEEE (2022) +3. Bengio, S., Vinyals, O., Jaitly, N., Shazeer, N.: Scheduled sampling for sequence prediction with recurrent neural networks. Advances in neural information processing systems 28 (2015) +4. Cai, Z., Ren, D., Zeng, A., Lin, Z., Yu, T., Wang, W., Fan, X., Gao, Y., Yu, Y., Pan, L., Hong, F., Zhang, M., Loy, C.C., Yang, L., Liu, Z.: Humman: Multi-modal 4d human dataset for versatile sensing and modeling. In: Avidan, S., Brostow, G., Cisse, M., Farinella, G.M., Hassner, T. (eds.) Computer Vision - ECCV 2022. pp. 557-577. Springer Nature Switzerland, Cham (2022) +5. Chen, L., Ng, R.: On the marriage of lp-norms and edit distance. In: Proceedings of the Thirtieth international conference on Very large data bases-Volume 30. pp. 792-803 (2004) +6. Chen, X., Jiang, B., Liu, W., Huang, Z., Fu, B., Chen, T., Yu, G.: Executing your commands via motion diffusion in latent space. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18000-18010 (2023) +7. Guo, C., Zou, S., Zuo, X., Wang, S., Ji, W., Li, X., Cheng, L.: Generating diverse and natural 3d human motions from text. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 5152-5161 (June 2022) +8. Guo, C., Zuo, X., Wang, S., Zou, S., Sun, Q., Deng, A., Gong, M., Cheng, L.: Action2motion: Conditioned generation of 3d human motions. In: Proceedings of the 28th ACM International Conference on Multimedia. pp. 2021-2029 (2020) +9. Guo, W., Du, Y., Shen, X., Lepetit, V., Alameda-Pineda, X., Moreno-Noguer, F.: Back to mlp: A simple baseline for human motion prediction. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 4809-4819 (2023) +0. Harvey, F.G., Yurick, M., Nowrouzezahrai, D., Pal, C.: Robust motion in-between. ACM Transactions on Graphics (TOG) 39(4), 60-1 (2020) +1. Hassan, M., Ceylan, D., Villegas, R., Saito, J., Yang, J., Zhou, Y., Black, M.J.: Stochastic scene-aware motion prediction. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 11374-11384 (2021) +2. Hassan, M., Choutas, V., Tzionas, D., Black, M.J.: Resolving 3D human pose ambiguities with 3D scene constraints. In: International Conference on Computer Vision. pp. 2282-2292 (Oct 2019), https://prox.is.tue.mpg.de + +13. Hassan, M., Ghosh, P., Tesch, J., Tzionas, D., Black, M.J.: Populating 3d scenes by learning human-scene interaction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14708-14718 (2021) +14. Hassan, M., Guo, Y., Wang, T., Black, M., Fidler, S., Peng, X.B.: Synthesizing physical character-scene interactions. In: ACM SIGGRAPH 2023 Conference Proceedings. SIGGRAPH '23, Association for Computing Machinery, New York, NY, USA (2023). https://doi.org/10.1145/3588432.3591525, https://doi.org/10.1145/3588432.3591525 +15. Hernandez, A., Gall, J., Moreno-Noguer, F.: Human motion prediction via spatiotemporal inpainting. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 7134-7143 (2019) +16. Huang, S., Wang, Z., Li, P., Jia, B., Liu, T., Zhu, Y., Liang, W., Zhu, S.C.: Diffusion-based generation, optimization, and planning in 3d scenes. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16750-16761 (2023) +17. Hwang, Y.K., Ahuja, N., et al.: A potential field approach to path planning. IEEE transactions on robotics and automation 8(1), 23-32 (1992) +18. Ji, Y., Xu, F., Yang, Y., Shen, F., Shen, H.T., Zheng, W.S.: A large-scale rgb-d database for arbitrary-view human action recognition. In: Proceedings of the 26th ACM international Conference on Multimedia. pp. 1510-1518 (2018) +19. Lee, J., Joo, H.: Locomotion-action-manipulation: Synthesizing human-scene interactions in complex 3d environments. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 9663-9674 (October 2023) +20. Li, J., Wu, J., Liu, C.K.: Object motion guided human motion synthesis. ACM Transactions on Graphics (TOG) 42(6), 1-11 (2023) +21. Li, R., Yang, S., Ross, D.A., Kanazawa, A.: Ai choreographer: Music conditioned 3d dance generation with aist++. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 13401-13412 (2021) +22. Li, R., Yang, S., Ross, D.A., Kanazawa, A.: Learn to dance with aist++: Music conditioned 3d dance generation. arXiv preprint arXiv:2101.08779 2(3) (2021) +23. Li, Y.L., Liu, X., Lu, H., Wang, S., Liu, J., Li, J., Lu, C.: Detailed 2d-3d joint representation for human-object interaction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10166–10175 (2020) +24. Lin, J., Zeng, A., Lu, S., Cai, Y., Zhang, R., Wang, H., Zhang, L.: Motion-x: A large-scale 3d expressive whole-body human motion dataset. arXiv preprint arXiv:2307.00818 (2023) +25. Loper, M., Mahmood, N., Romero, J., Pons-Moll, G., Black, M.J.: Smpl: A skinned multi-person linear model. ACM transactions on graphics (TOG) 34(6), 1-16 (2015) +26. Lucas, T., Baradel, F., Weinzaepfel, P., Rogez, G.: Posegpt: quantization-based 3d human motion generation and forecasting. In: Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part VI. pp. 417-435. Springer (2022) +27. Mahmood, N., Ghorbani, N., Troje, N.F., Pons-Moll, G., Black, M.J.: Amass: Archive of motion capture as surface shapes. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 5442-5451 (2019) +28. Nie, Y., Dai, A., Han, X., Nießner, M.: Pose2room: understanding 3d scenes from human activities. In: European Conference on Computer Vision. pp. 425-443. Springer (2022) + +29. Pavlakos, G., Choutas, V., Ghorbani, N., Bolkart, T., Osman, A.A.A., Tzionas, D., Black, M.J.: Expressive body capture: 3D hands, face, and body from a single image. In: Proceedings IEEE Conf. on Computer Vision and Pattern Recognition (CVPR). pp. 10975-10985 (2019) +30. Petrovich, M., Black, M.J., Varol, G.: Action-conditioned 3d human motion synthesis with transformer vae. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 10985-10995 (2021) +31. Petrovich, M., Black, M.J., Varol, G.: Temos: Generating diverse human motions from textual descriptions. In: Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXII. pp. 480-497. Springer (2022) +32. Petrovich, M., Black, M.J., Varol, G.: Tmr: Text-to-motion retrieval using contrastive 3d human motion synthesis. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 9488-9497 (October 2023) +33. Plappert, M., Mandery, C., Asfour, T.: The kit motion-language dataset. Big data 4(4), 236-252 (2016) +34. Prokudin, S., Lassner, C., Romero, J.: Efficient learning on point clouds with basis point sets. In: Proceedings of the IEEE International Conference on Computer Vision. pp. 4332-4341 (2019) +35. Puig, X., Undersander, E., Szot, A., Cote, M.D., Partsey, R., Yang, J., Desai, R., Clegg, A.W., Hlavac, M., Min, T., Gervet, T., Vondruš, V., Berges, V.P., Turner, J., Maksymets, O., Kira, Z., Kalakrishnan, M., Malik, J., Chaplot, D.S., Jain, U., Batra, D., Rai, A., Mottaghi, R.: Habitat 3.0: A co-habitat for humans, avatars and robots (2023) +36. Starke, S., Zhang, H., Komura, T., Saito, J.: Neural state machine for character-scene interactions. ACM Trans. Graph. 38(6), 209-1 (2019) +37. Straub, J., Whelan, T., Ma, L., Chen, Y., Wijmans, E., Green, S., Engel, J.J., Mur-Artal, R., Ren, C., Verma, S., et al.: The replica dataset: A digital replica of indoor spaces. arXiv preprint arXiv:1906.05797 (2019) +38. Szot, A., Clegg, A., Undersander, E., Wijmans, E., Zhao, Y., Turner, J., Maestre, N., Mukadam, M., Chaplot, D., Maksymets, O., Gokaslan, A., Vondrus, V., Dharur, S., Meier, F., Galuba, W., Chang, A., Kira, Z., Koltun, V., Malik, J., Savva, M., Batra, D.: Habitat 2.0: Training home assistants to rearrange their habitat. In: Advances in Neural Information Processing Systems (NeurIPS) (2021) +39. Taheri, O., Choutas, V., Black, M.J., Tzionas, D.: Goal: Generating 4d whole-body motion for hand-object grasping. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13263-13273 (2022) +40. Tessler, C., Kasten, Y., Guo, Y., Mannor, S., Chechik, G., Peng, X.B.: Calm: Conditional adversarial latent models for directable virtual characters. In: ACM SIGGRAPH 2023 Conference Proceedings. SIGGRAPH '23, Association for Computing Machinery, New York, NY, USA (2023). https://doi.org/10.1145/3588432.3591541, https://doi.org/10.1145/3588432.3591541 +41. Tevet, G., Gordon, B., Hertz, A., Bermano, A.H., Cohen-Or, D.: Motionclip: Exposing human motion generation to clip space. In: Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXII. pp. 358-374. Springer (2022) +42. Tevet, G., Raab, S., Gordon, B., Shafir, Y., Cohen-or, D., Bermano, A.H.: Human motion diffusion model. In: The Eleventh International Conference on Learning Representations (2022) + +43. Wang, J., Xu, H., Xu, J., Liu, S., Wang, X.: Synthesizing long-term 3d human motion and interaction in 3d scenes. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9401-9411 (2021) +44. Wang, J., Rong, Y., Liu, J., Yan, S., Lin, D., Dai, B.: Towards diverse and natural scene-aware 3d human motion synthesis. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 20460-20469 (2022) +45. Wang, Z., Chen, Y., Liu, T., Zhu, Y., Liang, W., Huang, S.: Humanise: Language-conditioned human motion generation in 3d scenes. Advances in Neural Information Processing Systems 35, 14959-14971 (2022) +46. Wu, Y., Wang, J., Zhang, Y., Zhang, S., Hilliges, O., Yu, F., Tang, S.: Saga: Stochastic whole-body grasping with contact. In: European Conference on Computer Vision. pp. 257-274. Springer (2022) +47. Xu, L., Song, Z., Wang, D., Su, J., Fang, Z., Ding, C., Gan, W., Yan, Y., Jin, X., Yang, X., Zeng, W., Wu, W.: Actformer: A gan-based transformer towards general action-conditioned 3d human motion generation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 2228-2238 (October 2023) +48. Ye, S., Wang, Y., Li, J., Park, D., Liu, C.K., Xu, H., Wu, J.: Scene synthesis from human motion. In: SIGGRAPH Asia 2022 Conference Papers (2022) +49. Yi, H., Huang, C.H.P., Tripathi, S., Hering, L., Thies, J., Black, M.J.: Mime: Human-aware 3d scene generation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 12965-12976 (June 2023) +50. Yi, H., Huang, C.H.P., Tzionas, D., Kocabas, M., Hassan, M., Tang, S., Thies, J., Black, M.J.: Human-aware object placement for visual environment reconstruction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3959-3970 (2022) +51. Zhang, J.Y., Pepose, S., Joo, H., Ramanan, D., Malik, J., Kanazawa, A.: Perceiving 3d human-object spatial arrangements from a single image in the wild. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XII 16. pp. 34-51. Springer (2020) +52. Zhang, J., Zhang, Y., Cun, X., Zhang, Y., Zhao, H., Lu, H., Shen, X., Shan, Y.: Generating human motion from textual descriptions with discrete representations. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 14730-14740 (June 2023) +53. Zhang, M., Cai, Z., Pan, L., Hong, F., Guo, X., Yang, L., Liu, Z.: Motiondiffuse: Text-driven human motion generation with diffusion model. arXiv preprint arXiv:2208.15001 (2022) +54. Zhang, M., Guo, X., Pan, L., Cai, Z., Hong, F., Li, H., Yang, L., Liu, Z.: Remodiffuse: Retrieval-augmented motion diffusion model. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 364-373 (October 2023) +55. Zhang, S., Zhang, Y., Ma, Q., Black, M.J., Tang, S.: Place: Proximity learning of articulation and contact in 3d environments. In: 2020 International Conference on 3D Vision (3DV). pp. 642-651. IEEE (2020) +56. Zhang, X., Bhatnagar, B.L., Starke, S., Guzov, V., Pons-Moll, G.: Couch: Towards controllable human-chair interactions. In: European Conference on Computer Vision. pp. 518-535. Springer (2022) +57. Zhang, Y., Hassan, M., Neumann, H., Black, M.J., Tang, S.: Generating 3d people in scenes without people. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 6194-6204 (2020) + +58. Zhang, Y., Tang, S.: The wanderings of odysseus in 3d scenes. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 20481-20491 (2022) +59. Zhao, K., Wang, S., Zhang, Y., Beeler, T., Tang, S.: Compositional human-scene interaction synthesis with semantic control. In: European Conference on Computer Vision. pp. 311-327. Springer (2022) +60. Zhao, K., Zhang, Y., Wang, S., Beeler, T., Tang, S.: Synthesizing diverse human motions in 3d indoor scenes. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 14738-14749 (October 2023) +61. Zheng, Y., Yang, Y., Mo, K., Li, J., Yu, T., Liu, Y., Liu, C.K., Guibas, L.J.: Gimo: Gaze-informed human motion prediction in context. In: European Conference on Computer Vision. pp. 676-694. Springer (2022) +62. Zhou, Y., Barnes, C., Lu, J., Yang, J., Li, H.: On the continuity of rotation representations in neural networks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5745-5753 (2019) \ No newline at end of file diff --git a/2024/Revisit Human-Scene Interaction via Space Occupancy/images.zip b/2024/Revisit Human-Scene Interaction via Space Occupancy/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..5746df2bcd190c61c9ccfbbcf926fcb8b1a2f28e --- /dev/null +++ b/2024/Revisit Human-Scene Interaction via Space Occupancy/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:057ca4171f81d8ea639b3313601f06280ef1aa897946ec758d1886d81131414d +size 382036 diff --git a/2024/Revisit Human-Scene Interaction via Space Occupancy/layout.json b/2024/Revisit Human-Scene Interaction via Space Occupancy/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..b6994254badf0590ac7ce72579bb73c01bb7d09d --- /dev/null +++ b/2024/Revisit Human-Scene Interaction via Space Occupancy/layout.json @@ -0,0 +1,10918 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 156, + 111, + 459, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 156, + 111, + 459, + 148 + ], + "spans": [ + { + "bbox": [ + 156, + 111, + 459, + 148 + ], + "type": "text", + "content": "Revisit Human-Scene Interaction via Space Occupancy" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 140, + 167, + 474, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 140, + 167, + 474, + 192 + ], + "spans": [ + { + "bbox": [ + 140, + 167, + 474, + 192 + ], + "type": "text", + "content": "Xinpeng Liu1, Haowen Hou1,2, Yanchao Yang3, Yong-Lu Li1, and Cewu Lu1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 238, + 201, + 374, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 238, + 201, + 374, + 213 + ], + "spans": [ + { + "bbox": [ + 238, + 201, + 374, + 213 + ], + "type": "text", + "content": "1 Shanghai Jiao Tong University" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 261, + 213, + 351, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 261, + 213, + 351, + 224 + ], + "spans": [ + { + "bbox": [ + 261, + 213, + 351, + 224 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 261, + 213, + 351, + 224 + ], + "type": "text", + "content": " Soochow University" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 242, + 224, + 372, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 242, + 224, + 372, + 236 + ], + "spans": [ + { + "bbox": [ + 242, + 224, + 372, + 236 + ], + "type": "text", + "content": "3 The University of Hong Kong" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 153, + 236, + 459, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 153, + 236, + 459, + 258 + ], + "spans": [ + { + "bbox": [ + 153, + 236, + 459, + 258 + ], + "type": "text", + "content": "xinpengliu0907@gmail.com, haowenhou@outlook.com, yanchaoy@hku.hk, {yonglu_li, lucewu}@sjtu.edu.cn" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 159, + 288, + 455, + 531 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 288, + 455, + 531 + ], + "spans": [ + { + "bbox": [ + 159, + 288, + 455, + 531 + ], + "type": "text", + "content": "Abstract. Human-scene Interaction (HSI) generation is a challenging task and crucial for various downstream tasks. However, one of the major obstacles is its limited data scale. High-quality data with simultaneously captured human and 3D environments is hard to acquire, resulting in limited data diversity and complexity. In this work, we argue that interaction with a scene is essentially interacting with the space occupancy of the scene from an abstract physical perspective, leading us to a unified novel view of Human-Occupancy Interaction. By treating pure motion sequences as records of humans interacting with invisible scene occupancy, we can aggregate motion-only data into a large-scale paired human-occupancy interaction database: Motion Occupancy Base (MOB). Thus, the need for costly paired motion-scene datasets with high-quality scene scans can be substantially alleviated. With this new unified view of Human-Occupancy interaction, a single motion controller is proposed to reach the target state given the surrounding occupancy. Once trained on MOB with complex occupancy layout, which is stringent to human movements, the controller could handle cramped scenes and generalize well to general scenes with limited complexity like regular living rooms. With no GT 3D scenes for training, our method can generate realistic and stable HSI motions in diverse scenarios, including both static and dynamic scenes. The project is available at https://foruck.github.io/occu-page/." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 160, + 540, + 405, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 540, + 405, + 552 + ], + "spans": [ + { + "bbox": [ + 160, + 540, + 405, + 552 + ], + "type": "text", + "content": "Keywords: Human-Scene Interaction " + }, + { + "bbox": [ + 160, + 540, + 405, + 552 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 160, + 540, + 405, + 552 + ], + "type": "text", + "content": " Motion Generation" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 573, + 230, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 573, + 230, + 586 + ], + "spans": [ + { + "bbox": [ + 132, + 573, + 230, + 586 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 599, + 482, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 599, + 482, + 635 + ], + "spans": [ + { + "bbox": [ + 130, + 599, + 482, + 635 + ], + "type": "text", + "content": "Human-Scene Interaction (HSI) generation has been an active field recently with its application in animation [36] and embodied AI [35, 38]. For ideal HSI generation, models should be able to control the flexible articulated human body" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 132, + 643, + 223, + 653 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 643, + 223, + 653 + ], + "spans": [ + { + "bbox": [ + 132, + 643, + 223, + 653 + ], + "type": "text", + "content": "* Equal contribution." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 654, + 235, + 665 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 654, + 235, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 654, + 235, + 665 + ], + "type": "text", + "content": "\\*\\* Corresponding author." + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 114, + 481, + 235 + ], + "blocks": [ + { + "bbox": [ + 133, + 114, + 481, + 235 + ], + "lines": [ + { + "bbox": [ + 133, + 114, + 481, + 235 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 481, + 235 + ], + "type": "image", + "image_path": "e5f11cd6ab03abe5894adc5853a9ba71a584ea6e0172351860ef03795388e774.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 244, + 482, + 289 + ], + "lines": [ + { + "bbox": [ + 130, + 244, + 482, + 289 + ], + "spans": [ + { + "bbox": [ + 130, + 244, + 482, + 289 + ], + "type": "text", + "content": "Fig. 1: We propose that interacting with scenes is essentially interacting with its space occupancy for static HSI. In this view, we can unify motion-only data into a unified human-occupancy knowledge base and train a versatile Human-Occupancy Interaction controller upon it, achieving stable generation under various scenarios." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 316, + 480, + 351 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 316, + 480, + 351 + ], + "spans": [ + { + "bbox": [ + 130, + 316, + 480, + 351 + ], + "type": "text", + "content": "to conduct various motions in highly complex 3D environments, considering geometry constraints, motion feasibility, and diversity. In this paper, we focus on interactions without changing the scene, which we refer to as static HSI." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 352, + 481, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 352, + 481, + 449 + ], + "spans": [ + { + "bbox": [ + 130, + 352, + 481, + 449 + ], + "type": "text", + "content": "The limited data acquisition is a major obstacle for static HSI generation. Simultaneous capture of human and 3D scenes is expensive. Some [11,56] collected data only for a narrow range of target objects. Widely adopted PROX [12] contains only 60 clips with limited diversity and complexity. HUMANISE [45] synthesized data by fitting motion from large-scale motion base AMASS [27] into various scenes. CIRCLE [1] put the scenes in VR applications to reduce the cost. Most efforts take explicit scenes as necessary assets. Instead, we dig into the essence of static HSI and revisit it from a space occupancy view." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 449, + 482, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 449, + 482, + 592 + ], + "spans": [ + { + "bbox": [ + 130, + 449, + 482, + 592 + ], + "type": "text", + "content": "Space occupancy is usually identified as an elementary constraint for HSI: humans and scenes cannot occupy the same space simultaneously, which prevents humans from penetrating scenes. We argue that occupancy is inherently the primary factor, especially for static HSI, which is essentially interaction with the scene occupancy. As in Fig. 1, sitting on a chair is practically sitting on a chair-shaped solid being. The chair's occupancy allows us to rest our weight upon it, offering options to sit or stand, i.e., affordances, while its height and shape naturally invite sitting. Other elements like category labels (e.g., identifying it as a \"chair\" semantically) influence our likelihood of choosing to sit. The scene occupancy, therefore, offers a range of affordances and delineates potential interaction types, establishing an initial probability distribution for interactions. This is primarily accessible to the model and should be learned well at first." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 482, + 666 + ], + "type": "text", + "content": "In light of this, the requirements of realistic scenes for HSI learning could be loosened, inspiring us to revisit large-scale motion datasets. Instead of treating motion-only data as locomotion knowledge bases like existing efforts [58,60], we transform them into a comprehensive Human-Occupancy Interaction database, named Motion Occupancy Base (MOB). For any given motion sequence, we first identify the space occupied by the human at any time in the sequence. The" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 205, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 205, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 205, + 100 + ], + "type": "text", + "content": "Liu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 480, + 187 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 480, + 187 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 480, + 187 + ], + "type": "text", + "content": "remaining space is conceptualized as the \"scene\". In this framework, the motion becomes a record of interaction with these defined space occupancies. As a result, MOB aggregates enormous paired human-occupancy data from existing large-scale motion-only data. Compared to current datasets [1, 12, 45], MOB presents over-complex and stringent geometric constraints, pushing the model to the limit in interacting with space occupancy." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 188, + 481, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 188, + 481, + 462 + ], + "spans": [ + { + "bbox": [ + 130, + 188, + 481, + 462 + ], + "type": "text", + "content": "How do we exploit the aggregated MOB for static HSI? Previous approaches generally provided two different answers. Some [1,16,45,61] adopt an end-to-end pipeline from scene to motion, which is straightforward but inflexible without the ability to interact with dynamic scenes. Some [43, 44, 60] typically decouple HSI into interaction and locomotion, adopting a two-stage paradigm: first generating key interaction frames and then synthesizing locomotion between them, with the help of methods like path-finding algorithms. However, this decomposition overlooks the integrated nature of locomotion and interaction, like simultaneous arm movement and walking to navigate obstacles. Our approach, rooted in space occupancy, reconceptualizes locomotion and interaction as a unified Human-Occupancy Interaction. We develop an auto-regressive HumanOccupancy Interaction controller for motion generation. The controller processes various inputs, including surrounding occupancy and target locations, as control signals at each time step, producing future motion predictions. Additionally, we introduce a field regulation module for enhanced collision avoidance. There are three major advantages of our controller. First, by learning from the challenging cases in MOB, it could handle complex scenes but also generalize well to regular simpler scenarios. Second, its versatility accommodates any static HSI without the need to switch between models for locomotion and interactions, even for HOI motion synthesis. Third, its auto-regressive nature allows it to adapt to dynamic environments such as an automatic door, a capability lacked in previous methods. In extensive experiments, our controller performs effectively across various settings, even without training on real 3D scene data." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 464, + 481, + 523 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 464, + 481, + 523 + ], + "spans": [ + { + "bbox": [ + 130, + 464, + 481, + 523 + ], + "type": "text", + "content": "Overall, our contributions are three-fold: 1) We exhume Human-Occupancy Interaction as a major component of static HSI. 2) By digging out the inherent Human-Occupancy relationships of pure motions, we convert motion-only data into a paired Human-Occupancy database MOB, expanding the data scale. 3) We develop a versatile controller for flexible and stable HSI motion generation." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 543, + 242, + 555 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 543, + 242, + 555 + ], + "spans": [ + { + "bbox": [ + 131, + 543, + 242, + 555 + ], + "type": "text", + "content": "2 Related Works" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 570, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 570, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 570, + 481, + 666 + ], + "type": "text", + "content": "Human-Scene Interaction Generation. Generating natural human motion sequences has received increasing attention with the emerging MoCap datasets [4, 7, 18, 24, 27, 33]. Early efforts were paid to prefix/suffix-conditioned motion generation [2, 9, 10, 15], known as motion prediction and interpolation. There also have been efforts to pursue generating motion with respect to specific action semantics, either represented in action categories [8, 30, 47] or natural language [6, 7, 26, 31, 32, 41, 42, 52–54]. Audio-conditioned motion generation was also explored, like music for dance generation [21, 22]. However, these approaches tend" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 223, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 223, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 223, + 91, + 447, + 102 + ], + "type": "text", + "content": "Revisit Human-Scene Interaction via Space Occupancy" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 390 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 390 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 390 + ], + "type": "text", + "content": "to focus on human motion only, without placing the generated motion in a grounded 3D environment. Recently, progress has been made to advance HSI generation. Early efforts [23,51] first reconstructed human-object interaction in 3D. Some [13,55,57] proceeded by placing SMPL-X [29] meshes in scanned 3D scenes. Semantic controls were also introduced [59]. For HSI motion generation, some [11,36,56] focused on interacting with a single target object, mostly chairs. GOAL [39] and SAGA [46] generated full-body pre-grasp motions. There have also been efforts [14,40] for the physical control of specific interactions. With paired motion-scene datasets [1,12], approaches that synthesize HSI motion in cluttered scenes appeared. Two-stage based algorithms [43,44] were proposed to first generate static key-frame poses, and then interpolate them for a coherent motion. GAMMA [58] proposed an RL framework for locomotion on flat grounds with the help of off-the-shelf path-finding algorithms. DIMOS [60] extended GAMMA with extra interaction policies. In contrast, the follow-ups [1,16,45,61] adopted a simple end-to-end one-stage paradigm, while focusing more on boosting data. GIMO [61] collected data with gaze information. HUMANISE [45] synthesized pseudo-data from large-scale motion bases and 3D scan bases. CIRCLE [1] collected data with a VR application. LAMA [19] coupled the RL framework with a motion matching algorithm, generated HSI motions with only a unified policy, and leveraged motion editing for optimization. Our major difference from previous efforts is the Human-Occupancy Interaction view, which alleviates the data hunger, and also enables a unified controller. Moreover, we are the first to generate motion in a dynamic scene, which previous efforts failed to handle." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 391, + 482, + 452 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 391, + 482, + 452 + ], + "spans": [ + { + "bbox": [ + 130, + 391, + 482, + 452 + ], + "type": "text", + "content": "Scene Creation from Motion. Pose2Room [28] predicted 3D object bounding boxes from human motion. MOVER [50] reconstructed 3D scenes from videos that record human-scene interaction. SUMMON [48] and MIME [49] placed 3D object models concerning human motion. These methods could be adopted to boost HSI data, though usually limited to a small range of indoor furniture." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 471, + 299, + 485 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 471, + 299, + 485 + ], + "spans": [ + { + "bbox": [ + 132, + 471, + 299, + 485 + ], + "type": "text", + "content": "3 Motion Occupancy Base" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 497, + 482, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 497, + 482, + 628 + ], + "spans": [ + { + "bbox": [ + 130, + 497, + 482, + 628 + ], + "type": "text", + "content": "HSI generation could be formulated as " + }, + { + "bbox": [ + 130, + 497, + 482, + 628 + ], + "type": "inline_equation", + "content": "M = \\mathcal{G}(S)" + }, + { + "bbox": [ + 130, + 497, + 482, + 628 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 497, + 482, + 628 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 130, + 497, + 482, + 628 + ], + "type": "text", + "content": " is the generated motion, " + }, + { + "bbox": [ + 130, + 497, + 482, + 628 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 130, + 497, + 482, + 628 + ], + "type": "text", + "content": " is the motion generator, and " + }, + { + "bbox": [ + 130, + 497, + 482, + 628 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 130, + 497, + 482, + 628 + ], + "type": "text", + "content": " is a 3D scene. Conventionally, the scene is expected to be as realistic as possible, requiring expensive acquisition. We find that these scenes could function almost equally to their occupancy for static HSI. Thus, high-quality scenes might become less necessary here. Given this, we revisit the large-scale motion-only datasets. Intuitively, motion only contains body movement information. But if we switch our sight to the vacant space around the moving person, considering the space as another form of occupancy, we could find that these motions possess information on how humans interact with certain space occupancies. To this end, we aggregate motion-only datasets into the human-occupancy interaction database: Motion Occupancy Base (MOB)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 630, + 482, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 630, + 482, + 667 + ], + "spans": [ + { + "bbox": [ + 130, + 630, + 482, + 667 + ], + "type": "text", + "content": "The overall pipeline is shown in Fig. 2. Motion with SMPL(-X) [25,29] representation are collected, re-sampled to 30 FPS, and rotated so that the gravity direction points to the negative z-axis. For motion sequence " + }, + { + "bbox": [ + 130, + 630, + 482, + 667 + ], + "type": "inline_equation", + "content": "X = \\{x^i\\}_{i=1}^t" + }, + { + "bbox": [ + 130, + 630, + 482, + 667 + ], + "type": "text", + "content": ", where" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 205, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 205, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 205, + 100 + ], + "type": "text", + "content": "Liu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 150, + 114, + 465, + 222 + ], + "blocks": [ + { + "bbox": [ + 150, + 114, + 465, + 222 + ], + "lines": [ + { + "bbox": [ + 150, + 114, + 465, + 222 + ], + "spans": [ + { + "bbox": [ + 150, + 114, + 465, + 222 + ], + "type": "image", + "image_path": "c49f40fd5050f99401eb9c13f41457627a671cb695e7cc0f89bc1cece52ed7a0.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 229, + 482, + 264 + ], + "lines": [ + { + "bbox": [ + 131, + 229, + 482, + 264 + ], + "spans": [ + { + "bbox": [ + 131, + 229, + 482, + 264 + ], + "type": "text", + "content": "Fig. 3: The architecture of our versatile motion controller. Given motion state and histories " + }, + { + "bbox": [ + 131, + 229, + 482, + 264 + ], + "type": "inline_equation", + "content": "x^{t}, h^{t}" + }, + { + "bbox": [ + 131, + 229, + 482, + 264 + ], + "type": "text", + "content": ", and control signals " + }, + { + "bbox": [ + 131, + 229, + 482, + 264 + ], + "type": "inline_equation", + "content": "c_{o}^{t}, c_{t}^{t}" + }, + { + "bbox": [ + 131, + 229, + 482, + 264 + ], + "type": "text", + "content": ", the controller auto-regressively generates the next motion frame w.r.t. the target pose and the canonical occupancy." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 284, + 482, + 321 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 284, + 482, + 321 + ], + "spans": [ + { + "bbox": [ + 130, + 284, + 482, + 321 + ], + "type": "inline_equation", + "content": "x_{i}" + }, + { + "bbox": [ + 130, + 284, + 482, + 321 + ], + "type": "text", + "content": " is the SMPL(-X) representation in the " + }, + { + "bbox": [ + 130, + 284, + 482, + 321 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 284, + 482, + 321 + ], + "type": "text", + "content": "-th frame, " + }, + { + "bbox": [ + 130, + 284, + 482, + 321 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 284, + 482, + 321 + ], + "type": "text", + "content": " is the number of frames, we first obtain mesh sequence " + }, + { + "bbox": [ + 130, + 284, + 482, + 321 + ], + "type": "inline_equation", + "content": "X_{m} = \\{x_{m}^{i}\\}_{i = 1}^{t}" + }, + { + "bbox": [ + 130, + 284, + 482, + 321 + ], + "type": "text", + "content": " via SMPL(-X) model, where " + }, + { + "bbox": [ + 130, + 284, + 482, + 321 + ], + "type": "inline_equation", + "content": "x_{m}^{i} = (f^{i},v^{i})" + }, + { + "bbox": [ + 130, + 284, + 482, + 321 + ], + "type": "text", + "content": " is the mesh for the " + }, + { + "bbox": [ + 130, + 284, + 482, + 321 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 130, + 284, + 482, + 321 + ], + "type": "text", + "content": "-th frame, " + }, + { + "bbox": [ + 130, + 284, + 482, + 321 + ], + "type": "inline_equation", + "content": "f^{i},v^{i}" + }, + { + "bbox": [ + 130, + 284, + 482, + 321 + ], + "type": "text", + "content": " are the mesh faces and surface" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 320, + 299, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 320, + 299, + 525 + ], + "spans": [ + { + "bbox": [ + 130, + 320, + 299, + 525 + ], + "type": "text", + "content": "vertices. A bounding box of " + }, + { + "bbox": [ + 130, + 320, + 299, + 525 + ], + "type": "inline_equation", + "content": "X_{m}" + }, + { + "bbox": [ + 130, + 320, + 299, + 525 + ], + "type": "text", + "content": " is obtained based on vertices " + }, + { + "bbox": [ + 130, + 320, + 299, + 525 + ], + "type": "inline_equation", + "content": "V = \\{v_{i}\\}_{i=1}^{t}" + }, + { + "bbox": [ + 130, + 320, + 299, + 525 + ], + "type": "text", + "content": ". Given the predefined grid unit size " + }, + { + "bbox": [ + 130, + 320, + 299, + 525 + ], + "type": "inline_equation", + "content": "u" + }, + { + "bbox": [ + 130, + 320, + 299, + 525 + ], + "type": "text", + "content": ", the box region is voxelized into motion occupancy volume " + }, + { + "bbox": [ + 130, + 320, + 299, + 525 + ], + "type": "inline_equation", + "content": "\\hat{O} \\in \\{0,1\\}^{[H \\times W \\times D]}" + }, + { + "bbox": [ + 130, + 320, + 299, + 525 + ], + "type": "text", + "content": ", computed in two steps: 1) " + }, + { + "bbox": [ + 130, + 320, + 299, + 525 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 130, + 320, + 299, + 525 + ], + "type": "text", + "content": " occupy the corresponding voxels in " + }, + { + "bbox": [ + 130, + 320, + 299, + 525 + ], + "type": "inline_equation", + "content": "\\hat{O}" + }, + { + "bbox": [ + 130, + 320, + 299, + 525 + ], + "type": "text", + "content": ". 2) Given the centers of the un-occupied voxels as " + }, + { + "bbox": [ + 130, + 320, + 299, + 525 + ], + "type": "inline_equation", + "content": "C = \\{c_{i}\\}_{i=1}^{|C|}" + }, + { + "bbox": [ + 130, + 320, + 299, + 525 + ], + "type": "text", + "content": ", if " + }, + { + "bbox": [ + 130, + 320, + 299, + 525 + ], + "type": "inline_equation", + "content": "c_{i}" + }, + { + "bbox": [ + 130, + 320, + 299, + 525 + ], + "type": "text", + "content": " is inside any human mesh " + }, + { + "bbox": [ + 130, + 320, + 299, + 525 + ], + "type": "inline_equation", + "content": "x_{m}^{i}" + }, + { + "bbox": [ + 130, + 320, + 299, + 525 + ], + "type": "text", + "content": ", the corresponding voxel is occupied. The pseudo-scene occupancy " + }, + { + "bbox": [ + 130, + 320, + 299, + 525 + ], + "type": "inline_equation", + "content": "O" + }, + { + "bbox": [ + 130, + 320, + 299, + 525 + ], + "type": "text", + "content": " is then obtained by negating the motion occupancy, formulated as " + }, + { + "bbox": [ + 130, + 320, + 299, + 525 + ], + "type": "inline_equation", + "content": "O = 1 - \\hat{O}" + }, + { + "bbox": [ + 130, + 320, + 299, + 525 + ], + "type": "text", + "content": ". Thus, we convert motion-only sample " + }, + { + "bbox": [ + 130, + 320, + 299, + 525 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 130, + 320, + 299, + 525 + ], + "type": "text", + "content": " into a motion-occupancy pair " + }, + { + "bbox": [ + 130, + 320, + 299, + 525 + ], + "type": "inline_equation", + "content": "(X,O)" + }, + { + "bbox": [ + 130, + 320, + 299, + 525 + ], + "type": "text", + "content": " in MOB. MOB" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 306, + 344, + 482, + 459 + ], + "blocks": [ + { + "bbox": [ + 306, + 344, + 482, + 459 + ], + "lines": [ + { + "bbox": [ + 306, + 344, + 482, + 459 + ], + "spans": [ + { + "bbox": [ + 306, + 344, + 482, + 459 + ], + "type": "image", + "image_path": "9f04fe83a768594cd8978f1701a9470940a2412e939bee049ba6b5b61721b096.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 467, + 482, + 502 + ], + "lines": [ + { + "bbox": [ + 304, + 467, + 482, + 502 + ], + "spans": [ + { + "bbox": [ + 304, + 467, + 482, + 502 + ], + "type": "text", + "content": "Fig. 2: The construction process of Motion Occupancy Base. Note that the ceilings of the occupancy are hidden for clarity." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 525, + 481, + 536 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 525, + 481, + 536 + ], + "spans": [ + { + "bbox": [ + 130, + 525, + 481, + 536 + ], + "type": "text", + "content": "aggregates 13 datasets with 98k instances. More details are in the supplementary." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 536, + 482, + 585 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 536, + 482, + 585 + ], + "spans": [ + { + "bbox": [ + 130, + 536, + 482, + 585 + ], + "type": "text", + "content": "As shown in Fig. 2, humans in MOB are acting with rather strict geometric constraints, which is uncommon in regular scenes covered by existing datasets [1, 12, 45]. However, we show that by learning from hard cases in MOB, our model could grow its ability to handle both complex and regular scenarios well." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 601, + 318, + 613 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 601, + 318, + 613 + ], + "spans": [ + { + "bbox": [ + 131, + 601, + 318, + 613 + ], + "type": "text", + "content": "4 Versatile Motion Controller" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 131, + 624, + 228, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 624, + 228, + 635 + ], + "spans": [ + { + "bbox": [ + 131, + 624, + 228, + 635 + ], + "type": "text", + "content": "4.1 Preliminaries" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "content": "To instantiate " + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "inline_equation", + "content": "M = \\mathcal{G}(S)" + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "content": " for HSI generation, previous efforts either adopt a one-shot transformer generator or decompose " + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "content": " into two separate generators, one for" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 223, + 91, + 448, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 223, + 91, + 448, + 103 + ], + "spans": [ + { + "bbox": [ + 223, + 91, + 448, + 103 + ], + "type": "text", + "content": "Revisit Human-Scene Interaction via Space Occupancy" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "content": "interaction key-frame generation, the other for key-frame interpolation. Instead, we formulate our generator as an auto-regressive controller. At time step " + }, + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "content": ", given historical motion state " + }, + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "inline_equation", + "content": "h^t" + }, + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "content": ", current pose state " + }, + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "inline_equation", + "content": "x^t" + }, + { + "bbox": [ + 130, + 116, + 479, + 152 + ], + "type": "text", + "content": ", we expect a controller as" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 268, + 160, + 481, + 174 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 268, + 160, + 481, + 174 + ], + "spans": [ + { + "bbox": [ + 268, + 160, + 481, + 174 + ], + "type": "interline_equation", + "content": "\\hat {y} ^ {t} = \\mathbf {G} \\left(h ^ {t}, x ^ {t}, c ^ {t}\\right), \\tag {1}", + "image_path": "9662c2b294ff5167b6466629c6b133b016efd979424f369a18866fe143e3110c.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 182, + 479, + 230 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 182, + 479, + 230 + ], + "spans": [ + { + "bbox": [ + 130, + 182, + 479, + 230 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 182, + 479, + 230 + ], + "type": "inline_equation", + "content": "\\hat{y}^t" + }, + { + "bbox": [ + 130, + 182, + 479, + 230 + ], + "type": "text", + "content": " is the predicted future motion, and " + }, + { + "bbox": [ + 130, + 182, + 479, + 230 + ], + "type": "inline_equation", + "content": "c^t" + }, + { + "bbox": [ + 130, + 182, + 479, + 230 + ], + "type": "text", + "content": " is the control signals. An overview is in Fig. 3. Sec. 4.2 introduces motion state representation " + }, + { + "bbox": [ + 130, + 182, + 479, + 230 + ], + "type": "inline_equation", + "content": "h^t, x^t, y^t" + }, + { + "bbox": [ + 130, + 182, + 479, + 230 + ], + "type": "text", + "content": ". The control signals are covered in Sec. 4.3. The controller architecture is detailed in Sec. 4.4. Sec. 4.5 demonstrates the training and inference procedure." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 247, + 307, + 258 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 247, + 307, + 258 + ], + "spans": [ + { + "bbox": [ + 132, + 247, + 307, + 258 + ], + "type": "text", + "content": "4.2 Motion State Representation" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 266, + 480, + 349 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 266, + 480, + 349 + ], + "spans": [ + { + "bbox": [ + 130, + 266, + 480, + 349 + ], + "type": "text", + "content": "Notations. In the global coordinate, we represent motion sequence " + }, + { + "bbox": [ + 130, + 266, + 480, + 349 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 130, + 266, + 480, + 349 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 130, + 266, + 480, + 349 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 130, + 266, + 480, + 349 + ], + "type": "text", + "content": " frames as " + }, + { + "bbox": [ + 130, + 266, + 480, + 349 + ], + "type": "inline_equation", + "content": "X = \\{x^{t}\\}_{t = 1}^{T}" + }, + { + "bbox": [ + 130, + 266, + 480, + 349 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 130, + 266, + 480, + 349 + ], + "type": "inline_equation", + "content": "x^{t} = (p_{r}^{t},\\theta_{r}^{t},r^{t})" + }, + { + "bbox": [ + 130, + 266, + 480, + 349 + ], + "type": "text", + "content": " is SMPL [25] parameters, with root location " + }, + { + "bbox": [ + 130, + 266, + 480, + 349 + ], + "type": "inline_equation", + "content": "p_r^t\\in \\mathcal{R}^3" + }, + { + "bbox": [ + 130, + 266, + 480, + 349 + ], + "type": "text", + "content": ", root orientation " + }, + { + "bbox": [ + 130, + 266, + 480, + 349 + ], + "type": "inline_equation", + "content": "\\theta_r^t\\in \\mathcal{R}^6" + }, + { + "bbox": [ + 130, + 266, + 480, + 349 + ], + "type": "text", + "content": ", and the orientation of " + }, + { + "bbox": [ + 130, + 266, + 480, + 349 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 130, + 266, + 480, + 349 + ], + "type": "text", + "content": " limbs as " + }, + { + "bbox": [ + 130, + 266, + 480, + 349 + ], + "type": "inline_equation", + "content": "\\theta^t\\in \\mathcal{R}^{j\\times 6}" + }, + { + "bbox": [ + 130, + 266, + 480, + 349 + ], + "type": "text", + "content": ". Based on this, the location of " + }, + { + "bbox": [ + 130, + 266, + 480, + 349 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 130, + 266, + 480, + 349 + ], + "type": "text", + "content": " joints " + }, + { + "bbox": [ + 130, + 266, + 480, + 349 + ], + "type": "inline_equation", + "content": "p^t\\in \\mathcal{R}^{j\\times 3}" + }, + { + "bbox": [ + 130, + 266, + 480, + 349 + ], + "type": "text", + "content": " could also be computed via SMPL. For simplicity, we use " + }, + { + "bbox": [ + 130, + 266, + 480, + 349 + ], + "type": "inline_equation", + "content": "\\cdot^{-w}" + }, + { + "bbox": [ + 130, + 266, + 480, + 349 + ], + "type": "text", + "content": " to represent the value from frame " + }, + { + "bbox": [ + 130, + 266, + 480, + 349 + ], + "type": "inline_equation", + "content": "t - w" + }, + { + "bbox": [ + 130, + 266, + 480, + 349 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 130, + 266, + 480, + 349 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 266, + 480, + 349 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 266, + 480, + 349 + ], + "type": "inline_equation", + "content": "\\cdot^{+w}" + }, + { + "bbox": [ + 130, + 266, + 480, + 349 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 130, + 266, + 480, + 349 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 266, + 480, + 349 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 130, + 266, + 480, + 349 + ], + "type": "inline_equation", + "content": "t + w" + }, + { + "bbox": [ + 130, + 266, + 480, + 349 + ], + "type": "text", + "content": ". The 6D form from [62] is adopted for all rotation and orientation representations." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 146, + 350, + 350, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 350, + 350, + 361 + ], + "spans": [ + { + "bbox": [ + 146, + 350, + 350, + 361 + ], + "type": "text", + "content": "Current Pose State could be formulated as" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 225, + 370, + 480, + 384 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 225, + 370, + 480, + 384 + ], + "spans": [ + { + "bbox": [ + 225, + 370, + 480, + 384 + ], + "type": "interline_equation", + "content": "x ^ {t} = \\operatorname {C a n o} \\left(\\left\\{p _ {r} ^ {t}, p ^ {t}, \\theta_ {r} ^ {t}, \\theta^ {t}, \\dot {p} _ {r} ^ {t}, \\dot {p} ^ {t}, \\dot {\\theta} _ {r} ^ {t} \\right\\}, t\\right), \\tag {2}", + "image_path": "1cfac098d3a09a970d1a512febb04d09acc34f02b9411348929af6456ad6af10.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 392, + 479, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 392, + 479, + 415 + ], + "spans": [ + { + "bbox": [ + 130, + 392, + 479, + 415 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 130, + 392, + 479, + 415 + ], + "type": "inline_equation", + "content": "\\cdot^t" + }, + { + "bbox": [ + 130, + 392, + 479, + 415 + ], + "type": "text", + "content": " indicates the first derivative at " + }, + { + "bbox": [ + 130, + 392, + 479, + 415 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 392, + 479, + 415 + ], + "type": "text", + "content": ". Cano " + }, + { + "bbox": [ + 130, + 392, + 479, + 415 + ], + "type": "inline_equation", + "content": "(\\cdot, t)" + }, + { + "bbox": [ + 130, + 392, + 479, + 415 + ], + "type": "text", + "content": " canonicalizes the input w.r.t. the root location and facing direction at " + }, + { + "bbox": [ + 130, + 392, + 479, + 415 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 392, + 479, + 415 + ], + "type": "text", + "content": ", ensuring consistency across frames." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 146, + 416, + 465, + 427 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 416, + 465, + 427 + ], + "spans": [ + { + "bbox": [ + 146, + 416, + 465, + 427 + ], + "type": "text", + "content": "Historical Motion State " + }, + { + "bbox": [ + 146, + 416, + 465, + 427 + ], + "type": "inline_equation", + "content": "h^t" + }, + { + "bbox": [ + 146, + 416, + 465, + 427 + ], + "type": "text", + "content": " captures motion from frame " + }, + { + "bbox": [ + 146, + 416, + 465, + 427 + ], + "type": "inline_equation", + "content": "t - w" + }, + { + "bbox": [ + 146, + 416, + 465, + 427 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 146, + 416, + 465, + 427 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 146, + 416, + 465, + 427 + ], + "type": "text", + "content": " as" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 205, + 436, + 480, + 449 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 205, + 436, + 480, + 449 + ], + "spans": [ + { + "bbox": [ + 205, + 436, + 480, + 449 + ], + "type": "interline_equation", + "content": "h ^ {t} = \\operatorname {C a n o} \\left(\\left\\{p _ {r} ^ {- w}, p ^ {- w}, \\theta_ {r} ^ {- w}, \\dot {p} _ {r} ^ {- w}, \\dot {p} ^ {- w}, \\dot {\\theta} _ {r} ^ {- w} \\right\\}, t\\right), \\tag {3}", + "image_path": "26ef89bb1e77e335a578e7c2913197af4a17c0b4846fcb1bad5ec8fa57bf9a76.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 130, + 458, + 480, + 493 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 458, + 480, + 493 + ], + "spans": [ + { + "bbox": [ + 130, + 458, + 480, + 493 + ], + "type": "text", + "content": "with trajectories " + }, + { + "bbox": [ + 130, + 458, + 480, + 493 + ], + "type": "inline_equation", + "content": "p_r^{-w}, p^{-w}" + }, + { + "bbox": [ + 130, + 458, + 480, + 493 + ], + "type": "text", + "content": ", root orientation " + }, + { + "bbox": [ + 130, + 458, + 480, + 493 + ], + "type": "inline_equation", + "content": "\\theta_r^{-w}" + }, + { + "bbox": [ + 130, + 458, + 480, + 493 + ], + "type": "text", + "content": ", velocities " + }, + { + "bbox": [ + 130, + 458, + 480, + 493 + ], + "type": "inline_equation", + "content": "\\dot{p}_r^{-w}, \\dot{p}^{-w}" + }, + { + "bbox": [ + 130, + 458, + 480, + 493 + ], + "type": "text", + "content": ", and root angular velocity " + }, + { + "bbox": [ + 130, + 458, + 480, + 493 + ], + "type": "inline_equation", + "content": "\\dot{\\theta}_r^{-w}" + }, + { + "bbox": [ + 130, + 458, + 480, + 493 + ], + "type": "text", + "content": ". For conciseness, if not specified, only the " + }, + { + "bbox": [ + 130, + 458, + 480, + 493 + ], + "type": "inline_equation", + "content": "p^{-w}" + }, + { + "bbox": [ + 130, + 458, + 480, + 493 + ], + "type": "text", + "content": " of the hands and feet is adopted in practice." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 130, + 494, + 480, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 494, + 480, + 581 + ], + "spans": [ + { + "bbox": [ + 130, + 494, + 480, + 581 + ], + "type": "text", + "content": "Future Prediction " + }, + { + "bbox": [ + 130, + 494, + 480, + 581 + ], + "type": "inline_equation", + "content": "\\hat{y}^t" + }, + { + "bbox": [ + 130, + 494, + 480, + 581 + ], + "type": "text", + "content": " is composed of two parts as " + }, + { + "bbox": [ + 130, + 494, + 480, + 581 + ], + "type": "inline_equation", + "content": "\\hat{y}^t = \\{\\hat{x}^{t + 1},\\hat{u}^t\\}" + }, + { + "bbox": [ + 130, + 494, + 480, + 581 + ], + "type": "text", + "content": ". Note that " + }, + { + "bbox": [ + 130, + 494, + 480, + 581 + ], + "type": "inline_equation", + "content": "\\hat{x}^{t + 1}" + }, + { + "bbox": [ + 130, + 494, + 480, + 581 + ], + "type": "text", + "content": " is in the egocentric canonical of frame " + }, + { + "bbox": [ + 130, + 494, + 480, + 581 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 494, + 480, + 581 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 130, + 494, + 480, + 581 + ], + "type": "inline_equation", + "content": "\\hat{p}_r,\\hat{\\dot{\\theta}}\\subset \\hat{x}^{t + 1}" + }, + { + "bbox": [ + 130, + 494, + 480, + 581 + ], + "type": "text", + "content": " are adopted to transit from " + }, + { + "bbox": [ + 130, + 494, + 480, + 581 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 494, + 480, + 581 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 130, + 494, + 480, + 581 + ], + "type": "inline_equation", + "content": "t + 1" + }, + { + "bbox": [ + 130, + 494, + 480, + 581 + ], + "type": "text", + "content": ". Mirroring the historical motion state " + }, + { + "bbox": [ + 130, + 494, + 480, + 581 + ], + "type": "inline_equation", + "content": "h^t" + }, + { + "bbox": [ + 130, + 494, + 480, + 581 + ], + "type": "text", + "content": ", we also incorporate a comprehensive set of future trajectories, facing directions and their changing rates as " + }, + { + "bbox": [ + 130, + 494, + 480, + 581 + ], + "type": "inline_equation", + "content": "\\hat{u}^t = \\{p_r^{+f},p^{+f},\\theta^{+f},\\dot{p}_r^{+f},\\dot{p}^{+f},\\dot{\\theta}^{+f}\\}" + }, + { + "bbox": [ + 130, + 494, + 480, + 581 + ], + "type": "text", + "content": " for a window size of " + }, + { + "bbox": [ + 130, + 494, + 480, + 581 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 130, + 494, + 480, + 581 + ], + "type": "text", + "content": ", all in the canonical view of " + }, + { + "bbox": [ + 130, + 494, + 480, + 581 + ], + "type": "inline_equation", + "content": "t + 1" + }, + { + "bbox": [ + 130, + 494, + 480, + 581 + ], + "type": "text", + "content": ". Empirically, integrating " + }, + { + "bbox": [ + 130, + 494, + 480, + 581 + ], + "type": "inline_equation", + "content": "\\hat{u}^t" + }, + { + "bbox": [ + 130, + 494, + 480, + 581 + ], + "type": "text", + "content": " significantly enhances the model's precision in navigating toward the target." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 597, + 238, + 609 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 597, + 238, + 609 + ], + "spans": [ + { + "bbox": [ + 132, + 597, + 238, + 609 + ], + "type": "text", + "content": "4.3 Control Signals" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 130, + 617, + 479, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 479, + 640 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 479, + 640 + ], + "type": "text", + "content": "In this section, we introduce how the different control signals are encoded. Note that our models can operate without these optional signals." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 130, + 642, + 480, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 642, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 642, + 480, + 665 + ], + "type": "text", + "content": "Canonical Occupancy. A simple voxel-based representation is adopted to encode how the canonical space is occupied by the scenes. We utilize a cubic" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 205, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 205, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 205, + 101 + ], + "type": "text", + "content": "Liu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "content": "point grid " + }, + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "inline_equation", + "content": "o^t \\in \\{0,1\\}^{s \\times s \\times s}" + }, + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "content": " characterized by the grid size " + }, + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "content": " and the grid unit size " + }, + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "inline_equation", + "content": "u" + }, + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "content": " at frame " + }, + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "content": ". The grid is positioned in the human's canonical space, with its center offset by " + }, + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "inline_equation", + "content": "s/4" + }, + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "content": " in the human-facing direction. Each cell contains a binary value, indicating whether the cell is intersecting with the global scene occupancy " + }, + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "inline_equation", + "content": "O" + }, + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "content": ". The canonical occupancy is thus represented as " + }, + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "inline_equation", + "content": "c_o^t = \\text{Flatten}(o^t) \\in \\{0,1\\}^{s^3}" + }, + { + "bbox": [ + 130, + 116, + 482, + 224 + ], + "type": "text", + "content": ". Note that because the occupancy is calculated in real-time, this control signal can naturally adapt to dynamic scenes where each timestep is treated statically. We will show in Section 5 that even though our training data is all static, our model exhibits a smooth generalization to dynamic scenes." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 224, + 483, + 310 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 224, + 483, + 310 + ], + "spans": [ + { + "bbox": [ + 130, + 224, + 483, + 310 + ], + "type": "text", + "content": "Target Pose is also considered a flexible control signal, which can be given in two ways. The first is the positions of the root, hands, and feet, each defined in relation to the canonical space of the current pose, represented as " + }, + { + "bbox": [ + 130, + 224, + 483, + 310 + ], + "type": "inline_equation", + "content": "c_t^t \\in \\mathcal{R}^{5 \\times 3}" + }, + { + "bbox": [ + 130, + 224, + 483, + 310 + ], + "type": "text", + "content": ". In practice, these could be assigned manually or generated from off-the-shelf models [59]. Also, partial or blank targets are acceptable. The second is the future trajectory of end-effectors, denoted as " + }, + { + "bbox": [ + 130, + 224, + 483, + 310 + ], + "type": "inline_equation", + "content": "c_t^{+f}" + }, + { + "bbox": [ + 130, + 224, + 483, + 310 + ], + "type": "text", + "content": " for a future window size of " + }, + { + "bbox": [ + 130, + 224, + 483, + 310 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 130, + 224, + 483, + 310 + ], + "type": "text", + "content": ". This paves the way for different applications in Section 5." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 331, + 280, + 341 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 331, + 280, + 341 + ], + "spans": [ + { + "bbox": [ + 132, + 331, + 280, + 341 + ], + "type": "text", + "content": "4.4 Controller Architecture" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 354, + 367, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 354, + 367, + 533 + ], + "spans": [ + { + "bbox": [ + 130, + 354, + 367, + 533 + ], + "type": "text", + "content": "We adopt a lightweight transformer-based architecture. " + }, + { + "bbox": [ + 130, + 354, + 367, + 533 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 130, + 354, + 367, + 533 + ], + "type": "text", + "content": " fully connected layers first project the " + }, + { + "bbox": [ + 130, + 354, + 367, + 533 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 130, + 354, + 367, + 533 + ], + "type": "text", + "content": " input features from " + }, + { + "bbox": [ + 130, + 354, + 367, + 533 + ], + "type": "inline_equation", + "content": "\\{h^t,x^t,c^t\\}" + }, + { + "bbox": [ + 130, + 354, + 367, + 533 + ], + "type": "text", + "content": " to the latent " + }, + { + "bbox": [ + 130, + 354, + 367, + 533 + ], + "type": "inline_equation", + "content": "z\\in \\mathcal{R}^{n\\times d}" + }, + { + "bbox": [ + 130, + 354, + 367, + 533 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 130, + 354, + 367, + 533 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 130, + 354, + 367, + 533 + ], + "type": "text", + "content": " is then fed into encoder " + }, + { + "bbox": [ + 130, + 354, + 367, + 533 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 130, + 354, + 367, + 533 + ], + "type": "text", + "content": " to distill dynamic information from the motion states and control signals. We devise " + }, + { + "bbox": [ + 130, + 354, + 367, + 533 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 130, + 354, + 367, + 533 + ], + "type": "text", + "content": " as a sequence of Transformer encoders, which takes " + }, + { + "bbox": [ + 130, + 354, + 367, + 533 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 130, + 354, + 367, + 533 + ], + "type": "text", + "content": " as input and outputs refined dynamic information " + }, + { + "bbox": [ + 130, + 354, + 367, + 533 + ], + "type": "inline_equation", + "content": "\\hat{z} = E(z)\\in \\mathcal{R}^{n\\times d}" + }, + { + "bbox": [ + 130, + 354, + 367, + 533 + ], + "type": "text", + "content": ". Finally, we adopt a transformer decoder " + }, + { + "bbox": [ + 130, + 354, + 367, + 533 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 130, + 354, + 367, + 533 + ], + "type": "text", + "content": " followed by a linear layer to condense the dynamic information, resulting in the final predictions. With learnable query " + }, + { + "bbox": [ + 130, + 354, + 367, + 533 + ], + "type": "inline_equation", + "content": "q\\in \\mathcal{R}^d" + }, + { + "bbox": [ + 130, + 354, + 367, + 533 + ], + "type": "text", + "content": " and the dynamic information " + }, + { + "bbox": [ + 130, + 354, + 367, + 533 + ], + "type": "inline_equation", + "content": "\\hat{z}" + }, + { + "bbox": [ + 130, + 354, + 367, + 533 + ], + "type": "text", + "content": " serving as K and V, respectively, the output of the transformer is further concatenated with a random noise " + }, + { + "bbox": [ + 130, + 354, + 367, + 533 + ], + "type": "inline_equation", + "content": "\\epsilon \\in \\mathcal{R}^d" + }, + { + "bbox": [ + 130, + 354, + 367, + 533 + ], + "type": "text", + "content": " for randomness. The linear layer then outputs the final prediction as " + }, + { + "bbox": [ + 130, + 354, + 367, + 533 + ], + "type": "inline_equation", + "content": "\\hat{y}^t = D(\\hat{z},q,\\epsilon)" + }, + { + "bbox": [ + 130, + 354, + 367, + 533 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 534, + 369, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 534, + 369, + 652 + ], + "spans": [ + { + "bbox": [ + 130, + 534, + 369, + 652 + ], + "type": "text", + "content": "Furthermore, to enhance collision avoidance, we introduce a novel module called Field Regulation. The basic idea is that humans tend to move faster in obstacle-sparse areas and decelerate in complex environments or when an obstacle is nearby. At very short distances, especially with an imminent collision happening, a complete halt is necessary. Given this and taking inspiration from the potential field in robot motion planning [17], we propose Occupancy Field, which exerts negligible influence on humans at a distance but" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 653, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 653, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 653, + 481, + 666 + ], + "type": "text", + "content": "escalates to a sturdy repulsion as proximity increases, thereby compelling the" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 375, + 378, + 481, + 483 + ], + "blocks": [ + { + "bbox": [ + 375, + 378, + 481, + 483 + ], + "lines": [ + { + "bbox": [ + 375, + 378, + 481, + 483 + ], + "spans": [ + { + "bbox": [ + 375, + 378, + 481, + 483 + ], + "type": "image", + "image_path": "30e2668dbbe56429deaffa35f1c05c951373bdb753827e069578e8503f4ffb3c.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 373, + 491, + 482, + 625 + ], + "lines": [ + { + "bbox": [ + 373, + 491, + 482, + 625 + ], + "spans": [ + { + "bbox": [ + 373, + 491, + 482, + 625 + ], + "type": "text", + "content": "Fig.4: Occupied voxel centers " + }, + { + "bbox": [ + 373, + 491, + 482, + 625 + ], + "type": "inline_equation", + "content": "\\{P_i\\}" + }, + { + "bbox": [ + 373, + 491, + 482, + 625 + ], + "type": "text", + "content": " are located at the top left of the figure. The initial velocity vector, " + }, + { + "bbox": [ + 373, + 491, + 482, + 625 + ], + "type": "inline_equation", + "content": "\\dot{p}_j" + }, + { + "bbox": [ + 373, + 491, + 482, + 625 + ], + "type": "text", + "content": ", is directed upwards, posing a risk of collision. To mitigate this, the occupancy field introduces a corrective velocity component " + }, + { + "bbox": [ + 373, + 491, + 482, + 625 + ], + "type": "inline_equation", + "content": "\\Delta \\dot{p}_j" + }, + { + "bbox": [ + 373, + 491, + 482, + 625 + ], + "type": "text", + "content": ". This redirects the final velocity " + }, + { + "bbox": [ + 373, + 491, + 482, + 625 + ], + "type": "inline_equation", + "content": "\\dot{p}_j^{out}" + }, + { + "bbox": [ + 373, + 491, + 482, + 625 + ], + "type": "text", + "content": " to avoid collision." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 223, + 91, + 448, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 223, + 91, + 448, + 103 + ], + "spans": [ + { + "bbox": [ + 223, + 91, + 448, + 103 + ], + "type": "text", + "content": "Revisit Human-Scene Interaction via Space Occupancy" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 91, + 481, + 100 + ], + "type": "text", + "content": "7" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 480, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 480, + 140 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 480, + 140 + ], + "type": "text", + "content": "human to decelerate and avoid collision. The occupancy field influences human motion only when a human is in motion toward it. The field functions as" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 261, + 146, + 481, + 162 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 261, + 146, + 481, + 162 + ], + "spans": [ + { + "bbox": [ + 261, + 146, + 481, + 162 + ], + "type": "interline_equation", + "content": "\\Delta \\dot {p} _ {j} ^ {t + 1} = \\mathcal {F} \\left(p _ {j} ^ {t + 1}, c _ {o} ^ {t}\\right). \\tag {4}", + "image_path": "98ce9a6489735ca507a4b0695c88fcab6bf6be59e9553ca4648b5766f73c876a.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 167, + 479, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 167, + 479, + 228 + ], + "spans": [ + { + "bbox": [ + 132, + 167, + 479, + 228 + ], + "type": "text", + "content": "The field " + }, + { + "bbox": [ + 132, + 167, + 479, + 228 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 132, + 167, + 479, + 228 + ], + "type": "text", + "content": " calculates a correction " + }, + { + "bbox": [ + 132, + 167, + 479, + 228 + ], + "type": "inline_equation", + "content": "\\Delta \\dot{p}^{t + 1}" + }, + { + "bbox": [ + 132, + 167, + 479, + 228 + ], + "type": "text", + "content": " w.r.t. predicted joint velocity " + }, + { + "bbox": [ + 132, + 167, + 479, + 228 + ], + "type": "inline_equation", + "content": "\\dot{p}_j^{t + 1} \\in \\{\\dot{p}_r^{t + 1}, \\dot{p}^{t + 1}\\}" + }, + { + "bbox": [ + 132, + 167, + 479, + 228 + ], + "type": "text", + "content": " and canonical occupancy " + }, + { + "bbox": [ + 132, + 167, + 479, + 228 + ], + "type": "inline_equation", + "content": "c_o^t" + }, + { + "bbox": [ + 132, + 167, + 479, + 228 + ], + "type": "text", + "content": ", which is then added to " + }, + { + "bbox": [ + 132, + 167, + 479, + 228 + ], + "type": "inline_equation", + "content": "\\dot{p}^{t + 1}" + }, + { + "bbox": [ + 132, + 167, + 479, + 228 + ], + "type": "text", + "content": ". For simplicity, the superscripts indicating the time steps are omitted in the following equations. In detail, given the center points of all the occupied voxels in " + }, + { + "bbox": [ + 132, + 167, + 479, + 228 + ], + "type": "inline_equation", + "content": "c_o" + }, + { + "bbox": [ + 132, + 167, + 479, + 228 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 132, + 167, + 479, + 228 + ], + "type": "inline_equation", + "content": "P \\subset \\mathcal{R}^{n_p \\times 3}" + }, + { + "bbox": [ + 132, + 167, + 479, + 228 + ], + "type": "text", + "content": ", the correction is calculated as" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 211, + 236, + 481, + 264 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 211, + 236, + 481, + 264 + ], + "spans": [ + { + "bbox": [ + 211, + 236, + 481, + 264 + ], + "type": "interline_equation", + "content": "\\Delta \\dot {p} _ {j} = \\sum_ {P _ {i} \\in P} - k r _ {i} \\max \\left(0, \\frac {1}{\\| v _ {i j} \\| _ {a} - \\gamma} - b\\right), \\tag {5}", + "image_path": "885246ac75b59f432272428937dad054275d0ac4ae3ee03adcc83ea363e83601.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 219, + 268, + 480, + 281 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 219, + 268, + 480, + 281 + ], + "spans": [ + { + "bbox": [ + 219, + 268, + 480, + 281 + ], + "type": "interline_equation", + "content": "r _ {i} = \\dot {p} _ {j} \\max (0, \\langle \\dot {p} _ {j}, v _ {i j} \\rangle), \\tag {6}", + "image_path": "bef187d23efedfaa92117d97626577b762951066d7a5e50f19d01e1f50b7719e.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 287, + 481, + 370 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 287, + 481, + 370 + ], + "spans": [ + { + "bbox": [ + 132, + 287, + 481, + 370 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 132, + 287, + 481, + 370 + ], + "type": "inline_equation", + "content": "v_{ij} = P_i - p_j" + }, + { + "bbox": [ + 132, + 287, + 481, + 370 + ], + "type": "text", + "content": " is the vector from the joint " + }, + { + "bbox": [ + 132, + 287, + 481, + 370 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 132, + 287, + 481, + 370 + ], + "type": "text", + "content": " to the voxel center, " + }, + { + "bbox": [ + 132, + 287, + 481, + 370 + ], + "type": "inline_equation", + "content": "\\langle \\cdot, \\cdot \\rangle" + }, + { + "bbox": [ + 132, + 287, + 481, + 370 + ], + "type": "text", + "content": " indicates the cosine between the two vectors, " + }, + { + "bbox": [ + 132, + 287, + 481, + 370 + ], + "type": "inline_equation", + "content": "\\| \\|_a" + }, + { + "bbox": [ + 132, + 287, + 481, + 370 + ], + "type": "text", + "content": " denotes " + }, + { + "bbox": [ + 132, + 287, + 481, + 370 + ], + "type": "inline_equation", + "content": "a" + }, + { + "bbox": [ + 132, + 287, + 481, + 370 + ], + "type": "text", + "content": "-norm, and " + }, + { + "bbox": [ + 132, + 287, + 481, + 370 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 132, + 287, + 481, + 370 + ], + "type": "text", + "content": " is the stiffness factor. " + }, + { + "bbox": [ + 132, + 287, + 481, + 370 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 132, + 287, + 481, + 370 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 132, + 287, + 481, + 370 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 132, + 287, + 481, + 370 + ], + "type": "text", + "content": " are thresholds to constrain the distance range of effective voxels, where " + }, + { + "bbox": [ + 132, + 287, + 481, + 370 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 132, + 287, + 481, + 370 + ], + "type": "text", + "content": " reflects the human body's width, and voxels within this range are discounted for stability. An illustration of Field Regulation is shown in Fig. 4. The field regulation is incorporated in the training of the controller to adapt to it, avoiding potential foot sliding caused due to direct modification on velocities." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 386, + 276, + 398 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 386, + 276, + 398 + ], + "spans": [ + { + "bbox": [ + 132, + 386, + 276, + 398 + ], + "type": "text", + "content": "4.5 Training and Inference" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 403, + 479, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 403, + 479, + 464 + ], + "spans": [ + { + "bbox": [ + 132, + 403, + 479, + 464 + ], + "type": "text", + "content": "Training Objectives The controller is trained in a fully-supervised manner. We employ a mixture of L1 and L2 losses as basic objectives. Besides, penetration loss and occupancy field loss are designed to hinder undesired collisions. Specifically, L1 loss is utilized for rotation-related output values " + }, + { + "bbox": [ + 132, + 403, + 479, + 464 + ], + "type": "inline_equation", + "content": "\\hat{y}_{\\theta}^{t}\\in \\hat{y}^{t}" + }, + { + "bbox": [ + 132, + 403, + 479, + 464 + ], + "type": "text", + "content": ", while L2 loss is applied to the rest outputs " + }, + { + "bbox": [ + 132, + 403, + 479, + 464 + ], + "type": "inline_equation", + "content": "\\hat{y}_p^t" + }, + { + "bbox": [ + 132, + 403, + 479, + 464 + ], + "type": "text", + "content": ". The mixture loss is formulated as" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 237, + 472, + 480, + 486 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 237, + 472, + 480, + 486 + ], + "spans": [ + { + "bbox": [ + 237, + 472, + 480, + 486 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {m i x} = \\left\\| y _ {\\theta} ^ {t} - \\hat {y} _ {\\theta} ^ {t} \\right\\| _ {1} + \\left\\| y _ {p} ^ {t} - \\hat {y} _ {p} ^ {t} \\right\\| _ {2}. \\tag {7}", + "image_path": "68ea4b92f887f6cc84a7e7231afae04241d88e5b941aff0a8c8ba0fa93b369a3.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 491, + 448, + 503 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 491, + 448, + 503 + ], + "spans": [ + { + "bbox": [ + 132, + 491, + 448, + 503 + ], + "type": "text", + "content": "Given the predicted joint location " + }, + { + "bbox": [ + 132, + 491, + 448, + 503 + ], + "type": "inline_equation", + "content": "\\hat{p}^{t + 1}" + }, + { + "bbox": [ + 132, + 491, + 448, + 503 + ], + "type": "text", + "content": ", penetration loss is computed as" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 230, + 510, + 480, + 541 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 230, + 510, + 480, + 541 + ], + "spans": [ + { + "bbox": [ + 230, + 510, + 480, + 541 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {p e n} = \\sum_ {i = 0} ^ {j} O \\left(\\hat {p} _ {j} ^ {t + 1}\\right) \\| r e f \\left(\\hat {p} _ {j} ^ {t + 1}, O\\right) \\| _ {2}, \\tag {8}", + "image_path": "09ace097efc0db740de20a86da1f91bc8f2212e7e06d50677f32c32235019cec.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 548, + 481, + 634 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 548, + 481, + 634 + ], + "spans": [ + { + "bbox": [ + 132, + 548, + 481, + 634 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 132, + 548, + 481, + 634 + ], + "type": "inline_equation", + "content": "O(\\cdot)" + }, + { + "bbox": [ + 132, + 548, + 481, + 634 + ], + "type": "text", + "content": " returns 1 if the point is occupied in " + }, + { + "bbox": [ + 132, + 548, + 481, + 634 + ], + "type": "inline_equation", + "content": "O" + }, + { + "bbox": [ + 132, + 548, + 481, + 634 + ], + "type": "text", + "content": " else 0, and " + }, + { + "bbox": [ + 132, + 548, + 481, + 634 + ], + "type": "inline_equation", + "content": "ref(\\hat{p}_j^{t + 1},O)" + }, + { + "bbox": [ + 132, + 548, + 481, + 634 + ], + "type": "text", + "content": " calculates the closest unoccupied voxel center in " + }, + { + "bbox": [ + 132, + 548, + 481, + 634 + ], + "type": "inline_equation", + "content": "O" + }, + { + "bbox": [ + 132, + 548, + 481, + 634 + ], + "type": "text", + "content": " for joint " + }, + { + "bbox": [ + 132, + 548, + 481, + 634 + ], + "type": "inline_equation", + "content": "\\hat{p}_j^{t + 1}" + }, + { + "bbox": [ + 132, + 548, + 481, + 634 + ], + "type": "text", + "content": ". Occupancy field loss has two components. First, to deter undesired interactions, we penalize the proportion of the occupancy field effect relative to the output root velocity. Second, observing occasional unnatural spikes in velocity, where the human figure seemingly attempts to navigate through an obstacle, we punish excessive velocities. The occupancy field loss is formulated as" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 249, + 640, + 480, + 669 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 249, + 640, + 480, + 669 + ], + "spans": [ + { + "bbox": [ + 249, + 640, + 480, + 669 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {f i e l d}} = \\left(\\frac {| \\Delta \\dot {p} |}{| \\dot {p} |}\\right) ^ {2} + | \\dot {p} | ^ {2}. \\tag {9}", + "image_path": "0e7e77b7766ba16d50d4ddc1e15f82c0cd234588b814903d39e90588c15e6cb1.jpg" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 140, + 100 + ], + "type": "text", + "content": "8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 205, + 101 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 205, + 101 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 205, + 101 + ], + "type": "text", + "content": "Liu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 187, + 114, + 430, + 202 + ], + "blocks": [ + { + "bbox": [ + 187, + 114, + 430, + 202 + ], + "lines": [ + { + "bbox": [ + 187, + 114, + 430, + 202 + ], + "spans": [ + { + "bbox": [ + 187, + 114, + 430, + 202 + ], + "type": "table", + "html": "
MethodSuc. (%) ↑DT (cm) ↓Time (s)FS (%)PEN ↓ERP
GT100.000.003.003.360.00-
NSM [36]11.7632.795.3311.04134.497.26
SAMP [11]13.9452.496.549.87145.746.98
DIMOS [60]15.7348.976.8210.52123.126.82
Ours w/o Occupancy71.2214.662.256.8946.054.13
Ours77.9115.303.628.1315.214.24
Ours w/o Lpen74.1414.893.547.8931.934.96
Ours w/o Lfield75.3415.692.8611.6525.835.70
Ours w/o FR77.3814.852.346.5733.494.18
Ours w/ BPS73.6416.833.359.3436.724.72
", + "image_path": "7708eee9e0cc8944ecac7348858a1abce4a4dda3f12efb60052da42dea5a2d81.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 225, + 202, + 388, + 213 + ], + "lines": [ + { + "bbox": [ + 225, + 202, + 388, + 213 + ], + "spans": [ + { + "bbox": [ + 225, + 202, + 388, + 213 + ], + "type": "text", + "content": "Table 1: Quantitative results on MOB." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 132, + 246, + 373, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 246, + 373, + 258 + ], + "spans": [ + { + "bbox": [ + 132, + 246, + 373, + 258 + ], + "type": "text", + "content": "The final loss is computed with loss coefficients " + }, + { + "bbox": [ + 132, + 246, + 373, + 258 + ], + "type": "inline_equation", + "content": "\\alpha, \\beta" + }, + { + "bbox": [ + 132, + 246, + 373, + 258 + ], + "type": "text", + "content": " as" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 244, + 268, + 480, + 279 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 244, + 268, + 480, + 279 + ], + "spans": [ + { + "bbox": [ + 244, + 268, + 480, + 279 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathcal {L} _ {m i x} + \\alpha \\mathcal {L} _ {p e n} + \\beta \\mathcal {L} _ {\\text {f i e l d}}. \\tag {10}", + "image_path": "c445deef31ce4e955eac2a1fa059d0e00df635df781d1379b8df4d4a73a9fdf6.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 288, + 482, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 288, + 482, + 407 + ], + "spans": [ + { + "bbox": [ + 130, + 288, + 482, + 407 + ], + "type": "text", + "content": "Auto-regressive Inference. We perform motion generation in an auto-regressive manner, enabling the model to adapt to dynamic environmental changes. Since the model solely manipulates canonical information, we track the global root position " + }, + { + "bbox": [ + 130, + 288, + 482, + 407 + ], + "type": "inline_equation", + "content": "p_{g}" + }, + { + "bbox": [ + 130, + 288, + 482, + 407 + ], + "type": "text", + "content": " and the facing direction " + }, + { + "bbox": [ + 130, + 288, + 482, + 407 + ], + "type": "inline_equation", + "content": "\\theta_{g}" + }, + { + "bbox": [ + 130, + 288, + 482, + 407 + ], + "type": "text", + "content": " of the human in the world coordinate system. Initially, we set past and current velocities " + }, + { + "bbox": [ + 130, + 288, + 482, + 407 + ], + "type": "inline_equation", + "content": "\\dot{p}^{0}" + }, + { + "bbox": [ + 130, + 288, + 482, + 407 + ], + "type": "text", + "content": " to zero. Before each prediction at frame " + }, + { + "bbox": [ + 130, + 288, + 482, + 407 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 130, + 288, + 482, + 407 + ], + "type": "text", + "content": ", we compute the control signals (if there are any), i.e., the relative target position " + }, + { + "bbox": [ + 130, + 288, + 482, + 407 + ], + "type": "inline_equation", + "content": "c_{t}^{t}" + }, + { + "bbox": [ + 130, + 288, + 482, + 407 + ], + "type": "text", + "content": " and the canonical space occupancy " + }, + { + "bbox": [ + 130, + 288, + 482, + 407 + ], + "type": "inline_equation", + "content": "c_{o}^{t}" + }, + { + "bbox": [ + 130, + 288, + 482, + 407 + ], + "type": "text", + "content": ", based on the current values of " + }, + { + "bbox": [ + 130, + 288, + 482, + 407 + ], + "type": "inline_equation", + "content": "p_{g}^{t}" + }, + { + "bbox": [ + 130, + 288, + 482, + 407 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 288, + 482, + 407 + ], + "type": "inline_equation", + "content": "\\theta_{g}^{t}" + }, + { + "bbox": [ + 130, + 288, + 482, + 407 + ], + "type": "text", + "content": ". After the prediction, we update " + }, + { + "bbox": [ + 130, + 288, + 482, + 407 + ], + "type": "inline_equation", + "content": "p_{g}^{t}" + }, + { + "bbox": [ + 130, + 288, + 482, + 407 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 130, + 288, + 482, + 407 + ], + "type": "inline_equation", + "content": "\\theta_{g}^{t}" + }, + { + "bbox": [ + 130, + 288, + 482, + 407 + ], + "type": "text", + "content": " with the model's output, along with the current pose state " + }, + { + "bbox": [ + 130, + 288, + 482, + 407 + ], + "type": "inline_equation", + "content": "c^{t + 1}" + }, + { + "bbox": [ + 130, + 288, + 482, + 407 + ], + "type": "text", + "content": ". Concurrently, we refresh the historical motion state " + }, + { + "bbox": [ + 130, + 288, + 482, + 407 + ], + "type": "inline_equation", + "content": "h^{t + 1}" + }, + { + "bbox": [ + 130, + 288, + 482, + 407 + ], + "type": "text", + "content": " to prepare for the next input." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 426, + 230, + 440 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 426, + 230, + 440 + ], + "spans": [ + { + "bbox": [ + 132, + 426, + 230, + 440 + ], + "type": "text", + "content": "5 Experiments" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 449, + 482, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 449, + 482, + 582 + ], + "spans": [ + { + "bbox": [ + 130, + 449, + 482, + 582 + ], + "type": "text", + "content": "Implementation Details. The controller consists of a two-layer transformer encoder and a single-layer transformer decoder, both with a latent dimension of 512 and 8 heads. The canonical occupancy grid size is set as (25, 25, 25), with a unit size of " + }, + { + "bbox": [ + 130, + 449, + 482, + 582 + ], + "type": "inline_equation", + "content": "8\\mathrm{cm}" + }, + { + "bbox": [ + 130, + 449, + 482, + 582 + ], + "type": "text", + "content": ". The controller operates at 10 FPS with the history and future window sizes both set as 1, which empirically brings more stable outcomes. The controller is trained on the MOB, with 1,393 sequences excluded from training as the test set. We train the controller for 75k iterations, with a cosine learning rate decay restart strategy. The initial learning rate is set as 1e-4. The control signals are randomly masked for augmentation during training. Scheduled Sampling [3] is adopted. The whole training takes approximately 6 hours on a single 12G NVIDIA Titan Xp. More details are in the supplementary." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 598, + 359, + 611 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 598, + 359, + 611 + ], + "spans": [ + { + "bbox": [ + 132, + 598, + 359, + 611 + ], + "type": "text", + "content": "5.1 Interacting with Occupancy from MOB" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 617, + 481, + 666 + ], + "type": "text", + "content": "We first evaluate the ability to interact with occupancy volumes with a goal-reaching task on MOB. Given an occupancy volume, a starting pose, and a target end-effector pose, the model is required to generate transition motion from the starting pose to the target pose. We split the test set into 30k 3-second" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 223, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 223, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 223, + 91, + 447, + 102 + ], + "type": "text", + "content": "Revisit Human-Scene Interaction via Space Occupancy" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 474, + 92, + 480, + 100 + ], + "type": "text", + "content": "9" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 142, + 114, + 471, + 311 + ], + "blocks": [ + { + "bbox": [ + 142, + 114, + 471, + 311 + ], + "lines": [ + { + "bbox": [ + 142, + 114, + 471, + 311 + ], + "spans": [ + { + "bbox": [ + 142, + 114, + 471, + 311 + ], + "type": "image", + "image_path": "e9616a1ae9e7663d319be7e5a38fcc42cffcfdf67a603753a4251b8e942e315f.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 319, + 482, + 365 + ], + "lines": [ + { + "bbox": [ + 130, + 319, + 482, + 365 + ], + "spans": [ + { + "bbox": [ + 130, + 319, + 482, + 365 + ], + "type": "text", + "content": "Fig. 5: Our controller can naturally reach the target (white to blue) given the complex occupancy. Furthermore, it can stabilize the motions (red to white) around the targets after reaching them. Previous SOTA [60] suffers from severe penetration and fails to reach the target. For simplicity, the ceilings of the occupancy are hidden." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 389, + 482, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 389, + 482, + 628 + ], + "spans": [ + { + "bbox": [ + 130, + 389, + 482, + 628 + ], + "type": "text", + "content": "clips, adopting the first frame as the starting pose and the final frame as the target pose. The models are required to generate motion for 10s towards the target. There are two reasons for this setting. First, we would like to provide enough time for the models to accomplish the tasks. Second, we are also curious about how the models would behave after the target is reached and no new target is given, or when a very close target is given. The results are evaluated from three aspects. Task Completion is determined by whether the target pose is reached. The target is identified as completed if the average Euclidean distance between the generated and target pose is lower than 20 centimeters and less than 50 voxels are penetrated. We report the success rate (Suc.). The average minimum distance to the target pose (DT) and the time taken to reach the target (Time) are reported for reference. Note that Time is only calculated for successful samples. Motion Quality is evaluated via foot sliding (FS) and penetration (PEN) detection. A frame is identified as FS with its foot velocity " + }, + { + "bbox": [ + 130, + 389, + 482, + 628 + ], + "type": "inline_equation", + "content": ">7.5\\mathrm{cm / s}" + }, + { + "bbox": [ + 130, + 389, + 482, + 628 + ], + "type": "text", + "content": " following GAMMA [58]. We report the percentage of frames with FS. For PEN, we report the number of voxels penetrated per frame. Similarity to Dataset is measured by the similarity between the predicted and GT target end-effector trajectory. We report Edit distance for the real penalty (ERP) [5] since the trajectories might have different lengths. It is only given as a reference, which is not the lower the better." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 629, + 482, + 666 + ], + "type": "text", + "content": "Quantitative results are shown in Tab. 1. For DIMOS [60], we convert the target pose into corresponding marker sets and report the best results from different policies for each test sample. For w/o Occ., we remove the canonical" + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "10" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 205, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 205, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 205, + 100 + ], + "type": "text", + "content": "Liu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 137, + 114, + 219, + 186 + ], + "blocks": [ + { + "bbox": [ + 137, + 114, + 219, + 186 + ], + "lines": [ + { + "bbox": [ + 137, + 114, + 219, + 186 + ], + "spans": [ + { + "bbox": [ + 137, + 114, + 219, + 186 + ], + "type": "image", + "image_path": "87245a5bd86dad665600291e9b5485cc67104ca506032de232652b4493d7de8b.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 133, + 186, + 223, + 199 + ], + "lines": [ + { + "bbox": [ + 133, + 186, + 223, + 199 + ], + "spans": [ + { + "bbox": [ + 133, + 186, + 223, + 199 + ], + "type": "text", + "content": "Space Occupancy overlay DIMOS Walkable Area from NavMesh" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 226, + 114, + 306, + 186 + ], + "blocks": [ + { + "bbox": [ + 226, + 114, + 306, + 186 + ], + "lines": [ + { + "bbox": [ + 226, + 114, + 306, + 186 + ], + "spans": [ + { + "bbox": [ + 226, + 114, + 306, + 186 + ], + "type": "image", + "image_path": "d334ea9aa2a44a4d217eb82acbab90ea5f056d02f16ea15cb93a3c936e47bd0a.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 231, + 189, + 301, + 196 + ], + "lines": [ + { + "bbox": [ + 231, + 189, + 301, + 196 + ], + "spans": [ + { + "bbox": [ + 231, + 189, + 301, + 196 + ], + "type": "text", + "content": "DIMOS path planning results" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 307, + 125, + 335, + 134 + ], + "lines": [ + { + "bbox": [ + 307, + 125, + 335, + 134 + ], + "spans": [ + { + "bbox": [ + 307, + 125, + 335, + 134 + ], + "type": "text", + "content": "Sit Down" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 339, + 114, + 481, + 200 + ], + "blocks": [ + { + "bbox": [ + 339, + 114, + 481, + 200 + ], + "lines": [ + { + "bbox": [ + 339, + 114, + 481, + 200 + ], + "spans": [ + { + "bbox": [ + 339, + 114, + 481, + 200 + ], + "type": "image", + "image_path": "5e84eab1357ea931c3eeaa4888955163dffd3b6d6a527245a9c42e09ed722657.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 208, + 482, + 253 + ], + "lines": [ + { + "bbox": [ + 130, + 208, + 482, + 253 + ], + "spans": [ + { + "bbox": [ + 130, + 208, + 482, + 253 + ], + "type": "text", + "content": "Fig. 6: A typical DIMOS [60] failure case. Fig. 7: With the unified view of human-The NavMesh algorithm produces sub- occupancy interaction, the same controller optimal walkable areas, resulting in path- in Fig. 5 produces diverse motions, includ-finding failures for " + }, + { + "bbox": [ + 130, + 208, + 482, + 253 + ], + "type": "inline_equation", + "content": "69.43\\%" + }, + { + "bbox": [ + 130, + 208, + 482, + 253 + ], + "type": "text", + "content": " MOB samples. ing sitting down, crawling, and getting up." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 281, + 482, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 281, + 482, + 437 + ], + "spans": [ + { + "bbox": [ + 130, + 281, + 482, + 437 + ], + "type": "text", + "content": "occupancy-related components. Previous efforts frequently fail to reach the goal successfully in complex scenarios, which we will elucidate later with qualitative results. Even when the goal is reached, severe penetration emerges, resulting in low success rates in Tab. 1. Our models outperform DIMOS by a considerable margin with a single controller compared to multiple policies. This reveals that the straightforward voxel-based representation is capable of encoding holistic information for human-occupancy interaction. The model without occupancy manages to approach the target as closely and quickly as possible. However, it suffers from more severe interpenetration with the occupancy, resulting in a lower success rate. Instead, our model with consideration of occupancy behaves more conservatively, spending more time to accomplish the tasks with less penetration. A slightly higher foot sliding rate is also observed, which is a trade-off for the penetration-avoiding ability that field regulation brings." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 438, + 482, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 438, + 482, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 438, + 482, + 666 + ], + "type": "text", + "content": "Qualitative results are shown in Fig. 5, 7. Given the complex occupancy, previous state-of-the-art DIMOS [60] frequently fails to reach the goal. Even when it reaches the goal, severe penetration is usually observed. Instead, our controller generates feasible motions to reach the targets. Previous methods' typical locomotion-interaction decomposition [44, 58, 60] tends to over-simplify both. For locomotion, most methods assume a rigid character (usually simplified as a rigid cylinder in " + }, + { + "bbox": [ + 130, + 438, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathrm{A}^*" + }, + { + "bbox": [ + 130, + 438, + 482, + 666 + ], + "type": "text", + "content": " [11, 44] and NavMesh algorithms [60]), overseeing the flexibility of humans. It might be okay for spacious rooms, however, when posed in a cramped scene or scene with complex geometric structures like in Fig. 5, models could fail in path-finding and thus degenerate. As shown in Fig. 6, for " + }, + { + "bbox": [ + 130, + 438, + 482, + 666 + ], + "type": "inline_equation", + "content": "69.43\\%" + }, + { + "bbox": [ + 130, + 438, + 482, + 666 + ], + "type": "text", + "content": " MOB samples, the NavMesh algorithm adopted by DIMOS failed to generate a feasible path from the start position to the target position. The statistic for " + }, + { + "bbox": [ + 130, + 438, + 482, + 666 + ], + "type": "inline_equation", + "content": "\\mathrm{A}^*" + }, + { + "bbox": [ + 130, + 438, + 482, + 666 + ], + "type": "text", + "content": " from SAMP [11] is " + }, + { + "bbox": [ + 130, + 438, + 482, + 666 + ], + "type": "inline_equation", + "content": "70.29\\%" + }, + { + "bbox": [ + 130, + 438, + 482, + 666 + ], + "type": "text", + "content": ". This explains the unsatisfying quantitative results of DIMOS. For interaction, the categories are typically limited to predefined ones. In contrast, with our unified view of Human-Occupancy interaction, our controller applies to various complex scenarios beyond simplified locomotion and interaction. Furthermore, diverse motions could be performed by the same controller (Fig. 7). By comparing our controllers with and without occupancy, we find that training with occupancy helps to avoid penetration significantly." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 223, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 223, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 223, + 91, + 447, + 102 + ], + "type": "text", + "content": "Revisit Human-Scene Interaction via Space Occupancy" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 479, + 100 + ], + "type": "text", + "content": "11" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 170, + 114, + 447, + 156 + ], + "blocks": [ + { + "bbox": [ + 170, + 114, + 447, + 156 + ], + "lines": [ + { + "bbox": [ + 170, + 114, + 447, + 156 + ], + "spans": [ + { + "bbox": [ + 170, + 114, + 447, + 156 + ], + "type": "table", + "html": "
ModelSuc. (%)Dist. to goal (cm)Collision depth (cm)FS (%)
GT100.000.008.742.45
Ours (w/ MOB only)73.017.199.5613.83
Ours (w/ CIRCLE only)75.526.8813.7411.76
Ours (w/ MOB + CIRCLE)78.976.7311.8512.87
", + "image_path": "bcaacbbd69fa041030a9d5a6713e2485d6997e79e69fa17b2e2181fe2acff799.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 237, + 157, + 375, + 167 + ], + "lines": [ + { + "bbox": [ + 237, + 157, + 375, + 167 + ], + "spans": [ + { + "bbox": [ + 237, + 157, + 375, + 167 + ], + "type": "text", + "content": "Table 2: Results on CIRCLE [1]." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 157, + 188, + 278, + 284 + ], + "blocks": [ + { + "bbox": [ + 157, + 188, + 278, + 284 + ], + "lines": [ + { + "bbox": [ + 157, + 188, + 278, + 284 + ], + "spans": [ + { + "bbox": [ + 157, + 188, + 278, + 284 + ], + "type": "image", + "image_path": "ad6c5920472bc89dea76c37b86bcb6c6374941503bec0779fe3945dca69b5970.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 295, + 483, + 340 + ], + "lines": [ + { + "bbox": [ + 130, + 295, + 483, + 340 + ], + "spans": [ + { + "bbox": [ + 130, + 295, + 483, + 340 + ], + "type": "text", + "content": "Fig. 8: We can generate long-term motion. Fig. 9: We synthesize Human-Object Interactions in realistic rooms: A human circulation action, with hand trajectory as the control navigates a cabinet, sits on a sofa, perches signal. The human successfully moves a floor on a bed, and then resumes walking. lamp to a new place." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 329, + 188, + 454, + 287 + ], + "blocks": [ + { + "bbox": [ + 329, + 188, + 454, + 287 + ], + "lines": [ + { + "bbox": [ + 329, + 188, + 454, + 287 + ], + "spans": [ + { + "bbox": [ + 329, + 188, + 454, + 287 + ], + "type": "image", + "image_path": "fc65b660786abf47cdc3ce3fdb42a77038104bc6c1a9346444a5bf7a395ba770.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 364, + 480, + 388 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 364, + 480, + 388 + ], + "spans": [ + { + "bbox": [ + 130, + 364, + 480, + 388 + ], + "type": "text", + "content": "Also, the controller with occupancy could stabilize the motion around the target after reaching it, while its counterpart tends to wander further inertially." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 406, + 312, + 418 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 406, + 312, + 418 + ], + "spans": [ + { + "bbox": [ + 131, + 406, + 312, + 418 + ], + "type": "text", + "content": "5.2 Interacting with Static Scenes" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 426, + 480, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 426, + 480, + 449 + ], + "spans": [ + { + "bbox": [ + 130, + 426, + 480, + 449 + ], + "type": "text", + "content": "Once our controller is trained on the more complex MOB, it could be effortlessly applied to realistic synthesized rooms and scene scans." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 450, + 481, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 450, + 481, + 628 + ], + "spans": [ + { + "bbox": [ + 130, + 450, + 481, + 628 + ], + "type": "text", + "content": "Quantitative results is reported on CIRCLE [1]. Given the initial full-body pose and the target right wrist position, the model is required to generate a right-hand reaching motion. Following [1], success rate, distance to the goal, cumulative collision depth, and foot sliding are reported as metrics. Besides the controller trained on MOB (excluding CIRCLE), we also train our controller on CIRCLE-only and MOB+CIRCLE. We randomly split CIRCLE into a 2,565-sample train set (for the ablative baseline only) and a 453-sample test set since the original split has not been made public yet. The results in Tab. 2 reveal that our controller could handle various tasks in complex natural scenes well. Also, its potential to advance HOI-related motion, like hand-reaching before grasping, is showcased. It is noticeable that the MOB-only model demonstrates a competitive success rate and better collision avoidance, revealing the efficacy of MOB with its advanced complexity. When incorporating MOB with CIRCLE, superior task completion is observed, showing that MOB could function as an effective auxiliary resource for the conventional HSI datasets." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 130, + 630, + 481, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 630, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 630, + 481, + 665 + ], + "type": "text", + "content": "Qualitative results. We test our model in a synthetic room from DIMOS [60] in Fig. 8 and a realistic scan from Replica [37] in Fig. 1. The occupancy is transferred from the scene SDFs. Our model produces convincing motions, nav-" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 205, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 205, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 205, + 100 + ], + "type": "text", + "content": "Liu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 155, + 114, + 459, + 175 + ], + "blocks": [ + { + "bbox": [ + 155, + 114, + 459, + 175 + ], + "lines": [ + { + "bbox": [ + 155, + 114, + 459, + 175 + ], + "spans": [ + { + "bbox": [ + 155, + 114, + 459, + 175 + ], + "type": "image", + "image_path": "095890d1c8cf55261fd0b148674fa8869e946b23356fefdb98c2c43f721a70c0.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 185, + 480, + 197 + ], + "lines": [ + { + "bbox": [ + 132, + 185, + 480, + 197 + ], + "spans": [ + { + "bbox": [ + 132, + 185, + 480, + 197 + ], + "type": "text", + "content": "Fig. 10: Our model can avoid dynamic obstacles and reasonably reach the target circle." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 220, + 482, + 353 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 220, + 482, + 353 + ], + "spans": [ + { + "bbox": [ + 130, + 220, + 482, + 353 + ], + "type": "text", + "content": "igates around obstacles, and transitions seamlessly between locomotion and interaction. Remarkably, it adapts to sitting on various objects, such as sofas, beds, and stools, revealing its versatility in recognizing different occupancies as potential seats. The test room features sparse furnishings, highlighting the model's ability to operate in open spaces diverging from MOB samples. Conversely, the Replica scene presents a more cluttered setup. Here, the model adeptly maneuvers through tight spaces and produces plausible movements. Furthermore, our model can generate long-term motions. The motions in Fig. 1 and Fig. 8 are generated in a single attempt by merely altering the target, both spanning around 30 seconds, which are clipped from longer motions for clear illustration in the limited space. More results are in the supplementary materials." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 371, + 285, + 384 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 371, + 285, + 384 + ], + "spans": [ + { + "bbox": [ + 132, + 371, + 285, + 384 + ], + "type": "text", + "content": "5.3 Interacting with Objects" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 392, + 481, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 392, + 481, + 453 + ], + "spans": [ + { + "bbox": [ + 130, + 392, + 481, + 453 + ], + "type": "text", + "content": "Our flexible control signal for target pose enables HOI motion synthesis. By fine-tuning on OMOMO [20], given the initial pose and hand trajectories, we synthesize plausible motions to interact with objects as in Figure 9, which exhibits rational full-body motion w.r.t. hand and object trajectories. More results are in the supplementary." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 472, + 328, + 484 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 472, + 328, + 484 + ], + "spans": [ + { + "bbox": [ + 132, + 472, + 328, + 484 + ], + "type": "text", + "content": "5.4 Interacting with Dynamic Scenes" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 493, + 482, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 493, + 482, + 601 + ], + "spans": [ + { + "bbox": [ + 130, + 493, + 482, + 601 + ], + "type": "text", + "content": "During training, the canonical occupancy perceived by our controller is dynamically changing w.r.t. to the human movement, enabling it to handle dynamic scenarios though trained on essentially static MOB. We demonstrate the interaction in dynamic scenes with a custom automatic door scenario in Fig. 10. With the target (the circle) in front of the human (a) blocked by a revolving door, we control the human to adapt its speed w.r.t. the door till enough clearance and halt precisely at the target. This sequence exemplifies the model's ability to produce realistic and contextually appropriate motion in dynamic environments. More details are in the supplementary." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 620, + 246, + 631 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 620, + 246, + 631 + ], + "spans": [ + { + "bbox": [ + 132, + 620, + 246, + 631 + ], + "type": "text", + "content": "5.5 Ablation Studies" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 130, + 641, + 481, + 666 + ], + "type": "text", + "content": "With our controller as the baseline, selected ablation studies on MOB are reported in Tab. 1. Complete results and details are included in the supplementary." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 223, + 91, + 447, + 103 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 223, + 91, + 447, + 103 + ], + "spans": [ + { + "bbox": [ + 223, + 91, + 447, + 103 + ], + "type": "text", + "content": "Revisit Human-Scene Interaction via Space Occupancy" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 91, + 481, + 100 + ], + "type": "text", + "content": "13" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 479, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 479, + 175 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 479, + 175 + ], + "type": "text", + "content": "Different loss terms. By removing penetration loss " + }, + { + "bbox": [ + 130, + 116, + 479, + 175 + ], + "type": "inline_equation", + "content": "L_{pen}" + }, + { + "bbox": [ + 130, + 116, + 479, + 175 + ], + "type": "text", + "content": ", a major degeneration in Success rate and PEN is observed. This shows that explicitly punishing penetration is necessary. The removal of " + }, + { + "bbox": [ + 130, + 116, + 479, + 175 + ], + "type": "inline_equation", + "content": "L_{field}" + }, + { + "bbox": [ + 130, + 116, + 479, + 175 + ], + "type": "text", + "content": " also hurts the success rate by " + }, + { + "bbox": [ + 130, + 116, + 479, + 175 + ], + "type": "inline_equation", + "content": "2.57\\%" + }, + { + "bbox": [ + 130, + 116, + 479, + 175 + ], + "type": "text", + "content": ", with PEN increasing by 10.62. FS also increases notably. We identify that without " + }, + { + "bbox": [ + 130, + 116, + 479, + 175 + ], + "type": "inline_equation", + "content": "L_{field}" + }, + { + "bbox": [ + 130, + 116, + 479, + 175 + ], + "type": "text", + "content": ", the field regulation would generate unnatural spikes." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 175, + 479, + 209 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 175, + 479, + 209 + ], + "spans": [ + { + "bbox": [ + 130, + 175, + 479, + 209 + ], + "type": "text", + "content": "Field Regulation (FR) is also evaluated. Eliminating the field regulation decreases the success rate slightly, with a notable PEN increase (+18.28) and FS decrease (-1.56%), revealing the tradeoff it conducts between PEN and FS." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 212, + 480, + 270 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 212, + 480, + 270 + ], + "spans": [ + { + "bbox": [ + 130, + 212, + 480, + 270 + ], + "type": "text", + "content": "Occupancy representation. We replaced the voxel-based representation of canonical occupancy with the more compact and efficient BPS [34]. The field regulation is also modified accordingly. Degradation is observed in most metrics. Despite its efficiency, BPS might ignore too much information and bring instability to the field regulation." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 288, + 217, + 300 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 288, + 217, + 300 + ], + "spans": [ + { + "bbox": [ + 131, + 288, + 217, + 300 + ], + "type": "text", + "content": "6 Discussion" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 311, + 480, + 478 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 311, + 480, + 478 + ], + "spans": [ + { + "bbox": [ + 130, + 311, + 480, + 478 + ], + "type": "text", + "content": "Limitations. First, when interacting with dynamic scenes, our controller could perform sub-optimally with slight collisions. This could be ascribed to the space occupancy changing rate gap between training data and testing scenes. During training, the changing rate is determined by the human locomotion speed only. While for dynamic scenes, there exists additional space occupancy speed. Mitigating this gap is worth exploring. Second, foot sliding is another artifact that could happen. However, the metrics are based on simple heuristics with gaps compared to real foot sliding. A reasonable and effective tool to address this would be promising. Third, the current MOB poses its main focus on static HSI. Despite the proposed pipeline that could apply to HOI motion synthesis, further exploration of the space-occupancy view of HOI would be an interesting future direction. Finally, the proposed versatile controller only accepts target poses, space occupancy, and joint trajectories as control signals. Extending the controller for more signals would be promising." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 494, + 479, + 541 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 494, + 479, + 541 + ], + "spans": [ + { + "bbox": [ + 130, + 494, + 479, + 541 + ], + "type": "text", + "content": "Broader Impacts. The exploration of HSI is beneficial for navigation in complex environments, with significant impacts on Embodied AI and Robotics. Also, multiple fields like animation production and interior design could be advanced. However, this might also be misused for false information." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 131, + 559, + 219, + 571 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 559, + 219, + 571 + ], + "spans": [ + { + "bbox": [ + 131, + 559, + 219, + 571 + ], + "type": "text", + "content": "7 Conclusion" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 581, + 479, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 581, + 479, + 665 + ], + "spans": [ + { + "bbox": [ + 130, + 581, + 479, + 665 + ], + "type": "text", + "content": "We revisited static HSI from a unified space occupancy view by identifying the major component of static HSI as Human-Occupancy Interaction. Large-scale Human Occupancy Interaction database MOB was aggregated from motion-only data, substantially improving diversity and complexity. Moreover, a versatile auto-regressive controller was trained on MOB for Human-Occupancy Interaction generation, which supports flexible and stable generation for static HSI under a spectrum of scenarios without training with real 3D scene data." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 205, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 205, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 205, + 100 + ], + "type": "text", + "content": "Liu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 133, + 114, + 240, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 114, + 240, + 129 + ], + "spans": [ + { + "bbox": [ + 133, + 114, + 240, + 129 + ], + "type": "text", + "content": "Acknowledgments" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 139, + 481, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 139, + 481, + 186 + ], + "spans": [ + { + "bbox": [ + 132, + 139, + 481, + 186 + ], + "type": "text", + "content": "This work is supported in part by the National Natural Science Foundation of China under Grants No.62306175, No.62302296, National Key Research and Development Project of China (No.2022ZD0160102, No.2021ZD0110704), Shanghai Artificial Intelligence Laboratory, XPLORER PRIZE grants." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 205, + 197, + 217 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 205, + 197, + 217 + ], + "spans": [ + { + "bbox": [ + 133, + 205, + 197, + 217 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 228, + 480, + 665 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 138, + 228, + 480, + 272 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 228, + 480, + 272 + ], + "spans": [ + { + "bbox": [ + 138, + 228, + 480, + 272 + ], + "type": "text", + "content": "1. Araújo, J.P., Li, J., Vetrivel, K., Agarwal, R., Wu, J., Gopinath, D., Clegg, A.W., Liu, K.: Circle: Capture in rich contextual environments. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 21211-21221 (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 274, + 480, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 274, + 480, + 304 + ], + "spans": [ + { + "bbox": [ + 138, + 274, + 480, + 304 + ], + "type": "text", + "content": "2. Athanasiou, N., Petrovich, M., Black, M.J., Varol, G.: Teach: Temporal action composition for 3d humans. In: 2022 International Conference on 3D Vision (3DV). pp. 414-423. IEEE (2022)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 306, + 480, + 337 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 306, + 480, + 337 + ], + "spans": [ + { + "bbox": [ + 138, + 306, + 480, + 337 + ], + "type": "text", + "content": "3. Bengio, S., Vinyals, O., Jaitly, N., Shazeer, N.: Scheduled sampling for sequence prediction with recurrent neural networks. Advances in neural information processing systems 28 (2015)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 338, + 480, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 338, + 480, + 392 + ], + "spans": [ + { + "bbox": [ + 138, + 338, + 480, + 392 + ], + "type": "text", + "content": "4. Cai, Z., Ren, D., Zeng, A., Lin, Z., Yu, T., Wang, W., Fan, X., Gao, Y., Yu, Y., Pan, L., Hong, F., Zhang, M., Loy, C.C., Yang, L., Liu, Z.: Humman: Multi-modal 4d human dataset for versatile sensing and modeling. In: Avidan, S., Brostow, G., Cisse, M., Farinella, G.M., Hassner, T. (eds.) Computer Vision - ECCV 2022. pp. 557-577. Springer Nature Switzerland, Cham (2022)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 138, + 394, + 480, + 424 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 394, + 480, + 424 + ], + "spans": [ + { + "bbox": [ + 138, + 394, + 480, + 424 + ], + "type": "text", + "content": "5. Chen, L., Ng, R.: On the marriage of lp-norms and edit distance. In: Proceedings of the Thirtieth international conference on Very large data bases-Volume 30. pp. 792-803 (2004)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 138, + 426, + 480, + 457 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 426, + 480, + 457 + ], + "spans": [ + { + "bbox": [ + 138, + 426, + 480, + 457 + ], + "type": "text", + "content": "6. Chen, X., Jiang, B., Liu, W., Huang, Z., Fu, B., Chen, T., Yu, G.: Executing your commands via motion diffusion in latent space. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 18000-18010 (2023)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 138, + 459, + 480, + 501 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 459, + 480, + 501 + ], + "spans": [ + { + "bbox": [ + 138, + 459, + 480, + 501 + ], + "type": "text", + "content": "7. Guo, C., Zou, S., Zuo, X., Wang, S., Ji, W., Li, X., Cheng, L.: Generating diverse and natural 3d human motions from text. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 5152-5161 (June 2022)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 138, + 502, + 480, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 502, + 480, + 533 + ], + "spans": [ + { + "bbox": [ + 138, + 502, + 480, + 533 + ], + "type": "text", + "content": "8. Guo, C., Zuo, X., Wang, S., Zou, S., Sun, Q., Deng, A., Gong, M., Cheng, L.: Action2motion: Conditioned generation of 3d human motions. In: Proceedings of the 28th ACM International Conference on Multimedia. pp. 2021-2029 (2020)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 535, + 480, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 535, + 480, + 578 + ], + "spans": [ + { + "bbox": [ + 138, + 535, + 480, + 578 + ], + "type": "text", + "content": "9. Guo, W., Du, Y., Shen, X., Lepetit, V., Alameda-Pineda, X., Moreno-Noguer, F.: Back to mlp: A simple baseline for human motion prediction. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 4809-4819 (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 138, + 579, + 480, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 579, + 480, + 599 + ], + "spans": [ + { + "bbox": [ + 138, + 579, + 480, + 599 + ], + "type": "text", + "content": "0. Harvey, F.G., Yurick, M., Nowrouzezahrai, D., Pal, C.: Robust motion in-between. ACM Transactions on Graphics (TOG) 39(4), 60-1 (2020)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 601, + 480, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 601, + 480, + 632 + ], + "spans": [ + { + "bbox": [ + 138, + 601, + 480, + 632 + ], + "type": "text", + "content": "1. Hassan, M., Ceylan, D., Villegas, R., Saito, J., Yang, J., Zhou, Y., Black, M.J.: Stochastic scene-aware motion prediction. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 11374-11384 (2021)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 138, + 633, + 480, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 633, + 480, + 665 + ], + "spans": [ + { + "bbox": [ + 138, + 633, + 480, + 665 + ], + "type": "text", + "content": "2. Hassan, M., Choutas, V., Tzionas, D., Black, M.J.: Resolving 3D human pose ambiguities with 3D scene constraints. In: International Conference on Computer Vision. pp. 2282-2292 (Oct 2019), https://prox.is.tue.mpg.de" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 223, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 223, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 223, + 91, + 447, + 102 + ], + "type": "text", + "content": "Revisit Human-Scene Interaction via Space Occupancy" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "15" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "type": "text", + "content": "13. Hassan, M., Ghosh, P., Tesch, J., Tzionas, D., Black, M.J.: Populating 3d scenes by learning human-scene interaction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14708-14718 (2021)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 150, + 481, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 150, + 481, + 205 + ], + "spans": [ + { + "bbox": [ + 133, + 150, + 481, + 205 + ], + "type": "text", + "content": "14. Hassan, M., Guo, Y., Wang, T., Black, M., Fidler, S., Peng, X.B.: Synthesizing physical character-scene interactions. In: ACM SIGGRAPH 2023 Conference Proceedings. SIGGRAPH '23, Association for Computing Machinery, New York, NY, USA (2023). https://doi.org/10.1145/3588432.3591525, https://doi.org/10.1145/3588432.3591525" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 206, + 481, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 206, + 481, + 239 + ], + "spans": [ + { + "bbox": [ + 132, + 206, + 481, + 239 + ], + "type": "text", + "content": "15. Hernandez, A., Gall, J., Moreno-Noguer, F.: Human motion prediction via spatiotemporal inpainting. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 7134-7143 (2019)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 239, + 481, + 283 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 239, + 481, + 283 + ], + "spans": [ + { + "bbox": [ + 132, + 239, + 481, + 283 + ], + "type": "text", + "content": "16. Huang, S., Wang, Z., Li, P., Jia, B., Liu, T., Zhu, Y., Liang, W., Zhu, S.C.: Diffusion-based generation, optimization, and planning in 3d scenes. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16750-16761 (2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 285, + 481, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 285, + 481, + 306 + ], + "spans": [ + { + "bbox": [ + 132, + 285, + 481, + 306 + ], + "type": "text", + "content": "17. Hwang, Y.K., Ahuja, N., et al.: A potential field approach to path planning. IEEE transactions on robotics and automation 8(1), 23-32 (1992)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 307, + 481, + 340 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 307, + 481, + 340 + ], + "spans": [ + { + "bbox": [ + 132, + 307, + 481, + 340 + ], + "type": "text", + "content": "18. Ji, Y., Xu, F., Yang, Y., Shen, F., Shen, H.T., Zheng, W.S.: A large-scale rgb-d database for arbitrary-view human action recognition. In: Proceedings of the 26th ACM international Conference on Multimedia. pp. 1510-1518 (2018)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 341, + 481, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 341, + 481, + 373 + ], + "spans": [ + { + "bbox": [ + 132, + 341, + 481, + 373 + ], + "type": "text", + "content": "19. Lee, J., Joo, H.: Locomotion-action-manipulation: Synthesizing human-scene interactions in complex 3d environments. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 9663-9674 (October 2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 374, + 481, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 374, + 481, + 396 + ], + "spans": [ + { + "bbox": [ + 132, + 374, + 481, + 396 + ], + "type": "text", + "content": "20. Li, J., Wu, J., Liu, C.K.: Object motion guided human motion synthesis. ACM Transactions on Graphics (TOG) 42(6), 1-11 (2023)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 397, + 481, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 397, + 481, + 430 + ], + "spans": [ + { + "bbox": [ + 132, + 397, + 481, + 430 + ], + "type": "text", + "content": "21. Li, R., Yang, S., Ross, D.A., Kanazawa, A.: Ai choreographer: Music conditioned 3d dance generation with aist++. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 13401-13412 (2021)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 430, + 481, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 430, + 481, + 453 + ], + "spans": [ + { + "bbox": [ + 132, + 430, + 481, + 453 + ], + "type": "text", + "content": "22. Li, R., Yang, S., Ross, D.A., Kanazawa, A.: Learn to dance with aist++: Music conditioned 3d dance generation. arXiv preprint arXiv:2101.08779 2(3) (2021)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 453, + 481, + 486 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 453, + 481, + 486 + ], + "spans": [ + { + "bbox": [ + 132, + 453, + 481, + 486 + ], + "type": "text", + "content": "23. Li, Y.L., Liu, X., Lu, H., Wang, S., Liu, J., Li, J., Lu, C.: Detailed 2d-3d joint representation for human-object interaction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10166–10175 (2020)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 487, + 481, + 518 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 487, + 481, + 518 + ], + "spans": [ + { + "bbox": [ + 132, + 487, + 481, + 518 + ], + "type": "text", + "content": "24. Lin, J., Zeng, A., Lu, S., Cai, Y., Zhang, R., Wang, H., Zhang, L.: Motion-x: A large-scale 3d expressive whole-body human motion dataset. arXiv preprint arXiv:2307.00818 (2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 520, + 481, + 552 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 520, + 481, + 552 + ], + "spans": [ + { + "bbox": [ + 132, + 520, + 481, + 552 + ], + "type": "text", + "content": "25. Loper, M., Mahmood, N., Romero, J., Pons-Moll, G., Black, M.J.: Smpl: A skinned multi-person linear model. ACM transactions on graphics (TOG) 34(6), 1-16 (2015)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 554, + 481, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 554, + 481, + 597 + ], + "spans": [ + { + "bbox": [ + 132, + 554, + 481, + 597 + ], + "type": "text", + "content": "26. Lucas, T., Baradel, F., Weinzaepfel, P., Rogez, G.: Posegpt: quantization-based 3d human motion generation and forecasting. In: Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part VI. pp. 417-435. Springer (2022)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 599, + 481, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 599, + 481, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 599, + 481, + 632 + ], + "type": "text", + "content": "27. Mahmood, N., Ghorbani, N., Troje, N.F., Pons-Moll, G., Black, M.J.: Amass: Archive of motion capture as surface shapes. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 5442-5451 (2019)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 632, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 481, + 665 + ], + "type": "text", + "content": "28. Nie, Y., Dai, A., Han, X., Nießner, M.: Pose2room: understanding 3d scenes from human activities. In: European Conference on Computer Vision. pp. 425-443. Springer (2022)" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 205, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 205, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 205, + 100 + ], + "type": "text", + "content": "Liu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 482, + 666 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 132, + 116, + 482, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 116, + 482, + 162 + ], + "spans": [ + { + "bbox": [ + 132, + 116, + 482, + 162 + ], + "type": "text", + "content": "29. Pavlakos, G., Choutas, V., Ghorbani, N., Bolkart, T., Osman, A.A.A., Tzionas, D., Black, M.J.: Expressive body capture: 3D hands, face, and body from a single image. In: Proceedings IEEE Conf. on Computer Vision and Pattern Recognition (CVPR). pp. 10975-10985 (2019)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 163, + 481, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 163, + 481, + 195 + ], + "spans": [ + { + "bbox": [ + 132, + 163, + 481, + 195 + ], + "type": "text", + "content": "30. Petrovich, M., Black, M.J., Varol, G.: Action-conditioned 3d human motion synthesis with transformer vae. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 10985-10995 (2021)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 196, + 481, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 196, + 481, + 239 + ], + "spans": [ + { + "bbox": [ + 132, + 196, + 481, + 239 + ], + "type": "text", + "content": "31. Petrovich, M., Black, M.J., Varol, G.: Temos: Generating diverse human motions from textual descriptions. In: Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXII. pp. 480-497. Springer (2022)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 240, + 481, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 240, + 481, + 274 + ], + "spans": [ + { + "bbox": [ + 132, + 240, + 481, + 274 + ], + "type": "text", + "content": "32. Petrovich, M., Black, M.J., Varol, G.: Tmr: Text-to-motion retrieval using contrastive 3d human motion synthesis. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 9488-9497 (October 2023)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 274, + 481, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 274, + 481, + 297 + ], + "spans": [ + { + "bbox": [ + 132, + 274, + 481, + 297 + ], + "type": "text", + "content": "33. Plappert, M., Mandery, C., Asfour, T.: The kit motion-language dataset. Big data 4(4), 236-252 (2016)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 297, + 481, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 297, + 481, + 330 + ], + "spans": [ + { + "bbox": [ + 132, + 297, + 481, + 330 + ], + "type": "text", + "content": "34. Prokudin, S., Lassner, C., Romero, J.: Efficient learning on point clouds with basis point sets. In: Proceedings of the IEEE International Conference on Computer Vision. pp. 4332-4341 (2019)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 331, + 481, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 331, + 481, + 385 + ], + "spans": [ + { + "bbox": [ + 132, + 331, + 481, + 385 + ], + "type": "text", + "content": "35. Puig, X., Undersander, E., Szot, A., Cote, M.D., Partsey, R., Yang, J., Desai, R., Clegg, A.W., Hlavac, M., Min, T., Gervet, T., Vondruš, V., Berges, V.P., Turner, J., Maksymets, O., Kira, Z., Kalakrishnan, M., Malik, J., Chaplot, D.S., Jain, U., Batra, D., Rai, A., Mottaghi, R.: Habitat 3.0: A co-habitat for humans, avatars and robots (2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 386, + 481, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 386, + 481, + 408 + ], + "spans": [ + { + "bbox": [ + 132, + 386, + 481, + 408 + ], + "type": "text", + "content": "36. Starke, S., Zhang, H., Komura, T., Saito, J.: Neural state machine for character-scene interactions. ACM Trans. Graph. 38(6), 209-1 (2019)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 409, + 481, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 409, + 481, + 442 + ], + "spans": [ + { + "bbox": [ + 132, + 409, + 481, + 442 + ], + "type": "text", + "content": "37. Straub, J., Whelan, T., Ma, L., Chen, Y., Wijmans, E., Green, S., Engel, J.J., Mur-Artal, R., Ren, C., Verma, S., et al.: The replica dataset: A digital replica of indoor spaces. arXiv preprint arXiv:1906.05797 (2019)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 443, + 481, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 443, + 481, + 498 + ], + "spans": [ + { + "bbox": [ + 132, + 443, + 481, + 498 + ], + "type": "text", + "content": "38. Szot, A., Clegg, A., Undersander, E., Wijmans, E., Zhao, Y., Turner, J., Maestre, N., Mukadam, M., Chaplot, D., Maksymets, O., Gokaslan, A., Vondrus, V., Dharur, S., Meier, F., Galuba, W., Chang, A., Kira, Z., Koltun, V., Malik, J., Savva, M., Batra, D.: Habitat 2.0: Training home assistants to rearrange their habitat. In: Advances in Neural Information Processing Systems (NeurIPS) (2021)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 498, + 481, + 532 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 498, + 481, + 532 + ], + "spans": [ + { + "bbox": [ + 132, + 498, + 481, + 532 + ], + "type": "text", + "content": "39. Taheri, O., Choutas, V., Black, M.J., Tzionas, D.: Goal: Generating 4d whole-body motion for hand-object grasping. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13263-13273 (2022)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 533, + 481, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 533, + 481, + 586 + ], + "spans": [ + { + "bbox": [ + 132, + 533, + 481, + 586 + ], + "type": "text", + "content": "40. Tessler, C., Kasten, Y., Guo, Y., Mannor, S., Chechik, G., Peng, X.B.: Calm: Conditional adversarial latent models for directable virtual characters. In: ACM SIGGRAPH 2023 Conference Proceedings. SIGGRAPH '23, Association for Computing Machinery, New York, NY, USA (2023). https://doi.org/10.1145/3588432.3591541, https://doi.org/10.1145/3588432.3591541" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 587, + 481, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 587, + 481, + 632 + ], + "spans": [ + { + "bbox": [ + 132, + 587, + 481, + 632 + ], + "type": "text", + "content": "41. Tevet, G., Gordon, B., Hertz, A., Bermano, A.H., Cohen-Or, D.: Motionclip: Exposing human motion generation to clip space. In: Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXII. pp. 358-374. Springer (2022)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 632, + 481, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 632, + 481, + 666 + ], + "spans": [ + { + "bbox": [ + 132, + 632, + 481, + 666 + ], + "type": "text", + "content": "42. Tevet, G., Raab, S., Gordon, B., Shafir, Y., Cohen-or, D., Bermano, A.H.: Human motion diffusion model. In: The Eleventh International Conference on Learning Representations (2022)" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 223, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 223, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 223, + 91, + 447, + 102 + ], + "type": "text", + "content": "Revisit Human-Scene Interaction via Space Occupancy" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 116, + 481, + 665 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "spans": [ + { + "bbox": [ + 133, + 116, + 481, + 149 + ], + "type": "text", + "content": "43. Wang, J., Xu, H., Xu, J., Liu, S., Wang, X.: Synthesizing long-term 3d human motion and interaction in 3d scenes. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9401-9411 (2021)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 149, + 481, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 149, + 481, + 182 + ], + "spans": [ + { + "bbox": [ + 133, + 149, + 481, + 182 + ], + "type": "text", + "content": "44. Wang, J., Rong, Y., Liu, J., Yan, S., Lin, D., Dai, B.: Towards diverse and natural scene-aware 3d human motion synthesis. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 20460-20469 (2022)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 182, + 481, + 213 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 182, + 481, + 213 + ], + "spans": [ + { + "bbox": [ + 132, + 182, + 481, + 213 + ], + "type": "text", + "content": "45. Wang, Z., Chen, Y., Liu, T., Zhu, Y., Liang, W., Huang, S.: Humanise: Language-conditioned human motion generation in 3d scenes. Advances in Neural Information Processing Systems 35, 14959-14971 (2022)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 213, + 481, + 246 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 213, + 481, + 246 + ], + "spans": [ + { + "bbox": [ + 132, + 213, + 481, + 246 + ], + "type": "text", + "content": "46. Wu, Y., Wang, J., Zhang, Y., Zhang, S., Hilliges, O., Yu, F., Tang, S.: Saga: Stochastic whole-body grasping with contact. In: European Conference on Computer Vision. pp. 257-274. Springer (2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 246, + 481, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 246, + 481, + 300 + ], + "spans": [ + { + "bbox": [ + 132, + 246, + 481, + 300 + ], + "type": "text", + "content": "47. Xu, L., Song, Z., Wang, D., Su, J., Fang, Z., Ding, C., Gan, W., Yan, Y., Jin, X., Yang, X., Zeng, W., Wu, W.: Actformer: A gan-based transformer towards general action-conditioned 3d human motion generation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 2228-2238 (October 2023)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 300, + 481, + 320 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 300, + 481, + 320 + ], + "spans": [ + { + "bbox": [ + 132, + 300, + 481, + 320 + ], + "type": "text", + "content": "48. Ye, S., Wang, Y., Li, J., Park, D., Liu, C.K., Xu, H., Wu, J.: Scene synthesis from human motion. In: SIGGRAPH Asia 2022 Conference Papers (2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 320, + 481, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 320, + 481, + 365 + ], + "spans": [ + { + "bbox": [ + 132, + 320, + 481, + 365 + ], + "type": "text", + "content": "49. Yi, H., Huang, C.H.P., Tripathi, S., Hering, L., Thies, J., Black, M.J.: Mime: Human-aware 3d scene generation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 12965-12976 (June 2023)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 365, + 481, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 365, + 481, + 407 + ], + "spans": [ + { + "bbox": [ + 132, + 365, + 481, + 407 + ], + "type": "text", + "content": "50. Yi, H., Huang, C.H.P., Tzionas, D., Kocabas, M., Hassan, M., Tang, S., Thies, J., Black, M.J.: Human-aware object placement for visual environment reconstruction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3959-3970 (2022)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 407, + 481, + 451 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 407, + 481, + 451 + ], + "spans": [ + { + "bbox": [ + 132, + 407, + 481, + 451 + ], + "type": "text", + "content": "51. Zhang, J.Y., Pepose, S., Joo, H., Ramanan, D., Malik, J., Kanazawa, A.: Perceiving 3d human-object spatial arrangements from a single image in the wild. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XII 16. pp. 34-51. Springer (2020)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 132, + 450, + 481, + 494 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 450, + 481, + 494 + ], + "spans": [ + { + "bbox": [ + 132, + 450, + 481, + 494 + ], + "type": "text", + "content": "52. Zhang, J., Zhang, Y., Cun, X., Zhang, Y., Zhao, H., Lu, H., Shen, X., Shan, Y.: Generating human motion from textual descriptions with discrete representations. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 14730-14740 (June 2023)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 493, + 481, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 493, + 481, + 525 + ], + "spans": [ + { + "bbox": [ + 132, + 493, + 481, + 525 + ], + "type": "text", + "content": "53. Zhang, M., Cai, Z., Pan, L., Hong, F., Guo, X., Yang, L., Liu, Z.: Motiondiffuse: Text-driven human motion generation with diffusion model. arXiv preprint arXiv:2208.15001 (2022)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 525, + 481, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 525, + 481, + 568 + ], + "spans": [ + { + "bbox": [ + 132, + 525, + 481, + 568 + ], + "type": "text", + "content": "54. Zhang, M., Guo, X., Pan, L., Cai, Z., Hong, F., Li, H., Yang, L., Liu, Z.: Remodiffuse: Retrieval-augmented motion diffusion model. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 364-373 (October 2023)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 132, + 568, + 481, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 568, + 481, + 601 + ], + "spans": [ + { + "bbox": [ + 132, + 568, + 481, + 601 + ], + "type": "text", + "content": "55. Zhang, S., Zhang, Y., Ma, Q., Black, M.J., Tang, S.: Place: Proximity learning of articulation and contact in 3d environments. In: 2020 International Conference on 3D Vision (3DV). pp. 642-651. IEEE (2020)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 132, + 601, + 481, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 601, + 481, + 633 + ], + "spans": [ + { + "bbox": [ + 132, + 601, + 481, + 633 + ], + "type": "text", + "content": "56. Zhang, X., Bhatnagar, B.L., Starke, S., Guzov, V., Pons-Moll, G.: Couch: Towards controllable human-chair interactions. In: European Conference on Computer Vision. pp. 518-535. Springer (2022)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 132, + 633, + 481, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 633, + 481, + 665 + ], + "spans": [ + { + "bbox": [ + 132, + 633, + 481, + 665 + ], + "type": "text", + "content": "57. Zhang, Y., Hassan, M., Neumann, H., Black, M.J., Tang, S.: Generating 3d people in scenes without people. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 6194-6204 (2020)" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "spans": [ + { + "bbox": [ + 133, + 91, + 144, + 100 + ], + "type": "text", + "content": "18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 166, + 91, + 205, + 100 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 91, + 205, + 100 + ], + "spans": [ + { + "bbox": [ + 166, + 91, + 205, + 100 + ], + "type": "text", + "content": "Liu et al." + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + }, + { + "para_blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 281 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 130, + 116, + 482, + 149 + ], + "type": "text", + "content": "58. Zhang, Y., Tang, S.: The wanderings of odysseus in 3d scenes. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 20481-20491 (2022)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 149, + 482, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 149, + 482, + 182 + ], + "spans": [ + { + "bbox": [ + 130, + 149, + 482, + 182 + ], + "type": "text", + "content": "59. Zhao, K., Wang, S., Zhang, Y., Beeler, T., Tang, S.: Compositional human-scene interaction synthesis with semantic control. In: European Conference on Computer Vision. pp. 311-327. Springer (2022)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 182, + 482, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 182, + 482, + 215 + ], + "spans": [ + { + "bbox": [ + 130, + 182, + 482, + 215 + ], + "type": "text", + "content": "60. Zhao, K., Zhang, Y., Wang, S., Beeler, T., Tang, S.: Synthesizing diverse human motions in 3d indoor scenes. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 14738-14749 (October 2023)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 215, + 482, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 215, + 482, + 248 + ], + "spans": [ + { + "bbox": [ + 130, + 215, + 482, + 248 + ], + "type": "text", + "content": "61. Zheng, Y., Yang, Y., Mo, K., Li, J., Yu, T., Liu, Y., Liu, C.K., Guibas, L.J.: Gimo: Gaze-informed human motion prediction in context. In: European Conference on Computer Vision. pp. 676-694. Springer (2022)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 248, + 482, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 248, + 482, + 281 + ], + "spans": [ + { + "bbox": [ + 130, + 248, + 482, + 281 + ], + "type": "text", + "content": "62. Zhou, Y., Barnes, C., Lu, J., Yang, J., Li, H.: On the continuity of rotation representations in neural networks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5745-5753 (2019)" + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 223, + 91, + 447, + 102 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 223, + 91, + 447, + 102 + ], + "spans": [ + { + "bbox": [ + 223, + 91, + 447, + 102 + ], + "type": "text", + "content": "Revisit Human-Scene Interaction via Space Occupancy" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 480, + 100 + ], + "type": "text", + "content": "19" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 18 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file